diff --git a/.gitattributes b/.gitattributes index e844e6b387aabfb1d51475db6073caa378d06027..e2c1c6eca54fed8ef0abf58d2593dc26a6532f3b 100644 --- a/.gitattributes +++ b/.gitattributes @@ -3961,3 +3961,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2023/TCOVIS_[[:space:]]Temporally[[:space:]]Consistent[[:space:]]Online[[:space:]]Video[[:space:]]Instance[[:space:]]Segmentation/806a2113-930c-496f-99e8-ccd02b9307f7_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/TEMPO_[[:space:]]Efficient[[:space:]]Multi-View[[:space:]]Pose[[:space:]]Estimation,[[:space:]]Tracking,[[:space:]]and[[:space:]]Forecasting/5ca0ee4c-d0bf-43fa-950c-91f169b1d361_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/TF-ICON_[[:space:]]Diffusion-Based[[:space:]]Training-Free[[:space:]]Cross-Domain[[:space:]]Image[[:space:]]Composition/082c8f3d-29c9-4e4b-9ab4-3aa55681552e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TIFA_[[:space:]]Accurate[[:space:]]and[[:space:]]Interpretable[[:space:]]Text-to-Image[[:space:]]Faithfulness[[:space:]]Evaluation[[:space:]]with[[:space:]]Question[[:space:]]Answering/ae7f1ffe-fe67-4917-84ae-5c196628c6e5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TIJO_[[:space:]]Trigger[[:space:]]Inversion[[:space:]]with[[:space:]]Joint[[:space:]]Optimization[[:space:]]for[[:space:]]Defending[[:space:]]Multimodal[[:space:]]Backdoored[[:space:]]Models/7bfffc83-27ef-4c73-aa19-9a15a6ee13fd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TM2D_[[:space:]]Bimodality[[:space:]]Driven[[:space:]]3D[[:space:]]Dance[[:space:]]Generation[[:space:]]via[[:space:]]Music-Text[[:space:]]Integration/1834d63b-e3e0-4b42-9900-788dbb6bee22_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TMA_[[:space:]]Temporal[[:space:]]Motion[[:space:]]Aggregation[[:space:]]for[[:space:]]Event-based[[:space:]]Optical[[:space:]]Flow/9d269c18-5110-4a81-a1b9-317c1455548e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TMR_[[:space:]]Text-to-Motion[[:space:]]Retrieval[[:space:]]Using[[:space:]]Contrastive[[:space:]]3D[[:space:]]Human[[:space:]]Motion[[:space:]]Synthesis/d1f3cbdb-e07a-4952-8052-a9aa7bed81da_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TORE_[[:space:]]Token[[:space:]]Reduction[[:space:]]for[[:space:]]Efficient[[:space:]]Human[[:space:]]Mesh[[:space:]]Recovery[[:space:]]with[[:space:]]Transformer/2d762d0d-bad4-403d-98f7-4210b7f7e39d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TRM-UAP_[[:space:]]Enhancing[[:space:]]the[[:space:]]Transferability[[:space:]]of[[:space:]]Data-Free[[:space:]]Universal[[:space:]]Adversarial[[:space:]]Perturbation[[:space:]]via[[:space:]]Truncated[[:space:]]Ratio[[:space:]]Maximization/6ac05fcd-de6a-4657-a3f2-a897a887dcbb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Take-A-Photo_[[:space:]]3D-to-2D[[:space:]]Generative[[:space:]]Pre-training[[:space:]]of[[:space:]]Point[[:space:]]Cloud[[:space:]]Models/e286768b-a5b2-4ea3-a9f1-4d06c6cc68f6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Talking[[:space:]]Head[[:space:]]Generation[[:space:]]with[[:space:]]Probabilistic[[:space:]]Audio-to-Visual[[:space:]]Diffusion[[:space:]]Priors/7d213d41-b6ba-4fcf-a556-dbfb8dc18cb6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Taming[[:space:]]Contrast[[:space:]]Maximization[[:space:]]for[[:space:]]Learning[[:space:]]Sequential,[[:space:]]Low-latency,[[:space:]]Event-based[[:space:]]Optical[[:space:]]Flow/ff7524e3-aa7d-4baa-a842-04712fa34f60_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Tangent[[:space:]]Model[[:space:]]Composition[[:space:]]for[[:space:]]Ensembling[[:space:]]and[[:space:]]Continual[[:space:]]Fine-tuning/1ae449c1-93a5-4652-bb27-6a83c99dd85c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Tangent[[:space:]]Sampson[[:space:]]Error_[[:space:]]Fast[[:space:]]Approximate[[:space:]]Two-view[[:space:]]Reprojection[[:space:]]Error[[:space:]]for[[:space:]]Central[[:space:]]Camera[[:space:]]Models/636b4c99-6986-4080-bf56-9a578e79f107_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Task[[:space:]]Agnostic[[:space:]]Restoration[[:space:]]of[[:space:]]Natural[[:space:]]Video[[:space:]]Dynamics/e82ebeb9-3e53-42eb-9d7f-02ba103a6afc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Task-Oriented[[:space:]]Multi-Modal[[:space:]]Mutual[[:space:]]Leaning[[:space:]]for[[:space:]]Vision-Language[[:space:]]Models/b6090cd5-35b5-49d1-9775-853efd91c055_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Task-aware[[:space:]]Adaptive[[:space:]]Learning[[:space:]]for[[:space:]]Cross-domain[[:space:]]Few-shot[[:space:]]Learning/d73ef7d6-b7de-4511-b8bf-4b4e20276122_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TaskExpert_[[:space:]]Dynamically[[:space:]]Assembling[[:space:]]Multi-Task[[:space:]]Representations[[:space:]]with[[:space:]]Memorial[[:space:]]Mixture-of-Experts/2e3b51ba-4260-4b59-a865-225826178569_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Taxonomy[[:space:]]Adaptive[[:space:]]Cross-Domain[[:space:]]Adaptation[[:space:]]in[[:space:]]Medical[[:space:]]Imaging[[:space:]]via[[:space:]]Optimization[[:space:]]Trajectory[[:space:]]Distillation/1f4cb2e9-7a0d-4b5a-99a2-329016955f96_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TeD-SPAD_[[:space:]]Temporal[[:space:]]Distinctiveness[[:space:]]for[[:space:]]Self-Supervised[[:space:]]Privacy-Preservation[[:space:]]for[[:space:]]Video[[:space:]]Anomaly[[:space:]]Detection/a12231ff-0413-4e22-8eb5-06598dd6fed6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Teaching[[:space:]]CLIP[[:space:]]to[[:space:]]Count[[:space:]]to[[:space:]]Ten/56d47f2c-86af-49f0-9e72-8decf5e81c8a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Tem-Adapter_[[:space:]]Adapting[[:space:]]Image-Text[[:space:]]Pretraining[[:space:]]for[[:space:]]Video[[:space:]]Question[[:space:]]Answer/4a5fd117-e0c8-457c-9985-c705ba827466_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Template[[:space:]]Inversion[[:space:]]Attack[[:space:]]against[[:space:]]Face[[:space:]]Recognition[[:space:]]Systems[[:space:]]using[[:space:]]3D[[:space:]]Face[[:space:]]Reconstruction/203e5250-94d9-4833-ba8d-b9c548251baa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Template-guided[[:space:]]Hierarchical[[:space:]]Feature[[:space:]]Restoration[[:space:]]for[[:space:]]Anomaly[[:space:]]Detection/b5ce36c9-1639-4be6-bce9-ead12d7207f0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Temporal[[:space:]]Collection[[:space:]]and[[:space:]]Distribution[[:space:]]for[[:space:]]Referring[[:space:]]Video[[:space:]]Object[[:space:]]Segmentation/3b08deeb-33eb-47b0-94a4-e0af89436d2d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Temporal[[:space:]]Enhanced[[:space:]]Training[[:space:]]of[[:space:]]Multi-view[[:space:]]3D[[:space:]]Object[[:space:]]Detector[[:space:]]via[[:space:]]Historical[[:space:]]Object[[:space:]]Prediction/553f0662-4b8d-4936-ab47-ad291ff03dd2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Temporal-Coded[[:space:]]Spiking[[:space:]]Neural[[:space:]]Networks[[:space:]]with[[:space:]]Dynamic[[:space:]]Firing[[:space:]]Threshold_[[:space:]]Learning[[:space:]]with[[:space:]]Event-Driven[[:space:]]Backpropagation/6f1f26ae-7b19-4337-978a-c5b559913901_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Test[[:space:]]Time[[:space:]]Adaptation[[:space:]]for[[:space:]]Blind[[:space:]]Image[[:space:]]Quality[[:space:]]Assessment/4f23b86f-f4db-4257-aee1-6b77c4f73657_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Test-time[[:space:]]Personalizable[[:space:]]Forecasting[[:space:]]of[[:space:]]3D[[:space:]]Human[[:space:]]Poses/205eafc7-d672-437b-9b70-81a2995fa72d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Tetra-NeRF_[[:space:]]Representing[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]Using[[:space:]]Tetrahedra/8c410ac7-5eb0-4f17-a4bb-daadb2446edc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TexFusion_[[:space:]]Synthesizing[[:space:]]3D[[:space:]]Textures[[:space:]]with[[:space:]]Text-Guided[[:space:]]Image[[:space:]]Diffusion[[:space:]]Models/698eb5a8-8d74-4b19-b9a7-1fe068b4d1fd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Text-Conditioned[[:space:]]Sampling[[:space:]]Framework[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Generation[[:space:]]with[[:space:]]Masked[[:space:]]Generative[[:space:]]Models/c18683e6-ab91-4cf4-bf6b-203837440b00_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Text-Driven[[:space:]]Generative[[:space:]]Domain[[:space:]]Adaptation[[:space:]]with[[:space:]]Spectral[[:space:]]Consistency[[:space:]]Regularization/e13295e4-a033-4d38-a3a5-88f3b3385e81_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Text2Performer_[[:space:]]Text-Driven[[:space:]]Human[[:space:]]Video[[:space:]]Generation/45c9b878-7941-4c19-aa7b-acd77ad00875_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Text2Room_[[:space:]]Extracting[[:space:]]Textured[[:space:]]3D[[:space:]]Meshes[[:space:]]from[[:space:]]2D[[:space:]]Text-to-Image[[:space:]]Models/ec345001-5311-431e-90bd-6b91f4b47f2a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Text2Tex_[[:space:]]Text-driven[[:space:]]Texture[[:space:]]Synthesis[[:space:]]via[[:space:]]Diffusion[[:space:]]Models/8023656a-1169-49dd-b8cd-4db39941377d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Text2Video-Zero_[[:space:]]Text-to-Image[[:space:]]Diffusion[[:space:]]Models[[:space:]]are[[:space:]]Zero-Shot[[:space:]]Video[[:space:]]Generators/56c5c84c-dd92-4857-8492-3e83266b25fc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TextManiA_[[:space:]]Enriching[[:space:]]Visual[[:space:]]Feature[[:space:]]by[[:space:]]Text-driven[[:space:]]Manifold[[:space:]]Augmentation/21a8728d-cd5e-44be-95fc-d79147ae763b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TextPSG_[[:space:]]Panoptic[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation[[:space:]]from[[:space:]]Textual[[:space:]]Descriptions/a412d306-b042-443c-958a-505db74fd5e7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Texture[[:space:]]Generation[[:space:]]on[[:space:]]3D[[:space:]]Meshes[[:space:]]with[[:space:]]Point-UV[[:space:]]Diffusion/295801ea-f88d-446b-aeea-24167c36963d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Texture[[:space:]]Learning[[:space:]]Domain[[:space:]]Randomization[[:space:]]for[[:space:]]Domain[[:space:]]Generalized[[:space:]]Segmentation/5378f03d-031c-45ab-a522-6e62326e01af_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Devil[[:space:]]is[[:space:]]in[[:space:]]the[[:space:]]Crack[[:space:]]Orientation_[[:space:]]A[[:space:]]New[[:space:]]Perspective[[:space:]]for[[:space:]]Crack[[:space:]]Detection/b4b142a9-63c6-4b31-9bbd-a4e1e31c2d72_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Devil[[:space:]]is[[:space:]]in[[:space:]]the[[:space:]]Upsampling_[[:space:]]Architectural[[:space:]]Decisions[[:space:]]Made[[:space:]]Simpler[[:space:]]for[[:space:]]Denoising[[:space:]]with[[:space:]]Deep[[:space:]]Image[[:space:]]Prior/6b883d6c-8e38-4d25-996f-79941abad307_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Effectiveness[[:space:]]of[[:space:]]MAE[[:space:]]Pre-Pretraining[[:space:]]for[[:space:]]Billion-Scale[[:space:]]Pretraining/5b230860-aa61-4f33-9fe7-2407ba62f713_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Euclidean[[:space:]]Space[[:space:]]is[[:space:]]Evil_[[:space:]]Hyperbolic[[:space:]]Attribute[[:space:]]Editing[[:space:]]for[[:space:]]Few-shot[[:space:]]Image[[:space:]]Generation/e705b187-21ba-4812-b0e4-1feb752f1cdd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Making[[:space:]]and[[:space:]]Breaking[[:space:]]of[[:space:]]Camouflage/b706bafb-dae1-4647-ad2f-702e05fcaa82_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Perils[[:space:]]of[[:space:]]Learning[[:space:]]From[[:space:]]Unlabeled[[:space:]]Data_[[:space:]]Backdoor[[:space:]]Attacks[[:space:]]on[[:space:]]Semi-supervised[[:space:]]Learning/b27d12fb-927b-406e-bce3-95c8be702c01_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Power[[:space:]]of[[:space:]]Sound[[:space:]](TPoS)_[[:space:]]Audio[[:space:]]Reactive[[:space:]]Video[[:space:]]Generation[[:space:]]with[[:space:]]Stable[[:space:]]Diffusion/3b9298a9-5799-4904-96da-bef6d287733c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Stable[[:space:]]Signature_[[:space:]]Rooting[[:space:]]Watermarks[[:space:]]in[[:space:]]Latent[[:space:]]Diffusion[[:space:]]Models/c2da89cd-17c8-420c-b57b-7af434285ab7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Unreasonable[[:space:]]Effectiveness[[:space:]]of[[:space:]]Large[[:space:]]Language-Vision[[:space:]]Models[[:space:]]for[[:space:]]Source-Free[[:space:]]Video[[:space:]]Domain[[:space:]]Adaptation/c596fbdc-cd07-43b2-9a08-9301d722c491_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/The[[:space:]]Victim[[:space:]]and[[:space:]]The[[:space:]]Beneficiary_[[:space:]]Exploiting[[:space:]]a[[:space:]]Poisoned[[:space:]]Model[[:space:]]to[[:space:]]Train[[:space:]]a[[:space:]]Clean[[:space:]]Model[[:space:]]on[[:space:]]Poisoned[[:space:]]Data/fdfec7d9-d3ec-4468-9f42-c0ec2550345d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Theoretical[[:space:]]and[[:space:]]Numerical[[:space:]]Analysis[[:space:]]of[[:space:]]3D[[:space:]]Reconstruction[[:space:]]Using[[:space:]]Point[[:space:]]and[[:space:]]Line[[:space:]]Incidences/5fbfedc8-4ebc-4e53-8cf0-a4029a8a49ba_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Thinking[[:space:]]Image[[:space:]]Color[[:space:]]Aesthetics[[:space:]]Assessment_[[:space:]]Models,[[:space:]]Datasets[[:space:]]and[[:space:]]Benchmarks/42e3e8c9-44a4-4bef-b505-a2c5b22ff20f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TiDAL_[[:space:]]Learning[[:space:]]Training[[:space:]]Dynamics[[:space:]]for[[:space:]]Active[[:space:]]Learning/cbb37e36-1eb0-42b7-b0b1-3b2793b016b9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TiDy-PSFs_[[:space:]]Computational[[:space:]]Imaging[[:space:]]with[[:space:]]Time-Averaged[[:space:]]Dynamic[[:space:]]Point-Spread-Functions/0fbb2c88-7b44-409c-9798-461b2922052e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Tiled[[:space:]]Multiplane[[:space:]]Images[[:space:]]for[[:space:]]Practical[[:space:]]3D[[:space:]]Photography/332ad0e6-c530-4dba-82ae-058961745db1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Time[[:space:]]Does[[:space:]]Tell_[[:space:]]Self-Supervised[[:space:]]Time-Tuning[[:space:]]of[[:space:]]Dense[[:space:]]Image[[:space:]]Representations/1b69a328-51e9-4f52-9cd0-ceb8e6de4113_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Time-to-Contact[[:space:]]Map[[:space:]]by[[:space:]]Joint[[:space:]]Estimation[[:space:]]of[[:space:]]Up-to-Scale[[:space:]]Inverse[[:space:]]Depth[[:space:]]and[[:space:]]Global[[:space:]]Motion[[:space:]]using[[:space:]]a[[:space:]]Single[[:space:]]Event[[:space:]]Camera/0f2857ef-2537-43f4-b49d-a3319a025266_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Tiny[[:space:]]Updater_[[:space:]]Towards[[:space:]]Efficient[[:space:]]Neural[[:space:]]Network-Driven[[:space:]]Software[[:space:]]Updating/23a6116f-5b95-4dcd-97e5-95528a6d1fff_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TinyCLIP_[[:space:]]CLIP[[:space:]]Distillation[[:space:]]via[[:space:]]Affinity[[:space:]]Mimicking[[:space:]]and[[:space:]]Weight[[:space:]]Inheritance/7a0fee57-111a-443f-8a63-68f0c3d1c0b8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/To[[:space:]]Adapt[[:space:]]or[[:space:]]Not[[:space:]]to[[:space:]]Adapt_[[:space:]]Real-Time[[:space:]]Adaptation[[:space:]]for[[:space:]]Semantic[[:space:]]Segmentation/d5cc5a6e-449b-4085-98b6-42b0071a0627_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Token-Label[[:space:]]Alignment[[:space:]]for[[:space:]]Vision[[:space:]]Transformers/ef46bf1f-34a3-4e70-bf1a-b3f5f27b821d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Too[[:space:]]Large;[[:space:]]Data[[:space:]]Reduction[[:space:]]for[[:space:]]Vision-Language[[:space:]]Pre-Training/bf4cd6c9-c260-42c2-88a5-f694382bd718_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/ToonTalker_[[:space:]]Cross-Domain[[:space:]]Face[[:space:]]Reenactment/d3014894-36d6-4bc9-bdc4-403449ecf9bd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/TopoSeg_[[:space:]]Topology-Aware[[:space:]]Nuclear[[:space:]]Instance[[:space:]]Segmentation/a134e8ff-997a-47c5-810d-0149806d46fc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Total-Recon_[[:space:]]Deformable[[:space:]]Scene[[:space:]]Reconstruction[[:space:]]for[[:space:]]Embodied[[:space:]]View[[:space:]]Synthesis/49fdb57e-3f25-4f0f-85ac-577ce80d4f48_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/ae7f1ffe-fe67-4917-84ae-5c196628c6e5_content_list.json b/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/ae7f1ffe-fe67-4917-84ae-5c196628c6e5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..02bb18bf0ee91b6ebb08f2c781a0e4a16714de5e --- /dev/null +++ b/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/ae7f1ffe-fe67-4917-84ae-5c196628c6e5_content_list.json @@ -0,0 +1,3206 @@ +[ + { + "type": "text", + "text": "TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering", + "text_level": 1, + "bbox": [ + 174, + 130, + 795, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yushi Hu1 Benlin Liu1 Jungo Kasai1 Yizhong Wang1 Mari Ostendorf1 Ranjay Krishna1,2 Noah A. Smith1,2 1University of Washington 2Allen Institute for AI https://tifa-benchmark.github.io/", + "bbox": [ + 230, + 202, + 730, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite thousands of researchers, engineers, and artists actively working on improving text-to-image generation models, systems often fail to produce images that accurately align with the text inputs. We introduce TIFA (Text-to-Image Faithfulness evaluation with question Answering), an automatic evaluation metric that measures the faithfulness of a generated image to its text input via visual question answering (VQA). Specifically, given a text input, we automatically generate several question-answer pairs using a language model. We calculate image faithfulness by checking whether existing VQA models can answer these questions using the generated image. TIFA is a reference-free metric that allows for fine-grained and interpretable evaluations of generated images. TIFA also has better correlations with human judgments than existing metrics. Based on this approach, we introduce TIFA v1.0, a benchmark consisting of 4K diverse text inputs and 25K questions across 12 categories (object, counting, etc.). We present a comprehensive evaluation of existing text-to-image models using TIFA v1.0 and highlight the limitations and challenges of current models. For instance, we find that current text-to-image models, despite doing well on color and material, still struggle in counting, spatial relations, and composing multiple objects. We hope our benchmark will help carefully measure the research progress in text-to-image synthesis and provide valuable insights for further research.1", + "bbox": [ + 75, + 340, + 473, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 746, + 209, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While we welcome artistic freedom when we commission art from artists, images produced by deep generative models [44, 46, 43, 47, 61] should conform closely to our requests. Despite the advances in generative models, it is still challenging for models to produce images faithful to users' intentions [40, 11, 30, 35, 36]. For example, current", + "bbox": [ + 75, + 772, + 470, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0534cb1a249c7e271671e89f343dc2e4cfa429059bd379d85675ff68f55c136b.jpg", + "image_caption": [ + "Stable Diffusion v1.5" + ], + "image_footnote": [ + "Text Input: A person sitting on a horse in air over gate in grass with people and trees in background." + ], + "bbox": [ + 557, + 321, + 687, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/984f040b7734452145bb0a652872438939cbafdb1886a69c1c3b86df988b4220.jpg", + "image_caption": [ + "Stable Diffusion v2.1" + ], + "image_footnote": [], + "bbox": [ + 707, + 321, + 839, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GPT-3 generated + verified QAs (pre-generated in TIFA v1.0 benchmark)", + "text_level": 1, + "bbox": [ + 542, + 454, + 848, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Question: what is the animal?", + "bbox": [ + 529, + 469, + 676, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Answer inferred from text: horse", + "bbox": [ + 687, + 469, + 851, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VQA:", + "bbox": [ + 553, + 489, + 586, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Horse", + "bbox": [ + 606, + 489, + 638, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2bf43cc660fd0cbecdd7a6a05464fa8a5a8de1811ed94b32e0d27b25c31d869c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 487, + 669, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Horse", + "bbox": [ + 761, + 489, + 792, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a103acb97b718e129bf14bd098cc04876cefcfa5e0ea8305c5e4819e6d873079.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 488, + 821, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Question: is there a gate?", + "bbox": [ + 529, + 510, + 656, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VQA:", + "bbox": [ + 553, + 526, + 586, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "No", + "bbox": [ + 612, + 527, + 630, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d87f16d37f52a8358d4b071322c5ce2b06a396dd438a98129ab3fecb1f8942dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 527, + 661, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Answer inferred from text: yes", + "bbox": [ + 689, + 510, + 841, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yes", + "bbox": [ + 766, + 527, + 787, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d0ee3238ee5c9725828c47e0a85ad4a3bd73aa5ab323447b1f80b863c4c3cc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 526, + 821, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Question: is the horse in air?", + "bbox": [ + 529, + 547, + 671, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VQA:", + "bbox": [ + 553, + 561, + 586, + 574 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5522b1600f668c49eac7cac3072645156e09fd921715ae997a7bbf79243f1d91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 561, + 630, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/12d21720e3a232654f85b354ee1e1e2e64a08e693896616df2d1b7e4d45a986e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 561, + 661, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd0192131f3afc4d22357421e1c9860357c0a4db85eb95034afaede286fe48f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 561, + 687, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a109c7eb64024dfb4799df213666bd0459271ee1743504b90ae74869342987f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 561, + 717, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a7d0228e334eb274d7795337a069b7531a0c9550f8e7779439ece13b3af090e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 583, + 545, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "TIFA 71.4", + "bbox": [ + 549, + 595, + 633, + 608 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/121579668e855fc1d5a60de5ef1fabadaf8939a1c28a315c435cc855e89249b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 608, + 630, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ined", + "bbox": [ + 638, + 612, + 658, + 621 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e2f53cc61ec55a112cf9006c4fdac613ac1e83011e8281fbca5e9a47ada3ddd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 612, + 687, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate", + "bbox": [ + 689, + 612, + 728, + 622 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/68c35f09e9d481791bf1b3bf74ab11598c76571421eaf99f060cfb1be6bc50aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 612, + 756, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Interpretable", + "bbox": [ + 759, + 612, + 818, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "v.s. other automatic image-text align metrics", + "text_level": 1, + "bbox": [ + 586, + 631, + 805, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CLIPScore", + "bbox": [ + 531, + 662, + 588, + 672 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/64ecd3c515b2c89600f1c498d79e8e943ff752319db31f78b2b77da92d1e0cfb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 646, + 656, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SPICE", + "bbox": [ + 702, + 661, + 740, + 671 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6d78a922c8fab6c04bbc48f56edfc903042ac82931bf4bb8d972f5dde7096d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 652, + 779, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 782, + 655, + 800, + 665 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sim. (caption, text input)", + "bbox": [ + 754, + 674, + 844, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CLIP:24.3 SPICE:22.2", + "bbox": [ + 531, + 691, + 655, + 702 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a4976d5848b68e554a64a723fd69f112fc97a4dfe505b52d15689e7e7604e2dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 691, + 696, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CLIP: 21.3 SPICE: 11.8", + "bbox": [ + 723, + 691, + 848, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. Illustration of how TIFA works, and comparison with the widely-used CLIPScore and SPICE metrics. Given the text input, TIFA uses GPT-3 to generate several question-answer pairs, and a QA model filters them (3 out of 14 questions for this text input are shown). TIFA measures whether VQA models can accurately answer these questions given the generated image. In this example, TIFA indicates that the image generated by Stable Diffusion v2.1 is better than that by v1.5, while CLIP and SPICE yield the opposite result. The text input is from the MSCOCO validation set.", + "bbox": [ + 496, + 709, + 893, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "models often fail to compose multiple objects [40, 11, 35], bind attributes to the wrong objects [11], and struggle in generating visual text [36]. Today, there are efforts to address", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": ". All data and a pip-installable evaluation package are available on the project page.", + "bbox": [ + 75, + 875, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20406", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3cc23a705a88ae78a1e31f868561e4e485461978a9e8587b0d15aab95d63115c.jpg", + "image_caption": [ + "Figure 2. (a) Overview of how TIFA evaluates the faithfulness of a synthesized image. TIFA uses a language model (LM), a question-answering (QA) model, and a visual-question-answering (VQA) model. Given a text input, we generate several question-answer pairs with the LM and then filter them via the QA model. To evaluate the faithfulness of a synthesized image to the text input, a VQA model answers these visual questions using the image, and we check the answers for correctness. (b) TIFA v1.0 benchmark. While TIFA is applicable to any text prompt, to allow direct comparison across different studies, and for ease of use, we introduce the TIFA v1.0 benchmark, a repository of text inputs along with pre-generated question-answer tuples with answer choices. To evaluate a text-to-image model, a user first produces the images for the text inputs in TIFA v1.0 and then performs VQA with our provided tools on generated images to compute TIFA." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 560, + 246 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/002e14dbc71ea718f6bd90b0b7376dd20db7e1bea3d060e78ff49d5bea7299a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 89, + 869, + 247 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "these challenges: researchers are imposing linguistic structure with diffusion guidance to produce images with multiple objects [11]; others are designing reward models trained using human feedback to better align generations with user intention [30]. However, progress is difficult to quantify without accurate and interpretable evaluation measures that explain when and how models struggle.", + "bbox": [ + 75, + 358, + 470, + 464 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A critical bottleneck, therefore, is the lack of reliable automatic evaluation metrics for text-to-image generation faithfulness. One of the popular metrics is CLIPScore [17], which measures the cosine similarity between the CLIP embeddings [42] of the text input and the generated image. However, since CLIP is not effective at counting objects [42], or reasoning compositionally [37], CLIPScore is unreliable and often inaccurate. Another family of evaluation metrics uses image captions, in which an image captioning model first converts the image into text, and then the image caption is evaluated by comparing it against the text input. Unfortunately, using captioning models is insufficient since they might decide to ignore salient information in images or focus on other non-essential image regions [24]; for example, a captioning model might say that the images in Figure 1 are \"a field of grass with trees in the background\". Moreover, evaluating text (caption) generation is inherently challenging [23, 26]. Another recent text-to-image evaluation is DALL-Eval [6], which employs object detection to determine if the objects in the texts are in the generated images. However, this approach only works on synthesized text and measures faithfulness along the limited axes of objects, counting, colors, and spatial relationships but misses activities, geolocation, weather, time, materials, shapes, sizes, and other potential categories we often ask about when we recall images from memory [29].", + "bbox": [ + 75, + 470, + 470, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above challenges, we introduce TIFA, a new metric to evaluate text-to-image generation faithfulness.", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach is illustrated in Figure 2. Given a repository of text inputs, we automatically generate question-answer pairs for each text via a language model (here, GPT-3 [3]). A question-answering (QA) system (here, UnifiedQA [25]) is subsequently used to verify and filter these question-answer pairs. To evaluate a generated image, we use a visual-question-answering (VQA) system (here, mPLUGlarge [31], BLIP-2 [32], etc.) to answer the questions given the generated image. We measure the image's faithfulness to the text input as the accuracy of the answers generated by the VQA system. While the accuracy of TIFA is dependent on the accuracy of the VQA model, our experiments show that TIFA has much higher correlation with human judgments than CLIPScore (Spearman's $\\rho = 0.60$ vs. 0.33) and captioning-based approaches (Spearman's $\\rho = 0.60$ vs. 0.34). Additionally, since the LMs and VQA models will continue to improve, we hypothesize that TIFA will continue to be more reliable over time. Also, our metrics can automatically detect when elements are missing in the generation: in Figure 2, TIFA detects that the generated image does not contain a TV.", + "bbox": [ + 496, + 358, + 893, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To promote the use of our new evaluation metric, we release TIFA v1.0, a large-scale text-to-image generation benchmark containing 4K diverse text inputs, sampled from the MSCOCO captions [34], DrawBench [47], PartiPrompts [61], and PaintSkill [6]. Each input comes with a pre-generated set of question-answer pairs, resulting in 25K questions covering 4.5K distinct elements. These questions have been automatically generated and pre-filtered using a question-answering model. This benchmark also comes with different VQA models [57, 28, 58, 32, 31, 21] that can be used to evaluate generative models and can be easily extended to use future VQA models when they become available.", + "bbox": [ + 496, + 681, + 893, + 876 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct a comprehensive evaluation of current text-", + "bbox": [ + 519, + 885, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20407", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to-image models using TIFA v1.0. Thanks to TIFA's ability to detect fine-grained unfaithfulness in images, we find that current state-of-the-art models are good at rendering common objects, animals, and colors, but still struggle in composing multiple objects, reasoning about spatial relations, and binding the correct activity for each entity. In addition, our ablation experiments show that TIFA is robust to different VQA models. Future researchers can use TIFA v1.0 to compare their text-to-image models' faithfulness across different studies. Also, future generative models may focus on addressing the weaknesses of current models that TIFA discovered. In addition, with TIFA, users can customize evaluations with their own text inputs and questions [10]; for example, a future TIFA benchmark could focus on counting or scene text.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 330, + 218, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We compare TIFA to other image and language generation evaluation metrics.", + "bbox": [ + 75, + 356, + 470, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prior image generation evaluation Prior work usually compares image generation models via pairwise comparison by humans. How to design automatic evaluation metrics to approximate human assessment of the quality of machine-generated images has always been a major challenge in computer vision. There are two aspects to evaluate, namely image quality and image-text alignment. Inception Score [48] and FID [18] are the most widely adopted metrics for image quality. They compare the features of the generated images and gold images extracted from a pre-trained Inception-V3 model [52] to evaluate the fidelity and diversity of generated images. However, they rely on ground-truth images and are based on a classification model, which makes them not suitable for complex datasets [12]. For image-text alignment, prior metrics are mainly based on CLIP [42], captioning, and object detection models. CLIPScore [17] and CLIP-R [39] are based on the cosine similarity of image and text CLIP [42] embeddings. [6, 19, 20] first convert the images using a captioning model, and then compare the image caption with the text using metrics like CIDEr [55] and SPICE [1]. SOA [19] and DALL-Eval [6] employ object detection models to determine if objects, attributes, and relations in the text input are in the generated image. However, this approach only works on synthesized text inputs and measures faithfulness on limited axes (object, counting, color, spatial relation), missing elements like material, shape, activities, and context. In contrast, thanks to the flexibility of questions, TIFA works on any text inputs and evaluates faithfulness across a broad spectrum of dimensions.", + "bbox": [ + 75, + 387, + 472, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Summarization evaluation in NLP TIFA is inspired by the summarization evaluation methods based on question answering (QA) [56, 50]. Given a summary, a language model generates a set of questions about the text. A QA model checks if the same answer can be inferred from the", + "bbox": [ + 75, + 824, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "text and the summary. These QA-based metrics have much higher correlations with human judgments on the factual consistency of summarization than other automatic metrics [56, 50]. TIFA can be seen as treating the text input for the text-to-image model as a summary of the generated image.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The TIFA Metric", + "text_level": 1, + "bbox": [ + 498, + 180, + 669, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a framework for automatically estimating the faithfulness of an image to its text prompt. Given a text input $T$ , we aim to measure the faithfulness of the generated image $I$ . An overview of our metric is illustrated in Figure 2. From $T$ , we generate $N$ multiple-choice question-answer tuples $\\{Q_i, C_i, A_i\\}_{i=1}^N$ , in which $Q_i$ is a question, $C_i$ is a set of answer choices, and $A_i \\in C_i$ is the gold answer. The answer $A_i$ can be inferred given $T, Q_i$ , and $C_i$ . Next, for each question $Q_i$ , we use a VQA model to produce an answer $A_i^{\\mathrm{VQA}} = \\max_{a \\in C_i} p(a \\mid I, Q_i)$ . We define the faithfulness between the text $T$ and image $I$ as the VQA accuracy:", + "bbox": [ + 496, + 205, + 893, + 372 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf a i t h f u l n e s s (T, I) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbb {1} \\left[ A _ {i} ^ {\\mathrm {V Q A}} = A _ {i} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 382, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The range of our faithfulness score is $[0, 1]$ . It is maximized when we have a performant VQA model, and the image $I$ accurately covers the information in the text $T$ so that for any question $Q$ , which can be answered given $T$ can also be answered given $I$ . Several key design decisions will be addressed in later sections: how to generate questions ( $\\S 3.1$ ), how to control the question quality ( $\\S 3.2$ ), and how to answer those questions ( $\\S 3.3$ ). Finally, we give a step-by-step qualitative example of TIFA in Figure 4.", + "bbox": [ + 496, + 425, + 893, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Question-Answer Generation", + "text_level": 1, + "bbox": [ + 500, + 571, + 761, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our main challenge is to generate diverse questions that cover all elements of the text input evenly. We also simplify the question-generation pipeline into a single GPT-3 [3] completion, so that TIFA can exploit the power of recent language models (LM) and work with updated black-box LMs (e.g., ChatGPT) in the future.", + "bbox": [ + 496, + 594, + 890, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by prior work [4], given a text prompt $T$ , we generate the question-answer tuples $\\{Q_i, C_i, A_i\\}_{i=1}^N$ via the pipeline illustrated in Figure 3. Different from prior work, which relies on multiple components, our pipeline is completed by a single inference run via in-context learning with GPT-3 [3, 59, 22, 41, 51], thereby avoiding the need for intermediate human annotations. We annotate 15 examples and use them as in-context examples for GPT-3 to follow. Here we take the text \"A photo of three dogs\" as an example. Each in-context example contains the following steps:", + "bbox": [ + 496, + 685, + 893, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Element extraction Given text prompt $T$ , GPT-3 will first extract all elements $\\{v_i\\}_{i=1}^m$ following prior work [4] (for the in-context examples, we perform element extraction manually). The elements include noun phrases (including named", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "20408", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6a1664ffedb426de384e4b27c58ce42eaa33029af954edf04727c5235e2a36cf.jpg", + "image_caption": [ + "Figure 3. Our question-answer pair generation pipeline. The whole pipeline can be executed via a single inference of GPT-3 via in-context learning. Given the text prompt, GPT-3 first extracts the elements and then generates two questions for each element. The GPT-3 output is then parsed and filtered by UnifiedQA." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 444, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "entities), verbs, adjectives, adverbs, and parse tree spans with no more than 3 words altogether. For the above example, the elements are photo, three, dogs.", + "bbox": [ + 75, + 513, + 470, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Element category classification For each element $v_{i}$ , following [29], we classify the elements into one of the following 12 categories: object, activity, animal, food, counting, color, material, spatial, location, shape, attribute, and other. As shown in Figure 3, text generated from GPT-3 contains the category corresponding to each question. For example, \"three\" is \"counting\", and \"dogs\" is classified as \"animal.\" This step allows a detailed analysis of the text-to-image model's ability in each category.", + "bbox": [ + 75, + 566, + 470, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Question generation conditioned on elements For each element $v_{i}$ , we generate two questions. The first is a question that should be answered \"yes\" for a faithful generated image, and the second question has $v_{i}$ as its answer. For example, two questions are generated for the element \"three\". The first is \"are there three dogs?\", and the choices are {yes, no}. The second is \"how many dogs are there?\", and the choices are {1,2,3,4}. These two types of questions make our evaluations diverse and robust to surface-level differences.", + "bbox": [ + 75, + 710, + 470, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Completing above steps by prompting GPT-3 once As mentioned earlier, for each text $T$ , the whole pipeline can be completed by one GPT-3 inference. We annotated 15", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in-context examples that cover all types of questions. The full prompt is in the Appendix. The prompt format is shown in Figure 3. Our in-context examples follow the same format, and identical examples are used for all text inputs, leading to a fixed and limited amount of human annotation cost. We use code-davinci-002 engine for question generation, and the decoding temperature is 0.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Question Filtering", + "text_level": 1, + "bbox": [ + 500, + 205, + 678, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure the quality of generated images, we use UnifiedQA [25] to verify the GPT-3 generated question-answer pairs and filter out the ones that GPT-3 and UnifiedQA do not agree on. UnifiedQA $^2$ is a state-of-the-art multi-task question-answering model that can answer both multiple-choice and free-form questions. Denote the UnifiedQA model as $QA$ . Given the text $T$ , question $Q_{i}$ , choices $C_{i}$ , and answer $A_{i}$ , Let $A_{i}^{f} = QA(T,Q_{i})$ be the free-form answer, and $A_{i}^{mc} = QA(T,Q_{i},C_{i})$ be the multiple-choice answer. We keep the question if $A_{i} = A_{i}^{mc}$ and the word-level $F_{1}$ score between $A_{i}^{f}$ and $A_{i}$ is greater than 0.7. We conduct a human evaluation on 1000 filtered question-answer pairs. Only 7 are considered not reasonable (e.g., generated choices do not include a correct answer). Details are in Appendix C.", + "bbox": [ + 496, + 229, + 893, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recommended VQA model Based on considerations over the accuracy, correlation with human judgments, and run time, we currently suggest using mPLUG-large as the VQA model for TIFA. Analysis is given in Section 5.4. Like the LM and QA components, the VQA component can be updated in the future as the technology improves.", + "bbox": [ + 496, + 448, + 893, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. VQA Models", + "text_level": 1, + "bbox": [ + 500, + 549, + 637, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since our questions contain a diverse set of visual elements (e.g., activity, art style), we use open-domain pre-trained vision-language models as our VQA model (rather than closed-class classification models fine-tuned on VQAv2 [14]). We provide tools to easily perform VQA on arbitrary images and questions, based on 5 state-of-the-art VQA models trained with distinct data and strategies.", + "bbox": [ + 496, + 571, + 893, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision-language models The general pre-trained vision-language model are GIT-large [57], VILT-B/32 [28], OFA-large [58], and mPLUG-large [31]. These models are pretrained on a large amount of image-text pairs, and downstream image-to-text tasks like image captioning and visual question answering. Notice that these models have not been trained to answer multiple-choice questions. For each question, we first decode the free-form answer and then choose the choice that has the highest similarity with the decoded answer, measured by SBERT [45]. Another model we use is BLIP-2 FlanT5-XL [32], in which a VIT [8] is connected with a frozen FlanT5 [7] via a lightweight transformer. This", + "bbox": [ + 496, + 684, + 893, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2Model checkpoint we use: https://huggingface.co/allenai/unifiedqa-v2-t5-large-1363200.", + "bbox": [ + 500, + 875, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20409", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text Input: A wet dog standing on a beach next to the ocean.", + "bbox": [ + 187, + 94, + 477, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 1: Question Generation + Filtering", + "text_level": 1, + "bbox": [ + 184, + 114, + 444, + 128 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ef4c9091aa61a427fc97e2b449681e134f288f7ee835c8dbc9c239353f60aabb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
About dog (animal/human)About ocean (location)About wet (attribute)
Q: Is this a dog?Q: Is there an ocean?Q: Is the dog wet?
A: YesA: YesA: Yes
Q: What animal is in the picture?Q: What type of place is this?Q: Is the dog wet or dry?
A: DogA: OceanA: Wet
About beach (location)About standing (activity)About next to (spatial relation)
Q: Is this a beach?Q: Is the dog standing?Q: Is the dog next to the ocean?
A: YesA: YesA: Yes
Q: What type of place is this?Q: What is the dog doing?Q: Is the dog next to or in the ocean?
A: BeachA: StandingA: Next to
", + "bbox": [ + 186, + 131, + 785, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 2: Visual Question Answering", + "text_level": 1, + "bbox": [ + 184, + 279, + 413, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d7f364f9ab8707cafc9c8b69df7c0068c1d8080b00eeda6d0f7fb9e5bbbfb393.jpg", + "image_caption": [ + "Stable Diffusion v1.5 image" + ], + "image_footnote": [], + "bbox": [ + 192, + 303, + 338, + 415 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0c1b2e85e33bca6cb3dac0200c8c07b2849ac37bc9f82b0bed02e94745197458.jpg", + "image_caption": [ + "Figure 4. Step-by-step qualitative example of TIFA metric. Given a text input, we first generate question-answer pairs and filter them. We strike through the questions filtered out by UnifiedQA. Then we run VQA models on the generated image to get the TIFA score." + ], + "image_footnote": [], + "bbox": [ + 196, + 449, + 225, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TIFA $7 / 10 = 0.7$", + "bbox": [ + 225, + 460, + 321, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "About dog (animal/human)", + "bbox": [ + 372, + 297, + 501, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: Is this a dog?", + "bbox": [ + 372, + 310, + 452, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A:Yes", + "bbox": [ + 372, + 321, + 406, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: What animal is in the picture?", + "bbox": [ + 372, + 335, + 531, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A: Dog", + "bbox": [ + 372, + 347, + 410, + 358 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5c7ee897bef39786145b2eb99461704d55798546ae145b39d7efdfbb077c88d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 316, + 517, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "About beach (location)", + "bbox": [ + 372, + 372, + 482, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: Is this a beach?", + "bbox": [ + 372, + 383, + 462, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A:Yes", + "bbox": [ + 372, + 396, + 406, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: What type of place is this?", + "bbox": [ + 372, + 407, + 514, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A:Beach", + "bbox": [ + 372, + 421, + 419, + 431 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a1527cc74f269bdf8891c2550978fbd1444cc40fff7bc6fbb0e2f3f664a8425e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 393, + 517, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "About ocean (location)", + "bbox": [ + 372, + 446, + 482, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: Is there an ocean?", + "bbox": [ + 372, + 457, + 473, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A:Yes", + "bbox": [ + 372, + 469, + 406, + 479 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/474e633457676a3587a2eff43a3d20eff1b2be1e0187a536884295bf2f1def45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 469, + 517, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "About standing (activity)", + "bbox": [ + 588, + 297, + 709, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: Is the dog standing?", + "bbox": [ + 588, + 310, + 700, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A: No", + "bbox": [ + 588, + 321, + 620, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "About wet (attribute)", + "bbox": [ + 588, + 345, + 689, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: Is the dog wet?", + "bbox": [ + 588, + 359, + 678, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A: No", + "bbox": [ + 588, + 371, + 620, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: Is the dog wet or dry?", + "bbox": [ + 588, + 383, + 709, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A: Dry", + "bbox": [ + 588, + 396, + 625, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "About next to (spatial relation)", + "bbox": [ + 588, + 420, + 733, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: Is the dog next to the ocean?", + "bbox": [ + 588, + 431, + 740, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A:Yes", + "bbox": [ + 588, + 444, + 622, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: Is the dog next to or in the ocean?", + "bbox": [ + 588, + 455, + 764, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A: Next to", + "bbox": [ + 588, + 469, + 640, + 479 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/154a1a6c0ae1168f586f12765bc1e45361922d68000e5b0eb7f41d5c0f5d602b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 321, + 772, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8f127a7f3ce2a7c1055b518c078a2a0abb59062804dd758d71a96d0aaeff9ebb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 368, + 772, + 383 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4ed20ead1959143d1542d950ff7df12fcc1713ed4b9df1e67f62cfa02cbe1fed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 393, + 772, + 407 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3af46fd7916c03519a8ac648491bd202dda28b23ee4f783ed31e744a6ab95dc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 441, + 772, + 455 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/da37f76ecdc14b69667ea3479244e558c11efe12da0ddbf6bf7dbb3cc8a55bd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 455, + 764, + 467 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2d49daa3451cc2ec8cda8af22572489f4e3be5e0679e6d69f362b8157a3f9d2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 469, + 772, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "model allows for performing multiple-choice VQA directly due to the flexibility of the LM.", + "bbox": [ + 75, + 534, + 468, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. FIFA v1.0: Benchmark for Text-to-Image Generation Faithfulness", + "text_level": 1, + "bbox": [ + 76, + 580, + 468, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce TIFA v1.0, a text-to-image generation faithfulness benchmark based on the evaluation method discussed in Section 3. The benchmark consists of 4,081 diverse text inputs paired with 25,829 question-answer pairs. Each question is classified into one of the categories discussed in Section 3.1. The benchmark also comes with Python pip-installable APIs to perform VQA with various state-of-the-art VQA models on arbitrary visual questions. The overall TIFA for each text-to-image model is computed by averaging TIFA scores of images generated from each text input in the benchmark.", + "bbox": [ + 75, + 623, + 470, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Text Collections", + "text_level": 1, + "bbox": [ + 76, + 800, + 236, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We collect 4,081 text inputs to benchmark text-to-image models' generation ability on diverse tasks. 2,000 text inputs are image captions from COCO validation set [34]. These captions have corresponding gold images. Since text-to-image models are often used to create abstract art, we also", + "bbox": [ + 75, + 824, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "collect 2,081 text inputs from previous works that do not correspond to any real image. All text inputs we use contain $\\geq 3$ words. We include 161 from DrawBench used in Imagen [47] (texts that are categorized as \"misspellings\" and \"rare words\" are removed); 1420 from PartiPrompt used in Parti [61] (texts in category \"abstract\" are removed); and 500 texts from PaintSkill used in DALL-Eval [6].", + "bbox": [ + 496, + 534, + 890, + 640 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cea8c2de7c988e0cee88fa94af21426a2267c1406064d78be40fcc353c752244.jpg", + "table_caption": [ + "Table 1. Statistics of FIFA v1.0." + ], + "table_footnote": [], + "table_body": "
Statistics
# of prompts4,081
- # of COCO captions2,000
- # of DrawBench, PartiPrompt, PaintSkill prompts2,081
# of questions25,829
- # of binary questions17,226
- # of multiple-choice questions8,603
avg. # of questions per prompt6.3
avg. # of words per prompt10.5
avg. # of elements per prompt4.3
", + "bbox": [ + 500, + 664, + 890, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Statistics and Diversity", + "text_level": 1, + "bbox": [ + 500, + 845, + 712, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 shows the basic statistics of the TIFA v1.0 benchmark. We demonstrate TIFA v1.0's diversity in Figure 5.", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20410", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a yellow t-shirt with a dog on it\"", + "bbox": [ + 91, + 95, + 183, + 114 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"a panda bear with aviator glasses on its head\"", + "bbox": [ + 89, + 128, + 184, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"A red colored car\"", + "bbox": [ + 88, + 162, + 186, + 171 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"One cat and one dog sitting on the grass\"", + "bbox": [ + 96, + 186, + 179, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6c61a26c3e469c31513795353603f160bebc79eae50b011355684b2a26063ae0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 205, + 95, + 233, + 116 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object", + "bbox": [ + 236, + 98, + 284, + 109 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "table,train,bus ...", + "bbox": [ + 236, + 109, + 326, + 116 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/84d5b5f257ac863bf4c194bda68a306d7cb6d83d35f3404dc03a3945095fc19a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 205, + 128, + 230, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial", + "bbox": [ + 236, + 128, + 290, + 138 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "above,below,right ..", + "bbox": [ + 236, + 138, + 333, + 147 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a87d4ea750f0a2a6aaf32c66dd68948e0719305c054ece6e1914e6bb092c4fd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 205, + 159, + 233, + 178 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Counting", + "bbox": [ + 236, + 157, + 300, + 169 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "two,three,several", + "bbox": [ + 236, + 170, + 336, + 178 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ecaa51dfa75b45fd02357fc3ec25e9187a5de3f1e14f26be11c52d1743718cc9.jpg", + "image_caption": [ + "Figure 5. Statistics and diversity of TIFA v1.0. The text inputs contain elements from 12 categories (e.g., object, spatial, and counting). We show the most common elements from each category. In addition, we also show some example text inputs on the sides." + ], + "image_footnote": [], + "bbox": [ + 205, + 189, + 230, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Food", + "bbox": [ + 236, + 189, + 269, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pizza, bananas ...", + "bbox": [ + 236, + 200, + 313, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a592dbfe7c077069367c89fbfd0094178622657e535606df89a3252a9b221c31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 98, + 370, + 116 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Animal/Human", + "bbox": [ + 375, + 98, + 470, + 108 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "man,woman,dog ...", + "bbox": [ + 379, + 109, + 457, + 117 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c7df1793d05328685c27b9822b2fb87d5c02eb4b10d09e2380b8a9dc5d6980f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 98, + 509, + 117 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attribute", + "bbox": [ + 517, + 98, + 586, + 108 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "large, small, young ...", + "bbox": [ + 517, + 109, + 617, + 117 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f23bfa4bb47ef71f24f07f2ef2984f4140817fd4e18026b42d2df91e241681f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 125, + 460, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TIFA v1.0", + "bbox": [ + 464, + 137, + 550, + 150 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4,081 text inputs 25,829 questions \n4,550 distinct elements 12 categories", + "bbox": [ + 393, + 157, + 583, + 178 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/002405c7513a53fd78d1b9f5eb5e76ff2af1dc6807e318d290b766de13f393c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 189, + 370, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Material", + "bbox": [ + 379, + 188, + 442, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "wooden,metal,glass", + "bbox": [ + 375, + 200, + 468, + 209 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3902b528b991b62460dda4e572fb9ef6e9ec734600dec1732ceb3fdb230250bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 188, + 514, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Shape", + "bbox": [ + 521, + 189, + 560, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "circular,square", + "bbox": [ + 519, + 200, + 601, + 209 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fccdb2ed2a7a22286cc81991007956b07612bbdb2d0f31087d2cbd69302a72e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 98, + 651, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Activity", + "bbox": [ + 658, + 98, + 723, + 108 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sitting.standing ...", + "bbox": [ + 658, + 109, + 756, + 117 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c7b57b159055b870cc30565355e60f8b78442d2cb03e1076794543b0623eeb09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 128, + 651, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Location", + "bbox": [ + 658, + 128, + 723, + 138 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "field,street,beach ...", + "bbox": [ + 656, + 140, + 759, + 148 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/abe62ee34e8e14d8d4f75fa01eeb7b06c591355111013fb98f786d9077ec3e5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 159, + 653, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Color", + "bbox": [ + 661, + 159, + 699, + 169 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "red,white,blue ...", + "bbox": [ + 660, + 171, + 743, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afb9daf7febb133252aba567033da63eb2106c6c7ea5461fe92df0678b7cab77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 190, + 653, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Other", + "bbox": [ + 661, + 189, + 700, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "punk, starry night", + "bbox": [ + 658, + 200, + 750, + 209 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"a painting of a fox", + "bbox": [ + 769, + 95, + 874, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in the style", + "bbox": [ + 781, + 104, + 851, + 113 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "starry night", + "bbox": [ + 781, + 113, + 851, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"A man running to", + "bbox": [ + 782, + 130, + 872, + 138 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "hit a tennis ball", + "bbox": [ + 782, + 138, + 870, + 147 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"a flag with a", + "bbox": [ + 787, + 162, + 864, + 170 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dinosaur on it\"", + "bbox": [ + 787, + 171, + 864, + 178 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"a photo of suitcase", + "bbox": [ + 777, + 188, + 880, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and chair; chair is", + "bbox": [ + 781, + 196, + 877, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "right to suitcase\"", + "bbox": [ + 781, + 205, + 877, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TIFA v1.0 contains questions about 4,550 distinct elements, which are categorized into 12 categories. The number of times each type of element occurs in the text input is object (7,854), animal/human (3,501), attribute (3,399), activity (2,851), spatial (2,265), location (1,840), color (1,743), counting (986), food (911), material (209), shape (69), and other (201). The \"other\" category includes notions often used in abstract art, such as \"starry night\" and \"steampunk.\" The accuracy of VQA on a particular genre measures the text-to-image model's ability in the corresponding aspect. Please refer to Appendix D for more details on TIFA v1.0.", + "bbox": [ + 75, + 261, + 472, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Finetuned Open-Source Language Model for Question Generation", + "text_level": 1, + "bbox": [ + 76, + 441, + 468, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For TIFA v1.0, we use GPT-3 to generate the questions. While benchmarking with TIFA v1.0 is a deterministic process, using TIFA to create a new benchmark might not be deterministic as the underlying question generator (GPT-3 in our case) might change privately. To promote deterministic benchmark generation, we fine-tune and release a LLaMA 2 (7B) [54] model that parses the captions and generates questions for arbitrary texts, using TIFA v1.0 questions as training examples.", + "bbox": [ + 75, + 482, + 470, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 638, + 209, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first show that TIFA has substantially higher correlations with human judgments than prior metrics on text-to-image faithfulness (§5.1). Then we present a comprehensive evaluation of existing text-to-image models using TIFA v1.0, highlighting the challenges of current text-to-image models (§5.3). Finally, we conduct an analysis of TIFA's robustness against different VQA models (§5.4). For all experiments, we use mPLUG as the VQA model for TIFA unless stated otherwise. The models we evaluate include AttnGAN [60], X-LXMERT [5], VQ-Diffusion [16], minDALL-E [27], and Stable Diffusion v1.1, v1.5, and v2.1 [46]. Details are in Appendix E.", + "bbox": [ + 75, + 665, + 472, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Correlation with Human Judgements", + "text_level": 1, + "bbox": [ + 500, + 260, + 821, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To compare TIFA with prior evaluation metrics, we first conduct human evaluations of the text-to-image models on the 1-5 Likert scale on text-to-image faithfulness. Then we compare TIFA with other metrics based on their correlation with human judgments.", + "bbox": [ + 496, + 284, + 890, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Likert scale on text-to-image faithfulness Annotators are asked to answer on a scale of 1 (worst) to 5 (best) to the question \"Does the image match the text?\" The detailed annotation guidelines are in Appendix C. Annotators are asked to focus on text-to-image faithfulness rather than image quality. The Likert scale should be based on how many elements in the text prompt are missed or misrepresented in the image. Objects are more important than attributes, relations, and activities. If an object is missed in the image, then all related attributes, activities, relations, etc. are also considered lost. An example is given in Figure 6.", + "bbox": [ + 496, + 368, + 893, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We collect annotations of 800 generated images on 160 text inputs from TIFA v1.0. For each prompt, we sample an image from the 5 most recent generative models we evaluated, i.e., minDALL-E, VQ-Diffusion, Stable Diffusion v1.1, v1.5, and v2.1. We collect 2 annotations per image and average over the scores as the single \"faithfulness\" score. The inter-annotator agreement measured by Krippendorf's $\\alpha$ is 0.67, indicating \"substantial\" agreement.", + "bbox": [ + 496, + 535, + 893, + 655 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dfc882a6b0fa1891f1c05c879f224d3362053a2e8e95be4b73c58bac9be2b93e.jpg", + "image_caption": [ + "Prompt:", + "Figure 6. Illustration of our Likert scale annotation guideline. Annotators are asked to give a score of 1 to 5 based on how many elements in the text prompt are missed or misrepresented in the image. The missed elements are underlined." + ], + "image_footnote": [], + "bbox": [ + 555, + 667, + 710, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A person sitting on a horse in air over gate in grass with people and trees in background.", + "bbox": [ + 720, + 684, + 834, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Miss/Misrepresent", + "bbox": [ + 720, + 762, + 833, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 out of 10 elements!", + "bbox": [ + 715, + 775, + 836, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines We compare our evaluation with two families of reference-free metrics on text-image match introduced in", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3LLaMA 2 question generation model checkpoint:", + "bbox": [ + 98, + 862, + 362, + 876 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "https://huggingface.co/tifa-benchmark/llama2_", + "bbox": [ + 78, + 876, + 434, + 887 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "tifa_question_generation", + "bbox": [ + 78, + 888, + 269, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20411", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Section 2. The first is the caption-based method. We use the state-of-the-art BLIP-2 FlanT5-XL [32] as the captioning model. The second approach is CLIPScore [17, 42]. We use CLIP (VIT-B/32) [42] to compute the score.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2e400e825c157cbf30e6d94e927257c62dea065af745323490c1d98750d47b21.jpg", + "table_caption": [ + "Table 2. Correlations between each evaluation metric and human judgment on text-to-image faithfulness, measured by Spearman's $\\rho$ and Kendall's $\\tau$ ." + ], + "table_footnote": [], + "table_body": "
Spearman's ρKendall's τ
Caption-Based
BLEU-418.318.8
ROUGE-L32.924.5
METEOR34.027.4
SPICE32.823.2
CLIPScore33.223.1
Ours
TIFA (VILT)49.338.2
TIFA (OFA)49.637.2
TIFA (GIT)54.542.6
TIFA (BLIP-2)55.943.6
TIFA (mPLUG)59.747.2
", + "bbox": [ + 116, + 202, + 426, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TIFA has a much higher correlation with human judgments than prior metrics. The correlations between each evaluation metric and human judgment are shown in Table 2. For caption-based evaluations, we use metrics BLEU-4 [38], ROUGE-L [33], METEOR [2], and SPICE [1]. TIFA has higher correlations with human judgments than all previous evaluation metrics on all VQA models. TIFA (mPLUG) yields the highest correlation with human judgments among all VQA models.", + "bbox": [ + 75, + 428, + 470, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Benchmarking Text-to-Image Models", + "text_level": 1, + "bbox": [ + 76, + 575, + 400, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 7 shows the average TIFA score text-to-image models get on TIFA v1.0. The detailed scores with each VQA model on each element type are provided in Appendix B. We can see a clear trend of how text-to-image models evolve over time. There is a jump in TIFA score after DALL-E [44] is released, from about $60\\%$ to $75\\%$ . Qualitative examples of our evaluation metric are in Appendix A.", + "bbox": [ + 75, + 599, + 470, + 705 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6444ce6e713d0de1c4dd28348aee4160542840b0ce2899eeeea58b43bcdfcaf.jpg", + "image_caption": [ + "Figure 7. Average FIFA score of text-to-image models on the FIFA v1.0 benchmark. The horizontal axis shows their release dates." + ], + "image_footnote": [], + "bbox": [ + 89, + 719, + 455, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Findings on Current Text-to-Image Models", + "text_level": 1, + "bbox": [ + 500, + 90, + 867, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 8 shows accuracy on each type of question in TIFA v1.0 for Stable Diffusion v1.1, v1.5, and v2.1. The score on each type reflects the text-to-image models' faithfulness in each type of visual element. To the best of our knowledge, TIFA is the only automatic evaluation method that can provide such a detailed fine-grained analysis of image generation. We separate the scores on COCO captions and other text inputs. For COCO captions, we also include the accuracy on the ground-truth images for reference. We summarize our findings in the following paragraphs.", + "bbox": [ + 496, + 113, + 893, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generating images from captions vs. free-form text From Figure 8, we can see that VQA accuracy is higher on the COCO captions than on other text inputs. The reason is that COCO captions correspond to real images, while other text inputs may correspond to compositions that cannot be found in real-world photos (e.g. \"a blue apple\").", + "bbox": [ + 496, + 273, + 892, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "What elements are text-to-image models struggling with?", + "text_level": 1, + "bbox": [ + 496, + 375, + 892, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the scores of each category in Figure 8, we can see that Stable Diffusion models are performing well on material, animal/human, color, and location in terms of text-to-image faithfulness. However, they yield low accuracy on questions involving shapes, counting, and spatial relations. \"Other\" mainly contains abstract art notions, and models are also struggling with them. There is also a big gap between the synthesized images and real images on the COCO captions. Future work can explore various directions (e.g., training data/loss and model architecture) to improve text-to-image models' faithfulness in these aspects.", + "bbox": [ + 496, + 388, + 893, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Why are ground-truth images not getting perfect scores?", + "text_level": 1, + "bbox": [ + 496, + 565, + 892, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ground-truth images in COCO do not get perfect scores because 1) the COCO captions contain a substantial amount of noise from crowd workers [24] and 2) VQA models are not perfect. Real images have higher accuracy in all categories except material, color and location, where differences are small. It is left to future work to determine whether this is simply due to noise or it is an area where assessment can be improved.", + "bbox": [ + 496, + 579, + 893, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stable Diffusion is evolving. We can see the consistent trend that Stable Diffusion models are improving in their later versions in most of the element categories. The exceptions are \"shape\" for both prompt sources, \"other\" and \"food\" for the free-form text inputs without gold images.", + "bbox": [ + 496, + 709, + 892, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Composing multiple objects is challenging. Figure 9 shows how the number of entities (objects, animals/humans, food) in the text input affects the average TIFA score. When there are more than 5 entities, The TIFA score starts to drop rapidly for all text-to-image models, consistent with similar findings in other vision-language evaluations [15, 13]. For reference, we also add the real images in COCO in this figure.", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20412", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/efcb7be6c7f5995a2fb4aff98010b343b41cdc22eb123bc70161d7238ea3d443.jpg", + "image_caption": [ + "VQA accuracy (by category) on COCO captions" + ], + "image_footnote": [ + "Stable Diffusion v1.1 Stable Diffusion v1.5 Stable Diffusion v2.1 COCO Ground-Truth" + ], + "bbox": [ + 101, + 130, + 869, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b54bec0300be6452b7e8f6d98172c093c1b8c0479f62ef377ac91e13b57a970.jpg", + "image_caption": [ + "VQA accuracy (by category) on free-form text", + "Figure 8. Accuracy on each type of question in the FIFA v1.0 benchmark. The text-to-image models are Stable Diffusion v1.1, v1.5, and v2.1. We order the categories by the average score Stable Diffusion v2.1 gets on corresponding questions. For COCO captions, we also include the accuracy of the ground-truth images for reference." + ], + "image_footnote": [], + "bbox": [ + 101, + 220, + 867, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd62f80e860bca3bec250310b2051b16b28f57dd17450724aa49c51322947c3d.jpg", + "image_caption": [ + "Figure 9. TIFA vs. numbers of entities (objects, animals/humans, and food) in the text input. The accuracy starts to drop when more than 5 entities are added to the text, showing that compositionality is hard for text-to-image models. Meanwhile, TIFA scores for COCO ground-truth (GT) images remain consistent." + ], + "image_footnote": [], + "bbox": [ + 89, + 356, + 449, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The TIFA score on real images is rather consistent and does not change as the number of entities increases. This quantitatively shows that composing multiple objects is challenging for current text-to-image models. One possible reason is that the CLIP text embedding, which is used to train Stable Diffusion, lacks compositionality, as investigated in [37].", + "bbox": [ + 75, + 626, + 468, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Analysis of VQA Models", + "text_level": 1, + "bbox": [ + 76, + 729, + 302, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One major concern of TIFA is that VQA models can introduce some errors. Table 2 shows that TIFA has a much higher correlation with human judgment than the previous metrics, regardless of the choice of the VQA models; here we conduct a more detailed analysis.", + "bbox": [ + 75, + 753, + 468, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sensitivity of TIFA to VQA models Figure 10 shows several recent text-to-image models' TIFA scores on the COCO captions in TIFA v1.0, measured by different VQA models. We also include the TIFA scores on the ground-truth COCO", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a0cef53a35d4ba2de499f15708435e4066ee30f3b94183cc9d6f39d93e7fab58.jpg", + "image_caption": [ + "Figure 10. Several text-to-image models' TIFA score on COCO captions, measured by different VQA models. We also include the accuracy of ground-truth images for reference." + ], + "image_footnote": [], + "bbox": [ + 509, + 354, + 879, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "images for reference. TIFA scores computed by different VQA models show a similar trend on these text-to-image models. Also, the ground-truth images get the highest TIFA score. We also computed Spearman's $\\rho$ of TIFA scores given by different VQA models. The pairwise correlation between all VQA models is greater than 0.6.", + "bbox": [ + 496, + 545, + 892, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Humans performing VQA To conduct further analysis on the VQA models, we ask annotators to answer the multiple-choice visual questions in TIFA v1.0. These annotations help us evaluate the accuracy of each VQA model. For multiple-choice questions, we add the option \"None of the above\" for human evaluation. Annotation guidelines are in Appendix C.", + "bbox": [ + 496, + 646, + 893, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We collect annotations of 1029 questions on 126 images. Each question is answered by two annotators. The inter-annotator agreement measured by Krippendorf's $\\alpha$ is 0.88. A third annotator is involved if two annotators disagree, and the final answer is chosen by the majority vote.", + "bbox": [ + 496, + 737, + 893, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Which VQA model should we use?", + "text_level": 1, + "bbox": [ + 498, + 823, + 746, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 reports the accuracy of each VQA model and the correlation between TIFA scores calculated by VQA model answers and human answers. We observe that higher model performance is directly related to the TIFA score's", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20413", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1ac4f68e319a0190fb52028e18db4d41728531a7408c22c16543410fd57702e9.jpg", + "table_caption": [ + "Table 3. Comparison of VQA models. The first row is the VQA accuracy, using the human VQA answers as reference. The second row is Spearman's correlation between TIFA scores calculated by each VQA model and the human VQA." + ], + "table_footnote": [], + "table_body": "
VILTOFAGITBLIP-2mPLUG
VQA Acc.76.177.179.181.084.5
TIFA Corr.60.963.772.575.676.8
", + "bbox": [ + 88, + 142, + 454, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "correlation with human judgments. mPLUG has the highest accuracy.", + "bbox": [ + 75, + 226, + 468, + 257 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Another important factor to consider is the runtime. We measure the inference speed of each VQA model on NVIDIA A40 GPU with batch size 1 over the Stable Diffusion v2.1 images (768 × 768 pixels). For one question, VILT takes 0.08s on average; OFA, GIT, and mPLUG all take about 0.25s; BLIP-2 takes 0.73s. Based on the above results, we choose mPLUG as the default VQA model for TIFA v1.0 because it is the most accurate while being reasonably fast.", + "bbox": [ + 75, + 257, + 468, + 378 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Separation of Text-to-Image Errors and VQA Errors Suppose an image gets a wrong answer given a visual question. Then the image generation or the VQA model might have made an error. Based on the human VQA results, we separate these two kinds of errors in Figure A. If human VQA gives the wrong answer, then we suspect the generated image has an error. Otherwise, the image is correct but the VQA model is making an error. Figure A shows that the majority of errors are made by the text-to-image models. For mPLUG, less than $25\\%$ errors are due to the VQA model. This suggests that the TIFA framework is a viable evaluation method despite its inherent challenges.", + "bbox": [ + 75, + 388, + 470, + 570 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5c4e0034b937bd063a8987e75ea85a0d28ce9c67dc0139cca89462041cd65461.jpg", + "image_caption": [ + "Figure 11. Source of the error when VQA gets the wrong answer." + ], + "image_footnote": [], + "bbox": [ + 89, + 583, + 455, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 76, + 708, + 191, + 724 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Why does TIFA work better than CLIPScore? Our experimental results and human evaluations in Section 5 show that TIFA is more accurate than prior metrics (CLIP [42] and captioning-based approaches) for evaluating text-to-image faithfulness. We hypothesize that the major challenge of these prior metrics is that they summarize the image outputs and text inputs into a single representation (embedding/caption). In contrast, TIFA exploits the power of the language models to decompose the text input into fine-grained probes, which allows VQA to capture more nuanced aspects of the text input and the generated image.", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "How do current VQA models perform on TIFA? One limitation is that TIFA requires VQA models to work reasonably well, which is true given the current models and TIFA v1.0, as shown in Section 5. Nevertheless, the assumption might not hold for current models in domains like anime and abstract art. TIFA is a modularized evaluation framework. The VQA models used within the framework can be updated as stronger VQA models become available in the future. For instance, we plan to incorporate GPT-4 once its image API is made public since it is likely to improve TIFA. Another possible solution is to ensemble multiple image understanding models. For example, one may employ expert models on art concepts. We leave this for future work.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Other limitations. Another limitation of TIFA is its runtime. Answering multiple visual questions is slower than one CLIP inference. In the scenario described in Section 5.4, mPLUG takes 1.6s to evaluate one image (without batching). Also, our question generation pipeline needs one inference on a modern language model for each text input. The run time is not a critical issue for benchmarking purposes, but may not be computationally feasible for the kind of large-scale data filtering done, for example, in LAION-5B [49]. Nevertheless, we would like to point out that our evaluation is much faster than the image generation process of diffusion models. Thus, we believe it is feasible to perform reranking and reinforcement learning with TIFA on diffusion models.", + "bbox": [ + 496, + 290, + 893, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 502, + 625, + 517 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present TIFA, a new automatic text-to-image faithfulness evaluation metric using VQA. Compared with prior metrics, TIFA is fine-grained, interpretable, and better aligned with human judgments. Based on this metric, we introduce the TIFA v1.0, a large-scale text-to-image benchmark containing 4K prompts and 25K questions. We conduct a comprehensive study of current text-to-image models using TIFA v1.0 and highlight the limitations of current generative models. We quantitatively show that current image generation models still struggle in counting, spatial relations, and composing multiple objects. Finally, we conduct extensive analysis and human evaluation, demonstrating that TIFA is robust to different VQA models. We hope TIFA will help evaluate future work on image generation and become increasingly sophisticated as it is upgraded with new LM, QA, and VQA components.", + "bbox": [ + 496, + 527, + 895, + 770 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "8. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 784, + 687, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Zirong Ye, Cheng-Yu Hsieh, Oscar Michel, Enhao Zhang, Jiafei Duan, Weijia Shi, Jieyu Zhang, the TIAL group, the ARK group, and the GRAIL lab at UW for their helpful feedback on this work. We would also thank all anonymous reviewers for providing insightful and constructive comments on our work.", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20414", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In European conference on computer vision, pages 382-398. Springer, 2016. 3, 7", + "[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7", + "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2, 3, 13", + "[4] Soravit Changpinyo, Doron Kukliansky, Idan Szpektor, Xi Chen, Nan Ding, and Radu Soricut. All you may need for vqa are image captions. In North American Chapter of the Association for Computational Linguistics, 2022. 3", + "[5] Jaemin Cho, Jiasen Lu, Dustin Schwenk, Hannaneh Hajishirzi, and Aniruddha Kembhavi. X-lxmert: Paint, caption and answer questions with multi-modal transformers. ArXiv, abs/2009.11278, 2020. 6, 14, 15", + "[6] Jaemin Cho, Abhaysinh Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers. ArXiv, abs/2202.04053, 2022. 2, 3, 5", + "[7] Hyung Won Chung, Le Hou, S. Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Wei Yu, Vincent Zhao, Yanping Huang, Andrew M. Dai, Hongkun Yu, Slav Petrov, Ed Huai hsin Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc Le, and Jason Wei. Scaling instruction-finetuned language models. ArXiv, abs/2210.11416, 2022. 4", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 4", + "[9] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868-12878, 2020. 14", + "[10] Kawin Ethayarajh and Dan Jurafsky. Utility is in the eye of the user: A critique of NLP leaderboards. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 3", + "[11] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, P. Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. *ArXiv*, abs/2212.05032, 2022. 1, 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Stanislav Frolov, Tobias Hinz, Federico Raue, Jorn Hees, and Andreas R. Dengel. Adversarial text-to-image synthesis: A review. Neural networks : the official journal of the International Neural Network Society, 144:187-209, 2021. 3", + "[13] Mona Gandhi, Mustafa Omer Gul, Eva Prakash, Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Measuring compositional consistency for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5046-5055, 2022. 7", + "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4", + "[15] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11287-11297, 2021. 7", + "[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10686-10696, 2021. 6, 14, 15", + "[17] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Joseph Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Conference on Empirical Methods in Natural Language Processing, 2021. 2, 3, 7", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017. 3", + "[19] Tobias Hinz, Stefan Heinrich, and Stefan Wermter. Semantic object accuracy for generative text-to-image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:1552-1565, 2019. 3", + "[20] Seunghoon Hong, Dingdong Yang, Jongwook Choi, and Honglak Lee. Inferring semantic layout for hierarchical text-to-image synthesis. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7986-7994, 2018. 3", + "[21] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. 2", + "[22] Yushi Hu, Chia-Hsuan Lee, Tianbao Xie, Tao Yu, Noah A Smith, and Mari Ostendorf. In-context learning for few-shot dialogue state tracking. arXiv preprint arXiv:2203.08568, 2022. 3", + "[23] Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras, Lavinia Dunagan, Jacob Morrison, Alexander Fabbri, Yejin Choi, and Noah A. Smith. Bidimensional leaderboards: Generate and evaluate language hand in hand. In North American Chapter of the Association for Computational Linguistics, 2022. 2", + "[24] Jungo Kasai, Keisuke Sakaguchi, Lavinia Dunagan, Jacob Daniel Morrison, Ronan Le Bras, Yejin Choi, and Noah A. Smith. Transparent human evaluation for image captioning." + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20415", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In North American Chapter of the Association for Computational Linguistics, 2022. 2, 7", + "[25] Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single qa system. In Findings, 2020. 2, 4, 13", + "[26] Daniel Khashabi, Gabriel Stanovsky, Jonathan Bragg, Nicholas Lourie, Jungo Kasai, Yejin Choi, Noah A. Smith, and Daniel Weld. GENIE: Toward reproducible and standardized human evaluation for text generation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 2022. 2", + "[27] Saehoon Kim, Sanghun Cho, Chiheon Kim, Doyup Lee, and Woonhyuk Baek. mindall-e on conceptual captions, 2021. 6, 14, 15", + "[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 2, 4", + "[29] Ranjay Krishna, Michael S. Bernstein, and Li Fei-Fei. Information maximizing visual question generation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2008–2018, 2019. 2, 4", + "[30] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 1, 2", + "[31] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng da Cao, Ji Zhang, Songfang Huang, Feiran Huang, Jingren Zhou, and Luo Si. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. In Conference on Empirical Methods in Natural Language Processing, 2022. 2, 4, 13", + "[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597, 2023. 2, 4, 7", + "[33] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain, July 2004. Association for Computational Linguistics. 7", + "[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2, 5, 13, 15", + "[35] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum. Compositional visual generation with composable diffusion models. ArXiv, abs/2206.01714, 2022. 1", + "[36] Rosanne Liu, Daniel H Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, R. J. Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. ArXiv, abs/2212.10562, 2022. 1" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. Crepe: Can vision-language foundation models reason compositionally? ArXiv, abs/2212.07796, 2022. 2, 8", + "[38] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 7", + "[39] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 3", + "[40] Vitali Petsiuk, Alexander E Siemann, Saisamrit Surbehera, Zad Chin, Keith Tyser, Gregory Hunter, Arvind Raghavan, Yann Hicke, Bryan A Plummer, Ori Kerret, et al. Human evaluation of text-to-image models on a multi-task benchmark. arXiv preprint arXiv:2211.12112, 2022. 1", + "[41] Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, and Mike Lewis. Measuring and narrowing the compositionality gap in language models, 2022. 3", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 7, 9, 14", + "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1", + "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 1, 7, 15", + "[45] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. ArXiv, abs/1908.10084, 2019. 4", + "[46] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 1, 6, 14, 16", + "[47] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1, 2, 5, 13, 15", + "[48] Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. ArXiv, abs/1606.03498, 2016. 3", + "[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next gener" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20416", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ation image-text models. arXiv preprint arXiv:2210.08402, 2022.9", + "[50] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 3", + "[51] Hongjin Su, Jungo Kasai, Chen Henry Wu, Weijia Shi, Tianlu Wang, Jiayi Xin, Rui Zhang, Mari Ostendorf, Luke Zettlemoyer, Noah A. Smith, and Tao Yu. Selective annotation makes language models better few-shot learners. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[52] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2826, 2015. 3", + "[53] Hao Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. ArXiv, abs/1908.07490, 2019. 15", + "[54] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 6", + "[55] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2015. 3", + "[56] Alex Wang, Kyunghyun Cho, and Mike Lewis. Asking and answering questions to evaluate the factual consistency of summaries. In Annual Meeting of the Association for Computational Linguistics, 2020. 3", + "[57] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2, 4", + "[58] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR, abs/2202.03052, 2022. 2, 4", + "[59] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 3", + "[60] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316-1324, 2017. 6, 14, 15", + "[61] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "ref_text", + "text": "Yang, Burcu Karagol Ayan, Benton C. Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. ArXiv, abs/2206.10789, 2022. 1, 2, 5", + "bbox": [ + 532, + 90, + 893, + 159 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20417", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/ae7f1ffe-fe67-4917-84ae-5c196628c6e5_model.json b/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/ae7f1ffe-fe67-4917-84ae-5c196628c6e5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e68daf3e8ce3b312055ea39dffec1864994ce080 --- /dev/null +++ b/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/ae7f1ffe-fe67-4917-84ae-5c196628c6e5_model.json @@ -0,0 +1,3920 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.175, + 0.131, + 0.797, + 0.177 + ], + "angle": 0, + "content": "TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.203, + 0.731, + 0.277 + ], + "angle": 0, + "content": "Yushi Hu1 Benlin Liu1 Jungo Kasai1 Yizhong Wang1 Mari Ostendorf1 Ranjay Krishna1,2 Noah A. Smith1,2 1University of Washington 2Allen Institute for AI https://tifa-benchmark.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.474, + 0.735 + ], + "angle": 0, + "content": "Despite thousands of researchers, engineers, and artists actively working on improving text-to-image generation models, systems often fail to produce images that accurately align with the text inputs. We introduce TIFA (Text-to-Image Faithfulness evaluation with question Answering), an automatic evaluation metric that measures the faithfulness of a generated image to its text input via visual question answering (VQA). Specifically, given a text input, we automatically generate several question-answer pairs using a language model. We calculate image faithfulness by checking whether existing VQA models can answer these questions using the generated image. TIFA is a reference-free metric that allows for fine-grained and interpretable evaluations of generated images. TIFA also has better correlations with human judgments than existing metrics. Based on this approach, we introduce TIFA v1.0, a benchmark consisting of 4K diverse text inputs and 25K questions across 12 categories (object, counting, etc.). We present a comprehensive evaluation of existing text-to-image models using TIFA v1.0 and highlight the limitations and challenges of current models. For instance, we find that current text-to-image models, despite doing well on color and material, still struggle in counting, spatial relations, and composing multiple objects. We hope our benchmark will help carefully measure the research progress in text-to-image synthesis and provide valuable insights for further research.1" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.747, + 0.21, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.773, + 0.471, + 0.864 + ], + "angle": 0, + "content": "While we welcome artistic freedom when we commission art from artists, images produced by deep generative models [44, 46, 43, 47, 61] should conform closely to our requests. Despite the advances in generative models, it is still challenging for models to produce images faithful to users' intentions [40, 11, 30, 35, 36]. For example, current" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.47, + 0.903 + ], + "angle": 0, + "content": ". All data and a pip-installable evaluation package are available on the project page." + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.312, + 0.678, + 0.323 + ], + "angle": 0, + "content": "Stable Diffusion v1.5" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.323, + 0.689, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.312, + 0.831, + 0.322 + ], + "angle": 0, + "content": "Stable Diffusion v2.1" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.322, + 0.841, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.541, + 0.424, + 0.851, + 0.449 + ], + "angle": 0, + "content": "Text Input: A person sitting on a horse in air over gate in grass with people and trees in background." + }, + { + "type": "title", + "bbox": [ + 0.544, + 0.455, + 0.849, + 0.467 + ], + "angle": 0, + "content": "GPT-3 generated + verified QAs (pre-generated in TIFA v1.0 benchmark)" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.47, + 0.677, + 0.482 + ], + "angle": 0, + "content": "Question: what is the animal?" + }, + { + "type": "text", + "bbox": [ + 0.688, + 0.47, + 0.852, + 0.482 + ], + "angle": 0, + "content": "Answer inferred from text: horse" + }, + { + "type": "text", + "bbox": [ + 0.554, + 0.49, + 0.587, + 0.502 + ], + "angle": 0, + "content": "VQA:" + }, + { + "type": "text", + "bbox": [ + 0.607, + 0.491, + 0.639, + 0.502 + ], + "angle": 0, + "content": "Horse" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.488, + 0.67, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.762, + 0.49, + 0.793, + 0.501 + ], + "angle": 0, + "content": "Horse" + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.489, + 0.823, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.511, + 0.658, + 0.522 + ], + "angle": 0, + "content": "Question: is there a gate?" + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.527, + 0.587, + 0.539 + ], + "angle": 0, + "content": "VQA:" + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.528, + 0.631, + 0.539 + ], + "angle": 0, + "content": "No" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.528, + 0.662, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.69, + 0.511, + 0.842, + 0.522 + ], + "angle": 0, + "content": "Answer inferred from text: yes" + }, + { + "type": "text", + "bbox": [ + 0.767, + 0.528, + 0.788, + 0.539 + ], + "angle": 0, + "content": "Yes" + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.527, + 0.823, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.549, + 0.673, + 0.56 + ], + "angle": 0, + "content": "Question: is the horse in air?" + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.563, + 0.587, + 0.575 + ], + "angle": 0, + "content": "VQA:" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.563, + 0.631, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.562, + 0.663, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.562, + 0.689, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.562, + 0.718, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.584, + 0.546, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.55, + 0.596, + 0.635, + 0.609 + ], + "angle": 0, + "content": "TIFA 71.4" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.609, + 0.631, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.639, + 0.613, + 0.659, + 0.622 + ], + "angle": 0, + "content": "ined" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.613, + 0.688, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.69, + 0.613, + 0.73, + 0.623 + ], + "angle": 0, + "content": "Accurate" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.613, + 0.757, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.76, + 0.613, + 0.82, + 0.623 + ], + "angle": 0, + "content": "Interpretable" + }, + { + "type": "title", + "bbox": [ + 0.587, + 0.632, + 0.807, + 0.642 + ], + "angle": 0, + "content": "v.s. other automatic image-text align metrics" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.663, + 0.589, + 0.673 + ], + "angle": 0, + "content": "CLIPScore" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.647, + 0.657, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.704, + 0.662, + 0.741, + 0.672 + ], + "angle": 0, + "content": "SPICE" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.653, + 0.78, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.783, + 0.656, + 0.801, + 0.666 + ], + "angle": 0, + "content": "#" + }, + { + "type": "text", + "bbox": [ + 0.756, + 0.675, + 0.845, + 0.683 + ], + "angle": 0, + "content": "sim. (caption, text input)" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.693, + 0.656, + 0.703 + ], + "angle": 0, + "content": "CLIP:24.3 SPICE:22.2" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.693, + 0.697, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.725, + 0.693, + 0.849, + 0.703 + ], + "angle": 0, + "content": "CLIP: 21.3 SPICE: 11.8" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.895, + 0.836 + ], + "angle": 0, + "content": "Figure 1. Illustration of how TIFA works, and comparison with the widely-used CLIPScore and SPICE metrics. Given the text input, TIFA uses GPT-3 to generate several question-answer pairs, and a QA model filters them (3 out of 14 questions for this text input are shown). TIFA measures whether VQA models can accurately answer these questions given the generated image. In this example, TIFA indicates that the image generated by Stable Diffusion v2.1 is better than that by v1.5, while CLIP and SPICE yield the opposite result. The text input is from the MSCOCO validation set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "models often fail to compose multiple objects [40, 11, 35], bind attributes to the wrong objects [11], and struggle in generating visual text [36]. Today, there are efforts to address" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20406" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.561, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.09, + 0.87, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.251, + 0.895, + 0.349 + ], + "angle": 0, + "content": "Figure 2. (a) Overview of how TIFA evaluates the faithfulness of a synthesized image. TIFA uses a language model (LM), a question-answering (QA) model, and a visual-question-answering (VQA) model. Given a text input, we generate several question-answer pairs with the LM and then filter them via the QA model. To evaluate the faithfulness of a synthesized image to the text input, a VQA model answers these visual questions using the image, and we check the answers for correctness. (b) TIFA v1.0 benchmark. While TIFA is applicable to any text prompt, to allow direct comparison across different studies, and for ease of use, we introduce the TIFA v1.0 benchmark, a repository of text inputs along with pre-generated question-answer tuples with answer choices. To evaluate a text-to-image model, a user first produces the images for the text inputs in TIFA v1.0 and then performs VQA with our provided tools on generated images to compute TIFA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.471, + 0.465 + ], + "angle": 0, + "content": "these challenges: researchers are imposing linguistic structure with diffusion guidance to produce images with multiple objects [11]; others are designing reward models trained using human feedback to better align generations with user intention [30]. However, progress is difficult to quantify without accurate and interpretable evaluation measures that explain when and how models struggle." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.471, + 0.864 + ], + "angle": 0, + "content": "A critical bottleneck, therefore, is the lack of reliable automatic evaluation metrics for text-to-image generation faithfulness. One of the popular metrics is CLIPScore [17], which measures the cosine similarity between the CLIP embeddings [42] of the text input and the generated image. However, since CLIP is not effective at counting objects [42], or reasoning compositionally [37], CLIPScore is unreliable and often inaccurate. Another family of evaluation metrics uses image captions, in which an image captioning model first converts the image into text, and then the image caption is evaluated by comparing it against the text input. Unfortunately, using captioning models is insufficient since they might decide to ignore salient information in images or focus on other non-essential image regions [24]; for example, a captioning model might say that the images in Figure 1 are \"a field of grass with trees in the background\". Moreover, evaluating text (caption) generation is inherently challenging [23, 26]. Another recent text-to-image evaluation is DALL-Eval [6], which employs object detection to determine if the objects in the texts are in the generated images. However, this approach only works on synthesized text and measures faithfulness along the limited axes of objects, counting, colors, and spatial relationships but misses activities, geolocation, weather, time, materials, shapes, sizes, and other potential categories we often ask about when we recall images from memory [29]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To address the above challenges, we introduce TIFA, a new metric to evaluate text-to-image generation faithfulness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.895, + 0.676 + ], + "angle": 0, + "content": "Our approach is illustrated in Figure 2. Given a repository of text inputs, we automatically generate question-answer pairs for each text via a language model (here, GPT-3 [3]). A question-answering (QA) system (here, UnifiedQA [25]) is subsequently used to verify and filter these question-answer pairs. To evaluate a generated image, we use a visual-question-answering (VQA) system (here, mPLUGlarge [31], BLIP-2 [32], etc.) to answer the questions given the generated image. We measure the image's faithfulness to the text input as the accuracy of the answers generated by the VQA system. While the accuracy of TIFA is dependent on the accuracy of the VQA model, our experiments show that TIFA has much higher correlation with human judgments than CLIPScore (Spearman's \\(\\rho = 0.60\\) vs. 0.33) and captioning-based approaches (Spearman's \\(\\rho = 0.60\\) vs. 0.34). Additionally, since the LMs and VQA models will continue to improve, we hypothesize that TIFA will continue to be more reliable over time. Also, our metrics can automatically detect when elements are missing in the generation: in Figure 2, TIFA detects that the generated image does not contain a TV." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.895, + 0.877 + ], + "angle": 0, + "content": "To promote the use of our new evaluation metric, we release TIFA v1.0, a large-scale text-to-image generation benchmark containing 4K diverse text inputs, sampled from the MSCOCO captions [34], DrawBench [47], PartiPrompts [61], and PaintSkill [6]. Each input comes with a pre-generated set of question-answer pairs, resulting in 25K questions covering 4.5K distinct elements. These questions have been automatically generated and pre-filtered using a question-answering model. This benchmark also comes with different VQA models [57, 28, 58, 32, 31, 21] that can be used to evaluate generative models and can be easily extended to use future VQA models when they become available." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We conduct a comprehensive evaluation of current text-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20407" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "to-image models using TIFA v1.0. Thanks to TIFA's ability to detect fine-grained unfaithfulness in images, we find that current state-of-the-art models are good at rendering common objects, animals, and colors, but still struggle in composing multiple objects, reasoning about spatial relations, and binding the correct activity for each entity. In addition, our ablation experiments show that TIFA is robust to different VQA models. Future researchers can use TIFA v1.0 to compare their text-to-image models' faithfulness across different studies. Also, future generative models may focus on addressing the weaknesses of current models that TIFA discovered. In addition, with TIFA, users can customize evaluations with their own text inputs and questions [10]; for example, a future TIFA benchmark could focus on counting or scene text." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.332, + 0.22, + 0.348 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.471, + 0.387 + ], + "angle": 0, + "content": "We compare TIFA to other image and language generation evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.473, + 0.825 + ], + "angle": 0, + "content": "Prior image generation evaluation Prior work usually compares image generation models via pairwise comparison by humans. How to design automatic evaluation metrics to approximate human assessment of the quality of machine-generated images has always been a major challenge in computer vision. There are two aspects to evaluate, namely image quality and image-text alignment. Inception Score [48] and FID [18] are the most widely adopted metrics for image quality. They compare the features of the generated images and gold images extracted from a pre-trained Inception-V3 model [52] to evaluate the fidelity and diversity of generated images. However, they rely on ground-truth images and are based on a classification model, which makes them not suitable for complex datasets [12]. For image-text alignment, prior metrics are mainly based on CLIP [42], captioning, and object detection models. CLIPScore [17] and CLIP-R [39] are based on the cosine similarity of image and text CLIP [42] embeddings. [6, 19, 20] first convert the images using a captioning model, and then compare the image caption with the text using metrics like CIDEr [55] and SPICE [1]. SOA [19] and DALL-Eval [6] employ object detection models to determine if objects, attributes, and relations in the text input are in the generated image. However, this approach only works on synthesized text inputs and measures faithfulness on limited axes (object, counting, color, spatial relation), missing elements like material, shape, activities, and context. In contrast, thanks to the flexibility of questions, TIFA works on any text inputs and evaluates faithfulness across a broad spectrum of dimensions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Summarization evaluation in NLP TIFA is inspired by the summarization evaluation methods based on question answering (QA) [56, 50]. Given a summary, a language model generates a set of questions about the text. A QA model checks if the same answer can be inferred from the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "text and the summary. These QA-based metrics have much higher correlations with human judgments on the factual consistency of summarization than other automatic metrics [56, 50]. TIFA can be seen as treating the text input for the text-to-image model as a summary of the generated image." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.181, + 0.67, + 0.197 + ], + "angle": 0, + "content": "3. The TIFA Metric" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.206, + 0.895, + 0.373 + ], + "angle": 0, + "content": "We introduce a framework for automatically estimating the faithfulness of an image to its text prompt. Given a text input \\( T \\), we aim to measure the faithfulness of the generated image \\( I \\). An overview of our metric is illustrated in Figure 2. From \\( T \\), we generate \\( N \\) multiple-choice question-answer tuples \\( \\{Q_i, C_i, A_i\\}_{i=1}^N \\), in which \\( Q_i \\) is a question, \\( C_i \\) is a set of answer choices, and \\( A_i \\in C_i \\) is the gold answer. The answer \\( A_i \\) can be inferred given \\( T, Q_i \\), and \\( C_i \\). Next, for each question \\( Q_i \\), we use a VQA model to produce an answer \\( A_i^{\\mathrm{VQA}} = \\max_{a \\in C_i} p(a \\mid I, Q_i) \\). We define the faithfulness between the text \\( T \\) and image \\( I \\) as the VQA accuracy:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.383, + 0.892, + 0.424 + ], + "angle": 0, + "content": "\\[\nf a i t h f u l n e s s (T, I) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbb {1} \\left[ A _ {i} ^ {\\mathrm {V Q A}} = A _ {i} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.895, + 0.563 + ], + "angle": 0, + "content": "The range of our faithfulness score is \\([0, 1]\\). It is maximized when we have a performant VQA model, and the image \\(I\\) accurately covers the information in the text \\(T\\) so that for any question \\(Q\\), which can be answered given \\(T\\) can also be answered given \\(I\\). Several key design decisions will be addressed in later sections: how to generate questions (\\(\\S 3.1\\)), how to control the question quality (\\(\\S 3.2\\)), and how to answer those questions (\\(\\S 3.3\\)). Finally, we give a step-by-step qualitative example of TIFA in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.572, + 0.763, + 0.587 + ], + "angle": 0, + "content": "3.1. Question-Answer Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.684 + ], + "angle": 0, + "content": "Our main challenge is to generate diverse questions that cover all elements of the text input evenly. We also simplify the question-generation pipeline into a single GPT-3 [3] completion, so that TIFA can exploit the power of recent language models (LM) and work with updated black-box LMs (e.g., ChatGPT) in the future." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.895, + 0.837 + ], + "angle": 0, + "content": "Inspired by prior work [4], given a text prompt \\( T \\), we generate the question-answer tuples \\( \\{Q_i, C_i, A_i\\}_{i=1}^N \\) via the pipeline illustrated in Figure 3. Different from prior work, which relies on multiple components, our pipeline is completed by a single inference run via in-context learning with GPT-3 [3, 59, 22, 41, 51], thereby avoiding the need for intermediate human annotations. We annotate 15 examples and use them as in-context examples for GPT-3 to follow. Here we take the text \"A photo of three dogs\" as an example. Each in-context example contains the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Element extraction Given text prompt \\( T \\), GPT-3 will first extract all elements \\( \\{v_i\\}_{i=1}^m \\) following prior work [4] (for the in-context examples, we perform element extraction manually). The elements include noun phrases (including named" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20408" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.446, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.432, + 0.472, + 0.504 + ], + "angle": 0, + "content": "Figure 3. Our question-answer pair generation pipeline. The whole pipeline can be executed via a single inference of GPT-3 via in-context learning. Given the text prompt, GPT-3 first extracts the elements and then generates two questions for each element. The GPT-3 output is then parsed and filtered by UnifiedQA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.472, + 0.56 + ], + "angle": 0, + "content": "entities), verbs, adjectives, adverbs, and parse tree spans with no more than 3 words altogether. For the above example, the elements are photo, three, dogs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.472, + 0.705 + ], + "angle": 0, + "content": "Element category classification For each element \\( v_{i} \\), following [29], we classify the elements into one of the following 12 categories: object, activity, animal, food, counting, color, material, spatial, location, shape, attribute, and other. As shown in Figure 3, text generated from GPT-3 contains the category corresponding to each question. For example, \"three\" is \"counting\", and \"dogs\" is classified as \"animal.\" This step allows a detailed analysis of the text-to-image model's ability in each category." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.472, + 0.849 + ], + "angle": 0, + "content": "Question generation conditioned on elements For each element \\( v_{i} \\), we generate two questions. The first is a question that should be answered \"yes\" for a faithful generated image, and the second question has \\( v_{i} \\) as its answer. For example, two questions are generated for the element \"three\". The first is \"are there three dogs?\", and the choices are {yes, no}. The second is \"how many dogs are there?\", and the choices are {1,2,3,4}. These two types of questions make our evaluations diverse and robust to surface-level differences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Completing above steps by prompting GPT-3 once As mentioned earlier, for each text \\(T\\) , the whole pipeline can be completed by one GPT-3 inference. We annotated 15" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "in-context examples that cover all types of questions. The full prompt is in the Appendix. The prompt format is shown in Figure 3. Our in-context examples follow the same format, and identical examples are used for all text inputs, leading to a fixed and limited amount of human annotation cost. We use code-davinci-002 engine for question generation, and the decoding temperature is 0." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.207, + 0.679, + 0.224 + ], + "angle": 0, + "content": "3.2. Question Filtering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.231, + 0.895, + 0.445 + ], + "angle": 0, + "content": "To ensure the quality of generated images, we use UnifiedQA [25] to verify the GPT-3 generated question-answer pairs and filter out the ones that GPT-3 and UnifiedQA do not agree on. UnifiedQA\\(^2\\) is a state-of-the-art multi-task question-answering model that can answer both multiple-choice and free-form questions. Denote the UnifiedQA model as \\( QA \\). Given the text \\( T \\), question \\( Q_{i} \\), choices \\( C_{i} \\), and answer \\( A_{i} \\), Let \\( A_{i}^{f} = QA(T,Q_{i}) \\) be the free-form answer, and \\( A_{i}^{mc} = QA(T,Q_{i},C_{i}) \\) be the multiple-choice answer. We keep the question if \\( A_{i} = A_{i}^{mc} \\) and the word-level \\( F_{1} \\) score between \\( A_{i}^{f} \\) and \\( A_{i} \\) is greater than 0.7. We conduct a human evaluation on 1000 filtered question-answer pairs. Only 7 are considered not reasonable (e.g., generated choices do not include a correct answer). Details are in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.894, + 0.541 + ], + "angle": 0, + "content": "Recommended VQA model Based on considerations over the accuracy, correlation with human judgments, and run time, we currently suggest using mPLUG-large as the VQA model for TIFA. Analysis is given in Section 5.4. Like the LM and QA components, the VQA component can be updated in the future as the technology improves." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.55, + 0.638, + 0.567 + ], + "angle": 0, + "content": "3.3. VQA Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.573, + 0.894, + 0.68 + ], + "angle": 0, + "content": "Since our questions contain a diverse set of visual elements (e.g., activity, art style), we use open-domain pre-trained vision-language models as our VQA model (rather than closed-class classification models fine-tuned on VQAv2 [14]). We provide tools to easily perform VQA on arbitrary images and questions, based on 5 state-of-the-art VQA models trained with distinct data and strategies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.895, + 0.867 + ], + "angle": 0, + "content": "Vision-language models The general pre-trained vision-language model are GIT-large [57], VILT-B/32 [28], OFA-large [58], and mPLUG-large [31]. These models are pretrained on a large amount of image-text pairs, and downstream image-to-text tasks like image captioning and visual question answering. Notice that these models have not been trained to answer multiple-choice questions. For each question, we first decode the free-form answer and then choose the choice that has the highest similarity with the decoded answer, measured by SBERT [45]. Another model we use is BLIP-2 FlanT5-XL [32], in which a VIT [8] is connected with a frozen FlanT5 [7] via a lightweight transformer. This" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.894, + 0.901 + ], + "angle": 0, + "content": "2Model checkpoint we use: https://huggingface.co/allenai/unifiedqa-v2-t5-large-1363200." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20409" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.188, + 0.095, + 0.478, + 0.107 + ], + "angle": 0, + "content": "Text Input: A wet dog standing on a beach next to the ocean." + }, + { + "type": "title", + "bbox": [ + 0.185, + 0.115, + 0.446, + 0.129 + ], + "angle": 0, + "content": "Step 1: Question Generation + Filtering" + }, + { + "type": "table", + "bbox": [ + 0.187, + 0.132, + 0.787, + 0.272 + ], + "angle": 0, + "content": "
About dog (animal/human)About ocean (location)About wet (attribute)
Q: Is this a dog?Q: Is there an ocean?Q: Is the dog wet?
A: YesA: YesA: Yes
Q: What animal is in the picture?Q: What type of place is this?Q: Is the dog wet or dry?
A: DogA: OceanA: Wet
About beach (location)About standing (activity)About next to (spatial relation)
Q: Is this a beach?Q: Is the dog standing?Q: Is the dog next to the ocean?
A: YesA: YesA: Yes
Q: What type of place is this?Q: What is the dog doing?Q: Is the dog next to or in the ocean?
A: BeachA: StandingA: Next to
" + }, + { + "type": "title", + "bbox": [ + 0.185, + 0.28, + 0.414, + 0.294 + ], + "angle": 0, + "content": "Step 2: Visual Question Answering" + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.304, + 0.339, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.429, + 0.335, + 0.44 + ], + "angle": 0, + "content": "Stable Diffusion v1.5 image" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.45, + 0.226, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.462, + 0.322, + 0.472 + ], + "angle": 0, + "content": "TIFA \\(7 / 10 = 0.7\\)" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.299, + 0.503, + 0.31 + ], + "angle": 0, + "content": "About dog (animal/human)" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.311, + 0.453, + 0.323 + ], + "angle": 0, + "content": "Q: Is this a dog?" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.323, + 0.408, + 0.334 + ], + "angle": 0, + "content": "A:Yes" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.336, + 0.532, + 0.347 + ], + "angle": 0, + "content": "Q: What animal is in the picture?" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.348, + 0.411, + 0.359 + ], + "angle": 0, + "content": "A: Dog" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.318, + 0.519, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.373, + 0.483, + 0.384 + ], + "angle": 0, + "content": "About beach (location)" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.385, + 0.463, + 0.395 + ], + "angle": 0, + "content": "Q: Is this a beach?" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.397, + 0.408, + 0.407 + ], + "angle": 0, + "content": "A:Yes" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.409, + 0.515, + 0.421 + ], + "angle": 0, + "content": "Q: What type of place is this?" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.422, + 0.42, + 0.432 + ], + "angle": 0, + "content": "A:Beach" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.394, + 0.519, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.447, + 0.483, + 0.457 + ], + "angle": 0, + "content": "About ocean (location)" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.458, + 0.475, + 0.469 + ], + "angle": 0, + "content": "Q: Is there an ocean?" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.47, + 0.408, + 0.481 + ], + "angle": 0, + "content": "A:Yes" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.47, + 0.519, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.299, + 0.71, + 0.31 + ], + "angle": 0, + "content": "About standing (activity)" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.311, + 0.702, + 0.322 + ], + "angle": 0, + "content": "Q: Is the dog standing?" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.323, + 0.621, + 0.333 + ], + "angle": 0, + "content": "A: No" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.347, + 0.691, + 0.358 + ], + "angle": 0, + "content": "About wet (attribute)" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.36, + 0.679, + 0.371 + ], + "angle": 0, + "content": "Q: Is the dog wet?" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.372, + 0.621, + 0.382 + ], + "angle": 0, + "content": "A: No" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.384, + 0.71, + 0.395 + ], + "angle": 0, + "content": "Q: Is the dog wet or dry?" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.397, + 0.626, + 0.408 + ], + "angle": 0, + "content": "A: Dry" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.421, + 0.735, + 0.432 + ], + "angle": 0, + "content": "About next to (spatial relation)" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.433, + 0.741, + 0.444 + ], + "angle": 0, + "content": "Q: Is the dog next to the ocean?" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.445, + 0.624, + 0.456 + ], + "angle": 0, + "content": "A:Yes" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.457, + 0.766, + 0.468 + ], + "angle": 0, + "content": "Q: Is the dog next to or in the ocean?" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.47, + 0.642, + 0.48 + ], + "angle": 0, + "content": "A: Next to" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.323, + 0.774, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.369, + 0.774, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.394, + 0.774, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.443, + 0.774, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.457, + 0.766, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.47, + 0.774, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "list", + "bbox": [ + 0.755, + 0.443, + 0.774, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.496, + 0.891, + 0.525 + ], + "angle": 0, + "content": "Figure 4. Step-by-step qualitative example of TIFA metric. Given a text input, we first generate question-answer pairs and filter them. We strike through the questions filtered out by UnifiedQA. Then we run VQA models on the generated image to get the TIFA score." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.566 + ], + "angle": 0, + "content": "model allows for performing multiple-choice VQA directly due to the flexibility of the LM." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.581, + 0.469, + 0.614 + ], + "angle": 0, + "content": "4. FIFA v1.0: Benchmark for Text-to-Image Generation Faithfulness" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.471, + 0.79 + ], + "angle": 0, + "content": "In this section, we introduce TIFA v1.0, a text-to-image generation faithfulness benchmark based on the evaluation method discussed in Section 3. The benchmark consists of 4,081 diverse text inputs paired with 25,829 question-answer pairs. Each question is classified into one of the categories discussed in Section 3.1. The benchmark also comes with Python pip-installable APIs to perform VQA with various state-of-the-art VQA models on arbitrary visual questions. The overall TIFA for each text-to-image model is computed by averaging TIFA scores of images generated from each text input in the benchmark." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.237, + 0.816 + ], + "angle": 0, + "content": "4.1. Text Collections" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We collect 4,081 text inputs to benchmark text-to-image models' generation ability on diverse tasks. 2,000 text inputs are image captions from COCO validation set [34]. These captions have corresponding gold images. Since text-to-image models are often used to create abstract art, we also" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.641 + ], + "angle": 0, + "content": "collect 2,081 text inputs from previous works that do not correspond to any real image. All text inputs we use contain \\(\\geq 3\\) words. We include 161 from DrawBench used in Imagen [47] (texts that are categorized as \"misspellings\" and \"rare words\" are removed); 1420 from PartiPrompt used in Parti [61] (texts in category \"abstract\" are removed); and 500 texts from PaintSkill used in DALL-Eval [6]." + }, + { + "type": "table_caption", + "bbox": [ + 0.595, + 0.652, + 0.793, + 0.665 + ], + "angle": 0, + "content": "Table 1. Statistics of FIFA v1.0." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.665, + 0.892, + 0.83 + ], + "angle": 0, + "content": "
Statistics
# of prompts4,081
- # of COCO captions2,000
- # of DrawBench, PartiPrompt, PaintSkill prompts2,081
# of questions25,829
- # of binary questions17,226
- # of multiple-choice questions8,603
avg. # of questions per prompt6.3
avg. # of words per prompt10.5
avg. # of elements per prompt4.3
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.713, + 0.863 + ], + "angle": 0, + "content": "4.2. Statistics and Diversity" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Table 1 shows the basic statistics of the TIFA v1.0 benchmark. We demonstrate TIFA v1.0's diversity in Figure 5." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20410" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.092, + 0.097, + 0.184, + 0.115 + ], + "angle": 0, + "content": "a yellow t-shirt with a dog on it\"" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.129, + 0.186, + 0.153 + ], + "angle": 0, + "content": "\"a panda bear with aviator glasses on its head\"" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.163, + 0.187, + 0.172 + ], + "angle": 0, + "content": "\"A red colored car\"" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.188, + 0.18, + 0.212 + ], + "angle": 0, + "content": "\"One cat and one dog sitting on the grass\"" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.097, + 0.234, + 0.117 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.099, + 0.285, + 0.11 + ], + "angle": 0, + "content": "Object" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.11, + 0.328, + 0.117 + ], + "angle": 0, + "content": "table,train,bus ..." + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.129, + 0.231, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.129, + 0.292, + 0.14 + ], + "angle": 0, + "content": "Spatial" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.14, + 0.334, + 0.148 + ], + "angle": 0, + "content": "above,below,right .." + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.16, + 0.234, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.159, + 0.301, + 0.17 + ], + "angle": 0, + "content": "Counting" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.171, + 0.338, + 0.179 + ], + "angle": 0, + "content": "two,three,several" + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.19, + 0.231, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.19, + 0.271, + 0.2 + ], + "angle": 0, + "content": "Food" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.201, + 0.314, + 0.209 + ], + "angle": 0, + "content": "pizza, bananas ..." + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.099, + 0.371, + 0.117 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.099, + 0.472, + 0.109 + ], + "angle": 0, + "content": "Animal/Human" + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.11, + 0.458, + 0.118 + ], + "angle": 0, + "content": "man,woman,dog ..." + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.099, + 0.511, + 0.118 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.099, + 0.588, + 0.109 + ], + "angle": 0, + "content": "Attribute" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.11, + 0.618, + 0.118 + ], + "angle": 0, + "content": "large, small, young ..." + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.126, + 0.461, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.465, + 0.138, + 0.551, + 0.151 + ], + "angle": 0, + "content": "TIFA v1.0" + }, + { + "type": "text", + "bbox": [ + 0.395, + 0.158, + 0.584, + 0.179 + ], + "angle": 0, + "content": "4,081 text inputs 25,829 questions \n4,550 distinct elements 12 categories" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.19, + 0.371, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.189, + 0.443, + 0.2 + ], + "angle": 0, + "content": "Material" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.201, + 0.469, + 0.21 + ], + "angle": 0, + "content": "wooden,metal,glass" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.189, + 0.515, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.19, + 0.562, + 0.2 + ], + "angle": 0, + "content": "Shape" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.201, + 0.602, + 0.21 + ], + "angle": 0, + "content": "circular,square" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.099, + 0.652, + 0.12 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.66, + 0.099, + 0.724, + 0.109 + ], + "angle": 0, + "content": "Activity" + }, + { + "type": "text", + "bbox": [ + 0.66, + 0.11, + 0.758, + 0.118 + ], + "angle": 0, + "content": "sitting.standing ..." + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.13, + 0.652, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.66, + 0.13, + 0.724, + 0.14 + ], + "angle": 0, + "content": "Location" + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.141, + 0.761, + 0.15 + ], + "angle": 0, + "content": "field,street,beach ..." + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.16, + 0.655, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.662, + 0.16, + 0.7, + 0.17 + ], + "angle": 0, + "content": "Color" + }, + { + "type": "text", + "bbox": [ + 0.661, + 0.172, + 0.745, + 0.18 + ], + "angle": 0, + "content": "red,white,blue ..." + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.191, + 0.654, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.662, + 0.19, + 0.701, + 0.2 + ], + "angle": 0, + "content": "Other" + }, + { + "type": "text", + "bbox": [ + 0.66, + 0.201, + 0.75, + 0.21 + ], + "angle": 0, + "content": "punk, starry night" + }, + { + "type": "text", + "bbox": [ + 0.771, + 0.097, + 0.875, + 0.106 + ], + "angle": 0, + "content": "\"a painting of a fox" + }, + { + "type": "text", + "bbox": [ + 0.782, + 0.106, + 0.852, + 0.114 + ], + "angle": 0, + "content": "in the style" + }, + { + "type": "text", + "bbox": [ + 0.782, + 0.114, + 0.852, + 0.121 + ], + "angle": 0, + "content": "starry night" + }, + { + "type": "text", + "bbox": [ + 0.783, + 0.131, + 0.873, + 0.139 + ], + "angle": 0, + "content": "\"A man running to" + }, + { + "type": "text", + "bbox": [ + 0.783, + 0.14, + 0.872, + 0.148 + ], + "angle": 0, + "content": "hit a tennis ball" + }, + { + "type": "text", + "bbox": [ + 0.789, + 0.163, + 0.866, + 0.171 + ], + "angle": 0, + "content": "\"a flag with a" + }, + { + "type": "text", + "bbox": [ + 0.789, + 0.172, + 0.866, + 0.179 + ], + "angle": 0, + "content": "dinosaur on it\"" + }, + { + "type": "text", + "bbox": [ + 0.778, + 0.189, + 0.882, + 0.197 + ], + "angle": 0, + "content": "\"a photo of suitcase" + }, + { + "type": "text", + "bbox": [ + 0.782, + 0.198, + 0.879, + 0.206 + ], + "angle": 0, + "content": "and chair; chair is" + }, + { + "type": "text", + "bbox": [ + 0.782, + 0.206, + 0.879, + 0.213 + ], + "angle": 0, + "content": "right to suitcase\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.223, + 0.892, + 0.251 + ], + "angle": 0, + "content": "Figure 5. Statistics and diversity of TIFA v1.0. The text inputs contain elements from 12 categories (e.g., object, spatial, and counting). We show the most common elements from each category. In addition, we also show some example text inputs on the sides." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.429 + ], + "angle": 0, + "content": "TIFA v1.0 contains questions about 4,550 distinct elements, which are categorized into 12 categories. The number of times each type of element occurs in the text input is object (7,854), animal/human (3,501), attribute (3,399), activity (2,851), spatial (2,265), location (1,840), color (1,743), counting (986), food (911), material (209), shape (69), and other (201). The \"other\" category includes notions often used in abstract art, such as \"starry night\" and \"steampunk.\" The accuracy of VQA on a particular genre measures the text-to-image model's ability in the corresponding aspect. Please refer to Appendix D for more details on TIFA v1.0." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.443, + 0.469, + 0.473 + ], + "angle": 0, + "content": "4.3. Finetuned Open-Source Language Model for Question Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.472, + 0.62 + ], + "angle": 0, + "content": "For TIFA v1.0, we use GPT-3 to generate the questions. While benchmarking with TIFA v1.0 is a deterministic process, using TIFA to create a new benchmark might not be deterministic as the underlying question generator (GPT-3 in our case) might change privately. To promote deterministic benchmark generation, we fine-tune and release a LLaMA 2 (7B) [54] model that parses the captions and generates questions for arbitrary texts, using TIFA v1.0 questions as training examples." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.639, + 0.21, + 0.656 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.473, + 0.848 + ], + "angle": 0, + "content": "In this section, we first show that TIFA has substantially higher correlations with human judgments than prior metrics on text-to-image faithfulness (§5.1). Then we present a comprehensive evaluation of existing text-to-image models using TIFA v1.0, highlighting the challenges of current text-to-image models (§5.3). Finally, we conduct an analysis of TIFA's robustness against different VQA models (§5.4). For all experiments, we use mPLUG as the VQA model for TIFA unless stated otherwise. The models we evaluate include AttnGAN [60], X-LXMERT [5], VQ-Diffusion [16], minDALL-E [27], and Stable Diffusion v1.1, v1.5, and v2.1 [46]. Details are in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.261, + 0.823, + 0.278 + ], + "angle": 0, + "content": "5.1. Correlation with Human Judgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.285, + 0.892, + 0.36 + ], + "angle": 0, + "content": "To compare TIFA with prior evaluation metrics, we first conduct human evaluations of the text-to-image models on the 1-5 Likert scale on text-to-image faithfulness. Then we compare TIFA with other metrics based on their correlation with human judgments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.895, + 0.535 + ], + "angle": 0, + "content": "Likert scale on text-to-image faithfulness Annotators are asked to answer on a scale of 1 (worst) to 5 (best) to the question \"Does the image match the text?\" The detailed annotation guidelines are in Appendix C. Annotators are asked to focus on text-to-image faithfulness rather than image quality. The Likert scale should be based on how many elements in the text prompt are missed or misrepresented in the image. Objects are more important than attributes, relations, and activities. If an object is missed in the image, then all related attributes, activities, relations, etc. are also considered lost. An example is given in Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.895, + 0.656 + ], + "angle": 0, + "content": "We collect annotations of 800 generated images on 160 text inputs from TIFA v1.0. For each prompt, we sample an image from the 5 most recent generative models we evaluated, i.e., minDALL-E, VQ-Diffusion, Stable Diffusion v1.1, v1.5, and v2.1. We collect 2 annotations per image and average over the scores as the single \"faithfulness\" score. The inter-annotator agreement measured by Krippendorf's \\(\\alpha\\) is 0.67, indicating \"substantial\" agreement." + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.668, + 0.712, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.671, + 0.77, + 0.683 + ], + "angle": 0, + "content": "Prompt:" + }, + { + "type": "text", + "bbox": [ + 0.721, + 0.685, + 0.835, + 0.755 + ], + "angle": 0, + "content": "A person sitting on a horse in air over gate in grass with people and trees in background." + }, + { + "type": "text", + "bbox": [ + 0.721, + 0.763, + 0.834, + 0.774 + ], + "angle": 0, + "content": "Miss/Misrepresent" + }, + { + "type": "text", + "bbox": [ + 0.717, + 0.776, + 0.837, + 0.787 + ], + "angle": 0, + "content": "4 out of 10 elements!" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.793, + 0.894, + 0.85 + ], + "angle": 0, + "content": "Figure 6. Illustration of our Likert scale annotation guideline. Annotators are asked to give a score of 1 to 5 based on how many elements in the text prompt are missed or misrepresented in the image. The missed elements are underlined." + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.863, + 0.364, + 0.877 + ], + "angle": 0, + "content": "3LLaMA 2 question generation model checkpoint:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.079, + 0.877, + 0.435, + 0.888 + ], + "angle": 0, + "content": "https://huggingface.co/tifa-benchmark/llama2_" + }, + { + "type": "page_footnote", + "bbox": [ + 0.079, + 0.889, + 0.271, + 0.9 + ], + "angle": 0, + "content": "tifa_question_generation" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.863, + 0.435, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Baselines We compare our evaluation with two families of reference-free metrics on text-image match introduced in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20411" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.471, + 0.152 + ], + "angle": 0, + "content": "Section 2. The first is the caption-based method. We use the state-of-the-art BLIP-2 FlanT5-XL [32] as the captioning model. The second approach is CLIPScore [17, 42]. We use CLIP (VIT-B/32) [42] to compute the score." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.164, + 0.47, + 0.204 + ], + "angle": 0, + "content": "Table 2. Correlations between each evaluation metric and human judgment on text-to-image faithfulness, measured by Spearman's \\(\\rho\\) and Kendall's \\(\\tau\\)." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.203, + 0.428, + 0.413 + ], + "angle": 0, + "content": "
Spearman's ρKendall's τ
Caption-Based
BLEU-418.318.8
ROUGE-L32.924.5
METEOR34.027.4
SPICE32.823.2
CLIPScore33.223.1
Ours
TIFA (VILT)49.338.2
TIFA (OFA)49.637.2
TIFA (GIT)54.542.6
TIFA (BLIP-2)55.943.6
TIFA (mPLUG)59.747.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.472, + 0.563 + ], + "angle": 0, + "content": "TIFA has a much higher correlation with human judgments than prior metrics. The correlations between each evaluation metric and human judgment are shown in Table 2. For caption-based evaluations, we use metrics BLEU-4 [38], ROUGE-L [33], METEOR [2], and SPICE [1]. TIFA has higher correlations with human judgments than all previous evaluation metrics on all VQA models. TIFA (mPLUG) yields the highest correlation with human judgments among all VQA models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.576, + 0.401, + 0.592 + ], + "angle": 0, + "content": "5.2. Benchmarking Text-to-Image Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.471, + 0.706 + ], + "angle": 0, + "content": "Figure 7 shows the average TIFA score text-to-image models get on TIFA v1.0. The detailed scores with each VQA model on each element type are provided in Appendix B. We can see a clear trend of how text-to-image models evolve over time. There is a jump in TIFA score after DALL-E [44] is released, from about \\(60\\%\\) to \\(75\\%\\). Qualitative examples of our evaluation metric are in Appendix A." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.72, + 0.457, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.899 + ], + "angle": 0, + "content": "Figure 7. Average FIFA score of text-to-image models on the FIFA v1.0 benchmark. The horizontal axis shows their release dates." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.868, + 0.108 + ], + "angle": 0, + "content": "5.3. Findings on Current Text-to-Image Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.894, + 0.267 + ], + "angle": 0, + "content": "Figure 8 shows accuracy on each type of question in TIFA v1.0 for Stable Diffusion v1.1, v1.5, and v2.1. The score on each type reflects the text-to-image models' faithfulness in each type of visual element. To the best of our knowledge, TIFA is the only automatic evaluation method that can provide such a detailed fine-grained analysis of image generation. We separate the scores on COCO captions and other text inputs. For COCO captions, we also include the accuracy on the ground-truth images for reference. We summarize our findings in the following paragraphs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.893, + 0.366 + ], + "angle": 0, + "content": "Generating images from captions vs. free-form text From Figure 8, we can see that VQA accuracy is higher on the COCO captions than on other text inputs. The reason is that COCO captions correspond to real images, while other text inputs may correspond to compositions that cannot be found in real-world photos (e.g. \"a blue apple\")." + }, + { + "type": "title", + "bbox": [ + 0.498, + 0.375, + 0.893, + 0.39 + ], + "angle": 0, + "content": "What elements are text-to-image models struggling with?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.894, + 0.556 + ], + "angle": 0, + "content": "Based on the scores of each category in Figure 8, we can see that Stable Diffusion models are performing well on material, animal/human, color, and location in terms of text-to-image faithfulness. However, they yield low accuracy on questions involving shapes, counting, and spatial relations. \"Other\" mainly contains abstract art notions, and models are also struggling with them. There is also a big gap between the synthesized images and real images on the COCO captions. Future work can explore various directions (e.g., training data/loss and model architecture) to improve text-to-image models' faithfulness in these aspects." + }, + { + "type": "title", + "bbox": [ + 0.498, + 0.566, + 0.893, + 0.58 + ], + "angle": 0, + "content": "Why are ground-truth images not getting perfect scores?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.894, + 0.701 + ], + "angle": 0, + "content": "Ground-truth images in COCO do not get perfect scores because 1) the COCO captions contain a substantial amount of noise from crowd workers [24] and 2) VQA models are not perfect. Real images have higher accuracy in all categories except material, color and location, where differences are small. It is left to future work to determine whether this is simply due to noise or it is an area where assessment can be improved." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.893, + 0.787 + ], + "angle": 0, + "content": "Stable Diffusion is evolving. We can see the consistent trend that Stable Diffusion models are improving in their later versions in most of the element categories. The exceptions are \"shape\" for both prompt sources, \"other\" and \"food\" for the free-form text inputs without gold images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Composing multiple objects is challenging. Figure 9 shows how the number of entities (objects, animals/humans, food) in the text input affects the average TIFA score. When there are more than 5 entities, The TIFA score starts to drop rapidly for all text-to-image models, consistent with similar findings in other vision-language evaluations [15, 13]. For reference, we also add the real images in COCO in this figure." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20412" + } + ], + [ + { + "type": "image_footnote", + "bbox": [ + 0.23, + 0.09, + 0.743, + 0.103 + ], + "angle": 0, + "content": "Stable Diffusion v1.1 Stable Diffusion v1.5 Stable Diffusion v2.1 COCO Ground-Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.116, + 0.624, + 0.13 + ], + "angle": 0, + "content": "VQA accuracy (by category) on COCO captions" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.131, + 0.87, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.207, + 0.619, + 0.221 + ], + "angle": 0, + "content": "VQA accuracy (by category) on free-form text" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.221, + 0.869, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.293, + 0.894, + 0.336 + ], + "angle": 0, + "content": "Figure 8. Accuracy on each type of question in the FIFA v1.0 benchmark. The text-to-image models are Stable Diffusion v1.1, v1.5, and v2.1. We order the categories by the average score Stable Diffusion v2.1 gets on corresponding questions. For COCO captions, we also include the accuracy of the ground-truth images for reference." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.357, + 0.45, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.53, + 0.471, + 0.601 + ], + "angle": 0, + "content": "Figure 9. TIFA vs. numbers of entities (objects, animals/humans, and food) in the text input. The accuracy starts to drop when more than 5 entities are added to the text, showing that compositionality is hard for text-to-image models. Meanwhile, TIFA scores for COCO ground-truth (GT) images remain consistent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.719 + ], + "angle": 0, + "content": "The TIFA score on real images is rather consistent and does not change as the number of entities increases. This quantitatively shows that composing multiple objects is challenging for current text-to-image models. One possible reason is that the CLIP text embedding, which is used to train Stable Diffusion, lacks compositionality, as investigated in [37]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.731, + 0.303, + 0.746 + ], + "angle": 0, + "content": "5.4. Analysis of VQA Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.755, + 0.47, + 0.831 + ], + "angle": 0, + "content": "One major concern of TIFA is that VQA models can introduce some errors. Table 2 shows that TIFA has a much higher correlation with human judgment than the previous metrics, regardless of the choice of the VQA models; here we conduct a more detailed analysis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Sensitivity of TIFA to VQA models Figure 10 shows several recent text-to-image models' TIFA scores on the COCO captions in TIFA v1.0, measured by different VQA models. We also include the TIFA scores on the ground-truth COCO" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.356, + 0.88, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.492, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Figure 10. Several text-to-image models' TIFA score on COCO captions, measured by different VQA models. We also include the accuracy of ground-truth images for reference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.893, + 0.637 + ], + "angle": 0, + "content": "images for reference. TIFA scores computed by different VQA models show a similar trend on these text-to-image models. Also, the ground-truth images get the highest TIFA score. We also computed Spearman's \\(\\rho\\) of TIFA scores given by different VQA models. The pairwise correlation between all VQA models is greater than 0.6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.895, + 0.737 + ], + "angle": 0, + "content": "Humans performing VQA To conduct further analysis on the VQA models, we ask annotators to answer the multiple-choice visual questions in TIFA v1.0. These annotations help us evaluate the accuracy of each VQA model. For multiple-choice questions, we add the option \"None of the above\" for human evaluation. Annotation guidelines are in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.738, + 0.894, + 0.815 + ], + "angle": 0, + "content": "We collect annotations of 1029 questions on 126 images. Each question is answered by two annotators. The inter-annotator agreement measured by Krippendorf's \\(\\alpha\\) is 0.88. A third annotator is involved if two annotators disagree, and the final answer is chosen by the majority vote." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.824, + 0.747, + 0.839 + ], + "angle": 0, + "content": "Which VQA model should we use?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Table 3 reports the accuracy of each VQA model and the correlation between TIFA scores calculated by VQA model answers and human answers. We observe that higher model performance is directly related to the TIFA score's" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20413" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.472, + 0.142 + ], + "angle": 0, + "content": "Table 3. Comparison of VQA models. The first row is the VQA accuracy, using the human VQA answers as reference. The second row is Spearman's correlation between TIFA scores calculated by each VQA model and the human VQA." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.143, + 0.455, + 0.201 + ], + "angle": 0, + "content": "
VILTOFAGITBLIP-2mPLUG
VQA Acc.76.177.179.181.084.5
TIFA Corr.60.963.772.575.676.8
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.47, + 0.258 + ], + "angle": 0, + "content": "correlation with human judgments. mPLUG has the highest accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.47, + 0.38 + ], + "angle": 0, + "content": "Another important factor to consider is the runtime. We measure the inference speed of each VQA model on NVIDIA A40 GPU with batch size 1 over the Stable Diffusion v2.1 images (768 × 768 pixels). For one question, VILT takes 0.08s on average; OFA, GIT, and mPLUG all take about 0.25s; BLIP-2 takes 0.73s. Based on the above results, we choose mPLUG as the default VQA model for TIFA v1.0 because it is the most accurate while being reasonably fast." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.472, + 0.571 + ], + "angle": 0, + "content": "Separation of Text-to-Image Errors and VQA Errors Suppose an image gets a wrong answer given a visual question. Then the image generation or the VQA model might have made an error. Based on the human VQA results, we separate these two kinds of errors in Figure A. If human VQA gives the wrong answer, then we suspect the generated image has an error. Otherwise, the image is correct but the VQA model is making an error. Figure A shows that the majority of errors are made by the text-to-image models. For mPLUG, less than \\(25\\%\\) errors are due to the VQA model. This suggests that the TIFA framework is a viable evaluation method despite its inherent challenges." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.584, + 0.457, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.677, + 0.47, + 0.692 + ], + "angle": 0, + "content": "Figure 11. Source of the error when VQA gets the wrong answer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.709, + 0.192, + 0.725 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Why does TIFA work better than CLIPScore? Our experimental results and human evaluations in Section 5 show that TIFA is more accurate than prior metrics (CLIP [42] and captioning-based approaches) for evaluating text-to-image faithfulness. We hypothesize that the major challenge of these prior metrics is that they summarize the image outputs and text inputs into a single representation (embedding/caption). In contrast, TIFA exploits the power of the language models to decompose the text input into fine-grained probes, which allows VQA to capture more nuanced aspects of the text input and the generated image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.288 + ], + "angle": 0, + "content": "How do current VQA models perform on TIFA? One limitation is that TIFA requires VQA models to work reasonably well, which is true given the current models and TIFA v1.0, as shown in Section 5. Nevertheless, the assumption might not hold for current models in domains like anime and abstract art. TIFA is a modularized evaluation framework. The VQA models used within the framework can be updated as stronger VQA models become available in the future. For instance, we plan to incorporate GPT-4 once its image API is made public since it is likely to improve TIFA. Another possible solution is to ensemble multiple image understanding models. For example, one may employ expert models on art concepts. We leave this for future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.895, + 0.489 + ], + "angle": 0, + "content": "Other limitations. Another limitation of TIFA is its runtime. Answering multiple visual questions is slower than one CLIP inference. In the scenario described in Section 5.4, mPLUG takes 1.6s to evaluate one image (without batching). Also, our question generation pipeline needs one inference on a modern language model for each text input. The run time is not a critical issue for benchmarking purposes, but may not be computationally feasible for the kind of large-scale data filtering done, for example, in LAION-5B [49]. Nevertheless, we would like to point out that our evaluation is much faster than the image generation process of diffusion models. Thus, we believe it is feasible to perform reranking and reinforcement learning with TIFA on diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.503, + 0.627, + 0.518 + ], + "angle": 0, + "content": "7. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.897, + 0.771 + ], + "angle": 0, + "content": "We present TIFA, a new automatic text-to-image faithfulness evaluation metric using VQA. Compared with prior metrics, TIFA is fine-grained, interpretable, and better aligned with human judgments. Based on this metric, we introduce the TIFA v1.0, a large-scale text-to-image benchmark containing 4K prompts and 25K questions. We conduct a comprehensive study of current text-to-image models using TIFA v1.0 and highlight the limitations of current generative models. We quantitatively show that current image generation models still struggle in counting, spatial relations, and composing multiple objects. Finally, we conduct extensive analysis and human evaluation, demonstrating that TIFA is robust to different VQA models. We hope TIFA will help evaluate future work on image generation and become increasingly sophisticated as it is upgraded with new LM, QA, and VQA components." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.688, + 0.803 + ], + "angle": 0, + "content": "8. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We thank Zirong Ye, Cheng-Yu Hsieh, Oscar Michel, Enhao Zhang, Jiafei Duan, Weijia Shi, Jieyu Zhang, the TIAL group, the ARK group, and the GRAIL lab at UW for their helpful feedback on this work. We would also thank all anonymous reviewers for providing insightful and constructive comments on our work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20414" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.171 + ], + "angle": 0, + "content": "[1] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In European conference on computer vision, pages 382-398. Springer, 2016. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.472, + 0.312 + ], + "angle": 0, + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2, 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.472, + 0.368 + ], + "angle": 0, + "content": "[4] Soravit Changpinyo, Doron Kukliansky, Idan Szpektor, Xi Chen, Nan Ding, and Radu Soricut. All you may need for vqa are image captions. In North American Chapter of the Association for Computational Linguistics, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.472, + 0.424 + ], + "angle": 0, + "content": "[5] Jaemin Cho, Jiasen Lu, Dustin Schwenk, Hannaneh Hajishirzi, and Aniruddha Kembhavi. X-lxmert: Paint, caption and answer questions with multi-modal transformers. ArXiv, abs/2009.11278, 2020. 6, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.472, + 0.479 + ], + "angle": 0, + "content": "[6] Jaemin Cho, Abhaysinh Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers. ArXiv, abs/2202.04053, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.472, + 0.62 + ], + "angle": 0, + "content": "[7] Hyung Won Chung, Le Hou, S. Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Wei Yu, Vincent Zhao, Yanping Huang, Andrew M. Dai, Hongkun Yu, Slav Petrov, Ed Huai hsin Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc Le, and Jason Wei. Scaling instruction-finetuned language models. ArXiv, abs/2210.11416, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.472, + 0.717 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.718, + 0.472, + 0.774 + ], + "angle": 0, + "content": "[9] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868-12878, 2020. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[10] Kawin Ethayarajh and Dan Jurafsky. Utility is in the eye of the user: A critique of NLP leaderboards. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[11] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, P. Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. *ArXiv*, abs/2212.05032, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.148 + ], + "angle": 0, + "content": "[12] Stanislav Frolov, Tobias Hinz, Federico Raue, Jorn Hees, and Andreas R. Dengel. Adversarial text-to-image synthesis: A review. Neural networks : the official journal of the International Neural Network Society, 144:187-209, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[13] Mona Gandhi, Mustafa Omer Gul, Eva Prakash, Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Measuring compositional consistency for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5046-5055, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.894, + 0.303 + ], + "angle": 0, + "content": "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[15] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11287-11297, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10686-10696, 2021. 6, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.894, + 0.516 + ], + "angle": 0, + "content": "[17] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Joseph Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Conference on Empirical Methods in Natural Language Processing, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.517, + 0.894, + 0.572 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.894, + 0.629 + ], + "angle": 0, + "content": "[19] Tobias Hinz, Stefan Heinrich, and Stefan Wermter. Semantic object accuracy for generative text-to-image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:1552-1565, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.631, + 0.894, + 0.687 + ], + "angle": 0, + "content": "[20] Seunghoon Hong, Dingdong Yang, Jongwook Choi, and Honglak Lee. Inferring semantic layout for hierarchical text-to-image synthesis. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7986-7994, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.894, + 0.731 + ], + "angle": 0, + "content": "[21] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.894, + 0.786 + ], + "angle": 0, + "content": "[22] Yushi Hu, Chia-Hsuan Lee, Tianbao Xie, Tao Yu, Noah A Smith, and Mari Ostendorf. In-context learning for few-shot dialogue state tracking. arXiv preprint arXiv:2203.08568, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.894, + 0.858 + ], + "angle": 0, + "content": "[23] Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras, Lavinia Dunagan, Jacob Morrison, Alexander Fabbri, Yejin Choi, and Noah A. Smith. Bidimensional leaderboards: Generate and evaluate language hand in hand. In North American Chapter of the Association for Computational Linguistics, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[24] Jungo Kasai, Keisuke Sakaguchi, Lavinia Dunagan, Jacob Daniel Morrison, Ronan Le Bras, Yejin Choi, and Noah A. Smith. Transparent human evaluation for image captioning." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20415" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.472, + 0.12 + ], + "angle": 0, + "content": "In North American Chapter of the Association for Computational Linguistics, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.177 + ], + "angle": 0, + "content": "[25] Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single qa system. In Findings, 2020. 2, 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.472, + 0.262 + ], + "angle": 0, + "content": "[26] Daniel Khashabi, Gabriel Stanovsky, Jonathan Bragg, Nicholas Lourie, Jungo Kasai, Yejin Choi, Noah A. Smith, and Daniel Weld. GENIE: Toward reproducible and standardized human evaluation for text generation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.472, + 0.304 + ], + "angle": 0, + "content": "[27] Saehoon Kim, Sanghun Cho, Chiheon Kim, Doyup Lee, and Woonhyuk Baek. mindall-e on conceptual captions, 2021. 6, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.472, + 0.362 + ], + "angle": 0, + "content": "[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.472, + 0.419 + ], + "angle": 0, + "content": "[29] Ranjay Krishna, Michael S. Bernstein, and Li Fei-Fei. Information maximizing visual question generation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2008–2018, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.472, + 0.489 + ], + "angle": 0, + "content": "[30] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.472, + 0.587 + ], + "angle": 0, + "content": "[31] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng da Cao, Ji Zhang, Songfang Huang, Feiran Huang, Jingren Zhou, and Luo Si. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. In Conference on Empirical Methods in Natural Language Processing, 2022. 2, 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.472, + 0.645 + ], + "angle": 0, + "content": "[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597, 2023. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.472, + 0.702 + ], + "angle": 0, + "content": "[33] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain, July 2004. Association for Computational Linguistics. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2, 5, 13, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[35] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum. Compositional visual generation with composable diffusion models. ArXiv, abs/2206.01714, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[36] Rosanne Liu, Daniel H Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, R. J. Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. ArXiv, abs/2212.10562, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[37] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. Crepe: Can vision-language foundation models reason compositionally? ArXiv, abs/2212.07796, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[38] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[39] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.894, + 0.36 + ], + "angle": 0, + "content": "[40] Vitali Petsiuk, Alexander E Siemann, Saisamrit Surbehera, Zad Chin, Keith Tyser, Gregory Hunter, Arvind Raghavan, Yann Hicke, Bryan A Plummer, Ori Kerret, et al. Human evaluation of text-to-image models on a multi-task benchmark. arXiv preprint arXiv:2211.12112, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.894, + 0.404 + ], + "angle": 0, + "content": "[41] Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, and Mike Lewis. Measuring and narrowing the compositionality gap in language models, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.894, + 0.488 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 7, 9, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.894, + 0.531 + ], + "angle": 0, + "content": "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.894, + 0.587 + ], + "angle": 0, + "content": "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 1, 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.894, + 0.631 + ], + "angle": 0, + "content": "[45] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. ArXiv, abs/1908.10084, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.894, + 0.702 + ], + "angle": 0, + "content": "[46] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 1, 6, 14, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.894, + 0.8 + ], + "angle": 0, + "content": "[47] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1, 2, 5, 13, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.894, + 0.844 + ], + "angle": 0, + "content": "[48] Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. ArXiv, abs/1606.03498, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next gener" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20416" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.12 + ], + "angle": 0, + "content": "ation image-text models. arXiv preprint arXiv:2210.08402, 2022.9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[50] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.472, + 0.274 + ], + "angle": 0, + "content": "[51] Hongjin Su, Jungo Kasai, Chen Henry Wu, Weijia Shi, Tianlu Wang, Jiayi Xin, Rui Zhang, Mari Ostendorf, Luke Zettlemoyer, Noah A. Smith, and Tao Yu. Selective annotation makes language models better few-shot learners. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.472, + 0.346 + ], + "angle": 0, + "content": "[52] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2826, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[53] Hao Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. ArXiv, abs/1908.07490, 2019. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.472, + 0.459 + ], + "angle": 0, + "content": "[54] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.472, + 0.517 + ], + "angle": 0, + "content": "[55] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.472, + 0.574 + ], + "angle": 0, + "content": "[56] Alex Wang, Kyunghyun Cho, and Mike Lewis. Asking and answering questions to evaluate the factual consistency of summaries. In Annual Meeting of the Association for Computational Linguistics, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.472, + 0.631 + ], + "angle": 0, + "content": "[57] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.472, + 0.701 + ], + "angle": 0, + "content": "[58] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR, abs/2202.03052, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.472, + 0.787 + ], + "angle": 0, + "content": "[59] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.472, + 0.87 + ], + "angle": 0, + "content": "[60] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316-1324, 2017. 6, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[61] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.534, + 0.092, + 0.894, + 0.16 + ], + "angle": 0, + "content": "Yang, Burcu Karagol Ayan, Benton C. Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. ArXiv, abs/2206.10789, 2022. 1, 2, 5" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20417" + } + ] +] \ No newline at end of file diff --git a/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/ae7f1ffe-fe67-4917-84ae-5c196628c6e5_origin.pdf b/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/ae7f1ffe-fe67-4917-84ae-5c196628c6e5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..172174e4cebca46f384a688d099505ea4b0fa60f --- /dev/null +++ b/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/ae7f1ffe-fe67-4917-84ae-5c196628c6e5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c30a8af439c351e53883ccd5d235be6033f71421f464bd30fdea3f082bbf9667 +size 5812529 diff --git a/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/full.md b/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4f888ee45ff75fcc73573a8698662e53bc2bc52f --- /dev/null +++ b/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/full.md @@ -0,0 +1,582 @@ +# TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering + +Yushi Hu1 Benlin Liu1 Jungo Kasai1 Yizhong Wang1 Mari Ostendorf1 Ranjay Krishna1,2 Noah A. Smith1,2 1University of Washington 2Allen Institute for AI https://tifa-benchmark.github.io/ + +# Abstract + +Despite thousands of researchers, engineers, and artists actively working on improving text-to-image generation models, systems often fail to produce images that accurately align with the text inputs. We introduce TIFA (Text-to-Image Faithfulness evaluation with question Answering), an automatic evaluation metric that measures the faithfulness of a generated image to its text input via visual question answering (VQA). Specifically, given a text input, we automatically generate several question-answer pairs using a language model. We calculate image faithfulness by checking whether existing VQA models can answer these questions using the generated image. TIFA is a reference-free metric that allows for fine-grained and interpretable evaluations of generated images. TIFA also has better correlations with human judgments than existing metrics. Based on this approach, we introduce TIFA v1.0, a benchmark consisting of 4K diverse text inputs and 25K questions across 12 categories (object, counting, etc.). We present a comprehensive evaluation of existing text-to-image models using TIFA v1.0 and highlight the limitations and challenges of current models. For instance, we find that current text-to-image models, despite doing well on color and material, still struggle in counting, spatial relations, and composing multiple objects. We hope our benchmark will help carefully measure the research progress in text-to-image synthesis and provide valuable insights for further research.1 + +# 1. Introduction + +While we welcome artistic freedom when we commission art from artists, images produced by deep generative models [44, 46, 43, 47, 61] should conform closely to our requests. Despite the advances in generative models, it is still challenging for models to produce images faithful to users' intentions [40, 11, 30, 35, 36]. For example, current + +Stable Diffusion v1.5 +![](images/0534cb1a249c7e271671e89f343dc2e4cfa429059bd379d85675ff68f55c136b.jpg) +Text Input: A person sitting on a horse in air over gate in grass with people and trees in background. + +![](images/984f040b7734452145bb0a652872438939cbafdb1886a69c1c3b86df988b4220.jpg) +Stable Diffusion v2.1 + +# GPT-3 generated + verified QAs (pre-generated in TIFA v1.0 benchmark) + +Question: what is the animal? + +Answer inferred from text: horse + +VQA: + +Horse + +![](images/2bf43cc660fd0cbecdd7a6a05464fa8a5a8de1811ed94b32e0d27b25c31d869c.jpg) + +Horse + +![](images/a103acb97b718e129bf14bd098cc04876cefcfa5e0ea8305c5e4819e6d873079.jpg) + +Question: is there a gate? + +VQA: + +No + +![](images/d87f16d37f52a8358d4b071322c5ce2b06a396dd438a98129ab3fecb1f8942dc.jpg) + +Answer inferred from text: yes + +Yes + +![](images/1d0ee3238ee5c9725828c47e0a85ad4a3bd73aa5ab323447b1f80b863c4c3cc7.jpg) + +Question: is the horse in air? + +VQA: + +![](images/5522b1600f668c49eac7cac3072645156e09fd921715ae997a7bbf79243f1d91.jpg) + +![](images/12d21720e3a232654f85b354ee1e1e2e64a08e693896616df2d1b7e4d45a986e.jpg) + +![](images/bd0192131f3afc4d22357421e1c9860357c0a4db85eb95034afaede286fe48f0.jpg) + +![](images/6a109c7eb64024dfb4799df213666bd0459271ee1743504b90ae74869342987f.jpg) + +![](images/5a7d0228e334eb274d7795337a069b7531a0c9550f8e7779439ece13b3af090e.jpg) + +TIFA 71.4 + +![](images/121579668e855fc1d5a60de5ef1fabadaf8939a1c28a315c435cc855e89249b9.jpg) + +ined + +![](images/3e2f53cc61ec55a112cf9006c4fdac613ac1e83011e8281fbca5e9a47ada3ddd.jpg) + +Accurate + +![](images/68c35f09e9d481791bf1b3bf74ab11598c76571421eaf99f060cfb1be6bc50aa.jpg) + +Interpretable + +# v.s. other automatic image-text align metrics + +CLIPScore + +![](images/64ecd3c515b2c89600f1c498d79e8e943ff752319db31f78b2b77da92d1e0cfb.jpg) + +SPICE + +![](images/c6d78a922c8fab6c04bbc48f56edfc903042ac82931bf4bb8d972f5dde7096d5.jpg) + +# + +sim. (caption, text input) + +CLIP:24.3 SPICE:22.2 + +![](images/a4976d5848b68e554a64a723fd69f112fc97a4dfe505b52d15689e7e7604e2dc.jpg) + +CLIP: 21.3 SPICE: 11.8 + +Figure 1. Illustration of how TIFA works, and comparison with the widely-used CLIPScore and SPICE metrics. Given the text input, TIFA uses GPT-3 to generate several question-answer pairs, and a QA model filters them (3 out of 14 questions for this text input are shown). TIFA measures whether VQA models can accurately answer these questions given the generated image. In this example, TIFA indicates that the image generated by Stable Diffusion v2.1 is better than that by v1.5, while CLIP and SPICE yield the opposite result. The text input is from the MSCOCO validation set. + +models often fail to compose multiple objects [40, 11, 35], bind attributes to the wrong objects [11], and struggle in generating visual text [36]. Today, there are efforts to address + +![](images/3cc23a705a88ae78a1e31f868561e4e485461978a9e8587b0d15aab95d63115c.jpg) +Figure 2. (a) Overview of how TIFA evaluates the faithfulness of a synthesized image. TIFA uses a language model (LM), a question-answering (QA) model, and a visual-question-answering (VQA) model. Given a text input, we generate several question-answer pairs with the LM and then filter them via the QA model. To evaluate the faithfulness of a synthesized image to the text input, a VQA model answers these visual questions using the image, and we check the answers for correctness. (b) TIFA v1.0 benchmark. While TIFA is applicable to any text prompt, to allow direct comparison across different studies, and for ease of use, we introduce the TIFA v1.0 benchmark, a repository of text inputs along with pre-generated question-answer tuples with answer choices. To evaluate a text-to-image model, a user first produces the images for the text inputs in TIFA v1.0 and then performs VQA with our provided tools on generated images to compute TIFA. + +![](images/002e14dbc71ea718f6bd90b0b7376dd20db7e1bea3d060e78ff49d5bea7299a4.jpg) + +these challenges: researchers are imposing linguistic structure with diffusion guidance to produce images with multiple objects [11]; others are designing reward models trained using human feedback to better align generations with user intention [30]. However, progress is difficult to quantify without accurate and interpretable evaluation measures that explain when and how models struggle. + +A critical bottleneck, therefore, is the lack of reliable automatic evaluation metrics for text-to-image generation faithfulness. One of the popular metrics is CLIPScore [17], which measures the cosine similarity between the CLIP embeddings [42] of the text input and the generated image. However, since CLIP is not effective at counting objects [42], or reasoning compositionally [37], CLIPScore is unreliable and often inaccurate. Another family of evaluation metrics uses image captions, in which an image captioning model first converts the image into text, and then the image caption is evaluated by comparing it against the text input. Unfortunately, using captioning models is insufficient since they might decide to ignore salient information in images or focus on other non-essential image regions [24]; for example, a captioning model might say that the images in Figure 1 are "a field of grass with trees in the background". Moreover, evaluating text (caption) generation is inherently challenging [23, 26]. Another recent text-to-image evaluation is DALL-Eval [6], which employs object detection to determine if the objects in the texts are in the generated images. However, this approach only works on synthesized text and measures faithfulness along the limited axes of objects, counting, colors, and spatial relationships but misses activities, geolocation, weather, time, materials, shapes, sizes, and other potential categories we often ask about when we recall images from memory [29]. + +To address the above challenges, we introduce TIFA, a new metric to evaluate text-to-image generation faithfulness. + +Our approach is illustrated in Figure 2. Given a repository of text inputs, we automatically generate question-answer pairs for each text via a language model (here, GPT-3 [3]). A question-answering (QA) system (here, UnifiedQA [25]) is subsequently used to verify and filter these question-answer pairs. To evaluate a generated image, we use a visual-question-answering (VQA) system (here, mPLUGlarge [31], BLIP-2 [32], etc.) to answer the questions given the generated image. We measure the image's faithfulness to the text input as the accuracy of the answers generated by the VQA system. While the accuracy of TIFA is dependent on the accuracy of the VQA model, our experiments show that TIFA has much higher correlation with human judgments than CLIPScore (Spearman's $\rho = 0.60$ vs. 0.33) and captioning-based approaches (Spearman's $\rho = 0.60$ vs. 0.34). Additionally, since the LMs and VQA models will continue to improve, we hypothesize that TIFA will continue to be more reliable over time. Also, our metrics can automatically detect when elements are missing in the generation: in Figure 2, TIFA detects that the generated image does not contain a TV. + +To promote the use of our new evaluation metric, we release TIFA v1.0, a large-scale text-to-image generation benchmark containing 4K diverse text inputs, sampled from the MSCOCO captions [34], DrawBench [47], PartiPrompts [61], and PaintSkill [6]. Each input comes with a pre-generated set of question-answer pairs, resulting in 25K questions covering 4.5K distinct elements. These questions have been automatically generated and pre-filtered using a question-answering model. This benchmark also comes with different VQA models [57, 28, 58, 32, 31, 21] that can be used to evaluate generative models and can be easily extended to use future VQA models when they become available. + +We conduct a comprehensive evaluation of current text- + +to-image models using TIFA v1.0. Thanks to TIFA's ability to detect fine-grained unfaithfulness in images, we find that current state-of-the-art models are good at rendering common objects, animals, and colors, but still struggle in composing multiple objects, reasoning about spatial relations, and binding the correct activity for each entity. In addition, our ablation experiments show that TIFA is robust to different VQA models. Future researchers can use TIFA v1.0 to compare their text-to-image models' faithfulness across different studies. Also, future generative models may focus on addressing the weaknesses of current models that TIFA discovered. In addition, with TIFA, users can customize evaluations with their own text inputs and questions [10]; for example, a future TIFA benchmark could focus on counting or scene text. + +# 2. Related Work + +We compare TIFA to other image and language generation evaluation metrics. + +Prior image generation evaluation Prior work usually compares image generation models via pairwise comparison by humans. How to design automatic evaluation metrics to approximate human assessment of the quality of machine-generated images has always been a major challenge in computer vision. There are two aspects to evaluate, namely image quality and image-text alignment. Inception Score [48] and FID [18] are the most widely adopted metrics for image quality. They compare the features of the generated images and gold images extracted from a pre-trained Inception-V3 model [52] to evaluate the fidelity and diversity of generated images. However, they rely on ground-truth images and are based on a classification model, which makes them not suitable for complex datasets [12]. For image-text alignment, prior metrics are mainly based on CLIP [42], captioning, and object detection models. CLIPScore [17] and CLIP-R [39] are based on the cosine similarity of image and text CLIP [42] embeddings. [6, 19, 20] first convert the images using a captioning model, and then compare the image caption with the text using metrics like CIDEr [55] and SPICE [1]. SOA [19] and DALL-Eval [6] employ object detection models to determine if objects, attributes, and relations in the text input are in the generated image. However, this approach only works on synthesized text inputs and measures faithfulness on limited axes (object, counting, color, spatial relation), missing elements like material, shape, activities, and context. In contrast, thanks to the flexibility of questions, TIFA works on any text inputs and evaluates faithfulness across a broad spectrum of dimensions. + +Summarization evaluation in NLP TIFA is inspired by the summarization evaluation methods based on question answering (QA) [56, 50]. Given a summary, a language model generates a set of questions about the text. A QA model checks if the same answer can be inferred from the + +text and the summary. These QA-based metrics have much higher correlations with human judgments on the factual consistency of summarization than other automatic metrics [56, 50]. TIFA can be seen as treating the text input for the text-to-image model as a summary of the generated image. + +# 3. The TIFA Metric + +We introduce a framework for automatically estimating the faithfulness of an image to its text prompt. Given a text input $T$ , we aim to measure the faithfulness of the generated image $I$ . An overview of our metric is illustrated in Figure 2. From $T$ , we generate $N$ multiple-choice question-answer tuples $\{Q_i, C_i, A_i\}_{i=1}^N$ , in which $Q_i$ is a question, $C_i$ is a set of answer choices, and $A_i \in C_i$ is the gold answer. The answer $A_i$ can be inferred given $T, Q_i$ , and $C_i$ . Next, for each question $Q_i$ , we use a VQA model to produce an answer $A_i^{\mathrm{VQA}} = \max_{a \in C_i} p(a \mid I, Q_i)$ . We define the faithfulness between the text $T$ and image $I$ as the VQA accuracy: + +$$ +f a i t h f u l n e s s (T, I) = \frac {1}{N} \sum_ {i = 1} ^ {N} \mathbb {1} \left[ A _ {i} ^ {\mathrm {V Q A}} = A _ {i} \right] \tag {1} +$$ + +The range of our faithfulness score is $[0, 1]$ . It is maximized when we have a performant VQA model, and the image $I$ accurately covers the information in the text $T$ so that for any question $Q$ , which can be answered given $T$ can also be answered given $I$ . Several key design decisions will be addressed in later sections: how to generate questions ( $\S 3.1$ ), how to control the question quality ( $\S 3.2$ ), and how to answer those questions ( $\S 3.3$ ). Finally, we give a step-by-step qualitative example of TIFA in Figure 4. + +# 3.1. Question-Answer Generation + +Our main challenge is to generate diverse questions that cover all elements of the text input evenly. We also simplify the question-generation pipeline into a single GPT-3 [3] completion, so that TIFA can exploit the power of recent language models (LM) and work with updated black-box LMs (e.g., ChatGPT) in the future. + +Inspired by prior work [4], given a text prompt $T$ , we generate the question-answer tuples $\{Q_i, C_i, A_i\}_{i=1}^N$ via the pipeline illustrated in Figure 3. Different from prior work, which relies on multiple components, our pipeline is completed by a single inference run via in-context learning with GPT-3 [3, 59, 22, 41, 51], thereby avoiding the need for intermediate human annotations. We annotate 15 examples and use them as in-context examples for GPT-3 to follow. Here we take the text "A photo of three dogs" as an example. Each in-context example contains the following steps: + +Element extraction Given text prompt $T$ , GPT-3 will first extract all elements $\{v_i\}_{i=1}^m$ following prior work [4] (for the in-context examples, we perform element extraction manually). The elements include noun phrases (including named + +![](images/6a1664ffedb426de384e4b27c58ce42eaa33029af954edf04727c5235e2a36cf.jpg) +Figure 3. Our question-answer pair generation pipeline. The whole pipeline can be executed via a single inference of GPT-3 via in-context learning. Given the text prompt, GPT-3 first extracts the elements and then generates two questions for each element. The GPT-3 output is then parsed and filtered by UnifiedQA. + +entities), verbs, adjectives, adverbs, and parse tree spans with no more than 3 words altogether. For the above example, the elements are photo, three, dogs. + +Element category classification For each element $v_{i}$ , following [29], we classify the elements into one of the following 12 categories: object, activity, animal, food, counting, color, material, spatial, location, shape, attribute, and other. As shown in Figure 3, text generated from GPT-3 contains the category corresponding to each question. For example, "three" is "counting", and "dogs" is classified as "animal." This step allows a detailed analysis of the text-to-image model's ability in each category. + +Question generation conditioned on elements For each element $v_{i}$ , we generate two questions. The first is a question that should be answered "yes" for a faithful generated image, and the second question has $v_{i}$ as its answer. For example, two questions are generated for the element "three". The first is "are there three dogs?", and the choices are {yes, no}. The second is "how many dogs are there?", and the choices are {1,2,3,4}. These two types of questions make our evaluations diverse and robust to surface-level differences. + +Completing above steps by prompting GPT-3 once As mentioned earlier, for each text $T$ , the whole pipeline can be completed by one GPT-3 inference. We annotated 15 + +in-context examples that cover all types of questions. The full prompt is in the Appendix. The prompt format is shown in Figure 3. Our in-context examples follow the same format, and identical examples are used for all text inputs, leading to a fixed and limited amount of human annotation cost. We use code-davinci-002 engine for question generation, and the decoding temperature is 0. + +# 3.2. Question Filtering + +To ensure the quality of generated images, we use UnifiedQA [25] to verify the GPT-3 generated question-answer pairs and filter out the ones that GPT-3 and UnifiedQA do not agree on. UnifiedQA $^2$ is a state-of-the-art multi-task question-answering model that can answer both multiple-choice and free-form questions. Denote the UnifiedQA model as $QA$ . Given the text $T$ , question $Q_{i}$ , choices $C_{i}$ , and answer $A_{i}$ , Let $A_{i}^{f} = QA(T,Q_{i})$ be the free-form answer, and $A_{i}^{mc} = QA(T,Q_{i},C_{i})$ be the multiple-choice answer. We keep the question if $A_{i} = A_{i}^{mc}$ and the word-level $F_{1}$ score between $A_{i}^{f}$ and $A_{i}$ is greater than 0.7. We conduct a human evaluation on 1000 filtered question-answer pairs. Only 7 are considered not reasonable (e.g., generated choices do not include a correct answer). Details are in Appendix C. + +Recommended VQA model Based on considerations over the accuracy, correlation with human judgments, and run time, we currently suggest using mPLUG-large as the VQA model for TIFA. Analysis is given in Section 5.4. Like the LM and QA components, the VQA component can be updated in the future as the technology improves. + +# 3.3. VQA Models + +Since our questions contain a diverse set of visual elements (e.g., activity, art style), we use open-domain pre-trained vision-language models as our VQA model (rather than closed-class classification models fine-tuned on VQAv2 [14]). We provide tools to easily perform VQA on arbitrary images and questions, based on 5 state-of-the-art VQA models trained with distinct data and strategies. + +Vision-language models The general pre-trained vision-language model are GIT-large [57], VILT-B/32 [28], OFA-large [58], and mPLUG-large [31]. These models are pretrained on a large amount of image-text pairs, and downstream image-to-text tasks like image captioning and visual question answering. Notice that these models have not been trained to answer multiple-choice questions. For each question, we first decode the free-form answer and then choose the choice that has the highest similarity with the decoded answer, measured by SBERT [45]. Another model we use is BLIP-2 FlanT5-XL [32], in which a VIT [8] is connected with a frozen FlanT5 [7] via a lightweight transformer. This + +Text Input: A wet dog standing on a beach next to the ocean. + +# Step 1: Question Generation + Filtering + +
About dog (animal/human)About ocean (location)About wet (attribute)
Q: Is this a dog?Q: Is there an ocean?Q: Is the dog wet?
A: YesA: YesA: Yes
Q: What animal is in the picture?Q: What type of place is this?Q: Is the dog wet or dry?
A: DogA: OceanA: Wet
About beach (location)About standing (activity)About next to (spatial relation)
Q: Is this a beach?Q: Is the dog standing?Q: Is the dog next to the ocean?
A: YesA: YesA: Yes
Q: What type of place is this?Q: What is the dog doing?Q: Is the dog next to or in the ocean?
A: BeachA: StandingA: Next to
+ +# Step 2: Visual Question Answering + +![](images/d7f364f9ab8707cafc9c8b69df7c0068c1d8080b00eeda6d0f7fb9e5bbbfb393.jpg) +Stable Diffusion v1.5 image + +![](images/0c1b2e85e33bca6cb3dac0200c8c07b2849ac37bc9f82b0bed02e94745197458.jpg) +Figure 4. Step-by-step qualitative example of TIFA metric. Given a text input, we first generate question-answer pairs and filter them. We strike through the questions filtered out by UnifiedQA. Then we run VQA models on the generated image to get the TIFA score. + +TIFA $7 / 10 = 0.7$ + +About dog (animal/human) + +Q: Is this a dog? + +A:Yes + +Q: What animal is in the picture? + +A: Dog + +![](images/5c7ee897bef39786145b2eb99461704d55798546ae145b39d7efdfbb077c88d0.jpg) + +About beach (location) + +Q: Is this a beach? + +A:Yes + +Q: What type of place is this? + +A:Beach + +![](images/a1527cc74f269bdf8891c2550978fbd1444cc40fff7bc6fbb0e2f3f664a8425e.jpg) + +About ocean (location) + +Q: Is there an ocean? + +A:Yes + +![](images/474e633457676a3587a2eff43a3d20eff1b2be1e0187a536884295bf2f1def45.jpg) + +About standing (activity) + +Q: Is the dog standing? + +A: No + +About wet (attribute) + +Q: Is the dog wet? + +A: No + +Q: Is the dog wet or dry? + +A: Dry + +About next to (spatial relation) + +Q: Is the dog next to the ocean? + +A:Yes + +Q: Is the dog next to or in the ocean? + +A: Next to + +![](images/154a1a6c0ae1168f586f12765bc1e45361922d68000e5b0eb7f41d5c0f5d602b.jpg) + +![](images/8f127a7f3ce2a7c1055b518c078a2a0abb59062804dd758d71a96d0aaeff9ebb.jpg) + +![](images/4ed20ead1959143d1542d950ff7df12fcc1713ed4b9df1e67f62cfa02cbe1fed.jpg) + +![](images/3af46fd7916c03519a8ac648491bd202dda28b23ee4f783ed31e744a6ab95dc0.jpg) + +![](images/da37f76ecdc14b69667ea3479244e558c11efe12da0ddbf6bf7dbb3cc8a55bd4.jpg) + +![](images/2d49daa3451cc2ec8cda8af22572489f4e3be5e0679e6d69f362b8157a3f9d2d.jpg) + +model allows for performing multiple-choice VQA directly due to the flexibility of the LM. + +# 4. FIFA v1.0: Benchmark for Text-to-Image Generation Faithfulness + +In this section, we introduce TIFA v1.0, a text-to-image generation faithfulness benchmark based on the evaluation method discussed in Section 3. The benchmark consists of 4,081 diverse text inputs paired with 25,829 question-answer pairs. Each question is classified into one of the categories discussed in Section 3.1. The benchmark also comes with Python pip-installable APIs to perform VQA with various state-of-the-art VQA models on arbitrary visual questions. The overall TIFA for each text-to-image model is computed by averaging TIFA scores of images generated from each text input in the benchmark. + +# 4.1. Text Collections + +We collect 4,081 text inputs to benchmark text-to-image models' generation ability on diverse tasks. 2,000 text inputs are image captions from COCO validation set [34]. These captions have corresponding gold images. Since text-to-image models are often used to create abstract art, we also + +collect 2,081 text inputs from previous works that do not correspond to any real image. All text inputs we use contain $\geq 3$ words. We include 161 from DrawBench used in Imagen [47] (texts that are categorized as "misspellings" and "rare words" are removed); 1420 from PartiPrompt used in Parti [61] (texts in category "abstract" are removed); and 500 texts from PaintSkill used in DALL-Eval [6]. + +Table 1. Statistics of FIFA v1.0. + +
Statistics
# of prompts4,081
- # of COCO captions2,000
- # of DrawBench, PartiPrompt, PaintSkill prompts2,081
# of questions25,829
- # of binary questions17,226
- # of multiple-choice questions8,603
avg. # of questions per prompt6.3
avg. # of words per prompt10.5
avg. # of elements per prompt4.3
+ +# 4.2. Statistics and Diversity + +Table 1 shows the basic statistics of the TIFA v1.0 benchmark. We demonstrate TIFA v1.0's diversity in Figure 5. + +a yellow t-shirt with a dog on it" + +"a panda bear with aviator glasses on its head" + +"A red colored car" + +"One cat and one dog sitting on the grass" + +![](images/6c61a26c3e469c31513795353603f160bebc79eae50b011355684b2a26063ae0.jpg) + +Object + +table,train,bus ... + +![](images/84d5b5f257ac863bf4c194bda68a306d7cb6d83d35f3404dc03a3945095fc19a.jpg) + +Spatial + +above,below,right .. + +![](images/a87d4ea750f0a2a6aaf32c66dd68948e0719305c054ece6e1914e6bb092c4fd2.jpg) + +Counting + +two,three,several + +![](images/ecaa51dfa75b45fd02357fc3ec25e9187a5de3f1e14f26be11c52d1743718cc9.jpg) +Figure 5. Statistics and diversity of TIFA v1.0. The text inputs contain elements from 12 categories (e.g., object, spatial, and counting). We show the most common elements from each category. In addition, we also show some example text inputs on the sides. + +Food + +pizza, bananas ... + +![](images/a592dbfe7c077069367c89fbfd0094178622657e535606df89a3252a9b221c31.jpg) + +Animal/Human + +man,woman,dog ... + +![](images/c7df1793d05328685c27b9822b2fb87d5c02eb4b10d09e2380b8a9dc5d6980f5.jpg) + +Attribute + +large, small, young ... + +![](images/2f23bfa4bb47ef71f24f07f2ef2984f4140817fd4e18026b42d2df91e241681f.jpg) + +TIFA v1.0 + +4,081 text inputs 25,829 questions +4,550 distinct elements 12 categories + +![](images/002405c7513a53fd78d1b9f5eb5e76ff2af1dc6807e318d290b766de13f393c7.jpg) + +Material + +wooden,metal,glass + +![](images/3902b528b991b62460dda4e572fb9ef6e9ec734600dec1732ceb3fdb230250bd.jpg) + +Shape + +circular,square + +![](images/fccdb2ed2a7a22286cc81991007956b07612bbdb2d0f31087d2cbd69302a72e3.jpg) + +Activity + +sitting.standing ... + +![](images/c7b57b159055b870cc30565355e60f8b78442d2cb03e1076794543b0623eeb09.jpg) + +Location + +field,street,beach ... + +![](images/abe62ee34e8e14d8d4f75fa01eeb7b06c591355111013fb98f786d9077ec3e5a.jpg) + +Color + +red,white,blue ... + +![](images/afb9daf7febb133252aba567033da63eb2106c6c7ea5461fe92df0678b7cab77.jpg) + +Other + +punk, starry night + +"a painting of a fox + +in the style + +starry night + +"A man running to + +hit a tennis ball + +"a flag with a + +dinosaur on it" + +"a photo of suitcase + +and chair; chair is + +right to suitcase" + +TIFA v1.0 contains questions about 4,550 distinct elements, which are categorized into 12 categories. The number of times each type of element occurs in the text input is object (7,854), animal/human (3,501), attribute (3,399), activity (2,851), spatial (2,265), location (1,840), color (1,743), counting (986), food (911), material (209), shape (69), and other (201). The "other" category includes notions often used in abstract art, such as "starry night" and "steampunk." The accuracy of VQA on a particular genre measures the text-to-image model's ability in the corresponding aspect. Please refer to Appendix D for more details on TIFA v1.0. + +# 4.3. Finetuned Open-Source Language Model for Question Generation + +For TIFA v1.0, we use GPT-3 to generate the questions. While benchmarking with TIFA v1.0 is a deterministic process, using TIFA to create a new benchmark might not be deterministic as the underlying question generator (GPT-3 in our case) might change privately. To promote deterministic benchmark generation, we fine-tune and release a LLaMA 2 (7B) [54] model that parses the captions and generates questions for arbitrary texts, using TIFA v1.0 questions as training examples. + +# 5. Experiments + +In this section, we first show that TIFA has substantially higher correlations with human judgments than prior metrics on text-to-image faithfulness (§5.1). Then we present a comprehensive evaluation of existing text-to-image models using TIFA v1.0, highlighting the challenges of current text-to-image models (§5.3). Finally, we conduct an analysis of TIFA's robustness against different VQA models (§5.4). For all experiments, we use mPLUG as the VQA model for TIFA unless stated otherwise. The models we evaluate include AttnGAN [60], X-LXMERT [5], VQ-Diffusion [16], minDALL-E [27], and Stable Diffusion v1.1, v1.5, and v2.1 [46]. Details are in Appendix E. + +# 5.1. Correlation with Human Judgements + +To compare TIFA with prior evaluation metrics, we first conduct human evaluations of the text-to-image models on the 1-5 Likert scale on text-to-image faithfulness. Then we compare TIFA with other metrics based on their correlation with human judgments. + +Likert scale on text-to-image faithfulness Annotators are asked to answer on a scale of 1 (worst) to 5 (best) to the question "Does the image match the text?" The detailed annotation guidelines are in Appendix C. Annotators are asked to focus on text-to-image faithfulness rather than image quality. The Likert scale should be based on how many elements in the text prompt are missed or misrepresented in the image. Objects are more important than attributes, relations, and activities. If an object is missed in the image, then all related attributes, activities, relations, etc. are also considered lost. An example is given in Figure 6. + +We collect annotations of 800 generated images on 160 text inputs from TIFA v1.0. For each prompt, we sample an image from the 5 most recent generative models we evaluated, i.e., minDALL-E, VQ-Diffusion, Stable Diffusion v1.1, v1.5, and v2.1. We collect 2 annotations per image and average over the scores as the single "faithfulness" score. The inter-annotator agreement measured by Krippendorf's $\alpha$ is 0.67, indicating "substantial" agreement. + +![](images/dfc882a6b0fa1891f1c05c879f224d3362053a2e8e95be4b73c58bac9be2b93e.jpg) +Prompt: +Figure 6. Illustration of our Likert scale annotation guideline. Annotators are asked to give a score of 1 to 5 based on how many elements in the text prompt are missed or misrepresented in the image. The missed elements are underlined. + +A person sitting on a horse in air over gate in grass with people and trees in background. + +Miss/Misrepresent + +4 out of 10 elements! + +Baselines We compare our evaluation with two families of reference-free metrics on text-image match introduced in + +Section 2. The first is the caption-based method. We use the state-of-the-art BLIP-2 FlanT5-XL [32] as the captioning model. The second approach is CLIPScore [17, 42]. We use CLIP (VIT-B/32) [42] to compute the score. + +Table 2. Correlations between each evaluation metric and human judgment on text-to-image faithfulness, measured by Spearman's $\rho$ and Kendall's $\tau$ . + +
Spearman's ρKendall's τ
Caption-Based
BLEU-418.318.8
ROUGE-L32.924.5
METEOR34.027.4
SPICE32.823.2
CLIPScore33.223.1
Ours
TIFA (VILT)49.338.2
TIFA (OFA)49.637.2
TIFA (GIT)54.542.6
TIFA (BLIP-2)55.943.6
TIFA (mPLUG)59.747.2
+ +TIFA has a much higher correlation with human judgments than prior metrics. The correlations between each evaluation metric and human judgment are shown in Table 2. For caption-based evaluations, we use metrics BLEU-4 [38], ROUGE-L [33], METEOR [2], and SPICE [1]. TIFA has higher correlations with human judgments than all previous evaluation metrics on all VQA models. TIFA (mPLUG) yields the highest correlation with human judgments among all VQA models. + +# 5.2. Benchmarking Text-to-Image Models + +Figure 7 shows the average TIFA score text-to-image models get on TIFA v1.0. The detailed scores with each VQA model on each element type are provided in Appendix B. We can see a clear trend of how text-to-image models evolve over time. There is a jump in TIFA score after DALL-E [44] is released, from about $60\%$ to $75\%$ . Qualitative examples of our evaluation metric are in Appendix A. + +![](images/f6444ce6e713d0de1c4dd28348aee4160542840b0ce2899eeeea58b43bcdfcaf.jpg) +Figure 7. Average FIFA score of text-to-image models on the FIFA v1.0 benchmark. The horizontal axis shows their release dates. + +# 5.3. Findings on Current Text-to-Image Models + +Figure 8 shows accuracy on each type of question in TIFA v1.0 for Stable Diffusion v1.1, v1.5, and v2.1. The score on each type reflects the text-to-image models' faithfulness in each type of visual element. To the best of our knowledge, TIFA is the only automatic evaluation method that can provide such a detailed fine-grained analysis of image generation. We separate the scores on COCO captions and other text inputs. For COCO captions, we also include the accuracy on the ground-truth images for reference. We summarize our findings in the following paragraphs. + +Generating images from captions vs. free-form text From Figure 8, we can see that VQA accuracy is higher on the COCO captions than on other text inputs. The reason is that COCO captions correspond to real images, while other text inputs may correspond to compositions that cannot be found in real-world photos (e.g. "a blue apple"). + +# What elements are text-to-image models struggling with? + +Based on the scores of each category in Figure 8, we can see that Stable Diffusion models are performing well on material, animal/human, color, and location in terms of text-to-image faithfulness. However, they yield low accuracy on questions involving shapes, counting, and spatial relations. "Other" mainly contains abstract art notions, and models are also struggling with them. There is also a big gap between the synthesized images and real images on the COCO captions. Future work can explore various directions (e.g., training data/loss and model architecture) to improve text-to-image models' faithfulness in these aspects. + +# Why are ground-truth images not getting perfect scores? + +Ground-truth images in COCO do not get perfect scores because 1) the COCO captions contain a substantial amount of noise from crowd workers [24] and 2) VQA models are not perfect. Real images have higher accuracy in all categories except material, color and location, where differences are small. It is left to future work to determine whether this is simply due to noise or it is an area where assessment can be improved. + +Stable Diffusion is evolving. We can see the consistent trend that Stable Diffusion models are improving in their later versions in most of the element categories. The exceptions are "shape" for both prompt sources, "other" and "food" for the free-form text inputs without gold images. + +Composing multiple objects is challenging. Figure 9 shows how the number of entities (objects, animals/humans, food) in the text input affects the average TIFA score. When there are more than 5 entities, The TIFA score starts to drop rapidly for all text-to-image models, consistent with similar findings in other vision-language evaluations [15, 13]. For reference, we also add the real images in COCO in this figure. + +VQA accuracy (by category) on COCO captions +![](images/efcb7be6c7f5995a2fb4aff98010b343b41cdc22eb123bc70161d7238ea3d443.jpg) +Stable Diffusion v1.1 Stable Diffusion v1.5 Stable Diffusion v2.1 COCO Ground-Truth + +![](images/3b54bec0300be6452b7e8f6d98172c093c1b8c0479f62ef377ac91e13b57a970.jpg) +VQA accuracy (by category) on free-form text +Figure 8. Accuracy on each type of question in the FIFA v1.0 benchmark. The text-to-image models are Stable Diffusion v1.1, v1.5, and v2.1. We order the categories by the average score Stable Diffusion v2.1 gets on corresponding questions. For COCO captions, we also include the accuracy of the ground-truth images for reference. + +![](images/bd62f80e860bca3bec250310b2051b16b28f57dd17450724aa49c51322947c3d.jpg) +Figure 9. TIFA vs. numbers of entities (objects, animals/humans, and food) in the text input. The accuracy starts to drop when more than 5 entities are added to the text, showing that compositionality is hard for text-to-image models. Meanwhile, TIFA scores for COCO ground-truth (GT) images remain consistent. + +The TIFA score on real images is rather consistent and does not change as the number of entities increases. This quantitatively shows that composing multiple objects is challenging for current text-to-image models. One possible reason is that the CLIP text embedding, which is used to train Stable Diffusion, lacks compositionality, as investigated in [37]. + +# 5.4. Analysis of VQA Models + +One major concern of TIFA is that VQA models can introduce some errors. Table 2 shows that TIFA has a much higher correlation with human judgment than the previous metrics, regardless of the choice of the VQA models; here we conduct a more detailed analysis. + +Sensitivity of TIFA to VQA models Figure 10 shows several recent text-to-image models' TIFA scores on the COCO captions in TIFA v1.0, measured by different VQA models. We also include the TIFA scores on the ground-truth COCO + +![](images/a0cef53a35d4ba2de499f15708435e4066ee30f3b94183cc9d6f39d93e7fab58.jpg) +Figure 10. Several text-to-image models' TIFA score on COCO captions, measured by different VQA models. We also include the accuracy of ground-truth images for reference. + +images for reference. TIFA scores computed by different VQA models show a similar trend on these text-to-image models. Also, the ground-truth images get the highest TIFA score. We also computed Spearman's $\rho$ of TIFA scores given by different VQA models. The pairwise correlation between all VQA models is greater than 0.6. + +Humans performing VQA To conduct further analysis on the VQA models, we ask annotators to answer the multiple-choice visual questions in TIFA v1.0. These annotations help us evaluate the accuracy of each VQA model. For multiple-choice questions, we add the option "None of the above" for human evaluation. Annotation guidelines are in Appendix C. + +We collect annotations of 1029 questions on 126 images. Each question is answered by two annotators. The inter-annotator agreement measured by Krippendorf's $\alpha$ is 0.88. A third annotator is involved if two annotators disagree, and the final answer is chosen by the majority vote. + +# Which VQA model should we use? + +Table 3 reports the accuracy of each VQA model and the correlation between TIFA scores calculated by VQA model answers and human answers. We observe that higher model performance is directly related to the TIFA score's + +Table 3. Comparison of VQA models. The first row is the VQA accuracy, using the human VQA answers as reference. The second row is Spearman's correlation between TIFA scores calculated by each VQA model and the human VQA. + +
VILTOFAGITBLIP-2mPLUG
VQA Acc.76.177.179.181.084.5
TIFA Corr.60.963.772.575.676.8
+ +correlation with human judgments. mPLUG has the highest accuracy. + +Another important factor to consider is the runtime. We measure the inference speed of each VQA model on NVIDIA A40 GPU with batch size 1 over the Stable Diffusion v2.1 images (768 × 768 pixels). For one question, VILT takes 0.08s on average; OFA, GIT, and mPLUG all take about 0.25s; BLIP-2 takes 0.73s. Based on the above results, we choose mPLUG as the default VQA model for TIFA v1.0 because it is the most accurate while being reasonably fast. + +Separation of Text-to-Image Errors and VQA Errors Suppose an image gets a wrong answer given a visual question. Then the image generation or the VQA model might have made an error. Based on the human VQA results, we separate these two kinds of errors in Figure A. If human VQA gives the wrong answer, then we suspect the generated image has an error. Otherwise, the image is correct but the VQA model is making an error. Figure A shows that the majority of errors are made by the text-to-image models. For mPLUG, less than $25\%$ errors are due to the VQA model. This suggests that the TIFA framework is a viable evaluation method despite its inherent challenges. + +![](images/5c4e0034b937bd063a8987e75ea85a0d28ce9c67dc0139cca89462041cd65461.jpg) +Figure 11. Source of the error when VQA gets the wrong answer. + +# 6. Discussion + +Why does TIFA work better than CLIPScore? Our experimental results and human evaluations in Section 5 show that TIFA is more accurate than prior metrics (CLIP [42] and captioning-based approaches) for evaluating text-to-image faithfulness. We hypothesize that the major challenge of these prior metrics is that they summarize the image outputs and text inputs into a single representation (embedding/caption). In contrast, TIFA exploits the power of the language models to decompose the text input into fine-grained probes, which allows VQA to capture more nuanced aspects of the text input and the generated image. + +How do current VQA models perform on TIFA? One limitation is that TIFA requires VQA models to work reasonably well, which is true given the current models and TIFA v1.0, as shown in Section 5. Nevertheless, the assumption might not hold for current models in domains like anime and abstract art. TIFA is a modularized evaluation framework. The VQA models used within the framework can be updated as stronger VQA models become available in the future. For instance, we plan to incorporate GPT-4 once its image API is made public since it is likely to improve TIFA. Another possible solution is to ensemble multiple image understanding models. For example, one may employ expert models on art concepts. We leave this for future work. + +Other limitations. Another limitation of TIFA is its runtime. Answering multiple visual questions is slower than one CLIP inference. In the scenario described in Section 5.4, mPLUG takes 1.6s to evaluate one image (without batching). Also, our question generation pipeline needs one inference on a modern language model for each text input. The run time is not a critical issue for benchmarking purposes, but may not be computationally feasible for the kind of large-scale data filtering done, for example, in LAION-5B [49]. Nevertheless, we would like to point out that our evaluation is much faster than the image generation process of diffusion models. Thus, we believe it is feasible to perform reranking and reinforcement learning with TIFA on diffusion models. + +# 7. Conclusions + +We present TIFA, a new automatic text-to-image faithfulness evaluation metric using VQA. Compared with prior metrics, TIFA is fine-grained, interpretable, and better aligned with human judgments. Based on this metric, we introduce the TIFA v1.0, a large-scale text-to-image benchmark containing 4K prompts and 25K questions. We conduct a comprehensive study of current text-to-image models using TIFA v1.0 and highlight the limitations of current generative models. We quantitatively show that current image generation models still struggle in counting, spatial relations, and composing multiple objects. Finally, we conduct extensive analysis and human evaluation, demonstrating that TIFA is robust to different VQA models. We hope TIFA will help evaluate future work on image generation and become increasingly sophisticated as it is upgraded with new LM, QA, and VQA components. + +# 8. Acknowledgements + +We thank Zirong Ye, Cheng-Yu Hsieh, Oscar Michel, Enhao Zhang, Jiafei Duan, Weijia Shi, Jieyu Zhang, the TIAL group, the ARK group, and the GRAIL lab at UW for their helpful feedback on this work. We would also thank all anonymous reviewers for providing insightful and constructive comments on our work. + +# References + +[1] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In European conference on computer vision, pages 382-398. Springer, 2016. 3, 7 +[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7 +[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2, 3, 13 +[4] Soravit Changpinyo, Doron Kukliansky, Idan Szpektor, Xi Chen, Nan Ding, and Radu Soricut. All you may need for vqa are image captions. In North American Chapter of the Association for Computational Linguistics, 2022. 3 +[5] Jaemin Cho, Jiasen Lu, Dustin Schwenk, Hannaneh Hajishirzi, and Aniruddha Kembhavi. X-lxmert: Paint, caption and answer questions with multi-modal transformers. ArXiv, abs/2009.11278, 2020. 6, 14, 15 +[6] Jaemin Cho, Abhaysinh Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers. ArXiv, abs/2202.04053, 2022. 2, 3, 5 +[7] Hyung Won Chung, Le Hou, S. Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Wei Yu, Vincent Zhao, Yanping Huang, Andrew M. Dai, Hongkun Yu, Slav Petrov, Ed Huai hsin Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc Le, and Jason Wei. Scaling instruction-finetuned language models. ArXiv, abs/2210.11416, 2022. 4 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 4 +[9] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868-12878, 2020. 14 +[10] Kawin Ethayarajh and Dan Jurafsky. Utility is in the eye of the user: A critique of NLP leaderboards. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 3 +[11] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, P. Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. *ArXiv*, abs/2212.05032, 2022. 1, 2 + +[12] Stanislav Frolov, Tobias Hinz, Federico Raue, Jorn Hees, and Andreas R. Dengel. Adversarial text-to-image synthesis: A review. Neural networks : the official journal of the International Neural Network Society, 144:187-209, 2021. 3 +[13] Mona Gandhi, Mustafa Omer Gul, Eva Prakash, Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Measuring compositional consistency for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5046-5055, 2022. 7 +[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4 +[15] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11287-11297, 2021. 7 +[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10686-10696, 2021. 6, 14, 15 +[17] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Joseph Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Conference on Empirical Methods in Natural Language Processing, 2021. 2, 3, 7 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017. 3 +[19] Tobias Hinz, Stefan Heinrich, and Stefan Wermter. Semantic object accuracy for generative text-to-image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:1552-1565, 2019. 3 +[20] Seunghoon Hong, Dingdong Yang, Jongwook Choi, and Honglak Lee. Inferring semantic layout for hierarchical text-to-image synthesis. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7986-7994, 2018. 3 +[21] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. 2 +[22] Yushi Hu, Chia-Hsuan Lee, Tianbao Xie, Tao Yu, Noah A Smith, and Mari Ostendorf. In-context learning for few-shot dialogue state tracking. arXiv preprint arXiv:2203.08568, 2022. 3 +[23] Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras, Lavinia Dunagan, Jacob Morrison, Alexander Fabbri, Yejin Choi, and Noah A. Smith. Bidimensional leaderboards: Generate and evaluate language hand in hand. In North American Chapter of the Association for Computational Linguistics, 2022. 2 +[24] Jungo Kasai, Keisuke Sakaguchi, Lavinia Dunagan, Jacob Daniel Morrison, Ronan Le Bras, Yejin Choi, and Noah A. Smith. Transparent human evaluation for image captioning. + +In North American Chapter of the Association for Computational Linguistics, 2022. 2, 7 +[25] Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single qa system. In Findings, 2020. 2, 4, 13 +[26] Daniel Khashabi, Gabriel Stanovsky, Jonathan Bragg, Nicholas Lourie, Jungo Kasai, Yejin Choi, Noah A. Smith, and Daniel Weld. GENIE: Toward reproducible and standardized human evaluation for text generation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 2022. 2 +[27] Saehoon Kim, Sanghun Cho, Chiheon Kim, Doyup Lee, and Woonhyuk Baek. mindall-e on conceptual captions, 2021. 6, 14, 15 +[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 2, 4 +[29] Ranjay Krishna, Michael S. Bernstein, and Li Fei-Fei. Information maximizing visual question generation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2008–2018, 2019. 2, 4 +[30] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 1, 2 +[31] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng da Cao, Ji Zhang, Songfang Huang, Feiran Huang, Jingren Zhou, and Luo Si. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. In Conference on Empirical Methods in Natural Language Processing, 2022. 2, 4, 13 +[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597, 2023. 2, 4, 7 +[33] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain, July 2004. Association for Computational Linguistics. 7 +[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2, 5, 13, 15 +[35] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum. Compositional visual generation with composable diffusion models. ArXiv, abs/2206.01714, 2022. 1 +[36] Rosanne Liu, Daniel H Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, R. J. Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. ArXiv, abs/2212.10562, 2022. 1 + +[37] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. Crepe: Can vision-language foundation models reason compositionally? ArXiv, abs/2212.07796, 2022. 2, 8 +[38] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 7 +[39] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 3 +[40] Vitali Petsiuk, Alexander E Siemann, Saisamrit Surbehera, Zad Chin, Keith Tyser, Gregory Hunter, Arvind Raghavan, Yann Hicke, Bryan A Plummer, Ori Kerret, et al. Human evaluation of text-to-image models on a multi-task benchmark. arXiv preprint arXiv:2211.12112, 2022. 1 +[41] Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, and Mike Lewis. Measuring and narrowing the compositionality gap in language models, 2022. 3 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 7, 9, 14 +[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1 +[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 1, 7, 15 +[45] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. ArXiv, abs/1908.10084, 2019. 4 +[46] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 1, 6, 14, 16 +[47] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1, 2, 5, 13, 15 +[48] Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. ArXiv, abs/1606.03498, 2016. 3 +[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next gener + +ation image-text models. arXiv preprint arXiv:2210.08402, 2022.9 +[50] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 3 +[51] Hongjin Su, Jungo Kasai, Chen Henry Wu, Weijia Shi, Tianlu Wang, Jiayi Xin, Rui Zhang, Mari Ostendorf, Luke Zettlemoyer, Noah A. Smith, and Tao Yu. Selective annotation makes language models better few-shot learners. In The Eleventh International Conference on Learning Representations, 2023. 3 +[52] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2826, 2015. 3 +[53] Hao Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. ArXiv, abs/1908.07490, 2019. 15 +[54] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 6 +[55] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2015. 3 +[56] Alex Wang, Kyunghyun Cho, and Mike Lewis. Asking and answering questions to evaluate the factual consistency of summaries. In Annual Meeting of the Association for Computational Linguistics, 2020. 3 +[57] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2, 4 +[58] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR, abs/2202.03052, 2022. 2, 4 +[59] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 3 +[60] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316-1324, 2017. 6, 14, 15 +[61] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei + +Yang, Burcu Karagol Ayan, Benton C. Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. ArXiv, abs/2206.10789, 2022. 1, 2, 5 \ No newline at end of file diff --git a/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/images.zip b/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..baf5c1b0d2b76e593ca66625c7bbe445fb65add9 --- /dev/null +++ b/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d26bfa34877c6879b9e05350c5ab08dc67990d18feb6f80bc07bce6934ed747 +size 483547 diff --git a/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/layout.json b/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..50428d6191a47d603b2a2952c992d86d5cd33e86 --- /dev/null +++ b/2023/TIFA_ Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering/layout.json @@ -0,0 +1,13434 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 487, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 487, + 140 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 487, + 140 + ], + "type": "text", + "content": "TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 160, + 447, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 160, + 447, + 219 + ], + "spans": [ + { + "bbox": [ + 141, + 160, + 447, + 219 + ], + "type": "text", + "content": "Yushi Hu1 Benlin Liu1 Jungo Kasai1 Yizhong Wang1 Mari Ostendorf1 Ranjay Krishna1,2 Noah A. Smith1,2 1University of Washington 2Allen Institute for AI https://tifa-benchmark.github.io/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 270, + 290, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 290, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 290, + 582 + ], + "type": "text", + "content": "Despite thousands of researchers, engineers, and artists actively working on improving text-to-image generation models, systems often fail to produce images that accurately align with the text inputs. We introduce TIFA (Text-to-Image Faithfulness evaluation with question Answering), an automatic evaluation metric that measures the faithfulness of a generated image to its text input via visual question answering (VQA). Specifically, given a text input, we automatically generate several question-answer pairs using a language model. We calculate image faithfulness by checking whether existing VQA models can answer these questions using the generated image. TIFA is a reference-free metric that allows for fine-grained and interpretable evaluations of generated images. TIFA also has better correlations with human judgments than existing metrics. Based on this approach, we introduce TIFA v1.0, a benchmark consisting of 4K diverse text inputs and 25K questions across 12 categories (object, counting, etc.). We present a comprehensive evaluation of existing text-to-image models using TIFA v1.0 and highlight the limitations and challenges of current models. For instance, we find that current text-to-image models, despite doing well on color and material, still struggle in counting, spatial relations, and composing multiple objects. We hope our benchmark will help carefully measure the research progress in text-to-image synthesis and provide valuable insights for further research.1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 612, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 612, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 612, + 288, + 684 + ], + "type": "text", + "content": "While we welcome artistic freedom when we commission art from artists, images produced by deep generative models [44, 46, 43, 47, 61] should conform closely to our requests. Despite the advances in generative models, it is still challenging for models to produce images faithful to users' intentions [40, 11, 30, 35, 36]. For example, current" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 341, + 255, + 421, + 334 + ], + "blocks": [ + { + "bbox": [ + 348, + 247, + 414, + 255 + ], + "lines": [ + { + "bbox": [ + 348, + 247, + 414, + 255 + ], + "spans": [ + { + "bbox": [ + 348, + 247, + 414, + 255 + ], + "type": "text", + "content": "Stable Diffusion v1.5" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 255, + 421, + 334 + ], + "lines": [ + { + "bbox": [ + 341, + 255, + 421, + 334 + ], + "spans": [ + { + "bbox": [ + 341, + 255, + 421, + 334 + ], + "type": "image", + "image_path": "0534cb1a249c7e271671e89f343dc2e4cfa429059bd379d85675ff68f55c136b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 335, + 520, + 355 + ], + "lines": [ + { + "bbox": [ + 331, + 335, + 520, + 355 + ], + "spans": [ + { + "bbox": [ + 331, + 335, + 520, + 355 + ], + "type": "text", + "content": "Text Input: A person sitting on a horse in air over gate in grass with people and trees in background." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 433, + 255, + 514, + 334 + ], + "blocks": [ + { + "bbox": [ + 443, + 247, + 508, + 255 + ], + "lines": [ + { + "bbox": [ + 443, + 247, + 508, + 255 + ], + "spans": [ + { + "bbox": [ + 443, + 247, + 508, + 255 + ], + "type": "text", + "content": "Stable Diffusion v2.1" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 433, + 255, + 514, + 334 + ], + "lines": [ + { + "bbox": [ + 433, + 255, + 514, + 334 + ], + "spans": [ + { + "bbox": [ + 433, + 255, + 514, + 334 + ], + "type": "image", + "image_path": "984f040b7734452145bb0a652872438939cbafdb1886a69c1c3b86df988b4220.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 332, + 360, + 519, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 360, + 519, + 369 + ], + "spans": [ + { + "bbox": [ + 332, + 360, + 519, + 369 + ], + "type": "text", + "content": "GPT-3 generated + verified QAs (pre-generated in TIFA v1.0 benchmark)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 372, + 414, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 372, + 414, + 381 + ], + "spans": [ + { + "bbox": [ + 324, + 372, + 414, + 381 + ], + "type": "text", + "content": "Question: what is the animal?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 421, + 372, + 521, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 372, + 521, + 381 + ], + "spans": [ + { + "bbox": [ + 421, + 372, + 521, + 381 + ], + "type": "text", + "content": "Answer inferred from text: horse" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 339, + 388, + 359, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 388, + 359, + 397 + ], + "spans": [ + { + "bbox": [ + 339, + 388, + 359, + 397 + ], + "type": "text", + "content": "VQA:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 371, + 388, + 391, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 388, + 391, + 397 + ], + "spans": [ + { + "bbox": [ + 371, + 388, + 391, + 397 + ], + "type": "text", + "content": "Horse" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 406, + 386, + 410, + 398 + ], + "blocks": [ + { + "bbox": [ + 406, + 386, + 410, + 398 + ], + "lines": [ + { + "bbox": [ + 406, + 386, + 410, + 398 + ], + "spans": [ + { + "bbox": [ + 406, + 386, + 410, + 398 + ], + "type": "image", + "image_path": "2bf43cc660fd0cbecdd7a6a05464fa8a5a8de1811ed94b32e0d27b25c31d869c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 466, + 388, + 485, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 388, + 485, + 396 + ], + "spans": [ + { + "bbox": [ + 466, + 388, + 485, + 396 + ], + "type": "text", + "content": "Horse" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 491, + 387, + 503, + 398 + ], + "blocks": [ + { + "bbox": [ + 491, + 387, + 503, + 398 + ], + "lines": [ + { + "bbox": [ + 491, + 387, + 503, + 398 + ], + "spans": [ + { + "bbox": [ + 491, + 387, + 503, + 398 + ], + "type": "image", + "image_path": "a103acb97b718e129bf14bd098cc04876cefcfa5e0ea8305c5e4819e6d873079.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 324, + 404, + 402, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 404, + 402, + 413 + ], + "spans": [ + { + "bbox": [ + 324, + 404, + 402, + 413 + ], + "type": "text", + "content": "Question: is there a gate?" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 339, + 417, + 359, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 417, + 359, + 426 + ], + "spans": [ + { + "bbox": [ + 339, + 417, + 359, + 426 + ], + "type": "text", + "content": "VQA:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 375, + 418, + 386, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 418, + 386, + 426 + ], + "spans": [ + { + "bbox": [ + 375, + 418, + 386, + 426 + ], + "type": "text", + "content": "No" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 394, + 418, + 405, + 427 + ], + "blocks": [ + { + "bbox": [ + 394, + 418, + 405, + 427 + ], + "lines": [ + { + "bbox": [ + 394, + 418, + 405, + 427 + ], + "spans": [ + { + "bbox": [ + 394, + 418, + 405, + 427 + ], + "type": "image", + "image_path": "d87f16d37f52a8358d4b071322c5ce2b06a396dd438a98129ab3fecb1f8942dc.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 422, + 404, + 515, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 404, + 515, + 413 + ], + "spans": [ + { + "bbox": [ + 422, + 404, + 515, + 413 + ], + "type": "text", + "content": "Answer inferred from text: yes" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 469, + 418, + 482, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 469, + 418, + 482, + 426 + ], + "spans": [ + { + "bbox": [ + 469, + 418, + 482, + 426 + ], + "type": "text", + "content": "Yes" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 492, + 417, + 503, + 427 + ], + "blocks": [ + { + "bbox": [ + 492, + 417, + 503, + 427 + ], + "lines": [ + { + "bbox": [ + 492, + 417, + 503, + 427 + ], + "spans": [ + { + "bbox": [ + 492, + 417, + 503, + 427 + ], + "type": "image", + "image_path": "1d0ee3238ee5c9725828c47e0a85ad4a3bd73aa5ab323447b1f80b863c4c3cc7.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 324, + 434, + 411, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 434, + 411, + 443 + ], + "spans": [ + { + "bbox": [ + 324, + 434, + 411, + 443 + ], + "type": "text", + "content": "Question: is the horse in air?" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 339, + 445, + 359, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 445, + 359, + 455 + ], + "spans": [ + { + "bbox": [ + 339, + 445, + 359, + 455 + ], + "type": "text", + "content": "VQA:" + } + ] + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 375, + 445, + 386, + 456 + ], + "blocks": [ + { + "bbox": [ + 375, + 445, + 386, + 456 + ], + "lines": [ + { + "bbox": [ + 375, + 445, + 386, + 456 + ], + "spans": [ + { + "bbox": [ + 375, + 445, + 386, + 456 + ], + "type": "image", + "image_path": "5522b1600f668c49eac7cac3072645156e09fd921715ae997a7bbf79243f1d91.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 394, + 445, + 405, + 456 + ], + "blocks": [ + { + "bbox": [ + 394, + 445, + 405, + 456 + ], + "lines": [ + { + "bbox": [ + 394, + 445, + 405, + 456 + ], + "spans": [ + { + "bbox": [ + 394, + 445, + 405, + 456 + ], + "type": "image", + "image_path": "12d21720e3a232654f85b354ee1e1e2e64a08e693896616df2d1b7e4d45a986e.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 410, + 445, + 421, + 456 + ], + "blocks": [ + { + "bbox": [ + 410, + 445, + 421, + 456 + ], + "lines": [ + { + "bbox": [ + 410, + 445, + 421, + 456 + ], + "spans": [ + { + "bbox": [ + 410, + 445, + 421, + 456 + ], + "type": "image", + "image_path": "bd0192131f3afc4d22357421e1c9860357c0a4db85eb95034afaede286fe48f0.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 422, + 445, + 439, + 456 + ], + "blocks": [ + { + "bbox": [ + 422, + 445, + 439, + 456 + ], + "lines": [ + { + "bbox": [ + 422, + 445, + 439, + 456 + ], + "spans": [ + { + "bbox": [ + 422, + 445, + 439, + 456 + ], + "type": "image", + "image_path": "6a109c7eb64024dfb4799df213666bd0459271ee1743504b90ae74869342987f.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 317, + 462, + 334, + 483 + ], + "blocks": [ + { + "bbox": [ + 317, + 462, + 334, + 483 + ], + "lines": [ + { + "bbox": [ + 317, + 462, + 334, + 483 + ], + "spans": [ + { + "bbox": [ + 317, + 462, + 334, + 483 + ], + "type": "image", + "image_path": "5a7d0228e334eb274d7795337a069b7531a0c9550f8e7779439ece13b3af090e.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "bbox": [ + 336, + 472, + 388, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 472, + 388, + 482 + ], + "spans": [ + { + "bbox": [ + 336, + 472, + 388, + 482 + ], + "type": "text", + "content": "TIFA 71.4" + } + ] + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 374, + 482, + 386, + 491 + ], + "blocks": [ + { + "bbox": [ + 374, + 482, + 386, + 491 + ], + "lines": [ + { + "bbox": [ + 374, + 482, + 386, + 491 + ], + "spans": [ + { + "bbox": [ + 374, + 482, + 386, + 491 + ], + "type": "image", + "image_path": "121579668e855fc1d5a60de5ef1fabadaf8939a1c28a315c435cc855e89249b9.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 391, + 485, + 403, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 485, + 403, + 492 + ], + "spans": [ + { + "bbox": [ + 391, + 485, + 403, + 492 + ], + "type": "text", + "content": "ined" + } + ] + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 408, + 485, + 421, + 492 + ], + "blocks": [ + { + "bbox": [ + 408, + 485, + 421, + 492 + ], + "lines": [ + { + "bbox": [ + 408, + 485, + 421, + 492 + ], + "spans": [ + { + "bbox": [ + 408, + 485, + 421, + 492 + ], + "type": "image", + "image_path": "3e2f53cc61ec55a112cf9006c4fdac613ac1e83011e8281fbca5e9a47ada3ddd.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 422, + 485, + 446, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 422, + 485, + 446, + 493 + ], + "spans": [ + { + "bbox": [ + 422, + 485, + 446, + 493 + ], + "type": "text", + "content": "Accurate" + } + ] + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 451, + 485, + 463, + 492 + ], + "blocks": [ + { + "bbox": [ + 451, + 485, + 463, + 492 + ], + "lines": [ + { + "bbox": [ + 451, + 485, + 463, + 492 + ], + "spans": [ + { + "bbox": [ + 451, + 485, + 463, + 492 + ], + "type": "image", + "image_path": "68c35f09e9d481791bf1b3bf74ab11598c76571421eaf99f060cfb1be6bc50aa.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "bbox": [ + 465, + 485, + 501, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 465, + 485, + 501, + 493 + ], + "spans": [ + { + "bbox": [ + 465, + 485, + 501, + 493 + ], + "type": "text", + "content": "Interpretable" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 359, + 500, + 493, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 500, + 493, + 508 + ], + "spans": [ + { + "bbox": [ + 359, + 500, + 493, + 508 + ], + "type": "text", + "content": "v.s. other automatic image-text align metrics" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 325, + 525, + 360, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 525, + 360, + 533 + ], + "spans": [ + { + "bbox": [ + 325, + 525, + 360, + 533 + ], + "type": "text", + "content": "CLIPScore" + } + ] + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 364, + 512, + 402, + 545 + ], + "blocks": [ + { + "bbox": [ + 364, + 512, + 402, + 545 + ], + "lines": [ + { + "bbox": [ + 364, + 512, + 402, + 545 + ], + "spans": [ + { + "bbox": [ + 364, + 512, + 402, + 545 + ], + "type": "image", + "image_path": "64ecd3c515b2c89600f1c498d79e8e943ff752319db31f78b2b77da92d1e0cfb.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "bbox": [ + 430, + 524, + 453, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 524, + 453, + 532 + ], + "spans": [ + { + "bbox": [ + 430, + 524, + 453, + 532 + ], + "type": "text", + "content": "SPICE" + } + ] + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 462, + 517, + 477, + 533 + ], + "blocks": [ + { + "bbox": [ + 462, + 517, + 477, + 533 + ], + "lines": [ + { + "bbox": [ + 462, + 517, + 477, + 533 + ], + "spans": [ + { + "bbox": [ + 462, + 517, + 477, + 533 + ], + "type": "image", + "image_path": "c6d78a922c8fab6c04bbc48f56edfc903042ac82931bf4bb8d972f5dde7096d5.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "bbox": [ + 479, + 519, + 490, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 479, + 519, + 490, + 527 + ], + "spans": [ + { + "bbox": [ + 479, + 519, + 490, + 527 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 462, + 534, + 517, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 462, + 534, + 517, + 540 + ], + "spans": [ + { + "bbox": [ + 462, + 534, + 517, + 540 + ], + "type": "text", + "content": "sim. (caption, text input)" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 325, + 548, + 401, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 548, + 401, + 556 + ], + "spans": [ + { + "bbox": [ + 325, + 548, + 401, + 556 + ], + "type": "text", + "content": "CLIP:24.3 SPICE:22.2" + } + ] + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 420, + 548, + 426, + 555 + ], + "blocks": [ + { + "bbox": [ + 420, + 548, + 426, + 555 + ], + "lines": [ + { + "bbox": [ + 420, + 548, + 426, + 555 + ], + "spans": [ + { + "bbox": [ + 420, + 548, + 426, + 555 + ], + "type": "image", + "image_path": "a4976d5848b68e554a64a723fd69f112fc97a4dfe505b52d15689e7e7604e2dc.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "bbox": [ + 443, + 548, + 519, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 443, + 548, + 519, + 556 + ], + "spans": [ + { + "bbox": [ + 443, + 548, + 519, + 556 + ], + "type": "text", + "content": "CLIP: 21.3 SPICE: 11.8" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 562, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 547, + 662 + ], + "type": "text", + "content": "Figure 1. Illustration of how TIFA works, and comparison with the widely-used CLIPScore and SPICE metrics. Given the text input, TIFA uses GPT-3 to generate several question-answer pairs, and a QA model filters them (3 out of 14 questions for this text input are shown). TIFA measures whether VQA models can accurately answer these questions given the generated image. In this example, TIFA indicates that the image generated by Stable Diffusion v2.1 is better than that by v1.5, while CLIP and SPICE yield the opposite result. The text input is from the MSCOCO validation set." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "models often fail to compose multiple objects [40, 11, 35], bind attributes to the wrong objects [11], and struggle in generating visual text [36]. Today, there are efforts to address" + } + ] + } + ], + "index": 54 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 287, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 715 + ], + "type": "text", + "content": ". All data and a pip-installable evaluation package are available on the project page." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20406" + } + ] + } + ], + "index": 55 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 343, + 195 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 343, + 195 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 343, + 195 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 343, + 195 + ], + "type": "image", + "image_path": "3cc23a705a88ae78a1e31f868561e4e485461978a9e8587b0d15aab95d63115c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 547, + 276 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 276 + ], + "type": "text", + "content": "Figure 2. (a) Overview of how TIFA evaluates the faithfulness of a synthesized image. TIFA uses a language model (LM), a question-answering (QA) model, and a visual-question-answering (VQA) model. Given a text input, we generate several question-answer pairs with the LM and then filter them via the QA model. To evaluate the faithfulness of a synthesized image to the text input, a VQA model answers these visual questions using the image, and we check the answers for correctness. (b) TIFA v1.0 benchmark. While TIFA is applicable to any text prompt, to allow direct comparison across different studies, and for ease of use, we introduce the TIFA v1.0 benchmark, a repository of text inputs along with pre-generated question-answer tuples with answer choices. To evaluate a text-to-image model, a user first produces the images for the text inputs in TIFA v1.0 and then performs VQA with our provided tools on generated images to compute TIFA." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 347, + 71, + 532, + 196 + ], + "blocks": [ + { + "bbox": [ + 347, + 71, + 532, + 196 + ], + "lines": [ + { + "bbox": [ + 347, + 71, + 532, + 196 + ], + "spans": [ + { + "bbox": [ + 347, + 71, + 532, + 196 + ], + "type": "image", + "image_path": "002e14dbc71ea718f6bd90b0b7376dd20db7e1bea3d060e78ff49d5bea7299a4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 284, + 288, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 288, + 368 + ], + "type": "text", + "content": "these challenges: researchers are imposing linguistic structure with diffusion guidance to produce images with multiple objects [11]; others are designing reward models trained using human feedback to better align generations with user intention [30]. However, progress is difficult to quantify without accurate and interpretable evaluation measures that explain when and how models struggle." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 373, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 288, + 684 + ], + "type": "text", + "content": "A critical bottleneck, therefore, is the lack of reliable automatic evaluation metrics for text-to-image generation faithfulness. One of the popular metrics is CLIPScore [17], which measures the cosine similarity between the CLIP embeddings [42] of the text input and the generated image. However, since CLIP is not effective at counting objects [42], or reasoning compositionally [37], CLIPScore is unreliable and often inaccurate. Another family of evaluation metrics uses image captions, in which an image captioning model first converts the image into text, and then the image caption is evaluated by comparing it against the text input. Unfortunately, using captioning models is insufficient since they might decide to ignore salient information in images or focus on other non-essential image regions [24]; for example, a captioning model might say that the images in Figure 1 are \"a field of grass with trees in the background\". Moreover, evaluating text (caption) generation is inherently challenging [23, 26]. Another recent text-to-image evaluation is DALL-Eval [6], which employs object detection to determine if the objects in the texts are in the generated images. However, this approach only works on synthesized text and measures faithfulness along the limited axes of objects, counting, colors, and spatial relationships but misses activities, geolocation, weather, time, materials, shapes, sizes, and other potential categories we often ask about when we recall images from memory [29]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "To address the above challenges, we introduce TIFA, a new metric to evaluate text-to-image generation faithfulness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 284, + 547, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 547, + 535 + ], + "type": "text", + "content": "Our approach is illustrated in Figure 2. Given a repository of text inputs, we automatically generate question-answer pairs for each text via a language model (here, GPT-3 [3]). A question-answering (QA) system (here, UnifiedQA [25]) is subsequently used to verify and filter these question-answer pairs. To evaluate a generated image, we use a visual-question-answering (VQA) system (here, mPLUGlarge [31], BLIP-2 [32], etc.) to answer the questions given the generated image. We measure the image's faithfulness to the text input as the accuracy of the answers generated by the VQA system. While the accuracy of TIFA is dependent on the accuracy of the VQA model, our experiments show that TIFA has much higher correlation with human judgments than CLIPScore (Spearman's " + }, + { + "bbox": [ + 304, + 284, + 547, + 535 + ], + "type": "inline_equation", + "content": "\\rho = 0.60" + }, + { + "bbox": [ + 304, + 284, + 547, + 535 + ], + "type": "text", + "content": " vs. 0.33) and captioning-based approaches (Spearman's " + }, + { + "bbox": [ + 304, + 284, + 547, + 535 + ], + "type": "inline_equation", + "content": "\\rho = 0.60" + }, + { + "bbox": [ + 304, + 284, + 547, + 535 + ], + "type": "text", + "content": " vs. 0.34). Additionally, since the LMs and VQA models will continue to improve, we hypothesize that TIFA will continue to be more reliable over time. Also, our metrics can automatically detect when elements are missing in the generation: in Figure 2, TIFA detects that the generated image does not contain a TV." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 540, + 547, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 547, + 694 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 547, + 694 + ], + "type": "text", + "content": "To promote the use of our new evaluation metric, we release TIFA v1.0, a large-scale text-to-image generation benchmark containing 4K diverse text inputs, sampled from the MSCOCO captions [34], DrawBench [47], PartiPrompts [61], and PaintSkill [6]. Each input comes with a pre-generated set of question-answer pairs, resulting in 25K questions covering 4.5K distinct elements. These questions have been automatically generated and pre-filtered using a question-answering model. This benchmark also comes with different VQA models [57, 28, 58, 32, 31, 21] that can be used to evaluate generative models and can be easily extended to use future VQA models when they become available." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 701, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 547, + 713 + ], + "type": "text", + "content": "We conduct a comprehensive evaluation of current text-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20407" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "to-image models using TIFA v1.0. Thanks to TIFA's ability to detect fine-grained unfaithfulness in images, we find that current state-of-the-art models are good at rendering common objects, animals, and colors, but still struggle in composing multiple objects, reasoning about spatial relations, and binding the correct activity for each entity. In addition, our ablation experiments show that TIFA is robust to different VQA models. Future researchers can use TIFA v1.0 to compare their text-to-image models' faithfulness across different studies. Also, future generative models may focus on addressing the weaknesses of current models that TIFA discovered. In addition, with TIFA, users can customize evaluations with their own text inputs and questions [10]; for example, a future TIFA benchmark could focus on counting or scene text." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 262, + 134, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 134, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 134, + 275 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 282, + 288, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 288, + 306 + ], + "type": "text", + "content": "We compare TIFA to other image and language generation evaluation metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 307, + 289, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 289, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 289, + 653 + ], + "type": "text", + "content": "Prior image generation evaluation Prior work usually compares image generation models via pairwise comparison by humans. How to design automatic evaluation metrics to approximate human assessment of the quality of machine-generated images has always been a major challenge in computer vision. There are two aspects to evaluate, namely image quality and image-text alignment. Inception Score [48] and FID [18] are the most widely adopted metrics for image quality. They compare the features of the generated images and gold images extracted from a pre-trained Inception-V3 model [52] to evaluate the fidelity and diversity of generated images. However, they rely on ground-truth images and are based on a classification model, which makes them not suitable for complex datasets [12]. For image-text alignment, prior metrics are mainly based on CLIP [42], captioning, and object detection models. CLIPScore [17] and CLIP-R [39] are based on the cosine similarity of image and text CLIP [42] embeddings. [6, 19, 20] first convert the images using a captioning model, and then compare the image caption with the text using metrics like CIDEr [55] and SPICE [1]. SOA [19] and DALL-Eval [6] employ object detection models to determine if objects, attributes, and relations in the text input are in the generated image. However, this approach only works on synthesized text inputs and measures faithfulness on limited axes (object, counting, color, spatial relation), missing elements like material, shape, activities, and context. In contrast, thanks to the flexibility of questions, TIFA works on any text inputs and evaluates faithfulness across a broad spectrum of dimensions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": "Summarization evaluation in NLP TIFA is inspired by the summarization evaluation methods based on question answering (QA) [56, 50]. Given a summary, a language model generates a set of questions about the text. A QA model checks if the same answer can be inferred from the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "text and the summary. These QA-based metrics have much higher correlations with human judgments on the factual consistency of summarization than other automatic metrics [56, 50]. TIFA can be seen as treating the text input for the text-to-image model as a summary of the generated image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 143, + 410, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 143, + 410, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 143, + 410, + 156 + ], + "type": "text", + "content": "3. The TIFA Metric" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": "We introduce a framework for automatically estimating the faithfulness of an image to its text prompt. Given a text input " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": ", we aim to measure the faithfulness of the generated image " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": ". An overview of our metric is illustrated in Figure 2. From " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": ", we generate " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": " multiple-choice question-answer tuples " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "\\{Q_i, C_i, A_i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "Q_i" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": " is a question, " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": " is a set of answer choices, and " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "A_i \\in C_i" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": " is the gold answer. The answer " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "A_i" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": " can be inferred given " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "T, Q_i" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": ". Next, for each question " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "Q_i" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": ", we use a VQA model to produce an answer " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "A_i^{\\mathrm{VQA}} = \\max_{a \\in C_i} p(a \\mid I, Q_i)" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": ". We define the faithfulness between the text " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": " and image " + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 163, + 547, + 295 + ], + "type": "text", + "content": " as the VQA accuracy:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 335, + 303, + 545, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 303, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 335, + 303, + 545, + 335 + ], + "type": "interline_equation", + "content": "f a i t h f u l n e s s (T, I) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbb {1} \\left[ A _ {i} ^ {\\mathrm {V Q A}} = A _ {i} \\right] \\tag {1}", + "image_path": "4e7ae0070b71bd2926a8dc8f548de83021656a00de6d28a40e060f406cde6421.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "text", + "content": "The range of our faithfulness score is " + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "inline_equation", + "content": "[0, 1]" + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "text", + "content": ". It is maximized when we have a performant VQA model, and the image " + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "text", + "content": " accurately covers the information in the text " + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "text", + "content": " so that for any question " + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "text", + "content": ", which can be answered given " + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "text", + "content": " can also be answered given " + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "text", + "content": ". Several key design decisions will be addressed in later sections: how to generate questions (" + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "inline_equation", + "content": "\\S 3.1" + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "text", + "content": "), how to control the question quality (" + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "inline_equation", + "content": "\\S 3.2" + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "text", + "content": "), and how to answer those questions (" + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "inline_equation", + "content": "\\S 3.3" + }, + { + "bbox": [ + 304, + 337, + 547, + 445 + ], + "type": "text", + "content": "). Finally, we give a step-by-step qualitative example of TIFA in Figure 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 453, + 466, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 466, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 466, + 464 + ], + "type": "text", + "content": "3.1. Question-Answer Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 471, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 541 + ], + "type": "text", + "content": "Our main challenge is to generate diverse questions that cover all elements of the text input evenly. We also simplify the question-generation pipeline into a single GPT-3 [3] completion, so that TIFA can exploit the power of recent language models (LM) and work with updated black-box LMs (e.g., ChatGPT) in the future." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 543, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 547, + 662 + ], + "type": "text", + "content": "Inspired by prior work [4], given a text prompt " + }, + { + "bbox": [ + 304, + 543, + 547, + 662 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 543, + 547, + 662 + ], + "type": "text", + "content": ", we generate the question-answer tuples " + }, + { + "bbox": [ + 304, + 543, + 547, + 662 + ], + "type": "inline_equation", + "content": "\\{Q_i, C_i, A_i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 543, + 547, + 662 + ], + "type": "text", + "content": " via the pipeline illustrated in Figure 3. Different from prior work, which relies on multiple components, our pipeline is completed by a single inference run via in-context learning with GPT-3 [3, 59, 22, 41, 51], thereby avoiding the need for intermediate human annotations. We annotate 15 examples and use them as in-context examples for GPT-3 to follow. Here we take the text \"A photo of three dogs\" as an example. Each in-context example contains the following steps:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Element extraction Given text prompt " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": ", GPT-3 will first extract all elements " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\{v_i\\}_{i=1}^m" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " following prior work [4] (for the in-context examples, we perform element extraction manually). The elements include noun phrases (including named" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20408" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 272, + 339 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 272, + 339 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 272, + 339 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 272, + 339 + ], + "type": "image", + "image_path": "6a1664ffedb426de384e4b27c58ce42eaa33029af954edf04727c5235e2a36cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 342, + 288, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 288, + 399 + ], + "type": "text", + "content": "Figure 3. Our question-answer pair generation pipeline. The whole pipeline can be executed via a single inference of GPT-3 via in-context learning. Given the text prompt, GPT-3 first extracts the elements and then generates two questions for each element. The GPT-3 output is then parsed and filtered by UnifiedQA." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 407, + 288, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 288, + 443 + ], + "type": "text", + "content": "entities), verbs, adjectives, adverbs, and parse tree spans with no more than 3 words altogether. For the above example, the elements are photo, three, dogs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 288, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 558 + ], + "type": "text", + "content": "Element category classification For each element " + }, + { + "bbox": [ + 46, + 449, + 288, + 558 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 449, + 288, + 558 + ], + "type": "text", + "content": ", following [29], we classify the elements into one of the following 12 categories: object, activity, animal, food, counting, color, material, spatial, location, shape, attribute, and other. As shown in Figure 3, text generated from GPT-3 contains the category corresponding to each question. For example, \"three\" is \"counting\", and \"dogs\" is classified as \"animal.\" This step allows a detailed analysis of the text-to-image model's ability in each category." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 563, + 288, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 288, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 288, + 672 + ], + "type": "text", + "content": "Question generation conditioned on elements For each element " + }, + { + "bbox": [ + 46, + 563, + 288, + 672 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 563, + 288, + 672 + ], + "type": "text", + "content": ", we generate two questions. The first is a question that should be answered \"yes\" for a faithful generated image, and the second question has " + }, + { + "bbox": [ + 46, + 563, + 288, + 672 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 563, + 288, + 672 + ], + "type": "text", + "content": " as its answer. For example, two questions are generated for the element \"three\". The first is \"are there three dogs?\", and the choices are {yes, no}. The second is \"how many dogs are there?\", and the choices are {1,2,3,4}. These two types of questions make our evaluations diverse and robust to surface-level differences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Completing above steps by prompting GPT-3 once As mentioned earlier, for each text " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " , the whole pipeline can be completed by one GPT-3 inference. We annotated 15" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "in-context examples that cover all types of questions. The full prompt is in the Appendix. The prompt format is shown in Figure 3. Our in-context examples follow the same format, and identical examples are used for all text inputs, leading to a fixed and limited amount of human annotation cost. We use code-davinci-002 engine for question generation, and the decoding temperature is 0." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 163, + 415, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 163, + 415, + 177 + ], + "spans": [ + { + "bbox": [ + 306, + 163, + 415, + 177 + ], + "type": "text", + "content": "3.2. Question Filtering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "content": "To ensure the quality of generated images, we use UnifiedQA [25] to verify the GPT-3 generated question-answer pairs and filter out the ones that GPT-3 and UnifiedQA do not agree on. UnifiedQA" + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "content": " is a state-of-the-art multi-task question-answering model that can answer both multiple-choice and free-form questions. Denote the UnifiedQA model as " + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "inline_equation", + "content": "QA" + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "content": ". Given the text " + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "content": ", question " + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "content": ", choices " + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "inline_equation", + "content": "C_{i}" + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "content": ", and answer " + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "content": ", Let " + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "inline_equation", + "content": "A_{i}^{f} = QA(T,Q_{i})" + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "content": " be the free-form answer, and " + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "inline_equation", + "content": "A_{i}^{mc} = QA(T,Q_{i},C_{i})" + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "content": " be the multiple-choice answer. We keep the question if " + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "inline_equation", + "content": "A_{i} = A_{i}^{mc}" + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "content": " and the word-level " + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "content": " score between " + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "inline_equation", + "content": "A_{i}^{f}" + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 304, + 182, + 547, + 352 + ], + "type": "text", + "content": " is greater than 0.7. We conduct a human evaluation on 1000 filtered question-answer pairs. Only 7 are considered not reasonable (e.g., generated choices do not include a correct answer). Details are in Appendix C." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 355, + 547, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 547, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 547, + 428 + ], + "type": "text", + "content": "Recommended VQA model Based on considerations over the accuracy, correlation with human judgments, and run time, we currently suggest using mPLUG-large as the VQA model for TIFA. Analysis is given in Section 5.4. Like the LM and QA components, the VQA component can be updated in the future as the technology improves." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 435, + 390, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 390, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 390, + 449 + ], + "type": "text", + "content": "3.3. VQA Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 453, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 547, + 538 + ], + "type": "text", + "content": "Since our questions contain a diverse set of visual elements (e.g., activity, art style), we use open-domain pre-trained vision-language models as our VQA model (rather than closed-class classification models fine-tuned on VQAv2 [14]). We provide tools to easily perform VQA on arbitrary images and questions, based on 5 state-of-the-art VQA models trained with distinct data and strategies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 542, + 547, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 547, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 547, + 686 + ], + "type": "text", + "content": "Vision-language models The general pre-trained vision-language model are GIT-large [57], VILT-B/32 [28], OFA-large [58], and mPLUG-large [31]. These models are pretrained on a large amount of image-text pairs, and downstream image-to-text tasks like image captioning and visual question answering. Notice that these models have not been trained to answer multiple-choice questions. For each question, we first decode the free-form answer and then choose the choice that has the highest similarity with the decoded answer, measured by SBERT [45]. Another model we use is BLIP-2 FlanT5-XL [32], in which a VIT [8] is connected with a frozen FlanT5 [7] via a lightweight transformer. This" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "text", + "content": "2Model checkpoint we use: https://huggingface.co/allenai/unifiedqa-v2-t5-large-1363200." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20409" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 115, + 75, + 292, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 75, + 292, + 84 + ], + "spans": [ + { + "bbox": [ + 115, + 75, + 292, + 84 + ], + "type": "text", + "content": "Text Input: A wet dog standing on a beach next to the ocean." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 91, + 272, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 91, + 272, + 102 + ], + "spans": [ + { + "bbox": [ + 113, + 91, + 272, + 102 + ], + "type": "text", + "content": "Step 1: Question Generation + Filtering" + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 114, + 104, + 481, + 215 + ], + "blocks": [ + { + "bbox": [ + 114, + 104, + 481, + 215 + ], + "lines": [ + { + "bbox": [ + 114, + 104, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 114, + 104, + 481, + 215 + ], + "type": "table", + "html": "
About dog (animal/human)About ocean (location)About wet (attribute)
Q: Is this a dog?Q: Is there an ocean?Q: Is the dog wet?
A: YesA: YesA: Yes
Q: What animal is in the picture?Q: What type of place is this?Q: Is the dog wet or dry?
A: DogA: OceanA: Wet
About beach (location)About standing (activity)About next to (spatial relation)
Q: Is this a beach?Q: Is the dog standing?Q: Is the dog next to the ocean?
A: YesA: YesA: Yes
Q: What type of place is this?Q: What is the dog doing?Q: Is the dog next to or in the ocean?
A: BeachA: StandingA: Next to
", + "image_path": "ef4c9091aa61a427fc97e2b449681e134f288f7ee835c8dbc9c239353f60aabb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 221, + 253, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 221, + 253, + 232 + ], + "spans": [ + { + "bbox": [ + 113, + 221, + 253, + 232 + ], + "type": "text", + "content": "Step 2: Visual Question Answering" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 118, + 240, + 207, + 329 + ], + "blocks": [ + { + "bbox": [ + 118, + 240, + 207, + 329 + ], + "lines": [ + { + "bbox": [ + 118, + 240, + 207, + 329 + ], + "spans": [ + { + "bbox": [ + 118, + 240, + 207, + 329 + ], + "type": "image", + "image_path": "d7f364f9ab8707cafc9c8b69df7c0068c1d8080b00eeda6d0f7fb9e5bbbfb393.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 339, + 205, + 348 + ], + "lines": [ + { + "bbox": [ + 119, + 339, + 205, + 348 + ], + "spans": [ + { + "bbox": [ + 119, + 339, + 205, + 348 + ], + "type": "text", + "content": "Stable Diffusion v1.5 image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 120, + 356, + 138, + 377 + ], + "blocks": [ + { + "bbox": [ + 120, + 356, + 138, + 377 + ], + "lines": [ + { + "bbox": [ + 120, + 356, + 138, + 377 + ], + "spans": [ + { + "bbox": [ + 120, + 356, + 138, + 377 + ], + "type": "image", + "image_path": "0c1b2e85e33bca6cb3dac0200c8c07b2849ac37bc9f82b0bed02e94745197458.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 392, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 392, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 545, + 415 + ], + "type": "text", + "content": "Figure 4. Step-by-step qualitative example of TIFA metric. Given a text input, we first generate question-answer pairs and filter them. We strike through the questions filtered out by UnifiedQA. Then we run VQA models on the generated image to get the TIFA score." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 365, + 197, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 365, + 197, + 373 + ], + "spans": [ + { + "bbox": [ + 138, + 365, + 197, + 373 + ], + "type": "text", + "content": "TIFA " + }, + { + "bbox": [ + 138, + 365, + 197, + 373 + ], + "type": "inline_equation", + "content": "7 / 10 = 0.7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 236, + 307, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 236, + 307, + 245 + ], + "spans": [ + { + "bbox": [ + 228, + 236, + 307, + 245 + ], + "type": "text", + "content": "About dog (animal/human)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 246, + 277, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 246, + 277, + 255 + ], + "spans": [ + { + "bbox": [ + 228, + 246, + 277, + 255 + ], + "type": "text", + "content": "Q: Is this a dog?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 228, + 255, + 249, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 255, + 249, + 264 + ], + "spans": [ + { + "bbox": [ + 228, + 255, + 249, + 264 + ], + "type": "text", + "content": "A:Yes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 228, + 266, + 325, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 266, + 325, + 274 + ], + "spans": [ + { + "bbox": [ + 228, + 266, + 325, + 274 + ], + "type": "text", + "content": "Q: What animal is in the picture?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 228, + 275, + 251, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 275, + 251, + 284 + ], + "spans": [ + { + "bbox": [ + 228, + 275, + 251, + 284 + ], + "type": "text", + "content": "A: Dog" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 251, + 317, + 263 + ], + "blocks": [ + { + "bbox": [ + 306, + 251, + 317, + 263 + ], + "lines": [ + { + "bbox": [ + 306, + 251, + 317, + 263 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 317, + 263 + ], + "type": "image", + "image_path": "5c7ee897bef39786145b2eb99461704d55798546ae145b39d7efdfbb077c88d0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 228, + 295, + 295, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 295, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 228, + 295, + 295, + 304 + ], + "type": "text", + "content": "About beach (location)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 228, + 304, + 283, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 304, + 283, + 312 + ], + "spans": [ + { + "bbox": [ + 228, + 304, + 283, + 312 + ], + "type": "text", + "content": "Q: Is this a beach?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 228, + 314, + 249, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 314, + 249, + 322 + ], + "spans": [ + { + "bbox": [ + 228, + 314, + 249, + 322 + ], + "type": "text", + "content": "A:Yes" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 228, + 323, + 315, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 323, + 315, + 333 + ], + "spans": [ + { + "bbox": [ + 228, + 323, + 315, + 333 + ], + "type": "text", + "content": "Q: What type of place is this?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 228, + 334, + 257, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 334, + 257, + 342 + ], + "spans": [ + { + "bbox": [ + 228, + 334, + 257, + 342 + ], + "type": "text", + "content": "A:Beach" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 306, + 312, + 317, + 323 + ], + "blocks": [ + { + "bbox": [ + 306, + 312, + 317, + 323 + ], + "lines": [ + { + "bbox": [ + 306, + 312, + 317, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 317, + 323 + ], + "type": "image", + "image_path": "a1527cc74f269bdf8891c2550978fbd1444cc40fff7bc6fbb0e2f3f664a8425e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 228, + 354, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 354, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 228, + 354, + 295, + 361 + ], + "type": "text", + "content": "About ocean (location)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 228, + 362, + 290, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 362, + 290, + 371 + ], + "spans": [ + { + "bbox": [ + 228, + 362, + 290, + 371 + ], + "type": "text", + "content": "Q: Is there an ocean?" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 228, + 372, + 249, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 372, + 249, + 380 + ], + "spans": [ + { + "bbox": [ + 228, + 372, + 249, + 380 + ], + "type": "text", + "content": "A:Yes" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 306, + 372, + 317, + 384 + ], + "blocks": [ + { + "bbox": [ + 306, + 372, + 317, + 384 + ], + "lines": [ + { + "bbox": [ + 306, + 372, + 317, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 317, + 384 + ], + "type": "image", + "image_path": "474e633457676a3587a2eff43a3d20eff1b2be1e0187a536884295bf2f1def45.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 360, + 236, + 434, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 236, + 434, + 245 + ], + "spans": [ + { + "bbox": [ + 360, + 236, + 434, + 245 + ], + "type": "text", + "content": "About standing (activity)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 360, + 246, + 429, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 246, + 429, + 255 + ], + "spans": [ + { + "bbox": [ + 360, + 246, + 429, + 255 + ], + "type": "text", + "content": "Q: Is the dog standing?" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 360, + 255, + 380, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 255, + 380, + 263 + ], + "spans": [ + { + "bbox": [ + 360, + 255, + 380, + 263 + ], + "type": "text", + "content": "A: No" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 360, + 274, + 422, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 274, + 422, + 283 + ], + "spans": [ + { + "bbox": [ + 360, + 274, + 422, + 283 + ], + "type": "text", + "content": "About wet (attribute)" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 360, + 285, + 415, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 285, + 415, + 293 + ], + "spans": [ + { + "bbox": [ + 360, + 285, + 415, + 293 + ], + "type": "text", + "content": "Q: Is the dog wet?" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 360, + 294, + 380, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 294, + 380, + 302 + ], + "spans": [ + { + "bbox": [ + 360, + 294, + 380, + 302 + ], + "type": "text", + "content": "A: No" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 360, + 304, + 434, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 304, + 434, + 312 + ], + "spans": [ + { + "bbox": [ + 360, + 304, + 434, + 312 + ], + "type": "text", + "content": "Q: Is the dog wet or dry?" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 360, + 314, + 383, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 314, + 383, + 323 + ], + "spans": [ + { + "bbox": [ + 360, + 314, + 383, + 323 + ], + "type": "text", + "content": "A: Dry" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 360, + 333, + 449, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 333, + 449, + 342 + ], + "spans": [ + { + "bbox": [ + 360, + 333, + 449, + 342 + ], + "type": "text", + "content": "About next to (spatial relation)" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 360, + 342, + 453, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 342, + 453, + 351 + ], + "spans": [ + { + "bbox": [ + 360, + 342, + 453, + 351 + ], + "type": "text", + "content": "Q: Is the dog next to the ocean?" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 360, + 352, + 381, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 352, + 381, + 361 + ], + "spans": [ + { + "bbox": [ + 360, + 352, + 381, + 361 + ], + "type": "text", + "content": "A:Yes" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 360, + 361, + 468, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 361, + 468, + 370 + ], + "spans": [ + { + "bbox": [ + 360, + 361, + 468, + 370 + ], + "type": "text", + "content": "Q: Is the dog next to or in the ocean?" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 360, + 372, + 392, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 372, + 392, + 380 + ], + "spans": [ + { + "bbox": [ + 360, + 372, + 392, + 380 + ], + "type": "text", + "content": "A: Next to" + } + ] + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 462, + 255, + 473, + 266 + ], + "blocks": [ + { + "bbox": [ + 462, + 255, + 473, + 266 + ], + "lines": [ + { + "bbox": [ + 462, + 255, + 473, + 266 + ], + "spans": [ + { + "bbox": [ + 462, + 255, + 473, + 266 + ], + "type": "image", + "image_path": "154a1a6c0ae1168f586f12765bc1e45361922d68000e5b0eb7f41d5c0f5d602b.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 462, + 292, + 473, + 304 + ], + "blocks": [ + { + "bbox": [ + 462, + 292, + 473, + 304 + ], + "lines": [ + { + "bbox": [ + 462, + 292, + 473, + 304 + ], + "spans": [ + { + "bbox": [ + 462, + 292, + 473, + 304 + ], + "type": "image", + "image_path": "8f127a7f3ce2a7c1055b518c078a2a0abb59062804dd758d71a96d0aaeff9ebb.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 462, + 312, + 473, + 323 + ], + "blocks": [ + { + "bbox": [ + 462, + 312, + 473, + 323 + ], + "lines": [ + { + "bbox": [ + 462, + 312, + 473, + 323 + ], + "spans": [ + { + "bbox": [ + 462, + 312, + 473, + 323 + ], + "type": "image", + "image_path": "4ed20ead1959143d1542d950ff7df12fcc1713ed4b9df1e67f62cfa02cbe1fed.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 462, + 350, + 473, + 361 + ], + "blocks": [ + { + "bbox": [ + 462, + 350, + 473, + 361 + ], + "lines": [ + { + "bbox": [ + 462, + 350, + 473, + 361 + ], + "spans": [ + { + "bbox": [ + 462, + 350, + 473, + 361 + ], + "type": "image", + "image_path": "3af46fd7916c03519a8ac648491bd202dda28b23ee4f783ed31e744a6ab95dc0.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 462, + 361, + 468, + 370 + ], + "blocks": [ + { + "bbox": [ + 462, + 361, + 468, + 370 + ], + "lines": [ + { + "bbox": [ + 462, + 361, + 468, + 370 + ], + "spans": [ + { + "bbox": [ + 462, + 361, + 468, + 370 + ], + "type": "image", + "image_path": "da37f76ecdc14b69667ea3479244e558c11efe12da0ddbf6bf7dbb3cc8a55bd4.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 462, + 372, + 473, + 383 + ], + "blocks": [ + { + "bbox": [ + 462, + 372, + 473, + 383 + ], + "lines": [ + { + "bbox": [ + 462, + 372, + 473, + 383 + ], + "spans": [ + { + "bbox": [ + 462, + 372, + 473, + 383 + ], + "type": "image", + "image_path": "2d49daa3451cc2ec8cda8af22572489f4e3be5e0679e6d69f362b8157a3f9d2d.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "bbox": [ + 46, + 423, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 448 + ], + "type": "text", + "content": "model allows for performing multiple-choice VQA directly due to the flexibility of the LM." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 47, + 460, + 287, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 287, + 486 + ], + "type": "text", + "content": "4. FIFA v1.0: Benchmark for Text-to-Image Generation Faithfulness" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 46, + 494, + 288, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 288, + 625 + ], + "type": "text", + "content": "In this section, we introduce TIFA v1.0, a text-to-image generation faithfulness benchmark based on the evaluation method discussed in Section 3. The benchmark consists of 4,081 diverse text inputs paired with 25,829 question-answer pairs. Each question is classified into one of the categories discussed in Section 3.1. The benchmark also comes with Python pip-installable APIs to perform VQA with various state-of-the-art VQA models on arbitrary visual questions. The overall TIFA for each text-to-image model is computed by averaging TIFA scores of images generated from each text input in the benchmark." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 47, + 634, + 145, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 145, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 145, + 646 + ], + "type": "text", + "content": "4.1. Text Collections" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": "We collect 4,081 text inputs to benchmark text-to-image models' generation ability on diverse tasks. 2,000 text inputs are image captions from COCO validation set [34]. These captions have corresponding gold images. Since text-to-image models are often used to create abstract art, we also" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 423, + 545, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 507 + ], + "type": "text", + "content": "collect 2,081 text inputs from previous works that do not correspond to any real image. All text inputs we use contain " + }, + { + "bbox": [ + 304, + 423, + 545, + 507 + ], + "type": "inline_equation", + "content": "\\geq 3" + }, + { + "bbox": [ + 304, + 423, + 545, + 507 + ], + "type": "text", + "content": " words. We include 161 from DrawBench used in Imagen [47] (texts that are categorized as \"misspellings\" and \"rare words\" are removed); 1420 from PartiPrompt used in Parti [61] (texts in category \"abstract\" are removed); and 500 texts from PaintSkill used in DALL-Eval [6]." + } + ] + } + ], + "index": 50 + }, + { + "type": "table", + "bbox": [ + 306, + 526, + 545, + 657 + ], + "blocks": [ + { + "bbox": [ + 364, + 516, + 485, + 526 + ], + "lines": [ + { + "bbox": [ + 364, + 516, + 485, + 526 + ], + "spans": [ + { + "bbox": [ + 364, + 516, + 485, + 526 + ], + "type": "text", + "content": "Table 1. Statistics of FIFA v1.0." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 526, + 545, + 657 + ], + "lines": [ + { + "bbox": [ + 306, + 526, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 545, + 657 + ], + "type": "table", + "html": "
Statistics
# of prompts4,081
- # of COCO captions2,000
- # of DrawBench, PartiPrompt, PaintSkill prompts2,081
# of questions25,829
- # of binary questions17,226
- # of multiple-choice questions8,603
avg. # of questions per prompt6.3
avg. # of words per prompt10.5
avg. # of elements per prompt4.3
", + "image_path": "cea8c2de7c988e0cee88fa94af21426a2267c1406064d78be40fcc353c752244.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "table_body" + } + ], + "index": 52 + }, + { + "bbox": [ + 306, + 670, + 436, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 436, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 436, + 683 + ], + "type": "text", + "content": "4.2. Statistics and Diversity" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Table 1 shows the basic statistics of the TIFA v1.0 benchmark. We demonstrate TIFA v1.0's diversity in Figure 5." + } + ] + } + ], + "index": 54 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20410" + } + ] + } + ], + "index": 55 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 76, + 112, + 91 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 76, + 112, + 91 + ], + "spans": [ + { + "bbox": [ + 56, + 76, + 112, + 91 + ], + "type": "text", + "content": "a yellow t-shirt with a dog on it\"" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 113, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 113, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 113, + 121 + ], + "type": "text", + "content": "\"a panda bear with aviator glasses on its head\"" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 129, + 114, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 129, + 114, + 136 + ], + "spans": [ + { + "bbox": [ + 54, + 129, + 114, + 136 + ], + "type": "text", + "content": "\"A red colored car\"" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 148, + 110, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 148, + 110, + 167 + ], + "spans": [ + { + "bbox": [ + 59, + 148, + 110, + 167 + ], + "type": "text", + "content": "\"One cat and one dog sitting on the grass\"" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 126, + 76, + 143, + 92 + ], + "blocks": [ + { + "bbox": [ + 126, + 76, + 143, + 92 + ], + "lines": [ + { + "bbox": [ + 126, + 76, + 143, + 92 + ], + "spans": [ + { + "bbox": [ + 126, + 76, + 143, + 92 + ], + "type": "image", + "image_path": "6c61a26c3e469c31513795353603f160bebc79eae50b011355684b2a26063ae0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 145, + 78, + 174, + 87 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 78, + 174, + 87 + ], + "spans": [ + { + "bbox": [ + 145, + 78, + 174, + 87 + ], + "type": "text", + "content": "Object" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 145, + 87, + 200, + 92 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 87, + 200, + 92 + ], + "spans": [ + { + "bbox": [ + 145, + 87, + 200, + 92 + ], + "type": "text", + "content": "table,train,bus ..." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 126, + 102, + 141, + 118 + ], + "blocks": [ + { + "bbox": [ + 126, + 102, + 141, + 118 + ], + "lines": [ + { + "bbox": [ + 126, + 102, + 141, + 118 + ], + "spans": [ + { + "bbox": [ + 126, + 102, + 141, + 118 + ], + "type": "image", + "image_path": "84d5b5f257ac863bf4c194bda68a306d7cb6d83d35f3404dc03a3945095fc19a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 145, + 102, + 178, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 102, + 178, + 110 + ], + "spans": [ + { + "bbox": [ + 145, + 102, + 178, + 110 + ], + "type": "text", + "content": "Spatial" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 145, + 110, + 204, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 110, + 204, + 117 + ], + "spans": [ + { + "bbox": [ + 145, + 110, + 204, + 117 + ], + "type": "text", + "content": "above,below,right .." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 126, + 126, + 143, + 141 + ], + "blocks": [ + { + "bbox": [ + 126, + 126, + 143, + 141 + ], + "lines": [ + { + "bbox": [ + 126, + 126, + 143, + 141 + ], + "spans": [ + { + "bbox": [ + 126, + 126, + 143, + 141 + ], + "type": "image", + "image_path": "a87d4ea750f0a2a6aaf32c66dd68948e0719305c054ece6e1914e6bb092c4fd2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 145, + 125, + 184, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 125, + 184, + 134 + ], + "spans": [ + { + "bbox": [ + 145, + 125, + 184, + 134 + ], + "type": "text", + "content": "Counting" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 145, + 135, + 206, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 135, + 206, + 141 + ], + "spans": [ + { + "bbox": [ + 145, + 135, + 206, + 141 + ], + "type": "text", + "content": "two,three,several" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 126, + 150, + 141, + 167 + ], + "blocks": [ + { + "bbox": [ + 126, + 150, + 141, + 167 + ], + "lines": [ + { + "bbox": [ + 126, + 150, + 141, + 167 + ], + "spans": [ + { + "bbox": [ + 126, + 150, + 141, + 167 + ], + "type": "image", + "image_path": "ecaa51dfa75b45fd02357fc3ec25e9187a5de3f1e14f26be11c52d1743718cc9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 176, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 545, + 198 + ], + "type": "text", + "content": "Figure 5. Statistics and diversity of TIFA v1.0. The text inputs contain elements from 12 categories (e.g., object, spatial, and counting). We show the most common elements from each category. In addition, we also show some example text inputs on the sides." + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 145, + 150, + 165, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 150, + 165, + 158 + ], + "spans": [ + { + "bbox": [ + 145, + 150, + 165, + 158 + ], + "type": "text", + "content": "Food" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 145, + 159, + 192, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 159, + 192, + 165 + ], + "spans": [ + { + "bbox": [ + 145, + 159, + 192, + 165 + ], + "type": "text", + "content": "pizza, bananas ..." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 211, + 78, + 227, + 92 + ], + "blocks": [ + { + "bbox": [ + 211, + 78, + 227, + 92 + ], + "lines": [ + { + "bbox": [ + 211, + 78, + 227, + 92 + ], + "spans": [ + { + "bbox": [ + 211, + 78, + 227, + 92 + ], + "type": "image", + "image_path": "a592dbfe7c077069367c89fbfd0094178622657e535606df89a3252a9b221c31.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 230, + 78, + 288, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 78, + 288, + 86 + ], + "spans": [ + { + "bbox": [ + 230, + 78, + 288, + 86 + ], + "type": "text", + "content": "Animal/Human" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 232, + 87, + 280, + 93 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 87, + 280, + 93 + ], + "spans": [ + { + "bbox": [ + 232, + 87, + 280, + 93 + ], + "type": "text", + "content": "man,woman,dog ..." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 298, + 78, + 312, + 93 + ], + "blocks": [ + { + "bbox": [ + 298, + 78, + 312, + 93 + ], + "lines": [ + { + "bbox": [ + 298, + 78, + 312, + 93 + ], + "spans": [ + { + "bbox": [ + 298, + 78, + 312, + 93 + ], + "type": "image", + "image_path": "c7df1793d05328685c27b9822b2fb87d5c02eb4b10d09e2380b8a9dc5d6980f5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 78, + 359, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 78, + 359, + 86 + ], + "spans": [ + { + "bbox": [ + 317, + 78, + 359, + 86 + ], + "type": "text", + "content": "Attribute" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 87, + 378, + 93 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 87, + 378, + 93 + ], + "spans": [ + { + "bbox": [ + 317, + 87, + 378, + 93 + ], + "type": "text", + "content": "large, small, young ..." + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 263, + 99, + 282, + 120 + ], + "blocks": [ + { + "bbox": [ + 263, + 99, + 282, + 120 + ], + "lines": [ + { + "bbox": [ + 263, + 99, + 282, + 120 + ], + "spans": [ + { + "bbox": [ + 263, + 99, + 282, + 120 + ], + "type": "image", + "image_path": "2f23bfa4bb47ef71f24f07f2ef2984f4140817fd4e18026b42d2df91e241681f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 284, + 109, + 337, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 109, + 337, + 119 + ], + "spans": [ + { + "bbox": [ + 284, + 109, + 337, + 119 + ], + "type": "text", + "content": "TIFA v1.0" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 241, + 125, + 357, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 125, + 357, + 141 + ], + "spans": [ + { + "bbox": [ + 241, + 125, + 357, + 141 + ], + "type": "text", + "content": "4,081 text inputs 25,829 questions \n4,550 distinct elements 12 categories" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 211, + 150, + 227, + 166 + ], + "blocks": [ + { + "bbox": [ + 211, + 150, + 227, + 166 + ], + "lines": [ + { + "bbox": [ + 211, + 150, + 227, + 166 + ], + "spans": [ + { + "bbox": [ + 211, + 150, + 227, + 166 + ], + "type": "image", + "image_path": "002405c7513a53fd78d1b9f5eb5e76ff2af1dc6807e318d290b766de13f393c7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 232, + 149, + 271, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 149, + 271, + 158 + ], + "spans": [ + { + "bbox": [ + 232, + 149, + 271, + 158 + ], + "type": "text", + "content": "Material" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 230, + 159, + 287, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 159, + 287, + 166 + ], + "spans": [ + { + "bbox": [ + 230, + 159, + 287, + 166 + ], + "type": "text", + "content": "wooden,metal,glass" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 300, + 149, + 315, + 158 + ], + "blocks": [ + { + "bbox": [ + 300, + 149, + 315, + 158 + ], + "lines": [ + { + "bbox": [ + 300, + 149, + 315, + 158 + ], + "spans": [ + { + "bbox": [ + 300, + 149, + 315, + 158 + ], + "type": "image", + "image_path": "3902b528b991b62460dda4e572fb9ef6e9ec734600dec1732ceb3fdb230250bd.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 319, + 150, + 343, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 150, + 343, + 158 + ], + "spans": [ + { + "bbox": [ + 319, + 150, + 343, + 158 + ], + "type": "text", + "content": "Shape" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 318, + 159, + 368, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 159, + 368, + 166 + ], + "spans": [ + { + "bbox": [ + 318, + 159, + 368, + 166 + ], + "type": "text", + "content": "circular,square" + } + ] + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 387, + 78, + 399, + 95 + ], + "blocks": [ + { + "bbox": [ + 387, + 78, + 399, + 95 + ], + "lines": [ + { + "bbox": [ + 387, + 78, + 399, + 95 + ], + "spans": [ + { + "bbox": [ + 387, + 78, + 399, + 95 + ], + "type": "image", + "image_path": "fccdb2ed2a7a22286cc81991007956b07612bbdb2d0f31087d2cbd69302a72e3.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 403, + 78, + 443, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 78, + 443, + 86 + ], + "spans": [ + { + "bbox": [ + 403, + 78, + 443, + 86 + ], + "type": "text", + "content": "Activity" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 403, + 87, + 463, + 93 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 87, + 463, + 93 + ], + "spans": [ + { + "bbox": [ + 403, + 87, + 463, + 93 + ], + "type": "text", + "content": "sitting.standing ..." + } + ] + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 387, + 102, + 399, + 118 + ], + "blocks": [ + { + "bbox": [ + 387, + 102, + 399, + 118 + ], + "lines": [ + { + "bbox": [ + 387, + 102, + 399, + 118 + ], + "spans": [ + { + "bbox": [ + 387, + 102, + 399, + 118 + ], + "type": "image", + "image_path": "c7b57b159055b870cc30565355e60f8b78442d2cb03e1076794543b0623eeb09.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 403, + 102, + 443, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 102, + 443, + 110 + ], + "spans": [ + { + "bbox": [ + 403, + 102, + 443, + 110 + ], + "type": "text", + "content": "Location" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 402, + 111, + 465, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 111, + 465, + 118 + ], + "spans": [ + { + "bbox": [ + 402, + 111, + 465, + 118 + ], + "type": "text", + "content": "field,street,beach ..." + } + ] + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 387, + 126, + 400, + 142 + ], + "blocks": [ + { + "bbox": [ + 387, + 126, + 400, + 142 + ], + "lines": [ + { + "bbox": [ + 387, + 126, + 400, + 142 + ], + "spans": [ + { + "bbox": [ + 387, + 126, + 400, + 142 + ], + "type": "image", + "image_path": "abe62ee34e8e14d8d4f75fa01eeb7b06c591355111013fb98f786d9077ec3e5a.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 405, + 126, + 428, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 126, + 428, + 134 + ], + "spans": [ + { + "bbox": [ + 405, + 126, + 428, + 134 + ], + "type": "text", + "content": "Color" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 404, + 136, + 455, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 136, + 455, + 142 + ], + "spans": [ + { + "bbox": [ + 404, + 136, + 455, + 142 + ], + "type": "text", + "content": "red,white,blue ..." + } + ] + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 386, + 151, + 400, + 167 + ], + "blocks": [ + { + "bbox": [ + 386, + 151, + 400, + 167 + ], + "lines": [ + { + "bbox": [ + 386, + 151, + 400, + 167 + ], + "spans": [ + { + "bbox": [ + 386, + 151, + 400, + 167 + ], + "type": "image", + "image_path": "afb9daf7febb133252aba567033da63eb2106c6c7ea5461fe92df0678b7cab77.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "bbox": [ + 405, + 150, + 429, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 150, + 429, + 158 + ], + "spans": [ + { + "bbox": [ + 405, + 150, + 429, + 158 + ], + "type": "text", + "content": "Other" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 403, + 159, + 459, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 159, + 459, + 166 + ], + "spans": [ + { + "bbox": [ + 403, + 159, + 459, + 166 + ], + "type": "text", + "content": "punk, starry night" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 471, + 76, + 535, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 471, + 76, + 535, + 83 + ], + "spans": [ + { + "bbox": [ + 471, + 76, + 535, + 83 + ], + "type": "text", + "content": "\"a painting of a fox" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 478, + 83, + 521, + 90 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 478, + 83, + 521, + 90 + ], + "spans": [ + { + "bbox": [ + 478, + 83, + 521, + 90 + ], + "type": "text", + "content": "in the style" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 478, + 90, + 521, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 478, + 90, + 521, + 95 + ], + "spans": [ + { + "bbox": [ + 478, + 90, + 521, + 95 + ], + "type": "text", + "content": "starry night" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 479, + 103, + 534, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 479, + 103, + 534, + 110 + ], + "spans": [ + { + "bbox": [ + 479, + 103, + 534, + 110 + ], + "type": "text", + "content": "\"A man running to" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 479, + 110, + 533, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 479, + 110, + 533, + 117 + ], + "spans": [ + { + "bbox": [ + 479, + 110, + 533, + 117 + ], + "type": "text", + "content": "hit a tennis ball" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 482, + 129, + 529, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 482, + 129, + 529, + 135 + ], + "spans": [ + { + "bbox": [ + 482, + 129, + 529, + 135 + ], + "type": "text", + "content": "\"a flag with a" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 482, + 136, + 529, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 482, + 136, + 529, + 141 + ], + "spans": [ + { + "bbox": [ + 482, + 136, + 529, + 141 + ], + "type": "text", + "content": "dinosaur on it\"" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 476, + 149, + 539, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 476, + 149, + 539, + 156 + ], + "spans": [ + { + "bbox": [ + 476, + 149, + 539, + 156 + ], + "type": "text", + "content": "\"a photo of suitcase" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 478, + 156, + 537, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 478, + 156, + 537, + 163 + ], + "spans": [ + { + "bbox": [ + 478, + 156, + 537, + 163 + ], + "type": "text", + "content": "and chair; chair is" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 478, + 163, + 537, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 478, + 163, + 537, + 168 + ], + "spans": [ + { + "bbox": [ + 478, + 163, + 537, + 168 + ], + "type": "text", + "content": "right to suitcase\"" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 46, + 207, + 289, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 339 + ], + "type": "text", + "content": "TIFA v1.0 contains questions about 4,550 distinct elements, which are categorized into 12 categories. The number of times each type of element occurs in the text input is object (7,854), animal/human (3,501), attribute (3,399), activity (2,851), spatial (2,265), location (1,840), color (1,743), counting (986), food (911), material (209), shape (69), and other (201). The \"other\" category includes notions often used in abstract art, such as \"starry night\" and \"steampunk.\" The accuracy of VQA on a particular genre measures the text-to-image model's ability in the corresponding aspect. Please refer to Appendix D for more details on TIFA v1.0." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 47, + 350, + 287, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 287, + 374 + ], + "type": "text", + "content": "4.3. Finetuned Open-Source Language Model for Question Generation" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 46, + 382, + 288, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 288, + 491 + ], + "type": "text", + "content": "For TIFA v1.0, we use GPT-3 to generate the questions. While benchmarking with TIFA v1.0 is a deterministic process, using TIFA to create a new benchmark might not be deterministic as the underlying question generator (GPT-3 in our case) might change privately. To promote deterministic benchmark generation, we fine-tune and release a LLaMA 2 (7B) [54] model that parses the captions and generates questions for arbitrary texts, using TIFA v1.0 questions as training examples." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 47, + 506, + 128, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 128, + 519 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 128, + 519 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 46, + 527, + 289, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 289, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 289, + 671 + ], + "type": "text", + "content": "In this section, we first show that TIFA has substantially higher correlations with human judgments than prior metrics on text-to-image faithfulness (§5.1). Then we present a comprehensive evaluation of existing text-to-image models using TIFA v1.0, highlighting the challenges of current text-to-image models (§5.3). Finally, we conduct an analysis of TIFA's robustness against different VQA models (§5.4). For all experiments, we use mPLUG as the VQA model for TIFA unless stated otherwise. The models we evaluate include AttnGAN [60], X-LXMERT [5], VQ-Diffusion [16], minDALL-E [27], and Stable Diffusion v1.1, v1.5, and v2.1 [46]. Details are in Appendix E." + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 306, + 206, + 503, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 206, + 503, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 206, + 503, + 220 + ], + "type": "text", + "content": "5.1. Correlation with Human Judgements" + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 225, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 225, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 225, + 545, + 285 + ], + "type": "text", + "content": "To compare TIFA with prior evaluation metrics, we first conduct human evaluations of the text-to-image models on the 1-5 Likert scale on text-to-image faithfulness. Then we compare TIFA with other metrics based on their correlation with human judgments." + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 292, + 547, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 547, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 547, + 423 + ], + "type": "text", + "content": "Likert scale on text-to-image faithfulness Annotators are asked to answer on a scale of 1 (worst) to 5 (best) to the question \"Does the image match the text?\" The detailed annotation guidelines are in Appendix C. Annotators are asked to focus on text-to-image faithfulness rather than image quality. The Likert scale should be based on how many elements in the text prompt are missed or misrepresented in the image. Objects are more important than attributes, relations, and activities. If an object is missed in the image, then all related attributes, activities, relations, etc. are also considered lost. An example is given in Figure 6." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 304, + 424, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 547, + 519 + ], + "type": "text", + "content": "We collect annotations of 800 generated images on 160 text inputs from TIFA v1.0. For each prompt, we sample an image from the 5 most recent generative models we evaluated, i.e., minDALL-E, VQ-Diffusion, Stable Diffusion v1.1, v1.5, and v2.1. We collect 2 annotations per image and average over the scores as the single \"faithfulness\" score. The inter-annotator agreement measured by Krippendorf's " + }, + { + "bbox": [ + 304, + 424, + 547, + 519 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 424, + 547, + 519 + ], + "type": "text", + "content": " is 0.67, indicating \"substantial\" agreement." + } + ] + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 340, + 529, + 435, + 624 + ], + "blocks": [ + { + "bbox": [ + 340, + 529, + 435, + 624 + ], + "lines": [ + { + "bbox": [ + 340, + 529, + 435, + 624 + ], + "spans": [ + { + "bbox": [ + 340, + 529, + 435, + 624 + ], + "type": "image", + "image_path": "dfc882a6b0fa1891f1c05c879f224d3362053a2e8e95be4b73c58bac9be2b93e.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 531, + 471, + 540 + ], + "lines": [ + { + "bbox": [ + 441, + 531, + 471, + 540 + ], + "spans": [ + { + "bbox": [ + 441, + 531, + 471, + 540 + ], + "type": "text", + "content": "Prompt:" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 628, + 547, + 673 + ], + "lines": [ + { + "bbox": [ + 304, + 628, + 547, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 547, + 673 + ], + "type": "text", + "content": "Figure 6. Illustration of our Likert scale annotation guideline. Annotators are asked to give a score of 1 to 5 based on how many elements in the text prompt are missed or misrepresented in the image. The missed elements are underlined." + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + } + ], + "index": 63 + }, + { + "bbox": [ + 441, + 542, + 511, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 441, + 542, + 511, + 597 + ], + "spans": [ + { + "bbox": [ + 441, + 542, + 511, + 597 + ], + "type": "text", + "content": "A person sitting on a horse in air over gate in grass with people and trees in background." + } + ] + } + ], + "index": 65 + }, + { + "bbox": [ + 441, + 604, + 510, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 441, + 604, + 510, + 613 + ], + "spans": [ + { + "bbox": [ + 441, + 604, + 510, + 613 + ], + "type": "text", + "content": "Miss/Misrepresent" + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 438, + 614, + 512, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 438, + 614, + 512, + 623 + ], + "spans": [ + { + "bbox": [ + 438, + 614, + 512, + 623 + ], + "type": "text", + "content": "4 out of 10 elements!" + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Baselines We compare our evaluation with two families of reference-free metrics on text-image match introduced in" + } + ] + } + ], + "index": 73 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 60, + 683, + 222, + 694 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 683, + 222, + 694 + ], + "spans": [ + { + "bbox": [ + 60, + 683, + 222, + 694 + ], + "type": "text", + "content": "3LLaMA 2 question generation model checkpoint:" + } + ] + } + ], + "index": 69 + }, + { + "bbox": [ + 48, + 694, + 266, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 694, + 266, + 703 + ], + "spans": [ + { + "bbox": [ + 48, + 694, + 266, + 703 + ], + "type": "text", + "content": "https://huggingface.co/tifa-benchmark/llama2_" + } + ] + } + ], + "index": 70 + }, + { + "bbox": [ + 48, + 704, + 165, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 704, + 165, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 704, + 165, + 712 + ], + "type": "text", + "content": "tifa_question_generation" + } + ] + } + ], + "index": 71 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20411" + } + ] + } + ], + "index": 74 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "Section 2. The first is the caption-based method. We use the state-of-the-art BLIP-2 FlanT5-XL [32] as the captioning model. The second approach is CLIPScore [17, 42]. We use CLIP (VIT-B/32) [42] to compute the score." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 71, + 160, + 261, + 327 + ], + "blocks": [ + { + "bbox": [ + 46, + 129, + 287, + 161 + ], + "lines": [ + { + "bbox": [ + 46, + 129, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 46, + 129, + 287, + 161 + ], + "type": "text", + "content": "Table 2. Correlations between each evaluation metric and human judgment on text-to-image faithfulness, measured by Spearman's " + }, + { + "bbox": [ + 46, + 129, + 287, + 161 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 129, + 287, + 161 + ], + "type": "text", + "content": " and Kendall's " + }, + { + "bbox": [ + 46, + 129, + 287, + 161 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 129, + 287, + 161 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 160, + 261, + 327 + ], + "lines": [ + { + "bbox": [ + 71, + 160, + 261, + 327 + ], + "spans": [ + { + "bbox": [ + 71, + 160, + 261, + 327 + ], + "type": "table", + "html": "
Spearman's ρKendall's τ
Caption-Based
BLEU-418.318.8
ROUGE-L32.924.5
METEOR34.027.4
SPICE32.823.2
CLIPScore33.223.1
Ours
TIFA (VILT)49.338.2
TIFA (OFA)49.637.2
TIFA (GIT)54.542.6
TIFA (BLIP-2)55.943.6
TIFA (mPLUG)59.747.2
", + "image_path": "2e400e825c157cbf30e6d94e927257c62dea065af745323490c1d98750d47b21.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 339, + 288, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 288, + 445 + ], + "type": "text", + "content": "TIFA has a much higher correlation with human judgments than prior metrics. The correlations between each evaluation metric and human judgment are shown in Table 2. For caption-based evaluations, we use metrics BLEU-4 [38], ROUGE-L [33], METEOR [2], and SPICE [1]. TIFA has higher correlations with human judgments than all previous evaluation metrics on all VQA models. TIFA (mPLUG) yields the highest correlation with human judgments among all VQA models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 456, + 245, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 245, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 245, + 468 + ], + "type": "text", + "content": "5.2. Benchmarking Text-to-Image Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 475, + 288, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 288, + 559 + ], + "type": "text", + "content": "Figure 7 shows the average TIFA score text-to-image models get on TIFA v1.0. The detailed scores with each VQA model on each element type are provided in Appendix B. We can see a clear trend of how text-to-image models evolve over time. There is a jump in TIFA score after DALL-E [44] is released, from about " + }, + { + "bbox": [ + 46, + 475, + 288, + 559 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 475, + 288, + 559 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 475, + 288, + 559 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 475, + 288, + 559 + ], + "type": "text", + "content": ". Qualitative examples of our evaluation metric are in Appendix A." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 55, + 570, + 279, + 687 + ], + "blocks": [ + { + "bbox": [ + 55, + 570, + 279, + 687 + ], + "lines": [ + { + "bbox": [ + 55, + 570, + 279, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 279, + 687 + ], + "type": "image", + "image_path": "f6444ce6e713d0de1c4dd28348aee4160542840b0ce2899eeeea58b43bcdfcaf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": "Figure 7. Average FIFA score of text-to-image models on the FIFA v1.0 benchmark. The horizontal axis shows their release dates." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 531, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 531, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 531, + 85 + ], + "type": "text", + "content": "5.3. Findings on Current Text-to-Image Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 90, + 547, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 211 + ], + "type": "text", + "content": "Figure 8 shows accuracy on each type of question in TIFA v1.0 for Stable Diffusion v1.1, v1.5, and v2.1. The score on each type reflects the text-to-image models' faithfulness in each type of visual element. To the best of our knowledge, TIFA is the only automatic evaluation method that can provide such a detailed fine-grained analysis of image generation. We separate the scores on COCO captions and other text inputs. For COCO captions, we also include the accuracy on the ground-truth images for reference. We summarize our findings in the following paragraphs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 217, + 546, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 546, + 289 + ], + "type": "text", + "content": "Generating images from captions vs. free-form text From Figure 8, we can see that VQA accuracy is higher on the COCO captions than on other text inputs. The reason is that COCO captions correspond to real images, while other text inputs may correspond to compositions that cannot be found in real-world photos (e.g. \"a blue apple\")." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 297, + 546, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 546, + 308 + ], + "type": "text", + "content": "What elements are text-to-image models struggling with?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 308, + 547, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 440 + ], + "type": "text", + "content": "Based on the scores of each category in Figure 8, we can see that Stable Diffusion models are performing well on material, animal/human, color, and location in terms of text-to-image faithfulness. However, they yield low accuracy on questions involving shapes, counting, and spatial relations. \"Other\" mainly contains abstract art notions, and models are also struggling with them. There is also a big gap between the synthesized images and real images on the COCO captions. Future work can explore various directions (e.g., training data/loss and model architecture) to improve text-to-image models' faithfulness in these aspects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 448, + 546, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 546, + 459 + ], + "type": "text", + "content": "Why are ground-truth images not getting perfect scores?" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 459, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 547, + 555 + ], + "type": "text", + "content": "Ground-truth images in COCO do not get perfect scores because 1) the COCO captions contain a substantial amount of noise from crowd workers [24] and 2) VQA models are not perfect. Real images have higher accuracy in all categories except material, color and location, where differences are small. It is left to future work to determine whether this is simply due to noise or it is an area where assessment can be improved." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 562, + 546, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 546, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 546, + 623 + ], + "type": "text", + "content": "Stable Diffusion is evolving. We can see the consistent trend that Stable Diffusion models are improving in their later versions in most of the element categories. The exceptions are \"shape\" for both prompt sources, \"other\" and \"food\" for the free-form text inputs without gold images." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "Composing multiple objects is challenging. Figure 9 shows how the number of entities (objects, animals/humans, food) in the text input affects the average TIFA score. When there are more than 5 entities, The TIFA score starts to drop rapidly for all text-to-image models, consistent with similar findings in other vision-language evaluations [15, 13]. For reference, we also add the real images in COCO in this figure." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20412" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 103, + 532, + 158 + ], + "blocks": [ + { + "bbox": [ + 140, + 71, + 454, + 81 + ], + "lines": [ + { + "bbox": [ + 140, + 71, + 454, + 81 + ], + "spans": [ + { + "bbox": [ + 140, + 71, + 454, + 81 + ], + "type": "text", + "content": "Stable Diffusion v1.1 Stable Diffusion v1.5 Stable Diffusion v2.1 COCO Ground-Truth" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 211, + 91, + 381, + 102 + ], + "lines": [ + { + "bbox": [ + 211, + 91, + 381, + 102 + ], + "spans": [ + { + "bbox": [ + 211, + 91, + 381, + 102 + ], + "type": "text", + "content": "VQA accuracy (by category) on COCO captions" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 103, + 532, + 158 + ], + "lines": [ + { + "bbox": [ + 62, + 103, + 532, + 158 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 532, + 158 + ], + "type": "image", + "image_path": "efcb7be6c7f5995a2fb4aff98010b343b41cdc22eb123bc70161d7238ea3d443.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 175, + 531, + 228 + ], + "blocks": [ + { + "bbox": [ + 214, + 163, + 378, + 175 + ], + "lines": [ + { + "bbox": [ + 214, + 163, + 378, + 175 + ], + "spans": [ + { + "bbox": [ + 214, + 163, + 378, + 175 + ], + "type": "text", + "content": "VQA accuracy (by category) on free-form text" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 175, + 531, + 228 + ], + "lines": [ + { + "bbox": [ + 62, + 175, + 531, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 531, + 228 + ], + "type": "image", + "image_path": "3b54bec0300be6452b7e8f6d98172c093c1b8c0479f62ef377ac91e13b57a970.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 547, + 266 + ], + "type": "text", + "content": "Figure 8. Accuracy on each type of question in the FIFA v1.0 benchmark. The text-to-image models are Stable Diffusion v1.1, v1.5, and v2.1. We order the categories by the average score Stable Diffusion v2.1 gets on corresponding questions. For COCO captions, we also include the accuracy of the ground-truth images for reference." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 55, + 282, + 275, + 416 + ], + "blocks": [ + { + "bbox": [ + 55, + 282, + 275, + 416 + ], + "lines": [ + { + "bbox": [ + 55, + 282, + 275, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 275, + 416 + ], + "type": "image", + "image_path": "bd62f80e860bca3bec250310b2051b16b28f57dd17450724aa49c51322947c3d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 419, + 288, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 419, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 288, + 475 + ], + "type": "text", + "content": "Figure 9. TIFA vs. numbers of entities (objects, animals/humans, and food) in the text input. The accuracy starts to drop when more than 5 entities are added to the text, showing that compositionality is hard for text-to-image models. Meanwhile, TIFA scores for COCO ground-truth (GT) images remain consistent." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 496, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 569 + ], + "type": "text", + "content": "The TIFA score on real images is rather consistent and does not change as the number of entities increases. This quantitatively shows that composing multiple objects is challenging for current text-to-image models. One possible reason is that the CLIP text embedding, which is used to train Stable Diffusion, lacks compositionality, as investigated in [37]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 578, + 185, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 185, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 185, + 590 + ], + "type": "text", + "content": "5.4. Analysis of VQA Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 597, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 658 + ], + "type": "text", + "content": "One major concern of TIFA is that VQA models can introduce some errors. Table 2 shows that TIFA has a much higher correlation with human judgment than the previous metrics, regardless of the choice of the VQA models; here we conduct a more detailed analysis." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Sensitivity of TIFA to VQA models Figure 10 shows several recent text-to-image models' TIFA scores on the COCO captions in TIFA v1.0, measured by different VQA models. We also include the TIFA scores on the ground-truth COCO" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 281, + 538, + 385 + ], + "blocks": [ + { + "bbox": [ + 312, + 281, + 538, + 385 + ], + "lines": [ + { + "bbox": [ + 312, + 281, + 538, + 385 + ], + "spans": [ + { + "bbox": [ + 312, + 281, + 538, + 385 + ], + "type": "image", + "image_path": "a0cef53a35d4ba2de499f15708435e4066ee30f3b94183cc9d6f39d93e7fab58.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 389, + 546, + 422 + ], + "lines": [ + { + "bbox": [ + 305, + 389, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 546, + 422 + ], + "type": "text", + "content": "Figure 10. Several text-to-image models' TIFA score on COCO captions, measured by different VQA models. We also include the accuracy of ground-truth images for reference." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 432, + 546, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 546, + 504 + ], + "type": "text", + "content": "images for reference. TIFA scores computed by different VQA models show a similar trend on these text-to-image models. Also, the ground-truth images get the highest TIFA score. We also computed Spearman's " + }, + { + "bbox": [ + 304, + 432, + 546, + 504 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 432, + 546, + 504 + ], + "type": "text", + "content": " of TIFA scores given by different VQA models. The pairwise correlation between all VQA models is greater than 0.6." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 512, + 547, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 547, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 547, + 583 + ], + "type": "text", + "content": "Humans performing VQA To conduct further analysis on the VQA models, we ask annotators to answer the multiple-choice visual questions in TIFA v1.0. These annotations help us evaluate the accuracy of each VQA model. For multiple-choice questions, we add the option \"None of the above\" for human evaluation. Annotation guidelines are in Appendix C." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 584, + 547, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 584, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 584, + 547, + 645 + ], + "type": "text", + "content": "We collect annotations of 1029 questions on 126 images. Each question is answered by two annotators. The inter-annotator agreement measured by Krippendorf's " + }, + { + "bbox": [ + 304, + 584, + 547, + 645 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 584, + 547, + 645 + ], + "type": "text", + "content": " is 0.88. A third annotator is involved if two annotators disagree, and the final answer is chosen by the majority vote." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 652, + 457, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 652, + 457, + 664 + ], + "spans": [ + { + "bbox": [ + 305, + 652, + 457, + 664 + ], + "type": "text", + "content": "Which VQA model should we use?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Table 3 reports the accuracy of each VQA model and the correlation between TIFA scores calculated by VQA model answers and human answers. We observe that higher model performance is directly related to the TIFA score's" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20413" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 113, + 278, + 159 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 288, + 112 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 288, + 112 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 288, + 112 + ], + "type": "text", + "content": "Table 3. Comparison of VQA models. The first row is the VQA accuracy, using the human VQA answers as reference. The second row is Spearman's correlation between TIFA scores calculated by each VQA model and the human VQA." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 113, + 278, + 159 + ], + "lines": [ + { + "bbox": [ + 54, + 113, + 278, + 159 + ], + "spans": [ + { + "bbox": [ + 54, + 113, + 278, + 159 + ], + "type": "table", + "html": "
VILTOFAGITBLIP-2mPLUG
VQA Acc.76.177.179.181.084.5
TIFA Corr.60.963.772.575.676.8
", + "image_path": "1ac4f68e319a0190fb52028e18db4d41728531a7408c22c16543410fd57702e9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 179, + 287, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 287, + 204 + ], + "type": "text", + "content": "correlation with human judgments. mPLUG has the highest accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 204, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 287, + 300 + ], + "type": "text", + "content": "Another important factor to consider is the runtime. We measure the inference speed of each VQA model on NVIDIA A40 GPU with batch size 1 over the Stable Diffusion v2.1 images (768 × 768 pixels). For one question, VILT takes 0.08s on average; OFA, GIT, and mPLUG all take about 0.25s; BLIP-2 takes 0.73s. Based on the above results, we choose mPLUG as the default VQA model for TIFA v1.0 because it is the most accurate while being reasonably fast." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 308, + 288, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 288, + 452 + ], + "type": "text", + "content": "Separation of Text-to-Image Errors and VQA Errors Suppose an image gets a wrong answer given a visual question. Then the image generation or the VQA model might have made an error. Based on the human VQA results, we separate these two kinds of errors in Figure A. If human VQA gives the wrong answer, then we suspect the generated image has an error. Otherwise, the image is correct but the VQA model is making an error. Figure A shows that the majority of errors are made by the text-to-image models. For mPLUG, less than " + }, + { + "bbox": [ + 46, + 308, + 288, + 452 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 308, + 288, + 452 + ], + "type": "text", + "content": " errors are due to the VQA model. This suggests that the TIFA framework is a viable evaluation method despite its inherent challenges." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 55, + 462, + 279, + 533 + ], + "blocks": [ + { + "bbox": [ + 55, + 462, + 279, + 533 + ], + "lines": [ + { + "bbox": [ + 55, + 462, + 279, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 279, + 533 + ], + "type": "image", + "image_path": "5c4e0034b937bd063a8987e75ea85a0d28ce9c67dc0139cca89462041cd65461.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 536, + 287, + 548 + ], + "lines": [ + { + "bbox": [ + 47, + 536, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 287, + 548 + ], + "type": "text", + "content": "Figure 11. Source of the error when VQA gets the wrong answer." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 561, + 117, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 117, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 117, + 574 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Why does TIFA work better than CLIPScore? Our experimental results and human evaluations in Section 5 show that TIFA is more accurate than prior metrics (CLIP [42] and captioning-based approaches) for evaluating text-to-image faithfulness. We hypothesize that the major challenge of these prior metrics is that they summarize the image outputs and text inputs into a single representation (embedding/caption). In contrast, TIFA exploits the power of the language models to decompose the text input into fine-grained probes, which allows VQA to capture more nuanced aspects of the text input and the generated image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "How do current VQA models perform on TIFA? One limitation is that TIFA requires VQA models to work reasonably well, which is true given the current models and TIFA v1.0, as shown in Section 5. Nevertheless, the assumption might not hold for current models in domains like anime and abstract art. TIFA is a modularized evaluation framework. The VQA models used within the framework can be updated as stronger VQA models become available in the future. For instance, we plan to incorporate GPT-4 once its image API is made public since it is likely to improve TIFA. Another possible solution is to ensemble multiple image understanding models. For example, one may employ expert models on art concepts. We leave this for future work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 230, + 547, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 547, + 387 + ], + "type": "text", + "content": "Other limitations. Another limitation of TIFA is its runtime. Answering multiple visual questions is slower than one CLIP inference. In the scenario described in Section 5.4, mPLUG takes 1.6s to evaluate one image (without batching). Also, our question generation pipeline needs one inference on a modern language model for each text input. The run time is not a critical issue for benchmarking purposes, but may not be computationally feasible for the kind of large-scale data filtering done, for example, in LAION-5B [49]. Nevertheless, we would like to point out that our evaluation is much faster than the image generation process of diffusion models. Thus, we believe it is feasible to perform reranking and reinforcement learning with TIFA on diffusion models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 398, + 383, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 398, + 383, + 410 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 383, + 410 + ], + "type": "text", + "content": "7. Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 418, + 548, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 548, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 548, + 610 + ], + "type": "text", + "content": "We present TIFA, a new automatic text-to-image faithfulness evaluation metric using VQA. Compared with prior metrics, TIFA is fine-grained, interpretable, and better aligned with human judgments. Based on this metric, we introduce the TIFA v1.0, a large-scale text-to-image benchmark containing 4K prompts and 25K questions. We conduct a comprehensive study of current text-to-image models using TIFA v1.0 and highlight the limitations of current generative models. We quantitatively show that current image generation models still struggle in counting, spatial relations, and composing multiple objects. Finally, we conduct extensive analysis and human evaluation, demonstrating that TIFA is robust to different VQA models. We hope TIFA will help evaluate future work on image generation and become increasingly sophisticated as it is upgraded with new LM, QA, and VQA components." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 621, + 421, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 421, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 421, + 635 + ], + "type": "text", + "content": "8. Acknowledgements" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "We thank Zirong Ye, Cheng-Yu Hsieh, Oscar Michel, Enhao Zhang, Jiafei Duan, Weijia Shi, Jieyu Zhang, the TIAL group, the ARK group, and the GRAIL lab at UW for their helpful feedback on this work. We would also thank all anonymous reviewers for providing insightful and constructive comments on our work." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20414" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 135 + ], + "type": "text", + "content": "[1] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In European conference on computer vision, pages 382-398. Springer, 2016. 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 191 + ], + "type": "text", + "content": "[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 247 + ], + "type": "text", + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2, 3, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "text", + "content": "[4] Soravit Changpinyo, Doron Kukliansky, Idan Szpektor, Xi Chen, Nan Ding, and Radu Soricut. All you may need for vqa are image captions. In North American Chapter of the Association for Computational Linguistics, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 292, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 288, + 335 + ], + "type": "text", + "content": "[5] Jaemin Cho, Jiasen Lu, Dustin Schwenk, Hannaneh Hajishirzi, and Aniruddha Kembhavi. X-lxmert: Paint, caption and answer questions with multi-modal transformers. ArXiv, abs/2009.11278, 2020. 6, 14, 15" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 336, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 379 + ], + "type": "text", + "content": "[6] Jaemin Cho, Abhaysinh Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers. ArXiv, abs/2202.04053, 2022. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 491 + ], + "type": "text", + "content": "[7] Hyung Won Chung, Le Hou, S. Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Wei Yu, Vincent Zhao, Yanping Huang, Andrew M. Dai, Hongkun Yu, Slav Petrov, Ed Huai hsin Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc Le, and Jason Wei. Scaling instruction-finetuned language models. ArXiv, abs/2210.11416, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 491, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 288, + 567 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 568, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 568, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 53, + 568, + 288, + 613 + ], + "type": "text", + "content": "[9] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868-12878, 2020. 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 658 + ], + "type": "text", + "content": "[10] Kawin Ethayarajh and Dan Jurafsky. Utility is in the eye of the user: A critique of NLP leaderboards. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[11] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Reddy Akula, P. Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. *ArXiv*, abs/2212.05032, 2022. 1, 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "text", + "content": "[12] Stanislav Frolov, Tobias Hinz, Federico Raue, Jorn Hees, and Andreas R. Dengel. Adversarial text-to-image synthesis: A review. Neural networks : the official journal of the International Neural Network Society, 144:187-209, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "type": "text", + "content": "[13] Mona Gandhi, Mustafa Omer Gul, Eva Prakash, Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Measuring compositional consistency for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5046-5055, 2022. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 185, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 547, + 239 + ], + "type": "text", + "content": "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 241, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 547, + 297 + ], + "type": "text", + "content": "[15] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11287-11297, 2021. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 297, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 547, + 352 + ], + "type": "text", + "content": "[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10686-10696, 2021. 6, 14, 15" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 353, + 547, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 547, + 408 + ], + "type": "text", + "content": "[17] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Joseph Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Conference on Empirical Methods in Natural Language Processing, 2021. 2, 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 409, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 409, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 409, + 547, + 453 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NIPS, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 454, + 547, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 454, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 454, + 547, + 498 + ], + "type": "text", + "content": "[19] Tobias Hinz, Stefan Heinrich, and Stefan Wermter. Semantic object accuracy for generative text-to-image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:1552-1565, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 499, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 499, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 499, + 547, + 544 + ], + "type": "text", + "content": "[20] Seunghoon Hong, Dingdong Yang, Jongwook Choi, and Honglak Lee. Inferring semantic layout for hierarchical text-to-image synthesis. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7986-7994, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 544, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 547, + 578 + ], + "type": "text", + "content": "[21] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 578, + 547, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 547, + 622 + ], + "type": "text", + "content": "[22] Yushi Hu, Chia-Hsuan Lee, Tianbao Xie, Tao Yu, Noah A Smith, and Mari Ostendorf. In-context learning for few-shot dialogue state tracking. arXiv preprint arXiv:2203.08568, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 624, + 547, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 547, + 679 + ], + "type": "text", + "content": "[23] Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras, Lavinia Dunagan, Jacob Morrison, Alexander Fabbri, Yejin Choi, and Noah A. Smith. Bidimensional leaderboards: Generate and evaluate language hand in hand. In North American Chapter of the Association for Computational Linguistics, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 680, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 547, + 714 + ], + "type": "text", + "content": "[24] Jungo Kasai, Keisuke Sakaguchi, Lavinia Dunagan, Jacob Daniel Morrison, Ronan Le Bras, Yejin Choi, and Noah A. Smith. Transparent human evaluation for image captioning." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20415" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "text", + "content": "In North American Chapter of the Association for Computational Linguistics, 2022. 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "text", + "content": "[25] Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single qa system. In Findings, 2020. 2, 4, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 207 + ], + "type": "text", + "content": "[26] Daniel Khashabi, Gabriel Stanovsky, Jonathan Bragg, Nicholas Lourie, Jungo Kasai, Yejin Choi, Noah A. Smith, and Daniel Weld. GENIE: Toward reproducible and standardized human evaluation for text generation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 240 + ], + "type": "text", + "content": "[27] Saehoon Kim, Sanghun Cho, Chiheon Kim, Doyup Lee, and Woonhyuk Baek. mindall-e on conceptual captions, 2021. 6, 14, 15" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 288, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 288, + 286 + ], + "type": "text", + "content": "[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 288, + 331 + ], + "type": "text", + "content": "[29] Ranjay Krishna, Michael S. Bernstein, and Li Fei-Fei. Information maximizing visual question generation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2008–2018, 2019. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 333, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 288, + 387 + ], + "type": "text", + "content": "[30] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 464 + ], + "type": "text", + "content": "[31] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng da Cao, Ji Zhang, Songfang Huang, Feiran Huang, Jingren Zhou, and Luo Si. mplug: Effective and efficient vision-language learning by cross-modal skip-connections. In Conference on Empirical Methods in Natural Language Processing, 2022. 2, 4, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "type": "text", + "content": "[32] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ArXiv, abs/2301.12597, 2023. 2, 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 555 + ], + "type": "text", + "content": "[33] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain, July 2004. Association for Computational Linguistics. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 612 + ], + "type": "text", + "content": "[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2, 5, 13, 15" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "text", + "content": "[35] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum. Compositional visual generation with composable diffusion models. ArXiv, abs/2206.01714, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[36] Rosanne Liu, Daniel H Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, R. J. Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. ArXiv, abs/2212.10562, 2022. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[37] Zixian Ma, Jerry Hong, Mustafa Omer Gul, Mona Gandhi, Irena Gao, and Ranjay Krishna. Crepe: Can vision-language foundation models reason compositionally? ArXiv, abs/2212.07796, 2022. 2, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": "[38] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 228 + ], + "type": "text", + "content": "[39] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 230, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 547, + 285 + ], + "type": "text", + "content": "[40] Vitali Petsiuk, Alexander E Siemann, Saisamrit Surbehera, Zad Chin, Keith Tyser, Gregory Hunter, Arvind Raghavan, Yann Hicke, Bryan A Plummer, Ori Kerret, et al. Human evaluation of text-to-image models on a multi-task benchmark. arXiv preprint arXiv:2211.12112, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 286, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 547, + 319 + ], + "type": "text", + "content": "[41] Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, and Mike Lewis. Measuring and narrowing the compositionality gap in language models, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 321, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 547, + 386 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 3, 7, 9, 14" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 388, + 547, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 547, + 420 + ], + "type": "text", + "content": "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 422, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 547, + 464 + ], + "type": "text", + "content": "[44] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 1, 7, 15" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 468, + 547, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 547, + 499 + ], + "type": "text", + "content": "[45] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. ArXiv, abs/1908.10084, 2019. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 501, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 547, + 555 + ], + "type": "text", + "content": "[46] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 1, 6, 14, 16" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 557, + 547, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 547, + 633 + ], + "type": "text", + "content": "[47] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1, 2, 5, 13, 15" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 635, + 547, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 547, + 668 + ], + "type": "text", + "content": "[48] Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. ArXiv, abs/1606.03498, 2016. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "text", + "content": "[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next gener" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20416" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 95 + ], + "type": "text", + "content": "ation image-text models. arXiv preprint arXiv:2210.08402, 2022.9" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 150 + ], + "type": "text", + "content": "[50] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 217 + ], + "type": "text", + "content": "[51] Hongjin Su, Jungo Kasai, Chen Henry Wu, Weijia Shi, Tianlu Wang, Jiayi Xin, Rui Zhang, Mari Ostendorf, Luke Zettlemoyer, Noah A. Smith, and Tao Yu. Selective annotation makes language models better few-shot learners. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 274 + ], + "type": "text", + "content": "[52] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2826, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 308 + ], + "type": "text", + "content": "[53] Hao Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. ArXiv, abs/1908.07490, 2019. 15" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 310, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 288, + 363 + ], + "type": "text", + "content": "[54] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 288, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 409 + ], + "type": "text", + "content": "[55] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 454 + ], + "type": "text", + "content": "[56] Alex Wang, Kyunghyun Cho, and Mike Lewis. Asking and answering questions to evaluate the factual consistency of summaries. In Annual Meeting of the Association for Computational Linguistics, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 288, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 499 + ], + "type": "text", + "content": "[57] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 288, + 555 + ], + "type": "text", + "content": "[58] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR, abs/2202.03052, 2022. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 288, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 623 + ], + "type": "text", + "content": "[59] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "text", + "content": "[60] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316-1324, 2017. 6, 14, 15" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[61] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 326, + 72, + 547, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 547, + 126 + ], + "type": "text", + "content": "Yang, Burcu Karagol Ayan, Benton C. Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. ArXiv, abs/2206.10789, 2022. 1, 2, 5" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20417" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/7bfffc83-27ef-4c73-aa19-9a15a6ee13fd_content_list.json b/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/7bfffc83-27ef-4c73-aa19-9a15a6ee13fd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0a27b38a06cbb51a86a453353d7159391d453bae --- /dev/null +++ b/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/7bfffc83-27ef-4c73-aa19-9a15a6ee13fd_content_list.json @@ -0,0 +1,1360 @@ +[ + { + "type": "text", + "text": "TIJO: Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models", + "text_level": 1, + "bbox": [ + 104, + 128, + 867, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Indranil Sur $^{1*}$ Karan Sikka $^{1}$ Matthew Walmer $^{2}$ Kaushik Koneripalli $^{1}$ Anirban Roy $^{1}$ Xiao Lin $^{1}$ Ajay Divakaran $^{1}$ Susmit Jha $^{1}$ SRI International $^{2}$ University of Maryland", + "bbox": [ + 184, + 202, + 776, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 297, + 313, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a Multimodal Backdoor Defense technique TIJO (Trigger Inversion using Joint Optimization). Recent work [48] has demonstrated successful backdoor attacks on multimodal models for the Visual Question Answering task. Their dual-key backdoor trigger is split across two modalities (image and text), such that the backdoor is activated if and only if the trigger is present in both modalities. We propose TIJO that defends against dual-key attacks through a joint optimization that reverse-engineers the trigger in both the image and text modalities. This joint optimization is challenging in multimodal models due to the disconnected nature of the visual pipeline which consists of an offline feature extractor, whose output is then fused with the text using a fusion module. The key insight enabling the joint optimization in TIJO is that the trigger inversion needs to be carried out in the object detection box feature space as opposed to the pixel space. We demonstrate the effectiveness of our method on the TrojVQA benchmark, where TIJO improves upon the state-of-the-art unimodal methods from an AUC of 0.6 to 0.92 on multimodal dual-key backdoors. Furthermore, our method also improves upon the unimodal baselines on unimodal backdoors. We present ablation studies and qualitative results to provide insights into our algorithm such as the critical importance of overlaying the inverted feature triggers on all visual features during trigger inversion. The prototype implementation of TIJO is available at https://github.com/SRI-CSL/TIJO.", + "bbox": [ + 75, + 330, + 473, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 771, + 209, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Neural Networks (DNNs) are vulnerable to adversarial attacks [47, 1, 16, 24]. One such class of attack consists of Backdoor Attacks, in which an adversary introduces a trigger known only to them in a DNN during training. Such a backdoored DNN will behave normally with typi", + "bbox": [ + 75, + 797, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ba32bbe2db80ccbd129d9e2c8e63106dbfde662160d2e9d7007d2accea5cc1f.jpg", + "image_caption": [ + "Figure 1. (Top) A dual-key backdoor attack for multimodal models [48], which is designed to activate if and only if the trigger is present in both the modalities. Such backdoors cannot be detected by unimodal defenses. (Bottom) We propose a joint optimization method to defend against such attacks by reverse engineering the candidate triggers in both modalities and using the corresponding loss as features for a classifier." + ], + "image_footnote": [], + "bbox": [ + 519, + 297, + 869, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cal in-distribution inputs but perform poorly (e.g. produce targeted misclassifications) on inputs stamped with a predefined trigger designed by the adversary [50, 21, 30].", + "bbox": [ + 496, + 655, + 892, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent work [48, 7] has introduced backdoors in multimodal domains such as Visual Question Answering (VQA) and Fake News Detection [7, 48]. In prior work [48], we have introduced a Dual-Key Backdoor Attack (shown in Figure 1), where the trigger is inserted in both the image and text modalities in such a manner that the backdoor is activated only when both modalities contain the trigger. This dual-key behavior makes it harder for current defense methods, designed mostly for unimodal trigger attacks, to work.", + "bbox": [ + 496, + 702, + 892, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There has been significant work developing defenses against backdoor attacks in the visual domain, in particular for the image classification task [45, 49, 6, 23]. Recent works have also explored defense in natural language pro", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: indranil.sur@sri.com", + "bbox": [ + 94, + 886, + 338, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "165", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cessing domains [38, 43, 34]. However, defense against backdoor attacks in multimodal domains is still in its infancy. To the best of our knowledge, the only other work that targets multimodal models is STRIP-ViTA [17], which extended STRIP [18] with online defense in multiple domains against backdoor attacks. Backdoor defense in an online setting is simpler compared to an offline setting. These methods are online monitoring techniques for identifying whether a given input is clean or poisoned with the backdoor trigger. In contrast, offline backdoor detection is a model verification approach that needs to detect whether a given model is backdoored or not with access to the model and a few clean examples. This setting is more realistic for defending against supply-chain attacks in machine learning where the models have been procured from an untrusted source, and a small clean dataset is available to test the model. We focus on multimodal defense in such an offline setting.", + "bbox": [ + 75, + 90, + 472, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a novel approach for defending against multimodal backdoor attacks, referred to as Trojan Inversion using Joint Optimization (TIJO), that reverse engineers the triggers in both modalities. Our approach is motivated by the Universal Adversarial Trigger (UAT) [47] that was proposed to identify naturally occurring universal triggers in pre-trained NLP models and has been extended in earlier works to identify Trojan triggers in NLP models. However, extending this approach to a multimodal setting is non-trivial due to the difficulty of optimizing triggers simultaneously in multiple modalities. Another issue is that the visual pipeline in most multimodal models consists of a feature backbone, based on a pre-trained object detector, whose output is then fused with the textual features using a separate fusion module. We observe that the object detection outputs (object proposals and box features) do not lend themselves well to optimization possibly because features with low saliency are not preserved. Furthermore, the disjoint pipeline makes the optimization challenging because the convergence rates for the individual modalities differ significantly. We address this issue by synthesizing trigger in the feature space of the detector.", + "bbox": [ + 75, + 363, + 472, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate TIJO on the TrojVQA dataset [48] that consists of over 800 VQA models spanning across 4 feature backbones and 10 model architectures. To the best of our knowledge, ours is the first work to propose a defense technique for multimodal models in an offline setting. Our results indicate strong improvement over prior unimodal methods. Our contributions are as follows:", + "bbox": [ + 75, + 696, + 470, + 800 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel approach for Multimodal Backdoor defense referred to as TIJO.", + "- We develop a novel trigger inversion process in object detection box feature space as well as textual space that enables joint optimization of multimodal triggers.", + "- We demonstrate TIJO on the TrojVQA dataset and show" + ], + "bbox": [ + 76, + 810, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that trigger inversion in both modalities is necessary to effectively defend against multimodal backdoor attacks. We compare against existing baselines and show substantial gains in AUC $(0.6\\to 0.92)$", + "bbox": [ + 509, + 90, + 890, + 150 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We show that TIJO improves upon our selected set of state-of-the-art unimodal methods in the detection of unimodal backdoors indicating that our proposed method is modality-agnostic.", + "- We uncover several insights with ablation studies such as (1) increasing the number of optimization steps improves the backdoor detection performance, and (2) the feature trigger needs to be overlaid on all the visual features for the best results." + ], + "bbox": [ + 500, + 152, + 890, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 310, + 640, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backdoor Attacks: Backdoor attacks are a type of targeted adversarial attack that were first introduced in [21]. Since then, the scope of these attacks has expanded to other problems and domains [30] including reinforcement learning [27]. Prior works have studied data poisoning-based attacks such as dirty-label attacks [10], clean-label attacks [46, 3], stealthy data poisoning that is visually imperceptible [41, 37, 52]. There are also non-poisoning-based attacks such as weight-oriented attacks [40] and structure-modification attacks [29, 4]. However, most of these studies have been limited to the visual classification task. Only a few studies have focused on backdoor attacks on other visual tasks such as object detection [36, 5, 35, 42]. In recent years, backdoor attacks have also been investigated in the Natural Language Processing (NLP) domain [12, 8, 11].", + "bbox": [ + 496, + 335, + 890, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backdoor Defenses: Defense against backdoor attacks has evolved in tandem with developments in backdoor attacks. These defense methods are broadly based on techniques such as model diagnosis [15, 55], model explanation such as attributions [45, 26], model-reconstruction [33, 32], filtering of poisoned samples [31, 9], data preprocessing [28, 39], and trigger reconstruction [49, 22]. Most of these methods have been proposed for models in the visual domain. There have been some recent works on backdoor defense in the NLP domain. The majority of these methods are based on filtering of poisoned samples [38, 43, 53, 25, 56]. Other works rely on ideas such as model diagnosis [14, 19], prepossessing-based [2], and trigger synthesis [34, 44].", + "bbox": [ + 496, + 580, + 890, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Modal Backdoor Attacks & Defenses: Recent studies have also extended data-poisoning based backdoor attacks into multimodal domains. Chen et al. [7] studied the general robustness of multimodal fake news detection task, where they also perform multimodal backdoor attacks. Walmer et al. [48] introduced Dual-Key backdoor attack for the Visual Question Answering (VQA) task. As shown in", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "166", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eb8ccf1996b4dd7f0a4de9c54beb1ebc2cce2d96f4da9f062351b11c59f9655e.jpg", + "image_caption": [ + "Figure 2. Shows key blocks of TIJO. (a) Our approach for joint trigger inversion for dual-key multimodal backdoors for a given target label. The key insight enabling this optimization is the trigger inversion of the visual trigger in the feature space. (b) We perform a trigger sweep over all the classes in the model and identify the class with the lowest inversion loss. (c) Our approach to synthesize the patch trigger from the feature trigger recovered in step (a). (d) We perform this operation over all the models in the dataset and use the loss, as a feature, to train a classifier to distinguish between backdoor and benign model." + ], + "image_footnote": [], + "bbox": [ + 158, + 88, + 807, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 1, this attack was designed to trigger the backdoor only when the trigger is present in both modalities, which makes the attack stealthier compared to a unimodal trigger.", + "bbox": [ + 75, + 477, + 468, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Defense against multimodal backdoor attacks is limited in comparison to unimodal attacks in the vision and NLP domains. Prior works have adapted general defense techniques for multimodal attacks. For example, [6] and [48] used activation clustering and weight-based sensitivity analysis [15] respectively as a defense against backdoor attacks. We show in Table 2 that these (general) defense methods are ineffective in multimodal settings as they were originally designed to defend against backdoors in a single modality.", + "bbox": [ + 75, + 523, + 468, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gao et al. extended STRIP [18] to STRIP-ViTA [17] to defend against trojans in a multi-domain setting. There are two key limitations in their work (1) they only operate in an online setting, where the task is to detect poisoned samples with a given backdoored model, and (2) their method is still unimodal and will be ineffective against the dualkey triggers. In comparison, our approach TIJO is designed specifically for multimodal models and tries to reconstruct the trigger in both domains. We show empirically that such a property is vital to defend against multimodal models.", + "bbox": [ + 75, + 660, + 468, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 828, + 184, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first discuss the threat model that we aim to defend against, then discuss the UAT method [47] and its extension to multimodal models, and present our method, TIJO.", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Threat Model", + "text_level": 1, + "bbox": [ + 500, + 476, + 640, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a multimodal model $f$ , we need to determine if $f$ is benign or backdoored. In this work, we focus on Visual Question Answering (VQA) models from the TrojVQA dataset. Let $\\mathcal{C}$ be the clean VQAv2 dataset [20] where each data entry is a triplet $(\\pmb{x},\\pmb{t},\\pmb{y})$ where $\\pmb{x}$ is the image, $\\pmb{t}$ is the tokenized question, and $y$ is the answer label. Most VQA models use a two-step process for generating the answer. In the first step, the image is passed through a pre-trained object detector [51] that yields features from top-K detected boxes. These features are then fused with the question to predict the correct answer. Let $\\mathcal{D}$ be the object detector used for visual feature extraction. The answer is generated using $f(t,\\mathcal{D}(\\pmb{x})) = y$ .", + "bbox": [ + 496, + 502, + 890, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our threat model, we assume that $\\mathcal{D}$ is benign and the adversary introduces the backdoor in the VQA model $f$ . This is also the threat model used in the TrojVQA dataset [48]. For a backdoored VQA model $f_{b}$ , the adversary introduces triggers $\\pmb{p}_{t}$ and $\\pmb{t}_{t}$ in both the image and text modalities respectively. $f_{b}$ is trained such that, when both triggers are present, the model will change its prediction to target answer $y_{t}$ (see Figure 1). In the TrojVQA dataset, $\\pmb{p}_{t}$ are small visual patches while $\\pmb{t}_{t}$ are natural words. The triggers and the model behavior are only known to the adversary.", + "bbox": [ + 496, + 700, + 890, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{M}$ be a policy that overlaps $\\pmb{p}_t$ on $\\pmb{x}$ and $\\mathcal{A}$ be a policy that appends $\\pmb{t}_t$ to $\\pmb{t}$ . Hence, for a backdoored VQA", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "167", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "model $f_{b}$ , we expect that", + "bbox": [ + 76, + 90, + 243, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {b} (\\mathcal {A} (\\boldsymbol {t}, \\boldsymbol {t} _ {t}), \\mathcal {D} (\\mathcal {M} (\\boldsymbol {x}, \\boldsymbol {p} _ {t}))) = y _ {t}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 117, + 380, + 133 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we focus on dual-key triggers [48], where the model changes its prediction only when both $\\mathcal{M}$ and $\\mathcal{A}$ are applied together.", + "bbox": [ + 76, + 143, + 468, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Trigger Inversion using UAT", + "text_level": 1, + "bbox": [ + 76, + 199, + 333, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TIJO is based on Universal Adversarial Triggers (UAT) [47], which extends Hotflip [13] from synthesizing adversarial tokens for a single input to all inputs in the dataset. As a result, obtained adversarial tokens are universal in nature. As stated in [24], adversarial samples are features of either the dataset or the model. Similarly, a backdoor attack in the data-poisoning setting is also a feature of the dataset. Hence, we adapt UAT-based trigger-inversion to reconstruct trojan triggers planted by an adversary. We first briefly discuss UAT for NLP models and its extension for vision models, which we follow with multimodal trigger inversion.", + "bbox": [ + 75, + 222, + 468, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq. 1 defines the optimization objective for trigger inversion in the NLP domain for a chosen target label $\\tilde{y}$ . Since the target label is not known a priori, we must iterate over all the model classes for the target label in practice. Here $\\mathcal{L}$ is the cross-entropy loss, and we optimize to minimize the expected loss over all samples in $S$ . In summary, we optimize to get the $t_{adv}$ that maximizes the likelihood of switching the class label to $\\tilde{y}$ for all samples in $S$ . Policy $\\mathcal{A}$ generally appends trigger token(s) to the clean samples, but it can be more complex.", + "bbox": [ + 75, + 388, + 468, + 540 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {t} _ {a d v}} \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {x} \\sim \\mathcal {S}} \\left[ \\mathcal {L} \\left(\\tilde {y}, f (\\mathcal {A} \\left(\\boldsymbol {t} _ {a d v}, \\boldsymbol {t}\\right), \\mathcal {D} (\\boldsymbol {x}))\\right) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 550, + 468, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the space of $t_{adv}$ is discrete, each optimization step is followed by a next token selection step. The next token is set by $t_{adv} \\gets t_i$ which minimizes the trigger inversion loss's first-order Taylor approximation around the current token embedding as given by Eq. 2. Here $\\mathcal{V}_f$ is the vocabulary of all tokens in $f$ , function $\\mathcal{E}_f$ gives the token embeddings and $\\nabla_{\\mathcal{E}_f(t_{adv})}\\mathcal{L}$ is the average gradient of the loss over a batch.", + "bbox": [ + 75, + 584, + 468, + 703 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {t} _ {i} \\in \\mathcal {V} _ {f}} \\left[ \\mathcal {E} _ {f} \\left(\\boldsymbol {t} _ {i}\\right) - \\mathcal {E} _ {f} \\left(\\boldsymbol {t} _ {a d v}\\right) \\right] ^ {\\intercal} \\nabla_ {\\mathcal {E} _ {f} \\left(\\boldsymbol {t} _ {a d v}\\right)} \\mathcal {L} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 714, + 468, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above optimization problem is solved efficiently by computing dot products between the gradient and the $\\nu_{f}$ embeddings and then using nearest neighbor or beam search to get the updated token $t_i$ [47]. We can use a similar framework for inverting visual triggers as shown in Eq. 3. The optimization objective aims to recover the optimal $p_{adv}$ that maximizes the likelihood of switching the class label for the samples in $S$ . The only difference is that we use projected gradient descent for patch $p_{adv}$ , overlaid on $x$ through policy $\\mathcal{M}$ , which needs to obey image constraints. This ap", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "proach is similar to prior trigger reconstruction-based methods such as Neural Cleansse [49].", + "bbox": [ + 498, + 90, + 888, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {p} _ {a d v}} {\\min } \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {x} \\sim \\mathcal {S}} \\left[ \\mathcal {L} (\\tilde {y}, f (\\boldsymbol {t}, \\mathcal {D} (\\mathcal {M} (\\boldsymbol {x}, \\boldsymbol {p} _ {a d v})))) \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 128, + 890, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multimodal Trigger Inversion with TIJO", + "text_level": 1, + "bbox": [ + 500, + 165, + 849, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now outline our approach for multimodal Trigger Inversion using Joint Optimization (TIJO) (shown in Figure 2). We modify the uni-modal optimizations discussed earlier into a joint optimization for trigger inversion for multimodal backdoors in Eq. 4. Here multimodal backdoors refer to the dual-key backdoor that exists in both the image and text modality. We optimize for both $\\pmb{t}_{adv}$ and $\\pmb{p}_{adv}$ to maximize the likelihood of switching the class label to $\\tilde{y}$ for all samples in $S$ .", + "bbox": [ + 496, + 188, + 890, + 325 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {t} _ {a d v}, \\boldsymbol {p} _ {a d v}} \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {x} \\sim \\mathcal {S}} \\left[ \\mathcal {L} \\left(\\tilde {y}, f \\left(\\mathcal {A} \\left(\\boldsymbol {t} _ {a d v}, \\boldsymbol {t}\\right), \\mathcal {D} \\left(\\mathcal {M} \\left(\\boldsymbol {p} _ {a d v}, \\boldsymbol {x}\\right)\\right)\\right)\\right) \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 330, + 890, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Solving Eq. 4 for multimodal (dual-key) backdoors is challenging. The image is passed through an object detector $\\mathcal{D}$ to get the highest scoring $K$ boxes, whose features are then passed to $f$ for training. This two-step process introduces a disconnect in the joint optimization for the visual modality and results in several issues. For example, when we stamp the patch on the image during optimization, the detector $\\mathcal{D}$ may not propose bounding boxes containing the patch $\\pmb{p}_{adv}$ . One solution would be to manually force the detector to sample a proposal around $\\pmb{p}_{adv}$ . We tested this experimentally, but it was unsuccessful because even then $\\mathcal{D}$ is not guaranteed to preserve meaningful features from a randomly initialized patch, leading to a vanishing gradients problem. Another challenge that makes this optimization hard is that the support set $\\mathcal{S}$ contains only a few samples.", + "bbox": [ + 496, + 368, + 892, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposed key idea: We propose to overcome this issue and enable the convergence for both the visual and textual trigger by performing trigger inversion for the visual triggers in the feature space of $\\mathcal{D}$ , while the textual trigger is optimized in the token space as done for UAT. We define $f_{adv}$ as the additive adversarial feature space signature and $\\mathcal{B}$ as the overlay policy by which we overlay $f_{adv}$ on box features from $\\mathcal{D}$ . The modified optimization objective is shown in Eq. 5, where we optimize $t_{adv}$ and $f_{adv}$ instead of $p_{adv}$ . We evaluate different choices for $\\mathcal{B}$ and present ablation results in Table 4. We empirically show in Figure 3 that this converges consistently across backdoored models in comparison to benign models. We have shown a detailed description of our approach in Figure 2. Similar to UAT, we optimize Eq. 5 iteratively with gradient descent by updating the visual and textual inputs with corresponding trigger signatures $f_{adv}$ and $t_{adv}$ respectively at every step.", + "bbox": [ + 496, + 595, + 892, + 852 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {t} _ {a d v}, \\boldsymbol {f} _ {a d v}} \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {x} \\sim \\mathcal {S}} \\left[ \\mathcal {L} \\left(\\tilde {y}, f \\left(\\mathcal {A} \\left(\\boldsymbol {t} _ {a d v}, \\boldsymbol {t}\\right), \\mathcal {B} \\left(\\mathcal {D} (\\boldsymbol {x}), \\boldsymbol {f} _ {a d v}\\right)\\right)\\right) \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 863, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "168", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5c35654038be906b515eb041efccf2d131482af00c66b04efbd9bfc0c646029f.jpg", + "image_caption": [ + "Figure 3. Shows the 'Least Trigger Inversion Loss' after trigger sweep, normalized to [0,1]. The blue and red dots are benign and backdoor models respectively. Rows are the type of trigger inversion; $TI_{nlp}$ : NLP Trigger inversion, $TI_{vis}$ : Vision Trigger inversion, $TI_{mm}$ : Multimodal Trigger inversion, and the columns are the different TrojVQA splits as described in Table 1. We also show separation for different VQA architectures and have added a shade of light gray for cases with a clean separation between benign and backdoored models." + ], + "image_footnote": [], + "bbox": [ + 138, + 88, + 831, + 364 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6e7f72a4e5acbedd14fbf14eef71576a891b5bc106082766a503757c4dd7fa94.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SplitNLPVisualTrain/TestTrigger Type
Tnlp✓X160/80Single Key NLP
TsolidXSolid160/80Single Key Vision
ToptimXOptimized160/80Single Key Vision
Tnlp+S✓Solid160/80Dual Key
Tnlp+O✓Optimized160/80Dual Key
T✓✓320/160Dual Key
", + "bbox": [ + 81, + 446, + 467, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Details about the TrojVQAdataset [48] and its splits.", + "bbox": [ + 76, + 551, + 437, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Trigger Patch Generation", + "text_level": 1, + "bbox": [ + 76, + 601, + 310, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also propose to recover the patch trigger $\\pmb{p}_{adv}$ based on the $\\pmb{f}_{adv}$ obtained using Eq. 5 (see Figure 2). We first compute the box proposals $\\pmb{b}_x \\gets \\mathcal{D}_{rpn}(\\mathcal{D}_{cnn}(\\pmb{x}))$ and box features $\\pmb{f}_x \\gets \\mathcal{D}_{roi}(\\mathcal{D}_{cnn}(\\pmb{x}),\\pmb{b}_x)$ on the clean image $\\pmb{x}$ . We also compute the box features $\\pmb{f}_{x_p} \\gets \\mathcal{D}_{roi}(\\mathcal{D}_{cnn}(\\mathcal{M}(\\pmb{x},\\pmb{p}_{adv})),\\pmb{b}_x)$ on the image stamped with $\\pmb{p}_{adv}$ . Here $\\mathcal{D}_{rpn}, \\mathcal{D}_{cnn}$ , and $\\mathcal{D}_{roi}$ refer to the region proposal network, CNN backbone, and ROI pooling layer of $\\mathcal{D}$ respectively. We overlay $\\pmb{f}_{adv}$ on $\\pmb{f}_x$ and then iteratively optimize $\\pmb{p}_{adv}$ to minimize the MSE loss between $\\pmb{f}_{x_p}$ and $\\mathcal{B}(\\pmb{f}_x,\\tilde{\\pmb{f}}_{adv})$ . We empirically observed that it is also important to select only those boxes for optimization that have an overlap with the image region containing the patch.", + "bbox": [ + 75, + 625, + 472, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Backdoored Model Classification", + "text_level": 1, + "bbox": [ + 76, + 832, + 366, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The optimization objective should ideally converge only if the model is backdoored and if the target label $\\tilde{y}$ is actually the poison label $y_{t}$ . We use this convergence property", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to train a classifier to separate backdoored and benign models. Since the poison label $y_{t}$ is unknown, we sweep over all the label space, $\\forall \\tilde{y} \\in \\mathcal{V}$ and repeat the trigger inversion process for each $\\tilde{y}$ (referred to as trigger sweep). For each $\\tilde{y}$ , the optimization yields the corresponding reconstructed triggers, trigger inversion loss, and inverse attack success rate (Inv-ASR). Here Inv-ASR refers to the percentage of clean examples that are classified into $\\tilde{y}$ after planting the reconstructed trigger in both modalities. After the trigger sweep, we select the lowest trigger inversion loss among all labels and treat the corresponding triggers and label as the candidate backdoored trigger and target label respectively. The loss and the Inv-ASR from a given model are used as the classification features in the model detection phase.", + "bbox": [ + 496, + 450, + 893, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first obtain the classification features for all the models in the dataset. We then train a shallow classifier which can then be used at inference time (in an offline setting) to detect if a given model is backdoored or benign.", + "bbox": [ + 496, + 662, + 893, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 738, + 633, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our approach in this section. We first discuss the dataset and metrics used for evaluation. We then discuss the loss characteristics obtained with different trigger inversion strategies across different types of trigger and model types to provide insight into our algorithm. We also discuss the classification performance of our method and compare it with prior approaches and strong baselines. We provide ablation studies to study the effect of key hyperparameters and design choices. Finally, we provide visualizations of", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "169", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4c745599737fd72cb3c842ff4bef0569063a9d71adf969b1b2299d1adaa60f74.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SplitGeneral Wt. AnalysisDBSUnimodal NCTABORTIJO \\( _{nlp} \\)Ours TIJO \\( _{vis} \\)TIJO \\( _{mm} \\)
Single Key\\( T_{nlp} \\)0.61±0.070.89±0.05--0.98±0.020.52±0.060.98±0.02
\\( T_{solid} \\)0.53±0.05-0.59±0.100.98±0.020.39±0.091.00±0.000.99±0.01
\\( T_{optim} \\)0.58±0.05-0.71±0.080.99±0.020.40±0.110.99±0.010.95±0.03
Dual Key\\( T_{nlp+S} \\)0.54±0.030.46±0.040.42±0.050.46±0.060.41±0.110.70±0.060.97±0.03
\\( T_{nlp+O} \\)0.60±0.130.45±0.010.50±0.090.52±0.030.43±0.120.57±0.070.86±0.10
\\( T \\)0.60±0.040.48±0.020.50±0.060.48±0.040.46±0.030.67±0.070.92±0.02
", + "bbox": [ + 96, + 88, + 870, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6eefd995075e4846c7adc3fe3f97766799b3b7b688b8988ca10d69984acb5fd9.jpg", + "image_caption": [ + "Figure 4. Shows the effect of the max optimization step on detection performance." + ], + "image_footnote": [], + "bbox": [ + 161, + 300, + 803, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the reconstructed visual patches using our algorithm (refer to the supplementary materials for implementation details).", + "bbox": [ + 75, + 483, + 470, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TrojVQA Dataset and Metric: We use the TrojVQA [48] dataset that was introduced recently and consists of both benign and poisoned VQA models. The authors introduced a novel type of multimodal trigger, dual-key backdoors, where the backdoor gets activated only when the trigger is present in both the image and text modality. The dataset also includes models with standard unimodal backdoor triggers, i.e. the trigger was introduced in either the text or image modality only. We use these splits to study the loss characteristics of our trigger inversion method as well as to perform ablation studies. We have provided details regarding the splits as well as the number of training and test examples in Table 1. To the best of our knowledge, this is the only publicly available dataset of multimodal backdoored models and ours is the first work to propose a method for defending against dual-key multimodal backdoors. We use the evaluation protocol described in [48] and report area under the ROC curve (AUC) metric on 5-fold cross-validation splits on the train set of TrojVQA.", + "bbox": [ + 75, + 532, + 468, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Trigger Inversion Loss Characteristics", + "text_level": 1, + "bbox": [ + 76, + 830, + 408, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show the loss characteristics of our trigger inversion approach in Figure 3. This loss is obtained after optimizing Eq. 5 and trigger-sweep (as discussed in Section 3.5). The", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8545c89da0fe6c7eaf13d011cccfb1d263804b0519e6fd47af50af276488f823.jpg", + "table_caption": [ + "Table 2. Shows AUC for different $TrojVQA$ splits with weight analysis, prior unimodal methods as well as three variants of our method- $\\mathrm{TIJO}_{\\mathrm{np}}$ , $\\mathrm{TIJO}_{\\mathrm{vis}}$ , and $\\mathrm{TIJO}_{\\mathrm{mm}}$ which optimize triggers in NLP, vision, and both modalities respectively. We see a clear improvement with $\\mathrm{TIJO}_{\\mathrm{mm}}$ for not only dual-key multimodal triggers but also for unimodal triggers. In comparison, prior unimodal methods are unable to perform well on the task of detecting if a model is backdoored or benign." + ], + "table_footnote": [], + "table_body": "
SplitModelInv-ASRLowest Loss
TnlpTIJOnlp0.94±0.050.98±0.02
TIJOmm0.54±0.030.98±0.02
TsolidTIJOvis0.91±0.051.00±0.00
TIJOmm0.56±0.040.99±0.01
ToptimTIJOvis0.90±0.040.99±0.01
TIJOmm0.54±0.020.95±0.03
TTIJOmm0.53±0.020.92±0.02
", + "bbox": [ + 529, + 479, + 859, + 611 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/10e66f413b613d60aff163ced0cc614cf6894a8697304dc43cb790b18c75ee1b.jpg", + "table_caption": [ + "Table 3. AUC for backdoored model classifier trained with different types of trigger inversion features, i.e. least loss features and maximum switch to target accuracy." + ], + "table_footnote": [], + "table_body": "
SplitTIJOvisTIJOmm
BoneBallBoneBall
Tsolid0.85±0.041.00±0.000.86±0.100.99±0.01
Toptim0.78±0.060.99±0.010.80±0.060.95±0.03
Tnlp+S0.47±0.080.70±0.060.77±0.050.97±0.03
Tnlp+O0.46±0.110.57±0.070.65±0.040.86±0.10
T0.52±0.040.67±0.070.72±0.070.92±0.02
", + "bbox": [ + 501, + 670, + 903, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. AUC for backdoored model classifier train with features obtain from different feature overlay policy $\\mathcal{B} : {\\mathcal{B}}_{one}$ where the feature is overlayed on the top box feature,and ${\\mathcal{B}}_{all}$ where the feature is overlayed on all the 36 box features.", + "bbox": [ + 498, + 787, + 890, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rows and columns in the figure correspond to the modalities involved in the trigger inversion optimization and $TrojVQA$", + "bbox": [ + 498, + 869, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "170", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a6410da2a585b02858be21a991b0fcebc8584a447ae6c01201a5742b69fe1e89.jpg", + "image_caption": [ + "Figure 5. Visualizes the generated image patches from $\\tilde{f}_{adv}$ using the trigger patch generation method described in Section 3.4. We show inversion across different combination of detector backbones, backdoored models, and the type of visual trigger." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 831, + 390 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/df7f4e234391a933cb1a40638699874d232169e2d336a085b9e590e270e091da.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SplitTIJOvisTIJOMm
λ=10-5λ=10-3λ=10-5λ=10-3
Tsolid0.97±0.030.97±0.020.91±0.040.89±0.03
Toptim0.96±0.030.96±0.030.89±0.070.90±0.03
Tnlp+S0.58±0.100.59±0.110.93±0.040.92±0.06
Tnlp+O0.47±0.110.47±0.120.87±0.070.87±0.08
T0.58±0.060.59±0.080.92±0.020.91±0.02
", + "bbox": [ + 78, + 445, + 486, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. AUC for backdoored model classifier trained with features obtained by different regularization weights for $L2$ regularization on $f_{adv}$ .", + "bbox": [ + 75, + 561, + 470, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "split respectively. It also shows the performance across different VQA models. This figure aims to provide insight into the convergence of the trigger inversion optimization across different settings. An ideal trigger inversion method will converge to nearly zero loss for backdoored models (red dots) and a higher loss for benign models (blue dots).", + "bbox": [ + 75, + 638, + 468, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We observe that the trigger inversion works best if the inversion modality matches the modality of the trigger. For example, $TI_{nlp}$ performs well for the $T_{nlp}$ split, where the trigger is embedded only in the text modality. Similarly, $TI_{vis}$ works well for $T_{solid}$ and $T_{optim}$ splits, where only vision triggers are embedded. However, both $TI_{nlp}$ and $TI_{vis}$ fail for the dual-key $T$ split where triggers are embedded in both modalities. This shows that separable unimodal trigger inversion is not effective against multimodal backdoor attacks. Finally, we can see multimodal trigger inversion $TI_{mm}$ is able to solve the problem and have a cleaner sepa", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ration between benign and backdoored models in the dualkey split. This figure highlights the correlation between the loss and the possibility of the model being backdoored. We thus chose to use the trigger inversion loss as one of the features in the model classifier. We also observe that $TI_{mm}$ is effective across most VQA models.", + "bbox": [ + 496, + 448, + 890, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We observed the phenomena of 'natural trojans' in multimodal models. Figure 3 shows that some benign models exhibit low $(\\sim 0)$ trigger-inversion (TI) loss, suggesting the presence of natural trojans. Models such as $\\mathrm{BAN}_4$ , $\\mathrm{BAN}_8$ , and $\\mathrm{BUTD}_e$ , are more prone to such natural trojans.", + "bbox": [ + 496, + 541, + 892, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Backdoorsed Model Classification Results", + "text_level": 1, + "bbox": [ + 498, + 632, + 849, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train a logistic regression classifier on the trigger inversion features as mentioned in Section 3.5. Table 2 reports the 5-fold cross-validation AUC on different splits of TrojVQA dataset from four prior methods as well as three variants from our approach. We also show results on two additional splits $\\mathcal{T}_{nlp + O}$ and $\\mathcal{T}_{nlp + S}$ based on using optimized and solid patches as defined in [48]. We clearly see that the unimodal variants of our method- $\\mathrm{TIJO}_{nlp}$ and $\\mathrm{TIJO}_{vis}-$ have almost perfect performance on their corresponding unimodal splits. For example, $\\mathrm{TIJO}_{nlp}$ achieves an AUC of 0.98 on split $\\mathcal{T}_{nlp}$ . However, their performance is low on the multimodal (dual-key) splits. $\\mathrm{TIJO}_{nlp}$ and $\\mathrm{TIJO}_{vis}$ achieve an AUC of 0.46 and 0.67 respectively on split $\\mathcal{T}$ . We also note that $\\mathrm{TIJO}_{vis}$ performs better than $\\mathrm{TIJO}_{nlp}$ on the multimodal splits. This is probably because there is a separation between benign and backdoored mod", + "bbox": [ + 496, + 659, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "171", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "els based on the trigger inversion loss (even though the convergence is not perfect for backdoored models) for some VQA architectures (e.g. $\\mathrm{MCAN}_S$ , $\\mathrm{MCAN}_L$ , $\\mathrm{NAS}_S$ , $\\mathrm{NAS}_L$ ) as evident in Figure 3. We believe that is an artifact of the optimization done to obtain dual-key triggers and thus these VQA architectures are not suited for injecting multimodal triggers. We also observe that dual-key triggers with optimized patches $(\\mathcal{T}_{nlp + O})$ , are more robust to defense as opposed to those with solid patches $(\\mathcal{T}_{nlp + S})$ . For example, the AUC of $\\mathrm{TIJO}_{vis}$ is substantially lower on $\\mathcal{T}_{nlp + O}$ (0.57) as compared to $\\mathcal{T}_{nlp + S}$ (0.70).", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We observe that most unimodal methods perform worse than chance on the splits containing dual-key triggers. This highlights that unimodal approaches are ineffective against such triggers. Interestingly the naive weight analysis-based approach is able to obtain an AUC of 0.6 on split $\\mathcal{T}$ . We finally observe that our approach $\\mathrm{TIJO}_{mm}$ outperforms all other approaches by a significant margin. $\\mathrm{TIJO}_{mm}$ obtains an AUC of 0.92 on split $\\mathcal{T}$ , compared to 0.67, 0.46, 0.60 by $\\mathrm{TIJO}_{vis}$ , $\\mathrm{TIJO}_{nlp}$ , and weight analysis respectively. We also note that $\\mathrm{TIJO}_{mm}$ performs well on all the splits, and thus could be used for modality agnostic trigger inversion.", + "bbox": [ + 75, + 257, + 472, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Experiments:", + "text_level": 1, + "bbox": [ + 76, + 433, + 287, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of classification feature: As discussed in Section 3.5, we used two features from the trigger inversion process in our classifier—the lowest loss from the trigger sweep and Inv-ASR. Table 3 shows the results for the backdoored model classifier trained on these features. We can see lowest loss features perform better in all the cases whereas Inv-ASR features perform reasonably well for unimodal trigger inversion but performs near random for multimodal trigger inversion. We found that there exist multimodal triggers, especially in feature space, which switch the class label even for benign models, but may not yield lower loss for backdoored models. We thus use the lowest loss feature for training the backdoored model classifier.", + "bbox": [ + 75, + 455, + 472, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature overlay: $\\mathcal{B}$ denotes the policy used to plant the feature trigger $f_{adv}$ on the visual inputs. We experiment with two policies: $\\mathcal{B}_{one}$ where optimized feature $f_{adv}$ is overlayed only on the top (based on objectness score) box feature from detector $\\mathcal{D}$ , and $\\mathcal{B}_{all}$ where the feature $f_{adv}$ is overlayed on all the 36 box features. Table 4 reports the results of these experiments. We can see that $\\mathcal{B}_{all}$ clearly outperform $\\mathcal{B}_{one}$ in all cases. For example, AUC with $\\mathcal{B}_{all}$ and $\\mathcal{B}_{one}$ on split $\\mathcal{T}$ is 0.92 and 0.72 respectively. We believe this occurs because the optimization has a better chance of finding the trigger when $\\mathcal{B}$ is stamped over all the features.", + "bbox": [ + 75, + 670, + 472, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of optimization steps and regularization: Figure 4 and Table 5 shows the effect of maximum optimization steps $T$ and regularization on detection performance.", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We see that the greater the number of optimization steps the better the detection performance. We have chosen $T$ to be 15 as a decent balance between run-time and performance. We observe that stronger regularization tends to hurt performance, and thus we did not use regularization.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Image Patch Generation Experiment:", + "text_level": 1, + "bbox": [ + 498, + 184, + 823, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We optimize for $\\pmb{p}_{adv}$ of size $64 \\times 64$ with $\\mathcal{M}$ overlaying the patch to center of the image (as described in Section 3.4). We optimize $\\pmb{p}_{adv}$ with Adam optimizer with a learning rate of 0.03, and betas as (0.5, 0.9) and use early stopping with a patience of 20 epochs. We optimize only over the clean image from the support set $S$ .", + "bbox": [ + 496, + 210, + 890, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 5 shows the generated patches for backdoored MFB VQA models [54]. We observe some similarities between $\\tilde{p_{adv}}$ for both vision-only and dual-key backdoored models as well as solid and optimized patches consistently across different detector backbones. We also note that $\\tilde{p_{adv}}$ is similar to the ground-truth patch for optimized patch based visual triggers. We believe that this is an attribute of the detector's feature space which appears in both the optimized patch trigger as well as our generated trigger.", + "bbox": [ + 496, + 304, + 890, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 462, + 617, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a novel defense technique TIJO (Trigger Inversion using Joint Optimization) to detect multimodal backdoor attacks. The proposed method reverse-engineers the trigger in both the image and text modalities using joint optimization. Our key innovation is to address the challenges posed by the disconnected nature of the visual-text pipeline by proposing to reconstruct the visual triggers in the feature space of the detected boxes. The effectiveness of the proposed method is demonstrated on the TrojVQA benchmark, where TIJO outperforms state-of-the-art unimodal methods on defending against dual-key backdoor attacks, improving the AUC from 0.6 to 0.92 on multimodal dual-key backdoors. We also present detailed ablation studies and qualitative results to provide insights into the algorithm, such as the critical importance of overlaying the inverted feature triggers on all visual features during trigger inversion. Our work is the first defense against multimodal backdoor attacks. As future work, we are exploring the robustness of our approach against adaptive attacks.", + "bbox": [ + 496, + 489, + 890, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This research was partially supported by the U.S. Army Research Laboratory Cooperative Research Agreement W911NF-17-2-0196, and the Intelligence Advanced Research Projects Agency (IARPA) TrojanAI and U.S. Army Research Office under the contract W911NF-20-C-0038. The content of this paper does not necessarily reflect the position or the policy of the Government, and no official endorsement should be inferred.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "172", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 6:14410-14430, 2018.", + "[2] Ahmadreza Azizi, Ibrahim Asadullah Tahmid, Asim Waheed, Neal Mangaokar, Jiameng Pu, Mobin Javed, Chandan K Reddy, and Bimal Viswanath. {T-Miner}: A generative approach to defend against trojan attacks on {DNN-based} text classification. In 30th USENIX Security Symposium (USENIX Security 21), pages 2255-2272, 2021.", + "[3] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A new backdoor attack in cnns by training set corruption without label poisoning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 101-105. IEEE, 2019.", + "[4] Jakub Breier, Xiaolu Hou, Martin Ochoa, and Jesus Solano. Foobar: Fault fooling backdoor attack on neural network training. IEEE Transactions on Dependable and Secure Computing, 2022.", + "[5] Shih-Han Chan, Yinpeng Dong, Jun Zhu, Xiaolu Zhang, and Jun Zhou. Baddet: Backdoor attacks on object detection. arXiv preprint arXiv:2205.14497, 2022.", + "[6] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian Molloy, and Biplav Srivastava. Detecting backdoor attacks on deep neural networks by activation clustering. arXiv preprint arXiv:1811.03728, 2018.", + "[7] Jinyin Chen, Chengyu Jia, Haibin Zheng, Ruoxi Chen, and Chenbo Fu. Is multi-modal necessarily better? robustness evaluation of multi-modal fake news detection. arXiv preprint arXiv:2206.08788, 2022.", + "[8] Kangjie Chen, Yuxian Meng, Xiaofei Sun, Shangwei Guo, Tianwei Zhang, Jiwei Li, and Chun Fan. Badpre: Task-agnostic backdoor attacks to pre-trained NLP foundation models. In International Conference on Learning Representations, 2022.", + "[9] Weixin Chen, Baoyuan Wu, and Haoqian Wang. Effective backdoor defense by exploiting sensitivity of poisoned samples. In Advances in Neural Information Processing Systems, 2022.", + "[10] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017.", + "[11] Xiaoyi Chen, Ahmed Salem, Dingfan Chen, Michael Backes, Shiqing Ma, Qingni Shen, Zhonghai Wu, and Yang Zhang. Badnl: Backdoor attacks against nlp models with semantic-preserving improvements. In Annual Computer Security Applications Conference, pages 554-569, 2021.", + "[12] Jiazhu Dai, Chuanshuai Chen, and Yufeng Li. A backdoor attack against lstm-based text classification systems. IEEE Access, 7:138872-138878, 2019.", + "[13] Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. Hotflip: White-box adversarial examples for text classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 31-36, 2018." + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Ming Fan, Ziliang Si, Xiaofei Xie, Yang Liu, and Ting Liu. Text backdoor detection using an interpretable rnn abstract model. IEEE Transactions on Information Forensics and Security, 16:4117-4132, 2021.", + "[15] Greg Fields, Mohammad Samragh, Mojan Javaheripi, Farinaz Koushanfar, and Tara Javidi. Trojan signatures in cnn weights. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-20, 2021.", + "[16] Lianli Gao, Qilong Zhang, Jingkuan Song, Xianglong Liu, and Heng Tao Shen. Patch-wise attack for fooling deep neural network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16, pages 307-322. Springer, 2020.", + "[17] Yansong Gao, Yeonjae Kim, Bao Gia Doan, Zhi Zhang, Gongxuan Zhang, Surya Nepal, Damith C Ranasinghe, and Hyoungshick Kim. Design and evaluation of a multidomain trojan detection method on deep neural networks. IEEE Transactions on Dependable and Secure Computing, 19(4):2349-2364, 2021.", + "[18] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 113-125, 2019.", + "[19] Diego Garcia-soto, Huili Chen, and Farinaz Koushanfar. Perd: Perturbation sensitivity-based neural trojan detection framework on nlp applications. arXiv preprint arXiv:2208.04943, 2022.", + "[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017.", + "[21] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017.", + "[22] Xiaoling Hu, Xiao Lin, Michael Cogswell, Yi Yao, Susmit Jha, and Chao Chen. Trigger hunting with a topological prior for trojan detection. In International Conference on Learning Representations, 2022.", + "[23] Xijie Huang, Moustafa Alzantot, and Mani Srivastava. Neuroninspect: Detecting backdoors in neural networks via output explanations. arXiv preprint arXiv:1911.07399, 2019.", + "[24] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019.", + "[25] Lesheng Jin, Zihan Wang, and Jingbo Shang. Wedef: Weakly supervised backdoor defense for text classification. arXiv preprint arXiv:2205.11803, 2022.", + "[26] Panagiota Kiourti, Wenchao Li, Anirban Roy, Karan Sikka, and Susmit Jha. Misa: Online defense of trojaned models using misattributions. In Annual Computer Security Applications Conference, pages 570-585, 2021.", + "[27] Panagiota Kiourti, Kacper Wardega, Susmit Jha, and Wenchao Li. Trojdrl: evaluation of backdoor attacks on deep reinforcement learning. In 2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2020." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "173", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Hyun Kwon. Defending deep neural networks against backdoor attack by using de-trigger autoencoder. IEEE Access, 2021.", + "[29] Yuanchun Li, Jiayi Hua, Haoyu Wang, Chunyang Chen, and Yunxin Liu. Deeppayload: Black-box backdoor attack on deep learning models through neural payload injection. In 2021 IEEE/ACM 43rd International Conference on Software Engineering (ICSE), pages 263-274. IEEE, 2021.", + "[30] Yiming Li, Yong Jiang, Zhifeng Li, and Shu-Tao Xia. Backdoor learning: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022.", + "[31] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems, 34:14900-14912, 2021.", + "[32] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing back-door triggers from deep neural networks. In International Conference on Learning Representations, 2021.", + "[33] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273-294. Springer, 2018.", + "[34] Yingqi Liu, Guangyu Shen, Guanhong Tao, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Piccolo: Exposing complex backdoors in nlp transformer models. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1561-1561. IEEE Computer Society, 2022.", + "[35] Hua Ma, Yinshan Li, Yansong Gao, Alsharif Abuadbba, Zhi Zhang, Anmin Fu, Hyoungshick Kim, Said F Al-Sarawi, Nepal Surya, and Derek Abbott. Dangerous cloaking: Natural trigger based backdoor attacks on object detectors in the physical world. arXiv preprint arXiv:2201.08619, 2022.", + "[36] Hua Ma, Yinshan Li, Yansong Gao, Zhi Zhang, Alsharif Abuadbba, Anmin Fu, Said F Al-Sarawi, Nepal Surya, and Derek Abbott. Macab: Model-agnostic clean-annotation backdoor to object detection with natural trigger in real-world. arXiv preprint arXiv:2209.02339, 2022.", + "[37] Tuan Anh Nguyen and Anh Tuan Tran. Wanet - imperceptible warping-based backdoor attack. In International Conference on Learning Representations, 2021.", + "[38] Fanchao Qi, Yangyi Chen, Mukai Li, Yuan Yao, Zhiyuan Liu, and Maosong Sun. Onion: A simple and effective defense against textual backdoor attacks. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9558-9566, 2021.", + "[39] Han Qiu, Yi Zeng, Shangwei Guo, Tianwei Zhang, Meikang Qiu, and Bhavani Thuraisingham. Deepsweep: An evaluation framework for mitigating dnn backdoor attacks using data augmentation. In Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security, pages 363-377, 2021.", + "[40] Adnan Siraj Rakin, Zhezhi He, and Deliang Fan. Tbt: Targeted neural network attack with bit Trojan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13198-13207, 2020." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11957-11965, 2020.", + "[42] Aniruddha Saha, Ajinkya Tejankar, Soroush Abbasi Koohpayegani, and Hamed Pirsivash. Backdoor attacks on self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13337-13346, 2022.", + "[43] Kun Shao, Junan Yang, Yang Ai, Hui Liu, and Yu Zhang. Bddr: An effective defense against textual backdoor attacks. Computers & Security, 110:102433, 2021.", + "[44] Guangyu Shen, Yingqi Liu, Guanhong Tao, Qiuling Xu, Zhuo Zhang, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Constrained optimization with dynamic bound-scaling for effective nlp backdoor defense. In International Conference on Machine Learning, pages 19879-19892. PMLR, 2022.", + "[45] Karan Sikka, Indranil Sur, Susmit Jha, Anirban Roy, and Ajay Divakaran. Detecting trojaned dnns using counterfactual attributions. arXiv preprint arXiv:2012.02275, 2020.", + "[46] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019.", + "[47] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing NLP. In Empirical Methods in Natural Language Processing, 2019.", + "[48] Matthew Walmer, Karan Sikka, Indranil Sur, Abhinav Shrivastava, and Susmit Jha. Dual-key multimodal backdoors for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15375-15385, 2022.", + "[49] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE Symposium on Security and Privacy (SP), pages 707-723. IEEE, 2019.", + "[50] Wenqi Wang, Run Wang, Lina Wang, Zhibo Wang, and Aoshuang Ye. Towards a robust deep neural network in texts: A survey. arXiv preprint arXiv:1902.07285, 2019.", + "[51] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.", + "[52] Mingfu Xue, Shifeng Ni, Yinghao Wu, Yushu Zhang, Jian Wang, and Weiqiang Liu. Imperceptible and multi-channel backdoor attack against deep neural networks. arXiv preprint arXiv:2201.13164, 2022.", + "[53] Wenkai Yang, Yankai Lin, Peng Li, Jie Zhou, and Xu Sun. Rap: Robustness-aware perturbations for defending against backdoor attacks on nlp models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8365-8381, 2021.", + "[54] Zhou Yu, Jun Yu, Jianping Fan, and Dacheng Tao. Multimodal factorized bilinear pooling with co-attention learning for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 1821–1830, 2017." + ], + "bbox": [ + 501, + 92, + 892, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "174", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Xiaoyu Zhang, Rohit Gupta, Ajmal Mian, Nazanin Rahnavard, and Mubarak Shah. Cassandra: Detecting trojaned networks from adversarial perturbations. IEEE Access, 9:135856-135867, 2021.", + "[56] Biru Zhu, Yujia Qin, Ganqu Cui, Yangyi Chen, Weilin Zhao, Chong Fu, Yangdong Deng, Zhiyuan Liu, Jingang Wang, Wei Wu, et al. Moderate-fitting as a natural backdoor defender for pre-trained language models. In Advances in Neural Information Processing Systems, 2022." + ], + "bbox": [ + 78, + 90, + 468, + 218 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "175", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/7bfffc83-27ef-4c73-aa19-9a15a6ee13fd_model.json b/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/7bfffc83-27ef-4c73-aa19-9a15a6ee13fd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3501847dda66347bc067cc6dd84be21e9420543 --- /dev/null +++ b/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/7bfffc83-27ef-4c73-aa19-9a15a6ee13fd_model.json @@ -0,0 +1,1993 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.868, + 0.175 + ], + "angle": 0, + "content": "TIJO: Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.203, + 0.777, + 0.265 + ], + "angle": 0, + "content": "Indranil Sur\\(^{1*}\\) Karan Sikka\\(^{1}\\) Matthew Walmer\\(^{2}\\) Kaushik Koneripalli\\(^{1}\\) Anirban Roy\\(^{1}\\) Xiao Lin\\(^{1}\\) Ajay Divakaran\\(^{1}\\) Susmit Jha\\(^{1}\\) SRI International \\(^{2}\\)University of Maryland" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.299, + 0.314, + 0.315 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.474, + 0.741 + ], + "angle": 0, + "content": "We present a Multimodal Backdoor Defense technique TIJO (Trigger Inversion using Joint Optimization). Recent work [48] has demonstrated successful backdoor attacks on multimodal models for the Visual Question Answering task. Their dual-key backdoor trigger is split across two modalities (image and text), such that the backdoor is activated if and only if the trigger is present in both modalities. We propose TIJO that defends against dual-key attacks through a joint optimization that reverse-engineers the trigger in both the image and text modalities. This joint optimization is challenging in multimodal models due to the disconnected nature of the visual pipeline which consists of an offline feature extractor, whose output is then fused with the text using a fusion module. The key insight enabling the joint optimization in TIJO is that the trigger inversion needs to be carried out in the object detection box feature space as opposed to the pixel space. We demonstrate the effectiveness of our method on the TrojVQA benchmark, where TIJO improves upon the state-of-the-art unimodal methods from an AUC of 0.6 to 0.92 on multimodal dual-key backdoors. Furthermore, our method also improves upon the unimodal baselines on unimodal backdoors. We present ablation studies and qualitative results to provide insights into our algorithm such as the critical importance of overlaying the inverted feature triggers on all visual features during trigger inversion. The prototype implementation of TIJO is available at https://github.com/SRI-CSL/TIJO." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.772, + 0.21, + 0.788 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.798, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Deep Neural Networks (DNNs) are vulnerable to adversarial attacks [47, 1, 16, 24]. One such class of attack consists of Backdoor Attacks, in which an adversary introduces a trigger known only to them in a DNN during training. Such a backdoored DNN will behave normally with typi" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.299, + 0.87, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.532, + 0.895, + 0.63 + ], + "angle": 0, + "content": "Figure 1. (Top) A dual-key backdoor attack for multimodal models [48], which is designed to activate if and only if the trigger is present in both the modalities. Such backdoors cannot be detected by unimodal defenses. (Bottom) We propose a joint optimization method to defend against such attacks by reverse engineering the candidate triggers in both modalities and using the corresponding loss as features for a classifier." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.893, + 0.701 + ], + "angle": 0, + "content": "cal in-distribution inputs but perform poorly (e.g. produce targeted misclassifications) on inputs stamped with a predefined trigger designed by the adversary [50, 21, 30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.893, + 0.839 + ], + "angle": 0, + "content": "Recent work [48, 7] has introduced backdoors in multimodal domains such as Visual Question Answering (VQA) and Fake News Detection [7, 48]. In prior work [48], we have introduced a Dual-Key Backdoor Attack (shown in Figure 1), where the trigger is inserted in both the image and text modalities in such a manner that the backdoor is activated only when both modalities contain the trigger. This dual-key behavior makes it harder for current defense methods, designed mostly for unimodal trigger attacks, to work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "There has been significant work developing defenses against backdoor attacks in the visual domain, in particular for the image classification task [45, 49, 6, 23]. Recent works have also explored defense in natural language pro" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.339, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author: indranil.sur@sri.com" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "165" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.363 + ], + "angle": 0, + "content": "cessing domains [38, 43, 34]. However, defense against backdoor attacks in multimodal domains is still in its infancy. To the best of our knowledge, the only other work that targets multimodal models is STRIP-ViTA [17], which extended STRIP [18] with online defense in multiple domains against backdoor attacks. Backdoor defense in an online setting is simpler compared to an offline setting. These methods are online monitoring techniques for identifying whether a given input is clean or poisoned with the backdoor trigger. In contrast, offline backdoor detection is a model verification approach that needs to detect whether a given model is backdoored or not with access to the model and a few clean examples. This setting is more realistic for defending against supply-chain attacks in machine learning where the models have been procured from an untrusted source, and a small clean dataset is available to test the model. We focus on multimodal defense in such an offline setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.695 + ], + "angle": 0, + "content": "In this work, we propose a novel approach for defending against multimodal backdoor attacks, referred to as Trojan Inversion using Joint Optimization (TIJO), that reverse engineers the triggers in both modalities. Our approach is motivated by the Universal Adversarial Trigger (UAT) [47] that was proposed to identify naturally occurring universal triggers in pre-trained NLP models and has been extended in earlier works to identify Trojan triggers in NLP models. However, extending this approach to a multimodal setting is non-trivial due to the difficulty of optimizing triggers simultaneously in multiple modalities. Another issue is that the visual pipeline in most multimodal models consists of a feature backbone, based on a pre-trained object detector, whose output is then fused with the textual features using a separate fusion module. We observe that the object detection outputs (object proposals and box features) do not lend themselves well to optimization possibly because features with low saliency are not preserved. Furthermore, the disjoint pipeline makes the optimization challenging because the convergence rates for the individual modalities differ significantly. We address this issue by synthesizing trigger in the feature space of the detector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.472, + 0.801 + ], + "angle": 0, + "content": "We evaluate TIJO on the TrojVQA dataset [48] that consists of over 800 VQA models spanning across 4 feature backbones and 10 model architectures. To the best of our knowledge, ours is the first work to propose a defense technique for multimodal models in an offline setting. Our results indicate strong improvement over prior unimodal methods. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.47, + 0.839 + ], + "angle": 0, + "content": "- We present a novel approach for Multimodal Backdoor defense referred to as TIJO." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.841, + 0.47, + 0.884 + ], + "angle": 0, + "content": "- We develop a novel trigger inversion process in object detection box feature space as well as textual space that enables joint optimization of multimodal triggers." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.9 + ], + "angle": 0, + "content": "- We demonstrate TIJO on the TrojVQA dataset and show" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.811, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "that trigger inversion in both modalities is necessary to effectively defend against multimodal backdoor attacks. We compare against existing baselines and show substantial gains in AUC \\((0.6\\to 0.92)\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.153, + 0.892, + 0.211 + ], + "angle": 0, + "content": "- We show that TIJO improves upon our selected set of state-of-the-art unimodal methods in the detection of unimodal backdoors indicating that our proposed method is modality-agnostic." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.213, + 0.892, + 0.287 + ], + "angle": 0, + "content": "- We uncover several insights with ablation studies such as (1) increasing the number of optimization steps improves the backdoor detection performance, and (2) the feature trigger needs to be overlaid on all the visual features for the best results." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.153, + 0.892, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.311, + 0.642, + 0.327 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.563 + ], + "angle": 0, + "content": "Backdoor Attacks: Backdoor attacks are a type of targeted adversarial attack that were first introduced in [21]. Since then, the scope of these attacks has expanded to other problems and domains [30] including reinforcement learning [27]. Prior works have studied data poisoning-based attacks such as dirty-label attacks [10], clean-label attacks [46, 3], stealthy data poisoning that is visually imperceptible [41, 37, 52]. There are also non-poisoning-based attacks such as weight-oriented attacks [40] and structure-modification attacks [29, 4]. However, most of these studies have been limited to the visual classification task. Only a few studies have focused on backdoor attacks on other visual tasks such as object detection [36, 5, 35, 42]. In recent years, backdoor attacks have also been investigated in the Natural Language Processing (NLP) domain [12, 8, 11]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Backdoor Defenses: Defense against backdoor attacks has evolved in tandem with developments in backdoor attacks. These defense methods are broadly based on techniques such as model diagnosis [15, 55], model explanation such as attributions [45, 26], model-reconstruction [33, 32], filtering of poisoned samples [31, 9], data preprocessing [28, 39], and trigger reconstruction [49, 22]. Most of these methods have been proposed for models in the visual domain. There have been some recent works on backdoor defense in the NLP domain. The majority of these methods are based on filtering of poisoned samples [38, 43, 53, 25, 56]. Other works rely on ideas such as model diagnosis [14, 19], prepossessing-based [2], and trigger synthesis [34, 44]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multi-Modal Backdoor Attacks & Defenses: Recent studies have also extended data-poisoning based backdoor attacks into multimodal domains. Chen et al. [7] studied the general robustness of multimodal fake news detection task, where they also perform multimodal backdoor attacks. Walmer et al. [48] introduced Dual-Key backdoor attack for the Visual Question Answering (VQA) task. As shown in" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "166" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.089, + 0.808, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.383, + 0.892, + 0.455 + ], + "angle": 0, + "content": "Figure 2. Shows key blocks of TIJO. (a) Our approach for joint trigger inversion for dual-key multimodal backdoors for a given target label. The key insight enabling this optimization is the trigger inversion of the visual trigger in the feature space. (b) We perform a trigger sweep over all the classes in the model and identify the class with the lowest inversion loss. (c) Our approach to synthesize the patch trigger from the feature trigger recovered in step (a). (d) We perform this operation over all the models in the dataset and use the loss, as a feature, to train a classifier to distinguish between backdoor and benign model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.469, + 0.523 + ], + "angle": 0, + "content": "Figure 1, this attack was designed to trigger the backdoor only when the trigger is present in both modalities, which makes the attack stealthier compared to a unimodal trigger." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.469, + 0.66 + ], + "angle": 0, + "content": "Defense against multimodal backdoor attacks is limited in comparison to unimodal attacks in the vision and NLP domains. Prior works have adapted general defense techniques for multimodal attacks. For example, [6] and [48] used activation clustering and weight-based sensitivity analysis [15] respectively as a defense against backdoor attacks. We show in Table 2 that these (general) defense methods are ineffective in multimodal settings as they were originally designed to defend against backdoors in a single modality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.469, + 0.812 + ], + "angle": 0, + "content": "Gao et al. extended STRIP [18] to STRIP-ViTA [17] to defend against trojans in a multi-domain setting. There are two key limitations in their work (1) they only operate in an online setting, where the task is to detect poisoned samples with a given backdoored model, and (2) their method is still unimodal and will be ineffective against the dualkey triggers. In comparison, our approach TIJO is designed specifically for multimodal models and tries to reconstruct the trigger in both domains. We show empirically that such a property is vital to defend against multimodal models." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.829, + 0.185, + 0.846 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We first discuss the threat model that we aim to defend against, then discuss the UAT method [47] and its extension to multimodal models, and present our method, TIJO." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.477, + 0.642, + 0.492 + ], + "angle": 0, + "content": "3.1. Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Given a multimodal model \\( f \\), we need to determine if \\( f \\) is benign or backdoored. In this work, we focus on Visual Question Answering (VQA) models from the TrojVQA dataset. Let \\( \\mathcal{C} \\) be the clean VQAv2 dataset [20] where each data entry is a triplet \\( (\\pmb{x},\\pmb{t},\\pmb{y}) \\) where \\( \\pmb{x} \\) is the image, \\( \\pmb{t} \\) is the tokenized question, and \\( y \\) is the answer label. Most VQA models use a two-step process for generating the answer. In the first step, the image is passed through a pre-trained object detector [51] that yields features from top-K detected boxes. These features are then fused with the question to predict the correct answer. Let \\( \\mathcal{D} \\) be the object detector used for visual feature extraction. The answer is generated using \\( f(t,\\mathcal{D}(\\pmb{x})) = y \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.868 + ], + "angle": 0, + "content": "In our threat model, we assume that \\(\\mathcal{D}\\) is benign and the adversary introduces the backdoor in the VQA model \\(f\\). This is also the threat model used in the TrojVQA dataset [48]. For a backdoored VQA model \\(f_{b}\\), the adversary introduces triggers \\(\\pmb{p}_{t}\\) and \\(\\pmb{t}_{t}\\) in both the image and text modalities respectively. \\(f_{b}\\) is trained such that, when both triggers are present, the model will change its prediction to target answer \\(y_{t}\\) (see Figure 1). In the TrojVQA dataset, \\(\\pmb{p}_{t}\\) are small visual patches while \\(\\pmb{t}_{t}\\) are natural words. The triggers and the model behavior are only known to the adversary." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Let \\(\\mathcal{M}\\) be a policy that overlaps \\(\\pmb{p}_t\\) on \\(\\pmb{x}\\) and \\(\\mathcal{A}\\) be a policy that appends \\(\\pmb{t}_t\\) to \\(\\pmb{t}\\). Hence, for a backdoored VQA" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "167" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.245, + 0.106 + ], + "angle": 0, + "content": "model \\(f_{b}\\) , we expect that" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.118, + 0.382, + 0.134 + ], + "angle": 0, + "content": "\\[\nf _ {b} (\\mathcal {A} (\\boldsymbol {t}, \\boldsymbol {t} _ {t}), \\mathcal {D} (\\mathcal {M} (\\boldsymbol {x}, \\boldsymbol {p} _ {t}))) = y _ {t}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.145, + 0.47, + 0.191 + ], + "angle": 0, + "content": "In this work, we focus on dual-key triggers [48], where the model changes its prediction only when both \\(\\mathcal{M}\\) and \\(\\mathcal{A}\\) are applied together." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.2, + 0.334, + 0.216 + ], + "angle": 0, + "content": "3.2. Trigger Inversion using UAT" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.223, + 0.469, + 0.388 + ], + "angle": 0, + "content": "TIJO is based on Universal Adversarial Triggers (UAT) [47], which extends Hotflip [13] from synthesizing adversarial tokens for a single input to all inputs in the dataset. As a result, obtained adversarial tokens are universal in nature. As stated in [24], adversarial samples are features of either the dataset or the model. Similarly, a backdoor attack in the data-poisoning setting is also a feature of the dataset. Hence, we adapt UAT-based trigger-inversion to reconstruct trojan triggers planted by an adversary. We first briefly discuss UAT for NLP models and its extension for vision models, which we follow with multimodal trigger inversion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.47, + 0.541 + ], + "angle": 0, + "content": "Eq. 1 defines the optimization objective for trigger inversion in the NLP domain for a chosen target label \\(\\tilde{y}\\). Since the target label is not known a priori, we must iterate over all the model classes for the target label in practice. Here \\(\\mathcal{L}\\) is the cross-entropy loss, and we optimize to minimize the expected loss over all samples in \\(S\\). In summary, we optimize to get the \\(t_{adv}\\) that maximizes the likelihood of switching the class label to \\(\\tilde{y}\\) for all samples in \\(S\\). Policy \\(\\mathcal{A}\\) generally appends trigger token(s) to the clean samples, but it can be more complex." + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.551, + 0.469, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {t} _ {a d v}} \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {x} \\sim \\mathcal {S}} \\left[ \\mathcal {L} \\left(\\tilde {y}, f (\\mathcal {A} \\left(\\boldsymbol {t} _ {a d v}, \\boldsymbol {t}\\right), \\mathcal {D} (\\boldsymbol {x}))\\right) \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.469, + 0.704 + ], + "angle": 0, + "content": "Since the space of \\( t_{adv} \\) is discrete, each optimization step is followed by a next token selection step. The next token is set by \\( t_{adv} \\gets t_i \\) which minimizes the trigger inversion loss's first-order Taylor approximation around the current token embedding as given by Eq. 2. Here \\( \\mathcal{V}_f \\) is the vocabulary of all tokens in \\( f \\), function \\( \\mathcal{E}_f \\) gives the token embeddings and \\( \\nabla_{\\mathcal{E}_f(t_{adv})}\\mathcal{L} \\) is the average gradient of the loss over a batch." + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.715, + 0.469, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {t} _ {i} \\in \\mathcal {V} _ {f}} \\left[ \\mathcal {E} _ {f} \\left(\\boldsymbol {t} _ {i}\\right) - \\mathcal {E} _ {f} \\left(\\boldsymbol {t} _ {a d v}\\right) \\right] ^ {\\intercal} \\nabla_ {\\mathcal {E} _ {f} \\left(\\boldsymbol {t} _ {a d v}\\right)} \\mathcal {L} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The above optimization problem is solved efficiently by computing dot products between the gradient and the \\(\\nu_{f}\\) embeddings and then using nearest neighbor or beam search to get the updated token \\(t_i\\) [47]. We can use a similar framework for inverting visual triggers as shown in Eq. 3. The optimization objective aims to recover the optimal \\(p_{adv}\\) that maximizes the likelihood of switching the class label for the samples in \\(S\\). The only difference is that we use projected gradient descent for patch \\(p_{adv}\\), overlaid on \\(x\\) through policy \\(\\mathcal{M}\\), which needs to obey image constraints. This ap" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.122 + ], + "angle": 0, + "content": "proach is similar to prior trigger reconstruction-based methods such as Neural Cleansse [49]." + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.129, + 0.892, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {p} _ {a d v}} {\\min } \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {x} \\sim \\mathcal {S}} \\left[ \\mathcal {L} (\\tilde {y}, f (\\boldsymbol {t}, \\mathcal {D} (\\mathcal {M} (\\boldsymbol {x}, \\boldsymbol {p} _ {a d v})))) \\right] \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.166, + 0.851, + 0.182 + ], + "angle": 0, + "content": "3.3. Multimodal Trigger Inversion with TIJO" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.189, + 0.892, + 0.326 + ], + "angle": 0, + "content": "We now outline our approach for multimodal Trigger Inversion using Joint Optimization (TIJO) (shown in Figure 2). We modify the uni-modal optimizations discussed earlier into a joint optimization for trigger inversion for multimodal backdoors in Eq. 4. Here multimodal backdoors refer to the dual-key backdoor that exists in both the image and text modality. We optimize for both \\( \\pmb{t}_{adv} \\) and \\( \\pmb{p}_{adv} \\) to maximize the likelihood of switching the class label to \\( \\tilde{y} \\) for all samples in \\( S \\)." + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.332, + 0.892, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {t} _ {a d v}, \\boldsymbol {p} _ {a d v}} \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {x} \\sim \\mathcal {S}} \\left[ \\mathcal {L} \\left(\\tilde {y}, f \\left(\\mathcal {A} \\left(\\boldsymbol {t} _ {a d v}, \\boldsymbol {t}\\right), \\mathcal {D} \\left(\\mathcal {M} \\left(\\boldsymbol {p} _ {a d v}, \\boldsymbol {x}\\right)\\right)\\right)\\right) \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.893, + 0.595 + ], + "angle": 0, + "content": "Solving Eq. 4 for multimodal (dual-key) backdoors is challenging. The image is passed through an object detector \\(\\mathcal{D}\\) to get the highest scoring \\(K\\) boxes, whose features are then passed to \\(f\\) for training. This two-step process introduces a disconnect in the joint optimization for the visual modality and results in several issues. For example, when we stamp the patch on the image during optimization, the detector \\(\\mathcal{D}\\) may not propose bounding boxes containing the patch \\(\\pmb{p}_{adv}\\). One solution would be to manually force the detector to sample a proposal around \\(\\pmb{p}_{adv}\\). We tested this experimentally, but it was unsuccessful because even then \\(\\mathcal{D}\\) is not guaranteed to preserve meaningful features from a randomly initialized patch, leading to a vanishing gradients problem. Another challenge that makes this optimization hard is that the support set \\(\\mathcal{S}\\) contains only a few samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.893, + 0.853 + ], + "angle": 0, + "content": "Proposed key idea: We propose to overcome this issue and enable the convergence for both the visual and textual trigger by performing trigger inversion for the visual triggers in the feature space of \\(\\mathcal{D}\\), while the textual trigger is optimized in the token space as done for UAT. We define \\(f_{adv}\\) as the additive adversarial feature space signature and \\(\\mathcal{B}\\) as the overlay policy by which we overlay \\(f_{adv}\\) on box features from \\(\\mathcal{D}\\). The modified optimization objective is shown in Eq. 5, where we optimize \\(t_{adv}\\) and \\(f_{adv}\\) instead of \\(p_{adv}\\). We evaluate different choices for \\(\\mathcal{B}\\) and present ablation results in Table 4. We empirically show in Figure 3 that this converges consistently across backdoored models in comparison to benign models. We have shown a detailed description of our approach in Figure 2. Similar to UAT, we optimize Eq. 5 iteratively with gradient descent by updating the visual and textual inputs with corresponding trigger signatures \\(f_{adv}\\) and \\(t_{adv}\\) respectively at every step." + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.864, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {t} _ {a d v}, \\boldsymbol {f} _ {a d v}} \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {x} \\sim \\mathcal {S}} \\left[ \\mathcal {L} \\left(\\tilde {y}, f \\left(\\mathcal {A} \\left(\\boldsymbol {t} _ {a d v}, \\boldsymbol {t}\\right), \\mathcal {B} \\left(\\mathcal {D} (\\boldsymbol {x}), \\boldsymbol {f} _ {a d v}\\right)\\right)\\right) \\right] \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "168" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.089, + 0.833, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.37, + 0.895, + 0.429 + ], + "angle": 0, + "content": "Figure 3. Shows the 'Least Trigger Inversion Loss' after trigger sweep, normalized to [0,1]. The blue and red dots are benign and backdoor models respectively. Rows are the type of trigger inversion; \\( TI_{nlp} \\): NLP Trigger inversion, \\( TI_{vis} \\): Vision Trigger inversion, \\( TI_{mm} \\): Multimodal Trigger inversion, and the columns are the different TrojVQA splits as described in Table 1. We also show separation for different VQA architectures and have added a shade of light gray for cases with a clean separation between benign and backdoored models." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.447, + 0.468, + 0.55 + ], + "angle": 0, + "content": "
SplitNLPVisualTrain/TestTrigger Type
Tnlp✓X160/80Single Key NLP
TsolidXSolid160/80Single Key Vision
ToptimXOptimized160/80Single Key Vision
Tnlp+S✓Solid160/80Dual Key
Tnlp+O✓Optimized160/80Dual Key
T✓✓320/160Dual Key
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.552, + 0.438, + 0.566 + ], + "angle": 0, + "content": "Table 1. Details about the TrojVQAdataset [48] and its splits." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.602, + 0.312, + 0.619 + ], + "angle": 0, + "content": "3.4. Trigger Patch Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.473, + 0.826 + ], + "angle": 0, + "content": "We also propose to recover the patch trigger \\( \\pmb{p}_{adv} \\) based on the \\( \\pmb{f}_{adv} \\) obtained using Eq. 5 (see Figure 2). We first compute the box proposals \\( \\pmb{b}_x \\gets \\mathcal{D}_{rpn}(\\mathcal{D}_{cnn}(\\pmb{x})) \\) and box features \\( \\pmb{f}_x \\gets \\mathcal{D}_{roi}(\\mathcal{D}_{cnn}(\\pmb{x}),\\pmb{b}_x) \\) on the clean image \\( \\pmb{x} \\). We also compute the box features \\( \\pmb{f}_{x_p} \\gets \\mathcal{D}_{roi}(\\mathcal{D}_{cnn}(\\mathcal{M}(\\pmb{x},\\pmb{p}_{adv})),\\pmb{b}_x) \\) on the image stamped with \\( \\pmb{p}_{adv} \\). Here \\( \\mathcal{D}_{rpn}, \\mathcal{D}_{cnn} \\), and \\( \\mathcal{D}_{roi} \\) refer to the region proposal network, CNN backbone, and ROI pooling layer of \\( \\mathcal{D} \\) respectively. We overlay \\( \\pmb{f}_{adv} \\) on \\( \\pmb{f}_x \\) and then iteratively optimize \\( \\pmb{p}_{adv} \\) to minimize the MSE loss between \\( \\pmb{f}_{x_p} \\) and \\( \\mathcal{B}(\\pmb{f}_x,\\tilde{\\pmb{f}}_{adv}) \\). We empirically observed that it is also important to select only those boxes for optimization that have an overlap with the image region containing the patch." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.833, + 0.367, + 0.849 + ], + "angle": 0, + "content": "3.5. Backdoored Model Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The optimization objective should ideally converge only if the model is backdoored and if the target label \\(\\tilde{y}\\) is actually the poison label \\(y_{t}\\). We use this convergence property" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.895, + 0.663 + ], + "angle": 0, + "content": "to train a classifier to separate backdoored and benign models. Since the poison label \\( y_{t} \\) is unknown, we sweep over all the label space, \\( \\forall \\tilde{y} \\in \\mathcal{V} \\) and repeat the trigger inversion process for each \\( \\tilde{y} \\) (referred to as trigger sweep). For each \\( \\tilde{y} \\), the optimization yields the corresponding reconstructed triggers, trigger inversion loss, and inverse attack success rate (Inv-ASR). Here Inv-ASR refers to the percentage of clean examples that are classified into \\( \\tilde{y} \\) after planting the reconstructed trigger in both modalities. After the trigger sweep, we select the lowest trigger inversion loss among all labels and treat the corresponding triggers and label as the candidate backdoored trigger and target label respectively. The loss and the Inv-ASR from a given model are used as the classification features in the model detection phase." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.895, + 0.725 + ], + "angle": 0, + "content": "We first obtain the classification features for all the models in the dataset. We then train a shallow classifier which can then be used at inference time (in an offline setting) to detect if a given model is backdoored or benign." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.739, + 0.634, + 0.757 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We evaluate our approach in this section. We first discuss the dataset and metrics used for evaluation. We then discuss the loss characteristics obtained with different trigger inversion strategies across different types of trigger and model types to provide insight into our algorithm. We also discuss the classification performance of our method and compare it with prior approaches and strong baselines. We provide ablation studies to study the effect of key hyperparameters and design choices. Finally, we provide visualizations of" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "169" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.872, + 0.226 + ], + "angle": 0, + "content": "
SplitGeneral Wt. AnalysisDBSUnimodal NCTABORTIJO \\( _{nlp} \\)Ours TIJO \\( _{vis} \\)TIJO \\( _{mm} \\)
Single Key\\( T_{nlp} \\)0.61±0.070.89±0.05--0.98±0.020.52±0.060.98±0.02
\\( T_{solid} \\)0.53±0.05-0.59±0.100.98±0.020.39±0.091.00±0.000.99±0.01
\\( T_{optim} \\)0.58±0.05-0.71±0.080.99±0.020.40±0.110.99±0.010.95±0.03
Dual Key\\( T_{nlp+S} \\)0.54±0.030.46±0.040.42±0.050.46±0.060.41±0.110.70±0.060.97±0.03
\\( T_{nlp+O} \\)0.60±0.130.45±0.010.50±0.090.52±0.030.43±0.120.57±0.070.86±0.10
\\( T \\)0.60±0.040.48±0.020.50±0.060.48±0.040.46±0.030.67±0.070.92±0.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.228, + 0.894, + 0.286 + ], + "angle": 0, + "content": "Table 2. Shows AUC for different \\(TrojVQA\\) splits with weight analysis, prior unimodal methods as well as three variants of our method- \\(\\mathrm{TIJO}_{\\mathrm{np}}\\), \\(\\mathrm{TIJO}_{\\mathrm{vis}}\\), and \\(\\mathrm{TIJO}_{\\mathrm{mm}}\\) which optimize triggers in NLP, vision, and both modalities respectively. We see a clear improvement with \\(\\mathrm{TIJO}_{\\mathrm{mm}}\\) for not only dual-key multimodal triggers but also for unimodal triggers. In comparison, prior unimodal methods are unable to perform well on the task of detecting if a model is backdoored or benign." + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.301, + 0.805, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.445, + 0.729, + 0.46 + ], + "angle": 0, + "content": "Figure 4. Shows the effect of the max optimization step on detection performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.471, + 0.515 + ], + "angle": 0, + "content": "the reconstructed visual patches using our algorithm (refer to the supplementary materials for implementation details)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.822 + ], + "angle": 0, + "content": "TrojVQA Dataset and Metric: We use the TrojVQA [48] dataset that was introduced recently and consists of both benign and poisoned VQA models. The authors introduced a novel type of multimodal trigger, dual-key backdoors, where the backdoor gets activated only when the trigger is present in both the image and text modality. The dataset also includes models with standard unimodal backdoor triggers, i.e. the trigger was introduced in either the text or image modality only. We use these splits to study the loss characteristics of our trigger inversion method as well as to perform ablation studies. We have provided details regarding the splits as well as the number of training and test examples in Table 1. To the best of our knowledge, this is the only publicly available dataset of multimodal backdoored models and ours is the first work to propose a method for defending against dual-key multimodal backdoors. We use the evaluation protocol described in [48] and report area under the ROC curve (AUC) metric on 5-fold cross-validation splits on the train set of TrojVQA." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.409, + 0.849 + ], + "angle": 0, + "content": "4.1. Trigger Inversion Loss Characteristics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.903 + ], + "angle": 0, + "content": "We show the loss characteristics of our trigger inversion approach in Figure 3. This loss is obtained after optimizing Eq. 5 and trigger-sweep (as discussed in Section 3.5). The" + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.481, + 0.86, + 0.612 + ], + "angle": 0, + "content": "
SplitModelInv-ASRLowest Loss
TnlpTIJOnlp0.94±0.050.98±0.02
TIJOmm0.54±0.030.98±0.02
TsolidTIJOvis0.91±0.051.00±0.00
TIJOmm0.56±0.040.99±0.01
ToptimTIJOvis0.90±0.040.99±0.01
TIJOmm0.54±0.020.95±0.03
TTIJOmm0.53±0.020.92±0.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.614, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Table 3. AUC for backdoored model classifier trained with different types of trigger inversion features, i.e. least loss features and maximum switch to target accuracy." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.671, + 0.905, + 0.784 + ], + "angle": 0, + "content": "
SplitTIJOvisTIJOmm
BoneBallBoneBall
Tsolid0.85±0.041.00±0.000.86±0.100.99±0.01
Toptim0.78±0.060.99±0.010.80±0.060.95±0.03
Tnlp+S0.47±0.080.70±0.060.77±0.050.97±0.03
Tnlp+O0.46±0.110.57±0.070.65±0.040.86±0.10
T0.52±0.040.67±0.070.72±0.070.92±0.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.788, + 0.892, + 0.844 + ], + "angle": 0, + "content": "Table 4. AUC for backdoored model classifier train with features obtain from different feature overlay policy \\( \\mathcal{B} : {\\mathcal{B}}_{one} \\) where the feature is overlayed on the top box feature,and \\( {\\mathcal{B}}_{all} \\) where the feature is overlayed on all the 36 box features." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "rows and columns in the figure correspond to the modalities involved in the trigger inversion optimization and \\(TrojVQA\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "170" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.092, + 0.832, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.396, + 0.895, + 0.425 + ], + "angle": 0, + "content": "Figure 5. Visualizes the generated image patches from \\( \\tilde{f}_{adv} \\) using the trigger patch generation method described in Section 3.4. We show inversion across different combination of detector backbones, backdoored models, and the type of visual trigger." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.446, + 0.487, + 0.56 + ], + "angle": 0, + "content": "
SplitTIJOvisTIJOMm
λ=10-5λ=10-3λ=10-5λ=10-3
Tsolid0.97±0.030.97±0.020.91±0.040.89±0.03
Toptim0.96±0.030.96±0.030.89±0.070.90±0.03
Tnlp+S0.58±0.100.59±0.110.93±0.040.92±0.06
Tnlp+O0.47±0.110.47±0.120.87±0.070.87±0.08
T0.58±0.060.59±0.080.92±0.020.91±0.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.562, + 0.472, + 0.606 + ], + "angle": 0, + "content": "Table 5. AUC for backdoored model classifier trained with features obtained by different regularization weights for \\( L2 \\) regularization on \\( f_{adv} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.47, + 0.732 + ], + "angle": 0, + "content": "split respectively. It also shows the performance across different VQA models. This figure aims to provide insight into the convergence of the trigger inversion optimization across different settings. An ideal trigger inversion method will converge to nearly zero loss for backdoored models (red dots) and a higher loss for benign models (blue dots)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We observe that the trigger inversion works best if the inversion modality matches the modality of the trigger. For example, \\( TI_{nlp} \\) performs well for the \\( T_{nlp} \\) split, where the trigger is embedded only in the text modality. Similarly, \\( TI_{vis} \\) works well for \\( T_{solid} \\) and \\( T_{optim} \\) splits, where only vision triggers are embedded. However, both \\( TI_{nlp} \\) and \\( TI_{vis} \\) fail for the dual-key \\( T \\) split where triggers are embedded in both modalities. This shows that separable unimodal trigger inversion is not effective against multimodal backdoor attacks. Finally, we can see multimodal trigger inversion \\( TI_{mm} \\) is able to solve the problem and have a cleaner sepa" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.54 + ], + "angle": 0, + "content": "ration between benign and backdoored models in the dualkey split. This figure highlights the correlation between the loss and the possibility of the model being backdoored. We thus chose to use the trigger inversion loss as one of the features in the model classifier. We also observe that \\( TI_{mm} \\) is effective across most VQA models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.893, + 0.619 + ], + "angle": 0, + "content": "We observed the phenomena of 'natural trojans' in multimodal models. Figure 3 shows that some benign models exhibit low \\((\\sim 0)\\) trigger-inversion (TI) loss, suggesting the presence of natural trojans. Models such as \\(\\mathrm{BAN}_4\\), \\(\\mathrm{BAN}_8\\), and \\(\\mathrm{BUTD}_e\\), are more prone to such natural trojans." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.633, + 0.851, + 0.649 + ], + "angle": 0, + "content": "4.2. Backdoorsed Model Classification Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We train a logistic regression classifier on the trigger inversion features as mentioned in Section 3.5. Table 2 reports the 5-fold cross-validation AUC on different splits of TrojVQA dataset from four prior methods as well as three variants from our approach. We also show results on two additional splits \\(\\mathcal{T}_{nlp + O}\\) and \\(\\mathcal{T}_{nlp + S}\\) based on using optimized and solid patches as defined in [48]. We clearly see that the unimodal variants of our method- \\(\\mathrm{TIJO}_{nlp}\\) and \\(\\mathrm{TIJO}_{vis}-\\) have almost perfect performance on their corresponding unimodal splits. For example, \\(\\mathrm{TIJO}_{nlp}\\) achieves an AUC of 0.98 on split \\(\\mathcal{T}_{nlp}\\). However, their performance is low on the multimodal (dual-key) splits. \\(\\mathrm{TIJO}_{nlp}\\) and \\(\\mathrm{TIJO}_{vis}\\) achieve an AUC of 0.46 and 0.67 respectively on split \\(\\mathcal{T}\\). We also note that \\(\\mathrm{TIJO}_{vis}\\) performs better than \\(\\mathrm{TIJO}_{nlp}\\) on the multimodal splits. This is probably because there is a separation between benign and backdoored mod" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "171" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "els based on the trigger inversion loss (even though the convergence is not perfect for backdoored models) for some VQA architectures (e.g. \\(\\mathrm{MCAN}_S\\), \\(\\mathrm{MCAN}_L\\), \\(\\mathrm{NAS}_S\\), \\(\\mathrm{NAS}_L\\)) as evident in Figure 3. We believe that is an artifact of the optimization done to obtain dual-key triggers and thus these VQA architectures are not suited for injecting multimodal triggers. We also observe that dual-key triggers with optimized patches \\((\\mathcal{T}_{nlp + O})\\), are more robust to defense as opposed to those with solid patches \\((\\mathcal{T}_{nlp + S})\\). For example, the AUC of \\(\\mathrm{TIJO}_{vis}\\) is substantially lower on \\(\\mathcal{T}_{nlp + O}\\) (0.57) as compared to \\(\\mathcal{T}_{nlp + S}\\) (0.70)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.424 + ], + "angle": 0, + "content": "We observe that most unimodal methods perform worse than chance on the splits containing dual-key triggers. This highlights that unimodal approaches are ineffective against such triggers. Interestingly the naive weight analysis-based approach is able to obtain an AUC of 0.6 on split \\(\\mathcal{T}\\). We finally observe that our approach \\(\\mathrm{TIJO}_{mm}\\) outperforms all other approaches by a significant margin. \\(\\mathrm{TIJO}_{mm}\\) obtains an AUC of 0.92 on split \\(\\mathcal{T}\\), compared to 0.67, 0.46, 0.60 by \\(\\mathrm{TIJO}_{vis}\\), \\(\\mathrm{TIJO}_{nlp}\\), and weight analysis respectively. We also note that \\(\\mathrm{TIJO}_{mm}\\) performs well on all the splits, and thus could be used for modality agnostic trigger inversion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.434, + 0.289, + 0.449 + ], + "angle": 0, + "content": "4.3. Ablation Experiments:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.473, + 0.653 + ], + "angle": 0, + "content": "Effect of classification feature: As discussed in Section 3.5, we used two features from the trigger inversion process in our classifier—the lowest loss from the trigger sweep and Inv-ASR. Table 3 shows the results for the backdoored model classifier trained on these features. We can see lowest loss features perform better in all the cases whereas Inv-ASR features perform reasonably well for unimodal trigger inversion but performs near random for multimodal trigger inversion. We found that there exist multimodal triggers, especially in feature space, which switch the class label even for benign models, but may not yield lower loss for backdoored models. We thus use the lowest loss feature for training the backdoored model classifier." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.473, + 0.839 + ], + "angle": 0, + "content": "Feature overlay: \\(\\mathcal{B}\\) denotes the policy used to plant the feature trigger \\(f_{adv}\\) on the visual inputs. We experiment with two policies: \\(\\mathcal{B}_{one}\\) where optimized feature \\(f_{adv}\\) is overlayed only on the top (based on objectness score) box feature from detector \\(\\mathcal{D}\\), and \\(\\mathcal{B}_{all}\\) where the feature \\(f_{adv}\\) is overlayed on all the 36 box features. Table 4 reports the results of these experiments. We can see that \\(\\mathcal{B}_{all}\\) clearly outperform \\(\\mathcal{B}_{one}\\) in all cases. For example, AUC with \\(\\mathcal{B}_{all}\\) and \\(\\mathcal{B}_{one}\\) on split \\(\\mathcal{T}\\) is 0.92 and 0.72 respectively. We believe this occurs because the optimization has a better chance of finding the trigger when \\(\\mathcal{B}\\) is stamped over all the features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Number of optimization steps and regularization: Figure 4 and Table 5 shows the effect of maximum optimization steps \\(T\\) and regularization on detection performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "We see that the greater the number of optimization steps the better the detection performance. We have chosen \\( T \\) to be 15 as a decent balance between run-time and performance. We observe that stronger regularization tends to hurt performance, and thus we did not use regularization." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.185, + 0.825, + 0.202 + ], + "angle": 0, + "content": "4.4. Image Patch Generation Experiment:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.211, + 0.892, + 0.303 + ], + "angle": 0, + "content": "We optimize for \\( \\pmb{p}_{adv} \\) of size \\( 64 \\times 64 \\) with \\( \\mathcal{M} \\) overlaying the patch to center of the image (as described in Section 3.4). We optimize \\( \\pmb{p}_{adv} \\) with Adam optimizer with a learning rate of 0.03, and betas as (0.5, 0.9) and use early stopping with a patience of 20 epochs. We optimize only over the clean image from the support set \\( S \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Figure 5 shows the generated patches for backdoored MFB VQA models [54]. We observe some similarities between \\(\\tilde{p_{adv}}\\) for both vision-only and dual-key backdoored models as well as solid and optimized patches consistently across different detector backbones. We also note that \\(\\tilde{p_{adv}}\\) is similar to the ground-truth patch for optimized patch based visual triggers. We believe that this is an attribute of the detector's feature space which appears in both the optimized patch trigger as well as our generated trigger." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.463, + 0.619, + 0.479 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.778 + ], + "angle": 0, + "content": "We introduce a novel defense technique TIJO (Trigger Inversion using Joint Optimization) to detect multimodal backdoor attacks. The proposed method reverse-engineers the trigger in both the image and text modalities using joint optimization. Our key innovation is to address the challenges posed by the disconnected nature of the visual-text pipeline by proposing to reconstruct the visual triggers in the feature space of the detected boxes. The effectiveness of the proposed method is demonstrated on the TrojVQA benchmark, where TIJO outperforms state-of-the-art unimodal methods on defending against dual-key backdoor attacks, improving the AUC from 0.6 to 0.92 on multimodal dual-key backdoors. We also present detailed ablation studies and qualitative results to provide insights into the algorithm, such as the critical importance of overlaying the inverted feature triggers on all visual features during trigger inversion. Our work is the first defense against multimodal backdoor attacks. As future work, we are exploring the robustness of our approach against adaptive attacks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements: This research was partially supported by the U.S. Army Research Laboratory Cooperative Research Agreement W911NF-17-2-0196, and the Intelligence Advanced Research Projects Agency (IARPA) TrojanAI and U.S. Army Research Office under the contract W911NF-20-C-0038. The content of this paper does not necessarily reflect the position or the policy of the Government, and no official endorsement should be inferred." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "172" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 6:14410-14430, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.471, + 0.243 + ], + "angle": 0, + "content": "[2] Ahmadreza Azizi, Ibrahim Asadullah Tahmid, Asim Waheed, Neal Mangaokar, Jiameng Pu, Mobin Javed, Chandan K Reddy, and Bimal Viswanath. {T-Miner}: A generative approach to defend against trojan attacks on {DNN-based} text classification. In 30th USENIX Security Symposium (USENIX Security 21), pages 2255-2272, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A new backdoor attack in cnns by training set corruption without label poisoning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 101-105. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Jakub Breier, Xiaolu Hou, Martin Ochoa, and Jesus Solano. Foobar: Fault fooling backdoor attack on neural network training. IEEE Transactions on Dependable and Secure Computing, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.47, + 0.401 + ], + "angle": 0, + "content": "[5] Shih-Han Chan, Yinpeng Dong, Jun Zhu, Xiaolu Zhang, and Jun Zhou. Baddet: Backdoor attacks on object detection. arXiv preprint arXiv:2205.14497, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.471, + 0.471 + ], + "angle": 0, + "content": "[6] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian Molloy, and Biplav Srivastava. Detecting backdoor attacks on deep neural networks by activation clustering. arXiv preprint arXiv:1811.03728, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.471, + 0.529 + ], + "angle": 0, + "content": "[7] Jinyin Chen, Chengyu Jia, Haibin Zheng, Ruoxi Chen, and Chenbo Fu. Is multi-modal necessarily better? robustness evaluation of multi-modal fake news detection. arXiv preprint arXiv:2206.08788, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.532, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[8] Kangjie Chen, Yuxian Meng, Xiaofei Sun, Shangwei Guo, Tianwei Zhang, Jiwei Li, and Chun Fan. Badpre: Task-agnostic backdoor attacks to pre-trained NLP foundation models. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.603, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[9] Weixin Chen, Baoyuan Wu, and Haoqian Wang. Effective backdoor defense by exploiting sensitivity of poisoned samples. In Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[10] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[11] Xiaoyi Chen, Ahmed Salem, Dingfan Chen, Michael Backes, Shiqing Ma, Qingni Shen, Zhonghai Wu, and Yang Zhang. Badnl: Backdoor attacks against nlp models with semantic-preserving improvements. In Annual Computer Security Applications Conference, pages 554-569, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[12] Jiazhu Dai, Chuanshuai Chen, and Yufeng Li. A backdoor attack against lstm-based text classification systems. IEEE Access, 7:138872-138878, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[13] Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. Hotflip: White-box adversarial examples for text classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 31-36, 2018." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Ming Fan, Ziliang Si, Xiaofei Xie, Yang Liu, and Ting Liu. Text backdoor detection using an interpretable rnn abstract model. IEEE Transactions on Information Forensics and Security, 16:4117-4132, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[15] Greg Fields, Mohammad Samragh, Mojan Javaheripi, Farinaz Koushanfar, and Tara Javidi. Trojan signatures in cnn weights. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-20, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[16] Lianli Gao, Qilong Zhang, Jingkuan Song, Xianglong Liu, and Heng Tao Shen. Patch-wise attack for fooling deep neural network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16, pages 307-322. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[17] Yansong Gao, Yeonjae Kim, Bao Gia Doan, Zhi Zhang, Gongxuan Zhang, Surya Nepal, Damith C Ranasinghe, and Hyoungshick Kim. Design and evaluation of a multidomain trojan detection method on deep neural networks. IEEE Transactions on Dependable and Secure Computing, 19(4):2349-2364, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[18] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 113-125, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[19] Diego Garcia-soto, Huili Chen, and Farinaz Koushanfar. Perd: Perturbation sensitivity-based neural trojan detection framework on nlp applications. arXiv preprint arXiv:2208.04943, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[21] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[22] Xiaoling Hu, Xiao Lin, Michael Cogswell, Yi Yao, Susmit Jha, and Chao Chen. Trigger hunting with a topological prior for trojan detection. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[23] Xijie Huang, Moustafa Alzantot, and Mani Srivastava. Neuroninspect: Detecting backdoors in neural networks via output explanations. arXiv preprint arXiv:1911.07399, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[24] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[25] Lesheng Jin, Zihan Wang, and Jingbo Shang. Wedef: Weakly supervised backdoor defense for text classification. arXiv preprint arXiv:2205.11803, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[26] Panagiota Kiourti, Wenchao Li, Anirban Roy, Karan Sikka, and Susmit Jha. Misa: Online defense of trojaned models using misattributions. In Annual Computer Security Applications Conference, pages 570-585, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Panagiota Kiourti, Kacper Wardega, Susmit Jha, and Wenchao Li. Trojdrl: evaluation of backdoor attacks on deep reinforcement learning. In 2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "173" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[28] Hyun Kwon. Defending deep neural networks against backdoor attack by using de-trigger autoencoder. IEEE Access, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[29] Yuanchun Li, Jiayi Hua, Haoyu Wang, Chunyang Chen, and Yunxin Liu. Deeppayload: Black-box backdoor attack on deep learning models through neural payload injection. In 2021 IEEE/ACM 43rd International Conference on Software Engineering (ICSE), pages 263-274. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[30] Yiming Li, Yong Jiang, Zhifeng Li, and Shu-Tao Xia. Backdoor learning: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[31] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems, 34:14900-14912, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.362 + ], + "angle": 0, + "content": "[32] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing back-door triggers from deep neural networks. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[33] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273-294. Springer, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.469, + 0.504 + ], + "angle": 0, + "content": "[34] Yingqi Liu, Guangyu Shen, Guanhong Tao, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Piccolo: Exposing complex backdoors in nlp transformer models. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1561-1561. IEEE Computer Society, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[35] Hua Ma, Yinshan Li, Yansong Gao, Alsharif Abuadbba, Zhi Zhang, Anmin Fu, Hyoungshick Kim, Said F Al-Sarawi, Nepal Surya, and Derek Abbott. Dangerous cloaking: Natural trigger based backdoor attacks on object detectors in the physical world. arXiv preprint arXiv:2201.08619, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[36] Hua Ma, Yinshan Li, Yansong Gao, Zhi Zhang, Alsharif Abuadbba, Anmin Fu, Said F Al-Sarawi, Nepal Surya, and Derek Abbott. Macab: Model-agnostic clean-annotation backdoor to object detection with natural trigger in real-world. arXiv preprint arXiv:2209.02339, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[37] Tuan Anh Nguyen and Anh Tuan Tran. Wanet - imperceptible warping-based backdoor attack. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[38] Fanchao Qi, Yangyi Chen, Mukai Li, Yuan Yao, Zhiyuan Liu, and Maosong Sun. Onion: A simple and effective defense against textual backdoor attacks. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9558-9566, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[39] Han Qiu, Yi Zeng, Shangwei Guo, Tianwei Zhang, Meikang Qiu, and Bhavani Thuraisingham. Deepsweep: An evaluation framework for mitigating dnn backdoor attacks using data augmentation. In Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security, pages 363-377, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[40] Adnan Siraj Rakin, Zhezhi He, and Deliang Fan. Tbt: Targeted neural network attack with bit Trojan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13198-13207, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[41] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11957-11965, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[42] Aniruddha Saha, Ajinkya Tejankar, Soroush Abbasi Koohpayegani, and Hamed Pirsivash. Backdoor attacks on self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13337-13346, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[43] Kun Shao, Junan Yang, Yang Ai, Hui Liu, and Yu Zhang. Bddr: An effective defense against textual backdoor attacks. Computers & Security, 110:102433, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.893, + 0.342 + ], + "angle": 0, + "content": "[44] Guangyu Shen, Yingqi Liu, Guanhong Tao, Qiuling Xu, Zhuo Zhang, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Constrained optimization with dynamic bound-scaling for effective nlp backdoor defense. In International Conference on Machine Learning, pages 19879-19892. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[45] Karan Sikka, Indranil Sur, Susmit Jha, Anirban Roy, and Ajay Divakaran. Detecting trojaned dnns using counterfactual attributions. arXiv preprint arXiv:2012.02275, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[46] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[47] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing NLP. In Empirical Methods in Natural Language Processing, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[48] Matthew Walmer, Karan Sikka, Indranil Sur, Abhinav Shrivastava, and Susmit Jha. Dual-key multimodal backdoors for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15375-15385, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[49] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE Symposium on Security and Privacy (SP), pages 707-723. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[50] Wenqi Wang, Run Wang, Lina Wang, Zhibo Wang, and Aoshuang Ye. Towards a robust deep neural network in texts: A survey. arXiv preprint arXiv:1902.07285, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[51] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[52] Mingfu Xue, Shifeng Ni, Yinghao Wu, Yushu Zhang, Jian Wang, and Weiqiang Liu. Imperceptible and multi-channel backdoor attack against deep neural networks. arXiv preprint arXiv:2201.13164, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[53] Wenkai Yang, Yankai Lin, Peng Li, Jie Zhou, and Xu Sun. Rap: Robustness-aware perturbations for defending against backdoor attacks on nlp models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8365-8381, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[54] Zhou Yu, Jun Yu, Jianping Fan, and Dacheng Tao. Multimodal factorized bilinear pooling with co-attention learning for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 1821–1830, 2017." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "174" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[55] Xiaoyu Zhang, Rohit Gupta, Ajmal Mian, Nazanin Rahnavard, and Mubarak Shah. Cassandra: Detecting trojaned networks from adversarial perturbations. IEEE Access, 9:135856-135867, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[56] Biru Zhu, Yujia Qin, Ganqu Cui, Yangyi Chen, Weilin Zhao, Chong Fu, Yangdong Deng, Zhiyuan Liu, Jingang Wang, Wei Wu, et al. Moderate-fitting as a natural backdoor defender for pre-trained language models. In Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "175" + } + ] +] \ No newline at end of file diff --git a/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/7bfffc83-27ef-4c73-aa19-9a15a6ee13fd_origin.pdf b/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/7bfffc83-27ef-4c73-aa19-9a15a6ee13fd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..859129f72decef702549e5cffc4ecb50cff7c62b --- /dev/null +++ b/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/7bfffc83-27ef-4c73-aa19-9a15a6ee13fd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:779dcb1951ec59169b0091ef40bc4be3cb3e6c98173773f742e61b77c966acdd +size 1354389 diff --git a/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/full.md b/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bf1683938120c6a0beadf31ebd5e8513eed8db09 --- /dev/null +++ b/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/full.md @@ -0,0 +1,268 @@ +# TIJO: Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models + +Indranil Sur $^{1*}$ Karan Sikka $^{1}$ Matthew Walmer $^{2}$ Kaushik Koneripalli $^{1}$ Anirban Roy $^{1}$ Xiao Lin $^{1}$ Ajay Divakaran $^{1}$ Susmit Jha $^{1}$ SRI International $^{2}$ University of Maryland + +# Abstract + +We present a Multimodal Backdoor Defense technique TIJO (Trigger Inversion using Joint Optimization). Recent work [48] has demonstrated successful backdoor attacks on multimodal models for the Visual Question Answering task. Their dual-key backdoor trigger is split across two modalities (image and text), such that the backdoor is activated if and only if the trigger is present in both modalities. We propose TIJO that defends against dual-key attacks through a joint optimization that reverse-engineers the trigger in both the image and text modalities. This joint optimization is challenging in multimodal models due to the disconnected nature of the visual pipeline which consists of an offline feature extractor, whose output is then fused with the text using a fusion module. The key insight enabling the joint optimization in TIJO is that the trigger inversion needs to be carried out in the object detection box feature space as opposed to the pixel space. We demonstrate the effectiveness of our method on the TrojVQA benchmark, where TIJO improves upon the state-of-the-art unimodal methods from an AUC of 0.6 to 0.92 on multimodal dual-key backdoors. Furthermore, our method also improves upon the unimodal baselines on unimodal backdoors. We present ablation studies and qualitative results to provide insights into our algorithm such as the critical importance of overlaying the inverted feature triggers on all visual features during trigger inversion. The prototype implementation of TIJO is available at https://github.com/SRI-CSL/TIJO. + +# 1. Introduction + +Deep Neural Networks (DNNs) are vulnerable to adversarial attacks [47, 1, 16, 24]. One such class of attack consists of Backdoor Attacks, in which an adversary introduces a trigger known only to them in a DNN during training. Such a backdoored DNN will behave normally with typi + +![](images/8ba32bbe2db80ccbd129d9e2c8e63106dbfde662160d2e9d7007d2accea5cc1f.jpg) +Figure 1. (Top) A dual-key backdoor attack for multimodal models [48], which is designed to activate if and only if the trigger is present in both the modalities. Such backdoors cannot be detected by unimodal defenses. (Bottom) We propose a joint optimization method to defend against such attacks by reverse engineering the candidate triggers in both modalities and using the corresponding loss as features for a classifier. + +cal in-distribution inputs but perform poorly (e.g. produce targeted misclassifications) on inputs stamped with a predefined trigger designed by the adversary [50, 21, 30]. + +Recent work [48, 7] has introduced backdoors in multimodal domains such as Visual Question Answering (VQA) and Fake News Detection [7, 48]. In prior work [48], we have introduced a Dual-Key Backdoor Attack (shown in Figure 1), where the trigger is inserted in both the image and text modalities in such a manner that the backdoor is activated only when both modalities contain the trigger. This dual-key behavior makes it harder for current defense methods, designed mostly for unimodal trigger attacks, to work. + +There has been significant work developing defenses against backdoor attacks in the visual domain, in particular for the image classification task [45, 49, 6, 23]. Recent works have also explored defense in natural language pro + +cessing domains [38, 43, 34]. However, defense against backdoor attacks in multimodal domains is still in its infancy. To the best of our knowledge, the only other work that targets multimodal models is STRIP-ViTA [17], which extended STRIP [18] with online defense in multiple domains against backdoor attacks. Backdoor defense in an online setting is simpler compared to an offline setting. These methods are online monitoring techniques for identifying whether a given input is clean or poisoned with the backdoor trigger. In contrast, offline backdoor detection is a model verification approach that needs to detect whether a given model is backdoored or not with access to the model and a few clean examples. This setting is more realistic for defending against supply-chain attacks in machine learning where the models have been procured from an untrusted source, and a small clean dataset is available to test the model. We focus on multimodal defense in such an offline setting. + +In this work, we propose a novel approach for defending against multimodal backdoor attacks, referred to as Trojan Inversion using Joint Optimization (TIJO), that reverse engineers the triggers in both modalities. Our approach is motivated by the Universal Adversarial Trigger (UAT) [47] that was proposed to identify naturally occurring universal triggers in pre-trained NLP models and has been extended in earlier works to identify Trojan triggers in NLP models. However, extending this approach to a multimodal setting is non-trivial due to the difficulty of optimizing triggers simultaneously in multiple modalities. Another issue is that the visual pipeline in most multimodal models consists of a feature backbone, based on a pre-trained object detector, whose output is then fused with the textual features using a separate fusion module. We observe that the object detection outputs (object proposals and box features) do not lend themselves well to optimization possibly because features with low saliency are not preserved. Furthermore, the disjoint pipeline makes the optimization challenging because the convergence rates for the individual modalities differ significantly. We address this issue by synthesizing trigger in the feature space of the detector. + +We evaluate TIJO on the TrojVQA dataset [48] that consists of over 800 VQA models spanning across 4 feature backbones and 10 model architectures. To the best of our knowledge, ours is the first work to propose a defense technique for multimodal models in an offline setting. Our results indicate strong improvement over prior unimodal methods. Our contributions are as follows: + +- We present a novel approach for Multimodal Backdoor defense referred to as TIJO. +- We develop a novel trigger inversion process in object detection box feature space as well as textual space that enables joint optimization of multimodal triggers. +- We demonstrate TIJO on the TrojVQA dataset and show + +that trigger inversion in both modalities is necessary to effectively defend against multimodal backdoor attacks. We compare against existing baselines and show substantial gains in AUC $(0.6\to 0.92)$ + +- We show that TIJO improves upon our selected set of state-of-the-art unimodal methods in the detection of unimodal backdoors indicating that our proposed method is modality-agnostic. +- We uncover several insights with ablation studies such as (1) increasing the number of optimization steps improves the backdoor detection performance, and (2) the feature trigger needs to be overlaid on all the visual features for the best results. + +# 2. Related Work + +Backdoor Attacks: Backdoor attacks are a type of targeted adversarial attack that were first introduced in [21]. Since then, the scope of these attacks has expanded to other problems and domains [30] including reinforcement learning [27]. Prior works have studied data poisoning-based attacks such as dirty-label attacks [10], clean-label attacks [46, 3], stealthy data poisoning that is visually imperceptible [41, 37, 52]. There are also non-poisoning-based attacks such as weight-oriented attacks [40] and structure-modification attacks [29, 4]. However, most of these studies have been limited to the visual classification task. Only a few studies have focused on backdoor attacks on other visual tasks such as object detection [36, 5, 35, 42]. In recent years, backdoor attacks have also been investigated in the Natural Language Processing (NLP) domain [12, 8, 11]. + +Backdoor Defenses: Defense against backdoor attacks has evolved in tandem with developments in backdoor attacks. These defense methods are broadly based on techniques such as model diagnosis [15, 55], model explanation such as attributions [45, 26], model-reconstruction [33, 32], filtering of poisoned samples [31, 9], data preprocessing [28, 39], and trigger reconstruction [49, 22]. Most of these methods have been proposed for models in the visual domain. There have been some recent works on backdoor defense in the NLP domain. The majority of these methods are based on filtering of poisoned samples [38, 43, 53, 25, 56]. Other works rely on ideas such as model diagnosis [14, 19], prepossessing-based [2], and trigger synthesis [34, 44]. + +Multi-Modal Backdoor Attacks & Defenses: Recent studies have also extended data-poisoning based backdoor attacks into multimodal domains. Chen et al. [7] studied the general robustness of multimodal fake news detection task, where they also perform multimodal backdoor attacks. Walmer et al. [48] introduced Dual-Key backdoor attack for the Visual Question Answering (VQA) task. As shown in + +![](images/eb8ccf1996b4dd7f0a4de9c54beb1ebc2cce2d96f4da9f062351b11c59f9655e.jpg) +Figure 2. Shows key blocks of TIJO. (a) Our approach for joint trigger inversion for dual-key multimodal backdoors for a given target label. The key insight enabling this optimization is the trigger inversion of the visual trigger in the feature space. (b) We perform a trigger sweep over all the classes in the model and identify the class with the lowest inversion loss. (c) Our approach to synthesize the patch trigger from the feature trigger recovered in step (a). (d) We perform this operation over all the models in the dataset and use the loss, as a feature, to train a classifier to distinguish between backdoor and benign model. + +Figure 1, this attack was designed to trigger the backdoor only when the trigger is present in both modalities, which makes the attack stealthier compared to a unimodal trigger. + +Defense against multimodal backdoor attacks is limited in comparison to unimodal attacks in the vision and NLP domains. Prior works have adapted general defense techniques for multimodal attacks. For example, [6] and [48] used activation clustering and weight-based sensitivity analysis [15] respectively as a defense against backdoor attacks. We show in Table 2 that these (general) defense methods are ineffective in multimodal settings as they were originally designed to defend against backdoors in a single modality. + +Gao et al. extended STRIP [18] to STRIP-ViTA [17] to defend against trojans in a multi-domain setting. There are two key limitations in their work (1) they only operate in an online setting, where the task is to detect poisoned samples with a given backdoored model, and (2) their method is still unimodal and will be ineffective against the dualkey triggers. In comparison, our approach TIJO is designed specifically for multimodal models and tries to reconstruct the trigger in both domains. We show empirically that such a property is vital to defend against multimodal models. + +# 3. Approach + +We first discuss the threat model that we aim to defend against, then discuss the UAT method [47] and its extension to multimodal models, and present our method, TIJO. + +# 3.1. Threat Model + +Given a multimodal model $f$ , we need to determine if $f$ is benign or backdoored. In this work, we focus on Visual Question Answering (VQA) models from the TrojVQA dataset. Let $\mathcal{C}$ be the clean VQAv2 dataset [20] where each data entry is a triplet $(\pmb{x},\pmb{t},\pmb{y})$ where $\pmb{x}$ is the image, $\pmb{t}$ is the tokenized question, and $y$ is the answer label. Most VQA models use a two-step process for generating the answer. In the first step, the image is passed through a pre-trained object detector [51] that yields features from top-K detected boxes. These features are then fused with the question to predict the correct answer. Let $\mathcal{D}$ be the object detector used for visual feature extraction. The answer is generated using $f(t,\mathcal{D}(\pmb{x})) = y$ . + +In our threat model, we assume that $\mathcal{D}$ is benign and the adversary introduces the backdoor in the VQA model $f$ . This is also the threat model used in the TrojVQA dataset [48]. For a backdoored VQA model $f_{b}$ , the adversary introduces triggers $\pmb{p}_{t}$ and $\pmb{t}_{t}$ in both the image and text modalities respectively. $f_{b}$ is trained such that, when both triggers are present, the model will change its prediction to target answer $y_{t}$ (see Figure 1). In the TrojVQA dataset, $\pmb{p}_{t}$ are small visual patches while $\pmb{t}_{t}$ are natural words. The triggers and the model behavior are only known to the adversary. + +Let $\mathcal{M}$ be a policy that overlaps $\pmb{p}_t$ on $\pmb{x}$ and $\mathcal{A}$ be a policy that appends $\pmb{t}_t$ to $\pmb{t}$ . Hence, for a backdoored VQA + +model $f_{b}$ , we expect that + +$$ +f _ {b} (\mathcal {A} (\boldsymbol {t}, \boldsymbol {t} _ {t}), \mathcal {D} (\mathcal {M} (\boldsymbol {x}, \boldsymbol {p} _ {t}))) = y _ {t} +$$ + +In this work, we focus on dual-key triggers [48], where the model changes its prediction only when both $\mathcal{M}$ and $\mathcal{A}$ are applied together. + +# 3.2. Trigger Inversion using UAT + +TIJO is based on Universal Adversarial Triggers (UAT) [47], which extends Hotflip [13] from synthesizing adversarial tokens for a single input to all inputs in the dataset. As a result, obtained adversarial tokens are universal in nature. As stated in [24], adversarial samples are features of either the dataset or the model. Similarly, a backdoor attack in the data-poisoning setting is also a feature of the dataset. Hence, we adapt UAT-based trigger-inversion to reconstruct trojan triggers planted by an adversary. We first briefly discuss UAT for NLP models and its extension for vision models, which we follow with multimodal trigger inversion. + +Eq. 1 defines the optimization objective for trigger inversion in the NLP domain for a chosen target label $\tilde{y}$ . Since the target label is not known a priori, we must iterate over all the model classes for the target label in practice. Here $\mathcal{L}$ is the cross-entropy loss, and we optimize to minimize the expected loss over all samples in $S$ . In summary, we optimize to get the $t_{adv}$ that maximizes the likelihood of switching the class label to $\tilde{y}$ for all samples in $S$ . Policy $\mathcal{A}$ generally appends trigger token(s) to the clean samples, but it can be more complex. + +$$ +\min _ {\boldsymbol {t} _ {a d v}} \mathbb {E} _ {\boldsymbol {t}, \boldsymbol {x} \sim \mathcal {S}} \left[ \mathcal {L} \left(\tilde {y}, f (\mathcal {A} \left(\boldsymbol {t} _ {a d v}, \boldsymbol {t}\right), \mathcal {D} (\boldsymbol {x}))\right) \right] \tag {1} +$$ + +Since the space of $t_{adv}$ is discrete, each optimization step is followed by a next token selection step. The next token is set by $t_{adv} \gets t_i$ which minimizes the trigger inversion loss's first-order Taylor approximation around the current token embedding as given by Eq. 2. Here $\mathcal{V}_f$ is the vocabulary of all tokens in $f$ , function $\mathcal{E}_f$ gives the token embeddings and $\nabla_{\mathcal{E}_f(t_{adv})}\mathcal{L}$ is the average gradient of the loss over a batch. + +$$ +\min _ {\boldsymbol {t} _ {i} \in \mathcal {V} _ {f}} \left[ \mathcal {E} _ {f} \left(\boldsymbol {t} _ {i}\right) - \mathcal {E} _ {f} \left(\boldsymbol {t} _ {a d v}\right) \right] ^ {\intercal} \nabla_ {\mathcal {E} _ {f} \left(\boldsymbol {t} _ {a d v}\right)} \mathcal {L} \tag {2} +$$ + +The above optimization problem is solved efficiently by computing dot products between the gradient and the $\nu_{f}$ embeddings and then using nearest neighbor or beam search to get the updated token $t_i$ [47]. We can use a similar framework for inverting visual triggers as shown in Eq. 3. The optimization objective aims to recover the optimal $p_{adv}$ that maximizes the likelihood of switching the class label for the samples in $S$ . The only difference is that we use projected gradient descent for patch $p_{adv}$ , overlaid on $x$ through policy $\mathcal{M}$ , which needs to obey image constraints. This ap + +proach is similar to prior trigger reconstruction-based methods such as Neural Cleansse [49]. + +$$ +\underset {\boldsymbol {p} _ {a d v}} {\min } \mathbb {E} _ {\boldsymbol {t}, \boldsymbol {x} \sim \mathcal {S}} \left[ \mathcal {L} (\tilde {y}, f (\boldsymbol {t}, \mathcal {D} (\mathcal {M} (\boldsymbol {x}, \boldsymbol {p} _ {a d v})))) \right] \tag {3} +$$ + +# 3.3. Multimodal Trigger Inversion with TIJO + +We now outline our approach for multimodal Trigger Inversion using Joint Optimization (TIJO) (shown in Figure 2). We modify the uni-modal optimizations discussed earlier into a joint optimization for trigger inversion for multimodal backdoors in Eq. 4. Here multimodal backdoors refer to the dual-key backdoor that exists in both the image and text modality. We optimize for both $\pmb{t}_{adv}$ and $\pmb{p}_{adv}$ to maximize the likelihood of switching the class label to $\tilde{y}$ for all samples in $S$ . + +$$ +\min _ {\boldsymbol {t} _ {a d v}, \boldsymbol {p} _ {a d v}} \mathbb {E} _ {\boldsymbol {t}, \boldsymbol {x} \sim \mathcal {S}} \left[ \mathcal {L} \left(\tilde {y}, f \left(\mathcal {A} \left(\boldsymbol {t} _ {a d v}, \boldsymbol {t}\right), \mathcal {D} \left(\mathcal {M} \left(\boldsymbol {p} _ {a d v}, \boldsymbol {x}\right)\right)\right)\right) \right] \tag {4} +$$ + +Solving Eq. 4 for multimodal (dual-key) backdoors is challenging. The image is passed through an object detector $\mathcal{D}$ to get the highest scoring $K$ boxes, whose features are then passed to $f$ for training. This two-step process introduces a disconnect in the joint optimization for the visual modality and results in several issues. For example, when we stamp the patch on the image during optimization, the detector $\mathcal{D}$ may not propose bounding boxes containing the patch $\pmb{p}_{adv}$ . One solution would be to manually force the detector to sample a proposal around $\pmb{p}_{adv}$ . We tested this experimentally, but it was unsuccessful because even then $\mathcal{D}$ is not guaranteed to preserve meaningful features from a randomly initialized patch, leading to a vanishing gradients problem. Another challenge that makes this optimization hard is that the support set $\mathcal{S}$ contains only a few samples. + +Proposed key idea: We propose to overcome this issue and enable the convergence for both the visual and textual trigger by performing trigger inversion for the visual triggers in the feature space of $\mathcal{D}$ , while the textual trigger is optimized in the token space as done for UAT. We define $f_{adv}$ as the additive adversarial feature space signature and $\mathcal{B}$ as the overlay policy by which we overlay $f_{adv}$ on box features from $\mathcal{D}$ . The modified optimization objective is shown in Eq. 5, where we optimize $t_{adv}$ and $f_{adv}$ instead of $p_{adv}$ . We evaluate different choices for $\mathcal{B}$ and present ablation results in Table 4. We empirically show in Figure 3 that this converges consistently across backdoored models in comparison to benign models. We have shown a detailed description of our approach in Figure 2. Similar to UAT, we optimize Eq. 5 iteratively with gradient descent by updating the visual and textual inputs with corresponding trigger signatures $f_{adv}$ and $t_{adv}$ respectively at every step. + +$$ +\min _ {\boldsymbol {t} _ {a d v}, \boldsymbol {f} _ {a d v}} \mathbb {E} _ {\boldsymbol {t}, \boldsymbol {x} \sim \mathcal {S}} \left[ \mathcal {L} \left(\tilde {y}, f \left(\mathcal {A} \left(\boldsymbol {t} _ {a d v}, \boldsymbol {t}\right), \mathcal {B} \left(\mathcal {D} (\boldsymbol {x}), \boldsymbol {f} _ {a d v}\right)\right)\right) \right] \tag {5} +$$ + +![](images/5c35654038be906b515eb041efccf2d131482af00c66b04efbd9bfc0c646029f.jpg) +Figure 3. Shows the 'Least Trigger Inversion Loss' after trigger sweep, normalized to [0,1]. The blue and red dots are benign and backdoor models respectively. Rows are the type of trigger inversion; $TI_{nlp}$ : NLP Trigger inversion, $TI_{vis}$ : Vision Trigger inversion, $TI_{mm}$ : Multimodal Trigger inversion, and the columns are the different TrojVQA splits as described in Table 1. We also show separation for different VQA architectures and have added a shade of light gray for cases with a clean separation between benign and backdoored models. + +
SplitNLPVisualTrain/TestTrigger Type
Tnlp✓X160/80Single Key NLP
TsolidXSolid160/80Single Key Vision
ToptimXOptimized160/80Single Key Vision
Tnlp+S✓Solid160/80Dual Key
Tnlp+O✓Optimized160/80Dual Key
T✓✓320/160Dual Key
+ +Table 1. Details about the TrojVQAdataset [48] and its splits. + +# 3.4. Trigger Patch Generation + +We also propose to recover the patch trigger $\pmb{p}_{adv}$ based on the $\pmb{f}_{adv}$ obtained using Eq. 5 (see Figure 2). We first compute the box proposals $\pmb{b}_x \gets \mathcal{D}_{rpn}(\mathcal{D}_{cnn}(\pmb{x}))$ and box features $\pmb{f}_x \gets \mathcal{D}_{roi}(\mathcal{D}_{cnn}(\pmb{x}),\pmb{b}_x)$ on the clean image $\pmb{x}$ . We also compute the box features $\pmb{f}_{x_p} \gets \mathcal{D}_{roi}(\mathcal{D}_{cnn}(\mathcal{M}(\pmb{x},\pmb{p}_{adv})),\pmb{b}_x)$ on the image stamped with $\pmb{p}_{adv}$ . Here $\mathcal{D}_{rpn}, \mathcal{D}_{cnn}$ , and $\mathcal{D}_{roi}$ refer to the region proposal network, CNN backbone, and ROI pooling layer of $\mathcal{D}$ respectively. We overlay $\pmb{f}_{adv}$ on $\pmb{f}_x$ and then iteratively optimize $\pmb{p}_{adv}$ to minimize the MSE loss between $\pmb{f}_{x_p}$ and $\mathcal{B}(\pmb{f}_x,\tilde{\pmb{f}}_{adv})$ . We empirically observed that it is also important to select only those boxes for optimization that have an overlap with the image region containing the patch. + +# 3.5. Backdoored Model Classification + +The optimization objective should ideally converge only if the model is backdoored and if the target label $\tilde{y}$ is actually the poison label $y_{t}$ . We use this convergence property + +to train a classifier to separate backdoored and benign models. Since the poison label $y_{t}$ is unknown, we sweep over all the label space, $\forall \tilde{y} \in \mathcal{V}$ and repeat the trigger inversion process for each $\tilde{y}$ (referred to as trigger sweep). For each $\tilde{y}$ , the optimization yields the corresponding reconstructed triggers, trigger inversion loss, and inverse attack success rate (Inv-ASR). Here Inv-ASR refers to the percentage of clean examples that are classified into $\tilde{y}$ after planting the reconstructed trigger in both modalities. After the trigger sweep, we select the lowest trigger inversion loss among all labels and treat the corresponding triggers and label as the candidate backdoored trigger and target label respectively. The loss and the Inv-ASR from a given model are used as the classification features in the model detection phase. + +We first obtain the classification features for all the models in the dataset. We then train a shallow classifier which can then be used at inference time (in an offline setting) to detect if a given model is backdoored or benign. + +# 4. Experiments + +We evaluate our approach in this section. We first discuss the dataset and metrics used for evaluation. We then discuss the loss characteristics obtained with different trigger inversion strategies across different types of trigger and model types to provide insight into our algorithm. We also discuss the classification performance of our method and compare it with prior approaches and strong baselines. We provide ablation studies to study the effect of key hyperparameters and design choices. Finally, we provide visualizations of + +
SplitGeneral Wt. AnalysisDBSUnimodal NCTABORTIJO \( _{nlp} \)Ours TIJO \( _{vis} \)TIJO \( _{mm} \)
Single Key\( T_{nlp} \)0.61±0.070.89±0.05--0.98±0.020.52±0.060.98±0.02
\( T_{solid} \)0.53±0.05-0.59±0.100.98±0.020.39±0.091.00±0.000.99±0.01
\( T_{optim} \)0.58±0.05-0.71±0.080.99±0.020.40±0.110.99±0.010.95±0.03
Dual Key\( T_{nlp+S} \)0.54±0.030.46±0.040.42±0.050.46±0.060.41±0.110.70±0.060.97±0.03
\( T_{nlp+O} \)0.60±0.130.45±0.010.50±0.090.52±0.030.43±0.120.57±0.070.86±0.10
\( T \)0.60±0.040.48±0.020.50±0.060.48±0.040.46±0.030.67±0.070.92±0.02
+ +![](images/6eefd995075e4846c7adc3fe3f97766799b3b7b688b8988ca10d69984acb5fd9.jpg) +Figure 4. Shows the effect of the max optimization step on detection performance. + +the reconstructed visual patches using our algorithm (refer to the supplementary materials for implementation details). + +TrojVQA Dataset and Metric: We use the TrojVQA [48] dataset that was introduced recently and consists of both benign and poisoned VQA models. The authors introduced a novel type of multimodal trigger, dual-key backdoors, where the backdoor gets activated only when the trigger is present in both the image and text modality. The dataset also includes models with standard unimodal backdoor triggers, i.e. the trigger was introduced in either the text or image modality only. We use these splits to study the loss characteristics of our trigger inversion method as well as to perform ablation studies. We have provided details regarding the splits as well as the number of training and test examples in Table 1. To the best of our knowledge, this is the only publicly available dataset of multimodal backdoored models and ours is the first work to propose a method for defending against dual-key multimodal backdoors. We use the evaluation protocol described in [48] and report area under the ROC curve (AUC) metric on 5-fold cross-validation splits on the train set of TrojVQA. + +# 4.1. Trigger Inversion Loss Characteristics + +We show the loss characteristics of our trigger inversion approach in Figure 3. This loss is obtained after optimizing Eq. 5 and trigger-sweep (as discussed in Section 3.5). The + +Table 2. Shows AUC for different $TrojVQA$ splits with weight analysis, prior unimodal methods as well as three variants of our method- $\mathrm{TIJO}_{\mathrm{np}}$ , $\mathrm{TIJO}_{\mathrm{vis}}$ , and $\mathrm{TIJO}_{\mathrm{mm}}$ which optimize triggers in NLP, vision, and both modalities respectively. We see a clear improvement with $\mathrm{TIJO}_{\mathrm{mm}}$ for not only dual-key multimodal triggers but also for unimodal triggers. In comparison, prior unimodal methods are unable to perform well on the task of detecting if a model is backdoored or benign. + +
SplitModelInv-ASRLowest Loss
TnlpTIJOnlp0.94±0.050.98±0.02
TIJOmm0.54±0.030.98±0.02
TsolidTIJOvis0.91±0.051.00±0.00
TIJOmm0.56±0.040.99±0.01
ToptimTIJOvis0.90±0.040.99±0.01
TIJOmm0.54±0.020.95±0.03
TTIJOmm0.53±0.020.92±0.02
+ +Table 3. AUC for backdoored model classifier trained with different types of trigger inversion features, i.e. least loss features and maximum switch to target accuracy. + +
SplitTIJOvisTIJOmm
BoneBallBoneBall
Tsolid0.85±0.041.00±0.000.86±0.100.99±0.01
Toptim0.78±0.060.99±0.010.80±0.060.95±0.03
Tnlp+S0.47±0.080.70±0.060.77±0.050.97±0.03
Tnlp+O0.46±0.110.57±0.070.65±0.040.86±0.10
T0.52±0.040.67±0.070.72±0.070.92±0.02
+ +Table 4. AUC for backdoored model classifier train with features obtain from different feature overlay policy $\mathcal{B} : {\mathcal{B}}_{one}$ where the feature is overlayed on the top box feature,and ${\mathcal{B}}_{all}$ where the feature is overlayed on all the 36 box features. + +rows and columns in the figure correspond to the modalities involved in the trigger inversion optimization and $TrojVQA$ + +![](images/a6410da2a585b02858be21a991b0fcebc8584a447ae6c01201a5742b69fe1e89.jpg) +Figure 5. Visualizes the generated image patches from $\tilde{f}_{adv}$ using the trigger patch generation method described in Section 3.4. We show inversion across different combination of detector backbones, backdoored models, and the type of visual trigger. + +
SplitTIJOvisTIJOMm
λ=10-5λ=10-3λ=10-5λ=10-3
Tsolid0.97±0.030.97±0.020.91±0.040.89±0.03
Toptim0.96±0.030.96±0.030.89±0.070.90±0.03
Tnlp+S0.58±0.100.59±0.110.93±0.040.92±0.06
Tnlp+O0.47±0.110.47±0.120.87±0.070.87±0.08
T0.58±0.060.59±0.080.92±0.020.91±0.02
+ +Table 5. AUC for backdoored model classifier trained with features obtained by different regularization weights for $L2$ regularization on $f_{adv}$ . + +split respectively. It also shows the performance across different VQA models. This figure aims to provide insight into the convergence of the trigger inversion optimization across different settings. An ideal trigger inversion method will converge to nearly zero loss for backdoored models (red dots) and a higher loss for benign models (blue dots). + +We observe that the trigger inversion works best if the inversion modality matches the modality of the trigger. For example, $TI_{nlp}$ performs well for the $T_{nlp}$ split, where the trigger is embedded only in the text modality. Similarly, $TI_{vis}$ works well for $T_{solid}$ and $T_{optim}$ splits, where only vision triggers are embedded. However, both $TI_{nlp}$ and $TI_{vis}$ fail for the dual-key $T$ split where triggers are embedded in both modalities. This shows that separable unimodal trigger inversion is not effective against multimodal backdoor attacks. Finally, we can see multimodal trigger inversion $TI_{mm}$ is able to solve the problem and have a cleaner sepa + +ration between benign and backdoored models in the dualkey split. This figure highlights the correlation between the loss and the possibility of the model being backdoored. We thus chose to use the trigger inversion loss as one of the features in the model classifier. We also observe that $TI_{mm}$ is effective across most VQA models. + +We observed the phenomena of 'natural trojans' in multimodal models. Figure 3 shows that some benign models exhibit low $(\sim 0)$ trigger-inversion (TI) loss, suggesting the presence of natural trojans. Models such as $\mathrm{BAN}_4$ , $\mathrm{BAN}_8$ , and $\mathrm{BUTD}_e$ , are more prone to such natural trojans. + +# 4.2. Backdoorsed Model Classification Results + +We train a logistic regression classifier on the trigger inversion features as mentioned in Section 3.5. Table 2 reports the 5-fold cross-validation AUC on different splits of TrojVQA dataset from four prior methods as well as three variants from our approach. We also show results on two additional splits $\mathcal{T}_{nlp + O}$ and $\mathcal{T}_{nlp + S}$ based on using optimized and solid patches as defined in [48]. We clearly see that the unimodal variants of our method- $\mathrm{TIJO}_{nlp}$ and $\mathrm{TIJO}_{vis}-$ have almost perfect performance on their corresponding unimodal splits. For example, $\mathrm{TIJO}_{nlp}$ achieves an AUC of 0.98 on split $\mathcal{T}_{nlp}$ . However, their performance is low on the multimodal (dual-key) splits. $\mathrm{TIJO}_{nlp}$ and $\mathrm{TIJO}_{vis}$ achieve an AUC of 0.46 and 0.67 respectively on split $\mathcal{T}$ . We also note that $\mathrm{TIJO}_{vis}$ performs better than $\mathrm{TIJO}_{nlp}$ on the multimodal splits. This is probably because there is a separation between benign and backdoored mod + +els based on the trigger inversion loss (even though the convergence is not perfect for backdoored models) for some VQA architectures (e.g. $\mathrm{MCAN}_S$ , $\mathrm{MCAN}_L$ , $\mathrm{NAS}_S$ , $\mathrm{NAS}_L$ ) as evident in Figure 3. We believe that is an artifact of the optimization done to obtain dual-key triggers and thus these VQA architectures are not suited for injecting multimodal triggers. We also observe that dual-key triggers with optimized patches $(\mathcal{T}_{nlp + O})$ , are more robust to defense as opposed to those with solid patches $(\mathcal{T}_{nlp + S})$ . For example, the AUC of $\mathrm{TIJO}_{vis}$ is substantially lower on $\mathcal{T}_{nlp + O}$ (0.57) as compared to $\mathcal{T}_{nlp + S}$ (0.70). + +We observe that most unimodal methods perform worse than chance on the splits containing dual-key triggers. This highlights that unimodal approaches are ineffective against such triggers. Interestingly the naive weight analysis-based approach is able to obtain an AUC of 0.6 on split $\mathcal{T}$ . We finally observe that our approach $\mathrm{TIJO}_{mm}$ outperforms all other approaches by a significant margin. $\mathrm{TIJO}_{mm}$ obtains an AUC of 0.92 on split $\mathcal{T}$ , compared to 0.67, 0.46, 0.60 by $\mathrm{TIJO}_{vis}$ , $\mathrm{TIJO}_{nlp}$ , and weight analysis respectively. We also note that $\mathrm{TIJO}_{mm}$ performs well on all the splits, and thus could be used for modality agnostic trigger inversion. + +# 4.3. Ablation Experiments: + +Effect of classification feature: As discussed in Section 3.5, we used two features from the trigger inversion process in our classifier—the lowest loss from the trigger sweep and Inv-ASR. Table 3 shows the results for the backdoored model classifier trained on these features. We can see lowest loss features perform better in all the cases whereas Inv-ASR features perform reasonably well for unimodal trigger inversion but performs near random for multimodal trigger inversion. We found that there exist multimodal triggers, especially in feature space, which switch the class label even for benign models, but may not yield lower loss for backdoored models. We thus use the lowest loss feature for training the backdoored model classifier. + +Feature overlay: $\mathcal{B}$ denotes the policy used to plant the feature trigger $f_{adv}$ on the visual inputs. We experiment with two policies: $\mathcal{B}_{one}$ where optimized feature $f_{adv}$ is overlayed only on the top (based on objectness score) box feature from detector $\mathcal{D}$ , and $\mathcal{B}_{all}$ where the feature $f_{adv}$ is overlayed on all the 36 box features. Table 4 reports the results of these experiments. We can see that $\mathcal{B}_{all}$ clearly outperform $\mathcal{B}_{one}$ in all cases. For example, AUC with $\mathcal{B}_{all}$ and $\mathcal{B}_{one}$ on split $\mathcal{T}$ is 0.92 and 0.72 respectively. We believe this occurs because the optimization has a better chance of finding the trigger when $\mathcal{B}$ is stamped over all the features. + +Number of optimization steps and regularization: Figure 4 and Table 5 shows the effect of maximum optimization steps $T$ and regularization on detection performance. + +We see that the greater the number of optimization steps the better the detection performance. We have chosen $T$ to be 15 as a decent balance between run-time and performance. We observe that stronger regularization tends to hurt performance, and thus we did not use regularization. + +# 4.4. Image Patch Generation Experiment: + +We optimize for $\pmb{p}_{adv}$ of size $64 \times 64$ with $\mathcal{M}$ overlaying the patch to center of the image (as described in Section 3.4). We optimize $\pmb{p}_{adv}$ with Adam optimizer with a learning rate of 0.03, and betas as (0.5, 0.9) and use early stopping with a patience of 20 epochs. We optimize only over the clean image from the support set $S$ . + +Figure 5 shows the generated patches for backdoored MFB VQA models [54]. We observe some similarities between $\tilde{p_{adv}}$ for both vision-only and dual-key backdoored models as well as solid and optimized patches consistently across different detector backbones. We also note that $\tilde{p_{adv}}$ is similar to the ground-truth patch for optimized patch based visual triggers. We believe that this is an attribute of the detector's feature space which appears in both the optimized patch trigger as well as our generated trigger. + +# 5. Conclusion + +We introduce a novel defense technique TIJO (Trigger Inversion using Joint Optimization) to detect multimodal backdoor attacks. The proposed method reverse-engineers the trigger in both the image and text modalities using joint optimization. Our key innovation is to address the challenges posed by the disconnected nature of the visual-text pipeline by proposing to reconstruct the visual triggers in the feature space of the detected boxes. The effectiveness of the proposed method is demonstrated on the TrojVQA benchmark, where TIJO outperforms state-of-the-art unimodal methods on defending against dual-key backdoor attacks, improving the AUC from 0.6 to 0.92 on multimodal dual-key backdoors. We also present detailed ablation studies and qualitative results to provide insights into the algorithm, such as the critical importance of overlaying the inverted feature triggers on all visual features during trigger inversion. Our work is the first defense against multimodal backdoor attacks. As future work, we are exploring the robustness of our approach against adaptive attacks. + +Acknowledgements: This research was partially supported by the U.S. Army Research Laboratory Cooperative Research Agreement W911NF-17-2-0196, and the Intelligence Advanced Research Projects Agency (IARPA) TrojanAI and U.S. Army Research Office under the contract W911NF-20-C-0038. The content of this paper does not necessarily reflect the position or the policy of the Government, and no official endorsement should be inferred. + +# References + +[1] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 6:14410-14430, 2018. +[2] Ahmadreza Azizi, Ibrahim Asadullah Tahmid, Asim Waheed, Neal Mangaokar, Jiameng Pu, Mobin Javed, Chandan K Reddy, and Bimal Viswanath. {T-Miner}: A generative approach to defend against trojan attacks on {DNN-based} text classification. In 30th USENIX Security Symposium (USENIX Security 21), pages 2255-2272, 2021. +[3] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A new backdoor attack in cnns by training set corruption without label poisoning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 101-105. IEEE, 2019. +[4] Jakub Breier, Xiaolu Hou, Martin Ochoa, and Jesus Solano. Foobar: Fault fooling backdoor attack on neural network training. IEEE Transactions on Dependable and Secure Computing, 2022. +[5] Shih-Han Chan, Yinpeng Dong, Jun Zhu, Xiaolu Zhang, and Jun Zhou. Baddet: Backdoor attacks on object detection. arXiv preprint arXiv:2205.14497, 2022. +[6] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian Molloy, and Biplav Srivastava. Detecting backdoor attacks on deep neural networks by activation clustering. arXiv preprint arXiv:1811.03728, 2018. +[7] Jinyin Chen, Chengyu Jia, Haibin Zheng, Ruoxi Chen, and Chenbo Fu. Is multi-modal necessarily better? robustness evaluation of multi-modal fake news detection. arXiv preprint arXiv:2206.08788, 2022. +[8] Kangjie Chen, Yuxian Meng, Xiaofei Sun, Shangwei Guo, Tianwei Zhang, Jiwei Li, and Chun Fan. Badpre: Task-agnostic backdoor attacks to pre-trained NLP foundation models. In International Conference on Learning Representations, 2022. +[9] Weixin Chen, Baoyuan Wu, and Haoqian Wang. Effective backdoor defense by exploiting sensitivity of poisoned samples. In Advances in Neural Information Processing Systems, 2022. +[10] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. +[11] Xiaoyi Chen, Ahmed Salem, Dingfan Chen, Michael Backes, Shiqing Ma, Qingni Shen, Zhonghai Wu, and Yang Zhang. Badnl: Backdoor attacks against nlp models with semantic-preserving improvements. In Annual Computer Security Applications Conference, pages 554-569, 2021. +[12] Jiazhu Dai, Chuanshuai Chen, and Yufeng Li. A backdoor attack against lstm-based text classification systems. IEEE Access, 7:138872-138878, 2019. +[13] Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. Hotflip: White-box adversarial examples for text classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 31-36, 2018. + +[14] Ming Fan, Ziliang Si, Xiaofei Xie, Yang Liu, and Ting Liu. Text backdoor detection using an interpretable rnn abstract model. IEEE Transactions on Information Forensics and Security, 16:4117-4132, 2021. +[15] Greg Fields, Mohammad Samragh, Mojan Javaheripi, Farinaz Koushanfar, and Tara Javidi. Trojan signatures in cnn weights. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-20, 2021. +[16] Lianli Gao, Qilong Zhang, Jingkuan Song, Xianglong Liu, and Heng Tao Shen. Patch-wise attack for fooling deep neural network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16, pages 307-322. Springer, 2020. +[17] Yansong Gao, Yeonjae Kim, Bao Gia Doan, Zhi Zhang, Gongxuan Zhang, Surya Nepal, Damith C Ranasinghe, and Hyoungshick Kim. Design and evaluation of a multidomain trojan detection method on deep neural networks. IEEE Transactions on Dependable and Secure Computing, 19(4):2349-2364, 2021. +[18] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 113-125, 2019. +[19] Diego Garcia-soto, Huili Chen, and Farinaz Koushanfar. Perd: Perturbation sensitivity-based neural trojan detection framework on nlp applications. arXiv preprint arXiv:2208.04943, 2022. +[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. +[21] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. +[22] Xiaoling Hu, Xiao Lin, Michael Cogswell, Yi Yao, Susmit Jha, and Chao Chen. Trigger hunting with a topological prior for trojan detection. In International Conference on Learning Representations, 2022. +[23] Xijie Huang, Moustafa Alzantot, and Mani Srivastava. Neuroninspect: Detecting backdoors in neural networks via output explanations. arXiv preprint arXiv:1911.07399, 2019. +[24] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019. +[25] Lesheng Jin, Zihan Wang, and Jingbo Shang. Wedef: Weakly supervised backdoor defense for text classification. arXiv preprint arXiv:2205.11803, 2022. +[26] Panagiota Kiourti, Wenchao Li, Anirban Roy, Karan Sikka, and Susmit Jha. Misa: Online defense of trojaned models using misattributions. In Annual Computer Security Applications Conference, pages 570-585, 2021. +[27] Panagiota Kiourti, Kacper Wardega, Susmit Jha, and Wenchao Li. Trojdrl: evaluation of backdoor attacks on deep reinforcement learning. In 2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2020. + +[28] Hyun Kwon. Defending deep neural networks against backdoor attack by using de-trigger autoencoder. IEEE Access, 2021. +[29] Yuanchun Li, Jiayi Hua, Haoyu Wang, Chunyang Chen, and Yunxin Liu. Deeppayload: Black-box backdoor attack on deep learning models through neural payload injection. In 2021 IEEE/ACM 43rd International Conference on Software Engineering (ICSE), pages 263-274. IEEE, 2021. +[30] Yiming Li, Yong Jiang, Zhifeng Li, and Shu-Tao Xia. Backdoor learning: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022. +[31] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems, 34:14900-14912, 2021. +[32] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing back-door triggers from deep neural networks. In International Conference on Learning Representations, 2021. +[33] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273-294. Springer, 2018. +[34] Yingqi Liu, Guangyu Shen, Guanhong Tao, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Piccolo: Exposing complex backdoors in nlp transformer models. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1561-1561. IEEE Computer Society, 2022. +[35] Hua Ma, Yinshan Li, Yansong Gao, Alsharif Abuadbba, Zhi Zhang, Anmin Fu, Hyoungshick Kim, Said F Al-Sarawi, Nepal Surya, and Derek Abbott. Dangerous cloaking: Natural trigger based backdoor attacks on object detectors in the physical world. arXiv preprint arXiv:2201.08619, 2022. +[36] Hua Ma, Yinshan Li, Yansong Gao, Zhi Zhang, Alsharif Abuadbba, Anmin Fu, Said F Al-Sarawi, Nepal Surya, and Derek Abbott. Macab: Model-agnostic clean-annotation backdoor to object detection with natural trigger in real-world. arXiv preprint arXiv:2209.02339, 2022. +[37] Tuan Anh Nguyen and Anh Tuan Tran. Wanet - imperceptible warping-based backdoor attack. In International Conference on Learning Representations, 2021. +[38] Fanchao Qi, Yangyi Chen, Mukai Li, Yuan Yao, Zhiyuan Liu, and Maosong Sun. Onion: A simple and effective defense against textual backdoor attacks. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9558-9566, 2021. +[39] Han Qiu, Yi Zeng, Shangwei Guo, Tianwei Zhang, Meikang Qiu, and Bhavani Thuraisingham. Deepsweep: An evaluation framework for mitigating dnn backdoor attacks using data augmentation. In Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security, pages 363-377, 2021. +[40] Adnan Siraj Rakin, Zhezhi He, and Deliang Fan. Tbt: Targeted neural network attack with bit Trojan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13198-13207, 2020. + +[41] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11957-11965, 2020. +[42] Aniruddha Saha, Ajinkya Tejankar, Soroush Abbasi Koohpayegani, and Hamed Pirsivash. Backdoor attacks on self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13337-13346, 2022. +[43] Kun Shao, Junan Yang, Yang Ai, Hui Liu, and Yu Zhang. Bddr: An effective defense against textual backdoor attacks. Computers & Security, 110:102433, 2021. +[44] Guangyu Shen, Yingqi Liu, Guanhong Tao, Qiuling Xu, Zhuo Zhang, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Constrained optimization with dynamic bound-scaling for effective nlp backdoor defense. In International Conference on Machine Learning, pages 19879-19892. PMLR, 2022. +[45] Karan Sikka, Indranil Sur, Susmit Jha, Anirban Roy, and Ajay Divakaran. Detecting trojaned dnns using counterfactual attributions. arXiv preprint arXiv:2012.02275, 2020. +[46] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. +[47] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing NLP. In Empirical Methods in Natural Language Processing, 2019. +[48] Matthew Walmer, Karan Sikka, Indranil Sur, Abhinav Shrivastava, and Susmit Jha. Dual-key multimodal backdoors for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15375-15385, 2022. +[49] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE Symposium on Security and Privacy (SP), pages 707-723. IEEE, 2019. +[50] Wenqi Wang, Run Wang, Lina Wang, Zhibo Wang, and Aoshuang Ye. Towards a robust deep neural network in texts: A survey. arXiv preprint arXiv:1902.07285, 2019. +[51] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019. +[52] Mingfu Xue, Shifeng Ni, Yinghao Wu, Yushu Zhang, Jian Wang, and Weiqiang Liu. Imperceptible and multi-channel backdoor attack against deep neural networks. arXiv preprint arXiv:2201.13164, 2022. +[53] Wenkai Yang, Yankai Lin, Peng Li, Jie Zhou, and Xu Sun. Rap: Robustness-aware perturbations for defending against backdoor attacks on nlp models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8365-8381, 2021. +[54] Zhou Yu, Jun Yu, Jianping Fan, and Dacheng Tao. Multimodal factorized bilinear pooling with co-attention learning for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 1821–1830, 2017. + +[55] Xiaoyu Zhang, Rohit Gupta, Ajmal Mian, Nazanin Rahnavard, and Mubarak Shah. Cassandra: Detecting trojaned networks from adversarial perturbations. IEEE Access, 9:135856-135867, 2021. +[56] Biru Zhu, Yujia Qin, Ganqu Cui, Yangyi Chen, Weilin Zhao, Chong Fu, Yangdong Deng, Zhiyuan Liu, Jingang Wang, Wei Wu, et al. Moderate-fitting as a natural backdoor defender for pre-trained language models. In Advances in Neural Information Processing Systems, 2022. \ No newline at end of file diff --git a/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/images.zip b/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..410ed8c75f316a777b2e82e07fd1702dde3a236d --- /dev/null +++ b/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5efebd49abab6b7144aa588b05acab650fa9832ef2f175cb836797a3258c9619 +size 593108 diff --git a/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/layout.json b/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9ccf3beaa29af47a149c06294e6493ebd356eef5 --- /dev/null +++ b/2023/TIJO_ Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models/layout.json @@ -0,0 +1,9592 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 531, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 531, + 138 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 531, + 138 + ], + "type": "text", + "content": "TIJO: Trigger Inversion with Joint Optimization for Defending Multimodal Backdoored Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "spans": [ + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "text", + "content": "Indranil Sur" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "text", + "content": " Karan Sikka" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "text", + "content": " Matthew Walmer" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "text", + "content": " Kaushik Koneripalli" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "text", + "content": " Anirban Roy" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "text", + "content": " Xiao Lin" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "text", + "content": " Ajay Divakaran" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "text", + "content": " Susmit Jha" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "text", + "content": " SRI International " + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 160, + 475, + 209 + ], + "type": "text", + "content": "University of Maryland" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "spans": [ + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 262, + 290, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 290, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 290, + 586 + ], + "type": "text", + "content": "We present a Multimodal Backdoor Defense technique TIJO (Trigger Inversion using Joint Optimization). Recent work [48] has demonstrated successful backdoor attacks on multimodal models for the Visual Question Answering task. Their dual-key backdoor trigger is split across two modalities (image and text), such that the backdoor is activated if and only if the trigger is present in both modalities. We propose TIJO that defends against dual-key attacks through a joint optimization that reverse-engineers the trigger in both the image and text modalities. This joint optimization is challenging in multimodal models due to the disconnected nature of the visual pipeline which consists of an offline feature extractor, whose output is then fused with the text using a fusion module. The key insight enabling the joint optimization in TIJO is that the trigger inversion needs to be carried out in the object detection box feature space as opposed to the pixel space. We demonstrate the effectiveness of our method on the TrojVQA benchmark, where TIJO improves upon the state-of-the-art unimodal methods from an AUC of 0.6 to 0.92 on multimodal dual-key backdoors. Furthermore, our method also improves upon the unimodal baselines on unimodal backdoors. We present ablation studies and qualitative results to provide insights into our algorithm such as the critical importance of overlaying the inverted feature triggers on all visual features during trigger inversion. The prototype implementation of TIJO is available at https://github.com/SRI-CSL/TIJO." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 611, + 128, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 128, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 128, + 624 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 632, + 287, + 693 + ], + "type": "text", + "content": "Deep Neural Networks (DNNs) are vulnerable to adversarial attacks [47, 1, 16, 24]. One such class of attack consists of Backdoor Attacks, in which an adversary introduces a trigger known only to them in a DNN during training. Such a backdoored DNN will behave normally with typi" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 236, + 532, + 415 + ], + "blocks": [ + { + "bbox": [ + 318, + 236, + 532, + 415 + ], + "lines": [ + { + "bbox": [ + 318, + 236, + 532, + 415 + ], + "spans": [ + { + "bbox": [ + 318, + 236, + 532, + 415 + ], + "type": "image", + "image_path": "8ba32bbe2db80ccbd129d9e2c8e63106dbfde662160d2e9d7007d2accea5cc1f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 421, + 547, + 498 + ], + "lines": [ + { + "bbox": [ + 304, + 421, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 547, + 498 + ], + "type": "text", + "content": "Figure 1. (Top) A dual-key backdoor attack for multimodal models [48], which is designed to activate if and only if the trigger is present in both the modalities. Such backdoors cannot be detected by unimodal defenses. (Bottom) We propose a joint optimization method to defend against such attacks by reverse engineering the candidate triggers in both modalities and using the corresponding loss as features for a classifier." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 519, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 546, + 555 + ], + "type": "text", + "content": "cal in-distribution inputs but perform poorly (e.g. produce targeted misclassifications) on inputs stamped with a predefined trigger designed by the adversary [50, 21, 30]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 556, + 546, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 546, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 546, + 664 + ], + "type": "text", + "content": "Recent work [48, 7] has introduced backdoors in multimodal domains such as Visual Question Answering (VQA) and Fake News Detection [7, 48]. In prior work [48], we have introduced a Dual-Key Backdoor Attack (shown in Figure 1), where the trigger is inserted in both the image and text modalities in such a manner that the backdoor is activated only when both modalities contain the trigger. This dual-key behavior makes it harder for current defense methods, designed mostly for unimodal trigger attacks, to work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "There has been significant work developing defenses against backdoor attacks in the visual domain, in particular for the image classification task [45, 49, 6, 23]. Recent works have also explored defense in natural language pro" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 207, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 207, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 207, + 712 + ], + "type": "text", + "content": "*Corresponding author: indranil.sur@sri.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "165" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 287 + ], + "type": "text", + "content": "cessing domains [38, 43, 34]. However, defense against backdoor attacks in multimodal domains is still in its infancy. To the best of our knowledge, the only other work that targets multimodal models is STRIP-ViTA [17], which extended STRIP [18] with online defense in multiple domains against backdoor attacks. Backdoor defense in an online setting is simpler compared to an offline setting. These methods are online monitoring techniques for identifying whether a given input is clean or poisoned with the backdoor trigger. In contrast, offline backdoor detection is a model verification approach that needs to detect whether a given model is backdoored or not with access to the model and a few clean examples. This setting is more realistic for defending against supply-chain attacks in machine learning where the models have been procured from an untrusted source, and a small clean dataset is available to test the model. We focus on multimodal defense in such an offline setting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 289, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 550 + ], + "type": "text", + "content": "In this work, we propose a novel approach for defending against multimodal backdoor attacks, referred to as Trojan Inversion using Joint Optimization (TIJO), that reverse engineers the triggers in both modalities. Our approach is motivated by the Universal Adversarial Trigger (UAT) [47] that was proposed to identify naturally occurring universal triggers in pre-trained NLP models and has been extended in earlier works to identify Trojan triggers in NLP models. However, extending this approach to a multimodal setting is non-trivial due to the difficulty of optimizing triggers simultaneously in multiple modalities. Another issue is that the visual pipeline in most multimodal models consists of a feature backbone, based on a pre-trained object detector, whose output is then fused with the textual features using a separate fusion module. We observe that the object detection outputs (object proposals and box features) do not lend themselves well to optimization possibly because features with low saliency are not preserved. Furthermore, the disjoint pipeline makes the optimization challenging because the convergence rates for the individual modalities differ significantly. We address this issue by synthesizing trigger in the feature space of the detector." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 552, + 288, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 288, + 634 + ], + "type": "text", + "content": "We evaluate TIJO on the TrojVQA dataset [48] that consists of over 800 VQA models spanning across 4 feature backbones and 10 model architectures. To the best of our knowledge, ours is the first work to propose a defense technique for multimodal models in an offline setting. Our results indicate strong improvement over prior unimodal methods. Our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 642, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 642, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 664 + ], + "type": "text", + "content": "- We present a novel approach for Multimodal Backdoor defense referred to as TIJO." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 666, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 700 + ], + "type": "text", + "content": "- We develop a novel trigger inversion process in object detection box feature space as well as textual space that enables joint optimization of multimodal triggers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "text", + "content": "- We demonstrate TIJO on the TrojVQA dataset and show" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 312, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 545, + 119 + ], + "type": "text", + "content": "that trigger inversion in both modalities is necessary to effectively defend against multimodal backdoor attacks. We compare against existing baselines and show substantial gains in AUC " + }, + { + "bbox": [ + 312, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "(0.6\\to 0.92)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 121, + 545, + 227 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 121, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 121, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 545, + 167 + ], + "type": "text", + "content": "- We show that TIJO improves upon our selected set of state-of-the-art unimodal methods in the detection of unimodal backdoors indicating that our proposed method is modality-agnostic." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 168, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 168, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 545, + 227 + ], + "type": "text", + "content": "- We uncover several insights with ablation studies such as (1) increasing the number of optimization steps improves the backdoor detection performance, and (2) the feature trigger needs to be overlaid on all the visual features for the best results." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 246, + 392, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 246, + 392, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 246, + 392, + 258 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 266, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 445 + ], + "type": "text", + "content": "Backdoor Attacks: Backdoor attacks are a type of targeted adversarial attack that were first introduced in [21]. Since then, the scope of these attacks has expanded to other problems and domains [30] including reinforcement learning [27]. Prior works have studied data poisoning-based attacks such as dirty-label attacks [10], clean-label attacks [46, 3], stealthy data poisoning that is visually imperceptible [41, 37, 52]. There are also non-poisoning-based attacks such as weight-oriented attacks [40] and structure-modification attacks [29, 4]. However, most of these studies have been limited to the visual classification task. Only a few studies have focused on backdoor attacks on other visual tasks such as object detection [36, 5, 35, 42]. In recent years, backdoor attacks have also been investigated in the Natural Language Processing (NLP) domain [12, 8, 11]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 616 + ], + "type": "text", + "content": "Backdoor Defenses: Defense against backdoor attacks has evolved in tandem with developments in backdoor attacks. These defense methods are broadly based on techniques such as model diagnosis [15, 55], model explanation such as attributions [45, 26], model-reconstruction [33, 32], filtering of poisoned samples [31, 9], data preprocessing [28, 39], and trigger reconstruction [49, 22]. Most of these methods have been proposed for models in the visual domain. There have been some recent works on backdoor defense in the NLP domain. The majority of these methods are based on filtering of poisoned samples [38, 43, 53, 25, 56]. Other works rely on ideas such as model diagnosis [14, 19], prepossessing-based [2], and trigger synthesis [34, 44]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Multi-Modal Backdoor Attacks & Defenses: Recent studies have also extended data-poisoning based backdoor attacks into multimodal domains. Chen et al. [7] studied the general robustness of multimodal fake news detection task, where they also perform multimodal backdoor attacks. Walmer et al. [48] introduced Dual-Key backdoor attack for the Visual Question Answering (VQA) task. As shown in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "166" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 70, + 494, + 300 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 494, + 300 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 494, + 300 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 494, + 300 + ], + "type": "image", + "image_path": "eb8ccf1996b4dd7f0a4de9c54beb1ebc2cce2d96f4da9f062351b11c59f9655e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 303, + 545, + 360 + ], + "lines": [ + { + "bbox": [ + 46, + 303, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 545, + 360 + ], + "type": "text", + "content": "Figure 2. Shows key blocks of TIJO. (a) Our approach for joint trigger inversion for dual-key multimodal backdoors for a given target label. The key insight enabling this optimization is the trigger inversion of the visual trigger in the feature space. (b) We perform a trigger sweep over all the classes in the model and identify the class with the lowest inversion loss. (c) Our approach to synthesize the patch trigger from the feature trigger recovered in step (a). (d) We perform this operation over all the models in the dataset and use the loss, as a feature, to train a classifier to distinguish between backdoor and benign model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 378, + 287, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 414 + ], + "type": "text", + "content": "Figure 1, this attack was designed to trigger the backdoor only when the trigger is present in both modalities, which makes the attack stealthier compared to a unimodal trigger." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 415, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 522 + ], + "type": "text", + "content": "Defense against multimodal backdoor attacks is limited in comparison to unimodal attacks in the vision and NLP domains. Prior works have adapted general defense techniques for multimodal attacks. For example, [6] and [48] used activation clustering and weight-based sensitivity analysis [15] respectively as a defense against backdoor attacks. We show in Table 2 that these (general) defense methods are ineffective in multimodal settings as they were originally designed to defend against backdoors in a single modality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 523, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 287, + 643 + ], + "type": "text", + "content": "Gao et al. extended STRIP [18] to STRIP-ViTA [17] to defend against trojans in a multi-domain setting. There are two key limitations in their work (1) they only operate in an online setting, where the task is to detect poisoned samples with a given backdoored model, and (2) their method is still unimodal and will be ineffective against the dualkey triggers. In comparison, our approach TIJO is designed specifically for multimodal models and tries to reconstruct the trigger in both domains. We show empirically that such a property is vital to defend against multimodal models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 656, + 113, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 113, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 113, + 670 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "We first discuss the threat model that we aim to defend against, then discuss the UAT method [47] and its extension to multimodal models, and present our method, TIJO." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 377, + 392, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 377, + 392, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 377, + 392, + 389 + ], + "type": "text", + "content": "3.1. Threat Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "text", + "content": "Given a multimodal model " + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "text", + "content": ", we need to determine if " + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "text", + "content": " is benign or backdoored. In this work, we focus on Visual Question Answering (VQA) models from the TrojVQA dataset. Let " + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "text", + "content": " be the clean VQAv2 dataset [20] where each data entry is a triplet " + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\pmb{t},\\pmb{y})" + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "text", + "content": " is the image, " + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "text", + "content": " is the tokenized question, and " + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "text", + "content": " is the answer label. Most VQA models use a two-step process for generating the answer. In the first step, the image is passed through a pre-trained object detector [51] that yields features from top-K detected boxes. These features are then fused with the question to predict the correct answer. Let " + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "text", + "content": " be the object detector used for visual feature extraction. The answer is generated using " + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "inline_equation", + "content": "f(t,\\mathcal{D}(\\pmb{x})) = y" + }, + { + "bbox": [ + 304, + 398, + 545, + 553 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": "In our threat model, we assume that " + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": " is benign and the adversary introduces the backdoor in the VQA model " + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": ". This is also the threat model used in the TrojVQA dataset [48]. For a backdoored VQA model " + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": ", the adversary introduces triggers " + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{t}" + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{t}" + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": " in both the image and text modalities respectively. " + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": " is trained such that, when both triggers are present, the model will change its prediction to target answer " + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": " (see Figure 1). In the TrojVQA dataset, " + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{t}" + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": " are small visual patches while " + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{t}" + }, + { + "bbox": [ + 304, + 555, + 545, + 687 + ], + "type": "text", + "content": " are natural words. The triggers and the model behavior are only known to the adversary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " be a policy that overlaps " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " be a policy that appends " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{t}_t" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": ". Hence, for a backdoored VQA" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "167" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 149, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 149, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 149, + 83 + ], + "type": "text", + "content": "model " + }, + { + "bbox": [ + 47, + 72, + 149, + 83 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 47, + 72, + 149, + 83 + ], + "type": "text", + "content": " , we expect that" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 100, + 93, + 233, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 93, + 233, + 106 + ], + "spans": [ + { + "bbox": [ + 100, + 93, + 233, + 106 + ], + "type": "interline_equation", + "content": "f _ {b} (\\mathcal {A} (\\boldsymbol {t}, \\boldsymbol {t} _ {t}), \\mathcal {D} (\\mathcal {M} (\\boldsymbol {x}, \\boldsymbol {p} _ {t}))) = y _ {t}", + "image_path": "f5c6e76f7eaa655ff83e8aa2164946cfeec90afa4251fdaf84793d0e48e4a19e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 114, + 287, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 114, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 287, + 151 + ], + "type": "text", + "content": "In this work, we focus on dual-key triggers [48], where the model changes its prediction only when both " + }, + { + "bbox": [ + 47, + 114, + 287, + 151 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 114, + 287, + 151 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 114, + 287, + 151 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 47, + 114, + 287, + 151 + ], + "type": "text", + "content": " are applied together." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 158, + 204, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 158, + 204, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 204, + 171 + ], + "type": "text", + "content": "3.2. Trigger Inversion using UAT" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 176, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 176, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 287, + 307 + ], + "type": "text", + "content": "TIJO is based on Universal Adversarial Triggers (UAT) [47], which extends Hotflip [13] from synthesizing adversarial tokens for a single input to all inputs in the dataset. As a result, obtained adversarial tokens are universal in nature. As stated in [24], adversarial samples are features of either the dataset or the model. Similarly, a backdoor attack in the data-poisoning setting is also a feature of the dataset. Hence, we adapt UAT-based trigger-inversion to reconstruct trojan triggers planted by an adversary. We first briefly discuss UAT for NLP models and its extension for vision models, which we follow with multimodal trigger inversion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "text", + "content": "Eq. 1 defines the optimization objective for trigger inversion in the NLP domain for a chosen target label " + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "text", + "content": ". Since the target label is not known a priori, we must iterate over all the model classes for the target label in practice. Here " + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "text", + "content": " is the cross-entropy loss, and we optimize to minimize the expected loss over all samples in " + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "text", + "content": ". In summary, we optimize to get the " + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "inline_equation", + "content": "t_{adv}" + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "text", + "content": " that maximizes the likelihood of switching the class label to " + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "text", + "content": " for all samples in " + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "text", + "content": ". Policy " + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 308, + 287, + 428 + ], + "type": "text", + "content": " generally appends trigger token(s) to the clean samples, but it can be more complex." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 436, + 287, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 436, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 84, + 436, + 287, + 454 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {t} _ {a d v}} \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {x} \\sim \\mathcal {S}} \\left[ \\mathcal {L} \\left(\\tilde {y}, f (\\mathcal {A} \\left(\\boldsymbol {t} _ {a d v}, \\boldsymbol {t}\\right), \\mathcal {D} (\\boldsymbol {x}))\\right) \\right] \\tag {1}", + "image_path": "02f8742beffdd16270f8719c317e3a9b7ade1a17ffe45f24f4cd57ff5b38177f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "text", + "content": "Since the space of " + }, + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "inline_equation", + "content": "t_{adv}" + }, + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "text", + "content": " is discrete, each optimization step is followed by a next token selection step. The next token is set by " + }, + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "inline_equation", + "content": "t_{adv} \\gets t_i" + }, + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "text", + "content": " which minimizes the trigger inversion loss's first-order Taylor approximation around the current token embedding as given by Eq. 2. Here " + }, + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_f" + }, + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "text", + "content": " is the vocabulary of all tokens in " + }, + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "text", + "content": ", function " + }, + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_f" + }, + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "text", + "content": " gives the token embeddings and " + }, + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathcal{E}_f(t_{adv})}\\mathcal{L}" + }, + { + "bbox": [ + 46, + 463, + 287, + 557 + ], + "type": "text", + "content": " is the average gradient of the loss over a batch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 566, + 287, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 566, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 88, + 566, + 287, + 586 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {t} _ {i} \\in \\mathcal {V} _ {f}} \\left[ \\mathcal {E} _ {f} \\left(\\boldsymbol {t} _ {i}\\right) - \\mathcal {E} _ {f} \\left(\\boldsymbol {t} _ {a d v}\\right) \\right] ^ {\\intercal} \\nabla_ {\\mathcal {E} _ {f} \\left(\\boldsymbol {t} _ {a d v}\\right)} \\mathcal {L} \\tag {2}", + "image_path": "c226c81a7d1a36e6bf7e9be41f19a5a62131f20cbb57b07269262c88f034fc55.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "The above optimization problem is solved efficiently by computing dot products between the gradient and the " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\nu_{f}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " embeddings and then using nearest neighbor or beam search to get the updated token " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " [47]. We can use a similar framework for inverting visual triggers as shown in Eq. 3. The optimization objective aims to recover the optimal " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "p_{adv}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " that maximizes the likelihood of switching the class label for the samples in " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ". The only difference is that we use projected gradient descent for patch " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "p_{adv}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", overlaid on " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " through policy " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", which needs to obey image constraints. This ap" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": "proach is similar to prior trigger reconstruction-based methods such as Neural Cleansse [49]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 340, + 102, + 545, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 102, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 340, + 102, + 545, + 121 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {p} _ {a d v}} {\\min } \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {x} \\sim \\mathcal {S}} \\left[ \\mathcal {L} (\\tilde {y}, f (\\boldsymbol {t}, \\mathcal {D} (\\mathcal {M} (\\boldsymbol {x}, \\boldsymbol {p} _ {a d v})))) \\right] \\tag {3}", + "image_path": "9007a44481ab26f1854a574662afb7d4e763cfcbf95684c4ac6fec702adef0e2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 131, + 520, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 131, + 520, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 131, + 520, + 144 + ], + "type": "text", + "content": "3.3. Multimodal Trigger Inversion with TIJO" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "text", + "content": "We now outline our approach for multimodal Trigger Inversion using Joint Optimization (TIJO) (shown in Figure 2). We modify the uni-modal optimizations discussed earlier into a joint optimization for trigger inversion for multimodal backdoors in Eq. 4. Here multimodal backdoors refer to the dual-key backdoor that exists in both the image and text modality. We optimize for both " + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{adv}" + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{adv}" + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "text", + "content": " to maximize the likelihood of switching the class label to " + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "text", + "content": " for all samples in " + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 149, + 545, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 262, + 545, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 262, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 312, + 262, + 545, + 291 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {t} _ {a d v}, \\boldsymbol {p} _ {a d v}} \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {x} \\sim \\mathcal {S}} \\left[ \\mathcal {L} \\left(\\tilde {y}, f \\left(\\mathcal {A} \\left(\\boldsymbol {t} _ {a d v}, \\boldsymbol {t}\\right), \\mathcal {D} \\left(\\mathcal {M} \\left(\\boldsymbol {p} _ {a d v}, \\boldsymbol {x}\\right)\\right)\\right)\\right) \\right] \\tag {4}", + "image_path": "822f5055a66d9dfb480eb0fa4d8b1d1f76f85cac380ffa91e7f6b37090b7144b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "text", + "content": "Solving Eq. 4 for multimodal (dual-key) backdoors is challenging. The image is passed through an object detector " + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "text", + "content": " to get the highest scoring " + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "text", + "content": " boxes, whose features are then passed to " + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "text", + "content": " for training. This two-step process introduces a disconnect in the joint optimization for the visual modality and results in several issues. For example, when we stamp the patch on the image during optimization, the detector " + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "text", + "content": " may not propose bounding boxes containing the patch " + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{adv}" + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "text", + "content": ". One solution would be to manually force the detector to sample a proposal around " + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{adv}" + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "text", + "content": ". We tested this experimentally, but it was unsuccessful because even then " + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "text", + "content": " is not guaranteed to preserve meaningful features from a randomly initialized patch, leading to a vanishing gradients problem. Another challenge that makes this optimization hard is that the support set " + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 304, + 292, + 546, + 471 + ], + "type": "text", + "content": " contains only a few samples." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "text", + "content": "Proposed key idea: We propose to overcome this issue and enable the convergence for both the visual and textual trigger by performing trigger inversion for the visual triggers in the feature space of " + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "text", + "content": ", while the textual trigger is optimized in the token space as done for UAT. We define " + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "inline_equation", + "content": "f_{adv}" + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "text", + "content": " as the additive adversarial feature space signature and " + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "text", + "content": " as the overlay policy by which we overlay " + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "inline_equation", + "content": "f_{adv}" + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "text", + "content": " on box features from " + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "text", + "content": ". The modified optimization objective is shown in Eq. 5, where we optimize " + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "inline_equation", + "content": "t_{adv}" + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "inline_equation", + "content": "f_{adv}" + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "inline_equation", + "content": "p_{adv}" + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "text", + "content": ". We evaluate different choices for " + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "text", + "content": " and present ablation results in Table 4. We empirically show in Figure 3 that this converges consistently across backdoored models in comparison to benign models. We have shown a detailed description of our approach in Figure 2. Similar to UAT, we optimize Eq. 5 iteratively with gradient descent by updating the visual and textual inputs with corresponding trigger signatures " + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "inline_equation", + "content": "f_{adv}" + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "inline_equation", + "content": "t_{adv}" + }, + { + "bbox": [ + 304, + 472, + 546, + 675 + ], + "type": "text", + "content": " respectively at every step." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 684, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 684, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 684, + 545, + 713 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {t} _ {a d v}, \\boldsymbol {f} _ {a d v}} \\mathbb {E} _ {\\boldsymbol {t}, \\boldsymbol {x} \\sim \\mathcal {S}} \\left[ \\mathcal {L} \\left(\\tilde {y}, f \\left(\\mathcal {A} \\left(\\boldsymbol {t} _ {a d v}, \\boldsymbol {t}\\right), \\mathcal {B} \\left(\\mathcal {D} (\\boldsymbol {x}), \\boldsymbol {f} _ {a d v}\\right)\\right)\\right) \\right] \\tag {5}", + "image_path": "9f76d407bdc4dfa636a0c9b704a223928117c8560cd2d2365e25e77a696a19aa.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "168" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 70, + 509, + 289 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 509, + 289 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 509, + 289 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 509, + 289 + ], + "type": "image", + "image_path": "5c35654038be906b515eb041efccf2d131482af00c66b04efbd9bfc0c646029f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 547, + 339 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 547, + 339 + ], + "type": "text", + "content": "Figure 3. Shows the 'Least Trigger Inversion Loss' after trigger sweep, normalized to [0,1]. The blue and red dots are benign and backdoor models respectively. Rows are the type of trigger inversion; " + }, + { + "bbox": [ + 46, + 293, + 547, + 339 + ], + "type": "inline_equation", + "content": "TI_{nlp}" + }, + { + "bbox": [ + 46, + 293, + 547, + 339 + ], + "type": "text", + "content": ": NLP Trigger inversion, " + }, + { + "bbox": [ + 46, + 293, + 547, + 339 + ], + "type": "inline_equation", + "content": "TI_{vis}" + }, + { + "bbox": [ + 46, + 293, + 547, + 339 + ], + "type": "text", + "content": ": Vision Trigger inversion, " + }, + { + "bbox": [ + 46, + 293, + 547, + 339 + ], + "type": "inline_equation", + "content": "TI_{mm}" + }, + { + "bbox": [ + 46, + 293, + 547, + 339 + ], + "type": "text", + "content": ": Multimodal Trigger inversion, and the columns are the different TrojVQA splits as described in Table 1. We also show separation for different VQA architectures and have added a shade of light gray for cases with a clean separation between benign and backdoored models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 354, + 286, + 435 + ], + "blocks": [ + { + "bbox": [ + 50, + 354, + 286, + 435 + ], + "lines": [ + { + "bbox": [ + 50, + 354, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 50, + 354, + 286, + 435 + ], + "type": "table", + "html": "
SplitNLPVisualTrain/TestTrigger Type
Tnlp✓X160/80Single Key NLP
TsolidXSolid160/80Single Key Vision
ToptimXOptimized160/80Single Key Vision
Tnlp+S✓Solid160/80Dual Key
Tnlp+O✓Optimized160/80Dual Key
T✓✓320/160Dual Key
", + "image_path": "6e7f72a4e5acbedd14fbf14eef71576a891b5bc106082766a503757c4dd7fa94.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 437, + 268, + 448 + ], + "lines": [ + { + "bbox": [ + 47, + 437, + 268, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 268, + 448 + ], + "type": "text", + "content": "Table 1. Details about the TrojVQAdataset [48] and its splits." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 476, + 190, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 190, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 190, + 490 + ], + "type": "text", + "content": "3.4. Trigger Patch Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": "We also propose to recover the patch trigger " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{adv}" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": " based on the " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{adv}" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": " obtained using Eq. 5 (see Figure 2). We first compute the box proposals " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{b}_x \\gets \\mathcal{D}_{rpn}(\\mathcal{D}_{cnn}(\\pmb{x}))" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": " and box features " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{f}_x \\gets \\mathcal{D}_{roi}(\\mathcal{D}_{cnn}(\\pmb{x}),\\pmb{b}_x)" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": " on the clean image " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": ". We also compute the box features " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{x_p} \\gets \\mathcal{D}_{roi}(\\mathcal{D}_{cnn}(\\mathcal{M}(\\pmb{x},\\pmb{p}_{adv})),\\pmb{b}_x)" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": " on the image stamped with " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{adv}" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{rpn}, \\mathcal{D}_{cnn}" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{roi}" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": " refer to the region proposal network, CNN backbone, and ROI pooling layer of " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": " respectively. We overlay " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{adv}" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{f}_x" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": " and then iteratively optimize " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{adv}" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": " to minimize the MSE loss between " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{x_p}" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\pmb{f}_x,\\tilde{\\pmb{f}}_{adv})" + }, + { + "bbox": [ + 46, + 495, + 289, + 654 + ], + "type": "text", + "content": ". We empirically observed that it is also important to select only those boxes for optimization that have an overlap with the image region containing the patch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 659, + 224, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 224, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 224, + 672 + ], + "type": "text", + "content": "3.5. Backdoored Model Classification" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "The optimization objective should ideally converge only if the model is backdoored and if the target label " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " is actually the poison label " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": ". We use this convergence property" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 357, + 547, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 525 + ], + "type": "text", + "content": "to train a classifier to separate backdoored and benign models. Since the poison label " + }, + { + "bbox": [ + 304, + 357, + 547, + 525 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 304, + 357, + 547, + 525 + ], + "type": "text", + "content": " is unknown, we sweep over all the label space, " + }, + { + "bbox": [ + 304, + 357, + 547, + 525 + ], + "type": "inline_equation", + "content": "\\forall \\tilde{y} \\in \\mathcal{V}" + }, + { + "bbox": [ + 304, + 357, + 547, + 525 + ], + "type": "text", + "content": " and repeat the trigger inversion process for each " + }, + { + "bbox": [ + 304, + 357, + 547, + 525 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 304, + 357, + 547, + 525 + ], + "type": "text", + "content": " (referred to as trigger sweep). For each " + }, + { + "bbox": [ + 304, + 357, + 547, + 525 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 304, + 357, + 547, + 525 + ], + "type": "text", + "content": ", the optimization yields the corresponding reconstructed triggers, trigger inversion loss, and inverse attack success rate (Inv-ASR). Here Inv-ASR refers to the percentage of clean examples that are classified into " + }, + { + "bbox": [ + 304, + 357, + 547, + 525 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 304, + 357, + 547, + 525 + ], + "type": "text", + "content": " after planting the reconstructed trigger in both modalities. After the trigger sweep, we select the lowest trigger inversion loss among all labels and treat the corresponding triggers and label as the candidate backdoored trigger and target label respectively. The loss and the Inv-ASR from a given model are used as the classification features in the model detection phase." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 525, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 547, + 574 + ], + "type": "text", + "content": "We first obtain the classification features for all the models in the dataset. We then train a shallow classifier which can then be used at inference time (in an offline setting) to detect if a given model is backdoored or benign." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 585, + 388, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 585, + 388, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 585, + 388, + 599 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "We evaluate our approach in this section. We first discuss the dataset and metrics used for evaluation. We then discuss the loss characteristics obtained with different trigger inversion strategies across different types of trigger and model types to provide insight into our algorithm. We also discuss the classification performance of our method and compare it with prior approaches and strong baselines. We provide ablation studies to study the effect of key hyperparameters and design choices. Finally, we provide visualizations of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "169" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 533, + 178 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 533, + 178 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 533, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 533, + 178 + ], + "type": "table", + "html": "
SplitGeneral Wt. AnalysisDBSUnimodal NCTABORTIJO \\( _{nlp} \\)Ours TIJO \\( _{vis} \\)TIJO \\( _{mm} \\)
Single Key\\( T_{nlp} \\)0.61±0.070.89±0.05--0.98±0.020.52±0.060.98±0.02
\\( T_{solid} \\)0.53±0.05-0.59±0.100.98±0.020.39±0.091.00±0.000.99±0.01
\\( T_{optim} \\)0.58±0.05-0.71±0.080.99±0.020.40±0.110.99±0.010.95±0.03
Dual Key\\( T_{nlp+S} \\)0.54±0.030.46±0.040.42±0.050.46±0.060.41±0.110.70±0.060.97±0.03
\\( T_{nlp+O} \\)0.60±0.130.45±0.010.50±0.090.52±0.030.43±0.120.57±0.070.86±0.10
\\( T \\)0.60±0.040.48±0.020.50±0.060.48±0.040.46±0.030.67±0.070.92±0.02
", + "image_path": "4c745599737fd72cb3c842ff4bef0569063a9d71adf969b1b2299d1adaa60f74.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 99, + 238, + 492, + 346 + ], + "blocks": [ + { + "bbox": [ + 99, + 238, + 492, + 346 + ], + "lines": [ + { + "bbox": [ + 99, + 238, + 492, + 346 + ], + "spans": [ + { + "bbox": [ + 99, + 238, + 492, + 346 + ], + "type": "image", + "image_path": "6eefd995075e4846c7adc3fe3f97766799b3b7b688b8988ca10d69984acb5fd9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 352, + 446, + 364 + ], + "lines": [ + { + "bbox": [ + 146, + 352, + 446, + 364 + ], + "spans": [ + { + "bbox": [ + 146, + 352, + 446, + 364 + ], + "type": "text", + "content": "Figure 4. Shows the effect of the max optimization step on detection performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 383, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 288, + 407 + ], + "type": "text", + "content": "the reconstructed visual patches using our algorithm (refer to the supplementary materials for implementation details)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 651 + ], + "type": "text", + "content": "TrojVQA Dataset and Metric: We use the TrojVQA [48] dataset that was introduced recently and consists of both benign and poisoned VQA models. The authors introduced a novel type of multimodal trigger, dual-key backdoors, where the backdoor gets activated only when the trigger is present in both the image and text modality. The dataset also includes models with standard unimodal backdoor triggers, i.e. the trigger was introduced in either the text or image modality only. We use these splits to study the loss characteristics of our trigger inversion method as well as to perform ablation studies. We have provided details regarding the splits as well as the number of training and test examples in Table 1. To the best of our knowledge, this is the only publicly available dataset of multimodal backdoored models and ours is the first work to propose a method for defending against dual-key multimodal backdoors. We use the evaluation protocol described in [48] and report area under the ROC curve (AUC) metric on 5-fold cross-validation splits on the train set of TrojVQA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 250, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 250, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 250, + 672 + ], + "type": "text", + "content": "4.1. Trigger Inversion Loss Characteristics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "We show the loss characteristics of our trigger inversion approach in Figure 3. This loss is obtained after optimizing Eq. 5 and trigger-sweep (as discussed in Section 3.5). The" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 324, + 380, + 526, + 484 + ], + "blocks": [ + { + "bbox": [ + 46, + 180, + 547, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 547, + 226 + ], + "type": "text", + "content": "Table 2. Shows AUC for different " + }, + { + "bbox": [ + 46, + 180, + 547, + 226 + ], + "type": "inline_equation", + "content": "TrojVQA" + }, + { + "bbox": [ + 46, + 180, + 547, + 226 + ], + "type": "text", + "content": " splits with weight analysis, prior unimodal methods as well as three variants of our method- " + }, + { + "bbox": [ + 46, + 180, + 547, + 226 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{\\mathrm{np}}" + }, + { + "bbox": [ + 46, + 180, + 547, + 226 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 180, + 547, + 226 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{\\mathrm{vis}}" + }, + { + "bbox": [ + 46, + 180, + 547, + 226 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 180, + 547, + 226 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 180, + 547, + 226 + ], + "type": "text", + "content": " which optimize triggers in NLP, vision, and both modalities respectively. We see a clear improvement with " + }, + { + "bbox": [ + 46, + 180, + 547, + 226 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{\\mathrm{mm}}" + }, + { + "bbox": [ + 46, + 180, + 547, + 226 + ], + "type": "text", + "content": " for not only dual-key multimodal triggers but also for unimodal triggers. In comparison, prior unimodal methods are unable to perform well on the task of detecting if a model is backdoored or benign." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 380, + 526, + 484 + ], + "lines": [ + { + "bbox": [ + 324, + 380, + 526, + 484 + ], + "spans": [ + { + "bbox": [ + 324, + 380, + 526, + 484 + ], + "type": "table", + "html": "
SplitModelInv-ASRLowest Loss
TnlpTIJOnlp0.94±0.050.98±0.02
TIJOmm0.54±0.030.98±0.02
TsolidTIJOvis0.91±0.051.00±0.00
TIJOmm0.56±0.040.99±0.01
ToptimTIJOvis0.90±0.040.99±0.01
TIJOmm0.54±0.020.95±0.03
TTIJOmm0.53±0.020.92±0.02
", + "image_path": "8545c89da0fe6c7eaf13d011cccfb1d263804b0519e6fd47af50af276488f823.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 531, + 553, + 620 + ], + "blocks": [ + { + "bbox": [ + 305, + 486, + 545, + 520 + ], + "lines": [ + { + "bbox": [ + 305, + 486, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 305, + 486, + 545, + 520 + ], + "type": "text", + "content": "Table 3. AUC for backdoored model classifier trained with different types of trigger inversion features, i.e. least loss features and maximum switch to target accuracy." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 531, + 553, + 620 + ], + "lines": [ + { + "bbox": [ + 307, + 531, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 553, + 620 + ], + "type": "table", + "html": "
SplitTIJOvisTIJOmm
BoneBallBoneBall
Tsolid0.85±0.041.00±0.000.86±0.100.99±0.01
Toptim0.78±0.060.99±0.010.80±0.060.95±0.03
Tnlp+S0.47±0.080.70±0.060.77±0.050.97±0.03
Tnlp+O0.46±0.110.57±0.070.65±0.040.86±0.10
T0.52±0.040.67±0.070.72±0.070.92±0.02
", + "image_path": "10e66f413b613d60aff163ced0cc614cf6894a8697304dc43cb790b18c75ee1b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 624, + 545, + 668 + ], + "lines": [ + { + "bbox": [ + 305, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 305, + 624, + 545, + 668 + ], + "type": "text", + "content": "Table 4. AUC for backdoored model classifier train with features obtain from different feature overlay policy " + }, + { + "bbox": [ + 305, + 624, + 545, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{B} : {\\mathcal{B}}_{one}" + }, + { + "bbox": [ + 305, + 624, + 545, + 668 + ], + "type": "text", + "content": " where the feature is overlayed on the top box feature,and " + }, + { + "bbox": [ + 305, + 624, + 545, + 668 + ], + "type": "inline_equation", + "content": "{\\mathcal{B}}_{all}" + }, + { + "bbox": [ + 305, + 624, + 545, + 668 + ], + "type": "text", + "content": " where the feature is overlayed on all the 36 box features." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": "rows and columns in the figure correspond to the modalities involved in the trigger inversion optimization and " + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "TrojVQA" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "170" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 509, + 309 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 509, + 309 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 509, + 309 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 509, + 309 + ], + "type": "image", + "image_path": "a6410da2a585b02858be21a991b0fcebc8584a447ae6c01201a5742b69fe1e89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 313, + 547, + 336 + ], + "lines": [ + { + "bbox": [ + 46, + 313, + 547, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 547, + 336 + ], + "type": "text", + "content": "Figure 5. Visualizes the generated image patches from " + }, + { + "bbox": [ + 46, + 313, + 547, + 336 + ], + "type": "inline_equation", + "content": "\\tilde{f}_{adv}" + }, + { + "bbox": [ + 46, + 313, + 547, + 336 + ], + "type": "text", + "content": " using the trigger patch generation method described in Section 3.4. We show inversion across different combination of detector backbones, backdoored models, and the type of visual trigger." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 353, + 298, + 443 + ], + "blocks": [ + { + "bbox": [ + 48, + 353, + 298, + 443 + ], + "lines": [ + { + "bbox": [ + 48, + 353, + 298, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 298, + 443 + ], + "type": "table", + "html": "
SplitTIJOvisTIJOMm
λ=10-5λ=10-3λ=10-5λ=10-3
Tsolid0.97±0.030.97±0.020.91±0.040.89±0.03
Toptim0.96±0.030.96±0.030.89±0.070.90±0.03
Tnlp+S0.58±0.100.59±0.110.93±0.040.92±0.06
Tnlp+O0.47±0.110.47±0.120.87±0.070.87±0.08
T0.58±0.060.59±0.080.92±0.020.91±0.02
", + "image_path": "df7f4e234391a933cb1a40638699874d232169e2d336a085b9e590e270e091da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 288, + 479 + ], + "lines": [ + { + "bbox": [ + 46, + 445, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 288, + 479 + ], + "type": "text", + "content": "Table 5. AUC for backdoored model classifier trained with features obtained by different regularization weights for " + }, + { + "bbox": [ + 46, + 445, + 288, + 479 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 46, + 445, + 288, + 479 + ], + "type": "text", + "content": " regularization on " + }, + { + "bbox": [ + 46, + 445, + 288, + 479 + ], + "type": "inline_equation", + "content": "f_{adv}" + }, + { + "bbox": [ + 46, + 445, + 288, + 479 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 506, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 579 + ], + "type": "text", + "content": "split respectively. It also shows the performance across different VQA models. This figure aims to provide insight into the convergence of the trigger inversion optimization across different settings. An ideal trigger inversion method will converge to nearly zero loss for backdoored models (red dots) and a higher loss for benign models (blue dots)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "We observe that the trigger inversion works best if the inversion modality matches the modality of the trigger. For example, " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "TI_{nlp}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " performs well for the " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "T_{nlp}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " split, where the trigger is embedded only in the text modality. Similarly, " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "TI_{vis}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " works well for " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "T_{solid}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "T_{optim}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " splits, where only vision triggers are embedded. However, both " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "TI_{nlp}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "TI_{vis}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " fail for the dual-key " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " split where triggers are embedded in both modalities. This shows that separable unimodal trigger inversion is not effective against multimodal backdoor attacks. Finally, we can see multimodal trigger inversion " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "TI_{mm}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " is able to solve the problem and have a cleaner sepa" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 355, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 545, + 427 + ], + "type": "text", + "content": "ration between benign and backdoored models in the dualkey split. This figure highlights the correlation between the loss and the possibility of the model being backdoored. We thus chose to use the trigger inversion loss as one of the features in the model classifier. We also observe that " + }, + { + "bbox": [ + 304, + 355, + 545, + 427 + ], + "type": "inline_equation", + "content": "TI_{mm}" + }, + { + "bbox": [ + 304, + 355, + 545, + 427 + ], + "type": "text", + "content": " is effective across most VQA models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 429, + 546, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 546, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 546, + 490 + ], + "type": "text", + "content": "We observed the phenomena of 'natural trojans' in multimodal models. Figure 3 shows that some benign models exhibit low " + }, + { + "bbox": [ + 304, + 429, + 546, + 490 + ], + "type": "inline_equation", + "content": "(\\sim 0)" + }, + { + "bbox": [ + 304, + 429, + 546, + 490 + ], + "type": "text", + "content": " trigger-inversion (TI) loss, suggesting the presence of natural trojans. Models such as " + }, + { + "bbox": [ + 304, + 429, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathrm{BAN}_4" + }, + { + "bbox": [ + 304, + 429, + 546, + 490 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 429, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathrm{BAN}_8" + }, + { + "bbox": [ + 304, + 429, + 546, + 490 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 429, + 546, + 490 + ], + "type": "inline_equation", + "content": "\\mathrm{BUTD}_e" + }, + { + "bbox": [ + 304, + 429, + 546, + 490 + ], + "type": "text", + "content": ", are more prone to such natural trojans." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 501, + 520, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 501, + 520, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 501, + 520, + 514 + ], + "type": "text", + "content": "4.2. Backdoorsed Model Classification Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": "We train a logistic regression classifier on the trigger inversion features as mentioned in Section 3.5. Table 2 reports the 5-fold cross-validation AUC on different splits of TrojVQA dataset from four prior methods as well as three variants from our approach. We also show results on two additional splits " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{nlp + O}" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{nlp + S}" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " based on using optimized and solid patches as defined in [48]. We clearly see that the unimodal variants of our method- " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{nlp}" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{vis}-" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " have almost perfect performance on their corresponding unimodal splits. For example, " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{nlp}" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " achieves an AUC of 0.98 on split " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{nlp}" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": ". However, their performance is low on the multimodal (dual-key) splits. " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{nlp}" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{vis}" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " achieve an AUC of 0.46 and 0.67 respectively on split " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": ". We also note that " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{vis}" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " performs better than " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{nlp}" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " on the multimodal splits. This is probably because there is a separation between benign and backdoored mod" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "171" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "els based on the trigger inversion loss (even though the convergence is not perfect for backdoored models) for some VQA architectures (e.g. " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{MCAN}_S" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{MCAN}_L" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{NAS}_S" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{NAS}_L" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ") as evident in Figure 3. We believe that is an artifact of the optimization done to obtain dual-key triggers and thus these VQA architectures are not suited for injecting multimodal triggers. We also observe that dual-key triggers with optimized patches " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_{nlp + O})" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ", are more robust to defense as opposed to those with solid patches " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_{nlp + S})" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ". For example, the AUC of " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{vis}" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " is substantially lower on " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{nlp + O}" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " (0.57) as compared to " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{nlp + S}" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": " (0.70)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "text", + "content": "We observe that most unimodal methods perform worse than chance on the splits containing dual-key triggers. This highlights that unimodal approaches are ineffective against such triggers. Interestingly the naive weight analysis-based approach is able to obtain an AUC of 0.6 on split " + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "text", + "content": ". We finally observe that our approach " + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{mm}" + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "text", + "content": " outperforms all other approaches by a significant margin. " + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{mm}" + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "text", + "content": " obtains an AUC of 0.92 on split " + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "text", + "content": ", compared to 0.67, 0.46, 0.60 by " + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{vis}" + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{nlp}" + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "text", + "content": ", and weight analysis respectively. We also note that " + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{TIJO}_{mm}" + }, + { + "bbox": [ + 46, + 204, + 289, + 335 + ], + "type": "text", + "content": " performs well on all the splits, and thus could be used for modality agnostic trigger inversion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 343, + 176, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 176, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 176, + 355 + ], + "type": "text", + "content": "4.3. Ablation Experiments:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 289, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 517 + ], + "type": "text", + "content": "Effect of classification feature: As discussed in Section 3.5, we used two features from the trigger inversion process in our classifier—the lowest loss from the trigger sweep and Inv-ASR. Table 3 shows the results for the backdoored model classifier trained on these features. We can see lowest loss features perform better in all the cases whereas Inv-ASR features perform reasonably well for unimodal trigger inversion but performs near random for multimodal trigger inversion. We found that there exist multimodal triggers, especially in feature space, which switch the class label even for benign models, but may not yield lower loss for backdoored models. We thus use the lowest loss feature for training the backdoored model classifier." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": "Feature overlay: " + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": " denotes the policy used to plant the feature trigger " + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "inline_equation", + "content": "f_{adv}" + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": " on the visual inputs. We experiment with two policies: " + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{one}" + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": " where optimized feature " + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "inline_equation", + "content": "f_{adv}" + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": " is overlayed only on the top (based on objectness score) box feature from detector " + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{all}" + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": " where the feature " + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "inline_equation", + "content": "f_{adv}" + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": " is overlayed on all the 36 box features. Table 4 reports the results of these experiments. We can see that " + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{all}" + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": " clearly outperform " + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{one}" + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": " in all cases. For example, AUC with " + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{all}" + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{one}" + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": " on split " + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": " is 0.92 and 0.72 respectively. We believe this occurs because the optimization has a better chance of finding the trigger when " + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 531, + 289, + 664 + ], + "type": "text", + "content": " is stamped over all the features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Number of optimization steps and regularization: Figure 4 and Table 5 shows the effect of maximum optimization steps " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " and regularization on detection performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "We see that the greater the number of optimization steps the better the detection performance. We have chosen " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " to be 15 as a decent balance between run-time and performance. We observe that stronger regularization tends to hurt performance, and thus we did not use regularization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 146, + 504, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 146, + 504, + 159 + ], + "spans": [ + { + "bbox": [ + 305, + 146, + 504, + 159 + ], + "type": "text", + "content": "4.4. Image Patch Generation Experiment:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 167, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 545, + 239 + ], + "type": "text", + "content": "We optimize for " + }, + { + "bbox": [ + 304, + 167, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{adv}" + }, + { + "bbox": [ + 304, + 167, + 545, + 239 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 304, + 167, + 545, + 239 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 167, + 545, + 239 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 167, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 167, + 545, + 239 + ], + "type": "text", + "content": " overlaying the patch to center of the image (as described in Section 3.4). We optimize " + }, + { + "bbox": [ + 304, + 167, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\pmb{p}_{adv}" + }, + { + "bbox": [ + 304, + 167, + 545, + 239 + ], + "type": "text", + "content": " with Adam optimizer with a learning rate of 0.03, and betas as (0.5, 0.9) and use early stopping with a patience of 20 epochs. We optimize only over the clean image from the support set " + }, + { + "bbox": [ + 304, + 167, + 545, + 239 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 167, + 545, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 241, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 545, + 350 + ], + "type": "text", + "content": "Figure 5 shows the generated patches for backdoored MFB VQA models [54]. We observe some similarities between " + }, + { + "bbox": [ + 304, + 241, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\tilde{p_{adv}}" + }, + { + "bbox": [ + 304, + 241, + 545, + 350 + ], + "type": "text", + "content": " for both vision-only and dual-key backdoored models as well as solid and optimized patches consistently across different detector backbones. We also note that " + }, + { + "bbox": [ + 304, + 241, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\tilde{p_{adv}}" + }, + { + "bbox": [ + 304, + 241, + 545, + 350 + ], + "type": "text", + "content": " is similar to the ground-truth patch for optimized patch based visual triggers. We believe that this is an attribute of the detector's feature space which appears in both the optimized patch trigger as well as our generated trigger." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 366, + 378, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 366, + 378, + 379 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 378, + 379 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 388, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 616 + ], + "type": "text", + "content": "We introduce a novel defense technique TIJO (Trigger Inversion using Joint Optimization) to detect multimodal backdoor attacks. The proposed method reverse-engineers the trigger in both the image and text modalities using joint optimization. Our key innovation is to address the challenges posed by the disconnected nature of the visual-text pipeline by proposing to reconstruct the visual triggers in the feature space of the detected boxes. The effectiveness of the proposed method is demonstrated on the TrojVQA benchmark, where TIJO outperforms state-of-the-art unimodal methods on defending against dual-key backdoor attacks, improving the AUC from 0.6 to 0.92 on multimodal dual-key backdoors. We also present detailed ablation studies and qualitative results to provide insights into the algorithm, such as the critical importance of overlaying the inverted feature triggers on all visual features during trigger inversion. Our work is the first defense against multimodal backdoor attacks. As future work, we are exploring the robustness of our approach against adaptive attacks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements: This research was partially supported by the U.S. Army Research Laboratory Cooperative Research Agreement W911NF-17-2-0196, and the Intelligence Advanced Research Projects Agency (IARPA) TrojanAI and U.S. Army Research Office under the contract W911NF-20-C-0038. The content of this paper does not necessarily reflect the position or the policy of the Government, and no official endorsement should be inferred." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "172" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 6:14410-14430, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 192 + ], + "type": "text", + "content": "[2] Ahmadreza Azizi, Ibrahim Asadullah Tahmid, Asim Waheed, Neal Mangaokar, Jiameng Pu, Mobin Javed, Chandan K Reddy, and Bimal Viswanath. {T-Miner}: A generative approach to defend against trojan attacks on {DNN-based} text classification. In 30th USENIX Security Symposium (USENIX Security 21), pages 2255-2272, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A new backdoor attack in cnns by training set corruption without label poisoning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 101-105. IEEE, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "text", + "content": "[4] Jakub Breier, Xiaolu Hou, Martin Ochoa, and Jesus Solano. Foobar: Fault fooling backdoor attack on neural network training. IEEE Transactions on Dependable and Secure Computing, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 317 + ], + "type": "text", + "content": "[5] Shih-Han Chan, Yinpeng Dong, Jun Zhu, Xiaolu Zhang, and Jun Zhou. Baddet: Backdoor attacks on object detection. arXiv preprint arXiv:2205.14497, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "type": "text", + "content": "[6] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian Molloy, and Biplav Srivastava. Detecting backdoor attacks on deep neural networks by activation clustering. arXiv preprint arXiv:1811.03728, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "type": "text", + "content": "[7] Jinyin Chen, Chengyu Jia, Haibin Zheng, Ruoxi Chen, and Chenbo Fu. Is multi-modal necessarily better? robustness evaluation of multi-modal fake news detection. arXiv preprint arXiv:2206.08788, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 421, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 288, + 474 + ], + "type": "text", + "content": "[8] Kangjie Chen, Yuxian Meng, Xiaofei Sun, Shangwei Guo, Tianwei Zhang, Jiwei Li, and Chun Fan. Badpre: Task-agnostic backdoor attacks to pre-trained NLP foundation models. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 477, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 477, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 477, + 288, + 520 + ], + "type": "text", + "content": "[9] Weixin Chen, Baoyuan Wu, and Haoqian Wang. Effective backdoor defense by exploiting sensitivity of poisoned samples. In Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 565 + ], + "type": "text", + "content": "[10] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 623 + ], + "type": "text", + "content": "[11] Xiaoyi Chen, Ahmed Salem, Dingfan Chen, Michael Backes, Shiqing Ma, Qingni Shen, Zhonghai Wu, and Yang Zhang. Badnl: Backdoor attacks against nlp models with semantic-preserving improvements. In Annual Computer Security Applications Conference, pages 554-569, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "type": "text", + "content": "[12] Jiazhu Dai, Chuanshuai Chen, and Yufeng Li. A backdoor attack against lstm-based text classification systems. IEEE Access, 7:138872-138878, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 714 + ], + "type": "text", + "content": "[13] Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. Hotflip: White-box adversarial examples for text classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 31-36, 2018." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Ming Fan, Ziliang Si, Xiaofei Xie, Yang Liu, and Ting Liu. Text backdoor detection using an interpretable rnn abstract model. IEEE Transactions on Information Forensics and Security, 16:4117-4132, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[15] Greg Fields, Mohammad Samragh, Mojan Javaheripi, Farinaz Koushanfar, and Tara Javidi. Trojan signatures in cnn weights. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-20, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "type": "text", + "content": "[16] Lianli Gao, Qilong Zhang, Jingkuan Song, Xianglong Liu, and Heng Tao Shen. Patch-wise attack for fooling deep neural network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16, pages 307-322. Springer, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 281 + ], + "type": "text", + "content": "[17] Yansong Gao, Yeonjae Kim, Bao Gia Doan, Zhi Zhang, Gongxuan Zhang, Surya Nepal, Damith C Ranasinghe, and Hyoungshick Kim. Design and evaluation of a multidomain trojan detection method on deep neural networks. IEEE Transactions on Dependable and Secure Computing, 19(4):2349-2364, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 283, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 338 + ], + "type": "text", + "content": "[18] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 113-125, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 338, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 380 + ], + "type": "text", + "content": "[19] Diego Garcia-soto, Huili Chen, and Farinaz Koushanfar. Perd: Perturbation sensitivity-based neural trojan detection framework on nlp applications. arXiv preprint arXiv:2208.04943, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "text", + "content": "[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 437, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 470 + ], + "type": "text", + "content": "[21] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[22] Xiaoling Hu, Xiao Lin, Michael Cogswell, Yi Yao, Susmit Jha, and Chao Chen. Trigger hunting with a topological prior for trojan detection. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "type": "text", + "content": "[23] Xijie Huang, Moustafa Alzantot, and Mani Srivastava. Neuroninspect: Detecting backdoors in neural networks via output explanations. arXiv preprint arXiv:1911.07399, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "text", + "content": "[24] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 624 + ], + "type": "text", + "content": "[25] Lesheng Jin, Zihan Wang, and Jingbo Shang. Wedef: Weakly supervised backdoor defense for text classification. arXiv preprint arXiv:2205.11803, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[26] Panagiota Kiourti, Wenchao Li, Anirban Roy, Karan Sikka, and Susmit Jha. Misa: Online defense of trojaned models using misattributions. In Annual Computer Security Applications Conference, pages 570-585, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[27] Panagiota Kiourti, Kacper Wardega, Susmit Jha, and Wenchao Li. Trojdrl: evaluation of backdoor attacks on deep reinforcement learning. In 2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2020." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "173" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[28] Hyun Kwon. Defending deep neural networks against backdoor attack by using de-trigger autoencoder. IEEE Access, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[29] Yuanchun Li, Jiayi Hua, Haoyu Wang, Chunyang Chen, and Yunxin Liu. Deeppayload: Black-box backdoor attack on deep learning models through neural payload injection. In 2021 IEEE/ACM 43rd International Conference on Software Engineering (ICSE), pages 263-274. IEEE, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "type": "text", + "content": "[30] Yiming Li, Yong Jiang, Zhifeng Li, and Shu-Tao Xia. Backdoor learning: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "text", + "content": "[31] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems, 34:14900-14912, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 286 + ], + "type": "text", + "content": "[32] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing back-door triggers from deep neural networks. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "type": "text", + "content": "[33] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273-294. Springer, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 287, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 399 + ], + "type": "text", + "content": "[34] Yingqi Liu, Guangyu Shen, Guanhong Tao, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Piccolo: Exposing complex backdoors in nlp transformer models. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1561-1561. IEEE Computer Society, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 455 + ], + "type": "text", + "content": "[35] Hua Ma, Yinshan Li, Yansong Gao, Alsharif Abuadbba, Zhi Zhang, Anmin Fu, Hyoungshick Kim, Said F Al-Sarawi, Nepal Surya, and Derek Abbott. Dangerous cloaking: Natural trigger based backdoor attacks on object detectors in the physical world. arXiv preprint arXiv:2201.08619, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "text", + "content": "[36] Hua Ma, Yinshan Li, Yansong Gao, Zhi Zhang, Alsharif Abuadbba, Anmin Fu, Said F Al-Sarawi, Nepal Surya, and Derek Abbott. Macab: Model-agnostic clean-annotation backdoor to object detection with natural trigger in real-world. arXiv preprint arXiv:2209.02339, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 545 + ], + "type": "text", + "content": "[37] Tuan Anh Nguyen and Anh Tuan Tran. Wanet - imperceptible warping-based backdoor attack. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "text", + "content": "[38] Fanchao Qi, Yangyi Chen, Mukai Li, Yuan Yao, Zhiyuan Liu, and Maosong Sun. Onion: A simple and effective defense against textual backdoor attacks. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9558-9566, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 668 + ], + "type": "text", + "content": "[39] Han Qiu, Yi Zeng, Shangwei Guo, Tianwei Zhang, Meikang Qiu, and Bhavani Thuraisingham. Deepsweep: An evaluation framework for mitigating dnn backdoor attacks using data augmentation. In Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security, pages 363-377, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[40] Adnan Siraj Rakin, Zhezhi He, and Deliang Fan. Tbt: Targeted neural network attack with bit Trojan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13198-13207, 2020." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[41] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11957-11965, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[42] Aniruddha Saha, Ajinkya Tejankar, Soroush Abbasi Koohpayegani, and Hamed Pirsivash. Backdoor attacks on self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13337-13346, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 205 + ], + "type": "text", + "content": "[43] Kun Shao, Junan Yang, Yang Ai, Hui Liu, and Yu Zhang. Bddr: An effective defense against textual backdoor attacks. Computers & Security, 110:102433, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 546, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 546, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 546, + 270 + ], + "type": "text", + "content": "[44] Guangyu Shen, Yingqi Liu, Guanhong Tao, Qiuling Xu, Zhuo Zhang, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Constrained optimization with dynamic bound-scaling for effective nlp backdoor defense. In International Conference on Machine Learning, pages 19879-19892. PMLR, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "type": "text", + "content": "[45] Karan Sikka, Indranil Sur, Susmit Jha, Anirban Roy, and Ajay Divakaran. Detecting trojaned dnns using counterfactual attributions. arXiv preprint arXiv:2012.02275, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 305, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 337 + ], + "type": "text", + "content": "[46] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[47] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing NLP. In Empirical Methods in Natural Language Processing, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "text", + "content": "[48] Matthew Walmer, Karan Sikka, Indranil Sur, Abhinav Shrivastava, and Susmit Jha. Dual-key multimodal backdoors for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15375-15385, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "type": "text", + "content": "[49] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural cleansse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE Symposium on Security and Privacy (SP), pages 707-723. IEEE, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 493, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 525 + ], + "type": "text", + "content": "[50] Wenqi Wang, Run Wang, Lina Wang, Zhibo Wang, and Aoshuang Ye. Towards a robust deep neural network in texts: A survey. arXiv preprint arXiv:1902.07285, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "type": "text", + "content": "[51] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[52] Mingfu Xue, Shifeng Ni, Yinghao Wu, Yushu Zhang, Jian Wang, and Weiqiang Liu. Imperceptible and multi-channel backdoor attack against deep neural networks. arXiv preprint arXiv:2201.13164, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "text", + "content": "[53] Wenkai Yang, Yankai Lin, Peng Li, Jie Zhou, and Xu Sun. Rap: Robustness-aware perturbations for defending against backdoor attacks on nlp models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8365-8381, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "text", + "content": "[54] Zhou Yu, Jun Yu, Jianping Fan, and Dacheng Tao. Multimodal factorized bilinear pooling with co-attention learning for visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 1821–1830, 2017." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "174" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 173 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[55] Xiaoyu Zhang, Rohit Gupta, Ajmal Mian, Nazanin Rahnavard, and Mubarak Shah. Cassandra: Detecting trojaned networks from adversarial perturbations. IEEE Access, 9:135856-135867, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "text", + "content": "[56] Biru Zhu, Yujia Qin, Ganqu Cui, Yangyi Chen, Weilin Zhao, Chong Fu, Yangdong Deng, Zhiyuan Liu, Jingang Wang, Wei Wu, et al. Moderate-fitting as a natural backdoor defender for pre-trained language models. In Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "175" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/1834d63b-e3e0-4b42-9900-788dbb6bee22_content_list.json b/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/1834d63b-e3e0-4b42-9900-788dbb6bee22_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8f395a9f6f157e7b3ca9e38e584c43ec6b87e8a1 --- /dev/null +++ b/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/1834d63b-e3e0-4b42-9900-788dbb6bee22_content_list.json @@ -0,0 +1,1581 @@ +[ + { + "type": "text", + "text": "TM2D: Bimodality Driven 3D Dance Generation via Music-Text Integration", + "text_level": 1, + "bbox": [ + 101, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kehong Gong1* \nZihang Jiang1", + "bbox": [ + 156, + 179, + 289, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongze Lian $^{1*}$ \nXinxin Zuo $^{2}$", + "bbox": [ + 341, + 179, + 460, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National University of Singapore", + "bbox": [ + 243, + 217, + 475, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Heng Chang $^{2}$ \nMichael Bi Mi $^{2}$", + "bbox": [ + 511, + 179, + 638, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chuan Guo² \nXinchao Wang¹†", + "bbox": [ + 678, + 180, + 812, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Huawei Technologies Co., Ltd.", + "bbox": [ + 506, + 217, + 723, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f6ebcdb8087e5d1836998117936b0bbe0d692f6ead0982b3e17b1d70968080c.jpg", + "image_caption": [ + "Figure 1: Generated 3D dance examples conditioned music and text with our method. Given an audio input and text input with a specific starting time and duration, our method is able to generate a sequence of dance motions that fit the music and text instruction. The character is from Mixamo [1]." + ], + "image_footnote": [], + "bbox": [ + 76, + 282, + 890, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 492, + 313, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a novel task for generating 3D dance movements that simultaneously incorporate both text and music modalities. Unlike existing works that generate dance movements using a single modality such as music, our goal is to produce richer dance movements guided by the instructive information provided by the text. However, the lack of paired motion data with both music and text modalities limits the ability to generate dance movements that integrate both. To alleviate this challenge, we propose to utilize a 3D human motion VQ-VAE to project the motions of the two datasets into a latent space consisting of quantized vectors, which effectively mix the motion tokens from the two datasets with different distributions for training. Additionally, we propose a cross-modal transformer to integrate text instructions into motion generation architecture for generating 3D dance movements without degrading the performance of music-conditioned dance generation. To better evaluate the quality of the generated motion, we introduce two novel metrics, namely Motion Prediction Distance (MPD) and Freezing Score (FS), to measure the coherence and freezing percentage of the generated motion. Extensive experiments show that our approach can gener", + "bbox": [ + 73, + 526, + 473, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ate realistic and coherent dance movements conditioned on both text and music while maintaining comparable performance with the two single modalities. Code is available at https://garfield-kh.github.io/TM2D/.", + "bbox": [ + 500, + 493, + 892, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 585, + 632, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The music-conditioned dance generation has become a topic of great interest in recent years. The ability to generate dance movements that are synchronized with music has numerous applications, such as behavior understanding, simulation, and benefiting the community of dancers and musicians [25, 9, 29, 44]. Although music has been used as a guidance to generate dance movements, another important modality cue, text (or language), which provides richer actions and more flexible motion guidance, and is studied in other tasks such as image classification [38], detection [15], segmentation [50], and text-driven image generation [42], has not been fully explored in dance generation. To this end, we first propose a novel task for generating 3D dance movements that simultaneously incorporate both text and music modalities, enabling the generated human to perform rich dancing movements in accordance with the music and text.", + "bbox": [ + 496, + 612, + 892, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Designing a system pipeline for this bimodality driven 3D dance generation task is non-trivial. There exist two", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution: gongkehong@u.nus.edu, dongze@nus.edu.sg", + "bbox": [ + 91, + 876, + 460, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author: xinchao@nus.edu.sg", + "bbox": [ + 91, + 888, + 331, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9942", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "significant challenges to be considered: i) the existing datasets only cater to either music-driven (music2dance) [46, 5, 57, 29] or text-driven (text2motion) [36, 16] human motion generation, and no paired 3D dance generation dataset exists that takes into account both music and text. While building a new large-scale paired 3D dance dataset based on music and text is possible, it is time-consuming with fully annotated 3D human motion [29]; ii) the integration of text into music-conditioned dance generation requires a suitable architecture. However, existing methods that use music as a driving force to generate dance movements might result in temporal-freezing frames or fail to generalize to in-the-wild scenarios [29, 44]. Therefore, simple integration of text into the existing music-conditioned architecture might pose a risk of degraded dance generation quality in our new task.", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the first challenge, we take advantage of existing music-dance and text-motion datasets for this new task. However, directly mixing motions from these two datasets would result in inferior performance since the motions from these two datasets are in completely different motion spaces. To overcome this, we propose to utilize a VQ-VAE architecture to project the motions into a consistent and shared latent space. In particular, we build up a shared codebook for all the motions from the training set, and motions from both datasets are now represented as discrete tokens that are implicitly constrained to fall into a shared latent space. For the second challenge, we propose to utilize a cross-modal transformer architecture that formulates both music2dance and text2motion as sequence-to-sequence translation tasks. This architecture directly translates audio and text features into motion tokens and enables bimodality driven ability by introducing a fusion strategy in the latent space with a shared motion decoder for both tasks. With the shared decoder, audio and text information can be efficiently fused during inference. Our entire cross-modal transformer architecture is both effective and efficient, allowing for the integration of text instructions to generate coherent 3D dance motions, as illustrated in Figure 1.", + "bbox": [ + 75, + 337, + 472, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To better evaluate the coherence of generated dance in our task, we propose a new evaluation metric, Motion Prediction Distance (MPD), which measures the distance between the predicted motion and the ground truth at the time of integrating text, thereby providing a more accurate evaluation of the coherence of frames. Additionally, we introduce a Freezing Score (FS) that quantifies the percentage of temporal freezing frames in dance generation, which is a common problem in music-conditioned dance generation. To better evaluate the performance of our method in real-world scenarios, we also collect some in-the-wild music data for evaluation. Our method successfully performs dance generation based on both text and music while maintaining comparable performance on the single modal-", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ity tasks (music2dance, text2motion) compared to other state-of-the-art methods.", + "bbox": [ + 496, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows: i) We propose an interesting task of utilizing both music and text for 3D dance generation and propose a pipeline named TM2D (Text-Music to Dance) for this task. ii) Rather than collecting a new training set, we effectively combine the existing music2dance and text2motion datasets and employ a VQ-VAE framework to encode motions from all training sets to a shared feature space. iii) We propose a cross-modal transformer as well as a bimodal feature fusion strategy to encode both audio and text features, which is both effective and efficient. iv) We propose two new metrics, MPD and FS, which efficiently reflect the quality of generated motion. v) We successfully generate realistic and coherent dance based on both music and text instructions while maintaining comparable performance on the single modality tasks (music2dance, text2motion).", + "bbox": [ + 496, + 122, + 893, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 378, + 640, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Music to Dance", + "text_level": 1, + "bbox": [ + 500, + 402, + 656, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Music2Dance is typically divided into 2D and 3D dance generation and has been explored for many years. Recent methods model 3D dance generation from the perspective of network architecture. For instance, Lee et al. [26] explore the convolutional neural network in casual dilated setting. Lee et al. [25] propose a dance unit with Variational Auto-Encoder (VAE). Ren et al. [41] employ recurrent neural network (RNN) and the graph convolution network is introduced in [12]. As for the 3D dance generation, [54, 3, 24] implement convolutional neural network, [45, 14] apply adversarial learning in generated dance. [5, 46, 51, 58, 23] implement Long short-term memory (LSTM), and [9] implement motion graph with learned music/dance embedding through matching approach [10]. More recently, transformer has been applied in dance generation [28, 29, 44, 27]. However, the previous methods usually produce temporal freezing frames when generating the long sequences, or are difficult to generalize to in-the-wild music, thereby not satisfiable when directly used for bimodal driven 3D dance generation. In this paper, we design a cross modal transformer which is effective and efficient to integrate text instruction to generate the 3D dance.", + "bbox": [ + 496, + 426, + 893, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Text to Motion", + "text_level": 1, + "bbox": [ + 500, + 771, + 651, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to music-driven motion generation, text is also utilized as instructions to generate motions. Text2motion can be categorized into action label based motion generation and language description based motion generation. The action label based methods [18, 34] generate motion conditioned on action labels. Action2Motion [18] implement a GRU-VAE to iteratively generate the next", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9943", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b2acb0fe0656441ef4b950020dc31cabf4e42c64014fc6fe9df4c6a086485f3a.jpg", + "image_caption": [ + "Figure 2: Our proposed pipeline for music-text conditioned 3D dance generation. Three stages from left to right: 3D human motion VQ-VAE, training stage of the cross-modal transformer, and inference stage of our pipeline. In the first stage, a VQ-VAE is trained with both motions from music2dance and text2motion data, which is then used to tokenize all motions. In the second stage, a dual path cross-modal transformer is employed for sequence2sequence translation tasks (i.e., audio to motion tokens, text to motion tokens), with a shared motion decoder. In the third stage, given audio and text inputs, the audio and text encoders first extract the corresponding features, which are then fused (late fusion) in the motion decoder to generate dance condition on both music and text." + ], + "image_footnote": [], + "bbox": [ + 106, + 108, + 315, + 388 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6a7a7e5cf65d81ecf02ec94346b5c0a3480756f34855069e0d6b1d76ef27083a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 112, + 607, + 387 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ef5b69fe33d5d813f5825ea47be9477e8023b6ae80745d50f61adfaf6c59bc33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 116, + 870, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "frame based on the action label and previous frames. ACTOR [34] implements a transformer-VAE to encode and decode the whole pose sequence in one-shot. Since the action label based methods are restricted in a small set of action labels, language description based methods are proposed for more flexible motion generation. [37, 31, 2, 16, 17] formulate the text-to-motion task as a machine translation problem, others [52, 4, 13, 35] learns a joint embedding space of text and motion. Among them, [37, 31, 2, 52, 4, 13] use RNN encoder-decoder to learn the mapping between text and motion. [17, 16] use transformer encoder with RNN decoder, and TM2T [17] further introduces motion to text as an inverse alignment for text to motion. TEMOS [35] and TEACH [6] use transformer for both encoder and decoder for one-shot generation. Different from these text2motion methods, we focus on how to integrate two modalities (i.e., music and text) together for 3D dance generation.", + "bbox": [ + 75, + 546, + 472, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 820, + 169, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. System Pipeline", + "text_level": 1, + "bbox": [ + 76, + 845, + 235, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall pipeline is shown in Figure 2, which consists of three stages. The first stage employs a 3D human motion", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VQ-VAE to encode both the motions of the music2dance and text2motion datasets to a shared codebook encoded by multiple vectors such that each dance motion can be represented as a discrete motion token. After that, the combinations of music, text, and motion token are fed into a cross-modal transformer for training, which effectively learn how to predict a sequential motion tokens according to the previous those as well as music and text. In the third stage, a random starting motion token is generated and inputted to the cross-modal transformer with the given music and text for the sequential motion token prediction, which will be decoded by the pre-trained 3D human motion VQ-VAE in the first stage to generate 3D dance.", + "bbox": [ + 496, + 546, + 893, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. 3D Human Motion VQ-VAE", + "text_level": 1, + "bbox": [ + 500, + 755, + 756, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since there are no paired 3D dance motions conditioned on both music and text, we try to tackle our task with the existing music2dance and text2motion datasets. Instead of directly mixing the motions from both datasets for training, we employ a VQ-VAE to project the motions into a consistent and shared latent space, which is represented with a codebook, as illustrated in Figure 2. To be specific, given a 3D human motion $M \\in \\mathbb{R}^{T \\times d_m}$ , where $T$ is the time length", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9944", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and $d_{m}$ is the dimension of the human motion, an encoder of VQ-VAE that consists of several 1-D convolutional layers projects $M$ to a latent vector $z\\in \\mathbb{R}^{T^{\\prime}\\times d}$ , where $T^{\\prime} = \\frac{T}{t}$ and $t$ is the time interval for downsampling and $d$ is the dimension of the latent vector. A learnable codebook $e\\in \\mathbb{R}^{K\\times d}$ describes all latent variable features of the whole dataset, where $K$ and $d$ are the length and dimension of the codebook, respectively. A quantized latent vector $z_{q}\\in \\mathbb{R}^{T^{\\prime}\\times d}$ will record the closest vector from codebook $e$ as follows", + "bbox": [ + 76, + 90, + 472, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {q, i} = \\arg \\min _ {e _ {j} \\in e} \\| z _ {i} - e _ {j} \\| \\in \\mathbb {R} ^ {d}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 236, + 468, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the motion token $t_m \\in \\mathbb{R}^{T' \\times 1}$ stores the index of the closest vector", + "bbox": [ + 76, + 272, + 468, + 301 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {m, i} = \\arg \\min _ {j} \\| z _ {i} - e _ {j} \\| \\in \\mathbb {R} ^ {1}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 313, + 468, + 337 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The quantized $z_{q}$ will be decoded with stacked convolutional layers to reconstruct the human motion $\\hat{M}$ , and $t_{m}$ will be used in the training stage of the cross-modal transformer.", + "bbox": [ + 76, + 345, + 468, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the training of 3D human motion VQ-VAE, we follow the strategy in [48] and the total loss contains a reconstruction loss for human motion regression, a codebook loss for the dictionary learning, and a commitment loss to stabilize the training process:", + "bbox": [ + 76, + 407, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v q} = \\| \\hat {M} - M \\| _ {1} + \\| \\operatorname {s g} [ e ] - e _ {q} \\| _ {2} ^ {2} + \\beta \\| e - \\operatorname {s g} [ e _ {q} ] \\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 493, + 468, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{sg}[\\cdot ]$ is 'stop gradient' and $\\beta$ is the factor term to adjust the weight of the commitment loss. A straight-through estimator is also employed to pass the gradient from the decoder to the encoder in back-propagation.", + "bbox": [ + 76, + 521, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we will show in Sec. 4, the motion tokens from both datasets encoded by 3D human motion VQ-VAE fall almost in the shared latent space, which shows the feasibility of using the separate music-conditioned and text-conditioned motion datasets for music-text conditioned 3D dance generation task.", + "bbox": [ + 76, + 580, + 468, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Cross-modal Transformer", + "text_level": 1, + "bbox": [ + 76, + 681, + 313, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The cross-modal transformer contains an audio encoder, a text encoder, and a motion decoder. Since we use two separate datasets for our task, the cross-modal transformer is divided into two branches, where one for music2dance and the other for text2motion. It takes the audio feature, the text feature, and the motion token $t_m$ encoded by 3D human motion VQ-VAE as inputs, and performs the sequence-to-sequence translation task to generate the future motion tokens.", + "bbox": [ + 76, + 704, + 468, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attention. Attention is introduced in Transformer [49] for natural language processing, and then applied to many domains [30, 8, 20, 39, 11, 53, 32, 55, 40] A $L$ -layer transformer typically consists of $L$ transformer blocks, which", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "contains a multi-head self-attention (MSA) and a feed forward network (FFN). Given the input $x \\in \\mathbb{R}^{N \\times c}$ , where $N$ is the sequence length and $c$ is the dimension of the input, the transformer block first maps it to keys $K \\in \\mathbb{R}^{N \\times c}$ , queries $Q \\in \\mathbb{R}^{N \\times c}$ , values $V \\in \\mathbb{R}^{N \\times c}$ with the linear projections, and then a self-attention operation is performed by", + "bbox": [ + 496, + 90, + 890, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {c}}\\right) V. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 191, + 890, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The output is followed by a normalization layer and a FFN. Audio encoder. Given a sequence of music, the audio encoder first extracts the raw audio features following Bailando [44], where a public audio processing toolbox, Librosa [22], is employed to obtain the mel frequency cepstral coefficients (MFCC), MFCC delta, constant- $Q$ chromagram, tempogram and onset strength. After that, an embedded layer followed by several transformer blocks containing self-attention operation e.g., Eq. (4), is used to generate the processed audio features $f_{a}\\in \\mathbb{R}^{T^{\\prime}\\times d}$", + "bbox": [ + 496, + 231, + 890, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text encoder. Given a text instruction, we first extract the text token with a GloVe [33], and then an embedded layer followed by several transformer blocks is performed to obtain a processed text feature $f_{t} \\in \\mathbb{R}^{n \\times d}$ , where $n$ is the length of the text feature.", + "bbox": [ + 496, + 383, + 890, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion decoder. Motion decoder takes the motion tokens encoded by the 3D human motion VQ-VAE as inputs and outputs the future motion tokens, which has the same spirit as the sequence-to-sequence translation task. Since the information after time $t$ is unknown at moment $t$ , a masked MSA is first employed to interact with motions at different times. The mask has the shape of a simple upper triangular matrix and is performed in Eq. (4). For the moment $t$ , only the motions before moment $t$ are able to perform self-attention operations. Motion decoder also maps the extracted audio feature $f_{a}$ and text feature $f_{t}$ via audio encoder and text encoder to $K$ and $Q$ , to perform cross-modal attention with motion tokens. To enable the music-text conditional dance generation, (i.e., feature fusion, we apply shared parameters), we use the motion decoder with shared parameters for model efficiency, as shown in Figure 2.", + "bbox": [ + 496, + 459, + 892, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Details of Music-text Fusion", + "text_level": 1, + "bbox": [ + 500, + 726, + 751, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The detailed architecture is listed in Figure 3. Given the audio feature and text feature, the motion decoder first processes the past motion tokens with a self-attention layer followed by addition, normalization, a cross-attention with audio and text features separately, addition, normalization, and a feed-forward layer are performed. Such a procedure is repeated $L$ ( $L = 6$ ) times to build a typical $L$ -layer transformer. Then we adopt a late fusion strategy to perform a weighted sum of features of audio and text at a specific time (i.e., effect range). The weight curve is shown in Figure 3,", + "bbox": [ + 496, + 750, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9945", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a152b169740aa65fb16ca2449b80664d5f90a4b443a91c4499c5ec47685bb5e9.jpg", + "image_caption": [ + "Figure 3: Details of music-text feature fusion: Given the audio and text features from the encoder, the decoder processes them separately in the early layers. Then the audiotext fusion layer is applied by the weighted sum at the effect range with weight cure. Finally, a linear projection layer and a softmax operation are applied to predict the music-text conditioned motion tokens." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 259, + 303 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a17d2629f92343d8b02e3e7d407bed0428a76891763c6caf14fc73570a7f1997.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 88, + 467, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we slightly increase the weight of the text feature by a half cosine curve until a peak value of 0.8 at the beginning (20% time of effect range), and decrease it by a half cosine curve at the end (20% time of effect range). The weight of audio feature $W_{audio}$ is 1 - $W_{text}$ to ensure the feature keeps the same scale. With the fused feature, a linear projection layer and a softmax operation are applied to predict the music-text conditioned motion tokens, which are then decoded by the decoder of human motion VQ-VAE to obtain the music-text conditioned 3D dance sequences.", + "bbox": [ + 75, + 445, + 468, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training and Inference", + "text_level": 1, + "bbox": [ + 76, + 604, + 290, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. To train the 3D human motion VQ-VAE, we crop 64 frames with a sliding window from the original motion sequences as inputs, i.e., $T = 64$ , for both music2dance and text2motion datasets. We use a three-layer encoder and decoder so that the time interval of downsampling $t$ is 8. We randomly sample motions from both datasets and employ the Adam optimizer with a batch size of 128 and a learning rate of $1e^{-4}$ to optimizer our 3D human motion VQ-VAE.", + "bbox": [ + 75, + 628, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train the cross-modal transformer, we split two branches into two streams to train the music2dance and text2motion tasks on both datasets. For the music2dance task, we use the same frequency as the motion token to sample the music vector so that the music feature has the same time length as the motion token. We perform the sequence to sequence translation with the motion decoder to obtain the prediction of motion tokens conditioned by the music input. As for the text2motion task, since the text and motion token have different lengths, we use a maximum text length", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of 84 for text2motion translation with a padding strategy. We employ 6 self-attention layers for the audio encoder, text encoder, and motion decoder with the hidden dimension of 512 and 8 heads. The cross entropy loss is adopted for both music2dance and text2motion tasks", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = - \\frac {1}{m} \\sum_ {i = 0} ^ {m - 1} \\sum_ {j = 0} ^ {C - 1} y _ {i j} \\log y _ {\\hat {i j}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 176, + 890, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{y}$ is the prediction of motion token and $y$ is the ground-truth. $m$ is the length of motion tokens and $C$ is the number of classes of motion tokens, i.e., the length of codebook of 3D human motion VQ-VAE. Both tasks are simultaneously optimized with a batch size of 64 and a learning rate of $1e^{-4}$ .", + "bbox": [ + 496, + 231, + 890, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. Our aim is to generate 3D dance with music-text integration. Therefore, at the inference stage, we first fuse the music feature and text feature extracted by the audio encoder and text encoder with a weighted sum. Specifically, given an audio feature with a time length of $T'$ and a text feature with the length of $n$ , we first feed them into the motion decoder for future motion token prediction. Then we adopt the late fusion strategy to have a weighted sum of the generated motion tokens from both audio and text features at the duration of the integrated text, followed by a linear projection layer and a softmax operation, to obtain the combined motion tokens. Finally, the combined motion tokens are decoded into a 3D dance sequence after going through the decoder of our 3D human motion VQ-VAE model.", + "bbox": [ + 496, + 321, + 892, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 549, + 633, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 500, + 574, + 705, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Since there are no paired 3D dance motions conditioned on both music and text for our task, we combine the existing music2dance and text2motion datasets. For the music2dance dataset, we employ the AIST++ dataset [29]. For the text2motion dataset, we employ the HumanML3D [16] dataset. Additionally, we also collect a new in-the-wild music dataset to evaluate the generalization ability of our method, which contains 82 wild music clips from Youtube. The total of 53 minutes duration (8x larger size than the AIST++ test set) and various styles and content of music, which faithfully lies out of the distribution of AIST++ (explained in supplementary material). The evaluation on our dataset better reflects the generalization ability of the models in real-world scenarios.", + "bbox": [ + 496, + 598, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. Since we are the first to propose this bimodality driven 3D dance generation task, there are no existing methods for comparisons. We implement a traditional motion editing algorithm, slerp [43], which aligns the last generated frame of the dance with the first frame of the motion by text with a transition window of 10 frames, and", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9946", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "applies spherical interpolation of quaternion to fill in the transition in between. To validate the effectiveness of our method on music2dance task, we compare our method to the previous music2dance works [28, 57, 21, 29, 44]. To validate the effectiveness of our method on text2motion task, we compare our method to the previous text2motion works [31, 4, 7, 13, 47, 25, 17].", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We adopt the same evaluation settings as suggested by FACT [29] and Bailando [44] to evaluate the dance generation quality, including Fréchet Inception Distances (FID) [19], Diversity, and Beat Align Score. For the text2motion quality, we adopt the same evaluation settings as suggested by TM2T [17], including R-precision, Multimodal-Dist, FID, Diversity, and MultiModality.", + "bbox": [ + 75, + 196, + 468, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we also propose two new evaluation metrics, Motion Prediction Distance (MPD) and Freezing Score. The former reflects the coherence of frames at the time of integrating text, and the latter reacts to the percentage of temporal freezing frames in the generated dance. For the Freezing Score, we introduce the Percentage of Freezing Frame (PFF) and $\\mathrm{AUC}_f$ , where the PFF is defined by measuring the percentage of frozen frames with two criteria: 1) the maximum joint velocity blew a threshold $(0.015\\mathrm{m / s})$ . 2) its duration exceeds a certain period (3s). The $\\mathrm{AUC}_f$ is defined by area under curve of PFF in the threshold range from 0 to 0.03. As for MPD, it is defined as", + "bbox": [ + 75, + 303, + 470, + 483 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M P D} = \\min _ {i} \\| f _ {i} \\left(M _ {t _ {0} \\rightarrow t _ {1}}\\right) - M _ {t _ {1} \\rightarrow t _ {2}} \\| _ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 493, + 468, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M$ is the dance motions, $f$ is a motion prediction model, $f_{i}(M_{t_{0}\\rightarrow t_{1}})$ is the i-th predicted possible future motion, and $t_0$ , $t_1$ , and $t_2$ are the timestamps. It means that the model predicts various potential possible motion from $t_1$ to $t_2$ with the motion from $t_0$ to $t_1$ . If this distance is small (i.e., the generated dance $M_{t_1\\to t_2}$ lays in the possible future), then the generated dance motion $M_{t_1\\to t_2}$ is more coherent at the time of integrating text. We adopt the motion prediction model DLow [56] in this evaluation metric as it is designed for diverse potential furture motion prediction.", + "bbox": [ + 75, + 522, + 468, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Perceptual Evaluation. Besides the above metric measuring, we also conduct extensive user studies on Amazon Mechanical Turk (AMT) to perceptually evaluate the visual effects of our generated 3D dance results. Particularly, given each pair of dance movements sampled from our method and others with the same music clip, we request 3 distinct users on AMT to present their preference regarding the music-dance alignment, motion realism, and mobility. We further set the bar of involved users that only the ones with Master recognition who also have finished more than 1,000 tasks with over $98\\%$ approve rate are considered. Overall, there are 55 users employed in our user studies that come from various regions, ages, races and gender. The results of the user study are more representative to show the effect of", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the generated dance in practice.", + "bbox": [ + 500, + 90, + 710, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation on Music-text Conditioned Dance Generation", + "text_level": 1, + "bbox": [ + 500, + 116, + 890, + 147 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We validate the results of music-text conditioned dance generation on the test set of AIST++ and our dataset, as illustrated in Table 2. Here we mainly evaluate the coherence of generated dance with MPD metric at the time where music meets text, i.e., the time point where text starts to take effect. We measure the coherence of pure music2dance generation as a baseline named a2d to reflect the influence of with/without text instruction. We use the past 25 motion frames to predict the future 30 frames, and calculate the MPD from future frame (ft) = 10 to ft = 30, respectively. Our method consistently outperforms slerp [43] and a2d baseline in both datasets, and gains a similar result compared to the ground-truth shown in Table 2, which verifies the naturalness of our generated motion. More importantly, user study experiments show that our method generates more realistic 3D dance compared to dance generation only conditioned music, and $70.0\\%$ Win rate even compared to the ground-truth shown in Table 2. The results of our user study indicate that the music-text conditioned dance generation received high ratings from participants, highlighting the importance of considering audience preferences in evaluating dance quality.", + "bbox": [ + 496, + 155, + 890, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation on Music Conditioned Dance Generation", + "text_level": 1, + "bbox": [ + 498, + 498, + 890, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the effectiveness of our architecture, we quantitatively compare our method with state-of-the-art those for music conditioned 3D dance generation. The results on the test set of AIST++ are shown in Table 1. We can find that our method outperforms the previous ones in terms of motion diversity $(\\mathrm{Div}_k$ and $\\mathrm{Div}_g)$ , while the performance of the motion quality $(\\mathrm{FID}_g)$ and beat align score is inferior under the condition that we do not use seed motion from ground-truth compared with FACT [29] and Bailando [44].", + "bbox": [ + 496, + 537, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The existing evaluation metrics are not sufficient to reflect the quality of generated dance in practice, which motivates us to propose a PFF and $\\mathrm{AUC}_f$ to evaluate the percentage of freezing frames. It is worth noting that our architecture outperforms FACT [29] and Bailando [44] in terms of PFF and $\\mathrm{AUC}_f$ , which shows that our method rarely generates frozen frames. Meanwhile, Li et al. [28] and DanceNet [57] also gain nearly zero freezing in PFF and $\\mathrm{AUC}_f$ . The reason is that the generated dances of Li et al. [28] are highly jittery making its velocity variation extremely high, which is also reported in [29, 44], and leads to the non-freezing issue. And for DanceNet [57], it generates dance in a repeat motion pattern, nearly zero freezing but with low diversity. Furthermore, the results of user study show that the generated 3D dance is more visually re", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9947", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7d6a3411797acf9123a7a4428ec5fadbcc84daac9027b02a6c0196557276d673.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DataMethodMotion QualityMotion DiversityFreezingBeat Align Score ↑User Study
FIDk ↓FIDg ↑Divk ↑Divg ↑PFF↓AUCf ↓Our Method Wins
AIST ++Ground-truth17.1010.608.197.450.000.000.237441.9%
Li et al. [28]86.4343.466.85*3.320.000.000.160798.3%
DanceNet [57]69.1825.492.862.850.000.980.143087.2%
DanceRevolution [21]73.4225.923.524.8711.0112.220.195070.5%
FACT [29]35.3522.115.946.1825.2921.590.220983.3%
Bailando [44]28.169.627.836.3414.9113.250.233265.0%
ours (only dance data)23.949.537.694.530.000.000.2127-
ours19.0120.099.456.360.000.000.2049-
Wild AudioFACT [29]70.3620.237.336.3432.6427.210.221189.6%
Bailando [44]50.5622.553.806.0417.5514.140.216670.8%
ours (only dance data)43.8513.088.524.760.780.760.1998-
ours27.6520.347.885.270.120.110.2290-
", + "bbox": [ + 81, + 88, + 885, + 329 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a66234f4fb7b95103ba555b9ff71aa93709a8e7798e72d2c425de91544be1db2.jpg", + "table_caption": [ + "Table 1: Music Conditioned Dance Generation: quantitative results on AIST++ and Wild Audio test set. The best and runner-up values are bold and underlined, respectively. Among compared methods, \"Li et al.\", DanceNet and FACT are multiplexing the same results of AIST++ benchmark [28], while DanceRevolution [21] is followed from Bailando [44]. $\\dagger$ FID $_k$ and DIV $_k$ are fetched from [29] while FID $_g$ and DIV $_g$ are fetched from [44]." + ], + "table_footnote": [], + "table_body": "
DataMethodMotion Prediction DistanceUser Study
ft=10ft=20ft=30Our Method Wins
AIST ++GT0.0480.0720.08870.0%
a2d0.0520.0840.10866.6%
slerp [43]0.0880.1220.13573.3%
ours0.0490.0800.102-
Wild Audioa2d0.0520.0880.11360.0%
slerp [43]0.1040.1480.17170.0%
ours0.0480.0790.107-
", + "bbox": [ + 80, + 425, + 460, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Music-text conditioned Dance Generation: quantitative results on AIST++ and wild audio test set. The best and runner-up values are bold and underlined, respectively.", + "bbox": [ + 75, + 607, + 468, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "alistic compared to other methods. Even in comparison to the ground truth, $41.9\\%$ of our generated dance is voted as the better in average. We also report the result trained with dance-only data, which shows comparable performance.", + "bbox": [ + 75, + 686, + 468, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to the evaluation on the test set of AIST++, we also show the experimental results on our in-the-wild dataset in Table 1. Our method outperforms FACT [29] and Bailando [44] in almost all metrics except for $\\mathrm{FID}_g\\mathrm{Div}_g$ . We can observe that both FACT [29] and Bailando [44] shows a large performance drop in terms of $\\mathrm{FID}_k$ , while ours maintain a small change. This is because FACT [29] and Bailando [44] requires seed motion, however, there is no ground-truth for in-the-wild scenario. With random sampled seed motion or token, their methods are not adapted", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "well for the in-the-wild scenarios. One can also notice that there exists freezing in our method on the in-the-wild dataset but the percentage of frozen frames is zero on the test set of AIST++, which results from the different distributions of AIST++ and our dataset. It also shows that our dataset is more challenging. We also report the result trained with dance-only data, which shows the advantage of our mix-training strategy in the wild scenario.", + "bbox": [ + 496, + 428, + 892, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Evaluation on Text Conditioned Motion Generation.", + "text_level": 1, + "bbox": [ + 498, + 556, + 890, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our text2motion approach in two different scenarios: inference with text only (t2m), and inference with text and music feature fusion. Table 3 shows the results. In the text-only setting, our approach achieves comparable performance with TM2T baseline(T) [17], which demonstrates that the mixed data/tasks training does not affect the quality of text2motion generation. We then apply our late fusion method by randomly sampling a music clip (from the whole AIST++ test set) and a text (from the whole t2m test set), and evaluate the generated dance clips following the same protocol as TM2T [17]. As Table 3 indicates, the text-dance consistency remains acceptable with a late fusion ratio of 0.8. For more details on the relation between the late fusion rate and text2motion result, please refer to the supplementary material.", + "bbox": [ + 496, + 595, + 892, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 832, + 683, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also show the qualitative results of our method for music-text conditioned 3D dance generation in Figure 4, Figure 5 and Figure 6. Specifically, Figure 4 shows the gen", + "bbox": [ + 498, + 854, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9948", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f64a580540a6691bd95518814da478e08ff1b6ea7d9dc1d35b6524a0ea33e490.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsR Precision↑FID↓MM Dist↓Diversity→MModality↑
Real motions0.5110.0022.9749.503
Seq2Seq [31]0.18011.755.5296.223
Language2Pose [4]0.24611.025.2967.676
Text2Gesture [7]0.1655.0126.0306.409
Hier [13]0.3016.5325.0128.332
MoCoGAN [47]0.03794.419.6430.462
Dance2Music [25]0.03366.988.1160.725
TM2T baseline(T) [17]0.3511.6694.0469.632
TM2T [17]0.4241.5013.4678.589
TM2D (t2m)0.3191.0214.0989.513
TM2D (LFR 0.8)0.3001.1054.3078.887
", + "bbox": [ + 86, + 88, + 455, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: The evaluation of text2motion on the HumanML3D dataset, where the results are averaged from 20 runs.", + "bbox": [ + 75, + 250, + 468, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "erated 3D dance with the same music sequence but with different text instructions. Figure 5 shows the generated 3D dance with the same text instruction but with a different start time. Figure 6 shows the generated 3D dance with the same text instruction but with different durations. As one can see, our approach maintains plausible visual results according to the text instructions for all three cases, which confirms that our approach is more flexible.", + "bbox": [ + 75, + 321, + 467, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. The Impact of Mixed Data", + "text_level": 1, + "bbox": [ + 76, + 452, + 316, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We employ two datasets (AIST++ and HumanML3D) to train our 3D human motion VQ-VAE so that the motion tokens drop in the same latent space. To empirically show this point, we conduct experiments to count the number of shared motion tokens. For a trained 3D human motion VQ-VAE with AIST++ and HumanML3D, there are 855 vectors and 912 vectors from codebook used to construct the motion token in AIST++ and HumanML3D. The total number of vectors contained in the codebook is 1024. among them, 846 vectors (98.9% in AIST++ and 92.8% in HumanML3D) are shared to generate the motion tokens, which shows the feasibility of using the separate music-conditioned and text-conditioned motion datasets for music-text conditioned 3D dance generation task. We further verify the latent space of the motions by VQ-VAE, and we perform t-SNE visualizations of the raw motion distribution (before VQ-VAE) and feature distribution encoded by VQ-VAE (after VQ-VAE) from a2d motion and t2m motion in Figure 7. Two motion datasets are mixed successfully, which provides the potential integration of bimodality dance generation. Refer to the supplementary materials for more details.", + "bbox": [ + 75, + 474, + 467, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Model Efficiency", + "text_level": 1, + "bbox": [ + 76, + 816, + 241, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also show the efficiency of our architecture in Table 4. In the inference stage, the complete model parameters and the inference time for 1s music are compared. Our architecture requires only about half model parameters and", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2da668e43b51e5fe4e3174a4b2d1539cb367aa1c010297da945da99eba7aa647.jpg", + "image_caption": [ + "Figure 4: Same music, different actions. These two generated dance share the same music and effect range (from 6s to 8s), with different text instructions: \"A person does a cartwheel\" (top), \"A person is spinning with arms spread out\" (bottom)." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 883, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6ed11b5d684d9dc22721a26d539a64cbab1593e3d2d49602d0d348a64e9d5d7.jpg", + "image_caption": [ + "Figure 5: Same action, different start time points. These two generated dance share the same music and text instruction (\"A person jumps up and down\") but different effect ranges: from 6 to 8s (top), from 7 to 9s (bottom)." + ], + "image_footnote": [], + "bbox": [ + 501, + 338, + 888, + 488 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9350bd249e2b5343d504bd69a319079e85b30d72d69ca01a62fb482cabbe418e.jpg", + "image_caption": [ + "Figure 6: Same action, different durations. These two generated dance share the same music and text instruction (\"A person is keeping jumping\") and action start time point (i.e., 6s) but different effect duration: 2s (top), 3s (bottom, i.e., jumped twice)." + ], + "image_footnote": [], + "bbox": [ + 501, + 578, + 888, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "inference time compared to Bailando [44], which attributes", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9949", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/01431764affd7d7ed49c8fcc7d29d4b6b9db99f838dee551d499ee0e5e534132.jpg", + "image_caption": [ + "(a) Before VQ-VAE" + ], + "image_footnote": [], + "bbox": [ + 106, + 113, + 263, + 204 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/bc4989c4236db7c2dc909ea9980d7c3906bb0ab3b3b4fe0f22db1813e69da301.jpg", + "image_caption": [ + "(b) After VQ-VAE", + "Figure 7: The t-SNE visualizations of motions before and after VQ-VAE (orange: motions in AIST++, blue: motions in HumanML3D)." + ], + "image_footnote": [], + "bbox": [ + 303, + 113, + 457, + 204 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/2c609239249cab11e4fe0994cf3903102fd26be37d2df369bb754062ee98ab57.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParametersInference time
FACT [29]120M8.300s/(1s music)
Bailando [44]152M0.236s/(1s music)
ours72M0.143s/(1s music)
", + "bbox": [ + 99, + 301, + 442, + 385 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 4: Comparisons of parameters and inference time.", + "bbox": [ + 86, + 395, + 455, + 410 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "that we do not need to encode separate features for the upper and lower half body. Compared to FACT [29], our architectures significantly reduce the inference time. The main reason is that FACT employs two transformers for the music, dance motion, and a cross-modal transformer, but we formulate it as the standard sequence-to-sequence translation task with one transformer only, which establishes the advantage of our architecture.", + "bbox": [ + 75, + 438, + 468, + 556 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 571, + 194, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, a novel task that simultaneously integrates both music and text instruction for 3D dance generation is proposed. Due to the lack of the paired 3D dance motions conditioned on both music and text, we resort to two existing datasets, i.e., music2dance and text2motion, to perform this task and employ a 3D human motion VQ-VAE to project the motions of the two datasets into a shared latent space so that the two datasets with different distributions can be effectively mixed for training. Moreover, we also propose a cross-modal transformer architecture to generate 3D dance without degrading the performance of music-conditioned dance generation. Two new evaluation metrics, MPD and FS, are proposed to reflect the quality of generated motion. Extensive experiments show that our method can generate dance motion that matches both music and text in a realistic and coherent way while maintaining comparable performance on two single modalities.", + "bbox": [ + 75, + 598, + 468, + 853 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. This project is supported by the Ministry of Education Singapore under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006),", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "and Singapore National Research Foundation (\"CogniVision - Energy-autonomous always-on cognitive and attentive cameras for distributed real-time vision with milliwatt power consumption\" grant NRF-CRP20-2017-0003) - www.green-ic.org/CogniVision.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 178, + 596, + 193 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] mixamo. https://www MIXamo.com/. Accessed: 2023-03-07. 1", + "[2] Hyemin Ahn, Timothy Ha, Yunho Choi, Hwiyeon Yoo, and Songhwai Oh. Text2action: Generative adversarial synthesis from language to action. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5915-5920. IEEE, 2018. 3", + "[3] Hyemin Ahn, Jaehun Kim, Kihyun Kim, and Songhwai Oh. Generative autoregressive networks for 3d dancing move synthesis from music. IEEE Robotics and Automation Letters, 5(2):3501-3508, 2020. 2", + "[4] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 3, 6, 8", + "[5] Omid Alemi, Jules Françoise, and Philippe Pasquier. Groovenet: Real-time music-driven dance movement generation using artificial neural networks. In Workshop on Machine Learning for Creativity, 23rd ACM SIGKDD Conference on Knowledge Discovery and Data Mining, volume 8, page 26, 2017. 2", + "[6] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. Teach: Temporal action compositions for 3d humans. In International Conference on 3D Vision (3DV), 2022. 3", + "[7] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In IEEE Virtual Reality and 3D User Interfaces (VR), pages 1-10. IEEE, 2021. 6, 8", + "[8] Hanting Chen, Yunhe Wang, Jianyuan Guo, and Dacheng Tao. VanillaNet: the Power of Minimalism in Deep Learning. arXiv preprint arXiv:2305.12972, 2023. 4", + "[9] Kang Chen, Zhipeng Tan, Jin Lei, Song-Hai Zhang, Yuan-Chen Guo, Weidong Zhang, and Shi-Min Hu. Choreomaster: choreography-oriented music-driven dance synthesis. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 1, 2", + "[10] Rukun Fan, Songhua Xu, and Weidong Geng. Example-based automatic music-driven conventional dance motion synthesis. IEEE transactions on visualization and computer graphics, 18(3):501-515, 2011. 2", + "[11] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. arXiv preprint arXiv:2305.10924, 2023. 4", + "[12] Joao P Ferreira, Thiago M Coutinho, Thiago L Gomes, José F Neto, and et al. Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics, 94:11-21, 2021. 2" + ], + "bbox": [ + 501, + 203, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9950", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1396–1406, 2021. 3, 6, 8", + "[14] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 2", + "[15] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 1", + "[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 3, 5", + "[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597, 2022. 3, 6, 7, 8", + "[18] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 2", + "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in neural information processing systems, volume 30, 2017. 6", + "[20] Duc N.M. Hoang, Shiwei Liu, Radu Marculescu, and Zhangyang Wang. Revisiting pruning at initialization through the lens of ramanujan graph. In International Conference on Learning Representations, 2023. 4", + "[21] Ruozhi Huang, Huang Hu, Wei Wu, Kei Sawada, Mi Zhang, and Daxin Jiang. Dance revolution: Long-term dance generation with music via curriculum learning. In International Conference on Learning Representations, 2021. 6, 7", + "[22] Yanghua Jin, Jiakai Zhang, Minjun Li, Yingtao Tian, Huachun Zhu, and Zhihao Fang. Towards the automatic anime characters creation with generative adversarial networks. arXiv preprint arXiv:1708.05509, 2017. 4", + "[23] Hsuan-Kai Kao and Li Su. Temporally guided music-to-body-movement generation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 147-155, 2020. 2", + "[24] Kosmas Kritsis, Aggelos Gkiokas, Aggelos Pikrakis, and Vassilis Katsouros. Danceconv: Dance motion generation with convolutional networks. IEEE Access, 10:44982-45000, 2022. 2", + "[25] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. Advances in neural information processing systems, 32, 2019. 1, 2, 6, 8" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Juheon Lee, Seohyun Kim, and Kyogu Lee. Listen to dance: Music-driven choreography generation using autoregressive encoder-decoder network. arXiv preprint arXiv:1811.00818, 2018. 2", + "[27] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 2", + "[28] Jiaman Li, Yihang Yin, Hang Chu, Yi Zhou, Tingwu Wang, Sanja Fidler, and Hao Li. Learning to generate diverse dance motions with transformer. arXiv preprint arXiv:2008.08171, 2020. 2, 6, 7", + "[29] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 1, 2, 5, 6, 7, 9", + "[30] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 4", + "[31] Angela S. Lin, Lemeng Wu, Rodolfo Corona, Kevin W. H. Tai, Qixing Huang, and Raymond J. Mooney. Generating animated videos of human activities from natural language descriptions. 2018. 3, 6, 8", + "[32] Xinyin Ma, Gongfan Fang, and Xinchao Wang. LLM-Pruner: On the Structural Pruning of Large Language Models. arXiv preprint arXiv:2305.11627, 2023. 4", + "[33] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 4", + "[34] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2, 3", + "[35] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, page 480-497, 2022. 3", + "[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236-252, 2016. 2", + "[37] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems, 109:13-26, 2018. 3", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9951", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Sucheng Ren, Xingyi Yang, Songhua Liu, and Xinchao Wang. SG-Former: Self-guided Transformer with Evolving Token Reallocation. In IEEE International Conference on Computer Vision, 2023. 4", + "[40] Sucheng Ren, Daquan Zhou, Shengfeng He, Jiashi Feng, and Xinchao Wang. Shunted Self-Attention via Multi-Scale Token Aggregation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 4", + "[41] Xuanchi Ren, Haoran Li, Zijian Huang, and Qifeng Chen. Self-supervised dance video synthesis conditioned on music. In Proceedings of the 28th ACM International Conference on Multimedia, pages 46-54, 2020. 2", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1", + "[43] Ken Shoemake. Animating rotation with quaternion curves. In Proceedings of the 12th annual conference on Computer graphics and interactive techniques, pages 245-254, 1985. 5, 6, 7", + "[44] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11050-11059, 2022. 1, 2, 4, 6, 7, 8, 9", + "[45] Guofei Sun, Yongkang Wong, Zhiyong Cheng, Mohan S Kankanhalli, Weidong Geng, and Xiangdong Li. Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Transactions on Multimedia, 23:497-509, 2020. 2", + "[46] Taoran Tang, Jia Jia, and Hanyang Mao. Dance with melody: An LSTM-autoencoder approach to music-oriented dance synthesis. In Proceedings of the 26th ACM international conference on Multimedia, pages 1598-1606, 2018. 2", + "[47] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 6, 8", + "[48] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4", + "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4", + "[50] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. 1", + "[51] Nelson Yalta, Shinji Watanabe, Kazuhiro Nakadai, and Tetsuya Ogata. Weakly-supervised deep recurrent neural networks for basic dance step generation. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2019. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Tatsuro Yamada, Hiroyuki Matsunaga, and Tetsuya Ogata. Paired recurrent autoencoders for bidirectional translation between robot actions and linguistic descriptions. IEEE Robotics and Automation Letters, 3(4):3441-3448, 2018. 3", + "[53] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Advances in neural information processing systems, 2022. 4", + "[54] Zijie Ye, Haozhe Wu, Jia Jia, Yaohua Bu, Wei Chen, Fanbo Meng, and Yanfeng Wang. Choreonet: Towards music to dance synthesis with choreographic action unit. In Proceedings of the 28th ACM International Conference on Multimedia, pages 744-752, 2020. 2", + "[55] Weihao Yu, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. Metaformer Baselines for Vision. arXiv preprint arXiv:2210.13452, 2022.4", + "[56] Ye Yuan and Kris Kitani. Dlow: Diversifying latent flows for diverse human motion prediction. In European Conference on Computer Vision, pages 346-364. Springer, 2020. 6", + "[57] Wenlin Zhuang, Congyi Wang, Jinxiang Chai, Yangang Wang, Ming Shao, and Siyu Xia. Music2dance: Dancenet for music-driven dance generation. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 18(2):1-21, 2022. 2, 6, 7", + "[58] Wenlin Zhuang, Yangang Wang, Joseph Robinson, Congyi Wang, Ming Shao, Yun Fu, and Siyu Xia. Towards 3d dance motion synthesis and control. arXiv preprint arXiv:2006.05743, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 486 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9952", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/1834d63b-e3e0-4b42-9900-788dbb6bee22_model.json b/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/1834d63b-e3e0-4b42-9900-788dbb6bee22_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8c17eeb7e06c01eae83d97cd46453cfc7a147203 --- /dev/null +++ b/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/1834d63b-e3e0-4b42-9900-788dbb6bee22_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.131, + 0.868, + 0.153 + ], + "angle": 0, + "content": "TM2D: Bimodality Driven 3D Dance Generation via Music-Text Integration" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.18, + 0.29, + 0.22 + ], + "angle": 0, + "content": "Kehong Gong1* \nZihang Jiang1" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.18, + 0.462, + 0.216 + ], + "angle": 0, + "content": "Dongze Lian\\(^{1*}\\) \nXinxin Zuo\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.218, + 0.476, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\) National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.18, + 0.64, + 0.216 + ], + "angle": 0, + "content": "Heng Chang\\(^{2}\\) \nMichael Bi Mi\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.679, + 0.181, + 0.813, + 0.218 + ], + "angle": 0, + "content": "Chuan Guo² \nXinchao Wang¹†" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.218, + 0.725, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\) Huawei Technologies Co., Ltd." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.283, + 0.891, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.434, + 0.893, + 0.48 + ], + "angle": 0, + "content": "Figure 1: Generated 3D dance examples conditioned music and text with our method. Given an audio input and text input with a specific starting time and duration, our method is able to generate a sequence of dance motions that fit the music and text instruction. The character is from Mixamo [1]." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.493, + 0.314, + 0.509 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.527, + 0.474, + 0.86 + ], + "angle": 0, + "content": "We propose a novel task for generating 3D dance movements that simultaneously incorporate both text and music modalities. Unlike existing works that generate dance movements using a single modality such as music, our goal is to produce richer dance movements guided by the instructive information provided by the text. However, the lack of paired motion data with both music and text modalities limits the ability to generate dance movements that integrate both. To alleviate this challenge, we propose to utilize a 3D human motion VQ-VAE to project the motions of the two datasets into a latent space consisting of quantized vectors, which effectively mix the motion tokens from the two datasets with different distributions for training. Additionally, we propose a cross-modal transformer to integrate text instructions into motion generation architecture for generating 3D dance movements without degrading the performance of music-conditioned dance generation. To better evaluate the quality of the generated motion, we introduce two novel metrics, namely Motion Prediction Distance (MPD) and Freezing Score (FS), to measure the coherence and freezing percentage of the generated motion. Extensive experiments show that our approach can gener" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.494, + 0.893, + 0.555 + ], + "angle": 0, + "content": "ate realistic and coherent dance movements conditioned on both text and music while maintaining comparable performance with the two single modalities. Code is available at https://garfield-kh.github.io/TM2D/." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.587, + 0.633, + 0.603 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.893, + 0.869 + ], + "angle": 0, + "content": "The music-conditioned dance generation has become a topic of great interest in recent years. The ability to generate dance movements that are synchronized with music has numerous applications, such as behavior understanding, simulation, and benefiting the community of dancers and musicians [25, 9, 29, 44]. Although music has been used as a guidance to generate dance movements, another important modality cue, text (or language), which provides richer actions and more flexible motion guidance, and is studied in other tasks such as image classification [38], detection [15], segmentation [50], and text-driven image generation [42], has not been fully explored in dance generation. To this end, we first propose a novel task for generating 3D dance movements that simultaneously incorporate both text and music modalities, enabling the generated human to perform rich dancing movements in accordance with the music and text." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Designing a system pipeline for this bimodality driven 3D dance generation task is non-trivial. There exist two" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.877, + 0.462, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution: gongkehong@u.nus.edu, dongze@nus.edu.sg" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.889, + 0.333, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author: xinchao@nus.edu.sg" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.871, + 0.462, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9942" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "significant challenges to be considered: i) the existing datasets only cater to either music-driven (music2dance) [46, 5, 57, 29] or text-driven (text2motion) [36, 16] human motion generation, and no paired 3D dance generation dataset exists that takes into account both music and text. While building a new large-scale paired 3D dance dataset based on music and text is possible, it is time-consuming with fully annotated 3D human motion [29]; ii) the integration of text into music-conditioned dance generation requires a suitable architecture. However, existing methods that use music as a driving force to generate dance movements might result in temporal-freezing frames or fail to generalize to in-the-wild scenarios [29, 44]. Therefore, simple integration of text into the existing music-conditioned architecture might pose a risk of degraded dance generation quality in our new task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.473, + 0.686 + ], + "angle": 0, + "content": "To address the first challenge, we take advantage of existing music-dance and text-motion datasets for this new task. However, directly mixing motions from these two datasets would result in inferior performance since the motions from these two datasets are in completely different motion spaces. To overcome this, we propose to utilize a VQ-VAE architecture to project the motions into a consistent and shared latent space. In particular, we build up a shared codebook for all the motions from the training set, and motions from both datasets are now represented as discrete tokens that are implicitly constrained to fall into a shared latent space. For the second challenge, we propose to utilize a cross-modal transformer architecture that formulates both music2dance and text2motion as sequence-to-sequence translation tasks. This architecture directly translates audio and text features into motion tokens and enables bimodality driven ability by introducing a fusion strategy in the latent space with a shared motion decoder for both tasks. With the shared decoder, audio and text information can be efficiently fused during inference. Our entire cross-modal transformer architecture is both effective and efficient, allowing for the integration of text instructions to generate coherent 3D dance motions, as illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To better evaluate the coherence of generated dance in our task, we propose a new evaluation metric, Motion Prediction Distance (MPD), which measures the distance between the predicted motion and the ground truth at the time of integrating text, thereby providing a more accurate evaluation of the coherence of frames. Additionally, we introduce a Freezing Score (FS) that quantifies the percentage of temporal freezing frames in dance generation, which is a common problem in music-conditioned dance generation. To better evaluate the performance of our method in real-world scenarios, we also collect some in-the-wild music data for evaluation. Our method successfully performs dance generation based on both text and music while maintaining comparable performance on the single modal-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ity tasks (music2dance, text2motion) compared to other state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.895, + 0.365 + ], + "angle": 0, + "content": "In summary, our contributions are as follows: i) We propose an interesting task of utilizing both music and text for 3D dance generation and propose a pipeline named TM2D (Text-Music to Dance) for this task. ii) Rather than collecting a new training set, we effectively combine the existing music2dance and text2motion datasets and employ a VQ-VAE framework to encode motions from all training sets to a shared feature space. iii) We propose a cross-modal transformer as well as a bimodal feature fusion strategy to encode both audio and text features, which is both effective and efficient. iv) We propose two new metrics, MPD and FS, which efficiently reflect the quality of generated motion. v) We successfully generate realistic and coherent dance based on both music and text instructions while maintaining comparable performance on the single modality tasks (music2dance, text2motion)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.379, + 0.642, + 0.395 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.404, + 0.657, + 0.42 + ], + "angle": 0, + "content": "2.1. Music to Dance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.895, + 0.761 + ], + "angle": 0, + "content": "Music2Dance is typically divided into 2D and 3D dance generation and has been explored for many years. Recent methods model 3D dance generation from the perspective of network architecture. For instance, Lee et al. [26] explore the convolutional neural network in casual dilated setting. Lee et al. [25] propose a dance unit with Variational Auto-Encoder (VAE). Ren et al. [41] employ recurrent neural network (RNN) and the graph convolution network is introduced in [12]. As for the 3D dance generation, [54, 3, 24] implement convolutional neural network, [45, 14] apply adversarial learning in generated dance. [5, 46, 51, 58, 23] implement Long short-term memory (LSTM), and [9] implement motion graph with learned music/dance embedding through matching approach [10]. More recently, transformer has been applied in dance generation [28, 29, 44, 27]. However, the previous methods usually produce temporal freezing frames when generating the long sequences, or are difficult to generalize to in-the-wild music, thereby not satisfiable when directly used for bimodal driven 3D dance generation. In this paper, we design a cross modal transformer which is effective and efficient to integrate text instruction to generate the 3D dance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.652, + 0.787 + ], + "angle": 0, + "content": "2.2. Text to Motion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In addition to music-driven motion generation, text is also utilized as instructions to generate motions. Text2motion can be categorized into action label based motion generation and language description based motion generation. The action label based methods [18, 34] generate motion conditioned on action labels. Action2Motion [18] implement a GRU-VAE to iteratively generate the next" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9943" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.109, + 0.316, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.113, + 0.609, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.117, + 0.872, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.415, + 0.896, + 0.522 + ], + "angle": 0, + "content": "Figure 2: Our proposed pipeline for music-text conditioned 3D dance generation. Three stages from left to right: 3D human motion VQ-VAE, training stage of the cross-modal transformer, and inference stage of our pipeline. In the first stage, a VQ-VAE is trained with both motions from music2dance and text2motion data, which is then used to tokenize all motions. In the second stage, a dual path cross-modal transformer is employed for sequence2sequence translation tasks (i.e., audio to motion tokens, text to motion tokens), with a shared motion decoder. In the third stage, given audio and text inputs, the audio and text encoders first extract the corresponding features, which are then fused (late fusion) in the motion decoder to generate dance condition on both music and text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.473, + 0.807 + ], + "angle": 0, + "content": "frame based on the action label and previous frames. ACTOR [34] implements a transformer-VAE to encode and decode the whole pose sequence in one-shot. Since the action label based methods are restricted in a small set of action labels, language description based methods are proposed for more flexible motion generation. [37, 31, 2, 16, 17] formulate the text-to-motion task as a machine translation problem, others [52, 4, 13, 35] learns a joint embedding space of text and motion. Among them, [37, 31, 2, 52, 4, 13] use RNN encoder-decoder to learn the mapping between text and motion. [17, 16] use transformer encoder with RNN decoder, and TM2T [17] further introduces motion to text as an inverse alignment for text to motion. TEMOS [35] and TEACH [6] use transformer for both encoder and decoder for one-shot generation. Different from these text2motion methods, we focus on how to integrate two modalities (i.e., music and text) together for 3D dance generation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.821, + 0.17, + 0.837 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.847, + 0.236, + 0.864 + ], + "angle": 0, + "content": "3.1. System Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The overall pipeline is shown in Figure 2, which consists of three stages. The first stage employs a 3D human motion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.895, + 0.745 + ], + "angle": 0, + "content": "VQ-VAE to encode both the motions of the music2dance and text2motion datasets to a shared codebook encoded by multiple vectors such that each dance motion can be represented as a discrete motion token. After that, the combinations of music, text, and motion token are fed into a cross-modal transformer for training, which effectively learn how to predict a sequential motion tokens according to the previous those as well as music and text. In the third stage, a random starting motion token is generated and inputted to the cross-modal transformer with the given music and text for the sequential motion token prediction, which will be decoded by the pre-trained 3D human motion VQ-VAE in the first stage to generate 3D dance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.757, + 0.773 + ], + "angle": 0, + "content": "3.2. 3D Human Motion VQ-VAE" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Since there are no paired 3D dance motions conditioned on both music and text, we try to tackle our task with the existing music2dance and text2motion datasets. Instead of directly mixing the motions from both datasets for training, we employ a VQ-VAE to project the motions into a consistent and shared latent space, which is represented with a codebook, as illustrated in Figure 2. To be specific, given a 3D human motion \\( M \\in \\mathbb{R}^{T \\times d_m} \\), where \\( T \\) is the time length" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9944" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "and \\(d_{m}\\) is the dimension of the human motion, an encoder of VQ-VAE that consists of several 1-D convolutional layers projects \\(M\\) to a latent vector \\(z\\in \\mathbb{R}^{T^{\\prime}\\times d}\\), where \\(T^{\\prime} = \\frac{T}{t}\\) and \\(t\\) is the time interval for downsampling and \\(d\\) is the dimension of the latent vector. A learnable codebook \\(e\\in \\mathbb{R}^{K\\times d}\\) describes all latent variable features of the whole dataset, where \\(K\\) and \\(d\\) are the length and dimension of the codebook, respectively. A quantized latent vector \\(z_{q}\\in \\mathbb{R}^{T^{\\prime}\\times d}\\) will record the closest vector from codebook \\(e\\) as follows" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.237, + 0.47, + 0.261 + ], + "angle": 0, + "content": "\\[\nz _ {q, i} = \\arg \\min _ {e _ {j} \\in e} \\| z _ {i} - e _ {j} \\| \\in \\mathbb {R} ^ {d}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.273, + 0.47, + 0.303 + ], + "angle": 0, + "content": "and the motion token \\(t_m \\in \\mathbb{R}^{T' \\times 1}\\) stores the index of the closest vector" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.314, + 0.47, + 0.338 + ], + "angle": 0, + "content": "\\[\nt _ {m, i} = \\arg \\min _ {j} \\| z _ {i} - e _ {j} \\| \\in \\mathbb {R} ^ {1}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.347, + 0.469, + 0.408 + ], + "angle": 0, + "content": "The quantized \\( z_{q} \\) will be decoded with stacked convolutional layers to reconstruct the human motion \\( \\hat{M} \\), and \\( t_{m} \\) will be used in the training stage of the cross-modal transformer." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.409, + 0.469, + 0.485 + ], + "angle": 0, + "content": "For the training of 3D human motion VQ-VAE, we follow the strategy in [48] and the total loss contains a reconstruction loss for human motion regression, a codebook loss for the dictionary learning, and a commitment loss to stabilize the training process:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.494, + 0.469, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v q} = \\| \\hat {M} - M \\| _ {1} + \\| \\operatorname {s g} [ e ] - e _ {q} \\| _ {2} ^ {2} + \\beta \\| e - \\operatorname {s g} [ e _ {q} ] \\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.522, + 0.469, + 0.581 + ], + "angle": 0, + "content": "where \\(\\mathrm{sg}[\\cdot ]\\) is 'stop gradient' and \\(\\beta\\) is the factor term to adjust the weight of the commitment loss. A straight-through estimator is also employed to pass the gradient from the decoder to the encoder in back-propagation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.582, + 0.47, + 0.673 + ], + "angle": 0, + "content": "As we will show in Sec. 4, the motion tokens from both datasets encoded by 3D human motion VQ-VAE fall almost in the shared latent space, which shows the feasibility of using the separate music-conditioned and text-conditioned motion datasets for music-text conditioned 3D dance generation task." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.682, + 0.314, + 0.696 + ], + "angle": 0, + "content": "3.3. Cross-modal Transformer" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.469, + 0.839 + ], + "angle": 0, + "content": "The cross-modal transformer contains an audio encoder, a text encoder, and a motion decoder. Since we use two separate datasets for our task, the cross-modal transformer is divided into two branches, where one for music2dance and the other for text2motion. It takes the audio feature, the text feature, and the motion token \\( t_m \\) encoded by 3D human motion VQ-VAE as inputs, and performs the sequence-to-sequence translation task to generate the future motion tokens." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Attention. Attention is introduced in Transformer [49] for natural language processing, and then applied to many domains [30, 8, 20, 39, 11, 53, 32, 55, 40] A \\(L\\)-layer transformer typically consists of \\(L\\) transformer blocks, which" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "contains a multi-head self-attention (MSA) and a feed forward network (FFN). Given the input \\( x \\in \\mathbb{R}^{N \\times c} \\), where \\( N \\) is the sequence length and \\( c \\) is the dimension of the input, the transformer block first maps it to keys \\( K \\in \\mathbb{R}^{N \\times c} \\), queries \\( Q \\in \\mathbb{R}^{N \\times c} \\), values \\( V \\in \\mathbb{R}^{N \\times c} \\) with the linear projections, and then a self-attention operation is performed by" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.193, + 0.892, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {c}}\\right) V. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.892, + 0.384 + ], + "angle": 0, + "content": "The output is followed by a normalization layer and a FFN. Audio encoder. Given a sequence of music, the audio encoder first extracts the raw audio features following Bailando [44], where a public audio processing toolbox, Librosa [22], is employed to obtain the mel frequency cepstral coefficients (MFCC), MFCC delta, constant- \\(Q\\) chromagram, tempogram and onset strength. After that, an embedded layer followed by several transformer blocks containing self-attention operation e.g., Eq. (4), is used to generate the processed audio features \\(f_{a}\\in \\mathbb{R}^{T^{\\prime}\\times d}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Text encoder. Given a text instruction, we first extract the text token with a GloVe [33], and then an embedded layer followed by several transformer blocks is performed to obtain a processed text feature \\( f_{t} \\in \\mathbb{R}^{n \\times d} \\), where \\( n \\) is the length of the text feature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.893, + 0.716 + ], + "angle": 0, + "content": "Motion decoder. Motion decoder takes the motion tokens encoded by the 3D human motion VQ-VAE as inputs and outputs the future motion tokens, which has the same spirit as the sequence-to-sequence translation task. Since the information after time \\( t \\) is unknown at moment \\( t \\), a masked MSA is first employed to interact with motions at different times. The mask has the shape of a simple upper triangular matrix and is performed in Eq. (4). For the moment \\( t \\), only the motions before moment \\( t \\) are able to perform self-attention operations. Motion decoder also maps the extracted audio feature \\( f_{a} \\) and text feature \\( f_{t} \\) via audio encoder and text encoder to \\( K \\) and \\( Q \\), to perform cross-modal attention with motion tokens. To enable the music-text conditional dance generation, (i.e., feature fusion, we apply shared parameters), we use the motion decoder with shared parameters for model efficiency, as shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.753, + 0.742 + ], + "angle": 0, + "content": "3.4. Details of Music-text Fusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The detailed architecture is listed in Figure 3. Given the audio feature and text feature, the motion decoder first processes the past motion tokens with a self-attention layer followed by addition, normalization, a cross-attention with audio and text features separately, addition, normalization, and a feed-forward layer are performed. Such a procedure is repeated \\( L \\) (\\( L = 6 \\)) times to build a typical \\( L \\)-layer transformer. Then we adopt a late fusion strategy to perform a weighted sum of features of audio and text at a specific time (i.e., effect range). The weight curve is shown in Figure 3," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9945" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.26, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.089, + 0.468, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.313, + 0.471, + 0.418 + ], + "angle": 0, + "content": "Figure 3: Details of music-text feature fusion: Given the audio and text features from the encoder, the decoder processes them separately in the early layers. Then the audiotext fusion layer is applied by the weighted sum at the effect range with weight cure. Finally, a linear projection layer and a softmax operation are applied to predict the music-text conditioned motion tokens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.598 + ], + "angle": 0, + "content": "where we slightly increase the weight of the text feature by a half cosine curve until a peak value of 0.8 at the beginning (20% time of effect range), and decrease it by a half cosine curve at the end (20% time of effect range). The weight of audio feature \\( W_{audio} \\) is 1 - \\( W_{text} \\) to ensure the feature keeps the same scale. With the fused feature, a linear projection layer and a softmax operation are applied to predict the music-text conditioned motion tokens, which are then decoded by the decoder of human motion VQ-VAE to obtain the music-text conditioned 3D dance sequences." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.291, + 0.621 + ], + "angle": 0, + "content": "3.5. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.75 + ], + "angle": 0, + "content": "Training. To train the 3D human motion VQ-VAE, we crop 64 frames with a sliding window from the original motion sequences as inputs, i.e., \\( T = 64 \\), for both music2dance and text2motion datasets. We use a three-layer encoder and decoder so that the time interval of downsampling \\( t \\) is 8. We randomly sample motions from both datasets and employ the Adam optimizer with a batch size of 128 and a learning rate of \\( 1e^{-4} \\) to optimizer our 3D human motion VQ-VAE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To train the cross-modal transformer, we split two branches into two streams to train the music2dance and text2motion tasks on both datasets. For the music2dance task, we use the same frequency as the motion token to sample the music vector so that the music feature has the same time length as the motion token. We perform the sequence to sequence translation with the motion decoder to obtain the prediction of motion tokens conditioned by the music input. As for the text2motion task, since the text and motion token have different lengths, we use a maximum text length" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "of 84 for text2motion translation with a padding strategy. We employ 6 self-attention layers for the audio encoder, text encoder, and motion decoder with the hidden dimension of 512 and 8 heads. The cross entropy loss is adopted for both music2dance and text2motion tasks" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.177, + 0.892, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = - \\frac {1}{m} \\sum_ {i = 0} ^ {m - 1} \\sum_ {j = 0} ^ {C - 1} y _ {i j} \\log y _ {\\hat {i j}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.892, + 0.321 + ], + "angle": 0, + "content": "where \\(\\hat{y}\\) is the prediction of motion token and \\(y\\) is the ground-truth. \\(m\\) is the length of motion tokens and \\(C\\) is the number of classes of motion tokens, i.e., the length of codebook of 3D human motion VQ-VAE. Both tasks are simultaneously optimized with a batch size of 64 and a learning rate of \\(1e^{-4}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.893, + 0.535 + ], + "angle": 0, + "content": "Inference. Our aim is to generate 3D dance with music-text integration. Therefore, at the inference stage, we first fuse the music feature and text feature extracted by the audio encoder and text encoder with a weighted sum. Specifically, given an audio feature with a time length of \\( T' \\) and a text feature with the length of \\( n \\), we first feed them into the motion decoder for future motion token prediction. Then we adopt the late fusion strategy to have a weighted sum of the generated motion tokens from both audio and text features at the duration of the integrated text, followed by a linear projection layer and a softmax operation, to obtain the combined motion tokens. Finally, the combined motion tokens are decoded into a 3D dance sequence after going through the decoder of our 3D human motion VQ-VAE model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.55, + 0.634, + 0.567 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.707, + 0.591 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Datasets. Since there are no paired 3D dance motions conditioned on both music and text for our task, we combine the existing music2dance and text2motion datasets. For the music2dance dataset, we employ the AIST++ dataset [29]. For the text2motion dataset, we employ the HumanML3D [16] dataset. Additionally, we also collect a new in-the-wild music dataset to evaluate the generalization ability of our method, which contains 82 wild music clips from Youtube. The total of 53 minutes duration (8x larger size than the AIST++ test set) and various styles and content of music, which faithfully lies out of the distribution of AIST++ (explained in supplementary material). The evaluation on our dataset better reflects the generalization ability of the models in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Baselines. Since we are the first to propose this bimodality driven 3D dance generation task, there are no existing methods for comparisons. We implement a traditional motion editing algorithm, slerp [43], which aligns the last generated frame of the dance with the first frame of the motion by text with a transition window of 10 frames, and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9946" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "applies spherical interpolation of quaternion to fill in the transition in between. To validate the effectiveness of our method on music2dance task, we compare our method to the previous music2dance works [28, 57, 21, 29, 44]. To validate the effectiveness of our method on text2motion task, we compare our method to the previous text2motion works [31, 4, 7, 13, 47, 25, 17]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.47, + 0.303 + ], + "angle": 0, + "content": "Evaluation metrics. We adopt the same evaluation settings as suggested by FACT [29] and Bailando [44] to evaluate the dance generation quality, including Fréchet Inception Distances (FID) [19], Diversity, and Beat Align Score. For the text2motion quality, we adopt the same evaluation settings as suggested by TM2T [17], including R-precision, Multimodal-Dist, FID, Diversity, and MultiModality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.471, + 0.484 + ], + "angle": 0, + "content": "In addition, we also propose two new evaluation metrics, Motion Prediction Distance (MPD) and Freezing Score. The former reflects the coherence of frames at the time of integrating text, and the latter reacts to the percentage of temporal freezing frames in the generated dance. For the Freezing Score, we introduce the Percentage of Freezing Frame (PFF) and \\(\\mathrm{AUC}_f\\), where the PFF is defined by measuring the percentage of frozen frames with two criteria: 1) the maximum joint velocity blew a threshold \\((0.015\\mathrm{m / s})\\). 2) its duration exceeds a certain period (3s). The \\(\\mathrm{AUC}_f\\) is defined by area under curve of PFF in the threshold range from 0 to 0.03. As for MPD, it is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.494, + 0.47, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M P D} = \\min _ {i} \\| f _ {i} \\left(M _ {t _ {0} \\rightarrow t _ {1}}\\right) - M _ {t _ {1} \\rightarrow t _ {2}} \\| _ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.688 + ], + "angle": 0, + "content": "where \\(M\\) is the dance motions, \\(f\\) is a motion prediction model, \\(f_{i}(M_{t_{0}\\rightarrow t_{1}})\\) is the i-th predicted possible future motion, and \\(t_0\\), \\(t_1\\), and \\(t_2\\) are the timestamps. It means that the model predicts various potential possible motion from \\(t_1\\) to \\(t_2\\) with the motion from \\(t_0\\) to \\(t_1\\). If this distance is small (i.e., the generated dance \\(M_{t_1\\to t_2}\\) lays in the possible future), then the generated dance motion \\(M_{t_1\\to t_2}\\) is more coherent at the time of integrating text. We adopt the motion prediction model DLow [56] in this evaluation metric as it is designed for diverse potential furture motion prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Perceptual Evaluation. Besides the above metric measuring, we also conduct extensive user studies on Amazon Mechanical Turk (AMT) to perceptually evaluate the visual effects of our generated 3D dance results. Particularly, given each pair of dance movements sampled from our method and others with the same music clip, we request 3 distinct users on AMT to present their preference regarding the music-dance alignment, motion realism, and mobility. We further set the bar of involved users that only the ones with Master recognition who also have finished more than 1,000 tasks with over \\(98\\%\\) approve rate are considered. Overall, there are 55 users employed in our user studies that come from various regions, ages, races and gender. The results of the user study are more representative to show the effect of" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.712, + 0.107 + ], + "angle": 0, + "content": "the generated dance in practice." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.117, + 0.892, + 0.148 + ], + "angle": 0, + "content": "4.2. Evaluation on Music-text Conditioned Dance Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.892, + 0.489 + ], + "angle": 0, + "content": "We validate the results of music-text conditioned dance generation on the test set of AIST++ and our dataset, as illustrated in Table 2. Here we mainly evaluate the coherence of generated dance with MPD metric at the time where music meets text, i.e., the time point where text starts to take effect. We measure the coherence of pure music2dance generation as a baseline named a2d to reflect the influence of with/without text instruction. We use the past 25 motion frames to predict the future 30 frames, and calculate the MPD from future frame (ft) = 10 to ft = 30, respectively. Our method consistently outperforms slerp [43] and a2d baseline in both datasets, and gains a similar result compared to the ground-truth shown in Table 2, which verifies the naturalness of our generated motion. More importantly, user study experiments show that our method generates more realistic 3D dance compared to dance generation only conditioned music, and \\(70.0\\%\\) Win rate even compared to the ground-truth shown in Table 2. The results of our user study indicate that the music-text conditioned dance generation received high ratings from participants, highlighting the importance of considering audience preferences in evaluating dance quality." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.499, + 0.892, + 0.53 + ], + "angle": 0, + "content": "4.3. Evaluation on Music Conditioned Dance Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.674 + ], + "angle": 0, + "content": "To validate the effectiveness of our architecture, we quantitatively compare our method with state-of-the-art those for music conditioned 3D dance generation. The results on the test set of AIST++ are shown in Table 1. We can find that our method outperforms the previous ones in terms of motion diversity \\((\\mathrm{Div}_k\\) and \\(\\mathrm{Div}_g)\\), while the performance of the motion quality \\((\\mathrm{FID}_g)\\) and beat align score is inferior under the condition that we do not use seed motion from ground-truth compared with FACT [29] and Bailando [44]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The existing evaluation metrics are not sufficient to reflect the quality of generated dance in practice, which motivates us to propose a PFF and \\(\\mathrm{AUC}_f\\) to evaluate the percentage of freezing frames. It is worth noting that our architecture outperforms FACT [29] and Bailando [44] in terms of PFF and \\(\\mathrm{AUC}_f\\), which shows that our method rarely generates frozen frames. Meanwhile, Li et al. [28] and DanceNet [57] also gain nearly zero freezing in PFF and \\(\\mathrm{AUC}_f\\). The reason is that the generated dances of Li et al. [28] are highly jittery making its velocity variation extremely high, which is also reported in [29, 44], and leads to the non-freezing issue. And for DanceNet [57], it generates dance in a repeat motion pattern, nearly zero freezing but with low diversity. Furthermore, the results of user study show that the generated 3D dance is more visually re" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9947" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.887, + 0.33 + ], + "angle": 0, + "content": "
DataMethodMotion QualityMotion DiversityFreezingBeat Align Score ↑User Study
FIDk ↓FIDg ↑Divk ↑Divg ↑PFF↓AUCf ↓Our Method Wins
AIST ++Ground-truth17.1010.608.197.450.000.000.237441.9%
Li et al. [28]86.4343.466.85*3.320.000.000.160798.3%
DanceNet [57]69.1825.492.862.850.000.980.143087.2%
DanceRevolution [21]73.4225.923.524.8711.0112.220.195070.5%
FACT [29]35.3522.115.946.1825.2921.590.220983.3%
Bailando [44]28.169.627.836.3414.9113.250.233265.0%
ours (only dance data)23.949.537.694.530.000.000.2127-
ours19.0120.099.456.360.000.000.2049-
Wild AudioFACT [29]70.3620.237.336.3432.6427.210.221189.6%
Bailando [44]50.5622.553.806.0417.5514.140.216670.8%
ours (only dance data)43.8513.088.524.760.780.760.1998-
ours27.6520.347.885.270.120.110.2290-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.34, + 0.893, + 0.404 + ], + "angle": 0, + "content": "Table 1: Music Conditioned Dance Generation: quantitative results on AIST++ and Wild Audio test set. The best and runner-up values are bold and underlined, respectively. Among compared methods, \"Li et al.\", DanceNet and FACT are multiplexing the same results of AIST++ benchmark [28], while DanceRevolution [21] is followed from Bailando [44]. \\(\\dagger\\) FID\\(_k\\) and DIV\\(_k\\) are fetched from [29] while FID\\(_g\\) and DIV\\(_g\\) are fetched from [44]." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.426, + 0.462, + 0.597 + ], + "angle": 0, + "content": "
DataMethodMotion Prediction DistanceUser Study
ft=10ft=20ft=30Our Method Wins
AIST ++GT0.0480.0720.08870.0%
a2d0.0520.0840.10866.6%
slerp [43]0.0880.1220.13573.3%
ours0.0490.0800.102-
Wild Audioa2d0.0520.0880.11360.0%
slerp [43]0.1040.1480.17170.0%
ours0.0480.0790.107-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.655 + ], + "angle": 0, + "content": "Table 2: Music-text conditioned Dance Generation: quantitative results on AIST++ and wild audio test set. The best and runner-up values are bold and underlined, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.469, + 0.748 + ], + "angle": 0, + "content": "alistic compared to other methods. Even in comparison to the ground truth, \\(41.9\\%\\) of our generated dance is voted as the better in average. We also report the result trained with dance-only data, which shows comparable performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In addition to the evaluation on the test set of AIST++, we also show the experimental results on our in-the-wild dataset in Table 1. Our method outperforms FACT [29] and Bailando [44] in almost all metrics except for \\(\\mathrm{FID}_g\\mathrm{Div}_g\\). We can observe that both FACT [29] and Bailando [44] shows a large performance drop in terms of \\(\\mathrm{FID}_k\\), while ours maintain a small change. This is because FACT [29] and Bailando [44] requires seed motion, however, there is no ground-truth for in-the-wild scenario. With random sampled seed motion or token, their methods are not adapted" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.893, + 0.55 + ], + "angle": 0, + "content": "well for the in-the-wild scenarios. One can also notice that there exists freezing in our method on the in-the-wild dataset but the percentage of frozen frames is zero on the test set of AIST++, which results from the different distributions of AIST++ and our dataset. It also shows that our dataset is more challenging. We also report the result trained with dance-only data, which shows the advantage of our mix-training strategy in the wild scenario." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.558, + 0.892, + 0.589 + ], + "angle": 0, + "content": "4.4. Evaluation on Text Conditioned Motion Generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.893, + 0.825 + ], + "angle": 0, + "content": "We evaluate our text2motion approach in two different scenarios: inference with text only (t2m), and inference with text and music feature fusion. Table 3 shows the results. In the text-only setting, our approach achieves comparable performance with TM2T baseline(T) [17], which demonstrates that the mixed data/tasks training does not affect the quality of text2motion generation. We then apply our late fusion method by randomly sampling a music clip (from the whole AIST++ test set) and a text (from the whole t2m test set), and evaluate the generated dance clips following the same protocol as TM2T [17]. As Table 3 indicates, the text-dance consistency remains acceptable with a late fusion ratio of 0.8. For more details on the relation between the late fusion rate and text2motion result, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.684, + 0.849 + ], + "angle": 0, + "content": "4.5. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.903 + ], + "angle": 0, + "content": "We also show the qualitative results of our method for music-text conditioned 3D dance generation in Figure 4, Figure 5 and Figure 6. Specifically, Figure 4 shows the gen" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9948" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.457, + 0.239 + ], + "angle": 0, + "content": "
MethodsR Precision↑FID↓MM Dist↓Diversity→MModality↑
Real motions0.5110.0022.9749.503
Seq2Seq [31]0.18011.755.5296.223
Language2Pose [4]0.24611.025.2967.676
Text2Gesture [7]0.1655.0126.0306.409
Hier [13]0.3016.5325.0128.332
MoCoGAN [47]0.03794.419.6430.462
Dance2Music [25]0.03366.988.1160.725
TM2T baseline(T) [17]0.3511.6694.0469.632
TM2T [17]0.4241.5013.4678.589
TM2D (t2m)0.3191.0214.0989.513
TM2D (LFR 0.8)0.3001.1054.3078.887
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.294 + ], + "angle": 0, + "content": "Table 3: The evaluation of text2motion on the HumanML3D dataset, where the results are averaged from 20 runs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.468, + 0.442 + ], + "angle": 0, + "content": "erated 3D dance with the same music sequence but with different text instructions. Figure 5 shows the generated 3D dance with the same text instruction but with a different start time. Figure 6 shows the generated 3D dance with the same text instruction but with different durations. As one can see, our approach maintains plausible visual results according to the text instructions for all three cases, which confirms that our approach is more flexible." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.453, + 0.317, + 0.468 + ], + "angle": 0, + "content": "4.6. The Impact of Mixed Data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.468, + 0.806 + ], + "angle": 0, + "content": "We employ two datasets (AIST++ and HumanML3D) to train our 3D human motion VQ-VAE so that the motion tokens drop in the same latent space. To empirically show this point, we conduct experiments to count the number of shared motion tokens. For a trained 3D human motion VQ-VAE with AIST++ and HumanML3D, there are 855 vectors and 912 vectors from codebook used to construct the motion token in AIST++ and HumanML3D. The total number of vectors contained in the codebook is 1024. among them, 846 vectors (98.9% in AIST++ and 92.8% in HumanML3D) are shared to generate the motion tokens, which shows the feasibility of using the separate music-conditioned and text-conditioned motion datasets for music-text conditioned 3D dance generation task. We further verify the latent space of the motions by VQ-VAE, and we perform t-SNE visualizations of the raw motion distribution (before VQ-VAE) and feature distribution encoded by VQ-VAE (after VQ-VAE) from a2d motion and t2m motion in Figure 7. Two motion datasets are mixed successfully, which provides the potential integration of bimodality dance generation. Refer to the supplementary materials for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.242, + 0.833 + ], + "angle": 0, + "content": "4.7. Model Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We also show the efficiency of our architecture in Table 4. In the inference stage, the complete model parameters and the inference time for 1s music are compared. Our architecture requires only about half model parameters and" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.093, + 0.885, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.245, + 0.892, + 0.32 + ], + "angle": 0, + "content": "Figure 4: Same music, different actions. These two generated dance share the same music and effect range (from 6s to 8s), with different text instructions: \"A person does a cartwheel\" (top), \"A person is spinning with arms spread out\" (bottom)." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.339, + 0.89, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.561 + ], + "angle": 0, + "content": "Figure 5: Same action, different start time points. These two generated dance share the same music and text instruction (\"A person jumps up and down\") but different effect ranges: from 6 to 8s (top), from 7 to 9s (bottom)." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.579, + 0.89, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Figure 6: Same action, different durations. These two generated dance share the same music and text instruction (\"A person is keeping jumping\") and action start time point (i.e., 6s) but different effect duration: 2s (top), 3s (bottom, i.e., jumped twice)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.891, + 0.901 + ], + "angle": 0, + "content": "inference time compared to Bailando [44], which attributes" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.114, + 0.264, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.218, + 0.227, + 0.23 + ], + "angle": 0, + "content": "(a) Before VQ-VAE" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.114, + 0.459, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.218, + 0.422, + 0.23 + ], + "angle": 0, + "content": "(b) After VQ-VAE" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.242, + 0.47, + 0.286 + ], + "angle": 0, + "content": "Figure 7: The t-SNE visualizations of motions before and after VQ-VAE (orange: motions in AIST++, blue: motions in HumanML3D)." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.302, + 0.443, + 0.386 + ], + "angle": 0, + "content": "
MethodParametersInference time
FACT [29]120M8.300s/(1s music)
Bailando [44]152M0.236s/(1s music)
ours72M0.143s/(1s music)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.088, + 0.396, + 0.457, + 0.411 + ], + "angle": 0, + "content": "Table 4: Comparisons of parameters and inference time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.558 + ], + "angle": 0, + "content": "that we do not need to encode separate features for the upper and lower half body. Compared to FACT [29], our architectures significantly reduce the inference time. The main reason is that FACT employs two transformers for the music, dance motion, and a cross-modal transformer, but we formulate it as the standard sequence-to-sequence translation task with one transformer only, which establishes the advantage of our architecture." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.196, + 0.589 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.854 + ], + "angle": 0, + "content": "In this paper, a novel task that simultaneously integrates both music and text instruction for 3D dance generation is proposed. Due to the lack of the paired 3D dance motions conditioned on both music and text, we resort to two existing datasets, i.e., music2dance and text2motion, to perform this task and employ a 3D human motion VQ-VAE to project the motions of the two datasets into a shared latent space so that the two datasets with different distributions can be effectively mixed for training. Moreover, we also propose a cross-modal transformer architecture to generate 3D dance without degrading the performance of music-conditioned dance generation. Two new evaluation metrics, MPD and FS, are proposed to reflect the quality of generated motion. Extensive experiments show that our method can generate dance motion that matches both music and text in a realistic and coherent way while maintaining comparable performance on two single modalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This project is supported by the Ministry of Education Singapore under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "and Singapore National Research Foundation (\"CogniVision - Energy-autonomous always-on cognitive and attentive cameras for distributed real-time vision with milliwatt power consumption\" grant NRF-CRP20-2017-0003) - www.green-ic.org/CogniVision." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.179, + 0.597, + 0.194 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.204, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[1] mixamo. https://www MIXamo.com/. Accessed: 2023-03-07. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.233, + 0.894, + 0.3 + ], + "angle": 0, + "content": "[2] Hyemin Ahn, Timothy Ha, Yunho Choi, Hwiyeon Yoo, and Songhwai Oh. Text2action: Generative adversarial synthesis from language to action. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5915-5920. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.302, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[3] Hyemin Ahn, Jaehun Kim, Kihyun Kim, and Songhwai Oh. Generative autoregressive networks for 3d dancing move synthesis from music. IEEE Robotics and Automation Letters, 5(2):3501-3508, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.359, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[4] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.414, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[5] Omid Alemi, Jules Françoise, and Philippe Pasquier. Groovenet: Real-time music-driven dance movement generation using artificial neural networks. In Workshop on Machine Learning for Creativity, 23rd ACM SIGKDD Conference on Knowledge Discovery and Data Mining, volume 8, page 26, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.498, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[6] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. Teach: Temporal action compositions for 3d humans. In International Conference on 3D Vision (3DV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.553, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[7] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In IEEE Virtual Reality and 3D User Interfaces (VR), pages 1-10. IEEE, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.637, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[8] Hanting Chen, Yunhe Wang, Jianyuan Guo, and Dacheng Tao. VanillaNet: the Power of Minimalism in Deep Learning. arXiv preprint arXiv:2305.12972, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[9] Kang Chen, Zhipeng Tan, Jin Lei, Song-Hai Zhang, Yuan-Chen Guo, Weidong Zhang, and Shi-Min Hu. Choreomaster: choreography-oriented music-driven dance synthesis. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[10] Rukun Fan, Songhua Xu, and Weidong Geng. Example-based automatic music-driven conventional dance motion synthesis. IEEE transactions on visualization and computer graphics, 18(3):501-515, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.79, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[11] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. arXiv preprint arXiv:2305.10924, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[12] Joao P Ferreira, Thiago M Coutinho, Thiago L Gomes, José F Neto, and et al. Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics, 94:11-21, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.204, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9950" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[13] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1396–1406, 2021. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[14] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[15] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.47, + 0.36 + ], + "angle": 0, + "content": "[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597, 2022. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[18] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in neural information processing systems, volume 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[20] Duc N.M. Hoang, Shiwei Liu, Radu Marculescu, and Zhangyang Wang. Revisiting pruning at initialization through the lens of ramanujan graph. In International Conference on Learning Representations, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[21] Ruozhi Huang, Huang Hu, Wei Wu, Kei Sawada, Mi Zhang, and Daxin Jiang. Dance revolution: Long-term dance generation with music via curriculum learning. In International Conference on Learning Representations, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[22] Yanghua Jin, Jiakai Zhang, Minjun Li, Yingtao Tian, Huachun Zhu, and Zhihao Fang. Towards the automatic anime characters creation with generative adversarial networks. arXiv preprint arXiv:1708.05509, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[23] Hsuan-Kai Kao and Li Su. Temporally guided music-to-body-movement generation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 147-155, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[24] Kosmas Kritsis, Aggelos Gkiokas, Aggelos Pikrakis, and Vassilis Katsouros. Danceconv: Dance motion generation with convolutional networks. IEEE Access, 10:44982-45000, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[25] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. Advances in neural information processing systems, 32, 2019. 1, 2, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[26] Juheon Lee, Seohyun Kim, and Kyogu Lee. Listen to dance: Music-driven choreography generation using autoregressive encoder-decoder network. arXiv preprint arXiv:1811.00818, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[27] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[28] Jiaman Li, Yihang Yin, Hang Chu, Yi Zhou, Tingwu Wang, Sanja Fidler, and Hao Li. Learning to generate diverse dance motions with transformer. arXiv preprint arXiv:2008.08171, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[29] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 1, 2, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[30] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[31] Angela S. Lin, Lemeng Wu, Rodolfo Corona, Kevin W. H. Tai, Qixing Huang, and Raymond J. Mooney. Generating animated videos of human activities from natural language descriptions. 2018. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[32] Xinyin Ma, Gongfan Fang, and Xinchao Wang. LLM-Pruner: On the Structural Pruning of Large Language Models. arXiv preprint arXiv:2305.11627, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.506, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[33] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[34] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[35] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, page 480-497, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236-252, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[37] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems, 109:13-26, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9951" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[39] Sucheng Ren, Xingyi Yang, Songhua Liu, and Xinchao Wang. SG-Former: Self-guided Transformer with Evolving Token Reallocation. In IEEE International Conference on Computer Vision, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[40] Sucheng Ren, Daquan Zhou, Shengfeng He, Jiashi Feng, and Xinchao Wang. Shunted Self-Attention via Multi-Scale Token Aggregation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[41] Xuanchi Ren, Haoran Li, Zijian Huang, and Qifeng Chen. Self-supervised dance video synthesis conditioned on music. In Proceedings of the 28th ACM International Conference on Multimedia, pages 46-54, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.47, + 0.329 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[43] Ken Shoemake. Animating rotation with quaternion curves. In Proceedings of the 12th annual conference on Computer graphics and interactive techniques, pages 245-254, 1985. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[44] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11050-11059, 2022. 1, 2, 4, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[45] Guofei Sun, Yongkang Wong, Zhiyong Cheng, Mohan S Kankanhalli, Weidong Geng, and Xiangdong Li. Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Transactions on Multimedia, 23:497-509, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[46] Taoran Tang, Jia Jia, and Hanyang Mao. Dance with melody: An LSTM-autoencoder approach to music-oriented dance synthesis. In Proceedings of the 26th ACM international conference on Multimedia, pages 1598-1606, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[47] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[48] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.762 + ], + "angle": 0, + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[50] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[51] Nelson Yalta, Shinji Watanabe, Kazuhiro Nakadai, and Tetsuya Ogata. Weakly-supervised deep recurrent neural networks for basic dance step generation. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[52] Tatsuro Yamada, Hiroyuki Matsunaga, and Tetsuya Ogata. Paired recurrent autoencoders for bidirectional translation between robot actions and linguistic descriptions. IEEE Robotics and Automation Letters, 3(4):3441-3448, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[53] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Advances in neural information processing systems, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[54] Zijie Ye, Haozhe Wu, Jia Jia, Yaohua Bu, Wei Chen, Fanbo Meng, and Yanfeng Wang. Choreonet: Towards music to dance synthesis with choreographic action unit. In Proceedings of the 28th ACM International Conference on Multimedia, pages 744-752, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[55] Weihao Yu, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. Metaformer Baselines for Vision. arXiv preprint arXiv:2210.13452, 2022.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[56] Ye Yuan and Kris Kitani. Dlow: Diversifying latent flows for diverse human motion prediction. In European Conference on Computer Vision, pages 346-364. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[57] Wenlin Zhuang, Congyi Wang, Jinxiang Chai, Yangang Wang, Ming Shao, and Siyu Xia. Music2dance: Dancenet for music-driven dance generation. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 18(2):1-21, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[58] Wenlin Zhuang, Yangang Wang, Joseph Robinson, Congyi Wang, Ming Shao, Yun Fu, and Siyu Xia. Towards 3d dance motion synthesis and control. arXiv preprint arXiv:2006.05743, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9952" + } + ] +] \ No newline at end of file diff --git a/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/1834d63b-e3e0-4b42-9900-788dbb6bee22_origin.pdf b/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/1834d63b-e3e0-4b42-9900-788dbb6bee22_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9800e48e12c7f8635a69fcaadd162d52c5adc7fa --- /dev/null +++ b/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/1834d63b-e3e0-4b42-9900-788dbb6bee22_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd5b131d713359626f2b94656dc4781fcaf6a16e14d0ff7facef3f42c7beb400 +size 2420335 diff --git a/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/full.md b/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6bc6b5e4ab4abf57fcdb2549bd5b60dae0a08bfe --- /dev/null +++ b/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/full.md @@ -0,0 +1,309 @@ +# TM2D: Bimodality Driven 3D Dance Generation via Music-Text Integration + +Kehong Gong1* +Zihang Jiang1 + +Dongze Lian $^{1*}$ +Xinxin Zuo $^{2}$ + +$^{1}$ National University of Singapore + +Heng Chang $^{2}$ +Michael Bi Mi $^{2}$ + +Chuan Guo² +Xinchao Wang¹† + +$^{2}$ Huawei Technologies Co., Ltd. + +![](images/1f6ebcdb8087e5d1836998117936b0bbe0d692f6ead0982b3e17b1d70968080c.jpg) +Figure 1: Generated 3D dance examples conditioned music and text with our method. Given an audio input and text input with a specific starting time and duration, our method is able to generate a sequence of dance motions that fit the music and text instruction. The character is from Mixamo [1]. + +# Abstract + +We propose a novel task for generating 3D dance movements that simultaneously incorporate both text and music modalities. Unlike existing works that generate dance movements using a single modality such as music, our goal is to produce richer dance movements guided by the instructive information provided by the text. However, the lack of paired motion data with both music and text modalities limits the ability to generate dance movements that integrate both. To alleviate this challenge, we propose to utilize a 3D human motion VQ-VAE to project the motions of the two datasets into a latent space consisting of quantized vectors, which effectively mix the motion tokens from the two datasets with different distributions for training. Additionally, we propose a cross-modal transformer to integrate text instructions into motion generation architecture for generating 3D dance movements without degrading the performance of music-conditioned dance generation. To better evaluate the quality of the generated motion, we introduce two novel metrics, namely Motion Prediction Distance (MPD) and Freezing Score (FS), to measure the coherence and freezing percentage of the generated motion. Extensive experiments show that our approach can gener + +ate realistic and coherent dance movements conditioned on both text and music while maintaining comparable performance with the two single modalities. Code is available at https://garfield-kh.github.io/TM2D/. + +# 1. Introduction + +The music-conditioned dance generation has become a topic of great interest in recent years. The ability to generate dance movements that are synchronized with music has numerous applications, such as behavior understanding, simulation, and benefiting the community of dancers and musicians [25, 9, 29, 44]. Although music has been used as a guidance to generate dance movements, another important modality cue, text (or language), which provides richer actions and more flexible motion guidance, and is studied in other tasks such as image classification [38], detection [15], segmentation [50], and text-driven image generation [42], has not been fully explored in dance generation. To this end, we first propose a novel task for generating 3D dance movements that simultaneously incorporate both text and music modalities, enabling the generated human to perform rich dancing movements in accordance with the music and text. + +Designing a system pipeline for this bimodality driven 3D dance generation task is non-trivial. There exist two + +significant challenges to be considered: i) the existing datasets only cater to either music-driven (music2dance) [46, 5, 57, 29] or text-driven (text2motion) [36, 16] human motion generation, and no paired 3D dance generation dataset exists that takes into account both music and text. While building a new large-scale paired 3D dance dataset based on music and text is possible, it is time-consuming with fully annotated 3D human motion [29]; ii) the integration of text into music-conditioned dance generation requires a suitable architecture. However, existing methods that use music as a driving force to generate dance movements might result in temporal-freezing frames or fail to generalize to in-the-wild scenarios [29, 44]. Therefore, simple integration of text into the existing music-conditioned architecture might pose a risk of degraded dance generation quality in our new task. + +To address the first challenge, we take advantage of existing music-dance and text-motion datasets for this new task. However, directly mixing motions from these two datasets would result in inferior performance since the motions from these two datasets are in completely different motion spaces. To overcome this, we propose to utilize a VQ-VAE architecture to project the motions into a consistent and shared latent space. In particular, we build up a shared codebook for all the motions from the training set, and motions from both datasets are now represented as discrete tokens that are implicitly constrained to fall into a shared latent space. For the second challenge, we propose to utilize a cross-modal transformer architecture that formulates both music2dance and text2motion as sequence-to-sequence translation tasks. This architecture directly translates audio and text features into motion tokens and enables bimodality driven ability by introducing a fusion strategy in the latent space with a shared motion decoder for both tasks. With the shared decoder, audio and text information can be efficiently fused during inference. Our entire cross-modal transformer architecture is both effective and efficient, allowing for the integration of text instructions to generate coherent 3D dance motions, as illustrated in Figure 1. + +To better evaluate the coherence of generated dance in our task, we propose a new evaluation metric, Motion Prediction Distance (MPD), which measures the distance between the predicted motion and the ground truth at the time of integrating text, thereby providing a more accurate evaluation of the coherence of frames. Additionally, we introduce a Freezing Score (FS) that quantifies the percentage of temporal freezing frames in dance generation, which is a common problem in music-conditioned dance generation. To better evaluate the performance of our method in real-world scenarios, we also collect some in-the-wild music data for evaluation. Our method successfully performs dance generation based on both text and music while maintaining comparable performance on the single modal- + +ity tasks (music2dance, text2motion) compared to other state-of-the-art methods. + +In summary, our contributions are as follows: i) We propose an interesting task of utilizing both music and text for 3D dance generation and propose a pipeline named TM2D (Text-Music to Dance) for this task. ii) Rather than collecting a new training set, we effectively combine the existing music2dance and text2motion datasets and employ a VQ-VAE framework to encode motions from all training sets to a shared feature space. iii) We propose a cross-modal transformer as well as a bimodal feature fusion strategy to encode both audio and text features, which is both effective and efficient. iv) We propose two new metrics, MPD and FS, which efficiently reflect the quality of generated motion. v) We successfully generate realistic and coherent dance based on both music and text instructions while maintaining comparable performance on the single modality tasks (music2dance, text2motion). + +# 2. Related Work + +# 2.1. Music to Dance + +Music2Dance is typically divided into 2D and 3D dance generation and has been explored for many years. Recent methods model 3D dance generation from the perspective of network architecture. For instance, Lee et al. [26] explore the convolutional neural network in casual dilated setting. Lee et al. [25] propose a dance unit with Variational Auto-Encoder (VAE). Ren et al. [41] employ recurrent neural network (RNN) and the graph convolution network is introduced in [12]. As for the 3D dance generation, [54, 3, 24] implement convolutional neural network, [45, 14] apply adversarial learning in generated dance. [5, 46, 51, 58, 23] implement Long short-term memory (LSTM), and [9] implement motion graph with learned music/dance embedding through matching approach [10]. More recently, transformer has been applied in dance generation [28, 29, 44, 27]. However, the previous methods usually produce temporal freezing frames when generating the long sequences, or are difficult to generalize to in-the-wild music, thereby not satisfiable when directly used for bimodal driven 3D dance generation. In this paper, we design a cross modal transformer which is effective and efficient to integrate text instruction to generate the 3D dance. + +# 2.2. Text to Motion + +In addition to music-driven motion generation, text is also utilized as instructions to generate motions. Text2motion can be categorized into action label based motion generation and language description based motion generation. The action label based methods [18, 34] generate motion conditioned on action labels. Action2Motion [18] implement a GRU-VAE to iteratively generate the next + +![](images/b2acb0fe0656441ef4b950020dc31cabf4e42c64014fc6fe9df4c6a086485f3a.jpg) +Figure 2: Our proposed pipeline for music-text conditioned 3D dance generation. Three stages from left to right: 3D human motion VQ-VAE, training stage of the cross-modal transformer, and inference stage of our pipeline. In the first stage, a VQ-VAE is trained with both motions from music2dance and text2motion data, which is then used to tokenize all motions. In the second stage, a dual path cross-modal transformer is employed for sequence2sequence translation tasks (i.e., audio to motion tokens, text to motion tokens), with a shared motion decoder. In the third stage, given audio and text inputs, the audio and text encoders first extract the corresponding features, which are then fused (late fusion) in the motion decoder to generate dance condition on both music and text. + +![](images/6a7a7e5cf65d81ecf02ec94346b5c0a3480756f34855069e0d6b1d76ef27083a.jpg) + +![](images/ef5b69fe33d5d813f5825ea47be9477e8023b6ae80745d50f61adfaf6c59bc33.jpg) + +frame based on the action label and previous frames. ACTOR [34] implements a transformer-VAE to encode and decode the whole pose sequence in one-shot. Since the action label based methods are restricted in a small set of action labels, language description based methods are proposed for more flexible motion generation. [37, 31, 2, 16, 17] formulate the text-to-motion task as a machine translation problem, others [52, 4, 13, 35] learns a joint embedding space of text and motion. Among them, [37, 31, 2, 52, 4, 13] use RNN encoder-decoder to learn the mapping between text and motion. [17, 16] use transformer encoder with RNN decoder, and TM2T [17] further introduces motion to text as an inverse alignment for text to motion. TEMOS [35] and TEACH [6] use transformer for both encoder and decoder for one-shot generation. Different from these text2motion methods, we focus on how to integrate two modalities (i.e., music and text) together for 3D dance generation. + +# 3. Method + +# 3.1. System Pipeline + +The overall pipeline is shown in Figure 2, which consists of three stages. The first stage employs a 3D human motion + +VQ-VAE to encode both the motions of the music2dance and text2motion datasets to a shared codebook encoded by multiple vectors such that each dance motion can be represented as a discrete motion token. After that, the combinations of music, text, and motion token are fed into a cross-modal transformer for training, which effectively learn how to predict a sequential motion tokens according to the previous those as well as music and text. In the third stage, a random starting motion token is generated and inputted to the cross-modal transformer with the given music and text for the sequential motion token prediction, which will be decoded by the pre-trained 3D human motion VQ-VAE in the first stage to generate 3D dance. + +# 3.2. 3D Human Motion VQ-VAE + +Since there are no paired 3D dance motions conditioned on both music and text, we try to tackle our task with the existing music2dance and text2motion datasets. Instead of directly mixing the motions from both datasets for training, we employ a VQ-VAE to project the motions into a consistent and shared latent space, which is represented with a codebook, as illustrated in Figure 2. To be specific, given a 3D human motion $M \in \mathbb{R}^{T \times d_m}$ , where $T$ is the time length + +and $d_{m}$ is the dimension of the human motion, an encoder of VQ-VAE that consists of several 1-D convolutional layers projects $M$ to a latent vector $z\in \mathbb{R}^{T^{\prime}\times d}$ , where $T^{\prime} = \frac{T}{t}$ and $t$ is the time interval for downsampling and $d$ is the dimension of the latent vector. A learnable codebook $e\in \mathbb{R}^{K\times d}$ describes all latent variable features of the whole dataset, where $K$ and $d$ are the length and dimension of the codebook, respectively. A quantized latent vector $z_{q}\in \mathbb{R}^{T^{\prime}\times d}$ will record the closest vector from codebook $e$ as follows + +$$ +z _ {q, i} = \arg \min _ {e _ {j} \in e} \| z _ {i} - e _ {j} \| \in \mathbb {R} ^ {d}, \tag {1} +$$ + +and the motion token $t_m \in \mathbb{R}^{T' \times 1}$ stores the index of the closest vector + +$$ +t _ {m, i} = \arg \min _ {j} \| z _ {i} - e _ {j} \| \in \mathbb {R} ^ {1}. \tag {2} +$$ + +The quantized $z_{q}$ will be decoded with stacked convolutional layers to reconstruct the human motion $\hat{M}$ , and $t_{m}$ will be used in the training stage of the cross-modal transformer. + +For the training of 3D human motion VQ-VAE, we follow the strategy in [48] and the total loss contains a reconstruction loss for human motion regression, a codebook loss for the dictionary learning, and a commitment loss to stabilize the training process: + +$$ +\mathcal {L} _ {v q} = \| \hat {M} - M \| _ {1} + \| \operatorname {s g} [ e ] - e _ {q} \| _ {2} ^ {2} + \beta \| e - \operatorname {s g} [ e _ {q} ] \| _ {2} ^ {2}, \tag {3} +$$ + +where $\mathrm{sg}[\cdot ]$ is 'stop gradient' and $\beta$ is the factor term to adjust the weight of the commitment loss. A straight-through estimator is also employed to pass the gradient from the decoder to the encoder in back-propagation. + +As we will show in Sec. 4, the motion tokens from both datasets encoded by 3D human motion VQ-VAE fall almost in the shared latent space, which shows the feasibility of using the separate music-conditioned and text-conditioned motion datasets for music-text conditioned 3D dance generation task. + +# 3.3. Cross-modal Transformer + +The cross-modal transformer contains an audio encoder, a text encoder, and a motion decoder. Since we use two separate datasets for our task, the cross-modal transformer is divided into two branches, where one for music2dance and the other for text2motion. It takes the audio feature, the text feature, and the motion token $t_m$ encoded by 3D human motion VQ-VAE as inputs, and performs the sequence-to-sequence translation task to generate the future motion tokens. + +Attention. Attention is introduced in Transformer [49] for natural language processing, and then applied to many domains [30, 8, 20, 39, 11, 53, 32, 55, 40] A $L$ -layer transformer typically consists of $L$ transformer blocks, which + +contains a multi-head self-attention (MSA) and a feed forward network (FFN). Given the input $x \in \mathbb{R}^{N \times c}$ , where $N$ is the sequence length and $c$ is the dimension of the input, the transformer block first maps it to keys $K \in \mathbb{R}^{N \times c}$ , queries $Q \in \mathbb{R}^{N \times c}$ , values $V \in \mathbb{R}^{N \times c}$ with the linear projections, and then a self-attention operation is performed by + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {c}}\right) V. \tag {4} +$$ + +The output is followed by a normalization layer and a FFN. Audio encoder. Given a sequence of music, the audio encoder first extracts the raw audio features following Bailando [44], where a public audio processing toolbox, Librosa [22], is employed to obtain the mel frequency cepstral coefficients (MFCC), MFCC delta, constant- $Q$ chromagram, tempogram and onset strength. After that, an embedded layer followed by several transformer blocks containing self-attention operation e.g., Eq. (4), is used to generate the processed audio features $f_{a}\in \mathbb{R}^{T^{\prime}\times d}$ + +Text encoder. Given a text instruction, we first extract the text token with a GloVe [33], and then an embedded layer followed by several transformer blocks is performed to obtain a processed text feature $f_{t} \in \mathbb{R}^{n \times d}$ , where $n$ is the length of the text feature. + +Motion decoder. Motion decoder takes the motion tokens encoded by the 3D human motion VQ-VAE as inputs and outputs the future motion tokens, which has the same spirit as the sequence-to-sequence translation task. Since the information after time $t$ is unknown at moment $t$ , a masked MSA is first employed to interact with motions at different times. The mask has the shape of a simple upper triangular matrix and is performed in Eq. (4). For the moment $t$ , only the motions before moment $t$ are able to perform self-attention operations. Motion decoder also maps the extracted audio feature $f_{a}$ and text feature $f_{t}$ via audio encoder and text encoder to $K$ and $Q$ , to perform cross-modal attention with motion tokens. To enable the music-text conditional dance generation, (i.e., feature fusion, we apply shared parameters), we use the motion decoder with shared parameters for model efficiency, as shown in Figure 2. + +# 3.4. Details of Music-text Fusion + +The detailed architecture is listed in Figure 3. Given the audio feature and text feature, the motion decoder first processes the past motion tokens with a self-attention layer followed by addition, normalization, a cross-attention with audio and text features separately, addition, normalization, and a feed-forward layer are performed. Such a procedure is repeated $L$ ( $L = 6$ ) times to build a typical $L$ -layer transformer. Then we adopt a late fusion strategy to perform a weighted sum of features of audio and text at a specific time (i.e., effect range). The weight curve is shown in Figure 3, + +![](images/a152b169740aa65fb16ca2449b80664d5f90a4b443a91c4499c5ec47685bb5e9.jpg) +Figure 3: Details of music-text feature fusion: Given the audio and text features from the encoder, the decoder processes them separately in the early layers. Then the audiotext fusion layer is applied by the weighted sum at the effect range with weight cure. Finally, a linear projection layer and a softmax operation are applied to predict the music-text conditioned motion tokens. + +![](images/a17d2629f92343d8b02e3e7d407bed0428a76891763c6caf14fc73570a7f1997.jpg) + +where we slightly increase the weight of the text feature by a half cosine curve until a peak value of 0.8 at the beginning (20% time of effect range), and decrease it by a half cosine curve at the end (20% time of effect range). The weight of audio feature $W_{audio}$ is 1 - $W_{text}$ to ensure the feature keeps the same scale. With the fused feature, a linear projection layer and a softmax operation are applied to predict the music-text conditioned motion tokens, which are then decoded by the decoder of human motion VQ-VAE to obtain the music-text conditioned 3D dance sequences. + +# 3.5. Training and Inference + +Training. To train the 3D human motion VQ-VAE, we crop 64 frames with a sliding window from the original motion sequences as inputs, i.e., $T = 64$ , for both music2dance and text2motion datasets. We use a three-layer encoder and decoder so that the time interval of downsampling $t$ is 8. We randomly sample motions from both datasets and employ the Adam optimizer with a batch size of 128 and a learning rate of $1e^{-4}$ to optimizer our 3D human motion VQ-VAE. + +To train the cross-modal transformer, we split two branches into two streams to train the music2dance and text2motion tasks on both datasets. For the music2dance task, we use the same frequency as the motion token to sample the music vector so that the music feature has the same time length as the motion token. We perform the sequence to sequence translation with the motion decoder to obtain the prediction of motion tokens conditioned by the music input. As for the text2motion task, since the text and motion token have different lengths, we use a maximum text length + +of 84 for text2motion translation with a padding strategy. We employ 6 self-attention layers for the audio encoder, text encoder, and motion decoder with the hidden dimension of 512 and 8 heads. The cross entropy loss is adopted for both music2dance and text2motion tasks + +$$ +\mathcal {L} _ {c l s} = - \frac {1}{m} \sum_ {i = 0} ^ {m - 1} \sum_ {j = 0} ^ {C - 1} y _ {i j} \log y _ {\hat {i j}}, \tag {5} +$$ + +where $\hat{y}$ is the prediction of motion token and $y$ is the ground-truth. $m$ is the length of motion tokens and $C$ is the number of classes of motion tokens, i.e., the length of codebook of 3D human motion VQ-VAE. Both tasks are simultaneously optimized with a batch size of 64 and a learning rate of $1e^{-4}$ . + +Inference. Our aim is to generate 3D dance with music-text integration. Therefore, at the inference stage, we first fuse the music feature and text feature extracted by the audio encoder and text encoder with a weighted sum. Specifically, given an audio feature with a time length of $T'$ and a text feature with the length of $n$ , we first feed them into the motion decoder for future motion token prediction. Then we adopt the late fusion strategy to have a weighted sum of the generated motion tokens from both audio and text features at the duration of the integrated text, followed by a linear projection layer and a softmax operation, to obtain the combined motion tokens. Finally, the combined motion tokens are decoded into a 3D dance sequence after going through the decoder of our 3D human motion VQ-VAE model. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. Since there are no paired 3D dance motions conditioned on both music and text for our task, we combine the existing music2dance and text2motion datasets. For the music2dance dataset, we employ the AIST++ dataset [29]. For the text2motion dataset, we employ the HumanML3D [16] dataset. Additionally, we also collect a new in-the-wild music dataset to evaluate the generalization ability of our method, which contains 82 wild music clips from Youtube. The total of 53 minutes duration (8x larger size than the AIST++ test set) and various styles and content of music, which faithfully lies out of the distribution of AIST++ (explained in supplementary material). The evaluation on our dataset better reflects the generalization ability of the models in real-world scenarios. + +Baselines. Since we are the first to propose this bimodality driven 3D dance generation task, there are no existing methods for comparisons. We implement a traditional motion editing algorithm, slerp [43], which aligns the last generated frame of the dance with the first frame of the motion by text with a transition window of 10 frames, and + +applies spherical interpolation of quaternion to fill in the transition in between. To validate the effectiveness of our method on music2dance task, we compare our method to the previous music2dance works [28, 57, 21, 29, 44]. To validate the effectiveness of our method on text2motion task, we compare our method to the previous text2motion works [31, 4, 7, 13, 47, 25, 17]. + +Evaluation metrics. We adopt the same evaluation settings as suggested by FACT [29] and Bailando [44] to evaluate the dance generation quality, including Fréchet Inception Distances (FID) [19], Diversity, and Beat Align Score. For the text2motion quality, we adopt the same evaluation settings as suggested by TM2T [17], including R-precision, Multimodal-Dist, FID, Diversity, and MultiModality. + +In addition, we also propose two new evaluation metrics, Motion Prediction Distance (MPD) and Freezing Score. The former reflects the coherence of frames at the time of integrating text, and the latter reacts to the percentage of temporal freezing frames in the generated dance. For the Freezing Score, we introduce the Percentage of Freezing Frame (PFF) and $\mathrm{AUC}_f$ , where the PFF is defined by measuring the percentage of frozen frames with two criteria: 1) the maximum joint velocity blew a threshold $(0.015\mathrm{m / s})$ . 2) its duration exceeds a certain period (3s). The $\mathrm{AUC}_f$ is defined by area under curve of PFF in the threshold range from 0 to 0.03. As for MPD, it is defined as + +$$ +\mathrm {M P D} = \min _ {i} \| f _ {i} \left(M _ {t _ {0} \rightarrow t _ {1}}\right) - M _ {t _ {1} \rightarrow t _ {2}} \| _ {2}, \tag {6} +$$ + +where $M$ is the dance motions, $f$ is a motion prediction model, $f_{i}(M_{t_{0}\rightarrow t_{1}})$ is the i-th predicted possible future motion, and $t_0$ , $t_1$ , and $t_2$ are the timestamps. It means that the model predicts various potential possible motion from $t_1$ to $t_2$ with the motion from $t_0$ to $t_1$ . If this distance is small (i.e., the generated dance $M_{t_1\to t_2}$ lays in the possible future), then the generated dance motion $M_{t_1\to t_2}$ is more coherent at the time of integrating text. We adopt the motion prediction model DLow [56] in this evaluation metric as it is designed for diverse potential furture motion prediction. + +Perceptual Evaluation. Besides the above metric measuring, we also conduct extensive user studies on Amazon Mechanical Turk (AMT) to perceptually evaluate the visual effects of our generated 3D dance results. Particularly, given each pair of dance movements sampled from our method and others with the same music clip, we request 3 distinct users on AMT to present their preference regarding the music-dance alignment, motion realism, and mobility. We further set the bar of involved users that only the ones with Master recognition who also have finished more than 1,000 tasks with over $98\%$ approve rate are considered. Overall, there are 55 users employed in our user studies that come from various regions, ages, races and gender. The results of the user study are more representative to show the effect of + +the generated dance in practice. + +# 4.2. Evaluation on Music-text Conditioned Dance Generation + +We validate the results of music-text conditioned dance generation on the test set of AIST++ and our dataset, as illustrated in Table 2. Here we mainly evaluate the coherence of generated dance with MPD metric at the time where music meets text, i.e., the time point where text starts to take effect. We measure the coherence of pure music2dance generation as a baseline named a2d to reflect the influence of with/without text instruction. We use the past 25 motion frames to predict the future 30 frames, and calculate the MPD from future frame (ft) = 10 to ft = 30, respectively. Our method consistently outperforms slerp [43] and a2d baseline in both datasets, and gains a similar result compared to the ground-truth shown in Table 2, which verifies the naturalness of our generated motion. More importantly, user study experiments show that our method generates more realistic 3D dance compared to dance generation only conditioned music, and $70.0\%$ Win rate even compared to the ground-truth shown in Table 2. The results of our user study indicate that the music-text conditioned dance generation received high ratings from participants, highlighting the importance of considering audience preferences in evaluating dance quality. + +# 4.3. Evaluation on Music Conditioned Dance Generation + +To validate the effectiveness of our architecture, we quantitatively compare our method with state-of-the-art those for music conditioned 3D dance generation. The results on the test set of AIST++ are shown in Table 1. We can find that our method outperforms the previous ones in terms of motion diversity $(\mathrm{Div}_k$ and $\mathrm{Div}_g)$ , while the performance of the motion quality $(\mathrm{FID}_g)$ and beat align score is inferior under the condition that we do not use seed motion from ground-truth compared with FACT [29] and Bailando [44]. + +The existing evaluation metrics are not sufficient to reflect the quality of generated dance in practice, which motivates us to propose a PFF and $\mathrm{AUC}_f$ to evaluate the percentage of freezing frames. It is worth noting that our architecture outperforms FACT [29] and Bailando [44] in terms of PFF and $\mathrm{AUC}_f$ , which shows that our method rarely generates frozen frames. Meanwhile, Li et al. [28] and DanceNet [57] also gain nearly zero freezing in PFF and $\mathrm{AUC}_f$ . The reason is that the generated dances of Li et al. [28] are highly jittery making its velocity variation extremely high, which is also reported in [29, 44], and leads to the non-freezing issue. And for DanceNet [57], it generates dance in a repeat motion pattern, nearly zero freezing but with low diversity. Furthermore, the results of user study show that the generated 3D dance is more visually re + +
DataMethodMotion QualityMotion DiversityFreezingBeat Align Score ↑User Study
FIDk ↓FIDg ↑Divk ↑Divg ↑PFF↓AUCf ↓Our Method Wins
AIST ++Ground-truth17.1010.608.197.450.000.000.237441.9%
Li et al. [28]86.4343.466.85*3.320.000.000.160798.3%
DanceNet [57]69.1825.492.862.850.000.980.143087.2%
DanceRevolution [21]73.4225.923.524.8711.0112.220.195070.5%
FACT [29]35.3522.115.946.1825.2921.590.220983.3%
Bailando [44]28.169.627.836.3414.9113.250.233265.0%
ours (only dance data)23.949.537.694.530.000.000.2127-
ours19.0120.099.456.360.000.000.2049-
Wild AudioFACT [29]70.3620.237.336.3432.6427.210.221189.6%
Bailando [44]50.5622.553.806.0417.5514.140.216670.8%
ours (only dance data)43.8513.088.524.760.780.760.1998-
ours27.6520.347.885.270.120.110.2290-
+ +Table 1: Music Conditioned Dance Generation: quantitative results on AIST++ and Wild Audio test set. The best and runner-up values are bold and underlined, respectively. Among compared methods, "Li et al.", DanceNet and FACT are multiplexing the same results of AIST++ benchmark [28], while DanceRevolution [21] is followed from Bailando [44]. $\dagger$ FID $_k$ and DIV $_k$ are fetched from [29] while FID $_g$ and DIV $_g$ are fetched from [44]. + +
DataMethodMotion Prediction DistanceUser Study
ft=10ft=20ft=30Our Method Wins
AIST ++GT0.0480.0720.08870.0%
a2d0.0520.0840.10866.6%
slerp [43]0.0880.1220.13573.3%
ours0.0490.0800.102-
Wild Audioa2d0.0520.0880.11360.0%
slerp [43]0.1040.1480.17170.0%
ours0.0480.0790.107-
+ +Table 2: Music-text conditioned Dance Generation: quantitative results on AIST++ and wild audio test set. The best and runner-up values are bold and underlined, respectively. + +alistic compared to other methods. Even in comparison to the ground truth, $41.9\%$ of our generated dance is voted as the better in average. We also report the result trained with dance-only data, which shows comparable performance. + +In addition to the evaluation on the test set of AIST++, we also show the experimental results on our in-the-wild dataset in Table 1. Our method outperforms FACT [29] and Bailando [44] in almost all metrics except for $\mathrm{FID}_g\mathrm{Div}_g$ . We can observe that both FACT [29] and Bailando [44] shows a large performance drop in terms of $\mathrm{FID}_k$ , while ours maintain a small change. This is because FACT [29] and Bailando [44] requires seed motion, however, there is no ground-truth for in-the-wild scenario. With random sampled seed motion or token, their methods are not adapted + +well for the in-the-wild scenarios. One can also notice that there exists freezing in our method on the in-the-wild dataset but the percentage of frozen frames is zero on the test set of AIST++, which results from the different distributions of AIST++ and our dataset. It also shows that our dataset is more challenging. We also report the result trained with dance-only data, which shows the advantage of our mix-training strategy in the wild scenario. + +# 4.4. Evaluation on Text Conditioned Motion Generation. + +We evaluate our text2motion approach in two different scenarios: inference with text only (t2m), and inference with text and music feature fusion. Table 3 shows the results. In the text-only setting, our approach achieves comparable performance with TM2T baseline(T) [17], which demonstrates that the mixed data/tasks training does not affect the quality of text2motion generation. We then apply our late fusion method by randomly sampling a music clip (from the whole AIST++ test set) and a text (from the whole t2m test set), and evaluate the generated dance clips following the same protocol as TM2T [17]. As Table 3 indicates, the text-dance consistency remains acceptable with a late fusion ratio of 0.8. For more details on the relation between the late fusion rate and text2motion result, please refer to the supplementary material. + +# 4.5. Qualitative Results + +We also show the qualitative results of our method for music-text conditioned 3D dance generation in Figure 4, Figure 5 and Figure 6. Specifically, Figure 4 shows the gen + +
MethodsR Precision↑FID↓MM Dist↓Diversity→MModality↑
Real motions0.5110.0022.9749.503
Seq2Seq [31]0.18011.755.5296.223
Language2Pose [4]0.24611.025.2967.676
Text2Gesture [7]0.1655.0126.0306.409
Hier [13]0.3016.5325.0128.332
MoCoGAN [47]0.03794.419.6430.462
Dance2Music [25]0.03366.988.1160.725
TM2T baseline(T) [17]0.3511.6694.0469.632
TM2T [17]0.4241.5013.4678.589
TM2D (t2m)0.3191.0214.0989.513
TM2D (LFR 0.8)0.3001.1054.3078.887
+ +Table 3: The evaluation of text2motion on the HumanML3D dataset, where the results are averaged from 20 runs. + +erated 3D dance with the same music sequence but with different text instructions. Figure 5 shows the generated 3D dance with the same text instruction but with a different start time. Figure 6 shows the generated 3D dance with the same text instruction but with different durations. As one can see, our approach maintains plausible visual results according to the text instructions for all three cases, which confirms that our approach is more flexible. + +# 4.6. The Impact of Mixed Data + +We employ two datasets (AIST++ and HumanML3D) to train our 3D human motion VQ-VAE so that the motion tokens drop in the same latent space. To empirically show this point, we conduct experiments to count the number of shared motion tokens. For a trained 3D human motion VQ-VAE with AIST++ and HumanML3D, there are 855 vectors and 912 vectors from codebook used to construct the motion token in AIST++ and HumanML3D. The total number of vectors contained in the codebook is 1024. among them, 846 vectors (98.9% in AIST++ and 92.8% in HumanML3D) are shared to generate the motion tokens, which shows the feasibility of using the separate music-conditioned and text-conditioned motion datasets for music-text conditioned 3D dance generation task. We further verify the latent space of the motions by VQ-VAE, and we perform t-SNE visualizations of the raw motion distribution (before VQ-VAE) and feature distribution encoded by VQ-VAE (after VQ-VAE) from a2d motion and t2m motion in Figure 7. Two motion datasets are mixed successfully, which provides the potential integration of bimodality dance generation. Refer to the supplementary materials for more details. + +# 4.7. Model Efficiency + +We also show the efficiency of our architecture in Table 4. In the inference stage, the complete model parameters and the inference time for 1s music are compared. Our architecture requires only about half model parameters and + +![](images/2da668e43b51e5fe4e3174a4b2d1539cb367aa1c010297da945da99eba7aa647.jpg) +Figure 4: Same music, different actions. These two generated dance share the same music and effect range (from 6s to 8s), with different text instructions: "A person does a cartwheel" (top), "A person is spinning with arms spread out" (bottom). + +![](images/d6ed11b5d684d9dc22721a26d539a64cbab1593e3d2d49602d0d348a64e9d5d7.jpg) +Figure 5: Same action, different start time points. These two generated dance share the same music and text instruction ("A person jumps up and down") but different effect ranges: from 6 to 8s (top), from 7 to 9s (bottom). + +![](images/9350bd249e2b5343d504bd69a319079e85b30d72d69ca01a62fb482cabbe418e.jpg) +Figure 6: Same action, different durations. These two generated dance share the same music and text instruction ("A person is keeping jumping") and action start time point (i.e., 6s) but different effect duration: 2s (top), 3s (bottom, i.e., jumped twice). + +inference time compared to Bailando [44], which attributes + +![](images/01431764affd7d7ed49c8fcc7d29d4b6b9db99f838dee551d499ee0e5e534132.jpg) +(a) Before VQ-VAE + +![](images/bc4989c4236db7c2dc909ea9980d7c3906bb0ab3b3b4fe0f22db1813e69da301.jpg) +(b) After VQ-VAE +Figure 7: The t-SNE visualizations of motions before and after VQ-VAE (orange: motions in AIST++, blue: motions in HumanML3D). + +
MethodParametersInference time
FACT [29]120M8.300s/(1s music)
Bailando [44]152M0.236s/(1s music)
ours72M0.143s/(1s music)
+ +Table 4: Comparisons of parameters and inference time. + +that we do not need to encode separate features for the upper and lower half body. Compared to FACT [29], our architectures significantly reduce the inference time. The main reason is that FACT employs two transformers for the music, dance motion, and a cross-modal transformer, but we formulate it as the standard sequence-to-sequence translation task with one transformer only, which establishes the advantage of our architecture. + +# 5. Conclusion + +In this paper, a novel task that simultaneously integrates both music and text instruction for 3D dance generation is proposed. Due to the lack of the paired 3D dance motions conditioned on both music and text, we resort to two existing datasets, i.e., music2dance and text2motion, to perform this task and employ a 3D human motion VQ-VAE to project the motions of the two datasets into a shared latent space so that the two datasets with different distributions can be effectively mixed for training. Moreover, we also propose a cross-modal transformer architecture to generate 3D dance without degrading the performance of music-conditioned dance generation. Two new evaluation metrics, MPD and FS, are proposed to reflect the quality of generated motion. Extensive experiments show that our method can generate dance motion that matches both music and text in a realistic and coherent way while maintaining comparable performance on two single modalities. + +Acknowledgement. This project is supported by the Ministry of Education Singapore under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006), + +and Singapore National Research Foundation ("CogniVision - Energy-autonomous always-on cognitive and attentive cameras for distributed real-time vision with milliwatt power consumption" grant NRF-CRP20-2017-0003) - www.green-ic.org/CogniVision. + +# References + +[1] mixamo. https://www MIXamo.com/. Accessed: 2023-03-07. 1 +[2] Hyemin Ahn, Timothy Ha, Yunho Choi, Hwiyeon Yoo, and Songhwai Oh. Text2action: Generative adversarial synthesis from language to action. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5915-5920. IEEE, 2018. 3 +[3] Hyemin Ahn, Jaehun Kim, Kihyun Kim, and Songhwai Oh. Generative autoregressive networks for 3d dancing move synthesis from music. IEEE Robotics and Automation Letters, 5(2):3501-3508, 2020. 2 +[4] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 3, 6, 8 +[5] Omid Alemi, Jules Françoise, and Philippe Pasquier. Groovenet: Real-time music-driven dance movement generation using artificial neural networks. In Workshop on Machine Learning for Creativity, 23rd ACM SIGKDD Conference on Knowledge Discovery and Data Mining, volume 8, page 26, 2017. 2 +[6] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. Teach: Temporal action compositions for 3d humans. In International Conference on 3D Vision (3DV), 2022. 3 +[7] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In IEEE Virtual Reality and 3D User Interfaces (VR), pages 1-10. IEEE, 2021. 6, 8 +[8] Hanting Chen, Yunhe Wang, Jianyuan Guo, and Dacheng Tao. VanillaNet: the Power of Minimalism in Deep Learning. arXiv preprint arXiv:2305.12972, 2023. 4 +[9] Kang Chen, Zhipeng Tan, Jin Lei, Song-Hai Zhang, Yuan-Chen Guo, Weidong Zhang, and Shi-Min Hu. Choreomaster: choreography-oriented music-driven dance synthesis. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 1, 2 +[10] Rukun Fan, Songhua Xu, and Weidong Geng. Example-based automatic music-driven conventional dance motion synthesis. IEEE transactions on visualization and computer graphics, 18(3):501-515, 2011. 2 +[11] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. arXiv preprint arXiv:2305.10924, 2023. 4 +[12] Joao P Ferreira, Thiago M Coutinho, Thiago L Gomes, José F Neto, and et al. Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics, 94:11-21, 2021. 2 + +[13] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1396–1406, 2021. 3, 6, 8 +[14] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 2 +[15] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 1 +[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 3, 5 +[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597, 2022. 3, 6, 7, 8 +[18] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 2 +[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in neural information processing systems, volume 30, 2017. 6 +[20] Duc N.M. Hoang, Shiwei Liu, Radu Marculescu, and Zhangyang Wang. Revisiting pruning at initialization through the lens of ramanujan graph. In International Conference on Learning Representations, 2023. 4 +[21] Ruozhi Huang, Huang Hu, Wei Wu, Kei Sawada, Mi Zhang, and Daxin Jiang. Dance revolution: Long-term dance generation with music via curriculum learning. In International Conference on Learning Representations, 2021. 6, 7 +[22] Yanghua Jin, Jiakai Zhang, Minjun Li, Yingtao Tian, Huachun Zhu, and Zhihao Fang. Towards the automatic anime characters creation with generative adversarial networks. arXiv preprint arXiv:1708.05509, 2017. 4 +[23] Hsuan-Kai Kao and Li Su. Temporally guided music-to-body-movement generation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 147-155, 2020. 2 +[24] Kosmas Kritsis, Aggelos Gkiokas, Aggelos Pikrakis, and Vassilis Katsouros. Danceconv: Dance motion generation with convolutional networks. IEEE Access, 10:44982-45000, 2022. 2 +[25] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. Advances in neural information processing systems, 32, 2019. 1, 2, 6, 8 + +[26] Juheon Lee, Seohyun Kim, and Kyogu Lee. Listen to dance: Music-driven choreography generation using autoregressive encoder-decoder network. arXiv preprint arXiv:1811.00818, 2018. 2 +[27] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 2 +[28] Jiaman Li, Yihang Yin, Hang Chu, Yi Zhou, Tingwu Wang, Sanja Fidler, and Hao Li. Learning to generate diverse dance motions with transformer. arXiv preprint arXiv:2008.08171, 2020. 2, 6, 7 +[29] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 1, 2, 5, 6, 7, 9 +[30] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 4 +[31] Angela S. Lin, Lemeng Wu, Rodolfo Corona, Kevin W. H. Tai, Qixing Huang, and Raymond J. Mooney. Generating animated videos of human activities from natural language descriptions. 2018. 3, 6, 8 +[32] Xinyin Ma, Gongfan Fang, and Xinchao Wang. LLM-Pruner: On the Structural Pruning of Large Language Models. arXiv preprint arXiv:2305.11627, 2023. 4 +[33] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 4 +[34] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2, 3 +[35] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, page 480-497, 2022. 3 +[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236-252, 2016. 2 +[37] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems, 109:13-26, 2018. 3 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1 + +[39] Sucheng Ren, Xingyi Yang, Songhua Liu, and Xinchao Wang. SG-Former: Self-guided Transformer with Evolving Token Reallocation. In IEEE International Conference on Computer Vision, 2023. 4 +[40] Sucheng Ren, Daquan Zhou, Shengfeng He, Jiashi Feng, and Xinchao Wang. Shunted Self-Attention via Multi-Scale Token Aggregation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 4 +[41] Xuanchi Ren, Haoran Li, Zijian Huang, and Qifeng Chen. Self-supervised dance video synthesis conditioned on music. In Proceedings of the 28th ACM International Conference on Multimedia, pages 46-54, 2020. 2 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1 +[43] Ken Shoemake. Animating rotation with quaternion curves. In Proceedings of the 12th annual conference on Computer graphics and interactive techniques, pages 245-254, 1985. 5, 6, 7 +[44] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11050-11059, 2022. 1, 2, 4, 6, 7, 8, 9 +[45] Guofei Sun, Yongkang Wong, Zhiyong Cheng, Mohan S Kankanhalli, Weidong Geng, and Xiangdong Li. Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Transactions on Multimedia, 23:497-509, 2020. 2 +[46] Taoran Tang, Jia Jia, and Hanyang Mao. Dance with melody: An LSTM-autoencoder approach to music-oriented dance synthesis. In Proceedings of the 26th ACM international conference on Multimedia, pages 1598-1606, 2018. 2 +[47] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 6, 8 +[48] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4 +[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4 +[50] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. 1 +[51] Nelson Yalta, Shinji Watanabe, Kazuhiro Nakadai, and Tetsuya Ogata. Weakly-supervised deep recurrent neural networks for basic dance step generation. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2019. 2 + +[52] Tatsuro Yamada, Hiroyuki Matsunaga, and Tetsuya Ogata. Paired recurrent autoencoders for bidirectional translation between robot actions and linguistic descriptions. IEEE Robotics and Automation Letters, 3(4):3441-3448, 2018. 3 +[53] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Advances in neural information processing systems, 2022. 4 +[54] Zijie Ye, Haozhe Wu, Jia Jia, Yaohua Bu, Wei Chen, Fanbo Meng, and Yanfeng Wang. Choreonet: Towards music to dance synthesis with choreographic action unit. In Proceedings of the 28th ACM International Conference on Multimedia, pages 744-752, 2020. 2 +[55] Weihao Yu, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. Metaformer Baselines for Vision. arXiv preprint arXiv:2210.13452, 2022.4 +[56] Ye Yuan and Kris Kitani. Dlow: Diversifying latent flows for diverse human motion prediction. In European Conference on Computer Vision, pages 346-364. Springer, 2020. 6 +[57] Wenlin Zhuang, Congyi Wang, Jinxiang Chai, Yangang Wang, Ming Shao, and Siyu Xia. Music2dance: Dancenet for music-driven dance generation. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 18(2):1-21, 2022. 2, 6, 7 +[58] Wenlin Zhuang, Yangang Wang, Joseph Robinson, Congyi Wang, Ming Shao, Yun Fu, and Siyu Xia. Towards 3d dance motion synthesis and control. arXiv preprint arXiv:2006.05743, 2020. 2 \ No newline at end of file diff --git a/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/images.zip b/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..36200819d3b003e51fbc575ddc4c289f52ede23e --- /dev/null +++ b/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ba38795b740a6901e8d4eaad4f907790e6b870c0f1254c84fd246b4538dc6d0 +size 552754 diff --git a/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/layout.json b/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fd51def5dfc4b2143fbbf400c213f8d54908fcf8 --- /dev/null +++ b/2023/TM2D_ Bimodality Driven 3D Dance Generation via Music-Text Integration/layout.json @@ -0,0 +1,8546 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 531, + 121 + ], + "type": "text", + "content": "TM2D: Bimodality Driven 3D Dance Generation via Music-Text Integration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 142, + 177, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 142, + 177, + 174 + ], + "spans": [ + { + "bbox": [ + 96, + 142, + 177, + 174 + ], + "type": "text", + "content": "Kehong Gong1* \nZihang Jiang1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 142, + 282, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 142, + 282, + 171 + ], + "spans": [ + { + "bbox": [ + 209, + 142, + 282, + 171 + ], + "type": "text", + "content": "Dongze Lian" + }, + { + "bbox": [ + 209, + 142, + 282, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 209, + 142, + 282, + 171 + ], + "type": "text", + "content": " \nXinxin Zuo" + }, + { + "bbox": [ + 209, + 142, + 282, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 172, + 291, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 172, + 291, + 186 + ], + "spans": [ + { + "bbox": [ + 149, + 172, + 291, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 172, + 291, + 186 + ], + "type": "text", + "content": " National University of Singapore" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 142, + 391, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 391, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 391, + 171 + ], + "type": "text", + "content": "Heng Chang" + }, + { + "bbox": [ + 313, + 142, + 391, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 313, + 142, + 391, + 171 + ], + "type": "text", + "content": " \nMichael Bi Mi" + }, + { + "bbox": [ + 313, + 142, + 391, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 415, + 143, + 497, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 143, + 497, + 172 + ], + "spans": [ + { + "bbox": [ + 415, + 143, + 497, + 172 + ], + "type": "text", + "content": "Chuan Guo² \nXinchao Wang¹†" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 172, + 443, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 172, + 443, + 185 + ], + "spans": [ + { + "bbox": [ + 310, + 172, + 443, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 310, + 172, + 443, + 185 + ], + "type": "text", + "content": " Huawei Technologies Co., Ltd." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 47, + 224, + 545, + 334 + ], + "blocks": [ + { + "bbox": [ + 47, + 224, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 47, + 224, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 545, + 334 + ], + "type": "image", + "image_path": "1f6ebcdb8087e5d1836998117936b0bbe0d692f6ead0982b3e17b1d70968080c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 343, + 546, + 380 + ], + "lines": [ + { + "bbox": [ + 46, + 343, + 546, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 546, + 380 + ], + "type": "text", + "content": "Figure 1: Generated 3D dance examples conditioned music and text with our method. Given an audio input and text input with a specific starting time and duration, our method is able to generate a sequence of dance motions that fit the music and text instruction. The character is from Mixamo [1]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 390, + 192, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 390, + 192, + 403 + ], + "spans": [ + { + "bbox": [ + 143, + 390, + 192, + 403 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 417, + 290, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 417, + 290, + 681 + ], + "spans": [ + { + "bbox": [ + 45, + 417, + 290, + 681 + ], + "type": "text", + "content": "We propose a novel task for generating 3D dance movements that simultaneously incorporate both text and music modalities. Unlike existing works that generate dance movements using a single modality such as music, our goal is to produce richer dance movements guided by the instructive information provided by the text. However, the lack of paired motion data with both music and text modalities limits the ability to generate dance movements that integrate both. To alleviate this challenge, we propose to utilize a 3D human motion VQ-VAE to project the motions of the two datasets into a latent space consisting of quantized vectors, which effectively mix the motion tokens from the two datasets with different distributions for training. Additionally, we propose a cross-modal transformer to integrate text instructions into motion generation architecture for generating 3D dance movements without degrading the performance of music-conditioned dance generation. To better evaluate the quality of the generated motion, we introduce two novel metrics, namely Motion Prediction Distance (MPD) and Freezing Score (FS), to measure the coherence and freezing percentage of the generated motion. Extensive experiments show that our approach can gener" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 391, + 546, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 391, + 546, + 439 + ], + "spans": [ + { + "bbox": [ + 306, + 391, + 546, + 439 + ], + "type": "text", + "content": "ate realistic and coherent dance movements conditioned on both text and music while maintaining comparable performance with the two single modalities. Code is available at https://garfield-kh.github.io/TM2D/." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 464, + 387, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 387, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 387, + 477 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 485, + 546, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 546, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 546, + 688 + ], + "type": "text", + "content": "The music-conditioned dance generation has become a topic of great interest in recent years. The ability to generate dance movements that are synchronized with music has numerous applications, such as behavior understanding, simulation, and benefiting the community of dancers and musicians [25, 9, 29, 44]. Although music has been used as a guidance to generate dance movements, another important modality cue, text (or language), which provides richer actions and more flexible motion guidance, and is studied in other tasks such as image classification [38], detection [15], segmentation [50], and text-driven image generation [42], has not been fully explored in dance generation. To this end, we first propose a novel task for generating 3D dance movements that simultaneously incorporate both text and music modalities, enabling the generated human to perform rich dancing movements in accordance with the music and text." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Designing a system pipeline for this bimodality driven 3D dance generation task is non-trivial. There exist two" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 694, + 282, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 694, + 282, + 703 + ], + "spans": [ + { + "bbox": [ + 56, + 694, + 282, + 703 + ], + "type": "text", + "content": "*Equal contribution: gongkehong@u.nus.edu, dongze@nus.edu.sg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 56, + 704, + 203, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 704, + 203, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 704, + 203, + 712 + ], + "type": "text", + "content": "†Corresponding author: xinchao@nus.edu.sg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9942" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "significant challenges to be considered: i) the existing datasets only cater to either music-driven (music2dance) [46, 5, 57, 29] or text-driven (text2motion) [36, 16] human motion generation, and no paired 3D dance generation dataset exists that takes into account both music and text. While building a new large-scale paired 3D dance dataset based on music and text is possible, it is time-consuming with fully annotated 3D human motion [29]; ii) the integration of text into music-conditioned dance generation requires a suitable architecture. However, existing methods that use music as a driving force to generate dance movements might result in temporal-freezing frames or fail to generalize to in-the-wild scenarios [29, 44]. Therefore, simple integration of text into the existing music-conditioned architecture might pose a risk of degraded dance generation quality in our new task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 267, + 289, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 289, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 289, + 543 + ], + "type": "text", + "content": "To address the first challenge, we take advantage of existing music-dance and text-motion datasets for this new task. However, directly mixing motions from these two datasets would result in inferior performance since the motions from these two datasets are in completely different motion spaces. To overcome this, we propose to utilize a VQ-VAE architecture to project the motions into a consistent and shared latent space. In particular, we build up a shared codebook for all the motions from the training set, and motions from both datasets are now represented as discrete tokens that are implicitly constrained to fall into a shared latent space. For the second challenge, we propose to utilize a cross-modal transformer architecture that formulates both music2dance and text2motion as sequence-to-sequence translation tasks. This architecture directly translates audio and text features into motion tokens and enables bimodality driven ability by introducing a fusion strategy in the latent space with a shared motion decoder for both tasks. With the shared decoder, audio and text information can be efficiently fused during inference. Our entire cross-modal transformer architecture is both effective and efficient, allowing for the integration of text instructions to generate coherent 3D dance motions, as illustrated in Figure 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "To better evaluate the coherence of generated dance in our task, we propose a new evaluation metric, Motion Prediction Distance (MPD), which measures the distance between the predicted motion and the ground truth at the time of integrating text, thereby providing a more accurate evaluation of the coherence of frames. Additionally, we introduce a Freezing Score (FS) that quantifies the percentage of temporal freezing frames in dance generation, which is a common problem in music-conditioned dance generation. To better evaluate the performance of our method in real-world scenarios, we also collect some in-the-wild music data for evaluation. Our method successfully performs dance generation based on both text and music while maintaining comparable performance on the single modal-" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": "ity tasks (music2dance, text2motion) compared to other state-of-the-art methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 97, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 547, + 289 + ], + "type": "text", + "content": "In summary, our contributions are as follows: i) We propose an interesting task of utilizing both music and text for 3D dance generation and propose a pipeline named TM2D (Text-Music to Dance) for this task. ii) Rather than collecting a new training set, we effectively combine the existing music2dance and text2motion datasets and employ a VQ-VAE framework to encode motions from all training sets to a shared feature space. iii) We propose a cross-modal transformer as well as a bimodal feature fusion strategy to encode both audio and text features, which is both effective and efficient. iv) We propose two new metrics, MPD and FS, which efficiently reflect the quality of generated motion. v) We successfully generate realistic and coherent dance based on both music and text instructions while maintaining comparable performance on the single modality tasks (music2dance, text2motion)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 300, + 392, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 392, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 392, + 312 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 319, + 402, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 402, + 332 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 402, + 332 + ], + "type": "text", + "content": "2.1. Music to Dance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 338, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 602 + ], + "type": "text", + "content": "Music2Dance is typically divided into 2D and 3D dance generation and has been explored for many years. Recent methods model 3D dance generation from the perspective of network architecture. For instance, Lee et al. [26] explore the convolutional neural network in casual dilated setting. Lee et al. [25] propose a dance unit with Variational Auto-Encoder (VAE). Ren et al. [41] employ recurrent neural network (RNN) and the graph convolution network is introduced in [12]. As for the 3D dance generation, [54, 3, 24] implement convolutional neural network, [45, 14] apply adversarial learning in generated dance. [5, 46, 51, 58, 23] implement Long short-term memory (LSTM), and [9] implement motion graph with learned music/dance embedding through matching approach [10]. More recently, transformer has been applied in dance generation [28, 29, 44, 27]. However, the previous methods usually produce temporal freezing frames when generating the long sequences, or are difficult to generalize to in-the-wild music, thereby not satisfiable when directly used for bimodal driven 3D dance generation. In this paper, we design a cross modal transformer which is effective and efficient to integrate text instruction to generate the 3D dance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 611, + 399, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 399, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 399, + 623 + ], + "type": "text", + "content": "2.2. Text to Motion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "In addition to music-driven motion generation, text is also utilized as instructions to generate motions. Text2motion can be categorized into action label based motion generation and language description based motion generation. The action label based methods [18, 34] generate motion conditioned on action labels. Action2Motion [18] implement a GRU-VAE to iteratively generate the next" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9943" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 86, + 193, + 308 + ], + "blocks": [ + { + "bbox": [ + 65, + 86, + 193, + 308 + ], + "lines": [ + { + "bbox": [ + 65, + 86, + 193, + 308 + ], + "spans": [ + { + "bbox": [ + 65, + 86, + 193, + 308 + ], + "type": "image", + "image_path": "b2acb0fe0656441ef4b950020dc31cabf4e42c64014fc6fe9df4c6a086485f3a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 328, + 548, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 328, + 548, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 548, + 413 + ], + "type": "text", + "content": "Figure 2: Our proposed pipeline for music-text conditioned 3D dance generation. Three stages from left to right: 3D human motion VQ-VAE, training stage of the cross-modal transformer, and inference stage of our pipeline. In the first stage, a VQ-VAE is trained with both motions from music2dance and text2motion data, which is then used to tokenize all motions. In the second stage, a dual path cross-modal transformer is employed for sequence2sequence translation tasks (i.e., audio to motion tokens, text to motion tokens), with a shared motion decoder. In the third stage, given audio and text inputs, the audio and text encoders first extract the corresponding features, which are then fused (late fusion) in the motion decoder to generate dance condition on both music and text." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 198, + 89, + 372, + 307 + ], + "blocks": [ + { + "bbox": [ + 198, + 89, + 372, + 307 + ], + "lines": [ + { + "bbox": [ + 198, + 89, + 372, + 307 + ], + "spans": [ + { + "bbox": [ + 198, + 89, + 372, + 307 + ], + "type": "image", + "image_path": "6a7a7e5cf65d81ecf02ec94346b5c0a3480756f34855069e0d6b1d76ef27083a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 374, + 92, + 533, + 298 + ], + "blocks": [ + { + "bbox": [ + 374, + 92, + 533, + 298 + ], + "lines": [ + { + "bbox": [ + 374, + 92, + 533, + 298 + ], + "spans": [ + { + "bbox": [ + 374, + 92, + 533, + 298 + ], + "type": "image", + "image_path": "ef5b69fe33d5d813f5825ea47be9477e8023b6ae80745d50f61adfaf6c59bc33.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 433, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 289, + 639 + ], + "type": "text", + "content": "frame based on the action label and previous frames. ACTOR [34] implements a transformer-VAE to encode and decode the whole pose sequence in one-shot. Since the action label based methods are restricted in a small set of action labels, language description based methods are proposed for more flexible motion generation. [37, 31, 2, 16, 17] formulate the text-to-motion task as a machine translation problem, others [52, 4, 13, 35] learns a joint embedding space of text and motion. Among them, [37, 31, 2, 52, 4, 13] use RNN encoder-decoder to learn the mapping between text and motion. [17, 16] use transformer encoder with RNN decoder, and TM2T [17] further introduces motion to text as an inverse alignment for text to motion. TEMOS [35] and TEACH [6] use transformer for both encoder and decoder for one-shot generation. Different from these text2motion methods, we focus on how to integrate two modalities (i.e., music and text) together for 3D dance generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 650, + 104, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 104, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 104, + 662 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 670, + 144, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 144, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 144, + 684 + ], + "type": "text", + "content": "3.1. System Pipeline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "The overall pipeline is shown in Figure 2, which consists of three stages. The first stage employs a 3D human motion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 433, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 547, + 590 + ], + "type": "text", + "content": "VQ-VAE to encode both the motions of the music2dance and text2motion datasets to a shared codebook encoded by multiple vectors such that each dance motion can be represented as a discrete motion token. After that, the combinations of music, text, and motion token are fed into a cross-modal transformer for training, which effectively learn how to predict a sequential motion tokens according to the previous those as well as music and text. In the third stage, a random starting motion token is generated and inputted to the cross-modal transformer with the given music and text for the sequential motion token prediction, which will be decoded by the pre-trained 3D human motion VQ-VAE in the first stage to generate 3D dance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 598, + 463, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 463, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 463, + 612 + ], + "type": "text", + "content": "3.2. 3D Human Motion VQ-VAE" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Since there are no paired 3D dance motions conditioned on both music and text, we try to tackle our task with the existing music2dance and text2motion datasets. Instead of directly mixing the motions from both datasets for training, we employ a VQ-VAE to project the motions into a consistent and shared latent space, which is represented with a codebook, as illustrated in Figure 2. To be specific, given a 3D human motion " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{T \\times d_m}" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " is the time length" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9944" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "d_{m}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " is the dimension of the human motion, an encoder of VQ-VAE that consists of several 1-D convolutional layers projects " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " to a latent vector " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "z\\in \\mathbb{R}^{T^{\\prime}\\times d}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "T^{\\prime} = \\frac{T}{t}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " is the time interval for downsampling and " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " is the dimension of the latent vector. A learnable codebook " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "e\\in \\mathbb{R}^{K\\times d}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " describes all latent variable features of the whole dataset, where " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " are the length and dimension of the codebook, respectively. A quantized latent vector " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "z_{q}\\in \\mathbb{R}^{T^{\\prime}\\times d}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " will record the closest vector from codebook " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": " as follows" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 101, + 187, + 287, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 187, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 101, + 187, + 287, + 206 + ], + "type": "interline_equation", + "content": "z _ {q, i} = \\arg \\min _ {e _ {j} \\in e} \\| z _ {i} - e _ {j} \\| \\in \\mathbb {R} ^ {d}, \\tag {1}", + "image_path": "596900dfbb61a9663af2a6033ed391a56e4fec52020a3cf2d638492e9ef0bf92.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 216, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 216, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 216, + 287, + 239 + ], + "type": "text", + "content": "and the motion token " + }, + { + "bbox": [ + 47, + 216, + 287, + 239 + ], + "type": "inline_equation", + "content": "t_m \\in \\mathbb{R}^{T' \\times 1}" + }, + { + "bbox": [ + 47, + 216, + 287, + 239 + ], + "type": "text", + "content": " stores the index of the closest vector" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 248, + 287, + 267 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 248, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 100, + 248, + 287, + 267 + ], + "type": "interline_equation", + "content": "t _ {m, i} = \\arg \\min _ {j} \\| z _ {i} - e _ {j} \\| \\in \\mathbb {R} ^ {1}. \\tag {2}", + "image_path": "b8d2375b3a1a390abd3127f483eb2e107055ab4033214b53837b554d8719cbc0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 274, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 287, + 323 + ], + "type": "text", + "content": "The quantized " + }, + { + "bbox": [ + 47, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "z_{q}" + }, + { + "bbox": [ + 47, + 274, + 287, + 323 + ], + "type": "text", + "content": " will be decoded with stacked convolutional layers to reconstruct the human motion " + }, + { + "bbox": [ + 47, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 47, + 274, + 287, + 323 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 274, + 287, + 323 + ], + "type": "inline_equation", + "content": "t_{m}" + }, + { + "bbox": [ + 47, + 274, + 287, + 323 + ], + "type": "text", + "content": " will be used in the training stage of the cross-modal transformer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 323, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 287, + 384 + ], + "type": "text", + "content": "For the training of 3D human motion VQ-VAE, we follow the strategy in [48] and the total loss contains a reconstruction loss for human motion regression, a codebook loss for the dictionary learning, and a commitment loss to stabilize the training process:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 391, + 287, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 391, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 52, + 391, + 287, + 406 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v q} = \\| \\hat {M} - M \\| _ {1} + \\| \\operatorname {s g} [ e ] - e _ {q} \\| _ {2} ^ {2} + \\beta \\| e - \\operatorname {s g} [ e _ {q} ] \\| _ {2} ^ {2}, \\tag {3}", + "image_path": "a3ac4b7f0fea35b14e80dd8968c35b46c5f743c9138e9b2396fb762dea9e542b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}[\\cdot ]" + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "text", + "content": " is 'stop gradient' and " + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 413, + 287, + 460 + ], + "type": "text", + "content": " is the factor term to adjust the weight of the commitment loss. A straight-through estimator is also employed to pass the gradient from the decoder to the encoder in back-propagation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 460, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 287, + 533 + ], + "type": "text", + "content": "As we will show in Sec. 4, the motion tokens from both datasets encoded by 3D human motion VQ-VAE fall almost in the shared latent space, which shows the feasibility of using the separate music-conditioned and text-conditioned motion datasets for music-text conditioned 3D dance generation task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 540, + 192, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 192, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 192, + 551 + ], + "type": "text", + "content": "3.3. Cross-modal Transformer" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 558, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 287, + 664 + ], + "type": "text", + "content": "The cross-modal transformer contains an audio encoder, a text encoder, and a motion decoder. Since we use two separate datasets for our task, the cross-modal transformer is divided into two branches, where one for music2dance and the other for text2motion. It takes the audio feature, the text feature, and the motion token " + }, + { + "bbox": [ + 47, + 558, + 287, + 664 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 47, + 558, + 287, + 664 + ], + "type": "text", + "content": " encoded by 3D human motion VQ-VAE as inputs, and performs the sequence-to-sequence translation task to generate the future motion tokens." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "Attention. Attention is introduced in Transformer [49] for natural language processing, and then applied to many domains [30, 8, 20, 39, 11, 53, 32, 55, 40] A " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "-layer transformer typically consists of " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " transformer blocks, which" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "contains a multi-head self-attention (MSA) and a feed forward network (FFN). Given the input " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{N \\times c}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " is the sequence length and " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " is the dimension of the input, the transformer block first maps it to keys " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{N \\times c}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": ", queries " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{N \\times c}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": ", values " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{N \\times c}" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " with the linear projections, and then a self-attention operation is performed by" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 336, + 152, + 545, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 152, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 336, + 152, + 545, + 180 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {c}}\\right) V. \\tag {4}", + "image_path": "b74d95ca3a40f40259245ce15aa453614b7272a64cb8b370ba1a94067623d2a1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 183, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 545, + 304 + ], + "type": "text", + "content": "The output is followed by a normalization layer and a FFN. Audio encoder. Given a sequence of music, the audio encoder first extracts the raw audio features following Bailando [44], where a public audio processing toolbox, Librosa [22], is employed to obtain the mel frequency cepstral coefficients (MFCC), MFCC delta, constant- " + }, + { + "bbox": [ + 304, + 183, + 545, + 304 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 183, + 545, + 304 + ], + "type": "text", + "content": " chromagram, tempogram and onset strength. After that, an embedded layer followed by several transformer blocks containing self-attention operation e.g., Eq. (4), is used to generate the processed audio features " + }, + { + "bbox": [ + 304, + 183, + 545, + 304 + ], + "type": "inline_equation", + "content": "f_{a}\\in \\mathbb{R}^{T^{\\prime}\\times d}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 304, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 363 + ], + "type": "text", + "content": "Text encoder. Given a text instruction, we first extract the text token with a GloVe [33], and then an embedded layer followed by several transformer blocks is performed to obtain a processed text feature " + }, + { + "bbox": [ + 304, + 304, + 545, + 363 + ], + "type": "inline_equation", + "content": "f_{t} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 304, + 304, + 545, + 363 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 304, + 545, + 363 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 304, + 545, + 363 + ], + "type": "text", + "content": " is the length of the text feature." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "text", + "content": "Motion decoder. Motion decoder takes the motion tokens encoded by the 3D human motion VQ-VAE as inputs and outputs the future motion tokens, which has the same spirit as the sequence-to-sequence translation task. Since the information after time " + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "text", + "content": " is unknown at moment " + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "text", + "content": ", a masked MSA is first employed to interact with motions at different times. The mask has the shape of a simple upper triangular matrix and is performed in Eq. (4). For the moment " + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "text", + "content": ", only the motions before moment " + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "text", + "content": " are able to perform self-attention operations. Motion decoder also maps the extracted audio feature " + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "text", + "content": " and text feature " + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "text", + "content": " via audio encoder and text encoder to " + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 364, + 546, + 567 + ], + "type": "text", + "content": ", to perform cross-modal attention with motion tokens. To enable the music-text conditional dance generation, (i.e., feature fusion, we apply shared parameters), we use the motion decoder with shared parameters for model efficiency, as shown in Figure 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 575, + 460, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 460, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 460, + 587 + ], + "type": "text", + "content": "3.4. Details of Music-text Fusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": "The detailed architecture is listed in Figure 3. Given the audio feature and text feature, the motion decoder first processes the past motion tokens with a self-attention layer followed by addition, normalization, a cross-attention with audio and text features separately, addition, normalization, and a feed-forward layer are performed. Such a procedure is repeated " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "L = 6" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": ") times to build a typical " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": "-layer transformer. Then we adopt a late fusion strategy to perform a weighted sum of features of audio and text at a specific time (i.e., effect range). The weight curve is shown in Figure 3," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9945" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 159, + 240 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 159, + 240 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 159, + 240 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 159, + 240 + ], + "type": "image", + "image_path": "a152b169740aa65fb16ca2449b80664d5f90a4b443a91c4499c5ec47685bb5e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 288, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 288, + 331 + ], + "type": "text", + "content": "Figure 3: Details of music-text feature fusion: Given the audio and text features from the encoder, the decoder processes them separately in the early layers. Then the audiotext fusion layer is applied by the weighted sum at the effect range with weight cure. Finally, a linear projection layer and a softmax operation are applied to predict the music-text conditioned motion tokens." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 161, + 70, + 286, + 239 + ], + "blocks": [ + { + "bbox": [ + 161, + 70, + 286, + 239 + ], + "lines": [ + { + "bbox": [ + 161, + 70, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 161, + 70, + 286, + 239 + ], + "type": "image", + "image_path": "a17d2629f92343d8b02e3e7d407bed0428a76891763c6caf14fc73570a7f1997.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 353, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 473 + ], + "type": "text", + "content": "where we slightly increase the weight of the text feature by a half cosine curve until a peak value of 0.8 at the beginning (20% time of effect range), and decrease it by a half cosine curve at the end (20% time of effect range). The weight of audio feature " + }, + { + "bbox": [ + 46, + 353, + 287, + 473 + ], + "type": "inline_equation", + "content": "W_{audio}" + }, + { + "bbox": [ + 46, + 353, + 287, + 473 + ], + "type": "text", + "content": " is 1 - " + }, + { + "bbox": [ + 46, + 353, + 287, + 473 + ], + "type": "inline_equation", + "content": "W_{text}" + }, + { + "bbox": [ + 46, + 353, + 287, + 473 + ], + "type": "text", + "content": " to ensure the feature keeps the same scale. With the fused feature, a linear projection layer and a softmax operation are applied to predict the music-text conditioned motion tokens, which are then decoded by the decoder of human motion VQ-VAE to obtain the music-text conditioned 3D dance sequences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 178, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 178, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 178, + 491 + ], + "type": "text", + "content": "3.5. Training and Inference" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "text", + "content": "Training. To train the 3D human motion VQ-VAE, we crop 64 frames with a sliding window from the original motion sequences as inputs, i.e., " + }, + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "inline_equation", + "content": "T = 64" + }, + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "text", + "content": ", for both music2dance and text2motion datasets. We use a three-layer encoder and decoder so that the time interval of downsampling " + }, + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "text", + "content": " is 8. We randomly sample motions from both datasets and employ the Adam optimizer with a batch size of 128 and a learning rate of " + }, + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 46, + 498, + 287, + 594 + ], + "type": "text", + "content": " to optimizer our 3D human motion VQ-VAE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "To train the cross-modal transformer, we split two branches into two streams to train the music2dance and text2motion tasks on both datasets. For the music2dance task, we use the same frequency as the motion token to sample the music vector so that the music feature has the same time length as the motion token. We perform the sequence to sequence translation with the motion decoder to obtain the prediction of motion tokens conditioned by the music input. As for the text2motion task, since the text and motion token have different lengths, we use a maximum text length" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "of 84 for text2motion translation with a padding strategy. We employ 6 self-attention layers for the audio encoder, text encoder, and motion decoder with the hidden dimension of 512 and 8 heads. The cross entropy loss is adopted for both music2dance and text2motion tasks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 358, + 140, + 545, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 140, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 358, + 140, + 545, + 174 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = - \\frac {1}{m} \\sum_ {i = 0} ^ {m - 1} \\sum_ {j = 0} ^ {C - 1} y _ {i j} \\log y _ {\\hat {i j}}, \\tag {5}", + "image_path": "96c50d8af5463951733278b42ee4286e5f5cb2391db6b2b3b83942c823ac0f90.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 183, + 545, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 545, + 254 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 183, + 545, + 254 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 183, + 545, + 254 + ], + "type": "text", + "content": " is the prediction of motion token and " + }, + { + "bbox": [ + 304, + 183, + 545, + 254 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 183, + 545, + 254 + ], + "type": "text", + "content": " is the ground-truth. " + }, + { + "bbox": [ + 304, + 183, + 545, + 254 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 183, + 545, + 254 + ], + "type": "text", + "content": " is the length of motion tokens and " + }, + { + "bbox": [ + 304, + 183, + 545, + 254 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 183, + 545, + 254 + ], + "type": "text", + "content": " is the number of classes of motion tokens, i.e., the length of codebook of 3D human motion VQ-VAE. Both tasks are simultaneously optimized with a batch size of 64 and a learning rate of " + }, + { + "bbox": [ + 304, + 183, + 545, + 254 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 304, + 183, + 545, + 254 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 255, + 546, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 546, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 546, + 423 + ], + "type": "text", + "content": "Inference. Our aim is to generate 3D dance with music-text integration. Therefore, at the inference stage, we first fuse the music feature and text feature extracted by the audio encoder and text encoder with a weighted sum. Specifically, given an audio feature with a time length of " + }, + { + "bbox": [ + 304, + 255, + 546, + 423 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 304, + 255, + 546, + 423 + ], + "type": "text", + "content": " and a text feature with the length of " + }, + { + "bbox": [ + 304, + 255, + 546, + 423 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 255, + 546, + 423 + ], + "type": "text", + "content": ", we first feed them into the motion decoder for future motion token prediction. Then we adopt the late fusion strategy to have a weighted sum of the generated motion tokens from both audio and text features at the duration of the integrated text, followed by a linear projection layer and a softmax operation, to obtain the combined motion tokens. Finally, the combined motion tokens are decoded into a 3D dance sequence after going through the decoder of our 3D human motion VQ-VAE model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 435, + 388, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 388, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 388, + 449 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 455, + 432, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 432, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 432, + 468 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 640 + ], + "type": "text", + "content": "Datasets. Since there are no paired 3D dance motions conditioned on both music and text for our task, we combine the existing music2dance and text2motion datasets. For the music2dance dataset, we employ the AIST++ dataset [29]. For the text2motion dataset, we employ the HumanML3D [16] dataset. Additionally, we also collect a new in-the-wild music dataset to evaluate the generalization ability of our method, which contains 82 wild music clips from Youtube. The total of 53 minutes duration (8x larger size than the AIST++ test set) and various styles and content of music, which faithfully lies out of the distribution of AIST++ (explained in supplementary material). The evaluation on our dataset better reflects the generalization ability of the models in real-world scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Baselines. Since we are the first to propose this bimodality driven 3D dance generation task, there are no existing methods for comparisons. We implement a traditional motion editing algorithm, slerp [43], which aligns the last generated frame of the dance with the first frame of the motion by text with a transition window of 10 frames, and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9946" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "applies spherical interpolation of quaternion to fill in the transition in between. To validate the effectiveness of our method on music2dance task, we compare our method to the previous music2dance works [28, 57, 21, 29, 44]. To validate the effectiveness of our method on text2motion task, we compare our method to the previous text2motion works [31, 4, 7, 13, 47, 25, 17]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 287, + 239 + ], + "type": "text", + "content": "Evaluation metrics. We adopt the same evaluation settings as suggested by FACT [29] and Bailando [44] to evaluate the dance generation quality, including Fréchet Inception Distances (FID) [19], Diversity, and Beat Align Score. For the text2motion quality, we adopt the same evaluation settings as suggested by TM2T [17], including R-precision, Multimodal-Dist, FID, Diversity, and MultiModality." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 240, + 288, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 288, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 288, + 383 + ], + "type": "text", + "content": "In addition, we also propose two new evaluation metrics, Motion Prediction Distance (MPD) and Freezing Score. The former reflects the coherence of frames at the time of integrating text, and the latter reacts to the percentage of temporal freezing frames in the generated dance. For the Freezing Score, we introduce the Percentage of Freezing Frame (PFF) and " + }, + { + "bbox": [ + 46, + 240, + 288, + 383 + ], + "type": "inline_equation", + "content": "\\mathrm{AUC}_f" + }, + { + "bbox": [ + 46, + 240, + 288, + 383 + ], + "type": "text", + "content": ", where the PFF is defined by measuring the percentage of frozen frames with two criteria: 1) the maximum joint velocity blew a threshold " + }, + { + "bbox": [ + 46, + 240, + 288, + 383 + ], + "type": "inline_equation", + "content": "(0.015\\mathrm{m / s})" + }, + { + "bbox": [ + 46, + 240, + 288, + 383 + ], + "type": "text", + "content": ". 2) its duration exceeds a certain period (3s). The " + }, + { + "bbox": [ + 46, + 240, + 288, + 383 + ], + "type": "inline_equation", + "content": "\\mathrm{AUC}_f" + }, + { + "bbox": [ + 46, + 240, + 288, + 383 + ], + "type": "text", + "content": " is defined by area under curve of PFF in the threshold range from 0 to 0.03. As for MPD, it is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 391, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 391, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 83, + 391, + 287, + 407 + ], + "type": "interline_equation", + "content": "\\mathrm {M P D} = \\min _ {i} \\| f _ {i} \\left(M _ {t _ {0} \\rightarrow t _ {1}}\\right) - M _ {t _ {1} \\rightarrow t _ {2}} \\| _ {2}, \\tag {6}", + "image_path": "6b0529935115496d206a88c976e130c628a104eecf5e15aefa5911dcc095d12e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": " is the dance motions, " + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": " is a motion prediction model, " + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "inline_equation", + "content": "f_{i}(M_{t_{0}\\rightarrow t_{1}})" + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": " is the i-th predicted possible future motion, and " + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": " are the timestamps. It means that the model predicts various potential possible motion from " + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": " with the motion from " + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": ". If this distance is small (i.e., the generated dance " + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "inline_equation", + "content": "M_{t_1\\to t_2}" + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": " lays in the possible future), then the generated dance motion " + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "inline_equation", + "content": "M_{t_1\\to t_2}" + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": " is more coherent at the time of integrating text. We adopt the motion prediction model DLow [56] in this evaluation metric as it is designed for diverse potential furture motion prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Perceptual Evaluation. Besides the above metric measuring, we also conduct extensive user studies on Amazon Mechanical Turk (AMT) to perceptually evaluate the visual effects of our generated 3D dance results. Particularly, given each pair of dance movements sampled from our method and others with the same music clip, we request 3 distinct users on AMT to present their preference regarding the music-dance alignment, motion realism, and mobility. We further set the bar of involved users that only the ones with Master recognition who also have finished more than 1,000 tasks with over " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " approve rate are considered. Overall, there are 55 users employed in our user studies that come from various regions, ages, races and gender. The results of the user study are more representative to show the effect of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 435, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 435, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 435, + 84 + ], + "type": "text", + "content": "the generated dance in practice." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 92, + 545, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 92, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 306, + 92, + 545, + 117 + ], + "type": "text", + "content": "4.2. Evaluation on Music-text Conditioned Dance Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 123, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 545, + 387 + ], + "type": "text", + "content": "We validate the results of music-text conditioned dance generation on the test set of AIST++ and our dataset, as illustrated in Table 2. Here we mainly evaluate the coherence of generated dance with MPD metric at the time where music meets text, i.e., the time point where text starts to take effect. We measure the coherence of pure music2dance generation as a baseline named a2d to reflect the influence of with/without text instruction. We use the past 25 motion frames to predict the future 30 frames, and calculate the MPD from future frame (ft) = 10 to ft = 30, respectively. Our method consistently outperforms slerp [43] and a2d baseline in both datasets, and gains a similar result compared to the ground-truth shown in Table 2, which verifies the naturalness of our generated motion. More importantly, user study experiments show that our method generates more realistic 3D dance compared to dance generation only conditioned music, and " + }, + { + "bbox": [ + 304, + 123, + 545, + 387 + ], + "type": "inline_equation", + "content": "70.0\\%" + }, + { + "bbox": [ + 304, + 123, + 545, + 387 + ], + "type": "text", + "content": " Win rate even compared to the ground-truth shown in Table 2. The results of our user study indicate that the music-text conditioned dance generation received high ratings from participants, highlighting the importance of considering audience preferences in evaluating dance quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 395, + 545, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 395, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 395, + 545, + 419 + ], + "type": "text", + "content": "4.3. Evaluation on Music Conditioned Dance Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": "To validate the effectiveness of our architecture, we quantitatively compare our method with state-of-the-art those for music conditioned 3D dance generation. The results on the test set of AIST++ are shown in Table 1. We can find that our method outperforms the previous ones in terms of motion diversity " + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "inline_equation", + "content": "(\\mathrm{Div}_k" + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_g)" + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": ", while the performance of the motion quality " + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "inline_equation", + "content": "(\\mathrm{FID}_g)" + }, + { + "bbox": [ + 304, + 426, + 545, + 533 + ], + "type": "text", + "content": " and beat align score is inferior under the condition that we do not use seed motion from ground-truth compared with FACT [29] and Bailando [44]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "The existing evaluation metrics are not sufficient to reflect the quality of generated dance in practice, which motivates us to propose a PFF and " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AUC}_f" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " to evaluate the percentage of freezing frames. It is worth noting that our architecture outperforms FACT [29] and Bailando [44] in terms of PFF and " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AUC}_f" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ", which shows that our method rarely generates frozen frames. Meanwhile, Li et al. [28] and DanceNet [57] also gain nearly zero freezing in PFF and " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AUC}_f" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ". The reason is that the generated dances of Li et al. [28] are highly jittery making its velocity variation extremely high, which is also reported in [29, 44], and leads to the non-freezing issue. And for DanceNet [57], it generates dance in a repeat motion pattern, nearly zero freezing but with low diversity. Furthermore, the results of user study show that the generated 3D dance is more visually re" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9947" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 542, + 261 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 542, + 261 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 542, + 261 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 542, + 261 + ], + "type": "table", + "html": "
DataMethodMotion QualityMotion DiversityFreezingBeat Align Score ↑User Study
FIDk ↓FIDg ↑Divk ↑Divg ↑PFF↓AUCf ↓Our Method Wins
AIST ++Ground-truth17.1010.608.197.450.000.000.237441.9%
Li et al. [28]86.4343.466.85*3.320.000.000.160798.3%
DanceNet [57]69.1825.492.862.850.000.980.143087.2%
DanceRevolution [21]73.4225.923.524.8711.0112.220.195070.5%
FACT [29]35.3522.115.946.1825.2921.590.220983.3%
Bailando [44]28.169.627.836.3414.9113.250.233265.0%
ours (only dance data)23.949.537.694.530.000.000.2127-
ours19.0120.099.456.360.000.000.2049-
Wild AudioFACT [29]70.3620.237.336.3432.6427.210.221189.6%
Bailando [44]50.5622.553.806.0417.5514.140.216670.8%
ours (only dance data)43.8513.088.524.760.780.760.1998-
ours27.6520.347.885.270.120.110.2290-
", + "image_path": "7d6a3411797acf9123a7a4428ec5fadbcc84daac9027b02a6c0196557276d673.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 337, + 282, + 472 + ], + "blocks": [ + { + "bbox": [ + 46, + 269, + 546, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 269, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 546, + 319 + ], + "type": "text", + "content": "Table 1: Music Conditioned Dance Generation: quantitative results on AIST++ and Wild Audio test set. The best and runner-up values are bold and underlined, respectively. Among compared methods, \"Li et al.\", DanceNet and FACT are multiplexing the same results of AIST++ benchmark [28], while DanceRevolution [21] is followed from Bailando [44]. " + }, + { + "bbox": [ + 46, + 269, + 546, + 319 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 269, + 546, + 319 + ], + "type": "text", + "content": " FID" + }, + { + "bbox": [ + 46, + 269, + 546, + 319 + ], + "type": "inline_equation", + "content": "_k" + }, + { + "bbox": [ + 46, + 269, + 546, + 319 + ], + "type": "text", + "content": " and DIV" + }, + { + "bbox": [ + 46, + 269, + 546, + 319 + ], + "type": "inline_equation", + "content": "_k" + }, + { + "bbox": [ + 46, + 269, + 546, + 319 + ], + "type": "text", + "content": " are fetched from [29] while FID" + }, + { + "bbox": [ + 46, + 269, + 546, + 319 + ], + "type": "inline_equation", + "content": "_g" + }, + { + "bbox": [ + 46, + 269, + 546, + 319 + ], + "type": "text", + "content": " and DIV" + }, + { + "bbox": [ + 46, + 269, + 546, + 319 + ], + "type": "inline_equation", + "content": "_g" + }, + { + "bbox": [ + 46, + 269, + 546, + 319 + ], + "type": "text", + "content": " are fetched from [44]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 337, + 282, + 472 + ], + "lines": [ + { + "bbox": [ + 49, + 337, + 282, + 472 + ], + "spans": [ + { + "bbox": [ + 49, + 337, + 282, + 472 + ], + "type": "table", + "html": "
DataMethodMotion Prediction DistanceUser Study
ft=10ft=20ft=30Our Method Wins
AIST ++GT0.0480.0720.08870.0%
a2d0.0520.0840.10866.6%
slerp [43]0.0880.1220.13573.3%
ours0.0490.0800.102-
Wild Audioa2d0.0520.0880.11360.0%
slerp [43]0.1040.1480.17170.0%
ours0.0480.0790.107-
", + "image_path": "a66234f4fb7b95103ba555b9ff71aa93709a8e7798e72d2c425de91544be1db2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 481, + 287, + 518 + ], + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 518 + ], + "type": "text", + "content": "Table 2: Music-text conditioned Dance Generation: quantitative results on AIST++ and wild audio test set. The best and runner-up values are bold and underlined, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 544, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 592 + ], + "type": "text", + "content": "alistic compared to other methods. Even in comparison to the ground truth, " + }, + { + "bbox": [ + 46, + 544, + 287, + 592 + ], + "type": "inline_equation", + "content": "41.9\\%" + }, + { + "bbox": [ + 46, + 544, + 287, + 592 + ], + "type": "text", + "content": " of our generated dance is voted as the better in average. We also report the result trained with dance-only data, which shows comparable performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "In addition to the evaluation on the test set of AIST++, we also show the experimental results on our in-the-wild dataset in Table 1. Our method outperforms FACT [29] and Bailando [44] in almost all metrics except for " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_g\\mathrm{Div}_g" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ". We can observe that both FACT [29] and Bailando [44] shows a large performance drop in terms of " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_k" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", while ours maintain a small change. This is because FACT [29] and Bailando [44] requires seed motion, however, there is no ground-truth for in-the-wild scenario. With random sampled seed motion or token, their methods are not adapted" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 339, + 546, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 546, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 546, + 435 + ], + "type": "text", + "content": "well for the in-the-wild scenarios. One can also notice that there exists freezing in our method on the in-the-wild dataset but the percentage of frozen frames is zero on the test set of AIST++, which results from the different distributions of AIST++ and our dataset. It also shows that our dataset is more challenging. We also report the result trained with dance-only data, which shows the advantage of our mix-training strategy in the wild scenario." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 441, + 545, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 466 + ], + "type": "text", + "content": "4.4. Evaluation on Text Conditioned Motion Generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 472, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 653 + ], + "type": "text", + "content": "We evaluate our text2motion approach in two different scenarios: inference with text only (t2m), and inference with text and music feature fusion. Table 3 shows the results. In the text-only setting, our approach achieves comparable performance with TM2T baseline(T) [17], which demonstrates that the mixed data/tasks training does not affect the quality of text2motion generation. We then apply our late fusion method by randomly sampling a music clip (from the whole AIST++ test set) and a text (from the whole t2m test set), and evaluate the generated dance clips following the same protocol as TM2T [17]. As Table 3 indicates, the text-dance consistency remains acceptable with a late fusion ratio of 0.8. For more details on the relation between the late fusion rate and text2motion result, please refer to the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 659, + 418, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 418, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 418, + 672 + ], + "type": "text", + "content": "4.5. Qualitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 715 + ], + "type": "text", + "content": "We also show the qualitative results of our method for music-text conditioned 3D dance generation in Figure 4, Figure 5 and Figure 6. Specifically, Figure 4 shows the gen" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9948" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 279, + 189 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 279, + 189 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 279, + 189 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 279, + 189 + ], + "type": "table", + "html": "
MethodsR Precision↑FID↓MM Dist↓Diversity→MModality↑
Real motions0.5110.0022.9749.503
Seq2Seq [31]0.18011.755.5296.223
Language2Pose [4]0.24611.025.2967.676
Text2Gesture [7]0.1655.0126.0306.409
Hier [13]0.3016.5325.0128.332
MoCoGAN [47]0.03794.419.6430.462
Dance2Music [25]0.03366.988.1160.725
TM2T baseline(T) [17]0.3511.6694.0469.632
TM2T [17]0.4241.5013.4678.589
TM2D (t2m)0.3191.0214.0989.513
TM2D (LFR 0.8)0.3001.1054.3078.887
", + "image_path": "f64a580540a6691bd95518814da478e08ff1b6ea7d9dc1d35b6524a0ea33e490.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 198, + 287, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 232 + ], + "type": "text", + "content": "Table 3: The evaluation of text2motion on the HumanML3D dataset, where the results are averaged from 20 runs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 255, + 286, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 286, + 350 + ], + "type": "text", + "content": "erated 3D dance with the same music sequence but with different text instructions. Figure 5 shows the generated 3D dance with the same text instruction but with a different start time. Figure 6 shows the generated 3D dance with the same text instruction but with different durations. As one can see, our approach maintains plausible visual results according to the text instructions for all three cases, which confirms that our approach is more flexible." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 358, + 194, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 194, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 194, + 370 + ], + "type": "text", + "content": "4.6. The Impact of Mixed Data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 286, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 286, + 638 + ], + "type": "text", + "content": "We employ two datasets (AIST++ and HumanML3D) to train our 3D human motion VQ-VAE so that the motion tokens drop in the same latent space. To empirically show this point, we conduct experiments to count the number of shared motion tokens. For a trained 3D human motion VQ-VAE with AIST++ and HumanML3D, there are 855 vectors and 912 vectors from codebook used to construct the motion token in AIST++ and HumanML3D. The total number of vectors contained in the codebook is 1024. among them, 846 vectors (98.9% in AIST++ and 92.8% in HumanML3D) are shared to generate the motion tokens, which shows the feasibility of using the separate music-conditioned and text-conditioned motion datasets for music-text conditioned 3D dance generation task. We further verify the latent space of the motions by VQ-VAE, and we perform t-SNE visualizations of the raw motion distribution (before VQ-VAE) and feature distribution encoded by VQ-VAE (after VQ-VAE) from a2d motion and t2m motion in Figure 7. Two motion datasets are mixed successfully, which provides the potential integration of bimodality dance generation. Refer to the supplementary materials for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 148, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 148, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 148, + 659 + ], + "type": "text", + "content": "4.7. Model Efficiency" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "We also show the efficiency of our architecture in Table 4. In the inference stage, the complete model parameters and the inference time for 1s music are compared. Our architecture requires only about half model parameters and" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 541, + 182 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 541, + 182 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 541, + 182 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 541, + 182 + ], + "type": "image", + "image_path": "2da668e43b51e5fe4e3174a4b2d1539cb367aa1c010297da945da99eba7aa647.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 194, + 545, + 253 + ], + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 253 + ], + "type": "text", + "content": "Figure 4: Same music, different actions. These two generated dance share the same music and effect range (from 6s to 8s), with different text instructions: \"A person does a cartwheel\" (top), \"A person is spinning with arms spread out\" (bottom)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 268, + 544, + 387 + ], + "blocks": [ + { + "bbox": [ + 307, + 268, + 544, + 387 + ], + "lines": [ + { + "bbox": [ + 307, + 268, + 544, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 544, + 387 + ], + "type": "image", + "image_path": "d6ed11b5d684d9dc22721a26d539a64cbab1593e3d2d49602d0d348a64e9d5d7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 396, + 545, + 444 + ], + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 444 + ], + "type": "text", + "content": "Figure 5: Same action, different start time points. These two generated dance share the same music and text instruction (\"A person jumps up and down\") but different effect ranges: from 6 to 8s (top), from 7 to 9s (bottom)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 458, + 544, + 608 + ], + "blocks": [ + { + "bbox": [ + 307, + 458, + 544, + 608 + ], + "lines": [ + { + "bbox": [ + 307, + 458, + 544, + 608 + ], + "spans": [ + { + "bbox": [ + 307, + 458, + 544, + 608 + ], + "type": "image", + "image_path": "9350bd249e2b5343d504bd69a319079e85b30d72d69ca01a62fb482cabbe418e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": "Figure 6: Same action, different durations. These two generated dance share the same music and text instruction (\"A person is keeping jumping\") and action start time point (i.e., 6s) but different effect duration: 2s (top), 3s (bottom, i.e., jumped twice)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "inference time compared to Bailando [44], which attributes" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9949" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 90, + 161, + 162 + ], + "blocks": [ + { + "bbox": [ + 65, + 90, + 161, + 162 + ], + "lines": [ + { + "bbox": [ + 65, + 90, + 161, + 162 + ], + "spans": [ + { + "bbox": [ + 65, + 90, + 161, + 162 + ], + "type": "image", + "image_path": "01431764affd7d7ed49c8fcc7d29d4b6b9db99f838dee551d499ee0e5e534132.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 172, + 138, + 182 + ], + "lines": [ + { + "bbox": [ + 80, + 172, + 138, + 182 + ], + "spans": [ + { + "bbox": [ + 80, + 172, + 138, + 182 + ], + "type": "text", + "content": "(a) Before VQ-VAE" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 90, + 280, + 162 + ], + "blocks": [ + { + "bbox": [ + 186, + 90, + 280, + 162 + ], + "lines": [ + { + "bbox": [ + 186, + 90, + 280, + 162 + ], + "spans": [ + { + "bbox": [ + 186, + 90, + 280, + 162 + ], + "type": "image", + "image_path": "bc4989c4236db7c2dc909ea9980d7c3906bb0ab3b3b4fe0f22db1813e69da301.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 172, + 258, + 182 + ], + "lines": [ + { + "bbox": [ + 203, + 172, + 258, + 182 + ], + "spans": [ + { + "bbox": [ + 203, + 172, + 258, + 182 + ], + "type": "text", + "content": "(b) After VQ-VAE" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 191, + 287, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 191, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 287, + 226 + ], + "type": "text", + "content": "Figure 7: The t-SNE visualizations of motions before and after VQ-VAE (orange: motions in AIST++, blue: motions in HumanML3D)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 61, + 239, + 271, + 305 + ], + "blocks": [ + { + "bbox": [ + 61, + 239, + 271, + 305 + ], + "lines": [ + { + "bbox": [ + 61, + 239, + 271, + 305 + ], + "spans": [ + { + "bbox": [ + 61, + 239, + 271, + 305 + ], + "type": "table", + "html": "
MethodParametersInference time
FACT [29]120M8.300s/(1s music)
Bailando [44]152M0.236s/(1s music)
ours72M0.143s/(1s music)
", + "image_path": "2c609239249cab11e4fe0994cf3903102fd26be37d2df369bb754062ee98ab57.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 279, + 325 + ], + "lines": [ + { + "bbox": [ + 53, + 313, + 279, + 325 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 279, + 325 + ], + "type": "text", + "content": "Table 4: Comparisons of parameters and inference time." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 347, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 441 + ], + "type": "text", + "content": "that we do not need to encode separate features for the upper and lower half body. Compared to FACT [29], our architectures significantly reduce the inference time. The main reason is that FACT employs two transformers for the music, dance motion, and a cross-modal transformer, but we formulate it as the standard sequence-to-sequence translation task with one transformer only, which establishes the advantage of our architecture." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 453, + 119, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 119, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 119, + 466 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 474, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 676 + ], + "type": "text", + "content": "In this paper, a novel task that simultaneously integrates both music and text instruction for 3D dance generation is proposed. Due to the lack of the paired 3D dance motions conditioned on both music and text, we resort to two existing datasets, i.e., music2dance and text2motion, to perform this task and employ a 3D human motion VQ-VAE to project the motions of the two datasets into a shared latent space so that the two datasets with different distributions can be effectively mixed for training. Moreover, we also propose a cross-modal transformer architecture to generate 3D dance without degrading the performance of music-conditioned dance generation. Two new evaluation metrics, MPD and FS, are proposed to reflect the quality of generated motion. Extensive experiments show that our method can generate dance motion that matches both music and text in a realistic and coherent way while maintaining comparable performance on two single modalities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Acknowledgement. This project is supported by the Ministry of Education Singapore under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006)," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "and Singapore National Research Foundation (\"CogniVision - Energy-autonomous always-on cognitive and attentive cameras for distributed real-time vision with milliwatt power consumption\" grant NRF-CRP20-2017-0003) - www.green-ic.org/CogniVision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 141, + 365, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 141, + 365, + 153 + ], + "spans": [ + { + "bbox": [ + 306, + 141, + 365, + 153 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 161, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 311, + 161, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 161, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 311, + 161, + 545, + 182 + ], + "type": "text", + "content": "[1] mixamo. https://www MIXamo.com/. Accessed: 2023-03-07. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 184, + 547, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 184, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 311, + 184, + 547, + 237 + ], + "type": "text", + "content": "[2] Hyemin Ahn, Timothy Ha, Yunho Choi, Hwiyeon Yoo, and Songhwai Oh. Text2action: Generative adversarial synthesis from language to action. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5915-5920. IEEE, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 311, + 239, + 545, + 282 + ], + "type": "text", + "content": "[3] Hyemin Ahn, Jaehun Kim, Kihyun Kim, and Songhwai Oh. Generative autoregressive networks for 3d dancing move synthesis from music. IEEE Robotics and Automation Letters, 5(2):3501-3508, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 284, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 284, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 311, + 284, + 545, + 326 + ], + "type": "text", + "content": "[4] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719-728. IEEE, 2019. 3, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 327, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 327, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 311, + 327, + 545, + 392 + ], + "type": "text", + "content": "[5] Omid Alemi, Jules Françoise, and Philippe Pasquier. Groovenet: Real-time music-driven dance movement generation using artificial neural networks. In Workshop on Machine Learning for Creativity, 23rd ACM SIGKDD Conference on Knowledge Discovery and Data Mining, volume 8, page 26, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 394, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 394, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 311, + 394, + 545, + 436 + ], + "type": "text", + "content": "[6] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. Teach: Temporal action compositions for 3d humans. In International Conference on 3D Vision (3DV), 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 311, + 437, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 437, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 311, + 437, + 545, + 502 + ], + "type": "text", + "content": "[7] Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In IEEE Virtual Reality and 3D User Interfaces (VR), pages 1-10. IEEE, 2021. 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 311, + 504, + 545, + 536 + ], + "type": "text", + "content": "[8] Hanting Chen, Yunhe Wang, Jianyuan Guo, and Dacheng Tao. VanillaNet: the Power of Minimalism in Deep Learning. arXiv preprint arXiv:2305.12972, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 311, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 311, + 537, + 545, + 581 + ], + "type": "text", + "content": "[9] Kang Chen, Zhipeng Tan, Jin Lei, Song-Hai Zhang, Yuan-Chen Guo, Weidong Zhang, and Shi-Min Hu. Choreomaster: choreography-oriented music-driven dance synthesis. ACM Transactions on Graphics (TOG), 40(4):1-13, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 624 + ], + "type": "text", + "content": "[10] Rukun Fan, Songhua Xu, and Weidong Geng. Example-based automatic music-driven conventional dance motion synthesis. IEEE transactions on visualization and computer graphics, 18(3):501-515, 2011. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 625, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 657 + ], + "type": "text", + "content": "[11] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. arXiv preprint arXiv:2305.10924, 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[12] Joao P Ferreira, Thiago M Coutinho, Thiago L Gomes, José F Neto, and et al. Learning to dance: A graph convolutional adversarial network to generate realistic dance motions from audio. Computers & Graphics, 94:11-21, 2021. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9950" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[13] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1396–1406, 2021. 3, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "text", + "content": "[14] Shiry Ginosar, Amir Bar, Gefen Kohavi, Caroline Chan, Andrew Owens, and Jitendra Malik. Learning individual styles of conversational gesture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 229 + ], + "type": "text", + "content": "[15] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "text", + "content": "[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "text", + "content": "[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597, 2022. 3, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 386 + ], + "type": "text", + "content": "[18] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 441 + ], + "type": "text", + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in neural information processing systems, volume 30, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "text", + "content": "[20] Duc N.M. Hoang, Shiwei Liu, Radu Marculescu, and Zhangyang Wang. Revisiting pruning at initialization through the lens of ramanujan graph. In International Conference on Learning Representations, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 489, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 533 + ], + "type": "text", + "content": "[21] Ruozhi Huang, Huang Hu, Wei Wu, Kei Sawada, Mi Zhang, and Daxin Jiang. Dance revolution: Long-term dance generation with music via curriculum learning. In International Conference on Learning Representations, 2021. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "text", + "content": "[22] Yanghua Jin, Jiakai Zhang, Minjun Li, Yingtao Tian, Huachun Zhu, and Zhihao Fang. Towards the automatic anime characters creation with generative adversarial networks. arXiv preprint arXiv:1708.05509, 2017. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 621 + ], + "type": "text", + "content": "[23] Hsuan-Kai Kao and Li Su. Temporally guided music-to-body-movement generation. In Proceedings of the 28th ACM International Conference on Multimedia, pages 147-155, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[24] Kosmas Kritsis, Aggelos Gkiokas, Aggelos Pikrakis, and Vassilis Katsouros. Danceconv: Dance motion generation with convolutional networks. IEEE Access, 10:44982-45000, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[25] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. Advances in neural information processing systems, 32, 2019. 1, 2, 6, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[26] Juheon Lee, Seohyun Kim, and Kyogu Lee. Listen to dance: Music-driven choreography generation using autoregressive encoder-decoder network. arXiv preprint arXiv:1811.00818, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[27] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1272-1279, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[28] Jiaman Li, Yihang Yin, Hang Chu, Yi Zhou, Tingwu Wang, Sanja Fidler, and Hao Li. Learning to generate diverse dance motions with transformer. arXiv preprint arXiv:2008.08171, 2020. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[29] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 1, 2, 5, 6, 7, 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[30] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "type": "text", + "content": "[31] Angela S. Lin, Lemeng Wu, Rodolfo Corona, Kevin W. H. Tai, Qixing Huang, and Raymond J. Mooney. Generating animated videos of human activities from natural language descriptions. 2018. 3, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 545, + 399 + ], + "type": "text", + "content": "[32] Xinyin Ma, Gongfan Fang, and Xinchao Wang. LLM-Pruner: On the Structural Pruning of Large Language Models. arXiv preprint arXiv:2305.11627, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 400, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 545, + 453 + ], + "type": "text", + "content": "[33] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "type": "text", + "content": "[34] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "text", + "content": "[35] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, page 480-497, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 557, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 589 + ], + "type": "text", + "content": "[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. *Big data*, 4(4):236-252, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 591, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 645 + ], + "type": "text", + "content": "[37] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems, 109:13-26, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 713 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "9951" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[39] Sucheng Ren, Xingyi Yang, Songhua Liu, and Xinchao Wang. SG-Former: Self-guided Transformer with Evolving Token Reallocation. In IEEE International Conference on Computer Vision, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "text", + "content": "[40] Sucheng Ren, Daquan Zhou, Shengfeng He, Jiashi Feng, and Xinchao Wang. Shunted Self-Attention via Multi-Scale Token Aggregation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 205 + ], + "type": "text", + "content": "[41] Xuanchi Ren, Haoran Li, Zijian Huang, and Qifeng Chen. Self-supervised dance video synthesis conditioned on music. In Proceedings of the 28th ACM International Conference on Multimedia, pages 46-54, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 260 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[43] Ken Shoemake. Animating rotation with quaternion curves. In Proceedings of the 12th annual conference on Computer graphics and interactive techniques, pages 245-254, 1985. 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 370 + ], + "type": "text", + "content": "[44] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11050-11059, 2022. 1, 2, 4, 6, 7, 8, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 372, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 424 + ], + "type": "text", + "content": "[45] Guofei Sun, Yongkang Wong, Zhiyong Cheng, Mohan S Kankanhalli, Weidong Geng, and Xiangdong Li. Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Transactions on Multimedia, 23:497-509, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "text", + "content": "[46] Taoran Tang, Jia Jia, and Hanyang Mao. Dance with melody: An LSTM-autoencoder approach to music-oriented dance synthesis. In Proceedings of the 26th ACM international conference on Multimedia, pages 1598-1606, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "type": "text", + "content": "[47] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 526, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 558 + ], + "type": "text", + "content": "[48] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 603 + ], + "type": "text", + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 604, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 287, + 658 + ], + "type": "text", + "content": "[50] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[51] Nelson Yalta, Shinji Watanabe, Kazuhiro Nakadai, and Tetsuya Ogata. Weakly-supervised deep recurrent neural networks for basic dance step generation. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2019. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 385 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[52] Tatsuro Yamada, Hiroyuki Matsunaga, and Tetsuya Ogata. Paired recurrent autoencoders for bidirectional translation between robot actions and linguistic descriptions. IEEE Robotics and Automation Letters, 3(4):3441-3448, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[53] Xingyi Yang, Daquan Zhou, Songhua Liu, Jingwen Ye, and Xinchao Wang. Deep model reassembly. In Advances in neural information processing systems, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[54] Zijie Ye, Haozhe Wu, Jia Jia, Yaohua Bu, Wei Chen, Fanbo Meng, and Yanfeng Wang. Choreonet: Towards music to dance synthesis with choreographic action unit. In Proceedings of the 28th ACM International Conference on Multimedia, pages 744-752, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 250 + ], + "type": "text", + "content": "[55] Weihao Yu, Chenyang Si, Pan Zhou, Mi Luo, Yichen Zhou, Jiashi Feng, Shuicheng Yan, and Xinchao Wang. Metaformer Baselines for Vision. arXiv preprint arXiv:2210.13452, 2022.4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 252, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 285 + ], + "type": "text", + "content": "[56] Ye Yuan and Kris Kitani. Dlow: Diversifying latent flows for diverse human motion prediction. In European Conference on Computer Vision, pages 346-364. Springer, 2020. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 341 + ], + "type": "text", + "content": "[57] Wenlin Zhuang, Congyi Wang, Jinxiang Chai, Yangang Wang, Ming Shao, and Siyu Xia. Music2dance: Dancenet for music-driven dance generation. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 18(2):1-21, 2022. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "text", + "content": "[58] Wenlin Zhuang, Yangang Wang, Joseph Robinson, Congyi Wang, Ming Shao, Yun Fu, and Siyu Xia. Towards 3d dance motion synthesis and control. arXiv preprint arXiv:2006.05743, 2020. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9952" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/9d269c18-5110-4a81-a1b9-317c1455548e_content_list.json b/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/9d269c18-5110-4a81-a1b9-317c1455548e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b76f8a1a0e6ed80b6d46e5b3390e5c61539c231a --- /dev/null +++ b/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/9d269c18-5110-4a81-a1b9-317c1455548e_content_list.json @@ -0,0 +1,1617 @@ +[ + { + "type": "text", + "text": "TMA: Temporal Motion Aggregation for Event-based Optical Flow", + "text_level": 1, + "bbox": [ + 145, + 130, + 823, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haotian Liu1 Guang Chen1* Sanqing Qu1 Yanping Zhang1 \nZhijun Li1 Alois Knoll2 Changjun Jiang1 \n1Tongji University 2Technical University of Munich", + "bbox": [ + 199, + 178, + 769, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras have the ability to record continuous and detailed trajectories of objects with high temporal resolution, thereby providing intuitive motion cues for optical flow estimation. Nevertheless, most existing learning-based approaches for event optical flow estimation directly remould the paradigm of conventional images by representing the consecutive event stream as static frames, ignoring the inherent temporal continuity of event data. In this paper, we argue that temporal continuity is a vital element of event-based optical flow and propose a novel Temporal Motion Aggregation (TMA) approach to unlock its potential. Technically, TMA comprises three components: an event splitting strategy to incorporate intermediate motion information underlying the temporal context, a linear lookup strategy to align temporally fine-grained motion features and a novel motion pattern aggregation module to emphasize consistent patterns for motion feature enhancement. By incorporating temporally fine-grained motion information, TMA can derive better flow estimates than existing methods at early stages, which not only enables TMA to obtain more accurate final predictions, but also greatly reduces the demand for a number of refinements. Extensive experiments on DSEC-Flow and MVSEC datasets verify the effectiveness and superiority of our TMA. Remarkably, compared to E-RAFT, TMA achieves a $6\\%$ improvement in accuracy and a $40\\%$ reduction in inference time on DSEC-Flow. Code will be available at https://github.com/ispc-lab/TMA.", + "bbox": [ + 75, + 300, + 473, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 744, + 209, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Optical flow aims to compute the velocity of objects on the image plane without geometry prior and is a fundamental topic in event-based vision [8, 42]. It plays an important role in many applications, such as ego-motion estimation [38, 45], image reconstruction [24], and video frame interpolation [32, 31]. Recently, learning-based methods [29, 18, 30] have dominated frame-based optical flow", + "bbox": [ + 75, + 770, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d79da3f60da344d5debf26737a7222933e2d1633fbf35168e314c06375fb14cc.jpg", + "image_caption": [ + "(a) reference image" + ], + "image_footnote": [], + "bbox": [ + 506, + 258, + 694, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76e22e787cc0093cd70f73f2ae77508cbad73d691a23df358884f3ee86f683c7.jpg", + "image_caption": [ + "(b) event frame" + ], + "image_footnote": [], + "bbox": [ + 696, + 258, + 885, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/197e01e0c9ab95123c67414acdb7f28e2441de1741495e0250013e52ab5bb6cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 358, + 694, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/71b7e75eee9838c46a76b72c51fcaf25d19f44bae976519f476782507c6b6319.jpg", + "image_caption": [ + "(c) E-RAFT output & motion feature (d) our output & motion feature" + ], + "image_footnote": [], + "bbox": [ + 508, + 446, + 694, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2cddd66dab12422f7745cde39ca77fb477f80355b39094fea5f61fc9c634c840.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 358, + 885, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1b82bcbcfe65b05eb791f6404309750ddc92df34c88fc0a65b6383fd574f6c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 446, + 883, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1: Accuracy illustrations. (a) Reference image from DSEC-Flow. (b) Corresponding event frame. (c) The final flow prediction and visualization of the motion feature from E-RAFT [11]. (d) The final flow prediction and visualization of the motion feature from our proposed method. We visualize the motion feature at the initial stage by taking the average across channels. E-RAFT fails to generate informative motion features and results in blurred boundaries. In contrast, our method utilizes temporally fine-grained motion information to address the information scarcity issue in motion features, generating high-quality predictions.", + "bbox": [ + 496, + 547, + 890, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "by employing correlation volumes to derive motion features for flow regression. Inspired by this, several event-based approaches [11, 34] adopt a similar paradigm by converting the consecutive event stream into grid-like representations. Despite encouraging progress, these methods still suffer from non-negligible limitations, due to the great differences between event data and conventional images. Distinct from dense and colorful conventional (RGB) images, event data features spatial sparsity and a lack of intensity information. It is prone to result in close matching scores and invalid regions in the correlation volume. Consequently, less informative motion features are derived and regress inaccurate", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*corresponding author: guangchen@tongii.edu.cn", + "bbox": [ + 94, + 886, + 359, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9685", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "predictions. Figure 1 illustrates this issue in E-RAFT [11].", + "bbox": [ + 76, + 90, + 464, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the saying goes, each coin has two sides. Though event cameras encode visual information in a sparse manner, they are capable of capturing continuous and detailed object trajectories with high temporal resolution, thereby providing rich motion cues for optical flow estimation. Several model-based methods [9, 27] benefit from the temporal continuity by warping events along point trajectories, achieving decent performance. Drawing on this observation, we contend that the temporal continuity is a vital element of event-based optical flow and propose to unleash its potential within a learning-based framework.", + "bbox": [ + 75, + 107, + 467, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To materialize our idea, we propose a novel Temporal Motion Aggregation (TMA) approach to explore the inherent temporal continuity of event data. Technically, we revamp the dominant learning-based architectures with three distinct components. We first introduce an event splitting strategy for temporally-dense correlation volumes computation. By splitting the event stream into multiple segments and extracting their features, feature similarities are compared between the first feature and all others, which record rich intermediate motion information. Aware of that correlation volumes record motions between different time spans, we then design a linear lookup strategy to sample each correlation volume based on the corresponding flow estimate to encode motion features. As a result, fine-grained motions are warped to the same pixels across motion features. Furthermore, we consider the incorrect motion patterns in intermediate motion features due to the manual lookup. Therefore, we propose a novel pattern aggregation module for motion feature enhancement by aggregating consistent patterns between each intermediate motion feature and the last one (without manual lookup). Thanks to the incorporation of temporally fine-grained motion information, TMA can effectively address the information scarcity in motion features and thus generate good flow estimates at early stages, which not only enable TMA to obtain accurate flow predictions, but also greatly reduces the demand for a number of time-consuming refinements.", + "bbox": [ + 75, + 273, + 468, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our TMA on DSEC-Flow [10] and MVSEC [44] datasets. Extensive experiments demonstrate the superior advantages of TMA.", + "bbox": [ + 76, + 681, + 468, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 96, + 727, + 433, + 742 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We argue that the temporal continuity of event data is a vital element of event-based optical flow and propose to unlock its potential in a learning-based framework.", + "- We propose a novel Temporal Motion Aggregation (TMA) approach, which comprises three components: an event splitting strategy, a linear lookup strategy and a motion pattern aggregation module.", + "- The incorporation of temporally fine-grained motion information enables TMA to achieve high accuracy" + ], + "bbox": [ + 94, + 750, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "while maintaining high efficiency. Compared with E-RAFT, TMA achieves a $6\\%$ improvement in accuracy and a $40\\%$ reduction in inference time on DSEC-Flow.", + "bbox": [ + 531, + 90, + 890, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 150, + 640, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning methods for optical flow estimation. Recently, correlation volume based approaches dominate the learning-based optical flow estimation [25, 17, 29, 18, 37, 14]. As a milestone, PwCNet [29] summarizes a simple principle in architecture construction: pyramidal processing, warping and correlation volumes computation. Following the well-established pipeline, LiteFlowNet series [16, 17, 15] propose cascaded feature pyramids with a lightweight convolutional burden. VCN [37] proposes 4D volumetric U-Net architectures to enlarge the receptive fields for correlation volumes. IRR [18] notes the reusability of pyramidal refinement for flow update and proposes a shared-weights refiner across all scales, which can be combined with mainstream learning methods.", + "bbox": [ + 496, + 176, + 890, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the correlation volumes used in previous works are computed locally, which cannot fully deal with long-term motions. RAFT [30] firstly introduces the full correlation volume to capture global visual similarities. Separable Flow [41] proposes a separable correlation volume module as a drop-in replacement to correlation volumes and uses non-local aggregation layers [39, 40] to refine the correlation volumes efficiently. Besides, transformer components show promising performance on optical flow [36, 19, 35]. GMA [19] notices the underlying motion clues in the context and introduces an attention layer to aggregate hidden motions, improving the accuracy based on RAFT architecture. GMFlow [36] reformulates optical flow as a global matching problem by attention blocks, suggesting a new paradigm for flow estimation.", + "bbox": [ + 496, + 387, + 890, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We draw inspiration from GMA [19] to design our motion pattern aggregation module. In contrast to GMA which aggregates hidden motion patterns in the spatial context for motion feature enhancement, we extend the pattern aggregation to the temporal dimension and leverage feature cross-similarities to enhance the motion feature.", + "bbox": [ + 496, + 614, + 890, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event-based optical flow estimation. A number of traditional works remould frame-based approaches into event-based optical flow estimation. Benosman et al. [4] extends the classic Lucas-Kanade method to event-based optical flow. Liu et al. [21, 22] implement event-based corner keypoint detection and novel block-matching optical flow. Gallego et al. [9] propose a unifying contrast maximization (CM) framework and MultiCM [27] extends it to complex scenes. Besides, plane fitting [1, 3], filter banks [6] and time surface matching [2] are proposed to boost performance.", + "bbox": [ + 496, + 704, + 890, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, learning-based methods have dominated frame-based optical flow, inspiring attempts at event-based optical flow. Early learning methods [43, 45, 38] represent", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9686", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bd2575db2ab1f802b8ae96be3731aacc30879bf0446c7a6eb2bb30f9a195f1ba.jpg", + "image_caption": [ + "Figure 2: The overall architecture of TMA. First, we split the event stream into segments and extract features with a share-weights feature extractor. Then, regarding the first feature as reference, we correlate all other features to compute temporally-dense correlation volumes, which contain rich embedded motion information. Next, aware of the different time spans of correlation volumes, we look up each correlation volume based on a corresponding flow estimate. All sampled correlation maps are delivered into a share-weights motion feature encoder to derive motion features. We further propose a novel motion pattern aggregation module to enrich the spatial information of motion features. By concatenating the enhanced motion features and utilizing several refinements, high-quality flow predictions are generated." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 883, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the event stream as a static frame and utilize the architecture of U-Net [26] to predict optical flow, which can only address small motions. Mainstream works [7, 34, 11] introduce two consecutive event frames and follow the correlation volume based pipeline (PWCNet and RAFT), achieving higher accuracy on large-scales datasets [44, 10]. E-RAFT [11, 12] extends RAFT [30] and introduces full correlation volume into event-based vision. DCEIFlow [34] combines conventional images and event data to boost higher accuracy. Note that our work and STE-FlowNet [7] share a similar motivation to introduce temporal continuity with fine-grained discretization. However, STE-FlowNet takes a frame-by-frame manner [18] to update flow, leading to an insufficient inference process.", + "bbox": [ + 75, + 357, + 472, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrary to previous learning-based methods, we do not merely adopt the classic correlation volume based pipeline. Instead, we regard the temporal continuity of event data as a vital component of optical flow and incorporate the intermediate motion information into a learning-based framework, which boosts higher accuracy and efficiency.", + "bbox": [ + 76, + 569, + 470, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 676, + 212, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input event stream $\\{\\mathcal{E}\\}$ from $t_0$ to $t_1$ , the goal of optical flow is to estimate a dense displacement field $\\mathbf{u}:\\mathbb{R}^2\\to \\mathbb{R}^2$ which maps each per-pixel $(x,y)$ at $t_0$ to its corresponding coordinate $(x^{\\prime},y^{\\prime})$ at $t_1$ . To address this, mainstream learning-based approaches compute correlation volumes from two event frames as a matching prior. Such principle inspires several methods [11, 34] by squeezing event data into consecutive frames. Unfortunately, the characteristics of event data, spatial sparsity and lack of intensity records, are prone to result in close matching scores and invalid regions in the correlation volume. Consequently, less informative motion features are derived and regress inaccurate predictions. Here we address the information scarcity", + "bbox": [ + 75, + 703, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of motion features from the attractive temporal continuity of event data. Temporally fine-grained motion information in event data is sufficient to supplement the spatial richness of motion features, yielding high-quality flow predictions.", + "bbox": [ + 496, + 356, + 890, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we first describe the preliminary and event representation, and then present the detailed architecture of the proposed TMA.", + "bbox": [ + 496, + 417, + 890, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary and Event Representation", + "text_level": 1, + "bbox": [ + 498, + 470, + 831, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event cameras record light intensity changes in an asynchronous way. Each pixel serves as an independent trigger and generates an event instantly whenever the log intensity changes exceeding a threshold $C$ . An event $\\boldsymbol{e}_k = (\\boldsymbol{x}_k, p_k, t_k)$ is a triplet containing triggered pixel location $\\boldsymbol{x}_k = (x_k, y_k)^T$ , timestamp $t_k$ and signed intensity change $(p_k = \\pm 1)$ .", + "bbox": [ + 496, + 493, + 890, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ease the feature extraction and preserve the temporal continuity of the event data, we firstly transform the event stream into a 3D volume $V(x,y,t)$ along the temporal dimension following existing works [45, 11, 24]. Given an event stream $\\{\\mathcal{E}\\}$ , $V(x,y,t)$ is generated as:", + "bbox": [ + 496, + 599, + 890, + 675 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nt _ {i} ^ {*} = (B - 1) \\left(t _ {i} - t _ {1}\\right) / \\left(t _ {N _ {e}} - t _ {1}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 683, + 890, + 700 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {V} (x, y, t) = \\sum_ {i} p _ {i} k _ {b} \\left(x - x _ {i}\\right) k _ {b} \\left(y - y _ {i}\\right) k _ {b} \\left(t - t _ {i} ^ {*}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 703, + 890, + 733 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nk _ {b} (a) = \\max (0, 1 - | a |), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 738, + 890, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t_i^*$ represents the $i$ -th time bin. $B$ and $N_e$ represent the length of time bins and the number of events, respectively. $k_b(a)$ is the bilinear interpolation function.", + "bbox": [ + 496, + 763, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. TMA Architecture", + "text_level": 1, + "bbox": [ + 500, + 816, + 679, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We develop our framework based on the successful RAFT architecture [30]. Figure 2 summarizes the architecture of the proposed TMA, which consists of three components, an event splitting strategy, a linear lookup strategy", + "bbox": [ + 496, + 839, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9687", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8da1b9b2f86637a5772622540d3d8a83337ccf577b5648551fa3ce8ccecb5f51.jpg", + "image_caption": [ + "Figure 3: Linear lookup strategy. We compute correlation volumes by comparing the similarity of first feature with all others, which are misaligned due to different time spans. To rectify this, we narrow down the lookup coordinates linearly to warp correlation maps to spatially aligned pixels." + ], + "image_footnote": [], + "bbox": [ + 98, + 87, + 444, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and a motion pattern aggregation module. With the proposed method, we successfully unlock the potential of temporal continuity to predict event-based optical flow within a learning-based framework.", + "bbox": [ + 75, + 332, + 468, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event splitting. Instead of viewing the continuous event stream as a static frame which tends to be a matter of common knowledge in recent works [45, 11, 34], we prefer to identify it as a high frame rate video clip, which records continuous and detailed trajectories of objects, namely, provides intuitive motion cues for optical flow estimation. To introduce rich intermediate motion information into our method, we first propose an event splitting strategy, which splits the event stream into multiple segments. Therefore, more intermediate event frames are retained for correlation volume computation.", + "bbox": [ + 75, + 393, + 468, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, given an event stream $\\{\\mathcal{E}\\}$ from $t_0$ to $t_1$ , we split it into $g$ segments equally and the time span between adjacent segments is $dt$ . Besides, we introduce an auxiliary event segment from $t_0 - dt$ to $t_0$ as reference. Then, $g + 1$ event representations are generated following Sec. 3.1. All $g + 1$ event representations are then delivered into a share-weights encoder to achieve features $\\pmb{F}_i \\in \\mathbb{R}^{H \\times W \\times C}, i = 0, 1, 2, \\dots, g,$ where $H, W$ and $D$ represent the height, width and channel of features. We use all features to compute temporally-dense correlation volumes $\\pmb{C}_i, i = 1, 2, \\dots, g,$ which are given by:", + "bbox": [ + 76, + 559, + 468, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {C} _ {i} = \\frac {\\boldsymbol {F} _ {0} \\boldsymbol {F} _ {i} ^ {T}}{\\sqrt {D}} \\in \\mathbb {R} ^ {H \\times W \\times H \\times W}, i = 1, 2, \\dots , g, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 734, + 468, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{F}_0$ is the first feature and $\\pmb{F}_i$ is the $i$ -th feature derived from the $i$ -th event frame.", + "bbox": [ + 76, + 779, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporally-dense correlation volumes compare feature similarities across multiple time spans, enabling refined records of relative motions. This construction is vastly different from the classic full correlation volume that only contains feature similarities between a fixed time span.", + "bbox": [ + 76, + 810, + 468, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Linear lookup. Recall that each correlation volume $C_i$", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/edd9caa0b9361b83c8177c5cf45bb2de7ef9eea9cdead27b15955c2036afefbb.jpg", + "image_caption": [ + "Figure 4: Motion pattern aggregation. We pair each intermediate motion feature to the last motion feature. Then each pair is sent to cross-attention blocks to aggregate consistent patterns. Hence, each intermediate motion feature carries information from the last one and consistent patterns receive emphasis." + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 861, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is achieved by correlating $F_{0}$ and $F_{i}$ , which means the recorded relation motions are misaligned in time. While our goal is to predict optical flow from $t_{0}$ to $t_{1}$ , we hope that all sampled correlation maps could present the relative motions within the same time span. Therefore, based on that correlation volumes are evenly constructed in time and magnitudes of optical flow change linearly over a short period, we narrow down lookup coordinates in a linear fashion to match the timestamp of each correlation volume.", + "bbox": [ + 496, + 351, + 890, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 displays the linear lookup strategy. In detail, given current flow estimate $\\mathbf{u}$ , we split it into $g$ segments to sample each correlation volume. The linear lookup operation can be expressed as:", + "bbox": [ + 496, + 487, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} \\mathbf {u} = \\left(\\mathbf {c o o r d s} _ {1} - \\mathbf {c o o r d s} _ {0}\\right) / g = \\mathbf {u} / g \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 556, + 890, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {c o r r} _ {i} = S \\left(\\boldsymbol {C} _ {i}, \\operatorname {c o o r d s} _ {0} + i \\mathrm {d u}\\right), i = 1, 2, \\dots , g, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 574, + 890, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{coord}_0$ is the initialized coordinates at $t_0$ (reference time), $S(\\cdot, \\cdot)$ represents the lookup operation, $\\mathbf{coord}_i$ are lookup coordinates for $C_i$ and $\\mathbf{corr}_i$ is the correlation map sampled from $C_i$ . After delivering all correlation maps into a share-weights motion feature encoder, aligned motion features are achieved.", + "bbox": [ + 496, + 598, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From the perspective of image contrast [9], the standard lookup is responsible for warping possible matching points between two images to the same pixel. Our linear lookup extends it by warping matching points of a temporally fine-grained motion to the same pixel. As a result, similar motions are presented at same pixels across correlation maps, showing high contrast.", + "bbox": [ + 496, + 689, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion pattern aggregation. A standard operation for next step is to concatenate the motion features as a whole to update flow and the improvements made so far already alleviate the information scarcity in motion features. However, we notice that the concatenation operation treats all motion features equally, which means the majority of motion features, intermediate ones, dominate the flow regression. Due", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9688", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4cc613ce347d8063a08a953b1258344670e8f9d30b3ff8cbf420ebacd78e806e.jpg", + "image_caption": [ + "Figure 5: Qualitative examples of DSEC-Flow. For each two columns, we show the reference image and event frame and compare our method with a state-of-the-art baseline E-RAFT [11]. Significant improvements are highlighted by red boxes." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 379, + 294 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c9d09ce8f6abd1e8a67f68c318998a498a71fed2457dfd334adf1bcfb3bfa69c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 88, + 633, + 294 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2a660dd95348afbefb349d5a459a791499afdd118c8ebd0d7b0e307863e2f783.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 89, + 883, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to the linear optical flow assumption and manual lookup operation, these motion features are not perfectly correct. A natural thought is to put more emphasis on the last motion feature that is without manual lookup and retain the patterns in intermediate motion features that are consistent with the last motion feature. To this end, we balance the weights between the last motion feature and others through cross-similarities computed by transformers [33, 19].", + "bbox": [ + 75, + 340, + 468, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 4 shows the overview of motion pattern aggregation module. Given a set of motion features $M\\mathbf{F}_i \\in \\mathbb{R}^{H \\times W \\times D'}$ , $i = 1,2,\\dots,g$ , we pair each intermediate motion feature to the last one. Then each pair $(M\\mathbf{F}_i,M\\mathbf{F}_g)$ is sent to cross-attention blocks to compare cross-similarities and aggregate consistent patterns. The enhanced motion feature $\\widehat{M}\\mathbf{F}_i$ is obtained as:", + "bbox": [ + 76, + 462, + 470, + 568 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {A} = \\operatorname {A t t e n t i o n} \\left(\\boldsymbol {M} \\boldsymbol {F} _ {i} W ^ {Q}, \\boldsymbol {M} \\boldsymbol {F} _ {g} W ^ {K}, \\boldsymbol {M} \\boldsymbol {F} _ {g} W ^ {V}\\right) (7) \\\\ \\widehat {\\boldsymbol {M F}} _ {i} = \\boldsymbol {M F} _ {i} + \\operatorname {M L P} \\left(\\left[ \\boldsymbol {M F} _ {i}, \\boldsymbol {A W} ^ {O} \\right]\\right), (8) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 579, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\operatorname{Attention}(Q, K, V) = \\operatorname{softmax}\\left(\\frac{QK^T}{\\sqrt{d_k}}\\right)V$ , $d_k$ is the dimension of $K$ . $W^Q$ , $W^K$ , $W^V$ are projection matrices. $\\mathrm{MLP}(\\cdot)$ is multi-layer perceptron and $W^O$ is another projection matrix. We omit the normalization functions in the manuscript for simplicity. Motion pattern aggregation module does not change the dimension of motion features, so multiple cross-attention layers can be stacked easily.", + "bbox": [ + 75, + 632, + 468, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once motion features are enhanced by the motion pattern aggregation module, each $\\widehat{MF}_i$ carries motion patterns that are consistent with $MF_{g}$ . Namely, the last motion feature receives high attention and consistent patterns in motion features are fully utilized.", + "bbox": [ + 75, + 748, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Above three components are seamlessly integrated into a unified pipeline: incorporating intermediate motion information, warping matching motions and emphasizing consistent motion patterns. Attributed to the innovative application of temporal continuity to a learning-based frame", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "work, our proposed method is capable of overcoming the information scarcity issue in motion features and generating high-quality flow estimates even at an early refinement stage. This enables our method to achieve accurate final predictions, while reducing the demand for a number of refinements.", + "bbox": [ + 496, + 340, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Supervision", + "text_level": 1, + "bbox": [ + 500, + 440, + 627, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following RAFT [30], we choose $L_{1}$ distance between the predictions and ground-truth as the supervision signal for our model. The loss function $\\mathcal{L}$ is formally defined as:", + "bbox": [ + 498, + 463, + 890, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {j = 1} ^ {N} \\gamma^ {N - j} \\left\\| \\mathbf {u} ^ {g t} - \\mathbf {u} ^ {j} \\right\\| _ {1}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 518, + 890, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{u}^{gt}$ denotes the flow ground-truth and $\\mathbf{u}^j$ the flow prediction at $j$ -th stage. $N$ is the total refinement stages. The weight $\\gamma$ balances different stages of predictions.", + "bbox": [ + 496, + 574, + 890, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 633, + 633, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and evaluation setup. Following previous works [11, 45, 7], we provide experimental details and extensive comparison results on two event-based datasets DSEC-Flow [10] and MVSEC [44]. For DSEC-Flow, we train models on the official training set and evaluate DSEC-Flow test set on the public benchmark. For MVSEC, we conduct two kinds of experiments. Basically, following previous methods [43, 27], we train methods on outdoor_day2 sequence with ground-truth corresponding to a time interval of $dt = 1$ and $dt = 4$ gray images and evaluate 800 frames of outdoor_day1 and three indoor_flying sequences. Equally importantly, considering the supervised learning setting of our method and the large gap between outdoor_day and indoor_flying data, we extend the range of training data by including one indoor_flying sequence each time to further boost the accuracy of supervised learning methods.", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9689", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c56e5dad9054712f0f502bd3edfa29b45ea2e49357da9e742db3436340291a59.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsEPE1PE3PETime(ms)
MBMultiCM [27]3.4776.630.9-
SLEV-FlowNet†[43]2.3255.418.65
E-RAFT [11]0.7912.52.752
Ours0.7410.92.330
", + "bbox": [ + 81, + 88, + 464, + 174 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Results on DSEC-Flow. Best accuracy in bold. $\\dagger$ denotes the result is taken from E-RAFT [11]. Model-based (MB) methods need no training data; and supervised learning (SL) methods need ground-truth.", + "bbox": [ + 75, + 180, + 468, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. The end-point-error (EPE) is used as core metric for prediction accuracy for both DSEC-Flow and MVSEC. To evaluate the robustness against large displacements, we compute the percentage of pixels with EPE greater than $M$ pixels $(M\\mathrm{PE},M = 1,3)$ for DSEC-Flow. Percentage of pixels with EPE above 3 and $5\\%$ of the magnitude of the flow ground-truth (\\% Outlier) are presented for MVSEC. On DSEC-Flow, metrics are measured over pixels with valid ground-truth. While on MVSEC, metrics are measured over pixels with valid ground-truth and at least one triggered event.", + "bbox": [ + 75, + 250, + 468, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. The proposed TMA is implemented using PyTorch. We set the event segment $g = 5$ . For experiments on DSEC-Flow, the channel of event representation $B$ for each segment is set to 3. For MVSEC, $B$ is set to 1 when $dt = 1$ and 3 when $dt = 4$ , respectively. The sum of channels keeps the same with the state-of-the-art E-RAFT [11]. We only stack 1 layer of motion pattern aggregation module. Other components are basically identical to RAFT's model, including feature extractor, motion feature encoder and flow updater. Note that we use a small feature extractor whose final channel is 128. We predict 6 levels of optical flow and the weight $\\gamma$ is set to 0.8 for supervision. In training, we exploit AdamW [23] optimizer and One-cycle learning rate scheduler [28]. We train models on DSEC-Flow for 200K steps and MVSEC for 100K steps, both with a learning rate of 0.0002 and a batch size of 6.", + "bbox": [ + 75, + 415, + 470, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. DSEC-Flow", + "text_level": 1, + "bbox": [ + 76, + 665, + 205, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Accuracy comparison. Evaluation results on DSEC-Flow benchmark1 are provided in Table 1. Compared with E-RAFT, our method improves the EPE from 0.79 to 0.74. Our method is also leading in 1PE and 3PE with a margin of 1.6 and 0.4, showing better robustness.", + "bbox": [ + 75, + 688, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results are exhibited in Figure 5. In challenging regions such as textureless areas (the wall in second two columns, the shrubs in last two columns), existing correlation volume contains similar matching scores and fails to generate informative motion features, leading to low-quality predictions. On the contrary, our method aggregates", + "bbox": [ + 75, + 763, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/913c95c892169db088b345359884f596481916bbb25a0f55db95311de59310a9.jpg", + "image_caption": [ + "Figure 6: EPE vs. number of iterations at inference. The figure exhibits the prediction results on DSEC-Flow at different stages during inference. Our method achieves an accuracy $(\\mathrm{EPE} = 0.81)$ comparable to E-RAFT with 2 iterations while E-RAFT needs 6 iterations." + ], + "image_footnote": [], + "bbox": [ + 509, + 90, + 879, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "temporally consistent motion patterns to enhance motion features, distinguishing different moving objects and generating high-quality predictions.", + "bbox": [ + 496, + 339, + 890, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Efficiency comparison. The last column of Table 1 exhibits the inference time. Our method enjoys faster inference speed compared with E-RAFT, which is mainly attributed to fewer iterative updates. The reasonable explanation is that our method utilizes temporal motion information to enrich motion features, thus being able to generate good predictions even at an initial stage. To prove this, we compare the accuracy with E-RAFT at early inference stages in Figure 6. Our method achieves better flow initialization at early stages, supporting our argument. The ablation experiment on different iterations (Table 4e) also demonstrates that our method requires fewer iterations (ours: 6, E-RAFT: 12) to achieve surpassing accuracy.", + "bbox": [ + 496, + 383, + 892, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.MVSEC", + "text_level": 1, + "bbox": [ + 500, + 589, + 601, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inter-domain evaluation. Table 2 reports the results on MVSEC training with outdoor_day2 sequence. The top of Table 2 reports the results corresponding to $dt = 1$ grayscale frame and the bottom corresponds to $dt = 4$ grayscale frames. Our method shows comparable accuracy on outdoor_day1 for $dt = 1$ and highest accuracy on outdoor_day1 for $dt = 4$ among all compared methods. Unfortunately, like two other supervised learning methods, DCEI-Flow [34] and E-RAFT [11], our method fails to achieve superior results on indoor_flying sequences. One possible reason is that model based methods and unsupervised learning methods usually have better generalization ability, while supervised learning methods are limited if training data has a significantly large gap with test data [36, 34]. Since outdoor_day sequences and indoor_flying sequences have a great domain gap, it is challenging for supervised learning methods to generalize well on indoor_flying sequences.", + "bbox": [ + 496, + 613, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intra-domain evaluation. To verify the possible cause and better analyze the accuracy of our method under su", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1The ground-truth of DSEC-Flow dataset is unavailable, so we use last three checkpoints to evaluate the online benchmark. https://dsec.ifi.uzh.ch/uzh/dsec-flow-optical-flow-benchmark/.", + "bbox": [ + 75, + 862, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9690", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f2ed09171887807ee79b93eaf93ab7360fc399a14ce4e0bfa1c36feaaad76e50.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Inputindoor_flying1indoor_flying2indoor_flying3outdoor_day1
EPE% OutlierEPE% OutlierEPE% OutlierEPE% Outlier
dt = 1
MBBrebion et al. [5]E0.520.100.985.500.712.100.530.20
MultiCM (Burgers') [27]E0.420.100.600.590.500.280.300.10
SSLEV-FlowNet [43]E1.032.201.7215.101.5311.900.490.20
Spike-FlowNet [20]E0.84-1.28-1.11-0.49-
STE-FlowNet [7]E0.570.100.791.600.721.300.420.00
USLHagenaars et al. [13]E0.600.511.178.060.935.640.470.25
Zhu et al. [45]E0.580.001.024.000.873.000.320.00
ECN [38]E------0.300.02
SLE-RAFT [11]E1.105.721.9430.791.6625.200.240.00
DCEIFlow [34]E + I0.751.550.902.100.801.770.220.00
OursE1.063.631.8127.291.5823.260.250.07
dt = 4
MBMultiCM (Burgers') [27]E1.6912.952.4926.352.0619.031.259.21
SSLEV-FlowNet [43]E2.2524.704.0545.303.4539.701.237.30
Spike-FlowNet [20]E2.24-3.83-3.18-1.09-
STE-FlowNet [7]E1.7714.702.5226.102.2322.100.993.90
USLZhu et al. [45]E2.1824.203.8546.803.1847.801.309.70
Hagenaars et al. [13]E2.1621.513.9040.723.0029.601.6912.50
SLDCEIFlow [34]E + I2.0821.473.4842.052.5129.730.893.19
E-RAFT [11]E2.8140.255.0964.194.4657.110.721.12
OursE2.4329.914.3252.743.6042.020.701.08
", + "bbox": [ + 84, + 88, + 885, + 489 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/75f3f968d5c41e898b98264cd0c03cb842560de7f92dd5f719cdcc3a4f726435.jpg", + "table_caption": [ + "Table 2: Evaluation results on MVSEC [44] training with outdoor_day2 sequence. Model-based (MB) methods need no training data; semi-supervised learning (SSL) methods use grayscale images for supervision; unsupervised learning (USL) methods only require events; and supervised learning (SL) methods need ground-truth. E means the method only uses events as input; and I means the method requires images as input." + ], + "table_footnote": [], + "table_body": "
dt = 4Train Set +indoor_flying1indoor_flying2indoor_flying3outdoor_day1
EPE% OutlierEPE% OutlierEPE% OutlierEPE% Outlier
E-RAFT [11]indoor_flying1--2.3526.241.8418.070.721.38
Ours--2.1520.411.6814.300.640.98
E-RAFT [11]indoor_flying21.418.14--1.509.370.711.13
Ours1.325.94--1.508.880.640.96
E-RAFT [11]indoor_flying31.398.381.7717.52--0.721.46
Ours1.335.821.6712.80--0.671.10
", + "bbox": [ + 83, + 574, + 883, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Evaluation results on MVSEC [44] training with outdoor_day2 and one indoor_flying sequence. + means the indoor_flying sequence to be included into training set.", + "bbox": [ + 75, + 717, + 892, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pervised learning setting, we incorporate one indoor_flying sequence into training set each time and make comparisons with the supervised learning state-of-the-art E-RAFT. Table 3 shows the evaluation results corresponding to $dt = 4$ grayscale frames (For $dt = 1$ , refer to appendix). Compared with the results in Table 2, E-RAFT and our method both achieve great improvement on indoor_flying sequences (EPE of indoor_flying2, ours: $4.32 \\rightarrow 1.67$ , E-RAFT: 5.09 → 1.77), confirming the presence of domain gap between", + "bbox": [ + 75, + 757, + 472, + 893 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "outdoor_day and indoor_flying data. Compared with the accuracy of E-RAFT with the same training set, on the one hand, our method shows better generalization ability, which generates lower EPE and $\\%$ Outlier without seeing indoor_flying sequences. On the other hand, our method achieves higher accuracy on four sequences by incorporating indoor_flying sequences into training set. Overall, our method achieves a new state-of-the-art result, demonstrating the effectiveness of our proposed method.", + "bbox": [ + 496, + 757, + 893, + 893 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9691", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/97129115d91c360c5c2197e004c27cc0f178f606faa3d1a7b3fb4bbc14118cd4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
splitsEPE1PE3PETime (ms)
10.7911.912.7516
30.7610.982.4518
5/70.7410.862.3030
0.7411.132.2650
", + "bbox": [ + 86, + 88, + 472, + 179 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5a487fbbb1f9ed44b18924752496a67e6595a83a0ec621e5fb3f7ca20210152a.jpg", + "table_caption": [ + "(a) Event splitting. Intermediate motion information contributes" + ], + "table_footnote": [], + "table_body": "
setupEPE1PE3PETime (ms)
basic0.7410.862.3030
+v_Proj0.7510.902.3530
circular0.7611.252.4430
", + "bbox": [ + 88, + 196, + 472, + 271 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/91450d64bd7e52d1fd6f36f751ba9bea71760a969f26434b948f0687f87533b2.jpg", + "table_caption": [ + "(c) Pattern aggregation module design." + ], + "table_footnote": [ + "(e) Iterations. TMA achieves good results with a few iterations." + ], + "table_body": "
iterationsEPE1PE3PETime (ms)
20.8311.042.3710
40.7611.452.4316
6/80.7410.862.3030
0.7410.812.3254
", + "bbox": [ + 86, + 301, + 470, + 393 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/04d50b167edbeda2715247b1aa4f2d38ca1d20c0c5df48ea3ae54815729c5ff3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
styleEPE1PE3PE
w/o0.7812.002.66
linear0.7410.862.30
same0.7611.452.48
", + "bbox": [ + 491, + 88, + 879, + 165 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/06c5d04f2929d106dea5f3a21455b1508f20f36bc4f943fb697fbf6f9afe2cca.jpg", + "table_caption": [ + "(b) Lookup style. Sampling correlation volumes with a set of linear coordinates improves the accuracy." + ], + "table_footnote": [], + "table_body": "
layersEPE1PE3PETime (ms)
00.7511.362.3320
1/20.7410.862.3030
0.7510.902.3238
", + "bbox": [ + 491, + 196, + 877, + 271 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/43d2044be5526ce3f3559a5f67c56a119d9dbacb6177299752ff814134cc6270.jpg", + "table_caption": [ + "(d) Layers of pattern aggregation module. Stacking 1 pattern aggregation module showcases high accuracy." + ], + "table_footnote": [], + "table_body": "
radiusEPE1PE3PE
10.7511.142.37
20.7611.162.39
3/40.7410.862.30
0.7510.932.30
", + "bbox": [ + 491, + 303, + 874, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(f) Searching radius. A radius of 3 leads to high accuracy.", + "bbox": [ + 509, + 395, + 859, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: TMA ablations. Models are trained on DSEC-Flow. Settings used in our final model are underlined.", + "bbox": [ + 122, + 412, + 844, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 76, + 436, + 187, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct a series of ablation experiments to validate the proposed improvements. All ablation models are trained on DSEC-Flow and evaluated on the public benchmark.", + "bbox": [ + 75, + 460, + 468, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Event splitting. We compare different numbers of event segments in Table 4a. We choose different splits in 1, 3, 5 and 7. With the split set to 1, our model degrades into a classic correlation volume based framework and exhibits similar accuracy in comparison with E-RAFT. More splits lead to higher accuracy but lower inference efficiency, so we use 5 splits to achieve accuracy-efficiency trade-off.", + "bbox": [ + 75, + 506, + 468, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lookup style. Based on the assumption that the optical flow changes linearly over a short period, we propose the linear lookup strategy. To validate it, we try to cancel the lookup operation or just use same sampling coordinates on intermediate correlation volumes. As Table 4b shows, sampling the correlation maps with a set of linear coordinates leads to high accuracy.", + "bbox": [ + 75, + 612, + 468, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pattern aggregation module design. We ablate different components of pattern aggregation module in Table 4c. By default, we set the projection matrix of value $W^V$ as an identity matrix because the motion features are encoded by a motion feature encoder beforehand. In addition, we replace the pairing mode of motion features with circular pairing. Circular pairing gives equal treatment to all motion features without highlighting the importance of the last motion feature, which fails to improve accuracy.", + "bbox": [ + 75, + 718, + 468, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Layers of pattern aggregation module. We compare the accuracy with different layers of pattern aggregation modules in Table 4d. We note that stacking 1 layer of pat", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tern aggregation module showcases high accuracy. More layers may lead to difficulties in optimization, which in turn causes a decrease in accuracy.", + "bbox": [ + 496, + 436, + 890, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Iterations. Iterative learning plays a role in numerous works [18, 30]. Results in Table 4e show that 6 iterative updates reach the sweet spot of accuracy and efficiency.", + "bbox": [ + 496, + 483, + 890, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Searching radius. We compare different radii for looking up correlation volumes in Table 4f. Searching radius of 3 leads to high accuracy. Different from RAFT [30] that achieves better predictions with an increased radius, our model does not benefit a lot from a greater radius.", + "bbox": [ + 496, + 527, + 890, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 618, + 617, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we delve into event-based optical flow estimation problem. Different from existing methods that represent the event stream as static frames to adopt the classic correlation volume based pipeline, we contend that the temporal continuity is a vital element of event-based optical flow. We propose a novel Temporal Motion Aggregation (TMA) approach to unlock its potential within a learning-based framework. By incorporating temporally fine-grained motion information, TMA generates high-quality flow estimates at early stages, which not only enables TMA to obtain accurate final predictions, but also greatly reduces the demand for a number of refinements. Extensive experiments on DESC-Flow and MVSEC datasets verify the effectiveness and superiority of TMA. Meanwhile, we notice the limitations of supervised learning in generalization ability and expect further improvement to extend our idea into unsupervised learning frameworks.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9692", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This work was supported in part by Shanghai Municipal Science and Technology Major Project under Grant 2018SHZDZX01, ZJ Lab, the Shanghai Center for Brain Science and Brain-Inspired Technology, in part by Shanghai Rising Star Program under Grant 21QC1400900, in part by Tongji-Qomolo Autonomous Driving Commercial Vehicle Joint Lab Project, and in part by Xiaomi Young Talents Program.", + "bbox": [ + 75, + 90, + 472, + 203 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 229, + 173, + 244 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Himanshu Akolkar, Sio-Hoi Ieng, and Ryad Benosman. Real-time high speed motion prediction using fast aperture-robust event-driven visual flow. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):361-372, 2020. 2", + "[2] Mohammed Almatrafi, Raymond Baldwin, Kiyoharu Aizawa, and Keigo Hirakawa. Distance surface for event-based optical flow. IEEE transactions on pattern analysis and machine intelligence, 42(7):1547-1556, 2020. 2", + "[3] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE transactions on neural networks and learning systems, 25(2):407-417, 2013. 2", + "[4] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, 2012. 2", + "[5] Vincent Brebion, Julien Moreau, and Franck Davoine. Realtime optical flow for vehicular perception with low-and high-resolution event cameras. IEEE Transactions on Intelligent Transportation Systems, 2021. 7", + "[6] Tobias Brosch, Stephan Tschechne, and Heiko Neumann. On event-based optical flow detection. Frontiers in neuroscience, 9:137, 2015. 2", + "[7] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-temporal recurrent networks for event-based optical flow estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 525-533, 2022. 3, 5, 7", + "[8] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1", + "[9] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3867-3876, 2018. 2, 4", + "[10] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 2, 3, 5", + "[11] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cam" + ], + "bbox": [ + 78, + 255, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "eras. In 2021 International Conference on 3D Vision (3DV), pages 197-206. IEEE, 2021. 1, 2, 3, 4, 5, 6, 7", + "[12] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from events and frames. arXiv preprint arXiv:2203.13674, 2022. 3", + "[13] Jesse Hagenaars, Federico Paredes-Vallés, and Guido De Croon. Self-supervised learning of event-based optical flow with spiking neural networks. Advances in Neural Information Processing Systems, 34:7167-7179, 2021. 7", + "[14] Markus Hofinger, Samuel Rota Bulò, Lorenzo Porzi, Arno Knapitsch, Thomas Pock, and Peter Kontschieder. Improving optical flow on a pyramid level. In European Conference on Computer Vision, pages 770-786. Springer, 2020. 2", + "[15] Tak-Wai Hui and Chen Change Loy. Liteflownet3: Resolving correspondence ambiguity for more accurate optical flow estimation. In European Conference on Computer Vision, pages 169-184. Springer, 2020. 2", + "[16] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8981-8989, 2018. 2", + "[17] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. A lightweight optical flow cnn—revisiting data fidelity and regularization. IEEE transactions on pattern analysis and machine intelligence, 43(8):2555-2569, 2020. 2", + "[18] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2019. 1, 2, 3, 8", + "[19] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9772-9781, 2021. 2, 5", + "[20] Chankyu Lee, Adarsh Kumar Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, and Kaushik Roy. Spikeflownet: event-based optical flow estimation with energy-efficient hybrid neural networks. In European Conference on Computer Vision, pages 366-382. Springer, 2020. 7", + "[21] Min Liu and Tobi Delbruck. Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. BMVC, 2018. 2", + "[22] Min Liu and Tobi Delbruck. Edflow: Event driven optical flow camera with keypoint detection and adaptive block matching. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5776-5789, 2022. 2", + "[23] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2017. 6", + "[24] Federico Paredes-Vallés and Guido CHE de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446-3455, 2021. 1, 3", + "[25] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4161-4170, 2017. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9693", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3", + "[27] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In European Conference on Computer Vision, pages 628-645. Springer, 2022. 2, 5, 6, 7", + "[28] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications, volume 11006, page 1100612. International Society for Optics and Photonics, 2019. 6", + "[29] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018. 1, 2", + "[30] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 1, 2, 3, 5, 8", + "[31] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 1", + "[32] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 1", + "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5", + "[34] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing, 31:7237-7251, 2022. 1, 3, 4, 6, 7", + "[35] Haofei Xu, Jiaolong Yang, Jianfei Cai, Juyong Zhang, and Xin Tong. High-resolution optical flow from 1d attention and correlation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10498-10507, 2021. 2", + "[36] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 2, 6", + "[37] Gengshan Yang and Deva Ramanan. Volumetric correspondence networks for optical flow. Advances in neural information processing systems, 32:794-805, 2019. 2", + "[38] Chengxi Ye, Anton Mitrokhin, Cornelia Fermüller, James A Yorke, and Yiannis Aloimonos. Unsupervised learning of" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dense optical flow, depth and egomotion with event-based sensors. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5831-5838. IEEE, 2020. 1, 2, 7", + "[39] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 185-194, 2019. 2", + "[40] Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In European Conference on Computer Vision, pages 420-439. Springer, 2020. 2", + "[41] Feihu Zhang, Oliver J Woodford, Victor Adrian Prisacariu, and Philip HS Torr. Separable flow: Learning motion cost volumes for optical flow estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10807-10817, 2021. 2", + "[42] Xu Zheng, Yexin Liu, Yunfan Lu, Tongyan Hua, Tianbo Pan, Weiming Zhang, Dacheng Tao, and Lin Wang. Deep learning for event-based vision: A comprehensive survey and benchmarks. arXiv preprint arXiv:2302.08890, 2023. 1", + "[43] Alex Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In Proceedings of Robotics: Science and Systems, Pittsburgh, Pennsylvania, June 2018, 2, 5, 6, 7", + "[44] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018. 2, 3, 5, 7", + "[45] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 1, 2, 3, 4, 5, 7" + ], + "bbox": [ + 501, + 92, + 893, + 613 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9694", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/9d269c18-5110-4a81-a1b9-317c1455548e_model.json b/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/9d269c18-5110-4a81-a1b9-317c1455548e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6144278292b461c2166b0a0faef15c82c7ac3b4a --- /dev/null +++ b/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/9d269c18-5110-4a81-a1b9-317c1455548e_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.131, + 0.825, + 0.154 + ], + "angle": 0, + "content": "TMA: Temporal Motion Aggregation for Event-based Optical Flow" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.179, + 0.77, + 0.234 + ], + "angle": 0, + "content": "Haotian Liu1 Guang Chen1* Sanqing Qu1 Yanping Zhang1 \nZhijun Li1 Alois Knoll2 Changjun Jiang1 \n1Tongji University 2Technical University of Munich" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.725 + ], + "angle": 0, + "content": "Event cameras have the ability to record continuous and detailed trajectories of objects with high temporal resolution, thereby providing intuitive motion cues for optical flow estimation. Nevertheless, most existing learning-based approaches for event optical flow estimation directly remould the paradigm of conventional images by representing the consecutive event stream as static frames, ignoring the inherent temporal continuity of event data. In this paper, we argue that temporal continuity is a vital element of event-based optical flow and propose a novel Temporal Motion Aggregation (TMA) approach to unlock its potential. Technically, TMA comprises three components: an event splitting strategy to incorporate intermediate motion information underlying the temporal context, a linear lookup strategy to align temporally fine-grained motion features and a novel motion pattern aggregation module to emphasize consistent patterns for motion feature enhancement. By incorporating temporally fine-grained motion information, TMA can derive better flow estimates than existing methods at early stages, which not only enables TMA to obtain more accurate final predictions, but also greatly reduces the demand for a number of refinements. Extensive experiments on DSEC-Flow and MVSEC datasets verify the effectiveness and superiority of our TMA. Remarkably, compared to E-RAFT, TMA achieves a \\(6\\%\\) improvement in accuracy and a \\(40\\%\\) reduction in inference time on DSEC-Flow. Code will be available at https://github.com/ispc-lab/TMA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.745, + 0.21, + 0.76 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.771, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Optical flow aims to compute the velocity of objects on the image plane without geometry prior and is a fundamental topic in event-based vision [8, 42]. It plays an important role in many applications, such as ego-motion estimation [38, 45], image reconstruction [24], and video frame interpolation [32, 31]. Recently, learning-based methods [29, 18, 30] have dominated frame-based optical flow" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.259, + 0.696, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.348, + 0.656, + 0.36 + ], + "angle": 0, + "content": "(a) reference image" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.259, + 0.886, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.348, + 0.834, + 0.359 + ], + "angle": 0, + "content": "(b) event frame" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.359, + 0.696, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.448, + 0.696, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.359, + 0.886, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.448, + 0.885, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.534, + 0.876, + 0.546 + ], + "angle": 0, + "content": "(c) E-RAFT output & motion feature (d) our output & motion feature" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Figure 1: Accuracy illustrations. (a) Reference image from DSEC-Flow. (b) Corresponding event frame. (c) The final flow prediction and visualization of the motion feature from E-RAFT [11]. (d) The final flow prediction and visualization of the motion feature from our proposed method. We visualize the motion feature at the initial stage by taking the average across channels. E-RAFT fails to generate informative motion features and results in blurred boundaries. In contrast, our method utilizes temporally fine-grained motion information to address the information scarcity issue in motion features, generating high-quality predictions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "by employing correlation volumes to derive motion features for flow regression. Inspired by this, several event-based approaches [11, 34] adopt a similar paradigm by converting the consecutive event stream into grid-like representations. Despite encouraging progress, these methods still suffer from non-negligible limitations, due to the great differences between event data and conventional images. Distinct from dense and colorful conventional (RGB) images, event data features spatial sparsity and a lack of intensity information. It is prone to result in close matching scores and invalid regions in the correlation volume. Consequently, less informative motion features are derived and regress inaccurate" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.361, + 0.901 + ], + "angle": 0, + "content": "*corresponding author: guangchen@tongii.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9685" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.465, + 0.107 + ], + "angle": 0, + "content": "predictions. Figure 1 illustrates this issue in E-RAFT [11]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.468, + 0.273 + ], + "angle": 0, + "content": "As the saying goes, each coin has two sides. Though event cameras encode visual information in a sparse manner, they are capable of capturing continuous and detailed object trajectories with high temporal resolution, thereby providing rich motion cues for optical flow estimation. Several model-based methods [9, 27] benefit from the temporal continuity by warping events along point trajectories, achieving decent performance. Drawing on this observation, we contend that the temporal continuity is a vital element of event-based optical flow and propose to unleash its potential within a learning-based framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.47, + 0.681 + ], + "angle": 0, + "content": "To materialize our idea, we propose a novel Temporal Motion Aggregation (TMA) approach to explore the inherent temporal continuity of event data. Technically, we revamp the dominant learning-based architectures with three distinct components. We first introduce an event splitting strategy for temporally-dense correlation volumes computation. By splitting the event stream into multiple segments and extracting their features, feature similarities are compared between the first feature and all others, which record rich intermediate motion information. Aware of that correlation volumes record motions between different time spans, we then design a linear lookup strategy to sample each correlation volume based on the corresponding flow estimate to encode motion features. As a result, fine-grained motions are warped to the same pixels across motion features. Furthermore, we consider the incorrect motion patterns in intermediate motion features due to the manual lookup. Therefore, we propose a novel pattern aggregation module for motion feature enhancement by aggregating consistent patterns between each intermediate motion feature and the last one (without manual lookup). Thanks to the incorporation of temporally fine-grained motion information, TMA can effectively address the information scarcity in motion features and thus generate good flow estimates at early stages, which not only enable TMA to obtain accurate flow predictions, but also greatly reduces the demand for a number of time-consuming refinements." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.682, + 0.47, + 0.727 + ], + "angle": 0, + "content": "We evaluate our TMA on DSEC-Flow [10] and MVSEC [44] datasets. Extensive experiments demonstrate the superior advantages of TMA." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.728, + 0.434, + 0.743 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.75, + 0.469, + 0.794 + ], + "angle": 0, + "content": "- We argue that the temporal continuity of event data is a vital element of event-based optical flow and propose to unlock its potential in a learning-based framework." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.803, + 0.469, + 0.863 + ], + "angle": 0, + "content": "- We propose a novel Temporal Motion Aggregation (TMA) approach, which comprises three components: an event splitting strategy, a linear lookup strategy and a motion pattern aggregation module." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- The incorporation of temporally fine-grained motion information enables TMA to achieve high accuracy" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "while maintaining high efficiency. Compared with E-RAFT, TMA achieves a \\(6\\%\\) improvement in accuracy and a \\(40\\%\\) reduction in inference time on DSEC-Flow." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.151, + 0.642, + 0.166 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.177, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Learning methods for optical flow estimation. Recently, correlation volume based approaches dominate the learning-based optical flow estimation [25, 17, 29, 18, 37, 14]. As a milestone, PwCNet [29] summarizes a simple principle in architecture construction: pyramidal processing, warping and correlation volumes computation. Following the well-established pipeline, LiteFlowNet series [16, 17, 15] propose cascaded feature pyramids with a lightweight convolutional burden. VCN [37] proposes 4D volumetric U-Net architectures to enlarge the receptive fields for correlation volumes. IRR [18] notes the reusability of pyramidal refinement for flow update and proposes a shared-weights refiner across all scales, which can be combined with mainstream learning methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.614 + ], + "angle": 0, + "content": "However, the correlation volumes used in previous works are computed locally, which cannot fully deal with long-term motions. RAFT [30] firstly introduces the full correlation volume to capture global visual similarities. Separable Flow [41] proposes a separable correlation volume module as a drop-in replacement to correlation volumes and uses non-local aggregation layers [39, 40] to refine the correlation volumes efficiently. Besides, transformer components show promising performance on optical flow [36, 19, 35]. GMA [19] notices the underlying motion clues in the context and introduces an attention layer to aggregate hidden motions, improving the accuracy based on RAFT architecture. GMFlow [36] reformulates optical flow as a global matching problem by attention blocks, suggesting a new paradigm for flow estimation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.892, + 0.703 + ], + "angle": 0, + "content": "We draw inspiration from GMA [19] to design our motion pattern aggregation module. In contrast to GMA which aggregates hidden motion patterns in the spatial context for motion feature enhancement, we extend the pattern aggregation to the temporal dimension and leverage feature cross-similarities to enhance the motion feature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Event-based optical flow estimation. A number of traditional works remould frame-based approaches into event-based optical flow estimation. Benosman et al. [4] extends the classic Lucas-Kanade method to event-based optical flow. Liu et al. [21, 22] implement event-based corner keypoint detection and novel block-matching optical flow. Gallego et al. [9] propose a unifying contrast maximization (CM) framework and MultiCM [27] extends it to complex scenes. Besides, plane fitting [1, 3], filter banks [6] and time surface matching [2] are proposed to boost performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "More recently, learning-based methods have dominated frame-based optical flow, inspiring attempts at event-based optical flow. Early learning methods [43, 45, 38] represent" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9686" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.885, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.242, + 0.895, + 0.35 + ], + "angle": 0, + "content": "Figure 2: The overall architecture of TMA. First, we split the event stream into segments and extract features with a share-weights feature extractor. Then, regarding the first feature as reference, we correlate all other features to compute temporally-dense correlation volumes, which contain rich embedded motion information. Next, aware of the different time spans of correlation volumes, we look up each correlation volume based on a corresponding flow estimate. All sampled correlation maps are delivered into a share-weights motion feature encoder to derive motion features. We further propose a novel motion pattern aggregation module to enrich the spatial information of motion features. By concatenating the enhanced motion features and utilizing several refinements, high-quality flow predictions are generated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.473, + 0.569 + ], + "angle": 0, + "content": "the event stream as a static frame and utilize the architecture of U-Net [26] to predict optical flow, which can only address small motions. Mainstream works [7, 34, 11] introduce two consecutive event frames and follow the correlation volume based pipeline (PWCNet and RAFT), achieving higher accuracy on large-scales datasets [44, 10]. E-RAFT [11, 12] extends RAFT [30] and introduces full correlation volume into event-based vision. DCEIFlow [34] combines conventional images and event data to boost higher accuracy. Note that our work and STE-FlowNet [7] share a similar motivation to introduce temporal continuity with fine-grained discretization. However, STE-FlowNet takes a frame-by-frame manner [18] to update flow, leading to an insufficient inference process." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.57, + 0.471, + 0.661 + ], + "angle": 0, + "content": "Contrary to previous learning-based methods, we do not merely adopt the classic correlation volume based pipeline. Instead, we regard the temporal continuity of event data as a vital component of optical flow and incorporate the intermediate motion information into a learning-based framework, which boosts higher accuracy and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.214, + 0.696 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Given an input event stream \\(\\{\\mathcal{E}\\}\\) from \\(t_0\\) to \\(t_1\\), the goal of optical flow is to estimate a dense displacement field \\(\\mathbf{u}:\\mathbb{R}^2\\to \\mathbb{R}^2\\) which maps each per-pixel \\((x,y)\\) at \\(t_0\\) to its corresponding coordinate \\((x^{\\prime},y^{\\prime})\\) at \\(t_1\\). To address this, mainstream learning-based approaches compute correlation volumes from two event frames as a matching prior. Such principle inspires several methods [11, 34] by squeezing event data into consecutive frames. Unfortunately, the characteristics of event data, spatial sparsity and lack of intensity records, are prone to result in close matching scores and invalid regions in the correlation volume. Consequently, less informative motion features are derived and regress inaccurate predictions. Here we address the information scarcity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.417 + ], + "angle": 0, + "content": "of motion features from the attractive temporal continuity of event data. Temporally fine-grained motion information in event data is sufficient to supplement the spatial richness of motion features, yielding high-quality flow predictions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.463 + ], + "angle": 0, + "content": "In the following, we first describe the preliminary and event representation, and then present the detailed architecture of the proposed TMA." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.471, + 0.833, + 0.487 + ], + "angle": 0, + "content": "3.1. Preliminary and Event Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.6 + ], + "angle": 0, + "content": "Event cameras record light intensity changes in an asynchronous way. Each pixel serves as an independent trigger and generates an event instantly whenever the log intensity changes exceeding a threshold \\( C \\). An event \\( \\boldsymbol{e}_k = (\\boldsymbol{x}_k, p_k, t_k) \\) is a triplet containing triggered pixel location \\( \\boldsymbol{x}_k = (x_k, y_k)^T \\), timestamp \\( t_k \\) and signed intensity change \\( (p_k = \\pm 1) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.676 + ], + "angle": 0, + "content": "To ease the feature extraction and preserve the temporal continuity of the event data, we firstly transform the event stream into a 3D volume \\( V(x,y,t) \\) along the temporal dimension following existing works [45, 11, 24]. Given an event stream \\( \\{\\mathcal{E}\\} \\), \\( V(x,y,t) \\) is generated as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.684, + 0.891, + 0.701 + ], + "angle": 0, + "content": "\\[\nt _ {i} ^ {*} = (B - 1) \\left(t _ {i} - t _ {1}\\right) / \\left(t _ {N _ {e}} - t _ {1}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.704, + 0.891, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {V} (x, y, t) = \\sum_ {i} p _ {i} k _ {b} \\left(x - x _ {i}\\right) k _ {b} \\left(y - y _ {i}\\right) k _ {b} \\left(t - t _ {i} ^ {*}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.739, + 0.891, + 0.755 + ], + "angle": 0, + "content": "\\[\nk _ {b} (a) = \\max (0, 1 - | a |), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.81 + ], + "angle": 0, + "content": "where \\( t_i^* \\) represents the \\( i \\)-th time bin. \\( B \\) and \\( N_e \\) represent the length of time bins and the number of events, respectively. \\( k_b(a) \\) is the bilinear interpolation function." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.681, + 0.832 + ], + "angle": 0, + "content": "3.2. TMA Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We develop our framework based on the successful RAFT architecture [30]. Figure 2 summarizes the architecture of the proposed TMA, which consists of three components, an event splitting strategy, a linear lookup strategy" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9687" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.088, + 0.445, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.249, + 0.47, + 0.325 + ], + "angle": 0, + "content": "Figure 3: Linear lookup strategy. We compute correlation volumes by comparing the similarity of first feature with all others, which are misaligned due to different time spans. To rectify this, we narrow down the lookup coordinates linearly to warp correlation maps to spatially aligned pixels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.469, + 0.392 + ], + "angle": 0, + "content": "and a motion pattern aggregation module. With the proposed method, we successfully unlock the potential of temporal continuity to predict event-based optical flow within a learning-based framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.469, + 0.558 + ], + "angle": 0, + "content": "Event splitting. Instead of viewing the continuous event stream as a static frame which tends to be a matter of common knowledge in recent works [45, 11, 34], we prefer to identify it as a high frame rate video clip, which records continuous and detailed trajectories of objects, namely, provides intuitive motion cues for optical flow estimation. To introduce rich intermediate motion information into our method, we first propose an event splitting strategy, which splits the event stream into multiple segments. Therefore, more intermediate event frames are retained for correlation volume computation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.56, + 0.469, + 0.726 + ], + "angle": 0, + "content": "In particular, given an event stream \\(\\{\\mathcal{E}\\}\\) from \\(t_0\\) to \\(t_1\\), we split it into \\(g\\) segments equally and the time span between adjacent segments is \\(dt\\). Besides, we introduce an auxiliary event segment from \\(t_0 - dt\\) to \\(t_0\\) as reference. Then, \\(g + 1\\) event representations are generated following Sec. 3.1. All \\(g + 1\\) event representations are then delivered into a share-weights encoder to achieve features \\(\\pmb{F}_i \\in \\mathbb{R}^{H \\times W \\times C}, i = 0, 1, 2, \\dots, g,\\) where \\(H, W\\) and \\(D\\) represent the height, width and channel of features. We use all features to compute temporally-dense correlation volumes \\(\\pmb{C}_i, i = 1, 2, \\dots, g,\\) which are given by:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.736, + 0.469, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {C} _ {i} = \\frac {\\boldsymbol {F} _ {0} \\boldsymbol {F} _ {i} ^ {T}}{\\sqrt {D}} \\in \\mathbb {R} ^ {H \\times W \\times H \\times W}, i = 1, 2, \\dots , g, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.469, + 0.809 + ], + "angle": 0, + "content": "where \\( \\pmb{F}_0 \\) is the first feature and \\( \\pmb{F}_i \\) is the \\( i \\)-th feature derived from the \\( i \\)-th event frame." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.469, + 0.885 + ], + "angle": 0, + "content": "Temporally-dense correlation volumes compare feature similarities across multiple time spans, enabling refined records of relative motions. This construction is vastly different from the classic full correlation volume that only contains feature similarities between a fixed time span." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Linear lookup. Recall that each correlation volume \\( C_i \\)" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.091, + 0.862, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.346 + ], + "angle": 0, + "content": "Figure 4: Motion pattern aggregation. We pair each intermediate motion feature to the last motion feature. Then each pair is sent to cross-attention blocks to aggregate consistent patterns. Hence, each intermediate motion feature carries information from the last one and consistent patterns receive emphasis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.488 + ], + "angle": 0, + "content": "is achieved by correlating \\( F_{0} \\) and \\( F_{i} \\), which means the recorded relation motions are misaligned in time. While our goal is to predict optical flow from \\( t_{0} \\) to \\( t_{1} \\), we hope that all sampled correlation maps could present the relative motions within the same time span. Therefore, based on that correlation volumes are evenly constructed in time and magnitudes of optical flow change linearly over a short period, we narrow down lookup coordinates in a linear fashion to match the timestamp of each correlation volume." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.548 + ], + "angle": 0, + "content": "Figure 3 displays the linear lookup strategy. In detail, given current flow estimate \\(\\mathbf{u}\\), we split it into \\(g\\) segments to sample each correlation volume. The linear lookup operation can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.557, + 0.891, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} \\mathbf {u} = \\left(\\mathbf {c o o r d s} _ {1} - \\mathbf {c o o r d s} _ {0}\\right) / g = \\mathbf {u} / g \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.575, + 0.891, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\operatorname {c o r r} _ {i} = S \\left(\\boldsymbol {C} _ {i}, \\operatorname {c o o r d s} _ {0} + i \\mathrm {d u}\\right), i = 1, 2, \\dots , g, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.891, + 0.688 + ], + "angle": 0, + "content": "where \\(\\mathbf{coord}_0\\) is the initialized coordinates at \\(t_0\\) (reference time), \\(S(\\cdot, \\cdot)\\) represents the lookup operation, \\(\\mathbf{coord}_i\\) are lookup coordinates for \\(C_i\\) and \\(\\mathbf{corr}_i\\) is the correlation map sampled from \\(C_i\\). After delivering all correlation maps into a share-weights motion feature encoder, aligned motion features are achieved." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.795 + ], + "angle": 0, + "content": "From the perspective of image contrast [9], the standard lookup is responsible for warping possible matching points between two images to the same pixel. Our linear lookup extends it by warping matching points of a temporally fine-grained motion to the same pixel. As a result, similar motions are presented at same pixels across correlation maps, showing high contrast." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Motion pattern aggregation. A standard operation for next step is to concatenate the motion features as a whole to update flow and the improvements made so far already alleviate the information scarcity in motion features. However, we notice that the concatenation operation treats all motion features equally, which means the majority of motion features, intermediate ones, dominate the flow regression. Due" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9688" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.087, + 0.38, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.089, + 0.635, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.09, + 0.885, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.893, + 0.33 + ], + "angle": 0, + "content": "Figure 5: Qualitative examples of DSEC-Flow. For each two columns, we show the reference image and event frame and compare our method with a state-of-the-art baseline E-RAFT [11]. Significant improvements are highlighted by red boxes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.47, + 0.462 + ], + "angle": 0, + "content": "to the linear optical flow assumption and manual lookup operation, these motion features are not perfectly correct. A natural thought is to put more emphasis on the last motion feature that is without manual lookup and retain the patterns in intermediate motion features that are consistent with the last motion feature. To this end, we balance the weights between the last motion feature and others through cross-similarities computed by transformers [33, 19]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.463, + 0.471, + 0.569 + ], + "angle": 0, + "content": "Figure 4 shows the overview of motion pattern aggregation module. Given a set of motion features \\( M\\mathbf{F}_i \\in \\mathbb{R}^{H \\times W \\times D'} \\), \\( i = 1,2,\\dots,g \\), we pair each intermediate motion feature to the last one. Then each pair \\( (M\\mathbf{F}_i,M\\mathbf{F}_g) \\) is sent to cross-attention blocks to compare cross-similarities and aggregate consistent patterns. The enhanced motion feature \\( \\widehat{M}\\mathbf{F}_i \\) is obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.58, + 0.469, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {A} = \\operatorname {A t t e n t i o n} \\left(\\boldsymbol {M} \\boldsymbol {F} _ {i} W ^ {Q}, \\boldsymbol {M} \\boldsymbol {F} _ {g} W ^ {K}, \\boldsymbol {M} \\boldsymbol {F} _ {g} W ^ {V}\\right) (7) \\\\ \\widehat {\\boldsymbol {M F}} _ {i} = \\boldsymbol {M F} _ {i} + \\operatorname {M L P} \\left(\\left[ \\boldsymbol {M F} _ {i}, \\boldsymbol {A W} ^ {O} \\right]\\right), (8) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.469, + 0.748 + ], + "angle": 0, + "content": "where \\( \\operatorname{Attention}(Q, K, V) = \\operatorname{softmax}\\left(\\frac{QK^T}{\\sqrt{d_k}}\\right)V \\), \\( d_k \\) is the dimension of \\( K \\). \\( W^Q \\), \\( W^K \\), \\( W^V \\) are projection matrices. \\( \\mathrm{MLP}(\\cdot) \\) is multi-layer perceptron and \\( W^O \\) is another projection matrix. We omit the normalization functions in the manuscript for simplicity. Motion pattern aggregation module does not change the dimension of motion features, so multiple cross-attention layers can be stacked easily." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.825 + ], + "angle": 0, + "content": "Once motion features are enhanced by the motion pattern aggregation module, each \\(\\widehat{MF}_i\\) carries motion patterns that are consistent with \\(MF_{g}\\). Namely, the last motion feature receives high attention and consistent patterns in motion features are fully utilized." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Above three components are seamlessly integrated into a unified pipeline: incorporating intermediate motion information, warping matching motions and emphasizing consistent motion patterns. Attributed to the innovative application of temporal continuity to a learning-based frame" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.432 + ], + "angle": 0, + "content": "work, our proposed method is capable of overcoming the information scarcity issue in motion features and generating high-quality flow estimates even at an early refinement stage. This enables our method to achieve accurate final predictions, while reducing the demand for a number of refinements." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.441, + 0.628, + 0.457 + ], + "angle": 0, + "content": "3.3. Supervision" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.464, + 0.892, + 0.51 + ], + "angle": 0, + "content": "Following RAFT [30], we choose \\( L_{1} \\) distance between the predictions and ground-truth as the supervision signal for our model. The loss function \\( \\mathcal{L} \\) is formally defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.52, + 0.892, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {j = 1} ^ {N} \\gamma^ {N - j} \\left\\| \\mathbf {u} ^ {g t} - \\mathbf {u} ^ {j} \\right\\| _ {1}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.891, + 0.621 + ], + "angle": 0, + "content": "where \\(\\mathbf{u}^{gt}\\) denotes the flow ground-truth and \\(\\mathbf{u}^j\\) the flow prediction at \\(j\\)-th stage. \\(N\\) is the total refinement stages. The weight \\(\\gamma\\) balances different stages of predictions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.634, + 0.651 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets and evaluation setup. Following previous works [11, 45, 7], we provide experimental details and extensive comparison results on two event-based datasets DSEC-Flow [10] and MVSEC [44]. For DSEC-Flow, we train models on the official training set and evaluate DSEC-Flow test set on the public benchmark. For MVSEC, we conduct two kinds of experiments. Basically, following previous methods [43, 27], we train methods on outdoor_day2 sequence with ground-truth corresponding to a time interval of \\( dt = 1 \\) and \\( dt = 4 \\) gray images and evaluate 800 frames of outdoor_day1 and three indoor_flying sequences. Equally importantly, considering the supervised learning setting of our method and the large gap between outdoor_day and indoor_flying data, we extend the range of training data by including one indoor_flying sequence each time to further boost the accuracy of supervised learning methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9689" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.465, + 0.175 + ], + "angle": 0, + "content": "
MethodsEPE1PE3PETime(ms)
MBMultiCM [27]3.4776.630.9-
SLEV-FlowNet†[43]2.3255.418.65
E-RAFT [11]0.7912.52.752
Ours0.7410.92.330
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.181, + 0.47, + 0.243 + ], + "angle": 0, + "content": "Table 1: Results on DSEC-Flow. Best accuracy in bold. \\(\\dagger\\) denotes the result is taken from E-RAFT [11]. Model-based (MB) methods need no training data; and supervised learning (SL) methods need ground-truth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.414 + ], + "angle": 0, + "content": "Metrics. The end-point-error (EPE) is used as core metric for prediction accuracy for both DSEC-Flow and MVSEC. To evaluate the robustness against large displacements, we compute the percentage of pixels with EPE greater than \\(M\\) pixels \\((M\\mathrm{PE},M = 1,3)\\) for DSEC-Flow. Percentage of pixels with EPE above 3 and \\(5\\%\\) of the magnitude of the flow ground-truth (\\% Outlier) are presented for MVSEC. On DSEC-Flow, metrics are measured over pixels with valid ground-truth. While on MVSEC, metrics are measured over pixels with valid ground-truth and at least one triggered event." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.471, + 0.657 + ], + "angle": 0, + "content": "Implementation details. The proposed TMA is implemented using PyTorch. We set the event segment \\( g = 5 \\). For experiments on DSEC-Flow, the channel of event representation \\( B \\) for each segment is set to 3. For MVSEC, \\( B \\) is set to 1 when \\( dt = 1 \\) and 3 when \\( dt = 4 \\), respectively. The sum of channels keeps the same with the state-of-the-art E-RAFT [11]. We only stack 1 layer of motion pattern aggregation module. Other components are basically identical to RAFT's model, including feature extractor, motion feature encoder and flow updater. Note that we use a small feature extractor whose final channel is 128. We predict 6 levels of optical flow and the weight \\( \\gamma \\) is set to 0.8 for supervision. In training, we exploit AdamW [23] optimizer and One-cycle learning rate scheduler [28]. We train models on DSEC-Flow for 200K steps and MVSEC for 100K steps, both with a learning rate of 0.0002 and a batch size of 6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.206, + 0.68 + ], + "angle": 0, + "content": "4.1. DSEC-Flow" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.764 + ], + "angle": 0, + "content": "Accuracy comparison. Evaluation results on DSEC-Flow benchmark1 are provided in Table 1. Compared with E-RAFT, our method improves the EPE from 0.79 to 0.74. Our method is also leading in 1PE and 3PE with a margin of 1.6 and 0.4, showing better robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.856 + ], + "angle": 0, + "content": "Qualitative results are exhibited in Figure 5. In challenging regions such as textureless areas (the wall in second two columns, the shrubs in last two columns), existing correlation volume contains similar matching scores and fails to generate informative motion features, leading to low-quality predictions. On the contrary, our method aggregates" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.091, + 0.88, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.334 + ], + "angle": 0, + "content": "Figure 6: EPE vs. number of iterations at inference. The figure exhibits the prediction results on DSEC-Flow at different stages during inference. Our method achieves an accuracy \\((\\mathrm{EPE} = 0.81)\\) comparable to E-RAFT with 2 iterations while E-RAFT needs 6 iterations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.384 + ], + "angle": 0, + "content": "temporally consistent motion patterns to enhance motion features, distinguishing different moving objects and generating high-quality predictions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.893, + 0.581 + ], + "angle": 0, + "content": "Efficiency comparison. The last column of Table 1 exhibits the inference time. Our method enjoys faster inference speed compared with E-RAFT, which is mainly attributed to fewer iterative updates. The reasonable explanation is that our method utilizes temporal motion information to enrich motion features, thus being able to generate good predictions even at an initial stage. To prove this, we compare the accuracy with E-RAFT at early inference stages in Figure 6. Our method achieves better flow initialization at early stages, supporting our argument. The ablation experiment on different iterations (Table 4e) also demonstrates that our method requires fewer iterations (ours: 6, E-RAFT: 12) to achieve surpassing accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.602, + 0.605 + ], + "angle": 0, + "content": "4.2.MVSEC" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Inter-domain evaluation. Table 2 reports the results on MVSEC training with outdoor_day2 sequence. The top of Table 2 reports the results corresponding to \\( dt = 1 \\) grayscale frame and the bottom corresponds to \\( dt = 4 \\) grayscale frames. Our method shows comparable accuracy on outdoor_day1 for \\( dt = 1 \\) and highest accuracy on outdoor_day1 for \\( dt = 4 \\) among all compared methods. Unfortunately, like two other supervised learning methods, DCEI-Flow [34] and E-RAFT [11], our method fails to achieve superior results on indoor_flying sequences. One possible reason is that model based methods and unsupervised learning methods usually have better generalization ability, while supervised learning methods are limited if training data has a significantly large gap with test data [36, 34]. Since outdoor_day sequences and indoor_flying sequences have a great domain gap, it is challenging for supervised learning methods to generalize well on indoor_flying sequences." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Intra-domain evaluation. To verify the possible cause and better analyze the accuracy of our method under su" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.469, + 0.901 + ], + "angle": 0, + "content": "1The ground-truth of DSEC-Flow dataset is unavailable, so we use last three checkpoints to evaluate the online benchmark. https://dsec.ifi.uzh.ch/uzh/dsec-flow-optical-flow-benchmark/." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9690" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.085, + 0.089, + 0.887, + 0.491 + ], + "angle": 0, + "content": "
Inputindoor_flying1indoor_flying2indoor_flying3outdoor_day1
EPE% OutlierEPE% OutlierEPE% OutlierEPE% Outlier
dt = 1
MBBrebion et al. [5]E0.520.100.985.500.712.100.530.20
MultiCM (Burgers') [27]E0.420.100.600.590.500.280.300.10
SSLEV-FlowNet [43]E1.032.201.7215.101.5311.900.490.20
Spike-FlowNet [20]E0.84-1.28-1.11-0.49-
STE-FlowNet [7]E0.570.100.791.600.721.300.420.00
USLHagenaars et al. [13]E0.600.511.178.060.935.640.470.25
Zhu et al. [45]E0.580.001.024.000.873.000.320.00
ECN [38]E------0.300.02
SLE-RAFT [11]E1.105.721.9430.791.6625.200.240.00
DCEIFlow [34]E + I0.751.550.902.100.801.770.220.00
OursE1.063.631.8127.291.5823.260.250.07
dt = 4
MBMultiCM (Burgers') [27]E1.6912.952.4926.352.0619.031.259.21
SSLEV-FlowNet [43]E2.2524.704.0545.303.4539.701.237.30
Spike-FlowNet [20]E2.24-3.83-3.18-1.09-
STE-FlowNet [7]E1.7714.702.5226.102.2322.100.993.90
USLZhu et al. [45]E2.1824.203.8546.803.1847.801.309.70
Hagenaars et al. [13]E2.1621.513.9040.723.0029.601.6912.50
SLDCEIFlow [34]E + I2.0821.473.4842.052.5129.730.893.19
E-RAFT [11]E2.8140.255.0964.194.4657.110.721.12
OursE2.4329.914.3252.743.6042.020.701.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.501, + 0.895, + 0.562 + ], + "angle": 0, + "content": "Table 2: Evaluation results on MVSEC [44] training with outdoor_day2 sequence. Model-based (MB) methods need no training data; semi-supervised learning (SSL) methods use grayscale images for supervision; unsupervised learning (USL) methods only require events; and supervised learning (SL) methods need ground-truth. E means the method only uses events as input; and I means the method requires images as input." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.575, + 0.885, + 0.717 + ], + "angle": 0, + "content": "
dt = 4Train Set +indoor_flying1indoor_flying2indoor_flying3outdoor_day1
EPE% OutlierEPE% OutlierEPE% OutlierEPE% Outlier
E-RAFT [11]indoor_flying1--2.3526.241.8418.070.721.38
Ours--2.1520.411.6814.300.640.98
E-RAFT [11]indoor_flying21.418.14--1.509.370.711.13
Ours1.325.94--1.508.880.640.96
E-RAFT [11]indoor_flying31.398.381.7717.52--0.721.46
Ours1.335.821.6712.80--0.671.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.718, + 0.893, + 0.751 + ], + "angle": 0, + "content": "Table 3: Evaluation results on MVSEC [44] training with outdoor_day2 and one indoor_flying sequence. + means the indoor_flying sequence to be included into training set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.758, + 0.473, + 0.895 + ], + "angle": 0, + "content": "pervised learning setting, we incorporate one indoor_flying sequence into training set each time and make comparisons with the supervised learning state-of-the-art E-RAFT. Table 3 shows the evaluation results corresponding to \\( dt = 4 \\) grayscale frames (For \\( dt = 1 \\), refer to appendix). Compared with the results in Table 2, E-RAFT and our method both achieve great improvement on indoor_flying sequences (EPE of indoor_flying2, ours: \\( 4.32 \\rightarrow 1.67 \\), E-RAFT: 5.09 → 1.77), confirming the presence of domain gap between" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.895, + 0.895 + ], + "angle": 0, + "content": "outdoor_day and indoor_flying data. Compared with the accuracy of E-RAFT with the same training set, on the one hand, our method shows better generalization ability, which generates lower EPE and \\(\\%\\) Outlier without seeing indoor_flying sequences. On the other hand, our method achieves higher accuracy on four sequences by incorporating indoor_flying sequences into training set. Overall, our method achieves a new state-of-the-art result, demonstrating the effectiveness of our proposed method." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "9691" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.473, + 0.18 + ], + "angle": 0, + "content": "
splitsEPE1PE3PETime (ms)
10.7911.912.7516
30.7610.982.4518
5/70.7410.862.3030
0.7411.132.2650
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.182, + 0.473, + 0.196 + ], + "angle": 0, + "content": "(a) Event splitting. Intermediate motion information contributes" + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.197, + 0.473, + 0.272 + ], + "angle": 0, + "content": "
setupEPE1PE3PETime (ms)
basic0.7410.862.3030
+v_Proj0.7510.902.3530
circular0.7611.252.4430
" + }, + { + "type": "table_caption", + "bbox": [ + 0.16, + 0.275, + 0.405, + 0.289 + ], + "angle": 0, + "content": "(c) Pattern aggregation module design." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.303, + 0.472, + 0.394 + ], + "angle": 0, + "content": "
iterationsEPE1PE3PETime (ms)
20.8311.042.3710
40.7611.452.4316
6/80.7410.862.3030
0.7410.812.3254
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.396, + 0.471, + 0.41 + ], + "angle": 0, + "content": "(e) Iterations. TMA achieves good results with a few iterations." + }, + { + "type": "table", + "bbox": [ + 0.493, + 0.089, + 0.88, + 0.166 + ], + "angle": 0, + "content": "
styleEPE1PE3PE
w/o0.7812.002.66
linear0.7410.862.30
same0.7611.452.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.485, + 0.168, + 0.889, + 0.195 + ], + "angle": 0, + "content": "(b) Lookup style. Sampling correlation volumes with a set of linear coordinates improves the accuracy." + }, + { + "type": "table", + "bbox": [ + 0.493, + 0.197, + 0.879, + 0.272 + ], + "angle": 0, + "content": "
layersEPE1PE3PETime (ms)
00.7511.362.3320
1/20.7410.862.3030
0.7510.902.3238
" + }, + { + "type": "table_caption", + "bbox": [ + 0.484, + 0.274, + 0.889, + 0.303 + ], + "angle": 0, + "content": "(d) Layers of pattern aggregation module. Stacking 1 pattern aggregation module showcases high accuracy." + }, + { + "type": "table", + "bbox": [ + 0.493, + 0.304, + 0.875, + 0.394 + ], + "angle": 0, + "content": "
radiusEPE1PE3PE
10.7511.142.37
20.7611.162.39
3/40.7410.862.30
0.7510.932.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.396, + 0.86, + 0.411 + ], + "angle": 0, + "content": "(f) Searching radius. A radius of 3 leads to high accuracy." + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.413, + 0.845, + 0.428 + ], + "angle": 0, + "content": "Table 4: TMA ablations. Models are trained on DSEC-Flow. Settings used in our final model are underlined." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.437, + 0.188, + 0.451 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.469, + 0.506 + ], + "angle": 0, + "content": "We conduct a series of ablation experiments to validate the proposed improvements. All ablation models are trained on DSEC-Flow and evaluated on the public benchmark." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.469, + 0.612 + ], + "angle": 0, + "content": "Event splitting. We compare different numbers of event segments in Table 4a. We choose different splits in 1, 3, 5 and 7. With the split set to 1, our model degrades into a classic correlation volume based framework and exhibits similar accuracy in comparison with E-RAFT. More splits lead to higher accuracy but lower inference efficiency, so we use 5 splits to achieve accuracy-efficiency trade-off." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.469, + 0.718 + ], + "angle": 0, + "content": "Lookup style. Based on the assumption that the optical flow changes linearly over a short period, we propose the linear lookup strategy. To validate it, we try to cancel the lookup operation or just use same sampling coordinates on intermediate correlation volumes. As Table 4b shows, sampling the correlation maps with a set of linear coordinates leads to high accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.854 + ], + "angle": 0, + "content": "Pattern aggregation module design. We ablate different components of pattern aggregation module in Table 4c. By default, we set the projection matrix of value \\( W^V \\) as an identity matrix because the motion features are encoded by a motion feature encoder beforehand. In addition, we replace the pairing mode of motion features with circular pairing. Circular pairing gives equal treatment to all motion features without highlighting the importance of the last motion feature, which fails to improve accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Layers of pattern aggregation module. We compare the accuracy with different layers of pattern aggregation modules in Table 4d. We note that stacking 1 layer of pat" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.892, + 0.482 + ], + "angle": 0, + "content": "tern aggregation module showcases high accuracy. More layers may lead to difficulties in optimization, which in turn causes a decrease in accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.528 + ], + "angle": 0, + "content": "Iterations. Iterative learning plays a role in numerous works [18, 30]. Results in Table 4e show that 6 iterative updates reach the sweet spot of accuracy and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.892, + 0.604 + ], + "angle": 0, + "content": "Searching radius. We compare different radii for looking up correlation volumes in Table 4f. Searching radius of 3 leads to high accuracy. Different from RAFT [30] that achieves better predictions with an increased radius, our model does not benefit a lot from a greater radius." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.619, + 0.634 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we delve into event-based optical flow estimation problem. Different from existing methods that represent the event stream as static frames to adopt the classic correlation volume based pipeline, we contend that the temporal continuity is a vital element of event-based optical flow. We propose a novel Temporal Motion Aggregation (TMA) approach to unlock its potential within a learning-based framework. By incorporating temporally fine-grained motion information, TMA generates high-quality flow estimates at early stages, which not only enables TMA to obtain accurate final predictions, but also greatly reduces the demand for a number of refinements. Extensive experiments on DESC-Flow and MVSEC datasets verify the effectiveness and superiority of TMA. Meanwhile, we notice the limitations of supervised learning in generalization ability and expect further improvement to extend our idea into unsupervised learning frameworks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9692" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.204 + ], + "angle": 0, + "content": "Acknowledgments: This work was supported in part by Shanghai Municipal Science and Technology Major Project under Grant 2018SHZDZX01, ZJ Lab, the Shanghai Center for Brain Science and Brain-Inspired Technology, in part by Shanghai Rising Star Program under Grant 21QC1400900, in part by Tongji-Qomolo Autonomous Driving Commercial Vehicle Joint Lab Project, and in part by Xiaomi Young Talents Program." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.23, + 0.174, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.47, + 0.324 + ], + "angle": 0, + "content": "[1] Himanshu Akolkar, Sio-Hoi Ieng, and Ryad Benosman. Real-time high speed motion prediction using fast aperture-robust event-driven visual flow. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):361-372, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.472, + 0.381 + ], + "angle": 0, + "content": "[2] Mohammed Almatrafi, Raymond Baldwin, Kiyoharu Aizawa, and Keigo Hirakawa. Distance surface for event-based optical flow. IEEE transactions on pattern analysis and machine intelligence, 42(7):1547-1556, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[3] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE transactions on neural networks and learning systems, 25(2):407-417, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.492 + ], + "angle": 0, + "content": "[4] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.47, + 0.549 + ], + "angle": 0, + "content": "[5] Vincent Brebion, Julien Moreau, and Franck Davoine. Realtime optical flow for vehicular perception with low-and high-resolution event cameras. IEEE Transactions on Intelligent Transportation Systems, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[6] Tobias Brosch, Stephan Tschechne, and Heiko Neumann. On event-based optical flow detection. Frontiers in neuroscience, 9:137, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[7] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-temporal recurrent networks for event-based optical flow estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 525-533, 2022. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.663, + 0.47, + 0.732 + ], + "angle": 0, + "content": "[8] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.733, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[9] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3867-3876, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[10] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cam" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.256, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "eras. In 2021 International Conference on 3D Vision (3DV), pages 197-206. IEEE, 2021. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[12] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from events and frames. arXiv preprint arXiv:2203.13674, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[13] Jesse Hagenaars, Federico Paredes-Vallés, and Guido De Croon. Self-supervised learning of event-based optical flow with spiking neural networks. Advances in Neural Information Processing Systems, 34:7167-7179, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[14] Markus Hofinger, Samuel Rota Bulò, Lorenzo Porzi, Arno Knapitsch, Thomas Pock, and Peter Kontschieder. Improving optical flow on a pyramid level. In European Conference on Computer Vision, pages 770-786. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[15] Tak-Wai Hui and Chen Change Loy. Liteflownet3: Resolving correspondence ambiguity for more accurate optical flow estimation. In European Conference on Computer Vision, pages 169-184. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[16] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8981-8989, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[17] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. A lightweight optical flow cnn—revisiting data fidelity and regularization. IEEE transactions on pattern analysis and machine intelligence, 43(8):2555-2569, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[18] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2019. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[19] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9772-9781, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.58, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[20] Chankyu Lee, Adarsh Kumar Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, and Kaushik Roy. Spikeflownet: event-based optical flow estimation with energy-efficient hybrid neural networks. In European Conference on Computer Vision, pages 366-382. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[21] Min Liu and Tobi Delbruck. Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. BMVC, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[22] Min Liu and Tobi Delbruck. Edflow: Event driven optical flow camera with keypoint detection and adaptive block matching. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5776-5789, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[23] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Federico Paredes-Vallés and Guido CHE de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446-3455, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4161-4170, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9693" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[26] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.215 + ], + "angle": 0, + "content": "[27] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In European Conference on Computer Vision, pages 628-645. Springer, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.3 + ], + "angle": 0, + "content": "[28] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications, volume 11006, page 1100612. International Society for Optics and Photonics, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[29] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[30] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.471, + 0.496 + ], + "angle": 0, + "content": "[31] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.47, + 0.579 + ], + "angle": 0, + "content": "[32] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.69 + ], + "angle": 0, + "content": "[34] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing, 31:7237-7251, 2022. 1, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[35] Haofei Xu, Jiaolong Yang, Jianfei Cai, Juyong Zhang, and Xin Tong. High-resolution optical flow from 1d attention and correlation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10498-10507, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[36] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[37] Gengshan Yang and Deva Ramanan. Volumetric correspondence networks for optical flow. Advances in neural information processing systems, 32:794-805, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[38] Chengxi Ye, Anton Mitrokhin, Cornelia Fermüller, James A Yorke, and Yiannis Aloimonos. Unsupervised learning of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "dense optical flow, depth and egomotion with event-based sensors. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5831-5838. IEEE, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[39] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 185-194, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[40] Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In European Conference on Computer Vision, pages 420-439. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[41] Feihu Zhang, Oliver J Woodford, Victor Adrian Prisacariu, and Philip HS Torr. Separable flow: Learning motion cost volumes for optical flow estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10807-10817, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[42] Xu Zheng, Yexin Liu, Yunfan Lu, Tongyan Hua, Tianbo Pan, Weiming Zhang, Dacheng Tao, and Lin Wang. Deep learning for event-based vision: A comprehensive survey and benchmarks. arXiv preprint arXiv:2302.08890, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[43] Alex Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In Proceedings of Robotics: Science and Systems, Pittsburgh, Pennsylvania, June 2018, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.473, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[44] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.544, + 0.894, + 0.614 + ], + "angle": 0, + "content": "[45] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 1, 2, 3, 4, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9694" + } + ] +] \ No newline at end of file diff --git a/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/9d269c18-5110-4a81-a1b9-317c1455548e_origin.pdf b/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/9d269c18-5110-4a81-a1b9-317c1455548e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6ab07a698146cc0fb9f1c1d10820249d8700735e --- /dev/null +++ b/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/9d269c18-5110-4a81-a1b9-317c1455548e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88fa59189bb6d0e751635ccdf9623d1a987401adb76a333c6f2116a84203f37e +size 1037859 diff --git a/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/full.md b/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/full.md new file mode 100644 index 0000000000000000000000000000000000000000..25d4b45ef21572c62369a8d6322834a46149482e --- /dev/null +++ b/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/full.md @@ -0,0 +1,318 @@ +# TMA: Temporal Motion Aggregation for Event-based Optical Flow + +Haotian Liu1 Guang Chen1* Sanqing Qu1 Yanping Zhang1 +Zhijun Li1 Alois Knoll2 Changjun Jiang1 +1Tongji University 2Technical University of Munich + +# Abstract + +Event cameras have the ability to record continuous and detailed trajectories of objects with high temporal resolution, thereby providing intuitive motion cues for optical flow estimation. Nevertheless, most existing learning-based approaches for event optical flow estimation directly remould the paradigm of conventional images by representing the consecutive event stream as static frames, ignoring the inherent temporal continuity of event data. In this paper, we argue that temporal continuity is a vital element of event-based optical flow and propose a novel Temporal Motion Aggregation (TMA) approach to unlock its potential. Technically, TMA comprises three components: an event splitting strategy to incorporate intermediate motion information underlying the temporal context, a linear lookup strategy to align temporally fine-grained motion features and a novel motion pattern aggregation module to emphasize consistent patterns for motion feature enhancement. By incorporating temporally fine-grained motion information, TMA can derive better flow estimates than existing methods at early stages, which not only enables TMA to obtain more accurate final predictions, but also greatly reduces the demand for a number of refinements. Extensive experiments on DSEC-Flow and MVSEC datasets verify the effectiveness and superiority of our TMA. Remarkably, compared to E-RAFT, TMA achieves a $6\%$ improvement in accuracy and a $40\%$ reduction in inference time on DSEC-Flow. Code will be available at https://github.com/ispc-lab/TMA. + +# 1. Introduction + +Optical flow aims to compute the velocity of objects on the image plane without geometry prior and is a fundamental topic in event-based vision [8, 42]. It plays an important role in many applications, such as ego-motion estimation [38, 45], image reconstruction [24], and video frame interpolation [32, 31]. Recently, learning-based methods [29, 18, 30] have dominated frame-based optical flow + +![](images/d79da3f60da344d5debf26737a7222933e2d1633fbf35168e314c06375fb14cc.jpg) +(a) reference image + +![](images/76e22e787cc0093cd70f73f2ae77508cbad73d691a23df358884f3ee86f683c7.jpg) +(b) event frame + +![](images/197e01e0c9ab95123c67414acdb7f28e2441de1741495e0250013e52ab5bb6cf.jpg) + +![](images/71b7e75eee9838c46a76b72c51fcaf25d19f44bae976519f476782507c6b6319.jpg) +(c) E-RAFT output & motion feature (d) our output & motion feature + +![](images/2cddd66dab12422f7745cde39ca77fb477f80355b39094fea5f61fc9c634c840.jpg) + +![](images/b1b82bcbcfe65b05eb791f6404309750ddc92df34c88fc0a65b6383fd574f6c3.jpg) + +Figure 1: Accuracy illustrations. (a) Reference image from DSEC-Flow. (b) Corresponding event frame. (c) The final flow prediction and visualization of the motion feature from E-RAFT [11]. (d) The final flow prediction and visualization of the motion feature from our proposed method. We visualize the motion feature at the initial stage by taking the average across channels. E-RAFT fails to generate informative motion features and results in blurred boundaries. In contrast, our method utilizes temporally fine-grained motion information to address the information scarcity issue in motion features, generating high-quality predictions. + +by employing correlation volumes to derive motion features for flow regression. Inspired by this, several event-based approaches [11, 34] adopt a similar paradigm by converting the consecutive event stream into grid-like representations. Despite encouraging progress, these methods still suffer from non-negligible limitations, due to the great differences between event data and conventional images. Distinct from dense and colorful conventional (RGB) images, event data features spatial sparsity and a lack of intensity information. It is prone to result in close matching scores and invalid regions in the correlation volume. Consequently, less informative motion features are derived and regress inaccurate + +predictions. Figure 1 illustrates this issue in E-RAFT [11]. + +As the saying goes, each coin has two sides. Though event cameras encode visual information in a sparse manner, they are capable of capturing continuous and detailed object trajectories with high temporal resolution, thereby providing rich motion cues for optical flow estimation. Several model-based methods [9, 27] benefit from the temporal continuity by warping events along point trajectories, achieving decent performance. Drawing on this observation, we contend that the temporal continuity is a vital element of event-based optical flow and propose to unleash its potential within a learning-based framework. + +To materialize our idea, we propose a novel Temporal Motion Aggregation (TMA) approach to explore the inherent temporal continuity of event data. Technically, we revamp the dominant learning-based architectures with three distinct components. We first introduce an event splitting strategy for temporally-dense correlation volumes computation. By splitting the event stream into multiple segments and extracting their features, feature similarities are compared between the first feature and all others, which record rich intermediate motion information. Aware of that correlation volumes record motions between different time spans, we then design a linear lookup strategy to sample each correlation volume based on the corresponding flow estimate to encode motion features. As a result, fine-grained motions are warped to the same pixels across motion features. Furthermore, we consider the incorrect motion patterns in intermediate motion features due to the manual lookup. Therefore, we propose a novel pattern aggregation module for motion feature enhancement by aggregating consistent patterns between each intermediate motion feature and the last one (without manual lookup). Thanks to the incorporation of temporally fine-grained motion information, TMA can effectively address the information scarcity in motion features and thus generate good flow estimates at early stages, which not only enable TMA to obtain accurate flow predictions, but also greatly reduces the demand for a number of time-consuming refinements. + +We evaluate our TMA on DSEC-Flow [10] and MVSEC [44] datasets. Extensive experiments demonstrate the superior advantages of TMA. + +In summary, our main contributions are as follows: + +- We argue that the temporal continuity of event data is a vital element of event-based optical flow and propose to unlock its potential in a learning-based framework. +- We propose a novel Temporal Motion Aggregation (TMA) approach, which comprises three components: an event splitting strategy, a linear lookup strategy and a motion pattern aggregation module. +- The incorporation of temporally fine-grained motion information enables TMA to achieve high accuracy + +while maintaining high efficiency. Compared with E-RAFT, TMA achieves a $6\%$ improvement in accuracy and a $40\%$ reduction in inference time on DSEC-Flow. + +# 2. Related Work + +Learning methods for optical flow estimation. Recently, correlation volume based approaches dominate the learning-based optical flow estimation [25, 17, 29, 18, 37, 14]. As a milestone, PwCNet [29] summarizes a simple principle in architecture construction: pyramidal processing, warping and correlation volumes computation. Following the well-established pipeline, LiteFlowNet series [16, 17, 15] propose cascaded feature pyramids with a lightweight convolutional burden. VCN [37] proposes 4D volumetric U-Net architectures to enlarge the receptive fields for correlation volumes. IRR [18] notes the reusability of pyramidal refinement for flow update and proposes a shared-weights refiner across all scales, which can be combined with mainstream learning methods. + +However, the correlation volumes used in previous works are computed locally, which cannot fully deal with long-term motions. RAFT [30] firstly introduces the full correlation volume to capture global visual similarities. Separable Flow [41] proposes a separable correlation volume module as a drop-in replacement to correlation volumes and uses non-local aggregation layers [39, 40] to refine the correlation volumes efficiently. Besides, transformer components show promising performance on optical flow [36, 19, 35]. GMA [19] notices the underlying motion clues in the context and introduces an attention layer to aggregate hidden motions, improving the accuracy based on RAFT architecture. GMFlow [36] reformulates optical flow as a global matching problem by attention blocks, suggesting a new paradigm for flow estimation. + +We draw inspiration from GMA [19] to design our motion pattern aggregation module. In contrast to GMA which aggregates hidden motion patterns in the spatial context for motion feature enhancement, we extend the pattern aggregation to the temporal dimension and leverage feature cross-similarities to enhance the motion feature. + +Event-based optical flow estimation. A number of traditional works remould frame-based approaches into event-based optical flow estimation. Benosman et al. [4] extends the classic Lucas-Kanade method to event-based optical flow. Liu et al. [21, 22] implement event-based corner keypoint detection and novel block-matching optical flow. Gallego et al. [9] propose a unifying contrast maximization (CM) framework and MultiCM [27] extends it to complex scenes. Besides, plane fitting [1, 3], filter banks [6] and time surface matching [2] are proposed to boost performance. + +More recently, learning-based methods have dominated frame-based optical flow, inspiring attempts at event-based optical flow. Early learning methods [43, 45, 38] represent + +![](images/bd2575db2ab1f802b8ae96be3731aacc30879bf0446c7a6eb2bb30f9a195f1ba.jpg) +Figure 2: The overall architecture of TMA. First, we split the event stream into segments and extract features with a share-weights feature extractor. Then, regarding the first feature as reference, we correlate all other features to compute temporally-dense correlation volumes, which contain rich embedded motion information. Next, aware of the different time spans of correlation volumes, we look up each correlation volume based on a corresponding flow estimate. All sampled correlation maps are delivered into a share-weights motion feature encoder to derive motion features. We further propose a novel motion pattern aggregation module to enrich the spatial information of motion features. By concatenating the enhanced motion features and utilizing several refinements, high-quality flow predictions are generated. + +the event stream as a static frame and utilize the architecture of U-Net [26] to predict optical flow, which can only address small motions. Mainstream works [7, 34, 11] introduce two consecutive event frames and follow the correlation volume based pipeline (PWCNet and RAFT), achieving higher accuracy on large-scales datasets [44, 10]. E-RAFT [11, 12] extends RAFT [30] and introduces full correlation volume into event-based vision. DCEIFlow [34] combines conventional images and event data to boost higher accuracy. Note that our work and STE-FlowNet [7] share a similar motivation to introduce temporal continuity with fine-grained discretization. However, STE-FlowNet takes a frame-by-frame manner [18] to update flow, leading to an insufficient inference process. + +Contrary to previous learning-based methods, we do not merely adopt the classic correlation volume based pipeline. Instead, we regard the temporal continuity of event data as a vital component of optical flow and incorporate the intermediate motion information into a learning-based framework, which boosts higher accuracy and efficiency. + +# 3. Methodology + +Given an input event stream $\{\mathcal{E}\}$ from $t_0$ to $t_1$ , the goal of optical flow is to estimate a dense displacement field $\mathbf{u}:\mathbb{R}^2\to \mathbb{R}^2$ which maps each per-pixel $(x,y)$ at $t_0$ to its corresponding coordinate $(x^{\prime},y^{\prime})$ at $t_1$ . To address this, mainstream learning-based approaches compute correlation volumes from two event frames as a matching prior. Such principle inspires several methods [11, 34] by squeezing event data into consecutive frames. Unfortunately, the characteristics of event data, spatial sparsity and lack of intensity records, are prone to result in close matching scores and invalid regions in the correlation volume. Consequently, less informative motion features are derived and regress inaccurate predictions. Here we address the information scarcity + +of motion features from the attractive temporal continuity of event data. Temporally fine-grained motion information in event data is sufficient to supplement the spatial richness of motion features, yielding high-quality flow predictions. + +In the following, we first describe the preliminary and event representation, and then present the detailed architecture of the proposed TMA. + +# 3.1. Preliminary and Event Representation + +Event cameras record light intensity changes in an asynchronous way. Each pixel serves as an independent trigger and generates an event instantly whenever the log intensity changes exceeding a threshold $C$ . An event $\boldsymbol{e}_k = (\boldsymbol{x}_k, p_k, t_k)$ is a triplet containing triggered pixel location $\boldsymbol{x}_k = (x_k, y_k)^T$ , timestamp $t_k$ and signed intensity change $(p_k = \pm 1)$ . + +To ease the feature extraction and preserve the temporal continuity of the event data, we firstly transform the event stream into a 3D volume $V(x,y,t)$ along the temporal dimension following existing works [45, 11, 24]. Given an event stream $\{\mathcal{E}\}$ , $V(x,y,t)$ is generated as: + +$$ +t _ {i} ^ {*} = (B - 1) \left(t _ {i} - t _ {1}\right) / \left(t _ {N _ {e}} - t _ {1}\right) \tag {1} +$$ + +$$ +\boldsymbol {V} (x, y, t) = \sum_ {i} p _ {i} k _ {b} \left(x - x _ {i}\right) k _ {b} \left(y - y _ {i}\right) k _ {b} \left(t - t _ {i} ^ {*}\right) \tag {2} +$$ + +$$ +k _ {b} (a) = \max (0, 1 - | a |), \tag {3} +$$ + +where $t_i^*$ represents the $i$ -th time bin. $B$ and $N_e$ represent the length of time bins and the number of events, respectively. $k_b(a)$ is the bilinear interpolation function. + +# 3.2. TMA Architecture + +We develop our framework based on the successful RAFT architecture [30]. Figure 2 summarizes the architecture of the proposed TMA, which consists of three components, an event splitting strategy, a linear lookup strategy + +![](images/8da1b9b2f86637a5772622540d3d8a83337ccf577b5648551fa3ce8ccecb5f51.jpg) +Figure 3: Linear lookup strategy. We compute correlation volumes by comparing the similarity of first feature with all others, which are misaligned due to different time spans. To rectify this, we narrow down the lookup coordinates linearly to warp correlation maps to spatially aligned pixels. + +and a motion pattern aggregation module. With the proposed method, we successfully unlock the potential of temporal continuity to predict event-based optical flow within a learning-based framework. + +Event splitting. Instead of viewing the continuous event stream as a static frame which tends to be a matter of common knowledge in recent works [45, 11, 34], we prefer to identify it as a high frame rate video clip, which records continuous and detailed trajectories of objects, namely, provides intuitive motion cues for optical flow estimation. To introduce rich intermediate motion information into our method, we first propose an event splitting strategy, which splits the event stream into multiple segments. Therefore, more intermediate event frames are retained for correlation volume computation. + +In particular, given an event stream $\{\mathcal{E}\}$ from $t_0$ to $t_1$ , we split it into $g$ segments equally and the time span between adjacent segments is $dt$ . Besides, we introduce an auxiliary event segment from $t_0 - dt$ to $t_0$ as reference. Then, $g + 1$ event representations are generated following Sec. 3.1. All $g + 1$ event representations are then delivered into a share-weights encoder to achieve features $\pmb{F}_i \in \mathbb{R}^{H \times W \times C}, i = 0, 1, 2, \dots, g,$ where $H, W$ and $D$ represent the height, width and channel of features. We use all features to compute temporally-dense correlation volumes $\pmb{C}_i, i = 1, 2, \dots, g,$ which are given by: + +$$ +\boldsymbol {C} _ {i} = \frac {\boldsymbol {F} _ {0} \boldsymbol {F} _ {i} ^ {T}}{\sqrt {D}} \in \mathbb {R} ^ {H \times W \times H \times W}, i = 1, 2, \dots , g, \tag {4} +$$ + +where $\pmb{F}_0$ is the first feature and $\pmb{F}_i$ is the $i$ -th feature derived from the $i$ -th event frame. + +Temporally-dense correlation volumes compare feature similarities across multiple time spans, enabling refined records of relative motions. This construction is vastly different from the classic full correlation volume that only contains feature similarities between a fixed time span. + +Linear lookup. Recall that each correlation volume $C_i$ + +![](images/edd9caa0b9361b83c8177c5cf45bb2de7ef9eea9cdead27b15955c2036afefbb.jpg) +Figure 4: Motion pattern aggregation. We pair each intermediate motion feature to the last motion feature. Then each pair is sent to cross-attention blocks to aggregate consistent patterns. Hence, each intermediate motion feature carries information from the last one and consistent patterns receive emphasis. + +is achieved by correlating $F_{0}$ and $F_{i}$ , which means the recorded relation motions are misaligned in time. While our goal is to predict optical flow from $t_{0}$ to $t_{1}$ , we hope that all sampled correlation maps could present the relative motions within the same time span. Therefore, based on that correlation volumes are evenly constructed in time and magnitudes of optical flow change linearly over a short period, we narrow down lookup coordinates in a linear fashion to match the timestamp of each correlation volume. + +Figure 3 displays the linear lookup strategy. In detail, given current flow estimate $\mathbf{u}$ , we split it into $g$ segments to sample each correlation volume. The linear lookup operation can be expressed as: + +$$ +\mathbf {d} \mathbf {u} = \left(\mathbf {c o o r d s} _ {1} - \mathbf {c o o r d s} _ {0}\right) / g = \mathbf {u} / g \tag {5} +$$ + +$$ +\operatorname {c o r r} _ {i} = S \left(\boldsymbol {C} _ {i}, \operatorname {c o o r d s} _ {0} + i \mathrm {d u}\right), i = 1, 2, \dots , g, \tag {6} +$$ + +where $\mathbf{coord}_0$ is the initialized coordinates at $t_0$ (reference time), $S(\cdot, \cdot)$ represents the lookup operation, $\mathbf{coord}_i$ are lookup coordinates for $C_i$ and $\mathbf{corr}_i$ is the correlation map sampled from $C_i$ . After delivering all correlation maps into a share-weights motion feature encoder, aligned motion features are achieved. + +From the perspective of image contrast [9], the standard lookup is responsible for warping possible matching points between two images to the same pixel. Our linear lookup extends it by warping matching points of a temporally fine-grained motion to the same pixel. As a result, similar motions are presented at same pixels across correlation maps, showing high contrast. + +Motion pattern aggregation. A standard operation for next step is to concatenate the motion features as a whole to update flow and the improvements made so far already alleviate the information scarcity in motion features. However, we notice that the concatenation operation treats all motion features equally, which means the majority of motion features, intermediate ones, dominate the flow regression. Due + +![](images/4cc613ce347d8063a08a953b1258344670e8f9d30b3ff8cbf420ebacd78e806e.jpg) +Figure 5: Qualitative examples of DSEC-Flow. For each two columns, we show the reference image and event frame and compare our method with a state-of-the-art baseline E-RAFT [11]. Significant improvements are highlighted by red boxes. + +![](images/c9d09ce8f6abd1e8a67f68c318998a498a71fed2457dfd334adf1bcfb3bfa69c.jpg) + +![](images/2a660dd95348afbefb349d5a459a791499afdd118c8ebd0d7b0e307863e2f783.jpg) + +to the linear optical flow assumption and manual lookup operation, these motion features are not perfectly correct. A natural thought is to put more emphasis on the last motion feature that is without manual lookup and retain the patterns in intermediate motion features that are consistent with the last motion feature. To this end, we balance the weights between the last motion feature and others through cross-similarities computed by transformers [33, 19]. + +Figure 4 shows the overview of motion pattern aggregation module. Given a set of motion features $M\mathbf{F}_i \in \mathbb{R}^{H \times W \times D'}$ , $i = 1,2,\dots,g$ , we pair each intermediate motion feature to the last one. Then each pair $(M\mathbf{F}_i,M\mathbf{F}_g)$ is sent to cross-attention blocks to compare cross-similarities and aggregate consistent patterns. The enhanced motion feature $\widehat{M}\mathbf{F}_i$ is obtained as: + +$$ +\begin{array}{l} \boldsymbol {A} = \operatorname {A t t e n t i o n} \left(\boldsymbol {M} \boldsymbol {F} _ {i} W ^ {Q}, \boldsymbol {M} \boldsymbol {F} _ {g} W ^ {K}, \boldsymbol {M} \boldsymbol {F} _ {g} W ^ {V}\right) (7) \\ \widehat {\boldsymbol {M F}} _ {i} = \boldsymbol {M F} _ {i} + \operatorname {M L P} \left(\left[ \boldsymbol {M F} _ {i}, \boldsymbol {A W} ^ {O} \right]\right), (8) \\ \end{array} +$$ + +where $\operatorname{Attention}(Q, K, V) = \operatorname{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V$ , $d_k$ is the dimension of $K$ . $W^Q$ , $W^K$ , $W^V$ are projection matrices. $\mathrm{MLP}(\cdot)$ is multi-layer perceptron and $W^O$ is another projection matrix. We omit the normalization functions in the manuscript for simplicity. Motion pattern aggregation module does not change the dimension of motion features, so multiple cross-attention layers can be stacked easily. + +Once motion features are enhanced by the motion pattern aggregation module, each $\widehat{MF}_i$ carries motion patterns that are consistent with $MF_{g}$ . Namely, the last motion feature receives high attention and consistent patterns in motion features are fully utilized. + +Above three components are seamlessly integrated into a unified pipeline: incorporating intermediate motion information, warping matching motions and emphasizing consistent motion patterns. Attributed to the innovative application of temporal continuity to a learning-based frame + +work, our proposed method is capable of overcoming the information scarcity issue in motion features and generating high-quality flow estimates even at an early refinement stage. This enables our method to achieve accurate final predictions, while reducing the demand for a number of refinements. + +# 3.3. Supervision + +Following RAFT [30], we choose $L_{1}$ distance between the predictions and ground-truth as the supervision signal for our model. The loss function $\mathcal{L}$ is formally defined as: + +$$ +\mathcal {L} = \sum_ {j = 1} ^ {N} \gamma^ {N - j} \left\| \mathbf {u} ^ {g t} - \mathbf {u} ^ {j} \right\| _ {1}, \tag {9} +$$ + +where $\mathbf{u}^{gt}$ denotes the flow ground-truth and $\mathbf{u}^j$ the flow prediction at $j$ -th stage. $N$ is the total refinement stages. The weight $\gamma$ balances different stages of predictions. + +# 4. Experiments + +Datasets and evaluation setup. Following previous works [11, 45, 7], we provide experimental details and extensive comparison results on two event-based datasets DSEC-Flow [10] and MVSEC [44]. For DSEC-Flow, we train models on the official training set and evaluate DSEC-Flow test set on the public benchmark. For MVSEC, we conduct two kinds of experiments. Basically, following previous methods [43, 27], we train methods on outdoor_day2 sequence with ground-truth corresponding to a time interval of $dt = 1$ and $dt = 4$ gray images and evaluate 800 frames of outdoor_day1 and three indoor_flying sequences. Equally importantly, considering the supervised learning setting of our method and the large gap between outdoor_day and indoor_flying data, we extend the range of training data by including one indoor_flying sequence each time to further boost the accuracy of supervised learning methods. + +
MethodsEPE1PE3PETime(ms)
MBMultiCM [27]3.4776.630.9-
SLEV-FlowNet†[43]2.3255.418.65
E-RAFT [11]0.7912.52.752
Ours0.7410.92.330
+ +Table 1: Results on DSEC-Flow. Best accuracy in bold. $\dagger$ denotes the result is taken from E-RAFT [11]. Model-based (MB) methods need no training data; and supervised learning (SL) methods need ground-truth. + +Metrics. The end-point-error (EPE) is used as core metric for prediction accuracy for both DSEC-Flow and MVSEC. To evaluate the robustness against large displacements, we compute the percentage of pixels with EPE greater than $M$ pixels $(M\mathrm{PE},M = 1,3)$ for DSEC-Flow. Percentage of pixels with EPE above 3 and $5\%$ of the magnitude of the flow ground-truth (\% Outlier) are presented for MVSEC. On DSEC-Flow, metrics are measured over pixels with valid ground-truth. While on MVSEC, metrics are measured over pixels with valid ground-truth and at least one triggered event. + +Implementation details. The proposed TMA is implemented using PyTorch. We set the event segment $g = 5$ . For experiments on DSEC-Flow, the channel of event representation $B$ for each segment is set to 3. For MVSEC, $B$ is set to 1 when $dt = 1$ and 3 when $dt = 4$ , respectively. The sum of channels keeps the same with the state-of-the-art E-RAFT [11]. We only stack 1 layer of motion pattern aggregation module. Other components are basically identical to RAFT's model, including feature extractor, motion feature encoder and flow updater. Note that we use a small feature extractor whose final channel is 128. We predict 6 levels of optical flow and the weight $\gamma$ is set to 0.8 for supervision. In training, we exploit AdamW [23] optimizer and One-cycle learning rate scheduler [28]. We train models on DSEC-Flow for 200K steps and MVSEC for 100K steps, both with a learning rate of 0.0002 and a batch size of 6. + +# 4.1. DSEC-Flow + +Accuracy comparison. Evaluation results on DSEC-Flow benchmark1 are provided in Table 1. Compared with E-RAFT, our method improves the EPE from 0.79 to 0.74. Our method is also leading in 1PE and 3PE with a margin of 1.6 and 0.4, showing better robustness. + +Qualitative results are exhibited in Figure 5. In challenging regions such as textureless areas (the wall in second two columns, the shrubs in last two columns), existing correlation volume contains similar matching scores and fails to generate informative motion features, leading to low-quality predictions. On the contrary, our method aggregates + +![](images/913c95c892169db088b345359884f596481916bbb25a0f55db95311de59310a9.jpg) +Figure 6: EPE vs. number of iterations at inference. The figure exhibits the prediction results on DSEC-Flow at different stages during inference. Our method achieves an accuracy $(\mathrm{EPE} = 0.81)$ comparable to E-RAFT with 2 iterations while E-RAFT needs 6 iterations. + +temporally consistent motion patterns to enhance motion features, distinguishing different moving objects and generating high-quality predictions. + +Efficiency comparison. The last column of Table 1 exhibits the inference time. Our method enjoys faster inference speed compared with E-RAFT, which is mainly attributed to fewer iterative updates. The reasonable explanation is that our method utilizes temporal motion information to enrich motion features, thus being able to generate good predictions even at an initial stage. To prove this, we compare the accuracy with E-RAFT at early inference stages in Figure 6. Our method achieves better flow initialization at early stages, supporting our argument. The ablation experiment on different iterations (Table 4e) also demonstrates that our method requires fewer iterations (ours: 6, E-RAFT: 12) to achieve surpassing accuracy. + +# 4.2.MVSEC + +Inter-domain evaluation. Table 2 reports the results on MVSEC training with outdoor_day2 sequence. The top of Table 2 reports the results corresponding to $dt = 1$ grayscale frame and the bottom corresponds to $dt = 4$ grayscale frames. Our method shows comparable accuracy on outdoor_day1 for $dt = 1$ and highest accuracy on outdoor_day1 for $dt = 4$ among all compared methods. Unfortunately, like two other supervised learning methods, DCEI-Flow [34] and E-RAFT [11], our method fails to achieve superior results on indoor_flying sequences. One possible reason is that model based methods and unsupervised learning methods usually have better generalization ability, while supervised learning methods are limited if training data has a significantly large gap with test data [36, 34]. Since outdoor_day sequences and indoor_flying sequences have a great domain gap, it is challenging for supervised learning methods to generalize well on indoor_flying sequences. + +Intra-domain evaluation. To verify the possible cause and better analyze the accuracy of our method under su + +
Inputindoor_flying1indoor_flying2indoor_flying3outdoor_day1
EPE% OutlierEPE% OutlierEPE% OutlierEPE% Outlier
dt = 1
MBBrebion et al. [5]E0.520.100.985.500.712.100.530.20
MultiCM (Burgers') [27]E0.420.100.600.590.500.280.300.10
SSLEV-FlowNet [43]E1.032.201.7215.101.5311.900.490.20
Spike-FlowNet [20]E0.84-1.28-1.11-0.49-
STE-FlowNet [7]E0.570.100.791.600.721.300.420.00
USLHagenaars et al. [13]E0.600.511.178.060.935.640.470.25
Zhu et al. [45]E0.580.001.024.000.873.000.320.00
ECN [38]E------0.300.02
SLE-RAFT [11]E1.105.721.9430.791.6625.200.240.00
DCEIFlow [34]E + I0.751.550.902.100.801.770.220.00
OursE1.063.631.8127.291.5823.260.250.07
dt = 4
MBMultiCM (Burgers') [27]E1.6912.952.4926.352.0619.031.259.21
SSLEV-FlowNet [43]E2.2524.704.0545.303.4539.701.237.30
Spike-FlowNet [20]E2.24-3.83-3.18-1.09-
STE-FlowNet [7]E1.7714.702.5226.102.2322.100.993.90
USLZhu et al. [45]E2.1824.203.8546.803.1847.801.309.70
Hagenaars et al. [13]E2.1621.513.9040.723.0029.601.6912.50
SLDCEIFlow [34]E + I2.0821.473.4842.052.5129.730.893.19
E-RAFT [11]E2.8140.255.0964.194.4657.110.721.12
OursE2.4329.914.3252.743.6042.020.701.08
+ +Table 2: Evaluation results on MVSEC [44] training with outdoor_day2 sequence. Model-based (MB) methods need no training data; semi-supervised learning (SSL) methods use grayscale images for supervision; unsupervised learning (USL) methods only require events; and supervised learning (SL) methods need ground-truth. E means the method only uses events as input; and I means the method requires images as input. + +
dt = 4Train Set +indoor_flying1indoor_flying2indoor_flying3outdoor_day1
EPE% OutlierEPE% OutlierEPE% OutlierEPE% Outlier
E-RAFT [11]indoor_flying1--2.3526.241.8418.070.721.38
Ours--2.1520.411.6814.300.640.98
E-RAFT [11]indoor_flying21.418.14--1.509.370.711.13
Ours1.325.94--1.508.880.640.96
E-RAFT [11]indoor_flying31.398.381.7717.52--0.721.46
Ours1.335.821.6712.80--0.671.10
+ +Table 3: Evaluation results on MVSEC [44] training with outdoor_day2 and one indoor_flying sequence. + means the indoor_flying sequence to be included into training set. + +pervised learning setting, we incorporate one indoor_flying sequence into training set each time and make comparisons with the supervised learning state-of-the-art E-RAFT. Table 3 shows the evaluation results corresponding to $dt = 4$ grayscale frames (For $dt = 1$ , refer to appendix). Compared with the results in Table 2, E-RAFT and our method both achieve great improvement on indoor_flying sequences (EPE of indoor_flying2, ours: $4.32 \rightarrow 1.67$ , E-RAFT: 5.09 → 1.77), confirming the presence of domain gap between + +outdoor_day and indoor_flying data. Compared with the accuracy of E-RAFT with the same training set, on the one hand, our method shows better generalization ability, which generates lower EPE and $\%$ Outlier without seeing indoor_flying sequences. On the other hand, our method achieves higher accuracy on four sequences by incorporating indoor_flying sequences into training set. Overall, our method achieves a new state-of-the-art result, demonstrating the effectiveness of our proposed method. + +
splitsEPE1PE3PETime (ms)
10.7911.912.7516
30.7610.982.4518
5/70.7410.862.3030
0.7411.132.2650
+ +(a) Event splitting. Intermediate motion information contributes + +
setupEPE1PE3PETime (ms)
basic0.7410.862.3030
+v_Proj0.7510.902.3530
circular0.7611.252.4430
+ +(c) Pattern aggregation module design. + +
iterationsEPE1PE3PETime (ms)
20.8311.042.3710
40.7611.452.4316
6/80.7410.862.3030
0.7410.812.3254
+ +(e) Iterations. TMA achieves good results with a few iterations. + +
styleEPE1PE3PE
w/o0.7812.002.66
linear0.7410.862.30
same0.7611.452.48
+ +(b) Lookup style. Sampling correlation volumes with a set of linear coordinates improves the accuracy. + +
layersEPE1PE3PETime (ms)
00.7511.362.3320
1/20.7410.862.3030
0.7510.902.3238
+ +(d) Layers of pattern aggregation module. Stacking 1 pattern aggregation module showcases high accuracy. + +
radiusEPE1PE3PE
10.7511.142.37
20.7611.162.39
3/40.7410.862.30
0.7510.932.30
+ +(f) Searching radius. A radius of 3 leads to high accuracy. + +Table 4: TMA ablations. Models are trained on DSEC-Flow. Settings used in our final model are underlined. + +# 4.3. Ablations + +We conduct a series of ablation experiments to validate the proposed improvements. All ablation models are trained on DSEC-Flow and evaluated on the public benchmark. + +Event splitting. We compare different numbers of event segments in Table 4a. We choose different splits in 1, 3, 5 and 7. With the split set to 1, our model degrades into a classic correlation volume based framework and exhibits similar accuracy in comparison with E-RAFT. More splits lead to higher accuracy but lower inference efficiency, so we use 5 splits to achieve accuracy-efficiency trade-off. + +Lookup style. Based on the assumption that the optical flow changes linearly over a short period, we propose the linear lookup strategy. To validate it, we try to cancel the lookup operation or just use same sampling coordinates on intermediate correlation volumes. As Table 4b shows, sampling the correlation maps with a set of linear coordinates leads to high accuracy. + +Pattern aggregation module design. We ablate different components of pattern aggregation module in Table 4c. By default, we set the projection matrix of value $W^V$ as an identity matrix because the motion features are encoded by a motion feature encoder beforehand. In addition, we replace the pairing mode of motion features with circular pairing. Circular pairing gives equal treatment to all motion features without highlighting the importance of the last motion feature, which fails to improve accuracy. + +Layers of pattern aggregation module. We compare the accuracy with different layers of pattern aggregation modules in Table 4d. We note that stacking 1 layer of pat + +tern aggregation module showcases high accuracy. More layers may lead to difficulties in optimization, which in turn causes a decrease in accuracy. + +Iterations. Iterative learning plays a role in numerous works [18, 30]. Results in Table 4e show that 6 iterative updates reach the sweet spot of accuracy and efficiency. + +Searching radius. We compare different radii for looking up correlation volumes in Table 4f. Searching radius of 3 leads to high accuracy. Different from RAFT [30] that achieves better predictions with an increased radius, our model does not benefit a lot from a greater radius. + +# 5. Conclusion + +In this paper, we delve into event-based optical flow estimation problem. Different from existing methods that represent the event stream as static frames to adopt the classic correlation volume based pipeline, we contend that the temporal continuity is a vital element of event-based optical flow. We propose a novel Temporal Motion Aggregation (TMA) approach to unlock its potential within a learning-based framework. By incorporating temporally fine-grained motion information, TMA generates high-quality flow estimates at early stages, which not only enables TMA to obtain accurate final predictions, but also greatly reduces the demand for a number of refinements. Extensive experiments on DESC-Flow and MVSEC datasets verify the effectiveness and superiority of TMA. Meanwhile, we notice the limitations of supervised learning in generalization ability and expect further improvement to extend our idea into unsupervised learning frameworks. + +Acknowledgments: This work was supported in part by Shanghai Municipal Science and Technology Major Project under Grant 2018SHZDZX01, ZJ Lab, the Shanghai Center for Brain Science and Brain-Inspired Technology, in part by Shanghai Rising Star Program under Grant 21QC1400900, in part by Tongji-Qomolo Autonomous Driving Commercial Vehicle Joint Lab Project, and in part by Xiaomi Young Talents Program. + +# References + +[1] Himanshu Akolkar, Sio-Hoi Ieng, and Ryad Benosman. Real-time high speed motion prediction using fast aperture-robust event-driven visual flow. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):361-372, 2020. 2 +[2] Mohammed Almatrafi, Raymond Baldwin, Kiyoharu Aizawa, and Keigo Hirakawa. Distance surface for event-based optical flow. IEEE transactions on pattern analysis and machine intelligence, 42(7):1547-1556, 2020. 2 +[3] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE transactions on neural networks and learning systems, 25(2):407-417, 2013. 2 +[4] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, 2012. 2 +[5] Vincent Brebion, Julien Moreau, and Franck Davoine. Realtime optical flow for vehicular perception with low-and high-resolution event cameras. IEEE Transactions on Intelligent Transportation Systems, 2021. 7 +[6] Tobias Brosch, Stephan Tschechne, and Heiko Neumann. On event-based optical flow detection. Frontiers in neuroscience, 9:137, 2015. 2 +[7] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-temporal recurrent networks for event-based optical flow estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 525-533, 2022. 3, 5, 7 +[8] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1 +[9] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3867-3876, 2018. 2, 4 +[10] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 2, 3, 5 +[11] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cam + +eras. In 2021 International Conference on 3D Vision (3DV), pages 197-206. IEEE, 2021. 1, 2, 3, 4, 5, 6, 7 +[12] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from events and frames. arXiv preprint arXiv:2203.13674, 2022. 3 +[13] Jesse Hagenaars, Federico Paredes-Vallés, and Guido De Croon. Self-supervised learning of event-based optical flow with spiking neural networks. Advances in Neural Information Processing Systems, 34:7167-7179, 2021. 7 +[14] Markus Hofinger, Samuel Rota Bulò, Lorenzo Porzi, Arno Knapitsch, Thomas Pock, and Peter Kontschieder. Improving optical flow on a pyramid level. In European Conference on Computer Vision, pages 770-786. Springer, 2020. 2 +[15] Tak-Wai Hui and Chen Change Loy. Liteflownet3: Resolving correspondence ambiguity for more accurate optical flow estimation. In European Conference on Computer Vision, pages 169-184. Springer, 2020. 2 +[16] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8981-8989, 2018. 2 +[17] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. A lightweight optical flow cnn—revisiting data fidelity and regularization. IEEE transactions on pattern analysis and machine intelligence, 43(8):2555-2569, 2020. 2 +[18] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2019. 1, 2, 3, 8 +[19] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9772-9781, 2021. 2, 5 +[20] Chankyu Lee, Adarsh Kumar Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, and Kaushik Roy. Spikeflownet: event-based optical flow estimation with energy-efficient hybrid neural networks. In European Conference on Computer Vision, pages 366-382. Springer, 2020. 7 +[21] Min Liu and Tobi Delbruck. Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. BMVC, 2018. 2 +[22] Min Liu and Tobi Delbruck. Edflow: Event driven optical flow camera with keypoint detection and adaptive block matching. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5776-5789, 2022. 2 +[23] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2017. 6 +[24] Federico Paredes-Vallés and Guido CHE de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446-3455, 2021. 1, 3 +[25] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4161-4170, 2017. 2 + +[26] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3 +[27] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In European Conference on Computer Vision, pages 628-645. Springer, 2022. 2, 5, 6, 7 +[28] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications, volume 11006, page 1100612. International Society for Optics and Photonics, 2019. 6 +[29] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018. 1, 2 +[30] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 1, 2, 3, 5, 8 +[31] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 1 +[32] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 1 +[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5 +[34] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing, 31:7237-7251, 2022. 1, 3, 4, 6, 7 +[35] Haofei Xu, Jiaolong Yang, Jianfei Cai, Juyong Zhang, and Xin Tong. High-resolution optical flow from 1d attention and correlation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10498-10507, 2021. 2 +[36] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 2, 6 +[37] Gengshan Yang and Deva Ramanan. Volumetric correspondence networks for optical flow. Advances in neural information processing systems, 32:794-805, 2019. 2 +[38] Chengxi Ye, Anton Mitrokhin, Cornelia Fermüller, James A Yorke, and Yiannis Aloimonos. Unsupervised learning of + +dense optical flow, depth and egomotion with event-based sensors. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5831-5838. IEEE, 2020. 1, 2, 7 +[39] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 185-194, 2019. 2 +[40] Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In European Conference on Computer Vision, pages 420-439. Springer, 2020. 2 +[41] Feihu Zhang, Oliver J Woodford, Victor Adrian Prisacariu, and Philip HS Torr. Separable flow: Learning motion cost volumes for optical flow estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10807-10817, 2021. 2 +[42] Xu Zheng, Yexin Liu, Yunfan Lu, Tongyan Hua, Tianbo Pan, Weiming Zhang, Dacheng Tao, and Lin Wang. Deep learning for event-based vision: A comprehensive survey and benchmarks. arXiv preprint arXiv:2302.08890, 2023. 1 +[43] Alex Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In Proceedings of Robotics: Science and Systems, Pittsburgh, Pennsylvania, June 2018, 2, 5, 6, 7 +[44] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018. 2, 3, 5, 7 +[45] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 1, 2, 3, 4, 5, 7 \ No newline at end of file diff --git a/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/images.zip b/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b33a22d698a310e131d0e3d4a72c78ddb87bea21 --- /dev/null +++ b/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efb40599d3221f8b189c81544d7b8e2606b0de68929d0fe1fbcb6d590cebec56 +size 583507 diff --git a/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/layout.json b/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0c5641b29ac90c25faa5950216982c30253889d2 --- /dev/null +++ b/2023/TMA_ Temporal Motion Aggregation for Event-based Optical Flow/layout.json @@ -0,0 +1,8740 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 504, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 504, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 504, + 121 + ], + "type": "text", + "content": "TMA: Temporal Motion Aggregation for Event-based Optical Flow" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 141, + 471, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 141, + 471, + 185 + ], + "spans": [ + { + "bbox": [ + 122, + 141, + 471, + 185 + ], + "type": "text", + "content": "Haotian Liu1 Guang Chen1* Sanqing Qu1 Yanping Zhang1 \nZhijun Li1 Alois Knoll2 Changjun Jiang1 \n1Tongji University 2Technical University of Munich" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 574 + ], + "type": "text", + "content": "Event cameras have the ability to record continuous and detailed trajectories of objects with high temporal resolution, thereby providing intuitive motion cues for optical flow estimation. Nevertheless, most existing learning-based approaches for event optical flow estimation directly remould the paradigm of conventional images by representing the consecutive event stream as static frames, ignoring the inherent temporal continuity of event data. In this paper, we argue that temporal continuity is a vital element of event-based optical flow and propose a novel Temporal Motion Aggregation (TMA) approach to unlock its potential. Technically, TMA comprises three components: an event splitting strategy to incorporate intermediate motion information underlying the temporal context, a linear lookup strategy to align temporally fine-grained motion features and a novel motion pattern aggregation module to emphasize consistent patterns for motion feature enhancement. By incorporating temporally fine-grained motion information, TMA can derive better flow estimates than existing methods at early stages, which not only enables TMA to obtain more accurate final predictions, but also greatly reduces the demand for a number of refinements. Extensive experiments on DSEC-Flow and MVSEC datasets verify the effectiveness and superiority of our TMA. Remarkably, compared to E-RAFT, TMA achieves a " + }, + { + "bbox": [ + 46, + 238, + 290, + 574 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 574 + ], + "type": "text", + "content": " improvement in accuracy and a " + }, + { + "bbox": [ + 46, + 238, + 290, + 574 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 574 + ], + "type": "text", + "content": " reduction in inference time on DSEC-Flow. Code will be available at https://github.com/ispc-lab/TMA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 590, + 128, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 128, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 128, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 610, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 610, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 610, + 287, + 694 + ], + "type": "text", + "content": "Optical flow aims to compute the velocity of objects on the image plane without geometry prior and is a fundamental topic in event-based vision [8, 42]. It plays an important role in many applications, such as ego-motion estimation [38, 45], image reconstruction [24], and video frame interpolation [32, 31]. Recently, learning-based methods [29, 18, 30] have dominated frame-based optical flow" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 205, + 425, + 275 + ], + "blocks": [ + { + "bbox": [ + 310, + 205, + 425, + 275 + ], + "lines": [ + { + "bbox": [ + 310, + 205, + 425, + 275 + ], + "spans": [ + { + "bbox": [ + 310, + 205, + 425, + 275 + ], + "type": "image", + "image_path": "d79da3f60da344d5debf26737a7222933e2d1633fbf35168e314c06375fb14cc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 275, + 401, + 285 + ], + "lines": [ + { + "bbox": [ + 335, + 275, + 401, + 285 + ], + "spans": [ + { + "bbox": [ + 335, + 275, + 401, + 285 + ], + "type": "text", + "content": "(a) reference image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 426, + 205, + 542, + 275 + ], + "blocks": [ + { + "bbox": [ + 426, + 205, + 542, + 275 + ], + "lines": [ + { + "bbox": [ + 426, + 205, + 542, + 275 + ], + "spans": [ + { + "bbox": [ + 426, + 205, + 542, + 275 + ], + "type": "image", + "image_path": "76e22e787cc0093cd70f73f2ae77508cbad73d691a23df358884f3ee86f683c7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 275, + 510, + 284 + ], + "lines": [ + { + "bbox": [ + 458, + 275, + 510, + 284 + ], + "spans": [ + { + "bbox": [ + 458, + 275, + 510, + 284 + ], + "type": "text", + "content": "(b) event frame" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 284, + 425, + 354 + ], + "blocks": [ + { + "bbox": [ + 310, + 284, + 425, + 354 + ], + "lines": [ + { + "bbox": [ + 310, + 284, + 425, + 354 + ], + "spans": [ + { + "bbox": [ + 310, + 284, + 425, + 354 + ], + "type": "image", + "image_path": "197e01e0c9ab95123c67414acdb7f28e2441de1741495e0250013e52ab5bb6cf.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 311, + 354, + 425, + 422 + ], + "blocks": [ + { + "bbox": [ + 311, + 354, + 425, + 422 + ], + "lines": [ + { + "bbox": [ + 311, + 354, + 425, + 422 + ], + "spans": [ + { + "bbox": [ + 311, + 354, + 425, + 422 + ], + "type": "image", + "image_path": "71b7e75eee9838c46a76b72c51fcaf25d19f44bae976519f476782507c6b6319.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 422, + 536, + 432 + ], + "lines": [ + { + "bbox": [ + 307, + 422, + 536, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 536, + 432 + ], + "type": "text", + "content": "(c) E-RAFT output & motion feature (d) our output & motion feature" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 427, + 284, + 542, + 354 + ], + "blocks": [ + { + "bbox": [ + 427, + 284, + 542, + 354 + ], + "lines": [ + { + "bbox": [ + 427, + 284, + 542, + 354 + ], + "spans": [ + { + "bbox": [ + 427, + 284, + 542, + 354 + ], + "type": "image", + "image_path": "2cddd66dab12422f7745cde39ca77fb477f80355b39094fea5f61fc9c634c840.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 427, + 354, + 541, + 422 + ], + "blocks": [ + { + "bbox": [ + 427, + 354, + 541, + 422 + ], + "lines": [ + { + "bbox": [ + 427, + 354, + 541, + 422 + ], + "spans": [ + { + "bbox": [ + 427, + 354, + 541, + 422 + ], + "type": "image", + "image_path": "b1b82bcbcfe65b05eb791f6404309750ddc92df34c88fc0a65b6383fd574f6c3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 567 + ], + "type": "text", + "content": "Figure 1: Accuracy illustrations. (a) Reference image from DSEC-Flow. (b) Corresponding event frame. (c) The final flow prediction and visualization of the motion feature from E-RAFT [11]. (d) The final flow prediction and visualization of the motion feature from our proposed method. We visualize the motion feature at the initial stage by taking the average across channels. E-RAFT fails to generate informative motion features and results in blurred boundaries. In contrast, our method utilizes temporally fine-grained motion information to address the information scarcity issue in motion features, generating high-quality predictions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "by employing correlation volumes to derive motion features for flow regression. Inspired by this, several event-based approaches [11, 34] adopt a similar paradigm by converting the consecutive event stream into grid-like representations. Despite encouraging progress, these methods still suffer from non-negligible limitations, due to the great differences between event data and conventional images. Distinct from dense and colorful conventional (RGB) images, event data features spatial sparsity and a lack of intensity information. It is prone to result in close matching scores and invalid regions in the correlation volume. Consequently, less informative motion features are derived and regress inaccurate" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 220, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 220, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 220, + 713 + ], + "type": "text", + "content": "*corresponding author: guangchen@tongii.edu.cn" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9685" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 284, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 284, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 284, + 84 + ], + "type": "text", + "content": "predictions. Figure 1 illustrates this issue in E-RAFT [11]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 286, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 286, + 216 + ], + "type": "text", + "content": "As the saying goes, each coin has two sides. Though event cameras encode visual information in a sparse manner, they are capable of capturing continuous and detailed object trajectories with high temporal resolution, thereby providing rich motion cues for optical flow estimation. Several model-based methods [9, 27] benefit from the temporal continuity by warping events along point trajectories, achieving decent performance. Drawing on this observation, we contend that the temporal continuity is a vital element of event-based optical flow and propose to unleash its potential within a learning-based framework." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 217, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 539 + ], + "type": "text", + "content": "To materialize our idea, we propose a novel Temporal Motion Aggregation (TMA) approach to explore the inherent temporal continuity of event data. Technically, we revamp the dominant learning-based architectures with three distinct components. We first introduce an event splitting strategy for temporally-dense correlation volumes computation. By splitting the event stream into multiple segments and extracting their features, feature similarities are compared between the first feature and all others, which record rich intermediate motion information. Aware of that correlation volumes record motions between different time spans, we then design a linear lookup strategy to sample each correlation volume based on the corresponding flow estimate to encode motion features. As a result, fine-grained motions are warped to the same pixels across motion features. Furthermore, we consider the incorrect motion patterns in intermediate motion features due to the manual lookup. Therefore, we propose a novel pattern aggregation module for motion feature enhancement by aggregating consistent patterns between each intermediate motion feature and the last one (without manual lookup). Thanks to the incorporation of temporally fine-grained motion information, TMA can effectively address the information scarcity in motion features and thus generate good flow estimates at early stages, which not only enable TMA to obtain accurate flow predictions, but also greatly reduces the demand for a number of time-consuming refinements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 540, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 287, + 575 + ], + "type": "text", + "content": "We evaluate our TMA on DSEC-Flow [10] and MVSEC [44] datasets. Extensive experiments demonstrate the superior advantages of TMA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 576, + 265, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 576, + 265, + 588 + ], + "spans": [ + { + "bbox": [ + 59, + 576, + 265, + 588 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 594, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 58, + 594, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 594, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 58, + 594, + 287, + 628 + ], + "type": "text", + "content": "- We argue that the temporal continuity of event data is a vital element of event-based optical flow and propose to unlock its potential in a learning-based framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 635, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 635, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 58, + 635, + 287, + 683 + ], + "type": "text", + "content": "- We propose a novel Temporal Motion Aggregation (TMA) approach, which comprises three components: an event splitting strategy, a linear lookup strategy and a motion pattern aggregation module." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 287, + 713 + ], + "type": "text", + "content": "- The incorporation of temporally fine-grained motion information enables TMA to achieve high accuracy" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 545, + 107 + ], + "type": "text", + "content": "while maintaining high efficiency. Compared with E-RAFT, TMA achieves a " + }, + { + "bbox": [ + 325, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 325, + 72, + 545, + 107 + ], + "type": "text", + "content": " improvement in accuracy and a " + }, + { + "bbox": [ + 325, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 325, + 72, + 545, + 107 + ], + "type": "text", + "content": " reduction in inference time on DSEC-Flow." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 119, + 392, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 119, + 392, + 131 + ], + "spans": [ + { + "bbox": [ + 306, + 119, + 392, + 131 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 140, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 140, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 545, + 305 + ], + "type": "text", + "content": "Learning methods for optical flow estimation. Recently, correlation volume based approaches dominate the learning-based optical flow estimation [25, 17, 29, 18, 37, 14]. As a milestone, PwCNet [29] summarizes a simple principle in architecture construction: pyramidal processing, warping and correlation volumes computation. Following the well-established pipeline, LiteFlowNet series [16, 17, 15] propose cascaded feature pyramids with a lightweight convolutional burden. VCN [37] proposes 4D volumetric U-Net architectures to enlarge the receptive fields for correlation volumes. IRR [18] notes the reusability of pyramidal refinement for flow update and proposes a shared-weights refiner across all scales, which can be combined with mainstream learning methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 307, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 486 + ], + "type": "text", + "content": "However, the correlation volumes used in previous works are computed locally, which cannot fully deal with long-term motions. RAFT [30] firstly introduces the full correlation volume to capture global visual similarities. Separable Flow [41] proposes a separable correlation volume module as a drop-in replacement to correlation volumes and uses non-local aggregation layers [39, 40] to refine the correlation volumes efficiently. Besides, transformer components show promising performance on optical flow [36, 19, 35]. GMA [19] notices the underlying motion clues in the context and introduces an attention layer to aggregate hidden motions, improving the accuracy based on RAFT architecture. GMFlow [36] reformulates optical flow as a global matching problem by attention blocks, suggesting a new paradigm for flow estimation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 487, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 556 + ], + "type": "text", + "content": "We draw inspiration from GMA [19] to design our motion pattern aggregation module. In contrast to GMA which aggregates hidden motion patterns in the spatial context for motion feature enhancement, we extend the pattern aggregation to the temporal dimension and leverage feature cross-similarities to enhance the motion feature." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": "Event-based optical flow estimation. A number of traditional works remould frame-based approaches into event-based optical flow estimation. Benosman et al. [4] extends the classic Lucas-Kanade method to event-based optical flow. Liu et al. [21, 22] implement event-based corner keypoint detection and novel block-matching optical flow. Gallego et al. [9] propose a unifying contrast maximization (CM) framework and MultiCM [27] extends it to complex scenes. Besides, plane fitting [1, 3], filter banks [6] and time surface matching [2] are proposed to boost performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "More recently, learning-based methods have dominated frame-based optical flow, inspiring attempts at event-based optical flow. Early learning methods [43, 45, 38] represent" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9686" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 541, + 189 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 541, + 189 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 541, + 189 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 541, + 189 + ], + "type": "image", + "image_path": "bd2575db2ab1f802b8ae96be3731aacc30879bf0446c7a6eb2bb30f9a195f1ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 191, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 191, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 547, + 277 + ], + "type": "text", + "content": "Figure 2: The overall architecture of TMA. First, we split the event stream into segments and extract features with a share-weights feature extractor. Then, regarding the first feature as reference, we correlate all other features to compute temporally-dense correlation volumes, which contain rich embedded motion information. Next, aware of the different time spans of correlation volumes, we look up each correlation volume based on a corresponding flow estimate. All sampled correlation maps are delivered into a share-weights motion feature encoder to derive motion features. We further propose a novel motion pattern aggregation module to enrich the spatial information of motion features. By concatenating the enhanced motion features and utilizing several refinements, high-quality flow predictions are generated." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 283, + 289, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 289, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 289, + 450 + ], + "type": "text", + "content": "the event stream as a static frame and utilize the architecture of U-Net [26] to predict optical flow, which can only address small motions. Mainstream works [7, 34, 11] introduce two consecutive event frames and follow the correlation volume based pipeline (PWCNet and RAFT), achieving higher accuracy on large-scales datasets [44, 10]. E-RAFT [11, 12] extends RAFT [30] and introduces full correlation volume into event-based vision. DCEIFlow [34] combines conventional images and event data to boost higher accuracy. Note that our work and STE-FlowNet [7] share a similar motivation to introduce temporal continuity with fine-grained discretization. However, STE-FlowNet takes a frame-by-frame manner [18] to update flow, leading to an insufficient inference process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 451, + 288, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 451, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 451, + 288, + 523 + ], + "type": "text", + "content": "Contrary to previous learning-based methods, we do not merely adopt the classic correlation volume based pipeline. Instead, we regard the temporal continuity of event data as a vital component of optical flow and incorporate the intermediate motion information into a learning-based framework, which boosts higher accuracy and efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 536, + 130, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 130, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 130, + 551 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": "Given an input event stream " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}\\}" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": ", the goal of optical flow is to estimate a dense displacement field " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{u}:\\mathbb{R}^2\\to \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " which maps each per-pixel " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " to its corresponding coordinate " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "(x^{\\prime},y^{\\prime})" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": ". To address this, mainstream learning-based approaches compute correlation volumes from two event frames as a matching prior. Such principle inspires several methods [11, 34] by squeezing event data into consecutive frames. Unfortunately, the characteristics of event data, spatial sparsity and lack of intensity records, are prone to result in close matching scores and invalid regions in the correlation volume. Consequently, less informative motion features are derived and regress inaccurate predictions. Here we address the information scarcity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 330 + ], + "type": "text", + "content": "of motion features from the attractive temporal continuity of event data. Temporally fine-grained motion information in event data is sufficient to supplement the spatial richness of motion features, yielding high-quality flow predictions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 331, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 366 + ], + "type": "text", + "content": "In the following, we first describe the preliminary and event representation, and then present the detailed architecture of the proposed TMA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 373, + 509, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 509, + 385 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 509, + 385 + ], + "type": "text", + "content": "3.1. Preliminary and Event Representation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 391, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 475 + ], + "type": "text", + "content": "Event cameras record light intensity changes in an asynchronous way. Each pixel serves as an independent trigger and generates an event instantly whenever the log intensity changes exceeding a threshold " + }, + { + "bbox": [ + 304, + 391, + 545, + 475 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 391, + 545, + 475 + ], + "type": "text", + "content": ". An event " + }, + { + "bbox": [ + 304, + 391, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\boldsymbol{e}_k = (\\boldsymbol{x}_k, p_k, t_k)" + }, + { + "bbox": [ + 304, + 391, + 545, + 475 + ], + "type": "text", + "content": " is a triplet containing triggered pixel location " + }, + { + "bbox": [ + 304, + 391, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_k = (x_k, y_k)^T" + }, + { + "bbox": [ + 304, + 391, + 545, + 475 + ], + "type": "text", + "content": ", timestamp " + }, + { + "bbox": [ + 304, + 391, + 545, + 475 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 304, + 391, + 545, + 475 + ], + "type": "text", + "content": " and signed intensity change " + }, + { + "bbox": [ + 304, + 391, + 545, + 475 + ], + "type": "inline_equation", + "content": "(p_k = \\pm 1)" + }, + { + "bbox": [ + 304, + 391, + 545, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 475, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 535 + ], + "type": "text", + "content": "To ease the feature extraction and preserve the temporal continuity of the event data, we firstly transform the event stream into a 3D volume " + }, + { + "bbox": [ + 304, + 475, + 545, + 535 + ], + "type": "inline_equation", + "content": "V(x,y,t)" + }, + { + "bbox": [ + 304, + 475, + 545, + 535 + ], + "type": "text", + "content": " along the temporal dimension following existing works [45, 11, 24]. Given an event stream " + }, + { + "bbox": [ + 304, + 475, + 545, + 535 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}\\}" + }, + { + "bbox": [ + 304, + 475, + 545, + 535 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 475, + 545, + 535 + ], + "type": "inline_equation", + "content": "V(x,y,t)" + }, + { + "bbox": [ + 304, + 475, + 545, + 535 + ], + "type": "text", + "content": " is generated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 541, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 541, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 315, + 541, + 545, + 555 + ], + "type": "interline_equation", + "content": "t _ {i} ^ {*} = (B - 1) \\left(t _ {i} - t _ {1}\\right) / \\left(t _ {N _ {e}} - t _ {1}\\right) \\tag {1}", + "image_path": "d7f1722553ea09e7f1f6c5348946cdc09472532805fdcb70fc2e65d0941b1268.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 557, + 545, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 557, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 315, + 557, + 545, + 581 + ], + "type": "interline_equation", + "content": "\\boldsymbol {V} (x, y, t) = \\sum_ {i} p _ {i} k _ {b} \\left(x - x _ {i}\\right) k _ {b} \\left(y - y _ {i}\\right) k _ {b} \\left(t - t _ {i} ^ {*}\\right) \\tag {2}", + "image_path": "656bd5e4a05166a56b1193016fe41647b31363b818c59c1893ade04a98fd078d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 585, + 545, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 585, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 315, + 585, + 545, + 597 + ], + "type": "interline_equation", + "content": "k _ {b} (a) = \\max (0, 1 - | a |), \\tag {3}", + "image_path": "9330590aec7aeaa1837c2ff8e6e4f2819c6c360da8ff25698bafeef554e3c790.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "inline_equation", + "content": "t_i^*" + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": "-th time bin. " + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "inline_equation", + "content": "N_e" + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": " represent the length of time bins and the number of events, respectively. " + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "inline_equation", + "content": "k_b(a)" + }, + { + "bbox": [ + 304, + 605, + 545, + 641 + ], + "type": "text", + "content": " is the bilinear interpolation function." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 647, + 416, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 416, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 416, + 658 + ], + "type": "text", + "content": "3.2. TMA Architecture" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": "We develop our framework based on the successful RAFT architecture [30]. Figure 2 summarizes the architecture of the proposed TMA, which consists of three components, an event splitting strategy, a linear lookup strategy" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9687" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 69, + 272, + 194 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 272, + 194 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 272, + 194 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 272, + 194 + ], + "type": "image", + "image_path": "8da1b9b2f86637a5772622540d3d8a83337ccf577b5648551fa3ce8ccecb5f51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 197, + 287, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 197, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 287, + 257 + ], + "type": "text", + "content": "Figure 3: Linear lookup strategy. We compute correlation volumes by comparing the similarity of first feature with all others, which are misaligned due to different time spans. To rectify this, we narrow down the lookup coordinates linearly to warp correlation maps to spatially aligned pixels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 263, + 287, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 310 + ], + "type": "text", + "content": "and a motion pattern aggregation module. With the proposed method, we successfully unlock the potential of temporal continuity to predict event-based optical flow within a learning-based framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 312, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 441 + ], + "type": "text", + "content": "Event splitting. Instead of viewing the continuous event stream as a static frame which tends to be a matter of common knowledge in recent works [45, 11, 34], we prefer to identify it as a high frame rate video clip, which records continuous and detailed trajectories of objects, namely, provides intuitive motion cues for optical flow estimation. To introduce rich intermediate motion information into our method, we first propose an event splitting strategy, which splits the event stream into multiple segments. Therefore, more intermediate event frames are retained for correlation volume computation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": "In particular, given an event stream " + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}\\}" + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": ", we split it into " + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": " segments equally and the time span between adjacent segments is " + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "dt" + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": ". Besides, we introduce an auxiliary event segment from " + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "t_0 - dt" + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": " as reference. Then, " + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "g + 1" + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": " event representations are generated following Sec. 3.1. All " + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "g + 1" + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": " event representations are then delivered into a share-weights encoder to achieve features " + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i \\in \\mathbb{R}^{H \\times W \\times C}, i = 0, 1, 2, \\dots, g," + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": " represent the height, width and channel of features. We use all features to compute temporally-dense correlation volumes " + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\pmb{C}_i, i = 1, 2, \\dots, g," + }, + { + "bbox": [ + 47, + 443, + 287, + 574 + ], + "type": "text", + "content": " which are given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 582, + 287, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 582, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 73, + 582, + 287, + 609 + ], + "type": "interline_equation", + "content": "\\boldsymbol {C} _ {i} = \\frac {\\boldsymbol {F} _ {0} \\boldsymbol {F} _ {i} ^ {T}}{\\sqrt {D}} \\in \\mathbb {R} ^ {H \\times W \\times H \\times W}, i = 1, 2, \\dots , g, \\tag {4}", + "image_path": "8a7c3df061bc251d227600f4528e3b31868bfec3086de2483eb743e9a272078a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\pmb{F}_0" + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "text", + "content": " is the first feature and " + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\pmb{F}_i" + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "text", + "content": "-th feature derived from the " + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 617, + 287, + 640 + ], + "type": "text", + "content": "-th event frame." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 642, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 700 + ], + "type": "text", + "content": "Temporally-dense correlation volumes compare feature similarities across multiple time spans, enabling refined records of relative motions. This construction is vastly different from the classic full correlation volume that only contains feature similarities between a fixed time span." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "Linear lookup. Recall that each correlation volume " + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "inline_equation", + "content": "C_i" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 527, + 199 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 527, + 199 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 527, + 199 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 527, + 199 + ], + "type": "image", + "image_path": "edd9caa0b9361b83c8177c5cf45bb2de7ef9eea9cdead27b15955c2036afefbb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 274 + ], + "type": "text", + "content": "Figure 4: Motion pattern aggregation. We pair each intermediate motion feature to the last motion feature. Then each pair is sent to cross-attention blocks to aggregate consistent patterns. Hence, each intermediate motion feature carries information from the last one and consistent patterns receive emphasis." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "type": "text", + "content": "is achieved by correlating " + }, + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "type": "text", + "content": ", which means the recorded relation motions are misaligned in time. While our goal is to predict optical flow from " + }, + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "type": "inline_equation", + "content": "t_{0}" + }, + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "type": "inline_equation", + "content": "t_{1}" + }, + { + "bbox": [ + 304, + 278, + 545, + 386 + ], + "type": "text", + "content": ", we hope that all sampled correlation maps could present the relative motions within the same time span. Therefore, based on that correlation volumes are evenly constructed in time and magnitudes of optical flow change linearly over a short period, we narrow down lookup coordinates in a linear fashion to match the timestamp of each correlation volume." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 386, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 434 + ], + "type": "text", + "content": "Figure 3 displays the linear lookup strategy. In detail, given current flow estimate " + }, + { + "bbox": [ + 304, + 386, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 304, + 386, + 545, + 434 + ], + "type": "text", + "content": ", we split it into " + }, + { + "bbox": [ + 304, + 386, + 545, + 434 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 386, + 545, + 434 + ], + "type": "text", + "content": " segments to sample each correlation volume. The linear lookup operation can be expressed as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 441, + 545, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 441, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 321, + 441, + 545, + 453 + ], + "type": "interline_equation", + "content": "\\mathbf {d} \\mathbf {u} = \\left(\\mathbf {c o o r d s} _ {1} - \\mathbf {c o o r d s} _ {0}\\right) / g = \\mathbf {u} / g \\tag {5}", + "image_path": "00945a535fb0c6761f6406e3a9dd6436aed045d54a326c07d6af9ae29654cac1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 455, + 545, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 455, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 321, + 455, + 545, + 468 + ], + "type": "interline_equation", + "content": "\\operatorname {c o r r} _ {i} = S \\left(\\boldsymbol {C} _ {i}, \\operatorname {c o o r d s} _ {0} + i \\mathrm {d u}\\right), i = 1, 2, \\dots , g, \\tag {6}", + "image_path": "38f03c203e97715ff55c95f28faf19e813b4a1d38fb4425abd883c796d74847d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{coord}_0" + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "text", + "content": " is the initialized coordinates at " + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "text", + "content": " (reference time), " + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "inline_equation", + "content": "S(\\cdot, \\cdot)" + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "text", + "content": " represents the lookup operation, " + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{coord}_i" + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "text", + "content": " are lookup coordinates for " + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{corr}_i" + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "text", + "content": " is the correlation map sampled from " + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 304, + 474, + 545, + 544 + ], + "type": "text", + "content": ". After delivering all correlation maps into a share-weights motion feature encoder, aligned motion features are achieved." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 546, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 629 + ], + "type": "text", + "content": "From the perspective of image contrast [9], the standard lookup is responsible for warping possible matching points between two images to the same pixel. Our linear lookup extends it by warping matching points of a temporally fine-grained motion to the same pixel. As a result, similar motions are presented at same pixels across correlation maps, showing high contrast." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Motion pattern aggregation. A standard operation for next step is to concatenate the motion features as a whole to update flow and the improvements made so far already alleviate the information scarcity in motion features. However, we notice that the concatenation operation treats all motion features equally, which means the majority of motion features, intermediate ones, dominate the flow regression. Due" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9688" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 232, + 233 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 232, + 233 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 232, + 233 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 232, + 233 + ], + "type": "image", + "image_path": "4cc613ce347d8063a08a953b1258344670e8f9d30b3ff8cbf420ebacd78e806e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 546, + 261 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 546, + 261 + ], + "type": "text", + "content": "Figure 5: Qualitative examples of DSEC-Flow. For each two columns, we show the reference image and event frame and compare our method with a state-of-the-art baseline E-RAFT [11]. Significant improvements are highlighted by red boxes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 244, + 70, + 388, + 233 + ], + "blocks": [ + { + "bbox": [ + 244, + 70, + 388, + 233 + ], + "lines": [ + { + "bbox": [ + 244, + 70, + 388, + 233 + ], + "spans": [ + { + "bbox": [ + 244, + 70, + 388, + 233 + ], + "type": "image", + "image_path": "c9d09ce8f6abd1e8a67f68c318998a498a71fed2457dfd334adf1bcfb3bfa69c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 399, + 71, + 541, + 233 + ], + "blocks": [ + { + "bbox": [ + 399, + 71, + 541, + 233 + ], + "lines": [ + { + "bbox": [ + 399, + 71, + 541, + 233 + ], + "spans": [ + { + "bbox": [ + 399, + 71, + 541, + 233 + ], + "type": "image", + "image_path": "2a660dd95348afbefb349d5a459a791499afdd118c8ebd0d7b0e307863e2f783.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 270, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 365 + ], + "type": "text", + "content": "to the linear optical flow assumption and manual lookup operation, these motion features are not perfectly correct. A natural thought is to put more emphasis on the last motion feature that is without manual lookup and retain the patterns in intermediate motion features that are consistent with the last motion feature. To this end, we balance the weights between the last motion feature and others through cross-similarities computed by transformers [33, 19]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 366, + 288, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 288, + 450 + ], + "type": "text", + "content": "Figure 4 shows the overview of motion pattern aggregation module. Given a set of motion features " + }, + { + "bbox": [ + 47, + 366, + 288, + 450 + ], + "type": "inline_equation", + "content": "M\\mathbf{F}_i \\in \\mathbb{R}^{H \\times W \\times D'}" + }, + { + "bbox": [ + 47, + 366, + 288, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 366, + 288, + 450 + ], + "type": "inline_equation", + "content": "i = 1,2,\\dots,g" + }, + { + "bbox": [ + 47, + 366, + 288, + 450 + ], + "type": "text", + "content": ", we pair each intermediate motion feature to the last one. Then each pair " + }, + { + "bbox": [ + 47, + 366, + 288, + 450 + ], + "type": "inline_equation", + "content": "(M\\mathbf{F}_i,M\\mathbf{F}_g)" + }, + { + "bbox": [ + 47, + 366, + 288, + 450 + ], + "type": "text", + "content": " is sent to cross-attention blocks to compare cross-similarities and aggregate consistent patterns. The enhanced motion feature " + }, + { + "bbox": [ + 47, + 366, + 288, + 450 + ], + "type": "inline_equation", + "content": "\\widehat{M}\\mathbf{F}_i" + }, + { + "bbox": [ + 47, + 366, + 288, + 450 + ], + "type": "text", + "content": " is obtained as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 459, + 287, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 459, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 59, + 459, + 287, + 491 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {A} = \\operatorname {A t t e n t i o n} \\left(\\boldsymbol {M} \\boldsymbol {F} _ {i} W ^ {Q}, \\boldsymbol {M} \\boldsymbol {F} _ {g} W ^ {K}, \\boldsymbol {M} \\boldsymbol {F} _ {g} W ^ {V}\\right) (7) \\\\ \\widehat {\\boldsymbol {M F}} _ {i} = \\boldsymbol {M F} _ {i} + \\operatorname {M L P} \\left(\\left[ \\boldsymbol {M F} _ {i}, \\boldsymbol {A W} ^ {O} \\right]\\right), (8) \\\\ \\end{array}", + "image_path": "8bce5df543981633bd654ad4c888794fc488203f5b764a7fba6e11e40058bd55.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\operatorname{Attention}(Q, K, V) = \\operatorname{softmax}\\left(\\frac{QK^T}{\\sqrt{d_k}}\\right)V" + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "inline_equation", + "content": "d_k" + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "text", + "content": " is the dimension of " + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "inline_equation", + "content": "W^Q" + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "inline_equation", + "content": "W^K" + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "inline_equation", + "content": "W^V" + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "text", + "content": " are projection matrices. " + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}(\\cdot)" + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "text", + "content": " is multi-layer perceptron and " + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "inline_equation", + "content": "W^O" + }, + { + "bbox": [ + 46, + 501, + 287, + 592 + ], + "type": "text", + "content": " is another projection matrix. We omit the normalization functions in the manuscript for simplicity. Motion pattern aggregation module does not change the dimension of motion features, so multiple cross-attention layers can be stacked easily." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 593, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 653 + ], + "type": "text", + "content": "Once motion features are enhanced by the motion pattern aggregation module, each " + }, + { + "bbox": [ + 46, + 593, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\widehat{MF}_i" + }, + { + "bbox": [ + 46, + 593, + 287, + 653 + ], + "type": "text", + "content": " carries motion patterns that are consistent with " + }, + { + "bbox": [ + 46, + 593, + 287, + 653 + ], + "type": "inline_equation", + "content": "MF_{g}" + }, + { + "bbox": [ + 46, + 593, + 287, + 653 + ], + "type": "text", + "content": ". Namely, the last motion feature receives high attention and consistent patterns in motion features are fully utilized." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Above three components are seamlessly integrated into a unified pipeline: incorporating intermediate motion information, warping matching motions and emphasizing consistent motion patterns. Attributed to the innovative application of temporal continuity to a learning-based frame" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 342 + ], + "type": "text", + "content": "work, our proposed method is capable of overcoming the information scarcity issue in motion features and generating high-quality flow estimates even at an early refinement stage. This enables our method to achieve accurate final predictions, while reducing the demand for a number of refinements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 349, + 384, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 349, + 384, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 349, + 384, + 361 + ], + "type": "text", + "content": "3.3. Supervision" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 367, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 367, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 305, + 367, + 545, + 403 + ], + "type": "text", + "content": "Following RAFT [30], we choose " + }, + { + "bbox": [ + 305, + 367, + 545, + 403 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 305, + 367, + 545, + 403 + ], + "type": "text", + "content": " distance between the predictions and ground-truth as the supervision signal for our model. The loss function " + }, + { + "bbox": [ + 305, + 367, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 305, + 367, + 545, + 403 + ], + "type": "text", + "content": " is formally defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 411, + 545, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 411, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 365, + 411, + 545, + 445 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {j = 1} ^ {N} \\gamma^ {N - j} \\left\\| \\mathbf {u} ^ {g t} - \\mathbf {u} ^ {j} \\right\\| _ {1}, \\tag {9}", + "image_path": "bf9919f19e242f2798ea0fae2602fe80f0be264fb17993a6e63fac3b4ca374cc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 455, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 455, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{u}^{gt}" + }, + { + "bbox": [ + 304, + 455, + 545, + 491 + ], + "type": "text", + "content": " denotes the flow ground-truth and " + }, + { + "bbox": [ + 304, + 455, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{u}^j" + }, + { + "bbox": [ + 304, + 455, + 545, + 491 + ], + "type": "text", + "content": " the flow prediction at " + }, + { + "bbox": [ + 304, + 455, + 545, + 491 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 455, + 545, + 491 + ], + "type": "text", + "content": "-th stage. " + }, + { + "bbox": [ + 304, + 455, + 545, + 491 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 455, + 545, + 491 + ], + "type": "text", + "content": " is the total refinement stages. The weight " + }, + { + "bbox": [ + 304, + 455, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 455, + 545, + 491 + ], + "type": "text", + "content": " balances different stages of predictions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 502, + 388, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 388, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 388, + 515 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "Datasets and evaluation setup. Following previous works [11, 45, 7], we provide experimental details and extensive comparison results on two event-based datasets DSEC-Flow [10] and MVSEC [44]. For DSEC-Flow, we train models on the official training set and evaluate DSEC-Flow test set on the public benchmark. For MVSEC, we conduct two kinds of experiments. Basically, following previous methods [43, 27], we train methods on outdoor_day2 sequence with ground-truth corresponding to a time interval of " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "dt = 1" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "dt = 4" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": " gray images and evaluate 800 frames of outdoor_day1 and three indoor_flying sequences. Equally importantly, considering the supervised learning setting of our method and the large gap between outdoor_day and indoor_flying data, we extend the range of training data by including one indoor_flying sequence each time to further boost the accuracy of supervised learning methods." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9689" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 284, + 138 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 284, + 138 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 284, + 138 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 284, + 138 + ], + "type": "table", + "html": "
MethodsEPE1PE3PETime(ms)
MBMultiCM [27]3.4776.630.9-
SLEV-FlowNet†[43]2.3255.418.65
E-RAFT [11]0.7912.52.752
Ours0.7410.92.330
", + "image_path": "c56e5dad9054712f0f502bd3edfa29b45ea2e49357da9e742db3436340291a59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 143, + 287, + 192 + ], + "lines": [ + { + "bbox": [ + 46, + 143, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 143, + 287, + 192 + ], + "type": "text", + "content": "Table 1: Results on DSEC-Flow. Best accuracy in bold. " + }, + { + "bbox": [ + 46, + 143, + 287, + 192 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 143, + 287, + 192 + ], + "type": "text", + "content": " denotes the result is taken from E-RAFT [11]. Model-based (MB) methods need no training data; and supervised learning (SL) methods need ground-truth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 198, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 327 + ], + "type": "text", + "content": "Metrics. The end-point-error (EPE) is used as core metric for prediction accuracy for both DSEC-Flow and MVSEC. To evaluate the robustness against large displacements, we compute the percentage of pixels with EPE greater than " + }, + { + "bbox": [ + 46, + 198, + 287, + 327 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 198, + 287, + 327 + ], + "type": "text", + "content": " pixels " + }, + { + "bbox": [ + 46, + 198, + 287, + 327 + ], + "type": "inline_equation", + "content": "(M\\mathrm{PE},M = 1,3)" + }, + { + "bbox": [ + 46, + 198, + 287, + 327 + ], + "type": "text", + "content": " for DSEC-Flow. Percentage of pixels with EPE above 3 and " + }, + { + "bbox": [ + 46, + 198, + 287, + 327 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 198, + 287, + 327 + ], + "type": "text", + "content": " of the magnitude of the flow ground-truth (\\% Outlier) are presented for MVSEC. On DSEC-Flow, metrics are measured over pixels with valid ground-truth. While on MVSEC, metrics are measured over pixels with valid ground-truth and at least one triggered event." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "text", + "content": "Implementation details. The proposed TMA is implemented using PyTorch. We set the event segment " + }, + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "inline_equation", + "content": "g = 5" + }, + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "text", + "content": ". For experiments on DSEC-Flow, the channel of event representation " + }, + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "text", + "content": " for each segment is set to 3. For MVSEC, " + }, + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "text", + "content": " is set to 1 when " + }, + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "inline_equation", + "content": "dt = 1" + }, + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "text", + "content": " and 3 when " + }, + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "inline_equation", + "content": "dt = 4" + }, + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "text", + "content": ", respectively. The sum of channels keeps the same with the state-of-the-art E-RAFT [11]. We only stack 1 layer of motion pattern aggregation module. Other components are basically identical to RAFT's model, including feature extractor, motion feature encoder and flow updater. Note that we use a small feature extractor whose final channel is 128. We predict 6 levels of optical flow and the weight " + }, + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 329, + 288, + 520 + ], + "type": "text", + "content": " is set to 0.8 for supervision. In training, we exploit AdamW [23] optimizer and One-cycle learning rate scheduler [28]. We train models on DSEC-Flow for 200K steps and MVSEC for 100K steps, both with a learning rate of 0.0002 and a batch size of 6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 527, + 126, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 126, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 126, + 538 + ], + "type": "text", + "content": "4.1. DSEC-Flow" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 605 + ], + "type": "text", + "content": "Accuracy comparison. Evaluation results on DSEC-Flow benchmark1 are provided in Table 1. Compared with E-RAFT, our method improves the EPE from 0.79 to 0.74. Our method is also leading in 1PE and 3PE with a margin of 1.6 and 0.4, showing better robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "content": "Qualitative results are exhibited in Figure 5. In challenging regions such as textureless areas (the wall in second two columns, the shrubs in last two columns), existing correlation volume contains similar matching scores and fails to generate informative motion features, leading to low-quality predictions. On the contrary, our method aggregates" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 538, + 201 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 538, + 201 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 538, + 201 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 538, + 201 + ], + "type": "image", + "image_path": "913c95c892169db088b345359884f596481916bbb25a0f55db95311de59310a9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 205, + 545, + 264 + ], + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 264 + ], + "type": "text", + "content": "Figure 6: EPE vs. number of iterations at inference. The figure exhibits the prediction results on DSEC-Flow at different stages during inference. Our method achieves an accuracy " + }, + { + "bbox": [ + 304, + 205, + 545, + 264 + ], + "type": "inline_equation", + "content": "(\\mathrm{EPE} = 0.81)" + }, + { + "bbox": [ + 304, + 205, + 545, + 264 + ], + "type": "text", + "content": " comparable to E-RAFT with 2 iterations while E-RAFT needs 6 iterations." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 269, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 304 + ], + "type": "text", + "content": "temporally consistent motion patterns to enhance motion features, distinguishing different moving objects and generating high-quality predictions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 304, + 546, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 460 + ], + "type": "text", + "content": "Efficiency comparison. The last column of Table 1 exhibits the inference time. Our method enjoys faster inference speed compared with E-RAFT, which is mainly attributed to fewer iterative updates. The reasonable explanation is that our method utilizes temporal motion information to enrich motion features, thus being able to generate good predictions even at an initial stage. To prove this, we compare the accuracy with E-RAFT at early inference stages in Figure 6. Our method achieves better flow initialization at early stages, supporting our argument. The ablation experiment on different iterations (Table 4e) also demonstrates that our method requires fewer iterations (ours: 6, E-RAFT: 12) to achieve surpassing accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 467, + 368, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 368, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 368, + 479 + ], + "type": "text", + "content": "4.2.MVSEC" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "text", + "content": "Inter-domain evaluation. Table 2 reports the results on MVSEC training with outdoor_day2 sequence. The top of Table 2 reports the results corresponding to " + }, + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "inline_equation", + "content": "dt = 1" + }, + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "text", + "content": " grayscale frame and the bottom corresponds to " + }, + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "inline_equation", + "content": "dt = 4" + }, + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "text", + "content": " grayscale frames. Our method shows comparable accuracy on outdoor_day1 for " + }, + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "inline_equation", + "content": "dt = 1" + }, + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "text", + "content": " and highest accuracy on outdoor_day1 for " + }, + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "inline_equation", + "content": "dt = 4" + }, + { + "bbox": [ + 304, + 486, + 545, + 689 + ], + "type": "text", + "content": " among all compared methods. Unfortunately, like two other supervised learning methods, DCEI-Flow [34] and E-RAFT [11], our method fails to achieve superior results on indoor_flying sequences. One possible reason is that model based methods and unsupervised learning methods usually have better generalization ability, while supervised learning methods are limited if training data has a significantly large gap with test data [36, 34]. Since outdoor_day sequences and indoor_flying sequences have a great domain gap, it is challenging for supervised learning methods to generalize well on indoor_flying sequences." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Intra-domain evaluation. To verify the possible cause and better analyze the accuracy of our method under su" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": "1The ground-truth of DSEC-Flow dataset is unavailable, so we use last three checkpoints to evaluate the online benchmark. https://dsec.ifi.uzh.ch/uzh/dsec-flow-optical-flow-benchmark/." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9690" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 542, + 388 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 542, + 388 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 542, + 388 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 542, + 388 + ], + "type": "table", + "html": "
Inputindoor_flying1indoor_flying2indoor_flying3outdoor_day1
EPE% OutlierEPE% OutlierEPE% OutlierEPE% Outlier
dt = 1
MBBrebion et al. [5]E0.520.100.985.500.712.100.530.20
MultiCM (Burgers') [27]E0.420.100.600.590.500.280.300.10
SSLEV-FlowNet [43]E1.032.201.7215.101.5311.900.490.20
Spike-FlowNet [20]E0.84-1.28-1.11-0.49-
STE-FlowNet [7]E0.570.100.791.600.721.300.420.00
USLHagenaars et al. [13]E0.600.511.178.060.935.640.470.25
Zhu et al. [45]E0.580.001.024.000.873.000.320.00
ECN [38]E------0.300.02
SLE-RAFT [11]E1.105.721.9430.791.6625.200.240.00
DCEIFlow [34]E + I0.751.550.902.100.801.770.220.00
OursE1.063.631.8127.291.5823.260.250.07
dt = 4
MBMultiCM (Burgers') [27]E1.6912.952.4926.352.0619.031.259.21
SSLEV-FlowNet [43]E2.2524.704.0545.303.4539.701.237.30
Spike-FlowNet [20]E2.24-3.83-3.18-1.09-
STE-FlowNet [7]E1.7714.702.5226.102.2322.100.993.90
USLZhu et al. [45]E2.1824.203.8546.803.1847.801.309.70
Hagenaars et al. [13]E2.1621.513.9040.723.0029.601.6912.50
SLDCEIFlow [34]E + I2.0821.473.4842.052.5129.730.893.19
E-RAFT [11]E2.8140.255.0964.194.4657.110.721.12
OursE2.4329.914.3252.743.6042.020.701.08
", + "image_path": "f2ed09171887807ee79b93eaf93ab7360fc399a14ce4e0bfa1c36feaaad76e50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 455, + 541, + 567 + ], + "blocks": [ + { + "bbox": [ + 46, + 396, + 547, + 445 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 547, + 445 + ], + "type": "text", + "content": "Table 2: Evaluation results on MVSEC [44] training with outdoor_day2 sequence. Model-based (MB) methods need no training data; semi-supervised learning (SSL) methods use grayscale images for supervision; unsupervised learning (USL) methods only require events; and supervised learning (SL) methods need ground-truth. E means the method only uses events as input; and I means the method requires images as input." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 455, + 541, + 567 + ], + "lines": [ + { + "bbox": [ + 51, + 455, + 541, + 567 + ], + "spans": [ + { + "bbox": [ + 51, + 455, + 541, + 567 + ], + "type": "table", + "html": "
dt = 4Train Set +indoor_flying1indoor_flying2indoor_flying3outdoor_day1
EPE% OutlierEPE% OutlierEPE% OutlierEPE% Outlier
E-RAFT [11]indoor_flying1--2.3526.241.8418.070.721.38
Ours--2.1520.411.6814.300.640.98
E-RAFT [11]indoor_flying21.418.14--1.509.370.711.13
Ours1.325.94--1.508.880.640.96
E-RAFT [11]indoor_flying31.398.381.7717.52--0.721.46
Ours1.335.821.6712.80--0.671.10
", + "image_path": "75f3f968d5c41e898b98264cd0c03cb842560de7f92dd5f719cdcc3a4f726435.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 568, + 546, + 594 + ], + "lines": [ + { + "bbox": [ + 46, + 568, + 546, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 546, + 594 + ], + "type": "text", + "content": "Table 3: Evaluation results on MVSEC [44] training with outdoor_day2 and one indoor_flying sequence. + means the indoor_flying sequence to be included into training set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 600, + 289, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 600, + 289, + 708 + ], + "spans": [ + { + "bbox": [ + 46, + 600, + 289, + 708 + ], + "type": "text", + "content": "pervised learning setting, we incorporate one indoor_flying sequence into training set each time and make comparisons with the supervised learning state-of-the-art E-RAFT. Table 3 shows the evaluation results corresponding to " + }, + { + "bbox": [ + 46, + 600, + 289, + 708 + ], + "type": "inline_equation", + "content": "dt = 4" + }, + { + "bbox": [ + 46, + 600, + 289, + 708 + ], + "type": "text", + "content": " grayscale frames (For " + }, + { + "bbox": [ + 46, + 600, + 289, + 708 + ], + "type": "inline_equation", + "content": "dt = 1" + }, + { + "bbox": [ + 46, + 600, + 289, + 708 + ], + "type": "text", + "content": ", refer to appendix). Compared with the results in Table 2, E-RAFT and our method both achieve great improvement on indoor_flying sequences (EPE of indoor_flying2, ours: " + }, + { + "bbox": [ + 46, + 600, + 289, + 708 + ], + "type": "inline_equation", + "content": "4.32 \\rightarrow 1.67" + }, + { + "bbox": [ + 46, + 600, + 289, + 708 + ], + "type": "text", + "content": ", E-RAFT: 5.09 → 1.77), confirming the presence of domain gap between" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 600, + 547, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 547, + 708 + ], + "type": "text", + "content": "outdoor_day and indoor_flying data. Compared with the accuracy of E-RAFT with the same training set, on the one hand, our method shows better generalization ability, which generates lower EPE and " + }, + { + "bbox": [ + 304, + 600, + 547, + 708 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 304, + 600, + 547, + 708 + ], + "type": "text", + "content": " Outlier without seeing indoor_flying sequences. On the other hand, our method achieves higher accuracy on four sequences by incorporating indoor_flying sequences into training set. Overall, our method achieves a new state-of-the-art result, demonstrating the effectiveness of our proposed method." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "9691" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 289, + 142 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 289, + 142 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 289, + 142 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 289, + 142 + ], + "type": "table", + "html": "
splitsEPE1PE3PETime (ms)
10.7911.912.7516
30.7610.982.4518
5/70.7410.862.3030
0.7411.132.2650
", + "image_path": "97129115d91c360c5c2197e004c27cc0f178f606faa3d1a7b3fb4bbc14118cd4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 54, + 156, + 289, + 215 + ], + "blocks": [ + { + "bbox": [ + 54, + 144, + 289, + 155 + ], + "lines": [ + { + "bbox": [ + 54, + 144, + 289, + 155 + ], + "spans": [ + { + "bbox": [ + 54, + 144, + 289, + 155 + ], + "type": "text", + "content": "(a) Event splitting. Intermediate motion information contributes" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 156, + 289, + 215 + ], + "lines": [ + { + "bbox": [ + 54, + 156, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 54, + 156, + 289, + 215 + ], + "type": "table", + "html": "
setupEPE1PE3PETime (ms)
basic0.7410.862.3030
+v_Proj0.7510.902.3530
circular0.7611.252.4430
", + "image_path": "5a487fbbb1f9ed44b18924752496a67e6595a83a0ec621e5fb3f7ca20210152a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 53, + 239, + 288, + 312 + ], + "blocks": [ + { + "bbox": [ + 97, + 217, + 247, + 228 + ], + "lines": [ + { + "bbox": [ + 97, + 217, + 247, + 228 + ], + "spans": [ + { + "bbox": [ + 97, + 217, + 247, + 228 + ], + "type": "text", + "content": "(c) Pattern aggregation module design." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 239, + 288, + 312 + ], + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 312 + ], + "type": "table", + "html": "
iterationsEPE1PE3PETime (ms)
20.8311.042.3710
40.7611.452.4316
6/80.7410.862.3030
0.7410.812.3254
", + "image_path": "91450d64bd7e52d1fd6f36f751ba9bea71760a969f26434b948f0687f87533b2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 313, + 288, + 324 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 288, + 324 + ], + "type": "text", + "content": "(e) Iterations. TMA achieves good results with a few iterations." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 301, + 70, + 538, + 131 + ], + "blocks": [ + { + "bbox": [ + 301, + 70, + 538, + 131 + ], + "lines": [ + { + "bbox": [ + 301, + 70, + 538, + 131 + ], + "spans": [ + { + "bbox": [ + 301, + 70, + 538, + 131 + ], + "type": "table", + "html": "
styleEPE1PE3PE
w/o0.7812.002.66
linear0.7410.862.30
same0.7611.452.48
", + "image_path": "04d50b167edbeda2715247b1aa4f2d38ca1d20c0c5df48ea3ae54815729c5ff3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 301, + 156, + 537, + 215 + ], + "blocks": [ + { + "bbox": [ + 296, + 133, + 544, + 154 + ], + "lines": [ + { + "bbox": [ + 296, + 133, + 544, + 154 + ], + "spans": [ + { + "bbox": [ + 296, + 133, + 544, + 154 + ], + "type": "text", + "content": "(b) Lookup style. Sampling correlation volumes with a set of linear coordinates improves the accuracy." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 301, + 156, + 537, + 215 + ], + "lines": [ + { + "bbox": [ + 301, + 156, + 537, + 215 + ], + "spans": [ + { + "bbox": [ + 301, + 156, + 537, + 215 + ], + "type": "table", + "html": "
layersEPE1PE3PETime (ms)
00.7511.362.3320
1/20.7410.862.3030
0.7510.902.3238
", + "image_path": "06c5d04f2929d106dea5f3a21455b1508f20f36bc4f943fb697fbf6f9afe2cca.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 301, + 240, + 535, + 312 + ], + "blocks": [ + { + "bbox": [ + 296, + 217, + 544, + 239 + ], + "lines": [ + { + "bbox": [ + 296, + 217, + 544, + 239 + ], + "spans": [ + { + "bbox": [ + 296, + 217, + 544, + 239 + ], + "type": "text", + "content": "(d) Layers of pattern aggregation module. Stacking 1 pattern aggregation module showcases high accuracy." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 301, + 240, + 535, + 312 + ], + "lines": [ + { + "bbox": [ + 301, + 240, + 535, + 312 + ], + "spans": [ + { + "bbox": [ + 301, + 240, + 535, + 312 + ], + "type": "table", + "html": "
radiusEPE1PE3PE
10.7511.142.37
20.7611.162.39
3/40.7410.862.30
0.7510.932.30
", + "image_path": "43d2044be5526ce3f3559a5f67c56a119d9dbacb6177299752ff814134cc6270.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 313, + 526, + 325 + ], + "lines": [ + { + "bbox": [ + 312, + 313, + 526, + 325 + ], + "spans": [ + { + "bbox": [ + 312, + 313, + 526, + 325 + ], + "type": "text", + "content": "(f) Searching radius. A radius of 3 leads to high accuracy." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 75, + 327, + 517, + 338 + ], + "lines": [ + { + "bbox": [ + 75, + 327, + 517, + 338 + ], + "spans": [ + { + "bbox": [ + 75, + 327, + 517, + 338 + ], + "type": "text", + "content": "Table 4: TMA ablations. Models are trained on DSEC-Flow. Settings used in our final model are underlined." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 346, + 115, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 115, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 115, + 357 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 365, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 400 + ], + "type": "text", + "content": "We conduct a series of ablation experiments to validate the proposed improvements. All ablation models are trained on DSEC-Flow and evaluated on the public benchmark." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 484 + ], + "type": "text", + "content": "Event splitting. We compare different numbers of event segments in Table 4a. We choose different splits in 1, 3, 5 and 7. With the split set to 1, our model degrades into a classic correlation volume based framework and exhibits similar accuracy in comparison with E-RAFT. More splits lead to higher accuracy but lower inference efficiency, so we use 5 splits to achieve accuracy-efficiency trade-off." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 485, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 568 + ], + "type": "text", + "content": "Lookup style. Based on the assumption that the optical flow changes linearly over a short period, we propose the linear lookup strategy. To validate it, we try to cancel the lookup operation or just use same sampling coordinates on intermediate correlation volumes. As Table 4b shows, sampling the correlation maps with a set of linear coordinates leads to high accuracy." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "type": "text", + "content": "Pattern aggregation module design. We ablate different components of pattern aggregation module in Table 4c. By default, we set the projection matrix of value " + }, + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "type": "inline_equation", + "content": "W^V" + }, + { + "bbox": [ + 46, + 569, + 287, + 676 + ], + "type": "text", + "content": " as an identity matrix because the motion features are encoded by a motion feature encoder beforehand. In addition, we replace the pairing mode of motion features with circular pairing. Circular pairing gives equal treatment to all motion features without highlighting the importance of the last motion feature, which fails to improve accuracy." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Layers of pattern aggregation module. We compare the accuracy with different layers of pattern aggregation modules in Table 4d. We note that stacking 1 layer of pat" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 346, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 381 + ], + "type": "text", + "content": "tern aggregation module showcases high accuracy. More layers may lead to difficulties in optimization, which in turn causes a decrease in accuracy." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 383, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 418 + ], + "type": "text", + "content": "Iterations. Iterative learning plays a role in numerous works [18, 30]. Results in Table 4e show that 6 iterative updates reach the sweet spot of accuracy and efficiency." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 418, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 478 + ], + "type": "text", + "content": "Searching radius. We compare different radii for looking up correlation volumes in Table 4f. Searching radius of 3 leads to high accuracy. Different from RAFT [30] that achieves better predictions with an increased radius, our model does not benefit a lot from a greater radius." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we delve into event-based optical flow estimation problem. Different from existing methods that represent the event stream as static frames to adopt the classic correlation volume based pipeline, we contend that the temporal continuity is a vital element of event-based optical flow. We propose a novel Temporal Motion Aggregation (TMA) approach to unlock its potential within a learning-based framework. By incorporating temporally fine-grained motion information, TMA generates high-quality flow estimates at early stages, which not only enables TMA to obtain accurate final predictions, but also greatly reduces the demand for a number of refinements. Extensive experiments on DESC-Flow and MVSEC datasets verify the effectiveness and superiority of TMA. Meanwhile, we notice the limitations of supervised learning in generalization ability and expect further improvement to extend our idea into unsupervised learning frameworks." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9692" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 161 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 161 + ], + "type": "text", + "content": "Acknowledgments: This work was supported in part by Shanghai Municipal Science and Technology Major Project under Grant 2018SHZDZX01, ZJ Lab, the Shanghai Center for Brain Science and Brain-Inspired Technology, in part by Shanghai Rising Star Program under Grant 21QC1400900, in part by Tongji-Qomolo Autonomous Driving Commercial Vehicle Joint Lab Project, and in part by Xiaomi Young Talents Program." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 182, + 106, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 182, + 106, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 182, + 106, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 202, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "type": "text", + "content": "[1] Himanshu Akolkar, Sio-Hoi Ieng, and Ryad Benosman. Real-time high speed motion prediction using fast aperture-robust event-driven visual flow. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):361-372, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "type": "text", + "content": "[2] Mohammed Almatrafi, Raymond Baldwin, Kiyoharu Aizawa, and Keigo Hirakawa. Distance surface for event-based optical flow. IEEE transactions on pattern analysis and machine intelligence, 42(7):1547-1556, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "type": "text", + "content": "[3] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE transactions on neural networks and learning systems, 25(2):407-417, 2013. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 346, + 287, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 389 + ], + "type": "text", + "content": "[4] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, 2012. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "text", + "content": "[5] Vincent Brebion, Julien Moreau, and Franck Davoine. Realtime optical flow for vehicular perception with low-and high-resolution event cameras. IEEE Transactions on Intelligent Transportation Systems, 2021. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 436, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 468 + ], + "type": "text", + "content": "[6] Tobias Brosch, Stephan Tschechne, and Heiko Neumann. On event-based optical flow detection. Frontiers in neuroscience, 9:137, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 469, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 524 + ], + "type": "text", + "content": "[7] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-temporal recurrent networks for event-based optical flow estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 525-533, 2022. 3, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 525, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 287, + 579 + ], + "type": "text", + "content": "[8] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 580, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 580, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 53, + 580, + 287, + 645 + ], + "type": "text", + "content": "[9] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3867-3876, 2018. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[10] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3):4947-4954, 2021. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[11] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-raft: Dense optical flow from event cam" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "eras. In 2021 International Conference on 3D Vision (3DV), pages 197-206. IEEE, 2021. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "text", + "content": "[12] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from events and frames. arXiv preprint arXiv:2203.13674, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 172 + ], + "type": "text", + "content": "[13] Jesse Hagenaars, Federico Paredes-Vallés, and Guido De Croon. Self-supervised learning of event-based optical flow with spiking neural networks. Advances in Neural Information Processing Systems, 34:7167-7179, 2021. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 173, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 217 + ], + "type": "text", + "content": "[14] Markus Hofinger, Samuel Rota Bulò, Lorenzo Porzi, Arno Knapitsch, Thomas Pock, and Peter Kontschieder. Improving optical flow on a pyramid level. In European Conference on Computer Vision, pages 770-786. Springer, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 260 + ], + "type": "text", + "content": "[15] Tak-Wai Hui and Chen Change Loy. Liteflownet3: Resolving correspondence ambiguity for more accurate optical flow estimation. In European Conference on Computer Vision, pages 169-184. Springer, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "text", + "content": "[16] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. Liteflownet: A lightweight convolutional neural network for optical flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8981-8989, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 360 + ], + "type": "text", + "content": "[17] Tak-Wai Hui, Xiaou Tang, and Chen Change Loy. A lightweight optical flow cnn—revisiting data fidelity and regularization. IEEE transactions on pattern analysis and machine intelligence, 43(8):2555-2569, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 404 + ], + "type": "text", + "content": "[18] Junhwa Hur and Stefan Roth. Iterative residual refinement for joint optical flow and occlusion estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2019. 1, 2, 3, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "text", + "content": "[19] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9772-9781, 2021. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 459, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 459, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 459, + 545, + 514 + ], + "type": "text", + "content": "[20] Chankyu Lee, Adarsh Kumar Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, and Kaushik Roy. Spikeflownet: event-based optical flow estimation with energy-efficient hybrid neural networks. In European Conference on Computer Vision, pages 366-382. Springer, 2020. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "text", + "content": "[21] Min Liu and Tobi Delbruck. Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. BMVC, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "text", + "content": "[22] Min Liu and Tobi Delbruck. Edflow: Event driven optical flow camera with keypoint detection and adaptive block matching. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5776-5789, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 592, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 613 + ], + "type": "text", + "content": "[23] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 669 + ], + "type": "text", + "content": "[24] Federico Paredes-Vallés and Guido CHE de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446-3455, 2021. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[25] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4161-4170, 2017. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9693" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[26] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 170 + ], + "type": "text", + "content": "[27] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In European Conference on Computer Vision, pages 628-645. Springer, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 237 + ], + "type": "text", + "content": "[28] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications, volume 11006, page 1100612. International Society for Optics and Photonics, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 238, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 287, + 293 + ], + "type": "text", + "content": "[29] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8934-8943, 2018. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 293, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 327 + ], + "type": "text", + "content": "[30] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European conference on computer vision, pages 402-419. Springer, 2020. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 392 + ], + "type": "text", + "content": "[31] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 458 + ], + "type": "text", + "content": "[32] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 503 + ], + "type": "text", + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 503, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 546 + ], + "type": "text", + "content": "[34] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing, 31:7237-7251, 2022. 1, 3, 4, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 547, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 602 + ], + "type": "text", + "content": "[35] Haofei Xu, Jiaolong Yang, Jianfei Cai, Juyong Zhang, and Xin Tong. High-resolution optical flow from 1d attention and correlation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10498-10507, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 657 + ], + "type": "text", + "content": "[36] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8121-8130, 2022. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "text", + "content": "[37] Gengshan Yang and Deva Ramanan. Volumetric correspondence networks for optical flow. Advances in neural information processing systems, 32:794-805, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[38] Chengxi Ye, Anton Mitrokhin, Cornelia Fermüller, James A Yorke, and Yiannis Aloimonos. Unsupervised learning of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 486 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "dense optical flow, depth and egomotion with event-based sensors. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5831-5838. IEEE, 2020. 1, 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[39] Feihu Zhang, Victor Prisacariu, Ruigang Yang, and Philip HS Torr. Ga-net: Guided aggregation net for end-to-end stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 185-194, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 217 + ], + "type": "text", + "content": "[40] Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In European Conference on Computer Vision, pages 420-439. Springer, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 274 + ], + "type": "text", + "content": "[41] Feihu Zhang, Oliver J Woodford, Victor Adrian Prisacariu, and Philip HS Torr. Separable flow: Learning motion cost volumes for optical flow estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10807-10817, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 274, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 319 + ], + "type": "text", + "content": "[42] Xu Zheng, Yexin Liu, Yunfan Lu, Tongyan Hua, Tianbo Pan, Weiming Zhang, Dacheng Tao, and Lin Wang. Deep learning for event-based vision: A comprehensive survey and benchmarks. arXiv preprint arXiv:2302.08890, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "text", + "content": "[43] Alex Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Ev-flownet: Self-supervised optical flow estimation for event-based cameras. In Proceedings of Robotics: Science and Systems, Pittsburgh, Pennsylvania, June 2018, 2, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 374, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 374, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 374, + 545, + 430 + ], + "type": "text", + "content": "[44] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018. 2, 3, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 430, + 547, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 547, + 486 + ], + "type": "text", + "content": "[45] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9694" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/d1f3cbdb-e07a-4952-8052-a9aa7bed81da_content_list.json b/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/d1f3cbdb-e07a-4952-8052-a9aa7bed81da_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..925a99c2f12038f7b27e9e71f19eccda980b87f6 --- /dev/null +++ b/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/d1f3cbdb-e07a-4952-8052-a9aa7bed81da_content_list.json @@ -0,0 +1,1664 @@ +[ + { + "type": "text", + "text": "TMR: Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis", + "text_level": 1, + "bbox": [ + 107, + 130, + 861, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mathis Petrovich $^{1,2}$ Michael J. Black $^{2}$ Gül Varol $^{1}$ $^{1}$ LIGM, École des Ponts, Univ Gustave Eiffel, CNRS, France \n $^{2}$ Max Planck Institute for Intelligent Systems, Tübingen, Germany \nhttps://mathis.petrovich.fr/tmr", + "bbox": [ + 230, + 179, + 736, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 235, + 287, + 310, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present TMR, a simple yet effective approach for text to 3D human motion retrieval. While previous work has only treated retrieval as a proxy evaluation metric, we tackle it as a standalone task. Our method extends the state-of-the-art text-to-motion synthesis model TEMOS, and incorporates a contrastive loss to better structure the cross-modal latent space. We show that maintaining the motion generation loss, along with the contrastive training, is crucial to obtain good performance. We introduce a benchmark for evaluation and provide an in-depth analysis by reporting results on several protocols. Our extensive experiments on the KIT-ML and HumanML3D datasets show that TMR outperforms the prior work by a significant margin, for example reducing the median rank from 54 to 19. Finally, we showcase the potential of our approach on moment retrieval. Our code and models are publicly available at https://mathis.petrovich.fr/tmr.", + "bbox": [ + 75, + 321, + 473, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 599, + 202, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The language of movement cannot be translated into words. Barbara Mettler", + "bbox": [ + 107, + 627, + 454, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We ask the question whether a cross-modal space exists between 3D human motions and language. Our goal is to retrieve the most relevant 3D human motion from a gallery, given a natural language query that describes the desired motion (as illustrated in Figure 1). While text-to-image retrieval is a well-established problem within the broader vision & language field [39], there has been less focus on the related task of text-to-motion retrieval. Searching an existing motion capture dataset based on text input can often serve as a viable alternative to text-to-human-motion synthesis in many applications, while also providing the added benefit that the retrieved motion is guaranteed to be realistic. Additionally, once a cross-modal embedding is built to map text and motions into a joint representation space, both of the symmetrical text-to-motion and motion-to-text tasks", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3bd9579c967fe35c00393d88dd4a95ab700d969e4f4b8da77794fd0740ac96d8.jpg", + "image_caption": [ + "A person is stretching legs", + "Figure 1. Text-to-motion retrieval: We illustrate the task of text-based motion retrieval where the goal is to rank a gallery of motions according to their similarity to the given query in the form of a natural language description." + ], + "image_footnote": [], + "bbox": [ + 508, + 321, + 885, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "can be performed. Such a retrieval-based solution has a range of applications, including automatically indexing large motion capture collections, and helping to initialize the cumbersome text labeling process, by assigning the nearest text to each motion.", + "bbox": [ + 496, + 590, + 893, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Let us first differentiate text-to-motion retrieval from text-to-motion synthesis. Motion synthesis [5, 9, 36, 49] involves generating new data samples that go beyond the existing training set, while motion retrieval searches through existing motion capture collections. For certain applications, reusing motions from a collection may be sufficient, provided the collection is large enough to contain what the user is searching for. Unlike generative models for motion synthesis, which struggle to produce physically plausible, realistic sequences [5, 35, 36], a retrieval model has the advantage that it always returns a realistic motion. With this motivation, we pose the problem as a nearest neighbor search through a cross-modal text-motion space.", + "bbox": [ + 496, + 655, + 893, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early work performs search through motion databases to build motion graphs [4, 23] by finding paths between existing motions and synthesizing new motions by stitching motions together with generated transitions. If the motion database is", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9488", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "labeled with actions, the user can specify a series of actions to combine [4]. In contrast, our search database is not labeled with text. Motion matching [8], on the other hand, seeks to find the animation that best fits the current motion by searching a database of animations, doing motion-to-motion retrieval [45]. Our framework fundamentally differs from these lines of work in that our task is multi-modal, i.e., user query is text, which is compared against motions. The most similar work to ours is the very recent model from Guo et al. [15], which trains for a joint embedding space between the two modalities. This model is only used to provide a performance measure for motion synthesis tasks, by querying a generated motion within a gallery of 32 descriptions (i.e., motion-to-text retrieval), and counting how many times the correct text is retrieved1. While this can be considered as the first text-motion retrieval model in the literature, its main limitation is the low performance, in particular when the gallery contains fine-grained descriptions. We substantially improve over [15], by incorporating a joint synthesis & retrieval framework, as well as a more powerful contrastive training [34].", + "bbox": [ + 75, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We get inspiration from image-text models such as BLIP [26] and CoCa [54], which formulate a multi-task objective. Besides the standard dual-encoder matching (such as CLIP [39] with two unimodal encoders for image and text), [26, 54] also employ a text synthesis branch, performing image captioning. Such a generative capability potentially helps the model go beyond 'bag-of-words' understanding of vision-language concepts, observed for the naive contrastive models [11, 55]. In our case, we depart from TEMOS [36] which already has a synthesis branch to generate motions from text. We incorporate a cross-modal contrastive loss (i.e., InfoNCE [34]) in this framework to jointly train text-to-motion synthesis and text-to-motion retrieval tasks. We empirically demonstrate significant improvements with this approach when ablating the importance of each task.", + "bbox": [ + 75, + 378, + 472, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-motion data differs from its text-image counterparts particularly due to the nature of motion descriptions. In fact, for an off-the-shelf large language model, sentences describing different motions tend to be similar, since they fall within the same topic of human motions. For example, the text-text cosine similarities [46] after encoding motion descriptions from the KIT training set [37] are on average 0.71 on a scale between [0, 1], while this value is 0.56 (almost orthogonal) on a random subset of LAION [44] image descriptions with the same size. This poses several challenges. Typical motion datasets [15, 37, 38] contain similar motions with different accompanying texts, e.g., 'person walks', 'human walks', as well as similar texts with different motions, e.g., 'walk backwards', 'walk forwards'. With naive contrastive training [34], one would make all samples within a batch as negatives, except the corresponding label for a given anchor. In this work, we take into account the fact that there are potentially significant similarities between pairs within a batch. Specifically, we discard pairs that have a text-text", + "bbox": [ + 75, + 590, + 472, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "similarity in their labels that is above a certain threshold. Such careful negative sampling leads to performance improvements.", + "bbox": [ + 500, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we illustrate an additional use case for our retrieval model – zero-shot temporal localization – and highlight this task as a potential future avenue for research. Similar to temporal localization in videos with natural language queries [12, 13, 19, 25, 41], also referred to as “moment retrieval”, we showcase the grounding capability of our model by directly applying it on long motion sequences to retrieve corresponding moments. We illustrate results on the BABEL dataset [38], which typically contains a series of text annotations for each motion sequence. Note that the task is zero-shot, because the model has not been trained for localization, and at the same time has not seen BABEL labels, which come from a different domain (e.g., typically action-like descriptions instead of full sentences).", + "bbox": [ + 496, + 122, + 893, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are the following: (i) We address the little-studied problem of text-to-motion retrieval, and introduce a series of evaluation benchmarks with varying difficulty. (ii) We propose a joint synthesis and retrieval framework, as well as negative filtering, and obtain state-of-the-art performance on text-motion retrieval. (iii) We provide extensive experiments to analyze the effects of each component in controlled settings. Our code and models are publicly available2.", + "bbox": [ + 496, + 335, + 893, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 473, + 630, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present an overview of the most closely related work on text-to-motion synthesis and retrieval, as well as a brief discussion on cross-modal retrieval works.", + "bbox": [ + 496, + 500, + 890, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text and human motion. The research on human motion modeling has recently witnessed an increasing interest in bridging the gap between semantics and 3D human body motions, in particular for text-conditioned motion synthesis. Unlike unconstrained 3D human motion synthesis [53, 58, 59], action-conditioned [16, 35], or text-conditioned [5, 9, 14, 22, 24, 36, 49, 56] models add semantic controls to the generation process. The goal of these works is to generate either deterministically [2, 14, 27] or probabilistically [36], motion sequences that are faithful to the textual description inputs. Note that this is different than gesture synthesis from speech [17], in that the text describes the motion content. However, despite remarkable progress in text-to-image synthesis [40, 42], text-to-motion synthesis remains at a very nascent stage. The realism of the synthesized motions is limited, e.g., foot sliding artifacts [36]. We turn to text-to-motion retrieval as another alternative, and perhaps complementary approach to obtain motions for a given textual description. Our focus is therefore different from the work on synthesis. However, we do make use of a motion synthesis branch to aid the retrieval task.", + "bbox": [ + 496, + 570, + 893, + 872 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1While the paper [15] describes a motion-to-text retrieval metric, we notice that the provided code performs text-to-motion retrieval.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2https://mathis.petrovich.fr/tmr", + "bbox": [ + 517, + 886, + 759, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9489", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motion retrieval is relatively less explored. As briefly mentioned in Section 1, motion-to-motion retrieval (e.g., motion matching [8, 20, 45]) methods exist. However, the text-to-motion retrieval task is more challenging due to being cross-modal, i.e., nearest neighbor search across text and motion modalities. Within this category, the very recent work of Guo et al. [15] trains a retrieval model purely for evaluation purposes, and applies a margin-based contrastive loss [18], using Euclidean distance between all pairs within a batch.", + "bbox": [ + 75, + 90, + 470, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Two works are particularly relevant to ours. Firstly, we build on the TEMOS [36] text-to-motion synthesis model, which also has a cross-modal embedding space. However, text and motion embeddings are encouraged to be similar only across positive pairs. We therefore add a contrastive training strategy to incorporate negatives, consequently improving its retrieval capability from a large gallery of fine-grained motions. Secondly, we compare to the aforementioned method of [15], whose motion-to-text retrieval model is adopted for measuring text-to-motion generation performance automatically by other works [10, 49, 57].", + "bbox": [ + 75, + 227, + 473, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-modal retrieval. Among widely adopted vision & language retrieval models, some successful examples include CLIP [39], BLIP [26], and CoCa [54] for images, and MILNCE [33], Frozen [7], and CLIP4Clip [30] for videos. They all use variants of cross-modal contrastive learning techniques, such as InfoNCE [34], which we also employ in this work. As discussed in Section 1, we draw inspiration from BLIP [26] and CoCa [54], which add synthesis branches to standard retrieval frameworks. Our approach is similar in spirit to these works in that we perform a cross-modal vision & language retrieval task, but differ in focusing on 3D human motion retrieval, which, to the best of our knowledge, has not been benchmarked.", + "bbox": [ + 75, + 395, + 470, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Text-to-motion retrieval", + "text_level": 1, + "bbox": [ + 76, + 588, + 290, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the task and the terminology associated with text-to-motion retrieval (Section 3.1). Next, we present our model, named TMR, and its training protocol (Section 3.2). We then explain our simple approach for identifying and filtering incorrect negatives (Section 3.3), followed by a discussion of the implementation details (Section 3.4).", + "bbox": [ + 75, + 613, + 470, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Definitions", + "text_level": 1, + "bbox": [ + 76, + 710, + 196, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a natural language query $T$ , such as 'A person walks and then makes a right turn', the goal is to rank the motions from a database (i.e., the gallery) according to their semantic correspondence with the text query, and to retrieve the motion that matches best to the textual description. In other words, the task involves sorting the database so that the top ranked motions are the most relevant matches, i.e., creating a search engine to index motions. Additionally, we define the symmetric (and complementary) task, namely motion-to-text retrieval, where the aim is to retrieve the most suitable caption that matches a given motion from a database of texts.", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D human motion refers to a sequence of human poses. The task does not impose any limitations on the type of representation used, such as joint positions, rotations, or parametric models such as SMPL [28]. As detailed in Section 4.1, we choose to use the representation employed by Guo et al. [15] to facilitate comparisons with previous work. While our experiments are based on the SMPL skeleton, our method is applicable to any skeleton topology. One could alternatively use retargetting [1, 51, 52] to adapt to other body representations.", + "bbox": [ + 496, + 90, + 893, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text description refers to a sequence of words describing the action performed by a human in natural language. We do not restrict the format of the motion description. The text can be simply an action name (e.g., 'walk') or a full sentence (e.g., 'a human is walking'). The sentences can be fine-grained (e.g., 'a human is walking in a circle slowly'), and may contain one or several actions, simultaneously (e.g., 'walking while waving') or sequentially (e.g., 'walking then sitting').", + "bbox": [ + 496, + 227, + 893, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Joint training of retrieval and synthesis", + "text_level": 1, + "bbox": [ + 498, + 359, + 836, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce TMR, which extends the Transformer-based text-to-motion synthesis model TEMOS [36] by incorporating additional losses to make it suitable for the retrieval task. The architecture consists of two independent encoders for inputting motion and text, as well as a decoder that outputs motion (see Figure 2 for an overview). In the following, we review TEMOS [36] components and our added contrastive training.", + "bbox": [ + 496, + 383, + 890, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dual encoders. One approach to solving cross-modal retrieval tasks involves defining a similarity function between the two modalities. In our case, the two modalities are text and motion. The similarity function can be applied to compare a given query with each element in the database, and the maximum value would indicate the best match. In this paper, we follow the approach taken by previous work on metric learning, such as CLIP [39], by defining one encoder for each modality and then computing the cosine similarity between their respective embeddings. Such dual embedding has the advantage of fast inference time since the gallery embeddings can be computed and stored beforehand [32].", + "bbox": [ + 496, + 491, + 892, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model is built upon the components of TEMOS [36], which already provides a motion encoder and a text encoder, mapping them to a joint space (building on the idea from Language2Pose [2]); this serves as a strong baseline for our work. Both motion and text encoders are Transformer encoders [50] with additional learnable distribution parameters, as in the VAE-based ACTOR [35]. They are probabilistic in nature, outputting parameters of a Gaussian distribution $(\\mu$ and $\\Sigma)$ from which a latent vector $z\\in \\mathbb{R}^d$ can be sampled. While the text encoder takes text features from a pre-trained and frozen DistilBERT [43] network as input, the motion sequence is fed directly in the motion encoder. Note that when performing retrieval, we directly use the output embedding that corresponds to the mean token $(\\mu^{M}$ for motion, $\\mu^T$ for text).", + "bbox": [ + 496, + 672, + 893, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion decoder. TEMOS is trained for the task of motion", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9490", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b6aaf7f1f0ffd34b6555d615d240dfd6932730dbd2666a2d89ac1a9c9372187f.jpg", + "image_caption": [ + "Figure 2. Joint motion retrieval and synthesis: A simplified view of our TMR framework is presented, where we focus on the similarity matrix defined between text-motion pairs within a batch. Here, we show a batch of 3 samples for illustration purposes. The goal of the contrastive objective is to maximize the diagonal denoting positive pairs (green), and to minimize the off-diagonal negative pair items that have text similarity below a threshold (red). In this example, remaining similarities $S_{23}$ and $S_{32}$ are discarded from the loss computation because there is high text similarity between $T_{2}$ and $T_{3}$ . The rest of the model remains similar to TEMOS [36], which decodes a motion from both text $z^{T_i}$ and motion $z^{M_j}$ latent vectors. See text for further details." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 869, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "synthesis and comes equipped with a motion decoder branch. This decoder is identical to the one used in ACTOR [35], which supports variable-duration generation. More specifically, it takes a latent vector $z \\in \\mathbb{R}^d$ and a sinusoidal positional encoding as input, and generates a motion non-autoregressively through a single forward pass. We show in Section 4.2 that keeping this branch helps improve the results.", + "bbox": [ + 75, + 391, + 473, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TEMOS losses. We keep the same base set of losses of [36], defined as the weighted sum $\\mathcal{L}_{\\mathrm{TEMOS}} = \\mathcal{L}_{\\mathrm{R}} + \\lambda_{\\mathrm{KL}}\\mathcal{L}_{\\mathrm{KL}} + \\lambda_{\\mathrm{E}}\\mathcal{L}_{\\mathrm{E}}$ . In summary, a reconstruction loss term $\\mathcal{L}_{\\mathrm{R}}$ measures the motion reconstruction given text or motion input (via a smooth L1 loss). A Kullback-Leibler (KL) divergence loss term $\\mathcal{L}_{\\mathrm{KL}}$ is composed of four losses: two of them to regularize each encoded distributions $-\\mathcal{N}(\\mu^M,\\Sigma^M)$ for motion and $\\mathcal{N}(\\mu^T,\\Sigma^T)$ for text - to come from a normal distribution $\\mathcal{N}(0,I)$ . The other two enforce distribution similarity between the two modalities. A cross-modal embedding similarity loss $\\mathcal{L}_{\\mathrm{E}}$ enforces both text $z^{T}$ and motion $z^{M}$ latent codes to be similar to each other (with a smooth L1 loss). We set $\\lambda_{\\mathrm{KL}}$ and $\\lambda_{\\mathrm{E}}$ to $10^{-5}$ in our experiments as in [36].", + "bbox": [ + 75, + 508, + 473, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive training. While TEMOS has a cross-modal embedding space, its major drawback to be usable as an effective retrieval model is that it is never trained with negatives, but only positive motion-text pairs. To overcome this limitation, we incorporate a contrastive training with the usage of negative samples to better structure the latent space. Given a batch of N (positive) pairs of latent codes $(z_1^T,z_1^M),\\dots,(z_N^T,z_N^M)$ , we define any pair $(z_i^T,z_j^M)$ with $i\\neq j$ as negative. The similarity matrix $S$ computes the pairwise cosine similarities for all pairs in the batch $S_{ij} = \\cos (z_i^T,z_j^M)$ . In contrast to Guo et al. [15], who consider one random negative per batch (and use a margin loss), we adopt the more recent formulation of InfoNCE [34], which was proven effective in other work [7, 26, 39]. This loss term", + "bbox": [ + 75, + 703, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "can be defined as follows:", + "bbox": [ + 500, + 391, + 666, + 405 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {N C E}} = - \\frac {1}{2 N} \\sum_ {i} \\left(\\log \\frac {\\exp S _ {i i} / \\tau}{\\sum_ {j} \\exp S _ {i j} / \\tau} + \\log \\frac {\\exp S _ {i i} / \\tau}{\\sum_ {j} \\exp S _ {j i} / \\tau}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 412, + 892, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature hyperparameter.", + "bbox": [ + 498, + 459, + 772, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training loss. The total loss we use to train TMR is the weighted sum $\\mathcal{L}_{\\mathrm{TEMOS}} + \\lambda_{\\mathrm{NCE}}\\mathcal{L}_{\\mathrm{NCE}}$ where $\\lambda_{\\mathrm{NCE}}$ controls the importance of the contrastive loss.", + "bbox": [ + 498, + 474, + 892, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Filtering negatives", + "text_level": 1, + "bbox": [ + 500, + 529, + 679, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned in Section 1, the descriptions accompanying motion capture collections can be repetitive or similar across the training motions. We wish to prevent defining negatives between text-motion pairs that contain similar descriptions.", + "bbox": [ + 496, + 551, + 892, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consider for example the two text descriptions, \"A human making a counterclockwise turn\" and \"A person walks quarter a circle to the left\". In the KIT-ML benchmark [37], these two descriptions appear as two different annotations for the same motion. Due to the flexibility and ambiguity of natural language, different words may describe the same concepts (e.g., 'counterclockwise', 'circle to the left').", + "bbox": [ + 496, + 613, + 893, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, the random selection of batches can adversely affect the results because the model may have to push away two latent vectors that correspond to similar meanings. This can force the network to focus on unimportant details (e.g., 'someone' vs 'human'), ultimately resulting in decreased performance due to unstable behavior and reduced robustness to text variations.", + "bbox": [ + 496, + 719, + 893, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To alleviate this issue, we leverage an external large language model to provide sentence similarity scores. In particular, we use MPNet [47] to encode sentences and compute similarities between two text descriptions. We then determine whether to filter a pair of text descriptions $(t_{1}, t_{2})$ if their similarity", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9491", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is higher than a certain threshold, referring them as 'wrong negatives'. During training, we filter wrong negative pairs from the loss computation. Note that we refrain from defining them as positives, as the language model may also incorrectly say two descriptions are similar when they are not.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 174, + 292, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the AdamW optimizer [29] with a learning rate of $10^{-4}$ and a batch size of 32. Since the batch size can be an important hyperparameter for the InfoNCE loss, due to determining the number of negatives, we report experimental results with different values. The latent dimensionality of the embeddings is $d = 256$ . We set the temperature $\\tau$ to 0.1, and the weight of the contrastive loss term $\\lambda_{\\mathrm{NCE}}$ to 0.1. The threshold to filter negatives is set to 0.8. We provide experimental analyses to measure the sensitivity to these added hyperparameters.", + "bbox": [ + 75, + 196, + 470, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 345, + 202, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We start by describing the datasets and evaluation protocol used in the experiments (Section 4.1). We then report the performance of our model on our new retrieval benchmark along with comparison to prior work (Section 4.2). Next, we present our ablation study measuring the effects of the additional contrastive loss, the negative filtering, and the hyperparameters (Section 4.3). Finally, we provide qualitative results for retrieval (Section 4.4), and our use case of moment retrieval (Section 4.5).", + "bbox": [ + 75, + 371, + 470, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and evaluation", + "text_level": 1, + "bbox": [ + 76, + 500, + 294, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HumanML3D dataset (H3D) [15] provides natural language labels to describe the motions in AMASS [31] and HumanAct12 [16] motion capture collections. We follow the motion pre-processing procedure of [15], and apply the SMPL layer [28] to extract joint positions, canonicalize the skeletons to share the same topology (i.e., same bone lengths), then compute motion features (extracting local positions, velocities and foot contacts similar to Holden et al. [21]). The data is then augmented by mirroring left and right (both in motions and their corresponding texts). After this procedure, and following the official split, we obtain 23384, 1460, 4380 motions for the training, validation, and test sets, respectively. On average, each motion is annotated 3.0 times with different text. During training we randomly select one as the matching text, for testing we use the first text.", + "bbox": [ + 75, + 522, + 468, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "KIT Motion-Language dataset (KIT) [37] also come from motion capture data, with an emphasis on locomotion motions. It originally consists of 3911 motion sequences and 6278 text sentences. We pre-process the motions with the same procedure as in H3D. The data is split into 4888, 300, 830 motions for training, validation, and test sets, respectively. In this dataset, each motion is annotated 2.1 times on average.", + "bbox": [ + 75, + 750, + 470, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation protocol. We report standard retrieval performance measures: recall at several ranks, R@1, R@2, etc. for both text-to-motion and motion-to-text tasks (identical to the R", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Precision metrics adopted by Guo et al. [15]). Recall at rank $k$ measures the percentage of times the correct label is among the top $k$ results; therefore higher is better. We additionally report median rank (MedR), where lower is better. Note that the retrieval performance is evaluated on a gallery of unseen real motions (i.e., test set). A study of the synthesis performance of TMR can be found in Appendix B.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define several evaluation protocols, mainly changing the gallery set. (a) All the test set is used as a first protocol, without any modification. This set is partially problematic because there are repetitive texts across motions, or just minor differences (e.g., person vs human, walk vs walking). (b) All with threshold means we search over all the test set, but, in this case, we accept a retrieved motion as correct if its text label is similar to the query text above a threshold. For example, retrieving the motion corresponding to \"A human walks forward\" should be correct when the input query is \"Someone is walking forward\". We set a high threshold of 0.95 (scaled between [0, 1]) to remove very similar texts without removing too many fine-grained details. Appendix A provides statistics on how often similar text descriptions appear in the datasets. (c) Dissimilar subset refers to sampling 100 motion-text pairs whose texts are maximally far from each other (using an approximation of the quadratic knapsack problem [3]). This evaluation can be considered as an easy, but clean, subset of the previous ones. (d) Small batches is included to mimic the protocol described by Guo et al. [15], who randomly pick batches of 32 motion-text pairs and report the average performance. An ideal evaluation metric should not have randomness, and a gallery size of 32 is relatively easy compared to the previous protocols.", + "bbox": [ + 496, + 198, + 892, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. A new benchmark & comparison to prior work", + "text_level": 1, + "bbox": [ + 498, + 558, + 890, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present the performance of our model on this new retrieval benchmark, on H3D (Table 1) and KIT (Table 2) datasets, across all evaluation protocols. We also compare against prior work TEMOS [36] and Guo et al. [15]. For TEMOS, we retrain their model on both datasets to have a comparable benchmark since the original model differs in motion representation and lacks left/right data augmentation (and they only provide a KIT-pretrained model, not H3D). For [15], we take their publicly available models trained on these two datasets.", + "bbox": [ + 496, + 582, + 893, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TEMOS in particular is not designed to perform well on retrieval, since its cross-modal embedding space is only trained with positive pairs. However, Guo et al. train contrastively with negatives as well, using a margin loss [18]. For all 4 evaluation sets with varying difficulties, TMR outperforms the prior work, suggesting our model better captures the finegrained nature of motion descriptions. The model of [15] is adopted as part of motion synthesis evaluation in several works. TMR may therefore provide a better alternative. Our significant improvements over the state of the art can be credited to (i) jointly training for synthesis and retrieval, (ii) adopting the more recent contrastive objective, InfoNCE [34], while (iii) carefully", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9492", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3afcb079e8aed2e53f3d84be031af0eda3c7d94f7250a33f1a2adb7d3e509945.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ProtocolMethodsText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓
(a) AllTEMOS [36]2.124.095.878.2613.52173.03.864.546.949.3814.00183.25
Guo et al. [15]1.803.424.797.1212.4781.002.923.746.008.3612.9581.50
TMR5.6810.5914.0420.3430.9428.009.9512.4417.9523.5632.6928.50
(b) All with thresholdTEMOS [36]5.218.2211.1415.0922.1279.005.486.199.0012.0117.10129.0
Guo et al. [15]5.307.8310.7514.5922.5154.004.955.688.9311.6416.9469.50
TMR11.6015.3920.5027.7238.5219.0013.2015.7322.0327.6537.6321.50
(c) Dissimilar subsetTEMOS [36]33.0042.0049.0057.0066.004.0035.0044.0050.0056.0070.003.50
Guo et al. [15]34.0048.0057.0072.0084.003.0034.0047.0059.0072.0083.003.00
TMR47.0061.0071.0080.0086.002.0048.0063.0069.0080.0084.002.00
(d) Small batches [15]TEMOS [36]40.4953.5261.1470.9684.152.3339.9653.4961.7972.4085.892.33
Guo et al. [15]52.4871.0580.6589.6696.581.3952.0071.2181.1189.8796.781.38
TMR67.1681.3286.8191.4395.361.0467.9781.2086.3591.7095.271.03
", + "bbox": [ + 83, + 88, + 883, + 301 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ab545cb039191b6440d9d8e00b856d5bb7a972c639ac266d6e533def84821e18.jpg", + "table_caption": [ + "Table 1. Text-to-motion retrieval benchmark on HumanML3D: We establish four evaluation protocols as described in Section 4.1, with decreasing difficulty from (a) to (d). Our model TMR substantially outperforms the prior work of Guo et al. [15] and TEMOS [36], on the challenging H3D dataset." + ], + "table_footnote": [], + "table_body": "
ProtocolMethodsText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓
(a) AllTEMOS [36]7.1113.2517.5924.1035.6624.0011.6915.3020.1226.6336.3926.50
Guo et al. [15]3.376.9910.8416.8727.7128.004.946.5110.7216.1425.3028.50
TMR7.2313.9820.3628.3140.1217.0011.2013.8620.1228.0738.5518.00
(b) All with thresholdTEMOS [36]18.5524.3430.8442.2956.397.0017.7122.4128.8035.4247.1113.25
Guo et al. [15]13.2522.6529.7639.0449.5211.0010.4813.9820.4827.9538.5517.25
TMR24.5830.2441.9350.4860.365.0019.6423.7332.5341.2053.019.50
(c) Dissimilar subsetTEMOS [36]24.0040.0046.0054.0070.005.0033.0039.0045.0049.0064.006.50
Guo et al. [15]16.0029.0036.0048.0066.006.0024.0029.0036.0046.0066.007.00
TMR26.0046.0060.0070.0083.003.0034.0045.0060.0069.0082.003.50
(d) Small batches [15]TEMOS [36]43.8858.2567.0074.0084.752.0641.8855.8865.6275.2585.752.25
Guo et al. [15]42.2562.6275.1287.5096.121.8839.7562.7573.6286.8895.881.95
TMR49.2569.7578.2587.8895.001.5050.1267.1276.8888.8894.751.53
", + "bbox": [ + 83, + 357, + 883, + 570 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cd2b7e05c8e36c910e8020945fcda298b15c7e845990901e9b09351014ec8167.jpg", + "table_caption": [ + "Table 2. Text-to-motion retrieval benchmark on KIT-ML: As in Table 1, we report the four evaluation protocols, this time on the KIT dataset. Again, TMR significantly improves over Guo et al. [15] and TEMOS [36] across all protocols and metrics." + ], + "table_footnote": [], + "table_body": "
Motion \nRecons.InfoNCEMarginText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
XX✓15.0622.1725.7812.008.1911.5716.3919.50
X✓X19.7625.3036.876.0017.4719.7630.609.50
✓XX18.5524.3430.847.0017.7122.4128.8013.25
✓X✓19.8824.4634.467.0014.7019.7628.1912.50
✓✓X24.5830.2441.935.0019.6423.7332.539.50
", + "bbox": [ + 81, + 623, + 460, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Losses: We experiment with various loss definitions (i) with/without the motion reconstruction, and (ii) the choice of the contrastive loss between InfoNCE and margin-based. We see that InfoNCE [34] is a better alternative to the contrastive loss with Euclidean margin [18] (employed by Guo et al. [15]). The reconstruction loss through the motion decoder branch further boosts the results.", + "bbox": [ + 76, + 705, + 468, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "eliminating wrong negatives. In the following, we ablate these components in controlled experiments.", + "bbox": [ + 76, + 804, + 468, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 845, + 227, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The rest of the quantitative evaluation uses the ' (b) All with threshold' evaluation protocol, on the KIT dataset.", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Which losses matter? Table 3 compares several variants of TMR where we check (a) whether the jointly trained motion synthesis branch helps retrieval, and (b) how important the form of the contrastive loss is. When removing the synthesis branch and only using the contrastive loss, we perform a deterministic encoding (i.e., with a single token instead of two tokens $\\mu$ , $\\sigma$ ). First, we see that the motion synthesis branch certainly improves results compared with only training using a contrastive loss (e.g., 41.93 vs 36.87 R@3). This possibly forces the latent vector to capture the full content of the input text (i.e., instead of picking up on a subset of words, or bag-of-words [11, 55] upon finding a shortcut that satisfies the contrastive loss). Second, in the presence of a contrastive loss, the InfoNCE formulation is significantly better than the margin loss employed by previous work [15] (41.93 vs 34.46 R@1). Note that for this experiment, we keep the same negative filtering for both the margin loss and InfoNCE (we provide additional experiments without the negative filtering in Appendix B).", + "bbox": [ + 496, + 625, + 893, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9493", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1d329fd76a290f3b80e2dc5ffbd0392d811ceb522155dcf1b61bf853aa6bbf6e.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Filtering negatives: We compare several threshold values for filtering negatives from the loss comparison due to having similar texts. We observe that removing negatives based on text similarity above 0.8 (from a scale between [0,1]) performs well overall." + ], + "table_body": "
ThresholdText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
0.5519.4023.2530.489.0017.8321.6929.5214.00
0.6017.9526.8736.876.0020.6024.7031.8111.25
0.6523.0128.6736.397.0019.0421.6929.7611.50
0.7022.2929.6438.806.0018.1922.7732.059.00
0.7520.0027.1137.836.0020.2424.4634.229.50
0.8024.5830.2441.935.0019.6423.7332.539.50
0.8521.4525.7838.436.5020.8424.1033.379.50
0.9023.2530.1240.486.0020.0025.1833.139.50
0.9520.4826.9938.436.0019.2823.3731.9310.25
X22.1727.8336.027.0016.7521.3332.1711.50
", + "bbox": [ + 81, + 88, + 460, + 232 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effect of filtering negatives. As explained in Section 3.3, during training we filter out pairs whose texts are closer than a threshold in an embedding space, and do not count them in the contrastive loss computation. Note that we still keep each item in the batch for the motion synthesis objective. In Table 4, we perform experiments with a range of different values for this threshold selection. On a scale between [0, 1], a threshold of 0.8 has the best results, balancing keeping a sufficient number of negatives and removing the wrong ones. Without filtering at all, the performance remains at $36.02\\mathrm{R}@\\mathfrak{z}3$ (compared to 41.93). We provide statistics on the percentage of filtered pairs in Appendix A.", + "bbox": [ + 75, + 303, + 468, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameters of the contrastive training. We show the sensitivity of our model to several hyperparameters added when extending TEMOS: (i) temperature $\\tau$ of the cross entropy of InfoNCE [34] in Eq.1, (ii) the $\\lambda_{\\mathrm{NCE}}$ weighting parameter, and (iii) the batch size, which determines the number of negatives. We see in Table 5 that the model is indeed sensitive to the temperature, which is a common observation in other settings. The weight parameter and the batch size are relatively less important while also influencing the results to a certain extent. An experiment with the latent dimensionality hyperparameter can be found in Appendix B.", + "bbox": [ + 75, + 486, + 470, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative results", + "text_level": 1, + "bbox": [ + 76, + 664, + 253, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 3, we provide sample qualitative results for text-to-motion retrieval on the full test set of H3D. For each query text displayed on the left, top-5 retrieved motions are shown on the right along with their similarity scores. Note that the ground-truth text labels (at the bottom of each motion) for the retrieved motions are not used, and the gallery motions are unseen at training. For the first two examples with 'playing violin' and 'handstand', we retrieve the ground-truth motion at rank 1. We observe that the next ranked motions depict visually similar motions as well (e.g., 'cartwheel' involves standing on the hands). For the free-from prompt example 'Someone is swimming' (i.e., the exact text does not appear in the gallery), the three first motions resemble or involve the swimming action, whereas motions at ranks 4 and 5 are incorrect. We notice that", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1ee01f8a42c4ff91b6f4a83e2f98b1ea22ef122c6730c2bf4b5113f65f91fdc6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Temp.Text-motion retrievalMotion-text retrieval
τR@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
0.0019.5221.8127.2312.007.479.7616.5115.50
0.0121.4529.0438.806.0021.0827.1133.619.50
0.124.5830.2441.935.0019.6423.7332.539.50
1.01.081.933.61306.51.811.932.41372.0
", + "bbox": [ + 506, + 89, + 887, + 169 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a7d3f287d00acf66d9d250614857db39117aa75bd8c798951bc33f9cf62c3921.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
Weight λNCEText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
0.00118.5523.2536.757.0018.1924.3431.4511.50
0.0120.8426.9937.237.0018.9223.1332.1710.25
0.124.5830.2441.935.0019.6423.7332.539.50
1.019.5224.4634.467.0019.0424.3435.069.50
", + "bbox": [ + 506, + 189, + 885, + 271 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c9f8c1baabceadd3e65ba948a32c4af441ff8ff5145e879a712d7d0c146642e7.jpg", + "table_caption": [ + "(b)" + ], + "table_footnote": [ + "Table 5. Hyperparameters of the contrastive training: We measure the sensitivity to the parameters $\\tau$ (temperature), $\\lambda_{c}$ the weight of the contrastive loss, and the batch size. Note that the learning rate is proportionally altered when changing the batch size. We display a wide range of values to show the full trends." + ], + "table_body": "
Batch sizeText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
1625.4231.5740.126.0020.3624.1033.738.00
3224.5830.2441.935.0019.6423.7332.539.50
6420.2426.5138.196.0019.5224.2232.059.50
12818.5528.8036.757.0014.9418.4326.1411.50
", + "bbox": [ + 504, + 292, + 885, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(c)", + "bbox": [ + 687, + 373, + 704, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the incorrect motions have a low similarity $(< 0.6)$ , and the human bodies are rotated similarly as in swimming. More qualitative results can be seen in Appendix C, as well as our supplementary video on the project page.", + "bbox": [ + 496, + 474, + 890, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Use case: Moment retrieval", + "text_level": 1, + "bbox": [ + 500, + 545, + 746, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While our focus is retrieval, once our model is trained, it can be used for a different use case. Here, we test the limits of our approach, by qualitatively evaluating the capability of TMR on the task of temporally localizing a natural language query in a long 3D motion sequence. This is similar in spirit to moment retrieval in videos [12, 13, 19, 25, 41]. It is also related to categorical action localization in 3D motions [48]; however, our input is free-form text instead of symbolic action classes.", + "bbox": [ + 496, + 568, + 890, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 4, we show four examples, where we apply a model pre-trained on H3D on BABEL sequences. In each example, the queried text is displayed on the top. The x-axis denotes the frame number, the green rectangle represents the ground-truth location for the given action, and the dashed red line marks the localization with the maximum similarity. We simply compute the motion features in a sliding window manner. The similarity between the text label and a 20-frame window centered at each frame is shown in the y-axis as a 1D plot over time. Despite our model not being trained for temporal localization, we observe its grounding potential. Moreover, there exists a domain gap between BABEL label used at test time and H3D used for training. Quantitative evaluation and more qualitative examples are included in Appendix B.", + "bbox": [ + 496, + 689, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9494", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/064821e401a50ca69dde23dd5a446a4b51b8e4efa18757d1fd7b136ea203c749.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 118, + 200, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2bf02720a19f22ce9b82b0dcad9d9d8cb9e958c2e4657c6dcb42493d3a7da83.jpg", + "image_caption": [ + "A person appears to be playing the violin" + ], + "image_footnote": [], + "bbox": [ + 220, + 89, + 343, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95de5ac708fe12c406124aafeb2cd85604bbadcfa942bb983f28925b75979736.jpg", + "image_caption": [ + "Moving the hands and work some thing" + ], + "image_footnote": [], + "bbox": [ + 349, + 90, + 470, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/11dbe0118342598782f59ddb39f0816ae19701acad8cc74874486f3c7740b0b8.jpg", + "image_caption": [ + "$S = 0.85$", + "The man plays violin holding it in his right hand" + ], + "image_footnote": [], + "bbox": [ + 478, + 101, + 599, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a8ef19f55934eeb1450b68677175d78973c978e077db40c5ca1af7ebe474517d.jpg", + "image_caption": [ + "$S = 0.85$", + "The man plays the violin" + ], + "image_footnote": [], + "bbox": [ + 607, + 101, + 728, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84b3b304087e09182b8211dd9d308f76f0dea1531192733122e6995ce72207e0.jpg", + "image_caption": [ + "$S = 0.81$", + "The man holds something above his left shoulder and rubs it with his right hand" + ], + "image_footnote": [], + "bbox": [ + 736, + 101, + 857, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a87e77987a443f81c2ded9248616d3882553e376b11abe60c593628511da90c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 239, + 200, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a53020f79ba182d02c8ffa57d9e80687129bb6d6668dc5e2972ddce4c70f5cc0.jpg", + "image_caption": [ + "$S = 0.82$", + "A man steps forward and does a handstand" + ], + "image_footnote": [], + "bbox": [ + 220, + 217, + 343, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a899cfdc8d1e86545b6465259256ace6974da5bad8b9e8ae7e4f53821a02ccd6.jpg", + "image_caption": [ + "$S = 0.68$", + "Doing a cartwheel then jumping up and down." + ], + "image_footnote": [], + "bbox": [ + 351, + 217, + 470, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce655c700b0def4dd5399a9fa0672e60d0d1c5efeb27a382a7b8505d91e6e30f.jpg", + "image_caption": [ + "$S = 0.61$", + "The person does 2 cartwheels" + ], + "image_footnote": [], + "bbox": [ + 478, + 217, + 599, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00028c8c67deb4eced219a9704c277e46e53b827736780af3a969053dc255717.jpg", + "image_caption": [ + "$S = 0.48$", + "A person walks forward then turns completely around and does a cartwheel" + ], + "image_footnote": [], + "bbox": [ + 607, + 217, + 728, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25666dff1d750c88f5195a0fb8e01f23c420c365c65643b87981d1d82f9f87b2.jpg", + "image_caption": [ + "$S = 0.42$", + "The drunk guy struggles to walk down the street" + ], + "image_footnote": [], + "bbox": [ + 736, + 217, + 857, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ecc34bc462a299d4d3b13068a86d8c9240ca88185bfd2c5824cea64fcdda8c30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 88, + 354, + 189, + 385 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/563bf8f3a65becc7691d235a13bdfbe8dd2969478c21ba0e06dc8ab37377b706.jpg", + "image_caption": [ + "$S = 0.76$", + "The person was flying around like a fly" + ], + "image_footnote": [], + "bbox": [ + 220, + 334, + 341, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/948efe27aa5a168d1283927eba6a8a52508679771405fa71ef750c355641d53c.jpg", + "image_caption": [ + "$S = 0.71$", + "A person swam in free style" + ], + "image_footnote": [], + "bbox": [ + 351, + 334, + 470, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58a9e17f9f91e84c21fd1e6a282c0e564779e30e97fe9d0ce34c0ea1734ccda7.jpg", + "image_caption": [ + "$S = 0.62$", + "The person is preforming a swimming stroke know as the butterfly stroke. The arms swing from behind the head and reenter the water propelling the person forward." + ], + "image_footnote": [], + "bbox": [ + 478, + 334, + 599, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/027e487ee3bed9a7f267bf14b86ff3fae0e4ac966fa8160fcda2f450ad6b7e5d.jpg", + "image_caption": [ + "$S = 0.56$", + "A person lays on the ground" + ], + "image_footnote": [], + "bbox": [ + 607, + 334, + 728, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ffc9d2090b3d60d9023d6c26b88509a692df91c0a6d8c02e2567635b49f407fc.jpg", + "image_caption": [ + "$S = 0.52$", + "A man crawls forward on his stomach" + ], + "image_footnote": [], + "bbox": [ + 736, + 334, + 857, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/335b51424d89d8183ded2de2a693f99dd72e608153b97815ab28a6d08c710c08.jpg", + "image_caption": [ + "Figure 3. Qualitative retrieval results: We demonstrate example queries on the left, and corresponding retrieved motions on the right, ranked by text-motion similarity. The similarity values are displayed on the top. For each retrieved motion, we also show their accompanying ground-truth text label; note that we do not use these descriptions, but only provide them for analysis purposes. The motions from the gallery are all from the test set (unseen during training). In the first row, all top-5 retrieved motions correspond visually to 'playing violin' and the similarity scores are high $>0.80$ . In the second row, we correctly retrieve the 'handstand' motion at top-1, but the other motions mainly perform 'cartwheel' (which involves shortly standing on hands), but with a lower similarity score $<0.70$ . For the last example, we query a free-form text 'Someone is swimming', which does not exist in the gallery (but the word 'swim' does). The model successfully finds swimming motions among the top-3, and the other two motions involve the body parallel to the ground." + ], + "image_footnote": [], + "bbox": [ + 76, + 566, + 261, + 669 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ebf61b965f3627cbeeb951af27f629b49c414d8a813a4182e8c7875cfb206ede.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 568, + 429, + 669 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c0c7369183f4bd44168fe9db4c9952cce1984df268ab0362f3e12a0d8068122.jpg", + "image_caption": [ + "Figure 4. Moment retrieval: We plot the similarity between the temporally annotated BABEL text labels and the motions in a sliding window manner, and obtain a 1D signal over time (blue). We observe that a localization ability emerges from our model, even though it was not trained for temporal localization, and was not with the domain of BABEL labels. The ground-truth temporal span is denoted in green and the maximum similarity is marked with a dashed red line. More examples are provided in Appendix Figure A.2." + ], + "image_footnote": [], + "bbox": [ + 78, + 671, + 261, + 784 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/07058cae226e7cb70e2c403d32635699fdce8eba1c108a2f931801816fea53df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 672, + 429, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Limitations", + "text_level": 1, + "bbox": [ + 500, + 556, + 624, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our model comes with some limitations. Compared to the vast amount of data (e.g., 400M images [44]) in image-text collections used to train competitive foundation models, our motion-text training data can be considered extremely small (e.g., 23K motions in H3D). The generalization performance of motion retrieval models to in-the-wild motions is therefore limited. Data augmentations such as altering text can potentially help to a certain extent (cf. [6]); however, more motion capture is still needed. Another limitation concerns the case where one wishes to replace motion synthesis by retrieving a training motion. In this use case, the model requires the full encoded database (i.e., encoded training set) to be stored in memory, which can be memory-inefficient. Note that the encoded representation is considerably smaller than the original motion data. This could be dealt with using a hierarchical structuring of the motions and language descriptions (e.g. into upper-body motions, lower-body motions, sitting motions, standing motions, etc.).", + "bbox": [ + 496, + 580, + 893, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 844, + 612, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we focus on the relatively little-studied problem of motion retrieval with natural language queries. We", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9495", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "introduce TMR, a framework to jointly train text-to-motion retrieval and text-to-motion synthesis, with a special attention to the definition of negatives, taking into account the fine-grained nature of motion-language databases. We significantly improve over prior work, and provide a series of experiments highlighting the importance of each component.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Future work may consider incorporating a language synthesis branch, along with the motion synthesis branch, to build a symmetrical framework, which could bring further benefits.", + "bbox": [ + 76, + 181, + 470, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This work was granted access to the HPC resources of IDRIS under the allocation 2022-AD011012129R2 made by GENCI. GV acknowledges the ANR project CorVis ANR-21-CE23-0003-01. The authors would like to thank Lucas Ventura and Charles Raude.", + "bbox": [ + 76, + 234, + 468, + 303 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "MJB Disclosure: files.is.tue.mpg.de/black/CoI_ ICCV_2023.txt", + "bbox": [ + 76, + 305, + 468, + 332 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 347, + 168, + 363 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 2020. 3", + "[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2Pose: Natural language grounded pose forecasting. In International Conference on 3D Vision (3DV), 2019. 2, 3", + "[3] Méziane Aïder, Oussama Gacem, and Mhand Hifi. Branch and solve strategies-based algorithm for the quadratic multiple knapsack problem. Journal of the Operational Research Society, 2022. 5", + "[4] Okan Arikan, David A. Forsyth, and James F. O'Brien. Motion synthesis from annotations. ACM Transactions on Graphics (TOG), 2003. 1, 2", + "[5] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal action composition for 3D humans. In International Conference on 3D Vision (3DV), 2022. 1, 2", + "[6] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Güll Varol. SINC: Spatial composition of 3D human motions for simultaneous action generation. In International Conference on Computer Vision (ICCV), 2023. 8", + "[7] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In International Conference on Computer Vision (ICCV), 2021. 3, 4", + "[8] Michael Böttner and Simon Clavet. Motion matching - the road to next gen animation. In Nucl.ai, youtu.be/watch?v=z_wpgHFSWss&t=658s, 2015.2, 3", + "[9] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. In Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2", + "[10] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. MoFusion: A framework for denoising-diffusion-based motion synthesis. In Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[11] Sivan Doveh, Assaf Arbelle, Sivan Harary, Rameswar Panda, Roei Herzig, Eli Schwartz, Donghyun Kim, Raja Giryes, Rogerio Feris, Shimon Ullman, and Leonid Karlinsky. Teaching struc" + ], + "bbox": [ + 78, + 373, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tured vision & language concepts to vision & language models. In Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6", + "[12] Victor Escorcia, Mattia Soldan, Josef Sivic, Bernard Ghanem, and Bryan Russell. Temporal localization of moments in video collections with natural language. arXiv:1907.12763, 2019. 2, 7", + "[13] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In International Conference on Computer Vision (ICCV), 2017. 2, 7", + "[14] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In International Conference on Computer Vision (ICCV), 2021. 2", + "[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3D human motions from text. In Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6", + "[16] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2Motion: Conditioned generation of 3D human motions. In ACM International Conference on Multimedia (ACMMM), 2020. 2, 5", + "[17] Ikhsanul Habibie, Mohamed Elgharib, Kripashindu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. In SIGGRAPH, 2022. 2", + "[18] R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping. In Computer Vision and Pattern Recognition (CVPR), 2006. 3, 5, 6", + "[19] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In International Conference on Computer Vision (ICCV), 2017. 2, 7", + "[20] Daniel Holden, Oussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions on Graphics (TOG), 2020. 3", + "[21] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 2017. 5", + "[22] Sai Shashank Kalakonda, Shubh Maheshwari, Sarvadevabhatla, and Ravi Kiran. Action-GPT: Leveraging large-scale language models for improved and generalized zero shot action generation. arXiv:2211.15603, 2022. 2", + "[23] Lucas Kovar, Michael Gleicher, and Frédéric H. Pighin. Motion graphs. ACM Transactions on Graphics (TOG), 2002. 1", + "[24] Taeryung Lee, Gyeongsik Moon, and Kyoung Mu Lee. MultiAct: Long-term 3D human motion generation from multiple action labels. In AAAI Conference on Artificial Intelligence, 2023. 2", + "[25] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. TVR: A large-scale dataset for video-subtitle moment retrieval. In European Conference on Computer Vision (ECCV), 2020. 2, 7", + "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning (ICML), 2022. 2, 3, 4", + "[27] Angela S Lin, Lemeng Wu, Rodolfo Corona, Kevin Tai, Qixing Huang, and Raymond J Mooney. Generating animated videos of human activities from natural language descriptions. ViGIL NeurIPS Workshop, 2018. 2", + "[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person" + ], + "bbox": [ + 503, + 93, + 892, + 887 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9496", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "linear model. ACM Transactions on Graphics (TOG), 2015. 3, 5", + "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5", + "[30] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval and captioning. Neurocomputing, 2022. 3", + "[31] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision (ICCV), 2019. 5", + "[32] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers. In Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[33] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[34] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018. 2, 3, 4, 5, 6, 7", + "[35] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 1, 2, 3, 4", + "[36] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 5, 6", + "[37] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. *Big Data*, 2016. 2, 4, 5", + "[38] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english labels. In Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2, 3, 4", + "[40] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning (ICML), 2021. 2", + "[41] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. In Transactions of the Association for Computational Linguistics (TACL), 2013. 2, 7", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[43] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. In Energy Efficient Machine Learning and Cognitive Computing NeurIPS Workshop, 2019. 3" + ], + "bbox": [ + 78, + 90, + 468, + 890 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: open dataset of clip-filtered 400 million image-text pairs. In Data Centric AI NeurIPS Workshop, 2021. 2, 8", + "[45] H. Sidenbladh, M. J. Black, and L. Sigal. Implicit probabilistic models of human motion for synthesis and tracking. In European Conference on Computer Vision (ECCV), 2002. 2, 3", + "[46] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. MPNet: Masked and permuted pre-training for language understanding. In Neural Information Processing Systems (NeurIPS), 2020. 2", + "[47] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. MPNet: Masked and permuted pre-training for language understanding. In Neural Information Processing Systems (NeurIPS), 2020. 4", + "[48] Jiankai Sun, Bolei Zhou, Michael J. Black, and Arjun Chandrasekaran. LocATE: End-to-end localization of actions in 3D with transformers. arXiv:2203.10719, 2022. 7", + "[49] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In International Conference on Learning Representations (ICLR), 2023. 1, 2, 3", + "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Neural Information Processing Systems (NeurIPS), 2017. 3", + "[51] Ruben Villegas, Jimei Yang, Duygu Ceylan, and Honglak Lee. Neural kinematic networks for unsupervised motion retargeting. In Computer Vision and Pattern Recognition (CVPR), 2018. 3", + "[52] Jiashun Wang, Chao Wen, Yanwei Fu, Haitao Lin, Tianyun Zou, Xiangyang Xue, and Yinda Zhang. Neural pose transfer by spatially adaptive instance normalization. In Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[53] Sijie Yan, Zhizhong Li, Yuanjun Xiong, Huahan Yan, and Dahua Lin. Convolutional sequence generation for skeleton-based action synthesis. In International Conference on Computer Vision (ICCV), 2019. 2", + "[54] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. CoCa: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research (TMLR), 2022. 2, 3", + "[55] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In ICLR, 2022. 2, 6", + "[56] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2M-GPT: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2", + "[57] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. MotionDiffuse: Text-driven human motion generation with diffusion model. arXiv:2208.15001, 2022. 3", + "[58] Yan Zhang, Michael J. Black, and Siyu Tang. Perpetual motion: Generating unbounded human motion. arXiv:2007.13886, 2020. 2", + "[59] Rui Zhao, Hui Su, and Qiang Ji. Bayesian adversarial human motion synthesis. In CVPR, 2020. 2" + ], + "bbox": [ + 501, + 92, + 892, + 890 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9497", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/d1f3cbdb-e07a-4952-8052-a9aa7bed81da_model.json b/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/d1f3cbdb-e07a-4952-8052-a9aa7bed81da_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5651ada1ef077018d26646eb5121ea9d5c70a709 --- /dev/null +++ b/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/d1f3cbdb-e07a-4952-8052-a9aa7bed81da_model.json @@ -0,0 +1,2563 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.131, + 0.862, + 0.154 + ], + "angle": 0, + "content": "TMR: Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.18, + 0.737, + 0.254 + ], + "angle": 0, + "content": "Mathis Petrovich\\(^{1,2}\\) Michael J. Black\\(^{2}\\) Gül Varol\\(^{1}\\) \n\\(^{1}\\) LIGM, École des Ponts, Univ Gustave Eiffel, CNRS, France \n\\(^{2}\\) Max Planck Institute for Intelligent Systems, Tübingen, Germany \nhttps://mathis.petrovich.fr/tmr" + }, + { + "type": "title", + "bbox": [ + 0.236, + 0.288, + 0.312, + 0.304 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.474, + 0.565 + ], + "angle": 0, + "content": "In this paper, we present TMR, a simple yet effective approach for text to 3D human motion retrieval. While previous work has only treated retrieval as a proxy evaluation metric, we tackle it as a standalone task. Our method extends the state-of-the-art text-to-motion synthesis model TEMOS, and incorporates a contrastive loss to better structure the cross-modal latent space. We show that maintaining the motion generation loss, along with the contrastive training, is crucial to obtain good performance. We introduce a benchmark for evaluation and provide an in-depth analysis by reporting results on several protocols. Our extensive experiments on the KIT-ML and HumanML3D datasets show that TMR outperforms the prior work by a significant margin, for example reducing the median rank from 54 to 19. Finally, we showcase the potential of our approach on moment retrieval. Our code and models are publicly available at https://mathis.petrovich.fr/tmr." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.6, + 0.203, + 0.616 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.628, + 0.455, + 0.655 + ], + "angle": 0, + "content": "The language of movement cannot be translated into words. Barbara Mettler" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.472, + 0.901 + ], + "angle": 0, + "content": "We ask the question whether a cross-modal space exists between 3D human motions and language. Our goal is to retrieve the most relevant 3D human motion from a gallery, given a natural language query that describes the desired motion (as illustrated in Figure 1). While text-to-image retrieval is a well-established problem within the broader vision & language field [39], there has been less focus on the related task of text-to-motion retrieval. Searching an existing motion capture dataset based on text input can often serve as a viable alternative to text-to-human-motion synthesis in many applications, while also providing the added benefit that the retrieved motion is guaranteed to be realistic. Additionally, once a cross-modal embedding is built to map text and motions into a joint representation space, both of the symmetrical text-to-motion and motion-to-text tasks" + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.299, + 0.771, + 0.314 + ], + "angle": 0, + "content": "A person is stretching legs" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.322, + 0.887, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.515, + 0.893, + 0.571 + ], + "angle": 0, + "content": "Figure 1. Text-to-motion retrieval: We illustrate the task of text-based motion retrieval where the goal is to rank a gallery of motions according to their similarity to the given query in the form of a natural language description." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.895, + 0.654 + ], + "angle": 0, + "content": "can be performed. Such a retrieval-based solution has a range of applications, including automatically indexing large motion capture collections, and helping to initialize the cumbersome text labeling process, by assigning the nearest text to each motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.895, + 0.838 + ], + "angle": 0, + "content": "Let us first differentiate text-to-motion retrieval from text-to-motion synthesis. Motion synthesis [5, 9, 36, 49] involves generating new data samples that go beyond the existing training set, while motion retrieval searches through existing motion capture collections. For certain applications, reusing motions from a collection may be sufficient, provided the collection is large enough to contain what the user is searching for. Unlike generative models for motion synthesis, which struggle to produce physically plausible, realistic sequences [5, 35, 36], a retrieval model has the advantage that it always returns a realistic motion. With this motivation, we pose the problem as a nearest neighbor search through a cross-modal text-motion space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Early work performs search through motion databases to build motion graphs [4, 23] by finding paths between existing motions and synthesizing new motions by stitching motions together with generated transitions. If the motion database is" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9488" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.379 + ], + "angle": 0, + "content": "labeled with actions, the user can specify a series of actions to combine [4]. In contrast, our search database is not labeled with text. Motion matching [8], on the other hand, seeks to find the animation that best fits the current motion by searching a database of animations, doing motion-to-motion retrieval [45]. Our framework fundamentally differs from these lines of work in that our task is multi-modal, i.e., user query is text, which is compared against motions. The most similar work to ours is the very recent model from Guo et al. [15], which trains for a joint embedding space between the two modalities. This model is only used to provide a performance measure for motion synthesis tasks, by querying a generated motion within a gallery of 32 descriptions (i.e., motion-to-text retrieval), and counting how many times the correct text is retrieved1. While this can be considered as the first text-motion retrieval model in the literature, its main limitation is the low performance, in particular when the gallery contains fine-grained descriptions. We substantially improve over [15], by incorporating a joint synthesis & retrieval framework, as well as a more powerful contrastive training [34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.473, + 0.59 + ], + "angle": 0, + "content": "We get inspiration from image-text models such as BLIP [26] and CoCa [54], which formulate a multi-task objective. Besides the standard dual-encoder matching (such as CLIP [39] with two unimodal encoders for image and text), [26, 54] also employ a text synthesis branch, performing image captioning. Such a generative capability potentially helps the model go beyond 'bag-of-words' understanding of vision-language concepts, observed for the naive contrastive models [11, 55]. In our case, we depart from TEMOS [36] which already has a synthesis branch to generate motions from text. We incorporate a cross-modal contrastive loss (i.e., InfoNCE [34]) in this framework to jointly train text-to-motion synthesis and text-to-motion retrieval tasks. We empirically demonstrate significant improvements with this approach when ablating the importance of each task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.473, + 0.865 + ], + "angle": 0, + "content": "Text-motion data differs from its text-image counterparts particularly due to the nature of motion descriptions. In fact, for an off-the-shelf large language model, sentences describing different motions tend to be similar, since they fall within the same topic of human motions. For example, the text-text cosine similarities [46] after encoding motion descriptions from the KIT training set [37] are on average 0.71 on a scale between [0, 1], while this value is 0.56 (almost orthogonal) on a random subset of LAION [44] image descriptions with the same size. This poses several challenges. Typical motion datasets [15, 37, 38] contain similar motions with different accompanying texts, e.g., 'person walks', 'human walks', as well as similar texts with different motions, e.g., 'walk backwards', 'walk forwards'. With naive contrastive training [34], one would make all samples within a batch as negatives, except the corresponding label for a given anchor. In this work, we take into account the fact that there are potentially significant similarities between pairs within a batch. Specifically, we discard pairs that have a text-text" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "similarity in their labels that is above a certain threshold. Such careful negative sampling leads to performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.895, + 0.334 + ], + "angle": 0, + "content": "In this paper, we illustrate an additional use case for our retrieval model – zero-shot temporal localization – and highlight this task as a potential future avenue for research. Similar to temporal localization in videos with natural language queries [12, 13, 19, 25, 41], also referred to as “moment retrieval”, we showcase the grounding capability of our model by directly applying it on long motion sequences to retrieve corresponding moments. We illustrate results on the BABEL dataset [38], which typically contains a series of text annotations for each motion sequence. Note that the task is zero-shot, because the model has not been trained for localization, and at the same time has not seen BABEL labels, which come from a different domain (e.g., typically action-like descriptions instead of full sentences)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.895, + 0.457 + ], + "angle": 0, + "content": "Our contributions are the following: (i) We address the little-studied problem of text-to-motion retrieval, and introduce a series of evaluation benchmarks with varying difficulty. (ii) We propose a joint synthesis and retrieval framework, as well as negative filtering, and obtain state-of-the-art performance on text-motion retrieval. (iii) We provide extensive experiments to analyze the effects of each component in controlled settings. Our code and models are publicly available2." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.474, + 0.631, + 0.49 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.545 + ], + "angle": 0, + "content": "We present an overview of the most closely related work on text-to-motion synthesis and retrieval, as well as a brief discussion on cross-modal retrieval works." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.895, + 0.873 + ], + "angle": 0, + "content": "Text and human motion. The research on human motion modeling has recently witnessed an increasing interest in bridging the gap between semantics and 3D human body motions, in particular for text-conditioned motion synthesis. Unlike unconstrained 3D human motion synthesis [53, 58, 59], action-conditioned [16, 35], or text-conditioned [5, 9, 14, 22, 24, 36, 49, 56] models add semantic controls to the generation process. The goal of these works is to generate either deterministically [2, 14, 27] or probabilistically [36], motion sequences that are faithful to the textual description inputs. Note that this is different than gesture synthesis from speech [17], in that the text describes the motion content. However, despite remarkable progress in text-to-image synthesis [40, 42], text-to-motion synthesis remains at a very nascent stage. The realism of the synthesized motions is limited, e.g., foot sliding artifacts [36]. We turn to text-to-motion retrieval as another alternative, and perhaps complementary approach to obtain motions for a given textual description. Our focus is therefore different from the work on synthesis. However, we do make use of a motion synthesis branch to aid the retrieval task." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "1While the paper [15] describes a motion-to-text retrieval metric, we notice that the provided code performs text-to-motion retrieval." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.761, + 0.9 + ], + "angle": 0, + "content": "2https://mathis.petrovich.fr/tmr" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9489" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.227 + ], + "angle": 0, + "content": "Motion retrieval is relatively less explored. As briefly mentioned in Section 1, motion-to-motion retrieval (e.g., motion matching [8, 20, 45]) methods exist. However, the text-to-motion retrieval task is more challenging due to being cross-modal, i.e., nearest neighbor search across text and motion modalities. Within this category, the very recent work of Guo et al. [15] trains a retrieval model purely for evaluation purposes, and applies a margin-based contrastive loss [18], using Euclidean distance between all pairs within a batch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.474, + 0.381 + ], + "angle": 0, + "content": "Two works are particularly relevant to ours. Firstly, we build on the TEMOS [36] text-to-motion synthesis model, which also has a cross-modal embedding space. However, text and motion embeddings are encouraged to be similar only across positive pairs. We therefore add a contrastive training strategy to incorporate negatives, consequently improving its retrieval capability from a large gallery of fine-grained motions. Secondly, we compare to the aforementioned method of [15], whose motion-to-text retrieval model is adopted for measuring text-to-motion generation performance automatically by other works [10, 49, 57]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.471, + 0.577 + ], + "angle": 0, + "content": "Cross-modal retrieval. Among widely adopted vision & language retrieval models, some successful examples include CLIP [39], BLIP [26], and CoCa [54] for images, and MILNCE [33], Frozen [7], and CLIP4Clip [30] for videos. They all use variants of cross-modal contrastive learning techniques, such as InfoNCE [34], which we also employ in this work. As discussed in Section 1, we draw inspiration from BLIP [26] and CoCa [54], which add synthesis branches to standard retrieval frameworks. Our approach is similar in spirit to these works in that we perform a cross-modal vision & language retrieval task, but differ in focusing on 3D human motion retrieval, which, to the best of our knowledge, has not been benchmarked." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.291, + 0.605 + ], + "angle": 0, + "content": "3. Text-to-motion retrieval" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.706 + ], + "angle": 0, + "content": "In this section, we introduce the task and the terminology associated with text-to-motion retrieval (Section 3.1). Next, we present our model, named TMR, and its training protocol (Section 3.2). We then explain our simple approach for identifying and filtering incorrect negatives (Section 3.3), followed by a discussion of the implementation details (Section 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.197, + 0.726 + ], + "angle": 0, + "content": "3.1. Definitions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Given a natural language query \\( T \\), such as 'A person walks and then makes a right turn', the goal is to rank the motions from a database (i.e., the gallery) according to their semantic correspondence with the text query, and to retrieve the motion that matches best to the textual description. In other words, the task involves sorting the database so that the top ranked motions are the most relevant matches, i.e., creating a search engine to index motions. Additionally, we define the symmetric (and complementary) task, namely motion-to-text retrieval, where the aim is to retrieve the most suitable caption that matches a given motion from a database of texts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.227 + ], + "angle": 0, + "content": "3D human motion refers to a sequence of human poses. The task does not impose any limitations on the type of representation used, such as joint positions, rotations, or parametric models such as SMPL [28]. As detailed in Section 4.1, we choose to use the representation employed by Guo et al. [15] to facilitate comparisons with previous work. While our experiments are based on the SMPL skeleton, our method is applicable to any skeleton topology. One could alternatively use retargetting [1, 51, 52] to adapt to other body representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.895, + 0.351 + ], + "angle": 0, + "content": "Text description refers to a sequence of words describing the action performed by a human in natural language. We do not restrict the format of the motion description. The text can be simply an action name (e.g., 'walk') or a full sentence (e.g., 'a human is walking'). The sentences can be fine-grained (e.g., 'a human is walking in a circle slowly'), and may contain one or several actions, simultaneously (e.g., 'walking while waving') or sequentially (e.g., 'walking then sitting')." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.361, + 0.837, + 0.377 + ], + "angle": 0, + "content": "3.2. Joint training of retrieval and synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.491 + ], + "angle": 0, + "content": "We introduce TMR, which extends the Transformer-based text-to-motion synthesis model TEMOS [36] by incorporating additional losses to make it suitable for the retrieval task. The architecture consists of two independent encoders for inputting motion and text, as well as a decoder that outputs motion (see Figure 2 for an overview). In the following, we review TEMOS [36] components and our added contrastive training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.893, + 0.673 + ], + "angle": 0, + "content": "Dual encoders. One approach to solving cross-modal retrieval tasks involves defining a similarity function between the two modalities. In our case, the two modalities are text and motion. The similarity function can be applied to compare a given query with each element in the database, and the maximum value would indicate the best match. In this paper, we follow the approach taken by previous work on metric learning, such as CLIP [39], by defining one encoder for each modality and then computing the cosine similarity between their respective embeddings. Such dual embedding has the advantage of fast inference time since the gallery embeddings can be computed and stored beforehand [32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.895, + 0.885 + ], + "angle": 0, + "content": "Our model is built upon the components of TEMOS [36], which already provides a motion encoder and a text encoder, mapping them to a joint space (building on the idea from Language2Pose [2]); this serves as a strong baseline for our work. Both motion and text encoders are Transformer encoders [50] with additional learnable distribution parameters, as in the VAE-based ACTOR [35]. They are probabilistic in nature, outputting parameters of a Gaussian distribution \\((\\mu\\) and \\(\\Sigma)\\) from which a latent vector \\(z\\in \\mathbb{R}^d\\) can be sampled. While the text encoder takes text features from a pre-trained and frozen DistilBERT [43] network as input, the motion sequence is fed directly in the motion encoder. Note that when performing retrieval, we directly use the output embedding that corresponds to the mean token \\((\\mu^{M}\\) for motion, \\(\\mu^T\\) for text)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Motion decoder. TEMOS is trained for the task of motion" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9490" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.87, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.296, + 0.896, + 0.381 + ], + "angle": 0, + "content": "Figure 2. Joint motion retrieval and synthesis: A simplified view of our TMR framework is presented, where we focus on the similarity matrix defined between text-motion pairs within a batch. Here, we show a batch of 3 samples for illustration purposes. The goal of the contrastive objective is to maximize the diagonal denoting positive pairs (green), and to minimize the off-diagonal negative pair items that have text similarity below a threshold (red). In this example, remaining similarities \\( S_{23} \\) and \\( S_{32} \\) are discarded from the loss computation because there is high text similarity between \\( T_{2} \\) and \\( T_{3} \\). The rest of the model remains similar to TEMOS [36], which decodes a motion from both text \\( z^{T_i} \\) and motion \\( z^{M_j} \\) latent vectors. See text for further details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.475, + 0.499 + ], + "angle": 0, + "content": "synthesis and comes equipped with a motion decoder branch. This decoder is identical to the one used in ACTOR [35], which supports variable-duration generation. More specifically, it takes a latent vector \\( z \\in \\mathbb{R}^d \\) and a sinusoidal positional encoding as input, and generates a motion non-autoregressively through a single forward pass. We show in Section 4.2 that keeping this branch helps improve the results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.474, + 0.693 + ], + "angle": 0, + "content": "TEMOS losses. We keep the same base set of losses of [36], defined as the weighted sum \\(\\mathcal{L}_{\\mathrm{TEMOS}} = \\mathcal{L}_{\\mathrm{R}} + \\lambda_{\\mathrm{KL}}\\mathcal{L}_{\\mathrm{KL}} + \\lambda_{\\mathrm{E}}\\mathcal{L}_{\\mathrm{E}}\\). In summary, a reconstruction loss term \\(\\mathcal{L}_{\\mathrm{R}}\\) measures the motion reconstruction given text or motion input (via a smooth L1 loss). A Kullback-Leibler (KL) divergence loss term \\(\\mathcal{L}_{\\mathrm{KL}}\\) is composed of four losses: two of them to regularize each encoded distributions \\(-\\mathcal{N}(\\mu^M,\\Sigma^M)\\) for motion and \\(\\mathcal{N}(\\mu^T,\\Sigma^T)\\) for text - to come from a normal distribution \\(\\mathcal{N}(0,I)\\). The other two enforce distribution similarity between the two modalities. A cross-modal embedding similarity loss \\(\\mathcal{L}_{\\mathrm{E}}\\) enforces both text \\(z^{T}\\) and motion \\(z^{M}\\) latent codes to be similar to each other (with a smooth L1 loss). We set \\(\\lambda_{\\mathrm{KL}}\\) and \\(\\lambda_{\\mathrm{E}}\\) to \\(10^{-5}\\) in our experiments as in [36]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Contrastive training. While TEMOS has a cross-modal embedding space, its major drawback to be usable as an effective retrieval model is that it is never trained with negatives, but only positive motion-text pairs. To overcome this limitation, we incorporate a contrastive training with the usage of negative samples to better structure the latent space. Given a batch of N (positive) pairs of latent codes \\((z_1^T,z_1^M),\\dots,(z_N^T,z_N^M)\\), we define any pair \\((z_i^T,z_j^M)\\) with \\(i\\neq j\\) as negative. The similarity matrix \\(S\\) computes the pairwise cosine similarities for all pairs in the batch \\(S_{ij} = \\cos (z_i^T,z_j^M)\\). In contrast to Guo et al. [15], who consider one random negative per batch (and use a margin loss), we adopt the more recent formulation of InfoNCE [34], which was proven effective in other work [7, 26, 39]. This loss term" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.392, + 0.667, + 0.406 + ], + "angle": 0, + "content": "can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.413, + 0.893, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {N C E}} = - \\frac {1}{2 N} \\sum_ {i} \\left(\\log \\frac {\\exp S _ {i i} / \\tau}{\\sum_ {j} \\exp S _ {i j} / \\tau} + \\log \\frac {\\exp S _ {i i} / \\tau}{\\sum_ {j} \\exp S _ {j i} / \\tau}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.46, + 0.774, + 0.474 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.475, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Training loss. The total loss we use to train TMR is the weighted sum \\(\\mathcal{L}_{\\mathrm{TEMOS}} + \\lambda_{\\mathrm{NCE}}\\mathcal{L}_{\\mathrm{NCE}}\\) where \\(\\lambda_{\\mathrm{NCE}}\\) controls the importance of the contrastive loss." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.53, + 0.681, + 0.546 + ], + "angle": 0, + "content": "3.3. Filtering negatives" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.614 + ], + "angle": 0, + "content": "As mentioned in Section 1, the descriptions accompanying motion capture collections can be repetitive or similar across the training motions. We wish to prevent defining negatives between text-motion pairs that contain similar descriptions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.719 + ], + "angle": 0, + "content": "Consider for example the two text descriptions, \"A human making a counterclockwise turn\" and \"A person walks quarter a circle to the left\". In the KIT-ML benchmark [37], these two descriptions appear as two different annotations for the same motion. Due to the flexibility and ambiguity of natural language, different words may describe the same concepts (e.g., 'counterclockwise', 'circle to the left')." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.825 + ], + "angle": 0, + "content": "During training, the random selection of batches can adversely affect the results because the model may have to push away two latent vectors that correspond to similar meanings. This can force the network to focus on unimportant details (e.g., 'someone' vs 'human'), ultimately resulting in decreased performance due to unstable behavior and reduced robustness to text variations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To alleviate this issue, we leverage an external large language model to provide sentence similarity scores. In particular, we use MPNet [47] to encode sentences and compute similarities between two text descriptions. We then determine whether to filter a pair of text descriptions \\((t_{1}, t_{2})\\) if their similarity" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9491" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "is higher than a certain threshold, referring them as 'wrong negatives'. During training, we filter wrong negative pairs from the loss computation. Note that we refrain from defining them as positives, as the language model may also incorrectly say two descriptions are similar when they are not." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.175, + 0.293, + 0.192 + ], + "angle": 0, + "content": "3.4. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.471, + 0.336 + ], + "angle": 0, + "content": "We use the AdamW optimizer [29] with a learning rate of \\(10^{-4}\\) and a batch size of 32. Since the batch size can be an important hyperparameter for the InfoNCE loss, due to determining the number of negatives, we report experimental results with different values. The latent dimensionality of the embeddings is \\(d = 256\\). We set the temperature \\(\\tau\\) to 0.1, and the weight of the contrastive loss term \\(\\lambda_{\\mathrm{NCE}}\\) to 0.1. The threshold to filter negatives is set to 0.8. We provide experimental analyses to measure the sensitivity to these added hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.347, + 0.204, + 0.364 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.472, + 0.493 + ], + "angle": 0, + "content": "We start by describing the datasets and evaluation protocol used in the experiments (Section 4.1). We then report the performance of our model on our new retrieval benchmark along with comparison to prior work (Section 4.2). Next, we present our ablation study measuring the effects of the additional contrastive loss, the negative filtering, and the hyperparameters (Section 4.3). Finally, we provide qualitative results for retrieval (Section 4.4), and our use case of moment retrieval (Section 4.5)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.501, + 0.295, + 0.515 + ], + "angle": 0, + "content": "4.1. Datasets and evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.749 + ], + "angle": 0, + "content": "HumanML3D dataset (H3D) [15] provides natural language labels to describe the motions in AMASS [31] and HumanAct12 [16] motion capture collections. We follow the motion pre-processing procedure of [15], and apply the SMPL layer [28] to extract joint positions, canonicalize the skeletons to share the same topology (i.e., same bone lengths), then compute motion features (extracting local positions, velocities and foot contacts similar to Holden et al. [21]). The data is then augmented by mirroring left and right (both in motions and their corresponding texts). After this procedure, and following the official split, we obtain 23384, 1460, 4380 motions for the training, validation, and test sets, respectively. On average, each motion is annotated 3.0 times with different text. During training we randomly select one as the matching text, for testing we use the first text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.856 + ], + "angle": 0, + "content": "KIT Motion-Language dataset (KIT) [37] also come from motion capture data, with an emphasis on locomotion motions. It originally consists of 3911 motion sequences and 6278 text sentences. We pre-process the motions with the same procedure as in H3D. The data is split into 4888, 300, 830 motions for training, validation, and test sets, respectively. In this dataset, each motion is annotated 2.1 times on average." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Evaluation protocol. We report standard retrieval performance measures: recall at several ranks, R@1, R@2, etc. for both text-to-motion and motion-to-text tasks (identical to the R" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Precision metrics adopted by Guo et al. [15]). Recall at rank \\( k \\) measures the percentage of times the correct label is among the top \\( k \\) results; therefore higher is better. We additionally report median rank (MedR), where lower is better. Note that the retrieval performance is evaluated on a gallery of unseen real motions (i.e., test set). A study of the synthesis performance of TMR can be found in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.893, + 0.547 + ], + "angle": 0, + "content": "We define several evaluation protocols, mainly changing the gallery set. (a) All the test set is used as a first protocol, without any modification. This set is partially problematic because there are repetitive texts across motions, or just minor differences (e.g., person vs human, walk vs walking). (b) All with threshold means we search over all the test set, but, in this case, we accept a retrieved motion as correct if its text label is similar to the query text above a threshold. For example, retrieving the motion corresponding to \"A human walks forward\" should be correct when the input query is \"Someone is walking forward\". We set a high threshold of 0.95 (scaled between [0, 1]) to remove very similar texts without removing too many fine-grained details. Appendix A provides statistics on how often similar text descriptions appear in the datasets. (c) Dissimilar subset refers to sampling 100 motion-text pairs whose texts are maximally far from each other (using an approximation of the quadratic knapsack problem [3]). This evaluation can be considered as an easy, but clean, subset of the previous ones. (d) Small batches is included to mimic the protocol described by Guo et al. [15], who randomly pick batches of 32 motion-text pairs and report the average performance. An ideal evaluation metric should not have randomness, and a gallery size of 32 is relatively easy compared to the previous protocols." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.559, + 0.892, + 0.575 + ], + "angle": 0, + "content": "4.2. A new benchmark & comparison to prior work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.894, + 0.718 + ], + "angle": 0, + "content": "We present the performance of our model on this new retrieval benchmark, on H3D (Table 1) and KIT (Table 2) datasets, across all evaluation protocols. We also compare against prior work TEMOS [36] and Guo et al. [15]. For TEMOS, we retrain their model on both datasets to have a comparable benchmark since the original model differs in motion representation and lacks left/right data augmentation (and they only provide a KIT-pretrained model, not H3D). For [15], we take their publicly available models trained on these two datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.901 + ], + "angle": 0, + "content": "TEMOS in particular is not designed to perform well on retrieval, since its cross-modal embedding space is only trained with positive pairs. However, Guo et al. train contrastively with negatives as well, using a margin loss [18]. For all 4 evaluation sets with varying difficulties, TMR outperforms the prior work, suggesting our model better captures the finegrained nature of motion descriptions. The model of [15] is adopted as part of motion synthesis evaluation in several works. TMR may therefore provide a better alternative. Our significant improvements over the state of the art can be credited to (i) jointly training for synthesis and retrieval, (ii) adopting the more recent contrastive objective, InfoNCE [34], while (iii) carefully" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9492" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.885, + 0.302 + ], + "angle": 0, + "content": "
ProtocolMethodsText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓
(a) AllTEMOS [36]2.124.095.878.2613.52173.03.864.546.949.3814.00183.25
Guo et al. [15]1.803.424.797.1212.4781.002.923.746.008.3612.9581.50
TMR5.6810.5914.0420.3430.9428.009.9512.4417.9523.5632.6928.50
(b) All with thresholdTEMOS [36]5.218.2211.1415.0922.1279.005.486.199.0012.0117.10129.0
Guo et al. [15]5.307.8310.7514.5922.5154.004.955.688.9311.6416.9469.50
TMR11.6015.3920.5027.7238.5219.0013.2015.7322.0327.6537.6321.50
(c) Dissimilar subsetTEMOS [36]33.0042.0049.0057.0066.004.0035.0044.0050.0056.0070.003.50
Guo et al. [15]34.0048.0057.0072.0084.003.0034.0047.0059.0072.0083.003.00
TMR47.0061.0071.0080.0086.002.0048.0063.0069.0080.0084.002.00
(d) Small batches [15]TEMOS [36]40.4953.5261.1470.9684.152.3339.9653.4961.7972.4085.892.33
Guo et al. [15]52.4871.0580.6589.6696.581.3952.0071.2181.1189.8796.781.38
TMR67.1681.3286.8191.4395.361.0467.9781.2086.3591.7095.271.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.304, + 0.893, + 0.348 + ], + "angle": 0, + "content": "Table 1. Text-to-motion retrieval benchmark on HumanML3D: We establish four evaluation protocols as described in Section 4.1, with decreasing difficulty from (a) to (d). Our model TMR substantially outperforms the prior work of Guo et al. [15] and TEMOS [36], on the challenging H3D dataset." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.358, + 0.885, + 0.571 + ], + "angle": 0, + "content": "
ProtocolMethodsText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓
(a) AllTEMOS [36]7.1113.2517.5924.1035.6624.0011.6915.3020.1226.6336.3926.50
Guo et al. [15]3.376.9910.8416.8727.7128.004.946.5110.7216.1425.3028.50
TMR7.2313.9820.3628.3140.1217.0011.2013.8620.1228.0738.5518.00
(b) All with thresholdTEMOS [36]18.5524.3430.8442.2956.397.0017.7122.4128.8035.4247.1113.25
Guo et al. [15]13.2522.6529.7639.0449.5211.0010.4813.9820.4827.9538.5517.25
TMR24.5830.2441.9350.4860.365.0019.6423.7332.5341.2053.019.50
(c) Dissimilar subsetTEMOS [36]24.0040.0046.0054.0070.005.0033.0039.0045.0049.0064.006.50
Guo et al. [15]16.0029.0036.0048.0066.006.0024.0029.0036.0046.0066.007.00
TMR26.0046.0060.0070.0083.003.0034.0045.0060.0069.0082.003.50
(d) Small batches [15]TEMOS [36]43.8858.2567.0074.0084.752.0641.8855.8865.6275.2585.752.25
Guo et al. [15]42.2562.6275.1287.5096.121.8839.7562.7573.6286.8895.881.95
TMR49.2569.7578.2587.8895.001.5050.1267.1276.8888.8894.751.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.573, + 0.893, + 0.603 + ], + "angle": 0, + "content": "Table 2. Text-to-motion retrieval benchmark on KIT-ML: As in Table 1, we report the four evaluation protocols, this time on the KIT dataset. Again, TMR significantly improves over Guo et al. [15] and TEMOS [36] across all protocols and metrics." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.624, + 0.462, + 0.704 + ], + "angle": 0, + "content": "
Motion \nRecons.InfoNCEMarginText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
XX✓15.0622.1725.7812.008.1911.5716.3919.50
X✓X19.7625.3036.876.0017.4719.7630.609.50
✓XX18.5524.3430.847.0017.7122.4128.8013.25
✓X✓19.8824.4634.467.0014.7019.7628.1912.50
✓✓X24.5830.2441.935.0019.6423.7332.539.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.706, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Table 3. Losses: We experiment with various loss definitions (i) with/without the motion reconstruction, and (ii) the choice of the contrastive loss between InfoNCE and margin-based. We see that InfoNCE [34] is a better alternative to the contrastive loss with Euclidean margin [18] (employed by Guo et al. [15]). The reconstruction loss through the motion decoder branch further boosts the results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.805, + 0.47, + 0.837 + ], + "angle": 0, + "content": "eliminating wrong negatives. In the following, we ablate these components in controlled experiments." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.847, + 0.228, + 0.864 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The rest of the quantitative evaluation uses the ' (b) All with threshold' evaluation protocol, on the KIT dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.895, + 0.899 + ], + "angle": 0, + "content": "Which losses matter? Table 3 compares several variants of TMR where we check (a) whether the jointly trained motion synthesis branch helps retrieval, and (b) how important the form of the contrastive loss is. When removing the synthesis branch and only using the contrastive loss, we perform a deterministic encoding (i.e., with a single token instead of two tokens \\(\\mu\\), \\(\\sigma\\)). First, we see that the motion synthesis branch certainly improves results compared with only training using a contrastive loss (e.g., 41.93 vs 36.87 R@3). This possibly forces the latent vector to capture the full content of the input text (i.e., instead of picking up on a subset of words, or bag-of-words [11, 55] upon finding a shortcut that satisfies the contrastive loss). Second, in the presence of a contrastive loss, the InfoNCE formulation is significantly better than the margin loss employed by previous work [15] (41.93 vs 34.46 R@1). Note that for this experiment, we keep the same negative filtering for both the margin loss and InfoNCE (we provide additional experiments without the negative filtering in Appendix B)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9493" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.462, + 0.233 + ], + "angle": 0, + "content": "
ThresholdText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
0.5519.4023.2530.489.0017.8321.6929.5214.00
0.6017.9526.8736.876.0020.6024.7031.8111.25
0.6523.0128.6736.397.0019.0421.6929.7611.50
0.7022.2929.6438.806.0018.1922.7732.059.00
0.7520.0027.1137.836.0020.2424.4634.229.50
0.8024.5830.2441.935.0019.6423.7332.539.50
0.8521.4525.7838.436.5020.8424.1033.379.50
0.9023.2530.1240.486.0020.0025.1833.139.50
0.9520.4826.9938.436.0019.2823.3731.9310.25
X22.1727.8336.027.0016.7521.3332.1711.50
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.235, + 0.47, + 0.292 + ], + "angle": 0, + "content": "Table 4. Filtering negatives: We compare several threshold values for filtering negatives from the loss comparison due to having similar texts. We observe that removing negatives based on text similarity above 0.8 (from a scale between [0,1]) performs well overall." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.47, + 0.485 + ], + "angle": 0, + "content": "The effect of filtering negatives. As explained in Section 3.3, during training we filter out pairs whose texts are closer than a threshold in an embedding space, and do not count them in the contrastive loss computation. Note that we still keep each item in the batch for the motion synthesis objective. In Table 4, we perform experiments with a range of different values for this threshold selection. On a scale between [0, 1], a threshold of 0.8 has the best results, balancing keeping a sufficient number of negatives and removing the wrong ones. Without filtering at all, the performance remains at \\(36.02\\mathrm{R}@\\mathfrak{z}3\\) (compared to 41.93). We provide statistics on the percentage of filtered pairs in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.471, + 0.653 + ], + "angle": 0, + "content": "Hyperparameters of the contrastive training. We show the sensitivity of our model to several hyperparameters added when extending TEMOS: (i) temperature \\(\\tau\\) of the cross entropy of InfoNCE [34] in Eq.1, (ii) the \\(\\lambda_{\\mathrm{NCE}}\\) weighting parameter, and (iii) the batch size, which determines the number of negatives. We see in Table 5 that the model is indeed sensitive to the temperature, which is a common observation in other settings. The weight parameter and the batch size are relatively less important while also influencing the results to a certain extent. An experiment with the latent dimensionality hyperparameter can be found in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.254, + 0.681 + ], + "angle": 0, + "content": "4.4. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In Figure 3, we provide sample qualitative results for text-to-motion retrieval on the full test set of H3D. For each query text displayed on the left, top-5 retrieved motions are shown on the right along with their similarity scores. Note that the ground-truth text labels (at the bottom of each motion) for the retrieved motions are not used, and the gallery motions are unseen at training. For the first two examples with 'playing violin' and 'handstand', we retrieve the ground-truth motion at rank 1. We observe that the next ranked motions depict visually similar motions as well (e.g., 'cartwheel' involves standing on the hands). For the free-from prompt example 'Someone is swimming' (i.e., the exact text does not appear in the gallery), the three first motions resemble or involve the swimming action, whereas motions at ranks 4 and 5 are incorrect. We notice that" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.09, + 0.888, + 0.17 + ], + "angle": 0, + "content": "
Temp.Text-motion retrievalMotion-text retrieval
τR@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
0.0019.5221.8127.2312.007.479.7616.5115.50
0.0121.4529.0438.806.0021.0827.1133.619.50
0.124.5830.2441.935.0019.6423.7332.539.50
1.01.081.933.61306.51.811.932.41372.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.689, + 0.17, + 0.704, + 0.18 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.19, + 0.887, + 0.272 + ], + "angle": 0, + "content": "
Weight λNCEText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
0.00118.5523.2536.757.0018.1924.3431.4511.50
0.0120.8426.9937.237.0018.9223.1332.1710.25
0.124.5830.2441.935.0019.6423.7332.539.50
1.019.5224.4634.467.0019.0424.3435.069.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.689, + 0.272, + 0.704, + 0.282 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.293, + 0.887, + 0.374 + ], + "angle": 0, + "content": "
Batch sizeText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
1625.4231.5740.126.0020.3624.1033.738.00
3224.5830.2441.935.0019.6423.7332.539.50
6420.2426.5138.196.0019.5224.2232.059.50
12818.5528.8036.757.0014.9418.4326.1411.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.688, + 0.374, + 0.705, + 0.383 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.383, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Table 5. Hyperparameters of the contrastive training: We measure the sensitivity to the parameters \\(\\tau\\) (temperature), \\(\\lambda_{c}\\) the weight of the contrastive loss, and the batch size. Note that the learning rate is proportionally altered when changing the batch size. We display a wide range of values to show the full trends." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.538 + ], + "angle": 0, + "content": "the incorrect motions have a low similarity \\((< 0.6)\\), and the human bodies are rotated similarly as in swimming. More qualitative results can be seen in Appendix C, as well as our supplementary video on the project page." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.546, + 0.747, + 0.56 + ], + "angle": 0, + "content": "4.5. Use case: Moment retrieval" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.689 + ], + "angle": 0, + "content": "While our focus is retrieval, once our model is trained, it can be used for a different use case. Here, we test the limits of our approach, by qualitatively evaluating the capability of TMR on the task of temporally localizing a natural language query in a long 3D motion sequence. This is similar in spirit to moment retrieval in videos [12, 13, 19, 25, 41]. It is also related to categorical action localization in 3D motions [48]; however, our input is free-form text instead of symbolic action classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In Figure 4, we show four examples, where we apply a model pre-trained on H3D on BABEL sequences. In each example, the queried text is displayed on the top. The x-axis denotes the frame number, the green rectangle represents the ground-truth location for the given action, and the dashed red line marks the localization with the maximum similarity. We simply compute the motion features in a sliding window manner. The similarity between the text label and a 20-frame window centered at each frame is shown in the y-axis as a 1D plot over time. Despite our model not being trained for temporal localization, we observe its grounding potential. Moreover, there exists a domain gap between BABEL label used at test time and H3D used for training. Quantitative evaluation and more qualitative examples are included in Appendix B." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9494" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.119, + 0.201, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.09, + 0.344, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.179, + 0.327, + 0.197 + ], + "angle": 0, + "content": "A person appears to be playing the violin" + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.091, + 0.472, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.179, + 0.454, + 0.197 + ], + "angle": 0, + "content": "Moving the hands and work some thing" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.091, + 0.563, + 0.1 + ], + "angle": 0, + "content": "\\(S = 0.85\\)" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.102, + 0.601, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.179, + 0.591, + 0.197 + ], + "angle": 0, + "content": "The man plays violin holding it in his right hand" + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.091, + 0.692, + 0.1 + ], + "angle": 0, + "content": "\\(S = 0.85\\)" + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.102, + 0.73, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.179, + 0.697, + 0.196 + ], + "angle": 0, + "content": "The man plays the violin" + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.091, + 0.819, + 0.1 + ], + "angle": 0, + "content": "\\(S = 0.81\\)" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.102, + 0.858, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.178, + 0.854, + 0.2 + ], + "angle": 0, + "content": "The man holds something above his left shoulder and rubs it with his right hand" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.24, + 0.201, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.208, + 0.305, + 0.217 + ], + "angle": 0, + "content": "\\(S = 0.82\\)" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.218, + 0.344, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.296, + 0.329, + 0.313 + ], + "angle": 0, + "content": "A man steps forward and does a handstand" + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.208, + 0.433, + 0.217 + ], + "angle": 0, + "content": "\\(S = 0.68\\)" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.218, + 0.472, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.296, + 0.455, + 0.315 + ], + "angle": 0, + "content": "Doing a cartwheel then jumping up and down." + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.208, + 0.562, + 0.217 + ], + "angle": 0, + "content": "\\(S = 0.61\\)" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.218, + 0.601, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.296, + 0.571, + 0.313 + ], + "angle": 0, + "content": "The person does 2 cartwheels" + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.208, + 0.691, + 0.217 + ], + "angle": 0, + "content": "\\(S = 0.48\\)" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.218, + 0.73, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.293, + 0.723, + 0.317 + ], + "angle": 0, + "content": "A person walks forward then turns completely around and does a cartwheel" + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.208, + 0.819, + 0.217 + ], + "angle": 0, + "content": "\\(S = 0.42\\)" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.218, + 0.858, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.296, + 0.842, + 0.313 + ], + "angle": 0, + "content": "The drunk guy struggles to walk down the street" + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.355, + 0.191, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.325, + 0.305, + 0.334 + ], + "angle": 0, + "content": "\\(S = 0.76\\)" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.335, + 0.343, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.413, + 0.324, + 0.431 + ], + "angle": 0, + "content": "The person was flying around like a fly" + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.325, + 0.433, + 0.334 + ], + "angle": 0, + "content": "\\(S = 0.71\\)" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.335, + 0.472, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.413, + 0.441, + 0.43 + ], + "angle": 0, + "content": "A person swam in free style" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.325, + 0.563, + 0.334 + ], + "angle": 0, + "content": "\\(S = 0.62\\)" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.335, + 0.601, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.481, + 0.411, + 0.601, + 0.433 + ], + "angle": 0, + "content": "The person is preforming a swimming stroke know as the butterfly stroke. The arms swing from behind the head and reenter the water propelling the person forward." + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.325, + 0.691, + 0.334 + ], + "angle": 0, + "content": "\\(S = 0.56\\)" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.335, + 0.73, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.413, + 0.696, + 0.431 + ], + "angle": 0, + "content": "A person lays on the ground" + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.325, + 0.819, + 0.334 + ], + "angle": 0, + "content": "\\(S = 0.52\\)" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.335, + 0.858, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.413, + 0.837, + 0.43 + ], + "angle": 0, + "content": "A man crawls forward on his stomach" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.437, + 0.894, + 0.548 + ], + "angle": 0, + "content": "Figure 3. Qualitative retrieval results: We demonstrate example queries on the left, and corresponding retrieved motions on the right, ranked by text-motion similarity. The similarity values are displayed on the top. For each retrieved motion, we also show their accompanying ground-truth text label; note that we do not use these descriptions, but only provide them for analysis purposes. The motions from the gallery are all from the test set (unseen during training). In the first row, all top-5 retrieved motions correspond visually to 'playing violin' and the similarity scores are high \\(>0.80\\). In the second row, we correctly retrieve the 'handstand' motion at top-1, but the other motions mainly perform 'cartwheel' (which involves shortly standing on hands), but with a lower similarity score \\(<0.70\\). For the last example, we query a free-form text 'Someone is swimming', which does not exist in the gallery (but the word 'swim' does). The model successfully finds swimming motions among the top-3, and the other two motions involve the body parallel to the ground." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.568, + 0.262, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.569, + 0.43, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.672, + 0.262, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.673, + 0.43, + 0.774 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.789, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Figure 4. Moment retrieval: We plot the similarity between the temporally annotated BABEL text labels and the motions in a sliding window manner, and obtain a 1D signal over time (blue). We observe that a localization ability emerges from our model, even though it was not trained for temporal localization, and was not with the domain of BABEL labels. The ground-truth temporal span is denoted in green and the maximum similarity is marked with a dashed red line. More examples are provided in Appendix Figure A.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.625, + 0.572 + ], + "angle": 0, + "content": "4.6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.895, + 0.839 + ], + "angle": 0, + "content": "Our model comes with some limitations. Compared to the vast amount of data (e.g., 400M images [44]) in image-text collections used to train competitive foundation models, our motion-text training data can be considered extremely small (e.g., 23K motions in H3D). The generalization performance of motion retrieval models to in-the-wild motions is therefore limited. Data augmentations such as altering text can potentially help to a certain extent (cf. [6]); however, more motion capture is still needed. Another limitation concerns the case where one wishes to replace motion synthesis by retrieving a training motion. In this use case, the model requires the full encoded database (i.e., encoded training set) to be stored in memory, which can be memory-inefficient. Note that the encoded representation is considerably smaller than the original motion data. This could be dealt with using a hierarchical structuring of the motions and language descriptions (e.g. into upper-body motions, lower-body motions, sitting motions, standing motions, etc.)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.613, + 0.861 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In this paper, we focus on the relatively little-studied problem of motion retrieval with natural language queries. We" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9495" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "introduce TMR, a framework to jointly train text-to-motion retrieval and text-to-motion synthesis, with a special attention to the definition of negatives, taking into account the fine-grained nature of motion-language databases. We significantly improve over prior work, and provide a series of experiments highlighting the importance of each component." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.183, + 0.472, + 0.228 + ], + "angle": 0, + "content": "Future work may consider incorporating a language synthesis branch, along with the motion synthesis branch, to build a symmetrical framework, which could bring further benefits." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.236, + 0.47, + 0.304 + ], + "angle": 0, + "content": "Acknowledgements. This work was granted access to the HPC resources of IDRIS under the allocation 2022-AD011012129R2 made by GENCI. GV acknowledges the ANR project CorVis ANR-21-CE23-0003-01. The authors would like to thank Lucas Ventura and Charles Raude." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.306, + 0.469, + 0.333 + ], + "angle": 0, + "content": "MJB Disclosure: files.is.tue.mpg.de/black/CoI_ ICCV_2023.txt" + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.348, + 0.169, + 0.364 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.47, + 0.428 + ], + "angle": 0, + "content": "[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.472, + 0.468 + ], + "angle": 0, + "content": "[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2Pose: Natural language grounded pose forecasting. In International Conference on 3D Vision (3DV), 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.469, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[3] Méziane Aïder, Oussama Gacem, and Mhand Hifi. Branch and solve strategies-based algorithm for the quadratic multiple knapsack problem. Journal of the Operational Research Society, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.523, + 0.469, + 0.563 + ], + "angle": 0, + "content": "[4] Okan Arikan, David A. Forsyth, and James F. O'Brien. Motion synthesis from annotations. ACM Transactions on Graphics (TOG), 2003. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.564, + 0.47, + 0.603 + ], + "angle": 0, + "content": "[5] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal action composition for 3D humans. In International Conference on 3D Vision (3DV), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.604, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[6] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Güll Varol. SINC: Spatial composition of 3D human motions for simultaneous action generation. In International Conference on Computer Vision (ICCV), 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.659, + 0.469, + 0.711 + ], + "angle": 0, + "content": "[7] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In International Conference on Computer Vision (ICCV), 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.712, + 0.47, + 0.752 + ], + "angle": 0, + "content": "[8] Michael Böttner and Simon Clavet. Motion matching - the road to next gen animation. In Nucl.ai, youtu.be/watch?v=z_wpgHFSWss&t=658s, 2015.2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.753, + 0.469, + 0.806 + ], + "angle": 0, + "content": "[9] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. In Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.806, + 0.469, + 0.86 + ], + "angle": 0, + "content": "[10] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. MoFusion: A framework for denoising-diffusion-based motion synthesis. In Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Sivan Doveh, Assaf Arbelle, Sivan Harary, Rameswar Panda, Roei Herzig, Eli Schwartz, Donghyun Kim, Raja Giryes, Rogerio Feris, Shimon Ullman, and Leonid Karlinsky. Teaching struc" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.374, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.893, + 0.12 + ], + "angle": 0, + "content": "tured vision & language concepts to vision & language models. In Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.12, + 0.893, + 0.16 + ], + "angle": 0, + "content": "[12] Victor Escorcia, Mattia Soldan, Josef Sivic, Bernard Ghanem, and Bryan Russell. Temporal localization of moments in video collections with natural language. arXiv:1907.12763, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.161, + 0.893, + 0.2 + ], + "angle": 0, + "content": "[13] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In International Conference on Computer Vision (ICCV), 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.201, + 0.892, + 0.254 + ], + "angle": 0, + "content": "[14] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.255, + 0.892, + 0.308 + ], + "angle": 0, + "content": "[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3D human motions from text. In Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.309, + 0.893, + 0.363 + ], + "angle": 0, + "content": "[16] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2Motion: Conditioned generation of 3D human motions. In ACM International Conference on Multimedia (ACMMM), 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.893, + 0.417 + ], + "angle": 0, + "content": "[17] Ikhsanul Habibie, Mohamed Elgharib, Kripashindu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. In SIGGRAPH, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[18] R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping. In Computer Vision and Pattern Recognition (CVPR), 2006. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.458, + 0.893, + 0.511 + ], + "angle": 0, + "content": "[19] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In International Conference on Computer Vision (ICCV), 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[20] Daniel Holden, Oussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions on Graphics (TOG), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.552, + 0.892, + 0.592 + ], + "angle": 0, + "content": "[21] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.592, + 0.893, + 0.645 + ], + "angle": 0, + "content": "[22] Sai Shashank Kalakonda, Shubh Maheshwari, Sarvadevabhatla, and Ravi Kiran. Action-GPT: Leveraging large-scale language models for improved and generalized zero shot action generation. arXiv:2211.15603, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[23] Lucas Kovar, Michael Gleicher, and Frédéric H. Pighin. Motion graphs. ACM Transactions on Graphics (TOG), 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.673, + 0.893, + 0.712 + ], + "angle": 0, + "content": "[24] Taeryung Lee, Gyeongsik Moon, and Kyoung Mu Lee. MultiAct: Long-term 3D human motion generation from multiple action labels. In AAAI Conference on Artificial Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.712, + 0.893, + 0.753 + ], + "angle": 0, + "content": "[25] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. TVR: A large-scale dataset for video-subtitle moment retrieval. In European Conference on Computer Vision (ECCV), 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.753, + 0.893, + 0.807 + ], + "angle": 0, + "content": "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning (ICML), 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.808, + 0.893, + 0.86 + ], + "angle": 0, + "content": "[27] Angela S Lin, Lemeng Wu, Rodolfo Corona, Kevin Tai, Qixing Huang, and Raymond J Mooney. Generating animated videos of human activities from natural language descriptions. ViGIL NeurIPS Workshop, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.861, + 0.893, + 0.888 + ], + "angle": 0, + "content": "[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.893, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9496" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.106 + ], + "angle": 0, + "content": "linear model. ACM Transactions on Graphics (TOG), 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.106, + 0.469, + 0.146 + ], + "angle": 0, + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.147, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[30] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval and captioning. Neurocomputing, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.201, + 0.469, + 0.255 + ], + "angle": 0, + "content": "[31] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision (ICCV), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.255, + 0.469, + 0.309 + ], + "angle": 0, + "content": "[32] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers. In Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.469, + 0.363 + ], + "angle": 0, + "content": "[33] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[34] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[35] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[36] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[37] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. *Big Data*, 2016. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[38] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english labels. In Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.592, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[40] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning (ICML), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.728, + 0.469, + 0.782 + ], + "angle": 0, + "content": "[41] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. In Transactions of the Association for Computational Linguistics (TACL), 2013. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.782, + 0.469, + 0.836 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.836, + 0.469, + 0.891 + ], + "angle": 0, + "content": "[43] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. In Energy Efficient Machine Learning and Cognitive Computing NeurIPS Workshop, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[44] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: open dataset of clip-filtered 400 million image-text pairs. In Data Centric AI NeurIPS Workshop, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.161, + 0.892, + 0.202 + ], + "angle": 0, + "content": "[45] H. Sidenbladh, M. J. Black, and L. Sigal. Implicit probabilistic models of human motion for synthesis and tracking. In European Conference on Computer Vision (ECCV), 2002. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.202, + 0.892, + 0.255 + ], + "angle": 0, + "content": "[46] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. MPNet: Masked and permuted pre-training for language understanding. In Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.255, + 0.892, + 0.31 + ], + "angle": 0, + "content": "[47] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. MPNet: Masked and permuted pre-training for language understanding. In Neural Information Processing Systems (NeurIPS), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[48] Jiankai Sun, Bolei Zhou, Michael J. Black, and Arjun Chandrasekaran. LocATE: End-to-end localization of actions in 3D with transformers. arXiv:2203.10719, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[49] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In International Conference on Learning Representations (ICLR), 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Neural Information Processing Systems (NeurIPS), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.892, + 0.499 + ], + "angle": 0, + "content": "[51] Ruben Villegas, Jimei Yang, Duygu Ceylan, and Honglak Lee. Neural kinematic networks for unsupervised motion retargeting. In Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.499, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[52] Jiashun Wang, Chao Wen, Yanwei Fu, Haitao Lin, Tianyun Zou, Xiangyang Xue, and Yinda Zhang. Neural pose transfer by spatially adaptive instance normalization. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[53] Sijie Yan, Zhizhong Li, Yuanjun Xiong, Huahan Yan, and Dahua Lin. Convolutional sequence generation for skeleton-based action synthesis. In International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.607, + 0.892, + 0.661 + ], + "angle": 0, + "content": "[54] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. CoCa: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research (TMLR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[55] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In ICLR, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.715, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[56] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2M-GPT: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.769, + 0.892, + 0.823 + ], + "angle": 0, + "content": "[57] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. MotionDiffuse: Text-driven human motion generation with diffusion model. arXiv:2208.15001, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.823, + 0.892, + 0.863 + ], + "angle": 0, + "content": "[58] Yan Zhang, Michael J. Black, and Siyu Tang. Perpetual motion: Generating unbounded human motion. arXiv:2007.13886, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.863, + 0.892, + 0.891 + ], + "angle": 0, + "content": "[59] Rui Zhao, Hui Su, and Qiang Ji. Bayesian adversarial human motion synthesis. In CVPR, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9497" + } + ] +] \ No newline at end of file diff --git a/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/d1f3cbdb-e07a-4952-8052-a9aa7bed81da_origin.pdf b/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/d1f3cbdb-e07a-4952-8052-a9aa7bed81da_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..27a6dd920d7ba26d00ffb28c10553d6fdbc6643f --- /dev/null +++ b/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/d1f3cbdb-e07a-4952-8052-a9aa7bed81da_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4b1992ab54b9327a2fcd9bc5ac79f2144c3f673f28037a8f74a0cec32dc223a +size 1300269 diff --git a/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/full.md b/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d5f43ef7a0eda6808ea5e93635c891f824de2c0b --- /dev/null +++ b/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/full.md @@ -0,0 +1,340 @@ +# TMR: Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis + +Mathis Petrovich $^{1,2}$ Michael J. Black $^{2}$ Gül Varol $^{1}$ $^{1}$ LIGM, École des Ponts, Univ Gustave Eiffel, CNRS, France + $^{2}$ Max Planck Institute for Intelligent Systems, Tübingen, Germany +https://mathis.petrovich.fr/tmr + +# Abstract + +In this paper, we present TMR, a simple yet effective approach for text to 3D human motion retrieval. While previous work has only treated retrieval as a proxy evaluation metric, we tackle it as a standalone task. Our method extends the state-of-the-art text-to-motion synthesis model TEMOS, and incorporates a contrastive loss to better structure the cross-modal latent space. We show that maintaining the motion generation loss, along with the contrastive training, is crucial to obtain good performance. We introduce a benchmark for evaluation and provide an in-depth analysis by reporting results on several protocols. Our extensive experiments on the KIT-ML and HumanML3D datasets show that TMR outperforms the prior work by a significant margin, for example reducing the median rank from 54 to 19. Finally, we showcase the potential of our approach on moment retrieval. Our code and models are publicly available at https://mathis.petrovich.fr/tmr. + +# 1. Introduction + +The language of movement cannot be translated into words. Barbara Mettler + +We ask the question whether a cross-modal space exists between 3D human motions and language. Our goal is to retrieve the most relevant 3D human motion from a gallery, given a natural language query that describes the desired motion (as illustrated in Figure 1). While text-to-image retrieval is a well-established problem within the broader vision & language field [39], there has been less focus on the related task of text-to-motion retrieval. Searching an existing motion capture dataset based on text input can often serve as a viable alternative to text-to-human-motion synthesis in many applications, while also providing the added benefit that the retrieved motion is guaranteed to be realistic. Additionally, once a cross-modal embedding is built to map text and motions into a joint representation space, both of the symmetrical text-to-motion and motion-to-text tasks + +![](images/3bd9579c967fe35c00393d88dd4a95ab700d969e4f4b8da77794fd0740ac96d8.jpg) +A person is stretching legs +Figure 1. Text-to-motion retrieval: We illustrate the task of text-based motion retrieval where the goal is to rank a gallery of motions according to their similarity to the given query in the form of a natural language description. + +can be performed. Such a retrieval-based solution has a range of applications, including automatically indexing large motion capture collections, and helping to initialize the cumbersome text labeling process, by assigning the nearest text to each motion. + +Let us first differentiate text-to-motion retrieval from text-to-motion synthesis. Motion synthesis [5, 9, 36, 49] involves generating new data samples that go beyond the existing training set, while motion retrieval searches through existing motion capture collections. For certain applications, reusing motions from a collection may be sufficient, provided the collection is large enough to contain what the user is searching for. Unlike generative models for motion synthesis, which struggle to produce physically plausible, realistic sequences [5, 35, 36], a retrieval model has the advantage that it always returns a realistic motion. With this motivation, we pose the problem as a nearest neighbor search through a cross-modal text-motion space. + +Early work performs search through motion databases to build motion graphs [4, 23] by finding paths between existing motions and synthesizing new motions by stitching motions together with generated transitions. If the motion database is + +labeled with actions, the user can specify a series of actions to combine [4]. In contrast, our search database is not labeled with text. Motion matching [8], on the other hand, seeks to find the animation that best fits the current motion by searching a database of animations, doing motion-to-motion retrieval [45]. Our framework fundamentally differs from these lines of work in that our task is multi-modal, i.e., user query is text, which is compared against motions. The most similar work to ours is the very recent model from Guo et al. [15], which trains for a joint embedding space between the two modalities. This model is only used to provide a performance measure for motion synthesis tasks, by querying a generated motion within a gallery of 32 descriptions (i.e., motion-to-text retrieval), and counting how many times the correct text is retrieved1. While this can be considered as the first text-motion retrieval model in the literature, its main limitation is the low performance, in particular when the gallery contains fine-grained descriptions. We substantially improve over [15], by incorporating a joint synthesis & retrieval framework, as well as a more powerful contrastive training [34]. + +We get inspiration from image-text models such as BLIP [26] and CoCa [54], which formulate a multi-task objective. Besides the standard dual-encoder matching (such as CLIP [39] with two unimodal encoders for image and text), [26, 54] also employ a text synthesis branch, performing image captioning. Such a generative capability potentially helps the model go beyond 'bag-of-words' understanding of vision-language concepts, observed for the naive contrastive models [11, 55]. In our case, we depart from TEMOS [36] which already has a synthesis branch to generate motions from text. We incorporate a cross-modal contrastive loss (i.e., InfoNCE [34]) in this framework to jointly train text-to-motion synthesis and text-to-motion retrieval tasks. We empirically demonstrate significant improvements with this approach when ablating the importance of each task. + +Text-motion data differs from its text-image counterparts particularly due to the nature of motion descriptions. In fact, for an off-the-shelf large language model, sentences describing different motions tend to be similar, since they fall within the same topic of human motions. For example, the text-text cosine similarities [46] after encoding motion descriptions from the KIT training set [37] are on average 0.71 on a scale between [0, 1], while this value is 0.56 (almost orthogonal) on a random subset of LAION [44] image descriptions with the same size. This poses several challenges. Typical motion datasets [15, 37, 38] contain similar motions with different accompanying texts, e.g., 'person walks', 'human walks', as well as similar texts with different motions, e.g., 'walk backwards', 'walk forwards'. With naive contrastive training [34], one would make all samples within a batch as negatives, except the corresponding label for a given anchor. In this work, we take into account the fact that there are potentially significant similarities between pairs within a batch. Specifically, we discard pairs that have a text-text + +similarity in their labels that is above a certain threshold. Such careful negative sampling leads to performance improvements. + +In this paper, we illustrate an additional use case for our retrieval model – zero-shot temporal localization – and highlight this task as a potential future avenue for research. Similar to temporal localization in videos with natural language queries [12, 13, 19, 25, 41], also referred to as “moment retrieval”, we showcase the grounding capability of our model by directly applying it on long motion sequences to retrieve corresponding moments. We illustrate results on the BABEL dataset [38], which typically contains a series of text annotations for each motion sequence. Note that the task is zero-shot, because the model has not been trained for localization, and at the same time has not seen BABEL labels, which come from a different domain (e.g., typically action-like descriptions instead of full sentences). + +Our contributions are the following: (i) We address the little-studied problem of text-to-motion retrieval, and introduce a series of evaluation benchmarks with varying difficulty. (ii) We propose a joint synthesis and retrieval framework, as well as negative filtering, and obtain state-of-the-art performance on text-motion retrieval. (iii) We provide extensive experiments to analyze the effects of each component in controlled settings. Our code and models are publicly available2. + +# 2. Related work + +We present an overview of the most closely related work on text-to-motion synthesis and retrieval, as well as a brief discussion on cross-modal retrieval works. + +Text and human motion. The research on human motion modeling has recently witnessed an increasing interest in bridging the gap between semantics and 3D human body motions, in particular for text-conditioned motion synthesis. Unlike unconstrained 3D human motion synthesis [53, 58, 59], action-conditioned [16, 35], or text-conditioned [5, 9, 14, 22, 24, 36, 49, 56] models add semantic controls to the generation process. The goal of these works is to generate either deterministically [2, 14, 27] or probabilistically [36], motion sequences that are faithful to the textual description inputs. Note that this is different than gesture synthesis from speech [17], in that the text describes the motion content. However, despite remarkable progress in text-to-image synthesis [40, 42], text-to-motion synthesis remains at a very nascent stage. The realism of the synthesized motions is limited, e.g., foot sliding artifacts [36]. We turn to text-to-motion retrieval as another alternative, and perhaps complementary approach to obtain motions for a given textual description. Our focus is therefore different from the work on synthesis. However, we do make use of a motion synthesis branch to aid the retrieval task. + +Motion retrieval is relatively less explored. As briefly mentioned in Section 1, motion-to-motion retrieval (e.g., motion matching [8, 20, 45]) methods exist. However, the text-to-motion retrieval task is more challenging due to being cross-modal, i.e., nearest neighbor search across text and motion modalities. Within this category, the very recent work of Guo et al. [15] trains a retrieval model purely for evaluation purposes, and applies a margin-based contrastive loss [18], using Euclidean distance between all pairs within a batch. + +Two works are particularly relevant to ours. Firstly, we build on the TEMOS [36] text-to-motion synthesis model, which also has a cross-modal embedding space. However, text and motion embeddings are encouraged to be similar only across positive pairs. We therefore add a contrastive training strategy to incorporate negatives, consequently improving its retrieval capability from a large gallery of fine-grained motions. Secondly, we compare to the aforementioned method of [15], whose motion-to-text retrieval model is adopted for measuring text-to-motion generation performance automatically by other works [10, 49, 57]. + +Cross-modal retrieval. Among widely adopted vision & language retrieval models, some successful examples include CLIP [39], BLIP [26], and CoCa [54] for images, and MILNCE [33], Frozen [7], and CLIP4Clip [30] for videos. They all use variants of cross-modal contrastive learning techniques, such as InfoNCE [34], which we also employ in this work. As discussed in Section 1, we draw inspiration from BLIP [26] and CoCa [54], which add synthesis branches to standard retrieval frameworks. Our approach is similar in spirit to these works in that we perform a cross-modal vision & language retrieval task, but differ in focusing on 3D human motion retrieval, which, to the best of our knowledge, has not been benchmarked. + +# 3. Text-to-motion retrieval + +In this section, we introduce the task and the terminology associated with text-to-motion retrieval (Section 3.1). Next, we present our model, named TMR, and its training protocol (Section 3.2). We then explain our simple approach for identifying and filtering incorrect negatives (Section 3.3), followed by a discussion of the implementation details (Section 3.4). + +# 3.1. Definitions + +Given a natural language query $T$ , such as 'A person walks and then makes a right turn', the goal is to rank the motions from a database (i.e., the gallery) according to their semantic correspondence with the text query, and to retrieve the motion that matches best to the textual description. In other words, the task involves sorting the database so that the top ranked motions are the most relevant matches, i.e., creating a search engine to index motions. Additionally, we define the symmetric (and complementary) task, namely motion-to-text retrieval, where the aim is to retrieve the most suitable caption that matches a given motion from a database of texts. + +3D human motion refers to a sequence of human poses. The task does not impose any limitations on the type of representation used, such as joint positions, rotations, or parametric models such as SMPL [28]. As detailed in Section 4.1, we choose to use the representation employed by Guo et al. [15] to facilitate comparisons with previous work. While our experiments are based on the SMPL skeleton, our method is applicable to any skeleton topology. One could alternatively use retargetting [1, 51, 52] to adapt to other body representations. + +Text description refers to a sequence of words describing the action performed by a human in natural language. We do not restrict the format of the motion description. The text can be simply an action name (e.g., 'walk') or a full sentence (e.g., 'a human is walking'). The sentences can be fine-grained (e.g., 'a human is walking in a circle slowly'), and may contain one or several actions, simultaneously (e.g., 'walking while waving') or sequentially (e.g., 'walking then sitting'). + +# 3.2. Joint training of retrieval and synthesis + +We introduce TMR, which extends the Transformer-based text-to-motion synthesis model TEMOS [36] by incorporating additional losses to make it suitable for the retrieval task. The architecture consists of two independent encoders for inputting motion and text, as well as a decoder that outputs motion (see Figure 2 for an overview). In the following, we review TEMOS [36] components and our added contrastive training. + +Dual encoders. One approach to solving cross-modal retrieval tasks involves defining a similarity function between the two modalities. In our case, the two modalities are text and motion. The similarity function can be applied to compare a given query with each element in the database, and the maximum value would indicate the best match. In this paper, we follow the approach taken by previous work on metric learning, such as CLIP [39], by defining one encoder for each modality and then computing the cosine similarity between their respective embeddings. Such dual embedding has the advantage of fast inference time since the gallery embeddings can be computed and stored beforehand [32]. + +Our model is built upon the components of TEMOS [36], which already provides a motion encoder and a text encoder, mapping them to a joint space (building on the idea from Language2Pose [2]); this serves as a strong baseline for our work. Both motion and text encoders are Transformer encoders [50] with additional learnable distribution parameters, as in the VAE-based ACTOR [35]. They are probabilistic in nature, outputting parameters of a Gaussian distribution $(\mu$ and $\Sigma)$ from which a latent vector $z\in \mathbb{R}^d$ can be sampled. While the text encoder takes text features from a pre-trained and frozen DistilBERT [43] network as input, the motion sequence is fed directly in the motion encoder. Note that when performing retrieval, we directly use the output embedding that corresponds to the mean token $(\mu^{M}$ for motion, $\mu^T$ for text). + +Motion decoder. TEMOS is trained for the task of motion + +![](images/b6aaf7f1f0ffd34b6555d615d240dfd6932730dbd2666a2d89ac1a9c9372187f.jpg) +Figure 2. Joint motion retrieval and synthesis: A simplified view of our TMR framework is presented, where we focus on the similarity matrix defined between text-motion pairs within a batch. Here, we show a batch of 3 samples for illustration purposes. The goal of the contrastive objective is to maximize the diagonal denoting positive pairs (green), and to minimize the off-diagonal negative pair items that have text similarity below a threshold (red). In this example, remaining similarities $S_{23}$ and $S_{32}$ are discarded from the loss computation because there is high text similarity between $T_{2}$ and $T_{3}$ . The rest of the model remains similar to TEMOS [36], which decodes a motion from both text $z^{T_i}$ and motion $z^{M_j}$ latent vectors. See text for further details. + +synthesis and comes equipped with a motion decoder branch. This decoder is identical to the one used in ACTOR [35], which supports variable-duration generation. More specifically, it takes a latent vector $z \in \mathbb{R}^d$ and a sinusoidal positional encoding as input, and generates a motion non-autoregressively through a single forward pass. We show in Section 4.2 that keeping this branch helps improve the results. + +TEMOS losses. We keep the same base set of losses of [36], defined as the weighted sum $\mathcal{L}_{\mathrm{TEMOS}} = \mathcal{L}_{\mathrm{R}} + \lambda_{\mathrm{KL}}\mathcal{L}_{\mathrm{KL}} + \lambda_{\mathrm{E}}\mathcal{L}_{\mathrm{E}}$ . In summary, a reconstruction loss term $\mathcal{L}_{\mathrm{R}}$ measures the motion reconstruction given text or motion input (via a smooth L1 loss). A Kullback-Leibler (KL) divergence loss term $\mathcal{L}_{\mathrm{KL}}$ is composed of four losses: two of them to regularize each encoded distributions $-\mathcal{N}(\mu^M,\Sigma^M)$ for motion and $\mathcal{N}(\mu^T,\Sigma^T)$ for text - to come from a normal distribution $\mathcal{N}(0,I)$ . The other two enforce distribution similarity between the two modalities. A cross-modal embedding similarity loss $\mathcal{L}_{\mathrm{E}}$ enforces both text $z^{T}$ and motion $z^{M}$ latent codes to be similar to each other (with a smooth L1 loss). We set $\lambda_{\mathrm{KL}}$ and $\lambda_{\mathrm{E}}$ to $10^{-5}$ in our experiments as in [36]. + +Contrastive training. While TEMOS has a cross-modal embedding space, its major drawback to be usable as an effective retrieval model is that it is never trained with negatives, but only positive motion-text pairs. To overcome this limitation, we incorporate a contrastive training with the usage of negative samples to better structure the latent space. Given a batch of N (positive) pairs of latent codes $(z_1^T,z_1^M),\dots,(z_N^T,z_N^M)$ , we define any pair $(z_i^T,z_j^M)$ with $i\neq j$ as negative. The similarity matrix $S$ computes the pairwise cosine similarities for all pairs in the batch $S_{ij} = \cos (z_i^T,z_j^M)$ . In contrast to Guo et al. [15], who consider one random negative per batch (and use a margin loss), we adopt the more recent formulation of InfoNCE [34], which was proven effective in other work [7, 26, 39]. This loss term + +can be defined as follows: + +$$ +\mathcal {L} _ {\mathrm {N C E}} = - \frac {1}{2 N} \sum_ {i} \left(\log \frac {\exp S _ {i i} / \tau}{\sum_ {j} \exp S _ {i j} / \tau} + \log \frac {\exp S _ {i i} / \tau}{\sum_ {j} \exp S _ {j i} / \tau}\right), \tag {1} +$$ + +where $\tau$ is the temperature hyperparameter. + +Training loss. The total loss we use to train TMR is the weighted sum $\mathcal{L}_{\mathrm{TEMOS}} + \lambda_{\mathrm{NCE}}\mathcal{L}_{\mathrm{NCE}}$ where $\lambda_{\mathrm{NCE}}$ controls the importance of the contrastive loss. + +# 3.3. Filtering negatives + +As mentioned in Section 1, the descriptions accompanying motion capture collections can be repetitive or similar across the training motions. We wish to prevent defining negatives between text-motion pairs that contain similar descriptions. + +Consider for example the two text descriptions, "A human making a counterclockwise turn" and "A person walks quarter a circle to the left". In the KIT-ML benchmark [37], these two descriptions appear as two different annotations for the same motion. Due to the flexibility and ambiguity of natural language, different words may describe the same concepts (e.g., 'counterclockwise', 'circle to the left'). + +During training, the random selection of batches can adversely affect the results because the model may have to push away two latent vectors that correspond to similar meanings. This can force the network to focus on unimportant details (e.g., 'someone' vs 'human'), ultimately resulting in decreased performance due to unstable behavior and reduced robustness to text variations. + +To alleviate this issue, we leverage an external large language model to provide sentence similarity scores. In particular, we use MPNet [47] to encode sentences and compute similarities between two text descriptions. We then determine whether to filter a pair of text descriptions $(t_{1}, t_{2})$ if their similarity + +is higher than a certain threshold, referring them as 'wrong negatives'. During training, we filter wrong negative pairs from the loss computation. Note that we refrain from defining them as positives, as the language model may also incorrectly say two descriptions are similar when they are not. + +# 3.4. Implementation details + +We use the AdamW optimizer [29] with a learning rate of $10^{-4}$ and a batch size of 32. Since the batch size can be an important hyperparameter for the InfoNCE loss, due to determining the number of negatives, we report experimental results with different values. The latent dimensionality of the embeddings is $d = 256$ . We set the temperature $\tau$ to 0.1, and the weight of the contrastive loss term $\lambda_{\mathrm{NCE}}$ to 0.1. The threshold to filter negatives is set to 0.8. We provide experimental analyses to measure the sensitivity to these added hyperparameters. + +# 4. Experiments + +We start by describing the datasets and evaluation protocol used in the experiments (Section 4.1). We then report the performance of our model on our new retrieval benchmark along with comparison to prior work (Section 4.2). Next, we present our ablation study measuring the effects of the additional contrastive loss, the negative filtering, and the hyperparameters (Section 4.3). Finally, we provide qualitative results for retrieval (Section 4.4), and our use case of moment retrieval (Section 4.5). + +# 4.1. Datasets and evaluation + +HumanML3D dataset (H3D) [15] provides natural language labels to describe the motions in AMASS [31] and HumanAct12 [16] motion capture collections. We follow the motion pre-processing procedure of [15], and apply the SMPL layer [28] to extract joint positions, canonicalize the skeletons to share the same topology (i.e., same bone lengths), then compute motion features (extracting local positions, velocities and foot contacts similar to Holden et al. [21]). The data is then augmented by mirroring left and right (both in motions and their corresponding texts). After this procedure, and following the official split, we obtain 23384, 1460, 4380 motions for the training, validation, and test sets, respectively. On average, each motion is annotated 3.0 times with different text. During training we randomly select one as the matching text, for testing we use the first text. + +KIT Motion-Language dataset (KIT) [37] also come from motion capture data, with an emphasis on locomotion motions. It originally consists of 3911 motion sequences and 6278 text sentences. We pre-process the motions with the same procedure as in H3D. The data is split into 4888, 300, 830 motions for training, validation, and test sets, respectively. In this dataset, each motion is annotated 2.1 times on average. + +Evaluation protocol. We report standard retrieval performance measures: recall at several ranks, R@1, R@2, etc. for both text-to-motion and motion-to-text tasks (identical to the R + +Precision metrics adopted by Guo et al. [15]). Recall at rank $k$ measures the percentage of times the correct label is among the top $k$ results; therefore higher is better. We additionally report median rank (MedR), where lower is better. Note that the retrieval performance is evaluated on a gallery of unseen real motions (i.e., test set). A study of the synthesis performance of TMR can be found in Appendix B. + +We define several evaluation protocols, mainly changing the gallery set. (a) All the test set is used as a first protocol, without any modification. This set is partially problematic because there are repetitive texts across motions, or just minor differences (e.g., person vs human, walk vs walking). (b) All with threshold means we search over all the test set, but, in this case, we accept a retrieved motion as correct if its text label is similar to the query text above a threshold. For example, retrieving the motion corresponding to "A human walks forward" should be correct when the input query is "Someone is walking forward". We set a high threshold of 0.95 (scaled between [0, 1]) to remove very similar texts without removing too many fine-grained details. Appendix A provides statistics on how often similar text descriptions appear in the datasets. (c) Dissimilar subset refers to sampling 100 motion-text pairs whose texts are maximally far from each other (using an approximation of the quadratic knapsack problem [3]). This evaluation can be considered as an easy, but clean, subset of the previous ones. (d) Small batches is included to mimic the protocol described by Guo et al. [15], who randomly pick batches of 32 motion-text pairs and report the average performance. An ideal evaluation metric should not have randomness, and a gallery size of 32 is relatively easy compared to the previous protocols. + +# 4.2. A new benchmark & comparison to prior work + +We present the performance of our model on this new retrieval benchmark, on H3D (Table 1) and KIT (Table 2) datasets, across all evaluation protocols. We also compare against prior work TEMOS [36] and Guo et al. [15]. For TEMOS, we retrain their model on both datasets to have a comparable benchmark since the original model differs in motion representation and lacks left/right data augmentation (and they only provide a KIT-pretrained model, not H3D). For [15], we take their publicly available models trained on these two datasets. + +TEMOS in particular is not designed to perform well on retrieval, since its cross-modal embedding space is only trained with positive pairs. However, Guo et al. train contrastively with negatives as well, using a margin loss [18]. For all 4 evaluation sets with varying difficulties, TMR outperforms the prior work, suggesting our model better captures the finegrained nature of motion descriptions. The model of [15] is adopted as part of motion synthesis evaluation in several works. TMR may therefore provide a better alternative. Our significant improvements over the state of the art can be credited to (i) jointly training for synthesis and retrieval, (ii) adopting the more recent contrastive objective, InfoNCE [34], while (iii) carefully + +
ProtocolMethodsText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓
(a) AllTEMOS [36]2.124.095.878.2613.52173.03.864.546.949.3814.00183.25
Guo et al. [15]1.803.424.797.1212.4781.002.923.746.008.3612.9581.50
TMR5.6810.5914.0420.3430.9428.009.9512.4417.9523.5632.6928.50
(b) All with thresholdTEMOS [36]5.218.2211.1415.0922.1279.005.486.199.0012.0117.10129.0
Guo et al. [15]5.307.8310.7514.5922.5154.004.955.688.9311.6416.9469.50
TMR11.6015.3920.5027.7238.5219.0013.2015.7322.0327.6537.6321.50
(c) Dissimilar subsetTEMOS [36]33.0042.0049.0057.0066.004.0035.0044.0050.0056.0070.003.50
Guo et al. [15]34.0048.0057.0072.0084.003.0034.0047.0059.0072.0083.003.00
TMR47.0061.0071.0080.0086.002.0048.0063.0069.0080.0084.002.00
(d) Small batches [15]TEMOS [36]40.4953.5261.1470.9684.152.3339.9653.4961.7972.4085.892.33
Guo et al. [15]52.4871.0580.6589.6696.581.3952.0071.2181.1189.8796.781.38
TMR67.1681.3286.8191.4395.361.0467.9781.2086.3591.7095.271.03
+ +Table 1. Text-to-motion retrieval benchmark on HumanML3D: We establish four evaluation protocols as described in Section 4.1, with decreasing difficulty from (a) to (d). Our model TMR substantially outperforms the prior work of Guo et al. [15] and TEMOS [36], on the challenging H3D dataset. + +
ProtocolMethodsText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓
(a) AllTEMOS [36]7.1113.2517.5924.1035.6624.0011.6915.3020.1226.6336.3926.50
Guo et al. [15]3.376.9910.8416.8727.7128.004.946.5110.7216.1425.3028.50
TMR7.2313.9820.3628.3140.1217.0011.2013.8620.1228.0738.5518.00
(b) All with thresholdTEMOS [36]18.5524.3430.8442.2956.397.0017.7122.4128.8035.4247.1113.25
Guo et al. [15]13.2522.6529.7639.0449.5211.0010.4813.9820.4827.9538.5517.25
TMR24.5830.2441.9350.4860.365.0019.6423.7332.5341.2053.019.50
(c) Dissimilar subsetTEMOS [36]24.0040.0046.0054.0070.005.0033.0039.0045.0049.0064.006.50
Guo et al. [15]16.0029.0036.0048.0066.006.0024.0029.0036.0046.0066.007.00
TMR26.0046.0060.0070.0083.003.0034.0045.0060.0069.0082.003.50
(d) Small batches [15]TEMOS [36]43.8858.2567.0074.0084.752.0641.8855.8865.6275.2585.752.25
Guo et al. [15]42.2562.6275.1287.5096.121.8839.7562.7573.6286.8895.881.95
TMR49.2569.7578.2587.8895.001.5050.1267.1276.8888.8894.751.53
+ +Table 2. Text-to-motion retrieval benchmark on KIT-ML: As in Table 1, we report the four evaluation protocols, this time on the KIT dataset. Again, TMR significantly improves over Guo et al. [15] and TEMOS [36] across all protocols and metrics. + +
Motion +Recons.InfoNCEMarginText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
XX✓15.0622.1725.7812.008.1911.5716.3919.50
X✓X19.7625.3036.876.0017.4719.7630.609.50
✓XX18.5524.3430.847.0017.7122.4128.8013.25
✓X✓19.8824.4634.467.0014.7019.7628.1912.50
✓✓X24.5830.2441.935.0019.6423.7332.539.50
+ +Table 3. Losses: We experiment with various loss definitions (i) with/without the motion reconstruction, and (ii) the choice of the contrastive loss between InfoNCE and margin-based. We see that InfoNCE [34] is a better alternative to the contrastive loss with Euclidean margin [18] (employed by Guo et al. [15]). The reconstruction loss through the motion decoder branch further boosts the results. + +eliminating wrong negatives. In the following, we ablate these components in controlled experiments. + +# 4.3. Ablation study + +The rest of the quantitative evaluation uses the ' (b) All with threshold' evaluation protocol, on the KIT dataset. + +Which losses matter? Table 3 compares several variants of TMR where we check (a) whether the jointly trained motion synthesis branch helps retrieval, and (b) how important the form of the contrastive loss is. When removing the synthesis branch and only using the contrastive loss, we perform a deterministic encoding (i.e., with a single token instead of two tokens $\mu$ , $\sigma$ ). First, we see that the motion synthesis branch certainly improves results compared with only training using a contrastive loss (e.g., 41.93 vs 36.87 R@3). This possibly forces the latent vector to capture the full content of the input text (i.e., instead of picking up on a subset of words, or bag-of-words [11, 55] upon finding a shortcut that satisfies the contrastive loss). Second, in the presence of a contrastive loss, the InfoNCE formulation is significantly better than the margin loss employed by previous work [15] (41.93 vs 34.46 R@1). Note that for this experiment, we keep the same negative filtering for both the margin loss and InfoNCE (we provide additional experiments without the negative filtering in Appendix B). + +
ThresholdText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
0.5519.4023.2530.489.0017.8321.6929.5214.00
0.6017.9526.8736.876.0020.6024.7031.8111.25
0.6523.0128.6736.397.0019.0421.6929.7611.50
0.7022.2929.6438.806.0018.1922.7732.059.00
0.7520.0027.1137.836.0020.2424.4634.229.50
0.8024.5830.2441.935.0019.6423.7332.539.50
0.8521.4525.7838.436.5020.8424.1033.379.50
0.9023.2530.1240.486.0020.0025.1833.139.50
0.9520.4826.9938.436.0019.2823.3731.9310.25
X22.1727.8336.027.0016.7521.3332.1711.50
+ +Table 4. Filtering negatives: We compare several threshold values for filtering negatives from the loss comparison due to having similar texts. We observe that removing negatives based on text similarity above 0.8 (from a scale between [0,1]) performs well overall. + +The effect of filtering negatives. As explained in Section 3.3, during training we filter out pairs whose texts are closer than a threshold in an embedding space, and do not count them in the contrastive loss computation. Note that we still keep each item in the batch for the motion synthesis objective. In Table 4, we perform experiments with a range of different values for this threshold selection. On a scale between [0, 1], a threshold of 0.8 has the best results, balancing keeping a sufficient number of negatives and removing the wrong ones. Without filtering at all, the performance remains at $36.02\mathrm{R}@\mathfrak{z}3$ (compared to 41.93). We provide statistics on the percentage of filtered pairs in Appendix A. + +Hyperparameters of the contrastive training. We show the sensitivity of our model to several hyperparameters added when extending TEMOS: (i) temperature $\tau$ of the cross entropy of InfoNCE [34] in Eq.1, (ii) the $\lambda_{\mathrm{NCE}}$ weighting parameter, and (iii) the batch size, which determines the number of negatives. We see in Table 5 that the model is indeed sensitive to the temperature, which is a common observation in other settings. The weight parameter and the batch size are relatively less important while also influencing the results to a certain extent. An experiment with the latent dimensionality hyperparameter can be found in Appendix B. + +# 4.4. Qualitative results + +In Figure 3, we provide sample qualitative results for text-to-motion retrieval on the full test set of H3D. For each query text displayed on the left, top-5 retrieved motions are shown on the right along with their similarity scores. Note that the ground-truth text labels (at the bottom of each motion) for the retrieved motions are not used, and the gallery motions are unseen at training. For the first two examples with 'playing violin' and 'handstand', we retrieve the ground-truth motion at rank 1. We observe that the next ranked motions depict visually similar motions as well (e.g., 'cartwheel' involves standing on the hands). For the free-from prompt example 'Someone is swimming' (i.e., the exact text does not appear in the gallery), the three first motions resemble or involve the swimming action, whereas motions at ranks 4 and 5 are incorrect. We notice that + +
Temp.Text-motion retrievalMotion-text retrieval
τR@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
0.0019.5221.8127.2312.007.479.7616.5115.50
0.0121.4529.0438.806.0021.0827.1133.619.50
0.124.5830.2441.935.0019.6423.7332.539.50
1.01.081.933.61306.51.811.932.41372.0
+ +(a) + +
Weight λNCEText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
0.00118.5523.2536.757.0018.1924.3431.4511.50
0.0120.8426.9937.237.0018.9223.1332.1710.25
0.124.5830.2441.935.0019.6423.7332.539.50
1.019.5224.4634.467.0019.0424.3435.069.50
+ +(b) + +
Batch sizeText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
1625.4231.5740.126.0020.3624.1033.738.00
3224.5830.2441.935.0019.6423.7332.539.50
6420.2426.5138.196.0019.5224.2232.059.50
12818.5528.8036.757.0014.9418.4326.1411.50
+ +Table 5. Hyperparameters of the contrastive training: We measure the sensitivity to the parameters $\tau$ (temperature), $\lambda_{c}$ the weight of the contrastive loss, and the batch size. Note that the learning rate is proportionally altered when changing the batch size. We display a wide range of values to show the full trends. + +(c) + +the incorrect motions have a low similarity $(< 0.6)$ , and the human bodies are rotated similarly as in swimming. More qualitative results can be seen in Appendix C, as well as our supplementary video on the project page. + +# 4.5. Use case: Moment retrieval + +While our focus is retrieval, once our model is trained, it can be used for a different use case. Here, we test the limits of our approach, by qualitatively evaluating the capability of TMR on the task of temporally localizing a natural language query in a long 3D motion sequence. This is similar in spirit to moment retrieval in videos [12, 13, 19, 25, 41]. It is also related to categorical action localization in 3D motions [48]; however, our input is free-form text instead of symbolic action classes. + +In Figure 4, we show four examples, where we apply a model pre-trained on H3D on BABEL sequences. In each example, the queried text is displayed on the top. The x-axis denotes the frame number, the green rectangle represents the ground-truth location for the given action, and the dashed red line marks the localization with the maximum similarity. We simply compute the motion features in a sliding window manner. The similarity between the text label and a 20-frame window centered at each frame is shown in the y-axis as a 1D plot over time. Despite our model not being trained for temporal localization, we observe its grounding potential. Moreover, there exists a domain gap between BABEL label used at test time and H3D used for training. Quantitative evaluation and more qualitative examples are included in Appendix B. + +![](images/064821e401a50ca69dde23dd5a446a4b51b8e4efa18757d1fd7b136ea203c749.jpg) + +![](images/a2bf02720a19f22ce9b82b0dcad9d9d8cb9e958c2e4657c6dcb42493d3a7da83.jpg) +A person appears to be playing the violin + +![](images/95de5ac708fe12c406124aafeb2cd85604bbadcfa942bb983f28925b75979736.jpg) +Moving the hands and work some thing + +![](images/11dbe0118342598782f59ddb39f0816ae19701acad8cc74874486f3c7740b0b8.jpg) +$S = 0.85$ +The man plays violin holding it in his right hand + +![](images/a8ef19f55934eeb1450b68677175d78973c978e077db40c5ca1af7ebe474517d.jpg) +$S = 0.85$ +The man plays the violin + +![](images/84b3b304087e09182b8211dd9d308f76f0dea1531192733122e6995ce72207e0.jpg) +$S = 0.81$ +The man holds something above his left shoulder and rubs it with his right hand + +![](images/8a87e77987a443f81c2ded9248616d3882553e376b11abe60c593628511da90c.jpg) + +![](images/a53020f79ba182d02c8ffa57d9e80687129bb6d6668dc5e2972ddce4c70f5cc0.jpg) +$S = 0.82$ +A man steps forward and does a handstand + +![](images/a899cfdc8d1e86545b6465259256ace6974da5bad8b9e8ae7e4f53821a02ccd6.jpg) +$S = 0.68$ +Doing a cartwheel then jumping up and down. + +![](images/ce655c700b0def4dd5399a9fa0672e60d0d1c5efeb27a382a7b8505d91e6e30f.jpg) +$S = 0.61$ +The person does 2 cartwheels + +![](images/00028c8c67deb4eced219a9704c277e46e53b827736780af3a969053dc255717.jpg) +$S = 0.48$ +A person walks forward then turns completely around and does a cartwheel + +![](images/25666dff1d750c88f5195a0fb8e01f23c420c365c65643b87981d1d82f9f87b2.jpg) +$S = 0.42$ +The drunk guy struggles to walk down the street + +![](images/ecc34bc462a299d4d3b13068a86d8c9240ca88185bfd2c5824cea64fcdda8c30.jpg) + +![](images/563bf8f3a65becc7691d235a13bdfbe8dd2969478c21ba0e06dc8ab37377b706.jpg) +$S = 0.76$ +The person was flying around like a fly + +![](images/948efe27aa5a168d1283927eba6a8a52508679771405fa71ef750c355641d53c.jpg) +$S = 0.71$ +A person swam in free style + +![](images/58a9e17f9f91e84c21fd1e6a282c0e564779e30e97fe9d0ce34c0ea1734ccda7.jpg) +$S = 0.62$ +The person is preforming a swimming stroke know as the butterfly stroke. The arms swing from behind the head and reenter the water propelling the person forward. + +![](images/027e487ee3bed9a7f267bf14b86ff3fae0e4ac966fa8160fcda2f450ad6b7e5d.jpg) +$S = 0.56$ +A person lays on the ground + +![](images/ffc9d2090b3d60d9023d6c26b88509a692df91c0a6d8c02e2567635b49f407fc.jpg) +$S = 0.52$ +A man crawls forward on his stomach + +![](images/335b51424d89d8183ded2de2a693f99dd72e608153b97815ab28a6d08c710c08.jpg) +Figure 3. Qualitative retrieval results: We demonstrate example queries on the left, and corresponding retrieved motions on the right, ranked by text-motion similarity. The similarity values are displayed on the top. For each retrieved motion, we also show their accompanying ground-truth text label; note that we do not use these descriptions, but only provide them for analysis purposes. The motions from the gallery are all from the test set (unseen during training). In the first row, all top-5 retrieved motions correspond visually to 'playing violin' and the similarity scores are high $>0.80$ . In the second row, we correctly retrieve the 'handstand' motion at top-1, but the other motions mainly perform 'cartwheel' (which involves shortly standing on hands), but with a lower similarity score $<0.70$ . For the last example, we query a free-form text 'Someone is swimming', which does not exist in the gallery (but the word 'swim' does). The model successfully finds swimming motions among the top-3, and the other two motions involve the body parallel to the ground. + +![](images/ebf61b965f3627cbeeb951af27f629b49c414d8a813a4182e8c7875cfb206ede.jpg) + +![](images/8c0c7369183f4bd44168fe9db4c9952cce1984df268ab0362f3e12a0d8068122.jpg) +Figure 4. Moment retrieval: We plot the similarity between the temporally annotated BABEL text labels and the motions in a sliding window manner, and obtain a 1D signal over time (blue). We observe that a localization ability emerges from our model, even though it was not trained for temporal localization, and was not with the domain of BABEL labels. The ground-truth temporal span is denoted in green and the maximum similarity is marked with a dashed red line. More examples are provided in Appendix Figure A.2. + +![](images/07058cae226e7cb70e2c403d32635699fdce8eba1c108a2f931801816fea53df.jpg) + +# 4.6. Limitations + +Our model comes with some limitations. Compared to the vast amount of data (e.g., 400M images [44]) in image-text collections used to train competitive foundation models, our motion-text training data can be considered extremely small (e.g., 23K motions in H3D). The generalization performance of motion retrieval models to in-the-wild motions is therefore limited. Data augmentations such as altering text can potentially help to a certain extent (cf. [6]); however, more motion capture is still needed. Another limitation concerns the case where one wishes to replace motion synthesis by retrieving a training motion. In this use case, the model requires the full encoded database (i.e., encoded training set) to be stored in memory, which can be memory-inefficient. Note that the encoded representation is considerably smaller than the original motion data. This could be dealt with using a hierarchical structuring of the motions and language descriptions (e.g. into upper-body motions, lower-body motions, sitting motions, standing motions, etc.). + +# 5. Conclusion + +In this paper, we focus on the relatively little-studied problem of motion retrieval with natural language queries. We + +introduce TMR, a framework to jointly train text-to-motion retrieval and text-to-motion synthesis, with a special attention to the definition of negatives, taking into account the fine-grained nature of motion-language databases. We significantly improve over prior work, and provide a series of experiments highlighting the importance of each component. + +Future work may consider incorporating a language synthesis branch, along with the motion synthesis branch, to build a symmetrical framework, which could bring further benefits. + +Acknowledgements. This work was granted access to the HPC resources of IDRIS under the allocation 2022-AD011012129R2 made by GENCI. GV acknowledges the ANR project CorVis ANR-21-CE23-0003-01. The authors would like to thank Lucas Ventura and Charles Raude. + +MJB Disclosure: files.is.tue.mpg.de/black/CoI_ ICCV_2023.txt + +# References + +[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 2020. 3 +[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2Pose: Natural language grounded pose forecasting. In International Conference on 3D Vision (3DV), 2019. 2, 3 +[3] Méziane Aïder, Oussama Gacem, and Mhand Hifi. Branch and solve strategies-based algorithm for the quadratic multiple knapsack problem. Journal of the Operational Research Society, 2022. 5 +[4] Okan Arikan, David A. Forsyth, and James F. O'Brien. Motion synthesis from annotations. ACM Transactions on Graphics (TOG), 2003. 1, 2 +[5] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal action composition for 3D humans. In International Conference on 3D Vision (3DV), 2022. 1, 2 +[6] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Güll Varol. SINC: Spatial composition of 3D human motions for simultaneous action generation. In International Conference on Computer Vision (ICCV), 2023. 8 +[7] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In International Conference on Computer Vision (ICCV), 2021. 3, 4 +[8] Michael Böttner and Simon Clavet. Motion matching - the road to next gen animation. In Nucl.ai, youtu.be/watch?v=z_wpgHFSWss&t=658s, 2015.2, 3 +[9] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. In Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2 +[10] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. MoFusion: A framework for denoising-diffusion-based motion synthesis. In Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[11] Sivan Doveh, Assaf Arbelle, Sivan Harary, Rameswar Panda, Roei Herzig, Eli Schwartz, Donghyun Kim, Raja Giryes, Rogerio Feris, Shimon Ullman, and Leonid Karlinsky. Teaching struc + +tured vision & language concepts to vision & language models. In Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6 +[12] Victor Escorcia, Mattia Soldan, Josef Sivic, Bernard Ghanem, and Bryan Russell. Temporal localization of moments in video collections with natural language. arXiv:1907.12763, 2019. 2, 7 +[13] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In International Conference on Computer Vision (ICCV), 2017. 2, 7 +[14] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In International Conference on Computer Vision (ICCV), 2021. 2 +[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3D human motions from text. In Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6 +[16] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2Motion: Conditioned generation of 3D human motions. In ACM International Conference on Multimedia (ACMMM), 2020. 2, 5 +[17] Ikhsanul Habibie, Mohamed Elgharib, Kripashindu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. In SIGGRAPH, 2022. 2 +[18] R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping. In Computer Vision and Pattern Recognition (CVPR), 2006. 3, 5, 6 +[19] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In International Conference on Computer Vision (ICCV), 2017. 2, 7 +[20] Daniel Holden, Oussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions on Graphics (TOG), 2020. 3 +[21] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 2017. 5 +[22] Sai Shashank Kalakonda, Shubh Maheshwari, Sarvadevabhatla, and Ravi Kiran. Action-GPT: Leveraging large-scale language models for improved and generalized zero shot action generation. arXiv:2211.15603, 2022. 2 +[23] Lucas Kovar, Michael Gleicher, and Frédéric H. Pighin. Motion graphs. ACM Transactions on Graphics (TOG), 2002. 1 +[24] Taeryung Lee, Gyeongsik Moon, and Kyoung Mu Lee. MultiAct: Long-term 3D human motion generation from multiple action labels. In AAAI Conference on Artificial Intelligence, 2023. 2 +[25] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. TVR: A large-scale dataset for video-subtitle moment retrieval. In European Conference on Computer Vision (ECCV), 2020. 2, 7 +[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning (ICML), 2022. 2, 3, 4 +[27] Angela S Lin, Lemeng Wu, Rodolfo Corona, Kevin Tai, Qixing Huang, and Raymond J Mooney. Generating animated videos of human activities from natural language descriptions. ViGIL NeurIPS Workshop, 2018. 2 +[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person + +linear model. ACM Transactions on Graphics (TOG), 2015. 3, 5 +[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5 +[30] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval and captioning. Neurocomputing, 2022. 3 +[31] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision (ICCV), 2019. 5 +[32] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers. In Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[33] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[34] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018. 2, 3, 4, 5, 6, 7 +[35] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 1, 2, 3, 4 +[36] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 5, 6 +[37] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. *Big Data*, 2016. 2, 4, 5 +[38] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english labels. In Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2, 3, 4 +[40] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning (ICML), 2021. 2 +[41] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. In Transactions of the Association for Computational Linguistics (TACL), 2013. 2, 7 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[43] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. In Energy Efficient Machine Learning and Cognitive Computing NeurIPS Workshop, 2019. 3 + +[44] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: open dataset of clip-filtered 400 million image-text pairs. In Data Centric AI NeurIPS Workshop, 2021. 2, 8 +[45] H. Sidenbladh, M. J. Black, and L. Sigal. Implicit probabilistic models of human motion for synthesis and tracking. In European Conference on Computer Vision (ECCV), 2002. 2, 3 +[46] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. MPNet: Masked and permuted pre-training for language understanding. In Neural Information Processing Systems (NeurIPS), 2020. 2 +[47] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. MPNet: Masked and permuted pre-training for language understanding. In Neural Information Processing Systems (NeurIPS), 2020. 4 +[48] Jiankai Sun, Bolei Zhou, Michael J. Black, and Arjun Chandrasekaran. LocATE: End-to-end localization of actions in 3D with transformers. arXiv:2203.10719, 2022. 7 +[49] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In International Conference on Learning Representations (ICLR), 2023. 1, 2, 3 +[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Neural Information Processing Systems (NeurIPS), 2017. 3 +[51] Ruben Villegas, Jimei Yang, Duygu Ceylan, and Honglak Lee. Neural kinematic networks for unsupervised motion retargeting. In Computer Vision and Pattern Recognition (CVPR), 2018. 3 +[52] Jiashun Wang, Chao Wen, Yanwei Fu, Haitao Lin, Tianyun Zou, Xiangyang Xue, and Yinda Zhang. Neural pose transfer by spatially adaptive instance normalization. In Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[53] Sijie Yan, Zhizhong Li, Yuanjun Xiong, Huahan Yan, and Dahua Lin. Convolutional sequence generation for skeleton-based action synthesis. In International Conference on Computer Vision (ICCV), 2019. 2 +[54] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. CoCa: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research (TMLR), 2022. 2, 3 +[55] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In ICLR, 2022. 2, 6 +[56] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2M-GPT: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2 +[57] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. MotionDiffuse: Text-driven human motion generation with diffusion model. arXiv:2208.15001, 2022. 3 +[58] Yan Zhang, Michael J. Black, and Siyu Tang. Perpetual motion: Generating unbounded human motion. arXiv:2007.13886, 2020. 2 +[59] Rui Zhao, Hui Su, and Qiang Ji. Bayesian adversarial human motion synthesis. In CVPR, 2020. 2 \ No newline at end of file diff --git a/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/images.zip b/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9f125ba0dbf5f78cb4ebd6f014ed552db08cb24c --- /dev/null +++ b/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79ea25d3059ccba578a864cc80e5c7d4d90659322c0c25f1ed86f5659ef7d8aa +size 584549 diff --git a/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/layout.json b/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dc4e03fcd874329447f805617dc70394abf0e97a --- /dev/null +++ b/2023/TMR_ Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis/layout.json @@ -0,0 +1,9133 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 527, + 121 + ], + "type": "text", + "content": "TMR: Text-to-Motion Retrieval Using Contrastive 3D Human Motion Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 142, + 451, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 142, + 451, + 201 + ], + "spans": [ + { + "bbox": [ + 141, + 142, + 451, + 201 + ], + "type": "text", + "content": "Mathis Petrovich" + }, + { + "bbox": [ + 141, + 142, + 451, + 201 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 141, + 142, + 451, + 201 + ], + "type": "text", + "content": " Michael J. Black" + }, + { + "bbox": [ + 141, + 142, + 451, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 142, + 451, + 201 + ], + "type": "text", + "content": " Gül Varol" + }, + { + "bbox": [ + 141, + 142, + 451, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 142, + 451, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 142, + 451, + 201 + ], + "type": "text", + "content": " LIGM, École des Ponts, Univ Gustave Eiffel, CNRS, France \n" + }, + { + "bbox": [ + 141, + 142, + 451, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 142, + 451, + 201 + ], + "type": "text", + "content": " Max Planck Institute for Intelligent Systems, Tübingen, Germany \nhttps://mathis.petrovich.fr/tmr" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 144, + 228, + 190, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 228, + 190, + 240 + ], + "spans": [ + { + "bbox": [ + 144, + 228, + 190, + 240 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 255, + 290, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 290, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 290, + 447 + ], + "type": "text", + "content": "In this paper, we present TMR, a simple yet effective approach for text to 3D human motion retrieval. While previous work has only treated retrieval as a proxy evaluation metric, we tackle it as a standalone task. Our method extends the state-of-the-art text-to-motion synthesis model TEMOS, and incorporates a contrastive loss to better structure the cross-modal latent space. We show that maintaining the motion generation loss, along with the contrastive training, is crucial to obtain good performance. We introduce a benchmark for evaluation and provide an in-depth analysis by reporting results on several protocols. Our extensive experiments on the KIT-ML and HumanML3D datasets show that TMR outperforms the prior work by a significant margin, for example reducing the median rank from 54 to 19. Finally, we showcase the potential of our approach on moment retrieval. Our code and models are publicly available at https://mathis.petrovich.fr/tmr." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 475, + 124, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 124, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 124, + 487 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 497, + 278, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 497, + 278, + 518 + ], + "spans": [ + { + "bbox": [ + 66, + 497, + 278, + 518 + ], + "type": "text", + "content": "The language of movement cannot be translated into words. Barbara Mettler" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "We ask the question whether a cross-modal space exists between 3D human motions and language. Our goal is to retrieve the most relevant 3D human motion from a gallery, given a natural language query that describes the desired motion (as illustrated in Figure 1). While text-to-image retrieval is a well-established problem within the broader vision & language field [39], there has been less focus on the related task of text-to-motion retrieval. Searching an existing motion capture dataset based on text input can often serve as a viable alternative to text-to-human-motion synthesis in many applications, while also providing the added benefit that the retrieved motion is guaranteed to be realistic. Additionally, once a cross-modal embedding is built to map text and motions into a joint representation space, both of the symmetrical text-to-motion and motion-to-text tasks" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 311, + 255, + 542, + 403 + ], + "blocks": [ + { + "bbox": [ + 365, + 236, + 471, + 248 + ], + "lines": [ + { + "bbox": [ + 365, + 236, + 471, + 248 + ], + "spans": [ + { + "bbox": [ + 365, + 236, + 471, + 248 + ], + "type": "text", + "content": "A person is stretching legs" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 255, + 542, + 403 + ], + "lines": [ + { + "bbox": [ + 311, + 255, + 542, + 403 + ], + "spans": [ + { + "bbox": [ + 311, + 255, + 542, + 403 + ], + "type": "image", + "image_path": "3bd9579c967fe35c00393d88dd4a95ab700d969e4f4b8da77794fd0740ac96d8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 407, + 546, + 452 + ], + "lines": [ + { + "bbox": [ + 305, + 407, + 546, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 546, + 452 + ], + "type": "text", + "content": "Figure 1. Text-to-motion retrieval: We illustrate the task of text-based motion retrieval where the goal is to rank a gallery of motions according to their similarity to the given query in the form of a natural language description." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 468, + 547, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 547, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 547, + 517 + ], + "type": "text", + "content": "can be performed. Such a retrieval-based solution has a range of applications, including automatically indexing large motion capture collections, and helping to initialize the cumbersome text labeling process, by assigning the nearest text to each motion." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 519, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 547, + 663 + ], + "type": "text", + "content": "Let us first differentiate text-to-motion retrieval from text-to-motion synthesis. Motion synthesis [5, 9, 36, 49] involves generating new data samples that go beyond the existing training set, while motion retrieval searches through existing motion capture collections. For certain applications, reusing motions from a collection may be sufficient, provided the collection is large enough to contain what the user is searching for. Unlike generative models for motion synthesis, which struggle to produce physically plausible, realistic sequences [5, 35, 36], a retrieval model has the advantage that it always returns a realistic motion. With this motivation, we pose the problem as a nearest neighbor search through a cross-modal text-motion space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Early work performs search through motion databases to build motion graphs [4, 23] by finding paths between existing motions and synthesizing new motions by stitching motions together with generated transitions. If the motion database is" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9488" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "content": "labeled with actions, the user can specify a series of actions to combine [4]. In contrast, our search database is not labeled with text. Motion matching [8], on the other hand, seeks to find the animation that best fits the current motion by searching a database of animations, doing motion-to-motion retrieval [45]. Our framework fundamentally differs from these lines of work in that our task is multi-modal, i.e., user query is text, which is compared against motions. The most similar work to ours is the very recent model from Guo et al. [15], which trains for a joint embedding space between the two modalities. This model is only used to provide a performance measure for motion synthesis tasks, by querying a generated motion within a gallery of 32 descriptions (i.e., motion-to-text retrieval), and counting how many times the correct text is retrieved1. While this can be considered as the first text-motion retrieval model in the literature, its main limitation is the low performance, in particular when the gallery contains fine-grained descriptions. We substantially improve over [15], by incorporating a joint synthesis & retrieval framework, as well as a more powerful contrastive training [34]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 467 + ], + "type": "text", + "content": "We get inspiration from image-text models such as BLIP [26] and CoCa [54], which formulate a multi-task objective. Besides the standard dual-encoder matching (such as CLIP [39] with two unimodal encoders for image and text), [26, 54] also employ a text synthesis branch, performing image captioning. Such a generative capability potentially helps the model go beyond 'bag-of-words' understanding of vision-language concepts, observed for the naive contrastive models [11, 55]. In our case, we depart from TEMOS [36] which already has a synthesis branch to generate motions from text. We incorporate a cross-modal contrastive loss (i.e., InfoNCE [34]) in this framework to jointly train text-to-motion synthesis and text-to-motion retrieval tasks. We empirically demonstrate significant improvements with this approach when ablating the importance of each task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 468, + 289, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 289, + 685 + ], + "type": "text", + "content": "Text-motion data differs from its text-image counterparts particularly due to the nature of motion descriptions. In fact, for an off-the-shelf large language model, sentences describing different motions tend to be similar, since they fall within the same topic of human motions. For example, the text-text cosine similarities [46] after encoding motion descriptions from the KIT training set [37] are on average 0.71 on a scale between [0, 1], while this value is 0.56 (almost orthogonal) on a random subset of LAION [44] image descriptions with the same size. This poses several challenges. Typical motion datasets [15, 37, 38] contain similar motions with different accompanying texts, e.g., 'person walks', 'human walks', as well as similar texts with different motions, e.g., 'walk backwards', 'walk forwards'. With naive contrastive training [34], one would make all samples within a batch as negatives, except the corresponding label for a given anchor. In this work, we take into account the fact that there are potentially significant similarities between pairs within a batch. Specifically, we discard pairs that have a text-text" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "similarity in their labels that is above a certain threshold. Such careful negative sampling leads to performance improvements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 97, + 547, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 547, + 264 + ], + "type": "text", + "content": "In this paper, we illustrate an additional use case for our retrieval model – zero-shot temporal localization – and highlight this task as a potential future avenue for research. Similar to temporal localization in videos with natural language queries [12, 13, 19, 25, 41], also referred to as “moment retrieval”, we showcase the grounding capability of our model by directly applying it on long motion sequences to retrieve corresponding moments. We illustrate results on the BABEL dataset [38], which typically contains a series of text annotations for each motion sequence. Note that the task is zero-shot, because the model has not been trained for localization, and at the same time has not seen BABEL labels, which come from a different domain (e.g., typically action-like descriptions instead of full sentences)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 266, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 547, + 361 + ], + "type": "text", + "content": "Our contributions are the following: (i) We address the little-studied problem of text-to-motion retrieval, and introduce a series of evaluation benchmarks with varying difficulty. (ii) We propose a joint synthesis and retrieval framework, as well as negative filtering, and obtain state-of-the-art performance on text-motion retrieval. (iii) We provide extensive experiments to analyze the effects of each component in controlled settings. Our code and models are publicly available2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 375, + 386, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 375, + 386, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 386, + 388 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 431 + ], + "type": "text", + "content": "We present an overview of the most closely related work on text-to-motion synthesis and retrieval, as well as a brief discussion on cross-modal retrieval works." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 452, + 547, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 547, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 547, + 691 + ], + "type": "text", + "content": "Text and human motion. The research on human motion modeling has recently witnessed an increasing interest in bridging the gap between semantics and 3D human body motions, in particular for text-conditioned motion synthesis. Unlike unconstrained 3D human motion synthesis [53, 58, 59], action-conditioned [16, 35], or text-conditioned [5, 9, 14, 22, 24, 36, 49, 56] models add semantic controls to the generation process. The goal of these works is to generate either deterministically [2, 14, 27] or probabilistically [36], motion sequences that are faithful to the textual description inputs. Note that this is different than gesture synthesis from speech [17], in that the text describes the motion content. However, despite remarkable progress in text-to-image synthesis [40, 42], text-to-motion synthesis remains at a very nascent stage. The realism of the synthesized motions is limited, e.g., foot sliding artifacts [36]. We turn to text-to-motion retrieval as another alternative, and perhaps complementary approach to obtain motions for a given textual description. Our focus is therefore different from the work on synthesis. However, we do make use of a motion synthesis branch to aid the retrieval task." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "1While the paper [15] describes a motion-to-text retrieval metric, we notice that the provided code performs text-to-motion retrieval." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 702, + 465, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 465, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 465, + 712 + ], + "type": "text", + "content": "2https://mathis.petrovich.fr/tmr" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9489" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 179 + ], + "type": "text", + "content": "Motion retrieval is relatively less explored. As briefly mentioned in Section 1, motion-to-motion retrieval (e.g., motion matching [8, 20, 45]) methods exist. However, the text-to-motion retrieval task is more challenging due to being cross-modal, i.e., nearest neighbor search across text and motion modalities. Within this category, the very recent work of Guo et al. [15] trains a retrieval model purely for evaluation purposes, and applies a margin-based contrastive loss [18], using Euclidean distance between all pairs within a batch." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 290, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 290, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 290, + 301 + ], + "type": "text", + "content": "Two works are particularly relevant to ours. Firstly, we build on the TEMOS [36] text-to-motion synthesis model, which also has a cross-modal embedding space. However, text and motion embeddings are encouraged to be similar only across positive pairs. We therefore add a contrastive training strategy to incorporate negatives, consequently improving its retrieval capability from a large gallery of fine-grained motions. Secondly, we compare to the aforementioned method of [15], whose motion-to-text retrieval model is adopted for measuring text-to-motion generation performance automatically by other works [10, 49, 57]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 313, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 288, + 456 + ], + "type": "text", + "content": "Cross-modal retrieval. Among widely adopted vision & language retrieval models, some successful examples include CLIP [39], BLIP [26], and CoCa [54] for images, and MILNCE [33], Frozen [7], and CLIP4Clip [30] for videos. They all use variants of cross-modal contrastive learning techniques, such as InfoNCE [34], which we also employ in this work. As discussed in Section 1, we draw inspiration from BLIP [26] and CoCa [54], which add synthesis branches to standard retrieval frameworks. Our approach is similar in spirit to these works in that we perform a cross-modal vision & language retrieval task, but differ in focusing on 3D human motion retrieval, which, to the best of our knowledge, has not been benchmarked." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 466, + 178, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 178, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 178, + 479 + ], + "type": "text", + "content": "3. Text-to-motion retrieval" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 288, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 559 + ], + "type": "text", + "content": "In this section, we introduce the task and the terminology associated with text-to-motion retrieval (Section 3.1). Next, we present our model, named TMR, and its training protocol (Section 3.2). We then explain our simple approach for identifying and filtering incorrect negatives (Section 3.3), followed by a discussion of the implementation details (Section 3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 563, + 120, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 120, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 120, + 574 + ], + "type": "text", + "content": "3.1. Definitions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Given a natural language query " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": ", such as 'A person walks and then makes a right turn', the goal is to rank the motions from a database (i.e., the gallery) according to their semantic correspondence with the text query, and to retrieve the motion that matches best to the textual description. In other words, the task involves sorting the database so that the top ranked motions are the most relevant matches, i.e., creating a search engine to index motions. Additionally, we define the symmetric (and complementary) task, namely motion-to-text retrieval, where the aim is to retrieve the most suitable caption that matches a given motion from a database of texts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": "3D human motion refers to a sequence of human poses. The task does not impose any limitations on the type of representation used, such as joint positions, rotations, or parametric models such as SMPL [28]. As detailed in Section 4.1, we choose to use the representation employed by Guo et al. [15] to facilitate comparisons with previous work. While our experiments are based on the SMPL skeleton, our method is applicable to any skeleton topology. One could alternatively use retargetting [1, 51, 52] to adapt to other body representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 180, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 547, + 277 + ], + "type": "text", + "content": "Text description refers to a sequence of words describing the action performed by a human in natural language. We do not restrict the format of the motion description. The text can be simply an action name (e.g., 'walk') or a full sentence (e.g., 'a human is walking'). The sentences can be fine-grained (e.g., 'a human is walking in a circle slowly'), and may contain one or several actions, simultaneously (e.g., 'walking while waving') or sequentially (e.g., 'walking then sitting')." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 285, + 512, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 512, + 298 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 512, + 298 + ], + "type": "text", + "content": "3.2. Joint training of retrieval and synthesis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 304, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 388 + ], + "type": "text", + "content": "We introduce TMR, which extends the Transformer-based text-to-motion synthesis model TEMOS [36] by incorporating additional losses to make it suitable for the retrieval task. The architecture consists of two independent encoders for inputting motion and text, as well as a decoder that outputs motion (see Figure 2 for an overview). In the following, we review TEMOS [36] components and our added contrastive training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "text", + "content": "Dual encoders. One approach to solving cross-modal retrieval tasks involves defining a similarity function between the two modalities. In our case, the two modalities are text and motion. The similarity function can be applied to compare a given query with each element in the database, and the maximum value would indicate the best match. In this paper, we follow the approach taken by previous work on metric learning, such as CLIP [39], by defining one encoder for each modality and then computing the cosine similarity between their respective embeddings. Such dual embedding has the advantage of fast inference time since the gallery embeddings can be computed and stored beforehand [32]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "content": "Our model is built upon the components of TEMOS [36], which already provides a motion encoder and a text encoder, mapping them to a joint space (building on the idea from Language2Pose [2]); this serves as a strong baseline for our work. Both motion and text encoders are Transformer encoders [50] with additional learnable distribution parameters, as in the VAE-based ACTOR [35]. They are probabilistic in nature, outputting parameters of a Gaussian distribution " + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "inline_equation", + "content": "(\\mu" + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\Sigma)" + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "content": " from which a latent vector " + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "inline_equation", + "content": "z\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "content": " can be sampled. While the text encoder takes text features from a pre-trained and frozen DistilBERT [43] network as input, the motion sequence is fed directly in the motion encoder. Note that when performing retrieval, we directly use the output embedding that corresponds to the mean token " + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "inline_equation", + "content": "(\\mu^{M}" + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "content": " for motion, " + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mu^T" + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "content": " for text)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Motion decoder. TEMOS is trained for the task of motion" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9490" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 532, + 232 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 532, + 232 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 532, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 532, + 232 + ], + "type": "image", + "image_path": "b6aaf7f1f0ffd34b6555d615d240dfd6932730dbd2666a2d89ac1a9c9372187f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "type": "text", + "content": "Figure 2. Joint motion retrieval and synthesis: A simplified view of our TMR framework is presented, where we focus on the similarity matrix defined between text-motion pairs within a batch. Here, we show a batch of 3 samples for illustration purposes. The goal of the contrastive objective is to maximize the diagonal denoting positive pairs (green), and to minimize the off-diagonal negative pair items that have text similarity below a threshold (red). In this example, remaining similarities " + }, + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "type": "inline_equation", + "content": "S_{23}" + }, + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "type": "inline_equation", + "content": "S_{32}" + }, + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "type": "text", + "content": " are discarded from the loss computation because there is high text similarity between " + }, + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "type": "inline_equation", + "content": "T_{2}" + }, + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "type": "inline_equation", + "content": "T_{3}" + }, + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "type": "text", + "content": ". The rest of the model remains similar to TEMOS [36], which decodes a motion from both text " + }, + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "type": "inline_equation", + "content": "z^{T_i}" + }, + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "type": "text", + "content": " and motion " + }, + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "type": "inline_equation", + "content": "z^{M_j}" + }, + { + "bbox": [ + 46, + 234, + 548, + 301 + ], + "type": "text", + "content": " latent vectors. See text for further details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 310, + 290, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 290, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 290, + 395 + ], + "type": "text", + "content": "synthesis and comes equipped with a motion decoder branch. This decoder is identical to the one used in ACTOR [35], which supports variable-duration generation. More specifically, it takes a latent vector " + }, + { + "bbox": [ + 46, + 310, + 290, + 395 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 310, + 290, + 395 + ], + "type": "text", + "content": " and a sinusoidal positional encoding as input, and generates a motion non-autoregressively through a single forward pass. We show in Section 4.2 that keeping this branch helps improve the results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "content": "TEMOS losses. We keep the same base set of losses of [36], defined as the weighted sum " + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TEMOS}} = \\mathcal{L}_{\\mathrm{R}} + \\lambda_{\\mathrm{KL}}\\mathcal{L}_{\\mathrm{KL}} + \\lambda_{\\mathrm{E}}\\mathcal{L}_{\\mathrm{E}}" + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "content": ". In summary, a reconstruction loss term " + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{R}}" + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "content": " measures the motion reconstruction given text or motion input (via a smooth L1 loss). A Kullback-Leibler (KL) divergence loss term " + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}" + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "content": " is composed of four losses: two of them to regularize each encoded distributions " + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "inline_equation", + "content": "-\\mathcal{N}(\\mu^M,\\Sigma^M)" + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "content": " for motion and " + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu^T,\\Sigma^T)" + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "content": " for text - to come from a normal distribution " + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "content": ". The other two enforce distribution similarity between the two modalities. A cross-modal embedding similarity loss " + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{E}}" + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "content": " enforces both text " + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "inline_equation", + "content": "z^{T}" + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "content": " and motion " + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "inline_equation", + "content": "z^{M}" + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "content": " latent codes to be similar to each other (with a smooth L1 loss). We set " + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{KL}}" + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{E}}" + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 46, + 403, + 290, + 548 + ], + "type": "text", + "content": " in our experiments as in [36]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": "Contrastive training. While TEMOS has a cross-modal embedding space, its major drawback to be usable as an effective retrieval model is that it is never trained with negatives, but only positive motion-text pairs. To overcome this limitation, we incorporate a contrastive training with the usage of negative samples to better structure the latent space. Given a batch of N (positive) pairs of latent codes " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "(z_1^T,z_1^M),\\dots,(z_N^T,z_N^M)" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": ", we define any pair " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "(z_i^T,z_j^M)" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "i\\neq j" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " as negative. The similarity matrix " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": " computes the pairwise cosine similarities for all pairs in the batch " + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "inline_equation", + "content": "S_{ij} = \\cos (z_i^T,z_j^M)" + }, + { + "bbox": [ + 46, + 557, + 289, + 715 + ], + "type": "text", + "content": ". In contrast to Guo et al. [15], who consider one random negative per batch (and use a margin loss), we adopt the more recent formulation of InfoNCE [34], which was proven effective in other work [7, 26, 39]. This loss term" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 310, + 408, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 310, + 408, + 321 + ], + "spans": [ + { + "bbox": [ + 306, + 310, + 408, + 321 + ], + "type": "text", + "content": "can be defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 327, + 546, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 546, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {N C E}} = - \\frac {1}{2 N} \\sum_ {i} \\left(\\log \\frac {\\exp S _ {i i} / \\tau}{\\sum_ {j} \\exp S _ {i j} / \\tau} + \\log \\frac {\\exp S _ {i i} / \\tau}{\\sum_ {j} \\exp S _ {j i} / \\tau}\\right), \\tag {1}", + "image_path": "ac5ca49fe71b7d40fffec8cdb314aa101517a269fc3621c0e1dbadfb9105405b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 364, + 473, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 364, + 473, + 375 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 473, + 375 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 364, + 473, + 375 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 364, + 473, + 375 + ], + "type": "text", + "content": " is the temperature hyperparameter." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 376, + 546, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 546, + 411 + ], + "type": "text", + "content": "Training loss. The total loss we use to train TMR is the weighted sum " + }, + { + "bbox": [ + 305, + 376, + 546, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TEMOS}} + \\lambda_{\\mathrm{NCE}}\\mathcal{L}_{\\mathrm{NCE}}" + }, + { + "bbox": [ + 305, + 376, + 546, + 411 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 376, + 546, + 411 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{NCE}}" + }, + { + "bbox": [ + 305, + 376, + 546, + 411 + ], + "type": "text", + "content": " controls the importance of the contrastive loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 419, + 416, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 416, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 416, + 432 + ], + "type": "text", + "content": "3.3. Filtering negatives" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 437, + 546, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 486 + ], + "type": "text", + "content": "As mentioned in Section 1, the descriptions accompanying motion capture collections can be repetitive or similar across the training motions. We wish to prevent defining negatives between text-motion pairs that contain similar descriptions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 569 + ], + "type": "text", + "content": "Consider for example the two text descriptions, \"A human making a counterclockwise turn\" and \"A person walks quarter a circle to the left\". In the KIT-ML benchmark [37], these two descriptions appear as two different annotations for the same motion. Due to the flexibility and ambiguity of natural language, different words may describe the same concepts (e.g., 'counterclockwise', 'circle to the left')." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 653 + ], + "type": "text", + "content": "During training, the random selection of batches can adversely affect the results because the model may have to push away two latent vectors that correspond to similar meanings. This can force the network to focus on unimportant details (e.g., 'someone' vs 'human'), ultimately resulting in decreased performance due to unstable behavior and reduced robustness to text variations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "To alleviate this issue, we leverage an external large language model to provide sentence similarity scores. In particular, we use MPNet [47] to encode sentences and compute similarities between two text descriptions. We then determine whether to filter a pair of text descriptions " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "(t_{1}, t_{2})" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " if their similarity" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9491" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "is higher than a certain threshold, referring them as 'wrong negatives'. During training, we filter wrong negative pairs from the loss computation. Note that we refrain from defining them as positives, as the language model may also incorrectly say two descriptions are similar when they are not." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 138, + 179, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 138, + 179, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 138, + 179, + 152 + ], + "type": "text", + "content": "3.4. Implementation details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 156, + 288, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 288, + 266 + ], + "type": "text", + "content": "We use the AdamW optimizer [29] with a learning rate of " + }, + { + "bbox": [ + 46, + 156, + 288, + 266 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 156, + 288, + 266 + ], + "type": "text", + "content": " and a batch size of 32. Since the batch size can be an important hyperparameter for the InfoNCE loss, due to determining the number of negatives, we report experimental results with different values. The latent dimensionality of the embeddings is " + }, + { + "bbox": [ + 46, + 156, + 288, + 266 + ], + "type": "inline_equation", + "content": "d = 256" + }, + { + "bbox": [ + 46, + 156, + 288, + 266 + ], + "type": "text", + "content": ". We set the temperature " + }, + { + "bbox": [ + 46, + 156, + 288, + 266 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 156, + 288, + 266 + ], + "type": "text", + "content": " to 0.1, and the weight of the contrastive loss term " + }, + { + "bbox": [ + 46, + 156, + 288, + 266 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{NCE}}" + }, + { + "bbox": [ + 46, + 156, + 288, + 266 + ], + "type": "text", + "content": " to 0.1. The threshold to filter negatives is set to 0.8. We provide experimental analyses to measure the sensitivity to these added hyperparameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 274, + 124, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 124, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 124, + 288 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 294, + 288, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 288, + 390 + ], + "type": "text", + "content": "We start by describing the datasets and evaluation protocol used in the experiments (Section 4.1). We then report the performance of our model on our new retrieval benchmark along with comparison to prior work (Section 4.2). Next, we present our ablation study measuring the effects of the additional contrastive loss, the negative filtering, and the hyperparameters (Section 4.3). Finally, we provide qualitative results for retrieval (Section 4.4), and our use case of moment retrieval (Section 4.5)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 396, + 180, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 180, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 180, + 407 + ], + "type": "text", + "content": "4.1. Datasets and evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 414, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 593 + ], + "type": "text", + "content": "HumanML3D dataset (H3D) [15] provides natural language labels to describe the motions in AMASS [31] and HumanAct12 [16] motion capture collections. We follow the motion pre-processing procedure of [15], and apply the SMPL layer [28] to extract joint positions, canonicalize the skeletons to share the same topology (i.e., same bone lengths), then compute motion features (extracting local positions, velocities and foot contacts similar to Holden et al. [21]). The data is then augmented by mirroring left and right (both in motions and their corresponding texts). After this procedure, and following the official split, we obtain 23384, 1460, 4380 motions for the training, validation, and test sets, respectively. On average, each motion is annotated 3.0 times with different text. During training we randomly select one as the matching text, for testing we use the first text." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 677 + ], + "type": "text", + "content": "KIT Motion-Language dataset (KIT) [37] also come from motion capture data, with an emphasis on locomotion motions. It originally consists of 3911 motion sequences and 6278 text sentences. We pre-process the motions with the same procedure as in H3D. The data is split into 4888, 300, 830 motions for training, validation, and test sets, respectively. In this dataset, each motion is annotated 2.1 times on average." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Evaluation protocol. We report standard retrieval performance measures: recall at several ranks, R@1, R@2, etc. for both text-to-motion and motion-to-text tasks (identical to the R" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "Precision metrics adopted by Guo et al. [15]). Recall at rank " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " measures the percentage of times the correct label is among the top " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " results; therefore higher is better. We additionally report median rank (MedR), where lower is better. Note that the retrieval performance is evaluated on a gallery of unseen real motions (i.e., test set). A study of the synthesis performance of TMR can be found in Appendix B." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 157, + 546, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 546, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 546, + 433 + ], + "type": "text", + "content": "We define several evaluation protocols, mainly changing the gallery set. (a) All the test set is used as a first protocol, without any modification. This set is partially problematic because there are repetitive texts across motions, or just minor differences (e.g., person vs human, walk vs walking). (b) All with threshold means we search over all the test set, but, in this case, we accept a retrieved motion as correct if its text label is similar to the query text above a threshold. For example, retrieving the motion corresponding to \"A human walks forward\" should be correct when the input query is \"Someone is walking forward\". We set a high threshold of 0.95 (scaled between [0, 1]) to remove very similar texts without removing too many fine-grained details. Appendix A provides statistics on how often similar text descriptions appear in the datasets. (c) Dissimilar subset refers to sampling 100 motion-text pairs whose texts are maximally far from each other (using an approximation of the quadratic knapsack problem [3]). This evaluation can be considered as an easy, but clean, subset of the previous ones. (d) Small batches is included to mimic the protocol described by Guo et al. [15], who randomly pick batches of 32 motion-text pairs and report the average performance. An ideal evaluation metric should not have randomness, and a gallery size of 32 is relatively easy compared to the previous protocols." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 442, + 545, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 442, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 442, + 545, + 455 + ], + "type": "text", + "content": "4.2. A new benchmark & comparison to prior work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 461, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 568 + ], + "type": "text", + "content": "We present the performance of our model on this new retrieval benchmark, on H3D (Table 1) and KIT (Table 2) datasets, across all evaluation protocols. We also compare against prior work TEMOS [36] and Guo et al. [15]. For TEMOS, we retrain their model on both datasets to have a comparable benchmark since the original model differs in motion representation and lacks left/right data augmentation (and they only provide a KIT-pretrained model, not H3D). For [15], we take their publicly available models trained on these two datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "TEMOS in particular is not designed to perform well on retrieval, since its cross-modal embedding space is only trained with positive pairs. However, Guo et al. train contrastively with negatives as well, using a margin loss [18]. For all 4 evaluation sets with varying difficulties, TMR outperforms the prior work, suggesting our model better captures the finegrained nature of motion descriptions. The model of [15] is adopted as part of motion synthesis evaluation in several works. TMR may therefore provide a better alternative. Our significant improvements over the state of the art can be credited to (i) jointly training for synthesis and retrieval, (ii) adopting the more recent contrastive objective, InfoNCE [34], while (iii) carefully" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9492" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 541, + 239 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 541, + 239 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 541, + 239 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 541, + 239 + ], + "type": "table", + "html": "
ProtocolMethodsText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓
(a) AllTEMOS [36]2.124.095.878.2613.52173.03.864.546.949.3814.00183.25
Guo et al. [15]1.803.424.797.1212.4781.002.923.746.008.3612.9581.50
TMR5.6810.5914.0420.3430.9428.009.9512.4417.9523.5632.6928.50
(b) All with thresholdTEMOS [36]5.218.2211.1415.0922.1279.005.486.199.0012.0117.10129.0
Guo et al. [15]5.307.8310.7514.5922.5154.004.955.688.9311.6416.9469.50
TMR11.6015.3920.5027.7238.5219.0013.2015.7322.0327.6537.6321.50
(c) Dissimilar subsetTEMOS [36]33.0042.0049.0057.0066.004.0035.0044.0050.0056.0070.003.50
Guo et al. [15]34.0048.0057.0072.0084.003.0034.0047.0059.0072.0083.003.00
TMR47.0061.0071.0080.0086.002.0048.0063.0069.0080.0084.002.00
(d) Small batches [15]TEMOS [36]40.4953.5261.1470.9684.152.3339.9653.4961.7972.4085.892.33
Guo et al. [15]52.4871.0580.6589.6696.581.3952.0071.2181.1189.8796.781.38
TMR67.1681.3286.8191.4395.361.0467.9781.2086.3591.7095.271.03
", + "image_path": "3afcb079e8aed2e53f3d84be031af0eda3c7d94f7250a33f1a2adb7d3e509945.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 283, + 541, + 452 + ], + "blocks": [ + { + "bbox": [ + 47, + 240, + 546, + 275 + ], + "lines": [ + { + "bbox": [ + 47, + 240, + 546, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 546, + 275 + ], + "type": "text", + "content": "Table 1. Text-to-motion retrieval benchmark on HumanML3D: We establish four evaluation protocols as described in Section 4.1, with decreasing difficulty from (a) to (d). Our model TMR substantially outperforms the prior work of Guo et al. [15] and TEMOS [36], on the challenging H3D dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 283, + 541, + 452 + ], + "lines": [ + { + "bbox": [ + 51, + 283, + 541, + 452 + ], + "spans": [ + { + "bbox": [ + 51, + 283, + 541, + 452 + ], + "type": "table", + "html": "
ProtocolMethodsText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓R@1↑R@2↑R@3↑R@5↑R@10↑MedR↓
(a) AllTEMOS [36]7.1113.2517.5924.1035.6624.0011.6915.3020.1226.6336.3926.50
Guo et al. [15]3.376.9910.8416.8727.7128.004.946.5110.7216.1425.3028.50
TMR7.2313.9820.3628.3140.1217.0011.2013.8620.1228.0738.5518.00
(b) All with thresholdTEMOS [36]18.5524.3430.8442.2956.397.0017.7122.4128.8035.4247.1113.25
Guo et al. [15]13.2522.6529.7639.0449.5211.0010.4813.9820.4827.9538.5517.25
TMR24.5830.2441.9350.4860.365.0019.6423.7332.5341.2053.019.50
(c) Dissimilar subsetTEMOS [36]24.0040.0046.0054.0070.005.0033.0039.0045.0049.0064.006.50
Guo et al. [15]16.0029.0036.0048.0066.006.0024.0029.0036.0046.0066.007.00
TMR26.0046.0060.0070.0083.003.0034.0045.0060.0069.0082.003.50
(d) Small batches [15]TEMOS [36]43.8858.2567.0074.0084.752.0641.8855.8865.6275.2585.752.25
Guo et al. [15]42.2562.6275.1287.5096.121.8839.7562.7573.6286.8895.881.95
TMR49.2569.7578.2587.8895.001.5050.1267.1276.8888.8894.751.53
", + "image_path": "ab545cb039191b6440d9d8e00b856d5bb7a972c639ac266d6e533def84821e18.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 494, + 282, + 557 + ], + "blocks": [ + { + "bbox": [ + 47, + 453, + 546, + 477 + ], + "lines": [ + { + "bbox": [ + 47, + 453, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 546, + 477 + ], + "type": "text", + "content": "Table 2. Text-to-motion retrieval benchmark on KIT-ML: As in Table 1, we report the four evaluation protocols, this time on the KIT dataset. Again, TMR significantly improves over Guo et al. [15] and TEMOS [36] across all protocols and metrics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 494, + 282, + 557 + ], + "lines": [ + { + "bbox": [ + 50, + 494, + 282, + 557 + ], + "spans": [ + { + "bbox": [ + 50, + 494, + 282, + 557 + ], + "type": "table", + "html": "
Motion \nRecons.InfoNCEMarginText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
XX✓15.0622.1725.7812.008.1911.5716.3919.50
X✓X19.7625.3036.876.0017.4719.7630.609.50
✓XX18.5524.3430.847.0017.7122.4128.8013.25
✓X✓19.8824.4634.467.0014.7019.7628.1912.50
✓✓X24.5830.2441.935.0019.6423.7332.539.50
", + "image_path": "cd2b7e05c8e36c910e8020945fcda298b15c7e845990901e9b09351014ec8167.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 559, + 287, + 625 + ], + "lines": [ + { + "bbox": [ + 47, + 559, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 287, + 625 + ], + "type": "text", + "content": "Table 3. Losses: We experiment with various loss definitions (i) with/without the motion reconstruction, and (ii) the choice of the contrastive loss between InfoNCE and margin-based. We see that InfoNCE [34] is a better alternative to the contrastive loss with Euclidean margin [18] (employed by Guo et al. [15]). The reconstruction loss through the motion decoder branch further boosts the results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 637, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 637, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 637, + 287, + 662 + ], + "type": "text", + "content": "eliminating wrong negatives. In the following, we ablate these components in controlled experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 139, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 139, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 139, + 684 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "The rest of the quantitative evaluation uses the ' (b) All with threshold' evaluation protocol, on the KIT dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 495, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 712 + ], + "type": "text", + "content": "Which losses matter? Table 3 compares several variants of TMR where we check (a) whether the jointly trained motion synthesis branch helps retrieval, and (b) how important the form of the contrastive loss is. When removing the synthesis branch and only using the contrastive loss, we perform a deterministic encoding (i.e., with a single token instead of two tokens " + }, + { + "bbox": [ + 304, + 495, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 495, + 547, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 495, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 495, + 547, + 712 + ], + "type": "text", + "content": "). First, we see that the motion synthesis branch certainly improves results compared with only training using a contrastive loss (e.g., 41.93 vs 36.87 R@3). This possibly forces the latent vector to capture the full content of the input text (i.e., instead of picking up on a subset of words, or bag-of-words [11, 55] upon finding a shortcut that satisfies the contrastive loss). Second, in the presence of a contrastive loss, the InfoNCE formulation is significantly better than the margin loss employed by previous work [15] (41.93 vs 34.46 R@1). Note that for this experiment, we keep the same negative filtering for both the margin loss and InfoNCE (we provide additional experiments without the negative filtering in Appendix B)." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9493" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 282, + 184 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 282, + 184 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 282, + 184 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 282, + 184 + ], + "type": "table", + "html": "
ThresholdText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
0.5519.4023.2530.489.0017.8321.6929.5214.00
0.6017.9526.8736.876.0020.6024.7031.8111.25
0.6523.0128.6736.397.0019.0421.6929.7611.50
0.7022.2929.6438.806.0018.1922.7732.059.00
0.7520.0027.1137.836.0020.2424.4634.229.50
0.8024.5830.2441.935.0019.6423.7332.539.50
0.8521.4525.7838.436.5020.8424.1033.379.50
0.9023.2530.1240.486.0020.0025.1833.139.50
0.9520.4826.9938.436.0019.2823.3731.9310.25
X22.1727.8336.027.0016.7521.3332.1711.50
", + "image_path": "1d329fd76a290f3b80e2dc5ffbd0392d811ceb522155dcf1b61bf853aa6bbf6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 186, + 287, + 231 + ], + "lines": [ + { + "bbox": [ + 47, + 186, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 287, + 231 + ], + "type": "text", + "content": "Table 4. Filtering negatives: We compare several threshold values for filtering negatives from the loss comparison due to having similar texts. We observe that removing negatives based on text similarity above 0.8 (from a scale between [0,1]) performs well overall." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 240, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 384 + ], + "type": "text", + "content": "The effect of filtering negatives. As explained in Section 3.3, during training we filter out pairs whose texts are closer than a threshold in an embedding space, and do not count them in the contrastive loss computation. Note that we still keep each item in the batch for the motion synthesis objective. In Table 4, we perform experiments with a range of different values for this threshold selection. On a scale between [0, 1], a threshold of 0.8 has the best results, balancing keeping a sufficient number of negatives and removing the wrong ones. Without filtering at all, the performance remains at " + }, + { + "bbox": [ + 46, + 240, + 287, + 384 + ], + "type": "inline_equation", + "content": "36.02\\mathrm{R}@\\mathfrak{z}3" + }, + { + "bbox": [ + 46, + 240, + 287, + 384 + ], + "type": "text", + "content": " (compared to 41.93). We provide statistics on the percentage of filtered pairs in Appendix A." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": "Hyperparameters of the contrastive training. We show the sensitivity of our model to several hyperparameters added when extending TEMOS: (i) temperature " + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": " of the cross entropy of InfoNCE [34] in Eq.1, (ii) the " + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{NCE}}" + }, + { + "bbox": [ + 46, + 385, + 288, + 517 + ], + "type": "text", + "content": " weighting parameter, and (iii) the batch size, which determines the number of negatives. We see in Table 5 that the model is indeed sensitive to the temperature, which is a common observation in other settings. The weight parameter and the batch size are relatively less important while also influencing the results to a certain extent. An experiment with the latent dimensionality hyperparameter can be found in Appendix B." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 526, + 155, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 155, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 155, + 539 + ], + "type": "text", + "content": "4.4. Qualitative results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "In Figure 3, we provide sample qualitative results for text-to-motion retrieval on the full test set of H3D. For each query text displayed on the left, top-5 retrieved motions are shown on the right along with their similarity scores. Note that the ground-truth text labels (at the bottom of each motion) for the retrieved motions are not used, and the gallery motions are unseen at training. For the first two examples with 'playing violin' and 'handstand', we retrieve the ground-truth motion at rank 1. We observe that the next ranked motions depict visually similar motions as well (e.g., 'cartwheel' involves standing on the hands). For the free-from prompt example 'Someone is swimming' (i.e., the exact text does not appear in the gallery), the three first motions resemble or involve the swimming action, whereas motions at ranks 4 and 5 are incorrect. We notice that" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 310, + 71, + 543, + 134 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 543, + 134 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 543, + 134 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 543, + 134 + ], + "type": "table", + "html": "
Temp.Text-motion retrievalMotion-text retrieval
τR@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
0.0019.5221.8127.2312.007.479.7616.5115.50
0.0121.4529.0438.806.0021.0827.1133.619.50
0.124.5830.2441.935.0019.6423.7332.539.50
1.01.081.933.61306.51.811.932.41372.0
", + "image_path": "1ee01f8a42c4ff91b6f4a83e2f98b1ea22ef122c6730c2bf4b5113f65f91fdc6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 150, + 542, + 215 + ], + "blocks": [ + { + "bbox": [ + 421, + 134, + 430, + 142 + ], + "lines": [ + { + "bbox": [ + 421, + 134, + 430, + 142 + ], + "spans": [ + { + "bbox": [ + 421, + 134, + 430, + 142 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 150, + 542, + 215 + ], + "lines": [ + { + "bbox": [ + 310, + 150, + 542, + 215 + ], + "spans": [ + { + "bbox": [ + 310, + 150, + 542, + 215 + ], + "type": "table", + "html": "
Weight λNCEText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
0.00118.5523.2536.757.0018.1924.3431.4511.50
0.0120.8426.9937.237.0018.9223.1332.1710.25
0.124.5830.2441.935.0019.6423.7332.539.50
1.019.5224.4634.467.0019.0424.3435.069.50
", + "image_path": "a7d3f287d00acf66d9d250614857db39117aa75bd8c798951bc33f9cf62c3921.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 232, + 542, + 296 + ], + "blocks": [ + { + "bbox": [ + 421, + 215, + 430, + 223 + ], + "lines": [ + { + "bbox": [ + 421, + 215, + 430, + 223 + ], + "spans": [ + { + "bbox": [ + 421, + 215, + 430, + 223 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 232, + 542, + 296 + ], + "lines": [ + { + "bbox": [ + 309, + 232, + 542, + 296 + ], + "spans": [ + { + "bbox": [ + 309, + 232, + 542, + 296 + ], + "type": "table", + "html": "
Batch sizeText-motion retrievalMotion-text retrieval
R@1↑R@2↑R@3↑MedR↓R@1↑R@2↑R@3↑MedR↓
1625.4231.5740.126.0020.3624.1033.738.00
3224.5830.2441.935.0019.6423.7332.539.50
6420.2426.5138.196.0019.5224.2232.059.50
12818.5528.8036.757.0014.9418.4326.1411.50
", + "image_path": "c9f8c1baabceadd3e65ba948a32c4af441ff8ff5145e879a712d7d0c146642e7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 303, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 305, + 303, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 545, + 358 + ], + "type": "text", + "content": "Table 5. Hyperparameters of the contrastive training: We measure the sensitivity to the parameters " + }, + { + "bbox": [ + 305, + 303, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 303, + 545, + 358 + ], + "type": "text", + "content": " (temperature), " + }, + { + "bbox": [ + 305, + 303, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 305, + 303, + 545, + 358 + ], + "type": "text", + "content": " the weight of the contrastive loss, and the batch size. Note that the learning rate is proportionally altered when changing the batch size. We display a wide range of values to show the full trends." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 421, + 296, + 431, + 303 + ], + "lines": [ + { + "bbox": [ + 421, + 296, + 431, + 303 + ], + "spans": [ + { + "bbox": [ + 421, + 296, + 431, + 303 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 376, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 426 + ], + "type": "text", + "content": "the incorrect motions have a low similarity " + }, + { + "bbox": [ + 304, + 376, + 545, + 426 + ], + "type": "inline_equation", + "content": "(< 0.6)" + }, + { + "bbox": [ + 304, + 376, + 545, + 426 + ], + "type": "text", + "content": ", and the human bodies are rotated similarly as in swimming. More qualitative results can be seen in Appendix C, as well as our supplementary video on the project page." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 432, + 457, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 457, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 457, + 443 + ], + "type": "text", + "content": "4.5. Use case: Moment retrieval" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 545 + ], + "type": "text", + "content": "While our focus is retrieval, once our model is trained, it can be used for a different use case. Here, we test the limits of our approach, by qualitatively evaluating the capability of TMR on the task of temporally localizing a natural language query in a long 3D motion sequence. This is similar in spirit to moment retrieval in videos [12, 13, 19, 25, 41]. It is also related to categorical action localization in 3D motions [48]; however, our input is free-form text instead of symbolic action classes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": "In Figure 4, we show four examples, where we apply a model pre-trained on H3D on BABEL sequences. In each example, the queried text is displayed on the top. The x-axis denotes the frame number, the green rectangle represents the ground-truth location for the given action, and the dashed red line marks the localization with the maximum similarity. We simply compute the motion features in a sliding window manner. The similarity between the text label and a 20-frame window centered at each frame is shown in the y-axis as a 1D plot over time. Despite our model not being trained for temporal localization, we observe its grounding potential. Moreover, there exists a domain gap between BABEL label used at test time and H3D used for training. Quantitative evaluation and more qualitative examples are included in Appendix B." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9494" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 94, + 123, + 123 + ], + "blocks": [ + { + "bbox": [ + 51, + 94, + 123, + 123 + ], + "lines": [ + { + "bbox": [ + 51, + 94, + 123, + 123 + ], + "spans": [ + { + "bbox": [ + 51, + 94, + 123, + 123 + ], + "type": "image", + "image_path": "064821e401a50ca69dde23dd5a446a4b51b8e4efa18757d1fd7b136ea203c749.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 135, + 71, + 210, + 139 + ], + "blocks": [ + { + "bbox": [ + 135, + 71, + 210, + 139 + ], + "lines": [ + { + "bbox": [ + 135, + 71, + 210, + 139 + ], + "spans": [ + { + "bbox": [ + 135, + 71, + 210, + 139 + ], + "type": "image", + "image_path": "a2bf02720a19f22ce9b82b0dcad9d9d8cb9e958c2e4657c6dcb42493d3a7da83.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 141, + 200, + 156 + ], + "lines": [ + { + "bbox": [ + 146, + 141, + 200, + 156 + ], + "spans": [ + { + "bbox": [ + 146, + 141, + 200, + 156 + ], + "type": "text", + "content": "A person appears to be playing the violin" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 214, + 72, + 288, + 139 + ], + "blocks": [ + { + "bbox": [ + 214, + 72, + 288, + 139 + ], + "lines": [ + { + "bbox": [ + 214, + 72, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 214, + 72, + 288, + 139 + ], + "type": "image", + "image_path": "95de5ac708fe12c406124aafeb2cd85604bbadcfa942bb983f28925b75979736.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 141, + 277, + 156 + ], + "lines": [ + { + "bbox": [ + 226, + 141, + 277, + 156 + ], + "spans": [ + { + "bbox": [ + 226, + 141, + 277, + 156 + ], + "type": "text", + "content": "Moving the hands and work some thing" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 293, + 80, + 367, + 139 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 344, + 79 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 344, + 79 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 344, + 79 + ], + "type": "inline_equation", + "content": "S = 0.85" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 293, + 80, + 367, + 139 + ], + "lines": [ + { + "bbox": [ + 293, + 80, + 367, + 139 + ], + "spans": [ + { + "bbox": [ + 293, + 80, + 367, + 139 + ], + "type": "image", + "image_path": "11dbe0118342598782f59ddb39f0816ae19701acad8cc74874486f3c7740b0b8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 141, + 361, + 156 + ], + "lines": [ + { + "bbox": [ + 299, + 141, + 361, + 156 + ], + "spans": [ + { + "bbox": [ + 299, + 141, + 361, + 156 + ], + "type": "text", + "content": "The man plays violin holding it in his right hand" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 372, + 80, + 446, + 139 + ], + "blocks": [ + { + "bbox": [ + 395, + 72, + 423, + 79 + ], + "lines": [ + { + "bbox": [ + 395, + 72, + 423, + 79 + ], + "spans": [ + { + "bbox": [ + 395, + 72, + 423, + 79 + ], + "type": "inline_equation", + "content": "S = 0.85" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 372, + 80, + 446, + 139 + ], + "lines": [ + { + "bbox": [ + 372, + 80, + 446, + 139 + ], + "spans": [ + { + "bbox": [ + 372, + 80, + 446, + 139 + ], + "type": "image", + "image_path": "a8ef19f55934eeb1450b68677175d78973c978e077db40c5ca1af7ebe474517d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 141, + 426, + 155 + ], + "lines": [ + { + "bbox": [ + 392, + 141, + 426, + 155 + ], + "spans": [ + { + "bbox": [ + 392, + 141, + 426, + 155 + ], + "type": "text", + "content": "The man plays the violin" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 451, + 80, + 525, + 139 + ], + "blocks": [ + { + "bbox": [ + 474, + 72, + 501, + 79 + ], + "lines": [ + { + "bbox": [ + 474, + 72, + 501, + 79 + ], + "spans": [ + { + "bbox": [ + 474, + 72, + 501, + 79 + ], + "type": "inline_equation", + "content": "S = 0.81" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 451, + 80, + 525, + 139 + ], + "lines": [ + { + "bbox": [ + 451, + 80, + 525, + 139 + ], + "spans": [ + { + "bbox": [ + 451, + 80, + 525, + 139 + ], + "type": "image", + "image_path": "84b3b304087e09182b8211dd9d308f76f0dea1531192733122e6995ce72207e0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 140, + 522, + 158 + ], + "lines": [ + { + "bbox": [ + 454, + 140, + 522, + 158 + ], + "spans": [ + { + "bbox": [ + 454, + 140, + 522, + 158 + ], + "type": "text", + "content": "The man holds something above his left shoulder and rubs it with his right hand" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 53, + 190, + 123, + 215 + ], + "blocks": [ + { + "bbox": [ + 53, + 190, + 123, + 215 + ], + "lines": [ + { + "bbox": [ + 53, + 190, + 123, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 190, + 123, + 215 + ], + "type": "image", + "image_path": "8a87e77987a443f81c2ded9248616d3882553e376b11abe60c593628511da90c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 135, + 172, + 210, + 232 + ], + "blocks": [ + { + "bbox": [ + 159, + 164, + 186, + 171 + ], + "lines": [ + { + "bbox": [ + 159, + 164, + 186, + 171 + ], + "spans": [ + { + "bbox": [ + 159, + 164, + 186, + 171 + ], + "type": "inline_equation", + "content": "S = 0.82" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 172, + 210, + 232 + ], + "lines": [ + { + "bbox": [ + 135, + 172, + 210, + 232 + ], + "spans": [ + { + "bbox": [ + 135, + 172, + 210, + 232 + ], + "type": "image", + "image_path": "a53020f79ba182d02c8ffa57d9e80687129bb6d6668dc5e2972ddce4c70f5cc0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 234, + 201, + 247 + ], + "lines": [ + { + "bbox": [ + 144, + 234, + 201, + 247 + ], + "spans": [ + { + "bbox": [ + 144, + 234, + 201, + 247 + ], + "type": "text", + "content": "A man steps forward and does a handstand" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 215, + 172, + 288, + 232 + ], + "blocks": [ + { + "bbox": [ + 238, + 164, + 264, + 171 + ], + "lines": [ + { + "bbox": [ + 238, + 164, + 264, + 171 + ], + "spans": [ + { + "bbox": [ + 238, + 164, + 264, + 171 + ], + "type": "inline_equation", + "content": "S = 0.68" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 215, + 172, + 288, + 232 + ], + "lines": [ + { + "bbox": [ + 215, + 172, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 215, + 172, + 288, + 232 + ], + "type": "image", + "image_path": "a899cfdc8d1e86545b6465259256ace6974da5bad8b9e8ae7e4f53821a02ccd6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 234, + 278, + 249 + ], + "lines": [ + { + "bbox": [ + 225, + 234, + 278, + 249 + ], + "spans": [ + { + "bbox": [ + 225, + 234, + 278, + 249 + ], + "type": "text", + "content": "Doing a cartwheel then jumping up and down." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 293, + 172, + 367, + 232 + ], + "blocks": [ + { + "bbox": [ + 317, + 164, + 343, + 171 + ], + "lines": [ + { + "bbox": [ + 317, + 164, + 343, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 343, + 171 + ], + "type": "inline_equation", + "content": "S = 0.61" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 293, + 172, + 367, + 232 + ], + "lines": [ + { + "bbox": [ + 293, + 172, + 367, + 232 + ], + "spans": [ + { + "bbox": [ + 293, + 172, + 367, + 232 + ], + "type": "image", + "image_path": "ce655c700b0def4dd5399a9fa0672e60d0d1c5efeb27a382a7b8505d91e6e30f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 234, + 349, + 247 + ], + "lines": [ + { + "bbox": [ + 312, + 234, + 349, + 247 + ], + "spans": [ + { + "bbox": [ + 312, + 234, + 349, + 247 + ], + "type": "text", + "content": "The person does 2 cartwheels" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 372, + 172, + 446, + 231 + ], + "blocks": [ + { + "bbox": [ + 395, + 164, + 422, + 171 + ], + "lines": [ + { + "bbox": [ + 395, + 164, + 422, + 171 + ], + "spans": [ + { + "bbox": [ + 395, + 164, + 422, + 171 + ], + "type": "inline_equation", + "content": "S = 0.48" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 372, + 172, + 446, + 231 + ], + "lines": [ + { + "bbox": [ + 372, + 172, + 446, + 231 + ], + "spans": [ + { + "bbox": [ + 372, + 172, + 446, + 231 + ], + "type": "image", + "image_path": "00028c8c67deb4eced219a9704c277e46e53b827736780af3a969053dc255717.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 232, + 442, + 251 + ], + "lines": [ + { + "bbox": [ + 376, + 232, + 442, + 251 + ], + "spans": [ + { + "bbox": [ + 376, + 232, + 442, + 251 + ], + "type": "text", + "content": "A person walks forward then turns completely around and does a cartwheel" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 451, + 172, + 525, + 231 + ], + "blocks": [ + { + "bbox": [ + 474, + 164, + 501, + 171 + ], + "lines": [ + { + "bbox": [ + 474, + 164, + 501, + 171 + ], + "spans": [ + { + "bbox": [ + 474, + 164, + 501, + 171 + ], + "type": "inline_equation", + "content": "S = 0.42" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 451, + 172, + 525, + 231 + ], + "lines": [ + { + "bbox": [ + 451, + 172, + 525, + 231 + ], + "spans": [ + { + "bbox": [ + 451, + 172, + 525, + 231 + ], + "type": "image", + "image_path": "25666dff1d750c88f5195a0fb8e01f23c420c365c65643b87981d1d82f9f87b2.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 234, + 515, + 247 + ], + "lines": [ + { + "bbox": [ + 460, + 234, + 515, + 247 + ], + "spans": [ + { + "bbox": [ + 460, + 234, + 515, + 247 + ], + "type": "text", + "content": "The drunk guy struggles to walk down the street" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 54, + 281, + 116, + 305 + ], + "blocks": [ + { + "bbox": [ + 54, + 281, + 116, + 305 + ], + "lines": [ + { + "bbox": [ + 54, + 281, + 116, + 305 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 116, + 305 + ], + "type": "image", + "image_path": "ecc34bc462a299d4d3b13068a86d8c9240ca88185bfd2c5824cea64fcdda8c30.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 135, + 265, + 209, + 323 + ], + "blocks": [ + { + "bbox": [ + 159, + 257, + 186, + 264 + ], + "lines": [ + { + "bbox": [ + 159, + 257, + 186, + 264 + ], + "spans": [ + { + "bbox": [ + 159, + 257, + 186, + 264 + ], + "type": "inline_equation", + "content": "S = 0.76" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 265, + 209, + 323 + ], + "lines": [ + { + "bbox": [ + 135, + 265, + 209, + 323 + ], + "spans": [ + { + "bbox": [ + 135, + 265, + 209, + 323 + ], + "type": "image", + "image_path": "563bf8f3a65becc7691d235a13bdfbe8dd2969478c21ba0e06dc8ab37377b706.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 327, + 198, + 341 + ], + "lines": [ + { + "bbox": [ + 148, + 327, + 198, + 341 + ], + "spans": [ + { + "bbox": [ + 148, + 327, + 198, + 341 + ], + "type": "text", + "content": "The person was flying around like a fly" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 215, + 265, + 288, + 323 + ], + "blocks": [ + { + "bbox": [ + 238, + 257, + 264, + 264 + ], + "lines": [ + { + "bbox": [ + 238, + 257, + 264, + 264 + ], + "spans": [ + { + "bbox": [ + 238, + 257, + 264, + 264 + ], + "type": "inline_equation", + "content": "S = 0.71" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 215, + 265, + 288, + 323 + ], + "lines": [ + { + "bbox": [ + 215, + 265, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 215, + 265, + 288, + 323 + ], + "type": "image", + "image_path": "948efe27aa5a168d1283927eba6a8a52508679771405fa71ef750c355641d53c.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 327, + 269, + 340 + ], + "lines": [ + { + "bbox": [ + 234, + 327, + 269, + 340 + ], + "spans": [ + { + "bbox": [ + 234, + 327, + 269, + 340 + ], + "type": "text", + "content": "A person swam in free style" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 293, + 265, + 367, + 323 + ], + "blocks": [ + { + "bbox": [ + 317, + 257, + 344, + 264 + ], + "lines": [ + { + "bbox": [ + 317, + 257, + 344, + 264 + ], + "spans": [ + { + "bbox": [ + 317, + 257, + 344, + 264 + ], + "type": "inline_equation", + "content": "S = 0.62" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 293, + 265, + 367, + 323 + ], + "lines": [ + { + "bbox": [ + 293, + 265, + 367, + 323 + ], + "spans": [ + { + "bbox": [ + 293, + 265, + 367, + 323 + ], + "type": "image", + "image_path": "58a9e17f9f91e84c21fd1e6a282c0e564779e30e97fe9d0ce34c0ea1734ccda7.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 325, + 367, + 342 + ], + "lines": [ + { + "bbox": [ + 294, + 325, + 367, + 342 + ], + "spans": [ + { + "bbox": [ + 294, + 325, + 367, + 342 + ], + "type": "text", + "content": "The person is preforming a swimming stroke know as the butterfly stroke. The arms swing from behind the head and reenter the water propelling the person forward." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 372, + 265, + 446, + 323 + ], + "blocks": [ + { + "bbox": [ + 395, + 257, + 422, + 264 + ], + "lines": [ + { + "bbox": [ + 395, + 257, + 422, + 264 + ], + "spans": [ + { + "bbox": [ + 395, + 257, + 422, + 264 + ], + "type": "inline_equation", + "content": "S = 0.56" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 372, + 265, + 446, + 323 + ], + "lines": [ + { + "bbox": [ + 372, + 265, + 446, + 323 + ], + "spans": [ + { + "bbox": [ + 372, + 265, + 446, + 323 + ], + "type": "image", + "image_path": "027e487ee3bed9a7f267bf14b86ff3fae0e4ac966fa8160fcda2f450ad6b7e5d.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 327, + 425, + 341 + ], + "lines": [ + { + "bbox": [ + 392, + 327, + 425, + 341 + ], + "spans": [ + { + "bbox": [ + 392, + 327, + 425, + 341 + ], + "type": "text", + "content": "A person lays on the ground" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 451, + 265, + 525, + 323 + ], + "blocks": [ + { + "bbox": [ + 474, + 257, + 501, + 264 + ], + "lines": [ + { + "bbox": [ + 474, + 257, + 501, + 264 + ], + "spans": [ + { + "bbox": [ + 474, + 257, + 501, + 264 + ], + "type": "inline_equation", + "content": "S = 0.52" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 451, + 265, + 525, + 323 + ], + "lines": [ + { + "bbox": [ + 451, + 265, + 525, + 323 + ], + "spans": [ + { + "bbox": [ + 451, + 265, + 525, + 323 + ], + "type": "image", + "image_path": "ffc9d2090b3d60d9023d6c26b88509a692df91c0a6d8c02e2567635b49f407fc.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 327, + 512, + 340 + ], + "lines": [ + { + "bbox": [ + 460, + 327, + 512, + 340 + ], + "spans": [ + { + "bbox": [ + 460, + 327, + 512, + 340 + ], + "type": "text", + "content": "A man crawls forward on his stomach" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 47, + 449, + 160, + 530 + ], + "blocks": [ + { + "bbox": [ + 46, + 346, + 547, + 434 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 547, + 434 + ], + "type": "text", + "content": "Figure 3. Qualitative retrieval results: We demonstrate example queries on the left, and corresponding retrieved motions on the right, ranked by text-motion similarity. The similarity values are displayed on the top. For each retrieved motion, we also show their accompanying ground-truth text label; note that we do not use these descriptions, but only provide them for analysis purposes. The motions from the gallery are all from the test set (unseen during training). In the first row, all top-5 retrieved motions correspond visually to 'playing violin' and the similarity scores are high " + }, + { + "bbox": [ + 46, + 346, + 547, + 434 + ], + "type": "inline_equation", + "content": ">0.80" + }, + { + "bbox": [ + 46, + 346, + 547, + 434 + ], + "type": "text", + "content": ". In the second row, we correctly retrieve the 'handstand' motion at top-1, but the other motions mainly perform 'cartwheel' (which involves shortly standing on hands), but with a lower similarity score " + }, + { + "bbox": [ + 46, + 346, + 547, + 434 + ], + "type": "inline_equation", + "content": "<0.70" + }, + { + "bbox": [ + 46, + 346, + 547, + 434 + ], + "type": "text", + "content": ". For the last example, we query a free-form text 'Someone is swimming', which does not exist in the gallery (but the word 'swim' does). The model successfully finds swimming motions among the top-3, and the other two motions involve the body parallel to the ground." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 449, + 160, + 530 + ], + "lines": [ + { + "bbox": [ + 47, + 449, + 160, + 530 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 160, + 530 + ], + "type": "image", + "image_path": "335b51424d89d8183ded2de2a693f99dd72e608153b97815ab28a6d08c710c08.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 161, + 450, + 263, + 530 + ], + "blocks": [ + { + "bbox": [ + 161, + 450, + 263, + 530 + ], + "lines": [ + { + "bbox": [ + 161, + 450, + 263, + 530 + ], + "spans": [ + { + "bbox": [ + 161, + 450, + 263, + 530 + ], + "type": "image", + "image_path": "ebf61b965f3627cbeeb951af27f629b49c414d8a813a4182e8c7875cfb206ede.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 48, + 532, + 160, + 621 + ], + "blocks": [ + { + "bbox": [ + 48, + 532, + 160, + 621 + ], + "lines": [ + { + "bbox": [ + 48, + 532, + 160, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 160, + 621 + ], + "type": "image", + "image_path": "8c0c7369183f4bd44168fe9db4c9952cce1984df268ab0362f3e12a0d8068122.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 624, + 287, + 712 + ], + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 712 + ], + "type": "text", + "content": "Figure 4. Moment retrieval: We plot the similarity between the temporally annotated BABEL text labels and the motions in a sliding window manner, and obtain a 1D signal over time (blue). We observe that a localization ability emerges from our model, even though it was not trained for temporal localization, and was not with the domain of BABEL labels. The ground-truth temporal span is denoted in green and the maximum similarity is marked with a dashed red line. More examples are provided in Appendix Figure A.2." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 161, + 533, + 263, + 613 + ], + "blocks": [ + { + "bbox": [ + 161, + 533, + 263, + 613 + ], + "lines": [ + { + "bbox": [ + 161, + 533, + 263, + 613 + ], + "spans": [ + { + "bbox": [ + 161, + 533, + 263, + 613 + ], + "type": "image", + "image_path": "07058cae226e7cb70e2c403d32635699fdce8eba1c108a2f931801816fea53df.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "bbox": [ + 306, + 441, + 382, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 382, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 382, + 453 + ], + "type": "text", + "content": "4.6. Limitations" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 460, + 547, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 664 + ], + "type": "text", + "content": "Our model comes with some limitations. Compared to the vast amount of data (e.g., 400M images [44]) in image-text collections used to train competitive foundation models, our motion-text training data can be considered extremely small (e.g., 23K motions in H3D). The generalization performance of motion retrieval models to in-the-wild motions is therefore limited. Data augmentations such as altering text can potentially help to a certain extent (cf. [6]); however, more motion capture is still needed. Another limitation concerns the case where one wishes to replace motion synthesis by retrieving a training motion. In this use case, the model requires the full encoded database (i.e., encoded training set) to be stored in memory, which can be memory-inefficient. Note that the encoded representation is considerably smaller than the original motion data. This could be dealt with using a hierarchical structuring of the motions and language descriptions (e.g. into upper-body motions, lower-body motions, sitting motions, standing motions, etc.)." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 306, + 669, + 375, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 375, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 375, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "In this paper, we focus on the relatively little-studied problem of motion retrieval with natural language queries. We" + } + ] + } + ], + "index": 55 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9495" + } + ] + } + ], + "index": 56 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "introduce TMR, a framework to jointly train text-to-motion retrieval and text-to-motion synthesis, with a special attention to the definition of negatives, taking into account the fine-grained nature of motion-language databases. We significantly improve over prior work, and provide a series of experiments highlighting the importance of each component." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 144, + 288, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 144, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 288, + 180 + ], + "type": "text", + "content": "Future work may consider incorporating a language synthesis branch, along with the motion synthesis branch, to build a symmetrical framework, which could bring further benefits." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 186, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 287, + 240 + ], + "type": "text", + "content": "Acknowledgements. This work was granted access to the HPC resources of IDRIS under the allocation 2022-AD011012129R2 made by GENCI. GV acknowledges the ANR project CorVis ANR-21-CE23-0003-01. The authors would like to thank Lucas Ventura and Charles Raude." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 242, + 287, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 242, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 287, + 263 + ], + "type": "text", + "content": "MJB Disclosure: files.is.tue.mpg.de/black/CoI_ ICCV_2023.txt" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 103, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 103, + 288 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 103, + 288 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 296, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 296, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 338 + ], + "type": "text", + "content": "[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 339, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 288, + 370 + ], + "type": "text", + "content": "[2] Chaitanya Ahuja and Louis-Philippe Morency. Language2Pose: Natural language grounded pose forecasting. In International Conference on 3D Vision (3DV), 2019. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 371, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 371, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 371, + 287, + 413 + ], + "type": "text", + "content": "[3] Méziane Aïder, Oussama Gacem, and Mhand Hifi. Branch and solve strategies-based algorithm for the quadratic multiple knapsack problem. Journal of the Operational Research Society, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 445 + ], + "type": "text", + "content": "[4] Okan Arikan, David A. Forsyth, and James F. O'Brien. Motion synthesis from annotations. ACM Transactions on Graphics (TOG), 2003. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 446, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 446, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 53, + 446, + 287, + 477 + ], + "type": "text", + "content": "[5] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gül Varol. TEACH: Temporal action composition for 3D humans. In International Conference on 3D Vision (3DV), 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 478, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 287, + 520 + ], + "type": "text", + "content": "[6] Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Güll Varol. SINC: Spatial composition of 3D human motions for simultaneous action generation. In International Conference on Computer Vision (ICCV), 2023. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 521, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 287, + 563 + ], + "type": "text", + "content": "[7] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In International Conference on Computer Vision (ICCV), 2021. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 563, + 287, + 595 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 563, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 53, + 563, + 287, + 595 + ], + "type": "text", + "content": "[8] Michael Böttner and Simon Clavet. Motion matching - the road to next gen animation. In Nucl.ai, youtu.be/watch?v=z_wpgHFSWss&t=658s, 2015.2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 596, + 287, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 596, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 53, + 596, + 287, + 638 + ], + "type": "text", + "content": "[9] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. In Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 638, + 287, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 638, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 48, + 638, + 287, + 681 + ], + "type": "text", + "content": "[10] Rishabh Dabral, Muhammad Hamza Mughal, Vladislav Golyanik, and Christian Theobalt. MoFusion: A framework for denoising-diffusion-based motion synthesis. In Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[11] Sivan Doveh, Assaf Arbelle, Sivan Harary, Rameswar Panda, Roei Herzig, Eli Schwartz, Donghyun Kim, Raja Giryes, Rogerio Feris, Shimon Ullman, and Leonid Karlinsky. Teaching struc" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 546, + 703 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 74, + 546, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 546, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 546, + 95 + ], + "type": "text", + "content": "tured vision & language concepts to vision & language models. In Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 95, + 546, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 546, + 126 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 546, + 126 + ], + "type": "text", + "content": "[12] Victor Escorcia, Mattia Soldan, Josef Sivic, Bernard Ghanem, and Bryan Russell. Temporal localization of moments in video collections with natural language. arXiv:1907.12763, 2019. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 127, + 546, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 127, + 546, + 158 + ], + "spans": [ + { + "bbox": [ + 308, + 127, + 546, + 158 + ], + "type": "text", + "content": "[13] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In International Conference on Computer Vision (ICCV), 2017. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 159, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 159, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 308, + 159, + 545, + 201 + ], + "type": "text", + "content": "[14] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "type": "text", + "content": "[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3D human motions from text. In Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 244, + 546, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 244, + 546, + 287 + ], + "spans": [ + { + "bbox": [ + 308, + 244, + 546, + 287 + ], + "type": "text", + "content": "[16] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2Motion: Conditioned generation of 3D human motions. In ACM International Conference on Multimedia (ACMMM), 2020. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 288, + 546, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 546, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 546, + 330 + ], + "type": "text", + "content": "[17] Ikhsanul Habibie, Mohamed Elgharib, Kripashindu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. In SIGGRAPH, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 331, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 361 + ], + "type": "text", + "content": "[18] R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping. In Computer Vision and Pattern Recognition (CVPR), 2006. 3, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 362, + 546, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 362, + 546, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 362, + 546, + 404 + ], + "type": "text", + "content": "[19] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In International Conference on Computer Vision (ICCV), 2017. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "type": "text", + "content": "[20] Daniel Holden, Oussama Kanoun, Maksym Perepichka, and Tiberiu Popa. Learned motion matching. ACM Transactions on Graphics (TOG), 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 437, + 545, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 468 + ], + "type": "text", + "content": "[21] Daniel Holden, Taku Komura, and Jun Saito. Phase-functional neural networks for character control. ACM Transactions on Graphics (TOG), 2017. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 468, + 546, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 546, + 510 + ], + "type": "text", + "content": "[22] Sai Shashank Kalakonda, Shubh Maheshwari, Sarvadevabhatla, and Ravi Kiran. Action-GPT: Leveraging large-scale language models for improved and generalized zero shot action generation. arXiv:2211.15603, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 511, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 532 + ], + "type": "text", + "content": "[23] Lucas Kovar, Michael Gleicher, and Frédéric H. Pighin. Motion graphs. ACM Transactions on Graphics (TOG), 2002. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 533, + 546, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 546, + 563 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 546, + 563 + ], + "type": "text", + "content": "[24] Taeryung Lee, Gyeongsik Moon, and Kyoung Mu Lee. MultiAct: Long-term 3D human motion generation from multiple action labels. In AAAI Conference on Artificial Intelligence, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 563, + 546, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 563, + 546, + 596 + ], + "spans": [ + { + "bbox": [ + 308, + 563, + 546, + 596 + ], + "type": "text", + "content": "[25] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. TVR: A large-scale dataset for video-subtitle moment retrieval. In European Conference on Computer Vision (ECCV), 2020. 2, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 596, + 546, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 596, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 308, + 596, + 546, + 639 + ], + "type": "text", + "content": "[26] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning (ICML), 2022. 2, 3, 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 639, + 546, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 639, + 546, + 681 + ], + "spans": [ + { + "bbox": [ + 308, + 639, + 546, + 681 + ], + "type": "text", + "content": "[27] Angela S Lin, Lemeng Wu, Rodolfo Corona, Kevin Tai, Qixing Huang, and Raymond J Mooney. Generating animated videos of human activities from natural language descriptions. ViGIL NeurIPS Workshop, 2018. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 681, + 546, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 546, + 703 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 546, + 703 + ], + "type": "text", + "content": "[28] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9496" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 705 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 83 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 83 + ], + "type": "text", + "content": "linear model. ACM Transactions on Graphics (TOG), 2015. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 83, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 83, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 287, + 115 + ], + "type": "text", + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 116, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 116, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 48, + 116, + 287, + 158 + ], + "type": "text", + "content": "[30] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval and captioning. Neurocomputing, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 159, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 159, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 159, + 287, + 201 + ], + "type": "text", + "content": "[31] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision (ICCV), 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 201, + 287, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 201, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 287, + 244 + ], + "type": "text", + "content": "[32] Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. Thinking fast and slow: Efficient text-to-visual retrieval with transformers. In Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 244, + 287, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 244, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 287, + 287 + ], + "type": "text", + "content": "[33] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 319 + ], + "type": "text", + "content": "[34] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv:1807.03748, 2018. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 319, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 361 + ], + "type": "text", + "content": "[35] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 361, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 404 + ], + "type": "text", + "content": "[36] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 404, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 426 + ], + "type": "text", + "content": "[37] Matthias Plappert, Christian Mandery, and Tamim Asfour. The KIT motion-language dataset. *Big Data*, 2016. 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "type": "text", + "content": "[38] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english labels. In Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 468, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 533 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "type": "text", + "content": "[40] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning (ICML), 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 576, + 287, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 576, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 48, + 576, + 287, + 619 + ], + "type": "text", + "content": "[41] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. In Transactions of the Association for Computational Linguistics (TACL), 2013. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 619, + 287, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 619, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 48, + 619, + 287, + 662 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 662, + 287, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 662, + 287, + 705 + ], + "spans": [ + { + "bbox": [ + 48, + 662, + 287, + 705 + ], + "type": "text", + "content": "[43] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. In Energy Efficient Machine Learning and Cognitive Computing NeurIPS Workshop, 2019. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 705 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[44] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: open dataset of clip-filtered 400 million image-text pairs. In Data Centric AI NeurIPS Workshop, 2021. 2, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 127, + 545, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 127, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 127, + 545, + 159 + ], + "type": "text", + "content": "[45] H. Sidenbladh, M. J. Black, and L. Sigal. Implicit probabilistic models of human motion for synthesis and tracking. In European Conference on Computer Vision (ECCV), 2002. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 159, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 159, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 545, + 201 + ], + "type": "text", + "content": "[46] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. MPNet: Masked and permuted pre-training for language understanding. In Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 201, + 545, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 201, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 307, + 201, + 545, + 245 + ], + "type": "text", + "content": "[47] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. MPNet: Masked and permuted pre-training for language understanding. In Neural Information Processing Systems (NeurIPS), 2020. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 277 + ], + "type": "text", + "content": "[48] Jiankai Sun, Bolei Zhou, Michael J. Black, and Arjun Chandrasekaran. LocATE: End-to-end localization of actions in 3D with transformers. arXiv:2203.10719, 2022. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "type": "text", + "content": "[49] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In International Conference on Learning Representations (ICLR), 2023. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "text", + "content": "[50] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Neural Information Processing Systems (NeurIPS), 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 363, + 545, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 545, + 395 + ], + "type": "text", + "content": "[51] Ruben Villegas, Jimei Yang, Duygu Ceylan, and Honglak Lee. Neural kinematic networks for unsupervised motion retargeting. In Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 395, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 395, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 395, + 545, + 437 + ], + "type": "text", + "content": "[52] Jiashun Wang, Chao Wen, Yanwei Fu, Haitao Lin, Tianyun Zou, Xiangyang Xue, and Yinda Zhang. Neural pose transfer by spatially adaptive instance normalization. In Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "text", + "content": "[53] Sijie Yan, Zhizhong Li, Yuanjun Xiong, Huahan Yan, and Dahua Lin. Convolutional sequence generation for skeleton-based action synthesis. In International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 480, + 545, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 480, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 307, + 480, + 545, + 523 + ], + "type": "text", + "content": "[54] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. CoCa: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research (TMLR), 2022. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "text", + "content": "[55] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In ICLR, 2022. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 566, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 566, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 545, + 609 + ], + "type": "text", + "content": "[56] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2M-GPT: Generating human motion from textual descriptions with discrete representations. In CVPR, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 609, + 545, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 609, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 307, + 609, + 545, + 651 + ], + "type": "text", + "content": "[57] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. MotionDiffuse: Text-driven human motion generation with diffusion model. arXiv:2208.15001, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 651, + 545, + 683 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 651, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 307, + 651, + 545, + 683 + ], + "type": "text", + "content": "[58] Yan Zhang, Michael J. Black, and Siyu Tang. Perpetual motion: Generating unbounded human motion. arXiv:2007.13886, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 683, + 545, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 683, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 307, + 683, + 545, + 705 + ], + "type": "text", + "content": "[59] Rui Zhao, Hui Su, and Qiang Ji. Bayesian adversarial human motion synthesis. In CVPR, 2020. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9497" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/2d762d0d-bad4-403d-98f7-4210b7f7e39d_content_list.json b/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/2d762d0d-bad4-403d-98f7-4210b7f7e39d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3593ccba11ccfdcb0a17d2148ea59d93adbefe97 --- /dev/null +++ b/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/2d762d0d-bad4-403d-98f7-4210b7f7e39d_content_list.json @@ -0,0 +1,2033 @@ +[ + { + "type": "text", + "text": "TORE: Token Reduction for Efficient Human Mesh Recovery with Transformer", + "text_level": 1, + "bbox": [ + 81, + 130, + 888, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiyang Dou $^{1\\dagger}$ Qingxuan Wu $^{2\\dagger}$ Cheng Lin $^{3\\dagger}$ Zeyu Cao $^{4\\dagger}$ Qiangqiang Wu $^{5}$ Weilin Wan $^{1}$ Taku Komura $^{1}$ Wenping Wang $^{6}$", + "bbox": [ + 148, + 178, + 813, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ The University of Hong Kong $^{2}$ University of Oxford $^{3}$ Tencent Games $^{4}$ University of Cambridge $^{5}$ City University of Hong Kong $^{6}$ Texas A&M University", + "bbox": [ + 137, + 227, + 831, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 299, + 313, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce a set of simple yet effective TOken REDuction (TORE) strategies for Transformer-based Human Mesh Recovery from monocular images. Current SOTA performance is achieved by Transformer-based structures. However, they suffer from high model complexity and computation cost caused by redundant tokens. We propose token reduction strategies based on two important aspects, i.e., the 3D geometry structure and 2D image feature, where we hierarchically recover the mesh geometry with priors from body structure and conduct token clustering to pass fewer but more discriminative image feature tokens to the Transformer. Our method massively reduces the number of tokens involved in high-complexity interactions in the Transformer. This leads to a significantly reduced computational cost while still achieving competitive or even higher accuracy in shape recovery. Extensive experiments across a wide range of benchmarks validate the superior effectiveness of the proposed method. We further demonstrate the generalizability of our method on hand mesh recovery. Visit our project page at https://frank-zy-dou.github.io/projects/Tore/index.html.", + "bbox": [ + 75, + 330, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 657, + 209, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human Mesh Recovery (HMR) has been extensively researched in recent years, given its wide real-world applications [15, 81, 11, 62, 69, 17, 42, 84]. This task becomes more challenging when the input is a monocular 2D image, due to the large pose and shape variation, large appearance variation, partial observation, and self-occlusion.", + "bbox": [ + 75, + 681, + 468, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There has been steady progress in 3D human mesh recovery [30, 50, 36, 24, 60, 39, 40, 8, 5]. Recently, Transformer [67] has shown state-of-the-art (SOTA) results on a wide variety of tasks due to its strong capability of capturing long-range dependency for more accurate predictions [3, 67, 76, 74]. Using tokens constructed from local features extracted by a convolutional neural network", + "bbox": [ + 75, + 773, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6fd3e346e3f7d408e93da1cbab0be16ccb81da2e55f3508c9e0ef783609ead4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 297, + 830, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d807dfac98eb535d4728e112321405cfcdf9915259009386c5d1749ad6f7f942.jpg", + "image_caption": [ + "Figure 1. Throughput v.s. Accuracy and GFLOPs v.s. Accuracy on Human3.6M [23]. Our method dramatically saves GFLOPs and improves throughput while maintaining highly competitive accuracy. The x-axis of the bottom GFLOPs figure is reversed for demonstration. Eb0 and R50 represent EfficientNet-b0 [65] and ResNet-50 [20] backbones, respectively." + ], + "image_footnote": [], + "bbox": [ + 537, + 398, + 830, + 493 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/1eef7f6686a2c120120c2c9359cd220b2129d59c5af7a977532bf1868fb3889b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
● METRO baseline ▲ FastMETRO baseline ● FastMETRO(S) baseline
METRO-TORE ▼ FastMETRO-TORE ● FastMETRO(S)-TORE
", + "bbox": [ + 504, + 502, + 885, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(CNN) [20, 70] to query the joint and mesh vertex positions, Transformer-based methods [39, 40, 8] achieved SOTA performance.", + "bbox": [ + 496, + 611, + 890, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, improved performance comes with costs: the increased expressivity of Transformers comes with quadratically increasing computational costs since all pairwise interactions are taken into account [14]. The space and time complexity of a QKV-attention operation is known to be $O(M^2)$ , where $M$ is the number of tokens. The token number thus plays a vital role in time efficiency: a large number of tokens inevitably leads to a heavy computation burden. Unfortunately, almost all the existing Transformer-based SOTA methods for HMR [39, 40, 8] are suffering from redundant tokens. This incurs a high model complexity and computational costs, which prevents the current Transformer-based HMR methods from achieving their full potential in real-world applications.", + "bbox": [ + 496, + 657, + 892, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we make key observations from two important aspects: the 3D geometry structure and 2D image", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger ,\\ddagger$ denote equal contributions.", + "bbox": [ + 101, + 887, + 267, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15143", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "feature, to reveal the problem of token redundancy. First, to recover the 3D geometry, all existing methods use both mesh vertices and skeleton joints as tokens for the feature interaction between input and body shape. Whereas a body mesh contains numerous vertices, they can be abstracted by a small number of joints on the skeleton. For instance, when animating a SMPL [44] avatar, the skeleton joints, together with a blend shape binding the joints and corresponding mesh vertices of the local body part, are able to describe various body meshes. Therefore, the joints can already be viewed as an underlying structure of a body shape, which intrinsically encodes the human mesh geometry. Second, for image-based input, most existing methods indiscriminately use all the feature patches to capture pose, shape and appearance variance. However, although the human body exhibits large variance, the important features for shape inference are dominantly clustered within the body area in an RGB image. Most features, e.g., image background, are not informative, thus bringing about redundancy.", + "bbox": [ + 76, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the aforementioned insights, we argue that the Pareto-front of accuracy and efficiency for Transformer-based HMR could be further improved by reducing the number of tokens [59, 47, 56]. To this end, we introduce a set of simple yet effective token reduction strategies mainly from two aspects corresponding to our observations. First, for 3D mesh recovery, instead of querying both vertices and joints with input features simultaneously, we consider learning a small set of body tokens at the skeleton level for each body part. To recover corresponding mesh vertices, we use an efficient Neural Shape Regressor (NSR) to infer the mesh from the body features encoded by these tokens. This query process can also be interpreted as an attention matrix decomposition, by which we effectively leverage the geometric insights encoded at the skeleton level to infer the mesh structure hierarchically. Second, for the input image feature, we introduce a learnable token pruner to prune the tokens of patch-based features extracted by a CNN. We employ a clustering-based strategy to identify discriminative features, which results in two appealing properties: 1) the end-to-end learning of the pruner is unsupervised, avoiding the need for additional data labeling; 2) it learns semantically consistent features across various images, thus further benefiting the geometry reasoning and enhancing the capability of generalizability. These token reduction strategies substantially reduce the number of query tokens involved in the computation without sacrificing the important information. An overview is shown in Figure 2.", + "bbox": [ + 76, + 383, + 472, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments across wide benchmarks [23, 68, 85], including both the human body and hand mesh recovery, to validate the proposed method. Compared to SOTA methods, our framework faithfully recovers body meshes with fewer tokens, which considerably reduces memory and computation overhead while maintain", + "bbox": [ + 76, + 810, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing competitive geometric accuracy.", + "bbox": [ + 500, + 92, + 740, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contribution is three-fold:", + "bbox": [ + 517, + 107, + 803, + 121 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reveal the issues of token redundancy in the existing Transformer-based methods for HMR.", + "- We propose effective strategies for token reduction by incorporating the insights from the 3D geometry structure and 2D image feature into the Transformer design.", + "- Our method achieves SOTA performance on various benchmarks with less computation cost. For instance, for the Transformer Encoder structure [39] and the Transformer Encoder-Decoder structure [8] with ResNet-50 [20] backbone, our method maintains competitive accuracy while saving $82.9\\%$ , $50.5\\%$ GFLOPs and improving $139.1\\%$ , $39.8\\%$ throughput, respectively; see Figure 1 for an overview." + ], + "bbox": [ + 517, + 132, + 893, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 361, + 640, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Mesh Recovery (HMR) from monocular images has achieved great progress in the past years [4, 30, 50, 36, 10, 40, 24, 60, 5, 39, 8, 82, 71, 58]. Given a monocular RGB image, the goal of HMR is to recover the 3D body shape, typically using a human body model, e.g., SMPL [44]. We refer readers to [66] for a comprehensive review. Existing methods fall into two categories: Non-Transformer-based and Transformer-based approaches.", + "bbox": [ + 496, + 387, + 893, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-Transformer-based Human Mesh Recovery To recover the human body shape, one could predict a few parameters, i.e., joint rotation and body shape, to drive the parametric model (parametric approach) or directly regress the vertex positions of the body (non-parametric approach). Parametric approaches SMPLify [4] estimates human pose and shape by fitting SMPL to the detected 2D key points [55] by optimization. Kanazawa et al. [24] adopt adversarial prior knowledge of the 3D body shape into the neural network for HMR. SPIN [30] then combines regression-based and optimization-based methods during the training loop. Intermediate features such as key points [54], pixel-to-surface correspondences [77] and texture consistency [52], have been exploited. HybrIK [36] proposes a hybrid inverse kinematics via twist-and-swing decomposition [2]. Regressing the body parameters from the image is a highly non-linear mapping [50], which thus limits the performance of parametric-based approaches.", + "bbox": [ + 496, + 523, + 893, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-parametric approaches Non-parametric approaches [50, 32, 9] recover spatial coordinates of a body shape directly from the image features. I2L-MeshNet [50] predicts the per-pixel likelihood on 1D heatmaps for vertex coordinates for a trade-off between accuracy and computational cost. GraphCMR [32] and Pose2Mesh [9] employ graph convolutional neural network (GCNN) [27]", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15144", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3214ac64bb94513573c69970a13aa59ff493cbc8bc8344123a8780d69e2899da.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed framework. Our goal is to reduce tokens for Transformer Encoder and Decoder which are critical modules in the whole pipeline. Image Token Pruner (ITP) and Neural Shape Regressor (NSR) are two lightweight components." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 861, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to regress mesh positions by modeling local interactions among vertices from the image [32] and 2D pose [9], respectively. Although steady progress has been made, SOTA accuracy is produced by Transformer-based methods.", + "bbox": [ + 75, + 338, + 468, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer-based Human Mesh Recovery Recently, Transformer structures have been successfully adopted in various vision tasks [12, 43, 6, 51, 66, 80]. In HMR, the SOTA performance is achieved by Transformer-based (attention mechanism) approaches [39, 29, 40, 8]. Specifically, METRO [39] recovers body mesh using a Transformer Encoder [67] to model vertex-vertex and vertex-joint interactions conducting dimension reduction from the image features to 3D shape. MeshGraphormer [40] injects Graph Convolutions [27] into the Transformer Encoder [39] blocks to improve local interactions among vertices. The recent work FastMETRO [8] employs a Transformer Encoder-Decoder architecture disentangling the image encoding and mesh estimation for model reduction and acceleration. However, all Transformer-based approaches [39, 40, 8] suffer from redundant tokens, causing heavy interaction between body geometry and image features, which is still cumbersome and computationally intensive. In this paper, we design strategies driven by the 3D body geometry and 2D image feature for token reduction for both Transformer Encoder and Transformer Encoder-Decoder structures.", + "bbox": [ + 75, + 398, + 470, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Token Reduction for Transformers Token reduction for Transformer structures has been causing people's attention nowadays [56, 47, 59, 78, 16, 38, 46, 7]. In particular, DynamicViT [56] reduces less informative tokens hierarchically with supervision from a teacher network to save computational costs during inference. TokenLearner [59] generates a smaller number of tokens with spatial attention adaptively. Patch-to-Cluster [16] proposes to cluster over image patches with a token selector. Our method follows this way and adopts a token selection layer for clustering over input tokens, reducing tokens and improving general", + "bbox": [ + 75, + 731, + 468, + 896 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ization capability in challenging scenarios. Token pruning techniques have also been explored in single-view 2D pose estimation and multi-view 3D pose estimation [47] by selecting tokens based on the attention value in the Transformer. Differently, in HMR, besides the key points, body geometry, i.e., mesh vertices, also introduces a burden to the Transformer and needs to be carefully considered.", + "bbox": [ + 496, + 338, + 892, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 459, + 635, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ the popular body parametric model SMPL [44] to represent a human body. Formally, given a monocular image $\\mathcal{I} \\in \\mathbb{R}^{224 \\times 224}$ , our goal is to recover the joint positions $\\mathcal{J} \\in \\mathbb{R}^{J \\times 3}$ and mesh vertices $\\mathcal{V} \\in \\mathbb{R}^{V \\times 3}$ where $J = 14$ , $V = 6890$ are the numbers of joints and the mesh vertices, respectively. In addition to the human body, we also demonstrate our method on hand mesh recovery with MANO model [57] to show its generalizability.", + "bbox": [ + 496, + 484, + 890, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing Transformer-based HMR methods fall into two categories: a single Transformer Encoder structure [39, 40] and a Transformer Encoder-Decoder structure [8]. Accordingly, we propose two strategies to reduce tokens from 3D and 2D levels, namely, Geometry Token Reduction (GTR) and Image Token Pruning (ITP), which can accommodate different Transformer structures to improve their efficiency.", + "bbox": [ + 496, + 607, + 890, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Geometry Token Reduction (GTR)", + "text_level": 1, + "bbox": [ + 500, + 724, + 803, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As an essential module, all the existing Transformer-based methods for HMR [39, 40, 8] have to model the interaction between body geometry and image features with an attention mechanism. Therefore, an effective token reduction method to improve the efficiency for inferring mesh geometry can benefit the design of Transformer based methods in a wide scope.", + "bbox": [ + 496, + 748, + 890, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We observe existing methods invariably use both skeleton joints and mesh vertices as tokens for regressing their spatial coordinates. The numerous tokens inevitably lead to", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15145", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "significant computational costs for a Transformer. In fact, as an underlying structure of the human body, the skeleton joints already provide strong priors for perceiving the body geometry. A human body avatar, like SMPL [44], can be driven by a small number of joints with the blend shape. This inspires us to regress the human body mesh in a hierarchical manner.", + "bbox": [ + 75, + 90, + 467, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our key idea is to decompose a heavy Transformer into smaller Transformer modules, each of which involves fewer tokens, avoiding expensive computation for image feature interaction. Specifically, our approach only queries the body tokens, whose count is equivalent to the number of joints. Since the number of joints is much less than the number of mesh vertices, we can significantly reduce the token number in complex 2D-3D feature interactions and this leads to a substantial enhancement in efficiency.", + "bbox": [ + 75, + 196, + 467, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To faithfully recover body shape from the body tokens, we introduce a Neural Shape Regressor (NSR) module, which is a light Transformer structure, to query the vertex tokens from their interaction with body features. Let $F_{J} = \\{f_{1}^{j}, f_{2}^{j}, \\dots, f_{J}^{j}\\}$ denotes a set of learned body features and $T_{\\mathcal{V}} = \\{t_{1}^{v}, t_{2}^{v}, \\dots, t_{V}^{v}\\}$ where $t_{i}^{v}$ represents a set of vertex tokens which are the query tokens. We conduct cross-attention between query vertex tokens and body features to model the interaction between vertices and learned body features. Note that non-adjacent vertices are masked to improve the efficiency [8].", + "bbox": [ + 75, + 334, + 467, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The feature size within NSR is even smaller compared with the main Transformer. Thanks to the informative geometric feature encoded in the body tokens, despite not incorporating image tokens, NSR still faithfully recovers the surface vertices. Another interesting discovery is that the learned attention scores reflect the correlation between joints and vertices at a body part level, similar to body blending weights (see Sec. 4.7.3). In this way, we can effectively reduce the cost of redundant interaction by decomposing the entire body into encoded constituent parts and efficiently recovering the whole body shape.", + "bbox": [ + 75, + 501, + 467, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Image Token Pruning (ITP)", + "text_level": 1, + "bbox": [ + 76, + 679, + 326, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the geometric insights, for a Transformer Encoder-Decoder structure, image tokens also affect the computation overhead. However, the existing methods adopting this Transformer structure fail to avoid the token redundancy issue [8] because all available image feature patches are indiscriminately involved in human mesh recovery. Actually, some features, e.g., image background, are not informative, thus introducing redundancy and increasing the computational cost of the Transformer.", + "bbox": [ + 75, + 703, + 467, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To tackle this issue, we introduce an effective token pruning strategy, namely, Image Token Pruning (ITP). Our key insight is that the informative features in an image for inferring 3D geometry are overwhelmingly clustered within", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the region of the human body. Inspired by recent advances in token pruning [16, 59, 56, 47], we propose to aggregate features into a small number of meaningful clusters.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $0 < \\rho < 1$ denote the predefined token pruning ratio. Our goal is to learn a projection that maps feature map $F_{\\mathcal{I}} \\in R^{HW \\times c}$ extracted from the given image with HW tokens to $Z_{\\mathcal{I}} \\in R^{T \\times c}$ with $T$ tokens, where $T = \\lfloor \\rho HW \\rfloor$ and $c$ is the feature dimension. The small number of clusters $Z_{\\mathcal{I}}$ are expected to capture the fewer but more discriminative features in $F_{\\mathcal{I}}$ . Inspired by [16, 59], we implement the projection as a learnable CNN module. We first apply Conv2D to the input feature, where the kernel size, stride and zero padding of Conv2D are $3 \\times 3$ , 1 and 1, reducing the feature dimension from $c$ to $c' = c/4$ with GELU [21] activation function applied. We further map the dimension $c$ to $N$ by MLP and Softmax, producing $M \\in R^{HW \\times N}$ . Finally, the clustered token set is given by $Z_{\\mathcal{I}} = M^T \\cdot F_{\\mathcal{I}}$ . The mapping matrix $M^T$ , in essence, produces a clustering over origin tokens. Each element $m_{ij} \\in M^T$ depicts the contribution of $j$ token to $i$ -th cluster. Note that LayerNorm is employed after the clustering.", + "bbox": [ + 496, + 136, + 890, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Token Reduction Supervision In order to encourage ITP to pay more attention to those discriminative feature regions, i.e., body parts. during token reduction, we introduce Token Reduction Supervision. Given the weak-perspective camera estimated the neural network, we compute the 2D projection of the ground truth 3D vertices $\\hat{\\mathcal{V}}$ as $s\\Pi (\\hat{\\mathcal{V}}) + t$ where $\\Pi$ represents an orthographic projection, $s$ and $t$ are scale and translation estimated by the network. The projected results are downsampled to a discrete $H\\times W$ grid corresponding to each token followed by a binary indicator function", + "bbox": [ + 496, + 407, + 890, + 559 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {d} (x) = \\left\\{ \\begin{array}{l} 1, \\text {i f t h e c e l l c o n t a i n s a p r o j e c t e d p o i n t .} \\\\ 0, \\text {i f t h e c e l l d o e s n o t c o n t a i n a p r o j e c t e d p o i n t .} \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 571, + 888, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then the supervision for token pruning is given by", + "bbox": [ + 500, + 625, + 834, + 640 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {P} = - \\frac {1}{N H W} \\sum_ {i} ^ {N} \\left(F _ {d} \\left(s \\Pi (\\hat {\\mathcal {V}}) + t\\right) \\cdot M _ {[:, i ]}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 651, + 888, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this supervision, the weights of background tokens in $M[:i]$ for each $i$ -th cluster will be penalized during training, which thus encourages the ITP to learn the discriminative features.", + "bbox": [ + 496, + 703, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed ITP has several noticeable advantages. First, in contrast to other methods relying on explicit supervision, e.g., using a pretrained teacher network to facilitate the pruning [56, 78, 33], our token pruner is trained in an unsupervised manner; Note that no ground truth masking is required for Token Reduction Supervision. Second, unlike previous clustering-based pruning methods [16] that use pre-defined class labels, ITP aims to adaptively identify discriminative features within the body region. Therefore,", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15146", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ITP is not limited to fixed semantic labels but is able to induce higher-level semantics according to the learning target.", + "bbox": [ + 75, + 90, + 467, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we find the clustered informative features learned by ITP improve the model generalizability, especially for challenging datasets, e.g., in-the-wild dataset 3DPW [68] as shown in Table 3. We elaborate on these properties in Sec. 4.3.2 and Sec. 4.7.1 with extensive experiments.", + "bbox": [ + 75, + 122, + 467, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Loss Functions", + "text_level": 1, + "bbox": [ + 76, + 226, + 228, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We supervise the network using L1 distance between predicted mesh vertices at three sampling levels: $L_{\\mathcal{V}3D} = ||\\mathcal{V}_{3D}^l - \\hat{\\mathcal{V}}^l||_1 + ||\\mathcal{V}_{3D}^m - \\hat{\\mathcal{V}}^m||_1 + ||\\mathcal{V}_{3D}^h - \\hat{\\mathcal{V}}_h||_1$ , where $*, *^m, *^h$ denote low, middle and high body mesh resolution with vertex numbers to be 431, 1723 and 6890 for an SMPL body. The ground-truth 3D joints are used for supervising predicted 3D joints and the ones regressed from the vertices $\\mathcal{V}_{3D}^h$ with a SMPL regression matrix $\\mathcal{M}$ : $L_{\\mathcal{J}3D} = ||\\mathcal{J}_{3D} - \\hat{\\mathcal{J}}_{3D}||_1, L_{\\mathcal{J}3D}^R = ||\\mathcal{M}(\\mathcal{V}) - \\hat{\\mathcal{J}}_{3D}||_1$ . A 2D projection loss $L_{J2D}$ is used during the network training, where we employ a weak perspective camera model to project 3D joints to 2D for supervision: $L_{\\mathcal{J}2D}^R = ||(s\\Pi(\\mathcal{M}(\\mathcal{V})) + t) - \\hat{\\mathcal{J}}_{2D}||_1$ . $s, t$ are scale and translation estimated by the network. Overall, together with the Token Pruning Supervision $L_P$ , the total loss is:", + "bbox": [ + 75, + 250, + 467, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L = \\alpha \\left[ \\lambda_ {\\mathcal {J} 3 D} \\left(L _ {\\mathcal {J} 3 D} ^ {R} + L _ {\\mathcal {J} 3 D}\\right) + \\lambda_ {\\mathcal {V} 3 D} \\left(L _ {\\mathcal {V} 3 D}\\right) + \\lambda_ {P} L _ {P} \\right] \\\\ + \\beta \\lambda_ {\\mathcal {J} 2 D} L _ {\\mathcal {J} 2 D} ^ {R}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 80, + 494, + 464, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ indicate the availability of the supervision. We set $\\lambda_{P},\\lambda_{J2D},\\lambda_{V3D},\\lambda_{J3D}$ to be 1, 100, 100, 1000.", + "bbox": [ + 75, + 545, + 467, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 593, + 281, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 76, + 619, + 274, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our model in two scenarios: human body mesh recovery and hand mesh recovery. We adopt commonly-used metrics [25, 31, 39, 40, 8]: Mean Per-Joint Position Error (MPJPE), MPJPE after further alignment, i.e., Procrustes Analysis (PAMPJPE) and Mean Per-Vertex Error (MPVE). For human body mesh recovery, our network is trained with Human3.6M [23], MuCo-3DHP[48], UP-3D [34], COCO [41], MPI [1]. Following previous works [39, 40, 8], we use the pseudo mesh data in Human3.6M [23] for training, splitting subjects S1, S5, S6, S7, S8 for training and S9, S11 for testing. We also report our performance on 3DPW [68], a more challenging in-the-wild dataset. To further evaluate the generalization ability of the proposed method, we test our method on hand mesh recovery on FreiHAND [85] dataset. For analyzing the efficiency of our method, we report GFLOPs and throughput (image per second), strictly following [56, 49, 59, 47, 38].", + "bbox": [ + 75, + 643, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6d6a85bd44eae7799d565e87ca85e08e185db8f6d2cffb3c246040a24710d867.jpg", + "table_caption": [ + "Table 1. Comparison with the Transformer Encoder structure METRO (M) [39] on Human3.6M [23]. We test with ResNet-50 (R50) [20] and HRNet-W64 (H64) [70] as backbones. GFLOPs $^{\\mathrm{T}}$ is GFLOPs of the transformer." + ], + "table_footnote": [], + "table_body": "
MethodGFLOPs ↓GFLOPsT ↓Throughput (im/s) ↑PAMPJPE ↓
M-H64 [39]56.527.5141.036.7
M-H64+GTR30.2 (-46.5%)0.8 (-97.1%)210.1 (+49.0%)37.1 (+1.1%)
M-R50 [39]31.627.5247.040.6
M-R50+GTR5.4 (-82.9%)0.8 (-97.1%)590.6 (+139.1%)42.0 (+3.4%)
", + "bbox": [ + 506, + 152, + 890, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 251, + 715, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Training We implement the network using PyTorch. For the Transformer Encoder-Decoder structure, we set the learning rate to be $1 \\times 10^{-4}$ . We use the AdamW optimizer [45] and train for 60 epochs, with a batch size of 16 per GPU on 4 Nvidia A100 GPUs. When comparing with Transformer Encoder structure METRO [39] (see Table 1), we follow the setting of METRO where we train the models with a batch size of 30 per GPU on 8 Nvidia A100 GPUs in total. We adopt Adam optimizer [26] and train the models for 200 epochs. See more details in Appendix A.", + "bbox": [ + 496, + 273, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Performance Evaluation To analyze the performance on a consumer-level GPU device, we measure the throughput on an NVIDIA RTX 3090 GPU with 24G VRAM. When comparing with encoder-decoder Transformer structure FastMETRO [8], we set the batch size to 32 following PPT [47]. The comparison with Transformer Encoder structure METRO [39] uses 16 as the batch size. Note that the batch size is limited by the size of the original METRO [39] model. To factor out the influence of the batch size, we provide a more detailed performance report in Appendix B.", + "bbox": [ + 496, + 426, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Performance on Human Mesh Recovery", + "text_level": 1, + "bbox": [ + 500, + 589, + 841, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Currently, there only exist two types of representative Transformer structures for HMR: Encoder-only structure (METRO [39]) and Encoder-Decoder structure (Fast-METRO [8]). To demonstrate the effectiveness of our method to both Transformer structures, we first conduct experiments by adding GTR to two structures: METRO and FastMETRO. Since ITP is designed for the encoder-decoder structure, we further report the results on FastMETRO with both ITP and GTR in Sec. 4.3.2. The overall comparison is shown in Table 5.", + "bbox": [ + 496, + 612, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3.1 Transformer Encoder Structure", + "text_level": 1, + "bbox": [ + 500, + 785, + 774, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Table 1, GTR effectively reduces the computation costs while still producing competitive accuracy results. For the encoder-based Transformer model METRO [39] with HRNet-W64 [70] as a backbone, applying GTR saves the GFLOPs of the whole model for $46.5\\%$ and the Transformer part for $97.1\\%$ , with throughput im", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15147", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c72ac0f1fe60fbb90346be7b9d8e80d2a7426fad6767b10662b67670932e94e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 88, + 274, + 164 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6676cc78dad18567d6712dc6dac93b1e8c5de2febd20136c471f9af99917d57b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 88, + 470, + 164 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/937cc42f1d112f837be254d8251447a65fbf875c0f437235c9f5adf65b2a1ec5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 165, + 274, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/53f6f812e9fb9899cd68417a64be2476d1903201d6bca5d307f192fe04706e42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 165, + 470, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b85f6a8c29ca2a1d742d96eddacd23191c9a2bcf11861871627f1d09e94b8d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 241, + 274, + 315 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6bc09f241d8d32243e21f229da1894d05931658d2b3be1f676eafe9c0919f6f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 241, + 470, + 314 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d50d40c956529558d6af8c3f6a4b5913652bba5873391472df7dbeb5aceb14c5.jpg", + "image_caption": [ + "Figure 3. Qualitative results of GTR equipped Encoder-Decoder structure [39] (H64) on Human3.6M [23] and 3DPW [68]." + ], + "image_footnote": [], + "bbox": [ + 80, + 316, + 274, + 392 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3cc684d60fb5b007ce93d3b584e9d23cf924864c11a0b5456af91da231c16a6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 315, + 470, + 392 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d2e137caa4342ffd2d2b2942f9b039651e0ebea5d575d2218a609771dd560823.jpg", + "table_caption": [ + "Table 2. Comparison with the Transformer Encoder-Decoder structure FastMETRO (FM) [8] on Human3.6M [23]. We test with EfficientNet-b0 (Eb0) [65], ResNet-50 (R50) [20] and HRNet-W64 (H64) [70] as backbones. GFLOPs $^{\\mathrm{T}}$ stand for GFLOPs for the transformer." + ], + "table_footnote": [], + "table_body": "
MethodGFLOPs ↓GFLOPsT ↓Throughput (im/s) ↑PAMPJPE ↓
FM-H64 [8]35.76.6221.533.7
FM-H64+GTR30.2 (-15.4%)0.7 (-89.4%)249.2 (+12.5%)34.8 (+3.2%)
FM-R50 [8]10.96.6576.037.3
FM-R50+GTR5.4 (-50.5%)0.7 (-89.4%)805.3 (+39.8%)38.6 (+3.4%)
FM(S)-R50 [8]6.42.2953.539.4
FM(S)-R50+GTR4.6 (-28.1%)0.3 (-86.4%)1128.9 (+18.4%)38.6 (-2.0%)
FM-Eb07.16.6517.645.8
FM-Eb0+GTR1.7 (-76.1%)0.7 (-89.4%)870.5 (+68.2%)44.2 (-3.5%)
", + "bbox": [ + 83, + 515, + 470, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "proved by $49\\%$ . For the ResNet-50 [20] backbone, GTR helps save the GFLOPs of the whole model for $82.9\\%$ and the Transformer part for $97.1\\%$ , with throughput improved by $139.1\\%$ . These results validate the effectiveness of the proposed GTR on the Transformer Encoder structure. Qualitative results of METRO-H64+GTR can be found in Figure 3, in which the model produces high-quality results in human mesh recovery over various input monocular images.", + "bbox": [ + 75, + 655, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2 Transformer Encoder-Decoder Structure", + "text_level": 1, + "bbox": [ + 76, + 815, + 416, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We experiment on the Transformer Encoder-Decoder structure FastMETRO [8] with its two variants: models with 1 and 3 Encoder-Decoder layers. For 1-layer FastMETRO, we denote it as FastMETRO(S). In Table 2, GTR equipped", + "bbox": [ + 76, + 839, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FastMETRO [39] produces competitive accuracy while reducing $15.4\\%$ in GFLOPs for the whole model, $89.4\\%$ in GFLOPs for the Transformer part and improving $12.5\\%$ in throughput with an HRNet-W64 [70] CNN backbone. When testing with ResNet-50 [20], GTR helps save $50.5\\%$ in GFLOPs for the whole model, $89.4\\%$ in GFLOPs for the Transformer part and improves $39.8\\%$ in throughput. Notably, our FastMETRO(S)-R50+GTR and FastMETRO(S)-Eb0+GTR yield 38.6, 44.2 mm PAMPJPE respectively, surpassing the corresponding baselines while greatly saving computational costs.", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then present the performance of the Transformer Encoder-Decoder structure with GTR and ITP using ResNet-50 [20] and EfficientNet-b0 [65] CNN backbones.", + "bbox": [ + 496, + 257, + 890, + 303 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9a18612dadff735d11b3e67cf87e7a70b4ad10c99e45a26c82cac4f1b2469058.jpg", + "table_caption": [ + "Table 3. Influence of ITP for monocular 3D human mesh recovery on 3DPW [68]." + ], + "table_footnote": [], + "table_body": "
MethodMPVEMPJPEPAMPJPE
FastMETRO-H64+GTR91.375.446.7
FastMETRO-H64+GTR+ITP@20%88.272.344.4
", + "bbox": [ + 506, + 353, + 890, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 4, TORE (GTR+ITP) is effective for the Encoder-Decoder Transformer structure. Specifically, when pruning with $20\\%$ , $50\\%$ in ITP, the models save the GLOPs of the Transformer structure by $14.3\\%$ , $14.3\\%$ and $28.5\\%$ , $42.9\\%$ while receiving competitive accuracy or even higher accuracy i.e., FM-Eb0+GTR+ITP@ $20\\%$ improves PAMPJPE compared with the baseline.", + "bbox": [ + 496, + 422, + 890, + 527 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fc02a879db89a21c14b13db460cfda2b6199276b61914f13dd19f61e3f241e76.jpg", + "table_caption": [ + "Table 5. Comparison with the SOTA methods for monocular 3D human mesh recovery on 3DPW [68] and Human3.6M [23]." + ], + "table_footnote": [], + "table_body": "
Method3DPWHuman3.6M
MPVEMPJPEPAMPJPEMPJPEPAMPJPE
HMR [25]-130.076.788.056.8
GraphCMR [32]--70.2-50.1
SPIN [30]116.496.959.262.541.1
I2LMeshNet [50]-93.257.755.741.1
PyMAF [83]110.192.858.957.740.5
ROMP-R50 [64]105.689.353.5--
PARE-R50 [29]99.782.952.3--
DSR-R50 [13]99.585.751.760.940.3
METRO-R50 [39]---56.540.6
METRO-H64 [39]88.277.147.954.036.7
METRO-H64+GTR87.975.546.657.637.1
FastMETRO-R50 [8]90.677.948.353.937.3
FastMETRO-R50+GTR+ITP@20%99.282.452.359.840.5
FastMETRO-H64 [8]84.173.544.652.233.7
FastMETRO-H64+GTR+ITP@20%88.272.344.459.636.4
FastMETRO-Eb0 [8]112.593.860.269.245.8
FastMETRO-Eb0+GTR+ITP@20%112.593.760.163.243.9
", + "bbox": [ + 504, + 579, + 895, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we find that ITP helps improve the accuracy on the challenging in-the-wild 3DPW [68] dataset, shown in Table 3. Specifically, when further equipped with ITP, the model performance in MPVE, MPJPE and PAMPJPE are improved by $3.1\\mathrm{mm}$ , $3.1\\mathrm{mm}$ and $2.3\\mathrm{mm}$ , respectively.", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15148", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fe5b946d16d80ad2af1b368e6289b67549f3330c6f87a49ec7226d0f1d86de46.jpg", + "table_caption": [ + "Table 4. Statistics of all proposed components (GTR + ITP) for Encoder-Decoder structure FastMETRO (FM) [8] on Human3.6M [23]. The backbones are EfficientNet-b0 (Eb0) [65] and ResNet-50 (R50) [20]. GFLOPs $^{\\mathrm{T}}$ stands for GFLOPs of the Transformer." + ], + "table_footnote": [], + "table_body": "
Method + Pruning Rate#TokensGFLOPs ↓GFLOPsT ↓Throughput (im/s) ↑PAMPJPE ↓
FM-R50+GTR495.40.7805.338.6
FM-R50+GTR+ITP@20%395.3 (-1.8%)0.6 (-14.3%)794.7 (-1.3%)40.5 (+4.9%)
FM-R50+GTR+ITP@50%245.1 (-5.5%)0.5 (-28.5%)806.1 (+0.1%)40.7 (+5.4%)
FM-Eb0+GTR491.70.7870.544.2
FM-Eb0+GTR+ITP@20%391.6 (-5.9%)0.6 (-14.3%)876.3 (+0.7%)43.9 (-0.%)
FM-Eb0+GTR+ITP@50%241.4 (-17.6%)0.4 (-42.9%)870.4 (-0.1%)44.0 (-0.5%)
", + "bbox": [ + 84, + 125, + 885, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd7d259b3a4cc5ec1d2550a7f956935ea521df3a69c821beeb1b6abe86c3bb36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 242, + 281, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/447312e7e09283642e1ed1d103858963b70c2f4e84fddf848c121f9c91832c5e.jpg", + "image_caption": [ + "Figure 4. Qualitative results of FastMETRO+GTP+ITP@20% on Human3.6M [23] and 3DPW [68]." + ], + "image_footnote": [], + "bbox": [ + 80, + 321, + 279, + 401 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/026c783d04408d6546c7bc6e26b8cabce8c3bd2f498696141238b3e6398cdba6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 242, + 483, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/946903e54861b1c25d1239531e680fc684e787f6db8bfdb2a85e6e961fa79315.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 321, + 483, + 401 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/839ec8ee3ca390e05b2e3871156819dcdc0e886108d9453e28aa494b2cd44e92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 242, + 686, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11554b0651fe67d123710420d241da859bf90726756bd50d4403c74cf77eb6ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 321, + 686, + 401 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dbcbfe70555b0bead4b216f4abc7ad20b29cadf163692c197ceb1108cd237f10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 242, + 890, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd2865b51f78814e662d1217c64f530a25bcd5ed8852928b0dc9177b8c37ae76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 321, + 890, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These results indicate that the ITP module learns more discriminative features and thus enhances the capability of generalization, allowing methods with ITP to achieve better performance in more challenging in-the-wild scenarios. The strong representation capability of ITP can also be seen in the clustering process, which promotes competitive accuracy with fewer tokens while producing semantically meaningful clustering results, as demonstrated in Sec. 4.7.1.", + "bbox": [ + 75, + 444, + 468, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct comparisons with existing SOTA approaches (both Transformer and Non-Transformer based) on monocular human mesh recovery. The results are summarized in Table 5. It is notable that, with effective token reduction techniques, our method produces competitive or higher accuracy on Human3.6M [23] and 3DPW [68] datasets. For instance, on 3DPW, our METRO-H64+GTR achieves $87.9\\mathrm{mm}$ MPVE, $75.5\\mathrm{mm}$ MPJPE and $46.6\\mathrm{mm}$ PAMPJPE surpassing METRO-H64. In general, token reduction results in information loss, which leads to a slight drop in accuracy. For MPVE, using an HRNet-W64 backbone, GTR typically causes a larger error increase of $2.6\\mathrm{mm}$ (from 84.1 to 86.7), while ITP is 1.5 mm (from 86.7 to 88.2). This suggests GTR sacrifices more accuracy for efficiency, while ITP has less impact. Qualitative results of FastMETRO+GTR+ITP@20% are visualized in Figure 4, where our method produces accurate and robust human mesh recovery from monocular images. In the inset, we also provide a qualitative comparison of FastMETRO+GTR+ITP@20% (w/ GTR+ITP) and the baseline model FastMETRO (w/o GTR+ITP),", + "bbox": [ + 75, + 565, + 470, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the joint-vertex attention is similar to the blending weights in SMPL, which properly captures the shape structure. However, the model w/o GTR+ITP redundantly correlates local joints with distant vertices, leading to additional interaction costs. In summary, extensive experiments validate the effectiveness of proposed strategies for token reduction across different", + "bbox": [ + 496, + 444, + 738, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Transformer structures (Encoder- w/ GTR+ITP w/o GTR+ITP based METRO [39] and Encoder-Decoder-based Fast-METRO [8]) and different Transformer model sizes (Fast-METRO and FastMETRO(S)).", + "bbox": [ + 496, + 609, + 890, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. ITP v.s. TokenLearner [59]", + "text_level": 1, + "bbox": [ + 500, + 679, + 743, + 695 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/74aa84dafe894e0cae915d51f8a555f536c90f31b236fc64cca0d3495cb06703.jpg", + "table_caption": [ + "Table 6. Comparison with TokenLearning [59] in ITP for token reduction on 3DPW [68] and Human3.6M [23]." + ], + "table_footnote": [], + "table_body": "
MethodGFLOPs3DPWHuman3.6M
MPVEMPJPEPAMPJPEMPJPEPAMPJPE
TokenLearner [59]5.799.382.452.661.245.4
Image Token Pruning5.399.282.452.359.840.5
", + "bbox": [ + 506, + 762, + 890, + 813 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare ITP with another popular pruning strategy Tokenlearner [59] for HMR on Human3.6M and 3DPW in Table 6. We use Encoder-Decoder structure [8] with ResNet-50 [20] as a backbone. The pruning rate is $20\\%$ . In", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/13333b54e4bf922d542975797abbb2441bedec462c18fb5b7ec7a5b6300e280a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 446, + 883, + 604 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15149", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a82b688f14eabb1b9abc281376466415eb929411338a0fed2d33583d20ea8626.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 282, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4cb2e743f5cd93c723a94d3ade0c3b2baefe0cc076ddb267ea71e2e9782cef19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 88, + 483, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0808a5a7604bb6d31de8de63e26009c6d52088476f90a7affae7351627a9a40f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 88, + 687, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b94ffb196aa260807334e7493627359a983ae801a5cfc48d1f5727571b9fd91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 88, + 890, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a58e49d5c51325fe02ac4f8a10df0d57d82993c21c1886b4aeb97f5e6c8332fa.jpg", + "image_caption": [ + "Figure 5. Qualitative results on FreiHAND [85] by FastMETRO+H64+GTR+ITP@20% model." + ], + "image_footnote": [], + "bbox": [ + 80, + 167, + 281, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a79a07f186f6410fd460ab244e76cd6f47964fff8edee0a83f4a12f308c4f474.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 167, + 483, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12ec44111f8919a2f3c4a696ae2af2475487e668e72a0417ac1fdd9af0026d0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 167, + 687, + 246 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0525dc7d8c878ec74c33ab57bae424ba84c5241ab106b1940a8aa949f80c7542.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 167, + 890, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8, ITP saves more computational costs while achieving the highest accuracy in terms of both human mesh recovery and joint estimation.", + "bbox": [ + 75, + 289, + 468, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. NSR v.s. GCN [27] and MLP", + "text_level": 1, + "bbox": [ + 76, + 344, + 336, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We discuss the effectiveness of NSR in GTR. We compare NSR with other implementations, including MultiLayer Perceptron and Graph Convolutional Network [27] following Pose2Mesh [9]. We condition the backbone on ResNet-50 [20] using Encoder-Decoder structure [8]. In Table 7, NSR achieves higher performance on Human3.6M and 3DPW. This indicates that the attention mechanism in NSR provides a stronger modeling capability of vertices given the learned body features, which thus improves the quality of recovered mesh vertices.", + "bbox": [ + 75, + 368, + 468, + 518 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8686277e6f99048a923622e11ca64b9bb6ba61e7708295025f5666ea17f1f0b1.jpg", + "table_caption": [ + "Table 7. Comparison of different network structures of NSR for GTR on 3DPW [68] and Human3.6M [23]." + ], + "table_footnote": [], + "table_body": "
Model3DPWHuman3.6M
MPVEMPJPEPAMPJPEMPJPEPAMPJPE
Multi-Layer Perceptron99.080.449.657.738.9
Graph Convolutional Network98.881.549.858.238.8
Neural Shape Regressor95.979.249.257.238.6
", + "bbox": [ + 84, + 569, + 468, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Generalization on Hand Mesh Recovery", + "text_level": 1, + "bbox": [ + 76, + 669, + 419, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To investigate the generalizability of our framework, we conduct an experiment on monocular hand mesh recovery, which is summarized in Table 8. Following FastMETRO [8], we report PAMPJPE, F-score@15mm (F@15mm) on FreiHand [85] together with GFLOPs $^{\\mathrm{T}}$ and throughput. The qualitative results are provided in Figure 5. The hand vertex-joint interactions are visualized in Appendix C.", + "bbox": [ + 75, + 691, + 467, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Further Discussion", + "text_level": 1, + "bbox": [ + 76, + 821, + 261, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7.1 Analysis of Image Token Pruning", + "text_level": 1, + "bbox": [ + 76, + 845, + 362, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image Token Pruning, in essence, achieves body-aware clustering results encoding discriminative body features.", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c1432171cdfc64acbf21b77fca661d6da5e888dc0a2e3c0402ec27a8067af5ff.jpg", + "table_caption": [ + "Table 8. Comparison with the SOTA methods on hand mesh recovery on FreiHAND [85]." + ], + "table_footnote": [], + "table_body": "
Method\\( \\mathrm{{GFLOPs}}{}^{\\mathrm{T}} \\downarrow \\) Throughput \\( \\uparrow \\) PAMPJPE \\( \\downarrow \\mathrm{F}@{15}\\mathrm{\\;{mm}} \\uparrow \\)
METRO-H64 [39]13.1186.76.80.981
FastMETRO-H64 [8]3.3228.46.50.982
FastMETRO-H64+GTR+ITP@20%1.0260.26.70.980
", + "bbox": [ + 504, + 323, + 890, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/981d7bad6fc317071ac01a34e788242e9e05512f5c3ac3d0181418b4056ad828.jpg", + "image_caption": [ + "Figure 6. Visualization of learned semantics by ITP. Note that the clustering results are consistent across different identities, i.e., different clusters correspond to different body joints." + ], + "image_footnote": [], + "bbox": [ + 501, + 391, + 890, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize the heatmap of the predicted clustering scores in Figure 6, where the model is with ITP at $50\\%$ pruning rate (resulting in 24 image tokens) and GTR. The backbone is ResNet-50. As shown in Figure 6, clustering-based pruning maps original tokens to a fewer number of clusters with semantics corresponding to the body joints. Note that the semantics is consistent across different identities.", + "bbox": [ + 496, + 662, + 890, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7.2 ITP with Token Reduction Supervision", + "text_level": 1, + "bbox": [ + 498, + 786, + 823, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 7, we visualize the scores predicted by ITP in HMR. The mask supervision is generated by projecting mesh vertices to image patches using estimated camera parameters as shown in Figure 7 (b)(c). ITP effectively learns to pay attention to a human body region in the given image. Quantitatively, the projection supervision improves PAM", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15150", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fc54262c033eabde71b001267b344aab5970e91b012da652442cb257b20f2a9b.jpg", + "image_caption": [ + "Figure 7. Visualization of learned semantics by Image Token Pruner. (a) projected joints. (b) projected mesh vertices. (c) mask supervision. (d) scores predicted by ITP." + ], + "image_footnote": [], + "bbox": [ + 80, + 87, + 472, + 252 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4d8665f371240e1a786092e55e385b39a9e7f13f3bb6e64c7b0eb002872b5d06.jpg", + "image_caption": [ + "Input R-Elbow R-Knee L-Elbow L-Knee Head", + "Figure 8. Visualization of cross-attention between joint and vertices. Samples are from Human3.6M and 3DPW." + ], + "image_footnote": [], + "bbox": [ + 80, + 311, + 468, + 463 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "PJPE from 41.6 to 40.6 on Human3.6M.", + "bbox": [ + 76, + 537, + 346, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.7.3 Vertex-Body Feature Interactions", + "text_level": 1, + "bbox": [ + 76, + 571, + 364, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We show the interactions between vertices and joints modeled by cross attention within Neural Shape Regressor; See Figure 8, where the heatmap is obtained by averaging attention scores across all heads of the multi-head cross-attention between query vertices and body features. As shown in Figure 8, the interactions of mesh vertices and joints are at the body part level, e.g., elbow, knee, head, which are similar to the way of the blending of a human body model, i.e., SMPL [44] and thus validates our claims.", + "bbox": [ + 75, + 595, + 468, + 731 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.8. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 76, + 739, + 341, + 755 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "When an extremely high pruning rate is applied to ITP, the accuracy of the model drops dramatically, e.g., when the token number is pruned to be one, the accuracy of the model drops dramatically (12.1%) from 38.6 to 43.3 PAMPJPE. For GTR, since we regress joints and vertices progressively, the quality of the recovered vertices by NSR depends on the learned body features. More failure cases are in Appendix D. In future work, one of the promising directions could be applying the shown enhanced efficiency of", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "HMR from monocular images to methods exhibiting high complexity for improving the model efficiency, especially in tasks such as human-environment/object interaction that perceives environments [37, 22, 18, 75, 73, 19, 61], as well as HMR from videos that involve temporal information [28, 53, 35, 63, 79, 72].", + "bbox": [ + 496, + 90, + 892, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 196, + 619, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we investigate the issues of token redundancy in the existing Transformer-based methods for both body and hand mesh recovery tasks. To tackle the problem, we introduce two effective token reduction strategies for Transformers by incorporating insights from both 3D geometry structure and 2D image features. Specifically, we recover the body shape in a hierarchical manner and cluster for image features to feed fewer but more discriminative tokens to the Transformer. Our method dramatically reduces the high-complexity interactions in the Transformer, improving the Pareto-front of accuracy and efficiency. Extensive experiments validate the proposed strategies.", + "bbox": [ + 496, + 222, + 892, + 404 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 417, + 687, + 435 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors would like to thank Gene Team for the fruitful discussion and the anonymous reviewers for their valuable comments and suggestions. This research is supported by Meta Reality Labs, Innovation and Technology Commission (Ref: ITS/319/21FP) and Research Grant Council (Ref: 17210222).", + "bbox": [ + 496, + 444, + 890, + 534 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 549, + 596, + 565 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014.", + "[2] Paolo Baerlocher and Ronan Boulic. Parametrization and range of motion of the ball-and-socket joint. In Deformable avatars, pages 180–190. Springer, 2001.", + "[3] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014.", + "[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In European conference on computer vision, pages 561-578. Springer, 2016.", + "[5] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, et al. Humman: Multi-modal 4d human dataset for versatile sensing and modeling. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 557-577. Springer, 2022." + ], + "bbox": [ + 506, + 574, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15151", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020.", + "[7] Xuanyao Chen, Zhijian Liu, Haotian Tang, Li Yi, Hang Zhao, and Song Han. Sparsevit: Revisiting activation sparsity for efficient high-resolution vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2061-2070, 2023.", + "[8] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3d human mesh recovery with transformers. In European Conference on Computer Vision (ECCV), 2022.", + "[9] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision, pages 769-787. Springer, 2020.", + "[10] MMHuman3D Contributors. Openmmlab 3d human parametric model toolbox and benchmark. https://github.com/open-mmlab/mmhuman3d, 2021.", + "[11] Haoye Dong, Xiaodan Liang, Xiaohui Shen, Bochao Wang, Hanjiang Lai, Jia Zhu, Zhiting Hu, and Jian Yin. Towards multi-pose guided virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9026-9035, 2019.", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[13] Sai Kumar Dwivedi, Nikos Athanasiou, Muhammed Kocabas, and Michael J Black. Learning to regress bodies from images using differentiable semantic rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11250-11259, 2021.", + "[14] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021.", + "[15] Chuang Gan, Yi Yang, Linchao Zhu, Deli Zhao, and Yueting Zhuang. Recognizing an action using its name: A knowledge-based approach. International Journal of Computer Vision, 120(1):61-77, 2016.", + "[16] Ryan Grainger, Thomas Paniagua, Xi Song, and Tianfu Wu. Learning patch-to-cluster attention in vision transformer. arXiv preprint arXiv:2203.11987, 2022.", + "[17] Yong Guo, Zhiyang Dou, Nan Zhang, Xiyue Liu, Boni Su, Yuguo Li, and Yinping Zhang. Student close contact behavior and COVID-19 transmission in China's classrooms. PNAS Nexus, 2(5):pgad142, 05 2023.", + "[18] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019.", + "[19] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3d scenes" + ], + "bbox": [ + 78, + 92, + 467, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14708-14718, 2021.", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[21] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016.", + "[22] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274-13285, 2022.", + "[23] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013.", + "[24] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018.", + "[25] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7122-7131, 2018.", + "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015.", + "[27] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016.", + "[28] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5253-5263, 2020.", + "[29] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11127-11137, 2021.", + "[30] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019.", + "[31] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2261, 2019.", + "[32] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4501-4510, 2019.", + "[33] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren," + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15152", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 620-640. Springer, 2022.", + "[34] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6050-6059, 2017.", + "[35] Gun-Hee Lee and Seong-Whan Lee. Uncertainty-aware human mesh recovery from video by learning part-based 3d dynamics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12375–12384, 2021.", + "[36] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3383-3393, 2021.", + "[37] Zhi Li, Soshi Shimada, Bernt Schiele, Christian Theobalt, and Vladislav Golyanik. Mocapdeform: Monocular 3d human motion capture in deformable scenes. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022.", + "[38] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022.", + "[39] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1954-1963, 2021.", + "[40] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphomer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12939-12948, 2021.", + "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014.", + "[42] Xiyue Liu, Zhiyang Dou, Lei Wang, Boni Su, Tianyi Jin, Yong Guo, Jianjian Wei, and Nan Zhang. Close contact behavior-based Covid-19 transmission and interventions in a subway system. Journal of Hazardous Materials, 436:129233, 2022.", + "[43] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021.", + "[44] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015.", + "[45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Haoyu Ma, Liangjian Chen, Deying Kong, Zhe Wang, Xingwei Liu, Hao Tang, Xiangyi Yan, Yusheng Xie, Shih-Yao Lin, and Xiaohui Xie. Transfusion: Cross-view fusion with transformer for 3d human pose estimation. arXiv preprint arXiv:2110.09554, 2021.", + "[47] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: token-pruned pose transformer for monocular and multi-view human pose estimation. In ECCV, 2022.", + "[48] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 2018 International Conference on 3D Vision (3DV), pages 120-130. IEEE, 2018.", + "[49] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12309-12318, 2022.", + "[50] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision, pages 752-768. Springer, 2020.", + "[51] Paschalis Panteleris and Antonis Argyros. Pe-former: Pose estimation transformer. In International Conference on Pattern Recognition and Artificial Intelligence, pages 3-14. Springer, 2022.", + "[52] Georgios Pavlakos, Nikos Kolotouros, and Kostas Daniilidis. Texturepose: Supervising human mesh estimation with texture consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 803-812, 2019.", + "[53] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human mesh recovery from multiple shots. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1485-1495, 2022.", + "[54] Georgios Pavlakos, Luyang Zhu, Xiaowei Zhou, and Kostas Daniilidis. Learning to estimate 3d human pose and shape from a single color image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 459-468, 2018.", + "[55] Leonid Pishchulin, Eldar Insafutdinov, Siyu Tang, Bjoern Andres, Mykhaylo Andriluka, Peter V Gehler, and Bernt Schiele. Deepcut: Joint subset partition and labeling for multi person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4929-4937, 2016.", + "[56] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021.", + "[57] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), Nov. 2017." + ], + "bbox": [ + 501, + 92, + 892, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15153", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Yu Rong, Jingbo Wang, Ziwei Liu, and Chen Change Loy. Monocular 3d reconstruction of interacting hands via collision-aware factorized refinements. In 2021 International Conference on 3D Vision (3DV), pages 432-441. IEEE, 2021.", + "[59] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. 2021.", + "[60] Akash Sengupta, Ignas Budvytis, and Roberto Cipolla. Hierarchical kinematic probability distributions for 3d human shape and pose estimation from images in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11219-11229, 2021.", + "[61] Zehong Shen, Zhi Cen, Sida Peng, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Learning human mesh recovery in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17038-17047, 2023.", + "[62] Deepak Sridhar, Niamul Quader, Srikanth Muralidharan, Yaoxin Li, Peng Dai, and Juwei Lu. Class semantics-based attention for action detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13739-13748, 2021.", + "[63] Zhaoqi Su, Weilin Wan, Tao Yu, Lingjie Liu, Lu Fang, Wenping Wang, and Yebin Liu. Mulaycap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics, 28(4):1862-1879, 2020.", + "[64] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11179-11188, 2021.", + "[65] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019.", + "[66] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A survey. arXiv preprint arXiv:2203.01923, 2022.", + "[67] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017.", + "[68] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), sep 2018.", + "[69] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20460-20469, 2022.", + "[70] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on pattern analysis and machine intelligence, 43(10):3349-3364, 2020.", + "[71] Wenjia Wang, Yongtao Ge, Haiyi Mei, Zhongang Cai, Qingping Sun, Yanjun Wang, Chunhua Shen, Lei Yang, and Taku Komura. Zolly: Zoom focal length correctly for perspective-distorted human mesh reconstruction. arXiv preprint arXiv:2303.13796, 2023.", + "[72] Wen-Li Wei, Jen-Chun Lin, Tyng-Luh Liu, and Hong-Yuan Mark Liao. Capturing humans in motion: Temporal-attentive 3d human pose and shape estimation from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13211–13220, 2022.", + "[73] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Progressive unsupervised learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2993-3002, June 2021.", + "[74] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B. Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14561-14571, June 2023.", + "[75] Yan Xia, Qiangqiang Wu, Wei Li, Antoni B Chan, and Uwe Stilla. A lightweight and detector-free 3d single object tracker on point clouds. IEEE Transactions on Intelligent Transportation Systems, 2023.", + "[76] Yan Xia, Yusheng Xu, Shuang Li, Rui Wang, Juan Du, Daniel Cremers, and Uwe Stilla. Soe-net: A self-attention and orientation encoding network for point cloud based place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11348–11357, 2021.", + "[77] Yuanlu Xu, Song-Chun Zhu, and Tony Tung. Denserac: Joint 3d pose and shape estimation by dense render-and-compare. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7760-7770, 2019.", + "[78] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch on imagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 558-567, 2021.", + "[79] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11038-11049, 2022.", + "[80] Wang Zeng, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11101-11111, 2022.", + "[81] Yuanhao Zhai, Le Wang, Wei Tang, Qilin Zhang, Junsong Yuan, and Gang Hua. Two-stream consensus network for weakly-supervised temporal action localization. In Euro-" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "15154", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pean conference on computer vision, pages 37-54. Springer, 2020.", + "[82] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11446-11456, 2021.", + "[83] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE International Conference on Computer Vision, 2021.", + "[84] Nan Zhang, Li Liu, Zhiyang Dou, Xiyue Liu, Xueze Yang, Doudou Miao, Yong Guo, Silan Gu, Yuguo Li, Hua Qian, and Jianjian Wei. Close contact behaviors of university and school students in 10 indoor environments. Journal of Hazardous Materials, 458:132069, 2023.", + "[85] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019." + ], + "bbox": [ + 78, + 92, + 467, + 414 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "15155", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/2d762d0d-bad4-403d-98f7-4210b7f7e39d_model.json b/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/2d762d0d-bad4-403d-98f7-4210b7f7e39d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..56a103ed158b2efc9d7f42b96ee06ec2bf82334a --- /dev/null +++ b/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/2d762d0d-bad4-403d-98f7-4210b7f7e39d_model.json @@ -0,0 +1,2998 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.131, + 0.89, + 0.152 + ], + "angle": 0, + "content": "TORE: Token Reduction for Efficient Human Mesh Recovery with Transformer" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.179, + 0.815, + 0.224 + ], + "angle": 0, + "content": "Zhiyang Dou\\(^{1\\dagger}\\) Qingxuan Wu\\(^{2\\dagger}\\) Cheng Lin\\(^{3\\dagger}\\) Zeyu Cao\\(^{4\\dagger}\\) Qiangqiang Wu\\(^{5}\\) Weilin Wan\\(^{1}\\) Taku Komura\\(^{1}\\) Wenping Wang\\(^{6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.228, + 0.832, + 0.267 + ], + "angle": 0, + "content": "\\(^{1}\\)The University of Hong Kong \\(^{2}\\)University of Oxford \\(^{3}\\)Tencent Games \\(^{4}\\)University of Cambridge \\(^{5}\\)City University of Hong Kong \\(^{6}\\)Texas A&M University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.3, + 0.314, + 0.316 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.474, + 0.651 + ], + "angle": 0, + "content": "In this paper, we introduce a set of simple yet effective TOken REDuction (TORE) strategies for Transformer-based Human Mesh Recovery from monocular images. Current SOTA performance is achieved by Transformer-based structures. However, they suffer from high model complexity and computation cost caused by redundant tokens. We propose token reduction strategies based on two important aspects, i.e., the 3D geometry structure and 2D image feature, where we hierarchically recover the mesh geometry with priors from body structure and conduct token clustering to pass fewer but more discriminative image feature tokens to the Transformer. Our method massively reduces the number of tokens involved in high-complexity interactions in the Transformer. This leads to a significantly reduced computational cost while still achieving competitive or even higher accuracy in shape recovery. Extensive experiments across a wide range of benchmarks validate the superior effectiveness of the proposed method. We further demonstrate the generalizability of our method on hand mesh recovery. Visit our project page at https://frank-zy-dou.github.io/projects/Tore/index.html." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.659, + 0.21, + 0.674 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Human Mesh Recovery (HMR) has been extensively researched in recent years, given its wide real-world applications [15, 81, 11, 62, 69, 17, 42, 84]. This task becomes more challenging when the input is a monocular 2D image, due to the large pose and shape variation, large appearance variation, partial observation, and self-occlusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.471, + 0.88 + ], + "angle": 0, + "content": "There has been steady progress in 3D human mesh recovery [30, 50, 36, 24, 60, 39, 40, 8, 5]. Recently, Transformer [67] has shown state-of-the-art (SOTA) results on a wide variety of tasks due to its strong capability of capturing long-range dependency for more accurate predictions [3, 67, 76, 74]. Using tokens constructed from local features extracted by a convolutional neural network" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.299, + 0.831, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.399, + 0.831, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.503, + 0.887, + 0.525 + ], + "angle": 0, + "content": "
● METRO baseline ▲ FastMETRO baseline ● FastMETRO(S) baseline
METRO-TORE ▼ FastMETRO-TORE ● FastMETRO(S)-TORE
" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.611 + ], + "angle": 0, + "content": "Figure 1. Throughput v.s. Accuracy and GFLOPs v.s. Accuracy on Human3.6M [23]. Our method dramatically saves GFLOPs and improves throughput while maintaining highly competitive accuracy. The x-axis of the bottom GFLOPs figure is reversed for demonstration. Eb0 and R50 represent EfficientNet-b0 [65] and ResNet-50 [20] backbones, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.657 + ], + "angle": 0, + "content": "(CNN) [20, 70] to query the joint and mesh vertex positions, Transformer-based methods [39, 40, 8] achieved SOTA performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.87 + ], + "angle": 0, + "content": "However, improved performance comes with costs: the increased expressivity of Transformers comes with quadratically increasing computational costs since all pairwise interactions are taken into account [14]. The space and time complexity of a QKV-attention operation is known to be \\( O(M^2) \\), where \\( M \\) is the number of tokens. The token number thus plays a vital role in time efficiency: a large number of tokens inevitably leads to a heavy computation burden. Unfortunately, almost all the existing Transformer-based SOTA methods for HMR [39, 40, 8] are suffering from redundant tokens. This incurs a high model complexity and computational costs, which prevents the current Transformer-based HMR methods from achieving their full potential in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In this paper, we make key observations from two important aspects: the 3D geometry structure and 2D image" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.269, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger ,\\ddagger\\) denote equal contributions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15143" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.38 + ], + "angle": 0, + "content": "feature, to reveal the problem of token redundancy. First, to recover the 3D geometry, all existing methods use both mesh vertices and skeleton joints as tokens for the feature interaction between input and body shape. Whereas a body mesh contains numerous vertices, they can be abstracted by a small number of joints on the skeleton. For instance, when animating a SMPL [44] avatar, the skeleton joints, together with a blend shape binding the joints and corresponding mesh vertices of the local body part, are able to describe various body meshes. Therefore, the joints can already be viewed as an underlying structure of a body shape, which intrinsically encodes the human mesh geometry. Second, for image-based input, most existing methods indiscriminately use all the feature patches to capture pose, shape and appearance variance. However, although the human body exhibits large variance, the important features for shape inference are dominantly clustered within the body area in an RGB image. Most features, e.g., image background, are not informative, thus bringing about redundancy." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.384, + 0.473, + 0.807 + ], + "angle": 0, + "content": "Given the aforementioned insights, we argue that the Pareto-front of accuracy and efficiency for Transformer-based HMR could be further improved by reducing the number of tokens [59, 47, 56]. To this end, we introduce a set of simple yet effective token reduction strategies mainly from two aspects corresponding to our observations. First, for 3D mesh recovery, instead of querying both vertices and joints with input features simultaneously, we consider learning a small set of body tokens at the skeleton level for each body part. To recover corresponding mesh vertices, we use an efficient Neural Shape Regressor (NSR) to infer the mesh from the body features encoded by these tokens. This query process can also be interpreted as an attention matrix decomposition, by which we effectively leverage the geometric insights encoded at the skeleton level to infer the mesh structure hierarchically. Second, for the input image feature, we introduce a learnable token pruner to prune the tokens of patch-based features extracted by a CNN. We employ a clustering-based strategy to identify discriminative features, which results in two appealing properties: 1) the end-to-end learning of the pruner is unsupervised, avoiding the need for additional data labeling; 2) it learns semantically consistent features across various images, thus further benefiting the geometry reasoning and enhancing the capability of generalizability. These token reduction strategies substantially reduce the number of query tokens involved in the computation without sacrificing the important information. An overview is shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We conduct extensive experiments across wide benchmarks [23, 68, 85], including both the human body and hand mesh recovery, to validate the proposed method. Compared to SOTA methods, our framework faithfully recovers body meshes with fewer tokens, which considerably reduces memory and computation overhead while maintain" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.093, + 0.741, + 0.107 + ], + "angle": 0, + "content": "ing competitive geometric accuracy." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.108, + 0.804, + 0.122 + ], + "angle": 0, + "content": "In summary, our contribution is three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.133, + 0.892, + 0.163 + ], + "angle": 0, + "content": "- We reveal the issues of token redundancy in the existing Transformer-based methods for HMR." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.174, + 0.894, + 0.22 + ], + "angle": 0, + "content": "- We propose effective strategies for token reduction by incorporating the insights from the 3D geometry structure and 2D image feature into the Transformer design." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.231, + 0.894, + 0.352 + ], + "angle": 0, + "content": "- Our method achieves SOTA performance on various benchmarks with less computation cost. For instance, for the Transformer Encoder structure [39] and the Transformer Encoder-Decoder structure [8] with ResNet-50 [20] backbone, our method maintains competitive accuracy while saving \\(82.9\\%\\), \\(50.5\\%\\) GFLOPs and improving \\(139.1\\%\\), \\(39.8\\%\\) throughput, respectively; see Figure 1 for an overview." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.133, + 0.894, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.362, + 0.642, + 0.379 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.522 + ], + "angle": 0, + "content": "Human Mesh Recovery (HMR) from monocular images has achieved great progress in the past years [4, 30, 50, 36, 10, 40, 24, 60, 5, 39, 8, 82, 71, 58]. Given a monocular RGB image, the goal of HMR is to recover the 3D body shape, typically using a human body model, e.g., SMPL [44]. We refer readers to [66] for a comprehensive review. Existing methods fall into two categories: Non-Transformer-based and Transformer-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.895, + 0.795 + ], + "angle": 0, + "content": "Non-Transformer-based Human Mesh Recovery To recover the human body shape, one could predict a few parameters, i.e., joint rotation and body shape, to drive the parametric model (parametric approach) or directly regress the vertex positions of the body (non-parametric approach). Parametric approaches SMPLify [4] estimates human pose and shape by fitting SMPL to the detected 2D key points [55] by optimization. Kanazawa et al. [24] adopt adversarial prior knowledge of the 3D body shape into the neural network for HMR. SPIN [30] then combines regression-based and optimization-based methods during the training loop. Intermediate features such as key points [54], pixel-to-surface correspondences [77] and texture consistency [52], have been exploited. HybrIK [36] proposes a hybrid inverse kinematics via twist-and-swing decomposition [2]. Regressing the body parameters from the image is a highly non-linear mapping [50], which thus limits the performance of parametric-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Non-parametric approaches Non-parametric approaches [50, 32, 9] recover spatial coordinates of a body shape directly from the image features. I2L-MeshNet [50] predicts the per-pixel likelihood on 1D heatmaps for vertex coordinates for a trade-off between accuracy and computational cost. GraphCMR [32] and Pose2Mesh [9] employ graph convolutional neural network (GCNN) [27]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15144" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.086, + 0.862, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.893, + 0.315 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed framework. Our goal is to reduce tokens for Transformer Encoder and Decoder which are critical modules in the whole pipeline. Image Token Pruner (ITP) and Neural Shape Regressor (NSR) are two lightweight components." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.47, + 0.399 + ], + "angle": 0, + "content": "to regress mesh positions by modeling local interactions among vertices from the image [32] and 2D pose [9], respectively. Although steady progress has been made, SOTA accuracy is produced by Transformer-based methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.73 + ], + "angle": 0, + "content": "Transformer-based Human Mesh Recovery Recently, Transformer structures have been successfully adopted in various vision tasks [12, 43, 6, 51, 66, 80]. In HMR, the SOTA performance is achieved by Transformer-based (attention mechanism) approaches [39, 29, 40, 8]. Specifically, METRO [39] recovers body mesh using a Transformer Encoder [67] to model vertex-vertex and vertex-joint interactions conducting dimension reduction from the image features to 3D shape. MeshGraphormer [40] injects Graph Convolutions [27] into the Transformer Encoder [39] blocks to improve local interactions among vertices. The recent work FastMETRO [8] employs a Transformer Encoder-Decoder architecture disentangling the image encoding and mesh estimation for model reduction and acceleration. However, all Transformer-based approaches [39, 40, 8] suffer from redundant tokens, causing heavy interaction between body geometry and image features, which is still cumbersome and computationally intensive. In this paper, we design strategies driven by the 3D body geometry and 2D image feature for token reduction for both Transformer Encoder and Transformer Encoder-Decoder structures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.897 + ], + "angle": 0, + "content": "Token Reduction for Transformers Token reduction for Transformer structures has been causing people's attention nowadays [56, 47, 59, 78, 16, 38, 46, 7]. In particular, DynamicViT [56] reduces less informative tokens hierarchically with supervision from a teacher network to save computational costs during inference. TokenLearner [59] generates a smaller number of tokens with spatial attention adaptively. Patch-to-Cluster [16] proposes to cluster over image patches with a token selector. Our method follows this way and adopts a token selection layer for clustering over input tokens, reducing tokens and improving general" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.893, + 0.445 + ], + "angle": 0, + "content": "ization capability in challenging scenarios. Token pruning techniques have also been explored in single-view 2D pose estimation and multi-view 3D pose estimation [47] by selecting tokens based on the attention value in the Transformer. Differently, in HMR, besides the key points, body geometry, i.e., mesh vertices, also introduces a burden to the Transformer and needs to be carefully considered." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.46, + 0.637, + 0.478 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.608 + ], + "angle": 0, + "content": "We employ the popular body parametric model SMPL [44] to represent a human body. Formally, given a monocular image \\(\\mathcal{I} \\in \\mathbb{R}^{224 \\times 224}\\), our goal is to recover the joint positions \\(\\mathcal{J} \\in \\mathbb{R}^{J \\times 3}\\) and mesh vertices \\(\\mathcal{V} \\in \\mathbb{R}^{V \\times 3}\\) where \\(J = 14\\), \\(V = 6890\\) are the numbers of joints and the mesh vertices, respectively. In addition to the human body, we also demonstrate our method on hand mesh recovery with MANO model [57] to show its generalizability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.714 + ], + "angle": 0, + "content": "Existing Transformer-based HMR methods fall into two categories: a single Transformer Encoder structure [39, 40] and a Transformer Encoder-Decoder structure [8]. Accordingly, we propose two strategies to reduce tokens from 3D and 2D levels, namely, Geometry Token Reduction (GTR) and Image Token Pruning (ITP), which can accommodate different Transformer structures to improve their efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.805, + 0.741 + ], + "angle": 0, + "content": "3.1. Geometry Token Reduction (GTR)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.854 + ], + "angle": 0, + "content": "As an essential module, all the existing Transformer-based methods for HMR [39, 40, 8] have to model the interaction between body geometry and image features with an attention mechanism. Therefore, an effective token reduction method to improve the efficiency for inferring mesh geometry can benefit the design of Transformer based methods in a wide scope." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We observe existing methods invariably use both skeleton joints and mesh vertices as tokens for regressing their spatial coordinates. The numerous tokens inevitably lead to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15145" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.196 + ], + "angle": 0, + "content": "significant computational costs for a Transformer. In fact, as an underlying structure of the human body, the skeleton joints already provide strong priors for perceiving the body geometry. A human body avatar, like SMPL [44], can be driven by a small number of joints with the blend shape. This inspires us to regress the human body mesh in a hierarchical manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.468, + 0.334 + ], + "angle": 0, + "content": "Our key idea is to decompose a heavy Transformer into smaller Transformer modules, each of which involves fewer tokens, avoiding expensive computation for image feature interaction. Specifically, our approach only queries the body tokens, whose count is equivalent to the number of joints. Since the number of joints is much less than the number of mesh vertices, we can significantly reduce the token number in complex 2D-3D feature interactions and this leads to a substantial enhancement in efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.468, + 0.501 + ], + "angle": 0, + "content": "To faithfully recover body shape from the body tokens, we introduce a Neural Shape Regressor (NSR) module, which is a light Transformer structure, to query the vertex tokens from their interaction with body features. Let \\( F_{J} = \\{f_{1}^{j}, f_{2}^{j}, \\dots, f_{J}^{j}\\} \\) denotes a set of learned body features and \\( T_{\\mathcal{V}} = \\{t_{1}^{v}, t_{2}^{v}, \\dots, t_{V}^{v}\\} \\) where \\( t_{i}^{v} \\) represents a set of vertex tokens which are the query tokens. We conduct cross-attention between query vertex tokens and body features to model the interaction between vertices and learned body features. Note that non-adjacent vertices are masked to improve the efficiency [8]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.468, + 0.669 + ], + "angle": 0, + "content": "The feature size within NSR is even smaller compared with the main Transformer. Thanks to the informative geometric feature encoded in the body tokens, despite not incorporating image tokens, NSR still faithfully recovers the surface vertices. Another interesting discovery is that the learned attention scores reflect the correlation between joints and vertices at a body part level, similar to body blending weights (see Sec. 4.7.3). In this way, we can effectively reduce the cost of redundant interaction by decomposing the entire body into encoded constituent parts and efficiently recovering the whole body shape." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.327, + 0.696 + ], + "angle": 0, + "content": "3.2. Image Token Pruning (ITP)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.468, + 0.839 + ], + "angle": 0, + "content": "In addition to the geometric insights, for a Transformer Encoder-Decoder structure, image tokens also affect the computation overhead. However, the existing methods adopting this Transformer structure fail to avoid the token redundancy issue [8] because all available image feature patches are indiscriminately involved in human mesh recovery. Actually, some features, e.g., image background, are not informative, thus introducing redundancy and increasing the computational cost of the Transformer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To tackle this issue, we introduce an effective token pruning strategy, namely, Image Token Pruning (ITP). Our key insight is that the informative features in an image for inferring 3D geometry are overwhelmingly clustered within" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.891, + 0.137 + ], + "angle": 0, + "content": "the region of the human body. Inspired by recent advances in token pruning [16, 59, 56, 47], we propose to aggregate features into a small number of meaningful clusters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.409 + ], + "angle": 0, + "content": "Let \\(0 < \\rho < 1\\) denote the predefined token pruning ratio. Our goal is to learn a projection that maps feature map \\(F_{\\mathcal{I}} \\in R^{HW \\times c}\\) extracted from the given image with HW tokens to \\(Z_{\\mathcal{I}} \\in R^{T \\times c}\\) with \\(T\\) tokens, where \\(T = \\lfloor \\rho HW \\rfloor\\) and \\(c\\) is the feature dimension. The small number of clusters \\(Z_{\\mathcal{I}}\\) are expected to capture the fewer but more discriminative features in \\(F_{\\mathcal{I}}\\). Inspired by [16, 59], we implement the projection as a learnable CNN module. We first apply Conv2D to the input feature, where the kernel size, stride and zero padding of Conv2D are \\(3 \\times 3\\), 1 and 1, reducing the feature dimension from \\(c\\) to \\(c' = c/4\\) with GELU [21] activation function applied. We further map the dimension \\(c\\) to \\(N\\) by MLP and Softmax, producing \\(M \\in R^{HW \\times N}\\). Finally, the clustered token set is given by \\(Z_{\\mathcal{I}} = M^T \\cdot F_{\\mathcal{I}}\\). The mapping matrix \\(M^T\\), in essence, produces a clustering over origin tokens. Each element \\(m_{ij} \\in M^T\\) depicts the contribution of \\(j\\) token to \\(i\\)-th cluster. Note that LayerNorm is employed after the clustering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Token Reduction Supervision In order to encourage ITP to pay more attention to those discriminative feature regions, i.e., body parts. during token reduction, we introduce Token Reduction Supervision. Given the weak-perspective camera estimated the neural network, we compute the 2D projection of the ground truth 3D vertices \\(\\hat{\\mathcal{V}}\\) as \\(s\\Pi (\\hat{\\mathcal{V}}) + t\\) where \\(\\Pi\\) represents an orthographic projection, \\(s\\) and \\(t\\) are scale and translation estimated by the network. The projected results are downsampled to a discrete \\(H\\times W\\) grid corresponding to each token followed by a binary indicator function" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.572, + 0.89, + 0.625 + ], + "angle": 0, + "content": "\\[\nF _ {d} (x) = \\left\\{ \\begin{array}{l} 1, \\text {i f t h e c e l l c o n t a i n s a p r o j e c t e d p o i n t .} \\\\ 0, \\text {i f t h e c e l l d o e s n o t c o n t a i n a p r o j e c t e d p o i n t .} \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.625, + 0.835, + 0.641 + ], + "angle": 0, + "content": "Then the supervision for token pruning is given by" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.652, + 0.89, + 0.693 + ], + "angle": 0, + "content": "\\[\nL _ {P} = - \\frac {1}{N H W} \\sum_ {i} ^ {N} \\left(F _ {d} \\left(s \\Pi (\\hat {\\mathcal {V}}) + t\\right) \\cdot M _ {[:, i ]}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.891, + 0.763 + ], + "angle": 0, + "content": "With this supervision, the weights of background tokens in \\( M[:i] \\) for each \\( i \\)-th cluster will be penalized during training, which thus encourages the ITP to learn the discriminative features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The proposed ITP has several noticeable advantages. First, in contrast to other methods relying on explicit supervision, e.g., using a pretrained teacher network to facilitate the pruning [56, 78, 33], our token pruner is trained in an unsupervised manner; Note that no ground truth masking is required for Token Reduction Supervision. Second, unlike previous clustering-based pruning methods [16] that use pre-defined class labels, ITP aims to adaptively identify discriminative features within the body region. Therefore," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15146" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "ITP is not limited to fixed semantic labels but is able to induce higher-level semantics according to the learning target." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.123, + 0.468, + 0.212 + ], + "angle": 0, + "content": "Finally, we find the clustered informative features learned by ITP improve the model generalizability, especially for challenging datasets, e.g., in-the-wild dataset 3DPW [68] as shown in Table 3. We elaborate on these properties in Sec. 4.3.2 and Sec. 4.7.1 with extensive experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.227, + 0.229, + 0.241 + ], + "angle": 0, + "content": "3.3. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.251, + 0.468, + 0.481 + ], + "angle": 0, + "content": "We supervise the network using L1 distance between predicted mesh vertices at three sampling levels: \\(L_{\\mathcal{V}3D} = ||\\mathcal{V}_{3D}^l - \\hat{\\mathcal{V}}^l||_1 + ||\\mathcal{V}_{3D}^m - \\hat{\\mathcal{V}}^m||_1 + ||\\mathcal{V}_{3D}^h - \\hat{\\mathcal{V}}_h||_1\\), where \\(*, *^m, *^h\\) denote low, middle and high body mesh resolution with vertex numbers to be 431, 1723 and 6890 for an SMPL body. The ground-truth 3D joints are used for supervising predicted 3D joints and the ones regressed from the vertices \\(\\mathcal{V}_{3D}^h\\) with a SMPL regression matrix \\(\\mathcal{M}\\): \\(L_{\\mathcal{J}3D} = ||\\mathcal{J}_{3D} - \\hat{\\mathcal{J}}_{3D}||_1, L_{\\mathcal{J}3D}^R = ||\\mathcal{M}(\\mathcal{V}) - \\hat{\\mathcal{J}}_{3D}||_1\\). A 2D projection loss \\(L_{J2D}\\) is used during the network training, where we employ a weak perspective camera model to project 3D joints to 2D for supervision: \\(L_{\\mathcal{J}2D}^R = ||(s\\Pi(\\mathcal{M}(\\mathcal{V})) + t) - \\hat{\\mathcal{J}}_{2D}||_1\\). \\(s, t\\) are scale and translation estimated by the network. Overall, together with the Token Pruning Supervision \\(L_P\\), the total loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.081, + 0.495, + 0.465, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L = \\alpha \\left[ \\lambda_ {\\mathcal {J} 3 D} \\left(L _ {\\mathcal {J} 3 D} ^ {R} + L _ {\\mathcal {J} 3 D}\\right) + \\lambda_ {\\mathcal {V} 3 D} \\left(L _ {\\mathcal {V} 3 D}\\right) + \\lambda_ {P} L _ {P} \\right] \\\\ + \\beta \\lambda_ {\\mathcal {J} 2 D} L _ {\\mathcal {J} 2 D} ^ {R}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.468, + 0.576 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) indicate the availability of the supervision. We set \\(\\lambda_{P},\\lambda_{J2D},\\lambda_{V3D},\\lambda_{J3D}\\) to be 1, 100, 100, 1000." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.594, + 0.282, + 0.611 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.276, + 0.635 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We evaluate our model in two scenarios: human body mesh recovery and hand mesh recovery. We adopt commonly-used metrics [25, 31, 39, 40, 8]: Mean Per-Joint Position Error (MPJPE), MPJPE after further alignment, i.e., Procrustes Analysis (PAMPJPE) and Mean Per-Vertex Error (MPVE). For human body mesh recovery, our network is trained with Human3.6M [23], MuCo-3DHP[48], UP-3D [34], COCO [41], MPI [1]. Following previous works [39, 40, 8], we use the pseudo mesh data in Human3.6M [23] for training, splitting subjects S1, S5, S6, S7, S8 for training and S9, S11 for testing. We also report our performance on 3DPW [68], a more challenging in-the-wild dataset. To further evaluate the generalization ability of the proposed method, we test our method on hand mesh recovery on FreiHAND [85] dataset. For analyzing the efficiency of our method, we report GFLOPs and throughput (image per second), strictly following [56, 49, 59, 47, 38]." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.088, + 0.892, + 0.143 + ], + "angle": 0, + "content": "Table 1. Comparison with the Transformer Encoder structure METRO (M) [39] on Human3.6M [23]. We test with ResNet-50 (R50) [20] and HRNet-W64 (H64) [70] as backbones. GFLOPs\\(^{\\mathrm{T}}\\) is GFLOPs of the transformer." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.154, + 0.892, + 0.214 + ], + "angle": 0, + "content": "
MethodGFLOPs ↓GFLOPsT ↓Throughput (im/s) ↑PAMPJPE ↓
M-H64 [39]56.527.5141.036.7
M-H64+GTR30.2 (-46.5%)0.8 (-97.1%)210.1 (+49.0%)37.1 (+1.1%)
M-R50 [39]31.627.5247.040.6
M-R50+GTR5.4 (-82.9%)0.8 (-97.1%)590.6 (+139.1%)42.0 (+3.4%)
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.252, + 0.716, + 0.268 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.892, + 0.427 + ], + "angle": 0, + "content": "Network Training We implement the network using PyTorch. For the Transformer Encoder-Decoder structure, we set the learning rate to be \\(1 \\times 10^{-4}\\). We use the AdamW optimizer [45] and train for 60 epochs, with a batch size of 16 per GPU on 4 Nvidia A100 GPUs. When comparing with Transformer Encoder structure METRO [39] (see Table 1), we follow the setting of METRO where we train the models with a batch size of 30 per GPU on 8 Nvidia A100 GPUs in total. We adopt Adam optimizer [26] and train the models for 200 epochs. See more details in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Performance Evaluation To analyze the performance on a consumer-level GPU device, we measure the throughput on an NVIDIA RTX 3090 GPU with 24G VRAM. When comparing with encoder-decoder Transformer structure FastMETRO [8], we set the batch size to 32 following PPT [47]. The comparison with Transformer Encoder structure METRO [39] uses 16 as the batch size. Note that the batch size is limited by the size of the original METRO [39] model. To factor out the influence of the batch size, we provide a more detailed performance report in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.842, + 0.605 + ], + "angle": 0, + "content": "4.3. Performance on Human Mesh Recovery" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Currently, there only exist two types of representative Transformer structures for HMR: Encoder-only structure (METRO [39]) and Encoder-Decoder structure (Fast-METRO [8]). To demonstrate the effectiveness of our method to both Transformer structures, we first conduct experiments by adding GTR to two structures: METRO and FastMETRO. Since ITP is designed for the encoder-decoder structure, we further report the results on FastMETRO with both ITP and GTR in Sec. 4.3.2. The overall comparison is shown in Table 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.776, + 0.8 + ], + "angle": 0, + "content": "4.3.1 Transformer Encoder Structure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, GTR effectively reduces the computation costs while still producing competitive accuracy results. For the encoder-based Transformer model METRO [39] with HRNet-W64 [70] as a backbone, applying GTR saves the GFLOPs of the whole model for \\(46.5\\%\\) and the Transformer part for \\(97.1\\%\\), with throughput im" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15147" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.275, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.089, + 0.471, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.166, + 0.275, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.166, + 0.471, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.242, + 0.275, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.242, + 0.471, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.317, + 0.275, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.316, + 0.471, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.396, + 0.47, + 0.424 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of GTR equipped Encoder-Decoder structure [39] (H64) on Human3.6M [23] and 3DPW [68]." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.436, + 0.47, + 0.505 + ], + "angle": 0, + "content": "Table 2. Comparison with the Transformer Encoder-Decoder structure FastMETRO (FM) [8] on Human3.6M [23]. We test with EfficientNet-b0 (Eb0) [65], ResNet-50 (R50) [20] and HRNet-W64 (H64) [70] as backbones. GFLOPs\\(^{\\mathrm{T}}\\) stand for GFLOPs for the transformer." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.516, + 0.472, + 0.617 + ], + "angle": 0, + "content": "
MethodGFLOPs ↓GFLOPsT ↓Throughput (im/s) ↑PAMPJPE ↓
FM-H64 [8]35.76.6221.533.7
FM-H64+GTR30.2 (-15.4%)0.7 (-89.4%)249.2 (+12.5%)34.8 (+3.2%)
FM-R50 [8]10.96.6576.037.3
FM-R50+GTR5.4 (-50.5%)0.7 (-89.4%)805.3 (+39.8%)38.6 (+3.4%)
FM(S)-R50 [8]6.42.2953.539.4
FM(S)-R50+GTR4.6 (-28.1%)0.3 (-86.4%)1128.9 (+18.4%)38.6 (-2.0%)
FM-Eb07.16.6517.645.8
FM-Eb0+GTR1.7 (-76.1%)0.7 (-89.4%)870.5 (+68.2%)44.2 (-3.5%)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.779 + ], + "angle": 0, + "content": "proved by \\(49\\%\\). For the ResNet-50 [20] backbone, GTR helps save the GFLOPs of the whole model for \\(82.9\\%\\) and the Transformer part for \\(97.1\\%\\), with throughput improved by \\(139.1\\%\\). These results validate the effectiveness of the proposed GTR on the Transformer Encoder structure. Qualitative results of METRO-H64+GTR can be found in Figure 3, in which the model produces high-quality results in human mesh recovery over various input monocular images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.418, + 0.83 + ], + "angle": 0, + "content": "4.3.2 Transformer Encoder-Decoder Structure" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We experiment on the Transformer Encoder-Decoder structure FastMETRO [8] with its two variants: models with 1 and 3 Encoder-Decoder layers. For 1-layer FastMETRO, we denote it as FastMETRO(S). In Table 2, GTR equipped" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "FastMETRO [39] produces competitive accuracy while reducing \\(15.4\\%\\) in GFLOPs for the whole model, \\(89.4\\%\\) in GFLOPs for the Transformer part and improving \\(12.5\\%\\) in throughput with an HRNet-W64 [70] CNN backbone. When testing with ResNet-50 [20], GTR helps save \\(50.5\\%\\) in GFLOPs for the whole model, \\(89.4\\%\\) in GFLOPs for the Transformer part and improves \\(39.8\\%\\) in throughput. Notably, our FastMETRO(S)-R50+GTR and FastMETRO(S)-Eb0+GTR yield 38.6, 44.2 mm PAMPJPE respectively, surpassing the corresponding baselines while greatly saving computational costs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.304 + ], + "angle": 0, + "content": "We then present the performance of the Transformer Encoder-Decoder structure with GTR and ITP using ResNet-50 [20] and EfficientNet-b0 [65] CNN backbones." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.317, + 0.892, + 0.345 + ], + "angle": 0, + "content": "Table 3. Influence of ITP for monocular 3D human mesh recovery on 3DPW [68]." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.354, + 0.892, + 0.391 + ], + "angle": 0, + "content": "
MethodMPVEMPJPEPAMPJPE
FastMETRO-H64+GTR91.375.446.7
FastMETRO-H64+GTR+ITP@20%88.272.344.4
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.892, + 0.528 + ], + "angle": 0, + "content": "As shown in Table 4, TORE (GTR+ITP) is effective for the Encoder-Decoder Transformer structure. Specifically, when pruning with \\(20\\%\\), \\(50\\%\\) in ITP, the models save the GLOPs of the Transformer structure by \\(14.3\\%\\), \\(14.3\\%\\) and \\(28.5\\%\\), \\(42.9\\%\\) while receiving competitive accuracy or even higher accuracy i.e., FM-Eb0+GTR+ITP@\\(20\\%\\) improves PAMPJPE compared with the baseline." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.571 + ], + "angle": 0, + "content": "Table 5. Comparison with the SOTA methods for monocular 3D human mesh recovery on 3DPW [68] and Human3.6M [23]." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.58, + 0.897, + 0.793 + ], + "angle": 0, + "content": "
Method3DPWHuman3.6M
MPVEMPJPEPAMPJPEMPJPEPAMPJPE
HMR [25]-130.076.788.056.8
GraphCMR [32]--70.2-50.1
SPIN [30]116.496.959.262.541.1
I2LMeshNet [50]-93.257.755.741.1
PyMAF [83]110.192.858.957.740.5
ROMP-R50 [64]105.689.353.5--
PARE-R50 [29]99.782.952.3--
DSR-R50 [13]99.585.751.760.940.3
METRO-R50 [39]---56.540.6
METRO-H64 [39]88.277.147.954.036.7
METRO-H64+GTR87.975.546.657.637.1
FastMETRO-R50 [8]90.677.948.353.937.3
FastMETRO-R50+GTR+ITP@20%99.282.452.359.840.5
FastMETRO-H64 [8]84.173.544.652.233.7
FastMETRO-H64+GTR+ITP@20%88.272.344.459.636.4
FastMETRO-Eb0 [8]112.593.860.269.245.8
FastMETRO-Eb0+GTR+ITP@20%112.593.760.163.243.9
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In addition, we find that ITP helps improve the accuracy on the challenging in-the-wild 3DPW [68] dataset, shown in Table 3. Specifically, when further equipped with ITP, the model performance in MPVE, MPJPE and PAMPJPE are improved by \\(3.1\\mathrm{mm}\\), \\(3.1\\mathrm{mm}\\) and \\(2.3\\mathrm{mm}\\), respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15148" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.891, + 0.116 + ], + "angle": 0, + "content": "Table 4. Statistics of all proposed components (GTR + ITP) for Encoder-Decoder structure FastMETRO (FM) [8] on Human3.6M [23]. The backbones are EfficientNet-b0 (Eb0) [65] and ResNet-50 (R50) [20]. GFLOPs\\(^{\\mathrm{T}}\\) stands for GFLOPs of the Transformer." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.125, + 0.887, + 0.219 + ], + "angle": 0, + "content": "
Method + Pruning Rate#TokensGFLOPs ↓GFLOPsT ↓Throughput (im/s) ↑PAMPJPE ↓
FM-R50+GTR495.40.7805.338.6
FM-R50+GTR+ITP@20%395.3 (-1.8%)0.6 (-14.3%)794.7 (-1.3%)40.5 (+4.9%)
FM-R50+GTR+ITP@50%245.1 (-5.5%)0.5 (-28.5%)806.1 (+0.1%)40.7 (+5.4%)
FM-Eb0+GTR491.70.7870.544.2
FM-Eb0+GTR+ITP@20%391.6 (-5.9%)0.6 (-14.3%)876.3 (+0.7%)43.9 (-0.%)
FM-Eb0+GTR+ITP@50%241.4 (-17.6%)0.4 (-42.9%)870.4 (-0.1%)44.0 (-0.5%)
" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.243, + 0.282, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.323, + 0.281, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.243, + 0.484, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.323, + 0.484, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.243, + 0.687, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.323, + 0.687, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.243, + 0.891, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.323, + 0.891, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.406, + 0.782, + 0.421 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of FastMETRO+GTP+ITP@20% on Human3.6M [23] and 3DPW [68]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.566 + ], + "angle": 0, + "content": "These results indicate that the ITP module learns more discriminative features and thus enhances the capability of generalization, allowing methods with ITP to achieve better performance in more challenging in-the-wild scenarios. The strong representation capability of ITP can also be seen in the clustering process, which promotes competitive accuracy with fewer tokens while producing semantically meaningful clustering results, as demonstrated in Sec. 4.7.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.471, + 0.884 + ], + "angle": 0, + "content": "We conduct comparisons with existing SOTA approaches (both Transformer and Non-Transformer based) on monocular human mesh recovery. The results are summarized in Table 5. It is notable that, with effective token reduction techniques, our method produces competitive or higher accuracy on Human3.6M [23] and 3DPW [68] datasets. For instance, on 3DPW, our METRO-H64+GTR achieves \\(87.9\\mathrm{mm}\\) MPVE, \\(75.5\\mathrm{mm}\\) MPJPE and \\(46.6\\mathrm{mm}\\) PAMPJPE surpassing METRO-H64. In general, token reduction results in information loss, which leads to a slight drop in accuracy. For MPVE, using an HRNet-W64 backbone, GTR typically causes a larger error increase of \\(2.6\\mathrm{mm}\\) (from 84.1 to 86.7), while ITP is 1.5 mm (from 86.7 to 88.2). This suggests GTR sacrifices more accuracy for efficiency, while ITP has less impact. Qualitative results of FastMETRO+GTR+ITP@20% are visualized in Figure 4, where our method produces accurate and robust human mesh recovery from monocular images. In the inset, we also provide a qualitative comparison of FastMETRO+GTR+ITP@20% (w/ GTR+ITP) and the baseline model FastMETRO (w/o GTR+ITP)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.74, + 0.61 + ], + "angle": 0, + "content": "where the joint-vertex attention is similar to the blending weights in SMPL, which properly captures the shape structure. However, the model w/o GTR+ITP redundantly correlates local joints with distant vertices, leading to additional interaction costs. In summary, extensive experiments validate the effectiveness of proposed strategies for token reduction across different" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.891, + 0.672 + ], + "angle": 0, + "content": "Transformer structures (Encoder- w/ GTR+ITP w/o GTR+ITP based METRO [39] and Encoder-Decoder-based Fast-METRO [8]) and different Transformer model sizes (Fast-METRO and FastMETRO(S))." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.745, + 0.696 + ], + "angle": 0, + "content": "4.4. ITP v.s. TokenLearner [59]" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.713, + 0.892, + 0.742 + ], + "angle": 0, + "content": "Table 6. Comparison with TokenLearning [59] in ITP for token reduction on 3DPW [68] and Human3.6M [23]." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.763, + 0.892, + 0.814 + ], + "angle": 0, + "content": "
MethodGFLOPs3DPWHuman3.6M
MPVEMPJPEPAMPJPEMPJPEPAMPJPE
TokenLearner [59]5.799.382.452.661.245.4
Image Token Pruning5.399.282.452.359.840.5
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We compare ITP with another popular pruning strategy Tokenlearner [59] for HMR on Human3.6M and 3DPW in Table 6. We use Encoder-Decoder structure [8] with ResNet-50 [20] as a backbone. The pruning rate is \\(20\\%\\). In" + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.447, + 0.884, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15149" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.283, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.089, + 0.485, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.089, + 0.688, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.089, + 0.892, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.169, + 0.282, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.169, + 0.485, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.169, + 0.688, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.169, + 0.892, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.251, + 0.768, + 0.266 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on FreiHAND [85] by FastMETRO+H64+GTR+ITP@20% model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.47, + 0.336 + ], + "angle": 0, + "content": "Table 8, ITP saves more computational costs while achieving the highest accuracy in terms of both human mesh recovery and joint estimation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.345, + 0.338, + 0.361 + ], + "angle": 0, + "content": "4.5. NSR v.s. GCN [27] and MLP" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.52 + ], + "angle": 0, + "content": "We discuss the effectiveness of NSR in GTR. We compare NSR with other implementations, including MultiLayer Perceptron and Graph Convolutional Network [27] following Pose2Mesh [9]. We condition the backbone on ResNet-50 [20] using Encoder-Decoder structure [8]. In Table 7, NSR achieves higher performance on Human3.6M and 3DPW. This indicates that the attention mechanism in NSR provides a stronger modeling capability of vertices given the learned body features, which thus improves the quality of recovered mesh vertices." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.532, + 0.47, + 0.56 + ], + "angle": 0, + "content": "Table 7. Comparison of different network structures of NSR for GTR on 3DPW [68] and Human3.6M [23]." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.57, + 0.47, + 0.634 + ], + "angle": 0, + "content": "
Model3DPWHuman3.6M
MPVEMPJPEPAMPJPEMPJPEPAMPJPE
Multi-Layer Perceptron99.080.449.657.738.9
Graph Convolutional Network98.881.549.858.238.8
Neural Shape Regressor95.979.249.257.238.6
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.67, + 0.421, + 0.687 + ], + "angle": 0, + "content": "4.6. Generalization on Hand Mesh Recovery" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.468, + 0.814 + ], + "angle": 0, + "content": "To investigate the generalizability of our framework, we conduct an experiment on monocular hand mesh recovery, which is summarized in Table 8. Following FastMETRO [8], we report PAMPJPE, F-score@15mm (F@15mm) on FreiHand [85] together with GFLOPs\\(^{\\mathrm{T}}\\) and throughput. The qualitative results are provided in Figure 5. The hand vertex-joint interactions are visualized in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.823, + 0.262, + 0.838 + ], + "angle": 0, + "content": "4.7. Further Discussion" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.364, + 0.863 + ], + "angle": 0, + "content": "4.7.1 Analysis of Image Token Pruning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Image Token Pruning, in essence, achieves body-aware clustering results encoding discriminative body features." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.287, + 0.892, + 0.314 + ], + "angle": 0, + "content": "Table 8. Comparison with the SOTA methods on hand mesh recovery on FreiHAND [85]." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.324, + 0.892, + 0.37 + ], + "angle": 0, + "content": "
Method\\( \\mathrm{{GFLOPs}}{}^{\\mathrm{T}} \\downarrow \\) Throughput \\( \\uparrow \\) PAMPJPE \\( \\downarrow \\mathrm{F}@{15}\\mathrm{\\;{mm}} \\uparrow \\)
METRO-H64 [39]13.1186.76.80.981
FastMETRO-H64 [8]3.3228.46.50.982
FastMETRO-H64+GTR+ITP@20%1.0260.26.70.980
" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.597, + 0.892, + 0.639 + ], + "angle": 0, + "content": "Figure 6. Visualization of learned semantics by ITP. Note that the clustering results are consistent across different identities, i.e., different clusters correspond to different body joints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.892, + 0.768 + ], + "angle": 0, + "content": "We visualize the heatmap of the predicted clustering scores in Figure 6, where the model is with ITP at \\(50\\%\\) pruning rate (resulting in 24 image tokens) and GTR. The backbone is ResNet-50. As shown in Figure 6, clustering-based pruning maps original tokens to a fewer number of clusters with semantics corresponding to the body joints. Note that the semantics is consistent across different identities." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.787, + 0.825, + 0.802 + ], + "angle": 0, + "content": "4.7.2 ITP with Token Reduction Supervision" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Figure 7, we visualize the scores predicted by ITP in HMR. The mask supervision is generated by projecting mesh vertices to image patches using estimated camera parameters as shown in Figure 7 (b)(c). ITP effectively learns to pay attention to a human body region in the given image. Quantitatively, the projection supervision improves PAM" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15150" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.088, + 0.473, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.258, + 0.47, + 0.302 + ], + "angle": 0, + "content": "Figure 7. Visualization of learned semantics by Image Token Pruner. (a) projected joints. (b) projected mesh vertices. (c) mask supervision. (d) scores predicted by ITP." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.313, + 0.47, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.465, + 0.458, + 0.479 + ], + "angle": 0, + "content": "Input R-Elbow R-Knee L-Elbow L-Knee Head" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.485, + 0.47, + 0.514 + ], + "angle": 0, + "content": "Figure 8. Visualization of cross-attention between joint and vertices. Samples are from Human3.6M and 3DPW." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.539, + 0.347, + 0.553 + ], + "angle": 0, + "content": "PJPE from 41.6 to 40.6 on Human3.6M." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.365, + 0.587 + ], + "angle": 0, + "content": "4.7.3 Vertex-Body Feature Interactions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.732 + ], + "angle": 0, + "content": "We show the interactions between vertices and joints modeled by cross attention within Neural Shape Regressor; See Figure 8, where the heatmap is obtained by averaging attention scores across all heads of the multi-head cross-attention between query vertices and body features. As shown in Figure 8, the interactions of mesh vertices and joints are at the body part level, e.g., elbow, knee, head, which are similar to the way of the blending of a human body model, i.e., SMPL [44] and thus validates our claims." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.342, + 0.756 + ], + "angle": 0, + "content": "4.8. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "When an extremely high pruning rate is applied to ITP, the accuracy of the model drops dramatically, e.g., when the token number is pruned to be one, the accuracy of the model drops dramatically (12.1%) from 38.6 to 43.3 PAMPJPE. For GTR, since we regress joints and vertices progressively, the quality of the recovered vertices by NSR depends on the learned body features. More failure cases are in Appendix D. In future work, one of the promising directions could be applying the shown enhanced efficiency of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.183 + ], + "angle": 0, + "content": "HMR from monocular images to methods exhibiting high complexity for improving the model efficiency, especially in tasks such as human-environment/object interaction that perceives environments [37, 22, 18, 75, 73, 19, 61], as well as HMR from videos that involve temporal information [28, 53, 35, 63, 79, 72]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.197, + 0.62, + 0.213 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.223, + 0.893, + 0.405 + ], + "angle": 0, + "content": "In this paper, we investigate the issues of token redundancy in the existing Transformer-based methods for both body and hand mesh recovery tasks. To tackle the problem, we introduce two effective token reduction strategies for Transformers by incorporating insights from both 3D geometry structure and 2D image features. Specifically, we recover the body shape in a hierarchical manner and cluster for image features to feed fewer but more discriminative tokens to the Transformer. Our method dramatically reduces the high-complexity interactions in the Transformer, improving the Pareto-front of accuracy and efficiency. Extensive experiments validate the proposed strategies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.419, + 0.688, + 0.436 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.535 + ], + "angle": 0, + "content": "The authors would like to thank Gene Team for the fruitful discussion and the anonymous reviewers for their valuable comments and suggestions. This research is supported by Meta Reality Labs, Innovation and Technology Commission (Ref: ITS/319/21FP) and Research Grant Council (Ref: 17210222)." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.55, + 0.597, + 0.566 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.575, + 0.893, + 0.645 + ], + "angle": 0, + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.647, + 0.894, + 0.688 + ], + "angle": 0, + "content": "[2] Paolo Baerlocher and Ronan Boulic. Parametrization and range of motion of the ball-and-socket joint. In Deformable avatars, pages 180–190. Springer, 2001." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.69, + 0.894, + 0.731 + ], + "angle": 0, + "content": "[3] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.733, + 0.894, + 0.801 + ], + "angle": 0, + "content": "[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In European conference on computer vision, pages 561-578. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.804, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[5] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, et al. Humman: Multi-modal 4d human dataset for versatile sensing and modeling. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 557-577. Springer, 2022." + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.575, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15151" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.468, + 0.147 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.468, + 0.218 + ], + "angle": 0, + "content": "[7] Xuanyao Chen, Zhijian Liu, Haotian Tang, Li Yi, Hang Zhao, and Song Han. Sparsevit: Revisiting activation sparsity for efficient high-resolution vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2061-2070, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.219, + 0.468, + 0.273 + ], + "angle": 0, + "content": "[8] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3d human mesh recovery with transformers. In European Conference on Computer Vision (ECCV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.468, + 0.329 + ], + "angle": 0, + "content": "[9] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision, pages 769-787. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.468, + 0.371 + ], + "angle": 0, + "content": "[10] MMHuman3D Contributors. Openmmlab 3d human parametric model toolbox and benchmark. https://github.com/open-mmlab/mmhuman3d, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.468, + 0.44 + ], + "angle": 0, + "content": "[11] Haoye Dong, Xiaodan Liang, Xiaohui Shen, Bochao Wang, Hanjiang Lai, Jia Zhu, Zhiting Hu, and Jian Yin. Towards multi-pose guided virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9026-9035, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.468, + 0.523 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.468, + 0.594 + ], + "angle": 0, + "content": "[13] Sai Kumar Dwivedi, Nikos Athanasiou, Muhammed Kocabas, and Michael J Black. Learning to regress bodies from images using differentiable semantic rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11250-11259, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.468, + 0.649 + ], + "angle": 0, + "content": "[14] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[15] Chuang Gan, Yi Yang, Linchao Zhu, Deli Zhao, and Yueting Zhuang. Recognizing an action using its name: A knowledge-based approach. International Journal of Computer Vision, 120(1):61-77, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.747 + ], + "angle": 0, + "content": "[16] Ryan Grainger, Thomas Paniagua, Xi Song, and Tianfu Wu. Learning patch-to-cluster attention in vision transformer. arXiv preprint arXiv:2203.11987, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.468, + 0.803 + ], + "angle": 0, + "content": "[17] Yong Guo, Zhiyang Dou, Nan Zhang, Xiyue Liu, Boni Su, Yuguo Li, and Yinping Zhang. Student close contact behavior and COVID-19 transmission in China's classrooms. PNAS Nexus, 2(5):pgad142, 05 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[18] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[19] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3d scenes" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14708-14718, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[21] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[22] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274-13285, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[23] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[24] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.427, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[25] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7122-7131, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[27] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[28] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5253-5263, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[29] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11127-11137, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[30] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[31] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2261, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[32] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4501-4510, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[33] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren," + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15152" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.162 + ], + "angle": 0, + "content": "Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 620-640. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.471, + 0.234 + ], + "angle": 0, + "content": "[34] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6050-6059, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[35] Gun-Hee Lee and Seong-Whan Lee. Uncertainty-aware human mesh recovery from video by learning part-based 3d dynamics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12375–12384, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[36] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3383-3393, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.471, + 0.43 + ], + "angle": 0, + "content": "[37] Zhi Li, Soshi Shimada, Bernt Schiele, Christian Theobalt, and Vladislav Golyanik. Mocapdeform: Monocular 3d human motion capture in deformable scenes. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.471, + 0.488 + ], + "angle": 0, + "content": "[38] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.471, + 0.545 + ], + "angle": 0, + "content": "[39] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1954-1963, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.471, + 0.589 + ], + "angle": 0, + "content": "[40] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphomer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12939-12948, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.471, + 0.729 + ], + "angle": 0, + "content": "[42] Xiyue Liu, Zhiyang Dou, Lei Wang, Boni Su, Tianyi Jin, Yong Guo, Jianjian Wei, and Nan Zhang. Close contact behavior-based Covid-19 transmission and interventions in a subway system. Journal of Hazardous Materials, 436:129233, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.801 + ], + "angle": 0, + "content": "[43] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[44] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[46] Haoyu Ma, Liangjian Chen, Deying Kong, Zhe Wang, Xingwei Liu, Hao Tang, Xiangyi Yan, Yusheng Xie, Shih-Yao Lin, and Xiaohui Xie. Transfusion: Cross-view fusion with transformer for 3d human pose estimation. arXiv preprint arXiv:2110.09554, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.893, + 0.219 + ], + "angle": 0, + "content": "[47] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: token-pruned pose transformer for monocular and multi-view human pose estimation. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.893, + 0.303 + ], + "angle": 0, + "content": "[48] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 2018 International Conference on 3D Vision (3DV), pages 120-130. IEEE, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.893, + 0.375 + ], + "angle": 0, + "content": "[49] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12309-12318, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.893, + 0.432 + ], + "angle": 0, + "content": "[50] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision, pages 752-768. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.893, + 0.49 + ], + "angle": 0, + "content": "[51] Paschalis Panteleris and Antonis Argyros. Pe-former: Pose estimation transformer. In International Conference on Pattern Recognition and Artificial Intelligence, pages 3-14. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.893, + 0.559 + ], + "angle": 0, + "content": "[52] Georgios Pavlakos, Nikos Kolotouros, and Kostas Daniilidis. Texturepose: Supervising human mesh estimation with texture consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 803-812, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.893, + 0.617 + ], + "angle": 0, + "content": "[53] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human mesh recovery from multiple shots. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1485-1495, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.893, + 0.687 + ], + "angle": 0, + "content": "[54] Georgios Pavlakos, Luyang Zhu, Xiaowei Zhou, and Kostas Daniilidis. Learning to estimate 3d human pose and shape from a single color image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 459-468, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.893, + 0.773 + ], + "angle": 0, + "content": "[55] Leonid Pishchulin, Eldar Insafutdinov, Siyu Tang, Bjoern Andres, Mykhaylo Andriluka, Peter V Gehler, and Bernt Schiele. Deepcut: Joint subset partition and labeling for multi person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4929-4937, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.893, + 0.842 + ], + "angle": 0, + "content": "[56] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[57] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), Nov. 2017." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15153" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[58] Yu Rong, Jingbo Wang, Ziwei Liu, and Chen Change Loy. Monocular 3d reconstruction of interacting hands via collision-aware factorized refinements. In 2021 International Conference on 3D Vision (3DV), pages 432-441. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[59] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[60] Akash Sengupta, Ignas Budvytis, and Roberto Cipolla. Hierarchical kinematic probability distributions for 3d human shape and pose estimation from images in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11219-11229, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[61] Zehong Shen, Zhi Cen, Sida Peng, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Learning human mesh recovery in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17038-17047, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[62] Deepak Sridhar, Niamul Quader, Srikanth Muralidharan, Yaoxin Li, Peng Dai, and Juwei Lu. Class semantics-based attention for action detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13739-13748, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.47, + 0.487 + ], + "angle": 0, + "content": "[63] Zhaoqi Su, Weilin Wan, Tao Yu, Lingjie Liu, Lu Fang, Wenping Wang, and Yebin Liu. Mulaycap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics, 28(4):1862-1879, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.47, + 0.545 + ], + "angle": 0, + "content": "[64] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11179-11188, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.47, + 0.601 + ], + "angle": 0, + "content": "[65] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[66] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A survey. arXiv preprint arXiv:2203.01923, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[67] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[68] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), sep 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[69] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20460-20469, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[70] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on pattern analysis and machine intelligence, 43(10):3349-3364, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.191 + ], + "angle": 0, + "content": "[71] Wenjia Wang, Yongtao Ge, Haiyi Mei, Zhongang Cai, Qingping Sun, Yanjun Wang, Chunhua Shen, Lei Yang, and Taku Komura. Zolly: Zoom focal length correctly for perspective-distorted human mesh reconstruction. arXiv preprint arXiv:2303.13796, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[72] Wen-Li Wei, Jen-Chun Lin, Tyng-Luh Liu, and Hong-Yuan Mark Liao. Capturing humans in motion: Temporal-attentive 3d human pose and shape estimation from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13211–13220, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.894, + 0.333 + ], + "angle": 0, + "content": "[73] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Progressive unsupervised learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2993-3002, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.894, + 0.417 + ], + "angle": 0, + "content": "[74] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B. Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14561-14571, June 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.894, + 0.475 + ], + "angle": 0, + "content": "[75] Yan Xia, Qiangqiang Wu, Wei Li, Antoni B Chan, and Uwe Stilla. A lightweight and detector-free 3d single object tracker on point clouds. IEEE Transactions on Intelligent Transportation Systems, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.894, + 0.559 + ], + "angle": 0, + "content": "[76] Yan Xia, Yusheng Xu, Shuang Li, Rui Wang, Juan Du, Daniel Cremers, and Uwe Stilla. Soe-net: A self-attention and orientation encoding network for point cloud based place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11348–11357, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.894, + 0.617 + ], + "angle": 0, + "content": "[77] Yuanlu Xu, Song-Chun Zhu, and Tony Tung. Denserac: Joint 3d pose and shape estimation by dense render-and-compare. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7760-7770, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.894, + 0.701 + ], + "angle": 0, + "content": "[78] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch on imagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 558-567, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.894, + 0.772 + ], + "angle": 0, + "content": "[79] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11038-11049, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[80] Wang Zeng, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11101-11111, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[81] Yuanhao Zhai, Le Wang, Wei Tang, Qilin Zhang, Junsong Yuan, and Gang Hua. Two-stream consensus network for weakly-supervised temporal action localization. In Euro-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15154" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "pean conference on computer vision, pages 37-54. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.203 + ], + "angle": 0, + "content": "[82] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11446-11456, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.275 + ], + "angle": 0, + "content": "[83] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.468, + 0.344 + ], + "angle": 0, + "content": "[84] Nan Zhang, Li Liu, Zhiyang Dou, Xiyue Liu, Xueze Yang, Doudou Miao, Yong Guo, Silan Gu, Yuguo Li, Hua Qian, and Jianjian Wei. Close contact behaviors of university and school students in 10 indoor environments. Journal of Hazardous Materials, 458:132069, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.347, + 0.468, + 0.415 + ], + "angle": 0, + "content": "[85] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "15155" + } + ] +] \ No newline at end of file diff --git a/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/2d762d0d-bad4-403d-98f7-4210b7f7e39d_origin.pdf b/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/2d762d0d-bad4-403d-98f7-4210b7f7e39d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..53550b29a786ceedbf5dba5a184322a52c0c0fa6 --- /dev/null +++ b/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/2d762d0d-bad4-403d-98f7-4210b7f7e39d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21df6fad53835d30098592e64a1cd6fc44e17ab84fe6619cb70143df88f12719 +size 1025515 diff --git a/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/full.md b/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7d2a51d781e3739cfb84d23e25f09b35005f59a5 --- /dev/null +++ b/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/full.md @@ -0,0 +1,403 @@ +# TORE: Token Reduction for Efficient Human Mesh Recovery with Transformer + +Zhiyang Dou $^{1\dagger}$ Qingxuan Wu $^{2\dagger}$ Cheng Lin $^{3\dagger}$ Zeyu Cao $^{4\dagger}$ Qiangqiang Wu $^{5}$ Weilin Wan $^{1}$ Taku Komura $^{1}$ Wenping Wang $^{6}$ + +$^{1}$ The University of Hong Kong $^{2}$ University of Oxford $^{3}$ Tencent Games $^{4}$ University of Cambridge $^{5}$ City University of Hong Kong $^{6}$ Texas A&M University + +# Abstract + +In this paper, we introduce a set of simple yet effective TOken REDuction (TORE) strategies for Transformer-based Human Mesh Recovery from monocular images. Current SOTA performance is achieved by Transformer-based structures. However, they suffer from high model complexity and computation cost caused by redundant tokens. We propose token reduction strategies based on two important aspects, i.e., the 3D geometry structure and 2D image feature, where we hierarchically recover the mesh geometry with priors from body structure and conduct token clustering to pass fewer but more discriminative image feature tokens to the Transformer. Our method massively reduces the number of tokens involved in high-complexity interactions in the Transformer. This leads to a significantly reduced computational cost while still achieving competitive or even higher accuracy in shape recovery. Extensive experiments across a wide range of benchmarks validate the superior effectiveness of the proposed method. We further demonstrate the generalizability of our method on hand mesh recovery. Visit our project page at https://frank-zy-dou.github.io/projects/Tore/index.html. + +# 1. Introduction + +Human Mesh Recovery (HMR) has been extensively researched in recent years, given its wide real-world applications [15, 81, 11, 62, 69, 17, 42, 84]. This task becomes more challenging when the input is a monocular 2D image, due to the large pose and shape variation, large appearance variation, partial observation, and self-occlusion. + +There has been steady progress in 3D human mesh recovery [30, 50, 36, 24, 60, 39, 40, 8, 5]. Recently, Transformer [67] has shown state-of-the-art (SOTA) results on a wide variety of tasks due to its strong capability of capturing long-range dependency for more accurate predictions [3, 67, 76, 74]. Using tokens constructed from local features extracted by a convolutional neural network + +![](images/c6fd3e346e3f7d408e93da1cbab0be16ccb81da2e55f3508c9e0ef783609ead4.jpg) + +![](images/d807dfac98eb535d4728e112321405cfcdf9915259009386c5d1749ad6f7f942.jpg) +Figure 1. Throughput v.s. Accuracy and GFLOPs v.s. Accuracy on Human3.6M [23]. Our method dramatically saves GFLOPs and improves throughput while maintaining highly competitive accuracy. The x-axis of the bottom GFLOPs figure is reversed for demonstration. Eb0 and R50 represent EfficientNet-b0 [65] and ResNet-50 [20] backbones, respectively. + +
● METRO baseline ▲ FastMETRO baseline ● FastMETRO(S) baseline
METRO-TORE ▼ FastMETRO-TORE ● FastMETRO(S)-TORE
+ +(CNN) [20, 70] to query the joint and mesh vertex positions, Transformer-based methods [39, 40, 8] achieved SOTA performance. + +However, improved performance comes with costs: the increased expressivity of Transformers comes with quadratically increasing computational costs since all pairwise interactions are taken into account [14]. The space and time complexity of a QKV-attention operation is known to be $O(M^2)$ , where $M$ is the number of tokens. The token number thus plays a vital role in time efficiency: a large number of tokens inevitably leads to a heavy computation burden. Unfortunately, almost all the existing Transformer-based SOTA methods for HMR [39, 40, 8] are suffering from redundant tokens. This incurs a high model complexity and computational costs, which prevents the current Transformer-based HMR methods from achieving their full potential in real-world applications. + +In this paper, we make key observations from two important aspects: the 3D geometry structure and 2D image + +feature, to reveal the problem of token redundancy. First, to recover the 3D geometry, all existing methods use both mesh vertices and skeleton joints as tokens for the feature interaction between input and body shape. Whereas a body mesh contains numerous vertices, they can be abstracted by a small number of joints on the skeleton. For instance, when animating a SMPL [44] avatar, the skeleton joints, together with a blend shape binding the joints and corresponding mesh vertices of the local body part, are able to describe various body meshes. Therefore, the joints can already be viewed as an underlying structure of a body shape, which intrinsically encodes the human mesh geometry. Second, for image-based input, most existing methods indiscriminately use all the feature patches to capture pose, shape and appearance variance. However, although the human body exhibits large variance, the important features for shape inference are dominantly clustered within the body area in an RGB image. Most features, e.g., image background, are not informative, thus bringing about redundancy. + +Given the aforementioned insights, we argue that the Pareto-front of accuracy and efficiency for Transformer-based HMR could be further improved by reducing the number of tokens [59, 47, 56]. To this end, we introduce a set of simple yet effective token reduction strategies mainly from two aspects corresponding to our observations. First, for 3D mesh recovery, instead of querying both vertices and joints with input features simultaneously, we consider learning a small set of body tokens at the skeleton level for each body part. To recover corresponding mesh vertices, we use an efficient Neural Shape Regressor (NSR) to infer the mesh from the body features encoded by these tokens. This query process can also be interpreted as an attention matrix decomposition, by which we effectively leverage the geometric insights encoded at the skeleton level to infer the mesh structure hierarchically. Second, for the input image feature, we introduce a learnable token pruner to prune the tokens of patch-based features extracted by a CNN. We employ a clustering-based strategy to identify discriminative features, which results in two appealing properties: 1) the end-to-end learning of the pruner is unsupervised, avoiding the need for additional data labeling; 2) it learns semantically consistent features across various images, thus further benefiting the geometry reasoning and enhancing the capability of generalizability. These token reduction strategies substantially reduce the number of query tokens involved in the computation without sacrificing the important information. An overview is shown in Figure 2. + +We conduct extensive experiments across wide benchmarks [23, 68, 85], including both the human body and hand mesh recovery, to validate the proposed method. Compared to SOTA methods, our framework faithfully recovers body meshes with fewer tokens, which considerably reduces memory and computation overhead while maintain + +ing competitive geometric accuracy. + +In summary, our contribution is three-fold: + +- We reveal the issues of token redundancy in the existing Transformer-based methods for HMR. +- We propose effective strategies for token reduction by incorporating the insights from the 3D geometry structure and 2D image feature into the Transformer design. +- Our method achieves SOTA performance on various benchmarks with less computation cost. For instance, for the Transformer Encoder structure [39] and the Transformer Encoder-Decoder structure [8] with ResNet-50 [20] backbone, our method maintains competitive accuracy while saving $82.9\%$ , $50.5\%$ GFLOPs and improving $139.1\%$ , $39.8\%$ throughput, respectively; see Figure 1 for an overview. + +# 2. Related Work + +Human Mesh Recovery (HMR) from monocular images has achieved great progress in the past years [4, 30, 50, 36, 10, 40, 24, 60, 5, 39, 8, 82, 71, 58]. Given a monocular RGB image, the goal of HMR is to recover the 3D body shape, typically using a human body model, e.g., SMPL [44]. We refer readers to [66] for a comprehensive review. Existing methods fall into two categories: Non-Transformer-based and Transformer-based approaches. + +Non-Transformer-based Human Mesh Recovery To recover the human body shape, one could predict a few parameters, i.e., joint rotation and body shape, to drive the parametric model (parametric approach) or directly regress the vertex positions of the body (non-parametric approach). Parametric approaches SMPLify [4] estimates human pose and shape by fitting SMPL to the detected 2D key points [55] by optimization. Kanazawa et al. [24] adopt adversarial prior knowledge of the 3D body shape into the neural network for HMR. SPIN [30] then combines regression-based and optimization-based methods during the training loop. Intermediate features such as key points [54], pixel-to-surface correspondences [77] and texture consistency [52], have been exploited. HybrIK [36] proposes a hybrid inverse kinematics via twist-and-swing decomposition [2]. Regressing the body parameters from the image is a highly non-linear mapping [50], which thus limits the performance of parametric-based approaches. + +Non-parametric approaches Non-parametric approaches [50, 32, 9] recover spatial coordinates of a body shape directly from the image features. I2L-MeshNet [50] predicts the per-pixel likelihood on 1D heatmaps for vertex coordinates for a trade-off between accuracy and computational cost. GraphCMR [32] and Pose2Mesh [9] employ graph convolutional neural network (GCNN) [27] + +![](images/3214ac64bb94513573c69970a13aa59ff493cbc8bc8344123a8780d69e2899da.jpg) +Figure 2. Overview of the proposed framework. Our goal is to reduce tokens for Transformer Encoder and Decoder which are critical modules in the whole pipeline. Image Token Pruner (ITP) and Neural Shape Regressor (NSR) are two lightweight components. + +to regress mesh positions by modeling local interactions among vertices from the image [32] and 2D pose [9], respectively. Although steady progress has been made, SOTA accuracy is produced by Transformer-based methods. + +Transformer-based Human Mesh Recovery Recently, Transformer structures have been successfully adopted in various vision tasks [12, 43, 6, 51, 66, 80]. In HMR, the SOTA performance is achieved by Transformer-based (attention mechanism) approaches [39, 29, 40, 8]. Specifically, METRO [39] recovers body mesh using a Transformer Encoder [67] to model vertex-vertex and vertex-joint interactions conducting dimension reduction from the image features to 3D shape. MeshGraphormer [40] injects Graph Convolutions [27] into the Transformer Encoder [39] blocks to improve local interactions among vertices. The recent work FastMETRO [8] employs a Transformer Encoder-Decoder architecture disentangling the image encoding and mesh estimation for model reduction and acceleration. However, all Transformer-based approaches [39, 40, 8] suffer from redundant tokens, causing heavy interaction between body geometry and image features, which is still cumbersome and computationally intensive. In this paper, we design strategies driven by the 3D body geometry and 2D image feature for token reduction for both Transformer Encoder and Transformer Encoder-Decoder structures. + +Token Reduction for Transformers Token reduction for Transformer structures has been causing people's attention nowadays [56, 47, 59, 78, 16, 38, 46, 7]. In particular, DynamicViT [56] reduces less informative tokens hierarchically with supervision from a teacher network to save computational costs during inference. TokenLearner [59] generates a smaller number of tokens with spatial attention adaptively. Patch-to-Cluster [16] proposes to cluster over image patches with a token selector. Our method follows this way and adopts a token selection layer for clustering over input tokens, reducing tokens and improving general + +ization capability in challenging scenarios. Token pruning techniques have also been explored in single-view 2D pose estimation and multi-view 3D pose estimation [47] by selecting tokens based on the attention value in the Transformer. Differently, in HMR, besides the key points, body geometry, i.e., mesh vertices, also introduces a burden to the Transformer and needs to be carefully considered. + +# 3. Methodology + +We employ the popular body parametric model SMPL [44] to represent a human body. Formally, given a monocular image $\mathcal{I} \in \mathbb{R}^{224 \times 224}$ , our goal is to recover the joint positions $\mathcal{J} \in \mathbb{R}^{J \times 3}$ and mesh vertices $\mathcal{V} \in \mathbb{R}^{V \times 3}$ where $J = 14$ , $V = 6890$ are the numbers of joints and the mesh vertices, respectively. In addition to the human body, we also demonstrate our method on hand mesh recovery with MANO model [57] to show its generalizability. + +Existing Transformer-based HMR methods fall into two categories: a single Transformer Encoder structure [39, 40] and a Transformer Encoder-Decoder structure [8]. Accordingly, we propose two strategies to reduce tokens from 3D and 2D levels, namely, Geometry Token Reduction (GTR) and Image Token Pruning (ITP), which can accommodate different Transformer structures to improve their efficiency. + +# 3.1. Geometry Token Reduction (GTR) + +As an essential module, all the existing Transformer-based methods for HMR [39, 40, 8] have to model the interaction between body geometry and image features with an attention mechanism. Therefore, an effective token reduction method to improve the efficiency for inferring mesh geometry can benefit the design of Transformer based methods in a wide scope. + +We observe existing methods invariably use both skeleton joints and mesh vertices as tokens for regressing their spatial coordinates. The numerous tokens inevitably lead to + +significant computational costs for a Transformer. In fact, as an underlying structure of the human body, the skeleton joints already provide strong priors for perceiving the body geometry. A human body avatar, like SMPL [44], can be driven by a small number of joints with the blend shape. This inspires us to regress the human body mesh in a hierarchical manner. + +Our key idea is to decompose a heavy Transformer into smaller Transformer modules, each of which involves fewer tokens, avoiding expensive computation for image feature interaction. Specifically, our approach only queries the body tokens, whose count is equivalent to the number of joints. Since the number of joints is much less than the number of mesh vertices, we can significantly reduce the token number in complex 2D-3D feature interactions and this leads to a substantial enhancement in efficiency. + +To faithfully recover body shape from the body tokens, we introduce a Neural Shape Regressor (NSR) module, which is a light Transformer structure, to query the vertex tokens from their interaction with body features. Let $F_{J} = \{f_{1}^{j}, f_{2}^{j}, \dots, f_{J}^{j}\}$ denotes a set of learned body features and $T_{\mathcal{V}} = \{t_{1}^{v}, t_{2}^{v}, \dots, t_{V}^{v}\}$ where $t_{i}^{v}$ represents a set of vertex tokens which are the query tokens. We conduct cross-attention between query vertex tokens and body features to model the interaction between vertices and learned body features. Note that non-adjacent vertices are masked to improve the efficiency [8]. + +The feature size within NSR is even smaller compared with the main Transformer. Thanks to the informative geometric feature encoded in the body tokens, despite not incorporating image tokens, NSR still faithfully recovers the surface vertices. Another interesting discovery is that the learned attention scores reflect the correlation between joints and vertices at a body part level, similar to body blending weights (see Sec. 4.7.3). In this way, we can effectively reduce the cost of redundant interaction by decomposing the entire body into encoded constituent parts and efficiently recovering the whole body shape. + +# 3.2. Image Token Pruning (ITP) + +In addition to the geometric insights, for a Transformer Encoder-Decoder structure, image tokens also affect the computation overhead. However, the existing methods adopting this Transformer structure fail to avoid the token redundancy issue [8] because all available image feature patches are indiscriminately involved in human mesh recovery. Actually, some features, e.g., image background, are not informative, thus introducing redundancy and increasing the computational cost of the Transformer. + +To tackle this issue, we introduce an effective token pruning strategy, namely, Image Token Pruning (ITP). Our key insight is that the informative features in an image for inferring 3D geometry are overwhelmingly clustered within + +the region of the human body. Inspired by recent advances in token pruning [16, 59, 56, 47], we propose to aggregate features into a small number of meaningful clusters. + +Let $0 < \rho < 1$ denote the predefined token pruning ratio. Our goal is to learn a projection that maps feature map $F_{\mathcal{I}} \in R^{HW \times c}$ extracted from the given image with HW tokens to $Z_{\mathcal{I}} \in R^{T \times c}$ with $T$ tokens, where $T = \lfloor \rho HW \rfloor$ and $c$ is the feature dimension. The small number of clusters $Z_{\mathcal{I}}$ are expected to capture the fewer but more discriminative features in $F_{\mathcal{I}}$ . Inspired by [16, 59], we implement the projection as a learnable CNN module. We first apply Conv2D to the input feature, where the kernel size, stride and zero padding of Conv2D are $3 \times 3$ , 1 and 1, reducing the feature dimension from $c$ to $c' = c/4$ with GELU [21] activation function applied. We further map the dimension $c$ to $N$ by MLP and Softmax, producing $M \in R^{HW \times N}$ . Finally, the clustered token set is given by $Z_{\mathcal{I}} = M^T \cdot F_{\mathcal{I}}$ . The mapping matrix $M^T$ , in essence, produces a clustering over origin tokens. Each element $m_{ij} \in M^T$ depicts the contribution of $j$ token to $i$ -th cluster. Note that LayerNorm is employed after the clustering. + +Token Reduction Supervision In order to encourage ITP to pay more attention to those discriminative feature regions, i.e., body parts. during token reduction, we introduce Token Reduction Supervision. Given the weak-perspective camera estimated the neural network, we compute the 2D projection of the ground truth 3D vertices $\hat{\mathcal{V}}$ as $s\Pi (\hat{\mathcal{V}}) + t$ where $\Pi$ represents an orthographic projection, $s$ and $t$ are scale and translation estimated by the network. The projected results are downsampled to a discrete $H\times W$ grid corresponding to each token followed by a binary indicator function + +$$ +F _ {d} (x) = \left\{ \begin{array}{l} 1, \text {i f t h e c e l l c o n t a i n s a p r o j e c t e d p o i n t .} \\ 0, \text {i f t h e c e l l d o e s n o t c o n t a i n a p r o j e c t e d p o i n t .} \end{array} \right. \tag {1} +$$ + +Then the supervision for token pruning is given by + +$$ +L _ {P} = - \frac {1}{N H W} \sum_ {i} ^ {N} \left(F _ {d} \left(s \Pi (\hat {\mathcal {V}}) + t\right) \cdot M _ {[:, i ]}\right). \tag {2} +$$ + +With this supervision, the weights of background tokens in $M[:i]$ for each $i$ -th cluster will be penalized during training, which thus encourages the ITP to learn the discriminative features. + +The proposed ITP has several noticeable advantages. First, in contrast to other methods relying on explicit supervision, e.g., using a pretrained teacher network to facilitate the pruning [56, 78, 33], our token pruner is trained in an unsupervised manner; Note that no ground truth masking is required for Token Reduction Supervision. Second, unlike previous clustering-based pruning methods [16] that use pre-defined class labels, ITP aims to adaptively identify discriminative features within the body region. Therefore, + +ITP is not limited to fixed semantic labels but is able to induce higher-level semantics according to the learning target. + +Finally, we find the clustered informative features learned by ITP improve the model generalizability, especially for challenging datasets, e.g., in-the-wild dataset 3DPW [68] as shown in Table 3. We elaborate on these properties in Sec. 4.3.2 and Sec. 4.7.1 with extensive experiments. + +# 3.3. Loss Functions + +We supervise the network using L1 distance between predicted mesh vertices at three sampling levels: $L_{\mathcal{V}3D} = ||\mathcal{V}_{3D}^l - \hat{\mathcal{V}}^l||_1 + ||\mathcal{V}_{3D}^m - \hat{\mathcal{V}}^m||_1 + ||\mathcal{V}_{3D}^h - \hat{\mathcal{V}}_h||_1$ , where $*, *^m, *^h$ denote low, middle and high body mesh resolution with vertex numbers to be 431, 1723 and 6890 for an SMPL body. The ground-truth 3D joints are used for supervising predicted 3D joints and the ones regressed from the vertices $\mathcal{V}_{3D}^h$ with a SMPL regression matrix $\mathcal{M}$ : $L_{\mathcal{J}3D} = ||\mathcal{J}_{3D} - \hat{\mathcal{J}}_{3D}||_1, L_{\mathcal{J}3D}^R = ||\mathcal{M}(\mathcal{V}) - \hat{\mathcal{J}}_{3D}||_1$ . A 2D projection loss $L_{J2D}$ is used during the network training, where we employ a weak perspective camera model to project 3D joints to 2D for supervision: $L_{\mathcal{J}2D}^R = ||(s\Pi(\mathcal{M}(\mathcal{V})) + t) - \hat{\mathcal{J}}_{2D}||_1$ . $s, t$ are scale and translation estimated by the network. Overall, together with the Token Pruning Supervision $L_P$ , the total loss is: + +$$ +\begin{array}{l} L = \alpha \left[ \lambda_ {\mathcal {J} 3 D} \left(L _ {\mathcal {J} 3 D} ^ {R} + L _ {\mathcal {J} 3 D}\right) + \lambda_ {\mathcal {V} 3 D} \left(L _ {\mathcal {V} 3 D}\right) + \lambda_ {P} L _ {P} \right] \\ + \beta \lambda_ {\mathcal {J} 2 D} L _ {\mathcal {J} 2 D} ^ {R}, \\ \end{array} +$$ + +where $\alpha$ and $\beta$ indicate the availability of the supervision. We set $\lambda_{P},\lambda_{J2D},\lambda_{V3D},\lambda_{J3D}$ to be 1, 100, 100, 1000. + +# 4. Experimental Results + +# 4.1. Datasets and Metrics + +We evaluate our model in two scenarios: human body mesh recovery and hand mesh recovery. We adopt commonly-used metrics [25, 31, 39, 40, 8]: Mean Per-Joint Position Error (MPJPE), MPJPE after further alignment, i.e., Procrustes Analysis (PAMPJPE) and Mean Per-Vertex Error (MPVE). For human body mesh recovery, our network is trained with Human3.6M [23], MuCo-3DHP[48], UP-3D [34], COCO [41], MPI [1]. Following previous works [39, 40, 8], we use the pseudo mesh data in Human3.6M [23] for training, splitting subjects S1, S5, S6, S7, S8 for training and S9, S11 for testing. We also report our performance on 3DPW [68], a more challenging in-the-wild dataset. To further evaluate the generalization ability of the proposed method, we test our method on hand mesh recovery on FreiHAND [85] dataset. For analyzing the efficiency of our method, we report GFLOPs and throughput (image per second), strictly following [56, 49, 59, 47, 38]. + +Table 1. Comparison with the Transformer Encoder structure METRO (M) [39] on Human3.6M [23]. We test with ResNet-50 (R50) [20] and HRNet-W64 (H64) [70] as backbones. GFLOPs $^{\mathrm{T}}$ is GFLOPs of the transformer. + +
MethodGFLOPs ↓GFLOPsT ↓Throughput (im/s) ↑PAMPJPE ↓
M-H64 [39]56.527.5141.036.7
M-H64+GTR30.2 (-46.5%)0.8 (-97.1%)210.1 (+49.0%)37.1 (+1.1%)
M-R50 [39]31.627.5247.040.6
M-R50+GTR5.4 (-82.9%)0.8 (-97.1%)590.6 (+139.1%)42.0 (+3.4%)
+ +# 4.2. Implementation Details + +Network Training We implement the network using PyTorch. For the Transformer Encoder-Decoder structure, we set the learning rate to be $1 \times 10^{-4}$ . We use the AdamW optimizer [45] and train for 60 epochs, with a batch size of 16 per GPU on 4 Nvidia A100 GPUs. When comparing with Transformer Encoder structure METRO [39] (see Table 1), we follow the setting of METRO where we train the models with a batch size of 30 per GPU on 8 Nvidia A100 GPUs in total. We adopt Adam optimizer [26] and train the models for 200 epochs. See more details in Appendix A. + +Performance Evaluation To analyze the performance on a consumer-level GPU device, we measure the throughput on an NVIDIA RTX 3090 GPU with 24G VRAM. When comparing with encoder-decoder Transformer structure FastMETRO [8], we set the batch size to 32 following PPT [47]. The comparison with Transformer Encoder structure METRO [39] uses 16 as the batch size. Note that the batch size is limited by the size of the original METRO [39] model. To factor out the influence of the batch size, we provide a more detailed performance report in Appendix B. + +# 4.3. Performance on Human Mesh Recovery + +Currently, there only exist two types of representative Transformer structures for HMR: Encoder-only structure (METRO [39]) and Encoder-Decoder structure (Fast-METRO [8]). To demonstrate the effectiveness of our method to both Transformer structures, we first conduct experiments by adding GTR to two structures: METRO and FastMETRO. Since ITP is designed for the encoder-decoder structure, we further report the results on FastMETRO with both ITP and GTR in Sec. 4.3.2. The overall comparison is shown in Table 5. + +# 4.3.1 Transformer Encoder Structure + +As shown in Table 1, GTR effectively reduces the computation costs while still producing competitive accuracy results. For the encoder-based Transformer model METRO [39] with HRNet-W64 [70] as a backbone, applying GTR saves the GFLOPs of the whole model for $46.5\%$ and the Transformer part for $97.1\%$ , with throughput im + +![](images/c72ac0f1fe60fbb90346be7b9d8e80d2a7426fad6767b10662b67670932e94e2.jpg) + +![](images/6676cc78dad18567d6712dc6dac93b1e8c5de2febd20136c471f9af99917d57b.jpg) + +![](images/937cc42f1d112f837be254d8251447a65fbf875c0f437235c9f5adf65b2a1ec5.jpg) + +![](images/53f6f812e9fb9899cd68417a64be2476d1903201d6bca5d307f192fe04706e42.jpg) + +![](images/2b85f6a8c29ca2a1d742d96eddacd23191c9a2bcf11861871627f1d09e94b8d1.jpg) + +![](images/6bc09f241d8d32243e21f229da1894d05931658d2b3be1f676eafe9c0919f6f4.jpg) + +![](images/d50d40c956529558d6af8c3f6a4b5913652bba5873391472df7dbeb5aceb14c5.jpg) +Figure 3. Qualitative results of GTR equipped Encoder-Decoder structure [39] (H64) on Human3.6M [23] and 3DPW [68]. + +![](images/3cc684d60fb5b007ce93d3b584e9d23cf924864c11a0b5456af91da231c16a6c.jpg) + +Table 2. Comparison with the Transformer Encoder-Decoder structure FastMETRO (FM) [8] on Human3.6M [23]. We test with EfficientNet-b0 (Eb0) [65], ResNet-50 (R50) [20] and HRNet-W64 (H64) [70] as backbones. GFLOPs $^{\mathrm{T}}$ stand for GFLOPs for the transformer. + +
MethodGFLOPs ↓GFLOPsT ↓Throughput (im/s) ↑PAMPJPE ↓
FM-H64 [8]35.76.6221.533.7
FM-H64+GTR30.2 (-15.4%)0.7 (-89.4%)249.2 (+12.5%)34.8 (+3.2%)
FM-R50 [8]10.96.6576.037.3
FM-R50+GTR5.4 (-50.5%)0.7 (-89.4%)805.3 (+39.8%)38.6 (+3.4%)
FM(S)-R50 [8]6.42.2953.539.4
FM(S)-R50+GTR4.6 (-28.1%)0.3 (-86.4%)1128.9 (+18.4%)38.6 (-2.0%)
FM-Eb07.16.6517.645.8
FM-Eb0+GTR1.7 (-76.1%)0.7 (-89.4%)870.5 (+68.2%)44.2 (-3.5%)
+ +proved by $49\%$ . For the ResNet-50 [20] backbone, GTR helps save the GFLOPs of the whole model for $82.9\%$ and the Transformer part for $97.1\%$ , with throughput improved by $139.1\%$ . These results validate the effectiveness of the proposed GTR on the Transformer Encoder structure. Qualitative results of METRO-H64+GTR can be found in Figure 3, in which the model produces high-quality results in human mesh recovery over various input monocular images. + +# 4.3.2 Transformer Encoder-Decoder Structure + +We experiment on the Transformer Encoder-Decoder structure FastMETRO [8] with its two variants: models with 1 and 3 Encoder-Decoder layers. For 1-layer FastMETRO, we denote it as FastMETRO(S). In Table 2, GTR equipped + +FastMETRO [39] produces competitive accuracy while reducing $15.4\%$ in GFLOPs for the whole model, $89.4\%$ in GFLOPs for the Transformer part and improving $12.5\%$ in throughput with an HRNet-W64 [70] CNN backbone. When testing with ResNet-50 [20], GTR helps save $50.5\%$ in GFLOPs for the whole model, $89.4\%$ in GFLOPs for the Transformer part and improves $39.8\%$ in throughput. Notably, our FastMETRO(S)-R50+GTR and FastMETRO(S)-Eb0+GTR yield 38.6, 44.2 mm PAMPJPE respectively, surpassing the corresponding baselines while greatly saving computational costs. + +We then present the performance of the Transformer Encoder-Decoder structure with GTR and ITP using ResNet-50 [20] and EfficientNet-b0 [65] CNN backbones. + +Table 3. Influence of ITP for monocular 3D human mesh recovery on 3DPW [68]. + +
MethodMPVEMPJPEPAMPJPE
FastMETRO-H64+GTR91.375.446.7
FastMETRO-H64+GTR+ITP@20%88.272.344.4
+ +As shown in Table 4, TORE (GTR+ITP) is effective for the Encoder-Decoder Transformer structure. Specifically, when pruning with $20\%$ , $50\%$ in ITP, the models save the GLOPs of the Transformer structure by $14.3\%$ , $14.3\%$ and $28.5\%$ , $42.9\%$ while receiving competitive accuracy or even higher accuracy i.e., FM-Eb0+GTR+ITP@ $20\%$ improves PAMPJPE compared with the baseline. + +Table 5. Comparison with the SOTA methods for monocular 3D human mesh recovery on 3DPW [68] and Human3.6M [23]. + +
Method3DPWHuman3.6M
MPVEMPJPEPAMPJPEMPJPEPAMPJPE
HMR [25]-130.076.788.056.8
GraphCMR [32]--70.2-50.1
SPIN [30]116.496.959.262.541.1
I2LMeshNet [50]-93.257.755.741.1
PyMAF [83]110.192.858.957.740.5
ROMP-R50 [64]105.689.353.5--
PARE-R50 [29]99.782.952.3--
DSR-R50 [13]99.585.751.760.940.3
METRO-R50 [39]---56.540.6
METRO-H64 [39]88.277.147.954.036.7
METRO-H64+GTR87.975.546.657.637.1
FastMETRO-R50 [8]90.677.948.353.937.3
FastMETRO-R50+GTR+ITP@20%99.282.452.359.840.5
FastMETRO-H64 [8]84.173.544.652.233.7
FastMETRO-H64+GTR+ITP@20%88.272.344.459.636.4
FastMETRO-Eb0 [8]112.593.860.269.245.8
FastMETRO-Eb0+GTR+ITP@20%112.593.760.163.243.9
+ +In addition, we find that ITP helps improve the accuracy on the challenging in-the-wild 3DPW [68] dataset, shown in Table 3. Specifically, when further equipped with ITP, the model performance in MPVE, MPJPE and PAMPJPE are improved by $3.1\mathrm{mm}$ , $3.1\mathrm{mm}$ and $2.3\mathrm{mm}$ , respectively. + +Table 4. Statistics of all proposed components (GTR + ITP) for Encoder-Decoder structure FastMETRO (FM) [8] on Human3.6M [23]. The backbones are EfficientNet-b0 (Eb0) [65] and ResNet-50 (R50) [20]. GFLOPs $^{\mathrm{T}}$ stands for GFLOPs of the Transformer. + +
Method + Pruning Rate#TokensGFLOPs ↓GFLOPsT ↓Throughput (im/s) ↑PAMPJPE ↓
FM-R50+GTR495.40.7805.338.6
FM-R50+GTR+ITP@20%395.3 (-1.8%)0.6 (-14.3%)794.7 (-1.3%)40.5 (+4.9%)
FM-R50+GTR+ITP@50%245.1 (-5.5%)0.5 (-28.5%)806.1 (+0.1%)40.7 (+5.4%)
FM-Eb0+GTR491.70.7870.544.2
FM-Eb0+GTR+ITP@20%391.6 (-5.9%)0.6 (-14.3%)876.3 (+0.7%)43.9 (-0.%)
FM-Eb0+GTR+ITP@50%241.4 (-17.6%)0.4 (-42.9%)870.4 (-0.1%)44.0 (-0.5%)
+ +![](images/cd7d259b3a4cc5ec1d2550a7f956935ea521df3a69c821beeb1b6abe86c3bb36.jpg) + +![](images/447312e7e09283642e1ed1d103858963b70c2f4e84fddf848c121f9c91832c5e.jpg) +Figure 4. Qualitative results of FastMETRO+GTP+ITP@20% on Human3.6M [23] and 3DPW [68]. + +![](images/026c783d04408d6546c7bc6e26b8cabce8c3bd2f498696141238b3e6398cdba6.jpg) + +![](images/946903e54861b1c25d1239531e680fc684e787f6db8bfdb2a85e6e961fa79315.jpg) + +![](images/839ec8ee3ca390e05b2e3871156819dcdc0e886108d9453e28aa494b2cd44e92.jpg) + +![](images/11554b0651fe67d123710420d241da859bf90726756bd50d4403c74cf77eb6ae.jpg) + +![](images/dbcbfe70555b0bead4b216f4abc7ad20b29cadf163692c197ceb1108cd237f10.jpg) + +![](images/bd2865b51f78814e662d1217c64f530a25bcd5ed8852928b0dc9177b8c37ae76.jpg) + +These results indicate that the ITP module learns more discriminative features and thus enhances the capability of generalization, allowing methods with ITP to achieve better performance in more challenging in-the-wild scenarios. The strong representation capability of ITP can also be seen in the clustering process, which promotes competitive accuracy with fewer tokens while producing semantically meaningful clustering results, as demonstrated in Sec. 4.7.1. + +We conduct comparisons with existing SOTA approaches (both Transformer and Non-Transformer based) on monocular human mesh recovery. The results are summarized in Table 5. It is notable that, with effective token reduction techniques, our method produces competitive or higher accuracy on Human3.6M [23] and 3DPW [68] datasets. For instance, on 3DPW, our METRO-H64+GTR achieves $87.9\mathrm{mm}$ MPVE, $75.5\mathrm{mm}$ MPJPE and $46.6\mathrm{mm}$ PAMPJPE surpassing METRO-H64. In general, token reduction results in information loss, which leads to a slight drop in accuracy. For MPVE, using an HRNet-W64 backbone, GTR typically causes a larger error increase of $2.6\mathrm{mm}$ (from 84.1 to 86.7), while ITP is 1.5 mm (from 86.7 to 88.2). This suggests GTR sacrifices more accuracy for efficiency, while ITP has less impact. Qualitative results of FastMETRO+GTR+ITP@20% are visualized in Figure 4, where our method produces accurate and robust human mesh recovery from monocular images. In the inset, we also provide a qualitative comparison of FastMETRO+GTR+ITP@20% (w/ GTR+ITP) and the baseline model FastMETRO (w/o GTR+ITP), + +where the joint-vertex attention is similar to the blending weights in SMPL, which properly captures the shape structure. However, the model w/o GTR+ITP redundantly correlates local joints with distant vertices, leading to additional interaction costs. In summary, extensive experiments validate the effectiveness of proposed strategies for token reduction across different + +Transformer structures (Encoder- w/ GTR+ITP w/o GTR+ITP based METRO [39] and Encoder-Decoder-based Fast-METRO [8]) and different Transformer model sizes (Fast-METRO and FastMETRO(S)). + +# 4.4. ITP v.s. TokenLearner [59] + +Table 6. Comparison with TokenLearning [59] in ITP for token reduction on 3DPW [68] and Human3.6M [23]. + +
MethodGFLOPs3DPWHuman3.6M
MPVEMPJPEPAMPJPEMPJPEPAMPJPE
TokenLearner [59]5.799.382.452.661.245.4
Image Token Pruning5.399.282.452.359.840.5
+ +We compare ITP with another popular pruning strategy Tokenlearner [59] for HMR on Human3.6M and 3DPW in Table 6. We use Encoder-Decoder structure [8] with ResNet-50 [20] as a backbone. The pruning rate is $20\%$ . In + +![](images/13333b54e4bf922d542975797abbb2441bedec462c18fb5b7ec7a5b6300e280a.jpg) + +![](images/a82b688f14eabb1b9abc281376466415eb929411338a0fed2d33583d20ea8626.jpg) + +![](images/4cb2e743f5cd93c723a94d3ade0c3b2baefe0cc076ddb267ea71e2e9782cef19.jpg) + +![](images/0808a5a7604bb6d31de8de63e26009c6d52088476f90a7affae7351627a9a40f.jpg) + +![](images/0b94ffb196aa260807334e7493627359a983ae801a5cfc48d1f5727571b9fd91.jpg) + +![](images/a58e49d5c51325fe02ac4f8a10df0d57d82993c21c1886b4aeb97f5e6c8332fa.jpg) +Figure 5. Qualitative results on FreiHAND [85] by FastMETRO+H64+GTR+ITP@20% model. + +![](images/a79a07f186f6410fd460ab244e76cd6f47964fff8edee0a83f4a12f308c4f474.jpg) + +![](images/12ec44111f8919a2f3c4a696ae2af2475487e668e72a0417ac1fdd9af0026d0d.jpg) + +![](images/0525dc7d8c878ec74c33ab57bae424ba84c5241ab106b1940a8aa949f80c7542.jpg) + +Table 8, ITP saves more computational costs while achieving the highest accuracy in terms of both human mesh recovery and joint estimation. + +# 4.5. NSR v.s. GCN [27] and MLP + +We discuss the effectiveness of NSR in GTR. We compare NSR with other implementations, including MultiLayer Perceptron and Graph Convolutional Network [27] following Pose2Mesh [9]. We condition the backbone on ResNet-50 [20] using Encoder-Decoder structure [8]. In Table 7, NSR achieves higher performance on Human3.6M and 3DPW. This indicates that the attention mechanism in NSR provides a stronger modeling capability of vertices given the learned body features, which thus improves the quality of recovered mesh vertices. + +Table 7. Comparison of different network structures of NSR for GTR on 3DPW [68] and Human3.6M [23]. + +
Model3DPWHuman3.6M
MPVEMPJPEPAMPJPEMPJPEPAMPJPE
Multi-Layer Perceptron99.080.449.657.738.9
Graph Convolutional Network98.881.549.858.238.8
Neural Shape Regressor95.979.249.257.238.6
+ +# 4.6. Generalization on Hand Mesh Recovery + +To investigate the generalizability of our framework, we conduct an experiment on monocular hand mesh recovery, which is summarized in Table 8. Following FastMETRO [8], we report PAMPJPE, F-score@15mm (F@15mm) on FreiHand [85] together with GFLOPs $^{\mathrm{T}}$ and throughput. The qualitative results are provided in Figure 5. The hand vertex-joint interactions are visualized in Appendix C. + +# 4.7. Further Discussion + +# 4.7.1 Analysis of Image Token Pruning + +Image Token Pruning, in essence, achieves body-aware clustering results encoding discriminative body features. + +Table 8. Comparison with the SOTA methods on hand mesh recovery on FreiHAND [85]. + +
Method\( \mathrm{{GFLOPs}}{}^{\mathrm{T}} \downarrow \) Throughput \( \uparrow \) PAMPJPE \( \downarrow \mathrm{F}@{15}\mathrm{\;{mm}} \uparrow \)
METRO-H64 [39]13.1186.76.80.981
FastMETRO-H64 [8]3.3228.46.50.982
FastMETRO-H64+GTR+ITP@20%1.0260.26.70.980
+ +![](images/981d7bad6fc317071ac01a34e788242e9e05512f5c3ac3d0181418b4056ad828.jpg) +Figure 6. Visualization of learned semantics by ITP. Note that the clustering results are consistent across different identities, i.e., different clusters correspond to different body joints. + +We visualize the heatmap of the predicted clustering scores in Figure 6, where the model is with ITP at $50\%$ pruning rate (resulting in 24 image tokens) and GTR. The backbone is ResNet-50. As shown in Figure 6, clustering-based pruning maps original tokens to a fewer number of clusters with semantics corresponding to the body joints. Note that the semantics is consistent across different identities. + +# 4.7.2 ITP with Token Reduction Supervision + +In Figure 7, we visualize the scores predicted by ITP in HMR. The mask supervision is generated by projecting mesh vertices to image patches using estimated camera parameters as shown in Figure 7 (b)(c). ITP effectively learns to pay attention to a human body region in the given image. Quantitatively, the projection supervision improves PAM + +![](images/fc54262c033eabde71b001267b344aab5970e91b012da652442cb257b20f2a9b.jpg) +Figure 7. Visualization of learned semantics by Image Token Pruner. (a) projected joints. (b) projected mesh vertices. (c) mask supervision. (d) scores predicted by ITP. + +![](images/4d8665f371240e1a786092e55e385b39a9e7f13f3bb6e64c7b0eb002872b5d06.jpg) +Input R-Elbow R-Knee L-Elbow L-Knee Head +Figure 8. Visualization of cross-attention between joint and vertices. Samples are from Human3.6M and 3DPW. + +PJPE from 41.6 to 40.6 on Human3.6M. + +# 4.7.3 Vertex-Body Feature Interactions + +We show the interactions between vertices and joints modeled by cross attention within Neural Shape Regressor; See Figure 8, where the heatmap is obtained by averaging attention scores across all heads of the multi-head cross-attention between query vertices and body features. As shown in Figure 8, the interactions of mesh vertices and joints are at the body part level, e.g., elbow, knee, head, which are similar to the way of the blending of a human body model, i.e., SMPL [44] and thus validates our claims. + +# 4.8. Limitations and Future Work + +When an extremely high pruning rate is applied to ITP, the accuracy of the model drops dramatically, e.g., when the token number is pruned to be one, the accuracy of the model drops dramatically (12.1%) from 38.6 to 43.3 PAMPJPE. For GTR, since we regress joints and vertices progressively, the quality of the recovered vertices by NSR depends on the learned body features. More failure cases are in Appendix D. In future work, one of the promising directions could be applying the shown enhanced efficiency of + +HMR from monocular images to methods exhibiting high complexity for improving the model efficiency, especially in tasks such as human-environment/object interaction that perceives environments [37, 22, 18, 75, 73, 19, 61], as well as HMR from videos that involve temporal information [28, 53, 35, 63, 79, 72]. + +# 5. Conclusion + +In this paper, we investigate the issues of token redundancy in the existing Transformer-based methods for both body and hand mesh recovery tasks. To tackle the problem, we introduce two effective token reduction strategies for Transformers by incorporating insights from both 3D geometry structure and 2D image features. Specifically, we recover the body shape in a hierarchical manner and cluster for image features to feed fewer but more discriminative tokens to the Transformer. Our method dramatically reduces the high-complexity interactions in the Transformer, improving the Pareto-front of accuracy and efficiency. Extensive experiments validate the proposed strategies. + +# 6. Acknowledgements + +The authors would like to thank Gene Team for the fruitful discussion and the anonymous reviewers for their valuable comments and suggestions. This research is supported by Meta Reality Labs, Innovation and Technology Commission (Ref: ITS/319/21FP) and Research Grant Council (Ref: 17210222). + +# References + +[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. +[2] Paolo Baerlocher and Ronan Boulic. Parametrization and range of motion of the ball-and-socket joint. In Deformable avatars, pages 180–190. Springer, 2001. +[3] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. +[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In European conference on computer vision, pages 561-578. Springer, 2016. +[5] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, et al. Humman: Multi-modal 4d human dataset for versatile sensing and modeling. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 557-577. Springer, 2022. + +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. +[7] Xuanyao Chen, Zhijian Liu, Haotian Tang, Li Yi, Hang Zhao, and Song Han. Sparsevit: Revisiting activation sparsity for efficient high-resolution vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2061-2070, 2023. +[8] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3d human mesh recovery with transformers. In European Conference on Computer Vision (ECCV), 2022. +[9] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision, pages 769-787. Springer, 2020. +[10] MMHuman3D Contributors. Openmmlab 3d human parametric model toolbox and benchmark. https://github.com/open-mmlab/mmhuman3d, 2021. +[11] Haoye Dong, Xiaodan Liang, Xiaohui Shen, Bochao Wang, Hanjiang Lai, Jia Zhu, Zhiting Hu, and Jian Yin. Towards multi-pose guided virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9026-9035, 2019. +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[13] Sai Kumar Dwivedi, Nikos Athanasiou, Muhammed Kocabas, and Michael J Black. Learning to regress bodies from images using differentiable semantic rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11250-11259, 2021. +[14] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. +[15] Chuang Gan, Yi Yang, Linchao Zhu, Deli Zhao, and Yueting Zhuang. Recognizing an action using its name: A knowledge-based approach. International Journal of Computer Vision, 120(1):61-77, 2016. +[16] Ryan Grainger, Thomas Paniagua, Xi Song, and Tianfu Wu. Learning patch-to-cluster attention in vision transformer. arXiv preprint arXiv:2203.11987, 2022. +[17] Yong Guo, Zhiyang Dou, Nan Zhang, Xiyue Liu, Boni Su, Yuguo Li, and Yinping Zhang. Student close contact behavior and COVID-19 transmission in China's classrooms. PNAS Nexus, 2(5):pgad142, 05 2023. +[18] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. +[19] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3d scenes + +by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14708-14718, 2021. +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[21] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. +[22] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274-13285, 2022. +[23] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013. +[24] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. +[25] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7122-7131, 2018. +[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. +[27] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. +[28] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5253-5263, 2020. +[29] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11127-11137, 2021. +[30] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019. +[31] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2261, 2019. +[32] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4501-4510, 2019. +[33] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, + +Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 620-640. Springer, 2022. +[34] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6050-6059, 2017. +[35] Gun-Hee Lee and Seong-Whan Lee. Uncertainty-aware human mesh recovery from video by learning part-based 3d dynamics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12375–12384, 2021. +[36] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3383-3393, 2021. +[37] Zhi Li, Soshi Shimada, Bernt Schiele, Christian Theobalt, and Vladislav Golyanik. Mocapdeform: Monocular 3d human motion capture in deformable scenes. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. +[38] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022. +[39] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1954-1963, 2021. +[40] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphomer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12939-12948, 2021. +[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. +[42] Xiyue Liu, Zhiyang Dou, Lei Wang, Boni Su, Tianyi Jin, Yong Guo, Jianjian Wei, and Nan Zhang. Close contact behavior-based Covid-19 transmission and interventions in a subway system. Journal of Hazardous Materials, 436:129233, 2022. +[43] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. +[44] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. +[45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. + +[46] Haoyu Ma, Liangjian Chen, Deying Kong, Zhe Wang, Xingwei Liu, Hao Tang, Xiangyi Yan, Yusheng Xie, Shih-Yao Lin, and Xiaohui Xie. Transfusion: Cross-view fusion with transformer for 3d human pose estimation. arXiv preprint arXiv:2110.09554, 2021. +[47] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: token-pruned pose transformer for monocular and multi-view human pose estimation. In ECCV, 2022. +[48] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 2018 International Conference on 3D Vision (3DV), pages 120-130. IEEE, 2018. +[49] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12309-12318, 2022. +[50] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision, pages 752-768. Springer, 2020. +[51] Paschalis Panteleris and Antonis Argyros. Pe-former: Pose estimation transformer. In International Conference on Pattern Recognition and Artificial Intelligence, pages 3-14. Springer, 2022. +[52] Georgios Pavlakos, Nikos Kolotouros, and Kostas Daniilidis. Texturepose: Supervising human mesh estimation with texture consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 803-812, 2019. +[53] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human mesh recovery from multiple shots. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1485-1495, 2022. +[54] Georgios Pavlakos, Luyang Zhu, Xiaowei Zhou, and Kostas Daniilidis. Learning to estimate 3d human pose and shape from a single color image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 459-468, 2018. +[55] Leonid Pishchulin, Eldar Insafutdinov, Siyu Tang, Bjoern Andres, Mykhaylo Andriluka, Peter V Gehler, and Bernt Schiele. Deepcut: Joint subset partition and labeling for multi person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4929-4937, 2016. +[56] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. +[57] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), Nov. 2017. + +[58] Yu Rong, Jingbo Wang, Ziwei Liu, and Chen Change Loy. Monocular 3d reconstruction of interacting hands via collision-aware factorized refinements. In 2021 International Conference on 3D Vision (3DV), pages 432-441. IEEE, 2021. +[59] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. 2021. +[60] Akash Sengupta, Ignas Budvytis, and Roberto Cipolla. Hierarchical kinematic probability distributions for 3d human shape and pose estimation from images in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11219-11229, 2021. +[61] Zehong Shen, Zhi Cen, Sida Peng, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Learning human mesh recovery in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17038-17047, 2023. +[62] Deepak Sridhar, Niamul Quader, Srikanth Muralidharan, Yaoxin Li, Peng Dai, and Juwei Lu. Class semantics-based attention for action detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13739-13748, 2021. +[63] Zhaoqi Su, Weilin Wan, Tao Yu, Lingjie Liu, Lu Fang, Wenping Wang, and Yebin Liu. Mulaycap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics, 28(4):1862-1879, 2020. +[64] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11179-11188, 2021. +[65] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. +[66] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A survey. arXiv preprint arXiv:2203.01923, 2022. +[67] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. +[68] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), sep 2018. +[69] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20460-20469, 2022. +[70] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions + +on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. +[71] Wenjia Wang, Yongtao Ge, Haiyi Mei, Zhongang Cai, Qingping Sun, Yanjun Wang, Chunhua Shen, Lei Yang, and Taku Komura. Zolly: Zoom focal length correctly for perspective-distorted human mesh reconstruction. arXiv preprint arXiv:2303.13796, 2023. +[72] Wen-Li Wei, Jen-Chun Lin, Tyng-Luh Liu, and Hong-Yuan Mark Liao. Capturing humans in motion: Temporal-attentive 3d human pose and shape estimation from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13211–13220, 2022. +[73] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Progressive unsupervised learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2993-3002, June 2021. +[74] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B. Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14561-14571, June 2023. +[75] Yan Xia, Qiangqiang Wu, Wei Li, Antoni B Chan, and Uwe Stilla. A lightweight and detector-free 3d single object tracker on point clouds. IEEE Transactions on Intelligent Transportation Systems, 2023. +[76] Yan Xia, Yusheng Xu, Shuang Li, Rui Wang, Juan Du, Daniel Cremers, and Uwe Stilla. Soe-net: A self-attention and orientation encoding network for point cloud based place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11348–11357, 2021. +[77] Yuanlu Xu, Song-Chun Zhu, and Tony Tung. Denserac: Joint 3d pose and shape estimation by dense render-and-compare. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7760-7770, 2019. +[78] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch on imagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 558-567, 2021. +[79] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11038-11049, 2022. +[80] Wang Zeng, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11101-11111, 2022. +[81] Yuanhao Zhai, Le Wang, Wei Tang, Qilin Zhang, Junsong Yuan, and Gang Hua. Two-stream consensus network for weakly-supervised temporal action localization. In Euro- + +pean conference on computer vision, pages 37-54. Springer, 2020. +[82] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11446-11456, 2021. +[83] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE International Conference on Computer Vision, 2021. +[84] Nan Zhang, Li Liu, Zhiyang Dou, Xiyue Liu, Xueze Yang, Doudou Miao, Yong Guo, Silan Gu, Yuguo Li, Hua Qian, and Jianjian Wei. Close contact behaviors of university and school students in 10 indoor environments. Journal of Hazardous Materials, 458:132069, 2023. +[85] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019. \ No newline at end of file diff --git a/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/images.zip b/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..628987b604d885fca0dcaa81355c2255cb9e0ae0 --- /dev/null +++ b/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a1fa2c55c4b95733442a609afe7920b99f4eb1e4c74b18d782bec22cdc966c7 +size 797360 diff --git a/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/layout.json b/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4e6e59d5ce10705adaef28558edd931678fe8bc3 --- /dev/null +++ b/2023/TORE_ Token Reduction for Efficient Human Mesh Recovery with Transformer/layout.json @@ -0,0 +1,11370 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 50, + 103, + 544, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 103, + 544, + 120 + ], + "spans": [ + { + "bbox": [ + 50, + 103, + 544, + 120 + ], + "type": "text", + "content": "TORE: Token Reduction for Efficient Human Mesh Recovery with Transformer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "spans": [ + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "text", + "content": "Zhiyang Dou" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "text", + "content": " Qingxuan Wu" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "text", + "content": " Cheng Lin" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "text", + "content": " Zeyu Cao" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "text", + "content": " Qiangqiang Wu" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "text", + "content": " Weilin Wan" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "text", + "content": " Taku Komura" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "text", + "content": " Wenping Wang" + }, + { + "bbox": [ + 91, + 141, + 498, + 177 + ], + "type": "inline_equation", + "content": "^{6}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "spans": [ + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "type": "text", + "content": "The University of Hong Kong " + }, + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "type": "text", + "content": "University of Oxford " + }, + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "type": "text", + "content": "Tencent Games " + }, + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "type": "text", + "content": "University of Cambridge " + }, + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "type": "text", + "content": "City University of Hong Kong " + }, + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 84, + 180, + 509, + 211 + ], + "type": "text", + "content": "Texas A&M University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 237, + 192, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 237, + 192, + 250 + ], + "spans": [ + { + "bbox": [ + 143, + 237, + 192, + 250 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 262, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 290, + 515 + ], + "type": "text", + "content": "In this paper, we introduce a set of simple yet effective TOken REDuction (TORE) strategies for Transformer-based Human Mesh Recovery from monocular images. Current SOTA performance is achieved by Transformer-based structures. However, they suffer from high model complexity and computation cost caused by redundant tokens. We propose token reduction strategies based on two important aspects, i.e., the 3D geometry structure and 2D image feature, where we hierarchically recover the mesh geometry with priors from body structure and conduct token clustering to pass fewer but more discriminative image feature tokens to the Transformer. Our method massively reduces the number of tokens involved in high-complexity interactions in the Transformer. This leads to a significantly reduced computational cost while still achieving competitive or even higher accuracy in shape recovery. Extensive experiments across a wide range of benchmarks validate the superior effectiveness of the proposed method. We further demonstrate the generalizability of our method on hand mesh recovery. Visit our project page at https://frank-zy-dou.github.io/projects/Tore/index.html." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 521, + 128, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 128, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 128, + 533 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 540, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 612 + ], + "type": "text", + "content": "Human Mesh Recovery (HMR) has been extensively researched in recent years, given its wide real-world applications [15, 81, 11, 62, 69, 17, 42, 84]. This task becomes more challenging when the input is a monocular 2D image, due to the large pose and shape variation, large appearance variation, partial observation, and self-occlusion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 613, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 288, + 696 + ], + "type": "text", + "content": "There has been steady progress in 3D human mesh recovery [30, 50, 36, 24, 60, 39, 40, 8, 5]. Recently, Transformer [67] has shown state-of-the-art (SOTA) results on a wide variety of tasks due to its strong capability of capturing long-range dependency for more accurate predictions [3, 67, 76, 74]. Using tokens constructed from local features extracted by a convolutional neural network" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 329, + 236, + 508, + 312 + ], + "blocks": [ + { + "bbox": [ + 329, + 236, + 508, + 312 + ], + "lines": [ + { + "bbox": [ + 329, + 236, + 508, + 312 + ], + "spans": [ + { + "bbox": [ + 329, + 236, + 508, + 312 + ], + "type": "image", + "image_path": "c6fd3e346e3f7d408e93da1cbab0be16ccb81da2e55f3508c9e0ef783609ead4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 329, + 316, + 508, + 391 + ], + "blocks": [ + { + "bbox": [ + 329, + 316, + 508, + 391 + ], + "lines": [ + { + "bbox": [ + 329, + 316, + 508, + 391 + ], + "spans": [ + { + "bbox": [ + 329, + 316, + 508, + 391 + ], + "type": "image", + "image_path": "d807dfac98eb535d4728e112321405cfcdf9915259009386c5d1749ad6f7f942.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 418, + 545, + 483 + ], + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 483 + ], + "type": "text", + "content": "Figure 1. Throughput v.s. Accuracy and GFLOPs v.s. Accuracy on Human3.6M [23]. Our method dramatically saves GFLOPs and improves throughput while maintaining highly competitive accuracy. The x-axis of the bottom GFLOPs figure is reversed for demonstration. Eb0 and R50 represent EfficientNet-b0 [65] and ResNet-50 [20] backbones, respectively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 398, + 542, + 415 + ], + "blocks": [ + { + "bbox": [ + 309, + 398, + 542, + 415 + ], + "lines": [ + { + "bbox": [ + 309, + 398, + 542, + 415 + ], + "spans": [ + { + "bbox": [ + 309, + 398, + 542, + 415 + ], + "type": "table", + "html": "
● METRO baseline ▲ FastMETRO baseline ● FastMETRO(S) baseline
METRO-TORE ▼ FastMETRO-TORE ● FastMETRO(S)-TORE
", + "image_path": "1eef7f6686a2c120120c2c9359cd220b2129d59c5af7a977532bf1868fb3889b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 484, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 520 + ], + "type": "text", + "content": "(CNN) [20, 70] to query the joint and mesh vertex positions, Transformer-based methods [39, 40, 8] achieved SOTA performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "text", + "content": "However, improved performance comes with costs: the increased expressivity of Transformers comes with quadratically increasing computational costs since all pairwise interactions are taken into account [14]. The space and time complexity of a QKV-attention operation is known to be " + }, + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "inline_equation", + "content": "O(M^2)" + }, + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 521, + 546, + 689 + ], + "type": "text", + "content": " is the number of tokens. The token number thus plays a vital role in time efficiency: a large number of tokens inevitably leads to a heavy computation burden. Unfortunately, almost all the existing Transformer-based SOTA methods for HMR [39, 40, 8] are suffering from redundant tokens. This incurs a high model complexity and computational costs, which prevents the current Transformer-based HMR methods from achieving their full potential in real-world applications." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "In this paper, we make key observations from two important aspects: the 3D geometry structure and 2D image" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 164, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 164, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 164, + 713 + ], + "type": "inline_equation", + "content": "\\dagger ,\\ddagger" + }, + { + "bbox": [ + 62, + 703, + 164, + 713 + ], + "type": "text", + "content": " denote equal contributions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15143" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "content": "feature, to reveal the problem of token redundancy. First, to recover the 3D geometry, all existing methods use both mesh vertices and skeleton joints as tokens for the feature interaction between input and body shape. Whereas a body mesh contains numerous vertices, they can be abstracted by a small number of joints on the skeleton. For instance, when animating a SMPL [44] avatar, the skeleton joints, together with a blend shape binding the joints and corresponding mesh vertices of the local body part, are able to describe various body meshes. Therefore, the joints can already be viewed as an underlying structure of a body shape, which intrinsically encodes the human mesh geometry. Second, for image-based input, most existing methods indiscriminately use all the feature patches to capture pose, shape and appearance variance. However, although the human body exhibits large variance, the important features for shape inference are dominantly clustered within the body area in an RGB image. Most features, e.g., image background, are not informative, thus bringing about redundancy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 304, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 289, + 639 + ], + "type": "text", + "content": "Given the aforementioned insights, we argue that the Pareto-front of accuracy and efficiency for Transformer-based HMR could be further improved by reducing the number of tokens [59, 47, 56]. To this end, we introduce a set of simple yet effective token reduction strategies mainly from two aspects corresponding to our observations. First, for 3D mesh recovery, instead of querying both vertices and joints with input features simultaneously, we consider learning a small set of body tokens at the skeleton level for each body part. To recover corresponding mesh vertices, we use an efficient Neural Shape Regressor (NSR) to infer the mesh from the body features encoded by these tokens. This query process can also be interpreted as an attention matrix decomposition, by which we effectively leverage the geometric insights encoded at the skeleton level to infer the mesh structure hierarchically. Second, for the input image feature, we introduce a learnable token pruner to prune the tokens of patch-based features extracted by a CNN. We employ a clustering-based strategy to identify discriminative features, which results in two appealing properties: 1) the end-to-end learning of the pruner is unsupervised, avoiding the need for additional data labeling; 2) it learns semantically consistent features across various images, thus further benefiting the geometry reasoning and enhancing the capability of generalizability. These token reduction strategies substantially reduce the number of query tokens involved in the computation without sacrificing the important information. An overview is shown in Figure 2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 642, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 289, + 715 + ], + "type": "text", + "content": "We conduct extensive experiments across wide benchmarks [23, 68, 85], including both the human body and hand mesh recovery, to validate the proposed method. Compared to SOTA methods, our framework faithfully recovers body meshes with fewer tokens, which considerably reduces memory and computation overhead while maintain" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 73, + 453, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 73, + 453, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 453, + 84 + ], + "type": "text", + "content": "ing competitive geometric accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 85, + 492, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 85, + 492, + 96 + ], + "spans": [ + { + "bbox": [ + 317, + 85, + 492, + 96 + ], + "type": "text", + "content": "In summary, our contribution is three-fold:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 105, + 547, + 278 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 105, + 545, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 105, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 105, + 545, + 129 + ], + "type": "text", + "content": "- We reveal the issues of token redundancy in the existing Transformer-based methods for HMR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 137, + 547, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 137, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 137, + 547, + 174 + ], + "type": "text", + "content": "- We propose effective strategies for token reduction by incorporating the insights from the 3D geometry structure and 2D image feature into the Transformer design." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 182, + 547, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 182, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 317, + 182, + 547, + 278 + ], + "type": "text", + "content": "- Our method achieves SOTA performance on various benchmarks with less computation cost. For instance, for the Transformer Encoder structure [39] and the Transformer Encoder-Decoder structure [8] with ResNet-50 [20] backbone, our method maintains competitive accuracy while saving " + }, + { + "bbox": [ + 317, + 182, + 547, + 278 + ], + "type": "inline_equation", + "content": "82.9\\%" + }, + { + "bbox": [ + 317, + 182, + 547, + 278 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 317, + 182, + 547, + 278 + ], + "type": "inline_equation", + "content": "50.5\\%" + }, + { + "bbox": [ + 317, + 182, + 547, + 278 + ], + "type": "text", + "content": " GFLOPs and improving " + }, + { + "bbox": [ + 317, + 182, + 547, + 278 + ], + "type": "inline_equation", + "content": "139.1\\%" + }, + { + "bbox": [ + 317, + 182, + 547, + 278 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 317, + 182, + 547, + 278 + ], + "type": "inline_equation", + "content": "39.8\\%" + }, + { + "bbox": [ + 317, + 182, + 547, + 278 + ], + "type": "text", + "content": " throughput, respectively; see Figure 1 for an overview." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 286, + 392, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 286, + 392, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 286, + 392, + 300 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 307, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 413 + ], + "type": "text", + "content": "Human Mesh Recovery (HMR) from monocular images has achieved great progress in the past years [4, 30, 50, 36, 10, 40, 24, 60, 5, 39, 8, 82, 71, 58]. Given a monocular RGB image, the goal of HMR is to recover the 3D body shape, typically using a human body model, e.g., SMPL [44]. We refer readers to [66] for a comprehensive review. Existing methods fall into two categories: Non-Transformer-based and Transformer-based approaches." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 415, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 629 + ], + "type": "text", + "content": "Non-Transformer-based Human Mesh Recovery To recover the human body shape, one could predict a few parameters, i.e., joint rotation and body shape, to drive the parametric model (parametric approach) or directly regress the vertex positions of the body (non-parametric approach). Parametric approaches SMPLify [4] estimates human pose and shape by fitting SMPL to the detected 2D key points [55] by optimization. Kanazawa et al. [24] adopt adversarial prior knowledge of the 3D body shape into the neural network for HMR. SPIN [30] then combines regression-based and optimization-based methods during the training loop. Intermediate features such as key points [54], pixel-to-surface correspondences [77] and texture consistency [52], have been exploited. HybrIK [36] proposes a hybrid inverse kinematics via twist-and-swing decomposition [2]. Regressing the body parameters from the image is a highly non-linear mapping [50], which thus limits the performance of parametric-based approaches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Non-parametric approaches Non-parametric approaches [50, 32, 9] recover spatial coordinates of a body shape directly from the image features. I2L-MeshNet [50] predicts the per-pixel likelihood on 1D heatmaps for vertex coordinates for a trade-off between accuracy and computational cost. GraphCMR [32] and Pose2Mesh [9] employ graph convolutional neural network (GCNN) [27]" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15144" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 527, + 223 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 527, + 223 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 527, + 223 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 527, + 223 + ], + "type": "image", + "image_path": "3214ac64bb94513573c69970a13aa59ff493cbc8bc8344123a8780d69e2899da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 225, + 546, + 249 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 546, + 249 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 546, + 249 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed framework. Our goal is to reduce tokens for Transformer Encoder and Decoder which are critical modules in the whole pipeline. Image Token Pruner (ITP) and Neural Shape Regressor (NSR) are two lightweight components." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 268, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 287, + 316 + ], + "type": "text", + "content": "to regress mesh positions by modeling local interactions among vertices from the image [32] and 2D pose [9], respectively. Although steady progress has been made, SOTA accuracy is produced by Transformer-based methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 288, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 578 + ], + "type": "text", + "content": "Transformer-based Human Mesh Recovery Recently, Transformer structures have been successfully adopted in various vision tasks [12, 43, 6, 51, 66, 80]. In HMR, the SOTA performance is achieved by Transformer-based (attention mechanism) approaches [39, 29, 40, 8]. Specifically, METRO [39] recovers body mesh using a Transformer Encoder [67] to model vertex-vertex and vertex-joint interactions conducting dimension reduction from the image features to 3D shape. MeshGraphormer [40] injects Graph Convolutions [27] into the Transformer Encoder [39] blocks to improve local interactions among vertices. The recent work FastMETRO [8] employs a Transformer Encoder-Decoder architecture disentangling the image encoding and mesh estimation for model reduction and acceleration. However, all Transformer-based approaches [39, 40, 8] suffer from redundant tokens, causing heavy interaction between body geometry and image features, which is still cumbersome and computationally intensive. In this paper, we design strategies driven by the 3D body geometry and 2D image feature for token reduction for both Transformer Encoder and Transformer Encoder-Decoder structures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 579, + 287, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 710 + ], + "type": "text", + "content": "Token Reduction for Transformers Token reduction for Transformer structures has been causing people's attention nowadays [56, 47, 59, 78, 16, 38, 46, 7]. In particular, DynamicViT [56] reduces less informative tokens hierarchically with supervision from a teacher network to save computational costs during inference. TokenLearner [59] generates a smaller number of tokens with spatial attention adaptively. Patch-to-Cluster [16] proposes to cluster over image patches with a token selector. Our method follows this way and adopts a token selection layer for clustering over input tokens, reducing tokens and improving general" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 268, + 546, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 546, + 352 + ], + "type": "text", + "content": "ization capability in challenging scenarios. Token pruning techniques have also been explored in single-view 2D pose estimation and multi-view 3D pose estimation [47] by selecting tokens based on the attention value in the Transformer. Differently, in HMR, besides the key points, body geometry, i.e., mesh vertices, also introduces a burden to the Transformer and needs to be carefully considered." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 364, + 389, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 364, + 389, + 378 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 389, + 378 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 384, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 481 + ], + "type": "text", + "content": "We employ the popular body parametric model SMPL [44] to represent a human body. Formally, given a monocular image " + }, + { + "bbox": [ + 304, + 384, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\in \\mathbb{R}^{224 \\times 224}" + }, + { + "bbox": [ + 304, + 384, + 545, + 481 + ], + "type": "text", + "content": ", our goal is to recover the joint positions " + }, + { + "bbox": [ + 304, + 384, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\in \\mathbb{R}^{J \\times 3}" + }, + { + "bbox": [ + 304, + 384, + 545, + 481 + ], + "type": "text", + "content": " and mesh vertices " + }, + { + "bbox": [ + 304, + 384, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\in \\mathbb{R}^{V \\times 3}" + }, + { + "bbox": [ + 304, + 384, + 545, + 481 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 384, + 545, + 481 + ], + "type": "inline_equation", + "content": "J = 14" + }, + { + "bbox": [ + 304, + 384, + 545, + 481 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 384, + 545, + 481 + ], + "type": "inline_equation", + "content": "V = 6890" + }, + { + "bbox": [ + 304, + 384, + 545, + 481 + ], + "type": "text", + "content": " are the numbers of joints and the mesh vertices, respectively. In addition to the human body, we also demonstrate our method on hand mesh recovery with MANO model [57] to show its generalizability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 481, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 565 + ], + "type": "text", + "content": "Existing Transformer-based HMR methods fall into two categories: a single Transformer Encoder structure [39, 40] and a Transformer Encoder-Decoder structure [8]. Accordingly, we propose two strategies to reduce tokens from 3D and 2D levels, namely, Geometry Token Reduction (GTR) and Image Token Pruning (ITP), which can accommodate different Transformer structures to improve their efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 574, + 492, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 492, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 492, + 586 + ], + "type": "text", + "content": "3.1. Geometry Token Reduction (GTR)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 593, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 676 + ], + "type": "text", + "content": "As an essential module, all the existing Transformer-based methods for HMR [39, 40, 8] have to model the interaction between body geometry and image features with an attention mechanism. Therefore, an effective token reduction method to improve the efficiency for inferring mesh geometry can benefit the design of Transformer based methods in a wide scope." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "We observe existing methods invariably use both skeleton joints and mesh vertices as tokens for regressing their spatial coordinates. The numerous tokens inevitably lead to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15145" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 155 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 155 + ], + "type": "text", + "content": "significant computational costs for a Transformer. In fact, as an underlying structure of the human body, the skeleton joints already provide strong priors for perceiving the body geometry. A human body avatar, like SMPL [44], can be driven by a small number of joints with the blend shape. This inspires us to regress the human body mesh in a hierarchical manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 286, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 286, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 286, + 264 + ], + "type": "text", + "content": "Our key idea is to decompose a heavy Transformer into smaller Transformer modules, each of which involves fewer tokens, avoiding expensive computation for image feature interaction. Specifically, our approach only queries the body tokens, whose count is equivalent to the number of joints. Since the number of joints is much less than the number of mesh vertices, we can significantly reduce the token number in complex 2D-3D feature interactions and this leads to a substantial enhancement in efficiency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 265, + 286, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 286, + 396 + ], + "type": "text", + "content": "To faithfully recover body shape from the body tokens, we introduce a Neural Shape Regressor (NSR) module, which is a light Transformer structure, to query the vertex tokens from their interaction with body features. Let " + }, + { + "bbox": [ + 46, + 265, + 286, + 396 + ], + "type": "inline_equation", + "content": "F_{J} = \\{f_{1}^{j}, f_{2}^{j}, \\dots, f_{J}^{j}\\}" + }, + { + "bbox": [ + 46, + 265, + 286, + 396 + ], + "type": "text", + "content": " denotes a set of learned body features and " + }, + { + "bbox": [ + 46, + 265, + 286, + 396 + ], + "type": "inline_equation", + "content": "T_{\\mathcal{V}} = \\{t_{1}^{v}, t_{2}^{v}, \\dots, t_{V}^{v}\\}" + }, + { + "bbox": [ + 46, + 265, + 286, + 396 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 265, + 286, + 396 + ], + "type": "inline_equation", + "content": "t_{i}^{v}" + }, + { + "bbox": [ + 46, + 265, + 286, + 396 + ], + "type": "text", + "content": " represents a set of vertex tokens which are the query tokens. We conduct cross-attention between query vertex tokens and body features to model the interaction between vertices and learned body features. Note that non-adjacent vertices are masked to improve the efficiency [8]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 397, + 286, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 286, + 529 + ], + "type": "text", + "content": "The feature size within NSR is even smaller compared with the main Transformer. Thanks to the informative geometric feature encoded in the body tokens, despite not incorporating image tokens, NSR still faithfully recovers the surface vertices. Another interesting discovery is that the learned attention scores reflect the correlation between joints and vertices at a body part level, similar to body blending weights (see Sec. 4.7.3). In this way, we can effectively reduce the cost of redundant interaction by decomposing the entire body into encoded constituent parts and efficiently recovering the whole body shape." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 538, + 200, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 200, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 200, + 551 + ], + "type": "text", + "content": "3.2. Image Token Pruning (ITP)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 286, + 664 + ], + "type": "text", + "content": "In addition to the geometric insights, for a Transformer Encoder-Decoder structure, image tokens also affect the computation overhead. However, the existing methods adopting this Transformer structure fail to avoid the token redundancy issue [8] because all available image feature patches are indiscriminately involved in human mesh recovery. Actually, some features, e.g., image background, are not informative, thus introducing redundancy and increasing the computational cost of the Transformer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "To tackle this issue, we introduce an effective token pruning strategy, namely, Image Token Pruning (ITP). Our key insight is that the informative features in an image for inferring 3D geometry are overwhelmingly clustered within" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "the region of the human body. Inspired by recent advances in token pruning [16, 59, 56, 47], we propose to aggregate features into a small number of meaningful clusters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "0 < \\rho < 1" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": " denote the predefined token pruning ratio. Our goal is to learn a projection that maps feature map " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "F_{\\mathcal{I}} \\in R^{HW \\times c}" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": " extracted from the given image with HW tokens to " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "Z_{\\mathcal{I}} \\in R^{T \\times c}" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": " tokens, where " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "T = \\lfloor \\rho HW \\rfloor" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": " is the feature dimension. The small number of clusters " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "Z_{\\mathcal{I}}" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": " are expected to capture the fewer but more discriminative features in " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "F_{\\mathcal{I}}" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": ". Inspired by [16, 59], we implement the projection as a learnable CNN module. We first apply Conv2D to the input feature, where the kernel size, stride and zero padding of Conv2D are " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": ", 1 and 1, reducing the feature dimension from " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "c' = c/4" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": " with GELU [21] activation function applied. We further map the dimension " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": " by MLP and Softmax, producing " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "M \\in R^{HW \\times N}" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": ". Finally, the clustered token set is given by " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "Z_{\\mathcal{I}} = M^T \\cdot F_{\\mathcal{I}}" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": ". The mapping matrix " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "M^T" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": ", in essence, produces a clustering over origin tokens. Each element " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "m_{ij} \\in M^T" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": " depicts the contribution of " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": " token to " + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 108, + 545, + 323 + ], + "type": "text", + "content": "-th cluster. Note that LayerNorm is employed after the clustering." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "content": "Token Reduction Supervision In order to encourage ITP to pay more attention to those discriminative feature regions, i.e., body parts. during token reduction, we introduce Token Reduction Supervision. Given the weak-perspective camera estimated the neural network, we compute the 2D projection of the ground truth 3D vertices " + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}" + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "inline_equation", + "content": "s\\Pi (\\hat{\\mathcal{V}}) + t" + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "content": " represents an orthographic projection, " + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "content": " are scale and translation estimated by the network. The projected results are downsampled to a discrete " + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 304, + 323, + 545, + 443 + ], + "type": "text", + "content": " grid corresponding to each token followed by a binary indicator function" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 453, + 544, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 544, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 544, + 495 + ], + "type": "interline_equation", + "content": "F _ {d} (x) = \\left\\{ \\begin{array}{l} 1, \\text {i f t h e c e l l c o n t a i n s a p r o j e c t e d p o i n t .} \\\\ 0, \\text {i f t h e c e l l d o e s n o t c o n t a i n a p r o j e c t e d p o i n t .} \\end{array} \\right. \\tag {1}", + "image_path": "1c76b11da429e13dc4dde6be21b4c3c0175bec656e34b7a0434e4eaa32ce3d21.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 495, + 511, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 511, + 507 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 511, + 507 + ], + "type": "text", + "content": "Then the supervision for token pruning is given by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 332, + 516, + 544, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 516, + 544, + 548 + ], + "spans": [ + { + "bbox": [ + 332, + 516, + 544, + 548 + ], + "type": "interline_equation", + "content": "L _ {P} = - \\frac {1}{N H W} \\sum_ {i} ^ {N} \\left(F _ {d} \\left(s \\Pi (\\hat {\\mathcal {V}}) + t\\right) \\cdot M _ {[:, i ]}\\right). \\tag {2}", + "image_path": "9b11e81ed17038d63f11926e2ee9d863e7f843b5fe944f2bb7933dd3a0d213c2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 557, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 604 + ], + "type": "text", + "content": "With this supervision, the weights of background tokens in " + }, + { + "bbox": [ + 304, + 557, + 545, + 604 + ], + "type": "inline_equation", + "content": "M[:i]" + }, + { + "bbox": [ + 304, + 557, + 545, + 604 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 304, + 557, + 545, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 557, + 545, + 604 + ], + "type": "text", + "content": "-th cluster will be penalized during training, which thus encourages the ITP to learn the discriminative features." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "The proposed ITP has several noticeable advantages. First, in contrast to other methods relying on explicit supervision, e.g., using a pretrained teacher network to facilitate the pruning [56, 78, 33], our token pruner is trained in an unsupervised manner; Note that no ground truth masking is required for Token Reduction Supervision. Second, unlike previous clustering-based pruning methods [16] that use pre-defined class labels, ITP aims to adaptively identify discriminative features within the body region. Therefore," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15146" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": "ITP is not limited to fixed semantic labels but is able to induce higher-level semantics according to the learning target." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 97, + 286, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 97, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 97, + 286, + 167 + ], + "type": "text", + "content": "Finally, we find the clustered informative features learned by ITP improve the model generalizability, especially for challenging datasets, e.g., in-the-wild dataset 3DPW [68] as shown in Table 3. We elaborate on these properties in Sec. 4.3.2 and Sec. 4.7.1 with extensive experiments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 179, + 140, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 179, + 140, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 179, + 140, + 190 + ], + "type": "text", + "content": "3.3. Loss Functions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "text", + "content": "We supervise the network using L1 distance between predicted mesh vertices at three sampling levels: " + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{V}3D} = ||\\mathcal{V}_{3D}^l - \\hat{\\mathcal{V}}^l||_1 + ||\\mathcal{V}_{3D}^m - \\hat{\\mathcal{V}}^m||_1 + ||\\mathcal{V}_{3D}^h - \\hat{\\mathcal{V}}_h||_1" + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "inline_equation", + "content": "*, *^m, *^h" + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "text", + "content": " denote low, middle and high body mesh resolution with vertex numbers to be 431, 1723 and 6890 for an SMPL body. The ground-truth 3D joints are used for supervising predicted 3D joints and the ones regressed from the vertices " + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{3D}^h" + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "text", + "content": " with a SMPL regression matrix " + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{J}3D} = ||\\mathcal{J}_{3D} - \\hat{\\mathcal{J}}_{3D}||_1, L_{\\mathcal{J}3D}^R = ||\\mathcal{M}(\\mathcal{V}) - \\hat{\\mathcal{J}}_{3D}||_1" + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "text", + "content": ". A 2D projection loss " + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "inline_equation", + "content": "L_{J2D}" + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "text", + "content": " is used during the network training, where we employ a weak perspective camera model to project 3D joints to 2D for supervision: " + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "inline_equation", + "content": "L_{\\mathcal{J}2D}^R = ||(s\\Pi(\\mathcal{M}(\\mathcal{V})) + t) - \\hat{\\mathcal{J}}_{2D}||_1" + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "inline_equation", + "content": "s, t" + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "text", + "content": " are scale and translation estimated by the network. Overall, together with the Token Pruning Supervision " + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "inline_equation", + "content": "L_P" + }, + { + "bbox": [ + 46, + 198, + 286, + 380 + ], + "type": "text", + "content": ", the total loss is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 392, + 284, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 392, + 284, + 422 + ], + "spans": [ + { + "bbox": [ + 49, + 392, + 284, + 422 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L = \\alpha \\left[ \\lambda_ {\\mathcal {J} 3 D} \\left(L _ {\\mathcal {J} 3 D} ^ {R} + L _ {\\mathcal {J} 3 D}\\right) + \\lambda_ {\\mathcal {V} 3 D} \\left(L _ {\\mathcal {V} 3 D}\\right) + \\lambda_ {P} L _ {P} \\right] \\\\ + \\beta \\lambda_ {\\mathcal {J} 2 D} L _ {\\mathcal {J} 2 D} ^ {R}, \\\\ \\end{array}", + "image_path": "ff2a054c6f96b4ea60fbafc6f19ff31e5207657da55129a0a2fdd70571a2d2e2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 432, + 286, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 286, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 432, + 286, + 456 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 432, + 286, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 432, + 286, + 456 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 432, + 286, + 456 + ], + "type": "text", + "content": " indicate the availability of the supervision. We set " + }, + { + "bbox": [ + 46, + 432, + 286, + 456 + ], + "type": "inline_equation", + "content": "\\lambda_{P},\\lambda_{J2D},\\lambda_{V3D},\\lambda_{J3D}" + }, + { + "bbox": [ + 46, + 432, + 286, + 456 + ], + "type": "text", + "content": " to be 1, 100, 100, 1000." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 470, + 172, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 470, + 172, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 470, + 172, + 483 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 491, + 168, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 168, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 168, + 502 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": "We evaluate our model in two scenarios: human body mesh recovery and hand mesh recovery. We adopt commonly-used metrics [25, 31, 39, 40, 8]: Mean Per-Joint Position Error (MPJPE), MPJPE after further alignment, i.e., Procrustes Analysis (PAMPJPE) and Mean Per-Vertex Error (MPVE). For human body mesh recovery, our network is trained with Human3.6M [23], MuCo-3DHP[48], UP-3D [34], COCO [41], MPI [1]. Following previous works [39, 40, 8], we use the pseudo mesh data in Human3.6M [23] for training, splitting subjects S1, S5, S6, S7, S8 for training and S9, S11 for testing. We also report our performance on 3DPW [68], a more challenging in-the-wild dataset. To further evaluate the generalization ability of the proposed method, we test our method on hand mesh recovery on FreiHAND [85] dataset. For analyzing the efficiency of our method, we report GFLOPs and throughput (image per second), strictly following [56, 49, 59, 47, 38]." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 121, + 545, + 169 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 545, + 113 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 545, + 113 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 545, + 113 + ], + "type": "text", + "content": "Table 1. Comparison with the Transformer Encoder structure METRO (M) [39] on Human3.6M [23]. We test with ResNet-50 (R50) [20] and HRNet-W64 (H64) [70] as backbones. GFLOPs" + }, + { + "bbox": [ + 305, + 69, + 545, + 113 + ], + "type": "inline_equation", + "content": "^{\\mathrm{T}}" + }, + { + "bbox": [ + 305, + 69, + 545, + 113 + ], + "type": "text", + "content": " is GFLOPs of the transformer." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 121, + 545, + 169 + ], + "lines": [ + { + "bbox": [ + 310, + 121, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 310, + 121, + 545, + 169 + ], + "type": "table", + "html": "
MethodGFLOPs ↓GFLOPsT ↓Throughput (im/s) ↑PAMPJPE ↓
M-H64 [39]56.527.5141.036.7
M-H64+GTR30.2 (-46.5%)0.8 (-97.1%)210.1 (+49.0%)37.1 (+1.1%)
M-R50 [39]31.627.5247.040.6
M-R50+GTR5.4 (-82.9%)0.8 (-97.1%)590.6 (+139.1%)42.0 (+3.4%)
", + "image_path": "6d6a85bd44eae7799d565e87ca85e08e185db8f6d2cffb3c246040a24710d867.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 199, + 438, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 199, + 438, + 212 + ], + "spans": [ + { + "bbox": [ + 306, + 199, + 438, + 212 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 217, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 338 + ], + "type": "text", + "content": "Network Training We implement the network using PyTorch. For the Transformer Encoder-Decoder structure, we set the learning rate to be " + }, + { + "bbox": [ + 304, + 217, + 545, + 338 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 217, + 545, + 338 + ], + "type": "text", + "content": ". We use the AdamW optimizer [45] and train for 60 epochs, with a batch size of 16 per GPU on 4 Nvidia A100 GPUs. When comparing with Transformer Encoder structure METRO [39] (see Table 1), we follow the setting of METRO where we train the models with a batch size of 30 per GPU on 8 Nvidia A100 GPUs in total. We adopt Adam optimizer [26] and train the models for 200 epochs. See more details in Appendix A." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 338, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 458 + ], + "type": "text", + "content": "Performance Evaluation To analyze the performance on a consumer-level GPU device, we measure the throughput on an NVIDIA RTX 3090 GPU with 24G VRAM. When comparing with encoder-decoder Transformer structure FastMETRO [8], we set the batch size to 32 following PPT [47]. The comparison with Transformer Encoder structure METRO [39] uses 16 as the batch size. Note that the batch size is limited by the size of the original METRO [39] model. To factor out the influence of the batch size, we provide a more detailed performance report in Appendix B." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 467, + 515, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 515, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 515, + 479 + ], + "type": "text", + "content": "4.3. Performance on Human Mesh Recovery" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 604 + ], + "type": "text", + "content": "Currently, there only exist two types of representative Transformer structures for HMR: Encoder-only structure (METRO [39]) and Encoder-Decoder structure (Fast-METRO [8]). To demonstrate the effectiveness of our method to both Transformer structures, we first conduct experiments by adding GTR to two structures: METRO and FastMETRO. Since ITP is designed for the encoder-decoder structure, we further report the results on FastMETRO with both ITP and GTR in Sec. 4.3.2. The overall comparison is shown in Table 5." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 622, + 474, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 474, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 474, + 633 + ], + "type": "text", + "content": "4.3.1 Transformer Encoder Structure" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "As shown in Table 1, GTR effectively reduces the computation costs while still producing competitive accuracy results. For the encoder-based Transformer model METRO [39] with HRNet-W64 [70] as a backbone, applying GTR saves the GFLOPs of the whole model for " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "46.5\\%" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " and the Transformer part for " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "97.1\\%" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": ", with throughput im" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15147" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 168, + 130 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 168, + 130 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 168, + 130 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 168, + 130 + ], + "type": "image", + "image_path": "c72ac0f1fe60fbb90346be7b9d8e80d2a7426fad6767b10662b67670932e94e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 169, + 70, + 288, + 130 + ], + "blocks": [ + { + "bbox": [ + 169, + 70, + 288, + 130 + ], + "lines": [ + { + "bbox": [ + 169, + 70, + 288, + 130 + ], + "spans": [ + { + "bbox": [ + 169, + 70, + 288, + 130 + ], + "type": "image", + "image_path": "6676cc78dad18567d6712dc6dac93b1e8c5de2febd20136c471f9af99917d57b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 131, + 168, + 190 + ], + "blocks": [ + { + "bbox": [ + 49, + 131, + 168, + 190 + ], + "lines": [ + { + "bbox": [ + 49, + 131, + 168, + 190 + ], + "spans": [ + { + "bbox": [ + 49, + 131, + 168, + 190 + ], + "type": "image", + "image_path": "937cc42f1d112f837be254d8251447a65fbf875c0f437235c9f5adf65b2a1ec5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 131, + 288, + 190 + ], + "blocks": [ + { + "bbox": [ + 168, + 131, + 288, + 190 + ], + "lines": [ + { + "bbox": [ + 168, + 131, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 168, + 131, + 288, + 190 + ], + "type": "image", + "image_path": "53f6f812e9fb9899cd68417a64be2476d1903201d6bca5d307f192fe04706e42.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 49, + 191, + 168, + 250 + ], + "blocks": [ + { + "bbox": [ + 49, + 191, + 168, + 250 + ], + "lines": [ + { + "bbox": [ + 49, + 191, + 168, + 250 + ], + "spans": [ + { + "bbox": [ + 49, + 191, + 168, + 250 + ], + "type": "image", + "image_path": "2b85f6a8c29ca2a1d742d96eddacd23191c9a2bcf11861871627f1d09e94b8d1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 191, + 288, + 249 + ], + "blocks": [ + { + "bbox": [ + 168, + 191, + 288, + 249 + ], + "lines": [ + { + "bbox": [ + 168, + 191, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 168, + 191, + 288, + 249 + ], + "type": "image", + "image_path": "6bc09f241d8d32243e21f229da1894d05931658d2b3be1f676eafe9c0919f6f4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 49, + 251, + 168, + 311 + ], + "blocks": [ + { + "bbox": [ + 49, + 251, + 168, + 311 + ], + "lines": [ + { + "bbox": [ + 49, + 251, + 168, + 311 + ], + "spans": [ + { + "bbox": [ + 49, + 251, + 168, + 311 + ], + "type": "image", + "image_path": "d50d40c956529558d6af8c3f6a4b5913652bba5873391472df7dbeb5aceb14c5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 313, + 287, + 335 + ], + "lines": [ + { + "bbox": [ + 47, + 313, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 287, + 335 + ], + "type": "text", + "content": "Figure 3. Qualitative results of GTR equipped Encoder-Decoder structure [39] (H64) on Human3.6M [23] and 3DPW [68]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 168, + 250, + 288, + 311 + ], + "blocks": [ + { + "bbox": [ + 168, + 250, + 288, + 311 + ], + "lines": [ + { + "bbox": [ + 168, + 250, + 288, + 311 + ], + "spans": [ + { + "bbox": [ + 168, + 250, + 288, + 311 + ], + "type": "image", + "image_path": "3cc684d60fb5b007ce93d3b584e9d23cf924864c11a0b5456af91da231c16a6c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 51, + 408, + 288, + 488 + ], + "blocks": [ + { + "bbox": [ + 47, + 345, + 287, + 399 + ], + "lines": [ + { + "bbox": [ + 47, + 345, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 345, + 287, + 399 + ], + "type": "text", + "content": "Table 2. Comparison with the Transformer Encoder-Decoder structure FastMETRO (FM) [8] on Human3.6M [23]. We test with EfficientNet-b0 (Eb0) [65], ResNet-50 (R50) [20] and HRNet-W64 (H64) [70] as backbones. GFLOPs" + }, + { + "bbox": [ + 47, + 345, + 287, + 399 + ], + "type": "inline_equation", + "content": "^{\\mathrm{T}}" + }, + { + "bbox": [ + 47, + 345, + 287, + 399 + ], + "type": "text", + "content": " stand for GFLOPs for the transformer." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 408, + 288, + 488 + ], + "lines": [ + { + "bbox": [ + 51, + 408, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 51, + 408, + 288, + 488 + ], + "type": "table", + "html": "
MethodGFLOPs ↓GFLOPsT ↓Throughput (im/s) ↑PAMPJPE ↓
FM-H64 [8]35.76.6221.533.7
FM-H64+GTR30.2 (-15.4%)0.7 (-89.4%)249.2 (+12.5%)34.8 (+3.2%)
FM-R50 [8]10.96.6576.037.3
FM-R50+GTR5.4 (-50.5%)0.7 (-89.4%)805.3 (+39.8%)38.6 (+3.4%)
FM(S)-R50 [8]6.42.2953.539.4
FM(S)-R50+GTR4.6 (-28.1%)0.3 (-86.4%)1128.9 (+18.4%)38.6 (-2.0%)
FM-Eb07.16.6517.645.8
FM-Eb0+GTR1.7 (-76.1%)0.7 (-89.4%)870.5 (+68.2%)44.2 (-3.5%)
", + "image_path": "d2e137caa4342ffd2d2b2942f9b039651e0ebea5d575d2218a609771dd560823.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "type": "text", + "content": "proved by " + }, + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "type": "inline_equation", + "content": "49\\%" + }, + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "type": "text", + "content": ". For the ResNet-50 [20] backbone, GTR helps save the GFLOPs of the whole model for " + }, + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "type": "inline_equation", + "content": "82.9\\%" + }, + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "type": "text", + "content": " and the Transformer part for " + }, + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "type": "inline_equation", + "content": "97.1\\%" + }, + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "type": "text", + "content": ", with throughput improved by " + }, + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "type": "inline_equation", + "content": "139.1\\%" + }, + { + "bbox": [ + 46, + 519, + 287, + 616 + ], + "type": "text", + "content": ". These results validate the effectiveness of the proposed GTR on the Transformer Encoder structure. Qualitative results of METRO-H64+GTR can be found in Figure 3, in which the model produces high-quality results in human mesh recovery over various input monocular images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 646, + 255, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 255, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 255, + 657 + ], + "type": "text", + "content": "4.3.2 Transformer Encoder-Decoder Structure" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "text", + "content": "We experiment on the Transformer Encoder-Decoder structure FastMETRO [8] with its two variants: models with 1 and 3 Encoder-Decoder layers. For 1-layer FastMETRO, we denote it as FastMETRO(S). In Table 2, GTR equipped" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "FastMETRO [39] produces competitive accuracy while reducing " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "15.4\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " in GFLOPs for the whole model, " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "89.4\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " in GFLOPs for the Transformer part and improving " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "12.5\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " in throughput with an HRNet-W64 [70] CNN backbone. When testing with ResNet-50 [20], GTR helps save " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "50.5\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " in GFLOPs for the whole model, " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "89.4\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " in GFLOPs for the Transformer part and improves " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "39.8\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " in throughput. Notably, our FastMETRO(S)-R50+GTR and FastMETRO(S)-Eb0+GTR yield 38.6, 44.2 mm PAMPJPE respectively, surpassing the corresponding baselines while greatly saving computational costs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 204, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 240 + ], + "type": "text", + "content": "We then present the performance of the Transformer Encoder-Decoder structure with GTR and ITP using ResNet-50 [20] and EfficientNet-b0 [65] CNN backbones." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 310, + 280, + 545, + 309 + ], + "blocks": [ + { + "bbox": [ + 304, + 251, + 545, + 273 + ], + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 273 + ], + "type": "text", + "content": "Table 3. Influence of ITP for monocular 3D human mesh recovery on 3DPW [68]." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 280, + 545, + 309 + ], + "lines": [ + { + "bbox": [ + 310, + 280, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 310, + 280, + 545, + 309 + ], + "type": "table", + "html": "
MethodMPVEMPJPEPAMPJPE
FastMETRO-H64+GTR91.375.446.7
FastMETRO-H64+GTR+ITP@20%88.272.344.4
", + "image_path": "9a18612dadff735d11b3e67cf87e7a70b4ad10c99e45a26c82cac4f1b2469058.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "text", + "content": "As shown in Table 4, TORE (GTR+ITP) is effective for the Encoder-Decoder Transformer structure. Specifically, when pruning with " + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "text", + "content": " in ITP, the models save the GLOPs of the Transformer structure by " + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "inline_equation", + "content": "14.3\\%" + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "inline_equation", + "content": "14.3\\%" + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "inline_equation", + "content": "28.5\\%" + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "inline_equation", + "content": "42.9\\%" + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "text", + "content": " while receiving competitive accuracy or even higher accuracy i.e., FM-Eb0+GTR+ITP@" + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 335, + 545, + 418 + ], + "type": "text", + "content": " improves PAMPJPE compared with the baseline." + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 309, + 459, + 548, + 628 + ], + "blocks": [ + { + "bbox": [ + 304, + 429, + 545, + 452 + ], + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 452 + ], + "type": "text", + "content": "Table 5. Comparison with the SOTA methods for monocular 3D human mesh recovery on 3DPW [68] and Human3.6M [23]." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 459, + 548, + 628 + ], + "lines": [ + { + "bbox": [ + 309, + 459, + 548, + 628 + ], + "spans": [ + { + "bbox": [ + 309, + 459, + 548, + 628 + ], + "type": "table", + "html": "
Method3DPWHuman3.6M
MPVEMPJPEPAMPJPEMPJPEPAMPJPE
HMR [25]-130.076.788.056.8
GraphCMR [32]--70.2-50.1
SPIN [30]116.496.959.262.541.1
I2LMeshNet [50]-93.257.755.741.1
PyMAF [83]110.192.858.957.740.5
ROMP-R50 [64]105.689.353.5--
PARE-R50 [29]99.782.952.3--
DSR-R50 [13]99.585.751.760.940.3
METRO-R50 [39]---56.540.6
METRO-H64 [39]88.277.147.954.036.7
METRO-H64+GTR87.975.546.657.637.1
FastMETRO-R50 [8]90.677.948.353.937.3
FastMETRO-R50+GTR+ITP@20%99.282.452.359.840.5
FastMETRO-H64 [8]84.173.544.652.233.7
FastMETRO-H64+GTR+ITP@20%88.272.344.459.636.4
FastMETRO-Eb0 [8]112.593.860.269.245.8
FastMETRO-Eb0+GTR+ITP@20%112.593.760.163.243.9
", + "image_path": "fc02a879db89a21c14b13db460cfda2b6199276b61914f13dd19f61e3f241e76.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "In addition, we find that ITP helps improve the accuracy on the challenging in-the-wild 3DPW [68] dataset, shown in Table 3. Specifically, when further equipped with ITP, the model performance in MPVE, MPJPE and PAMPJPE are improved by " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "3.1\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "3.1\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "inline_equation", + "content": "2.3\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15148" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 99, + 542, + 173 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "type": "text", + "content": "Table 4. Statistics of all proposed components (GTR + ITP) for Encoder-Decoder structure FastMETRO (FM) [8] on Human3.6M [23]. The backbones are EfficientNet-b0 (Eb0) [65] and ResNet-50 (R50) [20]. GFLOPs" + }, + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "type": "inline_equation", + "content": "^{\\mathrm{T}}" + }, + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "type": "text", + "content": " stands for GFLOPs of the Transformer." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 99, + 542, + 173 + ], + "lines": [ + { + "bbox": [ + 52, + 99, + 542, + 173 + ], + "spans": [ + { + "bbox": [ + 52, + 99, + 542, + 173 + ], + "type": "table", + "html": "
Method + Pruning Rate#TokensGFLOPs ↓GFLOPsT ↓Throughput (im/s) ↑PAMPJPE ↓
FM-R50+GTR495.40.7805.338.6
FM-R50+GTR+ITP@20%395.3 (-1.8%)0.6 (-14.3%)794.7 (-1.3%)40.5 (+4.9%)
FM-R50+GTR+ITP@50%245.1 (-5.5%)0.5 (-28.5%)806.1 (+0.1%)40.7 (+5.4%)
FM-Eb0+GTR491.70.7870.544.2
FM-Eb0+GTR+ITP@20%391.6 (-5.9%)0.6 (-14.3%)876.3 (+0.7%)43.9 (-0.%)
FM-Eb0+GTR+ITP@50%241.4 (-17.6%)0.4 (-42.9%)870.4 (-0.1%)44.0 (-0.5%)
", + "image_path": "fe5b946d16d80ad2af1b368e6289b67549f3330c6f87a49ec7226d0f1d86de46.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 192, + 172, + 255 + ], + "blocks": [ + { + "bbox": [ + 49, + 192, + 172, + 255 + ], + "lines": [ + { + "bbox": [ + 49, + 192, + 172, + 255 + ], + "spans": [ + { + "bbox": [ + 49, + 192, + 172, + 255 + ], + "type": "image", + "image_path": "cd7d259b3a4cc5ec1d2550a7f956935ea521df3a69c821beeb1b6abe86c3bb36.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 255, + 171, + 318 + ], + "blocks": [ + { + "bbox": [ + 49, + 255, + 171, + 318 + ], + "lines": [ + { + "bbox": [ + 49, + 255, + 171, + 318 + ], + "spans": [ + { + "bbox": [ + 49, + 255, + 171, + 318 + ], + "type": "image", + "image_path": "447312e7e09283642e1ed1d103858963b70c2f4e84fddf848c121f9c91832c5e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 321, + 478, + 333 + ], + "lines": [ + { + "bbox": [ + 113, + 321, + 478, + 333 + ], + "spans": [ + { + "bbox": [ + 113, + 321, + 478, + 333 + ], + "type": "text", + "content": "Figure 4. Qualitative results of FastMETRO+GTP+ITP@20% on Human3.6M [23] and 3DPW [68]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 173, + 192, + 296, + 255 + ], + "blocks": [ + { + "bbox": [ + 173, + 192, + 296, + 255 + ], + "lines": [ + { + "bbox": [ + 173, + 192, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 173, + 192, + 296, + 255 + ], + "type": "image", + "image_path": "026c783d04408d6546c7bc6e26b8cabce8c3bd2f498696141238b3e6398cdba6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 173, + 255, + 296, + 318 + ], + "blocks": [ + { + "bbox": [ + 173, + 255, + 296, + 318 + ], + "lines": [ + { + "bbox": [ + 173, + 255, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 173, + 255, + 296, + 318 + ], + "type": "image", + "image_path": "946903e54861b1c25d1239531e680fc684e787f6db8bfdb2a85e6e961fa79315.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 298, + 192, + 420, + 255 + ], + "blocks": [ + { + "bbox": [ + 298, + 192, + 420, + 255 + ], + "lines": [ + { + "bbox": [ + 298, + 192, + 420, + 255 + ], + "spans": [ + { + "bbox": [ + 298, + 192, + 420, + 255 + ], + "type": "image", + "image_path": "839ec8ee3ca390e05b2e3871156819dcdc0e886108d9453e28aa494b2cd44e92.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 298, + 255, + 420, + 318 + ], + "blocks": [ + { + "bbox": [ + 298, + 255, + 420, + 318 + ], + "lines": [ + { + "bbox": [ + 298, + 255, + 420, + 318 + ], + "spans": [ + { + "bbox": [ + 298, + 255, + 420, + 318 + ], + "type": "image", + "image_path": "11554b0651fe67d123710420d241da859bf90726756bd50d4403c74cf77eb6ae.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 421, + 192, + 545, + 253 + ], + "blocks": [ + { + "bbox": [ + 421, + 192, + 545, + 253 + ], + "lines": [ + { + "bbox": [ + 421, + 192, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 421, + 192, + 545, + 253 + ], + "type": "image", + "image_path": "dbcbfe70555b0bead4b216f4abc7ad20b29cadf163692c197ceb1108cd237f10.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 422, + 255, + 545, + 318 + ], + "blocks": [ + { + "bbox": [ + 422, + 255, + 545, + 318 + ], + "lines": [ + { + "bbox": [ + 422, + 255, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 422, + 255, + 545, + 318 + ], + "type": "image", + "image_path": "bd2865b51f78814e662d1217c64f530a25bcd5ed8852928b0dc9177b8c37ae76.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 352, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 448 + ], + "type": "text", + "content": "These results indicate that the ITP module learns more discriminative features and thus enhances the capability of generalization, allowing methods with ITP to achieve better performance in more challenging in-the-wild scenarios. The strong representation capability of ITP can also be seen in the clustering process, which promotes competitive accuracy with fewer tokens while producing semantically meaningful clustering results, as demonstrated in Sec. 4.7.1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 448, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 288, + 700 + ], + "type": "text", + "content": "We conduct comparisons with existing SOTA approaches (both Transformer and Non-Transformer based) on monocular human mesh recovery. The results are summarized in Table 5. It is notable that, with effective token reduction techniques, our method produces competitive or higher accuracy on Human3.6M [23] and 3DPW [68] datasets. For instance, on 3DPW, our METRO-H64+GTR achieves " + }, + { + "bbox": [ + 46, + 448, + 288, + 700 + ], + "type": "inline_equation", + "content": "87.9\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 448, + 288, + 700 + ], + "type": "text", + "content": " MPVE, " + }, + { + "bbox": [ + 46, + 448, + 288, + 700 + ], + "type": "inline_equation", + "content": "75.5\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 448, + 288, + 700 + ], + "type": "text", + "content": " MPJPE and " + }, + { + "bbox": [ + 46, + 448, + 288, + 700 + ], + "type": "inline_equation", + "content": "46.6\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 448, + 288, + 700 + ], + "type": "text", + "content": " PAMPJPE surpassing METRO-H64. In general, token reduction results in information loss, which leads to a slight drop in accuracy. For MPVE, using an HRNet-W64 backbone, GTR typically causes a larger error increase of " + }, + { + "bbox": [ + 46, + 448, + 288, + 700 + ], + "type": "inline_equation", + "content": "2.6\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 448, + 288, + 700 + ], + "type": "text", + "content": " (from 84.1 to 86.7), while ITP is 1.5 mm (from 86.7 to 88.2). This suggests GTR sacrifices more accuracy for efficiency, while ITP has less impact. Qualitative results of FastMETRO+GTR+ITP@20% are visualized in Figure 4, where our method produces accurate and robust human mesh recovery from monocular images. In the inset, we also provide a qualitative comparison of FastMETRO+GTR+ITP@20% (w/ GTR+ITP) and the baseline model FastMETRO (w/o GTR+ITP)," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 352, + 452, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 452, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 452, + 483 + ], + "type": "text", + "content": "where the joint-vertex attention is similar to the blending weights in SMPL, which properly captures the shape structure. However, the model w/o GTR+ITP redundantly correlates local joints with distant vertices, leading to additional interaction costs. In summary, extensive experiments validate the effectiveness of proposed strategies for token reduction across different" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 483, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 532 + ], + "type": "text", + "content": "Transformer structures (Encoder- w/ GTR+ITP w/o GTR+ITP based METRO [39] and Encoder-Decoder-based Fast-METRO [8]) and different Transformer model sizes (Fast-METRO and FastMETRO(S))." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 538, + 455, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 455, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 455, + 551 + ], + "type": "text", + "content": "4.4. ITP v.s. TokenLearner [59]" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 310, + 604, + 545, + 644 + ], + "blocks": [ + { + "bbox": [ + 306, + 564, + 545, + 587 + ], + "lines": [ + { + "bbox": [ + 306, + 564, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 564, + 545, + 587 + ], + "type": "text", + "content": "Table 6. Comparison with TokenLearning [59] in ITP for token reduction on 3DPW [68] and Human3.6M [23]." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 604, + 545, + 644 + ], + "lines": [ + { + "bbox": [ + 310, + 604, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 310, + 604, + 545, + 644 + ], + "type": "table", + "html": "
MethodGFLOPs3DPWHuman3.6M
MPVEMPJPEPAMPJPEMPJPEPAMPJPE
TokenLearner [59]5.799.382.452.661.245.4
Image Token Pruning5.399.282.452.359.840.5
", + "image_path": "74aa84dafe894e0cae915d51f8a555f536c90f31b236fc64cca0d3495cb06703.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "We compare ITP with another popular pruning strategy Tokenlearner [59] for HMR on Human3.6M and 3DPW in Table 6. We use Encoder-Decoder structure [8] with ResNet-50 [20] as a backbone. The pruning rate is " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ". In" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 457, + 354, + 541, + 479 + ], + "blocks": [ + { + "bbox": [ + 457, + 354, + 541, + 479 + ], + "lines": [ + { + "bbox": [ + 457, + 354, + 541, + 479 + ], + "spans": [ + { + "bbox": [ + 457, + 354, + 541, + 479 + ], + "type": "image", + "image_path": "13333b54e4bf922d542975797abbb2441bedec462c18fb5b7ec7a5b6300e280a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "15149" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 173, + 133 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 173, + 133 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 173, + 133 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 173, + 133 + ], + "type": "image", + "image_path": "a82b688f14eabb1b9abc281376466415eb929411338a0fed2d33583d20ea8626.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 296, + 133 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 296, + 133 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 296, + 133 + ], + "type": "image", + "image_path": "4cb2e743f5cd93c723a94d3ade0c3b2baefe0cc076ddb267ea71e2e9782cef19.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 296, + 70, + 421, + 133 + ], + "blocks": [ + { + "bbox": [ + 296, + 70, + 421, + 133 + ], + "lines": [ + { + "bbox": [ + 296, + 70, + 421, + 133 + ], + "spans": [ + { + "bbox": [ + 296, + 70, + 421, + 133 + ], + "type": "image", + "image_path": "0808a5a7604bb6d31de8de63e26009c6d52088476f90a7affae7351627a9a40f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 421, + 70, + 545, + 133 + ], + "blocks": [ + { + "bbox": [ + 421, + 70, + 545, + 133 + ], + "lines": [ + { + "bbox": [ + 421, + 70, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 421, + 70, + 545, + 133 + ], + "type": "image", + "image_path": "0b94ffb196aa260807334e7493627359a983ae801a5cfc48d1f5727571b9fd91.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 49, + 133, + 172, + 196 + ], + "blocks": [ + { + "bbox": [ + 49, + 133, + 172, + 196 + ], + "lines": [ + { + "bbox": [ + 49, + 133, + 172, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 133, + 172, + 196 + ], + "type": "image", + "image_path": "a58e49d5c51325fe02ac4f8a10df0d57d82993c21c1886b4aeb97f5e6c8332fa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 198, + 470, + 210 + ], + "lines": [ + { + "bbox": [ + 122, + 198, + 470, + 210 + ], + "spans": [ + { + "bbox": [ + 122, + 198, + 470, + 210 + ], + "type": "text", + "content": "Figure 5. Qualitative results on FreiHAND [85] by FastMETRO+H64+GTR+ITP@20% model." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 173, + 133, + 296, + 195 + ], + "blocks": [ + { + "bbox": [ + 173, + 133, + 296, + 195 + ], + "lines": [ + { + "bbox": [ + 173, + 133, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 173, + 133, + 296, + 195 + ], + "type": "image", + "image_path": "a79a07f186f6410fd460ab244e76cd6f47964fff8edee0a83f4a12f308c4f474.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 296, + 133, + 421, + 195 + ], + "blocks": [ + { + "bbox": [ + 296, + 133, + 421, + 195 + ], + "lines": [ + { + "bbox": [ + 296, + 133, + 421, + 195 + ], + "spans": [ + { + "bbox": [ + 296, + 133, + 421, + 195 + ], + "type": "image", + "image_path": "12ec44111f8919a2f3c4a696ae2af2475487e668e72a0417ac1fdd9af0026d0d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 421, + 133, + 545, + 195 + ], + "blocks": [ + { + "bbox": [ + 421, + 133, + 545, + 195 + ], + "lines": [ + { + "bbox": [ + 421, + 133, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 421, + 133, + 545, + 195 + ], + "type": "image", + "image_path": "0525dc7d8c878ec74c33ab57bae424ba84c5241ab106b1940a8aa949f80c7542.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 229, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 287, + 266 + ], + "type": "text", + "content": "Table 8, ITP saves more computational costs while achieving the highest accuracy in terms of both human mesh recovery and joint estimation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 273, + 206, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 273, + 206, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 273, + 206, + 285 + ], + "type": "text", + "content": "4.5. NSR v.s. GCN [27] and MLP" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 292, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 411 + ], + "type": "text", + "content": "We discuss the effectiveness of NSR in GTR. We compare NSR with other implementations, including MultiLayer Perceptron and Graph Convolutional Network [27] following Pose2Mesh [9]. We condition the backbone on ResNet-50 [20] using Encoder-Decoder structure [8]. In Table 7, NSR achieves higher performance on Human3.6M and 3DPW. This indicates that the attention mechanism in NSR provides a stronger modeling capability of vertices given the learned body features, which thus improves the quality of recovered mesh vertices." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 52, + 451, + 287, + 502 + ], + "blocks": [ + { + "bbox": [ + 47, + 421, + 287, + 443 + ], + "lines": [ + { + "bbox": [ + 47, + 421, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 287, + 443 + ], + "type": "text", + "content": "Table 7. Comparison of different network structures of NSR for GTR on 3DPW [68] and Human3.6M [23]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 451, + 287, + 502 + ], + "lines": [ + { + "bbox": [ + 52, + 451, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 52, + 451, + 287, + 502 + ], + "type": "table", + "html": "
Model3DPWHuman3.6M
MPVEMPJPEPAMPJPEMPJPEPAMPJPE
Multi-Layer Perceptron99.080.449.657.738.9
Graph Convolutional Network98.881.549.858.238.8
Neural Shape Regressor95.979.249.257.238.6
", + "image_path": "8686277e6f99048a923622e11ca64b9bb6ba61e7708295025f5666ea17f1f0b1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 530, + 257, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 257, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 257, + 544 + ], + "type": "text", + "content": "4.6. Generalization on Hand Mesh Recovery" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 548, + 286, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 286, + 644 + ], + "type": "text", + "content": "To investigate the generalizability of our framework, we conduct an experiment on monocular hand mesh recovery, which is summarized in Table 8. Following FastMETRO [8], we report PAMPJPE, F-score@15mm (F@15mm) on FreiHand [85] together with GFLOPs" + }, + { + "bbox": [ + 46, + 548, + 286, + 644 + ], + "type": "inline_equation", + "content": "^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 548, + 286, + 644 + ], + "type": "text", + "content": " and throughput. The qualitative results are provided in Figure 5. The hand vertex-joint interactions are visualized in Appendix C." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 651, + 160, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 160, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 160, + 663 + ], + "type": "text", + "content": "4.7. Further Discussion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 670, + 222, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 222, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 222, + 683 + ], + "type": "text", + "content": "4.7.1 Analysis of Image Token Pruning" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Image Token Pruning, in essence, achieves body-aware clustering results encoding discriminative body features." + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 309, + 256, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 305, + 227, + 545, + 248 + ], + "lines": [ + { + "bbox": [ + 305, + 227, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 305, + 227, + 545, + 248 + ], + "type": "text", + "content": "Table 8. Comparison with the SOTA methods on hand mesh recovery on FreiHAND [85]." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 256, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 309, + 256, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 309, + 256, + 545, + 293 + ], + "type": "table", + "html": "
Method\\( \\mathrm{{GFLOPs}}{}^{\\mathrm{T}} \\downarrow \\) Throughput \\( \\uparrow \\) PAMPJPE \\( \\downarrow \\mathrm{F}@{15}\\mathrm{\\;{mm}} \\uparrow \\)
METRO-H64 [39]13.1186.76.80.981
FastMETRO-H64 [8]3.3228.46.50.982
FastMETRO-H64+GTR+ITP@20%1.0260.26.70.980
", + "image_path": "c1432171cdfc64acbf21b77fca661d6da5e888dc0a2e3c0402ec27a8067af5ff.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 307, + 310, + 545, + 467 + ], + "blocks": [ + { + "bbox": [ + 307, + 310, + 545, + 467 + ], + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 467 + ], + "type": "image", + "image_path": "981d7bad6fc317071ac01a34e788242e9e05512f5c3ac3d0181418b4056ad828.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 472, + 545, + 506 + ], + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 506 + ], + "type": "text", + "content": "Figure 6. Visualization of learned semantics by ITP. Note that the clustering results are consistent across different identities, i.e., different clusters correspond to different body joints." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 525, + 545, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 608 + ], + "type": "text", + "content": "We visualize the heatmap of the predicted clustering scores in Figure 6, where the model is with ITP at " + }, + { + "bbox": [ + 304, + 525, + 545, + 608 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 525, + 545, + 608 + ], + "type": "text", + "content": " pruning rate (resulting in 24 image tokens) and GTR. The backbone is ResNet-50. As shown in Figure 6, clustering-based pruning maps original tokens to a fewer number of clusters with semantics corresponding to the body joints. Note that the semantics is consistent across different identities." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 623, + 504, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 623, + 504, + 635 + ], + "spans": [ + { + "bbox": [ + 305, + 623, + 504, + 635 + ], + "type": "text", + "content": "4.7.2 ITP with Token Reduction Supervision" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "In Figure 7, we visualize the scores predicted by ITP in HMR. The mask supervision is generated by projecting mesh vertices to image patches using estimated camera parameters as shown in Figure 7 (b)(c). ITP effectively learns to pay attention to a human body region in the given image. Quantitatively, the projection supervision improves PAM" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15150" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 69, + 289, + 200 + ], + "blocks": [ + { + "bbox": [ + 49, + 69, + 289, + 200 + ], + "lines": [ + { + "bbox": [ + 49, + 69, + 289, + 200 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 289, + 200 + ], + "type": "image", + "image_path": "fc54262c033eabde71b001267b344aab5970e91b012da652442cb257b20f2a9b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 204, + 287, + 239 + ], + "lines": [ + { + "bbox": [ + 47, + 204, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 287, + 239 + ], + "type": "text", + "content": "Figure 7. Visualization of learned semantics by Image Token Pruner. (a) projected joints. (b) projected mesh vertices. (c) mask supervision. (d) scores predicted by ITP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 247, + 287, + 367 + ], + "blocks": [ + { + "bbox": [ + 49, + 247, + 287, + 367 + ], + "lines": [ + { + "bbox": [ + 49, + 247, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 49, + 247, + 287, + 367 + ], + "type": "image", + "image_path": "4d8665f371240e1a786092e55e385b39a9e7f13f3bb6e64c7b0eb002872b5d06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 368, + 280, + 379 + ], + "lines": [ + { + "bbox": [ + 59, + 368, + 280, + 379 + ], + "spans": [ + { + "bbox": [ + 59, + 368, + 280, + 379 + ], + "type": "text", + "content": "Input R-Elbow R-Knee L-Elbow L-Knee Head" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 384, + 287, + 407 + ], + "lines": [ + { + "bbox": [ + 47, + 384, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 287, + 407 + ], + "type": "text", + "content": "Figure 8. Visualization of cross-attention between joint and vertices. Samples are from Human3.6M and 3DPW." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 426, + 212, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 212, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 212, + 437 + ], + "type": "text", + "content": "PJPE from 41.6 to 40.6 on Human3.6M." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 453, + 223, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 223, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 223, + 464 + ], + "type": "text", + "content": "4.7.3 Vertex-Body Feature Interactions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 579 + ], + "type": "text", + "content": "We show the interactions between vertices and joints modeled by cross attention within Neural Shape Regressor; See Figure 8, where the heatmap is obtained by averaging attention scores across all heads of the multi-head cross-attention between query vertices and body features. As shown in Figure 8, the interactions of mesh vertices and joints are at the body part level, e.g., elbow, knee, head, which are similar to the way of the blending of a human body model, i.e., SMPL [44] and thus validates our claims." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 586, + 209, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 209, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 209, + 598 + ], + "type": "text", + "content": "4.8. Limitations and Future Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "When an extremely high pruning rate is applied to ITP, the accuracy of the model drops dramatically, e.g., when the token number is pruned to be one, the accuracy of the model drops dramatically (12.1%) from 38.6 to 43.3 PAMPJPE. For GTR, since we regress joints and vertices progressively, the quality of the recovered vertices by NSR depends on the learned body features. More failure cases are in Appendix D. In future work, one of the promising directions could be applying the shown enhanced efficiency of" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 546, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 144 + ], + "type": "text", + "content": "HMR from monocular images to methods exhibiting high complexity for improving the model efficiency, especially in tasks such as human-environment/object interaction that perceives environments [37, 22, 18, 75, 73, 19, 61], as well as HMR from videos that involve temporal information [28, 53, 35, 63, 79, 72]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 156, + 379, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 379, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 379, + 168 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 176, + 546, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 176, + 546, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 176, + 546, + 320 + ], + "type": "text", + "content": "In this paper, we investigate the issues of token redundancy in the existing Transformer-based methods for both body and hand mesh recovery tasks. To tackle the problem, we introduce two effective token reduction strategies for Transformers by incorporating insights from both 3D geometry structure and 2D image features. Specifically, we recover the body shape in a hierarchical manner and cluster for image features to feed fewer but more discriminative tokens to the Transformer. Our method dramatically reduces the high-complexity interactions in the Transformer, improving the Pareto-front of accuracy and efficiency. Extensive experiments validate the proposed strategies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 331, + 421, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 331, + 421, + 345 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 421, + 345 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 352, + 545, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 423 + ], + "type": "text", + "content": "The authors would like to thank Gene Team for the fruitful discussion and the anonymous reviewers for their valuable comments and suggestions. This research is supported by Meta Reality Labs, Innovation and Technology Commission (Ref: ITS/319/21FP) and Research Grant Council (Ref: 17210222)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 435, + 365, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 365, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 365, + 448 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 310, + 455, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 310, + 455, + 546, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 455, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 310, + 455, + 546, + 510 + ], + "type": "text", + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 512, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 512, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 311, + 512, + 547, + 544 + ], + "type": "text", + "content": "[2] Paolo Baerlocher and Ronan Boulic. Parametrization and range of motion of the ball-and-socket joint. In Deformable avatars, pages 180–190. Springer, 2001." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 546, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 546, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 311, + 546, + 547, + 578 + ], + "type": "text", + "content": "[3] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 311, + 580, + 547, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 580, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 311, + 580, + 547, + 634 + ], + "type": "text", + "content": "[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In European conference on computer vision, pages 561-578. Springer, 2016." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 636, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 636, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 311, + 636, + 547, + 713 + ], + "type": "text", + "content": "[5] Zhongang Cai, Daxuan Ren, Ailing Zeng, Zhengyu Lin, Tao Yu, Wenjia Wang, Xiangyu Fan, Yang Gao, Yifan Yu, Liang Pan, et al. Humman: Multi-modal 4d human dataset for versatile sensing and modeling. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 557-577. Springer, 2022." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15151" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 73, + 286, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 286, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 286, + 116 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 286, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 286, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 286, + 172 + ], + "type": "text", + "content": "[7] Xuanyao Chen, Zhijian Liu, Haotian Tang, Li Yi, Hang Zhao, and Song Han. Sparsevit: Revisiting activation sparsity for efficient high-resolution vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2061-2070, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 173, + 286, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 173, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 173, + 286, + 216 + ], + "type": "text", + "content": "[8] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3d human mesh recovery with transformers. In European Conference on Computer Vision (ECCV), 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 217, + 286, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 286, + 260 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 286, + 260 + ], + "type": "text", + "content": "[9] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In European Conference on Computer Vision, pages 769-787. Springer, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 286, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 286, + 293 + ], + "type": "text", + "content": "[10] MMHuman3D Contributors. Openmmlab 3d human parametric model toolbox and benchmark. https://github.com/open-mmlab/mmhuman3d, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 286, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 286, + 348 + ], + "type": "text", + "content": "[11] Haoye Dong, Xiaodan Liang, Xiaohui Shen, Bochao Wang, Hanjiang Lai, Jia Zhu, Zhiting Hu, and Jian Yin. Towards multi-pose guided virtual try-on network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9026-9035, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 286, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 286, + 414 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "type": "text", + "content": "[13] Sai Kumar Dwivedi, Nikos Athanasiou, Muhammed Kocabas, and Michael J Black. Learning to regress bodies from images using differentiable semantic rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11250-11259, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 286, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 286, + 514 + ], + "type": "text", + "content": "[14] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 557 + ], + "type": "text", + "content": "[15] Chuang Gan, Yi Yang, Linchao Zhu, Deli Zhao, and Yueting Zhuang. Recognizing an action using its name: A knowledge-based approach. International Journal of Computer Vision, 120(1):61-77, 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 286, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 591 + ], + "type": "text", + "content": "[16] Ryan Grainger, Thomas Paniagua, Xi Song, and Tianfu Wu. Learning patch-to-cluster attention in vision transformer. arXiv preprint arXiv:2203.11987, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 286, + 635 + ], + "type": "text", + "content": "[17] Yong Guo, Zhiyang Dou, Nan Zhang, Xiyue Liu, Boni Su, Yuguo Li, and Yinping Zhang. Student close contact behavior and COVID-19 transmission in China's classrooms. PNAS Nexus, 2(5):pgad142, 05 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 689 + ], + "type": "text", + "content": "[18] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[19] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3d scenes" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14708-14718, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 150, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 172 + ], + "type": "text", + "content": "[21] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 173, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 238 + ], + "type": "text", + "content": "[22] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274-13285, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "text", + "content": "[23] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36(7):1325-1339, 2013." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 294, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 336 + ], + "type": "text", + "content": "[24] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[25] Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7122-7131, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 382, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 403 + ], + "type": "text", + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "type": "text", + "content": "[27] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "text", + "content": "[28] Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5253-5263, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "type": "text", + "content": "[29] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. Pare: Part attention regressor for 3d human body estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11127-11137, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "text", + "content": "[30] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In ICCV, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "text", + "content": "[31] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2261, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "text", + "content": "[32] Nikos Kolotouros, Georgios Pavlakos, and Kostas Dani-ilidis. Convolutional mesh regression for single-image human shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4501-4510, 2019." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[33] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15152" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 128 + ], + "type": "text", + "content": "Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XI, pages 620-640. Springer, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "text", + "content": "[34] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J Black, and Peter V Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6050-6059, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 229 + ], + "type": "text", + "content": "[35] Gun-Hee Lee and Seong-Whan Lee. Uncertainty-aware human mesh recovery from video by learning part-based 3d dynamics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12375–12384, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 285 + ], + "type": "text", + "content": "[36] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3383-3393, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 288, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 340 + ], + "type": "text", + "content": "[37] Zhi Li, Soshi Shimada, Bernt Schiele, Christian Theobalt, and Vladislav Golyanik. Mocapdeform: Monocular 3d human motion capture in deformable scenes. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 386 + ], + "type": "text", + "content": "[38] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "type": "text", + "content": "[39] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1954-1963, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 288, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 466 + ], + "type": "text", + "content": "[40] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphomer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12939-12948, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 468, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 288, + 521 + ], + "type": "text", + "content": "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 577 + ], + "type": "text", + "content": "[42] Xiyue Liu, Zhiyang Dou, Lei Wang, Boni Su, Tianyi Jin, Yong Guo, Jianjian Wei, and Nan Zhang. Close contact behavior-based Covid-19 transmission and interventions in a subway system. Journal of Hazardous Materials, 436:129233, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "type": "text", + "content": "[43] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "text", + "content": "[44] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[46] Haoyu Ma, Liangjian Chen, Deying Kong, Zhe Wang, Xingwei Liu, Hao Tang, Xiangyi Yan, Yusheng Xie, Shih-Yao Lin, and Xiaohui Xie. Transfusion: Cross-view fusion with transformer for 3d human pose estimation. arXiv preprint arXiv:2110.09554, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 173 + ], + "type": "text", + "content": "[47] Haoyu Ma, Zhe Wang, Yifei Chen, Deying Kong, Liangjian Chen, Xingwei Liu, Xiangyi Yan, Hao Tang, and Xiaohui Xie. Ppt: token-pruned pose transformer for monocular and multi-view human pose estimation. In ECCV, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 239 + ], + "type": "text", + "content": "[48] Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In 2018 International Conference on 3D Vision (3DV), pages 120-130. IEEE, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 242, + 546, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 546, + 297 + ], + "type": "text", + "content": "[49] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12309-12318, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 298, + 546, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 546, + 342 + ], + "type": "text", + "content": "[50] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In European Conference on Computer Vision, pages 752-768. Springer, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 343, + 546, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 546, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 546, + 388 + ], + "type": "text", + "content": "[51] Paschalis Panteleris and Antonis Argyros. Pe-former: Pose estimation transformer. In International Conference on Pattern Recognition and Artificial Intelligence, pages 3-14. Springer, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 388, + 546, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 546, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 546, + 442 + ], + "type": "text", + "content": "[52] Georgios Pavlakos, Nikos Kolotouros, and Kostas Daniilidis. Texturepose: Supervising human mesh estimation with texture consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 803-812, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 445, + 546, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 546, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 546, + 488 + ], + "type": "text", + "content": "[53] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human mesh recovery from multiple shots. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1485-1495, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 490, + 546, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 546, + 544 + ], + "type": "text", + "content": "[54] Georgios Pavlakos, Luyang Zhu, Xiaowei Zhou, and Kostas Daniilidis. Learning to estimate 3d human pose and shape from a single color image. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 459-468, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 546, + 546, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 546, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 546, + 612 + ], + "type": "text", + "content": "[55] Leonid Pishchulin, Eldar Insafutdinov, Siyu Tang, Bjoern Andres, Mykhaylo Andriluka, Peter V Gehler, and Bernt Schiele. Deepcut: Joint subset partition and labeling for multi person pose estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4929-4937, 2016." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 613, + 546, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 666 + ], + "type": "text", + "content": "[56] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 670, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 546, + 713 + ], + "type": "text", + "content": "[57] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), Nov. 2017." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15153" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[58] Yu Rong, Jingbo Wang, Ziwei Liu, and Chen Change Loy. Monocular 3d reconstruction of interacting hands via collision-aware factorized refinements. In 2021 International Conference on 3D Vision (3DV), pages 432-441. IEEE, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "type": "text", + "content": "[59] Michael Ryoo, AJ Piergiovanni, Anurag Arnab, Mostafa Dehghani, and Anelia Angelova. Tokenlearner: Adaptive space-time tokenization for videos. 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 218 + ], + "type": "text", + "content": "[60] Akash Sengupta, Ignas Budvytis, and Roberto Cipolla. Hierarchical kinematic probability distributions for 3d human shape and pose estimation from images in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11219-11229, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "type": "text", + "content": "[61] Zehong Shen, Zhi Cen, Sida Peng, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Learning human mesh recovery in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17038-17047, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "type": "text", + "content": "[62] Deepak Sridhar, Niamul Quader, Srikanth Muralidharan, Yaoxin Li, Peng Dai, and Juwei Lu. Class semantics-based attention for action detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13739-13748, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 385 + ], + "type": "text", + "content": "[63] Zhaoqi Su, Weilin Wan, Tao Yu, Lingjie Liu, Lu Fang, Wenping Wang, and Yebin Liu. Mulaycap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics, 28(4):1862-1879, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 431 + ], + "type": "text", + "content": "[64] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Michael J Black, and Tao Mei. Monocular, one-stage, regression of multiple 3d people. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11179-11188, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 475 + ], + "type": "text", + "content": "[65] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 478, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 510 + ], + "type": "text", + "content": "[66] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A survey. arXiv preprint arXiv:2203.01923, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 555 + ], + "type": "text", + "content": "[67] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 610 + ], + "type": "text", + "content": "[68] Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In European Conference on Computer Vision (ECCV), sep 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[69] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20460-20469, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[70] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "on pattern analysis and machine intelligence, 43(10):3349-3364, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 151 + ], + "type": "text", + "content": "[71] Wenjia Wang, Yongtao Ge, Haiyi Mei, Zhongang Cai, Qingping Sun, Yanjun Wang, Chunhua Shen, Lei Yang, and Taku Komura. Zolly: Zoom focal length correctly for perspective-distorted human mesh reconstruction. arXiv preprint arXiv:2303.13796, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 217 + ], + "type": "text", + "content": "[72] Wen-Li Wei, Jen-Chun Lin, Tyng-Luh Liu, and Hong-Yuan Mark Liao. Capturing humans in motion: Temporal-attentive 3d human pose and shape estimation from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13211–13220, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 547, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 547, + 263 + ], + "type": "text", + "content": "[73] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Progressive unsupervised learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2993-3002, June 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 265, + 547, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 547, + 330 + ], + "type": "text", + "content": "[74] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B. Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14561-14571, June 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 332, + 547, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 547, + 376 + ], + "type": "text", + "content": "[75] Yan Xia, Qiangqiang Wu, Wei Li, Antoni B Chan, and Uwe Stilla. A lightweight and detector-free 3d single object tracker on point clouds. IEEE Transactions on Intelligent Transportation Systems, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 377, + 547, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 547, + 442 + ], + "type": "text", + "content": "[76] Yan Xia, Yusheng Xu, Shuang Li, Rui Wang, Juan Du, Daniel Cremers, and Uwe Stilla. Soe-net: A self-attention and orientation encoding network for point cloud based place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11348–11357, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 445, + 547, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 547, + 488 + ], + "type": "text", + "content": "[77] Yuanlu Xu, Song-Chun Zhu, and Tony Tung. Denserac: Joint 3d pose and shape estimation by dense render-and-compare. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7760-7770, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 490, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 547, + 555 + ], + "type": "text", + "content": "[78] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch on imagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 558-567, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 557, + 547, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 547, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 547, + 611 + ], + "type": "text", + "content": "[79] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11038-11049, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 613, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 547, + 678 + ], + "type": "text", + "content": "[80] Wang Zeng, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11101-11111, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 680, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 547, + 713 + ], + "type": "text", + "content": "[81] Yuanhao Zhai, Le Wang, Wei Tang, Qilin Zhang, Junsong Yuan, and Gang Hua. Two-stream consensus network for weakly-supervised temporal action localization. In Euro-" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15154" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 328 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "text", + "content": "pean conference on computer vision, pages 37-54. Springer, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 160 + ], + "type": "text", + "content": "[82] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11446-11456, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 217 + ], + "type": "text", + "content": "[83] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 220, + 286, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 286, + 272 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 286, + 272 + ], + "type": "text", + "content": "[84] Nan Zhang, Li Liu, Zhiyang Dou, Xiyue Liu, Xueze Yang, Doudou Miao, Yong Guo, Silan Gu, Yuguo Li, Hua Qian, and Jianjian Wei. Close contact behaviors of university and school students in 10 indoor environments. Journal of Hazardous Materials, 458:132069, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 274, + 286, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 274, + 286, + 328 + ], + "spans": [ + { + "bbox": [ + 49, + 274, + 286, + 328 + ], + "type": "text", + "content": "[85] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 813-822, 2019." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15155" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/6ac05fcd-de6a-4657-a3f2-a897a887dcbb_content_list.json b/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/6ac05fcd-de6a-4657-a3f2-a897a887dcbb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bc21523b45108628b4f67e096344ab3e6ba1307e --- /dev/null +++ b/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/6ac05fcd-de6a-4657-a3f2-a897a887dcbb_content_list.json @@ -0,0 +1,1922 @@ +[ + { + "type": "text", + "text": "TRM-UAP: Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization", + "text_level": 1, + "bbox": [ + 91, + 128, + 879, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, Di Ming* \nSchool of Computer Science and Engineering, Chongqing University of Technology \nChongqing, China", + "bbox": [ + 151, + 203, + 815, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lyrl99804@qq.com, {xfeng, ylwang, yangwu, diming}@cqut.edu.cn", + "bbox": [ + 215, + 258, + 754, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aiming at crafting a single universal adversarial perturbation (UAP) to fool CNN models for various data samples, universal attack enables a more efficient and accurate evaluation for the robustness of CNN models. Early universal attacks craft UAPs depending on data priors. For more practical applications, the data-free universal attacks that make UAPs from random noises have aroused much attention recently. However, existing data-free UAP methods perturb all the CNN feature layers equally via the maximization of the CNN activation, leading to poor transferability. In this paper, we propose a novel data-free universal attack without depending on any real data samples through truncated ratio maximization, which we term as TRM-UAP. Specifically, different from the maximization of the positive activation in convolution layers, we propose to optimize the UAP generation from the ratio of positive and negative activations. To further enhance the transferability of universal attack, TRM-UAP not only performs the ratio maximization merely on low-level generic features via the truncation strategy, but also incorporates a curriculum optimization algorithm that can effectively learn the diversity of artificial images. Extensive experiments on the ImageNet dataset verify that TRM-UAP achieves a state-of-the-art average fooling rate and excellent transferability on different CNN models as compared to other data-free UAP methods. Code is available at https://github.com/RandolphCarter0/TRMUAP.", + "bbox": [ + 76, + 340, + 472, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 777, + 209, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early research [26] shows that tiny and imperceptible perturbations can seriously disturb the prediction results of deep neural networks (DNNs), especially for image recognition tasks [3]. Adversarial Examples (AEs), crafted by adding tiny perturbations deliberately to benign samples,", + "bbox": [ + 76, + 801, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are not only imperceptible in computer vision tasks but also prone to transfer among the DNN models. Therefore, AEs have been regarded as a serious threat to DNN models since the development of deep learning [1].", + "bbox": [ + 500, + 310, + 890, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To explore the impact of AEs, many methods are proposed to craft the adversarial perturbation that is highly transferable to various DNN models (i.e., with a high fooling rate on other DNN models). However, the AEs generated by these works are explicitly designed for some specific samples and often fail to perturb other samples that are even from the same dataset. Different from the aforementioned image-specific attack, the universal attack that generates image-agnostic universal adversarial perturbation (UAP) was proposed in [16]. The UAP in a universal attack setting is trained from prior knowledge, such as substitute data, surrogate model, etc. By adding the UAP to benign samples, universal attacks can generate numerous AEs all together in a very short period. Furthermore, the universal attack can fool most DNN models that are trained from similar datasets, and can greatly reduce the computational cost of crafting AEs, making adversaries more applicable to real scenarios than the image-specific attacks [4, 6, 34].", + "bbox": [ + 496, + 371, + 892, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Whereas, no matter the image-specific attacks or the universal attacks, well-annotated training data (e.g., [21, 9, 7]) or substitute data (e.g., [16, 12, 22, 19]) is required to generate AEs. In practice, obtaining a well-labeled large-scale dataset is challenging and costly, especially for some applications with critical security demand, where less prior knowledge is available. Recently, researchers have investigated data-free universal attack methods [18, 17, 14, 20, 33], where AEs are generated directly from random noises rather than data priors. Compared with previous methods that disturb the gradient or maximize the classification loss in data-dependent scenarios, current data-free UAP methods explore the feature-based perturbation, which tries to maximize the activation (e.g., ReLU) of convolutional neural network (CNN) features. It is shown that feature-based UAP methods achieved highly efficient and applicable universal attack without using any data priors [4, 6, 34]. How", + "bbox": [ + 496, + 643, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author: Di Ming", + "bbox": [ + 94, + 886, + 276, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4762", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ever, current data-free UAP methods only consider the positive activation [18, 17] and perturb all layers of CNN features equally [20, 14]. As a result, UAPs crafted by surrogate models are hard to transfer to target models [6, 34].", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards enhancing the transferability of data-free UAPs, we propose a novel data-free universal attack method called TRM-UAP, which formulates the UAP generation as a truncated ratio maximization problem. In particular, TRM-UAP attempts to over-fire positive neurons so that extracted features from multiple CNN layers can be fully disrupted. Besides the maximization of positive neurons, the proper activation on negative neurons may also be helpful. Hence, TRM-UAP performs a ratio maximization of positive and negative activations, which facilitates some negative neurons transiting to be positive. Moreover, some feature-based adversarial attacks [29, 35] reveal that multiple layers of CNN features are not equally important for disturbing the image classification [11, 28]. To further improve the transferability of universal attack, TRM-UAP performs ratio maximization only on parts of CNN activation layers. Specifically, a truncation strategy is proposed to truncate the activation of high-level convolutions and retain the activation of shallow convolutions to train perturbations. The intuition is that low-level CNN features are shown to provide more generic (data-independent) feature representations than high-level semantic features [30]. Thus the disturbed low-level features are expected to be able to transfer attacks across different CNN models, leading to highly transferable UAPs. Furthermore, a curriculum learning [2] based optimization algorithm is introduced to utilize artificial images and explore the diversity of input.", + "bbox": [ + 75, + 152, + 472, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In general, the main contributions of our work can be summarized as follows:", + "bbox": [ + 76, + 561, + 468, + 589 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel data-free universal attack method to craft image-agnostic adversarial perturbations without utilizing any real data samples during training.", + "- To the best of our knowledge, we are the first to formulate the data-free universal attack as a truncated ratio maximization problem. The proposed TRM-UAP enhances the transferability of UAPs from three perspectives: (i) the maximization on CNN features for crafting UAPs is enhanced by both maximization of positive activation and minimization of negative activation. (ii) the ratio maximization on truncated feature layers improves the generalizability of universal attacks to CNN models. (iii) artificial images are incorporated into curriculum optimization algorithm to strengthen the UAP dominance on feature activations.", + "- Extensive experiments on the ImageNet dataset show that TRM-UAP obtains the highest average fooling rate compared with other data-free UAP methods, indicating that truncated ratio maximization can greatly improve the transferability of universal attack." + ], + "bbox": [ + 94, + 594, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-specific attacks: The attacks that try to craft perturbations for corresponding images by utilizing training or substitute data are termed as image-specific attacks. Typical image-specific attacks, such as Fast Gradient Sign Method (FGSM) [9], I-FGSM and MI-FGSM [7], Project Gradient Descent (PGD) [15], and Output Diversified Sampling (ODS) [27], etc., usually make use of the gradient information from well-trained target CNN model (i.e., white-box attack) or surrogate model (i.e., black-box attack) to disturb images. However, since gradient-based attacks often require a number of iterations, the computational cost of making AEs becomes extremely expensive as the number of examples increases. Recently, feature-based attacks that try to craft AEs from the perspective of CNN features have received broad concern [29, 8, 35]. For example, Neuron Attribution based Attack (NAA) [35] conducted feature-level attacks by deploying a neuron attribution method to estimate the importance of neurons at multiple CNN layers.", + "bbox": [ + 496, + 114, + 890, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data-dependent universal attacks: The universal attack aims at training a single perturbation that is independent of any specific sample, i.e., UAP. By adding a single UAP to various benign samples, numerous AEs can be generated very efficiently. Until recently, most of the universal attacks are still data-dependent, meaning that the perturbations are learned from data priors, such as training data, substitute data, etc. For example, the primary universal attack [16] tried to seek a universal perturbation for a set of training samples. On the other hand, there are some works [22, 19] applying generative adversarial models to craft perturbations. Recently, Zhang et al. [32] proposed to generate UAPs by training perturbations on the proxy datasets.", + "bbox": [ + 496, + 387, + 890, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data-free universal attacks: The data-free universal attack makes UAP via random noise rather than prior knowledge from data directly. It is considered the most applicable attack for real applications in the adversarial attack field. However, since the information of both data and target models is unknown, the data-free universal attack is extremely challenging and only a few recent works focus on it. Mopuri et al. [18] first propose a data-free universal attack, i.e., the Fast Feature Fool (FFF) method, by maximizing the feature activations across all the CNN layers. In addition, they further propose a Generalizable Data-free UAP (GDUAP) to improve the FFF attack via optimizing the UAP training process with a saturation check strategy [17]. Besides, AAA [20] learns a generative model to make UAPs based on the pre-defined class impression. However, as the number of image categories increases, the cost of acquiring class impressions will also broaden. Prior-Driven Uncertainty Approximation (PD-UA) [14] proposes to craft UAPs by maximizing the uncertainty approximation of the model. Cosine-UAP [33] introduces the cosine similarity to train UAPs in a self-supervised way. Because of the shortage of", + "bbox": [ + 496, + 584, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4763", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "prior knowledge, current data-free UAP methods generally present poor transferability across various CNN models.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 133, + 210, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the motivation of maximizing the ratio of activations. After that, we present the details of the proposed truncated ratio maximization method to craft the UAP in a strictly data-free fashion.", + "bbox": [ + 76, + 159, + 468, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 228, + 205, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To fool CNN models $f$ , UAP [16] tries to find the universal perturbation $\\pmb{v}$ which maximizes the classification loss $\\mathcal{L}$ (e.g., cross entropy) over the data distribution $\\mathbb{D}$ :", + "bbox": [ + 76, + 251, + 468, + 296 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\boldsymbol {v} \\sim \\mathbb {S}} \\underset {(\\boldsymbol {x}, \\boldsymbol {y}) \\sim \\mathbb {D}} {\\mathbb {E}} [ \\mathcal {L} (f (\\boldsymbol {v} + \\boldsymbol {x}), \\boldsymbol {y})) ] \\quad \\text {s . t .} \\| \\boldsymbol {v} \\| _ {p} \\leq \\epsilon \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 306, + 468, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{S}$ is the perturbation space, $(\\pmb{x},\\pmb{y})$ are the data and the class label (or one-hot ground truth) sampled from $\\mathbb{D}$ , $f(\\cdot)$ is the output of CNN models, and $\\epsilon$ is the constant to constrain the perturbation $\\pmb{v}$ in the $\\ell_p$ -norm bound ( $e.g., p = 0,1,2,\\infty$ ). To be consistent with previous works [18, 17, 32], we adopt the $\\ell_{\\infty}$ -norm to realize the imperceptibility of perturbation for all analyzes throughout this paper.", + "bbox": [ + 76, + 340, + 468, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the training data of the target model is usually unavailable and hard to access, making the data-dependence approaches not suitable for practical adversarial attacks. As a consequence, Mopuri et al. [18] propose the Fast Feature Fool (FFF) method to maximize the CNN activations in a strictly data-free fashion:", + "bbox": [ + 75, + 446, + 468, + 536 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\boldsymbol {v}} \\quad | | \\mathcal {A} ^ {(i)} (\\boldsymbol {v}) | | _ {2}, \\quad \\text {f o r} i = 1, 2, \\dots , L \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 546, + 468, + 573 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s . t .} \\quad | | \\boldsymbol {v} | | _ {\\infty} \\leq \\epsilon\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 569, + 251, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{A}^{(i)}(\\pmb{v}) = \\mathrm{Activation}(\\mathcal{C}^{(i)}(\\pmb{v}))$ (e.g.,ReLU activation), $\\mathcal{C}^{(i)}(\\pmb{v})$ is the output of the $i$ -th convolution layer, and $L$ is the total number of convolution layers. It can be seen that this typical data-free universal attack method aims at accumulating the errors gradually through multiple convolution layers to enlarge the classification loss, finally leading to the misclassification of perturbed samples $\\pmb{v} + \\pmb{x}$ .", + "bbox": [ + 76, + 593, + 468, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Maximizing the Ratio of Activations", + "text_level": 1, + "bbox": [ + 76, + 709, + 390, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from previous data-free universal methods [18, 17] that only maximize the activation of convolution layer (see Fig. 1 (a)), we propose to craft the universal adversarial perturbation $\\pmb{v}$ via maximizing the ratio of activations:", + "bbox": [ + 76, + 732, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\boldsymbol {v}} \\frac {| | \\mathcal {C} _ {+} ^ {(i)} (\\boldsymbol {v}) | | _ {2}}{| | \\mathcal {C} _ {-} ^ {(i)} (\\boldsymbol {v}) | | _ {2}}, \\quad \\text {f o r} i = 1, 2, \\dots , L \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 801, + 468, + 840 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s . t .} \\quad | | \\boldsymbol {v} | | _ {\\infty} \\leq \\epsilon\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 844, + 251, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where positive activation and negative activation in $i$ -th convolution layer are defined as $\\mathcal{C}_{+}^{(i)}(\\pmb{v}) = \\max (\\mathcal{C}^{(i)}(\\pmb{v}),0)$ and", + "bbox": [ + 76, + 868, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cd4db687608d88f9c2802209819768be2d60e16af792fcbe04bb63b9ff227737.jpg", + "image_caption": [ + "Figure 1. The illustration of previous positive maximization [18, 17] and our proposed ratio maximization." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 875, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathcal{C}_{-}^{(i)}(\\pmb {v}) = \\min (\\mathcal{C}^{(i)}(\\pmb {v}),0)$ respectively, and their magnitudes are measured by $\\ell_2$ -norm.", + "bbox": [ + 498, + 333, + 888, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Towards further enlarging the fooling probability, this ratio formulation (3) tries to increase the magnitude and the number of positive activation via maximizing positive activation and minimizing negative activation simultaneously. Specifically speaking, ratio maximization is a dynamic process. Firstly, minimizing the negative activation could lead to the change of sign in some negative features whose values are close to zero, i.e., jumping from a tiny negative value to a tiny positive value, as shown in Fig. 1 (b). Afterward, at the next iteration, maximizing the positive activation will continue to increase the magnitude of above-mentioned features from a tiny positive value to a large positive value.", + "bbox": [ + 496, + 366, + 890, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, with the decrease in the number of negative activation and the increase in the number of positive activation, our proposed formulation (3) can achieve a higher ratio of positive activation as compared to previous data-free universal attacks, e.g., the formulation (2). Meanwhile, it also indicates that our ratio maximization could craft the UAP with more significant attack intensity so as to increase the fooling probability further.", + "bbox": [ + 496, + 547, + 890, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Truncated Ratio Maximization", + "text_level": 1, + "bbox": [ + 500, + 680, + 772, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Features extracted from shallow to deep convolution layers behave differently in crafting the adversarial perturbation. LAFEAT [31] analyzes the influence of each intermediate convolution layer on the misclassification, and attacks features only in a specific layer. On the other hand, GD-UAP [17] thoroughly studied the relative change in feature activations of adversarial examples as compared to original images at multiple CNN layers. However, GD-UAP attacks the feature activations extracted from all convolution layers together, without taking layer-wise feature patterns into consideration. According to [30], low-level features in shallow layers have generic patterns, while high-level features in deep layers have specific patterns. It can be seen that", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4764", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3d806e9afd92dc1567563c9adf926db09e929a69e828aaa62d22d4334f4e8c45.jpg", + "image_caption": [ + "Figure 2. Truncated positive and negative activation losses, where cumulative shaded areas are used to craft the UAP." + ], + "image_footnote": [], + "bbox": [ + 135, + 90, + 410, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "not all convolution layers contribute positively to increasing the positive activation and the objective loss. Therefore, based on ratio maximization defined in formulation (3), we propose a truncated ratio maximization method to further enhance the intensity and the transferability of UAPs.", + "bbox": [ + 75, + 325, + 467, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For notational simplicity, we begin by defining the ratio maximization problem in $i$ -th convolution layer as $\\mathcal{L}^{(i)}(\\pmb{v}) = \\mathcal{L}_{+}^{(i)}(\\pmb{v}) / \\mathcal{L}_{-}^{(i)}(\\pmb{v})$ , where $\\mathcal{L}_{+}^{(i)}(\\pmb{v}) = ||\\mathcal{C}_{+}^{(i)}(\\pmb{v})||_{2}$ and $\\mathcal{L}_{-}^{(i)}(\\pmb{v}) = ||\\mathcal{C}_{-}^{(i)}(\\pmb{v})||_{2}$ . To maximize the ratio of activations in convolution layers altogether, the overall loss function of ratio maximization is reformulated as", + "bbox": [ + 76, + 401, + 467, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\boldsymbol {v}) = \\sum_ {i = 1} ^ {L} \\log \\mathcal {L} ^ {(i)} (\\boldsymbol {v}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 501, + 468, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where log rescales the activation to prevent gradient explosions. Different from other methods that craft UAPs based on all convolution layers or a specific convolution layer, we propose to truncate the computation of positive and negative activations in deep convolution layers, since specific feature patterns are not beneficial for improving the quality of UAPs. Assuming that the truncation rate $(TR)$ is defined as $TR = \\lfloor (L - l) / L\\rfloor \\%$ , we make truncation on $(l + 1)$ -th convolution layer by setting $\\mathcal{L}_{+}^{(i)}(\\pmb {v}) = \\tau$ $(i > l)$ and $\\mathcal{L}_{-}^{(i)}(\\pmb {v}) = \\tau$ $(i > l)$ , where $\\tau$ is a small positive number such as $1e - 9$ . As a result, $\\mathcal{L}(\\pmb {v})$ can be equivalently rewritten as $\\mathcal{L}(\\pmb {v}) = \\sum_{i = 1}^{L}\\log \\mathcal{L}^{(i)}(\\pmb {v}) = \\sum_{i = 1}^{l}\\log \\mathcal{L}^{(i)}(\\pmb {v})$", + "bbox": [ + 75, + 549, + 467, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, $\\mathcal{L}_{+}^{(i)}(\\pmb{v})$ and $\\mathcal{L}_{-}^{(i)}(\\pmb{v})$ not only have different scales but also behave differently in various layers for maximizing the ratio loss, which could lead to a degradation of the attack performance. To resolve these problems, we first reformulate the ratio loss in $i$ -th convolution layer as", + "bbox": [ + 75, + 739, + 467, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\alpha} ^ {(i)} (\\boldsymbol {v}) = \\frac {\\mathcal {L} _ {+} ^ {(i)} (\\boldsymbol {v})}{(\\mathcal {L} _ {-} ^ {(i)} (\\boldsymbol {v})) ^ {\\alpha}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 823, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is a scaling hyperparameter to adjust the relative importance between positive and negative activation losses.", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd1b9880d0864a22bc1b11b2a6e89447b521da93c5cfcf74790729c2e5b06279.jpg", + "image_caption": [ + "Figure 3. The mean positive activation loss in various convolution layers with different truncation settings." + ], + "image_footnote": [], + "bbox": [ + 532, + 92, + 851, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Secondly, positive and negative truncation rates are defined as $PTR = \\lfloor (L - l') / L \\rfloor \\%$ and $NTR = \\lfloor (L - l'') / L \\rfloor \\%$ respectively. That is to say, we make truncation on different layers for positive and negative activation losses by setting $\\mathcal{L}_{+}^{(i)}(\\pmb{v}) = \\tau (i > l')$ and $\\mathcal{L}_{-}^{(i)}(\\pmb{v}) = \\tau (i > l'')$ , see Fig. 2. Finally, the overall loss function $\\mathcal{L}(\\pmb{v})$ can be reduced to", + "bbox": [ + 498, + 327, + 890, + 421 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\boldsymbol {v}) = \\sum_ {i = 1} ^ {L} \\log \\mathcal {L} _ {\\alpha} ^ {(i)} (\\boldsymbol {v}) \\\\ = \\sum_ {i = 1} ^ {l ^ {\\prime}} \\log \\mathcal {L} _ {+} ^ {(i)} (\\boldsymbol {v}) - \\alpha \\cdot \\sum_ {i = 1} ^ {l ^ {\\prime \\prime}} \\log \\mathcal {L} _ {-} ^ {(i)} (\\boldsymbol {v}) + c \\\\ \\propto \\sum_ {i = 1} ^ {l ^ {\\prime}} \\log \\mathcal {L} _ {+} ^ {(i)} (\\boldsymbol {v}) - \\alpha \\cdot \\sum_ {i = 1} ^ {l ^ {\\prime \\prime}} \\log \\mathcal {L} _ {-} ^ {(i)} (\\boldsymbol {v}), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 430, + 888, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c = ((1 - \\alpha)\\cdot L - l^{\\prime} + \\alpha \\cdot l^{\\prime \\prime})\\cdot \\log \\tau$ , i.e., a constant.", + "bbox": [ + 500, + 517, + 888, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consequently, for data-free universal adversarial attack, we propose a truncated ratio maximization method (TRMUAP) to craft the perturbation $\\pmb{v}$ satisfying", + "bbox": [ + 498, + 532, + 888, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max _ {\\boldsymbol {v}} \\sum_ {i = 1} ^ {l ^ {\\prime}} \\log \\left| \\left| \\mathcal {C} _ {+} ^ {(i)} (\\boldsymbol {v}) \\right| \\right| _ {2} - \\alpha \\cdot \\sum_ {i = 1} ^ {l ^ {\\prime \\prime}} \\log \\left| \\left| \\mathcal {C} _ {-} ^ {(i)} (\\boldsymbol {v}) \\right| \\right| _ {2} \\tag {7} \\\\ \\text {s . t .} \\quad | | \\boldsymbol {v} | | _ {\\infty} \\leq \\epsilon \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 588, + 888, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{v}$ is constrained by $\\ell_{\\infty}$ -norm and bound $\\epsilon$ . To verify the previous claim, we compute the mean positive activation loss in various convolution layers via feeding UAPs into CNN models (e.g., VGG19). As shown in Fig. 3, TRM-UAP $(PTR \\neq NTR$ , i.e., $l' \\neq l'' < L)$ generates a larger cumulative positive activation loss than RM-UAP $(l' = l'' = L)$ and TRM-UAP $(PTR = NTR$ , i.e., $l' = l'' < L)$ , which could lead to a further increase in fooling rates.", + "bbox": [ + 498, + 656, + 890, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Curriculum Optimization Algorithm", + "text_level": 1, + "bbox": [ + 500, + 786, + 816, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For purpose of improving the diversity of inputs, we utilize Gaussian noise and jigsaw image [33] to generate artificial images, which are fed into surrogate models together with the perturbation. To resemble training samples in real-world datasets, the mean filter is used to smooth the edge between different regions in jigsaw images. However, it", + "bbox": [ + 498, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4765", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "will be difficult for the optimization algorithm to converge if artificial images with complex patterns are provided at the beginning of the training procedure. Inspired by curriculum learning [2], we start the training with simple artificial images and then gradually feed more complex artificial images into surrogate models. Specifically, the generation of artificial images is controlled by the distribution parameter and the training iteration, defined as:", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {1} \\prec D _ {2} \\prec \\dots \\prec D _ {n},\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 219, + 346, + 234 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {t} = \\left\\{\\boldsymbol {x} \\mid \\boldsymbol {x} \\sim P \\left(\\theta_ {0}, t\\right) \\right\\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 234, + 467, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D_{t}$ is a set of artificial images $\\pmb{x}$ sampled at $t$ -th iteration $(t = 1,2,\\dots ,n)$ , $P$ is the distribution of artificial samples with the default parameter $\\theta_0$ , and $D_{a} \\prec D_{b}$ denotes that the pattern of artificial images in $D_{b}$ is more complex than $D_{a}$ . In detail, as the number of training iterations $t$ increases, the distribution parameter $\\theta_0$ (e.g., the standard deviation of Gaussian distribution and the frequency of jigsaw images) is gradually increasing. For $t$ -th iteration, our curriculum optimization algorithm is maximizing", + "bbox": [ + 75, + 261, + 472, + 398 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = \\frac {1}{| D _ {t} |} \\sum_ {\\boldsymbol {x} \\in D _ {t}} \\left(\\sum_ {i = 1} ^ {l ^ {\\prime}} \\log \\mathcal {L} _ {+} ^ {(i)} (\\boldsymbol {v} + \\boldsymbol {x}) - \\alpha \\cdot \\sum_ {i = 1} ^ {l ^ {\\prime \\prime}} \\log \\mathcal {L} _ {-} ^ {(i)} (\\boldsymbol {v} + \\boldsymbol {x})\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 406, + 467, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|D_t|$ is the number of artificial images in the set $D_t$ .", + "bbox": [ + 75, + 462, + 464, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The entire procedure of our proposed TRM-UAP method is summarized in Algorithm 1. During the training process, the perturbation vector $\\pmb{v}$ is randomly initialized by the uniform distribution $\\mathcal{U}$ , and a surrogate dataset is used to validate the convergence of optimization for every $H$ iterations. $l^{\\prime}$ and $l''$ represent the retained layers of positive and negative activations corresponding to truncation rates $PTR$ and $NTR$ respectively. The scaling hyperparameter $\\alpha$ adjusts the ratio between truncated positive and negative activation losses. The algorithm is converged only if either the iteration number reaches the maximum $T$ or the fooling rate test reaches the threshold $F_{max}$ . Similar to GD-UAP [17], we run a saturation rate test for each iteration to dynamically compress the perturbation vector only if the saturation rate $\\hat{r}$ is smaller than the predefined threshold $r$ .", + "bbox": [ + 75, + 477, + 472, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Connection to Other Data-Free UAP Methods", + "text_level": 1, + "bbox": [ + 76, + 710, + 465, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When positive and negative truncation rates are set as $PTR = 0\\%$ and $NTR = 100\\%$ (i.e., $l' = L, l'' = 0$ ), this means all the negative activations are truncated and all the positive activations are retained. Thus, the proposed TRM-UAP defined in Eq. (7) is reduced to GD-UAP [17]. It can be seen that GD-UAP is only a special case of TRM-UAP under the data-free universal attack setting. As compared to GD-UAP, our TRM-UAP improves both intensity and transferability of crafted UAPs via truncated ratio maximization. The detail of their differences will be further discussed in the experiments.", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Curriculum optimization algorithm for solving truncated ratio maximization problem to craft universal adversarial perturbations", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: surrogate CNN model $f$ , limitation value $\\epsilon$ , learning rate $\\eta$ , positive and negative truncation rates $PTR$ , $NTR$ , scaling hyperparameter $\\alpha$ , maximum iteration number $T$ , convergence threshold $F_{max}$ , validation test hyperparameter $H$ , saturation threshold $r$", + "bbox": [ + 498, + 137, + 893, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: universal adversarial perturbation $\\pmb{v}$", + "bbox": [ + 500, + 214, + 797, + 228 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize $\\pmb{v}_0\\sim \\mathcal{U}(-\\epsilon ,\\epsilon)$ $t = 0,F = 0$", + "2: while $t < T$ and $F < F_{\\text{max}}$ do", + "3: $t = t + 1$", + "4: Generate the artificial image set $D_{t}$ via Eq. (8)", + "5: Compute the gradient $\\nabla \\mathcal{L}_t$ of the loss $\\mathcal{L}_t$ in Eq. (9)", + "6: Update $\\pmb{v}_t = \\pmb{v}_{t-1} + \\eta \\cdot \\nabla \\mathcal{L}_t$", + "7: Clip $\\pmb{v}_t = \\min (\\epsilon ,\\max (\\pmb {v}_t, - \\epsilon))$", + "8: Compute the saturation rate $\\hat{r}$ and adjust $\\pmb{v}_t$ if $\\hat{r} < r$", + "9: Conduct the fooling rate test $FR$ if $t\\% H == 0$ and $F = F + 1$ if $FR$ not the best fooling rate" + ], + "bbox": [ + 509, + 231, + 890, + 381 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "10: end while", + "11: return $\\mathbf{v}_t$" + ], + "bbox": [ + 504, + 383, + 599, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 444, + 632, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setup: By following the experiment setup in existing data-free universal methods [18, 17, 20, 14], we evaluated the proposed TRM-UAP method on the validation set of ImageNet [23] with the classical pre-trained CNN models including AlexNet [13], VGG16 [24], VGG19 [24], ResNet152 [10] and GoogleNet [25].", + "bbox": [ + 496, + 470, + 890, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Criteria: We used the fooling rate $(FR)$ proposed by data-free universal methods [18, 17] as the evaluation metric. Particularly, a higher fooling rate represented higher attack success rate and better transferability. In addition, we used the logit loss of C&W attack [5] to further evaluate the transferability of data-free UAPs.", + "bbox": [ + 496, + 564, + 890, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparative Methods: We made a comparison with data-free universal approaches, including FFF [18], AAA [20], GD-UAP [17], PD-UA [14] and Cosine-UAP [33]. Note that GD-UAP [17] trained models under different setups, for a fair comparison, we reproduced the results of GD-UAP [17] by making the setup consistent with TRM-UAP.", + "bbox": [ + 496, + 657, + 890, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details: All of our experiments were implemented on PyTorch with a single NVIDIA GeForce RTX 3090Ti GPU. Following the common setting [18, 17], we set $\\epsilon = 10 / 255$ to restrict changes of the perturbation. The maximum iteration $T$ was set as 10000, and the saturation threshold $r$ was set to $0.001\\%$ . The value range of hyperparameters $PTR, NTR \\in [0, 1]$ denotes the truncation rate of positive and negative activations respectively. We set the scaling hyperparameter $\\alpha$ and truncation rates with appropriate values for different CNN models.", + "bbox": [ + 496, + 750, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4766", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/74e3c5defd1bd3781f259e0fecfb7781e8fdada96f5ae0ddc1dac94346a44370.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AttackAlexNetVGG16VGG19ResNet152GoogleNetAverage
FFF80.9247.1043.62-56.44-
AAA89.0471.5972.8460.7275.2873.89
GD-UAP85.2490.0187.3445.9645.8764.65
PD-UA-70.6964.9846.3967.12-
Cosine-UAP91.0789.4886.8165.3587.5784.08
TRM-UAP(Ours)93.53±0.0794.30±0.1591.35±0.3067.46±0.3585.32±0.0486.39
", + "bbox": [ + 132, + 88, + 834, + 209 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/48080ed138239a037176c0698ad4f11197de1d091466e7fd54eb6269d9dc5b19.jpg", + "table_caption": [ + "Table 1. The fooling rate $(FR)$ of our proposed TRM-UAP and other data-free universal methods. We show the mean and standard deviation of $FR$ with five runs." + ], + "table_footnote": [], + "table_body": "
AlexNetVGG16VGG19ResNet152GoogleNet
AlexNet93.53±0.0760.10±0.2457.08±0.1527.31±0.3032.70±0.22
VGG1647.53±0.5194.30±0.1289.68±0.1461.43±0.4053.95±0.59
VGG1946.01±0.4489.82±0.1591.35±0.3047.19±0.6646.48±0.78
ResNet15253.56±0.7577.20±0.3573.30±0.4167.46±0.3557.54±0.50
GoogleNet60.10±1.1679.66±0.9579.98±1.0658.85±1.9485.32±0.04
", + "bbox": [ + 191, + 256, + 777, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The transferability of the perturbation crafted by our TRM-UAP method (rows: target models; columns: surrogate models). The fooling rate $(FR)$ on the diagonal represents white-box attacks (i.e., UAP is crafted by the target model itself), and other off-diagonal values are black-box attacks (i.e., UAP is crafted by non-target models.). All results of mean and standard deviation are computed on average with five turns.", + "bbox": [ + 75, + 373, + 893, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 76, + 460, + 217, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed TRM-UAP is applied to craft UAPs on five CNN models and create AEs to attack these CNN models separately on the ImageNet validation set. We set $\\alpha \\in \\{1.0, 1.0, 0.5, 1.5, 1.3\\}$ , $PTR \\in \\{0.0, 0.2, 0.2, 0.7, 0.4\\}$ , $NTR \\in \\{0.8, 0.3, 0.2, 0.8, 0.3\\}$ , corresponding to AlexNet, VGG16, VGG19, ResNet152, GoogleNet. The attack performance compared with other data-free universal methods is shown in Table 1. Note that the results of GD-UAP were reproduced by our best effort on PyTorch for a fair comparison. As compared to other methods, our TRM-UAP method improves the fooling rate in most CNN models. Although Cosine-UAP achieves a higher $FR$ in GoogleNet, TRM-UAP improves the $FR$ around $1 \\sim 5\\%$ on other four models and further improves the average $FR$ of all five models more than $2\\%$ . Besides, it is shown in Fig. 4 that the visualization of crafted UAPs has clearly particular image patterns with abundant features. We also visualized AEs of TRM-UAP attack in Fig. 6 and found that high probabilities of AEs were given to the labels of incorrect class.", + "bbox": [ + 75, + 487, + 472, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further used the UAP made by surrogate models to attack other target models to verify the transferability. Table 2 represented the attack performance in different attack settings. The models in rows denote target models, and the models in columns denote surrogate models for crafting the UAP. The results on the diagonal denote the white-box attack, while the rest are transferable attacks in the black-box setting. Most TRM-UAP attacks perform excellently in", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "transferability, which indicates that the universal perturbations crafted by TRM-UAP are generalized well to transfer attacks on target models.", + "bbox": [ + 496, + 460, + 893, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluating the Transferability", + "text_level": 1, + "bbox": [ + 498, + 513, + 769, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further compare the effectiveness of universal attacks, we used the logit loss of C&W attack [5] to evaluate the transferability of GD-UAP, Cosine-UAP, and TRM-UAP. The logit loss is defined as $loss = (\\max_{j \\neq t} F(\\boldsymbol{v} + \\boldsymbol{x})_j) - F(\\boldsymbol{v} + \\boldsymbol{x})_t$ , where $F(\\cdot)_j$ is the $j$ -th output in the logits layer of network and $t$ is the index of ground truth. A larger logit loss value represents better transferability of attacks to CNN models. The logit loss of original images is calculated as a baseline. Note that the results of GD-UAP and Cosine-UAP are reproduced with our best effort on PyTorch. As shown in Fig. 5, TRM-UAP achieves a higher logit loss than GD-UAP and Cosine-UAP in most cases. Additionally, the discrepancy among GD-UAP, Cosine-UAP, and TRM-UAP becomes significant in deep CNN models (e.g., ResNet152 and GoogleNet). Therefore, TRM-UAP achieves relatively higher logit losses in both white-box and black-box attack settings, which indicates a better transferability across various CNN models.", + "bbox": [ + 496, + 537, + 893, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Parameter Study", + "text_level": 1, + "bbox": [ + 498, + 816, + 669, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, the influence of ratio maximization on attack performance is studied. Here we mainly focused on validating the effectiveness of truncated negative activation in deep CNN models (e.g., ResNet152 and GoogleNet). The experi", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4767", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/65c9d802a59f4dcb81e3bb2f629100341f6cd3fcdc0b714697823fb1fac9f3fb.jpg", + "image_caption": [ + "(a) AlexNet" + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 223, + 193 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b3d3196e40adb2a8c67e651bcb9c67e61ae87efae8b7629b9d7eebc61232f8a.jpg", + "image_caption": [ + "(b) VGG16" + ], + "image_footnote": [], + "bbox": [ + 251, + 90, + 388, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d0f65dd6835d1346f5582a0f4930288b9eedb01e701bb03ed2c2121bcc69f9b.jpg", + "image_caption": [ + "(c) VGG19" + ], + "image_footnote": [], + "bbox": [ + 418, + 90, + 553, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c573be2aaf3325f72df9cf636b44987fe6db48ffb6c2e2550e3c8b93bebc8bc.jpg", + "image_caption": [ + "(d) ResNet152" + ], + "image_footnote": [], + "bbox": [ + 584, + 90, + 718, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/07d9ff150d5f58d2a1fba9577c128d092d352195bb49c0301f60596993b9fa6a.jpg", + "image_caption": [ + "(e) GoogleNet" + ], + "image_footnote": [], + "bbox": [ + 748, + 89, + 883, + 193 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9840ca6ae6a1cec36052939a5a34a5c39b25a980b35c684409de51b5abd6c74.jpg", + "image_caption": [ + "Figure 4. The visualization of UAPs crafted by our TRM-UAP method. For the best visualization in color, we enhance the value of the pixel to [0, 255].", + "(a) AlexNet" + ], + "image_footnote": [], + "bbox": [ + 83, + 262, + 241, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/653907a0c232758cb095a21ffe5995ca6c5497a2b86518feee531c915091ac34.jpg", + "image_caption": [ + "(b) VGG16" + ], + "image_footnote": [], + "bbox": [ + 243, + 262, + 405, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0f231656c84bd49878ea6b3134009c957cfc20c7b313d705c8a8718447b12cda.jpg", + "image_caption": [ + "(c) VGG19" + ], + "image_footnote": [], + "bbox": [ + 408, + 262, + 570, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f14feee229e1d871c0c9f9a4bf3510d048914bbc72da18f37fdb2a00fb66fa2b.jpg", + "image_caption": [ + "(d) ResNet152" + ], + "image_footnote": [], + "bbox": [ + 573, + 262, + 728, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36beb49379502e532a7e458717f469680dd7428a62426c9cf841c1ad959c2396.jpg", + "image_caption": [ + "(e) GoogleNet" + ], + "image_footnote": [], + "bbox": [ + 730, + 262, + 887, + 335 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d200e91fbd33143f9c0679bf72f2adaf5a50e390a19fae2b177fe7024a2f0c6.jpg", + "image_caption": [ + "Figure 5. The logit loss of original samples and adversarial samples crafted by GD-UAP, Cosine-UAP, our TRM-UAP methods. In each sub-figure, the subtitle indicates the surrogate model for crafting UAPs, and target models are shown on the horizontal axis. Original samples are classified correctly by the target model and exploited to craft AEs for comparisons. For each attack setting, we compute the average of the logit loss over all the samples.", + "bullet train(98.15%)" + ], + "image_footnote": [], + "bbox": [ + 86, + 431, + 174, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a432cc4dffc8d6ca71a39e3cb5fe1820ebee297308efd5c4adfbacffc937c49.jpg", + "image_caption": [ + "puffer(99.89%)" + ], + "image_footnote": [], + "bbox": [ + 181, + 433, + 267, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0494d384130a691d241b67c8c68102120a3dfff002e4c448a3a996c3597b8a7.jpg", + "image_caption": [ + "quilt(45.63%)", + "Figure 6. Examples of our TRM-UAP attack tested on VGG16 and the perturbation crafted by VGG19 (top: original examples; bottom: adversarial examples). The annotation represents the classification label and the probability predicted by CNNs." + ], + "image_footnote": [], + "bbox": [ + 88, + 505, + 174, + 570 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0a84fbccd57c6376f0c62579ad8e800087dc0eafca838fd508e726eadfeb3b6.jpg", + "image_caption": [ + "theater curtain(72.82%)" + ], + "image_footnote": [], + "bbox": [ + 181, + 505, + 271, + 570 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2229b9cbff59ff4b52fb8a35ccf0888fea7ff1ec65ab48d6a669fe35af48b83c.jpg", + "image_caption": [ + "monastery(34.69%)" + ], + "image_footnote": [], + "bbox": [ + 274, + 433, + 361, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06e6dafd85668f5dc017d538c8984867aed7a23d311e63f3c5ad50d920fa6f07.jpg", + "image_caption": [ + "church(43.55%)" + ], + "image_footnote": [], + "bbox": [ + 274, + 505, + 361, + 570 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/80fb138084f9b9a0ba6f8d92ce61ed8391114c0a95a22fd87cd9e3195cda1559.jpg", + "image_caption": [ + "shoji(87.62%)" + ], + "image_footnote": [], + "bbox": [ + 367, + 433, + 454, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fea73f939316b40e8544950656a81ac387fca478f012d0152d5a5d38dd1aa6c3.jpg", + "image_caption": [ + "shower curtain(94.35%)" + ], + "image_footnote": [], + "bbox": [ + 367, + 505, + 457, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "imental setting is the same as Section 4.1. Based on this, we also crafted UAPs by truncated positive activation only, i.e., $PTR$ remains unchanged and $NTR = 100\\%$ . As shown in Fig. 7, the ablation experiment indicates that truncated negative activation combined with truncated positive activation improves the $FR$ by a large margin (around $8 \\sim 18\\%$ ) as compared to standalone truncated positive activation, thus enhancing both intensity and transferability of UAPs significantly.", + "bbox": [ + 75, + 672, + 468, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To explore the influence of truncation strategy, we computed the fooling rate of attacks in CNN models by varying truncation rates. Specifically, the proportion of retained activation increased by $10\\%$ for each step. That means UAP learned the feature information from shallow to deep convolution layers gradually. Due to space limitations, UAPs", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/54360931ab4da5e24809ad0f987e60634e820290f30d4fddee5b2fc4c6a85944.jpg", + "image_caption": [ + "(a) ResNet152", + "Figure 7. The ablation study of truncated negative activations on CNN models. Positive: the attack only uses the truncated positive activation. Positive&Negative: the attack integrates truncated positive and negative activations." + ], + "image_footnote": [], + "bbox": [ + 511, + 444, + 694, + 540 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a0325d7a358b602399802ea52d8d43043b08f33eb2521616b840fe45c56fe47.jpg", + "image_caption": [ + "(b) GoogleNet" + ], + "image_footnote": [], + "bbox": [ + 699, + 444, + 883, + 540 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "were only tested in white-box settings. The results of all CNN models were presented in Fig. 8, where the parameter space was explored w.r.t. NTR and PTR. As can be seen, the optimal result was obtained on the attack setting of the truncated activation rather than the whole activation. Especially in deep CNN models, only part of positive and negative activations were retained to craft UAPs. From the analysis of truncation in Fig. 8, we observe that the shallow convolution layer extracts generic features which are better at transferring the attack across different models.", + "bbox": [ + 496, + 638, + 890, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Exploring the Property of UAP", + "text_level": 1, + "bbox": [ + 500, + 800, + 774, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the following, the property of universal perturbations is thoroughly investigated. We studied the distinction between original example and adversarial examples made by image-specific attack (MI-FGSM) and data-free universal attacks (GD-UAP and TRM-UAP). As shown in Fig. 9, AEs", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4768", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/572da65a74803e16285449d42a3de05c833fced37a732be51e6527405adca038.jpg", + "image_caption": [ + "(a) AlexNet" + ], + "image_footnote": [], + "bbox": [ + 91, + 94, + 223, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f9584c2a5a9435aa9c1cfe179bf4704a78f252e1f5274770805964feb08086e0.jpg", + "image_caption": [ + "(b) VGG16" + ], + "image_footnote": [], + "bbox": [ + 256, + 94, + 387, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/87e3bfd800fe78c910306a2e98dce874140db9cd2cc92ade1b8e808d20971ab2.jpg", + "image_caption": [ + "(c) VGG19" + ], + "image_footnote": [], + "bbox": [ + 421, + 92, + 552, + 171 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b391c6653a3ce10e43f466c0221f525adc1ebc0ac56fce94b04cc371894a638f.jpg", + "image_caption": [ + "(d) ResNet152" + ], + "image_footnote": [], + "bbox": [ + 584, + 94, + 715, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/920e604745f5203df0ea6803dfda4ff9d12dedefe1c86835061fe8dbfe052133.jpg", + "image_caption": [ + "(e) GoogleNet" + ], + "image_footnote": [], + "bbox": [ + 750, + 94, + 879, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a4b5e47d7ded467512cf5dec0c5bfe14807c8f647eed16c01d4b176527ca6108.jpg", + "image_caption": [ + "Figure 8. The parameter study of truncation rates on CNN models in white-box attack setting. The $x$ , $y$ , and $z$ axis are NTR, PTR, and FR, respectively. Note that there are only five convolution layers in AlexNet, truncation results (e.g., $10\\%$ , $30\\%$ , $50\\%$ , etc.) are omitted since these values are consistent with the ones at their right coordinate." + ], + "image_footnote": [], + "bbox": [ + 84, + 257, + 171, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6e036f3f9eece8a241eb2c3235ead53938aeb250022b273360956be79176ce6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 258, + 267, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/891ea19d7f31bcdd2f5b5590e25f2f8fc377d0e4731f741e6638a74733302f89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 258, + 364, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/31dc80df69a5682a54422bbd9b097440f2f58e52df4477e33cb6207ce04b108e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 258, + 460, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/438c031f7e905448fb9b5d8ce6dd3a1b954c6bedb6ebdcb0f46d163df157fe45.jpg", + "image_caption": [ + "Figure 9. The distinction of feature maps among original image, image-specific attack, and universal attacks (top: feature maps extracted from a shallow convolution layer; bottom: original example and AEs made by VGG19)." + ], + "image_footnote": [], + "bbox": [ + 84, + 340, + 171, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5bc5ff70594099e09d7459d6747fc33c69384834846216e9a5078d68df1bbf0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 340, + 267, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bba6eb8b5f34e8f463ebbf59b2e27b1cf6ad621434c1670bf79a0dc52faf5d93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 340, + 364, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/34635c586aa6340dfa8f236ea295eec759a14ab0f58e24b2467881216340d752.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 340, + 460, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are indistinguishable compared with original example, and the imperceptible noise added to the input effectively perturbed feature maps in a shallow convolution layer. Furthermore, data-free universal attacks disturbed the feature map seriously in contrast to the image-specific attack. Nevertheless, the feature map of TRM-UAP had better regularity and repeatability than GD-UAP. We can conclude that features extracted by shallow convolution layers profoundly impact the final classification of CNN models and are beneficial for universal attack to craft the transferable perturbation.", + "bbox": [ + 75, + 491, + 467, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further verify the above viewpoints, we trained UAPs in completely opposite truncation strategies (i.e., retaining former $80\\%$ layers versus retaining latter $20\\%$ layers), and counted the distribution of class labels of AEs. The result in Fig. 10 (a) was consistent with the hypothesis [33] that the top-1 logits prediction of UAP dominates the prediction of AEs. Both labels were theater curtain. Contrarily, it was shown in Fig. 10 (b) that the phenomenon of top-1 label domination disappeared, where the UAP label was obelisk and the AE label was brass. Thus, the shallow convolution layer is more beneficial to transfer universal attack with generic features through a variety of models.", + "bbox": [ + 75, + 643, + 467, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 843, + 194, + 858 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed a novel data-free universal attack, called TRM-UAP, to reformulate the UAP gener-", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/01d53a6840e63a9d1b5bf12eaf66ab779888517bb3fbab19ae86c318d80a4959.jpg", + "image_caption": [ + "(a) UAP from low-level (former $80\\%$ ) convolution layers" + ], + "image_footnote": [], + "bbox": [ + 519, + 263, + 697, + 351 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c911b1a7b9b5b49dfa3f7a4e0c0296d9ed77dd75f12afc39f3dfbb359a82f5a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 263, + 879, + 351 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7d129a42526e3fd9d856f53c514a5d25820b3343a1d11f7aa656c92dbfe0e2c5.jpg", + "image_caption": [ + "(b) UAP from high-level (latter $20\\%$ ) convolution layers" + ], + "image_footnote": [], + "bbox": [ + 517, + 386, + 697, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/90d691a8ad3cbe797146d9aeffdfe746ba085acfa32aba42863f52dac3afeec5.jpg", + "image_caption": [ + "Figure 10. The analysis of the label distribution of UAPs and AEs (both made by VGG16). In each subfigure, the left is the top-10 logits value of UAP $(\\pmb{v})$ , and the right is the distribution of top-10 predicated labels of AEs $(\\pmb{v} + \\pmb{x})$ ." + ], + "image_footnote": [], + "bbox": [ + 699, + 387, + 874, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ation task as a truncated ratio optimization problem. Compared with previous methods, our TRM-UAP method integrates positive activation maximization with negative activation minimization. Towards further improving the transferability of UAPs, we propose a truncation strategy that computes positive and negative activation losses from low-level convolutions, as well as a curriculum optimization algorithm to fully mine the diversity of artificial images. Experimental results on the ImageNet dataset validate the better transferability of our TRM-UAP attack than other data-free UAP attacks on different CNN models. Additionally, UAPs learned from generic features in shallow convolution layers can dominate the model prediction and transfer the universal attack.", + "bbox": [ + 496, + 580, + 890, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work is partially supported by the Scientific Research Foundation of Chongqing University of Technology under Grant No.2022ZDZ026, Natural Science Foundation of Chongqing, China under Grant No.CSTB2022NSCQ-MSX0493, the Key project of Chongqing Technology Innovation and Application Development under Grant No.cstc2021jscx-dxwtBX0018.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4769", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 6:14410-14430, 2018.", + "[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009.", + "[3] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Šrndić, Pavel Laskov, Giorgio Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 387-402. Springer, 2013.", + "[4] Battista Biggio and Fabio Roli. Wild patterns: Ten years after the rise of adversarial machine learning. Pattern Recognition, 84:317-331, 2018.", + "[5] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy, pages 39-57, 2017.", + "[6] Ashutosh Chaubey, Nikhil Agrawal, Kavya Barnwal, Keerat K Guliani, and Pramod Mehta. Universal adversarial perturbations: A survey. arXiv preprint arXiv:2005.08087, 2020.", + "[7] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018.", + "[8] Aditya Ganeshan, Vivek BS, and R Venkatesh Babu. Fda: Feature disruptive attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8069-8079, 2019.", + "[9] Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015.", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016.", + "[11] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in Neural Information Processing Systems, 32, 2019.", + "[12] Valentin Khrulkov and Ivan Oseledets. Art of singular vectors and universal adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8562-8570, 2018.", + "[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 2012.", + "[14] Hong Liu, Rongrong Ji, Jie Li, Baochang Zhang, Yue Gao, Yongjian Wu, and Feiyue Huang. Universal adversarial perturbation via prior driven uncertainty approximation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2941-2949, 2019." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018.", + "[16] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1765-1773, 2017.", + "[17] Konda Reddy Mopuri, Aditya Ganeshan, and R Venkatesh Babu. Generalizable data-free objective for crafting universal adversarial perturbations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(10):2452-2465, 2018.", + "[18] Konda Reddy Mopuri, Utsav Garg, and R Venkatesh Babu. Fast feature fool: A data independent approach to universal adversarial perturbations. In Proceedings of the British Machine Vision Conference, 2017.", + "[19] Konda Reddy Mopuri, Utkarsh Ojha, Utsav Garg, and R Venkatesh Babu. Nag: Network for adversary generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 742-751, 2018.", + "[20] Konda Reddy Mopuri, Phani Krishna Uppala, and R Venkatesh Babu. Ask, acquire, and attack: Data-free uap generation using class impressions. In Proceedings of the European Conference on Computer Vision, pages 20-35, 2018.", + "[21] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Symposium on Security and Privacy, pages 372-387, 2016.", + "[22] Omid Poursaeed, Isay Katsman, Bicheng Gao, and Serge Belongie. Generative adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4422-4431, 2018.", + "[23] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211–252, 2015.", + "[24] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015.", + "[25] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015.", + "[26] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014.", + "[27] Yusuke Tashiro, Yang Song, and Stefano Ermon. Diversity can be transferred: Output diversification for white-and black-box attacks. Advances in Neural Information Processing Systems, 33:4536-4548, 2020." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4770", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In International Conference on Learning Representations, 2019.", + "[29] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7639-7648, 2021.", + "[30] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in Neural Information Processing Systems, 27, 2014.", + "[31] Yunrui Yu, Xitong Gao, and Cheng-Zhong Xu. Lafeat: piercing through adversarial defenses with latent features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5735-5745, 2021.", + "[32] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In So Kweon. Understanding adversarial examples from the mutual influence of images and perturbations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14521-14530, 2020.", + "[33] Chaoning Zhang, Philipp Benz, Adil Karjauv, and In So Kweon. Data-free universal adversarial perturbation and black-box attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7868-7877, 2021.", + "[34] Chaoning Zhang, Philipp Benz, Chenguo Lin, Adil Karjauv, Jing Wu, and In So Kweon. A survey on universal adversarial attack. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages 4687-4694, 2021.", + "[35] Jianping Zhang, Weibin Wu, Jen-tse Huang, Yizhan Huang, Wenxuan Wang, Yuxin Su, and Michael R Lyu. Improving adversarial transferability via neuron attribution-based attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14993-15002, 2022." + ], + "bbox": [ + 78, + 90, + 470, + 611 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4771", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/6ac05fcd-de6a-4657-a3f2-a897a887dcbb_model.json b/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/6ac05fcd-de6a-4657-a3f2-a897a887dcbb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..44d79914ad199c84cb6e85d07a0a299f4428e2b0 --- /dev/null +++ b/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/6ac05fcd-de6a-4657-a3f2-a897a887dcbb_model.json @@ -0,0 +1,2651 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.88, + 0.175 + ], + "angle": 0, + "content": "TRM-UAP: Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.204, + 0.816, + 0.257 + ], + "angle": 0, + "content": "Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, Di Ming* \nSchool of Computer Science and Engineering, Chongqing University of Technology \nChongqing, China" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.26, + 0.756, + 0.275 + ], + "angle": 0, + "content": "lyrl99804@qq.com, {xfeng, ylwang, yangwu, diming}@cqut.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.342, + 0.473, + 0.751 + ], + "angle": 0, + "content": "Aiming at crafting a single universal adversarial perturbation (UAP) to fool CNN models for various data samples, universal attack enables a more efficient and accurate evaluation for the robustness of CNN models. Early universal attacks craft UAPs depending on data priors. For more practical applications, the data-free universal attacks that make UAPs from random noises have aroused much attention recently. However, existing data-free UAP methods perturb all the CNN feature layers equally via the maximization of the CNN activation, leading to poor transferability. In this paper, we propose a novel data-free universal attack without depending on any real data samples through truncated ratio maximization, which we term as TRM-UAP. Specifically, different from the maximization of the positive activation in convolution layers, we propose to optimize the UAP generation from the ratio of positive and negative activations. To further enhance the transferability of universal attack, TRM-UAP not only performs the ratio maximization merely on low-level generic features via the truncation strategy, but also incorporates a curriculum optimization algorithm that can effectively learn the diversity of artificial images. Extensive experiments on the ImageNet dataset verify that TRM-UAP achieves a state-of-the-art average fooling rate and excellent transferability on different CNN models as compared to other data-free UAP methods. Code is available at https://github.com/RandolphCarter0/TRMUAP." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.778, + 0.21, + 0.794 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.803, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Early research [26] shows that tiny and imperceptible perturbations can seriously disturb the prediction results of deep neural networks (DNNs), especially for image recognition tasks [3]. Adversarial Examples (AEs), crafted by adding tiny perturbations deliberately to benign samples," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.311, + 0.892, + 0.371 + ], + "angle": 0, + "content": "are not only imperceptible in computer vision tasks but also prone to transfer among the DNN models. Therefore, AEs have been regarded as a serious threat to DNN models since the development of deep learning [1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.893, + 0.644 + ], + "angle": 0, + "content": "To explore the impact of AEs, many methods are proposed to craft the adversarial perturbation that is highly transferable to various DNN models (i.e., with a high fooling rate on other DNN models). However, the AEs generated by these works are explicitly designed for some specific samples and often fail to perturb other samples that are even from the same dataset. Different from the aforementioned image-specific attack, the universal attack that generates image-agnostic universal adversarial perturbation (UAP) was proposed in [16]. The UAP in a universal attack setting is trained from prior knowledge, such as substitute data, surrogate model, etc. By adding the UAP to benign samples, universal attacks can generate numerous AEs all together in a very short period. Furthermore, the universal attack can fool most DNN models that are trained from similar datasets, and can greatly reduce the computational cost of crafting AEs, making adversaries more applicable to real scenarios than the image-specific attacks [4, 6, 34]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Whereas, no matter the image-specific attacks or the universal attacks, well-annotated training data (e.g., [21, 9, 7]) or substitute data (e.g., [16, 12, 22, 19]) is required to generate AEs. In practice, obtaining a well-labeled large-scale dataset is challenging and costly, especially for some applications with critical security demand, where less prior knowledge is available. Recently, researchers have investigated data-free universal attack methods [18, 17, 14, 20, 33], where AEs are generated directly from random noises rather than data priors. Compared with previous methods that disturb the gradient or maximize the classification loss in data-dependent scenarios, current data-free UAP methods explore the feature-based perturbation, which tries to maximize the activation (e.g., ReLU) of convolutional neural network (CNN) features. It is shown that feature-based UAP methods achieved highly efficient and applicable universal attack without using any data priors [4, 6, 34]. How" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.277, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author: Di Ming" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4762" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "ever, current data-free UAP methods only consider the positive activation [18, 17] and perturb all layers of CNN features equally [20, 14]. As a result, UAPs crafted by surrogate models are hard to transfer to target models [6, 34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.473, + 0.56 + ], + "angle": 0, + "content": "Towards enhancing the transferability of data-free UAPs, we propose a novel data-free universal attack method called TRM-UAP, which formulates the UAP generation as a truncated ratio maximization problem. In particular, TRM-UAP attempts to over-fire positive neurons so that extracted features from multiple CNN layers can be fully disrupted. Besides the maximization of positive neurons, the proper activation on negative neurons may also be helpful. Hence, TRM-UAP performs a ratio maximization of positive and negative activations, which facilitates some negative neurons transiting to be positive. Moreover, some feature-based adversarial attacks [29, 35] reveal that multiple layers of CNN features are not equally important for disturbing the image classification [11, 28]. To further improve the transferability of universal attack, TRM-UAP performs ratio maximization only on parts of CNN activation layers. Specifically, a truncation strategy is proposed to truncate the activation of high-level convolutions and retain the activation of shallow convolutions to train perturbations. The intuition is that low-level CNN features are shown to provide more generic (data-independent) feature representations than high-level semantic features [30]. Thus the disturbed low-level features are expected to be able to transfer attacks across different CNN models, leading to highly transferable UAPs. Furthermore, a curriculum learning [2] based optimization algorithm is introduced to utilize artificial images and explore the diversity of input." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.562, + 0.47, + 0.59 + ], + "angle": 0, + "content": "In general, the main contributions of our work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.595, + 0.468, + 0.639 + ], + "angle": 0, + "content": "- We propose a novel data-free universal attack method to craft image-agnostic adversarial perturbations without utilizing any real data samples during training." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.642, + 0.468, + 0.821 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to formulate the data-free universal attack as a truncated ratio maximization problem. The proposed TRM-UAP enhances the transferability of UAPs from three perspectives: (i) the maximization on CNN features for crafting UAPs is enhanced by both maximization of positive activation and minimization of negative activation. (ii) the ratio maximization on truncated feature layers improves the generalizability of universal attacks to CNN models. (iii) artificial images are incorporated into curriculum optimization algorithm to strengthen the UAP dominance on feature activations." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.826, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- Extensive experiments on the ImageNet dataset show that TRM-UAP obtains the highest average fooling rate compared with other data-free UAP methods, indicating that truncated ratio maximization can greatly improve the transferability of universal attack." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.595, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Image-specific attacks: The attacks that try to craft perturbations for corresponding images by utilizing training or substitute data are termed as image-specific attacks. Typical image-specific attacks, such as Fast Gradient Sign Method (FGSM) [9], I-FGSM and MI-FGSM [7], Project Gradient Descent (PGD) [15], and Output Diversified Sampling (ODS) [27], etc., usually make use of the gradient information from well-trained target CNN model (i.e., white-box attack) or surrogate model (i.e., black-box attack) to disturb images. However, since gradient-based attacks often require a number of iterations, the computational cost of making AEs becomes extremely expensive as the number of examples increases. Recently, feature-based attacks that try to craft AEs from the perspective of CNN features have received broad concern [29, 8, 35]. For example, Neuron Attribution based Attack (NAA) [35] conducted feature-level attacks by deploying a neuron attribution method to estimate the importance of neurons at multiple CNN layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Data-dependent universal attacks: The universal attack aims at training a single perturbation that is independent of any specific sample, i.e., UAP. By adding a single UAP to various benign samples, numerous AEs can be generated very efficiently. Until recently, most of the universal attacks are still data-dependent, meaning that the perturbations are learned from data priors, such as training data, substitute data, etc. For example, the primary universal attack [16] tried to seek a universal perturbation for a set of training samples. On the other hand, there are some works [22, 19] applying generative adversarial models to craft perturbations. Recently, Zhang et al. [32] proposed to generate UAPs by training perturbations on the proxy datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Data-free universal attacks: The data-free universal attack makes UAP via random noise rather than prior knowledge from data directly. It is considered the most applicable attack for real applications in the adversarial attack field. However, since the information of both data and target models is unknown, the data-free universal attack is extremely challenging and only a few recent works focus on it. Mopuri et al. [18] first propose a data-free universal attack, i.e., the Fast Feature Fool (FFF) method, by maximizing the feature activations across all the CNN layers. In addition, they further propose a Generalizable Data-free UAP (GDUAP) to improve the FFF attack via optimizing the UAP training process with a saturation check strategy [17]. Besides, AAA [20] learns a generative model to make UAPs based on the pre-defined class impression. However, as the number of image categories increases, the cost of acquiring class impressions will also broaden. Prior-Driven Uncertainty Approximation (PD-UA) [14] proposes to craft UAPs by maximizing the uncertainty approximation of the model. Cosine-UAP [33] introduces the cosine similarity to train UAPs in a self-supervised way. Because of the shortage of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4763" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "prior knowledge, current data-free UAP methods generally present poor transferability across various CNN models." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.134, + 0.212, + 0.151 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.16, + 0.469, + 0.22 + ], + "angle": 0, + "content": "In this section, we first introduce the motivation of maximizing the ratio of activations. After that, we present the details of the proposed truncated ratio maximization method to craft the UAP in a strictly data-free fashion." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.229, + 0.206, + 0.244 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.252, + 0.469, + 0.297 + ], + "angle": 0, + "content": "To fool CNN models \\( f \\), UAP [16] tries to find the universal perturbation \\( \\pmb{v} \\) which maximizes the classification loss \\( \\mathcal{L} \\) (e.g., cross entropy) over the data distribution \\( \\mathbb{D} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.307, + 0.469, + 0.333 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\boldsymbol {v} \\sim \\mathbb {S}} \\underset {(\\boldsymbol {x}, \\boldsymbol {y}) \\sim \\mathbb {D}} {\\mathbb {E}} [ \\mathcal {L} (f (\\boldsymbol {v} + \\boldsymbol {x}), \\boldsymbol {y})) ] \\quad \\text {s . t .} \\| \\boldsymbol {v} \\| _ {p} \\leq \\epsilon \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.342, + 0.469, + 0.447 + ], + "angle": 0, + "content": "where \\(\\mathbb{S}\\) is the perturbation space, \\((\\pmb{x},\\pmb{y})\\) are the data and the class label (or one-hot ground truth) sampled from \\(\\mathbb{D}\\), \\(f(\\cdot)\\) is the output of CNN models, and \\(\\epsilon\\) is the constant to constrain the perturbation \\(\\pmb{v}\\) in the \\(\\ell_p\\)-norm bound (\\(e.g., p = 0,1,2,\\infty\\)). To be consistent with previous works [18, 17, 32], we adopt the \\(\\ell_{\\infty}\\)-norm to realize the imperceptibility of perturbation for all analyzes throughout this paper." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.469, + 0.537 + ], + "angle": 0, + "content": "However, the training data of the target model is usually unavailable and hard to access, making the data-dependence approaches not suitable for practical adversarial attacks. As a consequence, Mopuri et al. [18] propose the Fast Feature Fool (FFF) method to maximize the CNN activations in a strictly data-free fashion:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.547, + 0.469, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\boldsymbol {v}} \\quad | | \\mathcal {A} ^ {(i)} (\\boldsymbol {v}) | | _ {2}, \\quad \\text {f o r} i = 1, 2, \\dots , L \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.57, + 0.253, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\text {s . t .} \\quad | | \\boldsymbol {v} | | _ {\\infty} \\leq \\epsilon\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.594, + 0.469, + 0.701 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}^{(i)}(\\pmb{v}) = \\mathrm{Activation}(\\mathcal{C}^{(i)}(\\pmb{v}))\\) (e.g.,ReLU activation), \\(\\mathcal{C}^{(i)}(\\pmb{v})\\) is the output of the \\(i\\)-th convolution layer, and \\(L\\) is the total number of convolution layers. It can be seen that this typical data-free universal attack method aims at accumulating the errors gradually through multiple convolution layers to enlarge the classification loss, finally leading to the misclassification of perturbed samples \\(\\pmb{v} + \\pmb{x}\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.71, + 0.391, + 0.725 + ], + "angle": 0, + "content": "3.2. Maximizing the Ratio of Activations" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.733, + 0.469, + 0.793 + ], + "angle": 0, + "content": "Different from previous data-free universal methods [18, 17] that only maximize the activation of convolution layer (see Fig. 1 (a)), we propose to craft the universal adversarial perturbation \\(\\pmb{v}\\) via maximizing the ratio of activations:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.803, + 0.469, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\boldsymbol {v}} \\frac {| | \\mathcal {C} _ {+} ^ {(i)} (\\boldsymbol {v}) | | _ {2}}{| | \\mathcal {C} _ {-} ^ {(i)} (\\boldsymbol {v}) | | _ {2}}, \\quad \\text {f o r} i = 1, 2, \\dots , L \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.845, + 0.252, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\text {s . t .} \\quad | | \\boldsymbol {v} | | _ {\\infty} \\leq \\epsilon\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.869, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where positive activation and negative activation in \\(i\\)-th convolution layer are defined as \\(\\mathcal{C}_{+}^{(i)}(\\pmb{v}) = \\max (\\mathcal{C}^{(i)}(\\pmb{v}),0)\\) and" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.092, + 0.877, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.286, + 0.891, + 0.316 + ], + "angle": 0, + "content": "Figure 1. The illustration of previous positive maximization [18, 17] and our proposed ratio maximization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.334, + 0.89, + 0.366 + ], + "angle": 0, + "content": "\\(\\mathcal{C}_{-}^{(i)}(\\pmb {v}) = \\min (\\mathcal{C}^{(i)}(\\pmb {v}),0)\\) respectively, and their magnitudes are measured by \\(\\ell_2\\) -norm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.892, + 0.548 + ], + "angle": 0, + "content": "Towards further enlarging the fooling probability, this ratio formulation (3) tries to increase the magnitude and the number of positive activation via maximizing positive activation and minimizing negative activation simultaneously. Specifically speaking, ratio maximization is a dynamic process. Firstly, minimizing the negative activation could lead to the change of sign in some negative features whose values are close to zero, i.e., jumping from a tiny negative value to a tiny positive value, as shown in Fig. 1 (b). Afterward, at the next iteration, maximizing the positive activation will continue to increase the magnitude of above-mentioned features from a tiny positive value to a large positive value." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.67 + ], + "angle": 0, + "content": "During training, with the decrease in the number of negative activation and the increase in the number of positive activation, our proposed formulation (3) can achieve a higher ratio of positive activation as compared to previous data-free universal attacks, e.g., the formulation (2). Meanwhile, it also indicates that our ratio maximization could craft the UAP with more significant attack intensity so as to increase the fooling probability further." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.773, + 0.695 + ], + "angle": 0, + "content": "3.3. Truncated Ratio Maximization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Features extracted from shallow to deep convolution layers behave differently in crafting the adversarial perturbation. LAFEAT [31] analyzes the influence of each intermediate convolution layer on the misclassification, and attacks features only in a specific layer. On the other hand, GD-UAP [17] thoroughly studied the relative change in feature activations of adversarial examples as compared to original images at multiple CNN layers. However, GD-UAP attacks the feature activations extracted from all convolution layers together, without taking layer-wise feature patterns into consideration. According to [30], low-level features in shallow layers have generic patterns, while high-level features in deep layers have specific patterns. It can be seen that" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4764" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.092, + 0.411, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.272, + 0.468, + 0.303 + ], + "angle": 0, + "content": "Figure 2. Truncated positive and negative activation losses, where cumulative shaded areas are used to craft the UAP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.468, + 0.401 + ], + "angle": 0, + "content": "not all convolution layers contribute positively to increasing the positive activation and the objective loss. Therefore, based on ratio maximization defined in formulation (3), we propose a truncated ratio maximization method to further enhance the intensity and the transferability of UAPs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.402, + 0.468, + 0.497 + ], + "angle": 0, + "content": "For notational simplicity, we begin by defining the ratio maximization problem in \\(i\\)-th convolution layer as \\(\\mathcal{L}^{(i)}(\\pmb{v}) = \\mathcal{L}_{+}^{(i)}(\\pmb{v}) / \\mathcal{L}_{-}^{(i)}(\\pmb{v})\\), where \\(\\mathcal{L}_{+}^{(i)}(\\pmb{v}) = ||\\mathcal{C}_{+}^{(i)}(\\pmb{v})||_{2}\\) and \\(\\mathcal{L}_{-}^{(i)}(\\pmb{v}) = ||\\mathcal{C}_{-}^{(i)}(\\pmb{v})||_{2}\\). To maximize the ratio of activations in convolution layers altogether, the overall loss function of ratio maximization is reformulated as" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.502, + 0.469, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\boldsymbol {v}) = \\sum_ {i = 1} ^ {L} \\log \\mathcal {L} ^ {(i)} (\\boldsymbol {v}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.468, + 0.739 + ], + "angle": 0, + "content": "where log rescales the activation to prevent gradient explosions. Different from other methods that craft UAPs based on all convolution layers or a specific convolution layer, we propose to truncate the computation of positive and negative activations in deep convolution layers, since specific feature patterns are not beneficial for improving the quality of UAPs. Assuming that the truncation rate \\((TR)\\) is defined as \\(TR = \\lfloor (L - l) / L\\rfloor \\%\\) , we make truncation on \\((l + 1)\\) -th convolution layer by setting \\(\\mathcal{L}_{+}^{(i)}(\\pmb {v}) = \\tau\\) \\((i > l)\\) and \\(\\mathcal{L}_{-}^{(i)}(\\pmb {v}) = \\tau\\) \\((i > l)\\) , where \\(\\tau\\) is a small positive number such as \\(1e - 9\\) . As a result, \\(\\mathcal{L}(\\pmb {v})\\) can be equivalently rewritten as \\(\\mathcal{L}(\\pmb {v}) = \\sum_{i = 1}^{L}\\log \\mathcal{L}^{(i)}(\\pmb {v}) = \\sum_{i = 1}^{l}\\log \\mathcal{L}^{(i)}(\\pmb {v})\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.74, + 0.468, + 0.817 + ], + "angle": 0, + "content": "However, \\(\\mathcal{L}_{+}^{(i)}(\\pmb{v})\\) and \\(\\mathcal{L}_{-}^{(i)}(\\pmb{v})\\) not only have different scales but also behave differently in various layers for maximizing the ratio loss, which could lead to a degradation of the attack performance. To resolve these problems, we first reformulate the ratio loss in \\(i\\)-th convolution layer as" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.824, + 0.469, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\alpha} ^ {(i)} (\\boldsymbol {v}) = \\frac {\\mathcal {L} _ {+} ^ {(i)} (\\boldsymbol {v})}{(\\mathcal {L} _ {-} ^ {(i)} (\\boldsymbol {v})) ^ {\\alpha}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a scaling hyperparameter to adjust the relative importance between positive and negative activation losses." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.093, + 0.852, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.303 + ], + "angle": 0, + "content": "Figure 3. The mean positive activation loss in various convolution layers with different truncation settings." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.328, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Secondly, positive and negative truncation rates are defined as \\( PTR = \\lfloor (L - l') / L \\rfloor \\% \\) and \\( NTR = \\lfloor (L - l'') / L \\rfloor \\% \\) respectively. That is to say, we make truncation on different layers for positive and negative activation losses by setting \\( \\mathcal{L}_{+}^{(i)}(\\pmb{v}) = \\tau (i > l') \\) and \\( \\mathcal{L}_{-}^{(i)}(\\pmb{v}) = \\tau (i > l'') \\), see Fig. 2. Finally, the overall loss function \\( \\mathcal{L}(\\pmb{v}) \\) can be reduced to" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.431, + 0.89, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\boldsymbol {v}) = \\sum_ {i = 1} ^ {L} \\log \\mathcal {L} _ {\\alpha} ^ {(i)} (\\boldsymbol {v}) \\\\ = \\sum_ {i = 1} ^ {l ^ {\\prime}} \\log \\mathcal {L} _ {+} ^ {(i)} (\\boldsymbol {v}) - \\alpha \\cdot \\sum_ {i = 1} ^ {l ^ {\\prime \\prime}} \\log \\mathcal {L} _ {-} ^ {(i)} (\\boldsymbol {v}) + c \\\\ \\propto \\sum_ {i = 1} ^ {l ^ {\\prime}} \\log \\mathcal {L} _ {+} ^ {(i)} (\\boldsymbol {v}) - \\alpha \\cdot \\sum_ {i = 1} ^ {l ^ {\\prime \\prime}} \\log \\mathcal {L} _ {-} ^ {(i)} (\\boldsymbol {v}), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.518, + 0.89, + 0.534 + ], + "angle": 0, + "content": "where \\(c = ((1 - \\alpha)\\cdot L - l^{\\prime} + \\alpha \\cdot l^{\\prime \\prime})\\cdot \\log \\tau\\) , i.e., a constant." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.534, + 0.89, + 0.58 + ], + "angle": 0, + "content": "Consequently, for data-free universal adversarial attack, we propose a truncated ratio maximization method (TRMUAP) to craft the perturbation \\(\\pmb{v}\\) satisfying" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.589, + 0.89, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max _ {\\boldsymbol {v}} \\sum_ {i = 1} ^ {l ^ {\\prime}} \\log \\left| \\left| \\mathcal {C} _ {+} ^ {(i)} (\\boldsymbol {v}) \\right| \\right| _ {2} - \\alpha \\cdot \\sum_ {i = 1} ^ {l ^ {\\prime \\prime}} \\log \\left| \\left| \\mathcal {C} _ {-} ^ {(i)} (\\boldsymbol {v}) \\right| \\right| _ {2} \\tag {7} \\\\ \\text {s . t .} \\quad | | \\boldsymbol {v} | | _ {\\infty} \\leq \\epsilon \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.657, + 0.892, + 0.779 + ], + "angle": 0, + "content": "where \\(\\pmb{v}\\) is constrained by \\(\\ell_{\\infty}\\)-norm and bound \\(\\epsilon\\). To verify the previous claim, we compute the mean positive activation loss in various convolution layers via feeding UAPs into CNN models (e.g., VGG19). As shown in Fig. 3, TRM-UAP \\((PTR \\neq NTR\\), i.e., \\(l' \\neq l'' < L)\\) generates a larger cumulative positive activation loss than RM-UAP \\((l' = l'' = L)\\) and TRM-UAP \\((PTR = NTR\\), i.e., \\(l' = l'' < L)\\), which could lead to a further increase in fooling rates." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.818, + 0.804 + ], + "angle": 0, + "content": "3.4. Curriculum Optimization Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For purpose of improving the diversity of inputs, we utilize Gaussian noise and jigsaw image [33] to generate artificial images, which are fed into surrogate models together with the perturbation. To resemble training samples in real-world datasets, the mean filter is used to smooth the edge between different regions in jigsaw images. However, it" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4765" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "will be difficult for the optimization algorithm to converge if artificial images with complex patterns are provided at the beginning of the training procedure. Inspired by curriculum learning [2], we start the training with simple artificial images and then gradually feed more complex artificial images into surrogate models. Specifically, the generation of artificial images is controlled by the distribution parameter and the training iteration, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.22, + 0.348, + 0.236 + ], + "angle": 0, + "content": "\\[\nD _ {1} \\prec D _ {2} \\prec \\dots \\prec D _ {n},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.236, + 0.468, + 0.255 + ], + "angle": 0, + "content": "\\[\nD _ {t} = \\left\\{\\boldsymbol {x} \\mid \\boldsymbol {x} \\sim P \\left(\\theta_ {0}, t\\right) \\right\\}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.399 + ], + "angle": 0, + "content": "where \\(D_{t}\\) is a set of artificial images \\(\\pmb{x}\\) sampled at \\(t\\)-th iteration \\((t = 1,2,\\dots ,n)\\), \\(P\\) is the distribution of artificial samples with the default parameter \\(\\theta_0\\), and \\(D_{a} \\prec D_{b}\\) denotes that the pattern of artificial images in \\(D_{b}\\) is more complex than \\(D_{a}\\). In detail, as the number of training iterations \\(t\\) increases, the distribution parameter \\(\\theta_0\\) (e.g., the standard deviation of Gaussian distribution and the frequency of jigsaw images) is gradually increasing. For \\(t\\)-th iteration, our curriculum optimization algorithm is maximizing" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.407, + 0.468, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = \\frac {1}{| D _ {t} |} \\sum_ {\\boldsymbol {x} \\in D _ {t}} \\left(\\sum_ {i = 1} ^ {l ^ {\\prime}} \\log \\mathcal {L} _ {+} ^ {(i)} (\\boldsymbol {v} + \\boldsymbol {x}) - \\alpha \\cdot \\sum_ {i = 1} ^ {l ^ {\\prime \\prime}} \\log \\mathcal {L} _ {-} ^ {(i)} (\\boldsymbol {v} + \\boldsymbol {x})\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.465, + 0.478 + ], + "angle": 0, + "content": "where \\( |D_t| \\) is the number of artificial images in the set \\( D_t \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.473, + 0.705 + ], + "angle": 0, + "content": "The entire procedure of our proposed TRM-UAP method is summarized in Algorithm 1. During the training process, the perturbation vector \\(\\pmb{v}\\) is randomly initialized by the uniform distribution \\(\\mathcal{U}\\), and a surrogate dataset is used to validate the convergence of optimization for every \\(H\\) iterations. \\(l^{\\prime}\\) and \\(l''\\) represent the retained layers of positive and negative activations corresponding to truncation rates \\(PTR\\) and \\(NTR\\) respectively. The scaling hyperparameter \\(\\alpha\\) adjusts the ratio between truncated positive and negative activation losses. The algorithm is converged only if either the iteration number reaches the maximum \\(T\\) or the fooling rate test reaches the threshold \\(F_{max}\\). Similar to GD-UAP [17], we run a saturation rate test for each iteration to dynamically compress the perturbation vector only if the saturation rate \\(\\hat{r}\\) is smaller than the predefined threshold \\(r\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.466, + 0.727 + ], + "angle": 0, + "content": "3.5. Connection to Other Data-Free UAP Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "When positive and negative truncation rates are set as \\( PTR = 0\\% \\) and \\( NTR = 100\\% \\) (i.e., \\( l' = L, l'' = 0 \\)), this means all the negative activations are truncated and all the positive activations are retained. Thus, the proposed TRM-UAP defined in Eq. (7) is reduced to GD-UAP [17]. It can be seen that GD-UAP is only a special case of TRM-UAP under the data-free universal attack setting. As compared to GD-UAP, our TRM-UAP improves both intensity and transferability of crafted UAPs via truncated ratio maximization. The detail of their differences will be further discussed in the experiments." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "Algorithm 1 Curriculum optimization algorithm for solving truncated ratio maximization problem to craft universal adversarial perturbations" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.138, + 0.894, + 0.213 + ], + "angle": 0, + "content": "Input: surrogate CNN model \\( f \\), limitation value \\( \\epsilon \\), learning rate \\( \\eta \\), positive and negative truncation rates \\( PTR \\), \\( NTR \\), scaling hyperparameter \\( \\alpha \\), maximum iteration number \\( T \\), convergence threshold \\( F_{max} \\), validation test hyperparameter \\( H \\), saturation threshold \\( r \\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.215, + 0.799, + 0.229 + ], + "angle": 0, + "content": "Output: universal adversarial perturbation \\(\\pmb{v}\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.232, + 0.786, + 0.246 + ], + "angle": 0, + "content": "1: Initialize \\(\\pmb{v}_0\\sim \\mathcal{U}(-\\epsilon ,\\epsilon)\\) \\(t = 0,F = 0\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.248, + 0.742, + 0.262 + ], + "angle": 0, + "content": "2: while \\( t < T \\) and \\( F < F_{\\text{max}} \\) do" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.264, + 0.61, + 0.275 + ], + "angle": 0, + "content": "3: \\( t = t + 1 \\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.277, + 0.853, + 0.291 + ], + "angle": 0, + "content": "4: Generate the artificial image set \\( D_{t} \\) via Eq. (8)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.293, + 0.882, + 0.307 + ], + "angle": 0, + "content": "5: Compute the gradient \\(\\nabla \\mathcal{L}_t\\) of the loss \\(\\mathcal{L}_t\\) in Eq. (9)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.308, + 0.739, + 0.322 + ], + "angle": 0, + "content": "6: Update \\(\\pmb{v}_t = \\pmb{v}_{t-1} + \\eta \\cdot \\nabla \\mathcal{L}_t\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.324, + 0.756, + 0.338 + ], + "angle": 0, + "content": "7: Clip \\(\\pmb{v}_t = \\min (\\epsilon ,\\max (\\pmb {v}_t, - \\epsilon))\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.339, + 0.886, + 0.352 + ], + "angle": 0, + "content": "8: Compute the saturation rate \\(\\hat{r}\\) and adjust \\(\\pmb{v}_t\\) if \\(\\hat{r} < r\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.353, + 0.892, + 0.382 + ], + "angle": 0, + "content": "9: Conduct the fooling rate test \\( FR \\) if \\( t\\% H == 0 \\) and \\( F = F + 1 \\) if \\( FR \\) not the best fooling rate" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.232, + 0.892, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.384, + 0.601, + 0.396 + ], + "angle": 0, + "content": "10: end while" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.399, + 0.597, + 0.412 + ], + "angle": 0, + "content": "11: return \\( \\mathbf{v}_t \\)" + }, + { + "type": "list", + "bbox": [ + 0.505, + 0.384, + 0.601, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.445, + 0.633, + 0.463 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.564 + ], + "angle": 0, + "content": "Setup: By following the experiment setup in existing data-free universal methods [18, 17, 20, 14], we evaluated the proposed TRM-UAP method on the validation set of ImageNet [23] with the classical pre-trained CNN models including AlexNet [13], VGG16 [24], VGG19 [24], ResNet152 [10] and GoogleNet [25]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Evaluation Criteria: We used the fooling rate \\((FR)\\) proposed by data-free universal methods [18, 17] as the evaluation metric. Particularly, a higher fooling rate represented higher attack success rate and better transferability. In addition, we used the logit loss of C&W attack [5] to further evaluate the transferability of data-free UAPs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Comparative Methods: We made a comparison with data-free universal approaches, including FFF [18], AAA [20], GD-UAP [17], PD-UA [14] and Cosine-UAP [33]. Note that GD-UAP [17] trained models under different setups, for a fair comparison, we reproduced the results of GD-UAP [17] by making the setup consistent with TRM-UAP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Implementation Details: All of our experiments were implemented on PyTorch with a single NVIDIA GeForce RTX 3090Ti GPU. Following the common setting [18, 17], we set \\(\\epsilon = 10 / 255\\) to restrict changes of the perturbation. The maximum iteration \\(T\\) was set as 10000, and the saturation threshold \\(r\\) was set to \\(0.001\\%\\). The value range of hyperparameters \\(PTR, NTR \\in [0, 1]\\) denotes the truncation rate of positive and negative activations respectively. We set the scaling hyperparameter \\(\\alpha\\) and truncation rates with appropriate values for different CNN models." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4766" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.133, + 0.089, + 0.835, + 0.21 + ], + "angle": 0, + "content": "
AttackAlexNetVGG16VGG19ResNet152GoogleNetAverage
FFF80.9247.1043.62-56.44-
AAA89.0471.5972.8460.7275.2873.89
GD-UAP85.2490.0187.3445.9645.8764.65
PD-UA-70.6964.9846.3967.12-
Cosine-UAP91.0789.4886.8165.3587.5784.08
TRM-UAP(Ours)93.53±0.0794.30±0.1591.35±0.3067.46±0.3585.32±0.0486.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.213, + 0.895, + 0.245 + ], + "angle": 0, + "content": "Table 1. The fooling rate \\((FR)\\) of our proposed TRM-UAP and other data-free universal methods. We show the mean and standard deviation of \\(FR\\) with five runs." + }, + { + "type": "table", + "bbox": [ + 0.192, + 0.257, + 0.778, + 0.372 + ], + "angle": 0, + "content": "
AlexNetVGG16VGG19ResNet152GoogleNet
AlexNet93.53±0.0760.10±0.2457.08±0.1527.31±0.3032.70±0.22
VGG1647.53±0.5194.30±0.1289.68±0.1461.43±0.4053.95±0.59
VGG1946.01±0.4489.82±0.1591.35±0.3047.19±0.6646.48±0.78
ResNet15253.56±0.7577.20±0.3573.30±0.4167.46±0.3557.54±0.50
GoogleNet60.10±1.1679.66±0.9579.98±1.0658.85±1.9485.32±0.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.374, + 0.895, + 0.437 + ], + "angle": 0, + "content": "Table 2. The transferability of the perturbation crafted by our TRM-UAP method (rows: target models; columns: surrogate models). The fooling rate \\((FR)\\) on the diagonal represents white-box attacks (i.e., UAP is crafted by the target model itself), and other off-diagonal values are black-box attacks (i.e., UAP is crafted by non-target models.). All results of mean and standard deviation are computed on average with five turns." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.461, + 0.218, + 0.475 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.473, + 0.777 + ], + "angle": 0, + "content": "Our proposed TRM-UAP is applied to craft UAPs on five CNN models and create AEs to attack these CNN models separately on the ImageNet validation set. We set \\(\\alpha \\in \\{1.0, 1.0, 0.5, 1.5, 1.3\\}\\), \\(PTR \\in \\{0.0, 0.2, 0.2, 0.7, 0.4\\}\\), \\(NTR \\in \\{0.8, 0.3, 0.2, 0.8, 0.3\\}\\), corresponding to AlexNet, VGG16, VGG19, ResNet152, GoogleNet. The attack performance compared with other data-free universal methods is shown in Table 1. Note that the results of GD-UAP were reproduced by our best effort on PyTorch for a fair comparison. As compared to other methods, our TRM-UAP method improves the fooling rate in most CNN models. Although Cosine-UAP achieves a higher \\(FR\\) in GoogleNet, TRM-UAP improves the \\(FR\\) around \\(1 \\sim 5\\%\\) on other four models and further improves the average \\(FR\\) of all five models more than \\(2\\%\\). Besides, it is shown in Fig. 4 that the visualization of crafted UAPs has clearly particular image patterns with abundant features. We also visualized AEs of TRM-UAP attack in Fig. 6 and found that high probabilities of AEs were given to the labels of incorrect class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.904 + ], + "angle": 0, + "content": "We further used the UAP made by surrogate models to attack other target models to verify the transferability. Table 2 represented the attack performance in different attack settings. The models in rows denote target models, and the models in columns denote surrogate models for crafting the UAP. The results on the diagonal denote the white-box attack, while the rest are transferable attacks in the black-box setting. Most TRM-UAP attacks perform excellently in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.895, + 0.507 + ], + "angle": 0, + "content": "transferability, which indicates that the universal perturbations crafted by TRM-UAP are generalized well to transfer attacks on target models." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.514, + 0.771, + 0.531 + ], + "angle": 0, + "content": "4.2. Evaluating the Transferability" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.81 + ], + "angle": 0, + "content": "To further compare the effectiveness of universal attacks, we used the logit loss of C&W attack [5] to evaluate the transferability of GD-UAP, Cosine-UAP, and TRM-UAP. The logit loss is defined as \\( loss = (\\max_{j \\neq t} F(\\boldsymbol{v} + \\boldsymbol{x})_j) - F(\\boldsymbol{v} + \\boldsymbol{x})_t \\), where \\( F(\\cdot)_j \\) is the \\( j \\)-th output in the logits layer of network and \\( t \\) is the index of ground truth. A larger logit loss value represents better transferability of attacks to CNN models. The logit loss of original images is calculated as a baseline. Note that the results of GD-UAP and Cosine-UAP are reproduced with our best effort on PyTorch. As shown in Fig. 5, TRM-UAP achieves a higher logit loss than GD-UAP and Cosine-UAP in most cases. Additionally, the discrepancy among GD-UAP, Cosine-UAP, and TRM-UAP becomes significant in deep CNN models (e.g., ResNet152 and GoogleNet). Therefore, TRM-UAP achieves relatively higher logit losses in both white-box and black-box attack settings, which indicates a better transferability across various CNN models." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.818, + 0.671, + 0.835 + ], + "angle": 0, + "content": "4.3. Parameter Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "First, the influence of ratio maximization on attack performance is studied. Here we mainly focused on validating the effectiveness of truncated negative activation in deep CNN models (e.g., ResNet152 and GoogleNet). The experi" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "4767" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.225, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.199, + 0.187, + 0.21 + ], + "angle": 0, + "content": "(a) AlexNet" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.091, + 0.39, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.198, + 0.352, + 0.21 + ], + "angle": 0, + "content": "(b) VGG16" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.091, + 0.554, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.198, + 0.516, + 0.21 + ], + "angle": 0, + "content": "(c) VGG19" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.091, + 0.719, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.198, + 0.691, + 0.21 + ], + "angle": 0, + "content": "(d) ResNet152" + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.09, + 0.885, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.198, + 0.855, + 0.21 + ], + "angle": 0, + "content": "(e) GoogleNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.214, + 0.892, + 0.245 + ], + "angle": 0, + "content": "Figure 4. The visualization of UAPs crafted by our TRM-UAP method. For the best visualization in color, we enhance the value of the pixel to [0, 255]." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.263, + 0.242, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.34, + 0.195, + 0.352 + ], + "angle": 0, + "content": "(a) AlexNet" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.263, + 0.406, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.34, + 0.356, + 0.352 + ], + "angle": 0, + "content": "(b) VGG16" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.263, + 0.571, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.34, + 0.519, + 0.352 + ], + "angle": 0, + "content": "(c) VGG19" + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.263, + 0.73, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.34, + 0.69, + 0.352 + ], + "angle": 0, + "content": "(d) ResNet152" + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.263, + 0.888, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.77, + 0.34, + 0.848, + 0.352 + ], + "angle": 0, + "content": "(e) GoogleNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.355, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Figure 5. The logit loss of original samples and adversarial samples crafted by GD-UAP, Cosine-UAP, our TRM-UAP methods. In each sub-figure, the subtitle indicates the surrogate model for crafting UAPs, and target models are shown on the horizontal axis. Original samples are classified correctly by the target model and exploited to craft AEs for comparisons. For each attack setting, we compute the average of the logit loss over all the samples." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.433, + 0.175, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.5, + 0.174, + 0.506 + ], + "angle": 0, + "content": "bullet train(98.15%)" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.434, + 0.269, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.5, + 0.257, + 0.506 + ], + "angle": 0, + "content": "puffer(99.89%)" + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.506, + 0.175, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.57, + 0.16, + 0.577 + ], + "angle": 0, + "content": "quilt(45.63%)" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.506, + 0.272, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.571, + 0.274, + 0.577 + ], + "angle": 0, + "content": "theater curtain(72.82%)" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.434, + 0.362, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.5, + 0.361, + 0.506 + ], + "angle": 0, + "content": "monastery(34.69%)" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.506, + 0.362, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.571, + 0.352, + 0.577 + ], + "angle": 0, + "content": "church(43.55%)" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.434, + 0.455, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.5, + 0.439, + 0.506 + ], + "angle": 0, + "content": "shoji(87.62%)" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.506, + 0.458, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.571, + 0.462, + 0.577 + ], + "angle": 0, + "content": "shower curtain(94.35%)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.581, + 0.469, + 0.657 + ], + "angle": 0, + "content": "Figure 6. Examples of our TRM-UAP attack tested on VGG16 and the perturbation crafted by VGG19 (top: original examples; bottom: adversarial examples). The annotation represents the classification label and the probability predicted by CNNs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.469, + 0.81 + ], + "angle": 0, + "content": "imental setting is the same as Section 4.1. Based on this, we also crafted UAPs by truncated positive activation only, i.e., \\(PTR\\) remains unchanged and \\(NTR = 100\\%\\). As shown in Fig. 7, the ablation experiment indicates that truncated negative activation combined with truncated positive activation improves the \\(FR\\) by a large margin (around \\(8 \\sim 18\\%\\)) as compared to standalone truncated positive activation, thus enhancing both intensity and transferability of UAPs significantly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To explore the influence of truncation strategy, we computed the fooling rate of attacks in CNN models by varying truncation rates. Specifically, the proportion of retained activation increased by \\(10\\%\\) for each step. That means UAP learned the feature information from shallow to deep convolution layers gradually. Due to space limitations, UAPs" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.445, + 0.695, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.544, + 0.642, + 0.556 + ], + "angle": 0, + "content": "(a) ResNet152" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.445, + 0.884, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.544, + 0.829, + 0.557 + ], + "angle": 0, + "content": "(b) GoogleNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.56, + 0.892, + 0.621 + ], + "angle": 0, + "content": "Figure 7. The ablation study of truncated negative activations on CNN models. Positive: the attack only uses the truncated positive activation. Positive&Negative: the attack integrates truncated positive and negative activations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.79 + ], + "angle": 0, + "content": "were only tested in white-box settings. The results of all CNN models were presented in Fig. 8, where the parameter space was explored w.r.t. NTR and PTR. As can be seen, the optimal result was obtained on the attack setting of the truncated activation rather than the whole activation. Especially in deep CNN models, only part of positive and negative activations were retained to craft UAPs. From the analysis of truncation in Fig. 8, we observe that the shallow convolution layer extracts generic features which are better at transferring the attack across different models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.776, + 0.818 + ], + "angle": 0, + "content": "4.4. Exploring the Property of UAP" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the following, the property of universal perturbations is thoroughly investigated. We studied the distinction between original example and adversarial examples made by image-specific attack (MI-FGSM) and data-free universal attacks (GD-UAP and TRM-UAP). As shown in Fig. 9, AEs" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4768" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.095, + 0.225, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.181, + 0.188, + 0.193 + ], + "angle": 0, + "content": "(a) AlexNet" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.095, + 0.388, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.181, + 0.352, + 0.193 + ], + "angle": 0, + "content": "(b) VGG16" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.093, + 0.553, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.181, + 0.516, + 0.193 + ], + "angle": 0, + "content": "(c) VGG19" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.095, + 0.716, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.181, + 0.689, + 0.193 + ], + "angle": 0, + "content": "(d) ResNet152" + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.095, + 0.88, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.181, + 0.853, + 0.193 + ], + "angle": 0, + "content": "(e) GoogleNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.197, + 0.894, + 0.242 + ], + "angle": 0, + "content": "Figure 8. The parameter study of truncation rates on CNN models in white-box attack setting. The \\( x \\), \\( y \\), and \\( z \\) axis are NTR, PTR, and FR, respectively. Note that there are only five convolution layers in AlexNet, truncation results (e.g., \\( 10\\% \\), \\( 30\\% \\), \\( 50\\% \\), etc.) are omitted since these values are consistent with the ones at their right coordinate." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.258, + 0.173, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.259, + 0.269, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.259, + 0.365, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.259, + 0.462, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.342, + 0.173, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.342, + 0.268, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.342, + 0.365, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.342, + 0.461, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.411, + 0.468, + 0.471 + ], + "angle": 0, + "content": "Figure 9. The distinction of feature maps among original image, image-specific attack, and universal attacks (top: feature maps extracted from a shallow convolution layer; bottom: original example and AEs made by VGG19)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.468, + 0.643 + ], + "angle": 0, + "content": "are indistinguishable compared with original example, and the imperceptible noise added to the input effectively perturbed feature maps in a shallow convolution layer. Furthermore, data-free universal attacks disturbed the feature map seriously in contrast to the image-specific attack. Nevertheless, the feature map of TRM-UAP had better regularity and repeatability than GD-UAP. We can conclude that features extracted by shallow convolution layers profoundly impact the final classification of CNN models and are beneficial for universal attack to craft the transferable perturbation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.468, + 0.826 + ], + "angle": 0, + "content": "To further verify the above viewpoints, we trained UAPs in completely opposite truncation strategies (i.e., retaining former \\(80\\%\\) layers versus retaining latter \\(20\\%\\) layers), and counted the distribution of class labels of AEs. The result in Fig. 10 (a) was consistent with the hypothesis [33] that the top-1 logits prediction of UAP dominates the prediction of AEs. Both labels were theater curtain. Contrarily, it was shown in Fig. 10 (b) that the phenomenon of top-1 label domination disappeared, where the UAP label was obelisk and the AE label was brass. Thus, the shallow convolution layer is more beneficial to transfer universal attack with generic features through a variety of models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.844, + 0.195, + 0.859 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In this paper, we have proposed a novel data-free universal attack, called TRM-UAP, to reformulate the UAP gener-" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.264, + 0.698, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.264, + 0.88, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.361, + 0.86, + 0.373 + ], + "angle": 0, + "content": "(a) UAP from low-level (former \\(80\\%\\) ) convolution layers" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.387, + 0.699, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.388, + 0.875, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.48, + 0.845, + 0.492 + ], + "angle": 0, + "content": "(b) UAP from high-level (latter \\(20\\%\\) ) convolution layers" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.495, + 0.892, + 0.556 + ], + "angle": 0, + "content": "Figure 10. The analysis of the label distribution of UAPs and AEs (both made by VGG16). In each subfigure, the left is the top-10 logits value of UAP \\((\\pmb{v})\\), and the right is the distribution of top-10 predicated labels of AEs \\((\\pmb{v} + \\pmb{x})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.791 + ], + "angle": 0, + "content": "ation task as a truncated ratio optimization problem. Compared with previous methods, our TRM-UAP method integrates positive activation maximization with negative activation minimization. Towards further improving the transferability of UAPs, we propose a truncation strategy that computes positive and negative activation losses from low-level convolutions, as well as a curriculum optimization algorithm to fully mine the diversity of artificial images. Experimental results on the ImageNet dataset validate the better transferability of our TRM-UAP attack than other data-free UAP attacks on different CNN models. Additionally, UAPs learned from generic features in shallow convolution layers can dominate the model prediction and transfer the universal attack." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work is partially supported by the Scientific Research Foundation of Chongqing University of Technology under Grant No.2022ZDZ026, Natural Science Foundation of Chongqing, China under Grant No.CSTB2022NSCQ-MSX0493, the Key project of Chongqing Technology Innovation and Application Development under Grant No.cstc2021jscx-dxwtBX0018." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4769" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 6:14410-14430, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.471, + 0.212 + ], + "angle": 0, + "content": "[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.214, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Šrndić, Pavel Laskov, Giorgio Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 387-402. Springer, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.339 + ], + "angle": 0, + "content": "[4] Battista Biggio and Fabio Roli. Wild patterns: Ten years after the rise of adversarial machine learning. Pattern Recognition, 84:317-331, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.471, + 0.382 + ], + "angle": 0, + "content": "[5] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy, pages 39-57, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.471, + 0.436 + ], + "angle": 0, + "content": "[6] Ashutosh Chaubey, Nikhil Agrawal, Kavya Barnwal, Keerat K Guliani, and Pramod Mehta. Universal adversarial perturbations: A survey. arXiv preprint arXiv:2005.08087, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[7] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[8] Aditya Ganeshan, Vivek BS, and R Venkatesh Babu. Fda: Feature disruptive attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8069-8079, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[9] Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[11] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in Neural Information Processing Systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[12] Valentin Khrulkov and Ivan Oseledets. Art of singular vectors and universal adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8562-8570, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Hong Liu, Rongrong Ji, Jie Li, Baochang Zhang, Yue Gao, Yongjian Wu, and Feiyue Huang. Universal adversarial perturbation via prior driven uncertainty approximation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2941-2949, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[15] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[16] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1765-1773, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[17] Konda Reddy Mopuri, Aditya Ganeshan, and R Venkatesh Babu. Generalizable data-free objective for crafting universal adversarial perturbations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(10):2452-2465, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[18] Konda Reddy Mopuri, Utsav Garg, and R Venkatesh Babu. Fast feature fool: A data independent approach to universal adversarial perturbations. In Proceedings of the British Machine Vision Conference, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[19] Konda Reddy Mopuri, Utkarsh Ojha, Utsav Garg, and R Venkatesh Babu. Nag: Network for adversary generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 742-751, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[20] Konda Reddy Mopuri, Phani Krishna Uppala, and R Venkatesh Babu. Ask, acquire, and attack: Data-free uap generation using class impressions. In Proceedings of the European Conference on Computer Vision, pages 20-35, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[21] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Symposium on Security and Privacy, pages 372-387, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[22] Omid Poursaeed, Isay Katsman, Bicheng Gao, and Serge Belongie. Generative adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4422-4431, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[23] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211–252, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[24] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[25] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Yusuke Tashiro, Yang Song, and Stefano Ermon. Diversity can be transferred: Output diversification for white-and black-box attacks. Advances in Neural Information Processing Systems, 33:4536-4548, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4770" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.217 + ], + "angle": 0, + "content": "[29] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7639-7648, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[30] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in Neural Information Processing Systems, 27, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.318 + ], + "angle": 0, + "content": "[31] Yunrui Yu, Xitong Gao, and Cheng-Zhong Xu. Lafeat: piercing through adversarial defenses with latent features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5735-5745, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[32] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In So Kweon. Understanding adversarial examples from the mutual influence of images and perturbations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14521-14530, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[33] Chaoning Zhang, Philipp Benz, Adil Karjauv, and In So Kweon. Data-free universal adversarial perturbation and black-box attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7868-7877, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.47, + 0.528 + ], + "angle": 0, + "content": "[34] Chaoning Zhang, Philipp Benz, Chenguo Lin, Adil Karjauv, Jing Wu, and In So Kweon. A survey on universal adversarial attack. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages 4687-4694, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.47, + 0.612 + ], + "angle": 0, + "content": "[35] Jianping Zhang, Weibin Wu, Jen-tse Huang, Yizhan Huang, Wenxuan Wang, Yuxin Su, and Michael R Lyu. Improving adversarial transferability via neuron attribution-based attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14993-15002, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "4771" + } + ] +] \ No newline at end of file diff --git a/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/6ac05fcd-de6a-4657-a3f2-a897a887dcbb_origin.pdf b/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/6ac05fcd-de6a-4657-a3f2-a897a887dcbb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a4e9a842ff17537f232ef686aadb68c8ab7fcd6c --- /dev/null +++ b/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/6ac05fcd-de6a-4657-a3f2-a897a887dcbb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28a19447fec1632b084b64eb117436671e38a8ee5e9232f74346f32f6dd43ee8 +size 4849537 diff --git a/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/full.md b/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8002763e68d94539742023952b430ca4a042e674 --- /dev/null +++ b/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/full.md @@ -0,0 +1,385 @@ +# TRM-UAP: Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization + +Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, Di Ming* +School of Computer Science and Engineering, Chongqing University of Technology +Chongqing, China + +lyrl99804@qq.com, {xfeng, ylwang, yangwu, diming}@cqut.edu.cn + +# Abstract + +Aiming at crafting a single universal adversarial perturbation (UAP) to fool CNN models for various data samples, universal attack enables a more efficient and accurate evaluation for the robustness of CNN models. Early universal attacks craft UAPs depending on data priors. For more practical applications, the data-free universal attacks that make UAPs from random noises have aroused much attention recently. However, existing data-free UAP methods perturb all the CNN feature layers equally via the maximization of the CNN activation, leading to poor transferability. In this paper, we propose a novel data-free universal attack without depending on any real data samples through truncated ratio maximization, which we term as TRM-UAP. Specifically, different from the maximization of the positive activation in convolution layers, we propose to optimize the UAP generation from the ratio of positive and negative activations. To further enhance the transferability of universal attack, TRM-UAP not only performs the ratio maximization merely on low-level generic features via the truncation strategy, but also incorporates a curriculum optimization algorithm that can effectively learn the diversity of artificial images. Extensive experiments on the ImageNet dataset verify that TRM-UAP achieves a state-of-the-art average fooling rate and excellent transferability on different CNN models as compared to other data-free UAP methods. Code is available at https://github.com/RandolphCarter0/TRMUAP. + +# 1. Introduction + +Early research [26] shows that tiny and imperceptible perturbations can seriously disturb the prediction results of deep neural networks (DNNs), especially for image recognition tasks [3]. Adversarial Examples (AEs), crafted by adding tiny perturbations deliberately to benign samples, + +are not only imperceptible in computer vision tasks but also prone to transfer among the DNN models. Therefore, AEs have been regarded as a serious threat to DNN models since the development of deep learning [1]. + +To explore the impact of AEs, many methods are proposed to craft the adversarial perturbation that is highly transferable to various DNN models (i.e., with a high fooling rate on other DNN models). However, the AEs generated by these works are explicitly designed for some specific samples and often fail to perturb other samples that are even from the same dataset. Different from the aforementioned image-specific attack, the universal attack that generates image-agnostic universal adversarial perturbation (UAP) was proposed in [16]. The UAP in a universal attack setting is trained from prior knowledge, such as substitute data, surrogate model, etc. By adding the UAP to benign samples, universal attacks can generate numerous AEs all together in a very short period. Furthermore, the universal attack can fool most DNN models that are trained from similar datasets, and can greatly reduce the computational cost of crafting AEs, making adversaries more applicable to real scenarios than the image-specific attacks [4, 6, 34]. + +Whereas, no matter the image-specific attacks or the universal attacks, well-annotated training data (e.g., [21, 9, 7]) or substitute data (e.g., [16, 12, 22, 19]) is required to generate AEs. In practice, obtaining a well-labeled large-scale dataset is challenging and costly, especially for some applications with critical security demand, where less prior knowledge is available. Recently, researchers have investigated data-free universal attack methods [18, 17, 14, 20, 33], where AEs are generated directly from random noises rather than data priors. Compared with previous methods that disturb the gradient or maximize the classification loss in data-dependent scenarios, current data-free UAP methods explore the feature-based perturbation, which tries to maximize the activation (e.g., ReLU) of convolutional neural network (CNN) features. It is shown that feature-based UAP methods achieved highly efficient and applicable universal attack without using any data priors [4, 6, 34]. How + +ever, current data-free UAP methods only consider the positive activation [18, 17] and perturb all layers of CNN features equally [20, 14]. As a result, UAPs crafted by surrogate models are hard to transfer to target models [6, 34]. + +Towards enhancing the transferability of data-free UAPs, we propose a novel data-free universal attack method called TRM-UAP, which formulates the UAP generation as a truncated ratio maximization problem. In particular, TRM-UAP attempts to over-fire positive neurons so that extracted features from multiple CNN layers can be fully disrupted. Besides the maximization of positive neurons, the proper activation on negative neurons may also be helpful. Hence, TRM-UAP performs a ratio maximization of positive and negative activations, which facilitates some negative neurons transiting to be positive. Moreover, some feature-based adversarial attacks [29, 35] reveal that multiple layers of CNN features are not equally important for disturbing the image classification [11, 28]. To further improve the transferability of universal attack, TRM-UAP performs ratio maximization only on parts of CNN activation layers. Specifically, a truncation strategy is proposed to truncate the activation of high-level convolutions and retain the activation of shallow convolutions to train perturbations. The intuition is that low-level CNN features are shown to provide more generic (data-independent) feature representations than high-level semantic features [30]. Thus the disturbed low-level features are expected to be able to transfer attacks across different CNN models, leading to highly transferable UAPs. Furthermore, a curriculum learning [2] based optimization algorithm is introduced to utilize artificial images and explore the diversity of input. + +In general, the main contributions of our work can be summarized as follows: + +- We propose a novel data-free universal attack method to craft image-agnostic adversarial perturbations without utilizing any real data samples during training. +- To the best of our knowledge, we are the first to formulate the data-free universal attack as a truncated ratio maximization problem. The proposed TRM-UAP enhances the transferability of UAPs from three perspectives: (i) the maximization on CNN features for crafting UAPs is enhanced by both maximization of positive activation and minimization of negative activation. (ii) the ratio maximization on truncated feature layers improves the generalizability of universal attacks to CNN models. (iii) artificial images are incorporated into curriculum optimization algorithm to strengthen the UAP dominance on feature activations. +- Extensive experiments on the ImageNet dataset show that TRM-UAP obtains the highest average fooling rate compared with other data-free UAP methods, indicating that truncated ratio maximization can greatly improve the transferability of universal attack. + +# 2. Related Work + +Image-specific attacks: The attacks that try to craft perturbations for corresponding images by utilizing training or substitute data are termed as image-specific attacks. Typical image-specific attacks, such as Fast Gradient Sign Method (FGSM) [9], I-FGSM and MI-FGSM [7], Project Gradient Descent (PGD) [15], and Output Diversified Sampling (ODS) [27], etc., usually make use of the gradient information from well-trained target CNN model (i.e., white-box attack) or surrogate model (i.e., black-box attack) to disturb images. However, since gradient-based attacks often require a number of iterations, the computational cost of making AEs becomes extremely expensive as the number of examples increases. Recently, feature-based attacks that try to craft AEs from the perspective of CNN features have received broad concern [29, 8, 35]. For example, Neuron Attribution based Attack (NAA) [35] conducted feature-level attacks by deploying a neuron attribution method to estimate the importance of neurons at multiple CNN layers. + +Data-dependent universal attacks: The universal attack aims at training a single perturbation that is independent of any specific sample, i.e., UAP. By adding a single UAP to various benign samples, numerous AEs can be generated very efficiently. Until recently, most of the universal attacks are still data-dependent, meaning that the perturbations are learned from data priors, such as training data, substitute data, etc. For example, the primary universal attack [16] tried to seek a universal perturbation for a set of training samples. On the other hand, there are some works [22, 19] applying generative adversarial models to craft perturbations. Recently, Zhang et al. [32] proposed to generate UAPs by training perturbations on the proxy datasets. + +Data-free universal attacks: The data-free universal attack makes UAP via random noise rather than prior knowledge from data directly. It is considered the most applicable attack for real applications in the adversarial attack field. However, since the information of both data and target models is unknown, the data-free universal attack is extremely challenging and only a few recent works focus on it. Mopuri et al. [18] first propose a data-free universal attack, i.e., the Fast Feature Fool (FFF) method, by maximizing the feature activations across all the CNN layers. In addition, they further propose a Generalizable Data-free UAP (GDUAP) to improve the FFF attack via optimizing the UAP training process with a saturation check strategy [17]. Besides, AAA [20] learns a generative model to make UAPs based on the pre-defined class impression. However, as the number of image categories increases, the cost of acquiring class impressions will also broaden. Prior-Driven Uncertainty Approximation (PD-UA) [14] proposes to craft UAPs by maximizing the uncertainty approximation of the model. Cosine-UAP [33] introduces the cosine similarity to train UAPs in a self-supervised way. Because of the shortage of + +prior knowledge, current data-free UAP methods generally present poor transferability across various CNN models. + +# 3. Methodology + +In this section, we first introduce the motivation of maximizing the ratio of activations. After that, we present the details of the proposed truncated ratio maximization method to craft the UAP in a strictly data-free fashion. + +# 3.1. Preliminary + +To fool CNN models $f$ , UAP [16] tries to find the universal perturbation $\pmb{v}$ which maximizes the classification loss $\mathcal{L}$ (e.g., cross entropy) over the data distribution $\mathbb{D}$ : + +$$ +\max _ {\boldsymbol {v} \sim \mathbb {S}} \underset {(\boldsymbol {x}, \boldsymbol {y}) \sim \mathbb {D}} {\mathbb {E}} [ \mathcal {L} (f (\boldsymbol {v} + \boldsymbol {x}), \boldsymbol {y})) ] \quad \text {s . t .} \| \boldsymbol {v} \| _ {p} \leq \epsilon \tag {1} +$$ + +where $\mathbb{S}$ is the perturbation space, $(\pmb{x},\pmb{y})$ are the data and the class label (or one-hot ground truth) sampled from $\mathbb{D}$ , $f(\cdot)$ is the output of CNN models, and $\epsilon$ is the constant to constrain the perturbation $\pmb{v}$ in the $\ell_p$ -norm bound ( $e.g., p = 0,1,2,\infty$ ). To be consistent with previous works [18, 17, 32], we adopt the $\ell_{\infty}$ -norm to realize the imperceptibility of perturbation for all analyzes throughout this paper. + +However, the training data of the target model is usually unavailable and hard to access, making the data-dependence approaches not suitable for practical adversarial attacks. As a consequence, Mopuri et al. [18] propose the Fast Feature Fool (FFF) method to maximize the CNN activations in a strictly data-free fashion: + +$$ +\max _ {\boldsymbol {v}} \quad | | \mathcal {A} ^ {(i)} (\boldsymbol {v}) | | _ {2}, \quad \text {f o r} i = 1, 2, \dots , L \tag {2} +$$ + +$$ +\text {s . t .} \quad | | \boldsymbol {v} | | _ {\infty} \leq \epsilon +$$ + +where $\mathcal{A}^{(i)}(\pmb{v}) = \mathrm{Activation}(\mathcal{C}^{(i)}(\pmb{v}))$ (e.g.,ReLU activation), $\mathcal{C}^{(i)}(\pmb{v})$ is the output of the $i$ -th convolution layer, and $L$ is the total number of convolution layers. It can be seen that this typical data-free universal attack method aims at accumulating the errors gradually through multiple convolution layers to enlarge the classification loss, finally leading to the misclassification of perturbed samples $\pmb{v} + \pmb{x}$ . + +# 3.2. Maximizing the Ratio of Activations + +Different from previous data-free universal methods [18, 17] that only maximize the activation of convolution layer (see Fig. 1 (a)), we propose to craft the universal adversarial perturbation $\pmb{v}$ via maximizing the ratio of activations: + +$$ +\max _ {\boldsymbol {v}} \frac {| | \mathcal {C} _ {+} ^ {(i)} (\boldsymbol {v}) | | _ {2}}{| | \mathcal {C} _ {-} ^ {(i)} (\boldsymbol {v}) | | _ {2}}, \quad \text {f o r} i = 1, 2, \dots , L \tag {3} +$$ + +$$ +\text {s . t .} \quad | | \boldsymbol {v} | | _ {\infty} \leq \epsilon +$$ + +where positive activation and negative activation in $i$ -th convolution layer are defined as $\mathcal{C}_{+}^{(i)}(\pmb{v}) = \max (\mathcal{C}^{(i)}(\pmb{v}),0)$ and + +![](images/cd4db687608d88f9c2802209819768be2d60e16af792fcbe04bb63b9ff227737.jpg) +Figure 1. The illustration of previous positive maximization [18, 17] and our proposed ratio maximization. + +$\mathcal{C}_{-}^{(i)}(\pmb {v}) = \min (\mathcal{C}^{(i)}(\pmb {v}),0)$ respectively, and their magnitudes are measured by $\ell_2$ -norm. + +Towards further enlarging the fooling probability, this ratio formulation (3) tries to increase the magnitude and the number of positive activation via maximizing positive activation and minimizing negative activation simultaneously. Specifically speaking, ratio maximization is a dynamic process. Firstly, minimizing the negative activation could lead to the change of sign in some negative features whose values are close to zero, i.e., jumping from a tiny negative value to a tiny positive value, as shown in Fig. 1 (b). Afterward, at the next iteration, maximizing the positive activation will continue to increase the magnitude of above-mentioned features from a tiny positive value to a large positive value. + +During training, with the decrease in the number of negative activation and the increase in the number of positive activation, our proposed formulation (3) can achieve a higher ratio of positive activation as compared to previous data-free universal attacks, e.g., the formulation (2). Meanwhile, it also indicates that our ratio maximization could craft the UAP with more significant attack intensity so as to increase the fooling probability further. + +# 3.3. Truncated Ratio Maximization + +Features extracted from shallow to deep convolution layers behave differently in crafting the adversarial perturbation. LAFEAT [31] analyzes the influence of each intermediate convolution layer on the misclassification, and attacks features only in a specific layer. On the other hand, GD-UAP [17] thoroughly studied the relative change in feature activations of adversarial examples as compared to original images at multiple CNN layers. However, GD-UAP attacks the feature activations extracted from all convolution layers together, without taking layer-wise feature patterns into consideration. According to [30], low-level features in shallow layers have generic patterns, while high-level features in deep layers have specific patterns. It can be seen that + +![](images/3d806e9afd92dc1567563c9adf926db09e929a69e828aaa62d22d4334f4e8c45.jpg) +Figure 2. Truncated positive and negative activation losses, where cumulative shaded areas are used to craft the UAP. + +not all convolution layers contribute positively to increasing the positive activation and the objective loss. Therefore, based on ratio maximization defined in formulation (3), we propose a truncated ratio maximization method to further enhance the intensity and the transferability of UAPs. + +For notational simplicity, we begin by defining the ratio maximization problem in $i$ -th convolution layer as $\mathcal{L}^{(i)}(\pmb{v}) = \mathcal{L}_{+}^{(i)}(\pmb{v}) / \mathcal{L}_{-}^{(i)}(\pmb{v})$ , where $\mathcal{L}_{+}^{(i)}(\pmb{v}) = ||\mathcal{C}_{+}^{(i)}(\pmb{v})||_{2}$ and $\mathcal{L}_{-}^{(i)}(\pmb{v}) = ||\mathcal{C}_{-}^{(i)}(\pmb{v})||_{2}$ . To maximize the ratio of activations in convolution layers altogether, the overall loss function of ratio maximization is reformulated as + +$$ +\mathcal {L} (\boldsymbol {v}) = \sum_ {i = 1} ^ {L} \log \mathcal {L} ^ {(i)} (\boldsymbol {v}) \tag {4} +$$ + +where log rescales the activation to prevent gradient explosions. Different from other methods that craft UAPs based on all convolution layers or a specific convolution layer, we propose to truncate the computation of positive and negative activations in deep convolution layers, since specific feature patterns are not beneficial for improving the quality of UAPs. Assuming that the truncation rate $(TR)$ is defined as $TR = \lfloor (L - l) / L\rfloor \%$ , we make truncation on $(l + 1)$ -th convolution layer by setting $\mathcal{L}_{+}^{(i)}(\pmb {v}) = \tau$ $(i > l)$ and $\mathcal{L}_{-}^{(i)}(\pmb {v}) = \tau$ $(i > l)$ , where $\tau$ is a small positive number such as $1e - 9$ . As a result, $\mathcal{L}(\pmb {v})$ can be equivalently rewritten as $\mathcal{L}(\pmb {v}) = \sum_{i = 1}^{L}\log \mathcal{L}^{(i)}(\pmb {v}) = \sum_{i = 1}^{l}\log \mathcal{L}^{(i)}(\pmb {v})$ + +However, $\mathcal{L}_{+}^{(i)}(\pmb{v})$ and $\mathcal{L}_{-}^{(i)}(\pmb{v})$ not only have different scales but also behave differently in various layers for maximizing the ratio loss, which could lead to a degradation of the attack performance. To resolve these problems, we first reformulate the ratio loss in $i$ -th convolution layer as + +$$ +\mathcal {L} _ {\alpha} ^ {(i)} (\boldsymbol {v}) = \frac {\mathcal {L} _ {+} ^ {(i)} (\boldsymbol {v})}{(\mathcal {L} _ {-} ^ {(i)} (\boldsymbol {v})) ^ {\alpha}}, \tag {5} +$$ + +where $\alpha$ is a scaling hyperparameter to adjust the relative importance between positive and negative activation losses. + +![](images/cd1b9880d0864a22bc1b11b2a6e89447b521da93c5cfcf74790729c2e5b06279.jpg) +Figure 3. The mean positive activation loss in various convolution layers with different truncation settings. + +Secondly, positive and negative truncation rates are defined as $PTR = \lfloor (L - l') / L \rfloor \%$ and $NTR = \lfloor (L - l'') / L \rfloor \%$ respectively. That is to say, we make truncation on different layers for positive and negative activation losses by setting $\mathcal{L}_{+}^{(i)}(\pmb{v}) = \tau (i > l')$ and $\mathcal{L}_{-}^{(i)}(\pmb{v}) = \tau (i > l'')$ , see Fig. 2. Finally, the overall loss function $\mathcal{L}(\pmb{v})$ can be reduced to + +$$ +\begin{array}{l} \mathcal {L} (\boldsymbol {v}) = \sum_ {i = 1} ^ {L} \log \mathcal {L} _ {\alpha} ^ {(i)} (\boldsymbol {v}) \\ = \sum_ {i = 1} ^ {l ^ {\prime}} \log \mathcal {L} _ {+} ^ {(i)} (\boldsymbol {v}) - \alpha \cdot \sum_ {i = 1} ^ {l ^ {\prime \prime}} \log \mathcal {L} _ {-} ^ {(i)} (\boldsymbol {v}) + c \\ \propto \sum_ {i = 1} ^ {l ^ {\prime}} \log \mathcal {L} _ {+} ^ {(i)} (\boldsymbol {v}) - \alpha \cdot \sum_ {i = 1} ^ {l ^ {\prime \prime}} \log \mathcal {L} _ {-} ^ {(i)} (\boldsymbol {v}), \tag {6} \\ \end{array} +$$ + +where $c = ((1 - \alpha)\cdot L - l^{\prime} + \alpha \cdot l^{\prime \prime})\cdot \log \tau$ , i.e., a constant. + +Consequently, for data-free universal adversarial attack, we propose a truncated ratio maximization method (TRMUAP) to craft the perturbation $\pmb{v}$ satisfying + +$$ +\begin{array}{l} \max _ {\boldsymbol {v}} \sum_ {i = 1} ^ {l ^ {\prime}} \log \left| \left| \mathcal {C} _ {+} ^ {(i)} (\boldsymbol {v}) \right| \right| _ {2} - \alpha \cdot \sum_ {i = 1} ^ {l ^ {\prime \prime}} \log \left| \left| \mathcal {C} _ {-} ^ {(i)} (\boldsymbol {v}) \right| \right| _ {2} \tag {7} \\ \text {s . t .} \quad | | \boldsymbol {v} | | _ {\infty} \leq \epsilon \\ \end{array} +$$ + +where $\pmb{v}$ is constrained by $\ell_{\infty}$ -norm and bound $\epsilon$ . To verify the previous claim, we compute the mean positive activation loss in various convolution layers via feeding UAPs into CNN models (e.g., VGG19). As shown in Fig. 3, TRM-UAP $(PTR \neq NTR$ , i.e., $l' \neq l'' < L)$ generates a larger cumulative positive activation loss than RM-UAP $(l' = l'' = L)$ and TRM-UAP $(PTR = NTR$ , i.e., $l' = l'' < L)$ , which could lead to a further increase in fooling rates. + +# 3.4. Curriculum Optimization Algorithm + +For purpose of improving the diversity of inputs, we utilize Gaussian noise and jigsaw image [33] to generate artificial images, which are fed into surrogate models together with the perturbation. To resemble training samples in real-world datasets, the mean filter is used to smooth the edge between different regions in jigsaw images. However, it + +will be difficult for the optimization algorithm to converge if artificial images with complex patterns are provided at the beginning of the training procedure. Inspired by curriculum learning [2], we start the training with simple artificial images and then gradually feed more complex artificial images into surrogate models. Specifically, the generation of artificial images is controlled by the distribution parameter and the training iteration, defined as: + +$$ +D _ {1} \prec D _ {2} \prec \dots \prec D _ {n}, +$$ + +$$ +D _ {t} = \left\{\boldsymbol {x} \mid \boldsymbol {x} \sim P \left(\theta_ {0}, t\right) \right\}, \tag {8} +$$ + +where $D_{t}$ is a set of artificial images $\pmb{x}$ sampled at $t$ -th iteration $(t = 1,2,\dots ,n)$ , $P$ is the distribution of artificial samples with the default parameter $\theta_0$ , and $D_{a} \prec D_{b}$ denotes that the pattern of artificial images in $D_{b}$ is more complex than $D_{a}$ . In detail, as the number of training iterations $t$ increases, the distribution parameter $\theta_0$ (e.g., the standard deviation of Gaussian distribution and the frequency of jigsaw images) is gradually increasing. For $t$ -th iteration, our curriculum optimization algorithm is maximizing + +$$ +\mathcal {L} _ {t} = \frac {1}{| D _ {t} |} \sum_ {\boldsymbol {x} \in D _ {t}} \left(\sum_ {i = 1} ^ {l ^ {\prime}} \log \mathcal {L} _ {+} ^ {(i)} (\boldsymbol {v} + \boldsymbol {x}) - \alpha \cdot \sum_ {i = 1} ^ {l ^ {\prime \prime}} \log \mathcal {L} _ {-} ^ {(i)} (\boldsymbol {v} + \boldsymbol {x})\right) \tag {9} +$$ + +where $|D_t|$ is the number of artificial images in the set $D_t$ . + +The entire procedure of our proposed TRM-UAP method is summarized in Algorithm 1. During the training process, the perturbation vector $\pmb{v}$ is randomly initialized by the uniform distribution $\mathcal{U}$ , and a surrogate dataset is used to validate the convergence of optimization for every $H$ iterations. $l^{\prime}$ and $l''$ represent the retained layers of positive and negative activations corresponding to truncation rates $PTR$ and $NTR$ respectively. The scaling hyperparameter $\alpha$ adjusts the ratio between truncated positive and negative activation losses. The algorithm is converged only if either the iteration number reaches the maximum $T$ or the fooling rate test reaches the threshold $F_{max}$ . Similar to GD-UAP [17], we run a saturation rate test for each iteration to dynamically compress the perturbation vector only if the saturation rate $\hat{r}$ is smaller than the predefined threshold $r$ . + +# 3.5. Connection to Other Data-Free UAP Methods + +When positive and negative truncation rates are set as $PTR = 0\%$ and $NTR = 100\%$ (i.e., $l' = L, l'' = 0$ ), this means all the negative activations are truncated and all the positive activations are retained. Thus, the proposed TRM-UAP defined in Eq. (7) is reduced to GD-UAP [17]. It can be seen that GD-UAP is only a special case of TRM-UAP under the data-free universal attack setting. As compared to GD-UAP, our TRM-UAP improves both intensity and transferability of crafted UAPs via truncated ratio maximization. The detail of their differences will be further discussed in the experiments. + +Algorithm 1 Curriculum optimization algorithm for solving truncated ratio maximization problem to craft universal adversarial perturbations + +Input: surrogate CNN model $f$ , limitation value $\epsilon$ , learning rate $\eta$ , positive and negative truncation rates $PTR$ , $NTR$ , scaling hyperparameter $\alpha$ , maximum iteration number $T$ , convergence threshold $F_{max}$ , validation test hyperparameter $H$ , saturation threshold $r$ + +Output: universal adversarial perturbation $\pmb{v}$ + +1: Initialize $\pmb{v}_0\sim \mathcal{U}(-\epsilon ,\epsilon)$ $t = 0,F = 0$ +2: while $t < T$ and $F < F_{\text{max}}$ do +3: $t = t + 1$ +4: Generate the artificial image set $D_{t}$ via Eq. (8) +5: Compute the gradient $\nabla \mathcal{L}_t$ of the loss $\mathcal{L}_t$ in Eq. (9) +6: Update $\pmb{v}_t = \pmb{v}_{t-1} + \eta \cdot \nabla \mathcal{L}_t$ +7: Clip $\pmb{v}_t = \min (\epsilon ,\max (\pmb {v}_t, - \epsilon))$ +8: Compute the saturation rate $\hat{r}$ and adjust $\pmb{v}_t$ if $\hat{r} < r$ +9: Conduct the fooling rate test $FR$ if $t\% H == 0$ and $F = F + 1$ if $FR$ not the best fooling rate + +10: end while +11: return $\mathbf{v}_t$ + +# 4. Experiments + +Setup: By following the experiment setup in existing data-free universal methods [18, 17, 20, 14], we evaluated the proposed TRM-UAP method on the validation set of ImageNet [23] with the classical pre-trained CNN models including AlexNet [13], VGG16 [24], VGG19 [24], ResNet152 [10] and GoogleNet [25]. + +Evaluation Criteria: We used the fooling rate $(FR)$ proposed by data-free universal methods [18, 17] as the evaluation metric. Particularly, a higher fooling rate represented higher attack success rate and better transferability. In addition, we used the logit loss of C&W attack [5] to further evaluate the transferability of data-free UAPs. + +Comparative Methods: We made a comparison with data-free universal approaches, including FFF [18], AAA [20], GD-UAP [17], PD-UA [14] and Cosine-UAP [33]. Note that GD-UAP [17] trained models under different setups, for a fair comparison, we reproduced the results of GD-UAP [17] by making the setup consistent with TRM-UAP. + +Implementation Details: All of our experiments were implemented on PyTorch with a single NVIDIA GeForce RTX 3090Ti GPU. Following the common setting [18, 17], we set $\epsilon = 10 / 255$ to restrict changes of the perturbation. The maximum iteration $T$ was set as 10000, and the saturation threshold $r$ was set to $0.001\%$ . The value range of hyperparameters $PTR, NTR \in [0, 1]$ denotes the truncation rate of positive and negative activations respectively. We set the scaling hyperparameter $\alpha$ and truncation rates with appropriate values for different CNN models. + +
AttackAlexNetVGG16VGG19ResNet152GoogleNetAverage
FFF80.9247.1043.62-56.44-
AAA89.0471.5972.8460.7275.2873.89
GD-UAP85.2490.0187.3445.9645.8764.65
PD-UA-70.6964.9846.3967.12-
Cosine-UAP91.0789.4886.8165.3587.5784.08
TRM-UAP(Ours)93.53±0.0794.30±0.1591.35±0.3067.46±0.3585.32±0.0486.39
+ +Table 1. The fooling rate $(FR)$ of our proposed TRM-UAP and other data-free universal methods. We show the mean and standard deviation of $FR$ with five runs. + +
AlexNetVGG16VGG19ResNet152GoogleNet
AlexNet93.53±0.0760.10±0.2457.08±0.1527.31±0.3032.70±0.22
VGG1647.53±0.5194.30±0.1289.68±0.1461.43±0.4053.95±0.59
VGG1946.01±0.4489.82±0.1591.35±0.3047.19±0.6646.48±0.78
ResNet15253.56±0.7577.20±0.3573.30±0.4167.46±0.3557.54±0.50
GoogleNet60.10±1.1679.66±0.9579.98±1.0658.85±1.9485.32±0.04
+ +Table 2. The transferability of the perturbation crafted by our TRM-UAP method (rows: target models; columns: surrogate models). The fooling rate $(FR)$ on the diagonal represents white-box attacks (i.e., UAP is crafted by the target model itself), and other off-diagonal values are black-box attacks (i.e., UAP is crafted by non-target models.). All results of mean and standard deviation are computed on average with five turns. + +# 4.1. Main Results + +Our proposed TRM-UAP is applied to craft UAPs on five CNN models and create AEs to attack these CNN models separately on the ImageNet validation set. We set $\alpha \in \{1.0, 1.0, 0.5, 1.5, 1.3\}$ , $PTR \in \{0.0, 0.2, 0.2, 0.7, 0.4\}$ , $NTR \in \{0.8, 0.3, 0.2, 0.8, 0.3\}$ , corresponding to AlexNet, VGG16, VGG19, ResNet152, GoogleNet. The attack performance compared with other data-free universal methods is shown in Table 1. Note that the results of GD-UAP were reproduced by our best effort on PyTorch for a fair comparison. As compared to other methods, our TRM-UAP method improves the fooling rate in most CNN models. Although Cosine-UAP achieves a higher $FR$ in GoogleNet, TRM-UAP improves the $FR$ around $1 \sim 5\%$ on other four models and further improves the average $FR$ of all five models more than $2\%$ . Besides, it is shown in Fig. 4 that the visualization of crafted UAPs has clearly particular image patterns with abundant features. We also visualized AEs of TRM-UAP attack in Fig. 6 and found that high probabilities of AEs were given to the labels of incorrect class. + +We further used the UAP made by surrogate models to attack other target models to verify the transferability. Table 2 represented the attack performance in different attack settings. The models in rows denote target models, and the models in columns denote surrogate models for crafting the UAP. The results on the diagonal denote the white-box attack, while the rest are transferable attacks in the black-box setting. Most TRM-UAP attacks perform excellently in + +transferability, which indicates that the universal perturbations crafted by TRM-UAP are generalized well to transfer attacks on target models. + +# 4.2. Evaluating the Transferability + +To further compare the effectiveness of universal attacks, we used the logit loss of C&W attack [5] to evaluate the transferability of GD-UAP, Cosine-UAP, and TRM-UAP. The logit loss is defined as $loss = (\max_{j \neq t} F(\boldsymbol{v} + \boldsymbol{x})_j) - F(\boldsymbol{v} + \boldsymbol{x})_t$ , where $F(\cdot)_j$ is the $j$ -th output in the logits layer of network and $t$ is the index of ground truth. A larger logit loss value represents better transferability of attacks to CNN models. The logit loss of original images is calculated as a baseline. Note that the results of GD-UAP and Cosine-UAP are reproduced with our best effort on PyTorch. As shown in Fig. 5, TRM-UAP achieves a higher logit loss than GD-UAP and Cosine-UAP in most cases. Additionally, the discrepancy among GD-UAP, Cosine-UAP, and TRM-UAP becomes significant in deep CNN models (e.g., ResNet152 and GoogleNet). Therefore, TRM-UAP achieves relatively higher logit losses in both white-box and black-box attack settings, which indicates a better transferability across various CNN models. + +# 4.3. Parameter Study + +First, the influence of ratio maximization on attack performance is studied. Here we mainly focused on validating the effectiveness of truncated negative activation in deep CNN models (e.g., ResNet152 and GoogleNet). The experi + +![](images/65c9d802a59f4dcb81e3bb2f629100341f6cd3fcdc0b714697823fb1fac9f3fb.jpg) +(a) AlexNet + +![](images/8b3d3196e40adb2a8c67e651bcb9c67e61ae87efae8b7629b9d7eebc61232f8a.jpg) +(b) VGG16 + +![](images/4d0f65dd6835d1346f5582a0f4930288b9eedb01e701bb03ed2c2121bcc69f9b.jpg) +(c) VGG19 + +![](images/1c573be2aaf3325f72df9cf636b44987fe6db48ffb6c2e2550e3c8b93bebc8bc.jpg) +(d) ResNet152 + +![](images/07d9ff150d5f58d2a1fba9577c128d092d352195bb49c0301f60596993b9fa6a.jpg) +(e) GoogleNet + +![](images/d9840ca6ae6a1cec36052939a5a34a5c39b25a980b35c684409de51b5abd6c74.jpg) +Figure 4. The visualization of UAPs crafted by our TRM-UAP method. For the best visualization in color, we enhance the value of the pixel to [0, 255]. +(a) AlexNet + +![](images/653907a0c232758cb095a21ffe5995ca6c5497a2b86518feee531c915091ac34.jpg) +(b) VGG16 + +![](images/0f231656c84bd49878ea6b3134009c957cfc20c7b313d705c8a8718447b12cda.jpg) +(c) VGG19 + +![](images/f14feee229e1d871c0c9f9a4bf3510d048914bbc72da18f37fdb2a00fb66fa2b.jpg) +(d) ResNet152 + +![](images/36beb49379502e532a7e458717f469680dd7428a62426c9cf841c1ad959c2396.jpg) +(e) GoogleNet + +![](images/2d200e91fbd33143f9c0679bf72f2adaf5a50e390a19fae2b177fe7024a2f0c6.jpg) +Figure 5. The logit loss of original samples and adversarial samples crafted by GD-UAP, Cosine-UAP, our TRM-UAP methods. In each sub-figure, the subtitle indicates the surrogate model for crafting UAPs, and target models are shown on the horizontal axis. Original samples are classified correctly by the target model and exploited to craft AEs for comparisons. For each attack setting, we compute the average of the logit loss over all the samples. +bullet train(98.15%) + +![](images/0a432cc4dffc8d6ca71a39e3cb5fe1820ebee297308efd5c4adfbacffc937c49.jpg) +puffer(99.89%) + +![](images/d0494d384130a691d241b67c8c68102120a3dfff002e4c448a3a996c3597b8a7.jpg) +quilt(45.63%) +Figure 6. Examples of our TRM-UAP attack tested on VGG16 and the perturbation crafted by VGG19 (top: original examples; bottom: adversarial examples). The annotation represents the classification label and the probability predicted by CNNs. + +![](images/d0a84fbccd57c6376f0c62579ad8e800087dc0eafca838fd508e726eadfeb3b6.jpg) +theater curtain(72.82%) + +![](images/2229b9cbff59ff4b52fb8a35ccf0888fea7ff1ec65ab48d6a669fe35af48b83c.jpg) +monastery(34.69%) + +![](images/06e6dafd85668f5dc017d538c8984867aed7a23d311e63f3c5ad50d920fa6f07.jpg) +church(43.55%) + +![](images/80fb138084f9b9a0ba6f8d92ce61ed8391114c0a95a22fd87cd9e3195cda1559.jpg) +shoji(87.62%) + +![](images/fea73f939316b40e8544950656a81ac387fca478f012d0152d5a5d38dd1aa6c3.jpg) +shower curtain(94.35%) + +imental setting is the same as Section 4.1. Based on this, we also crafted UAPs by truncated positive activation only, i.e., $PTR$ remains unchanged and $NTR = 100\%$ . As shown in Fig. 7, the ablation experiment indicates that truncated negative activation combined with truncated positive activation improves the $FR$ by a large margin (around $8 \sim 18\%$ ) as compared to standalone truncated positive activation, thus enhancing both intensity and transferability of UAPs significantly. + +To explore the influence of truncation strategy, we computed the fooling rate of attacks in CNN models by varying truncation rates. Specifically, the proportion of retained activation increased by $10\%$ for each step. That means UAP learned the feature information from shallow to deep convolution layers gradually. Due to space limitations, UAPs + +![](images/54360931ab4da5e24809ad0f987e60634e820290f30d4fddee5b2fc4c6a85944.jpg) +(a) ResNet152 +Figure 7. The ablation study of truncated negative activations on CNN models. Positive: the attack only uses the truncated positive activation. Positive&Negative: the attack integrates truncated positive and negative activations. + +![](images/4a0325d7a358b602399802ea52d8d43043b08f33eb2521616b840fe45c56fe47.jpg) +(b) GoogleNet + +were only tested in white-box settings. The results of all CNN models were presented in Fig. 8, where the parameter space was explored w.r.t. NTR and PTR. As can be seen, the optimal result was obtained on the attack setting of the truncated activation rather than the whole activation. Especially in deep CNN models, only part of positive and negative activations were retained to craft UAPs. From the analysis of truncation in Fig. 8, we observe that the shallow convolution layer extracts generic features which are better at transferring the attack across different models. + +# 4.4. Exploring the Property of UAP + +In the following, the property of universal perturbations is thoroughly investigated. We studied the distinction between original example and adversarial examples made by image-specific attack (MI-FGSM) and data-free universal attacks (GD-UAP and TRM-UAP). As shown in Fig. 9, AEs + +![](images/572da65a74803e16285449d42a3de05c833fced37a732be51e6527405adca038.jpg) +(a) AlexNet + +![](images/f9584c2a5a9435aa9c1cfe179bf4704a78f252e1f5274770805964feb08086e0.jpg) +(b) VGG16 + +![](images/87e3bfd800fe78c910306a2e98dce874140db9cd2cc92ade1b8e808d20971ab2.jpg) +(c) VGG19 + +![](images/b391c6653a3ce10e43f466c0221f525adc1ebc0ac56fce94b04cc371894a638f.jpg) +(d) ResNet152 + +![](images/920e604745f5203df0ea6803dfda4ff9d12dedefe1c86835061fe8dbfe052133.jpg) +(e) GoogleNet + +![](images/a4b5e47d7ded467512cf5dec0c5bfe14807c8f647eed16c01d4b176527ca6108.jpg) +Figure 8. The parameter study of truncation rates on CNN models in white-box attack setting. The $x$ , $y$ , and $z$ axis are NTR, PTR, and FR, respectively. Note that there are only five convolution layers in AlexNet, truncation results (e.g., $10\%$ , $30\%$ , $50\%$ , etc.) are omitted since these values are consistent with the ones at their right coordinate. + +![](images/a6e036f3f9eece8a241eb2c3235ead53938aeb250022b273360956be79176ce6.jpg) + +![](images/891ea19d7f31bcdd2f5b5590e25f2f8fc377d0e4731f741e6638a74733302f89.jpg) + +![](images/31dc80df69a5682a54422bbd9b097440f2f58e52df4477e33cb6207ce04b108e.jpg) + +![](images/438c031f7e905448fb9b5d8ce6dd3a1b954c6bedb6ebdcb0f46d163df157fe45.jpg) +Figure 9. The distinction of feature maps among original image, image-specific attack, and universal attacks (top: feature maps extracted from a shallow convolution layer; bottom: original example and AEs made by VGG19). + +![](images/5bc5ff70594099e09d7459d6747fc33c69384834846216e9a5078d68df1bbf0c.jpg) + +![](images/bba6eb8b5f34e8f463ebbf59b2e27b1cf6ad621434c1670bf79a0dc52faf5d93.jpg) + +![](images/34635c586aa6340dfa8f236ea295eec759a14ab0f58e24b2467881216340d752.jpg) + +are indistinguishable compared with original example, and the imperceptible noise added to the input effectively perturbed feature maps in a shallow convolution layer. Furthermore, data-free universal attacks disturbed the feature map seriously in contrast to the image-specific attack. Nevertheless, the feature map of TRM-UAP had better regularity and repeatability than GD-UAP. We can conclude that features extracted by shallow convolution layers profoundly impact the final classification of CNN models and are beneficial for universal attack to craft the transferable perturbation. + +To further verify the above viewpoints, we trained UAPs in completely opposite truncation strategies (i.e., retaining former $80\%$ layers versus retaining latter $20\%$ layers), and counted the distribution of class labels of AEs. The result in Fig. 10 (a) was consistent with the hypothesis [33] that the top-1 logits prediction of UAP dominates the prediction of AEs. Both labels were theater curtain. Contrarily, it was shown in Fig. 10 (b) that the phenomenon of top-1 label domination disappeared, where the UAP label was obelisk and the AE label was brass. Thus, the shallow convolution layer is more beneficial to transfer universal attack with generic features through a variety of models. + +# 5. Conclusion + +In this paper, we have proposed a novel data-free universal attack, called TRM-UAP, to reformulate the UAP gener- + +![](images/01d53a6840e63a9d1b5bf12eaf66ab779888517bb3fbab19ae86c318d80a4959.jpg) +(a) UAP from low-level (former $80\%$ ) convolution layers + +![](images/c911b1a7b9b5b49dfa3f7a4e0c0296d9ed77dd75f12afc39f3dfbb359a82f5a5.jpg) + +![](images/7d129a42526e3fd9d856f53c514a5d25820b3343a1d11f7aa656c92dbfe0e2c5.jpg) +(b) UAP from high-level (latter $20\%$ ) convolution layers + +![](images/90d691a8ad3cbe797146d9aeffdfe746ba085acfa32aba42863f52dac3afeec5.jpg) +Figure 10. The analysis of the label distribution of UAPs and AEs (both made by VGG16). In each subfigure, the left is the top-10 logits value of UAP $(\pmb{v})$ , and the right is the distribution of top-10 predicated labels of AEs $(\pmb{v} + \pmb{x})$ . + +ation task as a truncated ratio optimization problem. Compared with previous methods, our TRM-UAP method integrates positive activation maximization with negative activation minimization. Towards further improving the transferability of UAPs, we propose a truncation strategy that computes positive and negative activation losses from low-level convolutions, as well as a curriculum optimization algorithm to fully mine the diversity of artificial images. Experimental results on the ImageNet dataset validate the better transferability of our TRM-UAP attack than other data-free UAP attacks on different CNN models. Additionally, UAPs learned from generic features in shallow convolution layers can dominate the model prediction and transfer the universal attack. + +Acknowledgement. This work is partially supported by the Scientific Research Foundation of Chongqing University of Technology under Grant No.2022ZDZ026, Natural Science Foundation of Chongqing, China under Grant No.CSTB2022NSCQ-MSX0493, the Key project of Chongqing Technology Innovation and Application Development under Grant No.cstc2021jscx-dxwtBX0018. + +# References + +[1] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 6:14410-14430, 2018. +[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009. +[3] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Šrndić, Pavel Laskov, Giorgio Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 387-402. Springer, 2013. +[4] Battista Biggio and Fabio Roli. Wild patterns: Ten years after the rise of adversarial machine learning. Pattern Recognition, 84:317-331, 2018. +[5] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy, pages 39-57, 2017. +[6] Ashutosh Chaubey, Nikhil Agrawal, Kavya Barnwal, Keerat K Guliani, and Pramod Mehta. Universal adversarial perturbations: A survey. arXiv preprint arXiv:2005.08087, 2020. +[7] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018. +[8] Aditya Ganeshan, Vivek BS, and R Venkatesh Babu. Fda: Feature disruptive attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8069-8079, 2019. +[9] Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. +[11] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in Neural Information Processing Systems, 32, 2019. +[12] Valentin Khrulkov and Ivan Oseledets. Art of singular vectors and universal adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8562-8570, 2018. +[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 2012. +[14] Hong Liu, Rongrong Ji, Jie Li, Baochang Zhang, Yue Gao, Yongjian Wu, and Feiyue Huang. Universal adversarial perturbation via prior driven uncertainty approximation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2941-2949, 2019. + +[15] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. +[16] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1765-1773, 2017. +[17] Konda Reddy Mopuri, Aditya Ganeshan, and R Venkatesh Babu. Generalizable data-free objective for crafting universal adversarial perturbations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(10):2452-2465, 2018. +[18] Konda Reddy Mopuri, Utsav Garg, and R Venkatesh Babu. Fast feature fool: A data independent approach to universal adversarial perturbations. In Proceedings of the British Machine Vision Conference, 2017. +[19] Konda Reddy Mopuri, Utkarsh Ojha, Utsav Garg, and R Venkatesh Babu. Nag: Network for adversary generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 742-751, 2018. +[20] Konda Reddy Mopuri, Phani Krishna Uppala, and R Venkatesh Babu. Ask, acquire, and attack: Data-free uap generation using class impressions. In Proceedings of the European Conference on Computer Vision, pages 20-35, 2018. +[21] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Symposium on Security and Privacy, pages 372-387, 2016. +[22] Omid Poursaeed, Isay Katsman, Bicheng Gao, and Serge Belongie. Generative adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4422-4431, 2018. +[23] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211–252, 2015. +[24] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015. +[25] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015. +[26] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. +[27] Yusuke Tashiro, Yang Song, and Stefano Ermon. Diversity can be transferred: Output diversification for white-and black-box attacks. Advances in Neural Information Processing Systems, 33:4536-4548, 2020. + +[28] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In International Conference on Learning Representations, 2019. +[29] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7639-7648, 2021. +[30] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in Neural Information Processing Systems, 27, 2014. +[31] Yunrui Yu, Xitong Gao, and Cheng-Zhong Xu. Lafeat: piercing through adversarial defenses with latent features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5735-5745, 2021. +[32] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In So Kweon. Understanding adversarial examples from the mutual influence of images and perturbations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14521-14530, 2020. +[33] Chaoning Zhang, Philipp Benz, Adil Karjauv, and In So Kweon. Data-free universal adversarial perturbation and black-box attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7868-7877, 2021. +[34] Chaoning Zhang, Philipp Benz, Chenguo Lin, Adil Karjauv, Jing Wu, and In So Kweon. A survey on universal adversarial attack. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages 4687-4694, 2021. +[35] Jianping Zhang, Weibin Wu, Jen-tse Huang, Yizhan Huang, Wenxuan Wang, Yuxin Su, and Michael R Lyu. Improving adversarial transferability via neuron attribution-based attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14993-15002, 2022. \ No newline at end of file diff --git a/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/images.zip b/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a89d18f15e1bbad7a21b43d85063593284173093 --- /dev/null +++ b/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:163164afa90c10a137946792251e479b1e4320e5e38cf57bae230b1b59daf628 +size 672590 diff --git a/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/layout.json b/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..98afd1a014a2af652d17da2ba7d25d5afeb07c87 --- /dev/null +++ b/2023/TRM-UAP_ Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization/layout.json @@ -0,0 +1,11357 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 538, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 538, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 538, + 138 + ], + "type": "text", + "content": "TRM-UAP: Enhancing the Transferability of Data-Free Universal Adversarial Perturbation via Truncated Ratio Maximization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 161, + 499, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 161, + 499, + 203 + ], + "spans": [ + { + "bbox": [ + 93, + 161, + 499, + 203 + ], + "type": "text", + "content": "Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, Di Ming* \nSchool of Computer Science and Engineering, Chongqing University of Technology \nChongqing, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 205, + 462, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 462, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 462, + 217 + ], + "type": "text", + "content": "lyrl99804@qq.com, {xfeng, ylwang, yangwu, diming}@cqut.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 270, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 289, + 594 + ], + "type": "text", + "content": "Aiming at crafting a single universal adversarial perturbation (UAP) to fool CNN models for various data samples, universal attack enables a more efficient and accurate evaluation for the robustness of CNN models. Early universal attacks craft UAPs depending on data priors. For more practical applications, the data-free universal attacks that make UAPs from random noises have aroused much attention recently. However, existing data-free UAP methods perturb all the CNN feature layers equally via the maximization of the CNN activation, leading to poor transferability. In this paper, we propose a novel data-free universal attack without depending on any real data samples through truncated ratio maximization, which we term as TRM-UAP. Specifically, different from the maximization of the positive activation in convolution layers, we propose to optimize the UAP generation from the ratio of positive and negative activations. To further enhance the transferability of universal attack, TRM-UAP not only performs the ratio maximization merely on low-level generic features via the truncation strategy, but also incorporates a curriculum optimization algorithm that can effectively learn the diversity of artificial images. Extensive experiments on the ImageNet dataset verify that TRM-UAP achieves a state-of-the-art average fooling rate and excellent transferability on different CNN models as compared to other data-free UAP methods. Code is available at https://github.com/RandolphCarter0/TRMUAP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "content": "Early research [26] shows that tiny and imperceptible perturbations can seriously disturb the prediction results of deep neural networks (DNNs), especially for image recognition tasks [3]. Adversarial Examples (AEs), crafted by adding tiny perturbations deliberately to benign samples," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 246, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 246, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 246, + 545, + 293 + ], + "type": "text", + "content": "are not only imperceptible in computer vision tasks but also prone to transfer among the DNN models. Therefore, AEs have been regarded as a serious threat to DNN models since the development of deep learning [1]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 294, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 546, + 510 + ], + "type": "text", + "content": "To explore the impact of AEs, many methods are proposed to craft the adversarial perturbation that is highly transferable to various DNN models (i.e., with a high fooling rate on other DNN models). However, the AEs generated by these works are explicitly designed for some specific samples and often fail to perturb other samples that are even from the same dataset. Different from the aforementioned image-specific attack, the universal attack that generates image-agnostic universal adversarial perturbation (UAP) was proposed in [16]. The UAP in a universal attack setting is trained from prior knowledge, such as substitute data, surrogate model, etc. By adding the UAP to benign samples, universal attacks can generate numerous AEs all together in a very short period. Furthermore, the universal attack can fool most DNN models that are trained from similar datasets, and can greatly reduce the computational cost of crafting AEs, making adversaries more applicable to real scenarios than the image-specific attacks [4, 6, 34]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": "Whereas, no matter the image-specific attacks or the universal attacks, well-annotated training data (e.g., [21, 9, 7]) or substitute data (e.g., [16, 12, 22, 19]) is required to generate AEs. In practice, obtaining a well-labeled large-scale dataset is challenging and costly, especially for some applications with critical security demand, where less prior knowledge is available. Recently, researchers have investigated data-free universal attack methods [18, 17, 14, 20, 33], where AEs are generated directly from random noises rather than data priors. Compared with previous methods that disturb the gradient or maximize the classification loss in data-dependent scenarios, current data-free UAP methods explore the feature-based perturbation, which tries to maximize the activation (e.g., ReLU) of convolutional neural network (CNN) features. It is shown that feature-based UAP methods achieved highly efficient and applicable universal attack without using any data priors [4, 6, 34]. How" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 169, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 169, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 169, + 713 + ], + "type": "text", + "content": "*Corresponding Author: Di Ming" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4762" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "ever, current data-free UAP methods only consider the positive activation [18, 17] and perturb all layers of CNN features equally [20, 14]. As a result, UAPs crafted by surrogate models are hard to transfer to target models [6, 34]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 289, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 289, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 289, + 443 + ], + "type": "text", + "content": "Towards enhancing the transferability of data-free UAPs, we propose a novel data-free universal attack method called TRM-UAP, which formulates the UAP generation as a truncated ratio maximization problem. In particular, TRM-UAP attempts to over-fire positive neurons so that extracted features from multiple CNN layers can be fully disrupted. Besides the maximization of positive neurons, the proper activation on negative neurons may also be helpful. Hence, TRM-UAP performs a ratio maximization of positive and negative activations, which facilitates some negative neurons transiting to be positive. Moreover, some feature-based adversarial attacks [29, 35] reveal that multiple layers of CNN features are not equally important for disturbing the image classification [11, 28]. To further improve the transferability of universal attack, TRM-UAP performs ratio maximization only on parts of CNN activation layers. Specifically, a truncation strategy is proposed to truncate the activation of high-level convolutions and retain the activation of shallow convolutions to train perturbations. The intuition is that low-level CNN features are shown to provide more generic (data-independent) feature representations than high-level semantic features [30]. Thus the disturbed low-level features are expected to be able to transfer attacks across different CNN models, leading to highly transferable UAPs. Furthermore, a curriculum learning [2] based optimization algorithm is introduced to utilize artificial images and explore the diversity of input." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 445, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 287, + 467 + ], + "type": "text", + "content": "In general, the main contributions of our work can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 471, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 471, + 286, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 471, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 58, + 471, + 286, + 506 + ], + "type": "text", + "content": "- We propose a novel data-free universal attack method to craft image-agnostic adversarial perturbations without utilizing any real data samples during training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 508, + 286, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 508, + 286, + 650 + ], + "spans": [ + { + "bbox": [ + 58, + 508, + 286, + 650 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to formulate the data-free universal attack as a truncated ratio maximization problem. The proposed TRM-UAP enhances the transferability of UAPs from three perspectives: (i) the maximization on CNN features for crafting UAPs is enhanced by both maximization of positive activation and minimization of negative activation. (ii) the ratio maximization on truncated feature layers improves the generalizability of universal attacks to CNN models. (iii) artificial images are incorporated into curriculum optimization algorithm to strengthen the UAP dominance on feature activations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 654, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 654, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 654, + 286, + 712 + ], + "type": "text", + "content": "- Extensive experiments on the ImageNet dataset show that TRM-UAP obtains the highest average fooling rate compared with other data-free UAP methods, indicating that truncated ratio maximization can greatly improve the transferability of universal attack." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 91, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 305 + ], + "type": "text", + "content": "Image-specific attacks: The attacks that try to craft perturbations for corresponding images by utilizing training or substitute data are termed as image-specific attacks. Typical image-specific attacks, such as Fast Gradient Sign Method (FGSM) [9], I-FGSM and MI-FGSM [7], Project Gradient Descent (PGD) [15], and Output Diversified Sampling (ODS) [27], etc., usually make use of the gradient information from well-trained target CNN model (i.e., white-box attack) or surrogate model (i.e., black-box attack) to disturb images. However, since gradient-based attacks often require a number of iterations, the computational cost of making AEs becomes extremely expensive as the number of examples increases. Recently, feature-based attacks that try to craft AEs from the perspective of CNN features have received broad concern [29, 8, 35]. For example, Neuron Attribution based Attack (NAA) [35] conducted feature-level attacks by deploying a neuron attribution method to estimate the importance of neurons at multiple CNN layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 307, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 462 + ], + "type": "text", + "content": "Data-dependent universal attacks: The universal attack aims at training a single perturbation that is independent of any specific sample, i.e., UAP. By adding a single UAP to various benign samples, numerous AEs can be generated very efficiently. Until recently, most of the universal attacks are still data-dependent, meaning that the perturbations are learned from data priors, such as training data, substitute data, etc. For example, the primary universal attack [16] tried to seek a universal perturbation for a set of training samples. On the other hand, there are some works [22, 19] applying generative adversarial models to craft perturbations. Recently, Zhang et al. [32] proposed to generate UAPs by training perturbations on the proxy datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 713 + ], + "type": "text", + "content": "Data-free universal attacks: The data-free universal attack makes UAP via random noise rather than prior knowledge from data directly. It is considered the most applicable attack for real applications in the adversarial attack field. However, since the information of both data and target models is unknown, the data-free universal attack is extremely challenging and only a few recent works focus on it. Mopuri et al. [18] first propose a data-free universal attack, i.e., the Fast Feature Fool (FFF) method, by maximizing the feature activations across all the CNN layers. In addition, they further propose a Generalizable Data-free UAP (GDUAP) to improve the FFF attack via optimizing the UAP training process with a saturation check strategy [17]. Besides, AAA [20] learns a generative model to make UAPs based on the pre-defined class impression. However, as the number of image categories increases, the cost of acquiring class impressions will also broaden. Prior-Driven Uncertainty Approximation (PD-UA) [14] proposes to craft UAPs by maximizing the uncertainty approximation of the model. Cosine-UAP [33] introduces the cosine similarity to train UAPs in a self-supervised way. Because of the shortage of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4763" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "prior knowledge, current data-free UAP methods generally present poor transferability across various CNN models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 106, + 129, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 129, + 119 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 129, + 119 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 126, + 287, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 126, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 47, + 126, + 287, + 174 + ], + "type": "text", + "content": "In this section, we first introduce the motivation of maximizing the ratio of activations. After that, we present the details of the proposed truncated ratio maximization method to craft the UAP in a strictly data-free fashion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 181, + 126, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 181, + 126, + 193 + ], + "spans": [ + { + "bbox": [ + 47, + 181, + 126, + 193 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 199, + 287, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 199, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 199, + 287, + 235 + ], + "type": "text", + "content": "To fool CNN models " + }, + { + "bbox": [ + 47, + 199, + 287, + 235 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 199, + 287, + 235 + ], + "type": "text", + "content": ", UAP [16] tries to find the universal perturbation " + }, + { + "bbox": [ + 47, + 199, + 287, + 235 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 47, + 199, + 287, + 235 + ], + "type": "text", + "content": " which maximizes the classification loss " + }, + { + "bbox": [ + 47, + 199, + 287, + 235 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 47, + 199, + 287, + 235 + ], + "type": "text", + "content": " (e.g., cross entropy) over the data distribution " + }, + { + "bbox": [ + 47, + 199, + 287, + 235 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 47, + 199, + 287, + 235 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 243, + 287, + 263 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 243, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 58, + 243, + 287, + 263 + ], + "type": "interline_equation", + "content": "\\max _ {\\boldsymbol {v} \\sim \\mathbb {S}} \\underset {(\\boldsymbol {x}, \\boldsymbol {y}) \\sim \\mathbb {D}} {\\mathbb {E}} [ \\mathcal {L} (f (\\boldsymbol {v} + \\boldsymbol {x}), \\boldsymbol {y})) ] \\quad \\text {s . t .} \\| \\boldsymbol {v} \\| _ {p} \\leq \\epsilon \\tag {1}", + "image_path": "332a12953dad3fc8b860f11f31557a5b8df3e55476e8a7a18008a8bc280edf42.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\mathbb{S}" + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "text", + "content": " is the perturbation space, " + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "inline_equation", + "content": "(\\pmb{x},\\pmb{y})" + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "text", + "content": " are the data and the class label (or one-hot ground truth) sampled from " + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "text", + "content": " is the output of CNN models, and " + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "text", + "content": " is the constant to constrain the perturbation " + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "text", + "content": "-norm bound (" + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "inline_equation", + "content": "e.g., p = 0,1,2,\\infty" + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "text", + "content": "). To be consistent with previous works [18, 17, 32], we adopt the " + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 47, + 270, + 287, + 354 + ], + "type": "text", + "content": "-norm to realize the imperceptibility of perturbation for all analyzes throughout this paper." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 354, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 425 + ], + "type": "text", + "content": "However, the training data of the target model is usually unavailable and hard to access, making the data-dependence approaches not suitable for practical adversarial attacks. As a consequence, Mopuri et al. [18] propose the Fast Feature Fool (FFF) method to maximize the CNN activations in a strictly data-free fashion:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 79, + 433, + 287, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 433, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 79, + 433, + 287, + 454 + ], + "type": "interline_equation", + "content": "\\max _ {\\boldsymbol {v}} \\quad | | \\mathcal {A} ^ {(i)} (\\boldsymbol {v}) | | _ {2}, \\quad \\text {f o r} i = 1, 2, \\dots , L \\tag {2}", + "image_path": "608dac96c8f71d364577d576b645484f24c3bba8da81f7a6b8a6d8e41d4a23df.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 451, + 154, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 451, + 154, + 462 + ], + "spans": [ + { + "bbox": [ + 83, + 451, + 154, + 462 + ], + "type": "interline_equation", + "content": "\\text {s . t .} \\quad | | \\boldsymbol {v} | | _ {\\infty} \\leq \\epsilon", + "image_path": "74f099f92fe8a39e8d1de701928d199694e93f0b93cc79ab7583681c3a81f377.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 470, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 470, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 470, + 287, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 470, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{(i)}(\\pmb{v}) = \\mathrm{Activation}(\\mathcal{C}^{(i)}(\\pmb{v}))" + }, + { + "bbox": [ + 47, + 470, + 287, + 555 + ], + "type": "text", + "content": " (e.g.,ReLU activation), " + }, + { + "bbox": [ + 47, + 470, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(i)}(\\pmb{v})" + }, + { + "bbox": [ + 47, + 470, + 287, + 555 + ], + "type": "text", + "content": " is the output of the " + }, + { + "bbox": [ + 47, + 470, + 287, + 555 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 470, + 287, + 555 + ], + "type": "text", + "content": "-th convolution layer, and " + }, + { + "bbox": [ + 47, + 470, + 287, + 555 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 470, + 287, + 555 + ], + "type": "text", + "content": " is the total number of convolution layers. It can be seen that this typical data-free universal attack method aims at accumulating the errors gradually through multiple convolution layers to enlarge the classification loss, finally leading to the misclassification of perturbed samples " + }, + { + "bbox": [ + 47, + 470, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{v} + \\pmb{x}" + }, + { + "bbox": [ + 47, + 470, + 287, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 562, + 239, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 239, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 239, + 574 + ], + "type": "text", + "content": "3.2. Maximizing the Ratio of Activations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "type": "text", + "content": "Different from previous data-free universal methods [18, 17] that only maximize the activation of convolution layer (see Fig. 1 (a)), we propose to craft the universal adversarial perturbation " + }, + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 47, + 580, + 287, + 628 + ], + "type": "text", + "content": " via maximizing the ratio of activations:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 78, + 635, + 287, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 635, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 78, + 635, + 287, + 666 + ], + "type": "interline_equation", + "content": "\\max _ {\\boldsymbol {v}} \\frac {| | \\mathcal {C} _ {+} ^ {(i)} (\\boldsymbol {v}) | | _ {2}}{| | \\mathcal {C} _ {-} ^ {(i)} (\\boldsymbol {v}) | | _ {2}}, \\quad \\text {f o r} i = 1, 2, \\dots , L \\tag {3}", + "image_path": "026aa44048575daaf9c7b008622c2426a7b0f4a227362bf8c1dd6f71cac60190.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 83, + 669, + 154, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 669, + 154, + 681 + ], + "spans": [ + { + "bbox": [ + 83, + 669, + 154, + 681 + ], + "type": "interline_equation", + "content": "\\text {s . t .} \\quad | | \\boldsymbol {v} | | _ {\\infty} \\leq \\epsilon", + "image_path": "c1887c409e81f74b0fea62675e3f9fdc3da597c92468a8726ff87c240c787a63.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 688, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 688, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 688, + 287, + 714 + ], + "type": "text", + "content": "where positive activation and negative activation in " + }, + { + "bbox": [ + 47, + 688, + 287, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 688, + 287, + 714 + ], + "type": "text", + "content": "-th convolution layer are defined as " + }, + { + "bbox": [ + 47, + 688, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{+}^{(i)}(\\pmb{v}) = \\max (\\mathcal{C}^{(i)}(\\pmb{v}),0)" + }, + { + "bbox": [ + 47, + 688, + 287, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 536, + 223 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 536, + 223 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 536, + 223 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 536, + 223 + ], + "type": "image", + "image_path": "cd4db687608d88f9c2802209819768be2d60e16af792fcbe04bb63b9ff227737.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 226, + 545, + 250 + ], + "lines": [ + { + "bbox": [ + 305, + 226, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 305, + 226, + 545, + 250 + ], + "type": "text", + "content": "Figure 1. The illustration of previous positive maximization [18, 17] and our proposed ratio maximization." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 264, + 544, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 264, + 544, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 264, + 544, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{-}^{(i)}(\\pmb {v}) = \\min (\\mathcal{C}^{(i)}(\\pmb {v}),0)" + }, + { + "bbox": [ + 305, + 264, + 544, + 289 + ], + "type": "text", + "content": " respectively, and their magnitudes are measured by " + }, + { + "bbox": [ + 305, + 264, + 544, + 289 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 305, + 264, + 544, + 289 + ], + "type": "text", + "content": " -norm." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 545, + 434 + ], + "type": "text", + "content": "Towards further enlarging the fooling probability, this ratio formulation (3) tries to increase the magnitude and the number of positive activation via maximizing positive activation and minimizing negative activation simultaneously. Specifically speaking, ratio maximization is a dynamic process. Firstly, minimizing the negative activation could lead to the change of sign in some negative features whose values are close to zero, i.e., jumping from a tiny negative value to a tiny positive value, as shown in Fig. 1 (b). Afterward, at the next iteration, maximizing the positive activation will continue to increase the magnitude of above-mentioned features from a tiny positive value to a large positive value." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 434, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 530 + ], + "type": "text", + "content": "During training, with the decrease in the number of negative activation and the increase in the number of positive activation, our proposed formulation (3) can achieve a higher ratio of positive activation as compared to previous data-free universal attacks, e.g., the formulation (2). Meanwhile, it also indicates that our ratio maximization could craft the UAP with more significant attack intensity so as to increase the fooling probability further." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 539, + 473, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 473, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 473, + 550 + ], + "type": "text", + "content": "3.3. Truncated Ratio Maximization" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Features extracted from shallow to deep convolution layers behave differently in crafting the adversarial perturbation. LAFEAT [31] analyzes the influence of each intermediate convolution layer on the misclassification, and attacks features only in a specific layer. On the other hand, GD-UAP [17] thoroughly studied the relative change in feature activations of adversarial examples as compared to original images at multiple CNN layers. However, GD-UAP attacks the feature activations extracted from all convolution layers together, without taking layer-wise feature patterns into consideration. According to [30], low-level features in shallow layers have generic patterns, while high-level features in deep layers have specific patterns. It can be seen that" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4764" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 72, + 251, + 212 + ], + "blocks": [ + { + "bbox": [ + 83, + 72, + 251, + 212 + ], + "lines": [ + { + "bbox": [ + 83, + 72, + 251, + 212 + ], + "spans": [ + { + "bbox": [ + 83, + 72, + 251, + 212 + ], + "type": "image", + "image_path": "3d806e9afd92dc1567563c9adf926db09e929a69e828aaa62d22d4334f4e8c45.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 215, + 286, + 239 + ], + "lines": [ + { + "bbox": [ + 47, + 215, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 286, + 239 + ], + "type": "text", + "content": "Figure 2. Truncated positive and negative activation losses, where cumulative shaded areas are used to craft the UAP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 258, + 286, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 286, + 317 + ], + "type": "text", + "content": "not all convolution layers contribute positively to increasing the positive activation and the objective loss. Therefore, based on ratio maximization defined in formulation (3), we propose a truncated ratio maximization method to further enhance the intensity and the transferability of UAPs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 318, + 286, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 286, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 286, + 393 + ], + "type": "text", + "content": "For notational simplicity, we begin by defining the ratio maximization problem in " + }, + { + "bbox": [ + 47, + 318, + 286, + 393 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 318, + 286, + 393 + ], + "type": "text", + "content": "-th convolution layer as " + }, + { + "bbox": [ + 47, + 318, + 286, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{(i)}(\\pmb{v}) = \\mathcal{L}_{+}^{(i)}(\\pmb{v}) / \\mathcal{L}_{-}^{(i)}(\\pmb{v})" + }, + { + "bbox": [ + 47, + 318, + 286, + 393 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 318, + 286, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{+}^{(i)}(\\pmb{v}) = ||\\mathcal{C}_{+}^{(i)}(\\pmb{v})||_{2}" + }, + { + "bbox": [ + 47, + 318, + 286, + 393 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 318, + 286, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{-}^{(i)}(\\pmb{v}) = ||\\mathcal{C}_{-}^{(i)}(\\pmb{v})||_{2}" + }, + { + "bbox": [ + 47, + 318, + 286, + 393 + ], + "type": "text", + "content": ". To maximize the ratio of activations in convolution layers altogether, the overall loss function of ratio maximization is reformulated as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 397, + 287, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 397, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 118, + 397, + 287, + 430 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\boldsymbol {v}) = \\sum_ {i = 1} ^ {L} \\log \\mathcal {L} ^ {(i)} (\\boldsymbol {v}) \\tag {4}", + "image_path": "4530e4108f5fa6dc6d6ce359bd4b6e8407a54bf3090f8c6bde37dec1f1401c4a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "text", + "content": "where log rescales the activation to prevent gradient explosions. Different from other methods that craft UAPs based on all convolution layers or a specific convolution layer, we propose to truncate the computation of positive and negative activations in deep convolution layers, since specific feature patterns are not beneficial for improving the quality of UAPs. Assuming that the truncation rate " + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "inline_equation", + "content": "(TR)" + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "inline_equation", + "content": "TR = \\lfloor (L - l) / L\\rfloor \\%" + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "text", + "content": " , we make truncation on " + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "inline_equation", + "content": "(l + 1)" + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "text", + "content": " -th convolution layer by setting " + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{+}^{(i)}(\\pmb {v}) = \\tau" + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "inline_equation", + "content": "(i > l)" + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{-}^{(i)}(\\pmb {v}) = \\tau" + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "inline_equation", + "content": "(i > l)" + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "text", + "content": " is a small positive number such as " + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "inline_equation", + "content": "1e - 9" + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "text", + "content": " . As a result, " + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\pmb {v})" + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "text", + "content": " can be equivalently rewritten as " + }, + { + "bbox": [ + 46, + 435, + 286, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\pmb {v}) = \\sum_{i = 1}^{L}\\log \\mathcal{L}^{(i)}(\\pmb {v}) = \\sum_{i = 1}^{l}\\log \\mathcal{L}^{(i)}(\\pmb {v})" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "text", + "content": "However, " + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{+}^{(i)}(\\pmb{v})" + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{-}^{(i)}(\\pmb{v})" + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "text", + "content": " not only have different scales but also behave differently in various layers for maximizing the ratio loss, which could lead to a degradation of the attack performance. To resolve these problems, we first reformulate the ratio loss in " + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 586, + 286, + 647 + ], + "type": "text", + "content": "-th convolution layer as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 652, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 652, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 116, + 652, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\alpha} ^ {(i)} (\\boldsymbol {v}) = \\frac {\\mathcal {L} _ {+} ^ {(i)} (\\boldsymbol {v})}{(\\mathcal {L} _ {-} ^ {(i)} (\\boldsymbol {v})) ^ {\\alpha}}, \\tag {5}", + "image_path": "9e8cb8cdd6955d3ec9a11f971378b1bc50324434a40054b000fc3e77837a59a0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": " is a scaling hyperparameter to adjust the relative importance between positive and negative activation losses." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 326, + 73, + 521, + 213 + ], + "blocks": [ + { + "bbox": [ + 326, + 73, + 521, + 213 + ], + "lines": [ + { + "bbox": [ + 326, + 73, + 521, + 213 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 521, + 213 + ], + "type": "image", + "image_path": "cd1b9880d0864a22bc1b11b2a6e89447b521da93c5cfcf74790729c2e5b06279.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 216, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 305, + 216, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 545, + 239 + ], + "type": "text", + "content": "Figure 3. The mean positive activation loss in various convolution layers with different truncation settings." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 259, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 259, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 259, + 545, + 334 + ], + "type": "text", + "content": "Secondly, positive and negative truncation rates are defined as " + }, + { + "bbox": [ + 305, + 259, + 545, + 334 + ], + "type": "inline_equation", + "content": "PTR = \\lfloor (L - l') / L \\rfloor \\%" + }, + { + "bbox": [ + 305, + 259, + 545, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 259, + 545, + 334 + ], + "type": "inline_equation", + "content": "NTR = \\lfloor (L - l'') / L \\rfloor \\%" + }, + { + "bbox": [ + 305, + 259, + 545, + 334 + ], + "type": "text", + "content": " respectively. That is to say, we make truncation on different layers for positive and negative activation losses by setting " + }, + { + "bbox": [ + 305, + 259, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{+}^{(i)}(\\pmb{v}) = \\tau (i > l')" + }, + { + "bbox": [ + 305, + 259, + 545, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 259, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{-}^{(i)}(\\pmb{v}) = \\tau (i > l'')" + }, + { + "bbox": [ + 305, + 259, + 545, + 334 + ], + "type": "text", + "content": ", see Fig. 2. Finally, the overall loss function " + }, + { + "bbox": [ + 305, + 259, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\pmb{v})" + }, + { + "bbox": [ + 305, + 259, + 545, + 334 + ], + "type": "text", + "content": " can be reduced to" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 341, + 544, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 544, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 544, + 410 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\boldsymbol {v}) = \\sum_ {i = 1} ^ {L} \\log \\mathcal {L} _ {\\alpha} ^ {(i)} (\\boldsymbol {v}) \\\\ = \\sum_ {i = 1} ^ {l ^ {\\prime}} \\log \\mathcal {L} _ {+} ^ {(i)} (\\boldsymbol {v}) - \\alpha \\cdot \\sum_ {i = 1} ^ {l ^ {\\prime \\prime}} \\log \\mathcal {L} _ {-} ^ {(i)} (\\boldsymbol {v}) + c \\\\ \\propto \\sum_ {i = 1} ^ {l ^ {\\prime}} \\log \\mathcal {L} _ {+} ^ {(i)} (\\boldsymbol {v}) - \\alpha \\cdot \\sum_ {i = 1} ^ {l ^ {\\prime \\prime}} \\log \\mathcal {L} _ {-} ^ {(i)} (\\boldsymbol {v}), \\tag {6} \\\\ \\end{array}", + "image_path": "42a8dc72136b02a905fa0d15e3ee0b59da7942e62b2f3a93b68b6cac56413066.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 410, + 544, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 544, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 544, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 410, + 544, + 422 + ], + "type": "inline_equation", + "content": "c = ((1 - \\alpha)\\cdot L - l^{\\prime} + \\alpha \\cdot l^{\\prime \\prime})\\cdot \\log \\tau" + }, + { + "bbox": [ + 306, + 410, + 544, + 422 + ], + "type": "text", + "content": " , i.e., a constant." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 422, + 544, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 544, + 459 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 544, + 459 + ], + "type": "text", + "content": "Consequently, for data-free universal adversarial attack, we propose a truncated ratio maximization method (TRMUAP) to craft the perturbation " + }, + { + "bbox": [ + 305, + 422, + 544, + 459 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 305, + 422, + 544, + 459 + ], + "type": "text", + "content": " satisfying" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 466, + 544, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 544, + 512 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 544, + 512 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max _ {\\boldsymbol {v}} \\sum_ {i = 1} ^ {l ^ {\\prime}} \\log \\left| \\left| \\mathcal {C} _ {+} ^ {(i)} (\\boldsymbol {v}) \\right| \\right| _ {2} - \\alpha \\cdot \\sum_ {i = 1} ^ {l ^ {\\prime \\prime}} \\log \\left| \\left| \\mathcal {C} _ {-} ^ {(i)} (\\boldsymbol {v}) \\right| \\right| _ {2} \\tag {7} \\\\ \\text {s . t .} \\quad | | \\boldsymbol {v} | | _ {\\infty} \\leq \\epsilon \\\\ \\end{array}", + "image_path": "885763fe88063b2f6445e9db04215c828e20a7437ffaf61d776448dc1c447a97.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "text", + "content": " is constrained by " + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "text", + "content": "-norm and bound " + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "text", + "content": ". To verify the previous claim, we compute the mean positive activation loss in various convolution layers via feeding UAPs into CNN models (e.g., VGG19). As shown in Fig. 3, TRM-UAP " + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "(PTR \\neq NTR" + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "l' \\neq l'' < L)" + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "text", + "content": " generates a larger cumulative positive activation loss than RM-UAP " + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "(l' = l'' = L)" + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "text", + "content": " and TRM-UAP " + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "(PTR = NTR" + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "inline_equation", + "content": "l' = l'' < L)" + }, + { + "bbox": [ + 305, + 520, + 545, + 616 + ], + "type": "text", + "content": ", which could lead to a further increase in fooling rates." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 623, + 500, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 500, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 500, + 636 + ], + "type": "text", + "content": "3.4. Curriculum Optimization Algorithm" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 545, + 713 + ], + "type": "text", + "content": "For purpose of improving the diversity of inputs, we utilize Gaussian noise and jigsaw image [33] to generate artificial images, which are fed into surrogate models together with the perturbation. To resemble training samples in real-world datasets, the mean filter is used to smooth the edge between different regions in jigsaw images. However, it" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4765" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "will be difficult for the optimization algorithm to converge if artificial images with complex patterns are provided at the beginning of the training procedure. Inspired by curriculum learning [2], we start the training with simple artificial images and then gradually feed more complex artificial images into surrogate models. Specifically, the generation of artificial images is controlled by the distribution parameter and the training iteration, defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 174, + 212, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 174, + 212, + 186 + ], + "spans": [ + { + "bbox": [ + 116, + 174, + 212, + 186 + ], + "type": "interline_equation", + "content": "D _ {1} \\prec D _ {2} \\prec \\dots \\prec D _ {n},", + "image_path": "821930ca2c7d99b79644b83ee3e46292163b7b5c8749994d8b02f70d9b6597f7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 186, + 286, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 186, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 116, + 186, + 286, + 201 + ], + "type": "interline_equation", + "content": "D _ {t} = \\left\\{\\boldsymbol {x} \\mid \\boldsymbol {x} \\sim P \\left(\\theta_ {0}, t\\right) \\right\\}, \\tag {8}", + "image_path": "587c380691c120e64a933910e0749a5f0c0c936082e0922d2822ec198f06ddd9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "content": " is a set of artificial images " + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "content": " sampled at " + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "content": "-th iteration " + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "inline_equation", + "content": "(t = 1,2,\\dots ,n)" + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "content": " is the distribution of artificial samples with the default parameter " + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "inline_equation", + "content": "D_{a} \\prec D_{b}" + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "content": " denotes that the pattern of artificial images in " + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "inline_equation", + "content": "D_{b}" + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "content": " is more complex than " + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "inline_equation", + "content": "D_{a}" + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "content": ". In detail, as the number of training iterations " + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "content": " increases, the distribution parameter " + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "content": " (e.g., the standard deviation of Gaussian distribution and the frequency of jigsaw images) is gradually increasing. For " + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 207, + 289, + 316 + ], + "type": "text", + "content": "-th iteration, our curriculum optimization algorithm is maximizing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 322, + 286, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 286, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = \\frac {1}{| D _ {t} |} \\sum_ {\\boldsymbol {x} \\in D _ {t}} \\left(\\sum_ {i = 1} ^ {l ^ {\\prime}} \\log \\mathcal {L} _ {+} ^ {(i)} (\\boldsymbol {v} + \\boldsymbol {x}) - \\alpha \\cdot \\sum_ {i = 1} ^ {l ^ {\\prime \\prime}} \\log \\mathcal {L} _ {-} ^ {(i)} (\\boldsymbol {v} + \\boldsymbol {x})\\right) \\tag {9}", + "image_path": "f4b1f51ac3871d0b5c46a5536e0db392338ebe313989db7db6fa6d75882d5b51.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 366, + 284, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 284, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 284, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 366, + 284, + 378 + ], + "type": "inline_equation", + "content": "|D_t|" + }, + { + "bbox": [ + 46, + 366, + 284, + 378 + ], + "type": "text", + "content": " is the number of artificial images in the set " + }, + { + "bbox": [ + 46, + 366, + 284, + 378 + ], + "type": "inline_equation", + "content": "D_t" + }, + { + "bbox": [ + 46, + 366, + 284, + 378 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "content": "The entire procedure of our proposed TRM-UAP method is summarized in Algorithm 1. During the training process, the perturbation vector " + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "content": " is randomly initialized by the uniform distribution " + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "content": ", and a surrogate dataset is used to validate the convergence of optimization for every " + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "content": " iterations. " + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "inline_equation", + "content": "l^{\\prime}" + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "inline_equation", + "content": "l''" + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "content": " represent the retained layers of positive and negative activations corresponding to truncation rates " + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "inline_equation", + "content": "PTR" + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "inline_equation", + "content": "NTR" + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "content": " respectively. The scaling hyperparameter " + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "content": " adjusts the ratio between truncated positive and negative activation losses. The algorithm is converged only if either the iteration number reaches the maximum " + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "content": " or the fooling rate test reaches the threshold " + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "inline_equation", + "content": "F_{max}" + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "content": ". Similar to GD-UAP [17], we run a saturation rate test for each iteration to dynamically compress the perturbation vector only if the saturation rate " + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "inline_equation", + "content": "\\hat{r}" + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "content": " is smaller than the predefined threshold " + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 378, + 289, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 563, + 285, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 285, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 285, + 575 + ], + "type": "text", + "content": "3.5. Connection to Other Data-Free UAP Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "When positive and negative truncation rates are set as " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "PTR = 0\\%" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "NTR = 100\\%" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "l' = L, l'' = 0" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "), this means all the negative activations are truncated and all the positive activations are retained. Thus, the proposed TRM-UAP defined in Eq. (7) is reduced to GD-UAP [17]. It can be seen that GD-UAP is only a special case of TRM-UAP under the data-free universal attack setting. As compared to GD-UAP, our TRM-UAP improves both intensity and transferability of crafted UAPs via truncated ratio maximization. The detail of their differences will be further discussed in the experiments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "Algorithm 1 Curriculum optimization algorithm for solving truncated ratio maximization problem to craft universal adversarial perturbations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "text", + "content": "Input: surrogate CNN model " + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "text", + "content": ", limitation value " + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "text", + "content": ", learning rate " + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "text", + "content": ", positive and negative truncation rates " + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "inline_equation", + "content": "PTR" + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "inline_equation", + "content": "NTR" + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "text", + "content": ", scaling hyperparameter " + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "text", + "content": ", maximum iteration number " + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "text", + "content": ", convergence threshold " + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "inline_equation", + "content": "F_{max}" + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "text", + "content": ", validation test hyperparameter " + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "text", + "content": ", saturation threshold " + }, + { + "bbox": [ + 305, + 109, + 547, + 168 + ], + "type": "inline_equation", + "content": "r" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 170, + 488, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 170, + 488, + 181 + ], + "spans": [ + { + "bbox": [ + 306, + 170, + 488, + 181 + ], + "type": "text", + "content": "Output: universal adversarial perturbation " + }, + { + "bbox": [ + 306, + 170, + 488, + 181 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 183, + 545, + 302 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 312, + 183, + 481, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 183, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 183, + 481, + 194 + ], + "type": "text", + "content": "1: Initialize " + }, + { + "bbox": [ + 312, + 183, + 481, + 194 + ], + "type": "inline_equation", + "content": "\\pmb{v}_0\\sim \\mathcal{U}(-\\epsilon ,\\epsilon)" + }, + { + "bbox": [ + 312, + 183, + 481, + 194 + ], + "type": "inline_equation", + "content": "t = 0,F = 0" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 196, + 454, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 196, + 454, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 196, + 454, + 207 + ], + "type": "text", + "content": "2: while " + }, + { + "bbox": [ + 312, + 196, + 454, + 207 + ], + "type": "inline_equation", + "content": "t < T" + }, + { + "bbox": [ + 312, + 196, + 454, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 196, + 454, + 207 + ], + "type": "inline_equation", + "content": "F < F_{\\text{max}}" + }, + { + "bbox": [ + 312, + 196, + 454, + 207 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 209, + 373, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 209, + 373, + 217 + ], + "spans": [ + { + "bbox": [ + 312, + 209, + 373, + 217 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 312, + 209, + 373, + 217 + ], + "type": "inline_equation", + "content": "t = t + 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 219, + 522, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 219, + 522, + 230 + ], + "spans": [ + { + "bbox": [ + 312, + 219, + 522, + 230 + ], + "type": "text", + "content": "4: Generate the artificial image set " + }, + { + "bbox": [ + 312, + 219, + 522, + 230 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 312, + 219, + 522, + 230 + ], + "type": "text", + "content": " via Eq. (8)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 232, + 539, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 232, + 539, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 232, + 539, + 243 + ], + "type": "text", + "content": "5: Compute the gradient " + }, + { + "bbox": [ + 312, + 232, + 539, + 243 + ], + "type": "inline_equation", + "content": "\\nabla \\mathcal{L}_t" + }, + { + "bbox": [ + 312, + 232, + 539, + 243 + ], + "type": "text", + "content": " of the loss " + }, + { + "bbox": [ + 312, + 232, + 539, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 312, + 232, + 539, + 243 + ], + "type": "text", + "content": " in Eq. (9)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 243, + 452, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 243, + 452, + 255 + ], + "spans": [ + { + "bbox": [ + 312, + 243, + 452, + 255 + ], + "type": "text", + "content": "6: Update " + }, + { + "bbox": [ + 312, + 243, + 452, + 255 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t = \\pmb{v}_{t-1} + \\eta \\cdot \\nabla \\mathcal{L}_t" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 256, + 462, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 256, + 462, + 267 + ], + "spans": [ + { + "bbox": [ + 312, + 256, + 462, + 267 + ], + "type": "text", + "content": "7: Clip " + }, + { + "bbox": [ + 312, + 256, + 462, + 267 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t = \\min (\\epsilon ,\\max (\\pmb {v}_t, - \\epsilon))" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 268, + 542, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 268, + 542, + 278 + ], + "spans": [ + { + "bbox": [ + 312, + 268, + 542, + 278 + ], + "type": "text", + "content": "8: Compute the saturation rate " + }, + { + "bbox": [ + 312, + 268, + 542, + 278 + ], + "type": "inline_equation", + "content": "\\hat{r}" + }, + { + "bbox": [ + 312, + 268, + 542, + 278 + ], + "type": "text", + "content": " and adjust " + }, + { + "bbox": [ + 312, + 268, + 542, + 278 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t" + }, + { + "bbox": [ + 312, + 268, + 542, + 278 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 312, + 268, + 542, + 278 + ], + "type": "inline_equation", + "content": "\\hat{r} < r" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 279, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 279, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 312, + 279, + 545, + 302 + ], + "type": "text", + "content": "9: Conduct the fooling rate test " + }, + { + "bbox": [ + 312, + 279, + 545, + 302 + ], + "type": "inline_equation", + "content": "FR" + }, + { + "bbox": [ + 312, + 279, + 545, + 302 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 312, + 279, + 545, + 302 + ], + "type": "inline_equation", + "content": "t\\% H == 0" + }, + { + "bbox": [ + 312, + 279, + 545, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 279, + 545, + 302 + ], + "type": "inline_equation", + "content": "F = F + 1" + }, + { + "bbox": [ + 312, + 279, + 545, + 302 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 312, + 279, + 545, + 302 + ], + "type": "inline_equation", + "content": "FR" + }, + { + "bbox": [ + 312, + 279, + 545, + 302 + ], + "type": "text", + "content": " not the best fooling rate" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 309, + 304, + 367, + 326 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 309, + 304, + 367, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 304, + 367, + 313 + ], + "spans": [ + { + "bbox": [ + 309, + 304, + 367, + 313 + ], + "type": "text", + "content": "10: end while" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 309, + 316, + 365, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 316, + 365, + 326 + ], + "spans": [ + { + "bbox": [ + 309, + 316, + 365, + 326 + ], + "type": "text", + "content": "11: return " + }, + { + "bbox": [ + 309, + 316, + 365, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 352, + 387, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 352, + 387, + 366 + ], + "spans": [ + { + "bbox": [ + 306, + 352, + 387, + 366 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 373, + 545, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 446 + ], + "type": "text", + "content": "Setup: By following the experiment setup in existing data-free universal methods [18, 17, 20, 14], we evaluated the proposed TRM-UAP method on the validation set of ImageNet [23] with the classical pre-trained CNN models including AlexNet [13], VGG16 [24], VGG19 [24], ResNet152 [10] and GoogleNet [25]." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 447, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 519 + ], + "type": "text", + "content": "Evaluation Criteria: We used the fooling rate " + }, + { + "bbox": [ + 304, + 447, + 545, + 519 + ], + "type": "inline_equation", + "content": "(FR)" + }, + { + "bbox": [ + 304, + 447, + 545, + 519 + ], + "type": "text", + "content": " proposed by data-free universal methods [18, 17] as the evaluation metric. Particularly, a higher fooling rate represented higher attack success rate and better transferability. In addition, we used the logit loss of C&W attack [5] to further evaluate the transferability of data-free UAPs." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 521, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 592 + ], + "type": "text", + "content": "Comparative Methods: We made a comparison with data-free universal approaches, including FFF [18], AAA [20], GD-UAP [17], PD-UA [14] and Cosine-UAP [33]. Note that GD-UAP [17] trained models under different setups, for a fair comparison, we reproduced the results of GD-UAP [17] by making the setup consistent with TRM-UAP." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": "Implementation Details: All of our experiments were implemented on PyTorch with a single NVIDIA GeForce RTX 3090Ti GPU. Following the common setting [18, 17], we set " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon = 10 / 255" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": " to restrict changes of the perturbation. The maximum iteration " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": " was set as 10000, and the saturation threshold " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": " was set to " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "0.001\\%" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": ". The value range of hyperparameters " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "PTR, NTR \\in [0, 1]" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": " denotes the truncation rate of positive and negative activations respectively. We set the scaling hyperparameter " + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": " and truncation rates with appropriate values for different CNN models." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4766" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 81, + 70, + 511, + 166 + ], + "blocks": [ + { + "bbox": [ + 81, + 70, + 511, + 166 + ], + "lines": [ + { + "bbox": [ + 81, + 70, + 511, + 166 + ], + "spans": [ + { + "bbox": [ + 81, + 70, + 511, + 166 + ], + "type": "table", + "html": "
AttackAlexNetVGG16VGG19ResNet152GoogleNetAverage
FFF80.9247.1043.62-56.44-
AAA89.0471.5972.8460.7275.2873.89
GD-UAP85.2490.0187.3445.9645.8764.65
PD-UA-70.6964.9846.3967.12-
Cosine-UAP91.0789.4886.8165.3587.5784.08
TRM-UAP(Ours)93.53±0.0794.30±0.1591.35±0.3067.46±0.3585.32±0.0486.39
", + "image_path": "74e3c5defd1bd3781f259e0fecfb7781e8fdada96f5ae0ddc1dac94346a44370.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 117, + 203, + 476, + 294 + ], + "blocks": [ + { + "bbox": [ + 46, + 168, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 547, + 194 + ], + "type": "text", + "content": "Table 1. The fooling rate " + }, + { + "bbox": [ + 46, + 168, + 547, + 194 + ], + "type": "inline_equation", + "content": "(FR)" + }, + { + "bbox": [ + 46, + 168, + 547, + 194 + ], + "type": "text", + "content": " of our proposed TRM-UAP and other data-free universal methods. We show the mean and standard deviation of " + }, + { + "bbox": [ + 46, + 168, + 547, + 194 + ], + "type": "inline_equation", + "content": "FR" + }, + { + "bbox": [ + 46, + 168, + 547, + 194 + ], + "type": "text", + "content": " with five runs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 117, + 203, + 476, + 294 + ], + "lines": [ + { + "bbox": [ + 117, + 203, + 476, + 294 + ], + "spans": [ + { + "bbox": [ + 117, + 203, + 476, + 294 + ], + "type": "table", + "html": "
AlexNetVGG16VGG19ResNet152GoogleNet
AlexNet93.53±0.0760.10±0.2457.08±0.1527.31±0.3032.70±0.22
VGG1647.53±0.5194.30±0.1289.68±0.1461.43±0.4053.95±0.59
VGG1946.01±0.4489.82±0.1591.35±0.3047.19±0.6646.48±0.78
ResNet15253.56±0.7577.20±0.3573.30±0.4167.46±0.3557.54±0.50
GoogleNet60.10±1.1679.66±0.9579.98±1.0658.85±1.9485.32±0.04
", + "image_path": "48080ed138239a037176c0698ad4f11197de1d091466e7fd54eb6269d9dc5b19.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 296, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 296, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 547, + 346 + ], + "type": "text", + "content": "Table 2. The transferability of the perturbation crafted by our TRM-UAP method (rows: target models; columns: surrogate models). The fooling rate " + }, + { + "bbox": [ + 46, + 296, + 547, + 346 + ], + "type": "inline_equation", + "content": "(FR)" + }, + { + "bbox": [ + 46, + 296, + 547, + 346 + ], + "type": "text", + "content": " on the diagonal represents white-box attacks (i.e., UAP is crafted by the target model itself), and other off-diagonal values are black-box attacks (i.e., UAP is crafted by non-target models.). All results of mean and standard deviation are computed on average with five turns." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 365, + 133, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 133, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 133, + 376 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "text", + "content": "Our proposed TRM-UAP is applied to craft UAPs on five CNN models and create AEs to attack these CNN models separately on the ImageNet validation set. We set " + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\{1.0, 1.0, 0.5, 1.5, 1.3\\}" + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "inline_equation", + "content": "PTR \\in \\{0.0, 0.2, 0.2, 0.7, 0.4\\}" + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "inline_equation", + "content": "NTR \\in \\{0.8, 0.3, 0.2, 0.8, 0.3\\}" + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "text", + "content": ", corresponding to AlexNet, VGG16, VGG19, ResNet152, GoogleNet. The attack performance compared with other data-free universal methods is shown in Table 1. Note that the results of GD-UAP were reproduced by our best effort on PyTorch for a fair comparison. As compared to other methods, our TRM-UAP method improves the fooling rate in most CNN models. Although Cosine-UAP achieves a higher " + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "inline_equation", + "content": "FR" + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "text", + "content": " in GoogleNet, TRM-UAP improves the " + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "inline_equation", + "content": "FR" + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "text", + "content": " around " + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "inline_equation", + "content": "1 \\sim 5\\%" + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "text", + "content": " on other four models and further improves the average " + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "inline_equation", + "content": "FR" + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "text", + "content": " of all five models more than " + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 386, + 289, + 615 + ], + "type": "text", + "content": ". Besides, it is shown in Fig. 4 that the visualization of crafted UAPs has clearly particular image patterns with abundant features. We also visualized AEs of TRM-UAP attack in Fig. 6 and found that high probabilities of AEs were given to the labels of incorrect class." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "We further used the UAP made by surrogate models to attack other target models to verify the transferability. Table 2 represented the attack performance in different attack settings. The models in rows denote target models, and the models in columns denote surrogate models for crafting the UAP. The results on the diagonal denote the white-box attack, while the rest are transferable attacks in the black-box setting. Most TRM-UAP attacks perform excellently in" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 365, + 547, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 401 + ], + "type": "text", + "content": "transferability, which indicates that the universal perturbations crafted by TRM-UAP are generalized well to transfer attacks on target models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 407, + 471, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 471, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 471, + 420 + ], + "type": "text", + "content": "4.2. Evaluating the Transferability" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 426, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 641 + ], + "type": "text", + "content": "To further compare the effectiveness of universal attacks, we used the logit loss of C&W attack [5] to evaluate the transferability of GD-UAP, Cosine-UAP, and TRM-UAP. The logit loss is defined as " + }, + { + "bbox": [ + 304, + 426, + 547, + 641 + ], + "type": "inline_equation", + "content": "loss = (\\max_{j \\neq t} F(\\boldsymbol{v} + \\boldsymbol{x})_j) - F(\\boldsymbol{v} + \\boldsymbol{x})_t" + }, + { + "bbox": [ + 304, + 426, + 547, + 641 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 426, + 547, + 641 + ], + "type": "inline_equation", + "content": "F(\\cdot)_j" + }, + { + "bbox": [ + 304, + 426, + 547, + 641 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 426, + 547, + 641 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 426, + 547, + 641 + ], + "type": "text", + "content": "-th output in the logits layer of network and " + }, + { + "bbox": [ + 304, + 426, + 547, + 641 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 426, + 547, + 641 + ], + "type": "text", + "content": " is the index of ground truth. A larger logit loss value represents better transferability of attacks to CNN models. The logit loss of original images is calculated as a baseline. Note that the results of GD-UAP and Cosine-UAP are reproduced with our best effort on PyTorch. As shown in Fig. 5, TRM-UAP achieves a higher logit loss than GD-UAP and Cosine-UAP in most cases. Additionally, the discrepancy among GD-UAP, Cosine-UAP, and TRM-UAP becomes significant in deep CNN models (e.g., ResNet152 and GoogleNet). Therefore, TRM-UAP achieves relatively higher logit losses in both white-box and black-box attack settings, which indicates a better transferability across various CNN models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 647, + 410, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 647, + 410, + 661 + ], + "spans": [ + { + "bbox": [ + 305, + 647, + 410, + 661 + ], + "type": "text", + "content": "4.3. Parameter Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "First, the influence of ratio maximization on attack performance is studied. Here we mainly focused on validating the effectiveness of truncated negative activation in deep CNN models (e.g., ResNet152 and GoogleNet). The experi" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "4767" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 137, + 153 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 137, + 153 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 137, + 153 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 137, + 153 + ], + "type": "image", + "image_path": "65c9d802a59f4dcb81e3bb2f629100341f6cd3fcdc0b714697823fb1fac9f3fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 157, + 114, + 166 + ], + "lines": [ + { + "bbox": [ + 75, + 157, + 114, + 166 + ], + "spans": [ + { + "bbox": [ + 75, + 157, + 114, + 166 + ], + "type": "text", + "content": "(a) AlexNet" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 154, + 72, + 238, + 154 + ], + "blocks": [ + { + "bbox": [ + 154, + 72, + 238, + 154 + ], + "lines": [ + { + "bbox": [ + 154, + 72, + 238, + 154 + ], + "spans": [ + { + "bbox": [ + 154, + 72, + 238, + 154 + ], + "type": "image", + "image_path": "8b3d3196e40adb2a8c67e651bcb9c67e61ae87efae8b7629b9d7eebc61232f8a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 156, + 215, + 166 + ], + "lines": [ + { + "bbox": [ + 176, + 156, + 215, + 166 + ], + "spans": [ + { + "bbox": [ + 176, + 156, + 215, + 166 + ], + "type": "text", + "content": "(b) VGG16" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 256, + 72, + 339, + 154 + ], + "blocks": [ + { + "bbox": [ + 256, + 72, + 339, + 154 + ], + "lines": [ + { + "bbox": [ + 256, + 72, + 339, + 154 + ], + "spans": [ + { + "bbox": [ + 256, + 72, + 339, + 154 + ], + "type": "image", + "image_path": "4d0f65dd6835d1346f5582a0f4930288b9eedb01e701bb03ed2c2121bcc69f9b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 156, + 315, + 166 + ], + "lines": [ + { + "bbox": [ + 277, + 156, + 315, + 166 + ], + "spans": [ + { + "bbox": [ + 277, + 156, + 315, + 166 + ], + "type": "text", + "content": "(c) VGG19" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 358, + 72, + 440, + 154 + ], + "blocks": [ + { + "bbox": [ + 358, + 72, + 440, + 154 + ], + "lines": [ + { + "bbox": [ + 358, + 72, + 440, + 154 + ], + "spans": [ + { + "bbox": [ + 358, + 72, + 440, + 154 + ], + "type": "image", + "image_path": "1c573be2aaf3325f72df9cf636b44987fe6db48ffb6c2e2550e3c8b93bebc8bc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 156, + 422, + 166 + ], + "lines": [ + { + "bbox": [ + 373, + 156, + 422, + 166 + ], + "spans": [ + { + "bbox": [ + 373, + 156, + 422, + 166 + ], + "type": "text", + "content": "(d) ResNet152" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 458, + 71, + 541, + 153 + ], + "blocks": [ + { + "bbox": [ + 458, + 71, + 541, + 153 + ], + "lines": [ + { + "bbox": [ + 458, + 71, + 541, + 153 + ], + "spans": [ + { + "bbox": [ + 458, + 71, + 541, + 153 + ], + "type": "image", + "image_path": "07d9ff150d5f58d2a1fba9577c128d092d352195bb49c0301f60596993b9fa6a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 156, + 523, + 166 + ], + "lines": [ + { + "bbox": [ + 475, + 156, + 523, + 166 + ], + "spans": [ + { + "bbox": [ + 475, + 156, + 523, + 166 + ], + "type": "text", + "content": "(e) GoogleNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 51, + 208, + 148, + 266 + ], + "blocks": [ + { + "bbox": [ + 46, + 169, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 169, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 545, + 194 + ], + "type": "text", + "content": "Figure 4. The visualization of UAPs crafted by our TRM-UAP method. For the best visualization in color, we enhance the value of the pixel to [0, 255]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 208, + 148, + 266 + ], + "lines": [ + { + "bbox": [ + 51, + 208, + 148, + 266 + ], + "spans": [ + { + "bbox": [ + 51, + 208, + 148, + 266 + ], + "type": "image", + "image_path": "d9840ca6ae6a1cec36052939a5a34a5c39b25a980b35c684409de51b5abd6c74.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 269, + 119, + 278 + ], + "lines": [ + { + "bbox": [ + 79, + 269, + 119, + 278 + ], + "spans": [ + { + "bbox": [ + 79, + 269, + 119, + 278 + ], + "type": "text", + "content": "(a) AlexNet" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 149, + 208, + 248, + 266 + ], + "blocks": [ + { + "bbox": [ + 149, + 208, + 248, + 266 + ], + "lines": [ + { + "bbox": [ + 149, + 208, + 248, + 266 + ], + "spans": [ + { + "bbox": [ + 149, + 208, + 248, + 266 + ], + "type": "image", + "image_path": "653907a0c232758cb095a21ffe5995ca6c5497a2b86518feee531c915091ac34.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 269, + 217, + 278 + ], + "lines": [ + { + "bbox": [ + 179, + 269, + 217, + 278 + ], + "spans": [ + { + "bbox": [ + 179, + 269, + 217, + 278 + ], + "type": "text", + "content": "(b) VGG16" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 250, + 208, + 349, + 266 + ], + "blocks": [ + { + "bbox": [ + 250, + 208, + 349, + 266 + ], + "lines": [ + { + "bbox": [ + 250, + 208, + 349, + 266 + ], + "spans": [ + { + "bbox": [ + 250, + 208, + 349, + 266 + ], + "type": "image", + "image_path": "0f231656c84bd49878ea6b3134009c957cfc20c7b313d705c8a8718447b12cda.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 269, + 317, + 278 + ], + "lines": [ + { + "bbox": [ + 280, + 269, + 317, + 278 + ], + "spans": [ + { + "bbox": [ + 280, + 269, + 317, + 278 + ], + "type": "text", + "content": "(c) VGG19" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 351, + 208, + 446, + 266 + ], + "blocks": [ + { + "bbox": [ + 351, + 208, + 446, + 266 + ], + "lines": [ + { + "bbox": [ + 351, + 208, + 446, + 266 + ], + "spans": [ + { + "bbox": [ + 351, + 208, + 446, + 266 + ], + "type": "image", + "image_path": "f14feee229e1d871c0c9f9a4bf3510d048914bbc72da18f37fdb2a00fb66fa2b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 269, + 422, + 278 + ], + "lines": [ + { + "bbox": [ + 373, + 269, + 422, + 278 + ], + "spans": [ + { + "bbox": [ + 373, + 269, + 422, + 278 + ], + "type": "text", + "content": "(d) ResNet152" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 447, + 208, + 543, + 266 + ], + "blocks": [ + { + "bbox": [ + 447, + 208, + 543, + 266 + ], + "lines": [ + { + "bbox": [ + 447, + 208, + 543, + 266 + ], + "spans": [ + { + "bbox": [ + 447, + 208, + 543, + 266 + ], + "type": "image", + "image_path": "36beb49379502e532a7e458717f469680dd7428a62426c9cf841c1ad959c2396.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 269, + 518, + 278 + ], + "lines": [ + { + "bbox": [ + 471, + 269, + 518, + 278 + ], + "spans": [ + { + "bbox": [ + 471, + 269, + 518, + 278 + ], + "type": "text", + "content": "(e) GoogleNet" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 53, + 342, + 107, + 396 + ], + "blocks": [ + { + "bbox": [ + 46, + 281, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 545, + 329 + ], + "type": "text", + "content": "Figure 5. The logit loss of original samples and adversarial samples crafted by GD-UAP, Cosine-UAP, our TRM-UAP methods. In each sub-figure, the subtitle indicates the surrogate model for crafting UAPs, and target models are shown on the horizontal axis. Original samples are classified correctly by the target model and exploited to craft AEs for comparisons. For each attack setting, we compute the average of the logit loss over all the samples." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 342, + 107, + 396 + ], + "lines": [ + { + "bbox": [ + 53, + 342, + 107, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 107, + 396 + ], + "type": "image", + "image_path": "2d200e91fbd33143f9c0679bf72f2adaf5a50e390a19fae2b177fe7024a2f0c6.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 396, + 106, + 400 + ], + "lines": [ + { + "bbox": [ + 55, + 396, + 106, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 106, + 400 + ], + "type": "text", + "content": "bullet train(98.15%)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 111, + 343, + 164, + 396 + ], + "blocks": [ + { + "bbox": [ + 111, + 343, + 164, + 396 + ], + "lines": [ + { + "bbox": [ + 111, + 343, + 164, + 396 + ], + "spans": [ + { + "bbox": [ + 111, + 343, + 164, + 396 + ], + "type": "image", + "image_path": "0a432cc4dffc8d6ca71a39e3cb5fe1820ebee297308efd5c4adfbacffc937c49.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 396, + 157, + 400 + ], + "lines": [ + { + "bbox": [ + 119, + 396, + 157, + 400 + ], + "spans": [ + { + "bbox": [ + 119, + 396, + 157, + 400 + ], + "type": "text", + "content": "puffer(99.89%)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 54, + 400, + 107, + 452 + ], + "blocks": [ + { + "bbox": [ + 54, + 400, + 107, + 452 + ], + "lines": [ + { + "bbox": [ + 54, + 400, + 107, + 452 + ], + "spans": [ + { + "bbox": [ + 54, + 400, + 107, + 452 + ], + "type": "image", + "image_path": "d0494d384130a691d241b67c8c68102120a3dfff002e4c448a3a996c3597b8a7.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 451, + 97, + 456 + ], + "lines": [ + { + "bbox": [ + 64, + 451, + 97, + 456 + ], + "spans": [ + { + "bbox": [ + 64, + 451, + 97, + 456 + ], + "type": "text", + "content": "quilt(45.63%)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "type": "text", + "content": "Figure 6. Examples of our TRM-UAP attack tested on VGG16 and the perturbation crafted by VGG19 (top: original examples; bottom: adversarial examples). The annotation represents the classification label and the probability predicted by CNNs." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 111, + 400, + 166, + 452 + ], + "blocks": [ + { + "bbox": [ + 111, + 400, + 166, + 452 + ], + "lines": [ + { + "bbox": [ + 111, + 400, + 166, + 452 + ], + "spans": [ + { + "bbox": [ + 111, + 400, + 166, + 452 + ], + "type": "image", + "image_path": "d0a84fbccd57c6376f0c62579ad8e800087dc0eafca838fd508e726eadfeb3b6.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 452, + 167, + 456 + ], + "lines": [ + { + "bbox": [ + 108, + 452, + 167, + 456 + ], + "spans": [ + { + "bbox": [ + 108, + 452, + 167, + 456 + ], + "type": "text", + "content": "theater curtain(72.82%)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 168, + 343, + 221, + 396 + ], + "blocks": [ + { + "bbox": [ + 168, + 343, + 221, + 396 + ], + "lines": [ + { + "bbox": [ + 168, + 343, + 221, + 396 + ], + "spans": [ + { + "bbox": [ + 168, + 343, + 221, + 396 + ], + "type": "image", + "image_path": "2229b9cbff59ff4b52fb8a35ccf0888fea7ff1ec65ab48d6a669fe35af48b83c.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 396, + 220, + 400 + ], + "lines": [ + { + "bbox": [ + 169, + 396, + 220, + 400 + ], + "spans": [ + { + "bbox": [ + 169, + 396, + 220, + 400 + ], + "type": "text", + "content": "monastery(34.69%)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 168, + 400, + 221, + 452 + ], + "blocks": [ + { + "bbox": [ + 168, + 400, + 221, + 452 + ], + "lines": [ + { + "bbox": [ + 168, + 400, + 221, + 452 + ], + "spans": [ + { + "bbox": [ + 168, + 400, + 221, + 452 + ], + "type": "image", + "image_path": "06e6dafd85668f5dc017d538c8984867aed7a23d311e63f3c5ad50d920fa6f07.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 452, + 215, + 456 + ], + "lines": [ + { + "bbox": [ + 175, + 452, + 215, + 456 + ], + "spans": [ + { + "bbox": [ + 175, + 452, + 215, + 456 + ], + "type": "text", + "content": "church(43.55%)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 225, + 343, + 278, + 396 + ], + "blocks": [ + { + "bbox": [ + 225, + 343, + 278, + 396 + ], + "lines": [ + { + "bbox": [ + 225, + 343, + 278, + 396 + ], + "spans": [ + { + "bbox": [ + 225, + 343, + 278, + 396 + ], + "type": "image", + "image_path": "80fb138084f9b9a0ba6f8d92ce61ed8391114c0a95a22fd87cd9e3195cda1559.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 396, + 268, + 400 + ], + "lines": [ + { + "bbox": [ + 235, + 396, + 268, + 400 + ], + "spans": [ + { + "bbox": [ + 235, + 396, + 268, + 400 + ], + "type": "text", + "content": "shoji(87.62%)" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 225, + 400, + 280, + 452 + ], + "blocks": [ + { + "bbox": [ + 225, + 400, + 280, + 452 + ], + "lines": [ + { + "bbox": [ + 225, + 400, + 280, + 452 + ], + "spans": [ + { + "bbox": [ + 225, + 400, + 280, + 452 + ], + "type": "image", + "image_path": "fea73f939316b40e8544950656a81ac387fca478f012d0152d5a5d38dd1aa6c3.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 452, + 282, + 456 + ], + "lines": [ + { + "bbox": [ + 222, + 452, + 282, + 456 + ], + "spans": [ + { + "bbox": [ + 222, + 452, + 282, + 456 + ], + "type": "text", + "content": "shower curtain(94.35%)" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": "imental setting is the same as Section 4.1. Based on this, we also crafted UAPs by truncated positive activation only, i.e., " + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "inline_equation", + "content": "PTR" + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": " remains unchanged and " + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "inline_equation", + "content": "NTR = 100\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": ". As shown in Fig. 7, the ablation experiment indicates that truncated negative activation combined with truncated positive activation improves the " + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "inline_equation", + "content": "FR" + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": " by a large margin (around " + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "inline_equation", + "content": "8 \\sim 18\\%" + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": ") as compared to standalone truncated positive activation, thus enhancing both intensity and transferability of UAPs significantly." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "To explore the influence of truncation strategy, we computed the fooling rate of attacks in CNN models by varying truncation rates. Specifically, the proportion of retained activation increased by " + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": " for each step. That means UAP learned the feature information from shallow to deep convolution layers gradually. Due to space limitations, UAPs" + } + ] + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 313, + 352, + 425, + 428 + ], + "blocks": [ + { + "bbox": [ + 313, + 352, + 425, + 428 + ], + "lines": [ + { + "bbox": [ + 313, + 352, + 425, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 425, + 428 + ], + "type": "image", + "image_path": "54360931ab4da5e24809ad0f987e60634e820290f30d4fddee5b2fc4c6a85944.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 430, + 392, + 440 + ], + "lines": [ + { + "bbox": [ + 345, + 430, + 392, + 440 + ], + "spans": [ + { + "bbox": [ + 345, + 430, + 392, + 440 + ], + "type": "text", + "content": "(a) ResNet152" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 443, + 545, + 491 + ], + "lines": [ + { + "bbox": [ + 305, + 443, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 443, + 545, + 491 + ], + "type": "text", + "content": "Figure 7. The ablation study of truncated negative activations on CNN models. Positive: the attack only uses the truncated positive activation. Positive&Negative: the attack integrates truncated positive and negative activations." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 428, + 352, + 541, + 428 + ], + "blocks": [ + { + "bbox": [ + 428, + 352, + 541, + 428 + ], + "lines": [ + { + "bbox": [ + 428, + 352, + 541, + 428 + ], + "spans": [ + { + "bbox": [ + 428, + 352, + 541, + 428 + ], + "type": "image", + "image_path": "4a0325d7a358b602399802ea52d8d43043b08f33eb2521616b840fe45c56fe47.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 430, + 507, + 441 + ], + "lines": [ + { + "bbox": [ + 459, + 430, + 507, + 441 + ], + "spans": [ + { + "bbox": [ + 459, + 430, + 507, + 441 + ], + "type": "text", + "content": "(b) GoogleNet" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "bbox": [ + 304, + 506, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 625 + ], + "type": "text", + "content": "were only tested in white-box settings. The results of all CNN models were presented in Fig. 8, where the parameter space was explored w.r.t. NTR and PTR. As can be seen, the optimal result was obtained on the attack setting of the truncated activation rather than the whole activation. Especially in deep CNN models, only part of positive and negative activations were retained to craft UAPs. From the analysis of truncation in Fig. 8, we observe that the shallow convolution layer extracts generic features which are better at transferring the attack across different models." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 306, + 634, + 474, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 474, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 474, + 647 + ], + "type": "text", + "content": "4.4. Exploring the Property of UAP" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "In the following, the property of universal perturbations is thoroughly investigated. We studied the distinction between original example and adversarial examples made by image-specific attack (MI-FGSM) and data-free universal attacks (GD-UAP and TRM-UAP). As shown in Fig. 9, AEs" + } + ] + } + ], + "index": 48 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4768" + } + ] + } + ], + "index": 49 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 75, + 137, + 137 + ], + "blocks": [ + { + "bbox": [ + 56, + 75, + 137, + 137 + ], + "lines": [ + { + "bbox": [ + 56, + 75, + 137, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 75, + 137, + 137 + ], + "type": "image", + "image_path": "572da65a74803e16285449d42a3de05c833fced37a732be51e6527405adca038.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 143, + 115, + 152 + ], + "lines": [ + { + "bbox": [ + 76, + 143, + 115, + 152 + ], + "spans": [ + { + "bbox": [ + 76, + 143, + 115, + 152 + ], + "type": "text", + "content": "(a) AlexNet" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 157, + 75, + 237, + 137 + ], + "blocks": [ + { + "bbox": [ + 157, + 75, + 237, + 137 + ], + "lines": [ + { + "bbox": [ + 157, + 75, + 237, + 137 + ], + "spans": [ + { + "bbox": [ + 157, + 75, + 237, + 137 + ], + "type": "image", + "image_path": "f9584c2a5a9435aa9c1cfe179bf4704a78f252e1f5274770805964feb08086e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 143, + 215, + 152 + ], + "lines": [ + { + "bbox": [ + 178, + 143, + 215, + 152 + ], + "spans": [ + { + "bbox": [ + 178, + 143, + 215, + 152 + ], + "type": "text", + "content": "(b) VGG16" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 258, + 73, + 338, + 136 + ], + "blocks": [ + { + "bbox": [ + 258, + 73, + 338, + 136 + ], + "lines": [ + { + "bbox": [ + 258, + 73, + 338, + 136 + ], + "spans": [ + { + "bbox": [ + 258, + 73, + 338, + 136 + ], + "type": "image", + "image_path": "87e3bfd800fe78c910306a2e98dce874140db9cd2cc92ade1b8e808d20971ab2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 143, + 315, + 152 + ], + "lines": [ + { + "bbox": [ + 278, + 143, + 315, + 152 + ], + "spans": [ + { + "bbox": [ + 278, + 143, + 315, + 152 + ], + "type": "text", + "content": "(c) VGG19" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 358, + 75, + 438, + 137 + ], + "blocks": [ + { + "bbox": [ + 358, + 75, + 438, + 137 + ], + "lines": [ + { + "bbox": [ + 358, + 75, + 438, + 137 + ], + "spans": [ + { + "bbox": [ + 358, + 75, + 438, + 137 + ], + "type": "image", + "image_path": "b391c6653a3ce10e43f466c0221f525adc1ebc0ac56fce94b04cc371894a638f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 143, + 421, + 152 + ], + "lines": [ + { + "bbox": [ + 373, + 143, + 421, + 152 + ], + "spans": [ + { + "bbox": [ + 373, + 143, + 421, + 152 + ], + "type": "text", + "content": "(d) ResNet152" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 459, + 75, + 538, + 137 + ], + "blocks": [ + { + "bbox": [ + 459, + 75, + 538, + 137 + ], + "lines": [ + { + "bbox": [ + 459, + 75, + 538, + 137 + ], + "spans": [ + { + "bbox": [ + 459, + 75, + 538, + 137 + ], + "type": "image", + "image_path": "920e604745f5203df0ea6803dfda4ff9d12dedefe1c86835061fe8dbfe052133.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 143, + 522, + 152 + ], + "lines": [ + { + "bbox": [ + 474, + 143, + 522, + 152 + ], + "spans": [ + { + "bbox": [ + 474, + 143, + 522, + 152 + ], + "type": "text", + "content": "(e) GoogleNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 52, + 204, + 105, + 264 + ], + "blocks": [ + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "lines": [ + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "text", + "content": "Figure 8. The parameter study of truncation rates on CNN models in white-box attack setting. The " + }, + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "text", + "content": " axis are NTR, PTR, and FR, respectively. Note that there are only five convolution layers in AlexNet, truncation results (e.g., " + }, + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "text", + "content": ", etc.) are omitted since these values are consistent with the ones at their right coordinate." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 204, + 105, + 264 + ], + "lines": [ + { + "bbox": [ + 52, + 204, + 105, + 264 + ], + "spans": [ + { + "bbox": [ + 52, + 204, + 105, + 264 + ], + "type": "image", + "image_path": "a4b5e47d7ded467512cf5dec0c5bfe14807c8f647eed16c01d4b176527ca6108.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 111, + 205, + 164, + 264 + ], + "blocks": [ + { + "bbox": [ + 111, + 205, + 164, + 264 + ], + "lines": [ + { + "bbox": [ + 111, + 205, + 164, + 264 + ], + "spans": [ + { + "bbox": [ + 111, + 205, + 164, + 264 + ], + "type": "image", + "image_path": "a6e036f3f9eece8a241eb2c3235ead53938aeb250022b273360956be79176ce6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 170, + 205, + 223, + 264 + ], + "blocks": [ + { + "bbox": [ + 170, + 205, + 223, + 264 + ], + "lines": [ + { + "bbox": [ + 170, + 205, + 223, + 264 + ], + "spans": [ + { + "bbox": [ + 170, + 205, + 223, + 264 + ], + "type": "image", + "image_path": "891ea19d7f31bcdd2f5b5590e25f2f8fc377d0e4731f741e6638a74733302f89.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 228, + 205, + 282, + 264 + ], + "blocks": [ + { + "bbox": [ + 228, + 205, + 282, + 264 + ], + "lines": [ + { + "bbox": [ + 228, + 205, + 282, + 264 + ], + "spans": [ + { + "bbox": [ + 228, + 205, + 282, + 264 + ], + "type": "image", + "image_path": "31dc80df69a5682a54422bbd9b097440f2f58e52df4477e33cb6207ce04b108e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 52, + 270, + 105, + 323 + ], + "blocks": [ + { + "bbox": [ + 52, + 270, + 105, + 323 + ], + "lines": [ + { + "bbox": [ + 52, + 270, + 105, + 323 + ], + "spans": [ + { + "bbox": [ + 52, + 270, + 105, + 323 + ], + "type": "image", + "image_path": "438c031f7e905448fb9b5d8ce6dd3a1b954c6bedb6ebdcb0f46d163df157fe45.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 325, + 286, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 325, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 286, + 373 + ], + "type": "text", + "content": "Figure 9. The distinction of feature maps among original image, image-specific attack, and universal attacks (top: feature maps extracted from a shallow convolution layer; bottom: original example and AEs made by VGG19)." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 111, + 270, + 164, + 323 + ], + "blocks": [ + { + "bbox": [ + 111, + 270, + 164, + 323 + ], + "lines": [ + { + "bbox": [ + 111, + 270, + 164, + 323 + ], + "spans": [ + { + "bbox": [ + 111, + 270, + 164, + 323 + ], + "type": "image", + "image_path": "5bc5ff70594099e09d7459d6747fc33c69384834846216e9a5078d68df1bbf0c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 170, + 270, + 223, + 323 + ], + "blocks": [ + { + "bbox": [ + 170, + 270, + 223, + 323 + ], + "lines": [ + { + "bbox": [ + 170, + 270, + 223, + 323 + ], + "spans": [ + { + "bbox": [ + 170, + 270, + 223, + 323 + ], + "type": "image", + "image_path": "bba6eb8b5f34e8f463ebbf59b2e27b1cf6ad621434c1670bf79a0dc52faf5d93.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 229, + 270, + 282, + 323 + ], + "blocks": [ + { + "bbox": [ + 229, + 270, + 282, + 323 + ], + "lines": [ + { + "bbox": [ + 229, + 270, + 282, + 323 + ], + "spans": [ + { + "bbox": [ + 229, + 270, + 282, + 323 + ], + "type": "image", + "image_path": "34635c586aa6340dfa8f236ea295eec759a14ab0f58e24b2467881216340d752.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 389, + 286, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 286, + 509 + ], + "type": "text", + "content": "are indistinguishable compared with original example, and the imperceptible noise added to the input effectively perturbed feature maps in a shallow convolution layer. Furthermore, data-free universal attacks disturbed the feature map seriously in contrast to the image-specific attack. Nevertheless, the feature map of TRM-UAP had better regularity and repeatability than GD-UAP. We can conclude that features extracted by shallow convolution layers profoundly impact the final classification of CNN models and are beneficial for universal attack to craft the transferable perturbation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 510, + 286, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 286, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 286, + 654 + ], + "type": "text", + "content": "To further verify the above viewpoints, we trained UAPs in completely opposite truncation strategies (i.e., retaining former " + }, + { + "bbox": [ + 46, + 510, + 286, + 654 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 510, + 286, + 654 + ], + "type": "text", + "content": " layers versus retaining latter " + }, + { + "bbox": [ + 46, + 510, + 286, + 654 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 510, + 286, + 654 + ], + "type": "text", + "content": " layers), and counted the distribution of class labels of AEs. The result in Fig. 10 (a) was consistent with the hypothesis [33] that the top-1 logits prediction of UAP dominates the prediction of AEs. Both labels were theater curtain. Contrarily, it was shown in Fig. 10 (b) that the phenomenon of top-1 label domination disappeared, where the UAP label was obelisk and the AE label was brass. Thus, the shallow convolution layer is more beneficial to transfer universal attack with generic features through a variety of models." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 668, + 119, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 668, + 119, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 668, + 119, + 680 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "In this paper, we have proposed a novel data-free universal attack, called TRM-UAP, to reformulate the UAP gener-" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 318, + 209, + 427, + 278 + ], + "blocks": [ + { + "bbox": [ + 318, + 209, + 427, + 278 + ], + "lines": [ + { + "bbox": [ + 318, + 209, + 427, + 278 + ], + "spans": [ + { + "bbox": [ + 318, + 209, + 427, + 278 + ], + "type": "image", + "image_path": "01d53a6840e63a9d1b5bf12eaf66ab779888517bb3fbab19ae86c318d80a4959.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 285, + 526, + 295 + ], + "lines": [ + { + "bbox": [ + 342, + 285, + 526, + 295 + ], + "spans": [ + { + "bbox": [ + 342, + 285, + 526, + 295 + ], + "type": "text", + "content": "(a) UAP from low-level (former " + }, + { + "bbox": [ + 342, + 285, + 526, + 295 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 342, + 285, + 526, + 295 + ], + "type": "text", + "content": " ) convolution layers" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 427, + 209, + 538, + 278 + ], + "blocks": [ + { + "bbox": [ + 427, + 209, + 538, + 278 + ], + "lines": [ + { + "bbox": [ + 427, + 209, + 538, + 278 + ], + "spans": [ + { + "bbox": [ + 427, + 209, + 538, + 278 + ], + "type": "image", + "image_path": "c911b1a7b9b5b49dfa3f7a4e0c0296d9ed77dd75f12afc39f3dfbb359a82f5a5.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 317, + 306, + 427, + 376 + ], + "blocks": [ + { + "bbox": [ + 317, + 306, + 427, + 376 + ], + "lines": [ + { + "bbox": [ + 317, + 306, + 427, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 306, + 427, + 376 + ], + "type": "image", + "image_path": "7d129a42526e3fd9d856f53c514a5d25820b3343a1d11f7aa656c92dbfe0e2c5.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 380, + 517, + 389 + ], + "lines": [ + { + "bbox": [ + 335, + 380, + 517, + 389 + ], + "spans": [ + { + "bbox": [ + 335, + 380, + 517, + 389 + ], + "type": "text", + "content": "(b) UAP from high-level (latter " + }, + { + "bbox": [ + 335, + 380, + 517, + 389 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 335, + 380, + 517, + 389 + ], + "type": "text", + "content": " ) convolution layers" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 428, + 307, + 535, + 376 + ], + "blocks": [ + { + "bbox": [ + 428, + 307, + 535, + 376 + ], + "lines": [ + { + "bbox": [ + 428, + 307, + 535, + 376 + ], + "spans": [ + { + "bbox": [ + 428, + 307, + 535, + 376 + ], + "type": "image", + "image_path": "90d691a8ad3cbe797146d9aeffdfe746ba085acfa32aba42863f52dac3afeec5.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 392, + 545, + 440 + ], + "lines": [ + { + "bbox": [ + 305, + 392, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 545, + 440 + ], + "type": "text", + "content": "Figure 10. The analysis of the label distribution of UAPs and AEs (both made by VGG16). In each subfigure, the left is the top-10 logits value of UAP " + }, + { + "bbox": [ + 305, + 392, + 545, + 440 + ], + "type": "inline_equation", + "content": "(\\pmb{v})" + }, + { + "bbox": [ + 305, + 392, + 545, + 440 + ], + "type": "text", + "content": ", and the right is the distribution of top-10 predicated labels of AEs " + }, + { + "bbox": [ + 305, + 392, + 545, + 440 + ], + "type": "inline_equation", + "content": "(\\pmb{v} + \\pmb{x})" + }, + { + "bbox": [ + 305, + 392, + 545, + 440 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 626 + ], + "type": "text", + "content": "ation task as a truncated ratio optimization problem. Compared with previous methods, our TRM-UAP method integrates positive activation maximization with negative activation minimization. Towards further improving the transferability of UAPs, we propose a truncation strategy that computes positive and negative activation losses from low-level convolutions, as well as a curriculum optimization algorithm to fully mine the diversity of artificial images. Experimental results on the ImageNet dataset validate the better transferability of our TRM-UAP attack than other data-free UAP attacks on different CNN models. Additionally, UAPs learned from generic features in shallow convolution layers can dominate the model prediction and transfer the universal attack." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work is partially supported by the Scientific Research Foundation of Chongqing University of Technology under Grant No.2022ZDZ026, Natural Science Foundation of Chongqing, China under Grant No.CSTB2022NSCQ-MSX0493, the Key project of Chongqing Technology Innovation and Application Development under Grant No.cstc2021jscx-dxwtBX0018." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4769" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 6:14410-14430, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "type": "text", + "content": "[2] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41-48, 2009." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 169, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 169, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 169, + 288, + 235 + ], + "type": "text", + "content": "[3] Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Šrndić, Pavel Laskov, Giorgio Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 387-402. Springer, 2013." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 268 + ], + "type": "text", + "content": "[4] Battista Biggio and Fabio Roli. Wild patterns: Ten years after the rise of adversarial machine learning. Pattern Recognition, 84:317-331, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 288, + 302 + ], + "type": "text", + "content": "[5] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE Symposium on Security and Privacy, pages 39-57, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 288, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 288, + 345 + ], + "type": "text", + "content": "[6] Ashutosh Chaubey, Nikhil Agrawal, Kavya Barnwal, Keerat K Guliani, and Pramod Mehta. Universal adversarial perturbations: A survey. arXiv preprint arXiv:2005.08087, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 347, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 288, + 401 + ], + "type": "text", + "content": "[7] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9185-9193, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 445 + ], + "type": "text", + "content": "[8] Aditya Ganeshan, Vivek BS, and R Venkatesh Babu. Fda: Feature disruptive attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8069-8079, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 480 + ], + "type": "text", + "content": "[9] Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 481, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 524 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[11] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features. Advances in Neural Information Processing Systems, 32, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "text", + "content": "[12] Valentin Khrulkov and Ivan Oseledets. Art of singular vectors and universal adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8562-8570, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 656 + ], + "type": "text", + "content": "[13] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 2012." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[14] Hong Liu, Rongrong Ji, Jie Li, Baochang Zhang, Yue Gao, Yongjian Wu, and Feiyue Huang. Universal adversarial perturbation via prior driven uncertainty approximation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2941-2949, 2019." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[15] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[16] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1765-1773, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[17] Konda Reddy Mopuri, Aditya Ganeshan, and R Venkatesh Babu. Generalizable data-free objective for crafting universal adversarial perturbations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(10):2452-2465, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "text", + "content": "[18] Konda Reddy Mopuri, Utsav Garg, and R Venkatesh Babu. Fast feature fool: A data independent approach to universal adversarial perturbations. In Proceedings of the British Machine Vision Conference, 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "text", + "content": "[19] Konda Reddy Mopuri, Utkarsh Ojha, Utsav Garg, and R Venkatesh Babu. Nag: Network for adversary generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 742-751, 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 363 + ], + "type": "text", + "content": "[20] Konda Reddy Mopuri, Phani Krishna Uppala, and R Venkatesh Babu. Ask, acquire, and attack: Data-free uap generation using class impressions. In Proceedings of the European Conference on Computer Vision, pages 20-35, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "type": "text", + "content": "[21] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In 2016 IEEE European Symposium on Security and Privacy, pages 372-387, 2016." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "text", + "content": "[22] Omid Poursaeed, Isay Katsman, Bicheng Gao, and Serge Belongie. Generative adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4422-4431, 2018." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 468, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 532 + ], + "type": "text", + "content": "[23] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211–252, 2015." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 567 + ], + "type": "text", + "content": "[24] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2015." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "type": "text", + "content": "[25] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "text", + "content": "[26] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[27] Yusuke Tashiro, Yang Song, and Stefano Ermon. Diversity can be transferred: Output diversification for white-and black-box attacks. Advances in Neural Information Processing Systems, 33:4536-4548, 2020." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4770" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 484 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 171 + ], + "type": "text", + "content": "[29] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7639-7648, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 206 + ], + "type": "text", + "content": "[30] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? Advances in Neural Information Processing Systems, 27, 2014." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[31] Yunrui Yu, Xitong Gao, and Cheng-Zhong Xu. Lafeat: piercing through adversarial defenses with latent features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5735-5745, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "type": "text", + "content": "[32] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In So Kweon. Understanding adversarial examples from the mutual influence of images and perturbations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14521-14530, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "text", + "content": "[33] Chaoning Zhang, Philipp Benz, Adil Karjauv, and In So Kweon. Data-free universal adversarial perturbation and black-box attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7868-7877, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 364, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 287, + 418 + ], + "type": "text", + "content": "[34] Chaoning Zhang, Philipp Benz, Chenguo Lin, Adil Karjauv, Jing Wu, and In So Kweon. A survey on universal adversarial attack. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages 4687-4694, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 420, + 287, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 287, + 484 + ], + "type": "text", + "content": "[35] Jianping Zhang, Weibin Wu, Jen-tse Huang, Yizhan Huang, Wenxuan Wang, Yuxin Su, and Michael R Lyu. Improving adversarial transferability via neuron attribution-based attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14993-15002, 2022." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4771" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/e286768b-a5b2-4ea3-a9f1-4d06c6cc68f6_content_list.json b/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/e286768b-a5b2-4ea3-a9f1-4d06c6cc68f6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9aa115214d2d1f9b1c7b60673b49768db3136709 --- /dev/null +++ b/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/e286768b-a5b2-4ea3-a9f1-4d06c6cc68f6_content_list.json @@ -0,0 +1,1601 @@ +[ + { + "type": "text", + "text": "Take-A-Photo: 3D-to-2D Generative Pre-training of Point Cloud Models", + "text_level": 1, + "bbox": [ + 120, + 130, + 849, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyi Wang $^{1,2*}$ Xumin Yu $^{1,2*}$ Yongming Rao $^{1,2}$ Jie Zhou $^{1,2}$ Jiwen Lu $^{1,2\\dagger}$ \\\n $^{1}$ Department of Automation, Tsinghua University $^{2}$ BNRist\\\n{wziyi22, yuxm20}@mails.tsinghua.edu.cn;\\\nraoyongming95@gmail.com; {jzhou, lujiwen}@tsinghua.edu.cn", + "bbox": [ + 174, + 178, + 792, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 285, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the overwhelming trend of mask image modeling led by MAE, generative pre-training has shown a remarkable potential to boost the performance of fundamental models in 2D vision. However, in 3D vision, the over-reliance on Transformer-based backbones and the unordered nature of point clouds have restricted the further development of generative pre-training. In this paper, we propose a novel 3D-to-2D generative pre-training method that is adaptable to any point cloud model. We propose to generate view images from different instructed poses via the cross-attention mechanism as the pre-training scheme. Generating view images has more precise supervision than its point cloud counterpart, thus assisting 3D backbones to have a finer comprehension of the geometrical structure and stereoscopic relations of the point cloud. Experimental results have proved the superiority of our proposed 3D-to-2D generative pre-training over previous pre-training methods. Our method is also effective in boosting the performance of architecture-oriented approaches, achieving state-of-the-art performance when fine-tuning on ScanObjectNN classification and ShapeNet-Part segmentation tasks. Code is available at https://github.com/wangzy22/TakeAPhoto.", + "bbox": [ + 75, + 318, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 681, + 207, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nowadays, pre-training fundamental models with self-supervised mechanisms has witnessed a thriving development in the computer vision community, given its low requirement in data annotation and its high transferability to downstream applications. Self-supervised pre-training aims to fully exploit the statistical and structural knowledge from abundant annotation-free data and empowers the fundamental models with robust representation ability. In 2D vision, self-supervised pre-training has shown strong potential and achieved remarkable progress on diverse backbones in various downstream tasks. Successful pre-training strategies", + "bbox": [ + 75, + 708, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9587d5af785ce68735a5c199e73878cb3927a7b823a8c136bd6799aff9b872a9.jpg", + "image_caption": [ + "Figure 1: Principle illustration of 3D-to-2D generative pre-training. The photograph module explicitly encodes pose condition into 3D features from the backbone, and the 2D generator decodes pose-conditioned features into different view images." + ], + "image_footnote": [], + "bbox": [ + 509, + 284, + 893, + 377 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in 2D domain such as contrastive learning [19, 7] and mask modeling [18, 3] have also been adopted to 3D point cloud analysis [58, 30, 23] in recent research.", + "bbox": [ + 498, + 501, + 892, + 546 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the pre-training paradigm hasn't become the prevailing trend in 3D learning and architectural design is still the mainstream to reach a new state-of-the-art performance, which is considerably different from the dominant status of pre-training in 2D domain. In object-level analysis, generative pre-training inspired by MAE [18] has been studied thoroughly, but their performances still lag behind architecture-based methods like PointNeXt [35]. Two factors mainly contribute to the inferior status of generative pretraining in 3D learning. Since point clouds are unordered sets of point coordinates, there is no direct and precise supervision like one-to-one MSE loss between generated point clouds and their corresponding ground truth. Chamfer Distance supervision for point clouds only calculates a rough set-to-set matching and its imprecision has been widely discussed in [24, 50, 20]. Additionally, existing advanced generative pre-training methods in object analysis are limited to the Transformer-based backbone, and fail to be extended to other sophisticated point cloud models.", + "bbox": [ + 496, + 550, + 893, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To alleviate the aforementioned problems, we propose a 3D-to-2D generative pre-training method for point cloud analysis that has higher preciseness in supervision and broader adaptation to different backbones. Instead of re", + "bbox": [ + 498, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. †Corresponding author.", + "bbox": [ + 94, + 886, + 338, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5640", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "constructing point clouds as previous literature [30, 60], we propose to generate view images of the input point cloud given the instructed camera poses. This is similar to taking photos of a realistic object from different perspectives to fully present its structure or internal relations. Therefore, we name our model Take-A-Photo, in short TAP. More specifically, we propose a pose-dependent photograph module that utilizes the cross-attention mechanism to explicitly encode pose conditions with 3D features from the backbone. Then a 2D generator decodes pose-conditioned features into view images that are supervised by rendered ground truth images. The principle illustration of TAP is shown in Figure 1. In the pose-dependent photograph module, we do not provide detailed projection relations from 3D points to 2D pixels, thus the cross-attention layers are encouraged to learn by themselves what those view images look like conditioned on given poses. Since the projection layout, part occlusion relation, faces colors that represent light reflections are largely distinct among view images, the proposed 3D-to-2D generative pre-training is a challenging pretext task that obliges 3D backbone to gain higher representation ability of geometrical knowledge and stereoscopic relations.", + "bbox": [ + 80, + 90, + 468, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on various backbones and downstream tasks to verify the effectiveness and superiority of our proposed 3D-to-2D generative pre-training method. When pre-trained on synthetic ShapeNet [6] and transferred to real-world ScanObjectNN [45] classification, TAP brings consistent improvement to different backbone models and successfully outperforms previous point cloud generative pre-training methods based on Transformers backbone. With PointMLP [27] as the backbone, TAP achieves state-of-the-art performance on ScanObjectNN classification and ShapeNetPart [56] part segmentation among methods that do not include any pre-trained image or text model. We also conduct thorough ablation studies to discuss the architectural design of the TAP model and verify the individual contribution of each component.", + "bbox": [ + 80, + 424, + 468, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In conclusion, the contributions of our paper can be summarized as follows: (1) We propose TAP, the first 3D-to-2D generative pre-training method that is adaptable to any point cloud model. TAP pre-training helps to exploit the potential of point cloud models on geometric structure comprehension and stereoscopic relation understanding. (2) We propose a Photograph Module where we derive mathematical formulations to encode pose conditions as query tokens in cross-attention layers. (3) TAP surpasses previous generative pre-training methods on the Transformers backbone and achieves state-of-the-art performance on ScanObjectNN classification and ShapeNetPart segmentation.", + "bbox": [ + 80, + 650, + 468, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 80, + 845, + 215, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Point Cloud Analysis. Point cloud analysis is a fundamental and important task in the realm of 3D vision. Current", + "bbox": [ + 80, + 869, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "literature has developed two principal methodologies to extract structural representations from 3D point clouds, namely point-based and voxel-based methods. Point-based methods [32, 33, 48, 44, 35, 27] process unordered points directly, introducing various techniques for local information aggregation. Existing point-based methods can be categorized into three types: SetAbstraction-based [32, 33, 35], DynamicGraph-based [48], and Attention-based [57, 58, 30, 23, 60], all of which focus on modeling the relationships between points. Owing to their exceptional ability to effectively preserve fine-grained geometric information, point-based methods are frequently employed for object-level tasks. On the other hand, voxel-based methods [28, 21, 39] partition the 3D space into ordered voxels and employ 3D convolutions for feature extraction. Voxel-based methods are primarily based on SparseConvolution [10, 16], which enables efficient convolution operations in 3D space through sparse convolutions. In exchange for faster processing speeds, voxel-based methods sacrifice their capacity to capture detailed local structures, making them more suitable for large-scale scene-level tasks rather than object-level tasks.", + "bbox": [ + 501, + 90, + 892, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Point Cloud Pre-training. Pre-training has always been a focal point of research in the field of deep learning. Generally speaking, we usually distinguish pre-training methods based on the amount of annotation required, namely full-supervised pre-training [14, 59, 5], weakly-supervised pretraining [43, 4, 53], and unsupervised pre-training [19, 7, 8]. Among these methods, unsupervised pre-training has become the most popular approach, mainly due to its excellent transferability across tasks and its notable advantage of not relying on labeled data. Numerous researchers have proposed a variety of pretext tasks for unsupervised pre-training. Based on the type of pretext task employed, there are two prevailing pretext tasks for unsupervised pre-training. The first is contrastive learning, as exemplified by MoCo [19] and SimCLR [7]. The other method entails utilizing generative tasks to restore the data from partially or disrupted inputs, such as MAE [18] and BEiT [3].", + "bbox": [ + 501, + 417, + 892, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by pre-training strategies in the image domain, there are more and more unsupervised pre-training methods being proposed for point cloud pre-training. PointContrast [54] embraces the principle of contrastive learning, whereas PointBERT [58] and PointMAE [30] integrate reconstruction pretext tasks. However, existing generative-based pre-training methods for point clouds only consider a single modality. In this paper, we propose a cross-modal generative-based pre-training strategy to achieve more effective pre-training.", + "bbox": [ + 501, + 678, + 892, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cross-Modal Learning. Recently, cross-modal learning has been a popular research topic, aiming at learning from multiple modalities such as images, audio and point clouds. It has the potential to enhance the performance of various tasks,", + "bbox": [ + 501, + 840, + 892, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5641", + "bbox": [ + 483, + 946, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b18fe96d01a81d672dbef021c98642a46a696cbb528a7d7d794300d83fdd850a.jpg", + "image_caption": [ + "Figure 2: The pipeline of TAP pre-training method. We design a query generator to encode pose conditions and implement cross-attention layers to transform 3D point cloud features $F_{3\\mathrm{d}}$ to 2D view image features $F_{2\\mathrm{d}}$ according to pose instruction. The predicted pose-dependent view image from 2D generator is supervised by ground truth view image via MSE loss." + ], + "image_footnote": [], + "bbox": [ + 86, + 90, + 890, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "including visual recognition, speech recognition, and point cloud analysis. A variety of methods have been proposed for cross-modal learning, including multi-task learning [11], conditional generation [40], pre-training [34, 13] and tuning [49, 55].", + "bbox": [ + 75, + 321, + 470, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the realm of point cloud analysis, much previous work has explored this learning paradigm. Some literature leverages 2D data for 3D point cloud analysis, such as MVTN [17], MVCNN [42] and CrossPoint [2], proving that the multi-view images and the correspondence between images and points can be helpful for the 3D object understanding. Another line of the research, such as Image2Point [55] and P2P [49] take effort to adapt the models from 2D vision into 3D point cloud analysis, fully exploiting the relationship of 2D and 3D understanding. In this paper, we continue this learning paradigm in 3D vision domain, and for the first time propose the cross-modal generative pre-training scheme for point cloud models.", + "bbox": [ + 75, + 397, + 473, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. 3D-to-2D Generative Pre-training", + "text_level": 1, + "bbox": [ + 76, + 609, + 383, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary: Generative Pre-training", + "text_level": 1, + "bbox": [ + 76, + 633, + 401, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative pre-training is a fundamental branch of pretraining methods that aims at reconstructing integral and complete data given partial or disrupted input. Mathematically, suppose $x$ is a sample from raw data with no annotation. The pre-processing step $T(\\cdot)$ either erases part of $x$ randomly or splits $x$ into pieces and intermingles them to get $\\tilde{x} = T(x)$ . The generative pre-training model $M$ is designed to restore from those broken input $\\hat{x} = M(T(x))$ and the training loss function is designed to measure the reconstruction distance $\\mathcal{L} = D(\\hat{x},x)$ . In point cloud object analysis, earlier generative pre-training methods propose various pretext tasks as $T$ , including deformation [1], jigsaw puzzles [41] and depth projection [47] to produce disarrayed or partial point clouds. Recently, inspired by MAE [18] in the image domain, generative pre-training in 3D domain mainly focuses on implementing random masking as $T$ and utilizing", + "bbox": [ + 75, + 659, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformers model as $M$ for reconstruction [58, 30, 23, 60]. The reconstruction distance $D$ is usually measured by the classical $l_{2}$ Chamfer Distance:", + "bbox": [ + 496, + 321, + 893, + 367 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD (\\hat {x}, x) = \\frac {1}{| \\hat {x} |} \\sum_ {a \\in \\hat {x}} \\min _ {b \\in x} \\| a - b \\| _ {2} ^ {2} + \\frac {1}{| x |} \\sum_ {b \\in x} \\min _ {a \\in \\hat {x}} \\| a - b \\| _ {2} ^ {2} (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 372, + 893, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides Chamfer Distance between point clouds, some methods also exploit feature distance between latents [58] or occupancy value distance [23] as the loss function.", + "bbox": [ + 496, + 416, + 893, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The exact reason why generative pre-training would help enhance the representation ability of backbone models still remains an open question. However, abundant experimental results have conveyed that predicting missing parts according to known parts demands high reasoning ability and global comprehension capacity of the model. What's more, generative pre-training is more efficient and suitable for point cloud object analysis than contrastive pre-training, given that contrastive pre-training typically requires a large amount of training data to avoid trivial overfitting solutions but there has always been a data-starvation problem in point cloud object research field.", + "bbox": [ + 496, + 462, + 893, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overall Pipeline", + "text_level": 1, + "bbox": [ + 500, + 650, + 661, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from the aforementioned generative pre-training methods that focus on uni-modal point cloud reconstruction, we propose a novel cross-modal pre-training approach of generating view images from instructed camera poses.", + "bbox": [ + 496, + 674, + 893, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of our proposed TAP pre-training model is depicted in Figure 2. Our model takes as an input point cloud $P \\in \\mathbb{R}^{N \\times 3}$ , where $N$ is the number of points in the input point cloud. The basic building block of TAP mainly consists of: 1) a 3D Backbone that extracts 3D geometric features $F_{3\\mathrm{d}} \\in \\mathbb{R}^{n \\times C_{3\\mathrm{d}}}$ , where $n$ is the number of downsampled center points and $C_{3\\mathrm{d}}$ is the geometric feature dimension; 2) a pose-dependent Photograph Module that takes as inputs $F_{3\\mathrm{d}}$ and pose matrix $R \\in \\mathbb{R}^{3 \\times 3}$ , and predicts view image features $F_{2\\mathrm{d}}^{R} \\in \\mathbb{R}^{h \\times w \\times C_{2\\mathrm{d}}}$ conditioned on $R$ , where $h, w$ are height and width of predicted view image", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5642", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "feature map; 3) an $2D$ Generator that decodes $F_{2\\mathrm{d}}^{R}$ into an RGB image $I_{\\mathrm{gen}}^{R} \\in \\mathbb{R}^{H \\times W \\times 3}$ , where $H, W$ are height and width of the output view image.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we place no restriction on $F_{3\\mathrm{d}}$ , the 3D Backbone can be arbitrarily chosen and adopted. Therefore, our TAP is more flexible and compatible than existing generative pre-training methods that are limited to Transformer-based architecture. Experimental results in Section 4 will later verify that TAP brings consistent improvement to all kinds of point cloud models. The technical designs of the pose-dependent Photograph Module will be thoroughly discussed in Section 3.3. The 2D Generator consists of four Transpose Convolution layers to progressively upsample image resolution and decode RGB colors of each pixel.", + "bbox": [ + 75, + 137, + 470, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Photograph Module", + "text_level": 1, + "bbox": [ + 76, + 313, + 267, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architectural Design. As illustrated in Figure 2, we leverage cross-attention mechanism from Transformers [46] to build our pose-dependent Photograph Module.", + "bbox": [ + 75, + 337, + 470, + 383 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 393, + 468, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{k}$ is the scaling factor, and $Q,K,V$ are queries, keys and values matrix. More specifically, we design a Query Generator $\\Phi$ to encode camera pose conditions into query tokens: $Q = \\Phi (R)\\in \\mathbb{R}^{hw\\times C_{2\\mathrm{d}}}$ . We also design a Memory Builder $\\Theta$ to construct $K$ and $V$ from 3D geometric features: $K = V = \\Theta (F_{3\\mathrm{d}})\\in \\mathbb{R}^{m\\times C_{2\\mathrm{d}}}$ , where $m$ is the number of memory tokens. The output sequence of the cross attention layers will be rearranged from $hw\\times C_{2\\mathrm{d}}$ to $h\\times w\\times C_{2\\mathrm{d}}$ , forming the predicted view image features $F_{2\\mathrm{d}}^R$ .", + "bbox": [ + 75, + 438, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the cross-attention calculation process, we do not explicitly provide any projection clues of which 3D points would project to which 2D pixel. Instead, the Photograph Module learns by itself how to arrange unordered 3D feature points to ordered 2D pixel grids, purely based on semantic similarities between 3D geometric features and our delicately-designed queries that reveal pose information. Since one sample will only have one set of memory tokens in 3D space but its view images from different poses are quite distinct from each other, learning to predict precise view images from instructed poses in a data-driven manner is not a trivial task. Therefore, during the end-to-end optimization process, the 3D backbone is trained to have a stronger perception of the object's overall geometric structure and gain a higher representative ability of the stereoscopic relations. In this way, our proposed 3D-to-2D generative pre-training would help exploit the potential and enhance the strength of 3D backbone models.", + "bbox": [ + 75, + 575, + 470, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Query Generator. The query generator $\\Phi$ is designed to encode pose condition $R$ into 2D grid of shape $h \\times w$ . In object analysis, common practice is leveraging parallel light", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "shading to project 3D objects onto 2D grids, and pose matrix $R$ here is used to rotate objects into desired angles before projection. Therefore, each 2D grid actually represents an optical line that starts from infinity, passes through 3D objects and ends at the 2D plane. As a consequence, we choose the direction and the origin points that the optical line goes through as the delegate of the query grid.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before deriving formulations of optical lines for each grid, let us first revisit the parallel light shading process for better comprehension. Given 3D coordinates $\\mathbf{x} = (x,y,z)$ of a point cloud $P$ and pose matrix $R$ , rotation is first performed to align the object to the ideal pose position:", + "bbox": [ + 496, + 196, + 893, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\prime} = \\left(x ^ {\\prime}, y ^ {\\prime}, z ^ {\\prime}\\right) = R \\mathbf {x} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 281, + 890, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we just omit the final dimension $z^{\\prime}$ and evenly split the first two dimensions $(x^{\\prime},y^{\\prime})$ into 2D grids $(u,v)$ :", + "bbox": [ + 496, + 306, + 890, + 337 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nu = \\frac {x ^ {\\prime} - x _ {0}}{g _ {h}} + o _ {h}, \\quad v = \\frac {y ^ {\\prime} - y _ {0}}{g _ {w}} + o _ {w} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 345, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(x_0, y_0)$ is the minimum value of $(x', y')$ , $(g_h, g_w)$ is the grid size, $(o_h, o_w)$ is the offset value to place the projected object at the center of the image. $0 \\leq u \\leq h - 1$ , $0 \\leq v \\leq w - 1$ and $(u, v)$ is a sampled pixel coordinate from the 2D grid.", + "bbox": [ + 496, + 385, + 890, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now let us begin to derive formulations of the optical line that passes through the query grid. We only know $(u, v)$ for each grid and we want to reversely trace which 3D points $(x, y, z)$ are on the same optical line during parallel light projection. According to Eq. 4:", + "bbox": [ + 496, + 460, + 890, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x ^ {\\prime} = g _ {h} u + x _ {0} - o _ {h} = \\Psi_ {h} (u) \\\\ y ^ {\\prime} = g _ {w} v + y _ {0} - o _ {w} = \\Psi_ {w} (v) \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 544, + 890, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If we denote $A = R^{-1}$ and $A_{ij}$ as the element at $i^{th}$ row and $j^{th}$ column, then according to Eq. 3:", + "bbox": [ + 496, + 589, + 890, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx = A _ {1 1} \\Psi_ {h} (u) + A _ {1 2} \\Psi_ {w} (v) + A _ {1 3} z ^ {\\prime} = \\Omega_ {x} (u, v) + A _ {1 3} z ^ {\\prime}\n$$\n", + "text_format": "latex", + "bbox": [ + 496, + 628, + 890, + 645 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny = A _ {2 1} \\Psi_ {h} (u) + A _ {2 2} \\Psi_ {w} (v) + A _ {2 3} z ^ {\\prime} = \\Omega_ {y} (u, v) + A _ {2 3} z ^ {\\prime}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 648, + 890, + 664 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz = A _ {3 1} \\Psi_ {h} (u) + A _ {3 2} \\Psi_ {w} (v) + A _ {3 3} z ^ {\\prime} = \\Omega_ {z} (u, v) + A _ {3 3} z ^ {\\prime} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 667, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to the definition of line's parametric equation, Eq. 6 represents a line passing through the origin point $O: (\\Omega_x(u,v), \\Omega_y(u,v), \\Omega_z(u,v))$ with optical line direction $\\mathbf{d} = (A_{13}, A_{23}, A_{33})$ , where $\\Omega_x, \\Omega_y, \\Omega_z$ are xyz coordinates of $O$ and their formulations are conditioned on $u, v$ . Therefore, we concatenate the coordinate of origin point $O$ , normalized direction $\\mathbf{d}^{\\dagger} = \\mathbf{d} / \\| \\mathbf{d}\\|_2$ and normalized position $(u/h, v/w)$ as positional embedding together to be the initial state of our query. A multi-layer-perceptron (MLP) module is later leveraged to map the 8-dim initial query to higher dimensional space.", + "bbox": [ + 496, + 696, + 893, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory Builder. The memory builder takes $F_{3\\mathrm{d}}$ as input to prepare for initial state of $K, V$ in cross-attention layers. We", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5643", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "first concatenate aligned 3D coordinate $P_{\\mathrm{3d}}$ with 3D features to enhance the geometric knowledge of $F_{\\mathrm{3d}}$ :", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} _ {3 \\mathrm {d}} = \\operatorname {M L P} \\left(\\operatorname {c a t} \\left(F _ {3 \\mathrm {d}}, P _ {3 \\mathrm {d}}\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 132, + 470, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, we initialize a learnable memory token $T_{\\mathrm{pad}}$ as the pad token and concatenate it with $\\hat{F}_{3\\mathrm{d}}$ to obtain the initial state of $K, V$ . The reason for concatenating a learnable pad token $T_{\\mathrm{pad}}$ is that there are white background areas on the projected image (as shown in Figure 2). As $F_{3\\mathrm{d}}$ only encodes foreground objects, we further need a learnable pad token to represent background regions. Otherwise, the cross-attention layers will be confused to learn how to combine foreground tokens into background features and this will inevitably diminish the pre-training effectiveness.", + "bbox": [ + 75, + 162, + 470, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Objective Function", + "text_level": 1, + "bbox": [ + 76, + 324, + 261, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform per-pixel supervision with Mean Squared Error (MSE) loss between generated view image $I_{\\mathrm{gen}}^{R}$ and ground truth image $I_{\\mathrm{gt}}^{R}$ , aligned by camera pose $R$ . For simplicity, we will omit $R$ in later formulations. As the background of the rendered ground truth images is all white and reveals little information, we further design a compound loss to balance the weight between foreground regions and background regions:", + "bbox": [ + 75, + 348, + 470, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(I _ {\\text {g e n}}, I _ {\\text {g t}}\\right) = w ^ {\\mathrm {f g}} \\mathcal {D} ^ {\\mathrm {f g}} + w ^ {\\mathrm {b g}} \\mathcal {D} ^ {\\mathrm {b g}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 481, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} ^ {k} \\left(I _ {\\text {g e n}} ^ {k}, I _ {\\mathrm {g t}} ^ {k}\\right) = \\frac {1}{H W} \\sum_ {h, w} \\left(I _ {\\text {g e n}} ^ {k} (h, w) - I _ {\\mathrm {g t}} ^ {k} (h, w)\\right) ^ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 511, + 468, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $k = \\mathrm{fg}$ (foreground), $\\mathrm{bg}$ (background) and $w^{\\mathrm{fg}}$ , $w^{\\mathrm{bg}}$ are loss weights for foreground and background, respectively. Such per-pixel supervision is more precise than the ambiguous set-to-set Chamfer Distance introduced in Eq. 1.", + "bbox": [ + 75, + 555, + 468, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 630, + 209, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the setups of our pretraining scheme. Then we evaluate our pre-training method on various point cloud backbones by fine-tuning them on different downstream tasks, such as point cloud classification on ModelNet and ScanObjectNN datasets, and part segmentation on the ShapeNetPart dataset. Finally, we provide in-depth ablation studies for the architectural design of our proposed TAP pre-training pipeline.", + "bbox": [ + 75, + 655, + 470, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Pre-training Setups", + "text_level": 1, + "bbox": [ + 76, + 786, + 264, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Setups To align with previous research practices [58, 30, 54], we choose ShapeNet [6] that contains more than 50 thousand CAD models as our pre-training datasets. We sampled 1024 points from each 3D CAD model to form the point clouds, consistent with previous work. Since ShapeNet does not provide images for each point cloud, we use the", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/776be80e0544f6f2a35b36e2b765985afa3c92353ae7bd111b6b14d83f75274a.jpg", + "table_caption": [ + "Table 1: Classification results on the ScanObjectNN dataset. We report the overall accuracy $(\\%)$ . The results with $\\dagger$ are reproduced by PointNeXt [35] repository." + ], + "table_footnote": [], + "table_body": "
MethodOBJ_BGOBJ_ONLYPB_T50_RS
Hierarchical Models with TAP Pre-training
†DGCNN [48]--86.1
+ TAP--86.6 (+0.5)
†PointNet++ [33]--86.2
+ TAP--86.8 (+0.6)
†PointMLP [27]--87.4
+ TAP--88.5 (+1.1)
Standard Transformers with Generative Pre-training
w/o pre-training [46]79.8680.5577.24
OcCo [47]84.8585.5478.79
Point-BERT [58]87.4388.1283.07
MaskPoint [23]89.3088.1084.30
Point-MAE [30]90.0288.2985.18
TAP (Ours)90.3689.5085.67
", + "bbox": [ + 504, + 146, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "rendered image from 12 surrounding viewpoints generated by MVCNN [42]. During our pre-training, the models are exclusively pre-trained with the training split following the practice of previous work [58].", + "bbox": [ + 496, + 415, + 890, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture Setups We conduct experiments on various point cloud encoders, including PointNet++ [33], DGCNN [48], PointMLP [27] and Transformers [58] for point cloud object classification. During the pre-training stage, the photograph module takes encoded point cloud features and pose conditions as inputs to generate a $32 \\times$ downsampled view image feature map of size $7 \\times 7$ from a specific viewpoint. Then the 2D generator progressively upsamples the image feature map to decode RGB view images of size $224 \\times 224$ . We do not alter the architecture of the point cloud backbone since the photograph module and the 2D generator are exclusively used during the pretraining phase and are dropped during the fine-tuning stage. In our experiment, the photograph module is a six-layer cross-attention block, with attention layer channels limited to 256 to enhance efficiency. During the pre-training task on ShapeNet, we utilized four simple transpose convolutions to upsample the reconstructed 2D feature map and predict the RGB value for each pixel.", + "bbox": [ + 496, + 483, + 893, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details The experiments of TAP pretraining and finetuning on various downstream tasks are implemented with PyTorch [31]. We utilize AdamW [26] optimizer and the CosineAnnealing learning rate scheduler [25] to pre-train the point cloud backbone for 100 epochs. We set the initial learning rate as $5e^{-4}$ and weight decay as $5e^{-2}$ . In our experiment, we train various point cloud backbones with 32 batch sizes on a single Nvidia 3090Ti GPU. The", + "bbox": [ + 496, + 779, + 895, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5644", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/60bd2d3e38885583d7db2103f5becc7e1372f5a0d87b33943477a97082083179.jpg", + "table_caption": [ + "Table 2: Results comparisons with previous methods on the ScanObjectNN and ModelNet40 datasets. The model parameters number (#Params) and overall accuracy (\\%) are reported. † denotes our reproduced results of PointMLP on the ModelNet40 dataset. Methods with * introduce extra knowledge from pre-trained image models or pre-trained vision-language models. We do not compete with them for fair comparison and only list them for reference." + ], + "table_footnote": [], + "table_body": "
Method#Params (M)ScanObjectNNModelNet40
Supervised Learning Only
PointNet [32]3.568.089.2
PointNet++ [33]1.577.990.7
Transformer [46]22.177.2491.4
DGCNN [48]1.878.192.9
PointCNN [22]0.678.592.2
DRNet [36]-80.393.1
SimpleView [15]-80.5±0.393.9
GBNet [37]8.881.093.8
PRA-Net [9]2.381.093.7
MVTN [17]11.282.893.8
RepSurf-U [38]1.584.394.4
PointMLP [27]12.685.4±0.393.5†
PointNeXt [35]1.487.7±0.493.7
Transformers with Pre-training
OcCo [47]22.178.892.1
Point-BERT [58]22.183.193.2
MaskPoint [23]22.184.393.8
Point-MAE [30]22.185.293.8
Point-M2AE [60]15.386.494.0
With Pre-trained Image Model
ACT [13]*22.188.293.7
I2P-MAE [61]*-90.193.7
ReCon [34]*44.390.694.1
Our Proposed TAP Pre-training
PointMLP + TAP12.688.594.0
", + "bbox": [ + 80, + 222, + 467, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "drop path rate of the cross-attention layer is set to 0.1. The foreground and background loss weights $w^{\\mathrm{fg}}$ , $w^{\\mathrm{bg}}$ are set to 20 and 1. The detailed architecture of our simple 2D generator is: TConv(256, 128, 5, 4) → TConv(128, 64, 3, 2) → TConv(64, 32, 3, 2) → TConv(32, 3, 3, 2), where TConv stands for Transpose Convolution and the four numbers in the tuple denotes $(C_{in}, C_{out}, \\text{Kernel}, \\text{Stride})$ respectively. During the fine-tuning stage, we perform a learning rate warming up for point cloud backbones with 10 epochs, and keep other settings unchanged for a fair comparison.", + "bbox": [ + 75, + 641, + 470, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Downstream Tasks", + "text_level": 1, + "bbox": [ + 76, + 801, + 259, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we report the experimental results of various downstream tasks. We follow the previous work to conduct experiments of object classification on real-world ScanObjectNN and synthetic ModelNet40 datasets. We also verify the effectiveness of our pre-training method on the", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ab0c1890cc27997e1a36eee33a8013973e244ab527bce5bc78a7f80cab1c1d48.jpg", + "image_caption": [ + "Figure 3: The visualization results of our proposed 3D-to-2D generative pre-training. The first row displays view images generated by our TAP pre-training pipeline and the second row shows ground truth images. Our TAP can produce view images with appropriate shapes and reflection colors, demonstrating its ability in capturing geometric structure and stereoscopic knowledge." + ], + "image_footnote": [], + "bbox": [ + 501, + 94, + 890, + 204 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "part segmentation task with the ShapeNetPart dataset.", + "bbox": [ + 498, + 359, + 854, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Object Classification", + "text_level": 1, + "bbox": [ + 498, + 391, + 699, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Main Results. To evaluate the effectiveness of our proposed TAP, we implement it with various point cloud architectures, including classical baselines such as PointNet++, DGCNN, and PointMLP, as well as widely used Standard Transformers backbone for existing generative pre-training methods. We follow the common practice to experiment our model on three variants of the ScanObjectNN dataset: 1) OBJ-ONLY: cropping object without any background; 2) OBJ-BG: containing the background and object; 3) PB-T50-RS: adopting various augmentations to the objects. We reported the results comparing with existing pre-training methods in Table 1.", + "bbox": [ + 496, + 416, + 893, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in the upper part of the table, we pretrain the graph-based architecture DGCNN, set-abstraction-based architecture PointNet++, and MLP-based architecture PointMLP with TAP, and observe consistent improvements across models. The results strongly convey that our proposed TAP can be successfully applied to various types of point cloud models and the proposed novel 3D-to-2D generative pre-training is effective regardless of the backbone architecture. Considering that nearly all existing generative pre-training methods are specially designed for Transformer-based architecture, our TAP is much superior in its wider adaptation and higher flexibility. Additionally, we also provide a detailed and fair comparison with previous work by implementing TAP with the Standard Transformers architecture in the lower part of the table, where no hierarchical designs or inductive bias is included. Our TAP outperforms previous pre-training methods in all three split settings, providing strong evidence that our 3D-to-2D generative pretraining strategy can also benefit attention-based architectures and surpass uni-modal generative pre-training competitors. It is worth noting that although TAP and many previous", + "bbox": [ + 496, + 584, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5645", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0b27ca5a37e7fde9b3a05957d4de0b866f0716395adf606bbfa1e35a75c1df4d.jpg", + "table_caption": [ + "Table 3: Part segmentation results on the ShapeNetPart dataset. We report the mean IoU across all part categories $\\mathrm{mIoU}_C$ , the mean IoU across all instances $\\mathrm{mIoU}_I$ , and the IoU for each category." + ], + "table_footnote": [], + "table_body": "
MethodsmIoUCmIoUIaerobagcapcarchairearphoneguitarknifelamplaptopmotormugpistolrocketskateboardtable
Supervised Representation Learning Only
PointNet [32]80.483.783.478.782.574.989.673.091.585.980.895.365.293.081.257.972.880.6
PointNet++ [33]81.985.182.479.087.777.390.871.891.085.983.795.371.694.181.358.776.482.6
DGCNN [48]82.385.284.083.486.777.890.674.791.287.582.895.766.394.981.163.574.582.6
PointMLP [27]84.686.183.583.487.580.590.378.292.288.182.696.277.595.885.464.683.384.3
KPConv [44]85.186.484.686.387.281.191.177.892.688.482.796.278.195.885.469.082.083.6
Transformers with Uni-modal Generative Pre-training
Point-BERT [58]84.185.684.384.888.079.891.081.791.687.985.295.675.694.784.363.476.381.5
Point-MAE [30]84.286.184.385.088.380.591.378.592.187.486.196.175.294.684.763.577.182.4
MaskPoint [23]84.486.084.285.688.180.391.279.591.987.886.295.376.995.085.364.476.981.8
Point-M2AE [60]84.986.5----------------
Our Proposed 3D-to-2D Generative Pre-training
PointMLP+TAP85.286.984.886.189.582.592.175.992.388.785.696.579.896.085.966.278.183.2
", + "bbox": [ + 76, + 125, + 893, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pre-training approaches have significantly improved the performance of Transformers in point cloud tasks, increasing the accuracy from 77.24 to 85.67, they still lag far behind advanced point cloud networks such as PointMLP. Therefore, TAP's applicability to various models is an important characteristic that would benefit future research.", + "bbox": [ + 75, + 349, + 472, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons with Previous Methods. To clearly demonstrate the high performance of our proposed TAP pre-training in object classification tasks, we compare TAP with existing methods on both synthetic ModelNet and real-world ScanObjectNN (the hardest PB-T50-RS variant) datasets in Table 2. We categorize existing methods into two types: (1) architecture-oriented methods, which focus on developing novel model architectures for 3D point clouds and do not involve any pre-training techniques, and (2) pre-training methods, which pay more attention to the pre-training strategy and whose backbone model are mostly limited to Transformer-base architecture. It's worth noticing that methods marked with an asterisk (*) incorporate additional knowledge from pre-trained image models or pre-trained vision-language models. To ensure a fair and unbiased comparison, we refrain from directly comparing our method with these approaches. However, we include them in the listing for reference purposes, acknowledging their existence and potential relevance in related research.", + "bbox": [ + 75, + 446, + 472, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the experimental results, we can see that accompanied by PointMLP backbone model, our proposed TAP pre-training achieves the best classification accuracy on ScanObjectNN and ModelNet40 among existing models (with no pre-trained knowledge from image or language like P2P [49]), demonstrating the effectiveness of our approach and validating the superiority of our 3D-to-2D cross-modal generative pre-training method over previous generative pretraining methods. Moreover, we also note that our proposed method has brought higher performance improvements on the ScanObjectNN dataset than on ModelNet40. This may", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "be attributed to the reason that the cross-modal generative pre-training has enhanced the network's ability to understand point clouds from different views, which is beneficial for a more robust understanding of the real-scan data with more noise and disturbance in the ScanObjectNN dataset.", + "bbox": [ + 496, + 349, + 893, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization Results. Figure 3 shows the visualization results of TAP. The first row shows the generated view images while the second row displays the ground truth images for reference. The TAP method can successfully predict the accurate shape of the object and the RGB colors that represent light reflections in rendered images. Therefore, TAP is capable of capturing the geometric structure of 3D objects and reasoning occlusion relations from specific camera poses.", + "bbox": [ + 496, + 431, + 895, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Part Segmentation", + "text_level": 1, + "bbox": [ + 500, + 573, + 684, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performing dense prediction is always a more challenging task compared with classification. In this section, we evaluate the local distinguishability of our proposed TAP pre-training method, fine-tuning the pre-trained point cloud model on the ShapeNetPart dataset for the part segmentation task. Quantitative results are shown in Table 3. We implement PointMLP as the backbone model and compare our TAP results with two mainstreams of previous literature. The upper row displays classical architecture-oriented methods that focus on network design and are trained from scratch. The lower row shows members of the generative pre-training family that rely on Transformer-based architectures.", + "bbox": [ + 496, + 598, + 893, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "According to results comparisons, our TAP pre-training significantly improves the part segmentation performance of the PointMLP backbone, increasing class mIoU by 0.6 and instance mIoU by 0.8. More importantly, our TAP pretraining achieves state-of-the-art performance on both class mIoU and instance mIoU, surpassing leading works in both tracks. Specifically, TAP exceeds the performance of PointM2AE on instance mIoU by 0.4. This satisfactory perfor", + "bbox": [ + 496, + 780, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5646", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance serves as strong evidence to convey that our proposed TAP pre-training is superior to previous uni-model generative pre-training mechanisms in dense prediction tasks. This may be attributed to the factor that our supervision in 2D with MSE loss is more precise than the ambiguous Chamfer Distance in 3D reconstruction. Therefore, models with TAP pre-training obtain more accurate comprehension of local geometry and detail awareness, which contributes to mIoU gain in dense prediction tasks. What's more, TAP outperforms KPCov on instance mIoU by 0.5, demonstrating that the proposed 3D-to-2D generative pre-training method can fully exploit the potential of the point cloud model and help it have a better perception of objects' geometric structure. As TAP is adaptable to any architecture, future improvements in architectural design can also benefit from TAP pre-training.", + "bbox": [ + 75, + 90, + 472, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.3 Few-shot Classification", + "text_level": 1, + "bbox": [ + 76, + 339, + 292, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following Point-BERT [58], we conduct few-shot classification with Standard Transformers on ModelNet40 [52] dataset. As shown in Table 4, we report mean overall accuracy and standard deviation $(\\mathrm{mOA} \\pm \\mathrm{std})$ on 10 pre-defined data folders for each few-shot setting. The way and shot in Table 4 specify the number of categories and the number of training examples per category, respectively.", + "bbox": [ + 75, + 364, + 470, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the results, TAP achieves the highest mean overall accuracy across all few-shot settings when compared to previous generative pre-training approaches. Furthermore, TAP exhibits significantly lower standard deviations than those reported in the existing literature for the majority of few-shot settings, which signifies its robust performance and consistent superiority. This indicates that TAP is not only capable of achieving high mean overall accuracy but also exhibits reliability and robustness across various few-shot settings. Such stability is crucial in real-world applications, where consistency and predictability are vital for practical deployment.", + "bbox": [ + 75, + 472, + 470, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Scene-level Dense Predictions", + "text_level": 1, + "bbox": [ + 76, + 664, + 339, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess the effectiveness of TAP in handling scene-level dense prediction tasks, we carry out experiments on more complicated scene-level object detection and semantic segmentation on the ScanNetV2 [12] dataset. For the object detection task, we adopt 3DETR [29] and pre-train its encoder on the object-level dataset ShapeNet [6] with TAP. Average precision at 0.25 and 0.5 IoU thresholds are reported. Regarding semantic segmentation, we employ the PointTransformerV2 [51](PTv2) model and pre-train it on the ScanNetV2 dataset with TAP. We report mean IoU for evaluation metric. It is worth mentioning that PTv2 represents the current state-of-the-art approach with open-source code availability.", + "bbox": [ + 75, + 688, + 470, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on the results presented in Table 5, TAP consis", + "bbox": [ + 96, + 885, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b112527a4f069ca40185e2c4982ebbff230031afbe9765283fc79b8e43e2677f.jpg", + "table_caption": [ + "Table 4: Few-shot Classification with Standard Transformers on ModelNet40 dataset. We report mean overall accuracy and standard deviation on 10 pre-defined data folders for each setting. Best results are marked bold." + ], + "table_footnote": [], + "table_body": "
Method5-way10-way
10-shot20-shot10-shot20-shot
w/o pre-training87.8 ± 5.293.3 ± 4.384.6 ± 5.589.4 ± 6.3
Point-BERT [58]94.6 ± 3.196.3 ± 2.791.0 ± 5.492.7 ± 5.1
MaskPoint [23]95.0 ± 3.797.2 ± 1.791.4 ± 4.093.4 ± 3.5
Point-MAE [30]96.3 ± 2.597.8 ± 1.892.6 ± 4.195.0 ± 3.0
TAP97.3 ± 1.897.8 ± 1.793.1 ± 2.695.8 ± 1.0
", + "bbox": [ + 504, + 160, + 890, + 289 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f8a3b75ebae62c471b61602142a7a8635157dfb128e0a5558a780ac0fc9fd694.jpg", + "table_caption": [ + "Table 5: Scene-level object detection and semantic segmentation on ScanNetV2 [12]. Average precision at 0.25 IoU thresholds $(\\mathrm{AP}_{0.25})$ and 0.5 IoU thresholds $(\\mathrm{AP}_{0.5})$ of detection and mean Intersection-over-Union (mIoU) of semantic segmentation are reported." + ], + "table_footnote": [], + "table_body": "
MethodDet (3DETR [29])Seg (PTv2 [51])
AP0.25AP0.5mIoU
Baseline62.137.972.4
+TAP63.0 (+0.9)41.4 (+3.5)72.6 (+0.2)
", + "bbox": [ + 504, + 388, + 890, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tently enhances the performance of all baselines, thereby showcasing its efficacy in tackling more intricate scene-level dense prediction tasks. Remarkably, even with the encoder solely pre-trained on an object-level dataset for scene-level detection task, significant improvements are observed in both $\\mathrm{AP}_{0.25}$ and $\\mathrm{AP}_{0.5}$ metrics. This suggests that the learned representations from TAP effectively capture relevant information and generalize well to complex scenes, even when the pre-training data is limited to object-level collections. Such generalization capabilities are valuable in scenarios where obtaining large-scale fully annotated scene-level datasets may be challenging or expensive.", + "bbox": [ + 496, + 497, + 893, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 688, + 663, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To investigate the architectural design of our proposed Photograph Module in TAP pre-training pipeline, we conduct extensive ablation studies on the ScanObjectNN dataset with PointMLP as the backbone model.", + "bbox": [ + 496, + 710, + 893, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Photograph Module Architectural Designs. In Photograph Module, we implement cross-attention layers to generate view image feature maps conditioned on pose instruction. We believe that letting the module learn by itself how to rearrange 3D point features in 2D grids will enhance the representation ability of the 3D backbone. Therefore, we conduct ablation studies to verify this hypothesis. As shown in Table 6a, we implement Model $\\mathrm{A}_1$ with no attention lay", + "bbox": [ + 496, + 779, + 895, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5647", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6: Ablation studies on Photograph module in TAP pre-training pipeline. We choose PointMLP as the backbone model and conduct ablation studies on the ScanObjectNN dataset from two aspects: overall architectural designs and query designs. In Table(a), we first investigate the effectiveness of cross-attention design compared with direct projection (Model A1) and direct project with self-attention (Model A2). Then we analyze the influence of the different number of attention layers and feature channels in Model B and Model C. Finally, we discuss whether pad token in memory builder is beneficial in Model D. In Table(b), we conduct further experiments to compare different approaches for query designing: (1) Using learnable query based on the given viewpoints (Model E) or use mathematical formulations derived in Eq. 6. (2) The information we need when we mathematically encode pose information into init query status: (i) Origin: the coordinate of origin point $O$ that the optical line passes through. (ii) Direction: The normalized direction of the optical line. (iii) PE: The position embedding for each grid.", + "bbox": [ + 75, + 89, + 893, + 218 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/6965b9dc9c3650fd98f132215ee79a22973b88811af3e6a623836104ede7473f.jpg", + "table_caption": [ + "(a) Overall Architectural Designs." + ], + "table_footnote": [], + "table_body": "
ModelAttention TypeLayerNum.ChannelsMem.PadAcc.(%)
A1None-256X87.6 (-0.9)
A2SelfAttn6 layers256X87.8 (-0.7)
BCrossAttn2 layers256✓87.9 (-0.6)
CCrossAttn6 layers512✓87.8 (-0.7)
DCrossAttn6 layers256X88.3 (-0.2)
TAPCrossAttn6 layers256✓88.5
", + "bbox": [ + 81, + 234, + 519, + 359 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/8c927dcad61263ee7d6c6511ec8a5f7df254fba0d0da748cc1cbe5b343715a37.jpg", + "table_caption": [ + "(b) Query Designs." + ], + "table_footnote": [], + "table_body": "
ModelQuery TypeOriginDirectionPEAcc.(%)
ELearnableXXX87.5 (-1.0)
F1FormulaX✓✓86.5 (-2.0)
F2Formula✓XX87.8 (-0.7)
F3Formula✓✓X88.0 (-0.5)
F4Formula✓X✓88.1 (-0.4)
TAPFormula✓✓✓88.5
", + "bbox": [ + 534, + 234, + 893, + 362 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ers, directly projecting 3D feature points to 2D grids based on Eq. 3 and Eq. 4. This results in a much simpler pretraining task, as the projection relation has been directly told. Additionally, in Model $\\mathrm{A}_2$ , we add self-attention layers after explicit projection to help the model capture longer-range correlations. Pose knowledge is encoded as a pose token that is concatenated to projected grids, similar to the CLS token in classification Transformers. According to quantitative results comparison with TAP that implements cross-attention layers, fine-tuning results of pre-training methods in $\\mathrm{A}_1$ and $\\mathrm{A}_2$ version show inferiority. Therefore, the cross-attention architecture we designed to entirely LEARN the projection relation is the most suitable choice for the proposed 3D-to-2D generative pre-training.", + "bbox": [ + 75, + 387, + 472, + 598 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "What's more, we discuss the number of cross-attention layers, the dimension of feature channels and whether to concatenate pad token in memory builder in Model $B, C, D$ . According to the results, more cross-attention layers show stronger representation ability, while too large channel number will lead to performance decrease caused by over-fitting. The performance gain from Model $D$ to TAP also verifies that the pad token design in the memory builder is essential.", + "bbox": [ + 75, + 601, + 472, + 724 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Query Generator Designs. In the query generator, we derive the mathematical formulation of the optical lines passing through 2D grids. We propose to concatenate the coordinate of origin point $O$ , normalized direction $\\mathbf{d}^{\\dagger}$ and position embedding $(u / h, v / w)$ as the initial state of queries. In Table 6b, we first compare this mathematical design with totally learnable queries that takes pose matrix $R$ as input and implements MLP layers to predict query for each grid. As shown in Model $E$ , learnable queries cannot satisfactorily encode pose information, while our derived formulation for query construction is both clearer in physical meaning and", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "more competitive in fine-tuning accuracy.", + "bbox": [ + 498, + 387, + 777, + 402 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In ablation $\\mathrm{F_1}$ to $\\mathrm{F_4}$ , we progressively discuss the three components of query generation. Quantitative comparison with TAP verifies that every component is indispensable for query generation, where coordinates of origin points are of the most importance.", + "bbox": [ + 496, + 402, + 893, + 479 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 498, + 627, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we have proposed a novel 3D-to-2D generative pre-training method TAP that is adaptable to any point cloud model. We implemented the cross-attention mechanism to generate view images of point clouds from instructed camera poses. To better encode pose conditions and generate physically meaningful queries, we derived mathematical formulations of optical lines. The proposed TAP pre-training had higher preciseness in supervision and broader adaptation to different backbones, compared with directly reconstructing point clouds in previous methods. Experimental results conveyed that the TAP pre-training can help the backbone models better capture the structural knowledge and stereoscopic relations. Fine-tuning results of TAP pre-training achieve state-of-the-art performance on ScanObjectNN classification and ShapeNetPart segmentation, among methods that do not include any pre-trained image or text models. We believe the cross-modal generative pre-training paradigm will be a promising direction for future research.", + "bbox": [ + 496, + 523, + 895, + 796 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 809, + 660, + 825 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0114903 and in part by the National Natural Science Foundation of China under Grant 62125603.", + "bbox": [ + 498, + 834, + 893, + 895 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5648", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Idan Achituve, Haggai Maron, and Gal Chechik. Self-supervised learning for domain adaptation on point clouds. In WACV, 2021.", + "[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In CVPR, pages 9902-9912, June 2022.", + "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.", + "[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. *NeruIPS*, 2019.", + "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017.", + "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015.", + "[7] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020.", + "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020.", + "[9] Silin Cheng, Xiwu Chen, Xinwei He, Zhe Liu, and Xiang Bai. Pra-net: Point relation-aware network for 3d point cloud analysis. IEEE TIP, 30:4436-4448, 2021.", + "[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019.", + "[11] Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning. In ICML, 2008.", + "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017.", + "[13] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022.", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021.", + "[15] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In ICML, 2021." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018.", + "[17] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. MVTN: multi-view transformation network for 3d shape recognition. In ICCV, 2021.", + "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll'ar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021.", + "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020.", + "[20] Tianxin Huang, Zhonggan Ding, Jiangning Zhang, Ying Tai, Zhenyu Zhang, Mingang Chen, Chengjie Wang, and Yong Liu. Learning to measure the point cloud reconstruction loss in a representation space. In CVPR, 2023.", + "[21] Roman Klokov and Victor Lempitsky. Escape from cells: Deep kd-networks for the recognition of 3d point cloud models. In ICCV, 2017.", + "[22] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. NeurIPS, 2018.", + "[23] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV. Springer, 2022.", + "[24] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and Shi-Min Hu. Morphing and sampling network for dense point cloud completion. In AAAI, 2020.", + "[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016.", + "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018.", + "[27] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In ICLR, 2022.", + "[28] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In IROS. IEEE, 2015.", + "[29] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, 2021.", + "[30] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV. Springer, 2022.", + "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019.", + "[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017.", + "[33] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++ deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017." + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5649", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. arXiv preprint arXiv:2302.02318, 2023.", + "[35] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022.", + "[36] Shi Qiu, Saeed Anwar, and Nick Barnes. Dense-resolution network for point cloud classification and segmentation. In WACV, 2021.", + "[37] Shi Qiu, Saeed Anwar, and Nick Barnes. Geometric backprojection network for point cloud classification. IEEE TMM, 2022.", + "[38] Haoxi Ran, Jun Liu, and Chengjie Wang. Surface representation for point clouds. In CVPR, 2022.", + "[39] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Oct-net: Learning deep 3d representations at high resolutions. In CVPR, 2017.", + "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022.", + "[41] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. NeurIPS, 2019.", + "[42] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In ICCV, 2015.", + "[43] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 2017.", + "[44] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. ICCV, 2019.", + "[45] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019.", + "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017.", + "[47] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021.", + "[48] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. TOG, 2019.", + "[49] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. In NeurIPS, 2022.", + "[50] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Balanced chamfer distance as a comprehensive metric for point cloud completion. In NeurIPS, 2021.", + "[51] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015.", + "[53] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020.", + "[54] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020.", + "[55] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. arXiv preprint arXiv:2106.04180, 2021.", + "[56] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ToG, 2016.", + "[57] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In ICCV, 2021.", + "[58] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022.", + "[59] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, 2022.", + "[60] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. arXiv preprint arXiv:2205.14401, 2022.", + "[61] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In CVPR, 2023." + ], + "bbox": [ + 501, + 92, + 893, + 571 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5650", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/e286768b-a5b2-4ea3-a9f1-4d06c6cc68f6_model.json b/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/e286768b-a5b2-4ea3-a9f1-4d06c6cc68f6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aa5b49b43643beb0d44f71c4ac9bf5b939eab470 --- /dev/null +++ b/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/e286768b-a5b2-4ea3-a9f1-4d06c6cc68f6_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.131, + 0.85, + 0.152 + ], + "angle": 0, + "content": "Take-A-Photo: 3D-to-2D Generative Pre-training of Point Cloud Models" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.179, + 0.794, + 0.253 + ], + "angle": 0, + "content": "Ziyi Wang\\(^{1,2*}\\) Xumin Yu\\(^{1,2*}\\) Yongming Rao\\(^{1,2}\\) Jie Zhou\\(^{1,2}\\) Jiwen Lu\\(^{1,2\\dagger}\\)\\\n\\(^{1}\\)Department of Automation, Tsinghua University \\(^{2}\\)BNRist\\\n{wziyi22, yuxm20}@mails.tsinghua.edu.cn;\\\nraoyongming95@gmail.com; {jzhou, lujiwen}@tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.286, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.651 + ], + "angle": 0, + "content": "With the overwhelming trend of mask image modeling led by MAE, generative pre-training has shown a remarkable potential to boost the performance of fundamental models in 2D vision. However, in 3D vision, the over-reliance on Transformer-based backbones and the unordered nature of point clouds have restricted the further development of generative pre-training. In this paper, we propose a novel 3D-to-2D generative pre-training method that is adaptable to any point cloud model. We propose to generate view images from different instructed poses via the cross-attention mechanism as the pre-training scheme. Generating view images has more precise supervision than its point cloud counterpart, thus assisting 3D backbones to have a finer comprehension of the geometrical structure and stereoscopic relations of the point cloud. Experimental results have proved the superiority of our proposed 3D-to-2D generative pre-training over previous pre-training methods. Our method is also effective in boosting the performance of architecture-oriented approaches, achieving state-of-the-art performance when fine-tuning on ScanObjectNN classification and ShapeNet-Part segmentation tasks. Code is available at https://github.com/wangzy22/TakeAPhoto." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.682, + 0.208, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.709, + 0.471, + 0.876 + ], + "angle": 0, + "content": "Nowadays, pre-training fundamental models with self-supervised mechanisms has witnessed a thriving development in the computer vision community, given its low requirement in data annotation and its high transferability to downstream applications. Self-supervised pre-training aims to fully exploit the statistical and structural knowledge from abundant annotation-free data and empowers the fundamental models with robust representation ability. In 2D vision, self-supervised pre-training has shown strong potential and achieved remarkable progress on diverse backbones in various downstream tasks. Successful pre-training strategies" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.285, + 0.895, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.388, + 0.895, + 0.464 + ], + "angle": 0, + "content": "Figure 1: Principle illustration of 3D-to-2D generative pre-training. The photograph module explicitly encodes pose condition into 3D features from the backbone, and the 2D generator decodes pose-conditioned features into different view images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.502, + 0.893, + 0.547 + ], + "angle": 0, + "content": "in 2D domain such as contrastive learning [19, 7] and mask modeling [18, 3] have also been adopted to 3D point cloud analysis [58, 30, 23] in recent research." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.895, + 0.837 + ], + "angle": 0, + "content": "However, the pre-training paradigm hasn't become the prevailing trend in 3D learning and architectural design is still the mainstream to reach a new state-of-the-art performance, which is considerably different from the dominant status of pre-training in 2D domain. In object-level analysis, generative pre-training inspired by MAE [18] has been studied thoroughly, but their performances still lag behind architecture-based methods like PointNeXt [35]. Two factors mainly contribute to the inferior status of generative pretraining in 3D learning. Since point clouds are unordered sets of point coordinates, there is no direct and precise supervision like one-to-one MSE loss between generated point clouds and their corresponding ground truth. Chamfer Distance supervision for point clouds only calculates a rough set-to-set matching and its imprecision has been widely discussed in [24, 50, 20]. Additionally, existing advanced generative pre-training methods in object analysis are limited to the Transformer-based backbone, and fail to be extended to other sophisticated point cloud models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To alleviate the aforementioned problems, we propose a 3D-to-2D generative pre-training method for point cloud analysis that has higher preciseness in supervision and broader adaptation to different backbones. Instead of re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.339, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. †Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5640" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.47, + 0.423 + ], + "angle": 0, + "content": "constructing point clouds as previous literature [30, 60], we propose to generate view images of the input point cloud given the instructed camera poses. This is similar to taking photos of a realistic object from different perspectives to fully present its structure or internal relations. Therefore, we name our model Take-A-Photo, in short TAP. More specifically, we propose a pose-dependent photograph module that utilizes the cross-attention mechanism to explicitly encode pose conditions with 3D features from the backbone. Then a 2D generator decodes pose-conditioned features into view images that are supervised by rendered ground truth images. The principle illustration of TAP is shown in Figure 1. In the pose-dependent photograph module, we do not provide detailed projection relations from 3D points to 2D pixels, thus the cross-attention layers are encouraged to learn by themselves what those view images look like conditioned on given poses. Since the projection layout, part occlusion relation, faces colors that represent light reflections are largely distinct among view images, the proposed 3D-to-2D generative pre-training is a challenging pretext task that obliges 3D backbone to gain higher representation ability of geometrical knowledge and stereoscopic relations." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.425, + 0.47, + 0.649 + ], + "angle": 0, + "content": "We conduct extensive experiments on various backbones and downstream tasks to verify the effectiveness and superiority of our proposed 3D-to-2D generative pre-training method. When pre-trained on synthetic ShapeNet [6] and transferred to real-world ScanObjectNN [45] classification, TAP brings consistent improvement to different backbone models and successfully outperforms previous point cloud generative pre-training methods based on Transformers backbone. With PointMLP [27] as the backbone, TAP achieves state-of-the-art performance on ScanObjectNN classification and ShapeNetPart [56] part segmentation among methods that do not include any pre-trained image or text model. We also conduct thorough ablation studies to discuss the architectural design of the TAP model and verify the individual contribution of each component." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.651, + 0.47, + 0.831 + ], + "angle": 0, + "content": "In conclusion, the contributions of our paper can be summarized as follows: (1) We propose TAP, the first 3D-to-2D generative pre-training method that is adaptable to any point cloud model. TAP pre-training helps to exploit the potential of point cloud models on geometric structure comprehension and stereoscopic relation understanding. (2) We propose a Photograph Module where we derive mathematical formulations to encode pose conditions as query tokens in cross-attention layers. (3) TAP surpasses previous generative pre-training methods on the Transformers backbone and achieves state-of-the-art performance on ScanObjectNN classification and ShapeNetPart segmentation." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.846, + 0.217, + 0.861 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.871, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Point Cloud Analysis. Point cloud analysis is a fundamental and important task in the realm of 3D vision. Current" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.408 + ], + "angle": 0, + "content": "literature has developed two principal methodologies to extract structural representations from 3D point clouds, namely point-based and voxel-based methods. Point-based methods [32, 33, 48, 44, 35, 27] process unordered points directly, introducing various techniques for local information aggregation. Existing point-based methods can be categorized into three types: SetAbstraction-based [32, 33, 35], DynamicGraph-based [48], and Attention-based [57, 58, 30, 23, 60], all of which focus on modeling the relationships between points. Owing to their exceptional ability to effectively preserve fine-grained geometric information, point-based methods are frequently employed for object-level tasks. On the other hand, voxel-based methods [28, 21, 39] partition the 3D space into ordered voxels and employ 3D convolutions for feature extraction. Voxel-based methods are primarily based on SparseConvolution [10, 16], which enables efficient convolution operations in 3D space through sparse convolutions. In exchange for faster processing speeds, voxel-based methods sacrifice their capacity to capture detailed local structures, making them more suitable for large-scale scene-level tasks rather than object-level tasks." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.419, + 0.893, + 0.675 + ], + "angle": 0, + "content": "Point Cloud Pre-training. Pre-training has always been a focal point of research in the field of deep learning. Generally speaking, we usually distinguish pre-training methods based on the amount of annotation required, namely full-supervised pre-training [14, 59, 5], weakly-supervised pretraining [43, 4, 53], and unsupervised pre-training [19, 7, 8]. Among these methods, unsupervised pre-training has become the most popular approach, mainly due to its excellent transferability across tasks and its notable advantage of not relying on labeled data. Numerous researchers have proposed a variety of pretext tasks for unsupervised pre-training. Based on the type of pretext task employed, there are two prevailing pretext tasks for unsupervised pre-training. The first is contrastive learning, as exemplified by MoCo [19] and SimCLR [7]. The other method entails utilizing generative tasks to restore the data from partially or disrupted inputs, such as MAE [18] and BEiT [3]." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.679, + 0.893, + 0.83 + ], + "angle": 0, + "content": "Inspired by pre-training strategies in the image domain, there are more and more unsupervised pre-training methods being proposed for point cloud pre-training. PointContrast [54] embraces the principle of contrastive learning, whereas PointBERT [58] and PointMAE [30] integrate reconstruction pretext tasks. However, existing generative-based pre-training methods for point clouds only consider a single modality. In this paper, we propose a cross-modal generative-based pre-training strategy to achieve more effective pre-training." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.841, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Cross-Modal Learning. Recently, cross-modal learning has been a popular research topic, aiming at learning from multiple modalities such as images, audio and point clouds. It has the potential to enhance the performance of various tasks," + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.513, + 0.957 + ], + "angle": 0, + "content": "5641" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.092, + 0.892, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.237, + 0.897, + 0.285 + ], + "angle": 0, + "content": "Figure 2: The pipeline of TAP pre-training method. We design a query generator to encode pose conditions and implement cross-attention layers to transform 3D point cloud features \\( F_{3\\mathrm{d}} \\) to 2D view image features \\( F_{2\\mathrm{d}} \\) according to pose instruction. The predicted pose-dependent view image from 2D generator is supervised by ground truth view image via MSE loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.471, + 0.397 + ], + "angle": 0, + "content": "including visual recognition, speech recognition, and point cloud analysis. A variety of methods have been proposed for cross-modal learning, including multi-task learning [11], conditional generation [40], pre-training [34, 13] and tuning [49, 55]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.474, + 0.595 + ], + "angle": 0, + "content": "In the realm of point cloud analysis, much previous work has explored this learning paradigm. Some literature leverages 2D data for 3D point cloud analysis, such as MVTN [17], MVCNN [42] and CrossPoint [2], proving that the multi-view images and the correspondence between images and points can be helpful for the 3D object understanding. Another line of the research, such as Image2Point [55] and P2P [49] take effort to adapt the models from 2D vision into 3D point cloud analysis, fully exploiting the relationship of 2D and 3D understanding. In this paper, we continue this learning paradigm in 3D vision domain, and for the first time propose the cross-modal generative pre-training scheme for point cloud models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.61, + 0.385, + 0.628 + ], + "angle": 0, + "content": "3. 3D-to-2D Generative Pre-training" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.403, + 0.652 + ], + "angle": 0, + "content": "3.1. Preliminary: Generative Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Generative pre-training is a fundamental branch of pretraining methods that aims at reconstructing integral and complete data given partial or disrupted input. Mathematically, suppose \\( x \\) is a sample from raw data with no annotation. The pre-processing step \\( T(\\cdot) \\) either erases part of \\( x \\) randomly or splits \\( x \\) into pieces and intermingles them to get \\( \\tilde{x} = T(x) \\). The generative pre-training model \\( M \\) is designed to restore from those broken input \\( \\hat{x} = M(T(x)) \\) and the training loss function is designed to measure the reconstruction distance \\( \\mathcal{L} = D(\\hat{x},x) \\). In point cloud object analysis, earlier generative pre-training methods propose various pretext tasks as \\( T \\), including deformation [1], jigsaw puzzles [41] and depth projection [47] to produce disarrayed or partial point clouds. Recently, inspired by MAE [18] in the image domain, generative pre-training in 3D domain mainly focuses on implementing random masking as \\( T \\) and utilizing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.895, + 0.368 + ], + "angle": 0, + "content": "Transformers model as \\(M\\) for reconstruction [58, 30, 23, 60]. The reconstruction distance \\(D\\) is usually measured by the classical \\(l_{2}\\) Chamfer Distance:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.373, + 0.895, + 0.41 + ], + "angle": 0, + "content": "\\[\nD (\\hat {x}, x) = \\frac {1}{| \\hat {x} |} \\sum_ {a \\in \\hat {x}} \\min _ {b \\in x} \\| a - b \\| _ {2} ^ {2} + \\frac {1}{| x |} \\sum_ {b \\in x} \\min _ {a \\in \\hat {x}} \\| a - b \\| _ {2} ^ {2} (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.895, + 0.462 + ], + "angle": 0, + "content": "Besides Chamfer Distance between point clouds, some methods also exploit feature distance between latents [58] or occupancy value distance [23] as the loss function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.645 + ], + "angle": 0, + "content": "The exact reason why generative pre-training would help enhance the representation ability of backbone models still remains an open question. However, abundant experimental results have conveyed that predicting missing parts according to known parts demands high reasoning ability and global comprehension capacity of the model. What's more, generative pre-training is more efficient and suitable for point cloud object analysis than contrastive pre-training, given that contrastive pre-training typically requires a large amount of training data to avoid trivial overfitting solutions but there has always been a data-starvation problem in point cloud object research field." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.662, + 0.668 + ], + "angle": 0, + "content": "3.2. Overall Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.735 + ], + "angle": 0, + "content": "Different from the aforementioned generative pre-training methods that focus on uni-modal point cloud reconstruction, we propose a novel cross-modal pre-training approach of generating view images from instructed camera poses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.897, + 0.903 + ], + "angle": 0, + "content": "The overall architecture of our proposed TAP pre-training model is depicted in Figure 2. Our model takes as an input point cloud \\( P \\in \\mathbb{R}^{N \\times 3} \\), where \\( N \\) is the number of points in the input point cloud. The basic building block of TAP mainly consists of: 1) a 3D Backbone that extracts 3D geometric features \\( F_{3\\mathrm{d}} \\in \\mathbb{R}^{n \\times C_{3\\mathrm{d}}} \\), where \\( n \\) is the number of downsampled center points and \\( C_{3\\mathrm{d}} \\) is the geometric feature dimension; 2) a pose-dependent Photograph Module that takes as inputs \\( F_{3\\mathrm{d}} \\) and pose matrix \\( R \\in \\mathbb{R}^{3 \\times 3} \\), and predicts view image features \\( F_{2\\mathrm{d}}^{R} \\in \\mathbb{R}^{h \\times w \\times C_{2\\mathrm{d}}} \\) conditioned on \\( R \\), where \\( h, w \\) are height and width of predicted view image" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5642" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.137 + ], + "angle": 0, + "content": "feature map; 3) an \\(2D\\) Generator that decodes \\(F_{2\\mathrm{d}}^{R}\\) into an RGB image \\(I_{\\mathrm{gen}}^{R} \\in \\mathbb{R}^{H \\times W \\times 3}\\), where \\(H, W\\) are height and width of the output view image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.471, + 0.304 + ], + "angle": 0, + "content": "As we place no restriction on \\( F_{3\\mathrm{d}} \\), the 3D Backbone can be arbitrarily chosen and adopted. Therefore, our TAP is more flexible and compatible than existing generative pre-training methods that are limited to Transformer-based architecture. Experimental results in Section 4 will later verify that TAP brings consistent improvement to all kinds of point cloud models. The technical designs of the pose-dependent Photograph Module will be thoroughly discussed in Section 3.3. The 2D Generator consists of four Transpose Convolution layers to progressively upsample image resolution and decode RGB colors of each pixel." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.314, + 0.269, + 0.33 + ], + "angle": 0, + "content": "3.3. Photograph Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.471, + 0.384 + ], + "angle": 0, + "content": "Architectural Design. As illustrated in Figure 2, we leverage cross-attention mechanism from Transformers [46] to build our pose-dependent Photograph Module." + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.395, + 0.47, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.576 + ], + "angle": 0, + "content": "where \\(d_{k}\\) is the scaling factor, and \\(Q,K,V\\) are queries, keys and values matrix. More specifically, we design a Query Generator \\(\\Phi\\) to encode camera pose conditions into query tokens: \\(Q = \\Phi (R)\\in \\mathbb{R}^{hw\\times C_{2\\mathrm{d}}}\\). We also design a Memory Builder \\(\\Theta\\) to construct \\(K\\) and \\(V\\) from 3D geometric features: \\(K = V = \\Theta (F_{3\\mathrm{d}})\\in \\mathbb{R}^{m\\times C_{2\\mathrm{d}}}\\), where \\(m\\) is the number of memory tokens. The output sequence of the cross attention layers will be rearranged from \\(hw\\times C_{2\\mathrm{d}}\\) to \\(h\\times w\\times C_{2\\mathrm{d}}\\), forming the predicted view image features \\(F_{2\\mathrm{d}}^R\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.471, + 0.848 + ], + "angle": 0, + "content": "During the cross-attention calculation process, we do not explicitly provide any projection clues of which 3D points would project to which 2D pixel. Instead, the Photograph Module learns by itself how to arrange unordered 3D feature points to ordered 2D pixel grids, purely based on semantic similarities between 3D geometric features and our delicately-designed queries that reveal pose information. Since one sample will only have one set of memory tokens in 3D space but its view images from different poses are quite distinct from each other, learning to predict precise view images from instructed poses in a data-driven manner is not a trivial task. Therefore, during the end-to-end optimization process, the 3D backbone is trained to have a stronger perception of the object's overall geometric structure and gain a higher representative ability of the stereoscopic relations. In this way, our proposed 3D-to-2D generative pre-training would help exploit the potential and enhance the strength of 3D backbone models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Query Generator. The query generator \\(\\Phi\\) is designed to encode pose condition \\(R\\) into 2D grid of shape \\(h \\times w\\). In object analysis, common practice is leveraging parallel light" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.198 + ], + "angle": 0, + "content": "shading to project 3D objects onto 2D grids, and pose matrix \\( R \\) here is used to rotate objects into desired angles before projection. Therefore, each 2D grid actually represents an optical line that starts from infinity, passes through 3D objects and ends at the 2D plane. As a consequence, we choose the direction and the origin points that the optical line goes through as the delegate of the query grid." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.894, + 0.273 + ], + "angle": 0, + "content": "Before deriving formulations of optical lines for each grid, let us first revisit the parallel light shading process for better comprehension. Given 3D coordinates \\(\\mathbf{x} = (x,y,z)\\) of a point cloud \\(P\\) and pose matrix \\(R\\), rotation is first performed to align the object to the ideal pose position:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.282, + 0.892, + 0.298 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\prime} = \\left(x ^ {\\prime}, y ^ {\\prime}, z ^ {\\prime}\\right) = R \\mathbf {x} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.338 + ], + "angle": 0, + "content": "Then we just omit the final dimension \\(z^{\\prime}\\) and evenly split the first two dimensions \\((x^{\\prime},y^{\\prime})\\) into 2D grids \\((u,v)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.346, + 0.892, + 0.379 + ], + "angle": 0, + "content": "\\[\nu = \\frac {x ^ {\\prime} - x _ {0}}{g _ {h}} + o _ {h}, \\quad v = \\frac {y ^ {\\prime} - y _ {0}}{g _ {w}} + o _ {w} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.892, + 0.461 + ], + "angle": 0, + "content": "where \\((x_0, y_0)\\) is the minimum value of \\((x', y')\\), \\((g_h, g_w)\\) is the grid size, \\((o_h, o_w)\\) is the offset value to place the projected object at the center of the image. \\(0 \\leq u \\leq h - 1\\), \\(0 \\leq v \\leq w - 1\\) and \\((u, v)\\) is a sampled pixel coordinate from the 2D grid." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.537 + ], + "angle": 0, + "content": "Now let us begin to derive formulations of the optical line that passes through the query grid. We only know \\((u, v)\\) for each grid and we want to reversely trace which 3D points \\((x, y, z)\\) are on the same optical line during parallel light projection. According to Eq. 4:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.545, + 0.892, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x ^ {\\prime} = g _ {h} u + x _ {0} - o _ {h} = \\Psi_ {h} (u) \\\\ y ^ {\\prime} = g _ {w} v + y _ {0} - o _ {w} = \\Psi_ {w} (v) \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.622 + ], + "angle": 0, + "content": "If we denote \\( A = R^{-1} \\) and \\( A_{ij} \\) as the element at \\( i^{th} \\) row and \\( j^{th} \\) column, then according to Eq. 3:" + }, + { + "type": "equation", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.646 + ], + "angle": 0, + "content": "\\[\nx = A _ {1 1} \\Psi_ {h} (u) + A _ {1 2} \\Psi_ {w} (v) + A _ {1 3} z ^ {\\prime} = \\Omega_ {x} (u, v) + A _ {1 3} z ^ {\\prime}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.649, + 0.892, + 0.665 + ], + "angle": 0, + "content": "\\[\ny = A _ {2 1} \\Psi_ {h} (u) + A _ {2 2} \\Psi_ {w} (v) + A _ {2 3} z ^ {\\prime} = \\Omega_ {y} (u, v) + A _ {2 3} z ^ {\\prime}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.668, + 0.892, + 0.697 + ], + "angle": 0, + "content": "\\[\nz = A _ {3 1} \\Psi_ {h} (u) + A _ {3 2} \\Psi_ {w} (v) + A _ {3 3} z ^ {\\prime} = \\Omega_ {z} (u, v) + A _ {3 3} z ^ {\\prime} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.894, + 0.864 + ], + "angle": 0, + "content": "According to the definition of line's parametric equation, Eq. 6 represents a line passing through the origin point \\( O: (\\Omega_x(u,v), \\Omega_y(u,v), \\Omega_z(u,v)) \\) with optical line direction \\( \\mathbf{d} = (A_{13}, A_{23}, A_{33}) \\), where \\( \\Omega_x, \\Omega_y, \\Omega_z \\) are xyz coordinates of \\( O \\) and their formulations are conditioned on \\( u, v \\). Therefore, we concatenate the coordinate of origin point \\( O \\), normalized direction \\( \\mathbf{d}^{\\dagger} = \\mathbf{d} / \\| \\mathbf{d}\\|_2 \\) and normalized position \\( (u/h, v/w) \\) as positional embedding together to be the initial state of our query. A multi-layer-perceptron (MLP) module is later leveraged to map the 8-dim initial query to higher dimensional space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Memory Builder. The memory builder takes \\( F_{3\\mathrm{d}} \\) as input to prepare for initial state of \\( K, V \\) in cross-attention layers. We" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5643" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.122 + ], + "angle": 0, + "content": "first concatenate aligned 3D coordinate \\(P_{\\mathrm{3d}}\\) with 3D features to enhance the geometric knowledge of \\(F_{\\mathrm{3d}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.133, + 0.471, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\hat {F} _ {3 \\mathrm {d}} = \\operatorname {M L P} \\left(\\operatorname {c a t} \\left(F _ {3 \\mathrm {d}}, P _ {3 \\mathrm {d}}\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.163, + 0.471, + 0.316 + ], + "angle": 0, + "content": "Additionally, we initialize a learnable memory token \\( T_{\\mathrm{pad}} \\) as the pad token and concatenate it with \\( \\hat{F}_{3\\mathrm{d}} \\) to obtain the initial state of \\( K, V \\). The reason for concatenating a learnable pad token \\( T_{\\mathrm{pad}} \\) is that there are white background areas on the projected image (as shown in Figure 2). As \\( F_{3\\mathrm{d}} \\) only encodes foreground objects, we further need a learnable pad token to represent background regions. Otherwise, the cross-attention layers will be confused to learn how to combine foreground tokens into background features and this will inevitably diminish the pre-training effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.325, + 0.262, + 0.342 + ], + "angle": 0, + "content": "3.4. Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.471, + 0.471 + ], + "angle": 0, + "content": "We perform per-pixel supervision with Mean Squared Error (MSE) loss between generated view image \\( I_{\\mathrm{gen}}^{R} \\) and ground truth image \\( I_{\\mathrm{gt}}^{R} \\), aligned by camera pose \\( R \\). For simplicity, we will omit \\( R \\) in later formulations. As the background of the rendered ground truth images is all white and reveals little information, we further design a compound loss to balance the weight between foreground regions and background regions:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.482, + 0.47, + 0.501 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(I _ {\\text {g e n}}, I _ {\\text {g t}}\\right) = w ^ {\\mathrm {f g}} \\mathcal {D} ^ {\\mathrm {f g}} + w ^ {\\mathrm {b g}} \\mathcal {D} ^ {\\mathrm {b g}} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.512, + 0.47, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} ^ {k} \\left(I _ {\\text {g e n}} ^ {k}, I _ {\\mathrm {g t}} ^ {k}\\right) = \\frac {1}{H W} \\sum_ {h, w} \\left(I _ {\\text {g e n}} ^ {k} (h, w) - I _ {\\mathrm {g t}} ^ {k} (h, w)\\right) ^ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.47, + 0.618 + ], + "angle": 0, + "content": "where \\( k = \\mathrm{fg} \\) (foreground), \\( \\mathrm{bg} \\) (background) and \\( w^{\\mathrm{fg}} \\), \\( w^{\\mathrm{bg}} \\) are loss weights for foreground and background, respectively. Such per-pixel supervision is more precise than the ambiguous set-to-set Chamfer Distance introduced in Eq. 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.631, + 0.21, + 0.648 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.471, + 0.779 + ], + "angle": 0, + "content": "In this section, we first introduce the setups of our pretraining scheme. Then we evaluate our pre-training method on various point cloud backbones by fine-tuning them on different downstream tasks, such as point cloud classification on ModelNet and ScanObjectNN datasets, and part segmentation on the ShapeNetPart dataset. Finally, we provide in-depth ablation studies for the architectural design of our proposed TAP pre-training pipeline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.265, + 0.804 + ], + "angle": 0, + "content": "4.1. Pre-training Setups" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Data Setups To align with previous research practices [58, 30, 54], we choose ShapeNet [6] that contains more than 50 thousand CAD models as our pre-training datasets. We sampled 1024 points from each 3D CAD model to form the point clouds, consistent with previous work. Since ShapeNet does not provide images for each point cloud, we use the" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.136 + ], + "angle": 0, + "content": "Table 1: Classification results on the ScanObjectNN dataset. We report the overall accuracy \\((\\%)\\). The results with \\(\\dagger\\) are reproduced by PointNeXt [35] repository." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.147, + 0.892, + 0.384 + ], + "angle": 0, + "content": "
MethodOBJ_BGOBJ_ONLYPB_T50_RS
Hierarchical Models with TAP Pre-training
†DGCNN [48]--86.1
+ TAP--86.6 (+0.5)
†PointNet++ [33]--86.2
+ TAP--86.8 (+0.6)
†PointMLP [27]--87.4
+ TAP--88.5 (+1.1)
Standard Transformers with Generative Pre-training
w/o pre-training [46]79.8680.5577.24
OcCo [47]84.8585.5478.79
Point-BERT [58]87.4388.1283.07
MaskPoint [23]89.3088.1084.30
Point-MAE [30]90.0288.2985.18
TAP (Ours)90.3689.5085.67
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.478 + ], + "angle": 0, + "content": "rendered image from 12 surrounding viewpoints generated by MVCNN [42]. During our pre-training, the models are exclusively pre-trained with the training split following the practice of previous work [58]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.895, + 0.773 + ], + "angle": 0, + "content": "Architecture Setups We conduct experiments on various point cloud encoders, including PointNet++ [33], DGCNN [48], PointMLP [27] and Transformers [58] for point cloud object classification. During the pre-training stage, the photograph module takes encoded point cloud features and pose conditions as inputs to generate a \\(32 \\times\\) downsampled view image feature map of size \\(7 \\times 7\\) from a specific viewpoint. Then the 2D generator progressively upsamples the image feature map to decode RGB view images of size \\(224 \\times 224\\). We do not alter the architecture of the point cloud backbone since the photograph module and the 2D generator are exclusively used during the pretraining phase and are dropped during the fine-tuning stage. In our experiment, the photograph module is a six-layer cross-attention block, with attention layer channels limited to 256 to enhance efficiency. During the pre-training task on ShapeNet, we utilized four simple transpose convolutions to upsample the reconstructed 2D feature map and predict the RGB value for each pixel." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Implementation Details The experiments of TAP pretraining and finetuning on various downstream tasks are implemented with PyTorch [31]. We utilize AdamW [26] optimizer and the CosineAnnealing learning rate scheduler [25] to pre-train the point cloud backbone for 100 epochs. We set the initial learning rate as \\(5e^{-4}\\) and weight decay as \\(5e^{-2}\\). In our experiment, we train various point cloud backbones with 32 batch sizes on a single Nvidia 3090Ti GPU. The" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5644" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.473, + 0.211 + ], + "angle": 0, + "content": "Table 2: Results comparisons with previous methods on the ScanObjectNN and ModelNet40 datasets. The model parameters number (#Params) and overall accuracy (\\%) are reported. † denotes our reproduced results of PointMLP on the ModelNet40 dataset. Methods with * introduce extra knowledge from pre-trained image models or pre-trained vision-language models. We do not compete with them for fair comparison and only list them for reference." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.223, + 0.468, + 0.613 + ], + "angle": 0, + "content": "
Method#Params (M)ScanObjectNNModelNet40
Supervised Learning Only
PointNet [32]3.568.089.2
PointNet++ [33]1.577.990.7
Transformer [46]22.177.2491.4
DGCNN [48]1.878.192.9
PointCNN [22]0.678.592.2
DRNet [36]-80.393.1
SimpleView [15]-80.5±0.393.9
GBNet [37]8.881.093.8
PRA-Net [9]2.381.093.7
MVTN [17]11.282.893.8
RepSurf-U [38]1.584.394.4
PointMLP [27]12.685.4±0.393.5†
PointNeXt [35]1.487.7±0.493.7
Transformers with Pre-training
OcCo [47]22.178.892.1
Point-BERT [58]22.183.193.2
MaskPoint [23]22.184.393.8
Point-MAE [30]22.185.293.8
Point-M2AE [60]15.386.494.0
With Pre-trained Image Model
ACT [13]*22.188.293.7
I2P-MAE [61]*-90.193.7
ReCon [34]*44.390.694.1
Our Proposed TAP Pre-training
PointMLP + TAP12.688.594.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.472, + 0.794 + ], + "angle": 0, + "content": "drop path rate of the cross-attention layer is set to 0.1. The foreground and background loss weights \\( w^{\\mathrm{fg}} \\), \\( w^{\\mathrm{bg}} \\) are set to 20 and 1. The detailed architecture of our simple 2D generator is: TConv(256, 128, 5, 4) → TConv(128, 64, 3, 2) → TConv(64, 32, 3, 2) → TConv(32, 3, 3, 2), where TConv stands for Transpose Convolution and the four numbers in the tuple denotes \\( (C_{in}, C_{out}, \\text{Kernel}, \\text{Stride}) \\) respectively. During the fine-tuning stage, we perform a learning rate warming up for point cloud backbones with 10 epochs, and keep other settings unchanged for a fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.26, + 0.817 + ], + "angle": 0, + "content": "4.2. Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In this section, we report the experimental results of various downstream tasks. We follow the previous work to conduct experiments of object classification on real-world ScanObjectNN and synthetic ModelNet40 datasets. We also verify the effectiveness of our pre-training method on the" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.095, + 0.892, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.222, + 0.897, + 0.329 + ], + "angle": 0, + "content": "Figure 3: The visualization results of our proposed 3D-to-2D generative pre-training. The first row displays view images generated by our TAP pre-training pipeline and the second row shows ground truth images. Our TAP can produce view images with appropriate shapes and reflection colors, demonstrating its ability in capturing geometric structure and stereoscopic knowledge." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.361, + 0.856, + 0.377 + ], + "angle": 0, + "content": "part segmentation task with the ShapeNetPart dataset." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.392, + 0.7, + 0.407 + ], + "angle": 0, + "content": "4.2.1 Object Classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.895, + 0.583 + ], + "angle": 0, + "content": "Main Results. To evaluate the effectiveness of our proposed TAP, we implement it with various point cloud architectures, including classical baselines such as PointNet++, DGCNN, and PointMLP, as well as widely used Standard Transformers backbone for existing generative pre-training methods. We follow the common practice to experiment our model on three variants of the ScanObjectNN dataset: 1) OBJ-ONLY: cropping object without any background; 2) OBJ-BG: containing the background and object; 3) PB-T50-RS: adopting various augmentations to the objects. We reported the results comparing with existing pre-training methods in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.897, + 0.903 + ], + "angle": 0, + "content": "As shown in the upper part of the table, we pretrain the graph-based architecture DGCNN, set-abstraction-based architecture PointNet++, and MLP-based architecture PointMLP with TAP, and observe consistent improvements across models. The results strongly convey that our proposed TAP can be successfully applied to various types of point cloud models and the proposed novel 3D-to-2D generative pre-training is effective regardless of the backbone architecture. Considering that nearly all existing generative pre-training methods are specially designed for Transformer-based architecture, our TAP is much superior in its wider adaptation and higher flexibility. Additionally, we also provide a detailed and fair comparison with previous work by implementing TAP with the Standard Transformers architecture in the lower part of the table, where no hierarchical designs or inductive bias is included. Our TAP outperforms previous pre-training methods in all three split settings, providing strong evidence that our 3D-to-2D generative pretraining strategy can also benefit attention-based architectures and surpass uni-modal generative pre-training competitors. It is worth noting that although TAP and many previous" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5645" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.121 + ], + "angle": 0, + "content": "Table 3: Part segmentation results on the ShapeNetPart dataset. We report the mean IoU across all part categories \\(\\mathrm{mIoU}_C\\), the mean IoU across all instances \\(\\mathrm{mIoU}_I\\), and the IoU for each category." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.125, + 0.895, + 0.326 + ], + "angle": 0, + "content": "
MethodsmIoUCmIoUIaerobagcapcarchairearphoneguitarknifelamplaptopmotormugpistolrocketskateboardtable
Supervised Representation Learning Only
PointNet [32]80.483.783.478.782.574.989.673.091.585.980.895.365.293.081.257.972.880.6
PointNet++ [33]81.985.182.479.087.777.390.871.891.085.983.795.371.694.181.358.776.482.6
DGCNN [48]82.385.284.083.486.777.890.674.791.287.582.895.766.394.981.163.574.582.6
PointMLP [27]84.686.183.583.487.580.590.378.292.288.182.696.277.595.885.464.683.384.3
KPConv [44]85.186.484.686.387.281.191.177.892.688.482.796.278.195.885.469.082.083.6
Transformers with Uni-modal Generative Pre-training
Point-BERT [58]84.185.684.384.888.079.891.081.791.687.985.295.675.694.784.363.476.381.5
Point-MAE [30]84.286.184.385.088.380.591.378.592.187.486.196.175.294.684.763.577.182.4
MaskPoint [23]84.486.084.285.688.180.391.279.591.987.886.295.376.995.085.364.476.981.8
Point-M2AE [60]84.986.5----------------
Our Proposed 3D-to-2D Generative Pre-training
PointMLP+TAP85.286.984.886.189.582.592.175.992.388.785.696.579.896.085.966.278.183.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.473, + 0.44 + ], + "angle": 0, + "content": "pre-training approaches have significantly improved the performance of Transformers in point cloud tasks, increasing the accuracy from 77.24 to 85.67, they still lag far behind advanced point cloud networks such as PointMLP. Therefore, TAP's applicability to various models is an important characteristic that would benefit future research." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.473, + 0.734 + ], + "angle": 0, + "content": "Comparisons with Previous Methods. To clearly demonstrate the high performance of our proposed TAP pre-training in object classification tasks, we compare TAP with existing methods on both synthetic ModelNet and real-world ScanObjectNN (the hardest PB-T50-RS variant) datasets in Table 2. We categorize existing methods into two types: (1) architecture-oriented methods, which focus on developing novel model architectures for 3D point clouds and do not involve any pre-training techniques, and (2) pre-training methods, which pay more attention to the pre-training strategy and whose backbone model are mostly limited to Transformer-base architecture. It's worth noticing that methods marked with an asterisk (*) incorporate additional knowledge from pre-trained image models or pre-trained vision-language models. To ensure a fair and unbiased comparison, we refrain from directly comparing our method with these approaches. However, we include them in the listing for reference purposes, acknowledging their existence and potential relevance in related research." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.473, + 0.903 + ], + "angle": 0, + "content": "From the experimental results, we can see that accompanied by PointMLP backbone model, our proposed TAP pre-training achieves the best classification accuracy on ScanObjectNN and ModelNet40 among existing models (with no pre-trained knowledge from image or language like P2P [49]), demonstrating the effectiveness of our approach and validating the superiority of our 3D-to-2D cross-modal generative pre-training method over previous generative pretraining methods. Moreover, we also note that our proposed method has brought higher performance improvements on the ScanObjectNN dataset than on ModelNet40. This may" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.895, + 0.426 + ], + "angle": 0, + "content": "be attributed to the reason that the cross-modal generative pre-training has enhanced the network's ability to understand point clouds from different views, which is beneficial for a more robust understanding of the real-scan data with more noise and disturbance in the ScanObjectNN dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.897, + 0.555 + ], + "angle": 0, + "content": "Visualization Results. Figure 3 shows the visualization results of TAP. The first row shows the generated view images while the second row displays the ground truth images for reference. The TAP method can successfully predict the accurate shape of the object and the RGB colors that represent light reflections in rendered images. Therefore, TAP is capable of capturing the geometric structure of 3D objects and reasoning occlusion relations from specific camera poses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.685, + 0.59 + ], + "angle": 0, + "content": "4.2.2 Part Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.78 + ], + "angle": 0, + "content": "Performing dense prediction is always a more challenging task compared with classification. In this section, we evaluate the local distinguishability of our proposed TAP pre-training method, fine-tuning the pre-trained point cloud model on the ShapeNetPart dataset for the part segmentation task. Quantitative results are shown in Table 3. We implement PointMLP as the backbone model and compare our TAP results with two mainstreams of previous literature. The upper row displays classical architecture-oriented methods that focus on network design and are trained from scratch. The lower row shows members of the generative pre-training family that rely on Transformer-based architectures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.897, + 0.903 + ], + "angle": 0, + "content": "According to results comparisons, our TAP pre-training significantly improves the part segmentation performance of the PointMLP backbone, increasing class mIoU by 0.6 and instance mIoU by 0.8. More importantly, our TAP pretraining achieves state-of-the-art performance on both class mIoU and instance mIoU, surpassing leading works in both tracks. Specifically, TAP exceeds the performance of PointM2AE on instance mIoU by 0.4. This satisfactory perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5646" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.319 + ], + "angle": 0, + "content": "mance serves as strong evidence to convey that our proposed TAP pre-training is superior to previous uni-model generative pre-training mechanisms in dense prediction tasks. This may be attributed to the factor that our supervision in 2D with MSE loss is more precise than the ambiguous Chamfer Distance in 3D reconstruction. Therefore, models with TAP pre-training obtain more accurate comprehension of local geometry and detail awareness, which contributes to mIoU gain in dense prediction tasks. What's more, TAP outperforms KPCov on instance mIoU by 0.5, demonstrating that the proposed 3D-to-2D generative pre-training method can fully exploit the potential of the point cloud model and help it have a better perception of objects' geometric structure. As TAP is adaptable to any architecture, future improvements in architectural design can also benefit from TAP pre-training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.34, + 0.293, + 0.355 + ], + "angle": 0, + "content": "4.2.3 Few-shot Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.471, + 0.472 + ], + "angle": 0, + "content": "Following Point-BERT [58], we conduct few-shot classification with Standard Transformers on ModelNet40 [52] dataset. As shown in Table 4, we report mean overall accuracy and standard deviation \\((\\mathrm{mOA} \\pm \\mathrm{std})\\) on 10 pre-defined data folders for each few-shot setting. The way and shot in Table 4 specify the number of categories and the number of training examples per category, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.471, + 0.653 + ], + "angle": 0, + "content": "From the results, TAP achieves the highest mean overall accuracy across all few-shot settings when compared to previous generative pre-training approaches. Furthermore, TAP exhibits significantly lower standard deviations than those reported in the existing literature for the majority of few-shot settings, which signifies its robust performance and consistent superiority. This indicates that TAP is not only capable of achieving high mean overall accuracy but also exhibits reliability and robustness across various few-shot settings. Such stability is crucial in real-world applications, where consistency and predictability are vital for practical deployment." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.34, + 0.68 + ], + "angle": 0, + "content": "4.3. Scene-level Dense Predictions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.884 + ], + "angle": 0, + "content": "To assess the effectiveness of TAP in handling scene-level dense prediction tasks, we carry out experiments on more complicated scene-level object detection and semantic segmentation on the ScanNetV2 [12] dataset. For the object detection task, we adopt 3DETR [29] and pre-train its encoder on the object-level dataset ShapeNet [6] with TAP. Average precision at 0.25 and 0.5 IoU thresholds are reported. Regarding semantic segmentation, we employ the PointTransformerV2 [51](PTv2) model and pre-train it on the ScanNetV2 dataset with TAP. We report mean IoU for evaluation metric. It is worth mentioning that PTv2 represents the current state-of-the-art approach with open-source code availability." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Based on the results presented in Table 5, TAP consis" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.895, + 0.151 + ], + "angle": 0, + "content": "Table 4: Few-shot Classification with Standard Transformers on ModelNet40 dataset. We report mean overall accuracy and standard deviation on 10 pre-defined data folders for each setting. Best results are marked bold." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.161, + 0.891, + 0.29 + ], + "angle": 0, + "content": "
Method5-way10-way
10-shot20-shot10-shot20-shot
w/o pre-training87.8 ± 5.293.3 ± 4.384.6 ± 5.589.4 ± 6.3
Point-BERT [58]94.6 ± 3.196.3 ± 2.791.0 ± 5.492.7 ± 5.1
MaskPoint [23]95.0 ± 3.797.2 ± 1.791.4 ± 4.093.4 ± 3.5
Point-MAE [30]96.3 ± 2.597.8 ± 1.892.6 ± 4.195.0 ± 3.0
TAP97.3 ± 1.897.8 ± 1.793.1 ± 2.695.8 ± 1.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.302, + 0.895, + 0.379 + ], + "angle": 0, + "content": "Table 5: Scene-level object detection and semantic segmentation on ScanNetV2 [12]. Average precision at 0.25 IoU thresholds \\((\\mathrm{AP}_{0.25})\\) and 0.5 IoU thresholds \\((\\mathrm{AP}_{0.5})\\) of detection and mean Intersection-over-Union (mIoU) of semantic segmentation are reported." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.389, + 0.892, + 0.473 + ], + "angle": 0, + "content": "
MethodDet (3DETR [29])Seg (PTv2 [51])
AP0.25AP0.5mIoU
Baseline62.137.972.4
+TAP63.0 (+0.9)41.4 (+3.5)72.6 (+0.2)
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.895, + 0.68 + ], + "angle": 0, + "content": "tently enhances the performance of all baselines, thereby showcasing its efficacy in tackling more intricate scene-level dense prediction tasks. Remarkably, even with the encoder solely pre-trained on an object-level dataset for scene-level detection task, significant improvements are observed in both \\(\\mathrm{AP}_{0.25}\\) and \\(\\mathrm{AP}_{0.5}\\) metrics. This suggests that the learned representations from TAP effectively capture relevant information and generalize well to complex scenes, even when the pre-training data is limited to object-level collections. Such generalization capabilities are valuable in scenarios where obtaining large-scale fully annotated scene-level datasets may be challenging or expensive." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.689, + 0.665, + 0.704 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.895, + 0.773 + ], + "angle": 0, + "content": "To investigate the architectural design of our proposed Photograph Module in TAP pre-training pipeline, we conduct extensive ablation studies on the ScanObjectNN dataset with PointMLP as the backbone model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Photograph Module Architectural Designs. In Photograph Module, we implement cross-attention layers to generate view image feature maps conditioned on pose instruction. We believe that letting the module learn by itself how to rearrange 3D point features in 2D grids will enhance the representation ability of the 3D backbone. Therefore, we conduct ablation studies to verify this hypothesis. As shown in Table 6a, we implement Model \\(\\mathrm{A}_1\\) with no attention lay" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5647" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.895, + 0.219 + ], + "angle": 0, + "content": "Table 6: Ablation studies on Photograph module in TAP pre-training pipeline. We choose PointMLP as the backbone model and conduct ablation studies on the ScanObjectNN dataset from two aspects: overall architectural designs and query designs. In Table(a), we first investigate the effectiveness of cross-attention design compared with direct projection (Model A1) and direct project with self-attention (Model A2). Then we analyze the influence of the different number of attention layers and feature channels in Model B and Model C. Finally, we discuss whether pad token in memory builder is beneficial in Model D. In Table(b), we conduct further experiments to compare different approaches for query designing: (1) Using learnable query based on the given viewpoints (Model E) or use mathematical formulations derived in Eq. 6. (2) The information we need when we mathematically encode pose information into init query status: (i) Origin: the coordinate of origin point \\( O \\) that the optical line passes through. (ii) Direction: The normalized direction of the optical line. (iii) PE: The position embedding for each grid." + }, + { + "type": "table_caption", + "bbox": [ + 0.197, + 0.222, + 0.4, + 0.235 + ], + "angle": 0, + "content": "(a) Overall Architectural Designs." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.236, + 0.52, + 0.361 + ], + "angle": 0, + "content": "
ModelAttention TypeLayerNum.ChannelsMem.PadAcc.(%)
A1None-256X87.6 (-0.9)
A2SelfAttn6 layers256X87.8 (-0.7)
BCrossAttn2 layers256✓87.9 (-0.6)
CCrossAttn6 layers512✓87.8 (-0.7)
DCrossAttn6 layers256X88.3 (-0.2)
TAPCrossAttn6 layers256✓88.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.654, + 0.221, + 0.768, + 0.235 + ], + "angle": 0, + "content": "(b) Query Designs." + }, + { + "type": "table", + "bbox": [ + 0.535, + 0.235, + 0.894, + 0.363 + ], + "angle": 0, + "content": "
ModelQuery TypeOriginDirectionPEAcc.(%)
ELearnableXXX87.5 (-1.0)
F1FormulaX✓✓86.5 (-2.0)
F2Formula✓XX87.8 (-0.7)
F3Formula✓✓X88.0 (-0.5)
F4Formula✓X✓88.1 (-0.4)
TAPFormula✓✓✓88.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.473, + 0.599 + ], + "angle": 0, + "content": "ers, directly projecting 3D feature points to 2D grids based on Eq. 3 and Eq. 4. This results in a much simpler pretraining task, as the projection relation has been directly told. Additionally, in Model \\(\\mathrm{A}_2\\), we add self-attention layers after explicit projection to help the model capture longer-range correlations. Pose knowledge is encoded as a pose token that is concatenated to projected grids, similar to the CLS token in classification Transformers. According to quantitative results comparison with TAP that implements cross-attention layers, fine-tuning results of pre-training methods in \\(\\mathrm{A}_1\\) and \\(\\mathrm{A}_2\\) version show inferiority. Therefore, the cross-attention architecture we designed to entirely LEARN the projection relation is the most suitable choice for the proposed 3D-to-2D generative pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.473, + 0.725 + ], + "angle": 0, + "content": "What's more, we discuss the number of cross-attention layers, the dimension of feature channels and whether to concatenate pad token in memory builder in Model \\( B, C, D \\). According to the results, more cross-attention layers show stronger representation ability, while too large channel number will lead to performance decrease caused by over-fitting. The performance gain from Model \\( D \\) to TAP also verifies that the pad token design in the memory builder is essential." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Query Generator Designs. In the query generator, we derive the mathematical formulation of the optical lines passing through 2D grids. We propose to concatenate the coordinate of origin point \\( O \\), normalized direction \\( \\mathbf{d}^{\\dagger} \\) and position embedding \\( (u / h, v / w) \\) as the initial state of queries. In Table 6b, we first compare this mathematical design with totally learnable queries that takes pose matrix \\( R \\) as input and implements MLP layers to predict query for each grid. As shown in Model \\( E \\), learnable queries cannot satisfactorily encode pose information, while our derived formulation for query construction is both clearer in physical meaning and" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.388, + 0.779, + 0.403 + ], + "angle": 0, + "content": "more competitive in fine-tuning accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.895, + 0.48 + ], + "angle": 0, + "content": "In ablation \\(\\mathrm{F_1}\\) to \\(\\mathrm{F_4}\\), we progressively discuss the three components of query generation. Quantitative comparison with TAP verifies that every component is indispensable for query generation, where coordinates of origin points are of the most importance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.628, + 0.515 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.897, + 0.797 + ], + "angle": 0, + "content": "In this paper, we have proposed a novel 3D-to-2D generative pre-training method TAP that is adaptable to any point cloud model. We implemented the cross-attention mechanism to generate view images of point clouds from instructed camera poses. To better encode pose conditions and generate physically meaningful queries, we derived mathematical formulations of optical lines. The proposed TAP pre-training had higher preciseness in supervision and broader adaptation to different backbones, compared with directly reconstructing point clouds in previous methods. Experimental results conveyed that the TAP pre-training can help the backbone models better capture the structural knowledge and stereoscopic relations. Fine-tuning results of TAP pre-training achieve state-of-the-art performance on ScanObjectNN classification and ShapeNetPart segmentation, among methods that do not include any pre-trained image or text models. We believe the cross-modal generative pre-training paradigm will be a promising direction for future research." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.81, + 0.661, + 0.827 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.835, + 0.895, + 0.896 + ], + "angle": 0, + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0114903 and in part by the National Natural Science Foundation of China under Grant 62125603." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5648" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Idan Achituve, Haggai Maron, and Gal Chechik. Self-supervised learning for domain adaptation on point clouds. In WACV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In CVPR, pages 9902-9912, June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.27 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. *NeruIPS*, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.471, + 0.428 + ], + "angle": 0, + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.471, + 0.529 + ], + "angle": 0, + "content": "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.471, + 0.572 + ], + "angle": 0, + "content": "[9] Silin Cheng, Xiwu Chen, Xinwei He, Zhe Liu, and Xiang Bai. Pra-net: Point relation-aware network for 3d point cloud analysis. IEEE TIP, 30:4436-4448, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.47, + 0.614 + ], + "angle": 0, + "content": "[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.47, + 0.658 + ], + "angle": 0, + "content": "[11] Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning. In ICML, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[13] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In ICML, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[17] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. MVTN: multi-view transformation network for 3d shape recognition. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.894, + 0.221 + ], + "angle": 0, + "content": "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll'ar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.894, + 0.264 + ], + "angle": 0, + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.267, + 0.894, + 0.321 + ], + "angle": 0, + "content": "[20] Tianxin Huang, Zhonggan Ding, Jiangning Zhang, Ying Tai, Zhenyu Zhang, Mingang Chen, Chengjie Wang, and Yong Liu. Learning to measure the point cloud reconstruction loss in a representation space. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.894, + 0.364 + ], + "angle": 0, + "content": "[21] Roman Klokov and Victor Lempitsky. Escape from cells: Deep kd-networks for the recognition of 3d point cloud models. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.368, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[22] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. NeurIPS, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.894, + 0.452 + ], + "angle": 0, + "content": "[23] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[24] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and Shi-Min Hu. Morphing and sampling network for dense point cloud completion. In AAAI, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.541, + 0.892, + 0.569 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.571, + 0.892, + 0.612 + ], + "angle": 0, + "content": "[27] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[28] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In IROS. IEEE, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.658, + 0.892, + 0.698 + ], + "angle": 0, + "content": "[29] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[30] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.894, + 0.813 + ], + "angle": 0, + "content": "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[33] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++ deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5649" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[34] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. arXiv preprint arXiv:2302.02318, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[35] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[36] Shi Qiu, Saeed Anwar, and Nick Barnes. Dense-resolution network for point cloud classification and segmentation. In WACV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.472, + 0.288 + ], + "angle": 0, + "content": "[37] Shi Qiu, Saeed Anwar, and Nick Barnes. Geometric backprojection network for point cloud classification. IEEE TMM, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.472, + 0.319 + ], + "angle": 0, + "content": "[38] Haoxi Ran, Jun Liu, and Chengjie Wang. Surface representation for point clouds. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.472, + 0.36 + ], + "angle": 0, + "content": "[39] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Oct-net: Learning deep 3d representations at high resolutions. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.472, + 0.404 + ], + "angle": 0, + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.472, + 0.446 + ], + "angle": 0, + "content": "[41] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. NeurIPS, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.472, + 0.49 + ], + "angle": 0, + "content": "[42] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In ICCV, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.472, + 0.532 + ], + "angle": 0, + "content": "[43] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.472, + 0.588 + ], + "angle": 0, + "content": "[44] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.472, + 0.644 + ], + "angle": 0, + "content": "[45] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.472, + 0.688 + ], + "angle": 0, + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.472, + 0.731 + ], + "angle": 0, + "content": "[47] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[48] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. TOG, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.472, + 0.816 + ], + "angle": 0, + "content": "[49] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. In NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[50] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Balanced chamfer distance as a comprehensive metric for point cloud completion. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[51] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[52] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[53] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[54] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.29 + ], + "angle": 0, + "content": "[55] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. arXiv preprint arXiv:2106.04180, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.894, + 0.347 + ], + "angle": 0, + "content": "[56] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ToG, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.894, + 0.389 + ], + "angle": 0, + "content": "[57] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.894, + 0.432 + ], + "angle": 0, + "content": "[58] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.894, + 0.461 + ], + "angle": 0, + "content": "[59] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.894, + 0.518 + ], + "angle": 0, + "content": "[60] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. arXiv preprint arXiv:2205.14401, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.894, + 0.573 + ], + "angle": 0, + "content": "[61] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In CVPR, 2023." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5650" + } + ] +] \ No newline at end of file diff --git a/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/e286768b-a5b2-4ea3-a9f1-4d06c6cc68f6_origin.pdf b/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/e286768b-a5b2-4ea3-a9f1-4d06c6cc68f6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fcf774521d7817ea003ab865f9a27c4c0462e2ac --- /dev/null +++ b/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/e286768b-a5b2-4ea3-a9f1-4d06c6cc68f6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70076fce79b76b7dfd7d526a196d6f66d7da07784a7647f5536ab25d74f4b4e3 +size 598759 diff --git a/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/full.md b/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f37aef555aa2d1bc87d744e19e0e605f0a777410 --- /dev/null +++ b/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/full.md @@ -0,0 +1,334 @@ +# Take-A-Photo: 3D-to-2D Generative Pre-training of Point Cloud Models + +Ziyi Wang $^{1,2*}$ Xumin Yu $^{1,2*}$ Yongming Rao $^{1,2}$ Jie Zhou $^{1,2}$ Jiwen Lu $^{1,2\dagger}$ \ + $^{1}$ Department of Automation, Tsinghua University $^{2}$ BNRist\ +{wziyi22, yuxm20}@mails.tsinghua.edu.cn;\ +raoyongming95@gmail.com; {jzhou, lujiwen}@tsinghua.edu.cn + +# Abstract + +With the overwhelming trend of mask image modeling led by MAE, generative pre-training has shown a remarkable potential to boost the performance of fundamental models in 2D vision. However, in 3D vision, the over-reliance on Transformer-based backbones and the unordered nature of point clouds have restricted the further development of generative pre-training. In this paper, we propose a novel 3D-to-2D generative pre-training method that is adaptable to any point cloud model. We propose to generate view images from different instructed poses via the cross-attention mechanism as the pre-training scheme. Generating view images has more precise supervision than its point cloud counterpart, thus assisting 3D backbones to have a finer comprehension of the geometrical structure and stereoscopic relations of the point cloud. Experimental results have proved the superiority of our proposed 3D-to-2D generative pre-training over previous pre-training methods. Our method is also effective in boosting the performance of architecture-oriented approaches, achieving state-of-the-art performance when fine-tuning on ScanObjectNN classification and ShapeNet-Part segmentation tasks. Code is available at https://github.com/wangzy22/TakeAPhoto. + +# 1. Introduction + +Nowadays, pre-training fundamental models with self-supervised mechanisms has witnessed a thriving development in the computer vision community, given its low requirement in data annotation and its high transferability to downstream applications. Self-supervised pre-training aims to fully exploit the statistical and structural knowledge from abundant annotation-free data and empowers the fundamental models with robust representation ability. In 2D vision, self-supervised pre-training has shown strong potential and achieved remarkable progress on diverse backbones in various downstream tasks. Successful pre-training strategies + +![](images/9587d5af785ce68735a5c199e73878cb3927a7b823a8c136bd6799aff9b872a9.jpg) +Figure 1: Principle illustration of 3D-to-2D generative pre-training. The photograph module explicitly encodes pose condition into 3D features from the backbone, and the 2D generator decodes pose-conditioned features into different view images. + +in 2D domain such as contrastive learning [19, 7] and mask modeling [18, 3] have also been adopted to 3D point cloud analysis [58, 30, 23] in recent research. + +However, the pre-training paradigm hasn't become the prevailing trend in 3D learning and architectural design is still the mainstream to reach a new state-of-the-art performance, which is considerably different from the dominant status of pre-training in 2D domain. In object-level analysis, generative pre-training inspired by MAE [18] has been studied thoroughly, but their performances still lag behind architecture-based methods like PointNeXt [35]. Two factors mainly contribute to the inferior status of generative pretraining in 3D learning. Since point clouds are unordered sets of point coordinates, there is no direct and precise supervision like one-to-one MSE loss between generated point clouds and their corresponding ground truth. Chamfer Distance supervision for point clouds only calculates a rough set-to-set matching and its imprecision has been widely discussed in [24, 50, 20]. Additionally, existing advanced generative pre-training methods in object analysis are limited to the Transformer-based backbone, and fail to be extended to other sophisticated point cloud models. + +To alleviate the aforementioned problems, we propose a 3D-to-2D generative pre-training method for point cloud analysis that has higher preciseness in supervision and broader adaptation to different backbones. Instead of re + +constructing point clouds as previous literature [30, 60], we propose to generate view images of the input point cloud given the instructed camera poses. This is similar to taking photos of a realistic object from different perspectives to fully present its structure or internal relations. Therefore, we name our model Take-A-Photo, in short TAP. More specifically, we propose a pose-dependent photograph module that utilizes the cross-attention mechanism to explicitly encode pose conditions with 3D features from the backbone. Then a 2D generator decodes pose-conditioned features into view images that are supervised by rendered ground truth images. The principle illustration of TAP is shown in Figure 1. In the pose-dependent photograph module, we do not provide detailed projection relations from 3D points to 2D pixels, thus the cross-attention layers are encouraged to learn by themselves what those view images look like conditioned on given poses. Since the projection layout, part occlusion relation, faces colors that represent light reflections are largely distinct among view images, the proposed 3D-to-2D generative pre-training is a challenging pretext task that obliges 3D backbone to gain higher representation ability of geometrical knowledge and stereoscopic relations. + +We conduct extensive experiments on various backbones and downstream tasks to verify the effectiveness and superiority of our proposed 3D-to-2D generative pre-training method. When pre-trained on synthetic ShapeNet [6] and transferred to real-world ScanObjectNN [45] classification, TAP brings consistent improvement to different backbone models and successfully outperforms previous point cloud generative pre-training methods based on Transformers backbone. With PointMLP [27] as the backbone, TAP achieves state-of-the-art performance on ScanObjectNN classification and ShapeNetPart [56] part segmentation among methods that do not include any pre-trained image or text model. We also conduct thorough ablation studies to discuss the architectural design of the TAP model and verify the individual contribution of each component. + +In conclusion, the contributions of our paper can be summarized as follows: (1) We propose TAP, the first 3D-to-2D generative pre-training method that is adaptable to any point cloud model. TAP pre-training helps to exploit the potential of point cloud models on geometric structure comprehension and stereoscopic relation understanding. (2) We propose a Photograph Module where we derive mathematical formulations to encode pose conditions as query tokens in cross-attention layers. (3) TAP surpasses previous generative pre-training methods on the Transformers backbone and achieves state-of-the-art performance on ScanObjectNN classification and ShapeNetPart segmentation. + +# 2. Related Work + +Point Cloud Analysis. Point cloud analysis is a fundamental and important task in the realm of 3D vision. Current + +literature has developed two principal methodologies to extract structural representations from 3D point clouds, namely point-based and voxel-based methods. Point-based methods [32, 33, 48, 44, 35, 27] process unordered points directly, introducing various techniques for local information aggregation. Existing point-based methods can be categorized into three types: SetAbstraction-based [32, 33, 35], DynamicGraph-based [48], and Attention-based [57, 58, 30, 23, 60], all of which focus on modeling the relationships between points. Owing to their exceptional ability to effectively preserve fine-grained geometric information, point-based methods are frequently employed for object-level tasks. On the other hand, voxel-based methods [28, 21, 39] partition the 3D space into ordered voxels and employ 3D convolutions for feature extraction. Voxel-based methods are primarily based on SparseConvolution [10, 16], which enables efficient convolution operations in 3D space through sparse convolutions. In exchange for faster processing speeds, voxel-based methods sacrifice their capacity to capture detailed local structures, making them more suitable for large-scale scene-level tasks rather than object-level tasks. + +Point Cloud Pre-training. Pre-training has always been a focal point of research in the field of deep learning. Generally speaking, we usually distinguish pre-training methods based on the amount of annotation required, namely full-supervised pre-training [14, 59, 5], weakly-supervised pretraining [43, 4, 53], and unsupervised pre-training [19, 7, 8]. Among these methods, unsupervised pre-training has become the most popular approach, mainly due to its excellent transferability across tasks and its notable advantage of not relying on labeled data. Numerous researchers have proposed a variety of pretext tasks for unsupervised pre-training. Based on the type of pretext task employed, there are two prevailing pretext tasks for unsupervised pre-training. The first is contrastive learning, as exemplified by MoCo [19] and SimCLR [7]. The other method entails utilizing generative tasks to restore the data from partially or disrupted inputs, such as MAE [18] and BEiT [3]. + +Inspired by pre-training strategies in the image domain, there are more and more unsupervised pre-training methods being proposed for point cloud pre-training. PointContrast [54] embraces the principle of contrastive learning, whereas PointBERT [58] and PointMAE [30] integrate reconstruction pretext tasks. However, existing generative-based pre-training methods for point clouds only consider a single modality. In this paper, we propose a cross-modal generative-based pre-training strategy to achieve more effective pre-training. + +Cross-Modal Learning. Recently, cross-modal learning has been a popular research topic, aiming at learning from multiple modalities such as images, audio and point clouds. It has the potential to enhance the performance of various tasks, + +![](images/b18fe96d01a81d672dbef021c98642a46a696cbb528a7d7d794300d83fdd850a.jpg) +Figure 2: The pipeline of TAP pre-training method. We design a query generator to encode pose conditions and implement cross-attention layers to transform 3D point cloud features $F_{3\mathrm{d}}$ to 2D view image features $F_{2\mathrm{d}}$ according to pose instruction. The predicted pose-dependent view image from 2D generator is supervised by ground truth view image via MSE loss. + +including visual recognition, speech recognition, and point cloud analysis. A variety of methods have been proposed for cross-modal learning, including multi-task learning [11], conditional generation [40], pre-training [34, 13] and tuning [49, 55]. + +In the realm of point cloud analysis, much previous work has explored this learning paradigm. Some literature leverages 2D data for 3D point cloud analysis, such as MVTN [17], MVCNN [42] and CrossPoint [2], proving that the multi-view images and the correspondence between images and points can be helpful for the 3D object understanding. Another line of the research, such as Image2Point [55] and P2P [49] take effort to adapt the models from 2D vision into 3D point cloud analysis, fully exploiting the relationship of 2D and 3D understanding. In this paper, we continue this learning paradigm in 3D vision domain, and for the first time propose the cross-modal generative pre-training scheme for point cloud models. + +# 3. 3D-to-2D Generative Pre-training + +# 3.1. Preliminary: Generative Pre-training + +Generative pre-training is a fundamental branch of pretraining methods that aims at reconstructing integral and complete data given partial or disrupted input. Mathematically, suppose $x$ is a sample from raw data with no annotation. The pre-processing step $T(\cdot)$ either erases part of $x$ randomly or splits $x$ into pieces and intermingles them to get $\tilde{x} = T(x)$ . The generative pre-training model $M$ is designed to restore from those broken input $\hat{x} = M(T(x))$ and the training loss function is designed to measure the reconstruction distance $\mathcal{L} = D(\hat{x},x)$ . In point cloud object analysis, earlier generative pre-training methods propose various pretext tasks as $T$ , including deformation [1], jigsaw puzzles [41] and depth projection [47] to produce disarrayed or partial point clouds. Recently, inspired by MAE [18] in the image domain, generative pre-training in 3D domain mainly focuses on implementing random masking as $T$ and utilizing + +Transformers model as $M$ for reconstruction [58, 30, 23, 60]. The reconstruction distance $D$ is usually measured by the classical $l_{2}$ Chamfer Distance: + +$$ +D (\hat {x}, x) = \frac {1}{| \hat {x} |} \sum_ {a \in \hat {x}} \min _ {b \in x} \| a - b \| _ {2} ^ {2} + \frac {1}{| x |} \sum_ {b \in x} \min _ {a \in \hat {x}} \| a - b \| _ {2} ^ {2} (1) +$$ + +Besides Chamfer Distance between point clouds, some methods also exploit feature distance between latents [58] or occupancy value distance [23] as the loss function. + +The exact reason why generative pre-training would help enhance the representation ability of backbone models still remains an open question. However, abundant experimental results have conveyed that predicting missing parts according to known parts demands high reasoning ability and global comprehension capacity of the model. What's more, generative pre-training is more efficient and suitable for point cloud object analysis than contrastive pre-training, given that contrastive pre-training typically requires a large amount of training data to avoid trivial overfitting solutions but there has always been a data-starvation problem in point cloud object research field. + +# 3.2. Overall Pipeline + +Different from the aforementioned generative pre-training methods that focus on uni-modal point cloud reconstruction, we propose a novel cross-modal pre-training approach of generating view images from instructed camera poses. + +The overall architecture of our proposed TAP pre-training model is depicted in Figure 2. Our model takes as an input point cloud $P \in \mathbb{R}^{N \times 3}$ , where $N$ is the number of points in the input point cloud. The basic building block of TAP mainly consists of: 1) a 3D Backbone that extracts 3D geometric features $F_{3\mathrm{d}} \in \mathbb{R}^{n \times C_{3\mathrm{d}}}$ , where $n$ is the number of downsampled center points and $C_{3\mathrm{d}}$ is the geometric feature dimension; 2) a pose-dependent Photograph Module that takes as inputs $F_{3\mathrm{d}}$ and pose matrix $R \in \mathbb{R}^{3 \times 3}$ , and predicts view image features $F_{2\mathrm{d}}^{R} \in \mathbb{R}^{h \times w \times C_{2\mathrm{d}}}$ conditioned on $R$ , where $h, w$ are height and width of predicted view image + +feature map; 3) an $2D$ Generator that decodes $F_{2\mathrm{d}}^{R}$ into an RGB image $I_{\mathrm{gen}}^{R} \in \mathbb{R}^{H \times W \times 3}$ , where $H, W$ are height and width of the output view image. + +As we place no restriction on $F_{3\mathrm{d}}$ , the 3D Backbone can be arbitrarily chosen and adopted. Therefore, our TAP is more flexible and compatible than existing generative pre-training methods that are limited to Transformer-based architecture. Experimental results in Section 4 will later verify that TAP brings consistent improvement to all kinds of point cloud models. The technical designs of the pose-dependent Photograph Module will be thoroughly discussed in Section 3.3. The 2D Generator consists of four Transpose Convolution layers to progressively upsample image resolution and decode RGB colors of each pixel. + +# 3.3. Photograph Module + +Architectural Design. As illustrated in Figure 2, we leverage cross-attention mechanism from Transformers [46] to build our pose-dependent Photograph Module. + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) V \tag {2} +$$ + +where $d_{k}$ is the scaling factor, and $Q,K,V$ are queries, keys and values matrix. More specifically, we design a Query Generator $\Phi$ to encode camera pose conditions into query tokens: $Q = \Phi (R)\in \mathbb{R}^{hw\times C_{2\mathrm{d}}}$ . We also design a Memory Builder $\Theta$ to construct $K$ and $V$ from 3D geometric features: $K = V = \Theta (F_{3\mathrm{d}})\in \mathbb{R}^{m\times C_{2\mathrm{d}}}$ , where $m$ is the number of memory tokens. The output sequence of the cross attention layers will be rearranged from $hw\times C_{2\mathrm{d}}$ to $h\times w\times C_{2\mathrm{d}}$ , forming the predicted view image features $F_{2\mathrm{d}}^R$ . + +During the cross-attention calculation process, we do not explicitly provide any projection clues of which 3D points would project to which 2D pixel. Instead, the Photograph Module learns by itself how to arrange unordered 3D feature points to ordered 2D pixel grids, purely based on semantic similarities between 3D geometric features and our delicately-designed queries that reveal pose information. Since one sample will only have one set of memory tokens in 3D space but its view images from different poses are quite distinct from each other, learning to predict precise view images from instructed poses in a data-driven manner is not a trivial task. Therefore, during the end-to-end optimization process, the 3D backbone is trained to have a stronger perception of the object's overall geometric structure and gain a higher representative ability of the stereoscopic relations. In this way, our proposed 3D-to-2D generative pre-training would help exploit the potential and enhance the strength of 3D backbone models. + +Query Generator. The query generator $\Phi$ is designed to encode pose condition $R$ into 2D grid of shape $h \times w$ . In object analysis, common practice is leveraging parallel light + +shading to project 3D objects onto 2D grids, and pose matrix $R$ here is used to rotate objects into desired angles before projection. Therefore, each 2D grid actually represents an optical line that starts from infinity, passes through 3D objects and ends at the 2D plane. As a consequence, we choose the direction and the origin points that the optical line goes through as the delegate of the query grid. + +Before deriving formulations of optical lines for each grid, let us first revisit the parallel light shading process for better comprehension. Given 3D coordinates $\mathbf{x} = (x,y,z)$ of a point cloud $P$ and pose matrix $R$ , rotation is first performed to align the object to the ideal pose position: + +$$ +\mathbf {x} ^ {\prime} = \left(x ^ {\prime}, y ^ {\prime}, z ^ {\prime}\right) = R \mathbf {x} \tag {3} +$$ + +Then we just omit the final dimension $z^{\prime}$ and evenly split the first two dimensions $(x^{\prime},y^{\prime})$ into 2D grids $(u,v)$ : + +$$ +u = \frac {x ^ {\prime} - x _ {0}}{g _ {h}} + o _ {h}, \quad v = \frac {y ^ {\prime} - y _ {0}}{g _ {w}} + o _ {w} \tag {4} +$$ + +where $(x_0, y_0)$ is the minimum value of $(x', y')$ , $(g_h, g_w)$ is the grid size, $(o_h, o_w)$ is the offset value to place the projected object at the center of the image. $0 \leq u \leq h - 1$ , $0 \leq v \leq w - 1$ and $(u, v)$ is a sampled pixel coordinate from the 2D grid. + +Now let us begin to derive formulations of the optical line that passes through the query grid. We only know $(u, v)$ for each grid and we want to reversely trace which 3D points $(x, y, z)$ are on the same optical line during parallel light projection. According to Eq. 4: + +$$ +\begin{array}{l} x ^ {\prime} = g _ {h} u + x _ {0} - o _ {h} = \Psi_ {h} (u) \\ y ^ {\prime} = g _ {w} v + y _ {0} - o _ {w} = \Psi_ {w} (v) \tag {5} \\ \end{array} +$$ + +If we denote $A = R^{-1}$ and $A_{ij}$ as the element at $i^{th}$ row and $j^{th}$ column, then according to Eq. 3: + +$$ +x = A _ {1 1} \Psi_ {h} (u) + A _ {1 2} \Psi_ {w} (v) + A _ {1 3} z ^ {\prime} = \Omega_ {x} (u, v) + A _ {1 3} z ^ {\prime} +$$ + +$$ +y = A _ {2 1} \Psi_ {h} (u) + A _ {2 2} \Psi_ {w} (v) + A _ {2 3} z ^ {\prime} = \Omega_ {y} (u, v) + A _ {2 3} z ^ {\prime} +$$ + +$$ +z = A _ {3 1} \Psi_ {h} (u) + A _ {3 2} \Psi_ {w} (v) + A _ {3 3} z ^ {\prime} = \Omega_ {z} (u, v) + A _ {3 3} z ^ {\prime} \tag {6} +$$ + +According to the definition of line's parametric equation, Eq. 6 represents a line passing through the origin point $O: (\Omega_x(u,v), \Omega_y(u,v), \Omega_z(u,v))$ with optical line direction $\mathbf{d} = (A_{13}, A_{23}, A_{33})$ , where $\Omega_x, \Omega_y, \Omega_z$ are xyz coordinates of $O$ and their formulations are conditioned on $u, v$ . Therefore, we concatenate the coordinate of origin point $O$ , normalized direction $\mathbf{d}^{\dagger} = \mathbf{d} / \| \mathbf{d}\|_2$ and normalized position $(u/h, v/w)$ as positional embedding together to be the initial state of our query. A multi-layer-perceptron (MLP) module is later leveraged to map the 8-dim initial query to higher dimensional space. + +Memory Builder. The memory builder takes $F_{3\mathrm{d}}$ as input to prepare for initial state of $K, V$ in cross-attention layers. We + +first concatenate aligned 3D coordinate $P_{\mathrm{3d}}$ with 3D features to enhance the geometric knowledge of $F_{\mathrm{3d}}$ : + +$$ +\hat {F} _ {3 \mathrm {d}} = \operatorname {M L P} \left(\operatorname {c a t} \left(F _ {3 \mathrm {d}}, P _ {3 \mathrm {d}}\right)\right) \tag {7} +$$ + +Additionally, we initialize a learnable memory token $T_{\mathrm{pad}}$ as the pad token and concatenate it with $\hat{F}_{3\mathrm{d}}$ to obtain the initial state of $K, V$ . The reason for concatenating a learnable pad token $T_{\mathrm{pad}}$ is that there are white background areas on the projected image (as shown in Figure 2). As $F_{3\mathrm{d}}$ only encodes foreground objects, we further need a learnable pad token to represent background regions. Otherwise, the cross-attention layers will be confused to learn how to combine foreground tokens into background features and this will inevitably diminish the pre-training effectiveness. + +# 3.4. Objective Function + +We perform per-pixel supervision with Mean Squared Error (MSE) loss between generated view image $I_{\mathrm{gen}}^{R}$ and ground truth image $I_{\mathrm{gt}}^{R}$ , aligned by camera pose $R$ . For simplicity, we will omit $R$ in later formulations. As the background of the rendered ground truth images is all white and reveals little information, we further design a compound loss to balance the weight between foreground regions and background regions: + +$$ +\mathcal {L} \left(I _ {\text {g e n}}, I _ {\text {g t}}\right) = w ^ {\mathrm {f g}} \mathcal {D} ^ {\mathrm {f g}} + w ^ {\mathrm {b g}} \mathcal {D} ^ {\mathrm {b g}} \tag {8} +$$ + +$$ +\mathcal {D} ^ {k} \left(I _ {\text {g e n}} ^ {k}, I _ {\mathrm {g t}} ^ {k}\right) = \frac {1}{H W} \sum_ {h, w} \left(I _ {\text {g e n}} ^ {k} (h, w) - I _ {\mathrm {g t}} ^ {k} (h, w)\right) ^ {2} \tag {9} +$$ + +where $k = \mathrm{fg}$ (foreground), $\mathrm{bg}$ (background) and $w^{\mathrm{fg}}$ , $w^{\mathrm{bg}}$ are loss weights for foreground and background, respectively. Such per-pixel supervision is more precise than the ambiguous set-to-set Chamfer Distance introduced in Eq. 1. + +# 4. Experiments + +In this section, we first introduce the setups of our pretraining scheme. Then we evaluate our pre-training method on various point cloud backbones by fine-tuning them on different downstream tasks, such as point cloud classification on ModelNet and ScanObjectNN datasets, and part segmentation on the ShapeNetPart dataset. Finally, we provide in-depth ablation studies for the architectural design of our proposed TAP pre-training pipeline. + +# 4.1. Pre-training Setups + +Data Setups To align with previous research practices [58, 30, 54], we choose ShapeNet [6] that contains more than 50 thousand CAD models as our pre-training datasets. We sampled 1024 points from each 3D CAD model to form the point clouds, consistent with previous work. Since ShapeNet does not provide images for each point cloud, we use the + +Table 1: Classification results on the ScanObjectNN dataset. We report the overall accuracy $(\%)$ . The results with $\dagger$ are reproduced by PointNeXt [35] repository. + +
MethodOBJ_BGOBJ_ONLYPB_T50_RS
Hierarchical Models with TAP Pre-training
†DGCNN [48]--86.1
+ TAP--86.6 (+0.5)
†PointNet++ [33]--86.2
+ TAP--86.8 (+0.6)
†PointMLP [27]--87.4
+ TAP--88.5 (+1.1)
Standard Transformers with Generative Pre-training
w/o pre-training [46]79.8680.5577.24
OcCo [47]84.8585.5478.79
Point-BERT [58]87.4388.1283.07
MaskPoint [23]89.3088.1084.30
Point-MAE [30]90.0288.2985.18
TAP (Ours)90.3689.5085.67
+ +rendered image from 12 surrounding viewpoints generated by MVCNN [42]. During our pre-training, the models are exclusively pre-trained with the training split following the practice of previous work [58]. + +Architecture Setups We conduct experiments on various point cloud encoders, including PointNet++ [33], DGCNN [48], PointMLP [27] and Transformers [58] for point cloud object classification. During the pre-training stage, the photograph module takes encoded point cloud features and pose conditions as inputs to generate a $32 \times$ downsampled view image feature map of size $7 \times 7$ from a specific viewpoint. Then the 2D generator progressively upsamples the image feature map to decode RGB view images of size $224 \times 224$ . We do not alter the architecture of the point cloud backbone since the photograph module and the 2D generator are exclusively used during the pretraining phase and are dropped during the fine-tuning stage. In our experiment, the photograph module is a six-layer cross-attention block, with attention layer channels limited to 256 to enhance efficiency. During the pre-training task on ShapeNet, we utilized four simple transpose convolutions to upsample the reconstructed 2D feature map and predict the RGB value for each pixel. + +Implementation Details The experiments of TAP pretraining and finetuning on various downstream tasks are implemented with PyTorch [31]. We utilize AdamW [26] optimizer and the CosineAnnealing learning rate scheduler [25] to pre-train the point cloud backbone for 100 epochs. We set the initial learning rate as $5e^{-4}$ and weight decay as $5e^{-2}$ . In our experiment, we train various point cloud backbones with 32 batch sizes on a single Nvidia 3090Ti GPU. The + +Table 2: Results comparisons with previous methods on the ScanObjectNN and ModelNet40 datasets. The model parameters number (#Params) and overall accuracy (\%) are reported. † denotes our reproduced results of PointMLP on the ModelNet40 dataset. Methods with * introduce extra knowledge from pre-trained image models or pre-trained vision-language models. We do not compete with them for fair comparison and only list them for reference. + +
Method#Params (M)ScanObjectNNModelNet40
Supervised Learning Only
PointNet [32]3.568.089.2
PointNet++ [33]1.577.990.7
Transformer [46]22.177.2491.4
DGCNN [48]1.878.192.9
PointCNN [22]0.678.592.2
DRNet [36]-80.393.1
SimpleView [15]-80.5±0.393.9
GBNet [37]8.881.093.8
PRA-Net [9]2.381.093.7
MVTN [17]11.282.893.8
RepSurf-U [38]1.584.394.4
PointMLP [27]12.685.4±0.393.5†
PointNeXt [35]1.487.7±0.493.7
Transformers with Pre-training
OcCo [47]22.178.892.1
Point-BERT [58]22.183.193.2
MaskPoint [23]22.184.393.8
Point-MAE [30]22.185.293.8
Point-M2AE [60]15.386.494.0
With Pre-trained Image Model
ACT [13]*22.188.293.7
I2P-MAE [61]*-90.193.7
ReCon [34]*44.390.694.1
Our Proposed TAP Pre-training
PointMLP + TAP12.688.594.0
+ +drop path rate of the cross-attention layer is set to 0.1. The foreground and background loss weights $w^{\mathrm{fg}}$ , $w^{\mathrm{bg}}$ are set to 20 and 1. The detailed architecture of our simple 2D generator is: TConv(256, 128, 5, 4) → TConv(128, 64, 3, 2) → TConv(64, 32, 3, 2) → TConv(32, 3, 3, 2), where TConv stands for Transpose Convolution and the four numbers in the tuple denotes $(C_{in}, C_{out}, \text{Kernel}, \text{Stride})$ respectively. During the fine-tuning stage, we perform a learning rate warming up for point cloud backbones with 10 epochs, and keep other settings unchanged for a fair comparison. + +# 4.2. Downstream Tasks + +In this section, we report the experimental results of various downstream tasks. We follow the previous work to conduct experiments of object classification on real-world ScanObjectNN and synthetic ModelNet40 datasets. We also verify the effectiveness of our pre-training method on the + +![](images/ab0c1890cc27997e1a36eee33a8013973e244ab527bce5bc78a7f80cab1c1d48.jpg) +Figure 3: The visualization results of our proposed 3D-to-2D generative pre-training. The first row displays view images generated by our TAP pre-training pipeline and the second row shows ground truth images. Our TAP can produce view images with appropriate shapes and reflection colors, demonstrating its ability in capturing geometric structure and stereoscopic knowledge. + +part segmentation task with the ShapeNetPart dataset. + +# 4.2.1 Object Classification + +Main Results. To evaluate the effectiveness of our proposed TAP, we implement it with various point cloud architectures, including classical baselines such as PointNet++, DGCNN, and PointMLP, as well as widely used Standard Transformers backbone for existing generative pre-training methods. We follow the common practice to experiment our model on three variants of the ScanObjectNN dataset: 1) OBJ-ONLY: cropping object without any background; 2) OBJ-BG: containing the background and object; 3) PB-T50-RS: adopting various augmentations to the objects. We reported the results comparing with existing pre-training methods in Table 1. + +As shown in the upper part of the table, we pretrain the graph-based architecture DGCNN, set-abstraction-based architecture PointNet++, and MLP-based architecture PointMLP with TAP, and observe consistent improvements across models. The results strongly convey that our proposed TAP can be successfully applied to various types of point cloud models and the proposed novel 3D-to-2D generative pre-training is effective regardless of the backbone architecture. Considering that nearly all existing generative pre-training methods are specially designed for Transformer-based architecture, our TAP is much superior in its wider adaptation and higher flexibility. Additionally, we also provide a detailed and fair comparison with previous work by implementing TAP with the Standard Transformers architecture in the lower part of the table, where no hierarchical designs or inductive bias is included. Our TAP outperforms previous pre-training methods in all three split settings, providing strong evidence that our 3D-to-2D generative pretraining strategy can also benefit attention-based architectures and surpass uni-modal generative pre-training competitors. It is worth noting that although TAP and many previous + +Table 3: Part segmentation results on the ShapeNetPart dataset. We report the mean IoU across all part categories $\mathrm{mIoU}_C$ , the mean IoU across all instances $\mathrm{mIoU}_I$ , and the IoU for each category. + +
MethodsmIoUCmIoUIaerobagcapcarchairearphoneguitarknifelamplaptopmotormugpistolrocketskateboardtable
Supervised Representation Learning Only
PointNet [32]80.483.783.478.782.574.989.673.091.585.980.895.365.293.081.257.972.880.6
PointNet++ [33]81.985.182.479.087.777.390.871.891.085.983.795.371.694.181.358.776.482.6
DGCNN [48]82.385.284.083.486.777.890.674.791.287.582.895.766.394.981.163.574.582.6
PointMLP [27]84.686.183.583.487.580.590.378.292.288.182.696.277.595.885.464.683.384.3
KPConv [44]85.186.484.686.387.281.191.177.892.688.482.796.278.195.885.469.082.083.6
Transformers with Uni-modal Generative Pre-training
Point-BERT [58]84.185.684.384.888.079.891.081.791.687.985.295.675.694.784.363.476.381.5
Point-MAE [30]84.286.184.385.088.380.591.378.592.187.486.196.175.294.684.763.577.182.4
MaskPoint [23]84.486.084.285.688.180.391.279.591.987.886.295.376.995.085.364.476.981.8
Point-M2AE [60]84.986.5----------------
Our Proposed 3D-to-2D Generative Pre-training
PointMLP+TAP85.286.984.886.189.582.592.175.992.388.785.696.579.896.085.966.278.183.2
+ +pre-training approaches have significantly improved the performance of Transformers in point cloud tasks, increasing the accuracy from 77.24 to 85.67, they still lag far behind advanced point cloud networks such as PointMLP. Therefore, TAP's applicability to various models is an important characteristic that would benefit future research. + +Comparisons with Previous Methods. To clearly demonstrate the high performance of our proposed TAP pre-training in object classification tasks, we compare TAP with existing methods on both synthetic ModelNet and real-world ScanObjectNN (the hardest PB-T50-RS variant) datasets in Table 2. We categorize existing methods into two types: (1) architecture-oriented methods, which focus on developing novel model architectures for 3D point clouds and do not involve any pre-training techniques, and (2) pre-training methods, which pay more attention to the pre-training strategy and whose backbone model are mostly limited to Transformer-base architecture. It's worth noticing that methods marked with an asterisk (*) incorporate additional knowledge from pre-trained image models or pre-trained vision-language models. To ensure a fair and unbiased comparison, we refrain from directly comparing our method with these approaches. However, we include them in the listing for reference purposes, acknowledging their existence and potential relevance in related research. + +From the experimental results, we can see that accompanied by PointMLP backbone model, our proposed TAP pre-training achieves the best classification accuracy on ScanObjectNN and ModelNet40 among existing models (with no pre-trained knowledge from image or language like P2P [49]), demonstrating the effectiveness of our approach and validating the superiority of our 3D-to-2D cross-modal generative pre-training method over previous generative pretraining methods. Moreover, we also note that our proposed method has brought higher performance improvements on the ScanObjectNN dataset than on ModelNet40. This may + +be attributed to the reason that the cross-modal generative pre-training has enhanced the network's ability to understand point clouds from different views, which is beneficial for a more robust understanding of the real-scan data with more noise and disturbance in the ScanObjectNN dataset. + +Visualization Results. Figure 3 shows the visualization results of TAP. The first row shows the generated view images while the second row displays the ground truth images for reference. The TAP method can successfully predict the accurate shape of the object and the RGB colors that represent light reflections in rendered images. Therefore, TAP is capable of capturing the geometric structure of 3D objects and reasoning occlusion relations from specific camera poses. + +# 4.2.2 Part Segmentation + +Performing dense prediction is always a more challenging task compared with classification. In this section, we evaluate the local distinguishability of our proposed TAP pre-training method, fine-tuning the pre-trained point cloud model on the ShapeNetPart dataset for the part segmentation task. Quantitative results are shown in Table 3. We implement PointMLP as the backbone model and compare our TAP results with two mainstreams of previous literature. The upper row displays classical architecture-oriented methods that focus on network design and are trained from scratch. The lower row shows members of the generative pre-training family that rely on Transformer-based architectures. + +According to results comparisons, our TAP pre-training significantly improves the part segmentation performance of the PointMLP backbone, increasing class mIoU by 0.6 and instance mIoU by 0.8. More importantly, our TAP pretraining achieves state-of-the-art performance on both class mIoU and instance mIoU, surpassing leading works in both tracks. Specifically, TAP exceeds the performance of PointM2AE on instance mIoU by 0.4. This satisfactory perfor + +mance serves as strong evidence to convey that our proposed TAP pre-training is superior to previous uni-model generative pre-training mechanisms in dense prediction tasks. This may be attributed to the factor that our supervision in 2D with MSE loss is more precise than the ambiguous Chamfer Distance in 3D reconstruction. Therefore, models with TAP pre-training obtain more accurate comprehension of local geometry and detail awareness, which contributes to mIoU gain in dense prediction tasks. What's more, TAP outperforms KPCov on instance mIoU by 0.5, demonstrating that the proposed 3D-to-2D generative pre-training method can fully exploit the potential of the point cloud model and help it have a better perception of objects' geometric structure. As TAP is adaptable to any architecture, future improvements in architectural design can also benefit from TAP pre-training. + +# 4.2.3 Few-shot Classification + +Following Point-BERT [58], we conduct few-shot classification with Standard Transformers on ModelNet40 [52] dataset. As shown in Table 4, we report mean overall accuracy and standard deviation $(\mathrm{mOA} \pm \mathrm{std})$ on 10 pre-defined data folders for each few-shot setting. The way and shot in Table 4 specify the number of categories and the number of training examples per category, respectively. + +From the results, TAP achieves the highest mean overall accuracy across all few-shot settings when compared to previous generative pre-training approaches. Furthermore, TAP exhibits significantly lower standard deviations than those reported in the existing literature for the majority of few-shot settings, which signifies its robust performance and consistent superiority. This indicates that TAP is not only capable of achieving high mean overall accuracy but also exhibits reliability and robustness across various few-shot settings. Such stability is crucial in real-world applications, where consistency and predictability are vital for practical deployment. + +# 4.3. Scene-level Dense Predictions + +To assess the effectiveness of TAP in handling scene-level dense prediction tasks, we carry out experiments on more complicated scene-level object detection and semantic segmentation on the ScanNetV2 [12] dataset. For the object detection task, we adopt 3DETR [29] and pre-train its encoder on the object-level dataset ShapeNet [6] with TAP. Average precision at 0.25 and 0.5 IoU thresholds are reported. Regarding semantic segmentation, we employ the PointTransformerV2 [51](PTv2) model and pre-train it on the ScanNetV2 dataset with TAP. We report mean IoU for evaluation metric. It is worth mentioning that PTv2 represents the current state-of-the-art approach with open-source code availability. + +Based on the results presented in Table 5, TAP consis + +Table 4: Few-shot Classification with Standard Transformers on ModelNet40 dataset. We report mean overall accuracy and standard deviation on 10 pre-defined data folders for each setting. Best results are marked bold. + +
Method5-way10-way
10-shot20-shot10-shot20-shot
w/o pre-training87.8 ± 5.293.3 ± 4.384.6 ± 5.589.4 ± 6.3
Point-BERT [58]94.6 ± 3.196.3 ± 2.791.0 ± 5.492.7 ± 5.1
MaskPoint [23]95.0 ± 3.797.2 ± 1.791.4 ± 4.093.4 ± 3.5
Point-MAE [30]96.3 ± 2.597.8 ± 1.892.6 ± 4.195.0 ± 3.0
TAP97.3 ± 1.897.8 ± 1.793.1 ± 2.695.8 ± 1.0
+ +Table 5: Scene-level object detection and semantic segmentation on ScanNetV2 [12]. Average precision at 0.25 IoU thresholds $(\mathrm{AP}_{0.25})$ and 0.5 IoU thresholds $(\mathrm{AP}_{0.5})$ of detection and mean Intersection-over-Union (mIoU) of semantic segmentation are reported. + +
MethodDet (3DETR [29])Seg (PTv2 [51])
AP0.25AP0.5mIoU
Baseline62.137.972.4
+TAP63.0 (+0.9)41.4 (+3.5)72.6 (+0.2)
+ +tently enhances the performance of all baselines, thereby showcasing its efficacy in tackling more intricate scene-level dense prediction tasks. Remarkably, even with the encoder solely pre-trained on an object-level dataset for scene-level detection task, significant improvements are observed in both $\mathrm{AP}_{0.25}$ and $\mathrm{AP}_{0.5}$ metrics. This suggests that the learned representations from TAP effectively capture relevant information and generalize well to complex scenes, even when the pre-training data is limited to object-level collections. Such generalization capabilities are valuable in scenarios where obtaining large-scale fully annotated scene-level datasets may be challenging or expensive. + +# 4.4. Ablation Studies + +To investigate the architectural design of our proposed Photograph Module in TAP pre-training pipeline, we conduct extensive ablation studies on the ScanObjectNN dataset with PointMLP as the backbone model. + +Photograph Module Architectural Designs. In Photograph Module, we implement cross-attention layers to generate view image feature maps conditioned on pose instruction. We believe that letting the module learn by itself how to rearrange 3D point features in 2D grids will enhance the representation ability of the 3D backbone. Therefore, we conduct ablation studies to verify this hypothesis. As shown in Table 6a, we implement Model $\mathrm{A}_1$ with no attention lay + +Table 6: Ablation studies on Photograph module in TAP pre-training pipeline. We choose PointMLP as the backbone model and conduct ablation studies on the ScanObjectNN dataset from two aspects: overall architectural designs and query designs. In Table(a), we first investigate the effectiveness of cross-attention design compared with direct projection (Model A1) and direct project with self-attention (Model A2). Then we analyze the influence of the different number of attention layers and feature channels in Model B and Model C. Finally, we discuss whether pad token in memory builder is beneficial in Model D. In Table(b), we conduct further experiments to compare different approaches for query designing: (1) Using learnable query based on the given viewpoints (Model E) or use mathematical formulations derived in Eq. 6. (2) The information we need when we mathematically encode pose information into init query status: (i) Origin: the coordinate of origin point $O$ that the optical line passes through. (ii) Direction: The normalized direction of the optical line. (iii) PE: The position embedding for each grid. + +(a) Overall Architectural Designs. + +
ModelAttention TypeLayerNum.ChannelsMem.PadAcc.(%)
A1None-256X87.6 (-0.9)
A2SelfAttn6 layers256X87.8 (-0.7)
BCrossAttn2 layers256✓87.9 (-0.6)
CCrossAttn6 layers512✓87.8 (-0.7)
DCrossAttn6 layers256X88.3 (-0.2)
TAPCrossAttn6 layers256✓88.5
+ +(b) Query Designs. + +
ModelQuery TypeOriginDirectionPEAcc.(%)
ELearnableXXX87.5 (-1.0)
F1FormulaX✓✓86.5 (-2.0)
F2Formula✓XX87.8 (-0.7)
F3Formula✓✓X88.0 (-0.5)
F4Formula✓X✓88.1 (-0.4)
TAPFormula✓✓✓88.5
+ +ers, directly projecting 3D feature points to 2D grids based on Eq. 3 and Eq. 4. This results in a much simpler pretraining task, as the projection relation has been directly told. Additionally, in Model $\mathrm{A}_2$ , we add self-attention layers after explicit projection to help the model capture longer-range correlations. Pose knowledge is encoded as a pose token that is concatenated to projected grids, similar to the CLS token in classification Transformers. According to quantitative results comparison with TAP that implements cross-attention layers, fine-tuning results of pre-training methods in $\mathrm{A}_1$ and $\mathrm{A}_2$ version show inferiority. Therefore, the cross-attention architecture we designed to entirely LEARN the projection relation is the most suitable choice for the proposed 3D-to-2D generative pre-training. + +What's more, we discuss the number of cross-attention layers, the dimension of feature channels and whether to concatenate pad token in memory builder in Model $B, C, D$ . According to the results, more cross-attention layers show stronger representation ability, while too large channel number will lead to performance decrease caused by over-fitting. The performance gain from Model $D$ to TAP also verifies that the pad token design in the memory builder is essential. + +Query Generator Designs. In the query generator, we derive the mathematical formulation of the optical lines passing through 2D grids. We propose to concatenate the coordinate of origin point $O$ , normalized direction $\mathbf{d}^{\dagger}$ and position embedding $(u / h, v / w)$ as the initial state of queries. In Table 6b, we first compare this mathematical design with totally learnable queries that takes pose matrix $R$ as input and implements MLP layers to predict query for each grid. As shown in Model $E$ , learnable queries cannot satisfactorily encode pose information, while our derived formulation for query construction is both clearer in physical meaning and + +more competitive in fine-tuning accuracy. + +In ablation $\mathrm{F_1}$ to $\mathrm{F_4}$ , we progressively discuss the three components of query generation. Quantitative comparison with TAP verifies that every component is indispensable for query generation, where coordinates of origin points are of the most importance. + +# 5. Conclusions + +In this paper, we have proposed a novel 3D-to-2D generative pre-training method TAP that is adaptable to any point cloud model. We implemented the cross-attention mechanism to generate view images of point clouds from instructed camera poses. To better encode pose conditions and generate physically meaningful queries, we derived mathematical formulations of optical lines. The proposed TAP pre-training had higher preciseness in supervision and broader adaptation to different backbones, compared with directly reconstructing point clouds in previous methods. Experimental results conveyed that the TAP pre-training can help the backbone models better capture the structural knowledge and stereoscopic relations. Fine-tuning results of TAP pre-training achieve state-of-the-art performance on ScanObjectNN classification and ShapeNetPart segmentation, among methods that do not include any pre-trained image or text models. We believe the cross-modal generative pre-training paradigm will be a promising direction for future research. + +# Acknowledgement + +This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0114903 and in part by the National Natural Science Foundation of China under Grant 62125603. + +# References + +[1] Idan Achituve, Haggai Maron, and Gal Chechik. Self-supervised learning for domain adaptation on point clouds. In WACV, 2021. +[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In CVPR, pages 9902-9912, June 2022. +[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. +[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. *NeruIPS*, 2019. +[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. +[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. +[7] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020. +[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. +[9] Silin Cheng, Xiwu Chen, Xinwei He, Zhe Liu, and Xiang Bai. Pra-net: Point relation-aware network for 3d point cloud analysis. IEEE TIP, 30:4436-4448, 2021. +[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. +[11] Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning. In ICML, 2008. +[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. +[13] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022. +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. +[15] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In ICML, 2021. + +[16] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. +[17] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. MVTN: multi-view transformation network for 3d shape recognition. In ICCV, 2021. +[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll'ar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. +[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. +[20] Tianxin Huang, Zhonggan Ding, Jiangning Zhang, Ying Tai, Zhenyu Zhang, Mingang Chen, Chengjie Wang, and Yong Liu. Learning to measure the point cloud reconstruction loss in a representation space. In CVPR, 2023. +[21] Roman Klokov and Victor Lempitsky. Escape from cells: Deep kd-networks for the recognition of 3d point cloud models. In ICCV, 2017. +[22] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. NeurIPS, 2018. +[23] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV. Springer, 2022. +[24] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and Shi-Min Hu. Morphing and sampling network for dense point cloud completion. In AAAI, 2020. +[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. +[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018. +[27] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In ICLR, 2022. +[28] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In IROS. IEEE, 2015. +[29] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, 2021. +[30] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV. Springer, 2022. +[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. +[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. +[33] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++ deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. + +[34] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. arXiv preprint arXiv:2302.02318, 2023. +[35] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. +[36] Shi Qiu, Saeed Anwar, and Nick Barnes. Dense-resolution network for point cloud classification and segmentation. In WACV, 2021. +[37] Shi Qiu, Saeed Anwar, and Nick Barnes. Geometric backprojection network for point cloud classification. IEEE TMM, 2022. +[38] Haoxi Ran, Jun Liu, and Chengjie Wang. Surface representation for point clouds. In CVPR, 2022. +[39] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Oct-net: Learning deep 3d representations at high resolutions. In CVPR, 2017. +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. +[41] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. NeurIPS, 2019. +[42] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In ICCV, 2015. +[43] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 2017. +[44] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. ICCV, 2019. +[45] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. +[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. +[47] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. +[48] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. TOG, 2019. +[49] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. In NeurIPS, 2022. +[50] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Balanced chamfer distance as a comprehensive metric for point cloud completion. In NeurIPS, 2021. +[51] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. + +[52] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. +[53] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. +[54] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. +[55] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. arXiv preprint arXiv:2106.04180, 2021. +[56] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ToG, 2016. +[57] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In ICCV, 2021. +[58] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. +[59] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, 2022. +[60] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. arXiv preprint arXiv:2205.14401, 2022. +[61] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In CVPR, 2023. \ No newline at end of file diff --git a/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/images.zip b/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c55c2dad5cec4c4bbf353347600f34250fe3dd17 --- /dev/null +++ b/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8e7cdd25819426061a8b9f9ce734a928019aac6a6d41ac226798e7c73eb20fe +size 475001 diff --git a/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/layout.json b/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d1cf65b97ee112b0ffa51bc66e22f5e5f2ee2efe --- /dev/null +++ b/2023/Take-A-Photo_ 3D-to-2D Generative Pre-training of Point Cloud Models/layout.json @@ -0,0 +1,9448 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 103, + 520, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 103, + 520, + 120 + ], + "spans": [ + { + "bbox": [ + 74, + 103, + 520, + 120 + ], + "type": "text", + "content": "Take-A-Photo: 3D-to-2D Generative Pre-training of Point Cloud Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "spans": [ + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "text", + "content": "Ziyi Wang" + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "text", + "content": " Xumin Yu" + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "text", + "content": " Yongming Rao" + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "text", + "content": " Jie Zhou" + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "text", + "content": " Jiwen Lu" + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "text", + "content": "\\\n" + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "text", + "content": "Department of Automation, Tsinghua University " + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 141, + 485, + 200 + ], + "type": "text", + "content": "BNRist\\\n{wziyi22, yuxm20}@mails.tsinghua.edu.cn;\\\nraoyongming95@gmail.com; {jzhou, lujiwen}@tsinghua.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 252, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 515 + ], + "type": "text", + "content": "With the overwhelming trend of mask image modeling led by MAE, generative pre-training has shown a remarkable potential to boost the performance of fundamental models in 2D vision. However, in 3D vision, the over-reliance on Transformer-based backbones and the unordered nature of point clouds have restricted the further development of generative pre-training. In this paper, we propose a novel 3D-to-2D generative pre-training method that is adaptable to any point cloud model. We propose to generate view images from different instructed poses via the cross-attention mechanism as the pre-training scheme. Generating view images has more precise supervision than its point cloud counterpart, thus assisting 3D backbones to have a finer comprehension of the geometrical structure and stereoscopic relations of the point cloud. Experimental results have proved the superiority of our proposed 3D-to-2D generative pre-training over previous pre-training methods. Our method is also effective in boosting the performance of architecture-oriented approaches, achieving state-of-the-art performance when fine-tuning on ScanObjectNN classification and ShapeNet-Part segmentation tasks. Code is available at https://github.com/wangzy22/TakeAPhoto." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 540, + 127, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 127, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 127, + 553 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 561, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 561, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 561, + 288, + 693 + ], + "type": "text", + "content": "Nowadays, pre-training fundamental models with self-supervised mechanisms has witnessed a thriving development in the computer vision community, given its low requirement in data annotation and its high transferability to downstream applications. Self-supervised pre-training aims to fully exploit the statistical and structural knowledge from abundant annotation-free data and empowers the fundamental models with robust representation ability. In 2D vision, self-supervised pre-training has shown strong potential and achieved remarkable progress on diverse backbones in various downstream tasks. Successful pre-training strategies" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 312, + 225, + 547, + 299 + ], + "blocks": [ + { + "bbox": [ + 312, + 225, + 547, + 299 + ], + "lines": [ + { + "bbox": [ + 312, + 225, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 312, + 225, + 547, + 299 + ], + "type": "image", + "image_path": "9587d5af785ce68735a5c199e73878cb3927a7b823a8c136bd6799aff9b872a9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 307, + 547, + 367 + ], + "lines": [ + { + "bbox": [ + 305, + 307, + 547, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 547, + 367 + ], + "type": "text", + "content": "Figure 1: Principle illustration of 3D-to-2D generative pre-training. The photograph module explicitly encodes pose condition into 3D features from the backbone, and the 2D generator decodes pose-conditioned features into different view images." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 397, + 546, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 397, + 546, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 397, + 546, + 433 + ], + "type": "text", + "content": "in 2D domain such as contrastive learning [19, 7] and mask modeling [18, 3] have also been adopted to 3D point cloud analysis [58, 30, 23] in recent research." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 436, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 547, + 662 + ], + "type": "text", + "content": "However, the pre-training paradigm hasn't become the prevailing trend in 3D learning and architectural design is still the mainstream to reach a new state-of-the-art performance, which is considerably different from the dominant status of pre-training in 2D domain. In object-level analysis, generative pre-training inspired by MAE [18] has been studied thoroughly, but their performances still lag behind architecture-based methods like PointNeXt [35]. Two factors mainly contribute to the inferior status of generative pretraining in 3D learning. Since point clouds are unordered sets of point coordinates, there is no direct and precise supervision like one-to-one MSE loss between generated point clouds and their corresponding ground truth. Chamfer Distance supervision for point clouds only calculates a rough set-to-set matching and its imprecision has been widely discussed in [24, 50, 20]. Additionally, existing advanced generative pre-training methods in object analysis are limited to the Transformer-based backbone, and fail to be extended to other sophisticated point cloud models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 713 + ], + "type": "text", + "content": "To alleviate the aforementioned problems, we propose a 3D-to-2D generative pre-training method for point cloud analysis that has higher preciseness in supervision and broader adaptation to different backbones. Instead of re" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 207, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 207, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 207, + 713 + ], + "type": "text", + "content": "*Equal contribution. †Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5640" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 287, + 335 + ], + "type": "text", + "content": "constructing point clouds as previous literature [30, 60], we propose to generate view images of the input point cloud given the instructed camera poses. This is similar to taking photos of a realistic object from different perspectives to fully present its structure or internal relations. Therefore, we name our model Take-A-Photo, in short TAP. More specifically, we propose a pose-dependent photograph module that utilizes the cross-attention mechanism to explicitly encode pose conditions with 3D features from the backbone. Then a 2D generator decodes pose-conditioned features into view images that are supervised by rendered ground truth images. The principle illustration of TAP is shown in Figure 1. In the pose-dependent photograph module, we do not provide detailed projection relations from 3D points to 2D pixels, thus the cross-attention layers are encouraged to learn by themselves what those view images look like conditioned on given poses. Since the projection layout, part occlusion relation, faces colors that represent light reflections are largely distinct among view images, the proposed 3D-to-2D generative pre-training is a challenging pretext task that obliges 3D backbone to gain higher representation ability of geometrical knowledge and stereoscopic relations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 336, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 336, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 49, + 336, + 287, + 514 + ], + "type": "text", + "content": "We conduct extensive experiments on various backbones and downstream tasks to verify the effectiveness and superiority of our proposed 3D-to-2D generative pre-training method. When pre-trained on synthetic ShapeNet [6] and transferred to real-world ScanObjectNN [45] classification, TAP brings consistent improvement to different backbone models and successfully outperforms previous point cloud generative pre-training methods based on Transformers backbone. With PointMLP [27] as the backbone, TAP achieves state-of-the-art performance on ScanObjectNN classification and ShapeNetPart [56] part segmentation among methods that do not include any pre-trained image or text model. We also conduct thorough ablation studies to discuss the architectural design of the TAP model and verify the individual contribution of each component." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 515, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 515, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 49, + 515, + 287, + 658 + ], + "type": "text", + "content": "In conclusion, the contributions of our paper can be summarized as follows: (1) We propose TAP, the first 3D-to-2D generative pre-training method that is adaptable to any point cloud model. TAP pre-training helps to exploit the potential of point cloud models on geometric structure comprehension and stereoscopic relation understanding. (2) We propose a Photograph Module where we derive mathematical formulations to encode pose conditions as query tokens in cross-attention layers. (3) TAP surpasses previous generative pre-training methods on the Transformers backbone and achieves state-of-the-art performance on ScanObjectNN classification and ShapeNetPart segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 670, + 132, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 132, + 681 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 132, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 689, + 287, + 712 + ], + "type": "text", + "content": "Point Cloud Analysis. Point cloud analysis is a fundamental and important task in the realm of 3D vision. Current" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 307, + 72, + 546, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 323 + ], + "type": "text", + "content": "literature has developed two principal methodologies to extract structural representations from 3D point clouds, namely point-based and voxel-based methods. Point-based methods [32, 33, 48, 44, 35, 27] process unordered points directly, introducing various techniques for local information aggregation. Existing point-based methods can be categorized into three types: SetAbstraction-based [32, 33, 35], DynamicGraph-based [48], and Attention-based [57, 58, 30, 23, 60], all of which focus on modeling the relationships between points. Owing to their exceptional ability to effectively preserve fine-grained geometric information, point-based methods are frequently employed for object-level tasks. On the other hand, voxel-based methods [28, 21, 39] partition the 3D space into ordered voxels and employ 3D convolutions for feature extraction. Voxel-based methods are primarily based on SparseConvolution [10, 16], which enables efficient convolution operations in 3D space through sparse convolutions. In exchange for faster processing speeds, voxel-based methods sacrifice their capacity to capture detailed local structures, making them more suitable for large-scale scene-level tasks rather than object-level tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 331, + 546, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 546, + 534 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 546, + 534 + ], + "type": "text", + "content": "Point Cloud Pre-training. Pre-training has always been a focal point of research in the field of deep learning. Generally speaking, we usually distinguish pre-training methods based on the amount of annotation required, namely full-supervised pre-training [14, 59, 5], weakly-supervised pretraining [43, 4, 53], and unsupervised pre-training [19, 7, 8]. Among these methods, unsupervised pre-training has become the most popular approach, mainly due to its excellent transferability across tasks and its notable advantage of not relying on labeled data. Numerous researchers have proposed a variety of pretext tasks for unsupervised pre-training. Based on the type of pretext task employed, there are two prevailing pretext tasks for unsupervised pre-training. The first is contrastive learning, as exemplified by MoCo [19] and SimCLR [7]. The other method entails utilizing generative tasks to restore the data from partially or disrupted inputs, such as MAE [18] and BEiT [3]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 537, + 546, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 546, + 657 + ], + "type": "text", + "content": "Inspired by pre-training strategies in the image domain, there are more and more unsupervised pre-training methods being proposed for point cloud pre-training. PointContrast [54] embraces the principle of contrastive learning, whereas PointBERT [58] and PointMAE [30] integrate reconstruction pretext tasks. However, existing generative-based pre-training methods for point clouds only consider a single modality. In this paper, we propose a cross-modal generative-based pre-training strategy to achieve more effective pre-training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 666, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 666, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 666, + 546, + 712 + ], + "type": "text", + "content": "Cross-Modal Learning. Recently, cross-modal learning has been a popular research topic, aiming at learning from multiple modalities such as images, audio and point clouds. It has the potential to enhance the performance of various tasks," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 313, + 757 + ], + "type": "text", + "content": "5641" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 72, + 545, + 175 + ], + "blocks": [ + { + "bbox": [ + 53, + 72, + 545, + 175 + ], + "lines": [ + { + "bbox": [ + 53, + 72, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 545, + 175 + ], + "type": "image", + "image_path": "b18fe96d01a81d672dbef021c98642a46a696cbb528a7d7d794300d83fdd850a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 187, + 548, + 225 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 548, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 548, + 225 + ], + "type": "text", + "content": "Figure 2: The pipeline of TAP pre-training method. We design a query generator to encode pose conditions and implement cross-attention layers to transform 3D point cloud features " + }, + { + "bbox": [ + 46, + 187, + 548, + 225 + ], + "type": "inline_equation", + "content": "F_{3\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 187, + 548, + 225 + ], + "type": "text", + "content": " to 2D view image features " + }, + { + "bbox": [ + 46, + 187, + 548, + 225 + ], + "type": "inline_equation", + "content": "F_{2\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 187, + 548, + 225 + ], + "type": "text", + "content": " according to pose instruction. The predicted pose-dependent view image from 2D generator is supervised by ground truth view image via MSE loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 255, + 288, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 288, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 288, + 314 + ], + "type": "text", + "content": "including visual recognition, speech recognition, and point cloud analysis. A variety of methods have been proposed for cross-modal learning, including multi-task learning [11], conditional generation [40], pre-training [34, 13] and tuning [49, 55]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 315, + 290, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 290, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 290, + 471 + ], + "type": "text", + "content": "In the realm of point cloud analysis, much previous work has explored this learning paradigm. Some literature leverages 2D data for 3D point cloud analysis, such as MVTN [17], MVCNN [42] and CrossPoint [2], proving that the multi-view images and the correspondence between images and points can be helpful for the 3D object understanding. Another line of the research, such as Image2Point [55] and P2P [49] take effort to adapt the models from 2D vision into 3D point cloud analysis, fully exploiting the relationship of 2D and 3D understanding. In this paper, we continue this learning paradigm in 3D vision domain, and for the first time propose the cross-modal generative pre-training scheme for point cloud models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 483, + 235, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 235, + 497 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 235, + 497 + ], + "type": "text", + "content": "3. 3D-to-2D Generative Pre-training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 502, + 246, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 246, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 246, + 516 + ], + "type": "text", + "content": "3.1. Preliminary: Generative Pre-training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": "Generative pre-training is a fundamental branch of pretraining methods that aims at reconstructing integral and complete data given partial or disrupted input. Mathematically, suppose " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " is a sample from raw data with no annotation. The pre-processing step " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " either erases part of " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " randomly or splits " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " into pieces and intermingles them to get " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{x} = T(x)" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": ". The generative pre-training model " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " is designed to restore from those broken input " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\hat{x} = M(T(x))" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " and the training loss function is designed to measure the reconstruction distance " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = D(\\hat{x},x)" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": ". In point cloud object analysis, earlier generative pre-training methods propose various pretext tasks as " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": ", including deformation [1], jigsaw puzzles [41] and depth projection [47] to produce disarrayed or partial point clouds. Recently, inspired by MAE [18] in the image domain, generative pre-training in 3D domain mainly focuses on implementing random masking as " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " and utilizing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 255, + 547, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 291 + ], + "type": "text", + "content": "Transformers model as " + }, + { + "bbox": [ + 304, + 255, + 547, + 291 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 255, + 547, + 291 + ], + "type": "text", + "content": " for reconstruction [58, 30, 23, 60]. The reconstruction distance " + }, + { + "bbox": [ + 304, + 255, + 547, + 291 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 255, + 547, + 291 + ], + "type": "text", + "content": " is usually measured by the classical " + }, + { + "bbox": [ + 304, + 255, + 547, + 291 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 255, + 547, + 291 + ], + "type": "text", + "content": " Chamfer Distance:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 295, + 547, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 295, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 310, + 295, + 547, + 324 + ], + "type": "interline_equation", + "content": "D (\\hat {x}, x) = \\frac {1}{| \\hat {x} |} \\sum_ {a \\in \\hat {x}} \\min _ {b \\in x} \\| a - b \\| _ {2} ^ {2} + \\frac {1}{| x |} \\sum_ {b \\in x} \\min _ {a \\in \\hat {x}} \\| a - b \\| _ {2} ^ {2} (1)", + "image_path": "c5cf0d3844ff2189d527279faea6bbe43632e1f0cd1a9d6508c297bf9965efb7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 330, + 547, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 547, + 365 + ], + "type": "text", + "content": "Besides Chamfer Distance between point clouds, some methods also exploit feature distance between latents [58] or occupancy value distance [23] as the loss function." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "text", + "content": "The exact reason why generative pre-training would help enhance the representation ability of backbone models still remains an open question. However, abundant experimental results have conveyed that predicting missing parts according to known parts demands high reasoning ability and global comprehension capacity of the model. What's more, generative pre-training is more efficient and suitable for point cloud object analysis than contrastive pre-training, given that contrastive pre-training typically requires a large amount of training data to avoid trivial overfitting solutions but there has always been a data-starvation problem in point cloud object research field." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 515, + 405, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 405, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 405, + 529 + ], + "type": "text", + "content": "3.2. Overall Pipeline" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 582 + ], + "type": "text", + "content": "Different from the aforementioned generative pre-training methods that focus on uni-modal point cloud reconstruction, we propose a novel cross-modal pre-training approach of generating view images from instructed camera poses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "The overall architecture of our proposed TAP pre-training model is depicted in Figure 2. Our model takes as an input point cloud " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": " is the number of points in the input point cloud. The basic building block of TAP mainly consists of: 1) a 3D Backbone that extracts 3D geometric features " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "F_{3\\mathrm{d}} \\in \\mathbb{R}^{n \\times C_{3\\mathrm{d}}}" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": " is the number of downsampled center points and " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "C_{3\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": " is the geometric feature dimension; 2) a pose-dependent Photograph Module that takes as inputs " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "F_{3\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": " and pose matrix " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": ", and predicts view image features " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "F_{2\\mathrm{d}}^{R} \\in \\mathbb{R}^{h \\times w \\times C_{2\\mathrm{d}}}" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "inline_equation", + "content": "h, w" + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": " are height and width of predicted view image" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5642" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "feature map; 3) an " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " Generator that decodes " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "F_{2\\mathrm{d}}^{R}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " into an RGB image " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{gen}}^{R} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " are height and width of the output view image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 288, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 288, + 240 + ], + "type": "text", + "content": "As we place no restriction on " + }, + { + "bbox": [ + 46, + 109, + 288, + 240 + ], + "type": "inline_equation", + "content": "F_{3\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 109, + 288, + 240 + ], + "type": "text", + "content": ", the 3D Backbone can be arbitrarily chosen and adopted. Therefore, our TAP is more flexible and compatible than existing generative pre-training methods that are limited to Transformer-based architecture. Experimental results in Section 4 will later verify that TAP brings consistent improvement to all kinds of point cloud models. The technical designs of the pose-dependent Photograph Module will be thoroughly discussed in Section 3.3. The 2D Generator consists of four Transpose Convolution layers to progressively upsample image resolution and decode RGB colors of each pixel." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 248, + 164, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 164, + 261 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 164, + 261 + ], + "type": "text", + "content": "3.3. Photograph Module" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 267, + 288, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 288, + 304 + ], + "type": "text", + "content": "Architectural Design. As illustrated in Figure 2, we leverage cross-attention mechanism from Transformers [46] to build our pose-dependent Photograph Module." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 312, + 287, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 312, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 77, + 312, + 287, + 340 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {2}", + "image_path": "39b7ea0ad7cbd0770c53fd5e6d3b08d7e544aef35d5efb1b9e146a83b28e59e5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "content": " is the scaling factor, and " + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "inline_equation", + "content": "Q,K,V" + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "content": " are queries, keys and values matrix. More specifically, we design a Query Generator " + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "content": " to encode camera pose conditions into query tokens: " + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "inline_equation", + "content": "Q = \\Phi (R)\\in \\mathbb{R}^{hw\\times C_{2\\mathrm{d}}}" + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "content": ". We also design a Memory Builder " + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "content": " to construct " + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "content": " from 3D geometric features: " + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "inline_equation", + "content": "K = V = \\Theta (F_{3\\mathrm{d}})\\in \\mathbb{R}^{m\\times C_{2\\mathrm{d}}}" + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "content": " is the number of memory tokens. The output sequence of the cross attention layers will be rearranged from " + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "inline_equation", + "content": "hw\\times C_{2\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "inline_equation", + "content": "h\\times w\\times C_{2\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "content": ", forming the predicted view image features " + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "inline_equation", + "content": "F_{2\\mathrm{d}}^R" + }, + { + "bbox": [ + 46, + 347, + 287, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 456, + 288, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 288, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 288, + 671 + ], + "type": "text", + "content": "During the cross-attention calculation process, we do not explicitly provide any projection clues of which 3D points would project to which 2D pixel. Instead, the Photograph Module learns by itself how to arrange unordered 3D feature points to ordered 2D pixel grids, purely based on semantic similarities between 3D geometric features and our delicately-designed queries that reveal pose information. Since one sample will only have one set of memory tokens in 3D space but its view images from different poses are quite distinct from each other, learning to predict precise view images from instructed poses in a data-driven manner is not a trivial task. Therefore, during the end-to-end optimization process, the 3D backbone is trained to have a stronger perception of the object's overall geometric structure and gain a higher representative ability of the stereoscopic relations. In this way, our proposed 3D-to-2D generative pre-training would help exploit the potential and enhance the strength of 3D backbone models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Query Generator. The query generator " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " is designed to encode pose condition " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " into 2D grid of shape " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": ". In object analysis, common practice is leveraging parallel light" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "shading to project 3D objects onto 2D grids, and pose matrix " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " here is used to rotate objects into desired angles before projection. Therefore, each 2D grid actually represents an optical line that starts from infinity, passes through 3D objects and ends at the 2D plane. As a consequence, we choose the direction and the origin points that the optical line goes through as the delegate of the query grid." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 156, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 547, + 216 + ], + "type": "text", + "content": "Before deriving formulations of optical lines for each grid, let us first revisit the parallel light shading process for better comprehension. Given 3D coordinates " + }, + { + "bbox": [ + 304, + 156, + 547, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (x,y,z)" + }, + { + "bbox": [ + 304, + 156, + 547, + 216 + ], + "type": "text", + "content": " of a point cloud " + }, + { + "bbox": [ + 304, + 156, + 547, + 216 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 156, + 547, + 216 + ], + "type": "text", + "content": " and pose matrix " + }, + { + "bbox": [ + 304, + 156, + 547, + 216 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 156, + 547, + 216 + ], + "type": "text", + "content": ", rotation is first performed to align the object to the ideal pose position:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 378, + 223, + 545, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 223, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 378, + 223, + 545, + 236 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\prime} = \\left(x ^ {\\prime}, y ^ {\\prime}, z ^ {\\prime}\\right) = R \\mathbf {x} \\tag {3}", + "image_path": "8223e68fe31c44575ec901b6d0ea48d1bba8f0e4dfee3ca84588c72d763527bb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 243, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 267 + ], + "type": "text", + "content": "Then we just omit the final dimension " + }, + { + "bbox": [ + 304, + 243, + 545, + 267 + ], + "type": "inline_equation", + "content": "z^{\\prime}" + }, + { + "bbox": [ + 304, + 243, + 545, + 267 + ], + "type": "text", + "content": " and evenly split the first two dimensions " + }, + { + "bbox": [ + 304, + 243, + 545, + 267 + ], + "type": "inline_equation", + "content": "(x^{\\prime},y^{\\prime})" + }, + { + "bbox": [ + 304, + 243, + 545, + 267 + ], + "type": "text", + "content": " into 2D grids " + }, + { + "bbox": [ + 304, + 243, + 545, + 267 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 304, + 243, + 545, + 267 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 342, + 274, + 545, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 274, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 342, + 274, + 545, + 300 + ], + "type": "interline_equation", + "content": "u = \\frac {x ^ {\\prime} - x _ {0}}{g _ {h}} + o _ {h}, \\quad v = \\frac {y ^ {\\prime} - y _ {0}}{g _ {w}} + o _ {w} \\tag {4}", + "image_path": "13ea234ef8313b6e5ca2d9853da37e575ec984a62c58b959b2e355f13519f608.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "inline_equation", + "content": "(x_0, y_0)" + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "text", + "content": " is the minimum value of " + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "inline_equation", + "content": "(x', y')" + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "inline_equation", + "content": "(g_h, g_w)" + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "text", + "content": " is the grid size, " + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "inline_equation", + "content": "(o_h, o_w)" + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "text", + "content": " is the offset value to place the projected object at the center of the image. " + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "inline_equation", + "content": "0 \\leq u \\leq h - 1" + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "inline_equation", + "content": "0 \\leq v \\leq w - 1" + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 304, + 305, + 545, + 365 + ], + "type": "text", + "content": " is a sampled pixel coordinate from the 2D grid." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 365, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 425 + ], + "type": "text", + "content": "Now let us begin to derive formulations of the optical line that passes through the query grid. We only know " + }, + { + "bbox": [ + 304, + 365, + 545, + 425 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 304, + 365, + 545, + 425 + ], + "type": "text", + "content": " for each grid and we want to reversely trace which 3D points " + }, + { + "bbox": [ + 304, + 365, + 545, + 425 + ], + "type": "inline_equation", + "content": "(x, y, z)" + }, + { + "bbox": [ + 304, + 365, + 545, + 425 + ], + "type": "text", + "content": " are on the same optical line during parallel light projection. According to Eq. 4:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 431, + 545, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 431, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 361, + 431, + 545, + 460 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x ^ {\\prime} = g _ {h} u + x _ {0} - o _ {h} = \\Psi_ {h} (u) \\\\ y ^ {\\prime} = g _ {w} v + y _ {0} - o _ {w} = \\Psi_ {w} (v) \\tag {5} \\\\ \\end{array}", + "image_path": "7e269c8836f523de24f45a68e881998ee0e5e88d65060ce4138229022a017171.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 467, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 492 + ], + "type": "text", + "content": "If we denote " + }, + { + "bbox": [ + 304, + 467, + 545, + 492 + ], + "type": "inline_equation", + "content": "A = R^{-1}" + }, + { + "bbox": [ + 304, + 467, + 545, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 467, + 545, + 492 + ], + "type": "inline_equation", + "content": "A_{ij}" + }, + { + "bbox": [ + 304, + 467, + 545, + 492 + ], + "type": "text", + "content": " as the element at " + }, + { + "bbox": [ + 304, + 467, + 545, + 492 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 467, + 545, + 492 + ], + "type": "text", + "content": " row and " + }, + { + "bbox": [ + 304, + 467, + 545, + 492 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 304, + 467, + 545, + 492 + ], + "type": "text", + "content": " column, then according to Eq. 3:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 498, + 545, + 511 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 511 + ], + "type": "interline_equation", + "content": "x = A _ {1 1} \\Psi_ {h} (u) + A _ {1 2} \\Psi_ {w} (v) + A _ {1 3} z ^ {\\prime} = \\Omega_ {x} (u, v) + A _ {1 3} z ^ {\\prime}", + "image_path": "9d7f260d9c8b14cca2c7ebf366e3ec07befdbe6fb31895bd4c1fdaecc8872280.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 514, + 545, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 545, + 526 + ], + "type": "interline_equation", + "content": "y = A _ {2 1} \\Psi_ {h} (u) + A _ {2 2} \\Psi_ {w} (v) + A _ {2 3} z ^ {\\prime} = \\Omega_ {y} (u, v) + A _ {2 3} z ^ {\\prime}", + "image_path": "72211fb4c32ed60631764274fd0199af7b992da027db1d010610909791a29208.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 529, + 545, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 529, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 529, + 545, + 552 + ], + "type": "interline_equation", + "content": "z = A _ {3 1} \\Psi_ {h} (u) + A _ {3 2} \\Psi_ {w} (v) + A _ {3 3} z ^ {\\prime} = \\Omega_ {z} (u, v) + A _ {3 3} z ^ {\\prime} \\tag {6}", + "image_path": "55815e0f729050dcffe69bb08d898b3537b61704d203b3053036b7eff8b4c85a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "text", + "content": "According to the definition of line's parametric equation, Eq. 6 represents a line passing through the origin point " + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "inline_equation", + "content": "O: (\\Omega_x(u,v), \\Omega_y(u,v), \\Omega_z(u,v))" + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "text", + "content": " with optical line direction " + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{d} = (A_{13}, A_{23}, A_{33})" + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "inline_equation", + "content": "\\Omega_x, \\Omega_y, \\Omega_z" + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "text", + "content": " are xyz coordinates of " + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "text", + "content": " and their formulations are conditioned on " + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "inline_equation", + "content": "u, v" + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "text", + "content": ". Therefore, we concatenate the coordinate of origin point " + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "text", + "content": ", normalized direction " + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{d}^{\\dagger} = \\mathbf{d} / \\| \\mathbf{d}\\|_2" + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "text", + "content": " and normalized position " + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "inline_equation", + "content": "(u/h, v/w)" + }, + { + "bbox": [ + 304, + 552, + 547, + 684 + ], + "type": "text", + "content": " as positional embedding together to be the initial state of our query. A multi-layer-perceptron (MLP) module is later leveraged to map the 8-dim initial query to higher dimensional space." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "Memory Builder. The memory builder takes " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{3\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " as input to prepare for initial state of " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "K, V" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " in cross-attention layers. We" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5643" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "first concatenate aligned 3D coordinate " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{3d}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " with 3D features to enhance the geometric knowledge of " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{3d}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 105, + 288, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 105, + 288, + 119 + ], + "spans": [ + { + "bbox": [ + 110, + 105, + 288, + 119 + ], + "type": "interline_equation", + "content": "\\hat {F} _ {3 \\mathrm {d}} = \\operatorname {M L P} \\left(\\operatorname {c a t} \\left(F _ {3 \\mathrm {d}}, P _ {3 \\mathrm {d}}\\right)\\right) \\tag {7}", + "image_path": "ed4457f7c3cdba366f6c89012f91a1fc866273c0d4076a9015d5e50124230b91.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 129, + 288, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 129, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 129, + 288, + 250 + ], + "type": "text", + "content": "Additionally, we initialize a learnable memory token " + }, + { + "bbox": [ + 46, + 129, + 288, + 250 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{pad}}" + }, + { + "bbox": [ + 46, + 129, + 288, + 250 + ], + "type": "text", + "content": " as the pad token and concatenate it with " + }, + { + "bbox": [ + 46, + 129, + 288, + 250 + ], + "type": "inline_equation", + "content": "\\hat{F}_{3\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 129, + 288, + 250 + ], + "type": "text", + "content": " to obtain the initial state of " + }, + { + "bbox": [ + 46, + 129, + 288, + 250 + ], + "type": "inline_equation", + "content": "K, V" + }, + { + "bbox": [ + 46, + 129, + 288, + 250 + ], + "type": "text", + "content": ". The reason for concatenating a learnable pad token " + }, + { + "bbox": [ + 46, + 129, + 288, + 250 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{pad}}" + }, + { + "bbox": [ + 46, + 129, + 288, + 250 + ], + "type": "text", + "content": " is that there are white background areas on the projected image (as shown in Figure 2). As " + }, + { + "bbox": [ + 46, + 129, + 288, + 250 + ], + "type": "inline_equation", + "content": "F_{3\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 129, + 288, + 250 + ], + "type": "text", + "content": " only encodes foreground objects, we further need a learnable pad token to represent background regions. Otherwise, the cross-attention layers will be confused to learn how to combine foreground tokens into background features and this will inevitably diminish the pre-training effectiveness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 257, + 160, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 160, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 160, + 270 + ], + "type": "text", + "content": "3.4. Objective Function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 276, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 288, + 373 + ], + "type": "text", + "content": "We perform per-pixel supervision with Mean Squared Error (MSE) loss between generated view image " + }, + { + "bbox": [ + 46, + 276, + 288, + 373 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{gen}}^{R}" + }, + { + "bbox": [ + 46, + 276, + 288, + 373 + ], + "type": "text", + "content": " and ground truth image " + }, + { + "bbox": [ + 46, + 276, + 288, + 373 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{gt}}^{R}" + }, + { + "bbox": [ + 46, + 276, + 288, + 373 + ], + "type": "text", + "content": ", aligned by camera pose " + }, + { + "bbox": [ + 46, + 276, + 288, + 373 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 276, + 288, + 373 + ], + "type": "text", + "content": ". For simplicity, we will omit " + }, + { + "bbox": [ + 46, + 276, + 288, + 373 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 276, + 288, + 373 + ], + "type": "text", + "content": " in later formulations. As the background of the rendered ground truth images is all white and reveals little information, we further design a compound loss to balance the weight between foreground regions and background regions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 381, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 381, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 101, + 381, + 287, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(I _ {\\text {g e n}}, I _ {\\text {g t}}\\right) = w ^ {\\mathrm {f g}} \\mathcal {D} ^ {\\mathrm {f g}} + w ^ {\\mathrm {b g}} \\mathcal {D} ^ {\\mathrm {b g}} \\tag {8}", + "image_path": "45de472d43234143268868841ceee9d79da14d56f09b0c38804245461e1d3f78.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 405, + 287, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 405, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 59, + 405, + 287, + 434 + ], + "type": "interline_equation", + "content": "\\mathcal {D} ^ {k} \\left(I _ {\\text {g e n}} ^ {k}, I _ {\\mathrm {g t}} ^ {k}\\right) = \\frac {1}{H W} \\sum_ {h, w} \\left(I _ {\\text {g e n}} ^ {k} (h, w) - I _ {\\mathrm {g t}} ^ {k} (h, w)\\right) ^ {2} \\tag {9}", + "image_path": "f2605bf49f3b2d46c47f15a377d7bdd1031e9081537fc0a5b0b0b564dc135f9f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 440, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 440, + 287, + 489 + ], + "type": "inline_equation", + "content": "k = \\mathrm{fg}" + }, + { + "bbox": [ + 46, + 440, + 287, + 489 + ], + "type": "text", + "content": " (foreground), " + }, + { + "bbox": [ + 46, + 440, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\mathrm{bg}" + }, + { + "bbox": [ + 46, + 440, + 287, + 489 + ], + "type": "text", + "content": " (background) and " + }, + { + "bbox": [ + 46, + 440, + 287, + 489 + ], + "type": "inline_equation", + "content": "w^{\\mathrm{fg}}" + }, + { + "bbox": [ + 46, + 440, + 287, + 489 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 440, + 287, + 489 + ], + "type": "inline_equation", + "content": "w^{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 440, + 287, + 489 + ], + "type": "text", + "content": " are loss weights for foreground and background, respectively. Such per-pixel supervision is more precise than the ambiguous set-to-set Chamfer Distance introduced in Eq. 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 499, + 128, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 128, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 128, + 513 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 519, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 288, + 616 + ], + "type": "text", + "content": "In this section, we first introduce the setups of our pretraining scheme. Then we evaluate our pre-training method on various point cloud backbones by fine-tuning them on different downstream tasks, such as point cloud classification on ModelNet and ScanObjectNN datasets, and part segmentation on the ShapeNetPart dataset. Finally, we provide in-depth ablation studies for the architectural design of our proposed TAP pre-training pipeline." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 623, + 162, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 162, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 162, + 636 + ], + "type": "text", + "content": "4.1. Pre-training Setups" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Data Setups To align with previous research practices [58, 30, 54], we choose ShapeNet [6] that contains more than 50 thousand CAD models as our pre-training datasets. We sampled 1024 points from each 3D CAD model to form the point clouds, consistent with previous work. Since ShapeNet does not provide images for each point cloud, we use the" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 309, + 116, + 545, + 304 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "type": "text", + "content": "Table 1: Classification results on the ScanObjectNN dataset. We report the overall accuracy " + }, + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "type": "text", + "content": ". The results with " + }, + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 70, + 545, + 107 + ], + "type": "text", + "content": " are reproduced by PointNeXt [35] repository." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 116, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 309, + 116, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 309, + 116, + 545, + 304 + ], + "type": "table", + "html": "
MethodOBJ_BGOBJ_ONLYPB_T50_RS
Hierarchical Models with TAP Pre-training
†DGCNN [48]--86.1
+ TAP--86.6 (+0.5)
†PointNet++ [33]--86.2
+ TAP--86.8 (+0.6)
†PointMLP [27]--87.4
+ TAP--88.5 (+1.1)
Standard Transformers with Generative Pre-training
w/o pre-training [46]79.8680.5577.24
OcCo [47]84.8585.5478.79
Point-BERT [58]87.4388.1283.07
MaskPoint [23]89.3088.1084.30
Point-MAE [30]90.0288.2985.18
TAP (Ours)90.3689.5085.67
", + "image_path": "776be80e0544f6f2a35b36e2b765985afa3c92353ae7bd111b6b14d83f75274a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 329, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 378 + ], + "type": "text", + "content": "rendered image from 12 surrounding viewpoints generated by MVCNN [42]. During our pre-training, the models are exclusively pre-trained with the training split following the practice of previous work [58]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 383, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 547, + 612 + ], + "type": "text", + "content": "Architecture Setups We conduct experiments on various point cloud encoders, including PointNet++ [33], DGCNN [48], PointMLP [27] and Transformers [58] for point cloud object classification. During the pre-training stage, the photograph module takes encoded point cloud features and pose conditions as inputs to generate a " + }, + { + "bbox": [ + 304, + 383, + 547, + 612 + ], + "type": "inline_equation", + "content": "32 \\times" + }, + { + "bbox": [ + 304, + 383, + 547, + 612 + ], + "type": "text", + "content": " downsampled view image feature map of size " + }, + { + "bbox": [ + 304, + 383, + 547, + 612 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 304, + 383, + 547, + 612 + ], + "type": "text", + "content": " from a specific viewpoint. Then the 2D generator progressively upsamples the image feature map to decode RGB view images of size " + }, + { + "bbox": [ + 304, + 383, + 547, + 612 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 383, + 547, + 612 + ], + "type": "text", + "content": ". We do not alter the architecture of the point cloud backbone since the photograph module and the 2D generator are exclusively used during the pretraining phase and are dropped during the fine-tuning stage. In our experiment, the photograph module is a six-layer cross-attention block, with attention layer channels limited to 256 to enhance efficiency. During the pre-training task on ShapeNet, we utilized four simple transpose convolutions to upsample the reconstructed 2D feature map and predict the RGB value for each pixel." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "type": "text", + "content": "Implementation Details The experiments of TAP pretraining and finetuning on various downstream tasks are implemented with PyTorch [31]. We utilize AdamW [26] optimizer and the CosineAnnealing learning rate scheduler [25] to pre-train the point cloud backbone for 100 epochs. We set the initial learning rate as " + }, + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "type": "inline_equation", + "content": "5e^{-4}" + }, + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "type": "text", + "content": " and weight decay as " + }, + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "type": "inline_equation", + "content": "5e^{-2}" + }, + { + "bbox": [ + 304, + 617, + 548, + 713 + ], + "type": "text", + "content": ". In our experiment, we train various point cloud backbones with 32 batch sizes on a single Nvidia 3090Ti GPU. The" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5644" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 176, + 286, + 485 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 167 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 167 + ], + "type": "text", + "content": "Table 2: Results comparisons with previous methods on the ScanObjectNN and ModelNet40 datasets. The model parameters number (#Params) and overall accuracy (\\%) are reported. † denotes our reproduced results of PointMLP on the ModelNet40 dataset. Methods with * introduce extra knowledge from pre-trained image models or pre-trained vision-language models. We do not compete with them for fair comparison and only list them for reference." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 176, + 286, + 485 + ], + "lines": [ + { + "bbox": [ + 49, + 176, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 49, + 176, + 286, + 485 + ], + "type": "table", + "html": "
Method#Params (M)ScanObjectNNModelNet40
Supervised Learning Only
PointNet [32]3.568.089.2
PointNet++ [33]1.577.990.7
Transformer [46]22.177.2491.4
DGCNN [48]1.878.192.9
PointCNN [22]0.678.592.2
DRNet [36]-80.393.1
SimpleView [15]-80.5±0.393.9
GBNet [37]8.881.093.8
PRA-Net [9]2.381.093.7
MVTN [17]11.282.893.8
RepSurf-U [38]1.584.394.4
PointMLP [27]12.685.4±0.393.5†
PointNeXt [35]1.487.7±0.493.7
Transformers with Pre-training
OcCo [47]22.178.892.1
Point-BERT [58]22.183.193.2
MaskPoint [23]22.184.393.8
Point-MAE [30]22.185.293.8
Point-M2AE [60]15.386.494.0
With Pre-trained Image Model
ACT [13]*22.188.293.7
I2P-MAE [61]*-90.193.7
ReCon [34]*44.390.694.1
Our Proposed TAP Pre-training
PointMLP + TAP12.688.594.0
", + "image_path": "60bd2d3e38885583d7db2103f5becc7e1372f5a0d87b33943477a97082083179.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 508, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 288, + 628 + ], + "type": "text", + "content": "drop path rate of the cross-attention layer is set to 0.1. The foreground and background loss weights " + }, + { + "bbox": [ + 46, + 508, + 288, + 628 + ], + "type": "inline_equation", + "content": "w^{\\mathrm{fg}}" + }, + { + "bbox": [ + 46, + 508, + 288, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 508, + 288, + 628 + ], + "type": "inline_equation", + "content": "w^{\\mathrm{bg}}" + }, + { + "bbox": [ + 46, + 508, + 288, + 628 + ], + "type": "text", + "content": " are set to 20 and 1. The detailed architecture of our simple 2D generator is: TConv(256, 128, 5, 4) → TConv(128, 64, 3, 2) → TConv(64, 32, 3, 2) → TConv(32, 3, 3, 2), where TConv stands for Transpose Convolution and the four numbers in the tuple denotes " + }, + { + "bbox": [ + 46, + 508, + 288, + 628 + ], + "type": "inline_equation", + "content": "(C_{in}, C_{out}, \\text{Kernel}, \\text{Stride})" + }, + { + "bbox": [ + 46, + 508, + 288, + 628 + ], + "type": "text", + "content": " respectively. During the fine-tuning stage, we perform a learning rate warming up for point cloud backbones with 10 epochs, and keep other settings unchanged for a fair comparison." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 635, + 159, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 159, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 159, + 647 + ], + "type": "text", + "content": "4.2. Downstream Tasks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "In this section, we report the experimental results of various downstream tasks. We follow the previous work to conduct experiments of object classification on real-world ScanObjectNN and synthetic ModelNet40 datasets. We also verify the effectiveness of our pre-training method on the" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 75, + 545, + 162 + ], + "blocks": [ + { + "bbox": [ + 307, + 75, + 545, + 162 + ], + "lines": [ + { + "bbox": [ + 307, + 75, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 75, + 545, + 162 + ], + "type": "image", + "image_path": "ab0c1890cc27997e1a36eee33a8013973e244ab527bce5bc78a7f80cab1c1d48.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 175, + 548, + 260 + ], + "lines": [ + { + "bbox": [ + 304, + 175, + 548, + 260 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 548, + 260 + ], + "type": "text", + "content": "Figure 3: The visualization results of our proposed 3D-to-2D generative pre-training. The first row displays view images generated by our TAP pre-training pipeline and the second row shows ground truth images. Our TAP can produce view images with appropriate shapes and reflection colors, demonstrating its ability in capturing geometric structure and stereoscopic knowledge." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 285, + 523, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 523, + 298 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 523, + 298 + ], + "type": "text", + "content": "part segmentation task with the ShapeNetPart dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 310, + 428, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 310, + 428, + 322 + ], + "spans": [ + { + "bbox": [ + 305, + 310, + 428, + 322 + ], + "type": "text", + "content": "4.2.1 Object Classification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 330, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 547, + 461 + ], + "type": "text", + "content": "Main Results. To evaluate the effectiveness of our proposed TAP, we implement it with various point cloud architectures, including classical baselines such as PointNet++, DGCNN, and PointMLP, as well as widely used Standard Transformers backbone for existing generative pre-training methods. We follow the common practice to experiment our model on three variants of the ScanObjectNN dataset: 1) OBJ-ONLY: cropping object without any background; 2) OBJ-BG: containing the background and object; 3) PB-T50-RS: adopting various augmentations to the objects. We reported the results comparing with existing pre-training methods in Table 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 463, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 548, + 715 + ], + "type": "text", + "content": "As shown in the upper part of the table, we pretrain the graph-based architecture DGCNN, set-abstraction-based architecture PointNet++, and MLP-based architecture PointMLP with TAP, and observe consistent improvements across models. The results strongly convey that our proposed TAP can be successfully applied to various types of point cloud models and the proposed novel 3D-to-2D generative pre-training is effective regardless of the backbone architecture. Considering that nearly all existing generative pre-training methods are specially designed for Transformer-based architecture, our TAP is much superior in its wider adaptation and higher flexibility. Additionally, we also provide a detailed and fair comparison with previous work by implementing TAP with the Standard Transformers architecture in the lower part of the table, where no hierarchical designs or inductive bias is included. Our TAP outperforms previous pre-training methods in all three split settings, providing strong evidence that our 3D-to-2D generative pretraining strategy can also benefit attention-based architectures and surpass uni-modal generative pre-training competitors. It is worth noting that although TAP and many previous" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5645" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 99, + 547, + 258 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 3: Part segmentation results on the ShapeNetPart dataset. We report the mean IoU across all part categories " + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}_C" + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": ", the mean IoU across all instances " + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}_I" + }, + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": ", and the IoU for each category." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 99, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 47, + 99, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 99, + 547, + 258 + ], + "type": "table", + "html": "
MethodsmIoUCmIoUIaerobagcapcarchairearphoneguitarknifelamplaptopmotormugpistolrocketskateboardtable
Supervised Representation Learning Only
PointNet [32]80.483.783.478.782.574.989.673.091.585.980.895.365.293.081.257.972.880.6
PointNet++ [33]81.985.182.479.087.777.390.871.891.085.983.795.371.694.181.358.776.482.6
DGCNN [48]82.385.284.083.486.777.890.674.791.287.582.895.766.394.981.163.574.582.6
PointMLP [27]84.686.183.583.487.580.590.378.292.288.182.696.277.595.885.464.683.384.3
KPConv [44]85.186.484.686.387.281.191.177.892.688.482.796.278.195.885.469.082.083.6
Transformers with Uni-modal Generative Pre-training
Point-BERT [58]84.185.684.384.888.079.891.081.791.687.985.295.675.694.784.363.476.381.5
Point-MAE [30]84.286.184.385.088.380.591.378.592.187.486.196.175.294.684.763.577.182.4
MaskPoint [23]84.486.084.285.688.180.391.279.591.987.886.295.376.995.085.364.476.981.8
Point-M2AE [60]84.986.5----------------
Our Proposed 3D-to-2D Generative Pre-training
PointMLP+TAP85.286.984.886.189.582.592.175.992.388.785.696.579.896.085.966.278.183.2
", + "image_path": "0b27ca5a37e7fde9b3a05957d4de0b866f0716395adf606bbfa1e35a75c1df4d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 277, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 348 + ], + "type": "text", + "content": "pre-training approaches have significantly improved the performance of Transformers in point cloud tasks, increasing the accuracy from 77.24 to 85.67, they still lag far behind advanced point cloud networks such as PointMLP. Therefore, TAP's applicability to various models is an important characteristic that would benefit future research." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 289, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 581 + ], + "type": "text", + "content": "Comparisons with Previous Methods. To clearly demonstrate the high performance of our proposed TAP pre-training in object classification tasks, we compare TAP with existing methods on both synthetic ModelNet and real-world ScanObjectNN (the hardest PB-T50-RS variant) datasets in Table 2. We categorize existing methods into two types: (1) architecture-oriented methods, which focus on developing novel model architectures for 3D point clouds and do not involve any pre-training techniques, and (2) pre-training methods, which pay more attention to the pre-training strategy and whose backbone model are mostly limited to Transformer-base architecture. It's worth noticing that methods marked with an asterisk (*) incorporate additional knowledge from pre-trained image models or pre-trained vision-language models. To ensure a fair and unbiased comparison, we refrain from directly comparing our method with these approaches. However, we include them in the listing for reference purposes, acknowledging their existence and potential relevance in related research." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "From the experimental results, we can see that accompanied by PointMLP backbone model, our proposed TAP pre-training achieves the best classification accuracy on ScanObjectNN and ModelNet40 among existing models (with no pre-trained knowledge from image or language like P2P [49]), demonstrating the effectiveness of our approach and validating the superiority of our 3D-to-2D cross-modal generative pre-training method over previous generative pretraining methods. Moreover, we also note that our proposed method has brought higher performance improvements on the ScanObjectNN dataset than on ModelNet40. This may" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 277, + 547, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 337 + ], + "type": "text", + "content": "be attributed to the reason that the cross-modal generative pre-training has enhanced the network's ability to understand point clouds from different views, which is beneficial for a more robust understanding of the real-scan data with more noise and disturbance in the ScanObjectNN dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 342, + 548, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 548, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 548, + 439 + ], + "type": "text", + "content": "Visualization Results. Figure 3 shows the visualization results of TAP. The first row shows the generated view images while the second row displays the ground truth images for reference. The TAP method can successfully predict the accurate shape of the object and the RGB colors that represent light reflections in rendered images. Therefore, TAP is capable of capturing the geometric structure of 3D objects and reasoning occlusion relations from specific camera poses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 454, + 419, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 419, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 419, + 467 + ], + "type": "text", + "content": "4.2.2 Part Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 474, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 617 + ], + "type": "text", + "content": "Performing dense prediction is always a more challenging task compared with classification. In this section, we evaluate the local distinguishability of our proposed TAP pre-training method, fine-tuning the pre-trained point cloud model on the ShapeNetPart dataset for the part segmentation task. Quantitative results are shown in Table 3. We implement PointMLP as the backbone model and compare our TAP results with two mainstreams of previous literature. The upper row displays classical architecture-oriented methods that focus on network design and are trained from scratch. The lower row shows members of the generative pre-training family that rely on Transformer-based architectures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 548, + 715 + ], + "type": "text", + "content": "According to results comparisons, our TAP pre-training significantly improves the part segmentation performance of the PointMLP backbone, increasing class mIoU by 0.6 and instance mIoU by 0.8. More importantly, our TAP pretraining achieves state-of-the-art performance on both class mIoU and instance mIoU, surpassing leading works in both tracks. Specifically, TAP exceeds the performance of PointM2AE on instance mIoU by 0.4. This satisfactory perfor" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5646" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": "mance serves as strong evidence to convey that our proposed TAP pre-training is superior to previous uni-model generative pre-training mechanisms in dense prediction tasks. This may be attributed to the factor that our supervision in 2D with MSE loss is more precise than the ambiguous Chamfer Distance in 3D reconstruction. Therefore, models with TAP pre-training obtain more accurate comprehension of local geometry and detail awareness, which contributes to mIoU gain in dense prediction tasks. What's more, TAP outperforms KPCov on instance mIoU by 0.5, demonstrating that the proposed 3D-to-2D generative pre-training method can fully exploit the potential of the point cloud model and help it have a better perception of objects' geometric structure. As TAP is adaptable to any architecture, future improvements in architectural design can also benefit from TAP pre-training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 269, + 179, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 179, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 179, + 281 + ], + "type": "text", + "content": "4.2.3 Few-shot Classification" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 288, + 373 + ], + "type": "text", + "content": "Following Point-BERT [58], we conduct few-shot classification with Standard Transformers on ModelNet40 [52] dataset. As shown in Table 4, we report mean overall accuracy and standard deviation " + }, + { + "bbox": [ + 46, + 289, + 288, + 373 + ], + "type": "inline_equation", + "content": "(\\mathrm{mOA} \\pm \\mathrm{std})" + }, + { + "bbox": [ + 46, + 289, + 288, + 373 + ], + "type": "text", + "content": " on 10 pre-defined data folders for each few-shot setting. The way and shot in Table 4 specify the number of categories and the number of training examples per category, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 374, + 288, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 288, + 517 + ], + "type": "text", + "content": "From the results, TAP achieves the highest mean overall accuracy across all few-shot settings when compared to previous generative pre-training approaches. Furthermore, TAP exhibits significantly lower standard deviations than those reported in the existing literature for the majority of few-shot settings, which signifies its robust performance and consistent superiority. This indicates that TAP is not only capable of achieving high mean overall accuracy but also exhibits reliability and robustness across various few-shot settings. Such stability is crucial in real-world applications, where consistency and predictability are vital for practical deployment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 526, + 208, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 208, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 208, + 538 + ], + "type": "text", + "content": "4.3. Scene-level Dense Predictions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 700 + ], + "type": "text", + "content": "To assess the effectiveness of TAP in handling scene-level dense prediction tasks, we carry out experiments on more complicated scene-level object detection and semantic segmentation on the ScanNetV2 [12] dataset. For the object detection task, we adopt 3DETR [29] and pre-train its encoder on the object-level dataset ShapeNet [6] with TAP. Average precision at 0.25 and 0.5 IoU thresholds are reported. Regarding semantic segmentation, we employ the PointTransformerV2 [51](PTv2) model and pre-train it on the ScanNetV2 dataset with TAP. We report mean IoU for evaluation metric. It is worth mentioning that PTv2 represents the current state-of-the-art approach with open-source code availability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 713 + ], + "type": "text", + "content": "Based on the results presented in Table 5, TAP consis" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 127, + 545, + 229 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 547, + 119 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 547, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 547, + 119 + ], + "type": "text", + "content": "Table 4: Few-shot Classification with Standard Transformers on ModelNet40 dataset. We report mean overall accuracy and standard deviation on 10 pre-defined data folders for each setting. Best results are marked bold." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 127, + 545, + 229 + ], + "lines": [ + { + "bbox": [ + 309, + 127, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 309, + 127, + 545, + 229 + ], + "type": "table", + "html": "
Method5-way10-way
10-shot20-shot10-shot20-shot
w/o pre-training87.8 ± 5.293.3 ± 4.384.6 ± 5.589.4 ± 6.3
Point-BERT [58]94.6 ± 3.196.3 ± 2.791.0 ± 5.492.7 ± 5.1
MaskPoint [23]95.0 ± 3.797.2 ± 1.791.4 ± 4.093.4 ± 3.5
Point-MAE [30]96.3 ± 2.597.8 ± 1.892.6 ± 4.195.0 ± 3.0
TAP97.3 ± 1.897.8 ± 1.793.1 ± 2.695.8 ± 1.0
", + "image_path": "b112527a4f069ca40185e2c4982ebbff230031afbe9765283fc79b8e43e2677f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 308, + 545, + 374 + ], + "blocks": [ + { + "bbox": [ + 304, + 239, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 304, + 239, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 547, + 300 + ], + "type": "text", + "content": "Table 5: Scene-level object detection and semantic segmentation on ScanNetV2 [12]. Average precision at 0.25 IoU thresholds " + }, + { + "bbox": [ + 304, + 239, + 547, + 300 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}_{0.25})" + }, + { + "bbox": [ + 304, + 239, + 547, + 300 + ], + "type": "text", + "content": " and 0.5 IoU thresholds " + }, + { + "bbox": [ + 304, + 239, + 547, + 300 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}_{0.5})" + }, + { + "bbox": [ + 304, + 239, + 547, + 300 + ], + "type": "text", + "content": " of detection and mean Intersection-over-Union (mIoU) of semantic segmentation are reported." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 308, + 545, + 374 + ], + "lines": [ + { + "bbox": [ + 309, + 308, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 309, + 308, + 545, + 374 + ], + "type": "table", + "html": "
MethodDet (3DETR [29])Seg (PTv2 [51])
AP0.25AP0.5mIoU
Baseline62.137.972.4
+TAP63.0 (+0.9)41.4 (+3.5)72.6 (+0.2)
", + "image_path": "f8a3b75ebae62c471b61602142a7a8635157dfb128e0a5558a780ac0fc9fd694.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 394, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 547, + 538 + ], + "type": "text", + "content": "tently enhances the performance of all baselines, thereby showcasing its efficacy in tackling more intricate scene-level dense prediction tasks. Remarkably, even with the encoder solely pre-trained on an object-level dataset for scene-level detection task, significant improvements are observed in both " + }, + { + "bbox": [ + 304, + 394, + 547, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{0.25}" + }, + { + "bbox": [ + 304, + 394, + 547, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 394, + 547, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{0.5}" + }, + { + "bbox": [ + 304, + 394, + 547, + 538 + ], + "type": "text", + "content": " metrics. This suggests that the learned representations from TAP effectively capture relevant information and generalize well to complex scenes, even when the pre-training data is limited to object-level collections. Such generalization capabilities are valuable in scenarios where obtaining large-scale fully annotated scene-level datasets may be challenging or expensive." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 545, + 406, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 545, + 406, + 557 + ], + "spans": [ + { + "bbox": [ + 306, + 545, + 406, + 557 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 563, + 547, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 547, + 612 + ], + "type": "text", + "content": "To investigate the architectural design of our proposed Photograph Module in TAP pre-training pipeline, we conduct extensive ablation studies on the ScanObjectNN dataset with PointMLP as the backbone model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "content": "Photograph Module Architectural Designs. In Photograph Module, we implement cross-attention layers to generate view image feature maps conditioned on pose instruction. We believe that letting the module learn by itself how to rearrange 3D point features in 2D grids will enhance the representation ability of the 3D backbone. Therefore, we conduct ablation studies to verify this hypothesis. As shown in Table 6a, we implement Model " + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_1" + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "content": " with no attention lay" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5647" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 547, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 547, + 173 + ], + "type": "text", + "content": "Table 6: Ablation studies on Photograph module in TAP pre-training pipeline. We choose PointMLP as the backbone model and conduct ablation studies on the ScanObjectNN dataset from two aspects: overall architectural designs and query designs. In Table(a), we first investigate the effectiveness of cross-attention design compared with direct projection (Model A1) and direct project with self-attention (Model A2). Then we analyze the influence of the different number of attention layers and feature channels in Model B and Model C. Finally, we discuss whether pad token in memory builder is beneficial in Model D. In Table(b), we conduct further experiments to compare different approaches for query designing: (1) Using learnable query based on the given viewpoints (Model E) or use mathematical formulations derived in Eq. 6. (2) The information we need when we mathematically encode pose information into init query status: (i) Origin: the coordinate of origin point " + }, + { + "bbox": [ + 46, + 71, + 547, + 173 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 71, + 547, + 173 + ], + "type": "text", + "content": " that the optical line passes through. (ii) Direction: The normalized direction of the optical line. (iii) PE: The position embedding for each grid." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 186, + 318, + 285 + ], + "blocks": [ + { + "bbox": [ + 120, + 175, + 244, + 186 + ], + "lines": [ + { + "bbox": [ + 120, + 175, + 244, + 186 + ], + "spans": [ + { + "bbox": [ + 120, + 175, + 244, + 186 + ], + "type": "text", + "content": "(a) Overall Architectural Designs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 186, + 318, + 285 + ], + "lines": [ + { + "bbox": [ + 50, + 186, + 318, + 285 + ], + "spans": [ + { + "bbox": [ + 50, + 186, + 318, + 285 + ], + "type": "table", + "html": "
ModelAttention TypeLayerNum.ChannelsMem.PadAcc.(%)
A1None-256X87.6 (-0.9)
A2SelfAttn6 layers256X87.8 (-0.7)
BCrossAttn2 layers256✓87.9 (-0.6)
CCrossAttn6 layers512✓87.8 (-0.7)
DCrossAttn6 layers256X88.3 (-0.2)
TAPCrossAttn6 layers256✓88.5
", + "image_path": "6965b9dc9c3650fd98f132215ee79a22973b88811af3e6a623836104ede7473f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 327, + 186, + 547, + 287 + ], + "blocks": [ + { + "bbox": [ + 400, + 175, + 470, + 186 + ], + "lines": [ + { + "bbox": [ + 400, + 175, + 470, + 186 + ], + "spans": [ + { + "bbox": [ + 400, + 175, + 470, + 186 + ], + "type": "text", + "content": "(b) Query Designs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 327, + 186, + 547, + 287 + ], + "lines": [ + { + "bbox": [ + 327, + 186, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 327, + 186, + 547, + 287 + ], + "type": "table", + "html": "
ModelQuery TypeOriginDirectionPEAcc.(%)
ELearnableXXX87.5 (-1.0)
F1FormulaX✓✓86.5 (-2.0)
F2Formula✓XX87.8 (-0.7)
F3Formula✓✓X88.0 (-0.5)
F4Formula✓X✓88.1 (-0.4)
TAPFormula✓✓✓88.5
", + "image_path": "8c927dcad61263ee7d6c6511ec8a5f7df254fba0d0da748cc1cbe5b343715a37.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 307, + 289, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 289, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 289, + 474 + ], + "type": "text", + "content": "ers, directly projecting 3D feature points to 2D grids based on Eq. 3 and Eq. 4. This results in a much simpler pretraining task, as the projection relation has been directly told. Additionally, in Model " + }, + { + "bbox": [ + 46, + 307, + 289, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_2" + }, + { + "bbox": [ + 46, + 307, + 289, + 474 + ], + "type": "text", + "content": ", we add self-attention layers after explicit projection to help the model capture longer-range correlations. Pose knowledge is encoded as a pose token that is concatenated to projected grids, similar to the CLS token in classification Transformers. According to quantitative results comparison with TAP that implements cross-attention layers, fine-tuning results of pre-training methods in " + }, + { + "bbox": [ + 46, + 307, + 289, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_1" + }, + { + "bbox": [ + 46, + 307, + 289, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 307, + 289, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_2" + }, + { + "bbox": [ + 46, + 307, + 289, + 474 + ], + "type": "text", + "content": " version show inferiority. Therefore, the cross-attention architecture we designed to entirely LEARN the projection relation is the most suitable choice for the proposed 3D-to-2D generative pre-training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 476, + 289, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 289, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 289, + 574 + ], + "type": "text", + "content": "What's more, we discuss the number of cross-attention layers, the dimension of feature channels and whether to concatenate pad token in memory builder in Model " + }, + { + "bbox": [ + 46, + 476, + 289, + 574 + ], + "type": "inline_equation", + "content": "B, C, D" + }, + { + "bbox": [ + 46, + 476, + 289, + 574 + ], + "type": "text", + "content": ". According to the results, more cross-attention layers show stronger representation ability, while too large channel number will lead to performance decrease caused by over-fitting. The performance gain from Model " + }, + { + "bbox": [ + 46, + 476, + 289, + 574 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 476, + 289, + 574 + ], + "type": "text", + "content": " to TAP also verifies that the pad token design in the memory builder is essential." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Query Generator Designs. In the query generator, we derive the mathematical formulation of the optical lines passing through 2D grids. We propose to concatenate the coordinate of origin point " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", normalized direction " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{d}^{\\dagger}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and position embedding " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(u / h, v / w)" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " as the initial state of queries. In Table 6b, we first compare this mathematical design with totally learnable queries that takes pose matrix " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " as input and implements MLP layers to predict query for each grid. As shown in Model " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", learnable queries cannot satisfactorily encode pose information, while our derived formulation for query construction is both clearer in physical meaning and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 307, + 476, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 307, + 476, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 476, + 319 + ], + "type": "text", + "content": "more competitive in fine-tuning accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 319, + 547, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 380 + ], + "type": "text", + "content": "In ablation " + }, + { + "bbox": [ + 304, + 319, + 547, + 380 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}" + }, + { + "bbox": [ + 304, + 319, + 547, + 380 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 319, + 547, + 380 + ], + "type": "inline_equation", + "content": "\\mathrm{F_4}" + }, + { + "bbox": [ + 304, + 319, + 547, + 380 + ], + "type": "text", + "content": ", we progressively discuss the three components of query generation. Quantitative comparison with TAP verifies that every component is indispensable for query generation, where coordinates of origin points are of the most importance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 395, + 384, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 384, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 384, + 407 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 415, + 548, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 548, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 548, + 631 + ], + "type": "text", + "content": "In this paper, we have proposed a novel 3D-to-2D generative pre-training method TAP that is adaptable to any point cloud model. We implemented the cross-attention mechanism to generate view images of point clouds from instructed camera poses. To better encode pose conditions and generate physically meaningful queries, we derived mathematical formulations of optical lines. The proposed TAP pre-training had higher preciseness in supervision and broader adaptation to different backbones, compared with directly reconstructing point clouds in previous methods. Experimental results conveyed that the TAP pre-training can help the backbone models better capture the structural knowledge and stereoscopic relations. Fine-tuning results of TAP pre-training achieve state-of-the-art performance on ScanObjectNN classification and ShapeNetPart segmentation, among methods that do not include any pre-trained image or text models. We believe the cross-modal generative pre-training paradigm will be a promising direction for future research." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 641, + 404, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 641, + 404, + 654 + ], + "spans": [ + { + "bbox": [ + 306, + 641, + 404, + 654 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 661, + 547, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 661, + 547, + 709 + ], + "spans": [ + { + "bbox": [ + 305, + 661, + 547, + 709 + ], + "type": "text", + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0114903 and in part by the National Natural Science Foundation of China under Grant 62125603." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5648" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Idan Achituve, Haggai Maron, and Gal Chechik. Self-supervised learning for domain adaptation on point clouds. In WACV, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "text", + "content": "[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In CVPR, pages 9902-9912, June 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 213 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 249 + ], + "type": "text", + "content": "[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. *NeruIPS*, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 282 + ], + "type": "text", + "content": "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 288, + 338 + ], + "type": "text", + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 288, + 384 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners. arXiv preprint arXiv:2006.10029, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 418 + ], + "type": "text", + "content": "[8] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 420, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 288, + 453 + ], + "type": "text", + "content": "[9] Silin Cheng, Xiwu Chen, Xinwei He, Zhe Liu, and Xiang Bai. Pra-net: Point relation-aware network for 3d point cloud analysis. IEEE TIP, 30:4436-4448, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 454, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 486 + ], + "type": "text", + "content": "[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 488, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 521 + ], + "type": "text", + "content": "[11] Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning. In ICML, 2008." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "type": "text", + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 621 + ], + "type": "text", + "content": "[13] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[15] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In ICML, 2021." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "type": "text", + "content": "[17] Abdullah Hamdi, Silvio Giancola, and Bernard Ghanem. MVTN: multi-view transformation network for 3d shape recognition. In ICCV, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 142, + 547, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 547, + 175 + ], + "type": "text", + "content": "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll'ar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 177, + 547, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 547, + 209 + ], + "type": "text", + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 211, + 547, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 211, + 547, + 254 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 547, + 254 + ], + "type": "text", + "content": "[20] Tianxin Huang, Zhonggan Ding, Jiangning Zhang, Ying Tai, Zhenyu Zhang, Mingang Chen, Chengjie Wang, and Yong Liu. Learning to measure the point cloud reconstruction loss in a representation space. In CVPR, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 256, + 547, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 547, + 288 + ], + "type": "text", + "content": "[21] Roman Klokov and Victor Lempitsky. Escape from cells: Deep kd-networks for the recognition of 3d point cloud models. In ICCV, 2017." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 291, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 291, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 545, + 323 + ], + "type": "text", + "content": "[22] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. NeurIPS, 2018." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 325, + 547, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 547, + 357 + ], + "type": "text", + "content": "[23] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In ECCV. Springer, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "type": "text", + "content": "[24] Minghua Liu, Lu Sheng, Sheng Yang, Jing Shao, and Shi-Min Hu. Morphing and sampling network for dense point cloud completion. In AAAI, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 426 + ], + "type": "text", + "content": "[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 428, + 545, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 428, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 307, + 428, + 545, + 450 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2018." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 452, + 545, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 452, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 307, + 452, + 545, + 484 + ], + "type": "text", + "content": "[27] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. In ICLR, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 486, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 545, + 518 + ], + "type": "text", + "content": "[28] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In IROS. IEEE, 2015." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 521, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 545, + 552 + ], + "type": "text", + "content": "[29] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, 2021." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 555, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 588 + ], + "type": "text", + "content": "[30] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV. Springer, 2022." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 590, + 547, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 643 + ], + "type": "text", + "content": "[31] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "text", + "content": "[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 680, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 547, + 713 + ], + "type": "text", + "content": "[33] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++ deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017." + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5649" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[34] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. arXiv preprint arXiv:2302.02318, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 162 + ], + "type": "text", + "content": "[35] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 195 + ], + "type": "text", + "content": "[36] Shi Qiu, Saeed Anwar, and Nick Barnes. Dense-resolution network for point cloud classification and segmentation. In WACV, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 228 + ], + "type": "text", + "content": "[37] Shi Qiu, Saeed Anwar, and Nick Barnes. Geometric backprojection network for point cloud classification. IEEE TMM, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 252 + ], + "type": "text", + "content": "[38] Haoxi Ran, Jun Liu, and Chengjie Wang. Surface representation for point clouds. In CVPR, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 253, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 288, + 285 + ], + "type": "text", + "content": "[39] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Oct-net: Learning deep 3d representations at high resolutions. In CVPR, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 286, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 319 + ], + "type": "text", + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 320, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 288, + 353 + ], + "type": "text", + "content": "[41] Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space. NeurIPS, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 354, + 288, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 388 + ], + "type": "text", + "content": "[42] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In ICCV, 2015." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 388, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 421 + ], + "type": "text", + "content": "[43] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "type": "text", + "content": "[44] Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. ICCV, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "type": "text", + "content": "[45] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 511, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 288, + 544 + ], + "type": "text", + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 545, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 578 + ], + "type": "text", + "content": "[47] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 579, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 612 + ], + "type": "text", + "content": "[48] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. TOG, 2019." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 613, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 646 + ], + "type": "text", + "content": "[49] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. In NeurIPS, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "type": "text", + "content": "[50] Tong Wu, Liang Pan, Junzhe Zhang, Tai Wang, Ziwei Liu, and Dahua Lin. Balanced chamfer distance as a comprehensive metric for point cloud completion. In NeurIPS, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[51] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 453 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[52] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 139 + ], + "type": "text", + "content": "[53] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 141, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 174 + ], + "type": "text", + "content": "[54] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 175, + 547, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 229 + ], + "type": "text", + "content": "[55] Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. arXiv preprint arXiv:2106.04180, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 230, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 547, + 274 + ], + "type": "text", + "content": "[56] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ToG, 2016." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 275, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 547, + 308 + ], + "type": "text", + "content": "[57] Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In ICCV, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 309, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 547, + 342 + ], + "type": "text", + "content": "[58] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 343, + 547, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 547, + 365 + ], + "type": "text", + "content": "[59] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 365, + 547, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 547, + 410 + ], + "type": "text", + "content": "[60] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. arXiv preprint arXiv:2205.14401, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 411, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 547, + 453 + ], + "type": "text", + "content": "[61] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In CVPR, 2023." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5650" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/7d213d41-b6ba-4fcf-a556-dbfb8dc18cb6_content_list.json b/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/7d213d41-b6ba-4fcf-a556-dbfb8dc18cb6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8f02effe1ea7a2950e82227a1deba3c2faa3142e --- /dev/null +++ b/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/7d213d41-b6ba-4fcf-a556-dbfb8dc18cb6_content_list.json @@ -0,0 +1,1554 @@ +[ + { + "type": "text", + "text": "Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors", + "text_level": 1, + "bbox": [ + 98, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhentao Yu $^{1*}$ Zixin Yin $^{1,2*†}$ Deyu Zhou $^{1,3*†}$ Duomin Wang $^{1}$ Finn Wong $^{1}$ Baoyuan Wang $^{1‡}$ Xiaobing.AI.", + "bbox": [ + 215, + 178, + 750, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ The Hong Kong University of Science and Technology. $^{3}$ The Hong Kong University of Science and Technology (Guangzhou).", + "bbox": [ + 205, + 233, + 759, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(yuzhentao, wangduomin, wangwenlan, wangbaoyuan)@xiaobing.ai zyinaf@connect.ust.hk, dzhou861@connect.hkust-gz.edu.cn", + "bbox": [ + 233, + 272, + 736, + 304 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bfba7cfa543a582cf526bb6a6de99c5ebfa6d99a1ed18db19c2e179cfdbc5791.jpg", + "image_caption": [ + "Figure 1. Given only an audio source and an arbitrary identity image, our system can generate a video with natural-looking and diverse facial motions (pose, expression, blink & gaze), while maintaining accurate audio-lip synchronization. Here we show randomly sampled sequences from our diffusion prior for two identities, note that the lip-irrelevant facial motion varies but the lip is still in-sync." + ], + "image_footnote": [], + "bbox": [ + 81, + 338, + 890, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 630, + 313, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce a novel framework for one-shot audio-driven talking head generation. Unlike prior works that require additional driving sources for controlled synthesis in a deterministic manner, we instead sample all holistic lip-irrelevant facial motions (i.e. pose, expression, blink, gaze, etc.) to semantically match the input audio while still maintaining both the photo-realism of audio-lip synchronization and overall naturalness. This is achieved by our newly proposed audio-to-visual diffusion prior trained on top of the mapping between audio and non-lip representations. Thanks to the probabilistic nature of the diffusion prior, one big advantage of our framework is it can synthe", + "bbox": [ + 75, + 661, + 470, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "size diverse facial motion sequences given the same audio clip, which is quite user-friendly for many real applications. Through comprehensive evaluations of public benchmarks, we conclude that (1) our diffusion prior outperforms autoregressive prior significantly on all the concerned metrics; (2) our overall system is competitive with prior works in terms of audio-lip synchronization but can effectively sample rich and natural-looking lip-irrelevant facial motions while still semantically harmonized with the audio input.", + "bbox": [ + 500, + 631, + 890, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 814, + 632, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Audio-driven face reenactment and talking head generation have received raising attention due to the broad killer applications in movie production, gaming, virtual digital avatars, and potentially more while we move toward the era", + "bbox": [ + 500, + 840, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors have contributed equally to this work.", + "bbox": [ + 94, + 851, + 379, + 864 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was done when Zixin Yin and Deyu Zhou were interns at XiaoBing.AI.", + "bbox": [ + 76, + 864, + 468, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding author.", + "bbox": [ + 94, + 888, + 220, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7645", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of the metaverse. The past literature tries to advance this area from various perspectives. One line of research focuses on improving the generation quality originating from regular GAN-based methods [64, 65] to pre-trained StyleGAN [3] and to the most recent Neural Radiance Field (NeRF) based methods [13, 42]. Orthogonal to direction, another line of works emphasizes the importance of disentangled representation from the audio [32, 21, 66] for controlled generations. i.e., [21] can predict the emotion from the audio input while [66] can decouple the speech signal into speaker identity and the phonetic content. A closely related set of works targeting more granular controlled talking head generation by providing additional input signals. For example, PC-AVS [65] requires a separate video to provide the pose signal in addition to the audio clip but does not support the control of expression and blinking. To remedy this, GC-AVT [27] requires inputting additional driving video for expression in addition to pose and audio driving clip. The technical challenge behind those approaches is how to faithfully transfer the desired driving signal (i.e., pose, expression, audio) into the results without affecting each other through intrinsic disentanglement. Although encouraging results have been reported, we argue such a setting is not practical for broader applications. It is quite challenging, or at least labor-intensive, for novice users to find the \"best\" individual driving sources for each controlled dimension to make the final talking head video overall look not only lifelike but also coherent from semantic and emotional perspectives. Therefore, it is both more practical and generalizable if the setting only requires an audio signal as the driving source and expects a data-driven model to sample reasonable other facial motions irrelevant to lips.", + "bbox": [ + 76, + 90, + 472, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By no means, we are the first to advocate an audio-only driving setup for talking head generation. Audio2head [54] predicts the poses from audio input with an LSTM network. [29] employs an autoregressive model by assuming the poses are jointly determined by the audio and past head motions along the sequence, although the results are encouraging, the model can't generalize due to their person-specific training strategy. FACIAL [61] could infer the blink, but requires a reference video input rather than a one-shot reference image. There are other related works along this direction, but they either only infer one facial attribute (i.e., emotion in EVP[21], pose from [54]) using ad-hoc methods, or they simply do not support inferring other facial motions [38], or treat it as a block-box mapping model [3] without explicitly respecting the one-to-many mapping nature between audio and other visual facial attributes (pose, expression, blink). A more principled solution is desired to consolidate this line of work.", + "bbox": [ + 76, + 579, + 472, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel framework that can holistically infer all the non-lip-related facial attributes from the audio input while maintaining accurate synchronization", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "between audio and the corresponding visual lip motions. This is achieved by two important learning steps in our pipeline. (1) A pre-trained identity-irrelevant facial motion representation can help decouple the lip and non-lip representations. To promote this learning, we employ a novel orthogonal loss on top of the modified facial reenactment framework [6]. The disentanglement enables to generate one-to-one mapping with lip representations to ensure the synchronization, and generate richer non-lip representations with a one-to-many mapping. (2) A novel audio-to-visual diffusion prior model is introduced to address the probabilistic sampling from audio representations to the abovelearned non-lip representation. This prior is expected to solve the one-to-many mapping and provide diverse results during the inference stage. The entire pipeline can be easily built up on top of the existing framework, such as PC-AVS[65] without heavily retraining every component. To sum up, we make the following contributions:", + "bbox": [ + 496, + 90, + 893, + 364 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To our best knowledge, we are the first to holistically predict non-lip facial motions based on audio input only, providing good usability for reenactment or dubbing applications without extra driving video sources. Our method addresses the intrinsic one-to-many challenge in a probabilistic way, allowing diverse and realistic facial motion generation under the same audio input, as shown in Fig. 1.", + "- We leverage the pre-trained visual identity-irrelevant facial motion representations. And further, learn disentangled lip-related and lip-irrelevant representations through a novel orthogonal loss on top of PC-AVS[65]. A powerful diffusion prior model is then introduced to effectively infer all lip-irrelevant facial motions for a given audio segment in the representation space.", + "- We systematically evaluate the naturalness and diversity of the results with new metrics, which paves a way for future studies. Meanwhile, results show that our method can produce natural-looking head poses and facial motions without hurting the audio-lip synchronization. Our model will be released." + ], + "bbox": [ + 500, + 373, + 890, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 709, + 650, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio-Visual Cross-modal Learning Cross-modal representation learning is a long-standing research topic, ranging from speech enhancement [11], speech source separation [26] to synchronization [8, 22] and other speech disentanglement [34, 21, 32]. Among them, EVP [21] used cross-modal supervision to disentangle speech content and emotion from the audio signal with landmark as the intermediate representation. Recently, CMC [49] discussed how multiview \"modality\" can be jointly unitized to boost intrinsic representation through contrastive learning, rather than predictive (or reconstruction) learning, it also demonstrated", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7646", + "bbox": [ + 482, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that the more views, the better. Concurrently, MMV [2] introduced different modality embedding graphs for effective cross-modal representations, again through contrastive learning. More recently, HCMoCo [16] extended similar ideas with a hierarchical strategy to learn different levels of representations for human-centric perception tasks. Although impressive results were reported, it is still unclear how it performs on face analysis tasks, not mention to on synthesis tasks. Our work shares a similar spirit but a different purpose, where we first pre-train a non-identity visual representation which then helps learn a lip and non-lip space, the latter further serves as the upper-bound learning target for subsequent audio-to-visual diffusion prior.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Face Reenactment & Talking Head Generation Face Reenactment is designed to transfer part or full facial motion from a driving source to the target video with good IDpreserved appearance and background. It can be further divided into two categories depending on whether the driving source is from video [17, 18, 5, 48, 24, 36, 59, 19, 60, 6, 56] or audio [52, 47, 62, 21, 55, 37, 63, 30]. Among them, audio-driven face reenactment generally aims to edit the mouth regions of the target video in order to match the input audio while leaving other facial attributes mostly unchanged, i.e., pose. EVP [21] tries to infer the non-rigid facial expression in addition to lip motion from the audio input. A closely related line of work is the audio-driven talking-head generations [38, 66, 61, 3] where only one target reference face is given, hence, other face attributes including pose, expression, blink and gaze have to be either explicitly given [65, 27] or partly inferred through statistical methods [66, 31, 61, 29, 3, 47]. Specifically, Lu et al. [29] employed an auto-regressive model while Min et al. [31] leveraged normalized flow prior to predicting a natural-looking pose sequence from the input audio, both showed encouraging results. Compared with them, we aim to infer more diverse facial motions including pose, expression, and even blink and gaze in a holistic manner through an audio-to-visual diffusion prior model.", + "bbox": [ + 75, + 299, + 472, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Generative Models The diffusion model [44, 15, 45], which is a likelihood-based model consisting of cascading denoising autoencoders, has recently shown great success in numerous generative tasks with different modalities including image [10, 35, 40, 41], audio [25], video [43], and motion [46]. To name a few, DDPM [15] explored the diffusion model for unconditional image generation. GLIDE [35] introduced text-conditional diffusion model and showed that classifier free guidance has better performance than CLIP [39] guidance. DALLE-2 [40] modified GLIDE to generate semantically consistent images conditioned on a CLIP image embedding, and proposed a diffusion prior that produces the image embedding given a text caption. MDM [46] utilized a classifier-free diffusion-based", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "generative model for text-to-motion and action-to-motion tasks, allowing motion completion and editing as well.", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 133, + 591, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given $L$ segments of audio $A_{1:L} = (A_1, A_2, \\dots, A_L)$ and a reference image $I_{ref}$ as inputs, our model M synthesizes video frames $\\hat{X}_{1:L} = (\\hat{X}_1, \\hat{X}_2, \\dots, \\hat{X}_L)$ having the same identity as $I_{ref}$ and lip motion synchronized to $A_{1:L}$ :", + "bbox": [ + 498, + 157, + 893, + 223 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {1: L} = \\mathbf {M} \\left(A _ {1: L}, I _ {r e f}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 229, + 890, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of our proposed framework is shown in Fig. 2, which consists of three major components:", + "bbox": [ + 498, + 256, + 890, + 287 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Lip & Non-lip Disentanglement Given a pre-trained identity- and appearance-irrelevant facial motion encoder $\\mathrm{E}_v$ , we first leverage it to learn two complementary features, $\\mathbf{f}_l^a$ for lip, and $\\mathbf{f}_{nl}^v$ for non-lip features including pose, expression, blink and gaze.", + "- Audio to Non-lip Diffusion Prior A diffusion prior network $\\mathrm{P}_{a2nl}$ models the one-to-many mapping from audio feature $\\mathbf{f}^a$ to the \"hallucinated\" non-lip feature $\\mathbf{f}_{nl}^a$ , which is trained to be close to its visual counterpart $\\mathbf{f}_{nl}^v$ , allowing audio-only facial driving at inference time.", + "- Audio-based Talking Head Generation Given an identity feature $\\mathbf{f}_{id}$ , a lip feature $\\mathbf{f}_l^a$ , and a non-lip feature $\\mathbf{f}_{nl}^a$ generated with the diffusion prior, we concatenate them together and feed into a GAN similar to LPD [6] or PC-AVS [65] to output the final reenacted video." + ], + "bbox": [ + 500, + 294, + 890, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Lip & Non-lip Disentanglement", + "text_level": 1, + "bbox": [ + 500, + 545, + 779, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The identity encoder $\\mathrm{E}_{id}$ is designed to capture identity and appearance information, while $\\mathrm{E}_v$ is instead to remove both while encoding all facial motions. Both encoders are pre-trained in a similar setting as LPD [6] and PC-AVS [65]. After pretraining, we conduct lip & non-lip disentanglement consisting of audio-visual contrastive learning for lip feature learning and decorrelation for non-lip feature learning, as well as face reconstruction, as described below:", + "bbox": [ + 496, + 568, + 890, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio-Visual Pretraining for Audio Lip Space An audio encoder $\\mathrm{E}_a$ is trained to provide accurate audio feature $\\mathbf{f}^a = \\mathrm{E}_a(A_{1:N})$ through contrastive learning [39] with $\\mathbf{f}_{cl}^v = \\mathrm{MLP}_{cl}(\\mathrm{E}_v(X_{1:N}))$ , where $N$ is the number of samples and $X_{1:N}$ are $N$ frames in the same video corresponding to audio $A_{1:N}$ . Then $\\mathbf{f}^a$ is further compressed to filter out lip-relevant information and get audio lip feature $\\mathbf{f}_l^a$ through $\\mathrm{MLP}_{a2l}$ since $\\mathbf{f}^a$ mainly contains lip related information, as shown in Sec. 4.3. The contrastive loss is defined as $L_{con}(m,n) = -\\frac{1}{N}\\sum_{i=1}^{N}\\log\\frac{\\exp(m_i/\\|m_i\\|_2 \\cdot n_i/\\|n_i\\|_2)}{\\sum_{j=1}^{N}\\exp(m_i/\\|m_i\\|_2 \\cdot n_j/\\|n_j\\|_2)}$ , resulting in a total contrastive loss between $\\mathbf{f}^a$ and $\\mathbf{f}_{cl}^v$ ,", + "bbox": [ + 496, + 696, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c l} = \\frac {1}{2} \\left[ L _ {\\text {c o n}} \\left(\\mathbf {f} ^ {a}, \\mathbf {f} _ {c l} ^ {v}\\right) + L _ {\\text {c o n}} \\left(\\mathbf {f} _ {c l} ^ {v}, \\mathbf {f} ^ {a}\\right) \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 875, + 890, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7647", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f1a80a4f86f0080f1fe4c303c1bdb2d75be6d9fd455043f7e18b5a635bc1e1c3.jpg", + "image_caption": [ + "Figure 2. The overall pipeline of our proposed framework. The dotted lines represent the loss functions that are used only in training, i.e., $L_{cl}$ , $L_{ol}$ and $L_{a2nl}$ . $L_{G}$ is the training loss for visual non-lip space. Note that there is a switch in the figure, which means the different forward processes in training and inference: in training, the concatenated feature $\\mathbf{f}_{cat} = cat(\\mathbf{f}_{id},\\mathbf{f}_{nl}^{v},\\mathbf{f}_{l}^{a})$ ; at inference stage, $\\mathbf{f}_{cat} = cat(\\mathbf{f}_{id},\\mathbf{f}_{nl}^{a},\\mathbf{f}_{l}^{a})$ . Thus, the visual motion part is not needed anymore at the inference stage." + ], + "image_footnote": [], + "bbox": [ + 89, + 94, + 888, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction Learning for Visual Non-lip Space To learn a good non-lip space, we propose an orthogonal loss to penalize the correlation between $\\mathbf{f}_{ol}^{v}$ and $\\mathbf{f}^{\\alpha}$ , which uses the well-learned audio space to disentangle lip-related motions. The orthogonal loss is used together with reconstruction since there should also be a completeness constraint to avoid mode collapse of lip-irrelevant features. In practice, we maintain two memory banks (MB) for storing $\\mathbf{f}_{ol}^{v}$ and $\\mathbf{f}^{\\alpha}$ in previous $K - 1$ steps, in order to have more samples than $N$ . We denote the feature dimension of $\\mathbf{f}^{\\alpha}$ as $d_{a}$ , and $\\mathbf{f}_{ol}^{v}$ as $d_{v}$ . The orthogonal loss is defined as follows:", + "bbox": [ + 75, + 392, + 472, + 559 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {o l} = \\frac {1}{d _ {v}} \\sum_ {i = 1} ^ {d _ {v}} \\sum_ {j = 1} ^ {d _ {a}} P _ {c o r} \\left(\\mathbf {f} _ {o l} ^ {v}, \\mathbf {f} ^ {a}\\right) _ {(i, j)} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 564, + 468, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{cor}(\\mathbf{f}_{ol}^v,\\mathbf{f}^a)\\in d_a\\times d_v$ computes the Pearson correlation coefficient between $\\mathbf{f}_{ol}^{v}$ and $\\mathbf{f}^a$ . Since the correlation between lip motions and audio features is strong, $L_{ol}$ (which encourages the linear uncorrelation between $\\mathbf{f}_{ol}^{v}$ and $\\mathbf{f}^a$ ) is effective to disentangle lip-related motions. Then, $\\mathrm{MLP}_{nl}$ further projects $\\mathbf{f}_{ol}^{v}$ into a smaller space to remove undesired content and get non-lip feature $\\mathbf{f}_{nl}^{v}$ .", + "bbox": [ + 75, + 613, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow the reconstruction loss $L_{\\mathrm{GAN}}$ , $L_{L1}$ and $L_{\\mathrm{VGG}}$ used in [65]. The generator $G$ receives the concatenation $\\mathbf{f}_{cat} = cat(\\mathbf{f}_{id}, \\mathbf{f}_{nl}^{v}, \\mathbf{f}_{l}^{a})$ as inputs and generates final images $I_{gen}$ with modulated convolution. The discriminator $D$ is trained jointly by generative adversarial learning [23, 12]. Note that we omit the batch-mean operation here for convenience, but in practice, all losses used here are calculated with the average in a batch. Different from [65], we propose an additional gaze loss $L_{gaze}$ by utilizing a pre-trained gaze encoder [1], which calculates the $L1$ distance between the gaze features of generated images and ground-truth (GT) images $I$ as follows:", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g a z e} = \\left\\| \\Phi (I) - \\Phi \\left(I _ {g e n}\\right) \\right\\| _ {1}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 385, + 892, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The total loss $L_{G}$ is given as:", + "bbox": [ + 519, + 405, + 718, + 420 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {G} = \\lambda_ {o l} \\cdot L _ {o l} + \\lambda_ {g a z e} \\cdot L _ {g a z e} + \\lambda_ {L 1} \\cdot L _ {L 1} \\tag {5} \\\\ + \\lambda_ {\\text {G A N}} \\cdot L _ {\\text {G A N}} + \\lambda_ {\\text {V G G}} \\cdot L _ {\\text {V G G}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 426, + 892, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ s are weights for losses.", + "bbox": [ + 500, + 465, + 718, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Audio2nonlip Diffusion Prior", + "text_level": 1, + "bbox": [ + 500, + 488, + 761, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Beside a strong relationship with lip motions, audio features have a complex non-linear relationship with non-lip motions and it is a one-to-many mapping problem. This is due to the fact that there are many reasonable facial motions that can be corresponded to the same audio input. As inspired by DALLE-2 [40] and MDM [46], we design an audio2nonlip diffusion prior network $\\mathrm{P}_{a2nl}$ to address this problem. Our proposed network is depicted in Fig. 3.", + "bbox": [ + 496, + 511, + 892, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion Model Diffusion model is designed based on the stochastic diffusion process. The forward diffusion process is defined below:", + "bbox": [ + 496, + 638, + 893, + 683 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(n _ {1: L} ^ {t} \\mid n _ {1: L} ^ {t - 1}\\right) = \\mathcal {N} \\left(n _ {1: L} ^ {t - 1}; \\sqrt {\\alpha^ {t}} n _ {1: L} ^ {t - 1}, (1 - \\alpha^ {t}) I\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 688, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our context, $n_{1:L} = (n_1,n_2,\\dots,n_L)$ is a sequence of non-lip feature $\\mathbf{f}_{nl}^v$ . $\\alpha^t$ is a hyper-parameter and $t\\sim [1,T]$ is the time step of the diffusion process. Eq. 6 can be approximated to $n_{1:L}^{T}\\sim \\mathcal{N}(0,1)$ if $\\alpha^t$ is small enough.", + "bbox": [ + 496, + 712, + 893, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the reversed diffusion process, $\\mathrm{P}_{a2nl}$ models the audio2nonlip distribution as $p(n_{1:L}^{0}|a_{1:L})$ , where $a_{1:L} = (a_1,a_2,\\dots,a_L)$ is a sequence of audio feature $\\mathbf{f}^a$ . Instead of predicting the noise as formulated by vanilla DDPM [45], $\\mathrm{P}_{a2nl}$ learns to predict the initial signal itself with the simplified objective function [45] described as follows:", + "bbox": [ + 496, + 773, + 893, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s i m p l e} = \\mathbb {E} _ {n ^ {0} \\sim q (n ^ {0} | \\mathbf {f} ^ {\\alpha}), t \\sim [ 1, T ]} \\left[ \\left\\| n ^ {0} - \\mathrm {P} _ {a 2 n l} \\left(n ^ {t}, t, \\mathbf {f} ^ {\\alpha}\\right) \\right\\| _ {2} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 867, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7648", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/942fce6204997a8e2459c4065a7ffff1084e5a0fbfa0f70974050a23eda17ad9.jpg", + "image_caption": [ + "Figure 3. Our audio2nonlip diffusion prior network $\\mathrm{P}_{a2nl}$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 444, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we use $n^0$ to represent $n_{1:L}^{0}$ for convenience.", + "bbox": [ + 76, + 335, + 416, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement $\\mathrm{P}_{a2nl}$ with a transformer encoder [51] architecture. As shown in Fig. 3, an audio feature $a_{1:L}$ is added to the time embedding $\\mathrm{Emb}_{\\text{time}}$ . Then, it's concatenated with noisy non-lip feature $n_{1:L}^{t}$ and added with positional embeddings $\\mathrm{Emb}_{pos}$ [51]. The model encodes the embeddings with bidirectional self-attention and feedforward layers then output denoised non-lip feature $n_{1:L}^{0}$ .", + "bbox": [ + 75, + 353, + 467, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Velocity Loss Note that our designed $\\mathrm{P}_{a2nl}$ denoises $n_{1:L}^{t}$ in a non-autoregressive way. To encourage the naturalness and coherence of generated non-lip motion, we borrow the velocity loss from MDM [46] defined as follows:", + "bbox": [ + 76, + 473, + 468, + 534 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {v e l} = \\frac {1}{L - 1} \\sum_ {i = 2} ^ {L} \\left\\| \\left(\\hat {n} _ {i} ^ {0} - \\hat {n} _ {i - 1} ^ {0}\\right) - \\left(n _ {i} ^ {0} - n _ {i - 1} ^ {0}\\right) \\right\\| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 547, + 468, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_i^0$ is the GT non-lip feature and $\\hat{n}_i^0$ is the predicted denoised non-lip feature at the $i$ -th position respectively. The intuition here is that the difference between adjacent non-lip features should be close to the difference in the GT. The total loss of this stage is defined as:", + "bbox": [ + 75, + 603, + 468, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a 2 n l} = L _ {\\text {s i m p l e}} + L _ {v e l}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 695, + 468, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Classifier-free Guidance We use classifier-free guidance [35] for conditioned diffusion generation. In training time, we randomly set the condition to $\\emptyset$ for $10\\%$ of the samples so that $\\mathrm{P}_{a2nl}(n^t,t,\\emptyset)$ approximates $n^0$ . At inference stage, the output of the $P_{a2nl}$ is extrapolated further in the direction of $\\mathrm{P}_{a2nl}(n^t,t,\\mathbf{f}^a)$ and away from $\\mathrm{P}_{a2nl}(n^t,t,\\emptyset)$ :", + "bbox": [ + 75, + 739, + 468, + 832 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P} _ {a 2 n l} \\left(n ^ {t}, t, \\mathbf {f} ^ {a}\\right) = s \\cdot \\mathrm {P} _ {a 2 n l} \\left(n ^ {t}, t, \\mathbf {f} ^ {a}\\right) + (1 - s) \\cdot \\mathrm {P} _ {a 2 n l} \\left(n ^ {t}, t, \\emptyset\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 845, + 467, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s$ is a scaling parameter while increasing it improves sample quality at the cost of diversity.", + "bbox": [ + 76, + 878, + 468, + 907 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sequential Mask Editing To deliver a smooth non-lip facial sequence with arbitrary length, we design a mechanism to ensure continuity between generated segments, and provide an editing method to do so. Specifically, we randomly mask $90\\%$ of the non-lip tokens $n^0$ and concatenate them with noisy non-lip tokens $n^t$ as the input of $\\mathrm{P}_{a2nl}$ at training stage. Practically, we found it easy for $\\mathrm{P}_{a2nl}$ to quickly learn how to fill in the masked region using hints from the non-lip features in the unmasked regions, generating continuous feature prediction without any extra design in training loss. As a result, it is possible to use $\\mathrm{P}_{a2nl}$ for audio-guided non-lip motion generation as well as non-lip motion editing at inference time, providing non-lip facial motion with good naturalism and diversity.", + "bbox": [ + 496, + 90, + 890, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Audio-based Talking Head Generation", + "text_level": 1, + "bbox": [ + 500, + 313, + 830, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The main difference at inference time is that non-lip visual feature $\\mathbf{f}_{nl}^{v}$ is replaced with $\\mathbf{f}_{nl}^{a}$ generated by $\\mathrm{P}_{a2nl}$ through reversed diffusion process. Thanks to the random noise introduced in the diffusion process, $\\mathrm{P}_{a2nl}$ can generate various visual non-lip features given the same audio input, resulting in diverse and reasonable reenactment videos without any extra needs of driving video sources.", + "bbox": [ + 496, + 337, + 890, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To generate a video sequence longer than the maximal length of the model input with smooth transitions, the first input non-lip feature to $\\mathrm{P}_{a2nl}$ is set to the last generated non-lip feature from $\\mathrm{P}_{a2nl}$ in the previous step, to ensure continuity in non-lip facial motion.", + "bbox": [ + 496, + 444, + 890, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 535, + 632, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed method was evaluated from two aspects: 1) synchronization between audio and lip motions; 2) naturalness and richness of lip-irrelevant facial motions. Both quantitative and qualitative experiments were conducted to showcase the superiority of our method. In addition, we also come up with a novel metric to measure the overall quality of generated facial motions.", + "bbox": [ + 496, + 561, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets Our model was trained on VoxCeleb2 and evaluated on both VoxCeleb1 and VoxCeleb2. Here are the details of the two datasets,", + "bbox": [ + 496, + 676, + 890, + 723 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- VoxCeleb1 [33]: The dataset consists of 100,000 utterances from 1,251 celebrities. 100 videos with 25 identities in total were randomly chosen for the test.", + "- VoxCeleb2 [7]: The dataset contains 1 million utterances with 6,112 identities. 500 test videos with 25 identities were randomly chosen for the test." + ], + "bbox": [ + 500, + 734, + 890, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparing Methods Some prior works such as PC-AVS [65], GC-AVT [28], and EAMM [20] require additional driving sources instead of predicting them based on", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7649", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7cf2f39c6995ed4aedca8d13bc8376ebde8bbefd54d0cdbb61461aa9da3948ac.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodA S SctgntVoxCeleb2VoxCeleb1
FID ↓Sc ↑Nc ↓FID ↓Sc ↑Nc ↓
Wav2Lip [38]X X7.8022.39.230.1844.78.800.13
EAMM [20]X X1.7626.14.751.7042.62.780.58
PC-AVS [65]X X7.3514.48.210.1235.48.420.15
MakeItTalk [66]✓ X7.3519.52.030.7240.22.160.71
Audio2head [54]✓ X7.351016.420.131046.650.10
Ours + AR✓ X7.3524.37.230.0249.77.050.04
Ours + Diff.✓ ✓7.3514.27.340.0035.07.310.01
", + "bbox": [ + 78, + 89, + 472, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. The quantitative results of synchronization and image quality under self-reenactment scenario. Bold means the best. A indicates the ability to generate non-lip motions with audio-only signals without other signals, while S means the ability to sample different results for each round.", + "bbox": [ + 75, + 244, + 468, + 315 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fa66ce24ff507467591c54de0277c266fe026e082abfdfd957815d6a51ee5d9f.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 83, + 327, + 189, + 415 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/169ffe2b44a4265ecc6b4f53f1dd08d6067f130267ffffd8e4f97d357b3d87c0.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 196, + 329, + 297, + 415 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/50344fa4daa4e9c0bbd39cf47f316cf198d205962fde5ad2713c8f3585d41199.jpg", + "image_caption": [ + "(c)", + "Figure 4. Distribution visualization of generated poses." + ], + "image_footnote": [], + "bbox": [ + 303, + 329, + 408, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "audio input. We only compare those methods on lip synchronization and image quality but emphasize that our setup is targeting a more practical scenario. Other works such as Wav2Lip [38] only focus on revising the lip region without touching other facial motions, which are not directly comparable either. MakeItTalk [66] and Audio2head [54] are two closely related works to ours in terms of problem setup, however, our method is designed to predict all the non-lip facial motions with diversity, rather than only limited poses.", + "bbox": [ + 75, + 462, + 468, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details Backbones of our models including $\\mathrm{E}_{id}$ , $\\mathrm{E}_v$ , $\\mathrm{E}_a$ , $\\mathrm{G}$ , and $\\mathrm{D}$ are borrowed from [65]. For the diffusion prior network $\\mathrm{P}_{a2nl}$ , we use a 8-layer transformer [51] with 512-d tokens and 1024-d fully feedforward layers. Note that our models were trained on VoxCeleb2 only but tested on both VoxCeleb1 and VoxCeleb2. All models were trained on 4 NVIDIA A100 GPUs.", + "bbox": [ + 75, + 608, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 724, + 299, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics We evaluate the performance of generated talking head from the following aspects: image quality, audio-lip sync accuracy, the variation and naturalness of non-lip facial motion. Frechet Inception Distance (FID) score [14] is used for the evaluation of image quality. A lower FID score indicates a lower distance between the distribution of generated images and real images. Following prior works [38, 65], we use SyncNet Error-Confidence $(\\mathbf{S}_c)$ as the indicator of audio-lip synchronization, where greater confidence indicates better synchronization. However, the", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "normalized confidence score NLSE-C (short for $\\mathbf{N}_c$ ) is used to address the concern raised in recent works [57, 53] regarding the strong relationship between $\\mathbf{S}_c$ and its training data, which can make it unfair to compare methods trained on different data. $\\mathbf{N}_c$ is defined as $\\mathbf{N}_c = \\frac{\\left|\\mathbf{S}_c^{gen} - \\mathbf{S}_c^{gt}\\right|}{\\mathbf{S}_c^{gt}}$ , where $\\mathbf{S}_c^{gen}$ is the confidence value of generated images and $\\mathbf{S}_c^{gt}$ is that of its training data.", + "bbox": [ + 496, + 90, + 890, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate facial motions such as pose, expression, and blink, we utilize a pre-trained 3D morphable face model [9] and include shape irrelevant 3DMM parameters to calculate the following metrics.", + "bbox": [ + 496, + 205, + 890, + 267 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Var: The variance of generated facial motions, i.e., the variance of the 3DMM coefficients for each video is calculated and then averaged over the test set. A closer Var to GT indicates a better match to the variation of real data.", + "- $\\mathbf{FID}_{fm}$ : FID score of 3DMM coefficients calculated as follows: $\\mathbf{FID}_{fm} = \\frac{1}{K}\\sum_{i=1}^{K}\\mathbf{FID}(\\beta_i)$ , where $\\beta_i$ is a sequence of 3DMM coefficients in the $i$ -th video.", + "- $\\mathbf{FID}_{\\Delta fm}$ : FID score of 3DMM coefficient difference between consecutive frames, i.e., $\\mathbf{FID}_{\\Delta fm}$ , which is similar to $\\mathbf{FID}_{fm}$ but the 3DMM coefficient difference between consecutive frames is measured, taking temporal naturalness into consideration.", + "- SND: Our new proposed metric, denoted as Sequence Naturalness Distance. It is the sum of $\\mathbf{FID}_{fm}$ and $\\mathbf{FID}_{\\Delta fm}$ indicating the difference of distribution between generated motion and GT motion, from both spatial and temporal perspectives, i.e., $\\mathbf{SND} = \\mathbf{FID}_{fm} + \\mathbf{FID}_{\\Delta fm}$ . The lower SND, the better naturalness." + ], + "bbox": [ + 500, + 276, + 890, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Results Our main results consist of two parts: audio-lip synchronization and image quality as shown in Table 1; richness and sequence naturalness which are shown in Table 2.", + "bbox": [ + 496, + 583, + 890, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It shows that our method with diffusion prior archives the best synchronization ability compared to other methods in Table 1. Note that although Wav2Lip [38] and PC-AVS [65] achieve the highest SyncNet scores, our best $N_{c}$ indicates the most substantial synchronization ability [57], which is demonstrated in Sec. 4.2. Meanwhile, the image quality of our model with diffusion prior is the best in both test sets.", + "bbox": [ + 496, + 643, + 890, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 2, our method shows significantly higher variance than the other two audio-driven methods and it is closer to the variance of real data. This indicates that our method can produce reasonable diverse head movements and expressions, rather than slight head movement. It's notable that auto-regressive prior, trained with causal attention and regression loss, has a higher variance than GT because it often generates extreme motions, showing a low naturalness score in Table 3. For naturalness, our method achieves the lowest $\\mathbf{FID}_{fm}$ , $\\mathbf{FID}_{\\Delta fm}$ as well as SND on", + "bbox": [ + 496, + 750, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7650", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c5e3e8c1fc0624cf461fd3a211375fea8e9ba026d1471fa455d940cf5a2ecf2f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVoxCeleb2VoxCeleb1
Var →FIDfm ↓FIDΔfm ↓SND ↓Var →FIDfm ↓FIDΔfm ↓SND ↓
GT1.98---1.88---
MakeItTalk [66]0.674.701.746.440.804.271.075.34
Audio2head [54]0.895.941.307.240.765.031.016.04
Ours + AR3.075.432.227.652.926.211.687.89
Ours + Diff.1.573.601.084.681.663.980.874.85
w/o Lvel2.094.021.275.292.284.651.115.76
training w/o editing3.096.762.008.763.066.961.708.66
", + "bbox": [ + 124, + 88, + 848, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. The quantitative results of variance and naturalness of generated non-lip motions under self-reenactment scenario. $\\rightarrow$ indicates a closer score to GT is better. Due to setting differences, i.e., Wav2Lip, EAMM, and PC-AVS, are not valid comparisons here.", + "bbox": [ + 75, + 239, + 893, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14c2cd94afeb0438c0797a72a0544111867c8871db3289bc32b5799eb11c448c.jpg", + "image_caption": [ + "Figure 5. Qualitative results of our method compared to other baselines. Each row shows nine uniformly sampled frames from videos. Here we use two audio sources to drive different identities respectively. Our method shows accurate lip-audio synchronization with diverse and natural poses and expressions." + ], + "image_footnote": [], + "bbox": [ + 125, + 273, + 846, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "both VoxCeleb1 and VoxCeleb2, surpassing other state-of-the-art by a large margin. Also, we randomly sample 5,000 poses for each method and employ t-SNE [50] for visual-", + "bbox": [ + 75, + 853, + 470, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ization in Fig. 4, which demonstrates that the distribution we generate is the most comprehensive to GT. More details and examples are shown in the supplementary. In summary,", + "bbox": [ + 498, + 853, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7651", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c3988fbacbbd83ba571a12b2d7208f0ebe079530e84d4a3e0418476ddce53860.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Human evaluation on generated samples on VoxCeleb2. * means that PC-AVS was evaluated on cross-reenactment scenario." + ], + "table_body": "
MethodSync AccuracyNaturalnessRichness
GT4.614.414.18
MakeItTalk [66]1.972.352.10
PC-AVS [65]*3.132.682.68
Audio2head [54]2.562.352.34
Ours + AR3.462.533.37
Ours + Diff.4.083.823.68
", + "bbox": [ + 84, + 88, + 459, + 213 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our method can generate rich facial motions with excellent realism (i.e., a lower distribution distance) and naturalness.", + "bbox": [ + 76, + 263, + 468, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 304, + 285, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Two audio-driven video sequences of our method verses other baselines are shown in Fig. 5. It's clearly shown that our method is capable of producing much more natural-looking and diverse facial motions than other baselines, including pose, expression, blink and gaze, while maintaining accurate audio-lip synchronization compared to the GT.", + "bbox": [ + 76, + 328, + 468, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study We invited 20 subjects for human evaluations, focusing on three aspects: 1) the accuracy of audio-lip synchronization; 2) the naturalness of non-lip motions and coherency between non-lip motions and audios; 3) the richness of non-lip motions. 50 videos are generated from audio and identity in the test set using the following methods: our diffusion prior, our auto-regressive prior, MakeItTalk [66], Audio2head [54] and PC-AVS [65] (driven by poses from another video). After shuffling the generated videos, each annotator rates from 1 (bad) to 5 (good) according to Mean Opinion Scores (MOS) rating protocol. Table 3 tells that our model with diffusion prior archives the best synchronization, naturalness and richness among all the models. Our model with auto-regressive prior archives relatively good richness compared to the remaining models, we attribute it to the exposure bias problem [4] brought by auto-regressive generation, which leads to high richness but low naturalness. Note that PC-AVS shows incompatible result in Table 3 as compared to Table 1 on $\\mathbf{S}_c$ , indicating that choosing pose driving signals from a random video may lead to synthesis results with bad audio-lip synchronization.", + "bbox": [ + 76, + 428, + 468, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It's interesting that our proposed SND scores reflect the naturalness of user study to some degree. Nevertheless, we leave it to future work for rigorous study of their correlation through more human evaluations.", + "bbox": [ + 76, + 744, + 468, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation", + "text_level": 1, + "bbox": [ + 76, + 816, + 181, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4 discusses the performance of lip & non-lip disentanglement. While tested on lip only metrics without $L_{ol}$ , SyncNet scores get lower, which indicates that G tends to generate mouth movements with information extracted", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4444654d62bae8cbe38faec3bc3143fd6bee5058f246ba58d3f1b06f30746917.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Evaluation of disentanglement between lip and non-lip on VoxCeleb2. $\\mathbf{B}_d$ and $\\mathrm{G}_d$ denote L2 distances of 2D landmark [58] of eyes and iris, respectively, while $\\mathrm{E}_d$ and $\\mathrm{P}_d$ denote L2 distances of 3DMM coefficients of poses and expressions accordingly. Note that $\\mathbf{f}_{nl}^v$ is fixed while tested on lip only metrics, and $\\mathbf{f}_l^a$ is fixed while tested on non-lip only metrics for fair comparisons." + ], + "table_body": "
MethodLip OnlyNon-lip Only
Sc↑Nc↓Bd↓Gd↓Ed↓Pd↓
GT7.35-----
w/o Lol2.840.610.00740.19600.05680.0021
w/o MB7.690.050.01120.03150.08750.0293
Ours7.480.030.00670.01630.05580.0019
", + "bbox": [ + 500, + 88, + 893, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from visual modality instead of audio. Thus, lip & non-lip space are not fully disentangled. Without MB, the performance of driving non-lip motions gets worse. Because it is hard to use a large batch to compute correlation, resulting in a not well decoupled non-lip space.", + "bbox": [ + 496, + 296, + 890, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted ablation study for $\\mathrm{P}_{a2nl}$ from three aspects: 1) auto-regressive prior versus diffusion prior; 2) with/without velocity loss $L_{vel}$ ; 3) training with/without editing mechanism. Diffusion prior shows significant improvements on naturalness scores including $\\mathrm{FID}_{fm}$ , $\\mathrm{FID}_{\\Delta fm}$ and SND, as compared to autoregressive prior shown in Table 2, with slightly better synchronization ability in Table 1. Fig. 6 shows that without $L_{vel}$ , the prediction of non-autoregressive diffusion prior model $\\mathrm{P}_{a2nl}$ becomes unstable, which leads to a larger variance and worse SND scores in Table 2. While training without the editing mechanism mentioned in Sec. 3.2, we applied editing only during sampling as in MDM [46], and observed that it jitters between adjacent frames as shown in Fig. 7, resulting in a poorer SND score in Table 2. Note that our model trained with editing mechanism generates smoother results.", + "bbox": [ + 496, + 371, + 892, + 613 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2ca78a9ffd31ae9d3d234671fd3a8dcb4774cc72a0df01d974879acd3986791.jpg", + "image_caption": [ + "Figure 6. Ablation study of our diffusion prior $\\mathrm{P}_{a2nl}$ w or w/o $L_{vel}$ . Each row shows five uniformly sampled frames." + ], + "image_footnote": [], + "bbox": [ + 521, + 618, + 875, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 799, + 625, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel talking head generation method based on a diffusion prior model, which can generate diverse and natural-looking talking head videos with only audio and an identity image as inputs. Such \"audio-driving is all you need\" setting is very friendly", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7652", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/732e1e805223738e2c730ac7ebaeea552c1544f45af3394be6d22a32e1367d87.jpg", + "image_caption": [ + "Figure 7. Ablation study of our diffusion prior $\\mathbf{P}_{a2nl}$ trained w or w/o editing. Each row shows five adjacent frame." + ], + "image_footnote": [], + "bbox": [ + 89, + 90, + 452, + 219 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "for reenactment and dubbing applications. Comprehensive evaluations including our newly proposed metrics validated the effectiveness of our system.", + "bbox": [ + 76, + 275, + 468, + 321 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations Our method mainly focuses on non-lip motion generation. We leave the improvement of rendering quality to future works.", + "bbox": [ + 76, + 328, + 468, + 372 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 386, + 171, + 400 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ahmed A Abdelrahman, Thorsten Hempel, Aly Khalifa, and Ayoub Al-Hamadi. L2cs-net: Fine-grained gaze estimation in unconstrained environments. arXiv preprint arXiv:2203.03339, 2022. 4", + "[2] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelović, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-Supervised MultiModal Versatile Networks. In NeurIPS, 2020. 3", + "[3] Mohammed M Alghamdi, He Wang, Andrew J Bulpitt, and David C Hogg. Talking head from speech audio using a pretrained image generator. arXiv preprint arXiv:2209.04252, 2022. 2, 3", + "[4] Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems, 28, 2015. 8", + "[5] Stella Bounareli, Vasileios Argyriou, and Georgios Tzimiropoulos. Finding directions in gan's latent space for neural face reenactment. arXiv preprint arXiv:2202.00046, 2022.3", + "[6] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 3", + "[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 5", + "[8] Soo-Whan Chung, Joon Son Chung, and Hong-Goo Kang. Perfect match: Improved cross-modal embeddings for audiovisual synchronisation. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3965-3969. IEEE, 2019. 2", + "[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with" + ], + "bbox": [ + 86, + 411, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 6", + "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3", + "[11] Aviv Gabbay, Asaph Shamir, and Shmuel Peleg. Visual speech enhancement. arXiv preprint arXiv:1711.08789, 2017.2", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 4", + "[13] Yudong Guo, Keyu Chen, Sen Liang, Yongjin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3", + "[16] Fangzhou Hong, Liang Pan, Zhongang Cai, and Ziwei Liu. Versatile multi-modal pre-training for human-centric perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16156-16166, 2022. 3", + "[17] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3397-3406, 2022. 3", + "[18] Gee-Sern Hsu, Chun-Hung Tsai, and Hung-Yi Wu. Dual-generator face reenactment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 642-650, 2022. 3", + "[19] Po-Hsiang Huang, Fu-En Yang, and Yu-Chiang Frank Wang. Learning identity-invariant motion representations for crossid face reenactment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7084–7092, 2020. 3", + "[20] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. arXiv preprint arXiv:2205.15278, 2022. 5, 6", + "[21] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3", + "[22] Venkatesh S Kadandale, Juan F Montesinos, and Gloria Haro. Vocalist: An audio-visual synchronisation model for lips and voices. arXiv preprint arXiv:2204.02090, 2022. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7653", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 4", + "[24] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Nießner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics, 37(4):163:1-14, Aug. 2018. 3", + "[25] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 3", + "[26] Jiyoung Lee, Soo-Whan Chung, Sunok Kim, Hong-Goo Kang, and Kwanghoon Sohn. Looking into your speech: Learning cross-modal affinity for audio-visual speech separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1336–1345, 2021. 2", + "[27] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3387-3396, June 2022. 2, 3", + "[28] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 5", + "[29] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2, 3", + "[30] Moustafa Meshry, Saksham Suri, Larry S Davis, and Abhnav Shrivastava. Learned spatial representations for few-shot talking-head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13829-13838, 2021. 3", + "[31] Dongchan Min, Minyoung Song, and Sung Ju Hwang. Stylistalker: One-shot style-based audio-driven talking head video generation. preprint arXiv:2208.10922, 2022. 3", + "[32] Gaurav Mittal and Baoyuan Wang. Animating face using disentangled audio representations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), March 2020. 2", + "[33] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. Voxceleb: a large-scale speaker identification dataset. arXiv preprint arXiv:1706.08612, 2017. 5", + "[34] Arsha Nagrani, Joon son Chung, Samuel Albanie, and Andrew Zisserman. Disentangled speech embeddings using cross-modal self-supervision, 2020. 2", + "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3, 5" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Andrew Zisserman Olivia Wiles, A. Sophia Koepke. X2face: A network for controlling face generation by using images, audio, and pose codes. In ECCV 2018, 2018. 3", + "[37] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In 36th AAAI Conference on Artificial Intelligence (AAAI 22). Association for the Advancement of Artificial Intelligence, 2022. 3", + "[38] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2, 3, 6", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3", + "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3, 4", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3", + "[42] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. arXiv preprint arXiv:2207.11770, 2022. 2", + "[43] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3", + "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3", + "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3,4", + "[46] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 3, 4, 5, 8", + "[47] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. ECCV 2020, 2020. 3", + "[48] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 3" + ], + "bbox": [ + 501, + 92, + 892, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7654", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. arXiv preprint arXiv:1906.05849, 2019. 2", + "[50] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7", + "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5, 6", + "[52] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. CoRR, abs/1906.06337, 2019. 3", + "[53] Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang. Progressive disentangled representation learning for fine-grained controllable talking head synthesis. arXiv preprint arXiv:2211.14506, 2022. 6", + "[54] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2, 6, 7, 8", + "[55] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2531–2539, 2022. 3", + "[56] Chao Xu, Jiangning Zhang, Yue Han, Guanzhong Tian, Xianfang Zeng, Ying Tai, Yabiao Wang, Chengjie Wang, and Yong Liu. Designing one unified framework for high-fidelity face reenactment and swapping. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 54-71. Springer, 2022. 3", + "[57] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022. 6", + "[58] Baosheng Yu and Dacheng Tao. Heatmap regression via randomized rounding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 8", + "[59] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 3", + "[60] Xianfang Zeng, Yusu Pan, Mengmeng Wang, Jiangning Zhang, and Yong Liu. Realistic face reenactment via self-supervised disentangling of identity and pose. pages 12757-12764. AAAI Press, 2020. 3", + "[61] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3867-3876, 2021. 2, 3", + "[62] Jiangning Zhang, Liang Liu, Zhucun Xue, and Yong Liu. Apb2face: Audio-guided face reenactment with auxiliary pose and blink signals. 2020. 3", + "[63] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661-3670, 2021. 3", + "[64] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9299–9306, 2019. 2", + "[65] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2, 3, 4, 5, 6, 8", + "[66] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 2, 3, 6, 7, 8" + ], + "bbox": [ + 501, + 90, + 893, + 345 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7655", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/7d213d41-b6ba-4fcf-a556-dbfb8dc18cb6_model.json b/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/7d213d41-b6ba-4fcf-a556-dbfb8dc18cb6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e61e71c41b15950ad6525f2a761a6cc7a477a0b9 --- /dev/null +++ b/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/7d213d41-b6ba-4fcf-a556-dbfb8dc18cb6_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.873, + 0.153 + ], + "angle": 0, + "content": "Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.179, + 0.751, + 0.234 + ], + "angle": 0, + "content": "Zhentao Yu\\(^{1*}\\) Zixin Yin\\(^{1,2*†}\\) Deyu Zhou\\(^{1,3*†}\\) Duomin Wang\\(^{1}\\) Finn Wong\\(^{1}\\) Baoyuan Wang\\(^{1‡}\\) Xiaobing.AI." + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.234, + 0.761, + 0.27 + ], + "angle": 0, + "content": "\\(^{2}\\)The Hong Kong University of Science and Technology. \\(^{3}\\)The Hong Kong University of Science and Technology (Guangzhou)." + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.273, + 0.737, + 0.305 + ], + "angle": 0, + "content": "(yuzhentao, wangduomin, wangwenlan, wangbaoyuan)@xiaobing.ai zyinaf@connect.ust.hk, dzhou861@connect.hkust-gz.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.339, + 0.891, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.57, + 0.894, + 0.613 + ], + "angle": 0, + "content": "Figure 1. Given only an audio source and an arbitrary identity image, our system can generate a video with natural-looking and diverse facial motions (pose, expression, blink & gaze), while maintaining accurate audio-lip synchronization. Here we show randomly sampled sequences from our diffusion prior for two identities, note that the lip-irrelevant facial motion varies but the lip is still in-sync." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.631, + 0.314, + 0.646 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.662, + 0.471, + 0.844 + ], + "angle": 0, + "content": "We introduce a novel framework for one-shot audio-driven talking head generation. Unlike prior works that require additional driving sources for controlled synthesis in a deterministic manner, we instead sample all holistic lip-irrelevant facial motions (i.e. pose, expression, blink, gaze, etc.) to semantically match the input audio while still maintaining both the photo-realism of audio-lip synchronization and overall naturalness. This is achieved by our newly proposed audio-to-visual diffusion prior trained on top of the mapping between audio and non-lip representations. Thanks to the probabilistic nature of the diffusion prior, one big advantage of our framework is it can synthe" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.632, + 0.892, + 0.768 + ], + "angle": 0, + "content": "size diverse facial motion sequences given the same audio clip, which is quite user-friendly for many real applications. Through comprehensive evaluations of public benchmarks, we conclude that (1) our diffusion prior outperforms autoregressive prior significantly on all the concerned metrics; (2) our overall system is competitive with prior works in terms of audio-lip synchronization but can effectively sample rich and natural-looking lip-irrelevant facial motions while still semantically harmonized with the audio input." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.815, + 0.633, + 0.831 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.842, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Audio-driven face reenactment and talking head generation have received raising attention due to the broad killer applications in movie production, gaming, virtual digital avatars, and potentially more while we move toward the era" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.852, + 0.38, + 0.865 + ], + "angle": 0, + "content": "*These authors have contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.865, + 0.47, + 0.889 + ], + "angle": 0, + "content": "This work was done when Zixin Yin and Deyu Zhou were interns at XiaoBing.AI." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.889, + 0.222, + 0.902 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.852, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7645" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.575 + ], + "angle": 0, + "content": "of the metaverse. The past literature tries to advance this area from various perspectives. One line of research focuses on improving the generation quality originating from regular GAN-based methods [64, 65] to pre-trained StyleGAN [3] and to the most recent Neural Radiance Field (NeRF) based methods [13, 42]. Orthogonal to direction, another line of works emphasizes the importance of disentangled representation from the audio [32, 21, 66] for controlled generations. i.e., [21] can predict the emotion from the audio input while [66] can decouple the speech signal into speaker identity and the phonetic content. A closely related set of works targeting more granular controlled talking head generation by providing additional input signals. For example, PC-AVS [65] requires a separate video to provide the pose signal in addition to the audio clip but does not support the control of expression and blinking. To remedy this, GC-AVT [27] requires inputting additional driving video for expression in addition to pose and audio driving clip. The technical challenge behind those approaches is how to faithfully transfer the desired driving signal (i.e., pose, expression, audio) into the results without affecting each other through intrinsic disentanglement. Although encouraging results have been reported, we argue such a setting is not practical for broader applications. It is quite challenging, or at least labor-intensive, for novice users to find the \"best\" individual driving sources for each controlled dimension to make the final talking head video overall look not only lifelike but also coherent from semantic and emotional perspectives. Therefore, it is both more practical and generalizable if the setting only requires an audio signal as the driving source and expects a data-driven model to sample reasonable other facial motions irrelevant to lips." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.58, + 0.473, + 0.852 + ], + "angle": 0, + "content": "By no means, we are the first to advocate an audio-only driving setup for talking head generation. Audio2head [54] predicts the poses from audio input with an LSTM network. [29] employs an autoregressive model by assuming the poses are jointly determined by the audio and past head motions along the sequence, although the results are encouraging, the model can't generalize due to their person-specific training strategy. FACIAL [61] could infer the blink, but requires a reference video input rather than a one-shot reference image. There are other related works along this direction, but they either only infer one facial attribute (i.e., emotion in EVP[21], pose from [54]) using ad-hoc methods, or they simply do not support inferring other facial motions [38], or treat it as a block-box mapping model [3] without explicitly respecting the one-to-many mapping nature between audio and other visual facial attributes (pose, expression, blink). A more principled solution is desired to consolidate this line of work." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce a novel framework that can holistically infer all the non-lip-related facial attributes from the audio input while maintaining accurate synchronization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.365 + ], + "angle": 0, + "content": "between audio and the corresponding visual lip motions. This is achieved by two important learning steps in our pipeline. (1) A pre-trained identity-irrelevant facial motion representation can help decouple the lip and non-lip representations. To promote this learning, we employ a novel orthogonal loss on top of the modified facial reenactment framework [6]. The disentanglement enables to generate one-to-one mapping with lip representations to ensure the synchronization, and generate richer non-lip representations with a one-to-many mapping. (2) A novel audio-to-visual diffusion prior model is introduced to address the probabilistic sampling from audio representations to the abovelearned non-lip representation. This prior is expected to solve the one-to-many mapping and provide diverse results during the inference stage. The entire pipeline can be easily built up on top of the existing framework, such as PC-AVS[65] without heavily retraining every component. To sum up, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.374, + 0.892, + 0.48 + ], + "angle": 0, + "content": "- To our best knowledge, we are the first to holistically predict non-lip facial motions based on audio input only, providing good usability for reenactment or dubbing applications without extra driving video sources. Our method addresses the intrinsic one-to-many challenge in a probabilistic way, allowing diverse and realistic facial motion generation under the same audio input, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.489, + 0.892, + 0.597 + ], + "angle": 0, + "content": "- We leverage the pre-trained visual identity-irrelevant facial motion representations. And further, learn disentangled lip-related and lip-irrelevant representations through a novel orthogonal loss on top of PC-AVS[65]. A powerful diffusion prior model is then introduced to effectively infer all lip-irrelevant facial motions for a given audio segment in the representation space." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.606, + 0.892, + 0.695 + ], + "angle": 0, + "content": "- We systematically evaluate the naturalness and diversity of the results with new metrics, which paves a way for future studies. Meanwhile, results show that our method can produce natural-looking head poses and facial motions without hurting the audio-lip synchronization. Our model will be released." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.374, + 0.892, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.651, + 0.726 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Audio-Visual Cross-modal Learning Cross-modal representation learning is a long-standing research topic, ranging from speech enhancement [11], speech source separation [26] to synchronization [8, 22] and other speech disentanglement [34, 21, 32]. Among them, EVP [21] used cross-modal supervision to disentangle speech content and emotion from the audio signal with landmark as the intermediate representation. Recently, CMC [49] discussed how multiview \"modality\" can be jointly unitized to boost intrinsic representation through contrastive learning, rather than predictive (or reconstruction) learning, it also demonstrated" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "7646" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.289 + ], + "angle": 0, + "content": "that the more views, the better. Concurrently, MMV [2] introduced different modality embedding graphs for effective cross-modal representations, again through contrastive learning. More recently, HCMoCo [16] extended similar ideas with a hierarchical strategy to learn different levels of representations for human-centric perception tasks. Although impressive results were reported, it is still unclear how it performs on face analysis tasks, not mention to on synthesis tasks. Our work shares a similar spirit but a different purpose, where we first pre-train a non-identity visual representation which then helps learn a lip and non-lip space, the latter further serves as the upper-bound learning target for subsequent audio-to-visual diffusion prior." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.473, + 0.68 + ], + "angle": 0, + "content": "Face Reenactment & Talking Head Generation Face Reenactment is designed to transfer part or full facial motion from a driving source to the target video with good IDpreserved appearance and background. It can be further divided into two categories depending on whether the driving source is from video [17, 18, 5, 48, 24, 36, 59, 19, 60, 6, 56] or audio [52, 47, 62, 21, 55, 37, 63, 30]. Among them, audio-driven face reenactment generally aims to edit the mouth regions of the target video in order to match the input audio while leaving other facial attributes mostly unchanged, i.e., pose. EVP [21] tries to infer the non-rigid facial expression in addition to lip motion from the audio input. A closely related line of work is the audio-driven talking-head generations [38, 66, 61, 3] where only one target reference face is given, hence, other face attributes including pose, expression, blink and gaze have to be either explicitly given [65, 27] or partly inferred through statistical methods [66, 31, 61, 29, 3, 47]. Specifically, Lu et al. [29] employed an auto-regressive model while Min et al. [31] leveraged normalized flow prior to predicting a natural-looking pose sequence from the input audio, both showed encouraging results. Compared with them, we aim to infer more diverse facial motions including pose, expression, and even blink and gaze in a holistic manner through an audio-to-visual diffusion prior model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Diffusion Generative Models The diffusion model [44, 15, 45], which is a likelihood-based model consisting of cascading denoising autoencoders, has recently shown great success in numerous generative tasks with different modalities including image [10, 35, 40, 41], audio [25], video [43], and motion [46]. To name a few, DDPM [15] explored the diffusion model for unconditional image generation. GLIDE [35] introduced text-conditional diffusion model and showed that classifier free guidance has better performance than CLIP [39] guidance. DALLE-2 [40] modified GLIDE to generate semantically consistent images conditioned on a CLIP image embedding, and proposed a diffusion prior that produces the image embedding given a text caption. MDM [46] utilized a classifier-free diffusion-based" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "generative model for text-to-motion and action-to-motion tasks, allowing motion completion and editing as well." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.134, + 0.593, + 0.15 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.159, + 0.894, + 0.224 + ], + "angle": 0, + "content": "Given \\(L\\) segments of audio \\(A_{1:L} = (A_1, A_2, \\dots, A_L)\\) and a reference image \\(I_{ref}\\) as inputs, our model M synthesizes video frames \\(\\hat{X}_{1:L} = (\\hat{X}_1, \\hat{X}_2, \\dots, \\hat{X}_L)\\) having the same identity as \\(I_{ref}\\) and lip motion synchronized to \\(A_{1:L}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.231, + 0.892, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {1: L} = \\mathbf {M} \\left(A _ {1: L}, I _ {r e f}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.257, + 0.892, + 0.289 + ], + "angle": 0, + "content": "An overview of our proposed framework is shown in Fig. 2, which consists of three major components:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.295, + 0.892, + 0.372 + ], + "angle": 0, + "content": "- Lip & Non-lip Disentanglement Given a pre-trained identity- and appearance-irrelevant facial motion encoder \\( \\mathrm{E}_v \\), we first leverage it to learn two complementary features, \\( \\mathbf{f}_l^a \\) for lip, and \\( \\mathbf{f}_{nl}^v \\) for non-lip features including pose, expression, blink and gaze." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.38, + 0.892, + 0.456 + ], + "angle": 0, + "content": "- Audio to Non-lip Diffusion Prior A diffusion prior network \\( \\mathrm{P}_{a2nl} \\) models the one-to-many mapping from audio feature \\( \\mathbf{f}^a \\) to the \"hallucinated\" non-lip feature \\( \\mathbf{f}_{nl}^a \\), which is trained to be close to its visual counterpart \\( \\mathbf{f}_{nl}^v \\), allowing audio-only facial driving at inference time." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.464, + 0.892, + 0.54 + ], + "angle": 0, + "content": "- Audio-based Talking Head Generation Given an identity feature \\(\\mathbf{f}_{id}\\), a lip feature \\(\\mathbf{f}_l^a\\), and a non-lip feature \\(\\mathbf{f}_{nl}^a\\) generated with the diffusion prior, we concatenate them together and feed into a GAN similar to LPD [6] or PC-AVS [65] to output the final reenacted video." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.295, + 0.892, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.546, + 0.78, + 0.563 + ], + "angle": 0, + "content": "3.1. Lip & Non-lip Disentanglement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.69 + ], + "angle": 0, + "content": "The identity encoder \\( \\mathrm{E}_{id} \\) is designed to capture identity and appearance information, while \\( \\mathrm{E}_v \\) is instead to remove both while encoding all facial motions. Both encoders are pre-trained in a similar setting as LPD [6] and PC-AVS [65]. After pretraining, we conduct lip & non-lip disentanglement consisting of audio-visual contrastive learning for lip feature learning and decorrelation for non-lip feature learning, as well as face reconstruction, as described below:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Audio-Visual Pretraining for Audio Lip Space An audio encoder \\( \\mathrm{E}_a \\) is trained to provide accurate audio feature \\( \\mathbf{f}^a = \\mathrm{E}_a(A_{1:N}) \\) through contrastive learning [39] with \\( \\mathbf{f}_{cl}^v = \\mathrm{MLP}_{cl}(\\mathrm{E}_v(X_{1:N})) \\), where \\( N \\) is the number of samples and \\( X_{1:N} \\) are \\( N \\) frames in the same video corresponding to audio \\( A_{1:N} \\). Then \\( \\mathbf{f}^a \\) is further compressed to filter out lip-relevant information and get audio lip feature \\( \\mathbf{f}_l^a \\) through \\( \\mathrm{MLP}_{a2l} \\) since \\( \\mathbf{f}^a \\) mainly contains lip related information, as shown in Sec. 4.3. The contrastive loss is defined as \\( L_{con}(m,n) = -\\frac{1}{N}\\sum_{i=1}^{N}\\log\\frac{\\exp(m_i/\\|m_i\\|_2 \\cdot n_i/\\|n_i\\|_2)}{\\sum_{j=1}^{N}\\exp(m_i/\\|m_i\\|_2 \\cdot n_j/\\|n_j\\|_2)} \\), resulting in a total contrastive loss between \\( \\mathbf{f}^a \\) and \\( \\mathbf{f}_{cl}^v \\)," + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.876, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\nL _ {c l} = \\frac {1}{2} \\left[ L _ {\\text {c o n}} \\left(\\mathbf {f} ^ {a}, \\mathbf {f} _ {c l} ^ {v}\\right) + L _ {\\text {c o n}} \\left(\\mathbf {f} _ {c l} ^ {v}, \\mathbf {f} ^ {a}\\right) \\right]. \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7647" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.095, + 0.889, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.896, + 0.348 + ], + "angle": 0, + "content": "Figure 2. The overall pipeline of our proposed framework. The dotted lines represent the loss functions that are used only in training, i.e., \\( L_{cl} \\), \\( L_{ol} \\) and \\( L_{a2nl} \\). \\( L_{G} \\) is the training loss for visual non-lip space. Note that there is a switch in the figure, which means the different forward processes in training and inference: in training, the concatenated feature \\( \\mathbf{f}_{cat} = cat(\\mathbf{f}_{id},\\mathbf{f}_{nl}^{v},\\mathbf{f}_{l}^{a}) \\); at inference stage, \\( \\mathbf{f}_{cat} = cat(\\mathbf{f}_{id},\\mathbf{f}_{nl}^{a},\\mathbf{f}_{l}^{a}) \\). Thus, the visual motion part is not needed anymore at the inference stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.473, + 0.56 + ], + "angle": 0, + "content": "Reconstruction Learning for Visual Non-lip Space To learn a good non-lip space, we propose an orthogonal loss to penalize the correlation between \\(\\mathbf{f}_{ol}^{v}\\) and \\(\\mathbf{f}^{\\alpha}\\), which uses the well-learned audio space to disentangle lip-related motions. The orthogonal loss is used together with reconstruction since there should also be a completeness constraint to avoid mode collapse of lip-irrelevant features. In practice, we maintain two memory banks (MB) for storing \\(\\mathbf{f}_{ol}^{v}\\) and \\(\\mathbf{f}^{\\alpha}\\) in previous \\(K - 1\\) steps, in order to have more samples than \\(N\\). We denote the feature dimension of \\(\\mathbf{f}^{\\alpha}\\) as \\(d_{a}\\), and \\(\\mathbf{f}_{ol}^{v}\\) as \\(d_{v}\\). The orthogonal loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.565, + 0.47, + 0.607 + ], + "angle": 0, + "content": "\\[\nL _ {o l} = \\frac {1}{d _ {v}} \\sum_ {i = 1} ^ {d _ {v}} \\sum_ {j = 1} ^ {d _ {a}} P _ {c o r} \\left(\\mathbf {f} _ {o l} ^ {v}, \\mathbf {f} ^ {a}\\right) _ {(i, j)} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.72 + ], + "angle": 0, + "content": "where \\(P_{cor}(\\mathbf{f}_{ol}^v,\\mathbf{f}^a)\\in d_a\\times d_v\\) computes the Pearson correlation coefficient between \\(\\mathbf{f}_{ol}^{v}\\) and \\(\\mathbf{f}^a\\). Since the correlation between lip motions and audio features is strong, \\(L_{ol}\\) (which encourages the linear uncorrelation between \\(\\mathbf{f}_{ol}^{v}\\) and \\(\\mathbf{f}^a\\)) is effective to disentangle lip-related motions. Then, \\(\\mathrm{MLP}_{nl}\\) further projects \\(\\mathbf{f}_{ol}^{v}\\) into a smaller space to remove undesired content and get non-lip feature \\(\\mathbf{f}_{nl}^{v}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We follow the reconstruction loss \\( L_{\\mathrm{GAN}} \\), \\( L_{L1} \\) and \\( L_{\\mathrm{VGG}} \\) used in [65]. The generator \\( G \\) receives the concatenation \\( \\mathbf{f}_{cat} = cat(\\mathbf{f}_{id}, \\mathbf{f}_{nl}^{v}, \\mathbf{f}_{l}^{a}) \\) as inputs and generates final images \\( I_{gen} \\) with modulated convolution. The discriminator \\( D \\) is trained jointly by generative adversarial learning [23, 12]. Note that we omit the batch-mean operation here for convenience, but in practice, all losses used here are calculated with the average in a batch. Different from [65], we propose an additional gaze loss \\( L_{gaze} \\) by utilizing a pre-trained gaze encoder [1], which calculates the \\( L1 \\) distance between the gaze features of generated images and ground-truth (GT) images \\( I \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.386, + 0.893, + 0.404 + ], + "angle": 0, + "content": "\\[\nL _ {g a z e} = \\left\\| \\Phi (I) - \\Phi \\left(I _ {g e n}\\right) \\right\\| _ {1}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.406, + 0.719, + 0.421 + ], + "angle": 0, + "content": "The total loss \\(L_{G}\\) is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.427, + 0.893, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {G} = \\lambda_ {o l} \\cdot L _ {o l} + \\lambda_ {g a z e} \\cdot L _ {g a z e} + \\lambda_ {L 1} \\cdot L _ {L 1} \\tag {5} \\\\ + \\lambda_ {\\text {G A N}} \\cdot L _ {\\text {G A N}} + \\lambda_ {\\text {V G G}} \\cdot L _ {\\text {V G G}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.467, + 0.719, + 0.482 + ], + "angle": 0, + "content": "where \\(\\lambda\\) s are weights for losses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.489, + 0.763, + 0.505 + ], + "angle": 0, + "content": "3.2. Audio2nonlip Diffusion Prior" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.893, + 0.634 + ], + "angle": 0, + "content": "Beside a strong relationship with lip motions, audio features have a complex non-linear relationship with non-lip motions and it is a one-to-many mapping problem. This is due to the fact that there are many reasonable facial motions that can be corresponded to the same audio input. As inspired by DALLE-2 [40] and MDM [46], we design an audio2nonlip diffusion prior network \\( \\mathrm{P}_{a2nl} \\) to address this problem. Our proposed network is depicted in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.894, + 0.684 + ], + "angle": 0, + "content": "Diffusion Model Diffusion model is designed based on the stochastic diffusion process. The forward diffusion process is defined below:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.689, + 0.892, + 0.709 + ], + "angle": 0, + "content": "\\[\nq \\left(n _ {1: L} ^ {t} \\mid n _ {1: L} ^ {t - 1}\\right) = \\mathcal {N} \\left(n _ {1: L} ^ {t - 1}; \\sqrt {\\alpha^ {t}} n _ {1: L} ^ {t - 1}, (1 - \\alpha^ {t}) I\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.894, + 0.774 + ], + "angle": 0, + "content": "In our context, \\(n_{1:L} = (n_1,n_2,\\dots,n_L)\\) is a sequence of non-lip feature \\(\\mathbf{f}_{nl}^v\\). \\(\\alpha^t\\) is a hyper-parameter and \\(t\\sim [1,T]\\) is the time step of the diffusion process. Eq. 6 can be approximated to \\(n_{1:L}^{T}\\sim \\mathcal{N}(0,1)\\) if \\(\\alpha^t\\) is small enough." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.774, + 0.894, + 0.865 + ], + "angle": 0, + "content": "During the reversed diffusion process, \\(\\mathrm{P}_{a2nl}\\) models the audio2nonlip distribution as \\(p(n_{1:L}^{0}|a_{1:L})\\), where \\(a_{1:L} = (a_1,a_2,\\dots,a_L)\\) is a sequence of audio feature \\(\\mathbf{f}^a\\). Instead of predicting the noise as formulated by vanilla DDPM [45], \\(\\mathrm{P}_{a2nl}\\) learns to predict the initial signal itself with the simplified objective function [45] described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.868, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\nL _ {s i m p l e} = \\mathbb {E} _ {n ^ {0} \\sim q (n ^ {0} | \\mathbf {f} ^ {\\alpha}), t \\sim [ 1, T ]} \\left[ \\left\\| n ^ {0} - \\mathrm {P} _ {a 2 n l} \\left(n ^ {t}, t, \\mathbf {f} ^ {\\alpha}\\right) \\right\\| _ {2} \\right], \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7648" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.445, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.295, + 0.428, + 0.31 + ], + "angle": 0, + "content": "Figure 3. Our audio2nonlip diffusion prior network \\(\\mathrm{P}_{a2nl}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.336, + 0.418, + 0.353 + ], + "angle": 0, + "content": "where we use \\(n^0\\) to represent \\(n_{1:L}^{0}\\) for convenience." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.468, + 0.46 + ], + "angle": 0, + "content": "We implement \\( \\mathrm{P}_{a2nl} \\) with a transformer encoder [51] architecture. As shown in Fig. 3, an audio feature \\( a_{1:L} \\) is added to the time embedding \\( \\mathrm{Emb}_{\\text{time}} \\). Then, it's concatenated with noisy non-lip feature \\( n_{1:L}^{t} \\) and added with positional embeddings \\( \\mathrm{Emb}_{pos} \\) [51]. The model encodes the embeddings with bidirectional self-attention and feedforward layers then output denoised non-lip feature \\( n_{1:L}^{0} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.474, + 0.469, + 0.535 + ], + "angle": 0, + "content": "Velocity Loss Note that our designed \\(\\mathrm{P}_{a2nl}\\) denoises \\(n_{1:L}^{t}\\) in a non-autoregressive way. To encourage the naturalness and coherence of generated non-lip motion, we borrow the velocity loss from MDM [46] defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.549, + 0.469, + 0.589 + ], + "angle": 0, + "content": "\\[\nL _ {v e l} = \\frac {1}{L - 1} \\sum_ {i = 2} ^ {L} \\left\\| \\left(\\hat {n} _ {i} ^ {0} - \\hat {n} _ {i - 1} ^ {0}\\right) - \\left(n _ {i} ^ {0} - n _ {i - 1} ^ {0}\\right) \\right\\| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.469, + 0.68 + ], + "angle": 0, + "content": "where \\( n_i^0 \\) is the GT non-lip feature and \\( \\hat{n}_i^0 \\) is the predicted denoised non-lip feature at the \\( i \\)-th position respectively. The intuition here is that the difference between adjacent non-lip features should be close to the difference in the GT. The total loss of this stage is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.696, + 0.469, + 0.712 + ], + "angle": 0, + "content": "\\[\nL _ {a 2 n l} = L _ {\\text {s i m p l e}} + L _ {v e l}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.741, + 0.469, + 0.833 + ], + "angle": 0, + "content": "Classifier-free Guidance We use classifier-free guidance [35] for conditioned diffusion generation. In training time, we randomly set the condition to \\(\\emptyset\\) for \\(10\\%\\) of the samples so that \\(\\mathrm{P}_{a2nl}(n^t,t,\\emptyset)\\) approximates \\(n^0\\). At inference stage, the output of the \\(P_{a2nl}\\) is extrapolated further in the direction of \\(\\mathrm{P}_{a2nl}(n^t,t,\\mathbf{f}^a)\\) and away from \\(\\mathrm{P}_{a2nl}(n^t,t,\\emptyset)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.846, + 0.468, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P} _ {a 2 n l} \\left(n ^ {t}, t, \\mathbf {f} ^ {a}\\right) = s \\cdot \\mathrm {P} _ {a 2 n l} \\left(n ^ {t}, t, \\mathbf {f} ^ {a}\\right) + (1 - s) \\cdot \\mathrm {P} _ {a 2 n l} \\left(n ^ {t}, t, \\emptyset\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.879, + 0.469, + 0.909 + ], + "angle": 0, + "content": "where \\( s \\) is a scaling parameter while increasing it improves sample quality at the cost of diversity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Sequential Mask Editing To deliver a smooth non-lip facial sequence with arbitrary length, we design a mechanism to ensure continuity between generated segments, and provide an editing method to do so. Specifically, we randomly mask \\(90\\%\\) of the non-lip tokens \\(n^0\\) and concatenate them with noisy non-lip tokens \\(n^t\\) as the input of \\(\\mathrm{P}_{a2nl}\\) at training stage. Practically, we found it easy for \\(\\mathrm{P}_{a2nl}\\) to quickly learn how to fill in the masked region using hints from the non-lip features in the unmasked regions, generating continuous feature prediction without any extra design in training loss. As a result, it is possible to use \\(\\mathrm{P}_{a2nl}\\) for audio-guided non-lip motion generation as well as non-lip motion editing at inference time, providing non-lip facial motion with good naturalism and diversity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.314, + 0.831, + 0.33 + ], + "angle": 0, + "content": "3.3. Audio-based Talking Head Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.444 + ], + "angle": 0, + "content": "The main difference at inference time is that non-lip visual feature \\(\\mathbf{f}_{nl}^{v}\\) is replaced with \\(\\mathbf{f}_{nl}^{a}\\) generated by \\(\\mathrm{P}_{a2nl}\\) through reversed diffusion process. Thanks to the random noise introduced in the diffusion process, \\(\\mathrm{P}_{a2nl}\\) can generate various visual non-lip features given the same audio input, resulting in diverse and reasonable reenactment videos without any extra needs of driving video sources." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.52 + ], + "angle": 0, + "content": "To generate a video sequence longer than the maximal length of the model input with smooth transitions, the first input non-lip feature to \\( \\mathrm{P}_{a2nl} \\) is set to the last generated non-lip feature from \\( \\mathrm{P}_{a2nl} \\) in the previous step, to ensure continuity in non-lip facial motion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.536, + 0.633, + 0.553 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.668 + ], + "angle": 0, + "content": "Our proposed method was evaluated from two aspects: 1) synchronization between audio and lip motions; 2) naturalness and richness of lip-irrelevant facial motions. Both quantitative and qualitative experiments were conducted to showcase the superiority of our method. In addition, we also come up with a novel metric to measure the overall quality of generated facial motions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.678, + 0.891, + 0.724 + ], + "angle": 0, + "content": "Datasets Our model was trained on VoxCeleb2 and evaluated on both VoxCeleb1 and VoxCeleb2. Here are the details of the two datasets," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.736, + 0.891, + 0.781 + ], + "angle": 0, + "content": "- VoxCeleb1 [33]: The dataset consists of 100,000 utterances from 1,251 celebrities. 100 videos with 25 identities in total were randomly chosen for the test." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.787, + 0.892, + 0.833 + ], + "angle": 0, + "content": "- VoxCeleb2 [7]: The dataset contains 1 million utterances with 6,112 identities. 500 test videos with 25 identities were randomly chosen for the test." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.736, + 0.892, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparing Methods Some prior works such as PC-AVS [65], GC-AVT [28], and EAMM [20] require additional driving sources instead of predicting them based on" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7649" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.09, + 0.473, + 0.244 + ], + "angle": 0, + "content": "
MethodA S SctgntVoxCeleb2VoxCeleb1
FID ↓Sc ↑Nc ↓FID ↓Sc ↑Nc ↓
Wav2Lip [38]X X7.8022.39.230.1844.78.800.13
EAMM [20]X X1.7626.14.751.7042.62.780.58
PC-AVS [65]X X7.3514.48.210.1235.48.420.15
MakeItTalk [66]✓ X7.3519.52.030.7240.22.160.71
Audio2head [54]✓ X7.351016.420.131046.650.10
Ours + AR✓ X7.3524.37.230.0249.77.050.04
Ours + Diff.✓ ✓7.3514.27.340.0035.07.310.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.246, + 0.47, + 0.316 + ], + "angle": 0, + "content": "Table 1. The quantitative results of synchronization and image quality under self-reenactment scenario. Bold means the best. A indicates the ability to generate non-lip motions with audio-only signals without other signals, while S means the ability to sample different results for each round." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.328, + 0.191, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.419, + 0.151, + 0.429 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.33, + 0.298, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.419, + 0.261, + 0.429 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.33, + 0.409, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.419, + 0.37, + 0.429 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.433, + 0.434, + 0.448 + ], + "angle": 0, + "content": "Figure 4. Distribution visualization of generated poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.6 + ], + "angle": 0, + "content": "audio input. We only compare those methods on lip synchronization and image quality but emphasize that our setup is targeting a more practical scenario. Other works such as Wav2Lip [38] only focus on revising the lip region without touching other facial motions, which are not directly comparable either. MakeItTalk [66] and Audio2head [54] are two closely related works to ours in terms of problem setup, however, our method is designed to predict all the non-lip facial motions with diversity, rather than only limited poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.716 + ], + "angle": 0, + "content": "Implementation Details Backbones of our models including \\( \\mathrm{E}_{id} \\), \\( \\mathrm{E}_v \\), \\( \\mathrm{E}_a \\), \\( \\mathrm{G} \\), and \\( \\mathrm{D} \\) are borrowed from [65]. For the diffusion prior network \\( \\mathrm{P}_{a2nl} \\), we use a 8-layer transformer [51] with 512-d tokens and 1024-d fully feedforward layers. Note that our models were trained on VoxCeleb2 only but tested on both VoxCeleb1 and VoxCeleb2. All models were trained on 4 NVIDIA A100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.3, + 0.743 + ], + "angle": 0, + "content": "4.1. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics We evaluate the performance of generated talking head from the following aspects: image quality, audio-lip sync accuracy, the variation and naturalness of non-lip facial motion. Frechet Inception Distance (FID) score [14] is used for the evaluation of image quality. A lower FID score indicates a lower distance between the distribution of generated images and real images. Following prior works [38, 65], we use SyncNet Error-Confidence \\((\\mathbf{S}_c)\\) as the indicator of audio-lip synchronization, where greater confidence indicates better synchronization. However, the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.207 + ], + "angle": 0, + "content": "normalized confidence score NLSE-C (short for \\(\\mathbf{N}_c\\)) is used to address the concern raised in recent works [57, 53] regarding the strong relationship between \\(\\mathbf{S}_c\\) and its training data, which can make it unfair to compare methods trained on different data. \\(\\mathbf{N}_c\\) is defined as \\(\\mathbf{N}_c = \\frac{\\left|\\mathbf{S}_c^{gen} - \\mathbf{S}_c^{gt}\\right|}{\\mathbf{S}_c^{gt}}\\), where \\(\\mathbf{S}_c^{gen}\\) is the confidence value of generated images and \\(\\mathbf{S}_c^{gt}\\) is that of its training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.207, + 0.892, + 0.268 + ], + "angle": 0, + "content": "To evaluate facial motions such as pose, expression, and blink, we utilize a pre-trained 3D morphable face model [9] and include shape irrelevant 3DMM parameters to calculate the following metrics." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.277, + 0.892, + 0.338 + ], + "angle": 0, + "content": "- Var: The variance of generated facial motions, i.e., the variance of the 3DMM coefficients for each video is calculated and then averaged over the test set. A closer Var to GT indicates a better match to the variation of real data." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.342, + 0.892, + 0.39 + ], + "angle": 0, + "content": "- \\(\\mathbf{FID}_{fm}\\): FID score of 3DMM coefficients calculated as follows: \\(\\mathbf{FID}_{fm} = \\frac{1}{K}\\sum_{i=1}^{K}\\mathbf{FID}(\\beta_i)\\), where \\(\\beta_i\\) is a sequence of 3DMM coefficients in the \\(i\\)-th video." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.395, + 0.892, + 0.469 + ], + "angle": 0, + "content": "- \\(\\mathbf{FID}_{\\Delta fm}\\): FID score of 3DMM coefficient difference between consecutive frames, i.e., \\(\\mathbf{FID}_{\\Delta fm}\\), which is similar to \\(\\mathbf{FID}_{fm}\\) but the 3DMM coefficient difference between consecutive frames is measured, taking temporal naturalness into consideration." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.476, + 0.892, + 0.567 + ], + "angle": 0, + "content": "- SND: Our new proposed metric, denoted as Sequence Naturalness Distance. It is the sum of \\(\\mathbf{FID}_{fm}\\) and \\(\\mathbf{FID}_{\\Delta fm}\\) indicating the difference of distribution between generated motion and GT motion, from both spatial and temporal perspectives, i.e., \\(\\mathbf{SND} = \\mathbf{FID}_{fm} + \\mathbf{FID}_{\\Delta fm}\\). The lower SND, the better naturalness." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.277, + 0.892, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Evaluation Results Our main results consist of two parts: audio-lip synchronization and image quality as shown in Table 1; richness and sequence naturalness which are shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.75 + ], + "angle": 0, + "content": "It shows that our method with diffusion prior archives the best synchronization ability compared to other methods in Table 1. Note that although Wav2Lip [38] and PC-AVS [65] achieve the highest SyncNet scores, our best \\( N_{c} \\) indicates the most substantial synchronization ability [57], which is demonstrated in Sec. 4.2. Meanwhile, the image quality of our model with diffusion prior is the best in both test sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In Table 2, our method shows significantly higher variance than the other two audio-driven methods and it is closer to the variance of real data. This indicates that our method can produce reasonable diverse head movements and expressions, rather than slight head movement. It's notable that auto-regressive prior, trained with causal attention and regression loss, has a higher variance than GT because it often generates extreme motions, showing a low naturalness score in Table 3. For naturalness, our method achieves the lowest \\(\\mathbf{FID}_{fm}\\), \\(\\mathbf{FID}_{\\Delta fm}\\) as well as SND on" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7650" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.125, + 0.089, + 0.849, + 0.24 + ], + "angle": 0, + "content": "
MethodVoxCeleb2VoxCeleb1
Var →FIDfm ↓FIDΔfm ↓SND ↓Var →FIDfm ↓FIDΔfm ↓SND ↓
GT1.98---1.88---
MakeItTalk [66]0.674.701.746.440.804.271.075.34
Audio2head [54]0.895.941.307.240.765.031.016.04
Ours + AR3.075.432.227.652.926.211.687.89
Ours + Diff.1.573.601.084.681.663.980.874.85
w/o Lvel2.094.021.275.292.284.651.115.76
training w/o editing3.096.762.008.763.066.961.708.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.24, + 0.894, + 0.269 + ], + "angle": 0, + "content": "Table 2. The quantitative results of variance and naturalness of generated non-lip motions under self-reenactment scenario. \\( \\rightarrow \\) indicates a closer score to GT is better. Due to setting differences, i.e., Wav2Lip, EAMM, and PC-AVS, are not valid comparisons here." + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.275, + 0.847, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.787, + 0.895, + 0.83 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of our method compared to other baselines. Each row shows nine uniformly sampled frames from videos. Here we use two audio sources to drive different identities respectively. Our method shows accurate lip-audio synchronization with diverse and natural poses and expressions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.854, + 0.472, + 0.9 + ], + "angle": 0, + "content": "both VoxCeleb1 and VoxCeleb2, surpassing other state-of-the-art by a large margin. Also, we randomly sample 5,000 poses for each method and employ t-SNE [50] for visual-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.854, + 0.895, + 0.901 + ], + "angle": 0, + "content": "ization in Fig. 4, which demonstrates that the distribution we generate is the most comprehensive to GT. More details and examples are shown in the supplementary. In summary," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "7651" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.46, + 0.214 + ], + "angle": 0, + "content": "
MethodSync AccuracyNaturalnessRichness
GT4.614.414.18
MakeItTalk [66]1.972.352.10
PC-AVS [65]*3.132.682.68
Audio2head [54]2.562.352.34
Ours + AR3.462.533.37
Ours + Diff.4.083.823.68
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.078, + 0.214, + 0.468, + 0.24 + ], + "angle": 0, + "content": "Table 3. Human evaluation on generated samples on VoxCeleb2. * means that PC-AVS was evaluated on cross-reenactment scenario." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.264, + 0.469, + 0.294 + ], + "angle": 0, + "content": "our method can generate rich facial motions with excellent realism (i.e., a lower distribution distance) and naturalness." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.305, + 0.287, + 0.32 + ], + "angle": 0, + "content": "4.2. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.329, + 0.469, + 0.419 + ], + "angle": 0, + "content": "Two audio-driven video sequences of our method verses other baselines are shown in Fig. 5. It's clearly shown that our method is capable of producing much more natural-looking and diverse facial motions than other baselines, including pose, expression, blink and gaze, while maintaining accurate audio-lip synchronization compared to the GT." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.429, + 0.469, + 0.745 + ], + "angle": 0, + "content": "User Study We invited 20 subjects for human evaluations, focusing on three aspects: 1) the accuracy of audio-lip synchronization; 2) the naturalness of non-lip motions and coherency between non-lip motions and audios; 3) the richness of non-lip motions. 50 videos are generated from audio and identity in the test set using the following methods: our diffusion prior, our auto-regressive prior, MakeItTalk [66], Audio2head [54] and PC-AVS [65] (driven by poses from another video). After shuffling the generated videos, each annotator rates from 1 (bad) to 5 (good) according to Mean Opinion Scores (MOS) rating protocol. Table 3 tells that our model with diffusion prior archives the best synchronization, naturalness and richness among all the models. Our model with auto-regressive prior archives relatively good richness compared to the remaining models, we attribute it to the exposure bias problem [4] brought by auto-regressive generation, which leads to high richness but low naturalness. Note that PC-AVS shows incompatible result in Table 3 as compared to Table 1 on \\(\\mathbf{S}_c\\), indicating that choosing pose driving signals from a random video may lead to synthesis results with bad audio-lip synchronization." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.746, + 0.469, + 0.806 + ], + "angle": 0, + "content": "It's interesting that our proposed SND scores reflect the naturalness of user study to some degree. Nevertheless, we leave it to future work for rigorous study of their correlation through more human evaluations." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.817, + 0.182, + 0.832 + ], + "angle": 0, + "content": "4.3. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Table 4 discusses the performance of lip & non-lip disentanglement. While tested on lip only metrics without \\( L_{ol} \\), SyncNet scores get lower, which indicates that G tends to generate mouth movements with information extracted" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.895, + 0.184 + ], + "angle": 0, + "content": "
MethodLip OnlyNon-lip Only
Sc↑Nc↓Bd↓Gd↓Ed↓Pd↓
GT7.35-----
w/o Lol2.840.610.00740.19600.05680.0021
w/o MB7.690.050.01120.03150.08750.0293
Ours7.480.030.00670.01630.05580.0019
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.187, + 0.892, + 0.271 + ], + "angle": 0, + "content": "Table 4. Evaluation of disentanglement between lip and non-lip on VoxCeleb2. \\( \\mathbf{B}_d \\) and \\( \\mathrm{G}_d \\) denote L2 distances of 2D landmark [58] of eyes and iris, respectively, while \\( \\mathrm{E}_d \\) and \\( \\mathrm{P}_d \\) denote L2 distances of 3DMM coefficients of poses and expressions accordingly. Note that \\( \\mathbf{f}_{nl}^v \\) is fixed while tested on lip only metrics, and \\( \\mathbf{f}_l^a \\) is fixed while tested on non-lip only metrics for fair comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.892, + 0.371 + ], + "angle": 0, + "content": "from visual modality instead of audio. Thus, lip & non-lip space are not fully disentangled. Without MB, the performance of driving non-lip motions gets worse. Because it is hard to use a large batch to compute correlation, resulting in a not well decoupled non-lip space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.893, + 0.614 + ], + "angle": 0, + "content": "We conducted ablation study for \\(\\mathrm{P}_{a2nl}\\) from three aspects: 1) auto-regressive prior versus diffusion prior; 2) with/without velocity loss \\(L_{vel}\\); 3) training with/without editing mechanism. Diffusion prior shows significant improvements on naturalness scores including \\(\\mathrm{FID}_{fm}\\), \\(\\mathrm{FID}_{\\Delta fm}\\) and SND, as compared to autoregressive prior shown in Table 2, with slightly better synchronization ability in Table 1. Fig. 6 shows that without \\(L_{vel}\\), the prediction of non-autoregressive diffusion prior model \\(\\mathrm{P}_{a2nl}\\) becomes unstable, which leads to a larger variance and worse SND scores in Table 2. While training without the editing mechanism mentioned in Sec. 3.2, we applied editing only during sampling as in MDM [46], and observed that it jitters between adjacent frames as shown in Fig. 7, resulting in a poorer SND score in Table 2. Note that our model trained with editing mechanism generates smoother results." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.619, + 0.876, + 0.747 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.753, + 0.892, + 0.781 + ], + "angle": 0, + "content": "Figure 6. Ablation study of our diffusion prior \\( \\mathrm{P}_{a2nl} \\) w or w/o \\( L_{vel} \\). Each row shows five uniformly sampled frames." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.627, + 0.815 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a novel talking head generation method based on a diffusion prior model, which can generate diverse and natural-looking talking head videos with only audio and an identity image as inputs. Such \"audio-driving is all you need\" setting is very friendly" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7652" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.092, + 0.454, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.226, + 0.47, + 0.253 + ], + "angle": 0, + "content": "Figure 7. Ablation study of our diffusion prior \\(\\mathbf{P}_{a2nl}\\) trained w or w/o editing. Each row shows five adjacent frame." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.276, + 0.469, + 0.322 + ], + "angle": 0, + "content": "for reenactment and dubbing applications. Comprehensive evaluations including our newly proposed metrics validated the effectiveness of our system." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.329, + 0.469, + 0.373 + ], + "angle": 0, + "content": "Limitations Our method mainly focuses on non-lip motion generation. We leave the improvement of rendering quality to future works." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.387, + 0.173, + 0.401 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.412, + 0.47, + 0.465 + ], + "angle": 0, + "content": "[1] Ahmed A Abdelrahman, Thorsten Hempel, Aly Khalifa, and Ayoub Al-Hamadi. L2cs-net: Fine-grained gaze estimation in unconstrained environments. arXiv preprint arXiv:2203.03339, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.468, + 0.469, + 0.534 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelović, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-Supervised MultiModal Versatile Networks. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.469, + 0.591 + ], + "angle": 0, + "content": "[3] Mohammed M Alghamdi, He Wang, Andrew J Bulpitt, and David C Hogg. Talking head from speech audio using a pretrained image generator. arXiv preprint arXiv:2209.04252, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.594, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[4] Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems, 28, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.65, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[5] Stella Bounareli, Vasileios Argyriou, and Georgios Tzimiropoulos. Finding directions in gan's latent space for neural face reenactment. arXiv preprint arXiv:2202.00046, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.706, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[6] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.762, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.804, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[8] Soo-Whan Chung, Joon Son Chung, and Hong-Goo Kang. Perfect match: Improved cross-modal embeddings for audiovisual synchronisation. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3965-3969. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.874, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.412, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[11] Aviv Gabbay, Asaph Shamir, and Shmuel Peleg. Visual speech enhancement. arXiv preprint arXiv:1711.08789, 2017.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[13] Yudong Guo, Keyu Chen, Sen Liang, Yongjin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[16] Fangzhou Hong, Liang Pan, Zhongang Cai, and Ziwei Liu. Versatile multi-modal pre-training for human-centric perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16156-16166, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[17] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3397-3406, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[18] Gee-Sern Hsu, Chun-Hung Tsai, and Hung-Yi Wu. Dual-generator face reenactment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 642-650, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[19] Po-Hsiang Huang, Fu-En Yang, and Yu-Chiang Frank Wang. Learning identity-invariant motion representations for crossid face reenactment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7084–7092, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[20] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. arXiv preprint arXiv:2205.15278, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[21] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[22] Venkatesh S Kadandale, Juan F Montesinos, and Gloria Haro. Vocalist: An audio-visual synchronisation model for lips and voices. arXiv preprint arXiv:2204.02090, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "7653" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.472, + 0.22 + ], + "angle": 0, + "content": "[24] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Nießner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics, 37(4):163:1-14, Aug. 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.471, + 0.264 + ], + "angle": 0, + "content": "[25] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.267, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[26] Jiyoung Lee, Soo-Whan Chung, Sunok Kim, Hong-Goo Kang, and Kwanghoon Sohn. Looking into your speech: Learning cross-modal affinity for audio-visual speech separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1336–1345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[27] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3387-3396, June 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.439, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[28] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[29] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.47, + 0.638 + ], + "angle": 0, + "content": "[30] Moustafa Meshry, Saksham Suri, Larry S Davis, and Abhnav Shrivastava. Learned spatial representations for few-shot talking-head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13829-13838, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.641, + 0.47, + 0.682 + ], + "angle": 0, + "content": "[31] Dongchan Min, Minyoung Song, and Sung Ju Hwang. Stylistalker: One-shot style-based audio-driven talking head video generation. preprint arXiv:2208.10922, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.47, + 0.74 + ], + "angle": 0, + "content": "[32] Gaurav Mittal and Baoyuan Wang. Animating face using disentangled audio representations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), March 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[33] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. Voxceleb: a large-scale speaker identification dataset. arXiv preprint arXiv:1706.08612, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[34] Arsha Nagrani, Joon son Chung, Samuel Albanie, and Andrew Zisserman. Disentangled speech embeddings using cross-modal self-supervision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[36] Andrew Zisserman Olivia Wiles, A. Sophia Koepke. X2face: A network for controlling face generation by using images, audio, and pose codes. In ECCV 2018, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[37] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In 36th AAAI Conference on Artificial Intelligence (AAAI 22). Association for the Advancement of Artificial Intelligence, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[38] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.893, + 0.488 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[42] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. arXiv preprint arXiv:2207.11770, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[43] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[46] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[47] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. ECCV 2020, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[48] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7654" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[49] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. arXiv preprint arXiv:1906.05849, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[50] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[52] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. CoRR, abs/1906.06337, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[53] Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang. Progressive disentangled representation learning for fine-grained controllable talking head synthesis. arXiv preprint arXiv:2211.14506, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[54] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.471, + 0.442 + ], + "angle": 0, + "content": "[55] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2531–2539, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.471, + 0.537 + ], + "angle": 0, + "content": "[56] Chao Xu, Jiangning Zhang, Yue Han, Guanzhong Tian, Xianfang Zeng, Ying Tai, Yabiao Wang, Chengjie Wang, and Yong Liu. Designing one unified framework for high-fidelity face reenactment and swapping. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 54-71. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[57] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[58] Baosheng Yu and Dacheng Tao. Heatmap regression via randomized rounding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[59] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[60] Xianfang Zeng, Yusu Pan, Mengmeng Wang, Jiangning Zhang, and Yong Liu. Realistic face reenactment via self-supervised disentangling of identity and pose. pages 12757-12764. AAAI Press, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[61] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3867-3876, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[62] Jiangning Zhang, Liang Liu, Zhucun Xue, and Yong Liu. Apb2face: Audio-guided face reenactment with auxiliary pose and blink signals. 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[63] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.895, + 0.134 + ], + "angle": 0, + "content": "a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661-3670, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.895, + 0.205 + ], + "angle": 0, + "content": "[64] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9299–9306, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.895, + 0.289 + ], + "angle": 0, + "content": "[65] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.895, + 0.346 + ], + "angle": 0, + "content": "[66] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 2, 3, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.895, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7655" + } + ] +] \ No newline at end of file diff --git a/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/7d213d41-b6ba-4fcf-a556-dbfb8dc18cb6_origin.pdf b/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/7d213d41-b6ba-4fcf-a556-dbfb8dc18cb6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2dafa72b4dcc67a0beecb143ccc05c91c9b2f7a3 --- /dev/null +++ b/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/7d213d41-b6ba-4fcf-a556-dbfb8dc18cb6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:891d930e8010ebfd39e77215ba5bd52b158c958645092c6a399454752933ab7f +size 10036170 diff --git a/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/full.md b/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4af141ee127173688836b025e878d71409a73b11 --- /dev/null +++ b/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/full.md @@ -0,0 +1,321 @@ +# Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors + +Zhentao Yu $^{1*}$ Zixin Yin $^{1,2*†}$ Deyu Zhou $^{1,3*†}$ Duomin Wang $^{1}$ Finn Wong $^{1}$ Baoyuan Wang $^{1‡}$ Xiaobing.AI. + +$^{2}$ The Hong Kong University of Science and Technology. $^{3}$ The Hong Kong University of Science and Technology (Guangzhou). + +(yuzhentao, wangduomin, wangwenlan, wangbaoyuan)@xiaobing.ai zyinaf@connect.ust.hk, dzhou861@connect.hkust-gz.edu.cn + +![](images/bfba7cfa543a582cf526bb6a6de99c5ebfa6d99a1ed18db19c2e179cfdbc5791.jpg) +Figure 1. Given only an audio source and an arbitrary identity image, our system can generate a video with natural-looking and diverse facial motions (pose, expression, blink & gaze), while maintaining accurate audio-lip synchronization. Here we show randomly sampled sequences from our diffusion prior for two identities, note that the lip-irrelevant facial motion varies but the lip is still in-sync. + +# Abstract + +We introduce a novel framework for one-shot audio-driven talking head generation. Unlike prior works that require additional driving sources for controlled synthesis in a deterministic manner, we instead sample all holistic lip-irrelevant facial motions (i.e. pose, expression, blink, gaze, etc.) to semantically match the input audio while still maintaining both the photo-realism of audio-lip synchronization and overall naturalness. This is achieved by our newly proposed audio-to-visual diffusion prior trained on top of the mapping between audio and non-lip representations. Thanks to the probabilistic nature of the diffusion prior, one big advantage of our framework is it can synthe + +size diverse facial motion sequences given the same audio clip, which is quite user-friendly for many real applications. Through comprehensive evaluations of public benchmarks, we conclude that (1) our diffusion prior outperforms autoregressive prior significantly on all the concerned metrics; (2) our overall system is competitive with prior works in terms of audio-lip synchronization but can effectively sample rich and natural-looking lip-irrelevant facial motions while still semantically harmonized with the audio input. + +# 1. Introduction + +Audio-driven face reenactment and talking head generation have received raising attention due to the broad killer applications in movie production, gaming, virtual digital avatars, and potentially more while we move toward the era + +of the metaverse. The past literature tries to advance this area from various perspectives. One line of research focuses on improving the generation quality originating from regular GAN-based methods [64, 65] to pre-trained StyleGAN [3] and to the most recent Neural Radiance Field (NeRF) based methods [13, 42]. Orthogonal to direction, another line of works emphasizes the importance of disentangled representation from the audio [32, 21, 66] for controlled generations. i.e., [21] can predict the emotion from the audio input while [66] can decouple the speech signal into speaker identity and the phonetic content. A closely related set of works targeting more granular controlled talking head generation by providing additional input signals. For example, PC-AVS [65] requires a separate video to provide the pose signal in addition to the audio clip but does not support the control of expression and blinking. To remedy this, GC-AVT [27] requires inputting additional driving video for expression in addition to pose and audio driving clip. The technical challenge behind those approaches is how to faithfully transfer the desired driving signal (i.e., pose, expression, audio) into the results without affecting each other through intrinsic disentanglement. Although encouraging results have been reported, we argue such a setting is not practical for broader applications. It is quite challenging, or at least labor-intensive, for novice users to find the "best" individual driving sources for each controlled dimension to make the final talking head video overall look not only lifelike but also coherent from semantic and emotional perspectives. Therefore, it is both more practical and generalizable if the setting only requires an audio signal as the driving source and expects a data-driven model to sample reasonable other facial motions irrelevant to lips. + +By no means, we are the first to advocate an audio-only driving setup for talking head generation. Audio2head [54] predicts the poses from audio input with an LSTM network. [29] employs an autoregressive model by assuming the poses are jointly determined by the audio and past head motions along the sequence, although the results are encouraging, the model can't generalize due to their person-specific training strategy. FACIAL [61] could infer the blink, but requires a reference video input rather than a one-shot reference image. There are other related works along this direction, but they either only infer one facial attribute (i.e., emotion in EVP[21], pose from [54]) using ad-hoc methods, or they simply do not support inferring other facial motions [38], or treat it as a block-box mapping model [3] without explicitly respecting the one-to-many mapping nature between audio and other visual facial attributes (pose, expression, blink). A more principled solution is desired to consolidate this line of work. + +In this paper, we introduce a novel framework that can holistically infer all the non-lip-related facial attributes from the audio input while maintaining accurate synchronization + +between audio and the corresponding visual lip motions. This is achieved by two important learning steps in our pipeline. (1) A pre-trained identity-irrelevant facial motion representation can help decouple the lip and non-lip representations. To promote this learning, we employ a novel orthogonal loss on top of the modified facial reenactment framework [6]. The disentanglement enables to generate one-to-one mapping with lip representations to ensure the synchronization, and generate richer non-lip representations with a one-to-many mapping. (2) A novel audio-to-visual diffusion prior model is introduced to address the probabilistic sampling from audio representations to the abovelearned non-lip representation. This prior is expected to solve the one-to-many mapping and provide diverse results during the inference stage. The entire pipeline can be easily built up on top of the existing framework, such as PC-AVS[65] without heavily retraining every component. To sum up, we make the following contributions: + +- To our best knowledge, we are the first to holistically predict non-lip facial motions based on audio input only, providing good usability for reenactment or dubbing applications without extra driving video sources. Our method addresses the intrinsic one-to-many challenge in a probabilistic way, allowing diverse and realistic facial motion generation under the same audio input, as shown in Fig. 1. +- We leverage the pre-trained visual identity-irrelevant facial motion representations. And further, learn disentangled lip-related and lip-irrelevant representations through a novel orthogonal loss on top of PC-AVS[65]. A powerful diffusion prior model is then introduced to effectively infer all lip-irrelevant facial motions for a given audio segment in the representation space. +- We systematically evaluate the naturalness and diversity of the results with new metrics, which paves a way for future studies. Meanwhile, results show that our method can produce natural-looking head poses and facial motions without hurting the audio-lip synchronization. Our model will be released. + +# 2. Related Works + +Audio-Visual Cross-modal Learning Cross-modal representation learning is a long-standing research topic, ranging from speech enhancement [11], speech source separation [26] to synchronization [8, 22] and other speech disentanglement [34, 21, 32]. Among them, EVP [21] used cross-modal supervision to disentangle speech content and emotion from the audio signal with landmark as the intermediate representation. Recently, CMC [49] discussed how multiview "modality" can be jointly unitized to boost intrinsic representation through contrastive learning, rather than predictive (or reconstruction) learning, it also demonstrated + +that the more views, the better. Concurrently, MMV [2] introduced different modality embedding graphs for effective cross-modal representations, again through contrastive learning. More recently, HCMoCo [16] extended similar ideas with a hierarchical strategy to learn different levels of representations for human-centric perception tasks. Although impressive results were reported, it is still unclear how it performs on face analysis tasks, not mention to on synthesis tasks. Our work shares a similar spirit but a different purpose, where we first pre-train a non-identity visual representation which then helps learn a lip and non-lip space, the latter further serves as the upper-bound learning target for subsequent audio-to-visual diffusion prior. + +Face Reenactment & Talking Head Generation Face Reenactment is designed to transfer part or full facial motion from a driving source to the target video with good IDpreserved appearance and background. It can be further divided into two categories depending on whether the driving source is from video [17, 18, 5, 48, 24, 36, 59, 19, 60, 6, 56] or audio [52, 47, 62, 21, 55, 37, 63, 30]. Among them, audio-driven face reenactment generally aims to edit the mouth regions of the target video in order to match the input audio while leaving other facial attributes mostly unchanged, i.e., pose. EVP [21] tries to infer the non-rigid facial expression in addition to lip motion from the audio input. A closely related line of work is the audio-driven talking-head generations [38, 66, 61, 3] where only one target reference face is given, hence, other face attributes including pose, expression, blink and gaze have to be either explicitly given [65, 27] or partly inferred through statistical methods [66, 31, 61, 29, 3, 47]. Specifically, Lu et al. [29] employed an auto-regressive model while Min et al. [31] leveraged normalized flow prior to predicting a natural-looking pose sequence from the input audio, both showed encouraging results. Compared with them, we aim to infer more diverse facial motions including pose, expression, and even blink and gaze in a holistic manner through an audio-to-visual diffusion prior model. + +Diffusion Generative Models The diffusion model [44, 15, 45], which is a likelihood-based model consisting of cascading denoising autoencoders, has recently shown great success in numerous generative tasks with different modalities including image [10, 35, 40, 41], audio [25], video [43], and motion [46]. To name a few, DDPM [15] explored the diffusion model for unconditional image generation. GLIDE [35] introduced text-conditional diffusion model and showed that classifier free guidance has better performance than CLIP [39] guidance. DALLE-2 [40] modified GLIDE to generate semantically consistent images conditioned on a CLIP image embedding, and proposed a diffusion prior that produces the image embedding given a text caption. MDM [46] utilized a classifier-free diffusion-based + +generative model for text-to-motion and action-to-motion tasks, allowing motion completion and editing as well. + +# 3. Method + +Given $L$ segments of audio $A_{1:L} = (A_1, A_2, \dots, A_L)$ and a reference image $I_{ref}$ as inputs, our model M synthesizes video frames $\hat{X}_{1:L} = (\hat{X}_1, \hat{X}_2, \dots, \hat{X}_L)$ having the same identity as $I_{ref}$ and lip motion synchronized to $A_{1:L}$ : + +$$ +\hat {X} _ {1: L} = \mathbf {M} \left(A _ {1: L}, I _ {r e f}\right). \tag {1} +$$ + +An overview of our proposed framework is shown in Fig. 2, which consists of three major components: + +- Lip & Non-lip Disentanglement Given a pre-trained identity- and appearance-irrelevant facial motion encoder $\mathrm{E}_v$ , we first leverage it to learn two complementary features, $\mathbf{f}_l^a$ for lip, and $\mathbf{f}_{nl}^v$ for non-lip features including pose, expression, blink and gaze. +- Audio to Non-lip Diffusion Prior A diffusion prior network $\mathrm{P}_{a2nl}$ models the one-to-many mapping from audio feature $\mathbf{f}^a$ to the "hallucinated" non-lip feature $\mathbf{f}_{nl}^a$ , which is trained to be close to its visual counterpart $\mathbf{f}_{nl}^v$ , allowing audio-only facial driving at inference time. +- Audio-based Talking Head Generation Given an identity feature $\mathbf{f}_{id}$ , a lip feature $\mathbf{f}_l^a$ , and a non-lip feature $\mathbf{f}_{nl}^a$ generated with the diffusion prior, we concatenate them together and feed into a GAN similar to LPD [6] or PC-AVS [65] to output the final reenacted video. + +# 3.1. Lip & Non-lip Disentanglement + +The identity encoder $\mathrm{E}_{id}$ is designed to capture identity and appearance information, while $\mathrm{E}_v$ is instead to remove both while encoding all facial motions. Both encoders are pre-trained in a similar setting as LPD [6] and PC-AVS [65]. After pretraining, we conduct lip & non-lip disentanglement consisting of audio-visual contrastive learning for lip feature learning and decorrelation for non-lip feature learning, as well as face reconstruction, as described below: + +Audio-Visual Pretraining for Audio Lip Space An audio encoder $\mathrm{E}_a$ is trained to provide accurate audio feature $\mathbf{f}^a = \mathrm{E}_a(A_{1:N})$ through contrastive learning [39] with $\mathbf{f}_{cl}^v = \mathrm{MLP}_{cl}(\mathrm{E}_v(X_{1:N}))$ , where $N$ is the number of samples and $X_{1:N}$ are $N$ frames in the same video corresponding to audio $A_{1:N}$ . Then $\mathbf{f}^a$ is further compressed to filter out lip-relevant information and get audio lip feature $\mathbf{f}_l^a$ through $\mathrm{MLP}_{a2l}$ since $\mathbf{f}^a$ mainly contains lip related information, as shown in Sec. 4.3. The contrastive loss is defined as $L_{con}(m,n) = -\frac{1}{N}\sum_{i=1}^{N}\log\frac{\exp(m_i/\|m_i\|_2 \cdot n_i/\|n_i\|_2)}{\sum_{j=1}^{N}\exp(m_i/\|m_i\|_2 \cdot n_j/\|n_j\|_2)}$ , resulting in a total contrastive loss between $\mathbf{f}^a$ and $\mathbf{f}_{cl}^v$ , + +$$ +L _ {c l} = \frac {1}{2} \left[ L _ {\text {c o n}} \left(\mathbf {f} ^ {a}, \mathbf {f} _ {c l} ^ {v}\right) + L _ {\text {c o n}} \left(\mathbf {f} _ {c l} ^ {v}, \mathbf {f} ^ {a}\right) \right]. \tag {2} +$$ + +![](images/f1a80a4f86f0080f1fe4c303c1bdb2d75be6d9fd455043f7e18b5a635bc1e1c3.jpg) +Figure 2. The overall pipeline of our proposed framework. The dotted lines represent the loss functions that are used only in training, i.e., $L_{cl}$ , $L_{ol}$ and $L_{a2nl}$ . $L_{G}$ is the training loss for visual non-lip space. Note that there is a switch in the figure, which means the different forward processes in training and inference: in training, the concatenated feature $\mathbf{f}_{cat} = cat(\mathbf{f}_{id},\mathbf{f}_{nl}^{v},\mathbf{f}_{l}^{a})$ ; at inference stage, $\mathbf{f}_{cat} = cat(\mathbf{f}_{id},\mathbf{f}_{nl}^{a},\mathbf{f}_{l}^{a})$ . Thus, the visual motion part is not needed anymore at the inference stage. + +Reconstruction Learning for Visual Non-lip Space To learn a good non-lip space, we propose an orthogonal loss to penalize the correlation between $\mathbf{f}_{ol}^{v}$ and $\mathbf{f}^{\alpha}$ , which uses the well-learned audio space to disentangle lip-related motions. The orthogonal loss is used together with reconstruction since there should also be a completeness constraint to avoid mode collapse of lip-irrelevant features. In practice, we maintain two memory banks (MB) for storing $\mathbf{f}_{ol}^{v}$ and $\mathbf{f}^{\alpha}$ in previous $K - 1$ steps, in order to have more samples than $N$ . We denote the feature dimension of $\mathbf{f}^{\alpha}$ as $d_{a}$ , and $\mathbf{f}_{ol}^{v}$ as $d_{v}$ . The orthogonal loss is defined as follows: + +$$ +L _ {o l} = \frac {1}{d _ {v}} \sum_ {i = 1} ^ {d _ {v}} \sum_ {j = 1} ^ {d _ {a}} P _ {c o r} \left(\mathbf {f} _ {o l} ^ {v}, \mathbf {f} ^ {a}\right) _ {(i, j)} ^ {2}, \tag {3} +$$ + +where $P_{cor}(\mathbf{f}_{ol}^v,\mathbf{f}^a)\in d_a\times d_v$ computes the Pearson correlation coefficient between $\mathbf{f}_{ol}^{v}$ and $\mathbf{f}^a$ . Since the correlation between lip motions and audio features is strong, $L_{ol}$ (which encourages the linear uncorrelation between $\mathbf{f}_{ol}^{v}$ and $\mathbf{f}^a$ ) is effective to disentangle lip-related motions. Then, $\mathrm{MLP}_{nl}$ further projects $\mathbf{f}_{ol}^{v}$ into a smaller space to remove undesired content and get non-lip feature $\mathbf{f}_{nl}^{v}$ . + +We follow the reconstruction loss $L_{\mathrm{GAN}}$ , $L_{L1}$ and $L_{\mathrm{VGG}}$ used in [65]. The generator $G$ receives the concatenation $\mathbf{f}_{cat} = cat(\mathbf{f}_{id}, \mathbf{f}_{nl}^{v}, \mathbf{f}_{l}^{a})$ as inputs and generates final images $I_{gen}$ with modulated convolution. The discriminator $D$ is trained jointly by generative adversarial learning [23, 12]. Note that we omit the batch-mean operation here for convenience, but in practice, all losses used here are calculated with the average in a batch. Different from [65], we propose an additional gaze loss $L_{gaze}$ by utilizing a pre-trained gaze encoder [1], which calculates the $L1$ distance between the gaze features of generated images and ground-truth (GT) images $I$ as follows: + +$$ +L _ {g a z e} = \left\| \Phi (I) - \Phi \left(I _ {g e n}\right) \right\| _ {1}. \tag {4} +$$ + +The total loss $L_{G}$ is given as: + +$$ +\begin{array}{l} L _ {G} = \lambda_ {o l} \cdot L _ {o l} + \lambda_ {g a z e} \cdot L _ {g a z e} + \lambda_ {L 1} \cdot L _ {L 1} \tag {5} \\ + \lambda_ {\text {G A N}} \cdot L _ {\text {G A N}} + \lambda_ {\text {V G G}} \cdot L _ {\text {V G G}}, \\ \end{array} +$$ + +where $\lambda$ s are weights for losses. + +# 3.2. Audio2nonlip Diffusion Prior + +Beside a strong relationship with lip motions, audio features have a complex non-linear relationship with non-lip motions and it is a one-to-many mapping problem. This is due to the fact that there are many reasonable facial motions that can be corresponded to the same audio input. As inspired by DALLE-2 [40] and MDM [46], we design an audio2nonlip diffusion prior network $\mathrm{P}_{a2nl}$ to address this problem. Our proposed network is depicted in Fig. 3. + +Diffusion Model Diffusion model is designed based on the stochastic diffusion process. The forward diffusion process is defined below: + +$$ +q \left(n _ {1: L} ^ {t} \mid n _ {1: L} ^ {t - 1}\right) = \mathcal {N} \left(n _ {1: L} ^ {t - 1}; \sqrt {\alpha^ {t}} n _ {1: L} ^ {t - 1}, (1 - \alpha^ {t}) I\right), \tag {6} +$$ + +In our context, $n_{1:L} = (n_1,n_2,\dots,n_L)$ is a sequence of non-lip feature $\mathbf{f}_{nl}^v$ . $\alpha^t$ is a hyper-parameter and $t\sim [1,T]$ is the time step of the diffusion process. Eq. 6 can be approximated to $n_{1:L}^{T}\sim \mathcal{N}(0,1)$ if $\alpha^t$ is small enough. + +During the reversed diffusion process, $\mathrm{P}_{a2nl}$ models the audio2nonlip distribution as $p(n_{1:L}^{0}|a_{1:L})$ , where $a_{1:L} = (a_1,a_2,\dots,a_L)$ is a sequence of audio feature $\mathbf{f}^a$ . Instead of predicting the noise as formulated by vanilla DDPM [45], $\mathrm{P}_{a2nl}$ learns to predict the initial signal itself with the simplified objective function [45] described as follows: + +$$ +L _ {s i m p l e} = \mathbb {E} _ {n ^ {0} \sim q (n ^ {0} | \mathbf {f} ^ {\alpha}), t \sim [ 1, T ]} \left[ \left\| n ^ {0} - \mathrm {P} _ {a 2 n l} \left(n ^ {t}, t, \mathbf {f} ^ {\alpha}\right) \right\| _ {2} \right], \tag {7} +$$ + +![](images/942fce6204997a8e2459c4065a7ffff1084e5a0fbfa0f70974050a23eda17ad9.jpg) +Figure 3. Our audio2nonlip diffusion prior network $\mathrm{P}_{a2nl}$ . + +where we use $n^0$ to represent $n_{1:L}^{0}$ for convenience. + +We implement $\mathrm{P}_{a2nl}$ with a transformer encoder [51] architecture. As shown in Fig. 3, an audio feature $a_{1:L}$ is added to the time embedding $\mathrm{Emb}_{\text{time}}$ . Then, it's concatenated with noisy non-lip feature $n_{1:L}^{t}$ and added with positional embeddings $\mathrm{Emb}_{pos}$ [51]. The model encodes the embeddings with bidirectional self-attention and feedforward layers then output denoised non-lip feature $n_{1:L}^{0}$ . + +Velocity Loss Note that our designed $\mathrm{P}_{a2nl}$ denoises $n_{1:L}^{t}$ in a non-autoregressive way. To encourage the naturalness and coherence of generated non-lip motion, we borrow the velocity loss from MDM [46] defined as follows: + +$$ +L _ {v e l} = \frac {1}{L - 1} \sum_ {i = 2} ^ {L} \left\| \left(\hat {n} _ {i} ^ {0} - \hat {n} _ {i - 1} ^ {0}\right) - \left(n _ {i} ^ {0} - n _ {i - 1} ^ {0}\right) \right\| _ {2}, \tag {8} +$$ + +where $n_i^0$ is the GT non-lip feature and $\hat{n}_i^0$ is the predicted denoised non-lip feature at the $i$ -th position respectively. The intuition here is that the difference between adjacent non-lip features should be close to the difference in the GT. The total loss of this stage is defined as: + +$$ +L _ {a 2 n l} = L _ {\text {s i m p l e}} + L _ {v e l}. \tag {9} +$$ + +Classifier-free Guidance We use classifier-free guidance [35] for conditioned diffusion generation. In training time, we randomly set the condition to $\emptyset$ for $10\%$ of the samples so that $\mathrm{P}_{a2nl}(n^t,t,\emptyset)$ approximates $n^0$ . At inference stage, the output of the $P_{a2nl}$ is extrapolated further in the direction of $\mathrm{P}_{a2nl}(n^t,t,\mathbf{f}^a)$ and away from $\mathrm{P}_{a2nl}(n^t,t,\emptyset)$ : + +$$ +\mathrm {P} _ {a 2 n l} \left(n ^ {t}, t, \mathbf {f} ^ {a}\right) = s \cdot \mathrm {P} _ {a 2 n l} \left(n ^ {t}, t, \mathbf {f} ^ {a}\right) + (1 - s) \cdot \mathrm {P} _ {a 2 n l} \left(n ^ {t}, t, \emptyset\right), \tag {10} +$$ + +where $s$ is a scaling parameter while increasing it improves sample quality at the cost of diversity. + +Sequential Mask Editing To deliver a smooth non-lip facial sequence with arbitrary length, we design a mechanism to ensure continuity between generated segments, and provide an editing method to do so. Specifically, we randomly mask $90\%$ of the non-lip tokens $n^0$ and concatenate them with noisy non-lip tokens $n^t$ as the input of $\mathrm{P}_{a2nl}$ at training stage. Practically, we found it easy for $\mathrm{P}_{a2nl}$ to quickly learn how to fill in the masked region using hints from the non-lip features in the unmasked regions, generating continuous feature prediction without any extra design in training loss. As a result, it is possible to use $\mathrm{P}_{a2nl}$ for audio-guided non-lip motion generation as well as non-lip motion editing at inference time, providing non-lip facial motion with good naturalism and diversity. + +# 3.3. Audio-based Talking Head Generation + +The main difference at inference time is that non-lip visual feature $\mathbf{f}_{nl}^{v}$ is replaced with $\mathbf{f}_{nl}^{a}$ generated by $\mathrm{P}_{a2nl}$ through reversed diffusion process. Thanks to the random noise introduced in the diffusion process, $\mathrm{P}_{a2nl}$ can generate various visual non-lip features given the same audio input, resulting in diverse and reasonable reenactment videos without any extra needs of driving video sources. + +To generate a video sequence longer than the maximal length of the model input with smooth transitions, the first input non-lip feature to $\mathrm{P}_{a2nl}$ is set to the last generated non-lip feature from $\mathrm{P}_{a2nl}$ in the previous step, to ensure continuity in non-lip facial motion. + +# 4. Experiments + +Our proposed method was evaluated from two aspects: 1) synchronization between audio and lip motions; 2) naturalness and richness of lip-irrelevant facial motions. Both quantitative and qualitative experiments were conducted to showcase the superiority of our method. In addition, we also come up with a novel metric to measure the overall quality of generated facial motions. + +Datasets Our model was trained on VoxCeleb2 and evaluated on both VoxCeleb1 and VoxCeleb2. Here are the details of the two datasets, + +- VoxCeleb1 [33]: The dataset consists of 100,000 utterances from 1,251 celebrities. 100 videos with 25 identities in total were randomly chosen for the test. +- VoxCeleb2 [7]: The dataset contains 1 million utterances with 6,112 identities. 500 test videos with 25 identities were randomly chosen for the test. + +Comparing Methods Some prior works such as PC-AVS [65], GC-AVT [28], and EAMM [20] require additional driving sources instead of predicting them based on + +
MethodA S SctgntVoxCeleb2VoxCeleb1
FID ↓Sc ↑Nc ↓FID ↓Sc ↑Nc ↓
Wav2Lip [38]X X7.8022.39.230.1844.78.800.13
EAMM [20]X X1.7626.14.751.7042.62.780.58
PC-AVS [65]X X7.3514.48.210.1235.48.420.15
MakeItTalk [66]✓ X7.3519.52.030.7240.22.160.71
Audio2head [54]✓ X7.351016.420.131046.650.10
Ours + AR✓ X7.3524.37.230.0249.77.050.04
Ours + Diff.✓ ✓7.3514.27.340.0035.07.310.01
+ +Table 1. The quantitative results of synchronization and image quality under self-reenactment scenario. Bold means the best. A indicates the ability to generate non-lip motions with audio-only signals without other signals, while S means the ability to sample different results for each round. + +![](images/fa66ce24ff507467591c54de0277c266fe026e082abfdfd957815d6a51ee5d9f.jpg) +(a) + +![](images/169ffe2b44a4265ecc6b4f53f1dd08d6067f130267ffffd8e4f97d357b3d87c0.jpg) +(b) + +![](images/50344fa4daa4e9c0bbd39cf47f316cf198d205962fde5ad2713c8f3585d41199.jpg) +(c) +Figure 4. Distribution visualization of generated poses. + +audio input. We only compare those methods on lip synchronization and image quality but emphasize that our setup is targeting a more practical scenario. Other works such as Wav2Lip [38] only focus on revising the lip region without touching other facial motions, which are not directly comparable either. MakeItTalk [66] and Audio2head [54] are two closely related works to ours in terms of problem setup, however, our method is designed to predict all the non-lip facial motions with diversity, rather than only limited poses. + +Implementation Details Backbones of our models including $\mathrm{E}_{id}$ , $\mathrm{E}_v$ , $\mathrm{E}_a$ , $\mathrm{G}$ , and $\mathrm{D}$ are borrowed from [65]. For the diffusion prior network $\mathrm{P}_{a2nl}$ , we use a 8-layer transformer [51] with 512-d tokens and 1024-d fully feedforward layers. Note that our models were trained on VoxCeleb2 only but tested on both VoxCeleb1 and VoxCeleb2. All models were trained on 4 NVIDIA A100 GPUs. + +# 4.1. Quantitative Evaluation + +Evaluation Metrics We evaluate the performance of generated talking head from the following aspects: image quality, audio-lip sync accuracy, the variation and naturalness of non-lip facial motion. Frechet Inception Distance (FID) score [14] is used for the evaluation of image quality. A lower FID score indicates a lower distance between the distribution of generated images and real images. Following prior works [38, 65], we use SyncNet Error-Confidence $(\mathbf{S}_c)$ as the indicator of audio-lip synchronization, where greater confidence indicates better synchronization. However, the + +normalized confidence score NLSE-C (short for $\mathbf{N}_c$ ) is used to address the concern raised in recent works [57, 53] regarding the strong relationship between $\mathbf{S}_c$ and its training data, which can make it unfair to compare methods trained on different data. $\mathbf{N}_c$ is defined as $\mathbf{N}_c = \frac{\left|\mathbf{S}_c^{gen} - \mathbf{S}_c^{gt}\right|}{\mathbf{S}_c^{gt}}$ , where $\mathbf{S}_c^{gen}$ is the confidence value of generated images and $\mathbf{S}_c^{gt}$ is that of its training data. + +To evaluate facial motions such as pose, expression, and blink, we utilize a pre-trained 3D morphable face model [9] and include shape irrelevant 3DMM parameters to calculate the following metrics. + +- Var: The variance of generated facial motions, i.e., the variance of the 3DMM coefficients for each video is calculated and then averaged over the test set. A closer Var to GT indicates a better match to the variation of real data. +- $\mathbf{FID}_{fm}$ : FID score of 3DMM coefficients calculated as follows: $\mathbf{FID}_{fm} = \frac{1}{K}\sum_{i=1}^{K}\mathbf{FID}(\beta_i)$ , where $\beta_i$ is a sequence of 3DMM coefficients in the $i$ -th video. +- $\mathbf{FID}_{\Delta fm}$ : FID score of 3DMM coefficient difference between consecutive frames, i.e., $\mathbf{FID}_{\Delta fm}$ , which is similar to $\mathbf{FID}_{fm}$ but the 3DMM coefficient difference between consecutive frames is measured, taking temporal naturalness into consideration. +- SND: Our new proposed metric, denoted as Sequence Naturalness Distance. It is the sum of $\mathbf{FID}_{fm}$ and $\mathbf{FID}_{\Delta fm}$ indicating the difference of distribution between generated motion and GT motion, from both spatial and temporal perspectives, i.e., $\mathbf{SND} = \mathbf{FID}_{fm} + \mathbf{FID}_{\Delta fm}$ . The lower SND, the better naturalness. + +Evaluation Results Our main results consist of two parts: audio-lip synchronization and image quality as shown in Table 1; richness and sequence naturalness which are shown in Table 2. + +It shows that our method with diffusion prior archives the best synchronization ability compared to other methods in Table 1. Note that although Wav2Lip [38] and PC-AVS [65] achieve the highest SyncNet scores, our best $N_{c}$ indicates the most substantial synchronization ability [57], which is demonstrated in Sec. 4.2. Meanwhile, the image quality of our model with diffusion prior is the best in both test sets. + +In Table 2, our method shows significantly higher variance than the other two audio-driven methods and it is closer to the variance of real data. This indicates that our method can produce reasonable diverse head movements and expressions, rather than slight head movement. It's notable that auto-regressive prior, trained with causal attention and regression loss, has a higher variance than GT because it often generates extreme motions, showing a low naturalness score in Table 3. For naturalness, our method achieves the lowest $\mathbf{FID}_{fm}$ , $\mathbf{FID}_{\Delta fm}$ as well as SND on + +
MethodVoxCeleb2VoxCeleb1
Var →FIDfm ↓FIDΔfm ↓SND ↓Var →FIDfm ↓FIDΔfm ↓SND ↓
GT1.98---1.88---
MakeItTalk [66]0.674.701.746.440.804.271.075.34
Audio2head [54]0.895.941.307.240.765.031.016.04
Ours + AR3.075.432.227.652.926.211.687.89
Ours + Diff.1.573.601.084.681.663.980.874.85
w/o Lvel2.094.021.275.292.284.651.115.76
training w/o editing3.096.762.008.763.066.961.708.66
+ +Table 2. The quantitative results of variance and naturalness of generated non-lip motions under self-reenactment scenario. $\rightarrow$ indicates a closer score to GT is better. Due to setting differences, i.e., Wav2Lip, EAMM, and PC-AVS, are not valid comparisons here. + +![](images/14c2cd94afeb0438c0797a72a0544111867c8871db3289bc32b5799eb11c448c.jpg) +Figure 5. Qualitative results of our method compared to other baselines. Each row shows nine uniformly sampled frames from videos. Here we use two audio sources to drive different identities respectively. Our method shows accurate lip-audio synchronization with diverse and natural poses and expressions. + +both VoxCeleb1 and VoxCeleb2, surpassing other state-of-the-art by a large margin. Also, we randomly sample 5,000 poses for each method and employ t-SNE [50] for visual- + +ization in Fig. 4, which demonstrates that the distribution we generate is the most comprehensive to GT. More details and examples are shown in the supplementary. In summary, + +
MethodSync AccuracyNaturalnessRichness
GT4.614.414.18
MakeItTalk [66]1.972.352.10
PC-AVS [65]*3.132.682.68
Audio2head [54]2.562.352.34
Ours + AR3.462.533.37
Ours + Diff.4.083.823.68
+ +Table 3. Human evaluation on generated samples on VoxCeleb2. * means that PC-AVS was evaluated on cross-reenactment scenario. + +our method can generate rich facial motions with excellent realism (i.e., a lower distribution distance) and naturalness. + +# 4.2. Qualitative Evaluation + +Two audio-driven video sequences of our method verses other baselines are shown in Fig. 5. It's clearly shown that our method is capable of producing much more natural-looking and diverse facial motions than other baselines, including pose, expression, blink and gaze, while maintaining accurate audio-lip synchronization compared to the GT. + +User Study We invited 20 subjects for human evaluations, focusing on three aspects: 1) the accuracy of audio-lip synchronization; 2) the naturalness of non-lip motions and coherency between non-lip motions and audios; 3) the richness of non-lip motions. 50 videos are generated from audio and identity in the test set using the following methods: our diffusion prior, our auto-regressive prior, MakeItTalk [66], Audio2head [54] and PC-AVS [65] (driven by poses from another video). After shuffling the generated videos, each annotator rates from 1 (bad) to 5 (good) according to Mean Opinion Scores (MOS) rating protocol. Table 3 tells that our model with diffusion prior archives the best synchronization, naturalness and richness among all the models. Our model with auto-regressive prior archives relatively good richness compared to the remaining models, we attribute it to the exposure bias problem [4] brought by auto-regressive generation, which leads to high richness but low naturalness. Note that PC-AVS shows incompatible result in Table 3 as compared to Table 1 on $\mathbf{S}_c$ , indicating that choosing pose driving signals from a random video may lead to synthesis results with bad audio-lip synchronization. + +It's interesting that our proposed SND scores reflect the naturalness of user study to some degree. Nevertheless, we leave it to future work for rigorous study of their correlation through more human evaluations. + +# 4.3. Ablation + +Table 4 discusses the performance of lip & non-lip disentanglement. While tested on lip only metrics without $L_{ol}$ , SyncNet scores get lower, which indicates that G tends to generate mouth movements with information extracted + +
MethodLip OnlyNon-lip Only
Sc↑Nc↓Bd↓Gd↓Ed↓Pd↓
GT7.35-----
w/o Lol2.840.610.00740.19600.05680.0021
w/o MB7.690.050.01120.03150.08750.0293
Ours7.480.030.00670.01630.05580.0019
+ +Table 4. Evaluation of disentanglement between lip and non-lip on VoxCeleb2. $\mathbf{B}_d$ and $\mathrm{G}_d$ denote L2 distances of 2D landmark [58] of eyes and iris, respectively, while $\mathrm{E}_d$ and $\mathrm{P}_d$ denote L2 distances of 3DMM coefficients of poses and expressions accordingly. Note that $\mathbf{f}_{nl}^v$ is fixed while tested on lip only metrics, and $\mathbf{f}_l^a$ is fixed while tested on non-lip only metrics for fair comparisons. + +from visual modality instead of audio. Thus, lip & non-lip space are not fully disentangled. Without MB, the performance of driving non-lip motions gets worse. Because it is hard to use a large batch to compute correlation, resulting in a not well decoupled non-lip space. + +We conducted ablation study for $\mathrm{P}_{a2nl}$ from three aspects: 1) auto-regressive prior versus diffusion prior; 2) with/without velocity loss $L_{vel}$ ; 3) training with/without editing mechanism. Diffusion prior shows significant improvements on naturalness scores including $\mathrm{FID}_{fm}$ , $\mathrm{FID}_{\Delta fm}$ and SND, as compared to autoregressive prior shown in Table 2, with slightly better synchronization ability in Table 1. Fig. 6 shows that without $L_{vel}$ , the prediction of non-autoregressive diffusion prior model $\mathrm{P}_{a2nl}$ becomes unstable, which leads to a larger variance and worse SND scores in Table 2. While training without the editing mechanism mentioned in Sec. 3.2, we applied editing only during sampling as in MDM [46], and observed that it jitters between adjacent frames as shown in Fig. 7, resulting in a poorer SND score in Table 2. Note that our model trained with editing mechanism generates smoother results. + +![](images/a2ca78a9ffd31ae9d3d234671fd3a8dcb4774cc72a0df01d974879acd3986791.jpg) +Figure 6. Ablation study of our diffusion prior $\mathrm{P}_{a2nl}$ w or w/o $L_{vel}$ . Each row shows five uniformly sampled frames. + +# 5. Conclusions + +In this paper, we introduce a novel talking head generation method based on a diffusion prior model, which can generate diverse and natural-looking talking head videos with only audio and an identity image as inputs. Such "audio-driving is all you need" setting is very friendly + +![](images/732e1e805223738e2c730ac7ebaeea552c1544f45af3394be6d22a32e1367d87.jpg) +Figure 7. Ablation study of our diffusion prior $\mathbf{P}_{a2nl}$ trained w or w/o editing. Each row shows five adjacent frame. + +for reenactment and dubbing applications. Comprehensive evaluations including our newly proposed metrics validated the effectiveness of our system. + +Limitations Our method mainly focuses on non-lip motion generation. We leave the improvement of rendering quality to future works. + +# References + +[1] Ahmed A Abdelrahman, Thorsten Hempel, Aly Khalifa, and Ayoub Al-Hamadi. L2cs-net: Fine-grained gaze estimation in unconstrained environments. arXiv preprint arXiv:2203.03339, 2022. 4 +[2] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelović, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-Supervised MultiModal Versatile Networks. In NeurIPS, 2020. 3 +[3] Mohammed M Alghamdi, He Wang, Andrew J Bulpitt, and David C Hogg. Talking head from speech audio using a pretrained image generator. arXiv preprint arXiv:2209.04252, 2022. 2, 3 +[4] Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems, 28, 2015. 8 +[5] Stella Bounareli, Vasileios Argyriou, and Georgios Tzimiropoulos. Finding directions in gan's latent space for neural face reenactment. arXiv preprint arXiv:2202.00046, 2022.3 +[6] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 3 +[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 5 +[8] Soo-Whan Chung, Joon Son Chung, and Hong-Goo Kang. Perfect match: Improved cross-modal embeddings for audiovisual synchronisation. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3965-3969. IEEE, 2019. 2 +[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with + +weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 6 +[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3 +[11] Aviv Gabbay, Asaph Shamir, and Shmuel Peleg. Visual speech enhancement. arXiv preprint arXiv:1711.08789, 2017.2 +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 4 +[13] Yudong Guo, Keyu Chen, Sen Liang, Yongjin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3 +[16] Fangzhou Hong, Liang Pan, Zhongang Cai, and Ziwei Liu. Versatile multi-modal pre-training for human-centric perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16156-16166, 2022. 3 +[17] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3397-3406, 2022. 3 +[18] Gee-Sern Hsu, Chun-Hung Tsai, and Hung-Yi Wu. Dual-generator face reenactment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 642-650, 2022. 3 +[19] Po-Hsiang Huang, Fu-En Yang, and Yu-Chiang Frank Wang. Learning identity-invariant motion representations for crossid face reenactment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7084–7092, 2020. 3 +[20] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. arXiv preprint arXiv:2205.15278, 2022. 5, 6 +[21] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3 +[22] Venkatesh S Kadandale, Juan F Montesinos, and Gloria Haro. Vocalist: An audio-visual synchronisation model for lips and voices. arXiv preprint arXiv:2204.02090, 2022. 2 + +[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 4 +[24] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Nießner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics, 37(4):163:1-14, Aug. 2018. 3 +[25] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 3 +[26] Jiyoung Lee, Soo-Whan Chung, Sunok Kim, Hong-Goo Kang, and Kwanghoon Sohn. Looking into your speech: Learning cross-modal affinity for audio-visual speech separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1336–1345, 2021. 2 +[27] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3387-3396, June 2022. 2, 3 +[28] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 5 +[29] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2, 3 +[30] Moustafa Meshry, Saksham Suri, Larry S Davis, and Abhnav Shrivastava. Learned spatial representations for few-shot talking-head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13829-13838, 2021. 3 +[31] Dongchan Min, Minyoung Song, and Sung Ju Hwang. Stylistalker: One-shot style-based audio-driven talking head video generation. preprint arXiv:2208.10922, 2022. 3 +[32] Gaurav Mittal and Baoyuan Wang. Animating face using disentangled audio representations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), March 2020. 2 +[33] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. Voxceleb: a large-scale speaker identification dataset. arXiv preprint arXiv:1706.08612, 2017. 5 +[34] Arsha Nagrani, Joon son Chung, Samuel Albanie, and Andrew Zisserman. Disentangled speech embeddings using cross-modal self-supervision, 2020. 2 +[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3, 5 + +[36] Andrew Zisserman Olivia Wiles, A. Sophia Koepke. X2face: A network for controlling face generation by using images, audio, and pose codes. In ECCV 2018, 2018. 3 +[37] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In 36th AAAI Conference on Artificial Intelligence (AAAI 22). Association for the Advancement of Artificial Intelligence, 2022. 3 +[38] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2, 3, 6 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3 +[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3, 4 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3 +[42] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. arXiv preprint arXiv:2207.11770, 2022. 2 +[43] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3 +[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3 +[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3,4 +[46] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 3, 4, 5, 8 +[47] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. ECCV 2020, 2020. 3 +[48] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 3 + +[49] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. arXiv preprint arXiv:1906.05849, 2019. 2 +[50] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7 +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5, 6 +[52] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. CoRR, abs/1906.06337, 2019. 3 +[53] Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang. Progressive disentangled representation learning for fine-grained controllable talking head synthesis. arXiv preprint arXiv:2211.14506, 2022. 6 +[54] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2, 6, 7, 8 +[55] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2531–2539, 2022. 3 +[56] Chao Xu, Jiangning Zhang, Yue Han, Guanzhong Tian, Xianfang Zeng, Ying Tai, Yabiao Wang, Chengjie Wang, and Yong Liu. Designing one unified framework for high-fidelity face reenactment and swapping. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 54-71. Springer, 2022. 3 +[57] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022. 6 +[58] Baosheng Yu and Dacheng Tao. Heatmap regression via randomized rounding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 8 +[59] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 3 +[60] Xianfang Zeng, Yusu Pan, Mengmeng Wang, Jiangning Zhang, and Yong Liu. Realistic face reenactment via self-supervised disentangling of identity and pose. pages 12757-12764. AAAI Press, 2020. 3 +[61] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3867-3876, 2021. 2, 3 +[62] Jiangning Zhang, Liang Liu, Zhucun Xue, and Yong Liu. Apb2face: Audio-guided face reenactment with auxiliary pose and blink signals. 2020. 3 +[63] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with + +a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661-3670, 2021. 3 +[64] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9299–9306, 2019. 2 +[65] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2, 3, 4, 5, 6, 8 +[66] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 2, 3, 6, 7, 8 \ No newline at end of file diff --git a/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/images.zip b/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..12973d7fd21ec9ddb992fdb44af70a95325449a8 --- /dev/null +++ b/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70d4981d69eaee6a3a5aba38c8bf5633d7aaad2d40935112fb522adf1ad48bcc +size 647445 diff --git a/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/layout.json b/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c03948f54a8af1a615d40cfed9e528427b844333 --- /dev/null +++ b/2023/Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors/layout.json @@ -0,0 +1,10608 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 534, + 121 + ], + "type": "text", + "content": "Talking Head Generation with Probabilistic Audio-to-Visual Diffusion Priors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "text", + "content": "Zhentao Yu" + }, + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "text", + "content": " Zixin Yin" + }, + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{1,2*†}" + }, + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "text", + "content": " Deyu Zhou" + }, + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{1,3*†}" + }, + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "text", + "content": " Duomin Wang" + }, + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "text", + "content": " Finn Wong" + }, + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "text", + "content": " Baoyuan Wang" + }, + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{1‡}" + }, + { + "bbox": [ + 132, + 141, + 459, + 185 + ], + "type": "text", + "content": " Xiaobing.AI." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 185, + 465, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 185, + 465, + 213 + ], + "spans": [ + { + "bbox": [ + 126, + 185, + 465, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 185, + 465, + 213 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology. " + }, + { + "bbox": [ + 126, + 185, + 465, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 126, + 185, + 465, + 213 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology (Guangzhou)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 216, + 451, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 216, + 451, + 241 + ], + "spans": [ + { + "bbox": [ + 143, + 216, + 451, + 241 + ], + "type": "text", + "content": "(yuzhentao, wangduomin, wangwenlan, wangbaoyuan)@xiaobing.ai zyinaf@connect.ust.hk, dzhou861@connect.hkust-gz.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 268, + 545, + 434 + ], + "blocks": [ + { + "bbox": [ + 50, + 268, + 545, + 434 + ], + "lines": [ + { + "bbox": [ + 50, + 268, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 50, + 268, + 545, + 434 + ], + "type": "image", + "image_path": "bfba7cfa543a582cf526bb6a6de99c5ebfa6d99a1ed18db19c2e179cfdbc5791.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 451, + 547, + 485 + ], + "lines": [ + { + "bbox": [ + 46, + 451, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 547, + 485 + ], + "type": "text", + "content": "Figure 1. Given only an audio source and an arbitrary identity image, our system can generate a video with natural-looking and diverse facial motions (pose, expression, blink & gaze), while maintaining accurate audio-lip synchronization. Here we show randomly sampled sequences from our diffusion prior for two identities, note that the lip-irrelevant facial motion varies but the lip is still in-sync." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 499, + 192, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 499, + 192, + 511 + ], + "spans": [ + { + "bbox": [ + 143, + 499, + 192, + 511 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 524, + 288, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 524, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 524, + 288, + 668 + ], + "type": "text", + "content": "We introduce a novel framework for one-shot audio-driven talking head generation. Unlike prior works that require additional driving sources for controlled synthesis in a deterministic manner, we instead sample all holistic lip-irrelevant facial motions (i.e. pose, expression, blink, gaze, etc.) to semantically match the input audio while still maintaining both the photo-realism of audio-lip synchronization and overall naturalness. This is achieved by our newly proposed audio-to-visual diffusion prior trained on top of the mapping between audio and non-lip representations. Thanks to the probabilistic nature of the diffusion prior, one big advantage of our framework is it can synthe" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 500, + 545, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 500, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 306, + 500, + 545, + 608 + ], + "type": "text", + "content": "size diverse facial motion sequences given the same audio clip, which is quite user-friendly for many real applications. Through comprehensive evaluations of public benchmarks, we conclude that (1) our diffusion prior outperforms autoregressive prior significantly on all the concerned metrics; (2) our overall system is competitive with prior works in terms of audio-lip synchronization but can effectively sample rich and natural-looking lip-irrelevant facial motions while still semantically harmonized with the audio input." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 645, + 387, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 387, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 387, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 547, + 715 + ], + "type": "text", + "content": "Audio-driven face reenactment and talking head generation have received raising attention due to the broad killer applications in movie production, gaming, virtual digital avatars, and potentially more while we move toward the era" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 674, + 232, + 685 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 674, + 232, + 685 + ], + "spans": [ + { + "bbox": [ + 58, + 674, + 232, + 685 + ], + "type": "text", + "content": "*These authors have contributed equally to this work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 685, + 287, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 685, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 47, + 685, + 287, + 704 + ], + "type": "text", + "content": "This work was done when Zixin Yin and Deyu Zhou were interns at XiaoBing.AI." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 58, + 704, + 135, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 704, + 135, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 704, + 135, + 714 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 58, + 704, + 135, + 714 + ], + "type": "text", + "content": " Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7645" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 455 + ], + "type": "text", + "content": "of the metaverse. The past literature tries to advance this area from various perspectives. One line of research focuses on improving the generation quality originating from regular GAN-based methods [64, 65] to pre-trained StyleGAN [3] and to the most recent Neural Radiance Field (NeRF) based methods [13, 42]. Orthogonal to direction, another line of works emphasizes the importance of disentangled representation from the audio [32, 21, 66] for controlled generations. i.e., [21] can predict the emotion from the audio input while [66] can decouple the speech signal into speaker identity and the phonetic content. A closely related set of works targeting more granular controlled talking head generation by providing additional input signals. For example, PC-AVS [65] requires a separate video to provide the pose signal in addition to the audio clip but does not support the control of expression and blinking. To remedy this, GC-AVT [27] requires inputting additional driving video for expression in addition to pose and audio driving clip. The technical challenge behind those approaches is how to faithfully transfer the desired driving signal (i.e., pose, expression, audio) into the results without affecting each other through intrinsic disentanglement. Although encouraging results have been reported, we argue such a setting is not practical for broader applications. It is quite challenging, or at least labor-intensive, for novice users to find the \"best\" individual driving sources for each controlled dimension to make the final talking head video overall look not only lifelike but also coherent from semantic and emotional perspectives. Therefore, it is both more practical and generalizable if the setting only requires an audio signal as the driving source and expects a data-driven model to sample reasonable other facial motions irrelevant to lips." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 459, + 289, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 289, + 674 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 289, + 674 + ], + "type": "text", + "content": "By no means, we are the first to advocate an audio-only driving setup for talking head generation. Audio2head [54] predicts the poses from audio input with an LSTM network. [29] employs an autoregressive model by assuming the poses are jointly determined by the audio and past head motions along the sequence, although the results are encouraging, the model can't generalize due to their person-specific training strategy. FACIAL [61] could infer the blink, but requires a reference video input rather than a one-shot reference image. There are other related works along this direction, but they either only infer one facial attribute (i.e., emotion in EVP[21], pose from [54]) using ad-hoc methods, or they simply do not support inferring other facial motions [38], or treat it as a block-box mapping model [3] without explicitly respecting the one-to-many mapping nature between audio and other visual facial attributes (pose, expression, blink). A more principled solution is desired to consolidate this line of work." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "In this paper, we introduce a novel framework that can holistically infer all the non-lip-related facial attributes from the audio input while maintaining accurate synchronization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": "between audio and the corresponding visual lip motions. This is achieved by two important learning steps in our pipeline. (1) A pre-trained identity-irrelevant facial motion representation can help decouple the lip and non-lip representations. To promote this learning, we employ a novel orthogonal loss on top of the modified facial reenactment framework [6]. The disentanglement enables to generate one-to-one mapping with lip representations to ensure the synchronization, and generate richer non-lip representations with a one-to-many mapping. (2) A novel audio-to-visual diffusion prior model is introduced to address the probabilistic sampling from audio representations to the abovelearned non-lip representation. This prior is expected to solve the one-to-many mapping and provide diverse results during the inference stage. The entire pipeline can be easily built up on top of the existing framework, such as PC-AVS[65] without heavily retraining every component. To sum up, we make the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 296, + 545, + 550 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 296, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 296, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 296, + 545, + 380 + ], + "type": "text", + "content": "- To our best knowledge, we are the first to holistically predict non-lip facial motions based on audio input only, providing good usability for reenactment or dubbing applications without extra driving video sources. Our method addresses the intrinsic one-to-many challenge in a probabilistic way, allowing diverse and realistic facial motion generation under the same audio input, as shown in Fig. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 387, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 387, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 387, + 545, + 472 + ], + "type": "text", + "content": "- We leverage the pre-trained visual identity-irrelevant facial motion representations. And further, learn disentangled lip-related and lip-irrelevant representations through a novel orthogonal loss on top of PC-AVS[65]. A powerful diffusion prior model is then introduced to effectively infer all lip-irrelevant facial motions for a given audio segment in the representation space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 479, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 545, + 550 + ], + "type": "text", + "content": "- We systematically evaluate the naturalness and diversity of the results with new metrics, which paves a way for future studies. Meanwhile, results show that our method can produce natural-looking head poses and facial motions without hurting the audio-lip synchronization. Our model will be released." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 562, + 398, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 398, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 398, + 574 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Audio-Visual Cross-modal Learning Cross-modal representation learning is a long-standing research topic, ranging from speech enhancement [11], speech source separation [26] to synchronization [8, 22] and other speech disentanglement [34, 21, 32]. Among them, EVP [21] used cross-modal supervision to disentangle speech content and emotion from the audio signal with landmark as the intermediate representation. Recently, CMC [49] discussed how multiview \"modality\" can be jointly unitized to boost intrinsic representation through contrastive learning, rather than predictive (or reconstruction) learning, it also demonstrated" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "type": "text", + "content": "7646" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "that the more views, the better. Concurrently, MMV [2] introduced different modality embedding graphs for effective cross-modal representations, again through contrastive learning. More recently, HCMoCo [16] extended similar ideas with a hierarchical strategy to learn different levels of representations for human-centric perception tasks. Although impressive results were reported, it is still unclear how it performs on face analysis tasks, not mention to on synthesis tasks. Our work shares a similar spirit but a different purpose, where we first pre-train a non-identity visual representation which then helps learn a lip and non-lip space, the latter further serves as the upper-bound learning target for subsequent audio-to-visual diffusion prior." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 237, + 289, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 289, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 289, + 538 + ], + "type": "text", + "content": "Face Reenactment & Talking Head Generation Face Reenactment is designed to transfer part or full facial motion from a driving source to the target video with good IDpreserved appearance and background. It can be further divided into two categories depending on whether the driving source is from video [17, 18, 5, 48, 24, 36, 59, 19, 60, 6, 56] or audio [52, 47, 62, 21, 55, 37, 63, 30]. Among them, audio-driven face reenactment generally aims to edit the mouth regions of the target video in order to match the input audio while leaving other facial attributes mostly unchanged, i.e., pose. EVP [21] tries to infer the non-rigid facial expression in addition to lip motion from the audio input. A closely related line of work is the audio-driven talking-head generations [38, 66, 61, 3] where only one target reference face is given, hence, other face attributes including pose, expression, blink and gaze have to be either explicitly given [65, 27] or partly inferred through statistical methods [66, 31, 61, 29, 3, 47]. Specifically, Lu et al. [29] employed an auto-regressive model while Min et al. [31] leveraged normalized flow prior to predicting a natural-looking pose sequence from the input audio, both showed encouraging results. Compared with them, we aim to infer more diverse facial motions including pose, expression, and even blink and gaze in a holistic manner through an audio-to-visual diffusion prior model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "Diffusion Generative Models The diffusion model [44, 15, 45], which is a likelihood-based model consisting of cascading denoising autoencoders, has recently shown great success in numerous generative tasks with different modalities including image [10, 35, 40, 41], audio [25], video [43], and motion [46]. To name a few, DDPM [15] explored the diffusion model for unconditional image generation. GLIDE [35] introduced text-conditional diffusion model and showed that classifier free guidance has better performance than CLIP [39] guidance. DALLE-2 [40] modified GLIDE to generate semantically consistent images conditioned on a CLIP image embedding, and proposed a diffusion prior that produces the image embedding given a text caption. MDM [46] utilized a classifier-free diffusion-based" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "generative model for text-to-motion and action-to-motion tasks, allowing motion completion and editing as well." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 106, + 362, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 106, + 362, + 118 + ], + "spans": [ + { + "bbox": [ + 306, + 106, + 362, + 118 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "text", + "content": " segments of audio " + }, + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "inline_equation", + "content": "A_{1:L} = (A_1, A_2, \\dots, A_L)" + }, + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "text", + "content": " and a reference image " + }, + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "text", + "content": " as inputs, our model M synthesizes video frames " + }, + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "inline_equation", + "content": "\\hat{X}_{1:L} = (\\hat{X}_1, \\hat{X}_2, \\dots, \\hat{X}_L)" + }, + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "text", + "content": " having the same identity as " + }, + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "text", + "content": " and lip motion synchronized to " + }, + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "inline_equation", + "content": "A_{1:L}" + }, + { + "bbox": [ + 305, + 125, + 547, + 177 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 376, + 182, + 545, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 182, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 376, + 182, + 545, + 198 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {1: L} = \\mathbf {M} \\left(A _ {1: L}, I _ {r e f}\\right). \\tag {1}", + "image_path": "a9068ded19be4a34e75dd0fad60d612fe7f3204ed40d7130d7b0ce45f8d58bd6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 203, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 203, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 203, + 545, + 228 + ], + "type": "text", + "content": "An overview of our proposed framework is shown in Fig. 2, which consists of three major components:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 233, + 545, + 427 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 233, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 233, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 233, + 545, + 294 + ], + "type": "text", + "content": "- Lip & Non-lip Disentanglement Given a pre-trained identity- and appearance-irrelevant facial motion encoder " + }, + { + "bbox": [ + 306, + 233, + 545, + 294 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_v" + }, + { + "bbox": [ + 306, + 233, + 545, + 294 + ], + "type": "text", + "content": ", we first leverage it to learn two complementary features, " + }, + { + "bbox": [ + 306, + 233, + 545, + 294 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_l^a" + }, + { + "bbox": [ + 306, + 233, + 545, + 294 + ], + "type": "text", + "content": " for lip, and " + }, + { + "bbox": [ + 306, + 233, + 545, + 294 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{nl}^v" + }, + { + "bbox": [ + 306, + 233, + 545, + 294 + ], + "type": "text", + "content": " for non-lip features including pose, expression, blink and gaze." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 300, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 545, + 361 + ], + "type": "text", + "content": "- Audio to Non-lip Diffusion Prior A diffusion prior network " + }, + { + "bbox": [ + 306, + 300, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 306, + 300, + 545, + 361 + ], + "type": "text", + "content": " models the one-to-many mapping from audio feature " + }, + { + "bbox": [ + 306, + 300, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^a" + }, + { + "bbox": [ + 306, + 300, + 545, + 361 + ], + "type": "text", + "content": " to the \"hallucinated\" non-lip feature " + }, + { + "bbox": [ + 306, + 300, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{nl}^a" + }, + { + "bbox": [ + 306, + 300, + 545, + 361 + ], + "type": "text", + "content": ", which is trained to be close to its visual counterpart " + }, + { + "bbox": [ + 306, + 300, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{nl}^v" + }, + { + "bbox": [ + 306, + 300, + 545, + 361 + ], + "type": "text", + "content": ", allowing audio-only facial driving at inference time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 367, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 367, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 306, + 367, + 545, + 427 + ], + "type": "text", + "content": "- Audio-based Talking Head Generation Given an identity feature " + }, + { + "bbox": [ + 306, + 367, + 545, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{id}" + }, + { + "bbox": [ + 306, + 367, + 545, + 427 + ], + "type": "text", + "content": ", a lip feature " + }, + { + "bbox": [ + 306, + 367, + 545, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_l^a" + }, + { + "bbox": [ + 306, + 367, + 545, + 427 + ], + "type": "text", + "content": ", and a non-lip feature " + }, + { + "bbox": [ + 306, + 367, + 545, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{nl}^a" + }, + { + "bbox": [ + 306, + 367, + 545, + 427 + ], + "type": "text", + "content": " generated with the diffusion prior, we concatenate them together and feed into a GAN similar to LPD [6] or PC-AVS [65] to output the final reenacted video." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 432, + 477, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 477, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 477, + 445 + ], + "type": "text", + "content": "3.1. Lip & Non-lip Disentanglement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 450, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 546 + ], + "type": "text", + "content": "The identity encoder " + }, + { + "bbox": [ + 304, + 450, + 545, + 546 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{id}" + }, + { + "bbox": [ + 304, + 450, + 545, + 546 + ], + "type": "text", + "content": " is designed to capture identity and appearance information, while " + }, + { + "bbox": [ + 304, + 450, + 545, + 546 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_v" + }, + { + "bbox": [ + 304, + 450, + 545, + 546 + ], + "type": "text", + "content": " is instead to remove both while encoding all facial motions. Both encoders are pre-trained in a similar setting as LPD [6] and PC-AVS [65]. After pretraining, we conduct lip & non-lip disentanglement consisting of audio-visual contrastive learning for lip feature learning and decorrelation for non-lip feature learning, as well as face reconstruction, as described below:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": "Audio-Visual Pretraining for Audio Lip Space An audio encoder " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_a" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": " is trained to provide accurate audio feature " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^a = \\mathrm{E}_a(A_{1:N})" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": " through contrastive learning [39] with " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{cl}^v = \\mathrm{MLP}_{cl}(\\mathrm{E}_v(X_{1:N}))" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": " is the number of samples and " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "X_{1:N}" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": " frames in the same video corresponding to audio " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "A_{1:N}" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^a" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": " is further compressed to filter out lip-relevant information and get audio lip feature " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_l^a" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{a2l}" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": " since " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^a" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": " mainly contains lip related information, as shown in Sec. 4.3. The contrastive loss is defined as " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "L_{con}(m,n) = -\\frac{1}{N}\\sum_{i=1}^{N}\\log\\frac{\\exp(m_i/\\|m_i\\|_2 \\cdot n_i/\\|n_i\\|_2)}{\\sum_{j=1}^{N}\\exp(m_i/\\|m_i\\|_2 \\cdot n_j/\\|n_j\\|_2)}" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": ", resulting in a total contrastive loss between " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^a" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{cl}^v" + }, + { + "bbox": [ + 304, + 552, + 545, + 689 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 347, + 693, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 693, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 347, + 693, + 545, + 717 + ], + "type": "interline_equation", + "content": "L _ {c l} = \\frac {1}{2} \\left[ L _ {\\text {c o n}} \\left(\\mathbf {f} ^ {a}, \\mathbf {f} _ {c l} ^ {v}\\right) + L _ {\\text {c o n}} \\left(\\mathbf {f} _ {c l} ^ {v}, \\mathbf {f} ^ {a}\\right) \\right]. \\tag {2}", + "image_path": "bbf314241bc72ca155ba5d553b846c213be46b30e4e17e0bd75eeb083bfa837c.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7647" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 75, + 544, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 75, + 544, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 75, + 544, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 75, + 544, + 228 + ], + "type": "image", + "image_path": "f1a80a4f86f0080f1fe4c303c1bdb2d75be6d9fd455043f7e18b5a635bc1e1c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "text", + "content": "Figure 2. The overall pipeline of our proposed framework. The dotted lines represent the loss functions that are used only in training, i.e., " + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "inline_equation", + "content": "L_{cl}" + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "inline_equation", + "content": "L_{ol}" + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "inline_equation", + "content": "L_{a2nl}" + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "inline_equation", + "content": "L_{G}" + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "text", + "content": " is the training loss for visual non-lip space. Note that there is a switch in the figure, which means the different forward processes in training and inference: in training, the concatenated feature " + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{cat} = cat(\\mathbf{f}_{id},\\mathbf{f}_{nl}^{v},\\mathbf{f}_{l}^{a})" + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "text", + "content": "; at inference stage, " + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{cat} = cat(\\mathbf{f}_{id},\\mathbf{f}_{nl}^{a},\\mathbf{f}_{l}^{a})" + }, + { + "bbox": [ + 46, + 229, + 548, + 275 + ], + "type": "text", + "content": ". Thus, the visual motion part is not needed anymore at the inference stage." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "text", + "content": "Reconstruction Learning for Visual Non-lip Space To learn a good non-lip space, we propose an orthogonal loss to penalize the correlation between " + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ol}^{v}" + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{\\alpha}" + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "text", + "content": ", which uses the well-learned audio space to disentangle lip-related motions. The orthogonal loss is used together with reconstruction since there should also be a completeness constraint to avoid mode collapse of lip-irrelevant features. In practice, we maintain two memory banks (MB) for storing " + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ol}^{v}" + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{\\alpha}" + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "text", + "content": " in previous " + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "text", + "content": " steps, in order to have more samples than " + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "text", + "content": ". We denote the feature dimension of " + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{\\alpha}" + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "inline_equation", + "content": "d_{a}" + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ol}^{v}" + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "inline_equation", + "content": "d_{v}" + }, + { + "bbox": [ + 46, + 311, + 289, + 443 + ], + "type": "text", + "content": ". The orthogonal loss is defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 447, + 287, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 447, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 94, + 447, + 287, + 480 + ], + "type": "interline_equation", + "content": "L _ {o l} = \\frac {1}{d _ {v}} \\sum_ {i = 1} ^ {d _ {v}} \\sum_ {j = 1} ^ {d _ {a}} P _ {c o r} \\left(\\mathbf {f} _ {o l} ^ {v}, \\mathbf {f} ^ {a}\\right) _ {(i, j)} ^ {2}, \\tag {3}", + "image_path": "8581c91470e97bb4f324f22b3de548d4fdec6736c200a1626545c24fae98be04.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "inline_equation", + "content": "P_{cor}(\\mathbf{f}_{ol}^v,\\mathbf{f}^a)\\in d_a\\times d_v" + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "content": " computes the Pearson correlation coefficient between " + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ol}^{v}" + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^a" + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "content": ". Since the correlation between lip motions and audio features is strong, " + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "inline_equation", + "content": "L_{ol}" + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "content": " (which encourages the linear uncorrelation between " + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ol}^{v}" + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^a" + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "content": ") is effective to disentangle lip-related motions. Then, " + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_{nl}" + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "content": " further projects " + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ol}^{v}" + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "content": " into a smaller space to remove undesired content and get non-lip feature " + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{nl}^{v}" + }, + { + "bbox": [ + 46, + 486, + 287, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "We follow the reconstruction loss " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{GAN}}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "L_{L1}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{VGG}}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " used in [65]. The generator " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " receives the concatenation " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{cat} = cat(\\mathbf{f}_{id}, \\mathbf{f}_{nl}^{v}, \\mathbf{f}_{l}^{a})" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " as inputs and generates final images " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "I_{gen}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " with modulated convolution. The discriminator " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " is trained jointly by generative adversarial learning [23, 12]. Note that we omit the batch-mean operation here for convenience, but in practice, all losses used here are calculated with the average in a batch. Different from [65], we propose an additional gaze loss " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "L_{gaze}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " by utilizing a pre-trained gaze encoder [1], which calculates the " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " distance between the gaze features of generated images and ground-truth (GT) images " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 364, + 305, + 546, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 305, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 364, + 305, + 546, + 319 + ], + "type": "interline_equation", + "content": "L _ {g a z e} = \\left\\| \\Phi (I) - \\Phi \\left(I _ {g e n}\\right) \\right\\| _ {1}. \\tag {4}", + "image_path": "acbf1b146f2d8d6b0d27cba5e0b265f0d7fc05634f10bf0a99640135b503faba.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 321, + 440, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 321, + 440, + 333 + ], + "spans": [ + { + "bbox": [ + 318, + 321, + 440, + 333 + ], + "type": "text", + "content": "The total loss " + }, + { + "bbox": [ + 318, + 321, + 440, + 333 + ], + "type": "inline_equation", + "content": "L_{G}" + }, + { + "bbox": [ + 318, + 321, + 440, + 333 + ], + "type": "text", + "content": " is given as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 335, + 338, + 546, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 338, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 335, + 338, + 546, + 365 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {G} = \\lambda_ {o l} \\cdot L _ {o l} + \\lambda_ {g a z e} \\cdot L _ {g a z e} + \\lambda_ {L 1} \\cdot L _ {L 1} \\tag {5} \\\\ + \\lambda_ {\\text {G A N}} \\cdot L _ {\\text {G A N}} + \\lambda_ {\\text {V G G}} \\cdot L _ {\\text {V G G}}, \\\\ \\end{array}", + "image_path": "c425085424d404e57cf11b891eadd0666e110267856ce5c80f174674b461da2a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 369, + 440, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 440, + 381 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 440, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 369, + 440, + 381 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 369, + 440, + 381 + ], + "type": "text", + "content": " s are weights for losses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 387, + 466, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 387, + 466, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 387, + 466, + 399 + ], + "type": "text", + "content": "3.2. Audio2nonlip Diffusion Prior" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 405, + 546, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 546, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 546, + 502 + ], + "type": "text", + "content": "Beside a strong relationship with lip motions, audio features have a complex non-linear relationship with non-lip motions and it is a one-to-many mapping problem. This is due to the fact that there are many reasonable facial motions that can be corresponded to the same audio input. As inspired by DALLE-2 [40] and MDM [46], we design an audio2nonlip diffusion prior network " + }, + { + "bbox": [ + 304, + 405, + 546, + 502 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 304, + 405, + 546, + 502 + ], + "type": "text", + "content": " to address this problem. Our proposed network is depicted in Fig. 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 506, + 547, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 541 + ], + "type": "text", + "content": "Diffusion Model Diffusion model is designed based on the stochastic diffusion process. The forward diffusion process is defined below:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 545, + 545, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 545, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 321, + 545, + 545, + 561 + ], + "type": "interline_equation", + "content": "q \\left(n _ {1: L} ^ {t} \\mid n _ {1: L} ^ {t - 1}\\right) = \\mathcal {N} \\left(n _ {1: L} ^ {t - 1}; \\sqrt {\\alpha^ {t}} n _ {1: L} ^ {t - 1}, (1 - \\alpha^ {t}) I\\right), \\tag {6}", + "image_path": "b23b86a2c92bc237476accbf101010b0fdb3e48f7032eb07ba2b37beb74454df.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "text", + "content": "In our context, " + }, + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "inline_equation", + "content": "n_{1:L} = (n_1,n_2,\\dots,n_L)" + }, + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "text", + "content": " is a sequence of non-lip feature " + }, + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{nl}^v" + }, + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "inline_equation", + "content": "\\alpha^t" + }, + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "text", + "content": " is a hyper-parameter and " + }, + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "inline_equation", + "content": "t\\sim [1,T]" + }, + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "text", + "content": " is the time step of the diffusion process. Eq. 6 can be approximated to " + }, + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "inline_equation", + "content": "n_{1:L}^{T}\\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "inline_equation", + "content": "\\alpha^t" + }, + { + "bbox": [ + 304, + 564, + 547, + 613 + ], + "type": "text", + "content": " is small enough." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 613, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 547, + 685 + ], + "type": "text", + "content": "During the reversed diffusion process, " + }, + { + "bbox": [ + 304, + 613, + 547, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 304, + 613, + 547, + 685 + ], + "type": "text", + "content": " models the audio2nonlip distribution as " + }, + { + "bbox": [ + 304, + 613, + 547, + 685 + ], + "type": "inline_equation", + "content": "p(n_{1:L}^{0}|a_{1:L})" + }, + { + "bbox": [ + 304, + 613, + 547, + 685 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 613, + 547, + 685 + ], + "type": "inline_equation", + "content": "a_{1:L} = (a_1,a_2,\\dots,a_L)" + }, + { + "bbox": [ + 304, + 613, + 547, + 685 + ], + "type": "text", + "content": " is a sequence of audio feature " + }, + { + "bbox": [ + 304, + 613, + 547, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^a" + }, + { + "bbox": [ + 304, + 613, + 547, + 685 + ], + "type": "text", + "content": ". Instead of predicting the noise as formulated by vanilla DDPM [45], " + }, + { + "bbox": [ + 304, + 613, + 547, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 304, + 613, + 547, + 685 + ], + "type": "text", + "content": " learns to predict the initial signal itself with the simplified objective function [45] described as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 309, + 687, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 687, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 309, + 687, + 545, + 713 + ], + "type": "interline_equation", + "content": "L _ {s i m p l e} = \\mathbb {E} _ {n ^ {0} \\sim q (n ^ {0} | \\mathbf {f} ^ {\\alpha}), t \\sim [ 1, T ]} \\left[ \\left\\| n ^ {0} - \\mathrm {P} _ {a 2 n l} \\left(n ^ {t}, t, \\mathbf {f} ^ {\\alpha}\\right) \\right\\| _ {2} \\right], \\tag {7}", + "image_path": "089564622f2c6097101ed268b6610b09a1d5cf6304ebf6bb8ec13ce15b0409b6.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7648" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 272, + 230 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 272, + 230 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 272, + 230 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 272, + 230 + ], + "type": "image", + "image_path": "942fce6204997a8e2459c4065a7ffff1084e5a0fbfa0f70974050a23eda17ad9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 233, + 261, + 245 + ], + "lines": [ + { + "bbox": [ + 47, + 233, + 261, + 245 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 261, + 245 + ], + "type": "text", + "content": "Figure 3. Our audio2nonlip diffusion prior network " + }, + { + "bbox": [ + 47, + 233, + 261, + 245 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 47, + 233, + 261, + 245 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 266, + 255, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 255, + 279 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 255, + 279 + ], + "type": "text", + "content": "where we use " + }, + { + "bbox": [ + 47, + 266, + 255, + 279 + ], + "type": "inline_equation", + "content": "n^0" + }, + { + "bbox": [ + 47, + 266, + 255, + 279 + ], + "type": "text", + "content": " to represent " + }, + { + "bbox": [ + 47, + 266, + 255, + 279 + ], + "type": "inline_equation", + "content": "n_{1:L}^{0}" + }, + { + "bbox": [ + 47, + 266, + 255, + 279 + ], + "type": "text", + "content": " for convenience." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "text", + "content": "We implement " + }, + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "text", + "content": " with a transformer encoder [51] architecture. As shown in Fig. 3, an audio feature " + }, + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "inline_equation", + "content": "a_{1:L}" + }, + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "text", + "content": " is added to the time embedding " + }, + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "inline_equation", + "content": "\\mathrm{Emb}_{\\text{time}}" + }, + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "text", + "content": ". Then, it's concatenated with noisy non-lip feature " + }, + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "inline_equation", + "content": "n_{1:L}^{t}" + }, + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "text", + "content": " and added with positional embeddings " + }, + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "inline_equation", + "content": "\\mathrm{Emb}_{pos}" + }, + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "text", + "content": " [51]. The model encodes the embeddings with bidirectional self-attention and feedforward layers then output denoised non-lip feature " + }, + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "inline_equation", + "content": "n_{1:L}^{0}" + }, + { + "bbox": [ + 46, + 280, + 286, + 364 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 375, + 287, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 375, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 375, + 287, + 423 + ], + "type": "text", + "content": "Velocity Loss Note that our designed " + }, + { + "bbox": [ + 47, + 375, + 287, + 423 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 47, + 375, + 287, + 423 + ], + "type": "text", + "content": " denoises " + }, + { + "bbox": [ + 47, + 375, + 287, + 423 + ], + "type": "inline_equation", + "content": "n_{1:L}^{t}" + }, + { + "bbox": [ + 47, + 375, + 287, + 423 + ], + "type": "text", + "content": " in a non-autoregressive way. To encourage the naturalness and coherence of generated non-lip motion, we borrow the velocity loss from MDM [46] defined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 434, + 287, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 287, + 466 + ], + "type": "interline_equation", + "content": "L _ {v e l} = \\frac {1}{L - 1} \\sum_ {i = 2} ^ {L} \\left\\| \\left(\\hat {n} _ {i} ^ {0} - \\hat {n} _ {i - 1} ^ {0}\\right) - \\left(n _ {i} ^ {0} - n _ {i - 1} ^ {0}\\right) \\right\\| _ {2}, \\tag {8}", + "image_path": "abcbfed36105ddc9f59da2a887f1876e964d0ec7fc05a4d00ea0c49435edd0b1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 478, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 287, + 538 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 478, + 287, + 538 + ], + "type": "inline_equation", + "content": "n_i^0" + }, + { + "bbox": [ + 46, + 478, + 287, + 538 + ], + "type": "text", + "content": " is the GT non-lip feature and " + }, + { + "bbox": [ + 46, + 478, + 287, + 538 + ], + "type": "inline_equation", + "content": "\\hat{n}_i^0" + }, + { + "bbox": [ + 46, + 478, + 287, + 538 + ], + "type": "text", + "content": " is the predicted denoised non-lip feature at the " + }, + { + "bbox": [ + 46, + 478, + 287, + 538 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 478, + 287, + 538 + ], + "type": "text", + "content": "-th position respectively. The intuition here is that the difference between adjacent non-lip features should be close to the difference in the GT. The total loss of this stage is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 551, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 551, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 115, + 551, + 287, + 563 + ], + "type": "interline_equation", + "content": "L _ {a 2 n l} = L _ {\\text {s i m p l e}} + L _ {v e l}. \\tag {9}", + "image_path": "7c51927c2f7c244a44d7f64301136c118fcfe3d6b31ac40decca491530bc0e68.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "text", + "content": "Classifier-free Guidance We use classifier-free guidance [35] for conditioned diffusion generation. In training time, we randomly set the condition to " + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\emptyset" + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "text", + "content": " of the samples so that " + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}(n^t,t,\\emptyset)" + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "text", + "content": " approximates " + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "inline_equation", + "content": "n^0" + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "text", + "content": ". At inference stage, the output of the " + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "inline_equation", + "content": "P_{a2nl}" + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "text", + "content": " is extrapolated further in the direction of " + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}(n^t,t,\\mathbf{f}^a)" + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "text", + "content": " and away from " + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}(n^t,t,\\emptyset)" + }, + { + "bbox": [ + 46, + 586, + 287, + 659 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 670, + 286, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 286, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 286, + 693 + ], + "type": "interline_equation", + "content": "\\mathrm {P} _ {a 2 n l} \\left(n ^ {t}, t, \\mathbf {f} ^ {a}\\right) = s \\cdot \\mathrm {P} _ {a 2 n l} \\left(n ^ {t}, t, \\mathbf {f} ^ {a}\\right) + (1 - s) \\cdot \\mathrm {P} _ {a 2 n l} \\left(n ^ {t}, t, \\emptyset\\right), \\tag {10}", + "image_path": "a8ad37e612048a43287f4e6a3d469ff48823d75f860d7d0e2ce42a173254b978.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 696, + 287, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 696, + 287, + 719 + ], + "spans": [ + { + "bbox": [ + 47, + 696, + 287, + 719 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 696, + 287, + 719 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 696, + 287, + 719 + ], + "type": "text", + "content": " is a scaling parameter while increasing it improves sample quality at the cost of diversity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": "Sequential Mask Editing To deliver a smooth non-lip facial sequence with arbitrary length, we design a mechanism to ensure continuity between generated segments, and provide an editing method to do so. Specifically, we randomly mask " + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": " of the non-lip tokens " + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "inline_equation", + "content": "n^0" + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": " and concatenate them with noisy non-lip tokens " + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "inline_equation", + "content": "n^t" + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": " as the input of " + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": " at training stage. Practically, we found it easy for " + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": " to quickly learn how to fill in the masked region using hints from the non-lip features in the unmasked regions, generating continuous feature prediction without any extra design in training loss. As a result, it is possible to use " + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": " for audio-guided non-lip motion generation as well as non-lip motion editing at inference time, providing non-lip facial motion with good naturalism and diversity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 248, + 508, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 248, + 508, + 261 + ], + "spans": [ + { + "bbox": [ + 306, + 248, + 508, + 261 + ], + "type": "text", + "content": "3.3. Audio-based Talking Head Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 267, + 545, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 351 + ], + "type": "text", + "content": "The main difference at inference time is that non-lip visual feature " + }, + { + "bbox": [ + 304, + 267, + 545, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{nl}^{v}" + }, + { + "bbox": [ + 304, + 267, + 545, + 351 + ], + "type": "text", + "content": " is replaced with " + }, + { + "bbox": [ + 304, + 267, + 545, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{nl}^{a}" + }, + { + "bbox": [ + 304, + 267, + 545, + 351 + ], + "type": "text", + "content": " generated by " + }, + { + "bbox": [ + 304, + 267, + 545, + 351 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 304, + 267, + 545, + 351 + ], + "type": "text", + "content": " through reversed diffusion process. Thanks to the random noise introduced in the diffusion process, " + }, + { + "bbox": [ + 304, + 267, + 545, + 351 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 304, + 267, + 545, + 351 + ], + "type": "text", + "content": " can generate various visual non-lip features given the same audio input, resulting in diverse and reasonable reenactment videos without any extra needs of driving video sources." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 352, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 411 + ], + "type": "text", + "content": "To generate a video sequence longer than the maximal length of the model input with smooth transitions, the first input non-lip feature to " + }, + { + "bbox": [ + 304, + 352, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 304, + 352, + 545, + 411 + ], + "type": "text", + "content": " is set to the last generated non-lip feature from " + }, + { + "bbox": [ + 304, + 352, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 304, + 352, + 545, + 411 + ], + "type": "text", + "content": " in the previous step, to ensure continuity in non-lip facial motion." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 424, + 387, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 424, + 387, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 424, + 387, + 437 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 445, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 529 + ], + "type": "text", + "content": "Our proposed method was evaluated from two aspects: 1) synchronization between audio and lip motions; 2) naturalness and richness of lip-irrelevant facial motions. Both quantitative and qualitative experiments were conducted to showcase the superiority of our method. In addition, we also come up with a novel metric to measure the overall quality of generated facial motions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 536, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 573 + ], + "type": "text", + "content": "Datasets Our model was trained on VoxCeleb2 and evaluated on both VoxCeleb1 and VoxCeleb2. Here are the details of the two datasets," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 582, + 545, + 659 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 306, + 582, + 545, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 545, + 618 + ], + "type": "text", + "content": "- VoxCeleb1 [33]: The dataset consists of 100,000 utterances from 1,251 celebrities. 100 videos with 25 identities in total were randomly chosen for the test." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 623, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 545, + 659 + ], + "type": "text", + "content": "- VoxCeleb2 [7]: The dataset contains 1 million utterances with 6,112 identities. 500 test videos with 25 identities were randomly chosen for the test." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Comparing Methods Some prior works such as PC-AVS [65], GC-AVT [28], and EAMM [20] require additional driving sources instead of predicting them based on" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7649" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 289, + 193 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 289, + 193 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 289, + 193 + ], + "type": "table", + "html": "
MethodA S SctgntVoxCeleb2VoxCeleb1
FID ↓Sc ↑Nc ↓FID ↓Sc ↑Nc ↓
Wav2Lip [38]X X7.8022.39.230.1844.78.800.13
EAMM [20]X X1.7626.14.751.7042.62.780.58
PC-AVS [65]X X7.3514.48.210.1235.48.420.15
MakeItTalk [66]✓ X7.3519.52.030.7240.22.160.71
Audio2head [54]✓ X7.351016.420.131046.650.10
Ours + AR✓ X7.3524.37.230.0249.77.050.04
Ours + Diff.✓ ✓7.3514.27.340.0035.07.310.01
", + "image_path": "7cf2f39c6995ed4aedca8d13bc8376ebde8bbefd54d0cdbb61461aa9da3948ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 287, + 250 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 287, + 250 + ], + "type": "text", + "content": "Table 1. The quantitative results of synchronization and image quality under self-reenactment scenario. Bold means the best. A indicates the ability to generate non-lip motions with audio-only signals without other signals, while S means the ability to sample different results for each round." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 51, + 259, + 116, + 329 + ], + "blocks": [ + { + "bbox": [ + 51, + 259, + 116, + 329 + ], + "lines": [ + { + "bbox": [ + 51, + 259, + 116, + 329 + ], + "spans": [ + { + "bbox": [ + 51, + 259, + 116, + 329 + ], + "type": "image", + "image_path": "fa66ce24ff507467591c54de0277c266fe026e082abfdfd957815d6a51ee5d9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 331, + 92, + 339 + ], + "lines": [ + { + "bbox": [ + 83, + 331, + 92, + 339 + ], + "spans": [ + { + "bbox": [ + 83, + 331, + 92, + 339 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 120, + 261, + 182, + 329 + ], + "blocks": [ + { + "bbox": [ + 120, + 261, + 182, + 329 + ], + "lines": [ + { + "bbox": [ + 120, + 261, + 182, + 329 + ], + "spans": [ + { + "bbox": [ + 120, + 261, + 182, + 329 + ], + "type": "image", + "image_path": "169ffe2b44a4265ecc6b4f53f1dd08d6067f130267ffffd8e4f97d357b3d87c0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 331, + 159, + 339 + ], + "lines": [ + { + "bbox": [ + 149, + 331, + 159, + 339 + ], + "spans": [ + { + "bbox": [ + 149, + 331, + 159, + 339 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 186, + 261, + 250, + 330 + ], + "blocks": [ + { + "bbox": [ + 186, + 261, + 250, + 330 + ], + "lines": [ + { + "bbox": [ + 186, + 261, + 250, + 330 + ], + "spans": [ + { + "bbox": [ + 186, + 261, + 250, + 330 + ], + "type": "image", + "image_path": "50344fa4daa4e9c0bbd39cf47f316cf198d205962fde5ad2713c8f3585d41199.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 331, + 226, + 339 + ], + "lines": [ + { + "bbox": [ + 216, + 331, + 226, + 339 + ], + "spans": [ + { + "bbox": [ + 216, + 331, + 226, + 339 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 342, + 265, + 354 + ], + "lines": [ + { + "bbox": [ + 65, + 342, + 265, + 354 + ], + "spans": [ + { + "bbox": [ + 65, + 342, + 265, + 354 + ], + "type": "text", + "content": "Figure 4. Distribution visualization of generated poses." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 366, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 475 + ], + "type": "text", + "content": "audio input. We only compare those methods on lip synchronization and image quality but emphasize that our setup is targeting a more practical scenario. Other works such as Wav2Lip [38] only focus on revising the lip region without touching other facial motions, which are not directly comparable either. MakeItTalk [66] and Audio2head [54] are two closely related works to ours in terms of problem setup, however, our method is designed to predict all the non-lip facial motions with diversity, rather than only limited poses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "text", + "content": "Implementation Details Backbones of our models including " + }, + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{id}" + }, + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_v" + }, + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_a" + }, + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{G}" + }, + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{D}" + }, + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "text", + "content": " are borrowed from [65]. For the diffusion prior network " + }, + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 46, + 482, + 287, + 567 + ], + "type": "text", + "content": ", we use a 8-layer transformer [51] with 512-d tokens and 1024-d fully feedforward layers. Note that our models were trained on VoxCeleb2 only but tested on both VoxCeleb1 and VoxCeleb2. All models were trained on 4 NVIDIA A100 GPUs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 574, + 183, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 183, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 183, + 588 + ], + "type": "text", + "content": "4.1. Quantitative Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Evaluation Metrics We evaluate the performance of generated talking head from the following aspects: image quality, audio-lip sync accuracy, the variation and naturalness of non-lip facial motion. Frechet Inception Distance (FID) score [14] is used for the evaluation of image quality. A lower FID score indicates a lower distance between the distribution of generated images and real images. Following prior works [38, 65], we use SyncNet Error-Confidence " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{S}_c)" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " as the indicator of audio-lip synchronization, where greater confidence indicates better synchronization. However, the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "text", + "content": "normalized confidence score NLSE-C (short for " + }, + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_c" + }, + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "text", + "content": ") is used to address the concern raised in recent works [57, 53] regarding the strong relationship between " + }, + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_c" + }, + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "text", + "content": " and its training data, which can make it unfair to compare methods trained on different data. " + }, + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_c" + }, + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_c = \\frac{\\left|\\mathbf{S}_c^{gen} - \\mathbf{S}_c^{gt}\\right|}{\\mathbf{S}_c^{gt}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_c^{gen}" + }, + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "text", + "content": " is the confidence value of generated images and " + }, + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_c^{gt}" + }, + { + "bbox": [ + 304, + 72, + 545, + 163 + ], + "type": "text", + "content": " is that of its training data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 163, + 545, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 545, + 212 + ], + "type": "text", + "content": "To evaluate facial motions such as pose, expression, and blink, we utilize a pre-trained 3D morphable face model [9] and include shape irrelevant 3DMM parameters to calculate the following metrics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 219, + 545, + 449 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 306, + 219, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 219, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 306, + 219, + 545, + 267 + ], + "type": "text", + "content": "- Var: The variance of generated facial motions, i.e., the variance of the 3DMM coefficients for each video is calculated and then averaged over the test set. A closer Var to GT indicates a better match to the variation of real data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{fm}" + }, + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "type": "text", + "content": ": FID score of 3DMM coefficients calculated as follows: " + }, + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{fm} = \\frac{1}{K}\\sum_{i=1}^{K}\\mathbf{FID}(\\beta_i)" + }, + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "type": "inline_equation", + "content": "\\beta_i" + }, + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "type": "text", + "content": " is a sequence of 3DMM coefficients in the " + }, + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 306, + 270, + 545, + 308 + ], + "type": "text", + "content": "-th video." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{\\Delta fm}" + }, + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "type": "text", + "content": ": FID score of 3DMM coefficient difference between consecutive frames, i.e., " + }, + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{\\Delta fm}" + }, + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "type": "text", + "content": ", which is similar to " + }, + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{fm}" + }, + { + "bbox": [ + 306, + 312, + 545, + 371 + ], + "type": "text", + "content": " but the 3DMM coefficient difference between consecutive frames is measured, taking temporal naturalness into consideration." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 376, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 545, + 449 + ], + "type": "text", + "content": "- SND: Our new proposed metric, denoted as Sequence Naturalness Distance. It is the sum of " + }, + { + "bbox": [ + 306, + 376, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{fm}" + }, + { + "bbox": [ + 306, + 376, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 376, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{\\Delta fm}" + }, + { + "bbox": [ + 306, + 376, + 545, + 449 + ], + "type": "text", + "content": " indicating the difference of distribution between generated motion and GT motion, from both spatial and temporal perspectives, i.e., " + }, + { + "bbox": [ + 306, + 376, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{SND} = \\mathbf{FID}_{fm} + \\mathbf{FID}_{\\Delta fm}" + }, + { + "bbox": [ + 306, + 376, + 545, + 449 + ], + "type": "text", + "content": ". The lower SND, the better naturalness." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 462, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 509 + ], + "type": "text", + "content": "Evaluation Results Our main results consist of two parts: audio-lip synchronization and image quality as shown in Table 1; richness and sequence naturalness which are shown in Table 2." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "text", + "content": "It shows that our method with diffusion prior archives the best synchronization ability compared to other methods in Table 1. Note that although Wav2Lip [38] and PC-AVS [65] achieve the highest SyncNet scores, our best " + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 304, + 510, + 545, + 594 + ], + "type": "text", + "content": " indicates the most substantial synchronization ability [57], which is demonstrated in Sec. 4.2. Meanwhile, the image quality of our model with diffusion prior is the best in both test sets." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "content": "In Table 2, our method shows significantly higher variance than the other two audio-driven methods and it is closer to the variance of real data. This indicates that our method can produce reasonable diverse head movements and expressions, rather than slight head movement. It's notable that auto-regressive prior, trained with causal attention and regression loss, has a higher variance than GT because it often generates extreme motions, showing a low naturalness score in Table 3. For naturalness, our method achieves the lowest " + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{fm}" + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{\\Delta fm}" + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "content": " as well as SND on" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7650" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 76, + 70, + 519, + 190 + ], + "blocks": [ + { + "bbox": [ + 76, + 70, + 519, + 190 + ], + "lines": [ + { + "bbox": [ + 76, + 70, + 519, + 190 + ], + "spans": [ + { + "bbox": [ + 76, + 70, + 519, + 190 + ], + "type": "table", + "html": "
MethodVoxCeleb2VoxCeleb1
Var →FIDfm ↓FIDΔfm ↓SND ↓Var →FIDfm ↓FIDΔfm ↓SND ↓
GT1.98---1.88---
MakeItTalk [66]0.674.701.746.440.804.271.075.34
Audio2head [54]0.895.941.307.240.765.031.016.04
Ours + AR3.075.432.227.652.926.211.687.89
Ours + Diff.1.573.601.084.681.663.980.874.85
w/o Lvel2.094.021.275.292.284.651.115.76
training w/o editing3.096.762.008.763.066.961.708.66
", + "image_path": "c5e3e8c1fc0624cf461fd3a211375fea8e9ba026d1471fa455d940cf5a2ecf2f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "type": "text", + "content": "Table 2. The quantitative results of variance and naturalness of generated non-lip motions under self-reenactment scenario. " + }, + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "type": "text", + "content": " indicates a closer score to GT is better. Due to setting differences, i.e., Wav2Lip, EAMM, and PC-AVS, are not valid comparisons here." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 77, + 217, + 518, + 619 + ], + "blocks": [ + { + "bbox": [ + 77, + 217, + 518, + 619 + ], + "lines": [ + { + "bbox": [ + 77, + 217, + 518, + 619 + ], + "spans": [ + { + "bbox": [ + 77, + 217, + 518, + 619 + ], + "type": "image", + "image_path": "14c2cd94afeb0438c0797a72a0544111867c8871db3289bc32b5799eb11c448c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 623, + 547, + 657 + ], + "lines": [ + { + "bbox": [ + 46, + 623, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 547, + 657 + ], + "type": "text", + "content": "Figure 5. Qualitative results of our method compared to other baselines. Each row shows nine uniformly sampled frames from videos. Here we use two audio sources to drive different identities respectively. Our method shows accurate lip-audio synchronization with diverse and natural poses and expressions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 676, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 676, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 676, + 288, + 712 + ], + "type": "text", + "content": "both VoxCeleb1 and VoxCeleb2, surpassing other state-of-the-art by a large margin. Also, we randomly sample 5,000 poses for each method and employ t-SNE [50] for visual-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 676, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 676, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 676, + 547, + 713 + ], + "type": "text", + "content": "ization in Fig. 4, which demonstrates that the distribution we generate is the most comprehensive to GT. More details and examples are shown in the supplementary. In summary," + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "7651" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 281, + 169 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 281, + 169 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 281, + 169 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 281, + 169 + ], + "type": "table", + "html": "
MethodSync AccuracyNaturalnessRichness
GT4.614.414.18
MakeItTalk [66]1.972.352.10
PC-AVS [65]*3.132.682.68
Audio2head [54]2.562.352.34
Ours + AR3.462.533.37
Ours + Diff.4.083.823.68
", + "image_path": "c3988fbacbbd83ba571a12b2d7208f0ebe079530e84d4a3e0418476ddce53860.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 169, + 286, + 190 + ], + "lines": [ + { + "bbox": [ + 47, + 169, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 169, + 286, + 190 + ], + "type": "text", + "content": "Table 3. Human evaluation on generated samples on VoxCeleb2. * means that PC-AVS was evaluated on cross-reenactment scenario." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 209, + 287, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 287, + 232 + ], + "type": "text", + "content": "our method can generate rich facial motions with excellent realism (i.e., a lower distribution distance) and naturalness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 241, + 175, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 241, + 175, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 175, + 253 + ], + "type": "text", + "content": "4.2. Qualitative Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 260, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 260, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 287, + 331 + ], + "type": "text", + "content": "Two audio-driven video sequences of our method verses other baselines are shown in Fig. 5. It's clearly shown that our method is capable of producing much more natural-looking and diverse facial motions than other baselines, including pose, expression, blink and gaze, while maintaining accurate audio-lip synchronization compared to the GT." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 339, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 339, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 339, + 287, + 590 + ], + "type": "text", + "content": "User Study We invited 20 subjects for human evaluations, focusing on three aspects: 1) the accuracy of audio-lip synchronization; 2) the naturalness of non-lip motions and coherency between non-lip motions and audios; 3) the richness of non-lip motions. 50 videos are generated from audio and identity in the test set using the following methods: our diffusion prior, our auto-regressive prior, MakeItTalk [66], Audio2head [54] and PC-AVS [65] (driven by poses from another video). After shuffling the generated videos, each annotator rates from 1 (bad) to 5 (good) according to Mean Opinion Scores (MOS) rating protocol. Table 3 tells that our model with diffusion prior archives the best synchronization, naturalness and richness among all the models. Our model with auto-regressive prior archives relatively good richness compared to the remaining models, we attribute it to the exposure bias problem [4] brought by auto-regressive generation, which leads to high richness but low naturalness. Note that PC-AVS shows incompatible result in Table 3 as compared to Table 1 on " + }, + { + "bbox": [ + 47, + 339, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_c" + }, + { + "bbox": [ + 47, + 339, + 287, + 590 + ], + "type": "text", + "content": ", indicating that choosing pose driving signals from a random video may lead to synthesis results with bad audio-lip synchronization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 590, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 287, + 638 + ], + "type": "text", + "content": "It's interesting that our proposed SND scores reflect the naturalness of user study to some degree. Nevertheless, we leave it to future work for rigorous study of their correlation through more human evaluations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 111, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 111, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 111, + 658 + ], + "type": "text", + "content": "4.3. Ablation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "Table 4 discusses the performance of lip & non-lip disentanglement. While tested on lip only metrics without " + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{ol}" + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": ", SyncNet scores get lower, which indicates that G tends to generate mouth movements with information extracted" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 547, + 145 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 547, + 145 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 547, + 145 + ], + "type": "table", + "html": "
MethodLip OnlyNon-lip Only
Sc↑Nc↓Bd↓Gd↓Ed↓Pd↓
GT7.35-----
w/o Lol2.840.610.00740.19600.05680.0021
w/o MB7.690.050.01120.03150.08750.0293
Ours7.480.030.00670.01630.05580.0019
", + "image_path": "4444654d62bae8cbe38faec3bc3143fd6bee5058f246ba58d3f1b06f30746917.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "lines": [ + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "type": "text", + "content": "Table 4. Evaluation of disentanglement between lip and non-lip on VoxCeleb2. " + }, + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_d" + }, + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "type": "inline_equation", + "content": "\\mathrm{G}_d" + }, + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "type": "text", + "content": " denote L2 distances of 2D landmark [58] of eyes and iris, respectively, while " + }, + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_d" + }, + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_d" + }, + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "type": "text", + "content": " denote L2 distances of 3DMM coefficients of poses and expressions accordingly. Note that " + }, + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{nl}^v" + }, + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "type": "text", + "content": " is fixed while tested on lip only metrics, and " + }, + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_l^a" + }, + { + "bbox": [ + 305, + 148, + 545, + 214 + ], + "type": "text", + "content": " is fixed while tested on non-lip only metrics for fair comparisons." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 545, + 293 + ], + "type": "text", + "content": "from visual modality instead of audio. Thus, lip & non-lip space are not fully disentangled. Without MB, the performance of driving non-lip motions gets worse. Because it is hard to use a large batch to compute correlation, resulting in a not well decoupled non-lip space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "text", + "content": "We conducted ablation study for " + }, + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "text", + "content": " from three aspects: 1) auto-regressive prior versus diffusion prior; 2) with/without velocity loss " + }, + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "inline_equation", + "content": "L_{vel}" + }, + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "text", + "content": "; 3) training with/without editing mechanism. Diffusion prior shows significant improvements on naturalness scores including " + }, + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{fm}" + }, + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\Delta fm}" + }, + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "text", + "content": " and SND, as compared to autoregressive prior shown in Table 2, with slightly better synchronization ability in Table 1. Fig. 6 shows that without " + }, + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "inline_equation", + "content": "L_{vel}" + }, + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "text", + "content": ", the prediction of non-autoregressive diffusion prior model " + }, + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 304, + 294, + 546, + 486 + ], + "type": "text", + "content": " becomes unstable, which leads to a larger variance and worse SND scores in Table 2. While training without the editing mechanism mentioned in Sec. 3.2, we applied editing only during sampling as in MDM [46], and observed that it jitters between adjacent frames as shown in Fig. 7, resulting in a poorer SND score in Table 2. Note that our model trained with editing mechanism generates smoother results." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 319, + 490, + 536, + 591 + ], + "blocks": [ + { + "bbox": [ + 319, + 490, + 536, + 591 + ], + "lines": [ + { + "bbox": [ + 319, + 490, + 536, + 591 + ], + "spans": [ + { + "bbox": [ + 319, + 490, + 536, + 591 + ], + "type": "image", + "image_path": "a2ca78a9ffd31ae9d3d234671fd3a8dcb4774cc72a0df01d974879acd3986791.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 596, + 545, + 618 + ], + "lines": [ + { + "bbox": [ + 306, + 596, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 306, + 596, + 545, + 618 + ], + "type": "text", + "content": "Figure 6. Ablation study of our diffusion prior " + }, + { + "bbox": [ + 306, + 596, + 545, + 618 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{a2nl}" + }, + { + "bbox": [ + 306, + 596, + 545, + 618 + ], + "type": "text", + "content": " w or w/o " + }, + { + "bbox": [ + 306, + 596, + 545, + 618 + ], + "type": "inline_equation", + "content": "L_{vel}" + }, + { + "bbox": [ + 306, + 596, + 545, + 618 + ], + "type": "text", + "content": ". Each row shows five uniformly sampled frames." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 633, + 383, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 383, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 383, + 645 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a novel talking head generation method based on a diffusion prior model, which can generate diverse and natural-looking talking head videos with only audio and an identity image as inputs. Such \"audio-driving is all you need\" setting is very friendly" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7652" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 72, + 277, + 174 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 277, + 174 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 277, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 277, + 174 + ], + "type": "image", + "image_path": "732e1e805223738e2c730ac7ebaeea552c1544f45af3394be6d22a32e1367d87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 178, + 287, + 200 + ], + "lines": [ + { + "bbox": [ + 47, + 178, + 287, + 200 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 287, + 200 + ], + "type": "text", + "content": "Figure 7. Ablation study of our diffusion prior " + }, + { + "bbox": [ + 47, + 178, + 287, + 200 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{a2nl}" + }, + { + "bbox": [ + 47, + 178, + 287, + 200 + ], + "type": "text", + "content": " trained w or w/o editing. Each row shows five adjacent frame." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 218, + 287, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 218, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 218, + 287, + 255 + ], + "type": "text", + "content": "for reenactment and dubbing applications. Comprehensive evaluations including our newly proposed metrics validated the effectiveness of our system." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 260, + 287, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 260, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 287, + 295 + ], + "type": "text", + "content": "Limitations Our method mainly focuses on non-lip motion generation. We leave the improvement of rendering quality to future works." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 306, + 105, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 306, + 105, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 306, + 105, + 317 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 326, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 326, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 326, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 326, + 287, + 368 + ], + "type": "text", + "content": "[1] Ahmed A Abdelrahman, Thorsten Hempel, Aly Khalifa, and Ayoub Al-Hamadi. L2cs-net: Fine-grained gaze estimation in unconstrained environments. arXiv preprint arXiv:2203.03339, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 370, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 370, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 370, + 287, + 422 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Adrià Recasens, Rosalia Schneider, Relja Arandjelović, Jason Ramapuram, Jeffrey De Fauw, Lucas Smaira, Sander Dieleman, and Andrew Zisserman. Self-Supervised MultiModal Versatile Networks. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "text", + "content": "[3] Mohammed M Alghamdi, He Wang, Andrew J Bulpitt, and David C Hogg. Talking head from speech audio using a pretrained image generator. arXiv preprint arXiv:2209.04252, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 470, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 470, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 53, + 470, + 287, + 514 + ], + "type": "text", + "content": "[4] Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems, 28, 2015. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "type": "text", + "content": "[5] Stella Bounareli, Vasileios Argyriou, and Georgios Tzimiropoulos. Finding directions in gan's latent space for neural face reenactment. arXiv preprint arXiv:2202.00046, 2022.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 559, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 287, + 601 + ], + "type": "text", + "content": "[6] Egor Burkov, Igor Pasechnik, Artur Grigorev, and Victor Lempitsky. Neural head reenactment with latent pose descriptors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 603, + 287, + 635 + ], + "type": "text", + "content": "[7] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 636, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 636, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 636, + 287, + 690 + ], + "type": "text", + "content": "[8] Soo-Whan Chung, Joon Son Chung, and Hong-Goo Kang. Perfect match: Improved cross-modal embeddings for audiovisual synchronisation. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3965-3969. IEEE, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 692, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 287, + 712 + ], + "type": "text", + "content": "[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0-0, 2019. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 150 + ], + "type": "text", + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 184 + ], + "type": "text", + "content": "[11] Aviv Gabbay, Asaph Shamir, and Shmuel Peleg. Visual speech enhancement. arXiv preprint arXiv:1711.08789, 2017.2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 274 + ], + "type": "text", + "content": "[13] Yudong Guo, Keyu Chen, Sen Liang, Yongjin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 330 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 366, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 420 + ], + "type": "text", + "content": "[16] Fangzhou Hong, Liang Pan, Zhongang Cai, and Ziwei Liu. Versatile multi-modal pre-training for human-centric perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16156-16166, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 422, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 475 + ], + "type": "text", + "content": "[17] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3397-3406, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 521 + ], + "type": "text", + "content": "[18] Gee-Sern Hsu, Chun-Hung Tsai, and Hung-Yi Wu. Dual-generator face reenactment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 642-650, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 523, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 577 + ], + "type": "text", + "content": "[19] Po-Hsiang Huang, Fu-En Yang, and Yu-Chiang Frank Wang. Learning identity-invariant motion representations for crossid face reenactment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7084–7092, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 580, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 545, + 623 + ], + "type": "text", + "content": "[20] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. arXiv preprint arXiv:2205.15278, 2022. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "text", + "content": "[21] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[22] Venkatesh S Kadandale, Juan F Montesinos, and Gloria Haro. Vocalist: An audio-visual synchronisation model for lips and voices. arXiv preprint arXiv:2204.02090, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7653" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 288, + 174 + ], + "type": "text", + "content": "[24] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Nießner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics, 37(4):163:1-14, Aug. 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 176, + 288, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 288, + 209 + ], + "type": "text", + "content": "[25] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 211, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 211, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 211, + 287, + 275 + ], + "type": "text", + "content": "[26] Jiyoung Lee, Soo-Whan Chung, Sunok Kim, Hong-Goo Kang, and Kwanghoon Sohn. Looking into your speech: Learning cross-modal affinity for audio-visual speech separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1336–1345, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 279, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 287, + 345 + ], + "type": "text", + "content": "[27] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3387-3396, June 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 347, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 347, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 347, + 287, + 413 + ], + "type": "text", + "content": "[28] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 415, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 448 + ], + "type": "text", + "content": "[29] Yuanxun Lu, Jinxiang Chai, and Xun Cao. Live speech portraits: real-time photorealistic talking-head animation. ACM Transactions on Graphics (TOG), 40(6):1-17, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 450, + 287, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 287, + 505 + ], + "type": "text", + "content": "[30] Moustafa Meshry, Saksham Suri, Larry S Davis, and Abhnav Shrivastava. Learned spatial representations for few-shot talking-head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13829-13838, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 507, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 507, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 507, + 287, + 540 + ], + "type": "text", + "content": "[31] Dongchan Min, Minyoung Song, and Sung Ju Hwang. Stylistalker: One-shot style-based audio-driven talking head video generation. preprint arXiv:2208.10922, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "text", + "content": "[32] Gaurav Mittal and Baoyuan Wang. Animating face using disentangled audio representations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), March 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "text", + "content": "[33] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. Voxceleb: a large-scale speaker identification dataset. arXiv preprint arXiv:1706.08612, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "type": "text", + "content": "[34] Arsha Nagrani, Joon son Chung, Samuel Albanie, and Andrew Zisserman. Disentangled speech embeddings using cross-modal self-supervision, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[35] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[36] Andrew Zisserman Olivia Wiles, A. Sophia Koepke. X2face: A network for controlling face generation by using images, audio, and pose codes. In ECCV 2018, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "text", + "content": "[37] Se Jin Park, Minsu Kim, Joanna Hong, Jeongsoo Choi, and Yong Man Ro. Synctalkface: Talking face generation with precise lip-syncing via audio-lip memory. In 36th AAAI Conference on Artificial Intelligence (AAAI 22). Association for the Advancement of Artificial Intelligence, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 164, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 217 + ], + "type": "text", + "content": "[38] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 285 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "text", + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 332, + 546, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 546, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 546, + 386 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[42] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. arXiv preprint arXiv:2207.11770, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 433, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 486 + ], + "type": "text", + "content": "[43] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 489, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 542 + ], + "type": "text", + "content": "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 544, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 577 + ], + "type": "text", + "content": "[45] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3,4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "type": "text", + "content": "[46] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 3, 4, 5, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "type": "text", + "content": "[47] Justus Thies, Mohamed Elgharib, Ayush Tewari, Christian Theobalt, and Matthias Nießner. Neural voice puppetry: Audio-driven facial reenactment. ECCV 2020, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[48] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7654" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[49] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. arXiv preprint arXiv:1906.05849, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "text", + "content": "[50] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 216 + ], + "type": "text", + "content": "[52] Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Realistic speech-driven facial animation with gans. CoRR, abs/1906.06337, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[53] Duomin Wang, Yu Deng, Zixin Yin, Heung-Yeung Shum, and Baoyuan Wang. Progressive disentangled representation learning for fine-grained controllable talking head synthesis. arXiv preprint arXiv:2211.14506, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "type": "text", + "content": "[54] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 288, + 350 + ], + "type": "text", + "content": "[55] Suzhen Wang, Lincheng Li, Yu Ding, and Xin Yu. One-shot talking face generation from single-speaker audio-visual correlation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2531–2539, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 425 + ], + "type": "text", + "content": "[56] Chao Xu, Jiangning Zhang, Yue Han, Guanzhong Tian, Xianfang Zeng, Ying Tai, Yabiao Wang, Chengjie Wang, and Yong Liu. Designing one unified framework for high-fidelity face reenactment and swapping. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 54-71. Springer, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "text", + "content": "[57] Shunyu Yao, RuiZhe Zhong, Yichao Yan, Guangtao Zhai, and Xiaokang Yang. Dfa-nerf: Personalized talking head generation via disentangled face attributes neural rendering. arXiv preprint arXiv:2201.00791, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 503 + ], + "type": "text", + "content": "[58] Baosheng Yu and Dacheng Tao. Heatmap regression via randomized rounding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 504, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 288, + 557 + ], + "type": "text", + "content": "[59] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "type": "text", + "content": "[60] Xianfang Zeng, Yusu Pan, Mengmeng Wang, Jiangning Zhang, and Yong Liu. Realistic face reenactment via self-supervised disentangling of identity and pose. pages 12757-12764. AAAI Press, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "type": "text", + "content": "[61] Chenxu Zhang, Yifan Zhao, Yifei Huang, Ming Zeng, Saifeng Ni, Madhukar Budagavi, and Xiaohu Guo. Facial: Synthesizing dynamic talking face with implicit attribute learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3867-3876, 2021. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "type": "text", + "content": "[62] Jiangning Zhang, Liang Liu, Zhucun Xue, and Yong Liu. Apb2face: Audio-guided face reenactment with auxiliary pose and blink signals. 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[63] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 274 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 326, + 72, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 547, + 106 + ], + "type": "text", + "content": "a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661-3670, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 106, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 547, + 162 + ], + "type": "text", + "content": "[64] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 9299–9306, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 228 + ], + "type": "text", + "content": "[65] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4176–4186, 2021. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 229, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 547, + 274 + ], + "type": "text", + "content": "[66] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7655" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/ff7524e3-aa7d-4baa-a842-04712fa34f60_content_list.json b/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/ff7524e3-aa7d-4baa-a842-04712fa34f60_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d9fa87beb0a880fcf3b58e4db260113810390ce8 --- /dev/null +++ b/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/ff7524e3-aa7d-4baa-a842-04712fa34f60_content_list.json @@ -0,0 +1,1398 @@ +[ + { + "type": "text", + "text": "Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow", + "text_level": 1, + "bbox": [ + 225, + 130, + 743, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federico Paredes-Valles $^{1,2}$ Kirk Y. W. Scheper $^{2}$ Christope De Wagter $^{1}$ Guido C. H. E. de Croon $^{1}$", + "bbox": [ + 84, + 202, + 908, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Micro Air Vehicle Laboratory, Delft University of Technology", + "bbox": [ + 243, + 226, + 751, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Stuttgart Laboratory 1, Sony Semiconductor Solutions Europe, Sony Europe B.V.", + "bbox": [ + 171, + 244, + 823, + 263 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8eb4e758df6147d95bc6ce1673c5196874b8dc4b404d80f7475b506b70aea81d.jpg", + "image_caption": [ + "Figure 1: Our pipeline estimates event-based optical flow by sequentially processing small partitions of the event stream with a recurrent model. We propose a novel self-supervised learning framework (left) based on a multi-timescale contrast maximization formulation that is able to exploit the high temporal resolution of event cameras via iterative warping to produce accurate optical flow predictions (right)." + ], + "image_footnote": [], + "bbox": [ + 78, + 290, + 509, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ebb4507e223e66d8b8413675910840cd4042c26e5346f46a69d49db706498050.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 292, + 890, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 587, + 313, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras have recently gained significant traction since they open up new avenues for low-latency and low-power solutions to complex computer vision problems. To unlock these solutions, it is necessary to develop algorithms that can leverage the unique nature of event data. However, the current state-of-the-art is still highly influenced by the frame-based literature, and usually fails to deliver on these promises. In this work, we take this into consideration and propose a novel self-supervised learning pipeline for the sequential estimation of event-based optical flow that allows for the scaling of the models to high inference frequencies. At its core, we have a continuously-running stateful neural model that is trained using a novel formulation of contrast maximization that makes it robust to nonlinearities and varying statistics in the input events. Results across multiple datasets confirm the effectiveness of our method, which establishes a new state of the art in terms of accuracy for approaches trained or optimized without ground truth.", + "bbox": [ + 73, + 618, + 473, + 891 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 587, + 632, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras capture per-pixel log-brightness changes at microsecond resolution [13]. This operating principle results in a sparse and asynchronous visual signal that, under constant illumination, directly encodes information about the apparent motion (i.e., optical flow) of contrast in the image space. These cameras offer several advantages, such as low latency and robustness to motion blur [13], and hence hold the potential of a high-bandwidth estimation of this optical flow information. However, the event-based nature of the generated visual signal poses a paradigm shift in the processing pipeline, and traditional, frame-based algorithms become suboptimal and often incompatible. Despite this, the majority of learning-based methods that have been proposed so far for event-based optical flow estimation are still highly influenced by frame-based approaches. This influence is normally reflected in two key aspects of their", + "bbox": [ + 496, + 612, + 893, + 854 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "The project's code and additional material can be found at https://mavlab.tudelft.nl/taming_event_flow/.", + "bbox": [ + 500, + 875, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9695", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pipelines: (i) the design of the network architecture, and (ii) the formulation of the loss function.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regarding architecture design, most literature methods format subsets of the input events as dense volumetric representations [45] that are processed at once by stateless (i.e., non-recurrent) models [17, 29, 37, 45]. Similarly to their frame-based counterparts [12, 38, 39], these models estimate the per-pixel displacement over the time-length of the event volume using only the information contained within it. Consequently, these volumes need to encode enough spatiotemporal information for motion to be discernible. However, if done over relatively long time periods, the subsequent models suffer from limitations such as high latency or having to deal with large pixel displacements [17, 18, 42].", + "bbox": [ + 75, + 123, + 467, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With respect to the loss function, multiple options have been explored due to the lack of real-world datasets with per-event ground truth. Pure supervised learning can be used with datasets such as MVSEC [43] or DSEC-Flow [17], but their ground truth only contains per-pixel displacement at low frequencies, which makes models difficult to train. On the other hand, a self-supervised learning (SSL) framework can be formulated using either the accompanying frames with the photometric error as a loss [11, 40, 44], or through an events-only contrast maximization [14, 15] proxy loss for motion compensation (i.e., event deblurring) [19, 26, 29, 35, 45]. However, despite not relying on ground truth, all the literature on SSL for optical flow assumes that the events move linearly within the time window of the loss, which ignores much of the potential of event cameras and their high temporal resolution (see Fig. 1, bottom left).", + "bbox": [ + 75, + 306, + 467, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we focus on the estimation of high frequency event-based optical flow and how this can be learned in an SSL fashion using contrast maximization with a relaxed linear motion assumption. To achieve this, we build upon the continuous-operation pipeline from Hagenaars et al. [19], which retrieves optical flow by sequentially processing small partitions of the event stream with a stateful (i.e., recurrent) model over time, instead of dealing with large volumes of input events. We augment (and train) this pipeline with a novel contrast maximization formulation that performs event motion compensation in an iterative manner at multiple temporal scales, as shown in Fig. 1. Using this framework, we achieve the best accuracy of all contrast-maximization-based approaches on multiple datasets, only being outperformed by pure supervised learning methods trained with ground-truth data.", + "bbox": [ + 75, + 549, + 467, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the extensions that we propose to the SSL framework in [19], i.e., our main contributions, are:", + "bbox": [ + 76, + 792, + 467, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- The first iterative event warping module in the context of contrast maximization (see Section 3.2). This module unlocks a novel multi-reference loss function that better captures the trajectory of scene points over time, thus improving the accuracy of the predictions.", + "bbox": [ + 94, + 825, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- The first multi-timescale approach to contrast maximization, which adds robustness, improves convergence, and reduces tuning requirements of loss-related hyperparameters (see Section 3.3).", + "bbox": [ + 517, + 90, + 890, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a result, we present the first self-supervised optical flow method for event cameras that relaxes the linear motion assumption, and hence that has the potential of exploiting the high temporal resolution of the sensor by producing estimates in a close-to continuous manner. We validate the proposed framework through extensive evaluations on multiple datasets. Additionally, we conduct ablation studies to show the effectiveness of each individual component.", + "bbox": [ + 498, + 152, + 890, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 286, + 635, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the aforementioned advantages of event cameras for optical flow estimation, extensive research has been carried out since these sensors were first introduced [1, 2, 4-8, 24, 27, 28, 35]. Regarding learning-based approaches, the first method was proposed by Zhu et al. in [44] with EV-FlowNet: a UNet-like [32] architecture trained with SSL, with the supervisory signal coming from the photometric error between subsequent frames captured with an accompanying camera. To avoid the need for a secondary vision sensor, in [45], Zhu et al. proposed an SSL framework around the contrast maximization for motion compensation idea from [14, 15], hence relying solely on event data. This pipeline was used and further improved in [19, 29, 35]. Other approaches trained EV-FlowNet in a supervised fashion with synthetic/real ground-truth data and showed higher accuracy levels through evaluations on public benchmarks [17, 37]. However, they also highlighted EV-FlowNet's inability to deal with large pixel displacements [17]. Because of this, inspired by the frame-based literature [39], Gehrig et al. proposed E-RAFT in [17]: the first architecture to introduce the use of correlation volumes in the event camera literature. This model, which was recently augmented with attention mechanisms [41], achieved state-of-the-art performance in multiple datasets.", + "bbox": [ + 498, + 311, + 890, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A novel perspective to learning event-based optical flow is to move away from event accumulation over long timespans, and instead rely on continuously-running stateful models that integrate information over time. The first methods of this kind were proposed by Paredes-Valles et al. [30] and Hagenaars et al. [19], but this idea has recently gained interest in the event camera literature [11, 31, 42]. The reason is that leveraging memory through sequential processing can potentially lead to lightweight and low-latency solutions that are also robust to large pixel displacements without the need for correlation volumes [42]. However, despite their potential of being scaled to high inference frequencies, all these solutions assume that events move linearly in the timespan of their loss function, and hence cannot capture the true trajectory of scene points over time.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9696", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When it comes to learning nonlinear pixel trajectories from event data, only the work of Gehrig et al. in [18] is to be highlighted. However, their approach uses multiple event and correlation volumes to fit Bézier curves to the trajectory of scene points, resulting in a high-latency solution. In contrast, in this work, we extend the continuous-operation pipeline from Hagenaars et al. [19] with an SSL framework in which discrete trajectories are regressed at high frequency by leveraging memory within the models. This approach allows, for the first time, to exploit the high temporal resolution of event cameras while capturing more accurately the trajectory of scene points over time thanks to the proposed iterative event warping mechanism.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Among non-learning-based approaches, the work of Shiba et al. in [35] bears particular relevance due to certain similarities with our framework. Specifically, Shiba et al. propose a tile-based method for event-based optical flow estimation that extends contrast maximization [14, 15] by incorporating (i) multiple spatial scales, (ii) a multi-reference focus loss, and (iii) a \"time-aware\" optical flow formulation. Similarly, our learning-based approach also employs a multi-scale strategy and utilizes a multi-reference focus loss. However, instead of making assumptions about the events' motion over extended time periods, we propose to learn their potentially nonlinear trajectories at high frequency by leveraging the iterative event warping module.", + "bbox": [ + 75, + 287, + 472, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 502, + 169, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of this work is to learn to sequentially estimate optical flow at high frequencies from a continuous stream of events. In such a pipeline, if the inference frequency is sufficiently high, events need to be processed nearly as soon as they are triggered by the sensor, with no pre-processing in between. Because of this, the integration of temporal information needs to happen in the network itself. To accomplish this, we propose the framework in Fig. 1, in which a stateful model is trained using our novel formulation of contrast maximization for sequential processing. The components of this framework are described in the following sections.", + "bbox": [ + 75, + 527, + 472, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Input format and contrast maximization", + "text_level": 1, + "bbox": [ + 76, + 708, + 423, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For an ideal camera, an event $\\boldsymbol{e}_i = (\\boldsymbol{x}_i, t_i, p_i)$ of polarity $p_i \\in \\{+, -\\}$ is triggered at pixel $\\boldsymbol{x}_i = (x_i, y_i)^T$ and time $t_i$ whenever the change in log-brightness since the last event at that pixel location reaches the contrast sensitivity threshold for that polarity [13].", + "bbox": [ + 75, + 732, + 468, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As in [19], we use a two-channel event count image as input representation, which gets populated with consecutive, non-overlapping, fine discrete partitions of the event stream $\\varepsilon_{k}^{\\mathrm{inp}}\\doteq \\{e_{i}\\}$ (further referred to as input partitions), each containing all the events in a time window of a certain duration, i.e., $t_i\\in [t_k^{\\mathrm{begin}},t_k^{\\mathrm{end}}]$ . This representation does not", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "contain temporal information by itself, and recurrent models are hence required for estimating of optical flow.", + "bbox": [ + 496, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regarding learning, we use the contrast maximization framework [14] to train, in an SSL fashion, to continuously estimate dense (i.e., per-pixel) optical flow from the event stream. Assuming brightness constancy, accurate flow information is encoded in the spatiotemporal misalignments (i.e., blur) among the events triggered by the same portion of a moving edge. To retrieve it, one has to compensate for this motion by geometrically transforming the events using a motion model. As in [19, 29, 35, 45], we transport each event to a reference time $t_{\\mathrm{ref}}$ through:", + "bbox": [ + 496, + 121, + 893, + 272 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {i} ^ {\\prime} = \\boldsymbol {x} _ {i} + \\left(t _ {\\text {r e f}} - t _ {i}\\right) \\boldsymbol {u} \\left(\\boldsymbol {x} _ {i}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 282, + 890, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{u}(\\pmb{x}) = (u(\\pmb{x}), v(\\pmb{x}))^T$ denotes the optical flow map used to transport each event from $t_i$ to $t_{\\mathrm{ref}}$ . The result of aggregating the transformed events is further referred to as the image of warped events (IWE) at $t_{\\mathrm{ref}}$ .", + "bbox": [ + 496, + 310, + 890, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the loss function of our SSL framework, we adopt the time-based focus objective function from [19]. Using the warped events at a given $t_{\\mathrm{ref}}$ , we generate an image of the per-pixel average timestamps for each polarity $p'$ via bilinear interpolation:", + "bbox": [ + 496, + 372, + 893, + 446 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} T _ {p ^ {\\prime}} (\\boldsymbol {x}; \\boldsymbol {u} | t _ {\\text {r e f}}) = \\frac {\\sum_ {j} \\kappa \\left(x - x _ {j} ^ {\\prime}\\right) \\kappa \\left(y - y _ {j} ^ {\\prime}\\right) \\bar {t} _ {j} \\left(t _ {\\text {r e f}} , t _ {j}\\right)}{\\sum_ {j} \\kappa \\left(x - x _ {j} ^ {\\prime}\\right) \\kappa \\left(y - y _ {j} ^ {\\prime}\\right) + \\epsilon} \\tag {2} \\\\ \\kappa (a) = \\max (0, 1 - | a |) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 455, + 890, + 513 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nj = \\{i \\mid p _ {i} = p ^ {\\prime} \\}, \\quad p ^ {\\prime} \\in \\{+, - \\}, \\quad \\epsilon \\approx 0\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 515, + 815, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{t}_i$ denotes the normalized timestamp contribution of the $i$ th event, according to Fig. 3 and Eq. 5.", + "bbox": [ + 496, + 542, + 893, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, the contrast maximization loss function at $t_{\\mathrm{ref}}$ is defined as the scaled sum of the squared temporal images:", + "bbox": [ + 496, + 574, + 890, + 604 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C M}} \\left(t _ {\\text {r e f}}\\right) = \\frac {\\sum_ {\\boldsymbol {x}} T _ {+} \\left(\\boldsymbol {x} ; \\boldsymbol {u} \\mid t _ {\\text {r e f}}\\right) ^ {2} + T _ {-} \\left(\\boldsymbol {x} ; \\boldsymbol {u} \\mid t _ {\\text {r e f}}\\right) ^ {2}}{\\sum_ {\\boldsymbol {x}} \\left[ n \\left(\\boldsymbol {x} ^ {\\prime}\\right) > 0 \\right] + \\epsilon} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 613, + 890, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $n(\\pmb{x}^{\\prime})$ denotes a per-pixel event count of the IWE. The lower the $\\mathcal{L}_{\\mathrm{CM}}$ , the better the event deblurring at $t_{\\mathrm{ref}}$ .", + "bbox": [ + 496, + 657, + 890, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As discussed in [14, 19, 36], for any focus objective function to be a robust supervisory signal for contrast maximization, the event partition used in the optimization needs to contain enough motion information (i.e., blur) so it can be compensated for. However, as in [19], that is not the case in our input partitions due to the fine discretization of the event stream that we are targeting. Therefore, only at training time, we define the so-called training partition (or event buffer in Fig. 1) $\\varepsilon_{k\\rightarrow k + R}^{\\mathrm{train}}\\doteq \\{\\varepsilon_i^{\\mathrm{inp}}\\}_{i = k}^{k + R}$ , which stacks together the events in the input partitions of $R$ successive forward passes through the network. Once these are performed, we use this partition and the estimated optical flow maps to compute the loss, and use truncated backpropagation through time to update the model parameters. After", + "bbox": [ + 496, + 689, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9697", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8eda2d9e871b781ae7697add87da18789d713eb2becd205c4eb6f65722bae766.jpg", + "image_caption": [ + "Figure 2: Incompatibility of multi-reference deblurring and linear event warping in the presence of nonlinearities in the pixel trajectories. Top: Events generated by a moving dot following circular motion (left), and optical flow color-coding scheme (right). Middle: Optical flow solutions required to produce sharp IWEs at different $t_{\\mathrm{ref}}$ using linear and iterative warping. While the former requires a different solution for each $t_{\\mathrm{ref}}$ , the proposed iterative warping can achieve this using a single solution. Arrows illustrate the direction of the required displacement $\\Delta \\pmb{x}_i = (t_{\\mathrm{ref}} - t_i)\\pmb{u}(\\pmb{x}_i)$ at each (discretized) spatial location. Bottom: Resulting IWEs at different $t_{\\mathrm{ref}}$ using the optimal optical flow map for $t_{\\mathrm{ref}} = \\mathrm{mid}$ ." + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 450, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this update, we detach the states of the network from the computational graph and clear the training partition and optical flow buffer. The importance of sequential processing, as an alternative to training stateless models in short input partitions, is corroborated in the supplementary material.", + "bbox": [ + 75, + 762, + 467, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Iterative event warping", + "text_level": 1, + "bbox": [ + 76, + 847, + 292, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to better approximate the trajectories of scene points over time, in this work we relax the linear motion as-", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ed19fc19c64c3b8d34fb8c174a5b4f787c59c9755326f2109289b1b3a79a8786.jpg", + "image_caption": [ + "Figure 3: Timestamp normalization profiles for the per-event contributions to the images of average timestamps in Eq. 2, for $R = 10$ . Left: Deblurring only done at the extremes of the training partition, as in [19, 29, 45]. Right: Deblurring done at all reference times, thanks to the proposed iterative event warping. Normalization profiles only shown for three $t_{\\mathrm{ref}}$ for a better visualization." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 887, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sumption in the SSL framework for event-based optical flow by augmenting the sequential-estimation pipeline from [19] with iterative event warping. Instead of transporting events to a given $t_{\\mathrm{ref}}$ assuming linear motion regardless of the length of the warping interval (as in [19, 29, 35, 45]), we perform a finer discretization of the event trajectories and assume that motion is only linear between optical flow estimates. Therefore, to express a group of events at a given $t_{\\mathrm{ref}}$ , we geometrically transform them using all the intermediate optical flow estimates through multiple iterations of Eq. 1. Fig. 2 shows the differences between the linear event warping in [19] and the proposed iterative augmentation, as well as the limitations of the former. Note that our iterative warping is fundamentally different from that of the work of Wu et al. in [42], where input events are deblurred before being passed to the models using residual optical flow estimates until convergence.", + "bbox": [ + 498, + 343, + 892, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Literature methods on event-based optical flow with contrast maximization compute the focus objective function at multiple reference times in the training partition (usually at the extremes) to prevent overfitting and/or scaling issues during backpropagation [19, 29, 35, 45]. However, since these approaches assume optical flow constancy in the span of their loss function, they suffer with nonlinearities in the pixel trajectories (see the blurry IWEs in Fig. 2, bottom). On the contrary, because of the finer discretization of the event trajectories coming with our sequential processing pipeline and the proposed iterative warping, we can use any (combination of) reference time(s) for the computation of the focus objective function. In fact, as shown in Fig. 3 (right), we propose the use of all the discretization points as reference times for event deblurring. Apart from the aforementioned regularizing benefits, having to produce sharp IWEs at any $t_{\\mathrm{ref}}$ forces the models to estimate a sequence of optical flow maps that is consistent with the velocity profile of the event stream. For a given training partition of", + "bbox": [ + 496, + 613, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9698", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b48ea9605462c4fee827fe64233a73e117cf6ac63dfb0de8c3b87ffefb877f92.jpg", + "image_caption": [ + "Figure 4: Multi-timescale approach to contrast maximization. For a given training partition of length $R$ , we fit multiple sub-partitions of different lengths (in the figure: one of length $R$ in red, two $R/2$ in blue, and four $R/4$ in green) and compute the loss in each of them according to Eq. 4. The global loss is computed as in Eq. 6. This figure only shows the timestamp normalization profiles of the central $t_{\\mathrm{ref}}$ of each sub-partition, but the losses are still computed at all reference times. An image representation of the accumulated input events in a sub-partition of each timescale is also shown." + ], + "image_footnote": [], + "bbox": [ + 80, + 95, + 480, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "length $R$ , the loss is computed as follows:", + "bbox": [ + 76, + 470, + 356, + 486 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C M}} ^ {R} = \\frac {1}{R + 1} \\sum_ {t _ {\\text {r e f}} = 0} ^ {R} \\mathcal {L} _ {\\mathrm {C M}} \\left(t _ {\\text {r e f}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 496, + 468, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which ensures that the IWEs (and the corresponding images of average timestamps in Eq. 2) at all reference times $t_{\\mathrm{ref}} \\in [0, R]$ contribute equally to the loss.", + "bbox": [ + 76, + 549, + 468, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As in [19, 29, 45], for $\\mathcal{L}_{\\mathrm{CM}}(t_{\\mathrm{ref}})$ in Eq. 3 to be a valid supervisory signal, events that are temporally close to the reference time $t_{\\mathrm{ref}}$ need to contribute to the temporal image in Eq. 2 with larger timestamp values than events that are far in time. Therefore, once the events have been transported to $t_{\\mathrm{ref}}$ , their timestamp is normalized prior to the computation of Eq. 2 as follows:", + "bbox": [ + 76, + 594, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {t} _ {i} \\left(t _ {\\text {r e f}}, t _ {i}\\right) = 1 - \\frac {\\left| t _ {\\text {r e f}} - t _ {i} \\right|}{R}, \\quad \\text {w i t h} t _ {i} \\in [ 0, R ] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 710, + 468, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which results in the normalization profiles in Fig. 3, for a given training partition of length $R$ .", + "bbox": [ + 76, + 750, + 468, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by [25], we mask individual events from the computation of the loss whenever we detect that they are transported outside the image space through the event warping process in the span of a deblurring window defined at the reference time $t_{\\mathrm{ref}}$ . This prevents our models from learning incorrect deformations at the image borders, and is motivated by the fact that the complete trajectory of the pixel is only partially observable in the training partition. An ab", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lation study on the impact of this masking strategy can be found in the supplementary material.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lastly, we do not augment the loss function in Eq. 4 with smoothing priors acting as regularization mechanisms. With iterative event warping, the error propagates through all the pixels covered in the warping process, regardless of whether they have input events or not. Therefore, the spatial coherence of the resulting optical flow maps is enhanced.", + "bbox": [ + 498, + 121, + 893, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Deblurring at multiple timescales", + "text_level": 1, + "bbox": [ + 500, + 222, + 792, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite the addition of iterative event warping, the success of our SSL framework still heavily depends on the hyperparameters that control the amount of motion information perceived by the networks in the span of a deblurring window. In our pipeline, these are: $\\mathrm{dt}_{\\mathrm{input}}$ , the timestep used to discretize the event stream; and $R$ , the number of forward passes, and hence optical flow maps, per loss. Thus, the effective length (in units of time) of the event window used for motion compensation is given by $\\mathrm{dt}_{\\mathrm{input}} \\times R$ . We hypothesize that, for each training dataset, there is an optimal length for this window that depends on the statistics of the data (e.g., event density, distribution of optical flow magnitudes) and model architecture, and that deviations from this optimal length lead to the learning of suboptimal solutions. E.g., shorter windows may converge to solutions that are more selective to fast rather than slow moving objects, and vice versa. Note that not only our method is sensitive to the tuning of these parameters, but also previous approaches based on contrast maximization [19, 29, 35, 45].", + "bbox": [ + 496, + 244, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To add a layer of robustness to the framework and relax its strong dependency on hyperparameter optimization, we propose the multi-timescale approach illustrated in Figs. 1 and 4. For a given training partition of length $R$ , instead of computing a single focus loss through Eq. 4, we compute this loss at $S$ temporal scales of length $R / 2^s$ , with $0 \\leq s \\leq S - 1$ , and combine them as follows:", + "bbox": [ + 496, + 532, + 893, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C M}} ^ {\\text {m u l t i}} = \\frac {1}{S} \\sum_ {s = 0} ^ {S - 1} \\frac {1}{2 ^ {s}} \\sum_ {p = 0} ^ {2 ^ {s} - 1} \\mathcal {L} _ {\\mathrm {C M}, p} ^ {R / 2 ^ {s}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 650, + 890, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown, we fit multiple non-overlapping sub-partitions in the training buffer if $s > 0$ . The subscript $p$ indicates their location in this buffer, starting from the earliest (see Fig. 4).", + "bbox": [ + 496, + 704, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that, through this multi-timescale approach to contrast maximization, our models need to converge to a solution that is suitable for all the timescales in the optimization, regardless of their length. An alternative formulation would be to incorporate per-pixel learnable masks (i.e., an attention module in the loss space) so that, depending on the input statistics, learning only happens at the most adequate scale. However, for this to happen, the loss function would have to be augmented to stimulate this behavior, and it is unclear how that would be done in practice.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9699", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1452b0e4edbc9fba7d891bbd856f0521d25a10cf94e5c12fc762238f3f7a6a94.jpg", + "image_caption": [ + "Figure 5: Schematic of the model architecture used in this work. It is characterized by $N_{\\mathcal{G}}$ recurrent encoders, $N_{\\mathcal{R}}$ residual blocks, and $N_{\\mathcal{G}}$ decoder layers. Optical flow estimates are produced at all decoder levels. In this diagram, $N_{\\mathcal{G}} = 2$ and $N_{\\mathcal{R}} = 1$ ." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 459, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Network architecture", + "text_level": 1, + "bbox": [ + 76, + 271, + 277, + 285 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the recurrent version of EV-FlowNet [44] proposed in $[19]^1$ (see Fig. 5). The events are represented as event count images (see Section 3.1), then passed through four encoders with strided convolutions followed by ConvGRUs [3] (channels doubling, starting from 64), two residual blocks [20], and then four decoders performing bilinear upsampling followed by convolution. After each decoder, there is a skip connection (using element-wise summation) from the corresponding encoder, as well as a depthwise convolution to produce estimates at lower scales, which are then concatenated with the activations of the previous decoder. Note that the proposed focus loss function (see Eq. 6) is applied to each intermediate optical flow estimate via upsampling. Lastly, all layers use $3\\times 3$ kernels and ReLU activations except for the prediction layers, which use TanH.", + "bbox": [ + 75, + 294, + 472, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 534, + 209, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on the DSEC-Flow [16, 17] and MVSEC [43] datasets. We evaluate the accuracy of the predictions based on the following metrics: (i) EPE (lower is better, ↓), the endpoint error; (ii) $\\%_{3\\mathrm{PE}}(\\downarrow)$ , the percentage of points with EPE greater than 3 pixels; (iii) FWL ( $\\uparrow$ ) [37], a deblurring quality metric based on the variance of the IWEs; and (iv) RSAT ( $\\downarrow$ ) [19], a deblurring quality metric based on the per-pixel average timestamps of the IWEs. We compare our solution to the published baselines, which range from supervised learning (SL) methods trained with ground truth, to SSL methods trained with grayscale images ( $\\mathrm{SSL}_{\\mathrm{F}}$ ) or events ( $\\mathrm{SSL}_{\\mathrm{E}}$ ), and model-based approaches (MB).", + "bbox": [ + 75, + 559, + 468, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train all our models on a subset of sequences from the training dataset of DSEC-Flow (only daylight recordings, see supplementary material). This corresponds to 19 minutes of training data, which we split into $572128 \\times 128$ (randomly-cropped) sequences of 2 seconds each. We use a batch size of 8 and train until convergence with the Adam optimizer [21] and a learning rate of 1e-5. To keep mem", + "bbox": [ + 75, + 741, + 468, + 847 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6151cc11d5d61687c1d5afe4a6a2e089ccd485066c3a3f26d6fb01ba747c1a0e.jpg", + "image_caption": [ + "Figure 6: Reconstruction of the pixel displacement of a scene point in the time window of a ground-truth sample from the multiple optical flow maps estimated in this period, i.e., $\\forall \\pmb{u}_k \\in [t_{\\mathrm{begin}}, t_{\\mathrm{end}}]$ . The error of the last optical flow estimate is magnified for clarity." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 893, + 220 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/88ca45d1e652da105f3e01188fe7e03bb1c4f09f3a78b632db5252f6a97fd9e0.jpg", + "table_caption": [], + "table_footnote": [ + "*Retrained by us on DSEC-Flow, linear warping." + ], + "table_body": "
EPE↓%3PE↓FWL↑RSAT↓
SLE-RAFT [17]0.792.681.330.87
EV-FlowNet, Gehrig et al. [17]2.3218.60--
Gehrig et al. [18]0.752.44--
IDNet [42]0.722.04--
TIDNet [42]0.843.41--
TMA [23]0.742.30--
Cuadrado et al. [9]1.7110.31--
E-Flowformer [22]0.762.68--
SSLEEV-FlowNet* [45]3.8631.451.300.85
ConvGRU-EV-FlowNet* [19]4.2733.271.550.90
dt = 0.01s, R = 2, S = 1 (Ours)9.6686.441.911.07
dt = 0.01s, R = 5, S = 1 (Ours)4.0552.221.580.97
dt = 0.01s, R = 10, S = 1 (Ours)2.3317.771.260.88
dt = 0.01s, R = 20, S = 1 (Ours)16.6333.671.061.10
dt = 0.01s, R = 10, S = 3 (Ours)2.8227.091.370.92
dt = 0.01s, R = 20, S = 4 (Ours)2.7323.731.240.90
MBShiba et al. [35]3.4730.861.370.89
", + "bbox": [ + 503, + 303, + 890, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Quantitative evaluation on the DSEC-Flow dataset [17]. Best in bold, runner up underlined. A breakdown of the results is provided in the supplementary material.", + "bbox": [ + 498, + 561, + 890, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ory usage within limits, we only propagate error gradients through up to 1e3 randomly-chosen events per millisecond of data. Despite this, note that we warp and use all the input events for the computation of the loss function.", + "bbox": [ + 498, + 625, + 890, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Evaluation procedure", + "text_level": 1, + "bbox": [ + 500, + 696, + 702, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When evaluating our sequential models, if $\\mathrm{dt_{gt}} > \\mathrm{dt_{input}}$ , we need to reconstruct the estimated per-pixel displacement in the ground-truth time window from the multiple optical flow maps estimated in this period. We do this by first averaging the (bilinearly interpolated) optical flow vectors that describe the trajectory of each scene point, and then by scaling the resulting optical flow vectors by $\\mathrm{dt_{gt}} / \\mathrm{dt_{input}}$ . This converts them from units of pixels/input to units of pixel displacement. An illustration of this reconstruction is shown in Fig. 6 for a scene point following a nonlinear trajectory. Note that our solution is subject to cumulative errors when evaluated through this reconstruction on benchmarks with", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1Our architecture is equivalent to ConvGRU-EV-FlowNet [19]. However, for the purpose of clarity within the rest of the paper, we will use this term to specifically refer to the original model from [19].", + "bbox": [ + 75, + 862, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9700", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b82cc80d8f8247aa65493bb77918a1efe77d1ff9f0da5a3212529b8f02586a7e.jpg", + "image_caption": [ + "Figure 7: Qualitative comparison of our method with the state-of-the-art E-RAFT architecture [17] and the model-based approach from Shiba et al. [35] on sequences from the test partition of the DSEC-Flow dataset [17]. Ground truth not included due to unavailability. The optical flow color coding can be found in Fig. 2 (top), and the corresponding IWEs in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 98, + 85, + 872, + 540 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ground truth provided at low rates (e.g., $10\\mathrm{Hz}$ in DSEC-Flow [17]). Therefore, it will compare unfavorably to other, non-sequential, methods that only produce a single optical flow estimate in the timespan of a ground-truth sample.", + "bbox": [ + 75, + 617, + 470, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Optical flow evaluation", + "text_level": 1, + "bbox": [ + 76, + 694, + 290, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on DSEC-Flow: Quantitative results of our evaluation on DSEC-Flow are presented in Table 1, and are supported by the qualitative comparison in Fig. 7. For this experiment, we trained multiple models with the same $\\mathrm{dt}_{\\mathrm{input}} = 0.01\\mathrm{s}$ (i.e., $\\times 10$ faster than DSEC's ground truth) but different lengths of the training partition, and with and without the multi-timescale approach. Multiple conclusions can be derived from the reported results. Firstly, our best performing model (i.e., $R = 10$ , $S = 1$ ) achieves the best accuracy of all contrast-maximization-based approaches on this dataset according to the EPE and the percentage of outliers. Specifically, it outperforms the baselines with an im", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "provement in the EPE in the $33\\% - 45\\%$ range, only being outperformed by SL methods trained with ground truth on the same dataset [9, 17, 22, 23, 42]. This confirms that (i) the timestamp-based loss function in Section 3.1 allows us to learn accurate event-based optical flow (contrary to the findings of [34]); and that (ii) our augmentations to the sequential pipeline in [19] lead to a significant improvement in the accuracy of the model (i.e., $45\\%$ improvement in the EPE).", + "bbox": [ + 496, + 617, + 890, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Secondly, these results also confirm our hypothesis that, for each training dataset, there is an optimal length for the training partition $R$ in terms of the EPE. According to Table 1, the optimal $R$ for this dataset, our model architecture, and our $\\mathrm{dt}_{\\mathrm{input}}$ is 10 (i.e., 0.1s of event data), with the EPE increasing if the training partition is made shorter or longer. However, as also shown in this table, we can relax the strong dependency of the contrast maximization framework on this parameter through the proposed multi-timescale approach.", + "bbox": [ + 496, + 763, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9701", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9c6a4d58e77599920c4ed2a0d6115fbe1fc498ddbb7a3cab9d673573aee2897a.jpg", + "table_caption": [], + "table_footnote": [ + "*Non-recurrent, volum. event repr. with 10 bins." + ], + "table_body": "
EPE↓%3PE↓FWL↑RSAT↓
dt = 0.1s*3.4834.720.980.87
dt = 0.05s3.0927.361.110.91
dt = 0.01s2.3317.771.260.88
dt = 0.005s2.3417.921.380.89
dt = 0.002s2.6621.832.040.90
", + "bbox": [ + 130, + 85, + 419, + 175 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our $S > 1$ models converged to solutions that slightly underperform our best performing single-scale model (EPE went up by $17\\% - 21\\%$ ), but were trained without the need to fine-tune the length of the training partition. Note that, despite this slight drop in accuracy, these multi-timescale solutions still outperform the other non-SL baselines. To further support these results, a visualization of the distribution of the EPE of our models as a function of the ground truth magnitude is provided in the supplementary material.", + "bbox": [ + 75, + 250, + 467, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lastly, Table 1 also allow us to conclude that deblurring quality metrics FWL [37] and RSAT [19] are not reliable indicators of the quality of the estimated optical flow. The reason for this is their inability to capture \"event collapse\" issues (as described in [33]), and would give favorable scores to undesirable solutions that warp all the events into a few pixels. According to our results, the FWL metric, being the spatial variance of the IWE relative to that of the identity warp, suffers more from this issue: the best FWL value is obtained with a model with 9.66 EPE.", + "bbox": [ + 75, + 386, + 467, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regarding qualitative results, Fig. 7 shows a comparison of our best performing model with the state-of-the-art E-RAFT architecture [17] and the contrast-maximization-based approach from Shiba et al. [35] on multiple sequences from the test partition of DSEC-Flow (i.e., ground truth is unavailable). These results confirm that our models are able to estimate high quality event-based optical flow despite not having access to ground-truth data during training, and also show the superiority of our method over the current best contrast-maximization-based approach [35]. Two limit cases in which our models provide suboptimal solutions are also shown in this figure: (i) sequences recorded at night (e.g., zurich_city_12_a) due to the presence of large amounts of events triggered by flashing lights and not by motion; and (ii) the car hood, which is also problematic for E-RAFT (i.e., does not capture it) and for [35]. Note that, in our case, (ii) is an artifact of the pixel displacements reconstructed from multiple optical flow estimates, and could be mitigated by having an occlusion handling mechanism in this reconstruction process.", + "bbox": [ + 75, + 537, + 467, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition to the evaluation in Table 1 and Fig. 7, we also conducted an experiment in which we trained multiple models with different $\\mathrm{dt}_{\\mathrm{input}}$ (ranging from 0.1s to 0.002s) but with the same amount of information in the training par", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/171d97d808feb2436182a3a948df99a423cc0ed960d33a654f82ba7f1fdb74a5.jpg", + "table_caption": [ + "Table 2: Impact of the input window length on the DSEC-Flow dataset [17]. Best in bold, runner up underlined." + ], + "table_footnote": [], + "table_body": "
EPE↓%3PE↓
SLEV-FlowNet+ [37]0.680.99
E-RAFT [17]0.241.70
EV-FlowNet, Gehrig et al. [17]0.310.00
TMA [23]0.250.07
Cuadrado et al. [9]0.85-
SSLFEV-FlowNet, Zhu et al. [44]0.490.20
Ziluo et al. [11]0.420.00
SSLEEV-FlowNet, Zhu et al. [45]0.320.00
EV-FlowNet, Paredes-Vallés et al. [29]0.925.40
EV-FlowNet, Shiba et al. [35]0.360.09
ConvGRU-EV-FlowNet [19]0.470.25
Ours0.270.05
MBAkolkar et al. [1]2.75-
Brebion et al. [7]0.530.20
Shiba et al. [35]0.300.11
", + "bbox": [ + 532, + 82, + 864, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: Quantitative evaluation on MVSEC's outdoor_1 sequence [43]. Best in bold, runner up underlined. Results on the indoor sequences can be found in the supplementary material.", + "bbox": [ + 498, + 311, + 890, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tition: 0.1s of event data. Results in Table 2 show that our sequential pipeline leads to an improvement in the accuracy of the predicted optical flow maps with respect to the stateless EV-FlowNet, which processes the 0.1s of event data at once. This improvement is due to the fact that the complexity of dealing with large pixel displacements gets reduced when processing the input data sequentially using shorter input windows. In addition to this, Table 2 also shows that the accuracy of our models is not compromised when estimating optical flow at higher frequencies, despite the high sparsity levels in the input data at those rates.", + "bbox": [ + 496, + 383, + 890, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on MVSEC: Quantitative results of our evaluation on the outdoor_day1 sequence from MVSEC are presented in Table 3, and are supported by the qualitative comparison in the supplementary material. For this experiment, since (i) there is no consensus in the literature with respect to the training dataset [17, 19, 29, 35, 37, 44, 45], and (ii) the outdoor_day2 sequence (i.e., the other daylight, automotive sequence) is only 9 minutes of duration during which the event camera is subject to high frequency vibrations [43], we decided to transfer one of our models trained on DSEC-Flow to MVSEC. More specifically, we chose the model trained with $\\mathrm{dt}_{\\mathrm{input}} = 0.005\\mathrm{s}$ and $R = 20$ from Table 2, as a model trained with a short input window on DSEC-Flow is expected to be robust to the slow motion statistics of MVSEC [17]. We deployed the model at the same frequency as the temporally-upsampled ground truth (i.e., $45\\mathrm{Hz}$ ). Results in Table 3 show that, by doing this, our model outperforms the great majority of methods in terms of the EPE (even some SL methods trained on this dataset), and is only surpassed by the current state-of-the-art E-RAFT architecture [17]. Besides reaffirming the high quality of the produced optical flow estimates, these results also confirm the generalizability of our method. For com", + "bbox": [ + 496, + 551, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9702", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pleteness, the results on the indoor evaluation sequences from MVSEC are provided in the supplementary material.", + "bbox": [ + 76, + 90, + 468, + 122 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 76, + 133, + 200, + 150 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The self-supervised method for event-based optical flow presented in this work, while demonstrating highly accurate and promising results, is not without limitations. Two critical challenges that need to be acknowledged are the brightness constancy assumption and the aperture problem. Firstly, the contrast maximization framework [14, 15] assumes constant illumination, leading our models to face difficulty in learning from events that are not due to motion in the image space but that arise from changes in illumination. Since this limitation is inherent to contrast maximization, it extends to other approaches based on the same principle [19, 29, 35]. Due to this assumption, we excluded sequences recorded at night from our training dataset (see supplementary material). Secondly, akin to many other optical flow methods, our approach is susceptible to the aperture problem. This indicates that only motion components normal to the orientation of an edge in the image space, also known as normal optical flow, can be reliably resolved [10]. Consequently, the proposed method might face challenges in accurately determining the true motion direction in certain ambiguous scenarios. The regularizing effect of the iterative event warping (see Section 3.2) and the multiple spatial scales at which dense optical flow is estimated in our architecture (see Section 3.4) are mechanisms in our proposed solution that collectively strive to counteract the aperture problem's influence.", + "bbox": [ + 75, + 160, + 472, + 553 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 566, + 194, + 580 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we presented the first learning-based approach to event-based optical flow estimation that is scalable to high inference frequencies while being able to accurately capture the true trajectory of scene points over time. The proposed pipeline is designed around a continuously-running stateful model that sequentially processes fine discrete partitions of the input event stream while integrating spatiotemporal information. We train this model through a novel, self-supervised, contrast maximization framework (i.e., event deblurring for supervision) that is characterized by an iterative event warping module and a multi-timescale loss function that add robustness and improve the accuracy of the predicted optical flow maps. We demonstrated the effectiveness of our approach on multiple datasets, where our models outperform the self-supervised and model-based baselines by large margins. Future research should look into how to learn to better combine the information from multiple timescales, as well as into the design of lightweight architectures that can keep up with real-time constraints.", + "bbox": [ + 75, + 590, + 472, + 880 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We believe that the proposed approach opens up avenues for future research, especially in the field of neuromorphic computing. Spiking networks running on neuromorphic hardware have the potential of exploiting the main benefits of event cameras, but for that they need to process the input events shortly after they arrive. Our proposed framework is a step toward this objective, as it enables the estimation of optical flow in a close to continuous manner, with all the integration of information happening within the model itself.", + "bbox": [ + 496, + 90, + 893, + 229 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 243, + 596, + 258 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Himanshu Akolkar, Sio-Hoi Ieng, and Ryad Benosman. Real-time high speed motion prediction using fast aperture-robust event-driven visual flow. IEEE Trans. Pattern Anal. Mach. Intell., 44(1):361-372, 2020. 2, 8", + "[2] Mohammed Almatrafi, Raymond Baldwin, Kiyoharu Aizawa, and Keigo Hirakawa. Distance surface for event-based optical flow. IEEE Trans. Pattern Anal. and Mach. Intell., 42(7):1547-1556, 2020. 2", + "[3] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. Int. Conf. Learn. Representations, 2015. 6", + "[4] Patrick Bardow, Andrew J. Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In IEEE Conf. Comput. Vis. Pattern Recog., pages 884-892, 2016. 2", + "[5] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE Trans. Neural Netw. Learn. Syst., 25(2):407-417, 2013. 2", + "[6] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, 2012. 2", + "[7] Vincent Brebion, Julien Moreau, and Franck Davoine. Realtime optical flow for vehicular perception with low-and high-resolution event cameras. IEEE Trans. on Intell. Transp. Systems, 2021. 2, 8", + "[8] Tobias Brosch, Stephan Tschechne, and Heiko Neumann. On event-based optical flow detection. Frontiers in Neuroscience, 9:137, 2015. 2", + "[9] Javier Cuadrado, Ulysse Rancon, Benoit R. Cotterreau, Francisco Barranco, and Timothee Masquelier. Optical flow estimation from event-based cameras and spiking neural networks. Frontiers in Neuroscience, 17:1160034, 2023. 6, 7, 8", + "[10] David B. de Jong, Federico Paredes-Vallés, and Guido C. H. E. de Croon. How do neural networks estimate optical flow? A neuropsychology-inspired study. IEEE Trans. Pattern Anal. Mach. Intell., 44(11):8290-8305, 2021. 9", + "[11] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-temporal recurrent networks for event-based optical flow estimation. In Proc. AAAI Conf. on Artificial Intell., volume 36, pages 525-533, 2022. 2, 8" + ], + "bbox": [ + 501, + 267, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9703", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. FlowNet: Learning optical flow with convolutional networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2758-2766, 2015. 2", + "[13] Guillermo Gallego, Tobi Delbruck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Trans. Pattern Anal. and Mach. Intell., 2020. 1, 3", + "[14] Guillermo Gallego, Mathias Gehrig, and Davide Scaramuzza. Focus is all you need: Loss functions for event-based vis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12280-12289, 2019. 2, 3, 9", + "[15] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3867-3876, 2018. 2, 3, 9", + "[16] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A stereo event camera dataset for driving scenarios. IEEE Robot. and Autom. Lett., 6(3):4947-4954, 2021. 6", + "[17] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In Int. Conf. 3D Vis., pages 197–206. IEEE, 2021. 2, 6, 7, 8", + "[18] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from events and frames. arXiv:2203.13674, 2022. 2, 3, 6", + "[19] Jesse J. Hagenaars, Federico Paredes-Valles, and Guido C. H. E. de Croon. Self-supervised learning of event-based optical flow with spiking neural networks. In Advances in Neural Information Process. Syst., volume 34, pages 7167–7179, 2021. 2, 3, 4, 5, 6, 7, 8, 9", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 6", + "[21] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Int. Conf. Learn. Representations, 2014. 6", + "[22] Yijin Li, Zhaoyang Huang, Shuo Chen, Xiaoyu Shi, Hongsheng Li, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Blinkflow: A dataset to push the limits of event-based optical flow estimation. arXiv:2303.07716, 2023. 6, 7", + "[23] Haotian Liu, Guang Chen, Sanqing Qu, Yanping Zhang, Zhijun Li, Alois Knoll, and Changjun Jiang. TMA: Temporal motion aggregation for event-based optical flow. arXiv:2303.11629, 2023. 6, 7, 8", + "[24] Min Liu and Tobi Delbruck. Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. In Brit. Mach. Vis. Conf., 2018. 2", + "[25] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proc. AAAI Conf. on Artificial Intell., volume 32, 2018. 5" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Anton Mitrokhin, Cornelia Fermüller, Chethan Parameshwara, and Yiannis Aloimonos. Event-based moving object detection and tracking. In IEEE/RSJ Int. Conf. Intell. Robots and Syst., 2018. 2", + "[27] Jun Nagata, Yusuke Sekikawa, and Yoshimitsu Aoki. Optical flow estimation by matching time surface with event-based cameras. Sensors, 21(4):1150, 2021. 2", + "[28] Garrick Orchard, Ryad Benosman, Ralph Etienne-Cummings, and Nitish V. Thakor. A spiking neural network architecture for visual motion estimation. In IEEE Biomedical Circuits Syst. Conf., pages 298-301. IEEE, 2013. 2", + "[29] Federico Paredes-Valles and Guido C. H. E. de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2, 3, 4, 5, 8, 9", + "[30] Federico Paredes-Vallés, Kirk Y. W. Scheper, and Guido C. H. E. de Croon. Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE Trans. Pattern Anal. and Mach. Intell., 42(8):2051-2064, 2020. 2", + "[31] Wachirawit Ponghiran, Chamika Mihiranga Liiyanagedera, and Kaushik Roy. Event-based temporally dense optical flow estimation with sequential neural networks. arXiv:2210.01244, 2022. 2", + "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Int. Conf. Medical Image Computing and Comput.-Assisted Intervention, pages 234–241. Springer, 2015. 2", + "[33] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event collapse in contrast maximization frameworks. Sensors, 22(14):5190, 2022. 8", + "[34] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. A fast geometric regularizer to mitigate event collapse in the contrast maximization framework. arXiv:2212.07350, 2022. 7", + "[35] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In Eur. Conf. Comput. Vis., 2022. 2, 3, 4, 5, 6, 7, 8, 9", + "[36] Timo Stoffregen and Lindsay Kleeman. Event cameras, contrast maximization and reward functions: An analysis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12300-12308, 2019. 3", + "[37] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In European Conf. Comput. Vis., 2020. 2, 6, 8", + "[38] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. PWC-Net: CNNs for optical flow using pyramid, warping, and cost volume. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8934–8943, 2018. 2", + "[39] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In Eur. Conf. Comput. Vis., pages 402-419. Springer, 2020. 2", + "[40] Z Wan, Y Dai, and Y Mao. Learning dense and continuous optical flow from an event camera. IEEE Trans. Image Process., 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9704", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Qimin Wang, Yongjun Zhang, Shaowu Yang, Zhe Liu, Lin Wang, and Luoxi Jing. E-HANet: Event-based hybrid attention network for optical flow estimation. In IEEE Int. Conf. Syst., Man, Cybernetics, pages 2189-2194. IEEE, 2022. 2", + "[42] Yilun Wu, Federico Paredes-Vallés, and Guido C. H. E. de Croon. Rethinking event-based optical flow: Iterative deblurring as an alternative to correlation volumes. arXiv:2211.13726, 2022. 2, 4, 6, 7", + "[43] Alex Z. Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3D perception. IEEE Robot. and Autom. Lett., 3(3):2032-2039, 2018. 2, 6, 8", + "[44] Alex Z. Zhu and Liangzhe Yuan. EV-FlowNet: Self-supervised optical flow estimation for event-based cameras. In Robot.: Science and Syst., 2018. 2, 6, 8", + "[45] Alex Z. Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 989-997, 2019. 2, 3, 4, 5, 6, 8" + ], + "bbox": [ + 78, + 90, + 468, + 373 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9705", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/ff7524e3-aa7d-4baa-a842-04712fa34f60_model.json b/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/ff7524e3-aa7d-4baa-a842-04712fa34f60_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e15006a12348d169562010bf8348d4f9992431be --- /dev/null +++ b/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/ff7524e3-aa7d-4baa-a842-04712fa34f60_model.json @@ -0,0 +1,1905 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.227, + 0.131, + 0.744, + 0.177 + ], + "angle": 0, + "content": "Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.203, + 0.91, + 0.223 + ], + "angle": 0, + "content": "Federico Paredes-Valles\\(^{1,2}\\) Kirk Y. W. Scheper\\(^{2}\\) Christope De Wagter\\(^{1}\\) Guido C. H. E. de Croon\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.227, + 0.753, + 0.246 + ], + "angle": 0, + "content": "1 Micro Air Vehicle Laboratory, Delft University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.246, + 0.825, + 0.264 + ], + "angle": 0, + "content": "2 Stuttgart Laboratory 1, Sony Semiconductor Solutions Europe, Sony Europe B.V." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.291, + 0.511, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.293, + 0.891, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.533, + 0.894, + 0.577 + ], + "angle": 0, + "content": "Figure 1: Our pipeline estimates event-based optical flow by sequentially processing small partitions of the event stream with a recurrent model. We propose a novel self-supervised learning framework (left) based on a multi-timescale contrast maximization formulation that is able to exploit the high temporal resolution of event cameras via iterative warping to produce accurate optical flow predictions (right)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.588, + 0.314, + 0.603 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.619, + 0.474, + 0.892 + ], + "angle": 0, + "content": "Event cameras have recently gained significant traction since they open up new avenues for low-latency and low-power solutions to complex computer vision problems. To unlock these solutions, it is necessary to develop algorithms that can leverage the unique nature of event data. However, the current state-of-the-art is still highly influenced by the frame-based literature, and usually fails to deliver on these promises. In this work, we take this into consideration and propose a novel self-supervised learning pipeline for the sequential estimation of event-based optical flow that allows for the scaling of the models to high inference frequencies. At its core, we have a continuously-running stateful neural model that is trained using a novel formulation of contrast maximization that makes it robust to nonlinearities and varying statistics in the input events. Results across multiple datasets confirm the effectiveness of our method, which establishes a new state of the art in terms of accuracy for approaches trained or optimized without ground truth." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.588, + 0.633, + 0.603 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.856 + ], + "angle": 0, + "content": "Event cameras capture per-pixel log-brightness changes at microsecond resolution [13]. This operating principle results in a sparse and asynchronous visual signal that, under constant illumination, directly encodes information about the apparent motion (i.e., optical flow) of contrast in the image space. These cameras offer several advantages, such as low latency and robustness to motion blur [13], and hence hold the potential of a high-bandwidth estimation of this optical flow information. However, the event-based nature of the generated visual signal poses a paradigm shift in the processing pipeline, and traditional, frame-based algorithms become suboptimal and often incompatible. Despite this, the majority of learning-based methods that have been proposed so far for event-based optical flow estimation are still highly influenced by frame-based approaches. This influence is normally reflected in two key aspects of their" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The project's code and additional material can be found at https://mavlab.tudelft.nl/taming_event_flow/." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9695" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "pipelines: (i) the design of the network architecture, and (ii) the formulation of the loss function." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.124, + 0.468, + 0.304 + ], + "angle": 0, + "content": "Regarding architecture design, most literature methods format subsets of the input events as dense volumetric representations [45] that are processed at once by stateless (i.e., non-recurrent) models [17, 29, 37, 45]. Similarly to their frame-based counterparts [12, 38, 39], these models estimate the per-pixel displacement over the time-length of the event volume using only the information contained within it. Consequently, these volumes need to encode enough spatiotemporal information for motion to be discernible. However, if done over relatively long time periods, the subsequent models suffer from limitations such as high latency or having to deal with large pixel displacements [17, 18, 42]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.468, + 0.548 + ], + "angle": 0, + "content": "With respect to the loss function, multiple options have been explored due to the lack of real-world datasets with per-event ground truth. Pure supervised learning can be used with datasets such as MVSEC [43] or DSEC-Flow [17], but their ground truth only contains per-pixel displacement at low frequencies, which makes models difficult to train. On the other hand, a self-supervised learning (SSL) framework can be formulated using either the accompanying frames with the photometric error as a loss [11, 40, 44], or through an events-only contrast maximization [14, 15] proxy loss for motion compensation (i.e., event deblurring) [19, 26, 29, 35, 45]. However, despite not relying on ground truth, all the literature on SSL for optical flow assumes that the events move linearly within the time window of the loss, which ignores much of the potential of event cameras and their high temporal resolution (see Fig. 1, bottom left)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.468, + 0.791 + ], + "angle": 0, + "content": "In this work, we focus on the estimation of high frequency event-based optical flow and how this can be learned in an SSL fashion using contrast maximization with a relaxed linear motion assumption. To achieve this, we build upon the continuous-operation pipeline from Hagenaars et al. [19], which retrieves optical flow by sequentially processing small partitions of the event stream with a stateful (i.e., recurrent) model over time, instead of dealing with large volumes of input events. We augment (and train) this pipeline with a novel contrast maximization formulation that performs event motion compensation in an iterative manner at multiple temporal scales, as shown in Fig. 1. Using this framework, we achieve the best accuracy of all contrast-maximization-based approaches on multiple datasets, only being outperformed by pure supervised learning methods trained with ground-truth data." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.794, + 0.468, + 0.823 + ], + "angle": 0, + "content": "In summary, the extensions that we propose to the SSL framework in [19], i.e., our main contributions, are:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- The first iterative event warping module in the context of contrast maximization (see Section 3.2). This module unlocks a novel multi-reference loss function that better captures the trajectory of scene points over time, thus improving the accuracy of the predictions." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.891, + 0.151 + ], + "angle": 0, + "content": "- The first multi-timescale approach to contrast maximization, which adds robustness, improves convergence, and reduces tuning requirements of loss-related hyperparameters (see Section 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.153, + 0.892, + 0.273 + ], + "angle": 0, + "content": "As a result, we present the first self-supervised optical flow method for event cameras that relaxes the linear motion assumption, and hence that has the potential of exploiting the high temporal resolution of the sensor by producing estimates in a close-to continuous manner. We validate the proposed framework through extensive evaluations on multiple datasets. Additionally, we conduct ablation studies to show the effectiveness of each individual component." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.287, + 0.637, + 0.303 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.313, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Due to the aforementioned advantages of event cameras for optical flow estimation, extensive research has been carried out since these sensors were first introduced [1, 2, 4-8, 24, 27, 28, 35]. Regarding learning-based approaches, the first method was proposed by Zhu et al. in [44] with EV-FlowNet: a UNet-like [32] architecture trained with SSL, with the supervisory signal coming from the photometric error between subsequent frames captured with an accompanying camera. To avoid the need for a secondary vision sensor, in [45], Zhu et al. proposed an SSL framework around the contrast maximization for motion compensation idea from [14, 15], hence relying solely on event data. This pipeline was used and further improved in [19, 29, 35]. Other approaches trained EV-FlowNet in a supervised fashion with synthetic/real ground-truth data and showed higher accuracy levels through evaluations on public benchmarks [17, 37]. However, they also highlighted EV-FlowNet's inability to deal with large pixel displacements [17]. Because of this, inspired by the frame-based literature [39], Gehrig et al. proposed E-RAFT in [17]: the first architecture to introduce the use of correlation volumes in the event camera literature. This model, which was recently augmented with attention mechanisms [41], achieved state-of-the-art performance in multiple datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "A novel perspective to learning event-based optical flow is to move away from event accumulation over long timespans, and instead rely on continuously-running stateful models that integrate information over time. The first methods of this kind were proposed by Paredes-Valles et al. [30] and Hagenaars et al. [19], but this idea has recently gained interest in the event camera literature [11, 31, 42]. The reason is that leveraging memory through sequential processing can potentially lead to lightweight and low-latency solutions that are also robust to large pixel displacements without the need for correlation volumes [42]. However, despite their potential of being scaled to high inference frequencies, all these solutions assume that events move linearly in the timespan of their loss function, and hence cannot capture the true trajectory of scene points over time." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9696" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "When it comes to learning nonlinear pixel trajectories from event data, only the work of Gehrig et al. in [18] is to be highlighted. However, their approach uses multiple event and correlation volumes to fit Bézier curves to the trajectory of scene points, resulting in a high-latency solution. In contrast, in this work, we extend the continuous-operation pipeline from Hagenaars et al. [19] with an SSL framework in which discrete trajectories are regressed at high frequency by leveraging memory within the models. This approach allows, for the first time, to exploit the high temporal resolution of event cameras while capturing more accurately the trajectory of scene points over time thanks to the proposed iterative event warping mechanism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.473, + 0.487 + ], + "angle": 0, + "content": "Among non-learning-based approaches, the work of Shiba et al. in [35] bears particular relevance due to certain similarities with our framework. Specifically, Shiba et al. propose a tile-based method for event-based optical flow estimation that extends contrast maximization [14, 15] by incorporating (i) multiple spatial scales, (ii) a multi-reference focus loss, and (iii) a \"time-aware\" optical flow formulation. Similarly, our learning-based approach also employs a multi-scale strategy and utilizes a multi-reference focus loss. However, instead of making assumptions about the events' motion over extended time periods, we propose to learn their potentially nonlinear trajectories at high frequency by leveraging the iterative event warping module." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.503, + 0.17, + 0.518 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.473, + 0.696 + ], + "angle": 0, + "content": "The goal of this work is to learn to sequentially estimate optical flow at high frequencies from a continuous stream of events. In such a pipeline, if the inference frequency is sufficiently high, events need to be processed nearly as soon as they are triggered by the sensor, with no pre-processing in between. Because of this, the integration of temporal information needs to happen in the network itself. To accomplish this, we propose the framework in Fig. 1, in which a stateful model is trained using our novel formulation of contrast maximization for sequential processing. The components of this framework are described in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.709, + 0.424, + 0.725 + ], + "angle": 0, + "content": "3.1. Input format and contrast maximization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.809 + ], + "angle": 0, + "content": "For an ideal camera, an event \\( \\boldsymbol{e}_i = (\\boldsymbol{x}_i, t_i, p_i) \\) of polarity \\( p_i \\in \\{+, -\\} \\) is triggered at pixel \\( \\boldsymbol{x}_i = (x_i, y_i)^T \\) and time \\( t_i \\) whenever the change in log-brightness since the last event at that pixel location reaches the contrast sensitivity threshold for that polarity [13]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As in [19], we use a two-channel event count image as input representation, which gets populated with consecutive, non-overlapping, fine discrete partitions of the event stream \\(\\varepsilon_{k}^{\\mathrm{inp}}\\doteq \\{e_{i}\\}\\) (further referred to as input partitions), each containing all the events in a time window of a certain duration, i.e., \\(t_i\\in [t_k^{\\mathrm{begin}},t_k^{\\mathrm{end}}]\\). This representation does not" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "contain temporal information by itself, and recurrent models are hence required for estimating of optical flow." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.894, + 0.273 + ], + "angle": 0, + "content": "Regarding learning, we use the contrast maximization framework [14] to train, in an SSL fashion, to continuously estimate dense (i.e., per-pixel) optical flow from the event stream. Assuming brightness constancy, accurate flow information is encoded in the spatiotemporal misalignments (i.e., blur) among the events triggered by the same portion of a moving edge. To retrieve it, one has to compensate for this motion by geometrically transforming the events using a motion model. As in [19, 29, 35, 45], we transport each event to a reference time \\( t_{\\mathrm{ref}} \\) through:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.284, + 0.892, + 0.301 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {i} ^ {\\prime} = \\boldsymbol {x} _ {i} + \\left(t _ {\\text {r e f}} - t _ {i}\\right) \\boldsymbol {u} \\left(\\boldsymbol {x} _ {i}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.372 + ], + "angle": 0, + "content": "where \\(\\pmb{u}(\\pmb{x}) = (u(\\pmb{x}), v(\\pmb{x}))^T\\) denotes the optical flow map used to transport each event from \\(t_i\\) to \\(t_{\\mathrm{ref}}\\). The result of aggregating the transformed events is further referred to as the image of warped events (IWE) at \\(t_{\\mathrm{ref}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.894, + 0.448 + ], + "angle": 0, + "content": "As the loss function of our SSL framework, we adopt the time-based focus objective function from [19]. Using the warped events at a given \\( t_{\\mathrm{ref}} \\), we generate an image of the per-pixel average timestamps for each polarity \\( p' \\) via bilinear interpolation:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.457, + 0.892, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} T _ {p ^ {\\prime}} (\\boldsymbol {x}; \\boldsymbol {u} | t _ {\\text {r e f}}) = \\frac {\\sum_ {j} \\kappa \\left(x - x _ {j} ^ {\\prime}\\right) \\kappa \\left(y - y _ {j} ^ {\\prime}\\right) \\bar {t} _ {j} \\left(t _ {\\text {r e f}} , t _ {j}\\right)}{\\sum_ {j} \\kappa \\left(x - x _ {j} ^ {\\prime}\\right) \\kappa \\left(y - y _ {j} ^ {\\prime}\\right) + \\epsilon} \\tag {2} \\\\ \\kappa (a) = \\max (0, 1 - | a |) \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.516, + 0.816, + 0.533 + ], + "angle": 0, + "content": "\\[\nj = \\{i \\mid p _ {i} = p ^ {\\prime} \\}, \\quad p ^ {\\prime} \\in \\{+, - \\}, \\quad \\epsilon \\approx 0\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.894, + 0.574 + ], + "angle": 0, + "content": "where \\(\\bar{t}_i\\) denotes the normalized timestamp contribution of the \\(i\\)th event, according to Fig. 3 and Eq. 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.605 + ], + "angle": 0, + "content": "Then, the contrast maximization loss function at \\( t_{\\mathrm{ref}} \\) is defined as the scaled sum of the squared temporal images:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.614, + 0.892, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C M}} \\left(t _ {\\text {r e f}}\\right) = \\frac {\\sum_ {\\boldsymbol {x}} T _ {+} \\left(\\boldsymbol {x} ; \\boldsymbol {u} \\mid t _ {\\text {r e f}}\\right) ^ {2} + T _ {-} \\left(\\boldsymbol {x} ; \\boldsymbol {u} \\mid t _ {\\text {r e f}}\\right) ^ {2}}{\\sum_ {\\boldsymbol {x}} \\left[ n \\left(\\boldsymbol {x} ^ {\\prime}\\right) > 0 \\right] + \\epsilon} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.891, + 0.69 + ], + "angle": 0, + "content": "where \\(n(\\pmb{x}^{\\prime})\\) denotes a per-pixel event count of the IWE. The lower the \\(\\mathcal{L}_{\\mathrm{CM}}\\), the better the event deblurring at \\(t_{\\mathrm{ref}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.901 + ], + "angle": 0, + "content": "As discussed in [14, 19, 36], for any focus objective function to be a robust supervisory signal for contrast maximization, the event partition used in the optimization needs to contain enough motion information (i.e., blur) so it can be compensated for. However, as in [19], that is not the case in our input partitions due to the fine discretization of the event stream that we are targeting. Therefore, only at training time, we define the so-called training partition (or event buffer in Fig. 1) \\(\\varepsilon_{k\\rightarrow k + R}^{\\mathrm{train}}\\doteq \\{\\varepsilon_i^{\\mathrm{inp}}\\}_{i = k}^{k + R}\\), which stacks together the events in the input partitions of \\(R\\) successive forward passes through the network. Once these are performed, we use this partition and the estimated optical flow maps to compute the loss, and use truncated backpropagation through time to update the model parameters. After" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9697" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.452, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.738 + ], + "angle": 0, + "content": "Figure 2: Incompatibility of multi-reference deblurring and linear event warping in the presence of nonlinearities in the pixel trajectories. Top: Events generated by a moving dot following circular motion (left), and optical flow color-coding scheme (right). Middle: Optical flow solutions required to produce sharp IWEs at different \\( t_{\\mathrm{ref}} \\) using linear and iterative warping. While the former requires a different solution for each \\( t_{\\mathrm{ref}} \\), the proposed iterative warping can achieve this using a single solution. Arrows illustrate the direction of the required displacement \\( \\Delta \\pmb{x}_i = (t_{\\mathrm{ref}} - t_i)\\pmb{u}(\\pmb{x}_i) \\) at each (discretized) spatial location. Bottom: Resulting IWEs at different \\( t_{\\mathrm{ref}} \\) using the optimal optical flow map for \\( t_{\\mathrm{ref}} = \\mathrm{mid} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.468, + 0.839 + ], + "angle": 0, + "content": "this update, we detach the states of the network from the computational graph and clear the training partition and optical flow buffer. The importance of sequential processing, as an alternative to training stateless models in short input partitions, is corroborated in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.294, + 0.865 + ], + "angle": 0, + "content": "3.2. Iterative event warping" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In order to better approximate the trajectories of scene points over time, in this work we relax the linear motion as-" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.888, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.21, + 0.892, + 0.294 + ], + "angle": 0, + "content": "Figure 3: Timestamp normalization profiles for the per-event contributions to the images of average timestamps in Eq. 2, for \\( R = 10 \\). Left: Deblurring only done at the extremes of the training partition, as in [19, 29, 45]. Right: Deblurring done at all reference times, thanks to the proposed iterative event warping. Normalization profiles only shown for three \\( t_{\\mathrm{ref}} \\) for a better visualization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.344, + 0.893, + 0.602 + ], + "angle": 0, + "content": "sumption in the SSL framework for event-based optical flow by augmenting the sequential-estimation pipeline from [19] with iterative event warping. Instead of transporting events to a given \\( t_{\\mathrm{ref}} \\) assuming linear motion regardless of the length of the warping interval (as in [19, 29, 35, 45]), we perform a finer discretization of the event trajectories and assume that motion is only linear between optical flow estimates. Therefore, to express a group of events at a given \\( t_{\\mathrm{ref}} \\), we geometrically transform them using all the intermediate optical flow estimates through multiple iterations of Eq. 1. Fig. 2 shows the differences between the linear event warping in [19] and the proposed iterative augmentation, as well as the limitations of the former. Note that our iterative warping is fundamentally different from that of the work of Wu et al. in [42], where input events are deblurred before being passed to the models using residual optical flow estimates until convergence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Literature methods on event-based optical flow with contrast maximization compute the focus objective function at multiple reference times in the training partition (usually at the extremes) to prevent overfitting and/or scaling issues during backpropagation [19, 29, 35, 45]. However, since these approaches assume optical flow constancy in the span of their loss function, they suffer with nonlinearities in the pixel trajectories (see the blurry IWEs in Fig. 2, bottom). On the contrary, because of the finer discretization of the event trajectories coming with our sequential processing pipeline and the proposed iterative warping, we can use any (combination of) reference time(s) for the computation of the focus objective function. In fact, as shown in Fig. 3 (right), we propose the use of all the discretization points as reference times for event deblurring. Apart from the aforementioned regularizing benefits, having to produce sharp IWEs at any \\( t_{\\mathrm{ref}} \\) forces the models to estimate a sequence of optical flow maps that is consistent with the velocity profile of the event stream. For a given training partition of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9698" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.097, + 0.482, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.32, + 0.474, + 0.446 + ], + "angle": 0, + "content": "Figure 4: Multi-timescale approach to contrast maximization. For a given training partition of length \\( R \\), we fit multiple sub-partitions of different lengths (in the figure: one of length \\( R \\) in red, two \\( R/2 \\) in blue, and four \\( R/4 \\) in green) and compute the loss in each of them according to Eq. 4. The global loss is computed as in Eq. 6. This figure only shows the timestamp normalization profiles of the central \\( t_{\\mathrm{ref}} \\) of each sub-partition, but the losses are still computed at all reference times. An image representation of the accumulated input events in a sub-partition of each timescale is also shown." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.471, + 0.357, + 0.487 + ], + "angle": 0, + "content": "length \\(R\\), the loss is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.497, + 0.47, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C M}} ^ {R} = \\frac {1}{R + 1} \\sum_ {t _ {\\text {r e f}} = 0} ^ {R} \\mathcal {L} _ {\\mathrm {C M}} \\left(t _ {\\text {r e f}}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.55, + 0.47, + 0.595 + ], + "angle": 0, + "content": "which ensures that the IWEs (and the corresponding images of average timestamps in Eq. 2) at all reference times \\( t_{\\mathrm{ref}} \\in [0, R] \\) contribute equally to the loss." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.595, + 0.47, + 0.701 + ], + "angle": 0, + "content": "As in [19, 29, 45], for \\(\\mathcal{L}_{\\mathrm{CM}}(t_{\\mathrm{ref}})\\) in Eq. 3 to be a valid supervisory signal, events that are temporally close to the reference time \\(t_{\\mathrm{ref}}\\) need to contribute to the temporal image in Eq. 2 with larger timestamp values than events that are far in time. Therefore, once the events have been transported to \\(t_{\\mathrm{ref}}\\), their timestamp is normalized prior to the computation of Eq. 2 as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.711, + 0.47, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\bar {t} _ {i} \\left(t _ {\\text {r e f}}, t _ {i}\\right) = 1 - \\frac {\\left| t _ {\\text {r e f}} - t _ {i} \\right|}{R}, \\quad \\text {w i t h} t _ {i} \\in [ 0, R ] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.47, + 0.78 + ], + "angle": 0, + "content": "which results in the normalization profiles in Fig. 3, for a given training partition of length \\( R \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Inspired by [25], we mask individual events from the computation of the loss whenever we detect that they are transported outside the image space through the event warping process in the span of a deblurring window defined at the reference time \\( t_{\\mathrm{ref}} \\). This prevents our models from learning incorrect deformations at the image borders, and is motivated by the fact that the complete trajectory of the pixel is only partially observable in the training partition. An ab" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "lation study on the impact of this masking strategy can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.122, + 0.894, + 0.213 + ], + "angle": 0, + "content": "Lastly, we do not augment the loss function in Eq. 4 with smoothing priors acting as regularization mechanisms. With iterative event warping, the error propagates through all the pixels covered in the warping process, regardless of whether they have input events or not. Therefore, the spatial coherence of the resulting optical flow maps is enhanced." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.223, + 0.794, + 0.239 + ], + "angle": 0, + "content": "3.3. Deblurring at multiple timescales" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.246, + 0.892, + 0.533 + ], + "angle": 0, + "content": "Despite the addition of iterative event warping, the success of our SSL framework still heavily depends on the hyperparameters that control the amount of motion information perceived by the networks in the span of a deblurring window. In our pipeline, these are: \\(\\mathrm{dt}_{\\mathrm{input}}\\), the timestep used to discretize the event stream; and \\(R\\), the number of forward passes, and hence optical flow maps, per loss. Thus, the effective length (in units of time) of the event window used for motion compensation is given by \\(\\mathrm{dt}_{\\mathrm{input}} \\times R\\). We hypothesize that, for each training dataset, there is an optimal length for this window that depends on the statistics of the data (e.g., event density, distribution of optical flow magnitudes) and model architecture, and that deviations from this optimal length lead to the learning of suboptimal solutions. E.g., shorter windows may converge to solutions that are more selective to fast rather than slow moving objects, and vice versa. Note that not only our method is sensitive to the tuning of these parameters, but also previous approaches based on contrast maximization [19, 29, 35, 45]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.894, + 0.639 + ], + "angle": 0, + "content": "To add a layer of robustness to the framework and relax its strong dependency on hyperparameter optimization, we propose the multi-timescale approach illustrated in Figs. 1 and 4. For a given training partition of length \\( R \\), instead of computing a single focus loss through Eq. 4, we compute this loss at \\( S \\) temporal scales of length \\( R / 2^s \\), with \\( 0 \\leq s \\leq S - 1 \\), and combine them as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.651, + 0.892, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C M}} ^ {\\text {m u l t i}} = \\frac {1}{S} \\sum_ {s = 0} ^ {S - 1} \\frac {1}{2 ^ {s}} \\sum_ {p = 0} ^ {2 ^ {s} - 1} \\mathcal {L} _ {\\mathrm {C M}, p} ^ {R / 2 ^ {s}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.75 + ], + "angle": 0, + "content": "As shown, we fit multiple non-overlapping sub-partitions in the training buffer if \\( s > 0 \\). The subscript \\( p \\) indicates their location in this buffer, starting from the earliest (see Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Note that, through this multi-timescale approach to contrast maximization, our models need to converge to a solution that is suitable for all the timescales in the optimization, regardless of their length. An alternative formulation would be to incorporate per-pixel learnable masks (i.e., an attention module in the loss space) so that, depending on the input statistics, learning only happens at the most adequate scale. However, for this to happen, the loss function would have to be augmented to stimulate this behavior, and it is unclear how that would be done in practice." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9699" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.46, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.19, + 0.47, + 0.248 + ], + "angle": 0, + "content": "Figure 5: Schematic of the model architecture used in this work. It is characterized by \\( N_{\\mathcal{G}} \\) recurrent encoders, \\( N_{\\mathcal{R}} \\) residual blocks, and \\( N_{\\mathcal{G}} \\) decoder layers. Optical flow estimates are produced at all decoder levels. In this diagram, \\( N_{\\mathcal{G}} = 2 \\) and \\( N_{\\mathcal{R}} = 1 \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.272, + 0.279, + 0.286 + ], + "angle": 0, + "content": "3.4. Network architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.473, + 0.522 + ], + "angle": 0, + "content": "We use the recurrent version of EV-FlowNet [44] proposed in \\([19]^1\\) (see Fig. 5). The events are represented as event count images (see Section 3.1), then passed through four encoders with strided convolutions followed by ConvGRUs [3] (channels doubling, starting from 64), two residual blocks [20], and then four decoders performing bilinear upsampling followed by convolution. After each decoder, there is a skip connection (using element-wise summation) from the corresponding encoder, as well as a depthwise convolution to produce estimates at lower scales, which are then concatenated with the activations of the previous decoder. Note that the proposed focus loss function (see Eq. 6) is applied to each intermediate optical flow estimate via upsampling. Lastly, all layers use \\(3\\times 3\\) kernels and ReLU activations except for the prediction layers, which use TanH." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.21, + 0.552 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.741 + ], + "angle": 0, + "content": "We evaluate our method on the DSEC-Flow [16, 17] and MVSEC [43] datasets. We evaluate the accuracy of the predictions based on the following metrics: (i) EPE (lower is better, ↓), the endpoint error; (ii) \\(\\%_{3\\mathrm{PE}}(\\downarrow)\\), the percentage of points with EPE greater than 3 pixels; (iii) FWL (\\(\\uparrow\\)) [37], a deblurring quality metric based on the variance of the IWEs; and (iv) RSAT (\\(\\downarrow\\)) [19], a deblurring quality metric based on the per-pixel average timestamps of the IWEs. We compare our solution to the published baselines, which range from supervised learning (SL) methods trained with ground truth, to SSL methods trained with grayscale images (\\(\\mathrm{SSL}_{\\mathrm{F}}\\)) or events (\\(\\mathrm{SSL}_{\\mathrm{E}}\\)), and model-based approaches (MB)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.848 + ], + "angle": 0, + "content": "We train all our models on a subset of sequences from the training dataset of DSEC-Flow (only daylight recordings, see supplementary material). This corresponds to 19 minutes of training data, which we split into \\(572128 \\times 128\\) (randomly-cropped) sequences of 2 seconds each. We use a batch size of 8 and train until convergence with the Adam optimizer [21] and a learning rate of 1e-5. To keep mem" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.092, + 0.895, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.234, + 0.895, + 0.29 + ], + "angle": 0, + "content": "Figure 6: Reconstruction of the pixel displacement of a scene point in the time window of a ground-truth sample from the multiple optical flow maps estimated in this period, i.e., \\(\\forall \\pmb{u}_k \\in [t_{\\mathrm{begin}}, t_{\\mathrm{end}}]\\). The error of the last optical flow estimate is magnified for clarity." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.304, + 0.892, + 0.54 + ], + "angle": 0, + "content": "
EPE↓%3PE↓FWL↑RSAT↓
SLE-RAFT [17]0.792.681.330.87
EV-FlowNet, Gehrig et al. [17]2.3218.60--
Gehrig et al. [18]0.752.44--
IDNet [42]0.722.04--
TIDNet [42]0.843.41--
TMA [23]0.742.30--
Cuadrado et al. [9]1.7110.31--
E-Flowformer [22]0.762.68--
SSLEEV-FlowNet* [45]3.8631.451.300.85
ConvGRU-EV-FlowNet* [19]4.2733.271.550.90
dt = 0.01s, R = 2, S = 1 (Ours)9.6686.441.911.07
dt = 0.01s, R = 5, S = 1 (Ours)4.0552.221.580.97
dt = 0.01s, R = 10, S = 1 (Ours)2.3317.771.260.88
dt = 0.01s, R = 20, S = 1 (Ours)16.6333.671.061.10
dt = 0.01s, R = 10, S = 3 (Ours)2.8227.091.370.92
dt = 0.01s, R = 20, S = 4 (Ours)2.7323.731.240.90
MBShiba et al. [35]3.4730.861.370.89
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.527, + 0.54, + 0.747, + 0.552 + ], + "angle": 0, + "content": "*Retrained by us on DSEC-Flow, linear warping." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.563, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation on the DSEC-Flow dataset [17]. Best in bold, runner up underlined. A breakdown of the results is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.626, + 0.892, + 0.687 + ], + "angle": 0, + "content": "ory usage within limits, we only propagate error gradients through up to 1e3 randomly-chosen events per millisecond of data. Despite this, note that we warp and use all the input events for the computation of the loss function." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.704, + 0.713 + ], + "angle": 0, + "content": "4.1. Evaluation procedure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.901 + ], + "angle": 0, + "content": "When evaluating our sequential models, if \\(\\mathrm{dt_{gt}} > \\mathrm{dt_{input}}\\), we need to reconstruct the estimated per-pixel displacement in the ground-truth time window from the multiple optical flow maps estimated in this period. We do this by first averaging the (bilinearly interpolated) optical flow vectors that describe the trajectory of each scene point, and then by scaling the resulting optical flow vectors by \\(\\mathrm{dt_{gt}} / \\mathrm{dt_{input}}\\). This converts them from units of pixels/input to units of pixel displacement. An illustration of this reconstruction is shown in Fig. 6 for a scene point following a nonlinear trajectory. Note that our solution is subject to cumulative errors when evaluated through this reconstruction on benchmarks with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.47, + 0.901 + ], + "angle": 0, + "content": "1Our architecture is equivalent to ConvGRU-EV-FlowNet [19]. However, for the purpose of clarity within the rest of the paper, we will use this term to specifically refer to the original model from [19]." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9700" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.086, + 0.873, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.55, + 0.893, + 0.593 + ], + "angle": 0, + "content": "Figure 7: Qualitative comparison of our method with the state-of-the-art E-RAFT architecture [17] and the model-based approach from Shiba et al. [35] on sequences from the test partition of the DSEC-Flow dataset [17]. Ground truth not included due to unavailability. The optical flow color coding can be found in Fig. 2 (top), and the corresponding IWEs in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.471, + 0.681 + ], + "angle": 0, + "content": "ground truth provided at low rates (e.g., \\(10\\mathrm{Hz}\\) in DSEC-Flow [17]). Therefore, it will compare unfavorably to other, non-sequential, methods that only produce a single optical flow estimate in the timespan of a ground-truth sample." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.291, + 0.71 + ], + "angle": 0, + "content": "4.2. Optical flow evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Evaluation on DSEC-Flow: Quantitative results of our evaluation on DSEC-Flow are presented in Table 1, and are supported by the qualitative comparison in Fig. 7. For this experiment, we trained multiple models with the same \\(\\mathrm{dt}_{\\mathrm{input}} = 0.01\\mathrm{s}\\) (i.e., \\(\\times 10\\) faster than DSEC's ground truth) but different lengths of the training partition, and with and without the multi-timescale approach. Multiple conclusions can be derived from the reported results. Firstly, our best performing model (i.e., \\(R = 10\\), \\(S = 1\\)) achieves the best accuracy of all contrast-maximization-based approaches on this dataset according to the EPE and the percentage of outliers. Specifically, it outperforms the baselines with an im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.754 + ], + "angle": 0, + "content": "provement in the EPE in the \\(33\\% - 45\\%\\) range, only being outperformed by SL methods trained with ground truth on the same dataset [9, 17, 22, 23, 42]. This confirms that (i) the timestamp-based loss function in Section 3.1 allows us to learn accurate event-based optical flow (contrary to the findings of [34]); and that (ii) our augmentations to the sequential pipeline in [19] lead to a significant improvement in the accuracy of the model (i.e., \\(45\\%\\) improvement in the EPE)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Secondly, these results also confirm our hypothesis that, for each training dataset, there is an optimal length for the training partition \\( R \\) in terms of the EPE. According to Table 1, the optimal \\( R \\) for this dataset, our model architecture, and our \\( \\mathrm{dt}_{\\mathrm{input}} \\) is 10 (i.e., 0.1s of event data), with the EPE increasing if the training partition is made shorter or longer. However, as also shown in this table, we can relax the strong dependency of the contrast maximization framework on this parameter through the proposed multi-timescale approach." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9701" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.131, + 0.086, + 0.421, + 0.176 + ], + "angle": 0, + "content": "
EPE↓%3PE↓FWL↑RSAT↓
dt = 0.1s*3.4834.720.980.87
dt = 0.05s3.0927.361.110.91
dt = 0.01s2.3317.771.260.88
dt = 0.005s2.3417.921.380.89
dt = 0.002s2.6621.832.040.90
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.138, + 0.178, + 0.376, + 0.189 + ], + "angle": 0, + "content": "*Non-recurrent, volum. event repr. with 10 bins." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.202, + 0.47, + 0.23 + ], + "angle": 0, + "content": "Table 2: Impact of the input window length on the DSEC-Flow dataset [17]. Best in bold, runner up underlined." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.251, + 0.468, + 0.386 + ], + "angle": 0, + "content": "Our \\( S > 1 \\) models converged to solutions that slightly underperform our best performing single-scale model (EPE went up by \\( 17\\% - 21\\% \\)), but were trained without the need to fine-tune the length of the training partition. Note that, despite this slight drop in accuracy, these multi-timescale solutions still outperform the other non-SL baselines. To further support these results, a visualization of the distribution of the EPE of our models as a function of the ground truth magnitude is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.468, + 0.537 + ], + "angle": 0, + "content": "Lastly, Table 1 also allow us to conclude that deblurring quality metrics FWL [37] and RSAT [19] are not reliable indicators of the quality of the estimated optical flow. The reason for this is their inability to capture \"event collapse\" issues (as described in [33]), and would give favorable scores to undesirable solutions that warp all the events into a few pixels. According to our results, the FWL metric, being the spatial variance of the IWE relative to that of the identity warp, suffers more from this issue: the best FWL value is obtained with a model with 9.66 EPE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.468, + 0.839 + ], + "angle": 0, + "content": "Regarding qualitative results, Fig. 7 shows a comparison of our best performing model with the state-of-the-art E-RAFT architecture [17] and the contrast-maximization-based approach from Shiba et al. [35] on multiple sequences from the test partition of DSEC-Flow (i.e., ground truth is unavailable). These results confirm that our models are able to estimate high quality event-based optical flow despite not having access to ground-truth data during training, and also show the superiority of our method over the current best contrast-maximization-based approach [35]. Two limit cases in which our models provide suboptimal solutions are also shown in this figure: (i) sequences recorded at night (e.g., zurich_city_12_a) due to the presence of large amounts of events triggered by flashing lights and not by motion; and (ii) the car hood, which is also problematic for E-RAFT (i.e., does not capture it) and for [35]. Note that, in our case, (ii) is an artifact of the pixel displacements reconstructed from multiple optical flow estimates, and could be mitigated by having an occlusion handling mechanism in this reconstruction process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In addition to the evaluation in Table 1 and Fig. 7, we also conducted an experiment in which we trained multiple models with different \\(\\mathrm{dt}_{\\mathrm{input}}\\) (ranging from 0.1s to 0.002s) but with the same amount of information in the training par" + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.083, + 0.865, + 0.303 + ], + "angle": 0, + "content": "
EPE↓%3PE↓
SLEV-FlowNet+ [37]0.680.99
E-RAFT [17]0.241.70
EV-FlowNet, Gehrig et al. [17]0.310.00
TMA [23]0.250.07
Cuadrado et al. [9]0.85-
SSLFEV-FlowNet, Zhu et al. [44]0.490.20
Ziluo et al. [11]0.420.00
SSLEEV-FlowNet, Zhu et al. [45]0.320.00
EV-FlowNet, Paredes-Vallés et al. [29]0.925.40
EV-FlowNet, Shiba et al. [35]0.360.09
ConvGRU-EV-FlowNet [19]0.470.25
Ours0.270.05
MBAkolkar et al. [1]2.75-
Brebion et al. [7]0.530.20
Shiba et al. [35]0.300.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.313, + 0.892, + 0.356 + ], + "angle": 0, + "content": "Table 3: Quantitative evaluation on MVSEC's outdoor_1 sequence [43]. Best in bold, runner up underlined. Results on the indoor sequences can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.552 + ], + "angle": 0, + "content": "tition: 0.1s of event data. Results in Table 2 show that our sequential pipeline leads to an improvement in the accuracy of the predicted optical flow maps with respect to the stateless EV-FlowNet, which processes the 0.1s of event data at once. This improvement is due to the fact that the complexity of dealing with large pixel displacements gets reduced when processing the input data sequentially using shorter input windows. In addition to this, Table 2 also shows that the accuracy of our models is not compromised when estimating optical flow at higher frequencies, despite the high sparsity levels in the input data at those rates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation on MVSEC: Quantitative results of our evaluation on the outdoor_day1 sequence from MVSEC are presented in Table 3, and are supported by the qualitative comparison in the supplementary material. For this experiment, since (i) there is no consensus in the literature with respect to the training dataset [17, 19, 29, 35, 37, 44, 45], and (ii) the outdoor_day2 sequence (i.e., the other daylight, automotive sequence) is only 9 minutes of duration during which the event camera is subject to high frequency vibrations [43], we decided to transfer one of our models trained on DSEC-Flow to MVSEC. More specifically, we chose the model trained with \\(\\mathrm{dt}_{\\mathrm{input}} = 0.005\\mathrm{s}\\) and \\(R = 20\\) from Table 2, as a model trained with a short input window on DSEC-Flow is expected to be robust to the slow motion statistics of MVSEC [17]. We deployed the model at the same frequency as the temporally-upsampled ground truth (i.e., \\(45\\mathrm{Hz}\\)). Results in Table 3 show that, by doing this, our model outperforms the great majority of methods in terms of the EPE (even some SL methods trained on this dataset), and is only surpassed by the current state-of-the-art E-RAFT architecture [17]. Besides reaffirming the high quality of the produced optical flow estimates, these results also confirm the generalizability of our method. For com" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9702" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "pleteness, the results on the indoor evaluation sequences from MVSEC are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.135, + 0.201, + 0.151 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.161, + 0.473, + 0.554 + ], + "angle": 0, + "content": "The self-supervised method for event-based optical flow presented in this work, while demonstrating highly accurate and promising results, is not without limitations. Two critical challenges that need to be acknowledged are the brightness constancy assumption and the aperture problem. Firstly, the contrast maximization framework [14, 15] assumes constant illumination, leading our models to face difficulty in learning from events that are not due to motion in the image space but that arise from changes in illumination. Since this limitation is inherent to contrast maximization, it extends to other approaches based on the same principle [19, 29, 35]. Due to this assumption, we excluded sequences recorded at night from our training dataset (see supplementary material). Secondly, akin to many other optical flow methods, our approach is susceptible to the aperture problem. This indicates that only motion components normal to the orientation of an edge in the image space, also known as normal optical flow, can be reliably resolved [10]. Consequently, the proposed method might face challenges in accurately determining the true motion direction in certain ambiguous scenarios. The regularizing effect of the iterative event warping (see Section 3.2) and the multiple spatial scales at which dense optical flow is estimated in our architecture (see Section 3.4) are mechanisms in our proposed solution that collectively strive to counteract the aperture problem's influence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.567, + 0.196, + 0.582 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.473, + 0.881 + ], + "angle": 0, + "content": "In this paper, we presented the first learning-based approach to event-based optical flow estimation that is scalable to high inference frequencies while being able to accurately capture the true trajectory of scene points over time. The proposed pipeline is designed around a continuously-running stateful model that sequentially processes fine discrete partitions of the input event stream while integrating spatiotemporal information. We train this model through a novel, self-supervised, contrast maximization framework (i.e., event deblurring for supervision) that is characterized by an iterative event warping module and a multi-timescale loss function that add robustness and improve the accuracy of the predicted optical flow maps. We demonstrated the effectiveness of our approach on multiple datasets, where our models outperform the self-supervised and model-based baselines by large margins. Future research should look into how to learn to better combine the information from multiple timescales, as well as into the design of lightweight architectures that can keep up with real-time constraints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.23 + ], + "angle": 0, + "content": "We believe that the proposed approach opens up avenues for future research, especially in the field of neuromorphic computing. Spiking networks running on neuromorphic hardware have the potential of exploiting the main benefits of event cameras, but for that they need to process the input events shortly after they arrive. Our proposed framework is a step toward this objective, as it enables the estimation of optical flow in a close to continuous manner, with all the integration of information happening within the model itself." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.244, + 0.597, + 0.259 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.507, + 0.268, + 0.895, + 0.326 + ], + "angle": 0, + "content": "[1] Himanshu Akolkar, Sio-Hoi Ieng, and Ryad Benosman. Real-time high speed motion prediction using fast aperture-robust event-driven visual flow. IEEE Trans. Pattern Anal. Mach. Intell., 44(1):361-372, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.327, + 0.895, + 0.383 + ], + "angle": 0, + "content": "[2] Mohammed Almatrafi, Raymond Baldwin, Kiyoharu Aizawa, and Keigo Hirakawa. Distance surface for event-based optical flow. IEEE Trans. Pattern Anal. and Mach. Intell., 42(7):1547-1556, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.384, + 0.895, + 0.439 + ], + "angle": 0, + "content": "[3] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. Int. Conf. Learn. Representations, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.442, + 0.895, + 0.498 + ], + "angle": 0, + "content": "[4] Patrick Bardow, Andrew J. Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In IEEE Conf. Comput. Vis. Pattern Recog., pages 884-892, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.5, + 0.895, + 0.542 + ], + "angle": 0, + "content": "[5] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE Trans. Neural Netw. Learn. Syst., 25(2):407-417, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.544, + 0.895, + 0.599 + ], + "angle": 0, + "content": "[6] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.601, + 0.895, + 0.657 + ], + "angle": 0, + "content": "[7] Vincent Brebion, Julien Moreau, and Franck Davoine. Realtime optical flow for vehicular perception with low-and high-resolution event cameras. IEEE Trans. on Intell. Transp. Systems, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.659, + 0.895, + 0.699 + ], + "angle": 0, + "content": "[8] Tobias Brosch, Stephan Tschechne, and Heiko Neumann. On event-based optical flow detection. Frontiers in Neuroscience, 9:137, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.703, + 0.895, + 0.771 + ], + "angle": 0, + "content": "[9] Javier Cuadrado, Ulysse Rancon, Benoit R. Cotterreau, Francisco Barranco, and Timothee Masquelier. Optical flow estimation from event-based cameras and spiking neural networks. Frontiers in Neuroscience, 17:1160034, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.774, + 0.895, + 0.829 + ], + "angle": 0, + "content": "[10] David B. de Jong, Federico Paredes-Vallés, and Guido C. H. E. de Croon. How do neural networks estimate optical flow? A neuropsychology-inspired study. IEEE Trans. Pattern Anal. Mach. Intell., 44(11):8290-8305, 2021. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.831, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[11] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-temporal recurrent networks for event-based optical flow estimation. In Proc. AAAI Conf. on Artificial Intell., volume 36, pages 525-533, 2022. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.268, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9703" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. FlowNet: Learning optical flow with convolutional networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2758-2766, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.47, + 0.247 + ], + "angle": 0, + "content": "[13] Guillermo Gallego, Tobi Delbruck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Trans. Pattern Anal. and Mach. Intell., 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[14] Guillermo Gallego, Mathias Gehrig, and Davide Scaramuzza. Focus is all you need: Loss functions for event-based vis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12280-12289, 2019. 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[15] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3867-3876, 2018. 2, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[16] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A stereo event camera dataset for driving scenarios. IEEE Robot. and Autom. Lett., 6(3):4947-4954, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[17] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In Int. Conf. 3D Vis., pages 197–206. IEEE, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[18] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from events and frames. arXiv:2203.13674, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.601 + ], + "angle": 0, + "content": "[19] Jesse J. Hagenaars, Federico Paredes-Valles, and Guido C. H. E. de Croon. Self-supervised learning of event-based optical flow with spiking neural networks. In Advances in Neural Information Process. Syst., volume 34, pages 7167–7179, 2021. 2, 3, 4, 5, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[21] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Int. Conf. Learn. Representations, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[22] Yijin Li, Zhaoyang Huang, Shuo Chen, Xiaoyu Shi, Hongsheng Li, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Blinkflow: A dataset to push the limits of event-based optical flow estimation. arXiv:2303.07716, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[23] Haotian Liu, Guang Chen, Sanqing Qu, Yanping Zhang, Zhijun Li, Alois Knoll, and Changjun Jiang. TMA: Temporal motion aggregation for event-based optical flow. arXiv:2303.11629, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[24] Min Liu and Tobi Delbruck. Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. In Brit. Mach. Vis. Conf., 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[25] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proc. AAAI Conf. on Artificial Intell., volume 32, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[26] Anton Mitrokhin, Cornelia Fermüller, Chethan Parameshwara, and Yiannis Aloimonos. Event-based moving object detection and tracking. In IEEE/RSJ Int. Conf. Intell. Robots and Syst., 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[27] Jun Nagata, Yusuke Sekikawa, and Yoshimitsu Aoki. Optical flow estimation by matching time surface with event-based cameras. Sensors, 21(4):1150, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[28] Garrick Orchard, Ryad Benosman, Ralph Etienne-Cummings, and Nitish V. Thakor. A spiking neural network architecture for visual motion estimation. In IEEE Biomedical Circuits Syst. Conf., pages 298-301. IEEE, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[29] Federico Paredes-Valles and Guido C. H. E. de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2, 3, 4, 5, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[30] Federico Paredes-Vallés, Kirk Y. W. Scheper, and Guido C. H. E. de Croon. Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE Trans. Pattern Anal. and Mach. Intell., 42(8):2051-2064, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[31] Wachirawit Ponghiran, Chamika Mihiranga Liiyanagedera, and Kaushik Roy. Event-based temporally dense optical flow estimation with sequential neural networks. arXiv:2210.01244, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Int. Conf. Medical Image Computing and Comput.-Assisted Intervention, pages 234–241. Springer, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[33] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event collapse in contrast maximization frameworks. Sensors, 22(14):5190, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.606 + ], + "angle": 0, + "content": "[34] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. A fast geometric regularizer to mitigate event collapse in the contrast maximization framework. arXiv:2212.07350, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[35] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In Eur. Conf. Comput. Vis., 2022. 2, 3, 4, 5, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[36] Timo Stoffregen and Lindsay Kleeman. Event cameras, contrast maximization and reward functions: An analysis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12300-12308, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[37] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In European Conf. Comput. Vis., 2020. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[38] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. PWC-Net: CNNs for optical flow using pyramid, warping, and cost volume. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8934–8943, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[39] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In Eur. Conf. Comput. Vis., pages 402-419. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[40] Z Wan, Y Dai, and Y Mao. Learning dense and continuous optical flow from an event camera. IEEE Trans. Image Process., 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9704" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[41] Qimin Wang, Yongjun Zhang, Shaowu Yang, Zhe Liu, Lin Wang, and Luoxi Jing. E-HANet: Event-based hybrid attention network for optical flow estimation. In IEEE Int. Conf. Syst., Man, Cybernetics, pages 2189-2194. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[42] Yilun Wu, Federico Paredes-Vallés, and Guido C. H. E. de Croon. Rethinking event-based optical flow: Iterative deblurring as an alternative to correlation volumes. arXiv:2211.13726, 2022. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.47, + 0.274 + ], + "angle": 0, + "content": "[43] Alex Z. Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3D perception. IEEE Robot. and Autom. Lett., 3(3):2032-2039, 2018. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.317 + ], + "angle": 0, + "content": "[44] Alex Z. Zhu and Liangzhe Yuan. EV-FlowNet: Self-supervised optical flow estimation for event-based cameras. In Robot.: Science and Syst., 2018. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[45] Alex Z. Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 989-997, 2019. 2, 3, 4, 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "9705" + } + ] +] \ No newline at end of file diff --git a/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/ff7524e3-aa7d-4baa-a842-04712fa34f60_origin.pdf b/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/ff7524e3-aa7d-4baa-a842-04712fa34f60_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..15c88dc8fddb12cf06fdd7c23c5e723a348a3b43 --- /dev/null +++ b/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/ff7524e3-aa7d-4baa-a842-04712fa34f60_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe84dc99dccfe2658bd3bd01a5a98ec025c74e255ebb2b344cd64c0c30d98e1b +size 9769429 diff --git a/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/full.md b/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/full.md new file mode 100644 index 0000000000000000000000000000000000000000..adb4076c04ca44d6656a4e0effe13e6960656d7a --- /dev/null +++ b/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/full.md @@ -0,0 +1,269 @@ +# Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow + +Federico Paredes-Valles $^{1,2}$ Kirk Y. W. Scheper $^{2}$ Christope De Wagter $^{1}$ Guido C. H. E. de Croon $^{1}$ + +1 Micro Air Vehicle Laboratory, Delft University of Technology + +2 Stuttgart Laboratory 1, Sony Semiconductor Solutions Europe, Sony Europe B.V. + +![](images/8eb4e758df6147d95bc6ce1673c5196874b8dc4b404d80f7475b506b70aea81d.jpg) +Figure 1: Our pipeline estimates event-based optical flow by sequentially processing small partitions of the event stream with a recurrent model. We propose a novel self-supervised learning framework (left) based on a multi-timescale contrast maximization formulation that is able to exploit the high temporal resolution of event cameras via iterative warping to produce accurate optical flow predictions (right). + +![](images/ebb4507e223e66d8b8413675910840cd4042c26e5346f46a69d49db706498050.jpg) + +# Abstract + +Event cameras have recently gained significant traction since they open up new avenues for low-latency and low-power solutions to complex computer vision problems. To unlock these solutions, it is necessary to develop algorithms that can leverage the unique nature of event data. However, the current state-of-the-art is still highly influenced by the frame-based literature, and usually fails to deliver on these promises. In this work, we take this into consideration and propose a novel self-supervised learning pipeline for the sequential estimation of event-based optical flow that allows for the scaling of the models to high inference frequencies. At its core, we have a continuously-running stateful neural model that is trained using a novel formulation of contrast maximization that makes it robust to nonlinearities and varying statistics in the input events. Results across multiple datasets confirm the effectiveness of our method, which establishes a new state of the art in terms of accuracy for approaches trained or optimized without ground truth. + +# 1. Introduction + +Event cameras capture per-pixel log-brightness changes at microsecond resolution [13]. This operating principle results in a sparse and asynchronous visual signal that, under constant illumination, directly encodes information about the apparent motion (i.e., optical flow) of contrast in the image space. These cameras offer several advantages, such as low latency and robustness to motion blur [13], and hence hold the potential of a high-bandwidth estimation of this optical flow information. However, the event-based nature of the generated visual signal poses a paradigm shift in the processing pipeline, and traditional, frame-based algorithms become suboptimal and often incompatible. Despite this, the majority of learning-based methods that have been proposed so far for event-based optical flow estimation are still highly influenced by frame-based approaches. This influence is normally reflected in two key aspects of their + +pipelines: (i) the design of the network architecture, and (ii) the formulation of the loss function. + +Regarding architecture design, most literature methods format subsets of the input events as dense volumetric representations [45] that are processed at once by stateless (i.e., non-recurrent) models [17, 29, 37, 45]. Similarly to their frame-based counterparts [12, 38, 39], these models estimate the per-pixel displacement over the time-length of the event volume using only the information contained within it. Consequently, these volumes need to encode enough spatiotemporal information for motion to be discernible. However, if done over relatively long time periods, the subsequent models suffer from limitations such as high latency or having to deal with large pixel displacements [17, 18, 42]. + +With respect to the loss function, multiple options have been explored due to the lack of real-world datasets with per-event ground truth. Pure supervised learning can be used with datasets such as MVSEC [43] or DSEC-Flow [17], but their ground truth only contains per-pixel displacement at low frequencies, which makes models difficult to train. On the other hand, a self-supervised learning (SSL) framework can be formulated using either the accompanying frames with the photometric error as a loss [11, 40, 44], or through an events-only contrast maximization [14, 15] proxy loss for motion compensation (i.e., event deblurring) [19, 26, 29, 35, 45]. However, despite not relying on ground truth, all the literature on SSL for optical flow assumes that the events move linearly within the time window of the loss, which ignores much of the potential of event cameras and their high temporal resolution (see Fig. 1, bottom left). + +In this work, we focus on the estimation of high frequency event-based optical flow and how this can be learned in an SSL fashion using contrast maximization with a relaxed linear motion assumption. To achieve this, we build upon the continuous-operation pipeline from Hagenaars et al. [19], which retrieves optical flow by sequentially processing small partitions of the event stream with a stateful (i.e., recurrent) model over time, instead of dealing with large volumes of input events. We augment (and train) this pipeline with a novel contrast maximization formulation that performs event motion compensation in an iterative manner at multiple temporal scales, as shown in Fig. 1. Using this framework, we achieve the best accuracy of all contrast-maximization-based approaches on multiple datasets, only being outperformed by pure supervised learning methods trained with ground-truth data. + +In summary, the extensions that we propose to the SSL framework in [19], i.e., our main contributions, are: + +- The first iterative event warping module in the context of contrast maximization (see Section 3.2). This module unlocks a novel multi-reference loss function that better captures the trajectory of scene points over time, thus improving the accuracy of the predictions. + +- The first multi-timescale approach to contrast maximization, which adds robustness, improves convergence, and reduces tuning requirements of loss-related hyperparameters (see Section 3.3). + +As a result, we present the first self-supervised optical flow method for event cameras that relaxes the linear motion assumption, and hence that has the potential of exploiting the high temporal resolution of the sensor by producing estimates in a close-to continuous manner. We validate the proposed framework through extensive evaluations on multiple datasets. Additionally, we conduct ablation studies to show the effectiveness of each individual component. + +# 2. Related work + +Due to the aforementioned advantages of event cameras for optical flow estimation, extensive research has been carried out since these sensors were first introduced [1, 2, 4-8, 24, 27, 28, 35]. Regarding learning-based approaches, the first method was proposed by Zhu et al. in [44] with EV-FlowNet: a UNet-like [32] architecture trained with SSL, with the supervisory signal coming from the photometric error between subsequent frames captured with an accompanying camera. To avoid the need for a secondary vision sensor, in [45], Zhu et al. proposed an SSL framework around the contrast maximization for motion compensation idea from [14, 15], hence relying solely on event data. This pipeline was used and further improved in [19, 29, 35]. Other approaches trained EV-FlowNet in a supervised fashion with synthetic/real ground-truth data and showed higher accuracy levels through evaluations on public benchmarks [17, 37]. However, they also highlighted EV-FlowNet's inability to deal with large pixel displacements [17]. Because of this, inspired by the frame-based literature [39], Gehrig et al. proposed E-RAFT in [17]: the first architecture to introduce the use of correlation volumes in the event camera literature. This model, which was recently augmented with attention mechanisms [41], achieved state-of-the-art performance in multiple datasets. + +A novel perspective to learning event-based optical flow is to move away from event accumulation over long timespans, and instead rely on continuously-running stateful models that integrate information over time. The first methods of this kind were proposed by Paredes-Valles et al. [30] and Hagenaars et al. [19], but this idea has recently gained interest in the event camera literature [11, 31, 42]. The reason is that leveraging memory through sequential processing can potentially lead to lightweight and low-latency solutions that are also robust to large pixel displacements without the need for correlation volumes [42]. However, despite their potential of being scaled to high inference frequencies, all these solutions assume that events move linearly in the timespan of their loss function, and hence cannot capture the true trajectory of scene points over time. + +When it comes to learning nonlinear pixel trajectories from event data, only the work of Gehrig et al. in [18] is to be highlighted. However, their approach uses multiple event and correlation volumes to fit Bézier curves to the trajectory of scene points, resulting in a high-latency solution. In contrast, in this work, we extend the continuous-operation pipeline from Hagenaars et al. [19] with an SSL framework in which discrete trajectories are regressed at high frequency by leveraging memory within the models. This approach allows, for the first time, to exploit the high temporal resolution of event cameras while capturing more accurately the trajectory of scene points over time thanks to the proposed iterative event warping mechanism. + +Among non-learning-based approaches, the work of Shiba et al. in [35] bears particular relevance due to certain similarities with our framework. Specifically, Shiba et al. propose a tile-based method for event-based optical flow estimation that extends contrast maximization [14, 15] by incorporating (i) multiple spatial scales, (ii) a multi-reference focus loss, and (iii) a "time-aware" optical flow formulation. Similarly, our learning-based approach also employs a multi-scale strategy and utilizes a multi-reference focus loss. However, instead of making assumptions about the events' motion over extended time periods, we propose to learn their potentially nonlinear trajectories at high frequency by leveraging the iterative event warping module. + +# 3. Method + +The goal of this work is to learn to sequentially estimate optical flow at high frequencies from a continuous stream of events. In such a pipeline, if the inference frequency is sufficiently high, events need to be processed nearly as soon as they are triggered by the sensor, with no pre-processing in between. Because of this, the integration of temporal information needs to happen in the network itself. To accomplish this, we propose the framework in Fig. 1, in which a stateful model is trained using our novel formulation of contrast maximization for sequential processing. The components of this framework are described in the following sections. + +# 3.1. Input format and contrast maximization + +For an ideal camera, an event $\boldsymbol{e}_i = (\boldsymbol{x}_i, t_i, p_i)$ of polarity $p_i \in \{+, -\}$ is triggered at pixel $\boldsymbol{x}_i = (x_i, y_i)^T$ and time $t_i$ whenever the change in log-brightness since the last event at that pixel location reaches the contrast sensitivity threshold for that polarity [13]. + +As in [19], we use a two-channel event count image as input representation, which gets populated with consecutive, non-overlapping, fine discrete partitions of the event stream $\varepsilon_{k}^{\mathrm{inp}}\doteq \{e_{i}\}$ (further referred to as input partitions), each containing all the events in a time window of a certain duration, i.e., $t_i\in [t_k^{\mathrm{begin}},t_k^{\mathrm{end}}]$ . This representation does not + +contain temporal information by itself, and recurrent models are hence required for estimating of optical flow. + +Regarding learning, we use the contrast maximization framework [14] to train, in an SSL fashion, to continuously estimate dense (i.e., per-pixel) optical flow from the event stream. Assuming brightness constancy, accurate flow information is encoded in the spatiotemporal misalignments (i.e., blur) among the events triggered by the same portion of a moving edge. To retrieve it, one has to compensate for this motion by geometrically transforming the events using a motion model. As in [19, 29, 35, 45], we transport each event to a reference time $t_{\mathrm{ref}}$ through: + +$$ +\boldsymbol {x} _ {i} ^ {\prime} = \boldsymbol {x} _ {i} + \left(t _ {\text {r e f}} - t _ {i}\right) \boldsymbol {u} \left(\boldsymbol {x} _ {i}\right) \tag {1} +$$ + +where $\pmb{u}(\pmb{x}) = (u(\pmb{x}), v(\pmb{x}))^T$ denotes the optical flow map used to transport each event from $t_i$ to $t_{\mathrm{ref}}$ . The result of aggregating the transformed events is further referred to as the image of warped events (IWE) at $t_{\mathrm{ref}}$ . + +As the loss function of our SSL framework, we adopt the time-based focus objective function from [19]. Using the warped events at a given $t_{\mathrm{ref}}$ , we generate an image of the per-pixel average timestamps for each polarity $p'$ via bilinear interpolation: + +$$ +\begin{array}{l} T _ {p ^ {\prime}} (\boldsymbol {x}; \boldsymbol {u} | t _ {\text {r e f}}) = \frac {\sum_ {j} \kappa \left(x - x _ {j} ^ {\prime}\right) \kappa \left(y - y _ {j} ^ {\prime}\right) \bar {t} _ {j} \left(t _ {\text {r e f}} , t _ {j}\right)}{\sum_ {j} \kappa \left(x - x _ {j} ^ {\prime}\right) \kappa \left(y - y _ {j} ^ {\prime}\right) + \epsilon} \tag {2} \\ \kappa (a) = \max (0, 1 - | a |) \\ \end{array} +$$ + +$$ +j = \{i \mid p _ {i} = p ^ {\prime} \}, \quad p ^ {\prime} \in \{+, - \}, \quad \epsilon \approx 0 +$$ + +where $\bar{t}_i$ denotes the normalized timestamp contribution of the $i$ th event, according to Fig. 3 and Eq. 5. + +Then, the contrast maximization loss function at $t_{\mathrm{ref}}$ is defined as the scaled sum of the squared temporal images: + +$$ +\mathcal {L} _ {\mathrm {C M}} \left(t _ {\text {r e f}}\right) = \frac {\sum_ {\boldsymbol {x}} T _ {+} \left(\boldsymbol {x} ; \boldsymbol {u} \mid t _ {\text {r e f}}\right) ^ {2} + T _ {-} \left(\boldsymbol {x} ; \boldsymbol {u} \mid t _ {\text {r e f}}\right) ^ {2}}{\sum_ {\boldsymbol {x}} \left[ n \left(\boldsymbol {x} ^ {\prime}\right) > 0 \right] + \epsilon} \tag {3} +$$ + +where $n(\pmb{x}^{\prime})$ denotes a per-pixel event count of the IWE. The lower the $\mathcal{L}_{\mathrm{CM}}$ , the better the event deblurring at $t_{\mathrm{ref}}$ . + +As discussed in [14, 19, 36], for any focus objective function to be a robust supervisory signal for contrast maximization, the event partition used in the optimization needs to contain enough motion information (i.e., blur) so it can be compensated for. However, as in [19], that is not the case in our input partitions due to the fine discretization of the event stream that we are targeting. Therefore, only at training time, we define the so-called training partition (or event buffer in Fig. 1) $\varepsilon_{k\rightarrow k + R}^{\mathrm{train}}\doteq \{\varepsilon_i^{\mathrm{inp}}\}_{i = k}^{k + R}$ , which stacks together the events in the input partitions of $R$ successive forward passes through the network. Once these are performed, we use this partition and the estimated optical flow maps to compute the loss, and use truncated backpropagation through time to update the model parameters. After + +![](images/8eda2d9e871b781ae7697add87da18789d713eb2becd205c4eb6f65722bae766.jpg) +Figure 2: Incompatibility of multi-reference deblurring and linear event warping in the presence of nonlinearities in the pixel trajectories. Top: Events generated by a moving dot following circular motion (left), and optical flow color-coding scheme (right). Middle: Optical flow solutions required to produce sharp IWEs at different $t_{\mathrm{ref}}$ using linear and iterative warping. While the former requires a different solution for each $t_{\mathrm{ref}}$ , the proposed iterative warping can achieve this using a single solution. Arrows illustrate the direction of the required displacement $\Delta \pmb{x}_i = (t_{\mathrm{ref}} - t_i)\pmb{u}(\pmb{x}_i)$ at each (discretized) spatial location. Bottom: Resulting IWEs at different $t_{\mathrm{ref}}$ using the optimal optical flow map for $t_{\mathrm{ref}} = \mathrm{mid}$ . + +this update, we detach the states of the network from the computational graph and clear the training partition and optical flow buffer. The importance of sequential processing, as an alternative to training stateless models in short input partitions, is corroborated in the supplementary material. + +# 3.2. Iterative event warping + +In order to better approximate the trajectories of scene points over time, in this work we relax the linear motion as- + +![](images/ed19fc19c64c3b8d34fb8c174a5b4f787c59c9755326f2109289b1b3a79a8786.jpg) +Figure 3: Timestamp normalization profiles for the per-event contributions to the images of average timestamps in Eq. 2, for $R = 10$ . Left: Deblurring only done at the extremes of the training partition, as in [19, 29, 45]. Right: Deblurring done at all reference times, thanks to the proposed iterative event warping. Normalization profiles only shown for three $t_{\mathrm{ref}}$ for a better visualization. + +sumption in the SSL framework for event-based optical flow by augmenting the sequential-estimation pipeline from [19] with iterative event warping. Instead of transporting events to a given $t_{\mathrm{ref}}$ assuming linear motion regardless of the length of the warping interval (as in [19, 29, 35, 45]), we perform a finer discretization of the event trajectories and assume that motion is only linear between optical flow estimates. Therefore, to express a group of events at a given $t_{\mathrm{ref}}$ , we geometrically transform them using all the intermediate optical flow estimates through multiple iterations of Eq. 1. Fig. 2 shows the differences between the linear event warping in [19] and the proposed iterative augmentation, as well as the limitations of the former. Note that our iterative warping is fundamentally different from that of the work of Wu et al. in [42], where input events are deblurred before being passed to the models using residual optical flow estimates until convergence. + +Literature methods on event-based optical flow with contrast maximization compute the focus objective function at multiple reference times in the training partition (usually at the extremes) to prevent overfitting and/or scaling issues during backpropagation [19, 29, 35, 45]. However, since these approaches assume optical flow constancy in the span of their loss function, they suffer with nonlinearities in the pixel trajectories (see the blurry IWEs in Fig. 2, bottom). On the contrary, because of the finer discretization of the event trajectories coming with our sequential processing pipeline and the proposed iterative warping, we can use any (combination of) reference time(s) for the computation of the focus objective function. In fact, as shown in Fig. 3 (right), we propose the use of all the discretization points as reference times for event deblurring. Apart from the aforementioned regularizing benefits, having to produce sharp IWEs at any $t_{\mathrm{ref}}$ forces the models to estimate a sequence of optical flow maps that is consistent with the velocity profile of the event stream. For a given training partition of + +![](images/b48ea9605462c4fee827fe64233a73e117cf6ac63dfb0de8c3b87ffefb877f92.jpg) +Figure 4: Multi-timescale approach to contrast maximization. For a given training partition of length $R$ , we fit multiple sub-partitions of different lengths (in the figure: one of length $R$ in red, two $R/2$ in blue, and four $R/4$ in green) and compute the loss in each of them according to Eq. 4. The global loss is computed as in Eq. 6. This figure only shows the timestamp normalization profiles of the central $t_{\mathrm{ref}}$ of each sub-partition, but the losses are still computed at all reference times. An image representation of the accumulated input events in a sub-partition of each timescale is also shown. + +length $R$ , the loss is computed as follows: + +$$ +\mathcal {L} _ {\mathrm {C M}} ^ {R} = \frac {1}{R + 1} \sum_ {t _ {\text {r e f}} = 0} ^ {R} \mathcal {L} _ {\mathrm {C M}} \left(t _ {\text {r e f}}\right) \tag {4} +$$ + +which ensures that the IWEs (and the corresponding images of average timestamps in Eq. 2) at all reference times $t_{\mathrm{ref}} \in [0, R]$ contribute equally to the loss. + +As in [19, 29, 45], for $\mathcal{L}_{\mathrm{CM}}(t_{\mathrm{ref}})$ in Eq. 3 to be a valid supervisory signal, events that are temporally close to the reference time $t_{\mathrm{ref}}$ need to contribute to the temporal image in Eq. 2 with larger timestamp values than events that are far in time. Therefore, once the events have been transported to $t_{\mathrm{ref}}$ , their timestamp is normalized prior to the computation of Eq. 2 as follows: + +$$ +\bar {t} _ {i} \left(t _ {\text {r e f}}, t _ {i}\right) = 1 - \frac {\left| t _ {\text {r e f}} - t _ {i} \right|}{R}, \quad \text {w i t h} t _ {i} \in [ 0, R ] \tag {5} +$$ + +which results in the normalization profiles in Fig. 3, for a given training partition of length $R$ . + +Inspired by [25], we mask individual events from the computation of the loss whenever we detect that they are transported outside the image space through the event warping process in the span of a deblurring window defined at the reference time $t_{\mathrm{ref}}$ . This prevents our models from learning incorrect deformations at the image borders, and is motivated by the fact that the complete trajectory of the pixel is only partially observable in the training partition. An ab + +lation study on the impact of this masking strategy can be found in the supplementary material. + +Lastly, we do not augment the loss function in Eq. 4 with smoothing priors acting as regularization mechanisms. With iterative event warping, the error propagates through all the pixels covered in the warping process, regardless of whether they have input events or not. Therefore, the spatial coherence of the resulting optical flow maps is enhanced. + +# 3.3. Deblurring at multiple timescales + +Despite the addition of iterative event warping, the success of our SSL framework still heavily depends on the hyperparameters that control the amount of motion information perceived by the networks in the span of a deblurring window. In our pipeline, these are: $\mathrm{dt}_{\mathrm{input}}$ , the timestep used to discretize the event stream; and $R$ , the number of forward passes, and hence optical flow maps, per loss. Thus, the effective length (in units of time) of the event window used for motion compensation is given by $\mathrm{dt}_{\mathrm{input}} \times R$ . We hypothesize that, for each training dataset, there is an optimal length for this window that depends on the statistics of the data (e.g., event density, distribution of optical flow magnitudes) and model architecture, and that deviations from this optimal length lead to the learning of suboptimal solutions. E.g., shorter windows may converge to solutions that are more selective to fast rather than slow moving objects, and vice versa. Note that not only our method is sensitive to the tuning of these parameters, but also previous approaches based on contrast maximization [19, 29, 35, 45]. + +To add a layer of robustness to the framework and relax its strong dependency on hyperparameter optimization, we propose the multi-timescale approach illustrated in Figs. 1 and 4. For a given training partition of length $R$ , instead of computing a single focus loss through Eq. 4, we compute this loss at $S$ temporal scales of length $R / 2^s$ , with $0 \leq s \leq S - 1$ , and combine them as follows: + +$$ +\mathcal {L} _ {\mathrm {C M}} ^ {\text {m u l t i}} = \frac {1}{S} \sum_ {s = 0} ^ {S - 1} \frac {1}{2 ^ {s}} \sum_ {p = 0} ^ {2 ^ {s} - 1} \mathcal {L} _ {\mathrm {C M}, p} ^ {R / 2 ^ {s}} \tag {6} +$$ + +As shown, we fit multiple non-overlapping sub-partitions in the training buffer if $s > 0$ . The subscript $p$ indicates their location in this buffer, starting from the earliest (see Fig. 4). + +Note that, through this multi-timescale approach to contrast maximization, our models need to converge to a solution that is suitable for all the timescales in the optimization, regardless of their length. An alternative formulation would be to incorporate per-pixel learnable masks (i.e., an attention module in the loss space) so that, depending on the input statistics, learning only happens at the most adequate scale. However, for this to happen, the loss function would have to be augmented to stimulate this behavior, and it is unclear how that would be done in practice. + +![](images/1452b0e4edbc9fba7d891bbd856f0521d25a10cf94e5c12fc762238f3f7a6a94.jpg) +Figure 5: Schematic of the model architecture used in this work. It is characterized by $N_{\mathcal{G}}$ recurrent encoders, $N_{\mathcal{R}}$ residual blocks, and $N_{\mathcal{G}}$ decoder layers. Optical flow estimates are produced at all decoder levels. In this diagram, $N_{\mathcal{G}} = 2$ and $N_{\mathcal{R}} = 1$ . + +# 3.4. Network architecture + +We use the recurrent version of EV-FlowNet [44] proposed in $[19]^1$ (see Fig. 5). The events are represented as event count images (see Section 3.1), then passed through four encoders with strided convolutions followed by ConvGRUs [3] (channels doubling, starting from 64), two residual blocks [20], and then four decoders performing bilinear upsampling followed by convolution. After each decoder, there is a skip connection (using element-wise summation) from the corresponding encoder, as well as a depthwise convolution to produce estimates at lower scales, which are then concatenated with the activations of the previous decoder. Note that the proposed focus loss function (see Eq. 6) is applied to each intermediate optical flow estimate via upsampling. Lastly, all layers use $3\times 3$ kernels and ReLU activations except for the prediction layers, which use TanH. + +# 4. Experiments + +We evaluate our method on the DSEC-Flow [16, 17] and MVSEC [43] datasets. We evaluate the accuracy of the predictions based on the following metrics: (i) EPE (lower is better, ↓), the endpoint error; (ii) $\%_{3\mathrm{PE}}(\downarrow)$ , the percentage of points with EPE greater than 3 pixels; (iii) FWL ( $\uparrow$ ) [37], a deblurring quality metric based on the variance of the IWEs; and (iv) RSAT ( $\downarrow$ ) [19], a deblurring quality metric based on the per-pixel average timestamps of the IWEs. We compare our solution to the published baselines, which range from supervised learning (SL) methods trained with ground truth, to SSL methods trained with grayscale images ( $\mathrm{SSL}_{\mathrm{F}}$ ) or events ( $\mathrm{SSL}_{\mathrm{E}}$ ), and model-based approaches (MB). + +We train all our models on a subset of sequences from the training dataset of DSEC-Flow (only daylight recordings, see supplementary material). This corresponds to 19 minutes of training data, which we split into $572128 \times 128$ (randomly-cropped) sequences of 2 seconds each. We use a batch size of 8 and train until convergence with the Adam optimizer [21] and a learning rate of 1e-5. To keep mem + +![](images/6151cc11d5d61687c1d5afe4a6a2e089ccd485066c3a3f26d6fb01ba747c1a0e.jpg) +Figure 6: Reconstruction of the pixel displacement of a scene point in the time window of a ground-truth sample from the multiple optical flow maps estimated in this period, i.e., $\forall \pmb{u}_k \in [t_{\mathrm{begin}}, t_{\mathrm{end}}]$ . The error of the last optical flow estimate is magnified for clarity. + +
EPE↓%3PE↓FWL↑RSAT↓
SLE-RAFT [17]0.792.681.330.87
EV-FlowNet, Gehrig et al. [17]2.3218.60--
Gehrig et al. [18]0.752.44--
IDNet [42]0.722.04--
TIDNet [42]0.843.41--
TMA [23]0.742.30--
Cuadrado et al. [9]1.7110.31--
E-Flowformer [22]0.762.68--
SSLEEV-FlowNet* [45]3.8631.451.300.85
ConvGRU-EV-FlowNet* [19]4.2733.271.550.90
dt = 0.01s, R = 2, S = 1 (Ours)9.6686.441.911.07
dt = 0.01s, R = 5, S = 1 (Ours)4.0552.221.580.97
dt = 0.01s, R = 10, S = 1 (Ours)2.3317.771.260.88
dt = 0.01s, R = 20, S = 1 (Ours)16.6333.671.061.10
dt = 0.01s, R = 10, S = 3 (Ours)2.8227.091.370.92
dt = 0.01s, R = 20, S = 4 (Ours)2.7323.731.240.90
MBShiba et al. [35]3.4730.861.370.89
+ +*Retrained by us on DSEC-Flow, linear warping. + +Table 1: Quantitative evaluation on the DSEC-Flow dataset [17]. Best in bold, runner up underlined. A breakdown of the results is provided in the supplementary material. + +ory usage within limits, we only propagate error gradients through up to 1e3 randomly-chosen events per millisecond of data. Despite this, note that we warp and use all the input events for the computation of the loss function. + +# 4.1. Evaluation procedure + +When evaluating our sequential models, if $\mathrm{dt_{gt}} > \mathrm{dt_{input}}$ , we need to reconstruct the estimated per-pixel displacement in the ground-truth time window from the multiple optical flow maps estimated in this period. We do this by first averaging the (bilinearly interpolated) optical flow vectors that describe the trajectory of each scene point, and then by scaling the resulting optical flow vectors by $\mathrm{dt_{gt}} / \mathrm{dt_{input}}$ . This converts them from units of pixels/input to units of pixel displacement. An illustration of this reconstruction is shown in Fig. 6 for a scene point following a nonlinear trajectory. Note that our solution is subject to cumulative errors when evaluated through this reconstruction on benchmarks with + +![](images/b82cc80d8f8247aa65493bb77918a1efe77d1ff9f0da5a3212529b8f02586a7e.jpg) +Figure 7: Qualitative comparison of our method with the state-of-the-art E-RAFT architecture [17] and the model-based approach from Shiba et al. [35] on sequences from the test partition of the DSEC-Flow dataset [17]. Ground truth not included due to unavailability. The optical flow color coding can be found in Fig. 2 (top), and the corresponding IWEs in the supplementary material. + +ground truth provided at low rates (e.g., $10\mathrm{Hz}$ in DSEC-Flow [17]). Therefore, it will compare unfavorably to other, non-sequential, methods that only produce a single optical flow estimate in the timespan of a ground-truth sample. + +# 4.2. Optical flow evaluation + +Evaluation on DSEC-Flow: Quantitative results of our evaluation on DSEC-Flow are presented in Table 1, and are supported by the qualitative comparison in Fig. 7. For this experiment, we trained multiple models with the same $\mathrm{dt}_{\mathrm{input}} = 0.01\mathrm{s}$ (i.e., $\times 10$ faster than DSEC's ground truth) but different lengths of the training partition, and with and without the multi-timescale approach. Multiple conclusions can be derived from the reported results. Firstly, our best performing model (i.e., $R = 10$ , $S = 1$ ) achieves the best accuracy of all contrast-maximization-based approaches on this dataset according to the EPE and the percentage of outliers. Specifically, it outperforms the baselines with an im + +provement in the EPE in the $33\% - 45\%$ range, only being outperformed by SL methods trained with ground truth on the same dataset [9, 17, 22, 23, 42]. This confirms that (i) the timestamp-based loss function in Section 3.1 allows us to learn accurate event-based optical flow (contrary to the findings of [34]); and that (ii) our augmentations to the sequential pipeline in [19] lead to a significant improvement in the accuracy of the model (i.e., $45\%$ improvement in the EPE). + +Secondly, these results also confirm our hypothesis that, for each training dataset, there is an optimal length for the training partition $R$ in terms of the EPE. According to Table 1, the optimal $R$ for this dataset, our model architecture, and our $\mathrm{dt}_{\mathrm{input}}$ is 10 (i.e., 0.1s of event data), with the EPE increasing if the training partition is made shorter or longer. However, as also shown in this table, we can relax the strong dependency of the contrast maximization framework on this parameter through the proposed multi-timescale approach. + +
EPE↓%3PE↓FWL↑RSAT↓
dt = 0.1s*3.4834.720.980.87
dt = 0.05s3.0927.361.110.91
dt = 0.01s2.3317.771.260.88
dt = 0.005s2.3417.921.380.89
dt = 0.002s2.6621.832.040.90
+ +*Non-recurrent, volum. event repr. with 10 bins. + +Our $S > 1$ models converged to solutions that slightly underperform our best performing single-scale model (EPE went up by $17\% - 21\%$ ), but were trained without the need to fine-tune the length of the training partition. Note that, despite this slight drop in accuracy, these multi-timescale solutions still outperform the other non-SL baselines. To further support these results, a visualization of the distribution of the EPE of our models as a function of the ground truth magnitude is provided in the supplementary material. + +Lastly, Table 1 also allow us to conclude that deblurring quality metrics FWL [37] and RSAT [19] are not reliable indicators of the quality of the estimated optical flow. The reason for this is their inability to capture "event collapse" issues (as described in [33]), and would give favorable scores to undesirable solutions that warp all the events into a few pixels. According to our results, the FWL metric, being the spatial variance of the IWE relative to that of the identity warp, suffers more from this issue: the best FWL value is obtained with a model with 9.66 EPE. + +Regarding qualitative results, Fig. 7 shows a comparison of our best performing model with the state-of-the-art E-RAFT architecture [17] and the contrast-maximization-based approach from Shiba et al. [35] on multiple sequences from the test partition of DSEC-Flow (i.e., ground truth is unavailable). These results confirm that our models are able to estimate high quality event-based optical flow despite not having access to ground-truth data during training, and also show the superiority of our method over the current best contrast-maximization-based approach [35]. Two limit cases in which our models provide suboptimal solutions are also shown in this figure: (i) sequences recorded at night (e.g., zurich_city_12_a) due to the presence of large amounts of events triggered by flashing lights and not by motion; and (ii) the car hood, which is also problematic for E-RAFT (i.e., does not capture it) and for [35]. Note that, in our case, (ii) is an artifact of the pixel displacements reconstructed from multiple optical flow estimates, and could be mitigated by having an occlusion handling mechanism in this reconstruction process. + +In addition to the evaluation in Table 1 and Fig. 7, we also conducted an experiment in which we trained multiple models with different $\mathrm{dt}_{\mathrm{input}}$ (ranging from 0.1s to 0.002s) but with the same amount of information in the training par + +Table 2: Impact of the input window length on the DSEC-Flow dataset [17]. Best in bold, runner up underlined. + +
EPE↓%3PE↓
SLEV-FlowNet+ [37]0.680.99
E-RAFT [17]0.241.70
EV-FlowNet, Gehrig et al. [17]0.310.00
TMA [23]0.250.07
Cuadrado et al. [9]0.85-
SSLFEV-FlowNet, Zhu et al. [44]0.490.20
Ziluo et al. [11]0.420.00
SSLEEV-FlowNet, Zhu et al. [45]0.320.00
EV-FlowNet, Paredes-Vallés et al. [29]0.925.40
EV-FlowNet, Shiba et al. [35]0.360.09
ConvGRU-EV-FlowNet [19]0.470.25
Ours0.270.05
MBAkolkar et al. [1]2.75-
Brebion et al. [7]0.530.20
Shiba et al. [35]0.300.11
+ +Table 3: Quantitative evaluation on MVSEC's outdoor_1 sequence [43]. Best in bold, runner up underlined. Results on the indoor sequences can be found in the supplementary material. + +tition: 0.1s of event data. Results in Table 2 show that our sequential pipeline leads to an improvement in the accuracy of the predicted optical flow maps with respect to the stateless EV-FlowNet, which processes the 0.1s of event data at once. This improvement is due to the fact that the complexity of dealing with large pixel displacements gets reduced when processing the input data sequentially using shorter input windows. In addition to this, Table 2 also shows that the accuracy of our models is not compromised when estimating optical flow at higher frequencies, despite the high sparsity levels in the input data at those rates. + +Evaluation on MVSEC: Quantitative results of our evaluation on the outdoor_day1 sequence from MVSEC are presented in Table 3, and are supported by the qualitative comparison in the supplementary material. For this experiment, since (i) there is no consensus in the literature with respect to the training dataset [17, 19, 29, 35, 37, 44, 45], and (ii) the outdoor_day2 sequence (i.e., the other daylight, automotive sequence) is only 9 minutes of duration during which the event camera is subject to high frequency vibrations [43], we decided to transfer one of our models trained on DSEC-Flow to MVSEC. More specifically, we chose the model trained with $\mathrm{dt}_{\mathrm{input}} = 0.005\mathrm{s}$ and $R = 20$ from Table 2, as a model trained with a short input window on DSEC-Flow is expected to be robust to the slow motion statistics of MVSEC [17]. We deployed the model at the same frequency as the temporally-upsampled ground truth (i.e., $45\mathrm{Hz}$ ). Results in Table 3 show that, by doing this, our model outperforms the great majority of methods in terms of the EPE (even some SL methods trained on this dataset), and is only surpassed by the current state-of-the-art E-RAFT architecture [17]. Besides reaffirming the high quality of the produced optical flow estimates, these results also confirm the generalizability of our method. For com + +pleteness, the results on the indoor evaluation sequences from MVSEC are provided in the supplementary material. + +# 5. Limitations + +The self-supervised method for event-based optical flow presented in this work, while demonstrating highly accurate and promising results, is not without limitations. Two critical challenges that need to be acknowledged are the brightness constancy assumption and the aperture problem. Firstly, the contrast maximization framework [14, 15] assumes constant illumination, leading our models to face difficulty in learning from events that are not due to motion in the image space but that arise from changes in illumination. Since this limitation is inherent to contrast maximization, it extends to other approaches based on the same principle [19, 29, 35]. Due to this assumption, we excluded sequences recorded at night from our training dataset (see supplementary material). Secondly, akin to many other optical flow methods, our approach is susceptible to the aperture problem. This indicates that only motion components normal to the orientation of an edge in the image space, also known as normal optical flow, can be reliably resolved [10]. Consequently, the proposed method might face challenges in accurately determining the true motion direction in certain ambiguous scenarios. The regularizing effect of the iterative event warping (see Section 3.2) and the multiple spatial scales at which dense optical flow is estimated in our architecture (see Section 3.4) are mechanisms in our proposed solution that collectively strive to counteract the aperture problem's influence. + +# 6. Conclusion + +In this paper, we presented the first learning-based approach to event-based optical flow estimation that is scalable to high inference frequencies while being able to accurately capture the true trajectory of scene points over time. The proposed pipeline is designed around a continuously-running stateful model that sequentially processes fine discrete partitions of the input event stream while integrating spatiotemporal information. We train this model through a novel, self-supervised, contrast maximization framework (i.e., event deblurring for supervision) that is characterized by an iterative event warping module and a multi-timescale loss function that add robustness and improve the accuracy of the predicted optical flow maps. We demonstrated the effectiveness of our approach on multiple datasets, where our models outperform the self-supervised and model-based baselines by large margins. Future research should look into how to learn to better combine the information from multiple timescales, as well as into the design of lightweight architectures that can keep up with real-time constraints. + +We believe that the proposed approach opens up avenues for future research, especially in the field of neuromorphic computing. Spiking networks running on neuromorphic hardware have the potential of exploiting the main benefits of event cameras, but for that they need to process the input events shortly after they arrive. Our proposed framework is a step toward this objective, as it enables the estimation of optical flow in a close to continuous manner, with all the integration of information happening within the model itself. + +# References + +[1] Himanshu Akolkar, Sio-Hoi Ieng, and Ryad Benosman. Real-time high speed motion prediction using fast aperture-robust event-driven visual flow. IEEE Trans. Pattern Anal. Mach. Intell., 44(1):361-372, 2020. 2, 8 +[2] Mohammed Almatrafi, Raymond Baldwin, Kiyoharu Aizawa, and Keigo Hirakawa. Distance surface for event-based optical flow. IEEE Trans. Pattern Anal. and Mach. Intell., 42(7):1547-1556, 2020. 2 +[3] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. Int. Conf. Learn. Representations, 2015. 6 +[4] Patrick Bardow, Andrew J. Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In IEEE Conf. Comput. Vis. Pattern Recog., pages 884-892, 2016. 2 +[5] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE Trans. Neural Netw. Learn. Syst., 25(2):407-417, 2013. 2 +[6] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, 2012. 2 +[7] Vincent Brebion, Julien Moreau, and Franck Davoine. Realtime optical flow for vehicular perception with low-and high-resolution event cameras. IEEE Trans. on Intell. Transp. Systems, 2021. 2, 8 +[8] Tobias Brosch, Stephan Tschechne, and Heiko Neumann. On event-based optical flow detection. Frontiers in Neuroscience, 9:137, 2015. 2 +[9] Javier Cuadrado, Ulysse Rancon, Benoit R. Cotterreau, Francisco Barranco, and Timothee Masquelier. Optical flow estimation from event-based cameras and spiking neural networks. Frontiers in Neuroscience, 17:1160034, 2023. 6, 7, 8 +[10] David B. de Jong, Federico Paredes-Vallés, and Guido C. H. E. de Croon. How do neural networks estimate optical flow? A neuropsychology-inspired study. IEEE Trans. Pattern Anal. Mach. Intell., 44(11):8290-8305, 2021. 9 +[11] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-temporal recurrent networks for event-based optical flow estimation. In Proc. AAAI Conf. on Artificial Intell., volume 36, pages 525-533, 2022. 2, 8 + +[12] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. FlowNet: Learning optical flow with convolutional networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2758-2766, 2015. 2 +[13] Guillermo Gallego, Tobi Delbruck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Trans. Pattern Anal. and Mach. Intell., 2020. 1, 3 +[14] Guillermo Gallego, Mathias Gehrig, and Davide Scaramuzza. Focus is all you need: Loss functions for event-based vis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12280-12289, 2019. 2, 3, 9 +[15] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3867-3876, 2018. 2, 3, 9 +[16] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A stereo event camera dataset for driving scenarios. IEEE Robot. and Autom. Lett., 6(3):4947-4954, 2021. 6 +[17] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In Int. Conf. 3D Vis., pages 197–206. IEEE, 2021. 2, 6, 7, 8 +[18] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from events and frames. arXiv:2203.13674, 2022. 2, 3, 6 +[19] Jesse J. Hagenaars, Federico Paredes-Valles, and Guido C. H. E. de Croon. Self-supervised learning of event-based optical flow with spiking neural networks. In Advances in Neural Information Process. Syst., volume 34, pages 7167–7179, 2021. 2, 3, 4, 5, 6, 7, 8, 9 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 6 +[21] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Int. Conf. Learn. Representations, 2014. 6 +[22] Yijin Li, Zhaoyang Huang, Shuo Chen, Xiaoyu Shi, Hongsheng Li, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Blinkflow: A dataset to push the limits of event-based optical flow estimation. arXiv:2303.07716, 2023. 6, 7 +[23] Haotian Liu, Guang Chen, Sanqing Qu, Yanping Zhang, Zhijun Li, Alois Knoll, and Changjun Jiang. TMA: Temporal motion aggregation for event-based optical flow. arXiv:2303.11629, 2023. 6, 7, 8 +[24] Min Liu and Tobi Delbruck. Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. In Brit. Mach. Vis. Conf., 2018. 2 +[25] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proc. AAAI Conf. on Artificial Intell., volume 32, 2018. 5 + +[26] Anton Mitrokhin, Cornelia Fermüller, Chethan Parameshwara, and Yiannis Aloimonos. Event-based moving object detection and tracking. In IEEE/RSJ Int. Conf. Intell. Robots and Syst., 2018. 2 +[27] Jun Nagata, Yusuke Sekikawa, and Yoshimitsu Aoki. Optical flow estimation by matching time surface with event-based cameras. Sensors, 21(4):1150, 2021. 2 +[28] Garrick Orchard, Ryad Benosman, Ralph Etienne-Cummings, and Nitish V. Thakor. A spiking neural network architecture for visual motion estimation. In IEEE Biomedical Circuits Syst. Conf., pages 298-301. IEEE, 2013. 2 +[29] Federico Paredes-Valles and Guido C. H. E. de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2, 3, 4, 5, 8, 9 +[30] Federico Paredes-Vallés, Kirk Y. W. Scheper, and Guido C. H. E. de Croon. Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE Trans. Pattern Anal. and Mach. Intell., 42(8):2051-2064, 2020. 2 +[31] Wachirawit Ponghiran, Chamika Mihiranga Liiyanagedera, and Kaushik Roy. Event-based temporally dense optical flow estimation with sequential neural networks. arXiv:2210.01244, 2022. 2 +[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Int. Conf. Medical Image Computing and Comput.-Assisted Intervention, pages 234–241. Springer, 2015. 2 +[33] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event collapse in contrast maximization frameworks. Sensors, 22(14):5190, 2022. 8 +[34] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. A fast geometric regularizer to mitigate event collapse in the contrast maximization framework. arXiv:2212.07350, 2022. 7 +[35] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In Eur. Conf. Comput. Vis., 2022. 2, 3, 4, 5, 6, 7, 8, 9 +[36] Timo Stoffregen and Lindsay Kleeman. Event cameras, contrast maximization and reward functions: An analysis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12300-12308, 2019. 3 +[37] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In European Conf. Comput. Vis., 2020. 2, 6, 8 +[38] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. PWC-Net: CNNs for optical flow using pyramid, warping, and cost volume. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8934–8943, 2018. 2 +[39] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In Eur. Conf. Comput. Vis., pages 402-419. Springer, 2020. 2 +[40] Z Wan, Y Dai, and Y Mao. Learning dense and continuous optical flow from an event camera. IEEE Trans. Image Process., 2022. 2 + +[41] Qimin Wang, Yongjun Zhang, Shaowu Yang, Zhe Liu, Lin Wang, and Luoxi Jing. E-HANet: Event-based hybrid attention network for optical flow estimation. In IEEE Int. Conf. Syst., Man, Cybernetics, pages 2189-2194. IEEE, 2022. 2 +[42] Yilun Wu, Federico Paredes-Vallés, and Guido C. H. E. de Croon. Rethinking event-based optical flow: Iterative deblurring as an alternative to correlation volumes. arXiv:2211.13726, 2022. 2, 4, 6, 7 +[43] Alex Z. Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3D perception. IEEE Robot. and Autom. Lett., 3(3):2032-2039, 2018. 2, 6, 8 +[44] Alex Z. Zhu and Liangzhe Yuan. EV-FlowNet: Self-supervised optical flow estimation for event-based cameras. In Robot.: Science and Syst., 2018. 2, 6, 8 +[45] Alex Z. Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 989-997, 2019. 2, 3, 4, 5, 6, 8 \ No newline at end of file diff --git a/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/images.zip b/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..302c2265db8cef519189b680d7d4504cc35f1576 --- /dev/null +++ b/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf6d74aaac2a20a4fc82941dbd934cca402a3e34ad1105a79ff8b59d27c69051 +size 556339 diff --git a/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/layout.json b/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8ab29c96c4fe9ea2a9677673a22da318080e231f --- /dev/null +++ b/2023/Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow/layout.json @@ -0,0 +1,7586 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 103, + 455, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 103, + 455, + 140 + ], + "spans": [ + { + "bbox": [ + 138, + 103, + 455, + 140 + ], + "type": "text", + "content": "Taming Contrast Maximization for Learning Sequential, Low-latency, Event-based Optical Flow" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 160, + 556, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 160, + 556, + 176 + ], + "spans": [ + { + "bbox": [ + 52, + 160, + 556, + 176 + ], + "type": "text", + "content": "Federico Paredes-Valles" + }, + { + "bbox": [ + 52, + 160, + 556, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 52, + 160, + 556, + 176 + ], + "type": "text", + "content": " Kirk Y. W. Scheper" + }, + { + "bbox": [ + 52, + 160, + 556, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 160, + 556, + 176 + ], + "type": "text", + "content": " Christope De Wagter" + }, + { + "bbox": [ + 52, + 160, + 556, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 160, + 556, + 176 + ], + "type": "text", + "content": " Guido C. H. E. de Croon" + }, + { + "bbox": [ + 52, + 160, + 556, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 179, + 460, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 179, + 460, + 194 + ], + "spans": [ + { + "bbox": [ + 149, + 179, + 460, + 194 + ], + "type": "text", + "content": "1 Micro Air Vehicle Laboratory, Delft University of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 194, + 504, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 194, + 504, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 504, + 209 + ], + "type": "text", + "content": "2 Stuttgart Laboratory 1, Sony Semiconductor Solutions Europe, Sony Europe B.V." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 48, + 230, + 312, + 411 + ], + "blocks": [ + { + "bbox": [ + 48, + 230, + 312, + 411 + ], + "lines": [ + { + "bbox": [ + 48, + 230, + 312, + 411 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 312, + 411 + ], + "type": "image", + "image_path": "8eb4e758df6147d95bc6ce1673c5196874b8dc4b404d80f7475b506b70aea81d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 422, + 547, + 456 + ], + "lines": [ + { + "bbox": [ + 45, + 422, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 45, + 422, + 547, + 456 + ], + "type": "text", + "content": "Figure 1: Our pipeline estimates event-based optical flow by sequentially processing small partitions of the event stream with a recurrent model. We propose a novel self-supervised learning framework (left) based on a multi-timescale contrast maximization formulation that is able to exploit the high temporal resolution of event cameras via iterative warping to produce accurate optical flow predictions (right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 232, + 545, + 403 + ], + "blocks": [ + { + "bbox": [ + 328, + 232, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 328, + 232, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 328, + 232, + 545, + 403 + ], + "type": "image", + "image_path": "ebb4507e223e66d8b8413675910840cd4042c26e5346f46a69d49db706498050.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 465, + 192, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 465, + 192, + 477 + ], + "spans": [ + { + "bbox": [ + 143, + 465, + 192, + 477 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 490, + 290, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 490, + 290, + 706 + ], + "spans": [ + { + "bbox": [ + 45, + 490, + 290, + 706 + ], + "type": "text", + "content": "Event cameras have recently gained significant traction since they open up new avenues for low-latency and low-power solutions to complex computer vision problems. To unlock these solutions, it is necessary to develop algorithms that can leverage the unique nature of event data. However, the current state-of-the-art is still highly influenced by the frame-based literature, and usually fails to deliver on these promises. In this work, we take this into consideration and propose a novel self-supervised learning pipeline for the sequential estimation of event-based optical flow that allows for the scaling of the models to high inference frequencies. At its core, we have a continuously-running stateful neural model that is trained using a novel formulation of contrast maximization that makes it robust to nonlinearities and varying statistics in the input events. Results across multiple datasets confirm the effectiveness of our method, which establishes a new state of the art in terms of accuracy for approaches trained or optimized without ground truth." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 465, + 387, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 465, + 387, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 465, + 387, + 477 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 485, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 677 + ], + "type": "text", + "content": "Event cameras capture per-pixel log-brightness changes at microsecond resolution [13]. This operating principle results in a sparse and asynchronous visual signal that, under constant illumination, directly encodes information about the apparent motion (i.e., optical flow) of contrast in the image space. These cameras offer several advantages, such as low latency and robustness to motion blur [13], and hence hold the potential of a high-bandwidth estimation of this optical flow information. However, the event-based nature of the generated visual signal poses a paradigm shift in the processing pipeline, and traditional, frame-based algorithms become suboptimal and often incompatible. Despite this, the majority of learning-based methods that have been proposed so far for event-based optical flow estimation are still highly influenced by frame-based approaches. This influence is normally reflected in two key aspects of their" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "text", + "content": "The project's code and additional material can be found at https://mavlab.tudelft.nl/taming_event_flow/." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "9695" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "pipelines: (i) the design of the network architecture, and (ii) the formulation of the loss function." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 98, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 98, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 98, + 286, + 240 + ], + "type": "text", + "content": "Regarding architecture design, most literature methods format subsets of the input events as dense volumetric representations [45] that are processed at once by stateless (i.e., non-recurrent) models [17, 29, 37, 45]. Similarly to their frame-based counterparts [12, 38, 39], these models estimate the per-pixel displacement over the time-length of the event volume using only the information contained within it. Consequently, these volumes need to encode enough spatiotemporal information for motion to be discernible. However, if done over relatively long time periods, the subsequent models suffer from limitations such as high latency or having to deal with large pixel displacements [17, 18, 42]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 243, + 286, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 286, + 434 + ], + "type": "text", + "content": "With respect to the loss function, multiple options have been explored due to the lack of real-world datasets with per-event ground truth. Pure supervised learning can be used with datasets such as MVSEC [43] or DSEC-Flow [17], but their ground truth only contains per-pixel displacement at low frequencies, which makes models difficult to train. On the other hand, a self-supervised learning (SSL) framework can be formulated using either the accompanying frames with the photometric error as a loss [11, 40, 44], or through an events-only contrast maximization [14, 15] proxy loss for motion compensation (i.e., event deblurring) [19, 26, 29, 35, 45]. However, despite not relying on ground truth, all the literature on SSL for optical flow assumes that the events move linearly within the time window of the loss, which ignores much of the potential of event cameras and their high temporal resolution (see Fig. 1, bottom left)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 435, + 286, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 286, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 286, + 626 + ], + "type": "text", + "content": "In this work, we focus on the estimation of high frequency event-based optical flow and how this can be learned in an SSL fashion using contrast maximization with a relaxed linear motion assumption. To achieve this, we build upon the continuous-operation pipeline from Hagenaars et al. [19], which retrieves optical flow by sequentially processing small partitions of the event stream with a stateful (i.e., recurrent) model over time, instead of dealing with large volumes of input events. We augment (and train) this pipeline with a novel contrast maximization formulation that performs event motion compensation in an iterative manner at multiple temporal scales, as shown in Fig. 1. Using this framework, we achieve the best accuracy of all contrast-maximization-based approaches on multiple datasets, only being outperformed by pure supervised learning methods trained with ground-truth data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 628, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 286, + 651 + ], + "type": "text", + "content": "In summary, the extensions that we propose to the SSL framework in [19], i.e., our main contributions, are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 654, + 286, + 713 + ], + "type": "text", + "content": "- The first iterative event warping module in the context of contrast maximization (see Section 3.2). This module unlocks a novel multi-reference loss function that better captures the trajectory of scene points over time, thus improving the accuracy of the predictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 119 + ], + "type": "text", + "content": "- The first multi-timescale approach to contrast maximization, which adds robustness, improves convergence, and reduces tuning requirements of loss-related hyperparameters (see Section 3.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 121, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 121, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 121, + 545, + 216 + ], + "type": "text", + "content": "As a result, we present the first self-supervised optical flow method for event cameras that relaxes the linear motion assumption, and hence that has the potential of exploiting the high temporal resolution of the sensor by producing estimates in a close-to continuous manner. We validate the proposed framework through extensive evaluations on multiple datasets. Additionally, we conduct ablation studies to show the effectiveness of each individual component." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 227, + 389, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 227, + 389, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 227, + 389, + 239 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 247, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 533 + ], + "type": "text", + "content": "Due to the aforementioned advantages of event cameras for optical flow estimation, extensive research has been carried out since these sensors were first introduced [1, 2, 4-8, 24, 27, 28, 35]. Regarding learning-based approaches, the first method was proposed by Zhu et al. in [44] with EV-FlowNet: a UNet-like [32] architecture trained with SSL, with the supervisory signal coming from the photometric error between subsequent frames captured with an accompanying camera. To avoid the need for a secondary vision sensor, in [45], Zhu et al. proposed an SSL framework around the contrast maximization for motion compensation idea from [14, 15], hence relying solely on event data. This pipeline was used and further improved in [19, 29, 35]. Other approaches trained EV-FlowNet in a supervised fashion with synthetic/real ground-truth data and showed higher accuracy levels through evaluations on public benchmarks [17, 37]. However, they also highlighted EV-FlowNet's inability to deal with large pixel displacements [17]. Because of this, inspired by the frame-based literature [39], Gehrig et al. proposed E-RAFT in [17]: the first architecture to introduce the use of correlation volumes in the event camera literature. This model, which was recently augmented with attention mechanisms [41], achieved state-of-the-art performance in multiple datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "A novel perspective to learning event-based optical flow is to move away from event accumulation over long timespans, and instead rely on continuously-running stateful models that integrate information over time. The first methods of this kind were proposed by Paredes-Valles et al. [30] and Hagenaars et al. [19], but this idea has recently gained interest in the event camera literature [11, 31, 42]. The reason is that leveraging memory through sequential processing can potentially lead to lightweight and low-latency solutions that are also robust to large pixel displacements without the need for correlation volumes [42]. However, despite their potential of being scaled to high inference frequencies, all these solutions assume that events move linearly in the timespan of their loss function, and hence cannot capture the true trajectory of scene points over time." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9696" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "When it comes to learning nonlinear pixel trajectories from event data, only the work of Gehrig et al. in [18] is to be highlighted. However, their approach uses multiple event and correlation volumes to fit Bézier curves to the trajectory of scene points, resulting in a high-latency solution. In contrast, in this work, we extend the continuous-operation pipeline from Hagenaars et al. [19] with an SSL framework in which discrete trajectories are regressed at high frequency by leveraging memory within the models. This approach allows, for the first time, to exploit the high temporal resolution of event cameras while capturing more accurately the trajectory of scene points over time thanks to the proposed iterative event warping mechanism." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 385 + ], + "type": "text", + "content": "Among non-learning-based approaches, the work of Shiba et al. in [35] bears particular relevance due to certain similarities with our framework. Specifically, Shiba et al. propose a tile-based method for event-based optical flow estimation that extends contrast maximization [14, 15] by incorporating (i) multiple spatial scales, (ii) a multi-reference focus loss, and (iii) a \"time-aware\" optical flow formulation. Similarly, our learning-based approach also employs a multi-scale strategy and utilizes a multi-reference focus loss. However, instead of making assumptions about the events' motion over extended time periods, we propose to learn their potentially nonlinear trajectories at high frequency by leveraging the iterative event warping module." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 398, + 104, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 398, + 104, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 398, + 104, + 410 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 418, + 289, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 289, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 289, + 551 + ], + "type": "text", + "content": "The goal of this work is to learn to sequentially estimate optical flow at high frequencies from a continuous stream of events. In such a pipeline, if the inference frequency is sufficiently high, events need to be processed nearly as soon as they are triggered by the sensor, with no pre-processing in between. Because of this, the integration of temporal information needs to happen in the network itself. To accomplish this, we propose the framework in Fig. 1, in which a stateful model is trained using our novel formulation of contrast maximization for sequential processing. The components of this framework are described in the following sections." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 561, + 259, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 259, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 259, + 574 + ], + "type": "text", + "content": "3.1. Input format and contrast maximization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "text", + "content": "For an ideal camera, an event " + }, + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\boldsymbol{e}_i = (\\boldsymbol{x}_i, t_i, p_i)" + }, + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "text", + "content": " of polarity " + }, + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "inline_equation", + "content": "p_i \\in \\{+, -\\}" + }, + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "text", + "content": " is triggered at pixel " + }, + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i = (x_i, y_i)^T" + }, + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 580, + 287, + 640 + ], + "type": "text", + "content": " whenever the change in log-brightness since the last event at that pixel location reaches the contrast sensitivity threshold for that polarity [13]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "As in [19], we use a two-channel event count image as input representation, which gets populated with consecutive, non-overlapping, fine discrete partitions of the event stream " + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\varepsilon_{k}^{\\mathrm{inp}}\\doteq \\{e_{i}\\}" + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": " (further referred to as input partitions), each containing all the events in a time window of a certain duration, i.e., " + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "inline_equation", + "content": "t_i\\in [t_k^{\\mathrm{begin}},t_k^{\\mathrm{end}}]" + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": ". This representation does not" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": "contain temporal information by itself, and recurrent models are hence required for estimating of optical flow." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 96, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 216 + ], + "type": "text", + "content": "Regarding learning, we use the contrast maximization framework [14] to train, in an SSL fashion, to continuously estimate dense (i.e., per-pixel) optical flow from the event stream. Assuming brightness constancy, accurate flow information is encoded in the spatiotemporal misalignments (i.e., blur) among the events triggered by the same portion of a moving edge. To retrieve it, one has to compensate for this motion by geometrically transforming the events using a motion model. As in [19, 29, 35, 45], we transport each event to a reference time " + }, + { + "bbox": [ + 304, + 96, + 547, + 216 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 96, + 547, + 216 + ], + "type": "text", + "content": " through:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 369, + 224, + 545, + 238 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 224, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 369, + 224, + 545, + 238 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {i} ^ {\\prime} = \\boldsymbol {x} _ {i} + \\left(t _ {\\text {r e f}} - t _ {i}\\right) \\boldsymbol {u} \\left(\\boldsymbol {x} _ {i}\\right) \\tag {1}", + "image_path": "057590ecbfb9e96893ddd449e1c7203bfc86283535a18c87908c325ee8e4b8f0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "type": "inline_equation", + "content": "\\pmb{u}(\\pmb{x}) = (u(\\pmb{x}), v(\\pmb{x}))^T" + }, + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "type": "text", + "content": " denotes the optical flow map used to transport each event from " + }, + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "type": "text", + "content": ". The result of aggregating the transformed events is further referred to as the image of warped events (IWE) at " + }, + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 295, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 547, + 354 + ], + "type": "text", + "content": "As the loss function of our SSL framework, we adopt the time-based focus objective function from [19]. Using the warped events at a given " + }, + { + "bbox": [ + 304, + 295, + 547, + 354 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 295, + 547, + 354 + ], + "type": "text", + "content": ", we generate an image of the per-pixel average timestamps for each polarity " + }, + { + "bbox": [ + 304, + 295, + 547, + 354 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 304, + 295, + 547, + 354 + ], + "type": "text", + "content": " via bilinear interpolation:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 361, + 545, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 545, + 407 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} T _ {p ^ {\\prime}} (\\boldsymbol {x}; \\boldsymbol {u} | t _ {\\text {r e f}}) = \\frac {\\sum_ {j} \\kappa \\left(x - x _ {j} ^ {\\prime}\\right) \\kappa \\left(y - y _ {j} ^ {\\prime}\\right) \\bar {t} _ {j} \\left(t _ {\\text {r e f}} , t _ {j}\\right)}{\\sum_ {j} \\kappa \\left(x - x _ {j} ^ {\\prime}\\right) \\kappa \\left(y - y _ {j} ^ {\\prime}\\right) + \\epsilon} \\tag {2} \\\\ \\kappa (a) = \\max (0, 1 - | a |) \\\\ \\end{array}", + "image_path": "ac9cb671ec4247d6cd8d6910e92972bac2cde0714061a80f336fbdb29e1ee400.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 408, + 499, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 408, + 499, + 422 + ], + "spans": [ + { + "bbox": [ + 315, + 408, + 499, + 422 + ], + "type": "interline_equation", + "content": "j = \\{i \\mid p _ {i} = p ^ {\\prime} \\}, \\quad p ^ {\\prime} \\in \\{+, - \\}, \\quad \\epsilon \\approx 0", + "image_path": "e73ad3c78cbbd9dc0c7832f9851efae0c9272bdf8dfcd191727a5be197b6fc23.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 430, + 547, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 454 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 430, + 547, + 454 + ], + "type": "inline_equation", + "content": "\\bar{t}_i" + }, + { + "bbox": [ + 304, + 430, + 547, + 454 + ], + "type": "text", + "content": " denotes the normalized timestamp contribution of the " + }, + { + "bbox": [ + 304, + 430, + 547, + 454 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 430, + 547, + 454 + ], + "type": "text", + "content": "th event, according to Fig. 3 and Eq. 5." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 455, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 479 + ], + "type": "text", + "content": "Then, the contrast maximization loss function at " + }, + { + "bbox": [ + 304, + 455, + 545, + 479 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 455, + 545, + 479 + ], + "type": "text", + "content": " is defined as the scaled sum of the squared temporal images:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 486, + 545, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 486, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 321, + 486, + 545, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C M}} \\left(t _ {\\text {r e f}}\\right) = \\frac {\\sum_ {\\boldsymbol {x}} T _ {+} \\left(\\boldsymbol {x} ; \\boldsymbol {u} \\mid t _ {\\text {r e f}}\\right) ^ {2} + T _ {-} \\left(\\boldsymbol {x} ; \\boldsymbol {u} \\mid t _ {\\text {r e f}}\\right) ^ {2}}{\\sum_ {\\boldsymbol {x}} \\left[ n \\left(\\boldsymbol {x} ^ {\\prime}\\right) > 0 \\right] + \\epsilon} \\tag {3}", + "image_path": "f90431d3ea2baf18fce953f076a45b681a5336ffe95924d6ca7fa24bf06ee9b6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 521, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 546 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 521, + 545, + 546 + ], + "type": "inline_equation", + "content": "n(\\pmb{x}^{\\prime})" + }, + { + "bbox": [ + 304, + 521, + 545, + 546 + ], + "type": "text", + "content": " denotes a per-pixel event count of the IWE. The lower the " + }, + { + "bbox": [ + 304, + 521, + 545, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CM}}" + }, + { + "bbox": [ + 304, + 521, + 545, + 546 + ], + "type": "text", + "content": ", the better the event deblurring at " + }, + { + "bbox": [ + 304, + 521, + 545, + 546 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 521, + 545, + 546 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "As discussed in [14, 19, 36], for any focus objective function to be a robust supervisory signal for contrast maximization, the event partition used in the optimization needs to contain enough motion information (i.e., blur) so it can be compensated for. However, as in [19], that is not the case in our input partitions due to the fine discretization of the event stream that we are targeting. Therefore, only at training time, we define the so-called training partition (or event buffer in Fig. 1) " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\varepsilon_{k\\rightarrow k + R}^{\\mathrm{train}}\\doteq \\{\\varepsilon_i^{\\mathrm{inp}}\\}_{i = k}^{k + R}" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": ", which stacks together the events in the input partitions of " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": " successive forward passes through the network. Once these are performed, we use this partition and the estimated optical flow maps to compute the loss, and use truncated backpropagation through time to update the model parameters. After" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9697" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 276, + 453 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 276, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 276, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 276, + 453 + ], + "type": "image", + "image_path": "8eda2d9e871b781ae7697add87da18789d713eb2becd205c4eb6f65722bae766.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 463, + 287, + 584 + ], + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 584 + ], + "type": "text", + "content": "Figure 2: Incompatibility of multi-reference deblurring and linear event warping in the presence of nonlinearities in the pixel trajectories. Top: Events generated by a moving dot following circular motion (left), and optical flow color-coding scheme (right). Middle: Optical flow solutions required to produce sharp IWEs at different " + }, + { + "bbox": [ + 46, + 463, + 287, + 584 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 463, + 287, + 584 + ], + "type": "text", + "content": " using linear and iterative warping. While the former requires a different solution for each " + }, + { + "bbox": [ + 46, + 463, + 287, + 584 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 463, + 287, + 584 + ], + "type": "text", + "content": ", the proposed iterative warping can achieve this using a single solution. Arrows illustrate the direction of the required displacement " + }, + { + "bbox": [ + 46, + 463, + 287, + 584 + ], + "type": "inline_equation", + "content": "\\Delta \\pmb{x}_i = (t_{\\mathrm{ref}} - t_i)\\pmb{u}(\\pmb{x}_i)" + }, + { + "bbox": [ + 46, + 463, + 287, + 584 + ], + "type": "text", + "content": " at each (discretized) spatial location. Bottom: Resulting IWEs at different " + }, + { + "bbox": [ + 46, + 463, + 287, + 584 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 463, + 287, + 584 + ], + "type": "text", + "content": " using the optimal optical flow map for " + }, + { + "bbox": [ + 46, + 463, + 287, + 584 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}} = \\mathrm{mid}" + }, + { + "bbox": [ + 46, + 463, + 287, + 584 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 604, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 286, + 664 + ], + "type": "text", + "content": "this update, we detach the states of the network from the computational graph and clear the training partition and optical flow buffer. The importance of sequential processing, as an alternative to training stateless models in short input partitions, is corroborated in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 671, + 179, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 179, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 179, + 685 + ], + "type": "text", + "content": "3.2. Iterative event warping" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "In order to better approximate the trajectories of scene points over time, in this work we relax the linear motion as-" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 543, + 157 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 543, + 157 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 543, + 157 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 543, + 157 + ], + "type": "image", + "image_path": "ed19fc19c64c3b8d34fb8c174a5b4f787c59c9755326f2109289b1b3a79a8786.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "text", + "content": "Figure 3: Timestamp normalization profiles for the per-event contributions to the images of average timestamps in Eq. 2, for " + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "inline_equation", + "content": "R = 10" + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "text", + "content": ". Left: Deblurring only done at the extremes of the training partition, as in [19, 29, 45]. Right: Deblurring done at all reference times, thanks to the proposed iterative event warping. Normalization profiles only shown for three " + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "text", + "content": " for a better visualization." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 272, + 546, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 272, + 546, + 476 + ], + "spans": [ + { + "bbox": [ + 305, + 272, + 546, + 476 + ], + "type": "text", + "content": "sumption in the SSL framework for event-based optical flow by augmenting the sequential-estimation pipeline from [19] with iterative event warping. Instead of transporting events to a given " + }, + { + "bbox": [ + 305, + 272, + 546, + 476 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 305, + 272, + 546, + 476 + ], + "type": "text", + "content": " assuming linear motion regardless of the length of the warping interval (as in [19, 29, 35, 45]), we perform a finer discretization of the event trajectories and assume that motion is only linear between optical flow estimates. Therefore, to express a group of events at a given " + }, + { + "bbox": [ + 305, + 272, + 546, + 476 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 305, + 272, + 546, + 476 + ], + "type": "text", + "content": ", we geometrically transform them using all the intermediate optical flow estimates through multiple iterations of Eq. 1. Fig. 2 shows the differences between the linear event warping in [19] and the proposed iterative augmentation, as well as the limitations of the former. Note that our iterative warping is fundamentally different from that of the work of Wu et al. in [42], where input events are deblurred before being passed to the models using residual optical flow estimates until convergence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": "Literature methods on event-based optical flow with contrast maximization compute the focus objective function at multiple reference times in the training partition (usually at the extremes) to prevent overfitting and/or scaling issues during backpropagation [19, 29, 35, 45]. However, since these approaches assume optical flow constancy in the span of their loss function, they suffer with nonlinearities in the pixel trajectories (see the blurry IWEs in Fig. 2, bottom). On the contrary, because of the finer discretization of the event trajectories coming with our sequential processing pipeline and the proposed iterative warping, we can use any (combination of) reference time(s) for the computation of the focus objective function. In fact, as shown in Fig. 3 (right), we propose the use of all the discretization points as reference times for event deblurring. Apart from the aforementioned regularizing benefits, having to produce sharp IWEs at any " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " forces the models to estimate a sequence of optical flow maps that is consistent with the velocity profile of the event stream. For a given training partition of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9698" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 76, + 294, + 242 + ], + "blocks": [ + { + "bbox": [ + 49, + 76, + 294, + 242 + ], + "lines": [ + { + "bbox": [ + 49, + 76, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 49, + 76, + 294, + 242 + ], + "type": "image", + "image_path": "b48ea9605462c4fee827fe64233a73e117cf6ac63dfb0de8c3b87ffefb877f92.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 253, + 290, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 253, + 290, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 290, + 353 + ], + "type": "text", + "content": "Figure 4: Multi-timescale approach to contrast maximization. For a given training partition of length " + }, + { + "bbox": [ + 46, + 253, + 290, + 353 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 253, + 290, + 353 + ], + "type": "text", + "content": ", we fit multiple sub-partitions of different lengths (in the figure: one of length " + }, + { + "bbox": [ + 46, + 253, + 290, + 353 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 253, + 290, + 353 + ], + "type": "text", + "content": " in red, two " + }, + { + "bbox": [ + 46, + 253, + 290, + 353 + ], + "type": "inline_equation", + "content": "R/2" + }, + { + "bbox": [ + 46, + 253, + 290, + 353 + ], + "type": "text", + "content": " in blue, and four " + }, + { + "bbox": [ + 46, + 253, + 290, + 353 + ], + "type": "inline_equation", + "content": "R/4" + }, + { + "bbox": [ + 46, + 253, + 290, + 353 + ], + "type": "text", + "content": " in green) and compute the loss in each of them according to Eq. 4. The global loss is computed as in Eq. 6. This figure only shows the timestamp normalization profiles of the central " + }, + { + "bbox": [ + 46, + 253, + 290, + 353 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 253, + 290, + 353 + ], + "type": "text", + "content": " of each sub-partition, but the losses are still computed at all reference times. An image representation of the accumulated input events in a sub-partition of each timescale is also shown." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 373, + 218, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 218, + 385 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 218, + 385 + ], + "type": "text", + "content": "length " + }, + { + "bbox": [ + 47, + 373, + 218, + 385 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 373, + 218, + 385 + ], + "type": "text", + "content": ", the loss is computed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 393, + 287, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 393, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 287, + 427 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C M}} ^ {R} = \\frac {1}{R + 1} \\sum_ {t _ {\\text {r e f}} = 0} ^ {R} \\mathcal {L} _ {\\mathrm {C M}} \\left(t _ {\\text {r e f}}\\right) \\tag {4}", + "image_path": "e329beb9d0214f10361396a8fa05d555e22d496ee36c5dfc6c686340d8fefdf0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 435, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 435, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 435, + 287, + 471 + ], + "type": "text", + "content": "which ensures that the IWEs (and the corresponding images of average timestamps in Eq. 2) at all reference times " + }, + { + "bbox": [ + 47, + 435, + 287, + 471 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}} \\in [0, R]" + }, + { + "bbox": [ + 47, + 435, + 287, + 471 + ], + "type": "text", + "content": " contribute equally to the loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 471, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 471, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 287, + 555 + ], + "type": "text", + "content": "As in [19, 29, 45], for " + }, + { + "bbox": [ + 47, + 471, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CM}}(t_{\\mathrm{ref}})" + }, + { + "bbox": [ + 47, + 471, + 287, + 555 + ], + "type": "text", + "content": " in Eq. 3 to be a valid supervisory signal, events that are temporally close to the reference time " + }, + { + "bbox": [ + 47, + 471, + 287, + 555 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 47, + 471, + 287, + 555 + ], + "type": "text", + "content": " need to contribute to the temporal image in Eq. 2 with larger timestamp values than events that are far in time. Therefore, once the events have been transported to " + }, + { + "bbox": [ + 47, + 471, + 287, + 555 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 47, + 471, + 287, + 555 + ], + "type": "text", + "content": ", their timestamp is normalized prior to the computation of Eq. 2 as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 563, + 287, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 563, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 73, + 563, + 287, + 587 + ], + "type": "interline_equation", + "content": "\\bar {t} _ {i} \\left(t _ {\\text {r e f}}, t _ {i}\\right) = 1 - \\frac {\\left| t _ {\\text {r e f}} - t _ {i} \\right|}{R}, \\quad \\text {w i t h} t _ {i} \\in [ 0, R ] \\tag {5}", + "image_path": "6affc92b51afa5117ec5dc7b66584ebbd350cd42d4dc599194a9eb18ae991318.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 594, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 617 + ], + "type": "text", + "content": "which results in the normalization profiles in Fig. 3, for a given training partition of length " + }, + { + "bbox": [ + 47, + 594, + 287, + 617 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 594, + 287, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Inspired by [25], we mask individual events from the computation of the loss whenever we detect that they are transported outside the image space through the event warping process in the span of a deblurring window defined at the reference time " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ". This prevents our models from learning incorrect deformations at the image borders, and is motivated by the fact that the complete trajectory of the pixel is only partially observable in the training partition. An ab" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "lation study on the impact of this masking strategy can be found in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 96, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 96, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 96, + 547, + 168 + ], + "type": "text", + "content": "Lastly, we do not augment the loss function in Eq. 4 with smoothing priors acting as regularization mechanisms. With iterative event warping, the error propagates through all the pixels covered in the warping process, regardless of whether they have input events or not. Therefore, the spatial coherence of the resulting optical flow maps is enhanced." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 176, + 485, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 176, + 485, + 189 + ], + "spans": [ + { + "bbox": [ + 306, + 176, + 485, + 189 + ], + "type": "text", + "content": "3.3. Deblurring at multiple timescales" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 194, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 422 + ], + "type": "text", + "content": "Despite the addition of iterative event warping, the success of our SSL framework still heavily depends on the hyperparameters that control the amount of motion information perceived by the networks in the span of a deblurring window. In our pipeline, these are: " + }, + { + "bbox": [ + 304, + 194, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathrm{dt}_{\\mathrm{input}}" + }, + { + "bbox": [ + 304, + 194, + 545, + 422 + ], + "type": "text", + "content": ", the timestep used to discretize the event stream; and " + }, + { + "bbox": [ + 304, + 194, + 545, + 422 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 194, + 545, + 422 + ], + "type": "text", + "content": ", the number of forward passes, and hence optical flow maps, per loss. Thus, the effective length (in units of time) of the event window used for motion compensation is given by " + }, + { + "bbox": [ + 304, + 194, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathrm{dt}_{\\mathrm{input}} \\times R" + }, + { + "bbox": [ + 304, + 194, + 545, + 422 + ], + "type": "text", + "content": ". We hypothesize that, for each training dataset, there is an optimal length for this window that depends on the statistics of the data (e.g., event density, distribution of optical flow magnitudes) and model architecture, and that deviations from this optimal length lead to the learning of suboptimal solutions. E.g., shorter windows may converge to solutions that are more selective to fast rather than slow moving objects, and vice versa. Note that not only our method is sensitive to the tuning of these parameters, but also previous approaches based on contrast maximization [19, 29, 35, 45]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "text", + "content": "To add a layer of robustness to the framework and relax its strong dependency on hyperparameter optimization, we propose the multi-timescale approach illustrated in Figs. 1 and 4. For a given training partition of length " + }, + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "text", + "content": ", instead of computing a single focus loss through Eq. 4, we compute this loss at " + }, + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "text", + "content": " temporal scales of length " + }, + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "inline_equation", + "content": "R / 2^s" + }, + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "inline_equation", + "content": "0 \\leq s \\leq S - 1" + }, + { + "bbox": [ + 304, + 422, + 547, + 506 + ], + "type": "text", + "content": ", and combine them as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 363, + 515, + 545, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 515, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 363, + 515, + 545, + 548 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C M}} ^ {\\text {m u l t i}} = \\frac {1}{S} \\sum_ {s = 0} ^ {S - 1} \\frac {1}{2 ^ {s}} \\sum_ {p = 0} ^ {2 ^ {s} - 1} \\mathcal {L} _ {\\mathrm {C M}, p} ^ {R / 2 ^ {s}} \\tag {6}", + "image_path": "90ff862ca5e19f2561e60eada6cf26f53ff8e8bea847549a05bf3205d2a156f2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "type": "text", + "content": "As shown, we fit multiple non-overlapping sub-partitions in the training buffer if " + }, + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "type": "inline_equation", + "content": "s > 0" + }, + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "type": "text", + "content": ". The subscript " + }, + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "type": "text", + "content": " indicates their location in this buffer, starting from the earliest (see Fig. 4)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Note that, through this multi-timescale approach to contrast maximization, our models need to converge to a solution that is suitable for all the timescales in the optimization, regardless of their length. An alternative formulation would be to incorporate per-pixel learnable masks (i.e., an attention module in the loss space) so that, depending on the input statistics, learning only happens at the most adequate scale. However, for this to happen, the loss function would have to be augmented to stimulate this behavior, and it is unclear how that would be done in practice." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9699" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 281, + 140 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 281, + 140 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 281, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 281, + 140 + ], + "type": "image", + "image_path": "1452b0e4edbc9fba7d891bbd856f0521d25a10cf94e5c12fc762238f3f7a6a94.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 150, + 287, + 196 + ], + "lines": [ + { + "bbox": [ + 46, + 150, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 46, + 150, + 287, + 196 + ], + "type": "text", + "content": "Figure 5: Schematic of the model architecture used in this work. It is characterized by " + }, + { + "bbox": [ + 46, + 150, + 287, + 196 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{G}}" + }, + { + "bbox": [ + 46, + 150, + 287, + 196 + ], + "type": "text", + "content": " recurrent encoders, " + }, + { + "bbox": [ + 46, + 150, + 287, + 196 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 150, + 287, + 196 + ], + "type": "text", + "content": " residual blocks, and " + }, + { + "bbox": [ + 46, + 150, + 287, + 196 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{G}}" + }, + { + "bbox": [ + 46, + 150, + 287, + 196 + ], + "type": "text", + "content": " decoder layers. Optical flow estimates are produced at all decoder levels. In this diagram, " + }, + { + "bbox": [ + 46, + 150, + 287, + 196 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{G}} = 2" + }, + { + "bbox": [ + 46, + 150, + 287, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 150, + 287, + 196 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{R}} = 1" + }, + { + "bbox": [ + 46, + 150, + 287, + 196 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 215, + 170, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 215, + 170, + 226 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 170, + 226 + ], + "type": "text", + "content": "3.4. Network architecture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 233, + 289, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 289, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 289, + 413 + ], + "type": "text", + "content": "We use the recurrent version of EV-FlowNet [44] proposed in " + }, + { + "bbox": [ + 46, + 233, + 289, + 413 + ], + "type": "inline_equation", + "content": "[19]^1" + }, + { + "bbox": [ + 46, + 233, + 289, + 413 + ], + "type": "text", + "content": " (see Fig. 5). The events are represented as event count images (see Section 3.1), then passed through four encoders with strided convolutions followed by ConvGRUs [3] (channels doubling, starting from 64), two residual blocks [20], and then four decoders performing bilinear upsampling followed by convolution. After each decoder, there is a skip connection (using element-wise summation) from the corresponding encoder, as well as a depthwise convolution to produce estimates at lower scales, which are then concatenated with the activations of the previous decoder. Note that the proposed focus loss function (see Eq. 6) is applied to each intermediate optical flow estimate via upsampling. Lastly, all layers use " + }, + { + "bbox": [ + 46, + 233, + 289, + 413 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 46, + 233, + 289, + 413 + ], + "type": "text", + "content": " kernels and ReLU activations except for the prediction layers, which use TanH." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 423, + 128, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 128, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 128, + 437 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 443, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 586 + ], + "type": "text", + "content": "We evaluate our method on the DSEC-Flow [16, 17] and MVSEC [43] datasets. We evaluate the accuracy of the predictions based on the following metrics: (i) EPE (lower is better, ↓), the endpoint error; (ii) " + }, + { + "bbox": [ + 46, + 443, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\%_{3\\mathrm{PE}}(\\downarrow)" + }, + { + "bbox": [ + 46, + 443, + 287, + 586 + ], + "type": "text", + "content": ", the percentage of points with EPE greater than 3 pixels; (iii) FWL (" + }, + { + "bbox": [ + 46, + 443, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 443, + 287, + 586 + ], + "type": "text", + "content": ") [37], a deblurring quality metric based on the variance of the IWEs; and (iv) RSAT (" + }, + { + "bbox": [ + 46, + 443, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 443, + 287, + 586 + ], + "type": "text", + "content": ") [19], a deblurring quality metric based on the per-pixel average timestamps of the IWEs. We compare our solution to the published baselines, which range from supervised learning (SL) methods trained with ground truth, to SSL methods trained with grayscale images (" + }, + { + "bbox": [ + 46, + 443, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{SSL}_{\\mathrm{F}}" + }, + { + "bbox": [ + 46, + 443, + 287, + 586 + ], + "type": "text", + "content": ") or events (" + }, + { + "bbox": [ + 46, + 443, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{SSL}_{\\mathrm{E}}" + }, + { + "bbox": [ + 46, + 443, + 287, + 586 + ], + "type": "text", + "content": "), and model-based approaches (MB)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 587, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 671 + ], + "type": "text", + "content": "We train all our models on a subset of sequences from the training dataset of DSEC-Flow (only daylight recordings, see supplementary material). This corresponds to 19 minutes of training data, which we split into " + }, + { + "bbox": [ + 46, + 587, + 287, + 671 + ], + "type": "inline_equation", + "content": "572128 \\times 128" + }, + { + "bbox": [ + 46, + 587, + 287, + 671 + ], + "type": "text", + "content": " (randomly-cropped) sequences of 2 seconds each. We use a batch size of 8 and train until convergence with the Adam optimizer [21] and a learning rate of 1e-5. To keep mem" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 547, + 175 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 547, + 175 + ], + "type": "image", + "image_path": "6151cc11d5d61687c1d5afe4a6a2e089ccd485066c3a3f26d6fb01ba747c1a0e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 185, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 305, + 185, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 305, + 185, + 547, + 229 + ], + "type": "text", + "content": "Figure 6: Reconstruction of the pixel displacement of a scene point in the time window of a ground-truth sample from the multiple optical flow maps estimated in this period, i.e., " + }, + { + "bbox": [ + 305, + 185, + 547, + 229 + ], + "type": "inline_equation", + "content": "\\forall \\pmb{u}_k \\in [t_{\\mathrm{begin}}, t_{\\mathrm{end}}]" + }, + { + "bbox": [ + 305, + 185, + 547, + 229 + ], + "type": "text", + "content": ". The error of the last optical flow estimate is magnified for clarity." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 240, + 545, + 427 + ], + "blocks": [ + { + "bbox": [ + 308, + 240, + 545, + 427 + ], + "lines": [ + { + "bbox": [ + 308, + 240, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 545, + 427 + ], + "type": "table", + "html": "
EPE↓%3PE↓FWL↑RSAT↓
SLE-RAFT [17]0.792.681.330.87
EV-FlowNet, Gehrig et al. [17]2.3218.60--
Gehrig et al. [18]0.752.44--
IDNet [42]0.722.04--
TIDNet [42]0.843.41--
TMA [23]0.742.30--
Cuadrado et al. [9]1.7110.31--
E-Flowformer [22]0.762.68--
SSLEEV-FlowNet* [45]3.8631.451.300.85
ConvGRU-EV-FlowNet* [19]4.2733.271.550.90
dt = 0.01s, R = 2, S = 1 (Ours)9.6686.441.911.07
dt = 0.01s, R = 5, S = 1 (Ours)4.0552.221.580.97
dt = 0.01s, R = 10, S = 1 (Ours)2.3317.771.260.88
dt = 0.01s, R = 20, S = 1 (Ours)16.6333.671.061.10
dt = 0.01s, R = 10, S = 3 (Ours)2.8227.091.370.92
dt = 0.01s, R = 20, S = 4 (Ours)2.7323.731.240.90
MBShiba et al. [35]3.4730.861.370.89
", + "image_path": "88ca45d1e652da105f3e01188fe7e03bb1c4f09f3a78b632db5252f6a97fd9e0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 322, + 427, + 457, + 437 + ], + "lines": [ + { + "bbox": [ + 322, + 427, + 457, + 437 + ], + "spans": [ + { + "bbox": [ + 322, + 427, + 457, + 437 + ], + "type": "text", + "content": "*Retrained by us on DSEC-Flow, linear warping." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 445, + 545, + 479 + ], + "lines": [ + { + "bbox": [ + 305, + 445, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 545, + 479 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation on the DSEC-Flow dataset [17]. Best in bold, runner up underlined. A breakdown of the results is provided in the supplementary material." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 495, + 545, + 544 + ], + "type": "text", + "content": "ory usage within limits, we only propagate error gradients through up to 1e3 randomly-chosen events per millisecond of data. Despite this, note that we warp and use all the input events for the computation of the loss function." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 552, + 430, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 430, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 430, + 564 + ], + "type": "text", + "content": "4.1. Evaluation procedure" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "When evaluating our sequential models, if " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{dt_{gt}} > \\mathrm{dt_{input}}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ", we need to reconstruct the estimated per-pixel displacement in the ground-truth time window from the multiple optical flow maps estimated in this period. We do this by first averaging the (bilinearly interpolated) optical flow vectors that describe the trajectory of each scene point, and then by scaling the resulting optical flow vectors by " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{dt_{gt}} / \\mathrm{dt_{input}}" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ". This converts them from units of pixels/input to units of pixel displacement. An illustration of this reconstruction is shown in Fig. 6 for a scene point following a nonlinear trajectory. Note that our solution is subject to cumulative errors when evaluated through this reconstruction on benchmarks with" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": "1Our architecture is equivalent to ConvGRU-EV-FlowNet [19]. However, for the purpose of clarity within the rest of the paper, we will use this term to specifically refer to the original model from [19]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9700" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 68, + 534, + 428 + ], + "blocks": [ + { + "bbox": [ + 60, + 68, + 534, + 428 + ], + "lines": [ + { + "bbox": [ + 60, + 68, + 534, + 428 + ], + "spans": [ + { + "bbox": [ + 60, + 68, + 534, + 428 + ], + "type": "image", + "image_path": "b82cc80d8f8247aa65493bb77918a1efe77d1ff9f0da5a3212529b8f02586a7e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 435, + 546, + 469 + ], + "lines": [ + { + "bbox": [ + 46, + 435, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 546, + 469 + ], + "type": "text", + "content": "Figure 7: Qualitative comparison of our method with the state-of-the-art E-RAFT architecture [17] and the model-based approach from Shiba et al. [35] on sequences from the test partition of the DSEC-Flow dataset [17]. Ground truth not included due to unavailability. The optical flow color coding can be found in Fig. 2 (top), and the corresponding IWEs in the supplementary material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 489, + 288, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 288, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 288, + 539 + ], + "type": "text", + "content": "ground truth provided at low rates (e.g., " + }, + { + "bbox": [ + 46, + 489, + 288, + 539 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 489, + 288, + 539 + ], + "type": "text", + "content": " in DSEC-Flow [17]). Therefore, it will compare unfavorably to other, non-sequential, methods that only produce a single optical flow estimate in the timespan of a ground-truth sample." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 550, + 178, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 178, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 178, + 562 + ], + "type": "text", + "content": "4.2. Optical flow evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Evaluation on DSEC-Flow: Quantitative results of our evaluation on DSEC-Flow are presented in Table 1, and are supported by the qualitative comparison in Fig. 7. For this experiment, we trained multiple models with the same " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{dt}_{\\mathrm{input}} = 0.01\\mathrm{s}" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " faster than DSEC's ground truth) but different lengths of the training partition, and with and without the multi-timescale approach. Multiple conclusions can be derived from the reported results. Firstly, our best performing model (i.e., " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "R = 10" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "S = 1" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": ") achieves the best accuracy of all contrast-maximization-based approaches on this dataset according to the EPE and the percentage of outliers. Specifically, it outperforms the baselines with an im" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 489, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 545, + 597 + ], + "type": "text", + "content": "provement in the EPE in the " + }, + { + "bbox": [ + 304, + 489, + 545, + 597 + ], + "type": "inline_equation", + "content": "33\\% - 45\\%" + }, + { + "bbox": [ + 304, + 489, + 545, + 597 + ], + "type": "text", + "content": " range, only being outperformed by SL methods trained with ground truth on the same dataset [9, 17, 22, 23, 42]. This confirms that (i) the timestamp-based loss function in Section 3.1 allows us to learn accurate event-based optical flow (contrary to the findings of [34]); and that (ii) our augmentations to the sequential pipeline in [19] lead to a significant improvement in the accuracy of the model (i.e., " + }, + { + "bbox": [ + 304, + 489, + 545, + 597 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 304, + 489, + 545, + 597 + ], + "type": "text", + "content": " improvement in the EPE)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "content": "Secondly, these results also confirm our hypothesis that, for each training dataset, there is an optimal length for the training partition " + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "content": " in terms of the EPE. According to Table 1, the optimal " + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "content": " for this dataset, our model architecture, and our " + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{dt}_{\\mathrm{input}}" + }, + { + "bbox": [ + 304, + 605, + 546, + 715 + ], + "type": "text", + "content": " is 10 (i.e., 0.1s of event data), with the EPE increasing if the training partition is made shorter or longer. However, as also shown in this table, we can relax the strong dependency of the contrast maximization framework on this parameter through the proposed multi-timescale approach." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9701" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 68, + 257, + 139 + ], + "blocks": [ + { + "bbox": [ + 80, + 68, + 257, + 139 + ], + "lines": [ + { + "bbox": [ + 80, + 68, + 257, + 139 + ], + "spans": [ + { + "bbox": [ + 80, + 68, + 257, + 139 + ], + "type": "table", + "html": "
EPE↓%3PE↓FWL↑RSAT↓
dt = 0.1s*3.4834.720.980.87
dt = 0.05s3.0927.361.110.91
dt = 0.01s2.3317.771.260.88
dt = 0.005s2.3417.921.380.89
dt = 0.002s2.6621.832.040.90
", + "image_path": "9c6a4d58e77599920c4ed2a0d6115fbe1fc498ddbb7a3cab9d673573aee2897a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 84, + 140, + 230, + 149 + ], + "lines": [ + { + "bbox": [ + 84, + 140, + 230, + 149 + ], + "spans": [ + { + "bbox": [ + 84, + 140, + 230, + 149 + ], + "type": "text", + "content": "*Non-recurrent, volum. event repr. with 10 bins." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 198, + 286, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 286, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 286, + 305 + ], + "type": "text", + "content": "Our " + }, + { + "bbox": [ + 46, + 198, + 286, + 305 + ], + "type": "inline_equation", + "content": "S > 1" + }, + { + "bbox": [ + 46, + 198, + 286, + 305 + ], + "type": "text", + "content": " models converged to solutions that slightly underperform our best performing single-scale model (EPE went up by " + }, + { + "bbox": [ + 46, + 198, + 286, + 305 + ], + "type": "inline_equation", + "content": "17\\% - 21\\%" + }, + { + "bbox": [ + 46, + 198, + 286, + 305 + ], + "type": "text", + "content": "), but were trained without the need to fine-tune the length of the training partition. Note that, despite this slight drop in accuracy, these multi-timescale solutions still outperform the other non-SL baselines. To further support these results, a visualization of the distribution of the EPE of our models as a function of the ground truth magnitude is provided in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 306, + 286, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 286, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 286, + 425 + ], + "type": "text", + "content": "Lastly, Table 1 also allow us to conclude that deblurring quality metrics FWL [37] and RSAT [19] are not reliable indicators of the quality of the estimated optical flow. The reason for this is their inability to capture \"event collapse\" issues (as described in [33]), and would give favorable scores to undesirable solutions that warp all the events into a few pixels. According to our results, the FWL metric, being the spatial variance of the IWE relative to that of the identity warp, suffers more from this issue: the best FWL value is obtained with a model with 9.66 EPE." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 286, + 664 + ], + "type": "text", + "content": "Regarding qualitative results, Fig. 7 shows a comparison of our best performing model with the state-of-the-art E-RAFT architecture [17] and the contrast-maximization-based approach from Shiba et al. [35] on multiple sequences from the test partition of DSEC-Flow (i.e., ground truth is unavailable). These results confirm that our models are able to estimate high quality event-based optical flow despite not having access to ground-truth data during training, and also show the superiority of our method over the current best contrast-maximization-based approach [35]. Two limit cases in which our models provide suboptimal solutions are also shown in this figure: (i) sequences recorded at night (e.g., zurich_city_12_a) due to the presence of large amounts of events triggered by flashing lights and not by motion; and (ii) the car hood, which is also problematic for E-RAFT (i.e., does not capture it) and for [35]. Note that, in our case, (ii) is an artifact of the pixel displacements reconstructed from multiple optical flow estimates, and could be mitigated by having an occlusion handling mechanism in this reconstruction process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "In addition to the evaluation in Table 1 and Fig. 7, we also conducted an experiment in which we trained multiple models with different " + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{dt}_{\\mathrm{input}}" + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": " (ranging from 0.1s to 0.002s) but with the same amount of information in the training par" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 65, + 529, + 239 + ], + "blocks": [ + { + "bbox": [ + 46, + 159, + 287, + 182 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 287, + 182 + ], + "type": "text", + "content": "Table 2: Impact of the input window length on the DSEC-Flow dataset [17]. Best in bold, runner up underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 65, + 529, + 239 + ], + "lines": [ + { + "bbox": [ + 326, + 65, + 529, + 239 + ], + "spans": [ + { + "bbox": [ + 326, + 65, + 529, + 239 + ], + "type": "table", + "html": "
EPE↓%3PE↓
SLEV-FlowNet+ [37]0.680.99
E-RAFT [17]0.241.70
EV-FlowNet, Gehrig et al. [17]0.310.00
TMA [23]0.250.07
Cuadrado et al. [9]0.85-
SSLFEV-FlowNet, Zhu et al. [44]0.490.20
Ziluo et al. [11]0.420.00
SSLEEV-FlowNet, Zhu et al. [45]0.320.00
EV-FlowNet, Paredes-Vallés et al. [29]0.925.40
EV-FlowNet, Shiba et al. [35]0.360.09
ConvGRU-EV-FlowNet [19]0.470.25
Ours0.270.05
MBAkolkar et al. [1]2.75-
Brebion et al. [7]0.530.20
Shiba et al. [35]0.300.11
", + "image_path": "171d97d808feb2436182a3a948df99a423cc0ed960d33a654f82ba7f1fdb74a5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 247, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 281 + ], + "type": "text", + "content": "Table 3: Quantitative evaluation on MVSEC's outdoor_1 sequence [43]. Best in bold, runner up underlined. Results on the indoor sequences can be found in the supplementary material." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 437 + ], + "type": "text", + "content": "tition: 0.1s of event data. Results in Table 2 show that our sequential pipeline leads to an improvement in the accuracy of the predicted optical flow maps with respect to the stateless EV-FlowNet, which processes the 0.1s of event data at once. This improvement is due to the fact that the complexity of dealing with large pixel displacements gets reduced when processing the input data sequentially using shorter input windows. In addition to this, Table 2 also shows that the accuracy of our models is not compromised when estimating optical flow at higher frequencies, despite the high sparsity levels in the input data at those rates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 437, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 713 + ], + "type": "text", + "content": "Evaluation on MVSEC: Quantitative results of our evaluation on the outdoor_day1 sequence from MVSEC are presented in Table 3, and are supported by the qualitative comparison in the supplementary material. For this experiment, since (i) there is no consensus in the literature with respect to the training dataset [17, 19, 29, 35, 37, 44, 45], and (ii) the outdoor_day2 sequence (i.e., the other daylight, automotive sequence) is only 9 minutes of duration during which the event camera is subject to high frequency vibrations [43], we decided to transfer one of our models trained on DSEC-Flow to MVSEC. More specifically, we chose the model trained with " + }, + { + "bbox": [ + 304, + 437, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{dt}_{\\mathrm{input}} = 0.005\\mathrm{s}" + }, + { + "bbox": [ + 304, + 437, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 545, + 713 + ], + "type": "inline_equation", + "content": "R = 20" + }, + { + "bbox": [ + 304, + 437, + 545, + 713 + ], + "type": "text", + "content": " from Table 2, as a model trained with a short input window on DSEC-Flow is expected to be robust to the slow motion statistics of MVSEC [17]. We deployed the model at the same frequency as the temporally-upsampled ground truth (i.e., " + }, + { + "bbox": [ + 304, + 437, + 545, + 713 + ], + "type": "inline_equation", + "content": "45\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 437, + 545, + 713 + ], + "type": "text", + "content": "). Results in Table 3 show that, by doing this, our model outperforms the great majority of methods in terms of the EPE (even some SL methods trained on this dataset), and is only surpassed by the current state-of-the-art E-RAFT architecture [17]. Besides reaffirming the high quality of the produced optical flow estimates, these results also confirm the generalizability of our method. For com" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9702" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": "pleteness, the results on the indoor evaluation sequences from MVSEC are provided in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 106, + 123, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 123, + 119 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 123, + 119 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 127, + 289, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 127, + 289, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 127, + 289, + 438 + ], + "type": "text", + "content": "The self-supervised method for event-based optical flow presented in this work, while demonstrating highly accurate and promising results, is not without limitations. Two critical challenges that need to be acknowledged are the brightness constancy assumption and the aperture problem. Firstly, the contrast maximization framework [14, 15] assumes constant illumination, leading our models to face difficulty in learning from events that are not due to motion in the image space but that arise from changes in illumination. Since this limitation is inherent to contrast maximization, it extends to other approaches based on the same principle [19, 29, 35]. Due to this assumption, we excluded sequences recorded at night from our training dataset (see supplementary material). Secondly, akin to many other optical flow methods, our approach is susceptible to the aperture problem. This indicates that only motion components normal to the orientation of an edge in the image space, also known as normal optical flow, can be reliably resolved [10]. Consequently, the proposed method might face challenges in accurately determining the true motion direction in certain ambiguous scenarios. The regularizing effect of the iterative event warping (see Section 3.2) and the multiple spatial scales at which dense optical flow is estimated in our architecture (see Section 3.4) are mechanisms in our proposed solution that collectively strive to counteract the aperture problem's influence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 449, + 119, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 119, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 119, + 460 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 468, + 289, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 289, + 697 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 289, + 697 + ], + "type": "text", + "content": "In this paper, we presented the first learning-based approach to event-based optical flow estimation that is scalable to high inference frequencies while being able to accurately capture the true trajectory of scene points over time. The proposed pipeline is designed around a continuously-running stateful model that sequentially processes fine discrete partitions of the input event stream while integrating spatiotemporal information. We train this model through a novel, self-supervised, contrast maximization framework (i.e., event deblurring for supervision) that is characterized by an iterative event warping module and a multi-timescale loss function that add robustness and improve the accuracy of the predicted optical flow maps. We demonstrated the effectiveness of our approach on multiple datasets, where our models outperform the self-supervised and model-based baselines by large margins. Future research should look into how to learn to better combine the information from multiple timescales, as well as into the design of lightweight architectures that can keep up with real-time constraints." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 182 + ], + "type": "text", + "content": "We believe that the proposed approach opens up avenues for future research, especially in the field of neuromorphic computing. Spiking networks running on neuromorphic hardware have the potential of exploiting the main benefits of event cameras, but for that they need to process the input events shortly after they arrive. Our proposed framework is a step toward this objective, as it enables the estimation of optical flow in a close to continuous manner, with all the integration of information happening within the model itself." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 193, + 365, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 193, + 365, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 193, + 365, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 212, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 310, + 212, + 547, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 212, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 310, + 212, + 547, + 258 + ], + "type": "text", + "content": "[1] Himanshu Akolkar, Sio-Hoi Ieng, and Ryad Benosman. Real-time high speed motion prediction using fast aperture-robust event-driven visual flow. IEEE Trans. Pattern Anal. Mach. Intell., 44(1):361-372, 2020. 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 258, + 547, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 258, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 310, + 258, + 547, + 303 + ], + "type": "text", + "content": "[2] Mohammed Almatrafi, Raymond Baldwin, Kiyoharu Aizawa, and Keigo Hirakawa. Distance surface for event-based optical flow. IEEE Trans. Pattern Anal. and Mach. Intell., 42(7):1547-1556, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 304, + 547, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 304, + 547, + 347 + ], + "spans": [ + { + "bbox": [ + 311, + 304, + 547, + 347 + ], + "type": "text", + "content": "[3] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. Int. Conf. Learn. Representations, 2015. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 350, + 547, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 350, + 547, + 394 + ], + "spans": [ + { + "bbox": [ + 311, + 350, + 547, + 394 + ], + "type": "text", + "content": "[4] Patrick Bardow, Andrew J. Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In IEEE Conf. Comput. Vis. Pattern Recog., pages 884-892, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 396, + 547, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 396, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 311, + 396, + 547, + 429 + ], + "type": "text", + "content": "[5] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE Trans. Neural Netw. Learn. Syst., 25(2):407-417, 2013. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 430, + 547, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 430, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 311, + 430, + 547, + 474 + ], + "type": "text", + "content": "[6] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous frameless event-based optical flow. Neural Networks, 27:32-37, 2012. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 475, + 547, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 475, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 311, + 475, + 547, + 520 + ], + "type": "text", + "content": "[7] Vincent Brebion, Julien Moreau, and Franck Davoine. Realtime optical flow for vehicular perception with low-and high-resolution event cameras. IEEE Trans. on Intell. Transp. Systems, 2021. 2, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 521, + 547, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 521, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 311, + 521, + 547, + 553 + ], + "type": "text", + "content": "[8] Tobias Brosch, Stephan Tschechne, and Heiko Neumann. On event-based optical flow detection. Frontiers in Neuroscience, 9:137, 2015. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 556, + 547, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 556, + 547, + 610 + ], + "spans": [ + { + "bbox": [ + 311, + 556, + 547, + 610 + ], + "type": "text", + "content": "[9] Javier Cuadrado, Ulysse Rancon, Benoit R. Cotterreau, Francisco Barranco, and Timothee Masquelier. Optical flow estimation from event-based cameras and spiking neural networks. Frontiers in Neuroscience, 17:1160034, 2023. 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 613, + 547, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 547, + 656 + ], + "type": "text", + "content": "[10] David B. de Jong, Federico Paredes-Vallés, and Guido C. H. E. de Croon. How do neural networks estimate optical flow? A neuropsychology-inspired study. IEEE Trans. Pattern Anal. Mach. Intell., 44(11):8290-8305, 2021. 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "text", + "content": "[11] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-temporal recurrent networks for event-based optical flow estimation. In Proc. AAAI Conf. on Artificial Intell., volume 36, pages 525-533, 2022. 2, 8" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9703" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. FlowNet: Learning optical flow with convolutional networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2758-2766, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 195 + ], + "type": "text", + "content": "[13] Guillermo Gallego, Tobi Delbruck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Trans. Pattern Anal. and Mach. Intell., 2020. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 239 + ], + "type": "text", + "content": "[14] Guillermo Gallego, Mathias Gehrig, and Davide Scaramuzza. Focus is all you need: Loss functions for event-based vis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12280-12289, 2019. 2, 3, 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 296 + ], + "type": "text", + "content": "[15] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3867-3876, 2018. 2, 3, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "text", + "content": "[16] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A stereo event camera dataset for driving scenarios. IEEE Robot. and Autom. Lett., 6(3):4947-4954, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 385 + ], + "type": "text", + "content": "[17] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In Int. Conf. 3D Vis., pages 197–206. IEEE, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "type": "text", + "content": "[18] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from events and frames. arXiv:2203.13674, 2022. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 475 + ], + "type": "text", + "content": "[19] Jesse J. Hagenaars, Federico Paredes-Valles, and Guido C. H. E. de Croon. Self-supervised learning of event-based optical flow with spiking neural networks. In Advances in Neural Information Process. Syst., volume 34, pages 7167–7179, 2021. 2, 3, 4, 5, 6, 7, 8, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 770-778, 2016. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 543 + ], + "type": "text", + "content": "[21] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Int. Conf. Learn. Representations, 2014. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "text", + "content": "[22] Yijin Li, Zhaoyang Huang, Shuo Chen, Xiaoyu Shi, Hongsheng Li, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Blinkflow: A dataset to push the limits of event-based optical flow estimation. arXiv:2303.07716, 2023. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[23] Haotian Liu, Guang Chen, Sanqing Qu, Yanping Zhang, Zhijun Li, Alois Knoll, and Changjun Jiang. TMA: Temporal motion aggregation for event-based optical flow. arXiv:2303.11629, 2023. 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[24] Min Liu and Tobi Delbruck. Adaptive time-slice block-matching optical flow algorithm for dynamic vision sensors. In Brit. Mach. Vis. Conf., 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[25] Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsupervised learning of optical flow with a bidirectional census loss. In Proc. AAAI Conf. on Artificial Intell., volume 32, 2018. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[26] Anton Mitrokhin, Cornelia Fermüller, Chethan Parameshwara, and Yiannis Aloimonos. Event-based moving object detection and tracking. In IEEE/RSJ Int. Conf. Intell. Robots and Syst., 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "text", + "content": "[27] Jun Nagata, Yusuke Sekikawa, and Yoshimitsu Aoki. Optical flow estimation by matching time surface with event-based cameras. Sensors, 21(4):1150, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 151, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 204 + ], + "type": "text", + "content": "[28] Garrick Orchard, Ryad Benosman, Ralph Etienne-Cummings, and Nitish V. Thakor. A spiking neural network architecture for visual motion estimation. In IEEE Biomedical Circuits Syst. Conf., pages 298-301. IEEE, 2013. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "type": "text", + "content": "[29] Federico Paredes-Valles and Guido C. H. E. de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. 2, 3, 4, 5, 8, 9" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "text", + "content": "[30] Federico Paredes-Vallés, Kirk Y. W. Scheper, and Guido C. H. E. de Croon. Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE Trans. Pattern Anal. and Mach. Intell., 42(8):2051-2064, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 358 + ], + "type": "text", + "content": "[31] Wachirawit Ponghiran, Chamika Mihiranga Liiyanagedera, and Kaushik Roy. Event-based temporally dense optical flow estimation with sequential neural networks. arXiv:2210.01244, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "type": "text", + "content": "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Int. Conf. Medical Image Computing and Comput.-Assisted Intervention, pages 234–241. Springer, 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 436 + ], + "type": "text", + "content": "[33] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event collapse in contrast maximization frameworks. Sensors, 22(14):5190, 2022. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 437, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 479 + ], + "type": "text", + "content": "[34] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. A fast geometric regularizer to mitigate event collapse in the contrast maximization framework. arXiv:2212.07350, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[35] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In Eur. Conf. Comput. Vis., 2022. 2, 3, 4, 5, 6, 7, 8, 9" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 557 + ], + "type": "text", + "content": "[36] Timo Stoffregen and Lindsay Kleeman. Event cameras, contrast maximization and reward functions: An analysis. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12300-12308, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[37] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In European Conf. Comput. Vis., 2020. 2, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 604, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 646 + ], + "type": "text", + "content": "[38] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. PWC-Net: CNNs for optical flow using pyramid, warping, and cost volume. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8934–8943, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "text", + "content": "[39] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In Eur. Conf. Comput. Vis., pages 402-419. Springer, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[40] Z Wan, Y Dai, and Y Mao. Learning dense and continuous optical flow from an event camera. IEEE Trans. Image Process., 2022. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9704" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 296 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[41] Qimin Wang, Yongjun Zhang, Shaowu Yang, Zhe Liu, Lin Wang, and Luoxi Jing. E-HANet: Event-based hybrid attention network for optical flow estimation. In IEEE Int. Conf. Syst., Man, Cybernetics, pages 2189-2194. IEEE, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "text", + "content": "[42] Yilun Wu, Federico Paredes-Vallés, and Guido C. H. E. de Croon. Rethinking event-based optical flow: Iterative deblurring as an alternative to correlation volumes. arXiv:2211.13726, 2022. 2, 4, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "text", + "content": "[43] Alex Z. Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3D perception. IEEE Robot. and Autom. Lett., 3(3):2032-2039, 2018. 2, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "type": "text", + "content": "[44] Alex Z. Zhu and Liangzhe Yuan. EV-FlowNet: Self-supervised optical flow estimation for event-based cameras. In Robot.: Science and Syst., 2018. 2, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 296 + ], + "type": "text", + "content": "[45] Alex Z. Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 989-997, 2019. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9705" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/1ae449c1-93a5-4652-bb27-6a83c99dd85c_content_list.json b/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/1ae449c1-93a5-4652-bb27-6a83c99dd85c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5bcf75c4e2bbf4c26ef2de985dcbf57165bb81a3 --- /dev/null +++ b/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/1ae449c1-93a5-4652-bb27-6a83c99dd85c_content_list.json @@ -0,0 +1,1674 @@ +[ + { + "type": "text", + "text": "Tangent Model Composition for Ensembling and Continual Fine-tuning", + "text_level": 1, + "bbox": [ + 124, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tian Yu Liu", + "bbox": [ + 261, + 181, + 361, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of California, Los Angeles", + "bbox": [ + 161, + 199, + 460, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tianyu@cs.ucla.edu", + "bbox": [ + 230, + 219, + 392, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stefano Soatto", + "bbox": [ + 598, + 181, + 718, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of California, Los Angeles", + "bbox": [ + 509, + 199, + 807, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "soatto@ucla.edu", + "bbox": [ + 589, + 219, + 725, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tangent Model Composition (TMC) is a method to combine component models independently fine-tuned around a pre-trained point. Component models are tangent vectors to the pre-trained model that can be added, scaled, or subtracted to support incremental learning, assembling, or unlearning. Component models are composed at inference time via scalar combination, reducing the cost of assembling to that of a single model. TMC improves accuracy by $4.2\\%$ compared to assembling non-linearly fine-tuned models at a $2.5 \\times 10 \\times$ reduction of inference cost, growing linearly with the number of component models. Each component model can be forgotten at zero cost, with no residual effect on the resulting inference. When used for continual fine-tuning, TMC is not constrained by sequential bias and can be executed in parallel on federated data. TMC outperforms recently published continual fine-tuning methods almost uniformly on each setting - task-incremental, class-incremental, and data-incremental - on a total of 13 experiments across 3 benchmark datasets, despite not using any replay buffer. TMC is designed for composing models that are local to a pre-trained embedding, but could be extended to more general settings. The code is available at: https://github.com/tianyu139/tangent-model-composition", + "bbox": [ + 75, + 300, + 473, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 693, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compositionality has long been considered central to cognition [5], possibly a reflection of compositionality in neural activity [50]. After decades of unsuccessful attempts to design compositional representations, deep neural networks trained on language data are beginning to exhibit emergent compositionality at scale, by capturing the compositional structure of the data. These models are now being used for visual inference, which has rekindled interest in the study of compositionality of visual representations. But beyond the data, any compositional structure that may be latent in the model is not directly accessible: One cannot simply \"compose\" weights or inner activations and expect", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "meaningful outcomes. The computational architecture of Transformers [52] has been leveraged extensively to co-opt the compositional structure of data through prompts or tokens [29, 55], but still the activations of trained models do not appear to be meaningfully composable. Compositionality of neural activity would allow one to combine activations from different models to capture novel concepts, or incorporate knowledge from different data without having to re-train or fine-tune the core models. This would enable open-universe classification and, more generally, combinatorial expansion of the hypothesis space. Continual learning could be performed simply by composing models trained on different data.", + "bbox": [ + 496, + 268, + 893, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we explore the simplest form of compositionality, that is linear combination. We leverage recent results on the linearization of deep neural networks around a pre-trained point, that can be trained by solving a convex optimization problem and yet perform on-par with nonlinear fine-tuning [1]. This suggests that the tangent space at pre-trained models may be used to linearly compose neural activations, or equivalently compose different models trained or fine-tuned on different datasets and/or for different tasks.", + "bbox": [ + 496, + 470, + 893, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We show that different models obtained from the linearization around a pre-trained point can be composed, combined, rescaled, and forgotten (\"unlearning\"), simply by scalar combinations. This fact can be used to perform ensembling at the inference cost of a single model (Sec. 3). It can also be used for continual fine-tuning, with each component model trained independently and in parallel, if necessary on federated data that can therefore be easily forgotten if needed.", + "bbox": [ + 496, + 625, + 893, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Linear combination is not viable for general concept composition nor arbitrary multi-task learning. In Sec. 5, we discuss limitations of our approach, which can only meaningfully compose models that are \"local\" to a pre-trained embedding. If there are models trained on tasks that are far in representation space, or even antagonistic, they are likely to live on different tangent planes, making linear combination inappropriate. One such example, described in the appendix, concerns models trained on real images (Ima", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18676", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/739c50f870b4986abb37e88572be03dd9e7c1b92a94c1631a7ee77c83ffb2bd9.jpg", + "image_caption": [ + "Figure 1. Composition of Fine-tuned Models: Purple indicates final models used for inference, yellow indicates specialist models fine-tuned on individual tasks. The paragon (a) is a model trained jointly on all tasks. However, in continual learning, different tasks are instantiated at different times, an re-training on their union is impractical. Ensembling (b) combines the output of different models trained separately on each task. Weight averaging (c) yields a \"model soup\" of fine-tuned non-linear models, which improves inference time; Tanget Model Composition (d) linearly composes models fine-tuned on the tangent space of a pre-trained model." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 851, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "geNet) and clip art or sketches (DomainNet). Nonetheless, in more common settings, TMC is competitive with general forms of ensembling such as averaging activations or logits of non-linearly fine-tuned models, and with more general forms of continual learning, including methods that use a replay buffer. For broader task coverage, one could use TMC to construct a tangent bundle of models around different pre-trained points, akin to a \"tangent model zoo.\" This concept is not too dissimilar from the architecture of some Foundation Models with a shared backbone and multiple distinct \"heavy heads\" [54], but far more compact, modular, efficient, interpretable, and easy to work with.", + "bbox": [ + 75, + 481, + 468, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the next section we place our contributions in the context of prior art, and in the following one we describe our method in more detail. Sec. 4 summarizes empirical evidence in support of our approach, and finally Sec. 5 discusses its main limitations.", + "bbox": [ + 75, + 662, + 468, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 753, + 215, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compositionality: Compositionality has been studied for decades as a means to expand the representative power of trained models, but has received increasing impetus in recent years thanks to large Transformer-based models that can be used with adaptable prompts. Prompt-tuning [29] has become commonplace to adapt large pre-trained models for specific downstream tasks. Composition of prompts has also been explored in [56, 55] for continual learning. More", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "general composition of parameters of trained deep networks has been explored for various purposes such as improving optimization and generalization, as we describe next.", + "bbox": [ + 496, + 481, + 890, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep network linearization: Deep networks are linearized using the first-order Taylor expansion around a pretrained weight [1, 18, 35, 49, 32]. [1] finetunes linearized networks modified to use Leaky-ReLU with gradient preconditioning to achieve similar performances to non-linear fine-tuning. We note that TMC is not just a matter of applying linearized networks to different datasets and averaging, since different scales would lead to imbalance as we describe in Sec. 3.3. Furthermore, unlike [1, 49], we do not require Leaky-ReLU nor gradient pre-conditioning.", + "bbox": [ + 496, + 530, + 892, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Averaging model weights: Weight averaging has been used in deep learning to improve generalization [25, 17], improve pre-training [12, 34], perform distributed finetuning [57], and increase robustness against distribution shift [59]. [25] averages points along the trajectory of SGD to find flatter minima, [53] averages \"late-phase weights\" obtained from the later stages of SGD. [24] uses weight interpolation for model patching, and [58] averages weights of models fine-tuned using different hyperparameters from pre-trained models to improve generalization.", + "bbox": [ + 496, + 685, + 890, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While weight averaging techniques for deep networks are related to ensembles [14, 19], in our case the two coincide since we operate in representation spaces that are linear in the parameters. We leverage this property to train on", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18677", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "disjoint subsets of data, and show that they can outperform existing weight averaging and ensembling techniques.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continual learning/fine-tuning: Continual learning [37] aims to adapt models to new tasks and data without degrading performance in previously trained tasks. Continual fine-tuning [49, 6] additionally requires minimal forgetting of a pre-training objective to ensure all tasks can benefit from it. Existing approaches can be broadly classified into parameter isolation, regularization, architecture, and replay-based.", + "bbox": [ + 75, + 121, + 467, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parameter isolation methods [32, 3, 47] operate in the network architecture space by allocating each task a different set of parameters, using techniques such as network pruning [33] to reduce additional memory requirements.", + "bbox": [ + 75, + 242, + 467, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regularization-based methods [30, 2, 26, 36, 28] incorporate additional terms into the training objective to prevent catastrophic forgetting. [26] penalizes the distance between weights of the previous and current task, and [45] penalizes changes in weights using Hessian approximations.", + "bbox": [ + 75, + 301, + 467, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Experience replay methods [44, 4, 32, 11, 43, 49, 7, 8, 26, 56, 6, 10, 9] assume the availability of a memory buffer to store samples from previous tasks. [49, 6] applies experience replay to continual fine-tuning. [49] shows that linearized models combined with replay and regularization alleviates catastrophic forgetting when fine-tuned, and [6] combines dark experience replay and an attention-based regularization loss to prevent catastrophic forgetting of the pre-training features, by encouraging similarity of features at each layer to the original pre-trained network. While replay methods have achieved state-of-the-art performance compared to non-replay ones, they require a sufficiently large memory buffer to store previously seen exemplars.", + "bbox": [ + 75, + 378, + 467, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A further line of work tackles exemplar-free class incremental learning (EFCIL) using regularization [60], distillation [22], and class-prototypes [61, 62, 40] to incrementally integrate new classes into a fixed-sized network without the memory costs of a replay buffer.", + "bbox": [ + 75, + 575, + 467, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We show that composition of tangent models trained on individual tasks is a viable method for replay-free continual fine-tuning, and can even outperform recently published replay-based methods. As a side benefit, our method also yields fully parallel training across tasks, and enables zero-cost unlearning.", + "bbox": [ + 75, + 650, + 467, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 753, + 166, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In continual learning, models are trained on a sequence of tasks indexed by $t = 1,\\dots ,T$ ,with the goal of performing well on all, under three main scenarios [51]: (1) Task-incremental, where the task identity $t$ (Task-ID) is known during inference, (2) Data/Domain-incremental, where Task-ID is not provided during inference, and (3) the most challenging Class-incremental, where Task-ID is unknown during inference and in addition the classes repre", + "bbox": [ + 75, + 780, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/1b505f6886c234254cd9cf4300e71010586f5f15e4916c3f10c7b1c0fd200a3a.jpg", + "table_caption": [ + "Table 1. A Comparison of Composition Methods: On average across 25 experiments, Tangent Model Ensembling (TME) improves accuracy by $5.1\\%$ over standard ensembling of non-linear fine-tuned models at a $2\\times$ increase of inference cost. Tangent Model Composition (TMC) also improved accuracy, by $4.2\\%$ , while reducing inference cost by $2.5\\times$ to $10\\times$ in our experiments, and growing linearly with the number of models in the ensemble. Best results for single and multi model inference are indicated in bold and italic respectively. Joint training paragon accuracy (Original Model/Tangent Model): Caltech-256 - $89.17\\% /86.77\\%$ , MIT-67 - $77.29\\% /77.74\\%$ , OxfordPets - $92.68\\% /93.68\\%$ ." + ], + "table_footnote": [], + "table_body": "
DatasetTasksSoupEns-LEns-SMTMCTME
Inference/MemoryCost:O(1)O(T)O(T)O(1)O(T)
C-Caltech-256 (Class-IL)584.0082.8083.0984.8285.53
1080.1076.8579.7682.3782.58
2074.8666.7675.4778.6679.30
C-MIT-67 (Class-IL)559.4856.7958.6869.4371.72
1049.3042.3650.5264.5365.72
2028.5123.6135.6248.9854.93
C-OxfordPnts (Class-IL)580.7181.1481.8684.4285.04
1071.3569.7177.3177.7579.97
D-Caltech-256 (Data-IL)585.9686.2786.5984.8685.13
1083.4884.0584.8083.6484.16
2079.6480.7283.1781.9982.94
D-MIT-67 (Data-IL)566.0066.5467.2174.8875.45
1058.8859.2060.4772.4173.41
2050.1051.9955.2070.2272.11
D-OxfordPnts (Data-IL)590.2290.8391.0293.1793.21
1087.6688.4288.8892.4892.72
2085.5785.9786.5891.2891.77
C-Caltech-256 (Task-IL)594.0194.7294.8494.3394.76
1095.0095.8796.0496.0396.23
2095.5696.5396.8497.2297.38
C-MIT-67 (Task-IL)583.3686.2486.3791.2791.79
1087.1990.4290.9794.4595.03
2088.4193.7393.8897.7198.09
C-OxfordPnts (Task-IL)597.2797.2797.3598.5798.74
1097.5997.9297.9899.0599.13
", + "bbox": [ + 503, + 239, + 888, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sented in each task are disjoint.", + "bbox": [ + 500, + 672, + 707, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We assume the setting of continual fine-tuning of a pretrained model. Naive fine-tuning causes catastrophic forgetting of prior tasks, including the one on which pre-training occurred, as the number of tasks increases [6]. Thus, only early tasks benefit from pre-training. Therefore, it is important for continual fine-tuning to ensure that all tasks can benefit from information acquired during pre-training.", + "bbox": [ + 496, + 696, + 890, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a unified method to tackle the main scenarios (1)-(3) by composing separately-trained tangent component models. We show that each task takes full advantage of pre-training, and catastrophic forgetting of prior tasks is averted since information from each previous task is encoded separately and only fused for inference.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18678", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Linearization", + "text_level": 1, + "bbox": [ + 76, + 90, + 272, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The effectiveness of model aggregation in weight space hinges on whether the models lie in within the same small-loss training basin of convergence. It is known that weights of different models fine-tuned on the same dataset can be connected by linear paths with constant loss (\"linear mode connectivity\") [16], but this does not generally occur for models fine-tuned on disjoint tasks/datasets, which is the case of interest to us. Therefore, we convexify the loss landscape locally by approximating the function represented by a deep (non-linear) network with its first-order Taylor approximation around the pre-trained weights. For any deep network $f$ and initial weights $w \\in \\mathbb{R}^m$ , we denote with $\\mathcal{H}(w)$ the set of all such models tangent to $f_w$ given by:", + "bbox": [ + 75, + 113, + 473, + 311 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} (w) = \\left\\{h _ {\\delta} (\\cdot) \\triangleq f _ {w} (\\cdot) + \\nabla_ {w} f _ {w} (\\cdot) \\cdot \\delta \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 335, + 468, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the tangent model $h_{\\delta}$ is linear in the parameters $\\delta$ , training it with standard losses such as mean-squared error (MSE) or empirical cross-entropy yields a convex loss landscape. In order to compose models, leveraging Jensen's inequality, we are guaranteed that for any dataset $\\mathcal{D}$ :", + "bbox": [ + 75, + 364, + 470, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL \\left(\\mathcal {D}, \\sum_ {i} ^ {T} h _ {\\theta_ {i} \\delta_ {i}}\\right) \\leq \\sum_ {i} \\theta_ {i} L \\left(\\mathcal {D}, h _ {\\delta_ {i}}\\right), \\quad \\sum_ {i} \\theta_ {i} = 1. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 452, + 468, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In other words, the loss resulting from a convex combination of tangent models in weight space will always be lower than or equal to the convex combination of their individual losses. This is particularly useful when combining models trained to some constant (possibly zero) loss, since it ensures that the composed model will have loss lower or equal to the loss of the component models.", + "bbox": [ + 75, + 503, + 468, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While tangent models might seem computationally expensive to train, we note that the Jacobian vector product $\\nabla_w f_w(\\cdot) \\cdot \\delta$ can be computed in a single forward pass as shown by [39, 35, 1]. This reduces the computational cost to that of a forward pass through the original network. Also note that while the tangent model has twice the degrees of freedom of the original model, fixing the pre-trained model at $w$ results in only a convex optimization in the remaining parameters, as many as the original model.", + "bbox": [ + 75, + 609, + 470, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the context of fine-tuning, $w$ is initialized by pretraining on large datasets such as ImageNet [13]. In the rest of this paper, we use the short-form notation $\\mathcal{H}$ and $h$ for $\\mathcal{H}(w)$ and $h_{\\delta}$ respectively.", + "bbox": [ + 75, + 744, + 468, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Tangent Model Composition for Replay-free Continual Fine-tuning", + "text_level": 1, + "bbox": [ + 76, + 814, + 468, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ensembling is an effective strategy for combining weak learners into a strong model [14]. For any finite $F \\subseteq \\mathbb{A}$ , where $\\mathbb{A}$ is a set of functions, and $\\lambda = \\{\\lambda_i\\}_{i=1}^{|F|}$ a set of", + "bbox": [ + 76, + 852, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ac03a9f1f8388b819c068a7ec9d8cb4a44135cd55a2075a1661e7b9e9c82c359.jpg", + "image_caption": [ + "Figure 2. Accuracy plot over models trained on C-Caltech-256, C-MIT-67, and C-Oxford Pets, split into 10 tasks each. Soft-max ensembling (TME) consistently yields the best results compared to logit ensembling (TMC) for each loss function. Composition of component models trained with the Rescaled Square Loss (RSL) outperforms those trained with empirical cross-entropy (CE) or MSE loss, trading off only $1.2\\%$ accuracy against TME to reduce inference time and memory requirement from $\\mathcal{O}(T)$ to $\\mathcal{O}(1)$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 879, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "weights, the logit ensemble operation is given by", + "bbox": [ + 500, + 435, + 823, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E n s e m b l e} (F, \\lambda) = \\sum_ {i = 1} ^ {| F |} \\lambda_ {i} f _ {i} (\\cdot) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 460, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the context of continual learning, a simple yet effective approach is to simply train a new model on each task $t$ , and ensemble the outputs at inference time. Unfortunately, this has the obvious shortcoming that both storage requirements and inference time grows along with the number of tasks.", + "bbox": [ + 496, + 513, + 890, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, consider the case where $\\mathbb{A} = \\mathcal{H}$ . Then for any finite $H \\subseteq \\mathcal{H}$ and associated weights $\\lambda$ , we can write", + "bbox": [ + 498, + 589, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {E n s e m b l e} (H, \\lambda) = \\sum_ {i = 1} ^ {| H |} \\lambda_ {i} h _ {i} (\\cdot) (4) \\\\ = f _ {w} (\\cdot) + \\sum_ {i = 1} ^ {| H |} \\nabla_ {w} f _ {w} (\\cdot) \\cdot \\lambda_ {i} \\delta_ {i} (5) \\\\ = h _ {\\sum \\lambda_ {i} \\delta_ {i}} (\\cdot) (6) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 628, + 890, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In other words, since $\\mathcal{H}$ defines a vector space with addition and scalar multiplication based on $\\delta$ , any linear combination (or ensemble) of models in $\\mathcal{H}$ is equivalent to a single model in $\\mathcal{H}$ derived simply by the linear combination over $\\delta$ . This reduces the inference time for any ensemble of $T$ tangent models from $\\mathcal{O}(T)$ to $\\mathcal{O}(1)$ .", + "bbox": [ + 496, + 750, + 893, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In continual learning, it is also important for storage space to not scale with $T$ , since the number of tasks can be arbitrarily large. However, since tasks arrive sequentially in the continual learning framework, we can further reduce", + "bbox": [ + 496, + 840, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18679", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d9b6c1e8863252bfbee63f7cd73b3ffb62a7eaa9b76e6aa5b11778def2a36790.jpg", + "table_caption": [ + "Table 2. Class-Incremental Learning: Comparison against existing methods using an ImageNet pre-trained ResNet-50. TMC is almost uniformly better than all replay-based and replay-free methods tested, outperforming the best replay-free (RF) method by $14.48\\%$ and even the best replay-based method [6] by $1.53\\%$ ." + ], + "table_footnote": [], + "table_body": "
Method / BufferTasksC-Caltech-256C-MIT-67
5102051020
ER [46] / 30070.14±0.2265.75±0.9265.490.6054.61±0.4149.57±1.4645.331.73
LWF [30] / None34.55±0.1822.76±0.1714.37±0.1425.14±0.5713.30±0.496.41±0.17
oEWC [48] / None28.71±1.3116.32±4.3715.59±3.8314.18±0.777.12±0.494.33±0.16
DER++ [7] / 30080.50±0.2677.84±1.0274.46±0.2164.56±0.5559.73 ±1.0250.28±1.27
Co2L [10] / 30020.71±2.8917.13±2.5314.620.6119.64±12.9321.21±0.5017.78±1.33
ER-ACE [9] / 30078.47±1.2573.85±0.5370.87±1.3864.94±0.5060.25±0.2652.48±0.29
TWF (RF) [6] / None77.95±0.5270.67±0.7262.79±0.8256.58±0.6245.03 ±0.7828.91±0.92
TWF [6] / 30081.34±0.9177.03±0.5372.44±0.4268.33±0.4664.21 ±0.6556.29±0.68
TMC / None84.82 ±0.0382.37 ±0.0878.66 ±0.1369.43±0.2164.53±0.2848.98 ±0.13
", + "bbox": [ + 117, + 128, + 844, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the memory capacity required for storage of models from $\\mathcal{O}(T)$ , to $\\mathcal{O}(1)$ by using a simple autoregressive model: We train on the first task $t_1$ to produce model $\\tilde{h}_1 = h_1$ . Then, for any subsequent task $t > 1$ , we train a new model $\\tilde{h}_t$ , and compose it with $h_{t-1}$ to give $h_t = \\lambda_{(t,1)}\\tilde{h}_t + \\lambda_{(t,2)}h_{t-1} := h_{\\delta'}$ where $\\delta' = \\frac{\\lambda_{(t,1)}\\tilde{\\delta}_t + \\lambda_{(t,2)}\\delta_{t-1}}{\\lambda_{(t,1)} + \\lambda_{(t,2)}}$ . Since we only require model $h_{t-1}$ to produce $h_t$ , all prior models up to and including $h_{t-1}$ can be discarded upon constructing $h_t$ , hence reducing memory requirement from $\\mathcal{O}(T)$ to $\\mathcal{O}(1)$ , equal to the size of a single model. In contrast to methods that expands the model at each task, or require a memory buffer of examples from previous tasks, this method does not incur additional storage requirements, nor assume availability of an extra replay buffer for storing past exemplars (i.e. our method is replay-free).", + "bbox": [ + 75, + 330, + 472, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 1 (Free Parallelization): Additionally, for any given convex loss function $L$ , the optimization objective", + "bbox": [ + 75, + 566, + 470, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\delta \\in \\mathbb {R} ^ {m}} \\sum_ {(x, y) \\in \\mathcal {D}} L \\left(h _ {\\delta} (x), y\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 603, + 468, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is convex and thus converges to a global optimum. This implies that each model $\\tilde{h}_t$ can be trained from a standard initialization as opposed to from $h_{t-1}$ . Hence, if multiple tasks arrive at the same time, such as when training in a federated manner, training can be parallelized across tasks and composed in a zero-cost, zero-shot manner to yield equivalent results (assuming equivalence of global optima) to sequential training and composition. To the best of our knowledge, apart from [41] which simply stores samples in a replay buffer and trains on them at test time, no other work in continual learning can be parallelized across tasks.", + "bbox": [ + 75, + 643, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 2 (Free Forgetting/Unlearning): If we relax memory constraints to store models from each task, unlearning any specific task $t$ (e.g., private data associated with specific samples) is straightforward and can be done in a zero-cost, zero-shot manner. Since subtraction is well-defined in the vector space $\\mathcal{H}$ , any task $i$ can be unlearned", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "from the model $h_T$ trained on $T$ tasks by simply subtracting the weights associated to $\\tilde{h}_i$ :", + "bbox": [ + 498, + 330, + 890, + 361 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {T \\backslash i} = h _ {T} - \\lambda_ {i, 1} \\prod_ {j = i + 1} ^ {T} \\lambda_ {(j, 2)} \\tilde {h} _ {i} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 372, + 890, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 3 (Unified Continual Fine-tuning): Our method does not assume any specific task splits, and is hence generalizable across all three forms of continual fine-tuning: task-incremental, class-incremental, and data-incremental.", + "bbox": [ + 498, + 424, + 890, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Scale Standardization", + "text_level": 1, + "bbox": [ + 500, + 494, + 694, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training with the cross-entropy loss does not ensure that the weights are comparable in scale. We can study the implications of this for composition on $\\mathcal{H}$ through the lens of logits ensembling. While logits ensembling has been shown to be effective [58], Fig. 2 shows that ensembling the normalized soft-max outputs (Tangent Model Ensembling, TME) is uniformly better than logit ensembling/weight averaging via TMC across various loss functions. However, TME cannot be represented by a linear combination of perturbations $\\delta_{i}$ and hence incurs $\\mathcal{O}(T)$ model storage costs and $\\mathcal{O}(T)$ inference time.", + "bbox": [ + 496, + 518, + 890, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead, we propose to induce a waterbed effect similar to soft-max ensembling using standardization. In particular, we train using the Rescaled Square Loss (RSL) [23]:", + "bbox": [ + 498, + 684, + 890, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s} (x, y) = \\frac {1}{K} \\left(\\alpha ([ f (x) ] _ {y} - \\beta) ^ {2} + \\sum_ {i = 1, i \\neq y} ^ {K} ([ f (x) ] _ {i}) ^ {2}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 739, + 890, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha, \\beta$ are hyper-parameters. This is the regular MSE loss when $\\alpha = \\beta = 1$ . Larger values of $\\beta$ reduce soft-max entropy in the output by scaling positive output class signals while keeping negative ones close to zero.", + "bbox": [ + 496, + 809, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use $\\beta$ to control the interaction between component models. For example, when tasks are highly dissimilar,", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18680", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7f2e3e13cdf73b2049efd0360c75b1a419ae252b5d7a0c3e226821c6f3d60911.jpg", + "table_caption": [ + "Table 3. Data-Incremental Learning on D-MIT-67, 4 tasks using ResNet-18. For fair comparison, we run our method under the settings proposed by [49]. TMC outperforms the best replay-free method [30] by $4.61\\%$ , and the best replay-based method [49] by $1.27\\%$ , while supporting parallel training across tasks (PL). If we sacrifice parallelization by initializing each model at task $t$ with $h_{t-1}$ (TMC-Seq), we further improve over TMC by $0.28\\%$ ." + ], + "table_footnote": [], + "table_body": "
MethodPL?Replay BufferAccuracy
SGDXNone63.48
LwF[30]XNone67.21
MAS[2]XNone62.49
OSLA[45]XNone64.08
EWC [26]X50063.68
DLCFT[49]X50070.55
TMC✓None71.82
TMC-SeqXNone72.10
Joint (Linear)--73.61
Joint (Paragon)--74.40
", + "bbox": [ + 86, + 184, + 450, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "such as in task- and class-incremental continual learning, we use larger values of $\\beta$ to reduce interference of component models trained on other tasks upon composition. On the other hand, when tasks are highly synergistic such as in data-incremental learning, lower values of $\\beta$ can be used to encourage interaction in the final composed model.", + "bbox": [ + 75, + 395, + 467, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 513, + 207, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We describe comparison baselines in Sec. 4.1, and our main results in Sec. 4.2, with additional studies in Sec. 4.3.", + "bbox": [ + 76, + 544, + 467, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method is replay-free. As such, its computational cost is at most half of what it would be if it used a replay buffer since replay-based methods typically sample separate batches of equal size from both the current task and replay buffer at each iteration [7, 6, 10, 46, 9]. Nonetheless, we compare our method to both replay-based and replay-free methods in our baselines. While replay-based methods have thus far outperformed replay-free ones, ours is uniformly more accurate than replay-based methods tested, in addition to replay-free ones.", + "bbox": [ + 75, + 579, + 467, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on Caltech-256 [20], MIT-67 [42], and OxfordPets [38]. For continual fine-tuning, we further split each training dataset into multiple disjoint subsets, each representing a single task, for fair comparison across experiments. We prefix sequential datasets split into tasks with disjoint labels (for task- and class-incremental learning) with C-, and random splits using a standardized random seed (for data-incremental learning) with D-. We set $\\alpha = 1, \\beta = 25$ for the former, $\\alpha = 1, \\beta = 5$ for the latter, and $\\alpha = 1, \\beta = 5$ for TME. Results in all tables and figures are averaged across 3 runs.", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0ef32265e31f3fe4ed170d117da0232be78600cf938bfe4440a5fca75766d9e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 92, + 834, + 256 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/273a8be4dd0983e861b7db82430be3541c7530047c70227af59d60cdc76eafcd.jpg", + "image_caption": [ + "Figure 3. Ablation on $\\beta$ averaged across (Top:) C-Caltech-256, C-MIT-67, and C-OxfordPets and (Bottom:) D-Caltech-256, D-MIT-67, and D-OxfordPets, split into 10 disjoint tasks. For class incremental datasets, larger values of $\\beta$ result in better generalization of the final composed model. For data-incremental datasets where tasks are synergistic, larger values of $\\beta$ instead harm the accuracy of both the component models and the composed model." + ], + "image_footnote": [], + "bbox": [ + 555, + 261, + 834, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Composition Baselines", + "text_level": 1, + "bbox": [ + 500, + 556, + 707, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In each of the following sections, we compare against the following baselines for model aggregation:", + "bbox": [ + 498, + 585, + 890, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Logits Ensemble (Ens-L): Averaging the output logits of non-linear component models.", + "bbox": [ + 498, + 622, + 890, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Softmax Ensemble (Ens-SM): Averaging the softmax output of non-linear component models.", + "bbox": [ + 498, + 657, + 888, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Weight composition of non-linear models (Soup): Averaging the weights of non-linear component models. This method is inspired by Model Soups [58] where weights of various non-linear models trained on the same task but with different hyper-parameters and augmentations are averaged to improve generalization at no additional inference cost.", + "bbox": [ + 496, + 694, + 890, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 500, + 811, + 637, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show that TMC outperforms existing methods in all three continual fine-tuning settings [51], from hardest to easiest: Class-Incremental (Sec. 4.2.1), Data-Incremental (Sec. 4.2.2), and Task-Incremental (Sec. 4.2.3).", + "bbox": [ + 498, + 839, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18681", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6a8a639407586607914e585f6184be3e9ea4bf07046c7e51d2102a12c719a783.jpg", + "table_caption": [ + "Table 4. Task-Incremental Learning: Comparison against existing methods using ImageNet pre-trained ResNet-50. We almost uniformly outperform all replay-based and replay-free methods that we compare against. In particular, we improve over the best replay-free (RF) method by $2.46\\%$ , and the best replay-based method by $1.38\\%$ [6]." + ], + "table_footnote": [], + "table_body": "
Method / BufferTasksC-Caltech-256C-MIT-67
5102051020
ER [46] / 30092.11±0.2694.33±0.4895.76±0.1385.84±0.0990.59±0.5195.13±0.27
LWF [30] / None93.73±0.1295.31±0.0896.29±0.0587.55±0.2090.82±0.0495.68±0.58
oEWC [48] / None83.97±0.7176.20±16.0791.49±0.3067.31±3.2268.45±4.9384.74±1.96
DER++ [7] / 30093.79±0.1394.92±0.1195.87±0.1786.83±0.2691.63±0.7594.54±0.62
Co2L [10] / 30034.51±3.0539.62±1.7045.94±1.4739.32±20.4764.40±1.3151.50±2.21
ER-ACE [9] / 30093.11±0.1294.63±0.0695.86±0.1087.13±0.1690.96±0.2894.92±0.60
TWF (RF) [6] / None94.45±0.0595.64±0.0696.36±0.1286.76±0.2588.27±0.5494.74±0.35
TWF [6] / 30094.38±0.0895.74±0.1796.59±0.0788.22±0.4192.13±0.2695.67±0.46
TMC / None94.33±0.0396.03±0.0297.22±0.0391.27±0.2194.45±0.1397.71±0.13
", + "bbox": [ + 119, + 128, + 846, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1 Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 76, + 330, + 328, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In class-incremental learning, datasets are partitioned so that label spaces of each task are disjoint.", + "bbox": [ + 75, + 367, + 470, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison against Composition Baselines: We compare against model composition baselines in Table 1. While TME yields the best generalization performance, this comes at a cost of $\\mathcal{O}(T)$ inference time. On the other hand, TMC achieves an $\\mathcal{O}(1)$ inference time with only a $1.73\\%$ trade-off in generalization accuracy on average, while outperforming weight averaging, logit ensembling, and even soft-max ensembling of non-linear models by an average of $7.83\\%$ , $11.37\\%$ , and $6.08\\%$ respectively, even though the latter two both require $\\mathcal{O}(T)$ inference time.", + "bbox": [ + 75, + 402, + 468, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison against existing methods: In Table 2, we compare against recent continual fine-tuning methods including both replay-free and replay-based methods. Here, replay-free methods [30, 48] perform poorly in the class-incremental setting relative to stronger replay-based methods [46, 9, 7, 6]. However, we show that not only do we improve over the best replay-free method by $14.48\\%$ , but also over the best replay-based method by $1.53\\%$ [6], corresponding to a $17.59\\%$ relative error reduction towards the paragon performance of joint training (Table 1).", + "bbox": [ + 75, + 561, + 468, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison against EFCIL methods: In Table 5, we compare against EFCIL methods on C-TinyImageNet [27]. The original EFCIL methods are evaluated under a different setting (denoted $\\dagger$ ), where the zeroth task contains $50\\%$ of the dataset. Hence, we further modified two of the best performing EFCIL methods, SSRE [62] and FeTrIL [40] to use the pre-trained ImageNet initialization, and run them under our setting of uniformly sharded data. TMC not only outperforms all methods ran under our harder setting, by an average of $15.3\\%$ over the next best method [40], but even yields equal or better performance when compared to all methods ran under $\\dagger$ .", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Data-Incremental Learning", + "text_level": 1, + "bbox": [ + 500, + 330, + 746, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Also referred to as domain-incremental learning, this setting assumes that tasks are split in a non-stratified manner, where each task has the same output space [51, 49].", + "bbox": [ + 496, + 354, + 890, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison against Composition Baselines: In Table 1, we compare against model composition baselines. Similar to the class-incremental setting, TME outperforms TMC by an average of $0.66\\%$ at a cost of $\\mathcal{O}(T)$ inference time. TMC uses $\\mathcal{O}(1)$ inference time and outperforms non-linear composition, logit ensemble, and even soft-max ensemble by an average of $7.04\\%$ , $6.32\\%$ , and $5.21\\%$ respectively.", + "bbox": [ + 496, + 401, + 892, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison against existing methods: In Table 3, we benchmark our model against results obtained by [49] on a pre-trained ResNet-18 model using the D-MIT-67 dataset split into 4 disjoint subsets. We improve over the best replay-free method [30] by $4.61\\%$ , and the best replay-based method [49] by $1.27\\%$ , corresponding to a $64\\%$ and $33\\%$ relative error reduction towards the joint training paragon respectively. TMC-Seq, which uses sequential initialization by initializing each component model $h_t$ with $h_{t - 1}$ instead of $h_0$ , further improves over TMC by $0.28\\%$ but at the cost of parallelizability.", + "bbox": [ + 496, + 507, + 893, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3 Task-Incremental Learning", + "text_level": 1, + "bbox": [ + 498, + 694, + 746, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In task-incremental learning, the output spaces of different tasks are disjoint, but the source task of each sample is known at inference time. Thus, we can restrict the model predictions to that of the given task during evaluation.", + "bbox": [ + 496, + 719, + 890, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison against Composition Baselines: Due to the disjoint output spaces of each task, ensembling outputs of specialist component models trained on individual tasks and restricting them to a target task essentially only considers the predictions of the specialist model trained on the target task (since, in general, specialist models are not effective discriminants for tasks other than those they are triend on). Hence, the independence of output spaces renders en", + "bbox": [ + 496, + 780, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18682", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8a74b4472ac9f08ce4efd40f4869e164c1548dabff5031dc235d697de0758e7b.jpg", + "table_caption": [ + "Table 5. Comparison against EFCIL methods using either ImageNet pre-trained ResNet-50, or under the original EFCIL setting where the zeroth task contains $50\\%$ of the dataset (denoted $\\dagger$ ). All experiments are ran on C-TinyImageNet. Italics denote the best method under $\\dagger$ , and bold denotes the best method under our harder setting. Under the former, TMC performs equally or better than all methods. Under the latter, TMC outperforms the best method by an average of $15.3\\%$ ." + ], + "table_footnote": [], + "table_body": "
TasksPASS† [61]IL2A† [60]SSRE† [62]FeTrIL† [40]SSRE[62]FeTrIL[40]TMC
549.647.350.454.833.744.157.6
1047.344.748.953.125.636.153.1
", + "bbox": [ + 138, + 142, + 823, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "semblies of specialist models highly effective. In contrast, composing non-linear models in weight space does not ensure minimal interference across different tasks. Indeed, we show in Table 1 that TME yields the best results, while the non-linear composition approach yields the worst in task-incremental learning. On the other hand, due to the equivalence of TMC to (logit) ensembling of linearized models, TMC outperforms not only weight averaging, but also logit and soft-max ensembles of non-linear models, by $3.78\\%$ , $1.83\\%$ , and $1.79\\%$ respectively.", + "bbox": [ + 75, + 218, + 470, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison against existing methods: Since each component model is trained only on task $t$ , by design, there is no interference from other tasks when training $\\tilde{h}_t$ . As such, the final composed model is equivalent to a specialist model when restricted to any task. However, in other state of the art continual learning methods such as those involving replay or loss regularization, this separation among trained for each task is violated. Hence, under the task-incremental setting, we show in Table 4 that TMC achieves the best performance across both replay-free and replay-based methods even without using any memory buffer, beating the next best method by $2.46\\%$ and $1.38\\%$ respectively [6].", + "bbox": [ + 75, + 369, + 468, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Further studies", + "text_level": 1, + "bbox": [ + 76, + 561, + 232, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present additional studies on the Rescaled Square Loss (Sec. 4.3.1), and compare against TMC-Seq to show that our method yields an effective and highly parallelizable training scheme for continual fine-tuning (Sec. 4.3.2).", + "bbox": [ + 75, + 584, + 468, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.1 Rescaled Square Loss", + "text_level": 1, + "bbox": [ + 76, + 665, + 282, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Sec. 3.3, we showed that $\\beta$ can be used to control the amount of interaction between tasks. In Fig. 3, we empirically show that in the class-incremental setting where tasks are highly dissimilar, larger values of $\\beta$ result in a better generalist model. On the other hand, under the data-incremental setting where tasks are more similar, information from other tasks can benefit the learning of any given task. Here, larger values of $\\beta$ instead harm generalization due to lowering the accuracy of each component model.", + "bbox": [ + 75, + 689, + 468, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2 Convexity and Parallelizability", + "text_level": 1, + "bbox": [ + 76, + 845, + 344, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given a sequence of tasks $1 \\ldots T$ and a pre-trained model $h_0$ , there are at least two natural ways to learn the compo", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f30f3019419ce173f201bee14ae28f270789489cb3996979412221091f617ea3.jpg", + "image_caption": [ + "Figure 4. Accuracy of composed models, where component models are initialized with $h_0$ (TMC) and $h_{t-1}$ (TMC-Seq). In the data-incremental setting (D), TMC-Seq outperforms TMC, while in the class-incremental setting (C), TMC performs significantly better. All results are averaged across D/C-Caltech-256, D/C-MIT-67, D/C-OxfordPets (we exclude C-OxfordPets in the 20-task setting since it does not contain enough classes)." + ], + "image_footnote": [], + "bbox": [ + 535, + 219, + 857, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "nent model $\\tilde{h}_t$ : (1) Train $\\tilde{h}_t$ with $h_{t-1}$ as initialization, and (2) Train $\\tilde{h}_t$ with $h_0$ as initialization. The first option (TMC-Seq) is most common among continual learning methods which impose regularization in the weights between the current and previous tasks or optimize using memory buffers consisting of examples obtained from previous tasks. As such, these methods are restricted to a sequential learning framework, and hence cannot be parallelized across tasks. In Sec. 3, we showed that, thanks to the convexity of the loss landscape, both options (1) and (2) are equivalent under our framework when we assume comparable generalization performance among global optima.", + "bbox": [ + 496, + 521, + 890, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In practice, for massively over-parameterized networks, such assumptions do not hold. However, we show empirically in Fig. 4 that under the data-incremental setting, compared to TMC-Seq, TMC only incurs a small loss in generalization performance to enable massive parallelization across tasks. This can potentially provide orders of magnitude speed-ups compared to existing continual learning methods. Moreover, under the class-incremental setting, TMC greatly outperforms TMC-Seq, showing that the pre-trained initialization yields solutions of better generalization than $h_{t - 1}$ . We attribute this to the fact that the hypothesis space of models which are globally optimal for the training data is vast and yields varying generalization.", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18683", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Thus, regardless of the loss landscape convexity, convergence to each point in this space depends on the \"direction of approach\". Hence, initialization at $h_{t-1}$ introduces a bias towards the previous tasks. Due to the high level of dissimilarity between tasks in the class-incremental setting, this bias harms the final model generalization performance.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 196, + 191, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our approach enables incremental learning under the assumption that increment tasks are \"close\" to the pre-trained model. We ensure proximity by training increment models independently on the tangent plane of the pre-trained model. The advantage is that pre-trained tasks are not forgotten, and increment tasks can be learned independently, in parallel, and easily forgotten if needed.", + "bbox": [ + 75, + 223, + 468, + 328 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Of course, this approach does not address the problem of incremental learning in full generality, when increment tasks can be arbitrary and arbitrarily unrelated. Nonetheless, we have shown that our method is competitive with existing methods in some of the most challenging settings, for instance when the hypothesis spaces of component tasks are disjoint. On the positive side, despite not requiring a replay buffer, our method outperforms replay-based ones under the assumptions in which we operate.", + "bbox": [ + 75, + 329, + 470, + 465 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A limitation of our approach is that it requires computing the linear span of pre-trained models, which can be challenging. However, for common deep network architectures such as the ResNet [21] or Transformer [52, 15] family, the Jacobian-Vector product can be efficiently evaluated [39, 35, 1, 31], reducing the computational complexity of training the tangent model to that of the same magnitude as a linear classifier on network activations. Hence, inference time for the linearized network is at most double that of the original model.", + "bbox": [ + 75, + 465, + 470, + 617 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Another limitation of our approach is assuming a first-order expansion around a pre-trained objective can well approximate the fine-tuned model. Such assumptions hold poorly when the pre-training and downstream tasks are highly unrelated. Nevertheless, an ImageNet [13] pretrained model already yields a sufficiently good local approximation for various real-world datasets [1, 31].", + "bbox": [ + 75, + 617, + 468, + 722 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While we mainly explored linear compositionality in the context of continual learning, we note that this framework can be easily generalized. We discussed possible applications towards federated learning and forgetting, since our method naturally yields a parallel training framework that fully compartmentalizes information obtained each task.", + "bbox": [ + 75, + 723, + 470, + 814 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 829, + 264, + 847 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We would like to thank Aditya Golatkar, Albert Zhao, and the anonymous reviewers for their feedback on the initial version of the paper. This work was supported by ARO", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "W911NF-17-1-0304 and ONR N00014-22-1-2252.", + "bbox": [ + 500, + 90, + 839, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 118, + 596, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alessandro Achille, Aditya Golatkar, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. Lqf: Linear quadratic fine-tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15729-15739, 2021. 1, 2, 4, 9", + "[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 3, 6", + "[3] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3366-3375, 2017. 3", + "[4] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Advances in neural information processing systems, 32, 2019. 3", + "[5] Elie Bienenstock and Stuart Geman. Compositionality in neural systems. In The handbook of brain theory and neural networks, pages 223-226. 1998. 1", + "[6] Matteo Boschini, Lorenzo Bonicelli, Angelo Porrello, Giovanni Bellitto, Matteo Pennisi, Simone Palazzo, Concetto Spampinato, and Simone Calderara. Transfer without forgetting. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIII, pages 692-709. Springer, 2022. 3, 5, 6, 7, 8", + "[7] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 3, 5, 6, 7", + "[8] Pietro Buzzega, Matteo Boschini, Angelo Porrello, and Simone Calderara. Rethinking experience replay: a bag of tricks for continual learning. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 2180-2187. IEEE, 2021. 3", + "[9] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuyte-laars, Joelle Pineau, and Eugene Belilovsky. New insights on reducing abrupt representation change in online continual learning. arXiv preprint arXiv:2203.03798, 2022. 3, 5, 6, 7", + "[10] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. In Proceedings of the IEEE/CVF International conference on computer vision, pages 9516-9525, 2021. 3, 5, 6, 7", + "[11] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 3", + "[12] Leshem Choshen, Elad Venezian, Noam Slonim, and Yoav Katz. Fusing finetuned models for better pretraining. arXiv preprint arXiv:2204.03044, 2022. 2", + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image" + ], + "bbox": [ + 501, + 143, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18684", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 9", + "[14] Thomas G Dietterich. Ensemble methods in machine learning. In International workshop on multiple classifier systems, pages 1-15. Springer, 2000. 2, 4", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 9", + "[16] Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lottery ticket hypothesis. In International Conference on Machine Learning, pages 3259-3269. PMLR, 2020. 4", + "[17] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. Advances in neural information processing systems, 31, 2018. 2", + "[18] Aditya Golatkar, Alessandro Achille, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. Mixed-privacy forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 792–801, 2021. 2", + "[19] Raphael Gontijo-Lopes, Yann Dauphin, and Ekin D Cubuk. No one representation to rule them all: Overlapping features of training methods. arXiv preprint arXiv:2110.12899, 2021. 2", + "[20] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 6", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 9", + "[22] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831-839, 2019. 3", + "[23] Like Hui and Mikhail Belkin. Evaluation of neural architectures trained with square loss vs cross-entropy in classification tasks. arXiv preprint arXiv:2006.07322, 2020. 5", + "[24] Gabriel Ilharco, Mitchell Wortsman, Samir Yitzhak Gadre, Shuran Song, Hannaneh Hajishirzi, Simon Kornblith, Ali Farhadi, and Ludwig Schmidt. Patching open-vocabulary models by interpolating weights. arXiv preprint arXiv:2208.05592, 2022. 2", + "[25] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 2", + "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3, 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 7", + "[28] Janghyeon Lee, Hyeong Gwon Hong, Donggyu Joo, and Junmo Kim. Continual learning with extended kronecker-factored approximate curvature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9001-9010, 2020. 3", + "[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 2", + "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 3, 5, 6, 7", + "[31] Tian Yu Liu, Aditya Golatkar, and Stefano Soatto. Tangent transformers for composition, privacy and removal. arXiv preprint arXiv:2307.08122, 2023. 9", + "[32] Tian Yu Liu, Aditya Golatkar, Stefano Soatto, and Alessandro Achille. Integral continual learning along the tangent vector field of tasks. arXiv preprint arXiv:2211.13108, 2022. 2, 3", + "[33] Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7765-7773, 2018. 3", + "[34] Michael Matena and Colin Raffel. Merging models with fisher-weighted averaging. arXiv preprint arXiv:2111.09832, 2021. 2", + "[35] Fangzhou Mu, Yingyu Liang, and Yin Li. Gradients as features for deep representation learning. arXiv preprint arXiv:2004.05529, 2020. 2, 4, 9", + "[36] Cuong V Nguyen, Yingzhen Li, Thang D Bui, and Richard E Turner. Variational continual learning. arXiv preprint arXiv:1710.10628, 2017. 3", + "[37] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural Networks, 113:54-71, 2019. 3", + "[38] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 6", + "[39] Barak A Pearlmutter. Fast exact multiplication by the hessian. Neural computation, 6(1):147-160, 1994. 4, 9", + "[40] Grégoire Petit, Adrian Popescu, Hugo Schindler, David Picard, and Bertrand Delezoide. Fetril: Feature translation for exemplar-free class-incremental learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3911–3920, 2023. 3, 7, 8", + "[41] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 524-540. Springer, 2020. 5", + "[42] Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In 2009 IEEE conference on computer vision and pattern recognition, pages 413-420. IEEE, 2009. 6" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18685", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 3", + "[44] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. arXiv preprint arXiv:1810.11910, 2018. 3", + "[45] Hippolyt Ritter, Aleksandar Botev, and David Barber. Online structured laplace approximations for overcoming catastrophic forgetting. Advances in Neural Information Processing Systems, 31, 2018. 3, 6", + "[46] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 5, 6, 7", + "[47] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 3", + "[48] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In International conference on machine learning, pages 4528-4537. PMLR, 2018. 5, 7", + "[49] Hyounguk Shon, Janghyeon Lee, Seung Hwan Kim, and Junmo Kim. Dlcft: Deep linear continual fine-tuning for general incremental learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 513-529. Springer, 2022. 2, 3, 6, 7", + "[50] Doris Y Tsao, Winrich A Freiwald, Tamara A Knutsen, Joseph B Mandeville, and Roger BH Tootell. Faces and objects in macaque cerebral cortex. Nature neuroscience, 6(9):989-995, 2003. 1", + "[51] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. arXiv preprint arXiv:1904.07734, 2019. 3, 6, 7", + "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 9", + "[53] Johannes Von Oswald, Seijin Kobayashi, Joao Sacramento, Alexander Meulemans, Christian Henning, and Benjamin F Grewe. Neural networks with late-phase weights. arXiv preprint arXiv:2007.12927, 2020. 2", + "[54] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 2", + "[55] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. arXiv preprint arXiv:2204.04799, 2022. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2, 3", + "[57] Mitchell Wortsman, Suchin Gururangan, Shen Li, Ali Farhadi, Ludwig Schmidt, Michael Rabbat, and Ari S Morcos. lo-fi: distributed fine-tuning without communication. arXiv preprint arXiv:2210.11948, 2022. 2", + "[58] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 2, 5, 6", + "[59] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 2", + "[60] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34:14306-14318, 2021. 3, 8", + "[61] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 3, 8", + "[62] Kai Zhu, Wei Zhai, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Self-sustaining representation expansion for nonexemplar class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9296-9305, 2022. 3, 7, 8" + ], + "bbox": [ + 501, + 90, + 893, + 612 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18686", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/1ae449c1-93a5-4652-bb27-6a83c99dd85c_model.json b/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/1ae449c1-93a5-4652-bb27-6a83c99dd85c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d7e1475b2d70f9d5a242f16f45d0e79c870eb775 --- /dev/null +++ b/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/1ae449c1-93a5-4652-bb27-6a83c99dd85c_model.json @@ -0,0 +1,2345 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.131, + 0.848, + 0.154 + ], + "angle": 0, + "content": "Tangent Model Composition for Ensembling and Continual Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.182, + 0.362, + 0.198 + ], + "angle": 0, + "content": "Tian Yu Liu" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.2, + 0.462, + 0.217 + ], + "angle": 0, + "content": "University of California, Los Angeles" + }, + { + "type": "text", + "bbox": [ + 0.231, + 0.22, + 0.393, + 0.234 + ], + "angle": 0, + "content": "tianyu@cs.ucla.edu" + }, + { + "type": "text", + "bbox": [ + 0.599, + 0.182, + 0.719, + 0.198 + ], + "angle": 0, + "content": "Stefano Soatto" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.2, + 0.808, + 0.217 + ], + "angle": 0, + "content": "University of California, Los Angeles" + }, + { + "type": "text", + "bbox": [ + 0.591, + 0.22, + 0.726, + 0.233 + ], + "angle": 0, + "content": "soatto@ucla.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.664 + ], + "angle": 0, + "content": "Tangent Model Composition (TMC) is a method to combine component models independently fine-tuned around a pre-trained point. Component models are tangent vectors to the pre-trained model that can be added, scaled, or subtracted to support incremental learning, assembling, or unlearning. Component models are composed at inference time via scalar combination, reducing the cost of assembling to that of a single model. TMC improves accuracy by \\(4.2\\%\\) compared to assembling non-linearly fine-tuned models at a \\(2.5 \\times 10 \\times\\) reduction of inference cost, growing linearly with the number of component models. Each component model can be forgotten at zero cost, with no residual effect on the resulting inference. When used for continual fine-tuning, TMC is not constrained by sequential bias and can be executed in parallel on federated data. TMC outperforms recently published continual fine-tuning methods almost uniformly on each setting - task-incremental, class-incremental, and data-incremental - on a total of 13 experiments across 3 benchmark datasets, despite not using any replay buffer. TMC is designed for composing models that are local to a pre-trained embedding, but could be extended to more general settings. The code is available at: https://github.com/tianyu139/tangent-model-composition" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.694, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Compositionality has long been considered central to cognition [5], possibly a reflection of compositionality in neural activity [50]. After decades of unsuccessful attempts to design compositional representations, deep neural networks trained on language data are beginning to exhibit emergent compositionality at scale, by capturing the compositional structure of the data. These models are now being used for visual inference, which has rekindled interest in the study of compositionality of visual representations. But beyond the data, any compositional structure that may be latent in the model is not directly accessible: One cannot simply \"compose\" weights or inner activations and expect" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.466 + ], + "angle": 0, + "content": "meaningful outcomes. The computational architecture of Transformers [52] has been leveraged extensively to co-opt the compositional structure of data through prompts or tokens [29, 55], but still the activations of trained models do not appear to be meaningfully composable. Compositionality of neural activity would allow one to combine activations from different models to capture novel concepts, or incorporate knowledge from different data without having to re-train or fine-tune the core models. This would enable open-universe classification and, more generally, combinatorial expansion of the hypothesis space. Continual learning could be performed simply by composing models trained on different data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.895, + 0.621 + ], + "angle": 0, + "content": "In this paper, we explore the simplest form of compositionality, that is linear combination. We leverage recent results on the linearization of deep neural networks around a pre-trained point, that can be trained by solving a convex optimization problem and yet perform on-par with nonlinear fine-tuning [1]. This suggests that the tangent space at pre-trained models may be used to linearly compose neural activations, or equivalently compose different models trained or fine-tuned on different datasets and/or for different tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.894, + 0.761 + ], + "angle": 0, + "content": "We show that different models obtained from the linearization around a pre-trained point can be composed, combined, rescaled, and forgotten (\"unlearning\"), simply by scalar combinations. This fact can be used to perform ensembling at the inference cost of a single model (Sec. 3). It can also be used for continual fine-tuning, with each component model trained independently and in parallel, if necessary on federated data that can therefore be easily forgotten if needed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Linear combination is not viable for general concept composition nor arbitrary multi-task learning. In Sec. 5, we discuss limitations of our approach, which can only meaningfully compose models that are \"local\" to a pre-trained embedding. If there are models trained on tasks that are far in representation space, or even antagonistic, they are likely to live on different tangent planes, making linear combination inappropriate. One such example, described in the appendix, concerns models trained on real images (Ima" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "18676" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.852, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.387, + 0.895, + 0.459 + ], + "angle": 0, + "content": "Figure 1. Composition of Fine-tuned Models: Purple indicates final models used for inference, yellow indicates specialist models fine-tuned on individual tasks. The paragon (a) is a model trained jointly on all tasks. However, in continual learning, different tasks are instantiated at different times, an re-training on their union is impractical. Ensembling (b) combines the output of different models trained separately on each task. Weight averaging (c) yields a \"model soup\" of fine-tuned non-linear models, which improves inference time; Tanget Model Composition (d) linearly composes models fine-tuned on the tangent space of a pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.663 + ], + "angle": 0, + "content": "geNet) and clip art or sketches (DomainNet). Nonetheless, in more common settings, TMC is competitive with general forms of ensembling such as averaging activations or logits of non-linearly fine-tuned models, and with more general forms of continual learning, including methods that use a replay buffer. For broader task coverage, one could use TMC to construct a tangent bundle of models around different pre-trained points, akin to a \"tangent model zoo.\" This concept is not too dissimilar from the architecture of some Foundation Models with a shared backbone and multiple distinct \"heavy heads\" [54], but far more compact, modular, efficient, interpretable, and easy to work with." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.47, + 0.739 + ], + "angle": 0, + "content": "In the next section we place our contributions in the context of prior art, and in the following one we describe our method in more detail. Sec. 4 summarizes empirical evidence in support of our approach, and finally Sec. 5 discusses its main limitations." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.216, + 0.77 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Compositionality: Compositionality has been studied for decades as a means to expand the representative power of trained models, but has received increasing impetus in recent years thanks to large Transformer-based models that can be used with adaptable prompts. Prompt-tuning [29] has become commonplace to adapt large pre-trained models for specific downstream tasks. Composition of prompts has also been explored in [56, 55] for continual learning. More" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.528 + ], + "angle": 0, + "content": "general composition of parameters of trained deep networks has been explored for various purposes such as improving optimization and generalization, as we describe next." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.893, + 0.682 + ], + "angle": 0, + "content": "Deep network linearization: Deep networks are linearized using the first-order Taylor expansion around a pretrained weight [1, 18, 35, 49, 32]. [1] finetunes linearized networks modified to use Leaky-ReLU with gradient preconditioning to achieve similar performances to non-linear fine-tuning. We note that TMC is not just a matter of applying linearized networks to different datasets and averaging, since different scales would lead to imbalance as we describe in Sec. 3.3. Furthermore, unlike [1, 49], we do not require Leaky-ReLU nor gradient pre-conditioning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Averaging model weights: Weight averaging has been used in deep learning to improve generalization [25, 17], improve pre-training [12, 34], perform distributed finetuning [57], and increase robustness against distribution shift [59]. [25] averages points along the trajectory of SGD to find flatter minima, [53] averages \"late-phase weights\" obtained from the later stages of SGD. [24] uses weight interpolation for model patching, and [58] averages weights of models fine-tuned using different hyperparameters from pre-trained models to improve generalization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "While weight averaging techniques for deep networks are related to ensembles [14, 19], in our case the two coincide since we operate in representation spaces that are linear in the parameters. We leverage this property to train on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18677" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "disjoint subsets of data, and show that they can outperform existing weight averaging and ensembling techniques." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.242 + ], + "angle": 0, + "content": "Continual learning/fine-tuning: Continual learning [37] aims to adapt models to new tasks and data without degrading performance in previously trained tasks. Continual fine-tuning [49, 6] additionally requires minimal forgetting of a pre-training objective to ensure all tasks can benefit from it. Existing approaches can be broadly classified into parameter isolation, regularization, architecture, and replay-based." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.468, + 0.303 + ], + "angle": 0, + "content": "Parameter isolation methods [32, 3, 47] operate in the network architecture space by allocating each task a different set of parameters, using techniques such as network pruning [33] to reduce additional memory requirements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.468, + 0.378 + ], + "angle": 0, + "content": "Regularization-based methods [30, 2, 26, 36, 28] incorporate additional terms into the training objective to prevent catastrophic forgetting. [26] penalizes the distance between weights of the previous and current task, and [45] penalizes changes in weights using Hessian approximations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.468, + 0.575 + ], + "angle": 0, + "content": "Experience replay methods [44, 4, 32, 11, 43, 49, 7, 8, 26, 56, 6, 10, 9] assume the availability of a memory buffer to store samples from previous tasks. [49, 6] applies experience replay to continual fine-tuning. [49] shows that linearized models combined with replay and regularization alleviates catastrophic forgetting when fine-tuned, and [6] combines dark experience replay and an attention-based regularization loss to prevent catastrophic forgetting of the pre-training features, by encouraging similarity of features at each layer to the original pre-trained network. While replay methods have achieved state-of-the-art performance compared to non-replay ones, they require a sufficiently large memory buffer to store previously seen exemplars." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.468, + 0.65 + ], + "angle": 0, + "content": "A further line of work tackles exemplar-free class incremental learning (EFCIL) using regularization [60], distillation [22], and class-prototypes [61, 62, 40] to incrementally integrate new classes into a fixed-sized network without the memory costs of a replay buffer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.468, + 0.741 + ], + "angle": 0, + "content": "We show that composition of tangent models trained on individual tasks is a viable method for replay-free continual fine-tuning, and can even outperform recently published replay-based methods. As a side benefit, our method also yields fully parallel training across tasks, and enables zero-cost unlearning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.168, + 0.77 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In continual learning, models are trained on a sequence of tasks indexed by \\( t = 1,\\dots ,T \\) ,with the goal of performing well on all, under three main scenarios [51]: (1) Task-incremental, where the task identity \\( t \\) (Task-ID) is known during inference, (2) Data/Domain-incremental, where Task-ID is not provided during inference, and (3) the most challenging Class-incremental, where Task-ID is unknown during inference and in addition the classes repre" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.087, + 0.892, + 0.239 + ], + "angle": 0, + "content": "Table 1. A Comparison of Composition Methods: On average across 25 experiments, Tangent Model Ensembling (TME) improves accuracy by \\(5.1\\%\\) over standard ensembling of non-linear fine-tuned models at a \\(2\\times\\) increase of inference cost. Tangent Model Composition (TMC) also improved accuracy, by \\(4.2\\%\\), while reducing inference cost by \\(2.5\\times\\) to \\(10\\times\\) in our experiments, and growing linearly with the number of models in the ensemble. Best results for single and multi model inference are indicated in bold and italic respectively. Joint training paragon accuracy (Original Model/Tangent Model): Caltech-256 - \\(89.17\\% /86.77\\%\\), MIT-67 - \\(77.29\\% /77.74\\%\\), OxfordPets - \\(92.68\\% /93.68\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.24, + 0.89, + 0.641 + ], + "angle": 0, + "content": "
DatasetTasksSoupEns-LEns-SMTMCTME
Inference/MemoryCost:O(1)O(T)O(T)O(1)O(T)
C-Caltech-256 (Class-IL)584.0082.8083.0984.8285.53
1080.1076.8579.7682.3782.58
2074.8666.7675.4778.6679.30
C-MIT-67 (Class-IL)559.4856.7958.6869.4371.72
1049.3042.3650.5264.5365.72
2028.5123.6135.6248.9854.93
C-OxfordPnts (Class-IL)580.7181.1481.8684.4285.04
1071.3569.7177.3177.7579.97
D-Caltech-256 (Data-IL)585.9686.2786.5984.8685.13
1083.4884.0584.8083.6484.16
2079.6480.7283.1781.9982.94
D-MIT-67 (Data-IL)566.0066.5467.2174.8875.45
1058.8859.2060.4772.4173.41
2050.1051.9955.2070.2272.11
D-OxfordPnts (Data-IL)590.2290.8391.0293.1793.21
1087.6688.4288.8892.4892.72
2085.5785.9786.5891.2891.77
C-Caltech-256 (Task-IL)594.0194.7294.8494.3394.76
1095.0095.8796.0496.0396.23
2095.5696.5396.8497.2297.38
C-MIT-67 (Task-IL)583.3686.2486.3791.2791.79
1087.1990.4290.9794.4595.03
2088.4193.7393.8897.7198.09
C-OxfordPnts (Task-IL)597.2797.2797.3598.5798.74
1097.5997.9297.9899.0599.13
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.673, + 0.708, + 0.687 + ], + "angle": 0, + "content": "sented in each task are disjoint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.802 + ], + "angle": 0, + "content": "We assume the setting of continual fine-tuning of a pretrained model. Naive fine-tuning causes catastrophic forgetting of prior tasks, including the one on which pre-training occurred, as the number of tasks increases [6]. Thus, only early tasks benefit from pre-training. Therefore, it is important for continual fine-tuning to ensure that all tasks can benefit from information acquired during pre-training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce a unified method to tackle the main scenarios (1)-(3) by composing separately-trained tangent component models. We show that each task takes full advantage of pre-training, and catastrophic forgetting of prior tasks is averted since information from each previous task is encoded separately and only fused for inference." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18678" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.273, + 0.106 + ], + "angle": 0, + "content": "3.1. Model Linearization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.474, + 0.313 + ], + "angle": 0, + "content": "The effectiveness of model aggregation in weight space hinges on whether the models lie in within the same small-loss training basin of convergence. It is known that weights of different models fine-tuned on the same dataset can be connected by linear paths with constant loss (\"linear mode connectivity\") [16], but this does not generally occur for models fine-tuned on disjoint tasks/datasets, which is the case of interest to us. Therefore, we convexify the loss landscape locally by approximating the function represented by a deep (non-linear) network with its first-order Taylor approximation around the pre-trained weights. For any deep network \\( f \\) and initial weights \\( w \\in \\mathbb{R}^m \\), we denote with \\( \\mathcal{H}(w) \\) the set of all such models tangent to \\( f_w \\) given by:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.337, + 0.47, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} (w) = \\left\\{h _ {\\delta} (\\cdot) \\triangleq f _ {w} (\\cdot) + \\nabla_ {w} f _ {w} (\\cdot) \\cdot \\delta \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.471, + 0.442 + ], + "angle": 0, + "content": "Since the tangent model \\( h_{\\delta} \\) is linear in the parameters \\( \\delta \\), training it with standard losses such as mean-squared error (MSE) or empirical cross-entropy yields a convex loss landscape. In order to compose models, leveraging Jensen's inequality, we are guaranteed that for any dataset \\( \\mathcal{D} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.453, + 0.47, + 0.495 + ], + "angle": 0, + "content": "\\[\nL \\left(\\mathcal {D}, \\sum_ {i} ^ {T} h _ {\\theta_ {i} \\delta_ {i}}\\right) \\leq \\sum_ {i} \\theta_ {i} L \\left(\\mathcal {D}, h _ {\\delta_ {i}}\\right), \\quad \\sum_ {i} \\theta_ {i} = 1. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.609 + ], + "angle": 0, + "content": "In other words, the loss resulting from a convex combination of tangent models in weight space will always be lower than or equal to the convex combination of their individual losses. This is particularly useful when combining models trained to some constant (possibly zero) loss, since it ensures that the composed model will have loss lower or equal to the loss of the component models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.472, + 0.746 + ], + "angle": 0, + "content": "While tangent models might seem computationally expensive to train, we note that the Jacobian vector product \\(\\nabla_w f_w(\\cdot) \\cdot \\delta\\) can be computed in a single forward pass as shown by [39, 35, 1]. This reduces the computational cost to that of a forward pass through the original network. Also note that while the tangent model has twice the degrees of freedom of the original model, fixing the pre-trained model at \\(w\\) results in only a convex optimization in the remaining parameters, as many as the original model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.47, + 0.807 + ], + "angle": 0, + "content": "In the context of fine-tuning, \\( w \\) is initialized by pretraining on large datasets such as ImageNet [13]. In the rest of this paper, we use the short-form notation \\( \\mathcal{H} \\) and \\( h \\) for \\( \\mathcal{H}(w) \\) and \\( h_{\\delta} \\) respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.47, + 0.847 + ], + "angle": 0, + "content": "3.2. Tangent Model Composition for Replay-free Continual Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.853, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Ensembling is an effective strategy for combining weak learners into a strong model [14]. For any finite \\(F \\subseteq \\mathbb{A}\\), where \\(\\mathbb{A}\\) is a set of functions, and \\(\\lambda = \\{\\lambda_i\\}_{i=1}^{|F|}\\) a set of" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.88, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.3, + 0.895, + 0.412 + ], + "angle": 0, + "content": "Figure 2. Accuracy plot over models trained on C-Caltech-256, C-MIT-67, and C-Oxford Pets, split into 10 tasks each. Soft-max ensembling (TME) consistently yields the best results compared to logit ensembling (TMC) for each loss function. Composition of component models trained with the Rescaled Square Loss (RSL) outperforms those trained with empirical cross-entropy (CE) or MSE loss, trading off only \\(1.2\\%\\) accuracy against TME to reduce inference time and memory requirement from \\(\\mathcal{O}(T)\\) to \\(\\mathcal{O}(1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.436, + 0.825, + 0.452 + ], + "angle": 0, + "content": "weights, the logit ensemble operation is given by" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.461, + 0.892, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E n s e m b l e} (F, \\lambda) = \\sum_ {i = 1} ^ {| F |} \\lambda_ {i} f _ {i} (\\cdot) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.59 + ], + "angle": 0, + "content": "In the context of continual learning, a simple yet effective approach is to simply train a new model on each task \\( t \\), and ensemble the outputs at inference time. Unfortunately, this has the obvious shortcoming that both storage requirements and inference time grows along with the number of tasks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.59, + 0.892, + 0.621 + ], + "angle": 0, + "content": "However, consider the case where \\(\\mathbb{A} = \\mathcal{H}\\). Then for any finite \\(H \\subseteq \\mathcal{H}\\) and associated weights \\(\\lambda\\), we can write" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.63, + 0.892, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {E n s e m b l e} (H, \\lambda) = \\sum_ {i = 1} ^ {| H |} \\lambda_ {i} h _ {i} (\\cdot) (4) \\\\ = f _ {w} (\\cdot) + \\sum_ {i = 1} ^ {| H |} \\nabla_ {w} f _ {w} (\\cdot) \\cdot \\lambda_ {i} \\delta_ {i} (5) \\\\ = h _ {\\sum \\lambda_ {i} \\delta_ {i}} (\\cdot) (6) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.84 + ], + "angle": 0, + "content": "In other words, since \\(\\mathcal{H}\\) defines a vector space with addition and scalar multiplication based on \\(\\delta\\), any linear combination (or ensemble) of models in \\(\\mathcal{H}\\) is equivalent to a single model in \\(\\mathcal{H}\\) derived simply by the linear combination over \\(\\delta\\). This reduces the inference time for any ensemble of \\(T\\) tangent models from \\(\\mathcal{O}(T)\\) to \\(\\mathcal{O}(1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In continual learning, it is also important for storage space to not scale with \\( T \\), since the number of tasks can be arbitrarily large. However, since tasks arrive sequentially in the continual learning framework, we can further reduce" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18679" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.087, + 0.895, + 0.129 + ], + "angle": 0, + "content": "Table 2. Class-Incremental Learning: Comparison against existing methods using an ImageNet pre-trained ResNet-50. TMC is almost uniformly better than all replay-based and replay-free methods tested, outperforming the best replay-free (RF) method by \\(14.48\\%\\) and even the best replay-based method [6] by \\(1.53\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.13, + 0.846, + 0.317 + ], + "angle": 0, + "content": "
Method / BufferTasksC-Caltech-256C-MIT-67
5102051020
ER [46] / 30070.14±0.2265.75±0.9265.490.6054.61±0.4149.57±1.4645.331.73
LWF [30] / None34.55±0.1822.76±0.1714.37±0.1425.14±0.5713.30±0.496.41±0.17
oEWC [48] / None28.71±1.3116.32±4.3715.59±3.8314.18±0.777.12±0.494.33±0.16
DER++ [7] / 30080.50±0.2677.84±1.0274.46±0.2164.56±0.5559.73 ±1.0250.28±1.27
Co2L [10] / 30020.71±2.8917.13±2.5314.620.6119.64±12.9321.21±0.5017.78±1.33
ER-ACE [9] / 30078.47±1.2573.85±0.5370.87±1.3864.94±0.5060.25±0.2652.48±0.29
TWF (RF) [6] / None77.95±0.5270.67±0.7262.79±0.8256.58±0.6245.03 ±0.7828.91±0.92
TWF [6] / 30081.34±0.9177.03±0.5372.44±0.4268.33±0.4664.21 ±0.6556.29±0.68
TMC / None84.82 ±0.0382.37 ±0.0878.66 ±0.1369.43±0.2164.53±0.2848.98 ±0.13
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.473, + 0.566 + ], + "angle": 0, + "content": "the memory capacity required for storage of models from \\(\\mathcal{O}(T)\\), to \\(\\mathcal{O}(1)\\) by using a simple autoregressive model: We train on the first task \\(t_1\\) to produce model \\(\\tilde{h}_1 = h_1\\). Then, for any subsequent task \\(t > 1\\), we train a new model \\(\\tilde{h}_t\\), and compose it with \\(h_{t-1}\\) to give \\(h_t = \\lambda_{(t,1)}\\tilde{h}_t + \\lambda_{(t,2)}h_{t-1} := h_{\\delta'}\\) where \\(\\delta' = \\frac{\\lambda_{(t,1)}\\tilde{\\delta}_t + \\lambda_{(t,2)}\\delta_{t-1}}{\\lambda_{(t,1)} + \\lambda_{(t,2)}}\\). Since we only require model \\(h_{t-1}\\) to produce \\(h_t\\), all prior models up to and including \\(h_{t-1}\\) can be discarded upon constructing \\(h_t\\), hence reducing memory requirement from \\(\\mathcal{O}(T)\\) to \\(\\mathcal{O}(1)\\), equal to the size of a single model. In contrast to methods that expands the model at each task, or require a memory buffer of examples from previous tasks, this method does not incur additional storage requirements, nor assume availability of an extra replay buffer for storing past exemplars (i.e. our method is replay-free)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.471, + 0.597 + ], + "angle": 0, + "content": "Remark 1 (Free Parallelization): Additionally, for any given convex loss function \\(L\\), the optimization objective" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.604, + 0.47, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\delta \\in \\mathbb {R} ^ {m}} \\sum_ {(x, y) \\in \\mathcal {D}} L \\left(h _ {\\delta} (x), y\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.81 + ], + "angle": 0, + "content": "is convex and thus converges to a global optimum. This implies that each model \\(\\tilde{h}_t\\) can be trained from a standard initialization as opposed to from \\(h_{t-1}\\). Hence, if multiple tasks arrive at the same time, such as when training in a federated manner, training can be parallelized across tasks and composed in a zero-cost, zero-shot manner to yield equivalent results (assuming equivalence of global optima) to sequential training and composition. To the best of our knowledge, apart from [41] which simply stores samples in a replay buffer and trains on them at test time, no other work in continual learning can be parallelized across tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Remark 2 (Free Forgetting/Unlearning): If we relax memory constraints to store models from each task, unlearning any specific task \\( t \\) (e.g., private data associated with specific samples) is straightforward and can be done in a zero-cost, zero-shot manner. Since subtraction is well-defined in the vector space \\( \\mathcal{H} \\), any task \\( i \\) can be unlearned" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.331, + 0.892, + 0.362 + ], + "angle": 0, + "content": "from the model \\( h_T \\) trained on \\( T \\) tasks by simply subtracting the weights associated to \\( \\tilde{h}_i \\):" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.373, + 0.892, + 0.415 + ], + "angle": 0, + "content": "\\[\nh _ {T \\backslash i} = h _ {T} - \\lambda_ {i, 1} \\prod_ {j = i + 1} ^ {T} \\lambda_ {(j, 2)} \\tilde {h} _ {i} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.425, + 0.892, + 0.486 + ], + "angle": 0, + "content": "Remark 3 (Unified Continual Fine-tuning): Our method does not assume any specific task splits, and is hence generalizable across all three forms of continual fine-tuning: task-incremental, class-incremental, and data-incremental." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.695, + 0.51 + ], + "angle": 0, + "content": "3.3. Scale Standardization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.684 + ], + "angle": 0, + "content": "Training with the cross-entropy loss does not ensure that the weights are comparable in scale. We can study the implications of this for composition on \\(\\mathcal{H}\\) through the lens of logits ensembling. While logits ensembling has been shown to be effective [58], Fig. 2 shows that ensembling the normalized soft-max outputs (Tangent Model Ensembling, TME) is uniformly better than logit ensembling/weight averaging via TMC across various loss functions. However, TME cannot be represented by a linear combination of perturbations \\(\\delta_{i}\\) and hence incurs \\(\\mathcal{O}(T)\\) model storage costs and \\(\\mathcal{O}(T)\\) inference time." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.685, + 0.892, + 0.73 + ], + "angle": 0, + "content": "Instead, we propose to induce a waterbed effect similar to soft-max ensembling using standardization. In particular, we train using the Rescaled Square Loss (RSL) [23]:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.741, + 0.892, + 0.799 + ], + "angle": 0, + "content": "\\[\nL _ {s} (x, y) = \\frac {1}{K} \\left(\\alpha ([ f (x) ] _ {y} - \\beta) ^ {2} + \\sum_ {i = 1, i \\neq y} ^ {K} ([ f (x) ] _ {i}) ^ {2}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.87 + ], + "angle": 0, + "content": "where \\(\\alpha, \\beta\\) are hyper-parameters. This is the regular MSE loss when \\(\\alpha = \\beta = 1\\). Larger values of \\(\\beta\\) reduce soft-max entropy in the output by scaling positive output class signals while keeping negative ones close to zero." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We use \\(\\beta\\) to control the interaction between component models. For example, when tasks are highly dissimilar," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18680" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.47, + 0.184 + ], + "angle": 0, + "content": "Table 3. Data-Incremental Learning on D-MIT-67, 4 tasks using ResNet-18. For fair comparison, we run our method under the settings proposed by [49]. TMC outperforms the best replay-free method [30] by \\(4.61\\%\\), and the best replay-based method [49] by \\(1.27\\%\\), while supporting parallel training across tasks (PL). If we sacrifice parallelization by initializing each model at task \\(t\\) with \\(h_{t-1}\\) (TMC-Seq), we further improve over TMC by \\(0.28\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.185, + 0.451, + 0.372 + ], + "angle": 0, + "content": "
MethodPL?Replay BufferAccuracy
SGDXNone63.48
LwF[30]XNone67.21
MAS[2]XNone62.49
OSLA[45]XNone64.08
EWC [26]X50063.68
DLCFT[49]X50070.55
TMC✓None71.82
TMC-SeqXNone72.10
Joint (Linear)--73.61
Joint (Paragon)--74.40
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.468, + 0.486 + ], + "angle": 0, + "content": "such as in task- and class-incremental continual learning, we use larger values of \\(\\beta\\) to reduce interference of component models trained on other tasks upon composition. On the other hand, when tasks are highly synergistic such as in data-incremental learning, lower values of \\(\\beta\\) can be used to encourage interaction in the final composed model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.514, + 0.208, + 0.531 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.545, + 0.468, + 0.574 + ], + "angle": 0, + "content": "We describe comparison baselines in Sec. 4.1, and our main results in Sec. 4.2, with additional studies in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.468, + 0.73 + ], + "angle": 0, + "content": "Our method is replay-free. As such, its computational cost is at most half of what it would be if it used a replay buffer since replay-based methods typically sample separate batches of equal size from both the current task and replay buffer at each iteration [7, 6, 10, 46, 9]. Nonetheless, we compare our method to both replay-based and replay-free methods in our baselines. While replay-based methods have thus far outperformed replay-free ones, ours is uniformly more accurate than replay-based methods tested, in addition to replay-free ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We evaluate our method on Caltech-256 [20], MIT-67 [42], and OxfordPets [38]. For continual fine-tuning, we further split each training dataset into multiple disjoint subsets, each representing a single task, for fair comparison across experiments. We prefix sequential datasets split into tasks with disjoint labels (for task- and class-incremental learning) with C-, and random splits using a standardized random seed (for data-incremental learning) with D-. We set \\(\\alpha = 1, \\beta = 25\\) for the former, \\(\\alpha = 1, \\beta = 5\\) for the latter, and \\(\\alpha = 1, \\beta = 5\\) for TME. Results in all tables and figures are averaged across 3 runs." + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.093, + 0.835, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.262, + 0.836, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.435, + 0.892, + 0.532 + ], + "angle": 0, + "content": "Figure 3. Ablation on \\(\\beta\\) averaged across (Top:) C-Caltech-256, C-MIT-67, and C-OxfordPets and (Bottom:) D-Caltech-256, D-MIT-67, and D-OxfordPets, split into 10 disjoint tasks. For class incremental datasets, larger values of \\(\\beta\\) result in better generalization of the final composed model. For data-incremental datasets where tasks are synergistic, larger values of \\(\\beta\\) instead harm the accuracy of both the component models and the composed model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.709, + 0.573 + ], + "angle": 0, + "content": "4.1. Composition Baselines" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.587, + 0.891, + 0.617 + ], + "angle": 0, + "content": "In each of the following sections, we compare against the following baselines for model aggregation:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.623, + 0.892, + 0.652 + ], + "angle": 0, + "content": "Logits Ensemble (Ens-L): Averaging the output logits of non-linear component models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.659, + 0.89, + 0.689 + ], + "angle": 0, + "content": "Softmax Ensemble (Ens-SM): Averaging the softmax output of non-linear component models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.785 + ], + "angle": 0, + "content": "Weight composition of non-linear models (Soup): Averaging the weights of non-linear component models. This method is inspired by Model Soups [58] where weights of various non-linear models trained on the same task but with different hyper-parameters and augmentations are averaged to improve generalization at no additional inference cost." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.812, + 0.638, + 0.826 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We show that TMC outperforms existing methods in all three continual fine-tuning settings [51], from hardest to easiest: Class-Incremental (Sec. 4.2.1), Data-Incremental (Sec. 4.2.2), and Task-Incremental (Sec. 4.2.3)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18681" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.088, + 0.894, + 0.129 + ], + "angle": 0, + "content": "Table 4. Task-Incremental Learning: Comparison against existing methods using ImageNet pre-trained ResNet-50. We almost uniformly outperform all replay-based and replay-free methods that we compare against. In particular, we improve over the best replay-free (RF) method by \\(2.46\\%\\), and the best replay-based method by \\(1.38\\%\\) [6]." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.13, + 0.847, + 0.317 + ], + "angle": 0, + "content": "
Method / BufferTasksC-Caltech-256C-MIT-67
5102051020
ER [46] / 30092.11±0.2694.33±0.4895.76±0.1385.84±0.0990.59±0.5195.13±0.27
LWF [30] / None93.73±0.1295.31±0.0896.29±0.0587.55±0.2090.82±0.0495.68±0.58
oEWC [48] / None83.97±0.7176.20±16.0791.49±0.3067.31±3.2268.45±4.9384.74±1.96
DER++ [7] / 30093.79±0.1394.92±0.1195.87±0.1786.83±0.2691.63±0.7594.54±0.62
Co2L [10] / 30034.51±3.0539.62±1.7045.94±1.4739.32±20.4764.40±1.3151.50±2.21
ER-ACE [9] / 30093.11±0.1294.63±0.0695.86±0.1087.13±0.1690.96±0.2894.92±0.60
TWF (RF) [6] / None94.45±0.0595.64±0.0696.36±0.1286.76±0.2588.27±0.5494.74±0.35
TWF [6] / 30094.38±0.0895.74±0.1796.59±0.0788.22±0.4192.13±0.2695.67±0.46
TMC / None94.33±0.0396.03±0.0297.22±0.0391.27±0.2194.45±0.1397.71±0.13
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.331, + 0.329, + 0.347 + ], + "angle": 0, + "content": "4.2.1 Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.471, + 0.399 + ], + "angle": 0, + "content": "In class-incremental learning, datasets are partitioned so that label spaces of each task are disjoint." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.47, + 0.556 + ], + "angle": 0, + "content": "Comparison against Composition Baselines: We compare against model composition baselines in Table 1. While TME yields the best generalization performance, this comes at a cost of \\(\\mathcal{O}(T)\\) inference time. On the other hand, TMC achieves an \\(\\mathcal{O}(1)\\) inference time with only a \\(1.73\\%\\) trade-off in generalization accuracy on average, while outperforming weight averaging, logit ensembling, and even soft-max ensembling of non-linear models by an average of \\(7.83\\%\\), \\(11.37\\%\\), and \\(6.08\\%\\) respectively, even though the latter two both require \\(\\mathcal{O}(T)\\) inference time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.47, + 0.715 + ], + "angle": 0, + "content": "Comparison against existing methods: In Table 2, we compare against recent continual fine-tuning methods including both replay-free and replay-based methods. Here, replay-free methods [30, 48] perform poorly in the class-incremental setting relative to stronger replay-based methods [46, 9, 7, 6]. However, we show that not only do we improve over the best replay-free method by \\(14.48\\%\\), but also over the best replay-based method by \\(1.53\\%\\) [6], corresponding to a \\(17.59\\%\\) relative error reduction towards the paragon performance of joint training (Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Comparison against EFCIL methods: In Table 5, we compare against EFCIL methods on C-TinyImageNet [27]. The original EFCIL methods are evaluated under a different setting (denoted \\(\\dagger\\)), where the zeroth task contains \\(50\\%\\) of the dataset. Hence, we further modified two of the best performing EFCIL methods, SSRE [62] and FeTrIL [40] to use the pre-trained ImageNet initialization, and run them under our setting of uniformly sharded data. TMC not only outperforms all methods ran under our harder setting, by an average of \\(15.3\\%\\) over the next best method [40], but even yields equal or better performance when compared to all methods ran under \\(\\dagger\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.331, + 0.747, + 0.347 + ], + "angle": 0, + "content": "4.2.2 Data-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.892, + 0.401 + ], + "angle": 0, + "content": "Also referred to as domain-incremental learning, this setting assumes that tasks are split in a non-stratified manner, where each task has the same output space [51, 49]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.893, + 0.507 + ], + "angle": 0, + "content": "Comparison against Composition Baselines: In Table 1, we compare against model composition baselines. Similar to the class-incremental setting, TME outperforms TMC by an average of \\(0.66\\%\\) at a cost of \\(\\mathcal{O}(T)\\) inference time. TMC uses \\(\\mathcal{O}(1)\\) inference time and outperforms non-linear composition, logit ensemble, and even soft-max ensemble by an average of \\(7.04\\%\\), \\(6.32\\%\\), and \\(5.21\\%\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.894, + 0.675 + ], + "angle": 0, + "content": "Comparison against existing methods: In Table 3, we benchmark our model against results obtained by [49] on a pre-trained ResNet-18 model using the D-MIT-67 dataset split into 4 disjoint subsets. We improve over the best replay-free method [30] by \\(4.61\\%\\), and the best replay-based method [49] by \\(1.27\\%\\), corresponding to a \\(64\\%\\) and \\(33\\%\\) relative error reduction towards the joint training paragon respectively. TMC-Seq, which uses sequential initialization by initializing each component model \\(h_t\\) with \\(h_{t - 1}\\) instead of \\(h_0\\), further improves over TMC by \\(0.28\\%\\) but at the cost of parallelizability." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.695, + 0.747, + 0.71 + ], + "angle": 0, + "content": "4.2.3 Task-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.78 + ], + "angle": 0, + "content": "In task-incremental learning, the output spaces of different tasks are disjoint, but the source task of each sample is known at inference time. Thus, we can restrict the model predictions to that of the given task during evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Comparison against Composition Baselines: Due to the disjoint output spaces of each task, ensembling outputs of specialist component models trained on individual tasks and restricting them to a target task essentially only considers the predictions of the specialist model trained on the target task (since, in general, specialist models are not effective discriminants for tasks other than those they are triend on). Hence, the independence of output spaces renders en" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18682" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.087, + 0.895, + 0.142 + ], + "angle": 0, + "content": "Table 5. Comparison against EFCIL methods using either ImageNet pre-trained ResNet-50, or under the original EFCIL setting where the zeroth task contains \\(50\\%\\) of the dataset (denoted \\(\\dagger\\)). All experiments are ran on C-TinyImageNet. Italics denote the best method under \\(\\dagger\\), and bold denotes the best method under our harder setting. Under the former, TMC performs equally or better than all methods. Under the latter, TMC outperforms the best method by an average of \\(15.3\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.143, + 0.825, + 0.205 + ], + "angle": 0, + "content": "
TasksPASS† [61]IL2A† [60]SSRE† [62]FeTrIL† [40]SSRE[62]FeTrIL[40]TMC
549.647.350.454.833.744.157.6
1047.344.748.953.125.636.153.1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.219, + 0.471, + 0.369 + ], + "angle": 0, + "content": "semblies of specialist models highly effective. In contrast, composing non-linear models in weight space does not ensure minimal interference across different tasks. Indeed, we show in Table 1 that TME yields the best results, while the non-linear composition approach yields the worst in task-incremental learning. On the other hand, due to the equivalence of TMC to (logit) ensembling of linearized models, TMC outperforms not only weight averaging, but also logit and soft-max ensembles of non-linear models, by \\(3.78\\%\\), \\(1.83\\%\\), and \\(1.79\\%\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.552 + ], + "angle": 0, + "content": "Comparison against existing methods: Since each component model is trained only on task \\( t \\), by design, there is no interference from other tasks when training \\( \\tilde{h}_t \\). As such, the final composed model is equivalent to a specialist model when restricted to any task. However, in other state of the art continual learning methods such as those involving replay or loss regularization, this separation among trained for each task is violated. Hence, under the task-incremental setting, we show in Table 4 that TMC achieves the best performance across both replay-free and replay-based methods even without using any memory buffer, beating the next best method by \\( 2.46\\% \\) and \\( 1.38\\% \\) respectively [6]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.562, + 0.233, + 0.576 + ], + "angle": 0, + "content": "4.3. Further studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.647 + ], + "angle": 0, + "content": "We present additional studies on the Rescaled Square Loss (Sec. 4.3.1), and compare against TMC-Seq to show that our method yields an effective and highly parallelizable training scheme for continual fine-tuning (Sec. 4.3.2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.283, + 0.682 + ], + "angle": 0, + "content": "4.3.1 Rescaled Square Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.469, + 0.827 + ], + "angle": 0, + "content": "In Sec. 3.3, we showed that \\(\\beta\\) can be used to control the amount of interaction between tasks. In Fig. 3, we empirically show that in the class-incremental setting where tasks are highly dissimilar, larger values of \\(\\beta\\) result in a better generalist model. On the other hand, under the data-incremental setting where tasks are more similar, information from other tasks can benefit the learning of any given task. Here, larger values of \\(\\beta\\) instead harm generalization due to lowering the accuracy of each component model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.345, + 0.862 + ], + "angle": 0, + "content": "4.3.2 Convexity and Parallelizability" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Given a sequence of tasks \\(1 \\ldots T\\) and a pre-trained model \\(h_0\\), there are at least two natural ways to learn the compo" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.22, + 0.859, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Figure 4. Accuracy of composed models, where component models are initialized with \\( h_0 \\) (TMC) and \\( h_{t-1} \\) (TMC-Seq). In the data-incremental setting (D), TMC-Seq outperforms TMC, while in the class-incremental setting (C), TMC performs significantly better. All results are averaged across D/C-Caltech-256, D/C-MIT-67, D/C-OxfordPets (we exclude C-OxfordPets in the 20-task setting since it does not contain enough classes)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.704 + ], + "angle": 0, + "content": "nent model \\(\\tilde{h}_t\\): (1) Train \\(\\tilde{h}_t\\) with \\(h_{t-1}\\) as initialization, and (2) Train \\(\\tilde{h}_t\\) with \\(h_0\\) as initialization. The first option (TMC-Seq) is most common among continual learning methods which impose regularization in the weights between the current and previous tasks or optimize using memory buffers consisting of examples obtained from previous tasks. As such, these methods are restricted to a sequential learning framework, and hence cannot be parallelized across tasks. In Sec. 3, we showed that, thanks to the convexity of the loss landscape, both options (1) and (2) are equivalent under our framework when we assume comparable generalization performance among global optima." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In practice, for massively over-parameterized networks, such assumptions do not hold. However, we show empirically in Fig. 4 that under the data-incremental setting, compared to TMC-Seq, TMC only incurs a small loss in generalization performance to enable massive parallelization across tasks. This can potentially provide orders of magnitude speed-ups compared to existing continual learning methods. Moreover, under the class-incremental setting, TMC greatly outperforms TMC-Seq, showing that the pre-trained initialization yields solutions of better generalization than \\( h_{t - 1} \\). We attribute this to the fact that the hypothesis space of models which are globally optimal for the training data is vast and yields varying generalization." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18683" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "Thus, regardless of the loss landscape convexity, convergence to each point in this space depends on the \"direction of approach\". Hence, initialization at \\( h_{t-1} \\) introduces a bias towards the previous tasks. Due to the high level of dissimilarity between tasks in the class-incremental setting, this bias harms the final model generalization performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.197, + 0.192, + 0.213 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.224, + 0.47, + 0.329 + ], + "angle": 0, + "content": "Our approach enables incremental learning under the assumption that increment tasks are \"close\" to the pre-trained model. We ensure proximity by training increment models independently on the tangent plane of the pre-trained model. The advantage is that pre-trained tasks are not forgotten, and increment tasks can be learned independently, in parallel, and easily forgotten if needed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.471, + 0.466 + ], + "angle": 0, + "content": "Of course, this approach does not address the problem of incremental learning in full generality, when increment tasks can be arbitrary and arbitrarily unrelated. Nonetheless, we have shown that our method is competitive with existing methods in some of the most challenging settings, for instance when the hypothesis spaces of component tasks are disjoint. On the positive side, despite not requiring a replay buffer, our method outperforms replay-based ones under the assumptions in which we operate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.471, + 0.618 + ], + "angle": 0, + "content": "A limitation of our approach is that it requires computing the linear span of pre-trained models, which can be challenging. However, for common deep network architectures such as the ResNet [21] or Transformer [52, 15] family, the Jacobian-Vector product can be efficiently evaluated [39, 35, 1, 31], reducing the computational complexity of training the tangent model to that of the same magnitude as a linear classifier on network activations. Hence, inference time for the linearized network is at most double that of the original model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.47, + 0.723 + ], + "angle": 0, + "content": "Another limitation of our approach is assuming a first-order expansion around a pre-trained objective can well approximate the fine-tuned model. Such assumptions hold poorly when the pre-training and downstream tasks are highly unrelated. Nevertheless, an ImageNet [13] pretrained model already yields a sufficiently good local approximation for various real-world datasets [1, 31]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.471, + 0.815 + ], + "angle": 0, + "content": "While we mainly explored linear compositionality in the context of continual learning, we note that this framework can be easily generalized. We discussed possible applications towards federated learning and forgetting, since our method naturally yields a parallel training framework that fully compartmentalizes information obtained each task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.266, + 0.848 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We would like to thank Aditya Golatkar, Albert Zhao, and the anonymous reviewers for their feedback on the initial version of the paper. This work was supported by ARO" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.092, + 0.84, + 0.106 + ], + "angle": 0, + "content": "W911NF-17-1-0304 and ONR N00014-22-1-2252." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.119, + 0.597, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.144, + 0.895, + 0.214 + ], + "angle": 0, + "content": "[1] Alessandro Achille, Aditya Golatkar, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. Lqf: Linear quadratic fine-tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15729-15739, 2021. 1, 2, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.215, + 0.895, + 0.284 + ], + "angle": 0, + "content": "[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.285, + 0.895, + 0.34 + ], + "angle": 0, + "content": "[3] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3366-3375, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.341, + 0.895, + 0.395 + ], + "angle": 0, + "content": "[4] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.397, + 0.895, + 0.438 + ], + "angle": 0, + "content": "[5] Elie Bienenstock and Stuart Geman. Compositionality in neural systems. In The handbook of brain theory and neural networks, pages 223-226. 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.439, + 0.895, + 0.534 + ], + "angle": 0, + "content": "[6] Matteo Boschini, Lorenzo Bonicelli, Angelo Porrello, Giovanni Bellitto, Matteo Pennisi, Simone Palazzo, Concetto Spampinato, and Simone Calderara. Transfer without forgetting. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIII, pages 692-709. Springer, 2022. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.536, + 0.895, + 0.605 + ], + "angle": 0, + "content": "[7] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.607, + 0.895, + 0.675 + ], + "angle": 0, + "content": "[8] Pietro Buzzega, Matteo Boschini, Angelo Porrello, and Simone Calderara. Rethinking experience replay: a bag of tricks for continual learning. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 2180-2187. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.677, + 0.895, + 0.732 + ], + "angle": 0, + "content": "[9] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuyte-laars, Joelle Pineau, and Eugene Belilovsky. New insights on reducing abrupt representation change in online continual learning. arXiv preprint arXiv:2203.03798, 2022. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.733, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[10] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. In Proceedings of the IEEE/CVF International conference on computer vision, pages 9516-9525, 2021. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.789, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[11] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.831, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[12] Leshem Choshen, Elad Venezian, Noam Slonim, and Yoav Katz. Fusing finetuned models for better pretraining. arXiv preprint arXiv:2204.03044, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.144, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18684" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.164 + ], + "angle": 0, + "content": "[14] Thomas G Dietterich. Ensemble methods in machine learning. In International workshop on multiple classifier systems, pages 1-15. Springer, 2000. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[16] Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lottery ticket hypothesis. In International Conference on Machine Learning, pages 3259-3269. PMLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.469, + 0.362 + ], + "angle": 0, + "content": "[17] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[18] Aditya Golatkar, Alessandro Achille, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. Mixed-privacy forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 792–801, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[19] Raphael Gontijo-Lopes, Yann Dauphin, and Ekin D Cubuk. No one representation to rule them all: Overlapping features of training methods. arXiv preprint arXiv:2110.12899, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.518 + ], + "angle": 0, + "content": "[20] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[22] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831-839, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[23] Like Hui and Mikhail Belkin. Evaluation of neural architectures trained with square loss vs cross-entropy in classification tasks. arXiv preprint arXiv:2006.07322, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[24] Gabriel Ilharco, Mitchell Wortsman, Samir Yitzhak Gadre, Shuran Song, Hannaneh Hajishirzi, Simon Kornblith, Ali Farhadi, and Ludwig Schmidt. Patching open-vocabulary models by interpolating weights. arXiv preprint arXiv:2208.05592, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[25] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[27] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[28] Janghyeon Lee, Hyeong Gwon Hong, Donggyu Joo, and Junmo Kim. Continual learning with extended kronecker-factored approximate curvature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9001-9010, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.238, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[31] Tian Yu Liu, Aditya Golatkar, and Stefano Soatto. Tangent transformers for composition, privacy and removal. arXiv preprint arXiv:2307.08122, 2023. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.325, + 0.892, + 0.379 + ], + "angle": 0, + "content": "[32] Tian Yu Liu, Aditya Golatkar, Stefano Soatto, and Alessandro Achille. Integral continual learning along the tangent vector field of tasks. arXiv preprint arXiv:2211.13108, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.382, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[33] Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7765-7773, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.44, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[34] Michael Matena and Colin Raffel. Merging models with fisher-weighted averaging. arXiv preprint arXiv:2111.09832, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[35] Fangzhou Mu, Yingyu Liang, and Yin Li. Gradients as features for deep representation learning. arXiv preprint arXiv:2004.05529, 2020. 2, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.528, + 0.892, + 0.569 + ], + "angle": 0, + "content": "[36] Cuong V Nguyen, Yingzhen Li, Thang D Bui, and Richard E Turner. Variational continual learning. arXiv preprint arXiv:1710.10628, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.571, + 0.892, + 0.625 + ], + "angle": 0, + "content": "[37] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural Networks, 113:54-71, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[38] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[39] Barak A Pearlmutter. Fast exact multiplication by the hessian. Neural computation, 6(1):147-160, 1994. 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[40] Grégoire Petit, Adrian Popescu, Hugo Schindler, David Picard, and Bertrand Delezoide. Fetril: Feature translation for exemplar-free class-incremental learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3911–3920, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[41] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 524-540. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[42] Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In 2009 IEEE conference on computer vision and pattern recognition, pages 413-420. IEEE, 2009. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18685" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[43] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[44] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. arXiv preprint arXiv:1810.11910, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.273 + ], + "angle": 0, + "content": "[45] Hippolyt Ritter, Aleksandar Botev, and David Barber. Online structured laplace approximations for overcoming catastrophic forgetting. Advances in Neural Information Processing Systems, 31, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[46] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[47] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[48] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In International conference on machine learning, pages 4528-4537. PMLR, 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.537 + ], + "angle": 0, + "content": "[49] Hyounguk Shon, Janghyeon Lee, Seung Hwan Kim, and Junmo Kim. Dlcft: Deep linear continual fine-tuning for general incremental learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 513-529. Springer, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.47, + 0.594 + ], + "angle": 0, + "content": "[50] Doris Y Tsao, Winrich A Freiwald, Tamara A Knutsen, Joseph B Mandeville, and Roger BH Tootell. Faces and objects in macaque cerebral cortex. Nature neuroscience, 6(9):989-995, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[51] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. arXiv preprint arXiv:1904.07734, 2019. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.692 + ], + "angle": 0, + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.748 + ], + "angle": 0, + "content": "[53] Johannes Von Oswald, Seijin Kobayashi, Joao Sacramento, Alexander Meulemans, Christian Henning, and Benjamin F Grewe. Neural networks with late-phase weights. arXiv preprint arXiv:2007.12927, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[54] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[55] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. arXiv preprint arXiv:2204.04799, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.175 + ], + "angle": 0, + "content": "[56] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[57] Mitchell Wortsman, Suchin Gururangan, Shen Li, Ali Farhadi, Ludwig Schmidt, Michael Rabbat, and Ari S Morcos. lo-fi: distributed fine-tuning without communication. arXiv preprint arXiv:2210.11948, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.893, + 0.33 + ], + "angle": 0, + "content": "[58] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.893, + 0.416 + ], + "angle": 0, + "content": "[59] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.416, + 0.893, + 0.47 + ], + "angle": 0, + "content": "[60] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34:14306-14318, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.471, + 0.893, + 0.541 + ], + "angle": 0, + "content": "[61] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.541, + 0.894, + 0.613 + ], + "angle": 0, + "content": "[62] Kai Zhu, Wei Zhai, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Self-sustaining representation expansion for nonexemplar class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9296-9305, 2022. 3, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18686" + } + ] +] \ No newline at end of file diff --git a/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/1ae449c1-93a5-4652-bb27-6a83c99dd85c_origin.pdf b/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/1ae449c1-93a5-4652-bb27-6a83c99dd85c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..90ddbb53aab9451caa90e96632f8559240b7092e --- /dev/null +++ b/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/1ae449c1-93a5-4652-bb27-6a83c99dd85c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90b9266a6fb896946416c748d0834a451546562ec7e11f83b91215785bc708f9 +size 671017 diff --git a/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/full.md b/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c26445656b09545dcc04f3613c6c5aedcdb30bec --- /dev/null +++ b/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/full.md @@ -0,0 +1,337 @@ +# Tangent Model Composition for Ensembling and Continual Fine-tuning + +Tian Yu Liu + +University of California, Los Angeles + +tianyu@cs.ucla.edu + +Stefano Soatto + +University of California, Los Angeles + +soatto@ucla.edu + +# Abstract + +Tangent Model Composition (TMC) is a method to combine component models independently fine-tuned around a pre-trained point. Component models are tangent vectors to the pre-trained model that can be added, scaled, or subtracted to support incremental learning, assembling, or unlearning. Component models are composed at inference time via scalar combination, reducing the cost of assembling to that of a single model. TMC improves accuracy by $4.2\%$ compared to assembling non-linearly fine-tuned models at a $2.5 \times 10 \times$ reduction of inference cost, growing linearly with the number of component models. Each component model can be forgotten at zero cost, with no residual effect on the resulting inference. When used for continual fine-tuning, TMC is not constrained by sequential bias and can be executed in parallel on federated data. TMC outperforms recently published continual fine-tuning methods almost uniformly on each setting - task-incremental, class-incremental, and data-incremental - on a total of 13 experiments across 3 benchmark datasets, despite not using any replay buffer. TMC is designed for composing models that are local to a pre-trained embedding, but could be extended to more general settings. The code is available at: https://github.com/tianyu139/tangent-model-composition + +# 1. Introduction + +Compositionality has long been considered central to cognition [5], possibly a reflection of compositionality in neural activity [50]. After decades of unsuccessful attempts to design compositional representations, deep neural networks trained on language data are beginning to exhibit emergent compositionality at scale, by capturing the compositional structure of the data. These models are now being used for visual inference, which has rekindled interest in the study of compositionality of visual representations. But beyond the data, any compositional structure that may be latent in the model is not directly accessible: One cannot simply "compose" weights or inner activations and expect + +meaningful outcomes. The computational architecture of Transformers [52] has been leveraged extensively to co-opt the compositional structure of data through prompts or tokens [29, 55], but still the activations of trained models do not appear to be meaningfully composable. Compositionality of neural activity would allow one to combine activations from different models to capture novel concepts, or incorporate knowledge from different data without having to re-train or fine-tune the core models. This would enable open-universe classification and, more generally, combinatorial expansion of the hypothesis space. Continual learning could be performed simply by composing models trained on different data. + +In this paper, we explore the simplest form of compositionality, that is linear combination. We leverage recent results on the linearization of deep neural networks around a pre-trained point, that can be trained by solving a convex optimization problem and yet perform on-par with nonlinear fine-tuning [1]. This suggests that the tangent space at pre-trained models may be used to linearly compose neural activations, or equivalently compose different models trained or fine-tuned on different datasets and/or for different tasks. + +We show that different models obtained from the linearization around a pre-trained point can be composed, combined, rescaled, and forgotten ("unlearning"), simply by scalar combinations. This fact can be used to perform ensembling at the inference cost of a single model (Sec. 3). It can also be used for continual fine-tuning, with each component model trained independently and in parallel, if necessary on federated data that can therefore be easily forgotten if needed. + +Linear combination is not viable for general concept composition nor arbitrary multi-task learning. In Sec. 5, we discuss limitations of our approach, which can only meaningfully compose models that are "local" to a pre-trained embedding. If there are models trained on tasks that are far in representation space, or even antagonistic, they are likely to live on different tangent planes, making linear combination inappropriate. One such example, described in the appendix, concerns models trained on real images (Ima + +![](images/739c50f870b4986abb37e88572be03dd9e7c1b92a94c1631a7ee77c83ffb2bd9.jpg) +Figure 1. Composition of Fine-tuned Models: Purple indicates final models used for inference, yellow indicates specialist models fine-tuned on individual tasks. The paragon (a) is a model trained jointly on all tasks. However, in continual learning, different tasks are instantiated at different times, an re-training on their union is impractical. Ensembling (b) combines the output of different models trained separately on each task. Weight averaging (c) yields a "model soup" of fine-tuned non-linear models, which improves inference time; Tanget Model Composition (d) linearly composes models fine-tuned on the tangent space of a pre-trained model. + +geNet) and clip art or sketches (DomainNet). Nonetheless, in more common settings, TMC is competitive with general forms of ensembling such as averaging activations or logits of non-linearly fine-tuned models, and with more general forms of continual learning, including methods that use a replay buffer. For broader task coverage, one could use TMC to construct a tangent bundle of models around different pre-trained points, akin to a "tangent model zoo." This concept is not too dissimilar from the architecture of some Foundation Models with a shared backbone and multiple distinct "heavy heads" [54], but far more compact, modular, efficient, interpretable, and easy to work with. + +In the next section we place our contributions in the context of prior art, and in the following one we describe our method in more detail. Sec. 4 summarizes empirical evidence in support of our approach, and finally Sec. 5 discusses its main limitations. + +# 2. Related work + +Compositionality: Compositionality has been studied for decades as a means to expand the representative power of trained models, but has received increasing impetus in recent years thanks to large Transformer-based models that can be used with adaptable prompts. Prompt-tuning [29] has become commonplace to adapt large pre-trained models for specific downstream tasks. Composition of prompts has also been explored in [56, 55] for continual learning. More + +general composition of parameters of trained deep networks has been explored for various purposes such as improving optimization and generalization, as we describe next. + +Deep network linearization: Deep networks are linearized using the first-order Taylor expansion around a pretrained weight [1, 18, 35, 49, 32]. [1] finetunes linearized networks modified to use Leaky-ReLU with gradient preconditioning to achieve similar performances to non-linear fine-tuning. We note that TMC is not just a matter of applying linearized networks to different datasets and averaging, since different scales would lead to imbalance as we describe in Sec. 3.3. Furthermore, unlike [1, 49], we do not require Leaky-ReLU nor gradient pre-conditioning. + +Averaging model weights: Weight averaging has been used in deep learning to improve generalization [25, 17], improve pre-training [12, 34], perform distributed finetuning [57], and increase robustness against distribution shift [59]. [25] averages points along the trajectory of SGD to find flatter minima, [53] averages "late-phase weights" obtained from the later stages of SGD. [24] uses weight interpolation for model patching, and [58] averages weights of models fine-tuned using different hyperparameters from pre-trained models to improve generalization. + +While weight averaging techniques for deep networks are related to ensembles [14, 19], in our case the two coincide since we operate in representation spaces that are linear in the parameters. We leverage this property to train on + +disjoint subsets of data, and show that they can outperform existing weight averaging and ensembling techniques. + +Continual learning/fine-tuning: Continual learning [37] aims to adapt models to new tasks and data without degrading performance in previously trained tasks. Continual fine-tuning [49, 6] additionally requires minimal forgetting of a pre-training objective to ensure all tasks can benefit from it. Existing approaches can be broadly classified into parameter isolation, regularization, architecture, and replay-based. + +Parameter isolation methods [32, 3, 47] operate in the network architecture space by allocating each task a different set of parameters, using techniques such as network pruning [33] to reduce additional memory requirements. + +Regularization-based methods [30, 2, 26, 36, 28] incorporate additional terms into the training objective to prevent catastrophic forgetting. [26] penalizes the distance between weights of the previous and current task, and [45] penalizes changes in weights using Hessian approximations. + +Experience replay methods [44, 4, 32, 11, 43, 49, 7, 8, 26, 56, 6, 10, 9] assume the availability of a memory buffer to store samples from previous tasks. [49, 6] applies experience replay to continual fine-tuning. [49] shows that linearized models combined with replay and regularization alleviates catastrophic forgetting when fine-tuned, and [6] combines dark experience replay and an attention-based regularization loss to prevent catastrophic forgetting of the pre-training features, by encouraging similarity of features at each layer to the original pre-trained network. While replay methods have achieved state-of-the-art performance compared to non-replay ones, they require a sufficiently large memory buffer to store previously seen exemplars. + +A further line of work tackles exemplar-free class incremental learning (EFCIL) using regularization [60], distillation [22], and class-prototypes [61, 62, 40] to incrementally integrate new classes into a fixed-sized network without the memory costs of a replay buffer. + +We show that composition of tangent models trained on individual tasks is a viable method for replay-free continual fine-tuning, and can even outperform recently published replay-based methods. As a side benefit, our method also yields fully parallel training across tasks, and enables zero-cost unlearning. + +# 3. Method + +In continual learning, models are trained on a sequence of tasks indexed by $t = 1,\dots ,T$ ,with the goal of performing well on all, under three main scenarios [51]: (1) Task-incremental, where the task identity $t$ (Task-ID) is known during inference, (2) Data/Domain-incremental, where Task-ID is not provided during inference, and (3) the most challenging Class-incremental, where Task-ID is unknown during inference and in addition the classes repre + +Table 1. A Comparison of Composition Methods: On average across 25 experiments, Tangent Model Ensembling (TME) improves accuracy by $5.1\%$ over standard ensembling of non-linear fine-tuned models at a $2\times$ increase of inference cost. Tangent Model Composition (TMC) also improved accuracy, by $4.2\%$ , while reducing inference cost by $2.5\times$ to $10\times$ in our experiments, and growing linearly with the number of models in the ensemble. Best results for single and multi model inference are indicated in bold and italic respectively. Joint training paragon accuracy (Original Model/Tangent Model): Caltech-256 - $89.17\% /86.77\%$ , MIT-67 - $77.29\% /77.74\%$ , OxfordPets - $92.68\% /93.68\%$ . + +
DatasetTasksSoupEns-LEns-SMTMCTME
Inference/MemoryCost:O(1)O(T)O(T)O(1)O(T)
C-Caltech-256 (Class-IL)584.0082.8083.0984.8285.53
1080.1076.8579.7682.3782.58
2074.8666.7675.4778.6679.30
C-MIT-67 (Class-IL)559.4856.7958.6869.4371.72
1049.3042.3650.5264.5365.72
2028.5123.6135.6248.9854.93
C-OxfordPnts (Class-IL)580.7181.1481.8684.4285.04
1071.3569.7177.3177.7579.97
D-Caltech-256 (Data-IL)585.9686.2786.5984.8685.13
1083.4884.0584.8083.6484.16
2079.6480.7283.1781.9982.94
D-MIT-67 (Data-IL)566.0066.5467.2174.8875.45
1058.8859.2060.4772.4173.41
2050.1051.9955.2070.2272.11
D-OxfordPnts (Data-IL)590.2290.8391.0293.1793.21
1087.6688.4288.8892.4892.72
2085.5785.9786.5891.2891.77
C-Caltech-256 (Task-IL)594.0194.7294.8494.3394.76
1095.0095.8796.0496.0396.23
2095.5696.5396.8497.2297.38
C-MIT-67 (Task-IL)583.3686.2486.3791.2791.79
1087.1990.4290.9794.4595.03
2088.4193.7393.8897.7198.09
C-OxfordPnts (Task-IL)597.2797.2797.3598.5798.74
1097.5997.9297.9899.0599.13
+ +sented in each task are disjoint. + +We assume the setting of continual fine-tuning of a pretrained model. Naive fine-tuning causes catastrophic forgetting of prior tasks, including the one on which pre-training occurred, as the number of tasks increases [6]. Thus, only early tasks benefit from pre-training. Therefore, it is important for continual fine-tuning to ensure that all tasks can benefit from information acquired during pre-training. + +We introduce a unified method to tackle the main scenarios (1)-(3) by composing separately-trained tangent component models. We show that each task takes full advantage of pre-training, and catastrophic forgetting of prior tasks is averted since information from each previous task is encoded separately and only fused for inference. + +# 3.1. Model Linearization + +The effectiveness of model aggregation in weight space hinges on whether the models lie in within the same small-loss training basin of convergence. It is known that weights of different models fine-tuned on the same dataset can be connected by linear paths with constant loss ("linear mode connectivity") [16], but this does not generally occur for models fine-tuned on disjoint tasks/datasets, which is the case of interest to us. Therefore, we convexify the loss landscape locally by approximating the function represented by a deep (non-linear) network with its first-order Taylor approximation around the pre-trained weights. For any deep network $f$ and initial weights $w \in \mathbb{R}^m$ , we denote with $\mathcal{H}(w)$ the set of all such models tangent to $f_w$ given by: + +$$ +\mathcal {H} (w) = \left\{h _ {\delta} (\cdot) \triangleq f _ {w} (\cdot) + \nabla_ {w} f _ {w} (\cdot) \cdot \delta \right\}. \tag {1} +$$ + +Since the tangent model $h_{\delta}$ is linear in the parameters $\delta$ , training it with standard losses such as mean-squared error (MSE) or empirical cross-entropy yields a convex loss landscape. In order to compose models, leveraging Jensen's inequality, we are guaranteed that for any dataset $\mathcal{D}$ : + +$$ +L \left(\mathcal {D}, \sum_ {i} ^ {T} h _ {\theta_ {i} \delta_ {i}}\right) \leq \sum_ {i} \theta_ {i} L \left(\mathcal {D}, h _ {\delta_ {i}}\right), \quad \sum_ {i} \theta_ {i} = 1. \tag {2} +$$ + +In other words, the loss resulting from a convex combination of tangent models in weight space will always be lower than or equal to the convex combination of their individual losses. This is particularly useful when combining models trained to some constant (possibly zero) loss, since it ensures that the composed model will have loss lower or equal to the loss of the component models. + +While tangent models might seem computationally expensive to train, we note that the Jacobian vector product $\nabla_w f_w(\cdot) \cdot \delta$ can be computed in a single forward pass as shown by [39, 35, 1]. This reduces the computational cost to that of a forward pass through the original network. Also note that while the tangent model has twice the degrees of freedom of the original model, fixing the pre-trained model at $w$ results in only a convex optimization in the remaining parameters, as many as the original model. + +In the context of fine-tuning, $w$ is initialized by pretraining on large datasets such as ImageNet [13]. In the rest of this paper, we use the short-form notation $\mathcal{H}$ and $h$ for $\mathcal{H}(w)$ and $h_{\delta}$ respectively. + +# 3.2. Tangent Model Composition for Replay-free Continual Fine-tuning + +Ensembling is an effective strategy for combining weak learners into a strong model [14]. For any finite $F \subseteq \mathbb{A}$ , where $\mathbb{A}$ is a set of functions, and $\lambda = \{\lambda_i\}_{i=1}^{|F|}$ a set of + +![](images/ac03a9f1f8388b819c068a7ec9d8cb4a44135cd55a2075a1661e7b9e9c82c359.jpg) +Figure 2. Accuracy plot over models trained on C-Caltech-256, C-MIT-67, and C-Oxford Pets, split into 10 tasks each. Soft-max ensembling (TME) consistently yields the best results compared to logit ensembling (TMC) for each loss function. Composition of component models trained with the Rescaled Square Loss (RSL) outperforms those trained with empirical cross-entropy (CE) or MSE loss, trading off only $1.2\%$ accuracy against TME to reduce inference time and memory requirement from $\mathcal{O}(T)$ to $\mathcal{O}(1)$ . + +weights, the logit ensemble operation is given by + +$$ +\operatorname {E n s e m b l e} (F, \lambda) = \sum_ {i = 1} ^ {| F |} \lambda_ {i} f _ {i} (\cdot) \tag {3} +$$ + +In the context of continual learning, a simple yet effective approach is to simply train a new model on each task $t$ , and ensemble the outputs at inference time. Unfortunately, this has the obvious shortcoming that both storage requirements and inference time grows along with the number of tasks. + +However, consider the case where $\mathbb{A} = \mathcal{H}$ . Then for any finite $H \subseteq \mathcal{H}$ and associated weights $\lambda$ , we can write + +$$ +\begin{array}{l} \operatorname {E n s e m b l e} (H, \lambda) = \sum_ {i = 1} ^ {| H |} \lambda_ {i} h _ {i} (\cdot) (4) \\ = f _ {w} (\cdot) + \sum_ {i = 1} ^ {| H |} \nabla_ {w} f _ {w} (\cdot) \cdot \lambda_ {i} \delta_ {i} (5) \\ = h _ {\sum \lambda_ {i} \delta_ {i}} (\cdot) (6) \\ \end{array} +$$ + +In other words, since $\mathcal{H}$ defines a vector space with addition and scalar multiplication based on $\delta$ , any linear combination (or ensemble) of models in $\mathcal{H}$ is equivalent to a single model in $\mathcal{H}$ derived simply by the linear combination over $\delta$ . This reduces the inference time for any ensemble of $T$ tangent models from $\mathcal{O}(T)$ to $\mathcal{O}(1)$ . + +In continual learning, it is also important for storage space to not scale with $T$ , since the number of tasks can be arbitrarily large. However, since tasks arrive sequentially in the continual learning framework, we can further reduce + +Table 2. Class-Incremental Learning: Comparison against existing methods using an ImageNet pre-trained ResNet-50. TMC is almost uniformly better than all replay-based and replay-free methods tested, outperforming the best replay-free (RF) method by $14.48\%$ and even the best replay-based method [6] by $1.53\%$ . + +
Method / BufferTasksC-Caltech-256C-MIT-67
5102051020
ER [46] / 30070.14±0.2265.75±0.9265.490.6054.61±0.4149.57±1.4645.331.73
LWF [30] / None34.55±0.1822.76±0.1714.37±0.1425.14±0.5713.30±0.496.41±0.17
oEWC [48] / None28.71±1.3116.32±4.3715.59±3.8314.18±0.777.12±0.494.33±0.16
DER++ [7] / 30080.50±0.2677.84±1.0274.46±0.2164.56±0.5559.73 ±1.0250.28±1.27
Co2L [10] / 30020.71±2.8917.13±2.5314.620.6119.64±12.9321.21±0.5017.78±1.33
ER-ACE [9] / 30078.47±1.2573.85±0.5370.87±1.3864.94±0.5060.25±0.2652.48±0.29
TWF (RF) [6] / None77.95±0.5270.67±0.7262.79±0.8256.58±0.6245.03 ±0.7828.91±0.92
TWF [6] / 30081.34±0.9177.03±0.5372.44±0.4268.33±0.4664.21 ±0.6556.29±0.68
TMC / None84.82 ±0.0382.37 ±0.0878.66 ±0.1369.43±0.2164.53±0.2848.98 ±0.13
+ +the memory capacity required for storage of models from $\mathcal{O}(T)$ , to $\mathcal{O}(1)$ by using a simple autoregressive model: We train on the first task $t_1$ to produce model $\tilde{h}_1 = h_1$ . Then, for any subsequent task $t > 1$ , we train a new model $\tilde{h}_t$ , and compose it with $h_{t-1}$ to give $h_t = \lambda_{(t,1)}\tilde{h}_t + \lambda_{(t,2)}h_{t-1} := h_{\delta'}$ where $\delta' = \frac{\lambda_{(t,1)}\tilde{\delta}_t + \lambda_{(t,2)}\delta_{t-1}}{\lambda_{(t,1)} + \lambda_{(t,2)}}$ . Since we only require model $h_{t-1}$ to produce $h_t$ , all prior models up to and including $h_{t-1}$ can be discarded upon constructing $h_t$ , hence reducing memory requirement from $\mathcal{O}(T)$ to $\mathcal{O}(1)$ , equal to the size of a single model. In contrast to methods that expands the model at each task, or require a memory buffer of examples from previous tasks, this method does not incur additional storage requirements, nor assume availability of an extra replay buffer for storing past exemplars (i.e. our method is replay-free). + +Remark 1 (Free Parallelization): Additionally, for any given convex loss function $L$ , the optimization objective + +$$ +\min _ {\delta \in \mathbb {R} ^ {m}} \sum_ {(x, y) \in \mathcal {D}} L \left(h _ {\delta} (x), y\right) \tag {7} +$$ + +is convex and thus converges to a global optimum. This implies that each model $\tilde{h}_t$ can be trained from a standard initialization as opposed to from $h_{t-1}$ . Hence, if multiple tasks arrive at the same time, such as when training in a federated manner, training can be parallelized across tasks and composed in a zero-cost, zero-shot manner to yield equivalent results (assuming equivalence of global optima) to sequential training and composition. To the best of our knowledge, apart from [41] which simply stores samples in a replay buffer and trains on them at test time, no other work in continual learning can be parallelized across tasks. + +Remark 2 (Free Forgetting/Unlearning): If we relax memory constraints to store models from each task, unlearning any specific task $t$ (e.g., private data associated with specific samples) is straightforward and can be done in a zero-cost, zero-shot manner. Since subtraction is well-defined in the vector space $\mathcal{H}$ , any task $i$ can be unlearned + +from the model $h_T$ trained on $T$ tasks by simply subtracting the weights associated to $\tilde{h}_i$ : + +$$ +h _ {T \backslash i} = h _ {T} - \lambda_ {i, 1} \prod_ {j = i + 1} ^ {T} \lambda_ {(j, 2)} \tilde {h} _ {i} \tag {8} +$$ + +Remark 3 (Unified Continual Fine-tuning): Our method does not assume any specific task splits, and is hence generalizable across all three forms of continual fine-tuning: task-incremental, class-incremental, and data-incremental. + +# 3.3. Scale Standardization + +Training with the cross-entropy loss does not ensure that the weights are comparable in scale. We can study the implications of this for composition on $\mathcal{H}$ through the lens of logits ensembling. While logits ensembling has been shown to be effective [58], Fig. 2 shows that ensembling the normalized soft-max outputs (Tangent Model Ensembling, TME) is uniformly better than logit ensembling/weight averaging via TMC across various loss functions. However, TME cannot be represented by a linear combination of perturbations $\delta_{i}$ and hence incurs $\mathcal{O}(T)$ model storage costs and $\mathcal{O}(T)$ inference time. + +Instead, we propose to induce a waterbed effect similar to soft-max ensembling using standardization. In particular, we train using the Rescaled Square Loss (RSL) [23]: + +$$ +L _ {s} (x, y) = \frac {1}{K} \left(\alpha ([ f (x) ] _ {y} - \beta) ^ {2} + \sum_ {i = 1, i \neq y} ^ {K} ([ f (x) ] _ {i}) ^ {2}\right) \tag {9} +$$ + +where $\alpha, \beta$ are hyper-parameters. This is the regular MSE loss when $\alpha = \beta = 1$ . Larger values of $\beta$ reduce soft-max entropy in the output by scaling positive output class signals while keeping negative ones close to zero. + +We use $\beta$ to control the interaction between component models. For example, when tasks are highly dissimilar, + +Table 3. Data-Incremental Learning on D-MIT-67, 4 tasks using ResNet-18. For fair comparison, we run our method under the settings proposed by [49]. TMC outperforms the best replay-free method [30] by $4.61\%$ , and the best replay-based method [49] by $1.27\%$ , while supporting parallel training across tasks (PL). If we sacrifice parallelization by initializing each model at task $t$ with $h_{t-1}$ (TMC-Seq), we further improve over TMC by $0.28\%$ . + +
MethodPL?Replay BufferAccuracy
SGDXNone63.48
LwF[30]XNone67.21
MAS[2]XNone62.49
OSLA[45]XNone64.08
EWC [26]X50063.68
DLCFT[49]X50070.55
TMC✓None71.82
TMC-SeqXNone72.10
Joint (Linear)--73.61
Joint (Paragon)--74.40
+ +such as in task- and class-incremental continual learning, we use larger values of $\beta$ to reduce interference of component models trained on other tasks upon composition. On the other hand, when tasks are highly synergistic such as in data-incremental learning, lower values of $\beta$ can be used to encourage interaction in the final composed model. + +# 4. Experiments + +We describe comparison baselines in Sec. 4.1, and our main results in Sec. 4.2, with additional studies in Sec. 4.3. + +Our method is replay-free. As such, its computational cost is at most half of what it would be if it used a replay buffer since replay-based methods typically sample separate batches of equal size from both the current task and replay buffer at each iteration [7, 6, 10, 46, 9]. Nonetheless, we compare our method to both replay-based and replay-free methods in our baselines. While replay-based methods have thus far outperformed replay-free ones, ours is uniformly more accurate than replay-based methods tested, in addition to replay-free ones. + +We evaluate our method on Caltech-256 [20], MIT-67 [42], and OxfordPets [38]. For continual fine-tuning, we further split each training dataset into multiple disjoint subsets, each representing a single task, for fair comparison across experiments. We prefix sequential datasets split into tasks with disjoint labels (for task- and class-incremental learning) with C-, and random splits using a standardized random seed (for data-incremental learning) with D-. We set $\alpha = 1, \beta = 25$ for the former, $\alpha = 1, \beta = 5$ for the latter, and $\alpha = 1, \beta = 5$ for TME. Results in all tables and figures are averaged across 3 runs. + +![](images/0ef32265e31f3fe4ed170d117da0232be78600cf938bfe4440a5fca75766d9e3.jpg) + +![](images/273a8be4dd0983e861b7db82430be3541c7530047c70227af59d60cdc76eafcd.jpg) +Figure 3. Ablation on $\beta$ averaged across (Top:) C-Caltech-256, C-MIT-67, and C-OxfordPets and (Bottom:) D-Caltech-256, D-MIT-67, and D-OxfordPets, split into 10 disjoint tasks. For class incremental datasets, larger values of $\beta$ result in better generalization of the final composed model. For data-incremental datasets where tasks are synergistic, larger values of $\beta$ instead harm the accuracy of both the component models and the composed model. + +# 4.1. Composition Baselines + +In each of the following sections, we compare against the following baselines for model aggregation: + +Logits Ensemble (Ens-L): Averaging the output logits of non-linear component models. + +Softmax Ensemble (Ens-SM): Averaging the softmax output of non-linear component models. + +Weight composition of non-linear models (Soup): Averaging the weights of non-linear component models. This method is inspired by Model Soups [58] where weights of various non-linear models trained on the same task but with different hyper-parameters and augmentations are averaged to improve generalization at no additional inference cost. + +# 4.2. Main Results + +We show that TMC outperforms existing methods in all three continual fine-tuning settings [51], from hardest to easiest: Class-Incremental (Sec. 4.2.1), Data-Incremental (Sec. 4.2.2), and Task-Incremental (Sec. 4.2.3). + +Table 4. Task-Incremental Learning: Comparison against existing methods using ImageNet pre-trained ResNet-50. We almost uniformly outperform all replay-based and replay-free methods that we compare against. In particular, we improve over the best replay-free (RF) method by $2.46\%$ , and the best replay-based method by $1.38\%$ [6]. + +
Method / BufferTasksC-Caltech-256C-MIT-67
5102051020
ER [46] / 30092.11±0.2694.33±0.4895.76±0.1385.84±0.0990.59±0.5195.13±0.27
LWF [30] / None93.73±0.1295.31±0.0896.29±0.0587.55±0.2090.82±0.0495.68±0.58
oEWC [48] / None83.97±0.7176.20±16.0791.49±0.3067.31±3.2268.45±4.9384.74±1.96
DER++ [7] / 30093.79±0.1394.92±0.1195.87±0.1786.83±0.2691.63±0.7594.54±0.62
Co2L [10] / 30034.51±3.0539.62±1.7045.94±1.4739.32±20.4764.40±1.3151.50±2.21
ER-ACE [9] / 30093.11±0.1294.63±0.0695.86±0.1087.13±0.1690.96±0.2894.92±0.60
TWF (RF) [6] / None94.45±0.0595.64±0.0696.36±0.1286.76±0.2588.27±0.5494.74±0.35
TWF [6] / 30094.38±0.0895.74±0.1796.59±0.0788.22±0.4192.13±0.2695.67±0.46
TMC / None94.33±0.0396.03±0.0297.22±0.0391.27±0.2194.45±0.1397.71±0.13
+ +# 4.2.1 Class-Incremental Learning + +In class-incremental learning, datasets are partitioned so that label spaces of each task are disjoint. + +Comparison against Composition Baselines: We compare against model composition baselines in Table 1. While TME yields the best generalization performance, this comes at a cost of $\mathcal{O}(T)$ inference time. On the other hand, TMC achieves an $\mathcal{O}(1)$ inference time with only a $1.73\%$ trade-off in generalization accuracy on average, while outperforming weight averaging, logit ensembling, and even soft-max ensembling of non-linear models by an average of $7.83\%$ , $11.37\%$ , and $6.08\%$ respectively, even though the latter two both require $\mathcal{O}(T)$ inference time. + +Comparison against existing methods: In Table 2, we compare against recent continual fine-tuning methods including both replay-free and replay-based methods. Here, replay-free methods [30, 48] perform poorly in the class-incremental setting relative to stronger replay-based methods [46, 9, 7, 6]. However, we show that not only do we improve over the best replay-free method by $14.48\%$ , but also over the best replay-based method by $1.53\%$ [6], corresponding to a $17.59\%$ relative error reduction towards the paragon performance of joint training (Table 1). + +Comparison against EFCIL methods: In Table 5, we compare against EFCIL methods on C-TinyImageNet [27]. The original EFCIL methods are evaluated under a different setting (denoted $\dagger$ ), where the zeroth task contains $50\%$ of the dataset. Hence, we further modified two of the best performing EFCIL methods, SSRE [62] and FeTrIL [40] to use the pre-trained ImageNet initialization, and run them under our setting of uniformly sharded data. TMC not only outperforms all methods ran under our harder setting, by an average of $15.3\%$ over the next best method [40], but even yields equal or better performance when compared to all methods ran under $\dagger$ . + +# 4.2.2 Data-Incremental Learning + +Also referred to as domain-incremental learning, this setting assumes that tasks are split in a non-stratified manner, where each task has the same output space [51, 49]. + +Comparison against Composition Baselines: In Table 1, we compare against model composition baselines. Similar to the class-incremental setting, TME outperforms TMC by an average of $0.66\%$ at a cost of $\mathcal{O}(T)$ inference time. TMC uses $\mathcal{O}(1)$ inference time and outperforms non-linear composition, logit ensemble, and even soft-max ensemble by an average of $7.04\%$ , $6.32\%$ , and $5.21\%$ respectively. + +Comparison against existing methods: In Table 3, we benchmark our model against results obtained by [49] on a pre-trained ResNet-18 model using the D-MIT-67 dataset split into 4 disjoint subsets. We improve over the best replay-free method [30] by $4.61\%$ , and the best replay-based method [49] by $1.27\%$ , corresponding to a $64\%$ and $33\%$ relative error reduction towards the joint training paragon respectively. TMC-Seq, which uses sequential initialization by initializing each component model $h_t$ with $h_{t - 1}$ instead of $h_0$ , further improves over TMC by $0.28\%$ but at the cost of parallelizability. + +# 4.2.3 Task-Incremental Learning + +In task-incremental learning, the output spaces of different tasks are disjoint, but the source task of each sample is known at inference time. Thus, we can restrict the model predictions to that of the given task during evaluation. + +Comparison against Composition Baselines: Due to the disjoint output spaces of each task, ensembling outputs of specialist component models trained on individual tasks and restricting them to a target task essentially only considers the predictions of the specialist model trained on the target task (since, in general, specialist models are not effective discriminants for tasks other than those they are triend on). Hence, the independence of output spaces renders en + +Table 5. Comparison against EFCIL methods using either ImageNet pre-trained ResNet-50, or under the original EFCIL setting where the zeroth task contains $50\%$ of the dataset (denoted $\dagger$ ). All experiments are ran on C-TinyImageNet. Italics denote the best method under $\dagger$ , and bold denotes the best method under our harder setting. Under the former, TMC performs equally or better than all methods. Under the latter, TMC outperforms the best method by an average of $15.3\%$ . + +
TasksPASS† [61]IL2A† [60]SSRE† [62]FeTrIL† [40]SSRE[62]FeTrIL[40]TMC
549.647.350.454.833.744.157.6
1047.344.748.953.125.636.153.1
+ +semblies of specialist models highly effective. In contrast, composing non-linear models in weight space does not ensure minimal interference across different tasks. Indeed, we show in Table 1 that TME yields the best results, while the non-linear composition approach yields the worst in task-incremental learning. On the other hand, due to the equivalence of TMC to (logit) ensembling of linearized models, TMC outperforms not only weight averaging, but also logit and soft-max ensembles of non-linear models, by $3.78\%$ , $1.83\%$ , and $1.79\%$ respectively. + +Comparison against existing methods: Since each component model is trained only on task $t$ , by design, there is no interference from other tasks when training $\tilde{h}_t$ . As such, the final composed model is equivalent to a specialist model when restricted to any task. However, in other state of the art continual learning methods such as those involving replay or loss regularization, this separation among trained for each task is violated. Hence, under the task-incremental setting, we show in Table 4 that TMC achieves the best performance across both replay-free and replay-based methods even without using any memory buffer, beating the next best method by $2.46\%$ and $1.38\%$ respectively [6]. + +# 4.3. Further studies + +We present additional studies on the Rescaled Square Loss (Sec. 4.3.1), and compare against TMC-Seq to show that our method yields an effective and highly parallelizable training scheme for continual fine-tuning (Sec. 4.3.2). + +# 4.3.1 Rescaled Square Loss + +In Sec. 3.3, we showed that $\beta$ can be used to control the amount of interaction between tasks. In Fig. 3, we empirically show that in the class-incremental setting where tasks are highly dissimilar, larger values of $\beta$ result in a better generalist model. On the other hand, under the data-incremental setting where tasks are more similar, information from other tasks can benefit the learning of any given task. Here, larger values of $\beta$ instead harm generalization due to lowering the accuracy of each component model. + +# 4.3.2 Convexity and Parallelizability + +Given a sequence of tasks $1 \ldots T$ and a pre-trained model $h_0$ , there are at least two natural ways to learn the compo + +![](images/f30f3019419ce173f201bee14ae28f270789489cb3996979412221091f617ea3.jpg) +Figure 4. Accuracy of composed models, where component models are initialized with $h_0$ (TMC) and $h_{t-1}$ (TMC-Seq). In the data-incremental setting (D), TMC-Seq outperforms TMC, while in the class-incremental setting (C), TMC performs significantly better. All results are averaged across D/C-Caltech-256, D/C-MIT-67, D/C-OxfordPets (we exclude C-OxfordPets in the 20-task setting since it does not contain enough classes). + +nent model $\tilde{h}_t$ : (1) Train $\tilde{h}_t$ with $h_{t-1}$ as initialization, and (2) Train $\tilde{h}_t$ with $h_0$ as initialization. The first option (TMC-Seq) is most common among continual learning methods which impose regularization in the weights between the current and previous tasks or optimize using memory buffers consisting of examples obtained from previous tasks. As such, these methods are restricted to a sequential learning framework, and hence cannot be parallelized across tasks. In Sec. 3, we showed that, thanks to the convexity of the loss landscape, both options (1) and (2) are equivalent under our framework when we assume comparable generalization performance among global optima. + +In practice, for massively over-parameterized networks, such assumptions do not hold. However, we show empirically in Fig. 4 that under the data-incremental setting, compared to TMC-Seq, TMC only incurs a small loss in generalization performance to enable massive parallelization across tasks. This can potentially provide orders of magnitude speed-ups compared to existing continual learning methods. Moreover, under the class-incremental setting, TMC greatly outperforms TMC-Seq, showing that the pre-trained initialization yields solutions of better generalization than $h_{t - 1}$ . We attribute this to the fact that the hypothesis space of models which are globally optimal for the training data is vast and yields varying generalization. + +Thus, regardless of the loss landscape convexity, convergence to each point in this space depends on the "direction of approach". Hence, initialization at $h_{t-1}$ introduces a bias towards the previous tasks. Due to the high level of dissimilarity between tasks in the class-incremental setting, this bias harms the final model generalization performance. + +# 5. Discussion + +Our approach enables incremental learning under the assumption that increment tasks are "close" to the pre-trained model. We ensure proximity by training increment models independently on the tangent plane of the pre-trained model. The advantage is that pre-trained tasks are not forgotten, and increment tasks can be learned independently, in parallel, and easily forgotten if needed. + +Of course, this approach does not address the problem of incremental learning in full generality, when increment tasks can be arbitrary and arbitrarily unrelated. Nonetheless, we have shown that our method is competitive with existing methods in some of the most challenging settings, for instance when the hypothesis spaces of component tasks are disjoint. On the positive side, despite not requiring a replay buffer, our method outperforms replay-based ones under the assumptions in which we operate. + +A limitation of our approach is that it requires computing the linear span of pre-trained models, which can be challenging. However, for common deep network architectures such as the ResNet [21] or Transformer [52, 15] family, the Jacobian-Vector product can be efficiently evaluated [39, 35, 1, 31], reducing the computational complexity of training the tangent model to that of the same magnitude as a linear classifier on network activations. Hence, inference time for the linearized network is at most double that of the original model. + +Another limitation of our approach is assuming a first-order expansion around a pre-trained objective can well approximate the fine-tuned model. Such assumptions hold poorly when the pre-training and downstream tasks are highly unrelated. Nevertheless, an ImageNet [13] pretrained model already yields a sufficiently good local approximation for various real-world datasets [1, 31]. + +While we mainly explored linear compositionality in the context of continual learning, we note that this framework can be easily generalized. We discussed possible applications towards federated learning and forgetting, since our method naturally yields a parallel training framework that fully compartmentalizes information obtained each task. + +# 6. Acknowledgements + +We would like to thank Aditya Golatkar, Albert Zhao, and the anonymous reviewers for their feedback on the initial version of the paper. This work was supported by ARO + +W911NF-17-1-0304 and ONR N00014-22-1-2252. + +# References + +[1] Alessandro Achille, Aditya Golatkar, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. Lqf: Linear quadratic fine-tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15729-15739, 2021. 1, 2, 4, 9 +[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 3, 6 +[3] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3366-3375, 2017. 3 +[4] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Advances in neural information processing systems, 32, 2019. 3 +[5] Elie Bienenstock and Stuart Geman. Compositionality in neural systems. In The handbook of brain theory and neural networks, pages 223-226. 1998. 1 +[6] Matteo Boschini, Lorenzo Bonicelli, Angelo Porrello, Giovanni Bellitto, Matteo Pennisi, Simone Palazzo, Concetto Spampinato, and Simone Calderara. Transfer without forgetting. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIII, pages 692-709. Springer, 2022. 3, 5, 6, 7, 8 +[7] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 3, 5, 6, 7 +[8] Pietro Buzzega, Matteo Boschini, Angelo Porrello, and Simone Calderara. Rethinking experience replay: a bag of tricks for continual learning. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 2180-2187. IEEE, 2021. 3 +[9] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuyte-laars, Joelle Pineau, and Eugene Belilovsky. New insights on reducing abrupt representation change in online continual learning. arXiv preprint arXiv:2203.03798, 2022. 3, 5, 6, 7 +[10] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. In Proceedings of the IEEE/CVF International conference on computer vision, pages 9516-9525, 2021. 3, 5, 6, 7 +[11] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 3 +[12] Leshem Choshen, Elad Venezian, Noam Slonim, and Yoav Katz. Fusing finetuned models for better pretraining. arXiv preprint arXiv:2204.03044, 2022. 2 +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image + +database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 9 +[14] Thomas G Dietterich. Ensemble methods in machine learning. In International workshop on multiple classifier systems, pages 1-15. Springer, 2000. 2, 4 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 9 +[16] Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lottery ticket hypothesis. In International Conference on Machine Learning, pages 3259-3269. PMLR, 2020. 4 +[17] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. Advances in neural information processing systems, 31, 2018. 2 +[18] Aditya Golatkar, Alessandro Achille, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. Mixed-privacy forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 792–801, 2021. 2 +[19] Raphael Gontijo-Lopes, Yann Dauphin, and Ekin D Cubuk. No one representation to rule them all: Overlapping features of training methods. arXiv preprint arXiv:2110.12899, 2021. 2 +[20] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 6 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 9 +[22] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831-839, 2019. 3 +[23] Like Hui and Mikhail Belkin. Evaluation of neural architectures trained with square loss vs cross-entropy in classification tasks. arXiv preprint arXiv:2006.07322, 2020. 5 +[24] Gabriel Ilharco, Mitchell Wortsman, Samir Yitzhak Gadre, Shuran Song, Hannaneh Hajishirzi, Simon Kornblith, Ali Farhadi, and Ludwig Schmidt. Patching open-vocabulary models by interpolating weights. arXiv preprint arXiv:2208.05592, 2022. 2 +[25] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 2 +[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3, 6 + +[27] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 7 +[28] Janghyeon Lee, Hyeong Gwon Hong, Donggyu Joo, and Junmo Kim. Continual learning with extended kronecker-factored approximate curvature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9001-9010, 2020. 3 +[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 2 +[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 3, 5, 6, 7 +[31] Tian Yu Liu, Aditya Golatkar, and Stefano Soatto. Tangent transformers for composition, privacy and removal. arXiv preprint arXiv:2307.08122, 2023. 9 +[32] Tian Yu Liu, Aditya Golatkar, Stefano Soatto, and Alessandro Achille. Integral continual learning along the tangent vector field of tasks. arXiv preprint arXiv:2211.13108, 2022. 2, 3 +[33] Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7765-7773, 2018. 3 +[34] Michael Matena and Colin Raffel. Merging models with fisher-weighted averaging. arXiv preprint arXiv:2111.09832, 2021. 2 +[35] Fangzhou Mu, Yingyu Liang, and Yin Li. Gradients as features for deep representation learning. arXiv preprint arXiv:2004.05529, 2020. 2, 4, 9 +[36] Cuong V Nguyen, Yingzhen Li, Thang D Bui, and Richard E Turner. Variational continual learning. arXiv preprint arXiv:1710.10628, 2017. 3 +[37] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural Networks, 113:54-71, 2019. 3 +[38] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 6 +[39] Barak A Pearlmutter. Fast exact multiplication by the hessian. Neural computation, 6(1):147-160, 1994. 4, 9 +[40] Grégoire Petit, Adrian Popescu, Hugo Schindler, David Picard, and Bertrand Delezoide. Fetril: Feature translation for exemplar-free class-incremental learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3911–3920, 2023. 3, 7, 8 +[41] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 524-540. Springer, 2020. 5 +[42] Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In 2009 IEEE conference on computer vision and pattern recognition, pages 413-420. IEEE, 2009. 6 + +[43] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 3 +[44] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. arXiv preprint arXiv:1810.11910, 2018. 3 +[45] Hippolyt Ritter, Aleksandar Botev, and David Barber. Online structured laplace approximations for overcoming catastrophic forgetting. Advances in Neural Information Processing Systems, 31, 2018. 3, 6 +[46] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 5, 6, 7 +[47] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 3 +[48] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In International conference on machine learning, pages 4528-4537. PMLR, 2018. 5, 7 +[49] Hyounguk Shon, Janghyeon Lee, Seung Hwan Kim, and Junmo Kim. Dlcft: Deep linear continual fine-tuning for general incremental learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 513-529. Springer, 2022. 2, 3, 6, 7 +[50] Doris Y Tsao, Winrich A Freiwald, Tamara A Knutsen, Joseph B Mandeville, and Roger BH Tootell. Faces and objects in macaque cerebral cortex. Nature neuroscience, 6(9):989-995, 2003. 1 +[51] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. arXiv preprint arXiv:1904.07734, 2019. 3, 6, 7 +[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 9 +[53] Johannes Von Oswald, Seijin Kobayashi, Joao Sacramento, Alexander Meulemans, Christian Henning, and Benjamin F Grewe. Neural networks with late-phase weights. arXiv preprint arXiv:2007.12927, 2020. 2 +[54] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 2 +[55] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. arXiv preprint arXiv:2204.04799, 2022. 1, 2 + +[56] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2, 3 +[57] Mitchell Wortsman, Suchin Gururangan, Shen Li, Ali Farhadi, Ludwig Schmidt, Michael Rabbat, and Ari S Morcos. lo-fi: distributed fine-tuning without communication. arXiv preprint arXiv:2210.11948, 2022. 2 +[58] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 2, 5, 6 +[59] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 2 +[60] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34:14306-14318, 2021. 3, 8 +[61] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 3, 8 +[62] Kai Zhu, Wei Zhai, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Self-sustaining representation expansion for nonexemplar class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9296-9305, 2022. 3, 7, 8 \ No newline at end of file diff --git a/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/images.zip b/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1c4416bcc8e4c296d60143ab15e139f4ddccce72 --- /dev/null +++ b/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a52445312b7904893cdb30f0b16b48cf46384af2e146fae3bbaf56f50426ebde +size 545109 diff --git a/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/layout.json b/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..59f25a857702ac737ee3684e7b964c24b9fed2be --- /dev/null +++ b/2023/Tangent Model Composition for Ensembling and Continual Fine-tuning/layout.json @@ -0,0 +1,10067 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 76, + 103, + 518, + 121 + ], + "type": "text", + "content": "Tangent Model Composition for Ensembling and Continual Fine-tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 144, + 221, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 144, + 221, + 156 + ], + "spans": [ + { + "bbox": [ + 160, + 144, + 221, + 156 + ], + "type": "text", + "content": "Tian Yu Liu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 158, + 282, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 158, + 282, + 171 + ], + "spans": [ + { + "bbox": [ + 99, + 158, + 282, + 171 + ], + "type": "text", + "content": "University of California, Los Angeles" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 174, + 240, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 174, + 240, + 185 + ], + "spans": [ + { + "bbox": [ + 141, + 174, + 240, + 185 + ], + "type": "text", + "content": "tianyu@cs.ucla.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 366, + 144, + 440, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 144, + 440, + 156 + ], + "spans": [ + { + "bbox": [ + 366, + 144, + 440, + 156 + ], + "type": "text", + "content": "Stefano Soatto" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 158, + 494, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 158, + 494, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 158, + 494, + 171 + ], + "type": "text", + "content": "University of California, Los Angeles" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 361, + 174, + 444, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 174, + 444, + 184 + ], + "spans": [ + { + "bbox": [ + 361, + 174, + 444, + 184 + ], + "type": "text", + "content": "soatto@ucla.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "content": "Tangent Model Composition (TMC) is a method to combine component models independently fine-tuned around a pre-trained point. Component models are tangent vectors to the pre-trained model that can be added, scaled, or subtracted to support incremental learning, assembling, or unlearning. Component models are composed at inference time via scalar combination, reducing the cost of assembling to that of a single model. TMC improves accuracy by " + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "content": " compared to assembling non-linearly fine-tuned models at a " + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "inline_equation", + "content": "2.5 \\times 10 \\times" + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "content": " reduction of inference cost, growing linearly with the number of component models. Each component model can be forgotten at zero cost, with no residual effect on the resulting inference. When used for continual fine-tuning, TMC is not constrained by sequential bias and can be executed in parallel on federated data. TMC outperforms recently published continual fine-tuning methods almost uniformly on each setting - task-incremental, class-incremental, and data-incremental - on a total of 13 experiments across 3 benchmark datasets, despite not using any replay buffer. TMC is designed for composing models that are local to a pre-trained embedding, but could be extended to more general settings. The code is available at: https://github.com/tianyu139/tangent-model-composition" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Compositionality has long been considered central to cognition [5], possibly a reflection of compositionality in neural activity [50]. After decades of unsuccessful attempts to design compositional representations, deep neural networks trained on language data are beginning to exhibit emergent compositionality at scale, by capturing the compositional structure of the data. These models are now being used for visual inference, which has rekindled interest in the study of compositionality of visual representations. But beyond the data, any compositional structure that may be latent in the model is not directly accessible: One cannot simply \"compose\" weights or inner activations and expect" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 213, + 547, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 369 + ], + "type": "text", + "content": "meaningful outcomes. The computational architecture of Transformers [52] has been leveraged extensively to co-opt the compositional structure of data through prompts or tokens [29, 55], but still the activations of trained models do not appear to be meaningfully composable. Compositionality of neural activity would allow one to combine activations from different models to capture novel concepts, or incorporate knowledge from different data without having to re-train or fine-tune the core models. This would enable open-universe classification and, more generally, combinatorial expansion of the hypothesis space. Continual learning could be performed simply by composing models trained on different data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 373, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 491 + ], + "type": "text", + "content": "In this paper, we explore the simplest form of compositionality, that is linear combination. We leverage recent results on the linearization of deep neural networks around a pre-trained point, that can be trained by solving a convex optimization problem and yet perform on-par with nonlinear fine-tuning [1]. This suggests that the tangent space at pre-trained models may be used to linearly compose neural activations, or equivalently compose different models trained or fine-tuned on different datasets and/or for different tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 495, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 602 + ], + "type": "text", + "content": "We show that different models obtained from the linearization around a pre-trained point can be composed, combined, rescaled, and forgotten (\"unlearning\"), simply by scalar combinations. This fact can be used to perform ensembling at the inference cost of a single model (Sec. 3). It can also be used for continual fine-tuning, with each component model trained independently and in parallel, if necessary on federated data that can therefore be easily forgotten if needed." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Linear combination is not viable for general concept composition nor arbitrary multi-task learning. In Sec. 5, we discuss limitations of our approach, which can only meaningfully compose models that are \"local\" to a pre-trained embedding. If there are models trained on tasks that are far in representation space, or even antagonistic, they are likely to live on different tangent planes, making linear combination inappropriate. One such example, described in the appendix, concerns models trained on real images (Ima" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18676" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 521, + 300 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 521, + 300 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 521, + 300 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 521, + 300 + ], + "type": "image", + "image_path": "739c50f870b4986abb37e88572be03dd9e7c1b92a94c1631a7ee77c83ffb2bd9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 306, + 547, + 363 + ], + "lines": [ + { + "bbox": [ + 46, + 306, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 547, + 363 + ], + "type": "text", + "content": "Figure 1. Composition of Fine-tuned Models: Purple indicates final models used for inference, yellow indicates specialist models fine-tuned on individual tasks. The paragon (a) is a model trained jointly on all tasks. However, in continual learning, different tasks are instantiated at different times, an re-training on their union is impractical. Ensembling (b) combines the output of different models trained separately on each task. Weight averaging (c) yields a \"model soup\" of fine-tuned non-linear models, which improves inference time; Tanget Model Composition (d) linearly composes models fine-tuned on the tangent space of a pre-trained model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 381, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 525 + ], + "type": "text", + "content": "geNet) and clip art or sketches (DomainNet). Nonetheless, in more common settings, TMC is competitive with general forms of ensembling such as averaging activations or logits of non-linearly fine-tuned models, and with more general forms of continual learning, including methods that use a replay buffer. For broader task coverage, one could use TMC to construct a tangent bundle of models around different pre-trained points, akin to a \"tangent model zoo.\" This concept is not too dissimilar from the architecture of some Foundation Models with a shared backbone and multiple distinct \"heavy heads\" [54], but far more compact, modular, efficient, interpretable, and easy to work with." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 525, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 585 + ], + "type": "text", + "content": "In the next section we place our contributions in the context of prior art, and in the following one we describe our method in more detail. Sec. 4 summarizes empirical evidence in support of our approach, and finally Sec. 5 discusses its main limitations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 597, + 132, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 132, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 132, + 609 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Compositionality: Compositionality has been studied for decades as a means to expand the representative power of trained models, but has received increasing impetus in recent years thanks to large Transformer-based models that can be used with adaptable prompts. Prompt-tuning [29] has become commonplace to adapt large pre-trained models for specific downstream tasks. Composition of prompts has also been explored in [56, 55] for continual learning. More" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 381, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 418 + ], + "type": "text", + "content": "general composition of parameters of trained deep networks has been explored for various purposes such as improving optimization and generalization, as we describe next." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 420, + 546, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 546, + 540 + ], + "type": "text", + "content": "Deep network linearization: Deep networks are linearized using the first-order Taylor expansion around a pretrained weight [1, 18, 35, 49, 32]. [1] finetunes linearized networks modified to use Leaky-ReLU with gradient preconditioning to achieve similar performances to non-linear fine-tuning. We note that TMC is not just a matter of applying linearized networks to different datasets and averaging, since different scales would lead to imbalance as we describe in Sec. 3.3. Furthermore, unlike [1, 49], we do not require Leaky-ReLU nor gradient pre-conditioning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 543, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 663 + ], + "type": "text", + "content": "Averaging model weights: Weight averaging has been used in deep learning to improve generalization [25, 17], improve pre-training [12, 34], perform distributed finetuning [57], and increase robustness against distribution shift [59]. [25] averages points along the trajectory of SGD to find flatter minima, [53] averages \"late-phase weights\" obtained from the later stages of SGD. [24] uses weight interpolation for model patching, and [58] averages weights of models fine-tuned using different hyperparameters from pre-trained models to improve generalization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "While weight averaging techniques for deep networks are related to ensembles [14, 19], in our case the two coincide since we operate in representation spaces that are linear in the parameters. We leverage this property to train on" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18677" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "disjoint subsets of data, and show that they can outperform existing weight averaging and ensembling techniques." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 191 + ], + "type": "text", + "content": "Continual learning/fine-tuning: Continual learning [37] aims to adapt models to new tasks and data without degrading performance in previously trained tasks. Continual fine-tuning [49, 6] additionally requires minimal forgetting of a pre-training objective to ensure all tasks can benefit from it. Existing approaches can be broadly classified into parameter isolation, regularization, architecture, and replay-based." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 192, + 286, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 286, + 239 + ], + "type": "text", + "content": "Parameter isolation methods [32, 3, 47] operate in the network architecture space by allocating each task a different set of parameters, using techniques such as network pruning [33] to reduce additional memory requirements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 239, + 286, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 286, + 299 + ], + "type": "text", + "content": "Regularization-based methods [30, 2, 26, 36, 28] incorporate additional terms into the training objective to prevent catastrophic forgetting. [26] penalizes the distance between weights of the previous and current task, and [45] penalizes changes in weights using Hessian approximations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 286, + 455 + ], + "type": "text", + "content": "Experience replay methods [44, 4, 32, 11, 43, 49, 7, 8, 26, 56, 6, 10, 9] assume the availability of a memory buffer to store samples from previous tasks. [49, 6] applies experience replay to continual fine-tuning. [49] shows that linearized models combined with replay and regularization alleviates catastrophic forgetting when fine-tuned, and [6] combines dark experience replay and an attention-based regularization loss to prevent catastrophic forgetting of the pre-training features, by encouraging similarity of features at each layer to the original pre-trained network. While replay methods have achieved state-of-the-art performance compared to non-replay ones, they require a sufficiently large memory buffer to store previously seen exemplars." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 286, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 286, + 514 + ], + "type": "text", + "content": "A further line of work tackles exemplar-free class incremental learning (EFCIL) using regularization [60], distillation [22], and class-prototypes [61, 62, 40] to incrementally integrate new classes into a fixed-sized network without the memory costs of a replay buffer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 515, + 286, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 286, + 586 + ], + "type": "text", + "content": "We show that composition of tangent models trained on individual tasks is a viable method for replay-free continual fine-tuning, and can even outperform recently published replay-based methods. As a side benefit, our method also yields fully parallel training across tasks, and enables zero-cost unlearning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 597, + 102, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 102, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 102, + 609 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": "In continual learning, models are trained on a sequence of tasks indexed by " + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "inline_equation", + "content": "t = 1,\\dots ,T" + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": " ,with the goal of performing well on all, under three main scenarios [51]: (1) Task-incremental, where the task identity " + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": " (Task-ID) is known during inference, (2) Data/Domain-incremental, where Task-ID is not provided during inference, and (3) the most challenging Class-incremental, where Task-ID is unknown during inference and in addition the classes repre" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 190, + 544, + 507 + ], + "blocks": [ + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "lines": [ + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "text", + "content": "Table 1. A Comparison of Composition Methods: On average across 25 experiments, Tangent Model Ensembling (TME) improves accuracy by " + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "text", + "content": " over standard ensembling of non-linear fine-tuned models at a " + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "text", + "content": " increase of inference cost. Tangent Model Composition (TMC) also improved accuracy, by " + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "text", + "content": ", while reducing inference cost by " + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "inline_equation", + "content": "2.5\\times" + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "text", + "content": " in our experiments, and growing linearly with the number of models in the ensemble. Best results for single and multi model inference are indicated in bold and italic respectively. Joint training paragon accuracy (Original Model/Tangent Model): Caltech-256 - " + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "inline_equation", + "content": "89.17\\% /86.77\\%" + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "text", + "content": ", MIT-67 - " + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "inline_equation", + "content": "77.29\\% /77.74\\%" + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "text", + "content": ", OxfordPets - " + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "inline_equation", + "content": "92.68\\% /93.68\\%" + }, + { + "bbox": [ + 305, + 68, + 545, + 189 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 190, + 544, + 507 + ], + "lines": [ + { + "bbox": [ + 308, + 190, + 544, + 507 + ], + "spans": [ + { + "bbox": [ + 308, + 190, + 544, + 507 + ], + "type": "table", + "html": "
DatasetTasksSoupEns-LEns-SMTMCTME
Inference/MemoryCost:O(1)O(T)O(T)O(1)O(T)
C-Caltech-256 (Class-IL)584.0082.8083.0984.8285.53
1080.1076.8579.7682.3782.58
2074.8666.7675.4778.6679.30
C-MIT-67 (Class-IL)559.4856.7958.6869.4371.72
1049.3042.3650.5264.5365.72
2028.5123.6135.6248.9854.93
C-OxfordPnts (Class-IL)580.7181.1481.8684.4285.04
1071.3569.7177.3177.7579.97
D-Caltech-256 (Data-IL)585.9686.2786.5984.8685.13
1083.4884.0584.8083.6484.16
2079.6480.7283.1781.9982.94
D-MIT-67 (Data-IL)566.0066.5467.2174.8875.45
1058.8859.2060.4772.4173.41
2050.1051.9955.2070.2272.11
D-OxfordPnts (Data-IL)590.2290.8391.0293.1793.21
1087.6688.4288.8892.4892.72
2085.5785.9786.5891.2891.77
C-Caltech-256 (Task-IL)594.0194.7294.8494.3394.76
1095.0095.8796.0496.0396.23
2095.5696.5396.8497.2297.38
C-MIT-67 (Task-IL)583.3686.2486.3791.2791.79
1087.1990.4290.9794.4595.03
2088.4193.7393.8897.7198.09
C-OxfordPnts (Task-IL)597.2797.2797.3598.5798.74
1097.5997.9297.9899.0599.13
", + "image_path": "1b505f6886c234254cd9cf4300e71010586f5f15e4916c3f10c7b1c0fd200a3a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 533, + 433, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 433, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 433, + 544 + ], + "type": "text", + "content": "sented in each task are disjoint." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 552, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 635 + ], + "type": "text", + "content": "We assume the setting of continual fine-tuning of a pretrained model. Naive fine-tuning causes catastrophic forgetting of prior tasks, including the one on which pre-training occurred, as the number of tasks increases [6]. Thus, only early tasks benefit from pre-training. Therefore, it is important for continual fine-tuning to ensure that all tasks can benefit from information acquired during pre-training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "We introduce a unified method to tackle the main scenarios (1)-(3) by composing separately-trained tangent component models. We show that each task takes full advantage of pre-training, and catastrophic forgetting of prior tasks is averted since information from each previous task is encoded separately and only fused for inference." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18678" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 167, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 167, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 167, + 83 + ], + "type": "text", + "content": "3.1. Model Linearization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 290, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 290, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 290, + 247 + ], + "type": "text", + "content": "The effectiveness of model aggregation in weight space hinges on whether the models lie in within the same small-loss training basin of convergence. It is known that weights of different models fine-tuned on the same dataset can be connected by linear paths with constant loss (\"linear mode connectivity\") [16], but this does not generally occur for models fine-tuned on disjoint tasks/datasets, which is the case of interest to us. Therefore, we convexify the loss landscape locally by approximating the function represented by a deep (non-linear) network with its first-order Taylor approximation around the pre-trained weights. For any deep network " + }, + { + "bbox": [ + 46, + 90, + 290, + 247 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 90, + 290, + 247 + ], + "type": "text", + "content": " and initial weights " + }, + { + "bbox": [ + 46, + 90, + 290, + 247 + ], + "type": "inline_equation", + "content": "w \\in \\mathbb{R}^m" + }, + { + "bbox": [ + 46, + 90, + 290, + 247 + ], + "type": "text", + "content": ", we denote with " + }, + { + "bbox": [ + 46, + 90, + 290, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(w)" + }, + { + "bbox": [ + 46, + 90, + 290, + 247 + ], + "type": "text", + "content": " the set of all such models tangent to " + }, + { + "bbox": [ + 46, + 90, + 290, + 247 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 46, + 90, + 290, + 247 + ], + "type": "text", + "content": " given by:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 82, + 266, + 287, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 266, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 82, + 266, + 287, + 281 + ], + "type": "interline_equation", + "content": "\\mathcal {H} (w) = \\left\\{h _ {\\delta} (\\cdot) \\triangleq f _ {w} (\\cdot) + \\nabla_ {w} f _ {w} (\\cdot) \\cdot \\delta \\right\\}. \\tag {1}", + "image_path": "5af978ac8a948643f30c13f8bef4930b2cab50310bf11e296b807ce0bf2b67a5.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 289, + 288, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 288, + 350 + ], + "type": "text", + "content": "Since the tangent model " + }, + { + "bbox": [ + 46, + 289, + 288, + 350 + ], + "type": "inline_equation", + "content": "h_{\\delta}" + }, + { + "bbox": [ + 46, + 289, + 288, + 350 + ], + "type": "text", + "content": " is linear in the parameters " + }, + { + "bbox": [ + 46, + 289, + 288, + 350 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 289, + 288, + 350 + ], + "type": "text", + "content": ", training it with standard losses such as mean-squared error (MSE) or empirical cross-entropy yields a convex loss landscape. In order to compose models, leveraging Jensen's inequality, we are guaranteed that for any dataset " + }, + { + "bbox": [ + 46, + 289, + 288, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 289, + 288, + 350 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 358, + 287, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 358, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 59, + 358, + 287, + 392 + ], + "type": "interline_equation", + "content": "L \\left(\\mathcal {D}, \\sum_ {i} ^ {T} h _ {\\theta_ {i} \\delta_ {i}}\\right) \\leq \\sum_ {i} \\theta_ {i} L \\left(\\mathcal {D}, h _ {\\delta_ {i}}\\right), \\quad \\sum_ {i} \\theta_ {i} = 1. \\tag {2}", + "image_path": "73e9408f9c6600cbc55f4bb9a9c38ef7c7b593ffd2991c260ddc73ba50fc97b2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 482 + ], + "type": "text", + "content": "In other words, the loss resulting from a convex combination of tangent models in weight space will always be lower than or equal to the convex combination of their individual losses. This is particularly useful when combining models trained to some constant (possibly zero) loss, since it ensures that the composed model will have loss lower or equal to the loss of the component models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 483, + 288, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 590 + ], + "type": "text", + "content": "While tangent models might seem computationally expensive to train, we note that the Jacobian vector product " + }, + { + "bbox": [ + 46, + 483, + 288, + 590 + ], + "type": "inline_equation", + "content": "\\nabla_w f_w(\\cdot) \\cdot \\delta" + }, + { + "bbox": [ + 46, + 483, + 288, + 590 + ], + "type": "text", + "content": " can be computed in a single forward pass as shown by [39, 35, 1]. This reduces the computational cost to that of a forward pass through the original network. Also note that while the tangent model has twice the degrees of freedom of the original model, fixing the pre-trained model at " + }, + { + "bbox": [ + 46, + 483, + 288, + 590 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 483, + 288, + 590 + ], + "type": "text", + "content": " results in only a convex optimization in the remaining parameters, as many as the original model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": "In the context of fine-tuning, " + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": " is initialized by pretraining on large datasets such as ImageNet [13]. In the rest of this paper, we use the short-form notation " + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(w)" + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "inline_equation", + "content": "h_{\\delta}" + }, + { + "bbox": [ + 46, + 590, + 287, + 639 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 645, + 287, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 287, + 670 + ], + "type": "text", + "content": "3.2. Tangent Model Composition for Replay-free Continual Fine-tuning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 675, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 675, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 675, + 288, + 715 + ], + "type": "text", + "content": "Ensembling is an effective strategy for combining weak learners into a strong model [14]. For any finite " + }, + { + "bbox": [ + 47, + 675, + 288, + 715 + ], + "type": "inline_equation", + "content": "F \\subseteq \\mathbb{A}" + }, + { + "bbox": [ + 47, + 675, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 675, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{A}" + }, + { + "bbox": [ + 47, + 675, + 288, + 715 + ], + "type": "text", + "content": " is a set of functions, and " + }, + { + "bbox": [ + 47, + 675, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\lambda = \\{\\lambda_i\\}_{i=1}^{|F|}" + }, + { + "bbox": [ + 47, + 675, + 288, + 715 + ], + "type": "text", + "content": " a set of" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 538, + 233 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 538, + 233 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 538, + 233 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 538, + 233 + ], + "type": "image", + "image_path": "ac03a9f1f8388b819c068a7ec9d8cb4a44135cd55a2075a1661e7b9e9c82c359.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 237, + 547, + 326 + ], + "lines": [ + { + "bbox": [ + 304, + 237, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 547, + 326 + ], + "type": "text", + "content": "Figure 2. Accuracy plot over models trained on C-Caltech-256, C-MIT-67, and C-Oxford Pets, split into 10 tasks each. Soft-max ensembling (TME) consistently yields the best results compared to logit ensembling (TMC) for each loss function. Composition of component models trained with the Rescaled Square Loss (RSL) outperforms those trained with empirical cross-entropy (CE) or MSE loss, trading off only " + }, + { + "bbox": [ + 304, + 237, + 547, + 326 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 237, + 547, + 326 + ], + "type": "text", + "content": " accuracy against TME to reduce inference time and memory requirement from " + }, + { + "bbox": [ + 304, + 237, + 547, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(T)" + }, + { + "bbox": [ + 304, + 237, + 547, + 326 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 237, + 547, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 304, + 237, + 547, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 345, + 504, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 345, + 504, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 504, + 357 + ], + "type": "text", + "content": "weights, the logit ensemble operation is given by" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 365, + 545, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 365, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 361, + 365, + 545, + 399 + ], + "type": "interline_equation", + "content": "\\operatorname {E n s e m b l e} (F, \\lambda) = \\sum_ {i = 1} ^ {| F |} \\lambda_ {i} f _ {i} (\\cdot) \\tag {3}", + "image_path": "f77b7c462a29c25e8fb564d0d069ad547fd13d80860671ea6dc9d4baa08286dd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 407, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 467 + ], + "type": "text", + "content": "In the context of continual learning, a simple yet effective approach is to simply train a new model on each task " + }, + { + "bbox": [ + 304, + 407, + 545, + 467 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 407, + 545, + 467 + ], + "type": "text", + "content": ", and ensemble the outputs at inference time. Unfortunately, this has the obvious shortcoming that both storage requirements and inference time grows along with the number of tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "text", + "content": "However, consider the case where " + }, + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathbb{A} = \\mathcal{H}" + }, + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "text", + "content": ". Then for any finite " + }, + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "inline_equation", + "content": "H \\subseteq \\mathcal{H}" + }, + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "text", + "content": " and associated weights " + }, + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 467, + 545, + 491 + ], + "type": "text", + "content": ", we can write" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 346, + 498, + 545, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 498, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 346, + 498, + 545, + 587 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {E n s e m b l e} (H, \\lambda) = \\sum_ {i = 1} ^ {| H |} \\lambda_ {i} h _ {i} (\\cdot) (4) \\\\ = f _ {w} (\\cdot) + \\sum_ {i = 1} ^ {| H |} \\nabla_ {w} f _ {w} (\\cdot) \\cdot \\lambda_ {i} \\delta_ {i} (5) \\\\ = h _ {\\sum \\lambda_ {i} \\delta_ {i}} (\\cdot) (6) \\\\ \\end{array}", + "image_path": "b4eb57b5e81d28afdafb1464202df01b4ea91da79d12fddb17d792894d0b8dd9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "content": "In other words, since " + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "content": " defines a vector space with addition and scalar multiplication based on " + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "content": ", any linear combination (or ensemble) of models in " + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "content": " is equivalent to a single model in " + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "content": " derived simply by the linear combination over " + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "content": ". This reduces the inference time for any ensemble of " + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "content": " tangent models from " + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(T)" + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": "In continual learning, it is also important for storage space to not scale with " + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": ", since the number of tasks can be arbitrarily large. However, since tasks arrive sequentially in the continual learning framework, we can further reduce" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18679" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 102, + 517, + 251 + ], + "blocks": [ + { + "bbox": [ + 46, + 68, + 547, + 102 + ], + "lines": [ + { + "bbox": [ + 46, + 68, + 547, + 102 + ], + "spans": [ + { + "bbox": [ + 46, + 68, + 547, + 102 + ], + "type": "text", + "content": "Table 2. Class-Incremental Learning: Comparison against existing methods using an ImageNet pre-trained ResNet-50. TMC is almost uniformly better than all replay-based and replay-free methods tested, outperforming the best replay-free (RF) method by " + }, + { + "bbox": [ + 46, + 68, + 547, + 102 + ], + "type": "inline_equation", + "content": "14.48\\%" + }, + { + "bbox": [ + 46, + 68, + 547, + 102 + ], + "type": "text", + "content": " and even the best replay-based method [6] by " + }, + { + "bbox": [ + 46, + 68, + 547, + 102 + ], + "type": "inline_equation", + "content": "1.53\\%" + }, + { + "bbox": [ + 46, + 68, + 547, + 102 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 102, + 517, + 251 + ], + "lines": [ + { + "bbox": [ + 72, + 102, + 517, + 251 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 517, + 251 + ], + "type": "table", + "html": "
Method / BufferTasksC-Caltech-256C-MIT-67
5102051020
ER [46] / 30070.14±0.2265.75±0.9265.490.6054.61±0.4149.57±1.4645.331.73
LWF [30] / None34.55±0.1822.76±0.1714.37±0.1425.14±0.5713.30±0.496.41±0.17
oEWC [48] / None28.71±1.3116.32±4.3715.59±3.8314.18±0.777.12±0.494.33±0.16
DER++ [7] / 30080.50±0.2677.84±1.0274.46±0.2164.56±0.5559.73 ±1.0250.28±1.27
Co2L [10] / 30020.71±2.8917.13±2.5314.620.6119.64±12.9321.21±0.5017.78±1.33
ER-ACE [9] / 30078.47±1.2573.85±0.5370.87±1.3864.94±0.5060.25±0.2652.48±0.29
TWF (RF) [6] / None77.95±0.5270.67±0.7262.79±0.8256.58±0.6245.03 ±0.7828.91±0.92
TWF [6] / 30081.34±0.9177.03±0.5372.44±0.4268.33±0.4664.21 ±0.6556.29±0.68
TMC / None84.82 ±0.0382.37 ±0.0878.66 ±0.1369.43±0.2164.53±0.2848.98 ±0.13
", + "image_path": "d9b6c1e8863252bfbee63f7cd73b3ffb62a7eaa9b76e6aa5b11778def2a36790.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": "the memory capacity required for storage of models from " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(T)" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": ", to " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": " by using a simple autoregressive model: We train on the first task " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": " to produce model " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "\\tilde{h}_1 = h_1" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": ". Then, for any subsequent task " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "t > 1" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": ", we train a new model " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "\\tilde{h}_t" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": ", and compose it with " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": " to give " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "h_t = \\lambda_{(t,1)}\\tilde{h}_t + \\lambda_{(t,2)}h_{t-1} := h_{\\delta'}" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "\\delta' = \\frac{\\lambda_{(t,1)}\\tilde{\\delta}_t + \\lambda_{(t,2)}\\delta_{t-1}}{\\lambda_{(t,1)} + \\lambda_{(t,2)}}" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": ". Since we only require model " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": ", all prior models up to and including " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": " can be discarded upon constructing " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": ", hence reducing memory requirement from " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(T)" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 46, + 262, + 289, + 448 + ], + "type": "text", + "content": ", equal to the size of a single model. In contrast to methods that expands the model at each task, or require a memory buffer of examples from previous tasks, this method does not incur additional storage requirements, nor assume availability of an extra replay buffer for storing past exemplars (i.e. our method is replay-free)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 288, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 472 + ], + "type": "text", + "content": "Remark 1 (Free Parallelization): Additionally, for any given convex loss function " + }, + { + "bbox": [ + 46, + 449, + 288, + 472 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 449, + 288, + 472 + ], + "type": "text", + "content": ", the optimization objective" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 478, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 478, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 114, + 478, + 287, + 506 + ], + "type": "interline_equation", + "content": "\\min _ {\\delta \\in \\mathbb {R} ^ {m}} \\sum_ {(x, y) \\in \\mathcal {D}} L \\left(h _ {\\delta} (x), y\\right) \\tag {7}", + "image_path": "fd6fc319d53e6174f8922c4ea7d7b94bc37bedd0d23861322676e6d0d68e348d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": "is convex and thus converges to a global optimum. This implies that each model " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\tilde{h}_t" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": " can be trained from a standard initialization as opposed to from " + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": ". Hence, if multiple tasks arrive at the same time, such as when training in a federated manner, training can be parallelized across tasks and composed in a zero-cost, zero-shot manner to yield equivalent results (assuming equivalence of global optima) to sequential training and composition. To the best of our knowledge, apart from [41] which simply stores samples in a replay buffer and trains on them at test time, no other work in continual learning can be parallelized across tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Remark 2 (Free Forgetting/Unlearning): If we relax memory constraints to store models from each task, unlearning any specific task " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " (e.g., private data associated with specific samples) is straightforward and can be done in a zero-cost, zero-shot manner. Since subtraction is well-defined in the vector space " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": ", any task " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " can be unlearned" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "type": "text", + "content": "from the model " + }, + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "type": "inline_equation", + "content": "h_T" + }, + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "type": "text", + "content": " trained on " + }, + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "type": "text", + "content": " tasks by simply subtracting the weights associated to " + }, + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "type": "inline_equation", + "content": "\\tilde{h}_i" + }, + { + "bbox": [ + 305, + 262, + 545, + 286 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 359, + 295, + 545, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 295, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 359, + 295, + 545, + 328 + ], + "type": "interline_equation", + "content": "h _ {T \\backslash i} = h _ {T} - \\lambda_ {i, 1} \\prod_ {j = i + 1} ^ {T} \\lambda_ {(j, 2)} \\tilde {h} _ {i} \\tag {8}", + "image_path": "fc6efdeb886e54b6a8c5aa3498a3ec6c7bb62c505efdfa3da83dfb4380136842.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 336, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 336, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 545, + 384 + ], + "type": "text", + "content": "Remark 3 (Unified Continual Fine-tuning): Our method does not assume any specific task splits, and is hence generalizable across all three forms of continual fine-tuning: task-incremental, class-incremental, and data-incremental." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 392, + 425, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 425, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 425, + 403 + ], + "type": "text", + "content": "3.3. Scale Standardization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 411, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 541 + ], + "type": "text", + "content": "Training with the cross-entropy loss does not ensure that the weights are comparable in scale. We can study the implications of this for composition on " + }, + { + "bbox": [ + 304, + 411, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 304, + 411, + 545, + 541 + ], + "type": "text", + "content": " through the lens of logits ensembling. While logits ensembling has been shown to be effective [58], Fig. 2 shows that ensembling the normalized soft-max outputs (Tangent Model Ensembling, TME) is uniformly better than logit ensembling/weight averaging via TMC across various loss functions. However, TME cannot be represented by a linear combination of perturbations " + }, + { + "bbox": [ + 304, + 411, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 304, + 411, + 545, + 541 + ], + "type": "text", + "content": " and hence incurs " + }, + { + "bbox": [ + 304, + 411, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(T)" + }, + { + "bbox": [ + 304, + 411, + 545, + 541 + ], + "type": "text", + "content": " model storage costs and " + }, + { + "bbox": [ + 304, + 411, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(T)" + }, + { + "bbox": [ + 304, + 411, + 545, + 541 + ], + "type": "text", + "content": " inference time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 542, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 542, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 542, + 545, + 578 + ], + "type": "text", + "content": "Instead, we propose to induce a waterbed effect similar to soft-max ensembling using standardization. In particular, we train using the Rescaled Square Loss (RSL) [23]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 586, + 545, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 586, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 315, + 586, + 545, + 632 + ], + "type": "interline_equation", + "content": "L _ {s} (x, y) = \\frac {1}{K} \\left(\\alpha ([ f (x) ] _ {y} - \\beta) ^ {2} + \\sum_ {i = 1, i \\neq y} ^ {K} ([ f (x) ] _ {i}) ^ {2}\\right) \\tag {9}", + "image_path": "ba84987ebeb3231437b9daa55e36d416519efbce6a467406a7a54523ef7bf1b7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": " are hyper-parameters. This is the regular MSE loss when " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\alpha = \\beta = 1" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": ". Larger values of " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": " reduce soft-max entropy in the output by scaling positive output class signals while keeping negative ones close to zero." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " to control the interaction between component models. For example, when tasks are highly dissimilar," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18680" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 146, + 276, + 294 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 145 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 145 + ], + "type": "text", + "content": "Table 3. Data-Incremental Learning on D-MIT-67, 4 tasks using ResNet-18. For fair comparison, we run our method under the settings proposed by [49]. TMC outperforms the best replay-free method [30] by " + }, + { + "bbox": [ + 47, + 69, + 287, + 145 + ], + "type": "inline_equation", + "content": "4.61\\%" + }, + { + "bbox": [ + 47, + 69, + 287, + 145 + ], + "type": "text", + "content": ", and the best replay-based method [49] by " + }, + { + "bbox": [ + 47, + 69, + 287, + 145 + ], + "type": "inline_equation", + "content": "1.27\\%" + }, + { + "bbox": [ + 47, + 69, + 287, + 145 + ], + "type": "text", + "content": ", while supporting parallel training across tasks (PL). If we sacrifice parallelization by initializing each model at task " + }, + { + "bbox": [ + 47, + 69, + 287, + 145 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 69, + 287, + 145 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 69, + 287, + 145 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 47, + 69, + 287, + 145 + ], + "type": "text", + "content": " (TMC-Seq), we further improve over TMC by " + }, + { + "bbox": [ + 47, + 69, + 287, + 145 + ], + "type": "inline_equation", + "content": "0.28\\%" + }, + { + "bbox": [ + 47, + 69, + 287, + 145 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 146, + 276, + 294 + ], + "lines": [ + { + "bbox": [ + 53, + 146, + 276, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 276, + 294 + ], + "type": "table", + "html": "
MethodPL?Replay BufferAccuracy
SGDXNone63.48
LwF[30]XNone67.21
MAS[2]XNone62.49
OSLA[45]XNone64.08
EWC [26]X50063.68
DLCFT[49]X50070.55
TMC✓None71.82
TMC-SeqXNone72.10
Joint (Linear)--73.61
Joint (Paragon)--74.40
", + "image_path": "7f2e3e13cdf73b2049efd0360c75b1a419ae252b5d7a0c3e226821c6f3d60911.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 313, + 286, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 286, + 384 + ], + "type": "text", + "content": "such as in task- and class-incremental continual learning, we use larger values of " + }, + { + "bbox": [ + 46, + 313, + 286, + 384 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 313, + 286, + 384 + ], + "type": "text", + "content": " to reduce interference of component models trained on other tasks upon composition. On the other hand, when tasks are highly synergistic such as in data-incremental learning, lower values of " + }, + { + "bbox": [ + 46, + 313, + 286, + 384 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 313, + 286, + 384 + ], + "type": "text", + "content": " can be used to encourage interaction in the final composed model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 407, + 127, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 127, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 127, + 420 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 431, + 286, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 286, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 286, + 454 + ], + "type": "text", + "content": "We describe comparison baselines in Sec. 4.1, and our main results in Sec. 4.2, with additional studies in Sec. 4.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 459, + 286, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 286, + 578 + ], + "type": "text", + "content": "Our method is replay-free. As such, its computational cost is at most half of what it would be if it used a replay buffer since replay-based methods typically sample separate batches of equal size from both the current task and replay buffer at each iteration [7, 6, 10, 46, 9]. Nonetheless, we compare our method to both replay-based and replay-free methods in our baselines. While replay-based methods have thus far outperformed replay-free ones, ours is uniformly more accurate than replay-based methods tested, in addition to replay-free ones." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "We evaluate our method on Caltech-256 [20], MIT-67 [42], and OxfordPets [38]. For continual fine-tuning, we further split each training dataset into multiple disjoint subsets, each representing a single task, for fair comparison across experiments. We prefix sequential datasets split into tasks with disjoint labels (for task- and class-incremental learning) with C-, and random splits using a standardized random seed (for data-incremental learning) with D-. We set " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 1, \\beta = 25" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " for the former, " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 1, \\beta = 5" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " for the latter, and " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 1, \\beta = 5" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " for TME. Results in all tables and figures are averaged across 3 runs." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 340, + 73, + 511, + 203 + ], + "blocks": [ + { + "bbox": [ + 340, + 73, + 511, + 203 + ], + "lines": [ + { + "bbox": [ + 340, + 73, + 511, + 203 + ], + "spans": [ + { + "bbox": [ + 340, + 73, + 511, + 203 + ], + "type": "image", + "image_path": "0ef32265e31f3fe4ed170d117da0232be78600cf938bfe4440a5fca75766d9e3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 340, + 207, + 511, + 339 + ], + "blocks": [ + { + "bbox": [ + 340, + 207, + 511, + 339 + ], + "lines": [ + { + "bbox": [ + 340, + 207, + 511, + 339 + ], + "spans": [ + { + "bbox": [ + 340, + 207, + 511, + 339 + ], + "type": "image", + "image_path": "273a8be4dd0983e861b7db82430be3541c7530047c70227af59d60cdc76eafcd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 344, + 545, + 421 + ], + "lines": [ + { + "bbox": [ + 304, + 344, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 545, + 421 + ], + "type": "text", + "content": "Figure 3. Ablation on " + }, + { + "bbox": [ + 304, + 344, + 545, + 421 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 344, + 545, + 421 + ], + "type": "text", + "content": " averaged across (Top:) C-Caltech-256, C-MIT-67, and C-OxfordPets and (Bottom:) D-Caltech-256, D-MIT-67, and D-OxfordPets, split into 10 disjoint tasks. For class incremental datasets, larger values of " + }, + { + "bbox": [ + 304, + 344, + 545, + 421 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 344, + 545, + 421 + ], + "type": "text", + "content": " result in better generalization of the final composed model. For data-incremental datasets where tasks are synergistic, larger values of " + }, + { + "bbox": [ + 304, + 344, + 545, + 421 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 344, + 545, + 421 + ], + "type": "text", + "content": " instead harm the accuracy of both the component models and the composed model." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 441, + 433, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 433, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 433, + 453 + ], + "type": "text", + "content": "4.1. Composition Baselines" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 464, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 464, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 305, + 464, + 545, + 488 + ], + "type": "text", + "content": "In each of the following sections, we compare against the following baselines for model aggregation:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 493, + 545, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 493, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 493, + 545, + 516 + ], + "type": "text", + "content": "Logits Ensemble (Ens-L): Averaging the output logits of non-linear component models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 521, + 544, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 544, + 545 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 544, + 545 + ], + "type": "text", + "content": "Softmax Ensemble (Ens-SM): Averaging the softmax output of non-linear component models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 550, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 621 + ], + "type": "text", + "content": "Weight composition of non-linear models (Soup): Averaging the weights of non-linear component models. This method is inspired by Model Soups [58] where weights of various non-linear models trained on the same task but with different hyper-parameters and augmentations are averaged to improve generalization at no additional inference cost." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 643, + 390, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 643, + 390, + 654 + ], + "spans": [ + { + "bbox": [ + 306, + 643, + 390, + 654 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 712 + ], + "type": "text", + "content": "We show that TMC outperforms existing methods in all three continual fine-tuning settings [51], from hardest to easiest: Class-Incremental (Sec. 4.2.1), Data-Incremental (Sec. 4.2.2), and Task-Incremental (Sec. 4.2.3)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18681" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 102, + 518, + 251 + ], + "blocks": [ + { + "bbox": [ + 46, + 69, + 547, + 102 + ], + "lines": [ + { + "bbox": [ + 46, + 69, + 547, + 102 + ], + "spans": [ + { + "bbox": [ + 46, + 69, + 547, + 102 + ], + "type": "text", + "content": "Table 4. Task-Incremental Learning: Comparison against existing methods using ImageNet pre-trained ResNet-50. We almost uniformly outperform all replay-based and replay-free methods that we compare against. In particular, we improve over the best replay-free (RF) method by " + }, + { + "bbox": [ + 46, + 69, + 547, + 102 + ], + "type": "inline_equation", + "content": "2.46\\%" + }, + { + "bbox": [ + 46, + 69, + 547, + 102 + ], + "type": "text", + "content": ", and the best replay-based method by " + }, + { + "bbox": [ + 46, + 69, + 547, + 102 + ], + "type": "inline_equation", + "content": "1.38\\%" + }, + { + "bbox": [ + 46, + 69, + 547, + 102 + ], + "type": "text", + "content": " [6]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 102, + 518, + 251 + ], + "lines": [ + { + "bbox": [ + 73, + 102, + 518, + 251 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 518, + 251 + ], + "type": "table", + "html": "
Method / BufferTasksC-Caltech-256C-MIT-67
5102051020
ER [46] / 30092.11±0.2694.33±0.4895.76±0.1385.84±0.0990.59±0.5195.13±0.27
LWF [30] / None93.73±0.1295.31±0.0896.29±0.0587.55±0.2090.82±0.0495.68±0.58
oEWC [48] / None83.97±0.7176.20±16.0791.49±0.3067.31±3.2268.45±4.9384.74±1.96
DER++ [7] / 30093.79±0.1394.92±0.1195.87±0.1786.83±0.2691.63±0.7594.54±0.62
Co2L [10] / 30034.51±3.0539.62±1.7045.94±1.4739.32±20.4764.40±1.3151.50±2.21
ER-ACE [9] / 30093.11±0.1294.63±0.0695.86±0.1087.13±0.1690.96±0.2894.92±0.60
TWF (RF) [6] / None94.45±0.0595.64±0.0696.36±0.1286.76±0.2588.27±0.5494.74±0.35
TWF [6] / 30094.38±0.0895.74±0.1796.59±0.0788.22±0.4192.13±0.2695.67±0.46
TMC / None94.33±0.0396.03±0.0297.22±0.0391.27±0.2194.45±0.1397.71±0.13
", + "image_path": "6a8a639407586607914e585f6184be3e9ea4bf07046c7e51d2102a12c719a783.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 262, + 201, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 201, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 201, + 274 + ], + "type": "text", + "content": "4.2.1 Class-Incremental Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 291, + 288, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 288, + 316 + ], + "type": "text", + "content": "In class-incremental learning, datasets are partitioned so that label spaces of each task are disjoint." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "text", + "content": "Comparison against Composition Baselines: We compare against model composition baselines in Table 1. While TME yields the best generalization performance, this comes at a cost of " + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(T)" + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "text", + "content": " inference time. On the other hand, TMC achieves an " + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "text", + "content": " inference time with only a " + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "inline_equation", + "content": "1.73\\%" + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "text", + "content": " trade-off in generalization accuracy on average, while outperforming weight averaging, logit ensembling, and even soft-max ensembling of non-linear models by an average of " + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "inline_equation", + "content": "7.83\\%" + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "inline_equation", + "content": "11.37\\%" + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "inline_equation", + "content": "6.08\\%" + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "text", + "content": " respectively, even though the latter two both require " + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(T)" + }, + { + "bbox": [ + 46, + 319, + 287, + 440 + ], + "type": "text", + "content": " inference time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 445, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 566 + ], + "type": "text", + "content": "Comparison against existing methods: In Table 2, we compare against recent continual fine-tuning methods including both replay-free and replay-based methods. Here, replay-free methods [30, 48] perform poorly in the class-incremental setting relative to stronger replay-based methods [46, 9, 7, 6]. However, we show that not only do we improve over the best replay-free method by " + }, + { + "bbox": [ + 46, + 445, + 287, + 566 + ], + "type": "inline_equation", + "content": "14.48\\%" + }, + { + "bbox": [ + 46, + 445, + 287, + 566 + ], + "type": "text", + "content": ", but also over the best replay-based method by " + }, + { + "bbox": [ + 46, + 445, + 287, + 566 + ], + "type": "inline_equation", + "content": "1.53\\%" + }, + { + "bbox": [ + 46, + 445, + 287, + 566 + ], + "type": "text", + "content": " [6], corresponding to a " + }, + { + "bbox": [ + 46, + 445, + 287, + 566 + ], + "type": "inline_equation", + "content": "17.59\\%" + }, + { + "bbox": [ + 46, + 445, + 287, + 566 + ], + "type": "text", + "content": " relative error reduction towards the paragon performance of joint training (Table 1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Comparison against EFCIL methods: In Table 5, we compare against EFCIL methods on C-TinyImageNet [27]. The original EFCIL methods are evaluated under a different setting (denoted " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "), where the zeroth task contains " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " of the dataset. Hence, we further modified two of the best performing EFCIL methods, SSRE [62] and FeTrIL [40] to use the pre-trained ImageNet initialization, and run them under our setting of uniformly sharded data. TMC not only outperforms all methods ran under our harder setting, by an average of " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "15.3\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " over the next best method [40], but even yields equal or better performance when compared to all methods ran under " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 262, + 457, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 457, + 274 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 457, + 274 + ], + "type": "text", + "content": "4.2.2 Data-Incremental Learning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 281, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 317 + ], + "type": "text", + "content": "Also referred to as domain-incremental learning, this setting assumes that tasks are split in a non-stratified manner, where each task has the same output space [51, 49]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "text", + "content": "Comparison against Composition Baselines: In Table 1, we compare against model composition baselines. Similar to the class-incremental setting, TME outperforms TMC by an average of " + }, + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "inline_equation", + "content": "0.66\\%" + }, + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "text", + "content": " at a cost of " + }, + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(T)" + }, + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "text", + "content": " inference time. TMC uses " + }, + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "text", + "content": " inference time and outperforms non-linear composition, logit ensemble, and even soft-max ensemble by an average of " + }, + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "inline_equation", + "content": "7.04\\%" + }, + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "inline_equation", + "content": "6.32\\%" + }, + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "inline_equation", + "content": "5.21\\%" + }, + { + "bbox": [ + 304, + 318, + 546, + 401 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "text", + "content": "Comparison against existing methods: In Table 3, we benchmark our model against results obtained by [49] on a pre-trained ResNet-18 model using the D-MIT-67 dataset split into 4 disjoint subsets. We improve over the best replay-free method [30] by " + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "inline_equation", + "content": "4.61\\%" + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "text", + "content": ", and the best replay-based method [49] by " + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "inline_equation", + "content": "1.27\\%" + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "text", + "content": ", corresponding to a " + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "inline_equation", + "content": "64\\%" + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "text", + "content": " relative error reduction towards the joint training paragon respectively. TMC-Seq, which uses sequential initialization by initializing each component model " + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "inline_equation", + "content": "h_{t - 1}" + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "text", + "content": ", further improves over TMC by " + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "inline_equation", + "content": "0.28\\%" + }, + { + "bbox": [ + 304, + 402, + 547, + 534 + ], + "type": "text", + "content": " but at the cost of parallelizability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 550, + 457, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 550, + 457, + 562 + ], + "spans": [ + { + "bbox": [ + 305, + 550, + 457, + 562 + ], + "type": "text", + "content": "4.2.3 Task-Incremental Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 617 + ], + "type": "text", + "content": "In task-incremental learning, the output spaces of different tasks are disjoint, but the source task of each sample is known at inference time. Thus, we can restrict the model predictions to that of the given task during evaluation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "content": "Comparison against Composition Baselines: Due to the disjoint output spaces of each task, ensembling outputs of specialist component models trained on individual tasks and restricting them to a target task essentially only considers the predictions of the specialist model trained on the target task (since, in general, specialist models are not effective discriminants for tasks other than those they are triend on). Hence, the independence of output spaces renders en" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18682" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 113, + 504, + 162 + ], + "blocks": [ + { + "bbox": [ + 46, + 68, + 547, + 112 + ], + "lines": [ + { + "bbox": [ + 46, + 68, + 547, + 112 + ], + "spans": [ + { + "bbox": [ + 46, + 68, + 547, + 112 + ], + "type": "text", + "content": "Table 5. Comparison against EFCIL methods using either ImageNet pre-trained ResNet-50, or under the original EFCIL setting where the zeroth task contains " + }, + { + "bbox": [ + 46, + 68, + 547, + 112 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 68, + 547, + 112 + ], + "type": "text", + "content": " of the dataset (denoted " + }, + { + "bbox": [ + 46, + 68, + 547, + 112 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 68, + 547, + 112 + ], + "type": "text", + "content": "). All experiments are ran on C-TinyImageNet. Italics denote the best method under " + }, + { + "bbox": [ + 46, + 68, + 547, + 112 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 68, + 547, + 112 + ], + "type": "text", + "content": ", and bold denotes the best method under our harder setting. Under the former, TMC performs equally or better than all methods. Under the latter, TMC outperforms the best method by an average of " + }, + { + "bbox": [ + 46, + 68, + 547, + 112 + ], + "type": "inline_equation", + "content": "15.3\\%" + }, + { + "bbox": [ + 46, + 68, + 547, + 112 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 113, + 504, + 162 + ], + "lines": [ + { + "bbox": [ + 85, + 113, + 504, + 162 + ], + "spans": [ + { + "bbox": [ + 85, + 113, + 504, + 162 + ], + "type": "table", + "html": "
TasksPASS† [61]IL2A† [60]SSRE† [62]FeTrIL† [40]SSRE[62]FeTrIL[40]TMC
549.647.350.454.833.744.157.6
1047.344.748.953.125.636.153.1
", + "image_path": "8a74b4472ac9f08ce4efd40f4869e164c1548dabff5031dc235d697de0758e7b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 173, + 288, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 173, + 288, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 173, + 288, + 292 + ], + "type": "text", + "content": "semblies of specialist models highly effective. In contrast, composing non-linear models in weight space does not ensure minimal interference across different tasks. Indeed, we show in Table 1 that TME yields the best results, while the non-linear composition approach yields the worst in task-incremental learning. On the other hand, due to the equivalence of TMC to (logit) ensembling of linearized models, TMC outperforms not only weight averaging, but also logit and soft-max ensembles of non-linear models, by " + }, + { + "bbox": [ + 46, + 173, + 288, + 292 + ], + "type": "inline_equation", + "content": "3.78\\%" + }, + { + "bbox": [ + 46, + 173, + 288, + 292 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 173, + 288, + 292 + ], + "type": "inline_equation", + "content": "1.83\\%" + }, + { + "bbox": [ + 46, + 173, + 288, + 292 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 173, + 288, + 292 + ], + "type": "inline_equation", + "content": "1.79\\%" + }, + { + "bbox": [ + 46, + 173, + 288, + 292 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 293, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 437 + ], + "type": "text", + "content": "Comparison against existing methods: Since each component model is trained only on task " + }, + { + "bbox": [ + 46, + 293, + 287, + 437 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 293, + 287, + 437 + ], + "type": "text", + "content": ", by design, there is no interference from other tasks when training " + }, + { + "bbox": [ + 46, + 293, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\tilde{h}_t" + }, + { + "bbox": [ + 46, + 293, + 287, + 437 + ], + "type": "text", + "content": ". As such, the final composed model is equivalent to a specialist model when restricted to any task. However, in other state of the art continual learning methods such as those involving replay or loss regularization, this separation among trained for each task is violated. Hence, under the task-incremental setting, we show in Table 4 that TMC achieves the best performance across both replay-free and replay-based methods even without using any memory buffer, beating the next best method by " + }, + { + "bbox": [ + 46, + 293, + 287, + 437 + ], + "type": "inline_equation", + "content": "2.46\\%" + }, + { + "bbox": [ + 46, + 293, + 287, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 293, + 287, + 437 + ], + "type": "inline_equation", + "content": "1.38\\%" + }, + { + "bbox": [ + 46, + 293, + 287, + 437 + ], + "type": "text", + "content": " respectively [6]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 445, + 142, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 142, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 142, + 456 + ], + "type": "text", + "content": "4.3. Further studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 463, + 287, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 512 + ], + "type": "text", + "content": "We present additional studies on the Rescaled Square Loss (Sec. 4.3.1), and compare against TMC-Seq to show that our method yields an effective and highly parallelizable training scheme for continual fine-tuning (Sec. 4.3.2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 527, + 173, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 173, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 173, + 540 + ], + "type": "text", + "content": "4.3.1 Rescaled Square Loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 546, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 654 + ], + "type": "text", + "content": "In Sec. 3.3, we showed that " + }, + { + "bbox": [ + 46, + 546, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 546, + 287, + 654 + ], + "type": "text", + "content": " can be used to control the amount of interaction between tasks. In Fig. 3, we empirically show that in the class-incremental setting where tasks are highly dissimilar, larger values of " + }, + { + "bbox": [ + 46, + 546, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 546, + 287, + 654 + ], + "type": "text", + "content": " result in a better generalist model. On the other hand, under the data-incremental setting where tasks are more similar, information from other tasks can benefit the learning of any given task. Here, larger values of " + }, + { + "bbox": [ + 46, + 546, + 287, + 654 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 546, + 287, + 654 + ], + "type": "text", + "content": " instead harm generalization due to lowering the accuracy of each component model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 670, + 211, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 211, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 211, + 682 + ], + "type": "text", + "content": "4.3.2 Convexity and Parallelizability" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Given a sequence of tasks " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "1 \\ldots T" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " and a pre-trained model " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": ", there are at least two natural ways to learn the compo" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 174, + 525, + 324 + ], + "blocks": [ + { + "bbox": [ + 328, + 174, + 525, + 324 + ], + "lines": [ + { + "bbox": [ + 328, + 174, + 525, + 324 + ], + "spans": [ + { + "bbox": [ + 328, + 174, + 525, + 324 + ], + "type": "image", + "image_path": "f30f3019419ce173f201bee14ae28f270789489cb3996979412221091f617ea3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 330, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 407 + ], + "type": "text", + "content": "Figure 4. Accuracy of composed models, where component models are initialized with " + }, + { + "bbox": [ + 304, + 330, + 545, + 407 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 304, + 330, + 545, + 407 + ], + "type": "text", + "content": " (TMC) and " + }, + { + "bbox": [ + 304, + 330, + 545, + 407 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 304, + 330, + 545, + 407 + ], + "type": "text", + "content": " (TMC-Seq). In the data-incremental setting (D), TMC-Seq outperforms TMC, while in the class-incremental setting (C), TMC performs significantly better. All results are averaged across D/C-Caltech-256, D/C-MIT-67, D/C-OxfordPets (we exclude C-OxfordPets in the 20-task setting since it does not contain enough classes)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "text", + "content": "nent model " + }, + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\tilde{h}_t" + }, + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "text", + "content": ": (1) Train " + }, + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\tilde{h}_t" + }, + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "text", + "content": " as initialization, and (2) Train " + }, + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\tilde{h}_t" + }, + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 304, + 413, + 545, + 557 + ], + "type": "text", + "content": " as initialization. The first option (TMC-Seq) is most common among continual learning methods which impose regularization in the weights between the current and previous tasks or optimize using memory buffers consisting of examples obtained from previous tasks. As such, these methods are restricted to a sequential learning framework, and hence cannot be parallelized across tasks. In Sec. 3, we showed that, thanks to the convexity of the loss landscape, both options (1) and (2) are equivalent under our framework when we assume comparable generalization performance among global optima." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "In practice, for massively over-parameterized networks, such assumptions do not hold. However, we show empirically in Fig. 4 that under the data-incremental setting, compared to TMC-Seq, TMC only incurs a small loss in generalization performance to enable massive parallelization across tasks. This can potentially provide orders of magnitude speed-ups compared to existing continual learning methods. Moreover, under the class-incremental setting, TMC greatly outperforms TMC-Seq, showing that the pre-trained initialization yields solutions of better generalization than " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_{t - 1}" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": ". We attribute this to the fact that the hypothesis space of models which are globally optimal for the training data is vast and yields varying generalization." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18683" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "Thus, regardless of the loss landscape convexity, convergence to each point in this space depends on the \"direction of approach\". Hence, initialization at " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " introduces a bias towards the previous tasks. Due to the high level of dissimilarity between tasks in the class-incremental setting, this bias harms the final model generalization performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 156, + 117, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 117, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 117, + 168 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 177, + 287, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 177, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 287, + 260 + ], + "type": "text", + "content": "Our approach enables incremental learning under the assumption that increment tasks are \"close\" to the pre-trained model. We ensure proximity by training increment models independently on the tangent plane of the pre-trained model. The advantage is that pre-trained tasks are not forgotten, and increment tasks can be learned independently, in parallel, and easily forgotten if needed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 261, + 288, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 288, + 369 + ], + "type": "text", + "content": "Of course, this approach does not address the problem of incremental learning in full generality, when increment tasks can be arbitrary and arbitrarily unrelated. Nonetheless, we have shown that our method is competitive with existing methods in some of the most challenging settings, for instance when the hypothesis spaces of component tasks are disjoint. On the positive side, despite not requiring a replay buffer, our method outperforms replay-based ones under the assumptions in which we operate." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 288, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 288, + 489 + ], + "type": "text", + "content": "A limitation of our approach is that it requires computing the linear span of pre-trained models, which can be challenging. However, for common deep network architectures such as the ResNet [21] or Transformer [52, 15] family, the Jacobian-Vector product can be efficiently evaluated [39, 35, 1, 31], reducing the computational complexity of training the tangent model to that of the same magnitude as a linear classifier on network activations. Hence, inference time for the linearized network is at most double that of the original model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "content": "Another limitation of our approach is assuming a first-order expansion around a pre-trained objective can well approximate the fine-tuned model. Such assumptions hold poorly when the pre-training and downstream tasks are highly unrelated. Nevertheless, an ImageNet [13] pretrained model already yields a sufficiently good local approximation for various real-world datasets [1, 31]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 573, + 288, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 288, + 645 + ], + "type": "text", + "content": "While we mainly explored linear compositionality in the context of continual learning, we note that this framework can be easily generalized. We discussed possible applications towards federated learning and forgetting, since our method naturally yields a parallel training framework that fully compartmentalizes information obtained each task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 657, + 162, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 162, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 162, + 671 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "We would like to thank Aditya Golatkar, Albert Zhao, and the anonymous reviewers for their feedback on the initial version of the paper. This work was supported by ARO" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 514, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 514, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 514, + 83 + ], + "type": "text", + "content": "W911NF-17-1-0304 and ONR N00014-22-1-2252." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 94, + 365, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 94, + 365, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 94, + 365, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 114, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 311, + 114, + 547, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 114, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 311, + 114, + 547, + 169 + ], + "type": "text", + "content": "[1] Alessandro Achille, Aditya Golatkar, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. Lqf: Linear quadratic fine-tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15729-15739, 2021. 1, 2, 4, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 170, + 547, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 170, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 311, + 170, + 547, + 224 + ], + "type": "text", + "content": "[2] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 225, + 547, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 225, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 311, + 225, + 547, + 269 + ], + "type": "text", + "content": "[3] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3366-3375, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 270, + 547, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 270, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 311, + 270, + 547, + 312 + ], + "type": "text", + "content": "[4] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 314, + 547, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 314, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 311, + 314, + 547, + 346 + ], + "type": "text", + "content": "[5] Elie Bienenstock and Stuart Geman. Compositionality in neural systems. In The handbook of brain theory and neural networks, pages 223-226. 1998. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 347, + 547, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 347, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 311, + 347, + 547, + 422 + ], + "type": "text", + "content": "[6] Matteo Boschini, Lorenzo Bonicelli, Angelo Porrello, Giovanni Bellitto, Matteo Pennisi, Simone Palazzo, Concetto Spampinato, and Simone Calderara. Transfer without forgetting. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIII, pages 692-709. Springer, 2022. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 424, + 547, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 424, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 311, + 424, + 547, + 479 + ], + "type": "text", + "content": "[7] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 3, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 480, + 547, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 480, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 311, + 480, + 547, + 534 + ], + "type": "text", + "content": "[8] Pietro Buzzega, Matteo Boschini, Angelo Porrello, and Simone Calderara. Rethinking experience replay: a bag of tricks for continual learning. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 2180-2187. IEEE, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 311, + 536, + 547, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 536, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 311, + 536, + 547, + 579 + ], + "type": "text", + "content": "[9] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuyte-laars, Joelle Pineau, and Eugene Belilovsky. New insights on reducing abrupt representation change in online continual learning. arXiv preprint arXiv:2203.03798, 2022. 3, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 580, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 623 + ], + "type": "text", + "content": "[10] Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. Co2l: Contrastive continual learning. In Proceedings of the IEEE/CVF International conference on computer vision, pages 9516-9525, 2021. 3, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 624, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 657 + ], + "type": "text", + "content": "[11] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "text", + "content": "[12] Leshem Choshen, Elad Venezian, Noam Slonim, and Yoav Katz. Fusing finetuned models for better pretraining. arXiv preprint arXiv:2204.03044, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18684" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 9" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[14] Thomas G Dietterich. Ensemble methods in machine learning. In International workshop on multiple classifier systems, pages 1-15. Springer, 2000. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 195 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 241 + ], + "type": "text", + "content": "[16] Jonathan Frankle, Gintare Karolina Dziugaite, Daniel Roy, and Michael Carbin. Linear mode connectivity and the lottery ticket hypothesis. In International Conference on Machine Learning, pages 3259-3269. PMLR, 2020. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 242, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 287, + 286 + ], + "type": "text", + "content": "[17] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry P Vetrov, and Andrew G Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 342 + ], + "type": "text", + "content": "[18] Aditya Golatkar, Alessandro Achille, Avinash Ravichandran, Marzia Polito, and Stefano Soatto. Mixed-privacy forgetting in deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 792–801, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "type": "text", + "content": "[19] Raphael Gontijo-Lopes, Yann Dauphin, and Ekin D Cubuk. No one representation to rule them all: Overlapping features of training methods. arXiv preprint arXiv:2110.12899, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 410 + ], + "type": "text", + "content": "[20] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "text", + "content": "[22] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831-839, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 545 + ], + "type": "text", + "content": "[23] Like Hui and Mikhail Belkin. Evaluation of neural architectures trained with square loss vs cross-entropy in classification tasks. arXiv preprint arXiv:2006.07322, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "text", + "content": "[24] Gabriel Ilharco, Mitchell Wortsman, Samir Yitzhak Gadre, Shuran Song, Hannaneh Hajishirzi, Simon Kornblith, Ali Farhadi, and Ludwig Schmidt. Patching open-vocabulary models by interpolating weights. arXiv preprint arXiv:2208.05592, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "text", + "content": "[25] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. arXiv preprint arXiv:1803.05407, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[27] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 152 + ], + "type": "text", + "content": "[28] Janghyeon Lee, Hyeong Gwon Hong, Donggyu Joo, and Junmo Kim. Continual learning with extended kronecker-factored approximate curvature. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9001-9010, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 186 + ], + "type": "text", + "content": "[29] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 188, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 188, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 188, + 545, + 220 + ], + "type": "text", + "content": "[30] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 3, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 255 + ], + "type": "text", + "content": "[31] Tian Yu Liu, Aditya Golatkar, and Stefano Soatto. Tangent transformers for composition, privacy and removal. arXiv preprint arXiv:2307.08122, 2023. 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 257, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 257, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 257, + 545, + 300 + ], + "type": "text", + "content": "[32] Tian Yu Liu, Aditya Golatkar, Stefano Soatto, and Alessandro Achille. Integral continual learning along the tangent vector field of tasks. arXiv preprint arXiv:2211.13108, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "type": "text", + "content": "[33] Arun Mallya and Svetlana Lazebnik. Packnet: Adding multiple tasks to a single network by iterative pruning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7765-7773, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 348, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 348, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 348, + 545, + 380 + ], + "type": "text", + "content": "[34] Michael Matena and Colin Raffel. Merging models with fisher-weighted averaging. arXiv preprint arXiv:2111.09832, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "text", + "content": "[35] Fangzhou Mu, Yingyu Liang, and Yin Li. Gradients as features for deep representation learning. arXiv preprint arXiv:2004.05529, 2020. 2, 4, 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 418, + 545, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 418, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 307, + 418, + 545, + 450 + ], + "type": "text", + "content": "[36] Cuong V Nguyen, Yingzhen Li, Thang D Bui, and Richard E Turner. Variational continual learning. arXiv preprint arXiv:1710.10628, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 452, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 452, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 452, + 545, + 495 + ], + "type": "text", + "content": "[37] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural Networks, 113:54-71, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 498, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 540 + ], + "type": "text", + "content": "[38] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 544, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 565 + ], + "type": "text", + "content": "[39] Barak A Pearlmutter. Fast exact multiplication by the hessian. Neural computation, 6(1):147-160, 1994. 4, 9" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 567, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 622 + ], + "type": "text", + "content": "[40] Grégoire Petit, Adrian Popescu, Hugo Schindler, David Picard, and Bertrand Delezoide. Fetril: Feature translation for exemplar-free class-incremental learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3911–3920, 2023. 3, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "text", + "content": "[41] Ameya Prabhu, Philip HS Torr, and Puneet K Dokania. Gdumb: A simple approach that questions our progress in continual learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 524-540. Springer, 2020. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[42] Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In 2009 IEEE conference on computer vision and pattern recognition, pages 413-420. IEEE, 2009. 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18685" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[43] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "text", + "content": "[44] Matthew Riemer, Ignacio Cases, Robert Ajemian, Miao Liu, Irina Rish, Yuhai Tu, and Gerald Tesauro. Learning to learn without forgetting by maximizing transfer and minimizing interference. arXiv preprint arXiv:1810.11910, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "text", + "content": "[45] Hippolyt Ritter, Aleksandar Botev, and David Barber. Online structured laplace approximations for overcoming catastrophic forgetting. Advances in Neural Information Processing Systems, 31, 2018. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "text", + "content": "[46] Anthony Robins. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7(2):123-146, 1995. 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "type": "text", + "content": "[47] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 358 + ], + "type": "text", + "content": "[48] Jonathan Schwarz, Wojciech Czarnecki, Jelena Luketina, Agnieszka Grabska-Barwinska, Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. Progress & compress: A scalable framework for continual learning. In International conference on machine learning, pages 4528-4537. PMLR, 2018. 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 425 + ], + "type": "text", + "content": "[49] Hyounguk Shon, Janghyeon Lee, Seung Hwan Kim, and Junmo Kim. Dlcft: Deep linear continual fine-tuning for general incremental learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 513-529. Springer, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "text", + "content": "[50] Doris Y Tsao, Winrich A Freiwald, Tamara A Knutsen, Joseph B Mandeville, and Roger BH Tootell. Faces and objects in macaque cerebral cortex. Nature neuroscience, 6(9):989-995, 2003. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 503 + ], + "type": "text", + "content": "[51] Gido M Van de Ven and Andreas S Tolias. Three scenarios for continual learning. arXiv preprint arXiv:1904.07734, 2019. 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 548 + ], + "type": "text", + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 592 + ], + "type": "text", + "content": "[53] Johannes Von Oswald, Seijin Kobayashi, Joao Sacramento, Alexander Meulemans, Christian Henning, and Benjamin F Grewe. Neural networks with late-phase weights. arXiv preprint arXiv:2007.12927, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 658 + ], + "type": "text", + "content": "[54] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 714 + ], + "type": "text", + "content": "[55] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. arXiv preprint arXiv:2204.04799, 2022. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 485 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 138 + ], + "type": "text", + "content": "[56] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 546, + 184 + ], + "type": "text", + "content": "[57] Mitchell Wortsman, Suchin Gururangan, Shen Li, Ali Farhadi, Ludwig Schmidt, Michael Rabbat, and Ari S Morcos. lo-fi: distributed fine-tuning without communication. arXiv preprint arXiv:2210.11948, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 185, + 546, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 546, + 261 + ], + "type": "text", + "content": "[58] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 2, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 261, + 546, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 546, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 546, + 329 + ], + "type": "text", + "content": "[59] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7959-7971, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 329, + 546, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 329, + 546, + 372 + ], + "spans": [ + { + "bbox": [ + 308, + 329, + 546, + 372 + ], + "type": "text", + "content": "[60] Fei Zhu, Zhen Cheng, Xu-Yao Zhang, and Cheng-lin Liu. Class-incremental learning via dual augmentation. Advances in Neural Information Processing Systems, 34:14306-14318, 2021. 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 373, + 546, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 373, + 546, + 428 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 546, + 428 + ], + "type": "text", + "content": "[61] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5871-5880, 2021. 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 428, + 547, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 428, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 547, + 485 + ], + "type": "text", + "content": "[62] Kai Zhu, Wei Zhai, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Self-sustaining representation expansion for nonexemplar class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9296-9305, 2022. 3, 7, 8" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18686" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/636b4c99-6986-4080-bf56-9a578e79f107_content_list.json b/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/636b4c99-6986-4080-bf56-9a578e79f107_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f71c369c20ef53a7c6655243f416f6a52b416065 --- /dev/null +++ b/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/636b4c99-6986-4080-bf56-9a578e79f107_content_list.json @@ -0,0 +1,2089 @@ +[ + { + "type": "text", + "text": "Tangent Sampson Error: Fast Approximate Two-view Reprojection Error for Central Camera Models", + "text_level": 1, + "bbox": [ + 99, + 131, + 870, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mikhail Terekhov ETH Zurich", + "bbox": [ + 259, + 203, + 406, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mterekhov@ethz.ch", + "bbox": [ + 256, + 242, + 410, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Viktor Larsson", + "bbox": [ + 527, + 203, + 655, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lund University", + "bbox": [ + 527, + 222, + 658, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "viktor.larsson@math.lth.se", + "bbox": [ + 473, + 242, + 707, + 255 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fafdb5c2548aa683e1fc4b744598dc60f15672810a3bf881ffb1b04cd5c31455.jpg", + "image_caption": [ + "Figure 1. The figure shows the level sets of the proposed Tangent Sampson error. In contrast to classical Sampson error, the proposed error function applies to arbitrary central camera models; in this case a 360 panoramic image and a fisheye camera. Given a relative pose and a pair of points (highlighted in red), we vary a point on one image while keeping the other fixed, and thus get the values on each image. A given color corresponds to a constant value of reprojection error, with purple indicating near-zero errors along the epipolar curve." + ], + "image_footnote": [], + "bbox": [ + 133, + 295, + 598, + 476 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2bc68f527d4fd5a4e586c6eff8e4d9739f7e44827fa75707f9ae1a038caf3244.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 306, + 838, + 476 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 560, + 313, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper we introduce the Tangent Sampson error, which is a generalization of the classical Sampson error in two-view geometry that allows for arbitrary central camera models. It only requires local gradients of the distortion map at the original correspondences (allowing for pre-computation) resulting in a negligible increase in computational cost when used in RANSAC or local refinement. The error effectively approximates the true-reprojection error for a large variety of cameras, including extremely wide field-of-view lenses that cannot be undistorted to a single pinhole image. We show experimentally that the new error outperforms competing approaches both when used for model scoring in RANSAC and for non-linear refinement of the relative camera pose.", + "bbox": [ + 73, + 592, + 473, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 829, + 209, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating two-view geometry from a set of sparse 2D-2D correspondences is a common sub-problem in many vision applications. Each correspondence $(\\pmb{p}_1, \\pmb{p}_2) \\in \\mathbb{R}^2 \\times \\mathbb{R}^2$", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "constrains the geometry via the epipolar constraint,", + "bbox": [ + 500, + 561, + 841, + 577 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\boldsymbol {p} _ {2}; 1\\right) ^ {\\top} \\boldsymbol {E} \\left(\\boldsymbol {p} _ {1}; 1\\right) = 0, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 584, + 890, + 601 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $\\pmb{E} = [t]_{\\times}\\pmb{R} \\in \\mathbb{R}^{3\\times 3}$ is the essential matrix and the points $p_1, p_2$ are on the normalized image plane. This constraint ensures that the 2D-2D correspondence can be exactly triangulated (i.e. with zero reprojection error) using the camera pair extracted from $\\pmb{E}$ . However when the measurement $(p_1, p_2)$ is inexact, it is natural to measure how close this correspondence is to being consistent with a given epipolar geometry. This is for example used for determining inlier matches during robust estimation, or as a loss to be minimized when refining an initial estimate of $\\pmb{E}$ .", + "bbox": [ + 496, + 606, + 892, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Under the assumption of i.i.d. Gaussian noise, the ML-estimate of the closest consistent correspondence, sometimes called the gold-standard estimate in the literature [12], is found by minimizing the squared reprojection error,", + "bbox": [ + 498, + 758, + 893, + 819 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {M L} ^ {2} = \\min _ {\\boldsymbol {X} \\in \\mathbb {R} ^ {3}} \\| \\boldsymbol {p} _ {1} - \\pi (\\boldsymbol {X}) \\| ^ {2} + \\| \\boldsymbol {p} _ {2} - \\pi (\\boldsymbol {R} \\boldsymbol {X} + t) \\| ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 839, + 890, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $\\pi : \\mathbb{R}^3 \\to \\mathbb{R}^2$ is the camera projection function. For identical pinhole cameras with unit aspect ratio and no", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3370", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "skew, this can be reformulated as finding the corrected normalized image points, i.e. the closest pair of 2D points that satisfy the epipolar constraint exactly,", + "bbox": [ + 76, + 90, + 470, + 137 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {E} _ {M L} ^ {2} = \\min _ {\\hat {\\boldsymbol {p}} _ {1}, \\hat {\\boldsymbol {p}} _ {2} \\in \\mathbb {R} ^ {2}} \\| \\boldsymbol {p} _ {1} - \\hat {\\boldsymbol {p}} _ {1} \\| ^ {2} + \\| \\boldsymbol {p} _ {2} - \\hat {\\boldsymbol {p}} _ {2} \\| ^ {2} \\tag {3} \\\\ \\begin{array}{l} \\text {s . t .} \\quad (\\hat {\\boldsymbol {p}} _ {2}; 1) ^ {\\intercal} \\boldsymbol {E} (\\hat {\\boldsymbol {p}} _ {1}; 1) = 0. \\end{array} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 147, + 468, + 193 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The closed-form solution to (3) can be found as a root to a degree-six polynomial (see [13]), or via iterative methods [20]. In robust estimators, e.g. RANSAC, it is often necessary to evaluate these residuals thousands (or potentially millions) of times for each image pair, which in practice means that cheaper approximate error metrics are favored.", + "bbox": [ + 75, + 205, + 468, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These approximate or surrogate error metrics can be divided into two categories: image-based (approximating the true reprojection error) or geometry-based (measuring some geometric quantity, e.g. angles, instead). In the first category we have the epipolar distance, i.e. distance from the point to the corresponding epipolar line. This can be interpreted as assuming zero noise on one of the keypoints. The error is often computed symmetrically and thus called the symmetric epipolar distance, having the form", + "bbox": [ + 75, + 297, + 468, + 434 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {S E D} ^ {2} = \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| \\boldsymbol {E} _ {1 2} \\boldsymbol {d} _ {1} \\right\\| ^ {2}} + \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| (\\boldsymbol {E} ^ {\\intercal}) _ {1 2} \\boldsymbol {d} _ {2} \\right\\| ^ {2}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 444, + 468, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $E_{12} \\in \\mathbb{R}^{2 \\times 3}$ refers to the first two rows of $\\pmb{E}$ , and $\\pmb{d}_k = (\\pmb{p}_k; 1) \\in \\mathbb{R}^3$ . Another image-based error is the Sampson error, which is a first-order approximation of $\\mathcal{E}_{ML}$ (see Sec. 2.1 for details):", + "bbox": [ + 76, + 489, + 468, + 551 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {S} ^ {2} = \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| \\boldsymbol {E} _ {1 2} \\boldsymbol {d} _ {1} \\right\\| ^ {2} + \\left\\| (\\boldsymbol {E} ^ {\\intercal}) _ {1 2} \\boldsymbol {d} _ {2} \\right\\| ^ {2}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 561, + 468, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Both of these errors can after appropriate rescaling (using the focal lengths) be expressed in pixels and thus are natural to interpret in terms of the keypoint detection noise levels.", + "bbox": [ + 76, + 606, + 468, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the second category, an example is the cosine error,", + "bbox": [ + 96, + 652, + 455, + 667 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {C S} ^ {2} = \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| \\boldsymbol {d} _ {2} \\right\\| ^ {2} \\left\\| \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right\\| ^ {2}} + \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| \\boldsymbol {E} ^ {\\intercal} \\boldsymbol {d} _ {2} \\right\\| ^ {2} \\left\\| \\boldsymbol {d} _ {1} \\right\\| ^ {2}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 679, + 468, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which (symmetrically) measures the cosine between the normal of the epipolar plane $\\pmb{E}\\pmb{d}_1 \\in \\mathbb{R}^3$ and corresponding point $d_2 \\in \\mathbb{R}^3$ . Another geometric error is the algebraic residual,", + "bbox": [ + 76, + 724, + 468, + 782 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {A L G} = \\left| \\boldsymbol {d} _ {2} ^ {\\mathrm {T}} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right|, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 785, + 468, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which was given a geometric interpretation in [19].", + "bbox": [ + 76, + 809, + 413, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most of the literature has focused on the pinhole camera case, and the image-based errors above $(\\mathcal{E}_{SED}$ and $\\mathcal{E}_S)$ measure the error in the image plane assuming this model. For non-pinhole cameras, we either have to undistort the image such that $\\mathcal{E}_{SED}$ and $\\mathcal{E}_S$ are applicable, or use the", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fb436acb963ff1442f198de14775181020643ebdeb8ace826562f59690f7b384.jpg", + "image_caption": [ + "Figure 2. Undistorting the image deforms the level-sets of the reprojection error and introduces a radial bias in the optimization." + ], + "image_footnote": [], + "bbox": [ + 506, + 88, + 694, + 198 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/63be9b5cc0d75ed8e40990fe8c129ea9506a70f1c529bf9f99521b86e25b1e1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 88, + 885, + 198 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "geometry-based error metrics (e.g. $\\mathcal{E}_{CS}$ or $\\mathcal{E}_{ALG}$ ). However, undistorting to a pinhole image (which is not always possible for large field-of-view cameras) will introduce a radial-bias into the error (see Figure 2).", + "bbox": [ + 498, + 244, + 890, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper we instead propose the Tangent Sampson error $\\mathcal{E}_{TS}$ , a version of the classical Sampson error that has been generalized to handle arbitrary camera models. Compared to current alternatives it has the following benefits;", + "bbox": [ + 498, + 305, + 890, + 364 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- It is cheap to evaluate (similar in cost to the original Sampson error) since it only requires computing the camera model forward-projection and Jacobian once (which can be pre-computed for each point).", + "- Since it measures errors in the original image, it is easier to set meaningful thresholds in pixels. This is in contrast to geometric error functions, where it is not always clear how to relate the thresholds (e.g. in angles) to the keypoint detection noise in pixels.", + "- It works for arbitrary central camera models, only requiring that we can compute a Jacobian of the camera projection function at each correspondence. This also allows the formulation to work for cameras modeled via look-up tables, where the Jacobian can be estimated via interpolation of finite differences." + ], + "bbox": [ + 517, + 375, + 890, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Section 2.1 we first recall the derivations for the original Sampson error. Next, in Section 3 we propose our generalization to arbitrary central camera models, and in Section 4 we experimentally show that the proposed error outperforms the competing methods on a variety of camera and lens configurations.", + "bbox": [ + 498, + 619, + 890, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 724, + 638, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Sampson error was introduced in [23] to approximate the geometric distance from a point to a conic. The proposed approach was to linearize the equation of the conic around the given point, rendering a closed-form solution for the new distance, now from the point to a line. This idea proved useful for estimating distances to other implicit surfaces, notably those set by constraints of a fundamental matrix [22].", + "bbox": [ + 498, + 750, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The fundamental matrix is only applicable to perspective cameras, and a lot of subsequent work in computer vision", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3371", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "focused on generalizing it to non-linear camera models, together with some notion of a reprojection error. A well-known example is the division model from Fitzgibbon [8]. It maps a point $\\pmb{p}$ on the normalized image plane to the corresponding bearing vector $\\pmb{d}$ by", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {d} = (\\boldsymbol {p}; 1 + \\lambda \\| \\boldsymbol {p} \\| ^ {2}). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 175, + 468, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The original paper [8] already discusses a way of approximating the gold-standard error for the new model: first, undistort the original points of the correspondence; next, apply the closed-form triangulation from [13], and distort the projections back onto the original image. The resulting points are then approximate minimizers of (2). A faster iterative scheme for the optimal error was later proposed in [16], which adapts the idea of [20], originally developed for pinhole cameras, to the division model.", + "bbox": [ + 75, + 202, + 468, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several camera models with closed-form expressions for epipolar curves were proposed. In [25], epipolar curves are represented by cubics, in the rational model [5] — by conics, and in [2] — by circles. With such models, we can use the Sampson approximation to estimate the distance from one point of the correspondence to the epipolar curve of the other, and vice versa. Note that this gives us an approximation of the symmetric epipolar distance, not the Sampson error, because for the latter we would need to optimize both point positions simultaneously, and this would move the epipolar curves as well.", + "bbox": [ + 75, + 338, + 468, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All these models were proposed with the goal of searching for the intrinsic parameters together with the epipolar geometry. Indeed, [2] and [5] even provide their own versions of a fundamental matrix on a lifted image space. These matrices can be estimated directly from point correspondences, and then decomposed to provide the distortion parameters and relative pose up to an irreducible ambiguity. We, on the other hand, constrain ourselves to the calibrated case, and thus can afford much greater generality.", + "bbox": [ + 75, + 505, + 468, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are also various heuristic generalizations of the symmetric epipolar distance to general non-linear camera models that have been used in practice. One approach is to project the bearing vector of one point of the correspondence onto the epipolar plane given by the other. The projected vector is then mapped back onto the image to provide an approximate corrected point, for which the error can be computed. Mathematically, this is expressed as", + "bbox": [ + 75, + 641, + 468, + 760 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {n} _ {1} = \\boldsymbol {E} \\boldsymbol {d} _ {1} / \\| \\boldsymbol {E} \\boldsymbol {d} _ {1} \\|, \\quad \\boldsymbol {n} _ {2} = \\boldsymbol {E} ^ {\\top} \\boldsymbol {d} _ {2} / \\| \\boldsymbol {E} ^ {\\top} \\boldsymbol {d} _ {2} \\|, \\tag {9} \\\\ \\mathcal {E} _ {\\pi S E D} ^ {2} = \\left\\| \\boldsymbol {p} _ {1} - \\pi \\left(\\boldsymbol {d} _ {1} - \\boldsymbol {n} _ {2} \\boldsymbol {n} _ {2} ^ {\\intercal} \\boldsymbol {d} _ {1}\\right) \\right\\| ^ {2} \\\\ + \\left\\| \\boldsymbol {p} _ {2} - \\pi \\left(\\boldsymbol {d} _ {2} - \\boldsymbol {n} _ {1} \\boldsymbol {n} _ {1} ^ {\\intercal} \\boldsymbol {d} _ {2}\\right) \\right\\| ^ {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 772, + 467, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p_1$ and $p_2$ are the points on the images, $d_1$ and $d_2$ are the corresponding bearing vectors, $n_1$ and $n_2$ are the normals of the epipolar planes. We have encountered this formulation in private codebases, but we are not aware of its", + "bbox": [ + 75, + 840, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "origin. We call it the projective symmetric epipolar error. While this does not give us exact distances to the epipolar curves even in the pinhole case, it often provides a reasonable approximation. The main problem with this approach is that it requires calculating $\\pi(p)$ every time, making it rather computationally heavy for some camera models.", + "bbox": [ + 496, + 90, + 890, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the probabilistic perspective, the geometric error is only justified as long as the noise in the correspondences stays standard Gaussian and i.i.d. A natural generalization is to allow for arbitrary covariances in point positions. For example, when the keypoints are searched on a multi-resolution image pyramid, the standard deviation of the position can be taken as the downscaling factor of the level on which the point was found. For conics and fundamental matrices, covariance-aware error formulations were derived in [3]. We will show how this approach is generalized to our proposed error formulation in Section 3.2.", + "bbox": [ + 496, + 181, + 890, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In [11] the authors propose another first-order approximation of the reprojection error. The method builds on the optimal triangulation method from [13] and differentates the sixth-order polynomial that has the optimal geometric error as its first root. The first-order approximation is then given by one step of the Newton-Raphson iteration on this polynomial starting from zero. In the paper, this error is only derived for conics. While it would also work for fundamental matrices, this does not generalize to the nonlinear camera model setting we are interested in.", + "bbox": [ + 496, + 347, + 890, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we would like to note that the Sampson error can also be derived for homographies [4]. Our method is applicable there as well, but with larger computational cost, which somewhat defeats its original purpose.", + "bbox": [ + 496, + 500, + 890, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. The Sampson Approximation", + "text_level": 1, + "bbox": [ + 500, + 568, + 761, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional Sampson error [22, 23] approximates the geometric error, i.e. the distance to the closest pair of points which satisfy the epipolar constraint. Given a correspondence $\\mathbf{z} = (\\pmb{p}_1, \\pmb{p}_2) \\in \\mathbb{R}^2 \\times \\mathbb{R}^2$ and essential matrix $\\pmb{E} \\in \\mathbb{R}^{3 \\times 3}$ , this can be formalized as", + "bbox": [ + 496, + 590, + 890, + 666 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {E} _ {M L} ^ {2} (\\boldsymbol {z}, \\boldsymbol {E}) = \\min _ {\\hat {\\boldsymbol {z}}} \\| \\hat {\\boldsymbol {z}} - \\boldsymbol {z} \\| ^ {2} \\tag {11} \\\\ \\begin{array}{c c} \\text {s . t .} & C (\\hat {z}) = 0, \\end{array} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 676, + 888, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{z}} = (\\hat{\\pmb{p}}_1,\\hat{\\pmb{p}}_2)\\in \\mathbb{R}^2\\times \\mathbb{R}^2$ is the corrected correspondence and the epipolar constraint is", + "bbox": [ + 496, + 729, + 888, + 760 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC (\\hat {\\boldsymbol {z}}) = \\left(\\hat {\\boldsymbol {p}} _ {2}; 1\\right) ^ {\\intercal} \\boldsymbol {E} \\left(\\hat {\\boldsymbol {p}} _ {1}; 1\\right) = 0. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 771, + 888, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since (11) does not admit a simple closed form solution, the idea in [23] is to linearize the epipolar constraint, $C(\\hat{z}) = 0$ at the original correspondence $\\mathcal{Z}$ , i.e.", + "bbox": [ + 496, + 799, + 888, + 843 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {S} ^ {2} (\\boldsymbol {z}, \\boldsymbol {E}) = \\min _ {\\hat {\\boldsymbol {z}}} \\| \\hat {\\boldsymbol {z}} - \\boldsymbol {z} \\| ^ {2} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 853, + 888, + 876 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s . t .} \\quad C (\\boldsymbol {z}) + \\boldsymbol {J} _ {C} (\\hat {\\boldsymbol {z}} - \\boldsymbol {z}) = 0,\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 881, + 841, + 896 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3372", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $J_{C} = \\partial C(z) / \\partial z$ is the Jacobian of the constraint, evaluated at $z$ . Introducing a Lagrangian for (13),", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\hat {z}, \\lambda) = \\| \\hat {z} - z \\| ^ {2} + \\lambda (C (z) + J _ {C} (\\hat {z} - z)), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 132, + 468, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we get the first-order constraints as", + "bbox": [ + 76, + 162, + 312, + 176 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} = z - \\frac {\\lambda}{2} \\boldsymbol {J} _ {C} ^ {\\intercal}, \\quad C (z) + \\boldsymbol {J} _ {C} (\\hat {z} - z) = 0. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 189, + 468, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inserting the first equation into the second yields", + "bbox": [ + 76, + 229, + 398, + 244 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC (\\boldsymbol {z}) - \\frac {\\lambda}{2} \\| \\boldsymbol {J} _ {C} \\| ^ {2} = 0 \\Rightarrow \\hat {\\boldsymbol {z}} = \\boldsymbol {z} - \\frac {C (\\boldsymbol {z})}{\\| \\boldsymbol {J} _ {C} \\| ^ {2}} \\boldsymbol {J} _ {C} ^ {\\intercal}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 255, + 468, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus the minimum in (13) is given by", + "bbox": [ + 76, + 299, + 330, + 314 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {S} ^ {2} = \\left\\| \\frac {C (\\boldsymbol {z})}{\\| \\boldsymbol {J} _ {C} \\| ^ {2}} \\boldsymbol {J} _ {C} ^ {\\intercal} \\right\\| ^ {2} = \\frac {C (\\boldsymbol {z}) ^ {2}}{\\| \\boldsymbol {J} _ {C} \\| ^ {2}}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 325, + 468, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Replacing $C(z)$ with the epipolar constraint, we arrive at the classical formula for the Sampson error:", + "bbox": [ + 76, + 372, + 468, + 402 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {S} ^ {2} = \\frac {\\left(\\left(\\boldsymbol {p} _ {2} ; 1\\right) ^ {\\intercal} \\boldsymbol {E} \\left(\\boldsymbol {p} _ {1} ; 1\\right)\\right) ^ {2}}{\\left\\| \\boldsymbol {E} _ {1 2} \\left(\\boldsymbol {p} _ {1} ; 1\\right) \\right\\| ^ {2} + \\left\\| \\left(\\boldsymbol {E} ^ {\\intercal}\\right) _ {1 2} \\left(\\boldsymbol {p} _ {2} ; 1\\right) \\right\\| ^ {2}}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 414, + 468, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Generalizing the Sampson Error", + "text_level": 1, + "bbox": [ + 76, + 458, + 375, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Sampson error derived in the previous section assumes a pinhole projection model. However, many cameras, in particular those with wide field-of-view, require a more complex model. One popular choice is the Kannala-Brandt [14] model in which a 3D point $\\pmb{d} = (X,Y,Z)^{\\mathrm{T}}$ in the camera coordinate system is projected as", + "bbox": [ + 76, + 484, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi (\\boldsymbol {d}) = \\frac {\\theta_ {d}}{\\sqrt {X ^ {2} + Y ^ {2}}} \\binom {f _ {x} X} {f _ {y} Y} + \\binom {c _ {x}} {c _ {y}} \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 585, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {d} = \\theta \\left(1 + k _ {1} \\theta^ {2} + k _ {2} \\theta^ {4} + k _ {3} \\theta^ {6} + k _ {4} \\theta^ {8}\\right) \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 621, + 468, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta = \\operatorname {a t a n 2} \\left(\\sqrt {X ^ {2} + Y ^ {2}}, Z\\right), \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 643, + 468, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k_{1},\\ldots ,k_{4},f_{x},f_{y},c_{x},c_{y}$ are the intrinsic parameters.", + "bbox": [ + 76, + 674, + 467, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now consider arbitrary central camera models and derive a generalization of the Sampson error. We assume that we have some forward projection function $\\pi : \\mathbb{R}^3 \\to \\Omega$ , where $\\Omega \\subset \\mathbb{R}^2$ is the image plane. In the case of the Kannala-Brandt model this is given by (19), but our derivations work for any differentiable model with central projections, i.e. $\\pi(\\lambda d) = \\pi(d) - \\forall \\lambda > 0$ . Similarly we define the (normalized) unprojection, $\\pi^{-1} : \\Omega \\to S^2$ as mapping image points onto the corresponding unit-vectors in 3D. The geometric error (11) then becomes $\\mathcal{E}_{ML}(z, E) =$", + "bbox": [ + 75, + 691, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\hat {z}} \\| \\hat {z} - z \\| ^ {2} \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 852, + 468, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s . t .} \\quad C (\\hat {\\boldsymbol {z}}) = \\pi^ {- 1} \\left(\\hat {\\boldsymbol {p}} _ {2}\\right) ^ {\\intercal} \\boldsymbol {E} \\pi^ {- 1} \\left(\\hat {\\boldsymbol {p}} _ {1}\\right) = 0. \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 878, + 468, + 896 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, most camera models only define the forward projection $\\pi$ , and unprojection $\\pi^{-1}$ is solved for iteratively. For example, (19) does not have a closed-form inverse. This makes it non-trivial to directly compute its Jacobians required for the Sampson approximation. To avoid this computation we rely on the inverse function theorem. The variant that we will use can be found in the book [6]. It is tailored to functions that map higher-dimensional inputs to lower-dimensional outputs. We rephrase it here for reader's convenience, adapting the notation to our needs and losing some generality:", + "bbox": [ + 500, + 90, + 890, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. [6, Theorem 1E.6] Let $\\pi : \\mathbb{R}^3 \\to \\mathbb{R}^2$ be continuously differentiable in a neighborhood around $d \\in \\mathbb{R}^3$ . Further, suppose that its Jacobian $\\pmb{J}$ at $d$ has full rank. Then, there exists a neighborhood $V$ of $\\bar{\\pmb{p}} = \\pi(d)$ and a continuously differentiable function $\\widetilde{\\pi^{-1}}: V \\to \\mathbb{R}^3$ such that $\\pi(\\widetilde{\\pi^{-1}}(p)) = p \\forall p \\in V$ and whose Jacobian is given by $\\pmb{J}^\\dagger$ , the Moore-Penrose pseudoinverse of $\\pmb{J}$ .", + "bbox": [ + 498, + 262, + 890, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The construction of $\\widetilde{\\pi^{-1}}$ in the proof from [6] also gives us that $\\widetilde{\\pi^{-1}} (\\bar{p}) = d$ . Note that $\\pi$ admits many right-inverses, but they only differ in lengths assigned to unprojected bearing vectors. Thus, $\\widetilde{\\pi^{-1}}$ that the theorem gives us need not coincide with $\\pi^{-1}$ that maps points to unit vectors. However, the following result shows that when we evaluate $J$ at a unit vector $d = \\pi^{-1}(\\bar{p})$ , $J^{\\dagger}$ actually gives us the Jacobian of $\\pi^{-1}$ :", + "bbox": [ + 498, + 380, + 890, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Corollary 1. Let $\\pmb{d} \\in S^2$ and $\\pmb{J} = \\frac{\\partial\\pi}{\\partial\\bar{\\pmb{d}}} \\in \\mathbb{R}^{2\\times 3}$ . Then the Jacobian of $\\pi^{-1}$ at $\\bar{\\pmb{p}} = \\pi (\\pmb{d})$ is given by $\\pmb{J}^{\\dagger}$ .", + "bbox": [ + 498, + 513, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. Since the only difference between $\\widetilde{\\pi^{-1}}$ from theorem 1 and $\\pi^{-1}$ is the norm of the output, we can write", + "bbox": [ + 498, + 554, + 890, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi^ {- 1} (\\boldsymbol {p}) = h \\left(\\widetilde {\\left(\\pi^ {- 1} (\\boldsymbol {p})\\right)} \\quad \\forall \\boldsymbol {p} \\in V, \\right. \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 593, + 890, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h: \\pmb{v} \\mapsto \\pmb{v} / \\| \\pmb{v} \\|$ is the normalization function. We also have $\\pmb{d} = \\widetilde{\\pi^{-1}}(\\bar{\\pmb{p}})$ . Thus, the Jacobian of $\\pi^{-1}$ at $\\bar{\\pmb{p}}$ is given by", + "bbox": [ + 498, + 619, + 890, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\pi^ {- 1}}{\\partial \\boldsymbol {p}} = \\left. \\frac {\\partial h}{\\partial \\boldsymbol {v}} \\right| _ {\\boldsymbol {v} = \\boldsymbol {d}} \\frac {\\partial \\widetilde {\\pi^ {- 1}}}{\\partial \\boldsymbol {p}} = (\\boldsymbol {I} - \\boldsymbol {d} \\boldsymbol {d} ^ {\\intercal}) \\boldsymbol {J} ^ {\\dagger} = \\boldsymbol {J} ^ {\\dagger}. \\tag {25}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 674, + 890, + 712 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d401e52b9febab44d26971f331f4aed318ccecd2838658691e008ee18dee3991.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 874, + 718, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the last equation we used $\\pmb{d}^{\\mathsf{T}}\\pmb{J}^{\\dagger} = \\mathbf{0}$ , which is true in general for any null-vector $\\pmb{d}$ of the matrix $J$ , but in our case can also be seen from the following lemma that gives a closed-form expression for $\\pmb{J}^{\\dagger}$ :", + "bbox": [ + 498, + 739, + 890, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lemma 1. Let $\\pmb{d} \\in \\mathbb{R}^3$ and $J = \\frac{\\partial\\pi}{\\partial\\pmb{d}} \\in \\mathbb{R}^{2\\times 3}$ . Then, $J^{\\dagger}$ is given by", + "bbox": [ + 498, + 806, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} = \\frac {1}{\\boldsymbol {d} \\cdot \\left(\\boldsymbol {g} _ {x} \\times \\boldsymbol {g} _ {y}\\right)} \\left[ \\left(\\boldsymbol {g} _ {y} \\times \\boldsymbol {d}\\right), \\quad \\left(\\boldsymbol {d} \\times \\boldsymbol {g} _ {x}\\right) \\right], \\tag {26}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 844, + 890, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{g}_x$ , $\\pmb{g}_y \\in \\mathbb{R}^3$ are the two rows of the Jacobian $\\pmb{J}$ .", + "bbox": [ + 500, + 883, + 867, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3373", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. Simple calculations show that $JM = I_{2 \\times 2}$ . It remains to show that $MJ \\in \\mathbb{R}^{3 \\times 3}$ is symmetric. For ease of notation, introduce $\\alpha = d \\cdot (g_x \\times g_y)$ . Then we have", + "bbox": [ + 76, + 90, + 472, + 138 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\alpha \\boldsymbol {M} \\boldsymbol {J} = (\\boldsymbol {g} _ {y} \\times \\boldsymbol {d}) \\boldsymbol {g} _ {x} ^ {\\intercal} + (\\boldsymbol {d} \\times \\boldsymbol {g} _ {x}) \\boldsymbol {g} _ {y} ^ {\\intercal} = (27) \\\\ = [ \\boldsymbol {d} ] _ {\\times} \\left(- \\boldsymbol {g} _ {y} \\boldsymbol {g} _ {x} ^ {\\intercal} + \\boldsymbol {g} _ {x} \\boldsymbol {g} _ {y} ^ {\\intercal}\\right) = [ \\boldsymbol {d} ] _ {\\times} [ \\boldsymbol {g} _ {y} \\times \\boldsymbol {g} _ {x} ] _ {\\times}. (28) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 147, + 468, + 185 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the projection function $\\pi$ does not depend on the norm of $\\pmb{d}$ , i.e. $\\pi(\\pmb{d}) = \\pi(\\lambda \\pmb{d}) \\forall \\lambda > 0$ , we have that $\\pmb{J} \\pmb{d} = \\mathbf{0}$ . Since $\\pmb{d}$ is then orthogonal to both $\\pmb{g}_x$ and $\\pmb{g}_y$ we have $\\pmb{g}_y \\times \\pmb{g}_x = \\gamma \\pmb{d}$ for some $\\gamma \\in \\mathbb{R}$ . Inserting into (28) we get", + "bbox": [ + 76, + 194, + 468, + 257 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} \\boldsymbol {J} = \\frac {\\gamma}{\\alpha} [ \\boldsymbol {d} ] _ {\\times} ^ {2}, \\tag {29}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 263, + 468, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which is a symmetric matrix.", + "bbox": [ + 76, + 299, + 272, + 313 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9fd1beca797ef476155a5745b8b3b1554b9fe4f9bd2437aff48b003af4503df7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 314, + 468, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The expression (26) can be used to efficiently compute the Jacobian $J_{C}$ of the constraint function $C$ from (23), and finally to arrive at the Sampson approximation of the reprojection error in any differentiable camera model:", + "bbox": [ + 76, + 338, + 468, + 398 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {T S} ^ {2} = \\frac {C (\\boldsymbol {z}) ^ {2}}{\\| \\boldsymbol {J} _ {C} \\| ^ {2}} = \\frac {\\left(\\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1}\\right) ^ {2}}{\\| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {J} _ {1} ^ {\\dagger} \\| ^ {2} + \\| \\boldsymbol {d} _ {1} ^ {\\intercal} \\boldsymbol {E} ^ {\\intercal} \\boldsymbol {J} _ {2} ^ {\\dagger} \\| ^ {2}}, \\tag {30}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 406, + 468, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{d}_i = \\pi^{-1}(\\pmb{p}_i)$ and the two Jacobians are evaluated at $\\pmb{d}_1$ and $\\pmb{d}_2$ respectively. In the rest of the paper we refer to this as the Tangent Sampson error. The name reflects the fact that after we linearize the unprojection $\\pi^{-1}$ , we are essentially optimizing in the tangent plane to $S^2$ at the bearing vectors of the undistorted points. Note that the formula is analogous to the original Sampson error (18) (replacing $\\pmb{d}_1$ and $\\pmb{d}_2$ with the image points, and replacing $J_1^\\dagger$ and $J_2^\\dagger$ with $[10; 01; 00]$ ). Further, since $J_1$ and $J_2$ are evaluated at the original points of the correspondence, they are independent of $\\pmb{E}$ and can be precomputed. Thus, during RANSAC and refinement the computational cost stays approximately the same as for the classical Sampson error.", + "bbox": [ + 76, + 452, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Backward Camera Models", + "text_level": 1, + "bbox": [ + 76, + 657, + 318, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While most camera models used in practice are forward-models, i.e. mapping from the world to the image, there also exist backward-models which instead take the image points as input. The division model (8) is perhaps the most popular example here. The derivations above can easily be adapted to allow for backward-models. The pseudoinverses $J_{k}^{\\dagger}$ are simply replaced with the $3 \\times 2$ Jacobians from the undistortion map,", + "bbox": [ + 76, + 680, + 468, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {T S} ^ {2} = \\frac {\\left(\\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1}\\right) ^ {2}}{\\left\\| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {J} _ {1} \\right\\| ^ {2} + \\left\\| \\boldsymbol {d} _ {1} ^ {\\intercal} \\boldsymbol {E} ^ {\\intercal} \\boldsymbol {J} _ {2} \\right\\| ^ {2}}. \\tag {31}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 809, + 468, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In particular, for the division model (8) we would have", + "bbox": [ + 76, + 851, + 439, + 866 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {J} = \\frac {\\partial h}{\\partial \\boldsymbol {d}} \\frac {\\partial \\boldsymbol {d}}{\\partial \\boldsymbol {x}} = \\frac {\\partial h}{\\partial \\boldsymbol {d}} \\left[ I _ {2 \\times 2}; 2 \\lambda \\boldsymbol {x} ^ {\\intercal} \\right] \\in \\mathbb {R} ^ {3 \\times 2}, \\tag {32}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 875, + 468, + 906 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we need to explicitly include the Jacobian $\\frac{\\partial h}{\\partial d}$ of the normalization function $h$ , since the unprojected $\\pmb{d}$ from (8) is not a unit vector.", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Measurement Uncertainty", + "text_level": 1, + "bbox": [ + 500, + 143, + 736, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the approach from [3], we can incorporate measurement uncertainty in the keypoint positions into the Tangent Sampson error. We will assume that the keypoint positions are affected by independent Gaussian noise", + "bbox": [ + 498, + 167, + 890, + 228 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {1} \\sim \\mathcal {N} \\left(\\bar {\\boldsymbol {p}} _ {1}, \\boldsymbol {\\Sigma} _ {1}\\right), \\quad \\boldsymbol {p} _ {2} \\sim \\mathcal {N} \\left(\\bar {\\boldsymbol {p}} _ {2}, \\boldsymbol {\\Sigma} _ {2}\\right) \\tag {33}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 234, + 890, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with known covariances. For observations $\\pmb{p}_1$ and $\\pmb{p}_2$ , the maximum-likelihood error for an essential matrix $E$ is", + "bbox": [ + 498, + 258, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {E} _ {M L} ^ {2} (\\boldsymbol {z}, \\boldsymbol {E}) = \\min _ {\\hat {z}} \\quad \\frac {1}{2} (\\boldsymbol {z} - \\hat {\\boldsymbol {z}}) ^ {\\intercal} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {z} - \\hat {\\boldsymbol {z}}) \\tag {34} \\\\ \\begin{array}{c c} \\text {s . t .} & C (\\hat {\\boldsymbol {z}}) = 0, \\end{array} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 292, + 890, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C$ is the constraint function (23) and $\\pmb{\\Sigma}$ is the joint covariance of $\\pmb{z} = [\\pmb{p}_1; \\pmb{p}_2]$ , i.e. $\\pmb{\\Sigma} = \\text{diag}(\\pmb{\\Sigma}_1, \\pmb{\\Sigma}_2)$ .", + "bbox": [ + 498, + 348, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The procedure from Section 2.1 yields in this case", + "bbox": [ + 519, + 378, + 851, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {T S} ^ {2} = \\frac {C (\\boldsymbol {z}) ^ {2}}{\\boldsymbol {J} _ {C} \\boldsymbol {\\Sigma} \\boldsymbol {J} _ {C} ^ {\\intercal}} = \\frac {\\left(\\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1}\\right) ^ {2}}{\\| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {J} _ {1} ^ {\\dagger} \\| _ {\\Sigma_ {1}} ^ {2} + \\| \\boldsymbol {d} _ {1} ^ {\\intercal} \\boldsymbol {E} ^ {\\intercal} \\boldsymbol {J} _ {2} ^ {\\dagger} \\| _ {\\Sigma_ {2}} ^ {2}}, \\tag {35}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 398, + 890, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\| \\pmb{v} \\|_{\\Sigma}^{2} = \\pmb{v}^{T} \\pmb{\\Sigma} \\pmb{v}$ . See the supplementary material for details on the derivation. This is a generalization of the last expression in [3, Section 3.4] to the arbitrary camera model case. We evaluate how it compares to the original Tangent Sampson error (30) in a synthetic setting in Section 4.4.1.", + "bbox": [ + 498, + 439, + 890, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Evaluation", + "text_level": 1, + "bbox": [ + 500, + 527, + 733, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the rest of the paper we experimentally evaluate the Tangent Sampson error. We compare against a selection of popular error metrics, summarized in Table 1.", + "bbox": [ + 498, + 551, + 890, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 604, + 715, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implemented all error formulations in $\\mathrm{C + + }$ using Eigen [9] as a linear algebra library and Ceres [1] for optimization. Since evaluation is run on Intel hardware, we additionally speed up the errors that do not require iterative methods or explicit camera projection $(\\mathcal{E}_{TS},\\mathcal{E}_S,\\mathcal{E}_{SED},$ $\\mathcal{E}_{CS}$ ,and $\\mathcal{E}_{ALG})$ by vectorizing their implementations with AVX and FMA instruction sets. In Table 1 we also list the approximate computational cost of evaluating the residual. Tangent Sampson error is significantly faster compared to other metrics which also measure the error in the original image, and has comparable, but slightly higher, runtime to the purely geometric and pinhole-based metrics. For the computational cost we include parts which cannot be precomputed (i.e. parts that depend on the essential matrix). For a comparison with the non-optimized errors (using pure Eigen-implementations), we refer the reader to the supplementary material, where we list the runtime for all formulations without explicit vectorization and some setup details.", + "bbox": [ + 498, + 628, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3374", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/36d35b93157ff82a80a66a676bd31e0b84fd9d50d8cce3eabee2e4aba956e87a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NameRuntime
Rel.Abs.
\\( \\mathcal{E}_{ALG} \\)(7)Algebraic error↓1.01.3 ns
\\( \\mathcal{E}_{CS} \\)(6)Cosine error↓1.82.4 ns
\\( \\mathcal{E}_{ML} \\)(3)(Pinhole) Repoj. error\\( \\blacktriangleright \\)28.036.7 ns
\\( \\mathcal{E}_{SED} \\)(4)(Pinhole) Sym. Epi. Dist.\\( \\blacktriangleright \\)1.72.3 ns
\\( \\mathcal{E}_{S} \\)(5)(Pinhole) Sampson error\\( \\blacktriangleright \\)1.41.9 ns
\\( \\mathcal{E}_{\\pi ML} \\)(22)Reproj. error\\( \\square \\)8861160 ns
\\( \\mathcal{E}_{\\pi SED} \\)(10)Proj. Sym. Epi. Dist.\\( \\square \\)92121 ns
\\( \\mathcal{E}_{TS} \\)(30)Tangent Sampson\\( \\square \\)2.22.8 ns
", + "bbox": [ + 81, + 88, + 467, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Error metrics compared in the experimental evaluation. The errors are grouped by the domain where the error is computed: geometric errors $(\\mathcal{L})$ , the undistorted image $(\\mathfrak{X})$ and the original image $(\\square)$ . Computational cost is shown relative to the algebraic error which is the cheapest.", + "bbox": [ + 75, + 253, + 468, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Datasets", + "text_level": 1, + "bbox": [ + 76, + 330, + 179, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the evaluation we consider two datasets. First the meta-calibration dataset from BabelCalib [21]. The dataset is a collection of checkerboard images from a wide variety of cameras, including some with extremely wide field-of-view $(>270^{\\circ})$ . For the evaluation we split the dataset into three groups depending on their field-of-view; $< 90^{\\circ}$ , $90 - 150^{\\circ}$ , and $>150^{\\circ}$ . As the original dataset is biased towards higher field-of-view cameras, we also include some additional checkerboard images taken with a standard DSLR (circa $50 - 70^{\\circ}$ ) to even out the distribution. These checkerboard datasets allow us to evaluate on a large number of different camera and lens setups. In Section 4.3 we show that the Tangent Sampson error consistently approximates the true reprojection error, regardless of the camera used, and in Section 4.4 we show that when used for local refinement it yields the most accurate epipolar geometries.", + "bbox": [ + 75, + 354, + 468, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the performance of the error metrics in the context of RANSAC we consider image pairs extracted from Structure-from-Motion datasets, where the poses from the reconstruction are taken as a psuedo-ground truth. As the experiments in Section 4.3-4.4 show that most error metrics behave similarly for low to medium field-of-view cameras $(< 90^{\\circ})$ , we focus our evaluation on cameras with larger field-of-view. For this we use two datasets (Grossmunster and Kirchenge) from Larsson et al. [17], which contain 373 and 369 images taken with a fisheye-lens.", + "bbox": [ + 75, + 595, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with True Reprojection Error", + "text_level": 1, + "bbox": [ + 76, + 756, + 441, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate how different error metrics compare with the true maximum-likelihood estimate, $\\mathcal{E}_{ML}$ (3) for pinhole cameras and $\\mathcal{E}_{\\pi ML}$ (22) for non-pinhole cameras. To compute the ML-estimate we perform two-view triangulation followed by non-linear optimization of the 3D point, minimizing the squared reprojection error (using the known camera intrinsics). Note that this approach is orders of magnitude slower compared to the metrics we evaluate (see Ta", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ble 1) and cannot reasonably be used in practical scenarios.", + "bbox": [ + 498, + 90, + 890, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the experiment we consider the collection of checkerboard images from [21]. We use the calibration toolbox from BabelCalib [21] to estimate the extrinsic and intrinsic parameters for each image. For each pair we collect the 2D-2D correspondences from corresponding corners. As the corner detector [10] is extremely accurate on these images (some having around $0.2\\mathrm{px}$ error during calibration), we add Gaussian noise with $1\\mathrm{px}$ standard deviation, to better differentiate the methods (as all methods give zero error for exact data).", + "bbox": [ + 496, + 107, + 890, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since it is not meaningful to directly compare the residual values, we instead compare their ordering. For this we make use of Kendall's $\\tau$ rank correlation coefficient [15]. It is defined for two rankings $(p_1,\\dots ,p_n)$ and $(q_{1},\\ldots ,q_{n})$ as", + "bbox": [ + 498, + 258, + 890, + 320 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tau (p, q) = \\frac {\\left| \\{\\text {c o n c o r d a n t p a i r s} \\} \\right| - \\left| \\{\\text {d i s c o r d a n t p a i r s} \\} \\right|}{\\binom {n} {2}}. \\tag {36}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 340, + 890, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A pair of indices $(i,j)\\in \\{1,\\ldots ,n\\} ^2$ where $i > j$ is called concordant if and only if", + "bbox": [ + 498, + 386, + 890, + 417 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left(p _ {i} > p _ {j} \\wedge q _ {i} > q _ {j}\\right) \\vee \\left(p _ {i} < p _ {j} \\wedge q _ {i} < q _ {j}\\right). \\tag {37}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 431, + 890, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Otherwise, the pair is discordant. Note that due to the normalization we have $\\tau(p,q) \\in [-1,1]$ , where identical rankings yield $\\tau = 1$ , and reversed $\\tau = -1$ .", + "bbox": [ + 496, + 460, + 890, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results are shown in Fig. 3, split into three groups based on the camera field-of-view. For low to medium field-of-view cameras $(< 90^{\\circ})$ we can see that all metrics except the algebraic residual behave similar to the true reprojection error. As the field-of-view increases, there is a more clear difference between the methods, with the pinhole-based errors performing the worst. Note however, that for all cameras the proposed Tangent Sampson error offers the best approximation of the true reprojection error.", + "bbox": [ + 496, + 507, + 890, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to Kendall-Tau distance, we compare the errors qualitatively. For this, we show the values of an error for a fixed correspondence on a pair of images from a panoramic and fisheye camera in Fig. 4. There, we keep the point on the fisheye image fixed, and vary the point on the panoramic image to show different error values coded by different colors. One can see that $\\mathcal{E}_{TS}$ error models $\\mathcal{E}_{\\pi ML}$ closely in the region where the latter is well-defined. $\\mathcal{E}_{TS}$ is also one of the two errors that do not degenerate around the epipole, and, at the same time, it correctly models the spread of values of $\\mathcal{E}_{\\pi ML}$ around the epipolar curve. Note that the shape of true reprojection error varies significantly depending on how close the correspondence is to the epipole. In the supplementary material we show more qualitative results. In the supplementary material we also show a comparison with [16], which proposed an optimal method for computing the $\\pi ML$ error (22) for the division model (8).", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3375", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f8f9863b3dffe9841e2d7a3e51d10e5305c5ca8f233618b3decbd19021312a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 88, + 97, + 379, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43382c54d7df2953af5c5c9ea9a3c5e819ce256ec61fc26f7437a18c94e6eabf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 95, + 614, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3cc5e8188087f251cb695991e8e1d4294fe8cb4cde52620bb4b989d2dd3358c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 95, + 851, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4afa0f4d629a848b416988c4ff23f50895fcd5deb1029428b8d7f4760eef655c.jpg", + "image_caption": [ + "Figure 3. The figure shows the Kendall $\\tau$ coefficient comparing different error metrics for the checkerboard data. For all cameras the proposed Tangent Sampson error correlates well with the true reprojection error $\\pi ML$ ." + ], + "image_footnote": [], + "bbox": [ + 88, + 316, + 285, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cba1d2bf71e6381b62eab4f5cb6c7e32dfab92369c76378c13175ba5acd16d06.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of different error functions. The thresholds for colormaps were selected to be the error at the point marked by the blue plus. The shape of the true reprojection error varies heavily depending on how close the correspondence is to the epipoles. In the supplementary material we show more examples. We also show $\\mathcal{E}_{\\pi SED}$ and errors computed on the undistorted image in the supplementary. The geometric error is only shown in the region where the 3D point after midpoint triangulation does not correspond to negative depths of bearing vectors (areas where this does not hold for one or both depths are highlighted)." + ], + "image_footnote": [], + "bbox": [ + 285, + 316, + 483, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea6147ceb46990171650d7039770608f9b014e2d790c4f61e6ba52ca5842ae4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 316, + 681, + 395 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06a0fdee873acd6009111bca766195b0a5328473d4045423eb20eb98ead0b4d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 316, + 880, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Accuracy in Pose Refinement", + "text_level": 1, + "bbox": [ + 76, + 483, + 338, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next we evaluate the error metrics on how they perform when used for pose-refinement. For the experiment we again consider the image pairs extracted from the checkerboard images. For each image pair we perform local optimization minimizing the proposed error metrics over all correspondences. To initialize the optimization we take the ground truth relative pose and randomly perturb both the rotation and translation by 1 degree. Additionally, point positions were perturbed with Gaussian noise with unit standard deviation along each dimension. For the two ML metrics we triangulate 3D points which are optimized jointly with the camera pose. We experience that this sometimes has problems with worse convergence. The results are shown in Table 2. Again we see that for low to medium field-of-view cameras, the difference between the error metrics is small. The Tangent Sampson error consistently provides the most accurate camera poses after refinement.", + "bbox": [ + 75, + 507, + 468, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Convergence Properties. The previous experiment evaluated the accuracy of the camera poses obtained when initialized close to the correct pose. We also experimented with varying the initial pose and measuring how often the optimization converged to the correct pose. Figure 5 shows the percentage of poses (across all pairs) that are within 5 degrees average rotation/translation error for varying errors in the initial pose. For large initial errors $(>25^{\\circ})$ , the proposed Tangent-Sampson error performs slightly better com", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/497cdf7c45381b2230692d8d70a7f185c1b300470d2dd2823f7a9d2d74ec0736.jpg", + "image_caption": [ + "Figure 5. Convergence properties. Successful trials (average rot. and trans. error $< 5^{\\circ}$ ) for varying initial pose errors (in degrees)." + ], + "image_footnote": [], + "bbox": [ + 508, + 484, + 890, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pared to the other metrics.", + "bbox": [ + 500, + 684, + 676, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1 Effect of Measurement Uncertainty", + "text_level": 1, + "bbox": [ + 500, + 723, + 802, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiments in [3] already show that taking point covariances into account in the cost function provides benefits for fundamental matrix estimation. Here, we conduct another synthetic test to show that this effect persists when estimating the essential matrix with highly nonlinear cameras. We generate 2000 instances with pairs of spherical cameras $(1000 \\times 2000\\mathrm{px})$ with a random relative pose. For each pair we project 100 random 3D-points uniformly drawn from the cube $[-10, 10]^3$ , and add Gaussian noise with covariance $\\Lambda$ .", + "bbox": [ + 496, + 748, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Covariances are generated as $\\Lambda = \\alpha Q\\mathrm{diag}(\\beta ,1 - \\beta)Q^{\\intercal}$", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3376", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8db0af217d6f9825c8c06953ded14f13b545eac87d1b865543b79b85d8481ed0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
< 90° (5332 pairs)90 - 150° (21927 pairs)> 150° (11860 pairs)
RotationTranslationRotationTranslationRotationTranslation
\\( \\mathcal{E}_{ALG} \\)0.379 / 0.3141.137 / 0.3281.528 / 0.3783.938 / 0.5881.582 / 0.5063.489 / 0.655
\\( \\mathcal{E}_S \\)0.376 / 0.3121.056 / 0.3271.464 / 0.3473.520 / 0.5751.634 / 0.5673.812 / 0.778
\\( \\mathcal{E}_{SED} \\)0.376 / 0.3121.065 / 0.3271.475 / 0.3523.554 / 0.5782.596 / 0.9865.667 / 1.342
\\( \\mathcal{E}_{COS} \\)0.376 / 0.3131.066 / 0.3261.454 / 0.3413.459 / 0.5581.447 / 0.4543.235 / 0.612
\\( \\mathcal{E}_{ML} \\)0.376 / 0.3121.103 / 0.3271.472 / 0.3583.733 / 0.5751.630 / 0.5633.771 / 0.766
\\( \\mathcal{E}_{\\pi ML} \\)0.375 / 0.3121.075 / 0.3271.451 / 0.3503.778 / 0.5571.418 / 0.4463.143 / 0.597
\\( \\mathcal{E}_{\\pi SED} \\)0.376 / 0.3131.074 / 0.3271.458 / 0.3433.490 / 0.5591.473 / 0.4503.249 / 0.607
\\( \\mathcal{E}_{TS} \\)0.376 / 0.3121.057 / 0.3271.443 / 0.3403.432 / 0.5571.465 / 0.4453.226 / 0.598
", + "bbox": [ + 102, + 88, + 867, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2a99a1fc063882dcc1eeac8ac24b68cbb4d615cb714de51983e69e9f10c5048.jpg", + "image_caption": [ + "Figure 6. Measurement Uncertainty. The distribution of pose errors (maximum of rotation and translation error) when using covariance-aware Tangent Sampson error for pose refinement versus non-covariance-aware in synthetic experiments. Boxes show quartiles of the respective distribution of errors and whiskers extend to the minimal and maximal value." + ], + "image_footnote": [], + "bbox": [ + 94, + 320, + 455, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $Q \\sim U(\\mathrm{SO}(2))$ is a random rotation on the plane, $\\beta \\sim U([0, 0.5])$ is a random eccentricity parameter, and $\\alpha$ is the overall scale of the noise. To emulate keypoint detection on a multi-scale image pyramid, all generated points were split into 4 equal sets, with $\\alpha$ within each set assigned to 1, 2, 4, and 8, respectively. Points and covariances are then passed to a Levenberg-Marquardt optimization of the relative pose using (35) as the residual function, with true pose serving as the initial estimate. Figure 6 demonstrates how the covariance-aware formulation (35) compares to the baseline (30). Covariance information is useful in the case when points are affected by noise of unequal scales.", + "bbox": [ + 75, + 529, + 468, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Evaluation in Robust Estimators", + "text_level": 1, + "bbox": [ + 76, + 724, + 364, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally we evaluate the methods in the context of RANSAC [7]. We implement a standard LO-RANSAC [18] estimator which uses the various error metrics both for model scoring (using MSAC scoring [24]) and refinement (using non-linear optimization on the inlier-set). For the experiment we consider image pairs from two Structure-from-Motion reconstructions (see Section 4.2). From each dataset we extract all image pairs which contain between 50 and 200 correct matches (according to the Structure-from-Motion ground-truth). For each pair we estimate the rela", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5072a274e019e3faa37997e5583984eff60168290dd6bd162990bd000d6f1b4f.jpg", + "table_caption": [ + "Table 2. Evaluation of camera pose refinement. Table shows the mean and median errors (in degrees) of the rotation and translation after refinement. The data is split into three groups based on the camera field-of-view. Best error is highlighted in bold. Second-best is underlined. For low-distortion cameras the choice of error makes almost no difference on the quality of pose." + ], + "table_footnote": [], + "table_body": "
GrossmunsterKirchenge
εR,t< 5°RTεR,t< 5°RT
ΕALG0.76175.039.60.31884.821.6
ΕCOS0.56181.550.40.25491.625.5
ΕSED1.16168.982.00.35586.444.0
ΕS0.92474.551.00.30089.225.3
ΕπSED0.56181.3211.20.25591.687.4
ΕTS0.55082.246.40.24891.924.2
", + "bbox": [ + 500, + 319, + 903, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Evaluation in robust estimators. Table shows the median pose error (maximum of rotation and translation error), percentage of pairs for which the estimation was successful (pose error less than 5 degrees), and runtime (in ms). Best value in the column is highlighted in bold. Second-best is underlined.", + "bbox": [ + 496, + 473, + 892, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tive pose using RANSAC and compute the error in the pose, as well as percentage of successful runs. For each method we tuned the inlier threshold on a separately extracted validation set. For more details on the experiment setup see the supplementary material. The results are presented in Table 3. Using Tangent Sampson error consistently gives slightly better performance compared to other error metrics.", + "bbox": [ + 496, + 561, + 892, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 696, + 627, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper we have presented a simple generalization to the classical Sampson error which allows for using arbitrary camera models. In experiments on a wide variety of camera and lens configurations, we have shown that the proposed error metric outperforms the alternative metrics that are used in practice. While the improvements are sometimes minor (e.g. for low-distortion pinhole cameras), they are consistent and come at essentially no additional cost.", + "bbox": [ + 496, + 729, + 893, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This project received funding from the strategic research program ELLIIT. MT was supported by the ESOP scholarship from the ETH Foundation.", + "bbox": [ + 496, + 856, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3377", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sameer Agarwal, Keir Mierle, and Others. Ceres solver. http://ceres-solver.org.5", + "[2] João Pedro Barreto and Kostas Daniilidis. Fundamental matrix for cameras with radial distortion. In International Conference on Computer Vision (ICCV), 2005. 3", + "[3] Wojciech Chojnacki, Michael J. Brooks, Anton Van Den Hengel, and Darren Gawley. On the fitting of surfaces to data with covariances. IEEE Trans. Pattern Analysis and Machine Intelligence (PAMI), 2000. 3, 5, 7", + "[4] Ondřej Chum, Tomáš Pajdla, and Peter Sturm. The geometric error for homographies. Computer Vision and Image Understanding (CVIU), 2005. 3", + "[5] David Claus and Andrew W Fitzgibbon. A rational function lens distortion model for general cameras. In Computer Vision and Pattern Recognition (CVPR), 2005. 3", + "[6] Asen L Dontchev and R Tyrrell Rockafellar. Implicit functions and solution mappings, volume 543. Springer, 2009. 4", + "[7] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 8", + "[8] Andrew W Fitzgibbon. Simultaneous linear estimation of multiple view geometry and lens distortion. In Computer Vision and Pattern Recognition (CVPR), 2001. 3", + "[9] Gaël Guennebaud, Benoit Jacob, et al. Eigen v3. http://eigen.tuxfamily.org, 2010. 5", + "[10] Hyowon Ha, Michal Perdoch, Hatem Alismail, In So Kweon, and Yaser Sheikh. Deltille grids for geometric camera calibration. In International Conference on Computer Vision (ICCV), 2017. 6", + "[11] Matthew Harker and Paul O'Leary. First order geometric distance (the myth of sampsonus). In *British Machine Vision Conference (BMVC)*, 2006. 3", + "[12] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1", + "[13] Richard I Hartley and Peter Sturm. Triangulation. Computer vision and image understanding, 68(2):146-157, 1997. 2, 3", + "[14] Juho Kannala and Sami S Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. 2006. 4", + "[15] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1/2):81-93, 1938. 6", + "[16] Zuzana Kukelova and Viktor Larsson. Radial distortion triangulation. In Computer Vision and Pattern Recognition (CVPR), 2019. 3, 6", + "[17] Viktor Larsson, Nicolas Zobernig, Kasim Taskin, and Marc Pollefeys. Calibration-free structure-from-motion with calibrated radial trifocal tensors. In European Conference on Computer Vision (ECCV), 2020. 6", + "[18] Karel Lebeda, Jiri Matas, and Ondrej Chum. Fixing the locally optimized ransac-full experimental evaluation. In British Machine Vision Conference (BMVC), 2012. 8" + ], + "bbox": [ + 78, + 116, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Seong Hun Lee and Javier Civera. Geometric interpretations of the normalized epipolar error. arXiv preprint arXiv:2008.01254, 2020. 2", + "[20] Peter Lindstrom. Triangulation made easy. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 1554-1561. IEEE, 2010. 2, 3", + "[21] Yaroslava Lochman, Kostiantyn Liepieshov, Jianhui Chen, Michal Perdoch, Christopher Zach, and James Pritts. Babel-calib: A universal approach to calibrating central cameras. In International Conference on Computer Vision (ICCV), 2021. 6", + "[22] Quan-Tuan Luong and Olivier D Faugeras. The fundamental matrix: Theory, algorithms, and stability analysis. International Journal of Computer Vision (IJCV), 1996. 2, 3", + "[23] Paul D Sampson. Fitting conic sections to \"very scattered\" data: An iterative refinement of the bookstein algorithm. Computer graphics and image processing, 1982. 2, 3", + "[24] Philip HS Torr and Andrew Zisserman. Mlesac: A new robust estimator with application to estimating image geometry. Computer Vision and Image Understanding (CVIU), 2000, 8", + "[25] Zhengyou Zhang. On the epipolar geometry between two images with lens distortion. In International Conference on Pattern Recognition (ICPR), 1996. 3" + ], + "bbox": [ + 501, + 92, + 890, + 430 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3378", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/636b4c99-6986-4080-bf56-9a578e79f107_model.json b/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/636b4c99-6986-4080-bf56-9a578e79f107_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cf926f31e40bbfe2b99f8b655d502037777dda7f --- /dev/null +++ b/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/636b4c99-6986-4080-bf56-9a578e79f107_model.json @@ -0,0 +1,2352 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.132, + 0.872, + 0.177 + ], + "angle": 0, + "content": "Tangent Sampson Error: Fast Approximate Two-view Reprojection Error for Central Camera Models" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.204, + 0.408, + 0.238 + ], + "angle": 0, + "content": "Mikhail Terekhov ETH Zurich" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.243, + 0.411, + 0.255 + ], + "angle": 0, + "content": "mterekhov@ethz.ch" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.204, + 0.656, + 0.22 + ], + "angle": 0, + "content": "Viktor Larsson" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.223, + 0.66, + 0.24 + ], + "angle": 0, + "content": "Lund University" + }, + { + "type": "text", + "bbox": [ + 0.475, + 0.243, + 0.708, + 0.256 + ], + "angle": 0, + "content": "viktor.larsson@math.lth.se" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.296, + 0.599, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.308, + 0.839, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.492, + 0.895, + 0.549 + ], + "angle": 0, + "content": "Figure 1. The figure shows the level sets of the proposed Tangent Sampson error. In contrast to classical Sampson error, the proposed error function applies to arbitrary central camera models; in this case a 360 panoramic image and a fisheye camera. Given a relative pose and a pair of points (highlighted in red), we vary a point on one image while keeping the other fixed, and thus get the values on each image. A given color corresponds to a constant value of reprojection error, with purple indicating near-zero errors along the epipolar curve." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.561, + 0.314, + 0.577 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.593, + 0.474, + 0.806 + ], + "angle": 0, + "content": "In this paper we introduce the Tangent Sampson error, which is a generalization of the classical Sampson error in two-view geometry that allows for arbitrary central camera models. It only requires local gradients of the distortion map at the original correspondences (allowing for pre-computation) resulting in a negligible increase in computational cost when used in RANSAC or local refinement. The error effectively approximates the true-reprojection error for a large variety of cameras, including extremely wide field-of-view lenses that cannot be undistorted to a single pinhole image. We show experimentally that the new error outperforms competing approaches both when used for model scoring in RANSAC and for non-linear refinement of the relative camera pose." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.83, + 0.21, + 0.846 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Estimating two-view geometry from a set of sparse 2D-2D correspondences is a common sub-problem in many vision applications. Each correspondence \\((\\pmb{p}_1, \\pmb{p}_2) \\in \\mathbb{R}^2 \\times \\mathbb{R}^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.563, + 0.843, + 0.578 + ], + "angle": 0, + "content": "constrains the geometry via the epipolar constraint," + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.585, + 0.892, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\left(\\boldsymbol {p} _ {2}; 1\\right) ^ {\\top} \\boldsymbol {E} \\left(\\boldsymbol {p} _ {1}; 1\\right) = 0, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.893, + 0.758 + ], + "angle": 0, + "content": "where \\(\\pmb{E} = [t]_{\\times}\\pmb{R} \\in \\mathbb{R}^{3\\times 3}\\) is the essential matrix and the points \\(p_1, p_2\\) are on the normalized image plane. This constraint ensures that the 2D-2D correspondence can be exactly triangulated (i.e. with zero reprojection error) using the camera pair extracted from \\(\\pmb{E}\\). However when the measurement \\((p_1, p_2)\\) is inexact, it is natural to measure how close this correspondence is to being consistent with a given epipolar geometry. This is for example used for determining inlier matches during robust estimation, or as a loss to be minimized when refining an initial estimate of \\(\\pmb{E}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.759, + 0.894, + 0.82 + ], + "angle": 0, + "content": "Under the assumption of i.i.d. Gaussian noise, the ML-estimate of the closest consistent correspondence, sometimes called the gold-standard estimate in the literature [12], is found by minimizing the squared reprojection error," + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.84, + 0.892, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {M L} ^ {2} = \\min _ {\\boldsymbol {X} \\in \\mathbb {R} ^ {3}} \\| \\boldsymbol {p} _ {1} - \\pi (\\boldsymbol {X}) \\| ^ {2} + \\| \\boldsymbol {p} _ {2} - \\pi (\\boldsymbol {R} \\boldsymbol {X} + t) \\| ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.87, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\pi : \\mathbb{R}^3 \\to \\mathbb{R}^2\\) is the camera projection function. For identical pinhole cameras with unit aspect ratio and no" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3370" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.138 + ], + "angle": 0, + "content": "skew, this can be reformulated as finding the corrected normalized image points, i.e. the closest pair of 2D points that satisfy the epipolar constraint exactly," + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.148, + 0.469, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {E} _ {M L} ^ {2} = \\min _ {\\hat {\\boldsymbol {p}} _ {1}, \\hat {\\boldsymbol {p}} _ {2} \\in \\mathbb {R} ^ {2}} \\| \\boldsymbol {p} _ {1} - \\hat {\\boldsymbol {p}} _ {1} \\| ^ {2} + \\| \\boldsymbol {p} _ {2} - \\hat {\\boldsymbol {p}} _ {2} \\| ^ {2} \\tag {3} \\\\ \\begin{array}{l} \\text {s . t .} \\quad (\\hat {\\boldsymbol {p}} _ {2}; 1) ^ {\\intercal} \\boldsymbol {E} (\\hat {\\boldsymbol {p}} _ {1}; 1) = 0. \\end{array} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.207, + 0.469, + 0.297 + ], + "angle": 0, + "content": "The closed-form solution to (3) can be found as a root to a degree-six polynomial (see [13]), or via iterative methods [20]. In robust estimators, e.g. RANSAC, it is often necessary to evaluate these residuals thousands (or potentially millions) of times for each image pair, which in practice means that cheaper approximate error metrics are favored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.469, + 0.435 + ], + "angle": 0, + "content": "These approximate or surrogate error metrics can be divided into two categories: image-based (approximating the true reprojection error) or geometry-based (measuring some geometric quantity, e.g. angles, instead). In the first category we have the epipolar distance, i.e. distance from the point to the corresponding epipolar line. This can be interpreted as assuming zero noise on one of the keypoints. The error is often computed symmetrically and thus called the symmetric epipolar distance, having the form" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.445, + 0.469, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {S E D} ^ {2} = \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| \\boldsymbol {E} _ {1 2} \\boldsymbol {d} _ {1} \\right\\| ^ {2}} + \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| (\\boldsymbol {E} ^ {\\intercal}) _ {1 2} \\boldsymbol {d} _ {2} \\right\\| ^ {2}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.491, + 0.469, + 0.552 + ], + "angle": 0, + "content": "where \\(E_{12} \\in \\mathbb{R}^{2 \\times 3}\\) refers to the first two rows of \\(\\pmb{E}\\), and \\(\\pmb{d}_k = (\\pmb{p}_k; 1) \\in \\mathbb{R}^3\\). Another image-based error is the Sampson error, which is a first-order approximation of \\(\\mathcal{E}_{ML}\\) (see Sec. 2.1 for details):" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.563, + 0.469, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {S} ^ {2} = \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| \\boldsymbol {E} _ {1 2} \\boldsymbol {d} _ {1} \\right\\| ^ {2} + \\left\\| (\\boldsymbol {E} ^ {\\intercal}) _ {1 2} \\boldsymbol {d} _ {2} \\right\\| ^ {2}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.607, + 0.469, + 0.653 + ], + "angle": 0, + "content": "Both of these errors can after appropriate rescaling (using the focal lengths) be expressed in pixels and thus are natural to interpret in terms of the keypoint detection noise levels." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.654, + 0.457, + 0.668 + ], + "angle": 0, + "content": "In the second category, an example is the cosine error," + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.68, + 0.469, + 0.714 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {C S} ^ {2} = \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| \\boldsymbol {d} _ {2} \\right\\| ^ {2} \\left\\| \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right\\| ^ {2}} + \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| \\boldsymbol {E} ^ {\\intercal} \\boldsymbol {d} _ {2} \\right\\| ^ {2} \\left\\| \\boldsymbol {d} _ {1} \\right\\| ^ {2}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.725, + 0.469, + 0.783 + ], + "angle": 0, + "content": "which (symmetrically) measures the cosine between the normal of the epipolar plane \\( \\pmb{E}\\pmb{d}_1 \\in \\mathbb{R}^3 \\) and corresponding point \\( d_2 \\in \\mathbb{R}^3 \\). Another geometric error is the algebraic residual," + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.786, + 0.469, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {A L G} = \\left| \\boldsymbol {d} _ {2} ^ {\\mathrm {T}} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right|, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.415, + 0.825 + ], + "angle": 0, + "content": "which was given a geometric interpretation in [19]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Most of the literature has focused on the pinhole camera case, and the image-based errors above \\((\\mathcal{E}_{SED}\\) and \\(\\mathcal{E}_S)\\) measure the error in the image plane assuming this model. For non-pinhole cameras, we either have to undistort the image such that \\(\\mathcal{E}_{SED}\\) and \\(\\mathcal{E}_S\\) are applicable, or use the" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.089, + 0.696, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.089, + 0.887, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.209, + 0.89, + 0.238 + ], + "angle": 0, + "content": "Figure 2. Undistorting the image deforms the level-sets of the reprojection error and introduces a radial bias in the optimization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.245, + 0.892, + 0.305 + ], + "angle": 0, + "content": "geometry-based error metrics (e.g. \\(\\mathcal{E}_{CS}\\) or \\(\\mathcal{E}_{ALG}\\)). However, undistorting to a pinhole image (which is not always possible for large field-of-view cameras) will introduce a radial-bias into the error (see Figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.306, + 0.892, + 0.366 + ], + "angle": 0, + "content": "In this paper we instead propose the Tangent Sampson error \\(\\mathcal{E}_{TS}\\), a version of the classical Sampson error that has been generalized to handle arbitrary camera models. Compared to current alternatives it has the following benefits;" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.376, + 0.892, + 0.436 + ], + "angle": 0, + "content": "- It is cheap to evaluate (similar in cost to the original Sampson error) since it only requires computing the camera model forward-projection and Jacobian once (which can be pre-computed for each point)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.441, + 0.892, + 0.516 + ], + "angle": 0, + "content": "- Since it measures errors in the original image, it is easier to set meaningful thresholds in pixels. This is in contrast to geometric error functions, where it is not always clear how to relate the thresholds (e.g. in angles) to the keypoint detection noise in pixels." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.521, + 0.892, + 0.611 + ], + "angle": 0, + "content": "- It works for arbitrary central camera models, only requiring that we can compute a Jacobian of the camera projection function at each correspondence. This also allows the formulation to work for cameras modeled via look-up tables, where the Jacobian can be estimated via interpolation of finite differences." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.376, + 0.892, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.621, + 0.892, + 0.711 + ], + "angle": 0, + "content": "In Section 2.1 we first recall the derivations for the original Sampson error. Next, in Section 3 we propose our generalization to arbitrary central camera models, and in Section 4 we experimentally show that the proposed error outperforms the competing methods on a variety of camera and lens configurations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.64, + 0.74 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.75, + 0.892, + 0.87 + ], + "angle": 0, + "content": "The Sampson error was introduced in [23] to approximate the geometric distance from a point to a conic. The proposed approach was to linearize the equation of the conic around the given point, rendering a closed-form solution for the new distance, now from the point to a line. This idea proved useful for estimating distances to other implicit surfaces, notably those set by constraints of a fundamental matrix [22]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The fundamental matrix is only applicable to perspective cameras, and a lot of subsequent work in computer vision" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3371" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "focused on generalizing it to non-linear camera models, together with some notion of a reprojection error. A well-known example is the division model from Fitzgibbon [8]. It maps a point \\( \\pmb{p} \\) on the normalized image plane to the corresponding bearing vector \\( \\pmb{d} \\) by" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.176, + 0.469, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {d} = (\\boldsymbol {p}; 1 + \\lambda \\| \\boldsymbol {p} \\| ^ {2}). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.203, + 0.469, + 0.337 + ], + "angle": 0, + "content": "The original paper [8] already discusses a way of approximating the gold-standard error for the new model: first, undistort the original points of the correspondence; next, apply the closed-form triangulation from [13], and distort the projections back onto the original image. The resulting points are then approximate minimizers of (2). A faster iterative scheme for the optimal error was later proposed in [16], which adapts the idea of [20], originally developed for pinhole cameras, to the division model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.469, + 0.504 + ], + "angle": 0, + "content": "Several camera models with closed-form expressions for epipolar curves were proposed. In [25], epipolar curves are represented by cubics, in the rational model [5] — by conics, and in [2] — by circles. With such models, we can use the Sampson approximation to estimate the distance from one point of the correspondence to the epipolar curve of the other, and vice versa. Note that this gives us an approximation of the symmetric epipolar distance, not the Sampson error, because for the latter we would need to optimize both point positions simultaneously, and this would move the epipolar curves as well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.469, + 0.64 + ], + "angle": 0, + "content": "All these models were proposed with the goal of searching for the intrinsic parameters together with the epipolar geometry. Indeed, [2] and [5] even provide their own versions of a fundamental matrix on a lifted image space. These matrices can be estimated directly from point correspondences, and then decomposed to provide the distortion parameters and relative pose up to an irreducible ambiguity. We, on the other hand, constrain ourselves to the calibrated case, and thus can afford much greater generality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.469, + 0.761 + ], + "angle": 0, + "content": "There are also various heuristic generalizations of the symmetric epipolar distance to general non-linear camera models that have been used in practice. One approach is to project the bearing vector of one point of the correspondence onto the epipolar plane given by the other. The projected vector is then mapped back onto the image to provide an approximate corrected point, for which the error can be computed. Mathematically, this is expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.773, + 0.468, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {n} _ {1} = \\boldsymbol {E} \\boldsymbol {d} _ {1} / \\| \\boldsymbol {E} \\boldsymbol {d} _ {1} \\|, \\quad \\boldsymbol {n} _ {2} = \\boldsymbol {E} ^ {\\top} \\boldsymbol {d} _ {2} / \\| \\boldsymbol {E} ^ {\\top} \\boldsymbol {d} _ {2} \\|, \\tag {9} \\\\ \\mathcal {E} _ {\\pi S E D} ^ {2} = \\left\\| \\boldsymbol {p} _ {1} - \\pi \\left(\\boldsymbol {d} _ {1} - \\boldsymbol {n} _ {2} \\boldsymbol {n} _ {2} ^ {\\intercal} \\boldsymbol {d} _ {1}\\right) \\right\\| ^ {2} \\\\ + \\left\\| \\boldsymbol {p} _ {2} - \\pi \\left(\\boldsymbol {d} _ {2} - \\boldsymbol {n} _ {1} \\boldsymbol {n} _ {1} ^ {\\intercal} \\boldsymbol {d} _ {2}\\right) \\right\\| ^ {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.9 + ], + "angle": 0, + "content": "where \\( p_1 \\) and \\( p_2 \\) are the points on the images, \\( d_1 \\) and \\( d_2 \\) are the corresponding bearing vectors, \\( n_1 \\) and \\( n_2 \\) are the normals of the epipolar planes. We have encountered this formulation in private codebases, but we are not aware of its" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.181 + ], + "angle": 0, + "content": "origin. We call it the projective symmetric epipolar error. While this does not give us exact distances to the epipolar curves even in the pinhole case, it often provides a reasonable approximation. The main problem with this approach is that it requires calculating \\(\\pi(p)\\) every time, making it rather computationally heavy for some camera models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.347 + ], + "angle": 0, + "content": "From the probabilistic perspective, the geometric error is only justified as long as the noise in the correspondences stays standard Gaussian and i.i.d. A natural generalization is to allow for arbitrary covariances in point positions. For example, when the keypoints are searched on a multi-resolution image pyramid, the standard deviation of the position can be taken as the downscaling factor of the level on which the point was found. For conics and fundamental matrices, covariance-aware error formulations were derived in [3]. We will show how this approach is generalized to our proposed error formulation in Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.499 + ], + "angle": 0, + "content": "In [11] the authors propose another first-order approximation of the reprojection error. The method builds on the optimal triangulation method from [13] and differentates the sixth-order polynomial that has the optimal geometric error as its first root. The first-order approximation is then given by one step of the Newton-Raphson iteration on this polynomial starting from zero. In the paper, this error is only derived for conics. While it would also work for fundamental matrices, this does not generalize to the nonlinear camera model setting we are interested in." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Finally, we would like to note that the Sampson error can also be derived for homographies [4]. Our method is applicable there as well, but with larger computational cost, which somewhat defeats its original purpose." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.569, + 0.762, + 0.585 + ], + "angle": 0, + "content": "2.1. The Sampson Approximation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.592, + 0.891, + 0.667 + ], + "angle": 0, + "content": "Traditional Sampson error [22, 23] approximates the geometric error, i.e. the distance to the closest pair of points which satisfy the epipolar constraint. Given a correspondence \\( \\mathbf{z} = (\\pmb{p}_1, \\pmb{p}_2) \\in \\mathbb{R}^2 \\times \\mathbb{R}^2 \\) and essential matrix \\( \\pmb{E} \\in \\mathbb{R}^{3 \\times 3} \\), this can be formalized as" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.677, + 0.89, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {E} _ {M L} ^ {2} (\\boldsymbol {z}, \\boldsymbol {E}) = \\min _ {\\hat {\\boldsymbol {z}}} \\| \\hat {\\boldsymbol {z}} - \\boldsymbol {z} \\| ^ {2} \\tag {11} \\\\ \\begin{array}{c c} \\text {s . t .} & C (\\hat {z}) = 0, \\end{array} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.89, + 0.761 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{z}} = (\\hat{\\pmb{p}}_1,\\hat{\\pmb{p}}_2)\\in \\mathbb{R}^2\\times \\mathbb{R}^2\\) is the corrected correspondence and the epipolar constraint is" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.772, + 0.89, + 0.789 + ], + "angle": 0, + "content": "\\[\nC (\\hat {\\boldsymbol {z}}) = \\left(\\hat {\\boldsymbol {p}} _ {2}; 1\\right) ^ {\\intercal} \\boldsymbol {E} \\left(\\hat {\\boldsymbol {p}} _ {1}; 1\\right) = 0. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.8, + 0.89, + 0.844 + ], + "angle": 0, + "content": "Since (11) does not admit a simple closed form solution, the idea in [23] is to linearize the epipolar constraint, \\(C(\\hat{z}) = 0\\) at the original correspondence \\(\\mathcal{Z}\\), i.e." + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.854, + 0.89, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {S} ^ {2} (\\boldsymbol {z}, \\boldsymbol {E}) = \\min _ {\\hat {\\boldsymbol {z}}} \\| \\hat {\\boldsymbol {z}} - \\boldsymbol {z} \\| ^ {2} \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.882, + 0.843, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\text {s . t .} \\quad C (\\boldsymbol {z}) + \\boldsymbol {J} _ {C} (\\hat {\\boldsymbol {z}} - \\boldsymbol {z}) = 0,\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3372" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "where \\( J_{C} = \\partial C(z) / \\partial z \\) is the Jacobian of the constraint, evaluated at \\( z \\). Introducing a Lagrangian for (13)," + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.133, + 0.47, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\hat {z}, \\lambda) = \\| \\hat {z} - z \\| ^ {2} + \\lambda (C (z) + J _ {C} (\\hat {z} - z)), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.164, + 0.313, + 0.178 + ], + "angle": 0, + "content": "we get the first-order constraints as" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.19, + 0.469, + 0.219 + ], + "angle": 0, + "content": "\\[\n\\hat {z} = z - \\frac {\\lambda}{2} \\boldsymbol {J} _ {C} ^ {\\intercal}, \\quad C (z) + \\boldsymbol {J} _ {C} (\\hat {z} - z) = 0. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.23, + 0.4, + 0.245 + ], + "angle": 0, + "content": "Inserting the first equation into the second yields" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.256, + 0.469, + 0.289 + ], + "angle": 0, + "content": "\\[\nC (\\boldsymbol {z}) - \\frac {\\lambda}{2} \\| \\boldsymbol {J} _ {C} \\| ^ {2} = 0 \\Rightarrow \\hat {\\boldsymbol {z}} = \\boldsymbol {z} - \\frac {C (\\boldsymbol {z})}{\\| \\boldsymbol {J} _ {C} \\| ^ {2}} \\boldsymbol {J} _ {C} ^ {\\intercal}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.3, + 0.331, + 0.315 + ], + "angle": 0, + "content": "Thus the minimum in (13) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.326, + 0.469, + 0.362 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {S} ^ {2} = \\left\\| \\frac {C (\\boldsymbol {z})}{\\| \\boldsymbol {J} _ {C} \\| ^ {2}} \\boldsymbol {J} _ {C} ^ {\\intercal} \\right\\| ^ {2} = \\frac {C (\\boldsymbol {z}) ^ {2}}{\\| \\boldsymbol {J} _ {C} \\| ^ {2}}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.373, + 0.47, + 0.403 + ], + "angle": 0, + "content": "Replacing \\( C(z) \\) with the epipolar constraint, we arrive at the classical formula for the Sampson error:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.415, + 0.469, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {S} ^ {2} = \\frac {\\left(\\left(\\boldsymbol {p} _ {2} ; 1\\right) ^ {\\intercal} \\boldsymbol {E} \\left(\\boldsymbol {p} _ {1} ; 1\\right)\\right) ^ {2}}{\\left\\| \\boldsymbol {E} _ {1 2} \\left(\\boldsymbol {p} _ {1} ; 1\\right) \\right\\| ^ {2} + \\left\\| \\left(\\boldsymbol {E} ^ {\\intercal}\\right) _ {1 2} \\left(\\boldsymbol {p} _ {2} ; 1\\right) \\right\\| ^ {2}}. \\tag {18}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.459, + 0.376, + 0.477 + ], + "angle": 0, + "content": "3. Generalizing the Sampson Error" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.486, + 0.47, + 0.576 + ], + "angle": 0, + "content": "The Sampson error derived in the previous section assumes a pinhole projection model. However, many cameras, in particular those with wide field-of-view, require a more complex model. One popular choice is the Kannala-Brandt [14] model in which a 3D point \\( \\pmb{d} = (X,Y,Z)^{\\mathrm{T}} \\) in the camera coordinate system is projected as" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.587, + 0.469, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\pi (\\boldsymbol {d}) = \\frac {\\theta_ {d}}{\\sqrt {X ^ {2} + Y ^ {2}}} \\binom {f _ {x} X} {f _ {y} Y} + \\binom {c _ {x}} {c _ {y}} \\tag {19}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.622, + 0.469, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\theta_ {d} = \\theta \\left(1 + k _ {1} \\theta^ {2} + k _ {2} \\theta^ {4} + k _ {3} \\theta^ {6} + k _ {4} \\theta^ {8}\\right) \\tag {20}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.644, + 0.469, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\theta = \\operatorname {a t a n 2} \\left(\\sqrt {X ^ {2} + Y ^ {2}}, Z\\right), \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.468, + 0.691 + ], + "angle": 0, + "content": "where \\(k_{1},\\ldots ,k_{4},f_{x},f_{y},c_{x},c_{y}\\) are the intrinsic parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.469, + 0.842 + ], + "angle": 0, + "content": "We now consider arbitrary central camera models and derive a generalization of the Sampson error. We assume that we have some forward projection function \\(\\pi : \\mathbb{R}^3 \\to \\Omega\\), where \\(\\Omega \\subset \\mathbb{R}^2\\) is the image plane. In the case of the Kannala-Brandt model this is given by (19), but our derivations work for any differentiable model with central projections, i.e. \\(\\pi(\\lambda d) = \\pi(d) - \\forall \\lambda > 0\\). Similarly we define the (normalized) unprojection, \\(\\pi^{-1} : \\Omega \\to S^2\\) as mapping image points onto the corresponding unit-vectors in 3D. The geometric error (11) then becomes \\(\\mathcal{E}_{ML}(z, E) =\\)" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.853, + 0.469, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\hat {z}} \\| \\hat {z} - z \\| ^ {2} \\tag {22}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.88, + 0.469, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\text {s . t .} \\quad C (\\hat {\\boldsymbol {z}}) = \\pi^ {- 1} \\left(\\hat {\\boldsymbol {p}} _ {2}\\right) ^ {\\intercal} \\boldsymbol {E} \\pi^ {- 1} \\left(\\hat {\\boldsymbol {p}} _ {1}\\right) = 0. \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "In practice, most camera models only define the forward projection \\(\\pi\\), and unprojection \\(\\pi^{-1}\\) is solved for iteratively. For example, (19) does not have a closed-form inverse. This makes it non-trivial to directly compute its Jacobians required for the Sampson approximation. To avoid this computation we rely on the inverse function theorem. The variant that we will use can be found in the book [6]. It is tailored to functions that map higher-dimensional inputs to lower-dimensional outputs. We rephrase it here for reader's convenience, adapting the notation to our needs and losing some generality:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.263, + 0.892, + 0.375 + ], + "angle": 0, + "content": "Theorem 1. [6, Theorem 1E.6] Let \\(\\pi : \\mathbb{R}^3 \\to \\mathbb{R}^2\\) be continuously differentiable in a neighborhood around \\(d \\in \\mathbb{R}^3\\). Further, suppose that its Jacobian \\(\\pmb{J}\\) at \\(d\\) has full rank. Then, there exists a neighborhood \\(V\\) of \\(\\bar{\\pmb{p}} = \\pi(d)\\) and a continuously differentiable function \\(\\widetilde{\\pi^{-1}}: V \\to \\mathbb{R}^3\\) such that \\(\\pi(\\widetilde{\\pi^{-1}}(p)) = p \\forall p \\in V\\) and whose Jacobian is given by \\(\\pmb{J}^\\dagger\\), the Moore-Penrose pseudoinverse of \\(\\pmb{J}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.381, + 0.892, + 0.507 + ], + "angle": 0, + "content": "The construction of \\(\\widetilde{\\pi^{-1}}\\) in the proof from [6] also gives us that \\(\\widetilde{\\pi^{-1}} (\\bar{p}) = d\\). Note that \\(\\pi\\) admits many right-inverses, but they only differ in lengths assigned to unprojected bearing vectors. Thus, \\(\\widetilde{\\pi^{-1}}\\) that the theorem gives us need not coincide with \\(\\pi^{-1}\\) that maps points to unit vectors. However, the following result shows that when we evaluate \\(J\\) at a unit vector \\(d = \\pi^{-1}(\\bar{p})\\), \\(J^{\\dagger}\\) actually gives us the Jacobian of \\(\\pi^{-1}\\):" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.514, + 0.892, + 0.548 + ], + "angle": 0, + "content": "Corollary 1. Let \\( \\pmb{d} \\in S^2 \\) and \\( \\pmb{J} = \\frac{\\partial\\pi}{\\partial\\bar{\\pmb{d}}} \\in \\mathbb{R}^{2\\times 3} \\). Then the Jacobian of \\( \\pi^{-1} \\) at \\( \\bar{\\pmb{p}} = \\pi (\\pmb{d}) \\) is given by \\( \\pmb{J}^{\\dagger} \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.555, + 0.892, + 0.587 + ], + "angle": 0, + "content": "Proof. Since the only difference between \\(\\widetilde{\\pi^{-1}}\\) from theorem 1 and \\(\\pi^{-1}\\) is the norm of the output, we can write" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.594, + 0.891, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\pi^ {- 1} (\\boldsymbol {p}) = h \\left(\\widetilde {\\left(\\pi^ {- 1} (\\boldsymbol {p})\\right)} \\quad \\forall \\boldsymbol {p} \\in V, \\right. \\tag {24}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.62, + 0.892, + 0.669 + ], + "angle": 0, + "content": "where \\( h: \\pmb{v} \\mapsto \\pmb{v} / \\| \\pmb{v} \\| \\) is the normalization function. We also have \\( \\pmb{d} = \\widetilde{\\pi^{-1}}(\\bar{\\pmb{p}}) \\). Thus, the Jacobian of \\( \\pi^{-1} \\) at \\( \\bar{\\pmb{p}} \\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.675, + 0.891, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\pi^ {- 1}}{\\partial \\boldsymbol {p}} = \\left. \\frac {\\partial h}{\\partial \\boldsymbol {v}} \\right| _ {\\boldsymbol {v} = \\boldsymbol {d}} \\frac {\\partial \\widetilde {\\pi^ {- 1}}}{\\partial \\boldsymbol {p}} = (\\boldsymbol {I} - \\boldsymbol {d} \\boldsymbol {d} ^ {\\intercal}) \\boldsymbol {J} ^ {\\dagger} = \\boldsymbol {J} ^ {\\dagger}. \\tag {25}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.875, + 0.719, + 0.892, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.741, + 0.892, + 0.802 + ], + "angle": 0, + "content": "In the last equation we used \\( \\pmb{d}^{\\mathsf{T}}\\pmb{J}^{\\dagger} = \\mathbf{0} \\), which is true in general for any null-vector \\( \\pmb{d} \\) of the matrix \\( J \\), but in our case can also be seen from the following lemma that gives a closed-form expression for \\( \\pmb{J}^{\\dagger} \\):" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.807, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Lemma 1. Let \\( \\pmb{d} \\in \\mathbb{R}^3 \\) and \\( J = \\frac{\\partial\\pi}{\\partial\\pmb{d}} \\in \\mathbb{R}^{2\\times 3} \\). Then, \\( J^{\\dagger} \\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.845, + 0.891, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} = \\frac {1}{\\boldsymbol {d} \\cdot \\left(\\boldsymbol {g} _ {x} \\times \\boldsymbol {g} _ {y}\\right)} \\left[ \\left(\\boldsymbol {g} _ {y} \\times \\boldsymbol {d}\\right), \\quad \\left(\\boldsymbol {d} \\times \\boldsymbol {g} _ {x}\\right) \\right], \\tag {26}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.885, + 0.868, + 0.904 + ], + "angle": 0, + "content": "where \\(\\pmb{g}_x\\), \\(\\pmb{g}_y \\in \\mathbb{R}^3\\) are the two rows of the Jacobian \\(\\pmb{J}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3373" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.14 + ], + "angle": 0, + "content": "Proof. Simple calculations show that \\( JM = I_{2 \\times 2} \\). It remains to show that \\( MJ \\in \\mathbb{R}^{3 \\times 3} \\) is symmetric. For ease of notation, introduce \\( \\alpha = d \\cdot (g_x \\times g_y) \\). Then we have" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.148, + 0.469, + 0.186 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\alpha \\boldsymbol {M} \\boldsymbol {J} = (\\boldsymbol {g} _ {y} \\times \\boldsymbol {d}) \\boldsymbol {g} _ {x} ^ {\\intercal} + (\\boldsymbol {d} \\times \\boldsymbol {g} _ {x}) \\boldsymbol {g} _ {y} ^ {\\intercal} = (27) \\\\ = [ \\boldsymbol {d} ] _ {\\times} \\left(- \\boldsymbol {g} _ {y} \\boldsymbol {g} _ {x} ^ {\\intercal} + \\boldsymbol {g} _ {x} \\boldsymbol {g} _ {y} ^ {\\intercal}\\right) = [ \\boldsymbol {d} ] _ {\\times} [ \\boldsymbol {g} _ {y} \\times \\boldsymbol {g} _ {x} ] _ {\\times}. (28) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.195, + 0.469, + 0.258 + ], + "angle": 0, + "content": "Since the projection function \\(\\pi\\) does not depend on the norm of \\(\\pmb{d}\\), i.e. \\(\\pi(\\pmb{d}) = \\pi(\\lambda \\pmb{d}) \\forall \\lambda > 0\\), we have that \\(\\pmb{J} \\pmb{d} = \\mathbf{0}\\). Since \\(\\pmb{d}\\) is then orthogonal to both \\(\\pmb{g}_x\\) and \\(\\pmb{g}_y\\) we have \\(\\pmb{g}_y \\times \\pmb{g}_x = \\gamma \\pmb{d}\\) for some \\(\\gamma \\in \\mathbb{R}\\). Inserting into (28) we get" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.265, + 0.469, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} \\boldsymbol {J} = \\frac {\\gamma}{\\alpha} [ \\boldsymbol {d} ] _ {\\times} ^ {2}, \\tag {29}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.3, + 0.274, + 0.314 + ], + "angle": 0, + "content": "which is a symmetric matrix." + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.315, + 0.469, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.339, + 0.47, + 0.4 + ], + "angle": 0, + "content": "The expression (26) can be used to efficiently compute the Jacobian \\( J_{C} \\) of the constraint function \\( C \\) from (23), and finally to arrive at the Sampson approximation of the reprojection error in any differentiable camera model:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.407, + 0.469, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {T S} ^ {2} = \\frac {C (\\boldsymbol {z}) ^ {2}}{\\| \\boldsymbol {J} _ {C} \\| ^ {2}} = \\frac {\\left(\\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1}\\right) ^ {2}}{\\| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {J} _ {1} ^ {\\dagger} \\| ^ {2} + \\| \\boldsymbol {d} _ {1} ^ {\\intercal} \\boldsymbol {E} ^ {\\intercal} \\boldsymbol {J} _ {2} ^ {\\dagger} \\| ^ {2}}, \\tag {30}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.453, + 0.47, + 0.65 + ], + "angle": 0, + "content": "where \\( \\pmb{d}_i = \\pi^{-1}(\\pmb{p}_i) \\) and the two Jacobians are evaluated at \\( \\pmb{d}_1 \\) and \\( \\pmb{d}_2 \\) respectively. In the rest of the paper we refer to this as the Tangent Sampson error. The name reflects the fact that after we linearize the unprojection \\( \\pi^{-1} \\), we are essentially optimizing in the tangent plane to \\( S^2 \\) at the bearing vectors of the undistorted points. Note that the formula is analogous to the original Sampson error (18) (replacing \\( \\pmb{d}_1 \\) and \\( \\pmb{d}_2 \\) with the image points, and replacing \\( J_1^\\dagger \\) and \\( J_2^\\dagger \\) with \\( [10; 01; 00] \\)). Further, since \\( J_1 \\) and \\( J_2 \\) are evaluated at the original points of the correspondence, they are independent of \\( \\pmb{E} \\) and can be precomputed. Thus, during RANSAC and refinement the computational cost stays approximately the same as for the classical Sampson error." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.658, + 0.32, + 0.673 + ], + "angle": 0, + "content": "3.1. Backward Camera Models" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.681, + 0.469, + 0.803 + ], + "angle": 0, + "content": "While most camera models used in practice are forward-models, i.e. mapping from the world to the image, there also exist backward-models which instead take the image points as input. The division model (8) is perhaps the most popular example here. The derivations above can easily be adapted to allow for backward-models. The pseudoinverses \\( J_{k}^{\\dagger} \\) are simply replaced with the \\( 3 \\times 2 \\) Jacobians from the undistortion map," + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.81, + 0.469, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {T S} ^ {2} = \\frac {\\left(\\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1}\\right) ^ {2}}{\\left\\| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {J} _ {1} \\right\\| ^ {2} + \\left\\| \\boldsymbol {d} _ {1} ^ {\\intercal} \\boldsymbol {E} ^ {\\intercal} \\boldsymbol {J} _ {2} \\right\\| ^ {2}}. \\tag {31}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.852, + 0.44, + 0.867 + ], + "angle": 0, + "content": "In particular, for the division model (8) we would have" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.875, + 0.469, + 0.907 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {J} = \\frac {\\partial h}{\\partial \\boldsymbol {d}} \\frac {\\partial \\boldsymbol {d}}{\\partial \\boldsymbol {x}} = \\frac {\\partial h}{\\partial \\boldsymbol {d}} \\left[ I _ {2 \\times 2}; 2 \\lambda \\boldsymbol {x} ^ {\\intercal} \\right] \\in \\mathbb {R} ^ {3 \\times 2}, \\tag {32}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.136 + ], + "angle": 0, + "content": "where we need to explicitly include the Jacobian \\(\\frac{\\partial h}{\\partial d}\\) of the normalization function \\(h\\), since the unprojected \\(\\pmb{d}\\) from (8) is not a unit vector." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.145, + 0.738, + 0.161 + ], + "angle": 0, + "content": "3.2. Measurement Uncertainty" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.168, + 0.892, + 0.229 + ], + "angle": 0, + "content": "Following the approach from [3], we can incorporate measurement uncertainty in the keypoint positions into the Tangent Sampson error. We will assume that the keypoint positions are affected by independent Gaussian noise" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.236, + 0.892, + 0.253 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {1} \\sim \\mathcal {N} \\left(\\bar {\\boldsymbol {p}} _ {1}, \\boldsymbol {\\Sigma} _ {1}\\right), \\quad \\boldsymbol {p} _ {2} \\sim \\mathcal {N} \\left(\\bar {\\boldsymbol {p}} _ {2}, \\boldsymbol {\\Sigma} _ {2}\\right) \\tag {33}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.26, + 0.892, + 0.289 + ], + "angle": 0, + "content": "with known covariances. For observations \\( \\pmb{p}_1 \\) and \\( \\pmb{p}_2 \\), the maximum-likelihood error for an essential matrix \\( E \\) is" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.294, + 0.891, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {E} _ {M L} ^ {2} (\\boldsymbol {z}, \\boldsymbol {E}) = \\min _ {\\hat {z}} \\quad \\frac {1}{2} (\\boldsymbol {z} - \\hat {\\boldsymbol {z}}) ^ {\\intercal} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {z} - \\hat {\\boldsymbol {z}}) \\tag {34} \\\\ \\begin{array}{c c} \\text {s . t .} & C (\\hat {\\boldsymbol {z}}) = 0, \\end{array} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.349, + 0.892, + 0.379 + ], + "angle": 0, + "content": "where \\(C\\) is the constraint function (23) and \\(\\pmb{\\Sigma}\\) is the joint covariance of \\(\\pmb{z} = [\\pmb{p}_1; \\pmb{p}_2]\\), i.e. \\(\\pmb{\\Sigma} = \\text{diag}(\\pmb{\\Sigma}_1, \\pmb{\\Sigma}_2)\\)." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.38, + 0.852, + 0.394 + ], + "angle": 0, + "content": "The procedure from Section 2.1 yields in this case" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.399, + 0.891, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {T S} ^ {2} = \\frac {C (\\boldsymbol {z}) ^ {2}}{\\boldsymbol {J} _ {C} \\boldsymbol {\\Sigma} \\boldsymbol {J} _ {C} ^ {\\intercal}} = \\frac {\\left(\\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1}\\right) ^ {2}}{\\| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {J} _ {1} ^ {\\dagger} \\| _ {\\Sigma_ {1}} ^ {2} + \\| \\boldsymbol {d} _ {1} ^ {\\intercal} \\boldsymbol {E} ^ {\\intercal} \\boldsymbol {J} _ {2} ^ {\\dagger} \\| _ {\\Sigma_ {2}} ^ {2}}, \\tag {35}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.44, + 0.892, + 0.516 + ], + "angle": 0, + "content": "where \\(\\| \\pmb{v} \\|_{\\Sigma}^{2} = \\pmb{v}^{T} \\pmb{\\Sigma} \\pmb{v}\\). See the supplementary material for details on the derivation. This is a generalization of the last expression in [3, Section 3.4] to the arbitrary camera model case. We evaluate how it compares to the original Tangent Sampson error (30) in a synthetic setting in Section 4.4.1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.735, + 0.544 + ], + "angle": 0, + "content": "4. Experimental Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.553, + 0.892, + 0.599 + ], + "angle": 0, + "content": "In the rest of the paper we experimentally evaluate the Tangent Sampson error. We compare against a selection of popular error metrics, summarized in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.717, + 0.622 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We implemented all error formulations in \\(\\mathrm{C + + }\\) using Eigen [9] as a linear algebra library and Ceres [1] for optimization. Since evaluation is run on Intel hardware, we additionally speed up the errors that do not require iterative methods or explicit camera projection \\((\\mathcal{E}_{TS},\\mathcal{E}_S,\\mathcal{E}_{SED},\\) \\(\\mathcal{E}_{CS}\\) ,and \\(\\mathcal{E}_{ALG})\\) by vectorizing their implementations with AVX and FMA instruction sets. In Table 1 we also list the approximate computational cost of evaluating the residual. Tangent Sampson error is significantly faster compared to other metrics which also measure the error in the original image, and has comparable, but slightly higher, runtime to the purely geometric and pinhole-based metrics. For the computational cost we include parts which cannot be precomputed (i.e. parts that depend on the essential matrix). For a comparison with the non-optimized errors (using pure Eigen-implementations), we refer the reader to the supplementary material, where we list the runtime for all formulations without explicit vectorization and some setup details." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3374" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.244 + ], + "angle": 0, + "content": "
NameRuntime
Rel.Abs.
\\( \\mathcal{E}_{ALG} \\)(7)Algebraic error↓1.01.3 ns
\\( \\mathcal{E}_{CS} \\)(6)Cosine error↓1.82.4 ns
\\( \\mathcal{E}_{ML} \\)(3)(Pinhole) Repoj. error\\( \\blacktriangleright \\)28.036.7 ns
\\( \\mathcal{E}_{SED} \\)(4)(Pinhole) Sym. Epi. Dist.\\( \\blacktriangleright \\)1.72.3 ns
\\( \\mathcal{E}_{S} \\)(5)(Pinhole) Sampson error\\( \\blacktriangleright \\)1.41.9 ns
\\( \\mathcal{E}_{\\pi ML} \\)(22)Reproj. error\\( \\square \\)8861160 ns
\\( \\mathcal{E}_{\\pi SED} \\)(10)Proj. Sym. Epi. Dist.\\( \\square \\)92121 ns
\\( \\mathcal{E}_{TS} \\)(30)Tangent Sampson\\( \\square \\)2.22.8 ns
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.255, + 0.47, + 0.326 + ], + "angle": 0, + "content": "Table 1. Error metrics compared in the experimental evaluation. The errors are grouped by the domain where the error is computed: geometric errors \\((\\mathcal{L})\\), the undistorted image \\((\\mathfrak{X})\\) and the original image \\((\\square)\\). Computational cost is shown relative to the algebraic error which is the cheapest." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.332, + 0.181, + 0.346 + ], + "angle": 0, + "content": "4.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.596 + ], + "angle": 0, + "content": "For the evaluation we consider two datasets. First the meta-calibration dataset from BabelCalib [21]. The dataset is a collection of checkerboard images from a wide variety of cameras, including some with extremely wide field-of-view \\((>270^{\\circ})\\). For the evaluation we split the dataset into three groups depending on their field-of-view; \\(< 90^{\\circ}\\), \\(90 - 150^{\\circ}\\), and \\(>150^{\\circ}\\). As the original dataset is biased towards higher field-of-view cameras, we also include some additional checkerboard images taken with a standard DSLR (circa \\(50 - 70^{\\circ}\\)) to even out the distribution. These checkerboard datasets allow us to evaluate on a large number of different camera and lens setups. In Section 4.3 we show that the Tangent Sampson error consistently approximates the true reprojection error, regardless of the camera used, and in Section 4.4 we show that when used for local refinement it yields the most accurate epipolar geometries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.749 + ], + "angle": 0, + "content": "To evaluate the performance of the error metrics in the context of RANSAC we consider image pairs extracted from Structure-from-Motion datasets, where the poses from the reconstruction are taken as a psuedo-ground truth. As the experiments in Section 4.3-4.4 show that most error metrics behave similarly for low to medium field-of-view cameras \\((< 90^{\\circ})\\), we focus our evaluation on cameras with larger field-of-view. For this we use two datasets (Grossmunster and Kirchenge) from Larsson et al. [17], which contain 373 and 369 images taken with a fisheye-lens." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.442, + 0.774 + ], + "angle": 0, + "content": "4.3. Comparison with True Reprojection Error" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We first evaluate how different error metrics compare with the true maximum-likelihood estimate, \\(\\mathcal{E}_{ML}\\) (3) for pinhole cameras and \\(\\mathcal{E}_{\\pi ML}\\) (22) for non-pinhole cameras. To compute the ML-estimate we perform two-view triangulation followed by non-linear optimization of the 3D point, minimizing the squared reprojection error (using the known camera intrinsics). Note that this approach is orders of magnitude slower compared to the metrics we evaluate (see Ta" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.107 + ], + "angle": 0, + "content": "ble 1) and cannot reasonably be used in practical scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.108, + 0.892, + 0.257 + ], + "angle": 0, + "content": "For the experiment we consider the collection of checkerboard images from [21]. We use the calibration toolbox from BabelCalib [21] to estimate the extrinsic and intrinsic parameters for each image. For each pair we collect the 2D-2D correspondences from corresponding corners. As the corner detector [10] is extremely accurate on these images (some having around \\(0.2\\mathrm{px}\\) error during calibration), we add Gaussian noise with \\(1\\mathrm{px}\\) standard deviation, to better differentiate the methods (as all methods give zero error for exact data)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.259, + 0.892, + 0.321 + ], + "angle": 0, + "content": "Since it is not meaningful to directly compare the residual values, we instead compare their ordering. For this we make use of Kendall's \\(\\tau\\) rank correlation coefficient [15]. It is defined for two rankings \\((p_1,\\dots ,p_n)\\) and \\((q_{1},\\ldots ,q_{n})\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.342, + 0.892, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\tau (p, q) = \\frac {\\left| \\{\\text {c o n c o r d a n t p a i r s} \\} \\right| - \\left| \\{\\text {d i s c o r d a n t p a i r s} \\} \\right|}{\\binom {n} {2}}. \\tag {36}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.387, + 0.892, + 0.419 + ], + "angle": 0, + "content": "A pair of indices \\((i,j)\\in \\{1,\\ldots ,n\\} ^2\\) where \\(i > j\\) is called concordant if and only if" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.433, + 0.892, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\left(p _ {i} > p _ {j} \\wedge q _ {i} > q _ {j}\\right) \\vee \\left(p _ {i} < p _ {j} \\wedge q _ {i} < q _ {j}\\right). \\tag {37}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.506 + ], + "angle": 0, + "content": "Otherwise, the pair is discordant. Note that due to the normalization we have \\(\\tau(p,q) \\in [-1,1]\\), where identical rankings yield \\(\\tau = 1\\), and reversed \\(\\tau = -1\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.644 + ], + "angle": 0, + "content": "The results are shown in Fig. 3, split into three groups based on the camera field-of-view. For low to medium field-of-view cameras \\((< 90^{\\circ})\\) we can see that all metrics except the algebraic residual behave similar to the true reprojection error. As the field-of-view increases, there is a more clear difference between the methods, with the pinhole-based errors performing the worst. Note however, that for all cameras the proposed Tangent Sampson error offers the best approximation of the true reprojection error." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In addition to Kendall-Tau distance, we compare the errors qualitatively. For this, we show the values of an error for a fixed correspondence on a pair of images from a panoramic and fisheye camera in Fig. 4. There, we keep the point on the fisheye image fixed, and vary the point on the panoramic image to show different error values coded by different colors. One can see that \\(\\mathcal{E}_{TS}\\) error models \\(\\mathcal{E}_{\\pi ML}\\) closely in the region where the latter is well-defined. \\(\\mathcal{E}_{TS}\\) is also one of the two errors that do not degenerate around the epipole, and, at the same time, it correctly models the spread of values of \\(\\mathcal{E}_{\\pi ML}\\) around the epipolar curve. Note that the shape of true reprojection error varies significantly depending on how close the correspondence is to the epipole. In the supplementary material we show more qualitative results. In the supplementary material we also show a comparison with [16], which proposed an optimal method for computing the \\(\\pi ML\\) error (22) for the division model (8)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3375" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.089, + 0.098, + 0.38, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.097, + 0.616, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.097, + 0.852, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.892, + 0.315 + ], + "angle": 0, + "content": "Figure 3. The figure shows the Kendall \\(\\tau\\) coefficient comparing different error metrics for the checkerboard data. For all cameras the proposed Tangent Sampson error correlates well with the true reprojection error \\(\\pi ML\\)." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.318, + 0.287, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.318, + 0.484, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.318, + 0.683, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.318, + 0.882, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.407, + 0.895, + 0.478 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of different error functions. The thresholds for colormaps were selected to be the error at the point marked by the blue plus. The shape of the true reprojection error varies heavily depending on how close the correspondence is to the epipoles. In the supplementary material we show more examples. We also show \\(\\mathcal{E}_{\\pi SED}\\) and errors computed on the undistorted image in the supplementary. The geometric error is only shown in the region where the 3D point after midpoint triangulation does not correspond to negative depths of bearing vectors (areas where this does not hold for one or both depths are highlighted)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.484, + 0.339, + 0.5 + ], + "angle": 0, + "content": "4.4. Accuracy in Pose Refinement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Next we evaluate the error metrics on how they perform when used for pose-refinement. For the experiment we again consider the image pairs extracted from the checkerboard images. For each image pair we perform local optimization minimizing the proposed error metrics over all correspondences. To initialize the optimization we take the ground truth relative pose and randomly perturb both the rotation and translation by 1 degree. Additionally, point positions were perturbed with Gaussian noise with unit standard deviation along each dimension. For the two ML metrics we triangulate 3D points which are optimized jointly with the camera pose. We experience that this sometimes has problems with worse convergence. The results are shown in Table 2. Again we see that for low to medium field-of-view cameras, the difference between the error metrics is small. The Tangent Sampson error consistently provides the most accurate camera poses after refinement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Convergence Properties. The previous experiment evaluated the accuracy of the camera poses obtained when initialized close to the correct pose. We also experimented with varying the initial pose and measuring how often the optimization converged to the correct pose. Figure 5 shows the percentage of poses (across all pairs) that are within 5 degrees average rotation/translation error for varying errors in the initial pose. For large initial errors \\((>25^{\\circ})\\), the proposed Tangent-Sampson error performs slightly better com" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.486, + 0.891, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.647, + 0.892, + 0.675 + ], + "angle": 0, + "content": "Figure 5. Convergence properties. Successful trials (average rot. and trans. error \\(< 5^{\\circ}\\)) for varying initial pose errors (in degrees)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.685, + 0.678, + 0.7 + ], + "angle": 0, + "content": "pared to the other metrics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.803, + 0.74 + ], + "angle": 0, + "content": "4.4.1 Effect of Measurement Uncertainty" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Experiments in [3] already show that taking point covariances into account in the cost function provides benefits for fundamental matrix estimation. Here, we conduct another synthetic test to show that this effect persists when estimating the essential matrix with highly nonlinear cameras. We generate 2000 instances with pairs of spherical cameras \\((1000 \\times 2000\\mathrm{px})\\) with a random relative pose. For each pair we project 100 random 3D-points uniformly drawn from the cube \\([-10, 10]^3\\), and add Gaussian noise with covariance \\(\\Lambda\\)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Covariances are generated as \\(\\Lambda = \\alpha Q\\mathrm{diag}(\\beta ,1 - \\beta)Q^{\\intercal}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3376" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.868, + 0.263 + ], + "angle": 0, + "content": "
< 90° (5332 pairs)90 - 150° (21927 pairs)> 150° (11860 pairs)
RotationTranslationRotationTranslationRotationTranslation
\\( \\mathcal{E}_{ALG} \\)0.379 / 0.3141.137 / 0.3281.528 / 0.3783.938 / 0.5881.582 / 0.5063.489 / 0.655
\\( \\mathcal{E}_S \\)0.376 / 0.3121.056 / 0.3271.464 / 0.3473.520 / 0.5751.634 / 0.5673.812 / 0.778
\\( \\mathcal{E}_{SED} \\)0.376 / 0.3121.065 / 0.3271.475 / 0.3523.554 / 0.5782.596 / 0.9865.667 / 1.342
\\( \\mathcal{E}_{COS} \\)0.376 / 0.3131.066 / 0.3261.454 / 0.3413.459 / 0.5581.447 / 0.4543.235 / 0.612
\\( \\mathcal{E}_{ML} \\)0.376 / 0.3121.103 / 0.3271.472 / 0.3583.733 / 0.5751.630 / 0.5633.771 / 0.766
\\( \\mathcal{E}_{\\pi ML} \\)0.375 / 0.3121.075 / 0.3271.451 / 0.3503.778 / 0.5571.418 / 0.4463.143 / 0.597
\\( \\mathcal{E}_{\\pi SED} \\)0.376 / 0.3131.074 / 0.3271.458 / 0.3433.490 / 0.5591.473 / 0.4503.249 / 0.607
\\( \\mathcal{E}_{TS} \\)0.376 / 0.3121.057 / 0.3271.443 / 0.3403.432 / 0.5571.465 / 0.4453.226 / 0.598
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.273, + 0.893, + 0.318 + ], + "angle": 0, + "content": "Table 2. Evaluation of camera pose refinement. Table shows the mean and median errors (in degrees) of the rotation and translation after refinement. The data is split into three groups based on the camera field-of-view. Best error is highlighted in bold. Second-best is underlined. For low-distortion cameras the choice of error makes almost no difference on the quality of pose." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.321, + 0.456, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.435, + 0.471, + 0.519 + ], + "angle": 0, + "content": "Figure 6. Measurement Uncertainty. The distribution of pose errors (maximum of rotation and translation error) when using covariance-aware Tangent Sampson error for pose refinement versus non-covariance-aware in synthetic experiments. Boxes show quartiles of the respective distribution of errors and whiskers extend to the minimal and maximal value." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.712 + ], + "angle": 0, + "content": "where \\( Q \\sim U(\\mathrm{SO}(2)) \\) is a random rotation on the plane, \\( \\beta \\sim U([0, 0.5]) \\) is a random eccentricity parameter, and \\( \\alpha \\) is the overall scale of the noise. To emulate keypoint detection on a multi-scale image pyramid, all generated points were split into 4 equal sets, with \\( \\alpha \\) within each set assigned to 1, 2, 4, and 8, respectively. Points and covariances are then passed to a Levenberg-Marquardt optimization of the relative pose using (35) as the residual function, with true pose serving as the initial estimate. Figure 6 demonstrates how the covariance-aware formulation (35) compares to the baseline (30). Covariance information is useful in the case when points are affected by noise of unequal scales." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.365, + 0.74 + ], + "angle": 0, + "content": "4.5. Evaluation in Robust Estimators" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Finally we evaluate the methods in the context of RANSAC [7]. We implement a standard LO-RANSAC [18] estimator which uses the various error metrics both for model scoring (using MSAC scoring [24]) and refinement (using non-linear optimization on the inlier-set). For the experiment we consider image pairs from two Structure-from-Motion reconstructions (see Section 4.2). From each dataset we extract all image pairs which contain between 50 and 200 correct matches (according to the Structure-from-Motion ground-truth). For each pair we estimate the rela" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.32, + 0.905, + 0.466 + ], + "angle": 0, + "content": "
GrossmunsterKirchenge
εR,t< 5°RTεR,t< 5°RT
ΕALG0.76175.039.60.31884.821.6
ΕCOS0.56181.550.40.25491.625.5
ΕSED1.16168.982.00.35586.444.0
ΕS0.92474.551.00.30089.225.3
ΕπSED0.56181.3211.20.25591.687.4
ΕTS0.55082.246.40.24891.924.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.474, + 0.893, + 0.545 + ], + "angle": 0, + "content": "Table 3. Evaluation in robust estimators. Table shows the median pose error (maximum of rotation and translation error), percentage of pairs for which the estimation was successful (pose error less than 5 degrees), and runtime (in ms). Best value in the column is highlighted in bold. Second-best is underlined." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.893, + 0.671 + ], + "angle": 0, + "content": "tive pose using RANSAC and compute the error in the pose, as well as percentage of successful runs. For each method we tuned the inlier threshold on a separately extracted validation set. For more details on the experiment setup see the supplementary material. The results are presented in Table 3. Using Tangent Sampson error consistently gives slightly better performance compared to other error metrics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.698, + 0.628, + 0.714 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.894, + 0.851 + ], + "angle": 0, + "content": "In this paper we have presented a simple generalization to the classical Sampson error which allows for using arbitrary camera models. In experiments on a wide variety of camera and lens configurations, we have shown that the proposed error metric outperforms the alternative metrics that are used in practice. While the improvements are sometimes minor (e.g. for low-distortion pinhole cameras), they are consistent and come at essentially no additional cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.857, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments: This project received funding from the strategic research program ELLIIT. MT was supported by the ESOP scholarship from the ETH Foundation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3377" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Sameer Agarwal, Keir Mierle, and Others. Ceres solver. http://ceres-solver.org.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.468, + 0.187 + ], + "angle": 0, + "content": "[2] João Pedro Barreto and Kostas Daniilidis. Fundamental matrix for cameras with radial distortion. In International Conference on Computer Vision (ICCV), 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.189, + 0.468, + 0.245 + ], + "angle": 0, + "content": "[3] Wojciech Chojnacki, Michael J. Brooks, Anton Van Den Hengel, and Darren Gawley. On the fitting of surfaces to data with covariances. IEEE Trans. Pattern Analysis and Machine Intelligence (PAMI), 2000. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.248, + 0.468, + 0.288 + ], + "angle": 0, + "content": "[4] Ondřej Chum, Tomáš Pajdla, and Peter Sturm. The geometric error for homographies. Computer Vision and Image Understanding (CVIU), 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.468, + 0.332 + ], + "angle": 0, + "content": "[5] David Claus and Andrew W Fitzgibbon. A rational function lens distortion model for general cameras. In Computer Vision and Pattern Recognition (CVPR), 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[6] Asen L Dontchev and R Tyrrell Rockafellar. Implicit functions and solution mappings, volume 543. Springer, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.378, + 0.468, + 0.433 + ], + "angle": 0, + "content": "[7] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.436, + 0.468, + 0.477 + ], + "angle": 0, + "content": "[8] Andrew W Fitzgibbon. Simultaneous linear estimation of multiple view geometry and lens distortion. In Computer Vision and Pattern Recognition (CVPR), 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.479, + 0.468, + 0.508 + ], + "angle": 0, + "content": "[9] Gaël Guennebaud, Benoit Jacob, et al. Eigen v3. http://eigen.tuxfamily.org, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.51, + 0.468, + 0.564 + ], + "angle": 0, + "content": "[10] Hyowon Ha, Michal Perdoch, Hatem Alismail, In So Kweon, and Yaser Sheikh. Deltille grids for geometric camera calibration. In International Conference on Computer Vision (ICCV), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.468, + 0.607 + ], + "angle": 0, + "content": "[11] Matthew Harker and Paul O'Leary. First order geometric distance (the myth of sampsonus). In *British Machine Vision Conference (BMVC)*, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.611, + 0.468, + 0.651 + ], + "angle": 0, + "content": "[12] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.655, + 0.468, + 0.682 + ], + "angle": 0, + "content": "[13] Richard I Hartley and Peter Sturm. Triangulation. Computer vision and image understanding, 68(2):146-157, 1997. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.685, + 0.468, + 0.725 + ], + "angle": 0, + "content": "[14] Juho Kannala and Sami S Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.728, + 0.468, + 0.755 + ], + "angle": 0, + "content": "[15] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1/2):81-93, 1938. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.468, + 0.799 + ], + "angle": 0, + "content": "[16] Zuzana Kukelova and Viktor Larsson. Radial distortion triangulation. In Computer Vision and Pattern Recognition (CVPR), 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[17] Viktor Larsson, Nicolas Zobernig, Kasim Taskin, and Marc Pollefeys. Calibration-free structure-from-motion with calibrated radial trifocal tensors. In European Conference on Computer Vision (ECCV), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[18] Karel Lebeda, Jiri Matas, and Ondrej Chum. Fixing the locally optimized ransac-full experimental evaluation. In British Machine Vision Conference (BMVC), 2012. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[19] Seong Hun Lee and Javier Civera. Geometric interpretations of the normalized epipolar error. arXiv preprint arXiv:2008.01254, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[20] Peter Lindstrom. Triangulation made easy. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 1554-1561. IEEE, 2010. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[21] Yaroslava Lochman, Kostiantyn Liepieshov, Jianhui Chen, Michal Perdoch, Christopher Zach, and James Pritts. Babel-calib: A universal approach to calibrating central cameras. In International Conference on Computer Vision (ICCV), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[22] Quan-Tuan Luong and Olivier D Faugeras. The fundamental matrix: Theory, algorithms, and stability analysis. International Journal of Computer Vision (IJCV), 1996. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[23] Paul D Sampson. Fitting conic sections to \"very scattered\" data: An iterative refinement of the bookstein algorithm. Computer graphics and image processing, 1982. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[24] Philip HS Torr and Andrew Zisserman. Mlesac: A new robust estimator with application to estimating image geometry. Computer Vision and Image Understanding (CVIU), 2000, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[25] Zhengyou Zhang. On the epipolar geometry between two images with lens distortion. In International Conference on Pattern Recognition (ICPR), 1996. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3378" + } + ] +] \ No newline at end of file diff --git a/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/636b4c99-6986-4080-bf56-9a578e79f107_origin.pdf b/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/636b4c99-6986-4080-bf56-9a578e79f107_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a44864bf3434288d2bccceea31ac80b7efe716d9 --- /dev/null +++ b/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/636b4c99-6986-4080-bf56-9a578e79f107_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bfc072029616b91d30af986894bf2cc11e36fa27e69ae3afe4039d662500dfd +size 9232844 diff --git a/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/full.md b/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a2ba6e0d78ace43e6e0becda498f7eaff8db2504 --- /dev/null +++ b/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/full.md @@ -0,0 +1,436 @@ +# Tangent Sampson Error: Fast Approximate Two-view Reprojection Error for Central Camera Models + +Mikhail Terekhov ETH Zurich + +mterekhov@ethz.ch + +Viktor Larsson + +Lund University + +viktor.larsson@math.lth.se + +![](images/fafdb5c2548aa683e1fc4b744598dc60f15672810a3bf881ffb1b04cd5c31455.jpg) +Figure 1. The figure shows the level sets of the proposed Tangent Sampson error. In contrast to classical Sampson error, the proposed error function applies to arbitrary central camera models; in this case a 360 panoramic image and a fisheye camera. Given a relative pose and a pair of points (highlighted in red), we vary a point on one image while keeping the other fixed, and thus get the values on each image. A given color corresponds to a constant value of reprojection error, with purple indicating near-zero errors along the epipolar curve. + +![](images/2bc68f527d4fd5a4e586c6eff8e4d9739f7e44827fa75707f9ae1a038caf3244.jpg) + +# Abstract + +In this paper we introduce the Tangent Sampson error, which is a generalization of the classical Sampson error in two-view geometry that allows for arbitrary central camera models. It only requires local gradients of the distortion map at the original correspondences (allowing for pre-computation) resulting in a negligible increase in computational cost when used in RANSAC or local refinement. The error effectively approximates the true-reprojection error for a large variety of cameras, including extremely wide field-of-view lenses that cannot be undistorted to a single pinhole image. We show experimentally that the new error outperforms competing approaches both when used for model scoring in RANSAC and for non-linear refinement of the relative camera pose. + +# 1. Introduction + +Estimating two-view geometry from a set of sparse 2D-2D correspondences is a common sub-problem in many vision applications. Each correspondence $(\pmb{p}_1, \pmb{p}_2) \in \mathbb{R}^2 \times \mathbb{R}^2$ + +constrains the geometry via the epipolar constraint, + +$$ +\left(\boldsymbol {p} _ {2}; 1\right) ^ {\top} \boldsymbol {E} \left(\boldsymbol {p} _ {1}; 1\right) = 0, \tag {1} +$$ + +where $\pmb{E} = [t]_{\times}\pmb{R} \in \mathbb{R}^{3\times 3}$ is the essential matrix and the points $p_1, p_2$ are on the normalized image plane. This constraint ensures that the 2D-2D correspondence can be exactly triangulated (i.e. with zero reprojection error) using the camera pair extracted from $\pmb{E}$ . However when the measurement $(p_1, p_2)$ is inexact, it is natural to measure how close this correspondence is to being consistent with a given epipolar geometry. This is for example used for determining inlier matches during robust estimation, or as a loss to be minimized when refining an initial estimate of $\pmb{E}$ . + +Under the assumption of i.i.d. Gaussian noise, the ML-estimate of the closest consistent correspondence, sometimes called the gold-standard estimate in the literature [12], is found by minimizing the squared reprojection error, + +$$ +\mathcal {E} _ {M L} ^ {2} = \min _ {\boldsymbol {X} \in \mathbb {R} ^ {3}} \| \boldsymbol {p} _ {1} - \pi (\boldsymbol {X}) \| ^ {2} + \| \boldsymbol {p} _ {2} - \pi (\boldsymbol {R} \boldsymbol {X} + t) \| ^ {2}, \tag {2} +$$ + +where $\pi : \mathbb{R}^3 \to \mathbb{R}^2$ is the camera projection function. For identical pinhole cameras with unit aspect ratio and no + +skew, this can be reformulated as finding the corrected normalized image points, i.e. the closest pair of 2D points that satisfy the epipolar constraint exactly, + +$$ +\begin{array}{l} \mathcal {E} _ {M L} ^ {2} = \min _ {\hat {\boldsymbol {p}} _ {1}, \hat {\boldsymbol {p}} _ {2} \in \mathbb {R} ^ {2}} \| \boldsymbol {p} _ {1} - \hat {\boldsymbol {p}} _ {1} \| ^ {2} + \| \boldsymbol {p} _ {2} - \hat {\boldsymbol {p}} _ {2} \| ^ {2} \tag {3} \\ \begin{array}{l} \text {s . t .} \quad (\hat {\boldsymbol {p}} _ {2}; 1) ^ {\intercal} \boldsymbol {E} (\hat {\boldsymbol {p}} _ {1}; 1) = 0. \end{array} \\ \end{array} +$$ + +The closed-form solution to (3) can be found as a root to a degree-six polynomial (see [13]), or via iterative methods [20]. In robust estimators, e.g. RANSAC, it is often necessary to evaluate these residuals thousands (or potentially millions) of times for each image pair, which in practice means that cheaper approximate error metrics are favored. + +These approximate or surrogate error metrics can be divided into two categories: image-based (approximating the true reprojection error) or geometry-based (measuring some geometric quantity, e.g. angles, instead). In the first category we have the epipolar distance, i.e. distance from the point to the corresponding epipolar line. This can be interpreted as assuming zero noise on one of the keypoints. The error is often computed symmetrically and thus called the symmetric epipolar distance, having the form + +$$ +\mathcal {E} _ {S E D} ^ {2} = \frac {\left| \boldsymbol {d} _ {2} ^ {\intercal} \boldsymbol {E} \boldsymbol {d} _ {1} \right| ^ {2}}{\left\| \boldsymbol {E} _ {1 2} \boldsymbol {d} _ {1} \right\| ^ {2}} + \frac {\left| \boldsymbol {d} _ {2} ^ {\intercal} \boldsymbol {E} \boldsymbol {d} _ {1} \right| ^ {2}}{\left\| (\boldsymbol {E} ^ {\intercal}) _ {1 2} \boldsymbol {d} _ {2} \right\| ^ {2}}, \tag {4} +$$ + +where $E_{12} \in \mathbb{R}^{2 \times 3}$ refers to the first two rows of $\pmb{E}$ , and $\pmb{d}_k = (\pmb{p}_k; 1) \in \mathbb{R}^3$ . Another image-based error is the Sampson error, which is a first-order approximation of $\mathcal{E}_{ML}$ (see Sec. 2.1 for details): + +$$ +\mathcal {E} _ {S} ^ {2} = \frac {\left| \boldsymbol {d} _ {2} ^ {\intercal} \boldsymbol {E} \boldsymbol {d} _ {1} \right| ^ {2}}{\left\| \boldsymbol {E} _ {1 2} \boldsymbol {d} _ {1} \right\| ^ {2} + \left\| (\boldsymbol {E} ^ {\intercal}) _ {1 2} \boldsymbol {d} _ {2} \right\| ^ {2}}. \tag {5} +$$ + +Both of these errors can after appropriate rescaling (using the focal lengths) be expressed in pixels and thus are natural to interpret in terms of the keypoint detection noise levels. + +In the second category, an example is the cosine error, + +$$ +\mathcal {E} _ {C S} ^ {2} = \frac {\left| \boldsymbol {d} _ {2} ^ {\intercal} \boldsymbol {E} \boldsymbol {d} _ {1} \right| ^ {2}}{\left\| \boldsymbol {d} _ {2} \right\| ^ {2} \left\| \boldsymbol {E} \boldsymbol {d} _ {1} \right\| ^ {2}} + \frac {\left| \boldsymbol {d} _ {2} ^ {\intercal} \boldsymbol {E} \boldsymbol {d} _ {1} \right| ^ {2}}{\left\| \boldsymbol {E} ^ {\intercal} \boldsymbol {d} _ {2} \right\| ^ {2} \left\| \boldsymbol {d} _ {1} \right\| ^ {2}}, \tag {6} +$$ + +which (symmetrically) measures the cosine between the normal of the epipolar plane $\pmb{E}\pmb{d}_1 \in \mathbb{R}^3$ and corresponding point $d_2 \in \mathbb{R}^3$ . Another geometric error is the algebraic residual, + +$$ +\mathcal {E} _ {A L G} = \left| \boldsymbol {d} _ {2} ^ {\mathrm {T}} \boldsymbol {E} \boldsymbol {d} _ {1} \right|, \tag {7} +$$ + +which was given a geometric interpretation in [19]. + +Most of the literature has focused on the pinhole camera case, and the image-based errors above $(\mathcal{E}_{SED}$ and $\mathcal{E}_S)$ measure the error in the image plane assuming this model. For non-pinhole cameras, we either have to undistort the image such that $\mathcal{E}_{SED}$ and $\mathcal{E}_S$ are applicable, or use the + +![](images/fb436acb963ff1442f198de14775181020643ebdeb8ace826562f59690f7b384.jpg) +Figure 2. Undistorting the image deforms the level-sets of the reprojection error and introduces a radial bias in the optimization. + +![](images/63be9b5cc0d75ed8e40990fe8c129ea9506a70f1c529bf9f99521b86e25b1e1e.jpg) + +geometry-based error metrics (e.g. $\mathcal{E}_{CS}$ or $\mathcal{E}_{ALG}$ ). However, undistorting to a pinhole image (which is not always possible for large field-of-view cameras) will introduce a radial-bias into the error (see Figure 2). + +In this paper we instead propose the Tangent Sampson error $\mathcal{E}_{TS}$ , a version of the classical Sampson error that has been generalized to handle arbitrary camera models. Compared to current alternatives it has the following benefits; + +- It is cheap to evaluate (similar in cost to the original Sampson error) since it only requires computing the camera model forward-projection and Jacobian once (which can be pre-computed for each point). +- Since it measures errors in the original image, it is easier to set meaningful thresholds in pixels. This is in contrast to geometric error functions, where it is not always clear how to relate the thresholds (e.g. in angles) to the keypoint detection noise in pixels. +- It works for arbitrary central camera models, only requiring that we can compute a Jacobian of the camera projection function at each correspondence. This also allows the formulation to work for cameras modeled via look-up tables, where the Jacobian can be estimated via interpolation of finite differences. + +In Section 2.1 we first recall the derivations for the original Sampson error. Next, in Section 3 we propose our generalization to arbitrary central camera models, and in Section 4 we experimentally show that the proposed error outperforms the competing methods on a variety of camera and lens configurations. + +# 2. Related Work + +The Sampson error was introduced in [23] to approximate the geometric distance from a point to a conic. The proposed approach was to linearize the equation of the conic around the given point, rendering a closed-form solution for the new distance, now from the point to a line. This idea proved useful for estimating distances to other implicit surfaces, notably those set by constraints of a fundamental matrix [22]. + +The fundamental matrix is only applicable to perspective cameras, and a lot of subsequent work in computer vision + +focused on generalizing it to non-linear camera models, together with some notion of a reprojection error. A well-known example is the division model from Fitzgibbon [8]. It maps a point $\pmb{p}$ on the normalized image plane to the corresponding bearing vector $\pmb{d}$ by + +$$ +\boldsymbol {d} = (\boldsymbol {p}; 1 + \lambda \| \boldsymbol {p} \| ^ {2}). \tag {8} +$$ + +The original paper [8] already discusses a way of approximating the gold-standard error for the new model: first, undistort the original points of the correspondence; next, apply the closed-form triangulation from [13], and distort the projections back onto the original image. The resulting points are then approximate minimizers of (2). A faster iterative scheme for the optimal error was later proposed in [16], which adapts the idea of [20], originally developed for pinhole cameras, to the division model. + +Several camera models with closed-form expressions for epipolar curves were proposed. In [25], epipolar curves are represented by cubics, in the rational model [5] — by conics, and in [2] — by circles. With such models, we can use the Sampson approximation to estimate the distance from one point of the correspondence to the epipolar curve of the other, and vice versa. Note that this gives us an approximation of the symmetric epipolar distance, not the Sampson error, because for the latter we would need to optimize both point positions simultaneously, and this would move the epipolar curves as well. + +All these models were proposed with the goal of searching for the intrinsic parameters together with the epipolar geometry. Indeed, [2] and [5] even provide their own versions of a fundamental matrix on a lifted image space. These matrices can be estimated directly from point correspondences, and then decomposed to provide the distortion parameters and relative pose up to an irreducible ambiguity. We, on the other hand, constrain ourselves to the calibrated case, and thus can afford much greater generality. + +There are also various heuristic generalizations of the symmetric epipolar distance to general non-linear camera models that have been used in practice. One approach is to project the bearing vector of one point of the correspondence onto the epipolar plane given by the other. The projected vector is then mapped back onto the image to provide an approximate corrected point, for which the error can be computed. Mathematically, this is expressed as + +$$ +\begin{array}{l} \boldsymbol {n} _ {1} = \boldsymbol {E} \boldsymbol {d} _ {1} / \| \boldsymbol {E} \boldsymbol {d} _ {1} \|, \quad \boldsymbol {n} _ {2} = \boldsymbol {E} ^ {\top} \boldsymbol {d} _ {2} / \| \boldsymbol {E} ^ {\top} \boldsymbol {d} _ {2} \|, \tag {9} \\ \mathcal {E} _ {\pi S E D} ^ {2} = \left\| \boldsymbol {p} _ {1} - \pi \left(\boldsymbol {d} _ {1} - \boldsymbol {n} _ {2} \boldsymbol {n} _ {2} ^ {\intercal} \boldsymbol {d} _ {1}\right) \right\| ^ {2} \\ + \left\| \boldsymbol {p} _ {2} - \pi \left(\boldsymbol {d} _ {2} - \boldsymbol {n} _ {1} \boldsymbol {n} _ {1} ^ {\intercal} \boldsymbol {d} _ {2}\right) \right\| ^ {2} \\ \end{array} +$$ + +where $p_1$ and $p_2$ are the points on the images, $d_1$ and $d_2$ are the corresponding bearing vectors, $n_1$ and $n_2$ are the normals of the epipolar planes. We have encountered this formulation in private codebases, but we are not aware of its + +origin. We call it the projective symmetric epipolar error. While this does not give us exact distances to the epipolar curves even in the pinhole case, it often provides a reasonable approximation. The main problem with this approach is that it requires calculating $\pi(p)$ every time, making it rather computationally heavy for some camera models. + +From the probabilistic perspective, the geometric error is only justified as long as the noise in the correspondences stays standard Gaussian and i.i.d. A natural generalization is to allow for arbitrary covariances in point positions. For example, when the keypoints are searched on a multi-resolution image pyramid, the standard deviation of the position can be taken as the downscaling factor of the level on which the point was found. For conics and fundamental matrices, covariance-aware error formulations were derived in [3]. We will show how this approach is generalized to our proposed error formulation in Section 3.2. + +In [11] the authors propose another first-order approximation of the reprojection error. The method builds on the optimal triangulation method from [13] and differentates the sixth-order polynomial that has the optimal geometric error as its first root. The first-order approximation is then given by one step of the Newton-Raphson iteration on this polynomial starting from zero. In the paper, this error is only derived for conics. While it would also work for fundamental matrices, this does not generalize to the nonlinear camera model setting we are interested in. + +Finally, we would like to note that the Sampson error can also be derived for homographies [4]. Our method is applicable there as well, but with larger computational cost, which somewhat defeats its original purpose. + +# 2.1. The Sampson Approximation + +Traditional Sampson error [22, 23] approximates the geometric error, i.e. the distance to the closest pair of points which satisfy the epipolar constraint. Given a correspondence $\mathbf{z} = (\pmb{p}_1, \pmb{p}_2) \in \mathbb{R}^2 \times \mathbb{R}^2$ and essential matrix $\pmb{E} \in \mathbb{R}^{3 \times 3}$ , this can be formalized as + +$$ +\begin{array}{l} \mathcal {E} _ {M L} ^ {2} (\boldsymbol {z}, \boldsymbol {E}) = \min _ {\hat {\boldsymbol {z}}} \| \hat {\boldsymbol {z}} - \boldsymbol {z} \| ^ {2} \tag {11} \\ \begin{array}{c c} \text {s . t .} & C (\hat {z}) = 0, \end{array} \\ \end{array} +$$ + +where $\hat{\pmb{z}} = (\hat{\pmb{p}}_1,\hat{\pmb{p}}_2)\in \mathbb{R}^2\times \mathbb{R}^2$ is the corrected correspondence and the epipolar constraint is + +$$ +C (\hat {\boldsymbol {z}}) = \left(\hat {\boldsymbol {p}} _ {2}; 1\right) ^ {\intercal} \boldsymbol {E} \left(\hat {\boldsymbol {p}} _ {1}; 1\right) = 0. \tag {12} +$$ + +Since (11) does not admit a simple closed form solution, the idea in [23] is to linearize the epipolar constraint, $C(\hat{z}) = 0$ at the original correspondence $\mathcal{Z}$ , i.e. + +$$ +\mathcal {E} _ {S} ^ {2} (\boldsymbol {z}, \boldsymbol {E}) = \min _ {\hat {\boldsymbol {z}}} \| \hat {\boldsymbol {z}} - \boldsymbol {z} \| ^ {2} \tag {13} +$$ + +$$ +\text {s . t .} \quad C (\boldsymbol {z}) + \boldsymbol {J} _ {C} (\hat {\boldsymbol {z}} - \boldsymbol {z}) = 0, +$$ + +where $J_{C} = \partial C(z) / \partial z$ is the Jacobian of the constraint, evaluated at $z$ . Introducing a Lagrangian for (13), + +$$ +\mathcal {L} (\hat {z}, \lambda) = \| \hat {z} - z \| ^ {2} + \lambda (C (z) + J _ {C} (\hat {z} - z)), \tag {14} +$$ + +we get the first-order constraints as + +$$ +\hat {z} = z - \frac {\lambda}{2} \boldsymbol {J} _ {C} ^ {\intercal}, \quad C (z) + \boldsymbol {J} _ {C} (\hat {z} - z) = 0. \tag {15} +$$ + +Inserting the first equation into the second yields + +$$ +C (\boldsymbol {z}) - \frac {\lambda}{2} \| \boldsymbol {J} _ {C} \| ^ {2} = 0 \Rightarrow \hat {\boldsymbol {z}} = \boldsymbol {z} - \frac {C (\boldsymbol {z})}{\| \boldsymbol {J} _ {C} \| ^ {2}} \boldsymbol {J} _ {C} ^ {\intercal}. \tag {16} +$$ + +Thus the minimum in (13) is given by + +$$ +\mathcal {E} _ {S} ^ {2} = \left\| \frac {C (\boldsymbol {z})}{\| \boldsymbol {J} _ {C} \| ^ {2}} \boldsymbol {J} _ {C} ^ {\intercal} \right\| ^ {2} = \frac {C (\boldsymbol {z}) ^ {2}}{\| \boldsymbol {J} _ {C} \| ^ {2}}. \tag {17} +$$ + +Replacing $C(z)$ with the epipolar constraint, we arrive at the classical formula for the Sampson error: + +$$ +\mathcal {E} _ {S} ^ {2} = \frac {\left(\left(\boldsymbol {p} _ {2} ; 1\right) ^ {\intercal} \boldsymbol {E} \left(\boldsymbol {p} _ {1} ; 1\right)\right) ^ {2}}{\left\| \boldsymbol {E} _ {1 2} \left(\boldsymbol {p} _ {1} ; 1\right) \right\| ^ {2} + \left\| \left(\boldsymbol {E} ^ {\intercal}\right) _ {1 2} \left(\boldsymbol {p} _ {2} ; 1\right) \right\| ^ {2}}. \tag {18} +$$ + +# 3. Generalizing the Sampson Error + +The Sampson error derived in the previous section assumes a pinhole projection model. However, many cameras, in particular those with wide field-of-view, require a more complex model. One popular choice is the Kannala-Brandt [14] model in which a 3D point $\pmb{d} = (X,Y,Z)^{\mathrm{T}}$ in the camera coordinate system is projected as + +$$ +\pi (\boldsymbol {d}) = \frac {\theta_ {d}}{\sqrt {X ^ {2} + Y ^ {2}}} \binom {f _ {x} X} {f _ {y} Y} + \binom {c _ {x}} {c _ {y}} \tag {19} +$$ + +$$ +\theta_ {d} = \theta \left(1 + k _ {1} \theta^ {2} + k _ {2} \theta^ {4} + k _ {3} \theta^ {6} + k _ {4} \theta^ {8}\right) \tag {20} +$$ + +$$ +\theta = \operatorname {a t a n 2} \left(\sqrt {X ^ {2} + Y ^ {2}}, Z\right), \tag {21} +$$ + +where $k_{1},\ldots ,k_{4},f_{x},f_{y},c_{x},c_{y}$ are the intrinsic parameters. + +We now consider arbitrary central camera models and derive a generalization of the Sampson error. We assume that we have some forward projection function $\pi : \mathbb{R}^3 \to \Omega$ , where $\Omega \subset \mathbb{R}^2$ is the image plane. In the case of the Kannala-Brandt model this is given by (19), but our derivations work for any differentiable model with central projections, i.e. $\pi(\lambda d) = \pi(d) - \forall \lambda > 0$ . Similarly we define the (normalized) unprojection, $\pi^{-1} : \Omega \to S^2$ as mapping image points onto the corresponding unit-vectors in 3D. The geometric error (11) then becomes $\mathcal{E}_{ML}(z, E) =$ + +$$ +\min _ {\hat {z}} \| \hat {z} - z \| ^ {2} \tag {22} +$$ + +$$ +\text {s . t .} \quad C (\hat {\boldsymbol {z}}) = \pi^ {- 1} \left(\hat {\boldsymbol {p}} _ {2}\right) ^ {\intercal} \boldsymbol {E} \pi^ {- 1} \left(\hat {\boldsymbol {p}} _ {1}\right) = 0. \tag {23} +$$ + +In practice, most camera models only define the forward projection $\pi$ , and unprojection $\pi^{-1}$ is solved for iteratively. For example, (19) does not have a closed-form inverse. This makes it non-trivial to directly compute its Jacobians required for the Sampson approximation. To avoid this computation we rely on the inverse function theorem. The variant that we will use can be found in the book [6]. It is tailored to functions that map higher-dimensional inputs to lower-dimensional outputs. We rephrase it here for reader's convenience, adapting the notation to our needs and losing some generality: + +Theorem 1. [6, Theorem 1E.6] Let $\pi : \mathbb{R}^3 \to \mathbb{R}^2$ be continuously differentiable in a neighborhood around $d \in \mathbb{R}^3$ . Further, suppose that its Jacobian $\pmb{J}$ at $d$ has full rank. Then, there exists a neighborhood $V$ of $\bar{\pmb{p}} = \pi(d)$ and a continuously differentiable function $\widetilde{\pi^{-1}}: V \to \mathbb{R}^3$ such that $\pi(\widetilde{\pi^{-1}}(p)) = p \forall p \in V$ and whose Jacobian is given by $\pmb{J}^\dagger$ , the Moore-Penrose pseudoinverse of $\pmb{J}$ . + +The construction of $\widetilde{\pi^{-1}}$ in the proof from [6] also gives us that $\widetilde{\pi^{-1}} (\bar{p}) = d$ . Note that $\pi$ admits many right-inverses, but they only differ in lengths assigned to unprojected bearing vectors. Thus, $\widetilde{\pi^{-1}}$ that the theorem gives us need not coincide with $\pi^{-1}$ that maps points to unit vectors. However, the following result shows that when we evaluate $J$ at a unit vector $d = \pi^{-1}(\bar{p})$ , $J^{\dagger}$ actually gives us the Jacobian of $\pi^{-1}$ : + +Corollary 1. Let $\pmb{d} \in S^2$ and $\pmb{J} = \frac{\partial\pi}{\partial\bar{\pmb{d}}} \in \mathbb{R}^{2\times 3}$ . Then the Jacobian of $\pi^{-1}$ at $\bar{\pmb{p}} = \pi (\pmb{d})$ is given by $\pmb{J}^{\dagger}$ . + +Proof. Since the only difference between $\widetilde{\pi^{-1}}$ from theorem 1 and $\pi^{-1}$ is the norm of the output, we can write + +$$ +\pi^ {- 1} (\boldsymbol {p}) = h \left(\widetilde {\left(\pi^ {- 1} (\boldsymbol {p})\right)} \quad \forall \boldsymbol {p} \in V, \right. \tag {24} +$$ + +where $h: \pmb{v} \mapsto \pmb{v} / \| \pmb{v} \|$ is the normalization function. We also have $\pmb{d} = \widetilde{\pi^{-1}}(\bar{\pmb{p}})$ . Thus, the Jacobian of $\pi^{-1}$ at $\bar{\pmb{p}}$ is given by + +$$ +\frac {\partial \pi^ {- 1}}{\partial \boldsymbol {p}} = \left. \frac {\partial h}{\partial \boldsymbol {v}} \right| _ {\boldsymbol {v} = \boldsymbol {d}} \frac {\partial \widetilde {\pi^ {- 1}}}{\partial \boldsymbol {p}} = (\boldsymbol {I} - \boldsymbol {d} \boldsymbol {d} ^ {\intercal}) \boldsymbol {J} ^ {\dagger} = \boldsymbol {J} ^ {\dagger}. \tag {25} +$$ + +![](images/d401e52b9febab44d26971f331f4aed318ccecd2838658691e008ee18dee3991.jpg) + +In the last equation we used $\pmb{d}^{\mathsf{T}}\pmb{J}^{\dagger} = \mathbf{0}$ , which is true in general for any null-vector $\pmb{d}$ of the matrix $J$ , but in our case can also be seen from the following lemma that gives a closed-form expression for $\pmb{J}^{\dagger}$ : + +Lemma 1. Let $\pmb{d} \in \mathbb{R}^3$ and $J = \frac{\partial\pi}{\partial\pmb{d}} \in \mathbb{R}^{2\times 3}$ . Then, $J^{\dagger}$ is given by + +$$ +\boldsymbol {M} = \frac {1}{\boldsymbol {d} \cdot \left(\boldsymbol {g} _ {x} \times \boldsymbol {g} _ {y}\right)} \left[ \left(\boldsymbol {g} _ {y} \times \boldsymbol {d}\right), \quad \left(\boldsymbol {d} \times \boldsymbol {g} _ {x}\right) \right], \tag {26} +$$ + +where $\pmb{g}_x$ , $\pmb{g}_y \in \mathbb{R}^3$ are the two rows of the Jacobian $\pmb{J}$ . + +Proof. Simple calculations show that $JM = I_{2 \times 2}$ . It remains to show that $MJ \in \mathbb{R}^{3 \times 3}$ is symmetric. For ease of notation, introduce $\alpha = d \cdot (g_x \times g_y)$ . Then we have + +$$ +\begin{array}{l} \alpha \boldsymbol {M} \boldsymbol {J} = (\boldsymbol {g} _ {y} \times \boldsymbol {d}) \boldsymbol {g} _ {x} ^ {\intercal} + (\boldsymbol {d} \times \boldsymbol {g} _ {x}) \boldsymbol {g} _ {y} ^ {\intercal} = (27) \\ = [ \boldsymbol {d} ] _ {\times} \left(- \boldsymbol {g} _ {y} \boldsymbol {g} _ {x} ^ {\intercal} + \boldsymbol {g} _ {x} \boldsymbol {g} _ {y} ^ {\intercal}\right) = [ \boldsymbol {d} ] _ {\times} [ \boldsymbol {g} _ {y} \times \boldsymbol {g} _ {x} ] _ {\times}. (28) \\ \end{array} +$$ + +Since the projection function $\pi$ does not depend on the norm of $\pmb{d}$ , i.e. $\pi(\pmb{d}) = \pi(\lambda \pmb{d}) \forall \lambda > 0$ , we have that $\pmb{J} \pmb{d} = \mathbf{0}$ . Since $\pmb{d}$ is then orthogonal to both $\pmb{g}_x$ and $\pmb{g}_y$ we have $\pmb{g}_y \times \pmb{g}_x = \gamma \pmb{d}$ for some $\gamma \in \mathbb{R}$ . Inserting into (28) we get + +$$ +\boldsymbol {M} \boldsymbol {J} = \frac {\gamma}{\alpha} [ \boldsymbol {d} ] _ {\times} ^ {2}, \tag {29} +$$ + +which is a symmetric matrix. + +![](images/9fd1beca797ef476155a5745b8b3b1554b9fe4f9bd2437aff48b003af4503df7.jpg) + +The expression (26) can be used to efficiently compute the Jacobian $J_{C}$ of the constraint function $C$ from (23), and finally to arrive at the Sampson approximation of the reprojection error in any differentiable camera model: + +$$ +\mathcal {E} _ {T S} ^ {2} = \frac {C (\boldsymbol {z}) ^ {2}}{\| \boldsymbol {J} _ {C} \| ^ {2}} = \frac {\left(\boldsymbol {d} _ {2} ^ {\intercal} \boldsymbol {E} \boldsymbol {d} _ {1}\right) ^ {2}}{\| \boldsymbol {d} _ {2} ^ {\intercal} \boldsymbol {E} \boldsymbol {J} _ {1} ^ {\dagger} \| ^ {2} + \| \boldsymbol {d} _ {1} ^ {\intercal} \boldsymbol {E} ^ {\intercal} \boldsymbol {J} _ {2} ^ {\dagger} \| ^ {2}}, \tag {30} +$$ + +where $\pmb{d}_i = \pi^{-1}(\pmb{p}_i)$ and the two Jacobians are evaluated at $\pmb{d}_1$ and $\pmb{d}_2$ respectively. In the rest of the paper we refer to this as the Tangent Sampson error. The name reflects the fact that after we linearize the unprojection $\pi^{-1}$ , we are essentially optimizing in the tangent plane to $S^2$ at the bearing vectors of the undistorted points. Note that the formula is analogous to the original Sampson error (18) (replacing $\pmb{d}_1$ and $\pmb{d}_2$ with the image points, and replacing $J_1^\dagger$ and $J_2^\dagger$ with $[10; 01; 00]$ ). Further, since $J_1$ and $J_2$ are evaluated at the original points of the correspondence, they are independent of $\pmb{E}$ and can be precomputed. Thus, during RANSAC and refinement the computational cost stays approximately the same as for the classical Sampson error. + +# 3.1. Backward Camera Models + +While most camera models used in practice are forward-models, i.e. mapping from the world to the image, there also exist backward-models which instead take the image points as input. The division model (8) is perhaps the most popular example here. The derivations above can easily be adapted to allow for backward-models. The pseudoinverses $J_{k}^{\dagger}$ are simply replaced with the $3 \times 2$ Jacobians from the undistortion map, + +$$ +\mathcal {E} _ {T S} ^ {2} = \frac {\left(\boldsymbol {d} _ {2} ^ {\intercal} \boldsymbol {E} \boldsymbol {d} _ {1}\right) ^ {2}}{\left\| \boldsymbol {d} _ {2} ^ {\intercal} \boldsymbol {E} \boldsymbol {J} _ {1} \right\| ^ {2} + \left\| \boldsymbol {d} _ {1} ^ {\intercal} \boldsymbol {E} ^ {\intercal} \boldsymbol {J} _ {2} \right\| ^ {2}}. \tag {31} +$$ + +In particular, for the division model (8) we would have + +$$ +\boldsymbol {J} = \frac {\partial h}{\partial \boldsymbol {d}} \frac {\partial \boldsymbol {d}}{\partial \boldsymbol {x}} = \frac {\partial h}{\partial \boldsymbol {d}} \left[ I _ {2 \times 2}; 2 \lambda \boldsymbol {x} ^ {\intercal} \right] \in \mathbb {R} ^ {3 \times 2}, \tag {32} +$$ + +where we need to explicitly include the Jacobian $\frac{\partial h}{\partial d}$ of the normalization function $h$ , since the unprojected $\pmb{d}$ from (8) is not a unit vector. + +# 3.2. Measurement Uncertainty + +Following the approach from [3], we can incorporate measurement uncertainty in the keypoint positions into the Tangent Sampson error. We will assume that the keypoint positions are affected by independent Gaussian noise + +$$ +\boldsymbol {p} _ {1} \sim \mathcal {N} \left(\bar {\boldsymbol {p}} _ {1}, \boldsymbol {\Sigma} _ {1}\right), \quad \boldsymbol {p} _ {2} \sim \mathcal {N} \left(\bar {\boldsymbol {p}} _ {2}, \boldsymbol {\Sigma} _ {2}\right) \tag {33} +$$ + +with known covariances. For observations $\pmb{p}_1$ and $\pmb{p}_2$ , the maximum-likelihood error for an essential matrix $E$ is + +$$ +\begin{array}{l} \mathcal {E} _ {M L} ^ {2} (\boldsymbol {z}, \boldsymbol {E}) = \min _ {\hat {z}} \quad \frac {1}{2} (\boldsymbol {z} - \hat {\boldsymbol {z}}) ^ {\intercal} \boldsymbol {\Sigma} ^ {- 1} (\boldsymbol {z} - \hat {\boldsymbol {z}}) \tag {34} \\ \begin{array}{c c} \text {s . t .} & C (\hat {\boldsymbol {z}}) = 0, \end{array} \\ \end{array} +$$ + +where $C$ is the constraint function (23) and $\pmb{\Sigma}$ is the joint covariance of $\pmb{z} = [\pmb{p}_1; \pmb{p}_2]$ , i.e. $\pmb{\Sigma} = \text{diag}(\pmb{\Sigma}_1, \pmb{\Sigma}_2)$ . + +The procedure from Section 2.1 yields in this case + +$$ +\mathcal {E} _ {T S} ^ {2} = \frac {C (\boldsymbol {z}) ^ {2}}{\boldsymbol {J} _ {C} \boldsymbol {\Sigma} \boldsymbol {J} _ {C} ^ {\intercal}} = \frac {\left(\boldsymbol {d} _ {2} ^ {\intercal} \boldsymbol {E} \boldsymbol {d} _ {1}\right) ^ {2}}{\| \boldsymbol {d} _ {2} ^ {\intercal} \boldsymbol {E} \boldsymbol {J} _ {1} ^ {\dagger} \| _ {\Sigma_ {1}} ^ {2} + \| \boldsymbol {d} _ {1} ^ {\intercal} \boldsymbol {E} ^ {\intercal} \boldsymbol {J} _ {2} ^ {\dagger} \| _ {\Sigma_ {2}} ^ {2}}, \tag {35} +$$ + +where $\| \pmb{v} \|_{\Sigma}^{2} = \pmb{v}^{T} \pmb{\Sigma} \pmb{v}$ . See the supplementary material for details on the derivation. This is a generalization of the last expression in [3, Section 3.4] to the arbitrary camera model case. We evaluate how it compares to the original Tangent Sampson error (30) in a synthetic setting in Section 4.4.1. + +# 4. Experimental Evaluation + +In the rest of the paper we experimentally evaluate the Tangent Sampson error. We compare against a selection of popular error metrics, summarized in Table 1. + +# 4.1. Implementation Details + +We implemented all error formulations in $\mathrm{C + + }$ using Eigen [9] as a linear algebra library and Ceres [1] for optimization. Since evaluation is run on Intel hardware, we additionally speed up the errors that do not require iterative methods or explicit camera projection $(\mathcal{E}_{TS},\mathcal{E}_S,\mathcal{E}_{SED},$ $\mathcal{E}_{CS}$ ,and $\mathcal{E}_{ALG})$ by vectorizing their implementations with AVX and FMA instruction sets. In Table 1 we also list the approximate computational cost of evaluating the residual. Tangent Sampson error is significantly faster compared to other metrics which also measure the error in the original image, and has comparable, but slightly higher, runtime to the purely geometric and pinhole-based metrics. For the computational cost we include parts which cannot be precomputed (i.e. parts that depend on the essential matrix). For a comparison with the non-optimized errors (using pure Eigen-implementations), we refer the reader to the supplementary material, where we list the runtime for all formulations without explicit vectorization and some setup details. + +
NameRuntime
Rel.Abs.
\( \mathcal{E}_{ALG} \)(7)Algebraic error↓1.01.3 ns
\( \mathcal{E}_{CS} \)(6)Cosine error↓1.82.4 ns
\( \mathcal{E}_{ML} \)(3)(Pinhole) Repoj. error\( \blacktriangleright \)28.036.7 ns
\( \mathcal{E}_{SED} \)(4)(Pinhole) Sym. Epi. Dist.\( \blacktriangleright \)1.72.3 ns
\( \mathcal{E}_{S} \)(5)(Pinhole) Sampson error\( \blacktriangleright \)1.41.9 ns
\( \mathcal{E}_{\pi ML} \)(22)Reproj. error\( \square \)8861160 ns
\( \mathcal{E}_{\pi SED} \)(10)Proj. Sym. Epi. Dist.\( \square \)92121 ns
\( \mathcal{E}_{TS} \)(30)Tangent Sampson\( \square \)2.22.8 ns
+ +Table 1. Error metrics compared in the experimental evaluation. The errors are grouped by the domain where the error is computed: geometric errors $(\mathcal{L})$ , the undistorted image $(\mathfrak{X})$ and the original image $(\square)$ . Computational cost is shown relative to the algebraic error which is the cheapest. + +# 4.2. Datasets + +For the evaluation we consider two datasets. First the meta-calibration dataset from BabelCalib [21]. The dataset is a collection of checkerboard images from a wide variety of cameras, including some with extremely wide field-of-view $(>270^{\circ})$ . For the evaluation we split the dataset into three groups depending on their field-of-view; $< 90^{\circ}$ , $90 - 150^{\circ}$ , and $>150^{\circ}$ . As the original dataset is biased towards higher field-of-view cameras, we also include some additional checkerboard images taken with a standard DSLR (circa $50 - 70^{\circ}$ ) to even out the distribution. These checkerboard datasets allow us to evaluate on a large number of different camera and lens setups. In Section 4.3 we show that the Tangent Sampson error consistently approximates the true reprojection error, regardless of the camera used, and in Section 4.4 we show that when used for local refinement it yields the most accurate epipolar geometries. + +To evaluate the performance of the error metrics in the context of RANSAC we consider image pairs extracted from Structure-from-Motion datasets, where the poses from the reconstruction are taken as a psuedo-ground truth. As the experiments in Section 4.3-4.4 show that most error metrics behave similarly for low to medium field-of-view cameras $(< 90^{\circ})$ , we focus our evaluation on cameras with larger field-of-view. For this we use two datasets (Grossmunster and Kirchenge) from Larsson et al. [17], which contain 373 and 369 images taken with a fisheye-lens. + +# 4.3. Comparison with True Reprojection Error + +We first evaluate how different error metrics compare with the true maximum-likelihood estimate, $\mathcal{E}_{ML}$ (3) for pinhole cameras and $\mathcal{E}_{\pi ML}$ (22) for non-pinhole cameras. To compute the ML-estimate we perform two-view triangulation followed by non-linear optimization of the 3D point, minimizing the squared reprojection error (using the known camera intrinsics). Note that this approach is orders of magnitude slower compared to the metrics we evaluate (see Ta + +ble 1) and cannot reasonably be used in practical scenarios. + +For the experiment we consider the collection of checkerboard images from [21]. We use the calibration toolbox from BabelCalib [21] to estimate the extrinsic and intrinsic parameters for each image. For each pair we collect the 2D-2D correspondences from corresponding corners. As the corner detector [10] is extremely accurate on these images (some having around $0.2\mathrm{px}$ error during calibration), we add Gaussian noise with $1\mathrm{px}$ standard deviation, to better differentiate the methods (as all methods give zero error for exact data). + +Since it is not meaningful to directly compare the residual values, we instead compare their ordering. For this we make use of Kendall's $\tau$ rank correlation coefficient [15]. It is defined for two rankings $(p_1,\dots ,p_n)$ and $(q_{1},\ldots ,q_{n})$ as + +$$ +\tau (p, q) = \frac {\left| \{\text {c o n c o r d a n t p a i r s} \} \right| - \left| \{\text {d i s c o r d a n t p a i r s} \} \right|}{\binom {n} {2}}. \tag {36} +$$ + +A pair of indices $(i,j)\in \{1,\ldots ,n\} ^2$ where $i > j$ is called concordant if and only if + +$$ +\left(p _ {i} > p _ {j} \wedge q _ {i} > q _ {j}\right) \vee \left(p _ {i} < p _ {j} \wedge q _ {i} < q _ {j}\right). \tag {37} +$$ + +Otherwise, the pair is discordant. Note that due to the normalization we have $\tau(p,q) \in [-1,1]$ , where identical rankings yield $\tau = 1$ , and reversed $\tau = -1$ . + +The results are shown in Fig. 3, split into three groups based on the camera field-of-view. For low to medium field-of-view cameras $(< 90^{\circ})$ we can see that all metrics except the algebraic residual behave similar to the true reprojection error. As the field-of-view increases, there is a more clear difference between the methods, with the pinhole-based errors performing the worst. Note however, that for all cameras the proposed Tangent Sampson error offers the best approximation of the true reprojection error. + +In addition to Kendall-Tau distance, we compare the errors qualitatively. For this, we show the values of an error for a fixed correspondence on a pair of images from a panoramic and fisheye camera in Fig. 4. There, we keep the point on the fisheye image fixed, and vary the point on the panoramic image to show different error values coded by different colors. One can see that $\mathcal{E}_{TS}$ error models $\mathcal{E}_{\pi ML}$ closely in the region where the latter is well-defined. $\mathcal{E}_{TS}$ is also one of the two errors that do not degenerate around the epipole, and, at the same time, it correctly models the spread of values of $\mathcal{E}_{\pi ML}$ around the epipolar curve. Note that the shape of true reprojection error varies significantly depending on how close the correspondence is to the epipole. In the supplementary material we show more qualitative results. In the supplementary material we also show a comparison with [16], which proposed an optimal method for computing the $\pi ML$ error (22) for the division model (8). + +![](images/9f8f9863b3dffe9841e2d7a3e51d10e5305c5ca8f233618b3decbd19021312a8.jpg) + +![](images/43382c54d7df2953af5c5c9ea9a3c5e819ce256ec61fc26f7437a18c94e6eabf.jpg) + +![](images/d3cc5e8188087f251cb695991e8e1d4294fe8cb4cde52620bb4b989d2dd3358c.jpg) + +![](images/4afa0f4d629a848b416988c4ff23f50895fcd5deb1029428b8d7f4760eef655c.jpg) +Figure 3. The figure shows the Kendall $\tau$ coefficient comparing different error metrics for the checkerboard data. For all cameras the proposed Tangent Sampson error correlates well with the true reprojection error $\pi ML$ . + +![](images/cba1d2bf71e6381b62eab4f5cb6c7e32dfab92369c76378c13175ba5acd16d06.jpg) +Figure 4. Qualitative comparison of different error functions. The thresholds for colormaps were selected to be the error at the point marked by the blue plus. The shape of the true reprojection error varies heavily depending on how close the correspondence is to the epipoles. In the supplementary material we show more examples. We also show $\mathcal{E}_{\pi SED}$ and errors computed on the undistorted image in the supplementary. The geometric error is only shown in the region where the 3D point after midpoint triangulation does not correspond to negative depths of bearing vectors (areas where this does not hold for one or both depths are highlighted). + +![](images/ea6147ceb46990171650d7039770608f9b014e2d790c4f61e6ba52ca5842ae4d.jpg) + +![](images/06a0fdee873acd6009111bca766195b0a5328473d4045423eb20eb98ead0b4d9.jpg) + +# 4.4. Accuracy in Pose Refinement + +Next we evaluate the error metrics on how they perform when used for pose-refinement. For the experiment we again consider the image pairs extracted from the checkerboard images. For each image pair we perform local optimization minimizing the proposed error metrics over all correspondences. To initialize the optimization we take the ground truth relative pose and randomly perturb both the rotation and translation by 1 degree. Additionally, point positions were perturbed with Gaussian noise with unit standard deviation along each dimension. For the two ML metrics we triangulate 3D points which are optimized jointly with the camera pose. We experience that this sometimes has problems with worse convergence. The results are shown in Table 2. Again we see that for low to medium field-of-view cameras, the difference between the error metrics is small. The Tangent Sampson error consistently provides the most accurate camera poses after refinement. + +Convergence Properties. The previous experiment evaluated the accuracy of the camera poses obtained when initialized close to the correct pose. We also experimented with varying the initial pose and measuring how often the optimization converged to the correct pose. Figure 5 shows the percentage of poses (across all pairs) that are within 5 degrees average rotation/translation error for varying errors in the initial pose. For large initial errors $(>25^{\circ})$ , the proposed Tangent-Sampson error performs slightly better com + +![](images/497cdf7c45381b2230692d8d70a7f185c1b300470d2dd2823f7a9d2d74ec0736.jpg) +Figure 5. Convergence properties. Successful trials (average rot. and trans. error $< 5^{\circ}$ ) for varying initial pose errors (in degrees). + +pared to the other metrics. + +# 4.4.1 Effect of Measurement Uncertainty + +Experiments in [3] already show that taking point covariances into account in the cost function provides benefits for fundamental matrix estimation. Here, we conduct another synthetic test to show that this effect persists when estimating the essential matrix with highly nonlinear cameras. We generate 2000 instances with pairs of spherical cameras $(1000 \times 2000\mathrm{px})$ with a random relative pose. For each pair we project 100 random 3D-points uniformly drawn from the cube $[-10, 10]^3$ , and add Gaussian noise with covariance $\Lambda$ . + +Covariances are generated as $\Lambda = \alpha Q\mathrm{diag}(\beta ,1 - \beta)Q^{\intercal}$ + +
< 90° (5332 pairs)90 - 150° (21927 pairs)> 150° (11860 pairs)
RotationTranslationRotationTranslationRotationTranslation
\( \mathcal{E}_{ALG} \)0.379 / 0.3141.137 / 0.3281.528 / 0.3783.938 / 0.5881.582 / 0.5063.489 / 0.655
\( \mathcal{E}_S \)0.376 / 0.3121.056 / 0.3271.464 / 0.3473.520 / 0.5751.634 / 0.5673.812 / 0.778
\( \mathcal{E}_{SED} \)0.376 / 0.3121.065 / 0.3271.475 / 0.3523.554 / 0.5782.596 / 0.9865.667 / 1.342
\( \mathcal{E}_{COS} \)0.376 / 0.3131.066 / 0.3261.454 / 0.3413.459 / 0.5581.447 / 0.4543.235 / 0.612
\( \mathcal{E}_{ML} \)0.376 / 0.3121.103 / 0.3271.472 / 0.3583.733 / 0.5751.630 / 0.5633.771 / 0.766
\( \mathcal{E}_{\pi ML} \)0.375 / 0.3121.075 / 0.3271.451 / 0.3503.778 / 0.5571.418 / 0.4463.143 / 0.597
\( \mathcal{E}_{\pi SED} \)0.376 / 0.3131.074 / 0.3271.458 / 0.3433.490 / 0.5591.473 / 0.4503.249 / 0.607
\( \mathcal{E}_{TS} \)0.376 / 0.3121.057 / 0.3271.443 / 0.3403.432 / 0.5571.465 / 0.4453.226 / 0.598
+ +![](images/d2a99a1fc063882dcc1eeac8ac24b68cbb4d615cb714de51983e69e9f10c5048.jpg) +Figure 6. Measurement Uncertainty. The distribution of pose errors (maximum of rotation and translation error) when using covariance-aware Tangent Sampson error for pose refinement versus non-covariance-aware in synthetic experiments. Boxes show quartiles of the respective distribution of errors and whiskers extend to the minimal and maximal value. + +where $Q \sim U(\mathrm{SO}(2))$ is a random rotation on the plane, $\beta \sim U([0, 0.5])$ is a random eccentricity parameter, and $\alpha$ is the overall scale of the noise. To emulate keypoint detection on a multi-scale image pyramid, all generated points were split into 4 equal sets, with $\alpha$ within each set assigned to 1, 2, 4, and 8, respectively. Points and covariances are then passed to a Levenberg-Marquardt optimization of the relative pose using (35) as the residual function, with true pose serving as the initial estimate. Figure 6 demonstrates how the covariance-aware formulation (35) compares to the baseline (30). Covariance information is useful in the case when points are affected by noise of unequal scales. + +# 4.5. Evaluation in Robust Estimators + +Finally we evaluate the methods in the context of RANSAC [7]. We implement a standard LO-RANSAC [18] estimator which uses the various error metrics both for model scoring (using MSAC scoring [24]) and refinement (using non-linear optimization on the inlier-set). For the experiment we consider image pairs from two Structure-from-Motion reconstructions (see Section 4.2). From each dataset we extract all image pairs which contain between 50 and 200 correct matches (according to the Structure-from-Motion ground-truth). For each pair we estimate the rela + +Table 2. Evaluation of camera pose refinement. Table shows the mean and median errors (in degrees) of the rotation and translation after refinement. The data is split into three groups based on the camera field-of-view. Best error is highlighted in bold. Second-best is underlined. For low-distortion cameras the choice of error makes almost no difference on the quality of pose. + +
GrossmunsterKirchenge
εR,t< 5°RTεR,t< 5°RT
ΕALG0.76175.039.60.31884.821.6
ΕCOS0.56181.550.40.25491.625.5
ΕSED1.16168.982.00.35586.444.0
ΕS0.92474.551.00.30089.225.3
ΕπSED0.56181.3211.20.25591.687.4
ΕTS0.55082.246.40.24891.924.2
+ +Table 3. Evaluation in robust estimators. Table shows the median pose error (maximum of rotation and translation error), percentage of pairs for which the estimation was successful (pose error less than 5 degrees), and runtime (in ms). Best value in the column is highlighted in bold. Second-best is underlined. + +tive pose using RANSAC and compute the error in the pose, as well as percentage of successful runs. For each method we tuned the inlier threshold on a separately extracted validation set. For more details on the experiment setup see the supplementary material. The results are presented in Table 3. Using Tangent Sampson error consistently gives slightly better performance compared to other error metrics. + +# 5. Conclusions + +In this paper we have presented a simple generalization to the classical Sampson error which allows for using arbitrary camera models. In experiments on a wide variety of camera and lens configurations, we have shown that the proposed error metric outperforms the alternative metrics that are used in practice. While the improvements are sometimes minor (e.g. for low-distortion pinhole cameras), they are consistent and come at essentially no additional cost. + +Acknowledgments: This project received funding from the strategic research program ELLIIT. MT was supported by the ESOP scholarship from the ETH Foundation. + +# References + +[1] Sameer Agarwal, Keir Mierle, and Others. Ceres solver. http://ceres-solver.org.5 +[2] João Pedro Barreto and Kostas Daniilidis. Fundamental matrix for cameras with radial distortion. In International Conference on Computer Vision (ICCV), 2005. 3 +[3] Wojciech Chojnacki, Michael J. Brooks, Anton Van Den Hengel, and Darren Gawley. On the fitting of surfaces to data with covariances. IEEE Trans. Pattern Analysis and Machine Intelligence (PAMI), 2000. 3, 5, 7 +[4] Ondřej Chum, Tomáš Pajdla, and Peter Sturm. The geometric error for homographies. Computer Vision and Image Understanding (CVIU), 2005. 3 +[5] David Claus and Andrew W Fitzgibbon. A rational function lens distortion model for general cameras. In Computer Vision and Pattern Recognition (CVPR), 2005. 3 +[6] Asen L Dontchev and R Tyrrell Rockafellar. Implicit functions and solution mappings, volume 543. Springer, 2009. 4 +[7] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 8 +[8] Andrew W Fitzgibbon. Simultaneous linear estimation of multiple view geometry and lens distortion. In Computer Vision and Pattern Recognition (CVPR), 2001. 3 +[9] Gaël Guennebaud, Benoit Jacob, et al. Eigen v3. http://eigen.tuxfamily.org, 2010. 5 +[10] Hyowon Ha, Michal Perdoch, Hatem Alismail, In So Kweon, and Yaser Sheikh. Deltille grids for geometric camera calibration. In International Conference on Computer Vision (ICCV), 2017. 6 +[11] Matthew Harker and Paul O'Leary. First order geometric distance (the myth of sampsonus). In *British Machine Vision Conference (BMVC)*, 2006. 3 +[12] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1 +[13] Richard I Hartley and Peter Sturm. Triangulation. Computer vision and image understanding, 68(2):146-157, 1997. 2, 3 +[14] Juho Kannala and Sami S Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. 2006. 4 +[15] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1/2):81-93, 1938. 6 +[16] Zuzana Kukelova and Viktor Larsson. Radial distortion triangulation. In Computer Vision and Pattern Recognition (CVPR), 2019. 3, 6 +[17] Viktor Larsson, Nicolas Zobernig, Kasim Taskin, and Marc Pollefeys. Calibration-free structure-from-motion with calibrated radial trifocal tensors. In European Conference on Computer Vision (ECCV), 2020. 6 +[18] Karel Lebeda, Jiri Matas, and Ondrej Chum. Fixing the locally optimized ransac-full experimental evaluation. In British Machine Vision Conference (BMVC), 2012. 8 + +[19] Seong Hun Lee and Javier Civera. Geometric interpretations of the normalized epipolar error. arXiv preprint arXiv:2008.01254, 2020. 2 +[20] Peter Lindstrom. Triangulation made easy. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 1554-1561. IEEE, 2010. 2, 3 +[21] Yaroslava Lochman, Kostiantyn Liepieshov, Jianhui Chen, Michal Perdoch, Christopher Zach, and James Pritts. Babel-calib: A universal approach to calibrating central cameras. In International Conference on Computer Vision (ICCV), 2021. 6 +[22] Quan-Tuan Luong and Olivier D Faugeras. The fundamental matrix: Theory, algorithms, and stability analysis. International Journal of Computer Vision (IJCV), 1996. 2, 3 +[23] Paul D Sampson. Fitting conic sections to "very scattered" data: An iterative refinement of the bookstein algorithm. Computer graphics and image processing, 1982. 2, 3 +[24] Philip HS Torr and Andrew Zisserman. Mlesac: A new robust estimator with application to estimating image geometry. Computer Vision and Image Understanding (CVIU), 2000, 8 +[25] Zhengyou Zhang. On the epipolar geometry between two images with lens distortion. In International Conference on Pattern Recognition (ICPR), 1996. 3 \ No newline at end of file diff --git a/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/images.zip b/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4e545c83a553ca45550a4540dd550d6853ded146 --- /dev/null +++ b/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:405eadfd3990612c1205ef7fcb6ceee68a583921fd96c4c4689fd28b59e78365 +size 670378 diff --git a/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/layout.json b/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b012b60f67b2dac0782c86d561558523afb8b416 --- /dev/null +++ b/2023/Tangent Sampson Error_ Fast Approximate Two-view Reprojection Error for Central Camera Models/layout.json @@ -0,0 +1,10659 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 104, + 533, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 104, + 533, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 104, + 533, + 140 + ], + "type": "text", + "content": "Tangent Sampson Error: Fast Approximate Two-view Reprojection Error for Central Camera Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 161, + 249, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 161, + 249, + 188 + ], + "spans": [ + { + "bbox": [ + 159, + 161, + 249, + 188 + ], + "type": "text", + "content": "Mikhail Terekhov ETH Zurich" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 192, + 251, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 192, + 251, + 201 + ], + "spans": [ + { + "bbox": [ + 157, + 192, + 251, + 201 + ], + "type": "text", + "content": "mterekhov@ethz.ch" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 323, + 161, + 401, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 161, + 401, + 174 + ], + "spans": [ + { + "bbox": [ + 323, + 161, + 401, + 174 + ], + "type": "text", + "content": "Viktor Larsson" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 323, + 176, + 403, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 176, + 403, + 190 + ], + "spans": [ + { + "bbox": [ + 323, + 176, + 403, + 190 + ], + "type": "text", + "content": "Lund University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 290, + 192, + 433, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 192, + 433, + 202 + ], + "spans": [ + { + "bbox": [ + 290, + 192, + 433, + 202 + ], + "type": "text", + "content": "viktor.larsson@math.lth.se" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 82, + 234, + 366, + 377 + ], + "blocks": [ + { + "bbox": [ + 82, + 234, + 366, + 377 + ], + "lines": [ + { + "bbox": [ + 82, + 234, + 366, + 377 + ], + "spans": [ + { + "bbox": [ + 82, + 234, + 366, + 377 + ], + "type": "image", + "image_path": "fafdb5c2548aa683e1fc4b744598dc60f15672810a3bf881ffb1b04cd5c31455.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 389, + 547, + 434 + ], + "lines": [ + { + "bbox": [ + 46, + 389, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 547, + 434 + ], + "type": "text", + "content": "Figure 1. The figure shows the level sets of the proposed Tangent Sampson error. In contrast to classical Sampson error, the proposed error function applies to arbitrary central camera models; in this case a 360 panoramic image and a fisheye camera. Given a relative pose and a pair of points (highlighted in red), we vary a point on one image while keeping the other fixed, and thus get the values on each image. A given color corresponds to a constant value of reprojection error, with purple indicating near-zero errors along the epipolar curve." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 369, + 243, + 513, + 377 + ], + "blocks": [ + { + "bbox": [ + 369, + 243, + 513, + 377 + ], + "lines": [ + { + "bbox": [ + 369, + 243, + 513, + 377 + ], + "spans": [ + { + "bbox": [ + 369, + 243, + 513, + 377 + ], + "type": "image", + "image_path": "2bc68f527d4fd5a4e586c6eff8e4d9739f7e44827fa75707f9ae1a038caf3244.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 444, + 192, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 444, + 192, + 456 + ], + "spans": [ + { + "bbox": [ + 143, + 444, + 192, + 456 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 45, + 469, + 290, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 469, + 290, + 638 + ], + "spans": [ + { + "bbox": [ + 45, + 469, + 290, + 638 + ], + "type": "text", + "content": "In this paper we introduce the Tangent Sampson error, which is a generalization of the classical Sampson error in two-view geometry that allows for arbitrary central camera models. It only requires local gradients of the distortion map at the original correspondences (allowing for pre-computation) resulting in a negligible increase in computational cost when used in RANSAC or local refinement. The error effectively approximates the true-reprojection error for a large variety of cameras, including extremely wide field-of-view lenses that cannot be undistorted to a single pinhole image. We show experimentally that the new error outperforms competing approaches both when used for model scoring in RANSAC and for non-linear refinement of the relative camera pose." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 657, + 128, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 128, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 128, + 670 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Estimating two-view geometry from a set of sparse 2D-2D correspondences is a common sub-problem in many vision applications. Each correspondence " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\pmb{p}_1, \\pmb{p}_2) \\in \\mathbb{R}^2 \\times \\mathbb{R}^2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 445, + 515, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 515, + 457 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 515, + 457 + ], + "type": "text", + "content": "constrains the geometry via the epipolar constraint," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 379, + 463, + 545, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 463, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 379, + 463, + 545, + 476 + ], + "type": "interline_equation", + "content": "\\left(\\boldsymbol {p} _ {2}; 1\\right) ^ {\\top} \\boldsymbol {E} \\left(\\boldsymbol {p} _ {1}; 1\\right) = 0, \\tag {1}", + "image_path": "db573487e6efed3b559e9049dafafd8fa237cd16eaa4cdab51815b451ee3753a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 480, + 546, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 546, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 546, + 600 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 480, + 546, + 600 + ], + "type": "inline_equation", + "content": "\\pmb{E} = [t]_{\\times}\\pmb{R} \\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 304, + 480, + 546, + 600 + ], + "type": "text", + "content": " is the essential matrix and the points " + }, + { + "bbox": [ + 304, + 480, + 546, + 600 + ], + "type": "inline_equation", + "content": "p_1, p_2" + }, + { + "bbox": [ + 304, + 480, + 546, + 600 + ], + "type": "text", + "content": " are on the normalized image plane. This constraint ensures that the 2D-2D correspondence can be exactly triangulated (i.e. with zero reprojection error) using the camera pair extracted from " + }, + { + "bbox": [ + 304, + 480, + 546, + 600 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 304, + 480, + 546, + 600 + ], + "type": "text", + "content": ". However when the measurement " + }, + { + "bbox": [ + 304, + 480, + 546, + 600 + ], + "type": "inline_equation", + "content": "(p_1, p_2)" + }, + { + "bbox": [ + 304, + 480, + 546, + 600 + ], + "type": "text", + "content": " is inexact, it is natural to measure how close this correspondence is to being consistent with a given epipolar geometry. This is for example used for determining inlier matches during robust estimation, or as a loss to be minimized when refining an initial estimate of " + }, + { + "bbox": [ + 304, + 480, + 546, + 600 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 304, + 480, + 546, + 600 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 601, + 547, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 601, + 547, + 649 + ], + "spans": [ + { + "bbox": [ + 305, + 601, + 547, + 649 + ], + "type": "text", + "content": "Under the assumption of i.i.d. Gaussian noise, the ML-estimate of the closest consistent correspondence, sometimes called the gold-standard estimate in the literature [12], is found by minimizing the squared reprojection error," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 310, + 665, + 545, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 665, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 310, + 665, + 545, + 684 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {M L} ^ {2} = \\min _ {\\boldsymbol {X} \\in \\mathbb {R} ^ {3}} \\| \\boldsymbol {p} _ {1} - \\pi (\\boldsymbol {X}) \\| ^ {2} + \\| \\boldsymbol {p} _ {2} - \\pi (\\boldsymbol {R} \\boldsymbol {X} + t) \\| ^ {2}, \\tag {2}", + "image_path": "11f62da5979f20cfb993608f30163321b35a289e82023373055924def3840aa6.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pi : \\mathbb{R}^3 \\to \\mathbb{R}^2" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " is the camera projection function. For identical pinhole cameras with unit aspect ratio and no" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3370" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 109 + ], + "type": "text", + "content": "skew, this can be reformulated as finding the corrected normalized image points, i.e. the closest pair of 2D points that satisfy the epipolar constraint exactly," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 72, + 117, + 287, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 117, + 287, + 153 + ], + "spans": [ + { + "bbox": [ + 72, + 117, + 287, + 153 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {E} _ {M L} ^ {2} = \\min _ {\\hat {\\boldsymbol {p}} _ {1}, \\hat {\\boldsymbol {p}} _ {2} \\in \\mathbb {R} ^ {2}} \\| \\boldsymbol {p} _ {1} - \\hat {\\boldsymbol {p}} _ {1} \\| ^ {2} + \\| \\boldsymbol {p} _ {2} - \\hat {\\boldsymbol {p}} _ {2} \\| ^ {2} \\tag {3} \\\\ \\begin{array}{l} \\text {s . t .} \\quad (\\hat {\\boldsymbol {p}} _ {2}; 1) ^ {\\intercal} \\boldsymbol {E} (\\hat {\\boldsymbol {p}} _ {1}; 1) = 0. \\end{array} \\\\ \\end{array}", + "image_path": "9ecd2133c2cd246cf1eb978c7840fa6fccea027febf3421842521e7f6ece24bd.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 163, + 287, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 163, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 287, + 235 + ], + "type": "text", + "content": "The closed-form solution to (3) can be found as a root to a degree-six polynomial (see [13]), or via iterative methods [20]. In robust estimators, e.g. RANSAC, it is often necessary to evaluate these residuals thousands (or potentially millions) of times for each image pair, which in practice means that cheaper approximate error metrics are favored." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 236, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 344 + ], + "type": "text", + "content": "These approximate or surrogate error metrics can be divided into two categories: image-based (approximating the true reprojection error) or geometry-based (measuring some geometric quantity, e.g. angles, instead). In the first category we have the epipolar distance, i.e. distance from the point to the corresponding epipolar line. This can be interpreted as assuming zero noise on one of the keypoints. The error is often computed symmetrically and thus called the symmetric epipolar distance, having the form" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 352, + 287, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 352, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 88, + 352, + 287, + 379 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {S E D} ^ {2} = \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| \\boldsymbol {E} _ {1 2} \\boldsymbol {d} _ {1} \\right\\| ^ {2}} + \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| (\\boldsymbol {E} ^ {\\intercal}) _ {1 2} \\boldsymbol {d} _ {2} \\right\\| ^ {2}}, \\tag {4}", + "image_path": "919edb7bcd38b9a6782311a01588dbbc9b72422038d1d7b2aefe7b4d6f2a33a9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "type": "inline_equation", + "content": "E_{12} \\in \\mathbb{R}^{2 \\times 3}" + }, + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "type": "text", + "content": " refers to the first two rows of " + }, + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\pmb{d}_k = (\\pmb{p}_k; 1) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "type": "text", + "content": ". Another image-based error is the Sampson error, which is a first-order approximation of " + }, + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{ML}" + }, + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "type": "text", + "content": " (see Sec. 2.1 for details):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 445, + 287, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 445, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 96, + 445, + 287, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {S} ^ {2} = \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| \\boldsymbol {E} _ {1 2} \\boldsymbol {d} _ {1} \\right\\| ^ {2} + \\left\\| (\\boldsymbol {E} ^ {\\intercal}) _ {1 2} \\boldsymbol {d} _ {2} \\right\\| ^ {2}}. \\tag {5}", + "image_path": "e281ccc6103a111de4f6ad8dad1dccdac5a00573487b40765910bea73b20a28f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 480, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 287, + 517 + ], + "type": "text", + "content": "Both of these errors can after appropriate rescaling (using the focal lengths) be expressed in pixels and thus are natural to interpret in terms of the keypoint detection noise levels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 517, + 279, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 517, + 279, + 529 + ], + "spans": [ + { + "bbox": [ + 59, + 517, + 279, + 529 + ], + "type": "text", + "content": "In the second category, an example is the cosine error," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 538, + 287, + 565 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 538, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 77, + 538, + 287, + 565 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {C S} ^ {2} = \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| \\boldsymbol {d} _ {2} \\right\\| ^ {2} \\left\\| \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right\\| ^ {2}} + \\frac {\\left| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right| ^ {2}}{\\left\\| \\boldsymbol {E} ^ {\\intercal} \\boldsymbol {d} _ {2} \\right\\| ^ {2} \\left\\| \\boldsymbol {d} _ {1} \\right\\| ^ {2}}, \\tag {6}", + "image_path": "c96ffcc4f7d27bbfa3085c1352e01b6fb9cbbd4b51120f75b34e7410d4d63d82.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 574, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 287, + 620 + ], + "type": "text", + "content": "which (symmetrically) measures the cosine between the normal of the epipolar plane " + }, + { + "bbox": [ + 47, + 574, + 287, + 620 + ], + "type": "inline_equation", + "content": "\\pmb{E}\\pmb{d}_1 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 574, + 287, + 620 + ], + "type": "text", + "content": " and corresponding point " + }, + { + "bbox": [ + 47, + 574, + 287, + 620 + ], + "type": "inline_equation", + "content": "d_2 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 574, + 287, + 620 + ], + "type": "text", + "content": ". Another geometric error is the algebraic residual," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 622, + 287, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 622, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 127, + 622, + 287, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {A L G} = \\left| \\boldsymbol {d} _ {2} ^ {\\mathrm {T}} \\boldsymbol {E} \\boldsymbol {d} _ {1} \\right|, \\tag {7}", + "image_path": "8b2b7d7354363bf0d7dcf41bc0f01333a0f0bf53ee617deab76c8cb78b381cf4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 641, + 253, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 253, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 253, + 653 + ], + "type": "text", + "content": "which was given a geometric interpretation in [19]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Most of the literature has focused on the pinhole camera case, and the image-based errors above " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{E}_{SED}" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_S)" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " measure the error in the image plane assuming this model. For non-pinhole cameras, we either have to undistort the image such that " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{SED}" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_S" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " are applicable, or use the" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 425, + 157 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 425, + 157 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 425, + 157 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 425, + 157 + ], + "type": "image", + "image_path": "fb436acb963ff1442f198de14775181020643ebdeb8ace826562f59690f7b384.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 165, + 544, + 188 + ], + "lines": [ + { + "bbox": [ + 305, + 165, + 544, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 165, + 544, + 188 + ], + "type": "text", + "content": "Figure 2. Undistorting the image deforms the level-sets of the reprojection error and introduces a radial bias in the optimization." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 542, + 157 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 542, + 157 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 542, + 157 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 542, + 157 + ], + "type": "image", + "image_path": "63be9b5cc0d75ed8e40990fe8c129ea9506a70f1c529bf9f99521b86e25b1e1e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 194, + 545, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 194, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 545, + 241 + ], + "type": "text", + "content": "geometry-based error metrics (e.g. " + }, + { + "bbox": [ + 305, + 194, + 545, + 241 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{CS}" + }, + { + "bbox": [ + 305, + 194, + 545, + 241 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 194, + 545, + 241 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{ALG}" + }, + { + "bbox": [ + 305, + 194, + 545, + 241 + ], + "type": "text", + "content": "). However, undistorting to a pinhole image (which is not always possible for large field-of-view cameras) will introduce a radial-bias into the error (see Figure 2)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 242, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 242, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 242, + 545, + 289 + ], + "type": "text", + "content": "In this paper we instead propose the Tangent Sampson error " + }, + { + "bbox": [ + 305, + 242, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{TS}" + }, + { + "bbox": [ + 305, + 242, + 545, + 289 + ], + "type": "text", + "content": ", a version of the classical Sampson error that has been generalized to handle arbitrary camera models. Compared to current alternatives it has the following benefits;" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 297, + 545, + 483 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 317, + 297, + 545, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 297, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 317, + 297, + 545, + 345 + ], + "type": "text", + "content": "- It is cheap to evaluate (similar in cost to the original Sampson error) since it only requires computing the camera model forward-projection and Jacobian once (which can be pre-computed for each point)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 349, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 349, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 349, + 545, + 408 + ], + "type": "text", + "content": "- Since it measures errors in the original image, it is easier to set meaningful thresholds in pixels. This is in contrast to geometric error functions, where it is not always clear how to relate the thresholds (e.g. in angles) to the keypoint detection noise in pixels." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 412, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 412, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 317, + 412, + 545, + 483 + ], + "type": "text", + "content": "- It works for arbitrary central camera models, only requiring that we can compute a Jacobian of the camera projection function at each correspondence. This also allows the formulation to work for cameras modeled via look-up tables, where the Jacobian can be estimated via interpolation of finite differences." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "type": "text", + "content": "In Section 2.1 we first recall the derivations for the original Sampson error. Next, in Section 3 we propose our generalization to arbitrary central camera models, and in Section 4 we experimentally show that the proposed error outperforms the competing methods on a variety of camera and lens configurations." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 574, + 391, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 391, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 391, + 586 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 594, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 689 + ], + "type": "text", + "content": "The Sampson error was introduced in [23] to approximate the geometric distance from a point to a conic. The proposed approach was to linearize the equation of the conic around the given point, rendering a closed-form solution for the new distance, now from the point to a line. This idea proved useful for estimating distances to other implicit surfaces, notably those set by constraints of a fundamental matrix [22]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "The fundamental matrix is only applicable to perspective cameras, and a lot of subsequent work in computer vision" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3371" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "focused on generalizing it to non-linear camera models, together with some notion of a reprojection error. A well-known example is the division model from Fitzgibbon [8]. It maps a point " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " on the normalized image plane to the corresponding bearing vector " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\pmb{d}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 121, + 139, + 287, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 139, + 287, + 153 + ], + "spans": [ + { + "bbox": [ + 121, + 139, + 287, + 153 + ], + "type": "interline_equation", + "content": "\\boldsymbol {d} = (\\boldsymbol {p}; 1 + \\lambda \\| \\boldsymbol {p} \\| ^ {2}). \\tag {8}", + "image_path": "40c6e98db70bc95cbcab26dcdccf1eff8aef59cd09149a6fa5cee22699e67085.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 160, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 160, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 160, + 287, + 266 + ], + "type": "text", + "content": "The original paper [8] already discusses a way of approximating the gold-standard error for the new model: first, undistort the original points of the correspondence; next, apply the closed-form triangulation from [13], and distort the projections back onto the original image. The resulting points are then approximate minimizers of (2). A faster iterative scheme for the optimal error was later proposed in [16], which adapts the idea of [20], originally developed for pinhole cameras, to the division model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 268, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 287, + 399 + ], + "type": "text", + "content": "Several camera models with closed-form expressions for epipolar curves were proposed. In [25], epipolar curves are represented by cubics, in the rational model [5] — by conics, and in [2] — by circles. With such models, we can use the Sampson approximation to estimate the distance from one point of the correspondence to the epipolar curve of the other, and vice versa. Note that this gives us an approximation of the symmetric epipolar distance, not the Sampson error, because for the latter we would need to optimize both point positions simultaneously, and this would move the epipolar curves as well." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 400, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 506 + ], + "type": "text", + "content": "All these models were proposed with the goal of searching for the intrinsic parameters together with the epipolar geometry. Indeed, [2] and [5] even provide their own versions of a fundamental matrix on a lifted image space. These matrices can be estimated directly from point correspondences, and then decomposed to provide the distortion parameters and relative pose up to an irreducible ambiguity. We, on the other hand, constrain ourselves to the calibrated case, and thus can afford much greater generality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 508, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 602 + ], + "type": "text", + "content": "There are also various heuristic generalizations of the symmetric epipolar distance to general non-linear camera models that have been used in practice. One approach is to project the bearing vector of one point of the correspondence onto the epipolar plane given by the other. The projected vector is then mapped back onto the image to provide an approximate corrected point, for which the error can be computed. Mathematically, this is expressed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 612, + 286, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 612, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 75, + 612, + 286, + 658 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {n} _ {1} = \\boldsymbol {E} \\boldsymbol {d} _ {1} / \\| \\boldsymbol {E} \\boldsymbol {d} _ {1} \\|, \\quad \\boldsymbol {n} _ {2} = \\boldsymbol {E} ^ {\\top} \\boldsymbol {d} _ {2} / \\| \\boldsymbol {E} ^ {\\top} \\boldsymbol {d} _ {2} \\|, \\tag {9} \\\\ \\mathcal {E} _ {\\pi S E D} ^ {2} = \\left\\| \\boldsymbol {p} _ {1} - \\pi \\left(\\boldsymbol {d} _ {1} - \\boldsymbol {n} _ {2} \\boldsymbol {n} _ {2} ^ {\\intercal} \\boldsymbol {d} _ {1}\\right) \\right\\| ^ {2} \\\\ + \\left\\| \\boldsymbol {p} _ {2} - \\pi \\left(\\boldsymbol {d} _ {2} - \\boldsymbol {n} _ {1} \\boldsymbol {n} _ {1} ^ {\\intercal} \\boldsymbol {d} _ {2}\\right) \\right\\| ^ {2} \\\\ \\end{array}", + "image_path": "e4c35cc0be69c7630362b4cdd737857c8a18ab7f6378b6ffcb29451d4630bd46.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "content": " are the points on the images, " + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "inline_equation", + "content": "d_1" + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "inline_equation", + "content": "d_2" + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "content": " are the corresponding bearing vectors, " + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "inline_equation", + "content": "n_1" + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "inline_equation", + "content": "n_2" + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "content": " are the normals of the epipolar planes. We have encountered this formulation in private codebases, but we are not aware of its" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": "origin. We call it the projective symmetric epipolar error. While this does not give us exact distances to the epipolar curves even in the pinhole case, it often provides a reasonable approximation. The main problem with this approach is that it requires calculating " + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "inline_equation", + "content": "\\pi(p)" + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": " every time, making it rather computationally heavy for some camera models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 144, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 274 + ], + "type": "text", + "content": "From the probabilistic perspective, the geometric error is only justified as long as the noise in the correspondences stays standard Gaussian and i.i.d. A natural generalization is to allow for arbitrary covariances in point positions. For example, when the keypoints are searched on a multi-resolution image pyramid, the standard deviation of the position can be taken as the downscaling factor of the level on which the point was found. For conics and fundamental matrices, covariance-aware error formulations were derived in [3]. We will show how this approach is generalized to our proposed error formulation in Section 3.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 275, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 395 + ], + "type": "text", + "content": "In [11] the authors propose another first-order approximation of the reprojection error. The method builds on the optimal triangulation method from [13] and differentates the sixth-order polynomial that has the optimal geometric error as its first root. The first-order approximation is then given by one step of the Newton-Raphson iteration on this polynomial starting from zero. In the paper, this error is only derived for conics. While it would also work for fundamental matrices, this does not generalize to the nonlinear camera model setting we are interested in." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 396, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 443 + ], + "type": "text", + "content": "Finally, we would like to note that the Sampson error can also be derived for homographies [4]. Our method is applicable there as well, but with larger computational cost, which somewhat defeats its original purpose." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 450, + 466, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 450, + 466, + 463 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 466, + 463 + ], + "type": "text", + "content": "2.1. The Sampson Approximation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "content": "Traditional Sampson error [22, 23] approximates the geometric error, i.e. the distance to the closest pair of points which satisfy the epipolar constraint. Given a correspondence " + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = (\\pmb{p}_1, \\pmb{p}_2) \\in \\mathbb{R}^2 \\times \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "content": " and essential matrix " + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "inline_equation", + "content": "\\pmb{E} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 304, + 468, + 545, + 528 + ], + "type": "text", + "content": ", this can be formalized as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 359, + 536, + 544, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 536, + 544, + 569 + ], + "spans": [ + { + "bbox": [ + 359, + 536, + 544, + 569 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {E} _ {M L} ^ {2} (\\boldsymbol {z}, \\boldsymbol {E}) = \\min _ {\\hat {\\boldsymbol {z}}} \\| \\hat {\\boldsymbol {z}} - \\boldsymbol {z} \\| ^ {2} \\tag {11} \\\\ \\begin{array}{c c} \\text {s . t .} & C (\\hat {z}) = 0, \\end{array} \\\\ \\end{array}", + "image_path": "de36ab1c2790f63080557162c527cd1d8d261d11ba556ff9b23b123c3730dc49.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 578, + 544, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 544, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 544, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 578, + 544, + 602 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{z}} = (\\hat{\\pmb{p}}_1,\\hat{\\pmb{p}}_2)\\in \\mathbb{R}^2\\times \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 578, + 544, + 602 + ], + "type": "text", + "content": " is the corrected correspondence and the epipolar constraint is" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 361, + 611, + 544, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 611, + 544, + 624 + ], + "spans": [ + { + "bbox": [ + 361, + 611, + 544, + 624 + ], + "type": "interline_equation", + "content": "C (\\hat {\\boldsymbol {z}}) = \\left(\\hat {\\boldsymbol {p}} _ {2}; 1\\right) ^ {\\intercal} \\boldsymbol {E} \\left(\\hat {\\boldsymbol {p}} _ {1}; 1\\right) = 0. \\tag {12}", + "image_path": "a0fc6693e610c02f0ad7f44cd2302ddc68e2c61811b31cea225fe3cb612e2a62.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 633, + 544, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 633, + 544, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 633, + 544, + 668 + ], + "type": "text", + "content": "Since (11) does not admit a simple closed form solution, the idea in [23] is to linearize the epipolar constraint, " + }, + { + "bbox": [ + 304, + 633, + 544, + 668 + ], + "type": "inline_equation", + "content": "C(\\hat{z}) = 0" + }, + { + "bbox": [ + 304, + 633, + 544, + 668 + ], + "type": "text", + "content": " at the original correspondence " + }, + { + "bbox": [ + 304, + 633, + 544, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 633, + 544, + 668 + ], + "type": "text", + "content": ", i.e." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 334, + 676, + 544, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 676, + 544, + 694 + ], + "spans": [ + { + "bbox": [ + 334, + 676, + 544, + 694 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {S} ^ {2} (\\boldsymbol {z}, \\boldsymbol {E}) = \\min _ {\\hat {\\boldsymbol {z}}} \\| \\hat {\\boldsymbol {z}} - \\boldsymbol {z} \\| ^ {2} \\tag {13}", + "image_path": "508cab5c39e0ccb513397acfba461900e5adf7f85fde1b9eb8ea72013a681305.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 389, + 698, + 515, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 698, + 515, + 710 + ], + "spans": [ + { + "bbox": [ + 389, + 698, + 515, + 710 + ], + "type": "interline_equation", + "content": "\\text {s . t .} \\quad C (\\boldsymbol {z}) + \\boldsymbol {J} _ {C} (\\hat {\\boldsymbol {z}} - \\boldsymbol {z}) = 0,", + "image_path": "f4c42979616d3a94396fbdbb638ab73abfd7d82599d93cd4b6bee054cca3eae8.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3372" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "J_{C} = \\partial C(z) / \\partial z" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " is the Jacobian of the constraint, evaluated at " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": ". Introducing a Lagrangian for (13)," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 105, + 287, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 105, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 61, + 105, + 287, + 119 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\hat {z}, \\lambda) = \\| \\hat {z} - z \\| ^ {2} + \\lambda (C (z) + J _ {C} (\\hat {z} - z)), \\tag {14}", + "image_path": "6e0f35643045e9f849495bca2a87ac1467c1dd4f0f2bfa8f2aeb13dc17000ddc.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 129, + 191, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 191, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 191, + 140 + ], + "type": "text", + "content": "we get the first-order constraints as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 150, + 287, + 173 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 150, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 69, + 150, + 287, + 173 + ], + "type": "interline_equation", + "content": "\\hat {z} = z - \\frac {\\lambda}{2} \\boldsymbol {J} _ {C} ^ {\\intercal}, \\quad C (z) + \\boldsymbol {J} _ {C} (\\hat {z} - z) = 0. \\tag {15}", + "image_path": "337ec50fd10eaee69f282b760bf876693f8f84d343467e39b80fd24f450e07f5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 182, + 244, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 244, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 244, + 194 + ], + "type": "text", + "content": "Inserting the first equation into the second yields" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 202, + 287, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 202, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 202, + 287, + 228 + ], + "type": "interline_equation", + "content": "C (\\boldsymbol {z}) - \\frac {\\lambda}{2} \\| \\boldsymbol {J} _ {C} \\| ^ {2} = 0 \\Rightarrow \\hat {\\boldsymbol {z}} = \\boldsymbol {z} - \\frac {C (\\boldsymbol {z})}{\\| \\boldsymbol {J} _ {C} \\| ^ {2}} \\boldsymbol {J} _ {C} ^ {\\intercal}. \\tag {16}", + "image_path": "bb6fc06938e6871e766322a32582f4fa560be7185d6661b0c9db7ba6d276ad73.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 237, + 202, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 202, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 202, + 249 + ], + "type": "text", + "content": "Thus the minimum in (13) is given by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 258, + 287, + 286 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 258, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 99, + 258, + 287, + 286 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {S} ^ {2} = \\left\\| \\frac {C (\\boldsymbol {z})}{\\| \\boldsymbol {J} _ {C} \\| ^ {2}} \\boldsymbol {J} _ {C} ^ {\\intercal} \\right\\| ^ {2} = \\frac {C (\\boldsymbol {z}) ^ {2}}{\\| \\boldsymbol {J} _ {C} \\| ^ {2}}. \\tag {17}", + "image_path": "e7cf71f5f2bfd5192ad648c5db39ac57a370c50977af7cd091f15c31072fffd5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 295, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 295, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 287, + 319 + ], + "type": "text", + "content": "Replacing " + }, + { + "bbox": [ + 47, + 295, + 287, + 319 + ], + "type": "inline_equation", + "content": "C(z)" + }, + { + "bbox": [ + 47, + 295, + 287, + 319 + ], + "type": "text", + "content": " with the epipolar constraint, we arrive at the classical formula for the Sampson error:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 328, + 287, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 328, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 70, + 328, + 287, + 355 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {S} ^ {2} = \\frac {\\left(\\left(\\boldsymbol {p} _ {2} ; 1\\right) ^ {\\intercal} \\boldsymbol {E} \\left(\\boldsymbol {p} _ {1} ; 1\\right)\\right) ^ {2}}{\\left\\| \\boldsymbol {E} _ {1 2} \\left(\\boldsymbol {p} _ {1} ; 1\\right) \\right\\| ^ {2} + \\left\\| \\left(\\boldsymbol {E} ^ {\\intercal}\\right) _ {1 2} \\left(\\boldsymbol {p} _ {2} ; 1\\right) \\right\\| ^ {2}}. \\tag {18}", + "image_path": "b5479a4e9027492ee149332bc98ab6017db8d89f6962ab0759f1a9c27b2789a8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 363, + 230, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 230, + 377 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 230, + 377 + ], + "type": "text", + "content": "3. Generalizing the Sampson Error" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 384, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 287, + 456 + ], + "type": "text", + "content": "The Sampson error derived in the previous section assumes a pinhole projection model. However, many cameras, in particular those with wide field-of-view, require a more complex model. One popular choice is the Kannala-Brandt [14] model in which a 3D point " + }, + { + "bbox": [ + 47, + 384, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{d} = (X,Y,Z)^{\\mathrm{T}}" + }, + { + "bbox": [ + 47, + 384, + 287, + 456 + ], + "type": "text", + "content": " in the camera coordinate system is projected as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 94, + 464, + 287, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 464, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 94, + 464, + 287, + 491 + ], + "type": "interline_equation", + "content": "\\pi (\\boldsymbol {d}) = \\frac {\\theta_ {d}}{\\sqrt {X ^ {2} + Y ^ {2}}} \\binom {f _ {x} X} {f _ {y} Y} + \\binom {c _ {x}} {c _ {y}} \\tag {19}", + "image_path": "40bcafdd0258d4253eb081a2ab8605fb041303971473fc83707fb7189085273b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 83, + 492, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 492, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 83, + 492, + 287, + 506 + ], + "type": "interline_equation", + "content": "\\theta_ {d} = \\theta \\left(1 + k _ {1} \\theta^ {2} + k _ {2} \\theta^ {4} + k _ {3} \\theta^ {6} + k _ {4} \\theta^ {8}\\right) \\tag {20}", + "image_path": "36ade9b8872a37826ed355952f86087d42b8841bf1b0639d0f806a46da2f4a23.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 510, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 510, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 137, + 510, + 287, + 525 + ], + "type": "interline_equation", + "content": "\\theta = \\operatorname {a t a n 2} \\left(\\sqrt {X ^ {2} + Y ^ {2}}, Z\\right), \\tag {21}", + "image_path": "e1faaa3452f259f5587bffaef64ec4fbcb6c0a2e9e43639a5124a9fa80695cde.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 534, + 286, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 286, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 286, + 547 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 534, + 286, + 547 + ], + "type": "inline_equation", + "content": "k_{1},\\ldots ,k_{4},f_{x},f_{y},c_{x},c_{y}" + }, + { + "bbox": [ + 47, + 534, + 286, + 547 + ], + "type": "text", + "content": " are the intrinsic parameters." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "text", + "content": "We now consider arbitrary central camera models and derive a generalization of the Sampson error. We assume that we have some forward projection function " + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\pi : \\mathbb{R}^3 \\to \\Omega" + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\Omega \\subset \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "text", + "content": " is the image plane. In the case of the Kannala-Brandt model this is given by (19), but our derivations work for any differentiable model with central projections, i.e. " + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\pi(\\lambda d) = \\pi(d) - \\forall \\lambda > 0" + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "text", + "content": ". Similarly we define the (normalized) unprojection, " + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\pi^{-1} : \\Omega \\to S^2" + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "text", + "content": " as mapping image points onto the corresponding unit-vectors in 3D. The geometric error (11) then becomes " + }, + { + "bbox": [ + 46, + 548, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{ML}(z, E) =" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 82, + 675, + 287, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 675, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 82, + 675, + 287, + 694 + ], + "type": "interline_equation", + "content": "\\min _ {\\hat {z}} \\| \\hat {z} - z \\| ^ {2} \\tag {22}", + "image_path": "635849ce3946d8b68ad39c9b7b6093e8a4b06cac6ad2159357dc0aaec7beefdb.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 85, + 696, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 696, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 85, + 696, + 287, + 710 + ], + "type": "interline_equation", + "content": "\\text {s . t .} \\quad C (\\hat {\\boldsymbol {z}}) = \\pi^ {- 1} \\left(\\hat {\\boldsymbol {p}} _ {2}\\right) ^ {\\intercal} \\boldsymbol {E} \\pi^ {- 1} \\left(\\hat {\\boldsymbol {p}} _ {1}\\right) = 0. \\tag {23}", + "image_path": "94ade997a4e1d4564544ac7e03b186b3f8ea92930b186d0471eaaffb139892e7.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "text", + "content": "In practice, most camera models only define the forward projection " + }, + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "text", + "content": ", and unprojection " + }, + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "\\pi^{-1}" + }, + { + "bbox": [ + 306, + 72, + 545, + 204 + ], + "type": "text", + "content": " is solved for iteratively. For example, (19) does not have a closed-form inverse. This makes it non-trivial to directly compute its Jacobians required for the Sampson approximation. To avoid this computation we rely on the inverse function theorem. The variant that we will use can be found in the book [6]. It is tailored to functions that map higher-dimensional inputs to lower-dimensional outputs. We rephrase it here for reader's convenience, adapting the notation to our needs and losing some generality:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "text", + "content": "Theorem 1. [6, Theorem 1E.6] Let " + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\pi : \\mathbb{R}^3 \\to \\mathbb{R}^2" + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "text", + "content": " be continuously differentiable in a neighborhood around " + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "inline_equation", + "content": "d \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "text", + "content": ". Further, suppose that its Jacobian " + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{J}" + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "text", + "content": " has full rank. Then, there exists a neighborhood " + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{p}} = \\pi(d)" + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "text", + "content": " and a continuously differentiable function " + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pi^{-1}}: V \\to \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\pi(\\widetilde{\\pi^{-1}}(p)) = p \\forall p \\in V" + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "text", + "content": " and whose Jacobian is given by " + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{J}^\\dagger" + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "text", + "content": ", the Moore-Penrose pseudoinverse of " + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{J}" + }, + { + "bbox": [ + 305, + 208, + 545, + 297 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "text", + "content": "The construction of " + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pi^{-1}}" + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "text", + "content": " in the proof from [6] also gives us that " + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pi^{-1}} (\\bar{p}) = d" + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "text", + "content": " admits many right-inverses, but they only differ in lengths assigned to unprojected bearing vectors. Thus, " + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pi^{-1}}" + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "text", + "content": " that the theorem gives us need not coincide with " + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\pi^{-1}" + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "text", + "content": " that maps points to unit vectors. However, the following result shows that when we evaluate " + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "text", + "content": " at a unit vector " + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "inline_equation", + "content": "d = \\pi^{-1}(\\bar{p})" + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "inline_equation", + "content": "J^{\\dagger}" + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "text", + "content": " actually gives us the Jacobian of " + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\pi^{-1}" + }, + { + "bbox": [ + 305, + 301, + 545, + 401 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 407, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 434 + ], + "type": "text", + "content": "Corollary 1. Let " + }, + { + "bbox": [ + 305, + 407, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{d} \\in S^2" + }, + { + "bbox": [ + 305, + 407, + 545, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 407, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{J} = \\frac{\\partial\\pi}{\\partial\\bar{\\pmb{d}}} \\in \\mathbb{R}^{2\\times 3}" + }, + { + "bbox": [ + 305, + 407, + 545, + 434 + ], + "type": "text", + "content": ". Then the Jacobian of " + }, + { + "bbox": [ + 305, + 407, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\pi^{-1}" + }, + { + "bbox": [ + 305, + 407, + 545, + 434 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 305, + 407, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{p}} = \\pi (\\pmb{d})" + }, + { + "bbox": [ + 305, + 407, + 545, + 434 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 305, + 407, + 545, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{J}^{\\dagger}" + }, + { + "bbox": [ + 305, + 407, + 545, + 434 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 439, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 439, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 439, + 545, + 464 + ], + "type": "text", + "content": "Proof. Since the only difference between " + }, + { + "bbox": [ + 305, + 439, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pi^{-1}}" + }, + { + "bbox": [ + 305, + 439, + 545, + 464 + ], + "type": "text", + "content": " from theorem 1 and " + }, + { + "bbox": [ + 305, + 439, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\pi^{-1}" + }, + { + "bbox": [ + 305, + 439, + 545, + 464 + ], + "type": "text", + "content": " is the norm of the output, we can write" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 358, + 470, + 545, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 470, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 358, + 470, + 545, + 485 + ], + "type": "interline_equation", + "content": "\\pi^ {- 1} (\\boldsymbol {p}) = h \\left(\\widetilde {\\left(\\pi^ {- 1} (\\boldsymbol {p})\\right)} \\quad \\forall \\boldsymbol {p} \\in V, \\right. \\tag {24}", + "image_path": "e8cc8f06d988f1b115733060ecc7ce1e1993f71d4e071add000edf64e97726ae.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 491, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 545, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 491, + 545, + 529 + ], + "type": "inline_equation", + "content": "h: \\pmb{v} \\mapsto \\pmb{v} / \\| \\pmb{v} \\|" + }, + { + "bbox": [ + 305, + 491, + 545, + 529 + ], + "type": "text", + "content": " is the normalization function. We also have " + }, + { + "bbox": [ + 305, + 491, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\pmb{d} = \\widetilde{\\pi^{-1}}(\\bar{\\pmb{p}})" + }, + { + "bbox": [ + 305, + 491, + 545, + 529 + ], + "type": "text", + "content": ". Thus, the Jacobian of " + }, + { + "bbox": [ + 305, + 491, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\pi^{-1}" + }, + { + "bbox": [ + 305, + 491, + 545, + 529 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 305, + 491, + 545, + 529 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{p}}" + }, + { + "bbox": [ + 305, + 491, + 545, + 529 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 534, + 545, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 534, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 318, + 534, + 545, + 564 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\pi^ {- 1}}{\\partial \\boldsymbol {p}} = \\left. \\frac {\\partial h}{\\partial \\boldsymbol {v}} \\right| _ {\\boldsymbol {v} = \\boldsymbol {d}} \\frac {\\partial \\widetilde {\\pi^ {- 1}}}{\\partial \\boldsymbol {p}} = (\\boldsymbol {I} - \\boldsymbol {d} \\boldsymbol {d} ^ {\\intercal}) \\boldsymbol {J} ^ {\\dagger} = \\boldsymbol {J} ^ {\\dagger}. \\tag {25}", + "image_path": "1530e909b9bdd4642f1ab3d3ad6bc29d0150fa9889bdbbffc8367f5dda44cc80.jpg" + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 535, + 569, + 545, + 578 + ], + "blocks": [ + { + "bbox": [ + 535, + 569, + 545, + 578 + ], + "lines": [ + { + "bbox": [ + 535, + 569, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 535, + 569, + 545, + 578 + ], + "type": "image", + "image_path": "d401e52b9febab44d26971f331f4aed318ccecd2838658691e008ee18dee3991.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 586, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 545, + 635 + ], + "type": "text", + "content": "In the last equation we used " + }, + { + "bbox": [ + 305, + 586, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{d}^{\\mathsf{T}}\\pmb{J}^{\\dagger} = \\mathbf{0}" + }, + { + "bbox": [ + 305, + 586, + 545, + 635 + ], + "type": "text", + "content": ", which is true in general for any null-vector " + }, + { + "bbox": [ + 305, + 586, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{d}" + }, + { + "bbox": [ + 305, + 586, + 545, + 635 + ], + "type": "text", + "content": " of the matrix " + }, + { + "bbox": [ + 305, + 586, + 545, + 635 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 305, + 586, + 545, + 635 + ], + "type": "text", + "content": ", but in our case can also be seen from the following lemma that gives a closed-form expression for " + }, + { + "bbox": [ + 305, + 586, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{J}^{\\dagger}" + }, + { + "bbox": [ + 305, + 586, + 545, + 635 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 639, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 639, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 639, + 545, + 665 + ], + "type": "text", + "content": "Lemma 1. Let " + }, + { + "bbox": [ + 305, + 639, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{d} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 639, + 545, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 639, + 545, + 665 + ], + "type": "inline_equation", + "content": "J = \\frac{\\partial\\pi}{\\partial\\pmb{d}} \\in \\mathbb{R}^{2\\times 3}" + }, + { + "bbox": [ + 305, + 639, + 545, + 665 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 305, + 639, + 545, + 665 + ], + "type": "inline_equation", + "content": "J^{\\dagger}" + }, + { + "bbox": [ + 305, + 639, + 545, + 665 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 324, + 669, + 545, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 669, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 324, + 669, + 545, + 696 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} = \\frac {1}{\\boldsymbol {d} \\cdot \\left(\\boldsymbol {g} _ {x} \\times \\boldsymbol {g} _ {y}\\right)} \\left[ \\left(\\boldsymbol {g} _ {y} \\times \\boldsymbol {d}\\right), \\quad \\left(\\boldsymbol {d} \\times \\boldsymbol {g} _ {x}\\right) \\right], \\tag {26}", + "image_path": "a23f2d2106bc1c556a32ab9daf60a96edbc4b8df4c7773bdac83d1fe13ae599f.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 700, + 531, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 700, + 531, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 700, + 531, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 700, + 531, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{g}_x" + }, + { + "bbox": [ + 306, + 700, + 531, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 700, + 531, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{g}_y \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 306, + 700, + 531, + 715 + ], + "type": "text", + "content": " are the two rows of the Jacobian " + }, + { + "bbox": [ + 306, + 700, + 531, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{J}" + }, + { + "bbox": [ + 306, + 700, + 531, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3373" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 110 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 110 + ], + "type": "text", + "content": "Proof. Simple calculations show that " + }, + { + "bbox": [ + 47, + 72, + 289, + 110 + ], + "type": "inline_equation", + "content": "JM = I_{2 \\times 2}" + }, + { + "bbox": [ + 47, + 72, + 289, + 110 + ], + "type": "text", + "content": ". It remains to show that " + }, + { + "bbox": [ + 47, + 72, + 289, + 110 + ], + "type": "inline_equation", + "content": "MJ \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 47, + 72, + 289, + 110 + ], + "type": "text", + "content": " is symmetric. For ease of notation, introduce " + }, + { + "bbox": [ + 47, + 72, + 289, + 110 + ], + "type": "inline_equation", + "content": "\\alpha = d \\cdot (g_x \\times g_y)" + }, + { + "bbox": [ + 47, + 72, + 289, + 110 + ], + "type": "text", + "content": ". Then we have" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 117, + 287, + 147 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 117, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 63, + 117, + 287, + 147 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\alpha \\boldsymbol {M} \\boldsymbol {J} = (\\boldsymbol {g} _ {y} \\times \\boldsymbol {d}) \\boldsymbol {g} _ {x} ^ {\\intercal} + (\\boldsymbol {d} \\times \\boldsymbol {g} _ {x}) \\boldsymbol {g} _ {y} ^ {\\intercal} = (27) \\\\ = [ \\boldsymbol {d} ] _ {\\times} \\left(- \\boldsymbol {g} _ {y} \\boldsymbol {g} _ {x} ^ {\\intercal} + \\boldsymbol {g} _ {x} \\boldsymbol {g} _ {y} ^ {\\intercal}\\right) = [ \\boldsymbol {d} ] _ {\\times} [ \\boldsymbol {g} _ {y} \\times \\boldsymbol {g} _ {x} ] _ {\\times}. (28) \\\\ \\end{array}", + "image_path": "61fb677052823837f8db3caa80de30feb44a90d0cd44b04e071fc6f1d0106802.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "text", + "content": "Since the projection function " + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "text", + "content": " does not depend on the norm of " + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\pmb{d}" + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\pi(\\pmb{d}) = \\pi(\\lambda \\pmb{d}) \\forall \\lambda > 0" + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "text", + "content": ", we have that " + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\pmb{J} \\pmb{d} = \\mathbf{0}" + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\pmb{d}" + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "text", + "content": " is then orthogonal to both " + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\pmb{g}_x" + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\pmb{g}_y" + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "text", + "content": " we have " + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\pmb{g}_y \\times \\pmb{g}_x = \\gamma \\pmb{d}" + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "text", + "content": " for some " + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}" + }, + { + "bbox": [ + 47, + 154, + 287, + 204 + ], + "type": "text", + "content": ". Inserting into (28) we get" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 209, + 287, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 209, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 134, + 209, + 287, + 232 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} \\boldsymbol {J} = \\frac {\\gamma}{\\alpha} [ \\boldsymbol {d} ] _ {\\times} ^ {2}, \\tag {29}", + "image_path": "b7afb16a716fca1a3cd4d59e111fafcd828c3a4aaf7b9ce4327ebf2f7b36ffe4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 237, + 167, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 167, + 248 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 167, + 248 + ], + "type": "text", + "content": "which is a symmetric matrix." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 277, + 249, + 287, + 258 + ], + "blocks": [ + { + "bbox": [ + 277, + 249, + 287, + 258 + ], + "lines": [ + { + "bbox": [ + 277, + 249, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 277, + 249, + 287, + 258 + ], + "type": "image", + "image_path": "9fd1beca797ef476155a5745b8b3b1554b9fe4f9bd2437aff48b003af4503df7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 268, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 268, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 268, + 287, + 316 + ], + "type": "text", + "content": "The expression (26) can be used to efficiently compute the Jacobian " + }, + { + "bbox": [ + 47, + 268, + 287, + 316 + ], + "type": "inline_equation", + "content": "J_{C}" + }, + { + "bbox": [ + 47, + 268, + 287, + 316 + ], + "type": "text", + "content": " of the constraint function " + }, + { + "bbox": [ + 47, + 268, + 287, + 316 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 268, + 287, + 316 + ], + "type": "text", + "content": " from (23), and finally to arrive at the Sampson approximation of the reprojection error in any differentiable camera model:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 322, + 287, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 322, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 62, + 322, + 287, + 352 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {T S} ^ {2} = \\frac {C (\\boldsymbol {z}) ^ {2}}{\\| \\boldsymbol {J} _ {C} \\| ^ {2}} = \\frac {\\left(\\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1}\\right) ^ {2}}{\\| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {J} _ {1} ^ {\\dagger} \\| ^ {2} + \\| \\boldsymbol {d} _ {1} ^ {\\intercal} \\boldsymbol {E} ^ {\\intercal} \\boldsymbol {J} _ {2} ^ {\\dagger} \\| ^ {2}}, \\tag {30}", + "image_path": "1846c3cae88ed4bde89b986a071aba5ddbe77b0b177617b115bd8f162ec2c5d6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\pmb{d}_i = \\pi^{-1}(\\pmb{p}_i)" + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": " and the two Jacobians are evaluated at " + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\pmb{d}_1" + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\pmb{d}_2" + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": " respectively. In the rest of the paper we refer to this as the Tangent Sampson error. The name reflects the fact that after we linearize the unprojection " + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\pi^{-1}" + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": ", we are essentially optimizing in the tangent plane to " + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": " at the bearing vectors of the undistorted points. Note that the formula is analogous to the original Sampson error (18) (replacing " + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\pmb{d}_1" + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\pmb{d}_2" + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": " with the image points, and replacing " + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "inline_equation", + "content": "J_1^\\dagger" + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "inline_equation", + "content": "J_2^\\dagger" + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "inline_equation", + "content": "[10; 01; 00]" + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": "). Further, since " + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "inline_equation", + "content": "J_1" + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "inline_equation", + "content": "J_2" + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": " are evaluated at the original points of the correspondence, they are independent of " + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\pmb{E}" + }, + { + "bbox": [ + 47, + 358, + 287, + 514 + ], + "type": "text", + "content": " and can be precomputed. Thus, during RANSAC and refinement the computational cost stays approximately the same as for the classical Sampson error." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 521, + 195, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 195, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 195, + 533 + ], + "type": "text", + "content": "3.1. Backward Camera Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 539, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 287, + 635 + ], + "type": "text", + "content": "While most camera models used in practice are forward-models, i.e. mapping from the world to the image, there also exist backward-models which instead take the image points as input. The division model (8) is perhaps the most popular example here. The derivations above can easily be adapted to allow for backward-models. The pseudoinverses " + }, + { + "bbox": [ + 47, + 539, + 287, + 635 + ], + "type": "inline_equation", + "content": "J_{k}^{\\dagger}" + }, + { + "bbox": [ + 47, + 539, + 287, + 635 + ], + "type": "text", + "content": " are simply replaced with the " + }, + { + "bbox": [ + 47, + 539, + 287, + 635 + ], + "type": "inline_equation", + "content": "3 \\times 2" + }, + { + "bbox": [ + 47, + 539, + 287, + 635 + ], + "type": "text", + "content": " Jacobians from the undistortion map," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 93, + 641, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 641, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 93, + 641, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {T S} ^ {2} = \\frac {\\left(\\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1}\\right) ^ {2}}{\\left\\| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {J} _ {1} \\right\\| ^ {2} + \\left\\| \\boldsymbol {d} _ {1} ^ {\\intercal} \\boldsymbol {E} ^ {\\intercal} \\boldsymbol {J} _ {2} \\right\\| ^ {2}}. \\tag {31}", + "image_path": "ff3b9acc1ce3e3a62e3b8b10bb65234de57a0974e07075e07268a01155e1b382.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 674, + 269, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 674, + 269, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 674, + 269, + 686 + ], + "type": "text", + "content": "In particular, for the division model (8) we would have" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 70, + 693, + 287, + 718 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 693, + 287, + 718 + ], + "spans": [ + { + "bbox": [ + 70, + 693, + 287, + 718 + ], + "type": "interline_equation", + "content": "\\boldsymbol {J} = \\frac {\\partial h}{\\partial \\boldsymbol {d}} \\frac {\\partial \\boldsymbol {d}}{\\partial \\boldsymbol {x}} = \\frac {\\partial h}{\\partial \\boldsymbol {d}} \\left[ I _ {2 \\times 2}; 2 \\lambda \\boldsymbol {x} ^ {\\intercal} \\right] \\in \\mathbb {R} ^ {3 \\times 2}, \\tag {32}", + "image_path": "e7e8b01f0b04073648e01ab836f7bd879f1cc6c31bf03dbf29557c40d2a5508e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "where we need to explicitly include the Jacobian " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\frac{\\partial h}{\\partial d}" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": " of the normalization function " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": ", since the unprojected " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\pmb{d}" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": " from (8) is not a unit vector." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 114, + 451, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 114, + 451, + 127 + ], + "spans": [ + { + "bbox": [ + 306, + 114, + 451, + 127 + ], + "type": "text", + "content": "3.2. Measurement Uncertainty" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 133, + 545, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 545, + 181 + ], + "type": "text", + "content": "Following the approach from [3], we can incorporate measurement uncertainty in the keypoint positions into the Tangent Sampson error. We will assume that the keypoint positions are affected by independent Gaussian noise" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 347, + 186, + 545, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 186, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 347, + 186, + 545, + 200 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {1} \\sim \\mathcal {N} \\left(\\bar {\\boldsymbol {p}} _ {1}, \\boldsymbol {\\Sigma} _ {1}\\right), \\quad \\boldsymbol {p} _ {2} \\sim \\mathcal {N} \\left(\\bar {\\boldsymbol {p}} _ {2}, \\boldsymbol {\\Sigma} _ {2}\\right) \\tag {33}", + "image_path": "bc88c385ea4010bb568343659d1df8466256e5a9e30b8b82ff890a807ed466cf.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 205, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 545, + 228 + ], + "type": "text", + "content": "with known covariances. For observations " + }, + { + "bbox": [ + 305, + 205, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{p}_1" + }, + { + "bbox": [ + 305, + 205, + 545, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 205, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{p}_2" + }, + { + "bbox": [ + 305, + 205, + 545, + 228 + ], + "type": "text", + "content": ", the maximum-likelihood error for an essential matrix " + }, + { + "bbox": [ + 305, + 205, + 545, + 228 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 305, + 205, + 545, + 228 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 324, + 232, + 545, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 232, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 324, + 232, + 545, + 270 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {E} _ {M L} ^ {2} (\\boldsymbol {z}, \\boldsymbol {E}) = \\min _ {\\hat {z}} \\quad \\frac {1}{2} (\\boldsymbol {z} - \\hat {\\boldsymbol {z}}) ^ {\\intercal} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {z} - \\hat {\\boldsymbol {z}}) \\tag {34} \\\\ \\begin{array}{c c} \\text {s . t .} & C (\\hat {\\boldsymbol {z}}) = 0, \\end{array} \\\\ \\end{array}", + "image_path": "f50947784101038b07cbb7856cabd6a08231654adede705893d660c99afeafed.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 276, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 276, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 276, + 545, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 276, + 545, + 300 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 276, + 545, + 300 + ], + "type": "text", + "content": " is the constraint function (23) and " + }, + { + "bbox": [ + 305, + 276, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 305, + 276, + 545, + 300 + ], + "type": "text", + "content": " is the joint covariance of " + }, + { + "bbox": [ + 305, + 276, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\pmb{z} = [\\pmb{p}_1; \\pmb{p}_2]" + }, + { + "bbox": [ + 305, + 276, + 545, + 300 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 305, + 276, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma} = \\text{diag}(\\pmb{\\Sigma}_1, \\pmb{\\Sigma}_2)" + }, + { + "bbox": [ + 305, + 276, + 545, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 300, + 521, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 300, + 521, + 312 + ], + "spans": [ + { + "bbox": [ + 318, + 300, + 521, + 312 + ], + "type": "text", + "content": "The procedure from Section 2.1 yields in this case" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 316, + 545, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 316, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 320, + 316, + 545, + 344 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {T S} ^ {2} = \\frac {C (\\boldsymbol {z}) ^ {2}}{\\boldsymbol {J} _ {C} \\boldsymbol {\\Sigma} \\boldsymbol {J} _ {C} ^ {\\intercal}} = \\frac {\\left(\\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {d} _ {1}\\right) ^ {2}}{\\| \\boldsymbol {d} _ {2} ^ {\\intercal} \\boldsymbol {E} \\boldsymbol {J} _ {1} ^ {\\dagger} \\| _ {\\Sigma_ {1}} ^ {2} + \\| \\boldsymbol {d} _ {1} ^ {\\intercal} \\boldsymbol {E} ^ {\\intercal} \\boldsymbol {J} _ {2} ^ {\\dagger} \\| _ {\\Sigma_ {2}} ^ {2}}, \\tag {35}", + "image_path": "d07a74754e7ace39780021ccb42a3ab9167edbb9154c45ee4bb3c488ac9cf111.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 348, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 348, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 348, + 545, + 408 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 348, + 545, + 408 + ], + "type": "inline_equation", + "content": "\\| \\pmb{v} \\|_{\\Sigma}^{2} = \\pmb{v}^{T} \\pmb{\\Sigma} \\pmb{v}" + }, + { + "bbox": [ + 305, + 348, + 545, + 408 + ], + "type": "text", + "content": ". See the supplementary material for details on the derivation. This is a generalization of the last expression in [3, Section 3.4] to the arbitrary camera model case. We evaluate how it compares to the original Tangent Sampson error (30) in a synthetic setting in Section 4.4.1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 418, + 449, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 449, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 449, + 430 + ], + "type": "text", + "content": "4. Experimental Evaluation" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 437, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 474 + ], + "type": "text", + "content": "In the rest of the paper we experimentally evaluate the Tangent Sampson error. We compare against a selection of popular error metrics, summarized in Table 1." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 479, + 438, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 438, + 492 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 438, + 492 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 498, + 545, + 713 + ], + "type": "text", + "content": "We implemented all error formulations in " + }, + { + "bbox": [ + 305, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{C + + }" + }, + { + "bbox": [ + 305, + 498, + 545, + 713 + ], + "type": "text", + "content": " using Eigen [9] as a linear algebra library and Ceres [1] for optimization. Since evaluation is run on Intel hardware, we additionally speed up the errors that do not require iterative methods or explicit camera projection " + }, + { + "bbox": [ + 305, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{E}_{TS},\\mathcal{E}_S,\\mathcal{E}_{SED}," + }, + { + "bbox": [ + 305, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{CS}" + }, + { + "bbox": [ + 305, + 498, + 545, + 713 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 305, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{ALG})" + }, + { + "bbox": [ + 305, + 498, + 545, + 713 + ], + "type": "text", + "content": " by vectorizing their implementations with AVX and FMA instruction sets. In Table 1 we also list the approximate computational cost of evaluating the residual. Tangent Sampson error is significantly faster compared to other metrics which also measure the error in the original image, and has comparable, but slightly higher, runtime to the purely geometric and pinhole-based metrics. For the computational cost we include parts which cannot be precomputed (i.e. parts that depend on the essential matrix). For a comparison with the non-optimized errors (using pure Eigen-implementations), we refer the reader to the supplementary material, where we list the runtime for all formulations without explicit vectorization and some setup details." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3374" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 193 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 193 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 193 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 193 + ], + "type": "table", + "html": "
NameRuntime
Rel.Abs.
\\( \\mathcal{E}_{ALG} \\)(7)Algebraic error↓1.01.3 ns
\\( \\mathcal{E}_{CS} \\)(6)Cosine error↓1.82.4 ns
\\( \\mathcal{E}_{ML} \\)(3)(Pinhole) Repoj. error\\( \\blacktriangleright \\)28.036.7 ns
\\( \\mathcal{E}_{SED} \\)(4)(Pinhole) Sym. Epi. Dist.\\( \\blacktriangleright \\)1.72.3 ns
\\( \\mathcal{E}_{S} \\)(5)(Pinhole) Sampson error\\( \\blacktriangleright \\)1.41.9 ns
\\( \\mathcal{E}_{\\pi ML} \\)(22)Reproj. error\\( \\square \\)8861160 ns
\\( \\mathcal{E}_{\\pi SED} \\)(10)Proj. Sym. Epi. Dist.\\( \\square \\)92121 ns
\\( \\mathcal{E}_{TS} \\)(30)Tangent Sampson\\( \\square \\)2.22.8 ns
", + "image_path": "36d35b93157ff82a80a66a676bd31e0b84fd9d50d8cce3eabee2e4aba956e87a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 201, + 287, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 258 + ], + "type": "text", + "content": "Table 1. Error metrics compared in the experimental evaluation. The errors are grouped by the domain where the error is computed: geometric errors " + }, + { + "bbox": [ + 46, + 201, + 287, + 258 + ], + "type": "inline_equation", + "content": "(\\mathcal{L})" + }, + { + "bbox": [ + 46, + 201, + 287, + 258 + ], + "type": "text", + "content": ", the undistorted image " + }, + { + "bbox": [ + 46, + 201, + 287, + 258 + ], + "type": "inline_equation", + "content": "(\\mathfrak{X})" + }, + { + "bbox": [ + 46, + 201, + 287, + 258 + ], + "type": "text", + "content": " and the original image " + }, + { + "bbox": [ + 46, + 201, + 287, + 258 + ], + "type": "inline_equation", + "content": "(\\square)" + }, + { + "bbox": [ + 46, + 201, + 287, + 258 + ], + "type": "text", + "content": ". Computational cost is shown relative to the algebraic error which is the cheapest." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 262, + 110, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 110, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 110, + 274 + ], + "type": "text", + "content": "4.2. Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": "For the evaluation we consider two datasets. First the meta-calibration dataset from BabelCalib [21]. The dataset is a collection of checkerboard images from a wide variety of cameras, including some with extremely wide field-of-view " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": "(>270^{\\circ})" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": ". For the evaluation we split the dataset into three groups depending on their field-of-view; " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": "< 90^{\\circ}" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": "90 - 150^{\\circ}" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": ">150^{\\circ}" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": ". As the original dataset is biased towards higher field-of-view cameras, we also include some additional checkerboard images taken with a standard DSLR (circa " + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "inline_equation", + "content": "50 - 70^{\\circ}" + }, + { + "bbox": [ + 46, + 281, + 287, + 472 + ], + "type": "text", + "content": ") to even out the distribution. These checkerboard datasets allow us to evaluate on a large number of different camera and lens setups. In Section 4.3 we show that the Tangent Sampson error consistently approximates the true reprojection error, regardless of the camera used, and in Section 4.4 we show that when used for local refinement it yields the most accurate epipolar geometries." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "content": "To evaluate the performance of the error metrics in the context of RANSAC we consider image pairs extracted from Structure-from-Motion datasets, where the poses from the reconstruction are taken as a psuedo-ground truth. As the experiments in Section 4.3-4.4 show that most error metrics behave similarly for low to medium field-of-view cameras " + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "inline_equation", + "content": "(< 90^{\\circ})" + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "content": ", we focus our evaluation on cameras with larger field-of-view. For this we use two datasets (Grossmunster and Kirchenge) from Larsson et al. [17], which contain 373 and 369 images taken with a fisheye-lens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 599, + 270, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 270, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 270, + 613 + ], + "type": "text", + "content": "4.3. Comparison with True Reprojection Error" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "We first evaluate how different error metrics compare with the true maximum-likelihood estimate, " + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{ML}" + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": " (3) for pinhole cameras and " + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\pi ML}" + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": " (22) for non-pinhole cameras. To compute the ML-estimate we perform two-view triangulation followed by non-linear optimization of the 3D point, minimizing the squared reprojection error (using the known camera intrinsics). Note that this approach is orders of magnitude slower compared to the metrics we evaluate (see Ta" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 84 + ], + "type": "text", + "content": "ble 1) and cannot reasonably be used in practical scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 85, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 85, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 545, + 203 + ], + "type": "text", + "content": "For the experiment we consider the collection of checkerboard images from [21]. We use the calibration toolbox from BabelCalib [21] to estimate the extrinsic and intrinsic parameters for each image. For each pair we collect the 2D-2D correspondences from corresponding corners. As the corner detector [10] is extremely accurate on these images (some having around " + }, + { + "bbox": [ + 304, + 85, + 545, + 203 + ], + "type": "inline_equation", + "content": "0.2\\mathrm{px}" + }, + { + "bbox": [ + 304, + 85, + 545, + 203 + ], + "type": "text", + "content": " error during calibration), we add Gaussian noise with " + }, + { + "bbox": [ + 304, + 85, + 545, + 203 + ], + "type": "inline_equation", + "content": "1\\mathrm{px}" + }, + { + "bbox": [ + 304, + 85, + 545, + 203 + ], + "type": "text", + "content": " standard deviation, to better differentiate the methods (as all methods give zero error for exact data)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 205, + 545, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 545, + 254 + ], + "type": "text", + "content": "Since it is not meaningful to directly compare the residual values, we instead compare their ordering. For this we make use of Kendall's " + }, + { + "bbox": [ + 305, + 205, + 545, + 254 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 205, + 545, + 254 + ], + "type": "text", + "content": " rank correlation coefficient [15]. It is defined for two rankings " + }, + { + "bbox": [ + 305, + 205, + 545, + 254 + ], + "type": "inline_equation", + "content": "(p_1,\\dots ,p_n)" + }, + { + "bbox": [ + 305, + 205, + 545, + 254 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 205, + 545, + 254 + ], + "type": "inline_equation", + "content": "(q_{1},\\ldots ,q_{n})" + }, + { + "bbox": [ + 305, + 205, + 545, + 254 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 270, + 545, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 270, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 270, + 545, + 297 + ], + "type": "interline_equation", + "content": "\\tau (p, q) = \\frac {\\left| \\{\\text {c o n c o r d a n t p a i r s} \\} \\right| - \\left| \\{\\text {d i s c o r d a n t p a i r s} \\} \\right|}{\\binom {n} {2}}. \\tag {36}", + "image_path": "6d9ac4c36837dde5bc0440b9c66cb3a5a7de87d7f5185eae14efb53ff0218c4f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "type": "text", + "content": "A pair of indices " + }, + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "type": "inline_equation", + "content": "(i,j)\\in \\{1,\\ldots ,n\\} ^2" + }, + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "type": "inline_equation", + "content": "i > j" + }, + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "type": "text", + "content": " is called concordant if and only if" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 329, + 342, + 545, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 342, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 329, + 342, + 545, + 356 + ], + "type": "interline_equation", + "content": "\\left(p _ {i} > p _ {j} \\wedge q _ {i} > q _ {j}\\right) \\vee \\left(p _ {i} < p _ {j} \\wedge q _ {i} < q _ {j}\\right). \\tag {37}", + "image_path": "1d67aaab7dadfb25adf5acebf96a3d84f12a32ff7edc8775b83e53c8a28e283a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "type": "text", + "content": "Otherwise, the pair is discordant. Note that due to the normalization we have " + }, + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "type": "inline_equation", + "content": "\\tau(p,q) \\in [-1,1]" + }, + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "type": "text", + "content": ", where identical rankings yield " + }, + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "type": "inline_equation", + "content": "\\tau = 1" + }, + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "type": "text", + "content": ", and reversed " + }, + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "type": "inline_equation", + "content": "\\tau = -1" + }, + { + "bbox": [ + 304, + 365, + 545, + 400 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 402, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 510 + ], + "type": "text", + "content": "The results are shown in Fig. 3, split into three groups based on the camera field-of-view. For low to medium field-of-view cameras " + }, + { + "bbox": [ + 304, + 402, + 545, + 510 + ], + "type": "inline_equation", + "content": "(< 90^{\\circ})" + }, + { + "bbox": [ + 304, + 402, + 545, + 510 + ], + "type": "text", + "content": " we can see that all metrics except the algebraic residual behave similar to the true reprojection error. As the field-of-view increases, there is a more clear difference between the methods, with the pinhole-based errors performing the worst. Note however, that for all cameras the proposed Tangent Sampson error offers the best approximation of the true reprojection error." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "In addition to Kendall-Tau distance, we compare the errors qualitatively. For this, we show the values of an error for a fixed correspondence on a pair of images from a panoramic and fisheye camera in Fig. 4. There, we keep the point on the fisheye image fixed, and vary the point on the panoramic image to show different error values coded by different colors. One can see that " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{TS}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " error models " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\pi ML}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " closely in the region where the latter is well-defined. " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{TS}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " is also one of the two errors that do not degenerate around the epipole, and, at the same time, it correctly models the spread of values of " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\pi ML}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " around the epipolar curve. Note that the shape of true reprojection error varies significantly depending on how close the correspondence is to the epipole. In the supplementary material we show more qualitative results. In the supplementary material we also show a comparison with [16], which proposed an optimal method for computing the " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pi ML" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " error (22) for the division model (8)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3375" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 77, + 232, + 216 + ], + "blocks": [ + { + "bbox": [ + 54, + 77, + 232, + 216 + ], + "lines": [ + { + "bbox": [ + 54, + 77, + 232, + 216 + ], + "spans": [ + { + "bbox": [ + 54, + 77, + 232, + 216 + ], + "type": "image", + "image_path": "9f8f9863b3dffe9841e2d7a3e51d10e5305c5ca8f233618b3decbd19021312a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 238, + 76, + 376, + 216 + ], + "blocks": [ + { + "bbox": [ + 238, + 76, + 376, + 216 + ], + "lines": [ + { + "bbox": [ + 238, + 76, + 376, + 216 + ], + "spans": [ + { + "bbox": [ + 238, + 76, + 376, + 216 + ], + "type": "image", + "image_path": "43382c54d7df2953af5c5c9ea9a3c5e819ce256ec61fc26f7437a18c94e6eabf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 384, + 76, + 521, + 216 + ], + "blocks": [ + { + "bbox": [ + 384, + 76, + 521, + 216 + ], + "lines": [ + { + "bbox": [ + 384, + 76, + 521, + 216 + ], + "spans": [ + { + "bbox": [ + 384, + 76, + 521, + 216 + ], + "type": "image", + "image_path": "d3cc5e8188087f251cb695991e8e1d4294fe8cb4cde52620bb4b989d2dd3358c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 54, + 251, + 175, + 313 + ], + "blocks": [ + { + "bbox": [ + 46, + 227, + 545, + 249 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 545, + 249 + ], + "type": "text", + "content": "Figure 3. The figure shows the Kendall " + }, + { + "bbox": [ + 46, + 227, + 545, + 249 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 227, + 545, + 249 + ], + "type": "text", + "content": " coefficient comparing different error metrics for the checkerboard data. For all cameras the proposed Tangent Sampson error correlates well with the true reprojection error " + }, + { + "bbox": [ + 46, + 227, + 545, + 249 + ], + "type": "inline_equation", + "content": "\\pi ML" + }, + { + "bbox": [ + 46, + 227, + 545, + 249 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 251, + 175, + 313 + ], + "lines": [ + { + "bbox": [ + 54, + 251, + 175, + 313 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 175, + 313 + ], + "type": "image", + "image_path": "4afa0f4d629a848b416988c4ff23f50895fcd5deb1029428b8d7f4760eef655c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 175, + 251, + 296, + 313 + ], + "blocks": [ + { + "bbox": [ + 175, + 251, + 296, + 313 + ], + "lines": [ + { + "bbox": [ + 175, + 251, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 175, + 251, + 296, + 313 + ], + "type": "image", + "image_path": "cba1d2bf71e6381b62eab4f5cb6c7e32dfab92369c76378c13175ba5acd16d06.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 322, + 547, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 322, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 547, + 378 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of different error functions. The thresholds for colormaps were selected to be the error at the point marked by the blue plus. The shape of the true reprojection error varies heavily depending on how close the correspondence is to the epipoles. In the supplementary material we show more examples. We also show " + }, + { + "bbox": [ + 46, + 322, + 547, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\pi SED}" + }, + { + "bbox": [ + 46, + 322, + 547, + 378 + ], + "type": "text", + "content": " and errors computed on the undistorted image in the supplementary. The geometric error is only shown in the region where the 3D point after midpoint triangulation does not correspond to negative depths of bearing vectors (areas where this does not hold for one or both depths are highlighted)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 297, + 251, + 417, + 313 + ], + "blocks": [ + { + "bbox": [ + 297, + 251, + 417, + 313 + ], + "lines": [ + { + "bbox": [ + 297, + 251, + 417, + 313 + ], + "spans": [ + { + "bbox": [ + 297, + 251, + 417, + 313 + ], + "type": "image", + "image_path": "ea6147ceb46990171650d7039770608f9b014e2d790c4f61e6ba52ca5842ae4d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 418, + 251, + 539, + 313 + ], + "blocks": [ + { + "bbox": [ + 418, + 251, + 539, + 313 + ], + "lines": [ + { + "bbox": [ + 418, + 251, + 539, + 313 + ], + "spans": [ + { + "bbox": [ + 418, + 251, + 539, + 313 + ], + "type": "image", + "image_path": "06a0fdee873acd6009111bca766195b0a5328473d4045423eb20eb98ead0b4d9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 383, + 207, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 207, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 207, + 396 + ], + "type": "text", + "content": "4.4. Accuracy in Pose Refinement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 402, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 605 + ], + "type": "text", + "content": "Next we evaluate the error metrics on how they perform when used for pose-refinement. For the experiment we again consider the image pairs extracted from the checkerboard images. For each image pair we perform local optimization minimizing the proposed error metrics over all correspondences. To initialize the optimization we take the ground truth relative pose and randomly perturb both the rotation and translation by 1 degree. Additionally, point positions were perturbed with Gaussian noise with unit standard deviation along each dimension. For the two ML metrics we triangulate 3D points which are optimized jointly with the camera pose. We experience that this sometimes has problems with worse convergence. The results are shown in Table 2. Again we see that for low to medium field-of-view cameras, the difference between the error metrics is small. The Tangent Sampson error consistently provides the most accurate camera poses after refinement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Convergence Properties. The previous experiment evaluated the accuracy of the camera poses obtained when initialized close to the correct pose. We also experimented with varying the initial pose and measuring how often the optimization converged to the correct pose. Figure 5 shows the percentage of poses (across all pairs) that are within 5 degrees average rotation/translation error for varying errors in the initial pose. For large initial errors " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "(>25^{\\circ})" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ", the proposed Tangent-Sampson error performs slightly better com" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 311, + 384, + 545, + 501 + ], + "blocks": [ + { + "bbox": [ + 311, + 384, + 545, + 501 + ], + "lines": [ + { + "bbox": [ + 311, + 384, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 311, + 384, + 545, + 501 + ], + "type": "image", + "image_path": "497cdf7c45381b2230692d8d70a7f185c1b300470d2dd2823f7a9d2d74ec0736.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 512, + 545, + 534 + ], + "lines": [ + { + "bbox": [ + 305, + 512, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 305, + 512, + 545, + 534 + ], + "type": "text", + "content": "Figure 5. Convergence properties. Successful trials (average rot. and trans. error " + }, + { + "bbox": [ + 305, + 512, + 545, + 534 + ], + "type": "inline_equation", + "content": "< 5^{\\circ}" + }, + { + "bbox": [ + 305, + 512, + 545, + 534 + ], + "type": "text", + "content": ") for varying initial pose errors (in degrees)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 542, + 414, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 542, + 414, + 554 + ], + "spans": [ + { + "bbox": [ + 306, + 542, + 414, + 554 + ], + "type": "text", + "content": "pared to the other metrics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 573, + 491, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 491, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 491, + 586 + ], + "type": "text", + "content": "4.4.1 Effect of Measurement Uncertainty" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 593, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 700 + ], + "type": "text", + "content": "Experiments in [3] already show that taking point covariances into account in the cost function provides benefits for fundamental matrix estimation. Here, we conduct another synthetic test to show that this effect persists when estimating the essential matrix with highly nonlinear cameras. We generate 2000 instances with pairs of spherical cameras " + }, + { + "bbox": [ + 304, + 593, + 545, + 700 + ], + "type": "inline_equation", + "content": "(1000 \\times 2000\\mathrm{px})" + }, + { + "bbox": [ + 304, + 593, + 545, + 700 + ], + "type": "text", + "content": " with a random relative pose. For each pair we project 100 random 3D-points uniformly drawn from the cube " + }, + { + "bbox": [ + 304, + 593, + 545, + 700 + ], + "type": "inline_equation", + "content": "[-10, 10]^3" + }, + { + "bbox": [ + 304, + 593, + 545, + 700 + ], + "type": "text", + "content": ", and add Gaussian noise with covariance " + }, + { + "bbox": [ + 304, + 593, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 304, + 593, + 545, + 700 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Covariances are generated as " + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\Lambda = \\alpha Q\\mathrm{diag}(\\beta ,1 - \\beta)Q^{\\intercal}" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3376" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 531, + 208 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 531, + 208 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 531, + 208 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 531, + 208 + ], + "type": "table", + "html": "
< 90° (5332 pairs)90 - 150° (21927 pairs)> 150° (11860 pairs)
RotationTranslationRotationTranslationRotationTranslation
\\( \\mathcal{E}_{ALG} \\)0.379 / 0.3141.137 / 0.3281.528 / 0.3783.938 / 0.5881.582 / 0.5063.489 / 0.655
\\( \\mathcal{E}_S \\)0.376 / 0.3121.056 / 0.3271.464 / 0.3473.520 / 0.5751.634 / 0.5673.812 / 0.778
\\( \\mathcal{E}_{SED} \\)0.376 / 0.3121.065 / 0.3271.475 / 0.3523.554 / 0.5782.596 / 0.9865.667 / 1.342
\\( \\mathcal{E}_{COS} \\)0.376 / 0.3131.066 / 0.3261.454 / 0.3413.459 / 0.5581.447 / 0.4543.235 / 0.612
\\( \\mathcal{E}_{ML} \\)0.376 / 0.3121.103 / 0.3271.472 / 0.3583.733 / 0.5751.630 / 0.5633.771 / 0.766
\\( \\mathcal{E}_{\\pi ML} \\)0.375 / 0.3121.075 / 0.3271.451 / 0.3503.778 / 0.5571.418 / 0.4463.143 / 0.597
\\( \\mathcal{E}_{\\pi SED} \\)0.376 / 0.3131.074 / 0.3271.458 / 0.3433.490 / 0.5591.473 / 0.4503.249 / 0.607
\\( \\mathcal{E}_{TS} \\)0.376 / 0.3121.057 / 0.3271.443 / 0.3403.432 / 0.5571.465 / 0.4453.226 / 0.598
", + "image_path": "8db0af217d6f9825c8c06953ded14f13b545eac87d1b865543b79b85d8481ed0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 254, + 279, + 334 + ], + "blocks": [ + { + "bbox": [ + 58, + 254, + 279, + 334 + ], + "lines": [ + { + "bbox": [ + 58, + 254, + 279, + 334 + ], + "spans": [ + { + "bbox": [ + 58, + 254, + 279, + 334 + ], + "type": "image", + "image_path": "d2a99a1fc063882dcc1eeac8ac24b68cbb4d615cb714de51983e69e9f10c5048.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 344, + 288, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 344, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 288, + 411 + ], + "type": "text", + "content": "Figure 6. Measurement Uncertainty. The distribution of pose errors (maximum of rotation and translation error) when using covariance-aware Tangent Sampson error for pose refinement versus non-covariance-aware in synthetic experiments. Boxes show quartiles of the respective distribution of errors and whiskers extend to the minimal and maximal value." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 419, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 419, + 287, + 563 + ], + "type": "inline_equation", + "content": "Q \\sim U(\\mathrm{SO}(2))" + }, + { + "bbox": [ + 46, + 419, + 287, + 563 + ], + "type": "text", + "content": " is a random rotation on the plane, " + }, + { + "bbox": [ + 46, + 419, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\beta \\sim U([0, 0.5])" + }, + { + "bbox": [ + 46, + 419, + 287, + 563 + ], + "type": "text", + "content": " is a random eccentricity parameter, and " + }, + { + "bbox": [ + 46, + 419, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 419, + 287, + 563 + ], + "type": "text", + "content": " is the overall scale of the noise. To emulate keypoint detection on a multi-scale image pyramid, all generated points were split into 4 equal sets, with " + }, + { + "bbox": [ + 46, + 419, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 419, + 287, + 563 + ], + "type": "text", + "content": " within each set assigned to 1, 2, 4, and 8, respectively. Points and covariances are then passed to a Levenberg-Marquardt optimization of the relative pose using (35) as the residual function, with true pose serving as the initial estimate. Figure 6 demonstrates how the covariance-aware formulation (35) compares to the baseline (30). Covariance information is useful in the case when points are affected by noise of unequal scales." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 223, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 223, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 223, + 586 + ], + "type": "text", + "content": "4.5. Evaluation in Robust Estimators" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Finally we evaluate the methods in the context of RANSAC [7]. We implement a standard LO-RANSAC [18] estimator which uses the various error metrics both for model scoring (using MSAC scoring [24]) and refinement (using non-linear optimization on the inlier-set). For the experiment we consider image pairs from two Structure-from-Motion reconstructions (see Section 4.2). From each dataset we extract all image pairs which contain between 50 and 200 correct matches (according to the Structure-from-Motion ground-truth). For each pair we estimate the rela" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 253, + 553, + 369 + ], + "blocks": [ + { + "bbox": [ + 46, + 216, + 546, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 546, + 251 + ], + "type": "text", + "content": "Table 2. Evaluation of camera pose refinement. Table shows the mean and median errors (in degrees) of the rotation and translation after refinement. The data is split into three groups based on the camera field-of-view. Best error is highlighted in bold. Second-best is underlined. For low-distortion cameras the choice of error makes almost no difference on the quality of pose." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 253, + 553, + 369 + ], + "lines": [ + { + "bbox": [ + 306, + 253, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 553, + 369 + ], + "type": "table", + "html": "
GrossmunsterKirchenge
εR,t< 5°RTεR,t< 5°RT
ΕALG0.76175.039.60.31884.821.6
ΕCOS0.56181.550.40.25491.625.5
ΕSED1.16168.982.00.35586.444.0
ΕS0.92474.551.00.30089.225.3
ΕπSED0.56181.3211.20.25591.687.4
ΕTS0.55082.246.40.24891.924.2
", + "image_path": "5072a274e019e3faa37997e5583984eff60168290dd6bd162990bd000d6f1b4f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 375, + 546, + 431 + ], + "lines": [ + { + "bbox": [ + 304, + 375, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 546, + 431 + ], + "type": "text", + "content": "Table 3. Evaluation in robust estimators. Table shows the median pose error (maximum of rotation and translation error), percentage of pairs for which the estimation was successful (pose error less than 5 degrees), and runtime (in ms). Best value in the column is highlighted in bold. Second-best is underlined." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 445, + 546, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 546, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 546, + 531 + ], + "type": "text", + "content": "tive pose using RANSAC and compute the error in the pose, as well as percentage of successful runs. For each method we tuned the inlier threshold on a separately extracted validation set. For more details on the experiment setup see the supplementary material. The results are presented in Table 3. Using Tangent Sampson error consistently gives slightly better performance compared to other error metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 552, + 384, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 384, + 565 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 384, + 565 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 578, + 547, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 547, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 547, + 673 + ], + "type": "text", + "content": "In this paper we have presented a simple generalization to the classical Sampson error which allows for using arbitrary camera models. In experiments on a wide variety of camera and lens configurations, we have shown that the proposed error metric outperforms the alternative metrics that are used in practice. While the improvements are sometimes minor (e.g. for low-distortion pinhole cameras), they are consistent and come at essentially no additional cost." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 678, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 678, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 678, + 546, + 713 + ], + "type": "text", + "content": "Acknowledgments: This project received funding from the strategic research program ELLIIT. MT was supported by the ESOP scholarship from the ETH Foundation." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3377" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 92, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 286, + 113 + ], + "type": "text", + "content": "[1] Sameer Agarwal, Keir Mierle, and Others. Ceres solver. http://ceres-solver.org.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 115, + 286, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 286, + 148 + ], + "type": "text", + "content": "[2] João Pedro Barreto and Kostas Daniilidis. Fundamental matrix for cameras with radial distortion. In International Conference on Computer Vision (ICCV), 2005. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 149, + 286, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 149, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 149, + 286, + 194 + ], + "type": "text", + "content": "[3] Wojciech Chojnacki, Michael J. Brooks, Anton Van Den Hengel, and Darren Gawley. On the fitting of surfaces to data with covariances. IEEE Trans. Pattern Analysis and Machine Intelligence (PAMI), 2000. 3, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 196, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 196, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 196, + 286, + 228 + ], + "type": "text", + "content": "[4] Ondřej Chum, Tomáš Pajdla, and Peter Sturm. The geometric error for homographies. Computer Vision and Image Understanding (CVIU), 2005. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 230, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 286, + 262 + ], + "type": "text", + "content": "[5] David Claus and Andrew W Fitzgibbon. A rational function lens distortion model for general cameras. In Computer Vision and Pattern Recognition (CVPR), 2005. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 264, + 286, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 286, + 296 + ], + "type": "text", + "content": "[6] Asen L Dontchev and R Tyrrell Rockafellar. Implicit functions and solution mappings, volume 543. Springer, 2009. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 299, + 286, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 299, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 299, + 286, + 342 + ], + "type": "text", + "content": "[7] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 345, + 286, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 345, + 286, + 377 + ], + "spans": [ + { + "bbox": [ + 53, + 345, + 286, + 377 + ], + "type": "text", + "content": "[8] Andrew W Fitzgibbon. Simultaneous linear estimation of multiple view geometry and lens distortion. In Computer Vision and Pattern Recognition (CVPR), 2001. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 379, + 286, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 379, + 286, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 379, + 286, + 402 + ], + "type": "text", + "content": "[9] Gaël Guennebaud, Benoit Jacob, et al. Eigen v3. http://eigen.tuxfamily.org, 2010. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 403, + 286, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 403, + 286, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 286, + 446 + ], + "type": "text", + "content": "[10] Hyowon Ha, Michal Perdoch, Hatem Alismail, In So Kweon, and Yaser Sheikh. Deltille grids for geometric camera calibration. In International Conference on Computer Vision (ICCV), 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 449, + 286, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 286, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 286, + 480 + ], + "type": "text", + "content": "[11] Matthew Harker and Paul O'Leary. First order geometric distance (the myth of sampsonus). In *British Machine Vision Conference (BMVC)*, 2006. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 483, + 286, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 483, + 286, + 515 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 286, + 515 + ], + "type": "text", + "content": "[12] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 518, + 286, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 518, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 518, + 286, + 540 + ], + "type": "text", + "content": "[13] Richard I Hartley and Peter Sturm. Triangulation. Computer vision and image understanding, 68(2):146-157, 1997. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 542, + 286, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 542, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 542, + 286, + 574 + ], + "type": "text", + "content": "[14] Juho Kannala and Sami S Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. 2006. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 576, + 286, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 576, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 576, + 286, + 597 + ], + "type": "text", + "content": "[15] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1/2):81-93, 1938. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 600, + 286, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 286, + 632 + ], + "type": "text", + "content": "[16] Zuzana Kukelova and Viktor Larsson. Radial distortion triangulation. In Computer Vision and Pattern Recognition (CVPR), 2019. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "type": "text", + "content": "[17] Viktor Larsson, Nicolas Zobernig, Kasim Taskin, and Marc Pollefeys. Calibration-free structure-from-motion with calibrated radial trifocal tensors. In European Conference on Computer Vision (ECCV), 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "text", + "content": "[18] Karel Lebeda, Jiri Matas, and Ondrej Chum. Fixing the locally optimized ransac-full experimental evaluation. In British Machine Vision Conference (BMVC), 2012. 8" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 341 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[19] Seong Hun Lee and Javier Civera. Geometric interpretations of the normalized epipolar error. arXiv preprint arXiv:2008.01254, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[20] Peter Lindstrom. Triangulation made easy. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 1554-1561. IEEE, 2010. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 141, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 194 + ], + "type": "text", + "content": "[21] Yaroslava Lochman, Kostiantyn Liepieshov, Jianhui Chen, Michal Perdoch, Christopher Zach, and James Pritts. Babel-calib: A universal approach to calibrating central cameras. In International Conference on Computer Vision (ICCV), 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "text", + "content": "[22] Quan-Tuan Luong and Olivier D Faugeras. The fundamental matrix: Theory, algorithms, and stability analysis. International Journal of Computer Vision (IJCV), 1996. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "text", + "content": "[23] Paul D Sampson. Fitting conic sections to \"very scattered\" data: An iterative refinement of the bookstein algorithm. Computer graphics and image processing, 1982. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 264, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 306 + ], + "type": "text", + "content": "[24] Philip HS Torr and Andrew Zisserman. Mlesac: A new robust estimator with application to estimating image geometry. Computer Vision and Image Understanding (CVIU), 2000, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 309, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 341 + ], + "type": "text", + "content": "[25] Zhengyou Zhang. On the epipolar geometry between two images with lens distortion. In International Conference on Pattern Recognition (ICPR), 1996. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3378" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Task Agnostic Restoration of Natural Video Dynamics/e82ebeb9-3e53-42eb-9d7f-02ba103a6afc_content_list.json b/2023/Task Agnostic Restoration of Natural Video Dynamics/e82ebeb9-3e53-42eb-9d7f-02ba103a6afc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eb740a42bf84f710a04d98c8c4db0fdf9dc30fce --- /dev/null +++ b/2023/Task Agnostic Restoration of Natural Video Dynamics/e82ebeb9-3e53-42eb-9d7f-02ba103a6afc_content_list.json @@ -0,0 +1,1492 @@ +[ + { + "type": "text", + "text": "Task Agnostic Restoration of Natural Video Dynamics", + "text_level": 1, + "bbox": [ + 210, + 130, + 759, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Muhammad Kashif Ali Dongjin Kim Tae Hyun Kim* \nDepartment of Computer Science, Hanyang University, Seoul, Korea \n{kashifali, dongjinkim, taehyunkim}@hanyang.ac.kr", + "bbox": [ + 212, + 180, + 756, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In many video restoration/translation tasks, image processing operations are naively extended to the video domain by processing each frame independently, disregarding the temporal connection of the video frames. This disregard for the temporal connection often leads to severe temporal inconsistencies. State-Of-The-Art (SOTA) techniques that address these inconsistencies rely on the availability of unprocessed videos to implicitly siphon and utilize consistent video dynamics to restore the temporal consistency of frame-wise processed videos which often jeopardizes the translation effect. We propose a general framework for this task that learns to infer and utilize consistent motion dynamics from inconsistent videos to mitigate the temporal flicker while preserving the perceptual quality for both the temporally neighboring and relatively distant frames without requiring the raw videos at test time. The proposed framework produces SOTA results on two benchmark datasets, DAVIS and videoo.net, processed by numerous image processing applications. The code and the trained models are available at https://github.com/MKashifAli/TARONVD.", + "bbox": [ + 76, + 301, + 473, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 650, + 209, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video sharing social media platforms like Snapchat and TikTok have introduced the common populace to a plethora of computer vision applications such as Style Transfer [9], Colorization [43], Denoising [42], and Dehazing [38]. With this wide-scale integration of these classical computer vision applications in such platforms, various image processing operations are naively extended to videos due to scarcity of annotated video datasets and their computational complexity of video processing methodologies. This naive extension of static image processing methodologies to videos disregards the temporal connection of the video progression and introduces severe temporal flickering in the videos. This temporal flicker can appear for various reasons; for in", + "bbox": [ + 75, + 676, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "stance, these image processing methods can produce drastically different results for temporally neighboring frames due to slight changes in their global or local content statistics, or it could also happen due to the multi-modality of the application as there could exist a number of valid solutions for images with similar content as highlighted in [3, 41]. Therefore, the extension of these image-to-image translation tasks to the video domain is an active area of research in computer vision, and methods that can help extend various image processing applications with little to no knowledge of the operations used to create videos are quite useful.", + "bbox": [ + 496, + 268, + 893, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, there are several task-dependent temporal consistency correction approaches available such as [4, 23, 34], but due to the complex nature of these applications, only a handful of approaches have been proposed to tackle the problem of blind temporal consistency correction. Bonneel et al. [3] initiated the investigation of blind temporal consistency correction with gradient-domain minimization of per-frame processed video with the unprocessed video to minimize the warping error between the frames. Lai et al. [16] extended the formulation of [3] with the help of recurrent Convolutional Neural Networks (CNN) and introduced a perceptual penalty in their formulation to restrict the deviation of perceptual content of the restored video from the frame-wise processed video. Deep Video Prior (DVP) [19] extended Deep Image Prior (DIP) [35] to the temporal dimension and proposed to formulate enforcing temporal consistency by training a CNN to generate processed video from the unprocessed video without utilizing optical flow. All of these approaches rely on the availability of unprocessed videos for implicit extraction of consistent motion dynamics to use as a restoration guide.", + "bbox": [ + 496, + 441, + 895, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Please note that all the previously proposed approaches for this task define it with the help of unprocessed videos (both task-specific and task-agnostic). Having access to unprocessed videos at test time helps the models in siphoning and transferring consistent motion dynamics to the inconsistent videos to decrease temporal flicker. Despite its efficacy, this implicit definition limits the applicability of the previously proposed approaches to only the videos for which their unprocessed counterparts are available and also", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*: Corresponding author.", + "bbox": [ + 75, + 886, + 210, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13534", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/7e3e6137179f94d068e6ef5086a0cf3bf1bf337ae6c040a1908a68d9a62dd432.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Key featuresBonneel et al. [3]Lai et al. [16]DVP [19]Ours
Does not require raw videos×××✓
Explicit video dynamics design✓××✓
Faithful perceptual restoration×××✓
Sequential processing✓✓×✓
Long term consistency××✓✓
Sharpness in processed videos×✓×✓
High-frequency flicker removal✓×✓✓
Test time training××✓×
Runtime (per frame avg.)2.4635*0.21463.63650.2236
", + "bbox": [ + 81, + 88, + 470, + 186 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1: Key differences in available approaches. \\*\\*\" in front of Bonnel et al. [3]'s runtime highlights environment difference and is only presented for reference only. Please note that the runtime is averaged over the same video sequence for all the methods.", + "bbox": [ + 75, + 196, + 468, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "introduce an inherent bias towards the unprocessed videos which can compromise the translation effect and quality of the processed videos (as presented in Fig. $8\\sim 10$", + "bbox": [ + 75, + 300, + 468, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to overcome these limitations, we aim to learn and infer consistent motion representation solely from temporally inconsistent videos. Doing so does not only alleviate the need for raw videos at test time, but also mitigates the inherent bias towards the raw videos that is commonly present in the currently available approaches for this task. For the task of learning these consistent motion representation, we fine-tune conventional optical flow estimations networks like [11, 31, 33] along with the restorative part of the network in an end-to-end manner. The proposed model pipeline is illustrated in Fig. 1. The proposed model achieves state-of-the-art qualitative and quantitative results. The proposed formulation can also handle the resolution mismatch problem in processed and raw videos and makes it possible to extend the Single Image Super-Resolution [17] methods to Video Super-Resolution methods without any modification. The detailed description of our formulation is presented in 3. We summarize our contributions as follows:", + "bbox": [ + 75, + 345, + 468, + 617 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Identify and propose tailored solutions for various challenges of naïve extensions of image processing applications to videos.", + "- Learn to infer consistent motion representations from temporally inconsistent videos.", + "- Utilize the learned consistent motion representations to propose a general framework for task agnostic temporal consistency correction that produce SOTA results." + ], + "bbox": [ + 76, + 618, + 467, + 737 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 753, + 218, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The literature is divided into two main streams to generate visually appealing videos through image transformation models. The first stream tackles the task of mitigating temporal inconsistencies with the reformulation of the task at hand. Whereas, The second stream focuses on developing post-processing models that refine the frame-wise processed videos by penalizing the temporal deviation of the processed video from its unprocessed counterpart. The", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "second stream is further divided into two sub-streams: task-specific and task-agnostic. The following passages summarize the details of both of the above-mentioned streams, respectively.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reformulation stream (video-to-video translation): These approaches are generally termed video-to-video translation applications. Generally, these approaches either consider multiple frames as input and produce multiple output frames or generate a single frame from multiple input frames in the form of frame recurrence such as [5, 23]. There are also cases like video style transfer where content information is propagated to the next time step with the help of optical flow to initiate the optimization of the next frame. This frame-recurrent methodology has also been proven effective in applications like video super resolution [29]. Designing these approaches for each task and training them from scratch is a taxing task and data scarcity can make these formulations unfeasible. Generally, such models often do not adapt well to different tasks; therefore, methods that can restore the temporal consistency of multiple frame-wise processed videos are being actively investigated.", + "bbox": [ + 496, + 152, + 890, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Post-processing methods: Bonneel et al. [3] initiated the investigation of task agnostic temporal consistency correction using a gradient-based optimization strategy in which temporally consistent (unprocessed) videos were used as restoration guides for correcting the temporal inconsistency of frame-wise processed videos. Their formulation motivated various task-dependent and task-agnostic, and temporal consistency correction approaches with slight variations of their original objective formulation. Despite the simplicity of their formulation, there only exist a handful of approaches that address task-agnostic temporal consistency correction due to the difficult nature of the task. Yao et al. [40] proposed a key-frame strategy that accounted for the motion of different objects in those key-frames to handle occlusion as well as temporal consistency. Lai et al.[16] proposed the first deep learning approach for this task by employing ConvLSTM [30] and a perceptual loss [13]. Lei et al. [19] proposed an extension of Deep Image Prior [35] and demonstrated its capability to mitigate the temporal flicker of per-frame processed videos.", + "bbox": [ + 496, + 422, + 890, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Apart from these general frameworks, various task-specific approaches have also been proposed, such as [5, 23, 34]. Defining these task-specific approaches is relatively straightforward. These approaches define a temporal extension around a backbone method and penalize the deviation of generated content with the help of motion information from the unprocessed videos.", + "bbox": [ + 496, + 729, + 890, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a general formulation for this task that alleviates the requirement of unprocessed videos at test time by learning to infer a consistent motion representation from the temporally inconsistent videos. This inferred", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13535", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "motion representation is then processed with the content of temporally inconsistent videos in a two branched network that restores the natural dynamics of per-frame processed videos. Our proposed network also contains a recurrent bottleneck module that helps consistent propagation of content throughout the video sequence and produces SOTA qualitative and quantitative results by only taking into account the temporally inconsistent videos at test time.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to aid the readers, we provide the following table to highlight the key differences between the currently available approaches for this task.", + "bbox": [ + 75, + 212, + 468, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 273, + 250, + 290 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/01617768a5374400703e76106386427c83011b67d8f0b9dab46041c43f9df958.jpg", + "image_caption": [ + "Figure 1: Model Architecture. An illustration of the proposed model architecture. The proposed model consists of a two-branch architecture that separately processes motion and content. The combined motion and content features concatenated along the channel dimension and are then passed through a Recurrent Bottleneck Block (RBB) and a decoder to generate the restored frame." + ], + "image_footnote": [], + "bbox": [ + 78, + 313, + 468, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the details of the proposed method. Consider a raw (temporally consistent) video $\\{I_0, I_1, I_2, \\dots, I_n\\}$ with $n$ frames and its frame-wise processed (temporally inconsistent) video $\\{P_0, P_1, P_2, \\dots, P_n\\}$ acquired by an image processing function $h$ as $\\{h(I_0), h(I_1), h(I_2), \\dots, h(I_n)\\}$ . The goal of this task is to restore the temporal consistency of the temporally inconsistent video and to produce a temporally consistent version $\\{O_0, O_1, O_2, \\dots, O_n\\}$ . For the sake of simplicity, these notations are chosen to be in line with the notations used in past literature [3, 16, 19].", + "bbox": [ + 75, + 657, + 468, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed formulation draws inspiration from the disentanglement of a video into its base components: content and motion representations as presented in [22]. In their work [22], they investigated the controlled generation of short length videos by mixing motion and con", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/07563e715b481b3ceacaa6abe2fbf34ed3d98d29043097b40b254c8d232f2402.jpg", + "image_caption": [ + "Figure 2: Temporal flicker in optical flow. An illustration of the decomposition of optical flow generated from temporally inconsistent frames to consistent optical flow and inconsistent flicker." + ], + "image_footnote": [], + "bbox": [ + 504, + 87, + 890, + 138 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tent components from various videos. In this work, we explore a similar concept of controlled disentanglement and re-entanglement strategy to restore natural video dynamics of the frame-wise processed videos. More specifically, we argue that the previously proposed approaches for this tasks implicitly employ this disentanglement and re-entanglement in their formulation with the help of unprocessed video frames. Conventionally, optical flow is used as an inter-frame motion representations. The estimated optical flow can also be used to approximate the content of future frames with the help of past frames provided that there does not exist occlusion between the frames (as presented in Eq. 1).", + "bbox": [ + 496, + 247, + 890, + 444 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {t + 1} \\approx w \\left(I _ {t}, o f _ {t + 1 \\Rightarrow t}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 443, + 890, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This interdependence of optical flow and content can lead to inconsistent motion representation as optical flow estimation networks are not trained explicitly to distinguish temporally consistent content from temporal flicker. Due to this limitation, the evaluated optical flow estimated from temporally inconsistent frames is often compromised with temporal flicker (as presented in Fig. 2). This compromise of the motion information makes it challenging to separate the consistent component of optical flow from the additive flicker. Therefore, the currently available techniques for this task often rely on siphoning implicit motion representations from the raw videos and often suffer in cases where the content of both processed and raw videos deviate significantly. Another problem with employing these implicit representations is the bias introduced due to the lack of control in mixing these implicit representations to restore videos which often leads to the mitigation of translation effect. Therefore, if a mechanism can be developed that produces similar motion representations from both unprocessed and framewise processed videos, better and unbiased restorative effects than those produced by the previously proposed approaches can be achieved without relying on the availability of unprocessed videos.", + "bbox": [ + 496, + 463, + 892, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Please note that another observation that can be made from Fig. 2 (right most figure) is that a significant amount of flicker in frame-wise processed videos, generally occurs near the motion boundaries. This observation helps us in defining objective functions capable of localizing and targeting temporal deviations as presented in Sec. 4.1.", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13536", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome these problems and to learn and utilize consistent motion representation, we integrate an optical flow estimation network into the restoration network as illustrated in Fig. 1. A similar integration of a small-scale optical flow estimation network (SpyNet [26]) for task-oriented flow estimation is presented in [37] for the task video denoising for its extremely small size. We utilize a relatively larger network (PWC-Net [31]) in our method and provide an ablation study on different sized optical flow models in the accompanying supplementary text. We further extend their idea to a two pass mechanism with an added encoder to encourage generation of consistent dynamics. The optical flow network in the proposed methodology is finetuned in an end-to-end manner without any special supervision.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed model takes in three frames $O_{t-1}$ ( $P_{t-1}$ for the first time step), $P_t$ and $P_{t+1}$ as input. These three frames are propagated through both the content and motion branches. The proposed model consists of a UNet [27] like structure with multiple encoder streams and a single decoder stream. The decoder contains skip connections from the content stream to encourage better reconstruction. The bottleneck part of the model contains a recurrent bottleneck block to transmit and manipulate the information of the generated frames to temporally distant frames. Please note that the multi-branch encoder(s)-bottleneck-decoder architecture used for the proposed model is employed due to its proven efficacy in video generation, translation and restoration tasks as described in [4, 8, 14, 15, 16, 23, 32, 34]. The main difference in the proposed architecture comes from the design of our motion branch which allows us to present a standalone solution for this task. The motion branch comprises of two passes through a PwCNet [31] with shared weights followed by a conventional encoder-like architecture. An illustration of the proposed model is presented in Fig. 1. This two-pass strategy in the motion branch along with the attached encoder allows the proposed model in regressing consistent motion dynamics from temporally inconsistent frames. Further architectural details of the proposed network are presented in the accompanying supplemental. Figure 3 presents conventional optical flow along with the learned motion representation generated by the integrated finetuned PwC-Net [31] from temporally consistent (left) and inconsistent frames (right). Please note that the \"modified\" flow can differ starkly from the conventional flow and can only be interpreted effectively by the network it is trained with as described in [37]. Contrasting to the conventional optical flow modified flow estimated from both consistent and inconsistent videos is quite similar, which suggests that the network can effectively utilize the inferred consistent motion representation to restore natural temporal dynamics in the per-frame videos by only taking into account the temporally inconsistent videos at test time. Another benefit of this disentangled representation", + "bbox": [ + 75, + 311, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/afb0d404e443e7db0b81ac4e7feb6d31e177f021f3e9caa503fa41bb70d840b2.jpg", + "image_caption": [ + "Figure 3: Conventional and Modified flow like representation. Comparison of original and Modified flow evaluated from temporally consistent (left) and inconsistent (right) frames. The modified flow evaluated from both the temporally consistent and inconsistent frames is quite similar." + ], + "image_footnote": [], + "bbox": [ + 542, + 90, + 857, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is that it allows the users to further improve the level of restored temporal consistency with the help of an iterative arrangement; please refer to the supplemental for the details of the iterative arrangement and the results generated with such arrangement. The objective functions used for the end-to-end training of the proposed network are described in detail in the next section.", + "bbox": [ + 496, + 315, + 893, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Objective Functions", + "text_level": 1, + "bbox": [ + 498, + 434, + 692, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section provides the details of the learning objectives used in the training phase of the proposed network. These objectives can be classified into two categories, local neighborhood and long-term objectives. Both of these optimization categories are presented below.", + "bbox": [ + 496, + 459, + 892, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Local Neighborhood Losses", + "text_level": 1, + "bbox": [ + 498, + 545, + 748, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we aim to find a general solution for correcting the temporal consistency of frame-wise processed videos. Due to the diversity of the applications addressed, a simplistic spatial content matching reconstruction loss cannot justify the generation of temporally consistent frames, as the content can vary significantly in applications like style transfer [9] and image adjustment [38]. This discrepancy in the image space of processed and raw videos hinders a straightforward definition of a feasible loss function for this task. Due to the challenging nature of this task, we propose a flow gradient loss that provides a supervision signal for the reconstruction of motion boundaries (where most of the inconsistencies/flicker appear; as highlighted in Fig. 2) with the help of optical flow acquired through the unprocessed videos. We further extend this loss to only compare the gradients of the optical flow acquired through the synthesized and raw frames. Doing so reduces the redundancy present in the high dimensional optical flow which contains a redundant global component for this task which is only useful in applications like video stabilization [1] as described in [39]. The proposed flow gradient loss encapsulates the local spatio-temporal information necessary", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13537", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to mitigate temporal flicker near the motion boundaries by removing the unnecessary high-dimensional components from the optical flow. Similar information bottle-necking techniques have been discussed thoroughly in vision based understanding and manipulation literature [7, 24]. The proposed spatio-temporal loss is defined as follows:", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f g} = \\sum_ {t = 2} ^ {T} \\left\\| \\nabla \\left(o f \\left(O _ {t}, O _ {t - 1}\\right)\\right), \\nabla \\left(o f \\left(I _ {t}, I _ {t - 1}\\right)\\right) \\right\\| _ {1}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 191, + 468, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathcal{L}_{fg}$ and of represents flow gradient loss and the optical flow estimation network (FlowNet2.0 [11]) respectively. Please note that any optical flow estimation network (SpyNet [26], PWC-Net [31], RAFT [33], etc.) can be used in both the motion branch and the definition of the loss functions, we choose FlowNet2.0 [11] for its relatively lower mean end-point-error, as the estimated optical flow can vary significantly for small scale networks (comparison provided in the supplementary text). In equation 2, the $\\nabla$ operator denotes the spatial gradient of the estimated optical flow. A similar loss in optical flow domain has been proposed in [21] where the raw optical flow of the generated video was compared with that of the ground truth video. Although effective for the task of video inpainting, the raw optical flow matching approach fails in applications like style transfer in which the entire content of the generated video and the raw video differ greatly, i.e., contains excessive spatio-temporal variations which contribute to optical flow estimation. In our experiments, the models trained with raw optical flow performed poorly, whereas the model trained with flow gradients converged in a relatively shorter time and produced sharper frames.", + "bbox": [ + 75, + 238, + 472, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Please note that this optical flow-based loss by itself is not sufficient for faithfully correcting the temporal consistency, as there could exist multiple solutions for optical flow equation, e.g., it does not take into account the appearance of the synthesized and raw frames. In order to address this issue and to guide the network for the generation of perceptually credible frames, a non-local optical flow based loss in the image space is employed as presented below:", + "bbox": [ + 75, + 571, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\sum_ {t = 2} ^ {T} C _ {t} \\left\\| O _ {t} - \\hat {O} _ {t} \\right\\| _ {1}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 702, + 468, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This warping based reconstruction loss function has been adopted from the [4], as many variants of this objective function with slight modifications have been proven effective for enforcing temporal smoothness in video translation tasks in [3, 5, 16, 23, 34]. Here $T$ represents the total number of frames in a sequence and $\\hat{O}_t$ represents the warped frame $O_{t-1}$ (as described in Eq. 1). The working of this loss function can be summarized as the propagation of content from preceding frames to the subsequent frames in image domain to ensure that the generated frames contain similar", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "content to the previous frames and the inter-frame motion dynamics of the synthesized frames are similar to that of the temporally consistent counterpart. This non-local optical flow based loss function also takes into account the occlusion problem of the Eq. 1 by masking the occluded and de-occluded content as defined below:", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC _ {t} = \\exp \\left(- \\alpha \\left\\| I _ {t} - \\hat {I} _ {t} \\right\\| _ {2} ^ {2}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 194, + 892, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The value for $\\alpha = 50$ is chosen according to the previous works where a similar strategy is used for evaluating occlusion [34, 28].", + "bbox": [ + 496, + 228, + 890, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An additional conventional short-term perceptual similarity loss [13] was also introduced in the training phase to minimize the deviation of the synthesized frames from the original processed frames. This loss is defined as follows:", + "bbox": [ + 496, + 273, + 890, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\sum_ {t} \\sum_ {l} \\| \\phi_ {l} (O _ {t}) - \\phi_ {l} (P _ {t}) \\| _ {1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 348, + 892, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\phi_l(.)$ represents layers of a VGG-16 network till the layer $relu_{-4.3}$ (trained on the ImageNet dataset [6]).", + "bbox": [ + 496, + 381, + 890, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Temporal Constancy Loss", + "text_level": 1, + "bbox": [ + 498, + 422, + 735, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss defined in Eq. 3 adequately defines the content propagation from the previous frame to the current frame but lacks the ability to enforce it in temporally distant instances. For instance, consider the task of colorization where a car appears in a portion of the frames and is assigned some random color, and it disappears and re-appears in a later instance. The frame-wise colorization method can assign it a different color in each sequence, and the loss function defined in Eq. 3 will be sufficient to enforce the color constancy of the car in each interval where the car is visible. This, however, does not ensure the constancy of the color assigned to the car in both intervals, i.e., the car could be assigned a solid blue color in the first interval and a solid red color in the next interval which is quite prominent in methods like [16, 34]. To address this problem of temporally distant yet similar instances, we introduce a recurrent module (ConvLSTM [30]) in our proposed model and add explicit constraints that force the model to generate temporally consistent frames. The previously proposed approaches for this task simply try to minimize the deviation of current frames from the first frame of the sequence, which is inadequate for videos containing largely varying frames. This variation of the content can render this loss ineffective, and the inconsistencies can only be weakly penalized with the help of motion dynamics siphoned from the raw videos. Therefore, we introduce a bi-directional penalty that enforces the temporal smoothness in both backward (through the local neighborhood losses) and forward (through the proposed long-term loss) direction. This bi-directional optimization strategy is inspired from the ping", + "bbox": [ + 496, + 446, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13538", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pong loss dynamic as presented in [5]. The proposed long-term loss is termed as constancy loss and is presented below:", + "bbox": [ + 75, + 90, + 468, + 133 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n s t a n c y}} = \\sum_ {p = 1} ^ {T - 2} \\sum_ {t = p + 2} ^ {T} C _ {t \\Rightarrow p} \\| O _ {t} - w \\left(O _ {p}, o f \\left(I _ {t - 1}, I _ {p}\\right)\\right) \\| _ {1}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 157, + 467, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here subscript $p$ highlights that the preceding distant images are compared with all the subsequent frames. Unlike the simplistic long-term temporal constraints used in [16, 34] which only penalized the deviation of content from the first frame, this loss term takes into account the inter-frame flicker of all the frames in a sequence and provides a strong supervision signal capable of penalizing temporal discrepancies throughout the sequence. Due to this stronger temporal penalty accompanied with the proposed flow gradient loss, the proposed network learns to produce smoother videos despite having access to only the frame-wise processed videos at test time. The final loss for each training sequence is given by,", + "bbox": [ + 75, + 208, + 468, + 404 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {f g} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {3} \\mathcal {L} _ {p} + \\lambda_ {4} \\mathcal {L} _ {\\text {c o n s t a n c y}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 412, + 468, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The $\\lambda(s)$ in the equation above define the contribution of each loss in the optimization phase. The details of these hyperparameters are provided in the supplementary text.", + "bbox": [ + 75, + 439, + 468, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments and Results", + "text_level": 1, + "bbox": [ + 76, + 497, + 310, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During our experimentation phase, we tested various optical flow estimation networks such as [11, 26, 31]. Flownet2.0 [11] has approximately 162M trainable parameters, and it produced relatively better results in lesser training iterations as compared to the other two variants. Having a vast number of parameters compromises both the training and testing time. Therefore, we opted for a medium-sized optical flow estimation network (PWC-Net [31]) and a higher number of optimization iterations. The results produced by both variants are discussed in the supplementary text. We also experimented with an iterative arrangement of the trained models where restored videos are again subjected to temporal consistency correction models and evaluated that the proposed model architecture consistently reduces warp error with every iteration (as presented in Fig. 4). Generally, the frames subjected to a high number of consistency correction iterations lose fine details. The generated results by various restoration iterations are provided in the supplemental. Please note that the comparative results provided in this paper are generated through a single restoration iteration.", + "bbox": [ + 75, + 523, + 468, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The dataset for training the proposed model contains videos from the DAVIS dataset for video segmentation [25] and videos gathered from Videovo.net by [16]. The training dataset contains videos processed through a diverse range", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dac6bec8e5826318e9aaef88af745197141633817afe45eae5c7eb9b4b555461.jpg", + "image_caption": [ + "Figure 4: Temporal warp error vs. iterations. Temporal warp error consistently decreases with a higher number of restoration iterations." + ], + "image_footnote": [], + "bbox": [ + 527, + 80, + 864, + 202 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of applications such as, Artistic Style Transfer [9, 13], Colorization [36, 44], Image Enhancement [10], Intrinsic Image Decomposition [2], and Image-to-Image Translation tasks [20, 44, 46]. The qualitative and quantitative results are described below.", + "bbox": [ + 498, + 266, + 892, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Quantitative Results", + "text_level": 1, + "bbox": [ + 500, + 349, + 692, + 364 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b5292eca0715747a6fe4a9079fa865df8fbe741c3e58b25b2a7f27a767178d53.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskDAVISVIDEO
Bonneel[3]Lai[16]OursDVP[19]*Bonneel[3]Lai[16]OursDVP[19]*
WCT [20]/antimono0.00290.00310.00260.00220.00150.00210.00150.0017
WCT [20]/asheville0.00470.00590.0047-0.00320.00430.0039-
WCT [20]/candy0.00340.00470.0035-0.00200.00320.0025-
WCT [20]/feathers0.00400.00400.0027-0.00270.00300.0021-
WCT [20]/sketch0.00360.00290.0021-0.00250.00220.0017-
WCT [20]/wave0.00330.00350.00270.00260.00230.00260.00210.0019
Fast-neural-style [13]/princess0.00430.00630.00420.00380.00350.00530.00420.0046
Fast-neural-style [13]/ludnie0.00210.00230.00220.00220.00120.00140.00140.0022
DBL [10]/expertA0.00170.00110.00100.00150.00140.00070.00060.0015
DBL [10]/expertB0.00160.00100.00080.00140.00110.00060.00040.0014
IID [2]/reflectance0.00150.00080.00070.00120.00110.00060.00060.0006
IID [2]/shading0.00140.00080.00080.00130.00080.00040.00040.0017
CycleGAN [46]/photo2ukiyoe0.00240.00190.00150.00130.00170.00130.00100.0018
CycleGAN [46]/photo2vangogh0.00260.00260.00190.00140.00200.00200.00150.0020
Colorization [43]0.00160.00110.00080.00130.00100.00060.00040.0012
Colorization [36]0.00150.00090.00070.00130.00100.00050.00030.0012
Average0.00270.00270.00210.0018*0.00180.00190.00150.0018*
", + "bbox": [ + 503, + 383, + 893, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Quantitative comparison of Temporal Warping Error. The comparison presented in this table is limited to the results produced by [3, 16] and the proposed method. Lower warp error suggests better temporal consistency. *** Indicates partially averaged results.", + "bbox": [ + 498, + 527, + 893, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative results are evaluated on the basis of temporal warp error. Tab. 2 shows the quantitative result produced by the methods proposed in [3, 16, 19] and our method. Please note that there are over 830 video sequences in the evaluation datasets and over 115,000 frames. Due to the extremely tedious and resource-intensive nature of DVP [19], only a portion of the evaluation datasets could be processed and the evaluated metrics in the DVP [19] column are averaged over only the processed videos for each task. In order to have a fair comparison, the metrics evaluated on the videos processed with DVP [19] are excluded from the large-scale quantitative comparison and are only presented for reference only. Furthermore, it is also noteworthy that in their paper [19], the authors do not evaluate the results on the entirety of the dataset and sample tasks for comparative results and report that their warp error is comparable to that of Bonneel et al. [3].", + "bbox": [ + 496, + 613, + 892, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The lower temporal warp error represents better temporal consistency. The lowest temporal warp error is highlighted", + "bbox": [ + 498, + 869, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13539", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in red, and the second best is highlighted in blue. It is worth mentioning that the temporal warp error does not take into account the perceptual quality of the produced frames and assigns a lower value to blurry videos (as produced by [19]). The quality of the results produced by each of the discussed methods can be verified through the accompanying supplementary video. It is evident from Tab. 2 that the proposed model produces state-of-the-art consistency results. Please note that, unlike the proposed method, all of the compared methods require the availability of raw/unprocessed videos for restoring temporal dynamics at both the training and test time.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further verify the efficacy of the proposed model through a specialized metric (as presented in [12]) for evaluating motion smoothness in videos, on the portion of evaluation datasets processed through the DVP [19] and present our findings below.", + "bbox": [ + 76, + 273, + 470, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is worth mentioning that some attempts to judge the perceptual quality through metrics like mean PSNR (performance degradation [19]) have been proposed for this task. We intentionally disregarded this metric and argue that this metric might be helpful to evaluate the performance of successive iterations of their prior-based method where each iteration leads to progressively better results, but it can be a bit misleading in judging the overall perceptual quality of restored videos for this task. A similar phenomenon can be observed with other local perceptual metrics such as LPIPS [45] as it is impossible to quantize a meaningful deviation of perceptual distance or PSNR from per-frame processed videos due to the lack of ground truth data.", + "bbox": [ + 75, + 349, + 468, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to be consistent with the prior works, we nonetheless present the LPIPS [45] results in 3. The proposed method performs competitively with the method proposed by [16] while achieving better temporal consistency. It is noteworthy that each of the discussed metrics does not take into account the aspects considered by the other metric and only provides limited insights about the results. Therefore, to properly evaluate the models on both perceptual quality and temporal consistency, thorough user studies are essential and are used as the primary metric for the evaluation of this task. We conducted two separate user studies to properly evaluate all of the discussed methods and present our findings in the next sub-section.", + "bbox": [ + 75, + 547, + 470, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1.1 Smoothness Metric (SM)", + "text_level": 1, + "bbox": [ + 76, + 768, + 305, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Due to the unavailability of ground truth videos, the quantitative metric results presented (both in the main paper and this supplemental) can only give us an approximate insight towards the nature of the restored videos. In order to properly evaluate our work, we present further large-scale quantitative results generated on the portion of the evaluation datasets processed through DVP [19] in this section. Fig. 5", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f13d647bdf06efb97bd10e615f84d66fff0271b94bca4ea6f02e59eaead9c425.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskDAVISVIDEO
Bonneel[3]Lai[16]OursDVP[19]*Bonneel[3]Lai[16]OursDVP[19]*
WCT [20]/antimony0.21660.04310.04850.33940.21800.04510.04470.2992
WCT [20]/asheville0.20560.03730.0449-0.26800.05140.0485-
WCT [20]/candy0.25880.04350.0497-0.21750.05340.0514-
WCT [20]/feathers0.20960.04670.0508-0.15190.05900.0540-
WCT [20]/sketch0.14510.04990.0511-0.18710.05170.0496-
WCT [20]/wave0.17770.04440.04850.34230.25830.07490.07930.3451
Fast-neural-style [13]/princess0.22380.06460.07440.47820.16100.05550.06250.2566
Fast-neural-style [13]/tudnie0.14730.05030.06090.36600.08000.05140.06040.1487
DBL [10]/expertA0.07950.03790.04950.14530.07420.05040.05840.1349
DBL [10]/expertB0.05300.03590.04810.13610.13590.06390.05780.1947
IID [2]/reflectance0.10110.04930.05360.13890.07830.06590.07310.2635
IID [2]/ shading0.06460.05270.06430.15580.11910.05060.05920.2302
CycleGAN [46]/photo2ukiyoe0.09080.03940.05020.23510.14620.05050.05750.2571
CycleGAN [46]/photo2vangogh0.12300.03900.04940.25650.12240.05190.05510.1701
Colorization [43]0.06160.04020.04680.15550.08670.04950.05660.1441
Colorization [36]0.03290.03730.04750.14740.13170.00810.02200.1558
Average0.13690.04450.05240.2414*0.15230.05210.05560.2167*
", + "bbox": [ + 504, + 88, + 893, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Quantitative comparison of LPIPS [45]. The comparison presented in this table is limited to the results produced by [3, 16] and the proposed method. Lower LPIPS suggests better perceptual similarity. \\*\\*\" Indicates partially averaged results.", + "bbox": [ + 496, + 231, + 893, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a42da043b741f38ce395a808bcfd27963c2b397d954d830bde866d267845bdb5.jpg", + "image_caption": [ + "Figure 5: Motion Smoothness Results. This figure presents the quantitative results acquired through the motion smoothness metric as described above. The lower motion smoothness represents smoother transitions in the videos." + ], + "image_footnote": [], + "bbox": [ + 506, + 311, + 879, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "presents the temporal smoothness results generated through the motion smoothness metric as presented in [12]. This metric tries to quantify the smoothness of videos over various window sizes through differential optical flow as follows:", + "bbox": [ + 496, + 508, + 893, + 583 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D = \\left(F _ {I _ {1} \\rightarrow I _ {2}} - F _ {I _ {0} \\rightarrow I _ {1}}\\right) - \\left(F _ {O _ {1} \\rightarrow O _ {2}} - F _ {O _ {0} \\rightarrow O _ {1}}\\right), \\\\ \\mathrm {M} (s) = \\log \\left(\\sum_ {d \\in D} \\mathbf {1} _ {[ s, s + 1)} \\left(\\| \\boldsymbol {d} \\| _ {2}\\right)\\right) - \\log (| D |), \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 589, + 890, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, $F_{a\\rightarrow b}$ denotes optical flow estimated from frame $a$ to frame $b$ . The preceding $I$ and $O$ refers to the raw/unprocessed and restored frames respectively. In Eq. 8, $d$ denotes each of the $[\\nabla u,\\nabla v]$ vectors in the differential optical flow and $s$ represents the various window sizes for evaluation and $|D|$ denotes the dimensionality of the differential optical flow. $\\mathbf{1}_S(d)$ is the masking factor which is equal to 1 if $d$ belongs to the set $S$ . The resultant $\\mathrm{M}(s)$ gives a measure of temporal smoothness over 3 consecutive frames. Fig. 5 presents the average results on the portion of evaluation datasets processed through [3, 16, 19] and the proposed model. A lower score suggests better smoothness. The proposed model again achieves SOTA results in terms of motion smoothness even over bigger window sizes as well.", + "bbox": [ + 496, + 657, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Please note that all the presented metrics are subjective", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13540", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in nature and do not account for the overall quality and consistency of the results due to the lack of ground truth; therefore, user studies should be considered as the main criteria for judging the results of all the proposed approaches for this task. We discuss the findings of the conducted user studies below.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1.2 User Study", + "text_level": 1, + "bbox": [ + 76, + 199, + 209, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a3c168b6d3b2e579a30a4930cf5c16755842618c99ed5f577aacee49081251e5.jpg", + "image_caption": [ + "Figure 6: User Study. This donut chart highlights the user preferences collected through the user study. A majority of the participants preferred the results produced by the proposed method." + ], + "image_footnote": [], + "bbox": [ + 187, + 217, + 361, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted two separate user studies; a comprehensive and a factorized user study, to properly evaluate the performance of the proposed model. The first user study consisted of 36 participants with 150 different scenes processed through each of the proposed models. Each participant was asked to judge 5 videos processed through each of the proposed methods for this task. The users were shown a randomly sampled scene from the video pool processed through [3, 16, 19] and this work, simultaneously. The participants were instructed to judge the videos on the basis of naturalness, quality, consistency, content, and style preservation. Figure 6 presents the findings of the first (comprehensive) user study. On average, $41\\%$ of the users preferred the videos restored by the proposed method. Furthermore, the results generated by the method proposed in Lai et al. [16] also received a significant preference in the user study but it is worth noting that all the previously proposed approaches require the unprocessed counterparts of the frame-wise processed videos at test time to restore the temporal consistency, whereas the proposed model restores the temporal consistency by only taking into account the inconsistent video at the test time.", + "bbox": [ + 75, + 450, + 470, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1.3 Factorized User Study", + "text_level": 1, + "bbox": [ + 76, + 801, + 285, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Apart from the comprehensive user study, we also conducted a secondary in-depth user study to evaluate the reasoning of participants and their opinion on their selections of videos processed with all the methods on challenging scenarios.", + "bbox": [ + 75, + 825, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this user study, each user was instructed to judge 15 randomly selected videos (5 videos from each of the applications mentioned above) processed by each of the proposed models for this task. Please note that in this user study users were also shown the per-frame processed video in an extended window to judge the videos on content and style preservation as well. Apart from recording their preference, the participants were also instructed to record the reasoning for their selection. $\\sim 86\\%$ of the participants remarked that they chose the videos based on lesser flickering and better quality preservation. The results of the second user study for each of the selected tasks are provided in Fig. 7. On average, $\\sim 47\\%$ of the users preferred the videos processed through the proposed model for these extremely challenging tasks.", + "bbox": [ + 496, + 90, + 890, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 330, + 681, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figures $8\\sim 10$ present some of the comparative qualitative results produced by the methods proposed in [3, 16, 19] and the proposed model. The results produced by [3] fail to retain the perceptual quality of video sequences where occlusion and deocclusion occur. Whereas the frames processed by DVP [19] are significantly blurry and lose the translation effect. In contrast to the comparison with DVP [19] and Boneel et al. [3], the comparison with the results produced through the method proposed by Lai et al. [16] can be a bit tricky. In order to aid the readers, we provide an epipolar comparison between the frames restored through their method and ours in Fig. 10. Due to the implicit definition of motion dynamics in their method [16], their method fails to faithfully retain the perceptual quality in the restored frames and often flips the colors/appearance of objects in the restored video according to the raw video. This bias towards the raw videos is quite prominent in methods like DVP [19] and Lai et al. [16]. Contrasting to their formulation, the explicit design of motion dynamics in the proposed method mitigates the need for raw videos at test time which consequently eliminates this bias and enables the proposed method to faithfully restore videos. Further qualitative comparisons and higher quality results, along with the results on super resolution [17], inpainting [18], and multiple restoration iterations are presented in the supplemental.", + "bbox": [ + 496, + 354, + 890, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Future work", + "text_level": 1, + "bbox": [ + 500, + 758, + 767, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed model has difficulties in restoring videos with very low frame rates and in the case of style-transferred videos, some finer inconsistencies like finer brush strokes are lost, which makes the resulting video seem a little dull (which is to be expected from all the proposed approaches for this task). A good direction for future work in this field can be the direct denoising of conventional optical flow with deterministic warp operators. The developed consis", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13541", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3909694f263ba7677bb389caab3c5d9704199e1305ea1958c1de3cae6c812a7d.jpg", + "image_caption": [ + "Figure 7: Factorized user study. The user preference scores for the tasks of style transfer, CycleGAN, and intrinsic image decomposition (left to right respectively). The majority of the users preferred the videos processed through the proposed method in each category of tasks." + ], + "image_footnote": [], + "bbox": [ + 158, + 85, + 369, + 304 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fd53144caa3953fbb0d72bac85fedc98a92d98d3c407f6e8bf75569569fa54b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 85, + 589, + 305 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/246c88f8bfb11b50682b25f88bd19e08e7558a745914c72db11177b55eb5ce9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 85, + 812, + 305 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c1f3e3962454ad5c55d7908768400076a39caa285da3a28e584f25af40eedaa9.jpg", + "image_caption": [ + "Figure 8: Comparison with Bonneel et al. [3] The method proposed by Bonneel et al. [3] has difficulty in handling occlusion and long term consistency and produce temporal artifacts." + ], + "image_footnote": [], + "bbox": [ + 81, + 378, + 467, + 503 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/59bb2b89a8456beded0c4b186847ea19fa7d594e0acc4113f2438f49b02aa5b6.jpg", + "image_caption": [ + "Figure 9: Comparison with DVP [19] DVP [19] produces blurry frames and compromises the translation effect. This phenomenon is prominent in videos where new content is generated e.g. style transfer [13], CycleGAN [46]." + ], + "image_footnote": [], + "bbox": [ + 81, + 580, + 467, + 708 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "tent flow models can also be used for a number of video processing applications like denoising and dehazing.", + "bbox": [ + 75, + 794, + 468, + 825 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 832, + 194, + 848 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we present a task-agnostic temporal consistency correction framework that restores natural video dynamics by inferring and utilizing consistent motion repre", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/aaf535c0e9477358f4a24cbca6f1b92554edd7f41226cd5b05b3ca2006c135ef.jpg", + "image_caption": [ + "Figure 10: Comparison with Lai et al. [16] This figure produces epipolar temporal comparison of videos restored by Lai et al. [16] and the proposed method. The method proposed in [16], due to its dependence on raw videos and implicit siphoning/mixing of motion dynamics from the raw video introduces bias towards the raw videos." + ], + "image_footnote": [], + "bbox": [ + 521, + 378, + 870, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "sentations from temporally inconsistent videos. The proposed strategy mitigates the need for siphoning video dynamics from the unprocessed videos at test time consequently mitigating the compromise in translation effects that is prominent in the currently proposed approaches for this task and expands the scope of its employability to a wider set of applications. Through extensive experimentation and user studies, we demonstrate that despite the limited access to data at test-time the proposed method compares favorably to the pre-existing methods available for this task and achieves SOTA performance.", + "bbox": [ + 496, + 571, + 890, + 739 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 744, + 668, + 763 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2022-0-00156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D meta-verse transformation) and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06.", + "bbox": [ + 496, + 770, + 890, + 891 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13542", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Muhammad Kashif Ali, Sangjoon Yu, and Tae Hyun Kim. Deep motion blind video stabilization. BMVC, 2021.", + "[2] Sean Bell, Kavita Bala, and Noah Snavely. Intrinsic images in the wild. ACM TOG, 33(4):1-12, 2014.", + "[3] Nicolas Bonneel, James Tompkin, Kalyan Sunkavalli, Deqing Sun, Sylvain Paris, and Hanspeter Pfister. Blind video temporal consistency. ACM TOG, 34(6):1-9, 2015.", + "[4] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent online video style transfer. In ICCV, pages 1105-1114, 2017.", + "[5] Mengyu Chu, You Xie, Laura Leal-Taixe, and Nils Thuerey. Temporally coherent gans for video super-resolution (tecogan). arXiv preprint arXiv:1811.09393, 1(2):3, 2018.", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009.", + "[7] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Counting out time: Class agnostic video repetition counting in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10387-10396, 2020.", + "[8] Wei Gao, Yijun Li, Yihang Yin, and Ming-Hsuan Yang. Fast video multi-style transfer. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3211-3219, 2020.", + "[9] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015.", + "[10] Michael Gharbi, Jiawen Chen, Jonathan T Barron, Samuel W Hasinoff, and Frédo Durand. Deep bilateral learning for real-time image enhancement. ACM TOG, 36(4):1-12, 2017.", + "[11] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, pages 2462-2470, 2017.", + "[12] Meiguang Jin, Zhe Hu, and Paolo Favaro. Learning to extract flawless slow motion from blurry videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8112-8121, 2019.", + "[13] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016.", + "[14] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep blind video decaptioning by temporal aggregation and recurrence. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4258–4267, 2019.", + "[15] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Recurrent temporal aggregation framework for deep video inpainting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42:1038–1052, 2019.", + "[16] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018.", + "[17] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken," + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, pages 4681-4690, 2017.", + "[18] Eunhye Lee, Jeongmu Kim, Jisu Kim, and Tae Hyun Kim. Restore from restored: Single-image inpainting. arXiv preprint arXiv:2102.08078, 2021.", + "[19] Chenyang Lei, Yazhou Xing, and Qifeng Chen. Blind video temporal consistency via deep video prior. NeurIPS, 2020.", + "[20] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. NeurIPS, 30, 2017.", + "[21] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. arXiv preprint arXiv:2204.02663, 2022.", + "[22] Xunyu Lin, Victor Campos, Xavier Giro-i Nieto, Jordi Torres, and Cristian Canton Ferrer. Disentangling motion, foreground and background features in videos. arXiv preprint arXiv:1707.04092, 2017.", + "[23] Yihao Liu, Hengyuan Zhao, Kelvin CK Chan, Xintao Wang, Chen Change Loy, Yu Qiao, and Chao Dong. Temporally consistent video colorization with deep feature propagation and self-regularization learning. arXiv preprint arXiv:2110.04562, 2021.", + "[24] Cheng Ma, Yongming Rao, Yean Cheng, Ce Chen, Jiwen Lu, and Jie Zhou. Structure-preserving super resolution with gradient guidance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7769-7778, 2020.", + "[25] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016.", + "[26] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In CVPR, pages 4161-4170, 2017.", + "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015.", + "[28] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In $GCPR$ , pages 26-36. Springer, 2016.", + "[29] Mehdi SM Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super-resolution. In $CVPR$ , pages 6626-6634, 2018.", + "[30] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. NeurIPS, 28, 2015.", + "[31] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In CVPR, pages 8934-8943, 2018.", + "[32] Ximeng Sun, Huijuan Xu, and Kate Saenko. Twostreamvan: Improving motion modeling in video generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2744-2753, 2020." + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13543", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419. Springer, 2020.", + "[34] Hugo Thimonier, Julien Despois, Robin Kips, and Matthieu Perrot. Learning long term style preserving blind video temporal consistency. In ICME, pages 1-6. IEEE, 2021.", + "[35] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In CVPR, pages 9446-9454, 2018.", + "[36] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In ECCV, pages 836-850. Springer, 2014.", + "[37] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127(8):1106-1125, 2019.", + "[38] Zhicheng Yan, Hao Zhang, Baoyuan Wang, Sylvain Paris, and Yizhou Yu. Automatic photo adjustment using deep neural networks. ACM TOG, 35(2):1-15, 2016.", + "[39] Yuchen Yang, Ye Xiang, Shuaicheng Liu, Lifang Wu, Boxuan Zhao, and Bing Zeng. Glm-net: Global and local motion estimation via task-oriented encoder-decoder structure. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4211-4219, 2021.", + "[40] Chun-Han Yao, Chia-Yang Chang, and Shao-Yi Chien. Occlusion-aware video temporal consistency. In ACMMM, pages 777-785, 2017.", + "[41] Bo Zhang, Mingming He, Jing Liao, Pedro V Sander, Lu Yuan, Amine Bermak, and Dong Chen. Deep exemplar-based video colorization. In CVPR, pages 8052-8061, 2019.", + "[42] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017.", + "[43] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, pages 649-666. Springer, 2016.", + "[44] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, pages 649-666. Springer, 2016.", + "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018.", + "[46] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2223-2232, 2017." + ], + "bbox": [ + 78, + 90, + 467, + 757 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13544", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Task Agnostic Restoration of Natural Video Dynamics/e82ebeb9-3e53-42eb-9d7f-02ba103a6afc_model.json b/2023/Task Agnostic Restoration of Natural Video Dynamics/e82ebeb9-3e53-42eb-9d7f-02ba103a6afc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cda90b60a86317bd81d86848463401ee1d0c1bb9 --- /dev/null +++ b/2023/Task Agnostic Restoration of Natural Video Dynamics/e82ebeb9-3e53-42eb-9d7f-02ba103a6afc_model.json @@ -0,0 +1,2037 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.212, + 0.131, + 0.76, + 0.154 + ], + "angle": 0, + "content": "Task Agnostic Restoration of Natural Video Dynamics" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.757, + 0.234 + ], + "angle": 0, + "content": "Muhammad Kashif Ali Dongjin Kim Tae Hyun Kim* \nDepartment of Computer Science, Hanyang University, Seoul, Korea \n{kashifali, dongjinkim, taehyunkim}@hanyang.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.474, + 0.619 + ], + "angle": 0, + "content": "In many video restoration/translation tasks, image processing operations are naively extended to the video domain by processing each frame independently, disregarding the temporal connection of the video frames. This disregard for the temporal connection often leads to severe temporal inconsistencies. State-Of-The-Art (SOTA) techniques that address these inconsistencies rely on the availability of unprocessed videos to implicitly siphon and utilize consistent video dynamics to restore the temporal consistency of frame-wise processed videos which often jeopardizes the translation effect. We propose a general framework for this task that learns to infer and utilize consistent motion dynamics from inconsistent videos to mitigate the temporal flicker while preserving the perceptual quality for both the temporally neighboring and relatively distant frames without requiring the raw videos at test time. The proposed framework produces SOTA results on two benchmark datasets, DAVIS and videoo.net, processed by numerous image processing applications. The code and the trained models are available at https://github.com/MKashifAli/TARONVD." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.651, + 0.21, + 0.667 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Video sharing social media platforms like Snapchat and TikTok have introduced the common populace to a plethora of computer vision applications such as Style Transfer [9], Colorization [43], Denoising [42], and Dehazing [38]. With this wide-scale integration of these classical computer vision applications in such platforms, various image processing operations are naively extended to videos due to scarcity of annotated video datasets and their computational complexity of video processing methodologies. This naive extension of static image processing methodologies to videos disregards the temporal connection of the video progression and introduces severe temporal flickering in the videos. This temporal flicker can appear for various reasons; for in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.438 + ], + "angle": 0, + "content": "stance, these image processing methods can produce drastically different results for temporally neighboring frames due to slight changes in their global or local content statistics, or it could also happen due to the multi-modality of the application as there could exist a number of valid solutions for images with similar content as highlighted in [3, 41]. Therefore, the extension of these image-to-image translation tasks to the video domain is an active area of research in computer vision, and methods that can help extend various image processing applications with little to no knowledge of the operations used to create videos are quite useful." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.896, + 0.761 + ], + "angle": 0, + "content": "Currently, there are several task-dependent temporal consistency correction approaches available such as [4, 23, 34], but due to the complex nature of these applications, only a handful of approaches have been proposed to tackle the problem of blind temporal consistency correction. Bonneel et al. [3] initiated the investigation of blind temporal consistency correction with gradient-domain minimization of per-frame processed video with the unprocessed video to minimize the warping error between the frames. Lai et al. [16] extended the formulation of [3] with the help of recurrent Convolutional Neural Networks (CNN) and introduced a perceptual penalty in their formulation to restrict the deviation of perceptual content of the restored video from the frame-wise processed video. Deep Video Prior (DVP) [19] extended Deep Image Prior (DIP) [35] to the temporal dimension and proposed to formulate enforcing temporal consistency by training a CNN to generate processed video from the unprocessed video without utilizing optical flow. All of these approaches rely on the availability of unprocessed videos for implicit extraction of consistent motion dynamics to use as a restoration guide." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Please note that all the previously proposed approaches for this task define it with the help of unprocessed videos (both task-specific and task-agnostic). Having access to unprocessed videos at test time helps the models in siphoning and transferring consistent motion dynamics to the inconsistent videos to decrease temporal flicker. Despite its efficacy, this implicit definition limits the applicability of the previously proposed approaches to only the videos for which their unprocessed counterparts are available and also" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.887, + 0.212, + 0.9 + ], + "angle": 0, + "content": "*: Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13534" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.472, + 0.187 + ], + "angle": 0, + "content": "
Key featuresBonneel et al. [3]Lai et al. [16]DVP [19]Ours
Does not require raw videos×××✓
Explicit video dynamics design✓××✓
Faithful perceptual restoration×××✓
Sequential processing✓✓×✓
Long term consistency××✓✓
Sharpness in processed videos×✓×✓
High-frequency flicker removal✓×✓✓
Test time training××✓×
Runtime (per frame avg.)2.4635*0.21463.63650.2236
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.197, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Table 1: Key differences in available approaches. \\*\\*\" in front of Bonnel et al. [3]'s runtime highlights environment difference and is only presented for reference only. Please note that the runtime is averaged over the same video sequence for all the methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.469, + 0.346 + ], + "angle": 0, + "content": "introduce an inherent bias towards the unprocessed videos which can compromise the translation effect and quality of the processed videos (as presented in Fig. \\(8\\sim 10\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.469, + 0.618 + ], + "angle": 0, + "content": "In order to overcome these limitations, we aim to learn and infer consistent motion representation solely from temporally inconsistent videos. Doing so does not only alleviate the need for raw videos at test time, but also mitigates the inherent bias towards the raw videos that is commonly present in the currently available approaches for this task. For the task of learning these consistent motion representation, we fine-tune conventional optical flow estimations networks like [11, 31, 33] along with the restorative part of the network in an end-to-end manner. The proposed model pipeline is illustrated in Fig. 1. The proposed model achieves state-of-the-art qualitative and quantitative results. The proposed formulation can also handle the resolution mismatch problem in processed and raw videos and makes it possible to extend the Single Image Super-Resolution [17] methods to Video Super-Resolution methods without any modification. The detailed description of our formulation is presented in 3. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.619, + 0.468, + 0.661 + ], + "angle": 0, + "content": "- Identify and propose tailored solutions for various challenges of naïve extensions of image processing applications to videos." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.664, + 0.468, + 0.693 + ], + "angle": 0, + "content": "- Learn to infer consistent motion representations from temporally inconsistent videos." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.694, + 0.468, + 0.738 + ], + "angle": 0, + "content": "- Utilize the learned consistent motion representations to propose a general framework for task agnostic temporal consistency correction that produce SOTA results." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.619, + 0.468, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.754, + 0.219, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The literature is divided into two main streams to generate visually appealing videos through image transformation models. The first stream tackles the task of mitigating temporal inconsistencies with the reformulation of the task at hand. Whereas, The second stream focuses on developing post-processing models that refine the frame-wise processed videos by penalizing the temporal deviation of the processed video from its unprocessed counterpart. The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "second stream is further divided into two sub-streams: task-specific and task-agnostic. The following passages summarize the details of both of the above-mentioned streams, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Reformulation stream (video-to-video translation): These approaches are generally termed video-to-video translation applications. Generally, these approaches either consider multiple frames as input and produce multiple output frames or generate a single frame from multiple input frames in the form of frame recurrence such as [5, 23]. There are also cases like video style transfer where content information is propagated to the next time step with the help of optical flow to initiate the optimization of the next frame. This frame-recurrent methodology has also been proven effective in applications like video super resolution [29]. Designing these approaches for each task and training them from scratch is a taxing task and data scarcity can make these formulations unfeasible. Generally, such models often do not adapt well to different tasks; therefore, methods that can restore the temporal consistency of multiple frame-wise processed videos are being actively investigated." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.726 + ], + "angle": 0, + "content": "Post-processing methods: Bonneel et al. [3] initiated the investigation of task agnostic temporal consistency correction using a gradient-based optimization strategy in which temporally consistent (unprocessed) videos were used as restoration guides for correcting the temporal inconsistency of frame-wise processed videos. Their formulation motivated various task-dependent and task-agnostic, and temporal consistency correction approaches with slight variations of their original objective formulation. Despite the simplicity of their formulation, there only exist a handful of approaches that address task-agnostic temporal consistency correction due to the difficult nature of the task. Yao et al. [40] proposed a key-frame strategy that accounted for the motion of different objects in those key-frames to handle occlusion as well as temporal consistency. Lai et al.[16] proposed the first deep learning approach for this task by employing ConvLSTM [30] and a perceptual loss [13]. Lei et al. [19] proposed an extension of Deep Image Prior [35] and demonstrated its capability to mitigate the temporal flicker of per-frame processed videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.892, + 0.836 + ], + "angle": 0, + "content": "Apart from these general frameworks, various task-specific approaches have also been proposed, such as [5, 23, 34]. Defining these task-specific approaches is relatively straightforward. These approaches define a temporal extension around a backbone method and penalize the deviation of generated content with the help of motion information from the unprocessed videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a general formulation for this task that alleviates the requirement of unprocessed videos at test time by learning to infer a consistent motion representation from the temporally inconsistent videos. This inferred" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13535" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "motion representation is then processed with the content of temporally inconsistent videos in a two branched network that restores the natural dynamics of per-frame processed videos. Our proposed network also contains a recurrent bottleneck module that helps consistent propagation of content throughout the video sequence and produces SOTA qualitative and quantitative results by only taking into account the temporally inconsistent videos at test time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.47, + 0.259 + ], + "angle": 0, + "content": "In order to aid the readers, we provide the following table to highlight the key differences between the currently available approaches for this task." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.274, + 0.251, + 0.291 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.314, + 0.47, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.643 + ], + "angle": 0, + "content": "Figure 1: Model Architecture. An illustration of the proposed model architecture. The proposed model consists of a two-branch architecture that separately processes motion and content. The combined motion and content features concatenated along the channel dimension and are then passed through a Recurrent Bottleneck Block (RBB) and a decoder to generate the restored frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.825 + ], + "angle": 0, + "content": "In this section, we describe the details of the proposed method. Consider a raw (temporally consistent) video \\(\\{I_0, I_1, I_2, \\dots, I_n\\}\\) with \\(n\\) frames and its frame-wise processed (temporally inconsistent) video \\(\\{P_0, P_1, P_2, \\dots, P_n\\}\\) acquired by an image processing function \\(h\\) as \\(\\{h(I_0), h(I_1), h(I_2), \\dots, h(I_n)\\}\\). The goal of this task is to restore the temporal consistency of the temporally inconsistent video and to produce a temporally consistent version \\(\\{O_0, O_1, O_2, \\dots, O_n\\}\\). For the sake of simplicity, these notations are chosen to be in line with the notations used in past literature [3, 16, 19]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The proposed formulation draws inspiration from the disentanglement of a video into its base components: content and motion representations as presented in [22]. In their work [22], they investigated the controlled generation of short length videos by mixing motion and con" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.088, + 0.891, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.161, + 0.894, + 0.221 + ], + "angle": 0, + "content": "Figure 2: Temporal flicker in optical flow. An illustration of the decomposition of optical flow generated from temporally inconsistent frames to consistent optical flow and inconsistent flicker." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.892, + 0.445 + ], + "angle": 0, + "content": "tent components from various videos. In this work, we explore a similar concept of controlled disentanglement and re-entanglement strategy to restore natural video dynamics of the frame-wise processed videos. More specifically, we argue that the previously proposed approaches for this tasks implicitly employ this disentanglement and re-entanglement in their formulation with the help of unprocessed video frames. Conventionally, optical flow is used as an inter-frame motion representations. The estimated optical flow can also be used to approximate the content of future frames with the help of past frames provided that there does not exist occlusion between the frames (as presented in Eq. 1)." + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.444, + 0.892, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {t + 1} \\approx w \\left(I _ {t}, o f _ {t + 1 \\Rightarrow t}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.893, + 0.81 + ], + "angle": 0, + "content": "This interdependence of optical flow and content can lead to inconsistent motion representation as optical flow estimation networks are not trained explicitly to distinguish temporally consistent content from temporal flicker. Due to this limitation, the evaluated optical flow estimated from temporally inconsistent frames is often compromised with temporal flicker (as presented in Fig. 2). This compromise of the motion information makes it challenging to separate the consistent component of optical flow from the additive flicker. Therefore, the currently available techniques for this task often rely on siphoning implicit motion representations from the raw videos and often suffer in cases where the content of both processed and raw videos deviate significantly. Another problem with employing these implicit representations is the bias introduced due to the lack of control in mixing these implicit representations to restore videos which often leads to the mitigation of translation effect. Therefore, if a mechanism can be developed that produces similar motion representations from both unprocessed and framewise processed videos, better and unbiased restorative effects than those produced by the previously proposed approaches can be achieved without relying on the availability of unprocessed videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Please note that another observation that can be made from Fig. 2 (right most figure) is that a significant amount of flicker in frame-wise processed videos, generally occurs near the motion boundaries. This observation helps us in defining objective functions capable of localizing and targeting temporal deviations as presented in Sec. 4.1." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13536" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "To overcome these problems and to learn and utilize consistent motion representation, we integrate an optical flow estimation network into the restoration network as illustrated in Fig. 1. A similar integration of a small-scale optical flow estimation network (SpyNet [26]) for task-oriented flow estimation is presented in [37] for the task video denoising for its extremely small size. We utilize a relatively larger network (PWC-Net [31]) in our method and provide an ablation study on different sized optical flow models in the accompanying supplementary text. We further extend their idea to a two pass mechanism with an added encoder to encourage generation of consistent dynamics. The optical flow network in the proposed methodology is finetuned in an end-to-end manner without any special supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.473, + 0.904 + ], + "angle": 0, + "content": "The proposed model takes in three frames \\( O_{t-1} \\) (\\( P_{t-1} \\) for the first time step), \\( P_t \\) and \\( P_{t+1} \\) as input. These three frames are propagated through both the content and motion branches. The proposed model consists of a UNet [27] like structure with multiple encoder streams and a single decoder stream. The decoder contains skip connections from the content stream to encourage better reconstruction. The bottleneck part of the model contains a recurrent bottleneck block to transmit and manipulate the information of the generated frames to temporally distant frames. Please note that the multi-branch encoder(s)-bottleneck-decoder architecture used for the proposed model is employed due to its proven efficacy in video generation, translation and restoration tasks as described in [4, 8, 14, 15, 16, 23, 32, 34]. The main difference in the proposed architecture comes from the design of our motion branch which allows us to present a standalone solution for this task. The motion branch comprises of two passes through a PwCNet [31] with shared weights followed by a conventional encoder-like architecture. An illustration of the proposed model is presented in Fig. 1. This two-pass strategy in the motion branch along with the attached encoder allows the proposed model in regressing consistent motion dynamics from temporally inconsistent frames. Further architectural details of the proposed network are presented in the accompanying supplemental. Figure 3 presents conventional optical flow along with the learned motion representation generated by the integrated finetuned PwC-Net [31] from temporally consistent (left) and inconsistent frames (right). Please note that the \"modified\" flow can differ starkly from the conventional flow and can only be interpreted effectively by the network it is trained with as described in [37]. Contrasting to the conventional optical flow modified flow estimated from both consistent and inconsistent videos is quite similar, which suggests that the network can effectively utilize the inferred consistent motion representation to restore natural temporal dynamics in the per-frame videos by only taking into account the temporally inconsistent videos at test time. Another benefit of this disentangled representation" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.091, + 0.858, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.229, + 0.895, + 0.307 + ], + "angle": 0, + "content": "Figure 3: Conventional and Modified flow like representation. Comparison of original and Modified flow evaluated from temporally consistent (left) and inconsistent (right) frames. The modified flow evaluated from both the temporally consistent and inconsistent frames is quite similar." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.316, + 0.895, + 0.421 + ], + "angle": 0, + "content": "is that it allows the users to further improve the level of restored temporal consistency with the help of an iterative arrangement; please refer to the supplemental for the details of the iterative arrangement and the results generated with such arrangement. The objective functions used for the end-to-end training of the proposed network are described in detail in the next section." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.435, + 0.694, + 0.453 + ], + "angle": 0, + "content": "4. Objective Functions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.893, + 0.537 + ], + "angle": 0, + "content": "This section provides the details of the learning objectives used in the training phase of the proposed network. These objectives can be classified into two categories, local neighborhood and long-term objectives. Both of these optimization categories are presented below." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.546, + 0.749, + 0.562 + ], + "angle": 0, + "content": "4.1. Local Neighborhood Losses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this work, we aim to find a general solution for correcting the temporal consistency of frame-wise processed videos. Due to the diversity of the applications addressed, a simplistic spatial content matching reconstruction loss cannot justify the generation of temporally consistent frames, as the content can vary significantly in applications like style transfer [9] and image adjustment [38]. This discrepancy in the image space of processed and raw videos hinders a straightforward definition of a feasible loss function for this task. Due to the challenging nature of this task, we propose a flow gradient loss that provides a supervision signal for the reconstruction of motion boundaries (where most of the inconsistencies/flicker appear; as highlighted in Fig. 2) with the help of optical flow acquired through the unprocessed videos. We further extend this loss to only compare the gradients of the optical flow acquired through the synthesized and raw frames. Doing so reduces the redundancy present in the high dimensional optical flow which contains a redundant global component for this task which is only useful in applications like video stabilization [1] as described in [39]. The proposed flow gradient loss encapsulates the local spatio-temporal information necessary" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13537" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "to mitigate temporal flicker near the motion boundaries by removing the unnecessary high-dimensional components from the optical flow. Similar information bottle-necking techniques have been discussed thoroughly in vision based understanding and manipulation literature [7, 24]. The proposed spatio-temporal loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.192, + 0.47, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f g} = \\sum_ {t = 2} ^ {T} \\left\\| \\nabla \\left(o f \\left(O _ {t}, O _ {t - 1}\\right)\\right), \\nabla \\left(o f \\left(I _ {t}, I _ {t - 1}\\right)\\right) \\right\\| _ {1}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.239, + 0.473, + 0.572 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{L}_{fg}\\) and of represents flow gradient loss and the optical flow estimation network (FlowNet2.0 [11]) respectively. Please note that any optical flow estimation network (SpyNet [26], PWC-Net [31], RAFT [33], etc.) can be used in both the motion branch and the definition of the loss functions, we choose FlowNet2.0 [11] for its relatively lower mean end-point-error, as the estimated optical flow can vary significantly for small scale networks (comparison provided in the supplementary text). In equation 2, the \\(\\nabla\\) operator denotes the spatial gradient of the estimated optical flow. A similar loss in optical flow domain has been proposed in [21] where the raw optical flow of the generated video was compared with that of the ground truth video. Although effective for the task of video inpainting, the raw optical flow matching approach fails in applications like style transfer in which the entire content of the generated video and the raw video differ greatly, i.e., contains excessive spatio-temporal variations which contribute to optical flow estimation. In our experiments, the models trained with raw optical flow performed poorly, whereas the model trained with flow gradients converged in a relatively shorter time and produced sharper frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.693 + ], + "angle": 0, + "content": "Please note that this optical flow-based loss by itself is not sufficient for faithfully correcting the temporal consistency, as there could exist multiple solutions for optical flow equation, e.g., it does not take into account the appearance of the synthesized and raw frames. In order to address this issue and to guide the network for the generation of perceptually credible frames, a non-local optical flow based loss in the image space is employed as presented below:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.703, + 0.47, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\sum_ {t = 2} ^ {T} C _ {t} \\left\\| O _ {t} - \\hat {O} _ {t} \\right\\| _ {1}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "This warping based reconstruction loss function has been adopted from the [4], as many variants of this objective function with slight modifications have been proven effective for enforcing temporal smoothness in video translation tasks in [3, 5, 16, 23, 34]. Here \\( T \\) represents the total number of frames in a sequence and \\( \\hat{O}_t \\) represents the warped frame \\( O_{t-1} \\) (as described in Eq. 1). The working of this loss function can be summarized as the propagation of content from preceding frames to the subsequent frames in image domain to ensure that the generated frames contain similar" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.183 + ], + "angle": 0, + "content": "content to the previous frames and the inter-frame motion dynamics of the synthesized frames are similar to that of the temporally consistent counterpart. This non-local optical flow based loss function also takes into account the occlusion problem of the Eq. 1 by masking the occluded and de-occluded content as defined below:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.195, + 0.893, + 0.226 + ], + "angle": 0, + "content": "\\[\nC _ {t} = \\exp \\left(- \\alpha \\left\\| I _ {t} - \\hat {I} _ {t} \\right\\| _ {2} ^ {2}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.229, + 0.892, + 0.274 + ], + "angle": 0, + "content": "The value for \\(\\alpha = 50\\) is chosen according to the previous works where a similar strategy is used for evaluating occlusion [34, 28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.892, + 0.337 + ], + "angle": 0, + "content": "An additional conventional short-term perceptual similarity loss [13] was also introduced in the training phase to minimize the deviation of the synthesized frames from the original processed frames. This loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.349, + 0.893, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\sum_ {t} \\sum_ {l} \\| \\phi_ {l} (O _ {t}) - \\phi_ {l} (P _ {t}) \\| _ {1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.892, + 0.414 + ], + "angle": 0, + "content": "Here \\(\\phi_l(.)\\) represents layers of a VGG-16 network till the layer \\(relu_{-4.3}\\) (trained on the ImageNet dataset [6])." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.424, + 0.736, + 0.44 + ], + "angle": 0, + "content": "4.2. Temporal Constancy Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The loss defined in Eq. 3 adequately defines the content propagation from the previous frame to the current frame but lacks the ability to enforce it in temporally distant instances. For instance, consider the task of colorization where a car appears in a portion of the frames and is assigned some random color, and it disappears and re-appears in a later instance. The frame-wise colorization method can assign it a different color in each sequence, and the loss function defined in Eq. 3 will be sufficient to enforce the color constancy of the car in each interval where the car is visible. This, however, does not ensure the constancy of the color assigned to the car in both intervals, i.e., the car could be assigned a solid blue color in the first interval and a solid red color in the next interval which is quite prominent in methods like [16, 34]. To address this problem of temporally distant yet similar instances, we introduce a recurrent module (ConvLSTM [30]) in our proposed model and add explicit constraints that force the model to generate temporally consistent frames. The previously proposed approaches for this task simply try to minimize the deviation of current frames from the first frame of the sequence, which is inadequate for videos containing largely varying frames. This variation of the content can render this loss ineffective, and the inconsistencies can only be weakly penalized with the help of motion dynamics siphoned from the raw videos. Therefore, we introduce a bi-directional penalty that enforces the temporal smoothness in both backward (through the local neighborhood losses) and forward (through the proposed long-term loss) direction. This bi-directional optimization strategy is inspired from the ping" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13538" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "pong loss dynamic as presented in [5]. The proposed long-term loss is termed as constancy loss and is presented below:" + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.158, + 0.468, + 0.208 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n s t a n c y}} = \\sum_ {p = 1} ^ {T - 2} \\sum_ {t = p + 2} ^ {T} C _ {t \\Rightarrow p} \\| O _ {t} - w \\left(O _ {p}, o f \\left(I _ {t - 1}, I _ {p}\\right)\\right) \\| _ {1}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.209, + 0.47, + 0.405 + ], + "angle": 0, + "content": "Here subscript \\( p \\) highlights that the preceding distant images are compared with all the subsequent frames. Unlike the simplistic long-term temporal constraints used in [16, 34] which only penalized the deviation of content from the first frame, this loss term takes into account the inter-frame flicker of all the frames in a sequence and provides a strong supervision signal capable of penalizing temporal discrepancies throughout the sequence. Due to this stronger temporal penalty accompanied with the proposed flow gradient loss, the proposed network learns to produce smoother videos despite having access to only the frame-wise processed videos at test time. The final loss for each training sequence is given by," + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.413, + 0.469, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {f g} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {3} \\mathcal {L} _ {p} + \\lambda_ {4} \\mathcal {L} _ {\\text {c o n s t a n c y}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.469, + 0.486 + ], + "angle": 0, + "content": "The \\(\\lambda(s)\\) in the equation above define the contribution of each loss in the optimization phase. The details of these hyperparameters are provided in the supplementary text." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.312, + 0.515 + ], + "angle": 0, + "content": "5. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.469, + 0.839 + ], + "angle": 0, + "content": "During our experimentation phase, we tested various optical flow estimation networks such as [11, 26, 31]. Flownet2.0 [11] has approximately 162M trainable parameters, and it produced relatively better results in lesser training iterations as compared to the other two variants. Having a vast number of parameters compromises both the training and testing time. Therefore, we opted for a medium-sized optical flow estimation network (PWC-Net [31]) and a higher number of optimization iterations. The results produced by both variants are discussed in the supplementary text. We also experimented with an iterative arrangement of the trained models where restored videos are again subjected to temporal consistency correction models and evaluated that the proposed model architecture consistently reduces warp error with every iteration (as presented in Fig. 4). Generally, the frames subjected to a high number of consistency correction iterations lose fine details. The generated results by various restoration iterations are provided in the supplemental. Please note that the comparative results provided in this paper are generated through a single restoration iteration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The dataset for training the proposed model contains videos from the DAVIS dataset for video segmentation [25] and videos gathered from Videovo.net by [16]. The training dataset contains videos processed through a diverse range" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.082, + 0.865, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.21, + 0.894, + 0.255 + ], + "angle": 0, + "content": "Figure 4: Temporal warp error vs. iterations. Temporal warp error consistently decreases with a higher number of restoration iterations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.267, + 0.893, + 0.34 + ], + "angle": 0, + "content": "of applications such as, Artistic Style Transfer [9, 13], Colorization [36, 44], Image Enhancement [10], Intrinsic Image Decomposition [2], and Image-to-Image Translation tasks [20, 44, 46]. The qualitative and quantitative results are described below." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.35, + 0.694, + 0.366 + ], + "angle": 0, + "content": "5.1. Quantitative Results" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.385, + 0.895, + 0.517 + ], + "angle": 0, + "content": "
TaskDAVISVIDEO
Bonneel[3]Lai[16]OursDVP[19]*Bonneel[3]Lai[16]OursDVP[19]*
WCT [20]/antimono0.00290.00310.00260.00220.00150.00210.00150.0017
WCT [20]/asheville0.00470.00590.0047-0.00320.00430.0039-
WCT [20]/candy0.00340.00470.0035-0.00200.00320.0025-
WCT [20]/feathers0.00400.00400.0027-0.00270.00300.0021-
WCT [20]/sketch0.00360.00290.0021-0.00250.00220.0017-
WCT [20]/wave0.00330.00350.00270.00260.00230.00260.00210.0019
Fast-neural-style [13]/princess0.00430.00630.00420.00380.00350.00530.00420.0046
Fast-neural-style [13]/ludnie0.00210.00230.00220.00220.00120.00140.00140.0022
DBL [10]/expertA0.00170.00110.00100.00150.00140.00070.00060.0015
DBL [10]/expertB0.00160.00100.00080.00140.00110.00060.00040.0014
IID [2]/reflectance0.00150.00080.00070.00120.00110.00060.00060.0006
IID [2]/shading0.00140.00080.00080.00130.00080.00040.00040.0017
CycleGAN [46]/photo2ukiyoe0.00240.00190.00150.00130.00170.00130.00100.0018
CycleGAN [46]/photo2vangogh0.00260.00260.00190.00140.00200.00200.00150.0020
Colorization [43]0.00160.00110.00080.00130.00100.00060.00040.0012
Colorization [36]0.00150.00090.00070.00130.00100.00050.00030.0012
Average0.00270.00270.00210.0018*0.00180.00190.00150.0018*
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.529, + 0.894, + 0.604 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison of Temporal Warping Error. The comparison presented in this table is limited to the results produced by [3, 16] and the proposed method. Lower warp error suggests better temporal consistency. *** Indicates partially averaged results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.87 + ], + "angle": 0, + "content": "The quantitative results are evaluated on the basis of temporal warp error. Tab. 2 shows the quantitative result produced by the methods proposed in [3, 16, 19] and our method. Please note that there are over 830 video sequences in the evaluation datasets and over 115,000 frames. Due to the extremely tedious and resource-intensive nature of DVP [19], only a portion of the evaluation datasets could be processed and the evaluated metrics in the DVP [19] column are averaged over only the processed videos for each task. In order to have a fair comparison, the metrics evaluated on the videos processed with DVP [19] are excluded from the large-scale quantitative comparison and are only presented for reference only. Furthermore, it is also noteworthy that in their paper [19], the authors do not evaluate the results on the entirety of the dataset and sample tasks for comparative results and report that their warp error is comparable to that of Bonneel et al. [3]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The lower temporal warp error represents better temporal consistency. The lowest temporal warp error is highlighted" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13539" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.272 + ], + "angle": 0, + "content": "in red, and the second best is highlighted in blue. It is worth mentioning that the temporal warp error does not take into account the perceptual quality of the produced frames and assigns a lower value to blurry videos (as produced by [19]). The quality of the results produced by each of the discussed methods can be verified through the accompanying supplementary video. It is evident from Tab. 2 that the proposed model produces state-of-the-art consistency results. Please note that, unlike the proposed method, all of the compared methods require the availability of raw/unprocessed videos for restoring temporal dynamics at both the training and test time." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.274, + 0.472, + 0.35 + ], + "angle": 0, + "content": "We further verify the efficacy of the proposed model through a specialized metric (as presented in [12]) for evaluating motion smoothness in videos, on the portion of evaluation datasets processed through the DVP [19] and present our findings below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.47, + 0.547 + ], + "angle": 0, + "content": "It is worth mentioning that some attempts to judge the perceptual quality through metrics like mean PSNR (performance degradation [19]) have been proposed for this task. We intentionally disregarded this metric and argue that this metric might be helpful to evaluate the performance of successive iterations of their prior-based method where each iteration leads to progressively better results, but it can be a bit misleading in judging the overall perceptual quality of restored videos for this task. A similar phenomenon can be observed with other local perceptual metrics such as LPIPS [45] as it is impossible to quantize a meaningful deviation of perceptual distance or PSNR from per-frame processed videos due to the lack of ground truth data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.471, + 0.745 + ], + "angle": 0, + "content": "In order to be consistent with the prior works, we nonetheless present the LPIPS [45] results in 3. The proposed method performs competitively with the method proposed by [16] while achieving better temporal consistency. It is noteworthy that each of the discussed metrics does not take into account the aspects considered by the other metric and only provides limited insights about the results. Therefore, to properly evaluate the models on both perceptual quality and temporal consistency, thorough user studies are essential and are used as the primary metric for the evaluation of this task. We conducted two separate user studies to properly evaluate all of the discussed methods and present our findings in the next sub-section." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.307, + 0.785 + ], + "angle": 0, + "content": "5.1.1 Smoothness Metric (SM)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Due to the unavailability of ground truth videos, the quantitative metric results presented (both in the main paper and this supplemental) can only give us an approximate insight towards the nature of the restored videos. In order to properly evaluate our work, we present further large-scale quantitative results generated on the portion of the evaluation datasets processed through DVP [19] in this section. Fig. 5" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.895, + 0.221 + ], + "angle": 0, + "content": "
TaskDAVISVIDEO
Bonneel[3]Lai[16]OursDVP[19]*Bonneel[3]Lai[16]OursDVP[19]*
WCT [20]/antimony0.21660.04310.04850.33940.21800.04510.04470.2992
WCT [20]/asheville0.20560.03730.0449-0.26800.05140.0485-
WCT [20]/candy0.25880.04350.0497-0.21750.05340.0514-
WCT [20]/feathers0.20960.04670.0508-0.15190.05900.0540-
WCT [20]/sketch0.14510.04990.0511-0.18710.05170.0496-
WCT [20]/wave0.17770.04440.04850.34230.25830.07490.07930.3451
Fast-neural-style [13]/princess0.22380.06460.07440.47820.16100.05550.06250.2566
Fast-neural-style [13]/tudnie0.14730.05030.06090.36600.08000.05140.06040.1487
DBL [10]/expertA0.07950.03790.04950.14530.07420.05040.05840.1349
DBL [10]/expertB0.05300.03590.04810.13610.13590.06390.05780.1947
IID [2]/reflectance0.10110.04930.05360.13890.07830.06590.07310.2635
IID [2]/ shading0.06460.05270.06430.15580.11910.05060.05920.2302
CycleGAN [46]/photo2ukiyoe0.09080.03940.05020.23510.14620.05050.05750.2571
CycleGAN [46]/photo2vangogh0.12300.03900.04940.25650.12240.05190.05510.1701
Colorization [43]0.06160.04020.04680.15550.08670.04950.05660.1441
Colorization [36]0.03290.03730.04750.14740.13170.00810.02200.1558
Average0.13690.04450.05240.2414*0.15230.05210.05560.2167*
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.895, + 0.308 + ], + "angle": 0, + "content": "Table 3: Quantitative comparison of LPIPS [45]. The comparison presented in this table is limited to the results produced by [3, 16] and the proposed method. Lower LPIPS suggests better perceptual similarity. \\*\\*\" Indicates partially averaged results." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.313, + 0.88, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.427, + 0.895, + 0.503 + ], + "angle": 0, + "content": "Figure 5: Motion Smoothness Results. This figure presents the quantitative results acquired through the motion smoothness metric as described above. The lower motion smoothness represents smoother transitions in the videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.895, + 0.584 + ], + "angle": 0, + "content": "presents the temporal smoothness results generated through the motion smoothness metric as presented in [12]. This metric tries to quantify the smoothness of videos over various window sizes through differential optical flow as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.59, + 0.891, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D = \\left(F _ {I _ {1} \\rightarrow I _ {2}} - F _ {I _ {0} \\rightarrow I _ {1}}\\right) - \\left(F _ {O _ {1} \\rightarrow O _ {2}} - F _ {O _ {0} \\rightarrow O _ {1}}\\right), \\\\ \\mathrm {M} (s) = \\log \\left(\\sum_ {d \\in D} \\mathbf {1} _ {[ s, s + 1)} \\left(\\| \\boldsymbol {d} \\| _ {2}\\right)\\right) - \\log (| D |), \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Here, \\( F_{a\\rightarrow b} \\) denotes optical flow estimated from frame \\( a \\) to frame \\( b \\). The preceding \\( I \\) and \\( O \\) refers to the raw/unprocessed and restored frames respectively. In Eq. 8, \\( d \\) denotes each of the \\( [\\nabla u,\\nabla v] \\) vectors in the differential optical flow and \\( s \\) represents the various window sizes for evaluation and \\( |D| \\) denotes the dimensionality of the differential optical flow. \\( \\mathbf{1}_S(d) \\) is the masking factor which is equal to 1 if \\( d \\) belongs to the set \\( S \\). The resultant \\( \\mathrm{M}(s) \\) gives a measure of temporal smoothness over 3 consecutive frames. Fig. 5 presents the average results on the portion of evaluation datasets processed through [3, 16, 19] and the proposed model. A lower score suggests better smoothness. The proposed model again achieves SOTA results in terms of motion smoothness even over bigger window sizes as well." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Please note that all the presented metrics are subjective" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13540" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "in nature and do not account for the overall quality and consistency of the results due to the lack of ground truth; therefore, user studies should be considered as the main criteria for judging the results of all the proposed approaches for this task. We discuss the findings of the conducted user studies below." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.2, + 0.21, + 0.215 + ], + "angle": 0, + "content": "5.1.2 User Study" + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.218, + 0.362, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.384, + 0.471, + 0.446 + ], + "angle": 0, + "content": "Figure 6: User Study. This donut chart highlights the user preferences collected through the user study. A majority of the participants preferred the results produced by the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.471, + 0.784 + ], + "angle": 0, + "content": "We conducted two separate user studies; a comprehensive and a factorized user study, to properly evaluate the performance of the proposed model. The first user study consisted of 36 participants with 150 different scenes processed through each of the proposed models. Each participant was asked to judge 5 videos processed through each of the proposed methods for this task. The users were shown a randomly sampled scene from the video pool processed through [3, 16, 19] and this work, simultaneously. The participants were instructed to judge the videos on the basis of naturalness, quality, consistency, content, and style preservation. Figure 6 presents the findings of the first (comprehensive) user study. On average, \\(41\\%\\) of the users preferred the videos restored by the proposed method. Furthermore, the results generated by the method proposed in Lai et al. [16] also received a significant preference in the user study but it is worth noting that all the previously proposed approaches require the unprocessed counterparts of the frame-wise processed videos at test time to restore the temporal consistency, whereas the proposed model restores the temporal consistency by only taking into account the inconsistent video at the test time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.287, + 0.817 + ], + "angle": 0, + "content": "5.1.3 Factorized User Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Apart from the comprehensive user study, we also conducted a secondary in-depth user study to evaluate the reasoning of participants and their opinion on their selections of videos processed with all the methods on challenging scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.319 + ], + "angle": 0, + "content": "In this user study, each user was instructed to judge 15 randomly selected videos (5 videos from each of the applications mentioned above) processed by each of the proposed models for this task. Please note that in this user study users were also shown the per-frame processed video in an extended window to judge the videos on content and style preservation as well. Apart from recording their preference, the participants were also instructed to record the reasoning for their selection. \\(\\sim 86\\%\\) of the participants remarked that they chose the videos based on lesser flickering and better quality preservation. The results of the second user study for each of the selected tasks are provided in Fig. 7. On average, \\(\\sim 47\\%\\) of the users preferred the videos processed through the proposed model for these extremely challenging tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.332, + 0.683, + 0.348 + ], + "angle": 0, + "content": "5.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Figures \\(8\\sim 10\\) present some of the comparative qualitative results produced by the methods proposed in [3, 16, 19] and the proposed model. The results produced by [3] fail to retain the perceptual quality of video sequences where occlusion and deocclusion occur. Whereas the frames processed by DVP [19] are significantly blurry and lose the translation effect. In contrast to the comparison with DVP [19] and Boneel et al. [3], the comparison with the results produced through the method proposed by Lai et al. [16] can be a bit tricky. In order to aid the readers, we provide an epipolar comparison between the frames restored through their method and ours in Fig. 10. Due to the implicit definition of motion dynamics in their method [16], their method fails to faithfully retain the perceptual quality in the restored frames and often flips the colors/appearance of objects in the restored video according to the raw video. This bias towards the raw videos is quite prominent in methods like DVP [19] and Lai et al. [16]. Contrasting to their formulation, the explicit design of motion dynamics in the proposed method mitigates the need for raw videos at test time which consequently eliminates this bias and enables the proposed method to faithfully restore videos. Further qualitative comparisons and higher quality results, along with the results on super resolution [17], inpainting [18], and multiple restoration iterations are presented in the supplemental." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.76, + 0.768, + 0.775 + ], + "angle": 0, + "content": "6. Limitations and Future work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The proposed model has difficulties in restoring videos with very low frame rates and in the case of style-transferred videos, some finer inconsistencies like finer brush strokes are lost, which makes the resulting video seem a little dull (which is to be expected from all the proposed approaches for this task). A good direction for future work in this field can be the direct denoising of conventional optical flow with deterministic warp operators. The developed consis" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13541" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.087, + 0.37, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.087, + 0.59, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.087, + 0.813, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.318, + 0.895, + 0.366 + ], + "angle": 0, + "content": "Figure 7: Factorized user study. The user preference scores for the tasks of style transfer, CycleGAN, and intrinsic image decomposition (left to right respectively). The majority of the users preferred the videos processed through the proposed method in each category of tasks." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.379, + 0.468, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.508, + 0.471, + 0.569 + ], + "angle": 0, + "content": "Figure 8: Comparison with Bonneel et al. [3] The method proposed by Bonneel et al. [3] has difficulty in handling occlusion and long term consistency and produce temporal artifacts." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.582, + 0.468, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.721, + 0.471, + 0.783 + ], + "angle": 0, + "content": "Figure 9: Comparison with DVP [19] DVP [19] produces blurry frames and compromises the translation effect. This phenomenon is prominent in videos where new content is generated e.g. style transfer [13], CycleGAN [46]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.827 + ], + "angle": 0, + "content": "tent flow models can also be used for a number of video processing applications like denoising and dehazing." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.833, + 0.196, + 0.849 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this work, we present a task-agnostic temporal consistency correction framework that restores natural video dynamics by inferring and utilizing consistent motion repre" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.379, + 0.871, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.565 + ], + "angle": 0, + "content": "Figure 10: Comparison with Lai et al. [16] This figure produces epipolar temporal comparison of videos restored by Lai et al. [16] and the proposed method. The method proposed in [16], due to its dependence on raw videos and implicit siphoning/mixing of motion dynamics from the raw video introduces bias towards the raw videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.74 + ], + "angle": 0, + "content": "sentations from temporally inconsistent videos. The proposed strategy mitigates the need for siphoning video dynamics from the unprocessed videos at test time consequently mitigating the compromise in translation effects that is prominent in the currently proposed approaches for this task and expands the scope of its employability to a wider set of applications. Through extensive experimentation and user studies, we demonstrate that despite the limited access to data at test-time the proposed method compares favorably to the pre-existing methods available for this task and achieves SOTA performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.746, + 0.669, + 0.764 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.771, + 0.892, + 0.892 + ], + "angle": 0, + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2022-0-00156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D meta-verse transformation) and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13542" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] Muhammad Kashif Ali, Sangjoon Yu, and Tae Hyun Kim. Deep motion blind video stabilization. BMVC, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.144, + 0.468, + 0.171 + ], + "angle": 0, + "content": "[2] Sean Bell, Kavita Bala, and Noah Snavely. Intrinsic images in the wild. ACM TOG, 33(4):1-12, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.468, + 0.213 + ], + "angle": 0, + "content": "[3] Nicolas Bonneel, James Tompkin, Kalyan Sunkavalli, Deqing Sun, Sylvain Paris, and Hanspeter Pfister. Blind video temporal consistency. ACM TOG, 34(6):1-9, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.468, + 0.255 + ], + "angle": 0, + "content": "[4] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent online video style transfer. In ICCV, pages 1105-1114, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.468, + 0.298 + ], + "angle": 0, + "content": "[5] Mengyu Chu, You Xie, Laura Leal-Taixe, and Nils Thuerey. Temporally coherent gans for video super-resolution (tecogan). arXiv preprint arXiv:1811.09393, 1(2):3, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.299, + 0.468, + 0.34 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.468, + 0.41 + ], + "angle": 0, + "content": "[7] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Counting out time: Class agnostic video repetition counting in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10387-10396, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.468, + 0.464 + ], + "angle": 0, + "content": "[8] Wei Gao, Yijun Li, Yihang Yin, and Ming-Hsuan Yang. Fast video multi-style transfer. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3211-3219, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.468, + 0.507 + ], + "angle": 0, + "content": "[9] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.509, + 0.468, + 0.55 + ], + "angle": 0, + "content": "[10] Michael Gharbi, Jiawen Chen, Jonathan T Barron, Samuel W Hasinoff, and Frédo Durand. Deep bilateral learning for real-time image enhancement. ACM TOG, 36(4):1-12, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[11] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, pages 2462-2470, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[12] Meiguang Jin, Zhe Hu, and Paolo Favaro. Learning to extract flawless slow motion from blurry videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8112-8121, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[13] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[14] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep blind video decaptioning by temporal aggregation and recurrence. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4258–4267, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[15] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Recurrent temporal aggregation framework for deep video inpainting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42:1038–1052, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.468, + 0.872 + ], + "angle": 0, + "content": "[16] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[17] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, pages 4681-4690, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[18] Eunhye Lee, Jeongmu Kim, Jisu Kim, and Tae Hyun Kim. Restore from restored: Single-image inpainting. arXiv preprint arXiv:2102.08078, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.18, + 0.892, + 0.208 + ], + "angle": 0, + "content": "[19] Chenyang Lei, Yazhou Xing, and Qifeng Chen. Blind video temporal consistency via deep video prior. NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.21, + 0.892, + 0.251 + ], + "angle": 0, + "content": "[20] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. NeurIPS, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.254, + 0.892, + 0.308 + ], + "angle": 0, + "content": "[21] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. arXiv preprint arXiv:2204.02663, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.311, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[22] Xunyu Lin, Victor Campos, Xavier Giro-i Nieto, Jordi Torres, and Cristian Canton Ferrer. Disentangling motion, foreground and background features in videos. arXiv preprint arXiv:1707.04092, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.369, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[23] Yihao Liu, Hengyuan Zhao, Kelvin CK Chan, Xintao Wang, Chen Change Loy, Yu Qiao, and Chao Dong. Temporally consistent video colorization with deep feature propagation and self-regularization learning. arXiv preprint arXiv:2110.04562, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.44, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[24] Cheng Ma, Yongming Rao, Yean Cheng, Ce Chen, Jiwen Lu, and Jie Zhou. Structure-preserving super resolution with gradient guidance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7769-7778, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[25] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.569, + 0.892, + 0.61 + ], + "angle": 0, + "content": "[26] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In CVPR, pages 4161-4170, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.612, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.657, + 0.892, + 0.698 + ], + "angle": 0, + "content": "[28] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In \\(GCPR\\), pages 26-36. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[29] Mehdi SM Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super-resolution. In \\(CVPR\\), pages 6626-6634, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.744, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[30] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. NeurIPS, 28, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[31] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In CVPR, pages 8934-8943, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Ximeng Sun, Huijuan Xu, and Kate Saenko. Twostreamvan: Improving motion modeling in video generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2744-2753, 2020." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13543" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[33] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[34] Hugo Thimonier, Julien Despois, Robin Kips, and Matthieu Perrot. Learning long term style preserving blind video temporal consistency. In ICME, pages 1-6. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.468, + 0.207 + ], + "angle": 0, + "content": "[35] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In CVPR, pages 9446-9454, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.208, + 0.468, + 0.248 + ], + "angle": 0, + "content": "[36] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In ECCV, pages 836-850. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.25, + 0.468, + 0.305 + ], + "angle": 0, + "content": "[37] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127(8):1106-1125, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.307, + 0.468, + 0.348 + ], + "angle": 0, + "content": "[38] Zhicheng Yan, Hao Zhang, Baoyuan Wang, Sylvain Paris, and Yizhou Yu. Automatic photo adjustment using deep neural networks. ACM TOG, 35(2):1-15, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.35, + 0.468, + 0.418 + ], + "angle": 0, + "content": "[39] Yuchen Yang, Ye Xiang, Shuaicheng Liu, Lifang Wu, Boxuan Zhao, and Bing Zeng. Glm-net: Global and local motion estimation via task-oriented encoder-decoder structure. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4211-4219, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.42, + 0.468, + 0.461 + ], + "angle": 0, + "content": "[40] Chun-Han Yao, Chia-Yang Chang, and Shao-Yi Chien. Occlusion-aware video temporal consistency. In ACMMM, pages 777-785, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.463, + 0.468, + 0.504 + ], + "angle": 0, + "content": "[41] Bo Zhang, Mingming He, Jing Liao, Pedro V Sander, Lu Yuan, Amine Bermak, and Dong Chen. Deep exemplar-based video colorization. In CVPR, pages 8052-8061, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.506, + 0.468, + 0.56 + ], + "angle": 0, + "content": "[42] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.563, + 0.468, + 0.602 + ], + "angle": 0, + "content": "[43] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, pages 649-666. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.604, + 0.468, + 0.645 + ], + "angle": 0, + "content": "[44] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, pages 649-666. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.648, + 0.468, + 0.701 + ], + "angle": 0, + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.704, + 0.468, + 0.758 + ], + "angle": 0, + "content": "[46] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2223-2232, 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13544" + } + ] +] \ No newline at end of file diff --git a/2023/Task Agnostic Restoration of Natural Video Dynamics/e82ebeb9-3e53-42eb-9d7f-02ba103a6afc_origin.pdf b/2023/Task Agnostic Restoration of Natural Video Dynamics/e82ebeb9-3e53-42eb-9d7f-02ba103a6afc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bfa71f3f8552726fbb8976d8463e294a27714017 --- /dev/null +++ b/2023/Task Agnostic Restoration of Natural Video Dynamics/e82ebeb9-3e53-42eb-9d7f-02ba103a6afc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61391b5176f5c8ab58a76ec92f73e1b0f039eb462380841573e69049442f20f9 +size 2231494 diff --git a/2023/Task Agnostic Restoration of Natural Video Dynamics/full.md b/2023/Task Agnostic Restoration of Natural Video Dynamics/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e27a511c75e83b57cb7509454fe41053e4e1602b --- /dev/null +++ b/2023/Task Agnostic Restoration of Natural Video Dynamics/full.md @@ -0,0 +1,291 @@ +# Task Agnostic Restoration of Natural Video Dynamics + +Muhammad Kashif Ali Dongjin Kim Tae Hyun Kim* +Department of Computer Science, Hanyang University, Seoul, Korea +{kashifali, dongjinkim, taehyunkim}@hanyang.ac.kr + +# Abstract + +In many video restoration/translation tasks, image processing operations are naively extended to the video domain by processing each frame independently, disregarding the temporal connection of the video frames. This disregard for the temporal connection often leads to severe temporal inconsistencies. State-Of-The-Art (SOTA) techniques that address these inconsistencies rely on the availability of unprocessed videos to implicitly siphon and utilize consistent video dynamics to restore the temporal consistency of frame-wise processed videos which often jeopardizes the translation effect. We propose a general framework for this task that learns to infer and utilize consistent motion dynamics from inconsistent videos to mitigate the temporal flicker while preserving the perceptual quality for both the temporally neighboring and relatively distant frames without requiring the raw videos at test time. The proposed framework produces SOTA results on two benchmark datasets, DAVIS and videoo.net, processed by numerous image processing applications. The code and the trained models are available at https://github.com/MKashifAli/TARONVD. + +# 1. Introduction + +Video sharing social media platforms like Snapchat and TikTok have introduced the common populace to a plethora of computer vision applications such as Style Transfer [9], Colorization [43], Denoising [42], and Dehazing [38]. With this wide-scale integration of these classical computer vision applications in such platforms, various image processing operations are naively extended to videos due to scarcity of annotated video datasets and their computational complexity of video processing methodologies. This naive extension of static image processing methodologies to videos disregards the temporal connection of the video progression and introduces severe temporal flickering in the videos. This temporal flicker can appear for various reasons; for in + +stance, these image processing methods can produce drastically different results for temporally neighboring frames due to slight changes in their global or local content statistics, or it could also happen due to the multi-modality of the application as there could exist a number of valid solutions for images with similar content as highlighted in [3, 41]. Therefore, the extension of these image-to-image translation tasks to the video domain is an active area of research in computer vision, and methods that can help extend various image processing applications with little to no knowledge of the operations used to create videos are quite useful. + +Currently, there are several task-dependent temporal consistency correction approaches available such as [4, 23, 34], but due to the complex nature of these applications, only a handful of approaches have been proposed to tackle the problem of blind temporal consistency correction. Bonneel et al. [3] initiated the investigation of blind temporal consistency correction with gradient-domain minimization of per-frame processed video with the unprocessed video to minimize the warping error between the frames. Lai et al. [16] extended the formulation of [3] with the help of recurrent Convolutional Neural Networks (CNN) and introduced a perceptual penalty in their formulation to restrict the deviation of perceptual content of the restored video from the frame-wise processed video. Deep Video Prior (DVP) [19] extended Deep Image Prior (DIP) [35] to the temporal dimension and proposed to formulate enforcing temporal consistency by training a CNN to generate processed video from the unprocessed video without utilizing optical flow. All of these approaches rely on the availability of unprocessed videos for implicit extraction of consistent motion dynamics to use as a restoration guide. + +Please note that all the previously proposed approaches for this task define it with the help of unprocessed videos (both task-specific and task-agnostic). Having access to unprocessed videos at test time helps the models in siphoning and transferring consistent motion dynamics to the inconsistent videos to decrease temporal flicker. Despite its efficacy, this implicit definition limits the applicability of the previously proposed approaches to only the videos for which their unprocessed counterparts are available and also + +
Key featuresBonneel et al. [3]Lai et al. [16]DVP [19]Ours
Does not require raw videos×××✓
Explicit video dynamics design✓××✓
Faithful perceptual restoration×××✓
Sequential processing✓✓×✓
Long term consistency××✓✓
Sharpness in processed videos×✓×✓
High-frequency flicker removal✓×✓✓
Test time training××✓×
Runtime (per frame avg.)2.4635*0.21463.63650.2236
+ +Table 1: Key differences in available approaches. \*\*" in front of Bonnel et al. [3]'s runtime highlights environment difference and is only presented for reference only. Please note that the runtime is averaged over the same video sequence for all the methods. + +introduce an inherent bias towards the unprocessed videos which can compromise the translation effect and quality of the processed videos (as presented in Fig. $8\sim 10$ + +In order to overcome these limitations, we aim to learn and infer consistent motion representation solely from temporally inconsistent videos. Doing so does not only alleviate the need for raw videos at test time, but also mitigates the inherent bias towards the raw videos that is commonly present in the currently available approaches for this task. For the task of learning these consistent motion representation, we fine-tune conventional optical flow estimations networks like [11, 31, 33] along with the restorative part of the network in an end-to-end manner. The proposed model pipeline is illustrated in Fig. 1. The proposed model achieves state-of-the-art qualitative and quantitative results. The proposed formulation can also handle the resolution mismatch problem in processed and raw videos and makes it possible to extend the Single Image Super-Resolution [17] methods to Video Super-Resolution methods without any modification. The detailed description of our formulation is presented in 3. We summarize our contributions as follows: + +- Identify and propose tailored solutions for various challenges of naïve extensions of image processing applications to videos. +- Learn to infer consistent motion representations from temporally inconsistent videos. +- Utilize the learned consistent motion representations to propose a general framework for task agnostic temporal consistency correction that produce SOTA results. + +# 2. Related Work + +The literature is divided into two main streams to generate visually appealing videos through image transformation models. The first stream tackles the task of mitigating temporal inconsistencies with the reformulation of the task at hand. Whereas, The second stream focuses on developing post-processing models that refine the frame-wise processed videos by penalizing the temporal deviation of the processed video from its unprocessed counterpart. The + +second stream is further divided into two sub-streams: task-specific and task-agnostic. The following passages summarize the details of both of the above-mentioned streams, respectively. + +Reformulation stream (video-to-video translation): These approaches are generally termed video-to-video translation applications. Generally, these approaches either consider multiple frames as input and produce multiple output frames or generate a single frame from multiple input frames in the form of frame recurrence such as [5, 23]. There are also cases like video style transfer where content information is propagated to the next time step with the help of optical flow to initiate the optimization of the next frame. This frame-recurrent methodology has also been proven effective in applications like video super resolution [29]. Designing these approaches for each task and training them from scratch is a taxing task and data scarcity can make these formulations unfeasible. Generally, such models often do not adapt well to different tasks; therefore, methods that can restore the temporal consistency of multiple frame-wise processed videos are being actively investigated. + +Post-processing methods: Bonneel et al. [3] initiated the investigation of task agnostic temporal consistency correction using a gradient-based optimization strategy in which temporally consistent (unprocessed) videos were used as restoration guides for correcting the temporal inconsistency of frame-wise processed videos. Their formulation motivated various task-dependent and task-agnostic, and temporal consistency correction approaches with slight variations of their original objective formulation. Despite the simplicity of their formulation, there only exist a handful of approaches that address task-agnostic temporal consistency correction due to the difficult nature of the task. Yao et al. [40] proposed a key-frame strategy that accounted for the motion of different objects in those key-frames to handle occlusion as well as temporal consistency. Lai et al.[16] proposed the first deep learning approach for this task by employing ConvLSTM [30] and a perceptual loss [13]. Lei et al. [19] proposed an extension of Deep Image Prior [35] and demonstrated its capability to mitigate the temporal flicker of per-frame processed videos. + +Apart from these general frameworks, various task-specific approaches have also been proposed, such as [5, 23, 34]. Defining these task-specific approaches is relatively straightforward. These approaches define a temporal extension around a backbone method and penalize the deviation of generated content with the help of motion information from the unprocessed videos. + +In this work, we propose a general formulation for this task that alleviates the requirement of unprocessed videos at test time by learning to infer a consistent motion representation from the temporally inconsistent videos. This inferred + +motion representation is then processed with the content of temporally inconsistent videos in a two branched network that restores the natural dynamics of per-frame processed videos. Our proposed network also contains a recurrent bottleneck module that helps consistent propagation of content throughout the video sequence and produces SOTA qualitative and quantitative results by only taking into account the temporally inconsistent videos at test time. + +In order to aid the readers, we provide the following table to highlight the key differences between the currently available approaches for this task. + +# 3. Proposed Method + +![](images/01617768a5374400703e76106386427c83011b67d8f0b9dab46041c43f9df958.jpg) +Figure 1: Model Architecture. An illustration of the proposed model architecture. The proposed model consists of a two-branch architecture that separately processes motion and content. The combined motion and content features concatenated along the channel dimension and are then passed through a Recurrent Bottleneck Block (RBB) and a decoder to generate the restored frame. + +In this section, we describe the details of the proposed method. Consider a raw (temporally consistent) video $\{I_0, I_1, I_2, \dots, I_n\}$ with $n$ frames and its frame-wise processed (temporally inconsistent) video $\{P_0, P_1, P_2, \dots, P_n\}$ acquired by an image processing function $h$ as $\{h(I_0), h(I_1), h(I_2), \dots, h(I_n)\}$ . The goal of this task is to restore the temporal consistency of the temporally inconsistent video and to produce a temporally consistent version $\{O_0, O_1, O_2, \dots, O_n\}$ . For the sake of simplicity, these notations are chosen to be in line with the notations used in past literature [3, 16, 19]. + +The proposed formulation draws inspiration from the disentanglement of a video into its base components: content and motion representations as presented in [22]. In their work [22], they investigated the controlled generation of short length videos by mixing motion and con + +![](images/07563e715b481b3ceacaa6abe2fbf34ed3d98d29043097b40b254c8d232f2402.jpg) +Figure 2: Temporal flicker in optical flow. An illustration of the decomposition of optical flow generated from temporally inconsistent frames to consistent optical flow and inconsistent flicker. + +tent components from various videos. In this work, we explore a similar concept of controlled disentanglement and re-entanglement strategy to restore natural video dynamics of the frame-wise processed videos. More specifically, we argue that the previously proposed approaches for this tasks implicitly employ this disentanglement and re-entanglement in their formulation with the help of unprocessed video frames. Conventionally, optical flow is used as an inter-frame motion representations. The estimated optical flow can also be used to approximate the content of future frames with the help of past frames provided that there does not exist occlusion between the frames (as presented in Eq. 1). + +$$ +\hat {I} _ {t + 1} \approx w \left(I _ {t}, o f _ {t + 1 \Rightarrow t}\right). \tag {1} +$$ + +This interdependence of optical flow and content can lead to inconsistent motion representation as optical flow estimation networks are not trained explicitly to distinguish temporally consistent content from temporal flicker. Due to this limitation, the evaluated optical flow estimated from temporally inconsistent frames is often compromised with temporal flicker (as presented in Fig. 2). This compromise of the motion information makes it challenging to separate the consistent component of optical flow from the additive flicker. Therefore, the currently available techniques for this task often rely on siphoning implicit motion representations from the raw videos and often suffer in cases where the content of both processed and raw videos deviate significantly. Another problem with employing these implicit representations is the bias introduced due to the lack of control in mixing these implicit representations to restore videos which often leads to the mitigation of translation effect. Therefore, if a mechanism can be developed that produces similar motion representations from both unprocessed and framewise processed videos, better and unbiased restorative effects than those produced by the previously proposed approaches can be achieved without relying on the availability of unprocessed videos. + +Please note that another observation that can be made from Fig. 2 (right most figure) is that a significant amount of flicker in frame-wise processed videos, generally occurs near the motion boundaries. This observation helps us in defining objective functions capable of localizing and targeting temporal deviations as presented in Sec. 4.1. + +To overcome these problems and to learn and utilize consistent motion representation, we integrate an optical flow estimation network into the restoration network as illustrated in Fig. 1. A similar integration of a small-scale optical flow estimation network (SpyNet [26]) for task-oriented flow estimation is presented in [37] for the task video denoising for its extremely small size. We utilize a relatively larger network (PWC-Net [31]) in our method and provide an ablation study on different sized optical flow models in the accompanying supplementary text. We further extend their idea to a two pass mechanism with an added encoder to encourage generation of consistent dynamics. The optical flow network in the proposed methodology is finetuned in an end-to-end manner without any special supervision. + +The proposed model takes in three frames $O_{t-1}$ ( $P_{t-1}$ for the first time step), $P_t$ and $P_{t+1}$ as input. These three frames are propagated through both the content and motion branches. The proposed model consists of a UNet [27] like structure with multiple encoder streams and a single decoder stream. The decoder contains skip connections from the content stream to encourage better reconstruction. The bottleneck part of the model contains a recurrent bottleneck block to transmit and manipulate the information of the generated frames to temporally distant frames. Please note that the multi-branch encoder(s)-bottleneck-decoder architecture used for the proposed model is employed due to its proven efficacy in video generation, translation and restoration tasks as described in [4, 8, 14, 15, 16, 23, 32, 34]. The main difference in the proposed architecture comes from the design of our motion branch which allows us to present a standalone solution for this task. The motion branch comprises of two passes through a PwCNet [31] with shared weights followed by a conventional encoder-like architecture. An illustration of the proposed model is presented in Fig. 1. This two-pass strategy in the motion branch along with the attached encoder allows the proposed model in regressing consistent motion dynamics from temporally inconsistent frames. Further architectural details of the proposed network are presented in the accompanying supplemental. Figure 3 presents conventional optical flow along with the learned motion representation generated by the integrated finetuned PwC-Net [31] from temporally consistent (left) and inconsistent frames (right). Please note that the "modified" flow can differ starkly from the conventional flow and can only be interpreted effectively by the network it is trained with as described in [37]. Contrasting to the conventional optical flow modified flow estimated from both consistent and inconsistent videos is quite similar, which suggests that the network can effectively utilize the inferred consistent motion representation to restore natural temporal dynamics in the per-frame videos by only taking into account the temporally inconsistent videos at test time. Another benefit of this disentangled representation + +![](images/afb0d404e443e7db0b81ac4e7feb6d31e177f021f3e9caa503fa41bb70d840b2.jpg) +Figure 3: Conventional and Modified flow like representation. Comparison of original and Modified flow evaluated from temporally consistent (left) and inconsistent (right) frames. The modified flow evaluated from both the temporally consistent and inconsistent frames is quite similar. + +is that it allows the users to further improve the level of restored temporal consistency with the help of an iterative arrangement; please refer to the supplemental for the details of the iterative arrangement and the results generated with such arrangement. The objective functions used for the end-to-end training of the proposed network are described in detail in the next section. + +# 4. Objective Functions + +This section provides the details of the learning objectives used in the training phase of the proposed network. These objectives can be classified into two categories, local neighborhood and long-term objectives. Both of these optimization categories are presented below. + +# 4.1. Local Neighborhood Losses + +In this work, we aim to find a general solution for correcting the temporal consistency of frame-wise processed videos. Due to the diversity of the applications addressed, a simplistic spatial content matching reconstruction loss cannot justify the generation of temporally consistent frames, as the content can vary significantly in applications like style transfer [9] and image adjustment [38]. This discrepancy in the image space of processed and raw videos hinders a straightforward definition of a feasible loss function for this task. Due to the challenging nature of this task, we propose a flow gradient loss that provides a supervision signal for the reconstruction of motion boundaries (where most of the inconsistencies/flicker appear; as highlighted in Fig. 2) with the help of optical flow acquired through the unprocessed videos. We further extend this loss to only compare the gradients of the optical flow acquired through the synthesized and raw frames. Doing so reduces the redundancy present in the high dimensional optical flow which contains a redundant global component for this task which is only useful in applications like video stabilization [1] as described in [39]. The proposed flow gradient loss encapsulates the local spatio-temporal information necessary + +to mitigate temporal flicker near the motion boundaries by removing the unnecessary high-dimensional components from the optical flow. Similar information bottle-necking techniques have been discussed thoroughly in vision based understanding and manipulation literature [7, 24]. The proposed spatio-temporal loss is defined as follows: + +$$ +\mathcal {L} _ {f g} = \sum_ {t = 2} ^ {T} \left\| \nabla \left(o f \left(O _ {t}, O _ {t - 1}\right)\right), \nabla \left(o f \left(I _ {t}, I _ {t - 1}\right)\right) \right\| _ {1}. \tag {2} +$$ + +Here, $\mathcal{L}_{fg}$ and of represents flow gradient loss and the optical flow estimation network (FlowNet2.0 [11]) respectively. Please note that any optical flow estimation network (SpyNet [26], PWC-Net [31], RAFT [33], etc.) can be used in both the motion branch and the definition of the loss functions, we choose FlowNet2.0 [11] for its relatively lower mean end-point-error, as the estimated optical flow can vary significantly for small scale networks (comparison provided in the supplementary text). In equation 2, the $\nabla$ operator denotes the spatial gradient of the estimated optical flow. A similar loss in optical flow domain has been proposed in [21] where the raw optical flow of the generated video was compared with that of the ground truth video. Although effective for the task of video inpainting, the raw optical flow matching approach fails in applications like style transfer in which the entire content of the generated video and the raw video differ greatly, i.e., contains excessive spatio-temporal variations which contribute to optical flow estimation. In our experiments, the models trained with raw optical flow performed poorly, whereas the model trained with flow gradients converged in a relatively shorter time and produced sharper frames. + +Please note that this optical flow-based loss by itself is not sufficient for faithfully correcting the temporal consistency, as there could exist multiple solutions for optical flow equation, e.g., it does not take into account the appearance of the synthesized and raw frames. In order to address this issue and to guide the network for the generation of perceptually credible frames, a non-local optical flow based loss in the image space is employed as presented below: + +$$ +\mathcal {L} _ {\text {r e c o n}} = \sum_ {t = 2} ^ {T} C _ {t} \left\| O _ {t} - \hat {O} _ {t} \right\| _ {1}, \tag {3} +$$ + +This warping based reconstruction loss function has been adopted from the [4], as many variants of this objective function with slight modifications have been proven effective for enforcing temporal smoothness in video translation tasks in [3, 5, 16, 23, 34]. Here $T$ represents the total number of frames in a sequence and $\hat{O}_t$ represents the warped frame $O_{t-1}$ (as described in Eq. 1). The working of this loss function can be summarized as the propagation of content from preceding frames to the subsequent frames in image domain to ensure that the generated frames contain similar + +content to the previous frames and the inter-frame motion dynamics of the synthesized frames are similar to that of the temporally consistent counterpart. This non-local optical flow based loss function also takes into account the occlusion problem of the Eq. 1 by masking the occluded and de-occluded content as defined below: + +$$ +C _ {t} = \exp \left(- \alpha \left\| I _ {t} - \hat {I} _ {t} \right\| _ {2} ^ {2}\right). \tag {4} +$$ + +The value for $\alpha = 50$ is chosen according to the previous works where a similar strategy is used for evaluating occlusion [34, 28]. + +An additional conventional short-term perceptual similarity loss [13] was also introduced in the training phase to minimize the deviation of the synthesized frames from the original processed frames. This loss is defined as follows: + +$$ +\mathcal {L} _ {p} = \sum_ {t} \sum_ {l} \| \phi_ {l} (O _ {t}) - \phi_ {l} (P _ {t}) \| _ {1}, \tag {5} +$$ + +Here $\phi_l(.)$ represents layers of a VGG-16 network till the layer $relu_{-4.3}$ (trained on the ImageNet dataset [6]). + +# 4.2. Temporal Constancy Loss + +The loss defined in Eq. 3 adequately defines the content propagation from the previous frame to the current frame but lacks the ability to enforce it in temporally distant instances. For instance, consider the task of colorization where a car appears in a portion of the frames and is assigned some random color, and it disappears and re-appears in a later instance. The frame-wise colorization method can assign it a different color in each sequence, and the loss function defined in Eq. 3 will be sufficient to enforce the color constancy of the car in each interval where the car is visible. This, however, does not ensure the constancy of the color assigned to the car in both intervals, i.e., the car could be assigned a solid blue color in the first interval and a solid red color in the next interval which is quite prominent in methods like [16, 34]. To address this problem of temporally distant yet similar instances, we introduce a recurrent module (ConvLSTM [30]) in our proposed model and add explicit constraints that force the model to generate temporally consistent frames. The previously proposed approaches for this task simply try to minimize the deviation of current frames from the first frame of the sequence, which is inadequate for videos containing largely varying frames. This variation of the content can render this loss ineffective, and the inconsistencies can only be weakly penalized with the help of motion dynamics siphoned from the raw videos. Therefore, we introduce a bi-directional penalty that enforces the temporal smoothness in both backward (through the local neighborhood losses) and forward (through the proposed long-term loss) direction. This bi-directional optimization strategy is inspired from the ping + +pong loss dynamic as presented in [5]. The proposed long-term loss is termed as constancy loss and is presented below: + +$$ +\mathcal {L} _ {\text {c o n s t a n c y}} = \sum_ {p = 1} ^ {T - 2} \sum_ {t = p + 2} ^ {T} C _ {t \Rightarrow p} \| O _ {t} - w \left(O _ {p}, o f \left(I _ {t - 1}, I _ {p}\right)\right) \| _ {1}. \tag {6} +$$ + +Here subscript $p$ highlights that the preceding distant images are compared with all the subsequent frames. Unlike the simplistic long-term temporal constraints used in [16, 34] which only penalized the deviation of content from the first frame, this loss term takes into account the inter-frame flicker of all the frames in a sequence and provides a strong supervision signal capable of penalizing temporal discrepancies throughout the sequence. Due to this stronger temporal penalty accompanied with the proposed flow gradient loss, the proposed network learns to produce smoother videos despite having access to only the frame-wise processed videos at test time. The final loss for each training sequence is given by, + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {1} \mathcal {L} _ {f g} + \lambda_ {2} \mathcal {L} _ {\text {r e c o n}} + \lambda_ {3} \mathcal {L} _ {p} + \lambda_ {4} \mathcal {L} _ {\text {c o n s t a n c y}}. \tag {7} +$$ + +The $\lambda(s)$ in the equation above define the contribution of each loss in the optimization phase. The details of these hyperparameters are provided in the supplementary text. + +# 5. Experiments and Results + +During our experimentation phase, we tested various optical flow estimation networks such as [11, 26, 31]. Flownet2.0 [11] has approximately 162M trainable parameters, and it produced relatively better results in lesser training iterations as compared to the other two variants. Having a vast number of parameters compromises both the training and testing time. Therefore, we opted for a medium-sized optical flow estimation network (PWC-Net [31]) and a higher number of optimization iterations. The results produced by both variants are discussed in the supplementary text. We also experimented with an iterative arrangement of the trained models where restored videos are again subjected to temporal consistency correction models and evaluated that the proposed model architecture consistently reduces warp error with every iteration (as presented in Fig. 4). Generally, the frames subjected to a high number of consistency correction iterations lose fine details. The generated results by various restoration iterations are provided in the supplemental. Please note that the comparative results provided in this paper are generated through a single restoration iteration. + +The dataset for training the proposed model contains videos from the DAVIS dataset for video segmentation [25] and videos gathered from Videovo.net by [16]. The training dataset contains videos processed through a diverse range + +![](images/dac6bec8e5826318e9aaef88af745197141633817afe45eae5c7eb9b4b555461.jpg) +Figure 4: Temporal warp error vs. iterations. Temporal warp error consistently decreases with a higher number of restoration iterations. + +of applications such as, Artistic Style Transfer [9, 13], Colorization [36, 44], Image Enhancement [10], Intrinsic Image Decomposition [2], and Image-to-Image Translation tasks [20, 44, 46]. The qualitative and quantitative results are described below. + +# 5.1. Quantitative Results + +
TaskDAVISVIDEO
Bonneel[3]Lai[16]OursDVP[19]*Bonneel[3]Lai[16]OursDVP[19]*
WCT [20]/antimono0.00290.00310.00260.00220.00150.00210.00150.0017
WCT [20]/asheville0.00470.00590.0047-0.00320.00430.0039-
WCT [20]/candy0.00340.00470.0035-0.00200.00320.0025-
WCT [20]/feathers0.00400.00400.0027-0.00270.00300.0021-
WCT [20]/sketch0.00360.00290.0021-0.00250.00220.0017-
WCT [20]/wave0.00330.00350.00270.00260.00230.00260.00210.0019
Fast-neural-style [13]/princess0.00430.00630.00420.00380.00350.00530.00420.0046
Fast-neural-style [13]/ludnie0.00210.00230.00220.00220.00120.00140.00140.0022
DBL [10]/expertA0.00170.00110.00100.00150.00140.00070.00060.0015
DBL [10]/expertB0.00160.00100.00080.00140.00110.00060.00040.0014
IID [2]/reflectance0.00150.00080.00070.00120.00110.00060.00060.0006
IID [2]/shading0.00140.00080.00080.00130.00080.00040.00040.0017
CycleGAN [46]/photo2ukiyoe0.00240.00190.00150.00130.00170.00130.00100.0018
CycleGAN [46]/photo2vangogh0.00260.00260.00190.00140.00200.00200.00150.0020
Colorization [43]0.00160.00110.00080.00130.00100.00060.00040.0012
Colorization [36]0.00150.00090.00070.00130.00100.00050.00030.0012
Average0.00270.00270.00210.0018*0.00180.00190.00150.0018*
+ +Table 2: Quantitative comparison of Temporal Warping Error. The comparison presented in this table is limited to the results produced by [3, 16] and the proposed method. Lower warp error suggests better temporal consistency. *** Indicates partially averaged results. + +The quantitative results are evaluated on the basis of temporal warp error. Tab. 2 shows the quantitative result produced by the methods proposed in [3, 16, 19] and our method. Please note that there are over 830 video sequences in the evaluation datasets and over 115,000 frames. Due to the extremely tedious and resource-intensive nature of DVP [19], only a portion of the evaluation datasets could be processed and the evaluated metrics in the DVP [19] column are averaged over only the processed videos for each task. In order to have a fair comparison, the metrics evaluated on the videos processed with DVP [19] are excluded from the large-scale quantitative comparison and are only presented for reference only. Furthermore, it is also noteworthy that in their paper [19], the authors do not evaluate the results on the entirety of the dataset and sample tasks for comparative results and report that their warp error is comparable to that of Bonneel et al. [3]. + +The lower temporal warp error represents better temporal consistency. The lowest temporal warp error is highlighted + +in red, and the second best is highlighted in blue. It is worth mentioning that the temporal warp error does not take into account the perceptual quality of the produced frames and assigns a lower value to blurry videos (as produced by [19]). The quality of the results produced by each of the discussed methods can be verified through the accompanying supplementary video. It is evident from Tab. 2 that the proposed model produces state-of-the-art consistency results. Please note that, unlike the proposed method, all of the compared methods require the availability of raw/unprocessed videos for restoring temporal dynamics at both the training and test time. + +We further verify the efficacy of the proposed model through a specialized metric (as presented in [12]) for evaluating motion smoothness in videos, on the portion of evaluation datasets processed through the DVP [19] and present our findings below. + +It is worth mentioning that some attempts to judge the perceptual quality through metrics like mean PSNR (performance degradation [19]) have been proposed for this task. We intentionally disregarded this metric and argue that this metric might be helpful to evaluate the performance of successive iterations of their prior-based method where each iteration leads to progressively better results, but it can be a bit misleading in judging the overall perceptual quality of restored videos for this task. A similar phenomenon can be observed with other local perceptual metrics such as LPIPS [45] as it is impossible to quantize a meaningful deviation of perceptual distance or PSNR from per-frame processed videos due to the lack of ground truth data. + +In order to be consistent with the prior works, we nonetheless present the LPIPS [45] results in 3. The proposed method performs competitively with the method proposed by [16] while achieving better temporal consistency. It is noteworthy that each of the discussed metrics does not take into account the aspects considered by the other metric and only provides limited insights about the results. Therefore, to properly evaluate the models on both perceptual quality and temporal consistency, thorough user studies are essential and are used as the primary metric for the evaluation of this task. We conducted two separate user studies to properly evaluate all of the discussed methods and present our findings in the next sub-section. + +# 5.1.1 Smoothness Metric (SM) + +Due to the unavailability of ground truth videos, the quantitative metric results presented (both in the main paper and this supplemental) can only give us an approximate insight towards the nature of the restored videos. In order to properly evaluate our work, we present further large-scale quantitative results generated on the portion of the evaluation datasets processed through DVP [19] in this section. Fig. 5 + +
TaskDAVISVIDEO
Bonneel[3]Lai[16]OursDVP[19]*Bonneel[3]Lai[16]OursDVP[19]*
WCT [20]/antimony0.21660.04310.04850.33940.21800.04510.04470.2992
WCT [20]/asheville0.20560.03730.0449-0.26800.05140.0485-
WCT [20]/candy0.25880.04350.0497-0.21750.05340.0514-
WCT [20]/feathers0.20960.04670.0508-0.15190.05900.0540-
WCT [20]/sketch0.14510.04990.0511-0.18710.05170.0496-
WCT [20]/wave0.17770.04440.04850.34230.25830.07490.07930.3451
Fast-neural-style [13]/princess0.22380.06460.07440.47820.16100.05550.06250.2566
Fast-neural-style [13]/tudnie0.14730.05030.06090.36600.08000.05140.06040.1487
DBL [10]/expertA0.07950.03790.04950.14530.07420.05040.05840.1349
DBL [10]/expertB0.05300.03590.04810.13610.13590.06390.05780.1947
IID [2]/reflectance0.10110.04930.05360.13890.07830.06590.07310.2635
IID [2]/ shading0.06460.05270.06430.15580.11910.05060.05920.2302
CycleGAN [46]/photo2ukiyoe0.09080.03940.05020.23510.14620.05050.05750.2571
CycleGAN [46]/photo2vangogh0.12300.03900.04940.25650.12240.05190.05510.1701
Colorization [43]0.06160.04020.04680.15550.08670.04950.05660.1441
Colorization [36]0.03290.03730.04750.14740.13170.00810.02200.1558
Average0.13690.04450.05240.2414*0.15230.05210.05560.2167*
+ +Table 3: Quantitative comparison of LPIPS [45]. The comparison presented in this table is limited to the results produced by [3, 16] and the proposed method. Lower LPIPS suggests better perceptual similarity. \*\*" Indicates partially averaged results. + +![](images/a42da043b741f38ce395a808bcfd27963c2b397d954d830bde866d267845bdb5.jpg) +Figure 5: Motion Smoothness Results. This figure presents the quantitative results acquired through the motion smoothness metric as described above. The lower motion smoothness represents smoother transitions in the videos. + +presents the temporal smoothness results generated through the motion smoothness metric as presented in [12]. This metric tries to quantify the smoothness of videos over various window sizes through differential optical flow as follows: + +$$ +\begin{array}{l} D = \left(F _ {I _ {1} \rightarrow I _ {2}} - F _ {I _ {0} \rightarrow I _ {1}}\right) - \left(F _ {O _ {1} \rightarrow O _ {2}} - F _ {O _ {0} \rightarrow O _ {1}}\right), \\ \mathrm {M} (s) = \log \left(\sum_ {d \in D} \mathbf {1} _ {[ s, s + 1)} \left(\| \boldsymbol {d} \| _ {2}\right)\right) - \log (| D |), \tag {8} \\ \end{array} +$$ + +Here, $F_{a\rightarrow b}$ denotes optical flow estimated from frame $a$ to frame $b$ . The preceding $I$ and $O$ refers to the raw/unprocessed and restored frames respectively. In Eq. 8, $d$ denotes each of the $[\nabla u,\nabla v]$ vectors in the differential optical flow and $s$ represents the various window sizes for evaluation and $|D|$ denotes the dimensionality of the differential optical flow. $\mathbf{1}_S(d)$ is the masking factor which is equal to 1 if $d$ belongs to the set $S$ . The resultant $\mathrm{M}(s)$ gives a measure of temporal smoothness over 3 consecutive frames. Fig. 5 presents the average results on the portion of evaluation datasets processed through [3, 16, 19] and the proposed model. A lower score suggests better smoothness. The proposed model again achieves SOTA results in terms of motion smoothness even over bigger window sizes as well. + +Please note that all the presented metrics are subjective + +in nature and do not account for the overall quality and consistency of the results due to the lack of ground truth; therefore, user studies should be considered as the main criteria for judging the results of all the proposed approaches for this task. We discuss the findings of the conducted user studies below. + +# 5.1.2 User Study + +![](images/a3c168b6d3b2e579a30a4930cf5c16755842618c99ed5f577aacee49081251e5.jpg) +Figure 6: User Study. This donut chart highlights the user preferences collected through the user study. A majority of the participants preferred the results produced by the proposed method. + +We conducted two separate user studies; a comprehensive and a factorized user study, to properly evaluate the performance of the proposed model. The first user study consisted of 36 participants with 150 different scenes processed through each of the proposed models. Each participant was asked to judge 5 videos processed through each of the proposed methods for this task. The users were shown a randomly sampled scene from the video pool processed through [3, 16, 19] and this work, simultaneously. The participants were instructed to judge the videos on the basis of naturalness, quality, consistency, content, and style preservation. Figure 6 presents the findings of the first (comprehensive) user study. On average, $41\%$ of the users preferred the videos restored by the proposed method. Furthermore, the results generated by the method proposed in Lai et al. [16] also received a significant preference in the user study but it is worth noting that all the previously proposed approaches require the unprocessed counterparts of the frame-wise processed videos at test time to restore the temporal consistency, whereas the proposed model restores the temporal consistency by only taking into account the inconsistent video at the test time. + +# 5.1.3 Factorized User Study + +Apart from the comprehensive user study, we also conducted a secondary in-depth user study to evaluate the reasoning of participants and their opinion on their selections of videos processed with all the methods on challenging scenarios. + +In this user study, each user was instructed to judge 15 randomly selected videos (5 videos from each of the applications mentioned above) processed by each of the proposed models for this task. Please note that in this user study users were also shown the per-frame processed video in an extended window to judge the videos on content and style preservation as well. Apart from recording their preference, the participants were also instructed to record the reasoning for their selection. $\sim 86\%$ of the participants remarked that they chose the videos based on lesser flickering and better quality preservation. The results of the second user study for each of the selected tasks are provided in Fig. 7. On average, $\sim 47\%$ of the users preferred the videos processed through the proposed model for these extremely challenging tasks. + +# 5.2. Qualitative Results + +Figures $8\sim 10$ present some of the comparative qualitative results produced by the methods proposed in [3, 16, 19] and the proposed model. The results produced by [3] fail to retain the perceptual quality of video sequences where occlusion and deocclusion occur. Whereas the frames processed by DVP [19] are significantly blurry and lose the translation effect. In contrast to the comparison with DVP [19] and Boneel et al. [3], the comparison with the results produced through the method proposed by Lai et al. [16] can be a bit tricky. In order to aid the readers, we provide an epipolar comparison between the frames restored through their method and ours in Fig. 10. Due to the implicit definition of motion dynamics in their method [16], their method fails to faithfully retain the perceptual quality in the restored frames and often flips the colors/appearance of objects in the restored video according to the raw video. This bias towards the raw videos is quite prominent in methods like DVP [19] and Lai et al. [16]. Contrasting to their formulation, the explicit design of motion dynamics in the proposed method mitigates the need for raw videos at test time which consequently eliminates this bias and enables the proposed method to faithfully restore videos. Further qualitative comparisons and higher quality results, along with the results on super resolution [17], inpainting [18], and multiple restoration iterations are presented in the supplemental. + +# 6. Limitations and Future work + +The proposed model has difficulties in restoring videos with very low frame rates and in the case of style-transferred videos, some finer inconsistencies like finer brush strokes are lost, which makes the resulting video seem a little dull (which is to be expected from all the proposed approaches for this task). A good direction for future work in this field can be the direct denoising of conventional optical flow with deterministic warp operators. The developed consis + +![](images/3909694f263ba7677bb389caab3c5d9704199e1305ea1958c1de3cae6c812a7d.jpg) +Figure 7: Factorized user study. The user preference scores for the tasks of style transfer, CycleGAN, and intrinsic image decomposition (left to right respectively). The majority of the users preferred the videos processed through the proposed method in each category of tasks. + +![](images/fd53144caa3953fbb0d72bac85fedc98a92d98d3c407f6e8bf75569569fa54b0.jpg) + +![](images/246c88f8bfb11b50682b25f88bd19e08e7558a745914c72db11177b55eb5ce9a.jpg) + +![](images/c1f3e3962454ad5c55d7908768400076a39caa285da3a28e584f25af40eedaa9.jpg) +Figure 8: Comparison with Bonneel et al. [3] The method proposed by Bonneel et al. [3] has difficulty in handling occlusion and long term consistency and produce temporal artifacts. + +![](images/59bb2b89a8456beded0c4b186847ea19fa7d594e0acc4113f2438f49b02aa5b6.jpg) +Figure 9: Comparison with DVP [19] DVP [19] produces blurry frames and compromises the translation effect. This phenomenon is prominent in videos where new content is generated e.g. style transfer [13], CycleGAN [46]. + +tent flow models can also be used for a number of video processing applications like denoising and dehazing. + +# 7. Conclusion + +In this work, we present a task-agnostic temporal consistency correction framework that restores natural video dynamics by inferring and utilizing consistent motion repre + +![](images/aaf535c0e9477358f4a24cbca6f1b92554edd7f41226cd5b05b3ca2006c135ef.jpg) +Figure 10: Comparison with Lai et al. [16] This figure produces epipolar temporal comparison of videos restored by Lai et al. [16] and the proposed method. The method proposed in [16], due to its dependence on raw videos and implicit siphoning/mixing of motion dynamics from the raw video introduces bias towards the raw videos. + +sentations from temporally inconsistent videos. The proposed strategy mitigates the need for siphoning video dynamics from the unprocessed videos at test time consequently mitigating the compromise in translation effects that is prominent in the currently proposed approaches for this task and expands the scope of its employability to a wider set of applications. Through extensive experimentation and user studies, we demonstrate that despite the limited access to data at test-time the proposed method compares favorably to the pre-existing methods available for this task and achieves SOTA performance. + +# Acknowledgements + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2022-0-00156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D meta-verse transformation) and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06. + +# References + +[1] Muhammad Kashif Ali, Sangjoon Yu, and Tae Hyun Kim. Deep motion blind video stabilization. BMVC, 2021. +[2] Sean Bell, Kavita Bala, and Noah Snavely. Intrinsic images in the wild. ACM TOG, 33(4):1-12, 2014. +[3] Nicolas Bonneel, James Tompkin, Kalyan Sunkavalli, Deqing Sun, Sylvain Paris, and Hanspeter Pfister. Blind video temporal consistency. ACM TOG, 34(6):1-9, 2015. +[4] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent online video style transfer. In ICCV, pages 1105-1114, 2017. +[5] Mengyu Chu, You Xie, Laura Leal-Taixe, and Nils Thuerey. Temporally coherent gans for video super-resolution (tecogan). arXiv preprint arXiv:1811.09393, 1(2):3, 2018. +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009. +[7] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Counting out time: Class agnostic video repetition counting in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10387-10396, 2020. +[8] Wei Gao, Yijun Li, Yihang Yin, and Ming-Hsuan Yang. Fast video multi-style transfer. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3211-3219, 2020. +[9] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015. +[10] Michael Gharbi, Jiawen Chen, Jonathan T Barron, Samuel W Hasinoff, and Frédo Durand. Deep bilateral learning for real-time image enhancement. ACM TOG, 36(4):1-12, 2017. +[11] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, pages 2462-2470, 2017. +[12] Meiguang Jin, Zhe Hu, and Paolo Favaro. Learning to extract flawless slow motion from blurry videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8112-8121, 2019. +[13] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016. +[14] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep blind video decaptioning by temporal aggregation and recurrence. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4258–4267, 2019. +[15] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Recurrent temporal aggregation framework for deep video inpainting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42:1038–1052, 2019. +[16] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018. +[17] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, + +Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, pages 4681-4690, 2017. +[18] Eunhye Lee, Jeongmu Kim, Jisu Kim, and Tae Hyun Kim. Restore from restored: Single-image inpainting. arXiv preprint arXiv:2102.08078, 2021. +[19] Chenyang Lei, Yazhou Xing, and Qifeng Chen. Blind video temporal consistency via deep video prior. NeurIPS, 2020. +[20] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. NeurIPS, 30, 2017. +[21] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. arXiv preprint arXiv:2204.02663, 2022. +[22] Xunyu Lin, Victor Campos, Xavier Giro-i Nieto, Jordi Torres, and Cristian Canton Ferrer. Disentangling motion, foreground and background features in videos. arXiv preprint arXiv:1707.04092, 2017. +[23] Yihao Liu, Hengyuan Zhao, Kelvin CK Chan, Xintao Wang, Chen Change Loy, Yu Qiao, and Chao Dong. Temporally consistent video colorization with deep feature propagation and self-regularization learning. arXiv preprint arXiv:2110.04562, 2021. +[24] Cheng Ma, Yongming Rao, Yean Cheng, Ce Chen, Jiwen Lu, and Jie Zhou. Structure-preserving super resolution with gradient guidance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7769-7778, 2020. +[25] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016. +[26] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In CVPR, pages 4161-4170, 2017. +[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. +[28] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In $GCPR$ , pages 26-36. Springer, 2016. +[29] Mehdi SM Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super-resolution. In $CVPR$ , pages 6626-6634, 2018. +[30] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. NeurIPS, 28, 2015. +[31] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In CVPR, pages 8934-8943, 2018. +[32] Ximeng Sun, Huijuan Xu, and Kate Saenko. Twostreamvan: Improving motion modeling in video generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2744-2753, 2020. + +[33] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419. Springer, 2020. +[34] Hugo Thimonier, Julien Despois, Robin Kips, and Matthieu Perrot. Learning long term style preserving blind video temporal consistency. In ICME, pages 1-6. IEEE, 2021. +[35] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In CVPR, pages 9446-9454, 2018. +[36] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In ECCV, pages 836-850. Springer, 2014. +[37] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127(8):1106-1125, 2019. +[38] Zhicheng Yan, Hao Zhang, Baoyuan Wang, Sylvain Paris, and Yizhou Yu. Automatic photo adjustment using deep neural networks. ACM TOG, 35(2):1-15, 2016. +[39] Yuchen Yang, Ye Xiang, Shuaicheng Liu, Lifang Wu, Boxuan Zhao, and Bing Zeng. Glm-net: Global and local motion estimation via task-oriented encoder-decoder structure. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4211-4219, 2021. +[40] Chun-Han Yao, Chia-Yang Chang, and Shao-Yi Chien. Occlusion-aware video temporal consistency. In ACMMM, pages 777-785, 2017. +[41] Bo Zhang, Mingming He, Jing Liao, Pedro V Sander, Lu Yuan, Amine Bermak, and Dong Chen. Deep exemplar-based video colorization. In CVPR, pages 8052-8061, 2019. +[42] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017. +[43] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, pages 649-666. Springer, 2016. +[44] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, pages 649-666. Springer, 2016. +[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. +[46] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2223-2232, 2017. \ No newline at end of file diff --git a/2023/Task Agnostic Restoration of Natural Video Dynamics/images.zip b/2023/Task Agnostic Restoration of Natural Video Dynamics/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..74b278c95c960da10ac1897631e95133ae0e61d0 --- /dev/null +++ b/2023/Task Agnostic Restoration of Natural Video Dynamics/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f3d8e71f71e262fdd761eb97e650cd6a242d3452939b74a9d05d7794ce00f02 +size 425070 diff --git a/2023/Task Agnostic Restoration of Natural Video Dynamics/layout.json b/2023/Task Agnostic Restoration of Natural Video Dynamics/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fc55bb319b3dbe4fc79e56e8ca52723361f258dd --- /dev/null +++ b/2023/Task Agnostic Restoration of Natural Video Dynamics/layout.json @@ -0,0 +1,7071 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 465, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 465, + 121 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 465, + 121 + ], + "type": "text", + "content": "Task Agnostic Restoration of Natural Video Dynamics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 463, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 463, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 463, + 185 + ], + "type": "text", + "content": "Muhammad Kashif Ali Dongjin Kim Tae Hyun Kim* \nDepartment of Computer Science, Hanyang University, Seoul, Korea \n{kashifali, dongjinkim, taehyunkim}@hanyang.ac.kr" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 239, + 290, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 290, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 290, + 490 + ], + "type": "text", + "content": "In many video restoration/translation tasks, image processing operations are naively extended to the video domain by processing each frame independently, disregarding the temporal connection of the video frames. This disregard for the temporal connection often leads to severe temporal inconsistencies. State-Of-The-Art (SOTA) techniques that address these inconsistencies rely on the availability of unprocessed videos to implicitly siphon and utilize consistent video dynamics to restore the temporal consistency of frame-wise processed videos which often jeopardizes the translation effect. We propose a general framework for this task that learns to infer and utilize consistent motion dynamics from inconsistent videos to mitigate the temporal flicker while preserving the perceptual quality for both the temporally neighboring and relatively distant frames without requiring the raw videos at test time. The proposed framework produces SOTA results on two benchmark datasets, DAVIS and videoo.net, processed by numerous image processing applications. The code and the trained models are available at https://github.com/MKashifAli/TARONVD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 536, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 288, + 693 + ], + "type": "text", + "content": "Video sharing social media platforms like Snapchat and TikTok have introduced the common populace to a plethora of computer vision applications such as Style Transfer [9], Colorization [43], Denoising [42], and Dehazing [38]. With this wide-scale integration of these classical computer vision applications in such platforms, various image processing operations are naively extended to videos due to scarcity of annotated video datasets and their computational complexity of video processing methodologies. This naive extension of static image processing methodologies to videos disregards the temporal connection of the video progression and introduces severe temporal flickering in the videos. This temporal flicker can appear for various reasons; for in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 213, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 346 + ], + "type": "text", + "content": "stance, these image processing methods can produce drastically different results for temporally neighboring frames due to slight changes in their global or local content statistics, or it could also happen due to the multi-modality of the application as there could exist a number of valid solutions for images with similar content as highlighted in [3, 41]. Therefore, the extension of these image-to-image translation tasks to the video domain is an active area of research in computer vision, and methods that can help extend various image processing applications with little to no knowledge of the operations used to create videos are quite useful." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 350, + 548, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 548, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 548, + 602 + ], + "type": "text", + "content": "Currently, there are several task-dependent temporal consistency correction approaches available such as [4, 23, 34], but due to the complex nature of these applications, only a handful of approaches have been proposed to tackle the problem of blind temporal consistency correction. Bonneel et al. [3] initiated the investigation of blind temporal consistency correction with gradient-domain minimization of per-frame processed video with the unprocessed video to minimize the warping error between the frames. Lai et al. [16] extended the formulation of [3] with the help of recurrent Convolutional Neural Networks (CNN) and introduced a perceptual penalty in their formulation to restrict the deviation of perceptual content of the restored video from the frame-wise processed video. Deep Video Prior (DVP) [19] extended Deep Image Prior (DIP) [35] to the temporal dimension and proposed to formulate enforcing temporal consistency by training a CNN to generate processed video from the unprocessed video without utilizing optical flow. All of these approaches rely on the availability of unprocessed videos for implicit extraction of consistent motion dynamics to use as a restoration guide." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "Please note that all the previously proposed approaches for this task define it with the help of unprocessed videos (both task-specific and task-agnostic). Having access to unprocessed videos at test time helps the models in siphoning and transferring consistent motion dynamics to the inconsistent videos to decrease temporal flicker. Despite its efficacy, this implicit definition limits the applicability of the previously proposed approaches to only the videos for which their unprocessed counterparts are available and also" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 702, + 129, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 702, + 129, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 702, + 129, + 712 + ], + "type": "text", + "content": "*: Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13534" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 288, + 148 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 288, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 288, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 288, + 148 + ], + "type": "table", + "html": "
Key featuresBonneel et al. [3]Lai et al. [16]DVP [19]Ours
Does not require raw videos×××✓
Explicit video dynamics design✓××✓
Faithful perceptual restoration×××✓
Sequential processing✓✓×✓
Long term consistency××✓✓
Sharpness in processed videos×✓×✓
High-frequency flicker removal✓×✓✓
Test time training××✓×
Runtime (per frame avg.)2.4635*0.21463.63650.2236
", + "image_path": "7e3e6137179f94d068e6ef5086a0cf3bf1bf337ae6c040a1908a68d9a62dd432.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 156, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 287, + 216 + ], + "type": "text", + "content": "Table 1: Key differences in available approaches. \\*\\*\" in front of Bonnel et al. [3]'s runtime highlights environment difference and is only presented for reference only. Please note that the runtime is averaged over the same video sequence for all the methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 238, + 287, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 287, + 274 + ], + "type": "text", + "content": "introduce an inherent bias towards the unprocessed videos which can compromise the translation effect and quality of the processed videos (as presented in Fig. " + }, + { + "bbox": [ + 46, + 238, + 287, + 274 + ], + "type": "inline_equation", + "content": "8\\sim 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 274, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 287, + 489 + ], + "type": "text", + "content": "In order to overcome these limitations, we aim to learn and infer consistent motion representation solely from temporally inconsistent videos. Doing so does not only alleviate the need for raw videos at test time, but also mitigates the inherent bias towards the raw videos that is commonly present in the currently available approaches for this task. For the task of learning these consistent motion representation, we fine-tune conventional optical flow estimations networks like [11, 31, 33] along with the restorative part of the network in an end-to-end manner. The proposed model pipeline is illustrated in Fig. 1. The proposed model achieves state-of-the-art qualitative and quantitative results. The proposed formulation can also handle the resolution mismatch problem in processed and raw videos and makes it possible to extend the Single Image Super-Resolution [17] methods to Video Super-Resolution methods without any modification. The detailed description of our formulation is presented in 3. We summarize our contributions as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 490, + 286, + 584 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 490, + 286, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 286, + 523 + ], + "type": "text", + "content": "- Identify and propose tailored solutions for various challenges of naïve extensions of image processing applications to videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 525, + 286, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 286, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 286, + 548 + ], + "type": "text", + "content": "- Learn to infer consistent motion representations from temporally inconsistent videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 549, + 286, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 286, + 584 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 286, + 584 + ], + "type": "text", + "content": "- Utilize the learned consistent motion representations to propose a general framework for task agnostic temporal consistency correction that produce SOTA results." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "The literature is divided into two main streams to generate visually appealing videos through image transformation models. The first stream tackles the task of mitigating temporal inconsistencies with the reformulation of the task at hand. Whereas, The second stream focuses on developing post-processing models that refine the frame-wise processed videos by penalizing the temporal deviation of the processed video from its unprocessed counterpart. The" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "second stream is further divided into two sub-streams: task-specific and task-agnostic. The following passages summarize the details of both of the above-mentioned streams, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 121, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 335 + ], + "type": "text", + "content": "Reformulation stream (video-to-video translation): These approaches are generally termed video-to-video translation applications. Generally, these approaches either consider multiple frames as input and produce multiple output frames or generate a single frame from multiple input frames in the form of frame recurrence such as [5, 23]. There are also cases like video style transfer where content information is propagated to the next time step with the help of optical flow to initiate the optimization of the next frame. This frame-recurrent methodology has also been proven effective in applications like video super resolution [29]. Designing these approaches for each task and training them from scratch is a taxing task and data scarcity can make these formulations unfeasible. Generally, such models often do not adapt well to different tasks; therefore, methods that can restore the temporal consistency of multiple frame-wise processed videos are being actively investigated." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 335, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 574 + ], + "type": "text", + "content": "Post-processing methods: Bonneel et al. [3] initiated the investigation of task agnostic temporal consistency correction using a gradient-based optimization strategy in which temporally consistent (unprocessed) videos were used as restoration guides for correcting the temporal inconsistency of frame-wise processed videos. Their formulation motivated various task-dependent and task-agnostic, and temporal consistency correction approaches with slight variations of their original objective formulation. Despite the simplicity of their formulation, there only exist a handful of approaches that address task-agnostic temporal consistency correction due to the difficult nature of the task. Yao et al. [40] proposed a key-frame strategy that accounted for the motion of different objects in those key-frames to handle occlusion as well as temporal consistency. Lai et al.[16] proposed the first deep learning approach for this task by employing ConvLSTM [30] and a perceptual loss [13]. Lei et al. [19] proposed an extension of Deep Image Prior [35] and demonstrated its capability to mitigate the temporal flicker of per-frame processed videos." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 578, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 662 + ], + "type": "text", + "content": "Apart from these general frameworks, various task-specific approaches have also been proposed, such as [5, 23, 34]. Defining these task-specific approaches is relatively straightforward. These approaches define a temporal extension around a backbone method and penalize the deviation of generated content with the help of motion information from the unprocessed videos." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "In this work, we propose a general formulation for this task that alleviates the requirement of unprocessed videos at test time by learning to infer a consistent motion representation from the temporally inconsistent videos. This inferred" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13535" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "motion representation is then processed with the content of temporally inconsistent videos in a two branched network that restores the natural dynamics of per-frame processed videos. Our proposed network also contains a recurrent bottleneck module that helps consistent propagation of content throughout the video sequence and produces SOTA qualitative and quantitative results by only taking into account the temporally inconsistent videos at test time." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 287, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 287, + 205 + ], + "type": "text", + "content": "In order to aid the readers, we provide the following table to highlight the key differences between the currently available approaches for this task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 217, + 153, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 153, + 230 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 153, + 230 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 248, + 287, + 404 + ], + "blocks": [ + { + "bbox": [ + 48, + 248, + 287, + 404 + ], + "lines": [ + { + "bbox": [ + 48, + 248, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 248, + 287, + 404 + ], + "type": "image", + "image_path": "01617768a5374400703e76106386427c83011b67d8f0b9dab46041c43f9df958.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 423, + 287, + 509 + ], + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 509 + ], + "type": "text", + "content": "Figure 1: Model Architecture. An illustration of the proposed model architecture. The proposed model consists of a two-branch architecture that separately processes motion and content. The combined motion and content features concatenated along the channel dimension and are then passed through a Recurrent Bottleneck Block (RBB) and a decoder to generate the restored frame." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "content": "In this section, we describe the details of the proposed method. Consider a raw (temporally consistent) video " + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\{I_0, I_1, I_2, \\dots, I_n\\}" + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "content": " frames and its frame-wise processed (temporally inconsistent) video " + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\{P_0, P_1, P_2, \\dots, P_n\\}" + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "content": " acquired by an image processing function " + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\{h(I_0), h(I_1), h(I_2), \\dots, h(I_n)\\}" + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "content": ". The goal of this task is to restore the temporal consistency of the temporally inconsistent video and to produce a temporally consistent version " + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\{O_0, O_1, O_2, \\dots, O_n\\}" + }, + { + "bbox": [ + 46, + 521, + 287, + 653 + ], + "type": "text", + "content": ". For the sake of simplicity, these notations are chosen to be in line with the notations used in past literature [3, 16, 19]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "The proposed formulation draws inspiration from the disentanglement of a video into its base components: content and motion representations as presented in [22]. In their work [22], they investigated the controlled generation of short length videos by mixing motion and con" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 69, + 545, + 110 + ], + "blocks": [ + { + "bbox": [ + 309, + 69, + 545, + 110 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 545, + 110 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 545, + 110 + ], + "type": "image", + "image_path": "07563e715b481b3ceacaa6abe2fbf34ed3d98d29043097b40b254c8d232f2402.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 127, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 304, + 127, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 547, + 175 + ], + "type": "text", + "content": "Figure 2: Temporal flicker in optical flow. An illustration of the decomposition of optical flow generated from temporally inconsistent frames to consistent optical flow and inconsistent flicker." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 196, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 545, + 352 + ], + "type": "text", + "content": "tent components from various videos. In this work, we explore a similar concept of controlled disentanglement and re-entanglement strategy to restore natural video dynamics of the frame-wise processed videos. More specifically, we argue that the previously proposed approaches for this tasks implicitly employ this disentanglement and re-entanglement in their formulation with the help of unprocessed video frames. Conventionally, optical flow is used as an inter-frame motion representations. The estimated optical flow can also be used to approximate the content of future frames with the help of past frames provided that there does not exist occlusion between the frames (as presented in Eq. 1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 378, + 351, + 545, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 351, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 378, + 351, + 545, + 363 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {t + 1} \\approx w \\left(I _ {t}, o f _ {t + 1 \\Rightarrow t}\\right). \\tag {1}", + "image_path": "f4185ef46409bb80b9bdc2706641bdb75199a0b56b7eaf3ee3cc1d2311126ec4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 367, + 546, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 546, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 546, + 641 + ], + "type": "text", + "content": "This interdependence of optical flow and content can lead to inconsistent motion representation as optical flow estimation networks are not trained explicitly to distinguish temporally consistent content from temporal flicker. Due to this limitation, the evaluated optical flow estimated from temporally inconsistent frames is often compromised with temporal flicker (as presented in Fig. 2). This compromise of the motion information makes it challenging to separate the consistent component of optical flow from the additive flicker. Therefore, the currently available techniques for this task often rely on siphoning implicit motion representations from the raw videos and often suffer in cases where the content of both processed and raw videos deviate significantly. Another problem with employing these implicit representations is the bias introduced due to the lack of control in mixing these implicit representations to restore videos which often leads to the mitigation of translation effect. Therefore, if a mechanism can be developed that produces similar motion representations from both unprocessed and framewise processed videos, better and unbiased restorative effects than those produced by the previously proposed approaches can be achieved without relying on the availability of unprocessed videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "Please note that another observation that can be made from Fig. 2 (right most figure) is that a significant amount of flicker in frame-wise processed videos, generally occurs near the motion boundaries. This observation helps us in defining objective functions capable of localizing and targeting temporal deviations as presented in Sec. 4.1." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13536" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "To overcome these problems and to learn and utilize consistent motion representation, we integrate an optical flow estimation network into the restoration network as illustrated in Fig. 1. A similar integration of a small-scale optical flow estimation network (SpyNet [26]) for task-oriented flow estimation is presented in [37] for the task video denoising for its extremely small size. We utilize a relatively larger network (PWC-Net [31]) in our method and provide an ablation study on different sized optical flow models in the accompanying supplementary text. We further extend their idea to a two pass mechanism with an added encoder to encourage generation of consistent dynamics. The optical flow network in the proposed methodology is finetuned in an end-to-end manner without any special supervision." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 247, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 289, + 715 + ], + "type": "text", + "content": "The proposed model takes in three frames " + }, + { + "bbox": [ + 46, + 247, + 289, + 715 + ], + "type": "inline_equation", + "content": "O_{t-1}" + }, + { + "bbox": [ + 46, + 247, + 289, + 715 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 247, + 289, + 715 + ], + "type": "inline_equation", + "content": "P_{t-1}" + }, + { + "bbox": [ + 46, + 247, + 289, + 715 + ], + "type": "text", + "content": " for the first time step), " + }, + { + "bbox": [ + 46, + 247, + 289, + 715 + ], + "type": "inline_equation", + "content": "P_t" + }, + { + "bbox": [ + 46, + 247, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 247, + 289, + 715 + ], + "type": "inline_equation", + "content": "P_{t+1}" + }, + { + "bbox": [ + 46, + 247, + 289, + 715 + ], + "type": "text", + "content": " as input. These three frames are propagated through both the content and motion branches. The proposed model consists of a UNet [27] like structure with multiple encoder streams and a single decoder stream. The decoder contains skip connections from the content stream to encourage better reconstruction. The bottleneck part of the model contains a recurrent bottleneck block to transmit and manipulate the information of the generated frames to temporally distant frames. Please note that the multi-branch encoder(s)-bottleneck-decoder architecture used for the proposed model is employed due to its proven efficacy in video generation, translation and restoration tasks as described in [4, 8, 14, 15, 16, 23, 32, 34]. The main difference in the proposed architecture comes from the design of our motion branch which allows us to present a standalone solution for this task. The motion branch comprises of two passes through a PwCNet [31] with shared weights followed by a conventional encoder-like architecture. An illustration of the proposed model is presented in Fig. 1. This two-pass strategy in the motion branch along with the attached encoder allows the proposed model in regressing consistent motion dynamics from temporally inconsistent frames. Further architectural details of the proposed network are presented in the accompanying supplemental. Figure 3 presents conventional optical flow along with the learned motion representation generated by the integrated finetuned PwC-Net [31] from temporally consistent (left) and inconsistent frames (right). Please note that the \"modified\" flow can differ starkly from the conventional flow and can only be interpreted effectively by the network it is trained with as described in [37]. Contrasting to the conventional optical flow modified flow estimated from both consistent and inconsistent videos is quite similar, which suggests that the network can effectively utilize the inferred consistent motion representation to restore natural temporal dynamics in the per-frame videos by only taking into account the temporally inconsistent videos at test time. Another benefit of this disentangled representation" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 332, + 72, + 525, + 174 + ], + "blocks": [ + { + "bbox": [ + 332, + 72, + 525, + 174 + ], + "lines": [ + { + "bbox": [ + 332, + 72, + 525, + 174 + ], + "spans": [ + { + "bbox": [ + 332, + 72, + 525, + 174 + ], + "type": "image", + "image_path": "afb0d404e443e7db0b81ac4e7feb6d31e177f021f3e9caa503fa41bb70d840b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 181, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 304, + 181, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 547, + 243 + ], + "type": "text", + "content": "Figure 3: Conventional and Modified flow like representation. Comparison of original and Modified flow evaluated from temporally consistent (left) and inconsistent (right) frames. The modified flow evaluated from both the temporally consistent and inconsistent frames is quite similar." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 250, + 547, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 250, + 547, + 333 + ], + "spans": [ + { + "bbox": [ + 304, + 250, + 547, + 333 + ], + "type": "text", + "content": "is that it allows the users to further improve the level of restored temporal consistency with the help of an iterative arrangement; please refer to the supplemental for the details of the iterative arrangement and the results generated with such arrangement. The objective functions used for the end-to-end training of the proposed network are described in detail in the next section." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 344, + 424, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 344, + 424, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 344, + 424, + 358 + ], + "type": "text", + "content": "4. Objective Functions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 364, + 546, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 546, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 546, + 425 + ], + "type": "text", + "content": "This section provides the details of the learning objectives used in the training phase of the proposed network. These objectives can be classified into two categories, local neighborhood and long-term objectives. Both of these optimization categories are presented below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 432, + 458, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 432, + 458, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 432, + 458, + 445 + ], + "type": "text", + "content": "4.1. Local Neighborhood Losses" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "In this work, we aim to find a general solution for correcting the temporal consistency of frame-wise processed videos. Due to the diversity of the applications addressed, a simplistic spatial content matching reconstruction loss cannot justify the generation of temporally consistent frames, as the content can vary significantly in applications like style transfer [9] and image adjustment [38]. This discrepancy in the image space of processed and raw videos hinders a straightforward definition of a feasible loss function for this task. Due to the challenging nature of this task, we propose a flow gradient loss that provides a supervision signal for the reconstruction of motion boundaries (where most of the inconsistencies/flicker appear; as highlighted in Fig. 2) with the help of optical flow acquired through the unprocessed videos. We further extend this loss to only compare the gradients of the optical flow acquired through the synthesized and raw frames. Doing so reduces the redundancy present in the high dimensional optical flow which contains a redundant global component for this task which is only useful in applications like video stabilization [1] as described in [39]. The proposed flow gradient loss encapsulates the local spatio-temporal information necessary" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13537" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "to mitigate temporal flicker near the motion boundaries by removing the unnecessary high-dimensional components from the optical flow. Similar information bottle-necking techniques have been discussed thoroughly in vision based understanding and manipulation literature [7, 24]. The proposed spatio-temporal loss is defined as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 152, + 287, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 287, + 183 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f g} = \\sum_ {t = 2} ^ {T} \\left\\| \\nabla \\left(o f \\left(O _ {t}, O _ {t - 1}\\right)\\right), \\nabla \\left(o f \\left(I _ {t}, I _ {t - 1}\\right)\\right) \\right\\| _ {1}. \\tag {2}", + "image_path": "9a900f4e735c0e3c38661d016029be25bdf1f740079fd2074aee0eb6984e50aa.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 189, + 289, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 189, + 289, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 289, + 453 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 189, + 289, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fg}" + }, + { + "bbox": [ + 46, + 189, + 289, + 453 + ], + "type": "text", + "content": " and of represents flow gradient loss and the optical flow estimation network (FlowNet2.0 [11]) respectively. Please note that any optical flow estimation network (SpyNet [26], PWC-Net [31], RAFT [33], etc.) can be used in both the motion branch and the definition of the loss functions, we choose FlowNet2.0 [11] for its relatively lower mean end-point-error, as the estimated optical flow can vary significantly for small scale networks (comparison provided in the supplementary text). In equation 2, the " + }, + { + "bbox": [ + 46, + 189, + 289, + 453 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 46, + 189, + 289, + 453 + ], + "type": "text", + "content": " operator denotes the spatial gradient of the estimated optical flow. A similar loss in optical flow domain has been proposed in [21] where the raw optical flow of the generated video was compared with that of the ground truth video. Although effective for the task of video inpainting, the raw optical flow matching approach fails in applications like style transfer in which the entire content of the generated video and the raw video differ greatly, i.e., contains excessive spatio-temporal variations which contribute to optical flow estimation. In our experiments, the models trained with raw optical flow performed poorly, whereas the model trained with flow gradients converged in a relatively shorter time and produced sharper frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": "Please note that this optical flow-based loss by itself is not sufficient for faithfully correcting the temporal consistency, as there could exist multiple solutions for optical flow equation, e.g., it does not take into account the appearance of the synthesized and raw frames. In order to address this issue and to guide the network for the generation of perceptually credible frames, a non-local optical flow based loss in the image space is employed as presented below:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 556, + 287, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 556, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 107, + 556, + 287, + 586 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\sum_ {t = 2} ^ {T} C _ {t} \\left\\| O _ {t} - \\hat {O} _ {t} \\right\\| _ {1}, \\tag {3}", + "image_path": "6125983f71156f471caaae37e2680c08c67c05ca455f93c11be604d16f4caee4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "This warping based reconstruction loss function has been adopted from the [4], as many variants of this objective function with slight modifications have been proven effective for enforcing temporal smoothness in video translation tasks in [3, 5, 16, 23, 34]. Here " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " represents the total number of frames in a sequence and " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{O}_t" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " represents the warped frame " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "O_{t-1}" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " (as described in Eq. 1). The working of this loss function can be summarized as the propagation of content from preceding frames to the subsequent frames in image domain to ensure that the generated frames contain similar" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "content to the previous frames and the inter-frame motion dynamics of the synthesized frames are similar to that of the temporally consistent counterpart. This non-local optical flow based loss function also takes into account the occlusion problem of the Eq. 1 by masking the occluded and de-occluded content as defined below:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 368, + 154, + 546, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 154, + 546, + 178 + ], + "spans": [ + { + "bbox": [ + 368, + 154, + 546, + 178 + ], + "type": "interline_equation", + "content": "C _ {t} = \\exp \\left(- \\alpha \\left\\| I _ {t} - \\hat {I} _ {t} \\right\\| _ {2} ^ {2}\\right). \\tag {4}", + "image_path": "aa7bd6f4381d60c3a048dee9a958394efd84cccbf91f21b053b9aa4169925596.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 181, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 545, + 217 + ], + "type": "text", + "content": "The value for " + }, + { + "bbox": [ + 304, + 181, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\alpha = 50" + }, + { + "bbox": [ + 304, + 181, + 545, + 217 + ], + "type": "text", + "content": " is chosen according to the previous works where a similar strategy is used for evaluating occlusion [34, 28]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 217, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 266 + ], + "type": "text", + "content": "An additional conventional short-term perceptual similarity loss [13] was also introduced in the training phase to minimize the deviation of the synthesized frames from the original processed frames. This loss is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 356, + 276, + 546, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 276, + 546, + 299 + ], + "spans": [ + { + "bbox": [ + 356, + 276, + 546, + 299 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\sum_ {t} \\sum_ {l} \\| \\phi_ {l} (O _ {t}) - \\phi_ {l} (P _ {t}) \\| _ {1}, \\tag {5}", + "image_path": "c79c7b80b678bef6cec9f73fc1c5f550696a933be53cbab172e117b3cdd132d7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 302, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 545, + 327 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 302, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\phi_l(.)" + }, + { + "bbox": [ + 304, + 302, + 545, + 327 + ], + "type": "text", + "content": " represents layers of a VGG-16 network till the layer " + }, + { + "bbox": [ + 304, + 302, + 545, + 327 + ], + "type": "inline_equation", + "content": "relu_{-4.3}" + }, + { + "bbox": [ + 304, + 302, + 545, + 327 + ], + "type": "text", + "content": " (trained on the ImageNet dataset [6])." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 335, + 450, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 450, + 348 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 450, + 348 + ], + "type": "text", + "content": "4.2. Temporal Constancy Loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 354, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 715 + ], + "type": "text", + "content": "The loss defined in Eq. 3 adequately defines the content propagation from the previous frame to the current frame but lacks the ability to enforce it in temporally distant instances. For instance, consider the task of colorization where a car appears in a portion of the frames and is assigned some random color, and it disappears and re-appears in a later instance. The frame-wise colorization method can assign it a different color in each sequence, and the loss function defined in Eq. 3 will be sufficient to enforce the color constancy of the car in each interval where the car is visible. This, however, does not ensure the constancy of the color assigned to the car in both intervals, i.e., the car could be assigned a solid blue color in the first interval and a solid red color in the next interval which is quite prominent in methods like [16, 34]. To address this problem of temporally distant yet similar instances, we introduce a recurrent module (ConvLSTM [30]) in our proposed model and add explicit constraints that force the model to generate temporally consistent frames. The previously proposed approaches for this task simply try to minimize the deviation of current frames from the first frame of the sequence, which is inadequate for videos containing largely varying frames. This variation of the content can render this loss ineffective, and the inconsistencies can only be weakly penalized with the help of motion dynamics siphoned from the raw videos. Therefore, we introduce a bi-directional penalty that enforces the temporal smoothness in both backward (through the local neighborhood losses) and forward (through the proposed long-term loss) direction. This bi-directional optimization strategy is inspired from the ping" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13538" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 106 + ], + "type": "text", + "content": "pong loss dynamic as presented in [5]. The proposed long-term loss is termed as constancy loss and is presented below:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 125, + 286, + 164 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 125, + 286, + 164 + ], + "spans": [ + { + "bbox": [ + 50, + 125, + 286, + 164 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n s t a n c y}} = \\sum_ {p = 1} ^ {T - 2} \\sum_ {t = p + 2} ^ {T} C _ {t \\Rightarrow p} \\| O _ {t} - w \\left(O _ {p}, o f \\left(I _ {t - 1}, I _ {p}\\right)\\right) \\| _ {1}. \\tag {6}", + "image_path": "3ded3dfd7693e68ef1d817d1e55eb5af78d6471ec67139a32fd15b78d5396c3f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 165, + 287, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 165, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 165, + 287, + 320 + ], + "type": "text", + "content": "Here subscript " + }, + { + "bbox": [ + 46, + 165, + 287, + 320 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 165, + 287, + 320 + ], + "type": "text", + "content": " highlights that the preceding distant images are compared with all the subsequent frames. Unlike the simplistic long-term temporal constraints used in [16, 34] which only penalized the deviation of content from the first frame, this loss term takes into account the inter-frame flicker of all the frames in a sequence and provides a strong supervision signal capable of penalizing temporal discrepancies throughout the sequence. Due to this stronger temporal penalty accompanied with the proposed flow gradient loss, the proposed network learns to produce smoother videos despite having access to only the frame-wise processed videos at test time. The final loss for each training sequence is given by," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 327, + 287, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 327, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 52, + 327, + 287, + 340 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {f g} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {3} \\mathcal {L} _ {p} + \\lambda_ {4} \\mathcal {L} _ {\\text {c o n s t a n c y}}. \\tag {7}", + "image_path": "c1daabc803d77ef2b4f48a03138a1ba3b8db8ad10273a1f9801e47b0c8409de7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 348, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 384 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 46, + 348, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\lambda(s)" + }, + { + "bbox": [ + 46, + 348, + 287, + 384 + ], + "type": "text", + "content": " in the equation above define the contribution of each loss in the optimization phase. The details of these hyperparameters are provided in the supplementary text." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 394, + 190, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 190, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 190, + 407 + ], + "type": "text", + "content": "5. Experiments and Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 415, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 664 + ], + "type": "text", + "content": "During our experimentation phase, we tested various optical flow estimation networks such as [11, 26, 31]. Flownet2.0 [11] has approximately 162M trainable parameters, and it produced relatively better results in lesser training iterations as compared to the other two variants. Having a vast number of parameters compromises both the training and testing time. Therefore, we opted for a medium-sized optical flow estimation network (PWC-Net [31]) and a higher number of optimization iterations. The results produced by both variants are discussed in the supplementary text. We also experimented with an iterative arrangement of the trained models where restored videos are again subjected to temporal consistency correction models and evaluated that the proposed model architecture consistently reduces warp error with every iteration (as presented in Fig. 4). Generally, the frames subjected to a high number of consistency correction iterations lose fine details. The generated results by various restoration iterations are provided in the supplemental. Please note that the comparative results provided in this paper are generated through a single restoration iteration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "The dataset for training the proposed model contains videos from the DAVIS dataset for video segmentation [25] and videos gathered from Videovo.net by [16]. The training dataset contains videos processed through a diverse range" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 323, + 64, + 529, + 160 + ], + "blocks": [ + { + "bbox": [ + 323, + 64, + 529, + 160 + ], + "lines": [ + { + "bbox": [ + 323, + 64, + 529, + 160 + ], + "spans": [ + { + "bbox": [ + 323, + 64, + 529, + 160 + ], + "type": "image", + "image_path": "dac6bec8e5826318e9aaef88af745197141633817afe45eae5c7eb9b4b555461.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 166, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 305, + 166, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 166, + 547, + 201 + ], + "type": "text", + "content": "Figure 4: Temporal warp error vs. iterations. Temporal warp error consistently decreases with a higher number of restoration iterations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 211, + 546, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 211, + 546, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 546, + 269 + ], + "type": "text", + "content": "of applications such as, Artistic Style Transfer [9, 13], Colorization [36, 44], Image Enhancement [10], Intrinsic Image Decomposition [2], and Image-to-Image Translation tasks [20, 44, 46]. The qualitative and quantitative results are described below." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 277, + 424, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 424, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 424, + 289 + ], + "type": "text", + "content": "5.1. Quantitative Results" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 308, + 304, + 547, + 409 + ], + "blocks": [ + { + "bbox": [ + 308, + 304, + 547, + 409 + ], + "lines": [ + { + "bbox": [ + 308, + 304, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 547, + 409 + ], + "type": "table", + "html": "
TaskDAVISVIDEO
Bonneel[3]Lai[16]OursDVP[19]*Bonneel[3]Lai[16]OursDVP[19]*
WCT [20]/antimono0.00290.00310.00260.00220.00150.00210.00150.0017
WCT [20]/asheville0.00470.00590.0047-0.00320.00430.0039-
WCT [20]/candy0.00340.00470.0035-0.00200.00320.0025-
WCT [20]/feathers0.00400.00400.0027-0.00270.00300.0021-
WCT [20]/sketch0.00360.00290.0021-0.00250.00220.0017-
WCT [20]/wave0.00330.00350.00270.00260.00230.00260.00210.0019
Fast-neural-style [13]/princess0.00430.00630.00420.00380.00350.00530.00420.0046
Fast-neural-style [13]/ludnie0.00210.00230.00220.00220.00120.00140.00140.0022
DBL [10]/expertA0.00170.00110.00100.00150.00140.00070.00060.0015
DBL [10]/expertB0.00160.00100.00080.00140.00110.00060.00040.0014
IID [2]/reflectance0.00150.00080.00070.00120.00110.00060.00060.0006
IID [2]/shading0.00140.00080.00080.00130.00080.00040.00040.0017
CycleGAN [46]/photo2ukiyoe0.00240.00190.00150.00130.00170.00130.00100.0018
CycleGAN [46]/photo2vangogh0.00260.00260.00190.00140.00200.00200.00150.0020
Colorization [43]0.00160.00110.00080.00130.00100.00060.00040.0012
Colorization [36]0.00150.00090.00070.00130.00100.00050.00030.0012
Average0.00270.00270.00210.0018*0.00180.00190.00150.0018*
", + "image_path": "b5292eca0715747a6fe4a9079fa865df8fbe741c3e58b25b2a7f27a767178d53.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 418, + 547, + 478 + ], + "lines": [ + { + "bbox": [ + 305, + 418, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 547, + 478 + ], + "type": "text", + "content": "Table 2: Quantitative comparison of Temporal Warping Error. The comparison presented in this table is limited to the results produced by [3, 16] and the proposed method. Lower warp error suggests better temporal consistency. *** Indicates partially averaged results." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 486, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 689 + ], + "type": "text", + "content": "The quantitative results are evaluated on the basis of temporal warp error. Tab. 2 shows the quantitative result produced by the methods proposed in [3, 16, 19] and our method. Please note that there are over 830 video sequences in the evaluation datasets and over 115,000 frames. Due to the extremely tedious and resource-intensive nature of DVP [19], only a portion of the evaluation datasets could be processed and the evaluated metrics in the DVP [19] column are averaged over only the processed videos for each task. In order to have a fair comparison, the metrics evaluated on the videos processed with DVP [19] are excluded from the large-scale quantitative comparison and are only presented for reference only. Furthermore, it is also noteworthy that in their paper [19], the authors do not evaluate the results on the entirety of the dataset and sample tasks for comparative results and report that their warp error is comparable to that of Bonneel et al. [3]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "text", + "content": "The lower temporal warp error represents better temporal consistency. The lowest temporal warp error is highlighted" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13539" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "in red, and the second best is highlighted in blue. It is worth mentioning that the temporal warp error does not take into account the perceptual quality of the produced frames and assigns a lower value to blurry videos (as produced by [19]). The quality of the results produced by each of the discussed methods can be verified through the accompanying supplementary video. It is evident from Tab. 2 that the proposed model produces state-of-the-art consistency results. Please note that, unlike the proposed method, all of the compared methods require the availability of raw/unprocessed videos for restoring temporal dynamics at both the training and test time." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 217, + 288, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 288, + 277 + ], + "type": "text", + "content": "We further verify the efficacy of the proposed model through a specialized metric (as presented in [12]) for evaluating motion smoothness in videos, on the portion of evaluation datasets processed through the DVP [19] and present our findings below." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 277, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 433 + ], + "type": "text", + "content": "It is worth mentioning that some attempts to judge the perceptual quality through metrics like mean PSNR (performance degradation [19]) have been proposed for this task. We intentionally disregarded this metric and argue that this metric might be helpful to evaluate the performance of successive iterations of their prior-based method where each iteration leads to progressively better results, but it can be a bit misleading in judging the overall perceptual quality of restored videos for this task. A similar phenomenon can be observed with other local perceptual metrics such as LPIPS [45] as it is impossible to quantize a meaningful deviation of perceptual distance or PSNR from per-frame processed videos due to the lack of ground truth data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 434, + 288, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 288, + 590 + ], + "type": "text", + "content": "In order to be consistent with the prior works, we nonetheless present the LPIPS [45] results in 3. The proposed method performs competitively with the method proposed by [16] while achieving better temporal consistency. It is noteworthy that each of the discussed metrics does not take into account the aspects considered by the other metric and only provides limited insights about the results. Therefore, to properly evaluate the models on both perceptual quality and temporal consistency, thorough user studies are essential and are used as the primary metric for the evaluation of this task. We conducted two separate user studies to properly evaluate all of the discussed methods and present our findings in the next sub-section." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 609, + 187, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 187, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 187, + 621 + ], + "type": "text", + "content": "5.1.1 Smoothness Metric (SM)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Due to the unavailability of ground truth videos, the quantitative metric results presented (both in the main paper and this supplemental) can only give us an approximate insight towards the nature of the restored videos. In order to properly evaluate our work, we present further large-scale quantitative results generated on the portion of the evaluation datasets processed through DVP [19] in this section. Fig. 5" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 547, + 175 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 547, + 175 + ], + "type": "table", + "html": "
TaskDAVISVIDEO
Bonneel[3]Lai[16]OursDVP[19]*Bonneel[3]Lai[16]OursDVP[19]*
WCT [20]/antimony0.21660.04310.04850.33940.21800.04510.04470.2992
WCT [20]/asheville0.20560.03730.0449-0.26800.05140.0485-
WCT [20]/candy0.25880.04350.0497-0.21750.05340.0514-
WCT [20]/feathers0.20960.04670.0508-0.15190.05900.0540-
WCT [20]/sketch0.14510.04990.0511-0.18710.05170.0496-
WCT [20]/wave0.17770.04440.04850.34230.25830.07490.07930.3451
Fast-neural-style [13]/princess0.22380.06460.07440.47820.16100.05550.06250.2566
Fast-neural-style [13]/tudnie0.14730.05030.06090.36600.08000.05140.06040.1487
DBL [10]/expertA0.07950.03790.04950.14530.07420.05040.05840.1349
DBL [10]/expertB0.05300.03590.04810.13610.13590.06390.05780.1947
IID [2]/reflectance0.10110.04930.05360.13890.07830.06590.07310.2635
IID [2]/ shading0.06460.05270.06430.15580.11910.05060.05920.2302
CycleGAN [46]/photo2ukiyoe0.09080.03940.05020.23510.14620.05050.05750.2571
CycleGAN [46]/photo2vangogh0.12300.03900.04940.25650.12240.05190.05510.1701
Colorization [43]0.06160.04020.04680.15550.08670.04950.05660.1441
Colorization [36]0.03290.03730.04750.14740.13170.00810.02200.1558
Average0.13690.04450.05240.2414*0.15230.05210.05560.2167*
", + "image_path": "f13d647bdf06efb97bd10e615f84d66fff0271b94bca4ea6f02e59eaead9c425.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 183, + 547, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 547, + 243 + ], + "type": "text", + "content": "Table 3: Quantitative comparison of LPIPS [45]. The comparison presented in this table is limited to the results produced by [3, 16] and the proposed method. Lower LPIPS suggests better perceptual similarity. \\*\\*\" Indicates partially averaged results." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 247, + 538, + 331 + ], + "blocks": [ + { + "bbox": [ + 310, + 247, + 538, + 331 + ], + "lines": [ + { + "bbox": [ + 310, + 247, + 538, + 331 + ], + "spans": [ + { + "bbox": [ + 310, + 247, + 538, + 331 + ], + "type": "image", + "image_path": "a42da043b741f38ce395a808bcfd27963c2b397d954d830bde866d267845bdb5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 338, + 547, + 398 + ], + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 398 + ], + "type": "text", + "content": "Figure 5: Motion Smoothness Results. This figure presents the quantitative results acquired through the motion smoothness metric as described above. The lower motion smoothness represents smoother transitions in the videos." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 403, + 547, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 547, + 462 + ], + "type": "text", + "content": "presents the temporal smoothness results generated through the motion smoothness metric as presented in [12]. This metric tries to quantify the smoothness of videos over various window sizes through differential optical flow as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 467, + 545, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 467, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 315, + 467, + 545, + 516 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D = \\left(F _ {I _ {1} \\rightarrow I _ {2}} - F _ {I _ {0} \\rightarrow I _ {1}}\\right) - \\left(F _ {O _ {1} \\rightarrow O _ {2}} - F _ {O _ {0} \\rightarrow O _ {1}}\\right), \\\\ \\mathrm {M} (s) = \\log \\left(\\sum_ {d \\in D} \\mathbf {1} _ {[ s, s + 1)} \\left(\\| \\boldsymbol {d} \\| _ {2}\\right)\\right) - \\log (| D |), \\tag {8} \\\\ \\end{array}", + "image_path": "8decd925b6eeb1c9b2767b3e2dc226f6503932fcbdbcdbd3e99f86d17080a6ae.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "F_{a\\rightarrow b}" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " denotes optical flow estimated from frame " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " to frame " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": ". The preceding " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " refers to the raw/unprocessed and restored frames respectively. In Eq. 8, " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " denotes each of the " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "[\\nabla u,\\nabla v]" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " vectors in the differential optical flow and " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " represents the various window sizes for evaluation and " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "|D|" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " denotes the dimensionality of the differential optical flow. " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_S(d)" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " is the masking factor which is equal to 1 if " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " belongs to the set " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": ". The resultant " + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathrm{M}(s)" + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": " gives a measure of temporal smoothness over 3 consecutive frames. Fig. 5 presents the average results on the portion of evaluation datasets processed through [3, 16, 19] and the proposed model. A lower score suggests better smoothness. The proposed model again achieves SOTA results in terms of motion smoothness even over bigger window sizes as well." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Please note that all the presented metrics are subjective" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13540" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "in nature and do not account for the overall quality and consistency of the results due to the lack of ground truth; therefore, user studies should be considered as the main criteria for judging the results of all the proposed approaches for this task. We discuss the findings of the conducted user studies below." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 158, + 128, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 158, + 128, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 128, + 170 + ], + "type": "text", + "content": "5.1.2 User Study" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 115, + 172, + 221, + 297 + ], + "blocks": [ + { + "bbox": [ + 115, + 172, + 221, + 297 + ], + "lines": [ + { + "bbox": [ + 115, + 172, + 221, + 297 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 221, + 297 + ], + "type": "image", + "image_path": "a3c168b6d3b2e579a30a4930cf5c16755842618c99ed5f577aacee49081251e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 304, + 288, + 353 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 353 + ], + "type": "text", + "content": "Figure 6: User Study. This donut chart highlights the user preferences collected through the user study. A majority of the participants preferred the results produced by the proposed method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 357, + 288, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 288, + 620 + ], + "type": "text", + "content": "We conducted two separate user studies; a comprehensive and a factorized user study, to properly evaluate the performance of the proposed model. The first user study consisted of 36 participants with 150 different scenes processed through each of the proposed models. Each participant was asked to judge 5 videos processed through each of the proposed methods for this task. The users were shown a randomly sampled scene from the video pool processed through [3, 16, 19] and this work, simultaneously. The participants were instructed to judge the videos on the basis of naturalness, quality, consistency, content, and style preservation. Figure 6 presents the findings of the first (comprehensive) user study. On average, " + }, + { + "bbox": [ + 46, + 357, + 288, + 620 + ], + "type": "inline_equation", + "content": "41\\%" + }, + { + "bbox": [ + 46, + 357, + 288, + 620 + ], + "type": "text", + "content": " of the users preferred the videos restored by the proposed method. Furthermore, the results generated by the method proposed in Lai et al. [16] also received a significant preference in the user study but it is worth noting that all the previously proposed approaches require the unprocessed counterparts of the frame-wise processed videos at test time to restore the temporal consistency, whereas the proposed model restores the temporal consistency by only taking into account the inconsistent video at the test time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 635, + 175, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 175, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 175, + 647 + ], + "type": "text", + "content": "5.1.3 Factorized User Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "content": "Apart from the comprehensive user study, we also conducted a secondary in-depth user study to evaluate the reasoning of participants and their opinion on their selections of videos processed with all the methods on challenging scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "text", + "content": "In this user study, each user was instructed to judge 15 randomly selected videos (5 videos from each of the applications mentioned above) processed by each of the proposed models for this task. Please note that in this user study users were also shown the per-frame processed video in an extended window to judge the videos on content and style preservation as well. Apart from recording their preference, the participants were also instructed to record the reasoning for their selection. " + }, + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "inline_equation", + "content": "\\sim 86\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "text", + "content": " of the participants remarked that they chose the videos based on lesser flickering and better quality preservation. The results of the second user study for each of the selected tasks are provided in Fig. 7. On average, " + }, + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "inline_equation", + "content": "\\sim 47\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 252 + ], + "type": "text", + "content": " of the users preferred the videos processed through the proposed model for these extremely challenging tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 262, + 417, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 417, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 417, + 275 + ], + "type": "text", + "content": "5.2. Qualitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 281, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 593 + ], + "type": "text", + "content": "Figures " + }, + { + "bbox": [ + 304, + 281, + 545, + 593 + ], + "type": "inline_equation", + "content": "8\\sim 10" + }, + { + "bbox": [ + 304, + 281, + 545, + 593 + ], + "type": "text", + "content": " present some of the comparative qualitative results produced by the methods proposed in [3, 16, 19] and the proposed model. The results produced by [3] fail to retain the perceptual quality of video sequences where occlusion and deocclusion occur. Whereas the frames processed by DVP [19] are significantly blurry and lose the translation effect. In contrast to the comparison with DVP [19] and Boneel et al. [3], the comparison with the results produced through the method proposed by Lai et al. [16] can be a bit tricky. In order to aid the readers, we provide an epipolar comparison between the frames restored through their method and ours in Fig. 10. Due to the implicit definition of motion dynamics in their method [16], their method fails to faithfully retain the perceptual quality in the restored frames and often flips the colors/appearance of objects in the restored video according to the raw video. This bias towards the raw videos is quite prominent in methods like DVP [19] and Lai et al. [16]. Contrasting to their formulation, the explicit design of motion dynamics in the proposed method mitigates the need for raw videos at test time which consequently eliminates this bias and enables the proposed method to faithfully restore videos. Further qualitative comparisons and higher quality results, along with the results on super resolution [17], inpainting [18], and multiple restoration iterations are presented in the supplemental." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 601, + 470, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 601, + 470, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 601, + 470, + 613 + ], + "type": "text", + "content": "6. Limitations and Future work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "The proposed model has difficulties in restoring videos with very low frame rates and in the case of style-transferred videos, some finer inconsistencies like finer brush strokes are lost, which makes the resulting video seem a little dull (which is to be expected from all the proposed approaches for this task). A good direction for future work in this field can be the direct denoising of conventional optical flow with deterministic warp operators. The developed consis" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13541" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 68, + 226, + 241 + ], + "blocks": [ + { + "bbox": [ + 97, + 68, + 226, + 241 + ], + "lines": [ + { + "bbox": [ + 97, + 68, + 226, + 241 + ], + "spans": [ + { + "bbox": [ + 97, + 68, + 226, + 241 + ], + "type": "image", + "image_path": "3909694f263ba7677bb389caab3c5d9704199e1305ea1958c1de3cae6c812a7d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 547, + 289 + ], + "type": "text", + "content": "Figure 7: Factorized user study. The user preference scores for the tasks of style transfer, CycleGAN, and intrinsic image decomposition (left to right respectively). The majority of the users preferred the videos processed through the proposed method in each category of tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 233, + 68, + 361, + 242 + ], + "blocks": [ + { + "bbox": [ + 233, + 68, + 361, + 242 + ], + "lines": [ + { + "bbox": [ + 233, + 68, + 361, + 242 + ], + "spans": [ + { + "bbox": [ + 233, + 68, + 361, + 242 + ], + "type": "image", + "image_path": "fd53144caa3953fbb0d72bac85fedc98a92d98d3c407f6e8bf75569569fa54b0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 367, + 68, + 497, + 242 + ], + "blocks": [ + { + "bbox": [ + 367, + 68, + 497, + 242 + ], + "lines": [ + { + "bbox": [ + 367, + 68, + 497, + 242 + ], + "spans": [ + { + "bbox": [ + 367, + 68, + 497, + 242 + ], + "type": "image", + "image_path": "246c88f8bfb11b50682b25f88bd19e08e7558a745914c72db11177b55eb5ce9a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 300, + 286, + 399 + ], + "blocks": [ + { + "bbox": [ + 50, + 300, + 286, + 399 + ], + "lines": [ + { + "bbox": [ + 50, + 300, + 286, + 399 + ], + "spans": [ + { + "bbox": [ + 50, + 300, + 286, + 399 + ], + "type": "image", + "image_path": "c1f3e3962454ad5c55d7908768400076a39caa285da3a28e584f25af40eedaa9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 402, + 288, + 450 + ], + "lines": [ + { + "bbox": [ + 46, + 402, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 288, + 450 + ], + "type": "text", + "content": "Figure 8: Comparison with Bonneel et al. [3] The method proposed by Bonneel et al. [3] has difficulty in handling occlusion and long term consistency and produce temporal artifacts." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 460, + 286, + 561 + ], + "blocks": [ + { + "bbox": [ + 50, + 460, + 286, + 561 + ], + "lines": [ + { + "bbox": [ + 50, + 460, + 286, + 561 + ], + "spans": [ + { + "bbox": [ + 50, + 460, + 286, + 561 + ], + "type": "image", + "image_path": "59bb2b89a8456beded0c4b186847ea19fa7d594e0acc4113f2438f49b02aa5b6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 571, + 288, + 620 + ], + "lines": [ + { + "bbox": [ + 46, + 571, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 288, + 620 + ], + "type": "text", + "content": "Figure 9: Comparison with DVP [19] DVP [19] produces blurry frames and compromises the translation effect. This phenomenon is prominent in videos where new content is generated e.g. style transfer [13], CycleGAN [46]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 654 + ], + "type": "text", + "content": "tent flow models can also be used for a number of video processing applications like denoising and dehazing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 659, + 119, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 119, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 119, + 672 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "In this work, we present a task-agnostic temporal consistency correction framework that restores natural video dynamics by inferring and utilizing consistent motion repre" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 300, + 533, + 371 + ], + "blocks": [ + { + "bbox": [ + 319, + 300, + 533, + 371 + ], + "lines": [ + { + "bbox": [ + 319, + 300, + 533, + 371 + ], + "spans": [ + { + "bbox": [ + 319, + 300, + 533, + 371 + ], + "type": "image", + "image_path": "aaf535c0e9477358f4a24cbca6f1b92554edd7f41226cd5b05b3ca2006c135ef.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 374, + 545, + 447 + ], + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 447 + ], + "type": "text", + "content": "Figure 10: Comparison with Lai et al. [16] This figure produces epipolar temporal comparison of videos restored by Lai et al. [16] and the proposed method. The method proposed in [16], due to its dependence on raw videos and implicit siphoning/mixing of motion dynamics from the raw video introduces bias towards the raw videos." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 453, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 586 + ], + "type": "text", + "content": "sentations from temporally inconsistent videos. The proposed strategy mitigates the need for siphoning video dynamics from the unprocessed videos at test time consequently mitigating the compromise in translation effects that is prominent in the currently proposed approaches for this task and expands the scope of its employability to a wider set of applications. Through extensive experimentation and user studies, we demonstrate that despite the limited access to data at test-time the proposed method compares favorably to the pre-existing methods available for this task and achieves SOTA performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 590, + 409, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 590, + 409, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 409, + 605 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 610, + 545, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 610, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 610, + 545, + 706 + ], + "type": "text", + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2022-0-00156, Fundamental research on continual meta-learning for quality enhancement of casual videos and their 3D meta-verse transformation) and Samsung Electronics Co., Ltd, and Samsung Research Funding Center of Samsung Electronics under Project Number SRFCIT1901-06." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13542" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] Muhammad Kashif Ali, Sangjoon Yu, and Tae Hyun Kim. Deep motion blind video stabilization. BMVC, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "text", + "content": "[2] Sean Bell, Kavita Bala, and Noah Snavely. Intrinsic images in the wild. ACM TOG, 33(4):1-12, 2014." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 136, + 286, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 286, + 168 + ], + "type": "text", + "content": "[3] Nicolas Bonneel, James Tompkin, Kalyan Sunkavalli, Deqing Sun, Sylvain Paris, and Hanspeter Pfister. Blind video temporal consistency. ACM TOG, 34(6):1-9, 2015." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "type": "text", + "content": "[4] Dongdong Chen, Jing Liao, Lu Yuan, Nenghai Yu, and Gang Hua. Coherent online video style transfer. In ICCV, pages 1105-1114, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 203, + 286, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 286, + 236 + ], + "type": "text", + "content": "[5] Mengyu Chu, You Xie, Laura Leal-Taixe, and Nils Thuerey. Temporally coherent gans for video super-resolution (tecogan). arXiv preprint arXiv:1811.09393, 1(2):3, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 236, + 286, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 286, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 286, + 269 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 270, + 286, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 286, + 324 + ], + "type": "text", + "content": "[7] Debidatta Dwibedi, Yusuf Aytar, Jonathan Tompson, Pierre Sermanet, and Andrew Zisserman. Counting out time: Class agnostic video repetition counting in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10387-10396, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 325, + 286, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 286, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 286, + 367 + ], + "type": "text", + "content": "[8] Wei Gao, Yijun Li, Yihang Yin, and Ming-Hsuan Yang. Fast video multi-style transfer. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3211-3219, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 369, + 286, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 286, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 286, + 401 + ], + "type": "text", + "content": "[9] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 403, + 286, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 403, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 286, + 435 + ], + "type": "text", + "content": "[10] Michael Gharbi, Jiawen Chen, Jonathan T Barron, Samuel W Hasinoff, and Frédo Durand. Deep bilateral learning for real-time image enhancement. ACM TOG, 36(4):1-12, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 436, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 286, + 479 + ], + "type": "text", + "content": "[11] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, pages 2462-2470, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 481, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 523 + ], + "type": "text", + "content": "[12] Meiguang Jin, Zhe Hu, and Paolo Favaro. Learning to extract flawless slow motion from blurry videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8112-8121, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 286, + 557 + ], + "type": "text", + "content": "[13] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 558, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 612 + ], + "type": "text", + "content": "[14] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep blind video decaptioning by temporal aggregation and recurrence. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4258–4267, 2019." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 613, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 656 + ], + "type": "text", + "content": "[15] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Recurrent temporal aggregation framework for deep video inpainting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42:1038–1052, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 658, + 286, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 690 + ], + "type": "text", + "content": "[16] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In ECCV, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[17] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken," + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In CVPR, pages 4681-4690, 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 140 + ], + "type": "text", + "content": "[18] Eunhye Lee, Jeongmu Kim, Jisu Kim, and Tae Hyun Kim. Restore from restored: Single-image inpainting. arXiv preprint arXiv:2102.08078, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 142, + 545, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 142, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 308, + 142, + 545, + 164 + ], + "type": "text", + "content": "[19] Chenyang Lei, Yazhou Xing, and Qifeng Chen. Blind video temporal consistency via deep video prior. NeurIPS, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 166, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 166, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 308, + 166, + 545, + 198 + ], + "type": "text", + "content": "[20] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. NeurIPS, 30, 2017." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 545, + 243 + ], + "type": "text", + "content": "[21] Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. arXiv preprint arXiv:2204.02663, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 246, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 246, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 308, + 246, + 545, + 289 + ], + "type": "text", + "content": "[22] Xunyu Lin, Victor Campos, Xavier Giro-i Nieto, Jordi Torres, and Cristian Canton Ferrer. Disentangling motion, foreground and background features in videos. arXiv preprint arXiv:1707.04092, 2017." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 292, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 292, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 292, + 545, + 346 + ], + "type": "text", + "content": "[23] Yihao Liu, Hengyuan Zhao, Kelvin CK Chan, Xintao Wang, Chen Change Loy, Yu Qiao, and Chao Dong. Temporally consistent video colorization with deep feature propagation and self-regularization learning. arXiv preprint arXiv:2110.04562, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 348, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 348, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 348, + 545, + 403 + ], + "type": "text", + "content": "[24] Cheng Ma, Yongming Rao, Yean Cheng, Ce Chen, Jiwen Lu, and Jie Zhou. Structure-preserving super resolution with gradient guidance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7769-7778, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 404, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 449 + ], + "type": "text", + "content": "[25] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, pages 724-732, 2016." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 450, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 450, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 308, + 450, + 545, + 483 + ], + "type": "text", + "content": "[26] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In CVPR, pages 4161-4170, 2017." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 484, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 484, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 308, + 484, + 545, + 518 + ], + "type": "text", + "content": "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 520, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 520, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 520, + 545, + 552 + ], + "type": "text", + "content": "[28] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos. In " + }, + { + "bbox": [ + 308, + 520, + 545, + 552 + ], + "type": "inline_equation", + "content": "GCPR" + }, + { + "bbox": [ + 308, + 520, + 545, + 552 + ], + "type": "text", + "content": ", pages 26-36. Springer, 2016." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 554, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 545, + 587 + ], + "type": "text", + "content": "[29] Mehdi SM Sajjadi, Raviteja Vemulapalli, and Matthew Brown. Frame-recurrent video super-resolution. In " + }, + { + "bbox": [ + 308, + 554, + 545, + 587 + ], + "type": "inline_equation", + "content": "CVPR" + }, + { + "bbox": [ + 308, + 554, + 545, + 587 + ], + "type": "text", + "content": ", pages 6626-6634, 2018." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 589, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 589, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 589, + 545, + 633 + ], + "type": "text", + "content": "[30] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. NeurIPS, 28, 2015." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 545, + 667 + ], + "type": "text", + "content": "[31] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In CVPR, pages 8934-8943, 2018." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[32] Ximeng Sun, Huijuan Xu, and Kate Saenko. Twostreamvan: Improving motion modeling in video generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2744-2753, 2020." + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13543" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 600 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[33] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419. Springer, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "text", + "content": "[34] Hugo Thimonier, Julien Despois, Robin Kips, and Matthieu Perrot. Learning long term style preserving blind video temporal consistency. In ICME, pages 1-6. IEEE, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 286, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 286, + 163 + ], + "type": "text", + "content": "[35] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In CVPR, pages 9446-9454, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 164, + 286, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 164, + 286, + 196 + ], + "spans": [ + { + "bbox": [ + 49, + 164, + 286, + 196 + ], + "type": "text", + "content": "[36] Michael Waechter, Nils Moehrle, and Michael Goesele. Let there be color! large-scale texturing of 3d reconstructions. In ECCV, pages 836-850. Springer, 2014." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 198, + 286, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 198, + 286, + 241 + ], + "spans": [ + { + "bbox": [ + 49, + 198, + 286, + 241 + ], + "type": "text", + "content": "[37] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127(8):1106-1125, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 243, + 286, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 243, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 286, + 275 + ], + "type": "text", + "content": "[38] Zhicheng Yan, Hao Zhang, Baoyuan Wang, Sylvain Paris, and Yizhou Yu. Automatic photo adjustment using deep neural networks. ACM TOG, 35(2):1-15, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 277, + 286, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 277, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 49, + 277, + 286, + 331 + ], + "type": "text", + "content": "[39] Yuchen Yang, Ye Xiang, Shuaicheng Liu, Lifang Wu, Boxuan Zhao, and Bing Zeng. Glm-net: Global and local motion estimation via task-oriented encoder-decoder structure. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4211-4219, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 332, + 286, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 332, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 49, + 332, + 286, + 365 + ], + "type": "text", + "content": "[40] Chun-Han Yao, Chia-Yang Chang, and Shao-Yi Chien. Occlusion-aware video temporal consistency. In ACMMM, pages 777-785, 2017." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 366, + 286, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 366, + 286, + 399 + ], + "spans": [ + { + "bbox": [ + 49, + 366, + 286, + 399 + ], + "type": "text", + "content": "[41] Bo Zhang, Mingming He, Jing Liao, Pedro V Sander, Lu Yuan, Amine Bermak, and Dong Chen. Deep exemplar-based video colorization. In CVPR, pages 8052-8061, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 400, + 286, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 400, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 49, + 400, + 286, + 443 + ], + "type": "text", + "content": "[42] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 445, + 286, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 286, + 476 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 286, + 476 + ], + "type": "text", + "content": "[43] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, pages 649-666. Springer, 2016." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 478, + 286, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 478, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 49, + 478, + 286, + 510 + ], + "type": "text", + "content": "[44] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, pages 649-666. Springer, 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 513, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 513, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 513, + 286, + 555 + ], + "type": "text", + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 557, + 286, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 557, + 286, + 600 + ], + "spans": [ + { + "bbox": [ + 49, + 557, + 286, + 600 + ], + "type": "text", + "content": "[46] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2223-2232, 2017." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13544" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/b6090cd5-35b5-49d1-9775-853efd91c055_content_list.json b/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/b6090cd5-35b5-49d1-9775-853efd91c055_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..86114486d9a4b7d1198c58103348121bfb5d4a40 --- /dev/null +++ b/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/b6090cd5-35b5-49d1-9775-853efd91c055_content_list.json @@ -0,0 +1,1720 @@ +[ + { + "type": "text", + "text": "Task-Oriented Multi-Modal Mutual Learning for Vision-Language Models", + "text_level": 1, + "bbox": [ + 107, + 128, + 861, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sifan Long $^{1,2,3\\dagger*}$ Zhen Zhao $^{3,4\\dagger*}$ Junkun Yuan $^{3,5\\dagger*}$ Zichang Tan $^{3}$ Jiangjiang Liu $^{3}$ Luping Zhou $^{4}$ Shengsheng Wang $^{1,2‡}$ Jingdong Wang $^{3‡}$", + "bbox": [ + 125, + 178, + 841, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ College of Computer Science and Technology, Jilin University, Jilin, China \n $^{2}$ Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Jilin, China", + "bbox": [ + 99, + 215, + 872, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Baidu VIS $^{4}$ University of Sydney $^{5}$ Zhejiang University longsf22@mails.jlu.edu.cn {zhen.zhao, luping.zhou}@sydney.edu.au yuanjk@zju.edu.cn wss@jlu.edu.cn {tanzichang, liujiangjiang, wangjingdong}@baidu.com", + "bbox": [ + 109, + 268, + 859, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 356, + 313, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompt learning has become one of the most efficient paradigms for adapting large pre-trained vision-language models to downstream tasks. Current state-of-the-art methods, like CoOp and ProDA, tend to adopt soft prompts to learn an appropriate prompt for each specific task. Recent CoCoOp further boosts the base-to-new generalization performance via an image-conditional prompt. However, it directly fuses identical image semantics to prompts of different labels and significantly weakens the discrimination among different classes as shown in our experiments. Motivated by this observation, we first propose a class-aware text prompt (CTP) to enrich generated prompts with label-related image information. Unlike CoCoOp, CTP can effectively involve image semantics and avoid introducing extra ambiguities into different prompts. On the other hand, instead of reserving the complete image representations, we propose text-guided feature tuning (TFT) to make the image branch attend to class-related representation. A contrastive loss is employed to align such augmented text and image representations on downstream tasks. In this way, the image-to-text CTP and text-to-image TFT can be mutually promoted to enhance the adaptation of VLMs for downstream tasks. Extensive experiments demonstrate that our method outperforms the existing methods by a significant margin. Especially, compared to CoCoOp, we achieve an average improvement of $4.03\\%$ on new classes and $3.19\\%$ on harmonic-mean over eleven classification benchmarks.", + "bbox": [ + 75, + 388, + 473, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 823, + 209, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, large vision-language models (VLM), such as CLIP [33] and ALIGN [15], which employ language as", + "bbox": [ + 75, + 848, + 468, + 880 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e60548bb39680cb70e3e638e536d9fedbf0ab699aae57b4f901a623b562a8cb5.jpg", + "image_caption": [ + "Figure 1: Comparisons between CoCoOp [50] and our method. The cosine distance between the positive and the negative prompts, which quantifies the class discrimination, and the average accuracy on benchmarks are reported." + ], + "image_footnote": [], + "bbox": [ + 500, + 354, + 906, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "supervision signal instead of discrete labels, have shown impressive generalization performance in a wide range of downstream vision tasks. Their multi-modal interaction nature delivers open-vocabulary support and achieves amazing zero-shot classification performance. Despite their impressive transferable abilities, as discussed in [25], it is essential to re-activate specific representation capabilities for optimal performance in certain downstream tasks. Considering their hundreds of millions or billions of parameters, attempting to fine-tune the entire model is impractical and even jeopardizes the well-established representation space [14]. To this end, many recent studies have centred on the efficient and effective adaptation of pre-trained and frozen large VLMs for the specific downstream tasks [28, 51, 50].", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † Interns at Baidu VIS. ‡ Corresponding authors.", + "bbox": [ + 94, + 886, + 465, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21959", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompt, a simple, compact, and viable strategy, has become the leading solution for deploying large pre-trained VLMs into certain downstream tasks. CLIP [33] utilizes hand-crafted prompts to achieve impressive zero- and few-shot classification performance. Nevertheless, manually-designed prompts require significant domain knowledge and can be highly time-consuming and sub-optimal for specific downstream tasks. To address this problem, later studies [28, 51] adopt soft prompts to learn an appropriate text prompt via optimizing a contrastive loss on different text labels. CoCoOp [50] further highlights the limitations of such static soft prompts and proposes learning image-dependent prompts conditioned on individual instances rather than fixed prompts. It achieves great performance gains on unseen classes by adding high-level image embedding to text prompts. However, compared to CoOp with static prompts, CoCoOp essentially fuses identical image semantics with different text labels, leading to inevitable learning ambiguity and resulting in an average performance drop of $2.22\\%$ on base classes on 11 datasets (see Table 1). For example, it may associate the dog image semantics with a prompt that references the [class] of a cat. When using the cosine distance to measure the differences between the positive and negative text prompts, as shown in Fig. 1, CoCoOp holds low distance values, suggesting that it brings significant learning ambiguities to text prompts. Therefore, we argue that text prompts should not only condition on distinct input images for better generalization abilities, but also adapt to different classes to eliminate the potential ambiguities.", + "bbox": [ + 76, + 90, + 472, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this goal, we propose Class-aware Text Prompts (CTP), which leverages label-related image information to generate finer prompts. Specifically, we first contact learnable context vectors and each class label to model the initial prompt sentences. Then we leverage these class prompt sentences to query their corresponding image regions and representations. Corresponding related image features are subsequently added to initial class prompt sentences to produce the final text prompts. In this way, generated image-dependent and class-aware prompts can better concentrate on the image information in a more precise manner. As shown in Fig. 1, our method enjoys better discrimination between positive and negative prompts and consistently outperforms CoCoOp on 11 classification datasets.", + "bbox": [ + 75, + 532, + 468, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, we identify a critical problem in these text prompt-based strategies: the image branch is ignored and not adjusted to specific downstream tasks. As shown in Fig. 2 (CoCoOp), on the task of identifying birds, the output image feature, without further tuning, can be distracted to leaves of the same color. Similarly, it also wrongly highlights the beer foam that is of a similar shape to recognize golf balls. Since the final recognition is jointly inferred by both text and image branches, such an issue may degrade the classification performance. Thus it is necessary", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f7f5394eb598b2671856cc56aa696bef15416101573a59ba6cb24821420bc5ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 89, + 625, + 188 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e38af2466fc343501b5baa10c409334eb80db5486b5af21f7b9fbb4c688f9545.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 90, + 743, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e9a7cf8de1641bb4206e1fbffe2fefc27df51a4b2f8ee6eee649980066ad3505.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 748, + 90, + 880, + 188 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3b30c3b4c2277835013d0ac2ba1f5a9e39a0ee9fe738bcc8bb821158d57fff41.jpg", + "image_caption": [ + "Figure 2: Comparisons of attention map visualization for CoCoOp and our method on ImageNet. Our method obtains better average accuracy of both base and new classes across 11 datasets by paying attention to task-related regions." + ], + "image_footnote": [], + "bbox": [ + 513, + 191, + 625, + 300 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0cda44d5167afcc175abc3c4f628891a8a03fc9fab3ad1f01e4d862cbbee4cf4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 193, + 743, + 300 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ebc7cca7075bb6fa3ef540e9e2e4d35a534f320a27c23f184c98be0d7fc0a7b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 748, + 193, + 880, + 300 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to tune the image features further so that the image branch can focus more on the tasks-related representation. We then propose Text-guided Feature Tuning (TFT), which leverages encoded text embedding to guide image representation more on task-related regions. As shown in Fig. 2 (ours), our method successfully focuses on task-related regions, i.e., birds and golf balls. We then leverage the contrastive loss function to further align class-aware text embedding and text-guided image features on certain downstream tasks.", + "bbox": [ + 496, + 398, + 890, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we propose a new task-oriented multimodal mutual-learning method, which well-integrates our designed class-aware text prompts and text-guided feature tuning for fast adaptation of frozen VLMs on downstream tasks. Image features can help construct image-dependent class-aware text prompts, leading to more discriminative text embedding. Simultaneously, improved text embedding can further guide the image branch attending to class-related representation. In this way, these two different modality branches can be tightly coupled and mutual-beneficial across the whole training process. Our main contributions are summarized in the following.", + "bbox": [ + 496, + 536, + 892, + 717 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose class-aware text prompts which generate prompts based on task-relevant image semantics instead of complete visual information. In this way, we improve the classification accuracy of unseen classes without introducing extra learning ambiguities.", + "- We propose text-guided feature tuning which enforces image branch to pay more attention to the task-related representation. As a result, the model avoids deviating attention to the task-irrelevant regions of the image.", + "- Benefiting from our mutual learning strategy, our" + ], + "bbox": [ + 517, + 728, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21960", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "method achieves SOTA results on four downstream tasks. Especially, ours significantly outperforms existing methods on the base-to-new generalization task.", + "bbox": [ + 109, + 90, + 467, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 162, + 215, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision language models (VLM). The current VLM can be roughly divided into four categories based on the training objectives: image-text matching [3, 20, 27], contrastive loss [19, 21, 22], masked language modeling [38, 39, 46], and masked image modeling [3, 27, 38]. As a milestone, CLIP utilizes 400 million image-text pairs to train a large-scale multi-modal model and demonstrates promising performance on a wide spectrum of tasks including few-shot and zero-shot visual recognition. Motivated by this work, numerous follow-ups have been proposed to improve the effectiveness (e.g., FLIP [24], A-CLIP [45], MaskCLIP [7], and SLIP [30]) or apply it to other domains (e.g., Dense-CLIP [34] and ActionCLIP [42]). The primary limitation of these methods is that hand-crafted prompts are dataset-sensitive and difficult to optimize. We design an automatic and learnable prompts method to enhance the generalization performance of pre-trained models on downstream tasks.", + "bbox": [ + 76, + 186, + 468, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt learning in NLP. As the scale and complexity of pre-trained language models continue to grow, fine-tuning for specific tasks is becoming increasingly expensive. In contrast, prompt-based approaches are an efficient and lightweight alternative that can be used to generate high-quality text with much lower computational requirements. The original prompts were manually designed prompt templates. While manually designing prompts is advantageous due to their intuitive and comprehensible nature, it also presents a significant challenge that demands extensive experimentation, experience, and language expertise, resulting in high costs. To overcome the limitations of manual prompt design, numerous studies have initiated research into automatically learn appropriate prompts. The automatic prompts can be categorized into two types: discrete prompts and continuous prompts. Discrete prompts consist of various approaches such as prompt mining [17], prompt paraphrasing [49, 10], gradient-based search [40], prompt generation [9] and prompt scoring [5]. On the other hand, continuous prompts include techniques such as prefix tuning [23], tuning initialized with discrete prompts [36] and hard-soft prompt hybrid tuning [26]. These methods have also been applied to the field of computer vision for prompt learning research. However, the task of prompt learning in computer vision is often considered more challenging than in natural language due to the relatively limited high-level semantic information present in visual data with raw pixels.", + "bbox": [ + 76, + 454, + 468, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt learning in vision language models. Prompt learning has been demonstrated to be an effective method", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for improving the performance of pre-trained language models on downstream tasks. Recently, prompt learning has gained increased attention in the context of vision language models. For example, CoOp [51] employs learnable vectors to model contextual words as prompts, and demonstrates that automatic prompts outperform hand-crafted prompts in downstream tasks. CoCoOp [50] extends CoOp by incorporating lightweight neural networks to dynamically generate prompts based on each image, thus mitigating sensitivity to class shifts. Different from the above methods, VP [1], VPT [43], and EVP [16] prompt with images. VP [1] directly combines learnable prompts and pixel-wise input images as new inputs to the model. EVP [43] shrinks the original image before padding the prompts around it, to avoid destroying the original image information. VPT [16] introduces a small amount learnable parameters into the input sequence of each transformer layer and learns them together with a linear head during fine-tuning. Building on the prompt learning approach of the text branch, we propose class-aware text prompt that generates image-dependent and class-aware prompts. Similarly, follow the feature tuning of image branch, we introduce text-guided tuning, which directs the image branch to focus on the task-relevant local regions rather than the global information.", + "bbox": [ + 496, + 90, + 890, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 470, + 589, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Comparisons of CLIP, CoOp, and CoCoOp", + "text_level": 1, + "bbox": [ + 500, + 498, + 867, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP comprises two encoders: an image encoder and a text encoder. The image encoder, denotes by $F(x)$ , converts an image $x \\in \\mathbb{R}^{3 \\times H \\times W}$ with height of $H$ and width of $W$ into a $d$ -dimensional image feature $f_x \\in \\mathbb{R}^{N \\times d}$ , where $N$ is the number of split patches. Meanwhile, the text encoder, denoted as $G(t)$ , generates an $d$ -dimensional text representation $g_t \\in \\mathbb{R}^{M \\times d}$ from natural language text $t$ , where $M$ is the number of classes. Two encoders are jointly trained using a contrastive loss function that maximizes the cosine similarity of matched pairs and minimizes that of the unmatched pairs. After training, CLIP can be directly used for zero-shot image recognition without requiring fine-tuning of the whole model. Since CLIP is pre-trained on whether an image matches a textual description, the hand-crafted prompt template is employed to convert raw labels into textual descriptions. The most common form of template in CLIP is \"a photo of a [CLASS]\", where the class token is replaced with specific class names such as \"cat\", \"dog\", \"car\", etc. We let the image features $f_x$ of an image $x$ be extracted by an image encoder and the text features $g_t$ be obtained by feeding the prompt description into the text encoder. The prediction task is defined as the classification of an image into one of $C$ categories, which are represented by the set $y \\in \\{1, \\dots, C\\}$ . Denote $y$ as the predicted category. Let $g_t^i$ be the $i$ -th dimension of text features $g_t$ , with image features", + "bbox": [ + 496, + 523, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21961", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7e2bafb580134e9fcc143ee5ebe227bc6853de91852f569c3261584a883539ff.jpg", + "image_caption": [ + "Figure 3: Comparisons of three representative prompt learning techniques and our method. The main differences lie in how the text and image branches focus on downstream tasks. CLIP artificially designs prompt templates. CoOp designs automatic prompts using learnable parameters. CoCoOp directly allows text branch to focus on images semantic through Meta-Net. We introduce Class-aware Text Prompts (CTP) and Text Feature Tuning (TFT) to the text and image branches, respectively. The CTP generates prompts based on class-related image information instead of using the identical image semantics like CoCoOp. The TFT enables the image branch to directly focus on downstream tasks. We leverage the contrastive loss function to align task-oriented text and images, making them promote each other for achieving better downstream generalization performance." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 428, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a57d25f8bf6d3de5b7ea94a9651d697290ac0c4115e83f379d47e10c4f72f977.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 433, + 88, + 890, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$f_{x}$ , we have the predicted probability of the $i$ -th class:", + "bbox": [ + 76, + 425, + 434, + 440 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f _ {x} , g _ {t} ^ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\cos \\left(f _ {x} , g _ {t} ^ {j}\\right) / \\tau\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 449, + 468, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot ,\\cdot)$ denotes the cosine similarity and $\\tau$ is the temperature parameter of the softmax function.", + "bbox": [ + 75, + 503, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CoOp replaces the hand-crafted prompts with automatically generated prompts. Specifically, CoOp introduces $k$ learnable context vectors $\\{v_{1},\\ldots ,v_{k}\\}$ to model the context words of the prompts. We define $c_{i}$ as the word embedding of the $i$ -th class name. Then, the prompt of $i$ -th class is denoted as $p_i = \\{v_1,\\dots,v_k,c_i\\}$ . Therefore, we have the predicted probability of the $i$ -th class using CoOp method:", + "bbox": [ + 75, + 539, + 468, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f _ {x} , G \\left(p _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\cos \\left(f _ {x} , G \\left(p _ {j}\\right)\\right) / \\tau\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 654, + 468, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $G(p_{i})$ is the text embedding from text encoder $G$ .", + "bbox": [ + 76, + 700, + 449, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CoCoOp extends CoOp by generating image-conditional prompts. Specifically, CoCoOp uses Meta-Net to generate the residual vector $\\pi$ based on each image. Each context token is now obtained by $v_{k}(x) = v_{k} + \\pi$ . The prompt of the $i$ -th class $c_{i}$ is defined as $p_i(x) = \\{v_1(x),\\dots ,v_k(x),c_i\\}$ . As a result, the prediction probability of the $i$ -th class is:", + "bbox": [ + 76, + 723, + 468, + 814 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f _ {x} , G \\left(p _ {i} (x)\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left( \\right.\\cos \\left(f _ {x} , G \\left(p _ {j} (x)\\right) / \\tau\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 821, + 468, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $G(p_i(x))$ is the text embedding conditional on the image $x$ from the text encoder $G$ .", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Our Task-Oriented Mutual Learning Method", + "text_level": 1, + "bbox": [ + 500, + 424, + 883, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method consists of two modules, i.e., Class-aware Text Prompts (CTP) and Text-guided Feature Tuning (TFT), as shown in Fig. 3. Compared to CoCoOp, we use CTP to generate class-aware prompts based on task-relevant local image regions instead of the global information. Besides, we use CTP to make the image branch directly pay attention to the task-related image region. We let the two modules be tightly coupled and mutual-beneficial across the training process by optimizing the contrastive loss function.", + "bbox": [ + 496, + 446, + 890, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CTP learns image conditioned discriminative prompts for finer paying attention to semantic-related regions of the images. Specifically, in order to obtain the text semantic-related regions of the image, we leverage the prompt $p$ and the image feature $f_{x}$ to calculate the attention matrix $A^{t}$ :", + "bbox": [ + 496, + 588, + 890, + 665 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA ^ {t} = p f _ {x} ^ {T}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 656, + 671, + 890, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A^t \\in \\mathbb{R}^{M \\times N}$ is the image-to-text attention map. $A_{i,j}^t$ represents the similarity between the $i$ -th class in the text prompts and the $j$ -th patch in the image. In this way, we can query the regions of the images that semantically related to the class information by the attention matrix $A^t$ . That is,", + "bbox": [ + 496, + 696, + 890, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {x} ^ {t} = \\operatorname {s o f t m a x} \\left(A ^ {t}\\right) f _ {x}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 781, + 890, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{x}^{t}$ is the regions correlated to the text of a specific class. We use it to obtain augmented class-aware prompts:", + "bbox": [ + 496, + 806, + 890, + 837 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np ^ {a} = p + f _ {x} ^ {t}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 844, + 890, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p^a$ is the text prompts enhanced by semanticallyrelevant image regions. Let $p_i^a$ be the $i$ -th dimension of $p^a$ ,", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21962", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c159916aeb2ec30ca1772bfd0a6cc852700f4e34e2702ae67d0efc8ed7715c87.jpg", + "table_caption": [ + "(a) Average over 11 datasets" + ], + "table_footnote": [], + "table_body": "
BaseNewHos
CLIP69.3474.2271.70
CoOp82.6963.2271.66
CoCoOp80.4771.6975.83
ProDA81.5672.3076.65
Ours83.0175.7279.02
", + "bbox": [ + 138, + 112, + 331, + 210 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/073489641f536df8bef392e777e2951c8ba965ec21c41b8e4c0283d11dcba289.jpg", + "table_caption": [ + "(b) ImageNet" + ], + "table_footnote": [], + "table_body": "
BaseNewHos
CLIP72.4368.1470.22
CoOp76.4767.8871.92
CoCoOp75.9870.4373.10
ProDA75.4070.2372.72
Ours77.4270.4473.77
", + "bbox": [ + 383, + 112, + 578, + 210 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/78509d0c261120b35eceeb7ef262ec08deb7c3008930f3ce05ae987cd626c553.jpg", + "table_caption": [ + "(c) Caltech101" + ], + "table_footnote": [], + "table_body": "
BaseNewHos
CLIP96.8494.0095.40
CoOp98.0089.8193.73
CoCoOp97.9693.8195.84
ProDA98.2793.2395.68
Ours98.3194.7596.50
", + "bbox": [ + 638, + 112, + 831, + 210 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e8ce3fd5a548b03855426d1e5ed6d09bae2a9c0a2c65d3aed1a748cc45c9c765.jpg", + "table_caption": [ + "(d) OxfordPets" + ], + "table_footnote": [], + "table_body": "
BaseNewHos
CLIP91.1797.2694.12
CoOp93.6795.2994.47
CoCoOp95.2097.6996.43
ProDA95.4397.8396.62
Ours95.8697.5596.70
", + "bbox": [ + 138, + 244, + 331, + 342 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f912b822cf107216189605348ece0ca763f8f65455b5003b6ba629dfb378de59.jpg", + "table_caption": [ + "(e) StanfordCars" + ], + "table_footnote": [], + "table_body": "
BaseNewHos
CLIP63.3774.8968.65
CoOp78.1260.4068.13
CoCoOp70.4973.5972.01
ProDA74.7071.2072.91
Ours76.2974.1775.22
", + "bbox": [ + 383, + 244, + 578, + 342 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/071f43b754916eafb0ce6339ad4eaf234458486d0b13721cae51879257d422cf.jpg", + "table_caption": [ + "(f) Flowers102" + ], + "table_footnote": [], + "table_body": "
BaseNewHos
CLIP72.0877.8074.83
CoOp97.6059.6774.06
CoCoOp94.8771.7581.71
ProDA97.7068.6880.66
Ours97.3677.7086.43
", + "bbox": [ + 638, + 244, + 831, + 342 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/43154d31c91896a9bad637cae47d366402f747f014c92e1a8aeb8b1eaa136f35.jpg", + "table_caption": [ + "(g) Food101" + ], + "table_footnote": [], + "table_body": "
BaseNewHos
CLIP90.1091.2290.66
CoOp88.3382.2685.19
CoCoOp90.7091.2990.99
ProDA90.3088.5789.43
Ours90.5492.3191.42
", + "bbox": [ + 138, + 376, + 331, + 474 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a9ab21687c8aed961cb6a9cc9cc5b40cb3902cc803b04f71bdab1518a0ed07b5.jpg", + "table_caption": [ + "(h) FGVCAircraft" + ], + "table_footnote": [], + "table_body": "
BaseNewHos
CLIP27.1936.2931.09
CoOp40.4422.3028.75
CoCoOp33.4123.7127.74
ProDA36.9034.1335.46
Ours39.4935.3737.32
", + "bbox": [ + 383, + 376, + 578, + 474 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/740126e6a45959c30423d6a0d11b601394d44018a1148beb525d360ac55f13a8.jpg", + "table_caption": [ + "(i) SUN397" + ], + "table_footnote": [], + "table_body": "
BaseNewHos
CLIP69.3675.3572.23
CoOp80.6065.8972.51
CoCoOp79.7476.8678.27
ProDA78.6776.9377.79
Ours82.1677.4979.76
", + "bbox": [ + 638, + 376, + 831, + 474 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1d5aef3a67e15e721b3d10805934a4be3beb918c71194b334bd9e190eacb864a.jpg", + "table_caption": [ + "(j) DTD" + ], + "table_footnote": [], + "table_body": "
BaseNewHos
CLIP53.2459.9056.37
CoOp79.4441.1854.24
CoCoOp77.0156.0064.85
ProDA80.6756.4866.44
Ours79.4761.5369.36
", + "bbox": [ + 133, + 508, + 326, + 607 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e5924c8948751be0af4c9408ef186c270431e7436e657651c34dbb4aa522cecb.jpg", + "table_caption": [ + "(k) EuroSAT" + ], + "table_footnote": [], + "table_body": "
BaseNewHos
CLIP56.4864.0560.03
CoOp92.1954.7468.69
CoCoOp87.4960.0471.21
ProDA83.9066.0073.88
Ours92.1473.8782.00
", + "bbox": [ + 385, + 508, + 578, + 608 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/42c3e9b17fb490a0a7055fcf3c97bb6bfd80e6e307b30a4e469f30e5a88c2d3d.jpg", + "table_caption": [ + "(1)UCF101" + ], + "table_footnote": [], + "table_body": "
BaseNewHos
CLIP70.5377.5073.85
CoOp84.6956.0567.46
CoCoOp82.3373.4577.64
ProDA85.2371.9778.04
Ours84.1277.7480.80
", + "bbox": [ + 638, + 508, + 833, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Results (%) of the base-to-new generalization task on 11 benchmark datasets. We report the accuracy with CLIP ViT-B/16 model on the base classes (Base), the unseen classes (New), and the harmonic mean of both of them (Hos).", + "bbox": [ + 75, + 618, + 893, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we then have the predicted probability of the $i$ -th class:", + "bbox": [ + 76, + 675, + 441, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f _ {x} , G \\left(p _ {i} ^ {a}\\right) / \\tau\\right)\\right.}{\\sum_ {j = 1} ^ {C} \\exp \\left( \\right.\\cos \\left(f _ {x} , G \\left(p _ {j} ^ {a}\\right) / \\tau\\right)} .} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 696, + 468, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We generate class-aware prompts instead of fusing identical image semantics with prompts of different classes, bringing category discrimination to the specific downstream tasks.", + "bbox": [ + 75, + 741, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TFT leverages text features to guide images to focus on task-related regions. Specifically, using the embeddings $g^{a}$ of the augmented prompts $p^{a}$ as input, we have attention:", + "bbox": [ + 75, + 792, + 468, + 838 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA ^ {x} = f _ {x} \\left(g ^ {a}\\right) ^ {T}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 844, + 468, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A^x \\in \\mathbb{R}^{N \\times M}$ denotes text-to-image attention map. $A_{i,j}^x$ represents the similarity between the $i$ -th patch in the", + "bbox": [ + 76, + 869, + 468, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "image and the $j$ -th class in the text representation. Similar to image-to-text, we use it to query the class-related part of the text correlated to the image, augmenting image features:", + "bbox": [ + 498, + 675, + 893, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {a} = \\operatorname {s o f t m a x} \\left(A ^ {x}\\right) g ^ {a} + f _ {x}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 727, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f^a$ is the augmented image embeddings. We thus let image branch focus on the tasks-related representation.", + "bbox": [ + 498, + 750, + 890, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Augmented contrastive loss function is then employed to further align class-aware text embedding and text-guided image features on specific downstream tasks. The predicted probability of the $i$ -th class, which is used to calculate the contrastive loss, after mutual augmentation is:", + "bbox": [ + 498, + 786, + 893, + 861 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f ^ {a} , g _ {i} ^ {a}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\cos \\left(f ^ {a} , g _ {i} ^ {a} / \\tau\\right)\\right)}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 866, + 890, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21963", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b81f29fe9de61578232259f33377c8c86cf683e8e61ebec5a1184b4c8911a78c.jpg", + "image_caption": [ + "Figure 4: Absolute improvement over CoCoOp in the base-to-new generalization task. Compared to CoCoOp, Our method achieves improvement on both base (left sub-figure) and new (right sub-figure) classes on most of the datasets." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 424, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a0bb075f465cafcb3179aa574b98a795062f9d0c9165364270cf7f56c8b495bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 87, + 851, + 284 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ab85fd60dd181cbaef192e783ea0ab702aba3a0ce320f3e015187b7af8beff47.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ImageNetCaltech101OxfordPedsStanfordCarsFlowers102Food101FGVCAircraftSUN397DTDEuroSATUCF101Average
CLIP68.6389.3688.9965.6770.4989.2327.1265.2946.0254.1769.8366.80
CoOp71.5195.5393.3174.2595.7087.2334.1874.8268.4677.8277.2977.28
CoCoOp71.0293.4393.9371.2187.3487.3932.0372.3263.8472.7877.4074.79
Ours72.9095.9093.9679.1096.7389.9538.7279.3772.4981.0083.4580.32
", + "bbox": [ + 81, + 339, + 890, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Results (%) of 16-shot learning task on 11 datasets.", + "bbox": [ + 279, + 468, + 686, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Task-targeted semantic information is transferred between the two branches by minimizing the augmented contrastive loss. We merge probability before and after augmentation:", + "bbox": [ + 75, + 511, + 468, + 556 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP (y = i \\mid x) = \\frac {\\exp \\left(\\left(\\cos \\left(f , g _ {i}\\right) + \\lambda \\left(\\cos \\left(f ^ {a} , g _ {i} ^ {a}\\right)\\right) / \\tau\\right)\\right.}{\\sum_ {j = 1} ^ {C} \\exp \\left( \\right.\\left(\\cos \\left(f , g _ {i}\\right) + \\lambda \\left(\\cos \\left(f ^ {a} , g _ {i} ^ {a}\\right)\\right) / \\tau\\right)} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 564, + 488, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is the balance hyper-parameter, which is analyzed in our experiments. We let the two different modalities tightly coupled and mutual beneficial across the whole training process by performing the contrastive optimization.", + "bbox": [ + 75, + 614, + 468, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 686, + 209, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of our method on four generalization tasks, including 1) generalization from base classes to new classes; 2) few-shot classification; 3) cross-dataset transfer; 4) domain generalization. After that, we provide extensive ablation studies and in-depth analyses.", + "bbox": [ + 75, + 713, + 468, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Following [33, 51], we use 11 image recognition datasets for the tasks of base-to-new generalization, few-shot classification and cross-dataset transfer. It contains generic image classification datasets (ImageNet [6] and Caltech101 [8]), fine-grained classification datasets (Oxford Pets [32], StanfordCars [18], Flowers102 [31], Food101 [2] and FGVCAircraft [29]), scene recognition ( SUN397", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[44]), action recognition (UCF101 [37]), texture classification (DTD [4]), and satellite imagery recognition (EuroSAT [11]). For the domain generalization task, we use ImageNet as the source dataset and select ImageNetV2 [35], ImageNet-Sketch [41], ImageNet-A [13], and ImageNet-R [12], which are the ImageNet variants, as the target.", + "bbox": [ + 496, + 511, + 892, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. By following [50, 51], we use the best visual backbone available in CLIP, i.e., ViT-B/16, throughout the experiments. We train 10 epochs using SGD optimizer with base learning rate of 0.002 and cosine decay schedule. We set the hyper-parameter $\\lambda$ in Eq. (11) to 0.2 for all experiments, and provide sensitivity analyses in Fig. 5. We run all the experiments three times with different random seeds and report the average classification accuracy.", + "bbox": [ + 496, + 612, + 893, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with 4 baselines. (1) Zero-shot CLIP [33] with hand-crafted prompts. (2) CoOp [51], using automatically generated prompts from few data. (3) CoCoOp [50], dynamically generating prompts conditioned on the images. (4) ProDA [28], which learns prompts from few data samples and mitigates the domain gap.", + "bbox": [ + 496, + 743, + 893, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Generalization From Base to New Classes", + "text_level": 1, + "bbox": [ + 498, + 845, + 856, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the previous works, we split the classes equally into two groups for each dataset: one as base and", + "bbox": [ + 498, + 869, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21964", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/915d29625ac35d523db30c1819ace31fa700c5b6e2a0308dd88274171fd707a2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SourceTarget
ImageNetCaletech101OxfordPedsStanfordCarsFlowers102Food101FGVCAircraftSUN397DTDEuroSATUCF101
CoOp71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.55
CoCoOp71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.21
Ours72.9095.7390.2265.1469.8986.3823.3266.4946.4747.2467.43
", + "bbox": [ + 81, + 88, + 890, + 207 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2f90f0dc5531ead41b65f708345b085826f102f172e2838f78746fdef4286f9e.jpg", + "table_caption": [ + "Table 3: Results of cross-dataset transfer task. Each method is trained on the source dataset and evaluated on the target." + ], + "table_footnote": [], + "table_body": "
SourceTarget
ImageNetImageNetV2ImageNet-SketchImageNet-AImageNet-R
CLIP66.7360.8346.1547.7773.96
CoOp71.5164.2047.9949.7175.21
CoCoOp71.0264.0748.7550.6376.18
Ours72.9064.5749.1150.9476.68
", + "bbox": [ + 81, + 244, + 888, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4: Results of domain generalization task. Each method is trained on ImageNet and evaluated on ImageNet variants.", + "bbox": [ + 78, + 364, + 887, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the other as new. The learnable modules are trained exclusively on the base classes, while evaluation is carried out separately on both the base and new classes to testify generalization ability. We report the results on 11 benchmarks in Table 1. Although compared to CoOp, CoCoOp significantly narrows generalization gap in unseen classed, but it decreases the accuracy in seen classes from $82.69\\%$ to $80.47\\%$ . We attribute it to the homogeneous prompts of CoCoOp, which weakens the discriminative semantics of different categories. In comparison, our method improves the accuracy in unseen classes from $80.47\\%$ to $83.01\\%$ by prompting each text label with corresponding image information. Benefit from the mutual learning of our CTP and TFT modules, our method further improves the accuracy in unseen classes from $71.69\\%$ to $75.72\\%$ , even surpasses the accuracy of CLIP hand-crafted prompts. We provide a detailed comparisons of CoCoOp and our method of per-dataset improvement in Fig. 4. Our method gains significant improvements over CoCoOp in both seen and unseen classes on 10 out of 11 recognition datasets. Surprisingly, our method significantly improves CoCoOp by more than $10\\%$ in unseen classes on SUN397 and Flowers102 datasets.", + "bbox": [ + 75, + 407, + 472, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Few-Shot Classification", + "text_level": 1, + "bbox": [ + 76, + 753, + 292, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report few-shot classification results in Table 2. Our method surpasses baseline methods on all datasets in the few-shot setting. Especially, our method outperforms CoCoOp by $9.39\\%$ , $8.65\\%$ , and $8.22\\%$ on Flowers102, DTD, and EuroSAT, respectively, and the average improvement over 11 datasets is $5.53\\%$ . Our method also achieve $2\\%$ on the challenging dataset of ImageNet. The above experiments show the great discriminative ability of our method.", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Cross-Dataset Transfer", + "text_level": 1, + "bbox": [ + 500, + 407, + 714, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then evaluate the generalization ability of our method on more challenging cross-dataset tasks. In this setting, we learn multi-modal prompts on ImageNet of 1000 classes. The effectiveness of the learned prompts is then tested on 10 datasets containing generic and fine-grained image classification, scene recognition, and texture classification. The results are reported in Table 3. Our method achieves the best average accuracy on the 11 datasets, especially ImageNet. It demonstrates the great transfer ability of our method.", + "bbox": [ + 496, + 430, + 890, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Domain Generalization", + "text_level": 1, + "bbox": [ + 500, + 574, + 715, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The domain generalization setting evaluates the generalization ability of the model on the target domain that is similar to but different from the source domain [48, 47]. Zero-shot CLIP introduces no additional training parameters and exhibits great robustness to naturally distribution shifts. Other methods use few samples to train learnable parameters, there is a risk of overfitting the source distribution. Therefore, we conduct experiments using ImageNet as the source domain and evaluate the ability of generalizing to unknown on four ImageNet variants. The results are shown in Table 4. Our method achieves significant performance on the 4 ImageNet variant datasets. It verifies that our method improves the classification ability of the source domain dataset while maintaining the generalization on the target domain.", + "bbox": [ + 496, + 598, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Analysis", + "text_level": 1, + "bbox": [ + 500, + 832, + 671, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of each module. To evaluate the effectiveness of Class-aware Text Prompts (CTP) and Text-guided Feature Tuning (TFT) of our method, we conduct ablation", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21965", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2adbc8f916b8cb58d045b8da4d43fbacec886fcad666d8d9e739f65f78d68af6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAverageImageNetCaltech101OxfordPeds
BaseNewHosBaseNewHosBaseNewHosBaseNewHos
A82.6963.2271.6676.4767.8871.9298.0089.8193.7393.6795.2994.47
B82.3872.4476.6476.9669.6273.1198.2394.2496.1995.6197.9796.78
C82.9372.9877.2477.2169.8673.3598.4492.7295.4995.4997.8196.64
Ours83.0175.7279.0277.4270.4473.7798.3194.7596.5095.8697.5596.70
MethodStanfordCarsFlowers102Food101FGVCAircraft
BaseNewHosBaseNewHosBaseNewHosBaseNewHos
A78.1260.4068.1397.6059.6774.0688.3382.2685.1940.4422.3028.75
B74.6273.6874.1596.7266.4278.7690.3091.4790.8836.4134.3935.37
C75.8474.5375.1897.3274.8684.6390.5691.6591.1037.8233.1735.34
Ours76.2974.1775.2297.3677.7086.4390.5492.3191.4239.4935.3737.32
MethodSUN397DTDEuroSATUCF101
BaseNewHosBaseNewHosBaseNewHosBaseNewHos
A80.6065.8972.5179.4441.1854.2492.1954.7468.6984.6956.0567.46
B81.7376.8979.2480.1851.7962.9391.7067.6277.8483.7272.7277.83
C82.2976.2479.1581.7154.7465.5690.1060.5272.4185.4076.6880.81
Ours82.1677.4979.7679.4761.5369.3692.1473.8782.0084.1277.7480.80
", + "bbox": [ + 80, + 88, + 888, + 363 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5ddb6e23144a422d60d603476eb4be8a7bd4903cf84458412d091b97f5f84dce.jpg", + "table_caption": [ + "Table 5: Ablation studies of our method on 11 datasets. Three ablation cases are considered: A: Ours w/o TVP w/o FTP. B: Ours w/o TVP. C: Ours w/o FTP. TVP is the text-reorganized vision prompt, and FTP is the fine-grained text prompt." + ], + "table_footnote": [], + "table_body": "
Prompt LearningFeature TuningAccuracy (%)
MLP-PLCTPMLP-FTTFT
✓71.66 (CoOp)
75.83 (+4.17)
✓76.64 (+4.98)
✓75.94 (+4.28)
✓77.24 (+5.58)
✓✓77.05 (+5.39)
✓✓79.02 (+7.36)
", + "bbox": [ + 81, + 430, + 465, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6: Comparison of different structures for prompt learning and feature tuning. The average results of harmonic mean of from-base-to-new generalization task on 11 datasets are reported. In compared to our attention design in CTP and TFT modules, MLP-PL and MLP-FT are designed using the Linear-ReLU-Linear block setting of [50]. Improvements over the baseline of CoOp, are marked in green.", + "bbox": [ + 75, + 561, + 468, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "experiments on 11 datasets, as reported in Table 5. In most cases, each module significantly improves the performance of the model. For average results, CTP and TFT improves the results by $5.58\\%$ and $4.98\\%$ , respectively, and the combination of them improves the results by $7.36\\%$ . It shows the effectiveness of the two branches of text-to-image and image-to-text, the mutual learning of the two modules further improves the performance on downstream tasks.", + "bbox": [ + 75, + 696, + 468, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison of different structure design of multi-modal mutual learning. To further provide in-depth analysis about our mutual learning, we further explore two vanilla structures: (1) MLP-PL: The image features are forwarded to a block of Linear-ReLU-Linear, borrowed from [50], and", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2ff833bdc7e9960031070fad5962b722dfa853110b5d03a62bf13b6b40ba2d08.jpg", + "image_caption": [ + "Figure 5: Sensitivity analysis of $\\lambda$ , with base, new, and host metrics, on UCF101 (left) and Caltech101 (right) datasets." + ], + "image_footnote": [], + "bbox": [ + 500, + 430, + 687, + 566 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b464df958c9d49e0af0522b8b1ee349247eb7f7e1f893f457bc10de6e5195241.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 430, + 888, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "then added to the text for augmenting it (the same to us). (2) MLP-FT: The text prompts are forwarded to the Linear-ReLU-Linear block, and then added to the image for augmenting it (the same to us). In comparison, our class-aware text prompts (CTP) module and text-guided feature tuning (TFT) module, adopt text-image attention to learn the augmented features instead of the Linear-ReLU-Linear block. We report the results of the different designs in Table 6. First, we find that combining MLP-PL & MLP-FT and CTP & TFT can both improve the results compared with using either of them. It indicates that both prompt learning and feature tuning are important to achieve better results. Second, compared with the design of Linear-ReLU-Linear block, our design of text-image attention further improves performance by $0.81\\%$ and $1.3\\%$ for prompt learning and feature tuning, respectively. It demonstrates the effectiveness of our design of attention, which helps the model to focus on class-aware and task-related semantics. Third, com", + "bbox": [ + 496, + 628, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21966", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pared with CoOp, both of the designs could improve the final results by large margins. The key factor of our mutual learning to achieve significant performance is the task-related alignment of vision and language in latent space.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sensitivity Analysis of $\\lambda$ . We evaluate the parameter sensitivity of $\\lambda$ of Eq. (11) in Fig. 5. The results suggest that the performance of our method is generally robust to $\\lambda$ , indicating a wide range of $\\lambda$ works well in downstream tasks.", + "bbox": [ + 75, + 157, + 468, + 218 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 229, + 194, + 244 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we introduce task-oriented multi-modal mutual learning for adapting large vision-language models to downstream vision tasks. We propose class-aware text prompt and text-guided feature tuning to unleash the potential of the vision-language model by re-activating its task-related representation abilities. Our method yields impressive generalization performance on a wide range of vision tasks and datasets. We hope the presented findings and insights in this paper could benefit the following works in designing more efficient and effective adaptation methods. For the future work, we think it is interesting to extend the adaptation of vision language models to more vision tasks, such as semantic segmentation, object detection, etc.", + "bbox": [ + 75, + 255, + 468, + 450 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations. Similar to CoCoOp, we learn image-conditioned representations, thus the batch-size of training need to be set to 1, which is not efficient enough for learning. We aim to solve this efficiency issue in the future work.", + "bbox": [ + 75, + 452, + 468, + 512 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledgements.", + "text_level": 1, + "bbox": [ + 76, + 523, + 267, + 540 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the Innovation Capacity Construction Project of Jilin Province Development and Reform Commission (2021FGWCXNLJSSZ10) and the Fundamental Research Funds for the Central Universities, JLU. Zichang Tan is the project lead.", + "bbox": [ + 75, + 550, + 468, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 638, + 171, + 652 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting largescale models. arXiv preprint arXiv:2203.17274, 1(3):4, 2022.3", + "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 446-461. Springer, 2014. 6", + "[3] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX, pages 104-120. Springer, 2020. 3", + "[4] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the" + ], + "bbox": [ + 86, + 662, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 6", + "[5] Joe Davison, Joshua Feldman, and Alexander M Rush. Commonsense knowledge mining from pretrained models. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pages 1173-1178, 2019. 3", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6", + "[7] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3", + "[8] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 6", + "[9] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pretrained language models better few-shot learners. arXiv preprint arXiv:2012.15723, 2020. 3", + "[10] Adi Haviv, Jonathan Berant, and Amir Globerson. Bertese: Learning to speak to bert. arXiv preprint arXiv:2103.05327, 2021. 3", + "[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6", + "[12] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 6", + "[13] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 6", + "[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 1", + "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1", + "[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Computer Vision-ECCV 2022: 17th" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21967", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 709-727. Springer, 2022. 3", + "[17] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438, 2020. 3", + "[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6", + "[19] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3", + "[20] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3", + "[21] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. Unimo: Towards unified-modal understanding and generation via cross-modal contrastive learning. arXiv preprint arXiv:2012.15409, 2020.3", + "[22] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 121-137. Springer, 2020. 3", + "[23] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3", + "[24] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. arXiv preprint arXiv:2212.00794, 2022. 3", + "[25] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35, 2023. 1", + "[26] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands, too. arXiv preprint arXiv:2103.10385, 2021. 3", + "[27] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3", + "[28] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 1, 2, 6", + "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI, pages 529-544. Springer, 2022. 3", + "[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 6", + "[32] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 6", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 6", + "[34] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 3", + "[35] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International conference on machine learning, pages 5389-5400. PMLR, 2019. 6", + "[36] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3", + "[37] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild. Center for Research in Computer Vision, 2(11), 2012. 6", + "[38] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019. 3", + "[39] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3", + "[40] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing nlp. arXiv preprint arXiv:1908.07125, 2019. 3", + "[41] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems, 32, 2019. 6", + "[42] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 3", + "[43] Junyang Wu, Xianhang Li, Chen Wei, Huiyu Wang, Alan Yuille, Yuyin Zhou, and Cihang Xie. Unleashing the power" + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21968", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of visual prompting at the pixel level. arXiv preprint arXiv:2212.10556, 2022.3", + "[44] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 6", + "[45] Yifan Yang, Weiquan Huang, Yixuan Wei, Houwen Peng, Xinyang Jiang, Huiqiang Jiang, Fangyun Wei, Yin Wang, Han Hu, Lili Qiu, et al. Attentive mask clip. arXiv preprint arXiv:2212.08653, 2022. 3", + "[46] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, 2021. 3", + "[47] Junkun Yuan, Xu Ma, Defang Chen, Kun Kuang, Fei Wu, and Lanfen Lin. Label-efficient domain generalization via collaborative exploration and generalization. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2361-2370, 2022. 7", + "[48] Junkun Yuan, Xu Ma, Defang Chen, Kun Kuang, Fei Wu, and Lanfen Lin. Domain-specific bias filtering for single labeled domain generalization. International Journal of Computer Vision, 131(2):552-571, 2023. 7", + "[49] Weizhe Yuan, Graham Neubig, and Pengfei Liu. Bartscore: Evaluating generated text as text generation. Advances in Neural Information Processing Systems, 34:27263-27277, 2021. 3", + "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 2, 3, 6, 8", + "[51] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 2, 3, 6" + ], + "bbox": [ + 78, + 90, + 468, + 627 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21969", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/b6090cd5-35b5-49d1-9775-853efd91c055_model.json b/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/b6090cd5-35b5-49d1-9775-853efd91c055_model.json new file mode 100644 index 0000000000000000000000000000000000000000..88e6b912364dc2e9a7eaa5a7e04103b33a912020 --- /dev/null +++ b/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/b6090cd5-35b5-49d1-9775-853efd91c055_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.13, + 0.862, + 0.154 + ], + "angle": 0, + "content": "Task-Oriented Multi-Modal Mutual Learning for Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.179, + 0.842, + 0.217 + ], + "angle": 0, + "content": "Sifan Long \\(^{1,2,3\\dagger*}\\) Zhen Zhao\\(^{3,4\\dagger*}\\) Junkun Yuan\\(^{3,5\\dagger*}\\) Zichang Tan\\(^{3}\\) Jiangjiang Liu\\(^{3}\\) Luping Zhou\\(^{4}\\) Shengsheng Wang\\(^{1,2‡}\\) Jingdong Wang\\(^{3‡}\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.217, + 0.874, + 0.269 + ], + "angle": 0, + "content": "\\(^{1}\\)College of Computer Science and Technology, Jilin University, Jilin, China \n\\(^{2}\\)Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Jilin, China" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.27, + 0.861, + 0.325 + ], + "angle": 0, + "content": "\\(^{3}\\)Baidu VIS \\(^{4}\\)University of Sydney \\(^{5}\\)Zhejiang University longsf22@mails.jlu.edu.cn {zhen.zhao, luping.zhou}@sydney.edu.au yuanjk@zju.edu.cn wss@jlu.edu.cn {tanzichang, liujiangjiang, wangjingdong}@baidu.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.357, + 0.314, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.474, + 0.798 + ], + "angle": 0, + "content": "Prompt learning has become one of the most efficient paradigms for adapting large pre-trained vision-language models to downstream tasks. Current state-of-the-art methods, like CoOp and ProDA, tend to adopt soft prompts to learn an appropriate prompt for each specific task. Recent CoCoOp further boosts the base-to-new generalization performance via an image-conditional prompt. However, it directly fuses identical image semantics to prompts of different labels and significantly weakens the discrimination among different classes as shown in our experiments. Motivated by this observation, we first propose a class-aware text prompt (CTP) to enrich generated prompts with label-related image information. Unlike CoCoOp, CTP can effectively involve image semantics and avoid introducing extra ambiguities into different prompts. On the other hand, instead of reserving the complete image representations, we propose text-guided feature tuning (TFT) to make the image branch attend to class-related representation. A contrastive loss is employed to align such augmented text and image representations on downstream tasks. In this way, the image-to-text CTP and text-to-image TFT can be mutually promoted to enhance the adaptation of VLMs for downstream tasks. Extensive experiments demonstrate that our method outperforms the existing methods by a significant margin. Especially, compared to CoCoOp, we achieve an average improvement of \\(4.03\\%\\) on new classes and \\(3.19\\%\\) on harmonic-mean over eleven classification benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.824, + 0.21, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.849, + 0.47, + 0.881 + ], + "angle": 0, + "content": "Recently, large vision-language models (VLM), such as CLIP [33] and ALIGN [15], which employ language as" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.355, + 0.907, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.597, + 0.895, + 0.659 + ], + "angle": 0, + "content": "Figure 1: Comparisons between CoCoOp [50] and our method. The cosine distance between the positive and the negative prompts, which quantifies the class discrimination, and the average accuracy on benchmarks are reported." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "supervision signal instead of discrete labels, have shown impressive generalization performance in a wide range of downstream vision tasks. Their multi-modal interaction nature delivers open-vocabulary support and achieves amazing zero-shot classification performance. Despite their impressive transferable abilities, as discussed in [25], it is essential to re-activate specific representation capabilities for optimal performance in certain downstream tasks. Considering their hundreds of millions or billions of parameters, attempting to fine-tune the entire model is impractical and even jeopardizes the well-established representation space [14]. To this end, many recent studies have centred on the efficient and effective adaptation of pre-trained and frozen large VLMs for the specific downstream tasks [28, 51, 50]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.466, + 0.902 + ], + "angle": 0, + "content": "*Equal contribution. † Interns at Baidu VIS. ‡ Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21959" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.531 + ], + "angle": 0, + "content": "Prompt, a simple, compact, and viable strategy, has become the leading solution for deploying large pre-trained VLMs into certain downstream tasks. CLIP [33] utilizes hand-crafted prompts to achieve impressive zero- and few-shot classification performance. Nevertheless, manually-designed prompts require significant domain knowledge and can be highly time-consuming and sub-optimal for specific downstream tasks. To address this problem, later studies [28, 51] adopt soft prompts to learn an appropriate text prompt via optimizing a contrastive loss on different text labels. CoCoOp [50] further highlights the limitations of such static soft prompts and proposes learning image-dependent prompts conditioned on individual instances rather than fixed prompts. It achieves great performance gains on unseen classes by adding high-level image embedding to text prompts. However, compared to CoOp with static prompts, CoCoOp essentially fuses identical image semantics with different text labels, leading to inevitable learning ambiguity and resulting in an average performance drop of \\(2.22\\%\\) on base classes on 11 datasets (see Table 1). For example, it may associate the dog image semantics with a prompt that references the [class] of a cat. When using the cosine distance to measure the differences between the positive and negative text prompts, as shown in Fig. 1, CoCoOp holds low distance values, suggesting that it brings significant learning ambiguities to text prompts. Therefore, we argue that text prompts should not only condition on distinct input images for better generalization abilities, but also adapt to different classes to eliminate the potential ambiguities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.746 + ], + "angle": 0, + "content": "To achieve this goal, we propose Class-aware Text Prompts (CTP), which leverages label-related image information to generate finer prompts. Specifically, we first contact learnable context vectors and each class label to model the initial prompt sentences. Then we leverage these class prompt sentences to query their corresponding image regions and representations. Corresponding related image features are subsequently added to initial class prompt sentences to produce the final text prompts. In this way, generated image-dependent and class-aware prompts can better concentrate on the image information in a more precise manner. As shown in Fig. 1, our method enjoys better discrimination between positive and negative prompts and consistently outperforms CoCoOp on 11 classification datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.904 + ], + "angle": 0, + "content": "On the other hand, we identify a critical problem in these text prompt-based strategies: the image branch is ignored and not adjusted to specific downstream tasks. As shown in Fig. 2 (CoCoOp), on the task of identifying birds, the output image feature, without further tuning, can be distracted to leaves of the same color. Similarly, it also wrongly highlights the beer foam that is of a similar shape to recognize golf balls. Since the final recognition is jointly inferred by both text and image branches, such an issue may degrade the classification performance. Thus it is necessary" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.09, + 0.627, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.091, + 0.744, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.091, + 0.882, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.193, + 0.627, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.194, + 0.744, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.194, + 0.882, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.312, + 0.895, + 0.374 + ], + "angle": 0, + "content": "Figure 2: Comparisons of attention map visualization for CoCoOp and our method on ImageNet. Our method obtains better average accuracy of both base and new classes across 11 datasets by paying attention to task-related regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.892, + 0.536 + ], + "angle": 0, + "content": "to tune the image features further so that the image branch can focus more on the tasks-related representation. We then propose Text-guided Feature Tuning (TFT), which leverages encoded text embedding to guide image representation more on task-related regions. As shown in Fig. 2 (ours), our method successfully focuses on task-related regions, i.e., birds and golf balls. We then leverage the contrastive loss function to further align class-aware text embedding and text-guided image features on certain downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.893, + 0.718 + ], + "angle": 0, + "content": "In summary, we propose a new task-oriented multimodal mutual-learning method, which well-integrates our designed class-aware text prompts and text-guided feature tuning for fast adaptation of frozen VLMs on downstream tasks. Image features can help construct image-dependent class-aware text prompts, leading to more discriminative text embedding. Simultaneously, improved text embedding can further guide the image branch attending to class-related representation. In this way, these two different modality branches can be tightly coupled and mutual-beneficial across the whole training process. Our main contributions are summarized in the following." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.729, + 0.892, + 0.804 + ], + "angle": 0, + "content": "- We propose class-aware text prompts which generate prompts based on task-relevant image semantics instead of complete visual information. In this way, we improve the classification accuracy of unseen classes without introducing extra learning ambiguities." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.815, + 0.892, + 0.875 + ], + "angle": 0, + "content": "- We propose text-guided feature tuning which enforces image branch to pay more attention to the task-related representation. As a result, the model avoids deviating attention to the task-irrelevant regions of the image." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.887, + 0.892, + 0.902 + ], + "angle": 0, + "content": "- Benefiting from our mutual learning strategy, our" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.729, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21960" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.138 + ], + "angle": 0, + "content": "method achieves SOTA results on four downstream tasks. Especially, ours significantly outperforms existing methods on the base-to-new generalization task." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.163, + 0.216, + 0.179 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.188, + 0.47, + 0.446 + ], + "angle": 0, + "content": "Vision language models (VLM). The current VLM can be roughly divided into four categories based on the training objectives: image-text matching [3, 20, 27], contrastive loss [19, 21, 22], masked language modeling [38, 39, 46], and masked image modeling [3, 27, 38]. As a milestone, CLIP utilizes 400 million image-text pairs to train a large-scale multi-modal model and demonstrates promising performance on a wide spectrum of tasks including few-shot and zero-shot visual recognition. Motivated by this work, numerous follow-ups have been proposed to improve the effectiveness (e.g., FLIP [24], A-CLIP [45], MaskCLIP [7], and SLIP [30]) or apply it to other domains (e.g., Dense-CLIP [34] and ActionCLIP [42]). The primary limitation of these methods is that hand-crafted prompts are dataset-sensitive and difficult to optimize. We design an automatic and learnable prompts method to enhance the generalization performance of pre-trained models on downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.455, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Prompt learning in NLP. As the scale and complexity of pre-trained language models continue to grow, fine-tuning for specific tasks is becoming increasingly expensive. In contrast, prompt-based approaches are an efficient and lightweight alternative that can be used to generate high-quality text with much lower computational requirements. The original prompts were manually designed prompt templates. While manually designing prompts is advantageous due to their intuitive and comprehensible nature, it also presents a significant challenge that demands extensive experimentation, experience, and language expertise, resulting in high costs. To overcome the limitations of manual prompt design, numerous studies have initiated research into automatically learn appropriate prompts. The automatic prompts can be categorized into two types: discrete prompts and continuous prompts. Discrete prompts consist of various approaches such as prompt mining [17], prompt paraphrasing [49, 10], gradient-based search [40], prompt generation [9] and prompt scoring [5]. On the other hand, continuous prompts include techniques such as prefix tuning [23], tuning initialized with discrete prompts [36] and hard-soft prompt hybrid tuning [26]. These methods have also been applied to the field of computer vision for prompt learning research. However, the task of prompt learning in computer vision is often considered more challenging than in natural language due to the relatively limited high-level semantic information present in visual data with raw pixels." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Prompt learning in vision language models. Prompt learning has been demonstrated to be an effective method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.455 + ], + "angle": 0, + "content": "for improving the performance of pre-trained language models on downstream tasks. Recently, prompt learning has gained increased attention in the context of vision language models. For example, CoOp [51] employs learnable vectors to model contextual words as prompts, and demonstrates that automatic prompts outperform hand-crafted prompts in downstream tasks. CoCoOp [50] extends CoOp by incorporating lightweight neural networks to dynamically generate prompts based on each image, thus mitigating sensitivity to class shifts. Different from the above methods, VP [1], VPT [43], and EVP [16] prompt with images. VP [1] directly combines learnable prompts and pixel-wise input images as new inputs to the model. EVP [43] shrinks the original image before padding the prompts around it, to avoid destroying the original image information. VPT [16] introduces a small amount learnable parameters into the input sequence of each transformer layer and learns them together with a linear head during fine-tuning. Building on the prompt learning approach of the text branch, we propose class-aware text prompt that generates image-dependent and class-aware prompts. Similarly, follow the feature tuning of image branch, we introduce text-guided tuning, which directs the image branch to focus on the task-relevant local regions rather than the global information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.472, + 0.591, + 0.488 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.868, + 0.515 + ], + "angle": 0, + "content": "3.1. Comparisons of CLIP, CoOp, and CoCoOp" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.903 + ], + "angle": 0, + "content": "CLIP comprises two encoders: an image encoder and a text encoder. The image encoder, denotes by \\( F(x) \\), converts an image \\( x \\in \\mathbb{R}^{3 \\times H \\times W} \\) with height of \\( H \\) and width of \\( W \\) into a \\( d \\)-dimensional image feature \\( f_x \\in \\mathbb{R}^{N \\times d} \\), where \\( N \\) is the number of split patches. Meanwhile, the text encoder, denoted as \\( G(t) \\), generates an \\( d \\)-dimensional text representation \\( g_t \\in \\mathbb{R}^{M \\times d} \\) from natural language text \\( t \\), where \\( M \\) is the number of classes. Two encoders are jointly trained using a contrastive loss function that maximizes the cosine similarity of matched pairs and minimizes that of the unmatched pairs. After training, CLIP can be directly used for zero-shot image recognition without requiring fine-tuning of the whole model. Since CLIP is pre-trained on whether an image matches a textual description, the hand-crafted prompt template is employed to convert raw labels into textual descriptions. The most common form of template in CLIP is \"a photo of a [CLASS]\", where the class token is replaced with specific class names such as \"cat\", \"dog\", \"car\", etc. We let the image features \\( f_x \\) of an image \\( x \\) be extracted by an image encoder and the text features \\( g_t \\) be obtained by feeding the prompt description into the text encoder. The prediction task is defined as the classification of an image into one of \\( C \\) categories, which are represented by the set \\( y \\in \\{1, \\dots, C\\} \\). Denote \\( y \\) as the predicted category. Let \\( g_t^i \\) be the \\( i \\)-th dimension of text features \\( g_t \\), with image features" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21961" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.429, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.089, + 0.891, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.895, + 0.4 + ], + "angle": 0, + "content": "Figure 3: Comparisons of three representative prompt learning techniques and our method. The main differences lie in how the text and image branches focus on downstream tasks. CLIP artificially designs prompt templates. CoOp designs automatic prompts using learnable parameters. CoCoOp directly allows text branch to focus on images semantic through Meta-Net. We introduce Class-aware Text Prompts (CTP) and Text Feature Tuning (TFT) to the text and image branches, respectively. The CTP generates prompts based on class-related image information instead of using the identical image semantics like CoCoOp. The TFT enables the image branch to directly focus on downstream tasks. We leverage the contrastive loss function to align task-oriented text and images, making them promote each other for achieving better downstream generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.426, + 0.435, + 0.441 + ], + "angle": 0, + "content": "\\(f_{x}\\), we have the predicted probability of the \\(i\\)-th class:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.45, + 0.47, + 0.495 + ], + "angle": 0, + "content": "\\[\nP (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f _ {x} , g _ {t} ^ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\cos \\left(f _ {x} , g _ {t} ^ {j}\\right) / \\tau\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.534 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot ,\\cdot)\\) denotes the cosine similarity and \\(\\tau\\) is the temperature parameter of the softmax function." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.47, + 0.648 + ], + "angle": 0, + "content": "CoOp replaces the hand-crafted prompts with automatically generated prompts. Specifically, CoOp introduces \\( k \\) learnable context vectors \\( \\{v_{1},\\ldots ,v_{k}\\} \\) to model the context words of the prompts. We define \\( c_{i} \\) as the word embedding of the \\( i \\)-th class name. Then, the prompt of \\( i \\)-th class is denoted as \\( p_i = \\{v_1,\\dots,v_k,c_i\\} \\). Therefore, we have the predicted probability of the \\( i \\)-th class using CoOp method:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.655, + 0.469, + 0.693 + ], + "angle": 0, + "content": "\\[\nP (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f _ {x} , G \\left(p _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\cos \\left(f _ {x} , G \\left(p _ {j}\\right)\\right) / \\tau\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.702, + 0.45, + 0.718 + ], + "angle": 0, + "content": "where \\( G(p_{i}) \\) is the text embedding from text encoder \\( G \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.724, + 0.469, + 0.815 + ], + "angle": 0, + "content": "CoCoOp extends CoOp by generating image-conditional prompts. Specifically, CoCoOp uses Meta-Net to generate the residual vector \\(\\pi\\) based on each image. Each context token is now obtained by \\(v_{k}(x) = v_{k} + \\pi\\). The prompt of the \\(i\\)-th class \\(c_{i}\\) is defined as \\(p_i(x) = \\{v_1(x),\\dots ,v_k(x),c_i\\}\\). As a result, the prediction probability of the \\(i\\)-th class is:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.823, + 0.469, + 0.862 + ], + "angle": 0, + "content": "\\[\nP (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f _ {x} , G \\left(p _ {i} (x)\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left( \\right.\\cos \\left(f _ {x} , G \\left(p _ {j} (x)\\right) / \\tau\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\( G(p_i(x)) \\) is the text embedding conditional on the image \\( x \\) from the text encoder \\( G \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.425, + 0.885, + 0.441 + ], + "angle": 0, + "content": "3.2. Our Task-Oriented Mutual Learning Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Our method consists of two modules, i.e., Class-aware Text Prompts (CTP) and Text-guided Feature Tuning (TFT), as shown in Fig. 3. Compared to CoCoOp, we use CTP to generate class-aware prompts based on task-relevant local image regions instead of the global information. Besides, we use CTP to make the image branch directly pay attention to the task-related image region. We let the two modules be tightly coupled and mutual-beneficial across the training process by optimizing the contrastive loss function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.666 + ], + "angle": 0, + "content": "CTP learns image conditioned discriminative prompts for finer paying attention to semantic-related regions of the images. Specifically, in order to obtain the text semantic-related regions of the image, we leverage the prompt \\( p \\) and the image feature \\( f_{x} \\) to calculate the attention matrix \\( A^{t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.657, + 0.672, + 0.891, + 0.691 + ], + "angle": 0, + "content": "\\[\nA ^ {t} = p f _ {x} ^ {T}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.774 + ], + "angle": 0, + "content": "where \\(A^t \\in \\mathbb{R}^{M \\times N}\\) is the image-to-text attention map. \\(A_{i,j}^t\\) represents the similarity between the \\(i\\)-th class in the text prompts and the \\(j\\)-th patch in the image. In this way, we can query the regions of the images that semantically related to the class information by the attention matrix \\(A^t\\). That is," + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.782, + 0.891, + 0.8 + ], + "angle": 0, + "content": "\\[\nf _ {x} ^ {t} = \\operatorname {s o f t m a x} \\left(A ^ {t}\\right) f _ {x}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.807, + 0.891, + 0.838 + ], + "angle": 0, + "content": "where \\( f_{x}^{t} \\) is the regions correlated to the text of a specific class. We use it to obtain augmented class-aware prompts:" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.845, + 0.891, + 0.863 + ], + "angle": 0, + "content": "\\[\np ^ {a} = p + f _ {x} ^ {t}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(p^a\\) is the text prompts enhanced by semanticallyrelevant image regions. Let \\(p_i^a\\) be the \\(i\\)-th dimension of \\(p^a\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21962" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.147, + 0.098, + 0.324, + 0.111 + ], + "angle": 0, + "content": "(a) Average over 11 datasets" + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.113, + 0.333, + 0.212 + ], + "angle": 0, + "content": "
BaseNewHos
CLIP69.3474.2271.70
CoOp82.6963.2271.66
CoCoOp80.4771.6975.83
ProDA81.5672.3076.65
Ours83.0175.7279.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.441, + 0.098, + 0.523, + 0.112 + ], + "angle": 0, + "content": "(b) ImageNet" + }, + { + "type": "table", + "bbox": [ + 0.385, + 0.113, + 0.58, + 0.212 + ], + "angle": 0, + "content": "
BaseNewHos
CLIP72.4368.1470.22
CoOp76.4767.8871.92
CoCoOp75.9870.4373.10
ProDA75.4070.2372.72
Ours77.4270.4473.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.691, + 0.098, + 0.78, + 0.111 + ], + "angle": 0, + "content": "(c) Caltech101" + }, + { + "type": "table", + "bbox": [ + 0.639, + 0.113, + 0.833, + 0.212 + ], + "angle": 0, + "content": "
BaseNewHos
CLIP96.8494.0095.40
CoOp98.0089.8193.73
CoCoOp97.9693.8195.84
ProDA98.2793.2395.68
Ours98.3194.7596.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.19, + 0.23, + 0.281, + 0.243 + ], + "angle": 0, + "content": "(d) OxfordPets" + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.245, + 0.333, + 0.343 + ], + "angle": 0, + "content": "
BaseNewHos
CLIP91.1797.2694.12
CoOp93.6795.2994.47
CoCoOp95.2097.6996.43
ProDA95.4397.8396.62
Ours95.8697.5596.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.432, + 0.23, + 0.533, + 0.243 + ], + "angle": 0, + "content": "(e) StanfordCars" + }, + { + "type": "table", + "bbox": [ + 0.385, + 0.245, + 0.579, + 0.343 + ], + "angle": 0, + "content": "
BaseNewHos
CLIP63.3774.8968.65
CoOp78.1260.4068.13
CoCoOp70.4973.5972.01
ProDA74.7071.2072.91
Ours76.2974.1775.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.691, + 0.23, + 0.781, + 0.243 + ], + "angle": 0, + "content": "(f) Flowers102" + }, + { + "type": "table", + "bbox": [ + 0.639, + 0.245, + 0.833, + 0.343 + ], + "angle": 0, + "content": "
BaseNewHos
CLIP72.0877.8074.83
CoOp97.6059.6774.06
CoCoOp94.8771.7581.71
ProDA97.7068.6880.66
Ours97.3677.7086.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.198, + 0.362, + 0.272, + 0.375 + ], + "angle": 0, + "content": "(g) Food101" + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.377, + 0.333, + 0.476 + ], + "angle": 0, + "content": "
BaseNewHos
CLIP90.1091.2290.66
CoOp88.3382.2685.19
CoCoOp90.7091.2990.99
ProDA90.3088.5789.43
Ours90.5492.3191.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.427, + 0.362, + 0.537, + 0.375 + ], + "angle": 0, + "content": "(h) FGVCAircraft" + }, + { + "type": "table", + "bbox": [ + 0.385, + 0.377, + 0.579, + 0.476 + ], + "angle": 0, + "content": "
BaseNewHos
CLIP27.1936.2931.09
CoOp40.4422.3028.75
CoCoOp33.4123.7127.74
ProDA36.9034.1335.46
Ours39.4935.3737.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.699, + 0.362, + 0.771, + 0.375 + ], + "angle": 0, + "content": "(i) SUN397" + }, + { + "type": "table", + "bbox": [ + 0.639, + 0.377, + 0.833, + 0.476 + ], + "angle": 0, + "content": "
BaseNewHos
CLIP69.3675.3572.23
CoOp80.6065.8972.51
CoCoOp79.7476.8678.27
ProDA78.6776.9377.79
Ours82.1677.4979.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.205, + 0.495, + 0.256, + 0.508 + ], + "angle": 0, + "content": "(j) DTD" + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.509, + 0.327, + 0.608 + ], + "angle": 0, + "content": "
BaseNewHos
CLIP53.2459.9056.37
CoOp79.4441.1854.24
CoCoOp77.0156.0064.85
ProDA80.6756.4866.44
Ours79.4761.5369.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.445, + 0.495, + 0.524, + 0.507 + ], + "angle": 0, + "content": "(k) EuroSAT" + }, + { + "type": "table", + "bbox": [ + 0.386, + 0.509, + 0.58, + 0.609 + ], + "angle": 0, + "content": "
BaseNewHos
CLIP56.4864.0560.03
CoOp92.1954.7468.69
CoCoOp87.4960.0471.21
ProDA83.9066.0073.88
Ours92.1473.8782.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.702, + 0.495, + 0.771, + 0.507 + ], + "angle": 0, + "content": "(1)UCF101" + }, + { + "type": "table", + "bbox": [ + 0.64, + 0.509, + 0.834, + 0.609 + ], + "angle": 0, + "content": "
BaseNewHos
CLIP70.5377.5073.85
CoOp84.6956.0567.46
CoCoOp82.3373.4577.64
ProDA85.2371.9778.04
Ours84.1277.7480.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.619, + 0.894, + 0.65 + ], + "angle": 0, + "content": "Table 1: Results (%) of the base-to-new generalization task on 11 benchmark datasets. We report the accuracy with CLIP ViT-B/16 model on the base classes (Base), the unseen classes (New), and the harmonic mean of both of them (Hos)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.676, + 0.442, + 0.692 + ], + "angle": 0, + "content": "we then have the predicted probability of the \\(i\\)-th class:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.698, + 0.47, + 0.737 + ], + "angle": 0, + "content": "\\[\nP (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f _ {x} , G \\left(p _ {i} ^ {a}\\right) / \\tau\\right)\\right.}{\\sum_ {j = 1} ^ {C} \\exp \\left( \\right.\\cos \\left(f _ {x} , G \\left(p _ {j} ^ {a}\\right) / \\tau\\right)} .} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.789 + ], + "angle": 0, + "content": "We generate class-aware prompts instead of fusing identical image semantics with prompts of different classes, bringing category discrimination to the specific downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.47, + 0.839 + ], + "angle": 0, + "content": "TFT leverages text features to guide images to focus on task-related regions. Specifically, using the embeddings \\( g^{a} \\) of the augmented prompts \\( p^{a} \\) as input, we have attention:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.845, + 0.469, + 0.864 + ], + "angle": 0, + "content": "\\[\nA ^ {x} = f _ {x} \\left(g ^ {a}\\right) ^ {T}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.47, + 0.905 + ], + "angle": 0, + "content": "where \\(A^x \\in \\mathbb{R}^{N \\times M}\\) denotes text-to-image attention map. \\(A_{i,j}^x\\) represents the similarity between the \\(i\\)-th patch in the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.676, + 0.894, + 0.722 + ], + "angle": 0, + "content": "image and the \\(j\\)-th class in the text representation. Similar to image-to-text, we use it to query the class-related part of the text correlated to the image, augmenting image features:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.728, + 0.892, + 0.746 + ], + "angle": 0, + "content": "\\[\nf ^ {a} = \\operatorname {s o f t m a x} \\left(A ^ {x}\\right) g ^ {a} + f _ {x}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.751, + 0.892, + 0.782 + ], + "angle": 0, + "content": "where \\( f^a \\) is the augmented image embeddings. We thus let image branch focus on the tasks-related representation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.787, + 0.894, + 0.862 + ], + "angle": 0, + "content": "Augmented contrastive loss function is then employed to further align class-aware text embedding and text-guided image features on specific downstream tasks. The predicted probability of the \\(i\\)-th class, which is used to calculate the contrastive loss, after mutual augmentation is:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.867, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\nP (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f ^ {a} , g _ {i} ^ {a}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\cos \\left(f ^ {a} , g _ {i} ^ {a} / \\tau\\right)\\right)}. \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21963" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.426, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.088, + 0.852, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.296, + 0.893, + 0.327 + ], + "angle": 0, + "content": "Figure 4: Absolute improvement over CoCoOp in the base-to-new generalization task. Compared to CoCoOp, Our method achieves improvement on both base (left sub-figure) and new (right sub-figure) classes on most of the datasets." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.34, + 0.891, + 0.46 + ], + "angle": 0, + "content": "
ImageNetCaltech101OxfordPedsStanfordCarsFlowers102Food101FGVCAircraftSUN397DTDEuroSATUCF101Average
CLIP68.6389.3688.9965.6770.4989.2327.1265.2946.0254.1769.8366.80
CoOp71.5195.5393.3174.2595.7087.2334.1874.8268.4677.8277.2977.28
CoCoOp71.0293.4393.9371.2187.3487.3932.0372.3263.8472.7877.4074.79
Ours72.9095.9093.9679.1096.7389.9538.7279.3772.4981.0083.4580.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.281, + 0.469, + 0.687, + 0.485 + ], + "angle": 0, + "content": "Table 2: Results (%) of 16-shot learning task on 11 datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.558 + ], + "angle": 0, + "content": "Task-targeted semantic information is transferred between the two branches by minimizing the augmented contrastive loss. We merge probability before and after augmentation:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.565, + 0.49, + 0.615 + ], + "angle": 0, + "content": "\\[\nP (y = i \\mid x) = \\frac {\\exp \\left(\\left(\\cos \\left(f , g _ {i}\\right) + \\lambda \\left(\\cos \\left(f ^ {a} , g _ {i} ^ {a}\\right)\\right) / \\tau\\right)\\right.}{\\sum_ {j = 1} ^ {C} \\exp \\left( \\right.\\left(\\cos \\left(f , g _ {i}\\right) + \\lambda \\left(\\cos \\left(f ^ {a} , g _ {i} ^ {a}\\right)\\right) / \\tau\\right)} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.47, + 0.678 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the balance hyper-parameter, which is analyzed in our experiments. We let the two different modalities tightly coupled and mutual beneficial across the whole training process by performing the contrastive optimization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.688, + 0.21, + 0.706 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.47, + 0.79 + ], + "angle": 0, + "content": "We evaluate the performance of our method on four generalization tasks, including 1) generalization from base classes to new classes; 2) few-shot classification; 3) cross-dataset transfer; 4) domain generalization. After that, we provide extensive ablation studies and in-depth analyses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Datasets. Following [33, 51], we use 11 image recognition datasets for the tasks of base-to-new generalization, few-shot classification and cross-dataset transfer. It contains generic image classification datasets (ImageNet [6] and Caltech101 [8]), fine-grained classification datasets (Oxford Pets [32], StanfordCars [18], Flowers102 [31], Food101 [2] and FGVCAircraft [29]), scene recognition ( SUN397" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.893, + 0.604 + ], + "angle": 0, + "content": "[44]), action recognition (UCF101 [37]), texture classification (DTD [4]), and satellite imagery recognition (EuroSAT [11]). For the domain generalization task, we use ImageNet as the source dataset and select ImageNetV2 [35], ImageNet-Sketch [41], ImageNet-A [13], and ImageNet-R [12], which are the ImageNet variants, as the target." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.734 + ], + "angle": 0, + "content": "Training Details. By following [50, 51], we use the best visual backbone available in CLIP, i.e., ViT-B/16, throughout the experiments. We train 10 epochs using SGD optimizer with base learning rate of 0.002 and cosine decay schedule. We set the hyper-parameter \\(\\lambda\\) in Eq. (11) to 0.2 for all experiments, and provide sensitivity analyses in Fig. 5. We run all the experiments three times with different random seeds and report the average classification accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.894, + 0.837 + ], + "angle": 0, + "content": "Baselines. We compare our method with 4 baselines. (1) Zero-shot CLIP [33] with hand-crafted prompts. (2) CoOp [51], using automatically generated prompts from few data. (3) CoCoOp [50], dynamically generating prompts conditioned on the images. (4) ProDA [28], which learns prompts from few data samples and mitigates the domain gap." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.847, + 0.857, + 0.862 + ], + "angle": 0, + "content": "4.1. Generalization From Base to New Classes" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Following the previous works, we split the classes equally into two groups for each dataset: one as base and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21964" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.208 + ], + "angle": 0, + "content": "
SourceTarget
ImageNetCaletech101OxfordPedsStanfordCarsFlowers102Food101FGVCAircraftSUN397DTDEuroSATUCF101
CoOp71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.55
CoCoOp71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.21
Ours72.9095.7390.2265.1469.8986.3823.3266.4946.4747.2467.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.218, + 0.882, + 0.234 + ], + "angle": 0, + "content": "Table 3: Results of cross-dataset transfer task. Each method is trained on the source dataset and evaluated on the target." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.246, + 0.889, + 0.355 + ], + "angle": 0, + "content": "
SourceTarget
ImageNetImageNetV2ImageNet-SketchImageNet-AImageNet-R
CLIP66.7360.8346.1547.7773.96
CoOp71.5164.2047.9949.7175.21
CoCoOp71.0264.0748.7550.6376.18
Ours72.9064.5749.1150.9476.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.366, + 0.888, + 0.382 + ], + "angle": 0, + "content": "Table 4: Results of domain generalization task. Each method is trained on ImageNet and evaluated on ImageNet variants." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.473, + 0.741 + ], + "angle": 0, + "content": "the other as new. The learnable modules are trained exclusively on the base classes, while evaluation is carried out separately on both the base and new classes to testify generalization ability. We report the results on 11 benchmarks in Table 1. Although compared to CoOp, CoCoOp significantly narrows generalization gap in unseen classed, but it decreases the accuracy in seen classes from \\(82.69\\%\\) to \\(80.47\\%\\). We attribute it to the homogeneous prompts of CoCoOp, which weakens the discriminative semantics of different categories. In comparison, our method improves the accuracy in unseen classes from \\(80.47\\%\\) to \\(83.01\\%\\) by prompting each text label with corresponding image information. Benefit from the mutual learning of our CTP and TFT modules, our method further improves the accuracy in unseen classes from \\(71.69\\%\\) to \\(75.72\\%\\), even surpasses the accuracy of CLIP hand-crafted prompts. We provide a detailed comparisons of CoCoOp and our method of per-dataset improvement in Fig. 4. Our method gains significant improvements over CoCoOp in both seen and unseen classes on 10 out of 11 recognition datasets. Surprisingly, our method significantly improves CoCoOp by more than \\(10\\%\\) in unseen classes on SUN397 and Flowers102 datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.294, + 0.77 + ], + "angle": 0, + "content": "4.2. Few-Shot Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We report few-shot classification results in Table 2. Our method surpasses baseline methods on all datasets in the few-shot setting. Especially, our method outperforms CoCoOp by \\(9.39\\%\\), \\(8.65\\%\\), and \\(8.22\\%\\) on Flowers102, DTD, and EuroSAT, respectively, and the average improvement over 11 datasets is \\(5.53\\%\\). Our method also achieve \\(2\\%\\) on the challenging dataset of ImageNet. The above experiments show the great discriminative ability of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.715, + 0.423 + ], + "angle": 0, + "content": "4.3. Cross-Dataset Transfer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.568 + ], + "angle": 0, + "content": "We then evaluate the generalization ability of our method on more challenging cross-dataset tasks. In this setting, we learn multi-modal prompts on ImageNet of 1000 classes. The effectiveness of the learned prompts is then tested on 10 datasets containing generic and fine-grained image classification, scene recognition, and texture classification. The results are reported in Table 3. Our method achieves the best average accuracy on the 11 datasets, especially ImageNet. It demonstrates the great transfer ability of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.717, + 0.59 + ], + "angle": 0, + "content": "4.4. Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.825 + ], + "angle": 0, + "content": "The domain generalization setting evaluates the generalization ability of the model on the target domain that is similar to but different from the source domain [48, 47]. Zero-shot CLIP introduces no additional training parameters and exhibits great robustness to naturally distribution shifts. Other methods use few samples to train learnable parameters, there is a risk of overfitting the source distribution. Therefore, we conduct experiments using ImageNet as the source domain and evaluate the ability of generalizing to unknown on four ImageNet variants. The results are shown in Table 4. Our method achieves significant performance on the 4 ImageNet variant datasets. It verifies that our method improves the classification ability of the source domain dataset while maintaining the generalization on the target domain." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.673, + 0.849 + ], + "angle": 0, + "content": "4.5. Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of each module. To evaluate the effectiveness of Class-aware Text Prompts (CTP) and Text-guided Feature Tuning (TFT) of our method, we conduct ablation" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21965" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.89, + 0.364 + ], + "angle": 0, + "content": "
MethodAverageImageNetCaltech101OxfordPeds
BaseNewHosBaseNewHosBaseNewHosBaseNewHos
A82.6963.2271.6676.4767.8871.9298.0089.8193.7393.6795.2994.47
B82.3872.4476.6476.9669.6273.1198.2394.2496.1995.6197.9796.78
C82.9372.9877.2477.2169.8673.3598.4492.7295.4995.4997.8196.64
Ours83.0175.7279.0277.4270.4473.7798.3194.7596.5095.8697.5596.70
MethodStanfordCarsFlowers102Food101FGVCAircraft
BaseNewHosBaseNewHosBaseNewHosBaseNewHos
A78.1260.4068.1397.6059.6774.0688.3382.2685.1940.4422.3028.75
B74.6273.6874.1596.7266.4278.7690.3091.4790.8836.4134.3935.37
C75.8474.5375.1897.3274.8684.6390.5691.6591.1037.8233.1735.34
Ours76.2974.1775.2297.3677.7086.4390.5492.3191.4239.4935.3737.32
MethodSUN397DTDEuroSATUCF101
BaseNewHosBaseNewHosBaseNewHosBaseNewHos
A80.6065.8972.5179.4441.1854.2492.1954.7468.6984.6956.0567.46
B81.7376.8979.2480.1851.7962.9391.7067.6277.8483.7272.7277.83
C82.2976.2479.1581.7154.7465.5690.1060.5272.4185.4076.6880.81
Ours82.1677.4979.7679.4761.5369.3692.1473.8782.0084.1277.7480.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.375, + 0.893, + 0.407 + ], + "angle": 0, + "content": "Table 5: Ablation studies of our method on 11 datasets. Three ablation cases are considered: A: Ours w/o TVP w/o FTP. B: Ours w/o TVP. C: Ours w/o FTP. TVP is the text-reorganized vision prompt, and FTP is the fine-grained text prompt." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.431, + 0.466, + 0.55 + ], + "angle": 0, + "content": "
Prompt LearningFeature TuningAccuracy (%)
MLP-PLCTPMLP-FTTFT
✓71.66 (CoOp)
75.83 (+4.17)
✓76.64 (+4.98)
✓75.94 (+4.28)
✓77.24 (+5.58)
✓✓77.05 (+5.39)
✓✓79.02 (+7.36)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.669 + ], + "angle": 0, + "content": "Table 6: Comparison of different structures for prompt learning and feature tuning. The average results of harmonic mean of from-base-to-new generalization task on 11 datasets are reported. In compared to our attention design in CTP and TFT modules, MLP-PL and MLP-FT are designed using the Linear-ReLU-Linear block setting of [50]. Improvements over the baseline of CoOp, are marked in green." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.818 + ], + "angle": 0, + "content": "experiments on 11 datasets, as reported in Table 5. In most cases, each module significantly improves the performance of the model. For average results, CTP and TFT improves the results by \\(5.58\\%\\) and \\(4.98\\%\\), respectively, and the combination of them improves the results by \\(7.36\\%\\). It shows the effectiveness of the two branches of text-to-image and image-to-text, the mutual learning of the two modules further improves the performance on downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Comparison of different structure design of multi-modal mutual learning. To further provide in-depth analysis about our mutual learning, we further explore two vanilla structures: (1) MLP-PL: The image features are forwarded to a block of Linear-ReLU-Linear, borrowed from [50], and" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.431, + 0.688, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.431, + 0.889, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.578, + 0.892, + 0.609 + ], + "angle": 0, + "content": "Figure 5: Sensitivity analysis of \\(\\lambda\\), with base, new, and host metrics, on UCF101 (left) and Caltech101 (right) datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.893, + 0.901 + ], + "angle": 0, + "content": "then added to the text for augmenting it (the same to us). (2) MLP-FT: The text prompts are forwarded to the Linear-ReLU-Linear block, and then added to the image for augmenting it (the same to us). In comparison, our class-aware text prompts (CTP) module and text-guided feature tuning (TFT) module, adopt text-image attention to learn the augmented features instead of the Linear-ReLU-Linear block. We report the results of the different designs in Table 6. First, we find that combining MLP-PL & MLP-FT and CTP & TFT can both improve the results compared with using either of them. It indicates that both prompt learning and feature tuning are important to achieve better results. Second, compared with the design of Linear-ReLU-Linear block, our design of text-image attention further improves performance by \\(0.81\\%\\) and \\(1.3\\%\\) for prompt learning and feature tuning, respectively. It demonstrates the effectiveness of our design of attention, which helps the model to focus on class-aware and task-related semantics. Third, com" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21966" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "pared with CoOp, both of the designs could improve the final results by large margins. The key factor of our mutual learning to achieve significant performance is the task-related alignment of vision and language in latent space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.158, + 0.47, + 0.219 + ], + "angle": 0, + "content": "Sensitivity Analysis of \\(\\lambda\\). We evaluate the parameter sensitivity of \\(\\lambda\\) of Eq. (11) in Fig. 5. The results suggest that the performance of our method is generally robust to \\(\\lambda\\), indicating a wide range of \\(\\lambda\\) works well in downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.231, + 0.196, + 0.246 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.256, + 0.47, + 0.451 + ], + "angle": 0, + "content": "In this paper, we introduce task-oriented multi-modal mutual learning for adapting large vision-language models to downstream vision tasks. We propose class-aware text prompt and text-guided feature tuning to unleash the potential of the vision-language model by re-activating its task-related representation abilities. Our method yields impressive generalization performance on a wide range of vision tasks and datasets. We hope the presented findings and insights in this paper could benefit the following works in designing more efficient and effective adaptation methods. For the future work, we think it is interesting to extend the adaptation of vision language models to more vision tasks, such as semantic segmentation, object detection, etc." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.469, + 0.513 + ], + "angle": 0, + "content": "Limitations. Similar to CoCoOp, we learn image-conditioned representations, thus the batch-size of training need to be set to 1, which is not efficient enough for learning. We aim to solve this efficiency issue in the future work." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.525, + 0.268, + 0.541 + ], + "angle": 0, + "content": "6. Acknowledgements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.469, + 0.626 + ], + "angle": 0, + "content": "This work is supported by the Innovation Capacity Construction Project of Jilin Province Development and Reform Commission (2021FGWCXNLJSSZ10) and the Fundamental Research Funds for the Central Universities, JLU. Zichang Tan is the project lead." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.639, + 0.173, + 0.654 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.663, + 0.47, + 0.717 + ], + "angle": 0, + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting largescale models. arXiv preprint arXiv:2203.17274, 1(3):4, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.72, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 446-461. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.79, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[3] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX, pages 104-120. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[4] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.663, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[5] Joe Davison, Joshua Feldman, and Alexander M Rush. Commonsense knowledge mining from pretrained models. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pages 1173-1178, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.207, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.264, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[7] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[8] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.406, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[9] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pretrained language models better few-shot learners. arXiv preprint arXiv:2012.15723, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[10] Adi Haviv, Jonathan Berant, and Amir Globerson. Bertese: Learning to speak to bert. arXiv preprint arXiv:2103.05327, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[12] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[13] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Computer Vision-ECCV 2022: 17th" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21967" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 709-727. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[17] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[19] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[20] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[21] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. Unimo: Towards unified-modal understanding and generation via cross-modal contrastive learning. arXiv preprint arXiv:2012.15409, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[22] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 121-137. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[23] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[24] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. arXiv preprint arXiv:2212.00794, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[25] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[26] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands, too. arXiv preprint arXiv:2103.10385, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[27] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[28] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[30] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI, pages 529-544. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[32] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[34] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[35] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International conference on machine learning, pages 5389-5400. PMLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[36] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[37] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild. Center for Research in Computer Vision, 2(11), 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[38] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[39] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[40] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing nlp. arXiv preprint arXiv:1908.07125, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[41] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[42] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[43] Junyang Wu, Xianhang Li, Chen Wei, Huiyu Wang, Alan Yuille, Yuyin Zhou, and Cihang Xie. Unleashing the power" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21968" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "of visual prompting at the pixel level. arXiv preprint arXiv:2212.10556, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[44] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.247 + ], + "angle": 0, + "content": "[45] Yifan Yang, Weiquan Huang, Yixuan Wei, Houwen Peng, Xinyang Jiang, Huiqiang Jiang, Fangyun Wei, Yin Wang, Han Hu, Lili Qiu, et al. Attentive mask clip. arXiv preprint arXiv:2212.08653, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.317 + ], + "angle": 0, + "content": "[46] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.319, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[47] Junkun Yuan, Xu Ma, Defang Chen, Kun Kuang, Fei Wu, and Lanfen Lin. Label-efficient domain generalization via collaborative exploration and generalization. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2361-2370, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[48] Junkun Yuan, Xu Ma, Defang Chen, Kun Kuang, Fei Wu, and Lanfen Lin. Domain-specific bias filtering for single labeled domain generalization. International Journal of Computer Vision, 131(2):552-571, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[49] Weizhe Yuan, Graham Neubig, and Pengfei Liu. Bartscore: Evaluating generated text as text generation. Advances in Neural Information Processing Systems, 34:27263-27277, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[51] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21969" + } + ] +] \ No newline at end of file diff --git a/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/b6090cd5-35b5-49d1-9775-853efd91c055_origin.pdf b/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/b6090cd5-35b5-49d1-9775-853efd91c055_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..20d5e9e7dce91b260b6eb7c972f99a4e2ccb310b --- /dev/null +++ b/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/b6090cd5-35b5-49d1-9775-853efd91c055_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d1fbfb3a9ce6fe5e1c7c7484fe2f9118b44adc670dc86a7bc4710d3b16af3b9 +size 735285 diff --git a/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/full.md b/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7e9b04e77e72c5f28f2e88258ac95a448b9e0d27 --- /dev/null +++ b/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/full.md @@ -0,0 +1,352 @@ +# Task-Oriented Multi-Modal Mutual Learning for Vision-Language Models + +Sifan Long $^{1,2,3\dagger*}$ Zhen Zhao $^{3,4\dagger*}$ Junkun Yuan $^{3,5\dagger*}$ Zichang Tan $^{3}$ Jiangjiang Liu $^{3}$ Luping Zhou $^{4}$ Shengsheng Wang $^{1,2‡}$ Jingdong Wang $^{3‡}$ + +$^{1}$ College of Computer Science and Technology, Jilin University, Jilin, China + $^{2}$ Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Jilin, China + +$^{3}$ Baidu VIS $^{4}$ University of Sydney $^{5}$ Zhejiang University longsf22@mails.jlu.edu.cn {zhen.zhao, luping.zhou}@sydney.edu.au yuanjk@zju.edu.cn wss@jlu.edu.cn {tanzichang, liujiangjiang, wangjingdong}@baidu.com + +# Abstract + +Prompt learning has become one of the most efficient paradigms for adapting large pre-trained vision-language models to downstream tasks. Current state-of-the-art methods, like CoOp and ProDA, tend to adopt soft prompts to learn an appropriate prompt for each specific task. Recent CoCoOp further boosts the base-to-new generalization performance via an image-conditional prompt. However, it directly fuses identical image semantics to prompts of different labels and significantly weakens the discrimination among different classes as shown in our experiments. Motivated by this observation, we first propose a class-aware text prompt (CTP) to enrich generated prompts with label-related image information. Unlike CoCoOp, CTP can effectively involve image semantics and avoid introducing extra ambiguities into different prompts. On the other hand, instead of reserving the complete image representations, we propose text-guided feature tuning (TFT) to make the image branch attend to class-related representation. A contrastive loss is employed to align such augmented text and image representations on downstream tasks. In this way, the image-to-text CTP and text-to-image TFT can be mutually promoted to enhance the adaptation of VLMs for downstream tasks. Extensive experiments demonstrate that our method outperforms the existing methods by a significant margin. Especially, compared to CoCoOp, we achieve an average improvement of $4.03\%$ on new classes and $3.19\%$ on harmonic-mean over eleven classification benchmarks. + +# 1. Introduction + +Recently, large vision-language models (VLM), such as CLIP [33] and ALIGN [15], which employ language as + +![](images/e60548bb39680cb70e3e638e536d9fedbf0ab699aae57b4f901a623b562a8cb5.jpg) +Figure 1: Comparisons between CoCoOp [50] and our method. The cosine distance between the positive and the negative prompts, which quantifies the class discrimination, and the average accuracy on benchmarks are reported. + +supervision signal instead of discrete labels, have shown impressive generalization performance in a wide range of downstream vision tasks. Their multi-modal interaction nature delivers open-vocabulary support and achieves amazing zero-shot classification performance. Despite their impressive transferable abilities, as discussed in [25], it is essential to re-activate specific representation capabilities for optimal performance in certain downstream tasks. Considering their hundreds of millions or billions of parameters, attempting to fine-tune the entire model is impractical and even jeopardizes the well-established representation space [14]. To this end, many recent studies have centred on the efficient and effective adaptation of pre-trained and frozen large VLMs for the specific downstream tasks [28, 51, 50]. + +Prompt, a simple, compact, and viable strategy, has become the leading solution for deploying large pre-trained VLMs into certain downstream tasks. CLIP [33] utilizes hand-crafted prompts to achieve impressive zero- and few-shot classification performance. Nevertheless, manually-designed prompts require significant domain knowledge and can be highly time-consuming and sub-optimal for specific downstream tasks. To address this problem, later studies [28, 51] adopt soft prompts to learn an appropriate text prompt via optimizing a contrastive loss on different text labels. CoCoOp [50] further highlights the limitations of such static soft prompts and proposes learning image-dependent prompts conditioned on individual instances rather than fixed prompts. It achieves great performance gains on unseen classes by adding high-level image embedding to text prompts. However, compared to CoOp with static prompts, CoCoOp essentially fuses identical image semantics with different text labels, leading to inevitable learning ambiguity and resulting in an average performance drop of $2.22\%$ on base classes on 11 datasets (see Table 1). For example, it may associate the dog image semantics with a prompt that references the [class] of a cat. When using the cosine distance to measure the differences between the positive and negative text prompts, as shown in Fig. 1, CoCoOp holds low distance values, suggesting that it brings significant learning ambiguities to text prompts. Therefore, we argue that text prompts should not only condition on distinct input images for better generalization abilities, but also adapt to different classes to eliminate the potential ambiguities. + +To achieve this goal, we propose Class-aware Text Prompts (CTP), which leverages label-related image information to generate finer prompts. Specifically, we first contact learnable context vectors and each class label to model the initial prompt sentences. Then we leverage these class prompt sentences to query their corresponding image regions and representations. Corresponding related image features are subsequently added to initial class prompt sentences to produce the final text prompts. In this way, generated image-dependent and class-aware prompts can better concentrate on the image information in a more precise manner. As shown in Fig. 1, our method enjoys better discrimination between positive and negative prompts and consistently outperforms CoCoOp on 11 classification datasets. + +On the other hand, we identify a critical problem in these text prompt-based strategies: the image branch is ignored and not adjusted to specific downstream tasks. As shown in Fig. 2 (CoCoOp), on the task of identifying birds, the output image feature, without further tuning, can be distracted to leaves of the same color. Similarly, it also wrongly highlights the beer foam that is of a similar shape to recognize golf balls. Since the final recognition is jointly inferred by both text and image branches, such an issue may degrade the classification performance. Thus it is necessary + +![](images/f7f5394eb598b2671856cc56aa696bef15416101573a59ba6cb24821420bc5ef.jpg) + +![](images/e38af2466fc343501b5baa10c409334eb80db5486b5af21f7b9fbb4c688f9545.jpg) + +![](images/e9a7cf8de1641bb4206e1fbffe2fefc27df51a4b2f8ee6eee649980066ad3505.jpg) + +![](images/3b30c3b4c2277835013d0ac2ba1f5a9e39a0ee9fe738bcc8bb821158d57fff41.jpg) +Figure 2: Comparisons of attention map visualization for CoCoOp and our method on ImageNet. Our method obtains better average accuracy of both base and new classes across 11 datasets by paying attention to task-related regions. + +![](images/0cda44d5167afcc175abc3c4f628891a8a03fc9fab3ad1f01e4d862cbbee4cf4.jpg) + +![](images/ebc7cca7075bb6fa3ef540e9e2e4d35a534f320a27c23f184c98be0d7fc0a7b5.jpg) + +to tune the image features further so that the image branch can focus more on the tasks-related representation. We then propose Text-guided Feature Tuning (TFT), which leverages encoded text embedding to guide image representation more on task-related regions. As shown in Fig. 2 (ours), our method successfully focuses on task-related regions, i.e., birds and golf balls. We then leverage the contrastive loss function to further align class-aware text embedding and text-guided image features on certain downstream tasks. + +In summary, we propose a new task-oriented multimodal mutual-learning method, which well-integrates our designed class-aware text prompts and text-guided feature tuning for fast adaptation of frozen VLMs on downstream tasks. Image features can help construct image-dependent class-aware text prompts, leading to more discriminative text embedding. Simultaneously, improved text embedding can further guide the image branch attending to class-related representation. In this way, these two different modality branches can be tightly coupled and mutual-beneficial across the whole training process. Our main contributions are summarized in the following. + +- We propose class-aware text prompts which generate prompts based on task-relevant image semantics instead of complete visual information. In this way, we improve the classification accuracy of unseen classes without introducing extra learning ambiguities. +- We propose text-guided feature tuning which enforces image branch to pay more attention to the task-related representation. As a result, the model avoids deviating attention to the task-irrelevant regions of the image. +- Benefiting from our mutual learning strategy, our + +method achieves SOTA results on four downstream tasks. Especially, ours significantly outperforms existing methods on the base-to-new generalization task. + +# 2. Related work + +Vision language models (VLM). The current VLM can be roughly divided into four categories based on the training objectives: image-text matching [3, 20, 27], contrastive loss [19, 21, 22], masked language modeling [38, 39, 46], and masked image modeling [3, 27, 38]. As a milestone, CLIP utilizes 400 million image-text pairs to train a large-scale multi-modal model and demonstrates promising performance on a wide spectrum of tasks including few-shot and zero-shot visual recognition. Motivated by this work, numerous follow-ups have been proposed to improve the effectiveness (e.g., FLIP [24], A-CLIP [45], MaskCLIP [7], and SLIP [30]) or apply it to other domains (e.g., Dense-CLIP [34] and ActionCLIP [42]). The primary limitation of these methods is that hand-crafted prompts are dataset-sensitive and difficult to optimize. We design an automatic and learnable prompts method to enhance the generalization performance of pre-trained models on downstream tasks. + +Prompt learning in NLP. As the scale and complexity of pre-trained language models continue to grow, fine-tuning for specific tasks is becoming increasingly expensive. In contrast, prompt-based approaches are an efficient and lightweight alternative that can be used to generate high-quality text with much lower computational requirements. The original prompts were manually designed prompt templates. While manually designing prompts is advantageous due to their intuitive and comprehensible nature, it also presents a significant challenge that demands extensive experimentation, experience, and language expertise, resulting in high costs. To overcome the limitations of manual prompt design, numerous studies have initiated research into automatically learn appropriate prompts. The automatic prompts can be categorized into two types: discrete prompts and continuous prompts. Discrete prompts consist of various approaches such as prompt mining [17], prompt paraphrasing [49, 10], gradient-based search [40], prompt generation [9] and prompt scoring [5]. On the other hand, continuous prompts include techniques such as prefix tuning [23], tuning initialized with discrete prompts [36] and hard-soft prompt hybrid tuning [26]. These methods have also been applied to the field of computer vision for prompt learning research. However, the task of prompt learning in computer vision is often considered more challenging than in natural language due to the relatively limited high-level semantic information present in visual data with raw pixels. + +Prompt learning in vision language models. Prompt learning has been demonstrated to be an effective method + +for improving the performance of pre-trained language models on downstream tasks. Recently, prompt learning has gained increased attention in the context of vision language models. For example, CoOp [51] employs learnable vectors to model contextual words as prompts, and demonstrates that automatic prompts outperform hand-crafted prompts in downstream tasks. CoCoOp [50] extends CoOp by incorporating lightweight neural networks to dynamically generate prompts based on each image, thus mitigating sensitivity to class shifts. Different from the above methods, VP [1], VPT [43], and EVP [16] prompt with images. VP [1] directly combines learnable prompts and pixel-wise input images as new inputs to the model. EVP [43] shrinks the original image before padding the prompts around it, to avoid destroying the original image information. VPT [16] introduces a small amount learnable parameters into the input sequence of each transformer layer and learns them together with a linear head during fine-tuning. Building on the prompt learning approach of the text branch, we propose class-aware text prompt that generates image-dependent and class-aware prompts. Similarly, follow the feature tuning of image branch, we introduce text-guided tuning, which directs the image branch to focus on the task-relevant local regions rather than the global information. + +# 3. Method + +# 3.1. Comparisons of CLIP, CoOp, and CoCoOp + +CLIP comprises two encoders: an image encoder and a text encoder. The image encoder, denotes by $F(x)$ , converts an image $x \in \mathbb{R}^{3 \times H \times W}$ with height of $H$ and width of $W$ into a $d$ -dimensional image feature $f_x \in \mathbb{R}^{N \times d}$ , where $N$ is the number of split patches. Meanwhile, the text encoder, denoted as $G(t)$ , generates an $d$ -dimensional text representation $g_t \in \mathbb{R}^{M \times d}$ from natural language text $t$ , where $M$ is the number of classes. Two encoders are jointly trained using a contrastive loss function that maximizes the cosine similarity of matched pairs and minimizes that of the unmatched pairs. After training, CLIP can be directly used for zero-shot image recognition without requiring fine-tuning of the whole model. Since CLIP is pre-trained on whether an image matches a textual description, the hand-crafted prompt template is employed to convert raw labels into textual descriptions. The most common form of template in CLIP is "a photo of a [CLASS]", where the class token is replaced with specific class names such as "cat", "dog", "car", etc. We let the image features $f_x$ of an image $x$ be extracted by an image encoder and the text features $g_t$ be obtained by feeding the prompt description into the text encoder. The prediction task is defined as the classification of an image into one of $C$ categories, which are represented by the set $y \in \{1, \dots, C\}$ . Denote $y$ as the predicted category. Let $g_t^i$ be the $i$ -th dimension of text features $g_t$ , with image features + +![](images/7e2bafb580134e9fcc143ee5ebe227bc6853de91852f569c3261584a883539ff.jpg) +Figure 3: Comparisons of three representative prompt learning techniques and our method. The main differences lie in how the text and image branches focus on downstream tasks. CLIP artificially designs prompt templates. CoOp designs automatic prompts using learnable parameters. CoCoOp directly allows text branch to focus on images semantic through Meta-Net. We introduce Class-aware Text Prompts (CTP) and Text Feature Tuning (TFT) to the text and image branches, respectively. The CTP generates prompts based on class-related image information instead of using the identical image semantics like CoCoOp. The TFT enables the image branch to directly focus on downstream tasks. We leverage the contrastive loss function to align task-oriented text and images, making them promote each other for achieving better downstream generalization performance. + +![](images/a57d25f8bf6d3de5b7ea94a9651d697290ac0c4115e83f379d47e10c4f72f977.jpg) + +$f_{x}$ , we have the predicted probability of the $i$ -th class: + +$$ +P (y = i \mid x) = \frac {\exp \left(\cos \left(f _ {x} , g _ {t} ^ {i}\right) / \tau\right)}{\sum_ {j = 1} ^ {C} \exp \left(\cos \left(f _ {x} , g _ {t} ^ {j}\right) / \tau\right)}, \tag {1} +$$ + +where $\cos (\cdot ,\cdot)$ denotes the cosine similarity and $\tau$ is the temperature parameter of the softmax function. + +CoOp replaces the hand-crafted prompts with automatically generated prompts. Specifically, CoOp introduces $k$ learnable context vectors $\{v_{1},\ldots ,v_{k}\}$ to model the context words of the prompts. We define $c_{i}$ as the word embedding of the $i$ -th class name. Then, the prompt of $i$ -th class is denoted as $p_i = \{v_1,\dots,v_k,c_i\}$ . Therefore, we have the predicted probability of the $i$ -th class using CoOp method: + +$$ +P (y = i \mid x) = \frac {\exp \left(\cos \left(f _ {x} , G \left(p _ {i}\right)\right) / \tau\right)}{\sum_ {j = 1} ^ {C} \exp \left(\cos \left(f _ {x} , G \left(p _ {j}\right)\right) / \tau\right)}, \tag {2} +$$ + +where $G(p_{i})$ is the text embedding from text encoder $G$ . + +CoCoOp extends CoOp by generating image-conditional prompts. Specifically, CoCoOp uses Meta-Net to generate the residual vector $\pi$ based on each image. Each context token is now obtained by $v_{k}(x) = v_{k} + \pi$ . The prompt of the $i$ -th class $c_{i}$ is defined as $p_i(x) = \{v_1(x),\dots ,v_k(x),c_i\}$ . As a result, the prediction probability of the $i$ -th class is: + +$$ +P (y = i \mid x) = \frac {\exp \left(\cos \left(f _ {x} , G \left(p _ {i} (x)\right)\right) / \tau\right)}{\sum_ {j = 1} ^ {C} \exp \left( \right.\cos \left(f _ {x} , G \left(p _ {j} (x)\right) / \tau\right)}, \tag {3} +$$ + +where $G(p_i(x))$ is the text embedding conditional on the image $x$ from the text encoder $G$ . + +# 3.2. Our Task-Oriented Mutual Learning Method + +Our method consists of two modules, i.e., Class-aware Text Prompts (CTP) and Text-guided Feature Tuning (TFT), as shown in Fig. 3. Compared to CoCoOp, we use CTP to generate class-aware prompts based on task-relevant local image regions instead of the global information. Besides, we use CTP to make the image branch directly pay attention to the task-related image region. We let the two modules be tightly coupled and mutual-beneficial across the training process by optimizing the contrastive loss function. + +CTP learns image conditioned discriminative prompts for finer paying attention to semantic-related regions of the images. Specifically, in order to obtain the text semantic-related regions of the image, we leverage the prompt $p$ and the image feature $f_{x}$ to calculate the attention matrix $A^{t}$ : + +$$ +A ^ {t} = p f _ {x} ^ {T}, \tag {4} +$$ + +where $A^t \in \mathbb{R}^{M \times N}$ is the image-to-text attention map. $A_{i,j}^t$ represents the similarity between the $i$ -th class in the text prompts and the $j$ -th patch in the image. In this way, we can query the regions of the images that semantically related to the class information by the attention matrix $A^t$ . That is, + +$$ +f _ {x} ^ {t} = \operatorname {s o f t m a x} \left(A ^ {t}\right) f _ {x}, \tag {5} +$$ + +where $f_{x}^{t}$ is the regions correlated to the text of a specific class. We use it to obtain augmented class-aware prompts: + +$$ +p ^ {a} = p + f _ {x} ^ {t}, \tag {6} +$$ + +where $p^a$ is the text prompts enhanced by semanticallyrelevant image regions. Let $p_i^a$ be the $i$ -th dimension of $p^a$ , + +(a) Average over 11 datasets + +
BaseNewHos
CLIP69.3474.2271.70
CoOp82.6963.2271.66
CoCoOp80.4771.6975.83
ProDA81.5672.3076.65
Ours83.0175.7279.02
+ +(b) ImageNet + +
BaseNewHos
CLIP72.4368.1470.22
CoOp76.4767.8871.92
CoCoOp75.9870.4373.10
ProDA75.4070.2372.72
Ours77.4270.4473.77
+ +(c) Caltech101 + +
BaseNewHos
CLIP96.8494.0095.40
CoOp98.0089.8193.73
CoCoOp97.9693.8195.84
ProDA98.2793.2395.68
Ours98.3194.7596.50
+ +(d) OxfordPets + +
BaseNewHos
CLIP91.1797.2694.12
CoOp93.6795.2994.47
CoCoOp95.2097.6996.43
ProDA95.4397.8396.62
Ours95.8697.5596.70
+ +(e) StanfordCars + +
BaseNewHos
CLIP63.3774.8968.65
CoOp78.1260.4068.13
CoCoOp70.4973.5972.01
ProDA74.7071.2072.91
Ours76.2974.1775.22
+ +(f) Flowers102 + +
BaseNewHos
CLIP72.0877.8074.83
CoOp97.6059.6774.06
CoCoOp94.8771.7581.71
ProDA97.7068.6880.66
Ours97.3677.7086.43
+ +(g) Food101 + +
BaseNewHos
CLIP90.1091.2290.66
CoOp88.3382.2685.19
CoCoOp90.7091.2990.99
ProDA90.3088.5789.43
Ours90.5492.3191.42
+ +(h) FGVCAircraft + +
BaseNewHos
CLIP27.1936.2931.09
CoOp40.4422.3028.75
CoCoOp33.4123.7127.74
ProDA36.9034.1335.46
Ours39.4935.3737.32
+ +(i) SUN397 + +
BaseNewHos
CLIP69.3675.3572.23
CoOp80.6065.8972.51
CoCoOp79.7476.8678.27
ProDA78.6776.9377.79
Ours82.1677.4979.76
+ +(j) DTD + +
BaseNewHos
CLIP53.2459.9056.37
CoOp79.4441.1854.24
CoCoOp77.0156.0064.85
ProDA80.6756.4866.44
Ours79.4761.5369.36
+ +(k) EuroSAT + +
BaseNewHos
CLIP56.4864.0560.03
CoOp92.1954.7468.69
CoCoOp87.4960.0471.21
ProDA83.9066.0073.88
Ours92.1473.8782.00
+ +(1)UCF101 + +
BaseNewHos
CLIP70.5377.5073.85
CoOp84.6956.0567.46
CoCoOp82.3373.4577.64
ProDA85.2371.9778.04
Ours84.1277.7480.80
+ +Table 1: Results (%) of the base-to-new generalization task on 11 benchmark datasets. We report the accuracy with CLIP ViT-B/16 model on the base classes (Base), the unseen classes (New), and the harmonic mean of both of them (Hos). + +we then have the predicted probability of the $i$ -th class: + +$$ +P (y = i \mid x) = \frac {\exp \left(\cos \left(f _ {x} , G \left(p _ {i} ^ {a}\right) / \tau\right)\right.}{\sum_ {j = 1} ^ {C} \exp \left( \right.\cos \left(f _ {x} , G \left(p _ {j} ^ {a}\right) / \tau\right)} .} \tag {7} +$$ + +We generate class-aware prompts instead of fusing identical image semantics with prompts of different classes, bringing category discrimination to the specific downstream tasks. + +TFT leverages text features to guide images to focus on task-related regions. Specifically, using the embeddings $g^{a}$ of the augmented prompts $p^{a}$ as input, we have attention: + +$$ +A ^ {x} = f _ {x} \left(g ^ {a}\right) ^ {T}, \tag {8} +$$ + +where $A^x \in \mathbb{R}^{N \times M}$ denotes text-to-image attention map. $A_{i,j}^x$ represents the similarity between the $i$ -th patch in the + +image and the $j$ -th class in the text representation. Similar to image-to-text, we use it to query the class-related part of the text correlated to the image, augmenting image features: + +$$ +f ^ {a} = \operatorname {s o f t m a x} \left(A ^ {x}\right) g ^ {a} + f _ {x}, \tag {9} +$$ + +where $f^a$ is the augmented image embeddings. We thus let image branch focus on the tasks-related representation. + +Augmented contrastive loss function is then employed to further align class-aware text embedding and text-guided image features on specific downstream tasks. The predicted probability of the $i$ -th class, which is used to calculate the contrastive loss, after mutual augmentation is: + +$$ +P (y = i \mid x) = \frac {\exp \left(\cos \left(f ^ {a} , g _ {i} ^ {a}\right) / \tau\right)}{\sum_ {j = 1} ^ {C} \exp \left(\cos \left(f ^ {a} , g _ {i} ^ {a} / \tau\right)\right)}. \tag {10} +$$ + +![](images/b81f29fe9de61578232259f33377c8c86cf683e8e61ebec5a1184b4c8911a78c.jpg) +Figure 4: Absolute improvement over CoCoOp in the base-to-new generalization task. Compared to CoCoOp, Our method achieves improvement on both base (left sub-figure) and new (right sub-figure) classes on most of the datasets. + +![](images/a0bb075f465cafcb3179aa574b98a795062f9d0c9165364270cf7f56c8b495bb.jpg) + +
ImageNetCaltech101OxfordPedsStanfordCarsFlowers102Food101FGVCAircraftSUN397DTDEuroSATUCF101Average
CLIP68.6389.3688.9965.6770.4989.2327.1265.2946.0254.1769.8366.80
CoOp71.5195.5393.3174.2595.7087.2334.1874.8268.4677.8277.2977.28
CoCoOp71.0293.4393.9371.2187.3487.3932.0372.3263.8472.7877.4074.79
Ours72.9095.9093.9679.1096.7389.9538.7279.3772.4981.0083.4580.32
+ +Table 2: Results (%) of 16-shot learning task on 11 datasets. + +Task-targeted semantic information is transferred between the two branches by minimizing the augmented contrastive loss. We merge probability before and after augmentation: + +$$ +P (y = i \mid x) = \frac {\exp \left(\left(\cos \left(f , g _ {i}\right) + \lambda \left(\cos \left(f ^ {a} , g _ {i} ^ {a}\right)\right) / \tau\right)\right.}{\sum_ {j = 1} ^ {C} \exp \left( \right.\left(\cos \left(f , g _ {i}\right) + \lambda \left(\cos \left(f ^ {a} , g _ {i} ^ {a}\right)\right) / \tau\right)} \tag {11} +$$ + +where $\lambda$ is the balance hyper-parameter, which is analyzed in our experiments. We let the two different modalities tightly coupled and mutual beneficial across the whole training process by performing the contrastive optimization. + +# 4. Experiments + +We evaluate the performance of our method on four generalization tasks, including 1) generalization from base classes to new classes; 2) few-shot classification; 3) cross-dataset transfer; 4) domain generalization. After that, we provide extensive ablation studies and in-depth analyses. + +Datasets. Following [33, 51], we use 11 image recognition datasets for the tasks of base-to-new generalization, few-shot classification and cross-dataset transfer. It contains generic image classification datasets (ImageNet [6] and Caltech101 [8]), fine-grained classification datasets (Oxford Pets [32], StanfordCars [18], Flowers102 [31], Food101 [2] and FGVCAircraft [29]), scene recognition ( SUN397 + +[44]), action recognition (UCF101 [37]), texture classification (DTD [4]), and satellite imagery recognition (EuroSAT [11]). For the domain generalization task, we use ImageNet as the source dataset and select ImageNetV2 [35], ImageNet-Sketch [41], ImageNet-A [13], and ImageNet-R [12], which are the ImageNet variants, as the target. + +Training Details. By following [50, 51], we use the best visual backbone available in CLIP, i.e., ViT-B/16, throughout the experiments. We train 10 epochs using SGD optimizer with base learning rate of 0.002 and cosine decay schedule. We set the hyper-parameter $\lambda$ in Eq. (11) to 0.2 for all experiments, and provide sensitivity analyses in Fig. 5. We run all the experiments three times with different random seeds and report the average classification accuracy. + +Baselines. We compare our method with 4 baselines. (1) Zero-shot CLIP [33] with hand-crafted prompts. (2) CoOp [51], using automatically generated prompts from few data. (3) CoCoOp [50], dynamically generating prompts conditioned on the images. (4) ProDA [28], which learns prompts from few data samples and mitigates the domain gap. + +# 4.1. Generalization From Base to New Classes + +Following the previous works, we split the classes equally into two groups for each dataset: one as base and + +
SourceTarget
ImageNetCaletech101OxfordPedsStanfordCarsFlowers102Food101FGVCAircraftSUN397DTDEuroSATUCF101
CoOp71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.55
CoCoOp71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.21
Ours72.9095.7390.2265.1469.8986.3823.3266.4946.4747.2467.43
+ +Table 3: Results of cross-dataset transfer task. Each method is trained on the source dataset and evaluated on the target. + +
SourceTarget
ImageNetImageNetV2ImageNet-SketchImageNet-AImageNet-R
CLIP66.7360.8346.1547.7773.96
CoOp71.5164.2047.9949.7175.21
CoCoOp71.0264.0748.7550.6376.18
Ours72.9064.5749.1150.9476.68
+ +Table 4: Results of domain generalization task. Each method is trained on ImageNet and evaluated on ImageNet variants. + +the other as new. The learnable modules are trained exclusively on the base classes, while evaluation is carried out separately on both the base and new classes to testify generalization ability. We report the results on 11 benchmarks in Table 1. Although compared to CoOp, CoCoOp significantly narrows generalization gap in unseen classed, but it decreases the accuracy in seen classes from $82.69\%$ to $80.47\%$ . We attribute it to the homogeneous prompts of CoCoOp, which weakens the discriminative semantics of different categories. In comparison, our method improves the accuracy in unseen classes from $80.47\%$ to $83.01\%$ by prompting each text label with corresponding image information. Benefit from the mutual learning of our CTP and TFT modules, our method further improves the accuracy in unseen classes from $71.69\%$ to $75.72\%$ , even surpasses the accuracy of CLIP hand-crafted prompts. We provide a detailed comparisons of CoCoOp and our method of per-dataset improvement in Fig. 4. Our method gains significant improvements over CoCoOp in both seen and unseen classes on 10 out of 11 recognition datasets. Surprisingly, our method significantly improves CoCoOp by more than $10\%$ in unseen classes on SUN397 and Flowers102 datasets. + +# 4.2. Few-Shot Classification + +We report few-shot classification results in Table 2. Our method surpasses baseline methods on all datasets in the few-shot setting. Especially, our method outperforms CoCoOp by $9.39\%$ , $8.65\%$ , and $8.22\%$ on Flowers102, DTD, and EuroSAT, respectively, and the average improvement over 11 datasets is $5.53\%$ . Our method also achieve $2\%$ on the challenging dataset of ImageNet. The above experiments show the great discriminative ability of our method. + +# 4.3. Cross-Dataset Transfer + +We then evaluate the generalization ability of our method on more challenging cross-dataset tasks. In this setting, we learn multi-modal prompts on ImageNet of 1000 classes. The effectiveness of the learned prompts is then tested on 10 datasets containing generic and fine-grained image classification, scene recognition, and texture classification. The results are reported in Table 3. Our method achieves the best average accuracy on the 11 datasets, especially ImageNet. It demonstrates the great transfer ability of our method. + +# 4.4. Domain Generalization + +The domain generalization setting evaluates the generalization ability of the model on the target domain that is similar to but different from the source domain [48, 47]. Zero-shot CLIP introduces no additional training parameters and exhibits great robustness to naturally distribution shifts. Other methods use few samples to train learnable parameters, there is a risk of overfitting the source distribution. Therefore, we conduct experiments using ImageNet as the source domain and evaluate the ability of generalizing to unknown on four ImageNet variants. The results are shown in Table 4. Our method achieves significant performance on the 4 ImageNet variant datasets. It verifies that our method improves the classification ability of the source domain dataset while maintaining the generalization on the target domain. + +# 4.5. Ablation Analysis + +Effectiveness of each module. To evaluate the effectiveness of Class-aware Text Prompts (CTP) and Text-guided Feature Tuning (TFT) of our method, we conduct ablation + +
MethodAverageImageNetCaltech101OxfordPeds
BaseNewHosBaseNewHosBaseNewHosBaseNewHos
A82.6963.2271.6676.4767.8871.9298.0089.8193.7393.6795.2994.47
B82.3872.4476.6476.9669.6273.1198.2394.2496.1995.6197.9796.78
C82.9372.9877.2477.2169.8673.3598.4492.7295.4995.4997.8196.64
Ours83.0175.7279.0277.4270.4473.7798.3194.7596.5095.8697.5596.70
MethodStanfordCarsFlowers102Food101FGVCAircraft
BaseNewHosBaseNewHosBaseNewHosBaseNewHos
A78.1260.4068.1397.6059.6774.0688.3382.2685.1940.4422.3028.75
B74.6273.6874.1596.7266.4278.7690.3091.4790.8836.4134.3935.37
C75.8474.5375.1897.3274.8684.6390.5691.6591.1037.8233.1735.34
Ours76.2974.1775.2297.3677.7086.4390.5492.3191.4239.4935.3737.32
MethodSUN397DTDEuroSATUCF101
BaseNewHosBaseNewHosBaseNewHosBaseNewHos
A80.6065.8972.5179.4441.1854.2492.1954.7468.6984.6956.0567.46
B81.7376.8979.2480.1851.7962.9391.7067.6277.8483.7272.7277.83
C82.2976.2479.1581.7154.7465.5690.1060.5272.4185.4076.6880.81
Ours82.1677.4979.7679.4761.5369.3692.1473.8782.0084.1277.7480.80
+ +Table 5: Ablation studies of our method on 11 datasets. Three ablation cases are considered: A: Ours w/o TVP w/o FTP. B: Ours w/o TVP. C: Ours w/o FTP. TVP is the text-reorganized vision prompt, and FTP is the fine-grained text prompt. + +
Prompt LearningFeature TuningAccuracy (%)
MLP-PLCTPMLP-FTTFT
✓71.66 (CoOp)
75.83 (+4.17)
✓76.64 (+4.98)
✓75.94 (+4.28)
✓77.24 (+5.58)
✓✓77.05 (+5.39)
✓✓79.02 (+7.36)
+ +Table 6: Comparison of different structures for prompt learning and feature tuning. The average results of harmonic mean of from-base-to-new generalization task on 11 datasets are reported. In compared to our attention design in CTP and TFT modules, MLP-PL and MLP-FT are designed using the Linear-ReLU-Linear block setting of [50]. Improvements over the baseline of CoOp, are marked in green. + +experiments on 11 datasets, as reported in Table 5. In most cases, each module significantly improves the performance of the model. For average results, CTP and TFT improves the results by $5.58\%$ and $4.98\%$ , respectively, and the combination of them improves the results by $7.36\%$ . It shows the effectiveness of the two branches of text-to-image and image-to-text, the mutual learning of the two modules further improves the performance on downstream tasks. + +Comparison of different structure design of multi-modal mutual learning. To further provide in-depth analysis about our mutual learning, we further explore two vanilla structures: (1) MLP-PL: The image features are forwarded to a block of Linear-ReLU-Linear, borrowed from [50], and + +![](images/2ff833bdc7e9960031070fad5962b722dfa853110b5d03a62bf13b6b40ba2d08.jpg) +Figure 5: Sensitivity analysis of $\lambda$ , with base, new, and host metrics, on UCF101 (left) and Caltech101 (right) datasets. + +![](images/b464df958c9d49e0af0522b8b1ee349247eb7f7e1f893f457bc10de6e5195241.jpg) + +then added to the text for augmenting it (the same to us). (2) MLP-FT: The text prompts are forwarded to the Linear-ReLU-Linear block, and then added to the image for augmenting it (the same to us). In comparison, our class-aware text prompts (CTP) module and text-guided feature tuning (TFT) module, adopt text-image attention to learn the augmented features instead of the Linear-ReLU-Linear block. We report the results of the different designs in Table 6. First, we find that combining MLP-PL & MLP-FT and CTP & TFT can both improve the results compared with using either of them. It indicates that both prompt learning and feature tuning are important to achieve better results. Second, compared with the design of Linear-ReLU-Linear block, our design of text-image attention further improves performance by $0.81\%$ and $1.3\%$ for prompt learning and feature tuning, respectively. It demonstrates the effectiveness of our design of attention, which helps the model to focus on class-aware and task-related semantics. Third, com + +pared with CoOp, both of the designs could improve the final results by large margins. The key factor of our mutual learning to achieve significant performance is the task-related alignment of vision and language in latent space. + +Sensitivity Analysis of $\lambda$ . We evaluate the parameter sensitivity of $\lambda$ of Eq. (11) in Fig. 5. The results suggest that the performance of our method is generally robust to $\lambda$ , indicating a wide range of $\lambda$ works well in downstream tasks. + +# 5. Conclusion + +In this paper, we introduce task-oriented multi-modal mutual learning for adapting large vision-language models to downstream vision tasks. We propose class-aware text prompt and text-guided feature tuning to unleash the potential of the vision-language model by re-activating its task-related representation abilities. Our method yields impressive generalization performance on a wide range of vision tasks and datasets. We hope the presented findings and insights in this paper could benefit the following works in designing more efficient and effective adaptation methods. For the future work, we think it is interesting to extend the adaptation of vision language models to more vision tasks, such as semantic segmentation, object detection, etc. + +Limitations. Similar to CoCoOp, we learn image-conditioned representations, thus the batch-size of training need to be set to 1, which is not efficient enough for learning. We aim to solve this efficiency issue in the future work. + +# 6. Acknowledgements. + +This work is supported by the Innovation Capacity Construction Project of Jilin Province Development and Reform Commission (2021FGWCXNLJSSZ10) and the Fundamental Research Funds for the Central Universities, JLU. Zichang Tan is the project lead. + +# References + +[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting largescale models. arXiv preprint arXiv:2203.17274, 1(3):4, 2022.3 +[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 446-461. Springer, 2014. 6 +[3] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX, pages 104-120. Springer, 2020. 3 +[4] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the + +wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 6 +[5] Joe Davison, Joshua Feldman, and Alexander M Rush. Commonsense knowledge mining from pretrained models. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pages 1173-1178, 2019. 3 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6 +[7] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3 +[8] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 6 +[9] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pretrained language models better few-shot learners. arXiv preprint arXiv:2012.15723, 2020. 3 +[10] Adi Haviv, Jonathan Berant, and Amir Globerson. Bertese: Learning to speak to bert. arXiv preprint arXiv:2103.05327, 2021. 3 +[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6 +[12] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 6 +[13] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 6 +[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 1 +[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1 +[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Computer Vision-ECCV 2022: 17th + +European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 709-727. Springer, 2022. 3 +[17] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438, 2020. 3 +[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6 +[19] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3 +[20] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3 +[21] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. Unimo: Towards unified-modal understanding and generation via cross-modal contrastive learning. arXiv preprint arXiv:2012.15409, 2020.3 +[22] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 121-137. Springer, 2020. 3 +[23] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3 +[24] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. arXiv preprint arXiv:2212.00794, 2022. 3 +[25] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35, 2023. 1 +[26] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands, too. arXiv preprint arXiv:2103.10385, 2021. 3 +[27] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3 +[28] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 1, 2, 6 +[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6 + +[30] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI, pages 529-544. Springer, 2022. 3 +[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 6 +[32] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 6 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 6 +[34] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 3 +[35] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International conference on machine learning, pages 5389-5400. PMLR, 2019. 6 +[36] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3 +[37] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild. Center for Research in Computer Vision, 2(11), 2012. 6 +[38] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019. 3 +[39] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3 +[40] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing nlp. arXiv preprint arXiv:1908.07125, 2019. 3 +[41] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems, 32, 2019. 6 +[42] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 3 +[43] Junyang Wu, Xianhang Li, Chen Wei, Huiyu Wang, Alan Yuille, Yuyin Zhou, and Cihang Xie. Unleashing the power + +of visual prompting at the pixel level. arXiv preprint arXiv:2212.10556, 2022.3 +[44] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 6 +[45] Yifan Yang, Weiquan Huang, Yixuan Wei, Houwen Peng, Xinyang Jiang, Huiqiang Jiang, Fangyun Wei, Yin Wang, Han Hu, Lili Qiu, et al. Attentive mask clip. arXiv preprint arXiv:2212.08653, 2022. 3 +[46] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, 2021. 3 +[47] Junkun Yuan, Xu Ma, Defang Chen, Kun Kuang, Fei Wu, and Lanfen Lin. Label-efficient domain generalization via collaborative exploration and generalization. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2361-2370, 2022. 7 +[48] Junkun Yuan, Xu Ma, Defang Chen, Kun Kuang, Fei Wu, and Lanfen Lin. Domain-specific bias filtering for single labeled domain generalization. International Journal of Computer Vision, 131(2):552-571, 2023. 7 +[49] Weizhe Yuan, Graham Neubig, and Pengfei Liu. Bartscore: Evaluating generated text as text generation. Advances in Neural Information Processing Systems, 34:27263-27277, 2021. 3 +[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 2, 3, 6, 8 +[51] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 2, 3, 6 \ No newline at end of file diff --git a/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/images.zip b/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f3d23e1b4b2404392b3c6e963cbebdcfe9763fd9 --- /dev/null +++ b/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a5979f3cc6e984a1dfbde55de1df652e6822d0010f025e7f098d98b2c496abc +size 785202 diff --git a/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/layout.json b/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7e883ac9879691e8195f9d35a0903b032157def6 --- /dev/null +++ b/2023/Task-Oriented Multi-Modal Mutual Leaning for Vision-Language Models/layout.json @@ -0,0 +1,9490 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 527, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 527, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 527, + 121 + ], + "type": "text", + "content": "Task-Oriented Multi-Modal Mutual Learning for Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "spans": [ + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "text", + "content": "Sifan Long " + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,3\\dagger*}" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "text", + "content": " Zhen Zhao" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "inline_equation", + "content": "^{3,4\\dagger*}" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "text", + "content": " Junkun Yuan" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "inline_equation", + "content": "^{3,5\\dagger*}" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "text", + "content": " Zichang Tan" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "text", + "content": " Jiangjiang Liu" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "text", + "content": " Luping Zhou" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "text", + "content": " Shengsheng Wang" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "inline_equation", + "content": "^{1,2‡}" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "text", + "content": " Jingdong Wang" + }, + { + "bbox": [ + 77, + 141, + 515, + 171 + ], + "type": "inline_equation", + "content": "^{3‡}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 171, + 534, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 534, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 534, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 171, + 534, + 213 + ], + "type": "text", + "content": "College of Computer Science and Technology, Jilin University, Jilin, China \n" + }, + { + "bbox": [ + 61, + 171, + 534, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 171, + 534, + 213 + ], + "type": "text", + "content": "Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Jilin, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 213, + 526, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 213, + 526, + 257 + ], + "spans": [ + { + "bbox": [ + 67, + 213, + 526, + 257 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 213, + 526, + 257 + ], + "type": "text", + "content": "Baidu VIS " + }, + { + "bbox": [ + 67, + 213, + 526, + 257 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 213, + 526, + 257 + ], + "type": "text", + "content": "University of Sydney " + }, + { + "bbox": [ + 67, + 213, + 526, + 257 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 67, + 213, + 526, + 257 + ], + "type": "text", + "content": "Zhejiang University longsf22@mails.jlu.edu.cn {zhen.zhao, luping.zhou}@sydney.edu.au yuanjk@zju.edu.cn wss@jlu.edu.cn {tanzichang, liujiangjiang, wangjingdong}@baidu.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "spans": [ + { + "bbox": [ + 143, + 282, + 192, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 308, + 290, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 290, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 290, + 632 + ], + "type": "text", + "content": "Prompt learning has become one of the most efficient paradigms for adapting large pre-trained vision-language models to downstream tasks. Current state-of-the-art methods, like CoOp and ProDA, tend to adopt soft prompts to learn an appropriate prompt for each specific task. Recent CoCoOp further boosts the base-to-new generalization performance via an image-conditional prompt. However, it directly fuses identical image semantics to prompts of different labels and significantly weakens the discrimination among different classes as shown in our experiments. Motivated by this observation, we first propose a class-aware text prompt (CTP) to enrich generated prompts with label-related image information. Unlike CoCoOp, CTP can effectively involve image semantics and avoid introducing extra ambiguities into different prompts. On the other hand, instead of reserving the complete image representations, we propose text-guided feature tuning (TFT) to make the image branch attend to class-related representation. A contrastive loss is employed to align such augmented text and image representations on downstream tasks. In this way, the image-to-text CTP and text-to-image TFT can be mutually promoted to enhance the adaptation of VLMs for downstream tasks. Extensive experiments demonstrate that our method outperforms the existing methods by a significant margin. Especially, compared to CoCoOp, we achieve an average improvement of " + }, + { + "bbox": [ + 46, + 308, + 290, + 632 + ], + "type": "inline_equation", + "content": "4.03\\%" + }, + { + "bbox": [ + 46, + 308, + 290, + 632 + ], + "type": "text", + "content": " on new classes and " + }, + { + "bbox": [ + 46, + 308, + 290, + 632 + ], + "type": "inline_equation", + "content": "3.19\\%" + }, + { + "bbox": [ + 46, + 308, + 290, + 632 + ], + "type": "text", + "content": " on harmonic-mean over eleven classification benchmarks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 128, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 672, + 287, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 672, + 287, + 697 + ], + "spans": [ + { + "bbox": [ + 46, + 672, + 287, + 697 + ], + "type": "text", + "content": "Recently, large vision-language models (VLM), such as CLIP [33] and ALIGN [15], which employ language as" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 281, + 555, + 466 + ], + "blocks": [ + { + "bbox": [ + 306, + 281, + 555, + 466 + ], + "lines": [ + { + "bbox": [ + 306, + 281, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 555, + 466 + ], + "type": "image", + "image_path": "e60548bb39680cb70e3e638e536d9fedbf0ab699aae57b4f901a623b562a8cb5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 521 + ], + "type": "text", + "content": "Figure 1: Comparisons between CoCoOp [50] and our method. The cosine distance between the positive and the negative prompts, which quantifies the class discrimination, and the average accuracy on benchmarks are reported." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "supervision signal instead of discrete labels, have shown impressive generalization performance in a wide range of downstream vision tasks. Their multi-modal interaction nature delivers open-vocabulary support and achieves amazing zero-shot classification performance. Despite their impressive transferable abilities, as discussed in [25], it is essential to re-activate specific representation capabilities for optimal performance in certain downstream tasks. Considering their hundreds of millions or billions of parameters, attempting to fine-tune the entire model is impractical and even jeopardizes the well-established representation space [14]. To this end, many recent studies have centred on the efficient and effective adaptation of pre-trained and frozen large VLMs for the specific downstream tasks [28, 51, 50]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 285, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 285, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 285, + 714 + ], + "type": "text", + "content": "*Equal contribution. † Interns at Baidu VIS. ‡ Corresponding authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21959" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "type": "text", + "content": "Prompt, a simple, compact, and viable strategy, has become the leading solution for deploying large pre-trained VLMs into certain downstream tasks. CLIP [33] utilizes hand-crafted prompts to achieve impressive zero- and few-shot classification performance. Nevertheless, manually-designed prompts require significant domain knowledge and can be highly time-consuming and sub-optimal for specific downstream tasks. To address this problem, later studies [28, 51] adopt soft prompts to learn an appropriate text prompt via optimizing a contrastive loss on different text labels. CoCoOp [50] further highlights the limitations of such static soft prompts and proposes learning image-dependent prompts conditioned on individual instances rather than fixed prompts. It achieves great performance gains on unseen classes by adding high-level image embedding to text prompts. However, compared to CoOp with static prompts, CoCoOp essentially fuses identical image semantics with different text labels, leading to inevitable learning ambiguity and resulting in an average performance drop of " + }, + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "type": "inline_equation", + "content": "2.22\\%" + }, + { + "bbox": [ + 47, + 72, + 289, + 420 + ], + "type": "text", + "content": " on base classes on 11 datasets (see Table 1). For example, it may associate the dog image semantics with a prompt that references the [class] of a cat. When using the cosine distance to measure the differences between the positive and negative text prompts, as shown in Fig. 1, CoCoOp holds low distance values, suggesting that it brings significant learning ambiguities to text prompts. Therefore, we argue that text prompts should not only condition on distinct input images for better generalization abilities, but also adapt to different classes to eliminate the potential ambiguities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 422, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 590 + ], + "type": "text", + "content": "To achieve this goal, we propose Class-aware Text Prompts (CTP), which leverages label-related image information to generate finer prompts. Specifically, we first contact learnable context vectors and each class label to model the initial prompt sentences. Then we leverage these class prompt sentences to query their corresponding image regions and representations. Corresponding related image features are subsequently added to initial class prompt sentences to produce the final text prompts. In this way, generated image-dependent and class-aware prompts can better concentrate on the image information in a more precise manner. As shown in Fig. 1, our method enjoys better discrimination between positive and negative prompts and consistently outperforms CoCoOp on 11 classification datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "On the other hand, we identify a critical problem in these text prompt-based strategies: the image branch is ignored and not adjusted to specific downstream tasks. As shown in Fig. 2 (CoCoOp), on the task of identifying birds, the output image feature, without further tuning, can be distracted to leaves of the same color. Similarly, it also wrongly highlights the beer foam that is of a similar shape to recognize golf balls. Since the final recognition is jointly inferred by both text and image branches, such an issue may degrade the classification performance. Thus it is necessary" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 383, + 149 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 383, + 149 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 383, + 149 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 383, + 149 + ], + "type": "image", + "image_path": "f7f5394eb598b2671856cc56aa696bef15416101573a59ba6cb24821420bc5ef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 388, + 72, + 455, + 148 + ], + "blocks": [ + { + "bbox": [ + 388, + 72, + 455, + 148 + ], + "lines": [ + { + "bbox": [ + 388, + 72, + 455, + 148 + ], + "spans": [ + { + "bbox": [ + 388, + 72, + 455, + 148 + ], + "type": "image", + "image_path": "e38af2466fc343501b5baa10c409334eb80db5486b5af21f7b9fbb4c688f9545.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 458, + 72, + 539, + 149 + ], + "blocks": [ + { + "bbox": [ + 458, + 72, + 539, + 149 + ], + "lines": [ + { + "bbox": [ + 458, + 72, + 539, + 149 + ], + "spans": [ + { + "bbox": [ + 458, + 72, + 539, + 149 + ], + "type": "image", + "image_path": "e9a7cf8de1641bb4206e1fbffe2fefc27df51a4b2f8ee6eee649980066ad3505.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 314, + 152, + 383, + 238 + ], + "blocks": [ + { + "bbox": [ + 314, + 152, + 383, + 238 + ], + "lines": [ + { + "bbox": [ + 314, + 152, + 383, + 238 + ], + "spans": [ + { + "bbox": [ + 314, + 152, + 383, + 238 + ], + "type": "image", + "image_path": "3b30c3b4c2277835013d0ac2ba1f5a9e39a0ee9fe738bcc8bb821158d57fff41.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 247, + 547, + 296 + ], + "lines": [ + { + "bbox": [ + 305, + 247, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 547, + 296 + ], + "type": "text", + "content": "Figure 2: Comparisons of attention map visualization for CoCoOp and our method on ImageNet. Our method obtains better average accuracy of both base and new classes across 11 datasets by paying attention to task-related regions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 388, + 153, + 455, + 238 + ], + "blocks": [ + { + "bbox": [ + 388, + 153, + 455, + 238 + ], + "lines": [ + { + "bbox": [ + 388, + 153, + 455, + 238 + ], + "spans": [ + { + "bbox": [ + 388, + 153, + 455, + 238 + ], + "type": "image", + "image_path": "0cda44d5167afcc175abc3c4f628891a8a03fc9fab3ad1f01e4d862cbbee4cf4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 458, + 153, + 539, + 238 + ], + "blocks": [ + { + "bbox": [ + 458, + 153, + 539, + 238 + ], + "lines": [ + { + "bbox": [ + 458, + 153, + 539, + 238 + ], + "spans": [ + { + "bbox": [ + 458, + 153, + 539, + 238 + ], + "type": "image", + "image_path": "ebc7cca7075bb6fa3ef540e9e2e4d35a534f320a27c23f184c98be0d7fc0a7b5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 316, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 424 + ], + "type": "text", + "content": "to tune the image features further so that the image branch can focus more on the tasks-related representation. We then propose Text-guided Feature Tuning (TFT), which leverages encoded text embedding to guide image representation more on task-related regions. As shown in Fig. 2 (ours), our method successfully focuses on task-related regions, i.e., birds and golf balls. We then leverage the contrastive loss function to further align class-aware text embedding and text-guided image features on certain downstream tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 425, + 546, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 546, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 546, + 568 + ], + "type": "text", + "content": "In summary, we propose a new task-oriented multimodal mutual-learning method, which well-integrates our designed class-aware text prompts and text-guided feature tuning for fast adaptation of frozen VLMs on downstream tasks. Image features can help construct image-dependent class-aware text prompts, leading to more discriminative text embedding. Simultaneously, improved text embedding can further guide the image branch attending to class-related representation. In this way, these two different modality branches can be tightly coupled and mutual-beneficial across the whole training process. Our main contributions are summarized in the following." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 577, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 317, + 577, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 577, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 317, + 577, + 545, + 636 + ], + "type": "text", + "content": "- We propose class-aware text prompts which generate prompts based on task-relevant image semantics instead of complete visual information. In this way, we improve the classification accuracy of unseen classes without introducing extra learning ambiguities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 645, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 645, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 317, + 645, + 545, + 693 + ], + "type": "text", + "content": "- We propose text-guided feature tuning which enforces image branch to pay more attention to the task-related representation. As a result, the model avoids deviating attention to the task-irrelevant regions of the image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 702, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 545, + 714 + ], + "type": "text", + "content": "- Benefiting from our mutual learning strategy, our" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21960" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 109 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 109 + ], + "type": "text", + "content": "method achieves SOTA results on four downstream tasks. Especially, ours significantly outperforms existing methods on the base-to-new generalization task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 129, + 132, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 132, + 141 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 132, + 141 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 148, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 148, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 287, + 353 + ], + "type": "text", + "content": "Vision language models (VLM). The current VLM can be roughly divided into four categories based on the training objectives: image-text matching [3, 20, 27], contrastive loss [19, 21, 22], masked language modeling [38, 39, 46], and masked image modeling [3, 27, 38]. As a milestone, CLIP utilizes 400 million image-text pairs to train a large-scale multi-modal model and demonstrates promising performance on a wide spectrum of tasks including few-shot and zero-shot visual recognition. Motivated by this work, numerous follow-ups have been proposed to improve the effectiveness (e.g., FLIP [24], A-CLIP [45], MaskCLIP [7], and SLIP [30]) or apply it to other domains (e.g., Dense-CLIP [34] and ActionCLIP [42]). The primary limitation of these methods is that hand-crafted prompts are dataset-sensitive and difficult to optimize. We design an automatic and learnable prompts method to enhance the generalization performance of pre-trained models on downstream tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 360, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 287, + 684 + ], + "type": "text", + "content": "Prompt learning in NLP. As the scale and complexity of pre-trained language models continue to grow, fine-tuning for specific tasks is becoming increasingly expensive. In contrast, prompt-based approaches are an efficient and lightweight alternative that can be used to generate high-quality text with much lower computational requirements. The original prompts were manually designed prompt templates. While manually designing prompts is advantageous due to their intuitive and comprehensible nature, it also presents a significant challenge that demands extensive experimentation, experience, and language expertise, resulting in high costs. To overcome the limitations of manual prompt design, numerous studies have initiated research into automatically learn appropriate prompts. The automatic prompts can be categorized into two types: discrete prompts and continuous prompts. Discrete prompts consist of various approaches such as prompt mining [17], prompt paraphrasing [49, 10], gradient-based search [40], prompt generation [9] and prompt scoring [5]. On the other hand, continuous prompts include techniques such as prefix tuning [23], tuning initialized with discrete prompts [36] and hard-soft prompt hybrid tuning [26]. These methods have also been applied to the field of computer vision for prompt learning research. However, the task of prompt learning in computer vision is often considered more challenging than in natural language due to the relatively limited high-level semantic information present in visual data with raw pixels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Prompt learning in vision language models. Prompt learning has been demonstrated to be an effective method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": "for improving the performance of pre-trained language models on downstream tasks. Recently, prompt learning has gained increased attention in the context of vision language models. For example, CoOp [51] employs learnable vectors to model contextual words as prompts, and demonstrates that automatic prompts outperform hand-crafted prompts in downstream tasks. CoCoOp [50] extends CoOp by incorporating lightweight neural networks to dynamically generate prompts based on each image, thus mitigating sensitivity to class shifts. Different from the above methods, VP [1], VPT [43], and EVP [16] prompt with images. VP [1] directly combines learnable prompts and pixel-wise input images as new inputs to the model. EVP [43] shrinks the original image before padding the prompts around it, to avoid destroying the original image information. VPT [16] introduces a small amount learnable parameters into the input sequence of each transformer layer and learns them together with a linear head during fine-tuning. Building on the prompt learning approach of the text branch, we propose class-aware text prompt that generates image-dependent and class-aware prompts. Similarly, follow the feature tuning of image branch, we introduce text-guided tuning, which directs the image branch to focus on the task-relevant local regions rather than the global information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 373, + 361, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 361, + 386 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 361, + 386 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 395, + 531, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 531, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 531, + 407 + ], + "type": "text", + "content": "3.1. Comparisons of CLIP, CoOp, and CoCoOp" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": "CLIP comprises two encoders: an image encoder and a text encoder. The image encoder, denotes by " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "F(x)" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": ", converts an image " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " with height of " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " and width of " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " into a " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": "-dimensional image feature " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "f_x \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " is the number of split patches. Meanwhile, the text encoder, denoted as " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "G(t)" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": ", generates an " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": "-dimensional text representation " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "g_t \\in \\mathbb{R}^{M \\times d}" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " from natural language text " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " is the number of classes. Two encoders are jointly trained using a contrastive loss function that maximizes the cosine similarity of matched pairs and minimizes that of the unmatched pairs. After training, CLIP can be directly used for zero-shot image recognition without requiring fine-tuning of the whole model. Since CLIP is pre-trained on whether an image matches a textual description, the hand-crafted prompt template is employed to convert raw labels into textual descriptions. The most common form of template in CLIP is \"a photo of a [CLASS]\", where the class token is replaced with specific class names such as \"cat\", \"dog\", \"car\", etc. We let the image features " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "f_x" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " of an image " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " be extracted by an image encoder and the text features " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " be obtained by feeding the prompt description into the text encoder. The prediction task is defined as the classification of an image into one of " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " categories, which are represented by the set " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "y \\in \\{1, \\dots, C\\}" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": ". Denote " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " as the predicted category. Let " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "g_t^i" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": " be the " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": "-th dimension of text features " + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 304, + 415, + 545, + 715 + ], + "type": "text", + "content": ", with image features" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21961" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 262, + 223 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 262, + 223 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 262, + 223 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 262, + 223 + ], + "type": "image", + "image_path": "7e2bafb580134e9fcc143ee5ebe227bc6853de91852f569c3261584a883539ff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 231, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 547, + 316 + ], + "type": "text", + "content": "Figure 3: Comparisons of three representative prompt learning techniques and our method. The main differences lie in how the text and image branches focus on downstream tasks. CLIP artificially designs prompt templates. CoOp designs automatic prompts using learnable parameters. CoCoOp directly allows text branch to focus on images semantic through Meta-Net. We introduce Class-aware Text Prompts (CTP) and Text Feature Tuning (TFT) to the text and image branches, respectively. The CTP generates prompts based on class-related image information instead of using the identical image semantics like CoCoOp. The TFT enables the image branch to directly focus on downstream tasks. We leverage the contrastive loss function to align task-oriented text and images, making them promote each other for achieving better downstream generalization performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 265, + 70, + 545, + 222 + ], + "blocks": [ + { + "bbox": [ + 265, + 70, + 545, + 222 + ], + "lines": [ + { + "bbox": [ + 265, + 70, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 265, + 70, + 545, + 222 + ], + "type": "image", + "image_path": "a57d25f8bf6d3de5b7ea94a9651d697290ac0c4115e83f379d47e10c4f72f977.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 337, + 266, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 337, + 266, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 266, + 349 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 47, + 337, + 266, + 349 + ], + "type": "text", + "content": ", we have the predicted probability of the " + }, + { + "bbox": [ + 47, + 337, + 266, + 349 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 337, + 266, + 349 + ], + "type": "text", + "content": "-th class:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 356, + 287, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 356, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 71, + 356, + 287, + 392 + ], + "type": "interline_equation", + "content": "P (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f _ {x} , g _ {t} ^ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\cos \\left(f _ {x} , g _ {t} ^ {j}\\right) / \\tau\\right)}, \\tag {1}", + "image_path": "30ffea2e259c1ac38fa2a7c4565f094b0c1515dfdbd4e297d19b891980a718e1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 399, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 399, + 287, + 422 + ], + "type": "text", + "content": " denotes the cosine similarity and " + }, + { + "bbox": [ + 46, + 399, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 399, + 287, + 422 + ], + "type": "text", + "content": " is the temperature parameter of the softmax function." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "text", + "content": "CoOp replaces the hand-crafted prompts with automatically generated prompts. Specifically, CoOp introduces " + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "text", + "content": " learnable context vectors " + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "inline_equation", + "content": "\\{v_{1},\\ldots ,v_{k}\\}" + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "text", + "content": " to model the context words of the prompts. We define " + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "text", + "content": " as the word embedding of the " + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "text", + "content": "-th class name. Then, the prompt of " + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "text", + "content": "-th class is denoted as " + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "inline_equation", + "content": "p_i = \\{v_1,\\dots,v_k,c_i\\}" + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "text", + "content": ". Therefore, we have the predicted probability of the " + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 427, + 287, + 513 + ], + "type": "text", + "content": "-th class using CoOp method:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 518, + 287, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 518, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 61, + 518, + 287, + 548 + ], + "type": "interline_equation", + "content": "P (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f _ {x} , G \\left(p _ {i}\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\cos \\left(f _ {x} , G \\left(p _ {j}\\right)\\right) / \\tau\\right)}, \\tag {2}", + "image_path": "76bb40ab4a247d29b1c84794b4bdf0213e12a88fdc13809f3949b0ad90300983.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 555, + 275, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 275, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 275, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 555, + 275, + 568 + ], + "type": "inline_equation", + "content": "G(p_{i})" + }, + { + "bbox": [ + 47, + 555, + 275, + 568 + ], + "type": "text", + "content": " is the text embedding from text encoder " + }, + { + "bbox": [ + 47, + 555, + 275, + 568 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 555, + 275, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "text", + "content": "CoCoOp extends CoOp by generating image-conditional prompts. Specifically, CoCoOp uses Meta-Net to generate the residual vector " + }, + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "text", + "content": " based on each image. Each context token is now obtained by " + }, + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "inline_equation", + "content": "v_{k}(x) = v_{k} + \\pi" + }, + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "text", + "content": ". The prompt of the " + }, + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "text", + "content": "-th class " + }, + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "inline_equation", + "content": "p_i(x) = \\{v_1(x),\\dots ,v_k(x),c_i\\}" + }, + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "text", + "content": ". As a result, the prediction probability of the " + }, + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 573, + 287, + 645 + ], + "type": "text", + "content": "-th class is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 651, + 287, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 287, + 682 + ], + "type": "interline_equation", + "content": "P (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f _ {x} , G \\left(p _ {i} (x)\\right)\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left( \\right.\\cos \\left(f _ {x} , G \\left(p _ {j} (x)\\right) / \\tau\\right)}, \\tag {3}", + "image_path": "af2bd741a07e37b706e2e872c6982e1bded74f6aead604d66d746a3ffa177451.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "G(p_i(x))" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " is the text embedding conditional on the image " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " from the text encoder " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 336, + 541, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 336, + 541, + 349 + ], + "spans": [ + { + "bbox": [ + 306, + 336, + 541, + 349 + ], + "type": "text", + "content": "3.2. Our Task-Oriented Mutual Learning Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 354, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 462 + ], + "type": "text", + "content": "Our method consists of two modules, i.e., Class-aware Text Prompts (CTP) and Text-guided Feature Tuning (TFT), as shown in Fig. 3. Compared to CoCoOp, we use CTP to generate class-aware prompts based on task-relevant local image regions instead of the global information. Besides, we use CTP to make the image branch directly pay attention to the task-related image region. We let the two modules be tightly coupled and mutual-beneficial across the training process by optimizing the contrastive loss function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 466, + 545, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 527 + ], + "type": "text", + "content": "CTP learns image conditioned discriminative prompts for finer paying attention to semantic-related regions of the images. Specifically, in order to obtain the text semantic-related regions of the image, we leverage the prompt " + }, + { + "bbox": [ + 304, + 466, + 545, + 527 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 466, + 545, + 527 + ], + "type": "text", + "content": " and the image feature " + }, + { + "bbox": [ + 304, + 466, + 545, + 527 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 304, + 466, + 545, + 527 + ], + "type": "text", + "content": " to calculate the attention matrix " + }, + { + "bbox": [ + 304, + 466, + 545, + 527 + ], + "type": "inline_equation", + "content": "A^{t}" + }, + { + "bbox": [ + 304, + 466, + 545, + 527 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 402, + 532, + 545, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 532, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 402, + 532, + 545, + 547 + ], + "type": "interline_equation", + "content": "A ^ {t} = p f _ {x} ^ {T}, \\tag {4}", + "image_path": "75c7a0bdc325c4b912279fe00e5aadc0460304627facd08b9cadadb689a27e2e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "inline_equation", + "content": "A^t \\in \\mathbb{R}^{M \\times N}" + }, + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "text", + "content": " is the image-to-text attention map. " + }, + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "inline_equation", + "content": "A_{i,j}^t" + }, + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "text", + "content": " represents the similarity between the " + }, + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "text", + "content": "-th class in the text prompts and the " + }, + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "text", + "content": "-th patch in the image. In this way, we can query the regions of the images that semantically related to the class information by the attention matrix " + }, + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "inline_equation", + "content": "A^t" + }, + { + "bbox": [ + 304, + 552, + 545, + 613 + ], + "type": "text", + "content": ". That is," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 377, + 619, + 545, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 619, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 377, + 619, + 545, + 633 + ], + "type": "interline_equation", + "content": "f _ {x} ^ {t} = \\operatorname {s o f t m a x} \\left(A ^ {t}\\right) f _ {x}, \\tag {5}", + "image_path": "51a27f9fa0d64f028b5a1d4b64ff8e28c7da4f2dda54246cc88fec3db0e84a1e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 639, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 663 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 639, + 545, + 663 + ], + "type": "inline_equation", + "content": "f_{x}^{t}" + }, + { + "bbox": [ + 304, + 639, + 545, + 663 + ], + "type": "text", + "content": " is the regions correlated to the text of a specific class. We use it to obtain augmented class-aware prompts:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 397, + 669, + 545, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 669, + 545, + 683 + ], + "spans": [ + { + "bbox": [ + 397, + 669, + 545, + 683 + ], + "type": "interline_equation", + "content": "p ^ {a} = p + f _ {x} ^ {t}, \\tag {6}", + "image_path": "49ebf25e55aa4ae216b51418419f30e311f73f7f8998a0d83dccfefcd7459101.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "p^a" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " is the text prompts enhanced by semanticallyrelevant image regions. Let " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "p_i^a" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " be the " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "-th dimension of " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "p^a" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21962" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 89, + 203, + 167 + ], + "blocks": [ + { + "bbox": [ + 89, + 77, + 198, + 87 + ], + "lines": [ + { + "bbox": [ + 89, + 77, + 198, + 87 + ], + "spans": [ + { + "bbox": [ + 89, + 77, + 198, + 87 + ], + "type": "text", + "content": "(a) Average over 11 datasets" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 89, + 203, + 167 + ], + "lines": [ + { + "bbox": [ + 85, + 89, + 203, + 167 + ], + "spans": [ + { + "bbox": [ + 85, + 89, + 203, + 167 + ], + "type": "table", + "html": "
BaseNewHos
CLIP69.3474.2271.70
CoOp82.6963.2271.66
CoCoOp80.4771.6975.83
ProDA81.5672.3076.65
Ours83.0175.7279.02
", + "image_path": "c159916aeb2ec30ca1772bfd0a6cc852700f4e34e2702ae67d0efc8ed7715c87.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 235, + 89, + 354, + 167 + ], + "blocks": [ + { + "bbox": [ + 269, + 77, + 320, + 88 + ], + "lines": [ + { + "bbox": [ + 269, + 77, + 320, + 88 + ], + "spans": [ + { + "bbox": [ + 269, + 77, + 320, + 88 + ], + "type": "text", + "content": "(b) ImageNet" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 235, + 89, + 354, + 167 + ], + "lines": [ + { + "bbox": [ + 235, + 89, + 354, + 167 + ], + "spans": [ + { + "bbox": [ + 235, + 89, + 354, + 167 + ], + "type": "table", + "html": "
BaseNewHos
CLIP72.4368.1470.22
CoOp76.4767.8871.92
CoCoOp75.9870.4373.10
ProDA75.4070.2372.72
Ours77.4270.4473.77
", + "image_path": "073489641f536df8bef392e777e2951c8ba965ec21c41b8e4c0283d11dcba289.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 391, + 89, + 509, + 167 + ], + "blocks": [ + { + "bbox": [ + 422, + 77, + 477, + 87 + ], + "lines": [ + { + "bbox": [ + 422, + 77, + 477, + 87 + ], + "spans": [ + { + "bbox": [ + 422, + 77, + 477, + 87 + ], + "type": "text", + "content": "(c) Caltech101" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 391, + 89, + 509, + 167 + ], + "lines": [ + { + "bbox": [ + 391, + 89, + 509, + 167 + ], + "spans": [ + { + "bbox": [ + 391, + 89, + 509, + 167 + ], + "type": "table", + "html": "
BaseNewHos
CLIP96.8494.0095.40
CoOp98.0089.8193.73
CoCoOp97.9693.8195.84
ProDA98.2793.2395.68
Ours98.3194.7596.50
", + "image_path": "78509d0c261120b35eceeb7ef262ec08deb7c3008930f3ce05ae987cd626c553.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 85, + 194, + 203, + 271 + ], + "blocks": [ + { + "bbox": [ + 116, + 182, + 171, + 192 + ], + "lines": [ + { + "bbox": [ + 116, + 182, + 171, + 192 + ], + "spans": [ + { + "bbox": [ + 116, + 182, + 171, + 192 + ], + "type": "text", + "content": "(d) OxfordPets" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 194, + 203, + 271 + ], + "lines": [ + { + "bbox": [ + 85, + 194, + 203, + 271 + ], + "spans": [ + { + "bbox": [ + 85, + 194, + 203, + 271 + ], + "type": "table", + "html": "
BaseNewHos
CLIP91.1797.2694.12
CoOp93.6795.2994.47
CoCoOp95.2097.6996.43
ProDA95.4397.8396.62
Ours95.8697.5596.70
", + "image_path": "e8ce3fd5a548b03855426d1e5ed6d09bae2a9c0a2c65d3aed1a748cc45c9c765.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 235, + 194, + 354, + 271 + ], + "blocks": [ + { + "bbox": [ + 264, + 182, + 326, + 192 + ], + "lines": [ + { + "bbox": [ + 264, + 182, + 326, + 192 + ], + "spans": [ + { + "bbox": [ + 264, + 182, + 326, + 192 + ], + "type": "text", + "content": "(e) StanfordCars" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 235, + 194, + 354, + 271 + ], + "lines": [ + { + "bbox": [ + 235, + 194, + 354, + 271 + ], + "spans": [ + { + "bbox": [ + 235, + 194, + 354, + 271 + ], + "type": "table", + "html": "
BaseNewHos
CLIP63.3774.8968.65
CoOp78.1260.4068.13
CoCoOp70.4973.5972.01
ProDA74.7071.2072.91
Ours76.2974.1775.22
", + "image_path": "f912b822cf107216189605348ece0ca763f8f65455b5003b6ba629dfb378de59.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 391, + 194, + 509, + 271 + ], + "blocks": [ + { + "bbox": [ + 422, + 182, + 477, + 192 + ], + "lines": [ + { + "bbox": [ + 422, + 182, + 477, + 192 + ], + "spans": [ + { + "bbox": [ + 422, + 182, + 477, + 192 + ], + "type": "text", + "content": "(f) Flowers102" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 391, + 194, + 509, + 271 + ], + "lines": [ + { + "bbox": [ + 391, + 194, + 509, + 271 + ], + "spans": [ + { + "bbox": [ + 391, + 194, + 509, + 271 + ], + "type": "table", + "html": "
BaseNewHos
CLIP72.0877.8074.83
CoOp97.6059.6774.06
CoCoOp94.8771.7581.71
ProDA97.7068.6880.66
Ours97.3677.7086.43
", + "image_path": "071f43b754916eafb0ce6339ad4eaf234458486d0b13721cae51879257d422cf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 85, + 298, + 203, + 376 + ], + "blocks": [ + { + "bbox": [ + 121, + 286, + 166, + 297 + ], + "lines": [ + { + "bbox": [ + 121, + 286, + 166, + 297 + ], + "spans": [ + { + "bbox": [ + 121, + 286, + 166, + 297 + ], + "type": "text", + "content": "(g) Food101" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 298, + 203, + 376 + ], + "lines": [ + { + "bbox": [ + 85, + 298, + 203, + 376 + ], + "spans": [ + { + "bbox": [ + 85, + 298, + 203, + 376 + ], + "type": "table", + "html": "
BaseNewHos
CLIP90.1091.2290.66
CoOp88.3382.2685.19
CoCoOp90.7091.2990.99
ProDA90.3088.5789.43
Ours90.5492.3191.42
", + "image_path": "43154d31c91896a9bad637cae47d366402f747f014c92e1a8aeb8b1eaa136f35.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 235, + 298, + 354, + 376 + ], + "blocks": [ + { + "bbox": [ + 261, + 286, + 328, + 297 + ], + "lines": [ + { + "bbox": [ + 261, + 286, + 328, + 297 + ], + "spans": [ + { + "bbox": [ + 261, + 286, + 328, + 297 + ], + "type": "text", + "content": "(h) FGVCAircraft" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 235, + 298, + 354, + 376 + ], + "lines": [ + { + "bbox": [ + 235, + 298, + 354, + 376 + ], + "spans": [ + { + "bbox": [ + 235, + 298, + 354, + 376 + ], + "type": "table", + "html": "
BaseNewHos
CLIP27.1936.2931.09
CoOp40.4422.3028.75
CoCoOp33.4123.7127.74
ProDA36.9034.1335.46
Ours39.4935.3737.32
", + "image_path": "a9ab21687c8aed961cb6a9cc9cc5b40cb3902cc803b04f71bdab1518a0ed07b5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 391, + 298, + 509, + 376 + ], + "blocks": [ + { + "bbox": [ + 427, + 286, + 471, + 297 + ], + "lines": [ + { + "bbox": [ + 427, + 286, + 471, + 297 + ], + "spans": [ + { + "bbox": [ + 427, + 286, + 471, + 297 + ], + "type": "text", + "content": "(i) SUN397" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 391, + 298, + 509, + 376 + ], + "lines": [ + { + "bbox": [ + 391, + 298, + 509, + 376 + ], + "spans": [ + { + "bbox": [ + 391, + 298, + 509, + 376 + ], + "type": "table", + "html": "
BaseNewHos
CLIP69.3675.3572.23
CoOp80.6065.8972.51
CoCoOp79.7476.8678.27
ProDA78.6776.9377.79
Ours82.1677.4979.76
", + "image_path": "740126e6a45959c30423d6a0d11b601394d44018a1148beb525d360ac55f13a8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 82, + 403, + 200, + 481 + ], + "blocks": [ + { + "bbox": [ + 125, + 392, + 156, + 402 + ], + "lines": [ + { + "bbox": [ + 125, + 392, + 156, + 402 + ], + "spans": [ + { + "bbox": [ + 125, + 392, + 156, + 402 + ], + "type": "text", + "content": "(j) DTD" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 403, + 200, + 481 + ], + "lines": [ + { + "bbox": [ + 82, + 403, + 200, + 481 + ], + "spans": [ + { + "bbox": [ + 82, + 403, + 200, + 481 + ], + "type": "table", + "html": "
BaseNewHos
CLIP53.2459.9056.37
CoOp79.4441.1854.24
CoCoOp77.0156.0064.85
ProDA80.6756.4866.44
Ours79.4761.5369.36
", + "image_path": "1d5aef3a67e15e721b3d10805934a4be3beb918c71194b334bd9e190eacb864a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 236, + 403, + 354, + 482 + ], + "blocks": [ + { + "bbox": [ + 272, + 392, + 320, + 401 + ], + "lines": [ + { + "bbox": [ + 272, + 392, + 320, + 401 + ], + "spans": [ + { + "bbox": [ + 272, + 392, + 320, + 401 + ], + "type": "text", + "content": "(k) EuroSAT" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 236, + 403, + 354, + 482 + ], + "lines": [ + { + "bbox": [ + 236, + 403, + 354, + 482 + ], + "spans": [ + { + "bbox": [ + 236, + 403, + 354, + 482 + ], + "type": "table", + "html": "
BaseNewHos
CLIP56.4864.0560.03
CoOp92.1954.7468.69
CoCoOp87.4960.0471.21
ProDA83.9066.0073.88
Ours92.1473.8782.00
", + "image_path": "e5924c8948751be0af4c9408ef186c270431e7436e657651c34dbb4aa522cecb.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 391, + 403, + 510, + 482 + ], + "blocks": [ + { + "bbox": [ + 429, + 392, + 471, + 401 + ], + "lines": [ + { + "bbox": [ + 429, + 392, + 471, + 401 + ], + "spans": [ + { + "bbox": [ + 429, + 392, + 471, + 401 + ], + "type": "text", + "content": "(1)UCF101" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 391, + 403, + 510, + 482 + ], + "lines": [ + { + "bbox": [ + 391, + 403, + 510, + 482 + ], + "spans": [ + { + "bbox": [ + 391, + 403, + 510, + 482 + ], + "type": "table", + "html": "
BaseNewHos
CLIP70.5377.5073.85
CoOp84.6956.0567.46
CoCoOp82.3373.4577.64
ProDA85.2371.9778.04
Ours84.1277.7480.80
", + "image_path": "42c3e9b17fb490a0a7055fcf3c97bb6bfd80e6e307b30a4e469f30e5a88c2d3d.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 490, + 547, + 514 + ], + "lines": [ + { + "bbox": [ + 46, + 490, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 547, + 514 + ], + "type": "text", + "content": "Table 1: Results (%) of the base-to-new generalization task on 11 benchmark datasets. We report the accuracy with CLIP ViT-B/16 model on the base classes (Base), the unseen classes (New), and the harmonic mean of both of them (Hos)." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 535, + 270, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 270, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 270, + 548 + ], + "type": "text", + "content": "we then have the predicted probability of the " + }, + { + "bbox": [ + 47, + 535, + 270, + 548 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 535, + 270, + 548 + ], + "type": "text", + "content": "-th class:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 63, + 552, + 287, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 552, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 63, + 552, + 287, + 583 + ], + "type": "interline_equation", + "content": "P (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f _ {x} , G \\left(p _ {i} ^ {a}\\right) / \\tau\\right)\\right.}{\\sum_ {j = 1} ^ {C} \\exp \\left( \\right.\\cos \\left(f _ {x} , G \\left(p _ {j} ^ {a}\\right) / \\tau\\right)} .} \\tag {7}", + "image_path": "ab5696fd627db0d096beb5e907f2fa0356e38c5aea940baab19de5060af80ec7.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 624 + ], + "type": "text", + "content": "We generate class-aware prompts instead of fusing identical image semantics with prompts of different classes, bringing category discrimination to the specific downstream tasks." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "text", + "content": "TFT leverages text features to guide images to focus on task-related regions. Specifically, using the embeddings " + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "inline_equation", + "content": "g^{a}" + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "text", + "content": " of the augmented prompts " + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "inline_equation", + "content": "p^{a}" + }, + { + "bbox": [ + 46, + 628, + 287, + 664 + ], + "type": "text", + "content": " as input, we have attention:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 134, + 669, + 287, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 669, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 134, + 669, + 287, + 684 + ], + "type": "interline_equation", + "content": "A ^ {x} = f _ {x} \\left(g ^ {a}\\right) ^ {T}, \\tag {8}", + "image_path": "5334f423e50d0fc635ba7d432684c7c6b38f38d7d90aa2db23b1fb176e517884.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 47, + 689, + 287, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 716 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 716 + ], + "type": "inline_equation", + "content": "A^x \\in \\mathbb{R}^{N \\times M}" + }, + { + "bbox": [ + 47, + 689, + 287, + 716 + ], + "type": "text", + "content": " denotes text-to-image attention map. " + }, + { + "bbox": [ + 47, + 689, + 287, + 716 + ], + "type": "inline_equation", + "content": "A_{i,j}^x" + }, + { + "bbox": [ + 47, + 689, + 287, + 716 + ], + "type": "text", + "content": " represents the similarity between the " + }, + { + "bbox": [ + 47, + 689, + 287, + 716 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 689, + 287, + 716 + ], + "type": "text", + "content": "-th patch in the" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 535, + 547, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 535, + 547, + 571 + ], + "spans": [ + { + "bbox": [ + 305, + 535, + 547, + 571 + ], + "type": "text", + "content": "image and the " + }, + { + "bbox": [ + 305, + 535, + 547, + 571 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 535, + 547, + 571 + ], + "type": "text", + "content": "-th class in the text representation. Similar to image-to-text, we use it to query the class-related part of the text correlated to the image, augmenting image features:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 367, + 576, + 545, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 576, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 367, + 576, + 545, + 590 + ], + "type": "interline_equation", + "content": "f ^ {a} = \\operatorname {s o f t m a x} \\left(A ^ {x}\\right) g ^ {a} + f _ {x}, \\tag {9}", + "image_path": "cc3b3508701ffa5c751caef786dfb7cfe1d249a10ee64fa792aa03e2327e2ffe.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 305, + 594, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 619 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 594, + 545, + 619 + ], + "type": "inline_equation", + "content": "f^a" + }, + { + "bbox": [ + 305, + 594, + 545, + 619 + ], + "type": "text", + "content": " is the augmented image embeddings. We thus let image branch focus on the tasks-related representation." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 305, + 623, + 547, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 623, + 547, + 682 + ], + "spans": [ + { + "bbox": [ + 305, + 623, + 547, + 682 + ], + "type": "text", + "content": "Augmented contrastive loss function is then employed to further align class-aware text embedding and text-guided image features on specific downstream tasks. The predicted probability of the " + }, + { + "bbox": [ + 305, + 623, + 547, + 682 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 623, + 547, + 682 + ], + "type": "text", + "content": "-th class, which is used to calculate the contrastive loss, after mutual augmentation is:" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 327, + 686, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 686, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 327, + 686, + 545, + 717 + ], + "type": "interline_equation", + "content": "P (y = i \\mid x) = \\frac {\\exp \\left(\\cos \\left(f ^ {a} , g _ {i} ^ {a}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(\\cos \\left(f ^ {a} , g _ {i} ^ {a} / \\tau\\right)\\right)}. \\tag {10}", + "image_path": "957e7b427c8a39f42ee06717635e94b6591d0d57443a656c5303f5372d545e21.jpg" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21963" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 260, + 224 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 260, + 224 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 260, + 224 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 260, + 224 + ], + "type": "image", + "image_path": "b81f29fe9de61578232259f33377c8c86cf683e8e61ebec5a1184b4c8911a78c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 234, + 546, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 234, + 546, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 546, + 258 + ], + "type": "text", + "content": "Figure 4: Absolute improvement over CoCoOp in the base-to-new generalization task. Compared to CoCoOp, Our method achieves improvement on both base (left sub-figure) and new (right sub-figure) classes on most of the datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 263, + 69, + 521, + 225 + ], + "blocks": [ + { + "bbox": [ + 263, + 69, + 521, + 225 + ], + "lines": [ + { + "bbox": [ + 263, + 69, + 521, + 225 + ], + "spans": [ + { + "bbox": [ + 263, + 69, + 521, + 225 + ], + "type": "image", + "image_path": "a0bb075f465cafcb3179aa574b98a795062f9d0c9165364270cf7f56c8b495bb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 269, + 545, + 364 + ], + "blocks": [ + { + "bbox": [ + 50, + 269, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 50, + 269, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 50, + 269, + 545, + 364 + ], + "type": "table", + "html": "
ImageNetCaltech101OxfordPedsStanfordCarsFlowers102Food101FGVCAircraftSUN397DTDEuroSATUCF101Average
CLIP68.6389.3688.9965.6770.4989.2327.1265.2946.0254.1769.8366.80
CoOp71.5195.5393.3174.2595.7087.2334.1874.8268.4677.8277.2977.28
CoCoOp71.0293.4393.9371.2187.3487.3932.0372.3263.8472.7877.4074.79
Ours72.9095.9093.9679.1096.7389.9538.7279.3772.4981.0083.4580.32
", + "image_path": "ab85fd60dd181cbaef192e783ea0ab702aba3a0ce320f3e015187b7af8beff47.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 171, + 371, + 420, + 384 + ], + "lines": [ + { + "bbox": [ + 171, + 371, + 420, + 384 + ], + "spans": [ + { + "bbox": [ + 171, + 371, + 420, + 384 + ], + "type": "text", + "content": "Table 2: Results (%) of 16-shot learning task on 11 datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 441 + ], + "type": "text", + "content": "Task-targeted semantic information is transferred between the two branches by minimizing the augmented contrastive loss. We merge probability before and after augmentation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 447, + 299, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 299, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 299, + 487 + ], + "type": "interline_equation", + "content": "P (y = i \\mid x) = \\frac {\\exp \\left(\\left(\\cos \\left(f , g _ {i}\\right) + \\lambda \\left(\\cos \\left(f ^ {a} , g _ {i} ^ {a}\\right)\\right) / \\tau\\right)\\right.}{\\sum_ {j = 1} ^ {C} \\exp \\left( \\right.\\left(\\cos \\left(f , g _ {i}\\right) + \\lambda \\left(\\cos \\left(f ^ {a} , g _ {i} ^ {a}\\right)\\right) / \\tau\\right)} \\tag {11}", + "image_path": "f0e266afd2e543a1e3c210bdc1416127a0c46e80bff3f47a9f704c703030995f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 487, + 287, + 536 + ], + "type": "text", + "content": " is the balance hyper-parameter, which is analyzed in our experiments. We let the two different modalities tightly coupled and mutual beneficial across the whole training process by performing the contrastive optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 544, + 128, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 128, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 128, + 559 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 565, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 287, + 625 + ], + "type": "text", + "content": "We evaluate the performance of our method on four generalization tasks, including 1) generalization from base classes to new classes; 2) few-shot classification; 3) cross-dataset transfer; 4) domain generalization. After that, we provide extensive ablation studies and in-depth analyses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Datasets. Following [33, 51], we use 11 image recognition datasets for the tasks of base-to-new generalization, few-shot classification and cross-dataset transfer. It contains generic image classification datasets (ImageNet [6] and Caltech101 [8]), fine-grained classification datasets (Oxford Pets [32], StanfordCars [18], Flowers102 [31], Food101 [2] and FGVCAircraft [29]), scene recognition ( SUN397" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 405, + 546, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 546, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 546, + 478 + ], + "type": "text", + "content": "[44]), action recognition (UCF101 [37]), texture classification (DTD [4]), and satellite imagery recognition (EuroSAT [11]). For the domain generalization task, we use ImageNet as the source dataset and select ImageNetV2 [35], ImageNet-Sketch [41], ImageNet-A [13], and ImageNet-R [12], which are the ImageNet variants, as the target." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 485, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 581 + ], + "type": "text", + "content": "Training Details. By following [50, 51], we use the best visual backbone available in CLIP, i.e., ViT-B/16, throughout the experiments. We train 10 epochs using SGD optimizer with base learning rate of 0.002 and cosine decay schedule. We set the hyper-parameter " + }, + { + "bbox": [ + 304, + 485, + 547, + 581 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 485, + 547, + 581 + ], + "type": "text", + "content": " in Eq. (11) to 0.2 for all experiments, and provide sensitivity analyses in Fig. 5. We run all the experiments three times with different random seeds and report the average classification accuracy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 589, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 547, + 662 + ], + "type": "text", + "content": "Baselines. We compare our method with 4 baselines. (1) Zero-shot CLIP [33] with hand-crafted prompts. (2) CoOp [51], using automatically generated prompts from few data. (3) CoCoOp [50], dynamically generating prompts conditioned on the images. (4) ProDA [28], which learns prompts from few data samples and mitigates the domain gap." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 670, + 524, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 670, + 524, + 682 + ], + "spans": [ + { + "bbox": [ + 305, + 670, + 524, + 682 + ], + "type": "text", + "content": "4.1. Generalization From Base to New Classes" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 714 + ], + "type": "text", + "content": "Following the previous works, we split the classes equally into two groups for each dataset: one as base and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21964" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 164 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 164 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 164 + ], + "type": "table", + "html": "
SourceTarget
ImageNetCaletech101OxfordPedsStanfordCarsFlowers102Food101FGVCAircraftSUN397DTDEuroSATUCF101
CoOp71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.55
CoCoOp71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.21
Ours72.9095.7390.2265.1469.8986.3823.3266.4946.4747.2467.43
", + "image_path": "915d29625ac35d523db30c1819ace31fa700c5b6e2a0308dd88274171fd707a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 194, + 544, + 281 + ], + "blocks": [ + { + "bbox": [ + 52, + 172, + 539, + 185 + ], + "lines": [ + { + "bbox": [ + 52, + 172, + 539, + 185 + ], + "spans": [ + { + "bbox": [ + 52, + 172, + 539, + 185 + ], + "type": "text", + "content": "Table 3: Results of cross-dataset transfer task. Each method is trained on the source dataset and evaluated on the target." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 194, + 544, + 281 + ], + "lines": [ + { + "bbox": [ + 50, + 194, + 544, + 281 + ], + "spans": [ + { + "bbox": [ + 50, + 194, + 544, + 281 + ], + "type": "table", + "html": "
SourceTarget
ImageNetImageNetV2ImageNet-SketchImageNet-AImageNet-R
CLIP66.7360.8346.1547.7773.96
CoOp71.5164.2047.9949.7175.21
CoCoOp71.0264.0748.7550.6376.18
Ours72.9064.5749.1150.9476.68
", + "image_path": "2f90f0dc5531ead41b65f708345b085826f102f172e2838f78746fdef4286f9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 289, + 543, + 302 + ], + "lines": [ + { + "bbox": [ + 48, + 289, + 543, + 302 + ], + "spans": [ + { + "bbox": [ + 48, + 289, + 543, + 302 + ], + "type": "text", + "content": "Table 4: Results of domain generalization task. Each method is trained on ImageNet and evaluated on ImageNet variants." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "text", + "content": "the other as new. The learnable modules are trained exclusively on the base classes, while evaluation is carried out separately on both the base and new classes to testify generalization ability. We report the results on 11 benchmarks in Table 1. Although compared to CoOp, CoCoOp significantly narrows generalization gap in unseen classed, but it decreases the accuracy in seen classes from " + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "inline_equation", + "content": "82.69\\%" + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "inline_equation", + "content": "80.47\\%" + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "text", + "content": ". We attribute it to the homogeneous prompts of CoCoOp, which weakens the discriminative semantics of different categories. In comparison, our method improves the accuracy in unseen classes from " + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "inline_equation", + "content": "80.47\\%" + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "inline_equation", + "content": "83.01\\%" + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "text", + "content": " by prompting each text label with corresponding image information. Benefit from the mutual learning of our CTP and TFT modules, our method further improves the accuracy in unseen classes from " + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "inline_equation", + "content": "71.69\\%" + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "inline_equation", + "content": "75.72\\%" + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "text", + "content": ", even surpasses the accuracy of CLIP hand-crafted prompts. We provide a detailed comparisons of CoCoOp and our method of per-dataset improvement in Fig. 4. Our method gains significant improvements over CoCoOp in both seen and unseen classes on 10 out of 11 recognition datasets. Surprisingly, our method significantly improves CoCoOp by more than " + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 323, + 289, + 586 + ], + "type": "text", + "content": " in unseen classes on SUN397 and Flowers102 datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 597, + 179, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 179, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 179, + 609 + ], + "type": "text", + "content": "4.2. Few-Shot Classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "We report few-shot classification results in Table 2. Our method surpasses baseline methods on all datasets in the few-shot setting. Especially, our method outperforms CoCoOp by " + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "inline_equation", + "content": "9.39\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "inline_equation", + "content": "8.65\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "inline_equation", + "content": "8.22\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": " on Flowers102, DTD, and EuroSAT, respectively, and the average improvement over 11 datasets is " + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "inline_equation", + "content": "5.53\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": ". Our method also achieve " + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": " on the challenging dataset of ImageNet. The above experiments show the great discriminative ability of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 323, + 437, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 437, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 437, + 335 + ], + "type": "text", + "content": "4.3. Cross-Dataset Transfer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 341, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 449 + ], + "type": "text", + "content": "We then evaluate the generalization ability of our method on more challenging cross-dataset tasks. In this setting, we learn multi-modal prompts on ImageNet of 1000 classes. The effectiveness of the learned prompts is then tested on 10 datasets containing generic and fine-grained image classification, scene recognition, and texture classification. The results are reported in Table 3. Our method achieves the best average accuracy on the 11 datasets, especially ImageNet. It demonstrates the great transfer ability of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 455, + 438, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 438, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 438, + 467 + ], + "type": "text", + "content": "4.4. Domain Generalization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 474, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 653 + ], + "type": "text", + "content": "The domain generalization setting evaluates the generalization ability of the model on the target domain that is similar to but different from the source domain [48, 47]. Zero-shot CLIP introduces no additional training parameters and exhibits great robustness to naturally distribution shifts. Other methods use few samples to train learnable parameters, there is a risk of overfitting the source distribution. Therefore, we conduct experiments using ImageNet as the source domain and evaluate the ability of generalizing to unknown on four ImageNet variants. The results are shown in Table 4. Our method achieves significant performance on the 4 ImageNet variant datasets. It verifies that our method improves the classification ability of the source domain dataset while maintaining the generalization on the target domain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 659, + 411, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 411, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 411, + 672 + ], + "type": "text", + "content": "4.5. Ablation Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Effectiveness of each module. To evaluate the effectiveness of Class-aware Text Prompts (CTP) and Text-guided Feature Tuning (TFT) of our method, we conduct ablation" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21965" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 544, + 288 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 544, + 288 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 544, + 288 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 544, + 288 + ], + "type": "table", + "html": "
MethodAverageImageNetCaltech101OxfordPeds
BaseNewHosBaseNewHosBaseNewHosBaseNewHos
A82.6963.2271.6676.4767.8871.9298.0089.8193.7393.6795.2994.47
B82.3872.4476.6476.9669.6273.1198.2394.2496.1995.6197.9796.78
C82.9372.9877.2477.2169.8673.3598.4492.7295.4995.4997.8196.64
Ours83.0175.7279.0277.4270.4473.7798.3194.7596.5095.8697.5596.70
MethodStanfordCarsFlowers102Food101FGVCAircraft
BaseNewHosBaseNewHosBaseNewHosBaseNewHos
A78.1260.4068.1397.6059.6774.0688.3382.2685.1940.4422.3028.75
B74.6273.6874.1596.7266.4278.7690.3091.4790.8836.4134.3935.37
C75.8474.5375.1897.3274.8684.6390.5691.6591.1037.8233.1735.34
Ours76.2974.1775.2297.3677.7086.4390.5492.3191.4239.4935.3737.32
MethodSUN397DTDEuroSATUCF101
BaseNewHosBaseNewHosBaseNewHosBaseNewHos
A80.6065.8972.5179.4441.1854.2492.1954.7468.6984.6956.0567.46
B81.7376.8979.2480.1851.7962.9391.7067.6277.8483.7272.7277.83
C82.2976.2479.1581.7154.7465.5690.1060.5272.4185.4076.6880.81
Ours82.1677.4979.7679.4761.5369.3692.1473.8782.0084.1277.7480.80
", + "image_path": "2adbc8f916b8cb58d045b8da4d43fbacec886fcad666d8d9e739f65f78d68af6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 341, + 285, + 435 + ], + "blocks": [ + { + "bbox": [ + 46, + 297, + 546, + 322 + ], + "lines": [ + { + "bbox": [ + 46, + 297, + 546, + 322 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 546, + 322 + ], + "type": "text", + "content": "Table 5: Ablation studies of our method on 11 datasets. Three ablation cases are considered: A: Ours w/o TVP w/o FTP. B: Ours w/o TVP. C: Ours w/o FTP. TVP is the text-reorganized vision prompt, and FTP is the fine-grained text prompt." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 341, + 285, + 435 + ], + "lines": [ + { + "bbox": [ + 50, + 341, + 285, + 435 + ], + "spans": [ + { + "bbox": [ + 50, + 341, + 285, + 435 + ], + "type": "table", + "html": "
Prompt LearningFeature TuningAccuracy (%)
MLP-PLCTPMLP-FTTFT
✓71.66 (CoOp)
75.83 (+4.17)
✓76.64 (+4.98)
✓75.94 (+4.28)
✓77.24 (+5.58)
✓✓77.05 (+5.39)
✓✓79.02 (+7.36)
", + "image_path": "5ddb6e23144a422d60d603476eb4be8a7bd4903cf84458412d091b97f5f84dce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 287, + 529 + ], + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 529 + ], + "type": "text", + "content": "Table 6: Comparison of different structures for prompt learning and feature tuning. The average results of harmonic mean of from-base-to-new generalization task on 11 datasets are reported. In compared to our attention design in CTP and TFT modules, MLP-PL and MLP-FT are designed using the Linear-ReLU-Linear block setting of [50]. Improvements over the baseline of CoOp, are marked in green." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 552, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 647 + ], + "type": "text", + "content": "experiments on 11 datasets, as reported in Table 5. In most cases, each module significantly improves the performance of the model. For average results, CTP and TFT improves the results by " + }, + { + "bbox": [ + 46, + 552, + 287, + 647 + ], + "type": "inline_equation", + "content": "5.58\\%" + }, + { + "bbox": [ + 46, + 552, + 287, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 552, + 287, + 647 + ], + "type": "inline_equation", + "content": "4.98\\%" + }, + { + "bbox": [ + 46, + 552, + 287, + 647 + ], + "type": "text", + "content": ", respectively, and the combination of them improves the results by " + }, + { + "bbox": [ + 46, + 552, + 287, + 647 + ], + "type": "inline_equation", + "content": "7.36\\%" + }, + { + "bbox": [ + 46, + 552, + 287, + 647 + ], + "type": "text", + "content": ". It shows the effectiveness of the two branches of text-to-image and image-to-text, the mutual learning of the two modules further improves the performance on downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Comparison of different structure design of multi-modal mutual learning. To further provide in-depth analysis about our mutual learning, we further explore two vanilla structures: (1) MLP-PL: The image features are forwarded to a block of Linear-ReLU-Linear, borrowed from [50], and" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 341, + 421, + 449 + ], + "blocks": [ + { + "bbox": [ + 306, + 341, + 421, + 449 + ], + "lines": [ + { + "bbox": [ + 306, + 341, + 421, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 421, + 449 + ], + "type": "image", + "image_path": "2ff833bdc7e9960031070fad5962b722dfa853110b5d03a62bf13b6b40ba2d08.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 457, + 545, + 482 + ], + "lines": [ + { + "bbox": [ + 305, + 457, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 305, + 457, + 545, + 482 + ], + "type": "text", + "content": "Figure 5: Sensitivity analysis of " + }, + { + "bbox": [ + 305, + 457, + 545, + 482 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 457, + 545, + 482 + ], + "type": "text", + "content": ", with base, new, and host metrics, on UCF101 (left) and Caltech101 (right) datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 429, + 341, + 544, + 449 + ], + "blocks": [ + { + "bbox": [ + 429, + 341, + 544, + 449 + ], + "lines": [ + { + "bbox": [ + 429, + 341, + 544, + 449 + ], + "spans": [ + { + "bbox": [ + 429, + 341, + 544, + 449 + ], + "type": "image", + "image_path": "b464df958c9d49e0af0522b8b1ee349247eb7f7e1f893f457bc10de6e5195241.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": "then added to the text for augmenting it (the same to us). (2) MLP-FT: The text prompts are forwarded to the Linear-ReLU-Linear block, and then added to the image for augmenting it (the same to us). In comparison, our class-aware text prompts (CTP) module and text-guided feature tuning (TFT) module, adopt text-image attention to learn the augmented features instead of the Linear-ReLU-Linear block. We report the results of the different designs in Table 6. First, we find that combining MLP-PL & MLP-FT and CTP & TFT can both improve the results compared with using either of them. It indicates that both prompt learning and feature tuning are important to achieve better results. Second, compared with the design of Linear-ReLU-Linear block, our design of text-image attention further improves performance by " + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "inline_equation", + "content": "0.81\\%" + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": " for prompt learning and feature tuning, respectively. It demonstrates the effectiveness of our design of attention, which helps the model to focus on class-aware and task-related semantics. Third, com" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21966" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "pared with CoOp, both of the designs could improve the final results by large margins. The key factor of our mutual learning to achieve significant performance is the task-related alignment of vision and language in latent space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 125, + 287, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 125, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 46, + 125, + 287, + 173 + ], + "type": "text", + "content": "Sensitivity Analysis of " + }, + { + "bbox": [ + 46, + 125, + 287, + 173 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 125, + 287, + 173 + ], + "type": "text", + "content": ". We evaluate the parameter sensitivity of " + }, + { + "bbox": [ + 46, + 125, + 287, + 173 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 125, + 287, + 173 + ], + "type": "text", + "content": " of Eq. (11) in Fig. 5. The results suggest that the performance of our method is generally robust to " + }, + { + "bbox": [ + 46, + 125, + 287, + 173 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 125, + 287, + 173 + ], + "type": "text", + "content": ", indicating a wide range of " + }, + { + "bbox": [ + 46, + 125, + 287, + 173 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 125, + 287, + 173 + ], + "type": "text", + "content": " works well in downstream tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 182, + 119, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 119, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 119, + 194 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 202, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 202, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 287, + 357 + ], + "type": "text", + "content": "In this paper, we introduce task-oriented multi-modal mutual learning for adapting large vision-language models to downstream vision tasks. We propose class-aware text prompt and text-guided feature tuning to unleash the potential of the vision-language model by re-activating its task-related representation abilities. Our method yields impressive generalization performance on a wide range of vision tasks and datasets. We hope the presented findings and insights in this paper could benefit the following works in designing more efficient and effective adaptation methods. For the future work, we think it is interesting to extend the adaptation of vision language models to more vision tasks, such as semantic segmentation, object detection, etc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 358, + 287, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 406 + ], + "type": "text", + "content": "Limitations. Similar to CoCoOp, we learn image-conditioned representations, thus the batch-size of training need to be set to 1, which is not efficient enough for learning. We aim to solve this efficiency issue in the future work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 415, + 164, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 164, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 164, + 428 + ], + "type": "text", + "content": "6. Acknowledgements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 436, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 495 + ], + "type": "text", + "content": "This work is supported by the Innovation Capacity Construction Project of Jilin Province Development and Reform Commission (2021FGWCXNLJSSZ10) and the Fundamental Research Funds for the Central Universities, JLU. Zichang Tan is the project lead." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 506, + 105, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 506, + 105, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 506, + 105, + 517 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 525, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 525, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 287, + 567 + ], + "type": "text", + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting largescale models. arXiv preprint arXiv:2203.17274, 1(3):4, 2022.3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 570, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 570, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 570, + 287, + 624 + ], + "type": "text", + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VI 13, pages 446-461. Springer, 2014. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 625, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 287, + 690 + ], + "type": "text", + "content": "[3] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX, pages 104-120. Springer, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 712 + ], + "type": "text", + "content": "[4] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 96, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 547, + 162 + ], + "type": "text", + "content": "[5] Joe Davison, Joshua Feldman, and Alexander M Rush. Commonsense knowledge mining from pretrained models. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pages 1173-1178, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 545, + 207 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 209, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 209, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 209, + 545, + 262 + ], + "type": "text", + "content": "[7] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 545, + 319 + ], + "type": "text", + "content": "[8] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recognition workshop, pages 178-178. IEEE, 2004. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 321, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 321, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 312, + 321, + 545, + 354 + ], + "type": "text", + "content": "[9] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pretrained language models better few-shot learners. arXiv preprint arXiv:2012.15723, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 355, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 387 + ], + "type": "text", + "content": "[10] Adi Haviv, Jonathan Berant, and Amir Globerson. Bertese: Learning to speak to bert. arXiv preprint arXiv:2103.05327, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "type": "text", + "content": "[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "type": "text", + "content": "[12] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[13] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "text", + "content": "[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 613, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 677 + ], + "type": "text", + "content": "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[16] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Computer Vision-ECCV 2022: 17th" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21967" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIII, pages 709-727. Springer, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[17] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "text", + "content": "[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 241 + ], + "type": "text", + "content": "[19] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "text", + "content": "[20] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 341 + ], + "type": "text", + "content": "[21] Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. Unimo: Towards unified-modal understanding and generation via cross-modal contrastive learning. arXiv preprint arXiv:2012.15409, 2020.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 419 + ], + "type": "text", + "content": "[22] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 121-137. Springer, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 453 + ], + "type": "text", + "content": "[23] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "text", + "content": "[24] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. arXiv preprint arXiv:2212.00794, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 543 + ], + "type": "text", + "content": "[25] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "text", + "content": "[26] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands, too. arXiv preprint arXiv:2103.10385, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "text", + "content": "[27] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[28] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 1, 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[30] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI, pages 529-544. Springer, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[32] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 220, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 284 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 286, + 545, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 351 + ], + "type": "text", + "content": "[34] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "text", + "content": "[35] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International conference on machine learning, pages 5389-5400. PMLR, 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 398, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 545, + 442 + ], + "type": "text", + "content": "[36] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 444, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 485 + ], + "type": "text", + "content": "[37] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. A dataset of 101 human action classes from videos in the wild. Center for Research in Computer Vision, 2(11), 2012. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 488, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 531 + ], + "type": "text", + "content": "[38] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "type": "text", + "content": "[39] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "type": "text", + "content": "[40] Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial triggers for attacking and analyzing nlp. arXiv preprint arXiv:1908.07125, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 612, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 656 + ], + "type": "text", + "content": "[41] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems, 32, 2019. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[42] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "text", + "content": "[43] Junyang Wu, Xianhang Li, Chen Wei, Huiyu Wang, Alan Yuille, Yuyin Zhou, and Cihang Xie. Unleashing the power" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21968" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 497 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "of visual prompting at the pixel level. arXiv preprint arXiv:2212.10556, 2022.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[44] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "text", + "content": "[45] Yifan Yang, Weiquan Huang, Yixuan Wei, Houwen Peng, Xinyang Jiang, Huiqiang Jiang, Fangyun Wei, Yin Wang, Han Hu, Lili Qiu, et al. Attentive mask clip. arXiv preprint arXiv:2212.08653, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 251 + ], + "type": "text", + "content": "[46] Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-vil: Knowledge enhanced vision-language representations through scene graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 252, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 252, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 252, + 287, + 308 + ], + "type": "text", + "content": "[47] Junkun Yuan, Xu Ma, Defang Chen, Kun Kuang, Fei Wu, and Lanfen Lin. Label-efficient domain generalization via collaborative exploration and generalization. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2361-2370, 2022. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 352 + ], + "type": "text", + "content": "[48] Junkun Yuan, Xu Ma, Defang Chen, Kun Kuang, Fei Wu, and Lanfen Lin. Domain-specific bias filtering for single labeled domain generalization. International Journal of Computer Vision, 131(2):552-571, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 396 + ], + "type": "text", + "content": "[49] Weizhe Yuan, Graham Neubig, and Pengfei Liu. Bartscore: Evaluating generated text as text generation. Advances in Neural Information Processing Systems, 34:27263-27277, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 398, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 398, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 398, + 287, + 453 + ], + "type": "text", + "content": "[50] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "type": "text", + "content": "[51] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21969" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/d73ef7d6-b7de-4511-b8bf-4b4e20276122_content_list.json b/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/d73ef7d6-b7de-4511-b8bf-4b4e20276122_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4366c2c0a3a4180d733a64688f33dc5636149d41 --- /dev/null +++ b/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/d73ef7d6-b7de-4511-b8bf-4b4e20276122_content_list.json @@ -0,0 +1,1580 @@ +[ + { + "type": "text", + "text": "Task-aware Adaptive Learning for Cross-domain Few-shot Learning", + "text_level": 1, + "bbox": [ + 138, + 130, + 831, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yurong Guo $^{1}$ , Ruoyi Du $^{1}$ , Yuan Dong $^{1}$ , Timothy Hospedales $^{2}$ , Yi-Zhe Song $^{3}$ , Zhanyu Ma $^{1*}$ $^{1}$ Beijing University of Posts and Telecommunications, China \n $^{2}$ University of Edinburgh, UK \n $^{3}$ University of Surrey, UK", + "bbox": [ + 127, + 179, + 846, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{guoyurong, duruoyi, yuandong, mazhanyu}@bupt.edu.cn, t.hospedales@ed.ac.uk, y.song@surrey.ac.uk", + "bbox": [ + 251, + 236, + 718, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although existing few-shot learning works yield promising results for in-domain queries, they still suffer from weak cross-domain generalization. Limited support data requires effective knowledge transfer, but domain-shift makes this harder. Towards this emerging challenge, researchers improved adaptation by introducing task-specific parameters, which are directly optimized and estimated for each task. However, adding a fixed number of additional parameters fails to consider the diverse domain shifts between target tasks and the source domain, limiting efficacy. In this paper, we first observe the dependence of task-specific parameter configuration on the target task. Abundant task-specific parameters may over-fit, and insufficient task-specific parameters may result in under-adaptation – but the optimal task-specific configuration varies for different test tasks. Based on these findings, we propose the Task-aware Adaptive Network (TA $^2$ -Net), which is trained by reinforcement learning to adaptively estimate the optimal task-specific parameter configuration for each test task. It learns, for example, that tasks with significant domain-shift usually have a larger need for task-specific parameters for adaptation. We evaluate our model on Meta-dataset. Empirical results show that our model outperforms existing state-of-the-art methods. Our code is available at https://github.com/PRIS-CV/TA2-Net.", + "bbox": [ + 75, + 335, + 473, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 743, + 209, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional deep learning models [42, 36, 12, 17] show excellent generalization performance when training on a large number of labeled samples. However, both abundant samples and reliable annotations are not always available in realistic applications, e.g., rare disease diagnosis, and fine-grained recognition. Few-shot learning [32, 33, 23, 4, 10], inspired by the fact that humans can quickly learn new", + "bbox": [ + 75, + 768, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5f61152cc6709e9fbbfc88f6be0b037b9db09271266362c3116aafc075c2f8a.jpg", + "image_caption": [ + "(a) Different tasks from MSCOCO" + ], + "image_footnote": [], + "bbox": [ + 506, + 304, + 728, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab4c6f2ce3a7baae99c0ab79ad9c702417e93954cbaa278aace44d8d178cf7e7.jpg", + "image_caption": [ + "(c) Different layers in a task from MSCOCO" + ], + "image_footnote": [], + "bbox": [ + 733, + 306, + 880, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f010ea80287bfdd4011a9b1f033015060138d6a5c34b5e6733ff7d306230aac.jpg", + "image_caption": [ + "(b) Different tasks from ilsvrc 2012" + ], + "image_footnote": [], + "bbox": [ + 509, + 382, + 728, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/37227961c6896da2ff9661fa2241f02fcc369060230080f2ef39cf143b9f9c18.jpg", + "image_caption": [ + "(d) Different layers in a task from ilsvrc_2012" + ], + "image_footnote": [], + "bbox": [ + 733, + 386, + 880, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5eb0ed051a37115f28ea7fb9a4414fa12d667abe339c130eaca44efffb10750d.jpg", + "image_caption": [ + "Figure 1. (a) and (b) reveal the accuracy of varying test tasks when inference was performed in different models, in which TA-Modules are attached to each layer in the backbone network. (c) and (d) show the accuracy when TA-Modules are attached to different layers in the backbone network. Four dashed boxes show the backbone and several TA-Modules." + ], + "image_footnote": [], + "bbox": [ + 506, + 465, + 581, + 530 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/00d37dc4197337b48a856d0070ce01391596ca45c2482ff87e350ed8bd6f63d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 468, + 658, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f69f2fcd8b8e87168852ce4f74b0fc88398e0ca3e48ff677b600c28204acade1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 468, + 769, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f6721d53d122fccde922e3bb27316351ef6ca068aa5018c03e0cc119d1f5a38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 468, + 883, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "knowledge, aims to adapt the model to new classes by only a few labeled samples for each one.", + "bbox": [ + 496, + 642, + 890, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, meta-learning-based few-shot learning methods have made great strides in the setting where the train and test tasks are sampled from the same domain [49, 22]. However, more and more studies have demonstrated that these existing works fail to generalize well to novel classes that are heterogeneous with the source domain [45, 9, 14]. This limitation is attributable to the fact that most previous few-shot learning models focus solely on how to quickly adapt to a set of novel classes with only a few labeled samples per class but less effort has been made to understand and address the domain shift problem between the target (test) task and training domain. Recently, to address this problem, a series of research referred to as cross-domain few-shot learning are proposed [45, 14, 35, 3, 7, 20, 26, 6].", + "bbox": [ + 496, + 672, + 892, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generally, these existing methods address the domain", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 94, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1590", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shift problem by introducing a set of task-specific parameters to enable the feature extractor or classifier to adapt to the feature distributions of new tasks. According to the parameter generation method, these methods can be loosely divided into the following two main streams. One kind of approach is to use an auxiliary network [35, 3, 43] to generate task-specific parameters. The auxiliary network is first meta-trained with multiple tasks from the source domain, and then it utilizes the support set to estimate task-specific parameters for the target task. Other methods [21] directly attach a set of task-specific parameters to the pre-trained model on the source domain and then optimize them on a few labeled samples of the target task to well generalize the knowledge to new classes. Although cross-domain few-shot learning performance has improved, the fixed number of task-specific parameters may be too rigid to account for diverse tasks, resulting in less generality.", + "bbox": [ + 75, + 90, + 472, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, from the perspective of parameter size, we studied the performance of test tasks when inference was performed in models equipped with varying sizes of task-specific parameters. Specifically, following the trend of introducing directly optimizing task-specific parameters [21], we designed several Task-specific Adapters (TA-Modules) with increasing parameters, which are respectively attached to the conv of pre-trained backbone network (e.g., ResNet-18) to construct cross-domain models. As shown in Figure 1, the optimal task-specific parameters policy varies depending on the target task. More specifically, for different test tasks, the optimal task-specific parameters required in different layers of the backbone network for effectively adapting pre-learned knowledge to the target task's feature distribution differ. Abundant task-specific parameters may over-fit on target tasks, whereas few task-specific parameters will show the under-adaptation problem. In order to effectively direct the feature extractor to adapt feature distributions, we must carefully design task-specific parameters in each layer for the target task.", + "bbox": [ + 75, + 349, + 472, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the findings, for improving the recognition performance of cross-domain few-shot learning, we propose the Task-aware Adaptive Network (TA $^2$ -Net) to learn the optimal task-specific parameters policy for target tasks adaptively. As shown in Figure 2, our TA $^2$ -Net model consists of an Action Generation Network (\"Agent\") trained by Reinforcement Learning to generate \"actions\" - adaptive Task-specific Adapter execution decisions aimed at the target task, and an Adapted Network built on the \"actions\" to infer the target task, meanwhile, which can be viewed as the \"Environment\" to provide a \"reward\" for the Action Generation Network. The TA $^2$ -Net can learn optimally adapted network graphs aimed at the target task to transfer pre-learned knowledge efficiently.", + "bbox": [ + 75, + 654, + 472, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 96, + 868, + 328, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(i) We propose the Task-aware Adaptive Network (TA $^2$ -", + "bbox": [ + 96, + 883, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Net), which can adaptively learn optimal task-specific parameters policy for each target task.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(ii) We evaluate our model on the Meta-dataset. Empirical results show that our model obviously outperforms existing state-of-the-art methods.", + "(iii) We further analyze parameter distribution and discover that domains with significant distribution shift (compared with the source domain) usually have a larger demand for task-specific parameters to effectively adapt pre-learned knowledge. Furthermore, layers in deep blocks necessitate more task-specific parameters than those in shallow blocks for learning task-specific features." + ], + "bbox": [ + 498, + 121, + 893, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 286, + 640, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cross-domain Few-shot Learning Existing cross-domain few-shot learning methods can be coarsely grouped into two categories: feature-selection-based methods and adaptation-based methods.", + "bbox": [ + 498, + 311, + 890, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several feature-selection-based methods, including SUR [7], URT [26], and URL [20], aim to achieve generalization by adaptively integrating feature representations from multiple training domains. SUR [7] and URT [26] rely on attention mechanisms to select appropriate domain-specific representations for a given few-shot learning task. While these methods have shown good performance, they require multiple forward networks during inference time. In contrast, Li et al. [20] proposed URL, an efficient method that distills knowledge from multiple domain networks to a single feature extractor. However, supervised training of multiple domains can result in supervision collapse, leading to the loss of information required for transfer to new tasks or domains. To address this issue, Doersch et al. [6] proposed using self-supervised learning to encourage the learning of general features from multiple training domains.", + "bbox": [ + 496, + 372, + 892, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The goal of adaptation-based methods is to adapt a feature extractor and classifier learned from a source domain to a target task through task-specific parameters. CNAPS [35] employed an adaptation network to generate task-specific parameters for each hidden layer of the feature extractor and classifier to transfer knowledge. Simple CNAPS [3] improved upon this approach by replacing the parametric classifier with a Mahalanobis distance-based distance metric, which increased classification accuracy while reducing the number of parameters. Triantafillou et al.[43] proposed the FLUTE, which learns domain-specific parameters using FiLM layers called universal templates. Li et al.[21] have demonstrated that attaching task-specific weights to a pretrained model and optimizing them from scratch directly on a small support set can enable good generalization to the target task. However, the shifts of feature distribution between different target tasks and the pre-learned domain vary, leading to the change of task-specific parameters demand. In this paper, we focus on adaptively generating task-specific", + "bbox": [ + 496, + 614, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1591", + "bbox": [ + 483, + 945, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/903c83177ec845267659ab2d6306439b2a6aa747dae4cbb6980c93006391ab16.jpg", + "image_caption": [ + "Figure 2. Task-aware Adaptive Network (TA $^2$ -Net). The \"Agent\" is the Action Generation Network, which predicts the \"action\" of Task-specific Adapter execution and interacts with the \"Environment\" (Env, the Adapted Network) providing the \"reward\" for the Agent. \"none\" represent that the model does not need the Task-specific Adapter to adapt feature." + ], + "image_footnote": [], + "bbox": [ + 122, + 90, + 851, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "graphs for each target task, to adapt the prior knowledge to the target task well.", + "bbox": [ + 75, + 422, + 470, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reinforcement Learning Reinforcement learning [1, 8] is primarily used to solve the sequence decision problem, in which the reward value obtained from agent-environment interaction is used as a feedback signal to continuously optimize agent strategies. There are three main approaches to solving Reinforcement Learning (RL) problems: methods based on value functions [30, 46, 48], methods based on policy search [41, 40], and the actor-critic approach [15, 39]. The deep reinforcement learning algorithm based on value functions approximates the value function or action-value function with a deep neural network and updates it with the temporal difference (TD) learning or Q-learning method. Among them, Deep Q-Network (DQN) [30] algorithm is representative. Furthermore, a series of variants have been proposed, including double DQN (DDQN) [46], which optimizes the problem of overestimation about Q value in the DQN algorithm, and Dueling DQN [48], which improves DQN from the network structure and can estimate Q function more accurately. Policy-based methods [41, 40] directly search for an optimal policy without maintaining a value function model. Parameters of the chosen policy are updated using either gradient-based or gradient-free optimization to maximize the expected return. The actor-critic approach [15] employs both value functions and policy search to trade-off the variance reduction of policy gradients and bias introduction from value function methods, in which the \"actor\" (policy) learns by using feedback from the \"critic\" (value function) [1]. In this paper, we employ policy-gradient-based Reinforcement Learning to train", + "bbox": [ + 75, + 463, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the Action Generation Network to adaptively generate optimal actions – Task-specific Adapter execution decisions for each target task.", + "bbox": [ + 498, + 422, + 893, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 483, + 591, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the problem setting and introduce our method.", + "bbox": [ + 498, + 508, + 890, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 500, + 550, + 629, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-shot task Few-shot learning classification aims to classify a large number of samples with only a small number of labeled ones in each class. Labeled samples are generally referred to as the support set $S = \\{(\\pmb{x}_i, y_i)\\}_{i=1}^{N_S}$ , while classified ones are known as the query set $Q = \\{(\\pmb{x}_i)\\}_{i=1}^{N_Q}$ . $y_i$ is the label of the sample $\\pmb{x}_i$ . A few-shot task is created by combining the support and query sets.", + "bbox": [ + 496, + 575, + 890, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-domain few-shot learning Compared with traditional few-shot learning, cross-domain few-shot learning is more challenging, since not only the labeled samples in the target task $D_{t}$ are few, but also the distribution of the target domain and the base domain $D_{b}$ may be highly heterogeneous. Therefore, we must strike a balance between model capacity and the reliability of the adaptation while avoiding over-fitting to the few labeled samples or under-adaptation for the target task.", + "bbox": [ + 496, + 684, + 890, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Task-aware Adaptive Network (TA²-Net)", + "text_level": 1, + "bbox": [ + 500, + 830, + 849, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the observations that varying tasks require variant task-specific parameters for effective adaptation, we design the $\\mathrm{TA}^2$ -Net adaptively learning the optimally", + "bbox": [ + 498, + 854, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1592", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "adapted network graph for each target task. As shown in Figure 2, the $\\mathrm{TA}^2$ -Net consists of Action Generation Network and Adapted Network. To directly optimize the Action Generation Network to generate optimal task-specific parameters for the Adapted Network, we cast the model in Reinforcement Learning. In particular, Action Generation Network can be viewed as an \"Agent\" that interacts with the \"Environment\" (Adapted Network). The Action Generation Network will result in the \"actions\" that are the prediction of Task-specific Adapter execution for the target task. Based on the generated \"actions\", we can build the unique Adapted Network, which will provide a \"reward\" for the agent - we denote this reward by $r$ . We will introduce the Action Generation Network, Adapted Network, and their optimization process.", + "bbox": [ + 75, + 90, + 472, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Action Generation Network", + "text_level": 1, + "bbox": [ + 76, + 338, + 326, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our focus in the Action Generation Network is to generate adaptive Task-specific Adapter execution decisions for each layer of the Adapted Network aimed at the target task. As shown in Figure 2, the Gate Network is integrated into each layer of the feature encoder $f_{\\phi}(\\cdot)$ pre-trained by the seen domain to generate the discrete probability for $K$ kinds of Task-specific Adapters. Concretely, the discrete probability in $l$ -th layer can be formulated as $\\{z_l^i\\}_{i=1}^K$ . Let $h_{l-1} \\in R^{N_S \\times C \\times W \\times H}$ denotes the input feature of $l$ -th feature encoder. In the Gate Network, the input feature is first fed to the Global Average Pooling (GAP) to compress the spatial dimension. Formally, the output $u_{l-1}$ can be achieved as", + "bbox": [ + 75, + 362, + 472, + 549 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {u} _ {l - 1} = G A P \\left(\\boldsymbol {h} _ {l - 1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 563, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $u_{l-1}$ can be written as $u_{l-1} = \\left[u_{l-1}^{1}, u_{l-1}^{2}, \\dots, u_{l-1}^{N_{S}}\\right]$ . $u_{l-1}^{i}$ represents the channel-level feature embedding of support sample $i$ and $u_{l-1}^{i} \\in R^{C}$ . Then, let $v_{l-1} = \\left[v_{l-1}^{1}, v_{l-1}^{2}, \\dots, v_{l-1}^{N}\\right]$ denotes the class prototypes in the target task, where $v_{l-1}^{n}$ refers to the class center of class $n$ and $N$ denotes the number of classes in the target task. Formally, the class center for class $n$ can be obtained by computing the mean vector of all embedded features in the class as", + "bbox": [ + 76, + 595, + 468, + 737 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {l - 1} ^ {n} = \\frac {1}{| S _ {n} |} \\sum_ {i \\in S _ {n}} \\boldsymbol {u} _ {l - 1} ^ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 746, + 468, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{n}$ denotes the set of examples labeled with class $n$ . The output $\\boldsymbol{v}_{l-1}$ is a collection of all class descriptors with statistics that express the entire target task. To capture the dependencies between features, we add a simple non-linear function of two fully-connected layers coupled with a ReLU activation function. The output of this operation is a vector containing unnormalized scores for executing the $K$ kinds", + "bbox": [ + 75, + 794, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of Task-specific Adapters.", + "bbox": [ + 500, + 90, + 676, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {z} _ {l} = \\mathbf {W} _ {l} ^ {2} \\sigma \\left(\\mathbf {W} _ {l} ^ {1} v _ {l - 1}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 116, + 892, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{z}_l\\in R^K$ $\\sigma$ refers to the $ReLU$ activation function. $\\mathbf{W}_l^1\\in R^{d\\times C}$ $\\mathbf{W}_l^2\\in R^{K\\times d}$ and $d$ is the dimension of the hidden layer. We can write $\\bar{z}_l$ as $\\bar{z}_l = [\\bar{z}_l^1,\\bar{z}_l^2,\\dots ,\\bar{z}_l^K ]$ Then, the Gate outputs a distribution over the $l$ -th Task-specific Adapter execution decisions $\\pmb {z}_l$ using the Softmax function", + "bbox": [ + 496, + 145, + 893, + 236 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {\\boldsymbol {\\ell}} = \\operatorname {s o f t m a x} (\\bar {\\boldsymbol {z}} _ {\\boldsymbol {\\ell}}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 250, + 892, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{l} = \\left[z_{l}^{1},z_{l}^{2},\\dots ,z_{l}^{K}\\right]$ . $z_{l}^{k}$ denotes the possibility of selecting $k$ -th Task-specific Adapter in the $l$ -th layer. And $z = [z_{1},z_{2},\\dots ,z_{L}]$ , $L$ is the number of layers in the backbone.", + "bbox": [ + 496, + 275, + 893, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training The generation of Task-specific Adapter execution decisions is cast as the Reinforcement Learning problem. The Action Generation Network aims to decide the most appropriate Task-specific Adapters for conducting few-shot learning on the current tasks. We utilize the predicted result of the Adapted Network constructed based on the current \"actions\" as the rewards for optimizing the Action Generation Network. Inspired by the self-critical sequence training approach [34], the goal of training is to minimize the negative expected reward", + "bbox": [ + 496, + 337, + 893, + 488 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL (\\theta) = - \\left(r \\left(\\boldsymbol {z} ^ {s}\\right) - r \\left(\\boldsymbol {z} ^ {m}\\right)\\right) \\log \\left(\\boldsymbol {z} ^ {s}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 500, + 890, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta$ is the parameters of the Gate Network. $\\theta = \\{\\mathbf{W}_l^1,\\mathbf{W}_l^2\\}$ and $l\\in \\{1,2,\\dots ,L\\}$ . $z^{s} = [z_{1}^{s},\\dots,z_{L}^{s}]$ and $z_{l}^{s}$ is the possibility of action randomly sampled from the $K$ actions - Task-specific Adapters, at the layer $l$ . $r(z^{s})$ is the reward obtained by the Adapted Network, which is constructed by the actions with $z^{s}$ discrete possibility. $z^{m} = [z_{1}^{m},\\dots,z_{L}^{m}]$ , and", + "bbox": [ + 498, + 527, + 893, + 635 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {l} ^ {m} = \\operatorname {a r g m a x} \\left(\\boldsymbol {z} _ {l}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 643, + 890, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{argmax}(\\cdot)$ returns the possibility of action equipped with maximum probability over the output distribution $z_{l}$ of the Gate in $l$ -th layer. $r(z^{m})$ again is the reference reward obtained in Adapted Network equipped with actions with $z^{m}$ discrete possibility. Accordingly, TA modules from the model that return a higher reward than $z^{m}$ will be \"pushed up\", or increased in probability, while TA modules that result in a lower reward will be suppressed.", + "bbox": [ + 496, + 669, + 893, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Adapted Network", + "text_level": 1, + "bbox": [ + 500, + 801, + 679, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2, the network architecture of the Adapted Network consists of a series of residual blocks, which are formed by the Adaptive Module (AM) followed by the BN layer and ReLU function. In the Adaptive Module, a $3 \\times 3$ conv is paralleled with the Task-specific Adapter,", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1593", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Task-aware Adaptive Learning", + "text_level": 1, + "bbox": [ + 78, + 90, + 370, + 106 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Train the backbone network $f_{\\phi}(\\cdot)$ on the source domain;", + "2: Build Action Generation Network $M_{act}$ on the backbone network; Initialization the parameters $\\theta$ of Gate Networks;", + "3: while training do", + "4: Random sample task $T_{i}$ in the source domain; $T_{i} = S \\bigcup Q, S = \\{(x_{i},y_{i})\\}_{i = 1}^{N_{S}}, Q = \\{(x_{i})\\}_{i = 1}^{N_{Q}}$ ;", + "5: $\\pmb{z} = f_{\\phi, \\theta}(x), x \\in S, \\pmb{z} = [\\pmb{z}_1, \\pmb{z}_2, \\dots, \\pmb{z}_L]$ ;", + "6: $\\pmb{z}^s = [Sample(\\pmb{z}_1), \\dots, Sample(\\pmb{z}_L)]; \\pmb{z}^m = [argmax(\\pmb{z}_1), \\dots, argmax(\\pmb{z}_L)];$", + "7: Build Adapted Network $M_{adapt}^{1}$ , $M_{adapt}^{2}$ based on the backbone network according to actions $z^{s}$ , $z^{m}$ respectively;", + "8: Initialization $\\psi, \\pi$ in $M_{adapt}^{1}$ and $M_{adapt}^{2}$ ;", + "9: Optimize $\\psi$ , $\\pi$ over $S$ based on $L_{adapt}(\\psi, \\pi) = \\frac{1}{N_S} \\sum_{(x,y) \\in S} l(f_{\\phi,\\psi,\\pi}(\\boldsymbol{x}), y)$ ;", + "10: Predict over $Q$ ; $r(\\pmb{z}^s) = N_{adapt}^1 / N_Q$ , $r(\\pmb{z}^m) = N_{adapt}^2 / N_Q$ ;", + "11: Optimize the parameters $\\theta$ of $M_{act}$ based on $L(\\theta) = -(r(\\pmb{z}^s) - r(\\pmb{z}^m))\\log (\\pmb{z}^s)$ ;", + "12: end while" + ], + "bbox": [ + 81, + 108, + 468, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Action Generation Network $M_{act}$", + "bbox": [ + 78, + 446, + 364, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which is the action predicted by the corresponding Gate Network of the Action Generation Network.", + "bbox": [ + 75, + 491, + 468, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $f_{\\phi_l}(\\cdot)$ denote the $3 \\times 3$ conv in the $l$ -th Adaptive Module, which is pre-trained by the seen domain. With $h_{l-1} \\in R^{N_S \\times C \\times W \\times H}$ as the input feature of $l$ -th Adaptive Module, the output is $h_l = f_{\\phi_l}(h_{l-1})$ when no Task-specific Adapter is selected according to the decision generated by the Action Generation Network, that is the feature does not need adaptation. When the $k$ -th Task-specific Adapter is chosen, which is constructed by paralleling $k$ task-specific layers in our design, the feature learned by the Task-specific Adapter will be combined with the feature extracted from the $3 \\times 3$ conv as", + "bbox": [ + 75, + 521, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {h} _ {\\boldsymbol {l}} = f _ {\\phi_ {\\boldsymbol {l}}} \\left(\\boldsymbol {h} _ {\\boldsymbol {l} - \\mathbf {1}}\\right) + \\frac {1}{k} \\sum_ {i = 1} ^ {k} f _ {\\psi_ {\\boldsymbol {l}} ^ {i}} \\left(\\boldsymbol {h} _ {\\boldsymbol {l} - \\mathbf {1}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 696, + 468, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{\\psi_l^i}(\\cdot)$ denotes the $i$ -th task-specific layer in the Task-specific Adapter, which is $1 \\times 1$ conv in the experiments.", + "bbox": [ + 75, + 747, + 467, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Upon the generated \"actions\" with the discrete possibility $z^s$ and $z^m$ in the Action Generation Network, we will construct two Adapted Networks $M_{adapt}^{1}$ and $M_{adapt}^{2}$ . In training, the Adapted Networks are trained with the cross-entropy loss over the support samples:", + "bbox": [ + 76, + 777, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a d a p t} (\\psi , \\pi) = \\frac {1}{N _ {S}} \\sum_ {(x, y) \\in S} l \\left(f _ {\\phi , \\psi , \\pi} (\\boldsymbol {x}), y\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 864, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{\\psi, \\pi}(\\cdot)$ denotes the output softmax probability vector of the linear classifier with parameters $\\pi$ , in which dimensionality equals the number of categories in the support set $S$ . Then, the reward for the agent is achieved by predicting the Adapted Network over the query samples $Q$ :", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr \\left(\\boldsymbol {z} ^ {\\boldsymbol {s}}\\right) = \\frac {N _ {\\text {a d a p t}} ^ {1}}{N _ {Q}} \\quad \\text {a n d} \\quad r \\left(\\boldsymbol {z} ^ {\\boldsymbol {m}}\\right) = \\frac {N _ {\\text {a d a p t}} ^ {2}}{N _ {Q}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 176, + 890, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{adapt}^{1}$ and $N_{adapt}^{2}$ denote the number of query samples correctly predicted by Adapted Networks $M_{adapt}^{1}$ and $M_{adapt}^{2}$ respectively. We show the Task-aware Adaptive Learning pseudo code in Algorithm 1.", + "bbox": [ + 496, + 224, + 890, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Testing", + "text_level": 1, + "bbox": [ + 500, + 306, + 604, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the support samples of testing will first go through the Action Generation Network to generate optimal Task-specific Adapter executions for the target task, which are the action with maximum probability. And then, based on the executions, the Adapted Network will be built and finetuned over the support set. Finally, the query samples will be predicted by the Adapted Network.", + "bbox": [ + 496, + 330, + 890, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 450, + 630, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this part, we first introduce the experimental setup and then compare our method with the state-of-the-art methods. Furthermore, we provide a series of analyses of our method.", + "bbox": [ + 496, + 474, + 890, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 500, + 530, + 686, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. Meta-Dataset [44] is the standard benchmark for evaluating the performance of few-shot classification. It consists of 13 image datasets: ILSVRC 2012 [36], Omniglot [18], FGVC-Aircraft [29], CUB-200-2011 [47], Describable Textures [5], Quick Draw [11], FGVCx Fungi [38], VGG Flower [31], Traffic Signs [13], MSCOCO [24], MNIST [19], CIFAR-10 [16], and CIFAR-100 [16]. For all datasets in the experiments, we used the standard split procedure as [44].", + "bbox": [ + 496, + 551, + 890, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. In the experiments, we considered two training settings: multi-domain learning and single-domain setting. In the multi-domain setting, the first eight datasets (ILSVRC 2012, Omniglot, Aircraft, CUB-200-2011, Textures, Quick Draw, Fungi, and VGG Flower) of the Meta-dataset are used to train the feature extractor. We followed the method in [20] to obtain the pre-trained model. And in the single-domain setting, only the ILSVRC 2012 dataset participated in the training of the feature extractor. As in [20], we adopted the ResNet-18 as the backbone no matter in the multi-domain setting or the single-domain setting. In the training, the $\\mathrm{TA}^2$ -Net is trained using the Adam optimizer. The hyper-parameters $K$ are set as $K = 3$ . We also provide the ablation study in Section 4.4.", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1594", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/56b5acf6eccb73f9f3fc1caca74125a82ff9933730808f44653b838c298bd820.jpg", + "table_caption": [ + "Table 1. Comparison to state-of-the-art methods in the multi-domain setting. Mean accuracy, $95\\%$ confidence interval are reported. The best results are marked in bold." + ], + "table_footnote": [], + "table_body": "
Test DatasetCNAPS [35]Simple CNAPS [3]Transductive CNAPS [2]SUR [7]URT [26]FLUTE [43]tri-M [27]URL [20]TSA [21]Ours
ImageNet50.8 ± 1.158.4 ± 1.157.9 ± 1.156.2 ± 1.056.8 ± 1.158.6 ± 1.051.8 ± 1.158.8 ± 1.159.5 ± 1.059.6 ± 1.0
Omniglot91.7 ± 0.591.6 ± 0.694.3 ± 0.494.1 ± 0.494.2 ± 0.492.0 ± 0.693.2 ± 0.594.5 ± 0.494.9 ± 0.495.5 ± 0.4
Aircraft83.7 ± 0.682.0 ± 0.784.7 ± 0.585.5 ± 0.585.8 ± 0.582.8 ± 0.787.2 ± 0.589.4 ± 0.489.9 ± 0.490.5 ± 0.4
Birds73.6 ± 0.974.8 ± 0.978.8 ± 0.771.0 ± 1.076.2 ± 0.875.3 ± 0.879.2 ± 0.880.7 ± 0.881.1 ± 0.881.4 ± 0.8
Textures59.5 ± 0.768.8 ± 0.966.2 ± 0.871.0 ± 0.871.6 ± 0.771.2 ± 0.868.8 ± 0.877.2 ± 0.777.5 ± 0.777.4 ± 0.7
Quick Draw74.7 ± 0.876.5 ± 0.877.9 ± 0.681.8 ± 0.682.4 ± 0.677.3 ± 0.779.5 ± 0.782.5 ± 0.681.7 ± 0.682.5 ± 0.6
Fungi50.2 ± 1.146.6 ± 1.048.9 ± 1.264.3 ± 0.964.0 ± 1.048.5 ± 1.058.1 ± 1.168.1 ± 0.966.3 ± 0.866.3 ± 0.9
VGG Flower88.9 ± 0.590.5 ± 0.592.3 ± 0.482.9 ± 0.887.9 ± 0.690.5 ± 0.591.6 ± 0.692.0 ± 0.592.2 ± 0.592.6 ± 0.4
Traffic Sign56.5 ± 1.157.2 ± 1.059.7 ± 1.151.0 ± 1.148.2 ± 1.163.0 ± 1.058.4 ± 1.163.3 ± 1.182.8 ± 1.087.4 ± 0.8
MSCOCO39.4 ± 1.048.9 ± 1.142.5 ± 1.152.0 ± 1.151.5 ± 1.152.8 ± 1.150.0 ± 1.057.3 ± 1.057.6 ± 1.057.9 ± 0.9
MNIST-94.6 ± 0.494.7 ± 0.394.3 ± 0.490.6 ± 0.596.2 ± 0.395.6 ± 0.594.7 ± 0.496.7 ± 0.497.0 ± 0.4
CIFAR-10-74.9 ± 0.773.6 ± 0.766.5 ± 0.967.0 ± 0.875.4 ± 0.878.6 ± 0.774.2 ± 0.882.9 ± 0.782.1 ± 0.8
CIFAR-100-61.3 ± 1.161.8 ± 1.056.9 ± 1.157.3 ± 1.062.0 ± 1.067.1 ± 1.063.5 ± 1.070.4 ± 0.970.9 ± 0.9
Average Seen71.673.775.175.977.474.576.280.480.480.7
Average Unseen-67.466.564.162.969.969.970.678.179.1
Average All-71.271.871.471.872.773.876.679.580.1
Average Rank-7.66.27.26.55.85.53.22.21.2
", + "bbox": [ + 76, + 119, + 893, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison to state-of-the-art methods", + "text_level": 1, + "bbox": [ + 76, + 467, + 416, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the experiment set in [21], we sampled tasks with varying numbers of ways and shots in the inference and conducted experiments in the multi-domain setting and single-domain setting.", + "bbox": [ + 76, + 489, + 468, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-domain setting. Here, the results of our method on 13 datasets of the Meta-Dataset are compared with the state-of-the-art methods. We also report average accuracy over seen domains, and unseen domains, as well as the average rank as in [21]. From the results in Table 1, our method obtains state-of-the-art results on 10 out of 13 datasets and achieves the best average accuracy on seen and unseen domains. It is worth noting that our method obtains significantly better average results than the second-best approach on the unseen datasets $(+1.0\\%)$ . The unseen datasets are heterogeneous with respect to the training domain, with a significant domain gap between them. In contrast to adopting the fixed number of task-specific parameters for all target tasks, e.g. CNAPS, Simple CNAPS, FLUTE, and TSA, the $\\mathrm{TA}^2$ -Net can adaptively learn the optimal task-specific parameter policies for unseen target tasks to efficiently transfer the pre-trained knowledge to the target feature distribution.", + "bbox": [ + 75, + 551, + 470, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Single-domain setting. We also evaluate our method in the single-domain setting, which is more challenging due to optimizing the Action Generation Network only on the ImageNet dataset. Results are shown in Table 2. First, compared with other single-domain training methods, our", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "method achieves significant superiority. For example, $\\mathrm{TA}^2$ -Net outperforms the most competing method (TSA [21]) by $+2.9\\%$ , $+0.8\\%$ , $+1.6\\%$ , and $+1\\%$ on Omniglot, Quick Draw, Traffic Signs and MNIST, respectively. Second, it is worth noting that, for the average accuracy on the unseen datasets, ours outperforms the second-best approach with a clear margin $(+3.1\\%)$ . And we clearly observe that $\\mathrm{TA}^2$ -Net achieves competitive accuracy on the ImageNet, compared with the best performance. Despite being trained on just one dataset, the Action Generation Network exhibits good generalization, producing the optimal task-specific parameter policies for the unseen domain in an excellently adaptive manner.", + "bbox": [ + 496, + 467, + 893, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Further Analysis", + "text_level": 1, + "bbox": [ + 500, + 678, + 668, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Varying-way Five-shot setting. The number of support samples ('shot') in the standard Meta-Dataset evaluation setting is variable and can range up to 100. In this section, we evaluate our method in the 5-shot setting which is more challenging due to fewer labeled samples and can better reflect the progressiveness of the model. Table 3 shows the results compared with other state-of-the-art methods. As can be seen, ours achieves better average accuracy on both seen and unseen domains. More specifically, our method outperforms other methods on the vast majority of domains (8 out of 13) and obtains the best average rank. Furthermore, the average accuracy on unseen datasets is $1.2\\%$ higher than the second-best method.", + "bbox": [ + 496, + 703, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1595", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6991f1dd1b4c750db17f5db984df17cb3f74542a21adf85b3d174517eeaf2a99.jpg", + "table_caption": [ + "Table 2. Comparison to state-of-the-art methods in the single-domain setting. Mean accuracy, $95\\%$ confidence interval are reported. The best results are marked in bold." + ], + "table_footnote": [], + "table_body": "
Test DatasetProtoNet [44]fo-Proto -MAML [44]ALFA+fo-Proto -MAML [44]BOHB [37]FLUTE [43]TSA [21]Ours
ImageNet50.5 ± 1.149.5 ± 1.152.8 ± 1.151.9 ± 1.146.9 ± 1.159.5 ± 1.159.3 ± 1.1
Omniglot60.0 ± 1.463.4 ± 1.361.9 ± 1.567.6 ± 1.261.6 ± 1.478.2 ± 1.281.1 ± 1.1
Aircraft53.1 ± 1.056.0 ± 1.063.4 ± 1.154.1 ± 0.948.5 ± 1.072.2 ± 1.072.6 ± 0.9
Birds68.8 ± 1.068.7 ± 1.069.8 ± 1.170.7 ± 0.947.9 ± 1.074.9 ± 0.975.1 ± 0.9
Textures66.6 ± 0.866.5 ± 0.870.8 ± 0.968.3 ± 0.863.8 ± 0.877.3 ± 0.776.8 ± 0.8
Quick Draw49.0 ± 1.151.5 ± 1.059.2 ± 1.250.3 ± 1.057.5 ± 1.067.6 ± 0.968.4 ± 0.9
Fungi39.7 ± 1.140.0 ± 1.141.5 ± 1.241.4 ± 1.131.8 ± 1.044.7 ± 1.045.3 ± 1.0
VGG Flower85.3 ± 0.887.2 ± 0.786.0 ± 0.887.3 ± 0.680.1 ± 0.990.9 ± 0.691.0 ± 0.6
Traffic Sign47.1 ± 1.148.8 ± 1.160.8 ± 1.351.8 ± 1.046.5 ± 1.182.5 ± 0.884.1 ± 0.7
MSCOCO41.0 ± 1.143.7 ± 1.148.1 ± 1.148.0 ± 1.041.4 ± 1.059.0 ± 1.058.0 ± 1.0
MNIST----80.8 ± 0.893.9 ± 0.694.9 ± 0.5
CIFAR-10----65.4 ± 0.882.1 ± 0.782.0 ± 0.7
CIFAR-100----52.7 ± 1.170.7 ± 0.970.8 ± 0.9
Average Seen50.549.552.851.946.959.559.3
Average Unseen56.758.462.460.053.271.975.0
Average All56.157.561.459.252.670.773.8
Average Rank5.95.23.44.15.61.71.3
", + "bbox": [ + 81, + 132, + 521, + 351 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b57a8cb11bf2b59d7109433dfd1dd3ce5f1e80ff36855cb1a58625f6ff2b1757.jpg", + "image_caption": [ + "Figure 3. Performance of widening or deepening the Task-specific Adapter (TA) in the multi-domain setting. The axes indicate the accuracy of methods on a particular dataset. Each color corresponds to a different method." + ], + "image_footnote": [], + "bbox": [ + 117, + 369, + 429, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Wider or deeper in the Task-specific Adapter (TA)? To explore the method for improving the adaptation of the model, we increase the parameters of the Task-specific Adapter (TA) in wide or depth respectively. Specifically, two types of TAs are designed: (a) $1 \\times 1$ Conv in parallel, (b) $1 \\times 1$ Conv in series. As shown in Figure 3, increasing the width of the TA allows the model to achieve better performance than deepening. In particular, among the 13 datasets, 5 datasets (Traffic Signs, MSCOCO, MNIST, CIFAR-10, and CIFAR-100) unseen in the training obviously benefited more from widening the Task-specific Adapter (TA). All of the above indicates that the best way to improve adaptability is to increase the width of the Task-specific Adapter (TA).", + "bbox": [ + 75, + 641, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of adaptive learning. In this section, we attempt to provide a more in-depth discussion of the efficacy of fixed networks as well as the proposed adaptive learned network. Figure 4 depicts the performance over the", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/48fba64b06e5a88e7081add47079e66c20c35b79d3451d76e4e2f0edcd8398f2.jpg", + "table_caption": [ + "Table 3. Results for Varying-Way Five-Shot setting. Mean accuracy, $95\\%$ confidence interval are reported. The best results are marked in bold." + ], + "table_footnote": [], + "table_body": "
Test DatasetSimple CNAPS [3]SUR [7]URT [26]URL [20]TSA [21]Ours
ImageNet47.2 ± 1.046.7 ± 1.048.6 ± 1.049.4 ± 1.048.3 ± 1.049.3 ± 1.0
Omniglot95.1 ± 0.395.8 ± 0.396.0 ± 0.396.0 ± 0.396.8 ± 0.396.6 ± 0.2
Aircraft74.6 ± 0.682.1 ± 0.681.2 ± 0.684.8 ± 0.585.5 ± 0.585.9 ± 0.4
Birds69.6 ± 0.762.8 ± 0.971.2 ± 0.776.0 ± 0.676.6 ± 0.677.3 ± 0.6
Textures57.5 ± 0.760.2 ± 0.765.2 ± 0.769.1 ± 0.668.3 ± 0.768.3 ± 0.6
Quick Draw70.9 ± 0.679.0 ± 0.579.2 ± 0.578.2 ± 0.577.9 ± 0.678.5 ± 0.5
Fungi50.3 ± 1.066.5 ± 0.866.9 ± 0.970.0 ± 0.870.4 ± 0.870.3 ± 0.8
VGG Flower86.5 ± 0.476.9 ± 0.682.4 ± 0.589.3 ± 0.489.5 ± 0.490.0 ± 0.4
Traffic Sign55.2 ± 0.844.9 ± 0.945.1 ± 0.957.5 ± 0.872.3 ± 0.676.7 ± 0.5
MSCOCO49.2 ± 0.848.1 ± 0.952.3 ± 0.956.1 ± 0.856.0 ± 0.856.0 ± 0.8
MNIST88.9 ± 0.490.1 ± 0.486.5 ± 0.589.7 ± 0.492.5 ± 0.493.3 ± 0.3
CIFAR-1066.1 ± 0.750.3 ± 1.061.4 ± 0.766.0 ± 0.772.0 ± 0.773.1 ± 0.7
CIFAR-10053.8 ± 0.946.4 ± 0.952.5 ± 0.957.0 ± 0.964.1 ± 0.864.1 ± 0.8
Average Seen69.071.273.876.676.777.0
Average Unseen62.656.059.665.271.472.6
Average All66.565.468.372.274.675.3
Average Rank5.254.12.82.21.5
", + "bbox": [ + 532, + 131, + 890, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f3ce285afcd3d5a16cf39dcd2bb0f416de8fe60500b3a5195099c0604f3d60f.jpg", + "image_caption": [ + "Figure 4. Effectiveness of adaptive learning." + ], + "image_footnote": [], + "bbox": [ + 542, + 369, + 851, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Meta-dataset in the multi-domain and single-domain setting, with Model-1/2/3 constructed by attaching Module-1/2/3 respectively as shown in Figure 1 to the backbone. It can be observed that, in comparison to the backbone network, the average accuracy of seen domains in both the multi-domain and single-domain settings decreased when using the fixed Task-specific Adapter to adaptation, but our method performed better. Furthermore, average accuracy on unseen domains has consistently improved overall in the fixed Task-specific Adapter and can be improved further through adaptive learning. This discovery inspires us to recognize that simply designing a fixed Task-specific Adapter does not work and forces us to recognize the necessity of adaptively learning optimal task-specific parameters policy for each target task.", + "bbox": [ + 496, + 625, + 890, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Reinforcement Learning. Our $\\mathrm{TA}^2$ -Net model adopts the Reinforcement Learning algorithm combined with Episodic Training to adaptively learn opti", + "bbox": [ + 496, + 854, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1596", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9bfdfb265709122d04a4792b62a6440a9c6df301b20ae31fc435066e90d430d1.jpg", + "image_caption": [ + "(a) MMD and TA parameters of different datasets in the Meta-Dataset" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 421, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dbf31d7b4fad8267be1b2013ca0b0319d2c49f5d90f327647149a1ed29016bcf.jpg", + "image_caption": [ + "(b) TA parameters distribution in different layers for the Fungi dataset" + ], + "image_footnote": [], + "bbox": [ + 436, + 90, + 658, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4cc7c970ac9a734ebf14b9c1d9c46c20d0af5d9edce2582e38dd5eaff4a0927d.jpg", + "image_caption": [ + "(c) TA parameters distribution in different layers for the CIFAR-10 dataset", + "Figure 5. (a): The Line chart shows the mean and standard deviation of TA parameters adaptively learned by the $\\mathrm{TA}^2$ -Net in target tasks sampled in different datasets. And the box plot shows the $25\\%$ quantiles of the MMD in different datasets while the notches indicate the median. (b) and (c): TA parameters distribution in different blocks of the task-aware adapted network for the Fungi and CIFAR-10 dataset. Sub-bars in the block denote different layers within the Block." + ], + "image_footnote": [], + "bbox": [ + 666, + 90, + 883, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mal task-specific parameters policy for target tasks. To explore the effectiveness of Reinforcement Learning, we designed the Simple $\\mathrm{TA}^2$ -Net trained by Episodic Training, which only includes the Adapted Network attaching a Gate Network in each layer for generating \"decisions\" - Chosen the Task-specific Adapter. As shown in Table 4, the $\\mathrm{TA}^2$ -Net optimized with the Reinforcement Learning outperforms the Simple $\\mathrm{TA}^2$ -Net, which shows that Reinforcement Learning can better learn the TA decisions for target tasks.", + "bbox": [ + 75, + 330, + 468, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Parameters distribution. This is an intuitive guess that the greater the difference (domain shift) between the source domain and the target domain features, the greater the demand for TA parameters. In this section, we first observe the relationship between the TA parameters and the domain shift. As widely used in previous works [28, 25], we adopt the Maximum Mean Discrepancy (MMD) to quantify the domain shift between the target task features and the source domain features extracted by the pre-trained backbone (in the multi-domain setting). The larger MMD indicates a larger domain shift between the test task and the source domain. As shown in Figure 5 (a), the Line chart and box plot shows the TA parameters and MMD distribution in different datasets, respectively. We can find that MMD for each dataset has a narrow distribution. And, on the whole, datasets with significant MMD have a larger demand for the TA parameters. Furthermore, linear regression and Pearson correlation analysis revealed that MMD is positively correlated with TA parameters, with the Pearson correlation coefficient $R = 0.9186$ . It is attributed to the greater heterogeneity, more task-specific parameters are required to effectively adapt pre-learned knowledge to the feature distribution of target tasks. We further explore the TA parameters distribution in different layers of the task-aware adapted network learned by our proposed TA $^2$ -Net. As shown in Figure 5 (b) and (c), layers in the deep blocks have more task-specific parameters for learning the task-specific fea", + "bbox": [ + 75, + 493, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ed222538aeb9b72f75aa1bb54336865d25749caa53c62e25372ed35142bf46d4.jpg", + "table_caption": [ + "Table 4. Effectiveness of Reinforcement Learning in the multidomain setting." + ], + "table_footnote": [], + "table_body": "
ModelReinforcement LearningSeen DomainUnseen DomainAll
Simple TA2-NetX75.567.972.6
TA2-Net✓77.072.675.3
", + "bbox": [ + 519, + 361, + 874, + 434 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b733b0b82c9cae51f4af305a93bd43b6e6086b95cc80c104f371b88065bdad18.jpg", + "table_caption": [ + "Table 5. Mean accuracy $(\\%)$ for different number of layers $K$ in the Task-specific Adapter." + ], + "table_footnote": [], + "table_body": "
K=1K=2K=3K=4
Seen Domain75.475.275.375.1
Unseen Domain66.670.370.568.2
All72.073.373.572.5
", + "bbox": [ + 539, + 481, + 854, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tures than those in the shallow blocks as demonstrated by both Fungi and CIFAR-10, which may be because the shallow layer contains more general information while the deep layer contains more domain-specific information [50].", + "bbox": [ + 496, + 578, + 890, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 650, + 653, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Maximum width $K$ of the Task-specific Adapter. Here, we discuss the maximum width $K$ of the Task-specific Adapter, which represents the upper bound of the task-specific parameters. Experiments are conducted on the Meta-dataset in the varying-way five-shot setting. Table 5 shows the performances of $K \\in [1,2,3,4]$ . We can observe that, as $K$ increases, the average classification accuracy on seen domains lightly decreases. For unseen domains, the average classification accuracy increases at first and then decreases as $K$ increases, with $K = 3$ producing the best results. Compared with seen domains, unseen domains are heterogeneous with the source domain and require more task-specific parameters to transfer the pre-learned knowledge. The average performance becomes poorer when using $K = 4$ , which may be due to too much TA, the model", + "bbox": [ + 496, + 672, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1597", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "overfitting, or the search space of the model structure expanding, increasing the optimization difficulty of reinforcement learning. Considering the performance of various target tasks and the calculating costs, all experiments in this paper adopt $K = 3$ .", + "bbox": [ + 75, + 90, + 472, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 180, + 196, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, for studying cross-domain few-shot learning problems, we explored the performance of test tasks when inference was performed in models equipped with a variant number of task-specific parameters. The exploration revealed that target tasks require variant task-specific parameters for outstanding adaptation. Motivated by this, we proposed a Task-aware Adaptive Network (TA $^2$ -Net), which can adaptively learn optimal task-specific parameters policy for each target task for exhaustively improving the recognition performance of cross-domain few-shot learning. By conducting extensive experiments, we proved that our TA $^2$ -Net could achieve the best results on Meta-dataset.", + "bbox": [ + 75, + 205, + 472, + 387 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 76, + 401, + 228, + 419 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by Beijing Natural Science Foundation Project No. Z200002, in part by National Natural Science Foundation of China (NSFC) No. U19B2036, 62225601, in part by the Program for Youth Innovative Research Team of BUPT No. 2023QNTD02, in part by scholarships from China Scholarship Council (CSC) under Grant CSC No. 202206470055, in part by BUPT Excellent Ph.D. Students Foundation No. CX2022152.", + "bbox": [ + 75, + 426, + 468, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 76, + 561, + 174, + 577 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kai Arulkumaran, Marc Peter Deisenroth, Miles Brundage, and Anil Anthony Bharath. Deep reinforcement learning: A brief survey. IEEE Signal Processing Magazine, 34(6):26-38, 2017. 3", + "[2] Peyman Bateni, Jarred Barber, Jan-Willem van de Meent, and Frank Wood. Enhancing few-shot image classification with unlabelled examples. In WACV, pages 2796-2805, 2022. 6", + "[3] Peyman Bateni, Raghav Goyal, Vaden Masrani, Frank Wood, and Leonid Signal. Improved few-shot visual classification. In CVPR, pages 14493-14502, 2020. 1, 2, 6, 7", + "[4] Luca Bertinetto, Joao F. Henriques, Philip Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In ICLR, 2019. 1", + "[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5", + "[6] Carl Doersch, Ankush Gupta, and Andrew Zisserman. CrossTransformers: spatially-aware few-shot transfer. NeurIPS, 33:21981-21993, 2020. 1, 2", + "[7] Nikita Dvornik, Cordelia Schmid, and Julien Mairal. Selecting relevant features from a multi-domain representation for" + ], + "bbox": [ + 86, + 587, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "few-shot classification. In ECCV, pages 769-786, 2020. 1, 2, 6, 7", + "[8] Sven Gronauer and Klaus Diepold. Multi-agent deep reinforcement learning: a survey. Artificial Intelligence Review, 55(2):895-943, 2022. 3", + "[9] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In ECCV, pages 124-141, 2020. 1", + "[10] Yurong Guo, Ruoyi Du, Xiaoxu Li, Jiyang Xie, Zhanyu Ma, and Yuan Dong. Learning calibrated class centers for few-shot classification by pair-wise similarity. IEEE TIP, 31:4543-4555, 2022. 1", + "[11] David Ha and Douglas Eck. A Neural Representation of Sketch Drawings. arXiv preprint arXiv:1704.03477, 2017. 5", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1", + "[13] Sebastian Houben, Johannes Stallkamp, Jan Salmen, Marc Schlipsing, and Christian Igel. Detection of traffic signs in real-world images: The german traffic sign detection benchmark. In IJCNN, pages 1-8, 2013. 5", + "[14] Yanxu Hu and Andy J. Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, pages 20-37, 2022. 1", + "[15] Vijay Konda and John Tsitsiklis. Actor-critic algorithms. NeurIPS, 12, 1999. 3", + "[16] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5", + "[17] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 1", + "[18] Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 350(6266):1332-1338, 2015. 5", + "[19] Yann LeCun, Corinna Cortes, and CJ Burges. MNIST handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, 2, 2010. 5", + "[20] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple domains for few-shot classification. In ICCV, pages 9526-9535, 2021. 1, 2, 5, 6, 7", + "[21] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Cross-domain few-shot learning with task-specific adapters. In CVPR, pages 7161-7170, 2022. 2, 6, 7", + "[22] Xiaoxu Li, Jijie Wu, Zhuo Sun, Zhanyu Ma, Jie Cao, and Jing-Hao Xue. BSNet: Bi-similarity network for few-shot fine-grained image classification. IEEE TIP, 30:1318-1331, 2020. 1", + "[23] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. MetaSGD: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 1", + "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 5" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1598", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Ge Liu, Linglan Zhao, and Xiangzhong Fang. PDA: Proxy-based domain adaptation for few-shot image recognition. Image and Vision Computing, 110:104164, 2021. 8", + "[26] Lu Liu, William Hamilton, Guodong Long, Jing Jiang, and Hugo Larochelle. A universal representation transformer layer for few-shot image classification. In ICLR, 2021. 1, 2, 6, 7", + "[27] Yanbin Liu, Juho Lee, Linchao Zhu, Ling Chen, Humphrey Shi, and Yi Yang. A multi-mode modulator for multi-domain few-shot classification. In ICCV, pages 8453-8462, 2021. 6", + "[28] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105, 2015. 8", + "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5", + "[30] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. 3", + "[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5", + "[32] Hang Qi, Matthew Brown, and David G. Lowe. Low-Shot learning with imprinted weights. In CVPR, pages 5822-5830, 2018. 1", + "[33] Sachin Ravi and H. Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 1", + "[34] Steven J Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel. Self-critical sequence training for image captioning. In CVPR, pages 7008-7024, 2017. 4", + "[35] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, and Richard E Turner. Fast and flexible multi-task classification using conditional neural adaptive processes. NeurIPS, 32, 2019. 1, 2, 6", + "[36] Olga Russakovsky, J. Deng, Hao Su, J. Krause, S. Satheesh, S. Ma, Zhiheng Huang, A. Karpathy, A. Khosla, Michael S. Bernstein, A. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 1, 5", + "[37] Tonmoy Saikia, Thomas Brox, and Cordelia Schmid. Optimized generic feature learning for few-shot classification across domains. arXiv preprint arXiv:2001.07926, 2020. 7", + "[38] Brigit Schroeder and Yin Cui. FGVCx Fungi Classification Challenge 2018. github.com/visipedia/fgvcx_fungi_comp, 2018.5", + "[39] John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438, 2015. 3", + "[40] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3", + "[41] David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. In ICML, pages 387-395, 2014. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1", + "[43] Eleni Triantafillou, Hugo Larochelle, Richard Zemel, and Vincent Dumoulin. Learning a universal template for few-shot dataset generalization. In ICML, pages 10424–10433, 2021. 2, 6, 7", + "[44] Eleni Triantafillou, Tyler Zhu, Vincent Dumoulin, Pascal Lamblin, Utku Evci, Kelvin Xu, Ross Goroshin, Carles Gelada, Kevin Swersky, Pierre-Antoine Manzagol, et al. Meta-dataset: A dataset of datasets for learning to learn from few examples. arXiv preprint arXiv:1903.03096, 2019. 5, 7", + "[45] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In ICLR, 2020. 1", + "[46] Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-learning. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016. 3", + "[47] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5", + "[48] Ziyu Wang, Tom Schaul, Matteo Hessel, Hado Hasselt, Marc Lanctot, and Nando Freitas. Dueling network architectures for deep reinforcement learning. In ICML, pages 1995–2003, 2016. 3", + "[49] Davis Wertheimer, Luming Tang, and Bharath Hariharan. Few-shot classification with feature map reconstruction networks. In CVPR, pages 8012-8021, 2021. 1", + "[50] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 8" + ], + "bbox": [ + 501, + 92, + 890, + 544 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1599", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/d73ef7d6-b7de-4511-b8bf-4b4e20276122_model.json b/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/d73ef7d6-b7de-4511-b8bf-4b4e20276122_model.json new file mode 100644 index 0000000000000000000000000000000000000000..35267610a9d3f43fc96a258b2f2e3c306151fbb7 --- /dev/null +++ b/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/d73ef7d6-b7de-4511-b8bf-4b4e20276122_model.json @@ -0,0 +1,2321 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.131, + 0.833, + 0.154 + ], + "angle": 0, + "content": "Task-aware Adaptive Learning for Cross-domain Few-shot Learning" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.18, + 0.848, + 0.234 + ], + "angle": 0, + "content": "Yurong Guo\\(^{1}\\), Ruoyi Du\\(^{1}\\), Yuan Dong\\(^{1}\\), Timothy Hospedales\\(^{2}\\), Yi-Zhe Song\\(^{3}\\), Zhanyu Ma\\(^{1*}\\) \n\\(^{1}\\)Beijing University of Posts and Telecommunications, China \n\\(^{2}\\)University of Edinburgh, UK \n\\(^{3}\\)University of Surrey, UK" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.237, + 0.719, + 0.27 + ], + "angle": 0, + "content": "{guoyurong, duruoyi, yuandong, mazhanyu}@bupt.edu.cn, t.hospedales@ed.ac.uk, y.song@surrey.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.474, + 0.714 + ], + "angle": 0, + "content": "Although existing few-shot learning works yield promising results for in-domain queries, they still suffer from weak cross-domain generalization. Limited support data requires effective knowledge transfer, but domain-shift makes this harder. Towards this emerging challenge, researchers improved adaptation by introducing task-specific parameters, which are directly optimized and estimated for each task. However, adding a fixed number of additional parameters fails to consider the diverse domain shifts between target tasks and the source domain, limiting efficacy. In this paper, we first observe the dependence of task-specific parameter configuration on the target task. Abundant task-specific parameters may over-fit, and insufficient task-specific parameters may result in under-adaptation – but the optimal task-specific configuration varies for different test tasks. Based on these findings, we propose the Task-aware Adaptive Network (TA\\(^2\\)-Net), which is trained by reinforcement learning to adaptively estimate the optimal task-specific parameter configuration for each test task. It learns, for example, that tasks with significant domain-shift usually have a larger need for task-specific parameters for adaptation. We evaluate our model on Meta-dataset. Empirical results show that our model outperforms existing state-of-the-art methods. Our code is available at https://github.com/PRIS-CV/TA2-Net." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.744, + 0.21, + 0.76 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Traditional deep learning models [42, 36, 12, 17] show excellent generalization performance when training on a large number of labeled samples. However, both abundant samples and reliable annotations are not always available in realistic applications, e.g., rare disease diagnosis, and fine-grained recognition. Few-shot learning [32, 33, 23, 4, 10], inspired by the fact that humans can quickly learn new" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.305, + 0.73, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.376, + 0.684, + 0.383 + ], + "angle": 0, + "content": "(a) Different tasks from MSCOCO" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.307, + 0.881, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.376, + 0.881, + 0.384 + ], + "angle": 0, + "content": "(c) Different layers in a task from MSCOCO" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.383, + 0.729, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.454, + 0.686, + 0.461 + ], + "angle": 0, + "content": "(b) Different tasks from ilsvrc 2012" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.387, + 0.881, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.454, + 0.884, + 0.462 + ], + "angle": 0, + "content": "(d) Different layers in a task from ilsvrc_2012" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.467, + 0.583, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.469, + 0.66, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.469, + 0.771, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.469, + 0.884, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.625 + ], + "angle": 0, + "content": "Figure 1. (a) and (b) reveal the accuracy of varying test tasks when inference was performed in different models, in which TA-Modules are attached to each layer in the backbone network. (c) and (d) show the accuracy when TA-Modules are attached to different layers in the backbone network. Four dashed boxes show the backbone and several TA-Modules." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.673 + ], + "angle": 0, + "content": "knowledge, aims to adapt the model to new classes by only a few labeled samples for each one." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.893, + 0.885 + ], + "angle": 0, + "content": "Recently, meta-learning-based few-shot learning methods have made great strides in the setting where the train and test tasks are sampled from the same domain [49, 22]. However, more and more studies have demonstrated that these existing works fail to generalize well to novel classes that are heterogeneous with the source domain [45, 9, 14]. This limitation is attributable to the fact that most previous few-shot learning models focus solely on how to quickly adapt to a set of novel classes with only a few labeled samples per class but less effort has been made to understand and address the domain shift problem between the target (test) task and training domain. Recently, to address this problem, a series of research referred to as cross-domain few-shot learning are proposed [45, 14, 35, 3, 7, 20, 26, 6]." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Generally, these existing methods address the domain" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.225, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1590" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.349 + ], + "angle": 0, + "content": "shift problem by introducing a set of task-specific parameters to enable the feature extractor or classifier to adapt to the feature distributions of new tasks. According to the parameter generation method, these methods can be loosely divided into the following two main streams. One kind of approach is to use an auxiliary network [35, 3, 43] to generate task-specific parameters. The auxiliary network is first meta-trained with multiple tasks from the source domain, and then it utilizes the support set to estimate task-specific parameters for the target task. Other methods [21] directly attach a set of task-specific parameters to the pre-trained model on the source domain and then optimize them on a few labeled samples of the target task to well generalize the knowledge to new classes. Although cross-domain few-shot learning performance has improved, the fixed number of task-specific parameters may be too rigid to account for diverse tasks, resulting in less generality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.473, + 0.653 + ], + "angle": 0, + "content": "In this paper, from the perspective of parameter size, we studied the performance of test tasks when inference was performed in models equipped with varying sizes of task-specific parameters. Specifically, following the trend of introducing directly optimizing task-specific parameters [21], we designed several Task-specific Adapters (TA-Modules) with increasing parameters, which are respectively attached to the conv of pre-trained backbone network (e.g., ResNet-18) to construct cross-domain models. As shown in Figure 1, the optimal task-specific parameters policy varies depending on the target task. More specifically, for different test tasks, the optimal task-specific parameters required in different layers of the backbone network for effectively adapting pre-learned knowledge to the target task's feature distribution differ. Abundant task-specific parameters may over-fit on target tasks, whereas few task-specific parameters will show the under-adaptation problem. In order to effectively direct the feature extractor to adapt feature distributions, we must carefully design task-specific parameters in each layer for the target task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.473, + 0.867 + ], + "angle": 0, + "content": "Motivated by the findings, for improving the recognition performance of cross-domain few-shot learning, we propose the Task-aware Adaptive Network (TA\\(^2\\)-Net) to learn the optimal task-specific parameters policy for target tasks adaptively. As shown in Figure 2, our TA\\(^2\\)-Net model consists of an Action Generation Network (\"Agent\") trained by Reinforcement Learning to generate \"actions\" - adaptive Task-specific Adapter execution decisions aimed at the target task, and an Adapted Network built on the \"actions\" to infer the target task, meanwhile, which can be viewed as the \"Environment\" to provide a \"reward\" for the Action Generation Network. The TA\\(^2\\)-Net can learn optimally adapted network graphs aimed at the target task to transfer pre-learned knowledge efficiently." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.869, + 0.329, + 0.883 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.885, + 0.47, + 0.902 + ], + "angle": 0, + "content": "(i) We propose the Task-aware Adaptive Network (TA\\(^2\\)-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Net), which can adaptively learn optimal task-specific parameters policy for each target task." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.122, + 0.892, + 0.167 + ], + "angle": 0, + "content": "(ii) We evaluate our model on the Meta-dataset. Empirical results show that our model obviously outperforms existing state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.168, + 0.894, + 0.273 + ], + "angle": 0, + "content": "(iii) We further analyze parameter distribution and discover that domains with significant distribution shift (compared with the source domain) usually have a larger demand for task-specific parameters to effectively adapt pre-learned knowledge. Furthermore, layers in deep blocks necessitate more task-specific parameters than those in shallow blocks for learning task-specific features." + }, + { + "type": "list", + "bbox": [ + 0.499, + 0.122, + 0.894, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.287, + 0.642, + 0.303 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.313, + 0.892, + 0.372 + ], + "angle": 0, + "content": "Cross-domain Few-shot Learning Existing cross-domain few-shot learning methods can be coarsely grouped into two categories: feature-selection-based methods and adaptation-based methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.893, + 0.614 + ], + "angle": 0, + "content": "Several feature-selection-based methods, including SUR [7], URT [26], and URL [20], aim to achieve generalization by adaptively integrating feature representations from multiple training domains. SUR [7] and URT [26] rely on attention mechanisms to select appropriate domain-specific representations for a given few-shot learning task. While these methods have shown good performance, they require multiple forward networks during inference time. In contrast, Li et al. [20] proposed URL, an efficient method that distills knowledge from multiple domain networks to a single feature extractor. However, supervised training of multiple domains can result in supervision collapse, leading to the loss of information required for transfer to new tasks or domains. To address this issue, Doersch et al. [6] proposed using self-supervised learning to encourage the learning of general features from multiple training domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The goal of adaptation-based methods is to adapt a feature extractor and classifier learned from a source domain to a target task through task-specific parameters. CNAPS [35] employed an adaptation network to generate task-specific parameters for each hidden layer of the feature extractor and classifier to transfer knowledge. Simple CNAPS [3] improved upon this approach by replacing the parametric classifier with a Mahalanobis distance-based distance metric, which increased classification accuracy while reducing the number of parameters. Triantafillou et al.[43] proposed the FLUTE, which learns domain-specific parameters using FiLM layers called universal templates. Li et al.[21] have demonstrated that attaching task-specific weights to a pretrained model and optimizing them from scratch directly on a small support set can enable good generalization to the target task. However, the shifts of feature distribution between different target tasks and the pre-learned domain vary, leading to the change of task-specific parameters demand. In this paper, we focus on adaptively generating task-specific" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1591" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.092, + 0.852, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.364, + 0.895, + 0.409 + ], + "angle": 0, + "content": "Figure 2. Task-aware Adaptive Network (TA\\(^2\\)-Net). The \"Agent\" is the Action Generation Network, which predicts the \"action\" of Task-specific Adapter execution and interacts with the \"Environment\" (Env, the Adapted Network) providing the \"reward\" for the Agent. \"none\" represent that the model does not need the Task-specific Adapter to adapt feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.471, + 0.455 + ], + "angle": 0, + "content": "graphs for each target task, to adapt the prior knowledge to the target task well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Reinforcement Learning Reinforcement learning [1, 8] is primarily used to solve the sequence decision problem, in which the reward value obtained from agent-environment interaction is used as a feedback signal to continuously optimize agent strategies. There are three main approaches to solving Reinforcement Learning (RL) problems: methods based on value functions [30, 46, 48], methods based on policy search [41, 40], and the actor-critic approach [15, 39]. The deep reinforcement learning algorithm based on value functions approximates the value function or action-value function with a deep neural network and updates it with the temporal difference (TD) learning or Q-learning method. Among them, Deep Q-Network (DQN) [30] algorithm is representative. Furthermore, a series of variants have been proposed, including double DQN (DDQN) [46], which optimizes the problem of overestimation about Q value in the DQN algorithm, and Dueling DQN [48], which improves DQN from the network structure and can estimate Q function more accurately. Policy-based methods [41, 40] directly search for an optimal policy without maintaining a value function model. Parameters of the chosen policy are updated using either gradient-based or gradient-free optimization to maximize the expected return. The actor-critic approach [15] employs both value functions and policy search to trade-off the variance reduction of policy gradients and bias introduction from value function methods, in which the \"actor\" (policy) learns by using feedback from the \"critic\" (value function) [1]. In this paper, we employ policy-gradient-based Reinforcement Learning to train" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.424, + 0.895, + 0.469 + ], + "angle": 0, + "content": "the Action Generation Network to adaptively generate optimal actions – Task-specific Adapter execution decisions for each target task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.484, + 0.593, + 0.5 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.51, + 0.892, + 0.541 + ], + "angle": 0, + "content": "In this section, we describe the problem setting and introduce our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.551, + 0.63, + 0.568 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.892, + 0.684 + ], + "angle": 0, + "content": "Few-shot task Few-shot learning classification aims to classify a large number of samples with only a small number of labeled ones in each class. Labeled samples are generally referred to as the support set \\( S = \\{(\\pmb{x}_i, y_i)\\}_{i=1}^{N_S} \\), while classified ones are known as the query set \\( Q = \\{(\\pmb{x}_i)\\}_{i=1}^{N_Q} \\). \\( y_i \\) is the label of the sample \\( \\pmb{x}_i \\). A few-shot task is created by combining the support and query sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.821 + ], + "angle": 0, + "content": "Cross-domain few-shot learning Compared with traditional few-shot learning, cross-domain few-shot learning is more challenging, since not only the labeled samples in the target task \\( D_{t} \\) are few, but also the distribution of the target domain and the base domain \\( D_{b} \\) may be highly heterogeneous. Therefore, we must strike a balance between model capacity and the reliability of the adaptation while avoiding over-fitting to the few labeled samples or under-adaptation for the target task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.851, + 0.848 + ], + "angle": 0, + "content": "3.2. Task-aware Adaptive Network (TA²-Net)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Inspired by the observations that varying tasks require variant task-specific parameters for effective adaptation, we design the \\(\\mathrm{TA}^2\\)-Net adaptively learning the optimally" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1592" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.319 + ], + "angle": 0, + "content": "adapted network graph for each target task. As shown in Figure 2, the \\(\\mathrm{TA}^2\\)-Net consists of Action Generation Network and Adapted Network. To directly optimize the Action Generation Network to generate optimal task-specific parameters for the Adapted Network, we cast the model in Reinforcement Learning. In particular, Action Generation Network can be viewed as an \"Agent\" that interacts with the \"Environment\" (Adapted Network). The Action Generation Network will result in the \"actions\" that are the prediction of Task-specific Adapter execution for the target task. Based on the generated \"actions\", we can build the unique Adapted Network, which will provide a \"reward\" for the agent - we denote this reward by \\(r\\). We will introduce the Action Generation Network, Adapted Network, and their optimization process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.339, + 0.327, + 0.354 + ], + "angle": 0, + "content": "3.2.1 Action Generation Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.473, + 0.55 + ], + "angle": 0, + "content": "Our focus in the Action Generation Network is to generate adaptive Task-specific Adapter execution decisions for each layer of the Adapted Network aimed at the target task. As shown in Figure 2, the Gate Network is integrated into each layer of the feature encoder \\( f_{\\phi}(\\cdot) \\) pre-trained by the seen domain to generate the discrete probability for \\( K \\) kinds of Task-specific Adapters. Concretely, the discrete probability in \\( l \\)-th layer can be formulated as \\( \\{z_l^i\\}_{i=1}^K \\). Let \\( h_{l-1} \\in R^{N_S \\times C \\times W \\times H} \\) denotes the input feature of \\( l \\)-th feature encoder. In the Gate Network, the input feature is first fed to the Global Average Pooling (GAP) to compress the spatial dimension. Formally, the output \\( u_{l-1} \\) can be achieved as" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.564, + 0.47, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {u} _ {l - 1} = G A P \\left(\\boldsymbol {h} _ {l - 1}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.596, + 0.47, + 0.738 + ], + "angle": 0, + "content": "where \\( u_{l-1} \\) can be written as \\( u_{l-1} = \\left[u_{l-1}^{1}, u_{l-1}^{2}, \\dots, u_{l-1}^{N_{S}}\\right] \\). \\( u_{l-1}^{i} \\) represents the channel-level feature embedding of support sample \\( i \\) and \\( u_{l-1}^{i} \\in R^{C} \\). Then, let \\( v_{l-1} = \\left[v_{l-1}^{1}, v_{l-1}^{2}, \\dots, v_{l-1}^{N}\\right] \\) denotes the class prototypes in the target task, where \\( v_{l-1}^{n} \\) refers to the class center of class \\( n \\) and \\( N \\) denotes the number of classes in the target task. Formally, the class center for class \\( n \\) can be obtained by computing the mean vector of all embedded features in the class as" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.747, + 0.469, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {l - 1} ^ {n} = \\frac {1}{| S _ {n} |} \\sum_ {i \\in S _ {n}} \\boldsymbol {u} _ {l - 1} ^ {i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\( S_{n} \\) denotes the set of examples labeled with class \\( n \\). The output \\( \\boldsymbol{v}_{l-1} \\) is a collection of all class descriptors with statistics that express the entire target task. To capture the dependencies between features, we add a simple non-linear function of two fully-connected layers coupled with a ReLU activation function. The output of this operation is a vector containing unnormalized scores for executing the \\( K \\) kinds" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.678, + 0.107 + ], + "angle": 0, + "content": "of Task-specific Adapters." + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.117, + 0.893, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\bar {z} _ {l} = \\mathbf {W} _ {l} ^ {2} \\sigma \\left(\\mathbf {W} _ {l} ^ {1} v _ {l - 1}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.146, + 0.894, + 0.237 + ], + "angle": 0, + "content": "where \\(\\bar{z}_l\\in R^K\\) \\(\\sigma\\) refers to the \\(ReLU\\) activation function. \\(\\mathbf{W}_l^1\\in R^{d\\times C}\\) \\(\\mathbf{W}_l^2\\in R^{K\\times d}\\) and \\(d\\) is the dimension of the hidden layer. We can write \\(\\bar{z}_l\\) as \\(\\bar{z}_l = [\\bar{z}_l^1,\\bar{z}_l^2,\\dots ,\\bar{z}_l^K ]\\) Then, the Gate outputs a distribution over the \\(l\\) -th Task-specific Adapter execution decisions \\(\\pmb {z}_l\\) using the Softmax function" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.25, + 0.893, + 0.266 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {\\boldsymbol {\\ell}} = \\operatorname {s o f t m a x} (\\bar {\\boldsymbol {z}} _ {\\boldsymbol {\\ell}}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.894, + 0.336 + ], + "angle": 0, + "content": "where \\( z_{l} = \\left[z_{l}^{1},z_{l}^{2},\\dots ,z_{l}^{K}\\right] \\). \\( z_{l}^{k} \\) denotes the possibility of selecting \\( k \\)-th Task-specific Adapter in the \\( l \\)-th layer. And \\( z = [z_{1},z_{2},\\dots ,z_{L}] \\), \\( L \\) is the number of layers in the backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.895, + 0.489 + ], + "angle": 0, + "content": "Training The generation of Task-specific Adapter execution decisions is cast as the Reinforcement Learning problem. The Action Generation Network aims to decide the most appropriate Task-specific Adapters for conducting few-shot learning on the current tasks. We utilize the predicted result of the Adapted Network constructed based on the current \"actions\" as the rewards for optimizing the Action Generation Network. Inspired by the self-critical sequence training approach [34], the goal of training is to minimize the negative expected reward" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.5, + 0.892, + 0.517 + ], + "angle": 0, + "content": "\\[\nL (\\theta) = - \\left(r \\left(\\boldsymbol {z} ^ {s}\\right) - r \\left(\\boldsymbol {z} ^ {m}\\right)\\right) \\log \\left(\\boldsymbol {z} ^ {s}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.529, + 0.894, + 0.636 + ], + "angle": 0, + "content": "where \\(\\theta\\) is the parameters of the Gate Network. \\(\\theta = \\{\\mathbf{W}_l^1,\\mathbf{W}_l^2\\}\\) and \\(l\\in \\{1,2,\\dots ,L\\}\\). \\(z^{s} = [z_{1}^{s},\\dots,z_{L}^{s}]\\) and \\(z_{l}^{s}\\) is the possibility of action randomly sampled from the \\(K\\) actions - Task-specific Adapters, at the layer \\(l\\). \\(r(z^{s})\\) is the reward obtained by the Adapted Network, which is constructed by the actions with \\(z^{s}\\) discrete possibility. \\(z^{m} = [z_{1}^{m},\\dots,z_{L}^{m}]\\), and" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.644, + 0.892, + 0.661 + ], + "angle": 0, + "content": "\\[\nz _ {l} ^ {m} = \\operatorname {a r g m a x} \\left(\\boldsymbol {z} _ {l}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.895, + 0.791 + ], + "angle": 0, + "content": "where \\( \\operatorname{argmax}(\\cdot) \\) returns the possibility of action equipped with maximum probability over the output distribution \\( z_{l} \\) of the Gate in \\( l \\)-th layer. \\( r(z^{m}) \\) again is the reference reward obtained in Adapted Network equipped with actions with \\( z^{m} \\) discrete possibility. Accordingly, TA modules from the model that return a higher reward than \\( z^{m} \\) will be \"pushed up\", or increased in probability, while TA modules that result in a lower reward will be suppressed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.681, + 0.817 + ], + "angle": 0, + "content": "3.2.2 Adapted Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "As shown in Figure 2, the network architecture of the Adapted Network consists of a series of residual blocks, which are formed by the Adaptive Module (AM) followed by the BN layer and ReLU function. In the Adaptive Module, a \\(3 \\times 3\\) conv is paralleled with the Task-specific Adapter," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1593" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.092, + 0.372, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Task-aware Adaptive Learning" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.109, + 0.469, + 0.138 + ], + "angle": 0, + "content": "1: Train the backbone network \\( f_{\\phi}(\\cdot) \\) on the source domain;" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.14, + 0.469, + 0.183 + ], + "angle": 0, + "content": "2: Build Action Generation Network \\( M_{act} \\) on the backbone network; Initialization the parameters \\( \\theta \\) of Gate Networks;" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.185, + 0.223, + 0.2 + ], + "angle": 0, + "content": "3: while training do" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.2, + 0.468, + 0.232 + ], + "angle": 0, + "content": "4: Random sample task \\( T_{i} \\) in the source domain; \\( T_{i} = S \\bigcup Q, S = \\{(x_{i},y_{i})\\}_{i = 1}^{N_{S}}, Q = \\{(x_{i})\\}_{i = 1}^{N_{Q}} \\);" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.232, + 0.413, + 0.247 + ], + "angle": 0, + "content": "5: \\( \\pmb{z} = f_{\\phi, \\theta}(x), x \\in S, \\pmb{z} = [\\pmb{z}_1, \\pmb{z}_2, \\dots, \\pmb{z}_L] \\);" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.247, + 0.468, + 0.277 + ], + "angle": 0, + "content": "6: \\( \\pmb{z}^s = [Sample(\\pmb{z}_1), \\dots, Sample(\\pmb{z}_L)]; \\pmb{z}^m = [argmax(\\pmb{z}_1), \\dots, argmax(\\pmb{z}_L)]; \\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.277, + 0.468, + 0.322 + ], + "angle": 0, + "content": "7: Build Adapted Network \\( M_{adapt}^{1} \\), \\( M_{adapt}^{2} \\) based on the backbone network according to actions \\( z^{s} \\), \\( z^{m} \\) respectively;" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.322, + 0.398, + 0.339 + ], + "angle": 0, + "content": "8: Initialization \\(\\psi, \\pi\\) in \\(M_{adapt}^{1}\\) and \\(M_{adapt}^{2}\\);" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.339, + 0.468, + 0.37 + ], + "angle": 0, + "content": "9: Optimize \\(\\psi\\), \\(\\pi\\) over \\(S\\) based on \\(L_{adapt}(\\psi, \\pi) = \\frac{1}{N_S} \\sum_{(x,y) \\in S} l(f_{\\phi,\\psi,\\pi}(\\boldsymbol{x}), y)\\);" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.37, + 0.469, + 0.403 + ], + "angle": 0, + "content": "10: Predict over \\(Q\\); \\(r(\\pmb{z}^s) = N_{adapt}^1 / N_Q\\), \\(r(\\pmb{z}^m) = N_{adapt}^2 / N_Q\\);" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.403, + 0.468, + 0.433 + ], + "angle": 0, + "content": "11: Optimize the parameters \\(\\theta\\) of \\(M_{act}\\) based on \\(L(\\theta) = -(r(\\pmb{z}^s) - r(\\pmb{z}^m))\\log (\\pmb{z}^s)\\);" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.433, + 0.177, + 0.445 + ], + "angle": 0, + "content": "12: end while" + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.109, + 0.469, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.447, + 0.366, + 0.462 + ], + "angle": 0, + "content": "Output: Action Generation Network \\( M_{act} \\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.469, + 0.521 + ], + "angle": 0, + "content": "which is the action predicted by the corresponding Gate Network of the Action Generation Network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.469, + 0.688 + ], + "angle": 0, + "content": "Let \\( f_{\\phi_l}(\\cdot) \\) denote the \\( 3 \\times 3 \\) conv in the \\( l \\)-th Adaptive Module, which is pre-trained by the seen domain. With \\( h_{l-1} \\in R^{N_S \\times C \\times W \\times H} \\) as the input feature of \\( l \\)-th Adaptive Module, the output is \\( h_l = f_{\\phi_l}(h_{l-1}) \\) when no Task-specific Adapter is selected according to the decision generated by the Action Generation Network, that is the feature does not need adaptation. When the \\( k \\)-th Task-specific Adapter is chosen, which is constructed by paralleling \\( k \\) task-specific layers in our design, the feature learned by the Task-specific Adapter will be combined with the feature extracted from the \\( 3 \\times 3 \\) conv as" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.697, + 0.469, + 0.738 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {h} _ {\\boldsymbol {l}} = f _ {\\phi_ {\\boldsymbol {l}}} \\left(\\boldsymbol {h} _ {\\boldsymbol {l} - \\mathbf {1}}\\right) + \\frac {1}{k} \\sum_ {i = 1} ^ {k} f _ {\\psi_ {\\boldsymbol {l}} ^ {i}} \\left(\\boldsymbol {h} _ {\\boldsymbol {l} - \\mathbf {1}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.468, + 0.779 + ], + "angle": 0, + "content": "where \\( f_{\\psi_l^i}(\\cdot) \\) denotes the \\( i \\)-th task-specific layer in the Task-specific Adapter, which is \\( 1 \\times 1 \\) conv in the experiments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.779, + 0.469, + 0.856 + ], + "angle": 0, + "content": "Training Upon the generated \"actions\" with the discrete possibility \\( z^s \\) and \\( z^m \\) in the Action Generation Network, we will construct two Adapted Networks \\( M_{adapt}^{1} \\) and \\( M_{adapt}^{2} \\). In training, the Adapted Networks are trained with the cross-entropy loss over the support samples:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.866, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\nL _ {a d a p t} (\\psi , \\pi) = \\frac {1}{N _ {S}} \\sum_ {(x, y) \\in S} l \\left(f _ {\\phi , \\psi , \\pi} (\\boldsymbol {x}), y\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "where \\( f_{\\psi, \\pi}(\\cdot) \\) denotes the output softmax probability vector of the linear classifier with parameters \\( \\pi \\), in which dimensionality equals the number of categories in the support set \\( S \\). Then, the reward for the agent is achieved by predicting the Adapted Network over the query samples \\( Q \\):" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.177, + 0.892, + 0.214 + ], + "angle": 0, + "content": "\\[\nr \\left(\\boldsymbol {z} ^ {\\boldsymbol {s}}\\right) = \\frac {N _ {\\text {a d a p t}} ^ {1}}{N _ {Q}} \\quad \\text {a n d} \\quad r \\left(\\boldsymbol {z} ^ {\\boldsymbol {m}}\\right) = \\frac {N _ {\\text {a d a p t}} ^ {2}}{N _ {Q}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.225, + 0.892, + 0.29 + ], + "angle": 0, + "content": "where \\( N_{adapt}^{1} \\) and \\( N_{adapt}^{2} \\) denote the number of query samples correctly predicted by Adapted Networks \\( M_{adapt}^{1} \\) and \\( M_{adapt}^{2} \\) respectively. We show the Task-aware Adaptive Learning pseudo code in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.308, + 0.606, + 0.323 + ], + "angle": 0, + "content": "3.2.3 Testing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.331, + 0.892, + 0.437 + ], + "angle": 0, + "content": "Note that the support samples of testing will first go through the Action Generation Network to generate optimal Task-specific Adapter executions for the target task, which are the action with maximum probability. And then, based on the executions, the Adapted Network will be built and finetuned over the support set. Finally, the query samples will be predicted by the Adapted Network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.451, + 0.632, + 0.467 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.521 + ], + "angle": 0, + "content": "In this part, we first introduce the experimental setup and then compare our method with the state-of-the-art methods. Furthermore, we provide a series of analyses of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.687, + 0.547 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Dataset. Meta-Dataset [44] is the standard benchmark for evaluating the performance of few-shot classification. It consists of 13 image datasets: ILSVRC 2012 [36], Omniglot [18], FGVC-Aircraft [29], CUB-200-2011 [47], Describable Textures [5], Quick Draw [11], FGVCx Fungi [38], VGG Flower [31], Traffic Signs [13], MSCOCO [24], MNIST [19], CIFAR-10 [16], and CIFAR-100 [16]. For all datasets in the experiments, we used the standard split procedure as [44]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation details. In the experiments, we considered two training settings: multi-domain learning and single-domain setting. In the multi-domain setting, the first eight datasets (ILSVRC 2012, Omniglot, Aircraft, CUB-200-2011, Textures, Quick Draw, Fungi, and VGG Flower) of the Meta-dataset are used to train the feature extractor. We followed the method in [20] to obtain the pre-trained model. And in the single-domain setting, only the ILSVRC 2012 dataset participated in the training of the feature extractor. As in [20], we adopted the ResNet-18 as the backbone no matter in the multi-domain setting or the single-domain setting. In the training, the \\(\\mathrm{TA}^2\\)-Net is trained using the Adam optimizer. The hyper-parameters \\(K\\) are set as \\(K = 3\\). We also provide the ablation study in Section 4.4." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1594" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.116 + ], + "angle": 0, + "content": "Table 1. Comparison to state-of-the-art methods in the multi-domain setting. Mean accuracy, \\(95\\%\\) confidence interval are reported. The best results are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.121, + 0.895, + 0.451 + ], + "angle": 0, + "content": "
Test DatasetCNAPS [35]Simple CNAPS [3]Transductive CNAPS [2]SUR [7]URT [26]FLUTE [43]tri-M [27]URL [20]TSA [21]Ours
ImageNet50.8 ± 1.158.4 ± 1.157.9 ± 1.156.2 ± 1.056.8 ± 1.158.6 ± 1.051.8 ± 1.158.8 ± 1.159.5 ± 1.059.6 ± 1.0
Omniglot91.7 ± 0.591.6 ± 0.694.3 ± 0.494.1 ± 0.494.2 ± 0.492.0 ± 0.693.2 ± 0.594.5 ± 0.494.9 ± 0.495.5 ± 0.4
Aircraft83.7 ± 0.682.0 ± 0.784.7 ± 0.585.5 ± 0.585.8 ± 0.582.8 ± 0.787.2 ± 0.589.4 ± 0.489.9 ± 0.490.5 ± 0.4
Birds73.6 ± 0.974.8 ± 0.978.8 ± 0.771.0 ± 1.076.2 ± 0.875.3 ± 0.879.2 ± 0.880.7 ± 0.881.1 ± 0.881.4 ± 0.8
Textures59.5 ± 0.768.8 ± 0.966.2 ± 0.871.0 ± 0.871.6 ± 0.771.2 ± 0.868.8 ± 0.877.2 ± 0.777.5 ± 0.777.4 ± 0.7
Quick Draw74.7 ± 0.876.5 ± 0.877.9 ± 0.681.8 ± 0.682.4 ± 0.677.3 ± 0.779.5 ± 0.782.5 ± 0.681.7 ± 0.682.5 ± 0.6
Fungi50.2 ± 1.146.6 ± 1.048.9 ± 1.264.3 ± 0.964.0 ± 1.048.5 ± 1.058.1 ± 1.168.1 ± 0.966.3 ± 0.866.3 ± 0.9
VGG Flower88.9 ± 0.590.5 ± 0.592.3 ± 0.482.9 ± 0.887.9 ± 0.690.5 ± 0.591.6 ± 0.692.0 ± 0.592.2 ± 0.592.6 ± 0.4
Traffic Sign56.5 ± 1.157.2 ± 1.059.7 ± 1.151.0 ± 1.148.2 ± 1.163.0 ± 1.058.4 ± 1.163.3 ± 1.182.8 ± 1.087.4 ± 0.8
MSCOCO39.4 ± 1.048.9 ± 1.142.5 ± 1.152.0 ± 1.151.5 ± 1.152.8 ± 1.150.0 ± 1.057.3 ± 1.057.6 ± 1.057.9 ± 0.9
MNIST-94.6 ± 0.494.7 ± 0.394.3 ± 0.490.6 ± 0.596.2 ± 0.395.6 ± 0.594.7 ± 0.496.7 ± 0.497.0 ± 0.4
CIFAR-10-74.9 ± 0.773.6 ± 0.766.5 ± 0.967.0 ± 0.875.4 ± 0.878.6 ± 0.774.2 ± 0.882.9 ± 0.782.1 ± 0.8
CIFAR-100-61.3 ± 1.161.8 ± 1.056.9 ± 1.157.3 ± 1.062.0 ± 1.067.1 ± 1.063.5 ± 1.070.4 ± 0.970.9 ± 0.9
Average Seen71.673.775.175.977.474.576.280.480.480.7
Average Unseen-67.466.564.162.969.969.970.678.179.1
Average All-71.271.871.471.872.773.876.679.580.1
Average Rank-7.66.27.26.55.85.53.22.21.2
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.468, + 0.418, + 0.484 + ], + "angle": 0, + "content": "4.2. Comparison to state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.491, + 0.47, + 0.553 + ], + "angle": 0, + "content": "Following the experiment set in [21], we sampled tasks with varying numbers of ways and shots in the inference and conducted experiments in the multi-domain setting and single-domain setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.471, + 0.824 + ], + "angle": 0, + "content": "Multi-domain setting. Here, the results of our method on 13 datasets of the Meta-Dataset are compared with the state-of-the-art methods. We also report average accuracy over seen domains, and unseen domains, as well as the average rank as in [21]. From the results in Table 1, our method obtains state-of-the-art results on 10 out of 13 datasets and achieves the best average accuracy on seen and unseen domains. It is worth noting that our method obtains significantly better average results than the second-best approach on the unseen datasets \\((+1.0\\%)\\). The unseen datasets are heterogeneous with respect to the training domain, with a significant domain gap between them. In contrast to adopting the fixed number of task-specific parameters for all target tasks, e.g. CNAPS, Simple CNAPS, FLUTE, and TSA, the \\(\\mathrm{TA}^2\\)-Net can adaptively learn the optimal task-specific parameter policies for unseen target tasks to efficiently transfer the pre-trained knowledge to the target feature distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Single-domain setting. We also evaluate our method in the single-domain setting, which is more challenging due to optimizing the Action Generation Network only on the ImageNet dataset. Results are shown in Table 2. First, compared with other single-domain training methods, our" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.895, + 0.665 + ], + "angle": 0, + "content": "method achieves significant superiority. For example, \\(\\mathrm{TA}^2\\)-Net outperforms the most competing method (TSA [21]) by \\(+2.9\\%\\), \\(+0.8\\%\\), \\(+1.6\\%\\), and \\(+1\\%\\) on Omniglot, Quick Draw, Traffic Signs and MNIST, respectively. Second, it is worth noting that, for the average accuracy on the unseen datasets, ours outperforms the second-best approach with a clear margin \\((+3.1\\%)\\). And we clearly observe that \\(\\mathrm{TA}^2\\)-Net achieves competitive accuracy on the ImageNet, compared with the best performance. Despite being trained on just one dataset, the Action Generation Network exhibits good generalization, producing the optimal task-specific parameter policies for the unseen domain in an excellently adaptive manner." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.669, + 0.696 + ], + "angle": 0, + "content": "4.3. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Varying-way Five-shot setting. The number of support samples ('shot') in the standard Meta-Dataset evaluation setting is variable and can range up to 100. In this section, we evaluate our method in the 5-shot setting which is more challenging due to fewer labeled samples and can better reflect the progressiveness of the model. Table 3 shows the results compared with other state-of-the-art methods. As can be seen, ours achieves better average accuracy on both seen and unseen domains. More specifically, our method outperforms other methods on the vast majority of domains (8 out of 13) and obtains the best average rank. Furthermore, the average accuracy on unseen datasets is \\(1.2\\%\\) higher than the second-best method." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1595" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.088, + 0.522, + 0.128 + ], + "angle": 0, + "content": "Table 2. Comparison to state-of-the-art methods in the single-domain setting. Mean accuracy, \\(95\\%\\) confidence interval are reported. The best results are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.133, + 0.522, + 0.352 + ], + "angle": 0, + "content": "
Test DatasetProtoNet [44]fo-Proto -MAML [44]ALFA+fo-Proto -MAML [44]BOHB [37]FLUTE [43]TSA [21]Ours
ImageNet50.5 ± 1.149.5 ± 1.152.8 ± 1.151.9 ± 1.146.9 ± 1.159.5 ± 1.159.3 ± 1.1
Omniglot60.0 ± 1.463.4 ± 1.361.9 ± 1.567.6 ± 1.261.6 ± 1.478.2 ± 1.281.1 ± 1.1
Aircraft53.1 ± 1.056.0 ± 1.063.4 ± 1.154.1 ± 0.948.5 ± 1.072.2 ± 1.072.6 ± 0.9
Birds68.8 ± 1.068.7 ± 1.069.8 ± 1.170.7 ± 0.947.9 ± 1.074.9 ± 0.975.1 ± 0.9
Textures66.6 ± 0.866.5 ± 0.870.8 ± 0.968.3 ± 0.863.8 ± 0.877.3 ± 0.776.8 ± 0.8
Quick Draw49.0 ± 1.151.5 ± 1.059.2 ± 1.250.3 ± 1.057.5 ± 1.067.6 ± 0.968.4 ± 0.9
Fungi39.7 ± 1.140.0 ± 1.141.5 ± 1.241.4 ± 1.131.8 ± 1.044.7 ± 1.045.3 ± 1.0
VGG Flower85.3 ± 0.887.2 ± 0.786.0 ± 0.887.3 ± 0.680.1 ± 0.990.9 ± 0.691.0 ± 0.6
Traffic Sign47.1 ± 1.148.8 ± 1.160.8 ± 1.351.8 ± 1.046.5 ± 1.182.5 ± 0.884.1 ± 0.7
MSCOCO41.0 ± 1.143.7 ± 1.148.1 ± 1.148.0 ± 1.041.4 ± 1.059.0 ± 1.058.0 ± 1.0
MNIST----80.8 ± 0.893.9 ± 0.694.9 ± 0.5
CIFAR-10----65.4 ± 0.882.1 ± 0.782.0 ± 0.7
CIFAR-100----52.7 ± 1.170.7 ± 0.970.8 ± 0.9
Average Seen50.549.552.851.946.959.559.3
Average Unseen56.758.462.460.053.271.975.0
Average All56.157.561.459.252.670.773.8
Average Rank5.95.23.44.15.61.71.3
" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.371, + 0.43, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.563, + 0.471, + 0.62 + ], + "angle": 0, + "content": "Figure 3. Performance of widening or deepening the Task-specific Adapter (TA) in the multi-domain setting. The axes indicate the accuracy of methods on a particular dataset. Each color corresponds to a different method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Wider or deeper in the Task-specific Adapter (TA)? To explore the method for improving the adaptation of the model, we increase the parameters of the Task-specific Adapter (TA) in wide or depth respectively. Specifically, two types of TAs are designed: (a) \\(1 \\times 1\\) Conv in parallel, (b) \\(1 \\times 1\\) Conv in series. As shown in Figure 3, increasing the width of the TA allows the model to achieve better performance than deepening. In particular, among the 13 datasets, 5 datasets (Traffic Signs, MSCOCO, MNIST, CIFAR-10, and CIFAR-100) unseen in the training obviously benefited more from widening the Task-specific Adapter (TA). All of the above indicates that the best way to improve adaptability is to increase the width of the Task-specific Adapter (TA)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of adaptive learning. In this section, we attempt to provide a more in-depth discussion of the efficacy of fixed networks as well as the proposed adaptive learned network. Figure 4 depicts the performance over the" + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.088, + 0.892, + 0.128 + ], + "angle": 0, + "content": "Table 3. Results for Varying-Way Five-Shot setting. Mean accuracy, \\(95\\%\\) confidence interval are reported. The best results are marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.132, + 0.892, + 0.351 + ], + "angle": 0, + "content": "
Test DatasetSimple CNAPS [3]SUR [7]URT [26]URL [20]TSA [21]Ours
ImageNet47.2 ± 1.046.7 ± 1.048.6 ± 1.049.4 ± 1.048.3 ± 1.049.3 ± 1.0
Omniglot95.1 ± 0.395.8 ± 0.396.0 ± 0.396.0 ± 0.396.8 ± 0.396.6 ± 0.2
Aircraft74.6 ± 0.682.1 ± 0.681.2 ± 0.684.8 ± 0.585.5 ± 0.585.9 ± 0.4
Birds69.6 ± 0.762.8 ± 0.971.2 ± 0.776.0 ± 0.676.6 ± 0.677.3 ± 0.6
Textures57.5 ± 0.760.2 ± 0.765.2 ± 0.769.1 ± 0.668.3 ± 0.768.3 ± 0.6
Quick Draw70.9 ± 0.679.0 ± 0.579.2 ± 0.578.2 ± 0.577.9 ± 0.678.5 ± 0.5
Fungi50.3 ± 1.066.5 ± 0.866.9 ± 0.970.0 ± 0.870.4 ± 0.870.3 ± 0.8
VGG Flower86.5 ± 0.476.9 ± 0.682.4 ± 0.589.3 ± 0.489.5 ± 0.490.0 ± 0.4
Traffic Sign55.2 ± 0.844.9 ± 0.945.1 ± 0.957.5 ± 0.872.3 ± 0.676.7 ± 0.5
MSCOCO49.2 ± 0.848.1 ± 0.952.3 ± 0.956.1 ± 0.856.0 ± 0.856.0 ± 0.8
MNIST88.9 ± 0.490.1 ± 0.486.5 ± 0.589.7 ± 0.492.5 ± 0.493.3 ± 0.3
CIFAR-1066.1 ± 0.750.3 ± 1.061.4 ± 0.766.0 ± 0.772.0 ± 0.773.1 ± 0.7
CIFAR-10053.8 ± 0.946.4 ± 0.952.5 ± 0.957.0 ± 0.964.1 ± 0.864.1 ± 0.8
Average Seen69.071.273.876.676.777.0
Average Unseen62.656.059.665.271.472.6
Average All66.565.468.372.274.675.3
Average Rank5.254.12.82.21.5
" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.371, + 0.852, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.584, + 0.828, + 0.598 + ], + "angle": 0, + "content": "Figure 4. Effectiveness of adaptive learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Meta-dataset in the multi-domain and single-domain setting, with Model-1/2/3 constructed by attaching Module-1/2/3 respectively as shown in Figure 1 to the backbone. It can be observed that, in comparison to the backbone network, the average accuracy of seen domains in both the multi-domain and single-domain settings decreased when using the fixed Task-specific Adapter to adaptation, but our method performed better. Furthermore, average accuracy on unseen domains has consistently improved overall in the fixed Task-specific Adapter and can be improved further through adaptive learning. This discovery inspires us to recognize that simply designing a fixed Task-specific Adapter does not work and forces us to recognize the necessity of adaptively learning optimal task-specific parameters policy for each target task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of Reinforcement Learning. Our \\(\\mathrm{TA}^2\\)-Net model adopts the Reinforcement Learning algorithm combined with Episodic Training to adaptively learn opti" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1596" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.422, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.226, + 0.327, + 0.249 + ], + "angle": 0, + "content": "(a) MMD and TA parameters of different datasets in the Meta-Dataset" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.091, + 0.659, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.226, + 0.66, + 0.25 + ], + "angle": 0, + "content": "(b) TA parameters distribution in different layers for the Fungi dataset" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.092, + 0.885, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.226, + 0.888, + 0.249 + ], + "angle": 0, + "content": "(c) TA parameters distribution in different layers for the CIFAR-10 dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.892, + 0.314 + ], + "angle": 0, + "content": "Figure 5. (a): The Line chart shows the mean and standard deviation of TA parameters adaptively learned by the \\(\\mathrm{TA}^2\\)-Net in target tasks sampled in different datasets. And the box plot shows the \\(25\\%\\) quantiles of the MMD in different datasets while the notches indicate the median. (b) and (c): TA parameters distribution in different blocks of the task-aware adapted network for the Fungi and CIFAR-10 dataset. Sub-bars in the block denote different layers within the Block." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.482 + ], + "angle": 0, + "content": "mal task-specific parameters policy for target tasks. To explore the effectiveness of Reinforcement Learning, we designed the Simple \\(\\mathrm{TA}^2\\)-Net trained by Episodic Training, which only includes the Adapted Network attaching a Gate Network in each layer for generating \"decisions\" - Chosen the Task-specific Adapter. As shown in Table 4, the \\(\\mathrm{TA}^2\\)-Net optimized with the Reinforcement Learning outperforms the Simple \\(\\mathrm{TA}^2\\)-Net, which shows that Reinforcement Learning can better learn the TA decisions for target tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Parameters distribution. This is an intuitive guess that the greater the difference (domain shift) between the source domain and the target domain features, the greater the demand for TA parameters. In this section, we first observe the relationship between the TA parameters and the domain shift. As widely used in previous works [28, 25], we adopt the Maximum Mean Discrepancy (MMD) to quantify the domain shift between the target task features and the source domain features extracted by the pre-trained backbone (in the multi-domain setting). The larger MMD indicates a larger domain shift between the test task and the source domain. As shown in Figure 5 (a), the Line chart and box plot shows the TA parameters and MMD distribution in different datasets, respectively. We can find that MMD for each dataset has a narrow distribution. And, on the whole, datasets with significant MMD have a larger demand for the TA parameters. Furthermore, linear regression and Pearson correlation analysis revealed that MMD is positively correlated with TA parameters, with the Pearson correlation coefficient \\( R = 0.9186 \\). It is attributed to the greater heterogeneity, more task-specific parameters are required to effectively adapt pre-learned knowledge to the feature distribution of target tasks. We further explore the TA parameters distribution in different layers of the task-aware adapted network learned by our proposed TA\\(^2\\)-Net. As shown in Figure 5 (b) and (c), layers in the deep blocks have more task-specific parameters for learning the task-specific fea" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.328, + 0.89, + 0.356 + ], + "angle": 0, + "content": "Table 4. Effectiveness of Reinforcement Learning in the multidomain setting." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.362, + 0.875, + 0.435 + ], + "angle": 0, + "content": "
ModelReinforcement LearningSeen DomainUnseen DomainAll
Simple TA2-NetX75.567.972.6
TA2-Net✓77.072.675.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.449, + 0.892, + 0.476 + ], + "angle": 0, + "content": "Table 5. Mean accuracy \\((\\%)\\) for different number of layers \\(K\\) in the Task-specific Adapter." + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.482, + 0.855, + 0.552 + ], + "angle": 0, + "content": "
K=1K=2K=3K=4
Seen Domain75.475.275.375.1
Unseen Domain66.670.370.568.2
All72.073.373.572.5
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.891, + 0.64 + ], + "angle": 0, + "content": "tures than those in the shallow blocks as demonstrated by both Fungi and CIFAR-10, which may be because the shallow layer contains more general information while the deep layer contains more domain-specific information [50]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.654, + 0.667 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Maximum width \\( K \\) of the Task-specific Adapter. Here, we discuss the maximum width \\( K \\) of the Task-specific Adapter, which represents the upper bound of the task-specific parameters. Experiments are conducted on the Meta-dataset in the varying-way five-shot setting. Table 5 shows the performances of \\( K \\in [1,2,3,4] \\). We can observe that, as \\( K \\) increases, the average classification accuracy on seen domains lightly decreases. For unseen domains, the average classification accuracy increases at first and then decreases as \\( K \\) increases, with \\( K = 3 \\) producing the best results. Compared with seen domains, unseen domains are heterogeneous with the source domain and require more task-specific parameters to transfer the pre-learned knowledge. The average performance becomes poorer when using \\( K = 4 \\), which may be due to too much TA, the model" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1597" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.168 + ], + "angle": 0, + "content": "overfitting, or the search space of the model structure expanding, increasing the optimization difficulty of reinforcement learning. Considering the performance of various target tasks and the calculating costs, all experiments in this paper adopt \\( K = 3 \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.181, + 0.197, + 0.197 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.207, + 0.473, + 0.388 + ], + "angle": 0, + "content": "In this paper, for studying cross-domain few-shot learning problems, we explored the performance of test tasks when inference was performed in models equipped with a variant number of task-specific parameters. The exploration revealed that target tasks require variant task-specific parameters for outstanding adaptation. Motivated by this, we proposed a Task-aware Adaptive Network (TA\\(^2\\)-Net), which can adaptively learn optimal task-specific parameters policy for each target task for exhaustively improving the recognition performance of cross-domain few-shot learning. By conducting extensive experiments, we proved that our TA\\(^2\\)-Net could achieve the best results on Meta-dataset." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.402, + 0.229, + 0.42 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.47, + 0.548 + ], + "angle": 0, + "content": "This work was supported in part by Beijing Natural Science Foundation Project No. Z200002, in part by National Natural Science Foundation of China (NSFC) No. U19B2036, 62225601, in part by the Program for Youth Innovative Research Team of BUPT No. 2023QNTD02, in part by scholarships from China Scholarship Council (CSC) under Grant CSC No. 202206470055, in part by BUPT Excellent Ph.D. Students Foundation No. CX2022152." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.563, + 0.175, + 0.578 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.47, + 0.643 + ], + "angle": 0, + "content": "[1] Kai Arulkumaran, Marc Peter Deisenroth, Miles Brundage, and Anil Anthony Bharath. Deep reinforcement learning: A brief survey. IEEE Signal Processing Magazine, 34(6):26-38, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[2] Peyman Bateni, Jarred Barber, Jan-Willem van de Meent, and Frank Wood. Enhancing few-shot image classification with unlabelled examples. In WACV, pages 2796-2805, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.702, + 0.468, + 0.743 + ], + "angle": 0, + "content": "[3] Peyman Bateni, Raghav Goyal, Vaden Masrani, Frank Wood, and Leonid Signal. Improved few-shot visual classification. In CVPR, pages 14493-14502, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.745, + 0.468, + 0.785 + ], + "angle": 0, + "content": "[4] Luca Bertinetto, Joao F. Henriques, Philip Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In ICLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.788, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.831, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[6] Carl Doersch, Ankush Gupta, and Andrew Zisserman. CrossTransformers: spatially-aware few-shot transfer. NeurIPS, 33:21981-21993, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.468, + 0.902 + ], + "angle": 0, + "content": "[7] Nikita Dvornik, Cordelia Schmid, and Julien Mairal. Selecting relevant features from a multi-domain representation for" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.588, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "few-shot classification. In ECCV, pages 769-786, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[8] Sven Gronauer and Klaus Diepold. Multi-agent deep reinforcement learning: a survey. Artificial Intelligence Review, 55(2):895-943, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.165, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[9] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In ECCV, pages 124-141, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[10] Yurong Guo, Ruoyi Du, Xiaoxu Li, Jiyang Xie, Zhanyu Ma, and Yuan Dong. Learning calibrated class centers for few-shot classification by pair-wise similarity. IEEE TIP, 31:4543-4555, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[11] David Ha and Douglas Eck. A Neural Representation of Sketch Drawings. arXiv preprint arXiv:1704.03477, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[13] Sebastian Houben, Johannes Stallkamp, Jan Salmen, Marc Schlipsing, and Christian Igel. Detection of traffic signs in real-world images: The german traffic sign detection benchmark. In IJCNN, pages 1-8, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[14] Yanxu Hu and Andy J. Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, pages 20-37, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[15] Vijay Konda and John Tsitsiklis. Actor-critic algorithms. NeurIPS, 12, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[16] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[17] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[18] Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 350(6266):1332-1338, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[19] Yann LeCun, Corinna Cortes, and CJ Burges. MNIST handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, 2, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.662, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[20] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple domains for few-shot classification. In ICCV, pages 9526-9535, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.705, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[21] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Cross-domain few-shot learning with task-specific adapters. In CVPR, pages 7161-7170, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[22] Xiaoxu Li, Jijie Wu, Zhuo Sun, Zhanyu Ma, Jie Cao, and Jing-Hao Xue. BSNet: Bi-similarity network for few-shot fine-grained image classification. IEEE TIP, 30:1318-1331, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[23] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. MetaSGD: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1598" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[25] Ge Liu, Linglan Zhao, and Xiangzhong Fang. PDA: Proxy-based domain adaptation for few-shot image recognition. Image and Vision Computing, 110:104164, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[26] Lu Liu, William Hamilton, Guodong Long, Jing Jiang, and Hugo Larochelle. A universal representation transformer layer for few-shot image classification. In ICLR, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[27] Yanbin Liu, Juho Lee, Linchao Zhu, Ling Chen, Humphrey Shi, and Yi Yang. A multi-mode modulator for multi-domain few-shot classification. In ICCV, pages 8453-8462, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[28] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[30] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.441 + ], + "angle": 0, + "content": "[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[32] Hang Qi, Matthew Brown, and David G. Lowe. Low-Shot learning with imprinted weights. In CVPR, pages 5822-5830, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[33] Sachin Ravi and H. Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.553 + ], + "angle": 0, + "content": "[34] Steven J Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel. Self-critical sequence training for image captioning. In CVPR, pages 7008-7024, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.554, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[35] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, and Richard E Turner. Fast and flexible multi-task classification using conditional neural adaptive processes. NeurIPS, 32, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[36] Olga Russakovsky, J. Deng, Hao Su, J. Krause, S. Satheesh, S. Ma, Zhiheng Huang, A. Karpathy, A. Khosla, Michael S. Bernstein, A. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[37] Tonmoy Saikia, Thomas Brox, and Cordelia Schmid. Optimized generic feature learning for few-shot classification across domains. arXiv preprint arXiv:2001.07926, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[38] Brigit Schroeder and Yin Cui. FGVCx Fungi Classification Challenge 2018. github.com/visipedia/fgvcx_fungi_comp, 2018.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[39] John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[40] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. In ICML, pages 387-395, 2014. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[43] Eleni Triantafillou, Hugo Larochelle, Richard Zemel, and Vincent Dumoulin. Learning a universal template for few-shot dataset generalization. In ICML, pages 10424–10433, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[44] Eleni Triantafillou, Tyler Zhu, Vincent Dumoulin, Pascal Lamblin, Utku Evci, Kelvin Xu, Ross Goroshin, Carles Gelada, Kevin Swersky, Pierre-Antoine Manzagol, et al. Meta-dataset: A dataset of datasets for learning to learn from few examples. arXiv preprint arXiv:1903.03096, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[45] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In ICLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[46] Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-learning. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[47] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[48] Ziyu Wang, Tom Schaul, Matteo Hessel, Hado Hasselt, Marc Lanctot, and Nando Freitas. Dueling network architectures for deep reinforcement learning. In ICML, pages 1995–2003, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[49] Davis Wertheimer, Luming Tang, and Bharath Hariharan. Few-shot classification with feature map reconstruction networks. In CVPR, pages 8012-8021, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[50] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1599" + } + ] +] \ No newline at end of file diff --git a/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/d73ef7d6-b7de-4511-b8bf-4b4e20276122_origin.pdf b/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/d73ef7d6-b7de-4511-b8bf-4b4e20276122_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..15a42590ba2923b2595463386796d88ab96a5a47 --- /dev/null +++ b/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/d73ef7d6-b7de-4511-b8bf-4b4e20276122_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:723573b6709cc6273b0b430955a92362b22b56e0cce04c4447f0b654b5c5a07e +size 1864009 diff --git a/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/full.md b/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b125d92477f3d60497fd97b57ae8874c453b2e58 --- /dev/null +++ b/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/full.md @@ -0,0 +1,328 @@ +# Task-aware Adaptive Learning for Cross-domain Few-shot Learning + +Yurong Guo $^{1}$ , Ruoyi Du $^{1}$ , Yuan Dong $^{1}$ , Timothy Hospedales $^{2}$ , Yi-Zhe Song $^{3}$ , Zhanyu Ma $^{1*}$ $^{1}$ Beijing University of Posts and Telecommunications, China + $^{2}$ University of Edinburgh, UK + $^{3}$ University of Surrey, UK + +{guoyurong, duruoyi, yuandong, mazhanyu}@bupt.edu.cn, t.hospedales@ed.ac.uk, y.song@surrey.ac.uk + +# Abstract + +Although existing few-shot learning works yield promising results for in-domain queries, they still suffer from weak cross-domain generalization. Limited support data requires effective knowledge transfer, but domain-shift makes this harder. Towards this emerging challenge, researchers improved adaptation by introducing task-specific parameters, which are directly optimized and estimated for each task. However, adding a fixed number of additional parameters fails to consider the diverse domain shifts between target tasks and the source domain, limiting efficacy. In this paper, we first observe the dependence of task-specific parameter configuration on the target task. Abundant task-specific parameters may over-fit, and insufficient task-specific parameters may result in under-adaptation – but the optimal task-specific configuration varies for different test tasks. Based on these findings, we propose the Task-aware Adaptive Network (TA $^2$ -Net), which is trained by reinforcement learning to adaptively estimate the optimal task-specific parameter configuration for each test task. It learns, for example, that tasks with significant domain-shift usually have a larger need for task-specific parameters for adaptation. We evaluate our model on Meta-dataset. Empirical results show that our model outperforms existing state-of-the-art methods. Our code is available at https://github.com/PRIS-CV/TA2-Net. + +# 1. Introduction + +Traditional deep learning models [42, 36, 12, 17] show excellent generalization performance when training on a large number of labeled samples. However, both abundant samples and reliable annotations are not always available in realistic applications, e.g., rare disease diagnosis, and fine-grained recognition. Few-shot learning [32, 33, 23, 4, 10], inspired by the fact that humans can quickly learn new + +![](images/a5f61152cc6709e9fbbfc88f6be0b037b9db09271266362c3116aafc075c2f8a.jpg) +(a) Different tasks from MSCOCO + +![](images/ab4c6f2ce3a7baae99c0ab79ad9c702417e93954cbaa278aace44d8d178cf7e7.jpg) +(c) Different layers in a task from MSCOCO + +![](images/6f010ea80287bfdd4011a9b1f033015060138d6a5c34b5e6733ff7d306230aac.jpg) +(b) Different tasks from ilsvrc 2012 + +![](images/37227961c6896da2ff9661fa2241f02fcc369060230080f2ef39cf143b9f9c18.jpg) +(d) Different layers in a task from ilsvrc_2012 + +![](images/5eb0ed051a37115f28ea7fb9a4414fa12d667abe339c130eaca44efffb10750d.jpg) +Figure 1. (a) and (b) reveal the accuracy of varying test tasks when inference was performed in different models, in which TA-Modules are attached to each layer in the backbone network. (c) and (d) show the accuracy when TA-Modules are attached to different layers in the backbone network. Four dashed boxes show the backbone and several TA-Modules. + +![](images/00d37dc4197337b48a856d0070ce01391596ca45c2482ff87e350ed8bd6f63d4.jpg) + +![](images/f69f2fcd8b8e87168852ce4f74b0fc88398e0ca3e48ff677b600c28204acade1.jpg) + +![](images/8f6721d53d122fccde922e3bb27316351ef6ca068aa5018c03e0cc119d1f5a38.jpg) + +knowledge, aims to adapt the model to new classes by only a few labeled samples for each one. + +Recently, meta-learning-based few-shot learning methods have made great strides in the setting where the train and test tasks are sampled from the same domain [49, 22]. However, more and more studies have demonstrated that these existing works fail to generalize well to novel classes that are heterogeneous with the source domain [45, 9, 14]. This limitation is attributable to the fact that most previous few-shot learning models focus solely on how to quickly adapt to a set of novel classes with only a few labeled samples per class but less effort has been made to understand and address the domain shift problem between the target (test) task and training domain. Recently, to address this problem, a series of research referred to as cross-domain few-shot learning are proposed [45, 14, 35, 3, 7, 20, 26, 6]. + +Generally, these existing methods address the domain + +shift problem by introducing a set of task-specific parameters to enable the feature extractor or classifier to adapt to the feature distributions of new tasks. According to the parameter generation method, these methods can be loosely divided into the following two main streams. One kind of approach is to use an auxiliary network [35, 3, 43] to generate task-specific parameters. The auxiliary network is first meta-trained with multiple tasks from the source domain, and then it utilizes the support set to estimate task-specific parameters for the target task. Other methods [21] directly attach a set of task-specific parameters to the pre-trained model on the source domain and then optimize them on a few labeled samples of the target task to well generalize the knowledge to new classes. Although cross-domain few-shot learning performance has improved, the fixed number of task-specific parameters may be too rigid to account for diverse tasks, resulting in less generality. + +In this paper, from the perspective of parameter size, we studied the performance of test tasks when inference was performed in models equipped with varying sizes of task-specific parameters. Specifically, following the trend of introducing directly optimizing task-specific parameters [21], we designed several Task-specific Adapters (TA-Modules) with increasing parameters, which are respectively attached to the conv of pre-trained backbone network (e.g., ResNet-18) to construct cross-domain models. As shown in Figure 1, the optimal task-specific parameters policy varies depending on the target task. More specifically, for different test tasks, the optimal task-specific parameters required in different layers of the backbone network for effectively adapting pre-learned knowledge to the target task's feature distribution differ. Abundant task-specific parameters may over-fit on target tasks, whereas few task-specific parameters will show the under-adaptation problem. In order to effectively direct the feature extractor to adapt feature distributions, we must carefully design task-specific parameters in each layer for the target task. + +Motivated by the findings, for improving the recognition performance of cross-domain few-shot learning, we propose the Task-aware Adaptive Network (TA $^2$ -Net) to learn the optimal task-specific parameters policy for target tasks adaptively. As shown in Figure 2, our TA $^2$ -Net model consists of an Action Generation Network ("Agent") trained by Reinforcement Learning to generate "actions" - adaptive Task-specific Adapter execution decisions aimed at the target task, and an Adapted Network built on the "actions" to infer the target task, meanwhile, which can be viewed as the "Environment" to provide a "reward" for the Action Generation Network. The TA $^2$ -Net can learn optimally adapted network graphs aimed at the target task to transfer pre-learned knowledge efficiently. + +In summary, our contributions are: + +(i) We propose the Task-aware Adaptive Network (TA $^2$ - + +Net), which can adaptively learn optimal task-specific parameters policy for each target task. + +(ii) We evaluate our model on the Meta-dataset. Empirical results show that our model obviously outperforms existing state-of-the-art methods. +(iii) We further analyze parameter distribution and discover that domains with significant distribution shift (compared with the source domain) usually have a larger demand for task-specific parameters to effectively adapt pre-learned knowledge. Furthermore, layers in deep blocks necessitate more task-specific parameters than those in shallow blocks for learning task-specific features. + +# 2. Related Work + +Cross-domain Few-shot Learning Existing cross-domain few-shot learning methods can be coarsely grouped into two categories: feature-selection-based methods and adaptation-based methods. + +Several feature-selection-based methods, including SUR [7], URT [26], and URL [20], aim to achieve generalization by adaptively integrating feature representations from multiple training domains. SUR [7] and URT [26] rely on attention mechanisms to select appropriate domain-specific representations for a given few-shot learning task. While these methods have shown good performance, they require multiple forward networks during inference time. In contrast, Li et al. [20] proposed URL, an efficient method that distills knowledge from multiple domain networks to a single feature extractor. However, supervised training of multiple domains can result in supervision collapse, leading to the loss of information required for transfer to new tasks or domains. To address this issue, Doersch et al. [6] proposed using self-supervised learning to encourage the learning of general features from multiple training domains. + +The goal of adaptation-based methods is to adapt a feature extractor and classifier learned from a source domain to a target task through task-specific parameters. CNAPS [35] employed an adaptation network to generate task-specific parameters for each hidden layer of the feature extractor and classifier to transfer knowledge. Simple CNAPS [3] improved upon this approach by replacing the parametric classifier with a Mahalanobis distance-based distance metric, which increased classification accuracy while reducing the number of parameters. Triantafillou et al.[43] proposed the FLUTE, which learns domain-specific parameters using FiLM layers called universal templates. Li et al.[21] have demonstrated that attaching task-specific weights to a pretrained model and optimizing them from scratch directly on a small support set can enable good generalization to the target task. However, the shifts of feature distribution between different target tasks and the pre-learned domain vary, leading to the change of task-specific parameters demand. In this paper, we focus on adaptively generating task-specific + +![](images/903c83177ec845267659ab2d6306439b2a6aa747dae4cbb6980c93006391ab16.jpg) +Figure 2. Task-aware Adaptive Network (TA $^2$ -Net). The "Agent" is the Action Generation Network, which predicts the "action" of Task-specific Adapter execution and interacts with the "Environment" (Env, the Adapted Network) providing the "reward" for the Agent. "none" represent that the model does not need the Task-specific Adapter to adapt feature. + +graphs for each target task, to adapt the prior knowledge to the target task well. + +Reinforcement Learning Reinforcement learning [1, 8] is primarily used to solve the sequence decision problem, in which the reward value obtained from agent-environment interaction is used as a feedback signal to continuously optimize agent strategies. There are three main approaches to solving Reinforcement Learning (RL) problems: methods based on value functions [30, 46, 48], methods based on policy search [41, 40], and the actor-critic approach [15, 39]. The deep reinforcement learning algorithm based on value functions approximates the value function or action-value function with a deep neural network and updates it with the temporal difference (TD) learning or Q-learning method. Among them, Deep Q-Network (DQN) [30] algorithm is representative. Furthermore, a series of variants have been proposed, including double DQN (DDQN) [46], which optimizes the problem of overestimation about Q value in the DQN algorithm, and Dueling DQN [48], which improves DQN from the network structure and can estimate Q function more accurately. Policy-based methods [41, 40] directly search for an optimal policy without maintaining a value function model. Parameters of the chosen policy are updated using either gradient-based or gradient-free optimization to maximize the expected return. The actor-critic approach [15] employs both value functions and policy search to trade-off the variance reduction of policy gradients and bias introduction from value function methods, in which the "actor" (policy) learns by using feedback from the "critic" (value function) [1]. In this paper, we employ policy-gradient-based Reinforcement Learning to train + +the Action Generation Network to adaptively generate optimal actions – Task-specific Adapter execution decisions for each target task. + +# 3. Method + +In this section, we describe the problem setting and introduce our method. + +# 3.1. Preliminary + +Few-shot task Few-shot learning classification aims to classify a large number of samples with only a small number of labeled ones in each class. Labeled samples are generally referred to as the support set $S = \{(\pmb{x}_i, y_i)\}_{i=1}^{N_S}$ , while classified ones are known as the query set $Q = \{(\pmb{x}_i)\}_{i=1}^{N_Q}$ . $y_i$ is the label of the sample $\pmb{x}_i$ . A few-shot task is created by combining the support and query sets. + +Cross-domain few-shot learning Compared with traditional few-shot learning, cross-domain few-shot learning is more challenging, since not only the labeled samples in the target task $D_{t}$ are few, but also the distribution of the target domain and the base domain $D_{b}$ may be highly heterogeneous. Therefore, we must strike a balance between model capacity and the reliability of the adaptation while avoiding over-fitting to the few labeled samples or under-adaptation for the target task. + +# 3.2. Task-aware Adaptive Network (TA²-Net) + +Inspired by the observations that varying tasks require variant task-specific parameters for effective adaptation, we design the $\mathrm{TA}^2$ -Net adaptively learning the optimally + +adapted network graph for each target task. As shown in Figure 2, the $\mathrm{TA}^2$ -Net consists of Action Generation Network and Adapted Network. To directly optimize the Action Generation Network to generate optimal task-specific parameters for the Adapted Network, we cast the model in Reinforcement Learning. In particular, Action Generation Network can be viewed as an "Agent" that interacts with the "Environment" (Adapted Network). The Action Generation Network will result in the "actions" that are the prediction of Task-specific Adapter execution for the target task. Based on the generated "actions", we can build the unique Adapted Network, which will provide a "reward" for the agent - we denote this reward by $r$ . We will introduce the Action Generation Network, Adapted Network, and their optimization process. + +# 3.2.1 Action Generation Network + +Our focus in the Action Generation Network is to generate adaptive Task-specific Adapter execution decisions for each layer of the Adapted Network aimed at the target task. As shown in Figure 2, the Gate Network is integrated into each layer of the feature encoder $f_{\phi}(\cdot)$ pre-trained by the seen domain to generate the discrete probability for $K$ kinds of Task-specific Adapters. Concretely, the discrete probability in $l$ -th layer can be formulated as $\{z_l^i\}_{i=1}^K$ . Let $h_{l-1} \in R^{N_S \times C \times W \times H}$ denotes the input feature of $l$ -th feature encoder. In the Gate Network, the input feature is first fed to the Global Average Pooling (GAP) to compress the spatial dimension. Formally, the output $u_{l-1}$ can be achieved as + +$$ +\boldsymbol {u} _ {l - 1} = G A P \left(\boldsymbol {h} _ {l - 1}\right), \tag {1} +$$ + +where $u_{l-1}$ can be written as $u_{l-1} = \left[u_{l-1}^{1}, u_{l-1}^{2}, \dots, u_{l-1}^{N_{S}}\right]$ . $u_{l-1}^{i}$ represents the channel-level feature embedding of support sample $i$ and $u_{l-1}^{i} \in R^{C}$ . Then, let $v_{l-1} = \left[v_{l-1}^{1}, v_{l-1}^{2}, \dots, v_{l-1}^{N}\right]$ denotes the class prototypes in the target task, where $v_{l-1}^{n}$ refers to the class center of class $n$ and $N$ denotes the number of classes in the target task. Formally, the class center for class $n$ can be obtained by computing the mean vector of all embedded features in the class as + +$$ +\boldsymbol {v} _ {l - 1} ^ {n} = \frac {1}{| S _ {n} |} \sum_ {i \in S _ {n}} \boldsymbol {u} _ {l - 1} ^ {i}, \tag {2} +$$ + +where $S_{n}$ denotes the set of examples labeled with class $n$ . The output $\boldsymbol{v}_{l-1}$ is a collection of all class descriptors with statistics that express the entire target task. To capture the dependencies between features, we add a simple non-linear function of two fully-connected layers coupled with a ReLU activation function. The output of this operation is a vector containing unnormalized scores for executing the $K$ kinds + +of Task-specific Adapters. + +$$ +\bar {z} _ {l} = \mathbf {W} _ {l} ^ {2} \sigma \left(\mathbf {W} _ {l} ^ {1} v _ {l - 1}\right), \tag {3} +$$ + +where $\bar{z}_l\in R^K$ $\sigma$ refers to the $ReLU$ activation function. $\mathbf{W}_l^1\in R^{d\times C}$ $\mathbf{W}_l^2\in R^{K\times d}$ and $d$ is the dimension of the hidden layer. We can write $\bar{z}_l$ as $\bar{z}_l = [\bar{z}_l^1,\bar{z}_l^2,\dots ,\bar{z}_l^K ]$ Then, the Gate outputs a distribution over the $l$ -th Task-specific Adapter execution decisions $\pmb {z}_l$ using the Softmax function + +$$ +\boldsymbol {z} _ {\boldsymbol {\ell}} = \operatorname {s o f t m a x} (\bar {\boldsymbol {z}} _ {\boldsymbol {\ell}}), \tag {4} +$$ + +where $z_{l} = \left[z_{l}^{1},z_{l}^{2},\dots ,z_{l}^{K}\right]$ . $z_{l}^{k}$ denotes the possibility of selecting $k$ -th Task-specific Adapter in the $l$ -th layer. And $z = [z_{1},z_{2},\dots ,z_{L}]$ , $L$ is the number of layers in the backbone. + +Training The generation of Task-specific Adapter execution decisions is cast as the Reinforcement Learning problem. The Action Generation Network aims to decide the most appropriate Task-specific Adapters for conducting few-shot learning on the current tasks. We utilize the predicted result of the Adapted Network constructed based on the current "actions" as the rewards for optimizing the Action Generation Network. Inspired by the self-critical sequence training approach [34], the goal of training is to minimize the negative expected reward + +$$ +L (\theta) = - \left(r \left(\boldsymbol {z} ^ {s}\right) - r \left(\boldsymbol {z} ^ {m}\right)\right) \log \left(\boldsymbol {z} ^ {s}\right), \tag {5} +$$ + +where $\theta$ is the parameters of the Gate Network. $\theta = \{\mathbf{W}_l^1,\mathbf{W}_l^2\}$ and $l\in \{1,2,\dots ,L\}$ . $z^{s} = [z_{1}^{s},\dots,z_{L}^{s}]$ and $z_{l}^{s}$ is the possibility of action randomly sampled from the $K$ actions - Task-specific Adapters, at the layer $l$ . $r(z^{s})$ is the reward obtained by the Adapted Network, which is constructed by the actions with $z^{s}$ discrete possibility. $z^{m} = [z_{1}^{m},\dots,z_{L}^{m}]$ , and + +$$ +z _ {l} ^ {m} = \operatorname {a r g m a x} \left(\boldsymbol {z} _ {l}\right). \tag {6} +$$ + +where $\operatorname{argmax}(\cdot)$ returns the possibility of action equipped with maximum probability over the output distribution $z_{l}$ of the Gate in $l$ -th layer. $r(z^{m})$ again is the reference reward obtained in Adapted Network equipped with actions with $z^{m}$ discrete possibility. Accordingly, TA modules from the model that return a higher reward than $z^{m}$ will be "pushed up", or increased in probability, while TA modules that result in a lower reward will be suppressed. + +# 3.2.2 Adapted Network + +As shown in Figure 2, the network architecture of the Adapted Network consists of a series of residual blocks, which are formed by the Adaptive Module (AM) followed by the BN layer and ReLU function. In the Adaptive Module, a $3 \times 3$ conv is paralleled with the Task-specific Adapter, + +# Algorithm 1 Task-aware Adaptive Learning + +1: Train the backbone network $f_{\phi}(\cdot)$ on the source domain; +2: Build Action Generation Network $M_{act}$ on the backbone network; Initialization the parameters $\theta$ of Gate Networks; +3: while training do +4: Random sample task $T_{i}$ in the source domain; $T_{i} = S \bigcup Q, S = \{(x_{i},y_{i})\}_{i = 1}^{N_{S}}, Q = \{(x_{i})\}_{i = 1}^{N_{Q}}$ ; +5: $\pmb{z} = f_{\phi, \theta}(x), x \in S, \pmb{z} = [\pmb{z}_1, \pmb{z}_2, \dots, \pmb{z}_L]$ ; +6: $\pmb{z}^s = [Sample(\pmb{z}_1), \dots, Sample(\pmb{z}_L)]; \pmb{z}^m = [argmax(\pmb{z}_1), \dots, argmax(\pmb{z}_L)];$ +7: Build Adapted Network $M_{adapt}^{1}$ , $M_{adapt}^{2}$ based on the backbone network according to actions $z^{s}$ , $z^{m}$ respectively; +8: Initialization $\psi, \pi$ in $M_{adapt}^{1}$ and $M_{adapt}^{2}$ ; +9: Optimize $\psi$ , $\pi$ over $S$ based on $L_{adapt}(\psi, \pi) = \frac{1}{N_S} \sum_{(x,y) \in S} l(f_{\phi,\psi,\pi}(\boldsymbol{x}), y)$ ; +10: Predict over $Q$ ; $r(\pmb{z}^s) = N_{adapt}^1 / N_Q$ , $r(\pmb{z}^m) = N_{adapt}^2 / N_Q$ ; +11: Optimize the parameters $\theta$ of $M_{act}$ based on $L(\theta) = -(r(\pmb{z}^s) - r(\pmb{z}^m))\log (\pmb{z}^s)$ ; +12: end while + +Output: Action Generation Network $M_{act}$ + +which is the action predicted by the corresponding Gate Network of the Action Generation Network. + +Let $f_{\phi_l}(\cdot)$ denote the $3 \times 3$ conv in the $l$ -th Adaptive Module, which is pre-trained by the seen domain. With $h_{l-1} \in R^{N_S \times C \times W \times H}$ as the input feature of $l$ -th Adaptive Module, the output is $h_l = f_{\phi_l}(h_{l-1})$ when no Task-specific Adapter is selected according to the decision generated by the Action Generation Network, that is the feature does not need adaptation. When the $k$ -th Task-specific Adapter is chosen, which is constructed by paralleling $k$ task-specific layers in our design, the feature learned by the Task-specific Adapter will be combined with the feature extracted from the $3 \times 3$ conv as + +$$ +\boldsymbol {h} _ {\boldsymbol {l}} = f _ {\phi_ {\boldsymbol {l}}} \left(\boldsymbol {h} _ {\boldsymbol {l} - \mathbf {1}}\right) + \frac {1}{k} \sum_ {i = 1} ^ {k} f _ {\psi_ {\boldsymbol {l}} ^ {i}} \left(\boldsymbol {h} _ {\boldsymbol {l} - \mathbf {1}}\right), \tag {7} +$$ + +where $f_{\psi_l^i}(\cdot)$ denotes the $i$ -th task-specific layer in the Task-specific Adapter, which is $1 \times 1$ conv in the experiments. + +Training Upon the generated "actions" with the discrete possibility $z^s$ and $z^m$ in the Action Generation Network, we will construct two Adapted Networks $M_{adapt}^{1}$ and $M_{adapt}^{2}$ . In training, the Adapted Networks are trained with the cross-entropy loss over the support samples: + +$$ +L _ {a d a p t} (\psi , \pi) = \frac {1}{N _ {S}} \sum_ {(x, y) \in S} l \left(f _ {\phi , \psi , \pi} (\boldsymbol {x}), y\right), \tag {8} +$$ + +where $f_{\psi, \pi}(\cdot)$ denotes the output softmax probability vector of the linear classifier with parameters $\pi$ , in which dimensionality equals the number of categories in the support set $S$ . Then, the reward for the agent is achieved by predicting the Adapted Network over the query samples $Q$ : + +$$ +r \left(\boldsymbol {z} ^ {\boldsymbol {s}}\right) = \frac {N _ {\text {a d a p t}} ^ {1}}{N _ {Q}} \quad \text {a n d} \quad r \left(\boldsymbol {z} ^ {\boldsymbol {m}}\right) = \frac {N _ {\text {a d a p t}} ^ {2}}{N _ {Q}}. \tag {9} +$$ + +where $N_{adapt}^{1}$ and $N_{adapt}^{2}$ denote the number of query samples correctly predicted by Adapted Networks $M_{adapt}^{1}$ and $M_{adapt}^{2}$ respectively. We show the Task-aware Adaptive Learning pseudo code in Algorithm 1. + +# 3.2.3 Testing + +Note that the support samples of testing will first go through the Action Generation Network to generate optimal Task-specific Adapter executions for the target task, which are the action with maximum probability. And then, based on the executions, the Adapted Network will be built and finetuned over the support set. Finally, the query samples will be predicted by the Adapted Network. + +# 4. Experiments + +In this part, we first introduce the experimental setup and then compare our method with the state-of-the-art methods. Furthermore, we provide a series of analyses of our method. + +# 4.1. Experimental setup + +Dataset. Meta-Dataset [44] is the standard benchmark for evaluating the performance of few-shot classification. It consists of 13 image datasets: ILSVRC 2012 [36], Omniglot [18], FGVC-Aircraft [29], CUB-200-2011 [47], Describable Textures [5], Quick Draw [11], FGVCx Fungi [38], VGG Flower [31], Traffic Signs [13], MSCOCO [24], MNIST [19], CIFAR-10 [16], and CIFAR-100 [16]. For all datasets in the experiments, we used the standard split procedure as [44]. + +Implementation details. In the experiments, we considered two training settings: multi-domain learning and single-domain setting. In the multi-domain setting, the first eight datasets (ILSVRC 2012, Omniglot, Aircraft, CUB-200-2011, Textures, Quick Draw, Fungi, and VGG Flower) of the Meta-dataset are used to train the feature extractor. We followed the method in [20] to obtain the pre-trained model. And in the single-domain setting, only the ILSVRC 2012 dataset participated in the training of the feature extractor. As in [20], we adopted the ResNet-18 as the backbone no matter in the multi-domain setting or the single-domain setting. In the training, the $\mathrm{TA}^2$ -Net is trained using the Adam optimizer. The hyper-parameters $K$ are set as $K = 3$ . We also provide the ablation study in Section 4.4. + +Table 1. Comparison to state-of-the-art methods in the multi-domain setting. Mean accuracy, $95\%$ confidence interval are reported. The best results are marked in bold. + +
Test DatasetCNAPS [35]Simple CNAPS [3]Transductive CNAPS [2]SUR [7]URT [26]FLUTE [43]tri-M [27]URL [20]TSA [21]Ours
ImageNet50.8 ± 1.158.4 ± 1.157.9 ± 1.156.2 ± 1.056.8 ± 1.158.6 ± 1.051.8 ± 1.158.8 ± 1.159.5 ± 1.059.6 ± 1.0
Omniglot91.7 ± 0.591.6 ± 0.694.3 ± 0.494.1 ± 0.494.2 ± 0.492.0 ± 0.693.2 ± 0.594.5 ± 0.494.9 ± 0.495.5 ± 0.4
Aircraft83.7 ± 0.682.0 ± 0.784.7 ± 0.585.5 ± 0.585.8 ± 0.582.8 ± 0.787.2 ± 0.589.4 ± 0.489.9 ± 0.490.5 ± 0.4
Birds73.6 ± 0.974.8 ± 0.978.8 ± 0.771.0 ± 1.076.2 ± 0.875.3 ± 0.879.2 ± 0.880.7 ± 0.881.1 ± 0.881.4 ± 0.8
Textures59.5 ± 0.768.8 ± 0.966.2 ± 0.871.0 ± 0.871.6 ± 0.771.2 ± 0.868.8 ± 0.877.2 ± 0.777.5 ± 0.777.4 ± 0.7
Quick Draw74.7 ± 0.876.5 ± 0.877.9 ± 0.681.8 ± 0.682.4 ± 0.677.3 ± 0.779.5 ± 0.782.5 ± 0.681.7 ± 0.682.5 ± 0.6
Fungi50.2 ± 1.146.6 ± 1.048.9 ± 1.264.3 ± 0.964.0 ± 1.048.5 ± 1.058.1 ± 1.168.1 ± 0.966.3 ± 0.866.3 ± 0.9
VGG Flower88.9 ± 0.590.5 ± 0.592.3 ± 0.482.9 ± 0.887.9 ± 0.690.5 ± 0.591.6 ± 0.692.0 ± 0.592.2 ± 0.592.6 ± 0.4
Traffic Sign56.5 ± 1.157.2 ± 1.059.7 ± 1.151.0 ± 1.148.2 ± 1.163.0 ± 1.058.4 ± 1.163.3 ± 1.182.8 ± 1.087.4 ± 0.8
MSCOCO39.4 ± 1.048.9 ± 1.142.5 ± 1.152.0 ± 1.151.5 ± 1.152.8 ± 1.150.0 ± 1.057.3 ± 1.057.6 ± 1.057.9 ± 0.9
MNIST-94.6 ± 0.494.7 ± 0.394.3 ± 0.490.6 ± 0.596.2 ± 0.395.6 ± 0.594.7 ± 0.496.7 ± 0.497.0 ± 0.4
CIFAR-10-74.9 ± 0.773.6 ± 0.766.5 ± 0.967.0 ± 0.875.4 ± 0.878.6 ± 0.774.2 ± 0.882.9 ± 0.782.1 ± 0.8
CIFAR-100-61.3 ± 1.161.8 ± 1.056.9 ± 1.157.3 ± 1.062.0 ± 1.067.1 ± 1.063.5 ± 1.070.4 ± 0.970.9 ± 0.9
Average Seen71.673.775.175.977.474.576.280.480.480.7
Average Unseen-67.466.564.162.969.969.970.678.179.1
Average All-71.271.871.471.872.773.876.679.580.1
Average Rank-7.66.27.26.55.85.53.22.21.2
+ +# 4.2. Comparison to state-of-the-art methods + +Following the experiment set in [21], we sampled tasks with varying numbers of ways and shots in the inference and conducted experiments in the multi-domain setting and single-domain setting. + +Multi-domain setting. Here, the results of our method on 13 datasets of the Meta-Dataset are compared with the state-of-the-art methods. We also report average accuracy over seen domains, and unseen domains, as well as the average rank as in [21]. From the results in Table 1, our method obtains state-of-the-art results on 10 out of 13 datasets and achieves the best average accuracy on seen and unseen domains. It is worth noting that our method obtains significantly better average results than the second-best approach on the unseen datasets $(+1.0\%)$ . The unseen datasets are heterogeneous with respect to the training domain, with a significant domain gap between them. In contrast to adopting the fixed number of task-specific parameters for all target tasks, e.g. CNAPS, Simple CNAPS, FLUTE, and TSA, the $\mathrm{TA}^2$ -Net can adaptively learn the optimal task-specific parameter policies for unseen target tasks to efficiently transfer the pre-trained knowledge to the target feature distribution. + +Single-domain setting. We also evaluate our method in the single-domain setting, which is more challenging due to optimizing the Action Generation Network only on the ImageNet dataset. Results are shown in Table 2. First, compared with other single-domain training methods, our + +method achieves significant superiority. For example, $\mathrm{TA}^2$ -Net outperforms the most competing method (TSA [21]) by $+2.9\%$ , $+0.8\%$ , $+1.6\%$ , and $+1\%$ on Omniglot, Quick Draw, Traffic Signs and MNIST, respectively. Second, it is worth noting that, for the average accuracy on the unseen datasets, ours outperforms the second-best approach with a clear margin $(+3.1\%)$ . And we clearly observe that $\mathrm{TA}^2$ -Net achieves competitive accuracy on the ImageNet, compared with the best performance. Despite being trained on just one dataset, the Action Generation Network exhibits good generalization, producing the optimal task-specific parameter policies for the unseen domain in an excellently adaptive manner. + +# 4.3. Further Analysis + +Varying-way Five-shot setting. The number of support samples ('shot') in the standard Meta-Dataset evaluation setting is variable and can range up to 100. In this section, we evaluate our method in the 5-shot setting which is more challenging due to fewer labeled samples and can better reflect the progressiveness of the model. Table 3 shows the results compared with other state-of-the-art methods. As can be seen, ours achieves better average accuracy on both seen and unseen domains. More specifically, our method outperforms other methods on the vast majority of domains (8 out of 13) and obtains the best average rank. Furthermore, the average accuracy on unseen datasets is $1.2\%$ higher than the second-best method. + +Table 2. Comparison to state-of-the-art methods in the single-domain setting. Mean accuracy, $95\%$ confidence interval are reported. The best results are marked in bold. + +
Test DatasetProtoNet [44]fo-Proto -MAML [44]ALFA+fo-Proto -MAML [44]BOHB [37]FLUTE [43]TSA [21]Ours
ImageNet50.5 ± 1.149.5 ± 1.152.8 ± 1.151.9 ± 1.146.9 ± 1.159.5 ± 1.159.3 ± 1.1
Omniglot60.0 ± 1.463.4 ± 1.361.9 ± 1.567.6 ± 1.261.6 ± 1.478.2 ± 1.281.1 ± 1.1
Aircraft53.1 ± 1.056.0 ± 1.063.4 ± 1.154.1 ± 0.948.5 ± 1.072.2 ± 1.072.6 ± 0.9
Birds68.8 ± 1.068.7 ± 1.069.8 ± 1.170.7 ± 0.947.9 ± 1.074.9 ± 0.975.1 ± 0.9
Textures66.6 ± 0.866.5 ± 0.870.8 ± 0.968.3 ± 0.863.8 ± 0.877.3 ± 0.776.8 ± 0.8
Quick Draw49.0 ± 1.151.5 ± 1.059.2 ± 1.250.3 ± 1.057.5 ± 1.067.6 ± 0.968.4 ± 0.9
Fungi39.7 ± 1.140.0 ± 1.141.5 ± 1.241.4 ± 1.131.8 ± 1.044.7 ± 1.045.3 ± 1.0
VGG Flower85.3 ± 0.887.2 ± 0.786.0 ± 0.887.3 ± 0.680.1 ± 0.990.9 ± 0.691.0 ± 0.6
Traffic Sign47.1 ± 1.148.8 ± 1.160.8 ± 1.351.8 ± 1.046.5 ± 1.182.5 ± 0.884.1 ± 0.7
MSCOCO41.0 ± 1.143.7 ± 1.148.1 ± 1.148.0 ± 1.041.4 ± 1.059.0 ± 1.058.0 ± 1.0
MNIST----80.8 ± 0.893.9 ± 0.694.9 ± 0.5
CIFAR-10----65.4 ± 0.882.1 ± 0.782.0 ± 0.7
CIFAR-100----52.7 ± 1.170.7 ± 0.970.8 ± 0.9
Average Seen50.549.552.851.946.959.559.3
Average Unseen56.758.462.460.053.271.975.0
Average All56.157.561.459.252.670.773.8
Average Rank5.95.23.44.15.61.71.3
+ +![](images/b57a8cb11bf2b59d7109433dfd1dd3ce5f1e80ff36855cb1a58625f6ff2b1757.jpg) +Figure 3. Performance of widening or deepening the Task-specific Adapter (TA) in the multi-domain setting. The axes indicate the accuracy of methods on a particular dataset. Each color corresponds to a different method. + +Wider or deeper in the Task-specific Adapter (TA)? To explore the method for improving the adaptation of the model, we increase the parameters of the Task-specific Adapter (TA) in wide or depth respectively. Specifically, two types of TAs are designed: (a) $1 \times 1$ Conv in parallel, (b) $1 \times 1$ Conv in series. As shown in Figure 3, increasing the width of the TA allows the model to achieve better performance than deepening. In particular, among the 13 datasets, 5 datasets (Traffic Signs, MSCOCO, MNIST, CIFAR-10, and CIFAR-100) unseen in the training obviously benefited more from widening the Task-specific Adapter (TA). All of the above indicates that the best way to improve adaptability is to increase the width of the Task-specific Adapter (TA). + +Effectiveness of adaptive learning. In this section, we attempt to provide a more in-depth discussion of the efficacy of fixed networks as well as the proposed adaptive learned network. Figure 4 depicts the performance over the + +Table 3. Results for Varying-Way Five-Shot setting. Mean accuracy, $95\%$ confidence interval are reported. The best results are marked in bold. + +
Test DatasetSimple CNAPS [3]SUR [7]URT [26]URL [20]TSA [21]Ours
ImageNet47.2 ± 1.046.7 ± 1.048.6 ± 1.049.4 ± 1.048.3 ± 1.049.3 ± 1.0
Omniglot95.1 ± 0.395.8 ± 0.396.0 ± 0.396.0 ± 0.396.8 ± 0.396.6 ± 0.2
Aircraft74.6 ± 0.682.1 ± 0.681.2 ± 0.684.8 ± 0.585.5 ± 0.585.9 ± 0.4
Birds69.6 ± 0.762.8 ± 0.971.2 ± 0.776.0 ± 0.676.6 ± 0.677.3 ± 0.6
Textures57.5 ± 0.760.2 ± 0.765.2 ± 0.769.1 ± 0.668.3 ± 0.768.3 ± 0.6
Quick Draw70.9 ± 0.679.0 ± 0.579.2 ± 0.578.2 ± 0.577.9 ± 0.678.5 ± 0.5
Fungi50.3 ± 1.066.5 ± 0.866.9 ± 0.970.0 ± 0.870.4 ± 0.870.3 ± 0.8
VGG Flower86.5 ± 0.476.9 ± 0.682.4 ± 0.589.3 ± 0.489.5 ± 0.490.0 ± 0.4
Traffic Sign55.2 ± 0.844.9 ± 0.945.1 ± 0.957.5 ± 0.872.3 ± 0.676.7 ± 0.5
MSCOCO49.2 ± 0.848.1 ± 0.952.3 ± 0.956.1 ± 0.856.0 ± 0.856.0 ± 0.8
MNIST88.9 ± 0.490.1 ± 0.486.5 ± 0.589.7 ± 0.492.5 ± 0.493.3 ± 0.3
CIFAR-1066.1 ± 0.750.3 ± 1.061.4 ± 0.766.0 ± 0.772.0 ± 0.773.1 ± 0.7
CIFAR-10053.8 ± 0.946.4 ± 0.952.5 ± 0.957.0 ± 0.964.1 ± 0.864.1 ± 0.8
Average Seen69.071.273.876.676.777.0
Average Unseen62.656.059.665.271.472.6
Average All66.565.468.372.274.675.3
Average Rank5.254.12.82.21.5
+ +![](images/1f3ce285afcd3d5a16cf39dcd2bb0f416de8fe60500b3a5195099c0604f3d60f.jpg) +Figure 4. Effectiveness of adaptive learning. + +Meta-dataset in the multi-domain and single-domain setting, with Model-1/2/3 constructed by attaching Module-1/2/3 respectively as shown in Figure 1 to the backbone. It can be observed that, in comparison to the backbone network, the average accuracy of seen domains in both the multi-domain and single-domain settings decreased when using the fixed Task-specific Adapter to adaptation, but our method performed better. Furthermore, average accuracy on unseen domains has consistently improved overall in the fixed Task-specific Adapter and can be improved further through adaptive learning. This discovery inspires us to recognize that simply designing a fixed Task-specific Adapter does not work and forces us to recognize the necessity of adaptively learning optimal task-specific parameters policy for each target task. + +Effectiveness of Reinforcement Learning. Our $\mathrm{TA}^2$ -Net model adopts the Reinforcement Learning algorithm combined with Episodic Training to adaptively learn opti + +![](images/9bfdfb265709122d04a4792b62a6440a9c6df301b20ae31fc435066e90d430d1.jpg) +(a) MMD and TA parameters of different datasets in the Meta-Dataset + +![](images/dbf31d7b4fad8267be1b2013ca0b0319d2c49f5d90f327647149a1ed29016bcf.jpg) +(b) TA parameters distribution in different layers for the Fungi dataset + +![](images/4cc7c970ac9a734ebf14b9c1d9c46c20d0af5d9edce2582e38dd5eaff4a0927d.jpg) +(c) TA parameters distribution in different layers for the CIFAR-10 dataset +Figure 5. (a): The Line chart shows the mean and standard deviation of TA parameters adaptively learned by the $\mathrm{TA}^2$ -Net in target tasks sampled in different datasets. And the box plot shows the $25\%$ quantiles of the MMD in different datasets while the notches indicate the median. (b) and (c): TA parameters distribution in different blocks of the task-aware adapted network for the Fungi and CIFAR-10 dataset. Sub-bars in the block denote different layers within the Block. + +mal task-specific parameters policy for target tasks. To explore the effectiveness of Reinforcement Learning, we designed the Simple $\mathrm{TA}^2$ -Net trained by Episodic Training, which only includes the Adapted Network attaching a Gate Network in each layer for generating "decisions" - Chosen the Task-specific Adapter. As shown in Table 4, the $\mathrm{TA}^2$ -Net optimized with the Reinforcement Learning outperforms the Simple $\mathrm{TA}^2$ -Net, which shows that Reinforcement Learning can better learn the TA decisions for target tasks. + +Parameters distribution. This is an intuitive guess that the greater the difference (domain shift) between the source domain and the target domain features, the greater the demand for TA parameters. In this section, we first observe the relationship between the TA parameters and the domain shift. As widely used in previous works [28, 25], we adopt the Maximum Mean Discrepancy (MMD) to quantify the domain shift between the target task features and the source domain features extracted by the pre-trained backbone (in the multi-domain setting). The larger MMD indicates a larger domain shift between the test task and the source domain. As shown in Figure 5 (a), the Line chart and box plot shows the TA parameters and MMD distribution in different datasets, respectively. We can find that MMD for each dataset has a narrow distribution. And, on the whole, datasets with significant MMD have a larger demand for the TA parameters. Furthermore, linear regression and Pearson correlation analysis revealed that MMD is positively correlated with TA parameters, with the Pearson correlation coefficient $R = 0.9186$ . It is attributed to the greater heterogeneity, more task-specific parameters are required to effectively adapt pre-learned knowledge to the feature distribution of target tasks. We further explore the TA parameters distribution in different layers of the task-aware adapted network learned by our proposed TA $^2$ -Net. As shown in Figure 5 (b) and (c), layers in the deep blocks have more task-specific parameters for learning the task-specific fea + +Table 4. Effectiveness of Reinforcement Learning in the multidomain setting. + +
ModelReinforcement LearningSeen DomainUnseen DomainAll
Simple TA2-NetX75.567.972.6
TA2-Net✓77.072.675.3
+ +Table 5. Mean accuracy $(\%)$ for different number of layers $K$ in the Task-specific Adapter. + +
K=1K=2K=3K=4
Seen Domain75.475.275.375.1
Unseen Domain66.670.370.568.2
All72.073.373.572.5
+ +tures than those in the shallow blocks as demonstrated by both Fungi and CIFAR-10, which may be because the shallow layer contains more general information while the deep layer contains more domain-specific information [50]. + +# 4.4. Ablation Study + +Maximum width $K$ of the Task-specific Adapter. Here, we discuss the maximum width $K$ of the Task-specific Adapter, which represents the upper bound of the task-specific parameters. Experiments are conducted on the Meta-dataset in the varying-way five-shot setting. Table 5 shows the performances of $K \in [1,2,3,4]$ . We can observe that, as $K$ increases, the average classification accuracy on seen domains lightly decreases. For unseen domains, the average classification accuracy increases at first and then decreases as $K$ increases, with $K = 3$ producing the best results. Compared with seen domains, unseen domains are heterogeneous with the source domain and require more task-specific parameters to transfer the pre-learned knowledge. The average performance becomes poorer when using $K = 4$ , which may be due to too much TA, the model + +overfitting, or the search space of the model structure expanding, increasing the optimization difficulty of reinforcement learning. Considering the performance of various target tasks and the calculating costs, all experiments in this paper adopt $K = 3$ . + +# 5. Conclusion + +In this paper, for studying cross-domain few-shot learning problems, we explored the performance of test tasks when inference was performed in models equipped with a variant number of task-specific parameters. The exploration revealed that target tasks require variant task-specific parameters for outstanding adaptation. Motivated by this, we proposed a Task-aware Adaptive Network (TA $^2$ -Net), which can adaptively learn optimal task-specific parameters policy for each target task for exhaustively improving the recognition performance of cross-domain few-shot learning. By conducting extensive experiments, we proved that our TA $^2$ -Net could achieve the best results on Meta-dataset. + +# Acknowledgment + +This work was supported in part by Beijing Natural Science Foundation Project No. Z200002, in part by National Natural Science Foundation of China (NSFC) No. U19B2036, 62225601, in part by the Program for Youth Innovative Research Team of BUPT No. 2023QNTD02, in part by scholarships from China Scholarship Council (CSC) under Grant CSC No. 202206470055, in part by BUPT Excellent Ph.D. Students Foundation No. CX2022152. + +# References + +[1] Kai Arulkumaran, Marc Peter Deisenroth, Miles Brundage, and Anil Anthony Bharath. Deep reinforcement learning: A brief survey. IEEE Signal Processing Magazine, 34(6):26-38, 2017. 3 +[2] Peyman Bateni, Jarred Barber, Jan-Willem van de Meent, and Frank Wood. Enhancing few-shot image classification with unlabelled examples. In WACV, pages 2796-2805, 2022. 6 +[3] Peyman Bateni, Raghav Goyal, Vaden Masrani, Frank Wood, and Leonid Signal. Improved few-shot visual classification. In CVPR, pages 14493-14502, 2020. 1, 2, 6, 7 +[4] Luca Bertinetto, Joao F. Henriques, Philip Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In ICLR, 2019. 1 +[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5 +[6] Carl Doersch, Ankush Gupta, and Andrew Zisserman. CrossTransformers: spatially-aware few-shot transfer. NeurIPS, 33:21981-21993, 2020. 1, 2 +[7] Nikita Dvornik, Cordelia Schmid, and Julien Mairal. Selecting relevant features from a multi-domain representation for + +few-shot classification. In ECCV, pages 769-786, 2020. 1, 2, 6, 7 +[8] Sven Gronauer and Klaus Diepold. Multi-agent deep reinforcement learning: a survey. Artificial Intelligence Review, 55(2):895-943, 2022. 3 +[9] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In ECCV, pages 124-141, 2020. 1 +[10] Yurong Guo, Ruoyi Du, Xiaoxu Li, Jiyang Xie, Zhanyu Ma, and Yuan Dong. Learning calibrated class centers for few-shot classification by pair-wise similarity. IEEE TIP, 31:4543-4555, 2022. 1 +[11] David Ha and Douglas Eck. A Neural Representation of Sketch Drawings. arXiv preprint arXiv:1704.03477, 2017. 5 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1 +[13] Sebastian Houben, Johannes Stallkamp, Jan Salmen, Marc Schlipsing, and Christian Igel. Detection of traffic signs in real-world images: The german traffic sign detection benchmark. In IJCNN, pages 1-8, 2013. 5 +[14] Yanxu Hu and Andy J. Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, pages 20-37, 2022. 1 +[15] Vijay Konda and John Tsitsiklis. Actor-critic algorithms. NeurIPS, 12, 1999. 3 +[16] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5 +[17] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 1 +[18] Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 350(6266):1332-1338, 2015. 5 +[19] Yann LeCun, Corinna Cortes, and CJ Burges. MNIST handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, 2, 2010. 5 +[20] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple domains for few-shot classification. In ICCV, pages 9526-9535, 2021. 1, 2, 5, 6, 7 +[21] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Cross-domain few-shot learning with task-specific adapters. In CVPR, pages 7161-7170, 2022. 2, 6, 7 +[22] Xiaoxu Li, Jijie Wu, Zhuo Sun, Zhanyu Ma, Jie Cao, and Jing-Hao Xue. BSNet: Bi-similarity network for few-shot fine-grained image classification. IEEE TIP, 30:1318-1331, 2020. 1 +[23] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. MetaSGD: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 1 +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 5 + +[25] Ge Liu, Linglan Zhao, and Xiangzhong Fang. PDA: Proxy-based domain adaptation for few-shot image recognition. Image and Vision Computing, 110:104164, 2021. 8 +[26] Lu Liu, William Hamilton, Guodong Long, Jing Jiang, and Hugo Larochelle. A universal representation transformer layer for few-shot image classification. In ICLR, 2021. 1, 2, 6, 7 +[27] Yanbin Liu, Juho Lee, Linchao Zhu, Ling Chen, Humphrey Shi, and Yi Yang. A multi-mode modulator for multi-domain few-shot classification. In ICCV, pages 8453-8462, 2021. 6 +[28] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105, 2015. 8 +[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5 +[30] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. 3 +[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5 +[32] Hang Qi, Matthew Brown, and David G. Lowe. Low-Shot learning with imprinted weights. In CVPR, pages 5822-5830, 2018. 1 +[33] Sachin Ravi and H. Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 1 +[34] Steven J Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel. Self-critical sequence training for image captioning. In CVPR, pages 7008-7024, 2017. 4 +[35] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, and Richard E Turner. Fast and flexible multi-task classification using conditional neural adaptive processes. NeurIPS, 32, 2019. 1, 2, 6 +[36] Olga Russakovsky, J. Deng, Hao Su, J. Krause, S. Satheesh, S. Ma, Zhiheng Huang, A. Karpathy, A. Khosla, Michael S. Bernstein, A. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 1, 5 +[37] Tonmoy Saikia, Thomas Brox, and Cordelia Schmid. Optimized generic feature learning for few-shot classification across domains. arXiv preprint arXiv:2001.07926, 2020. 7 +[38] Brigit Schroeder and Yin Cui. FGVCx Fungi Classification Challenge 2018. github.com/visipedia/fgvcx_fungi_comp, 2018.5 +[39] John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438, 2015. 3 +[40] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3 +[41] David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. In ICML, pages 387-395, 2014. 3 + +[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1 +[43] Eleni Triantafillou, Hugo Larochelle, Richard Zemel, and Vincent Dumoulin. Learning a universal template for few-shot dataset generalization. In ICML, pages 10424–10433, 2021. 2, 6, 7 +[44] Eleni Triantafillou, Tyler Zhu, Vincent Dumoulin, Pascal Lamblin, Utku Evci, Kelvin Xu, Ross Goroshin, Carles Gelada, Kevin Swersky, Pierre-Antoine Manzagol, et al. Meta-dataset: A dataset of datasets for learning to learn from few examples. arXiv preprint arXiv:1903.03096, 2019. 5, 7 +[45] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In ICLR, 2020. 1 +[46] Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-learning. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016. 3 +[47] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5 +[48] Ziyu Wang, Tom Schaul, Matteo Hessel, Hado Hasselt, Marc Lanctot, and Nando Freitas. Dueling network architectures for deep reinforcement learning. In ICML, pages 1995–2003, 2016. 3 +[49] Davis Wertheimer, Luming Tang, and Bharath Hariharan. Few-shot classification with feature map reconstruction networks. In CVPR, pages 8012-8021, 2021. 1 +[50] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 8 \ No newline at end of file diff --git a/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/images.zip b/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..26eee2965d93c12713c8368f4abe18fc505d3c36 --- /dev/null +++ b/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fab1705a018f677142f729dca9c73666ea5dd19f934697f114a22d1ac8c37d60 +size 595651 diff --git a/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/layout.json b/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bfacaa4332e2865fc6d41715058f0968a4cdb1ec --- /dev/null +++ b/2023/Task-aware Adaptive Learning for Cross-domain Few-shot Learning/layout.json @@ -0,0 +1,10391 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 509, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 509, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 509, + 121 + ], + "type": "text", + "content": "Task-aware Adaptive Learning for Cross-domain Few-shot Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "spans": [ + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "text", + "content": "Yurong Guo" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "text", + "content": ", Ruoyi Du" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "text", + "content": ", Yuan Dong" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "text", + "content": ", Timothy Hospedales" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "text", + "content": ", Yi-Zhe Song" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "text", + "content": ", Zhanyu Ma" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "text", + "content": "Beijing University of Posts and Telecommunications, China \n" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "text", + "content": "University of Edinburgh, UK \n" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 142, + 518, + 185 + ], + "type": "text", + "content": "University of Surrey, UK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 154, + 187, + 440, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 187, + 440, + 213 + ], + "spans": [ + { + "bbox": [ + 154, + 187, + 440, + 213 + ], + "type": "text", + "content": "{guoyurong, duruoyi, yuandong, mazhanyu}@bupt.edu.cn, t.hospedales@ed.ac.uk, y.song@surrey.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 266, + 290, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 565 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 565 + ], + "type": "text", + "content": "Although existing few-shot learning works yield promising results for in-domain queries, they still suffer from weak cross-domain generalization. Limited support data requires effective knowledge transfer, but domain-shift makes this harder. Towards this emerging challenge, researchers improved adaptation by introducing task-specific parameters, which are directly optimized and estimated for each task. However, adding a fixed number of additional parameters fails to consider the diverse domain shifts between target tasks and the source domain, limiting efficacy. In this paper, we first observe the dependence of task-specific parameter configuration on the target task. Abundant task-specific parameters may over-fit, and insufficient task-specific parameters may result in under-adaptation – but the optimal task-specific configuration varies for different test tasks. Based on these findings, we propose the Task-aware Adaptive Network (TA" + }, + { + "bbox": [ + 46, + 266, + 290, + 565 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 266, + 290, + 565 + ], + "type": "text", + "content": "-Net), which is trained by reinforcement learning to adaptively estimate the optimal task-specific parameter configuration for each test task. It learns, for example, that tasks with significant domain-shift usually have a larger need for task-specific parameters for adaptation. We evaluate our model on Meta-dataset. Empirical results show that our model outperforms existing state-of-the-art methods. Our code is available at https://github.com/PRIS-CV/TA2-Net." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 589, + 128, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 589, + 128, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 589, + 128, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "content": "Traditional deep learning models [42, 36, 12, 17] show excellent generalization performance when training on a large number of labeled samples. However, both abundant samples and reliable annotations are not always available in realistic applications, e.g., rare disease diagnosis, and fine-grained recognition. Few-shot learning [32, 33, 23, 4, 10], inspired by the fact that humans can quickly learn new" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 241, + 446, + 297 + ], + "blocks": [ + { + "bbox": [ + 310, + 241, + 446, + 297 + ], + "lines": [ + { + "bbox": [ + 310, + 241, + 446, + 297 + ], + "spans": [ + { + "bbox": [ + 310, + 241, + 446, + 297 + ], + "type": "image", + "image_path": "a5f61152cc6709e9fbbfc88f6be0b037b9db09271266362c3116aafc075c2f8a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 297, + 418, + 303 + ], + "lines": [ + { + "bbox": [ + 350, + 297, + 418, + 303 + ], + "spans": [ + { + "bbox": [ + 350, + 297, + 418, + 303 + ], + "type": "text", + "content": "(a) Different tasks from MSCOCO" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 449, + 243, + 539, + 297 + ], + "blocks": [ + { + "bbox": [ + 449, + 243, + 539, + 297 + ], + "lines": [ + { + "bbox": [ + 449, + 243, + 539, + 297 + ], + "spans": [ + { + "bbox": [ + 449, + 243, + 539, + 297 + ], + "type": "image", + "image_path": "ab4c6f2ce3a7baae99c0ab79ad9c702417e93954cbaa278aace44d8d178cf7e7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 297, + 539, + 304 + ], + "lines": [ + { + "bbox": [ + 451, + 297, + 539, + 304 + ], + "spans": [ + { + "bbox": [ + 451, + 297, + 539, + 304 + ], + "type": "text", + "content": "(c) Different layers in a task from MSCOCO" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 303, + 446, + 357 + ], + "blocks": [ + { + "bbox": [ + 312, + 303, + 446, + 357 + ], + "lines": [ + { + "bbox": [ + 312, + 303, + 446, + 357 + ], + "spans": [ + { + "bbox": [ + 312, + 303, + 446, + 357 + ], + "type": "image", + "image_path": "6f010ea80287bfdd4011a9b1f033015060138d6a5c34b5e6733ff7d306230aac.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 359, + 419, + 365 + ], + "lines": [ + { + "bbox": [ + 348, + 359, + 419, + 365 + ], + "spans": [ + { + "bbox": [ + 348, + 359, + 419, + 365 + ], + "type": "text", + "content": "(b) Different tasks from ilsvrc 2012" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 449, + 306, + 539, + 357 + ], + "blocks": [ + { + "bbox": [ + 449, + 306, + 539, + 357 + ], + "lines": [ + { + "bbox": [ + 449, + 306, + 539, + 357 + ], + "spans": [ + { + "bbox": [ + 449, + 306, + 539, + 357 + ], + "type": "image", + "image_path": "37227961c6896da2ff9661fa2241f02fcc369060230080f2ef39cf143b9f9c18.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 450, + 359, + 541, + 365 + ], + "lines": [ + { + "bbox": [ + 450, + 359, + 541, + 365 + ], + "spans": [ + { + "bbox": [ + 450, + 359, + 541, + 365 + ], + "type": "text", + "content": "(d) Different layers in a task from ilsvrc_2012" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 310, + 369, + 356, + 420 + ], + "blocks": [ + { + "bbox": [ + 310, + 369, + 356, + 420 + ], + "lines": [ + { + "bbox": [ + 310, + 369, + 356, + 420 + ], + "spans": [ + { + "bbox": [ + 310, + 369, + 356, + 420 + ], + "type": "image", + "image_path": "5eb0ed051a37115f28ea7fb9a4414fa12d667abe339c130eaca44efffb10750d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 429, + 545, + 495 + ], + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 495 + ], + "type": "text", + "content": "Figure 1. (a) and (b) reveal the accuracy of varying test tasks when inference was performed in different models, in which TA-Modules are attached to each layer in the backbone network. (c) and (d) show the accuracy when TA-Modules are attached to different layers in the backbone network. Four dashed boxes show the backbone and several TA-Modules." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 361, + 371, + 403, + 418 + ], + "blocks": [ + { + "bbox": [ + 361, + 371, + 403, + 418 + ], + "lines": [ + { + "bbox": [ + 361, + 371, + 403, + 418 + ], + "spans": [ + { + "bbox": [ + 361, + 371, + 403, + 418 + ], + "type": "image", + "image_path": "00d37dc4197337b48a856d0070ce01391596ca45c2482ff87e350ed8bd6f63d4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 408, + 371, + 471, + 418 + ], + "blocks": [ + { + "bbox": [ + 408, + 371, + 471, + 418 + ], + "lines": [ + { + "bbox": [ + 408, + 371, + 471, + 418 + ], + "spans": [ + { + "bbox": [ + 408, + 371, + 471, + 418 + ], + "type": "image", + "image_path": "f69f2fcd8b8e87168852ce4f74b0fc88398e0ca3e48ff677b600c28204acade1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 476, + 371, + 541, + 418 + ], + "blocks": [ + { + "bbox": [ + 476, + 371, + 541, + 418 + ], + "lines": [ + { + "bbox": [ + 476, + 371, + 541, + 418 + ], + "spans": [ + { + "bbox": [ + 476, + 371, + 541, + 418 + ], + "type": "image", + "image_path": "8f6721d53d122fccde922e3bb27316351ef6ca068aa5018c03e0cc119d1f5a38.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 509, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 533 + ], + "type": "text", + "content": "knowledge, aims to adapt the model to new classes by only a few labeled samples for each one." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 533, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 700 + ], + "type": "text", + "content": "Recently, meta-learning-based few-shot learning methods have made great strides in the setting where the train and test tasks are sampled from the same domain [49, 22]. However, more and more studies have demonstrated that these existing works fail to generalize well to novel classes that are heterogeneous with the source domain [45, 9, 14]. This limitation is attributable to the fact that most previous few-shot learning models focus solely on how to quickly adapt to a set of novel classes with only a few labeled samples per class but less effort has been made to understand and address the domain shift problem between the target (test) task and training domain. Recently, to address this problem, a series of research referred to as cross-domain few-shot learning are proposed [45, 14, 35, 3, 7, 20, 26, 6]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Generally, these existing methods address the domain" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1590" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "type": "text", + "content": "shift problem by introducing a set of task-specific parameters to enable the feature extractor or classifier to adapt to the feature distributions of new tasks. According to the parameter generation method, these methods can be loosely divided into the following two main streams. One kind of approach is to use an auxiliary network [35, 3, 43] to generate task-specific parameters. The auxiliary network is first meta-trained with multiple tasks from the source domain, and then it utilizes the support set to estimate task-specific parameters for the target task. Other methods [21] directly attach a set of task-specific parameters to the pre-trained model on the source domain and then optimize them on a few labeled samples of the target task to well generalize the knowledge to new classes. Although cross-domain few-shot learning performance has improved, the fixed number of task-specific parameters may be too rigid to account for diverse tasks, resulting in less generality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 289, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 517 + ], + "type": "text", + "content": "In this paper, from the perspective of parameter size, we studied the performance of test tasks when inference was performed in models equipped with varying sizes of task-specific parameters. Specifically, following the trend of introducing directly optimizing task-specific parameters [21], we designed several Task-specific Adapters (TA-Modules) with increasing parameters, which are respectively attached to the conv of pre-trained backbone network (e.g., ResNet-18) to construct cross-domain models. As shown in Figure 1, the optimal task-specific parameters policy varies depending on the target task. More specifically, for different test tasks, the optimal task-specific parameters required in different layers of the backbone network for effectively adapting pre-learned knowledge to the target task's feature distribution differ. Abundant task-specific parameters may over-fit on target tasks, whereas few task-specific parameters will show the under-adaptation problem. In order to effectively direct the feature extractor to adapt feature distributions, we must carefully design task-specific parameters in each layer for the target task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 518, + 289, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 289, + 686 + ], + "type": "text", + "content": "Motivated by the findings, for improving the recognition performance of cross-domain few-shot learning, we propose the Task-aware Adaptive Network (TA" + }, + { + "bbox": [ + 46, + 518, + 289, + 686 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 518, + 289, + 686 + ], + "type": "text", + "content": "-Net) to learn the optimal task-specific parameters policy for target tasks adaptively. As shown in Figure 2, our TA" + }, + { + "bbox": [ + 46, + 518, + 289, + 686 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 518, + 289, + 686 + ], + "type": "text", + "content": "-Net model consists of an Action Generation Network (\"Agent\") trained by Reinforcement Learning to generate \"actions\" - adaptive Task-specific Adapter execution decisions aimed at the target task, and an Adapted Network built on the \"actions\" to infer the target task, meanwhile, which can be viewed as the \"Environment\" to provide a \"reward\" for the Action Generation Network. The TA" + }, + { + "bbox": [ + 46, + 518, + 289, + 686 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 518, + 289, + 686 + ], + "type": "text", + "content": "-Net can learn optimally adapted network graphs aimed at the target task to transfer pre-learned knowledge efficiently." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 688, + 201, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 688, + 201, + 699 + ], + "spans": [ + { + "bbox": [ + 59, + 688, + 201, + 699 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 700, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 700, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 700, + 287, + 714 + ], + "type": "text", + "content": "(i) We propose the Task-aware Adaptive Network (TA" + }, + { + "bbox": [ + 59, + 700, + 287, + 714 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 59, + 700, + 287, + 714 + ], + "type": "text", + "content": "-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "Net), which can adaptively learn optimal task-specific parameters policy for each target task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 96, + 547, + 216 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 305, + 96, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 96, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 305, + 96, + 545, + 132 + ], + "type": "text", + "content": "(ii) We evaluate our model on the Meta-dataset. Empirical results show that our model obviously outperforms existing state-of-the-art methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 133, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 547, + 216 + ], + "type": "text", + "content": "(iii) We further analyze parameter distribution and discover that domains with significant distribution shift (compared with the source domain) usually have a larger demand for task-specific parameters to effectively adapt pre-learned knowledge. Furthermore, layers in deep blocks necessitate more task-specific parameters than those in shallow blocks for learning task-specific features." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 227, + 392, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 227, + 392, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 227, + 392, + 239 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 247, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 294 + ], + "type": "text", + "content": "Cross-domain Few-shot Learning Existing cross-domain few-shot learning methods can be coarsely grouped into two categories: feature-selection-based methods and adaptation-based methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 295, + 546, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 546, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 546, + 486 + ], + "type": "text", + "content": "Several feature-selection-based methods, including SUR [7], URT [26], and URL [20], aim to achieve generalization by adaptively integrating feature representations from multiple training domains. SUR [7] and URT [26] rely on attention mechanisms to select appropriate domain-specific representations for a given few-shot learning task. While these methods have shown good performance, they require multiple forward networks during inference time. In contrast, Li et al. [20] proposed URL, an efficient method that distills knowledge from multiple domain networks to a single feature extractor. However, supervised training of multiple domains can result in supervision collapse, leading to the loss of information required for transfer to new tasks or domains. To address this issue, Doersch et al. [6] proposed using self-supervised learning to encourage the learning of general features from multiple training domains." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 487, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 714 + ], + "type": "text", + "content": "The goal of adaptation-based methods is to adapt a feature extractor and classifier learned from a source domain to a target task through task-specific parameters. CNAPS [35] employed an adaptation network to generate task-specific parameters for each hidden layer of the feature extractor and classifier to transfer knowledge. Simple CNAPS [3] improved upon this approach by replacing the parametric classifier with a Mahalanobis distance-based distance metric, which increased classification accuracy while reducing the number of parameters. Triantafillou et al.[43] proposed the FLUTE, which learns domain-specific parameters using FiLM layers called universal templates. Li et al.[21] have demonstrated that attaching task-specific weights to a pretrained model and optimizing them from scratch directly on a small support set can enable good generalization to the target task. However, the shifts of feature distribution between different target tasks and the pre-learned domain vary, leading to the change of task-specific parameters demand. In this paper, we focus on adaptively generating task-specific" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "text", + "content": "1591" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 72, + 521, + 277 + ], + "blocks": [ + { + "bbox": [ + 75, + 72, + 521, + 277 + ], + "lines": [ + { + "bbox": [ + 75, + 72, + 521, + 277 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 521, + 277 + ], + "type": "image", + "image_path": "903c83177ec845267659ab2d6306439b2a6aa747dae4cbb6980c93006391ab16.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 288, + 547, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 288, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 547, + 323 + ], + "type": "text", + "content": "Figure 2. Task-aware Adaptive Network (TA" + }, + { + "bbox": [ + 46, + 288, + 547, + 323 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 288, + 547, + 323 + ], + "type": "text", + "content": "-Net). The \"Agent\" is the Action Generation Network, which predicts the \"action\" of Task-specific Adapter execution and interacts with the \"Environment\" (Env, the Adapted Network) providing the \"reward\" for the Agent. \"none\" represent that the model does not need the Task-specific Adapter to adapt feature." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 335, + 288, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 288, + 360 + ], + "type": "text", + "content": "graphs for each target task, to adapt the prior knowledge to the target task well." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 367, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 289, + 715 + ], + "type": "text", + "content": "Reinforcement Learning Reinforcement learning [1, 8] is primarily used to solve the sequence decision problem, in which the reward value obtained from agent-environment interaction is used as a feedback signal to continuously optimize agent strategies. There are three main approaches to solving Reinforcement Learning (RL) problems: methods based on value functions [30, 46, 48], methods based on policy search [41, 40], and the actor-critic approach [15, 39]. The deep reinforcement learning algorithm based on value functions approximates the value function or action-value function with a deep neural network and updates it with the temporal difference (TD) learning or Q-learning method. Among them, Deep Q-Network (DQN) [30] algorithm is representative. Furthermore, a series of variants have been proposed, including double DQN (DDQN) [46], which optimizes the problem of overestimation about Q value in the DQN algorithm, and Dueling DQN [48], which improves DQN from the network structure and can estimate Q function more accurately. Policy-based methods [41, 40] directly search for an optimal policy without maintaining a value function model. Parameters of the chosen policy are updated using either gradient-based or gradient-free optimization to maximize the expected return. The actor-critic approach [15] employs both value functions and policy search to trade-off the variance reduction of policy gradients and bias introduction from value function methods, in which the \"actor\" (policy) learns by using feedback from the \"critic\" (value function) [1]. In this paper, we employ policy-gradient-based Reinforcement Learning to train" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 335, + 547, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 547, + 371 + ], + "type": "text", + "content": "the Action Generation Network to adaptively generate optimal actions – Task-specific Adapter execution decisions for each target task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 383, + 362, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 383, + 362, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 383, + 362, + 396 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 403, + 545, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 403, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 545, + 428 + ], + "type": "text", + "content": "In this section, we describe the problem setting and introduce our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 436, + 385, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 436, + 385, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 436, + 385, + 449 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "type": "text", + "content": "Few-shot task Few-shot learning classification aims to classify a large number of samples with only a small number of labeled ones in each class. Labeled samples are generally referred to as the support set " + }, + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "type": "inline_equation", + "content": "S = \\{(\\pmb{x}_i, y_i)\\}_{i=1}^{N_S}" + }, + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "type": "text", + "content": ", while classified ones are known as the query set " + }, + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "type": "inline_equation", + "content": "Q = \\{(\\pmb{x}_i)\\}_{i=1}^{N_Q}" + }, + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "type": "text", + "content": " is the label of the sample " + }, + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 304, + 456, + 545, + 541 + ], + "type": "text", + "content": ". A few-shot task is created by combining the support and query sets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 542, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 650 + ], + "type": "text", + "content": "Cross-domain few-shot learning Compared with traditional few-shot learning, cross-domain few-shot learning is more challenging, since not only the labeled samples in the target task " + }, + { + "bbox": [ + 304, + 542, + 545, + 650 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 304, + 542, + 545, + 650 + ], + "type": "text", + "content": " are few, but also the distribution of the target domain and the base domain " + }, + { + "bbox": [ + 304, + 542, + 545, + 650 + ], + "type": "inline_equation", + "content": "D_{b}" + }, + { + "bbox": [ + 304, + 542, + 545, + 650 + ], + "type": "text", + "content": " may be highly heterogeneous. Therefore, we must strike a balance between model capacity and the reliability of the adaptation while avoiding over-fitting to the few labeled samples or under-adaptation for the target task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 658, + 520, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 520, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 520, + 671 + ], + "type": "text", + "content": "3.2. Task-aware Adaptive Network (TA²-Net)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "content": "Inspired by the observations that varying tasks require variant task-specific parameters for effective adaptation, we design the " + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^2" + }, + { + "bbox": [ + 305, + 677, + 545, + 715 + ], + "type": "text", + "content": "-Net adaptively learning the optimally" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1592" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": "adapted network graph for each target task. As shown in Figure 2, the " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^2" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": "-Net consists of Action Generation Network and Adapted Network. To directly optimize the Action Generation Network to generate optimal task-specific parameters for the Adapted Network, we cast the model in Reinforcement Learning. In particular, Action Generation Network can be viewed as an \"Agent\" that interacts with the \"Environment\" (Adapted Network). The Action Generation Network will result in the \"actions\" that are the prediction of Task-specific Adapter execution for the target task. Based on the generated \"actions\", we can build the unique Adapted Network, which will provide a \"reward\" for the agent - we denote this reward by " + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": ". We will introduce the Action Generation Network, Adapted Network, and their optimization process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 268, + 200, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 268, + 200, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 268, + 200, + 280 + ], + "type": "text", + "content": "3.2.1 Action Generation Network" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "text", + "content": "Our focus in the Action Generation Network is to generate adaptive Task-specific Adapter execution decisions for each layer of the Adapted Network aimed at the target task. As shown in Figure 2, the Gate Network is integrated into each layer of the feature encoder " + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "inline_equation", + "content": "f_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "text", + "content": " pre-trained by the seen domain to generate the discrete probability for " + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "text", + "content": " kinds of Task-specific Adapters. Concretely, the discrete probability in " + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "text", + "content": "-th layer can be formulated as " + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "inline_equation", + "content": "\\{z_l^i\\}_{i=1}^K" + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "inline_equation", + "content": "h_{l-1} \\in R^{N_S \\times C \\times W \\times H}" + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "text", + "content": " denotes the input feature of " + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "text", + "content": "-th feature encoder. In the Gate Network, the input feature is first fed to the Global Average Pooling (GAP) to compress the spatial dimension. Formally, the output " + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "inline_equation", + "content": "u_{l-1}" + }, + { + "bbox": [ + 46, + 287, + 289, + 435 + ], + "type": "text", + "content": " can be achieved as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 446, + 287, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 446, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 118, + 446, + 287, + 460 + ], + "type": "interline_equation", + "content": "\\boldsymbol {u} _ {l - 1} = G A P \\left(\\boldsymbol {h} _ {l - 1}\\right), \\tag {1}", + "image_path": "dba357a98af5243a0892f6c7b8dd56b07b9cc4f69ed8a83f5d2d1e958130f915.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "inline_equation", + "content": "u_{l-1}" + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "text", + "content": " can be written as " + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "inline_equation", + "content": "u_{l-1} = \\left[u_{l-1}^{1}, u_{l-1}^{2}, \\dots, u_{l-1}^{N_{S}}\\right]" + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "inline_equation", + "content": "u_{l-1}^{i}" + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "text", + "content": " represents the channel-level feature embedding of support sample " + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "inline_equation", + "content": "u_{l-1}^{i} \\in R^{C}" + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "text", + "content": ". Then, let " + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "inline_equation", + "content": "v_{l-1} = \\left[v_{l-1}^{1}, v_{l-1}^{2}, \\dots, v_{l-1}^{N}\\right]" + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "text", + "content": " denotes the class prototypes in the target task, where " + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "inline_equation", + "content": "v_{l-1}^{n}" + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "text", + "content": " refers to the class center of class " + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "text", + "content": " denotes the number of classes in the target task. Formally, the class center for class " + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 472, + 287, + 584 + ], + "type": "text", + "content": " can be obtained by computing the mean vector of all embedded features in the class as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 591, + 287, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 591, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 115, + 591, + 287, + 620 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {l - 1} ^ {n} = \\frac {1}{| S _ {n} |} \\sum_ {i \\in S _ {n}} \\boldsymbol {u} _ {l - 1} ^ {i}, \\tag {2}", + "image_path": "220364de26288853214114699412f0b1d58b10d03ea9f4979871c9f9e2350a33.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": " denotes the set of examples labeled with class " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": ". The output " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_{l-1}" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": " is a collection of all class descriptors with statistics that express the entire target task. To capture the dependencies between features, we add a simple non-linear function of two fully-connected layers coupled with a ReLU activation function. The output of this operation is a vector containing unnormalized scores for executing the " + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": " kinds" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 414, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 414, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 414, + 84 + ], + "type": "text", + "content": "of Task-specific Adapters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 375, + 92, + 546, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 92, + 546, + 107 + ], + "spans": [ + { + "bbox": [ + 375, + 92, + 546, + 107 + ], + "type": "interline_equation", + "content": "\\bar {z} _ {l} = \\mathbf {W} _ {l} ^ {2} \\sigma \\left(\\mathbf {W} _ {l} ^ {1} v _ {l - 1}\\right), \\tag {3}", + "image_path": "2e348bc11ba74ebbbba5b280109c91124fcff984afa361fbbd537e3a613d4886.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\bar{z}_l\\in R^K" + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "text", + "content": " refers to the " + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "inline_equation", + "content": "ReLU" + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "text", + "content": " activation function. " + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_l^1\\in R^{d\\times C}" + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_l^2\\in R^{K\\times d}" + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "text", + "content": " is the dimension of the hidden layer. We can write " + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\bar{z}_l" + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\bar{z}_l = [\\bar{z}_l^1,\\bar{z}_l^2,\\dots ,\\bar{z}_l^K ]" + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "text", + "content": " Then, the Gate outputs a distribution over the " + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "text", + "content": " -th Task-specific Adapter execution decisions " + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\pmb {z}_l" + }, + { + "bbox": [ + 304, + 115, + 547, + 187 + ], + "type": "text", + "content": " using the Softmax function" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 383, + 198, + 546, + 210 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 198, + 546, + 210 + ], + "spans": [ + { + "bbox": [ + 383, + 198, + 546, + 210 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {\\boldsymbol {\\ell}} = \\operatorname {s o f t m a x} (\\bar {\\boldsymbol {z}} _ {\\boldsymbol {\\ell}}), \\tag {4}", + "image_path": "cb7442bead9d3ef436a99e86f06acabac7f5379106bac7c87461a3ab2283c1c3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "inline_equation", + "content": "z_{l} = \\left[z_{l}^{1},z_{l}^{2},\\dots ,z_{l}^{K}\\right]" + }, + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "inline_equation", + "content": "z_{l}^{k}" + }, + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "text", + "content": " denotes the possibility of selecting " + }, + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "text", + "content": "-th Task-specific Adapter in the " + }, + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "text", + "content": "-th layer. And " + }, + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "inline_equation", + "content": "z = [z_{1},z_{2},\\dots ,z_{L}]" + }, + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 218, + 547, + 266 + ], + "type": "text", + "content": " is the number of layers in the backbone." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 267, + 547, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 547, + 387 + ], + "type": "text", + "content": "Training The generation of Task-specific Adapter execution decisions is cast as the Reinforcement Learning problem. The Action Generation Network aims to decide the most appropriate Task-specific Adapters for conducting few-shot learning on the current tasks. We utilize the predicted result of the Adapted Network constructed based on the current \"actions\" as the rewards for optimizing the Action Generation Network. Inspired by the self-critical sequence training approach [34], the goal of training is to minimize the negative expected reward" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 350, + 396, + 545, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 396, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 350, + 396, + 545, + 409 + ], + "type": "interline_equation", + "content": "L (\\theta) = - \\left(r \\left(\\boldsymbol {z} ^ {s}\\right) - r \\left(\\boldsymbol {z} ^ {m}\\right)\\right) \\log \\left(\\boldsymbol {z} ^ {s}\\right), \\tag {5}", + "image_path": "1a72381ce06287f48fa1373d0914a04b4bc6356958546a4cbe3bd2ee38c479bb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "text", + "content": " is the parameters of the Gate Network. " + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "\\theta = \\{\\mathbf{W}_l^1,\\mathbf{W}_l^2\\}" + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "l\\in \\{1,2,\\dots ,L\\}" + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "z^{s} = [z_{1}^{s},\\dots,z_{L}^{s}]" + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "z_{l}^{s}" + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "text", + "content": " is the possibility of action randomly sampled from the " + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "text", + "content": " actions - Task-specific Adapters, at the layer " + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "r(z^{s})" + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "text", + "content": " is the reward obtained by the Adapted Network, which is constructed by the actions with " + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "z^{s}" + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "text", + "content": " discrete possibility. " + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "inline_equation", + "content": "z^{m} = [z_{1}^{m},\\dots,z_{L}^{m}]" + }, + { + "bbox": [ + 305, + 418, + 547, + 503 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 383, + 510, + 545, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 510, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 383, + 510, + 545, + 523 + ], + "type": "interline_equation", + "content": "z _ {l} ^ {m} = \\operatorname {a r g m a x} \\left(\\boldsymbol {z} _ {l}\\right). \\tag {6}", + "image_path": "ca70a6f5e6f9d0bcdcde702265b8bbcff79dc3d21fa2df1d7dc0e73004cb9091.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\operatorname{argmax}(\\cdot)" + }, + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "text", + "content": " returns the possibility of action equipped with maximum probability over the output distribution " + }, + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "inline_equation", + "content": "z_{l}" + }, + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "text", + "content": " of the Gate in " + }, + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "text", + "content": "-th layer. " + }, + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "inline_equation", + "content": "r(z^{m})" + }, + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "text", + "content": " again is the reference reward obtained in Adapted Network equipped with actions with " + }, + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "inline_equation", + "content": "z^{m}" + }, + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "text", + "content": " discrete possibility. Accordingly, TA modules from the model that return a higher reward than " + }, + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "inline_equation", + "content": "z^{m}" + }, + { + "bbox": [ + 304, + 530, + 547, + 626 + ], + "type": "text", + "content": " will be \"pushed up\", or increased in probability, while TA modules that result in a lower reward will be suppressed." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 635, + 416, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 416, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 416, + 647 + ], + "type": "text", + "content": "3.2.2 Adapted Network" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "As shown in Figure 2, the network architecture of the Adapted Network consists of a series of residual blocks, which are formed by the Adaptive Module (AM) followed by the BN layer and ReLU function. In the Adaptive Module, a " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " conv is paralleled with the Task-specific Adapter," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1593" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 227, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 227, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 227, + 84 + ], + "type": "text", + "content": "Algorithm 1 Task-aware Adaptive Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 86, + 287, + 352 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 86, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 86, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 53, + 86, + 287, + 109 + ], + "type": "text", + "content": "1: Train the backbone network " + }, + { + "bbox": [ + 53, + 86, + 287, + 109 + ], + "type": "inline_equation", + "content": "f_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 53, + 86, + 287, + 109 + ], + "type": "text", + "content": " on the source domain;" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 110, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 110, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 110, + 287, + 144 + ], + "type": "text", + "content": "2: Build Action Generation Network " + }, + { + "bbox": [ + 53, + 110, + 287, + 144 + ], + "type": "inline_equation", + "content": "M_{act}" + }, + { + "bbox": [ + 53, + 110, + 287, + 144 + ], + "type": "text", + "content": " on the backbone network; Initialization the parameters " + }, + { + "bbox": [ + 53, + 110, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 53, + 110, + 287, + 144 + ], + "type": "text", + "content": " of Gate Networks;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 146, + 136, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 136, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 136, + 158 + ], + "type": "text", + "content": "3: while training do" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 158, + 286, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 286, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 286, + 183 + ], + "type": "text", + "content": "4: Random sample task " + }, + { + "bbox": [ + 53, + 158, + 286, + 183 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 53, + 158, + 286, + 183 + ], + "type": "text", + "content": " in the source domain; " + }, + { + "bbox": [ + 53, + 158, + 286, + 183 + ], + "type": "inline_equation", + "content": "T_{i} = S \\bigcup Q, S = \\{(x_{i},y_{i})\\}_{i = 1}^{N_{S}}, Q = \\{(x_{i})\\}_{i = 1}^{N_{Q}}" + }, + { + "bbox": [ + 53, + 158, + 286, + 183 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 183, + 252, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 183, + 252, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 183, + 252, + 195 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 53, + 183, + 252, + 195 + ], + "type": "inline_equation", + "content": "\\pmb{z} = f_{\\phi, \\theta}(x), x \\in S, \\pmb{z} = [\\pmb{z}_1, \\pmb{z}_2, \\dots, \\pmb{z}_L]" + }, + { + "bbox": [ + 53, + 183, + 252, + 195 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 195, + 286, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 286, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 286, + 219 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 53, + 195, + 286, + 219 + ], + "type": "inline_equation", + "content": "\\pmb{z}^s = [Sample(\\pmb{z}_1), \\dots, Sample(\\pmb{z}_L)]; \\pmb{z}^m = [argmax(\\pmb{z}_1), \\dots, argmax(\\pmb{z}_L)];" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 219, + 286, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 219, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 219, + 286, + 255 + ], + "type": "text", + "content": "7: Build Adapted Network " + }, + { + "bbox": [ + 53, + 219, + 286, + 255 + ], + "type": "inline_equation", + "content": "M_{adapt}^{1}" + }, + { + "bbox": [ + 53, + 219, + 286, + 255 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 53, + 219, + 286, + 255 + ], + "type": "inline_equation", + "content": "M_{adapt}^{2}" + }, + { + "bbox": [ + 53, + 219, + 286, + 255 + ], + "type": "text", + "content": " based on the backbone network according to actions " + }, + { + "bbox": [ + 53, + 219, + 286, + 255 + ], + "type": "inline_equation", + "content": "z^{s}" + }, + { + "bbox": [ + 53, + 219, + 286, + 255 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 53, + 219, + 286, + 255 + ], + "type": "inline_equation", + "content": "z^{m}" + }, + { + "bbox": [ + 53, + 219, + 286, + 255 + ], + "type": "text", + "content": " respectively;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 255, + 243, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 255, + 243, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 255, + 243, + 268 + ], + "type": "text", + "content": "8: Initialization " + }, + { + "bbox": [ + 53, + 255, + 243, + 268 + ], + "type": "inline_equation", + "content": "\\psi, \\pi" + }, + { + "bbox": [ + 53, + 255, + 243, + 268 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 53, + 255, + 243, + 268 + ], + "type": "inline_equation", + "content": "M_{adapt}^{1}" + }, + { + "bbox": [ + 53, + 255, + 243, + 268 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 53, + 255, + 243, + 268 + ], + "type": "inline_equation", + "content": "M_{adapt}^{2}" + }, + { + "bbox": [ + 53, + 255, + 243, + 268 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 268, + 286, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 268, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 268, + 286, + 293 + ], + "type": "text", + "content": "9: Optimize " + }, + { + "bbox": [ + 53, + 268, + 286, + 293 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 53, + 268, + 286, + 293 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 53, + 268, + 286, + 293 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 53, + 268, + 286, + 293 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 53, + 268, + 286, + 293 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 53, + 268, + 286, + 293 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 53, + 268, + 286, + 293 + ], + "type": "inline_equation", + "content": "L_{adapt}(\\psi, \\pi) = \\frac{1}{N_S} \\sum_{(x,y) \\in S} l(f_{\\phi,\\psi,\\pi}(\\boldsymbol{x}), y)" + }, + { + "bbox": [ + 53, + 268, + 286, + 293 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 50, + 293, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 293, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 50, + 293, + 287, + 319 + ], + "type": "text", + "content": "10: Predict over " + }, + { + "bbox": [ + 50, + 293, + 287, + 319 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 50, + 293, + 287, + 319 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 50, + 293, + 287, + 319 + ], + "type": "inline_equation", + "content": "r(\\pmb{z}^s) = N_{adapt}^1 / N_Q" + }, + { + "bbox": [ + 50, + 293, + 287, + 319 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 50, + 293, + 287, + 319 + ], + "type": "inline_equation", + "content": "r(\\pmb{z}^m) = N_{adapt}^2 / N_Q" + }, + { + "bbox": [ + 50, + 293, + 287, + 319 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 50, + 319, + 286, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 319, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 50, + 319, + 286, + 342 + ], + "type": "text", + "content": "11: Optimize the parameters " + }, + { + "bbox": [ + 50, + 319, + 286, + 342 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 50, + 319, + 286, + 342 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 50, + 319, + 286, + 342 + ], + "type": "inline_equation", + "content": "M_{act}" + }, + { + "bbox": [ + 50, + 319, + 286, + 342 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 50, + 319, + 286, + 342 + ], + "type": "inline_equation", + "content": "L(\\theta) = -(r(\\pmb{z}^s) - r(\\pmb{z}^m))\\log (\\pmb{z}^s)" + }, + { + "bbox": [ + 50, + 319, + 286, + 342 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 50, + 342, + 108, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 342, + 108, + 352 + ], + "spans": [ + { + "bbox": [ + 50, + 342, + 108, + 352 + ], + "type": "text", + "content": "12: end while" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 48, + 354, + 223, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 223, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 223, + 365 + ], + "type": "text", + "content": "Output: Action Generation Network " + }, + { + "bbox": [ + 48, + 354, + 223, + 365 + ], + "type": "inline_equation", + "content": "M_{act}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 389, + 287, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 412 + ], + "type": "text", + "content": "which is the action predicted by the corresponding Gate Network of the Action Generation Network." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "inline_equation", + "content": "f_{\\phi_l}(\\cdot)" + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "text", + "content": " conv in the " + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "text", + "content": "-th Adaptive Module, which is pre-trained by the seen domain. With " + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "inline_equation", + "content": "h_{l-1} \\in R^{N_S \\times C \\times W \\times H}" + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "text", + "content": " as the input feature of " + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "text", + "content": "-th Adaptive Module, the output is " + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "inline_equation", + "content": "h_l = f_{\\phi_l}(h_{l-1})" + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "text", + "content": " when no Task-specific Adapter is selected according to the decision generated by the Action Generation Network, that is the feature does not need adaptation. When the " + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "text", + "content": "-th Task-specific Adapter is chosen, which is constructed by paralleling " + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "text", + "content": " task-specific layers in our design, the feature learned by the Task-specific Adapter will be combined with the feature extracted from the " + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 413, + 287, + 544 + ], + "type": "text", + "content": " conv as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 87, + 552, + 287, + 584 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 552, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 87, + 552, + 287, + 584 + ], + "type": "interline_equation", + "content": "\\boldsymbol {h} _ {\\boldsymbol {l}} = f _ {\\phi_ {\\boldsymbol {l}}} \\left(\\boldsymbol {h} _ {\\boldsymbol {l} - \\mathbf {1}}\\right) + \\frac {1}{k} \\sum_ {i = 1} ^ {k} f _ {\\psi_ {\\boldsymbol {l}} ^ {i}} \\left(\\boldsymbol {h} _ {\\boldsymbol {l} - \\mathbf {1}}\\right), \\tag {7}", + "image_path": "f956fbc3def7878767b1a8016446a34d5108a0d4546025fe26b80b2e96277040.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 592, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 286, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 592, + 286, + 616 + ], + "type": "inline_equation", + "content": "f_{\\psi_l^i}(\\cdot)" + }, + { + "bbox": [ + 46, + 592, + 286, + 616 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 592, + 286, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 592, + 286, + 616 + ], + "type": "text", + "content": "-th task-specific layer in the Task-specific Adapter, which is " + }, + { + "bbox": [ + 46, + 592, + 286, + 616 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 592, + 286, + 616 + ], + "type": "text", + "content": " conv in the experiments." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 616, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 287, + 677 + ], + "type": "text", + "content": "Training Upon the generated \"actions\" with the discrete possibility " + }, + { + "bbox": [ + 47, + 616, + 287, + 677 + ], + "type": "inline_equation", + "content": "z^s" + }, + { + "bbox": [ + 47, + 616, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 616, + 287, + 677 + ], + "type": "inline_equation", + "content": "z^m" + }, + { + "bbox": [ + 47, + 616, + 287, + 677 + ], + "type": "text", + "content": " in the Action Generation Network, we will construct two Adapted Networks " + }, + { + "bbox": [ + 47, + 616, + 287, + 677 + ], + "type": "inline_equation", + "content": "M_{adapt}^{1}" + }, + { + "bbox": [ + 47, + 616, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 616, + 287, + 677 + ], + "type": "inline_equation", + "content": "M_{adapt}^{2}" + }, + { + "bbox": [ + 47, + 616, + 287, + 677 + ], + "type": "text", + "content": ". In training, the Adapted Networks are trained with the cross-entropy loss over the support samples:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 77, + 685, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 685, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 77, + 685, + 287, + 715 + ], + "type": "interline_equation", + "content": "L _ {a d a p t} (\\psi , \\pi) = \\frac {1}{N _ {S}} \\sum_ {(x, y) \\in S} l \\left(f _ {\\phi , \\psi , \\pi} (\\boldsymbol {x}), y\\right), \\tag {8}", + "image_path": "6b14db836232623be58192a74061f808b3f15bdb1f2230fb4d8ea4d6fc5fecf5.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "f_{\\psi, \\pi}(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " denotes the output softmax probability vector of the linear classifier with parameters " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", in which dimensionality equals the number of categories in the support set " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ". Then, the reward for the agent is achieved by predicting the Adapted Network over the query samples " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 140, + 545, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 140, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 318, + 140, + 545, + 169 + ], + "type": "interline_equation", + "content": "r \\left(\\boldsymbol {z} ^ {\\boldsymbol {s}}\\right) = \\frac {N _ {\\text {a d a p t}} ^ {1}}{N _ {Q}} \\quad \\text {a n d} \\quad r \\left(\\boldsymbol {z} ^ {\\boldsymbol {m}}\\right) = \\frac {N _ {\\text {a d a p t}} ^ {2}}{N _ {Q}}. \\tag {9}", + "image_path": "c91a53da1284a7c3329018cade8e384724e4fd57ba543af6e1dcb7213e073658.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 178, + 545, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 545, + 229 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 178, + 545, + 229 + ], + "type": "inline_equation", + "content": "N_{adapt}^{1}" + }, + { + "bbox": [ + 304, + 178, + 545, + 229 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 178, + 545, + 229 + ], + "type": "inline_equation", + "content": "N_{adapt}^{2}" + }, + { + "bbox": [ + 304, + 178, + 545, + 229 + ], + "type": "text", + "content": " denote the number of query samples correctly predicted by Adapted Networks " + }, + { + "bbox": [ + 304, + 178, + 545, + 229 + ], + "type": "inline_equation", + "content": "M_{adapt}^{1}" + }, + { + "bbox": [ + 304, + 178, + 545, + 229 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 178, + 545, + 229 + ], + "type": "inline_equation", + "content": "M_{adapt}^{2}" + }, + { + "bbox": [ + 304, + 178, + 545, + 229 + ], + "type": "text", + "content": " respectively. We show the Task-aware Adaptive Learning pseudo code in Algorithm 1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 243, + 370, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 370, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 370, + 255 + ], + "type": "text", + "content": "3.2.3 Testing" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 346 + ], + "type": "text", + "content": "Note that the support samples of testing will first go through the Action Generation Network to generate optimal Task-specific Adapter executions for the target task, which are the action with maximum probability. And then, based on the executions, the Adapted Network will be built and finetuned over the support set. Finally, the query samples will be predicted by the Adapted Network." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 357, + 386, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 357, + 386, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 357, + 386, + 369 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 376, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 412 + ], + "type": "text", + "content": "In this part, we first introduce the experimental setup and then compare our method with the state-of-the-art methods. Furthermore, we provide a series of analyses of our method." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 420, + 420, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 420, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 420, + 433 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 437, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 545 + ], + "type": "text", + "content": "Dataset. Meta-Dataset [44] is the standard benchmark for evaluating the performance of few-shot classification. It consists of 13 image datasets: ILSVRC 2012 [36], Omniglot [18], FGVC-Aircraft [29], CUB-200-2011 [47], Describable Textures [5], Quick Draw [11], FGVCx Fungi [38], VGG Flower [31], Traffic Signs [13], MSCOCO [24], MNIST [19], CIFAR-10 [16], and CIFAR-100 [16]. For all datasets in the experiments, we used the standard split procedure as [44]." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "Implementation details. In the experiments, we considered two training settings: multi-domain learning and single-domain setting. In the multi-domain setting, the first eight datasets (ILSVRC 2012, Omniglot, Aircraft, CUB-200-2011, Textures, Quick Draw, Fungi, and VGG Flower) of the Meta-dataset are used to train the feature extractor. We followed the method in [20] to obtain the pre-trained model. And in the single-domain setting, only the ILSVRC 2012 dataset participated in the training of the feature extractor. As in [20], we adopted the ResNet-18 as the backbone no matter in the multi-domain setting or the single-domain setting. In the training, the " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^2" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "-Net is trained using the Adam optimizer. The hyper-parameters " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " are set as " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": ". We also provide the ablation study in Section 4.4." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1594" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 95, + 547, + 357 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 91 + ], + "type": "text", + "content": "Table 1. Comparison to state-of-the-art methods in the multi-domain setting. Mean accuracy, " + }, + { + "bbox": [ + 47, + 70, + 545, + 91 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 47, + 70, + 545, + 91 + ], + "type": "text", + "content": " confidence interval are reported. The best results are marked in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 95, + 547, + 357 + ], + "lines": [ + { + "bbox": [ + 47, + 95, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 95, + 547, + 357 + ], + "type": "table", + "html": "
Test DatasetCNAPS [35]Simple CNAPS [3]Transductive CNAPS [2]SUR [7]URT [26]FLUTE [43]tri-M [27]URL [20]TSA [21]Ours
ImageNet50.8 ± 1.158.4 ± 1.157.9 ± 1.156.2 ± 1.056.8 ± 1.158.6 ± 1.051.8 ± 1.158.8 ± 1.159.5 ± 1.059.6 ± 1.0
Omniglot91.7 ± 0.591.6 ± 0.694.3 ± 0.494.1 ± 0.494.2 ± 0.492.0 ± 0.693.2 ± 0.594.5 ± 0.494.9 ± 0.495.5 ± 0.4
Aircraft83.7 ± 0.682.0 ± 0.784.7 ± 0.585.5 ± 0.585.8 ± 0.582.8 ± 0.787.2 ± 0.589.4 ± 0.489.9 ± 0.490.5 ± 0.4
Birds73.6 ± 0.974.8 ± 0.978.8 ± 0.771.0 ± 1.076.2 ± 0.875.3 ± 0.879.2 ± 0.880.7 ± 0.881.1 ± 0.881.4 ± 0.8
Textures59.5 ± 0.768.8 ± 0.966.2 ± 0.871.0 ± 0.871.6 ± 0.771.2 ± 0.868.8 ± 0.877.2 ± 0.777.5 ± 0.777.4 ± 0.7
Quick Draw74.7 ± 0.876.5 ± 0.877.9 ± 0.681.8 ± 0.682.4 ± 0.677.3 ± 0.779.5 ± 0.782.5 ± 0.681.7 ± 0.682.5 ± 0.6
Fungi50.2 ± 1.146.6 ± 1.048.9 ± 1.264.3 ± 0.964.0 ± 1.048.5 ± 1.058.1 ± 1.168.1 ± 0.966.3 ± 0.866.3 ± 0.9
VGG Flower88.9 ± 0.590.5 ± 0.592.3 ± 0.482.9 ± 0.887.9 ± 0.690.5 ± 0.591.6 ± 0.692.0 ± 0.592.2 ± 0.592.6 ± 0.4
Traffic Sign56.5 ± 1.157.2 ± 1.059.7 ± 1.151.0 ± 1.148.2 ± 1.163.0 ± 1.058.4 ± 1.163.3 ± 1.182.8 ± 1.087.4 ± 0.8
MSCOCO39.4 ± 1.048.9 ± 1.142.5 ± 1.152.0 ± 1.151.5 ± 1.152.8 ± 1.150.0 ± 1.057.3 ± 1.057.6 ± 1.057.9 ± 0.9
MNIST-94.6 ± 0.494.7 ± 0.394.3 ± 0.490.6 ± 0.596.2 ± 0.395.6 ± 0.594.7 ± 0.496.7 ± 0.497.0 ± 0.4
CIFAR-10-74.9 ± 0.773.6 ± 0.766.5 ± 0.967.0 ± 0.875.4 ± 0.878.6 ± 0.774.2 ± 0.882.9 ± 0.782.1 ± 0.8
CIFAR-100-61.3 ± 1.161.8 ± 1.056.9 ± 1.157.3 ± 1.062.0 ± 1.067.1 ± 1.063.5 ± 1.070.4 ± 0.970.9 ± 0.9
Average Seen71.673.775.175.977.474.576.280.480.480.7
Average Unseen-67.466.564.162.969.969.970.678.179.1
Average All-71.271.871.471.872.773.876.679.580.1
Average Rank-7.66.27.26.55.85.53.22.21.2
", + "image_path": "56b5acf6eccb73f9f3fc1caca74125a82ff9933730808f44653b838c298bd820.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 370, + 255, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 255, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 255, + 383 + ], + "type": "text", + "content": "4.2. Comparison to state-of-the-art methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 287, + 437 + ], + "type": "text", + "content": "Following the experiment set in [21], we sampled tasks with varying numbers of ways and shots in the inference and conducted experiments in the multi-domain setting and single-domain setting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 288, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 288, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 288, + 652 + ], + "type": "text", + "content": "Multi-domain setting. Here, the results of our method on 13 datasets of the Meta-Dataset are compared with the state-of-the-art methods. We also report average accuracy over seen domains, and unseen domains, as well as the average rank as in [21]. From the results in Table 1, our method obtains state-of-the-art results on 10 out of 13 datasets and achieves the best average accuracy on seen and unseen domains. It is worth noting that our method obtains significantly better average results than the second-best approach on the unseen datasets " + }, + { + "bbox": [ + 46, + 437, + 288, + 652 + ], + "type": "inline_equation", + "content": "(+1.0\\%)" + }, + { + "bbox": [ + 46, + 437, + 288, + 652 + ], + "type": "text", + "content": ". The unseen datasets are heterogeneous with respect to the training domain, with a significant domain gap between them. In contrast to adopting the fixed number of task-specific parameters for all target tasks, e.g. CNAPS, Simple CNAPS, FLUTE, and TSA, the " + }, + { + "bbox": [ + 46, + 437, + 288, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^2" + }, + { + "bbox": [ + 46, + 437, + 288, + 652 + ], + "type": "text", + "content": "-Net can adaptively learn the optimal task-specific parameter policies for unseen target tasks to efficiently transfer the pre-trained knowledge to the target feature distribution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Single-domain setting. We also evaluate our method in the single-domain setting, which is more challenging due to optimizing the Action Generation Network only on the ImageNet dataset. Results are shown in Table 2. First, compared with other single-domain training methods, our" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "text", + "content": "method achieves significant superiority. For example, " + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^2" + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "text", + "content": "-Net outperforms the most competing method (TSA [21]) by " + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "inline_equation", + "content": "+2.9\\%" + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "inline_equation", + "content": "+0.8\\%" + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "inline_equation", + "content": "+1.6\\%" + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "inline_equation", + "content": "+1\\%" + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "text", + "content": " on Omniglot, Quick Draw, Traffic Signs and MNIST, respectively. Second, it is worth noting that, for the average accuracy on the unseen datasets, ours outperforms the second-best approach with a clear margin " + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "inline_equation", + "content": "(+3.1\\%)" + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "text", + "content": ". And we clearly observe that " + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^2" + }, + { + "bbox": [ + 304, + 370, + 547, + 526 + ], + "type": "text", + "content": "-Net achieves competitive accuracy on the ImageNet, compared with the best performance. Despite being trained on just one dataset, the Action Generation Network exhibits good generalization, producing the optimal task-specific parameter policies for the unseen domain in an excellently adaptive manner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 537, + 409, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 409, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 409, + 551 + ], + "type": "text", + "content": "4.3. Further Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "Varying-way Five-shot setting. The number of support samples ('shot') in the standard Meta-Dataset evaluation setting is variable and can range up to 100. In this section, we evaluate our method in the 5-shot setting which is more challenging due to fewer labeled samples and can better reflect the progressiveness of the model. Table 3 shows the results compared with other state-of-the-art methods. As can be seen, ours achieves better average accuracy on both seen and unseen domains. More specifically, our method outperforms other methods on the vast majority of domains (8 out of 13) and obtains the best average rank. Furthermore, the average accuracy on unseen datasets is " + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": " higher than the second-best method." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1595" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 105, + 319, + 278 + ], + "blocks": [ + { + "bbox": [ + 49, + 69, + 319, + 101 + ], + "lines": [ + { + "bbox": [ + 49, + 69, + 319, + 101 + ], + "spans": [ + { + "bbox": [ + 49, + 69, + 319, + 101 + ], + "type": "text", + "content": "Table 2. Comparison to state-of-the-art methods in the single-domain setting. Mean accuracy, " + }, + { + "bbox": [ + 49, + 69, + 319, + 101 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 49, + 69, + 319, + 101 + ], + "type": "text", + "content": " confidence interval are reported. The best results are marked in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 105, + 319, + 278 + ], + "lines": [ + { + "bbox": [ + 50, + 105, + 319, + 278 + ], + "spans": [ + { + "bbox": [ + 50, + 105, + 319, + 278 + ], + "type": "table", + "html": "
Test DatasetProtoNet [44]fo-Proto -MAML [44]ALFA+fo-Proto -MAML [44]BOHB [37]FLUTE [43]TSA [21]Ours
ImageNet50.5 ± 1.149.5 ± 1.152.8 ± 1.151.9 ± 1.146.9 ± 1.159.5 ± 1.159.3 ± 1.1
Omniglot60.0 ± 1.463.4 ± 1.361.9 ± 1.567.6 ± 1.261.6 ± 1.478.2 ± 1.281.1 ± 1.1
Aircraft53.1 ± 1.056.0 ± 1.063.4 ± 1.154.1 ± 0.948.5 ± 1.072.2 ± 1.072.6 ± 0.9
Birds68.8 ± 1.068.7 ± 1.069.8 ± 1.170.7 ± 0.947.9 ± 1.074.9 ± 0.975.1 ± 0.9
Textures66.6 ± 0.866.5 ± 0.870.8 ± 0.968.3 ± 0.863.8 ± 0.877.3 ± 0.776.8 ± 0.8
Quick Draw49.0 ± 1.151.5 ± 1.059.2 ± 1.250.3 ± 1.057.5 ± 1.067.6 ± 0.968.4 ± 0.9
Fungi39.7 ± 1.140.0 ± 1.141.5 ± 1.241.4 ± 1.131.8 ± 1.044.7 ± 1.045.3 ± 1.0
VGG Flower85.3 ± 0.887.2 ± 0.786.0 ± 0.887.3 ± 0.680.1 ± 0.990.9 ± 0.691.0 ± 0.6
Traffic Sign47.1 ± 1.148.8 ± 1.160.8 ± 1.351.8 ± 1.046.5 ± 1.182.5 ± 0.884.1 ± 0.7
MSCOCO41.0 ± 1.143.7 ± 1.148.1 ± 1.148.0 ± 1.041.4 ± 1.059.0 ± 1.058.0 ± 1.0
MNIST----80.8 ± 0.893.9 ± 0.694.9 ± 0.5
CIFAR-10----65.4 ± 0.882.1 ± 0.782.0 ± 0.7
CIFAR-100----52.7 ± 1.170.7 ± 0.970.8 ± 0.9
Average Seen50.549.552.851.946.959.559.3
Average Unseen56.758.462.460.053.271.975.0
Average All56.157.561.459.252.670.773.8
Average Rank5.95.23.44.15.61.71.3
", + "image_path": "6991f1dd1b4c750db17f5db984df17cb3f74542a21adf85b3d174517eeaf2a99.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 72, + 293, + 263, + 439 + ], + "blocks": [ + { + "bbox": [ + 72, + 293, + 263, + 439 + ], + "lines": [ + { + "bbox": [ + 72, + 293, + 263, + 439 + ], + "spans": [ + { + "bbox": [ + 72, + 293, + 263, + 439 + ], + "type": "image", + "image_path": "b57a8cb11bf2b59d7109433dfd1dd3ce5f1e80ff36855cb1a58625f6ff2b1757.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 445, + 288, + 491 + ], + "lines": [ + { + "bbox": [ + 46, + 445, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 288, + 491 + ], + "type": "text", + "content": "Figure 3. Performance of widening or deepening the Task-specific Adapter (TA) in the multi-domain setting. The axes indicate the accuracy of methods on a particular dataset. Each color corresponds to a different method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "content": "Wider or deeper in the Task-specific Adapter (TA)? To explore the method for improving the adaptation of the model, we increase the parameters of the Task-specific Adapter (TA) in wide or depth respectively. Specifically, two types of TAs are designed: (a) " + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "content": " Conv in parallel, (b) " + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 508, + 287, + 664 + ], + "type": "text", + "content": " Conv in series. As shown in Figure 3, increasing the width of the TA allows the model to achieve better performance than deepening. In particular, among the 13 datasets, 5 datasets (Traffic Signs, MSCOCO, MNIST, CIFAR-10, and CIFAR-100) unseen in the training obviously benefited more from widening the Task-specific Adapter (TA). All of the above indicates that the best way to improve adaptability is to increase the width of the Task-specific Adapter (TA)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Effectiveness of adaptive learning. In this section, we attempt to provide a more in-depth discussion of the efficacy of fixed networks as well as the proposed adaptive learned network. Figure 4 depicts the performance over the" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 326, + 104, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 324, + 69, + 545, + 101 + ], + "lines": [ + { + "bbox": [ + 324, + 69, + 545, + 101 + ], + "spans": [ + { + "bbox": [ + 324, + 69, + 545, + 101 + ], + "type": "text", + "content": "Table 3. Results for Varying-Way Five-Shot setting. Mean accuracy, " + }, + { + "bbox": [ + 324, + 69, + 545, + 101 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 324, + 69, + 545, + 101 + ], + "type": "text", + "content": " confidence interval are reported. The best results are marked in bold." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 104, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 326, + 104, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 326, + 104, + 545, + 277 + ], + "type": "table", + "html": "
Test DatasetSimple CNAPS [3]SUR [7]URT [26]URL [20]TSA [21]Ours
ImageNet47.2 ± 1.046.7 ± 1.048.6 ± 1.049.4 ± 1.048.3 ± 1.049.3 ± 1.0
Omniglot95.1 ± 0.395.8 ± 0.396.0 ± 0.396.0 ± 0.396.8 ± 0.396.6 ± 0.2
Aircraft74.6 ± 0.682.1 ± 0.681.2 ± 0.684.8 ± 0.585.5 ± 0.585.9 ± 0.4
Birds69.6 ± 0.762.8 ± 0.971.2 ± 0.776.0 ± 0.676.6 ± 0.677.3 ± 0.6
Textures57.5 ± 0.760.2 ± 0.765.2 ± 0.769.1 ± 0.668.3 ± 0.768.3 ± 0.6
Quick Draw70.9 ± 0.679.0 ± 0.579.2 ± 0.578.2 ± 0.577.9 ± 0.678.5 ± 0.5
Fungi50.3 ± 1.066.5 ± 0.866.9 ± 0.970.0 ± 0.870.4 ± 0.870.3 ± 0.8
VGG Flower86.5 ± 0.476.9 ± 0.682.4 ± 0.589.3 ± 0.489.5 ± 0.490.0 ± 0.4
Traffic Sign55.2 ± 0.844.9 ± 0.945.1 ± 0.957.5 ± 0.872.3 ± 0.676.7 ± 0.5
MSCOCO49.2 ± 0.848.1 ± 0.952.3 ± 0.956.1 ± 0.856.0 ± 0.856.0 ± 0.8
MNIST88.9 ± 0.490.1 ± 0.486.5 ± 0.589.7 ± 0.492.5 ± 0.493.3 ± 0.3
CIFAR-1066.1 ± 0.750.3 ± 1.061.4 ± 0.766.0 ± 0.772.0 ± 0.773.1 ± 0.7
CIFAR-10053.8 ± 0.946.4 ± 0.952.5 ± 0.957.0 ± 0.964.1 ± 0.864.1 ± 0.8
Average Seen69.071.273.876.676.777.0
Average Unseen62.656.059.665.271.472.6
Average All66.565.468.372.274.675.3
Average Rank5.254.12.82.21.5
", + "image_path": "48fba64b06e5a88e7081add47079e66c20c35b79d3451d76e4e2f0edcd8398f2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 332, + 293, + 521, + 453 + ], + "blocks": [ + { + "bbox": [ + 332, + 293, + 521, + 453 + ], + "lines": [ + { + "bbox": [ + 332, + 293, + 521, + 453 + ], + "spans": [ + { + "bbox": [ + 332, + 293, + 521, + 453 + ], + "type": "image", + "image_path": "1f3ce285afcd3d5a16cf39dcd2bb0f416de8fe60500b3a5195099c0604f3d60f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 462, + 506, + 473 + ], + "lines": [ + { + "bbox": [ + 345, + 462, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 345, + 462, + 506, + 473 + ], + "type": "text", + "content": "Figure 4. Effectiveness of adaptive learning." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 495, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 674 + ], + "type": "text", + "content": "Meta-dataset in the multi-domain and single-domain setting, with Model-1/2/3 constructed by attaching Module-1/2/3 respectively as shown in Figure 1 to the backbone. It can be observed that, in comparison to the backbone network, the average accuracy of seen domains in both the multi-domain and single-domain settings decreased when using the fixed Task-specific Adapter to adaptation, but our method performed better. Furthermore, average accuracy on unseen domains has consistently improved overall in the fixed Task-specific Adapter and can be improved further through adaptive learning. This discovery inspires us to recognize that simply designing a fixed Task-specific Adapter does not work and forces us to recognize the necessity of adaptively learning optimal task-specific parameters policy for each target task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "text", + "content": "Effectiveness of Reinforcement Learning. Our " + }, + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^2" + }, + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "text", + "content": "-Net model adopts the Reinforcement Learning algorithm combined with Episodic Training to adaptively learn opti" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1596" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 258, + 177 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 258, + 177 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 258, + 177 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 258, + 177 + ], + "type": "image", + "image_path": "9bfdfb265709122d04a4792b62a6440a9c6df301b20ae31fc435066e90d430d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 178, + 200, + 197 + ], + "lines": [ + { + "bbox": [ + 77, + 178, + 200, + 197 + ], + "spans": [ + { + "bbox": [ + 77, + 178, + 200, + 197 + ], + "type": "text", + "content": "(a) MMD and TA parameters of different datasets in the Meta-Dataset" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 267, + 72, + 403, + 176 + ], + "blocks": [ + { + "bbox": [ + 267, + 72, + 403, + 176 + ], + "lines": [ + { + "bbox": [ + 267, + 72, + 403, + 176 + ], + "spans": [ + { + "bbox": [ + 267, + 72, + 403, + 176 + ], + "type": "image", + "image_path": "dbf31d7b4fad8267be1b2013ca0b0319d2c49f5d90f327647149a1ed29016bcf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 178, + 403, + 198 + ], + "lines": [ + { + "bbox": [ + 277, + 178, + 403, + 198 + ], + "spans": [ + { + "bbox": [ + 277, + 178, + 403, + 198 + ], + "type": "text", + "content": "(b) TA parameters distribution in different layers for the Fungi dataset" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 408, + 72, + 541, + 178 + ], + "blocks": [ + { + "bbox": [ + 408, + 72, + 541, + 178 + ], + "lines": [ + { + "bbox": [ + 408, + 72, + 541, + 178 + ], + "spans": [ + { + "bbox": [ + 408, + 72, + 541, + 178 + ], + "type": "image", + "image_path": "4cc7c970ac9a734ebf14b9c1d9c46c20d0af5d9edce2582e38dd5eaff4a0927d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 178, + 543, + 197 + ], + "lines": [ + { + "bbox": [ + 416, + 178, + 543, + 197 + ], + "spans": [ + { + "bbox": [ + 416, + 178, + 543, + 197 + ], + "type": "text", + "content": "(c) TA parameters distribution in different layers for the CIFAR-10 dataset" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 204, + 545, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 545, + 248 + ], + "type": "text", + "content": "Figure 5. (a): The Line chart shows the mean and standard deviation of TA parameters adaptively learned by the " + }, + { + "bbox": [ + 46, + 204, + 545, + 248 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^2" + }, + { + "bbox": [ + 46, + 204, + 545, + 248 + ], + "type": "text", + "content": "-Net in target tasks sampled in different datasets. And the box plot shows the " + }, + { + "bbox": [ + 46, + 204, + 545, + 248 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 204, + 545, + 248 + ], + "type": "text", + "content": " quantiles of the MMD in different datasets while the notches indicate the median. (b) and (c): TA parameters distribution in different blocks of the task-aware adapted network for the Fungi and CIFAR-10 dataset. Sub-bars in the block denote different layers within the Block." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 262, + 287, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 381 + ], + "type": "text", + "content": "mal task-specific parameters policy for target tasks. To explore the effectiveness of Reinforcement Learning, we designed the Simple " + }, + { + "bbox": [ + 46, + 262, + 287, + 381 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^2" + }, + { + "bbox": [ + 46, + 262, + 287, + 381 + ], + "type": "text", + "content": "-Net trained by Episodic Training, which only includes the Adapted Network attaching a Gate Network in each layer for generating \"decisions\" - Chosen the Task-specific Adapter. As shown in Table 4, the " + }, + { + "bbox": [ + 46, + 262, + 287, + 381 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^2" + }, + { + "bbox": [ + 46, + 262, + 287, + 381 + ], + "type": "text", + "content": "-Net optimized with the Reinforcement Learning outperforms the Simple " + }, + { + "bbox": [ + 46, + 262, + 287, + 381 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^2" + }, + { + "bbox": [ + 46, + 262, + 287, + 381 + ], + "type": "text", + "content": "-Net, which shows that Reinforcement Learning can better learn the TA decisions for target tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 391, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 288, + 713 + ], + "type": "text", + "content": "Parameters distribution. This is an intuitive guess that the greater the difference (domain shift) between the source domain and the target domain features, the greater the demand for TA parameters. In this section, we first observe the relationship between the TA parameters and the domain shift. As widely used in previous works [28, 25], we adopt the Maximum Mean Discrepancy (MMD) to quantify the domain shift between the target task features and the source domain features extracted by the pre-trained backbone (in the multi-domain setting). The larger MMD indicates a larger domain shift between the test task and the source domain. As shown in Figure 5 (a), the Line chart and box plot shows the TA parameters and MMD distribution in different datasets, respectively. We can find that MMD for each dataset has a narrow distribution. And, on the whole, datasets with significant MMD have a larger demand for the TA parameters. Furthermore, linear regression and Pearson correlation analysis revealed that MMD is positively correlated with TA parameters, with the Pearson correlation coefficient " + }, + { + "bbox": [ + 46, + 391, + 288, + 713 + ], + "type": "inline_equation", + "content": "R = 0.9186" + }, + { + "bbox": [ + 46, + 391, + 288, + 713 + ], + "type": "text", + "content": ". It is attributed to the greater heterogeneity, more task-specific parameters are required to effectively adapt pre-learned knowledge to the feature distribution of target tasks. We further explore the TA parameters distribution in different layers of the task-aware adapted network learned by our proposed TA" + }, + { + "bbox": [ + 46, + 391, + 288, + 713 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 391, + 288, + 713 + ], + "type": "text", + "content": "-Net. As shown in Figure 5 (b) and (c), layers in the deep blocks have more task-specific parameters for learning the task-specific fea" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 286, + 535, + 344 + ], + "blocks": [ + { + "bbox": [ + 305, + 259, + 544, + 281 + ], + "lines": [ + { + "bbox": [ + 305, + 259, + 544, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 259, + 544, + 281 + ], + "type": "text", + "content": "Table 4. Effectiveness of Reinforcement Learning in the multidomain setting." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 286, + 535, + 344 + ], + "lines": [ + { + "bbox": [ + 318, + 286, + 535, + 344 + ], + "spans": [ + { + "bbox": [ + 318, + 286, + 535, + 344 + ], + "type": "table", + "html": "
ModelReinforcement LearningSeen DomainUnseen DomainAll
Simple TA2-NetX75.567.972.6
TA2-Net✓77.072.675.3
", + "image_path": "ed222538aeb9b72f75aa1bb54336865d25749caa53c62e25372ed35142bf46d4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 330, + 381, + 523, + 437 + ], + "blocks": [ + { + "bbox": [ + 305, + 355, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 305, + 355, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 355, + 545, + 376 + ], + "type": "text", + "content": "Table 5. Mean accuracy " + }, + { + "bbox": [ + 305, + 355, + 545, + 376 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 355, + 545, + 376 + ], + "type": "text", + "content": " for different number of layers " + }, + { + "bbox": [ + 305, + 355, + 545, + 376 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 355, + 545, + 376 + ], + "type": "text", + "content": " in the Task-specific Adapter." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 381, + 523, + 437 + ], + "lines": [ + { + "bbox": [ + 330, + 381, + 523, + 437 + ], + "spans": [ + { + "bbox": [ + 330, + 381, + 523, + 437 + ], + "type": "table", + "html": "
K=1K=2K=3K=4
Seen Domain75.475.275.375.1
Unseen Domain66.670.370.568.2
All72.073.373.572.5
", + "image_path": "b733b0b82c9cae51f4af305a93bd43b6e6086b95cc80c104f371b88065bdad18.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 506 + ], + "type": "text", + "content": "tures than those in the shallow blocks as demonstrated by both Fungi and CIFAR-10, which may be because the shallow layer contains more general information while the deep layer contains more domain-specific information [50]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 515, + 400, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 400, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 400, + 528 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": "Maximum width " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": " of the Task-specific Adapter. Here, we discuss the maximum width " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": " of the Task-specific Adapter, which represents the upper bound of the task-specific parameters. Experiments are conducted on the Meta-dataset in the varying-way five-shot setting. Table 5 shows the performances of " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "K \\in [1,2,3,4]" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": ". We can observe that, as " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": " increases, the average classification accuracy on seen domains lightly decreases. For unseen domains, the average classification accuracy increases at first and then decreases as " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": " increases, with " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": " producing the best results. Compared with seen domains, unseen domains are heterogeneous with the source domain and require more task-specific parameters to transfer the pre-learned knowledge. The average performance becomes poorer when using " + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": ", which may be due to too much TA, the model" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1597" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": "overfitting, or the search space of the model structure expanding, increasing the optimization difficulty of reinforcement learning. Considering the performance of various target tasks and the calculating costs, all experiments in this paper adopt " + }, + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 143, + 120, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 143, + 120, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 143, + 120, + 156 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 163, + 289, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 163, + 289, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 289, + 307 + ], + "type": "text", + "content": "In this paper, for studying cross-domain few-shot learning problems, we explored the performance of test tasks when inference was performed in models equipped with a variant number of task-specific parameters. The exploration revealed that target tasks require variant task-specific parameters for outstanding adaptation. Motivated by this, we proposed a Task-aware Adaptive Network (TA" + }, + { + "bbox": [ + 46, + 163, + 289, + 307 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 163, + 289, + 307 + ], + "type": "text", + "content": "-Net), which can adaptively learn optimal task-specific parameters policy for each target task for exhaustively improving the recognition performance of cross-domain few-shot learning. By conducting extensive experiments, we proved that our TA" + }, + { + "bbox": [ + 46, + 163, + 289, + 307 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 163, + 289, + 307 + ], + "type": "text", + "content": "-Net could achieve the best results on Meta-dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 318, + 140, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 140, + 332 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 140, + 332 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 338, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 434 + ], + "type": "text", + "content": "This work was supported in part by Beijing Natural Science Foundation Project No. Z200002, in part by National Natural Science Foundation of China (NSFC) No. U19B2036, 62225601, in part by the Program for Youth Innovative Research Team of BUPT No. 2023QNTD02, in part by scholarships from China Scholarship Council (CSC) under Grant CSC No. 202206470055, in part by BUPT Excellent Ph.D. Students Foundation No. CX2022152." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 445, + 107, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 107, + 457 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 107, + 457 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 465, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 465, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 287, + 509 + ], + "type": "text", + "content": "[1] Kai Arulkumaran, Marc Peter Deisenroth, Miles Brundage, and Anil Anthony Bharath. Deep reinforcement learning: A brief survey. IEEE Signal Processing Magazine, 34(6):26-38, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 553 + ], + "type": "text", + "content": "[2] Peyman Bateni, Jarred Barber, Jan-Willem van de Meent, and Frank Wood. Enhancing few-shot image classification with unlabelled examples. In WACV, pages 2796-2805, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 555, + 286, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 286, + 588 + ], + "type": "text", + "content": "[3] Peyman Bateni, Raghav Goyal, Vaden Masrani, Frank Wood, and Leonid Signal. Improved few-shot visual classification. In CVPR, pages 14493-14502, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 590, + 286, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 286, + 621 + ], + "type": "text", + "content": "[4] Luca Bertinetto, Joao F. Henriques, Philip Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In ICLR, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 624, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 286, + 656 + ], + "type": "text", + "content": "[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 658, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 286, + 689 + ], + "type": "text", + "content": "[6] Carl Doersch, Ankush Gupta, and Andrew Zisserman. CrossTransformers: spatially-aware few-shot transfer. NeurIPS, 33:21981-21993, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 691, + 286, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 286, + 714 + ], + "type": "text", + "content": "[7] Nikita Dvornik, Cordelia Schmid, and Julien Mairal. Selecting relevant features from a multi-domain representation for" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "few-shot classification. In ECCV, pages 769-786, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 129 + ], + "type": "text", + "content": "[8] Sven Gronauer and Klaus Diepold. Multi-agent deep reinforcement learning: a survey. Artificial Intelligence Review, 55(2):895-943, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 130, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 130, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 312, + 130, + 545, + 174 + ], + "type": "text", + "content": "[9] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In ECCV, pages 124-141, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 217 + ], + "type": "text", + "content": "[10] Yurong Guo, Ruoyi Du, Xiaoxu Li, Jiyang Xie, Zhanyu Ma, and Yuan Dong. Learning calibrated class centers for few-shot classification by pair-wise similarity. IEEE TIP, 31:4543-4555, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 251 + ], + "type": "text", + "content": "[11] David Ha and Douglas Eck. A Neural Representation of Sketch Drawings. arXiv preprint arXiv:1704.03477, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "text", + "content": "[13] Sebastian Houben, Johannes Stallkamp, Jan Salmen, Marc Schlipsing, and Christian Igel. Detection of traffic signs in real-world images: The german traffic sign detection benchmark. In IJCNN, pages 1-8, 2013. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 332, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 364 + ], + "type": "text", + "content": "[14] Yanxu Hu and Andy J. Ma. Adversarial feature augmentation for cross-domain few-shot classification. In ECCV, pages 20-37, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 365, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 387 + ], + "type": "text", + "content": "[15] Vijay Konda and John Tsitsiklis. Actor-critic algorithms. NeurIPS, 12, 1999. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 388, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 410 + ], + "type": "text", + "content": "[16] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "type": "text", + "content": "[17] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "type": "text", + "content": "[18] Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 350(6266):1332-1338, 2015. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 490, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 522 + ], + "type": "text", + "content": "[19] Yann LeCun, Corinna Cortes, and CJ Burges. MNIST handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, 2, 2010. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 524, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 524, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 308, + 524, + 545, + 556 + ], + "type": "text", + "content": "[20] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple domains for few-shot classification. In ICCV, pages 9526-9535, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 558, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 558, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 558, + 545, + 590 + ], + "type": "text", + "content": "[21] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Cross-domain few-shot learning with task-specific adapters. In CVPR, pages 7161-7170, 2022. 2, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 591, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 633 + ], + "type": "text", + "content": "[22] Xiaoxu Li, Jijie Wu, Zhuo Sun, Zhanyu Ma, Jie Cao, and Jing-Hao Xue. BSNet: Bi-similarity network for few-shot fine-grained image classification. IEEE TIP, 30:1318-1331, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 635, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 669 + ], + "type": "text", + "content": "[23] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. MetaSGD: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 5" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1598" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[25] Ge Liu, Linglan Zhao, and Xiangzhong Fang. PDA: Proxy-based domain adaptation for few-shot image recognition. Image and Vision Computing, 110:104164, 2021. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "text", + "content": "[26] Lu Liu, William Hamilton, Guodong Long, Jing Jiang, and Hugo Larochelle. A universal representation transformer layer for few-shot image classification. In ICLR, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "text", + "content": "[27] Yanbin Liu, Juho Lee, Linchao Zhu, Ling Chen, Humphrey Shi, and Yi Yang. A multi-mode modulator for multi-domain few-shot classification. In ICCV, pages 8453-8462, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 183, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 287, + 216 + ], + "type": "text", + "content": "[28] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105, 2015. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 258 + ], + "type": "text", + "content": "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[30] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 349 + ], + "type": "text", + "content": "[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "text", + "content": "[32] Hang Qi, Matthew Brown, and David G. Lowe. Low-Shot learning with imprinted weights. In CVPR, pages 5822-5830, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 382, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 403 + ], + "type": "text", + "content": "[33] Sachin Ravi and H. Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 437 + ], + "type": "text", + "content": "[34] Steven J Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel. Self-critical sequence training for image captioning. In CVPR, pages 7008-7024, 2017. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 438, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 438, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 438, + 287, + 480 + ], + "type": "text", + "content": "[35] James Requeima, Jonathan Gordon, John Bronskill, Sebastian Nowozin, and Richard E Turner. Fast and flexible multi-task classification using conditional neural adaptive processes. NeurIPS, 32, 2019. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 481, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 535 + ], + "type": "text", + "content": "[36] Olga Russakovsky, J. Deng, Hao Su, J. Krause, S. Satheesh, S. Ma, Zhiheng Huang, A. Karpathy, A. Khosla, Michael S. Bernstein, A. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 1, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "text", + "content": "[37] Tonmoy Saikia, Thomas Brox, and Cordelia Schmid. Optimized generic feature learning for few-shot classification across domains. arXiv preprint arXiv:2001.07926, 2020. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "text", + "content": "[38] Brigit Schroeder and Yin Cui. FGVCx Fungi Classification Challenge 2018. github.com/visipedia/fgvcx_fungi_comp, 2018.5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "type": "text", + "content": "[39] John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High-dimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438, 2015. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[40] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[41] David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. In ICML, pages 387-395, 2014. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 431 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[43] Eleni Triantafillou, Hugo Larochelle, Richard Zemel, and Vincent Dumoulin. Learning a universal template for few-shot dataset generalization. In ICML, pages 10424–10433, 2021. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "text", + "content": "[44] Eleni Triantafillou, Tyler Zhu, Vincent Dumoulin, Pascal Lamblin, Utku Evci, Kelvin Xu, Ross Goroshin, Carles Gelada, Kevin Swersky, Pierre-Antoine Manzagol, et al. Meta-dataset: A dataset of datasets for learning to learn from few examples. arXiv preprint arXiv:1903.03096, 2019. 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "type": "text", + "content": "[45] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In ICLR, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "type": "text", + "content": "[46] Hado Van Hasselt, Arthur Guez, and David Silver. Deep reinforcement learning with double q-learning. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 285, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 318 + ], + "type": "text", + "content": "[47] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "text", + "content": "[48] Ziyu Wang, Tom Schaul, Matteo Hessel, Hado Hasselt, Marc Lanctot, and Nando Freitas. Dueling network architectures for deep reinforcement learning. In ICML, pages 1995–2003, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[49] Davis Wertheimer, Luming Tang, and Bharath Hariharan. Few-shot classification with feature map reconstruction networks. In CVPR, pages 8012-8021, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "text", + "content": "[50] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1599" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/2e3b51ba-4260-4b59-a865-225826178569_content_list.json b/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/2e3b51ba-4260-4b59-a865-225826178569_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..692f289ad7a51e187a8376084dbcf13e675c543a --- /dev/null +++ b/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/2e3b51ba-4260-4b59-a865-225826178569_content_list.json @@ -0,0 +1,1331 @@ +[ + { + "type": "text", + "text": "TaskExpert: Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts", + "text_level": 1, + "bbox": [ + 151, + 130, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hanrong Ye and Dan Xu\\* \nDepartment of Computer Science and Engineering, HKUST \nClear Water Bay, Kowloon, Hong Kong", + "bbox": [ + 246, + 202, + 722, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hyeae, danxu}@cse.ust.hk", + "bbox": [ + 374, + 258, + 594, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning discriminative task-specific features simultaneously for multiple distinct tasks is a fundamental problem in multi-task learning. Recent state-of-the-art models consider directly decoding task-specific features from one shared task-generic feature (e.g., feature from a backbone layer), and utilize carefully designed decoders to produce multi-task features. However, as the input feature is fully shared and each task decoder also shares decoding parameters for different input samples, it leads to a static feature decoding process, producing less discriminative task-specific representations. To tackle this limitation, we propose TaskExpert, a novel multi-task mixture-of-experts model that enables learning multiple representative task-generic feature spaces and decoding task-specific features in a dynamic manner. Specifically, TaskExpert introduces a set of expert networks to decompose the backbone feature into several representative task-generic features. Then, the task-specific features are decoded by using dynamic task-specific gating networks operating on the decomposed task-generic features. Furthermore, to establish long-range modeling of the task-specific representations from different layers of TaskExpert, we design a multitask feature memory that updates at each layer and acts as an additional feature expert for dynamic task-specific feature decoding. Extensive experiments demonstrate that our TaskExpert clearly outperforms previous best-performing methods on all 9 metrics of two competitive multi-task learning benchmarks for visual scene understanding (i.e., PASCAL-Context and NYUD-v2). Codes and models will be made publicly available.", + "bbox": [ + 75, + 342, + 472, + 795 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 828, + 207, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid development of deep learning models, many computer vision tasks can be effectively handled by deep networks, which strongly motivates researchers to de", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e2bee81ef2043ed6a48b3853a6685b448704df26edbcc8b6f19b9725b6e37485.jpg", + "image_caption": [ + "Figure 1: Motivative illustration of TaskExpert for dynamical decoding discriminative task-specific representations. A set of $K$ expert networks decompose the task-generic backbone feature into $K$ representative feature spaces. Then, the feature activations of an input from the multiple expert networks are dynamically assembled by the Task Gating network, to produce task-specific features for the $T$ tasks. TaskExpert further performs long-range modeling of task-specific representations by designing a multi-task feature memory. It is updated at the current layer based on the assembled task-specific representations, and is used as an additional feature expert in the next layer." + ], + "image_footnote": [], + "bbox": [ + 511, + 306, + 857, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "velop unified multi-task networks to perform joint learning and inference on multiple distinct tasks [1, 2, 3, 4]. Multi-Task Learning (MTL) can avoid repetitive training of single-task models and is able to simultaneously generate predictions for different tasks with only one-time inference. On the other hand, as different computer vision tasks usually share a basic common understanding of input images, learning multiple tasks concurrently can help improve the representation power of each task and boost the overall performance of multi-task models [5, 6].", + "bbox": [ + 496, + 729, + 890, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning discriminative task-specific features is a funda-", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21828", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mental problem for MTL models. To achieve this objective, recent state-of-the-art works [7, 2, 8, 1] consider a decoder-focused paradigm, in which a shared backbone model with pretrained parameters is applied to learn generic image representations, and then separate task decoders are carefully designed to generate task-specific features. Although this paradigm is straightforward to achieve task-specific feature decoding, and produces promising multi-task prediction results, these models have two aspects of potential limitations: (i) The input task-generic feature from a backbone layer is fully shared for the different task decoders, which requires a careful design of the decoders (e.g. capacity and structure) for the different tasks to achieve task-discriminative decoding. (ii) The parameters of each task decoder are typically shared for different input samples, resulting in a static decoding of task-specific features, while a sample-dependant decoder is highly beneficial for learning more discriminative task features, as the input samples can be diverse, and sample-related context information is particularly important for multi-task learning and inference. All these aspects hamper the previous models from learning effective task-specific features and predictions.", + "bbox": [ + 76, + 90, + 472, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the above-mentioned issues, this paper proposes a novel multi-task mixture-of-experts framework, coined as \"TaskExpert\", which dynamically performs the task-feature decoding for different inputs and tasks as depicted in Figure 1. Instead of using one shared backbone feature to decode different task features, the proposed TaskExpert introduces a set of task-generic expert networks that learn to decompose the backbone feature into a set of representative task-generic features. The feature decomposition enables the decoders to interact with finer-granularity task-generic feature spaces to produce more discriminative task features, as each expert can be responsible for modeling one representative feature space from the training data. Then, we dynamically assemble feature activations from different experts to produce task-specific representations, based on sample-dependent and task-specific gating scores predicted from a designed gating network. On the other hand, visual understanding tasks greatly benefit from representations of different network levels [9]. Thus, we further devise a \"multi-task feature memory\" to aggregate the dynamically decoded task-specific features produced at different network layers, and propose a \"Memorial Mixture-of-Experts (MMoE)\" module. For each task, MMoE utilizes its corresponding task feature from the multi-task feature memory as an additional feature expert, to improve task-specific feature decoding. In this way, MMoE enables long-range dynamic modeling of task-specific representation for each task throughout the entire backbone.", + "bbox": [ + 76, + 425, + 472, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contribution is three-fold:", + "bbox": [ + 96, + 848, + 413, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a multi-task mixture-of-experts model that allows effective decomposition of a task-generic fea", + "bbox": [ + 94, + 869, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ture from any backbone network layer and enables dynamically assembling discriminative multi-task representations by a task gating mechanism.", + "bbox": [ + 529, + 90, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We design a novel Memorial Mixture-of-Experts (MMoE) module by further introducing a multi-task feature memory, which can effectively interact (read and write) with the mixture of experts at different network layers to achieve long-range dynamic decoding of task-specific features by utilizing different levels of network representations.", + "bbox": [ + 517, + 143, + 890, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- The proposed TaskExpert is extensively validated on two challenging multi-task visual scene understanding benchmarks (i.e. Pascal-Context and NYUD-v2). The results clearly show the effectiveness of TaskExpert for dynamic multi-task representation learning, and also establish new state-of-the-art performances on all 9 metrics on the benchmarks, regarding the comparison with previous best-performing methods using both transformer- and CNN-based backbones.", + "bbox": [ + 517, + 258, + 893, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 407, + 650, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Task Learning in Computer Vision In the field of multi-task learning (MTL) for computer vision problems [10, 11, 12, 13, 14, 15, 16, 17], some research works explore effective solutions through a perspective of model optimization and improve the optimization process of MTL via different multi-task loss designs [3, 6, 18, 19, 20] and gradient manipulations [21, 22, 23, 24]. These optimization strategies have also been demonstrated beneficial for alleviating the issue of task competition in multi-task learning [6]. From another perspective, several existing works also investigate designing better multi-task model architectures. These related works can be roughly divided into two groups: encoder-focused and decoder-focused methods. The encoder-focused methods [25, 26, 27, 28] target designing multi-task encoders, while the decoder-focused methods typically share the encoder (i.e. the backbone) for different tasks and design sophisticated decoders for task-specific feature learning and prediction [1, 2, 29, 8, 30, 31]. A recent state-of-the-art decoder-focused method (i.e. In-vPT [7]) performs multi-task learning on powerful transformer architectures [32], and establishes a very strong baseline for multi-task visual scene understanding. It can effectively learn spatial and cross-task relationships globally. To adapt multi-task models to different tasks, researchers also propose to search for task-specific network structures for encoders [33, 34, 35] and decoders [2], or utilize task-specific learnable prompts to effectively capture task-specific information [36]. Despite promising performances of these works, they are still limited by static decoder network designs since their decoders cannot dynamically generate task-specific representations based on dif", + "bbox": [ + 496, + 431, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21829", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ferent inputs and tasks. In this paper, we propose TaskExpert, a novel and effective mixture-of-experts framework, which can achieve dynamic multi-task representation learning via feature decomposition and assembling through the entire backbone network.", + "bbox": [ + 75, + 90, + 470, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mixture-of-Experts Models With statistical motivation [37], Mixture-of-Experts (MoE) models are originally designed to control the dynamics of neural networks automatically [38, 39]. MoE learns a series of expert networks and a gating network. The outputs of expert networks are weighted by gating scores (or called \"gate values\") generated by a gating network before the weighting operation. In more recent works, some researchers use the gating scores as a criterion to sparsely select only one or a few experts. The sparse activation of experts enables a significant reduction of the computational cost when training large-scale models [40, 41, 42, 43]. Related to our work, [44] designs a multi-gate MoE to ensemble expert networks for different census analysis tasks with different gating networks, and [45, 46] propose sparse task-conditioned MoE networks for efficient multi-task learning, which generate predictions for only one task within a single forward network inference. In contrast to these works, our Memorial Mixture-of-Experts (MMoE) is a plug-and-play decoder module, which is designed for dynamically assembling discriminative task-specific features. Our TaskExpert can simultaneously generate multi-task predictions for all tasks in one single forward inference, leading to significantly higher multi-task training efficiency and enabling cross-task interaction learning. Moreover, MMoE models task-specific features throughout different network layers via the novel design of multi-task feature memory and help generate more discriminative task-specific features.", + "bbox": [ + 75, + 167, + 472, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. TaskExpert", + "text_level": 1, + "bbox": [ + 76, + 608, + 200, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Framework Overview", + "text_level": 1, + "bbox": [ + 76, + 633, + 282, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overall illustration of our TaskExpert framework is shown in Figure 2. The input image is first projected to a token sequence and then fed into a series of classic visual transformer layers. Our framework can also be flexibly used with a CNN backbone. The proposed Memorial Mixture-of-Experts (MMoE) module is embedded at different network layers, and each module takes the task-generic backbone feature of that layer as input. It dynamically assembles and outputs discriminative task-specific representations at the corresponding layer. The MMoE further interacts (reads and writes operations) with the designed multi-task feature memory, to perform a long-range dynamic multi-task feature decoding, which benefits from different levels of network features. The reads utilize the feature memory as an additional feature expert in the MMoE module, and the writes update the memory at the layer. Finally, after be", + "bbox": [ + 75, + 659, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/967f229000fb7786a71aa590a3c583170e9124f754048af17bf90b0a10cf03dc.jpg", + "image_caption": [ + "Figure 2: Overall architecture of the proposed TaskExpert model. Memorial Mixture-of-Experts (MMoE) modules are embedded at different backbone layers to dynamically decode task-specific feature for each task at the corresponding layer. The MMoE also interacts (reads and writes) with a designed multi-task feature memory to have a long-range dynamic multi-task feature decoding benefiting from different levels of network features. The reads utilize the memory as an additional feature expert in MMoE, and the writes update the memory at the layer. After being updated by the last network layer, the multi-task feature memory is used for the final prediction of each task." + ], + "image_footnote": [], + "bbox": [ + 504, + 93, + 885, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing updated by the outputs of all MMoE modules at different network layers, the task-specific features in multi-task feature memory are then used for generating the multi-task predictions with simple prediction heads. We elaborate on more details of our model in the following several sections.", + "bbox": [ + 496, + 568, + 893, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-Task Backbone Structure", + "text_level": 1, + "bbox": [ + 500, + 672, + 781, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider a widely-used vision transformer (ViT) model [47] with $L$ layers as the backbone of our multi-task MoE model. The input image is first divided into $H\\times W$ patches and processed by the patch embedding and positional embedding modules. Then, we obtain a sequence of patch tokens $\\mathbf{X}^0 = \\{\\mathbf{x}_1^0,\\mathbf{x}_2^0,\\dots,\\mathbf{x}_N^0\\}$ , where $N = H\\times W$ and the number of channels is $C$ . The patch tokens contain generic visual representations of the input image. We denote the $l$ -th transformer layer as a mapping function $f^{l}$ and the updating of image patch tokens at different transformer layers can be formulated as $\\mathbf{X}^l = f^l (\\mathbf{X}^{l - 1})$ , where $\\mathbf{X}^l\\in \\mathbb{R}^{N\\times C}$ denotes the output patch tokens of the transformer at the $l$ -th layer.", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21830", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aefd51ed303b4e3c26e501c19ae2b517f2b5501a124d6fd1bee9dc1b7c125d8d.jpg", + "image_caption": [ + "Figure 3: Illustration of the proposed Memorial Mixture-of-Experts (MMoE) module with the multi-task feature memory. MMoE accepts a task-generic feature from a backbone layer as input, and dynamically generates task-specific features for different tasks in a sample- and task-dependent manner. Specifically, MMoE decomposes the backbone feature to a set of $K$ representative task-generic features $(\\mathbf{R}_1^l, \\dots, \\mathbf{R}_K^l)$ via the $K$ expert networks. Then these features are used to assemble a set of $T$ task-specific features based on dynamically generated task-gating scores from the task gating networks. To establish long-range modeling of task-specific representations crossing different network layers, the multi-task feature memory is designed to interact with the expert networks. The multi-task feature memory serves as an additional feature expert to decode the task-specific representations, and the decoded representations are also further used to update the multi-task feature memory." + ], + "image_footnote": [], + "bbox": [ + 89, + 65, + 872, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Memorial Mixture-of-Experts (MMoE)", + "text_level": 1, + "bbox": [ + 76, + 482, + 413, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present the details of several core parts of the proposed memorial mixture-of-experts (MMoE) module, including backbone generic-feature decomposition with a set of network experts, context-aware task-specific gating, the multi-task feature memory to have long-range modeling on task-specific features through interactions (reads and writes) with the mixture of experts, dynamic task-specific feature assembling, and multi-task prediction heads.", + "bbox": [ + 75, + 506, + 470, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Backbone Feature Decomposition with Experts", + "text_level": 1, + "bbox": [ + 76, + 648, + 459, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 3, to enable learning discriminative task-specific representations from the task-generic backbone feature, we first define a set of expert networks that decompose the backbone feature into several representative task-generic features in sub-feature spaces, representing distinct patterns of the backbone features, as can be observed in Figure 6. It allows more fine-grained decoding of task-specific features. Specifically, at the $l$ -th backbone layer, let us assume that we use $K$ expert networks $\\{f_{e}^{l,i}(\\cdot):\\mathbb{R}^{N\\times C}\\to \\mathbb{R}^{N\\times C},i\\in [1,K]\\}$ to learn distinct task-generic feature projections, where superscript $l$ denotes the layer index and $i$ is the expert index. Each expert network employs the same Linear-BatchNorm-RELU-Linear structure but with different parameters. Through each expert network, we obtain a set of representative fea", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tures $\\{\\mathbf{R}_i^l\\in \\mathbb{R}^{N\\times C},i\\in [1,K]\\}$ by:", + "bbox": [ + 500, + 481, + 740, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {i} ^ {l} = f _ {e} ^ {l, i} \\left(\\mathbf {X} ^ {l}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 508, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{X}^l$ is the $l$ -th layer feature of the backbone.", + "bbox": [ + 500, + 539, + 831, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Context-Aware Task-Specific Gating", + "text_level": 1, + "bbox": [ + 500, + 574, + 808, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After we have obtained a set of representative task-generic features decomposed from a backbone feature for all the tasks, we aim at decoding discriminative task-specific features for different tasks by automatically assembling these representative features. Thus, we propose to adaptively control the contribution of each expert for each patch token by designing a task-specific gating network.", + "bbox": [ + 496, + 598, + 890, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each token in the representative task-generic features, we compute a task-gating score to control its contribution to generating the task-specific features. The task-gating scores of each task are dynamically generated by a task-specific gating network, using the features $\\mathbf{X}^l$ from the $l$ -th network layer. We term the task-specific gating network with a short name \"task gating\", and denote it as $f_{g}^{l,t}(\\cdot):\\mathbb{R}^{N\\times C\\to N\\times K}$ , where $t\\in [1,T]$ ( $T$ is the total number of tasks) is a task index. The task-gating scores $\\mathbf{G}^{l,t}\\in \\mathbb{R}^{N\\times K}$ is computed as $\\mathbf{G}^{l,t} = f_{g}^{l,t}(\\mathbf{X}^{l})$ .", + "bbox": [ + 496, + 704, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the design of the gating networks, previous Mixture-of-Experts methods [42, 43, 41] use a simple linear layer followed by a Softmax function, which does not take into", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21831", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/75375102ceba2b139b1590812e19a8b5f0682786b66443627bf8d7174ea4fcd8.jpg", + "image_caption": [ + "Figure 4: PCA visualization of the representative features by task-generic experts across different layers. In the deeper layers, the features become increasingly responsive to high-level semantic areas. Conversely, features at more shallow layers show a heightened sensitivity to low-level patterns." + ], + "image_footnote": [], + "bbox": [ + 80, + 73, + 883, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "account the context of each image patch token, while using only the corresponding patch token when computing the gating score. As various challenging computer vision tasks (e.g. semantic segmentation and monocular depth estimation) require modeling and understanding each image patch with context information, we thus design a context-aware gating strategy. Specifically, context-aware gating adopts convolution kernels to extend the receptive field of the gating networks, and thus it can involve a local context of tokens in the computation of gating scores. We first reshape the input token sequence $\\mathbf{X}$ (in $\\mathbb{R}^{N\\times C}$ ) as spatial maps (in $\\mathbb{R}^{H\\times W\\times C}$ ), and then compute the gating scores of $K$ experts with three layers, each having a $3\\times 3$ convolution with ReLU as intermediate activation function. We denote this computation procedure as \"CONV\", which will also be used later in Equation 3.", + "bbox": [ + 75, + 268, + 473, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 Multi-Task Feature Memory", + "text_level": 1, + "bbox": [ + 76, + 530, + 331, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To conduct long-range modeling of the multi-task representations decoded at different network layers that contain different levels of visual information, we design a multitask feature memory to interact with the MMoE modules through read and write operations. The read operation uses the task feature memory as an additional feature expert for task-specific feature decoding in MMoE. The write operation updates the multi-task feature memory by using the decoded task-specific features from MMoE. In this way, the multi-task feature memory can aggregate the decoded task-specific features crossing different network layers. Specifically, the read operation is an identity mapping. The write operation can be performed as follows: given a decoded task-specific feature of the $t$ -th task at the first layer, i.e. $\\mathbf{F}^{1,t}$ , the initialization of the multi-task feature memory $\\mathbf{M}^{1,t}$ is obtained by $\\mathbf{M}^{1,t} \\gets \\mathbf{F}^{1,t}$ . In the subsequent layers, suppose we decode task-specific features at every network layer, at the $l$ -th layer ( $l > 1$ ), we write the multi-task feature memory $\\{\\mathbf{M}^{l,t}, t \\in [1,T]\\}$ with the decoded task-specific feature $\\mathbf{F}^{l,t}$ weighted by a task-specific learnable momentum value $\\alpha^{l,t}$ as shown in Figure 3:", + "bbox": [ + 75, + 554, + 473, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} ^ {l, t} \\leftarrow \\mathbf {F} ^ {l, t} + \\alpha^ {l, t} \\mathbf {M} ^ {l - 1, t}, l > 1. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 883, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.4 Dynamically Assembling Task-Specific Features", + "text_level": 1, + "bbox": [ + 498, + 268, + 885, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the obtained representative task-generic features and task-specific gating scores as above-described, we can perform dynamic decoding of task-specific features, as the gating parameters are predicted in a sample-dependent manner. This is significantly different from existing decoder-focused multi-task structures, which share the decoding parameters for all the different input samples. Specifically, for the $t$ -th task, we multiply every token of the representative task-generic features $\\{\\mathbf{R}_i^l,i\\in [1,K]\\}$ with their corresponding task-gating score $\\{\\mathbf{G}_i^{l,t},i\\in [1,K]\\}$ and sum the tokens in the same spatial position to decode task-specific feature. Furthermore, to learn more discriminative features for each task, we utilize the corresponding task-specific feature in multi-task feature memory as an additional feature expert in MMoE, which contains long-range task-specific information learned throughout the entire network. It helps decode more effective and long-term task-specific features. As shown in Figure 3, for the MMoE at the $l$ -th transformer layer ( $l > 1$ ), we compute an additional gating score map $\\mathbf{G}_{K + 1}^{l,t}$ for the task-specific feature from the multi-task feature memory $\\mathbf{M}^{l - 1,t}$ , which is performed by adding one more output channel into the task gating network $f_{g}^{l,t}:\\mathbb{R}^{N\\times C}\\to \\mathbb{R}^{N\\times (K + 1)}$ . In summary, the task gating scores of the $t$ -th task and the $l$ -th layer, i.e., $\\mathbf{G}^{l,t} = \\{\\mathbf{G}_1^{l,t},\\mathbf{G}_2^{l,t},\\dots,\\mathbf{G}_{K + 1}^{l,t}\\}$ , is computed as:", + "bbox": [ + 496, + 291, + 892, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} ^ {l, t} = \\left\\{ \\begin{array}{l l} \\operatorname {S o f t m a x} \\left(\\operatorname {C O N V} ^ {K + 1} \\left(\\mathbf {X} ^ {l}\\right)\\right) & l > 1, \\\\ \\operatorname {S o f t m a x} \\left(\\operatorname {C O N V} ^ {K} \\left(\\mathbf {X} ^ {l}\\right)\\right) & l = 1, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 680, + 890, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the superscript $K$ of CONV denotes the number of output convolution kernels. And accordingly, the decoded task-specific features at the $l$ -th MMoE module is:", + "bbox": [ + 496, + 720, + 893, + 766 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} ^ {l, t} = \\left\\{ \\begin{array}{l l} \\sum_ {i = 1} ^ {K} \\mathbf {G} _ {i} ^ {l, t} \\times \\mathbf {R} _ {i} ^ {l} + \\mathbf {G} _ {K + 1} ^ {l, t} \\times \\mathbf {M} ^ {l - 1, t} & l > 1, \\\\ \\sum_ {i = 1} ^ {K} \\mathbf {G} _ {i} ^ {l, t} \\times \\mathbf {R} _ {i} ^ {l} & l = 1. \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 771, + 890, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alternative sparse activation of experts As an alternative, we can also use \"sparse connections\" in MMoE, which means that we only activate $K_{sel}$ selected experts to compute the task-specific features ( $K_{sel} < K$ ) for each token [41]. In this case, for each token, we find the top $K_{sel}$", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21832", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c6c7890ca9c9ec06deb4dcfa7ad740e344b73869e4cb9139fc056a0700537c8e.jpg", + "image_caption": [ + "Figure 5: PCA visualization of the decoded final task features generated by InvPT [7] and our TaskExpert. Our model generates more discriminative features for all tasks." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 467, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "experts to be activated based on the task-gating scores, and sum their outputs after being weighted by the task-gating scores. A comparison between the dense connection and sparse connection in MMoE is discussed in the experiments.", + "bbox": [ + 76, + 268, + 467, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Generation of Multi-Task Predictions", + "text_level": 1, + "bbox": [ + 76, + 338, + 398, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the final network layer $L$ , we use task-specific prediction heads to directly generate prediction outputs for each task from the multi-task feature memory $\\{\\mathbf{M}^{L,t}, t \\in [1, T]\\}$ . The prediction heads for different tasks have a similar structure, consisting of a $3 \\times 3$ convolution layer with batch normalization and ReLU activation, followed by a linear layer for the final prediction.", + "bbox": [ + 76, + 362, + 467, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 479, + 200, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we validate the effectiveness of our proposal through several aspects: (i) ablation study, (ii) comparison with SOTA methods, and (iii) qualitative analysis.", + "bbox": [ + 76, + 505, + 467, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 560, + 264, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tasks and Benchmarks For evaluation, we adopt two popular benchmarks in multi-task learning as previous works [7, 1], i.e. PASCAL-Context [48, 49] and NYUD-v2 [50] datasets. PASCAL-Context contains 4,998 samples in train split and 5,105 samples in test split. It provides labels for multiple important visual understanding tasks, including semantic segmentation, human parsing, and object boundary detection. Previous work on multi-task learning generates pseudo labels for another two tasks [5]: surface normals estimation and saliency detection. Meanwhile, NYUD-v2 contains 1,449 samples in train split and 795 samples in test split. It collects labels for semantic segmentation and monocular depth estimation, from which the labels of surface normal estimation and object boundary detection can be generated. We utilize all the tasks provided in these datasets for evaluation.", + "bbox": [ + 76, + 583, + 467, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics We use the same evaluation metrics as previous works [7]. Specifically, semantic segmentation (Semseg) and human parsing (Parsing) tasks are evaluated by the mean Intersection over Union (mIoU). Monocular depth estimation task (Depth) is evaluated by Root Mean", + "bbox": [ + 76, + 825, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/da73c8b4f6439380bfd4792102daab4c843c140212fa23ad2395ce8eac9170b9.jpg", + "table_caption": [ + "Table 1: Ablation study on PASCAL-Context with a ViT-Base backbone. \"MoE\" denotes using Mixture-of-Experts module; \"CG\" denotes using context-aware gating; \"Mem\" denotes using our multi-task feature memory. $\\downarrow$ ’ indicates lower better and ‘↑’ means higher better." + ], + "table_footnote": [], + "table_body": "
ModelSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑MTL Perf Δm ↑
Baseline76.4065.7084.6313.8369.50-3.70
w/ MoE77.6266.3784.6513.7770.40-2.88
w/ MoE+CG77.9966.8284.7613.6471.70-2.07
w/ MoE+CG+Mem78.4567.3884.9613.5572.30-1.46
", + "bbox": [ + 503, + 165, + 890, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Square Error (RMSE). Surface normal estimation task (Normal) is evaluated by the mean error (mErr) of angles. Saliency detection task (Saliency) is evaluated by maximal F-measure (maxF). Object boundary detection task (Boundary) is evaluated by the optimal-dataset-scale F-measure (odsF). We calculate the mean relative difference across all tasks compared to the single-task baseline, which is noted as Multi-task Performance (MTL Perf $\\Delta_{m}$ ) [5].", + "bbox": [ + 498, + 232, + 890, + 352 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details We use 5 expert networks in MMoE for PASCAL-Context and 4 expert networks for NYUD-v2, which are the same as the corresponding number of tasks. We add MMoE at the end of each transformer layer if not otherwise stated. For model training, we follow the setting of InvPT [7] and use Adam optimizer to train the models for 40,000 iterations with a batch size of 4. Polynomial learning rate scheduler is used with a learning rate of $2 \\times 10^{-5}$ and a weight decay rate of $1 \\times 10^{-6}$ . We adopt the loss functions and weights for different tasks in [7]. The ViT backbones are pre-trained on ImageNet-22K [51].", + "bbox": [ + 498, + 353, + 890, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 530, + 651, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Variants Definition To examine the effectiveness of TaskExpert under a fair setting, we compare it to the multi-task baseline model and its variants with the same backbone. (1) The baseline model (\"Baseline\") uses ViT-Base as backbone, and designs task-specific decoders in each transformer layer to generate layer-wise task-specific features. The task-specific features from all layers are added together as multi-scale input of the corresponding prediction head for each task. These layer-wise decoders share the same structure as the expert networks used in MMoE. The prediction heads for all the tasks have a similar structure: $\\mathrm{Conv}(3\\times 3)$ -BN-ReLU-Linear, which are the same as those used in TaskExpert. (2) \"TaskExpert w/ MoE\" adds the expert and gating networks of MMoE without context-aware gating to Baseline. It uses gating networks to control the contribution of each expert to each task in a sample-dependent manner without context-aware gating or multi-task feature memory. The number of experts is the same as the number of tasks. (3) \"TaskExpert w/ MoE+CG\" further uses our context-aware gating strategy, which improves the gating networks by introducing the contextual information of each token. (4) \"TaskExpert w/ MoE+CG+Mem\" is the full version of our TaskExpert using multi-task feature", + "bbox": [ + 498, + 553, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21833", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d7d2f108e35935847d8bf42d22a1760fda9c68894d59035d49c05ffd47068495.jpg", + "image_caption": [ + "Figure 6: Visualization of the task-gating scores by task-specific gating networks and the decomposed representative features from expert networks in MMoE modules. The gating scores for different experts show distinct distributions, while the representative features are discriminative to different visual elements. These results suggest different experts are able to decompose diverse visual information from the backbone feature." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7a7eee4a7e5bb9c92731e62b4c8fb1e027dcbeddc18976ec1f42aef2aa0cba2.jpg", + "image_caption": [ + "Figure 7: Qualitative comparison with SOTA method (i.e. InvPT [7]) on PASCAL-Context. Our TaskExpert generates better predictions with more accurate details as highlighted." + ], + "image_footnote": [], + "bbox": [ + 81, + 306, + 467, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "memory in the proposed MMoE module to establish modeling of long-term task-specific representation.", + "bbox": [ + 75, + 489, + 467, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key Results We analyze the performances of different model variants introduced above in Table 1. We find that using MoE brings a clear performance gain compared to the baseline model, which demonstrates the effectiveness of designing task-specific gating networks to aggregate the representative task-generic features for dynamically decoding task-specific features. For example, mIoU of semantic segmentation is improved from 76.40 to 77.62 (+1.22), and human parsing is improved from 65.70 to 66.37 (+0.67). Furthermore, the usage of context-aware gating strategy helps MoE achieves better performance by enabling the learning of contextual information for each token when computing the gating scores. For example, odsF of boundary detection is improved from 70.40 to 71.70 (+1.30) compared to only using MoE. Finally, with long-term modeling of task-specific representations via the proposed multi-task feature memory of Memorial Mixture-of-Experts, TaskExpert achieves significant performance improvement compared to the baseline model on different tasks, and the overall multitask performance $\\Delta_{m}$ is improved by +2.24.", + "bbox": [ + 75, + 521, + 467, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Study of Spare and Dense Connections in MoE The original MoE models use the dense connection as the gating mechanism: all experts are activated and contribute to the output [38, 39]. Recent works propose to utilize sparse connection as the gating mechanism: for each token, only a", + "bbox": [ + 75, + 824, + 467, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b12de7c3f03495262494cc966725a4a868e0ccb45b0b7634d314b5a5605d2040.jpg", + "table_caption": [ + "Table 2: Variants of TaskExpert on PASCAL-Context. The inference GPU memory is measured with batch size 1." + ], + "table_footnote": [], + "table_body": "
VariantSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑#ParamFLOPSInference GPU Mem
3 MMoE78.0466.5484.8513.6271.20174M597G8367M
6 MMoE78.1866.7984.9313.5971.70234M659G8851M
12 MMoE78.4567.3884.9613.5572.30347M782G12503M
ViT-S75.0462.6884.6814.2268.8055M204G4187M
ViT-B78.4567.3884.9613.5572.30347M782G12503M
", + "bbox": [ + 504, + 112, + 888, + 185 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/66395fec9962adb97f21b634e18b238b5aeec70bffed3b79e1e2cfee9c4c4efe.jpg", + "table_caption": [ + "Table 3: Comparison with multi-task MoE methods with ViT-S backbone on PASCAL-Context. The training GPU time is shown in the last column." + ], + "table_footnote": [], + "table_body": "
VariantSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑#ParamFLOPSGPU Hours
M3ViT [45]72.8062.1066.3014.5071.7042M420G~72
Mod-Squad [46]74.1062.7066.9013.7072.0050M420G~72
Ours75.0462.6884.6814.2268.8055M204G~5
", + "bbox": [ + 504, + 233, + 888, + 285 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cf71a362c7c727b01280fd9e24de5f1fdaeed2b6b14dde736d3b629baa8989e5.jpg", + "table_caption": [ + "Table 4: Comparison with the state-of-the-art methods with the same backbone on PASCAL-Context. Our TaskExpert clearly achieves superior performances on all tasks." + ], + "table_footnote": [], + "table_body": "
ModelFLOPs#ParamSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑
PAD-Net [8]124G81M53.6059.6065.8015.3072.50
MTI-Net [1]161G128M61.7060.1884.7814.2370.80
ATRC [2]216G96M67.6762.9382.2914.2472.42
PAD-Net w/ ViT-L [8]773G330M78.0167.1279.2114.3772.60
MTI-Net w/ ViT-L [1]774G851M78.3167.4084.7514.6773.00
ATRC w/ ViT-L [2]871G340M77.1166.8481.2014.2372.10
InvPT [7]669G423M79.0367.6184.8114.1573.00
TaskExpert (ours)622G420M80.6469.4284.8713.5673.30
", + "bbox": [ + 504, + 342, + 890, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "subset of experts contribute to the output [41], as explained in Section 3.3. In this experiment, we study the performances of using sparse connection (top 1 and top 2), and compare them with the default dense connection manner on PASCAL-Context. The results are demonstrated in Figure 8. It can be observed that activating more experts leads to better performance, and the dense connection achieves the best results. The reason is that it is able to freely utilize the full representative ability of all expert networks.", + "bbox": [ + 496, + 469, + 890, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Light-Weight Versions of TaskExpert To study the model performance and efficiency with a smaller parameter and memory budget, we conduct 2 sets of experiments in Table 3: (i) Using different numbers of MMoE modules on ViT-B backbone for PASCAL-Context. The MMoE modules are embedded evenly along the backbone's depth. (ii) Implementing TaskExpert on ViT-S, which is a smaller backbone. These smaller variants achieve reasonable performance with less parameter and memory consumption.", + "bbox": [ + 496, + 606, + 890, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 498, + 755, + 792, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We implement TaskExpert based on ViT-Large backbone with MMoE modules embedded at every two transformer layers, then we compare TaskExpert to recent SOTA methods in literature on both PASCAL-Context and NYUDv2 in Table 4 and Table 5. We re-implement several CNN-based SOTA multi-task learning methods on ViT-L backbone for a more comprehensive comparison on PASCAL-Context. Compared with the previous SOTA", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21834", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96a37487975fa33197668afd601913c587312b4572de77d778616d84ef74ef2e.jpg", + "image_caption": [ + "Figure 8: Comparison of TaskExpert with dense and sparse connection in MMoE. \"Top 1\" denotes only activating the expert with the highest gating score, \"Top 2\" means activating the experts with top 2 gating scores, and \"Dense\" activates all experts." + ], + "image_footnote": [], + "bbox": [ + 80, + 64, + 888, + 195 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8c733f4d6373526952daed207b83ba6f9414547a1caa83804153701beeadfc8f.jpg", + "table_caption": [ + "Table 5: Comparison with the state-of-the-art methods on NYUD-v2 dataset. Our TaskExpert achieves the best performances on all tasks." + ], + "table_footnote": [], + "table_body": "
ModelBackboneSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑
Cross-Stitch [25]HRNet1836.340.629020.8876.38
PAP [29]HRNet1836.720.617820.8276.42
PSD [31]HRNet1836.690.624620.8776.42
PAD-Net [8]HRNet1836.610.627020.8576.38
MTI-Net [1]HRNet4845.970.536520.2777.86
ATRC [2]HRNet4846.330.536320.1877.94
InvPT [7]ViT-L53.560.518319.0478.10
TaskExpert (ours)ViT-L55.350.515718.5478.40
", + "bbox": [ + 83, + 286, + 468, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model (i.e. InvPT [7]), TaskExpert achieves clearly better performances on all 9 metrics of these two challenging benchmarks with the same backbone. Particularly, on PASCAL-Context, we improve semantic segmentation by +1.61 and human parsing by +1.81. On NYUD-v2, semantic segmentation is improved by +1.79. It should be noted that, our gains are measured upon the very strong performance foundation established by InvPT (e.g. Semseg 79.03, Saliency 84.81 on PASCAL-Context). This superior performance clearly shows the effectiveness of the proposed TaskExpert. TaskExpert also maintains a higher efficiency with a reduction of 3 million in the number of parameters and 47 billion FLOPs compared to InvPT.", + "bbox": [ + 75, + 422, + 468, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As underscored in the Related Work section, our TaskExpert is designed for a different setting from $\\mathrm{M}^3\\mathrm{ViT}$ [45] and Mod-Squad [46]. Their models are constrained to generating a single task prediction per forward pass, while our model capably generates predictions for all tasks within a single forward pass. Therefore, the multi-task learning efficiency of our model is significantly higher than their methods. When utilizing the same backbone, it requires only half the FLOPS, as demonstrated in Table 3. Moreover, our model's training duration is markedly shorter since it learns all tasks concurrently in each iteration. While they necessitate around 18 hours of training with 4 A6000 GPUs as confirmed by their authors, our model completes training in under 5 hours using a single NVIDIA 3090 GPU.", + "bbox": [ + 75, + 619, + 470, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative study", + "text_level": 1, + "bbox": [ + 76, + 845, + 245, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Study of MMoE To understand better the learned gating scores and representative task-generic fea", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tures by the gating networks and expert networks in our MMoE modules, we present qualitative examples in Figure 6. As can be observed, TaskExpert effectively learns representative task-generic features with distinct activation distributions, focusing on different aspects of the image. Accordingly, the gating scores exhibit different response intensities to different areas for selectively assembling the experts' features for decoding task-specific features.", + "bbox": [ + 496, + 243, + 890, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PCA Visualization of Representative Features of Task-Experts at Different Layers We conduct principal component analysis (PCA) on the representative features at layers 1, 6, and 12. As shown in Fig. 4, we visualize the first three components of the feature PCA as a color map. It reveals that the task experts at deeper layers capture more high-level semantic information of the input. On the other hand, in the shallower layers, the task-feature experts focus more on low-level visual patterns (e.g., edges).", + "bbox": [ + 496, + 364, + 890, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Comparison of Decoded Task Features In Fig. 5, we also conduct PCA on the final task features generated by InvPT [7] and our TaskExpert, and draw the first three components. The features for different tasks from TaskExpert show clear task-specific characteristics and are much more discriminative than those of InvPT.", + "bbox": [ + 496, + 500, + 890, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Comparison of Prediction Results To intuitively examine the prediction quality of TaskExpert compared to the previous best-performing method [7], we visualize the predictions of the two models on the test split of PASCAL-Context in Figure 7. Our TaskExpert generates visually better multi-task predictions.", + "bbox": [ + 496, + 590, + 890, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 694, + 617, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present TaskExpert, a novel multitask mixture-of-expert model establishing new SOTA performances on two challenging benchmarks. TaskExpert can effectively perform dynamic decoding of task-specific features, from a set of representative task-generic features learned by expert networks. TaskExpert further models the long-term task-specific representations at different layers via a designed multi-task feature memory. Extensive experiments clearly prove the effectiveness and efficiency of our TaskExpert model, and we hope that our TaskExpert can also be beneficial for the future study of dynamic representation learning problems.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21835", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 78, + 90, + 236, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported in part by HKUST-SAIL joint research funding, the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, and HKUST Startup Fund No. R9253.", + "bbox": [ + 76, + 114, + 468, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 191, + 173, + 208 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In ECCV, 2020.", + "[2] David Bruggemann, Menelaos Kanakis, Anton Obukhov, Stamatos Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In ICCV, 2021.", + "[3] Amir R Zamir, Alexander Sax, Nikhil Cheerla, Rohan Suri, Zhangjie Cao, Jitendra Malik, and Leonidas J Guibas. Robust learning through cross-task consistency. In CVPR, 2020.", + "[4] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, 2017.", + "[5] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In CVPR, 2019.", + "[6] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, 2018.", + "[7] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. In ECCV, 2022.", + "[8] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In CVPR, 2018.", + "[9] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017.", + "[10] Yu Zhang and Qiang Yang. A survey on multi-task learning. TKDE, 2021.", + "[11] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. TPAMI, 2021.", + "[12] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018.", + "[13] Menelaos Kanakis, Thomas E Huang, David Brüggemann, Fisher Yu, and Luc Van Gool. Composite learning for robust and effective dense predictions. In WACV, 2023.", + "[14] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In ECCV, 2022." + ], + "bbox": [ + 78, + 217, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian Koring, Suman Saha, and Luc Van Gool. Three ways to improve semantic segmentation with self-supervised depth estimation. In CVPR, 2021.", + "[16] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In CVPR, 2019.", + "[17] Xiangtai Li, Henghui Ding, Wenwei Zhang, Haobo Yuan, Jiangmiao Pang, Guangliang Cheng, Kai Chen, Ziwei Liu, and Chen Change Loy. Transformer-based visual segmentation: A survey. arXiv preprint arXiv:2304.09854, 2023.", + "[18] Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. Conflict-averse gradient descent for multi-task learning. In NeurIPS, 2021.", + "[19] Siwei Yang, Hanrong Ye, and Dan Xu. Contrastive multitask dense prediction. In AAAI, 2023.", + "[20] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Learning multiple dense prediction tasks from partially annotated data. In CVPR, 2022.", + "[21] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. In NeurIPS, 2020.", + "[22] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In ICML, 2018.", + "[23] Zirui Wang, Yulia Tsvetkov, Orhan First, and Yuan Cao. Gradient vaccine: Investigating and improving multi-task optimization in massively multilingual models. In ICLR, 2020.", + "[24] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In NeurIPS, 2020.", + "[25] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In CVPR, 2016.", + "[26] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In CVPR, 2019.", + "[27] Lijun Zhang, Xiao Liu, and Hui Guan. Automtl: A programming framework for automating efficient multi-task learning. In NeurIPS, 2021.", + "[28] Yuan Gao, Haoping Bai, Zequn Jie, Jiayi Ma, Kui Jia, and Wei Liu. Mtl-nas: Task-agnostic neural architecture search towards general-purpose multi-task learning. In CVPR, 2020.", + "[29] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In CVPR, 2019.", + "[30] Xiaoya Zhang, Ling Zhou, Yong Li, Zhen Cui, Jin Xie, and Jian Yang. Transfer vision patterns for multi-task pixel learning. In ACMMM, 2021." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21836", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In CVPR, 2020.", + "[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017.", + "[33] Simon Vandenhende, Stamatios Georgoulis, Bert De Brabandere, and Luc Van Gool. Branched multi-task networks: deciding what layers to share. In ECCV, 2020.", + "[34] Yongxi Lu, Abhishek Kumar, Shuangfei Zhai, Yu Cheng, Tara Javidi, and Rogerio Feris. Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In CVPR, 2017.", + "[35] Pengsheng Guo, Chen-Yu Lee, and Daniel Ulbricht. Learning to branch for multi-task learning. In ICML, 2020.", + "[36] Hanrong Ye and Dan Xu. Taskprompt: Spatial-channel multi-task prompting for dense scene understanding. In ICLR, 2023.", + "[37] Michael I Jordan and Robert A Jacobs. Hierarchical mixtures of experts and the em algorithm. Neural Computation, 6(2):181-214, 1994.", + "[38] Robert A Jacobs and Michael I Jordan. Learning piecewise control strategies in a modular neural network architecture. IEEE Transactions on Systems, Man, and Cybernetics, 23(2):337-345, 1993.", + "[39] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. *Neural Computation*, 3(1):79-87, 1991.", + "[40] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013.", + "[41] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017.", + "[42] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. JMLR, 23:1-39, 2022.", + "[43] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. In ICLR, 2021.", + "[44] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In KDD, 2018.", + "[45] Hanxue Liang, Zhiwen Fan, Rishov Sarkar, Ziyu Jiang, Tianlong Chen, Kai Zou, Yu Cheng, Cong Hao, and Zhangyang Wang. M3vit: Mixture-of-experts vision transformer for efficient multi-task learning with model-accelerator co-design. In NeurIPS, 2022.", + "[46] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik Learned-Miller, and Chuang Gan. Mod-squad: Designing mixture of experts as modular multitask learners. arXiv preprint arXiv:2212.08066, 2022." + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021.", + "[48] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In CVPR, 2014.", + "[49] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010.", + "[50] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012.", + "[51] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + ], + "bbox": [ + 501, + 92, + 890, + 375 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21837", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/2e3b51ba-4260-4b59-a865-225826178569_model.json b/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/2e3b51ba-4260-4b59-a865-225826178569_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b9e4b7e8358f58165c1d567fe219f5e6972452d7 --- /dev/null +++ b/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/2e3b51ba-4260-4b59-a865-225826178569_model.json @@ -0,0 +1,1914 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.131, + 0.82, + 0.177 + ], + "angle": 0, + "content": "TaskExpert: Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.203, + 0.723, + 0.258 + ], + "angle": 0, + "content": "Hanrong Ye and Dan Xu\\* \nDepartment of Computer Science and Engineering, HKUST \nClear Water Bay, Kowloon, Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.26, + 0.595, + 0.275 + ], + "angle": 0, + "content": "{hyeae, danxu}@cse.ust.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.473, + 0.796 + ], + "angle": 0, + "content": "Learning discriminative task-specific features simultaneously for multiple distinct tasks is a fundamental problem in multi-task learning. Recent state-of-the-art models consider directly decoding task-specific features from one shared task-generic feature (e.g., feature from a backbone layer), and utilize carefully designed decoders to produce multi-task features. However, as the input feature is fully shared and each task decoder also shares decoding parameters for different input samples, it leads to a static feature decoding process, producing less discriminative task-specific representations. To tackle this limitation, we propose TaskExpert, a novel multi-task mixture-of-experts model that enables learning multiple representative task-generic feature spaces and decoding task-specific features in a dynamic manner. Specifically, TaskExpert introduces a set of expert networks to decompose the backbone feature into several representative task-generic features. Then, the task-specific features are decoded by using dynamic task-specific gating networks operating on the decomposed task-generic features. Furthermore, to establish long-range modeling of the task-specific representations from different layers of TaskExpert, we design a multitask feature memory that updates at each layer and acts as an additional feature expert for dynamic task-specific feature decoding. Extensive experiments demonstrate that our TaskExpert clearly outperforms previous best-performing methods on all 9 metrics of two competitive multi-task learning benchmarks for visual scene understanding (i.e., PASCAL-Context and NYUD-v2). Codes and models will be made publicly available." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.829, + 0.208, + 0.845 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "With the rapid development of deep learning models, many computer vision tasks can be effectively handled by deep networks, which strongly motivates researchers to de" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.308, + 0.859, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Figure 1: Motivative illustration of TaskExpert for dynamical decoding discriminative task-specific representations. A set of \\( K \\) expert networks decompose the task-generic backbone feature into \\( K \\) representative feature spaces. Then, the feature activations of an input from the multiple expert networks are dynamically assembled by the Task Gating network, to produce task-specific features for the \\( T \\) tasks. TaskExpert further performs long-range modeling of task-specific representations by designing a multi-task feature memory. It is updated at the current layer based on the assembled task-specific representations, and is used as an additional feature expert in the next layer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.892, + 0.881 + ], + "angle": 0, + "content": "velop unified multi-task networks to perform joint learning and inference on multiple distinct tasks [1, 2, 3, 4]. Multi-Task Learning (MTL) can avoid repetitive training of single-task models and is able to simultaneously generate predictions for different tasks with only one-time inference. On the other hand, as different computer vision tasks usually share a basic common understanding of input images, learning multiple tasks concurrently can help improve the representation power of each task and boost the overall performance of multi-task models [5, 6]." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Learning discriminative task-specific features is a funda-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21828" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.424 + ], + "angle": 0, + "content": "mental problem for MTL models. To achieve this objective, recent state-of-the-art works [7, 2, 8, 1] consider a decoder-focused paradigm, in which a shared backbone model with pretrained parameters is applied to learn generic image representations, and then separate task decoders are carefully designed to generate task-specific features. Although this paradigm is straightforward to achieve task-specific feature decoding, and produces promising multi-task prediction results, these models have two aspects of potential limitations: (i) The input task-generic feature from a backbone layer is fully shared for the different task decoders, which requires a careful design of the decoders (e.g. capacity and structure) for the different tasks to achieve task-discriminative decoding. (ii) The parameters of each task decoder are typically shared for different input samples, resulting in a static decoding of task-specific features, while a sample-dependant decoder is highly beneficial for learning more discriminative task features, as the input samples can be diverse, and sample-related context information is particularly important for multi-task learning and inference. All these aspects hamper the previous models from learning effective task-specific features and predictions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.426, + 0.473, + 0.848 + ], + "angle": 0, + "content": "To tackle the above-mentioned issues, this paper proposes a novel multi-task mixture-of-experts framework, coined as \"TaskExpert\", which dynamically performs the task-feature decoding for different inputs and tasks as depicted in Figure 1. Instead of using one shared backbone feature to decode different task features, the proposed TaskExpert introduces a set of task-generic expert networks that learn to decompose the backbone feature into a set of representative task-generic features. The feature decomposition enables the decoders to interact with finer-granularity task-generic feature spaces to produce more discriminative task features, as each expert can be responsible for modeling one representative feature space from the training data. Then, we dynamically assemble feature activations from different experts to produce task-specific representations, based on sample-dependent and task-specific gating scores predicted from a designed gating network. On the other hand, visual understanding tasks greatly benefit from representations of different network levels [9]. Thus, we further devise a \"multi-task feature memory\" to aggregate the dynamically decoded task-specific features produced at different network layers, and propose a \"Memorial Mixture-of-Experts (MMoE)\" module. For each task, MMoE utilizes its corresponding task feature from the multi-task feature memory as an additional feature expert, to improve task-specific feature decoding. In this way, MMoE enables long-range dynamic modeling of task-specific representation for each task throughout the entire backbone." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.849, + 0.415, + 0.864 + ], + "angle": 0, + "content": "In summary, the main contribution is three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We propose a multi-task mixture-of-experts model that allows effective decomposition of a task-generic fea" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "ture from any backbone network layer and enables dynamically assembling discriminative multi-task representations by a task gating mechanism." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.145, + 0.892, + 0.252 + ], + "angle": 0, + "content": "- We design a novel Memorial Mixture-of-Experts (MMoE) module by further introducing a multi-task feature memory, which can effectively interact (read and write) with the mixture of experts at different network layers to achieve long-range dynamic decoding of task-specific features by utilizing different levels of network representations." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.259, + 0.895, + 0.396 + ], + "angle": 0, + "content": "- The proposed TaskExpert is extensively validated on two challenging multi-task visual scene understanding benchmarks (i.e. Pascal-Context and NYUD-v2). The results clearly show the effectiveness of TaskExpert for dynamic multi-task representation learning, and also establish new state-of-the-art performances on all 9 metrics on the benchmarks, regarding the comparison with previous best-performing methods using both transformer- and CNN-based backbones." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.651, + 0.423 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Multi-Task Learning in Computer Vision In the field of multi-task learning (MTL) for computer vision problems [10, 11, 12, 13, 14, 15, 16, 17], some research works explore effective solutions through a perspective of model optimization and improve the optimization process of MTL via different multi-task loss designs [3, 6, 18, 19, 20] and gradient manipulations [21, 22, 23, 24]. These optimization strategies have also been demonstrated beneficial for alleviating the issue of task competition in multi-task learning [6]. From another perspective, several existing works also investigate designing better multi-task model architectures. These related works can be roughly divided into two groups: encoder-focused and decoder-focused methods. The encoder-focused methods [25, 26, 27, 28] target designing multi-task encoders, while the decoder-focused methods typically share the encoder (i.e. the backbone) for different tasks and design sophisticated decoders for task-specific feature learning and prediction [1, 2, 29, 8, 30, 31]. A recent state-of-the-art decoder-focused method (i.e. In-vPT [7]) performs multi-task learning on powerful transformer architectures [32], and establishes a very strong baseline for multi-task visual scene understanding. It can effectively learn spatial and cross-task relationships globally. To adapt multi-task models to different tasks, researchers also propose to search for task-specific network structures for encoders [33, 34, 35] and decoders [2], or utilize task-specific learnable prompts to effectively capture task-specific information [36]. Despite promising performances of these works, they are still limited by static decoder network designs since their decoders cannot dynamically generate task-specific representations based on dif" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21829" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.166 + ], + "angle": 0, + "content": "ferent inputs and tasks. In this paper, we propose TaskExpert, a novel and effective mixture-of-experts framework, which can achieve dynamic multi-task representation learning via feature decomposition and assembling through the entire backbone network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.473, + 0.593 + ], + "angle": 0, + "content": "Mixture-of-Experts Models With statistical motivation [37], Mixture-of-Experts (MoE) models are originally designed to control the dynamics of neural networks automatically [38, 39]. MoE learns a series of expert networks and a gating network. The outputs of expert networks are weighted by gating scores (or called \"gate values\") generated by a gating network before the weighting operation. In more recent works, some researchers use the gating scores as a criterion to sparsely select only one or a few experts. The sparse activation of experts enables a significant reduction of the computational cost when training large-scale models [40, 41, 42, 43]. Related to our work, [44] designs a multi-gate MoE to ensemble expert networks for different census analysis tasks with different gating networks, and [45, 46] propose sparse task-conditioned MoE networks for efficient multi-task learning, which generate predictions for only one task within a single forward network inference. In contrast to these works, our Memorial Mixture-of-Experts (MMoE) is a plug-and-play decoder module, which is designed for dynamically assembling discriminative task-specific features. Our TaskExpert can simultaneously generate multi-task predictions for all tasks in one single forward inference, leading to significantly higher multi-task training efficiency and enabling cross-task interaction learning. Moreover, MMoE models task-specific features throughout different network layers via the novel design of multi-task feature memory and help generate more discriminative task-specific features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.609, + 0.201, + 0.627 + ], + "angle": 0, + "content": "3. TaskExpert" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.283, + 0.65 + ], + "angle": 0, + "content": "3.1. Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.903 + ], + "angle": 0, + "content": "An overall illustration of our TaskExpert framework is shown in Figure 2. The input image is first projected to a token sequence and then fed into a series of classic visual transformer layers. Our framework can also be flexibly used with a CNN backbone. The proposed Memorial Mixture-of-Experts (MMoE) module is embedded at different network layers, and each module takes the task-generic backbone feature of that layer as input. It dynamically assembles and outputs discriminative task-specific representations at the corresponding layer. The MMoE further interacts (reads and writes operations) with the designed multi-task feature memory, to perform a long-range dynamic multi-task feature decoding, which benefits from different levels of network features. The reads utilize the feature memory as an additional feature expert in the MMoE module, and the writes update the memory at the layer. Finally, after be" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.094, + 0.887, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.366, + 0.895, + 0.548 + ], + "angle": 0, + "content": "Figure 2: Overall architecture of the proposed TaskExpert model. Memorial Mixture-of-Experts (MMoE) modules are embedded at different backbone layers to dynamically decode task-specific feature for each task at the corresponding layer. The MMoE also interacts (reads and writes) with a designed multi-task feature memory to have a long-range dynamic multi-task feature decoding benefiting from different levels of network features. The reads utilize the memory as an additional feature expert in MMoE, and the writes update the memory at the layer. After being updated by the last network layer, the multi-task feature memory is used for the final prediction of each task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.646 + ], + "angle": 0, + "content": "ing updated by the outputs of all MMoE modules at different network layers, the task-specific features in multi-task feature memory are then used for generating the multi-task predictions with simple prediction heads. We elaborate on more details of our model in the following several sections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.674, + 0.782, + 0.69 + ], + "angle": 0, + "content": "3.2. Multi-Task Backbone Structure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We consider a widely-used vision transformer (ViT) model [47] with \\(L\\) layers as the backbone of our multi-task MoE model. The input image is first divided into \\(H\\times W\\) patches and processed by the patch embedding and positional embedding modules. Then, we obtain a sequence of patch tokens \\(\\mathbf{X}^0 = \\{\\mathbf{x}_1^0,\\mathbf{x}_2^0,\\dots,\\mathbf{x}_N^0\\}\\), where \\(N = H\\times W\\) and the number of channels is \\(C\\). The patch tokens contain generic visual representations of the input image. We denote the \\(l\\)-th transformer layer as a mapping function \\(f^{l}\\) and the updating of image patch tokens at different transformer layers can be formulated as \\(\\mathbf{X}^l = f^l (\\mathbf{X}^{l - 1})\\), where \\(\\mathbf{X}^l\\in \\mathbb{R}^{N\\times C}\\) denotes the output patch tokens of the transformer at the \\(l\\)-th layer." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21830" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.066, + 0.873, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.348, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Figure 3: Illustration of the proposed Memorial Mixture-of-Experts (MMoE) module with the multi-task feature memory. MMoE accepts a task-generic feature from a backbone layer as input, and dynamically generates task-specific features for different tasks in a sample- and task-dependent manner. Specifically, MMoE decomposes the backbone feature to a set of \\( K \\) representative task-generic features \\( (\\mathbf{R}_1^l, \\dots, \\mathbf{R}_K^l) \\) via the \\( K \\) expert networks. Then these features are used to assemble a set of \\( T \\) task-specific features based on dynamically generated task-gating scores from the task gating networks. To establish long-range modeling of task-specific representations crossing different network layers, the multi-task feature memory is designed to interact with the expert networks. The multi-task feature memory serves as an additional feature expert to decode the task-specific representations, and the decoded representations are also further used to update the multi-task feature memory." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.483, + 0.415, + 0.5 + ], + "angle": 0, + "content": "3.3. Memorial Mixture-of-Experts (MMoE)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.471, + 0.629 + ], + "angle": 0, + "content": "We present the details of several core parts of the proposed memorial mixture-of-experts (MMoE) module, including backbone generic-feature decomposition with a set of network experts, context-aware task-specific gating, the multi-task feature memory to have long-range modeling on task-specific features through interactions (reads and writes) with the mixture of experts, dynamic task-specific feature assembling, and multi-task prediction heads." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.46, + 0.667 + ], + "angle": 0, + "content": "3.3.1 Backbone Feature Decomposition with Experts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.901 + ], + "angle": 0, + "content": "As shown in Figure 3, to enable learning discriminative task-specific representations from the task-generic backbone feature, we first define a set of expert networks that decompose the backbone feature into several representative task-generic features in sub-feature spaces, representing distinct patterns of the backbone features, as can be observed in Figure 6. It allows more fine-grained decoding of task-specific features. Specifically, at the \\(l\\)-th backbone layer, let us assume that we use \\(K\\) expert networks \\(\\{f_{e}^{l,i}(\\cdot):\\mathbb{R}^{N\\times C}\\to \\mathbb{R}^{N\\times C},i\\in [1,K]\\}\\) to learn distinct task-generic feature projections, where superscript \\(l\\) denotes the layer index and \\(i\\) is the expert index. Each expert network employs the same Linear-BatchNorm-RELU-Linear structure but with different parameters. Through each expert network, we obtain a set of representative fea" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.482, + 0.741, + 0.5 + ], + "angle": 0, + "content": "tures \\(\\{\\mathbf{R}_i^l\\in \\mathbb{R}^{N\\times C},i\\in [1,K]\\}\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.51, + 0.892, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {i} ^ {l} = f _ {e} ^ {l, i} \\left(\\mathbf {X} ^ {l}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.54, + 0.833, + 0.555 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}^l\\) is the \\(l\\)-th layer feature of the backbone." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.809, + 0.59 + ], + "angle": 0, + "content": "3.3.2 Context-Aware Task-Specific Gating" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.704 + ], + "angle": 0, + "content": "After we have obtained a set of representative task-generic features decomposed from a backbone feature for all the tasks, we aim at decoding discriminative task-specific features for different tasks by automatically assembling these representative features. Thus, we propose to adaptively control the contribution of each expert for each patch token by designing a task-specific gating network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.856 + ], + "angle": 0, + "content": "For each token in the representative task-generic features, we compute a task-gating score to control its contribution to generating the task-specific features. The task-gating scores of each task are dynamically generated by a task-specific gating network, using the features \\(\\mathbf{X}^l\\) from the \\(l\\)-th network layer. We term the task-specific gating network with a short name \"task gating\", and denote it as \\(f_{g}^{l,t}(\\cdot):\\mathbb{R}^{N\\times C\\to N\\times K}\\), where \\(t\\in [1,T]\\) (\\(T\\) is the total number of tasks) is a task index. The task-gating scores \\(\\mathbf{G}^{l,t}\\in \\mathbb{R}^{N\\times K}\\) is computed as \\(\\mathbf{G}^{l,t} = f_{g}^{l,t}(\\mathbf{X}^{l})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For the design of the gating networks, previous Mixture-of-Experts methods [42, 43, 41] use a simple linear layer followed by a Softmax function, which does not take into" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21831" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.074, + 0.885, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.209, + 0.895, + 0.255 + ], + "angle": 0, + "content": "Figure 4: PCA visualization of the representative features by task-generic experts across different layers. In the deeper layers, the features become increasingly responsive to high-level semantic areas. Conversely, features at more shallow layers show a heightened sensitivity to low-level patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.474, + 0.511 + ], + "angle": 0, + "content": "account the context of each image patch token, while using only the corresponding patch token when computing the gating score. As various challenging computer vision tasks (e.g. semantic segmentation and monocular depth estimation) require modeling and understanding each image patch with context information, we thus design a context-aware gating strategy. Specifically, context-aware gating adopts convolution kernels to extend the receptive field of the gating networks, and thus it can involve a local context of tokens in the computation of gating scores. We first reshape the input token sequence \\(\\mathbf{X}\\) (in \\(\\mathbb{R}^{N\\times C}\\)) as spatial maps (in \\(\\mathbb{R}^{H\\times W\\times C}\\)), and then compute the gating scores of \\(K\\) experts with three layers, each having a \\(3\\times 3\\) convolution with ReLU as intermediate activation function. We denote this computation procedure as \"CONV\", which will also be used later in Equation 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.332, + 0.547 + ], + "angle": 0, + "content": "3.3.3 Multi-Task Feature Memory" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.474, + 0.875 + ], + "angle": 0, + "content": "To conduct long-range modeling of the multi-task representations decoded at different network layers that contain different levels of visual information, we design a multitask feature memory to interact with the MMoE modules through read and write operations. The read operation uses the task feature memory as an additional feature expert for task-specific feature decoding in MMoE. The write operation updates the multi-task feature memory by using the decoded task-specific features from MMoE. In this way, the multi-task feature memory can aggregate the decoded task-specific features crossing different network layers. Specifically, the read operation is an identity mapping. The write operation can be performed as follows: given a decoded task-specific feature of the \\(t\\)-th task at the first layer, i.e. \\(\\mathbf{F}^{1,t}\\), the initialization of the multi-task feature memory \\(\\mathbf{M}^{1,t}\\) is obtained by \\(\\mathbf{M}^{1,t} \\gets \\mathbf{F}^{1,t}\\). In the subsequent layers, suppose we decode task-specific features at every network layer, at the \\(l\\)-th layer (\\(l > 1\\)), we write the multi-task feature memory \\(\\{\\mathbf{M}^{l,t}, t \\in [1,T]\\}\\) with the decoded task-specific feature \\(\\mathbf{F}^{l,t}\\) weighted by a task-specific learnable momentum value \\(\\alpha^{l,t}\\) as shown in Figure 3:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.884, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} ^ {l, t} \\leftarrow \\mathbf {F} ^ {l, t} + \\alpha^ {l, t} \\mathbf {M} ^ {l - 1, t}, l > 1. \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.269, + 0.887, + 0.285 + ], + "angle": 0, + "content": "3.3.4 Dynamically Assembling Task-Specific Features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.893, + 0.675 + ], + "angle": 0, + "content": "With the obtained representative task-generic features and task-specific gating scores as above-described, we can perform dynamic decoding of task-specific features, as the gating parameters are predicted in a sample-dependent manner. This is significantly different from existing decoder-focused multi-task structures, which share the decoding parameters for all the different input samples. Specifically, for the \\(t\\)-th task, we multiply every token of the representative task-generic features \\(\\{\\mathbf{R}_i^l,i\\in [1,K]\\}\\) with their corresponding task-gating score \\(\\{\\mathbf{G}_i^{l,t},i\\in [1,K]\\}\\) and sum the tokens in the same spatial position to decode task-specific feature. Furthermore, to learn more discriminative features for each task, we utilize the corresponding task-specific feature in multi-task feature memory as an additional feature expert in MMoE, which contains long-range task-specific information learned throughout the entire network. It helps decode more effective and long-term task-specific features. As shown in Figure 3, for the MMoE at the \\(l\\)-th transformer layer (\\(l > 1\\)), we compute an additional gating score map \\(\\mathbf{G}_{K + 1}^{l,t}\\) for the task-specific feature from the multi-task feature memory \\(\\mathbf{M}^{l - 1,t}\\), which is performed by adding one more output channel into the task gating network \\(f_{g}^{l,t}:\\mathbb{R}^{N\\times C}\\to \\mathbb{R}^{N\\times (K + 1)}\\). In summary, the task gating scores of the \\(t\\)-th task and the \\(l\\)-th layer, i.e., \\(\\mathbf{G}^{l,t} = \\{\\mathbf{G}_1^{l,t},\\mathbf{G}_2^{l,t},\\dots,\\mathbf{G}_{K + 1}^{l,t}\\}\\), is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.681, + 0.892, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} ^ {l, t} = \\left\\{ \\begin{array}{l l} \\operatorname {S o f t m a x} \\left(\\operatorname {C O N V} ^ {K + 1} \\left(\\mathbf {X} ^ {l}\\right)\\right) & l > 1, \\\\ \\operatorname {S o f t m a x} \\left(\\operatorname {C O N V} ^ {K} \\left(\\mathbf {X} ^ {l}\\right)\\right) & l = 1, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.721, + 0.895, + 0.767 + ], + "angle": 0, + "content": "where the superscript \\(K\\) of CONV denotes the number of output convolution kernels. And accordingly, the decoded task-specific features at the \\(l\\)-th MMoE module is:" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.772, + 0.892, + 0.826 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} ^ {l, t} = \\left\\{ \\begin{array}{l l} \\sum_ {i = 1} ^ {K} \\mathbf {G} _ {i} ^ {l, t} \\times \\mathbf {R} _ {i} ^ {l} + \\mathbf {G} _ {K + 1} ^ {l, t} \\times \\mathbf {M} ^ {l - 1, t} & l > 1, \\\\ \\sum_ {i = 1} ^ {K} \\mathbf {G} _ {i} ^ {l, t} \\times \\mathbf {R} _ {i} ^ {l} & l = 1. \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Alternative sparse activation of experts As an alternative, we can also use \"sparse connections\" in MMoE, which means that we only activate \\( K_{sel} \\) selected experts to compute the task-specific features (\\( K_{sel} < K \\)) for each token [41]. In this case, for each token, we find the top \\( K_{sel} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21832" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.468, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.211, + 0.468, + 0.256 + ], + "angle": 0, + "content": "Figure 5: PCA visualization of the decoded final task features generated by InvPT [7] and our TaskExpert. Our model generates more discriminative features for all tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.27, + 0.468, + 0.332 + ], + "angle": 0, + "content": "experts to be activated based on the task-gating scores, and sum their outputs after being weighted by the task-gating scores. A comparison between the dense connection and sparse connection in MMoE is discussed in the experiments." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.339, + 0.4, + 0.354 + ], + "angle": 0, + "content": "3.4. Generation of Multi-Task Predictions" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.363, + 0.468, + 0.468 + ], + "angle": 0, + "content": "At the final network layer \\(L\\), we use task-specific prediction heads to directly generate prediction outputs for each task from the multi-task feature memory \\(\\{\\mathbf{M}^{L,t}, t \\in [1, T]\\}\\). The prediction heads for different tasks have a similar structure, consisting of a \\(3 \\times 3\\) convolution layer with batch normalization and ReLU activation, followed by a linear layer for the final prediction." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.481, + 0.201, + 0.498 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.506, + 0.468, + 0.552 + ], + "angle": 0, + "content": "In this section, we validate the effectiveness of our proposal through several aspects: (i) ablation study, (ii) comparison with SOTA methods, and (iii) qualitative analysis." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.561, + 0.266, + 0.577 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.584, + 0.468, + 0.824 + ], + "angle": 0, + "content": "Tasks and Benchmarks For evaluation, we adopt two popular benchmarks in multi-task learning as previous works [7, 1], i.e. PASCAL-Context [48, 49] and NYUD-v2 [50] datasets. PASCAL-Context contains 4,998 samples in train split and 5,105 samples in test split. It provides labels for multiple important visual understanding tasks, including semantic segmentation, human parsing, and object boundary detection. Previous work on multi-task learning generates pseudo labels for another two tasks [5]: surface normals estimation and saliency detection. Meanwhile, NYUD-v2 contains 1,449 samples in train split and 795 samples in test split. It collects labels for semantic segmentation and monocular depth estimation, from which the labels of surface normal estimation and object boundary detection can be generated. We utilize all the tasks provided in these datasets for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Metrics We use the same evaluation metrics as previous works [7]. Specifically, semantic segmentation (Semseg) and human parsing (Parsing) tasks are evaluated by the mean Intersection over Union (mIoU). Monocular depth estimation task (Depth) is evaluated by Root Mean" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.09, + 0.892, + 0.164 + ], + "angle": 0, + "content": "Table 1: Ablation study on PASCAL-Context with a ViT-Base backbone. \"MoE\" denotes using Mixture-of-Experts module; \"CG\" denotes using context-aware gating; \"Mem\" denotes using our multi-task feature memory. \\( \\downarrow \\) ’ indicates lower better and ‘↑’ means higher better." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.166, + 0.892, + 0.23 + ], + "angle": 0, + "content": "
ModelSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑MTL Perf Δm ↑
Baseline76.4065.7084.6313.8369.50-3.70
w/ MoE77.6266.3784.6513.7770.40-2.88
w/ MoE+CG77.9966.8284.7613.6471.70-2.07
w/ MoE+CG+Mem78.4567.3884.9613.5572.30-1.46
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.233, + 0.892, + 0.353 + ], + "angle": 0, + "content": "Square Error (RMSE). Surface normal estimation task (Normal) is evaluated by the mean error (mErr) of angles. Saliency detection task (Saliency) is evaluated by maximal F-measure (maxF). Object boundary detection task (Boundary) is evaluated by the optimal-dataset-scale F-measure (odsF). We calculate the mean relative difference across all tasks compared to the single-task baseline, which is noted as Multi-task Performance (MTL Perf \\(\\Delta_{m}\\)) [5]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.354, + 0.892, + 0.52 + ], + "angle": 0, + "content": "Implementation Details We use 5 expert networks in MMoE for PASCAL-Context and 4 expert networks for NYUD-v2, which are the same as the corresponding number of tasks. We add MMoE at the end of each transformer layer if not otherwise stated. For model training, we follow the setting of InvPT [7] and use Adam optimizer to train the models for 40,000 iterations with a batch size of 4. Polynomial learning rate scheduler is used with a learning rate of \\(2 \\times 10^{-5}\\) and a weight decay rate of \\(1 \\times 10^{-6}\\). We adopt the loss functions and weights for different tasks in [7]. The ViT backbones are pre-trained on ImageNet-22K [51]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.652, + 0.546 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.554, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Model Variants Definition To examine the effectiveness of TaskExpert under a fair setting, we compare it to the multi-task baseline model and its variants with the same backbone. (1) The baseline model (\"Baseline\") uses ViT-Base as backbone, and designs task-specific decoders in each transformer layer to generate layer-wise task-specific features. The task-specific features from all layers are added together as multi-scale input of the corresponding prediction head for each task. These layer-wise decoders share the same structure as the expert networks used in MMoE. The prediction heads for all the tasks have a similar structure: \\(\\mathrm{Conv}(3\\times 3)\\)-BN-ReLU-Linear, which are the same as those used in TaskExpert. (2) \"TaskExpert w/ MoE\" adds the expert and gating networks of MMoE without context-aware gating to Baseline. It uses gating networks to control the contribution of each expert to each task in a sample-dependent manner without context-aware gating or multi-task feature memory. The number of experts is the same as the number of tasks. (3) \"TaskExpert w/ MoE+CG\" further uses our context-aware gating strategy, which improves the gating networks by introducing the contextual information of each token. (4) \"TaskExpert w/ MoE+CG+Mem\" is the full version of our TaskExpert using multi-task feature" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21833" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.18, + 0.47, + 0.3 + ], + "angle": 0, + "content": "Figure 6: Visualization of the task-gating scores by task-specific gating networks and the decomposed representative features from expert networks in MMoE modules. The gating scores for different experts show distinct distributions, while the representative features are discriminative to different visual elements. These results suggest different experts are able to decompose diverse visual information from the backbone feature." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.307, + 0.468, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.426, + 0.47, + 0.473 + ], + "angle": 0, + "content": "Figure 7: Qualitative comparison with SOTA method (i.e. InvPT [7]) on PASCAL-Context. Our TaskExpert generates better predictions with more accurate details as highlighted." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.468, + 0.52 + ], + "angle": 0, + "content": "memory in the proposed MMoE module to establish modeling of long-term task-specific representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.468, + 0.824 + ], + "angle": 0, + "content": "Key Results We analyze the performances of different model variants introduced above in Table 1. We find that using MoE brings a clear performance gain compared to the baseline model, which demonstrates the effectiveness of designing task-specific gating networks to aggregate the representative task-generic features for dynamically decoding task-specific features. For example, mIoU of semantic segmentation is improved from 76.40 to 77.62 (+1.22), and human parsing is improved from 65.70 to 66.37 (+0.67). Furthermore, the usage of context-aware gating strategy helps MoE achieves better performance by enabling the learning of contextual information for each token when computing the gating scores. For example, odsF of boundary detection is improved from 70.40 to 71.70 (+1.30) compared to only using MoE. Finally, with long-term modeling of task-specific representations via the proposed multi-task feature memory of Memorial Mixture-of-Experts, TaskExpert achieves significant performance improvement compared to the baseline model on different tasks, and the overall multitask performance \\(\\Delta_{m}\\) is improved by +2.24." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.903 + ], + "angle": 0, + "content": "Study of Spare and Dense Connections in MoE The original MoE models use the dense connection as the gating mechanism: all experts are activated and contribute to the output [38, 39]. Recent works propose to utilize sparse connection as the gating mechanism: for each token, only a" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.083, + 0.892, + 0.113 + ], + "angle": 0, + "content": "Table 2: Variants of TaskExpert on PASCAL-Context. The inference GPU memory is measured with batch size 1." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.113, + 0.89, + 0.186 + ], + "angle": 0, + "content": "
VariantSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑#ParamFLOPSInference GPU Mem
3 MMoE78.0466.5484.8513.6271.20174M597G8367M
6 MMoE78.1866.7984.9313.5971.70234M659G8851M
12 MMoE78.4567.3884.9613.5572.30347M782G12503M
ViT-S75.0462.6884.6814.2268.8055M204G4187M
ViT-B78.4567.3884.9613.5572.30347M782G12503M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.189, + 0.892, + 0.233 + ], + "angle": 0, + "content": "Table 3: Comparison with multi-task MoE methods with ViT-S backbone on PASCAL-Context. The training GPU time is shown in the last column." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.234, + 0.89, + 0.286 + ], + "angle": 0, + "content": "
VariantSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑#ParamFLOPSGPU Hours
M3ViT [45]72.8062.1066.3014.5071.7042M420G~72
Mod-Squad [46]74.1062.7066.9013.7072.0050M420G~72
Ours75.0462.6884.6814.2268.8055M204G~5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.298, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Table 4: Comparison with the state-of-the-art methods with the same backbone on PASCAL-Context. Our TaskExpert clearly achieves superior performances on all tasks." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.343, + 0.891, + 0.452 + ], + "angle": 0, + "content": "
ModelFLOPs#ParamSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑
PAD-Net [8]124G81M53.6059.6065.8015.3072.50
MTI-Net [1]161G128M61.7060.1884.7814.2370.80
ATRC [2]216G96M67.6762.9382.2914.2472.42
PAD-Net w/ ViT-L [8]773G330M78.0167.1279.2114.3772.60
MTI-Net w/ ViT-L [1]774G851M78.3167.4084.7514.6773.00
ATRC w/ ViT-L [2]871G340M77.1166.8481.2014.2372.10
InvPT [7]669G423M79.0367.6184.8114.1573.00
TaskExpert (ours)622G420M80.6469.4284.8713.5673.30
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.606 + ], + "angle": 0, + "content": "subset of experts contribute to the output [41], as explained in Section 3.3. In this experiment, we study the performances of using sparse connection (top 1 and top 2), and compare them with the default dense connection manner on PASCAL-Context. The results are demonstrated in Figure 8. It can be observed that activating more experts leads to better performance, and the dense connection achieves the best results. The reason is that it is able to freely utilize the full representative ability of all expert networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.892, + 0.743 + ], + "angle": 0, + "content": "Light-Weight Versions of TaskExpert To study the model performance and efficiency with a smaller parameter and memory budget, we conduct 2 sets of experiments in Table 3: (i) Using different numbers of MMoE modules on ViT-B backbone for PASCAL-Context. The MMoE modules are embedded evenly along the backbone's depth. (ii) Implementing TaskExpert on ViT-S, which is a smaller backbone. These smaller variants achieve reasonable performance with less parameter and memory consumption." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.756, + 0.794, + 0.772 + ], + "angle": 0, + "content": "4.3. Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We implement TaskExpert based on ViT-Large backbone with MMoE modules embedded at every two transformer layers, then we compare TaskExpert to recent SOTA methods in literature on both PASCAL-Context and NYUDv2 in Table 4 and Table 5. We re-implement several CNN-based SOTA multi-task learning methods on ViT-L backbone for a more comprehensive comparison on PASCAL-Context. Compared with the previous SOTA" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21834" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.065, + 0.89, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.199, + 0.895, + 0.23 + ], + "angle": 0, + "content": "Figure 8: Comparison of TaskExpert with dense and sparse connection in MMoE. \"Top 1\" denotes only activating the expert with the highest gating score, \"Top 2\" means activating the experts with top 2 gating scores, and \"Dense\" activates all experts." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.242, + 0.47, + 0.286 + ], + "angle": 0, + "content": "Table 5: Comparison with the state-of-the-art methods on NYUD-v2 dataset. Our TaskExpert achieves the best performances on all tasks." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.287, + 0.469, + 0.414 + ], + "angle": 0, + "content": "
ModelBackboneSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑
Cross-Stitch [25]HRNet1836.340.629020.8876.38
PAP [29]HRNet1836.720.617820.8276.42
PSD [31]HRNet1836.690.624620.8776.42
PAD-Net [8]HRNet1836.610.627020.8576.38
MTI-Net [1]HRNet4845.970.536520.2777.86
ATRC [2]HRNet4846.330.536320.1877.94
InvPT [7]ViT-L53.560.518319.0478.10
TaskExpert (ours)ViT-L55.350.515718.5478.40
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.47, + 0.619 + ], + "angle": 0, + "content": "model (i.e. InvPT [7]), TaskExpert achieves clearly better performances on all 9 metrics of these two challenging benchmarks with the same backbone. Particularly, on PASCAL-Context, we improve semantic segmentation by +1.61 and human parsing by +1.81. On NYUD-v2, semantic segmentation is improved by +1.79. It should be noted that, our gains are measured upon the very strong performance foundation established by InvPT (e.g. Semseg 79.03, Saliency 84.81 on PASCAL-Context). This superior performance clearly shows the effectiveness of the proposed TaskExpert. TaskExpert also maintains a higher efficiency with a reduction of 3 million in the number of parameters and 47 billion FLOPs compared to InvPT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.471, + 0.833 + ], + "angle": 0, + "content": "As underscored in the Related Work section, our TaskExpert is designed for a different setting from \\(\\mathrm{M}^3\\mathrm{ViT}\\) [45] and Mod-Squad [46]. Their models are constrained to generating a single task prediction per forward pass, while our model capably generates predictions for all tasks within a single forward pass. Therefore, the multi-task learning efficiency of our model is significantly higher than their methods. When utilizing the same backbone, it requires only half the FLOPS, as demonstrated in Table 3. Moreover, our model's training duration is markedly shorter since it learns all tasks concurrently in each iteration. While they necessitate around 18 hours of training with 4 A6000 GPUs as confirmed by their authors, our model completes training in under 5 hours using a single NVIDIA 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.246, + 0.863 + ], + "angle": 0, + "content": "4.4. Qualitative study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Qualitative Study of MMoE To understand better the learned gating scores and representative task-generic fea" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.365 + ], + "angle": 0, + "content": "tures by the gating networks and expert networks in our MMoE modules, we present qualitative examples in Figure 6. As can be observed, TaskExpert effectively learns representative task-generic features with distinct activation distributions, focusing on different aspects of the image. Accordingly, the gating scores exhibit different response intensities to different areas for selectively assembling the experts' features for decoding task-specific features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.501 + ], + "angle": 0, + "content": "PCA Visualization of Representative Features of Task-Experts at Different Layers We conduct principal component analysis (PCA) on the representative features at layers 1, 6, and 12. As shown in Fig. 4, we visualize the first three components of the feature PCA as a color map. It reveals that the task experts at deeper layers capture more high-level semantic information of the input. On the other hand, in the shallower layers, the task-feature experts focus more on low-level visual patterns (e.g., edges)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Qualitative Comparison of Decoded Task Features In Fig. 5, we also conduct PCA on the final task features generated by InvPT [7] and our TaskExpert, and draw the first three components. The features for different tasks from TaskExpert show clear task-specific characteristics and are much more discriminative than those of InvPT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Qualitative Comparison of Prediction Results To intuitively examine the prediction quality of TaskExpert compared to the previous best-performing method [7], we visualize the predictions of the two models on the test split of PASCAL-Context in Figure 7. Our TaskExpert generates visually better multi-task predictions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.619, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present TaskExpert, a novel multitask mixture-of-expert model establishing new SOTA performances on two challenging benchmarks. TaskExpert can effectively perform dynamic decoding of task-specific features, from a set of representative task-generic features learned by expert networks. TaskExpert further models the long-term task-specific representations at different layers via a designed multi-task feature memory. Extensive experiments clearly prove the effectiveness and efficiency of our TaskExpert model, and we hope that our TaskExpert can also be beneficial for the future study of dynamic representation learning problems." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21835" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.091, + 0.238, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.116, + 0.47, + 0.178 + ], + "angle": 0, + "content": "This research is supported in part by HKUST-SAIL joint research funding, the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, and HKUST Startup Fund No. R9253." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.193, + 0.174, + 0.209 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[1] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.266, + 0.472, + 0.323 + ], + "angle": 0, + "content": "[2] David Bruggemann, Menelaos Kanakis, Anton Obukhov, Stamatos Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.328, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[3] Amir R Zamir, Alexander Sax, Nikhil Cheerla, Rohan Suri, Zhangjie Cao, Jitendra Malik, and Leonidas J Guibas. Robust learning through cross-task consistency. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.377, + 0.47, + 0.42 + ], + "angle": 0, + "content": "[4] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.425, + 0.47, + 0.467 + ], + "angle": 0, + "content": "[5] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.473, + 0.47, + 0.516 + ], + "angle": 0, + "content": "[6] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.522, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[7] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.556, + 0.47, + 0.611 + ], + "angle": 0, + "content": "[8] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.618, + 0.47, + 0.66 + ], + "angle": 0, + "content": "[9] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.666, + 0.468, + 0.694 + ], + "angle": 0, + "content": "[10] Yu Zhang and Qiang Yang. A survey on multi-task learning. TKDE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.701, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[11] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. TPAMI, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.763, + 0.47, + 0.805 + ], + "angle": 0, + "content": "[12] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.811, + 0.47, + 0.853 + ], + "angle": 0, + "content": "[13] Menelaos Kanakis, Thomas E Huang, David Brüggemann, Fisher Yu, and Luc Van Gool. Composite learning for robust and effective dense predictions. In WACV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In ECCV, 2022." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.218, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian Koring, Suman Saha, and Luc Van Gool. Three ways to improve semantic segmentation with self-supervised depth estimation. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.153, + 0.892, + 0.181 + ], + "angle": 0, + "content": "[16] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.186, + 0.892, + 0.241 + ], + "angle": 0, + "content": "[17] Xiangtai Li, Henghui Ding, Wenwei Zhang, Haobo Yuan, Jiangmiao Pang, Guangliang Cheng, Kai Chen, Ziwei Liu, and Chen Change Loy. Transformer-based visual segmentation: A survey. arXiv preprint arXiv:2304.09854, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[18] Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. Conflict-averse gradient descent for multi-task learning. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[19] Siwei Yang, Hanrong Ye, and Dan Xu. Contrastive multitask dense prediction. In AAAI, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.325, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[20] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Learning multiple dense prediction tasks from partially annotated data. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.371, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[21] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.432, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[22] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In ICML, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.492, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[23] Zirui Wang, Yulia Tsvetkov, Orhan First, and Yuan Cao. Gradient vaccine: Investigating and improving multi-task optimization in massively multilingual models. In ICLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[24] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.599, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[25] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[26] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[27] Lijun Zhang, Xiao Liu, and Hui Guan. Automtl: A programming framework for automating efficient multi-task learning. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.753, + 0.892, + 0.807 + ], + "angle": 0, + "content": "[28] Yuan Gao, Haoping Bai, Zequn Jie, Jiayi Ma, Kui Jia, and Wei Liu. Mtl-nas: Task-agnostic neural architecture search towards general-purpose multi-task learning. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.813, + 0.892, + 0.854 + ], + "angle": 0, + "content": "[29] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Xiaoya Zhang, Ling Zhou, Yong Li, Zhen Cui, Jin Xie, and Jian Yang. Transfer vision patterns for multi-task pixel learning. In ACMMM, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21836" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[31] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.137, + 0.468, + 0.178 + ], + "angle": 0, + "content": "[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.182, + 0.468, + 0.223 + ], + "angle": 0, + "content": "[33] Simon Vandenhende, Stamatios Georgoulis, Bert De Brabandere, and Luc Van Gool. Branched multi-task networks: deciding what layers to share. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.226, + 0.468, + 0.281 + ], + "angle": 0, + "content": "[34] Yongxi Lu, Abhishek Kumar, Shuangfei Zhai, Yu Cheng, Tara Javidi, and Rogerio Feris. Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.285, + 0.468, + 0.313 + ], + "angle": 0, + "content": "[35] Pengsheng Guo, Chen-Yu Lee, and Daniel Ulbricht. Learning to branch for multi-task learning. In ICML, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.316, + 0.468, + 0.356 + ], + "angle": 0, + "content": "[36] Hanrong Ye and Dan Xu. Taskprompt: Spatial-channel multi-task prompting for dense scene understanding. In ICLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.36, + 0.468, + 0.401 + ], + "angle": 0, + "content": "[37] Michael I Jordan and Robert A Jacobs. Hierarchical mixtures of experts and the em algorithm. Neural Computation, 6(2):181-214, 1994." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.468, + 0.459 + ], + "angle": 0, + "content": "[38] Robert A Jacobs and Michael I Jordan. Learning piecewise control strategies in a modular neural network architecture. IEEE Transactions on Systems, Man, and Cybernetics, 23(2):337-345, 1993." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.464, + 0.468, + 0.505 + ], + "angle": 0, + "content": "[39] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. *Neural Computation*, 3(1):79-87, 1991." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.509, + 0.468, + 0.55 + ], + "angle": 0, + "content": "[40] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.553, + 0.468, + 0.608 + ], + "angle": 0, + "content": "[41] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.612, + 0.468, + 0.653 + ], + "angle": 0, + "content": "[42] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. JMLR, 23:1-39, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.656, + 0.468, + 0.724 + ], + "angle": 0, + "content": "[43] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.728, + 0.468, + 0.77 + ], + "angle": 0, + "content": "[44] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In KDD, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.774, + 0.468, + 0.841 + ], + "angle": 0, + "content": "[45] Hanxue Liang, Zhiwen Fan, Rishov Sarkar, Ziyu Jiang, Tianlong Chen, Kai Zou, Yu Cheng, Cong Hao, and Zhangyang Wang. M3vit: Mixture-of-experts vision transformer for efficient multi-task learning with model-accelerator co-design. In NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[46] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik Learned-Miller, and Chuang Gan. Mod-squad: Designing mixture of experts as modular multitask learners. arXiv preprint arXiv:2212.08066, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[47] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.236 + ], + "angle": 0, + "content": "[48] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In CVPR, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.241, + 0.892, + 0.283 + ], + "angle": 0, + "content": "[49] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[50] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[51] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21837" + } + ] +] \ No newline at end of file diff --git a/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/2e3b51ba-4260-4b59-a865-225826178569_origin.pdf b/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/2e3b51ba-4260-4b59-a865-225826178569_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bd21593a828c17ffe59aaf15b5b53a623271e181 --- /dev/null +++ b/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/2e3b51ba-4260-4b59-a865-225826178569_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0e1c858b0b7ebf8ecc52202c0c4386883d5ddb8ba253a8ca269de8c12c97afe +size 2535204 diff --git a/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/full.md b/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..19ee58df5eae98b11a1fbdbd4d12bdf9fa29104e --- /dev/null +++ b/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/full.md @@ -0,0 +1,266 @@ +# TaskExpert: Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts + +Hanrong Ye and Dan Xu\* +Department of Computer Science and Engineering, HKUST +Clear Water Bay, Kowloon, Hong Kong + +{hyeae, danxu}@cse.ust.hk + +# Abstract + +Learning discriminative task-specific features simultaneously for multiple distinct tasks is a fundamental problem in multi-task learning. Recent state-of-the-art models consider directly decoding task-specific features from one shared task-generic feature (e.g., feature from a backbone layer), and utilize carefully designed decoders to produce multi-task features. However, as the input feature is fully shared and each task decoder also shares decoding parameters for different input samples, it leads to a static feature decoding process, producing less discriminative task-specific representations. To tackle this limitation, we propose TaskExpert, a novel multi-task mixture-of-experts model that enables learning multiple representative task-generic feature spaces and decoding task-specific features in a dynamic manner. Specifically, TaskExpert introduces a set of expert networks to decompose the backbone feature into several representative task-generic features. Then, the task-specific features are decoded by using dynamic task-specific gating networks operating on the decomposed task-generic features. Furthermore, to establish long-range modeling of the task-specific representations from different layers of TaskExpert, we design a multitask feature memory that updates at each layer and acts as an additional feature expert for dynamic task-specific feature decoding. Extensive experiments demonstrate that our TaskExpert clearly outperforms previous best-performing methods on all 9 metrics of two competitive multi-task learning benchmarks for visual scene understanding (i.e., PASCAL-Context and NYUD-v2). Codes and models will be made publicly available. + +# 1. Introduction + +With the rapid development of deep learning models, many computer vision tasks can be effectively handled by deep networks, which strongly motivates researchers to de + +![](images/e2bee81ef2043ed6a48b3853a6685b448704df26edbcc8b6f19b9725b6e37485.jpg) +Figure 1: Motivative illustration of TaskExpert for dynamical decoding discriminative task-specific representations. A set of $K$ expert networks decompose the task-generic backbone feature into $K$ representative feature spaces. Then, the feature activations of an input from the multiple expert networks are dynamically assembled by the Task Gating network, to produce task-specific features for the $T$ tasks. TaskExpert further performs long-range modeling of task-specific representations by designing a multi-task feature memory. It is updated at the current layer based on the assembled task-specific representations, and is used as an additional feature expert in the next layer. + +velop unified multi-task networks to perform joint learning and inference on multiple distinct tasks [1, 2, 3, 4]. Multi-Task Learning (MTL) can avoid repetitive training of single-task models and is able to simultaneously generate predictions for different tasks with only one-time inference. On the other hand, as different computer vision tasks usually share a basic common understanding of input images, learning multiple tasks concurrently can help improve the representation power of each task and boost the overall performance of multi-task models [5, 6]. + +Learning discriminative task-specific features is a funda- + +mental problem for MTL models. To achieve this objective, recent state-of-the-art works [7, 2, 8, 1] consider a decoder-focused paradigm, in which a shared backbone model with pretrained parameters is applied to learn generic image representations, and then separate task decoders are carefully designed to generate task-specific features. Although this paradigm is straightforward to achieve task-specific feature decoding, and produces promising multi-task prediction results, these models have two aspects of potential limitations: (i) The input task-generic feature from a backbone layer is fully shared for the different task decoders, which requires a careful design of the decoders (e.g. capacity and structure) for the different tasks to achieve task-discriminative decoding. (ii) The parameters of each task decoder are typically shared for different input samples, resulting in a static decoding of task-specific features, while a sample-dependant decoder is highly beneficial for learning more discriminative task features, as the input samples can be diverse, and sample-related context information is particularly important for multi-task learning and inference. All these aspects hamper the previous models from learning effective task-specific features and predictions. + +To tackle the above-mentioned issues, this paper proposes a novel multi-task mixture-of-experts framework, coined as "TaskExpert", which dynamically performs the task-feature decoding for different inputs and tasks as depicted in Figure 1. Instead of using one shared backbone feature to decode different task features, the proposed TaskExpert introduces a set of task-generic expert networks that learn to decompose the backbone feature into a set of representative task-generic features. The feature decomposition enables the decoders to interact with finer-granularity task-generic feature spaces to produce more discriminative task features, as each expert can be responsible for modeling one representative feature space from the training data. Then, we dynamically assemble feature activations from different experts to produce task-specific representations, based on sample-dependent and task-specific gating scores predicted from a designed gating network. On the other hand, visual understanding tasks greatly benefit from representations of different network levels [9]. Thus, we further devise a "multi-task feature memory" to aggregate the dynamically decoded task-specific features produced at different network layers, and propose a "Memorial Mixture-of-Experts (MMoE)" module. For each task, MMoE utilizes its corresponding task feature from the multi-task feature memory as an additional feature expert, to improve task-specific feature decoding. In this way, MMoE enables long-range dynamic modeling of task-specific representation for each task throughout the entire backbone. + +In summary, the main contribution is three-fold: + +- We propose a multi-task mixture-of-experts model that allows effective decomposition of a task-generic fea + +ture from any backbone network layer and enables dynamically assembling discriminative multi-task representations by a task gating mechanism. + +- We design a novel Memorial Mixture-of-Experts (MMoE) module by further introducing a multi-task feature memory, which can effectively interact (read and write) with the mixture of experts at different network layers to achieve long-range dynamic decoding of task-specific features by utilizing different levels of network representations. + +- The proposed TaskExpert is extensively validated on two challenging multi-task visual scene understanding benchmarks (i.e. Pascal-Context and NYUD-v2). The results clearly show the effectiveness of TaskExpert for dynamic multi-task representation learning, and also establish new state-of-the-art performances on all 9 metrics on the benchmarks, regarding the comparison with previous best-performing methods using both transformer- and CNN-based backbones. + +# 2. Related Works + +Multi-Task Learning in Computer Vision In the field of multi-task learning (MTL) for computer vision problems [10, 11, 12, 13, 14, 15, 16, 17], some research works explore effective solutions through a perspective of model optimization and improve the optimization process of MTL via different multi-task loss designs [3, 6, 18, 19, 20] and gradient manipulations [21, 22, 23, 24]. These optimization strategies have also been demonstrated beneficial for alleviating the issue of task competition in multi-task learning [6]. From another perspective, several existing works also investigate designing better multi-task model architectures. These related works can be roughly divided into two groups: encoder-focused and decoder-focused methods. The encoder-focused methods [25, 26, 27, 28] target designing multi-task encoders, while the decoder-focused methods typically share the encoder (i.e. the backbone) for different tasks and design sophisticated decoders for task-specific feature learning and prediction [1, 2, 29, 8, 30, 31]. A recent state-of-the-art decoder-focused method (i.e. In-vPT [7]) performs multi-task learning on powerful transformer architectures [32], and establishes a very strong baseline for multi-task visual scene understanding. It can effectively learn spatial and cross-task relationships globally. To adapt multi-task models to different tasks, researchers also propose to search for task-specific network structures for encoders [33, 34, 35] and decoders [2], or utilize task-specific learnable prompts to effectively capture task-specific information [36]. Despite promising performances of these works, they are still limited by static decoder network designs since their decoders cannot dynamically generate task-specific representations based on dif + +ferent inputs and tasks. In this paper, we propose TaskExpert, a novel and effective mixture-of-experts framework, which can achieve dynamic multi-task representation learning via feature decomposition and assembling through the entire backbone network. + +Mixture-of-Experts Models With statistical motivation [37], Mixture-of-Experts (MoE) models are originally designed to control the dynamics of neural networks automatically [38, 39]. MoE learns a series of expert networks and a gating network. The outputs of expert networks are weighted by gating scores (or called "gate values") generated by a gating network before the weighting operation. In more recent works, some researchers use the gating scores as a criterion to sparsely select only one or a few experts. The sparse activation of experts enables a significant reduction of the computational cost when training large-scale models [40, 41, 42, 43]. Related to our work, [44] designs a multi-gate MoE to ensemble expert networks for different census analysis tasks with different gating networks, and [45, 46] propose sparse task-conditioned MoE networks for efficient multi-task learning, which generate predictions for only one task within a single forward network inference. In contrast to these works, our Memorial Mixture-of-Experts (MMoE) is a plug-and-play decoder module, which is designed for dynamically assembling discriminative task-specific features. Our TaskExpert can simultaneously generate multi-task predictions for all tasks in one single forward inference, leading to significantly higher multi-task training efficiency and enabling cross-task interaction learning. Moreover, MMoE models task-specific features throughout different network layers via the novel design of multi-task feature memory and help generate more discriminative task-specific features. + +# 3. TaskExpert + +# 3.1. Framework Overview + +An overall illustration of our TaskExpert framework is shown in Figure 2. The input image is first projected to a token sequence and then fed into a series of classic visual transformer layers. Our framework can also be flexibly used with a CNN backbone. The proposed Memorial Mixture-of-Experts (MMoE) module is embedded at different network layers, and each module takes the task-generic backbone feature of that layer as input. It dynamically assembles and outputs discriminative task-specific representations at the corresponding layer. The MMoE further interacts (reads and writes operations) with the designed multi-task feature memory, to perform a long-range dynamic multi-task feature decoding, which benefits from different levels of network features. The reads utilize the feature memory as an additional feature expert in the MMoE module, and the writes update the memory at the layer. Finally, after be + +![](images/967f229000fb7786a71aa590a3c583170e9124f754048af17bf90b0a10cf03dc.jpg) +Figure 2: Overall architecture of the proposed TaskExpert model. Memorial Mixture-of-Experts (MMoE) modules are embedded at different backbone layers to dynamically decode task-specific feature for each task at the corresponding layer. The MMoE also interacts (reads and writes) with a designed multi-task feature memory to have a long-range dynamic multi-task feature decoding benefiting from different levels of network features. The reads utilize the memory as an additional feature expert in MMoE, and the writes update the memory at the layer. After being updated by the last network layer, the multi-task feature memory is used for the final prediction of each task. + +ing updated by the outputs of all MMoE modules at different network layers, the task-specific features in multi-task feature memory are then used for generating the multi-task predictions with simple prediction heads. We elaborate on more details of our model in the following several sections. + +# 3.2. Multi-Task Backbone Structure + +We consider a widely-used vision transformer (ViT) model [47] with $L$ layers as the backbone of our multi-task MoE model. The input image is first divided into $H\times W$ patches and processed by the patch embedding and positional embedding modules. Then, we obtain a sequence of patch tokens $\mathbf{X}^0 = \{\mathbf{x}_1^0,\mathbf{x}_2^0,\dots,\mathbf{x}_N^0\}$ , where $N = H\times W$ and the number of channels is $C$ . The patch tokens contain generic visual representations of the input image. We denote the $l$ -th transformer layer as a mapping function $f^{l}$ and the updating of image patch tokens at different transformer layers can be formulated as $\mathbf{X}^l = f^l (\mathbf{X}^{l - 1})$ , where $\mathbf{X}^l\in \mathbb{R}^{N\times C}$ denotes the output patch tokens of the transformer at the $l$ -th layer. + +![](images/aefd51ed303b4e3c26e501c19ae2b517f2b5501a124d6fd1bee9dc1b7c125d8d.jpg) +Figure 3: Illustration of the proposed Memorial Mixture-of-Experts (MMoE) module with the multi-task feature memory. MMoE accepts a task-generic feature from a backbone layer as input, and dynamically generates task-specific features for different tasks in a sample- and task-dependent manner. Specifically, MMoE decomposes the backbone feature to a set of $K$ representative task-generic features $(\mathbf{R}_1^l, \dots, \mathbf{R}_K^l)$ via the $K$ expert networks. Then these features are used to assemble a set of $T$ task-specific features based on dynamically generated task-gating scores from the task gating networks. To establish long-range modeling of task-specific representations crossing different network layers, the multi-task feature memory is designed to interact with the expert networks. The multi-task feature memory serves as an additional feature expert to decode the task-specific representations, and the decoded representations are also further used to update the multi-task feature memory. + +# 3.3. Memorial Mixture-of-Experts (MMoE) + +We present the details of several core parts of the proposed memorial mixture-of-experts (MMoE) module, including backbone generic-feature decomposition with a set of network experts, context-aware task-specific gating, the multi-task feature memory to have long-range modeling on task-specific features through interactions (reads and writes) with the mixture of experts, dynamic task-specific feature assembling, and multi-task prediction heads. + +# 3.3.1 Backbone Feature Decomposition with Experts + +As shown in Figure 3, to enable learning discriminative task-specific representations from the task-generic backbone feature, we first define a set of expert networks that decompose the backbone feature into several representative task-generic features in sub-feature spaces, representing distinct patterns of the backbone features, as can be observed in Figure 6. It allows more fine-grained decoding of task-specific features. Specifically, at the $l$ -th backbone layer, let us assume that we use $K$ expert networks $\{f_{e}^{l,i}(\cdot):\mathbb{R}^{N\times C}\to \mathbb{R}^{N\times C},i\in [1,K]\}$ to learn distinct task-generic feature projections, where superscript $l$ denotes the layer index and $i$ is the expert index. Each expert network employs the same Linear-BatchNorm-RELU-Linear structure but with different parameters. Through each expert network, we obtain a set of representative fea + +tures $\{\mathbf{R}_i^l\in \mathbb{R}^{N\times C},i\in [1,K]\}$ by: + +$$ +\mathbf {R} _ {i} ^ {l} = f _ {e} ^ {l, i} \left(\mathbf {X} ^ {l}\right), \tag {1} +$$ + +where $\mathbf{X}^l$ is the $l$ -th layer feature of the backbone. + +# 3.3.2 Context-Aware Task-Specific Gating + +After we have obtained a set of representative task-generic features decomposed from a backbone feature for all the tasks, we aim at decoding discriminative task-specific features for different tasks by automatically assembling these representative features. Thus, we propose to adaptively control the contribution of each expert for each patch token by designing a task-specific gating network. + +For each token in the representative task-generic features, we compute a task-gating score to control its contribution to generating the task-specific features. The task-gating scores of each task are dynamically generated by a task-specific gating network, using the features $\mathbf{X}^l$ from the $l$ -th network layer. We term the task-specific gating network with a short name "task gating", and denote it as $f_{g}^{l,t}(\cdot):\mathbb{R}^{N\times C\to N\times K}$ , where $t\in [1,T]$ ( $T$ is the total number of tasks) is a task index. The task-gating scores $\mathbf{G}^{l,t}\in \mathbb{R}^{N\times K}$ is computed as $\mathbf{G}^{l,t} = f_{g}^{l,t}(\mathbf{X}^{l})$ . + +For the design of the gating networks, previous Mixture-of-Experts methods [42, 43, 41] use a simple linear layer followed by a Softmax function, which does not take into + +![](images/75375102ceba2b139b1590812e19a8b5f0682786b66443627bf8d7174ea4fcd8.jpg) +Figure 4: PCA visualization of the representative features by task-generic experts across different layers. In the deeper layers, the features become increasingly responsive to high-level semantic areas. Conversely, features at more shallow layers show a heightened sensitivity to low-level patterns. + +account the context of each image patch token, while using only the corresponding patch token when computing the gating score. As various challenging computer vision tasks (e.g. semantic segmentation and monocular depth estimation) require modeling and understanding each image patch with context information, we thus design a context-aware gating strategy. Specifically, context-aware gating adopts convolution kernels to extend the receptive field of the gating networks, and thus it can involve a local context of tokens in the computation of gating scores. We first reshape the input token sequence $\mathbf{X}$ (in $\mathbb{R}^{N\times C}$ ) as spatial maps (in $\mathbb{R}^{H\times W\times C}$ ), and then compute the gating scores of $K$ experts with three layers, each having a $3\times 3$ convolution with ReLU as intermediate activation function. We denote this computation procedure as "CONV", which will also be used later in Equation 3. + +# 3.3.3 Multi-Task Feature Memory + +To conduct long-range modeling of the multi-task representations decoded at different network layers that contain different levels of visual information, we design a multitask feature memory to interact with the MMoE modules through read and write operations. The read operation uses the task feature memory as an additional feature expert for task-specific feature decoding in MMoE. The write operation updates the multi-task feature memory by using the decoded task-specific features from MMoE. In this way, the multi-task feature memory can aggregate the decoded task-specific features crossing different network layers. Specifically, the read operation is an identity mapping. The write operation can be performed as follows: given a decoded task-specific feature of the $t$ -th task at the first layer, i.e. $\mathbf{F}^{1,t}$ , the initialization of the multi-task feature memory $\mathbf{M}^{1,t}$ is obtained by $\mathbf{M}^{1,t} \gets \mathbf{F}^{1,t}$ . In the subsequent layers, suppose we decode task-specific features at every network layer, at the $l$ -th layer ( $l > 1$ ), we write the multi-task feature memory $\{\mathbf{M}^{l,t}, t \in [1,T]\}$ with the decoded task-specific feature $\mathbf{F}^{l,t}$ weighted by a task-specific learnable momentum value $\alpha^{l,t}$ as shown in Figure 3: + +$$ +\mathbf {M} ^ {l, t} \leftarrow \mathbf {F} ^ {l, t} + \alpha^ {l, t} \mathbf {M} ^ {l - 1, t}, l > 1. \tag {2} +$$ + +# 3.3.4 Dynamically Assembling Task-Specific Features + +With the obtained representative task-generic features and task-specific gating scores as above-described, we can perform dynamic decoding of task-specific features, as the gating parameters are predicted in a sample-dependent manner. This is significantly different from existing decoder-focused multi-task structures, which share the decoding parameters for all the different input samples. Specifically, for the $t$ -th task, we multiply every token of the representative task-generic features $\{\mathbf{R}_i^l,i\in [1,K]\}$ with their corresponding task-gating score $\{\mathbf{G}_i^{l,t},i\in [1,K]\}$ and sum the tokens in the same spatial position to decode task-specific feature. Furthermore, to learn more discriminative features for each task, we utilize the corresponding task-specific feature in multi-task feature memory as an additional feature expert in MMoE, which contains long-range task-specific information learned throughout the entire network. It helps decode more effective and long-term task-specific features. As shown in Figure 3, for the MMoE at the $l$ -th transformer layer ( $l > 1$ ), we compute an additional gating score map $\mathbf{G}_{K + 1}^{l,t}$ for the task-specific feature from the multi-task feature memory $\mathbf{M}^{l - 1,t}$ , which is performed by adding one more output channel into the task gating network $f_{g}^{l,t}:\mathbb{R}^{N\times C}\to \mathbb{R}^{N\times (K + 1)}$ . In summary, the task gating scores of the $t$ -th task and the $l$ -th layer, i.e., $\mathbf{G}^{l,t} = \{\mathbf{G}_1^{l,t},\mathbf{G}_2^{l,t},\dots,\mathbf{G}_{K + 1}^{l,t}\}$ , is computed as: + +$$ +\mathbf {G} ^ {l, t} = \left\{ \begin{array}{l l} \operatorname {S o f t m a x} \left(\operatorname {C O N V} ^ {K + 1} \left(\mathbf {X} ^ {l}\right)\right) & l > 1, \\ \operatorname {S o f t m a x} \left(\operatorname {C O N V} ^ {K} \left(\mathbf {X} ^ {l}\right)\right) & l = 1, \end{array} \right. \tag {3} +$$ + +where the superscript $K$ of CONV denotes the number of output convolution kernels. And accordingly, the decoded task-specific features at the $l$ -th MMoE module is: + +$$ +\mathbf {F} ^ {l, t} = \left\{ \begin{array}{l l} \sum_ {i = 1} ^ {K} \mathbf {G} _ {i} ^ {l, t} \times \mathbf {R} _ {i} ^ {l} + \mathbf {G} _ {K + 1} ^ {l, t} \times \mathbf {M} ^ {l - 1, t} & l > 1, \\ \sum_ {i = 1} ^ {K} \mathbf {G} _ {i} ^ {l, t} \times \mathbf {R} _ {i} ^ {l} & l = 1. \end{array} \right. \tag {4} +$$ + +Alternative sparse activation of experts As an alternative, we can also use "sparse connections" in MMoE, which means that we only activate $K_{sel}$ selected experts to compute the task-specific features ( $K_{sel} < K$ ) for each token [41]. In this case, for each token, we find the top $K_{sel}$ + +![](images/c6c7890ca9c9ec06deb4dcfa7ad740e344b73869e4cb9139fc056a0700537c8e.jpg) +Figure 5: PCA visualization of the decoded final task features generated by InvPT [7] and our TaskExpert. Our model generates more discriminative features for all tasks. + +experts to be activated based on the task-gating scores, and sum their outputs after being weighted by the task-gating scores. A comparison between the dense connection and sparse connection in MMoE is discussed in the experiments. + +# 3.4. Generation of Multi-Task Predictions + +At the final network layer $L$ , we use task-specific prediction heads to directly generate prediction outputs for each task from the multi-task feature memory $\{\mathbf{M}^{L,t}, t \in [1, T]\}$ . The prediction heads for different tasks have a similar structure, consisting of a $3 \times 3$ convolution layer with batch normalization and ReLU activation, followed by a linear layer for the final prediction. + +# 4. Experiment + +In this section, we validate the effectiveness of our proposal through several aspects: (i) ablation study, (ii) comparison with SOTA methods, and (iii) qualitative analysis. + +# 4.1. Experimental Setup + +Tasks and Benchmarks For evaluation, we adopt two popular benchmarks in multi-task learning as previous works [7, 1], i.e. PASCAL-Context [48, 49] and NYUD-v2 [50] datasets. PASCAL-Context contains 4,998 samples in train split and 5,105 samples in test split. It provides labels for multiple important visual understanding tasks, including semantic segmentation, human parsing, and object boundary detection. Previous work on multi-task learning generates pseudo labels for another two tasks [5]: surface normals estimation and saliency detection. Meanwhile, NYUD-v2 contains 1,449 samples in train split and 795 samples in test split. It collects labels for semantic segmentation and monocular depth estimation, from which the labels of surface normal estimation and object boundary detection can be generated. We utilize all the tasks provided in these datasets for evaluation. + +Metrics We use the same evaluation metrics as previous works [7]. Specifically, semantic segmentation (Semseg) and human parsing (Parsing) tasks are evaluated by the mean Intersection over Union (mIoU). Monocular depth estimation task (Depth) is evaluated by Root Mean + +Table 1: Ablation study on PASCAL-Context with a ViT-Base backbone. "MoE" denotes using Mixture-of-Experts module; "CG" denotes using context-aware gating; "Mem" denotes using our multi-task feature memory. $\downarrow$ ’ indicates lower better and ‘↑’ means higher better. + +
ModelSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑MTL Perf Δm ↑
Baseline76.4065.7084.6313.8369.50-3.70
w/ MoE77.6266.3784.6513.7770.40-2.88
w/ MoE+CG77.9966.8284.7613.6471.70-2.07
w/ MoE+CG+Mem78.4567.3884.9613.5572.30-1.46
+ +Square Error (RMSE). Surface normal estimation task (Normal) is evaluated by the mean error (mErr) of angles. Saliency detection task (Saliency) is evaluated by maximal F-measure (maxF). Object boundary detection task (Boundary) is evaluated by the optimal-dataset-scale F-measure (odsF). We calculate the mean relative difference across all tasks compared to the single-task baseline, which is noted as Multi-task Performance (MTL Perf $\Delta_{m}$ ) [5]. + +Implementation Details We use 5 expert networks in MMoE for PASCAL-Context and 4 expert networks for NYUD-v2, which are the same as the corresponding number of tasks. We add MMoE at the end of each transformer layer if not otherwise stated. For model training, we follow the setting of InvPT [7] and use Adam optimizer to train the models for 40,000 iterations with a batch size of 4. Polynomial learning rate scheduler is used with a learning rate of $2 \times 10^{-5}$ and a weight decay rate of $1 \times 10^{-6}$ . We adopt the loss functions and weights for different tasks in [7]. The ViT backbones are pre-trained on ImageNet-22K [51]. + +# 4.2. Ablation Study + +Model Variants Definition To examine the effectiveness of TaskExpert under a fair setting, we compare it to the multi-task baseline model and its variants with the same backbone. (1) The baseline model ("Baseline") uses ViT-Base as backbone, and designs task-specific decoders in each transformer layer to generate layer-wise task-specific features. The task-specific features from all layers are added together as multi-scale input of the corresponding prediction head for each task. These layer-wise decoders share the same structure as the expert networks used in MMoE. The prediction heads for all the tasks have a similar structure: $\mathrm{Conv}(3\times 3)$ -BN-ReLU-Linear, which are the same as those used in TaskExpert. (2) "TaskExpert w/ MoE" adds the expert and gating networks of MMoE without context-aware gating to Baseline. It uses gating networks to control the contribution of each expert to each task in a sample-dependent manner without context-aware gating or multi-task feature memory. The number of experts is the same as the number of tasks. (3) "TaskExpert w/ MoE+CG" further uses our context-aware gating strategy, which improves the gating networks by introducing the contextual information of each token. (4) "TaskExpert w/ MoE+CG+Mem" is the full version of our TaskExpert using multi-task feature + +![](images/d7d2f108e35935847d8bf42d22a1760fda9c68894d59035d49c05ffd47068495.jpg) +Figure 6: Visualization of the task-gating scores by task-specific gating networks and the decomposed representative features from expert networks in MMoE modules. The gating scores for different experts show distinct distributions, while the representative features are discriminative to different visual elements. These results suggest different experts are able to decompose diverse visual information from the backbone feature. + +![](images/a7a7eee4a7e5bb9c92731e62b4c8fb1e027dcbeddc18976ec1f42aef2aa0cba2.jpg) +Figure 7: Qualitative comparison with SOTA method (i.e. InvPT [7]) on PASCAL-Context. Our TaskExpert generates better predictions with more accurate details as highlighted. + +memory in the proposed MMoE module to establish modeling of long-term task-specific representation. + +Key Results We analyze the performances of different model variants introduced above in Table 1. We find that using MoE brings a clear performance gain compared to the baseline model, which demonstrates the effectiveness of designing task-specific gating networks to aggregate the representative task-generic features for dynamically decoding task-specific features. For example, mIoU of semantic segmentation is improved from 76.40 to 77.62 (+1.22), and human parsing is improved from 65.70 to 66.37 (+0.67). Furthermore, the usage of context-aware gating strategy helps MoE achieves better performance by enabling the learning of contextual information for each token when computing the gating scores. For example, odsF of boundary detection is improved from 70.40 to 71.70 (+1.30) compared to only using MoE. Finally, with long-term modeling of task-specific representations via the proposed multi-task feature memory of Memorial Mixture-of-Experts, TaskExpert achieves significant performance improvement compared to the baseline model on different tasks, and the overall multitask performance $\Delta_{m}$ is improved by +2.24. + +Study of Spare and Dense Connections in MoE The original MoE models use the dense connection as the gating mechanism: all experts are activated and contribute to the output [38, 39]. Recent works propose to utilize sparse connection as the gating mechanism: for each token, only a + +Table 2: Variants of TaskExpert on PASCAL-Context. The inference GPU memory is measured with batch size 1. + +
VariantSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑#ParamFLOPSInference GPU Mem
3 MMoE78.0466.5484.8513.6271.20174M597G8367M
6 MMoE78.1866.7984.9313.5971.70234M659G8851M
12 MMoE78.4567.3884.9613.5572.30347M782G12503M
ViT-S75.0462.6884.6814.2268.8055M204G4187M
ViT-B78.4567.3884.9613.5572.30347M782G12503M
+ +Table 3: Comparison with multi-task MoE methods with ViT-S backbone on PASCAL-Context. The training GPU time is shown in the last column. + +
VariantSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑#ParamFLOPSGPU Hours
M3ViT [45]72.8062.1066.3014.5071.7042M420G~72
Mod-Squad [46]74.1062.7066.9013.7072.0050M420G~72
Ours75.0462.6884.6814.2268.8055M204G~5
+ +Table 4: Comparison with the state-of-the-art methods with the same backbone on PASCAL-Context. Our TaskExpert clearly achieves superior performances on all tasks. + +
ModelFLOPs#ParamSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑
PAD-Net [8]124G81M53.6059.6065.8015.3072.50
MTI-Net [1]161G128M61.7060.1884.7814.2370.80
ATRC [2]216G96M67.6762.9382.2914.2472.42
PAD-Net w/ ViT-L [8]773G330M78.0167.1279.2114.3772.60
MTI-Net w/ ViT-L [1]774G851M78.3167.4084.7514.6773.00
ATRC w/ ViT-L [2]871G340M77.1166.8481.2014.2372.10
InvPT [7]669G423M79.0367.6184.8114.1573.00
TaskExpert (ours)622G420M80.6469.4284.8713.5673.30
+ +subset of experts contribute to the output [41], as explained in Section 3.3. In this experiment, we study the performances of using sparse connection (top 1 and top 2), and compare them with the default dense connection manner on PASCAL-Context. The results are demonstrated in Figure 8. It can be observed that activating more experts leads to better performance, and the dense connection achieves the best results. The reason is that it is able to freely utilize the full representative ability of all expert networks. + +Light-Weight Versions of TaskExpert To study the model performance and efficiency with a smaller parameter and memory budget, we conduct 2 sets of experiments in Table 3: (i) Using different numbers of MMoE modules on ViT-B backbone for PASCAL-Context. The MMoE modules are embedded evenly along the backbone's depth. (ii) Implementing TaskExpert on ViT-S, which is a smaller backbone. These smaller variants achieve reasonable performance with less parameter and memory consumption. + +# 4.3. Comparison with SOTA Methods + +We implement TaskExpert based on ViT-Large backbone with MMoE modules embedded at every two transformer layers, then we compare TaskExpert to recent SOTA methods in literature on both PASCAL-Context and NYUDv2 in Table 4 and Table 5. We re-implement several CNN-based SOTA multi-task learning methods on ViT-L backbone for a more comprehensive comparison on PASCAL-Context. Compared with the previous SOTA + +![](images/96a37487975fa33197668afd601913c587312b4572de77d778616d84ef74ef2e.jpg) +Figure 8: Comparison of TaskExpert with dense and sparse connection in MMoE. "Top 1" denotes only activating the expert with the highest gating score, "Top 2" means activating the experts with top 2 gating scores, and "Dense" activates all experts. + +Table 5: Comparison with the state-of-the-art methods on NYUD-v2 dataset. Our TaskExpert achieves the best performances on all tasks. + +
ModelBackboneSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑
Cross-Stitch [25]HRNet1836.340.629020.8876.38
PAP [29]HRNet1836.720.617820.8276.42
PSD [31]HRNet1836.690.624620.8776.42
PAD-Net [8]HRNet1836.610.627020.8576.38
MTI-Net [1]HRNet4845.970.536520.2777.86
ATRC [2]HRNet4846.330.536320.1877.94
InvPT [7]ViT-L53.560.518319.0478.10
TaskExpert (ours)ViT-L55.350.515718.5478.40
+ +model (i.e. InvPT [7]), TaskExpert achieves clearly better performances on all 9 metrics of these two challenging benchmarks with the same backbone. Particularly, on PASCAL-Context, we improve semantic segmentation by +1.61 and human parsing by +1.81. On NYUD-v2, semantic segmentation is improved by +1.79. It should be noted that, our gains are measured upon the very strong performance foundation established by InvPT (e.g. Semseg 79.03, Saliency 84.81 on PASCAL-Context). This superior performance clearly shows the effectiveness of the proposed TaskExpert. TaskExpert also maintains a higher efficiency with a reduction of 3 million in the number of parameters and 47 billion FLOPs compared to InvPT. + +As underscored in the Related Work section, our TaskExpert is designed for a different setting from $\mathrm{M}^3\mathrm{ViT}$ [45] and Mod-Squad [46]. Their models are constrained to generating a single task prediction per forward pass, while our model capably generates predictions for all tasks within a single forward pass. Therefore, the multi-task learning efficiency of our model is significantly higher than their methods. When utilizing the same backbone, it requires only half the FLOPS, as demonstrated in Table 3. Moreover, our model's training duration is markedly shorter since it learns all tasks concurrently in each iteration. While they necessitate around 18 hours of training with 4 A6000 GPUs as confirmed by their authors, our model completes training in under 5 hours using a single NVIDIA 3090 GPU. + +# 4.4. Qualitative study + +Qualitative Study of MMoE To understand better the learned gating scores and representative task-generic fea + +tures by the gating networks and expert networks in our MMoE modules, we present qualitative examples in Figure 6. As can be observed, TaskExpert effectively learns representative task-generic features with distinct activation distributions, focusing on different aspects of the image. Accordingly, the gating scores exhibit different response intensities to different areas for selectively assembling the experts' features for decoding task-specific features. + +PCA Visualization of Representative Features of Task-Experts at Different Layers We conduct principal component analysis (PCA) on the representative features at layers 1, 6, and 12. As shown in Fig. 4, we visualize the first three components of the feature PCA as a color map. It reveals that the task experts at deeper layers capture more high-level semantic information of the input. On the other hand, in the shallower layers, the task-feature experts focus more on low-level visual patterns (e.g., edges). + +Qualitative Comparison of Decoded Task Features In Fig. 5, we also conduct PCA on the final task features generated by InvPT [7] and our TaskExpert, and draw the first three components. The features for different tasks from TaskExpert show clear task-specific characteristics and are much more discriminative than those of InvPT. + +Qualitative Comparison of Prediction Results To intuitively examine the prediction quality of TaskExpert compared to the previous best-performing method [7], we visualize the predictions of the two models on the test split of PASCAL-Context in Figure 7. Our TaskExpert generates visually better multi-task predictions. + +# 5. Conclusion + +In this paper, we present TaskExpert, a novel multitask mixture-of-expert model establishing new SOTA performances on two challenging benchmarks. TaskExpert can effectively perform dynamic decoding of task-specific features, from a set of representative task-generic features learned by expert networks. TaskExpert further models the long-term task-specific representations at different layers via a designed multi-task feature memory. Extensive experiments clearly prove the effectiveness and efficiency of our TaskExpert model, and we hope that our TaskExpert can also be beneficial for the future study of dynamic representation learning problems. + +# Acknowledgement + +This research is supported in part by HKUST-SAIL joint research funding, the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, and HKUST Startup Fund No. R9253. + +# References + +[1] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In ECCV, 2020. +[2] David Bruggemann, Menelaos Kanakis, Anton Obukhov, Stamatos Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In ICCV, 2021. +[3] Amir R Zamir, Alexander Sax, Nikhil Cheerla, Rohan Suri, Zhangjie Cao, Jitendra Malik, and Leonidas J Guibas. Robust learning through cross-task consistency. In CVPR, 2020. +[4] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, 2017. +[5] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In CVPR, 2019. +[6] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, 2018. +[7] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. In ECCV, 2022. +[8] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In CVPR, 2018. +[9] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. +[10] Yu Zhang and Qiang Yang. A survey on multi-task learning. TKDE, 2021. +[11] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. TPAMI, 2021. +[12] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018. +[13] Menelaos Kanakis, Thomas E Huang, David Brüggemann, Fisher Yu, and Luc Van Gool. Composite learning for robust and effective dense predictions. In WACV, 2023. +[14] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In ECCV, 2022. + +[15] Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian Koring, Suman Saha, and Luc Van Gool. Three ways to improve semantic segmentation with self-supervised depth estimation. In CVPR, 2021. +[16] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In CVPR, 2019. +[17] Xiangtai Li, Henghui Ding, Wenwei Zhang, Haobo Yuan, Jiangmiao Pang, Guangliang Cheng, Kai Chen, Ziwei Liu, and Chen Change Loy. Transformer-based visual segmentation: A survey. arXiv preprint arXiv:2304.09854, 2023. +[18] Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. Conflict-averse gradient descent for multi-task learning. In NeurIPS, 2021. +[19] Siwei Yang, Hanrong Ye, and Dan Xu. Contrastive multitask dense prediction. In AAAI, 2023. +[20] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Learning multiple dense prediction tasks from partially annotated data. In CVPR, 2022. +[21] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. In NeurIPS, 2020. +[22] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In ICML, 2018. +[23] Zirui Wang, Yulia Tsvetkov, Orhan First, and Yuan Cao. Gradient vaccine: Investigating and improving multi-task optimization in massively multilingual models. In ICLR, 2020. +[24] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In NeurIPS, 2020. +[25] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In CVPR, 2016. +[26] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In CVPR, 2019. +[27] Lijun Zhang, Xiao Liu, and Hui Guan. Automtl: A programming framework for automating efficient multi-task learning. In NeurIPS, 2021. +[28] Yuan Gao, Haoping Bai, Zequn Jie, Jiayi Ma, Kui Jia, and Wei Liu. Mtl-nas: Task-agnostic neural architecture search towards general-purpose multi-task learning. In CVPR, 2020. +[29] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In CVPR, 2019. +[30] Xiaoya Zhang, Ling Zhou, Yong Li, Zhen Cui, Jin Xie, and Jian Yang. Transfer vision patterns for multi-task pixel learning. In ACMMM, 2021. + +[31] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In CVPR, 2020. +[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. +[33] Simon Vandenhende, Stamatios Georgoulis, Bert De Brabandere, and Luc Van Gool. Branched multi-task networks: deciding what layers to share. In ECCV, 2020. +[34] Yongxi Lu, Abhishek Kumar, Shuangfei Zhai, Yu Cheng, Tara Javidi, and Rogerio Feris. Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In CVPR, 2017. +[35] Pengsheng Guo, Chen-Yu Lee, and Daniel Ulbricht. Learning to branch for multi-task learning. In ICML, 2020. +[36] Hanrong Ye and Dan Xu. Taskprompt: Spatial-channel multi-task prompting for dense scene understanding. In ICLR, 2023. +[37] Michael I Jordan and Robert A Jacobs. Hierarchical mixtures of experts and the em algorithm. Neural Computation, 6(2):181-214, 1994. +[38] Robert A Jacobs and Michael I Jordan. Learning piecewise control strategies in a modular neural network architecture. IEEE Transactions on Systems, Man, and Cybernetics, 23(2):337-345, 1993. +[39] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. *Neural Computation*, 3(1):79-87, 1991. +[40] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013. +[41] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. +[42] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. JMLR, 23:1-39, 2022. +[43] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. In ICLR, 2021. +[44] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In KDD, 2018. +[45] Hanxue Liang, Zhiwen Fan, Rishov Sarkar, Ziyu Jiang, Tianlong Chen, Kai Zou, Yu Cheng, Cong Hao, and Zhangyang Wang. M3vit: Mixture-of-experts vision transformer for efficient multi-task learning with model-accelerator co-design. In NeurIPS, 2022. +[46] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik Learned-Miller, and Chuang Gan. Mod-squad: Designing mixture of experts as modular multitask learners. arXiv preprint arXiv:2212.08066, 2022. + +[47] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. +[48] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In CVPR, 2014. +[49] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. +[50] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. +[51] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. \ No newline at end of file diff --git a/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/images.zip b/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ecc4a2c4874e4b484e7bc0c3067928b2d49442df --- /dev/null +++ b/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:296877417199112d62d901e3e287c24eedfe5bba61b13a2c2d1eb53472907188 +size 556263 diff --git a/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/layout.json b/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..550034f17bbfa2878b56ea77dbd7fc885e0e8fda --- /dev/null +++ b/2023/TaskExpert_ Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts/layout.json @@ -0,0 +1,7311 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 501, + 140 + ], + "type": "text", + "content": "TaskExpert: Dynamically Assembling Multi-Task Representations with Memorial Mixture-of-Experts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 160, + 442, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 160, + 442, + 204 + ], + "spans": [ + { + "bbox": [ + 151, + 160, + 442, + 204 + ], + "type": "text", + "content": "Hanrong Ye and Dan Xu\\* \nDepartment of Computer Science and Engineering, HKUST \nClear Water Bay, Kowloon, Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 229, + 205, + 364, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 205, + 364, + 217 + ], + "spans": [ + { + "bbox": [ + 229, + 205, + 364, + 217 + ], + "type": "text", + "content": "{hyeae, danxu}@cse.ust.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 271, + 289, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 289, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 289, + 630 + ], + "type": "text", + "content": "Learning discriminative task-specific features simultaneously for multiple distinct tasks is a fundamental problem in multi-task learning. Recent state-of-the-art models consider directly decoding task-specific features from one shared task-generic feature (e.g., feature from a backbone layer), and utilize carefully designed decoders to produce multi-task features. However, as the input feature is fully shared and each task decoder also shares decoding parameters for different input samples, it leads to a static feature decoding process, producing less discriminative task-specific representations. To tackle this limitation, we propose TaskExpert, a novel multi-task mixture-of-experts model that enables learning multiple representative task-generic feature spaces and decoding task-specific features in a dynamic manner. Specifically, TaskExpert introduces a set of expert networks to decompose the backbone feature into several representative task-generic features. Then, the task-specific features are decoded by using dynamic task-specific gating networks operating on the decomposed task-generic features. Furthermore, to establish long-range modeling of the task-specific representations from different layers of TaskExpert, we design a multitask feature memory that updates at each layer and acts as an additional feature expert for dynamic task-specific feature decoding. Extensive experiments demonstrate that our TaskExpert clearly outperforms previous best-performing methods on all 9 metrics of two competitive multi-task learning benchmarks for visual scene understanding (i.e., PASCAL-Context and NYUD-v2). Codes and models will be made publicly available." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 656, + 127, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 127, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 127, + 669 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "With the rapid development of deep learning models, many computer vision tasks can be effectively handled by deep networks, which strongly motivates researchers to de" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 243, + 525, + 418 + ], + "blocks": [ + { + "bbox": [ + 313, + 243, + 525, + 418 + ], + "lines": [ + { + "bbox": [ + 313, + 243, + 525, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 525, + 418 + ], + "type": "image", + "image_path": "e2bee81ef2043ed6a48b3853a6685b448704df26edbcc8b6f19b9725b6e37485.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 423, + 545, + 567 + ], + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 567 + ], + "type": "text", + "content": "Figure 1: Motivative illustration of TaskExpert for dynamical decoding discriminative task-specific representations. A set of " + }, + { + "bbox": [ + 304, + 423, + 545, + 567 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 423, + 545, + 567 + ], + "type": "text", + "content": " expert networks decompose the task-generic backbone feature into " + }, + { + "bbox": [ + 304, + 423, + 545, + 567 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 423, + 545, + 567 + ], + "type": "text", + "content": " representative feature spaces. Then, the feature activations of an input from the multiple expert networks are dynamically assembled by the Task Gating network, to produce task-specific features for the " + }, + { + "bbox": [ + 304, + 423, + 545, + 567 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 423, + 545, + 567 + ], + "type": "text", + "content": " tasks. TaskExpert further performs long-range modeling of task-specific representations by designing a multi-task feature memory. It is updated at the current layer based on the assembled task-specific representations, and is used as an additional feature expert in the next layer." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 578, + 545, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 697 + ], + "type": "text", + "content": "velop unified multi-task networks to perform joint learning and inference on multiple distinct tasks [1, 2, 3, 4]. Multi-Task Learning (MTL) can avoid repetitive training of single-task models and is able to simultaneously generate predictions for different tasks with only one-time inference. On the other hand, as different computer vision tasks usually share a basic common understanding of input images, learning multiple tasks concurrently can help improve the representation power of each task and boost the overall performance of multi-task models [5, 6]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Learning discriminative task-specific features is a funda-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21828" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 335 + ], + "type": "text", + "content": "mental problem for MTL models. To achieve this objective, recent state-of-the-art works [7, 2, 8, 1] consider a decoder-focused paradigm, in which a shared backbone model with pretrained parameters is applied to learn generic image representations, and then separate task decoders are carefully designed to generate task-specific features. Although this paradigm is straightforward to achieve task-specific feature decoding, and produces promising multi-task prediction results, these models have two aspects of potential limitations: (i) The input task-generic feature from a backbone layer is fully shared for the different task decoders, which requires a careful design of the decoders (e.g. capacity and structure) for the different tasks to achieve task-discriminative decoding. (ii) The parameters of each task decoder are typically shared for different input samples, resulting in a static decoding of task-specific features, while a sample-dependant decoder is highly beneficial for learning more discriminative task features, as the input samples can be diverse, and sample-related context information is particularly important for multi-task learning and inference. All these aspects hamper the previous models from learning effective task-specific features and predictions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 337, + 289, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 337, + 289, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 289, + 671 + ], + "type": "text", + "content": "To tackle the above-mentioned issues, this paper proposes a novel multi-task mixture-of-experts framework, coined as \"TaskExpert\", which dynamically performs the task-feature decoding for different inputs and tasks as depicted in Figure 1. Instead of using one shared backbone feature to decode different task features, the proposed TaskExpert introduces a set of task-generic expert networks that learn to decompose the backbone feature into a set of representative task-generic features. The feature decomposition enables the decoders to interact with finer-granularity task-generic feature spaces to produce more discriminative task features, as each expert can be responsible for modeling one representative feature space from the training data. Then, we dynamically assemble feature activations from different experts to produce task-specific representations, based on sample-dependent and task-specific gating scores predicted from a designed gating network. On the other hand, visual understanding tasks greatly benefit from representations of different network levels [9]. Thus, we further devise a \"multi-task feature memory\" to aggregate the dynamically decoded task-specific features produced at different network layers, and propose a \"Memorial Mixture-of-Experts (MMoE)\" module. For each task, MMoE utilizes its corresponding task feature from the multi-task feature memory as an additional feature expert, to improve task-specific feature decoding. In this way, MMoE enables long-range dynamic modeling of task-specific representation for each task throughout the entire backbone." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 672, + 253, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 672, + 253, + 684 + ], + "spans": [ + { + "bbox": [ + 59, + 672, + 253, + 684 + ], + "type": "text", + "content": "In summary, the main contribution is three-fold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "type": "text", + "content": "- We propose a multi-task mixture-of-experts model that allows effective decomposition of a task-generic fea" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 324, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 109 + ], + "type": "text", + "content": "ture from any backbone network layer and enables dynamically assembling discriminative multi-task representations by a task gating mechanism." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 114, + 545, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 114, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 317, + 114, + 545, + 199 + ], + "type": "text", + "content": "- We design a novel Memorial Mixture-of-Experts (MMoE) module by further introducing a multi-task feature memory, which can effectively interact (read and write) with the mixture of experts at different network layers to achieve long-range dynamic decoding of task-specific features by utilizing different levels of network representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 205, + 547, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 205, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 317, + 205, + 547, + 313 + ], + "type": "text", + "content": "- The proposed TaskExpert is extensively validated on two challenging multi-task visual scene understanding benchmarks (i.e. Pascal-Context and NYUD-v2). The results clearly show the effectiveness of TaskExpert for dynamic multi-task representation learning, and also establish new state-of-the-art performances on all 9 metrics on the benchmarks, regarding the comparison with previous best-performing methods using both transformer- and CNN-based backbones." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 323, + 398, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 398, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 398, + 335 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 342, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 715 + ], + "type": "text", + "content": "Multi-Task Learning in Computer Vision In the field of multi-task learning (MTL) for computer vision problems [10, 11, 12, 13, 14, 15, 16, 17], some research works explore effective solutions through a perspective of model optimization and improve the optimization process of MTL via different multi-task loss designs [3, 6, 18, 19, 20] and gradient manipulations [21, 22, 23, 24]. These optimization strategies have also been demonstrated beneficial for alleviating the issue of task competition in multi-task learning [6]. From another perspective, several existing works also investigate designing better multi-task model architectures. These related works can be roughly divided into two groups: encoder-focused and decoder-focused methods. The encoder-focused methods [25, 26, 27, 28] target designing multi-task encoders, while the decoder-focused methods typically share the encoder (i.e. the backbone) for different tasks and design sophisticated decoders for task-specific feature learning and prediction [1, 2, 29, 8, 30, 31]. A recent state-of-the-art decoder-focused method (i.e. In-vPT [7]) performs multi-task learning on powerful transformer architectures [32], and establishes a very strong baseline for multi-task visual scene understanding. It can effectively learn spatial and cross-task relationships globally. To adapt multi-task models to different tasks, researchers also propose to search for task-specific network structures for encoders [33, 34, 35] and decoders [2], or utilize task-specific learnable prompts to effectively capture task-specific information [36]. Despite promising performances of these works, they are still limited by static decoder network designs since their decoders cannot dynamically generate task-specific representations based on dif" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21829" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 131 + ], + "type": "text", + "content": "ferent inputs and tasks. In this paper, we propose TaskExpert, a novel and effective mixture-of-experts framework, which can achieve dynamic multi-task representation learning via feature decomposition and assembling through the entire backbone network." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 289, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 289, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 289, + 469 + ], + "type": "text", + "content": "Mixture-of-Experts Models With statistical motivation [37], Mixture-of-Experts (MoE) models are originally designed to control the dynamics of neural networks automatically [38, 39]. MoE learns a series of expert networks and a gating network. The outputs of expert networks are weighted by gating scores (or called \"gate values\") generated by a gating network before the weighting operation. In more recent works, some researchers use the gating scores as a criterion to sparsely select only one or a few experts. The sparse activation of experts enables a significant reduction of the computational cost when training large-scale models [40, 41, 42, 43]. Related to our work, [44] designs a multi-gate MoE to ensemble expert networks for different census analysis tasks with different gating networks, and [45, 46] propose sparse task-conditioned MoE networks for efficient multi-task learning, which generate predictions for only one task within a single forward network inference. In contrast to these works, our Memorial Mixture-of-Experts (MMoE) is a plug-and-play decoder module, which is designed for dynamically assembling discriminative task-specific features. Our TaskExpert can simultaneously generate multi-task predictions for all tasks in one single forward inference, leading to significantly higher multi-task training efficiency and enabling cross-task interaction learning. Moreover, MMoE models task-specific features throughout different network layers via the novel design of multi-task feature memory and help generate more discriminative task-specific features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 482, + 123, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 482, + 123, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 482, + 123, + 496 + ], + "type": "text", + "content": "3. TaskExpert" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 502, + 173, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 173, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 173, + 514 + ], + "type": "text", + "content": "3.1. Framework Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": "An overall illustration of our TaskExpert framework is shown in Figure 2. The input image is first projected to a token sequence and then fed into a series of classic visual transformer layers. Our framework can also be flexibly used with a CNN backbone. The proposed Memorial Mixture-of-Experts (MMoE) module is embedded at different network layers, and each module takes the task-generic backbone feature of that layer as input. It dynamically assembles and outputs discriminative task-specific representations at the corresponding layer. The MMoE further interacts (reads and writes operations) with the designed multi-task feature memory, to perform a long-range dynamic multi-task feature decoding, which benefits from different levels of network features. The reads utilize the feature memory as an additional feature expert in the MMoE module, and the writes update the memory at the layer. Finally, after be" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 74, + 542, + 284 + ], + "blocks": [ + { + "bbox": [ + 309, + 74, + 542, + 284 + ], + "lines": [ + { + "bbox": [ + 309, + 74, + 542, + 284 + ], + "spans": [ + { + "bbox": [ + 309, + 74, + 542, + 284 + ], + "type": "image", + "image_path": "967f229000fb7786a71aa590a3c583170e9124f754048af17bf90b0a10cf03dc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 289, + 547, + 434 + ], + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 434 + ], + "type": "text", + "content": "Figure 2: Overall architecture of the proposed TaskExpert model. Memorial Mixture-of-Experts (MMoE) modules are embedded at different backbone layers to dynamically decode task-specific feature for each task at the corresponding layer. The MMoE also interacts (reads and writes) with a designed multi-task feature memory to have a long-range dynamic multi-task feature decoding benefiting from different levels of network features. The reads utilize the memory as an additional feature expert in MMoE, and the writes update the memory at the layer. After being updated by the last network layer, the multi-task feature memory is used for the final prediction of each task." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 450, + 547, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 511 + ], + "type": "text", + "content": "ing updated by the outputs of all MMoE modules at different network layers, the task-specific features in multi-task feature memory are then used for generating the multi-task predictions with simple prediction heads. We elaborate on more details of our model in the following several sections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 533, + 478, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 478, + 546 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 478, + 546 + ], + "type": "text", + "content": "3.2. Multi-Task Backbone Structure" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "We consider a widely-used vision transformer (ViT) model [47] with " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": " layers as the backbone of our multi-task MoE model. The input image is first divided into " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": " patches and processed by the patch embedding and positional embedding modules. Then, we obtain a sequence of patch tokens " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^0 = \\{\\mathbf{x}_1^0,\\mathbf{x}_2^0,\\dots,\\mathbf{x}_N^0\\}" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "N = H\\times W" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": " and the number of channels is " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": ". The patch tokens contain generic visual representations of the input image. We denote the " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "-th transformer layer as a mapping function " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "f^{l}" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": " and the updating of image patch tokens at different transformer layers can be formulated as " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l = f^l (\\mathbf{X}^{l - 1})" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": " denotes the output patch tokens of the transformer at the " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "-th layer." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21830" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 52, + 534, + 270 + ], + "blocks": [ + { + "bbox": [ + 55, + 52, + 534, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 52, + 534, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 52, + 534, + 270 + ], + "type": "image", + "image_path": "aefd51ed303b4e3c26e501c19ae2b517f2b5501a124d6fd1bee9dc1b7c125d8d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 275, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 545, + 373 + ], + "type": "text", + "content": "Figure 3: Illustration of the proposed Memorial Mixture-of-Experts (MMoE) module with the multi-task feature memory. MMoE accepts a task-generic feature from a backbone layer as input, and dynamically generates task-specific features for different tasks in a sample- and task-dependent manner. Specifically, MMoE decomposes the backbone feature to a set of " + }, + { + "bbox": [ + 46, + 275, + 545, + 373 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 275, + 545, + 373 + ], + "type": "text", + "content": " representative task-generic features " + }, + { + "bbox": [ + 46, + 275, + 545, + 373 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}_1^l, \\dots, \\mathbf{R}_K^l)" + }, + { + "bbox": [ + 46, + 275, + 545, + 373 + ], + "type": "text", + "content": " via the " + }, + { + "bbox": [ + 46, + 275, + 545, + 373 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 275, + 545, + 373 + ], + "type": "text", + "content": " expert networks. Then these features are used to assemble a set of " + }, + { + "bbox": [ + 46, + 275, + 545, + 373 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 275, + 545, + 373 + ], + "type": "text", + "content": " task-specific features based on dynamically generated task-gating scores from the task gating networks. To establish long-range modeling of task-specific representations crossing different network layers, the multi-task feature memory is designed to interact with the expert networks. The multi-task feature memory serves as an additional feature expert to decode the task-specific representations, and the decoded representations are also further used to update the multi-task feature memory." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 382, + 253, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 253, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 253, + 396 + ], + "type": "text", + "content": "3.3. Memorial Mixture-of-Experts (MMoE)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 401, + 288, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 288, + 498 + ], + "type": "text", + "content": "We present the details of several core parts of the proposed memorial mixture-of-experts (MMoE) module, including backbone generic-feature decomposition with a set of network experts, context-aware task-specific gating, the multi-task feature memory to have long-range modeling on task-specific features through interactions (reads and writes) with the mixture of experts, dynamic task-specific feature assembling, and multi-task prediction heads." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 514, + 281, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 281, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 281, + 528 + ], + "type": "text", + "content": "3.3.1 Backbone Feature Decomposition with Experts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "As shown in Figure 3, to enable learning discriminative task-specific representations from the task-generic backbone feature, we first define a set of expert networks that decompose the backbone feature into several representative task-generic features in sub-feature spaces, representing distinct patterns of the backbone features, as can be observed in Figure 6. It allows more fine-grained decoding of task-specific features. Specifically, at the " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "-th backbone layer, let us assume that we use " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " expert networks " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\{f_{e}^{l,i}(\\cdot):\\mathbb{R}^{N\\times C}\\to \\mathbb{R}^{N\\times C},i\\in [1,K]\\}" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " to learn distinct task-generic feature projections, where superscript " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " denotes the layer index and " + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": " is the expert index. Each expert network employs the same Linear-BatchNorm-RELU-Linear structure but with different parameters. Through each expert network, we obtain a set of representative fea" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 381, + 453, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 453, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 453, + 396 + ], + "type": "text", + "content": "tures " + }, + { + "bbox": [ + 306, + 381, + 453, + 396 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{R}_i^l\\in \\mathbb{R}^{N\\times C},i\\in [1,K]\\}" + }, + { + "bbox": [ + 306, + 381, + 453, + 396 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 392, + 403, + 545, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 403, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 392, + 403, + 545, + 418 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {i} ^ {l} = f _ {e} ^ {l, i} \\left(\\mathbf {X} ^ {l}\\right), \\tag {1}", + "image_path": "ab377b0a4b69ed35849783cc146271a1232773bb7c356f46b7c8361c73f4d803.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 427, + 509, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 427, + 509, + 439 + ], + "spans": [ + { + "bbox": [ + 306, + 427, + 509, + 439 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 427, + 509, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l" + }, + { + "bbox": [ + 306, + 427, + 509, + 439 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 306, + 427, + 509, + 439 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 306, + 427, + 509, + 439 + ], + "type": "text", + "content": "-th layer feature of the backbone." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 455, + 495, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 495, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 495, + 467 + ], + "type": "text", + "content": "3.3.2 Context-Aware Task-Specific Gating" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 557 + ], + "type": "text", + "content": "After we have obtained a set of representative task-generic features decomposed from a backbone feature for all the tasks, we aim at decoding discriminative task-specific features for different tasks by automatically assembling these representative features. Thus, we propose to adaptively control the contribution of each expert for each patch token by designing a task-specific gating network." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": "For each token in the representative task-generic features, we compute a task-gating score to control its contribution to generating the task-specific features. The task-gating scores of each task are dynamically generated by a task-specific gating network, using the features " + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^l" + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": "-th network layer. We term the task-specific gating network with a short name \"task gating\", and denote it as " + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "inline_equation", + "content": "f_{g}^{l,t}(\\cdot):\\mathbb{R}^{N\\times C\\to N\\times K}" + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "inline_equation", + "content": "t\\in [1,T]" + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": " is the total number of tasks) is a task index. The task-gating scores " + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^{l,t}\\in \\mathbb{R}^{N\\times K}" + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^{l,t} = f_{g}^{l,t}(\\mathbf{X}^{l})" + }, + { + "bbox": [ + 304, + 558, + 545, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "For the design of the gating networks, previous Mixture-of-Experts methods [42, 43, 41] use a simple linear layer followed by a Softmax function, which does not take into" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21831" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 58, + 541, + 163 + ], + "blocks": [ + { + "bbox": [ + 49, + 58, + 541, + 163 + ], + "lines": [ + { + "bbox": [ + 49, + 58, + 541, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 58, + 541, + 163 + ], + "type": "image", + "image_path": "75375102ceba2b139b1590812e19a8b5f0682786b66443627bf8d7174ea4fcd8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 165, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 165, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 165, + 547, + 201 + ], + "type": "text", + "content": "Figure 4: PCA visualization of the representative features by task-generic experts across different layers. In the deeper layers, the features become increasingly responsive to high-level semantic areas. Conversely, features at more shallow layers show a heightened sensitivity to low-level patterns." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 213, + 290, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 290, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 290, + 404 + ], + "type": "text", + "content": "account the context of each image patch token, while using only the corresponding patch token when computing the gating score. As various challenging computer vision tasks (e.g. semantic segmentation and monocular depth estimation) require modeling and understanding each image patch with context information, we thus design a context-aware gating strategy. Specifically, context-aware gating adopts convolution kernels to extend the receptive field of the gating networks, and thus it can involve a local context of tokens in the computation of gating scores. We first reshape the input token sequence " + }, + { + "bbox": [ + 46, + 213, + 290, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 213, + 290, + 404 + ], + "type": "text", + "content": " (in " + }, + { + "bbox": [ + 46, + 213, + 290, + 404 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 46, + 213, + 290, + 404 + ], + "type": "text", + "content": ") as spatial maps (in " + }, + { + "bbox": [ + 46, + 213, + 290, + 404 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 46, + 213, + 290, + 404 + ], + "type": "text", + "content": "), and then compute the gating scores of " + }, + { + "bbox": [ + 46, + 213, + 290, + 404 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 213, + 290, + 404 + ], + "type": "text", + "content": " experts with three layers, each having a " + }, + { + "bbox": [ + 46, + 213, + 290, + 404 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 46, + 213, + 290, + 404 + ], + "type": "text", + "content": " convolution with ReLU as intermediate activation function. We denote this computation procedure as \"CONV\", which will also be used later in Equation 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 420, + 203, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 203, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 203, + 433 + ], + "type": "text", + "content": "3.3.3 Multi-Task Feature Memory" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "text", + "content": "To conduct long-range modeling of the multi-task representations decoded at different network layers that contain different levels of visual information, we design a multitask feature memory to interact with the MMoE modules through read and write operations. The read operation uses the task feature memory as an additional feature expert for task-specific feature decoding in MMoE. The write operation updates the multi-task feature memory by using the decoded task-specific features from MMoE. In this way, the multi-task feature memory can aggregate the decoded task-specific features crossing different network layers. Specifically, the read operation is an identity mapping. The write operation can be performed as follows: given a decoded task-specific feature of the " + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "text", + "content": "-th task at the first layer, i.e. " + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{1,t}" + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "text", + "content": ", the initialization of the multi-task feature memory " + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{1,t}" + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "text", + "content": " is obtained by " + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{1,t} \\gets \\mathbf{F}^{1,t}" + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "text", + "content": ". In the subsequent layers, suppose we decode task-specific features at every network layer, at the " + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "text", + "content": "-th layer (" + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "inline_equation", + "content": "l > 1" + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "text", + "content": "), we write the multi-task feature memory " + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}^{l,t}, t \\in [1,T]\\}" + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "text", + "content": " with the decoded task-specific feature " + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{l,t}" + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "text", + "content": " weighted by a task-specific learnable momentum value " + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "inline_equation", + "content": "\\alpha^{l,t}" + }, + { + "bbox": [ + 46, + 439, + 290, + 693 + ], + "type": "text", + "content": " as shown in Figure 3:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 94, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {M} ^ {l, t} \\leftarrow \\mathbf {F} ^ {l, t} + \\alpha^ {l, t} \\mathbf {M} ^ {l - 1, t}, l > 1. \\tag {2}", + "image_path": "eea50508712cd9765c26a4b6f6d5e602006beb76565dcc0fc6dc78e63a52037e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 213, + 542, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 542, + 225 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 542, + 225 + ], + "type": "text", + "content": "3.3.4 Dynamically Assembling Task-Specific Features" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "text", + "content": "With the obtained representative task-generic features and task-specific gating scores as above-described, we can perform dynamic decoding of task-specific features, as the gating parameters are predicted in a sample-dependent manner. This is significantly different from existing decoder-focused multi-task structures, which share the decoding parameters for all the different input samples. Specifically, for the " + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "text", + "content": "-th task, we multiply every token of the representative task-generic features " + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{R}_i^l,i\\in [1,K]\\}" + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "text", + "content": " with their corresponding task-gating score " + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{G}_i^{l,t},i\\in [1,K]\\}" + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "text", + "content": " and sum the tokens in the same spatial position to decode task-specific feature. Furthermore, to learn more discriminative features for each task, we utilize the corresponding task-specific feature in multi-task feature memory as an additional feature expert in MMoE, which contains long-range task-specific information learned throughout the entire network. It helps decode more effective and long-term task-specific features. As shown in Figure 3, for the MMoE at the " + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "text", + "content": "-th transformer layer (" + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "inline_equation", + "content": "l > 1" + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "text", + "content": "), we compute an additional gating score map " + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{K + 1}^{l,t}" + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "text", + "content": " for the task-specific feature from the multi-task feature memory " + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{l - 1,t}" + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "text", + "content": ", which is performed by adding one more output channel into the task gating network " + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "inline_equation", + "content": "f_{g}^{l,t}:\\mathbb{R}^{N\\times C}\\to \\mathbb{R}^{N\\times (K + 1)}" + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "text", + "content": ". In summary, the task gating scores of the " + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "text", + "content": "-th task and the " + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "text", + "content": "-th layer, i.e., " + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^{l,t} = \\{\\mathbf{G}_1^{l,t},\\mathbf{G}_2^{l,t},\\dots,\\mathbf{G}_{K + 1}^{l,t}\\}" + }, + { + "bbox": [ + 304, + 231, + 546, + 534 + ], + "type": "text", + "content": ", is computed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 539, + 545, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 539, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 321, + 539, + 545, + 567 + ], + "type": "interline_equation", + "content": "\\mathbf {G} ^ {l, t} = \\left\\{ \\begin{array}{l l} \\operatorname {S o f t m a x} \\left(\\operatorname {C O N V} ^ {K + 1} \\left(\\mathbf {X} ^ {l}\\right)\\right) & l > 1, \\\\ \\operatorname {S o f t m a x} \\left(\\operatorname {C O N V} ^ {K} \\left(\\mathbf {X} ^ {l}\\right)\\right) & l = 1, \\end{array} \\right. \\tag {3}", + "image_path": "db051b209bd01412bd9c7005a41d372d538b4234542c11e5aa3c38d8831c3da5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 571, + 547, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 547, + 607 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 547, + 607 + ], + "type": "text", + "content": "where the superscript " + }, + { + "bbox": [ + 304, + 571, + 547, + 607 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 571, + 547, + 607 + ], + "type": "text", + "content": " of CONV denotes the number of output convolution kernels. And accordingly, the decoded task-specific features at the " + }, + { + "bbox": [ + 304, + 571, + 547, + 607 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 571, + 547, + 607 + ], + "type": "text", + "content": "-th MMoE module is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "type": "interline_equation", + "content": "\\mathbf {F} ^ {l, t} = \\left\\{ \\begin{array}{l l} \\sum_ {i = 1} ^ {K} \\mathbf {G} _ {i} ^ {l, t} \\times \\mathbf {R} _ {i} ^ {l} + \\mathbf {G} _ {K + 1} ^ {l, t} \\times \\mathbf {M} ^ {l - 1, t} & l > 1, \\\\ \\sum_ {i = 1} ^ {K} \\mathbf {G} _ {i} ^ {l, t} \\times \\mathbf {R} _ {i} ^ {l} & l = 1. \\end{array} \\right. \\tag {4}", + "image_path": "8c47bdc4d26ded40b44fc60977b9ff945f690abbb2f45498fbfb280e23d7b7b2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Alternative sparse activation of experts As an alternative, we can also use \"sparse connections\" in MMoE, which means that we only activate " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "K_{sel}" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " selected experts to compute the task-specific features (" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "K_{sel} < K" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": ") for each token [41]. In this case, for each token, we find the top " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "K_{sel}" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21832" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 286, + 164 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 286, + 164 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 286, + 164 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 286, + 164 + ], + "type": "image", + "image_path": "c6c7890ca9c9ec06deb4dcfa7ad740e344b73869e4cb9139fc056a0700537c8e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 167, + 286, + 202 + ], + "lines": [ + { + "bbox": [ + 47, + 167, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 167, + 286, + 202 + ], + "type": "text", + "content": "Figure 5: PCA visualization of the decoded final task features generated by InvPT [7] and our TaskExpert. Our model generates more discriminative features for all tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 213, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 286, + 262 + ], + "type": "text", + "content": "experts to be activated based on the task-gating scores, and sum their outputs after being weighted by the task-gating scores. A comparison between the dense connection and sparse connection in MMoE is discussed in the experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 268, + 244, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 268, + 244, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 268, + 244, + 280 + ], + "type": "text", + "content": "3.4. Generation of Multi-Task Predictions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 287, + 286, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 287, + 286, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 287, + 286, + 370 + ], + "type": "text", + "content": "At the final network layer " + }, + { + "bbox": [ + 47, + 287, + 286, + 370 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 287, + 286, + 370 + ], + "type": "text", + "content": ", we use task-specific prediction heads to directly generate prediction outputs for each task from the multi-task feature memory " + }, + { + "bbox": [ + 47, + 287, + 286, + 370 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}^{L,t}, t \\in [1, T]\\}" + }, + { + "bbox": [ + 47, + 287, + 286, + 370 + ], + "type": "text", + "content": ". The prediction heads for different tasks have a similar structure, consisting of a " + }, + { + "bbox": [ + 47, + 287, + 286, + 370 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 47, + 287, + 286, + 370 + ], + "type": "text", + "content": " convolution layer with batch normalization and ReLU activation, followed by a linear layer for the final prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 380, + 123, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 123, + 394 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 123, + 394 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 400, + 286, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 400, + 286, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 400, + 286, + 437 + ], + "type": "text", + "content": "In this section, we validate the effectiveness of our proposal through several aspects: (i) ablation study, (ii) comparison with SOTA methods, and (iii) qualitative analysis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 444, + 162, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 162, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 162, + 456 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 462, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 286, + 652 + ], + "type": "text", + "content": "Tasks and Benchmarks For evaluation, we adopt two popular benchmarks in multi-task learning as previous works [7, 1], i.e. PASCAL-Context [48, 49] and NYUD-v2 [50] datasets. PASCAL-Context contains 4,998 samples in train split and 5,105 samples in test split. It provides labels for multiple important visual understanding tasks, including semantic segmentation, human parsing, and object boundary detection. Previous work on multi-task learning generates pseudo labels for another two tasks [5]: surface normals estimation and saliency detection. Meanwhile, NYUD-v2 contains 1,449 samples in train split and 795 samples in test split. It collects labels for semantic segmentation and monocular depth estimation, from which the labels of surface normal estimation and object boundary detection can be generated. We utilize all the tasks provided in these datasets for evaluation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 286, + 713 + ], + "type": "text", + "content": "Metrics We use the same evaluation metrics as previous works [7]. Specifically, semantic segmentation (Semseg) and human parsing (Parsing) tasks are evaluated by the mean Intersection over Union (mIoU). Monocular depth estimation task (Depth) is evaluated by Root Mean" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 131, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 545, + 129 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 545, + 129 + ], + "type": "text", + "content": "Table 1: Ablation study on PASCAL-Context with a ViT-Base backbone. \"MoE\" denotes using Mixture-of-Experts module; \"CG\" denotes using context-aware gating; \"Mem\" denotes using our multi-task feature memory. " + }, + { + "bbox": [ + 306, + 71, + 545, + 129 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 306, + 71, + 545, + 129 + ], + "type": "text", + "content": " ’ indicates lower better and ‘↑’ means higher better." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 131, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 308, + 131, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 131, + 545, + 182 + ], + "type": "table", + "html": "
ModelSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑MTL Perf Δm ↑
Baseline76.4065.7084.6313.8369.50-3.70
w/ MoE77.6266.3784.6513.7770.40-2.88
w/ MoE+CG77.9966.8284.7613.6471.70-2.07
w/ MoE+CG+Mem78.4567.3884.9613.5572.30-1.46
", + "image_path": "da73c8b4f6439380bfd4792102daab4c843c140212fa23ad2395ce8eac9170b9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 184, + 545, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 184, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 305, + 184, + 545, + 279 + ], + "type": "text", + "content": "Square Error (RMSE). Surface normal estimation task (Normal) is evaluated by the mean error (mErr) of angles. Saliency detection task (Saliency) is evaluated by maximal F-measure (maxF). Object boundary detection task (Boundary) is evaluated by the optimal-dataset-scale F-measure (odsF). We calculate the mean relative difference across all tasks compared to the single-task baseline, which is noted as Multi-task Performance (MTL Perf " + }, + { + "bbox": [ + 305, + 184, + 545, + 279 + ], + "type": "inline_equation", + "content": "\\Delta_{m}" + }, + { + "bbox": [ + 305, + 184, + 545, + 279 + ], + "type": "text", + "content": ") [5]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 280, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 280, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 280, + 545, + 411 + ], + "type": "text", + "content": "Implementation Details We use 5 expert networks in MMoE for PASCAL-Context and 4 expert networks for NYUD-v2, which are the same as the corresponding number of tasks. We add MMoE at the end of each transformer layer if not otherwise stated. For model training, we follow the setting of InvPT [7] and use Adam optimizer to train the models for 40,000 iterations with a batch size of 4. Polynomial learning rate scheduler is used with a learning rate of " + }, + { + "bbox": [ + 305, + 280, + 545, + 411 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 305, + 280, + 545, + 411 + ], + "type": "text", + "content": " and a weight decay rate of " + }, + { + "bbox": [ + 305, + 280, + 545, + 411 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 305, + 280, + 545, + 411 + ], + "type": "text", + "content": ". We adopt the loss functions and weights for different tasks in [7]. The ViT backbones are pre-trained on ImageNet-22K [51]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 420, + 399, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 399, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 399, + 432 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 438, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 438, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 545, + 713 + ], + "type": "text", + "content": "Model Variants Definition To examine the effectiveness of TaskExpert under a fair setting, we compare it to the multi-task baseline model and its variants with the same backbone. (1) The baseline model (\"Baseline\") uses ViT-Base as backbone, and designs task-specific decoders in each transformer layer to generate layer-wise task-specific features. The task-specific features from all layers are added together as multi-scale input of the corresponding prediction head for each task. These layer-wise decoders share the same structure as the expert networks used in MMoE. The prediction heads for all the tasks have a similar structure: " + }, + { + "bbox": [ + 305, + 438, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}(3\\times 3)" + }, + { + "bbox": [ + 305, + 438, + 545, + 713 + ], + "type": "text", + "content": "-BN-ReLU-Linear, which are the same as those used in TaskExpert. (2) \"TaskExpert w/ MoE\" adds the expert and gating networks of MMoE without context-aware gating to Baseline. It uses gating networks to control the contribution of each expert to each task in a sample-dependent manner without context-aware gating or multi-task feature memory. The number of experts is the same as the number of tasks. (3) \"TaskExpert w/ MoE+CG\" further uses our context-aware gating strategy, which improves the gating networks by introducing the contextual information of each token. (4) \"TaskExpert w/ MoE+CG+Mem\" is the full version of our TaskExpert using multi-task feature" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21833" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 139 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 139 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 139 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 139 + ], + "type": "image", + "image_path": "d7d2f108e35935847d8bf42d22a1760fda9c68894d59035d49c05ffd47068495.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 142, + 287, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 142, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 142, + 287, + 237 + ], + "type": "text", + "content": "Figure 6: Visualization of the task-gating scores by task-specific gating networks and the decomposed representative features from expert networks in MMoE modules. The gating scores for different experts show distinct distributions, while the representative features are discriminative to different visual elements. These results suggest different experts are able to decompose diverse visual information from the backbone feature." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 243, + 286, + 334 + ], + "blocks": [ + { + "bbox": [ + 50, + 243, + 286, + 334 + ], + "lines": [ + { + "bbox": [ + 50, + 243, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 50, + 243, + 286, + 334 + ], + "type": "image", + "image_path": "a7a7eee4a7e5bb9c92731e62b4c8fb1e027dcbeddc18976ec1f42aef2aa0cba2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 337, + 287, + 374 + ], + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 374 + ], + "type": "text", + "content": "Figure 7: Qualitative comparison with SOTA method (i.e. InvPT [7]) on PASCAL-Context. Our TaskExpert generates better predictions with more accurate details as highlighted." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 388, + 286, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 411 + ], + "type": "text", + "content": "memory in the proposed MMoE module to establish modeling of long-term task-specific representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 413, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 286, + 652 + ], + "type": "text", + "content": "Key Results We analyze the performances of different model variants introduced above in Table 1. We find that using MoE brings a clear performance gain compared to the baseline model, which demonstrates the effectiveness of designing task-specific gating networks to aggregate the representative task-generic features for dynamically decoding task-specific features. For example, mIoU of semantic segmentation is improved from 76.40 to 77.62 (+1.22), and human parsing is improved from 65.70 to 66.37 (+0.67). Furthermore, the usage of context-aware gating strategy helps MoE achieves better performance by enabling the learning of contextual information for each token when computing the gating scores. For example, odsF of boundary detection is improved from 70.40 to 71.70 (+1.30) compared to only using MoE. Finally, with long-term modeling of task-specific representations via the proposed multi-task feature memory of Memorial Mixture-of-Experts, TaskExpert achieves significant performance improvement compared to the baseline model on different tasks, and the overall multitask performance " + }, + { + "bbox": [ + 46, + 413, + 286, + 652 + ], + "type": "inline_equation", + "content": "\\Delta_{m}" + }, + { + "bbox": [ + 46, + 413, + 286, + 652 + ], + "type": "text", + "content": " is improved by +2.24." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 286, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 715 + ], + "type": "text", + "content": "Study of Spare and Dense Connections in MoE The original MoE models use the dense connection as the gating mechanism: all experts are activated and contribute to the output [38, 39]. Recent works propose to utilize sparse connection as the gating mechanism: for each token, only a" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 89, + 544, + 147 + ], + "blocks": [ + { + "bbox": [ + 305, + 65, + 545, + 89 + ], + "lines": [ + { + "bbox": [ + 305, + 65, + 545, + 89 + ], + "spans": [ + { + "bbox": [ + 305, + 65, + 545, + 89 + ], + "type": "text", + "content": "Table 2: Variants of TaskExpert on PASCAL-Context. The inference GPU memory is measured with batch size 1." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 89, + 544, + 147 + ], + "lines": [ + { + "bbox": [ + 309, + 89, + 544, + 147 + ], + "spans": [ + { + "bbox": [ + 309, + 89, + 544, + 147 + ], + "type": "table", + "html": "
VariantSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑#ParamFLOPSInference GPU Mem
3 MMoE78.0466.5484.8513.6271.20174M597G8367M
6 MMoE78.1866.7984.9313.5971.70234M659G8851M
12 MMoE78.4567.3884.9613.5572.30347M782G12503M
ViT-S75.0462.6884.6814.2268.8055M204G4187M
ViT-B78.4567.3884.9613.5572.30347M782G12503M
", + "image_path": "b12de7c3f03495262494cc966725a4a868e0ccb45b0b7634d314b5a5605d2040.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 185, + 544, + 226 + ], + "blocks": [ + { + "bbox": [ + 305, + 149, + 545, + 184 + ], + "lines": [ + { + "bbox": [ + 305, + 149, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 305, + 149, + 545, + 184 + ], + "type": "text", + "content": "Table 3: Comparison with multi-task MoE methods with ViT-S backbone on PASCAL-Context. The training GPU time is shown in the last column." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 185, + 544, + 226 + ], + "lines": [ + { + "bbox": [ + 309, + 185, + 544, + 226 + ], + "spans": [ + { + "bbox": [ + 309, + 185, + 544, + 226 + ], + "type": "table", + "html": "
VariantSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑#ParamFLOPSGPU Hours
M3ViT [45]72.8062.1066.3014.5071.7042M420G~72
Mod-Squad [46]74.1062.7066.9013.7072.0050M420G~72
Ours75.0462.6884.6814.2268.8055M204G~5
", + "image_path": "66395fec9962adb97f21b634e18b238b5aeec70bffed3b79e1e2cfee9c4c4efe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 271, + 545, + 357 + ], + "blocks": [ + { + "bbox": [ + 305, + 236, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 305, + 236, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 545, + 270 + ], + "type": "text", + "content": "Table 4: Comparison with the state-of-the-art methods with the same backbone on PASCAL-Context. Our TaskExpert clearly achieves superior performances on all tasks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 271, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 309, + 271, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 309, + 271, + 545, + 357 + ], + "type": "table", + "html": "
ModelFLOPs#ParamSemseg mIoU ↑Parsing mIoU ↑Saliency maxF ↑Normal mErr ↓Boundary odsF ↑
PAD-Net [8]124G81M53.6059.6065.8015.3072.50
MTI-Net [1]161G128M61.7060.1884.7814.2370.80
ATRC [2]216G96M67.6762.9382.2914.2472.42
PAD-Net w/ ViT-L [8]773G330M78.0167.1279.2114.3772.60
MTI-Net w/ ViT-L [1]774G851M78.3167.4084.7514.6773.00
ATRC w/ ViT-L [2]871G340M77.1166.8481.2014.2372.10
InvPT [7]669G423M79.0367.6184.8114.1573.00
TaskExpert (ours)622G420M80.6469.4284.8713.5673.30
", + "image_path": "cf71a362c7c727b01280fd9e24de5f1fdaeed2b6b14dde736d3b629baa8989e5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 372, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 479 + ], + "type": "text", + "content": "subset of experts contribute to the output [41], as explained in Section 3.3. In this experiment, we study the performances of using sparse connection (top 1 and top 2), and compare them with the default dense connection manner on PASCAL-Context. The results are demonstrated in Figure 8. It can be observed that activating more experts leads to better performance, and the dense connection achieves the best results. The reason is that it is able to freely utilize the full representative ability of all expert networks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 480, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 545, + 588 + ], + "type": "text", + "content": "Light-Weight Versions of TaskExpert To study the model performance and efficiency with a smaller parameter and memory budget, we conduct 2 sets of experiments in Table 3: (i) Using different numbers of MMoE modules on ViT-B backbone for PASCAL-Context. The MMoE modules are embedded evenly along the backbone's depth. (ii) Implementing TaskExpert on ViT-S, which is a smaller backbone. These smaller variants achieve reasonable performance with less parameter and memory consumption." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 598, + 485, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 598, + 485, + 611 + ], + "spans": [ + { + "bbox": [ + 305, + 598, + 485, + 611 + ], + "type": "text", + "content": "4.3. Comparison with SOTA Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We implement TaskExpert based on ViT-Large backbone with MMoE modules embedded at every two transformer layers, then we compare TaskExpert to recent SOTA methods in literature on both PASCAL-Context and NYUDv2 in Table 4 and Table 5. We re-implement several CNN-based SOTA multi-task learning methods on ViT-L backbone for a more comprehensive comparison on PASCAL-Context. Compared with the previous SOTA" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21834" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 51, + 544, + 155 + ], + "blocks": [ + { + "bbox": [ + 49, + 51, + 544, + 155 + ], + "lines": [ + { + "bbox": [ + 49, + 51, + 544, + 155 + ], + "spans": [ + { + "bbox": [ + 49, + 51, + 544, + 155 + ], + "type": "image", + "image_path": "96a37487975fa33197668afd601913c587312b4572de77d778616d84ef74ef2e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 157, + 547, + 182 + ], + "lines": [ + { + "bbox": [ + 46, + 157, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 547, + 182 + ], + "type": "text", + "content": "Figure 8: Comparison of TaskExpert with dense and sparse connection in MMoE. \"Top 1\" denotes only activating the expert with the highest gating score, \"Top 2\" means activating the experts with top 2 gating scores, and \"Dense\" activates all experts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 227, + 287, + 327 + ], + "blocks": [ + { + "bbox": [ + 46, + 191, + 287, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 191, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 287, + 226 + ], + "type": "text", + "content": "Table 5: Comparison with the state-of-the-art methods on NYUD-v2 dataset. Our TaskExpert achieves the best performances on all tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 227, + 287, + 327 + ], + "lines": [ + { + "bbox": [ + 51, + 227, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 51, + 227, + 287, + 327 + ], + "type": "table", + "html": "
ModelBackboneSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑
Cross-Stitch [25]HRNet1836.340.629020.8876.38
PAP [29]HRNet1836.720.617820.8276.42
PSD [31]HRNet1836.690.624620.8776.42
PAD-Net [8]HRNet1836.610.627020.8576.38
MTI-Net [1]HRNet4845.970.536520.2777.86
ATRC [2]HRNet4846.330.536320.1877.94
InvPT [7]ViT-L53.560.518319.0478.10
TaskExpert (ours)ViT-L55.350.515718.5478.40
", + "image_path": "8c733f4d6373526952daed207b83ba6f9414547a1caa83804153701beeadfc8f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 335, + 287, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 490 + ], + "type": "text", + "content": "model (i.e. InvPT [7]), TaskExpert achieves clearly better performances on all 9 metrics of these two challenging benchmarks with the same backbone. Particularly, on PASCAL-Context, we improve semantic segmentation by +1.61 and human parsing by +1.81. On NYUD-v2, semantic segmentation is improved by +1.79. It should be noted that, our gains are measured upon the very strong performance foundation established by InvPT (e.g. Semseg 79.03, Saliency 84.81 on PASCAL-Context). This superior performance clearly shows the effectiveness of the proposed TaskExpert. TaskExpert also maintains a higher efficiency with a reduction of 3 million in the number of parameters and 47 billion FLOPs compared to InvPT." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 491, + 288, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 288, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 288, + 659 + ], + "type": "text", + "content": "As underscored in the Related Work section, our TaskExpert is designed for a different setting from " + }, + { + "bbox": [ + 46, + 491, + 288, + 659 + ], + "type": "inline_equation", + "content": "\\mathrm{M}^3\\mathrm{ViT}" + }, + { + "bbox": [ + 46, + 491, + 288, + 659 + ], + "type": "text", + "content": " [45] and Mod-Squad [46]. Their models are constrained to generating a single task prediction per forward pass, while our model capably generates predictions for all tasks within a single forward pass. Therefore, the multi-task learning efficiency of our model is significantly higher than their methods. When utilizing the same backbone, it requires only half the FLOPS, as demonstrated in Table 3. Moreover, our model's training duration is markedly shorter since it learns all tasks concurrently in each iteration. While they necessitate around 18 hours of training with 4 A6000 GPUs as confirmed by their authors, our model completes training in under 5 hours using a single NVIDIA 3090 GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 670, + 150, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 150, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 150, + 683 + ], + "type": "text", + "content": "4.4. Qualitative study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Qualitative Study of MMoE To understand better the learned gating scores and representative task-generic fea" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 193, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 289 + ], + "type": "text", + "content": "tures by the gating networks and expert networks in our MMoE modules, we present qualitative examples in Figure 6. As can be observed, TaskExpert effectively learns representative task-generic features with distinct activation distributions, focusing on different aspects of the image. Accordingly, the gating scores exhibit different response intensities to different areas for selectively assembling the experts' features for decoding task-specific features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 289, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 396 + ], + "type": "text", + "content": "PCA Visualization of Representative Features of Task-Experts at Different Layers We conduct principal component analysis (PCA) on the representative features at layers 1, 6, and 12. As shown in Fig. 4, we visualize the first three components of the feature PCA as a color map. It reveals that the task experts at deeper layers capture more high-level semantic information of the input. On the other hand, in the shallower layers, the task-feature experts focus more on low-level visual patterns (e.g., edges)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 396, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 467 + ], + "type": "text", + "content": "Qualitative Comparison of Decoded Task Features In Fig. 5, we also conduct PCA on the final task features generated by InvPT [7] and our TaskExpert, and draw the first three components. The features for different tasks from TaskExpert show clear task-specific characteristics and are much more discriminative than those of InvPT." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 468, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 540 + ], + "type": "text", + "content": "Qualitative Comparison of Prediction Results To intuitively examine the prediction quality of TaskExpert compared to the previous best-performing method [7], we visualize the predictions of the two models on the test split of PASCAL-Context in Figure 7. Our TaskExpert generates visually better multi-task predictions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 378, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we present TaskExpert, a novel multitask mixture-of-expert model establishing new SOTA performances on two challenging benchmarks. TaskExpert can effectively perform dynamic decoding of task-specific features, from a set of representative task-generic features learned by expert networks. TaskExpert further models the long-term task-specific representations at different layers via a designed multi-task feature memory. Extensive experiments clearly prove the effectiveness and efficiency of our TaskExpert model, and we hope that our TaskExpert can also be beneficial for the future study of dynamic representation learning problems." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21835" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 145, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 145, + 85 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 145, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 140 + ], + "type": "text", + "content": "This research is supported in part by HKUST-SAIL joint research funding, the Early Career Scheme of the Research Grants Council (RGC) of the Hong Kong SAR under grant No. 26202321, and HKUST Startup Fund No. R9253." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 106, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 106, + 165 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 106, + 165 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 172, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 172, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 287, + 206 + ], + "type": "text", + "content": "[1] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In ECCV, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 210, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 210, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 210, + 288, + 255 + ], + "type": "text", + "content": "[2] David Bruggemann, Menelaos Kanakis, Anton Obukhov, Stamatos Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In ICCV, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 259, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 259, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 52, + 259, + 287, + 294 + ], + "type": "text", + "content": "[3] Amir R Zamir, Alexander Sax, Nikhil Cheerla, Rohan Suri, Zhangjie Cao, Jitendra Malik, and Leonidas J Guibas. Robust learning through cross-task consistency. In CVPR, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 298, + 287, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 298, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 52, + 298, + 287, + 332 + ], + "type": "text", + "content": "[4] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 336, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 336, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 52, + 336, + 287, + 369 + ], + "type": "text", + "content": "[5] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In CVPR, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 374, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 374, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 52, + 374, + 287, + 408 + ], + "type": "text", + "content": "[6] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 413, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 413, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 52, + 413, + 287, + 435 + ], + "type": "text", + "content": "[7] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. In ECCV, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 440, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 440, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 52, + 440, + 287, + 483 + ], + "type": "text", + "content": "[8] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In CVPR, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 52, + 489, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 489, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 52, + 489, + 287, + 522 + ], + "type": "text", + "content": "[9] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 527, + 286, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 527, + 286, + 549 + ], + "spans": [ + { + "bbox": [ + 48, + 527, + 286, + 549 + ], + "type": "text", + "content": "[10] Yu Zhang and Qiang Yang. A survey on multi-task learning. TKDE, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "text", + "content": "[11] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. TPAMI, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 604, + 287, + 637 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 287, + 637 + ], + "type": "text", + "content": "[12] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In CVPR, 2018." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 642, + 287, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 642, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 48, + 642, + 287, + 675 + ], + "type": "text", + "content": "[13] Menelaos Kanakis, Thomas E Huang, David Brüggemann, Fisher Yu, and Luc Van Gool. Composite learning for robust and effective dense predictions. In WACV, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[14] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In ECCV, 2022." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Lukas Hoyer, Dengxin Dai, Yuhua Chen, Adrian Koring, Suman Saha, and Luc Van Gool. Three ways to improve semantic segmentation with self-supervised depth estimation. In CVPR, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 121, + 545, + 143 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 121, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 307, + 121, + 545, + 143 + ], + "type": "text", + "content": "[16] Shikun Liu, Edward Johns, and Andrew J Davison. End-to-end multi-task learning with attention. In CVPR, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 147, + 545, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 147, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 307, + 147, + 545, + 190 + ], + "type": "text", + "content": "[17] Xiangtai Li, Henghui Ding, Wenwei Zhang, Haobo Yuan, Jiangmiao Pang, Guangliang Cheng, Kai Chen, Ziwei Liu, and Chen Change Loy. Transformer-based visual segmentation: A survey. arXiv preprint arXiv:2304.09854, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "type": "text", + "content": "[18] Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. Conflict-averse gradient descent for multi-task learning. In NeurIPS, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 231, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 253 + ], + "type": "text", + "content": "[19] Siwei Yang, Hanrong Ye, and Dan Xu. Contrastive multitask dense prediction. In AAAI, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 257, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 257, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 257, + 545, + 289 + ], + "type": "text", + "content": "[20] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Learning multiple dense prediction tasks from partially annotated data. In CVPR, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 293, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 293, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 545, + 338 + ], + "type": "text", + "content": "[21] Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gradient sign dropout. In NeurIPS, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "text", + "content": "[22] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In ICML, 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 389, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 389, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 389, + 545, + 432 + ], + "type": "text", + "content": "[23] Zirui Wang, Yulia Tsvetkov, Orhan First, and Yuan Cao. Gradient vaccine: Investigating and improving multi-task optimization in massively multilingual models. In ICLR, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 437, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 470 + ], + "type": "text", + "content": "[24] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. In NeurIPS, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 474, + 545, + 506 + ], + "type": "text", + "content": "[25] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In CVPR, 2016." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 511, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 554 + ], + "type": "text", + "content": "[26] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In CVPR, 2019." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 559, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 591 + ], + "type": "text", + "content": "[27] Lijun Zhang, Xiao Liu, and Hui Guan. Automtl: A programming framework for automating efficient multi-task learning. In NeurIPS, 2021." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 596, + 545, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 596, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 307, + 596, + 545, + 639 + ], + "type": "text", + "content": "[28] Yuan Gao, Haoping Bai, Zequn Jie, Jiayi Ma, Kui Jia, and Wei Liu. Mtl-nas: Task-agnostic neural architecture search towards general-purpose multi-task learning. In CVPR, 2020." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 643, + 545, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 643, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 307, + 643, + 545, + 676 + ], + "type": "text", + "content": "[29] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In CVPR, 2019." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[30] Xiaoya Zhang, Ling Zhou, Yong Li, Zhen Cui, Jin Xie, and Jian Yang. Transfer vision patterns for multi-task pixel learning. In ACMMM, 2021." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21836" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[31] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In CVPR, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 108, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 108, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 108, + 286, + 140 + ], + "type": "text", + "content": "[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 144, + 286, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 144, + 286, + 176 + ], + "spans": [ + { + "bbox": [ + 49, + 144, + 286, + 176 + ], + "type": "text", + "content": "[33] Simon Vandenhende, Stamatios Georgoulis, Bert De Brabandere, and Luc Van Gool. Branched multi-task networks: deciding what layers to share. In ECCV, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 178, + 286, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 178, + 286, + 222 + ], + "spans": [ + { + "bbox": [ + 49, + 178, + 286, + 222 + ], + "type": "text", + "content": "[34] Yongxi Lu, Abhishek Kumar, Shuangfei Zhai, Yu Cheng, Tara Javidi, and Rogerio Feris. Fully-adaptive feature sharing in multi-task networks with applications in person attribute classification. In CVPR, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 225, + 286, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 225, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 49, + 225, + 286, + 247 + ], + "type": "text", + "content": "[35] Pengsheng Guo, Chen-Yu Lee, and Daniel Ulbricht. Learning to branch for multi-task learning. In ICML, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 250, + 286, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 250, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 49, + 250, + 286, + 281 + ], + "type": "text", + "content": "[36] Hanrong Ye and Dan Xu. Taskprompt: Spatial-channel multi-task prompting for dense scene understanding. In ICLR, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 285, + 286, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 285, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 49, + 285, + 286, + 317 + ], + "type": "text", + "content": "[37] Michael I Jordan and Robert A Jacobs. Hierarchical mixtures of experts and the em algorithm. Neural Computation, 6(2):181-214, 1994." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 320, + 286, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 286, + 363 + ], + "type": "text", + "content": "[38] Robert A Jacobs and Michael I Jordan. Learning piecewise control strategies in a modular neural network architecture. IEEE Transactions on Systems, Man, and Cybernetics, 23(2):337-345, 1993." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 367, + 286, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 367, + 286, + 399 + ], + "spans": [ + { + "bbox": [ + 49, + 367, + 286, + 399 + ], + "type": "text", + "content": "[39] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. *Neural Computation*, 3(1):79-87, 1991." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 403, + 286, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 403, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 49, + 403, + 286, + 435 + ], + "type": "text", + "content": "[40] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 437, + 286, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 437, + 286, + 481 + ], + "spans": [ + { + "bbox": [ + 49, + 437, + 286, + 481 + ], + "type": "text", + "content": "[41] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 484, + 286, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 484, + 286, + 517 + ], + "spans": [ + { + "bbox": [ + 49, + 484, + 286, + 517 + ], + "type": "text", + "content": "[42] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. JMLR, 23:1-39, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 519, + 286, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 519, + 286, + 573 + ], + "spans": [ + { + "bbox": [ + 49, + 519, + 286, + 573 + ], + "type": "text", + "content": "[43] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. In ICLR, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 576, + 286, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 576, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 49, + 576, + 286, + 609 + ], + "type": "text", + "content": "[44] Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In KDD, 2018." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 613, + 286, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 286, + 666 + ], + "type": "text", + "content": "[45] Hanxue Liang, Zhiwen Fan, Rishov Sarkar, Ziyu Jiang, Tianlong Chen, Kai Zou, Yu Cheng, Cong Hao, and Zhangyang Wang. M3vit: Mixture-of-experts vision transformer for efficient multi-task learning with model-accelerator co-design. In NeurIPS, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "text", + "content": "[46] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik Learned-Miller, and Chuang Gan. Mod-squad: Designing mixture of experts as modular multitask learners. arXiv preprint arXiv:2212.08066, 2022." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 297 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[47] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "type": "text", + "content": "[48] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In CVPR, 2014." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 190, + 545, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 190, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 307, + 190, + 545, + 224 + ], + "type": "text", + "content": "[49] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 260 + ], + "type": "text", + "content": "[50] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[51] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21837" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/1f4cb2e9-7a0d-4b5a-99a2-329016955f96_content_list.json b/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/1f4cb2e9-7a0d-4b5a-99a2-329016955f96_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bf422b007cf8e039edd3f181a2596b3114114abd --- /dev/null +++ b/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/1f4cb2e9-7a0d-4b5a-99a2-329016955f96_content_list.json @@ -0,0 +1,1389 @@ +[ + { + "type": "text", + "text": "Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation", + "text_level": 1, + "bbox": [ + 129, + 130, + 841, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianan Fan $^{1}$ , Dongnan Liu $^{1}$ , Hang Chang $^{2}$ , Heng Huang $^{3}$ , Mei Chen $^{4}$ , and Weidong Cai $^{1}$", + "bbox": [ + 135, + 196, + 831, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ University of Sydney $^{2}$ Lawrence Berkeley National Laboratory $^{3}$ University of Maryland at College Park $^{4}$ Microsoft", + "bbox": [ + 89, + 226, + 880, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jfan6480@uni.sydney.edu.au dongnan.liu@sydney.edu.au hchang@lbl.gov", + "bbox": [ + 169, + 244, + 803, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "henghuanghh@gmail.com Mei.Chen@microsoft.com tom.cai@sydney.edu.au", + "bbox": [ + 169, + 257, + 797, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 316, + 313, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The success of automated medical image analysis depends on large-scale and expert-annotated training sets. Unsupervised domain adaptation (UDA) has been raised as a promising approach to alleviate the burden of labeled data collection. However, they generally operate under the closed-set adaptation setting assuming an identical label set between the source and target domains, which is over-restrictive in clinical practice where new classes commonly exist across datasets due to taxonomic inconsistency. While several methods have been presented to tackle both domain shifts and incoherent label sets, none of them take into account the common characteristics of the two issues and consider the learning dynamics along network training. In this work, we propose optimization trajectory distillation, a unified approach to address the two technical challenges from a new perspective. It exploits the low-rank nature of gradient space and devises a dual-stream distillation algorithm to regularize the learning dynamics of insufficiently annotated domain and classes with the external guidance obtained from reliable sources. Our approach resolves the issue of inadequate navigation along network optimization, which is the major obstacle in the taxonomy adaptive cross-domain adaptation scenario. We evaluate the proposed method extensively on several tasks towards various endpoints with clinical and open-world significance. The results demonstrate its effectiveness and improvements over previous methods. Code is available at https://github.com/camwew/TADA-MI.", + "bbox": [ + 76, + 347, + 473, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 799, + 209, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Automated and objective analysis of medical images is an important research topic and has been explored in various clinical applications [57, 13]. To relieve the burden of acquiring massive annotated data for model development on new domains, several unsupervised domain adapta", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8880daf55c116dc96f9c49849b0c6e7f4a517c78d336058b001a0e3aed7c4d52.jpg", + "image_caption": [ + "Figure 1. (a) Illustration of the issue of taxonomic inconsistency with nuclei recognition. The category label sets are incoherent across datasets. Different colours indicate the class each nucleus belongs to. The red asterisks denote novel classes only existing in Dataset 2. (b) Concept of the proposed method. We perform optimization trajectory distillation to deliver external navigation for learning target domain and new classes where the optimization steps tend to be restricted and unreliable." + ], + "image_footnote": [], + "bbox": [ + 506, + 316, + 890, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion (UDA) methods [7, 38, 37] are proposed to mitigate the data distribution bias [51] between a richly labeled source domain and a target domain with no explicit supervision.", + "bbox": [ + 496, + 685, + 892, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, these works assume a closed-set adaptation setting that the source and target domains should necessarily share the identical category label space and definition [44]. The restriction limits their practical applicability in the clinical wild [15]. Unlike natural images where the definitions of different entity categories (e.g., cat v.s. dog) are structured and globally unified, in the medical domain, inconsistency in taxonomy across different countries or institutes is a common issue compromising the feasibility of cross-domain adaptation [12]. Take nuclei recognition in histology images as an example, the ambiguous biologi", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21174", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cal characteristics of cells and distinct clinical usages of the analysis outcomes result in a lack of a unified categorization schema [15, 20]. With specific clinical purposes and downstream applications, different datasets could categorize nuclei with disparate granularities of biological concepts. Besides, as medical research evolves rapidly, novel cell and disease classes could be discovered and form datasets with continually expanding category label sets [41]. This motivates us to develop a generalized adaptation method with the capability to tackle both data distribution bias and category gap for more flexible UDA in clinical practice.", + "bbox": [ + 80, + 90, + 467, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we study the problem of taxonomy adaptive domain adaptation [19]. It assumes the target domain could adopt a label space different from the source domain. Following UDA, a labeled source dataset and unlabeled samples from the target domain are available during training. Additionally, to recognize the novel/fine-grained categories which are non-existent or unspecified in the source domain, a few samples of those target-private categories are annotated and utilized as exemplars. The technical challenge for this setting lies in how to alleviate domain shifts and concurrently learn new classes with very limited supervision. Recently, several methods are proposed towards a similar goal [32, 22]. However, they typically address the two issues individually in separate contexts and fail to design a unified paradigm according to their common characteristics [19], which could incur a subtle collision across issues since their objectives are non-relevant [60]. Besides, existing works either focus on cross-domain/class alignment in the feature space [40, 36] or resort to model output-based regularization such as self-supervision [45], whereas those approaches suffer from the equilibrium challenge of adversarial learning [1] and the low-quality pseudo-labels [54].", + "bbox": [ + 80, + 260, + 467, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present a unified framework for taxonomy adaptive cross-domain adaptation from a new perspective, i.e., via optimization trajectory distillation. Our strategy is motivated by a common challenge existing in both cross-domain and small-data learning regimes, which is the inadequate navigation in the course of network optimization. For cross-domain learning, the unstable feature alignment procedure and noisy pseudo-labels tend to induce error accumulation along network training [62]. Similarly, with limited support samples in new class learning, the optimization of network is inclined to step towards restricted local minima and cannot cover a globally-generalizable space [55]. To this end, we propose to exploit the optimization trajectory from a reliable \"teacher\" to provide external guidance for regularizing the learning dynamics of insufficiently annotated domain and classes, as illustrated in Fig. 1(b).", + "bbox": [ + 80, + 595, + 467, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method consists of two key components, i.e., cross-domain/class distillation and historical self-distillation. (i) Cross-domain and cross-class distillation aim to leverage the optimization trajectory from the richly labeled do", + "bbox": [ + 80, + 839, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "main and classes to serve as the \"teacher\". Motivated by the Neural Tangent Kernel (NTK) theory [27], we characterize the network optimization trajectory through gradient statistics. Then, given the observation that the subspace spanned by the gradients from most iterations is generally low-rank [35, 2], we design a gradient projection approach to suppress the noisy signals in stochastic gradients and rectify the distillation process. Thereafter constraints are imposed on the gradient statistics of target domain and new classes to calibrate their training dynamics towards domain-invariant and unbiased learning procedure. (ii) Historical self-distillation further drives the optimization paths of model to converge towards flat minima. It is found that the flatness of loss landscapes is strongly related to the model's robustness and generalizability [26], while how to take advantage of the insight to tackle domain shifts and limited supervision is under-explored. We propose to exploit the historical gradients to construct the informative low-rank subspaces and then perform gradient projection to alleviate loss sharpness. It compensates for the intense and out-of-order optimization updates incurred by inadequate regularization and leads to better generalization.", + "bbox": [ + 503, + 90, + 888, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our prime contributions are as follows: (1) We introduce a more generalized cross-domain adaptation paradigm for medical image analysis in which both data distribution bias and category gap exist across the source and target domains. (2) We leverage insights from recent learning theory research and propose a novel dual-stream optimization trajectory distillation method to provide external navigation in network training. We perform theoretical justifications from two perspectives to illustrate the merits of our method. (3) Experiments on various benchmarks validate the effectiveness and robustness of our proposed method and its improvements over existing approaches.", + "bbox": [ + 503, + 422, + 888, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 503, + 618, + 645, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain Adaptation Domain adaptation (DA) aims to mitigate the data distribution bias [16]. The classic semi-supervised/unsupervised DA focuses on the scenario where the target domain has an identical label space to the source domain [47, 14, 33]. The over-restricted precondition cannot suffice the demand of clinical usages [58]. There exist several efforts to take a step further than the closed-set adaptation, such as open-set DA [44] and universal DA [59]. However, the aim of those works is solely to detect the previously unseen classes, instead of learning to separately identify each new class with few supports. Some recent studies suggest to explicitly recognize the target-private categories. [32] performs feature projection and alignment based on the prototypical networks [50]. [19] combines pseudo-labeling and contrastive learning to combat domain bias and label gap. Nevertheless, those works are limited to tackling the two issues individually, which fails to exploit", + "bbox": [ + 503, + 643, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21175", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "their common characteristics and propose a unified solution. Cross-Domain Few-Shot Learning The technical challenge we need to resolve is similar to an emerging research topic, i.e., cross-domain few-shot learning [22]. The mainstream methods for this challenge include self-training [45, 25, 61], contrastive learning [11, 19], feature alignment/augmentation [53, 8, 24], and channel transformation [34, 39]. Those methods either focus on feature-level modulation or turn to output-level self-supervision and finetuning, which could bring instability along optimization [1, 54]. In this work, we propose a novel gradient-based framework to transfer knowledge in both cross-domain and few-shot settings. It implicitly characterizes the information in both feature space and output space.", + "bbox": [ + 75, + 90, + 467, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gradient Manipulation In current deep learning systems where gradient descent serves as the most popular training algorithm, the directions and amplitudes of network optimization steps are embedded in the gradient space [9]. Gradient manipulation refers to directly modulating the optimization gradients for improved learning process [42]. Recently, it is introduced to mitigate the data distribution shift and shows promises in UDA and domain generalization [14, 17, 49]. However, those works are still restricted by the closed-set condition and do not consider the low-rank nature of gradients [35], which make their approaches vulnerable to noises in the small-data learning regime.", + "bbox": [ + 75, + 304, + 467, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Flatness in Optimization In previous machine learning research, the connection between convergence to flat minima in optimization and model generalizability has long been established [23, 31]. Theoretical and empirical studies prove that the flatness of loss landscapes highly correlates with generalization capability [4]. [5] averages the network weights of multiple checkpoints to prompt generalization across domains. Different from their post-hoc averaging approach operating on fixed models, we propose to regularize the intermediate dynamics of network optimization to achieve flat minima.", + "bbox": [ + 75, + 487, + 467, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 669, + 166, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Statement", + "text_level": 1, + "bbox": [ + 76, + 695, + 261, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The taxonomy adaptive cross-domain adaptation problem can be formalized as follows. There exists a labeled source dataset $\\mathcal{D}_s = \\{(\\mathbf{x}_s,\\mathbf{y}_s)\\}$ , an unlabeled target dataset $\\mathcal{D}_t^u = \\{(\\mathbf{x}_t^u)\\}$ , and a few-shot labeled support set from the target domain $\\mathcal{D}_t^l = \\{(\\mathbf{x}_t^l,\\mathbf{y}_t^l)\\}$ . Samples from source and target datasets are drawn from different data distributions, i.e., $\\mathbf{x}_s\\sim \\mathcal{P}_s$ , $\\mathbf{x}_t\\sim \\mathcal{P}_t$ , $\\mathcal{P}_s\\neq \\mathcal{P}_t$ . The label sets of source and target datasets are denoted as $\\mathcal{C}_s$ and $\\mathcal{C}_t$ . Due to the novel/fine-grained categories only existing in the target dataset, we have $\\mathcal{C}_s\\neq \\mathcal{C}_t$ . The shared and target-private new classes are indicated as $\\overline{\\mathcal{C}} = \\mathcal{C}_s\\cap \\mathcal{C}_t$ and $\\bar{\\mathcal{C}} = \\mathcal{C}_t\\backslash \\overline{\\mathcal{C}}$ . We call $\\overline{\\mathcal{C}}$ existing in both datasets as anchor classes since", + "bbox": [ + 75, + 719, + 467, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "they can be used as anchors to facilitate knowledge transfer. The goal is to train a model jointly on $\\mathcal{D}_s, \\mathcal{D}_t^u, \\mathcal{D}_t^l$ which performs inferences on data from $\\mathcal{P}_t$ with labels in $\\mathcal{C}_t$ .", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Characterization of Optimization Trajectory", + "text_level": 1, + "bbox": [ + 498, + 143, + 877, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The technical challenge for the identified problem lies in overcoming both domain shifts and overfitting incurred by scarce training samples. As discussed in Section 1, those issues could lead to unreliable navigation during neural network optimization [18, 10]. Therefore, we propose to distill the optimization trajectory knowledge from well-annotated sources to regularize the training dynamics of target domain and new classes.", + "bbox": [ + 496, + 166, + 890, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, NTK theory [27] has been regarded as a tool to dictate the evolution of neural networks. For a network $f$ parameterized by $\\theta$ , its training dynamics can be described by a kernel $K$ . Given a training set with $N$ samples, the kernel can be defined as an $N \\times N$ matrix with $K(i,j;\\theta) = \\nabla_{\\theta}f(x^{i};\\theta)^{\\top} \\cdot \\nabla_{\\theta}f(x^{j};\\theta)$ , where $x^{i}$ and $x^{j}$ are two input samples. It indicates the strong correlations between the network's optimization trajectory and the gradients w.r.t. its parameters. We thus propose to collect a set of intermediate gradients during training and capture their statistics to characterize the optimization trajectory. Specifically, the second-order Taylor expansion shows that:", + "bbox": [ + 496, + 287, + 890, + 468 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f (x; \\theta) = f (x; \\theta^ {*}) + (\\theta - \\theta^ {*}) ^ {\\top} \\cdot \\nabla_ {\\theta} f (x; \\theta^ {*}) \\\\ + \\frac {1}{2} \\left(\\theta - \\theta^ {*}\\right) ^ {\\top} \\cdot \\nabla_ {\\theta} ^ {2} f (x; \\theta^ {*}) \\cdot \\left(\\theta - \\theta^ {*}\\right) \\tag {1} \\\\ + \\mathcal {O} \\left(\\left\\| \\theta - \\theta^ {*} \\right\\| ^ {2}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 474, + 890, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which motivates us to model the first- and second-order derivatives (Hessian) of $f$ to represent its learning dynamics. We use the mean of gradients to indicate the first-order derivative and follow [28] to approximate the second-order derivative with gradient variances.", + "bbox": [ + 496, + 549, + 890, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Cross-Domain and Cross-Class Distillation", + "text_level": 1, + "bbox": [ + 500, + 631, + 862, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gradient-based optimization algorithms depend on a large number of update steps supervised by sufficient labeled data [46]. In domain-shifted and small-data regimes, those algorithms suffer from noisy update signals and overfitting, without guarantee to converge towards generalizable minima [22]. To this end, we propose to distill the training dynamics from the label-rich source domain and anchor classes to the target domain and new classes for external regularization by leveraging their intermediate gradients. We firstly compute the gradients $\\pmb{g}$ for each domain and class by backpropagating the obtained losses. With a general backbone $\\mathbf{F}$ for feature extraction and a task-specific predictor $\\mathbf{C}$ for sample-/pixel-wise classification, the individual gradients can be expressed as:", + "bbox": [ + 496, + 655, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} := \\left[ \\frac {\\partial \\mathcal {L} (\\mathbf {C} (\\mathbf {F} (x)) , y)}{\\partial \\theta_ {\\mathbf {C} _ {1}}}, \\dots , \\frac {\\partial \\mathcal {L} (\\mathbf {C} (\\mathbf {F} (x)) , y)}{\\partial \\theta_ {\\mathbf {C} _ {w}}} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 869, + 890, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21176", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e755981b6e4ea3bed6166a0384259073797712d738a259dd2a3c521cd3161896.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed method. We devise the cross-domain and cross-class distillation module as well as the historical self-distillation module to transfer the model optimization knowledge from reliable sources by imposing constraints on the gradient descent steps w.r.t. the insufficiently supervised target domain and new classes." + ], + "image_footnote": [], + "bbox": [ + 101, + 93, + 869, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ is the loss function, $w$ indicates the total number of parameters in $\\mathbf{C}$ , $x$ and $y$ are training data and the corresponding label. For unlabeled samples $\\mathbf{x}_t^u$ in the target dataset, online pseudo-labels $\\tilde{\\mathbf{y}}_t^u$ are used for loss calculation. Then, as shown in Fig. 2, we aggregate the individual-level gradients in a memory buffer and retrieve the domain-and class-level gradient statistics to model the global first- and second-order derivatives. According to Section 3.2, for gradients w.r.t. domain/class $\\pi$ collected from a training set with $n_{\\pi}$ samples, we derive their mean and variance as:", + "bbox": [ + 75, + 398, + 472, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} _ {\\pi} ^ {m} = \\frac {1}{n _ {\\pi}} \\sum_ {i = 1} ^ {n _ {\\pi}} \\boldsymbol {g} _ {\\pi} ^ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 547, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} _ {\\pi} ^ {v} = \\frac {1}{n _ {\\pi} - 1} \\sum_ {i = 1} ^ {n _ {\\pi}} \\left(\\boldsymbol {g} _ {\\pi} ^ {i} - \\boldsymbol {g} _ {\\pi} ^ {m}\\right) ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 589, + 468, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Those metrics are measured for each domain and class separately.", + "bbox": [ + 75, + 619, + 468, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the gradients obtained from insufficiently annotated domain and classes are dubious and error-prone [55]. To alleviate the issue, we take inspiration from recent learning theory findings that stochastic gradients along the optimization steps are generally low-rank [35], which indicates that the optimization process is governed by the top gradient eigenspace. Specifically, we devise a gradient projection approach to suppress the noisy model update signals in particular to the target domain and new classes by projecting their gradients on the reliable subspaces identified by the source domain and anchor classes with sufficient supervision. Our approach involves three iterative steps, i.e., subspace construction, gradient projection, and gradient statistics matching. Firstly, to identify the principal eigenspace, we perform Singular Value Decomposition (SVD) on the aggregated gradients from the source domain and anchor classes. We denote gradients collected from source domain, target domain, anchor class, and new", + "bbox": [ + 75, + 650, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "class as $g_s$ , $g_t$ , $g_A$ , and $g_N$ . Given the sets of gradients $G_s = [g_s^1, g_s^2, \\dots, g_s^K]$ and $G_A = [g_A^1, g_A^2, \\dots, g_A^K]$ stored in the memory buffer of volume $K$ , we apply SVD:", + "bbox": [ + 498, + 398, + 892, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {G} _ {s} = \\boldsymbol {U} _ {s} \\boldsymbol {\\Sigma} _ {s} \\boldsymbol {V} _ {s} ^ {\\top}, \\quad \\boldsymbol {G} _ {A} = \\boldsymbol {U} _ {A} \\boldsymbol {\\Sigma} _ {A} \\boldsymbol {V} _ {A} ^ {\\top}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 446, + 890, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $U, V$ , and $\\Sigma$ contain left singular vectors $u^i$ , right singular vector $v^i$ , and non-negative singular values $\\sigma^i$ , respectively. To capture the overall characteristics, we use the gradients averaged along all anchor classes to represent $G_A$ . Next, we adopt the low-rank matrix approximation to select the top- $r$ significant left singular vectors in $U_s$ and $U_A$ based on the following criteria:", + "bbox": [ + 498, + 474, + 892, + 573 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\left(\\boldsymbol {G} _ {s}\\right) ^ {r _ {s}} \\right\\| _ {F} ^ {2} \\geq \\tau \\left\\| \\boldsymbol {G} _ {s} \\right\\| _ {F} ^ {2}, \\quad \\left\\| \\left(\\boldsymbol {G} _ {A}\\right) ^ {r _ {A}} \\right\\| _ {F} ^ {2} \\geq \\tau \\left\\| \\boldsymbol {G} _ {A} \\right\\| _ {F} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 580, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(\\pmb{G})^r = \\sum_{i=1}^{r} \\pmb{\\sigma}^i \\pmb{u}^i \\pmb{v}^{i^\\top}$ , $\\| \\cdot \\|_F$ denotes the Frobenius norm, $\\tau$ is a threshold to ensure most information is preserved and is set to 0.98. The principal subspace and the corresponding projection matrix are thereby constructed as $M_s = [u_s^1, u_s^2, \\dots, u_s^{r_s}]$ and $M_A = [u_A^1, u_A^2, \\dots, u_A^{r_A}]$ . Afterward, all gradients are projected on the identified subspace as shown in Fig. 2:", + "bbox": [ + 498, + 607, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} _ {s} ^ {*} = M _ {s} M _ {s} ^ {\\top} \\boldsymbol {g} _ {s}, \\quad \\boldsymbol {g} _ {t} ^ {*} = M _ {s} M _ {s} ^ {\\top} \\boldsymbol {g} _ {t}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 705, + 890, + 727 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} _ {A} ^ {*} = \\boldsymbol {M} _ {A} \\boldsymbol {M} _ {A} ^ {\\top} \\boldsymbol {g} _ {A}, \\quad \\boldsymbol {g} _ {N} ^ {*} = \\boldsymbol {M} _ {A} \\boldsymbol {M} _ {A} ^ {\\top} \\boldsymbol {g} _ {N}.\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 726, + 821, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then following Eq.(3)(4), we minimize the discrepancy between the statistics (i.e., mean and variance) of projected gradients to distill the learning dynamics from the source domain and anchor classes to the target domain and new classes. The overall training objective can be formulated as:", + "bbox": [ + 496, + 744, + 893, + 828 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta} \\mathcal {L} _ {E R M} + \\lambda \\left\\{\\| \\left(\\boldsymbol {g} _ {s} ^ {m *} - \\boldsymbol {g} _ {t} ^ {m *}\\right) \\| _ {2} ^ {2} + \\| \\left(\\boldsymbol {g} _ {s} ^ {v *} - \\boldsymbol {g} _ {t} ^ {v *}\\right) \\| _ {2} ^ {2} \\right. \\\\ \\left. + \\frac {1}{n _ {\\text {n e w}}} \\sum_ {i = 1} ^ {n _ {\\text {n e w}}} \\left[ \\| \\left(\\boldsymbol {g} _ {A} ^ {m *} - \\boldsymbol {g} _ {N i} ^ {m *}\\right) \\| _ {2} ^ {2} + \\| \\left(\\boldsymbol {g} _ {A} ^ {v *} - \\boldsymbol {g} _ {N i} ^ {v *}\\right) \\| _ {2} ^ {2} \\right] \\right\\}. \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 835, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21177", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\mathcal{L}_{ERM}$ denotes the empirical risk loss, which is implemented with cross-entropy loss for classification and Dice loss [43] for segmentation. $\\lambda$ is a balancing coefficient, $n_{new}$ is the number of new classes in the target domain. It is noted that there exists discrepancy between the anchor and new classes in semantic concepts, which makes feature alignment across classes harmful due to negative transfer [56]. However, aligning class-wise learning dynamics by enforcing the similarity between their gradients could contribute to lower empirical error on new classes. We prove this property theoretically in the supplementary material.", + "bbox": [ + 75, + 90, + 472, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Historical Self-distillation", + "text_level": 1, + "bbox": [ + 76, + 266, + 312, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inferior generalization behaviour is an outstanding challenge for taxonomy adaptive cross-domain adaptation [19]. When domain shifts and unseen classes exist, a trained model could only maintain its performance on test samples sharing multiple similar attributes with the training data but cannot generalize well to the disparate ones [63]. Motivated by the connection between flat minima and generalizability [31], we devise the historical self-distillation module to smooth the optimization path towards a well-generalizable solution by calibrating gradient descent steps.", + "bbox": [ + 75, + 289, + 468, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To reach flat loss landscapes [6], assuming a local loss function $\\psi$ and a network $f$ parameterized by $\\theta$ , instead of minimizing the original loss $\\psi(f(x;\\theta))$ only, the optimization objective should be:", + "bbox": [ + 75, + 440, + 470, + 501 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} - \\log \\int_ {\\left\\| \\theta - \\theta^ {*} \\right\\| \\leq \\Gamma} \\exp (- \\psi (f (x; \\theta^ {*})) - \\gamma \\| \\theta - \\theta^ {*} \\|) d \\theta^ {*}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 512, + 475, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x$ indicates random input samples, $\\Gamma$ is defined as the width of a local parameter valley, $\\gamma$ is a balancing weight. In other words, when $\\theta$ and $\\theta^{*}$ are neighbouring, the following criteria is required to be satisfied:", + "bbox": [ + 75, + 559, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\nabla_ {\\theta} \\psi (f (x; \\theta)) - \\nabla_ {\\theta} \\psi (f (x; \\theta^ {*})) \\right\\| \\leq \\beta \\| \\theta - \\theta^ {*} \\|, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 630, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is a smoothness factor. To this end, we propose to impose regulations on the optimizer steps by enforcing minima with uniformly-distributed gradients. Given that the gradients from most iterations are generally low-rank [35, 2], we exploit past gradients to identify the low-dimensional subspace which indicates the principal gradient distribution of local minima and then project current gradients on the constructed subspace to exclude sharp and noisy signals. Specifically, we collect and store individual gradients for each iteration along the training procedure, denoted as $G_{It} = [g_{It}^{t - T}, g_{It}^{t - T + 1}, \\dots, g_{It}^{t - 2}, g_{It}^{t - 1}]$ , where $t$ is the index of current iteration, $T$ is the size of memory buffer. Thereafter we perform SVD on the set of gradients $G_{It}$ to construct the historical subspace. Similar to Eq.(5)(6)(7), the top- $r$ left singular vectors of the decomposed gradient set are concatenated to formulate the projection matrix", + "bbox": [ + 75, + 659, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$M_{It}$ . Then for upcoming gradients obtained from backpropagation in each iteration, we project the gradients on the identified low-rank subspace and consider them as residuals. The subsequent optimizer step is conducted with the sum of the original and rectified gradients, as shown in Fig. 2. Take vanilla stochastic gradient descent as an example, with mini-batch gradients $\\tilde{g}$ in each iteration, the parameters $\\theta$ of the optimized network are updated by:", + "bbox": [ + 496, + 90, + 890, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta := \\theta - \\frac {1}{\\kappa} \\cdot \\eta \\cdot \\left(\\boldsymbol {M} _ {I t} \\boldsymbol {M} _ {I t} ^ {\\top} + \\kappa\\right) \\cdot \\tilde {\\boldsymbol {g}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 224, + 890, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\eta$ is the learning rate, $\\kappa$ is a trade-off parameter. As training proceeds, the principal subspace and the corresponding projection matrix are periodically renewed by applying SVD on the recently collected gradients. The overall training procedure is summarized in the supplementary material.", + "bbox": [ + 496, + 265, + 890, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Theoretical Analysis", + "text_level": 1, + "bbox": [ + 498, + 368, + 694, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we motivate our proposed method theoretically and provide mathematical insights on its merits.", + "bbox": [ + 496, + 393, + 890, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Joint Characterization of Feature and Output Space While our optimization trajectory distillation approach proposes a novel perspective that differs from existing works imposing regularization terms in feature and model-output space [32, 19], those information is still captured and implicitly modeled in the gradient-based framework. We prove this property for the classification and segmentation tasks under two commonly adopted loss functions. The detailed proof is presented in the supplemental material. It illustrates the superiority of our method as a unified framework that jointly characterizes the feature and output space as well as the learning dynamics.", + "bbox": [ + 496, + 424, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Impacts on Generalization Error In our method, we propose to minimize the discrepancy between the gradient statistics from different domains and classes in the identified principal subspace. We hereby prove its effectiveness towards a tighter generalization error bound on the target domain and new classes [3]. Details can be found in the supplemental material.", + "bbox": [ + 496, + 606, + 890, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 729, + 633, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Experiment Settings", + "text_level": 1, + "bbox": [ + 498, + 755, + 795, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our method and its potential clinical implications, we conduct experiments on five medical image datasets in regard to three important downstream tasks from different clinical scenarios. Extended experiments on more diverse tasks, including radiology and fundus analysis, as well as general visual task, are presented in the supplemental material to substantiate the broader applicability of our method.", + "bbox": [ + 496, + 779, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21178", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0961e345af2002045c55913cbcba95d2b5b708c6bb2408b042c19dcc90f7e1ae.jpg", + "table_caption": [ + "Table 1. Comparison results of our proposed method against other state-of-the-art methods for nuclei segmentation and recognition. The metrics with asterisk (*) are calculated on target-private new classes only. The subscripts denote the standard deviations. The best and second-best results are highlighted in bold and brown, respectively." + ], + "table_footnote": [], + "table_body": "
Methods5-shot10-shot
mF1mF1*mPQmPQ*mF1mF1*mPQmPQ*
Sup-only28.080.8921.460.3612.350.5312.230.4530.900.6625.230.8314.180.2414.940.32
Multi-task [48]33.851.2618.151.0818.580.4210.770.3535.150.9421.290.6219.140.2712.890.34
DANN [16]32.761.4116.521.1018.310.6610.400.5935.761.3523.120.8819.700.6014.140.42
CGDM [14]35.031.2821.040.9919.590.5013.220.3237.131.5724.221.1519.780.7113.900.40
LETR [40]27.781.4415.361.0216.680.5410.030.5134.851.5420.030.7318.180.5911.790.27
FT-CIDA [32]30.510.9216.550.8015.970.399.420.4032.631.2121.080.9517.090.4711.960.49
STARTUP [45]37.030.5822.440.7620.400.3313.820.3040.850.5025.920.9123.370.1916.640.24
DDN [25]37.920.8623.711.0520.620.5713.870.4340.160.7226.941.0822.790.3216.820.50
TSA [34]34.031.1719.551.1118.650.4312.060.3734.271.0522.581.2019.780.6114.240.73
TACS [19]36.731.6522.261.2918.130.8012.570.6239.291.4825.841.1322.270.6815.980.56
Ours40.260.9027.140.9721.780.4914.960.2243.880.5231.430.7824.810.2619.350.33
", + "bbox": [ + 142, + 133, + 828, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Nuclei Segmentation and Recognition Technically, this task can be formalized as a simultaneous instance segmentation and classification problem [21]. We evaluate our method by considering PanNuke [15] and Lizard [20] as the source and target datasets. They contain 481 and 291 visual fields cropped from whole-slide images with 189,744 and 495,179 annotated nuclei, respectively. There are six types of nuclei in Lizard and only two of them overlap with the five classes in PanNuke, which suggests four new classes in the target domain. We employ the widely used HoverNet [21] architecture with a standard ResNet-50 backbone as the base model. For quantitative evaluation, we follow [20] and report the average F1 and PQ (panoptic quality) scores for all classes (denoted as mF1 and mPQ) to provide a comprehensive assessment measuring the accuracy of nuclei detection, segmentation, and classification. Additionally, we also report the mF1* and mPQ* scores which are solely computed on new classes in the target domain.", + "bbox": [ + 75, + 351, + 472, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cancer Tissue Phenotyping In this setting, we perform adaptation from a two-class discrimination task to an eight-class one on CRC-TP [29] and Kather [30] datasets. Kather consists of $5000 \\times 150$ pathology image patches and includes six novel classes other than the two common classes also existing in the source domain. The experiments are conducted with ResNet-101 as backbone. For evaluation, we exploit the accuracy and F1 scores as metrics to indicate the classification performance. The average scores for all classes (mAcc and mF1) and particularly the ones on new classes only (mAcc* and mF1*) are reported.", + "bbox": [ + 75, + 626, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Skin Lesion Diagnosis We construct a fine-grained taxonomy adaptation setting from two coarse classes to seven subclasses with HAM10000 [52], which contains 10015 dermatoscopic images sampled from different body structures. The adopted experiment settings and evaluation metrics are the same as the cancer tissue phenotyping task. Since the label space of the source and target datasets do", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "not overlap, we only report the mAcc* and mF1* scores which are computed on target classes.", + "bbox": [ + 496, + 351, + 890, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous works in UDA [7], for all experiments, each dataset is randomly split into $80\\% / 20\\%$ for training/testing. To explicitly recognize the target-private new classes, we sample few (5/10) samples with corresponding labels to formulate the support set. The remaining target data is left unlabeled. More details can be found in the supplemental material.", + "bbox": [ + 496, + 382, + 892, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 500, + 498, + 591, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with state-of-the-art approaches for cross-domain adaptation, including UDA methods DANN [16] and CGDM [14] which do not consider the taxonomic inconsistency, as well as LETR [40], FT-CIDA[32], STARTUP [45], DDN [25], TSA [34], and TACS [19] which are designed for both domain shifts and new class learning. For UDA methods, we jointly perform domain alignment and train the model with the labeled target data. In addition, we conduct comparisons with multi-task learning approach [48], which trains the model on all domains and label sets simultaneously. \"Sup-only\" indicates the baseline where only the supervised loss on the labeled target dataset is used for training.", + "bbox": [ + 496, + 522, + 890, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows the comparison results on the cross-domain nuclei recognition and segmentation task. All methods adopt the same ResNet-50-HoverNet as base model following [21]. It can be observed that our proposed method outperforms the state-of-the-art cross-domain adaptation approaches by a large margin. We also notice an important point that compared with the \"Sup-only\" baseline, most competing methods fail to attain better performance on the metrics specific to new classes (i.e., mF1* and mPQ*). For example, under the 5-shot setting, despite a $4.68\\%$ improvement is achieved by DANN on mF1, it achieves inferior performance on mF1* which is $4.94\\%$ worse than \"Sup-only\".", + "bbox": [ + 496, + 719, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21179", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0d56a1fdcb44509487ba18d80e2e246a42e8e09a47b34545e404b401ac28822c.jpg", + "table_caption": [ + "Table 2. Comparison results of our proposed method against other state-of-the-art methods for cancer tissue phenotyping." + ], + "table_footnote": [], + "table_body": "
Methods5-shot10-shot
mAccmAcc*mF1mF1*mAccmAcc*mF1mF1*
Sup-only46.821.3042.261.2644.641.4942.831.0848.521.0245.931.6745.341.4546.151.83
Multi-task [48]48.350.7842.021.3445.551.1241.681.6348.560.6945.001.7246.030.6444.461.21
DANN [16]47.460.6742.531.2243.081.1443.131.3048.420.4845.311.2746.500.8144.020.64
CGDM [14]49.020.8243.531.5646.290.8743.580.6950.540.8444.571.3548.170.6645.230.40
LETR [40]45.471.9541.331.4743.171.8841.202.0245.940.5640.131.3041.161.1339.610.90
FT-CIDA [32]43.720.8839.060.7940.911.0238.730.8444.650.5338.800.9440.940.9638.961.07
STARTUP [45]49.520.6342.931.8343.591.3739.970.5949.400.5146.911.8746.380.7047.851.36
DDN [25]47.141.7543.250.9942.922.2940.760.5048.270.6744.081.2246.031.1445.441.18
TSA [34]48.221.4942.130.7844.391.8441.921.1347.181.0241.261.1045.341.7244.691.89
TACS [19]45.301.0740.281.1144.741.2642.040.9546.951.1643.691.7447.461.5147.511.25
Ours50.140.7647.661.5447.411.1046.451.1555.860.9851.271.0252.980.4952.120.94
", + "bbox": [ + 129, + 106, + 841, + 299 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0517a5d6408c1163df0cb619823b3f5bc4c40e816bfc30b9779f8972c73ca6a7.jpg", + "table_caption": [ + "Table 3. Comparison results of our proposed method against other state-of-the-art methods for skin lesion diagnosis." + ], + "table_footnote": [], + "table_body": "
Methods5-shot10-shot
mAcc*mF1*mAcc*mF1*
Sup-only35.710.7919.890.2238.500.3822.360.67
Multi-task [48]33.801.2019.060.3137.850.4023.320.32
DANN [16]37.220.5723.990.9240.060.9424.320.72
CGDM [14]41.501.0324.260.8047.220.7024.000.23
LETR [40]36.040.4221.390.6640.871.2223.500.49
FT-CIDA [32]31.950.7718.130.4834.600.9020.960.88
STARTUP [45]43.211.1527.170.9243.881.3525.590.60
DDN [25]36.191.0521.500.3442.461.2923.920.28
TSA [34]40.411.3624.051.0142.071.5125.441.06
TACS [19]45.050.4625.600.6546.730.6425.960.79
Ours50.790.6930.050.8152.140.9830.520.83
", + "bbox": [ + 93, + 353, + 454, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This phenomenon is incurred by the failure of cross-class generalization. In contrast, our method yields substantial improvements on those new class-specific metrics, which indicates the superior generalizability of our method. The observation is further verified through visual comparisons in the supplementary material.", + "bbox": [ + 75, + 578, + 468, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The overall quantitative results on the cancer tissue phenotyping and skin lesion diagnosis tasks are presented in Table 2 and 3, respectively. All methods are implemented based on the ResNet-101 backbone. Under each few-shot setting, our method shows higher classification scores compared with previous approaches. In particular, we achieve $0.62\\% \\sim 5.32\\%$ and $1.12\\% \\sim 4.81\\%$ improvements against the second-best results in terms of mAcc and mF1 for cancer tissue phenotyping, as well as $4.92\\% \\sim 7.58\\%$ and $2.88\\% \\sim 4.93\\%$ improvements in terms of mAcc* and mF1* for skin lesion diagnosis. The higher accuracy consistently attained by our method under different cross-domain adaptation settings and clinical tasks is attributed to the strong generalizability brought by the proposed optimization trajectory distillation approach.", + "bbox": [ + 75, + 674, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 500, + 323, + 720, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of Key Components Table 4 shows the ablation study results to illustrate the effectiveness of key components. \"projection\" denotes the gradient projection approach introduced in Section 3.3. For the \"w/o projection\" ablation, we perform cross-domain and cross-class distillation by directly minimizing the statistics discrepancy between unrectified gradients $(g_{s}, g_{t})$ and $(g_{A}, g_{N})$ in Eq. (7). It is observed that each component is indispensable and could improve the overall performance, which elaborates the importance of external navigation for learning target domain and new classes in network training. Interestingly, we notice that the impacts of cross-domain and cross-class distillation could be disparate across tasks. For example, cross-class distillation brings higher performance gain compared with cross-domain distillation on the cancer tissue phenotyping benchmark, while the situation is reversed for skin lesion diagnosis. It is caused by the different natures of each task. Class-imbalance is a serious issue for skin lesion diagnosis in that the class distribution is highly biased [52], which compromises the effectiveness of the cross-class distillation module. Differently, for cancer tissue phenotyping, classes are uniformly distributed [30] and could hence ensure the representativity of anchor classes and their exemplarity to be distilled to new classes.", + "bbox": [ + 496, + 351, + 890, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameter Sensitivity To investigate the effect of hyperparameters $\\lambda$ in Eq. (8) and $\\kappa$ in Eq. (11), we present the performance comparison by varying the choices of their values for cancer tissue phenotyping, as shown in Fig. 3. We only change one hyperparameter while keeping the other fixed at a time. The experimental results show that our framework is quite stable for hyperparameters in a wide interval, yet setting $\\lambda$ and $\\kappa$ to a large value is detrimental to the classification accuracy. Specifically, $\\lambda$ is a balancing coefficient between class discriminativeness and gradient alignment. When it is too large, the trained model is biased to matching learning dynamics and would lose es", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21180", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5940c44db642d070a68f39acef54e9d91bbb02109ccd4a9270e4e82f967130d4.jpg", + "table_caption": [ + "Table 4. Key component analysis of our method on three cross-domain adaptation benchmarks under 10-shot scheme. The best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodsNucleiCancerSkin
mF1mF1*mPQmPQ*mAccmAcc*mF1mF1*mAcc*mF1*
Full Framework43.8831.4324.8119.3555.8651.2752.9852.1252.1430.52
w/o cross-domain39.9630.0822.2918.1052.8149.9351.6250.6444.3425.28
w/o cross-class42.3027.8223.7617.0548.9047.2249.8544.7649.0628.24
w/o historical41.5430.4723.0017.5950.1748.0548.1945.1245.8325.97
w/o projection40.0628.8022.1017.3153.5845.7849.0149.2047.2928.02
", + "bbox": [ + 178, + 119, + 792, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d34469907c41fdddcf838ca7bce695303cbcbba49b1bdf427902020e289eccf9.jpg", + "image_caption": [ + "Figure 3. Performance comparison between different choices of hyperparameters $\\lambda$ and $\\kappa$ on the cancer tissue phenotyping benchmark under 10-shot scheme." + ], + "image_footnote": [], + "bbox": [ + 86, + 252, + 459, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sential semantic knowledge. For $\\kappa$ , it controls the proportions of unrectified gradients. Its large value indicates that most original gradients are preserved, which would incur noisy optimization steps and loss of flatness. Therefore, we decide to set $\\lambda = 10$ and $\\kappa = 100$ . It achieves the best performance under most metrics.", + "bbox": [ + 75, + 443, + 468, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impacts on Optimization Trajectory and Model Flatness To better understand the effectiveness of our method, we perform in-depth analysis about its impacts on model's training trajectory and flatness of the converged local minima. Firstly, as shown in Fig. 4(a), we visualize the optimization paths of our method and the multi-task baseline on the test accuracy surface with a contour plot. The details of plot procedure can refer to [26]. It can be observed that the optimization of the multi-task baseline fluctuates sharply and is restricted in a sub-optimal area with a relatively low test accuracy. This is induced by the biased navigation in the course of training due to domain shifts and category gap. With the devised dual-stream optimization trajectory distillation module, our method could by contrast suppress the noisy update signals and proceed consistently towards the optimum.", + "bbox": [ + 75, + 536, + 470, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides, we perform analysis about the solutions found by our method and the baseline in terms of flatness [5]. Specifically, the local flatness of a model parameter $\\theta$ is quantified by the expected changes of test accuracy between $\\theta$ and a neighbouring $\\theta^{*}$ with perturbation rate $\\varrho$ : $\\mathcal{F}_{\\varrho}(\\theta) = \\mathbb{E}_{\\|\\theta^{*}\\| = (1 + \\varrho)\\|\\theta\\|}\\left[\\mathrm{Acc}(\\theta^{*}) - \\mathrm{Acc}(\\theta)\\right]$ . We approximate the expected value by Monte-Carlo sampling over 50 samples. As demonstrated in Fig. 4(b), comparing with the", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/14d88eb8312627df860f13a9fe26421d4d73df2681158058767a931e55e216bf.jpg", + "image_caption": [ + "(a) Visualization of Optimization Trajectory" + ], + "image_footnote": [], + "bbox": [ + 504, + 250, + 705, + 398 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3014a8f72be3ae2f65d861f70aa9c8d2a8a4e91eaeb92f6de3b179fc94c3f87a.jpg", + "image_caption": [ + "(b) Flatness Analysis", + "Figure 4. Further analysis of our proposed method. (a) We plot the optimization trajectory of our method and the multi-task baseline on the test accuracy surface. (b) We visualize the local flatness of our method and the baseline via test accuracy gap by varying the model parameter perturbation rate. Each point is computed by Monte-Carlo approximation over 50 random samples. All the experiments are conducted on the cancer tissue phenotyping benchmark under 10-shot scheme." + ], + "image_footnote": [], + "bbox": [ + 707, + 250, + 888, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "baseline approach, our method shows stronger robustness against model parameter perturbation. It proves that our method successfully converges to minima with better local flatness and thus possesses superior generalizability [31].", + "bbox": [ + 496, + 556, + 890, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 633, + 617, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we study the taxonomy adaptive cross-domain adaptation paradigm, which allows incoherent label sets between source and target datasets. To jointly address the data distribution bias and category gap, we propose a unified framework which performs cross-domain and cross-class optimization trajectory distillation to calibrate the learning dynamics of insufficiently annotated domain and classes. Historical self-distillation is further devised to attain a well-generalizable solution via regulating gradient distributions. Extensive experiments on multiple benchmarks with different task contexts demonstrate the effectiveness and generalization capability of our method. In the future, we will extend this work to further discover the underlying mechanism of optimization trajectory distillation for model generalization and evaluate its potential in more adaptation scenarios.", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21181", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sanjeev Arora, Rong Ge, Yingyu Liang, Tengyu Ma, and Yi Zhang. Generalization and equilibrium in generative adversarial nets (gans). In International Conference on Machine Learning, pages 224-232. PMLR, 2017.", + "[2] Sheikh Shams Azam, Seyyedali Hosseinalipour, Qiang Qiu, and Christopher Brinton. Recycling model updates in federated learning: Are gradient subspaces low-rank? In International Conference on Learning Representations, 2022.", + "[3] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006.", + "[4] Debora Calderola, Barbara Caputo, and Marco Ciccone. Improving generalization in federated learning by seeking flat minima. In European Conference on Computer Vision, pages 654-672. Springer, 2022.", + "[5] Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Seunghyun Park, Yunsung Lee, and Sungrae Park. Swad: Domain generalization by seeking flat minima. Advances in Neural Information Processing Systems, 34:22405-22418, 2021.", + "[6] Pratik Chaudhari, Anna Choromanska, Stefano Soatto, Yann LeCun, Carlo Baldassi, Christian Borgs, Jennifer Chayes, Levent Sagun, and Riccardo Zecchina. Entropy-SGD: Biasing gradient descent into wide valleys. In International Conference on Learning Representations, 2017.", + "[7] Cheng Chen, Qi Dou, Hao Chen, Jing Qin, and Pheng-Ann Heng. Synergistic image and feature adaptation: Towards cross-modality domain adaptation for medical image segmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 865–872, 2019.", + "[8] Wentao Chen, Zhang Zhang, Wei Wang, Liang Wang, Zilei Wang, and Tieniu Tan. Cross-domain cross-set few-shot learning via learning compact and aligned representations. In European Conference on Computer Vision, pages 383-399. Springer, 2022.", + "[9] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018.", + "[10] Ali Cheraghian, Shafin Rahman, Sameera Ramasinghe, Pengfei Fang, Christian Simon, Lars Petersson, and Mehrtash Harandi. Synthesized feature based few-shot class incremental learning on a mixture of subspaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8661-8670, 2021.", + "[11] Debasmit Das, Sungrack Yun, and Fatih Porikli. ConfeSS: A framework for single source cross-domain few-shot learning. In International Conference on Learning Representations, 2022.", + "[12] Susanna Dodd, Mike Clarke, Lorne Becker, Chris Mavergames, Rebecca Fish, and Paula R Williamson. A taxonomy has been developed for outcomes in medical research to help improve knowledge discovery. Journal of clinical epidemiology, 96:84-92, 2018." + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] C Dominguez Conde, C Xu, LB Jarvis, DB Rainbow, SB Wells, T Gomes, SK Howlett, O Suchanek, K Polanski, HW King, et al. Cross-tissue immune cell analysis reveals tissue-specific features in humans. Science, 376(6594), 2022.", + "[14] Zhekai Du, Jingjing Li, Hongzu Su, Lei Zhu, and Ke Lu. Cross-domain gradient discrepancy minimization for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3937-3946, 2021.", + "[15] Jevgenij Gamper, Navid Alemi Koohbanani, Ksenija Benet, Ali Khuram, and Nasir Rajpoot. Pannuke: an open pancancer histology dataset for nuclei instance segmentation and classification. In European congress on digital pathology, pages 11-19. Springer, 2019.", + "[16] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016.", + "[17] Zhiqiang Gao, Shufei Zhang, Kaizhu Huang, Qiufeng Wang, and Chaoliang Zhong. Gradient distribution alignment certificates better adversarial domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8937-8946, 2021.", + "[18] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020.", + "[19] Rui Gong, Martin Danelljan, Dengxin Dai, Danda Pani Paudel, Ajad Chhatkuli, Fisher Yu, and Luc Van Gool. Tacs: Taxonomy adaptive cross-domain semantic segmentation. In European Conference on Computer Vision, pages 19-35. Springer, 2022.", + "[20] Simon Graham, Mostafa Jahanifar, Ayesha Azam, Mohammed Nimir, Yee-Wah Tsang, Katherine Dodd, Emily Hero, Harvir Sahota, Atisha Tank, Ksenija Benes, et al. Lizard: a large-scale dataset for colonic nuclear instance segmentation and classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 684-693, 2021.", + "[21] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:101563, 2019.", + "[22] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In European conference on computer vision, pages 124-141. Springer, 2020.", + "[23] Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural computation, 9(1):1-42, 1997.", + "[24] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In European Conference on Computer Vision, pages 20-37. Springer, 2022.", + "[25] Ashraful Islam, Chun-Fu Richard Chen, Rameswar Panda, Leonid Karlinsky, Rogerio Feris, and Richard J Radke. Dy" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21182", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "namic distillation network for cross-domain few-shot recognition with unlabeled data. Advances in Neural Information Processing Systems, 34:3584-3595, 2021.", + "[26] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. Proceedings of the Conference on Uncertainty in Artificial Intelligence, 2018.", + "[27] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018.", + "[28] Stanisław Jastrzebski, Zachary Kenton, Devansh Arpit, Nicolas Ballas, Asja Fischer, Yoshua Bengio, and Amos Storkey. Three factors influencing minima in sgd. arXiv preprint arXiv:1711.04623, 2017.", + "[29] Sajid Javed, Arif Mahmood, Muhammad Moazam Fraz, Navid Alemi Koohbanani, Ksenija Benes, Yee-Wah Tsang, Katherine Hewitt, David Epstein, David Snead, and Nasir Rajpoot. Cellular community detection for tissue phenotyping in colorectal cancer histology images. Medical image analysis, 63:101696, 2020.", + "[30] Jakob Nikolas Kather, Cleo-Aron Weis, Francesco Bianconi, Susanne M Melchers, Lothar R Schad, Timo Gaiser, Alexander Marx, and Frank Gerrit Zöllner. Multi-class texture analysis in colorectal cancer histology. Scientific reports, 6(1):1-11, 2016.", + "[31] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. In International Conference on Learning Representations, 2017.", + "[32] Jogendra Nath Kundu, Rahul Mysore Venkatesh, Naveen Venkat, Ambareesh Revanur, and R Venkatesh Babu. Class incremental domain adaptation. In European Conference on Computer Vision, pages 53-69. Springer, 2020.", + "[33] Canran Li, Dongnan Liu, Haoran Li, Zheng Zhang, Guangming Lu, Xiaojun Chang, and Weidong Cai. Domain adaptive nuclei instance segmentation and classification via category-aware feature alignment and pseudo-labelling. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 715-724. Springer, 2022.", + "[34] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Cross-domain few-shot learning with task-specific adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7161-7170, 2022.", + "[35] Xinyan Li, Qilong Gu, Yingxue Zhou, Tiancong Chen, and Arindam Banerjee. Hessian based analysis of sgd for deep nets: Dynamics and generalization. In Proceedings of the 2020 SIAM International Conference on Data Mining, pages 190-198. SIAM, 2020.", + "[36] Hanwen Liang, Qiong Zhang, Peng Dai, and Juwei Lu. Boosting the generalization capability in cross-domain few-shot learning via noise-enhanced supervised autoencoder. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9424-9434, 2021." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Dongnan Liu, Chaoyi Zhang, Yang Song, Heng Huang, Chenyu Wang, Michael Barnett, and Weidong Cai. Decompose to adapt: Cross-domain object detection via feature disentanglement. IEEE Transactions on Multimedia, 25:1333-1344, 2022.", + "[38] Dongnan Liu, Donghao Zhang, Yang Song, Fan Zhang, Lauren O'Donnell, Heng Huang, Mei Chen, and Weidong Cai. Unsupervised instance segmentation in microscopy images via panoptic domain adaptation and task re-weighting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4243-4252, 2020.", + "[39] Xu Luo, Jing Xu, and Zenglin Xu. Channel importance matters in few-shot image classification. In International Conference on Machine Learning, pages 14542-14559. PMLR, 2022.", + "[40] Zelun Luo, Yuliang Zou, Judy Hoffman, and Li F Fei-Fei. Label efficient learning of transferable representations acrossss domains and tasks. Advances in neural information processing systems, 30, 2017.", + "[41] Deborah Lupton. Medicine as culture: illness, disease and the body. Sage, 2012.", + "[42] Lucas Mansilla, Rodrigo Echeveste, Diego H Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6630-6638, 2021.", + "[43] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016.", + "[44] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In Proceedings of the IEEE international conference on computer vision, pages 754-763, 2017.", + "[45] Cheng Perng Phoo and Bharath Hariharan. Self-training for few-shot transfer across extreme task differences. In International Conference on Learning Representations, 2021.", + "[46] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations, 2017.", + "[47] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, Trevor Darell, and Kate Saenko. Semi-supervised domain adaptation via minimax entropy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8050-8058, 2019.", + "[48] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018.", + "[49] Yuge Shi, Jeffrey Seely, Philip Torr, Siddharth N, Awni Hannun, Nicolas Usunier, and Gabriel Synnaeve. Gradient matching for domain generalization. In International Conference on Learning Representations, 2022.", + "[50] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017.", + "[51] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011.", + "[52] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source der" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21183", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "matoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018.", + "[53] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In International Conference on Learning Representations, 2020.", + "[54] Qian Wang and Toby Breckon. Unsupervised domain adaptation via structured prediction based selective pseudolabeling. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 6243-6250, 2020.", + "[55] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020.", + "[56] Zirui Wang, Zihang Dai, Barnabás Póczos, and Jaime Carbonell. Characterizing and avoiding negative transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11293-11302, 2019.", + "[57] Julia K Winkler, Christine Fink, Ferdinand Toberer, Alexander Enk, Teresa Deinlein, Rainer Hofmann-Wellenhof, Luc Thomas, Aimilios Lallas, Andreas Blum, Wilhelm Stolz, et al. Association between surgical skin markings in dermoscopic images and diagnostic performance of a deep learning convolutional neural network for melanoma recognition. JAMA dermatology, 155(10):1135-1141, 2019.", + "[58] Mengya Xu, Mobarakol Islam, Chwee Ming Lim, and Hongliang Ren. Class-incremental domain adaptation with smoothing and calibration for surgical report generation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 269-278. Springer, 2021.", + "[59] Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Universal domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2720-2729, 2019.", + "[60] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020.", + "[61] Wang Yuan, Zhizhong Zhang, Cong Wang, Haichuan Song, Yuan Xie, and Lizhuang Ma. Task-level self-supervision for cross-domain few-shot learning. Proceedings of the AAAI Conference on Artificial Intelligence, 36(3):3215-3223, 2022.", + "[62] Qiming Zhang, Jing Zhang, Wei Liu, and Dacheng Tao. Category anchor-guided unsupervised domain adaptation for semantic segmentation. Advances in neural information processing systems, 32, 2019.", + "[63] Lin Zuo, Mengmeng Jing, Jingjing Li, Lei Zhu, Ke Lu, and Yang Yang. Challenging tough samples in unsupervised domain adaptation. Pattern Recognition, 110:107540, 2021." + ], + "bbox": [ + 78, + 92, + 468, + 797 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21184", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/1f4cb2e9-7a0d-4b5a-99a2-329016955f96_model.json b/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/1f4cb2e9-7a0d-4b5a-99a2-329016955f96_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9a91b5e3f2a42a457454c37587e2937d23e1afa5 --- /dev/null +++ b/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/1f4cb2e9-7a0d-4b5a-99a2-329016955f96_model.json @@ -0,0 +1,2092 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.131, + 0.842, + 0.177 + ], + "angle": 0, + "content": "Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.198, + 0.832, + 0.218 + ], + "angle": 0, + "content": "Jianan Fan\\(^{1}\\), Dongnan Liu\\(^{1}\\), Hang Chang\\(^{2}\\), Heng Huang\\(^{3}\\), Mei Chen\\(^{4}\\), and Weidong Cai\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.227, + 0.882, + 0.245 + ], + "angle": 0, + "content": "\\(^{1}\\)University of Sydney \\(^{2}\\)Lawrence Berkeley National Laboratory \\(^{3}\\)University of Maryland at College Park \\(^{4}\\)Microsoft" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.245, + 0.805, + 0.258 + ], + "angle": 0, + "content": "jfan6480@uni.sydney.edu.au dongnan.liu@sydney.edu.au hchang@lbl.gov" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.258, + 0.799, + 0.273 + ], + "angle": 0, + "content": "henghuanghh@gmail.com Mei.Chen@microsoft.com tom.cai@sydney.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.317, + 0.314, + 0.332 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.348, + 0.474, + 0.773 + ], + "angle": 0, + "content": "The success of automated medical image analysis depends on large-scale and expert-annotated training sets. Unsupervised domain adaptation (UDA) has been raised as a promising approach to alleviate the burden of labeled data collection. However, they generally operate under the closed-set adaptation setting assuming an identical label set between the source and target domains, which is over-restrictive in clinical practice where new classes commonly exist across datasets due to taxonomic inconsistency. While several methods have been presented to tackle both domain shifts and incoherent label sets, none of them take into account the common characteristics of the two issues and consider the learning dynamics along network training. In this work, we propose optimization trajectory distillation, a unified approach to address the two technical challenges from a new perspective. It exploits the low-rank nature of gradient space and devises a dual-stream distillation algorithm to regularize the learning dynamics of insufficiently annotated domain and classes with the external guidance obtained from reliable sources. Our approach resolves the issue of inadequate navigation along network optimization, which is the major obstacle in the taxonomy adaptive cross-domain adaptation scenario. We evaluate the proposed method extensively on several tasks towards various endpoints with clinical and open-world significance. The results demonstrate its effectiveness and improvements over previous methods. Code is available at https://github.com/camwew/TADA-MI." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.8, + 0.21, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Automated and objective analysis of medical images is an important research topic and has been explored in various clinical applications [57, 13]. To relieve the burden of acquiring massive annotated data for model development on new domains, several unsupervised domain adapta" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.317, + 0.891, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.544, + 0.895, + 0.656 + ], + "angle": 0, + "content": "Figure 1. (a) Illustration of the issue of taxonomic inconsistency with nuclei recognition. The category label sets are incoherent across datasets. Different colours indicate the class each nucleus belongs to. The red asterisks denote novel classes only existing in Dataset 2. (b) Concept of the proposed method. We perform optimization trajectory distillation to deliver external navigation for learning target domain and new classes where the optimization steps tend to be restricted and unreliable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.893, + 0.734 + ], + "angle": 0, + "content": "tion (UDA) methods [7, 38, 37] are proposed to mitigate the data distribution bias [51] between a richly labeled source domain and a target domain with no explicit supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "However, these works assume a closed-set adaptation setting that the source and target domains should necessarily share the identical category label space and definition [44]. The restriction limits their practical applicability in the clinical wild [15]. Unlike natural images where the definitions of different entity categories (e.g., cat v.s. dog) are structured and globally unified, in the medical domain, inconsistency in taxonomy across different countries or institutes is a common issue compromising the feasibility of cross-domain adaptation [12]. Take nuclei recognition in histology images as an example, the ambiguous biologi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21174" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.091, + 0.468, + 0.257 + ], + "angle": 0, + "content": "cal characteristics of cells and distinct clinical usages of the analysis outcomes result in a lack of a unified categorization schema [15, 20]. With specific clinical purposes and downstream applications, different datasets could categorize nuclei with disparate granularities of biological concepts. Besides, as medical research evolves rapidly, novel cell and disease classes could be discovered and form datasets with continually expanding category label sets [41]. This motivates us to develop a generalized adaptation method with the capability to tackle both data distribution bias and category gap for more flexible UDA in clinical practice." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.261, + 0.468, + 0.593 + ], + "angle": 0, + "content": "To this end, we study the problem of taxonomy adaptive domain adaptation [19]. It assumes the target domain could adopt a label space different from the source domain. Following UDA, a labeled source dataset and unlabeled samples from the target domain are available during training. Additionally, to recognize the novel/fine-grained categories which are non-existent or unspecified in the source domain, a few samples of those target-private categories are annotated and utilized as exemplars. The technical challenge for this setting lies in how to alleviate domain shifts and concurrently learn new classes with very limited supervision. Recently, several methods are proposed towards a similar goal [32, 22]. However, they typically address the two issues individually in separate contexts and fail to design a unified paradigm according to their common characteristics [19], which could incur a subtle collision across issues since their objectives are non-relevant [60]. Besides, existing works either focus on cross-domain/class alignment in the feature space [40, 36] or resort to model output-based regularization such as self-supervision [45], whereas those approaches suffer from the equilibrium challenge of adversarial learning [1] and the low-quality pseudo-labels [54]." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.596, + 0.468, + 0.837 + ], + "angle": 0, + "content": "In this work, we present a unified framework for taxonomy adaptive cross-domain adaptation from a new perspective, i.e., via optimization trajectory distillation. Our strategy is motivated by a common challenge existing in both cross-domain and small-data learning regimes, which is the inadequate navigation in the course of network optimization. For cross-domain learning, the unstable feature alignment procedure and noisy pseudo-labels tend to induce error accumulation along network training [62]. Similarly, with limited support samples in new class learning, the optimization of network is inclined to step towards restricted local minima and cannot cover a globally-generalizable space [55]. To this end, we propose to exploit the optimization trajectory from a reliable \"teacher\" to provide external guidance for regularizing the learning dynamics of insufficiently annotated domain and classes, as illustrated in Fig. 1(b)." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.84, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Our method consists of two key components, i.e., cross-domain/class distillation and historical self-distillation. (i) Cross-domain and cross-class distillation aim to leverage the optimization trajectory from the richly labeled do" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.091, + 0.89, + 0.422 + ], + "angle": 0, + "content": "main and classes to serve as the \"teacher\". Motivated by the Neural Tangent Kernel (NTK) theory [27], we characterize the network optimization trajectory through gradient statistics. Then, given the observation that the subspace spanned by the gradients from most iterations is generally low-rank [35, 2], we design a gradient projection approach to suppress the noisy signals in stochastic gradients and rectify the distillation process. Thereafter constraints are imposed on the gradient statistics of target domain and new classes to calibrate their training dynamics towards domain-invariant and unbiased learning procedure. (ii) Historical self-distillation further drives the optimization paths of model to converge towards flat minima. It is found that the flatness of loss landscapes is strongly related to the model's robustness and generalizability [26], while how to take advantage of the insight to tackle domain shifts and limited supervision is under-explored. We propose to exploit the historical gradients to construct the informative low-rank subspaces and then perform gradient projection to alleviate loss sharpness. It compensates for the intense and out-of-order optimization updates incurred by inadequate regularization and leads to better generalization." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.424, + 0.89, + 0.604 + ], + "angle": 0, + "content": "Our prime contributions are as follows: (1) We introduce a more generalized cross-domain adaptation paradigm for medical image analysis in which both data distribution bias and category gap exist across the source and target domains. (2) We leverage insights from recent learning theory research and propose a novel dual-stream optimization trajectory distillation method to provide external navigation in network training. We perform theoretical justifications from two perspectives to illustrate the merits of our method. (3) Experiments on various benchmarks validate the effectiveness and robustness of our proposed method and its improvements over existing approaches." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.619, + 0.647, + 0.634 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.644, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Domain Adaptation Domain adaptation (DA) aims to mitigate the data distribution bias [16]. The classic semi-supervised/unsupervised DA focuses on the scenario where the target domain has an identical label space to the source domain [47, 14, 33]. The over-restricted precondition cannot suffice the demand of clinical usages [58]. There exist several efforts to take a step further than the closed-set adaptation, such as open-set DA [44] and universal DA [59]. However, the aim of those works is solely to detect the previously unseen classes, instead of learning to separately identify each new class with few supports. Some recent studies suggest to explicitly recognize the target-private categories. [32] performs feature projection and alignment based on the prototypical networks [50]. [19] combines pseudo-labeling and contrastive learning to combat domain bias and label gap. Nevertheless, those works are limited to tackling the two issues individually, which fails to exploit" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21175" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.304 + ], + "angle": 0, + "content": "their common characteristics and propose a unified solution. Cross-Domain Few-Shot Learning The technical challenge we need to resolve is similar to an emerging research topic, i.e., cross-domain few-shot learning [22]. The mainstream methods for this challenge include self-training [45, 25, 61], contrastive learning [11, 19], feature alignment/augmentation [53, 8, 24], and channel transformation [34, 39]. Those methods either focus on feature-level modulation or turn to output-level self-supervision and finetuning, which could bring instability along optimization [1, 54]. In this work, we propose a novel gradient-based framework to transfer knowledge in both cross-domain and few-shot settings. It implicitly characterizes the information in both feature space and output space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.468, + 0.486 + ], + "angle": 0, + "content": "Gradient Manipulation In current deep learning systems where gradient descent serves as the most popular training algorithm, the directions and amplitudes of network optimization steps are embedded in the gradient space [9]. Gradient manipulation refers to directly modulating the optimization gradients for improved learning process [42]. Recently, it is introduced to mitigate the data distribution shift and shows promises in UDA and domain generalization [14, 17, 49]. However, those works are still restricted by the closed-set condition and do not consider the low-rank nature of gradients [35], which make their approaches vulnerable to noises in the small-data learning regime." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.468, + 0.651 + ], + "angle": 0, + "content": "Flatness in Optimization In previous machine learning research, the connection between convergence to flat minima in optimization and model generalizability has long been established [23, 31]. Theoretical and empirical studies prove that the flatness of loss landscapes highly correlates with generalization capability [4]. [5] averages the network weights of multiple checkpoints to prompt generalization across domains. Different from their post-hoc averaging approach operating on fixed models, we propose to regularize the intermediate dynamics of network optimization to achieve flat minima." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.67, + 0.167, + 0.685 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.262, + 0.71 + ], + "angle": 0, + "content": "3.1. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.9 + ], + "angle": 0, + "content": "The taxonomy adaptive cross-domain adaptation problem can be formalized as follows. There exists a labeled source dataset \\(\\mathcal{D}_s = \\{(\\mathbf{x}_s,\\mathbf{y}_s)\\}\\), an unlabeled target dataset \\(\\mathcal{D}_t^u = \\{(\\mathbf{x}_t^u)\\}\\), and a few-shot labeled support set from the target domain \\(\\mathcal{D}_t^l = \\{(\\mathbf{x}_t^l,\\mathbf{y}_t^l)\\}\\). Samples from source and target datasets are drawn from different data distributions, i.e., \\(\\mathbf{x}_s\\sim \\mathcal{P}_s\\), \\(\\mathbf{x}_t\\sim \\mathcal{P}_t\\), \\(\\mathcal{P}_s\\neq \\mathcal{P}_t\\). The label sets of source and target datasets are denoted as \\(\\mathcal{C}_s\\) and \\(\\mathcal{C}_t\\). Due to the novel/fine-grained categories only existing in the target dataset, we have \\(\\mathcal{C}_s\\neq \\mathcal{C}_t\\). The shared and target-private new classes are indicated as \\(\\overline{\\mathcal{C}} = \\mathcal{C}_s\\cap \\mathcal{C}_t\\) and \\(\\bar{\\mathcal{C}} = \\mathcal{C}_t\\backslash \\overline{\\mathcal{C}}\\). We call \\(\\overline{\\mathcal{C}}\\) existing in both datasets as anchor classes since" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "they can be used as anchors to facilitate knowledge transfer. The goal is to train a model jointly on \\(\\mathcal{D}_s, \\mathcal{D}_t^u, \\mathcal{D}_t^l\\) which performs inferences on data from \\(\\mathcal{P}_t\\) with labels in \\(\\mathcal{C}_t\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.144, + 0.878, + 0.16 + ], + "angle": 0, + "content": "3.2. Characterization of Optimization Trajectory" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.167, + 0.892, + 0.287 + ], + "angle": 0, + "content": "The technical challenge for the identified problem lies in overcoming both domain shifts and overfitting incurred by scarce training samples. As discussed in Section 1, those issues could lead to unreliable navigation during neural network optimization [18, 10]. Therefore, we propose to distill the optimization trajectory knowledge from well-annotated sources to regularize the training dynamics of target domain and new classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Recently, NTK theory [27] has been regarded as a tool to dictate the evolution of neural networks. For a network \\( f \\) parameterized by \\( \\theta \\), its training dynamics can be described by a kernel \\( K \\). Given a training set with \\( N \\) samples, the kernel can be defined as an \\( N \\times N \\) matrix with \\( K(i,j;\\theta) = \\nabla_{\\theta}f(x^{i};\\theta)^{\\top} \\cdot \\nabla_{\\theta}f(x^{j};\\theta) \\), where \\( x^{i} \\) and \\( x^{j} \\) are two input samples. It indicates the strong correlations between the network's optimization trajectory and the gradients w.r.t. its parameters. We thus propose to collect a set of intermediate gradients during training and capture their statistics to characterize the optimization trajectory. Specifically, the second-order Taylor expansion shows that:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.476, + 0.891, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f (x; \\theta) = f (x; \\theta^ {*}) + (\\theta - \\theta^ {*}) ^ {\\top} \\cdot \\nabla_ {\\theta} f (x; \\theta^ {*}) \\\\ + \\frac {1}{2} \\left(\\theta - \\theta^ {*}\\right) ^ {\\top} \\cdot \\nabla_ {\\theta} ^ {2} f (x; \\theta^ {*}) \\cdot \\left(\\theta - \\theta^ {*}\\right) \\tag {1} \\\\ + \\mathcal {O} \\left(\\left\\| \\theta - \\theta^ {*} \\right\\| ^ {2}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.624 + ], + "angle": 0, + "content": "which motivates us to model the first- and second-order derivatives (Hessian) of \\( f \\) to represent its learning dynamics. We use the mean of gradients to indicate the first-order derivative and follow [28] to approximate the second-order derivative with gradient variances." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.632, + 0.864, + 0.646 + ], + "angle": 0, + "content": "3.3. Cross-Domain and Cross-Class Distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Gradient-based optimization algorithms depend on a large number of update steps supervised by sufficient labeled data [46]. In domain-shifted and small-data regimes, those algorithms suffer from noisy update signals and overfitting, without guarantee to converge towards generalizable minima [22]. To this end, we propose to distill the training dynamics from the label-rich source domain and anchor classes to the target domain and new classes for external regularization by leveraging their intermediate gradients. We firstly compute the gradients \\( \\pmb{g} \\) for each domain and class by backpropagating the obtained losses. With a general backbone \\( \\mathbf{F} \\) for feature extraction and a task-specific predictor \\( \\mathbf{C} \\) for sample-/pixel-wise classification, the individual gradients can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.871, + 0.891, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} := \\left[ \\frac {\\partial \\mathcal {L} (\\mathbf {C} (\\mathbf {F} (x)) , y)}{\\partial \\theta_ {\\mathbf {C} _ {1}}}, \\dots , \\frac {\\partial \\mathcal {L} (\\mathbf {C} (\\mathbf {F} (x)) , y)}{\\partial \\theta_ {\\mathbf {C} _ {w}}} \\right], \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21176" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.094, + 0.87, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.331, + 0.895, + 0.377 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed method. We devise the cross-domain and cross-class distillation module as well as the historical self-distillation module to transfer the model optimization knowledge from reliable sources by imposing constraints on the gradient descent steps w.r.t. the insufficiently supervised target domain and new classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.473, + 0.538 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) is the loss function, \\(w\\) indicates the total number of parameters in \\(\\mathbf{C}\\), \\(x\\) and \\(y\\) are training data and the corresponding label. For unlabeled samples \\(\\mathbf{x}_t^u\\) in the target dataset, online pseudo-labels \\(\\tilde{\\mathbf{y}}_t^u\\) are used for loss calculation. Then, as shown in Fig. 2, we aggregate the individual-level gradients in a memory buffer and retrieve the domain-and class-level gradient statistics to model the global first- and second-order derivatives. According to Section 3.2, for gradients w.r.t. domain/class \\(\\pi\\) collected from a training set with \\(n_{\\pi}\\) samples, we derive their mean and variance as:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.549, + 0.47, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} _ {\\pi} ^ {m} = \\frac {1}{n _ {\\pi}} \\sum_ {i = 1} ^ {n _ {\\pi}} \\boldsymbol {g} _ {\\pi} ^ {i}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.59, + 0.47, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} _ {\\pi} ^ {v} = \\frac {1}{n _ {\\pi} - 1} \\sum_ {i = 1} ^ {n _ {\\pi}} \\left(\\boldsymbol {g} _ {\\pi} ^ {i} - \\boldsymbol {g} _ {\\pi} ^ {m}\\right) ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.65 + ], + "angle": 0, + "content": "Those metrics are measured for each domain and class separately." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.471, + 0.901 + ], + "angle": 0, + "content": "However, the gradients obtained from insufficiently annotated domain and classes are dubious and error-prone [55]. To alleviate the issue, we take inspiration from recent learning theory findings that stochastic gradients along the optimization steps are generally low-rank [35], which indicates that the optimization process is governed by the top gradient eigenspace. Specifically, we devise a gradient projection approach to suppress the noisy model update signals in particular to the target domain and new classes by projecting their gradients on the reliable subspaces identified by the source domain and anchor classes with sufficient supervision. Our approach involves three iterative steps, i.e., subspace construction, gradient projection, and gradient statistics matching. Firstly, to identify the principal eigenspace, we perform Singular Value Decomposition (SVD) on the aggregated gradients from the source domain and anchor classes. We denote gradients collected from source domain, target domain, anchor class, and new" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.399, + 0.893, + 0.442 + ], + "angle": 0, + "content": "class as \\( g_s \\), \\( g_t \\), \\( g_A \\), and \\( g_N \\). Given the sets of gradients \\( G_s = [g_s^1, g_s^2, \\dots, g_s^K] \\) and \\( G_A = [g_A^1, g_A^2, \\dots, g_A^K] \\) stored in the memory buffer of volume \\( K \\), we apply SVD:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.448, + 0.892, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {G} _ {s} = \\boldsymbol {U} _ {s} \\boldsymbol {\\Sigma} _ {s} \\boldsymbol {V} _ {s} ^ {\\top}, \\quad \\boldsymbol {G} _ {A} = \\boldsymbol {U} _ {A} \\boldsymbol {\\Sigma} _ {A} \\boldsymbol {V} _ {A} ^ {\\top}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.475, + 0.893, + 0.574 + ], + "angle": 0, + "content": "where \\(U, V\\), and \\(\\Sigma\\) contain left singular vectors \\(u^i\\), right singular vector \\(v^i\\), and non-negative singular values \\(\\sigma^i\\), respectively. To capture the overall characteristics, we use the gradients averaged along all anchor classes to represent \\(G_A\\). Next, we adopt the low-rank matrix approximation to select the top-\\(r\\) significant left singular vectors in \\(U_s\\) and \\(U_A\\) based on the following criteria:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.582, + 0.892, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\left(\\boldsymbol {G} _ {s}\\right) ^ {r _ {s}} \\right\\| _ {F} ^ {2} \\geq \\tau \\left\\| \\boldsymbol {G} _ {s} \\right\\| _ {F} ^ {2}, \\quad \\left\\| \\left(\\boldsymbol {G} _ {A}\\right) ^ {r _ {A}} \\right\\| _ {F} ^ {2} \\geq \\tau \\left\\| \\boldsymbol {G} _ {A} \\right\\| _ {F} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.608, + 0.892, + 0.701 + ], + "angle": 0, + "content": "where \\((\\pmb{G})^r = \\sum_{i=1}^{r} \\pmb{\\sigma}^i \\pmb{u}^i \\pmb{v}^{i^\\top}\\), \\(\\| \\cdot \\|_F\\) denotes the Frobenius norm, \\(\\tau\\) is a threshold to ensure most information is preserved and is set to 0.98. The principal subspace and the corresponding projection matrix are thereby constructed as \\(M_s = [u_s^1, u_s^2, \\dots, u_s^{r_s}]\\) and \\(M_A = [u_A^1, u_A^2, \\dots, u_A^{r_A}]\\). Afterward, all gradients are projected on the identified subspace as shown in Fig. 2:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.707, + 0.892, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} _ {s} ^ {*} = M _ {s} M _ {s} ^ {\\top} \\boldsymbol {g} _ {s}, \\quad \\boldsymbol {g} _ {t} ^ {*} = M _ {s} M _ {s} ^ {\\top} \\boldsymbol {g} _ {t}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.727, + 0.822, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} _ {A} ^ {*} = \\boldsymbol {M} _ {A} \\boldsymbol {M} _ {A} ^ {\\top} \\boldsymbol {g} _ {A}, \\quad \\boldsymbol {g} _ {N} ^ {*} = \\boldsymbol {M} _ {A} \\boldsymbol {M} _ {A} ^ {\\top} \\boldsymbol {g} _ {N}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.894, + 0.829 + ], + "angle": 0, + "content": "Then following Eq.(3)(4), we minimize the discrepancy between the statistics (i.e., mean and variance) of projected gradients to distill the learning dynamics from the source domain and anchor classes to the target domain and new classes. The overall training objective can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.837, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta} \\mathcal {L} _ {E R M} + \\lambda \\left\\{\\| \\left(\\boldsymbol {g} _ {s} ^ {m *} - \\boldsymbol {g} _ {t} ^ {m *}\\right) \\| _ {2} ^ {2} + \\| \\left(\\boldsymbol {g} _ {s} ^ {v *} - \\boldsymbol {g} _ {t} ^ {v *}\\right) \\| _ {2} ^ {2} \\right. \\\\ \\left. + \\frac {1}{n _ {\\text {n e w}}} \\sum_ {i = 1} ^ {n _ {\\text {n e w}}} \\left[ \\| \\left(\\boldsymbol {g} _ {A} ^ {m *} - \\boldsymbol {g} _ {N i} ^ {m *}\\right) \\| _ {2} ^ {2} + \\| \\left(\\boldsymbol {g} _ {A} ^ {v *} - \\boldsymbol {g} _ {N i} ^ {v *}\\right) \\| _ {2} ^ {2} \\right] \\right\\}. \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21177" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.259 + ], + "angle": 0, + "content": "Here \\(\\mathcal{L}_{ERM}\\) denotes the empirical risk loss, which is implemented with cross-entropy loss for classification and Dice loss [43] for segmentation. \\(\\lambda\\) is a balancing coefficient, \\(n_{new}\\) is the number of new classes in the target domain. It is noted that there exists discrepancy between the anchor and new classes in semantic concepts, which makes feature alignment across classes harmful due to negative transfer [56]. However, aligning class-wise learning dynamics by enforcing the similarity between their gradients could contribute to lower empirical error on new classes. We prove this property theoretically in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.267, + 0.313, + 0.283 + ], + "angle": 0, + "content": "3.4. Historical Self-distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.47, + 0.441 + ], + "angle": 0, + "content": "Inferior generalization behaviour is an outstanding challenge for taxonomy adaptive cross-domain adaptation [19]. When domain shifts and unseen classes exist, a trained model could only maintain its performance on test samples sharing multiple similar attributes with the training data but cannot generalize well to the disparate ones [63]. Motivated by the connection between flat minima and generalizability [31], we devise the historical self-distillation module to smooth the optimization path towards a well-generalizable solution by calibrating gradient descent steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.472, + 0.502 + ], + "angle": 0, + "content": "To reach flat loss landscapes [6], assuming a local loss function \\(\\psi\\) and a network \\(f\\) parameterized by \\(\\theta\\), instead of minimizing the original loss \\(\\psi(f(x;\\theta))\\) only, the optimization objective should be:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.513, + 0.476, + 0.558 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} - \\log \\int_ {\\left\\| \\theta - \\theta^ {*} \\right\\| \\leq \\Gamma} \\exp (- \\psi (f (x; \\theta^ {*})) - \\gamma \\| \\theta - \\theta^ {*} \\|) d \\theta^ {*}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.62 + ], + "angle": 0, + "content": "where \\( x \\) indicates random input samples, \\( \\Gamma \\) is defined as the width of a local parameter valley, \\( \\gamma \\) is a balancing weight. In other words, when \\( \\theta \\) and \\( \\theta^{*} \\) are neighbouring, the following criteria is required to be satisfied:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.631, + 0.47, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\nabla_ {\\theta} \\psi (f (x; \\theta)) - \\nabla_ {\\theta} \\psi (f (x; \\theta^ {*})) \\right\\| \\leq \\beta \\| \\theta - \\theta^ {*} \\|, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a smoothness factor. To this end, we propose to impose regulations on the optimizer steps by enforcing minima with uniformly-distributed gradients. Given that the gradients from most iterations are generally low-rank [35, 2], we exploit past gradients to identify the low-dimensional subspace which indicates the principal gradient distribution of local minima and then project current gradients on the constructed subspace to exclude sharp and noisy signals. Specifically, we collect and store individual gradients for each iteration along the training procedure, denoted as \\(G_{It} = [g_{It}^{t - T}, g_{It}^{t - T + 1}, \\dots, g_{It}^{t - 2}, g_{It}^{t - 1}]\\), where \\(t\\) is the index of current iteration, \\(T\\) is the size of memory buffer. Thereafter we perform SVD on the set of gradients \\(G_{It}\\) to construct the historical subspace. Similar to Eq.(5)(6)(7), the top-\\(r\\) left singular vectors of the decomposed gradient set are concatenated to formulate the projection matrix" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.214 + ], + "angle": 0, + "content": "\\(M_{It}\\). Then for upcoming gradients obtained from backpropagation in each iteration, we project the gradients on the identified low-rank subspace and consider them as residuals. The subsequent optimizer step is conducted with the sum of the original and rectified gradients, as shown in Fig. 2. Take vanilla stochastic gradient descent as an example, with mini-batch gradients \\(\\tilde{g}\\) in each iteration, the parameters \\(\\theta\\) of the optimized network are updated by:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.225, + 0.892, + 0.255 + ], + "angle": 0, + "content": "\\[\n\\theta := \\theta - \\frac {1}{\\kappa} \\cdot \\eta \\cdot \\left(\\boldsymbol {M} _ {I t} \\boldsymbol {M} _ {I t} ^ {\\top} + \\kappa\\right) \\cdot \\tilde {\\boldsymbol {g}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.356 + ], + "angle": 0, + "content": "where \\(\\eta\\) is the learning rate, \\(\\kappa\\) is a trade-off parameter. As training proceeds, the principal subspace and the corresponding projection matrix are periodically renewed by applying SVD on the recently collected gradients. The overall training procedure is summarized in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.369, + 0.695, + 0.386 + ], + "angle": 0, + "content": "3.5. Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.424 + ], + "angle": 0, + "content": "In this section, we motivate our proposed method theoretically and provide mathematical insights on its merits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Joint Characterization of Feature and Output Space While our optimization trajectory distillation approach proposes a novel perspective that differs from existing works imposing regularization terms in feature and model-output space [32, 19], those information is still captured and implicitly modeled in the gradient-based framework. We prove this property for the classification and segmentation tasks under two commonly adopted loss functions. The detailed proof is presented in the supplemental material. It illustrates the superiority of our method as a unified framework that jointly characterizes the feature and output space as well as the learning dynamics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.892, + 0.713 + ], + "angle": 0, + "content": "Impacts on Generalization Error In our method, we propose to minimize the discrepancy between the gradient statistics from different domains and classes in the identified principal subspace. We hereby prove its effectiveness towards a tighter generalization error bound on the target domain and new classes [3]. Details can be found in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.73, + 0.634, + 0.747 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.756, + 0.796, + 0.772 + ], + "angle": 0, + "content": "4.1. Datasets and Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our method and its potential clinical implications, we conduct experiments on five medical image datasets in regard to three important downstream tasks from different clinical scenarios. Extended experiments on more diverse tasks, including radiology and fundus analysis, as well as general visual task, are presented in the supplemental material to substantiate the broader applicability of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21178" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.141, + 0.089, + 0.828, + 0.133 + ], + "angle": 0, + "content": "Table 1. Comparison results of our proposed method against other state-of-the-art methods for nuclei segmentation and recognition. The metrics with asterisk (*) are calculated on target-private new classes only. The subscripts denote the standard deviations. The best and second-best results are highlighted in bold and brown, respectively." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.134, + 0.83, + 0.328 + ], + "angle": 0, + "content": "
Methods5-shot10-shot
mF1mF1*mPQmPQ*mF1mF1*mPQmPQ*
Sup-only28.080.8921.460.3612.350.5312.230.4530.900.6625.230.8314.180.2414.940.32
Multi-task [48]33.851.2618.151.0818.580.4210.770.3535.150.9421.290.6219.140.2712.890.34
DANN [16]32.761.4116.521.1018.310.6610.400.5935.761.3523.120.8819.700.6014.140.42
CGDM [14]35.031.2821.040.9919.590.5013.220.3237.131.5724.221.1519.780.7113.900.40
LETR [40]27.781.4415.361.0216.680.5410.030.5134.851.5420.030.7318.180.5911.790.27
FT-CIDA [32]30.510.9216.550.8015.970.399.420.4032.631.2121.080.9517.090.4711.960.49
STARTUP [45]37.030.5822.440.7620.400.3313.820.3040.850.5025.920.9123.370.1916.640.24
DDN [25]37.920.8623.711.0520.620.5713.870.4340.160.7226.941.0822.790.3216.820.50
TSA [34]34.031.1719.551.1118.650.4312.060.3734.271.0522.581.2019.780.6114.240.73
TACS [19]36.731.6522.261.2918.130.8012.570.6239.291.4825.841.1322.270.6815.980.56
Ours40.260.9027.140.9721.780.4914.960.2243.880.5231.430.7824.810.2619.350.33
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.473, + 0.625 + ], + "angle": 0, + "content": "Nuclei Segmentation and Recognition Technically, this task can be formalized as a simultaneous instance segmentation and classification problem [21]. We evaluate our method by considering PanNuke [15] and Lizard [20] as the source and target datasets. They contain 481 and 291 visual fields cropped from whole-slide images with 189,744 and 495,179 annotated nuclei, respectively. There are six types of nuclei in Lizard and only two of them overlap with the five classes in PanNuke, which suggests four new classes in the target domain. We employ the widely used HoverNet [21] architecture with a standard ResNet-50 backbone as the base model. For quantitative evaluation, we follow [20] and report the average F1 and PQ (panoptic quality) scores for all classes (denoted as mF1 and mPQ) to provide a comprehensive assessment measuring the accuracy of nuclei detection, segmentation, and classification. Additionally, we also report the mF1* and mPQ* scores which are solely computed on new classes in the target domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Cancer Tissue Phenotyping In this setting, we perform adaptation from a two-class discrimination task to an eight-class one on CRC-TP [29] and Kather [30] datasets. Kather consists of \\(5000 \\times 150\\) pathology image patches and includes six novel classes other than the two common classes also existing in the source domain. The experiments are conducted with ResNet-101 as backbone. For evaluation, we exploit the accuracy and F1 scores as metrics to indicate the classification performance. The average scores for all classes (mAcc and mF1) and particularly the ones on new classes only (mAcc* and mF1*) are reported." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Skin Lesion Diagnosis We construct a fine-grained taxonomy adaptation setting from two coarse classes to seven subclasses with HAM10000 [52], which contains 10015 dermatoscopic images sampled from different body structures. The adopted experiment settings and evaluation metrics are the same as the cancer tissue phenotyping task. Since the label space of the source and target datasets do" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.381 + ], + "angle": 0, + "content": "not overlap, we only report the mAcc* and mF1* scores which are computed on target classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.893, + 0.489 + ], + "angle": 0, + "content": "Following previous works in UDA [7], for all experiments, each dataset is randomly split into \\(80\\% / 20\\%\\) for training/testing. To explicitly recognize the target-private new classes, we sample few (5/10) samples with corresponding labels to formulate the support set. The remaining target data is left unlabeled. More details can be found in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.593, + 0.515 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.719 + ], + "angle": 0, + "content": "We compare our method with state-of-the-art approaches for cross-domain adaptation, including UDA methods DANN [16] and CGDM [14] which do not consider the taxonomic inconsistency, as well as LETR [40], FT-CIDA[32], STARTUP [45], DDN [25], TSA [34], and TACS [19] which are designed for both domain shifts and new class learning. For UDA methods, we jointly perform domain alignment and train the model with the labeled target data. In addition, we conduct comparisons with multi-task learning approach [48], which trains the model on all domains and label sets simultaneously. \"Sup-only\" indicates the baseline where only the supervised loss on the labeled target dataset is used for training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Table 1 shows the comparison results on the cross-domain nuclei recognition and segmentation task. All methods adopt the same ResNet-50-HoverNet as base model following [21]. It can be observed that our proposed method outperforms the state-of-the-art cross-domain adaptation approaches by a large margin. We also notice an important point that compared with the \"Sup-only\" baseline, most competing methods fail to attain better performance on the metrics specific to new classes (i.e., mF1* and mPQ*). For example, under the 5-shot setting, despite a \\(4.68\\%\\) improvement is achieved by DANN on mF1, it achieves inferior performance on mF1* which is \\(4.94\\%\\) worse than \"Sup-only\"." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21179" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.09, + 0.84, + 0.105 + ], + "angle": 0, + "content": "Table 2. Comparison results of our proposed method against other state-of-the-art methods for cancer tissue phenotyping." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.107, + 0.842, + 0.3 + ], + "angle": 0, + "content": "
Methods5-shot10-shot
mAccmAcc*mF1mF1*mAccmAcc*mF1mF1*
Sup-only46.821.3042.261.2644.641.4942.831.0848.521.0245.931.6745.341.4546.151.83
Multi-task [48]48.350.7842.021.3445.551.1241.681.6348.560.6945.001.7246.030.6444.461.21
DANN [16]47.460.6742.531.2243.081.1443.131.3048.420.4845.311.2746.500.8144.020.64
CGDM [14]49.020.8243.531.5646.290.8743.580.6950.540.8444.571.3548.170.6645.230.40
LETR [40]45.471.9541.331.4743.171.8841.202.0245.940.5640.131.3041.161.1339.610.90
FT-CIDA [32]43.720.8839.060.7940.911.0238.730.8444.650.5338.800.9440.940.9638.961.07
STARTUP [45]49.520.6342.931.8343.591.3739.970.5949.400.5146.911.8746.380.7047.851.36
DDN [25]47.141.7543.250.9942.922.2940.760.5048.270.6744.081.2246.031.1445.441.18
TSA [34]48.221.4942.130.7844.391.8441.921.1347.181.0241.261.1045.341.7244.691.89
TACS [19]45.301.0740.281.1144.741.2642.040.9546.951.1643.691.7447.461.5147.511.25
Ours50.140.7647.661.5447.411.1046.451.1555.860.9851.271.0252.980.4952.120.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.323, + 0.454, + 0.351 + ], + "angle": 0, + "content": "Table 3. Comparison results of our proposed method against other state-of-the-art methods for skin lesion diagnosis." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.354, + 0.455, + 0.546 + ], + "angle": 0, + "content": "
Methods5-shot10-shot
mAcc*mF1*mAcc*mF1*
Sup-only35.710.7919.890.2238.500.3822.360.67
Multi-task [48]33.801.2019.060.3137.850.4023.320.32
DANN [16]37.220.5723.990.9240.060.9424.320.72
CGDM [14]41.501.0324.260.8047.220.7024.000.23
LETR [40]36.040.4221.390.6640.871.2223.500.49
FT-CIDA [32]31.950.7718.130.4834.600.9020.960.88
STARTUP [45]43.211.1527.170.9243.881.3525.590.60
DDN [25]36.191.0521.500.3442.461.2923.920.28
TSA [34]40.411.3624.051.0142.071.5125.441.06
TACS [19]45.050.4625.600.6546.730.6425.960.79
Ours50.790.6930.050.8152.140.9830.520.83
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.47, + 0.672 + ], + "angle": 0, + "content": "This phenomenon is incurred by the failure of cross-class generalization. In contrast, our method yields substantial improvements on those new class-specific metrics, which indicates the superior generalizability of our method. The observation is further verified through visual comparisons in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The overall quantitative results on the cancer tissue phenotyping and skin lesion diagnosis tasks are presented in Table 2 and 3, respectively. All methods are implemented based on the ResNet-101 backbone. Under each few-shot setting, our method shows higher classification scores compared with previous approaches. In particular, we achieve \\(0.62\\% \\sim 5.32\\%\\) and \\(1.12\\% \\sim 4.81\\%\\) improvements against the second-best results in terms of mAcc and mF1 for cancer tissue phenotyping, as well as \\(4.92\\% \\sim 7.58\\%\\) and \\(2.88\\% \\sim 4.93\\%\\) improvements in terms of mAcc* and mF1* for skin lesion diagnosis. The higher accuracy consistently attained by our method under different cross-domain adaptation settings and clinical tasks is attributed to the strong generalizability brought by the proposed optimization trajectory distillation approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.324, + 0.722, + 0.34 + ], + "angle": 0, + "content": "4.3. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.715 + ], + "angle": 0, + "content": "Ablation of Key Components Table 4 shows the ablation study results to illustrate the effectiveness of key components. \"projection\" denotes the gradient projection approach introduced in Section 3.3. For the \"w/o projection\" ablation, we perform cross-domain and cross-class distillation by directly minimizing the statistics discrepancy between unrectified gradients \\((g_{s}, g_{t})\\) and \\((g_{A}, g_{N})\\) in Eq. (7). It is observed that each component is indispensable and could improve the overall performance, which elaborates the importance of external navigation for learning target domain and new classes in network training. Interestingly, we notice that the impacts of cross-domain and cross-class distillation could be disparate across tasks. For example, cross-class distillation brings higher performance gain compared with cross-domain distillation on the cancer tissue phenotyping benchmark, while the situation is reversed for skin lesion diagnosis. It is caused by the different natures of each task. Class-imbalance is a serious issue for skin lesion diagnosis in that the class distribution is highly biased [52], which compromises the effectiveness of the cross-class distillation module. Differently, for cancer tissue phenotyping, classes are uniformly distributed [30] and could hence ensure the representativity of anchor classes and their exemplarity to be distilled to new classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Hyperparameter Sensitivity To investigate the effect of hyperparameters \\(\\lambda\\) in Eq. (8) and \\(\\kappa\\) in Eq. (11), we present the performance comparison by varying the choices of their values for cancer tissue phenotyping, as shown in Fig. 3. We only change one hyperparameter while keeping the other fixed at a time. The experimental results show that our framework is quite stable for hyperparameters in a wide interval, yet setting \\(\\lambda\\) and \\(\\kappa\\) to a large value is detrimental to the classification accuracy. Specifically, \\(\\lambda\\) is a balancing coefficient between class discriminativeness and gradient alignment. When it is too large, the trained model is biased to matching learning dynamics and would lose es" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21180" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.177, + 0.09, + 0.794, + 0.118 + ], + "angle": 0, + "content": "Table 4. Key component analysis of our method on three cross-domain adaptation benchmarks under 10-shot scheme. The best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.179, + 0.12, + 0.794, + 0.23 + ], + "angle": 0, + "content": "
MethodsNucleiCancerSkin
mF1mF1*mPQmPQ*mAccmAcc*mF1mF1*mAcc*mF1*
Full Framework43.8831.4324.8119.3555.8651.2752.9852.1252.1430.52
w/o cross-domain39.9630.0822.2918.1052.8149.9351.6250.6444.3425.28
w/o cross-class42.3027.8223.7617.0548.9047.2249.8544.7649.0628.24
w/o historical41.5430.4723.0017.5950.1748.0548.1945.1245.8325.97
w/o projection40.0628.8022.1017.3153.5845.7849.0149.2047.2928.02
" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.253, + 0.46, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.376, + 0.471, + 0.417 + ], + "angle": 0, + "content": "Figure 3. Performance comparison between different choices of hyperparameters \\(\\lambda\\) and \\(\\kappa\\) on the cancer tissue phenotyping benchmark under 10-shot scheme." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.47, + 0.534 + ], + "angle": 0, + "content": "sential semantic knowledge. For \\(\\kappa\\), it controls the proportions of unrectified gradients. Its large value indicates that most original gradients are preserved, which would incur noisy optimization steps and loss of flatness. Therefore, we decide to set \\(\\lambda = 10\\) and \\(\\kappa = 100\\). It achieves the best performance under most metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.471, + 0.779 + ], + "angle": 0, + "content": "Impacts on Optimization Trajectory and Model Flatness To better understand the effectiveness of our method, we perform in-depth analysis about its impacts on model's training trajectory and flatness of the converged local minima. Firstly, as shown in Fig. 4(a), we visualize the optimization paths of our method and the multi-task baseline on the test accuracy surface with a contour plot. The details of plot procedure can refer to [26]. It can be observed that the optimization of the multi-task baseline fluctuates sharply and is restricted in a sub-optimal area with a relatively low test accuracy. This is induced by the biased navigation in the course of training due to domain shifts and category gap. With the devised dual-stream optimization trajectory distillation module, our method could by contrast suppress the noisy update signals and proceed consistently towards the optimum." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Besides, we perform analysis about the solutions found by our method and the baseline in terms of flatness [5]. Specifically, the local flatness of a model parameter \\(\\theta\\) is quantified by the expected changes of test accuracy between \\(\\theta\\) and a neighbouring \\(\\theta^{*}\\) with perturbation rate \\(\\varrho\\): \\(\\mathcal{F}_{\\varrho}(\\theta) = \\mathbb{E}_{\\|\\theta^{*}\\| = (1 + \\varrho)\\|\\theta\\|}\\left[\\mathrm{Acc}(\\theta^{*}) - \\mathrm{Acc}(\\theta)\\right]\\). We approximate the expected value by Monte-Carlo sampling over 50 samples. As demonstrated in Fig. 4(b), comparing with the" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.251, + 0.707, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.408, + 0.694, + 0.419 + ], + "angle": 0, + "content": "(a) Visualization of Optimization Trajectory" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.251, + 0.889, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.408, + 0.854, + 0.418 + ], + "angle": 0, + "content": "(b) Flatness Analysis" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.423, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Figure 4. Further analysis of our proposed method. (a) We plot the optimization trajectory of our method and the multi-task baseline on the test accuracy surface. (b) We visualize the local flatness of our method and the baseline via test accuracy gap by varying the model parameter perturbation rate. Each point is computed by Monte-Carlo approximation over 50 random samples. All the experiments are conducted on the cancer tissue phenotyping benchmark under 10-shot scheme." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.619 + ], + "angle": 0, + "content": "baseline approach, our method shows stronger robustness against model parameter perturbation. It proves that our method successfully converges to minima with better local flatness and thus possesses superior generalizability [31]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.619, + 0.649 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we study the taxonomy adaptive cross-domain adaptation paradigm, which allows incoherent label sets between source and target datasets. To jointly address the data distribution bias and category gap, we propose a unified framework which performs cross-domain and cross-class optimization trajectory distillation to calibrate the learning dynamics of insufficiently annotated domain and classes. Historical self-distillation is further devised to attain a well-generalizable solution via regulating gradient distributions. Extensive experiments on multiple benchmarks with different task contexts demonstrate the effectiveness and generalization capability of our method. In the future, we will extend this work to further discover the underlying mechanism of optimization trajectory distillation for model generalization and evaluate its potential in more adaptation scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21181" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Sanjeev Arora, Rong Ge, Yingyu Liang, Tengyu Ma, and Yi Zhang. Generalization and equilibrium in generative adversarial nets (gans). In International Conference on Machine Learning, pages 224-232. PMLR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[2] Sheikh Shams Azam, Seyyedali Hosseinalipour, Qiang Qiu, and Christopher Brinton. Recycling model updates in federated learning: Are gradient subspaces low-rank? In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[3] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.47, + 0.339 + ], + "angle": 0, + "content": "[4] Debora Calderola, Barbara Caputo, and Marco Ciccone. Improving generalization in federated learning by seeking flat minima. In European Conference on Computer Vision, pages 654-672. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.47, + 0.409 + ], + "angle": 0, + "content": "[5] Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Seunghyun Park, Yunsung Lee, and Sungrae Park. Swad: Domain generalization by seeking flat minima. Advances in Neural Information Processing Systems, 34:22405-22418, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.48 + ], + "angle": 0, + "content": "[6] Pratik Chaudhari, Anna Choromanska, Stefano Soatto, Yann LeCun, Carlo Baldassi, Christian Borgs, Jennifer Chayes, Levent Sagun, and Riccardo Zecchina. Entropy-SGD: Biasing gradient descent into wide valleys. In International Conference on Learning Representations, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[7] Cheng Chen, Qi Dou, Hao Chen, Jing Qin, and Pheng-Ann Heng. Synergistic image and feature adaptation: Towards cross-modality domain adaptation for medical image segmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 865–872, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[8] Wentao Chen, Zhang Zhang, Wei Wang, Liang Wang, Zilei Wang, and Tieniu Tan. Cross-domain cross-set few-shot learning via learning compact and aligned representations. In European Conference on Computer Vision, pages 383-399. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.47, + 0.69 + ], + "angle": 0, + "content": "[9] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[10] Ali Cheraghian, Shafin Rahman, Sameera Ramasinghe, Pengfei Fang, Christian Simon, Lars Petersson, and Mehrtash Harandi. Synthesized feature based few-shot class incremental learning on a mixture of subspaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8661-8670, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[11] Debasmit Das, Sungrack Yun, and Fatih Porikli. ConfeSS: A framework for single source cross-domain few-shot learning. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Susanna Dodd, Mike Clarke, Lorne Becker, Chris Mavergames, Rebecca Fish, and Paula R Williamson. A taxonomy has been developed for outcomes in medical research to help improve knowledge discovery. Journal of clinical epidemiology, 96:84-92, 2018." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[13] C Dominguez Conde, C Xu, LB Jarvis, DB Rainbow, SB Wells, T Gomes, SK Howlett, O Suchanek, K Polanski, HW King, et al. Cross-tissue immune cell analysis reveals tissue-specific features in humans. Science, 376(6594), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[14] Zhekai Du, Jingjing Li, Hongzu Su, Lei Zhu, and Ke Lu. Cross-domain gradient discrepancy minimization for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3937-3946, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[15] Jevgenij Gamper, Navid Alemi Koohbanani, Ksenija Benet, Ali Khuram, and Nasir Rajpoot. Pannuke: an open pancancer histology dataset for nuclei instance segmentation and classification. In European congress on digital pathology, pages 11-19. Springer, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[16] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[17] Zhiqiang Gao, Shufei Zhang, Kaizhu Huang, Qiufeng Wang, and Chaoliang Zhong. Gradient distribution alignment certificates better adversarial domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8937-8946, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[18] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[19] Rui Gong, Martin Danelljan, Dengxin Dai, Danda Pani Paudel, Ajad Chhatkuli, Fisher Yu, and Luc Van Gool. Tacs: Taxonomy adaptive cross-domain semantic segmentation. In European Conference on Computer Vision, pages 19-35. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.56, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[20] Simon Graham, Mostafa Jahanifar, Ayesha Azam, Mohammed Nimir, Yee-Wah Tsang, Katherine Dodd, Emily Hero, Harvir Sahota, Atisha Tank, Ksenija Benes, et al. Lizard: a large-scale dataset for colonic nuclear instance segmentation and classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 684-693, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.659, + 0.892, + 0.728 + ], + "angle": 0, + "content": "[21] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:101563, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.729, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[22] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In European conference on computer vision, pages 124-141. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.8, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[23] Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural computation, 9(1):1-42, 1997." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.83, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[24] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In European Conference on Computer Vision, pages 20-37. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Ashraful Islam, Chun-Fu Richard Chen, Rameswar Panda, Leonid Karlinsky, Rogerio Feris, and Richard J Radke. Dy" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21182" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "namic distillation network for cross-domain few-shot recognition with unlabeled data. Advances in Neural Information Processing Systems, 34:3584-3595, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[26] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. Proceedings of the Conference on Uncertainty in Artificial Intelligence, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[27] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.32 + ], + "angle": 0, + "content": "[28] Stanisław Jastrzebski, Zachary Kenton, Devansh Arpit, Nicolas Ballas, Asja Fischer, Yoshua Bengio, and Amos Storkey. Three factors influencing minima in sgd. arXiv preprint arXiv:1711.04623, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[29] Sajid Javed, Arif Mahmood, Muhammad Moazam Fraz, Navid Alemi Koohbanani, Ksenija Benes, Yee-Wah Tsang, Katherine Hewitt, David Epstein, David Snead, and Nasir Rajpoot. Cellular community detection for tissue phenotyping in colorectal cancer histology images. Medical image analysis, 63:101696, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[30] Jakob Nikolas Kather, Cleo-Aron Weis, Francesco Bianconi, Susanne M Melchers, Lothar R Schad, Timo Gaiser, Alexander Marx, and Frank Gerrit Zöllner. Multi-class texture analysis in colorectal cancer histology. Scientific reports, 6(1):1-11, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[31] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. In International Conference on Learning Representations, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.604 + ], + "angle": 0, + "content": "[32] Jogendra Nath Kundu, Rahul Mysore Venkatesh, Naveen Venkat, Ambareesh Revanur, and R Venkatesh Babu. Class incremental domain adaptation. In European Conference on Computer Vision, pages 53-69. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[33] Canran Li, Dongnan Liu, Haoran Li, Zheng Zhang, Guangming Lu, Xiaojun Chang, and Weidong Cai. Domain adaptive nuclei instance segmentation and classification via category-aware feature alignment and pseudo-labelling. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 715-724. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[34] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Cross-domain few-shot learning with task-specific adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7161-7170, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[35] Xinyan Li, Qilong Gu, Yingxue Zhou, Tiancong Chen, and Arindam Banerjee. Hessian based analysis of sgd for deep nets: Dynamics and generalization. In Proceedings of the 2020 SIAM International Conference on Data Mining, pages 190-198. SIAM, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[36] Hanwen Liang, Qiong Zhang, Peng Dai, and Juwei Lu. Boosting the generalization capability in cross-domain few-shot learning via noise-enhanced supervised autoencoder. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9424-9434, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[37] Dongnan Liu, Chaoyi Zhang, Yang Song, Heng Huang, Chenyu Wang, Michael Barnett, and Weidong Cai. Decompose to adapt: Cross-domain object detection via feature disentanglement. IEEE Transactions on Multimedia, 25:1333-1344, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[38] Dongnan Liu, Donghao Zhang, Yang Song, Fan Zhang, Lauren O'Donnell, Heng Huang, Mei Chen, and Weidong Cai. Unsupervised instance segmentation in microscopy images via panoptic domain adaptation and task re-weighting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4243-4252, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[39] Xu Luo, Jing Xu, and Zenglin Xu. Channel importance matters in few-shot image classification. In International Conference on Machine Learning, pages 14542-14559. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[40] Zelun Luo, Yuliang Zou, Judy Hoffman, and Li F Fei-Fei. Label efficient learning of transferable representations acrossss domains and tasks. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[41] Deborah Lupton. Medicine as culture: illness, disease and the body. Sage, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[42] Lucas Mansilla, Rodrigo Echeveste, Diego H Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6630-6638, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[43] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[44] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In Proceedings of the IEEE international conference on computer vision, pages 754-763, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[45] Cheng Perng Phoo and Bharath Hariharan. Self-training for few-shot transfer across extreme task differences. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[46] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[47] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, Trevor Darell, and Kate Saenko. Semi-supervised domain adaptation via minimax entropy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8050-8058, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[48] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[49] Yuge Shi, Jeffrey Seely, Philip Torr, Siddharth N, Awni Hannun, Nicolas Usunier, and Gabriel Synnaeve. Gradient matching for domain generalization. In International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[50] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[51] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source der" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21183" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "matoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[53] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In International Conference on Learning Representations, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[54] Qian Wang and Toby Breckon. Unsupervised domain adaptation via structured prediction based selective pseudolabeling. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 6243-6250, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[55] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[56] Zirui Wang, Zihang Dai, Barnabás Póczos, and Jaime Carbonell. Characterizing and avoiding negative transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11293-11302, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[57] Julia K Winkler, Christine Fink, Ferdinand Toberer, Alexander Enk, Teresa Deinlein, Rainer Hofmann-Wellenhof, Luc Thomas, Aimilios Lallas, Andreas Blum, Wilhelm Stolz, et al. Association between surgical skin markings in dermoscopic images and diagnostic performance of a deep learning convolutional neural network for melanoma recognition. JAMA dermatology, 155(10):1135-1141, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.469, + 0.514 + ], + "angle": 0, + "content": "[58] Mengya Xu, Mobarakol Islam, Chwee Ming Lim, and Hongliang Ren. Class-incremental domain adaptation with smoothing and calibration for surgical report generation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 269-278. Springer, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[59] Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Universal domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2720-2729, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[60] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.469, + 0.698 + ], + "angle": 0, + "content": "[61] Wang Yuan, Zhizhong Zhang, Cong Wang, Haichuan Song, Yuan Xie, and Lizhuang Ma. Task-level self-supervision for cross-domain few-shot learning. Proceedings of the AAAI Conference on Artificial Intelligence, 36(3):3215-3223, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[62] Qiming Zhang, Jing Zhang, Wei Liu, and Dacheng Tao. Category anchor-guided unsupervised domain adaptation for semantic segmentation. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.757, + 0.469, + 0.798 + ], + "angle": 0, + "content": "[63] Lin Zuo, Mengmeng Jing, Jingjing Li, Lei Zhu, Ke Lu, and Yang Yang. Challenging tough samples in unsupervised domain adaptation. Pattern Recognition, 110:107540, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21184" + } + ] +] \ No newline at end of file diff --git a/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/1f4cb2e9-7a0d-4b5a-99a2-329016955f96_origin.pdf b/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/1f4cb2e9-7a0d-4b5a-99a2-329016955f96_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1587bae1010aaef1902fda1c1e8ccbaca1c93c7f --- /dev/null +++ b/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/1f4cb2e9-7a0d-4b5a-99a2-329016955f96_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:352bb50a7931ddc5d8a18f7a8b024d78cb652e9aace862bb2a2496ed28155e47 +size 1826618 diff --git a/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/full.md b/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d1bc2c922600034bdd73c909e14ec19ca0165524 --- /dev/null +++ b/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/full.md @@ -0,0 +1,297 @@ +# Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation + +Jianan Fan $^{1}$ , Dongnan Liu $^{1}$ , Hang Chang $^{2}$ , Heng Huang $^{3}$ , Mei Chen $^{4}$ , and Weidong Cai $^{1}$ + +$^{1}$ University of Sydney $^{2}$ Lawrence Berkeley National Laboratory $^{3}$ University of Maryland at College Park $^{4}$ Microsoft + +jfan6480@uni.sydney.edu.au dongnan.liu@sydney.edu.au hchang@lbl.gov + +henghuanghh@gmail.com Mei.Chen@microsoft.com tom.cai@sydney.edu.au + +# Abstract + +The success of automated medical image analysis depends on large-scale and expert-annotated training sets. Unsupervised domain adaptation (UDA) has been raised as a promising approach to alleviate the burden of labeled data collection. However, they generally operate under the closed-set adaptation setting assuming an identical label set between the source and target domains, which is over-restrictive in clinical practice where new classes commonly exist across datasets due to taxonomic inconsistency. While several methods have been presented to tackle both domain shifts and incoherent label sets, none of them take into account the common characteristics of the two issues and consider the learning dynamics along network training. In this work, we propose optimization trajectory distillation, a unified approach to address the two technical challenges from a new perspective. It exploits the low-rank nature of gradient space and devises a dual-stream distillation algorithm to regularize the learning dynamics of insufficiently annotated domain and classes with the external guidance obtained from reliable sources. Our approach resolves the issue of inadequate navigation along network optimization, which is the major obstacle in the taxonomy adaptive cross-domain adaptation scenario. We evaluate the proposed method extensively on several tasks towards various endpoints with clinical and open-world significance. The results demonstrate its effectiveness and improvements over previous methods. Code is available at https://github.com/camwew/TADA-MI. + +# 1. Introduction + +Automated and objective analysis of medical images is an important research topic and has been explored in various clinical applications [57, 13]. To relieve the burden of acquiring massive annotated data for model development on new domains, several unsupervised domain adapta + +![](images/8880daf55c116dc96f9c49849b0c6e7f4a517c78d336058b001a0e3aed7c4d52.jpg) +Figure 1. (a) Illustration of the issue of taxonomic inconsistency with nuclei recognition. The category label sets are incoherent across datasets. Different colours indicate the class each nucleus belongs to. The red asterisks denote novel classes only existing in Dataset 2. (b) Concept of the proposed method. We perform optimization trajectory distillation to deliver external navigation for learning target domain and new classes where the optimization steps tend to be restricted and unreliable. + +tion (UDA) methods [7, 38, 37] are proposed to mitigate the data distribution bias [51] between a richly labeled source domain and a target domain with no explicit supervision. + +However, these works assume a closed-set adaptation setting that the source and target domains should necessarily share the identical category label space and definition [44]. The restriction limits their practical applicability in the clinical wild [15]. Unlike natural images where the definitions of different entity categories (e.g., cat v.s. dog) are structured and globally unified, in the medical domain, inconsistency in taxonomy across different countries or institutes is a common issue compromising the feasibility of cross-domain adaptation [12]. Take nuclei recognition in histology images as an example, the ambiguous biologi + +cal characteristics of cells and distinct clinical usages of the analysis outcomes result in a lack of a unified categorization schema [15, 20]. With specific clinical purposes and downstream applications, different datasets could categorize nuclei with disparate granularities of biological concepts. Besides, as medical research evolves rapidly, novel cell and disease classes could be discovered and form datasets with continually expanding category label sets [41]. This motivates us to develop a generalized adaptation method with the capability to tackle both data distribution bias and category gap for more flexible UDA in clinical practice. + +To this end, we study the problem of taxonomy adaptive domain adaptation [19]. It assumes the target domain could adopt a label space different from the source domain. Following UDA, a labeled source dataset and unlabeled samples from the target domain are available during training. Additionally, to recognize the novel/fine-grained categories which are non-existent or unspecified in the source domain, a few samples of those target-private categories are annotated and utilized as exemplars. The technical challenge for this setting lies in how to alleviate domain shifts and concurrently learn new classes with very limited supervision. Recently, several methods are proposed towards a similar goal [32, 22]. However, they typically address the two issues individually in separate contexts and fail to design a unified paradigm according to their common characteristics [19], which could incur a subtle collision across issues since their objectives are non-relevant [60]. Besides, existing works either focus on cross-domain/class alignment in the feature space [40, 36] or resort to model output-based regularization such as self-supervision [45], whereas those approaches suffer from the equilibrium challenge of adversarial learning [1] and the low-quality pseudo-labels [54]. + +In this work, we present a unified framework for taxonomy adaptive cross-domain adaptation from a new perspective, i.e., via optimization trajectory distillation. Our strategy is motivated by a common challenge existing in both cross-domain and small-data learning regimes, which is the inadequate navigation in the course of network optimization. For cross-domain learning, the unstable feature alignment procedure and noisy pseudo-labels tend to induce error accumulation along network training [62]. Similarly, with limited support samples in new class learning, the optimization of network is inclined to step towards restricted local minima and cannot cover a globally-generalizable space [55]. To this end, we propose to exploit the optimization trajectory from a reliable "teacher" to provide external guidance for regularizing the learning dynamics of insufficiently annotated domain and classes, as illustrated in Fig. 1(b). + +Our method consists of two key components, i.e., cross-domain/class distillation and historical self-distillation. (i) Cross-domain and cross-class distillation aim to leverage the optimization trajectory from the richly labeled do + +main and classes to serve as the "teacher". Motivated by the Neural Tangent Kernel (NTK) theory [27], we characterize the network optimization trajectory through gradient statistics. Then, given the observation that the subspace spanned by the gradients from most iterations is generally low-rank [35, 2], we design a gradient projection approach to suppress the noisy signals in stochastic gradients and rectify the distillation process. Thereafter constraints are imposed on the gradient statistics of target domain and new classes to calibrate their training dynamics towards domain-invariant and unbiased learning procedure. (ii) Historical self-distillation further drives the optimization paths of model to converge towards flat minima. It is found that the flatness of loss landscapes is strongly related to the model's robustness and generalizability [26], while how to take advantage of the insight to tackle domain shifts and limited supervision is under-explored. We propose to exploit the historical gradients to construct the informative low-rank subspaces and then perform gradient projection to alleviate loss sharpness. It compensates for the intense and out-of-order optimization updates incurred by inadequate regularization and leads to better generalization. + +Our prime contributions are as follows: (1) We introduce a more generalized cross-domain adaptation paradigm for medical image analysis in which both data distribution bias and category gap exist across the source and target domains. (2) We leverage insights from recent learning theory research and propose a novel dual-stream optimization trajectory distillation method to provide external navigation in network training. We perform theoretical justifications from two perspectives to illustrate the merits of our method. (3) Experiments on various benchmarks validate the effectiveness and robustness of our proposed method and its improvements over existing approaches. + +# 2. Related Works + +Domain Adaptation Domain adaptation (DA) aims to mitigate the data distribution bias [16]. The classic semi-supervised/unsupervised DA focuses on the scenario where the target domain has an identical label space to the source domain [47, 14, 33]. The over-restricted precondition cannot suffice the demand of clinical usages [58]. There exist several efforts to take a step further than the closed-set adaptation, such as open-set DA [44] and universal DA [59]. However, the aim of those works is solely to detect the previously unseen classes, instead of learning to separately identify each new class with few supports. Some recent studies suggest to explicitly recognize the target-private categories. [32] performs feature projection and alignment based on the prototypical networks [50]. [19] combines pseudo-labeling and contrastive learning to combat domain bias and label gap. Nevertheless, those works are limited to tackling the two issues individually, which fails to exploit + +their common characteristics and propose a unified solution. Cross-Domain Few-Shot Learning The technical challenge we need to resolve is similar to an emerging research topic, i.e., cross-domain few-shot learning [22]. The mainstream methods for this challenge include self-training [45, 25, 61], contrastive learning [11, 19], feature alignment/augmentation [53, 8, 24], and channel transformation [34, 39]. Those methods either focus on feature-level modulation or turn to output-level self-supervision and finetuning, which could bring instability along optimization [1, 54]. In this work, we propose a novel gradient-based framework to transfer knowledge in both cross-domain and few-shot settings. It implicitly characterizes the information in both feature space and output space. + +Gradient Manipulation In current deep learning systems where gradient descent serves as the most popular training algorithm, the directions and amplitudes of network optimization steps are embedded in the gradient space [9]. Gradient manipulation refers to directly modulating the optimization gradients for improved learning process [42]. Recently, it is introduced to mitigate the data distribution shift and shows promises in UDA and domain generalization [14, 17, 49]. However, those works are still restricted by the closed-set condition and do not consider the low-rank nature of gradients [35], which make their approaches vulnerable to noises in the small-data learning regime. + +Flatness in Optimization In previous machine learning research, the connection between convergence to flat minima in optimization and model generalizability has long been established [23, 31]. Theoretical and empirical studies prove that the flatness of loss landscapes highly correlates with generalization capability [4]. [5] averages the network weights of multiple checkpoints to prompt generalization across domains. Different from their post-hoc averaging approach operating on fixed models, we propose to regularize the intermediate dynamics of network optimization to achieve flat minima. + +# 3. Method + +# 3.1. Problem Statement + +The taxonomy adaptive cross-domain adaptation problem can be formalized as follows. There exists a labeled source dataset $\mathcal{D}_s = \{(\mathbf{x}_s,\mathbf{y}_s)\}$ , an unlabeled target dataset $\mathcal{D}_t^u = \{(\mathbf{x}_t^u)\}$ , and a few-shot labeled support set from the target domain $\mathcal{D}_t^l = \{(\mathbf{x}_t^l,\mathbf{y}_t^l)\}$ . Samples from source and target datasets are drawn from different data distributions, i.e., $\mathbf{x}_s\sim \mathcal{P}_s$ , $\mathbf{x}_t\sim \mathcal{P}_t$ , $\mathcal{P}_s\neq \mathcal{P}_t$ . The label sets of source and target datasets are denoted as $\mathcal{C}_s$ and $\mathcal{C}_t$ . Due to the novel/fine-grained categories only existing in the target dataset, we have $\mathcal{C}_s\neq \mathcal{C}_t$ . The shared and target-private new classes are indicated as $\overline{\mathcal{C}} = \mathcal{C}_s\cap \mathcal{C}_t$ and $\bar{\mathcal{C}} = \mathcal{C}_t\backslash \overline{\mathcal{C}}$ . We call $\overline{\mathcal{C}}$ existing in both datasets as anchor classes since + +they can be used as anchors to facilitate knowledge transfer. The goal is to train a model jointly on $\mathcal{D}_s, \mathcal{D}_t^u, \mathcal{D}_t^l$ which performs inferences on data from $\mathcal{P}_t$ with labels in $\mathcal{C}_t$ . + +# 3.2. Characterization of Optimization Trajectory + +The technical challenge for the identified problem lies in overcoming both domain shifts and overfitting incurred by scarce training samples. As discussed in Section 1, those issues could lead to unreliable navigation during neural network optimization [18, 10]. Therefore, we propose to distill the optimization trajectory knowledge from well-annotated sources to regularize the training dynamics of target domain and new classes. + +Recently, NTK theory [27] has been regarded as a tool to dictate the evolution of neural networks. For a network $f$ parameterized by $\theta$ , its training dynamics can be described by a kernel $K$ . Given a training set with $N$ samples, the kernel can be defined as an $N \times N$ matrix with $K(i,j;\theta) = \nabla_{\theta}f(x^{i};\theta)^{\top} \cdot \nabla_{\theta}f(x^{j};\theta)$ , where $x^{i}$ and $x^{j}$ are two input samples. It indicates the strong correlations between the network's optimization trajectory and the gradients w.r.t. its parameters. We thus propose to collect a set of intermediate gradients during training and capture their statistics to characterize the optimization trajectory. Specifically, the second-order Taylor expansion shows that: + +$$ +\begin{array}{l} f (x; \theta) = f (x; \theta^ {*}) + (\theta - \theta^ {*}) ^ {\top} \cdot \nabla_ {\theta} f (x; \theta^ {*}) \\ + \frac {1}{2} \left(\theta - \theta^ {*}\right) ^ {\top} \cdot \nabla_ {\theta} ^ {2} f (x; \theta^ {*}) \cdot \left(\theta - \theta^ {*}\right) \tag {1} \\ + \mathcal {O} \left(\left\| \theta - \theta^ {*} \right\| ^ {2}\right), \\ \end{array} +$$ + +which motivates us to model the first- and second-order derivatives (Hessian) of $f$ to represent its learning dynamics. We use the mean of gradients to indicate the first-order derivative and follow [28] to approximate the second-order derivative with gradient variances. + +# 3.3. Cross-Domain and Cross-Class Distillation + +Gradient-based optimization algorithms depend on a large number of update steps supervised by sufficient labeled data [46]. In domain-shifted and small-data regimes, those algorithms suffer from noisy update signals and overfitting, without guarantee to converge towards generalizable minima [22]. To this end, we propose to distill the training dynamics from the label-rich source domain and anchor classes to the target domain and new classes for external regularization by leveraging their intermediate gradients. We firstly compute the gradients $\pmb{g}$ for each domain and class by backpropagating the obtained losses. With a general backbone $\mathbf{F}$ for feature extraction and a task-specific predictor $\mathbf{C}$ for sample-/pixel-wise classification, the individual gradients can be expressed as: + +$$ +\boldsymbol {g} := \left[ \frac {\partial \mathcal {L} (\mathbf {C} (\mathbf {F} (x)) , y)}{\partial \theta_ {\mathbf {C} _ {1}}}, \dots , \frac {\partial \mathcal {L} (\mathbf {C} (\mathbf {F} (x)) , y)}{\partial \theta_ {\mathbf {C} _ {w}}} \right], \tag {2} +$$ + +![](images/e755981b6e4ea3bed6166a0384259073797712d738a259dd2a3c521cd3161896.jpg) +Figure 2. Overview of our proposed method. We devise the cross-domain and cross-class distillation module as well as the historical self-distillation module to transfer the model optimization knowledge from reliable sources by imposing constraints on the gradient descent steps w.r.t. the insufficiently supervised target domain and new classes. + +where $\mathcal{L}$ is the loss function, $w$ indicates the total number of parameters in $\mathbf{C}$ , $x$ and $y$ are training data and the corresponding label. For unlabeled samples $\mathbf{x}_t^u$ in the target dataset, online pseudo-labels $\tilde{\mathbf{y}}_t^u$ are used for loss calculation. Then, as shown in Fig. 2, we aggregate the individual-level gradients in a memory buffer and retrieve the domain-and class-level gradient statistics to model the global first- and second-order derivatives. According to Section 3.2, for gradients w.r.t. domain/class $\pi$ collected from a training set with $n_{\pi}$ samples, we derive their mean and variance as: + +$$ +\boldsymbol {g} _ {\pi} ^ {m} = \frac {1}{n _ {\pi}} \sum_ {i = 1} ^ {n _ {\pi}} \boldsymbol {g} _ {\pi} ^ {i}, \tag {3} +$$ + +$$ +\boldsymbol {g} _ {\pi} ^ {v} = \frac {1}{n _ {\pi} - 1} \sum_ {i = 1} ^ {n _ {\pi}} \left(\boldsymbol {g} _ {\pi} ^ {i} - \boldsymbol {g} _ {\pi} ^ {m}\right) ^ {2}. \tag {4} +$$ + +Those metrics are measured for each domain and class separately. + +However, the gradients obtained from insufficiently annotated domain and classes are dubious and error-prone [55]. To alleviate the issue, we take inspiration from recent learning theory findings that stochastic gradients along the optimization steps are generally low-rank [35], which indicates that the optimization process is governed by the top gradient eigenspace. Specifically, we devise a gradient projection approach to suppress the noisy model update signals in particular to the target domain and new classes by projecting their gradients on the reliable subspaces identified by the source domain and anchor classes with sufficient supervision. Our approach involves three iterative steps, i.e., subspace construction, gradient projection, and gradient statistics matching. Firstly, to identify the principal eigenspace, we perform Singular Value Decomposition (SVD) on the aggregated gradients from the source domain and anchor classes. We denote gradients collected from source domain, target domain, anchor class, and new + +class as $g_s$ , $g_t$ , $g_A$ , and $g_N$ . Given the sets of gradients $G_s = [g_s^1, g_s^2, \dots, g_s^K]$ and $G_A = [g_A^1, g_A^2, \dots, g_A^K]$ stored in the memory buffer of volume $K$ , we apply SVD: + +$$ +\boldsymbol {G} _ {s} = \boldsymbol {U} _ {s} \boldsymbol {\Sigma} _ {s} \boldsymbol {V} _ {s} ^ {\top}, \quad \boldsymbol {G} _ {A} = \boldsymbol {U} _ {A} \boldsymbol {\Sigma} _ {A} \boldsymbol {V} _ {A} ^ {\top}, \tag {5} +$$ + +where $U, V$ , and $\Sigma$ contain left singular vectors $u^i$ , right singular vector $v^i$ , and non-negative singular values $\sigma^i$ , respectively. To capture the overall characteristics, we use the gradients averaged along all anchor classes to represent $G_A$ . Next, we adopt the low-rank matrix approximation to select the top- $r$ significant left singular vectors in $U_s$ and $U_A$ based on the following criteria: + +$$ +\left\| \left(\boldsymbol {G} _ {s}\right) ^ {r _ {s}} \right\| _ {F} ^ {2} \geq \tau \left\| \boldsymbol {G} _ {s} \right\| _ {F} ^ {2}, \quad \left\| \left(\boldsymbol {G} _ {A}\right) ^ {r _ {A}} \right\| _ {F} ^ {2} \geq \tau \left\| \boldsymbol {G} _ {A} \right\| _ {F} ^ {2}, \tag {6} +$$ + +where $(\pmb{G})^r = \sum_{i=1}^{r} \pmb{\sigma}^i \pmb{u}^i \pmb{v}^{i^\top}$ , $\| \cdot \|_F$ denotes the Frobenius norm, $\tau$ is a threshold to ensure most information is preserved and is set to 0.98. The principal subspace and the corresponding projection matrix are thereby constructed as $M_s = [u_s^1, u_s^2, \dots, u_s^{r_s}]$ and $M_A = [u_A^1, u_A^2, \dots, u_A^{r_A}]$ . Afterward, all gradients are projected on the identified subspace as shown in Fig. 2: + +$$ +\boldsymbol {g} _ {s} ^ {*} = M _ {s} M _ {s} ^ {\top} \boldsymbol {g} _ {s}, \quad \boldsymbol {g} _ {t} ^ {*} = M _ {s} M _ {s} ^ {\top} \boldsymbol {g} _ {t}, \tag {7} +$$ + +$$ +\boldsymbol {g} _ {A} ^ {*} = \boldsymbol {M} _ {A} \boldsymbol {M} _ {A} ^ {\top} \boldsymbol {g} _ {A}, \quad \boldsymbol {g} _ {N} ^ {*} = \boldsymbol {M} _ {A} \boldsymbol {M} _ {A} ^ {\top} \boldsymbol {g} _ {N}. +$$ + +Then following Eq.(3)(4), we minimize the discrepancy between the statistics (i.e., mean and variance) of projected gradients to distill the learning dynamics from the source domain and anchor classes to the target domain and new classes. The overall training objective can be formulated as: + +$$ +\begin{array}{l} \min _ {\theta} \mathcal {L} _ {E R M} + \lambda \left\{\| \left(\boldsymbol {g} _ {s} ^ {m *} - \boldsymbol {g} _ {t} ^ {m *}\right) \| _ {2} ^ {2} + \| \left(\boldsymbol {g} _ {s} ^ {v *} - \boldsymbol {g} _ {t} ^ {v *}\right) \| _ {2} ^ {2} \right. \\ \left. + \frac {1}{n _ {\text {n e w}}} \sum_ {i = 1} ^ {n _ {\text {n e w}}} \left[ \| \left(\boldsymbol {g} _ {A} ^ {m *} - \boldsymbol {g} _ {N i} ^ {m *}\right) \| _ {2} ^ {2} + \| \left(\boldsymbol {g} _ {A} ^ {v *} - \boldsymbol {g} _ {N i} ^ {v *}\right) \| _ {2} ^ {2} \right] \right\}. \tag {8} \\ \end{array} +$$ + +Here $\mathcal{L}_{ERM}$ denotes the empirical risk loss, which is implemented with cross-entropy loss for classification and Dice loss [43] for segmentation. $\lambda$ is a balancing coefficient, $n_{new}$ is the number of new classes in the target domain. It is noted that there exists discrepancy between the anchor and new classes in semantic concepts, which makes feature alignment across classes harmful due to negative transfer [56]. However, aligning class-wise learning dynamics by enforcing the similarity between their gradients could contribute to lower empirical error on new classes. We prove this property theoretically in the supplementary material. + +# 3.4. Historical Self-distillation + +Inferior generalization behaviour is an outstanding challenge for taxonomy adaptive cross-domain adaptation [19]. When domain shifts and unseen classes exist, a trained model could only maintain its performance on test samples sharing multiple similar attributes with the training data but cannot generalize well to the disparate ones [63]. Motivated by the connection between flat minima and generalizability [31], we devise the historical self-distillation module to smooth the optimization path towards a well-generalizable solution by calibrating gradient descent steps. + +To reach flat loss landscapes [6], assuming a local loss function $\psi$ and a network $f$ parameterized by $\theta$ , instead of minimizing the original loss $\psi(f(x;\theta))$ only, the optimization objective should be: + +$$ +\min _ {\theta} - \log \int_ {\left\| \theta - \theta^ {*} \right\| \leq \Gamma} \exp (- \psi (f (x; \theta^ {*})) - \gamma \| \theta - \theta^ {*} \|) d \theta^ {*}, \tag {9} +$$ + +where $x$ indicates random input samples, $\Gamma$ is defined as the width of a local parameter valley, $\gamma$ is a balancing weight. In other words, when $\theta$ and $\theta^{*}$ are neighbouring, the following criteria is required to be satisfied: + +$$ +\left\| \nabla_ {\theta} \psi (f (x; \theta)) - \nabla_ {\theta} \psi (f (x; \theta^ {*})) \right\| \leq \beta \| \theta - \theta^ {*} \|, \tag {10} +$$ + +where $\beta$ is a smoothness factor. To this end, we propose to impose regulations on the optimizer steps by enforcing minima with uniformly-distributed gradients. Given that the gradients from most iterations are generally low-rank [35, 2], we exploit past gradients to identify the low-dimensional subspace which indicates the principal gradient distribution of local minima and then project current gradients on the constructed subspace to exclude sharp and noisy signals. Specifically, we collect and store individual gradients for each iteration along the training procedure, denoted as $G_{It} = [g_{It}^{t - T}, g_{It}^{t - T + 1}, \dots, g_{It}^{t - 2}, g_{It}^{t - 1}]$ , where $t$ is the index of current iteration, $T$ is the size of memory buffer. Thereafter we perform SVD on the set of gradients $G_{It}$ to construct the historical subspace. Similar to Eq.(5)(6)(7), the top- $r$ left singular vectors of the decomposed gradient set are concatenated to formulate the projection matrix + +$M_{It}$ . Then for upcoming gradients obtained from backpropagation in each iteration, we project the gradients on the identified low-rank subspace and consider them as residuals. The subsequent optimizer step is conducted with the sum of the original and rectified gradients, as shown in Fig. 2. Take vanilla stochastic gradient descent as an example, with mini-batch gradients $\tilde{g}$ in each iteration, the parameters $\theta$ of the optimized network are updated by: + +$$ +\theta := \theta - \frac {1}{\kappa} \cdot \eta \cdot \left(\boldsymbol {M} _ {I t} \boldsymbol {M} _ {I t} ^ {\top} + \kappa\right) \cdot \tilde {\boldsymbol {g}}, \tag {11} +$$ + +where $\eta$ is the learning rate, $\kappa$ is a trade-off parameter. As training proceeds, the principal subspace and the corresponding projection matrix are periodically renewed by applying SVD on the recently collected gradients. The overall training procedure is summarized in the supplementary material. + +# 3.5. Theoretical Analysis + +In this section, we motivate our proposed method theoretically and provide mathematical insights on its merits. + +Joint Characterization of Feature and Output Space While our optimization trajectory distillation approach proposes a novel perspective that differs from existing works imposing regularization terms in feature and model-output space [32, 19], those information is still captured and implicitly modeled in the gradient-based framework. We prove this property for the classification and segmentation tasks under two commonly adopted loss functions. The detailed proof is presented in the supplemental material. It illustrates the superiority of our method as a unified framework that jointly characterizes the feature and output space as well as the learning dynamics. + +Impacts on Generalization Error In our method, we propose to minimize the discrepancy between the gradient statistics from different domains and classes in the identified principal subspace. We hereby prove its effectiveness towards a tighter generalization error bound on the target domain and new classes [3]. Details can be found in the supplemental material. + +# 4. Experiments + +# 4.1. Datasets and Experiment Settings + +To evaluate the effectiveness of our method and its potential clinical implications, we conduct experiments on five medical image datasets in regard to three important downstream tasks from different clinical scenarios. Extended experiments on more diverse tasks, including radiology and fundus analysis, as well as general visual task, are presented in the supplemental material to substantiate the broader applicability of our method. + +Table 1. Comparison results of our proposed method against other state-of-the-art methods for nuclei segmentation and recognition. The metrics with asterisk (*) are calculated on target-private new classes only. The subscripts denote the standard deviations. The best and second-best results are highlighted in bold and brown, respectively. + +
Methods5-shot10-shot
mF1mF1*mPQmPQ*mF1mF1*mPQmPQ*
Sup-only28.080.8921.460.3612.350.5312.230.4530.900.6625.230.8314.180.2414.940.32
Multi-task [48]33.851.2618.151.0818.580.4210.770.3535.150.9421.290.6219.140.2712.890.34
DANN [16]32.761.4116.521.1018.310.6610.400.5935.761.3523.120.8819.700.6014.140.42
CGDM [14]35.031.2821.040.9919.590.5013.220.3237.131.5724.221.1519.780.7113.900.40
LETR [40]27.781.4415.361.0216.680.5410.030.5134.851.5420.030.7318.180.5911.790.27
FT-CIDA [32]30.510.9216.550.8015.970.399.420.4032.631.2121.080.9517.090.4711.960.49
STARTUP [45]37.030.5822.440.7620.400.3313.820.3040.850.5025.920.9123.370.1916.640.24
DDN [25]37.920.8623.711.0520.620.5713.870.4340.160.7226.941.0822.790.3216.820.50
TSA [34]34.031.1719.551.1118.650.4312.060.3734.271.0522.581.2019.780.6114.240.73
TACS [19]36.731.6522.261.2918.130.8012.570.6239.291.4825.841.1322.270.6815.980.56
Ours40.260.9027.140.9721.780.4914.960.2243.880.5231.430.7824.810.2619.350.33
+ +Nuclei Segmentation and Recognition Technically, this task can be formalized as a simultaneous instance segmentation and classification problem [21]. We evaluate our method by considering PanNuke [15] and Lizard [20] as the source and target datasets. They contain 481 and 291 visual fields cropped from whole-slide images with 189,744 and 495,179 annotated nuclei, respectively. There are six types of nuclei in Lizard and only two of them overlap with the five classes in PanNuke, which suggests four new classes in the target domain. We employ the widely used HoverNet [21] architecture with a standard ResNet-50 backbone as the base model. For quantitative evaluation, we follow [20] and report the average F1 and PQ (panoptic quality) scores for all classes (denoted as mF1 and mPQ) to provide a comprehensive assessment measuring the accuracy of nuclei detection, segmentation, and classification. Additionally, we also report the mF1* and mPQ* scores which are solely computed on new classes in the target domain. + +Cancer Tissue Phenotyping In this setting, we perform adaptation from a two-class discrimination task to an eight-class one on CRC-TP [29] and Kather [30] datasets. Kather consists of $5000 \times 150$ pathology image patches and includes six novel classes other than the two common classes also existing in the source domain. The experiments are conducted with ResNet-101 as backbone. For evaluation, we exploit the accuracy and F1 scores as metrics to indicate the classification performance. The average scores for all classes (mAcc and mF1) and particularly the ones on new classes only (mAcc* and mF1*) are reported. + +Skin Lesion Diagnosis We construct a fine-grained taxonomy adaptation setting from two coarse classes to seven subclasses with HAM10000 [52], which contains 10015 dermatoscopic images sampled from different body structures. The adopted experiment settings and evaluation metrics are the same as the cancer tissue phenotyping task. Since the label space of the source and target datasets do + +not overlap, we only report the mAcc* and mF1* scores which are computed on target classes. + +Following previous works in UDA [7], for all experiments, each dataset is randomly split into $80\% / 20\%$ for training/testing. To explicitly recognize the target-private new classes, we sample few (5/10) samples with corresponding labels to formulate the support set. The remaining target data is left unlabeled. More details can be found in the supplemental material. + +# 4.2. Results + +We compare our method with state-of-the-art approaches for cross-domain adaptation, including UDA methods DANN [16] and CGDM [14] which do not consider the taxonomic inconsistency, as well as LETR [40], FT-CIDA[32], STARTUP [45], DDN [25], TSA [34], and TACS [19] which are designed for both domain shifts and new class learning. For UDA methods, we jointly perform domain alignment and train the model with the labeled target data. In addition, we conduct comparisons with multi-task learning approach [48], which trains the model on all domains and label sets simultaneously. "Sup-only" indicates the baseline where only the supervised loss on the labeled target dataset is used for training. + +Table 1 shows the comparison results on the cross-domain nuclei recognition and segmentation task. All methods adopt the same ResNet-50-HoverNet as base model following [21]. It can be observed that our proposed method outperforms the state-of-the-art cross-domain adaptation approaches by a large margin. We also notice an important point that compared with the "Sup-only" baseline, most competing methods fail to attain better performance on the metrics specific to new classes (i.e., mF1* and mPQ*). For example, under the 5-shot setting, despite a $4.68\%$ improvement is achieved by DANN on mF1, it achieves inferior performance on mF1* which is $4.94\%$ worse than "Sup-only". + +Table 2. Comparison results of our proposed method against other state-of-the-art methods for cancer tissue phenotyping. + +
Methods5-shot10-shot
mAccmAcc*mF1mF1*mAccmAcc*mF1mF1*
Sup-only46.821.3042.261.2644.641.4942.831.0848.521.0245.931.6745.341.4546.151.83
Multi-task [48]48.350.7842.021.3445.551.1241.681.6348.560.6945.001.7246.030.6444.461.21
DANN [16]47.460.6742.531.2243.081.1443.131.3048.420.4845.311.2746.500.8144.020.64
CGDM [14]49.020.8243.531.5646.290.8743.580.6950.540.8444.571.3548.170.6645.230.40
LETR [40]45.471.9541.331.4743.171.8841.202.0245.940.5640.131.3041.161.1339.610.90
FT-CIDA [32]43.720.8839.060.7940.911.0238.730.8444.650.5338.800.9440.940.9638.961.07
STARTUP [45]49.520.6342.931.8343.591.3739.970.5949.400.5146.911.8746.380.7047.851.36
DDN [25]47.141.7543.250.9942.922.2940.760.5048.270.6744.081.2246.031.1445.441.18
TSA [34]48.221.4942.130.7844.391.8441.921.1347.181.0241.261.1045.341.7244.691.89
TACS [19]45.301.0740.281.1144.741.2642.040.9546.951.1643.691.7447.461.5147.511.25
Ours50.140.7647.661.5447.411.1046.451.1555.860.9851.271.0252.980.4952.120.94
+ +Table 3. Comparison results of our proposed method against other state-of-the-art methods for skin lesion diagnosis. + +
Methods5-shot10-shot
mAcc*mF1*mAcc*mF1*
Sup-only35.710.7919.890.2238.500.3822.360.67
Multi-task [48]33.801.2019.060.3137.850.4023.320.32
DANN [16]37.220.5723.990.9240.060.9424.320.72
CGDM [14]41.501.0324.260.8047.220.7024.000.23
LETR [40]36.040.4221.390.6640.871.2223.500.49
FT-CIDA [32]31.950.7718.130.4834.600.9020.960.88
STARTUP [45]43.211.1527.170.9243.881.3525.590.60
DDN [25]36.191.0521.500.3442.461.2923.920.28
TSA [34]40.411.3624.051.0142.071.5125.441.06
TACS [19]45.050.4625.600.6546.730.6425.960.79
Ours50.790.6930.050.8152.140.9830.520.83
+ +This phenomenon is incurred by the failure of cross-class generalization. In contrast, our method yields substantial improvements on those new class-specific metrics, which indicates the superior generalizability of our method. The observation is further verified through visual comparisons in the supplementary material. + +The overall quantitative results on the cancer tissue phenotyping and skin lesion diagnosis tasks are presented in Table 2 and 3, respectively. All methods are implemented based on the ResNet-101 backbone. Under each few-shot setting, our method shows higher classification scores compared with previous approaches. In particular, we achieve $0.62\% \sim 5.32\%$ and $1.12\% \sim 4.81\%$ improvements against the second-best results in terms of mAcc and mF1 for cancer tissue phenotyping, as well as $4.92\% \sim 7.58\%$ and $2.88\% \sim 4.93\%$ improvements in terms of mAcc* and mF1* for skin lesion diagnosis. The higher accuracy consistently attained by our method under different cross-domain adaptation settings and clinical tasks is attributed to the strong generalizability brought by the proposed optimization trajectory distillation approach. + +# 4.3. Analysis and Discussion + +Ablation of Key Components Table 4 shows the ablation study results to illustrate the effectiveness of key components. "projection" denotes the gradient projection approach introduced in Section 3.3. For the "w/o projection" ablation, we perform cross-domain and cross-class distillation by directly minimizing the statistics discrepancy between unrectified gradients $(g_{s}, g_{t})$ and $(g_{A}, g_{N})$ in Eq. (7). It is observed that each component is indispensable and could improve the overall performance, which elaborates the importance of external navigation for learning target domain and new classes in network training. Interestingly, we notice that the impacts of cross-domain and cross-class distillation could be disparate across tasks. For example, cross-class distillation brings higher performance gain compared with cross-domain distillation on the cancer tissue phenotyping benchmark, while the situation is reversed for skin lesion diagnosis. It is caused by the different natures of each task. Class-imbalance is a serious issue for skin lesion diagnosis in that the class distribution is highly biased [52], which compromises the effectiveness of the cross-class distillation module. Differently, for cancer tissue phenotyping, classes are uniformly distributed [30] and could hence ensure the representativity of anchor classes and their exemplarity to be distilled to new classes. + +Hyperparameter Sensitivity To investigate the effect of hyperparameters $\lambda$ in Eq. (8) and $\kappa$ in Eq. (11), we present the performance comparison by varying the choices of their values for cancer tissue phenotyping, as shown in Fig. 3. We only change one hyperparameter while keeping the other fixed at a time. The experimental results show that our framework is quite stable for hyperparameters in a wide interval, yet setting $\lambda$ and $\kappa$ to a large value is detrimental to the classification accuracy. Specifically, $\lambda$ is a balancing coefficient between class discriminativeness and gradient alignment. When it is too large, the trained model is biased to matching learning dynamics and would lose es + +Table 4. Key component analysis of our method on three cross-domain adaptation benchmarks under 10-shot scheme. The best results are highlighted in bold. + +
MethodsNucleiCancerSkin
mF1mF1*mPQmPQ*mAccmAcc*mF1mF1*mAcc*mF1*
Full Framework43.8831.4324.8119.3555.8651.2752.9852.1252.1430.52
w/o cross-domain39.9630.0822.2918.1052.8149.9351.6250.6444.3425.28
w/o cross-class42.3027.8223.7617.0548.9047.2249.8544.7649.0628.24
w/o historical41.5430.4723.0017.5950.1748.0548.1945.1245.8325.97
w/o projection40.0628.8022.1017.3153.5845.7849.0149.2047.2928.02
+ +![](images/d34469907c41fdddcf838ca7bce695303cbcbba49b1bdf427902020e289eccf9.jpg) +Figure 3. Performance comparison between different choices of hyperparameters $\lambda$ and $\kappa$ on the cancer tissue phenotyping benchmark under 10-shot scheme. + +sential semantic knowledge. For $\kappa$ , it controls the proportions of unrectified gradients. Its large value indicates that most original gradients are preserved, which would incur noisy optimization steps and loss of flatness. Therefore, we decide to set $\lambda = 10$ and $\kappa = 100$ . It achieves the best performance under most metrics. + +Impacts on Optimization Trajectory and Model Flatness To better understand the effectiveness of our method, we perform in-depth analysis about its impacts on model's training trajectory and flatness of the converged local minima. Firstly, as shown in Fig. 4(a), we visualize the optimization paths of our method and the multi-task baseline on the test accuracy surface with a contour plot. The details of plot procedure can refer to [26]. It can be observed that the optimization of the multi-task baseline fluctuates sharply and is restricted in a sub-optimal area with a relatively low test accuracy. This is induced by the biased navigation in the course of training due to domain shifts and category gap. With the devised dual-stream optimization trajectory distillation module, our method could by contrast suppress the noisy update signals and proceed consistently towards the optimum. + +Besides, we perform analysis about the solutions found by our method and the baseline in terms of flatness [5]. Specifically, the local flatness of a model parameter $\theta$ is quantified by the expected changes of test accuracy between $\theta$ and a neighbouring $\theta^{*}$ with perturbation rate $\varrho$ : $\mathcal{F}_{\varrho}(\theta) = \mathbb{E}_{\|\theta^{*}\| = (1 + \varrho)\|\theta\|}\left[\mathrm{Acc}(\theta^{*}) - \mathrm{Acc}(\theta)\right]$ . We approximate the expected value by Monte-Carlo sampling over 50 samples. As demonstrated in Fig. 4(b), comparing with the + +![](images/14d88eb8312627df860f13a9fe26421d4d73df2681158058767a931e55e216bf.jpg) +(a) Visualization of Optimization Trajectory + +![](images/3014a8f72be3ae2f65d861f70aa9c8d2a8a4e91eaeb92f6de3b179fc94c3f87a.jpg) +(b) Flatness Analysis +Figure 4. Further analysis of our proposed method. (a) We plot the optimization trajectory of our method and the multi-task baseline on the test accuracy surface. (b) We visualize the local flatness of our method and the baseline via test accuracy gap by varying the model parameter perturbation rate. Each point is computed by Monte-Carlo approximation over 50 random samples. All the experiments are conducted on the cancer tissue phenotyping benchmark under 10-shot scheme. + +baseline approach, our method shows stronger robustness against model parameter perturbation. It proves that our method successfully converges to minima with better local flatness and thus possesses superior generalizability [31]. + +# 5. Conclusion + +In this paper, we study the taxonomy adaptive cross-domain adaptation paradigm, which allows incoherent label sets between source and target datasets. To jointly address the data distribution bias and category gap, we propose a unified framework which performs cross-domain and cross-class optimization trajectory distillation to calibrate the learning dynamics of insufficiently annotated domain and classes. Historical self-distillation is further devised to attain a well-generalizable solution via regulating gradient distributions. Extensive experiments on multiple benchmarks with different task contexts demonstrate the effectiveness and generalization capability of our method. In the future, we will extend this work to further discover the underlying mechanism of optimization trajectory distillation for model generalization and evaluate its potential in more adaptation scenarios. + +# References + +[1] Sanjeev Arora, Rong Ge, Yingyu Liang, Tengyu Ma, and Yi Zhang. Generalization and equilibrium in generative adversarial nets (gans). In International Conference on Machine Learning, pages 224-232. PMLR, 2017. +[2] Sheikh Shams Azam, Seyyedali Hosseinalipour, Qiang Qiu, and Christopher Brinton. Recycling model updates in federated learning: Are gradient subspaces low-rank? In International Conference on Learning Representations, 2022. +[3] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006. +[4] Debora Calderola, Barbara Caputo, and Marco Ciccone. Improving generalization in federated learning by seeking flat minima. In European Conference on Computer Vision, pages 654-672. Springer, 2022. +[5] Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Seunghyun Park, Yunsung Lee, and Sungrae Park. Swad: Domain generalization by seeking flat minima. Advances in Neural Information Processing Systems, 34:22405-22418, 2021. +[6] Pratik Chaudhari, Anna Choromanska, Stefano Soatto, Yann LeCun, Carlo Baldassi, Christian Borgs, Jennifer Chayes, Levent Sagun, and Riccardo Zecchina. Entropy-SGD: Biasing gradient descent into wide valleys. In International Conference on Learning Representations, 2017. +[7] Cheng Chen, Qi Dou, Hao Chen, Jing Qin, and Pheng-Ann Heng. Synergistic image and feature adaptation: Towards cross-modality domain adaptation for medical image segmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 865–872, 2019. +[8] Wentao Chen, Zhang Zhang, Wei Wang, Liang Wang, Zilei Wang, and Tieniu Tan. Cross-domain cross-set few-shot learning via learning compact and aligned representations. In European Conference on Computer Vision, pages 383-399. Springer, 2022. +[9] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. +[10] Ali Cheraghian, Shafin Rahman, Sameera Ramasinghe, Pengfei Fang, Christian Simon, Lars Petersson, and Mehrtash Harandi. Synthesized feature based few-shot class incremental learning on a mixture of subspaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8661-8670, 2021. +[11] Debasmit Das, Sungrack Yun, and Fatih Porikli. ConfeSS: A framework for single source cross-domain few-shot learning. In International Conference on Learning Representations, 2022. +[12] Susanna Dodd, Mike Clarke, Lorne Becker, Chris Mavergames, Rebecca Fish, and Paula R Williamson. A taxonomy has been developed for outcomes in medical research to help improve knowledge discovery. Journal of clinical epidemiology, 96:84-92, 2018. + +[13] C Dominguez Conde, C Xu, LB Jarvis, DB Rainbow, SB Wells, T Gomes, SK Howlett, O Suchanek, K Polanski, HW King, et al. Cross-tissue immune cell analysis reveals tissue-specific features in humans. Science, 376(6594), 2022. +[14] Zhekai Du, Jingjing Li, Hongzu Su, Lei Zhu, and Ke Lu. Cross-domain gradient discrepancy minimization for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3937-3946, 2021. +[15] Jevgenij Gamper, Navid Alemi Koohbanani, Ksenija Benet, Ali Khuram, and Nasir Rajpoot. Pannuke: an open pancancer histology dataset for nuclei instance segmentation and classification. In European congress on digital pathology, pages 11-19. Springer, 2019. +[16] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. +[17] Zhiqiang Gao, Shufei Zhang, Kaizhu Huang, Qiufeng Wang, and Chaoliang Zhong. Gradient distribution alignment certificates better adversarial domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8937-8946, 2021. +[18] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020. +[19] Rui Gong, Martin Danelljan, Dengxin Dai, Danda Pani Paudel, Ajad Chhatkuli, Fisher Yu, and Luc Van Gool. Tacs: Taxonomy adaptive cross-domain semantic segmentation. In European Conference on Computer Vision, pages 19-35. Springer, 2022. +[20] Simon Graham, Mostafa Jahanifar, Ayesha Azam, Mohammed Nimir, Yee-Wah Tsang, Katherine Dodd, Emily Hero, Harvir Sahota, Atisha Tank, Ksenija Benes, et al. Lizard: a large-scale dataset for colonic nuclear instance segmentation and classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 684-693, 2021. +[21] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:101563, 2019. +[22] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In European conference on computer vision, pages 124-141. Springer, 2020. +[23] Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural computation, 9(1):1-42, 1997. +[24] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In European Conference on Computer Vision, pages 20-37. Springer, 2022. +[25] Ashraful Islam, Chun-Fu Richard Chen, Rameswar Panda, Leonid Karlinsky, Rogerio Feris, and Richard J Radke. Dy + +namic distillation network for cross-domain few-shot recognition with unlabeled data. Advances in Neural Information Processing Systems, 34:3584-3595, 2021. +[26] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. Proceedings of the Conference on Uncertainty in Artificial Intelligence, 2018. +[27] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. +[28] Stanisław Jastrzebski, Zachary Kenton, Devansh Arpit, Nicolas Ballas, Asja Fischer, Yoshua Bengio, and Amos Storkey. Three factors influencing minima in sgd. arXiv preprint arXiv:1711.04623, 2017. +[29] Sajid Javed, Arif Mahmood, Muhammad Moazam Fraz, Navid Alemi Koohbanani, Ksenija Benes, Yee-Wah Tsang, Katherine Hewitt, David Epstein, David Snead, and Nasir Rajpoot. Cellular community detection for tissue phenotyping in colorectal cancer histology images. Medical image analysis, 63:101696, 2020. +[30] Jakob Nikolas Kather, Cleo-Aron Weis, Francesco Bianconi, Susanne M Melchers, Lothar R Schad, Timo Gaiser, Alexander Marx, and Frank Gerrit Zöllner. Multi-class texture analysis in colorectal cancer histology. Scientific reports, 6(1):1-11, 2016. +[31] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. In International Conference on Learning Representations, 2017. +[32] Jogendra Nath Kundu, Rahul Mysore Venkatesh, Naveen Venkat, Ambareesh Revanur, and R Venkatesh Babu. Class incremental domain adaptation. In European Conference on Computer Vision, pages 53-69. Springer, 2020. +[33] Canran Li, Dongnan Liu, Haoran Li, Zheng Zhang, Guangming Lu, Xiaojun Chang, and Weidong Cai. Domain adaptive nuclei instance segmentation and classification via category-aware feature alignment and pseudo-labelling. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 715-724. Springer, 2022. +[34] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Cross-domain few-shot learning with task-specific adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7161-7170, 2022. +[35] Xinyan Li, Qilong Gu, Yingxue Zhou, Tiancong Chen, and Arindam Banerjee. Hessian based analysis of sgd for deep nets: Dynamics and generalization. In Proceedings of the 2020 SIAM International Conference on Data Mining, pages 190-198. SIAM, 2020. +[36] Hanwen Liang, Qiong Zhang, Peng Dai, and Juwei Lu. Boosting the generalization capability in cross-domain few-shot learning via noise-enhanced supervised autoencoder. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9424-9434, 2021. + +[37] Dongnan Liu, Chaoyi Zhang, Yang Song, Heng Huang, Chenyu Wang, Michael Barnett, and Weidong Cai. Decompose to adapt: Cross-domain object detection via feature disentanglement. IEEE Transactions on Multimedia, 25:1333-1344, 2022. +[38] Dongnan Liu, Donghao Zhang, Yang Song, Fan Zhang, Lauren O'Donnell, Heng Huang, Mei Chen, and Weidong Cai. Unsupervised instance segmentation in microscopy images via panoptic domain adaptation and task re-weighting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4243-4252, 2020. +[39] Xu Luo, Jing Xu, and Zenglin Xu. Channel importance matters in few-shot image classification. In International Conference on Machine Learning, pages 14542-14559. PMLR, 2022. +[40] Zelun Luo, Yuliang Zou, Judy Hoffman, and Li F Fei-Fei. Label efficient learning of transferable representations acrossss domains and tasks. Advances in neural information processing systems, 30, 2017. +[41] Deborah Lupton. Medicine as culture: illness, disease and the body. Sage, 2012. +[42] Lucas Mansilla, Rodrigo Echeveste, Diego H Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6630-6638, 2021. +[43] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. +[44] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In Proceedings of the IEEE international conference on computer vision, pages 754-763, 2017. +[45] Cheng Perng Phoo and Bharath Hariharan. Self-training for few-shot transfer across extreme task differences. In International Conference on Learning Representations, 2021. +[46] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations, 2017. +[47] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, Trevor Darell, and Kate Saenko. Semi-supervised domain adaptation via minimax entropy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8050-8058, 2019. +[48] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018. +[49] Yuge Shi, Jeffrey Seely, Philip Torr, Siddharth N, Awni Hannun, Nicolas Usunier, and Gabriel Synnaeve. Gradient matching for domain generalization. In International Conference on Learning Representations, 2022. +[50] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. +[51] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011. +[52] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source der + +matoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018. +[53] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In International Conference on Learning Representations, 2020. +[54] Qian Wang and Toby Breckon. Unsupervised domain adaptation via structured prediction based selective pseudolabeling. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 6243-6250, 2020. +[55] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020. +[56] Zirui Wang, Zihang Dai, Barnabás Póczos, and Jaime Carbonell. Characterizing and avoiding negative transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11293-11302, 2019. +[57] Julia K Winkler, Christine Fink, Ferdinand Toberer, Alexander Enk, Teresa Deinlein, Rainer Hofmann-Wellenhof, Luc Thomas, Aimilios Lallas, Andreas Blum, Wilhelm Stolz, et al. Association between surgical skin markings in dermoscopic images and diagnostic performance of a deep learning convolutional neural network for melanoma recognition. JAMA dermatology, 155(10):1135-1141, 2019. +[58] Mengya Xu, Mobarakol Islam, Chwee Ming Lim, and Hongliang Ren. Class-incremental domain adaptation with smoothing and calibration for surgical report generation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 269-278. Springer, 2021. +[59] Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Universal domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2720-2729, 2019. +[60] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. +[61] Wang Yuan, Zhizhong Zhang, Cong Wang, Haichuan Song, Yuan Xie, and Lizhuang Ma. Task-level self-supervision for cross-domain few-shot learning. Proceedings of the AAAI Conference on Artificial Intelligence, 36(3):3215-3223, 2022. +[62] Qiming Zhang, Jing Zhang, Wei Liu, and Dacheng Tao. Category anchor-guided unsupervised domain adaptation for semantic segmentation. Advances in neural information processing systems, 32, 2019. +[63] Lin Zuo, Mengmeng Jing, Jingjing Li, Lei Zhu, Ke Lu, and Yang Yang. Challenging tough samples in unsupervised domain adaptation. Pattern Recognition, 110:107540, 2021. \ No newline at end of file diff --git a/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/images.zip b/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..02255ba552c0f3e1aad8240ccd9ab270a9b103ce --- /dev/null +++ b/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60f44a1a6b618ef97fd99a771c9cd03fb6caca9a024f64e942891042885b0ac3 +size 568786 diff --git a/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/layout.json b/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e4aa779a84bce3407ba281fff2aa346dfdbf5380 --- /dev/null +++ b/2023/Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation/layout.json @@ -0,0 +1,8630 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 79, + 103, + 515, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 103, + 515, + 140 + ], + "spans": [ + { + "bbox": [ + 79, + 103, + 515, + 140 + ], + "type": "text", + "content": "Taxonomy Adaptive Cross-Domain Adaptation in Medical Imaging via Optimization Trajectory Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "spans": [ + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "type": "text", + "content": "Jianan Fan" + }, + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "type": "text", + "content": ", Dongnan Liu" + }, + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "type": "text", + "content": ", Hang Chang" + }, + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "type": "text", + "content": ", Heng Huang" + }, + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "type": "text", + "content": ", Mei Chen" + }, + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "type": "text", + "content": ", and Weidong Cai" + }, + { + "bbox": [ + 83, + 156, + 509, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 179, + 539, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 179, + 539, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 539, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 55, + 179, + 539, + 194 + ], + "type": "text", + "content": "University of Sydney " + }, + { + "bbox": [ + 55, + 179, + 539, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 55, + 179, + 539, + 194 + ], + "type": "text", + "content": "Lawrence Berkeley National Laboratory " + }, + { + "bbox": [ + 55, + 179, + 539, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 55, + 179, + 539, + 194 + ], + "type": "text", + "content": "University of Maryland at College Park " + }, + { + "bbox": [ + 55, + 179, + 539, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 55, + 179, + 539, + 194 + ], + "type": "text", + "content": "Microsoft" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 194, + 492, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 194, + 492, + 204 + ], + "spans": [ + { + "bbox": [ + 104, + 194, + 492, + 204 + ], + "type": "text", + "content": "jfan6480@uni.sydney.edu.au dongnan.liu@sydney.edu.au hchang@lbl.gov" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 204, + 488, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 204, + 488, + 216 + ], + "spans": [ + { + "bbox": [ + 104, + 204, + 488, + 216 + ], + "type": "text", + "content": "henghuanghh@gmail.com Mei.Chen@microsoft.com tom.cai@sydney.edu.au" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 251, + 192, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 251, + 192, + 262 + ], + "spans": [ + { + "bbox": [ + 143, + 251, + 192, + 262 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 275, + 290, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 290, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 290, + 612 + ], + "type": "text", + "content": "The success of automated medical image analysis depends on large-scale and expert-annotated training sets. Unsupervised domain adaptation (UDA) has been raised as a promising approach to alleviate the burden of labeled data collection. However, they generally operate under the closed-set adaptation setting assuming an identical label set between the source and target domains, which is over-restrictive in clinical practice where new classes commonly exist across datasets due to taxonomic inconsistency. While several methods have been presented to tackle both domain shifts and incoherent label sets, none of them take into account the common characteristics of the two issues and consider the learning dynamics along network training. In this work, we propose optimization trajectory distillation, a unified approach to address the two technical challenges from a new perspective. It exploits the low-rank nature of gradient space and devises a dual-stream distillation algorithm to regularize the learning dynamics of insufficiently annotated domain and classes with the external guidance obtained from reliable sources. Our approach resolves the issue of inadequate navigation along network optimization, which is the major obstacle in the taxonomy adaptive cross-domain adaptation scenario. We evaluate the proposed method extensively on several tasks towards various endpoints with clinical and open-world significance. The results demonstrate its effectiveness and improvements over previous methods. Code is available at https://github.com/camwew/TADA-MI." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 633, + 128, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 633, + 128, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 633, + 128, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Automated and objective analysis of medical images is an important research topic and has been explored in various clinical applications [57, 13]. To relieve the burden of acquiring massive annotated data for model development on new domains, several unsupervised domain adapta" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 251, + 545, + 426 + ], + "blocks": [ + { + "bbox": [ + 310, + 251, + 545, + 426 + ], + "lines": [ + { + "bbox": [ + 310, + 251, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 310, + 251, + 545, + 426 + ], + "type": "image", + "image_path": "8880daf55c116dc96f9c49849b0c6e7f4a517c78d336058b001a0e3aed7c4d52.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 430, + 547, + 519 + ], + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 519 + ], + "type": "text", + "content": "Figure 1. (a) Illustration of the issue of taxonomic inconsistency with nuclei recognition. The category label sets are incoherent across datasets. Different colours indicate the class each nucleus belongs to. The red asterisks denote novel classes only existing in Dataset 2. (b) Concept of the proposed method. We perform optimization trajectory distillation to deliver external navigation for learning target domain and new classes where the optimization steps tend to be restricted and unreliable." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 543, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 546, + 581 + ], + "type": "text", + "content": "tion (UDA) methods [7, 38, 37] are proposed to mitigate the data distribution bias [51] between a richly labeled source domain and a target domain with no explicit supervision." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "However, these works assume a closed-set adaptation setting that the source and target domains should necessarily share the identical category label space and definition [44]. The restriction limits their practical applicability in the clinical wild [15]. Unlike natural images where the definitions of different entity categories (e.g., cat v.s. dog) are structured and globally unified, in the medical domain, inconsistency in taxonomy across different countries or institutes is a common issue compromising the feasibility of cross-domain adaptation [12]. Take nuclei recognition in histology images as an example, the ambiguous biologi" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21174" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 203 + ], + "type": "text", + "content": "cal characteristics of cells and distinct clinical usages of the analysis outcomes result in a lack of a unified categorization schema [15, 20]. With specific clinical purposes and downstream applications, different datasets could categorize nuclei with disparate granularities of biological concepts. Besides, as medical research evolves rapidly, novel cell and disease classes could be discovered and form datasets with continually expanding category label sets [41]. This motivates us to develop a generalized adaptation method with the capability to tackle both data distribution bias and category gap for more flexible UDA in clinical practice." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 206, + 286, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 206, + 286, + 469 + ], + "spans": [ + { + "bbox": [ + 49, + 206, + 286, + 469 + ], + "type": "text", + "content": "To this end, we study the problem of taxonomy adaptive domain adaptation [19]. It assumes the target domain could adopt a label space different from the source domain. Following UDA, a labeled source dataset and unlabeled samples from the target domain are available during training. Additionally, to recognize the novel/fine-grained categories which are non-existent or unspecified in the source domain, a few samples of those target-private categories are annotated and utilized as exemplars. The technical challenge for this setting lies in how to alleviate domain shifts and concurrently learn new classes with very limited supervision. Recently, several methods are proposed towards a similar goal [32, 22]. However, they typically address the two issues individually in separate contexts and fail to design a unified paradigm according to their common characteristics [19], which could incur a subtle collision across issues since their objectives are non-relevant [60]. Besides, existing works either focus on cross-domain/class alignment in the feature space [40, 36] or resort to model output-based regularization such as self-supervision [45], whereas those approaches suffer from the equilibrium challenge of adversarial learning [1] and the low-quality pseudo-labels [54]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 472, + 286, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 472, + 286, + 662 + ], + "spans": [ + { + "bbox": [ + 49, + 472, + 286, + 662 + ], + "type": "text", + "content": "In this work, we present a unified framework for taxonomy adaptive cross-domain adaptation from a new perspective, i.e., via optimization trajectory distillation. Our strategy is motivated by a common challenge existing in both cross-domain and small-data learning regimes, which is the inadequate navigation in the course of network optimization. For cross-domain learning, the unstable feature alignment procedure and noisy pseudo-labels tend to induce error accumulation along network training [62]. Similarly, with limited support samples in new class learning, the optimization of network is inclined to step towards restricted local minima and cannot cover a globally-generalizable space [55]. To this end, we propose to exploit the optimization trajectory from a reliable \"teacher\" to provide external guidance for regularizing the learning dynamics of insufficiently annotated domain and classes, as illustrated in Fig. 1(b)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 665, + 286, + 712 + ], + "type": "text", + "content": "Our method consists of two key components, i.e., cross-domain/class distillation and historical self-distillation. (i) Cross-domain and cross-class distillation aim to leverage the optimization trajectory from the richly labeled do" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 72, + 544, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 334 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 334 + ], + "type": "text", + "content": "main and classes to serve as the \"teacher\". Motivated by the Neural Tangent Kernel (NTK) theory [27], we characterize the network optimization trajectory through gradient statistics. Then, given the observation that the subspace spanned by the gradients from most iterations is generally low-rank [35, 2], we design a gradient projection approach to suppress the noisy signals in stochastic gradients and rectify the distillation process. Thereafter constraints are imposed on the gradient statistics of target domain and new classes to calibrate their training dynamics towards domain-invariant and unbiased learning procedure. (ii) Historical self-distillation further drives the optimization paths of model to converge towards flat minima. It is found that the flatness of loss landscapes is strongly related to the model's robustness and generalizability [26], while how to take advantage of the insight to tackle domain shifts and limited supervision is under-explored. We propose to exploit the historical gradients to construct the informative low-rank subspaces and then perform gradient projection to alleviate loss sharpness. It compensates for the intense and out-of-order optimization updates incurred by inadequate regularization and leads to better generalization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 335, + 544, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 335, + 544, + 478 + ], + "spans": [ + { + "bbox": [ + 308, + 335, + 544, + 478 + ], + "type": "text", + "content": "Our prime contributions are as follows: (1) We introduce a more generalized cross-domain adaptation paradigm for medical image analysis in which both data distribution bias and category gap exist across the source and target domains. (2) We leverage insights from recent learning theory research and propose a novel dual-stream optimization trajectory distillation method to provide external navigation in network training. We perform theoretical justifications from two perspectives to illustrate the merits of our method. (3) Experiments on various benchmarks validate the effectiveness and robustness of our proposed method and its improvements over existing approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 490, + 395, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 395, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 395, + 502 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 510, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 544, + 712 + ], + "type": "text", + "content": "Domain Adaptation Domain adaptation (DA) aims to mitigate the data distribution bias [16]. The classic semi-supervised/unsupervised DA focuses on the scenario where the target domain has an identical label space to the source domain [47, 14, 33]. The over-restricted precondition cannot suffice the demand of clinical usages [58]. There exist several efforts to take a step further than the closed-set adaptation, such as open-set DA [44] and universal DA [59]. However, the aim of those works is solely to detect the previously unseen classes, instead of learning to separately identify each new class with few supports. Some recent studies suggest to explicitly recognize the target-private categories. [32] performs feature projection and alignment based on the prototypical networks [50]. [19] combines pseudo-labeling and contrastive learning to combat domain bias and label gap. Nevertheless, those works are limited to tackling the two issues individually, which fails to exploit" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21175" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 240 + ], + "type": "text", + "content": "their common characteristics and propose a unified solution. Cross-Domain Few-Shot Learning The technical challenge we need to resolve is similar to an emerging research topic, i.e., cross-domain few-shot learning [22]. The mainstream methods for this challenge include self-training [45, 25, 61], contrastive learning [11, 19], feature alignment/augmentation [53, 8, 24], and channel transformation [34, 39]. Those methods either focus on feature-level modulation or turn to output-level self-supervision and finetuning, which could bring instability along optimization [1, 54]. In this work, we propose a novel gradient-based framework to transfer knowledge in both cross-domain and few-shot settings. It implicitly characterizes the information in both feature space and output space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 241, + 286, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 286, + 384 + ], + "type": "text", + "content": "Gradient Manipulation In current deep learning systems where gradient descent serves as the most popular training algorithm, the directions and amplitudes of network optimization steps are embedded in the gradient space [9]. Gradient manipulation refers to directly modulating the optimization gradients for improved learning process [42]. Recently, it is introduced to mitigate the data distribution shift and shows promises in UDA and domain generalization [14, 17, 49]. However, those works are still restricted by the closed-set condition and do not consider the low-rank nature of gradients [35], which make their approaches vulnerable to noises in the small-data learning regime." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 386, + 286, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 286, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 286, + 515 + ], + "type": "text", + "content": "Flatness in Optimization In previous machine learning research, the connection between convergence to flat minima in optimization and model generalizability has long been established [23, 31]. Theoretical and empirical studies prove that the flatness of loss landscapes highly correlates with generalization capability [4]. [5] averages the network weights of multiple checkpoints to prompt generalization across domains. Different from their post-hoc averaging approach operating on fixed models, we propose to regularize the intermediate dynamics of network optimization to achieve flat minima." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 530, + 102, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 102, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 102, + 542 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 551, + 160, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 160, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 160, + 562 + ], + "type": "text", + "content": "3.1. Problem Statement" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": "The taxonomy adaptive cross-domain adaptation problem can be formalized as follows. There exists a labeled source dataset " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s = \\{(\\mathbf{x}_s,\\mathbf{y}_s)\\}" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": ", an unlabeled target dataset " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t^u = \\{(\\mathbf{x}_t^u)\\}" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": ", and a few-shot labeled support set from the target domain " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t^l = \\{(\\mathbf{x}_t^l,\\mathbf{y}_t^l)\\}" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": ". Samples from source and target datasets are drawn from different data distributions, i.e., " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_s\\sim \\mathcal{P}_s" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t\\sim \\mathcal{P}_t" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_s\\neq \\mathcal{P}_t" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": ". The label sets of source and target datasets are denoted as " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_s" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": ". Due to the novel/fine-grained categories only existing in the target dataset, we have " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_s\\neq \\mathcal{C}_t" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": ". The shared and target-private new classes are indicated as " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{C}} = \\mathcal{C}_s\\cap \\mathcal{C}_t" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{C}} = \\mathcal{C}_t\\backslash \\overline{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": ". We call " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": " existing in both datasets as anchor classes since" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "they can be used as anchors to facilitate knowledge transfer. The goal is to train a model jointly on " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s, \\mathcal{D}_t^u, \\mathcal{D}_t^l" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": " which performs inferences on data from " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": " with labels in " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 114, + 537, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 114, + 537, + 126 + ], + "spans": [ + { + "bbox": [ + 305, + 114, + 537, + 126 + ], + "type": "text", + "content": "3.2. Characterization of Optimization Trajectory" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 132, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 132, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 132, + 545, + 227 + ], + "type": "text", + "content": "The technical challenge for the identified problem lies in overcoming both domain shifts and overfitting incurred by scarce training samples. As discussed in Section 1, those issues could lead to unreliable navigation during neural network optimization [18, 10]. Therefore, we propose to distill the optimization trajectory knowledge from well-annotated sources to regularize the training dynamics of target domain and new classes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "text", + "content": "Recently, NTK theory [27] has been regarded as a tool to dictate the evolution of neural networks. For a network " + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "text", + "content": ", its training dynamics can be described by a kernel " + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "text", + "content": ". Given a training set with " + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "text", + "content": " samples, the kernel can be defined as an " + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "text", + "content": " matrix with " + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "inline_equation", + "content": "K(i,j;\\theta) = \\nabla_{\\theta}f(x^{i};\\theta)^{\\top} \\cdot \\nabla_{\\theta}f(x^{j};\\theta)" + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "inline_equation", + "content": "x^{i}" + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "inline_equation", + "content": "x^{j}" + }, + { + "bbox": [ + 304, + 228, + 545, + 371 + ], + "type": "text", + "content": " are two input samples. It indicates the strong correlations between the network's optimization trajectory and the gradients w.r.t. its parameters. We thus propose to collect a set of intermediate gradients during training and capture their statistics to characterize the optimization trajectory. Specifically, the second-order Taylor expansion shows that:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 334, + 376, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 376, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 334, + 376, + 545, + 430 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f (x; \\theta) = f (x; \\theta^ {*}) + (\\theta - \\theta^ {*}) ^ {\\top} \\cdot \\nabla_ {\\theta} f (x; \\theta^ {*}) \\\\ + \\frac {1}{2} \\left(\\theta - \\theta^ {*}\\right) ^ {\\top} \\cdot \\nabla_ {\\theta} ^ {2} f (x; \\theta^ {*}) \\cdot \\left(\\theta - \\theta^ {*}\\right) \\tag {1} \\\\ + \\mathcal {O} \\left(\\left\\| \\theta - \\theta^ {*} \\right\\| ^ {2}\\right), \\\\ \\end{array}", + "image_path": "d64a575fbdc25cc267cb52dd9cf7e9e8cf8db7904c770331a7a5f105daf65ae7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 435, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 494 + ], + "type": "text", + "content": "which motivates us to model the first- and second-order derivatives (Hessian) of " + }, + { + "bbox": [ + 304, + 435, + 545, + 494 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 435, + 545, + 494 + ], + "type": "text", + "content": " to represent its learning dynamics. We use the mean of gradients to indicate the first-order derivative and follow [28] to approximate the second-order derivative with gradient variances." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 500, + 528, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 500, + 528, + 511 + ], + "spans": [ + { + "bbox": [ + 306, + 500, + 528, + 511 + ], + "type": "text", + "content": "3.3. Cross-Domain and Cross-Class Distillation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 519, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 686 + ], + "type": "text", + "content": "Gradient-based optimization algorithms depend on a large number of update steps supervised by sufficient labeled data [46]. In domain-shifted and small-data regimes, those algorithms suffer from noisy update signals and overfitting, without guarantee to converge towards generalizable minima [22]. To this end, we propose to distill the training dynamics from the label-rich source domain and anchor classes to the target domain and new classes for external regularization by leveraging their intermediate gradients. We firstly compute the gradients " + }, + { + "bbox": [ + 304, + 519, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 304, + 519, + 545, + 686 + ], + "type": "text", + "content": " for each domain and class by backpropagating the obtained losses. With a general backbone " + }, + { + "bbox": [ + 304, + 519, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 519, + 545, + 686 + ], + "type": "text", + "content": " for feature extraction and a task-specific predictor " + }, + { + "bbox": [ + 304, + 519, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 304, + 519, + 545, + 686 + ], + "type": "text", + "content": " for sample-/pixel-wise classification, the individual gradients can be expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 689, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 689, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 321, + 689, + 545, + 717 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} := \\left[ \\frac {\\partial \\mathcal {L} (\\mathbf {C} (\\mathbf {F} (x)) , y)}{\\partial \\theta_ {\\mathbf {C} _ {1}}}, \\dots , \\frac {\\partial \\mathcal {L} (\\mathbf {C} (\\mathbf {F} (x)) , y)}{\\partial \\theta_ {\\mathbf {C} _ {w}}} \\right], \\tag {2}", + "image_path": "56b1ba21abb00a7ccf57c400498b660591769a0d24a650ec28a80a6c17953392.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21176" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 74, + 532, + 259 + ], + "blocks": [ + { + "bbox": [ + 62, + 74, + 532, + 259 + ], + "lines": [ + { + "bbox": [ + 62, + 74, + 532, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 74, + 532, + 259 + ], + "type": "image", + "image_path": "e755981b6e4ea3bed6166a0384259073797712d738a259dd2a3c521cd3161896.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 547, + 298 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed method. We devise the cross-domain and cross-class distillation module as well as the historical self-distillation module to transfer the model optimization knowledge from reliable sources by imposing constraints on the gradient descent steps w.r.t. the insufficiently supervised target domain and new classes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "text", + "content": " is the loss function, " + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "text", + "content": " indicates the total number of parameters in " + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "text", + "content": " are training data and the corresponding label. For unlabeled samples " + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^u" + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "text", + "content": " in the target dataset, online pseudo-labels " + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}_t^u" + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "text", + "content": " are used for loss calculation. Then, as shown in Fig. 2, we aggregate the individual-level gradients in a memory buffer and retrieve the domain-and class-level gradient statistics to model the global first- and second-order derivatives. According to Section 3.2, for gradients w.r.t. domain/class " + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "text", + "content": " collected from a training set with " + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "inline_equation", + "content": "n_{\\pi}" + }, + { + "bbox": [ + 46, + 316, + 289, + 426 + ], + "type": "text", + "content": " samples, we derive their mean and variance as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 434, + 287, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 434, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 125, + 434, + 287, + 456 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} _ {\\pi} ^ {m} = \\frac {1}{n _ {\\pi}} \\sum_ {i = 1} ^ {n _ {\\pi}} \\boldsymbol {g} _ {\\pi} ^ {i}, \\tag {3}", + "image_path": "29cc58df9ee7b4feacbec9f2c9914d58bb8a7ec3b0289d45a4afb85040732798.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 467, + 287, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 467, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 102, + 467, + 287, + 488 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} _ {\\pi} ^ {v} = \\frac {1}{n _ {\\pi} - 1} \\sum_ {i = 1} ^ {n _ {\\pi}} \\left(\\boldsymbol {g} _ {\\pi} ^ {i} - \\boldsymbol {g} _ {\\pi} ^ {m}\\right) ^ {2}. \\tag {4}", + "image_path": "06154619d400d8807d036e5d029f280b3c066767d79a080780791e26a3b1a842.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 491, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 514 + ], + "type": "text", + "content": "Those metrics are measured for each domain and class separately." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 515, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 288, + 713 + ], + "type": "text", + "content": "However, the gradients obtained from insufficiently annotated domain and classes are dubious and error-prone [55]. To alleviate the issue, we take inspiration from recent learning theory findings that stochastic gradients along the optimization steps are generally low-rank [35], which indicates that the optimization process is governed by the top gradient eigenspace. Specifically, we devise a gradient projection approach to suppress the noisy model update signals in particular to the target domain and new classes by projecting their gradients on the reliable subspaces identified by the source domain and anchor classes with sufficient supervision. Our approach involves three iterative steps, i.e., subspace construction, gradient projection, and gradient statistics matching. Firstly, to identify the principal eigenspace, we perform Singular Value Decomposition (SVD) on the aggregated gradients from the source domain and anchor classes. We denote gradients collected from source domain, target domain, anchor class, and new" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "text", + "content": "class as " + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "inline_equation", + "content": "g_s" + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "inline_equation", + "content": "g_A" + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "inline_equation", + "content": "g_N" + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "text", + "content": ". Given the sets of gradients " + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "inline_equation", + "content": "G_s = [g_s^1, g_s^2, \\dots, g_s^K]" + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "inline_equation", + "content": "G_A = [g_A^1, g_A^2, \\dots, g_A^K]" + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "text", + "content": " stored in the memory buffer of volume " + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 316, + 546, + 350 + ], + "type": "text", + "content": ", we apply SVD:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 346, + 354, + 545, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 354, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 346, + 354, + 545, + 369 + ], + "type": "interline_equation", + "content": "\\boldsymbol {G} _ {s} = \\boldsymbol {U} _ {s} \\boldsymbol {\\Sigma} _ {s} \\boldsymbol {V} _ {s} ^ {\\top}, \\quad \\boldsymbol {G} _ {A} = \\boldsymbol {U} _ {A} \\boldsymbol {\\Sigma} _ {A} \\boldsymbol {V} _ {A} ^ {\\top}, \\tag {5}", + "image_path": "04f891aae6e4b54f16830d20ad505d7dd1b04781ab42c7c3884be7c4e5e0d2ad.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "inline_equation", + "content": "U, V" + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "text", + "content": " contain left singular vectors " + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "inline_equation", + "content": "u^i" + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "text", + "content": ", right singular vector " + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "inline_equation", + "content": "v^i" + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "text", + "content": ", and non-negative singular values " + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "inline_equation", + "content": "\\sigma^i" + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "text", + "content": ", respectively. To capture the overall characteristics, we use the gradients averaged along all anchor classes to represent " + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "inline_equation", + "content": "G_A" + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "text", + "content": ". Next, we adopt the low-rank matrix approximation to select the top-" + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "text", + "content": " significant left singular vectors in " + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "inline_equation", + "content": "U_s" + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "inline_equation", + "content": "U_A" + }, + { + "bbox": [ + 305, + 376, + 546, + 454 + ], + "type": "text", + "content": " based on the following criteria:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 460, + 545, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 460, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 318, + 460, + 545, + 475 + ], + "type": "interline_equation", + "content": "\\left\\| \\left(\\boldsymbol {G} _ {s}\\right) ^ {r _ {s}} \\right\\| _ {F} ^ {2} \\geq \\tau \\left\\| \\boldsymbol {G} _ {s} \\right\\| _ {F} ^ {2}, \\quad \\left\\| \\left(\\boldsymbol {G} _ {A}\\right) ^ {r _ {A}} \\right\\| _ {F} ^ {2} \\geq \\tau \\left\\| \\boldsymbol {G} _ {A} \\right\\| _ {F} ^ {2}, \\tag {6}", + "image_path": "0d5a611faa0f9b30627517386dae09f4b5f38344e148b46b95e7b8259aa9aadf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 481, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 481, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 481, + 545, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 481, + 545, + 555 + ], + "type": "inline_equation", + "content": "(\\pmb{G})^r = \\sum_{i=1}^{r} \\pmb{\\sigma}^i \\pmb{u}^i \\pmb{v}^{i^\\top}" + }, + { + "bbox": [ + 305, + 481, + 545, + 555 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 481, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_F" + }, + { + "bbox": [ + 305, + 481, + 545, + 555 + ], + "type": "text", + "content": " denotes the Frobenius norm, " + }, + { + "bbox": [ + 305, + 481, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 481, + 545, + 555 + ], + "type": "text", + "content": " is a threshold to ensure most information is preserved and is set to 0.98. The principal subspace and the corresponding projection matrix are thereby constructed as " + }, + { + "bbox": [ + 305, + 481, + 545, + 555 + ], + "type": "inline_equation", + "content": "M_s = [u_s^1, u_s^2, \\dots, u_s^{r_s}]" + }, + { + "bbox": [ + 305, + 481, + 545, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 481, + 545, + 555 + ], + "type": "inline_equation", + "content": "M_A = [u_A^1, u_A^2, \\dots, u_A^{r_A}]" + }, + { + "bbox": [ + 305, + 481, + 545, + 555 + ], + "type": "text", + "content": ". Afterward, all gradients are projected on the identified subspace as shown in Fig. 2:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 351, + 559, + 545, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 559, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 351, + 559, + 545, + 576 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} _ {s} ^ {*} = M _ {s} M _ {s} ^ {\\top} \\boldsymbol {g} _ {s}, \\quad \\boldsymbol {g} _ {t} ^ {*} = M _ {s} M _ {s} ^ {\\top} \\boldsymbol {g} _ {t}, \\tag {7}", + "image_path": "ae23cbac2ab4501ad059da7e150d008190ffb130e91b728f3a1ceb30b52b0ece.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 575, + 503, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 575, + 503, + 587 + ], + "spans": [ + { + "bbox": [ + 348, + 575, + 503, + 587 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} _ {A} ^ {*} = \\boldsymbol {M} _ {A} \\boldsymbol {M} _ {A} ^ {\\top} \\boldsymbol {g} _ {A}, \\quad \\boldsymbol {g} _ {N} ^ {*} = \\boldsymbol {M} _ {A} \\boldsymbol {M} _ {A} ^ {\\top} \\boldsymbol {g} _ {N}.", + "image_path": "4560563a8bb18c4c35e61886025783e9d26da04bf331275a9abf322fe975a198.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 590, + 547, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 547, + 656 + ], + "type": "text", + "content": "Then following Eq.(3)(4), we minimize the discrepancy between the statistics (i.e., mean and variance) of projected gradients to distill the learning dynamics from the source domain and anchor classes to the target domain and new classes. The overall training objective can be formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 662, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta} \\mathcal {L} _ {E R M} + \\lambda \\left\\{\\| \\left(\\boldsymbol {g} _ {s} ^ {m *} - \\boldsymbol {g} _ {t} ^ {m *}\\right) \\| _ {2} ^ {2} + \\| \\left(\\boldsymbol {g} _ {s} ^ {v *} - \\boldsymbol {g} _ {t} ^ {v *}\\right) \\| _ {2} ^ {2} \\right. \\\\ \\left. + \\frac {1}{n _ {\\text {n e w}}} \\sum_ {i = 1} ^ {n _ {\\text {n e w}}} \\left[ \\| \\left(\\boldsymbol {g} _ {A} ^ {m *} - \\boldsymbol {g} _ {N i} ^ {m *}\\right) \\| _ {2} ^ {2} + \\| \\left(\\boldsymbol {g} _ {A} ^ {v *} - \\boldsymbol {g} _ {N i} ^ {v *}\\right) \\| _ {2} ^ {2} \\right] \\right\\}. \\tag {8} \\\\ \\end{array}", + "image_path": "d66acf9a8785bef03ac19bbd605310d51e11b8dc143a426e8b6c6ab992ab1813.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21177" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ERM}" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " denotes the empirical risk loss, which is implemented with cross-entropy loss for classification and Dice loss [43] for segmentation. " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " is a balancing coefficient, " + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "inline_equation", + "content": "n_{new}" + }, + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": " is the number of new classes in the target domain. It is noted that there exists discrepancy between the anchor and new classes in semantic concepts, which makes feature alignment across classes harmful due to negative transfer [56]. However, aligning class-wise learning dynamics by enforcing the similarity between their gradients could contribute to lower empirical error on new classes. We prove this property theoretically in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 211, + 191, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 211, + 191, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 191, + 224 + ], + "type": "text", + "content": "3.4. Historical Self-distillation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 229, + 287, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 287, + 349 + ], + "type": "text", + "content": "Inferior generalization behaviour is an outstanding challenge for taxonomy adaptive cross-domain adaptation [19]. When domain shifts and unseen classes exist, a trained model could only maintain its performance on test samples sharing multiple similar attributes with the training data but cannot generalize well to the disparate ones [63]. Motivated by the connection between flat minima and generalizability [31], we devise the historical self-distillation module to smooth the optimization path towards a well-generalizable solution by calibrating gradient descent steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 349, + 288, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 288, + 397 + ], + "type": "text", + "content": "To reach flat loss landscapes [6], assuming a local loss function " + }, + { + "bbox": [ + 46, + 349, + 288, + 397 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 349, + 288, + 397 + ], + "type": "text", + "content": " and a network " + }, + { + "bbox": [ + 46, + 349, + 288, + 397 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 349, + 288, + 397 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 46, + 349, + 288, + 397 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 349, + 288, + 397 + ], + "type": "text", + "content": ", instead of minimizing the original loss " + }, + { + "bbox": [ + 46, + 349, + 288, + 397 + ], + "type": "inline_equation", + "content": "\\psi(f(x;\\theta))" + }, + { + "bbox": [ + 46, + 349, + 288, + 397 + ], + "type": "text", + "content": " only, the optimization objective should be:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 406, + 291, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 291, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 291, + 441 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} - \\log \\int_ {\\left\\| \\theta - \\theta^ {*} \\right\\| \\leq \\Gamma} \\exp (- \\psi (f (x; \\theta^ {*})) - \\gamma \\| \\theta - \\theta^ {*} \\|) d \\theta^ {*}, \\tag {9}", + "image_path": "8825198c86743aabfb46e56d477532a101ce157a5c70b1f52e63338c2adbd201.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "type": "text", + "content": " indicates random input samples, " + }, + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "type": "text", + "content": " is defined as the width of a local parameter valley, " + }, + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "type": "text", + "content": " is a balancing weight. In other words, when " + }, + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 46, + 443, + 287, + 491 + ], + "type": "text", + "content": " are neighbouring, the following criteria is required to be satisfied:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 499, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 499, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 58, + 499, + 287, + 514 + ], + "type": "interline_equation", + "content": "\\left\\| \\nabla_ {\\theta} \\psi (f (x; \\theta)) - \\nabla_ {\\theta} \\psi (f (x; \\theta^ {*})) \\right\\| \\leq \\beta \\| \\theta - \\theta^ {*} \\|, \\tag {10}", + "image_path": "bd2f69bf55025bccd29ca8fe00b91f230249cff8864e48d47ab9dd195013f9e5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " is a smoothness factor. To this end, we propose to impose regulations on the optimizer steps by enforcing minima with uniformly-distributed gradients. Given that the gradients from most iterations are generally low-rank [35, 2], we exploit past gradients to identify the low-dimensional subspace which indicates the principal gradient distribution of local minima and then project current gradients on the constructed subspace to exclude sharp and noisy signals. Specifically, we collect and store individual gradients for each iteration along the training procedure, denoted as " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "G_{It} = [g_{It}^{t - T}, g_{It}^{t - T + 1}, \\dots, g_{It}^{t - 2}, g_{It}^{t - 1}]" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " is the index of current iteration, " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " is the size of memory buffer. Thereafter we perform SVD on the set of gradients " + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "G_{It}" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " to construct the historical subspace. Similar to Eq.(5)(6)(7), the top-" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": " left singular vectors of the decomposed gradient set are concatenated to formulate the projection matrix" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 169 + ], + "type": "inline_equation", + "content": "M_{It}" + }, + { + "bbox": [ + 304, + 72, + 545, + 169 + ], + "type": "text", + "content": ". Then for upcoming gradients obtained from backpropagation in each iteration, we project the gradients on the identified low-rank subspace and consider them as residuals. The subsequent optimizer step is conducted with the sum of the original and rectified gradients, as shown in Fig. 2. Take vanilla stochastic gradient descent as an example, with mini-batch gradients " + }, + { + "bbox": [ + 304, + 72, + 545, + 169 + ], + "type": "inline_equation", + "content": "\\tilde{g}" + }, + { + "bbox": [ + 304, + 72, + 545, + 169 + ], + "type": "text", + "content": " in each iteration, the parameters " + }, + { + "bbox": [ + 304, + 72, + 545, + 169 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 72, + 545, + 169 + ], + "type": "text", + "content": " of the optimized network are updated by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 349, + 178, + 545, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 178, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 349, + 178, + 545, + 201 + ], + "type": "interline_equation", + "content": "\\theta := \\theta - \\frac {1}{\\kappa} \\cdot \\eta \\cdot \\left(\\boldsymbol {M} _ {I t} \\boldsymbol {M} _ {I t} ^ {\\top} + \\kappa\\right) \\cdot \\tilde {\\boldsymbol {g}}, \\tag {11}", + "image_path": "c852165abbbf7d600a52c36715e63058b45bdcc5de54f583c7522de2b2cb078d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 210, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 210, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 210, + 545, + 281 + ], + "type": "text", + "content": " is the learning rate, " + }, + { + "bbox": [ + 304, + 210, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 210, + 545, + 281 + ], + "type": "text", + "content": " is a trade-off parameter. As training proceeds, the principal subspace and the corresponding projection matrix are periodically renewed by applying SVD on the recently collected gradients. The overall training procedure is summarized in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 292, + 425, + 305 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 292, + 425, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 425, + 305 + ], + "type": "text", + "content": "3.5. Theoretical Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 312, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 335 + ], + "type": "text", + "content": "In this section, we motivate our proposed method theoretically and provide mathematical insights on its merits." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 336, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 479 + ], + "type": "text", + "content": "Joint Characterization of Feature and Output Space While our optimization trajectory distillation approach proposes a novel perspective that differs from existing works imposing regularization terms in feature and model-output space [32, 19], those information is still captured and implicitly modeled in the gradient-based framework. We prove this property for the classification and segmentation tasks under two commonly adopted loss functions. The detailed proof is presented in the supplemental material. It illustrates the superiority of our method as a unified framework that jointly characterizes the feature and output space as well as the learning dynamics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 480, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 545, + 564 + ], + "type": "text", + "content": "Impacts on Generalization Error In our method, we propose to minimize the discrepancy between the gradient statistics from different domains and classes in the identified principal subspace. We hereby prove its effectiveness towards a tighter generalization error bound on the target domain and new classes [3]. Details can be found in the supplemental material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 578, + 388, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 578, + 388, + 591 + ], + "spans": [ + { + "bbox": [ + 305, + 578, + 388, + 591 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 598, + 487, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 598, + 487, + 611 + ], + "spans": [ + { + "bbox": [ + 305, + 598, + 487, + 611 + ], + "type": "text", + "content": "4.1. Datasets and Experiment Settings" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "To evaluate the effectiveness of our method and its potential clinical implications, we conduct experiments on five medical image datasets in regard to three important downstream tasks from different clinical scenarios. Extended experiments on more diverse tasks, including radiology and fundus analysis, as well as general visual task, are presented in the supplemental material to substantiate the broader applicability of our method." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21178" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 106, + 507, + 259 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 506, + 105 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 506, + 105 + ], + "type": "text", + "content": "Table 1. Comparison results of our proposed method against other state-of-the-art methods for nuclei segmentation and recognition. The metrics with asterisk (*) are calculated on target-private new classes only. The subscripts denote the standard deviations. The best and second-best results are highlighted in bold and brown, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 106, + 507, + 259 + ], + "lines": [ + { + "bbox": [ + 87, + 106, + 507, + 259 + ], + "spans": [ + { + "bbox": [ + 87, + 106, + 507, + 259 + ], + "type": "table", + "html": "
Methods5-shot10-shot
mF1mF1*mPQmPQ*mF1mF1*mPQmPQ*
Sup-only28.080.8921.460.3612.350.5312.230.4530.900.6625.230.8314.180.2414.940.32
Multi-task [48]33.851.2618.151.0818.580.4210.770.3535.150.9421.290.6219.140.2712.890.34
DANN [16]32.761.4116.521.1018.310.6610.400.5935.761.3523.120.8819.700.6014.140.42
CGDM [14]35.031.2821.040.9919.590.5013.220.3237.131.5724.221.1519.780.7113.900.40
LETR [40]27.781.4415.361.0216.680.5410.030.5134.851.5420.030.7318.180.5911.790.27
FT-CIDA [32]30.510.9216.550.8015.970.399.420.4032.631.2121.080.9517.090.4711.960.49
STARTUP [45]37.030.5822.440.7620.400.3313.820.3040.850.5025.920.9123.370.1916.640.24
DDN [25]37.920.8623.711.0520.620.5713.870.4340.160.7226.941.0822.790.3216.820.50
TSA [34]34.031.1719.551.1118.650.4312.060.3734.271.0522.581.2019.780.6114.240.73
TACS [19]36.731.6522.261.2918.130.8012.570.6239.291.4825.841.1322.270.6815.980.56
Ours40.260.9027.140.9721.780.4914.960.2243.880.5231.430.7824.810.2619.350.33
", + "image_path": "0961e345af2002045c55913cbcba95d2b5b708c6bb2408b042c19dcc90f7e1ae.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 278, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 289, + 495 + ], + "type": "text", + "content": "Nuclei Segmentation and Recognition Technically, this task can be formalized as a simultaneous instance segmentation and classification problem [21]. We evaluate our method by considering PanNuke [15] and Lizard [20] as the source and target datasets. They contain 481 and 291 visual fields cropped from whole-slide images with 189,744 and 495,179 annotated nuclei, respectively. There are six types of nuclei in Lizard and only two of them overlap with the five classes in PanNuke, which suggests four new classes in the target domain. We employ the widely used HoverNet [21] architecture with a standard ResNet-50 backbone as the base model. For quantitative evaluation, we follow [20] and report the average F1 and PQ (panoptic quality) scores for all classes (denoted as mF1 and mPQ) to provide a comprehensive assessment measuring the accuracy of nuclei detection, segmentation, and classification. Additionally, we also report the mF1* and mPQ* scores which are solely computed on new classes in the target domain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 496, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 628 + ], + "type": "text", + "content": "Cancer Tissue Phenotyping In this setting, we perform adaptation from a two-class discrimination task to an eight-class one on CRC-TP [29] and Kather [30] datasets. Kather consists of " + }, + { + "bbox": [ + 46, + 496, + 287, + 628 + ], + "type": "inline_equation", + "content": "5000 \\times 150" + }, + { + "bbox": [ + 46, + 496, + 287, + 628 + ], + "type": "text", + "content": " pathology image patches and includes six novel classes other than the two common classes also existing in the source domain. The experiments are conducted with ResNet-101 as backbone. For evaluation, we exploit the accuracy and F1 scores as metrics to indicate the classification performance. The average scores for all classes (mAcc and mF1) and particularly the ones on new classes only (mAcc* and mF1*) are reported." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Skin Lesion Diagnosis We construct a fine-grained taxonomy adaptation setting from two coarse classes to seven subclasses with HAM10000 [52], which contains 10015 dermatoscopic images sampled from different body structures. The adopted experiment settings and evaluation metrics are the same as the cancer tissue phenotyping task. Since the label space of the source and target datasets do" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 278, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 301 + ], + "type": "text", + "content": "not overlap, we only report the mAcc* and mF1* scores which are computed on target classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 303, + 546, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 546, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 546, + 387 + ], + "type": "text", + "content": "Following previous works in UDA [7], for all experiments, each dataset is randomly split into " + }, + { + "bbox": [ + 304, + 303, + 546, + 387 + ], + "type": "inline_equation", + "content": "80\\% / 20\\%" + }, + { + "bbox": [ + 304, + 303, + 546, + 387 + ], + "type": "text", + "content": " for training/testing. To explicitly recognize the target-private new classes, we sample few (5/10) samples with corresponding labels to formulate the support set. The remaining target data is left unlabeled. More details can be found in the supplemental material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 395, + 362, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 362, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 362, + 407 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 414, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 569 + ], + "type": "text", + "content": "We compare our method with state-of-the-art approaches for cross-domain adaptation, including UDA methods DANN [16] and CGDM [14] which do not consider the taxonomic inconsistency, as well as LETR [40], FT-CIDA[32], STARTUP [45], DDN [25], TSA [34], and TACS [19] which are designed for both domain shifts and new class learning. For UDA methods, we jointly perform domain alignment and train the model with the labeled target data. In addition, we conduct comparisons with multi-task learning approach [48], which trains the model on all domains and label sets simultaneously. \"Sup-only\" indicates the baseline where only the supervised loss on the labeled target dataset is used for training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 715 + ], + "type": "text", + "content": "Table 1 shows the comparison results on the cross-domain nuclei recognition and segmentation task. All methods adopt the same ResNet-50-HoverNet as base model following [21]. It can be observed that our proposed method outperforms the state-of-the-art cross-domain adaptation approaches by a large margin. We also notice an important point that compared with the \"Sup-only\" baseline, most competing methods fail to attain better performance on the metrics specific to new classes (i.e., mF1* and mPQ*). For example, under the 5-shot setting, despite a " + }, + { + "bbox": [ + 304, + 570, + 546, + 715 + ], + "type": "inline_equation", + "content": "4.68\\%" + }, + { + "bbox": [ + 304, + 570, + 546, + 715 + ], + "type": "text", + "content": " improvement is achieved by DANN on mF1, it achieves inferior performance on mF1* which is " + }, + { + "bbox": [ + 304, + 570, + 546, + 715 + ], + "type": "inline_equation", + "content": "4.94\\%" + }, + { + "bbox": [ + 304, + 570, + 546, + 715 + ], + "type": "text", + "content": " worse than \"Sup-only\"." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21179" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 84, + 515, + 237 + ], + "blocks": [ + { + "bbox": [ + 79, + 71, + 514, + 83 + ], + "lines": [ + { + "bbox": [ + 79, + 71, + 514, + 83 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 514, + 83 + ], + "type": "text", + "content": "Table 2. Comparison results of our proposed method against other state-of-the-art methods for cancer tissue phenotyping." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 84, + 515, + 237 + ], + "lines": [ + { + "bbox": [ + 79, + 84, + 515, + 237 + ], + "spans": [ + { + "bbox": [ + 79, + 84, + 515, + 237 + ], + "type": "table", + "html": "
Methods5-shot10-shot
mAccmAcc*mF1mF1*mAccmAcc*mF1mF1*
Sup-only46.821.3042.261.2644.641.4942.831.0848.521.0245.931.6745.341.4546.151.83
Multi-task [48]48.350.7842.021.3445.551.1241.681.6348.560.6945.001.7246.030.6444.461.21
DANN [16]47.460.6742.531.2243.081.1443.131.3048.420.4845.311.2746.500.8144.020.64
CGDM [14]49.020.8243.531.5646.290.8743.580.6950.540.8444.571.3548.170.6645.230.40
LETR [40]45.471.9541.331.4743.171.8841.202.0245.940.5640.131.3041.161.1339.610.90
FT-CIDA [32]43.720.8839.060.7940.911.0238.730.8444.650.5338.800.9440.940.9638.961.07
STARTUP [45]49.520.6342.931.8343.591.3739.970.5949.400.5146.911.8746.380.7047.851.36
DDN [25]47.141.7543.250.9942.922.2940.760.5048.270.6744.081.2246.031.1445.441.18
TSA [34]48.221.4942.130.7844.391.8441.921.1347.181.0241.261.1045.341.7244.691.89
TACS [19]45.301.0740.281.1144.741.2642.040.9546.951.1643.691.7447.461.5147.511.25
Ours50.140.7647.661.5447.411.1046.451.1555.860.9851.271.0252.980.4952.120.94
", + "image_path": "0d56a1fdcb44509487ba18d80e2e246a42e8e09a47b34545e404b401ac28822c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 280, + 278, + 432 + ], + "blocks": [ + { + "bbox": [ + 56, + 255, + 277, + 277 + ], + "lines": [ + { + "bbox": [ + 56, + 255, + 277, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 277, + 277 + ], + "type": "text", + "content": "Table 3. Comparison results of our proposed method against other state-of-the-art methods for skin lesion diagnosis." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 280, + 278, + 432 + ], + "lines": [ + { + "bbox": [ + 57, + 280, + 278, + 432 + ], + "spans": [ + { + "bbox": [ + 57, + 280, + 278, + 432 + ], + "type": "table", + "html": "
Methods5-shot10-shot
mAcc*mF1*mAcc*mF1*
Sup-only35.710.7919.890.2238.500.3822.360.67
Multi-task [48]33.801.2019.060.3137.850.4023.320.32
DANN [16]37.220.5723.990.9240.060.9424.320.72
CGDM [14]41.501.0324.260.8047.220.7024.000.23
LETR [40]36.040.4221.390.6640.871.2223.500.49
FT-CIDA [32]31.950.7718.130.4834.600.9020.960.88
STARTUP [45]43.211.1527.170.9243.881.3525.590.60
DDN [25]36.191.0521.500.3442.461.2923.920.28
TSA [34]40.411.3624.051.0142.071.5125.441.06
TACS [19]45.050.4625.600.6546.730.6425.960.79
Ours50.790.6930.050.8152.140.9830.520.83
", + "image_path": "0517a5d6408c1163df0cb619823b3f5bc4c40e816bfc30b9779f8972c73ca6a7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 458, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 532 + ], + "type": "text", + "content": "This phenomenon is incurred by the failure of cross-class generalization. In contrast, our method yields substantial improvements on those new class-specific metrics, which indicates the superior generalizability of our method. The observation is further verified through visual comparisons in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": "The overall quantitative results on the cancer tissue phenotyping and skin lesion diagnosis tasks are presented in Table 2 and 3, respectively. All methods are implemented based on the ResNet-101 backbone. Under each few-shot setting, our method shows higher classification scores compared with previous approaches. In particular, we achieve " + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "inline_equation", + "content": "0.62\\% \\sim 5.32\\%" + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "inline_equation", + "content": "1.12\\% \\sim 4.81\\%" + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": " improvements against the second-best results in terms of mAcc and mF1 for cancer tissue phenotyping, as well as " + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "inline_equation", + "content": "4.92\\% \\sim 7.58\\%" + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "inline_equation", + "content": "2.88\\% \\sim 4.93\\%" + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": " improvements in terms of mAcc* and mF1* for skin lesion diagnosis. The higher accuracy consistently attained by our method under different cross-domain adaptation settings and clinical tasks is attributed to the strong generalizability brought by the proposed optimization trajectory distillation approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 256, + 441, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 256, + 441, + 269 + ], + "spans": [ + { + "bbox": [ + 306, + 256, + 441, + 269 + ], + "type": "text", + "content": "4.3. Analysis and Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 278, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 566 + ], + "type": "text", + "content": "Ablation of Key Components Table 4 shows the ablation study results to illustrate the effectiveness of key components. \"projection\" denotes the gradient projection approach introduced in Section 3.3. For the \"w/o projection\" ablation, we perform cross-domain and cross-class distillation by directly minimizing the statistics discrepancy between unrectified gradients " + }, + { + "bbox": [ + 304, + 278, + 545, + 566 + ], + "type": "inline_equation", + "content": "(g_{s}, g_{t})" + }, + { + "bbox": [ + 304, + 278, + 545, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 278, + 545, + 566 + ], + "type": "inline_equation", + "content": "(g_{A}, g_{N})" + }, + { + "bbox": [ + 304, + 278, + 545, + 566 + ], + "type": "text", + "content": " in Eq. (7). It is observed that each component is indispensable and could improve the overall performance, which elaborates the importance of external navigation for learning target domain and new classes in network training. Interestingly, we notice that the impacts of cross-domain and cross-class distillation could be disparate across tasks. For example, cross-class distillation brings higher performance gain compared with cross-domain distillation on the cancer tissue phenotyping benchmark, while the situation is reversed for skin lesion diagnosis. It is caused by the different natures of each task. Class-imbalance is a serious issue for skin lesion diagnosis in that the class distribution is highly biased [52], which compromises the effectiveness of the cross-class distillation module. Differently, for cancer tissue phenotyping, classes are uniformly distributed [30] and could hence ensure the representativity of anchor classes and their exemplarity to be distilled to new classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Hyperparameter Sensitivity To investigate the effect of hyperparameters " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " in Eq. (8) and " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " in Eq. (11), we present the performance comparison by varying the choices of their values for cancer tissue phenotyping, as shown in Fig. 3. We only change one hyperparameter while keeping the other fixed at a time. The experimental results show that our framework is quite stable for hyperparameters in a wide interval, yet setting " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " to a large value is detrimental to the classification accuracy. Specifically, " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " is a balancing coefficient between class discriminativeness and gradient alignment. When it is too large, the trained model is biased to matching learning dynamics and would lose es" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21180" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 95, + 485, + 182 + ], + "blocks": [ + { + "bbox": [ + 108, + 71, + 485, + 93 + ], + "lines": [ + { + "bbox": [ + 108, + 71, + 485, + 93 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 485, + 93 + ], + "type": "text", + "content": "Table 4. Key component analysis of our method on three cross-domain adaptation benchmarks under 10-shot scheme. The best results are highlighted in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 109, + 95, + 485, + 182 + ], + "lines": [ + { + "bbox": [ + 109, + 95, + 485, + 182 + ], + "spans": [ + { + "bbox": [ + 109, + 95, + 485, + 182 + ], + "type": "table", + "html": "
MethodsNucleiCancerSkin
mF1mF1*mPQmPQ*mAccmAcc*mF1mF1*mAcc*mF1*
Full Framework43.8831.4324.8119.3555.8651.2752.9852.1252.1430.52
w/o cross-domain39.9630.0822.2918.1052.8149.9351.6250.6444.3425.28
w/o cross-class42.3027.8223.7617.0548.9047.2249.8544.7649.0628.24
w/o historical41.5430.4723.0017.5950.1748.0548.1945.1245.8325.97
w/o projection40.0628.8022.1017.3153.5845.7849.0149.2047.2928.02
", + "image_path": "5940c44db642d070a68f39acef54e9d91bbb02109ccd4a9270e4e82f967130d4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 200, + 281, + 293 + ], + "blocks": [ + { + "bbox": [ + 53, + 200, + 281, + 293 + ], + "lines": [ + { + "bbox": [ + 53, + 200, + 281, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 200, + 281, + 293 + ], + "type": "image", + "image_path": "d34469907c41fdddcf838ca7bce695303cbcbba49b1bdf427902020e289eccf9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 297, + 288, + 330 + ], + "lines": [ + { + "bbox": [ + 46, + 297, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 288, + 330 + ], + "type": "text", + "content": "Figure 3. Performance comparison between different choices of hyperparameters " + }, + { + "bbox": [ + 46, + 297, + 288, + 330 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 297, + 288, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 297, + 288, + 330 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 297, + 288, + 330 + ], + "type": "text", + "content": " on the cancer tissue phenotyping benchmark under 10-shot scheme." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 351, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 422 + ], + "type": "text", + "content": "sential semantic knowledge. For " + }, + { + "bbox": [ + 46, + 351, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 351, + 287, + 422 + ], + "type": "text", + "content": ", it controls the proportions of unrectified gradients. Its large value indicates that most original gradients are preserved, which would incur noisy optimization steps and loss of flatness. Therefore, we decide to set " + }, + { + "bbox": [ + 46, + 351, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\lambda = 10" + }, + { + "bbox": [ + 46, + 351, + 287, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 351, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\kappa = 100" + }, + { + "bbox": [ + 46, + 351, + 287, + 422 + ], + "type": "text", + "content": ". It achieves the best performance under most metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 425, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 288, + 616 + ], + "type": "text", + "content": "Impacts on Optimization Trajectory and Model Flatness To better understand the effectiveness of our method, we perform in-depth analysis about its impacts on model's training trajectory and flatness of the converged local minima. Firstly, as shown in Fig. 4(a), we visualize the optimization paths of our method and the multi-task baseline on the test accuracy surface with a contour plot. The details of plot procedure can refer to [26]. It can be observed that the optimization of the multi-task baseline fluctuates sharply and is restricted in a sub-optimal area with a relatively low test accuracy. This is induced by the biased navigation in the course of training due to domain shifts and category gap. With the devised dual-stream optimization trajectory distillation module, our method could by contrast suppress the noisy update signals and proceed consistently towards the optimum." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Besides, we perform analysis about the solutions found by our method and the baseline in terms of flatness [5]. Specifically, the local flatness of a model parameter " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " is quantified by the expected changes of test accuracy between " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " and a neighbouring " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " with perturbation rate " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\varrho" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\varrho}(\\theta) = \\mathbb{E}_{\\|\\theta^{*}\\| = (1 + \\varrho)\\|\\theta\\|}\\left[\\mathrm{Acc}(\\theta^{*}) - \\mathrm{Acc}(\\theta)\\right]" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ". We approximate the expected value by Monte-Carlo sampling over 50 samples. As demonstrated in Fig. 4(b), comparing with the" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 198, + 432, + 316 + ], + "blocks": [ + { + "bbox": [ + 309, + 198, + 432, + 316 + ], + "lines": [ + { + "bbox": [ + 309, + 198, + 432, + 316 + ], + "spans": [ + { + "bbox": [ + 309, + 198, + 432, + 316 + ], + "type": "image", + "image_path": "14d88eb8312627df860f13a9fe26421d4d73df2681158058767a931e55e216bf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 323, + 424, + 331 + ], + "lines": [ + { + "bbox": [ + 318, + 323, + 424, + 331 + ], + "spans": [ + { + "bbox": [ + 318, + 323, + 424, + 331 + ], + "type": "text", + "content": "(a) Visualization of Optimization Trajectory" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 433, + 198, + 544, + 320 + ], + "blocks": [ + { + "bbox": [ + 433, + 198, + 544, + 320 + ], + "lines": [ + { + "bbox": [ + 433, + 198, + 544, + 320 + ], + "spans": [ + { + "bbox": [ + 433, + 198, + 544, + 320 + ], + "type": "image", + "image_path": "3014a8f72be3ae2f65d861f70aa9c8d2a8a4e91eaeb92f6de3b179fc94c3f87a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 323, + 522, + 331 + ], + "lines": [ + { + "bbox": [ + 469, + 323, + 522, + 331 + ], + "spans": [ + { + "bbox": [ + 469, + 323, + 522, + 331 + ], + "type": "text", + "content": "(b) Flatness Analysis" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 335, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 422 + ], + "type": "text", + "content": "Figure 4. Further analysis of our proposed method. (a) We plot the optimization trajectory of our method and the multi-task baseline on the test accuracy surface. (b) We visualize the local flatness of our method and the baseline via test accuracy gap by varying the model parameter perturbation rate. Each point is computed by Monte-Carlo approximation over 50 random samples. All the experiments are conducted on the cancer tissue phenotyping benchmark under 10-shot scheme." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 441, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 490 + ], + "type": "text", + "content": "baseline approach, our method shows stronger robustness against model parameter perturbation. It proves that our method successfully converges to minima with better local flatness and thus possesses superior generalizability [31]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 502, + 378, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 378, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 378, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we study the taxonomy adaptive cross-domain adaptation paradigm, which allows incoherent label sets between source and target datasets. To jointly address the data distribution bias and category gap, we propose a unified framework which performs cross-domain and cross-class optimization trajectory distillation to calibrate the learning dynamics of insufficiently annotated domain and classes. Historical self-distillation is further devised to attain a well-generalizable solution via regulating gradient distributions. Extensive experiments on multiple benchmarks with different task contexts demonstrate the effectiveness and generalization capability of our method. In the future, we will extend this work to further discover the underlying mechanism of optimization trajectory distillation for model generalization and evaluate its potential in more adaptation scenarios." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21181" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Sanjeev Arora, Rong Ge, Yingyu Liang, Tengyu Ma, and Yi Zhang. Generalization and equilibrium in generative adversarial nets (gans). In International Conference on Machine Learning, pages 224-232. PMLR, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 179 + ], + "type": "text", + "content": "[2] Sheikh Shams Azam, Seyyedali Hosseinalipour, Qiang Qiu, and Christopher Brinton. Recycling model updates in federated learning: Are gradient subspaces low-rank? In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 223 + ], + "type": "text", + "content": "[3] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 268 + ], + "type": "text", + "content": "[4] Debora Calderola, Barbara Caputo, and Marco Ciccone. Improving generalization in federated learning by seeking flat minima. In European Conference on Computer Vision, pages 654-672. Springer, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "text", + "content": "[5] Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Seunghyun Park, Yunsung Lee, and Sungrae Park. Swad: Domain generalization by seeking flat minima. Advances in Neural Information Processing Systems, 34:22405-22418, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 380 + ], + "type": "text", + "content": "[6] Pratik Chaudhari, Anna Choromanska, Stefano Soatto, Yann LeCun, Carlo Baldassi, Christian Borgs, Jennifer Chayes, Levent Sagun, and Riccardo Zecchina. Entropy-SGD: Biasing gradient descent into wide valleys. In International Conference on Learning Representations, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 435 + ], + "type": "text", + "content": "[7] Cheng Chen, Qi Dou, Hao Chen, Jing Qin, and Pheng-Ann Heng. Synergistic image and feature adaptation: Towards cross-modality domain adaptation for medical image segmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 865–872, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "type": "text", + "content": "[8] Wentao Chen, Zhang Zhang, Wei Wang, Liang Wang, Zilei Wang, and Tieniu Tan. Cross-domain cross-set few-shot learning via learning compact and aligned representations. In European Conference on Computer Vision, pages 383-399. Springer, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 546 + ], + "type": "text", + "content": "[9] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 613 + ], + "type": "text", + "content": "[10] Ali Cheraghian, Shafin Rahman, Sameera Ramasinghe, Pengfei Fang, Christian Simon, Lars Petersson, and Mehrtash Harandi. Synthesized feature based few-shot class incremental learning on a mixture of subspaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8661-8670, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "text", + "content": "[11] Debasmit Das, Sungrack Yun, and Fatih Porikli. ConfeSS: A framework for single source cross-domain few-shot learning. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[12] Susanna Dodd, Mike Clarke, Lorne Becker, Chris Mavergames, Rebecca Fish, and Paula R Williamson. A taxonomy has been developed for outcomes in medical research to help improve knowledge discovery. Journal of clinical epidemiology, 96:84-92, 2018." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[13] C Dominguez Conde, C Xu, LB Jarvis, DB Rainbow, SB Wells, T Gomes, SK Howlett, O Suchanek, K Polanski, HW King, et al. Cross-tissue immune cell analysis reveals tissue-specific features in humans. Science, 376(6594), 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[14] Zhekai Du, Jingjing Li, Hongzu Su, Lei Zhu, and Ke Lu. Cross-domain gradient discrepancy minimization for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3937-3946, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[15] Jevgenij Gamper, Navid Alemi Koohbanani, Ksenija Benet, Ali Khuram, and Nasir Rajpoot. Pannuke: an open pancancer histology dataset for nuclei instance segmentation and classification. In European congress on digital pathology, pages 11-19. Springer, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 230, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 285 + ], + "type": "text", + "content": "[16] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 285, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 342 + ], + "type": "text", + "content": "[17] Zhiqiang Gao, Shufei Zhang, Kaizhu Huang, Qiufeng Wang, and Chaoliang Zhong. Gradient distribution alignment certificates better adversarial domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8937-8946, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 387 + ], + "type": "text", + "content": "[18] Yixiao Ge, Dapeng Chen, and Hongsheng Li. Mutual meanteaching: Pseudo label refinery for unsupervised domain adaptation on person re-identification. In International Conference on Learning Representations, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "type": "text", + "content": "[19] Rui Gong, Martin Danelljan, Dengxin Dai, Danda Pani Paudel, Ajad Chhatkuli, Fisher Yu, and Luc Van Gool. Tacs: Taxonomy adaptive cross-domain semantic segmentation. In European Conference on Computer Vision, pages 19-35. Springer, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 443, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 443, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 443, + 545, + 520 + ], + "type": "text", + "content": "[20] Simon Graham, Mostafa Jahanifar, Ayesha Azam, Mohammed Nimir, Yee-Wah Tsang, Katherine Dodd, Emily Hero, Harvir Sahota, Atisha Tank, Ksenija Benes, et al. Lizard: a large-scale dataset for colonic nuclear instance segmentation and classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 684-693, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 521, + 545, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 521, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 308, + 521, + 545, + 576 + ], + "type": "text", + "content": "[21] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:101563, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 577, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 577, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 577, + 545, + 632 + ], + "type": "text", + "content": "[22] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In European conference on computer vision, pages 124-141. Springer, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 633, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 633, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 633, + 545, + 656 + ], + "type": "text", + "content": "[23] Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural computation, 9(1):1-42, 1997." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 657, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 657, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 657, + 545, + 690 + ], + "type": "text", + "content": "[24] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In European Conference on Computer Vision, pages 20-37. Springer, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Ashraful Islam, Chun-Fu Richard Chen, Rameswar Panda, Leonid Karlinsky, Rogerio Feris, and Richard J Radke. Dy" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21182" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "namic distillation network for cross-domain few-shot recognition with unlabeled data. Advances in Neural Information Processing Systems, 34:3584-3595, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[26] Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization. Proceedings of the Conference on Uncertainty in Artificial Intelligence, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "text", + "content": "[27] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 253 + ], + "type": "text", + "content": "[28] Stanisław Jastrzebski, Zachary Kenton, Devansh Arpit, Nicolas Ballas, Asja Fischer, Yoshua Bengio, and Amos Storkey. Three factors influencing minima in sgd. arXiv preprint arXiv:1711.04623, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 319 + ], + "type": "text", + "content": "[29] Sajid Javed, Arif Mahmood, Muhammad Moazam Fraz, Navid Alemi Koohbanani, Ksenija Benes, Yee-Wah Tsang, Katherine Hewitt, David Epstein, David Snead, and Nasir Rajpoot. Cellular community detection for tissue phenotyping in colorectal cancer histology images. Medical image analysis, 63:101696, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "text", + "content": "[30] Jakob Nikolas Kather, Cleo-Aron Weis, Francesco Bianconi, Susanne M Melchers, Lothar R Schad, Timo Gaiser, Alexander Marx, and Frank Gerrit Zöllner. Multi-class texture analysis in colorectal cancer histology. Scientific reports, 6(1):1-11, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "type": "text", + "content": "[31] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. In International Conference on Learning Representations, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 287, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 478 + ], + "type": "text", + "content": "[32] Jogendra Nath Kundu, Rahul Mysore Venkatesh, Naveen Venkat, Ambareesh Revanur, and R Venkatesh Babu. Class incremental domain adaptation. In European Conference on Computer Vision, pages 53-69. Springer, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 479, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 555 + ], + "type": "text", + "content": "[33] Canran Li, Dongnan Liu, Haoran Li, Zheng Zhang, Guangming Lu, Xiaojun Chang, and Weidong Cai. Domain adaptive nuclei instance segmentation and classification via category-aware feature alignment and pseudo-labelling. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 715-724. Springer, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 556, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 601 + ], + "type": "text", + "content": "[34] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Cross-domain few-shot learning with task-specific adapters. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7161-7170, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "text", + "content": "[35] Xinyan Li, Qilong Gu, Yingxue Zhou, Tiancong Chen, and Arindam Banerjee. Hessian based analysis of sgd for deep nets: Dynamics and generalization. In Proceedings of the 2020 SIAM International Conference on Data Mining, pages 190-198. SIAM, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[36] Hanwen Liang, Qiong Zhang, Peng Dai, and Juwei Lu. Boosting the generalization capability in cross-domain few-shot learning via noise-enhanced supervised autoencoder. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9424-9434, 2021." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[37] Dongnan Liu, Chaoyi Zhang, Yang Song, Heng Huang, Chenyu Wang, Michael Barnett, and Weidong Cai. Decompose to adapt: Cross-domain object detection via feature disentanglement. IEEE Transactions on Multimedia, 25:1333-1344, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 195 + ], + "type": "text", + "content": "[38] Dongnan Liu, Donghao Zhang, Yang Song, Fan Zhang, Lauren O'Donnell, Heng Huang, Mei Chen, and Weidong Cai. Unsupervised instance segmentation in microscopy images via panoptic domain adaptation and task re-weighting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4243-4252, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 196, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 545, + 239 + ], + "type": "text", + "content": "[39] Xu Luo, Jing Xu, and Zenglin Xu. Channel importance matters in few-shot image classification. In International Conference on Machine Learning, pages 14542-14559. PMLR, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 240, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 545, + 284 + ], + "type": "text", + "content": "[40] Zelun Luo, Yuliang Zou, Judy Hoffman, and Li F Fei-Fei. Label efficient learning of transferable representations acrossss domains and tasks. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 285, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 545, + 308 + ], + "type": "text", + "content": "[41] Deborah Lupton. Medicine as culture: illness, disease and the body. Sage, 2012." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "text", + "content": "[42] Lucas Mansilla, Rodrigo Echeveste, Diego H Milone, and Enzo Ferrante. Domain generalization via gradient surgery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6630-6638, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "text", + "content": "[43] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 398, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 545, + 431 + ], + "type": "text", + "content": "[44] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In Proceedings of the IEEE international conference on computer vision, pages 754-763, 2017." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 432, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 464 + ], + "type": "text", + "content": "[45] Cheng Perng Phoo and Bharath Hariharan. Self-training for few-shot transfer across extreme task differences. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 466, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 498 + ], + "type": "text", + "content": "[46] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations, 2017." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 500, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 553 + ], + "type": "text", + "content": "[47] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, Trevor Darell, and Kate Saenko. Semi-supervised domain adaptation via minimax entropy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8050-8058, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 555, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 589 + ], + "type": "text", + "content": "[48] Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "text", + "content": "[49] Yuge Shi, Jeffrey Seely, Philip Torr, Siddharth N, Awni Hannun, Nicolas Usunier, and Gabriel Synnaeve. Gradient matching for domain generalization. In International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 634, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 667 + ], + "type": "text", + "content": "[50] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 690 + ], + "type": "text", + "content": "[51] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[52] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source der" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21183" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 632 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "matoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[53] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In International Conference on Learning Representations, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "text", + "content": "[54] Qian Wang and Toby Breckon. Unsupervised domain adaptation via structured prediction based selective pseudolabeling. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 6243-6250, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 218 + ], + "type": "text", + "content": "[55] Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur), 53(3):1-34, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 263 + ], + "type": "text", + "content": "[56] Zirui Wang, Zihang Dai, Barnabás Póczos, and Jaime Carbonell. Characterizing and avoiding negative transfer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11293-11302, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 265, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 341 + ], + "type": "text", + "content": "[57] Julia K Winkler, Christine Fink, Ferdinand Toberer, Alexander Enk, Teresa Deinlein, Rainer Hofmann-Wellenhof, Luc Thomas, Aimilios Lallas, Andreas Blum, Wilhelm Stolz, et al. Association between surgical skin markings in dermoscopic images and diagnostic performance of a deep learning convolutional neural network for melanoma recognition. JAMA dermatology, 155(10):1135-1141, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 342, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 407 + ], + "type": "text", + "content": "[58] Mengya Xu, Mobarakol Islam, Chwee Ming Lim, and Hongliang Ren. Class-incremental domain adaptation with smoothing and calibration for surgical report generation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 269-278. Springer, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "text", + "content": "[59] Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Universal domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2720-2729, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 497 + ], + "type": "text", + "content": "[60] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 498, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 552 + ], + "type": "text", + "content": "[61] Wang Yuan, Zhizhong Zhang, Cong Wang, Haichuan Song, Yuan Xie, and Lizhuang Ma. Task-level self-supervision for cross-domain few-shot learning. Proceedings of the AAAI Conference on Artificial Intelligence, 36(3):3215-3223, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "type": "text", + "content": "[62] Qiming Zhang, Jing Zhang, Wei Liu, and Dacheng Tao. Category anchor-guided unsupervised domain adaptation for semantic segmentation. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 599, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 287, + 632 + ], + "type": "text", + "content": "[63] Lin Zuo, Mengmeng Jing, Jingjing Li, Lei Zhu, Ke Lu, and Yang Yang. Challenging tough samples in unsupervised domain adaptation. Pattern Recognition, 110:107540, 2021." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21184" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/a12231ff-0413-4e22-8eb5-06598dd6fed6_content_list.json b/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/a12231ff-0413-4e22-8eb5-06598dd6fed6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8c08be74e125865f5f6cbf32f5fbc08439ecbc5c --- /dev/null +++ b/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/a12231ff-0413-4e22-8eb5-06598dd6fed6_content_list.json @@ -0,0 +1,1892 @@ +[ + { + "type": "text", + "text": "TeD-SPAD: Temporal Distinctiveness for Self-supervised Privacy-preservation for video Anomaly Detection", + "text_level": 1, + "bbox": [ + 91, + 128, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Joseph Fioresi, Ishan Rajendrakumar Dave, Mubarak Shah Center for Research in Computer Vision, University of Central Florida, Orlando, USA {joseph.fioresi, ishanrajendrakumar.dave}@ucf.edu, shah@crcv.ucf.edu Project Page: https://joefioresi718.github.io/TeD-SPAD_webpage/", + "bbox": [ + 143, + 203, + 823, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video anomaly detection (VAD) without human monitoring is a complex computer vision task that can have a positive impact on society if implemented successfully. While recent advances have made significant progress in solving this task, most existing approaches overlook a critical real-world concern: privacy. With the increasing popularity of artificial intelligence technologies, it becomes crucial to implement proper AI ethics into their development. Privacy leakage in VAD allows models to pick up and amplify unnecessary biases related to people's personal information, which may lead to undesirable decision making. In this paper, we propose TeD-SPAD, a privacy-aware video anomaly detection framework that destroys visual private information in a self-supervised manner. In particular, we propose the use of a temporally-distinct triplet loss to promote temporally discriminative features, which complements current weakly-supervised VAD methods. Using TeD-SPAD, we achieve a positive trade-off between privacy protection and utility anomaly detection performance on three popular weakly supervised VAD datasets: UCF-Crime, XD-Violence, and ShanghaiTech. Our proposed anonymization model reduces private attribute prediction by $32.25\\%$ while only reducing frame-level ROC AUC on the UCF-Crime anomaly detection dataset by $3.69\\%$ .", + "bbox": [ + 76, + 340, + 472, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 724, + 209, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Machine learning-driven technologies are increasingly being adopted by society. The progress in cloud computing has enabled the deployment of even computationally intensive technologies in the public space. One such application is video anomaly detection (VAD) in autonomous video analytics. VAD is a video understanding task that aims to identify the temporal location of anomalous events occurring in long continuous videos without human supervision. An anomaly can be defined as any unusual event, such as a traffic accident, an elderly person falling, or a fire. Proper", + "bbox": [ + 75, + 750, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b290adf0c9ef88329790cc0d2ec3335fae173addede23c54d89d8c434bdfeccc.jpg", + "image_caption": [ + "Figure 1: Single frame from video Shoplifting052_x264.mp4 of UCF-Crime [43] showing the types of private attributes (shown in red on the left image) leaked in visual data. In the right image we show the anonymized frame where these attributes are barely visible. The graph demonstrates the ability of our self-supervised framework to mitigate privacy leakage compared to prior work SPAct [13]. At a similar anomaly detection utility performance, our method prevents $32\\%$ of the visual private leakage compared to raw data." + ], + "image_footnote": [], + "bbox": [ + 501, + 306, + 728, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee303beaf532acdbce2d63df0f84598342211684dcc221ef88c392bea0bf8d05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 306, + 890, + 493 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "application of this technology can result in faster response times to anomalies, without the need for human resources to monitor camera feeds.", + "bbox": [ + 496, + 657, + 890, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, public adoption of such AI technologies brings justifiable concern about safety and their decision-making capabilities. Many of these concerns center around privacy violations and accuracy. VAD is an application where visual privacy leakage and data bias are exceedingly important issues. Sending videos to remote computers or cloud services to process results in unnecessary privacy leakage for people who are not directly involved in anomalous events. Additionally, an application employing a standard RGB video will incorporate any bias found in its training set, leading to potentially unfair decisions. An illustration of privacy leakage is shown in Fig. 1.", + "bbox": [ + 496, + 703, + 892, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, there have been interesting attempts to prevent", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13598", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "visual privacy leakage in action recognition. Some of the approaches utilize input downsizing-based solutions [40, 10, 27] and object-detection-dependent obfuscation-based formulations [38, 53]. Wu et al. [49] proposed an adversarial training-based framework where they train an anonymization function to remove privacy-preservation. Dave et al. [13] proposed a self-supervised privacy-preserving framework that does not require privacy labels and achieves similar performance to the supervised method [49]. Since many weakly-supervised anomaly detection (WSAD) methods rely on the pretrained features of action recognition, privacy-preserving action recognition seems like a promising candidate for privacy-preserving anomaly detection. However, detecting anomalies does not align well with privacy-preserved action recognition videos. The use of short videos in action recognition encourages the use of temporally-invariant features, where the features of clips at distinct timesteps should be the same. Conversely, detecting anomalies in long, untrimmed videos requires temporally-distinct reasoning, where features of clips at distinct timesteps of a video should be different, to determine whether events in the same scene are anomalous. This is why most existing anomaly detection methods focus on refining the features of pretrained video encoders to increase their temporal separability.", + "bbox": [ + 76, + 90, + 472, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To the best of our knowledge, privacy-preservation in video anomaly detection is an unexplored area in computer vision. Building on the existing self-supervised privacy-preserving action recognition framework [13], we propose a more aligned utility branch for anomaly detection. To achieve this, we use a novel temporally-distinct triplet loss to promote temporal distinctiveness during anonymization training. Once the anonymization function is learned through our proposed anonymization framework, we apply it to the anomaly dataset, which ensures privacy leakage mitigation in privacy-sensitive anomaly detection tasks. We use these anonymized features to train the current state-of-the-art WSAD method, MGFN [8].", + "bbox": [ + 76, + 468, + 472, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate privacy-preserving performance in anomaly detection, we adopt protocols from prior action recognition methods, where we report the utility performance on WSAD task on widely used anomaly datasets (UCF-Crime [43], XD-Violence [48], and ShanghaiTech [28]) and budget performance on privacy dataset VISPR [32].", + "bbox": [ + 76, + 665, + 472, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 96, + 756, + 423, + 770 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new problem of privacy-preservation in video anomaly detection, where we identify the privacy leakage issue in existing weakly supervised anomaly detection methods.", + "- To address this open problem, we propose TeD-SPAD, a framework based on self-supervised privacy-preservation with a temporally-distinct triplet loss to" + ], + "bbox": [ + 94, + 784, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "make the video anonymization process more suitable for anomaly detection.", + "bbox": [ + 529, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose evaluation protocols for privacy vs. anomaly trade-off, demonstrating that our proposed framework outperforms prior methods by significant margins across all anomaly detection benchmarks. On the widely used UCF-Crime dataset, our method is able to eliminate $32.25\\%$ of the privacy leakage at a cost of only a $3.96\\%$ reduction in frame-level AUC performance.", + "bbox": [ + 517, + 133, + 893, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 268, + 650, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Privacy Preservation", + "text_level": 1, + "bbox": [ + 500, + 295, + 651, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We observe that many works preserve visual privacy at capture time by using non-intrusive sensors such as thermal imaging, depth cameras, or event cameras [29, 18, 21, 1]. Other works allow for raw RGB visual information to be captured, but make an effort to protect the subject privacy in such a way that the data is still useful in a utility task. Earlier efforts aimed at dealing with visual privacy include image downsampling [11] or blocking/blurring privacy-related objects located using pretrained object detectors. Both of these obfuscation methods were shown to reduce utility results by more than they reduced privacy leakage [13, 24, 49].", + "bbox": [ + 496, + 310, + 893, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent developments have yielded numerous privacy-preserving approaches to action recognition [2, 51, 31]. Wu et al. released an action dataset with privacy labels, PA-HMDB [49]. They use an adversarial learning framework to obfuscate the privacy features using supervised privacy labels. MaSS [4] uses a similar framework to Wu et al. [49], except it adapts a compound loss to flexibly preserve certain attributes instead of destroying them. STPrivacy [24] upgraded the general framework to work with a transformer anonymizing block, masking entire video tubelets unnecessary for action recognition. Following Dave et al.'s SPAcT [13], we adopt a similar self-supervised adversarial anonymization framework without the use of the privacy labels, using NT-Xent [6] contrastive loss in the budget branch to mitigate spatial privacy leakage.", + "bbox": [ + 496, + 477, + 893, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Anomaly Detection", + "text_level": 1, + "bbox": [ + 500, + 704, + 638, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With such a high volume of available video footage, it is infeasible to create sufficient labelled data to solve supervised VAD. Therefore, many works explore unsupervised methods. These generally train a reconstruction model, then either reconstruct the current frame or try to predict the next frame, signaling an anomaly when reconstruction error is high [9, 35, 23, 33, 50, 44]. Giorgi et al. [16] used an autoencoder with differential privacy, generating anomaly scores from the noisy reconstructed images. This method helped retain some level of subject privacy, but was only evaluated on image quality metrics.", + "bbox": [ + 496, + 719, + 893, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sultani et al. [43] brought weak supervision to VAD,", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13599", + "bbox": [ + 480, + 924, + 519, + 936 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where anomalies are labelled at the video level. These authors introduce UCF-Crime, a large scale weakly supervised dataset. They propose formulating weakly supervised VAD as a multi instance learning (MIL) problem, showcasing the benefits of temporal smoothness loss & sparsity loss. With the exception of a select few works [54, 52], all following weakly supervised methods are considered anomaly feature representation learning as they improve upon MIL formulation [55, 45, 15, 25, 19, 46, 8, 41], which involves interpreting static video features extracted using an action classifier. Zhong et al. [54] propose a rearranged weakly supervised version of ShanghaiTech. Wu et al. [48] introduce XD-Violence, a large-scale weakly supervised dataset that includes audio, bringing multi-modal fusion to VAD. In this work, we choose to focus on the weakly supervised video anomaly detection setting due to its effectiveness and low annotation effort.", + "bbox": [ + 75, + 90, + 472, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most VAD works find it useful to model temporal relations between video segments [45, 15, 30, 48, 25, 46, 55, 5]. [30] are able to exploit dynamic variations in features between consecutive segments to help localize anomalies. The authors of RTFM [45] find that anomalous segments tend to have larger feature magnitudes ( $\\ell_2$ norm of extracted clip feature vector) than normal segments. They introduce a feature magnitude learning function to help identify anomalous segments with large feature magnitudes. MGFN [8] notices that varying scene attributes causes some normal videos to have larger feature magnitudes than anomalous, so they propose a feature magnitude contrastive loss to help capture these instances. In this work, we use the MGFN model to evaluate anomaly detection performance. Complementing these ideas, [47] demonstrates the effectiveness of explicitly encouraging feature discrimination. Intuitively, these observations can be aggregated by enforcing temporally-distinct feature representations. Building on the terminology introduced in [34] regarding invariant and distinctive representations, in this paper our goal is to enhance the distinctiveness of the representations within our utility branch.", + "bbox": [ + 75, + 347, + 472, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 678, + 169, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The central idea behind our proposed framework is to develop an anonymization function that can degrade privacy attributes during training without relying on privacy labels. Furthermore, this function must be able to maintain the performance of the weakly-supervised anomaly detection task. Fig. 2 displays a schematic diagram of the proposed framework. In Sec. 3.1, we provide a comprehensive discussion of the problem statement. Next, in Sec. 3.2, we describe the component of the framework and their initialization process. Sec. 3.3 outlines the anonymization function training, where we propose a temporally-distinct triplet loss to enhance the existing self-supervised privacy-preserving framework [13]. Once we learn the anonymization func", + "bbox": [ + 75, + 704, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion, in Sec. 3.4, we train the anomaly detection model using the privacy-preserved features obtained through our anonymization function. An overview of our complete framework is outlined in Algorithm Sec. 3.5.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Statement", + "text_level": 1, + "bbox": [ + 498, + 161, + 687, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define our problem statement similar to privacy-preserving action recognition frameworks [13, 47], but, with a different utility task. Suppose we have a video dataset $\\mathbb{D}_{\\text{anomaly}} = \\{X^{(i)}, Y^i\\}_{i=1}^N$ , where, $X^{(i)}$ is a video-instance, $N$ is the total number of samples and $Y^{(i)} \\in \\{0,1\\}$ is a binary label. Considering video-level anomaly detection as a utility task $T$ , and privacy attribute classification as the budget task $B$ , the aim of a privacy-preserving system is to maintain the performance of $T$ while reducing $B$ . To achieve this goal, the system learns an anonymization function $f_A$ , which modifies the original raw data. This goal of privacy-preservation could be fundamentally expressed as following criteria:", + "bbox": [ + 496, + 185, + 890, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Criterion-1: The performance of the utility task should not degrade from the original performance, i.e. loss $\\mathcal{L}_T$ value of the utility target model $f_T'$ should remain almost identical before and after applying the anonymization function.", + "bbox": [ + 496, + 381, + 890, + 441 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T} \\left(f _ {T} ^ {\\prime} \\left(f _ {A} ^ {*} (X)\\right), Y\\right) \\approx \\mathcal {L} _ {T} \\left(f _ {T} ^ {\\prime} (X), Y\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 454, + 890, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Criterion-2: Applying anonymization function should increase the loss $\\mathcal{L}_B$ for budget $B$ task of target budget model $f_B^\\prime$ .", + "bbox": [ + 496, + 483, + 890, + 529 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {B} \\left(f _ {B} ^ {\\prime} \\left(f _ {A} ^ {*} (X)\\right)\\right) \\gg \\mathcal {L} _ {B} \\left(f _ {B} ^ {\\prime} (X)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 529, + 890, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regarding weakly-supervised anomaly detection (WSAD), most of the existing methods require multi-stage training, meaning they are not end-to-end. This presents a challenge for incorporating it as a utility task in anonymization training. By contrast, privacy-preserving action recognition frameworks [13, 49] have an end-to-end utility task (i.e action recognition), making it more straightforward to include.", + "bbox": [ + 496, + 553, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since most WSAD methods rely on pretrained video encoder features from large-scale action recognition training, we can utilize the exact same anonymization process of privacy-preserving action recognition [13], utilizing action recognition as a proxy-utility task on a proxy-utility action dataset, which we denote as $\\mathbb{D}_{\\text {action }}$ . The limitation of such anonymization training is that by focusing solely on optimizing short clips for action recognition, it neglects the importance of temporally-distinct features. This oversight leads to a notable decline in anomaly detection performance, which relies heavily on these temporally-distinct features, as the training progresses. To resolve this issue, we reformulate the anonymization utility branch to enforce temporal distinctiveness to better align with the anomaly detection task. Hence, our $\\mathcal{L}_T$ is weighted sum of action", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "13600", + "bbox": [ + 480, + 924, + 519, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a5fe680f6fbe9983d2723bcddf3dd34b42af55eadf714f8ea112691e2fe248ec.jpg", + "image_caption": [ + "(b) Privacy-Preserved Anomaly Detection Workflow", + "Figure 2: Full TeD-SPAD framework consisting of the proxy anonymization training followed by the privacy-preserved anomaly detection. (a) shows this proxy training, where UNet is used to anonymize frames in such a way that reduces mutual information between frames while maintaining utility performance. We complement the standard cross-entropy loss with our proposed temporally-distinct triplet loss, which enforces a difference in clip features at distinct timesteps. After training the anonymizer and feature extractor, (b) shows the privacy-preserved workflow, where the anomaly dataset videos are passed through the proxy-trained $f_{A}$ , $f_{T}$ , then into any WSAD algorithm." + ], + "image_footnote": [], + "bbox": [ + 140, + 89, + 834, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "recognition loss $(\\mathcal{L}_{CE})$ and temporally-distinct triplet loss $(\\mathcal{L}_D)$ which is elaborated in Eq. 6.", + "bbox": [ + 75, + 496, + 468, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Anonymization Framework", + "text_level": 1, + "bbox": [ + 76, + 546, + 326, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our anonymization framework consists of 3 main components: (1) Anonymization function $(f_A)$ , which is a simple encoder-decoder model with a sigmoid activation. (2) Privacy removal model $(f_B)$ , which is an image encoder. (3) Utility model $(f_T)$ , which is a video encoder.", + "bbox": [ + 75, + 571, + 468, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Network Initialization First of all, our anonymization model is pretrained to initialize as an identity function. This pretraining involves the reconstruction of the frames from $\\mathbb{D}_{\\text {action }}$ using L1-reconstruction loss.", + "bbox": [ + 75, + 650, + 468, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L 1} = \\sum_ {c = 1} ^ {C} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} | x _ {c, h, w} - \\hat {x} _ {c, h, w} |, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 734, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{x}$ is the output of $f_{A}$ , $x$ is an input image, and $C, H, W$ corresponds to the channel, height, and width of the input image.", + "bbox": [ + 76, + 791, + 468, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Secondly, the privacy model $f_{B}$ is initialized with the self-supervised weights of SimCLR [7] on ImageNet [14]. The video encoder $f_{T}$ is pretrained with the standard action recognition weights from Kinetics400 dataset [3].", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Anonymization Training", + "text_level": 1, + "bbox": [ + 500, + 496, + 725, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Anonymization training is mainly comprised of a minimax optimization of utility loss and privacy loss.", + "bbox": [ + 498, + 518, + 890, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporally-distinct triplet loss as Utility Weakly-supervised video anomaly detection methods leverage temporal information to help localize anomalies. [30, 19, 45, 8] show the positive effect of feature magnitude ( $\\ell_2$ norm of clip feature vector) separability along the temporal dimension. In order to adopt the SPAcT anonymization framework to anomaly detection problems, we utilize a temporal distinctiveness objective in the utility branch. We design a temporally-distinct triplet loss, which increases the agreement between temporally-aligned clips of the same video and increases dissimilarity between representations of the clips which are temporally-misaligned. For anchor clip $\\mathbf{x}_t^{(i)}$ , we obtain the positive clip from the exact same timestamp, but with a differently augmented version denoted as $\\tilde{\\mathbf{x}}_t^{(i)}$ . Whereas, the negative clip is obtained from different timestamp $\\mathbf{x}_{t'}^{(i)}$ , where $t' \\neq t$ . This triplet of clips is passed through the utility model $f_T$ to achieve features denoted as $\\mathbf{z}_t^{(i)}$ , $\\tilde{\\mathbf{z}}_t^{(i)}$ , and $\\mathbf{z}_{t'}^{(i)}$ . The proposed temporally-distinct triplet loss can be expressed as follows:", + "bbox": [ + 496, + 550, + 892, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} ^ {(i)} = \\max \\{d (\\mathbf {z} _ {t} ^ {(i)}, \\tilde {\\mathbf {z}} _ {t} ^ {(i)}) - d (\\mathbf {z} _ {t} ^ {(i)}, \\mathbf {z} _ {t ^ {\\prime}} ^ {(i)}) + \\mu , 0 \\}, \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 853, + 890, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d(u_j,v_j) = \\|\\mathbf{u}_j - \\mathbf{v}_j\\|_2$ is Euclidean distance be", + "bbox": [ + 500, + 883, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13601", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tween two vectors $\\mathbf{u}$ and $\\mathbf{v}$ , and $\\mu$ is the controllable margin hyperparameter to determine how far to push and pull features in the latent space.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize this loss along with the standard cross-entropy action classification loss:", + "bbox": [ + 76, + 136, + 468, + 166 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C E} ^ {(i)} = - \\sum_ {c = 1} ^ {N _ {C}} \\mathbf {y} _ {c} ^ {(i)} \\log \\mathbf {p} _ {c} ^ {(i)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 176, + 468, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{C}$ is the total number of action classes of $\\mathbb{D}_{\\text {action }}$ , $\\mathbf{y}_c^{(i)}$ is the ground-truth label, and $\\mathbf{p}_c^{(i)}$ is the prediction vector by utility model $f_T$ .", + "bbox": [ + 76, + 223, + 468, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adding both temporal distinctiveness (Eq. 4) and action classification objective to our utility branch, our overall utility loss can be expressed as follows", + "bbox": [ + 76, + 268, + 468, + 314 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T} = \\mathcal {L} _ {C E} + \\omega * \\mathcal {L} _ {D}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 323, + 468, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega$ hyperparameter is the weight of temporally-distinct triplet loss with respect to cross-entropy loss.", + "bbox": [ + 75, + 349, + 468, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Privacy (i.e. budget) Loss $\\mathcal{L}_B$ We utilize the same self-supervised privacy loss from [13], which removes the private information by minimizing the agreement between the frames of the same video.", + "bbox": [ + 75, + 380, + 468, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Minimax Optimization After reformulating the utility loss, we use the minimax optimization process similar to [13]. It is a two-step iterative process that minimizes the utility loss and at the same time increases budget loss $\\mathcal{L}_B$ . At the end of this optimization, we obtain the learned anonymization function $(f_A)$ and utility video encoder $(f_T)$ .", + "bbox": [ + 75, + 440, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Privacy-preserving Anomaly Detection", + "text_level": 1, + "bbox": [ + 76, + 539, + 410, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to detect anomalies within videos from $\\mathbb{D}_{\\text{anomaly}}$ , we utilize the current state-of-the-art technique, Magnitude-Contrastive Glance-and-Focus Network (MGFN) [8]. Similar to other recent works in anomaly detection, MGFN requires fixed features for each video from a pretrained video encoder for anomaly detection training.", + "bbox": [ + 75, + 561, + 468, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimizing for Anomaly Detection MGFN anomaly detection is comprised of 4 main losses: (1) a standard sigmoid cross-entropy loss $L_{sce}$ for snippet classification accuracy, (2) a temporal smoothing loss $L_{ts}$ [43] to encourage smoothness between feature representations of consecutive segments, (3) a sparsity term $L_{sp}$ [43] to discourage false positive anomalies, and (4) a novel magnitude contrastive loss $L_{mc}$ to learn scene-adaptive feature distributions across videos, all which help to train a model $f_{AD}$ .", + "bbox": [ + 75, + 652, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training loss used in MGFN is compounded in the following equation:", + "bbox": [ + 76, + 789, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {A D} = L _ {s c e} + \\lambda_ {1} L _ {t s} + \\lambda_ {2} L _ {s p} + \\lambda_ {3} L _ {m c}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 829, + 468, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_1 = \\lambda_2 = 1$ , and $\\lambda_3 = 0.001$ . $f_{AD}$ outputs frame-level anomaly scores, which are used to calculate a final ROC AUC and AP for evaluation.", + "bbox": [ + 75, + 854, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Extraction In our privacy-preserving case, we cannot use $\\mathbb{D}_{\\text{anomaly}}$ directly for the feature extraction from the video encoder. We first anonymize each video $(X^i)$ of the dataset through the learned anonymization function $f_A$ to get an anonymized dataset. For the feature extraction, we utilize the learned utility video encoder $f_T$ . We denote this extracted set of anonymized features as $\\mathbb{F}_{\\text{anomaly}} = \\{f_T(f_A(X^i)) \\mid \\forall X^i \\in \\mathbb{D}_{\\text{anomaly}}\\}$ .", + "bbox": [ + 498, + 90, + 890, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Algorithm", + "text_level": 1, + "bbox": [ + 500, + 222, + 616, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let's consider the models $f_{A}, f_{T}, f_{B}, f_{AD}$ , parameterized by $\\theta_{A}, \\theta_{T}, \\theta_{B}$ , and $f_{AD}$ , respectively. $\\mathbb{D}_{\\text{action}}$ is the proxy action recognition dataset and $\\mathbb{D}_{\\text{anomaly}}$ is the downstream anomaly detection dataset. All training steps of our framework can be put together into a sophisticated form as algorithm 1.", + "bbox": [ + 498, + 244, + 890, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1: TeD-SPAD Framework", + "text_level": 1, + "bbox": [ + 509, + 354, + 761, + 369 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 Inputs:", + "2 Datasets: $\\mathbb{D}_{\\text{action}}$ , $\\mathbb{D}_{\\text{anomaly}}$", + "3 #Epochs: max_anon_epoch, max_ad_epoch", + "4 Learning Rates: $\\alpha_{AD},\\alpha_B,\\alpha_T$", + "5 Hyperparameters: $\\mu, \\omega$", + "6 Output: $\\theta_{AD},\\theta_{A}$", + "7 Model Initialization:", + "8 Initialize $\\theta_T$ with Kinetics400 weights [3];", + "9 Initialize $\\theta_B$ with SimCLR ImageNet weights [7].", + "10 Initialize $\\theta_{A}\\gets \\theta_{A} - \\alpha_{A}\\nabla_{\\theta_{A}}(\\mathcal{L}_{L1}(\\theta_{A}))$ (Ref. Eq. 3)", + "11 Anonymization Training:", + "12 for $e_0\\gets 1$ to max_anon_epoch do", + "13 Step-1 \n14 $\\theta_{A}\\gets \\theta_{A} - \\alpha_{A}\\nabla_{\\theta_{A}}(\\mathcal{L}_{T}(\\theta_{A},\\theta_{T}) - \\omega L_{B}(\\theta_{A},\\theta_{B}))$ \n15 Step-2 \n16 $\\theta_T\\gets \\theta_T - \\alpha_T\\nabla_{\\theta_T}(\\mathcal{L}_T(\\theta_T,\\theta_A))$ (Ref. Eq. 6) \n17 $\\theta_B\\gets \\theta_B - \\alpha_B\\nabla_{\\theta_B}(\\mathcal{L}_B(\\theta_B,\\theta_A))$", + "17 end", + "18 Feature Extraction on $\\mathbb{D}_{\\text{anomaly}}$ :", + "19 $\\mathbb{F}_{anomaly} = \\{f_T(f_A(X^i))\\mid \\forall X^i\\in \\mathbb{D}_{anomaly}\\}$", + "20 Privacy-Preserved Anomaly Detection Training:", + "21 for $e_0\\gets 1$ to max_ad_epoch do", + "22 $\\theta_{AD}\\gets \\theta_{AD} - \\alpha_{AD}\\nabla_{\\theta_{AD}}(L_{AD}(\\theta_{AD},\\mathbb{F}_{anomaly}))$", + "23 end" + ], + "bbox": [ + 503, + 375, + 866, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 776, + 630, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 801, + 601, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UCF-Crime [43] is the first large-scale weakly supervised video anomaly detection dataset. It contains 1,900 videos totaling 128 hours of untrimmed CCTV surveillance footage from a variety of different scenes. The videos contain 13 crime-based anomalies such as Arrest, Fighting, and", + "bbox": [ + 498, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13602", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Shoplifting in real world scenes.", + "bbox": [ + 76, + 90, + 294, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "XD-Violence [48] is currently the largest weakly supervised video anomaly detection dataset, totaling 217 hours of untrimmed video. All of its anomaly categories are related to violent activities. The videos are gathered from various types of cameras, movies, and games, resulting in a unique blend of scenes for increased difficulty.", + "bbox": [ + 75, + 106, + 468, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ShanghaiTech [28] is a medium-scale anomaly detection dataset containing videos covering 13 different scenes with various types of anomalies. While it was published as an unsupervised anomaly detection dataset, we use the weakly supervised rearrangement proposed by [54].", + "bbox": [ + 75, + 198, + 468, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VISPR [32] is an image dataset labelled with 68 privacy-related attributes, including gender, hair color, clothes, etc. It provides a multi-class classification problem for us to evaluate privacy-preservation on. The split we use for evaluation along with training details can found in Supp. Sec. B.", + "bbox": [ + 75, + 273, + 468, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UCF101 [42] is a very common dataset in action recognition, and its relative simplicity makes it practical for learning an anonymization model on, demonstrated in [13]. In this work, split-1 of UCF101 is used as $\\mathbb{D}_{\\text {action }}$ .", + "bbox": [ + 75, + 364, + 468, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 436, + 294, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Network Architecture Details $f_{A}$ is a UNet [39] model that transforms raw input frames into anonymized frames. I3D [3] is used for $f_{T}$ , to first learn anonymized action classification, then to extract anonymized features. A ResNet-50 [17] is our $f_{B}$ model during the anonymization training. $f_{AD}$ is MGFN [8], which consists of a shortcut convolution, a self-attentional convolution, and a feed-forward network.", + "bbox": [ + 75, + 460, + 468, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Process Details First we perform anonymization training for 80 epochs. Adam [22] optimizer is used for all models with a learning rate is 1e-4 with a batch size of 8. Loss weight $\\omega = 0.1$ and margin $\\mu = 1$ in a default setting. We train the MGFN [8] model using default hyperparameters. $f_{B}$ evaluation uses a batch size of 32 and a base learning rate of 1e-3, which follows a linear warmup and a step-based scheduler that drops by a factor of 1/5 upon loss stagnation.", + "bbox": [ + 75, + 566, + 468, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input Details For all experiments, we crop each image to a scale of 0.8, then resize to input resolution $224 \\times 224$ . Clips consist of 16 frames and are sampled from a random start with a skip-rate of 2. For anonymization training we utilize standard augmentations like random erase, random crop, horizontal flip and random color jitter. To maintain temporal consistency, augmentations are applied equally to every frame within each clip.", + "bbox": [ + 75, + 704, + 468, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Feature Extraction Given a raw input video $\\mathbf{X}^{\\mathrm{i}}\\in$ $\\mathbb{D}_{anomaly}$ , we first extract $S$ clips, where $S$ is the amount of non-overlapping 16-frame clips in $\\mathbf{X}$ . We sequentially pass each clip $S^{(j)}$ first through our anonymizer $f_{A}$ , then our feature extractor $f_{T}$ Extracted features $f_{T}(f_{A}(\\mathbf{X}^{i}))$", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "have the shape $S \\times C$ , where $C$ is the dimensionality of the feature vector. Specifically, features are extracted following the average pooling after the mix_5c layer of I3D and have a dimensionality of 2048.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocol and Performance Metrics To evaluate the learned anonymization function $f_{A}$ , we follow standard protocols of cross-dataset evaluation [13, 49]. In this protocol, testing videos of $\\mathbb{D}_{\\text{anomaly}}$ are anonymized by $f_{A}$ , and frame-level predictions are obtained through the $f_{T}$ and $f_{AD}$ . The calculated ROC AUC is used to evaluate performance on UCF-Crime and ShanghaiTech, and AP is used for XD-Violence. A higher AUC/AP is considered a more accurate anomaly localization. In order to evaluate the privacy leakage, the learned $f_{A}$ is utilized to anonymize the privacy dataset $\\mathbb{D}_{\\text{privacy}}$ to train and evaluate a target privacy model $f_{B}'$ . Privacy leakage is measured in terms of the performance of the target $f_{B}'$ on the test set of $\\mathbb{D}_{\\text{privacy}}$ . Since the privacy dataset is multi-label, the privacy leakage is measured in terms of mean average precision averaged across classes (cMAP).", + "bbox": [ + 496, + 152, + 892, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "More implementation details are in Supp. Sec. C.", + "bbox": [ + 500, + 393, + 826, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Privacy-Preserving Baselines", + "text_level": 1, + "bbox": [ + 500, + 417, + 758, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We run well-known self-supervised privacy-preservation techniques for video anomaly detection. In order to maintain a fair comparison across methods, we utilize the exact same network architectures and training process.", + "bbox": [ + 496, + 440, + 890, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Downsampling Baselines For Downsample-2x and Downsample-4x, we simply resize the input frames to a lower resolution by a factor of 2 (112 × 112) and 4 (56 × 56).", + "bbox": [ + 496, + 501, + 890, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object-Detector Based Obfuscation Baselines Obfuscation techniques are based on first detecting the person, followed by removing (i.e. blackening) or blurring them. Both obfuscation techniques use MS-COCO [26] pretrainedYOLO [37] object detector to obtain bounding boxes for person object class. We utilizeYOLOv51 implementation with yolov5x as backbone. The detected bounding boxes are assigned to pixel value 0 for the Blackening-based baseline. For the Blurring-based baselines, a Gaussian filter with kernel $k = 13$ and variance $\\sigma = 10$ is utilized.", + "bbox": [ + 496, + 546, + 890, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SPAct [13] Baseline We utilize official implementation2. For a fair comparison with our method, we utilize the exact same utility model I3D and privacy model ResNet-50.", + "bbox": [ + 496, + 696, + 890, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Evaluation on Benchmark Anomaly Datasets", + "text_level": 1, + "bbox": [ + 500, + 751, + 880, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare prior privacy-preserving methods to our method on 3 well-known anomaly detection benchmark datasets. Since privacy-preservation deals with both utility (i.e. anomaly) and privacy, we show results in form of a trade-off plot as presented in Fig. 3. Compared to the prior best method [13], our method is able to remove $19.9\\%$", + "bbox": [ + 496, + 775, + 890, + 864 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/ultralytics/yolov5", + "bbox": [ + 517, + 875, + 718, + 887 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2https://github.com/DAVEISHAN/SPAct", + "bbox": [ + 519, + 887, + 736, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13603", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/504be69d5bbc3416948eaa19b04a10cf9f53fccd18922b5a4f6d235e1df6b522.jpg", + "image_caption": [ + "(a) UCF-Crime" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 336, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d36aa73841e905a493ee5cbdd52644fa465295083822de88b62d6030bc0f157.jpg", + "image_caption": [ + "(b) XD-Violence", + "Figure 3: Trade-off plots between anomaly detection benchmarks [43, 48, 28] AUC and VISPR [32] privacy attribute prediction cMAP for different privacy-preserving methods. Optimal trade-off point is top left of plot (higher AD performance, lower PA prediction ability)." + ], + "image_footnote": [], + "bbox": [ + 344, + 88, + 602, + 236 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ee4ec2b383c638fdfda538a350b540ca47a49ae6b6b3853fd024765bda878aa.jpg", + "image_caption": [ + "(c) ShanghaiTech" + ], + "image_footnote": [], + "bbox": [ + 609, + 89, + 864, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "more privacy with a slightly better utility score $(1.19\\%)$ . This strongly supports our claim that promoting temporal distinctiveness during anonymization better aligns with anomaly detection tasks. Numeric data behind Fig. 3 plots can be found in Supp. Sec. D.", + "bbox": [ + 75, + 316, + 468, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 402, + 261, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 shows visual examples of the model outputs in different videos. We note that to the human eye, it is difficult to tell what is going on in each video, yet the anomaly detection model is still able to achieve high performance. Fig. 4a-4d include human subjects, yet none of their private attributes such as face and clothing are visible, and therefore cannot be used to make unfair decisions. In Fig. 4c, 4d, the background shop can still be seen, which is useful context for identifying the shoplifting and robbery anomalies.", + "bbox": [ + 75, + 428, + 468, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 is a graph of the anomaly score output from both the raw and anonymized video models for each frame of an input video. The blue box shows the ground truth anomalous frames. We see that the green anonymized anomaly scores are similar to the red raw scores while still increasing the anomaly scores on the ground truth anomaly, showing that our anonymization technique maintains utility performance. See Supp. Sec. D for more qualitative results.", + "bbox": [ + 75, + 564, + 470, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Evidence for Privacy Leakage at Feature-level", + "text_level": 1, + "bbox": [ + 76, + 695, + 467, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In anomaly feature representation learning, the anomaly detection algorithms do not directly work with the input videos, the videos are first passed through an action classifier to compute features. Even though the action recognition model sees the original videos, it is not certain whether the private information gets passed to the features. In order to confirm this, we create a simple fully connected network to predict VISPR private attributes, in the same fashion that we evaluate privacy for our other experiments. We stack the same VISPR image 16 times to create a video clip, then extract the clip features through $f_{T}$ . Our baseline uses unmodified input images passed through a pretrained Ki", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/732d05a70aec4776af6a6135129ebd6bf0e90df4919c5c5a0e1138d354ac4581.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Feature Extraction ModelVISPR
Privacy cMAP (%)Privacy Reduction (%)
Kinetics400 pretrained63.150.0
SPAct [13] Anonymized55.6011.96
Our Anonymized52.3017.18
", + "bbox": [ + 498, + 316, + 893, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "netics400 [20] I3D [3] model, with the other experiments using a paired anonymizer and tuned I3D model. Detailed explanations of this process are found in Supp. Sec. C. Through experimentation, we find that the action classifier latent features do in fact leak private information, therefore this private information gets passed into the anomaly detector. Table 1 demonstrates empirical evidence of this.", + "bbox": [ + 496, + 454, + 893, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 568, + 653, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of different utility losses $\\mathcal{L}_T$ We study the effect of different utility losses during the anonymization process on the final privacy vs. utility anomaly detection performance in Table 2. From row-(a,b), we see a clear reduction in privacy at no cost of anomaly performance; which demonstrates the effectiveness of our proposed temporally-distinct triplet loss $\\mathcal{L}_D$ .", + "bbox": [ + 496, + 592, + 893, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To this extent, we also implement the contrary objective to our $\\mathcal{L}_D$ which promotes temporal-invariance $\\mathcal{L}_I$ in the utility branch of the anonymization training. We implement", + "bbox": [ + 496, + 698, + 893, + 744 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/350633f4732a541278f5112e214ed22cfb2d90476f6d8f098b56d5667d387965.jpg", + "table_caption": [ + "Table 1: Quantitative evidence of action classification model features leaking privacy. Features from each model used to predict privacy attributes. Red indicates higher privacy leakage." + ], + "table_footnote": [], + "table_body": "
Utility Loss during Anonymization (L_T)VISPR Privacy cMAP(%)↓)UCF-Crime Anomaly AUC(%)↑)
(a)L_CE52.7173.93
(b)L_CE + L_D42.2174.81
(c)L_CE + L_I45.6469.52
", + "bbox": [ + 513, + 755, + 880, + 858 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Ablation with different utility losses during the anonymization process. Bold indicates best trade-off.", + "bbox": [ + 498, + 867, + 892, + 897 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13604", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8321a9e33e2b12f8faa319d453c7a10aface53ba71b28506d2eb99213154e55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 89, + 442, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6eebd3155d2efefa41ec86e1617a22c31c46130b303846c9024b2581ae7444c.jpg", + "image_caption": [ + "(a) Fighting018_x264.mp4", + "(c) Shoplifting048_x264.mp4" + ], + "image_footnote": [], + "bbox": [ + 80, + 204, + 441, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/250498cf4e879a6d3604d71777892b1ec7f3d003af4bce5fde5a956505898d71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 89, + 890, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfce969b4b19fb4c33655c199374dc0bd156dd5a8d809728239125e6f67c7093.jpg", + "image_caption": [ + "(b) Shooting040_x264.mp4", + "(d) Robbery023_x264.mp4" + ], + "image_footnote": [], + "bbox": [ + 529, + 204, + 890, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b45179a14f44024d90fe51c43c4228acab1da763fe6448c342fa8d13869d5be.jpg", + "image_caption": [ + "Figure 4: Qualitative results of our anonymization training on UCF-Crime [43] dataset. For each case, the top row shows the raw video frames, and the bottom row shows the frames after being passed through the anonymizer.", + "Figure 5: Frame-level anomaly score plot from XD-Violence. Green line shows our anonymized model, red line is the raw input model, both compared to the blue ground truth line. The below visualizations show uniformly sampled frames from the video." + ], + "image_footnote": [], + "bbox": [ + 200, + 366, + 772, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "this using well-known self-supervised works [36, 12]. From row-(c) we see that temporal-invariance objective is not well-suited for anomaly detection utility tasks and results in a significant drop of $6\\%$ . We provide extensive experiments with $\\mathcal{L}_I$ and its explanation in Supp. Sec. D.", + "bbox": [ + 76, + 690, + 468, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of different temporal distinctiveness objectives:", + "text_level": 1, + "bbox": [ + 76, + 773, + 454, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our objective of temporal distinctiveness can also be achieved through a contrastive loss. To this end we utilize the implementation of local-local temporal contrastive loss (LLTC) of [12]. It achieves $75.06\\%$ AUC at $42.44\\%$ cMAP (Table 3), which is closely comparable to our results using triplet loss. It is important to note that LLTC loss significantly increases the computation (GPU memory require", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ment) compared to triplet loss. It requires 8 clips (4 clips $\\times$ 2 augmented view) which results in 132.45G FLOPs compared to 3 clips of triplet loss requiring only 49.67G FLOPs in our experimental setup.", + "bbox": [ + 496, + 690, + 890, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Relative Weightage of $\\mathcal{L}_D$ Here we test the effect of changing the weight of the additional temporally-distinct triplet loss. Table 4 shows that weighting the loss at 0.1 achieves our best results of $32.25\\%$ relative increase in privacy with only a $3.69\\%$ reduction in utility performance. Without the enforced temporal distinctiveness, the utility model is limited by the quality of the reconstructed anonymized videos. The improper weighting of the temporal loss interferes with the classifier ability of the model,", + "bbox": [ + 496, + 763, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13605", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0c8aa71aad0e57f7edd7861f8c12dd69d6e0d61bf1203fc06d65dca53527da3b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Temporal Distinctiveness Loss LDVISPR Privacy cMAP(%)↓)UCF-Crime Anomaly AUC(%)(↑)
None52.7173.93
LLTC [12]42.4475.06
Triplet42.2174.81
", + "bbox": [ + 107, + 88, + 439, + 189 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b0ae30e1449879bfbf5455ea79a231aa6e877d9f646688b35b5d0cc07c2c834f.jpg", + "table_caption": [ + "Table 3: Comparison of different loss formulations to encourage temporal distinctiveness during the anonymization process. Both achieved similar results, with the triplet loss requiring significantly less computation." + ], + "table_footnote": [], + "table_body": "
TripletVISPRUCF-Crime
TemporalPrivacyAnomaly
Loss Weight ωcMAP(%) (↓)AUC(%) (↑)
052.7173.93
0.0153.8472.53
0.142.2174.81
1.055.2670.56
10.051.6769.27
", + "bbox": [ + 109, + 276, + 437, + 407 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "which can also harm the anonymization process. This suggests that action recognition loss $\\mathcal{L}_{CE}$ is still an important utility task for anomaly detection performance.", + "bbox": [ + 75, + 493, + 468, + 537 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effect of the margin in $\\mathcal{L}_D$ The proposed $\\mathcal{L}_D$ temporally-distinct triplet loss uses a margin hyperparameter $\\mu$ to allow for adjustments to the contrastive distance in the latent feature space. The only requirement is that $\\mu > 0$ . The intuition here is that a larger margin enforces greater feature spacing. As Table 5 shows, we empirically found that setting $\\mu = 1$ gives us the most robust results. A lower margin results in less temporally-distinct representations, which makes distinguishing between normal and anomalous features more difficult. On the other hand, increasing the $\\mu = 2$ results in more difficult temporal triplet loss (i.e a very high temporal distinctiveness) which may not align well with anomaly detection task.", + "bbox": [ + 75, + 537, + 470, + 734 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/fca4abafd1baf2b628be232cacfe59f40b14fc3a8951e444aa7727f5b348ec38.jpg", + "table_caption": [ + "Table 4: Comparison of using different loss weights of the temporally-distinct triplet loss during the anonymization process. The margin hyperparameter of the temporal triplet loss for each experiment was 1. Bold indicates best trade-off." + ], + "table_footnote": [], + "table_body": "
TripletVISPRUCF-Crime
TemporalPrivacyAnomaly
Loss Margin μcMAP(%) (↓)AUC(%) (↑)
0.549.7862.12
1.042.2174.81
2.058.8667.30
", + "bbox": [ + 109, + 747, + 437, + 848 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effect of temporal sampling in $\\mathcal{L}_D$ The proposed triplet loss (Eq. 4) forms the negative from the clip $\\mathbf{x}_{t'}^{(i)}$ of a different timestamp $t'$ . Distance between the timestamp of negative and anchor clip $t - t'$ is an important aspect to define temporal distinctiveness. We perform experiments with various distances as shown in Table 6. In our default setting, we use random distance as shown in the first row. From the second row, we can say that a smaller distance leads to a better anomaly score with a slight degradation in protecting privacy. At the same time, the third row suggests that enforcing temporal distinctiveness at higher distances leads to better privacy protection but at the cost of anomaly performance. This distance hyperparameter may be used as a tuning parameter to get different operating points of privacy vs. anomaly trade-off.", + "bbox": [ + 496, + 90, + 890, + 319 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/b5d9a59b3b2fd3db3906bfaf2404c4ddac3c7a21d9ee908bb1b5bde44b720e09.jpg", + "table_caption": [ + "Table 5: Comparison of using different hyperparameter margins of the temporally-distinct triplet loss during the anonymization process. Bold indicates default setup." + ], + "table_footnote": [], + "table_body": "
Negative Clip DistanceVISPR Privacy cMAP(%)（↓）UCF-Crime Anomaly AUC(%)(↑)
Random42.2174.81
846.3476.12
3228.6970.97
", + "bbox": [ + 550, + 330, + 843, + 433 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 6: Comparison of enforcing set clip sampling distance during the anonymization process. The margin hyperparameter of the temporal triplet loss for each experiment was 1.", + "bbox": [ + 496, + 449, + 890, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 518, + 619, + 535 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we highlight the importance of privacy, a previously neglected aspect of video anomaly detection. We present TeD-SPAD, a framework for applying Temporal Distinctiveness to Self-supervised Privacy-preserving video Anomaly Detection. TeD-SPAD demonstrates the effectiveness of using a temporally-distinct triplet loss while anonymizing an action recognition model, as it enhances feature representation temporal distinctiveness, which complements the downstream anomaly detection model. By effectively destroying spatial private information, we remove the model's ability to use this information in its decision-making process. As a future research direction, this framework can be extended to other tasks, such as spatiotemporal anomaly detection. The anonymizing encoder-decoder may also be made more powerful with techniques using recent masked image modeling. It is our hope that this work contributes to the development of more responsible and unbiased automated anomaly detection systems.", + "bbox": [ + 496, + 545, + 890, + 818 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 830, + 687, + 847 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Science Foundation (NSF) and Center for Smart Streetscapes (CS3) under NSF Cooperative Agreement No. EEC-2133516.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13606", + "bbox": [ + 480, + 924, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shafiq Ahmad, Pietro Morerio, and Alessio Del Bue. Person re-identification without identification via event anonymization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 2", + "[2] Shafiq Ahmad, Gianluca Scarpellini, Pietro Morerio, and Alessio Del Bue. Event-driven re-id: A new benchmark and method towards privacy-preserving person re-identification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 459–468, 2022. 2", + "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 4, 5, 6, 7", + "[4] Chun-Fu Chen, Shaohan Hu, Zhonghao Shi, Prateek Gulati, Bill Moriarty, Marco Pistoia, Vincenzo Piuri, and Pierangela Samarati. Mass: Multi-attribute selective suppression. arXiv preprint arXiv:2210.09904, 2022. 2", + "[5] Haoyang Chen, Xue Mei, Zhiyuan Ma, Xinhong Wu, and Yachuan Wei. Spatial-temporal graph attention network for video anomaly detection. Image and Vision Computing, 131:104629, Mar. 2023. 3", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations, June 2020. arXiv:2002.05709 [cs, stat]. 2", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 4, 5", + "[8] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 387-395, 2023. 2, 3, 4, 5, 6", + "[9] Yang Cong, Junsong Yuan, and Ji Liu. Sparse reconstruction cost for abnormal event detection. In CVPR 2011, pages 3449-3456, June 2011. ISSN: 1063-6919. 2", + "[10] Ji Dai, Behrouz Saghafi, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Towards privacy-preserving recognition of human activities. In 2015 IEEE international conference on image processing (ICIP), pages 4238-4242. IEEE, 2015. 2", + "[11] Ji Dai, Behrouz Saghafi, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Towards privacy-preserving recognition of human activities. In 2015 IEEE International Conference on Image Processing (ICIP), pages 4238-4242, Sept. 2015. 2", + "[12] Ishan Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. TCLR: Temporal contrastive learning for video representation. Computer Vision and Image Understanding, 219:103406, June 2022. 8, 9", + "[13] Ishan Rajendrakumar Dave, Chen Chen, and Mubarak Shah. Spact: Self-supervised privacy preservation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 5, 6, 7", + "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4", + "[15] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition, 2021. 3", + "[16] Giacomo Giorgi, Wisam Abbasi, and Andrea Saracino. Privacy-Preserving Analysis for Remote Video Anomaly Detection in Real Life Environments. Journal of Wireless Mobile Networks, Ubiquitous Computing, and Dependable Applications, 13(1):112-136, Mar. 2022. 2", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[18] Carlos Hinojosa, Miguel Marquez, Henry Arguello, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Privhar: Recognizing human actions from privacy-preserving lens. In European Conference on Computer Vision, pages 314–332. Springer, 2022. 2", + "[19] Chao Huang, Chengliang Liu, Jie Wen, Lian Wu, Yong Xu, Qiuping Jiang, and Yaowei Wang. Weakly Supervised Video Anomaly Detection via Self-Guided Temporal Discriminative Transformer. IEEE Transactions on Cybernetics, pages 1-14, 2022. Conference Name: IEEE Transactions on Cybernetics. 3, 4", + "[20] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The Kinetics Human Action Video Dataset, May 2017. arXiv:1705.06950 [cs]. 7", + "[21] Junho Kim, Young Min Kim, Yicheng Wu, Ramzi Zahred-dine, Weston A. Welge, Gurunandan Krishnan, Sizhuo Ma, and Jian Wang. Privacy-Preserving Visual Localization with Event Cameras, Dec. 2022. arXiv:2212.03177 [cs]. 2", + "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[23] B. Ravi Kiran, Dilip Mathew Thomas, and Ranjith Parakkal. An Overview of Deep Learning Based Methods for Unsupervised and Semi-Supervised Anomaly Detection in Videos. Journal of Imaging, 4(2):36, Feb. 2018. Number: 2 Publisher: Multidisciplinary Digital Publishing Institute. 2", + "[24] Ming Li, Jun Liu, Hehe Fan, Jia-Wei Liu, Jiahe Li, Mike Zheng Shou, and Jussi Keppo. STPrivacy: Spatio-Temporal Tubelet Sparsification and Anonymization for Privacy-preserving Action Recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 2", + "[25] Shuo Li, Fang Liu, and Licheng Jiao. Self-Training Multi-Sequence Learning with Transformer for Weakly Supervised Video Anomaly Detection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):1395–1403, June 2022. 3", + "[26] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13607", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 6", + "[27] Jixin Liu and Leilei Zhang. Indoor privacy-preserving action recognition via partially coupled convolutional neural network. In 2020 International Conference on Artificial Intelligence and Computer Engineering (ICAICE), pages 292-295. IEEE, 2020. 2", + "[28] W. Liu, D. Lian W. Luo, and S. Gao. Future frame prediction for anomaly detection - a new baseline. In 2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 6, 7", + "[29] Zelun Luo, Jun-Ting Hsieh, Niranjan Balachandar, Serena Yeung, Guido Pusiol, Jay Luxenberg, Grace Li, Li-Jia Li, N Lance Downing, Arnold Milstein, and Li Fei-Fei. Computer Vision-Based Descriptive Analytics of Seniors' Daily Activities for Long-Term Health Monitoring. 2", + "[30] Hui Lv, Chuanwei Zhou, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Localizing anomalies from weakly-labeled videos. IEEE transactions on image processing, 30:4505-4515, 2021. 3, 4", + "[31] Saemi Moon, Myeonghyeon Kim, Zhenyue Qin, Yang Liu, and Dongwoo Kim. Anonymization for skeleton action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 15028-15036, 2023. 2", + "[32] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Towards a visual privacy advisor: Understanding and predicting privacy risks in images. In IEEE International Conference on Computer Vision (ICCV), 2017. 2, 6, 7", + "[33] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning Memory-guided Normality for Anomaly Detection, Mar. 2020. arXiv:2003.13228 [cs]. 2", + "[34] Mandela Patrick, Yuki M Asano, Polina Kuznetsova, Ruth Fong, João F Henriques, Geoffrey Zweig, and Andrea Vedaldi. On compositions of transformations in contrastive self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9577-9587, 2021. 3", + "[35] Oluwatoyin P. Popoola and Kejun Wang. Video-Based Abnormal Human Behavior Recognition—A Review. IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), 42(6):865-878, Nov. 2012. Conference Name: IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews). 2", + "[36] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6964-6974, 2021. 8", + "[37] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You Only Look Once: Unified, Real-Time Object Detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, June 2016. ISSN: 1063-6919. 6", + "[38] Zhongzheng Ren, Yong Jae Lee, and Michael S Ryoo. Learning to anonymize faces for privacy preserving action" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "detection. In Proceedings of the European conference on computer vision (ECCV), pages 620-636, 2018. 2", + "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 6", + "[40] Michael S Ryoo, Brandon Rothrock, Charles Fleming, and Hyun Jong Yang. Privacy-preserving human activity recognition from extreme low resolution. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 2", + "[41] Hitesh Sapkota and Qi Yu. Bayesian nonparametric submodular video partition for robust anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3212-3221, 2022. 3", + "[42] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild, Dec. 2012. arXiv:1212.0402 [cs]. 6", + "[43] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-World Anomaly Detection in Surveillance Videos. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6479-6488, Salt Lake City, UT, June 2018. IEEE. 1, 2, 5, 7, 8", + "[44] Qiyue Sun and Yang Yang. Unsupervised video anomaly detection based on multi-timescale trajectory prediction. Computer Vision and Image Understanding, 227:103615, Jan. 2023. 2", + "[45] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975-4986, 2021. 3, 4", + "[46] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In ECCV, 2022. 3", + "[47] Peng Wu and Jing Liu. Learning Causal Temporal Relation and Feature Discrimination for Anomaly Detection. IEEE Transactions on Image Processing, 30:3513-3527, 2021. Conference Name: IEEE Transactions on Image Processing. 3", + "[48] Peng Wu, jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 6, 7", + "[49] Zhenyu Wu, Haotao Wang, Zhaowen Wang, Hailin Jin, and Zhangyang Wang. Privacy-preserving deep action recognition: An adversarial learning framework and a new dataset. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 2, 3, 6", + "[50] Hongchun Yuan, Zhenyu Cai, Hui Zhou, Yue Wang, and Xiangzhi Chen. TransAnomaly: Video Anomaly Detection Using Video Vision Transformer. IEEE Access, 9:123977-123986, 2021. Conference Name: IEEE Access. 2", + "[51] Sharare Zehtabian, Siavash Khodadadeh, Ladislau Bölönii, and Damla Turgut. Privacy-preserving learning of human" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13608", + "bbox": [ + 480, + 925, + 519, + 936 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "activity predictors in smart environments. In IEEE INFOCOM 2021-IEEE Conference on Computer Communications, pages 1-10. IEEE, 2021. 2", + "[52] Chen Zhang, Guorong Li, Yuankai Qi, Shuhui Wang, Laiyun Qing, Qingming Huang, and Ming-Hsuan Yang. Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16271-16280, 2023. 3", + "[53] Zhixiang Zhang, Thomas Cilloni, Charles Walter, and Charles Fleming. Multi-scale, class-generic, privacy-preserving video. *Electronics*, 10(10):1172, 2021. 2", + "[54] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1237-1246, 2019. 3, 6", + "[55] Yi Zhu and Shawn Newsam. Motion-Aware Feature for Improved Video Anomaly Detection, July 2019. arXiv:1907.10211 [cs, eess]. 3" + ], + "bbox": [ + 78, + 90, + 468, + 387 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13609", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/a12231ff-0413-4e22-8eb5-06598dd6fed6_model.json b/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/a12231ff-0413-4e22-8eb5-06598dd6fed6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d15b66805dad119fbc23769fbddd25b0f8ad53b2 --- /dev/null +++ b/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/a12231ff-0413-4e22-8eb5-06598dd6fed6_model.json @@ -0,0 +1,2809 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.88, + 0.177 + ], + "angle": 0, + "content": "TeD-SPAD: Temporal Distinctiveness for Self-supervised Privacy-preservation for video Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.204, + 0.825, + 0.278 + ], + "angle": 0, + "content": "Joseph Fioresi, Ishan Rajendrakumar Dave, Mubarak Shah Center for Research in Computer Vision, University of Central Florida, Orlando, USA {joseph.fioresi, ishanrajendrakumar.dave}@ucf.edu, shah@crcv.ucf.edu Project Page: https://joefioresi718.github.io/TeD-SPAD_webpage/" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.341, + 0.473, + 0.705 + ], + "angle": 0, + "content": "Video anomaly detection (VAD) without human monitoring is a complex computer vision task that can have a positive impact on society if implemented successfully. While recent advances have made significant progress in solving this task, most existing approaches overlook a critical real-world concern: privacy. With the increasing popularity of artificial intelligence technologies, it becomes crucial to implement proper AI ethics into their development. Privacy leakage in VAD allows models to pick up and amplify unnecessary biases related to people's personal information, which may lead to undesirable decision making. In this paper, we propose TeD-SPAD, a privacy-aware video anomaly detection framework that destroys visual private information in a self-supervised manner. In particular, we propose the use of a temporally-distinct triplet loss to promote temporally discriminative features, which complements current weakly-supervised VAD methods. Using TeD-SPAD, we achieve a positive trade-off between privacy protection and utility anomaly detection performance on three popular weakly supervised VAD datasets: UCF-Crime, XD-Violence, and ShanghaiTech. Our proposed anonymization model reduces private attribute prediction by \\(32.25\\%\\) while only reducing frame-level ROC AUC on the UCF-Crime anomaly detection dataset by \\(3.69\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.725, + 0.21, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Machine learning-driven technologies are increasingly being adopted by society. The progress in cloud computing has enabled the deployment of even computationally intensive technologies in the public space. One such application is video anomaly detection (VAD) in autonomous video analytics. VAD is a video understanding task that aims to identify the temporal location of anomalous events occurring in long continuous videos without human supervision. An anomaly can be defined as any unusual event, such as a traffic accident, an elderly person falling, or a fire. Proper" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.307, + 0.73, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.308, + 0.892, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.504, + 0.895, + 0.629 + ], + "angle": 0, + "content": "Figure 1: Single frame from video Shoplifting052_x264.mp4 of UCF-Crime [43] showing the types of private attributes (shown in red on the left image) leaked in visual data. In the right image we show the anonymized frame where these attributes are barely visible. The graph demonstrates the ability of our self-supervised framework to mitigate privacy leakage compared to prior work SPAct [13]. At a similar anomaly detection utility performance, our method prevents \\(32\\%\\) of the visual private leakage compared to raw data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.701 + ], + "angle": 0, + "content": "application of this technology can result in faster response times to anomalies, without the need for human resources to monitor camera feeds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.885 + ], + "angle": 0, + "content": "However, public adoption of such AI technologies brings justifiable concern about safety and their decision-making capabilities. Many of these concerns center around privacy violations and accuracy. VAD is an application where visual privacy leakage and data bias are exceedingly important issues. Sending videos to remote computers or cloud services to process results in unnecessary privacy leakage for people who are not directly involved in anomalous events. Additionally, an application employing a standard RGB video will incorporate any bias found in its training set, leading to potentially unfair decisions. An illustration of privacy leakage is shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Recently, there have been interesting attempts to prevent" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "13598" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.469 + ], + "angle": 0, + "content": "visual privacy leakage in action recognition. Some of the approaches utilize input downsizing-based solutions [40, 10, 27] and object-detection-dependent obfuscation-based formulations [38, 53]. Wu et al. [49] proposed an adversarial training-based framework where they train an anonymization function to remove privacy-preservation. Dave et al. [13] proposed a self-supervised privacy-preserving framework that does not require privacy labels and achieves similar performance to the supervised method [49]. Since many weakly-supervised anomaly detection (WSAD) methods rely on the pretrained features of action recognition, privacy-preserving action recognition seems like a promising candidate for privacy-preserving anomaly detection. However, detecting anomalies does not align well with privacy-preserved action recognition videos. The use of short videos in action recognition encourages the use of temporally-invariant features, where the features of clips at distinct timesteps should be the same. Conversely, detecting anomalies in long, untrimmed videos requires temporally-distinct reasoning, where features of clips at distinct timesteps of a video should be different, to determine whether events in the same scene are anomalous. This is why most existing anomaly detection methods focus on refining the features of pretrained video encoders to increase their temporal separability." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.469, + 0.473, + 0.666 + ], + "angle": 0, + "content": "To the best of our knowledge, privacy-preservation in video anomaly detection is an unexplored area in computer vision. Building on the existing self-supervised privacy-preserving action recognition framework [13], we propose a more aligned utility branch for anomaly detection. To achieve this, we use a novel temporally-distinct triplet loss to promote temporal distinctiveness during anonymization training. Once the anonymization function is learned through our proposed anonymization framework, we apply it to the anomaly dataset, which ensures privacy leakage mitigation in privacy-sensitive anomaly detection tasks. We use these anonymized features to train the current state-of-the-art WSAD method, MGFN [8]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.666, + 0.473, + 0.756 + ], + "angle": 0, + "content": "To evaluate privacy-preserving performance in anomaly detection, we adopt protocols from prior action recognition methods, where we report the utility performance on WSAD task on widely used anomaly datasets (UCF-Crime [43], XD-Violence [48], and ShanghaiTech [28]) and budget performance on privacy dataset VISPR [32]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.757, + 0.424, + 0.771 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.785, + 0.47, + 0.845 + ], + "angle": 0, + "content": "- We introduce a new problem of privacy-preservation in video anomaly detection, where we identify the privacy leakage issue in existing weakly supervised anomaly detection methods." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- To address this open problem, we propose TeD-SPAD, a framework based on self-supervised privacy-preservation with a temporally-distinct triplet loss to" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.785, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "make the video anonymization process more suitable for anomaly detection." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.134, + 0.895, + 0.255 + ], + "angle": 0, + "content": "- We propose evaluation protocols for privacy vs. anomaly trade-off, demonstrating that our proposed framework outperforms prior methods by significant margins across all anomaly detection benchmarks. On the widely used UCF-Crime dataset, our method is able to eliminate \\(32.25\\%\\) of the privacy leakage at a cost of only a \\(3.96\\%\\) reduction in frame-level AUC performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.269, + 0.651, + 0.285 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.296, + 0.652, + 0.31 + ], + "angle": 0, + "content": "Privacy Preservation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.895, + 0.477 + ], + "angle": 0, + "content": "We observe that many works preserve visual privacy at capture time by using non-intrusive sensors such as thermal imaging, depth cameras, or event cameras [29, 18, 21, 1]. Other works allow for raw RGB visual information to be captured, but make an effort to protect the subject privacy in such a way that the data is still useful in a utility task. Earlier efforts aimed at dealing with visual privacy include image downsampling [11] or blocking/blurring privacy-related objects located using pretrained object detectors. Both of these obfuscation methods were shown to reduce utility results by more than they reduced privacy leakage [13, 24, 49]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.704 + ], + "angle": 0, + "content": "Recent developments have yielded numerous privacy-preserving approaches to action recognition [2, 51, 31]. Wu et al. released an action dataset with privacy labels, PA-HMDB [49]. They use an adversarial learning framework to obfuscate the privacy features using supervised privacy labels. MaSS [4] uses a similar framework to Wu et al. [49], except it adapts a compound loss to flexibly preserve certain attributes instead of destroying them. STPrivacy [24] upgraded the general framework to work with a transformer anonymizing block, masking entire video tubelets unnecessary for action recognition. Following Dave et al.'s SPAcT [13], we adopt a similar self-supervised adversarial anonymization framework without the use of the privacy labels, using NT-Xent [6] contrastive loss in the budget branch to mitigate spatial privacy leakage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.705, + 0.64, + 0.719 + ], + "angle": 0, + "content": "Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.885 + ], + "angle": 0, + "content": "With such a high volume of available video footage, it is infeasible to create sufficient labelled data to solve supervised VAD. Therefore, many works explore unsupervised methods. These generally train a reconstruction model, then either reconstruct the current frame or try to predict the next frame, signaling an anomaly when reconstruction error is high [9, 35, 23, 33, 50, 44]. Giorgi et al. [16] used an autoencoder with differential privacy, generating anomaly scores from the noisy reconstructed images. This method helped retain some level of subject privacy, but was only evaluated on image quality metrics." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Sultani et al. [43] brought weak supervision to VAD," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "13599" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.346 + ], + "angle": 0, + "content": "where anomalies are labelled at the video level. These authors introduce UCF-Crime, a large scale weakly supervised dataset. They propose formulating weakly supervised VAD as a multi instance learning (MIL) problem, showcasing the benefits of temporal smoothness loss & sparsity loss. With the exception of a select few works [54, 52], all following weakly supervised methods are considered anomaly feature representation learning as they improve upon MIL formulation [55, 45, 15, 25, 19, 46, 8, 41], which involves interpreting static video features extracted using an action classifier. Zhong et al. [54] propose a rearranged weakly supervised version of ShanghaiTech. Wu et al. [48] introduce XD-Violence, a large-scale weakly supervised dataset that includes audio, bringing multi-modal fusion to VAD. In this work, we choose to focus on the weakly supervised video anomaly detection setting due to its effectiveness and low annotation effort." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.668 + ], + "angle": 0, + "content": "Most VAD works find it useful to model temporal relations between video segments [45, 15, 30, 48, 25, 46, 55, 5]. [30] are able to exploit dynamic variations in features between consecutive segments to help localize anomalies. The authors of RTFM [45] find that anomalous segments tend to have larger feature magnitudes (\\(\\ell_2\\) norm of extracted clip feature vector) than normal segments. They introduce a feature magnitude learning function to help identify anomalous segments with large feature magnitudes. MGFN [8] notices that varying scene attributes causes some normal videos to have larger feature magnitudes than anomalous, so they propose a feature magnitude contrastive loss to help capture these instances. In this work, we use the MGFN model to evaluate anomaly detection performance. Complementing these ideas, [47] demonstrates the effectiveness of explicitly encouraging feature discrimination. Intuitively, these observations can be aggregated by enforcing temporally-distinct feature representations. Building on the terminology introduced in [34] regarding invariant and distinctive representations, in this paper our goal is to enhance the distinctiveness of the representations within our utility branch." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.17, + 0.695 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.901 + ], + "angle": 0, + "content": "The central idea behind our proposed framework is to develop an anonymization function that can degrade privacy attributes during training without relying on privacy labels. Furthermore, this function must be able to maintain the performance of the weakly-supervised anomaly detection task. Fig. 2 displays a schematic diagram of the proposed framework. In Sec. 3.1, we provide a comprehensive discussion of the problem statement. Next, in Sec. 3.2, we describe the component of the framework and their initialization process. Sec. 3.3 outlines the anonymization function training, where we propose a temporally-distinct triplet loss to enhance the existing self-supervised privacy-preserving framework [13]. Once we learn the anonymization func" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "tion, in Sec. 3.4, we train the anomaly detection model using the privacy-preserved features obtained through our anonymization function. An overview of our complete framework is outlined in Algorithm Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.162, + 0.688, + 0.178 + ], + "angle": 0, + "content": "3.1. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.186, + 0.892, + 0.382 + ], + "angle": 0, + "content": "We define our problem statement similar to privacy-preserving action recognition frameworks [13, 47], but, with a different utility task. Suppose we have a video dataset \\(\\mathbb{D}_{\\text{anomaly}} = \\{X^{(i)}, Y^i\\}_{i=1}^N\\), where, \\(X^{(i)}\\) is a video-instance, \\(N\\) is the total number of samples and \\(Y^{(i)} \\in \\{0,1\\}\\) is a binary label. Considering video-level anomaly detection as a utility task \\(T\\), and privacy attribute classification as the budget task \\(B\\), the aim of a privacy-preserving system is to maintain the performance of \\(T\\) while reducing \\(B\\). To achieve this goal, the system learns an anonymization function \\(f_A\\), which modifies the original raw data. This goal of privacy-preservation could be fundamentally expressed as following criteria:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.892, + 0.443 + ], + "angle": 0, + "content": "Criterion-1: The performance of the utility task should not degrade from the original performance, i.e. loss \\(\\mathcal{L}_T\\) value of the utility target model \\(f_T'\\) should remain almost identical before and after applying the anonymization function." + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.455, + 0.892, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T} \\left(f _ {T} ^ {\\prime} \\left(f _ {A} ^ {*} (X)\\right), Y\\right) \\approx \\mathcal {L} _ {T} \\left(f _ {T} ^ {\\prime} (X), Y\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.53 + ], + "angle": 0, + "content": "Criterion-2: Applying anonymization function should increase the loss \\(\\mathcal{L}_B\\) for budget \\(B\\) task of target budget model \\(f_B^\\prime\\)." + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.53, + 0.892, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {B} \\left(f _ {B} ^ {\\prime} \\left(f _ {A} ^ {*} (X)\\right)\\right) \\gg \\mathcal {L} _ {B} \\left(f _ {B} ^ {\\prime} (X)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Regarding weakly-supervised anomaly detection (WSAD), most of the existing methods require multi-stage training, meaning they are not end-to-end. This presents a challenge for incorporating it as a utility task in anonymization training. By contrast, privacy-preserving action recognition frameworks [13, 49] have an end-to-end utility task (i.e action recognition), making it more straightforward to include." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Since most WSAD methods rely on pretrained video encoder features from large-scale action recognition training, we can utilize the exact same anonymization process of privacy-preserving action recognition [13], utilizing action recognition as a proxy-utility task on a proxy-utility action dataset, which we denote as \\(\\mathbb{D}_{\\text {action }}\\). The limitation of such anonymization training is that by focusing solely on optimizing short clips for action recognition, it neglects the importance of temporally-distinct features. This oversight leads to a notable decline in anomaly detection performance, which relies heavily on these temporally-distinct features, as the training progresses. To resolve this issue, we reformulate the anonymization utility branch to enforce temporal distinctiveness to better align with the anomaly detection task. Hence, our \\(\\mathcal{L}_T\\) is weighted sum of action" + }, + { + "type": "footer", + "bbox": [ + 0.481, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "13600" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.09, + 0.835, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.375, + 0.525, + 0.391 + ], + "angle": 0, + "content": "(b) Privacy-Preserved Anomaly Detection Workflow" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.401, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Figure 2: Full TeD-SPAD framework consisting of the proxy anonymization training followed by the privacy-preserved anomaly detection. (a) shows this proxy training, where UNet is used to anonymize frames in such a way that reduces mutual information between frames while maintaining utility performance. We complement the standard cross-entropy loss with our proposed temporally-distinct triplet loss, which enforces a difference in clip features at distinct timesteps. After training the anonymizer and feature extractor, (b) shows the privacy-preserved workflow, where the anomaly dataset videos are passed through the proxy-trained \\( f_{A} \\), \\( f_{T} \\), then into any WSAD algorithm." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.47, + 0.529 + ], + "angle": 0, + "content": "recognition loss \\((\\mathcal{L}_{CE})\\) and temporally-distinct triplet loss \\((\\mathcal{L}_D)\\) which is elaborated in Eq. 6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.547, + 0.327, + 0.563 + ], + "angle": 0, + "content": "3.2. Anonymization Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.469, + 0.649 + ], + "angle": 0, + "content": "Our anonymization framework consists of 3 main components: (1) Anonymization function \\((f_A)\\), which is a simple encoder-decoder model with a sigmoid activation. (2) Privacy removal model \\((f_B)\\), which is an image encoder. (3) Utility model \\((f_T)\\), which is a video encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.713 + ], + "angle": 0, + "content": "Network Initialization First of all, our anonymization model is pretrained to initialize as an identity function. This pretraining involves the reconstruction of the frames from \\(\\mathbb{D}_{\\text {action }}\\) using L1-reconstruction loss." + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.735, + 0.47, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L 1} = \\sum_ {c = 1} ^ {C} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} | x _ {c, h, w} - \\hat {x} _ {c, h, w} |, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.792, + 0.47, + 0.838 + ], + "angle": 0, + "content": "where \\(\\hat{x}\\) is the output of \\(f_{A}\\), \\(x\\) is an input image, and \\(C, H, W\\) corresponds to the channel, height, and width of the input image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Secondly, the privacy model \\( f_{B} \\) is initialized with the self-supervised weights of SimCLR [7] on ImageNet [14]. The video encoder \\( f_{T} \\) is pretrained with the standard action recognition weights from Kinetics400 dataset [3]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.497, + 0.726, + 0.513 + ], + "angle": 0, + "content": "3.3. Anonymization Training" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.52, + 0.891, + 0.551 + ], + "angle": 0, + "content": "Anonymization training is mainly comprised of a minimax optimization of utility loss and privacy loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.893, + 0.846 + ], + "angle": 0, + "content": "Temporally-distinct triplet loss as Utility Weakly-supervised video anomaly detection methods leverage temporal information to help localize anomalies. [30, 19, 45, 8] show the positive effect of feature magnitude (\\(\\ell_2\\) norm of clip feature vector) separability along the temporal dimension. In order to adopt the SPAcT anonymization framework to anomaly detection problems, we utilize a temporal distinctiveness objective in the utility branch. We design a temporally-distinct triplet loss, which increases the agreement between temporally-aligned clips of the same video and increases dissimilarity between representations of the clips which are temporally-misaligned. For anchor clip \\(\\mathbf{x}_t^{(i)}\\), we obtain the positive clip from the exact same timestamp, but with a differently augmented version denoted as \\(\\tilde{\\mathbf{x}}_t^{(i)}\\). Whereas, the negative clip is obtained from different timestamp \\(\\mathbf{x}_{t'}^{(i)}\\), where \\(t' \\neq t\\). This triplet of clips is passed through the utility model \\(f_T\\) to achieve features denoted as \\(\\mathbf{z}_t^{(i)}\\), \\(\\tilde{\\mathbf{z}}_t^{(i)}\\), and \\(\\mathbf{z}_{t'}^{(i)}\\). The proposed temporally-distinct triplet loss can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.854, + 0.892, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} ^ {(i)} = \\max \\{d (\\mathbf {z} _ {t} ^ {(i)}, \\tilde {\\mathbf {z}} _ {t} ^ {(i)}) - d (\\mathbf {z} _ {t} ^ {(i)}, \\mathbf {z} _ {t ^ {\\prime}} ^ {(i)}) + \\mu , 0 \\}, \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.885, + 0.891, + 0.904 + ], + "angle": 0, + "content": "where \\(d(u_j,v_j) = \\|\\mathbf{u}_j - \\mathbf{v}_j\\|_2\\) is Euclidean distance be" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "13601" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "tween two vectors \\(\\mathbf{u}\\) and \\(\\mathbf{v}\\), and \\(\\mu\\) is the controllable margin hyperparameter to determine how far to push and pull features in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.137, + 0.469, + 0.167 + ], + "angle": 0, + "content": "We utilize this loss along with the standard cross-entropy action classification loss:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.178, + 0.47, + 0.219 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C E} ^ {(i)} = - \\sum_ {c = 1} ^ {N _ {C}} \\mathbf {y} _ {c} ^ {(i)} \\log \\mathbf {p} _ {c} ^ {(i)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.224, + 0.469, + 0.269 + ], + "angle": 0, + "content": "where \\(N_{C}\\) is the total number of action classes of \\(\\mathbb{D}_{\\text {action }}\\), \\(\\mathbf{y}_c^{(i)}\\) is the ground-truth label, and \\(\\mathbf{p}_c^{(i)}\\) is the prediction vector by utility model \\(f_T\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.269, + 0.469, + 0.315 + ], + "angle": 0, + "content": "Adding both temporal distinctiveness (Eq. 4) and action classification objective to our utility branch, our overall utility loss can be expressed as follows" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.324, + 0.469, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T} = \\mathcal {L} _ {C E} + \\omega * \\mathcal {L} _ {D}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.469, + 0.381 + ], + "angle": 0, + "content": "where \\(\\omega\\) hyperparameter is the weight of temporally-distinct triplet loss with respect to cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.469, + 0.439 + ], + "angle": 0, + "content": "Privacy (i.e. budget) Loss \\(\\mathcal{L}_B\\) We utilize the same self-supervised privacy loss from [13], which removes the private information by minimizing the agreement between the frames of the same video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.469, + 0.533 + ], + "angle": 0, + "content": "Minimax Optimization After reformulating the utility loss, we use the minimax optimization process similar to [13]. It is a two-step iterative process that minimizes the utility loss and at the same time increases budget loss \\(\\mathcal{L}_B\\). At the end of this optimization, we obtain the learned anonymization function \\((f_A)\\) and utility video encoder \\((f_T)\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.54, + 0.411, + 0.557 + ], + "angle": 0, + "content": "3.4. Privacy-preserving Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.469, + 0.653 + ], + "angle": 0, + "content": "In order to detect anomalies within videos from \\(\\mathbb{D}_{\\text{anomaly}}\\), we utilize the current state-of-the-art technique, Magnitude-Contrastive Glance-and-Focus Network (MGFN) [8]. Similar to other recent works in anomaly detection, MGFN requires fixed features for each video from a pretrained video encoder for anomaly detection training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.654, + 0.469, + 0.789 + ], + "angle": 0, + "content": "Optimizing for Anomaly Detection MGFN anomaly detection is comprised of 4 main losses: (1) a standard sigmoid cross-entropy loss \\( L_{sce} \\) for snippet classification accuracy, (2) a temporal smoothing loss \\( L_{ts} \\) [43] to encourage smoothness between feature representations of consecutive segments, (3) a sparsity term \\( L_{sp} \\) [43] to discourage false positive anomalies, and (4) a novel magnitude contrastive loss \\( L_{mc} \\) to learn scene-adaptive feature distributions across videos, all which help to train a model \\( f_{AD} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.79, + 0.469, + 0.82 + ], + "angle": 0, + "content": "The training loss used in MGFN is compounded in the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.83, + 0.469, + 0.847 + ], + "angle": 0, + "content": "\\[\nL _ {A D} = L _ {s c e} + \\lambda_ {1} L _ {t s} + \\lambda_ {2} L _ {s p} + \\lambda_ {3} L _ {m c}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.9 + ], + "angle": 0, + "content": "where \\(\\lambda_1 = \\lambda_2 = 1\\), and \\(\\lambda_3 = 0.001\\). \\(f_{AD}\\) outputs frame-level anomaly scores, which are used to calculate a final ROC AUC and AP for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.214 + ], + "angle": 0, + "content": "Feature Extraction In our privacy-preserving case, we cannot use \\(\\mathbb{D}_{\\text{anomaly}}\\) directly for the feature extraction from the video encoder. We first anonymize each video \\((X^i)\\) of the dataset through the learned anonymization function \\(f_A\\) to get an anonymized dataset. For the feature extraction, we utilize the learned utility video encoder \\(f_T\\). We denote this extracted set of anonymized features as \\(\\mathbb{F}_{\\text{anomaly}} = \\{f_T(f_A(X^i)) \\mid \\forall X^i \\in \\mathbb{D}_{\\text{anomaly}}\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.223, + 0.617, + 0.239 + ], + "angle": 0, + "content": "3.5. Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.245, + 0.892, + 0.337 + ], + "angle": 0, + "content": "Let's consider the models \\( f_{A}, f_{T}, f_{B}, f_{AD} \\), parameterized by \\( \\theta_{A}, \\theta_{T}, \\theta_{B} \\), and \\( f_{AD} \\), respectively. \\( \\mathbb{D}_{\\text{action}} \\) is the proxy action recognition dataset and \\( \\mathbb{D}_{\\text{anomaly}} \\) is the downstream anomaly detection dataset. All training steps of our framework can be put together into a sophisticated form as algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.356, + 0.762, + 0.371 + ], + "angle": 0, + "content": "Algorithm 1: TeD-SPAD Framework" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.376, + 0.575, + 0.389 + ], + "angle": 0, + "content": "1 Inputs:" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.391, + 0.726, + 0.406 + ], + "angle": 0, + "content": "2 Datasets: \\(\\mathbb{D}_{\\text{action}}\\), \\(\\mathbb{D}_{\\text{anomaly}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.406, + 0.833, + 0.421 + ], + "angle": 0, + "content": "3 #Epochs: max_anon_epoch, max_ad_epoch" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.422, + 0.739, + 0.436 + ], + "angle": 0, + "content": "4 Learning Rates: \\(\\alpha_{AD},\\alpha_B,\\alpha_T\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.437, + 0.692, + 0.451 + ], + "angle": 0, + "content": "5 Hyperparameters: \\(\\mu, \\omega\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.452, + 0.64, + 0.466 + ], + "angle": 0, + "content": "6 Output: \\(\\theta_{AD},\\theta_{A}\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.467, + 0.658, + 0.479 + ], + "angle": 0, + "content": "7 Model Initialization:" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.482, + 0.803, + 0.496 + ], + "angle": 0, + "content": "8 Initialize \\(\\theta_T\\) with Kinetics400 weights [3];" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.497, + 0.85, + 0.511 + ], + "angle": 0, + "content": "9 Initialize \\(\\theta_B\\) with SimCLR ImageNet weights [7]." + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.512, + 0.867, + 0.527 + ], + "angle": 0, + "content": "10 Initialize \\(\\theta_{A}\\gets \\theta_{A} - \\alpha_{A}\\nabla_{\\theta_{A}}(\\mathcal{L}_{L1}(\\theta_{A}))\\) (Ref. Eq. 3)" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.528, + 0.69, + 0.542 + ], + "angle": 0, + "content": "11 Anonymization Training:" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.543, + 0.761, + 0.557 + ], + "angle": 0, + "content": "12 for \\(e_0\\gets 1\\) to max_anon_epoch do" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.558, + 0.857, + 0.634 + ], + "angle": 0, + "content": "13 Step-1 \n14 \\(\\theta_{A}\\gets \\theta_{A} - \\alpha_{A}\\nabla_{\\theta_{A}}(\\mathcal{L}_{T}(\\theta_{A},\\theta_{T}) - \\omega L_{B}(\\theta_{A},\\theta_{B}))\\) \n15 Step-2 \n16 \\(\\theta_T\\gets \\theta_T - \\alpha_T\\nabla_{\\theta_T}(\\mathcal{L}_T(\\theta_T,\\theta_A))\\) (Ref. Eq. 6) \n17 \\(\\theta_B\\gets \\theta_B - \\alpha_B\\nabla_{\\theta_B}(\\mathcal{L}_B(\\theta_B,\\theta_A))\\)" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.635, + 0.55, + 0.646 + ], + "angle": 0, + "content": "17 end" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.649, + 0.738, + 0.664 + ], + "angle": 0, + "content": "18 Feature Extraction on \\(\\mathbb{D}_{\\text{anomaly}}\\):" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.664, + 0.838, + 0.679 + ], + "angle": 0, + "content": "19 \\(\\mathbb{F}_{anomaly} = \\{f_T(f_A(X^i))\\mid \\forall X^i\\in \\mathbb{D}_{anomaly}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.68, + 0.838, + 0.694 + ], + "angle": 0, + "content": "20 Privacy-Preserved Anomaly Detection Training:" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.695, + 0.742, + 0.708 + ], + "angle": 0, + "content": "21 for \\(e_0\\gets 1\\) to max_ad_epoch do" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.709, + 0.857, + 0.724 + ], + "angle": 0, + "content": "22 \\(\\theta_{AD}\\gets \\theta_{AD} - \\alpha_{AD}\\nabla_{\\theta_{AD}}(L_{AD}(\\theta_{AD},\\mathbb{F}_{anomaly}))\\)" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.726, + 0.55, + 0.736 + ], + "angle": 0, + "content": "23 end" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.376, + 0.867, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.777, + 0.632, + 0.794 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.602, + 0.816 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "UCF-Crime [43] is the first large-scale weakly supervised video anomaly detection dataset. It contains 1,900 videos totaling 128 hours of untrimmed CCTV surveillance footage from a variety of different scenes. The videos contain 13 crime-based anomalies such as Arrest, Fighting, and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "13602" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.295, + 0.106 + ], + "angle": 0, + "content": "Shoplifting in real world scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.47, + 0.198 + ], + "angle": 0, + "content": "XD-Violence [48] is currently the largest weakly supervised video anomaly detection dataset, totaling 217 hours of untrimmed video. All of its anomaly categories are related to violent activities. The videos are gathered from various types of cameras, movies, and games, resulting in a unique blend of scenes for increased difficulty." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.469, + 0.274 + ], + "angle": 0, + "content": "ShanghaiTech [28] is a medium-scale anomaly detection dataset containing videos covering 13 different scenes with various types of anomalies. While it was published as an unsupervised anomaly detection dataset, we use the weakly supervised rearrangement proposed by [54]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.469, + 0.363 + ], + "angle": 0, + "content": "VISPR [32] is an image dataset labelled with 68 privacy-related attributes, including gender, hair color, clothes, etc. It provides a multi-class classification problem for us to evaluate privacy-preservation on. The split we use for evaluation along with training details can found in Supp. Sec. B." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.469, + 0.426 + ], + "angle": 0, + "content": "UCF101 [42] is a very common dataset in action recognition, and its relative simplicity makes it practical for learning an anonymization model on, demonstrated in [13]. In this work, split-1 of UCF101 is used as \\(\\mathbb{D}_{\\text {action }}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.438, + 0.295, + 0.453 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.469, + 0.566 + ], + "angle": 0, + "content": "Network Architecture Details \\(f_{A}\\) is a UNet [39] model that transforms raw input frames into anonymized frames. I3D [3] is used for \\(f_{T}\\), to first learn anonymized action classification, then to extract anonymized features. A ResNet-50 [17] is our \\(f_{B}\\) model during the anonymization training. \\(f_{AD}\\) is MGFN [8], which consists of a shortcut convolution, a self-attentional convolution, and a feed-forward network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.469, + 0.704 + ], + "angle": 0, + "content": "Training Process Details First we perform anonymization training for 80 epochs. Adam [22] optimizer is used for all models with a learning rate is 1e-4 with a batch size of 8. Loss weight \\(\\omega = 0.1\\) and margin \\(\\mu = 1\\) in a default setting. We train the MGFN [8] model using default hyperparameters. \\(f_{B}\\) evaluation uses a batch size of 32 and a base learning rate of 1e-3, which follows a linear warmup and a step-based scheduler that drops by a factor of 1/5 upon loss stagnation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Input Details For all experiments, we crop each image to a scale of 0.8, then resize to input resolution \\(224 \\times 224\\). Clips consist of 16 frames and are sampled from a random start with a skip-rate of 2. For anonymization training we utilize standard augmentations like random erase, random crop, horizontal flip and random color jitter. To maintain temporal consistency, augmentations are applied equally to every frame within each clip." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Feature Extraction Given a raw input video \\(\\mathbf{X}^{\\mathrm{i}}\\in\\) \\(\\mathbb{D}_{anomaly}\\) , we first extract \\(S\\) clips, where \\(S\\) is the amount of non-overlapping 16-frame clips in \\(\\mathbf{X}\\) . We sequentially pass each clip \\(S^{(j)}\\) first through our anonymizer \\(f_{A}\\) , then our feature extractor \\(f_{T}\\) Extracted features \\(f_{T}(f_{A}(\\mathbf{X}^{i}))\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "have the shape \\( S \\times C \\), where \\( C \\) is the dimensionality of the feature vector. Specifically, features are extracted following the average pooling after the mix_5c layer of I3D and have a dimensionality of 2048." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.893, + 0.393 + ], + "angle": 0, + "content": "Evaluation Protocol and Performance Metrics To evaluate the learned anonymization function \\( f_{A} \\), we follow standard protocols of cross-dataset evaluation [13, 49]. In this protocol, testing videos of \\( \\mathbb{D}_{\\text{anomaly}} \\) are anonymized by \\( f_{A} \\), and frame-level predictions are obtained through the \\( f_{T} \\) and \\( f_{AD} \\). The calculated ROC AUC is used to evaluate performance on UCF-Crime and ShanghaiTech, and AP is used for XD-Violence. A higher AUC/AP is considered a more accurate anomaly localization. In order to evaluate the privacy leakage, the learned \\( f_{A} \\) is utilized to anonymize the privacy dataset \\( \\mathbb{D}_{\\text{privacy}} \\) to train and evaluate a target privacy model \\( f_{B}' \\). Privacy leakage is measured in terms of the performance of the target \\( f_{B}' \\) on the test set of \\( \\mathbb{D}_{\\text{privacy}} \\). Since the privacy dataset is multi-label, the privacy leakage is measured in terms of mean average precision averaged across classes (cMAP)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.394, + 0.828, + 0.409 + ], + "angle": 0, + "content": "More implementation details are in Supp. Sec. C." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.418, + 0.759, + 0.433 + ], + "angle": 0, + "content": "4.3. Privacy-Preserving Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.892, + 0.501 + ], + "angle": 0, + "content": "We run well-known self-supervised privacy-preservation techniques for video anomaly detection. In order to maintain a fair comparison across methods, we utilize the exact same network architectures and training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.546 + ], + "angle": 0, + "content": "Downsampling Baselines For Downsample-2x and Downsample-4x, we simply resize the input frames to a lower resolution by a factor of 2 (112 × 112) and 4 (56 × 56)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.697 + ], + "angle": 0, + "content": "Object-Detector Based Obfuscation Baselines Obfuscation techniques are based on first detecting the person, followed by removing (i.e. blackening) or blurring them. Both obfuscation techniques use MS-COCO [26] pretrainedYOLO [37] object detector to obtain bounding boxes for person object class. We utilizeYOLOv51 implementation with yolov5x as backbone. The detected bounding boxes are assigned to pixel value 0 for the Blackening-based baseline. For the Blurring-based baselines, a Gaussian filter with kernel \\( k = 13 \\) and variance \\( \\sigma = 10 \\) is utilized." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.743 + ], + "angle": 0, + "content": "SPAct [13] Baseline We utilize official implementation2. For a fair comparison with our method, we utilize the exact same utility model I3D and privacy model ResNet-50." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.752, + 0.882, + 0.768 + ], + "angle": 0, + "content": "4.4. Evaluation on Benchmark Anomaly Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.776, + 0.892, + 0.866 + ], + "angle": 0, + "content": "We compare prior privacy-preserving methods to our method on 3 well-known anomaly detection benchmark datasets. Since privacy-preservation deals with both utility (i.e. anomaly) and privacy, we show results in form of a trade-off plot as presented in Fig. 3. Compared to the prior best method [13], our method is able to remove \\(19.9\\%\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.875, + 0.72, + 0.888 + ], + "angle": 0, + "content": "\\(^{1}\\)https://github.com/ultralytics/yolov5" + }, + { + "type": "page_footnote", + "bbox": [ + 0.52, + 0.888, + 0.737, + 0.901 + ], + "angle": 0, + "content": "2https://github.com/DAVEISHAN/SPAct" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.875, + 0.737, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "13603" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.338, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.246, + 0.258, + 0.259 + ], + "angle": 0, + "content": "(a) UCF-Crime" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.089, + 0.603, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.246, + 0.525, + 0.259 + ], + "angle": 0, + "content": "(b) XD-Violence" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.09, + 0.865, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.246, + 0.79, + 0.259 + ], + "angle": 0, + "content": "(c) ShanghaiTech" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.264, + 0.895, + 0.294 + ], + "angle": 0, + "content": "Figure 3: Trade-off plots between anomaly detection benchmarks [43, 48, 28] AUC and VISPR [32] privacy attribute prediction cMAP for different privacy-preserving methods. Optimal trade-off point is top left of plot (higher AD performance, lower PA prediction ability)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.47, + 0.395 + ], + "angle": 0, + "content": "more privacy with a slightly better utility score \\((1.19\\%)\\). This strongly supports our claim that promoting temporal distinctiveness during anonymization better aligns with anomaly detection tasks. Numeric data behind Fig. 3 plots can be found in Supp. Sec. D." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.404, + 0.262, + 0.421 + ], + "angle": 0, + "content": "4.5. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.565 + ], + "angle": 0, + "content": "Figure 4 shows visual examples of the model outputs in different videos. We note that to the human eye, it is difficult to tell what is going on in each video, yet the anomaly detection model is still able to achieve high performance. Fig. 4a-4d include human subjects, yet none of their private attributes such as face and clothing are visible, and therefore cannot be used to make unfair decisions. In Fig. 4c, 4d, the background shop can still be seen, which is useful context for identifying the shoplifting and robbery anomalies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.471, + 0.688 + ], + "angle": 0, + "content": "Figure 5 is a graph of the anomaly score output from both the raw and anonymized video models for each frame of an input video. The blue box shows the ground truth anomalous frames. We see that the green anonymized anomaly scores are similar to the red raw scores while still increasing the anomaly scores on the ground truth anomaly, showing that our anonymization technique maintains utility performance. See Supp. Sec. D for more qualitative results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.468, + 0.713 + ], + "angle": 0, + "content": "4.6. Evidence for Privacy Leakage at Feature-level" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In anomaly feature representation learning, the anomaly detection algorithms do not directly work with the input videos, the videos are first passed through an action classifier to compute features. Even though the action recognition model sees the original videos, it is not certain whether the private information gets passed to the features. In order to confirm this, we create a simple fully connected network to predict VISPR private attributes, in the same fashion that we evaluate privacy for our other experiments. We stack the same VISPR image 16 times to create a video clip, then extract the clip features through \\( f_{T} \\). Our baseline uses unmodified input images passed through a pretrained Ki" + }, + { + "type": "table", + "bbox": [ + 0.499, + 0.317, + 0.895, + 0.386 + ], + "angle": 0, + "content": "
Feature Extraction ModelVISPR
Privacy cMAP (%)Privacy Reduction (%)
Kinetics400 pretrained63.150.0
SPAct [13] Anonymized55.6011.96
Our Anonymized52.3017.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.397, + 0.895, + 0.44 + ], + "angle": 0, + "content": "Table 1: Quantitative evidence of action classification model features leaking privacy. Features from each model used to predict privacy attributes. Red indicates higher privacy leakage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.894, + 0.562 + ], + "angle": 0, + "content": "netics400 [20] I3D [3] model, with the other experiments using a paired anonymizer and tuned I3D model. Detailed explanations of this process are found in Supp. Sec. C. Through experimentation, we find that the action classifier latent features do in fact leak private information, therefore this private information gets passed into the anomaly detector. Table 1 demonstrates empirical evidence of this." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.569, + 0.655, + 0.586 + ], + "angle": 0, + "content": "4.7. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.894, + 0.699 + ], + "angle": 0, + "content": "Effect of different utility losses \\(\\mathcal{L}_T\\) We study the effect of different utility losses during the anonymization process on the final privacy vs. utility anomaly detection performance in Table 2. From row-(a,b), we see a clear reduction in privacy at no cost of anomaly performance; which demonstrates the effectiveness of our proposed temporally-distinct triplet loss \\(\\mathcal{L}_D\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.894, + 0.745 + ], + "angle": 0, + "content": "To this extent, we also implement the contrary objective to our \\(\\mathcal{L}_D\\) which promotes temporal-invariance \\(\\mathcal{L}_I\\) in the utility branch of the anonymization training. We implement" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.756, + 0.882, + 0.859 + ], + "angle": 0, + "content": "
Utility Loss during Anonymization (L_T)VISPR Privacy cMAP(%)↓)UCF-Crime Anomaly AUC(%)↑)
(a)L_CE52.7173.93
(b)L_CE + L_D42.2174.81
(c)L_CE + L_I45.6469.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.868, + 0.893, + 0.898 + ], + "angle": 0, + "content": "Table 2: Ablation with different utility losses during the anonymization process. Bold indicates best trade-off." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "13604" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.443, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.19, + 0.341, + 0.204 + ], + "angle": 0, + "content": "(a) Fighting018_x264.mp4" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.205, + 0.442, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.305, + 0.349, + 0.319 + ], + "angle": 0, + "content": "(c) Shoplifting048_x264.mp4" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.09, + 0.891, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.19, + 0.791, + 0.204 + ], + "angle": 0, + "content": "(b) Shooting040_x264.mp4" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.205, + 0.891, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.305, + 0.79, + 0.319 + ], + "angle": 0, + "content": "(d) Robbery023_x264.mp4" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.324, + 0.893, + 0.352 + ], + "angle": 0, + "content": "Figure 4: Qualitative results of our anonymization training on UCF-Crime [43] dataset. For each case, the top row shows the raw video frames, and the bottom row shows the frames after being passed through the anonymizer." + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.367, + 0.773, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.637, + 0.893, + 0.666 + ], + "angle": 0, + "content": "Figure 5: Frame-level anomaly score plot from XD-Violence. Green line shows our anonymized model, red line is the raw input model, both compared to the blue ground truth line. The below visualizations show uniformly sampled frames from the video." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.691, + 0.47, + 0.768 + ], + "angle": 0, + "content": "this using well-known self-supervised works [36, 12]. From row-(c) we see that temporal-invariance objective is not well-suited for anomaly detection utility tasks and results in a significant drop of \\(6\\%\\). We provide extensive experiments with \\(\\mathcal{L}_I\\) and its explanation in Supp. Sec. D." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.774, + 0.455, + 0.789 + ], + "angle": 0, + "content": "Effect of different temporal distinctiveness objectives:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our objective of temporal distinctiveness can also be achieved through a contrastive loss. To this end we utilize the implementation of local-local temporal contrastive loss (LLTC) of [12]. It achieves \\(75.06\\%\\) AUC at \\(42.44\\%\\) cMAP (Table 3), which is closely comparable to our results using triplet loss. It is important to note that LLTC loss significantly increases the computation (GPU memory require" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.691, + 0.892, + 0.752 + ], + "angle": 0, + "content": "ment) compared to triplet loss. It requires 8 clips (4 clips \\(\\times\\) 2 augmented view) which results in 132.45G FLOPs compared to 3 clips of triplet loss requiring only 49.67G FLOPs in our experimental setup." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Relative Weightage of \\(\\mathcal{L}_D\\) Here we test the effect of changing the weight of the additional temporally-distinct triplet loss. Table 4 shows that weighting the loss at 0.1 achieves our best results of \\(32.25\\%\\) relative increase in privacy with only a \\(3.69\\%\\) reduction in utility performance. Without the enforced temporal distinctiveness, the utility model is limited by the quality of the reconstructed anonymized videos. The improper weighting of the temporal loss interferes with the classifier ability of the model," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "13605" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.44, + 0.19 + ], + "angle": 0, + "content": "
Temporal Distinctiveness Loss LDVISPR Privacy cMAP(%)↓)UCF-Crime Anomaly AUC(%)(↑)
None52.7173.93
LLTC [12]42.4475.06
Triplet42.2174.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.206, + 0.472, + 0.263 + ], + "angle": 0, + "content": "Table 3: Comparison of different loss formulations to encourage temporal distinctiveness during the anonymization process. Both achieved similar results, with the triplet loss requiring significantly less computation." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.277, + 0.439, + 0.408 + ], + "angle": 0, + "content": "
TripletVISPRUCF-Crime
TemporalPrivacyAnomaly
Loss Weight ωcMAP(%) (↓)AUC(%) (↑)
052.7173.93
0.0153.8472.53
0.142.2174.81
1.055.2670.56
10.051.6769.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.421, + 0.472, + 0.477 + ], + "angle": 0, + "content": "Table 4: Comparison of using different loss weights of the temporally-distinct triplet loss during the anonymization process. The margin hyperparameter of the temporal triplet loss for each experiment was 1. Bold indicates best trade-off." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.47, + 0.539 + ], + "angle": 0, + "content": "which can also harm the anonymization process. This suggests that action recognition loss \\(\\mathcal{L}_{CE}\\) is still an important utility task for anomaly detection performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.736 + ], + "angle": 0, + "content": "Effect of the margin in \\(\\mathcal{L}_D\\) The proposed \\(\\mathcal{L}_D\\) temporally-distinct triplet loss uses a margin hyperparameter \\(\\mu\\) to allow for adjustments to the contrastive distance in the latent feature space. The only requirement is that \\(\\mu > 0\\). The intuition here is that a larger margin enforces greater feature spacing. As Table 5 shows, we empirically found that setting \\(\\mu = 1\\) gives us the most robust results. A lower margin results in less temporally-distinct representations, which makes distinguishing between normal and anomalous features more difficult. On the other hand, increasing the \\(\\mu = 2\\) results in more difficult temporal triplet loss (i.e a very high temporal distinctiveness) which may not align well with anomaly detection task." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.748, + 0.439, + 0.849 + ], + "angle": 0, + "content": "
TripletVISPRUCF-Crime
TemporalPrivacyAnomaly
Loss Margin μcMAP(%) (↓)AUC(%) (↑)
0.549.7862.12
1.042.2174.81
2.058.8667.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.855, + 0.472, + 0.898 + ], + "angle": 0, + "content": "Table 5: Comparison of using different hyperparameter margins of the temporally-distinct triplet loss during the anonymization process. Bold indicates default setup." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.32 + ], + "angle": 0, + "content": "Effect of temporal sampling in \\(\\mathcal{L}_D\\) The proposed triplet loss (Eq. 4) forms the negative from the clip \\(\\mathbf{x}_{t'}^{(i)}\\) of a different timestamp \\(t'\\). Distance between the timestamp of negative and anchor clip \\(t - t'\\) is an important aspect to define temporal distinctiveness. We perform experiments with various distances as shown in Table 6. In our default setting, we use random distance as shown in the first row. From the second row, we can say that a smaller distance leads to a better anomaly score with a slight degradation in protecting privacy. At the same time, the third row suggests that enforcing temporal distinctiveness at higher distances leads to better privacy protection but at the cost of anomaly performance. This distance hyperparameter may be used as a tuning parameter to get different operating points of privacy vs. anomaly trade-off." + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.332, + 0.844, + 0.434 + ], + "angle": 0, + "content": "
Negative Clip DistanceVISPR Privacy cMAP(%)（↓）UCF-Crime Anomaly AUC(%)(↑)
Random42.2174.81
846.3476.12
3228.6970.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.45, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Table 6: Comparison of enforcing set clip sampling distance during the anonymization process. The margin hyperparameter of the temporal triplet loss for each experiment was 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.52, + 0.62, + 0.536 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.819 + ], + "angle": 0, + "content": "In this paper, we highlight the importance of privacy, a previously neglected aspect of video anomaly detection. We present TeD-SPAD, a framework for applying Temporal Distinctiveness to Self-supervised Privacy-preserving video Anomaly Detection. TeD-SPAD demonstrates the effectiveness of using a temporally-distinct triplet loss while anonymizing an action recognition model, as it enhances feature representation temporal distinctiveness, which complements the downstream anomaly detection model. By effectively destroying spatial private information, we remove the model's ability to use this information in its decision-making process. As a future research direction, this framework can be extended to other tasks, such as spatiotemporal anomaly detection. The anonymizing encoder-decoder may also be made more powerful with techniques using recent masked image modeling. It is our hope that this work contributes to the development of more responsible and unbiased automated anomaly detection systems." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.688, + 0.848 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the National Science Foundation (NSF) and Center for Smart Streetscapes (CS3) under NSF Cooperative Agreement No. EEC-2133516." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "13606" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Shafiq Ahmad, Pietro Morerio, and Alessio Del Bue. Person re-identification without identification via event anonymization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Shafiq Ahmad, Gianluca Scarpellini, Pietro Morerio, and Alessio Del Bue. Event-driven re-id: A new benchmark and method towards privacy-preserving person re-identification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 459–468, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[4] Chun-Fu Chen, Shaohan Hu, Zhonghao Shi, Prateek Gulati, Bill Moriarty, Marco Pistoia, Vincenzo Piuri, and Pierangela Samarati. Mass: Multi-attribute selective suppression. arXiv preprint arXiv:2210.09904, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[5] Haoyang Chen, Xue Mei, Zhiyuan Ma, Xinhong Wu, and Yachuan Wei. Spatial-temporal graph attention network for video anomaly detection. Image and Vision Computing, 131:104629, Mar. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations, June 2020. arXiv:2002.05709 [cs, stat]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.47, + 0.599 + ], + "angle": 0, + "content": "[8] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 387-395, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.47, + 0.642 + ], + "angle": 0, + "content": "[9] Yang Cong, Junsong Yuan, and Ji Liu. Sparse reconstruction cost for abnormal event detection. In CVPR 2011, pages 3449-3456, June 2011. ISSN: 1063-6919. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.47, + 0.7 + ], + "angle": 0, + "content": "[10] Ji Dai, Behrouz Saghafi, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Towards privacy-preserving recognition of human activities. In 2015 IEEE international conference on image processing (ICIP), pages 4238-4242. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[11] Ji Dai, Behrouz Saghafi, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Towards privacy-preserving recognition of human activities. In 2015 IEEE International Conference on Image Processing (ICIP), pages 4238-4242, Sept. 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[12] Ishan Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. TCLR: Temporal contrastive learning for video representation. Computer Vision and Image Understanding, 219:103406, June 2022. 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[13] Ishan Rajendrakumar Dave, Chen Chen, and Mubarak Shah. Spact: Self-supervised privacy preservation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[15] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[16] Giacomo Giorgi, Wisam Abbasi, and Andrea Saracino. Privacy-Preserving Analysis for Remote Video Anomaly Detection in Real Life Environments. Journal of Wireless Mobile Networks, Ubiquitous Computing, and Dependable Applications, 13(1):112-136, Mar. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[18] Carlos Hinojosa, Miguel Marquez, Henry Arguello, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Privhar: Recognizing human actions from privacy-preserving lens. In European Conference on Computer Vision, pages 314–332. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[19] Chao Huang, Chengliang Liu, Jie Wen, Lian Wu, Yong Xu, Qiuping Jiang, and Yaowei Wang. Weakly Supervised Video Anomaly Detection via Self-Guided Temporal Discriminative Transformer. IEEE Transactions on Cybernetics, pages 1-14, 2022. Conference Name: IEEE Transactions on Cybernetics. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[20] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The Kinetics Human Action Video Dataset, May 2017. arXiv:1705.06950 [cs]. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[21] Junho Kim, Young Min Kim, Yicheng Wu, Ramzi Zahred-dine, Weston A. Welge, Gurunandan Krishnan, Sizhuo Ma, and Jian Wang. Privacy-Preserving Visual Localization with Event Cameras, Dec. 2022. arXiv:2212.03177 [cs]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[23] B. Ravi Kiran, Dilip Mathew Thomas, and Ranjith Parakkal. An Overview of Deep Learning Based Methods for Unsupervised and Semi-Supervised Anomaly Detection in Videos. Journal of Imaging, 4(2):36, Feb. 2018. Number: 2 Publisher: Multidisciplinary Digital Publishing Institute. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[24] Ming Li, Jun Liu, Hehe Fan, Jia-Wei Liu, Jiahe Li, Mike Zheng Shou, and Jussi Keppo. STPrivacy: Spatio-Temporal Tubelet Sparsification and Anonymization for Privacy-preserving Action Recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[25] Shuo Li, Fang Liu, and Licheng Jiao. Self-Training Multi-Sequence Learning with Transformer for Weakly Supervised Video Anomaly Detection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):1395–1403, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "13607" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[27] Jixin Liu and Leilei Zhang. Indoor privacy-preserving action recognition via partially coupled convolutional neural network. In 2020 International Conference on Artificial Intelligence and Computer Engineering (ICAICE), pages 292-295. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[28] W. Liu, D. Lian W. Luo, and S. Gao. Future frame prediction for anomaly detection - a new baseline. In 2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[29] Zelun Luo, Jun-Ting Hsieh, Niranjan Balachandar, Serena Yeung, Guido Pusiol, Jay Luxenberg, Grace Li, Li-Jia Li, N Lance Downing, Arnold Milstein, and Li Fei-Fei. Computer Vision-Based Descriptive Analytics of Seniors' Daily Activities for Long-Term Health Monitoring. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[30] Hui Lv, Chuanwei Zhou, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Localizing anomalies from weakly-labeled videos. IEEE transactions on image processing, 30:4505-4515, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.471, + 0.459 + ], + "angle": 0, + "content": "[31] Saemi Moon, Myeonghyeon Kim, Zhenyue Qin, Yang Liu, and Dongwoo Kim. Anonymization for skeleton action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 15028-15036, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.47, + 0.518 + ], + "angle": 0, + "content": "[32] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Towards a visual privacy advisor: Understanding and predicting privacy risks in images. In IEEE International Conference on Computer Vision (ICCV), 2017. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.47, + 0.56 + ], + "angle": 0, + "content": "[33] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning Memory-guided Normality for Anomaly Detection, Mar. 2020. arXiv:2003.13228 [cs]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[34] Mandela Patrick, Yuki M Asano, Polina Kuznetsova, Ruth Fong, João F Henriques, Geoffrey Zweig, and Andrea Vedaldi. On compositions of transformations in contrastive self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9577-9587, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.73 + ], + "angle": 0, + "content": "[35] Oluwatoyin P. Popoola and Kejun Wang. Video-Based Abnormal Human Behavior Recognition—A Review. IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), 42(6):865-878, Nov. 2012. Conference Name: IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews). 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[36] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6964-6974, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[37] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You Only Look Once: Unified, Real-Time Object Detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, June 2016. ISSN: 1063-6919. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[38] Zhongzheng Ren, Yong Jae Lee, and Michael S Ryoo. Learning to anonymize faces for privacy preserving action" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "detection. In Proceedings of the European conference on computer vision (ECCV), pages 620-636, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.894, + 0.26 + ], + "angle": 0, + "content": "[40] Michael S Ryoo, Brandon Rothrock, Charles Fleming, and Hyun Jong Yang. Privacy-preserving human activity recognition from extreme low resolution. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.894, + 0.316 + ], + "angle": 0, + "content": "[41] Hitesh Sapkota and Qi Yu. Bayesian nonparametric submodular video partition for robust anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3212-3221, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.317, + 0.894, + 0.357 + ], + "angle": 0, + "content": "[42] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild, Dec. 2012. arXiv:1212.0402 [cs]. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.359, + 0.894, + 0.426 + ], + "angle": 0, + "content": "[43] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-World Anomaly Detection in Surveillance Videos. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6479-6488, Salt Lake City, UT, June 2018. IEEE. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.894, + 0.481 + ], + "angle": 0, + "content": "[44] Qiyue Sun and Yang Yang. Unsupervised video anomaly detection based on multi-timescale trajectory prediction. Computer Vision and Image Understanding, 227:103615, Jan. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.894, + 0.564 + ], + "angle": 0, + "content": "[45] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975-4986, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[46] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[47] Peng Wu and Jing Liu. Learning Causal Temporal Relation and Feature Discrimination for Anomaly Detection. IEEE Transactions on Image Processing, 30:3513-3527, 2021. Conference Name: IEEE Transactions on Image Processing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[48] Peng Wu, jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[49] Zhenyu Wu, Haotao Wang, Zhaowen Wang, Hailin Jin, and Zhangyang Wang. Privacy-preserving deep action recognition: An adversarial learning framework and a new dataset. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[50] Hongchun Yuan, Zhenyu Cai, Hui Zhou, Yue Wang, and Xiangzhi Chen. TransAnomaly: Video Anomaly Detection Using Video Vision Transformer. IEEE Access, 9:123977-123986, 2021. Conference Name: IEEE Access. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[51] Sharare Zehtabian, Siavash Khodadadeh, Ladislau Bölönii, and Damla Turgut. Privacy-preserving learning of human" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.52, + 0.937 + ], + "angle": 0, + "content": "13608" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "activity predictors in smart environments. In IEEE INFOCOM 2021-IEEE Conference on Computer Communications, pages 1-10. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[52] Chen Zhang, Guorong Li, Yuankai Qi, Shuhui Wang, Laiyun Qing, Qingming Huang, and Ming-Hsuan Yang. Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16271-16280, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.221, + 0.468, + 0.261 + ], + "angle": 0, + "content": "[53] Zhixiang Zhang, Thomas Cilloni, Charles Walter, and Charles Fleming. Multi-scale, class-generic, privacy-preserving video. *Electronics*, 10(10):1172, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.468, + 0.344 + ], + "angle": 0, + "content": "[54] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1237-1246, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.347, + 0.468, + 0.388 + ], + "angle": 0, + "content": "[55] Yi Zhu and Shawn Newsam. Motion-Aware Feature for Improved Video Anomaly Detection, July 2019. arXiv:1907.10211 [cs, eess]. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "13609" + } + ] +] \ No newline at end of file diff --git a/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/a12231ff-0413-4e22-8eb5-06598dd6fed6_origin.pdf b/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/a12231ff-0413-4e22-8eb5-06598dd6fed6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5d3edd632589bf0d1f525dcf5b092c821ea6c440 --- /dev/null +++ b/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/a12231ff-0413-4e22-8eb5-06598dd6fed6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:681ff71bf4f57e20b6c8489b58ae63c9255d80d217e4a710ef64c9e5240828b7 +size 1490050 diff --git a/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/full.md b/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cd240227e3e9c4bdec7ed229bf27335d239b2691 --- /dev/null +++ b/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/full.md @@ -0,0 +1,388 @@ +# TeD-SPAD: Temporal Distinctiveness for Self-supervised Privacy-preservation for video Anomaly Detection + +Joseph Fioresi, Ishan Rajendrakumar Dave, Mubarak Shah Center for Research in Computer Vision, University of Central Florida, Orlando, USA {joseph.fioresi, ishanrajendrakumar.dave}@ucf.edu, shah@crcv.ucf.edu Project Page: https://joefioresi718.github.io/TeD-SPAD_webpage/ + +# Abstract + +Video anomaly detection (VAD) without human monitoring is a complex computer vision task that can have a positive impact on society if implemented successfully. While recent advances have made significant progress in solving this task, most existing approaches overlook a critical real-world concern: privacy. With the increasing popularity of artificial intelligence technologies, it becomes crucial to implement proper AI ethics into their development. Privacy leakage in VAD allows models to pick up and amplify unnecessary biases related to people's personal information, which may lead to undesirable decision making. In this paper, we propose TeD-SPAD, a privacy-aware video anomaly detection framework that destroys visual private information in a self-supervised manner. In particular, we propose the use of a temporally-distinct triplet loss to promote temporally discriminative features, which complements current weakly-supervised VAD methods. Using TeD-SPAD, we achieve a positive trade-off between privacy protection and utility anomaly detection performance on three popular weakly supervised VAD datasets: UCF-Crime, XD-Violence, and ShanghaiTech. Our proposed anonymization model reduces private attribute prediction by $32.25\%$ while only reducing frame-level ROC AUC on the UCF-Crime anomaly detection dataset by $3.69\%$ . + +# 1. Introduction + +Machine learning-driven technologies are increasingly being adopted by society. The progress in cloud computing has enabled the deployment of even computationally intensive technologies in the public space. One such application is video anomaly detection (VAD) in autonomous video analytics. VAD is a video understanding task that aims to identify the temporal location of anomalous events occurring in long continuous videos without human supervision. An anomaly can be defined as any unusual event, such as a traffic accident, an elderly person falling, or a fire. Proper + +![](images/b290adf0c9ef88329790cc0d2ec3335fae173addede23c54d89d8c434bdfeccc.jpg) +Figure 1: Single frame from video Shoplifting052_x264.mp4 of UCF-Crime [43] showing the types of private attributes (shown in red on the left image) leaked in visual data. In the right image we show the anonymized frame where these attributes are barely visible. The graph demonstrates the ability of our self-supervised framework to mitigate privacy leakage compared to prior work SPAct [13]. At a similar anomaly detection utility performance, our method prevents $32\%$ of the visual private leakage compared to raw data. + +![](images/ee303beaf532acdbce2d63df0f84598342211684dcc221ef88c392bea0bf8d05.jpg) + +application of this technology can result in faster response times to anomalies, without the need for human resources to monitor camera feeds. + +However, public adoption of such AI technologies brings justifiable concern about safety and their decision-making capabilities. Many of these concerns center around privacy violations and accuracy. VAD is an application where visual privacy leakage and data bias are exceedingly important issues. Sending videos to remote computers or cloud services to process results in unnecessary privacy leakage for people who are not directly involved in anomalous events. Additionally, an application employing a standard RGB video will incorporate any bias found in its training set, leading to potentially unfair decisions. An illustration of privacy leakage is shown in Fig. 1. + +Recently, there have been interesting attempts to prevent + +visual privacy leakage in action recognition. Some of the approaches utilize input downsizing-based solutions [40, 10, 27] and object-detection-dependent obfuscation-based formulations [38, 53]. Wu et al. [49] proposed an adversarial training-based framework where they train an anonymization function to remove privacy-preservation. Dave et al. [13] proposed a self-supervised privacy-preserving framework that does not require privacy labels and achieves similar performance to the supervised method [49]. Since many weakly-supervised anomaly detection (WSAD) methods rely on the pretrained features of action recognition, privacy-preserving action recognition seems like a promising candidate for privacy-preserving anomaly detection. However, detecting anomalies does not align well with privacy-preserved action recognition videos. The use of short videos in action recognition encourages the use of temporally-invariant features, where the features of clips at distinct timesteps should be the same. Conversely, detecting anomalies in long, untrimmed videos requires temporally-distinct reasoning, where features of clips at distinct timesteps of a video should be different, to determine whether events in the same scene are anomalous. This is why most existing anomaly detection methods focus on refining the features of pretrained video encoders to increase their temporal separability. + +To the best of our knowledge, privacy-preservation in video anomaly detection is an unexplored area in computer vision. Building on the existing self-supervised privacy-preserving action recognition framework [13], we propose a more aligned utility branch for anomaly detection. To achieve this, we use a novel temporally-distinct triplet loss to promote temporal distinctiveness during anonymization training. Once the anonymization function is learned through our proposed anonymization framework, we apply it to the anomaly dataset, which ensures privacy leakage mitigation in privacy-sensitive anomaly detection tasks. We use these anonymized features to train the current state-of-the-art WSAD method, MGFN [8]. + +To evaluate privacy-preserving performance in anomaly detection, we adopt protocols from prior action recognition methods, where we report the utility performance on WSAD task on widely used anomaly datasets (UCF-Crime [43], XD-Violence [48], and ShanghaiTech [28]) and budget performance on privacy dataset VISPR [32]. + +Our contributions can be summarized as follows: + +- We introduce a new problem of privacy-preservation in video anomaly detection, where we identify the privacy leakage issue in existing weakly supervised anomaly detection methods. +- To address this open problem, we propose TeD-SPAD, a framework based on self-supervised privacy-preservation with a temporally-distinct triplet loss to + +make the video anonymization process more suitable for anomaly detection. + +- We propose evaluation protocols for privacy vs. anomaly trade-off, demonstrating that our proposed framework outperforms prior methods by significant margins across all anomaly detection benchmarks. On the widely used UCF-Crime dataset, our method is able to eliminate $32.25\%$ of the privacy leakage at a cost of only a $3.96\%$ reduction in frame-level AUC performance. + +# 2. Related Works + +# Privacy Preservation + +We observe that many works preserve visual privacy at capture time by using non-intrusive sensors such as thermal imaging, depth cameras, or event cameras [29, 18, 21, 1]. Other works allow for raw RGB visual information to be captured, but make an effort to protect the subject privacy in such a way that the data is still useful in a utility task. Earlier efforts aimed at dealing with visual privacy include image downsampling [11] or blocking/blurring privacy-related objects located using pretrained object detectors. Both of these obfuscation methods were shown to reduce utility results by more than they reduced privacy leakage [13, 24, 49]. + +Recent developments have yielded numerous privacy-preserving approaches to action recognition [2, 51, 31]. Wu et al. released an action dataset with privacy labels, PA-HMDB [49]. They use an adversarial learning framework to obfuscate the privacy features using supervised privacy labels. MaSS [4] uses a similar framework to Wu et al. [49], except it adapts a compound loss to flexibly preserve certain attributes instead of destroying them. STPrivacy [24] upgraded the general framework to work with a transformer anonymizing block, masking entire video tubelets unnecessary for action recognition. Following Dave et al.'s SPAcT [13], we adopt a similar self-supervised adversarial anonymization framework without the use of the privacy labels, using NT-Xent [6] contrastive loss in the budget branch to mitigate spatial privacy leakage. + +# Anomaly Detection + +With such a high volume of available video footage, it is infeasible to create sufficient labelled data to solve supervised VAD. Therefore, many works explore unsupervised methods. These generally train a reconstruction model, then either reconstruct the current frame or try to predict the next frame, signaling an anomaly when reconstruction error is high [9, 35, 23, 33, 50, 44]. Giorgi et al. [16] used an autoencoder with differential privacy, generating anomaly scores from the noisy reconstructed images. This method helped retain some level of subject privacy, but was only evaluated on image quality metrics. + +Sultani et al. [43] brought weak supervision to VAD, + +where anomalies are labelled at the video level. These authors introduce UCF-Crime, a large scale weakly supervised dataset. They propose formulating weakly supervised VAD as a multi instance learning (MIL) problem, showcasing the benefits of temporal smoothness loss & sparsity loss. With the exception of a select few works [54, 52], all following weakly supervised methods are considered anomaly feature representation learning as they improve upon MIL formulation [55, 45, 15, 25, 19, 46, 8, 41], which involves interpreting static video features extracted using an action classifier. Zhong et al. [54] propose a rearranged weakly supervised version of ShanghaiTech. Wu et al. [48] introduce XD-Violence, a large-scale weakly supervised dataset that includes audio, bringing multi-modal fusion to VAD. In this work, we choose to focus on the weakly supervised video anomaly detection setting due to its effectiveness and low annotation effort. + +Most VAD works find it useful to model temporal relations between video segments [45, 15, 30, 48, 25, 46, 55, 5]. [30] are able to exploit dynamic variations in features between consecutive segments to help localize anomalies. The authors of RTFM [45] find that anomalous segments tend to have larger feature magnitudes ( $\ell_2$ norm of extracted clip feature vector) than normal segments. They introduce a feature magnitude learning function to help identify anomalous segments with large feature magnitudes. MGFN [8] notices that varying scene attributes causes some normal videos to have larger feature magnitudes than anomalous, so they propose a feature magnitude contrastive loss to help capture these instances. In this work, we use the MGFN model to evaluate anomaly detection performance. Complementing these ideas, [47] demonstrates the effectiveness of explicitly encouraging feature discrimination. Intuitively, these observations can be aggregated by enforcing temporally-distinct feature representations. Building on the terminology introduced in [34] regarding invariant and distinctive representations, in this paper our goal is to enhance the distinctiveness of the representations within our utility branch. + +# 3. Method + +The central idea behind our proposed framework is to develop an anonymization function that can degrade privacy attributes during training without relying on privacy labels. Furthermore, this function must be able to maintain the performance of the weakly-supervised anomaly detection task. Fig. 2 displays a schematic diagram of the proposed framework. In Sec. 3.1, we provide a comprehensive discussion of the problem statement. Next, in Sec. 3.2, we describe the component of the framework and their initialization process. Sec. 3.3 outlines the anonymization function training, where we propose a temporally-distinct triplet loss to enhance the existing self-supervised privacy-preserving framework [13]. Once we learn the anonymization func + +tion, in Sec. 3.4, we train the anomaly detection model using the privacy-preserved features obtained through our anonymization function. An overview of our complete framework is outlined in Algorithm Sec. 3.5. + +# 3.1. Problem Statement + +We define our problem statement similar to privacy-preserving action recognition frameworks [13, 47], but, with a different utility task. Suppose we have a video dataset $\mathbb{D}_{\text{anomaly}} = \{X^{(i)}, Y^i\}_{i=1}^N$ , where, $X^{(i)}$ is a video-instance, $N$ is the total number of samples and $Y^{(i)} \in \{0,1\}$ is a binary label. Considering video-level anomaly detection as a utility task $T$ , and privacy attribute classification as the budget task $B$ , the aim of a privacy-preserving system is to maintain the performance of $T$ while reducing $B$ . To achieve this goal, the system learns an anonymization function $f_A$ , which modifies the original raw data. This goal of privacy-preservation could be fundamentally expressed as following criteria: + +Criterion-1: The performance of the utility task should not degrade from the original performance, i.e. loss $\mathcal{L}_T$ value of the utility target model $f_T'$ should remain almost identical before and after applying the anonymization function. + +$$ +\mathcal {L} _ {T} \left(f _ {T} ^ {\prime} \left(f _ {A} ^ {*} (X)\right), Y\right) \approx \mathcal {L} _ {T} \left(f _ {T} ^ {\prime} (X), Y\right). \tag {1} +$$ + +Criterion-2: Applying anonymization function should increase the loss $\mathcal{L}_B$ for budget $B$ task of target budget model $f_B^\prime$ . + +$$ +\mathcal {L} _ {B} \left(f _ {B} ^ {\prime} \left(f _ {A} ^ {*} (X)\right)\right) \gg \mathcal {L} _ {B} \left(f _ {B} ^ {\prime} (X)\right). \tag {2} +$$ + +Regarding weakly-supervised anomaly detection (WSAD), most of the existing methods require multi-stage training, meaning they are not end-to-end. This presents a challenge for incorporating it as a utility task in anonymization training. By contrast, privacy-preserving action recognition frameworks [13, 49] have an end-to-end utility task (i.e action recognition), making it more straightforward to include. + +Since most WSAD methods rely on pretrained video encoder features from large-scale action recognition training, we can utilize the exact same anonymization process of privacy-preserving action recognition [13], utilizing action recognition as a proxy-utility task on a proxy-utility action dataset, which we denote as $\mathbb{D}_{\text {action }}$ . The limitation of such anonymization training is that by focusing solely on optimizing short clips for action recognition, it neglects the importance of temporally-distinct features. This oversight leads to a notable decline in anomaly detection performance, which relies heavily on these temporally-distinct features, as the training progresses. To resolve this issue, we reformulate the anonymization utility branch to enforce temporal distinctiveness to better align with the anomaly detection task. Hence, our $\mathcal{L}_T$ is weighted sum of action + +![](images/a5fe680f6fbe9983d2723bcddf3dd34b42af55eadf714f8ea112691e2fe248ec.jpg) +(b) Privacy-Preserved Anomaly Detection Workflow +Figure 2: Full TeD-SPAD framework consisting of the proxy anonymization training followed by the privacy-preserved anomaly detection. (a) shows this proxy training, where UNet is used to anonymize frames in such a way that reduces mutual information between frames while maintaining utility performance. We complement the standard cross-entropy loss with our proposed temporally-distinct triplet loss, which enforces a difference in clip features at distinct timesteps. After training the anonymizer and feature extractor, (b) shows the privacy-preserved workflow, where the anomaly dataset videos are passed through the proxy-trained $f_{A}$ , $f_{T}$ , then into any WSAD algorithm. + +recognition loss $(\mathcal{L}_{CE})$ and temporally-distinct triplet loss $(\mathcal{L}_D)$ which is elaborated in Eq. 6. + +# 3.2. Anonymization Framework + +Our anonymization framework consists of 3 main components: (1) Anonymization function $(f_A)$ , which is a simple encoder-decoder model with a sigmoid activation. (2) Privacy removal model $(f_B)$ , which is an image encoder. (3) Utility model $(f_T)$ , which is a video encoder. + +Network Initialization First of all, our anonymization model is pretrained to initialize as an identity function. This pretraining involves the reconstruction of the frames from $\mathbb{D}_{\text {action }}$ using L1-reconstruction loss. + +$$ +\mathcal {L} _ {L 1} = \sum_ {c = 1} ^ {C} \sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} | x _ {c, h, w} - \hat {x} _ {c, h, w} |, \tag {3} +$$ + +where $\hat{x}$ is the output of $f_{A}$ , $x$ is an input image, and $C, H, W$ corresponds to the channel, height, and width of the input image. + +Secondly, the privacy model $f_{B}$ is initialized with the self-supervised weights of SimCLR [7] on ImageNet [14]. The video encoder $f_{T}$ is pretrained with the standard action recognition weights from Kinetics400 dataset [3]. + +# 3.3. Anonymization Training + +Anonymization training is mainly comprised of a minimax optimization of utility loss and privacy loss. + +Temporally-distinct triplet loss as Utility Weakly-supervised video anomaly detection methods leverage temporal information to help localize anomalies. [30, 19, 45, 8] show the positive effect of feature magnitude ( $\ell_2$ norm of clip feature vector) separability along the temporal dimension. In order to adopt the SPAcT anonymization framework to anomaly detection problems, we utilize a temporal distinctiveness objective in the utility branch. We design a temporally-distinct triplet loss, which increases the agreement between temporally-aligned clips of the same video and increases dissimilarity between representations of the clips which are temporally-misaligned. For anchor clip $\mathbf{x}_t^{(i)}$ , we obtain the positive clip from the exact same timestamp, but with a differently augmented version denoted as $\tilde{\mathbf{x}}_t^{(i)}$ . Whereas, the negative clip is obtained from different timestamp $\mathbf{x}_{t'}^{(i)}$ , where $t' \neq t$ . This triplet of clips is passed through the utility model $f_T$ to achieve features denoted as $\mathbf{z}_t^{(i)}$ , $\tilde{\mathbf{z}}_t^{(i)}$ , and $\mathbf{z}_{t'}^{(i)}$ . The proposed temporally-distinct triplet loss can be expressed as follows: + +$$ +\mathcal {L} _ {D} ^ {(i)} = \max \{d (\mathbf {z} _ {t} ^ {(i)}, \tilde {\mathbf {z}} _ {t} ^ {(i)}) - d (\mathbf {z} _ {t} ^ {(i)}, \mathbf {z} _ {t ^ {\prime}} ^ {(i)}) + \mu , 0 \}, \qquad (4) +$$ + +where $d(u_j,v_j) = \|\mathbf{u}_j - \mathbf{v}_j\|_2$ is Euclidean distance be + +tween two vectors $\mathbf{u}$ and $\mathbf{v}$ , and $\mu$ is the controllable margin hyperparameter to determine how far to push and pull features in the latent space. + +We utilize this loss along with the standard cross-entropy action classification loss: + +$$ +\mathcal {L} _ {C E} ^ {(i)} = - \sum_ {c = 1} ^ {N _ {C}} \mathbf {y} _ {c} ^ {(i)} \log \mathbf {p} _ {c} ^ {(i)}, \tag {5} +$$ + +where $N_{C}$ is the total number of action classes of $\mathbb{D}_{\text {action }}$ , $\mathbf{y}_c^{(i)}$ is the ground-truth label, and $\mathbf{p}_c^{(i)}$ is the prediction vector by utility model $f_T$ . + +Adding both temporal distinctiveness (Eq. 4) and action classification objective to our utility branch, our overall utility loss can be expressed as follows + +$$ +\mathcal {L} _ {T} = \mathcal {L} _ {C E} + \omega * \mathcal {L} _ {D}, \tag {6} +$$ + +where $\omega$ hyperparameter is the weight of temporally-distinct triplet loss with respect to cross-entropy loss. + +Privacy (i.e. budget) Loss $\mathcal{L}_B$ We utilize the same self-supervised privacy loss from [13], which removes the private information by minimizing the agreement between the frames of the same video. + +Minimax Optimization After reformulating the utility loss, we use the minimax optimization process similar to [13]. It is a two-step iterative process that minimizes the utility loss and at the same time increases budget loss $\mathcal{L}_B$ . At the end of this optimization, we obtain the learned anonymization function $(f_A)$ and utility video encoder $(f_T)$ . + +# 3.4. Privacy-preserving Anomaly Detection + +In order to detect anomalies within videos from $\mathbb{D}_{\text{anomaly}}$ , we utilize the current state-of-the-art technique, Magnitude-Contrastive Glance-and-Focus Network (MGFN) [8]. Similar to other recent works in anomaly detection, MGFN requires fixed features for each video from a pretrained video encoder for anomaly detection training. + +Optimizing for Anomaly Detection MGFN anomaly detection is comprised of 4 main losses: (1) a standard sigmoid cross-entropy loss $L_{sce}$ for snippet classification accuracy, (2) a temporal smoothing loss $L_{ts}$ [43] to encourage smoothness between feature representations of consecutive segments, (3) a sparsity term $L_{sp}$ [43] to discourage false positive anomalies, and (4) a novel magnitude contrastive loss $L_{mc}$ to learn scene-adaptive feature distributions across videos, all which help to train a model $f_{AD}$ . + +The training loss used in MGFN is compounded in the following equation: + +$$ +L _ {A D} = L _ {s c e} + \lambda_ {1} L _ {t s} + \lambda_ {2} L _ {s p} + \lambda_ {3} L _ {m c}, \tag {7} +$$ + +where $\lambda_1 = \lambda_2 = 1$ , and $\lambda_3 = 0.001$ . $f_{AD}$ outputs frame-level anomaly scores, which are used to calculate a final ROC AUC and AP for evaluation. + +Feature Extraction In our privacy-preserving case, we cannot use $\mathbb{D}_{\text{anomaly}}$ directly for the feature extraction from the video encoder. We first anonymize each video $(X^i)$ of the dataset through the learned anonymization function $f_A$ to get an anonymized dataset. For the feature extraction, we utilize the learned utility video encoder $f_T$ . We denote this extracted set of anonymized features as $\mathbb{F}_{\text{anomaly}} = \{f_T(f_A(X^i)) \mid \forall X^i \in \mathbb{D}_{\text{anomaly}}\}$ . + +# 3.5. Algorithm + +Let's consider the models $f_{A}, f_{T}, f_{B}, f_{AD}$ , parameterized by $\theta_{A}, \theta_{T}, \theta_{B}$ , and $f_{AD}$ , respectively. $\mathbb{D}_{\text{action}}$ is the proxy action recognition dataset and $\mathbb{D}_{\text{anomaly}}$ is the downstream anomaly detection dataset. All training steps of our framework can be put together into a sophisticated form as algorithm 1. + +# Algorithm 1: TeD-SPAD Framework + +1 Inputs: +2 Datasets: $\mathbb{D}_{\text{action}}$ , $\mathbb{D}_{\text{anomaly}}$ +3 #Epochs: max_anon_epoch, max_ad_epoch +4 Learning Rates: $\alpha_{AD},\alpha_B,\alpha_T$ +5 Hyperparameters: $\mu, \omega$ +6 Output: $\theta_{AD},\theta_{A}$ +7 Model Initialization: +8 Initialize $\theta_T$ with Kinetics400 weights [3]; +9 Initialize $\theta_B$ with SimCLR ImageNet weights [7]. +10 Initialize $\theta_{A}\gets \theta_{A} - \alpha_{A}\nabla_{\theta_{A}}(\mathcal{L}_{L1}(\theta_{A}))$ (Ref. Eq. 3) +11 Anonymization Training: +12 for $e_0\gets 1$ to max_anon_epoch do +13 Step-1 +14 $\theta_{A}\gets \theta_{A} - \alpha_{A}\nabla_{\theta_{A}}(\mathcal{L}_{T}(\theta_{A},\theta_{T}) - \omega L_{B}(\theta_{A},\theta_{B}))$ +15 Step-2 +16 $\theta_T\gets \theta_T - \alpha_T\nabla_{\theta_T}(\mathcal{L}_T(\theta_T,\theta_A))$ (Ref. Eq. 6) +17 $\theta_B\gets \theta_B - \alpha_B\nabla_{\theta_B}(\mathcal{L}_B(\theta_B,\theta_A))$ +17 end +18 Feature Extraction on $\mathbb{D}_{\text{anomaly}}$ : +19 $\mathbb{F}_{anomaly} = \{f_T(f_A(X^i))\mid \forall X^i\in \mathbb{D}_{anomaly}\}$ +20 Privacy-Preserved Anomaly Detection Training: +21 for $e_0\gets 1$ to max_ad_epoch do +22 $\theta_{AD}\gets \theta_{AD} - \alpha_{AD}\nabla_{\theta_{AD}}(L_{AD}(\theta_{AD},\mathbb{F}_{anomaly}))$ +23 end + +# 4. Experiments + +# 4.1. Datasets + +UCF-Crime [43] is the first large-scale weakly supervised video anomaly detection dataset. It contains 1,900 videos totaling 128 hours of untrimmed CCTV surveillance footage from a variety of different scenes. The videos contain 13 crime-based anomalies such as Arrest, Fighting, and + +Shoplifting in real world scenes. + +XD-Violence [48] is currently the largest weakly supervised video anomaly detection dataset, totaling 217 hours of untrimmed video. All of its anomaly categories are related to violent activities. The videos are gathered from various types of cameras, movies, and games, resulting in a unique blend of scenes for increased difficulty. + +ShanghaiTech [28] is a medium-scale anomaly detection dataset containing videos covering 13 different scenes with various types of anomalies. While it was published as an unsupervised anomaly detection dataset, we use the weakly supervised rearrangement proposed by [54]. + +VISPR [32] is an image dataset labelled with 68 privacy-related attributes, including gender, hair color, clothes, etc. It provides a multi-class classification problem for us to evaluate privacy-preservation on. The split we use for evaluation along with training details can found in Supp. Sec. B. + +UCF101 [42] is a very common dataset in action recognition, and its relative simplicity makes it practical for learning an anonymization model on, demonstrated in [13]. In this work, split-1 of UCF101 is used as $\mathbb{D}_{\text {action }}$ . + +# 4.2. Implementation Details + +Network Architecture Details $f_{A}$ is a UNet [39] model that transforms raw input frames into anonymized frames. I3D [3] is used for $f_{T}$ , to first learn anonymized action classification, then to extract anonymized features. A ResNet-50 [17] is our $f_{B}$ model during the anonymization training. $f_{AD}$ is MGFN [8], which consists of a shortcut convolution, a self-attentional convolution, and a feed-forward network. + +Training Process Details First we perform anonymization training for 80 epochs. Adam [22] optimizer is used for all models with a learning rate is 1e-4 with a batch size of 8. Loss weight $\omega = 0.1$ and margin $\mu = 1$ in a default setting. We train the MGFN [8] model using default hyperparameters. $f_{B}$ evaluation uses a batch size of 32 and a base learning rate of 1e-3, which follows a linear warmup and a step-based scheduler that drops by a factor of 1/5 upon loss stagnation. + +Input Details For all experiments, we crop each image to a scale of 0.8, then resize to input resolution $224 \times 224$ . Clips consist of 16 frames and are sampled from a random start with a skip-rate of 2. For anonymization training we utilize standard augmentations like random erase, random crop, horizontal flip and random color jitter. To maintain temporal consistency, augmentations are applied equally to every frame within each clip. + +Feature Extraction Given a raw input video $\mathbf{X}^{\mathrm{i}}\in$ $\mathbb{D}_{anomaly}$ , we first extract $S$ clips, where $S$ is the amount of non-overlapping 16-frame clips in $\mathbf{X}$ . We sequentially pass each clip $S^{(j)}$ first through our anonymizer $f_{A}$ , then our feature extractor $f_{T}$ Extracted features $f_{T}(f_{A}(\mathbf{X}^{i}))$ + +have the shape $S \times C$ , where $C$ is the dimensionality of the feature vector. Specifically, features are extracted following the average pooling after the mix_5c layer of I3D and have a dimensionality of 2048. + +Evaluation Protocol and Performance Metrics To evaluate the learned anonymization function $f_{A}$ , we follow standard protocols of cross-dataset evaluation [13, 49]. In this protocol, testing videos of $\mathbb{D}_{\text{anomaly}}$ are anonymized by $f_{A}$ , and frame-level predictions are obtained through the $f_{T}$ and $f_{AD}$ . The calculated ROC AUC is used to evaluate performance on UCF-Crime and ShanghaiTech, and AP is used for XD-Violence. A higher AUC/AP is considered a more accurate anomaly localization. In order to evaluate the privacy leakage, the learned $f_{A}$ is utilized to anonymize the privacy dataset $\mathbb{D}_{\text{privacy}}$ to train and evaluate a target privacy model $f_{B}'$ . Privacy leakage is measured in terms of the performance of the target $f_{B}'$ on the test set of $\mathbb{D}_{\text{privacy}}$ . Since the privacy dataset is multi-label, the privacy leakage is measured in terms of mean average precision averaged across classes (cMAP). + +More implementation details are in Supp. Sec. C. + +# 4.3. Privacy-Preserving Baselines + +We run well-known self-supervised privacy-preservation techniques for video anomaly detection. In order to maintain a fair comparison across methods, we utilize the exact same network architectures and training process. + +Downsampling Baselines For Downsample-2x and Downsample-4x, we simply resize the input frames to a lower resolution by a factor of 2 (112 × 112) and 4 (56 × 56). + +Object-Detector Based Obfuscation Baselines Obfuscation techniques are based on first detecting the person, followed by removing (i.e. blackening) or blurring them. Both obfuscation techniques use MS-COCO [26] pretrainedYOLO [37] object detector to obtain bounding boxes for person object class. We utilizeYOLOv51 implementation with yolov5x as backbone. The detected bounding boxes are assigned to pixel value 0 for the Blackening-based baseline. For the Blurring-based baselines, a Gaussian filter with kernel $k = 13$ and variance $\sigma = 10$ is utilized. + +SPAct [13] Baseline We utilize official implementation2. For a fair comparison with our method, we utilize the exact same utility model I3D and privacy model ResNet-50. + +# 4.4. Evaluation on Benchmark Anomaly Datasets + +We compare prior privacy-preserving methods to our method on 3 well-known anomaly detection benchmark datasets. Since privacy-preservation deals with both utility (i.e. anomaly) and privacy, we show results in form of a trade-off plot as presented in Fig. 3. Compared to the prior best method [13], our method is able to remove $19.9\%$ + +![](images/504be69d5bbc3416948eaa19b04a10cf9f53fccd18922b5a4f6d235e1df6b522.jpg) +(a) UCF-Crime + +![](images/5d36aa73841e905a493ee5cbdd52644fa465295083822de88b62d6030bc0f157.jpg) +(b) XD-Violence +Figure 3: Trade-off plots between anomaly detection benchmarks [43, 48, 28] AUC and VISPR [32] privacy attribute prediction cMAP for different privacy-preserving methods. Optimal trade-off point is top left of plot (higher AD performance, lower PA prediction ability). + +![](images/5ee4ec2b383c638fdfda538a350b540ca47a49ae6b6b3853fd024765bda878aa.jpg) +(c) ShanghaiTech + +more privacy with a slightly better utility score $(1.19\%)$ . This strongly supports our claim that promoting temporal distinctiveness during anonymization better aligns with anomaly detection tasks. Numeric data behind Fig. 3 plots can be found in Supp. Sec. D. + +# 4.5. Qualitative Results + +Figure 4 shows visual examples of the model outputs in different videos. We note that to the human eye, it is difficult to tell what is going on in each video, yet the anomaly detection model is still able to achieve high performance. Fig. 4a-4d include human subjects, yet none of their private attributes such as face and clothing are visible, and therefore cannot be used to make unfair decisions. In Fig. 4c, 4d, the background shop can still be seen, which is useful context for identifying the shoplifting and robbery anomalies. + +Figure 5 is a graph of the anomaly score output from both the raw and anonymized video models for each frame of an input video. The blue box shows the ground truth anomalous frames. We see that the green anonymized anomaly scores are similar to the red raw scores while still increasing the anomaly scores on the ground truth anomaly, showing that our anonymization technique maintains utility performance. See Supp. Sec. D for more qualitative results. + +# 4.6. Evidence for Privacy Leakage at Feature-level + +In anomaly feature representation learning, the anomaly detection algorithms do not directly work with the input videos, the videos are first passed through an action classifier to compute features. Even though the action recognition model sees the original videos, it is not certain whether the private information gets passed to the features. In order to confirm this, we create a simple fully connected network to predict VISPR private attributes, in the same fashion that we evaluate privacy for our other experiments. We stack the same VISPR image 16 times to create a video clip, then extract the clip features through $f_{T}$ . Our baseline uses unmodified input images passed through a pretrained Ki + +
Feature Extraction ModelVISPR
Privacy cMAP (%)Privacy Reduction (%)
Kinetics400 pretrained63.150.0
SPAct [13] Anonymized55.6011.96
Our Anonymized52.3017.18
+ +netics400 [20] I3D [3] model, with the other experiments using a paired anonymizer and tuned I3D model. Detailed explanations of this process are found in Supp. Sec. C. Through experimentation, we find that the action classifier latent features do in fact leak private information, therefore this private information gets passed into the anomaly detector. Table 1 demonstrates empirical evidence of this. + +# 4.7. Ablation Study + +Effect of different utility losses $\mathcal{L}_T$ We study the effect of different utility losses during the anonymization process on the final privacy vs. utility anomaly detection performance in Table 2. From row-(a,b), we see a clear reduction in privacy at no cost of anomaly performance; which demonstrates the effectiveness of our proposed temporally-distinct triplet loss $\mathcal{L}_D$ . + +To this extent, we also implement the contrary objective to our $\mathcal{L}_D$ which promotes temporal-invariance $\mathcal{L}_I$ in the utility branch of the anonymization training. We implement + +Table 1: Quantitative evidence of action classification model features leaking privacy. Features from each model used to predict privacy attributes. Red indicates higher privacy leakage. + +
Utility Loss during Anonymization (L_T)VISPR Privacy cMAP(%)↓)UCF-Crime Anomaly AUC(%)↑)
(a)L_CE52.7173.93
(b)L_CE + L_D42.2174.81
(c)L_CE + L_I45.6469.52
+ +Table 2: Ablation with different utility losses during the anonymization process. Bold indicates best trade-off. + +![](images/a8321a9e33e2b12f8faa319d453c7a10aface53ba71b28506d2eb99213154e55.jpg) + +![](images/d6eebd3155d2efefa41ec86e1617a22c31c46130b303846c9024b2581ae7444c.jpg) +(a) Fighting018_x264.mp4 +(c) Shoplifting048_x264.mp4 + +![](images/250498cf4e879a6d3604d71777892b1ec7f3d003af4bce5fde5a956505898d71.jpg) + +![](images/cfce969b4b19fb4c33655c199374dc0bd156dd5a8d809728239125e6f67c7093.jpg) +(b) Shooting040_x264.mp4 +(d) Robbery023_x264.mp4 + +![](images/5b45179a14f44024d90fe51c43c4228acab1da763fe6448c342fa8d13869d5be.jpg) +Figure 4: Qualitative results of our anonymization training on UCF-Crime [43] dataset. For each case, the top row shows the raw video frames, and the bottom row shows the frames after being passed through the anonymizer. +Figure 5: Frame-level anomaly score plot from XD-Violence. Green line shows our anonymized model, red line is the raw input model, both compared to the blue ground truth line. The below visualizations show uniformly sampled frames from the video. + +this using well-known self-supervised works [36, 12]. From row-(c) we see that temporal-invariance objective is not well-suited for anomaly detection utility tasks and results in a significant drop of $6\%$ . We provide extensive experiments with $\mathcal{L}_I$ and its explanation in Supp. Sec. D. + +# Effect of different temporal distinctiveness objectives: + +Our objective of temporal distinctiveness can also be achieved through a contrastive loss. To this end we utilize the implementation of local-local temporal contrastive loss (LLTC) of [12]. It achieves $75.06\%$ AUC at $42.44\%$ cMAP (Table 3), which is closely comparable to our results using triplet loss. It is important to note that LLTC loss significantly increases the computation (GPU memory require + +ment) compared to triplet loss. It requires 8 clips (4 clips $\times$ 2 augmented view) which results in 132.45G FLOPs compared to 3 clips of triplet loss requiring only 49.67G FLOPs in our experimental setup. + +Relative Weightage of $\mathcal{L}_D$ Here we test the effect of changing the weight of the additional temporally-distinct triplet loss. Table 4 shows that weighting the loss at 0.1 achieves our best results of $32.25\%$ relative increase in privacy with only a $3.69\%$ reduction in utility performance. Without the enforced temporal distinctiveness, the utility model is limited by the quality of the reconstructed anonymized videos. The improper weighting of the temporal loss interferes with the classifier ability of the model, + +
Temporal Distinctiveness Loss LDVISPR Privacy cMAP(%)↓)UCF-Crime Anomaly AUC(%)(↑)
None52.7173.93
LLTC [12]42.4475.06
Triplet42.2174.81
+ +Table 3: Comparison of different loss formulations to encourage temporal distinctiveness during the anonymization process. Both achieved similar results, with the triplet loss requiring significantly less computation. + +
TripletVISPRUCF-Crime
TemporalPrivacyAnomaly
Loss Weight ωcMAP(%) (↓)AUC(%) (↑)
052.7173.93
0.0153.8472.53
0.142.2174.81
1.055.2670.56
10.051.6769.27
+ +which can also harm the anonymization process. This suggests that action recognition loss $\mathcal{L}_{CE}$ is still an important utility task for anomaly detection performance. + +Effect of the margin in $\mathcal{L}_D$ The proposed $\mathcal{L}_D$ temporally-distinct triplet loss uses a margin hyperparameter $\mu$ to allow for adjustments to the contrastive distance in the latent feature space. The only requirement is that $\mu > 0$ . The intuition here is that a larger margin enforces greater feature spacing. As Table 5 shows, we empirically found that setting $\mu = 1$ gives us the most robust results. A lower margin results in less temporally-distinct representations, which makes distinguishing between normal and anomalous features more difficult. On the other hand, increasing the $\mu = 2$ results in more difficult temporal triplet loss (i.e a very high temporal distinctiveness) which may not align well with anomaly detection task. + +Table 4: Comparison of using different loss weights of the temporally-distinct triplet loss during the anonymization process. The margin hyperparameter of the temporal triplet loss for each experiment was 1. Bold indicates best trade-off. + +
TripletVISPRUCF-Crime
TemporalPrivacyAnomaly
Loss Margin μcMAP(%) (↓)AUC(%) (↑)
0.549.7862.12
1.042.2174.81
2.058.8667.30
+ +Effect of temporal sampling in $\mathcal{L}_D$ The proposed triplet loss (Eq. 4) forms the negative from the clip $\mathbf{x}_{t'}^{(i)}$ of a different timestamp $t'$ . Distance between the timestamp of negative and anchor clip $t - t'$ is an important aspect to define temporal distinctiveness. We perform experiments with various distances as shown in Table 6. In our default setting, we use random distance as shown in the first row. From the second row, we can say that a smaller distance leads to a better anomaly score with a slight degradation in protecting privacy. At the same time, the third row suggests that enforcing temporal distinctiveness at higher distances leads to better privacy protection but at the cost of anomaly performance. This distance hyperparameter may be used as a tuning parameter to get different operating points of privacy vs. anomaly trade-off. + +Table 5: Comparison of using different hyperparameter margins of the temporally-distinct triplet loss during the anonymization process. Bold indicates default setup. + +
Negative Clip DistanceVISPR Privacy cMAP(%)（↓）UCF-Crime Anomaly AUC(%)(↑)
Random42.2174.81
846.3476.12
3228.6970.97
+ +Table 6: Comparison of enforcing set clip sampling distance during the anonymization process. The margin hyperparameter of the temporal triplet loss for each experiment was 1. + +# 5. Conclusion + +In this paper, we highlight the importance of privacy, a previously neglected aspect of video anomaly detection. We present TeD-SPAD, a framework for applying Temporal Distinctiveness to Self-supervised Privacy-preserving video Anomaly Detection. TeD-SPAD demonstrates the effectiveness of using a temporally-distinct triplet loss while anonymizing an action recognition model, as it enhances feature representation temporal distinctiveness, which complements the downstream anomaly detection model. By effectively destroying spatial private information, we remove the model's ability to use this information in its decision-making process. As a future research direction, this framework can be extended to other tasks, such as spatiotemporal anomaly detection. The anonymizing encoder-decoder may also be made more powerful with techniques using recent masked image modeling. It is our hope that this work contributes to the development of more responsible and unbiased automated anomaly detection systems. + +# 6. Acknowledgements + +This work was supported in part by the National Science Foundation (NSF) and Center for Smart Streetscapes (CS3) under NSF Cooperative Agreement No. EEC-2133516. + +# References + +[1] Shafiq Ahmad, Pietro Morerio, and Alessio Del Bue. Person re-identification without identification via event anonymization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 2 +[2] Shafiq Ahmad, Gianluca Scarpellini, Pietro Morerio, and Alessio Del Bue. Event-driven re-id: A new benchmark and method towards privacy-preserving person re-identification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 459–468, 2022. 2 +[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 4, 5, 6, 7 +[4] Chun-Fu Chen, Shaohan Hu, Zhonghao Shi, Prateek Gulati, Bill Moriarty, Marco Pistoia, Vincenzo Piuri, and Pierangela Samarati. Mass: Multi-attribute selective suppression. arXiv preprint arXiv:2210.09904, 2022. 2 +[5] Haoyang Chen, Xue Mei, Zhiyuan Ma, Xinhong Wu, and Yachuan Wei. Spatial-temporal graph attention network for video anomaly detection. Image and Vision Computing, 131:104629, Mar. 2023. 3 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations, June 2020. arXiv:2002.05709 [cs, stat]. 2 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 4, 5 +[8] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 387-395, 2023. 2, 3, 4, 5, 6 +[9] Yang Cong, Junsong Yuan, and Ji Liu. Sparse reconstruction cost for abnormal event detection. In CVPR 2011, pages 3449-3456, June 2011. ISSN: 1063-6919. 2 +[10] Ji Dai, Behrouz Saghafi, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Towards privacy-preserving recognition of human activities. In 2015 IEEE international conference on image processing (ICIP), pages 4238-4242. IEEE, 2015. 2 +[11] Ji Dai, Behrouz Saghafi, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Towards privacy-preserving recognition of human activities. In 2015 IEEE International Conference on Image Processing (ICIP), pages 4238-4242, Sept. 2015. 2 +[12] Ishan Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. TCLR: Temporal contrastive learning for video representation. Computer Vision and Image Understanding, 219:103406, June 2022. 8, 9 +[13] Ishan Rajendrakumar Dave, Chen Chen, and Mubarak Shah. Spact: Self-supervised privacy preservation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 5, 6, 7 +[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image + +database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4 +[15] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition, 2021. 3 +[16] Giacomo Giorgi, Wisam Abbasi, and Andrea Saracino. Privacy-Preserving Analysis for Remote Video Anomaly Detection in Real Life Environments. Journal of Wireless Mobile Networks, Ubiquitous Computing, and Dependable Applications, 13(1):112-136, Mar. 2022. 2 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[18] Carlos Hinojosa, Miguel Marquez, Henry Arguello, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Privhar: Recognizing human actions from privacy-preserving lens. In European Conference on Computer Vision, pages 314–332. Springer, 2022. 2 +[19] Chao Huang, Chengliang Liu, Jie Wen, Lian Wu, Yong Xu, Qiuping Jiang, and Yaowei Wang. Weakly Supervised Video Anomaly Detection via Self-Guided Temporal Discriminative Transformer. IEEE Transactions on Cybernetics, pages 1-14, 2022. Conference Name: IEEE Transactions on Cybernetics. 3, 4 +[20] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The Kinetics Human Action Video Dataset, May 2017. arXiv:1705.06950 [cs]. 7 +[21] Junho Kim, Young Min Kim, Yicheng Wu, Ramzi Zahred-dine, Weston A. Welge, Gurunandan Krishnan, Sizhuo Ma, and Jian Wang. Privacy-Preserving Visual Localization with Event Cameras, Dec. 2022. arXiv:2212.03177 [cs]. 2 +[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[23] B. Ravi Kiran, Dilip Mathew Thomas, and Ranjith Parakkal. An Overview of Deep Learning Based Methods for Unsupervised and Semi-Supervised Anomaly Detection in Videos. Journal of Imaging, 4(2):36, Feb. 2018. Number: 2 Publisher: Multidisciplinary Digital Publishing Institute. 2 +[24] Ming Li, Jun Liu, Hehe Fan, Jia-Wei Liu, Jiahe Li, Mike Zheng Shou, and Jussi Keppo. STPrivacy: Spatio-Temporal Tubelet Sparsification and Anonymization for Privacy-preserving Action Recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 2 +[25] Shuo Li, Fang Liu, and Licheng Jiao. Self-Training Multi-Sequence Learning with Transformer for Weakly Supervised Video Anomaly Detection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):1395–1403, June 2022. 3 +[26] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and + +C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 6 +[27] Jixin Liu and Leilei Zhang. Indoor privacy-preserving action recognition via partially coupled convolutional neural network. In 2020 International Conference on Artificial Intelligence and Computer Engineering (ICAICE), pages 292-295. IEEE, 2020. 2 +[28] W. Liu, D. Lian W. Luo, and S. Gao. Future frame prediction for anomaly detection - a new baseline. In 2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 6, 7 +[29] Zelun Luo, Jun-Ting Hsieh, Niranjan Balachandar, Serena Yeung, Guido Pusiol, Jay Luxenberg, Grace Li, Li-Jia Li, N Lance Downing, Arnold Milstein, and Li Fei-Fei. Computer Vision-Based Descriptive Analytics of Seniors' Daily Activities for Long-Term Health Monitoring. 2 +[30] Hui Lv, Chuanwei Zhou, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Localizing anomalies from weakly-labeled videos. IEEE transactions on image processing, 30:4505-4515, 2021. 3, 4 +[31] Saemi Moon, Myeonghyeon Kim, Zhenyue Qin, Yang Liu, and Dongwoo Kim. Anonymization for skeleton action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 15028-15036, 2023. 2 +[32] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Towards a visual privacy advisor: Understanding and predicting privacy risks in images. In IEEE International Conference on Computer Vision (ICCV), 2017. 2, 6, 7 +[33] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning Memory-guided Normality for Anomaly Detection, Mar. 2020. arXiv:2003.13228 [cs]. 2 +[34] Mandela Patrick, Yuki M Asano, Polina Kuznetsova, Ruth Fong, João F Henriques, Geoffrey Zweig, and Andrea Vedaldi. On compositions of transformations in contrastive self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9577-9587, 2021. 3 +[35] Oluwatoyin P. Popoola and Kejun Wang. Video-Based Abnormal Human Behavior Recognition—A Review. IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), 42(6):865-878, Nov. 2012. Conference Name: IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews). 2 +[36] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6964-6974, 2021. 8 +[37] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You Only Look Once: Unified, Real-Time Object Detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, June 2016. ISSN: 1063-6919. 6 +[38] Zhongzheng Ren, Yong Jae Lee, and Michael S Ryoo. Learning to anonymize faces for privacy preserving action + +detection. In Proceedings of the European conference on computer vision (ECCV), pages 620-636, 2018. 2 +[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 6 +[40] Michael S Ryoo, Brandon Rothrock, Charles Fleming, and Hyun Jong Yang. Privacy-preserving human activity recognition from extreme low resolution. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 2 +[41] Hitesh Sapkota and Qi Yu. Bayesian nonparametric submodular video partition for robust anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3212-3221, 2022. 3 +[42] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild, Dec. 2012. arXiv:1212.0402 [cs]. 6 +[43] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-World Anomaly Detection in Surveillance Videos. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6479-6488, Salt Lake City, UT, June 2018. IEEE. 1, 2, 5, 7, 8 +[44] Qiyue Sun and Yang Yang. Unsupervised video anomaly detection based on multi-timescale trajectory prediction. Computer Vision and Image Understanding, 227:103615, Jan. 2023. 2 +[45] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975-4986, 2021. 3, 4 +[46] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In ECCV, 2022. 3 +[47] Peng Wu and Jing Liu. Learning Causal Temporal Relation and Feature Discrimination for Anomaly Detection. IEEE Transactions on Image Processing, 30:3513-3527, 2021. Conference Name: IEEE Transactions on Image Processing. 3 +[48] Peng Wu, jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 6, 7 +[49] Zhenyu Wu, Haotao Wang, Zhaowen Wang, Hailin Jin, and Zhangyang Wang. Privacy-preserving deep action recognition: An adversarial learning framework and a new dataset. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 2, 3, 6 +[50] Hongchun Yuan, Zhenyu Cai, Hui Zhou, Yue Wang, and Xiangzhi Chen. TransAnomaly: Video Anomaly Detection Using Video Vision Transformer. IEEE Access, 9:123977-123986, 2021. Conference Name: IEEE Access. 2 +[51] Sharare Zehtabian, Siavash Khodadadeh, Ladislau Bölönii, and Damla Turgut. Privacy-preserving learning of human + +activity predictors in smart environments. In IEEE INFOCOM 2021-IEEE Conference on Computer Communications, pages 1-10. IEEE, 2021. 2 +[52] Chen Zhang, Guorong Li, Yuankai Qi, Shuhui Wang, Laiyun Qing, Qingming Huang, and Ming-Hsuan Yang. Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16271-16280, 2023. 3 +[53] Zhixiang Zhang, Thomas Cilloni, Charles Walter, and Charles Fleming. Multi-scale, class-generic, privacy-preserving video. *Electronics*, 10(10):1172, 2021. 2 +[54] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1237-1246, 2019. 3, 6 +[55] Yi Zhu and Shawn Newsam. Motion-Aware Feature for Improved Video Anomaly Detection, July 2019. arXiv:1907.10211 [cs, eess]. 3 \ No newline at end of file diff --git a/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/images.zip b/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..902ed0776c9418bc553f7ce0a308106cc8576ce5 --- /dev/null +++ b/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90551d37e6262228f59298e45ac06309618ed9dc659c99e6c9383d32875709f4 +size 487827 diff --git a/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/layout.json b/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..75a912f35224a8580f755ddf3bea5587a2e16573 --- /dev/null +++ b/2023/TeD-SPAD_ Temporal Distinctiveness for Self-Supervised Privacy-Preservation for Video Anomaly Detection/layout.json @@ -0,0 +1,11488 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 538, + 140 + ], + "type": "text", + "content": "TeD-SPAD: Temporal Distinctiveness for Self-supervised Privacy-preservation for video Anomaly Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 161, + 504, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 161, + 504, + 220 + ], + "spans": [ + { + "bbox": [ + 88, + 161, + 504, + 220 + ], + "type": "text", + "content": "Joseph Fioresi, Ishan Rajendrakumar Dave, Mubarak Shah Center for Research in Computer Vision, University of Central Florida, Orlando, USA {joseph.fioresi, ishanrajendrakumar.dave}@ucf.edu, shah@crcv.ucf.edu Project Page: https://joefioresi718.github.io/TeD-SPAD_webpage/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 270, + 289, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 289, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 289, + 558 + ], + "type": "text", + "content": "Video anomaly detection (VAD) without human monitoring is a complex computer vision task that can have a positive impact on society if implemented successfully. While recent advances have made significant progress in solving this task, most existing approaches overlook a critical real-world concern: privacy. With the increasing popularity of artificial intelligence technologies, it becomes crucial to implement proper AI ethics into their development. Privacy leakage in VAD allows models to pick up and amplify unnecessary biases related to people's personal information, which may lead to undesirable decision making. In this paper, we propose TeD-SPAD, a privacy-aware video anomaly detection framework that destroys visual private information in a self-supervised manner. In particular, we propose the use of a temporally-distinct triplet loss to promote temporally discriminative features, which complements current weakly-supervised VAD methods. Using TeD-SPAD, we achieve a positive trade-off between privacy protection and utility anomaly detection performance on three popular weakly supervised VAD datasets: UCF-Crime, XD-Violence, and ShanghaiTech. Our proposed anonymization model reduces private attribute prediction by " + }, + { + "bbox": [ + 47, + 270, + 289, + 558 + ], + "type": "inline_equation", + "content": "32.25\\%" + }, + { + "bbox": [ + 47, + 270, + 289, + 558 + ], + "type": "text", + "content": " while only reducing frame-level ROC AUC on the UCF-Crime anomaly detection dataset by " + }, + { + "bbox": [ + 47, + 270, + 289, + 558 + ], + "type": "inline_equation", + "content": "3.69\\%" + }, + { + "bbox": [ + 47, + 270, + 289, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 128, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": "Machine learning-driven technologies are increasingly being adopted by society. The progress in cloud computing has enabled the deployment of even computationally intensive technologies in the public space. One such application is video anomaly detection (VAD) in autonomous video analytics. VAD is a video understanding task that aims to identify the temporal location of anomalous events occurring in long continuous videos without human supervision. An anomaly can be defined as any unusual event, such as a traffic accident, an elderly person falling, or a fire. Proper" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 243, + 446, + 392 + ], + "blocks": [ + { + "bbox": [ + 307, + 243, + 446, + 392 + ], + "lines": [ + { + "bbox": [ + 307, + 243, + 446, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 446, + 392 + ], + "type": "image", + "image_path": "b290adf0c9ef88329790cc0d2ec3335fae173addede23c54d89d8c434bdfeccc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 399, + 547, + 498 + ], + "lines": [ + { + "bbox": [ + 305, + 399, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 547, + 498 + ], + "type": "text", + "content": "Figure 1: Single frame from video Shoplifting052_x264.mp4 of UCF-Crime [43] showing the types of private attributes (shown in red on the left image) leaked in visual data. In the right image we show the anonymized frame where these attributes are barely visible. The graph demonstrates the ability of our self-supervised framework to mitigate privacy leakage compared to prior work SPAct [13]. At a similar anomaly detection utility performance, our method prevents " + }, + { + "bbox": [ + 305, + 399, + 547, + 498 + ], + "type": "inline_equation", + "content": "32\\%" + }, + { + "bbox": [ + 305, + 399, + 547, + 498 + ], + "type": "text", + "content": " of the visual private leakage compared to raw data." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 446, + 243, + 545, + 391 + ], + "blocks": [ + { + "bbox": [ + 446, + 243, + 545, + 391 + ], + "lines": [ + { + "bbox": [ + 446, + 243, + 545, + 391 + ], + "spans": [ + { + "bbox": [ + 446, + 243, + 545, + 391 + ], + "type": "image", + "image_path": "ee303beaf532acdbce2d63df0f84598342211684dcc221ef88c392bea0bf8d05.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 555 + ], + "type": "text", + "content": "application of this technology can result in faster response times to anomalies, without the need for human resources to monitor camera feeds." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 557, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 700 + ], + "type": "text", + "content": "However, public adoption of such AI technologies brings justifiable concern about safety and their decision-making capabilities. Many of these concerns center around privacy violations and accuracy. VAD is an application where visual privacy leakage and data bias are exceedingly important issues. Sending videos to remote computers or cloud services to process results in unnecessary privacy leakage for people who are not directly involved in anomalous events. Additionally, an application employing a standard RGB video will incorporate any bias found in its training set, leading to potentially unfair decisions. An illustration of privacy leakage is shown in Fig. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Recently, there have been interesting attempts to prevent" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "13598" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 371 + ], + "type": "text", + "content": "visual privacy leakage in action recognition. Some of the approaches utilize input downsizing-based solutions [40, 10, 27] and object-detection-dependent obfuscation-based formulations [38, 53]. Wu et al. [49] proposed an adversarial training-based framework where they train an anonymization function to remove privacy-preservation. Dave et al. [13] proposed a self-supervised privacy-preserving framework that does not require privacy labels and achieves similar performance to the supervised method [49]. Since many weakly-supervised anomaly detection (WSAD) methods rely on the pretrained features of action recognition, privacy-preserving action recognition seems like a promising candidate for privacy-preserving anomaly detection. However, detecting anomalies does not align well with privacy-preserved action recognition videos. The use of short videos in action recognition encourages the use of temporally-invariant features, where the features of clips at distinct timesteps should be the same. Conversely, detecting anomalies in long, untrimmed videos requires temporally-distinct reasoning, where features of clips at distinct timesteps of a video should be different, to determine whether events in the same scene are anomalous. This is why most existing anomaly detection methods focus on refining the features of pretrained video encoders to increase their temporal separability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 371, + 289, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 289, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 289, + 527 + ], + "type": "text", + "content": "To the best of our knowledge, privacy-preservation in video anomaly detection is an unexplored area in computer vision. Building on the existing self-supervised privacy-preserving action recognition framework [13], we propose a more aligned utility branch for anomaly detection. To achieve this, we use a novel temporally-distinct triplet loss to promote temporal distinctiveness during anonymization training. Once the anonymization function is learned through our proposed anonymization framework, we apply it to the anomaly dataset, which ensures privacy leakage mitigation in privacy-sensitive anomaly detection tasks. We use these anonymized features to train the current state-of-the-art WSAD method, MGFN [8]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 527, + 289, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 289, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 289, + 598 + ], + "type": "text", + "content": "To evaluate privacy-preserving performance in anomaly detection, we adopt protocols from prior action recognition methods, where we report the utility performance on WSAD task on widely used anomaly datasets (UCF-Crime [43], XD-Violence [48], and ShanghaiTech [28]) and budget performance on privacy dataset VISPR [32]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 599, + 259, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 599, + 259, + 610 + ], + "spans": [ + { + "bbox": [ + 59, + 599, + 259, + 610 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 621, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 621, + 287, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 621, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 58, + 621, + 287, + 669 + ], + "type": "text", + "content": "- We introduce a new problem of privacy-preservation in video anomaly detection, where we identify the privacy leakage issue in existing weakly supervised anomaly detection methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 288, + 714 + ], + "type": "text", + "content": "- To address this open problem, we propose TeD-SPAD, a framework based on self-supervised privacy-preservation with a temporally-distinct triplet loss to" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 96 + ], + "type": "text", + "content": "make the video anonymization process more suitable for anomaly detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 106, + 547, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 547, + 201 + ], + "type": "text", + "content": "- We propose evaluation protocols for privacy vs. anomaly trade-off, demonstrating that our proposed framework outperforms prior methods by significant margins across all anomaly detection benchmarks. On the widely used UCF-Crime dataset, our method is able to eliminate " + }, + { + "bbox": [ + 317, + 106, + 547, + 201 + ], + "type": "inline_equation", + "content": "32.25\\%" + }, + { + "bbox": [ + 317, + 106, + 547, + 201 + ], + "type": "text", + "content": " of the privacy leakage at a cost of only a " + }, + { + "bbox": [ + 317, + 106, + 547, + 201 + ], + "type": "inline_equation", + "content": "3.96\\%" + }, + { + "bbox": [ + 317, + 106, + 547, + 201 + ], + "type": "text", + "content": " reduction in frame-level AUC performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 213, + 398, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 398, + 225 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 398, + 225 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 234, + 399, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 234, + 399, + 245 + ], + "spans": [ + { + "bbox": [ + 306, + 234, + 399, + 245 + ], + "type": "text", + "content": "Privacy Preservation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 246, + 547, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 547, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 547, + 377 + ], + "type": "text", + "content": "We observe that many works preserve visual privacy at capture time by using non-intrusive sensors such as thermal imaging, depth cameras, or event cameras [29, 18, 21, 1]. Other works allow for raw RGB visual information to be captured, but make an effort to protect the subject privacy in such a way that the data is still useful in a utility task. Earlier efforts aimed at dealing with visual privacy include image downsampling [11] or blocking/blurring privacy-related objects located using pretrained object detectors. Both of these obfuscation methods were shown to reduce utility results by more than they reduced privacy leakage [13, 24, 49]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 378, + 547, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 557 + ], + "type": "text", + "content": "Recent developments have yielded numerous privacy-preserving approaches to action recognition [2, 51, 31]. Wu et al. released an action dataset with privacy labels, PA-HMDB [49]. They use an adversarial learning framework to obfuscate the privacy features using supervised privacy labels. MaSS [4] uses a similar framework to Wu et al. [49], except it adapts a compound loss to flexibly preserve certain attributes instead of destroying them. STPrivacy [24] upgraded the general framework to work with a transformer anonymizing block, masking entire video tubelets unnecessary for action recognition. Following Dave et al.'s SPAcT [13], we adopt a similar self-supervised adversarial anonymization framework without the use of the privacy labels, using NT-Xent [6] contrastive loss in the budget branch to mitigate spatial privacy leakage." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 558, + 391, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 558, + 391, + 569 + ], + "spans": [ + { + "bbox": [ + 306, + 558, + 391, + 569 + ], + "type": "text", + "content": "Anomaly Detection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 700 + ], + "type": "text", + "content": "With such a high volume of available video footage, it is infeasible to create sufficient labelled data to solve supervised VAD. Therefore, many works explore unsupervised methods. These generally train a reconstruction model, then either reconstruct the current frame or try to predict the next frame, signaling an anomaly when reconstruction error is high [9, 35, 23, 33, 50, 44]. Giorgi et al. [16] used an autoencoder with differential privacy, generating anomaly scores from the noisy reconstructed images. This method helped retain some level of subject privacy, but was only evaluated on image quality metrics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Sultani et al. [43] brought weak supervision to VAD," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "text", + "content": "13599" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 274 + ], + "type": "text", + "content": "where anomalies are labelled at the video level. These authors introduce UCF-Crime, a large scale weakly supervised dataset. They propose formulating weakly supervised VAD as a multi instance learning (MIL) problem, showcasing the benefits of temporal smoothness loss & sparsity loss. With the exception of a select few works [54, 52], all following weakly supervised methods are considered anomaly feature representation learning as they improve upon MIL formulation [55, 45, 15, 25, 19, 46, 8, 41], which involves interpreting static video features extracted using an action classifier. Zhong et al. [54] propose a rearranged weakly supervised version of ShanghaiTech. Wu et al. [48] introduce XD-Violence, a large-scale weakly supervised dataset that includes audio, bringing multi-modal fusion to VAD. In this work, we choose to focus on the weakly supervised video anomaly detection setting due to its effectiveness and low annotation effort." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 529 + ], + "type": "text", + "content": "Most VAD works find it useful to model temporal relations between video segments [45, 15, 30, 48, 25, 46, 55, 5]. [30] are able to exploit dynamic variations in features between consecutive segments to help localize anomalies. The authors of RTFM [45] find that anomalous segments tend to have larger feature magnitudes (" + }, + { + "bbox": [ + 46, + 275, + 289, + 529 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 275, + 289, + 529 + ], + "type": "text", + "content": " norm of extracted clip feature vector) than normal segments. They introduce a feature magnitude learning function to help identify anomalous segments with large feature magnitudes. MGFN [8] notices that varying scene attributes causes some normal videos to have larger feature magnitudes than anomalous, so they propose a feature magnitude contrastive loss to help capture these instances. In this work, we use the MGFN model to evaluate anomaly detection performance. Complementing these ideas, [47] demonstrates the effectiveness of explicitly encouraging feature discrimination. Intuitively, these observations can be aggregated by enforcing temporally-distinct feature representations. Building on the terminology introduced in [34] regarding invariant and distinctive representations, in this paper our goal is to enhance the distinctiveness of the representations within our utility branch." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 537, + 104, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 104, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 104, + 550 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "content": "The central idea behind our proposed framework is to develop an anonymization function that can degrade privacy attributes during training without relying on privacy labels. Furthermore, this function must be able to maintain the performance of the weakly-supervised anomaly detection task. Fig. 2 displays a schematic diagram of the proposed framework. In Sec. 3.1, we provide a comprehensive discussion of the problem statement. Next, in Sec. 3.2, we describe the component of the framework and their initialization process. Sec. 3.3 outlines the anonymization function training, where we propose a temporally-distinct triplet loss to enhance the existing self-supervised privacy-preserving framework [13]. Once we learn the anonymization func" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "tion, in Sec. 3.4, we train the anomaly detection model using the privacy-preserved features obtained through our anonymization function. An overview of our complete framework is outlined in Algorithm Sec. 3.5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 128, + 421, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 128, + 421, + 140 + ], + "spans": [ + { + "bbox": [ + 305, + 128, + 421, + 140 + ], + "type": "text", + "content": "3.1. Problem Statement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "text", + "content": "We define our problem statement similar to privacy-preserving action recognition frameworks [13, 47], but, with a different utility task. Suppose we have a video dataset " + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text{anomaly}} = \\{X^{(i)}, Y^i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "text", + "content": ", where, " + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "inline_equation", + "content": "X^{(i)}" + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "text", + "content": " is a video-instance, " + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "text", + "content": " is the total number of samples and " + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "inline_equation", + "content": "Y^{(i)} \\in \\{0,1\\}" + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "text", + "content": " is a binary label. Considering video-level anomaly detection as a utility task " + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "text", + "content": ", and privacy attribute classification as the budget task " + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "text", + "content": ", the aim of a privacy-preserving system is to maintain the performance of " + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "text", + "content": " while reducing " + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "text", + "content": ". To achieve this goal, the system learns an anonymization function " + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "inline_equation", + "content": "f_A" + }, + { + "bbox": [ + 304, + 147, + 545, + 302 + ], + "type": "text", + "content": ", which modifies the original raw data. This goal of privacy-preservation could be fundamentally expressed as following criteria:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 302, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 545, + 350 + ], + "type": "text", + "content": "Criterion-1: The performance of the utility task should not degrade from the original performance, i.e. loss " + }, + { + "bbox": [ + 304, + 302, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_T" + }, + { + "bbox": [ + 304, + 302, + 545, + 350 + ], + "type": "text", + "content": " value of the utility target model " + }, + { + "bbox": [ + 304, + 302, + 545, + 350 + ], + "type": "inline_equation", + "content": "f_T'" + }, + { + "bbox": [ + 304, + 302, + 545, + 350 + ], + "type": "text", + "content": " should remain almost identical before and after applying the anonymization function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 345, + 360, + 545, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 360, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 345, + 360, + 545, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T} \\left(f _ {T} ^ {\\prime} \\left(f _ {A} ^ {*} (X)\\right), Y\\right) \\approx \\mathcal {L} _ {T} \\left(f _ {T} ^ {\\prime} (X), Y\\right). \\tag {1}", + "image_path": "e4b5cd3b6b3ec2d95898e8938a3b8ebccbd4015773b15818f17f34fe0caea97b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "type": "text", + "content": "Criterion-2: Applying anonymization function should increase the loss " + }, + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_B" + }, + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "type": "text", + "content": " for budget " + }, + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "type": "text", + "content": " task of target budget model " + }, + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "type": "inline_equation", + "content": "f_B^\\prime" + }, + { + "bbox": [ + 304, + 383, + 545, + 419 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 355, + 419, + 545, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 419, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 355, + 419, + 545, + 433 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {B} \\left(f _ {B} ^ {\\prime} \\left(f _ {A} ^ {*} (X)\\right)\\right) \\gg \\mathcal {L} _ {B} \\left(f _ {B} ^ {\\prime} (X)\\right). \\tag {2}", + "image_path": "c262f98d7911b48abc727701b392dea378c8ffe984525eeb51531c8e3dbc36dd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 438, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 533 + ], + "type": "text", + "content": "Regarding weakly-supervised anomaly detection (WSAD), most of the existing methods require multi-stage training, meaning they are not end-to-end. This presents a challenge for incorporating it as a utility task in anonymization training. By contrast, privacy-preserving action recognition frameworks [13, 49] have an end-to-end utility task (i.e action recognition), making it more straightforward to include." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "Since most WSAD methods rely on pretrained video encoder features from large-scale action recognition training, we can utilize the exact same anonymization process of privacy-preserving action recognition [13], utilizing action recognition as a proxy-utility task on a proxy-utility action dataset, which we denote as " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text {action }}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ". The limitation of such anonymization training is that by focusing solely on optimizing short clips for action recognition, it neglects the importance of temporally-distinct features. This oversight leads to a notable decline in anomaly detection performance, which relies heavily on these temporally-distinct features, as the training progresses. To resolve this issue, we reformulate the anonymization utility branch to enforce temporal distinctiveness to better align with the anomaly detection task. Hence, our " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_T" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " is weighted sum of action" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "text", + "content": "13600" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 71, + 511, + 294 + ], + "blocks": [ + { + "bbox": [ + 86, + 71, + 511, + 294 + ], + "lines": [ + { + "bbox": [ + 86, + 71, + 511, + 294 + ], + "spans": [ + { + "bbox": [ + 86, + 71, + 511, + 294 + ], + "type": "image", + "image_path": "a5fe680f6fbe9983d2723bcddf3dd34b42af55eadf714f8ea112691e2fe248ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 297, + 321, + 309 + ], + "lines": [ + { + "bbox": [ + 87, + 297, + 321, + 309 + ], + "spans": [ + { + "bbox": [ + 87, + 297, + 321, + 309 + ], + "type": "text", + "content": "(b) Privacy-Preserved Anomaly Detection Workflow" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 317, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 317, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 545, + 373 + ], + "type": "text", + "content": "Figure 2: Full TeD-SPAD framework consisting of the proxy anonymization training followed by the privacy-preserved anomaly detection. (a) shows this proxy training, where UNet is used to anonymize frames in such a way that reduces mutual information between frames while maintaining utility performance. We complement the standard cross-entropy loss with our proposed temporally-distinct triplet loss, which enforces a difference in clip features at distinct timesteps. After training the anonymizer and feature extractor, (b) shows the privacy-preserved workflow, where the anomaly dataset videos are passed through the proxy-trained " + }, + { + "bbox": [ + 46, + 317, + 545, + 373 + ], + "type": "inline_equation", + "content": "f_{A}" + }, + { + "bbox": [ + 46, + 317, + 545, + 373 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 317, + 545, + 373 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 46, + 317, + 545, + 373 + ], + "type": "text", + "content": ", then into any WSAD algorithm." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 393, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 418 + ], + "type": "text", + "content": "recognition loss " + }, + { + "bbox": [ + 46, + 393, + 287, + 418 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{CE})" + }, + { + "bbox": [ + 46, + 393, + 287, + 418 + ], + "type": "text", + "content": " and temporally-distinct triplet loss " + }, + { + "bbox": [ + 46, + 393, + 287, + 418 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_D)" + }, + { + "bbox": [ + 46, + 393, + 287, + 418 + ], + "type": "text", + "content": " which is elaborated in Eq. 6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 433, + 200, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 433, + 200, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 433, + 200, + 445 + ], + "type": "text", + "content": "3.2. Anonymization Framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "text", + "content": "Our anonymization framework consists of 3 main components: (1) Anonymization function " + }, + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "inline_equation", + "content": "(f_A)" + }, + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "text", + "content": ", which is a simple encoder-decoder model with a sigmoid activation. (2) Privacy removal model " + }, + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "inline_equation", + "content": "(f_B)" + }, + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "text", + "content": ", which is an image encoder. (3) Utility model " + }, + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "inline_equation", + "content": "(f_T)" + }, + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "text", + "content": ", which is a video encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "text", + "content": "Network Initialization First of all, our anonymization model is pretrained to initialize as an identity function. This pretraining involves the reconstruction of the frames from " + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text {action }}" + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "text", + "content": " using L1-reconstruction loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 582, + 287, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 582, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 89, + 582, + 287, + 616 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L 1} = \\sum_ {c = 1} ^ {C} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} | x _ {c, h, w} - \\hat {x} _ {c, h, w} |, \\tag {3}", + "image_path": "27f54d1317442f2c7a9a8c8f61f90067015d087c098a00a33d0d80461f1122db.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 627, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 287, + 663 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 627, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 47, + 627, + 287, + 663 + ], + "type": "text", + "content": " is the output of " + }, + { + "bbox": [ + 47, + 627, + 287, + 663 + ], + "type": "inline_equation", + "content": "f_{A}" + }, + { + "bbox": [ + 47, + 627, + 287, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 627, + 287, + 663 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 627, + 287, + 663 + ], + "type": "text", + "content": " is an input image, and " + }, + { + "bbox": [ + 47, + 627, + 287, + 663 + ], + "type": "inline_equation", + "content": "C, H, W" + }, + { + "bbox": [ + 47, + 627, + 287, + 663 + ], + "type": "text", + "content": " corresponds to the channel, height, and width of the input image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Secondly, the privacy model " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{B}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is initialized with the self-supervised weights of SimCLR [7] on ImageNet [14]. The video encoder " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is pretrained with the standard action recognition weights from Kinetics400 dataset [3]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 393, + 444, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 444, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 444, + 406 + ], + "type": "text", + "content": "3.3. Anonymization Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 411, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 545, + 436 + ], + "type": "text", + "content": "Anonymization training is mainly comprised of a minimax optimization of utility loss and privacy loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "text", + "content": "Temporally-distinct triplet loss as Utility Weakly-supervised video anomaly detection methods leverage temporal information to help localize anomalies. [30, 19, 45, 8] show the positive effect of feature magnitude (" + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "text", + "content": " norm of clip feature vector) separability along the temporal dimension. In order to adopt the SPAcT anonymization framework to anomaly detection problems, we utilize a temporal distinctiveness objective in the utility branch. We design a temporally-distinct triplet loss, which increases the agreement between temporally-aligned clips of the same video and increases dissimilarity between representations of the clips which are temporally-misaligned. For anchor clip " + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^{(i)}" + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "text", + "content": ", we obtain the positive clip from the exact same timestamp, but with a differently augmented version denoted as " + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_t^{(i)}" + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "text", + "content": ". Whereas, the negative clip is obtained from different timestamp " + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t'}^{(i)}" + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "inline_equation", + "content": "t' \\neq t" + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "text", + "content": ". This triplet of clips is passed through the utility model " + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "inline_equation", + "content": "f_T" + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "text", + "content": " to achieve features denoted as " + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{(i)}" + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}_t^{(i)}" + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t'}^{(i)}" + }, + { + "bbox": [ + 304, + 436, + 546, + 670 + ], + "type": "text", + "content": ". The proposed temporally-distinct triplet loss can be expressed as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 676, + 545, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 676, + 545, + 694 + ], + "spans": [ + { + "bbox": [ + 321, + 676, + 545, + 694 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} ^ {(i)} = \\max \\{d (\\mathbf {z} _ {t} ^ {(i)}, \\tilde {\\mathbf {z}} _ {t} ^ {(i)}) - d (\\mathbf {z} _ {t} ^ {(i)}, \\mathbf {z} _ {t ^ {\\prime}} ^ {(i)}) + \\mu , 0 \\}, \\qquad (4)", + "image_path": "354e60d364f736cc3769ad35936df272ad237a9f5ec3bd1d0179c0860efccc64.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 700, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 700, + 545, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 700, + 545, + 715 + ], + "type": "inline_equation", + "content": "d(u_j,v_j) = \\|\\mathbf{u}_j - \\mathbf{v}_j\\|_2" + }, + { + "bbox": [ + 306, + 700, + 545, + 715 + ], + "type": "text", + "content": " is Euclidean distance be" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "13601" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "tween two vectors " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": " is the controllable margin hyperparameter to determine how far to push and pull features in the latent space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 108, + 287, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 108, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 287, + 132 + ], + "type": "text", + "content": "We utilize this loss along with the standard cross-entropy action classification loss:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 140, + 287, + 173 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 140, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 110, + 140, + 287, + 173 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C E} ^ {(i)} = - \\sum_ {c = 1} ^ {N _ {C}} \\mathbf {y} _ {c} ^ {(i)} \\log \\mathbf {p} _ {c} ^ {(i)}, \\tag {5}", + "image_path": "84d52cfffbd549b31d935f83c934ebd41d305d7646d7662d0b7c73b7478cd210.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "inline_equation", + "content": "N_{C}" + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "text", + "content": " is the total number of action classes of " + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text {action }}" + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_c^{(i)}" + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "text", + "content": " is the ground-truth label, and " + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_c^{(i)}" + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "text", + "content": " is the prediction vector by utility model " + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "inline_equation", + "content": "f_T" + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 213, + 287, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 287, + 249 + ], + "type": "text", + "content": "Adding both temporal distinctiveness (Eq. 4) and action classification objective to our utility branch, our overall utility loss can be expressed as follows" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 119, + 256, + 287, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 256, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 119, + 256, + 287, + 269 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T} = \\mathcal {L} _ {C E} + \\omega * \\mathcal {L} _ {D}, \\tag {6}", + "image_path": "215f9412f0f6da73d02aa51e4a23b78e9902d7a6c68fd51f17f72998ae4b8f5a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 277, + 287, + 301 + ], + "type": "text", + "content": " hyperparameter is the weight of temporally-distinct triplet loss with respect to cross-entropy loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 301, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 347 + ], + "type": "text", + "content": "Privacy (i.e. budget) Loss " + }, + { + "bbox": [ + 46, + 301, + 287, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_B" + }, + { + "bbox": [ + 46, + 301, + 287, + 347 + ], + "type": "text", + "content": " We utilize the same self-supervised privacy loss from [13], which removes the private information by minimizing the agreement between the frames of the same video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 349, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 287, + 422 + ], + "type": "text", + "content": "Minimax Optimization After reformulating the utility loss, we use the minimax optimization process similar to [13]. It is a two-step iterative process that minimizes the utility loss and at the same time increases budget loss " + }, + { + "bbox": [ + 46, + 349, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_B" + }, + { + "bbox": [ + 46, + 349, + 287, + 422 + ], + "type": "text", + "content": ". At the end of this optimization, we obtain the learned anonymization function " + }, + { + "bbox": [ + 46, + 349, + 287, + 422 + ], + "type": "inline_equation", + "content": "(f_A)" + }, + { + "bbox": [ + 46, + 349, + 287, + 422 + ], + "type": "text", + "content": " and utility video encoder " + }, + { + "bbox": [ + 46, + 349, + 287, + 422 + ], + "type": "inline_equation", + "content": "(f_T)" + }, + { + "bbox": [ + 46, + 349, + 287, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 427, + 251, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 251, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 251, + 441 + ], + "type": "text", + "content": "3.4. Privacy-preserving Anomaly Detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 445, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 517 + ], + "type": "text", + "content": "In order to detect anomalies within videos from " + }, + { + "bbox": [ + 46, + 445, + 287, + 517 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text{anomaly}}" + }, + { + "bbox": [ + 46, + 445, + 287, + 517 + ], + "type": "text", + "content": ", we utilize the current state-of-the-art technique, Magnitude-Contrastive Glance-and-Focus Network (MGFN) [8]. Similar to other recent works in anomaly detection, MGFN requires fixed features for each video from a pretrained video encoder for anomaly detection training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "content": "Optimizing for Anomaly Detection MGFN anomaly detection is comprised of 4 main losses: (1) a standard sigmoid cross-entropy loss " + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "inline_equation", + "content": "L_{sce}" + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "content": " for snippet classification accuracy, (2) a temporal smoothing loss " + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "inline_equation", + "content": "L_{ts}" + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "content": " [43] to encourage smoothness between feature representations of consecutive segments, (3) a sparsity term " + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "inline_equation", + "content": "L_{sp}" + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "content": " [43] to discourage false positive anomalies, and (4) a novel magnitude contrastive loss " + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "inline_equation", + "content": "L_{mc}" + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "content": " to learn scene-adaptive feature distributions across videos, all which help to train a model " + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "inline_equation", + "content": "f_{AD}" + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 625, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 287, + 649 + ], + "type": "text", + "content": "The training loss used in MGFN is compounded in the following equation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 81, + 657, + 287, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 657, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 81, + 657, + 287, + 670 + ], + "type": "interline_equation", + "content": "L _ {A D} = L _ {s c e} + \\lambda_ {1} L _ {t s} + \\lambda_ {2} L _ {s p} + \\lambda_ {3} L _ {m c}, \\tag {7}", + "image_path": "62b625e30a6168e78fb9f9f2621e8011b090f053706a3c7dadae0425019bf00b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_1 = \\lambda_2 = 1" + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\lambda_3 = 0.001" + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "inline_equation", + "content": "f_{AD}" + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "content": " outputs frame-level anomaly scores, which are used to calculate a final ROC AUC and AP for evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 72, + 545, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 169 + ], + "type": "text", + "content": "Feature Extraction In our privacy-preserving case, we cannot use " + }, + { + "bbox": [ + 305, + 72, + 545, + 169 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text{anomaly}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 169 + ], + "type": "text", + "content": " directly for the feature extraction from the video encoder. We first anonymize each video " + }, + { + "bbox": [ + 305, + 72, + 545, + 169 + ], + "type": "inline_equation", + "content": "(X^i)" + }, + { + "bbox": [ + 305, + 72, + 545, + 169 + ], + "type": "text", + "content": " of the dataset through the learned anonymization function " + }, + { + "bbox": [ + 305, + 72, + 545, + 169 + ], + "type": "inline_equation", + "content": "f_A" + }, + { + "bbox": [ + 305, + 72, + 545, + 169 + ], + "type": "text", + "content": " to get an anonymized dataset. For the feature extraction, we utilize the learned utility video encoder " + }, + { + "bbox": [ + 305, + 72, + 545, + 169 + ], + "type": "inline_equation", + "content": "f_T" + }, + { + "bbox": [ + 305, + 72, + 545, + 169 + ], + "type": "text", + "content": ". We denote this extracted set of anonymized features as " + }, + { + "bbox": [ + 305, + 72, + 545, + 169 + ], + "type": "inline_equation", + "content": "\\mathbb{F}_{\\text{anomaly}} = \\{f_T(f_A(X^i)) \\mid \\forall X^i \\in \\mathbb{D}_{\\text{anomaly}}\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 169 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 176, + 377, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 176, + 377, + 189 + ], + "spans": [ + { + "bbox": [ + 306, + 176, + 377, + 189 + ], + "type": "text", + "content": "3.5. Algorithm" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 194, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 194, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 545, + 266 + ], + "type": "text", + "content": "Let's consider the models " + }, + { + "bbox": [ + 305, + 194, + 545, + 266 + ], + "type": "inline_equation", + "content": "f_{A}, f_{T}, f_{B}, f_{AD}" + }, + { + "bbox": [ + 305, + 194, + 545, + 266 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 305, + 194, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\theta_{A}, \\theta_{T}, \\theta_{B}" + }, + { + "bbox": [ + 305, + 194, + 545, + 266 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 194, + 545, + 266 + ], + "type": "inline_equation", + "content": "f_{AD}" + }, + { + "bbox": [ + 305, + 194, + 545, + 266 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 305, + 194, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text{action}}" + }, + { + "bbox": [ + 305, + 194, + 545, + 266 + ], + "type": "text", + "content": " is the proxy action recognition dataset and " + }, + { + "bbox": [ + 305, + 194, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text{anomaly}}" + }, + { + "bbox": [ + 305, + 194, + 545, + 266 + ], + "type": "text", + "content": " is the downstream anomaly detection dataset. All training steps of our framework can be put together into a sophisticated form as algorithm 1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 281, + 466, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 281, + 466, + 293 + ], + "spans": [ + { + "bbox": [ + 312, + 281, + 466, + 293 + ], + "type": "text", + "content": "Algorithm 1: TeD-SPAD Framework" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 297, + 530, + 582 + ], + "type": "list", + "angle": 0, + "index": 39, + "blocks": [ + { + "bbox": [ + 310, + 297, + 351, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 297, + 351, + 308 + ], + "spans": [ + { + "bbox": [ + 310, + 297, + 351, + 308 + ], + "type": "text", + "content": "1 Inputs:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 310, + 309, + 444, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 309, + 444, + 321 + ], + "spans": [ + { + "bbox": [ + 310, + 309, + 444, + 321 + ], + "type": "text", + "content": "2 Datasets: " + }, + { + "bbox": [ + 310, + 309, + 444, + 321 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text{action}}" + }, + { + "bbox": [ + 310, + 309, + 444, + 321 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 310, + 309, + 444, + 321 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text{anomaly}}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 310, + 321, + 509, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 321, + 509, + 333 + ], + "spans": [ + { + "bbox": [ + 310, + 321, + 509, + 333 + ], + "type": "text", + "content": "3 #Epochs: max_anon_epoch, max_ad_epoch" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 310, + 334, + 452, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 334, + 452, + 345 + ], + "spans": [ + { + "bbox": [ + 310, + 334, + 452, + 345 + ], + "type": "text", + "content": "4 Learning Rates: " + }, + { + "bbox": [ + 310, + 334, + 452, + 345 + ], + "type": "inline_equation", + "content": "\\alpha_{AD},\\alpha_B,\\alpha_T" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 310, + 346, + 423, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 346, + 423, + 357 + ], + "spans": [ + { + "bbox": [ + 310, + 346, + 423, + 357 + ], + "type": "text", + "content": "5 Hyperparameters: " + }, + { + "bbox": [ + 310, + 346, + 423, + 357 + ], + "type": "inline_equation", + "content": "\\mu, \\omega" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 310, + 357, + 391, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 357, + 391, + 369 + ], + "spans": [ + { + "bbox": [ + 310, + 357, + 391, + 369 + ], + "type": "text", + "content": "6 Output: " + }, + { + "bbox": [ + 310, + 357, + 391, + 369 + ], + "type": "inline_equation", + "content": "\\theta_{AD},\\theta_{A}" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 310, + 369, + 402, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 369, + 402, + 379 + ], + "spans": [ + { + "bbox": [ + 310, + 369, + 402, + 379 + ], + "type": "text", + "content": "7 Model Initialization:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 310, + 381, + 491, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 381, + 491, + 392 + ], + "spans": [ + { + "bbox": [ + 310, + 381, + 491, + 392 + ], + "type": "text", + "content": "8 Initialize " + }, + { + "bbox": [ + 310, + 381, + 491, + 392 + ], + "type": "inline_equation", + "content": "\\theta_T" + }, + { + "bbox": [ + 310, + 381, + 491, + 392 + ], + "type": "text", + "content": " with Kinetics400 weights [3];" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 310, + 393, + 520, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 393, + 520, + 404 + ], + "spans": [ + { + "bbox": [ + 310, + 393, + 520, + 404 + ], + "type": "text", + "content": "9 Initialize " + }, + { + "bbox": [ + 310, + 393, + 520, + 404 + ], + "type": "inline_equation", + "content": "\\theta_B" + }, + { + "bbox": [ + 310, + 393, + 520, + 404 + ], + "type": "text", + "content": " with SimCLR ImageNet weights [7]." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 309, + 405, + 530, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 405, + 530, + 417 + ], + "spans": [ + { + "bbox": [ + 309, + 405, + 530, + 417 + ], + "type": "text", + "content": "10 Initialize " + }, + { + "bbox": [ + 309, + 405, + 530, + 417 + ], + "type": "inline_equation", + "content": "\\theta_{A}\\gets \\theta_{A} - \\alpha_{A}\\nabla_{\\theta_{A}}(\\mathcal{L}_{L1}(\\theta_{A}))" + }, + { + "bbox": [ + 309, + 405, + 530, + 417 + ], + "type": "text", + "content": " (Ref. Eq. 3)" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 309, + 418, + 422, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 418, + 422, + 429 + ], + "spans": [ + { + "bbox": [ + 309, + 418, + 422, + 429 + ], + "type": "text", + "content": "11 Anonymization Training:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 430, + 465, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 430, + 465, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 430, + 465, + 441 + ], + "type": "text", + "content": "12 for " + }, + { + "bbox": [ + 308, + 430, + 465, + 441 + ], + "type": "inline_equation", + "content": "e_0\\gets 1" + }, + { + "bbox": [ + 308, + 430, + 465, + 441 + ], + "type": "text", + "content": " to max_anon_epoch do" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 441, + 524, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 441, + 524, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 441, + 524, + 502 + ], + "type": "text", + "content": "13 Step-1 \n14 " + }, + { + "bbox": [ + 308, + 441, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\theta_{A}\\gets \\theta_{A} - \\alpha_{A}\\nabla_{\\theta_{A}}(\\mathcal{L}_{T}(\\theta_{A},\\theta_{T}) - \\omega L_{B}(\\theta_{A},\\theta_{B}))" + }, + { + "bbox": [ + 308, + 441, + 524, + 502 + ], + "type": "text", + "content": " \n15 Step-2 \n16 " + }, + { + "bbox": [ + 308, + 441, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\theta_T\\gets \\theta_T - \\alpha_T\\nabla_{\\theta_T}(\\mathcal{L}_T(\\theta_T,\\theta_A))" + }, + { + "bbox": [ + 308, + 441, + 524, + 502 + ], + "type": "text", + "content": " (Ref. Eq. 6) \n17 " + }, + { + "bbox": [ + 308, + 441, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\theta_B\\gets \\theta_B - \\alpha_B\\nabla_{\\theta_B}(\\mathcal{L}_B(\\theta_B,\\theta_A))" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 502, + 336, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 502, + 336, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 502, + 336, + 511 + ], + "type": "text", + "content": "17 end" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 514, + 451, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 451, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 451, + 525 + ], + "type": "text", + "content": "18 Feature Extraction on " + }, + { + "bbox": [ + 308, + 514, + 451, + 525 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text{anomaly}}" + }, + { + "bbox": [ + 308, + 514, + 451, + 525 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 525, + 512, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 512, + 537 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 512, + 537 + ], + "type": "text", + "content": "19 " + }, + { + "bbox": [ + 308, + 525, + 512, + 537 + ], + "type": "inline_equation", + "content": "\\mathbb{F}_{anomaly} = \\{f_T(f_A(X^i))\\mid \\forall X^i\\in \\mathbb{D}_{anomaly}\\}" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 538, + 512, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 538, + 512, + 549 + ], + "spans": [ + { + "bbox": [ + 308, + 538, + 512, + 549 + ], + "type": "text", + "content": "20 Privacy-Preserved Anomaly Detection Training:" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 308, + 550, + 454, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 550, + 454, + 560 + ], + "spans": [ + { + "bbox": [ + 308, + 550, + 454, + 560 + ], + "type": "text", + "content": "21 for " + }, + { + "bbox": [ + 308, + 550, + 454, + 560 + ], + "type": "inline_equation", + "content": "e_0\\gets 1" + }, + { + "bbox": [ + 308, + 550, + 454, + 560 + ], + "type": "text", + "content": " to max_ad_epoch do" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 308, + 561, + 524, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 561, + 524, + 573 + ], + "spans": [ + { + "bbox": [ + 308, + 561, + 524, + 573 + ], + "type": "text", + "content": "22 " + }, + { + "bbox": [ + 308, + 561, + 524, + 573 + ], + "type": "inline_equation", + "content": "\\theta_{AD}\\gets \\theta_{AD} - \\alpha_{AD}\\nabla_{\\theta_{AD}}(L_{AD}(\\theta_{AD},\\mathbb{F}_{anomaly}))" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 308, + 574, + 336, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 574, + 336, + 582 + ], + "spans": [ + { + "bbox": [ + 308, + 574, + 336, + 582 + ], + "type": "text", + "content": "23 end" + } + ] + } + ], + "index": 38 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 615, + 386, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 615, + 386, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 615, + 386, + 628 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 306, + 635, + 368, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 368, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 368, + 646 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": "UCF-Crime [43] is the first large-scale weakly supervised video anomaly detection dataset. It contains 1,900 videos totaling 128 hours of untrimmed CCTV surveillance footage from a variety of different scenes. The videos contain 13 crime-based anomalies such as Arrest, Fighting, and" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "13602" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 180, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 180, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 180, + 83 + ], + "type": "text", + "content": "Shoplifting in real world scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": "XD-Violence [48] is currently the largest weakly supervised video anomaly detection dataset, totaling 217 hours of untrimmed video. All of its anomaly categories are related to violent activities. The videos are gathered from various types of cameras, movies, and games, resulting in a unique blend of scenes for increased difficulty." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 157, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 287, + 217 + ], + "type": "text", + "content": "ShanghaiTech [28] is a medium-scale anomaly detection dataset containing videos covering 13 different scenes with various types of anomalies. While it was published as an unsupervised anomaly detection dataset, we use the weakly supervised rearrangement proposed by [54]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 217, + 287, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 287 + ], + "type": "text", + "content": "VISPR [32] is an image dataset labelled with 68 privacy-related attributes, including gender, hair color, clothes, etc. It provides a multi-class classification problem for us to evaluate privacy-preservation on. The split we use for evaluation along with training details can found in Supp. Sec. B." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 289, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 337 + ], + "type": "text", + "content": "UCF101 [42] is a very common dataset in action recognition, and its relative simplicity makes it practical for learning an anonymization model on, demonstrated in [13]. In this work, split-1 of UCF101 is used as " + }, + { + "bbox": [ + 46, + 289, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text {action }}" + }, + { + "bbox": [ + 46, + 289, + 287, + 337 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 346, + 180, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 180, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 180, + 358 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 365, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 448 + ], + "type": "text", + "content": "Network Architecture Details " + }, + { + "bbox": [ + 46, + 365, + 287, + 448 + ], + "type": "inline_equation", + "content": "f_{A}" + }, + { + "bbox": [ + 46, + 365, + 287, + 448 + ], + "type": "text", + "content": " is a UNet [39] model that transforms raw input frames into anonymized frames. I3D [3] is used for " + }, + { + "bbox": [ + 46, + 365, + 287, + 448 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 46, + 365, + 287, + 448 + ], + "type": "text", + "content": ", to first learn anonymized action classification, then to extract anonymized features. A ResNet-50 [17] is our " + }, + { + "bbox": [ + 46, + 365, + 287, + 448 + ], + "type": "inline_equation", + "content": "f_{B}" + }, + { + "bbox": [ + 46, + 365, + 287, + 448 + ], + "type": "text", + "content": " model during the anonymization training. " + }, + { + "bbox": [ + 46, + 365, + 287, + 448 + ], + "type": "inline_equation", + "content": "f_{AD}" + }, + { + "bbox": [ + 46, + 365, + 287, + 448 + ], + "type": "text", + "content": " is MGFN [8], which consists of a shortcut convolution, a self-attentional convolution, and a feed-forward network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": "Training Process Details First we perform anonymization training for 80 epochs. Adam [22] optimizer is used for all models with a learning rate is 1e-4 with a batch size of 8. Loss weight " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\omega = 0.1" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": " and margin " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\mu = 1" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": " in a default setting. We train the MGFN [8] model using default hyperparameters. " + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "inline_equation", + "content": "f_{B}" + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": " evaluation uses a batch size of 32 and a base learning rate of 1e-3, which follows a linear warmup and a step-based scheduler that drops by a factor of 1/5 upon loss stagnation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 652 + ], + "type": "text", + "content": "Input Details For all experiments, we crop each image to a scale of 0.8, then resize to input resolution " + }, + { + "bbox": [ + 46, + 558, + 287, + 652 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 558, + 287, + 652 + ], + "type": "text", + "content": ". Clips consist of 16 frames and are sampled from a random start with a skip-rate of 2. For anonymization training we utilize standard augmentations like random erase, random crop, horizontal flip and random color jitter. To maintain temporal consistency, augmentations are applied equally to every frame within each clip." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "Feature Extraction Given a raw input video " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\mathrm{i}}\\in" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{anomaly}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " , we first extract " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " clips, where " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " is the amount of non-overlapping 16-frame clips in " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " . We sequentially pass each clip " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "S^{(j)}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " first through our anonymizer " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{A}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " , then our feature extractor " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " Extracted features " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{T}(f_{A}(\\mathbf{X}^{i}))" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "have the shape " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "S \\times C" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " is the dimensionality of the feature vector. Specifically, features are extracted following the average pooling after the mix_5c layer of I3D and have a dimensionality of 2048." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "text", + "content": "Evaluation Protocol and Performance Metrics To evaluate the learned anonymization function " + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "inline_equation", + "content": "f_{A}" + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "text", + "content": ", we follow standard protocols of cross-dataset evaluation [13, 49]. In this protocol, testing videos of " + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text{anomaly}}" + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "text", + "content": " are anonymized by " + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "inline_equation", + "content": "f_{A}" + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "text", + "content": ", and frame-level predictions are obtained through the " + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "inline_equation", + "content": "f_{AD}" + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "text", + "content": ". The calculated ROC AUC is used to evaluate performance on UCF-Crime and ShanghaiTech, and AP is used for XD-Violence. A higher AUC/AP is considered a more accurate anomaly localization. In order to evaluate the privacy leakage, the learned " + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "inline_equation", + "content": "f_{A}" + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "text", + "content": " is utilized to anonymize the privacy dataset " + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text{privacy}}" + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "text", + "content": " to train and evaluate a target privacy model " + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "inline_equation", + "content": "f_{B}'" + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "text", + "content": ". Privacy leakage is measured in terms of the performance of the target " + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "inline_equation", + "content": "f_{B}'" + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "text", + "content": " on the test set of " + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{\\text{privacy}}" + }, + { + "bbox": [ + 304, + 121, + 546, + 311 + ], + "type": "text", + "content": ". Since the privacy dataset is multi-label, the privacy leakage is measured in terms of mean average precision averaged across classes (cMAP)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 312, + 506, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 506, + 323 + ], + "type": "text", + "content": "More implementation details are in Supp. Sec. C." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 331, + 464, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 331, + 464, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 464, + 342 + ], + "type": "text", + "content": "4.3. Privacy-Preserving Baselines" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 349, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 545, + 396 + ], + "type": "text", + "content": "We run well-known self-supervised privacy-preservation techniques for video anomaly detection. In order to maintain a fair comparison across methods, we utilize the exact same network architectures and training process." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 397, + 545, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 432 + ], + "type": "text", + "content": "Downsampling Baselines For Downsample-2x and Downsample-4x, we simply resize the input frames to a lower resolution by a factor of 2 (112 × 112) and 4 (56 × 56)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 433, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 552 + ], + "type": "text", + "content": "Object-Detector Based Obfuscation Baselines Obfuscation techniques are based on first detecting the person, followed by removing (i.e. blackening) or blurring them. Both obfuscation techniques use MS-COCO [26] pretrainedYOLO [37] object detector to obtain bounding boxes for person object class. We utilizeYOLOv51 implementation with yolov5x as backbone. The detected bounding boxes are assigned to pixel value 0 for the Blackening-based baseline. For the Blurring-based baselines, a Gaussian filter with kernel " + }, + { + "bbox": [ + 304, + 433, + 545, + 552 + ], + "type": "inline_equation", + "content": "k = 13" + }, + { + "bbox": [ + 304, + 433, + 545, + 552 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 304, + 433, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\sigma = 10" + }, + { + "bbox": [ + 304, + 433, + 545, + 552 + ], + "type": "text", + "content": " is utilized." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 552, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 588 + ], + "type": "text", + "content": "SPAct [13] Baseline We utilize official implementation2. For a fair comparison with our method, we utilize the exact same utility model I3D and privacy model ResNet-50." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 595, + 539, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 595, + 539, + 608 + ], + "spans": [ + { + "bbox": [ + 306, + 595, + 539, + 608 + ], + "type": "text", + "content": "4.4. Evaluation on Benchmark Anomaly Datasets" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 614, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 614, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 614, + 545, + 685 + ], + "type": "text", + "content": "We compare prior privacy-preserving methods to our method on 3 well-known anomaly detection benchmark datasets. Since privacy-preservation deals with both utility (i.e. anomaly) and privacy, we show results in form of a trade-off plot as presented in Fig. 3. Compared to the prior best method [13], our method is able to remove " + }, + { + "bbox": [ + 304, + 614, + 545, + 685 + ], + "type": "inline_equation", + "content": "19.9\\%" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 693, + 440, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 693, + 440, + 703 + ], + "spans": [ + { + "bbox": [ + 317, + 693, + 440, + 703 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 317, + 693, + 440, + 703 + ], + "type": "text", + "content": "https://github.com/ultralytics/yolov5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 703, + 451, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 703, + 451, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 703, + 451, + 713 + ], + "type": "text", + "content": "2https://github.com/DAVEISHAN/SPAct" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "13603" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 206, + 186 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 206, + 186 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 206, + 186 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 206, + 186 + ], + "type": "image", + "image_path": "504be69d5bbc3416948eaa19b04a10cf9f53fccd18922b5a4f6d235e1df6b522.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 194, + 157, + 205 + ], + "lines": [ + { + "bbox": [ + 100, + 194, + 157, + 205 + ], + "spans": [ + { + "bbox": [ + 100, + 194, + 157, + 205 + ], + "type": "text", + "content": "(a) UCF-Crime" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 70, + 369, + 187 + ], + "blocks": [ + { + "bbox": [ + 211, + 70, + 369, + 187 + ], + "lines": [ + { + "bbox": [ + 211, + 70, + 369, + 187 + ], + "spans": [ + { + "bbox": [ + 211, + 70, + 369, + 187 + ], + "type": "image", + "image_path": "5d36aa73841e905a493ee5cbdd52644fa465295083822de88b62d6030bc0f157.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 194, + 321, + 205 + ], + "lines": [ + { + "bbox": [ + 258, + 194, + 321, + 205 + ], + "spans": [ + { + "bbox": [ + 258, + 194, + 321, + 205 + ], + "type": "text", + "content": "(b) XD-Violence" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 209, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 547, + 232 + ], + "type": "text", + "content": "Figure 3: Trade-off plots between anomaly detection benchmarks [43, 48, 28] AUC and VISPR [32] privacy attribute prediction cMAP for different privacy-preserving methods. Optimal trade-off point is top left of plot (higher AD performance, lower PA prediction ability)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 373, + 71, + 529, + 187 + ], + "blocks": [ + { + "bbox": [ + 373, + 71, + 529, + 187 + ], + "lines": [ + { + "bbox": [ + 373, + 71, + 529, + 187 + ], + "spans": [ + { + "bbox": [ + 373, + 71, + 529, + 187 + ], + "type": "image", + "image_path": "5ee4ec2b383c638fdfda538a350b540ca47a49ae6b6b3853fd024765bda878aa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 194, + 483, + 205 + ], + "lines": [ + { + "bbox": [ + 417, + 194, + 483, + 205 + ], + "spans": [ + { + "bbox": [ + 417, + 194, + 483, + 205 + ], + "type": "text", + "content": "(c) ShanghaiTech" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 251, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 312 + ], + "type": "text", + "content": "more privacy with a slightly better utility score " + }, + { + "bbox": [ + 46, + 251, + 287, + 312 + ], + "type": "inline_equation", + "content": "(1.19\\%)" + }, + { + "bbox": [ + 46, + 251, + 287, + 312 + ], + "type": "text", + "content": ". This strongly supports our claim that promoting temporal distinctiveness during anonymization better aligns with anomaly detection tasks. Numeric data behind Fig. 3 plots can be found in Supp. Sec. D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 319, + 160, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 160, + 333 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 160, + 333 + ], + "type": "text", + "content": "4.5. Qualitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 339, + 287, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 447 + ], + "type": "text", + "content": "Figure 4 shows visual examples of the model outputs in different videos. We note that to the human eye, it is difficult to tell what is going on in each video, yet the anomaly detection model is still able to achieve high performance. Fig. 4a-4d include human subjects, yet none of their private attributes such as face and clothing are visible, and therefore cannot be used to make unfair decisions. In Fig. 4c, 4d, the background shop can still be seen, which is useful context for identifying the shoplifting and robbery anomalies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 447, + 288, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 288, + 544 + ], + "type": "text", + "content": "Figure 5 is a graph of the anomaly score output from both the raw and anonymized video models for each frame of an input video. The blue box shows the ground truth anomalous frames. We see that the green anonymized anomaly scores are similar to the red raw scores while still increasing the anomaly scores on the ground truth anomaly, showing that our anonymization technique maintains utility performance. See Supp. Sec. D for more qualitative results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 551, + 286, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 286, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 286, + 564 + ], + "type": "text", + "content": "4.6. Evidence for Privacy Leakage at Feature-level" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "In anomaly feature representation learning, the anomaly detection algorithms do not directly work with the input videos, the videos are first passed through an action classifier to compute features. Even though the action recognition model sees the original videos, it is not certain whether the private information gets passed to the features. In order to confirm this, we create a simple fully connected network to predict VISPR private attributes, in the same fashion that we evaluate privacy for our other experiments. We stack the same VISPR image 16 times to create a video clip, then extract the clip features through " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": ". Our baseline uses unmodified input images passed through a pretrained Ki" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 305, + 251, + 547, + 305 + ], + "blocks": [ + { + "bbox": [ + 305, + 251, + 547, + 305 + ], + "lines": [ + { + "bbox": [ + 305, + 251, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 547, + 305 + ], + "type": "table", + "html": "
Feature Extraction ModelVISPR
Privacy cMAP (%)Privacy Reduction (%)
Kinetics400 pretrained63.150.0
SPAct [13] Anonymized55.6011.96
Our Anonymized52.3017.18
", + "image_path": "732d05a70aec4776af6a6135129ebd6bf0e90df4919c5c5a0e1138d354ac4581.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 360, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 547, + 445 + ], + "type": "text", + "content": "netics400 [20] I3D [3] model, with the other experiments using a paired anonymizer and tuned I3D model. Detailed explanations of this process are found in Supp. Sec. C. Through experimentation, we find that the action classifier latent features do in fact leak private information, therefore this private information gets passed into the anomaly detector. Table 1 demonstrates empirical evidence of this." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 450, + 400, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 450, + 400, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 400, + 464 + ], + "type": "text", + "content": "4.7. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 469, + 547, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 547, + 553 + ], + "type": "text", + "content": "Effect of different utility losses " + }, + { + "bbox": [ + 304, + 469, + 547, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_T" + }, + { + "bbox": [ + 304, + 469, + 547, + 553 + ], + "type": "text", + "content": " We study the effect of different utility losses during the anonymization process on the final privacy vs. utility anomaly detection performance in Table 2. From row-(a,b), we see a clear reduction in privacy at no cost of anomaly performance; which demonstrates the effectiveness of our proposed temporally-distinct triplet loss " + }, + { + "bbox": [ + 304, + 469, + 547, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 304, + 469, + 547, + 553 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 553, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 547, + 590 + ], + "type": "text", + "content": "To this extent, we also implement the contrary objective to our " + }, + { + "bbox": [ + 304, + 553, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 304, + 553, + 547, + 590 + ], + "type": "text", + "content": " which promotes temporal-invariance " + }, + { + "bbox": [ + 304, + 553, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_I" + }, + { + "bbox": [ + 304, + 553, + 547, + 590 + ], + "type": "text", + "content": " in the utility branch of the anonymization training. We implement" + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 314, + 598, + 539, + 680 + ], + "blocks": [ + { + "bbox": [ + 305, + 314, + 547, + 348 + ], + "lines": [ + { + "bbox": [ + 305, + 314, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 547, + 348 + ], + "type": "text", + "content": "Table 1: Quantitative evidence of action classification model features leaking privacy. Features from each model used to predict privacy attributes. Red indicates higher privacy leakage." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 598, + 539, + 680 + ], + "lines": [ + { + "bbox": [ + 314, + 598, + 539, + 680 + ], + "spans": [ + { + "bbox": [ + 314, + 598, + 539, + 680 + ], + "type": "table", + "html": "
Utility Loss during Anonymization (L_T)VISPR Privacy cMAP(%)↓)UCF-Crime Anomaly AUC(%)↑)
(a)L_CE52.7173.93
(b)L_CE + L_D42.2174.81
(c)L_CE + L_I45.6469.52
", + "image_path": "350633f4732a541278f5112e214ed22cfb2d90476f6d8f098b56d5667d387965.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 687, + 546, + 711 + ], + "lines": [ + { + "bbox": [ + 305, + 687, + 546, + 711 + ], + "spans": [ + { + "bbox": [ + 305, + 687, + 546, + 711 + ], + "type": "text", + "content": "Table 2: Ablation with different utility losses during the anonymization process. Bold indicates best trade-off." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "13604" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 271, + 145 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 271, + 145 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 271, + 145 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 271, + 145 + ], + "type": "image", + "image_path": "a8321a9e33e2b12f8faa319d453c7a10aface53ba71b28506d2eb99213154e55.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 162, + 270, + 236 + ], + "blocks": [ + { + "bbox": [ + 111, + 150, + 208, + 161 + ], + "lines": [ + { + "bbox": [ + 111, + 150, + 208, + 161 + ], + "spans": [ + { + "bbox": [ + 111, + 150, + 208, + 161 + ], + "type": "text", + "content": "(a) Fighting018_x264.mp4" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 162, + 270, + 236 + ], + "lines": [ + { + "bbox": [ + 49, + 162, + 270, + 236 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 270, + 236 + ], + "type": "image", + "image_path": "d6eebd3155d2efefa41ec86e1617a22c31c46130b303846c9024b2581ae7444c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 106, + 241, + 213, + 252 + ], + "lines": [ + { + "bbox": [ + 106, + 241, + 213, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 213, + 252 + ], + "type": "text", + "content": "(c) Shoplifting048_x264.mp4" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 324, + 71, + 545, + 145 + ], + "blocks": [ + { + "bbox": [ + 324, + 71, + 545, + 145 + ], + "lines": [ + { + "bbox": [ + 324, + 71, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 324, + 71, + 545, + 145 + ], + "type": "image", + "image_path": "250498cf4e879a6d3604d71777892b1ec7f3d003af4bce5fde5a956505898d71.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 324, + 162, + 545, + 236 + ], + "blocks": [ + { + "bbox": [ + 384, + 150, + 484, + 161 + ], + "lines": [ + { + "bbox": [ + 384, + 150, + 484, + 161 + ], + "spans": [ + { + "bbox": [ + 384, + 150, + 484, + 161 + ], + "type": "text", + "content": "(b) Shooting040_x264.mp4" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 324, + 162, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 324, + 162, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 324, + 162, + 545, + 236 + ], + "type": "image", + "image_path": "cfce969b4b19fb4c33655c199374dc0bd156dd5a8d809728239125e6f67c7093.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 241, + 483, + 252 + ], + "lines": [ + { + "bbox": [ + 385, + 241, + 483, + 252 + ], + "spans": [ + { + "bbox": [ + 385, + 241, + 483, + 252 + ], + "type": "text", + "content": "(d) Robbery023_x264.mp4" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 123, + 290, + 473, + 495 + ], + "blocks": [ + { + "bbox": [ + 46, + 256, + 546, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 546, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 546, + 278 + ], + "type": "text", + "content": "Figure 4: Qualitative results of our anonymization training on UCF-Crime [43] dataset. For each case, the top row shows the raw video frames, and the bottom row shows the frames after being passed through the anonymizer." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 123, + 290, + 473, + 495 + ], + "lines": [ + { + "bbox": [ + 123, + 290, + 473, + 495 + ], + "spans": [ + { + "bbox": [ + 123, + 290, + 473, + 495 + ], + "type": "image", + "image_path": "5b45179a14f44024d90fe51c43c4228acab1da763fe6448c342fa8d13869d5be.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 504, + 546, + 527 + ], + "lines": [ + { + "bbox": [ + 46, + 504, + 546, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 546, + 527 + ], + "type": "text", + "content": "Figure 5: Frame-level anomaly score plot from XD-Violence. Green line shows our anonymized model, red line is the raw input model, both compared to the blue ground truth line. The below visualizations show uniformly sampled frames from the video." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 547, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 547, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 547, + 287, + 608 + ], + "type": "text", + "content": "this using well-known self-supervised works [36, 12]. From row-(c) we see that temporal-invariance objective is not well-suited for anomaly detection utility tasks and results in a significant drop of " + }, + { + "bbox": [ + 47, + 547, + 287, + 608 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 47, + 547, + 287, + 608 + ], + "type": "text", + "content": ". We provide extensive experiments with " + }, + { + "bbox": [ + 47, + 547, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_I" + }, + { + "bbox": [ + 47, + 547, + 287, + 608 + ], + "type": "text", + "content": " and its explanation in Supp. Sec. D." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 613, + 278, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 278, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 278, + 624 + ], + "type": "text", + "content": "Effect of different temporal distinctiveness objectives:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Our objective of temporal distinctiveness can also be achieved through a contrastive loss. To this end we utilize the implementation of local-local temporal contrastive loss (LLTC) of [12]. It achieves " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "75.06\\%" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " AUC at " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "42.44\\%" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " cMAP (Table 3), which is closely comparable to our results using triplet loss. It is important to note that LLTC loss significantly increases the computation (GPU memory require" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 547, + 545, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 547, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 545, + 595 + ], + "type": "text", + "content": "ment) compared to triplet loss. It requires 8 clips (4 clips " + }, + { + "bbox": [ + 304, + 547, + 545, + 595 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 304, + 547, + 545, + 595 + ], + "type": "text", + "content": " 2 augmented view) which results in 132.45G FLOPs compared to 3 clips of triplet loss requiring only 49.67G FLOPs in our experimental setup." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": "Relative Weightage of " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " Here we test the effect of changing the weight of the additional temporally-distinct triplet loss. Table 4 shows that weighting the loss at 0.1 achieves our best results of " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "32.25\\%" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " relative increase in privacy with only a " + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "inline_equation", + "content": "3.69\\%" + }, + { + "bbox": [ + 304, + 605, + 546, + 713 + ], + "type": "text", + "content": " reduction in utility performance. Without the enforced temporal distinctiveness, the utility model is limited by the quality of the reconstructed anonymized videos. The improper weighting of the temporal loss interferes with the classifier ability of the model," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "13605" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 269, + 150 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 269, + 150 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 269, + 150 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 269, + 150 + ], + "type": "table", + "html": "
Temporal Distinctiveness Loss LDVISPR Privacy cMAP(%)↓)UCF-Crime Anomaly AUC(%)(↑)
None52.7173.93
LLTC [12]42.4475.06
Triplet42.2174.81
", + "image_path": "0c8aa71aad0e57f7edd7861f8c12dd69d6e0d61bf1203fc06d65dca53527da3b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 219, + 268, + 323 + ], + "blocks": [ + { + "bbox": [ + 46, + 163, + 288, + 208 + ], + "lines": [ + { + "bbox": [ + 46, + 163, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 288, + 208 + ], + "type": "text", + "content": "Table 3: Comparison of different loss formulations to encourage temporal distinctiveness during the anonymization process. Both achieved similar results, with the triplet loss requiring significantly less computation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 219, + 268, + 323 + ], + "lines": [ + { + "bbox": [ + 67, + 219, + 268, + 323 + ], + "spans": [ + { + "bbox": [ + 67, + 219, + 268, + 323 + ], + "type": "table", + "html": "
TripletVISPRUCF-Crime
TemporalPrivacyAnomaly
Loss Weight ωcMAP(%) (↓)AUC(%) (↑)
052.7173.93
0.0153.8472.53
0.142.2174.81
1.055.2670.56
10.051.6769.27
", + "image_path": "b0ae30e1449879bfbf5455ea79a231aa6e877d9f646688b35b5d0cc07c2c834f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 391, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 287, + 426 + ], + "type": "text", + "content": "which can also harm the anonymization process. This suggests that action recognition loss " + }, + { + "bbox": [ + 46, + 391, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 46, + 391, + 287, + 426 + ], + "type": "text", + "content": " is still an important utility task for anomaly detection performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "text", + "content": "Effect of the margin in " + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "text", + "content": " The proposed " + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "text", + "content": " temporally-distinct triplet loss uses a margin hyperparameter " + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "text", + "content": " to allow for adjustments to the contrastive distance in the latent feature space. The only requirement is that " + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "inline_equation", + "content": "\\mu > 0" + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "text", + "content": ". The intuition here is that a larger margin enforces greater feature spacing. As Table 5 shows, we empirically found that setting " + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "inline_equation", + "content": "\\mu = 1" + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "text", + "content": " gives us the most robust results. A lower margin results in less temporally-distinct representations, which makes distinguishing between normal and anomalous features more difficult. On the other hand, increasing the " + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "inline_equation", + "content": "\\mu = 2" + }, + { + "bbox": [ + 46, + 426, + 288, + 582 + ], + "type": "text", + "content": " results in more difficult temporal triplet loss (i.e a very high temporal distinctiveness) which may not align well with anomaly detection task." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 67, + 592, + 268, + 672 + ], + "blocks": [ + { + "bbox": [ + 46, + 333, + 288, + 377 + ], + "lines": [ + { + "bbox": [ + 46, + 333, + 288, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 288, + 377 + ], + "type": "text", + "content": "Table 4: Comparison of using different loss weights of the temporally-distinct triplet loss during the anonymization process. The margin hyperparameter of the temporal triplet loss for each experiment was 1. Bold indicates best trade-off." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 592, + 268, + 672 + ], + "lines": [ + { + "bbox": [ + 67, + 592, + 268, + 672 + ], + "spans": [ + { + "bbox": [ + 67, + 592, + 268, + 672 + ], + "type": "table", + "html": "
TripletVISPRUCF-Crime
TemporalPrivacyAnomaly
Loss Margin μcMAP(%) (↓)AUC(%) (↑)
0.549.7862.12
1.042.2174.81
2.058.8667.30
", + "image_path": "fca4abafd1baf2b628be232cacfe59f40b14fc3a8951e444aa7727f5b348ec38.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "content": "Effect of temporal sampling in " + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "content": " The proposed triplet loss (Eq. 4) forms the negative from the clip " + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t'}^{(i)}" + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "content": " of a different timestamp " + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "content": ". Distance between the timestamp of negative and anchor clip " + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "inline_equation", + "content": "t - t'" + }, + { + "bbox": [ + 304, + 72, + 545, + 253 + ], + "type": "text", + "content": " is an important aspect to define temporal distinctiveness. We perform experiments with various distances as shown in Table 6. In our default setting, we use random distance as shown in the first row. From the second row, we can say that a smaller distance leads to a better anomaly score with a slight degradation in protecting privacy. At the same time, the third row suggests that enforcing temporal distinctiveness at higher distances leads to better privacy protection but at the cost of anomaly performance. This distance hyperparameter may be used as a tuning parameter to get different operating points of privacy vs. anomaly trade-off." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 337, + 262, + 516, + 343 + ], + "blocks": [ + { + "bbox": [ + 46, + 677, + 288, + 711 + ], + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 711 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 711 + ], + "type": "text", + "content": "Table 5: Comparison of using different hyperparameter margins of the temporally-distinct triplet loss during the anonymization process. Bold indicates default setup." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 262, + 516, + 343 + ], + "lines": [ + { + "bbox": [ + 337, + 262, + 516, + 343 + ], + "spans": [ + { + "bbox": [ + 337, + 262, + 516, + 343 + ], + "type": "table", + "html": "
Negative Clip DistanceVISPR Privacy cMAP(%)（↓）UCF-Crime Anomaly AUC(%)(↑)
Random42.2174.81
846.3476.12
3228.6970.97
", + "image_path": "b5d9a59b3b2fd3db3906bfaf2404c4ddac3c7a21d9ee908bb1b5bde44b720e09.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 356, + 545, + 389 + ], + "lines": [ + { + "bbox": [ + 304, + 356, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 545, + 389 + ], + "type": "text", + "content": "Table 6: Comparison of enforcing set clip sampling distance during the anonymization process. The margin hyperparameter of the temporal triplet loss for each experiment was 1." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 411, + 379, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 379, + 424 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 379, + 424 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 432, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 648 + ], + "type": "text", + "content": "In this paper, we highlight the importance of privacy, a previously neglected aspect of video anomaly detection. We present TeD-SPAD, a framework for applying Temporal Distinctiveness to Self-supervised Privacy-preserving video Anomaly Detection. TeD-SPAD demonstrates the effectiveness of using a temporally-distinct triplet loss while anonymizing an action recognition model, as it enhances feature representation temporal distinctiveness, which complements the downstream anomaly detection model. By effectively destroying spatial private information, we remove the model's ability to use this information in its decision-making process. As a future research direction, this framework can be extended to other tasks, such as spatiotemporal anomaly detection. The anonymizing encoder-decoder may also be made more powerful with techniques using recent masked image modeling. It is our hope that this work contributes to the development of more responsible and unbiased automated anomaly detection systems." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 658, + 421, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 421, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 421, + 671 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in part by the National Science Foundation (NSF) and Center for Smart Streetscapes (CS3) under NSF Cooperative Agreement No. EEC-2133516." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "text", + "content": "13606" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Shafiq Ahmad, Pietro Morerio, and Alessio Del Bue. Person re-identification without identification via event anonymization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Shafiq Ahmad, Gianluca Scarpellini, Pietro Morerio, and Alessio Del Bue. Event-driven re-id: A new benchmark and method towards privacy-preserving person re-identification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 459–468, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 237 + ], + "type": "text", + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 4, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 282 + ], + "type": "text", + "content": "[4] Chun-Fu Chen, Shaohan Hu, Zhonghao Shi, Prateek Gulati, Bill Moriarty, Marco Pistoia, Vincenzo Piuri, and Pierangela Samarati. Mass: Multi-attribute selective suppression. arXiv preprint arXiv:2210.09904, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 327 + ], + "type": "text", + "content": "[5] Haoyang Chen, Xue Mei, Zhiyuan Ma, Xinhong Wu, and Yachuan Wei. Spatial-temporal graph attention network for video anomaly detection. Image and Vision Computing, 131:104629, Mar. 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 287, + 373 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations, June 2020. arXiv:2002.05709 [cs, stat]. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 374, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 287, + 407 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 409, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 287, + 474 + ], + "type": "text", + "content": "[8] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 387-395, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 287, + 508 + ], + "type": "text", + "content": "[9] Yang Cong, Junsong Yuan, and Ji Liu. Sparse reconstruction cost for abnormal event detection. In CVPR 2011, pages 3449-3456, June 2011. ISSN: 1063-6919. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "text", + "content": "[10] Ji Dai, Behrouz Saghafi, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Towards privacy-preserving recognition of human activities. In 2015 IEEE international conference on image processing (ICIP), pages 4238-4242. IEEE, 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "text", + "content": "[11] Ji Dai, Behrouz Saghafi, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Towards privacy-preserving recognition of human activities. In 2015 IEEE International Conference on Image Processing (ICIP), pages 4238-4242, Sept. 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[12] Ishan Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. TCLR: Temporal contrastive learning for video representation. Computer Vision and Image Understanding, 219:103406, June 2022. 8, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[13] Ishan Rajendrakumar Dave, Chen Chen, and Mubarak Shah. Spact: Self-supervised privacy preservation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 97, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 150 + ], + "type": "text", + "content": "[15] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE International Conference on Computer Vision and Pattern Recognition, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 152, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 207 + ], + "type": "text", + "content": "[16] Giacomo Giorgi, Wisam Abbasi, and Andrea Saracino. Privacy-Preserving Analysis for Remote Video Anomaly Detection in Real Life Environments. Journal of Wireless Mobile Networks, Ubiquitous Computing, and Dependable Applications, 13(1):112-136, Mar. 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[18] Carlos Hinojosa, Miguel Marquez, Henry Arguello, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Privhar: Recognizing human actions from privacy-preserving lens. In European Conference on Computer Vision, pages 314–332. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 310, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 374 + ], + "type": "text", + "content": "[19] Chao Huang, Chengliang Liu, Jie Wen, Lian Wu, Yong Xu, Qiuping Jiang, and Yaowei Wang. Weakly Supervised Video Anomaly Detection via Self-Guided Temporal Discriminative Transformer. IEEE Transactions on Cybernetics, pages 1-14, 2022. Conference Name: IEEE Transactions on Cybernetics. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 376, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 432 + ], + "type": "text", + "content": "[20] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natev, Mustafa Suleyman, and Andrew Zisserman. The Kinetics Human Action Video Dataset, May 2017. arXiv:1705.06950 [cs]. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 434, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 476 + ], + "type": "text", + "content": "[21] Junho Kim, Young Min Kim, Yicheng Wu, Ramzi Zahred-dine, Weston A. Welge, Gurunandan Krishnan, Sizhuo Ma, and Jian Wang. Privacy-Preserving Visual Localization with Event Cameras, Dec. 2022. arXiv:2212.03177 [cs]. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "text", + "content": "[22] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "text", + "content": "[23] B. Ravi Kiran, Dilip Mathew Thomas, and Ranjith Parakkal. An Overview of Deep Learning Based Methods for Unsupervised and Semi-Supervised Anomaly Detection in Videos. Journal of Imaging, 4(2):36, Feb. 2018. Number: 2 Publisher: Multidisciplinary Digital Publishing Institute. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 568, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 633 + ], + "type": "text", + "content": "[24] Ming Li, Jun Liu, Hehe Fan, Jia-Wei Liu, Jiahe Li, Mike Zheng Shou, and Jussi Keppo. STPrivacy: Spatio-Temporal Tubelet Sparsification and Anonymization for Privacy-preserving Action Recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[25] Shuo Li, Fang Liu, and Licheng Jiao. Self-Training Multi-Sequence Learning with Transformer for Weakly Supervised Video Anomaly Detection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):1395–1403, June 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "13607" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "text", + "content": "C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[27] Jixin Liu and Leilei Zhang. Indoor privacy-preserving action recognition via partially coupled convolutional neural network. In 2020 International Conference on Artificial Intelligence and Computer Engineering (ICAICE), pages 292-295. IEEE, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "text", + "content": "[28] W. Liu, D. Lian W. Luo, and S. Gao. Future frame prediction for anomaly detection - a new baseline. In 2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "text", + "content": "[29] Zelun Luo, Jun-Ting Hsieh, Niranjan Balachandar, Serena Yeung, Guido Pusiol, Jay Luxenberg, Grace Li, Li-Jia Li, N Lance Downing, Arnold Milstein, and Li Fei-Fei. Computer Vision-Based Descriptive Analytics of Seniors' Daily Activities for Long-Term Health Monitoring. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 308 + ], + "type": "text", + "content": "[30] Hui Lv, Chuanwei Zhou, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Localizing anomalies from weakly-labeled videos. IEEE transactions on image processing, 30:4505-4515, 2021. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 288, + 363 + ], + "type": "text", + "content": "[31] Saemi Moon, Myeonghyeon Kim, Zhenyue Qin, Yang Liu, and Dongwoo Kim. Anonymization for skeleton action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 15028-15036, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 410 + ], + "type": "text", + "content": "[32] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Towards a visual privacy advisor: Understanding and predicting privacy risks in images. In IEEE International Conference on Computer Vision (ICCV), 2017. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "text", + "content": "[33] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning Memory-guided Normality for Anomaly Detection, Mar. 2020. arXiv:2003.13228 [cs]. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "type": "text", + "content": "[34] Mandela Patrick, Yuki M Asano, Polina Kuznetsova, Ruth Fong, João F Henriques, Geoffrey Zweig, and Andrea Vedaldi. On compositions of transformations in contrastive self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9577-9587, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 578 + ], + "type": "text", + "content": "[35] Oluwatoyin P. Popoola and Kejun Wang. Video-Based Abnormal Human Behavior Recognition—A Review. IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), 42(6):865-878, Nov. 2012. Conference Name: IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews). 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "type": "text", + "content": "[36] Rui Qian, Tianjian Meng, Boqing Gong, Ming-Hsuan Yang, Huisheng Wang, Serge Belongie, and Yin Cui. Spatiotemporal contrastive video representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6964-6974, 2021. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[37] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You Only Look Once: Unified, Real-Time Object Detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 779-788, June 2016. ISSN: 1063-6919. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[38] Zhongzheng Ren, Yong Jae Lee, and Michael S Ryoo. Learning to anonymize faces for privacy preserving action" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "detection. In Proceedings of the European conference on computer vision (ECCV), pages 620-636, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 161 + ], + "type": "text", + "content": "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 162, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 547, + 205 + ], + "type": "text", + "content": "[40] Michael S Ryoo, Brandon Rothrock, Charles Fleming, and Hyun Jong Yang. Privacy-preserving human activity recognition from extreme low resolution. In Thirty-First AAAI Conference on Artificial Intelligence, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 206, + 547, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 547, + 250 + ], + "type": "text", + "content": "[41] Hitesh Sapkota and Qi Yu. Bayesian nonparametric submodular video partition for robust anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3212-3221, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 251, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 547, + 282 + ], + "type": "text", + "content": "[42] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild, Dec. 2012. arXiv:1212.0402 [cs]. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 284, + 547, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 284, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 547, + 337 + ], + "type": "text", + "content": "[43] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-World Anomaly Detection in Surveillance Videos. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6479-6488, Salt Lake City, UT, June 2018. IEEE. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 338, + 547, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 547, + 380 + ], + "type": "text", + "content": "[44] Qiyue Sun and Yang Yang. Unsupervised video anomaly detection based on multi-timescale trajectory prediction. Computer Vision and Image Understanding, 227:103615, Jan. 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 383, + 547, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 547, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 547, + 446 + ], + "type": "text", + "content": "[45] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975-4986, 2021. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 480 + ], + "type": "text", + "content": "[46] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In ECCV, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 482, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 535 + ], + "type": "text", + "content": "[47] Peng Wu and Jing Liu. Learning Causal Temporal Relation and Feature Discrimination for Anomaly Detection. IEEE Transactions on Image Processing, 30:3513-3527, 2021. Conference Name: IEEE Transactions on Image Processing. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 537, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 591 + ], + "type": "text", + "content": "[48] Peng Wu, jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 593, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 646 + ], + "type": "text", + "content": "[49] Zhenyu Wu, Haotao Wang, Zhaowen Wang, Hailin Jin, and Zhangyang Wang. Privacy-preserving deep action recognition: An adversarial learning framework and a new dataset. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[50] Hongchun Yuan, Zhenyu Cai, Hui Zhou, Yue Wang, and Xiangzhi Chen. TransAnomaly: Video Anomaly Detection Using Video Vision Transformer. IEEE Access, 9:123977-123986, 2021. Conference Name: IEEE Access. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[51] Sharare Zehtabian, Siavash Khodadadeh, Ladislau Bölönii, and Damla Turgut. Privacy-preserving learning of human" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "text", + "content": "13608" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 307 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 106 + ], + "type": "text", + "content": "activity predictors in smart environments. In IEEE INFOCOM 2021-IEEE Conference on Computer Communications, pages 1-10. IEEE, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "type": "text", + "content": "[52] Chen Zhang, Guorong Li, Yuankai Qi, Shuhui Wang, Laiyun Qing, Qingming Huang, and Ming-Hsuan Yang. Exploiting completeness and uncertainty of pseudo labels for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16271-16280, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 175, + 286, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 286, + 206 + ], + "type": "text", + "content": "[53] Zhixiang Zhang, Thomas Cilloni, Charles Walter, and Charles Fleming. Multi-scale, class-generic, privacy-preserving video. *Electronics*, 10(10):1172, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 286, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 286, + 272 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 286, + 272 + ], + "type": "text", + "content": "[54] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1237-1246, 2019. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 274, + 286, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 274, + 286, + 307 + ], + "spans": [ + { + "bbox": [ + 49, + 274, + 286, + 307 + ], + "type": "text", + "content": "[55] Yi Zhu and Shawn Newsam. Motion-Aware Feature for Improved Video Anomaly Detection, July 2019. arXiv:1907.10211 [cs, eess]. 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "13609" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Teaching CLIP to Count to Ten/56d47f2c-86af-49f0-9e72-8decf5e81c8a_content_list.json b/2023/Teaching CLIP to Count to Ten/56d47f2c-86af-49f0-9e72-8decf5e81c8a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0b6d80a75306af81007a71d53641417eb02000ea --- /dev/null +++ b/2023/Teaching CLIP to Count to Ten/56d47f2c-86af-49f0-9e72-8decf5e81c8a_content_list.json @@ -0,0 +1,1962 @@ +[ + { + "type": "text", + "text": "Teaching CLIP to Count to Ten", + "text_level": 1, + "bbox": [ + 325, + 130, + 645, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Roni Paiss $^{1,2}$", + "bbox": [ + 246, + 185, + 352, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inbar Mosseri1", + "bbox": [ + 274, + 205, + 393, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ariel Ephrat1", + "bbox": [ + 374, + 186, + 480, + 203 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Michal Irani1,3", + "bbox": [ + 436, + 205, + 553, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Omer Tov1", + "bbox": [ + 503, + 186, + 591, + 203 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shiran Zada", + "bbox": [ + 591, + 186, + 738, + 203 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tali Dekel1,3", + "bbox": [ + 591, + 204, + 692, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Weizmann Institute of Science", + "bbox": [ + 591, + 222, + 844, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Google Research", + "bbox": [ + 124, + 220, + 271, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91eee95232937f0e252058f7c8c5886e2d7a2d041b876888df891b7a006fd360.jpg", + "image_caption": [ + "(a) Image Retrieval Results (from LAION):" + ], + "image_footnote": [], + "bbox": [ + 127, + 270, + 405, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f8bfc496bef39364a209a44028010b02b8a5719d381c46c36f5b74ed665813f.jpg", + "image_caption": [ + "Figure 1: Inability of CLIP to count. We show that CLIP is insensitive to the number of objects in an image, and demonstrate the effectiveness of our Counting-Aware CLIP. (a) Image retrieval results using text captions of the form “a photo of ”, with different numbers and types of objects. Images that match the caption are marked with $\\checkmark$ ; images that do not are marked with $\\times$ . Our Counting-Aware model retrieves images that depict the requested number of objects, while the original CLIP often retrieves images that contain the wrong number of objects, or images where the number is explicitly written in the image (e.g. “nine hearts” - the image contains the number “9”, but has 11 hearts). (b) Attention maps demonstrating that our model attends to all matching object instances in the image, unlike the original CLIP." + ], + "image_footnote": [], + "bbox": [ + 418, + 270, + 671, + 486 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c98c803b7445f5f5ac68ae7b8587f31b5a883e8793b620c6a23af52beb17b6d.jpg", + "image_caption": [ + "(b) Relevancy Maps:" + ], + "image_footnote": [], + "bbox": [ + 692, + 282, + 852, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 614, + 313, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large vision-language models (VLMs), such as CLIP, learn rich joint image-text representations, facilitating advances in numerous downstream tasks, including zero-shot classification and text-to-image generation. Nevertheless, existing VLMs exhibit a prominent well-documented limitation - they fail to encapsulate compositional concepts such as counting. We introduce a simple yet effective method to improve the quantitative understanding of VLMs, while maintaining their overall performance on common benchmarks. Specifically, we propose a new counting-contrastive loss used to finetune a pre-trained VLM in tandem with its original objective. Our counting loss is deployed over automatically-created counterfactual examples, each consisting of an image and a caption containing an incorrect object count. For example, an image depicting three dogs is paired with the caption \"Six dogs playing in the yard\" as a negative example. Our loss encourages discrimination", + "bbox": [ + 75, + 643, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "between the correct caption and its counterfactual variant which serves as a hard negative example. To the best of our knowledge, this work is the first to extend CLIP's capabilities to object counting. Furthermore, we introduce \"CountBench\" - a new image-text counting benchmark for evaluating object counting capabilities. We demonstrate a significant improvement over state-of-the-art baseline models on this task. Finally, we leverage our counting-aware CLIP model for image retrieval and text-conditioned image generation, demonstrating that our model can produce specific counts of objects more reliably than existing ones.", + "bbox": [ + 496, + 616, + 893, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 799, + 632, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since the advent of CLIP [38], training large vision-language models (VLMs) has become a prominent paradigm for representation learning in computer vision. By observing huge corpora of paired images and captions crawled from the Web, these models learn powerful", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3170", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and rich joint image-text embedding spaces, which have been employed in numerous visual tasks, including classification [58, 59], segmentation [27, 53], motion generation [47], image captioning [30, 48], text-to-image generation [12, 29, 32, 41, 44] and image or video editing [4, 6, 9, 17, 25, 35, 49, 18, 7]. Recently, VLMs have also been a key component in text-to-image generative models [5, 39, 41, 43], which rely on their textual representations to encapsulate the semantic meaning of the input text.", + "bbox": [ + 75, + 90, + 467, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their power, prominent VLMs, such as CLIP [38] and BASIC [36], are known to possess a weak understanding of the number of objects present in an image [3, 36, 38]. This is demonstrated in Fig. 1, where, when given a caption of the template \"a photo of \", CLIP often fails to retrieve images that correctly match the described number. Downstream applications that rely on VLM-based representations inherit these limitations, e.g., image generation models struggle to reliably produce specific counts of objects [43, 51, 52].", + "bbox": [ + 75, + 228, + 467, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce a novel method that enhances the quantitative understanding of large-scale VLMs by encouraging them to produce representations that are sensitive to the number of objects in the image and text.", + "bbox": [ + 76, + 381, + 467, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We hypothesize that the reason existing VLMs fail to learn the concept of counting is severalfold: (i) Captions that accurately specify the number of objects become extremely rare in the data as the number of objects increases. For example, we found that for more than five objects, captions typically contain a general form of quantity, e.g., \"a group of...\" or \"many...\", rather than an accurate count. This can be attributed to the fact that people cannot instantly identify numerical quantities larger than four without explicitly counting the objects [24]. (ii) Numbers in the caption often refer to attributes that are NOT related to counting - e.g. age, time, address, temperature, model number (\"This is an iPhone 5\"), etc. (iii) The task of counting (associating the visible number of objects in an image with the number in the caption), is not explicitly enforced in current VLM training objectives. Therefore, current VLMs are able to count reasonably well only up to two or three (for which there are sufficient image-caption examples).", + "bbox": [ + 75, + 443, + 467, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We thus suggest to mitigate each of these problems by: (i) Creating suitable training data in which the captions contain accurate numbers of objects. (ii) Designing a training objective whereby understanding object counts is critical for discriminating between the correctly associated caption and incorrect ones.", + "bbox": [ + 75, + 717, + 467, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More specifically, as illustrated in Fig. 2, we automatically create a clean and diverse counting training set by curating image-text examples where the image depicts multiple objects and its caption is verified to express their count. We then finetune a pretrained VLM by formulating counting as a discriminative task – for each example, we create", + "bbox": [ + 75, + 809, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a counterfactual caption by swapping the spelled number associated with the object count with a different randomly selected number. The model's objective is then to associate the image correctly with its true count caption, discriminating it from the counterfactual one.", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate counting capabilities, we introduce \"Count-Bench\" - a carefully curated object counting benchmark, consisting of 540 diverse, high quality image-text examples. We evaluate our method on two prominent contrastive VLMs: CLIP [38] and BASIC [36], and demonstrate a significant improvement in accuracy in the task of zero-shot count classification over baseline models. Importantly, we achieve this while maintaining the original knowledge learned by the VLM, as demonstrated by an extensive evaluation of our model on standard zero-shot downstream tasks. The quantitative understanding of our model is further evident by our text-to-image retrieval results (e.g., Fig. 1(a)), as well as by the relevancy maps of our model, which demonstrate that the model correctly attends to all visible objects whose count is specified in the text (e.g., Fig. 1(b)). Finally, we train a large-scale text-to-image generative model [43] which incorporates our counting training set and counting-aware CLIP text encoder. The generated images from this model exhibit higher fidelity to the number of objects specified in the input prompts (Fig. 8).", + "bbox": [ + 496, + 167, + 890, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions are:", + "bbox": [ + 500, + 469, + 781, + 482 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A novel training framework for tackling the task of vision-language counting - an important limitation of current VLMs.", + "2. A new benchmark, \"CountBench\", carefully filtered and validated for evaluating VLMs on the counting task.", + "3. We apply our method to the widely-adopted VLMs CLIP and BASIC, demonstrating significant improvement on the counting task, while maintaining general (non-counting) performance on common benchmarks.", + "4. We utilize our counting-aware VLMs for downstream tasks including image retrieval and text-to-image generation, demonstrating more reliable results when the text prompt contains a specific number of objects." + ], + "bbox": [ + 501, + 489, + 890, + 689 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 709, + 635, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive vision-language models: Vision-language models have demonstrated impressive success in vision and multimodal tasks [2, 11, 36, 38, 46, 54, 37]. In this work, we focus on contrastive VLMs, such as CLIP [38] and BASIC [36], as they are widely used both for downstream tasks and as backbones for generative models [40, 43]. Both are trained with a contrastive objective, where matching text-image pairs should have a low cosine distance, and non-matching texts and images should be far apart. The representations computed by CLIP have proven to be very effective in vision and multimodal tasks, due to their zero-shot", + "bbox": [ + 496, + 734, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3171", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8a31b465668d00df40532343923c99b3a06dffdeb2d866729316ad01775add10.jpg", + "image_caption": [ + "(a) Obtaining Counting Training Set:" + ], + "image_footnote": [], + "bbox": [ + 81, + 108, + 470, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/73c8c96307071d976e9816dda85057c3279c8dc726282ab2e619184fa84eaa57.jpg", + "image_caption": [ + "(b) Teaching CLIP to count:", + "Figure 2: Method overview (a) We create a text-image counting training set in which each caption expresses the number of objects depicted in the corresponding image. This is done by using an off-the-shelf object detector to automatically identify text-image examples in which the text count matches the number of visible objects in the image (see Sec. 3.1). (b) We finetune a pre-trained CLIP model using our counting subset (a), through a dedicated contrastive objective $L_{count}$ , used in addition to the original (general) contrastive objective ( $L_{clip}$ ). Specifically, given a text-image example from our counting subset, we automatically create a counterfactual prompt by replacing the true object count in the original caption with an incorrect count; $L_{count}$ encourages the model to embed the image close to its original caption embedding (expressing the true object count) and far from its counterfactual count (see Sec. 3.2)." + ], + "image_footnote": [], + "bbox": [ + 73, + 222, + 478, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "capabilities and semantic nature, and have been used as a prominent component in numerous tasks and methods. BASIC [36] scaled up the size of the model, batch size and dataset, improving zero-shot accuracy on common benchmarks.", + "bbox": [ + 75, + 717, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Subitizing: \"Subitizing\" [24] is the ability of people to instantly recognize the number of objects at a glance - without actually counting them. It has been shown [23] that most humans can subitize only a small number of objects (up to $\\sim 4$ ). This cognitive phenomenon is reflected in the numbers people tend to specify in captions ( $\\leqslant 4$ ), and sub-", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sequentially in the counting abilities of VLMs, which struggle to count larger numbers. Subitizing was also referred to in computer vision [50, 56]. These papers aim to predict a very small number ( $\\leqslant 4$ ) of salient objects in an image, and are pure image-based (do not involve any text or VLMs).", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Counting in vision-language models: While demonstrating impressive recognition capabilities, large VLMs such as CLIP[38] and BASIC [36] are known to only partially capture the meaning of the text. Thus, they fail to understand the number of objects in an image [3, 36, 38]. Paiss et al. [34] demonstrated that CLIP attends to only a small subset of its input, mainly the nouns, and often ignores adjectives, numbers and prepositions.", + "bbox": [ + 496, + 185, + 890, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Counting has remained a stand-alone task under the domain of visual question answering (VQA), tackled with specifically designed architectures and techniques [1, 31, 57]. Our work defers from these prior efforts in several key aspects: (i) While previous efforts are restricted to VQA architectures and problem formulation, our goal is to improve the quantitative understanding of general-purpose contrastive VLMs (e.g., CLIP and BASIC), used in various vision and multimodal tasks where counting-aware solutions are not currently available. (ii) Our work can enhance the zero-shot counting capabilities of VLMs to unrestricted objects, unlike prior methods that are trained on specific domains, which can be problematic for new domains where no counting labels are available.", + "bbox": [ + 496, + 308, + 892, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 537, + 589, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to teach a pre-trained VLM (e.g., CLIP) to count, i.e., to improve its quantitative textual and visual understanding. Our framework, illustrated in Fig. 2, consists of two main stages. We first automatically create a counting training set, comprising clean and diverse images along with corresponding captions that describe the number of visible objects in the scene. We then leverage this dataset to finetune the VLM through a designated count-based contrastive loss that is used in tandem with the original generic image-text objective.", + "bbox": [ + 496, + 566, + 890, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More specifically, our key idea is to automatically generate counterfactual examples by swapping the true object count in the caption with a different random number. Our new counting loss encourages the model to embed an image close to its true count, as expressed by the original caption, while pushing it away from the embedding of the counterfactual count prompt. As the only difference between the correct caption and its counterfactual counterpart is a single word—the spelled number of objects—the model has to distinguish between the correct and incorrect count in order to succeed in its training task. Next, we describe our dataset creation and finetuning paradigm in detail.", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3172", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Creating an image-text counting train set", + "text_level": 1, + "bbox": [ + 76, + 90, + 429, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A naive approach for obtaining an image-text counting dataset is to filter a large-scale dataset for examples in which the caption contains a number. However, this approach results in a highly noisy dataset, since the numbers in the captions often refer to other numerical attributes that are unrelated to object counts. Such numerical attributes are age (\"A 7 year old girl\"), time (\"The time is 9 o'clock\"), addresses, model numbers, etc (see examples in supplementary). To ensure that the numbers in the captions indeed specify object counts, we employ several stages of automatic filtering in our data pipeline (Fig. 2 (a)):", + "bbox": [ + 75, + 125, + 472, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we filter out all examples whose caption does not contain a spelled number $\\in \\{\\text{\"two\",\\ldots,\"ten\"}\\}$ . We do so, as we observed that non-spelled numbers, or numbers higher than ten, mostly appear in conjunction with a measure of time, (e.g. dates) or addresses, rather than numbers of objects present in the image.", + "bbox": [ + 75, + 303, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the second stage, we verify that the spelled numbers serve as object counters, and that the counted objects are visible and detectable in the images. For example, for the caption \"A photo of three dogs\", we verify that the image indeed depicts three visible dogs, no more, and no less. This count verification is achieved automatically by first applying an off-the-shelf object detector [22], and counting the number of detections per object. We assume that the caption refers to the most prevalent object in the image. Thus, we retain only examples for which the number specified in the caption aligns with the number of instances of the maximally-detected object. We denote by $C$ our automatically filtered train set.", + "bbox": [ + 75, + 405, + 472, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Naturally, the filtered data $C$ is imbalanced. The number of examples that pass our filtering drops significantly as the count increases, e.g., the number of \"ten\" image-text pairs is around $1000 \\times$ smaller than \"two\". Training with such imbalanced data creates a bias—the loss can be reduced by classifying frequent numbers as the correct caption and rare numbers as counterfactual, regardless of the image content. Therefore, balancing the data is of essence. Due to scarcity of examples depicting more than six objects, we choose to balance the numbers \"two\" – \"six\" separately from the higher numbers \"seven\" – \"ten\". For each of the numbers \"two\" – \"six\", we sample around $37K$ samples, while for \"seven\" – \"ten\", we use all the samples passed by our filter. There are approximately $7K$ samples for \"seven\" down to around $1.5K$ samples for \"ten\". We found this approach to provide us with a diverse and relatively balanced training dataset, yet more sophisticated methods could be considered in the future. From this point on, $C$ will denote our filtered and balanced numbered training set.", + "bbox": [ + 75, + 613, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Teaching CLIP to count", + "text_level": 1, + "bbox": [ + 500, + 90, + 720, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to improve the quantitative understanding of a pre-trained VLM (e.g., CLIP), while preserving its real-world knowledge, as reflected by its zero-shot capabilities on commonly-evaluated benchmarks. Therefore, we use a combination of two loss functions:", + "bbox": [ + 498, + 113, + 893, + 188 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {C L I P} + \\lambda L _ {\\text {c o u n t}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 200, + 892, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{CLIP}$ is the regular contrastive loss of CLIP, $L_{count}$ is our counting-designated loss (described below), and $\\lambda$ is a hyperparameter used to weight the two losses.", + "bbox": [ + 496, + 227, + 890, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We finetune the model on two training sets: (i) A very large dataset collected from the Web that contains general in-the-wild images and captions. (ii) Our filtered numbered training set $C$ , described in Sec. 3.1, which contains samples where object counts are spelled out in the captions. While $L_{CLIP}$ is calculated on all samples, the counting loss $L_{count}$ is calculated only on samples from $C$ . For each image-text pair $(i_k, t_k) \\in C$ , a counterfactual caption $t_k^{CF}$ is automatically created by swapping the number in the caption $t_k$ with a different random number (e.g., the caption \"five dogs\" can be counterfactualized with \"eight dogs\"). At each training step, the triplets $(i_k, t_k, t_k^{CF})_{k=1}^N$ are then fed to CLIP's text and image encoders to obtain their embeddings $(ei_k, et_k, et_k^{CF})_{k=1}^N$ .", + "bbox": [ + 496, + 273, + 892, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, a contrastive loss $L_{count}$ is computed to enforce that the similarity score of the image is high with the original caption and low with the counterfactual caption:", + "bbox": [ + 496, + 484, + 893, + 530 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c o u n t}} = - \\frac {1}{N} \\sum_ {k = 1} ^ {N} \\log \\frac {\\exp \\left(e i _ {k} \\cdot e t _ {k}\\right)}{\\exp \\left(e i _ {k} \\cdot e t _ {k}\\right) + \\exp \\left(e i _ {k} \\cdot e t _ {k} ^ {C F}\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 551, + 890, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the original ground truth caption and counterfactual caption differ only by the number of objects specified in them, this loss encourages the model to learn the relationship between the specified spelled number and the number of the objects it refers to.", + "bbox": [ + 496, + 604, + 890, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the negative samples only in the counting objective $L_{count}$ , instead of adding them to the batch for the contrastive loss $L_{CLIP}$ in order to better weight their effect.", + "bbox": [ + 496, + 680, + 890, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. \"CountBench\" - an evaluation benchmark", + "text_level": 1, + "bbox": [ + 498, + 739, + 880, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce a new object-counting evaluation benchmark called CountBench. This benchmark was automatically curated (and manually verified) from the publicly available LAION-400M image-text dataset [45]. CountBench contains a total of 540 images containing between 2 and 10 instances of a particular object, where their corresponding captions reflect this number. This benchmark is used only for testing and evaluation. CountBench has no overlap with our training set $C$ .", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3173", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/60a53d01d6a9ce3c2b1c101e54229284346907e7b53a99bf0eed46556aee0dd5.jpg", + "image_caption": [ + "(a) Baseline CLIP (L/14)" + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 292, + 248 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f72c4513de1134111a33303371c9c4125b9deb7b71eb4fb0718f6d9ca0f9e3f4.jpg", + "image_caption": [ + "(b) Ours (L/14)" + ], + "image_footnote": [], + "bbox": [ + 295, + 87, + 472, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The images in CountBench were obtained by running our automatic filtering method described in Sec. 3.1 on the entire LAION-400M dataset. The filtering pipeline includes applying an object detector, whose benefits are twofold: (i) it facilitates verification of the object count specified in the caption, and (ii) it ensures that the counted objects are visible and detectable. The latter is important since benchmarks that do not disentangle the problem at hand from other challenges (e.g. occlusions and low resolution) may not reflect the performance on the actual task, since the difficulty can often stem from unrelated factors [14].", + "bbox": [ + 75, + 371, + 468, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This filtering produced over 158K images for the number \"two\", but only around 100 images for \"ten\", demonstrating again the severe number imbalance we encountered with our training set. After automatically balancing each number to 100-200 samples each, the entire dataset was manually verified to contain only pairs in which the spelled number in the caption matches the number of clearly visible objects in the image. The dataset was rebalanced after this stage, ending up with 60 image-text pairs per number $\\in\\{\\text{\"two\",\"..},\\text{\"ten\"}\\}$ , 540 in total. Samples from the dataset can be seen in Fig. 4.", + "bbox": [ + 75, + 537, + 468, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the CountBench benchmark to evaluate the counting abilities of the models trained with our method in Sec. 5.", + "bbox": [ + 75, + 705, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 768, + 209, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We thoroughly evaluate our method, both quantitatively and qualitatively, on object counting-related tasks using our CountBench benchmark. We further validate that the performance of our counting-aware models on a variety of general zero-shot classification benchmarks is retained [8, 13, 15, 16, 19, 20, 21, 26, 33, 42, 55]. To gain a better understanding of our models, we show visualizations", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/809322e4d93c9c1c9fcc88d8ca69afa6090ee5a3399ef0e8c40779bbf8074e6f.jpg", + "image_caption": [ + "\"two zebras in Cape Town\"" + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 606, + 160 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b4bf294d4124d4f9a8c295c122d52bf63f60ce1bc63a3fc69efe0af171d6ffa0.jpg", + "image_caption": [ + "\"three purebred chihuahuaas running...\"" + ], + "image_footnote": [], + "bbox": [ + 609, + 89, + 746, + 160 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e5eee1eaa287978d9ca669b958af5e975956b258f8c511dbff96408a952f7c1e.jpg", + "image_caption": [ + "\"...four antique pink and purple armchairs\"" + ], + "image_footnote": [], + "bbox": [ + 751, + 89, + 890, + 160 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f7020c2a1501ee4efb6c80d611fcdf289fb86300e156a6925d8088aaa5f1f2d9.jpg", + "image_caption": [ + "\"GT's five favourite Champagne...\"" + ], + "image_footnote": [], + "bbox": [ + 501, + 181, + 630, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fac855746b08ea777dc8fff3a873c7d6181ce5873289b6cbc2b3a02c18710ceb.jpg", + "image_caption": [ + "\"vintage silver plate tablespoons...set of six\"", + "\"seven Chewy Peanut Butter Cookies...\"" + ], + "image_footnote": [], + "bbox": [ + 638, + 181, + 776, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af29610451d12236218eff88d75dd5dcdee5352eceee7d27533794583d1e246a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 181, + 890, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2edc89da46c157df84fecdb8ce32e2025531e54496d17b28ba7b6476a7e015a1.jpg", + "image_caption": [ + "Figure 3: Confusion matrices on CountBench. Classification accuracy on CountBench, broken down into confusion matrices for the public CLIP ViT-L/14 (a), and our counting-aware CLIP ViT-L/14 model (b), demonstrating clear quantitative superiority of our model.", + "...eight colorful bright shiny red apples...", + "Figure 4: CountBench benchmark. Sample images and their corresponding captions from our new CountBench object counting benchmark. This benchmark was automatically curated (and manually verified) from the publicly-available LAION-400M dataset." + ], + "image_footnote": [], + "bbox": [ + 501, + 276, + 632, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2b4d7081091cf46a15d607ee3a58790cd8c805c541c9f8571f26a7e17e6715a9.jpg", + "image_caption": [ + "\"The view of the nine leftmost moai...\"" + ], + "image_footnote": [], + "bbox": [ + 638, + 276, + 777, + 349 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ad0fcb7059a51d822e0dc10016f65fdb018648ff4457989345d3075f4c6191f1.jpg", + "image_caption": [ + "\"All ten Christmas Poinsettia...\"" + ], + "image_footnote": [], + "bbox": [ + 785, + 276, + 890, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of text-image relevancy maps, along with per-word relevancy scores, demonstrating that our model indeed attends to the number of objects in the image and text. Finally, we apply our model to text-to-image retrieval and generation, producing specific numbers of objects more reliably than baseline models.", + "bbox": [ + 496, + 481, + 890, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Zero-shot counting accuracy", + "text_level": 1, + "bbox": [ + 500, + 587, + 754, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our models and baselines on CountBench on the task of classifying the number of objects in an image in a zero-shot manner. For each image in CountBench we augment the existing caption with eight other possible captions by replacing the number in its caption with all the numbers $\\in \\{\\text{two}, \\ldots, \\text{ten}\\}$ , and calculate the similarity score between the image and each of the nine captions. The number in the caption that obtains highest similarity score with the image is considered the predicted number.", + "bbox": [ + 496, + 612, + 890, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 reports the results of this evaluation on two prominent contrastive VLMs: CLIP-B/32 and BASIC-S. We report both the counting accuracy (selection of the correct number) and the mean deviation of the models' predictions from the correct numbers. For each of the architectures, we compare our model (configuration $E$ ) with two baseline configurations: (A) the official baseline model, and (B) the baseline model finetuned on our general text-image dataset used in our implementation, with the standard contrastive loss. Comparing the performance of these config", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3174", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5487f7e23302be46c5055ac4c71d9b951458dfa3f05a34640b51a96d8bae6b3c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CLIP-B/32BASIC-S
A Official BaselineB Internal BaselineC Ours (w/o Lcount)D Ours (Naive Filtering)E OursA Official BaselineB Internal BaselineC Ours (w/o Lcount)D Ours (Naive Filtering)E Ours
Accuracy ↑31.6732.9444.2649.8175.9317.9722.7530.5928.6869.02
Mean deviation from the correct number ↓1.531.440.971.280.492.132.021.291.870.64
", + "bbox": [ + 88, + 88, + 883, + 188 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3baf9b6a1f4af6a0e2678ef3ae3a1d6abefe220273dcf73af28db96d17c93d14.jpg", + "table_caption": [ + "Table 1: Quantitative counting results. Top-1 zero-shot accuracy and the mean absolute distance between the predicted numbers and the true numbers on CountBench. We compare several configurations: (A) The official CLIP [38] and BASIC [36] models. (B) The official baselines finetuned on our internal curated data. (C) Models trained with our filtered counting set, without $L_{count}$ (D) Models finetuned with $L_{count}$ on a naively filtered counting set (E) Our method, which is significantly superior to all other configurations, both in accuracy and deviation from correct number." + ], + "table_footnote": [], + "table_body": "
CLIP-B/32BASIC-S
Official BaselineInternal BaselineOursOfficial BaselineInternal BaselineOurs
ImageNet62.9364.9764.0659.7061.9661.18
CIFAR1063.9161.0060.6576.2284.6984.05
CIFAR10033.1032.4933.5645.3556.8055.89
Caltech10175.9982.5082.3678.1681.0381.05
EuroSAT45.2341.6637.6928.3945.8245.97
Food10183.0880.7280.5377.0877.8077.06
ImageNetA31.8530.8529.8117.6522.5521.68
ImageNetR69.3870.1770.3067.1167.6866.95
ImageNetV255.6556.5656.6252.2254.3553.60
Oxford Pets87.3587.7487.4180.6285.1584.87
Oxford Flowers66.1465.7367.3964.7466.4065.90
", + "bbox": [ + 83, + 290, + 460, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: Zero-shot accuracy on common benchmarks. We compare the zero-shot accuracy of our method and baselines on a variety of popular benchmarks. As seen, our method preserves the performance of the original model.", + "bbox": [ + 75, + 448, + 468, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "urations allows us to quantify the effect of using our own large-scale text-image dataset, which differs from the original unpublished data the models were trained on.", + "bbox": [ + 75, + 531, + 468, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As can be seen, our method $(E)$ achieves significantly superior counting accuracy compared to the baselines $(A, B)$ , attaining $2 - 3 \\times$ higher counting accuracy and more than $3 \\times$ lower mean deviation from the correct number. Results for different categories in CountBench are reported in App. 1.", + "bbox": [ + 75, + 580, + 468, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 also contains an ablation of the two components of our method: filtering a counting training set and finetuning with an additional loss $L_{count}$ . Models with configuration $C$ are finetuned on the filtered subset with no counting loss. The large gap in accuracy on CountBench between configurations $C$ and $E$ shows the importance of our loss for the improvement in counting abilities. Models with configuration $D$ are finetuned with the counting loss $L_{count}$ on an alternative counting subset, which consists of all the samples that contain spelled numbers $\\in \\{ \"two\",.., \"ten\" \\}$ without additional filtering. The significant difference in counting accuracy between configurations $D$ and $E$ demonstrates the importance of our restrictive filtering pipeline, as both configurations are finetuned with $L_{count}$ over the samples from a dedicated counting training set. As can be seen in Tab. 1, while the naively filtered data does improve performance", + "bbox": [ + 75, + 659, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/88ed71bec5f49597fb8e758243758382837c7068e11280ad99421afd496216da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 285, + 635, + 352 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d77c3f1fea7e20d39b76477ea44e25ebdc61b08744524187df25655b93fdfe5a.jpg", + "image_caption": [ + "\"Two ducks\" 0.23 0.77" + ], + "image_footnote": [], + "bbox": [ + 637, + 285, + 751, + 351 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9bc28b3af8480c287b33546b2bce9017a574af90c90c41f5dcfa86257fcc9e08.jpg", + "image_caption": [ + "\"Two ducks\" 0.45 0.55" + ], + "image_footnote": [], + "bbox": [ + 753, + 285, + 867, + 351 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3cf24cb2e1961fc90d2f81a01ace2126be093af29398a0f837802ff4f38369cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 372, + 635, + 426 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/69547da3ae2ee280ea58edeb7e3ba2b83426ef2c89e51ff0f73d63fdecef3fa6.jpg", + "image_caption": [ + "\"Three pumpkins\" 0.16 0.84" + ], + "image_footnote": [], + "bbox": [ + 637, + 372, + 751, + 426 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/25f91edf186465f58417bbebbe5a24a8956fc030015c019bc7f59f7d92ec3542.jpg", + "image_caption": [ + "\"Three pumpkins\" 0.45 0.55" + ], + "image_footnote": [], + "bbox": [ + 753, + 372, + 867, + 426 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b208bdb5d1566e2206f404ce99325a0cdb6408779e278fa188ad8e2e83af8eb.jpg", + "image_caption": [ + "Figure 5: Relevancy maps of both image and text. Visualization of the relevancy maps of both image and text, which represent, for each patch in the image and token in the text, how important it is to the prediction. Using our counting-aware CLIP model, the relevancy of the number (e.g., \"four\") in the text is increased. In addition, the model focuses on areas in the image that are relevant for counting." + ], + "image_footnote": [], + "bbox": [ + 521, + 448, + 635, + 501 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/36002ba95075fead18bba537240aefdb39b32682285019ba140700f92dc5c5fe.jpg", + "image_caption": [ + "Four cards 0.30 0.70" + ], + "image_footnote": [], + "bbox": [ + 637, + 448, + 751, + 501 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a86317d09edd9c60f83620a0061d575f796add80a7392b018526cbf4903c2bb1.jpg", + "image_caption": [ + "\"Four cards\" 0.43 0.57" + ], + "image_footnote": [], + "bbox": [ + 753, + 448, + 867, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "over a baseline trained without a dedicated counting subset, the obtained results are still significantly lower than those produced by our model. We attribute this gap in performance to mislabeled training samples in the naively filtered data, which are absent from our counting training set $C$ due to our filtering pipeline.", + "bbox": [ + 496, + 641, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Confusion matrices for the counting evaluation described above are shown in Fig. 3. For this experiment, we compare a public CLIP-L/14 model against our counting-aware version of it. As can be seen, our counting-aware CLIP model is significantly superior to the baseline across all numbers. It is worth noting that the baseline CLIP performs reasonably well for numbers within the subitizing range $(2 - 4)$ , but fails on larger numbers. This supports our claim that curated image captions with numbers beyond this range are rare, since specifying them requires manually counting the objects.", + "bbox": [ + 496, + 734, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3175", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9b288a39429169ee78a73b8e0159f6b682a0ec79bb79996a0d89b3ccf90dcfdd.jpg", + "image_caption": [ + "Figure 6: Relevancy maps for similarity between the image and different numbers. We compare the relevancy map of the input image with text prompts of several numbers (i.e. four to six) for both the baseline CLIP and our counting-aware CLIP. Our model focuses on the five stars when calculating the similarity with the prompt \"five\"." + ], + "image_footnote": [], + "bbox": [ + 81, + 80, + 467, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Count-based image retrieval", + "text_level": 1, + "bbox": [ + 76, + 300, + 331, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We consider the task of text-to-image retrieval where the text explicitly describes the desired count of objects. To obtain a diverse dataset that consists of varied numbers of objects, yet facilitates retrieval in reasonable time, we split the public LAION-400M dataset [45] into coarse categorical subsets by filtering samples where the caption contains a certain word (e.g., \"dogs\", \"animals\", \"cars\"), and perform retrieval on each of these subsets separately.", + "bbox": [ + 75, + 323, + 468, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For each category, we use the caption \"a photo of \" where number $\\in$ {\"two\",..,\"ten\"} (e.g. \"a photo of six dogs\"). For each caption, we retrieve the three images in the dataset that are predicted by the model to be most similar to the caption. Note that since there are no ground truth labels for the counts of objects, we present qualitative results. Fig. 7 shows the retrieved images using the original CLIP model and our counting-aware CLIP model. An extended figure can be found in the supplementary. As can be seen, when the requested number is larger than three, the images retrieved by the baseline model often depict arbitrary numbers of objects. Additionally, the baseline often retrieves the same images for several different requested numbers. This further implies that the baseline model mostly focuses on the existence of the described object in the image, and ignores the number in the caption. In contrast, our results depict accurate object counts in most cases.", + "bbox": [ + 75, + 445, + 472, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Performance on non-counting classification", + "text_level": 1, + "bbox": [ + 76, + 726, + 444, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify that our counting-aware models preserve the powerful image-text representation capabilities of the original models, we evaluate the zero-shot performance of our models on a variety of common classification benchmarks. Table 2 reports the zero-shot accuracy of our counting aware models against the baselines (corresponding to configurations $A$ , $B$ in Tab. 1). As can be seen, our models maintain similar overall accuracy. Additionally, comparing the official baseline and the internal baseline indicates that finetuning the models on our general text-image datasets leads to", + "bbox": [ + 75, + 750, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c953c2ed96e8d820c8bb4adc5cafd6d49100ff1e5e582ea8f5575099330d4fb.jpg", + "image_caption": [ + "\"A photo of {} shirts.\"", + "Figure 7: Top-3 count-based image retrieval Text-to-image retrieval results for different counts of shirts, ordered from left to right according to their similarity score (descending). Images that match the caption are marked with $\\checkmark$ and images that do not match it are marked with X. Extended results are in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 506, + 102, + 890, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "only a slight shift in the accuracy of the models on common benchmarks. An additional evaluation on the task of general zero-shot image retrieval is reported in the supplementary.", + "bbox": [ + 496, + 619, + 892, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Relevancy map visualization", + "text_level": 1, + "bbox": [ + 500, + 679, + 754, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To gain a better understanding of what our model learns, we use the explainability method of Chefer et al. [10] to visualize relevancy maps, which indicate the importance of different parts of the text and image to the model's similarity score prediction. Fig. 5 displays the normalized relevancy maps of several image-text pairs. Examining the relevancy maps of the text, it is apparent that the relevancy score of the spelled number in the caption is significantly higher for our model than the baseline model, which suggests that our model concentrates more on the mentioned number than the original one. Additionally, examining the relevancy maps of the images, it is evident that our model focuses on all pertinent objects in the image, whereas the original model", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3176", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c97bf8ffdabe56d3ed450c20f7de34cfc2f33c62c97b2998b11bb126d4da6f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 106, + 169, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e0001ba9b2ed6a87b7e609a22d9c4f2d9b8a8cc28900ae6f10c140e3d6258482.jpg", + "image_caption": [ + "\"{} colorful tulips\"" + ], + "image_footnote": [], + "bbox": [ + 173, + 106, + 256, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ebe34c920450ea7c10831de000a7d5cd2520856fb71157c21d005f885dd837b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 106, + 343, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/426d15c5a0c64de4dfbb85d6b1148ea3d66c789081fcc85f07ff3d4f821308b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 106, + 398, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e91967d31ccb05b26080542fe688696ae515cfc785cca4405ceb17c35b79b63e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 106, + 480, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0bc615c021b687a209861375174b46f94e9ec0e774b8d380d775da1d81298f0a.jpg", + "image_caption": [ + "\"six\"", + "Figure 8: Generated samples with Imagen using our counting-aware CLIP as backbone. The Imagen model benefits from the counting-aware representations produced by the our CLIP model, and is able to generate images that accurately follow the amounts specifies in the captions." + ], + "image_footnote": [], + "bbox": [ + 86, + 162, + 148, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db909db848b4f13445c4831d1595874f0455457fc232f3e9a674f867c06d46da.jpg", + "image_caption": [ + "\"seven\"" + ], + "image_footnote": [], + "bbox": [ + 151, + 162, + 232, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce72b3daa2f8576c510945ef95efe9da054c57d935a7ca5c39596912c2139328.jpg", + "image_caption": [ + "\"eight\"" + ], + "image_footnote": [], + "bbox": [ + 233, + 162, + 302, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65c1382adcc99112df7c077cde483b17a94438615e241eb15b235c073f30de74.jpg", + "image_caption": [ + "\"nine\"" + ], + "image_footnote": [], + "bbox": [ + 305, + 162, + 393, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e87c9cdbe4c81ff5d3134626a484e351b6117e00a05e29131994a2e1b2eac164.jpg", + "image_caption": [ + "\"ten\"" + ], + "image_footnote": [], + "bbox": [ + 397, + 162, + 482, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32b94397b9daf17b2b617470b432e3770b095bff07386d0e07abb93d2fbaa4aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 107, + 553, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4af404bbe7405baa102568802635fffcb01ef2b2f6a3fd907aed01a080d827d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 108, + 607, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f83638fe5c5d627631142fccc209b8cc667401c5bfb2a8364a57b21b95b6af94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 108, + 681, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/765b6d2321f0d93e25a9406b8d8994b7613ac292dc4423f35c35fdeebc84c074.jpg", + "image_caption": [ + "{}lamps" + ], + "image_footnote": [], + "bbox": [ + 681, + 107, + 735, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/180a81614f695a9d10633dc1bd3e2e071ed5ce9e975be81d420b52be09d7fe01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 107, + 823, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a450210894305ea521f3a9b8f81268d9ec90ce5a60e102f3e0ad8ba5bc765ad4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 825, + 108, + 879, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af9d53653afef40bc2ef1f6f8484ca738f5ea8277b792ff9f8b91f76dcfa6ec9.jpg", + "image_caption": [ + "\"six\"" + ], + "image_footnote": [], + "bbox": [ + 500, + 162, + 552, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/394468bffc7083cbb7d1b2dda123a7751ae18eca03374945b5a6c83026b3cf42.jpg", + "image_caption": [ + "\"seven\"" + ], + "image_footnote": [], + "bbox": [ + 553, + 162, + 676, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/be9fd6733a0a700645d1e2f720f0e9e5cc9539b8b034f9b7e0285521049cd0ad.jpg", + "image_caption": [ + "\"eight\"" + ], + "image_footnote": [], + "bbox": [ + 676, + 162, + 750, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/52030ed1d203acc98d76eee8221b8a8e2b112fea4a1e55d4373d67e830d5d514.jpg", + "image_caption": [ + "\"nine\"" + ], + "image_footnote": [], + "bbox": [ + 750, + 162, + 803, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aae2c2ac76cf3623d30f5c59e8730f7564d25830c9b608b9bcce0e398e7cb0d7.jpg", + "image_caption": [ + "\"ten\"" + ], + "image_footnote": [], + "bbox": [ + 803, + 162, + 880, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "primarily identifies a single instance of the described object.", + "bbox": [ + 75, + 287, + 467, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To verify that our model does not simply attend to all objects that appear in the image, we examined the relevancy maps in Fig. 6 using negative prompts (i.e. \"four\" when there are five elements in the image). Our model focuses only on relevant objects when the correct number is used, unlike the baseline CLIP that highlights all object types in the image. This demonstrates that our model learns to associate the number in the caption with the suitable number of objects, and does not exploit shortcuts or undesired content.", + "bbox": [ + 75, + 303, + 467, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Count-based text-to-image generation", + "text_level": 1, + "bbox": [ + 76, + 446, + 403, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text-to-image generative models that rely on CLIP for representations, are known to inherit its limitations [43, 52], thus fail to reliably produce specific counts of objects. In order to demonstrate the effectiveness of our counting-aware model for downstream image generation tasks, we train an Imagen [43] model conditioned on the textual embeddings of a pretrained CLIP-L/14, and another Imagen model conditioned on our counting-aware version of this model. To compare our model and the baseline, we synthesize 12 samples for each prompt in the counting category of the Draw-Bench benchmark [43]. For each sample, we check whether or not it contains the requested number of objects, as stated in its prompt.", + "bbox": [ + 75, + 470, + 467, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Imagen model trained with the baseline CLIP achieves a binary accuracy of $24.12\\%$ , while the Imagen trained with out counting-aware model reaches $40.35\\%$ accuracy. See the supplementary for additional evaluations.", + "bbox": [ + 75, + 667, + 467, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance on non-counting image generation: To verify that the generation quality is preserved for non-counting captions, we conducted a user study. We presented 52 participants with 20 pairs of images, each containing one image generated with an Imagen model trained with the baseline CLIP, and one with our counting-aware CLIP as backbone. The same random prompt from COCO [28] and seed were used for each pair. Participants were asked which image is better in terms of quality and realism. $54.7\\%$ of ratings favored our images, suggesting that our model does not degrade generation quality on general text captions.", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Limitations", + "text_level": 1, + "bbox": [ + 500, + 287, + 624, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method is limited by the insufficient existence of training data with images containing multiple instances of an object, along with a corresponding caption that correctly spells out this information. The effect of this data scarcity on our method increases with larger numbers (7, 8, etc.), as people tend to use \"a group of\" or \"many\" for large numbers of objects, instead of gruelingly counting them. Furthermore, many of the correct training pairs with higher numbers that do exist, contain relatively simplistic 2D collections of objects, as opposed to objects in a real-world scene. This is also the reason why our current method teaches CLIP to count only up to ten. Mitigating the above-mentioned data limitations, and generalizing to counting beyond 10, can possibly be achieved by concatenating multiple images with smaller numbers of objects. This will result in new training examples with more than ten natural-looking objects, and is part of our future work.", + "bbox": [ + 496, + 310, + 890, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 588, + 625, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present the first method to enhance VLMs with counting capabilities, while maintaining their overall performance on common benchmarks. This is an essential step towards enabling more accurate text-based image retrieval, classification and generation. A new counting-based contrastive loss is used to finetune a pre-trained VLM in tandem with its original objective. Our counting loss is deployed over automatically-created counterfactual examples, each consisting of an image and a caption containing an incorrect object count. We demonstrate a significant improvement over state-of-the-art baseline models in multiple tasks (classification, retrieval and generation) on several different datasets. Furthermore, we introduce CountBench - a new image-text counting benchmark for evaluating VLMs' object counting capabilities. This new benchmark contains in-the-wild images and captions, where the object counts are specified in the captions. Finally, our approach is not specific to counting, and can be generalized to other compositional concepts that VLMs fail to learn.", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3177", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In AAAI Conference on Artificial Intelligence, 2018. 3", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. ArXiv, abs/2204.14198, 2022. 2", + "[3] Jong Wook Kim Gretchen Krueger Sandhini Agarwal Alec Radford, Ilya Sutskever. Clip: Connecting text and images. https://openai.com/research/clip, 2021. 2, 3", + "[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18187-18197, 2022. 2", + "[5] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2", + "[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. ArXiv, abs/2204.02491, 2022. 2", + "[7] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 2", + "[8] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, 2014. 5", + "[9] Hila Chefer, Sagie Benaim, Roni Paiss, and Lior Wolf. Image-based clip-guided essence transfer. In ECCV, 2022. 2", + "[10] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 387-396, 2021. 7", + "[11] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiiovanni, Piotr Padlewski, Daniel M. Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model. ArXiv, abs/2209.06794, 2022. 2", + "[12] Katherine Crowson. Vqgan+clip. https://colab.research.google.com/drive/1L8oL-vLJXVcRzCFbPwOoMkPKJ8-aYdPN, 2021. 2" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 5", + "[14] Anuj Diwan, Layne Berry, Eunsol Choi, David F. Harwath, and Kyle Mahowald. Why is winoground hard? investigating failures in visuolinguistic compositionality. ArXiv, abs/2211.00768, 2022. 5", + "[15] Raveen Doon, Tarun Kumar Rawat, and Shweta Gautam. Cifar-10 classification using deep convolutional neural network. 2018 IEEE Punecon, pages 1–5, 2018. 5", + "[16] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. 2004 Conference on Computer Vision and Pattern Recognition Workshop, pages 178–178, 2004. 5", + "[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada. ACM Transactions on Graphics (TOG), 41:1 - 13, 2022. 2", + "[18] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arxiv:2307.10373, 2023. 2", + "[19] Patrick Helber, Benjamin Bischke, Andreas R. Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12:2217-2226, 2019. 5", + "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadayath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, October 2021. 5", + "[21] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. CVPR, 2021. 5", + "[22] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 4", + "[23] William Stanley Jevons. The power of numerical discrimination. Nature, 3:281-282, 1871. 3", + "[24] E. L. Kaufman and Miles W. Lord. The discrimination of visual number. The American journal of psychology, 62 4:498-525, 1949. 2, 3", + "[25] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2", + "[26] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5", + "[27] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. ArXiv, abs/2201.03546, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3178", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 8", + "[29] Xingchao Liu, Chengyue Gong, Lemeng Wu, Shujian Zhang, Haoran Su, and Qiang Liu. Fusedream: Training-free text-to-image generation with improved clip+gan space optimization. ArXiv, abs/2112.01573, 2021. 2", + "[30] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 2", + "[31] Duy-Kien Nguyen, Vedanuj Goswami, and Xinlei Chen. Movie: Revisiting modulated convolutions. 2021. 3", + "[32] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2", + "[33] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5", + "[34] Roni Paiss, Hila Chefer, and Lior Wolf. No token left behind: Explainability-aided image classification and generation. In ECCV, 2022. 3", + "[35] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2", + "[36] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, Mingxing Tan, and Quoc V. Le. Combined Scaling for Open-Vocabulary Image Classification. arXiv preprint arXiv:2111.10050, Nov. 2021. 2, 3, 6", + "[37] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6967-6977, June 2023. 2", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 6", + "[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 2", + "[42] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 5389-5400. PMLR, 09-15 Jun 2019. 5", + "[43] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, 2022. 2, 8", + "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 2", + "[45] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 4, 7", + "[46] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15617-15629, 2021. 2", + "[47] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022. 2", + "[48] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021. 2", + "[49] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Bachmann, Amit H. Bermano, Daniel Cohen-Or, Amir Roshan Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Trans. Graph., 41:86:1-86:11, 2022. 2", + "[50] Rijnder Wever and Tom F.H. Runia. Subitizing with variational autoencoders. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, September 2018. 3", + "[51] Yonghui Wu and David Fleet. How ai creates photorealistic images from text. https://blog.google/technology/research/how-ai-creates-photorealistic-images-from-text, 2022. 2", + "[52] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yin" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3179", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fei Yang, Burcu Karagol Ayan, Benton C. Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. ArXiv, abs/2206.10789, 2022. 2, 8", + "[53] Nir Zabari and Yedid Hoshen. Semantic segmentation in the-wild without seeing any segmentation examples. ArXiv, abs/2112.03185, 2021. 2", + "[54] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18123-18133, June 2022. 2", + "[55] Hui Zhang, Shenglong Zhou, Geoffrey Y. Li, and Naihua Xiu. 0/1 deep neural networks via block coordinate descent. ArXiv, abs/2206.09379, 2022. 5", + "[56] Jianming Zhang, Shugao Ma, Mehrnoosh Sameki, Stan Sclaroff, Margrit Betke, Zhe L. Lin, Xiaohui Shen, Brian L. Price, and Radomir Mech. Salient object subitizing. International Journal of Computer Vision, 124:169-186, 2015. 3", + "[57] Y. Zhang, Jonathon S. Hare, and Adam Prügel-Bennett. Learning to count objects in natural images for visual question answering. ArXiv, abs/1802.05766, 2018. 3", + "[58] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 530 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3180", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Teaching CLIP to Count to Ten/56d47f2c-86af-49f0-9e72-8decf5e81c8a_model.json b/2023/Teaching CLIP to Count to Ten/56d47f2c-86af-49f0-9e72-8decf5e81c8a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..97a6f1235db4a92388da43e66a7d1ae744950748 --- /dev/null +++ b/2023/Teaching CLIP to Count to Ten/56d47f2c-86af-49f0-9e72-8decf5e81c8a_model.json @@ -0,0 +1,2884 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.326, + 0.131, + 0.646, + 0.154 + ], + "angle": 0, + "content": "Teaching CLIP to Count to Ten" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.186, + 0.354, + 0.205 + ], + "angle": 0, + "content": "Roni Paiss\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.276, + 0.206, + 0.395, + 0.221 + ], + "angle": 0, + "content": "Inbar Mosseri1" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.187, + 0.481, + 0.204 + ], + "angle": 0, + "content": "Ariel Ephrat1" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.206, + 0.555, + 0.221 + ], + "angle": 0, + "content": "Michal Irani1,3" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.187, + 0.592, + 0.204 + ], + "angle": 0, + "content": "Omer Tov1" + }, + { + "type": "text", + "bbox": [ + 0.592, + 0.187, + 0.74, + 0.204 + ], + "angle": 0, + "content": "Shiran Zada" + }, + { + "type": "text", + "bbox": [ + 0.592, + 0.205, + 0.694, + 0.221 + ], + "angle": 0, + "content": "Tali Dekel1,3" + }, + { + "type": "text", + "bbox": [ + 0.593, + 0.223, + 0.845, + 0.239 + ], + "angle": 0, + "content": "3Weizmann Institute of Science" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.222, + 0.272, + 0.241 + ], + "angle": 0, + "content": "1Google Research" + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.25, + 0.439, + 0.267 + ], + "angle": 0, + "content": "(a) Image Retrieval Results (from LAION):" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.271, + 0.406, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.271, + 0.673, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.25, + 0.85, + 0.268 + ], + "angle": 0, + "content": "(b) Relevancy Maps:" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.283, + 0.853, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.497, + 0.895, + 0.604 + ], + "angle": 0, + "content": "Figure 1: Inability of CLIP to count. We show that CLIP is insensitive to the number of objects in an image, and demonstrate the effectiveness of our Counting-Aware CLIP. (a) Image retrieval results using text captions of the form “a photo of ”, with different numbers and types of objects. Images that match the caption are marked with \\(\\checkmark\\); images that do not are marked with \\(\\times\\). Our Counting-Aware model retrieves images that depict the requested number of objects, while the original CLIP often retrieves images that contain the wrong number of objects, or images where the number is explicitly written in the image (e.g. “nine hearts” - the image contains the number “9”, but has 11 hearts). (b) Attention maps demonstrating that our model attends to all matching object instances in the image, unlike the original CLIP." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.616, + 0.314, + 0.631 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Large vision-language models (VLMs), such as CLIP, learn rich joint image-text representations, facilitating advances in numerous downstream tasks, including zero-shot classification and text-to-image generation. Nevertheless, existing VLMs exhibit a prominent well-documented limitation - they fail to encapsulate compositional concepts such as counting. We introduce a simple yet effective method to improve the quantitative understanding of VLMs, while maintaining their overall performance on common benchmarks. Specifically, we propose a new counting-contrastive loss used to finetune a pre-trained VLM in tandem with its original objective. Our counting loss is deployed over automatically-created counterfactual examples, each consisting of an image and a caption containing an incorrect object count. For example, an image depicting three dogs is paired with the caption \"Six dogs playing in the yard\" as a negative example. Our loss encourages discrimination" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.895, + 0.785 + ], + "angle": 0, + "content": "between the correct caption and its counterfactual variant which serves as a hard negative example. To the best of our knowledge, this work is the first to extend CLIP's capabilities to object counting. Furthermore, we introduce \"CountBench\" - a new image-text counting benchmark for evaluating object counting capabilities. We demonstrate a significant improvement over state-of-the-art baseline models on this task. Finally, we leverage our counting-aware CLIP model for image retrieval and text-conditioned image generation, demonstrating that our model can produce specific counts of objects more reliably than existing ones." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.8, + 0.633, + 0.816 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Since the advent of CLIP [38], training large vision-language models (VLMs) has become a prominent paradigm for representation learning in computer vision. By observing huge corpora of paired images and captions crawled from the Web, these models learn powerful" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3170" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.227 + ], + "angle": 0, + "content": "and rich joint image-text embedding spaces, which have been employed in numerous visual tasks, including classification [58, 59], segmentation [27, 53], motion generation [47], image captioning [30, 48], text-to-image generation [12, 29, 32, 41, 44] and image or video editing [4, 6, 9, 17, 25, 35, 49, 18, 7]. Recently, VLMs have also been a key component in text-to-image generative models [5, 39, 41, 43], which rely on their textual representations to encapsulate the semantic meaning of the input text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.468, + 0.381 + ], + "angle": 0, + "content": "Despite their power, prominent VLMs, such as CLIP [38] and BASIC [36], are known to possess a weak understanding of the number of objects present in an image [3, 36, 38]. This is demonstrated in Fig. 1, where, when given a caption of the template \"a photo of \", CLIP often fails to retrieve images that correctly match the described number. Downstream applications that rely on VLM-based representations inherit these limitations, e.g., image generation models struggle to reliably produce specific counts of objects [43, 51, 52]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.382, + 0.468, + 0.442 + ], + "angle": 0, + "content": "In this work, we introduce a novel method that enhances the quantitative understanding of large-scale VLMs by encouraging them to produce representations that are sensitive to the number of objects in the image and text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.468, + 0.717 + ], + "angle": 0, + "content": "We hypothesize that the reason existing VLMs fail to learn the concept of counting is severalfold: (i) Captions that accurately specify the number of objects become extremely rare in the data as the number of objects increases. For example, we found that for more than five objects, captions typically contain a general form of quantity, e.g., \"a group of...\" or \"many...\", rather than an accurate count. This can be attributed to the fact that people cannot instantly identify numerical quantities larger than four without explicitly counting the objects [24]. (ii) Numbers in the caption often refer to attributes that are NOT related to counting - e.g. age, time, address, temperature, model number (\"This is an iPhone 5\"), etc. (iii) The task of counting (associating the visible number of objects in an image with the number in the caption), is not explicitly enforced in current VLM training objectives. Therefore, current VLMs are able to count reasonably well only up to two or three (for which there are sufficient image-caption examples)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.468, + 0.808 + ], + "angle": 0, + "content": "We thus suggest to mitigate each of these problems by: (i) Creating suitable training data in which the captions contain accurate numbers of objects. (ii) Designing a training objective whereby understanding object counts is critical for discriminating between the correctly associated caption and incorrect ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.468, + 0.901 + ], + "angle": 0, + "content": "More specifically, as illustrated in Fig. 2, we automatically create a clean and diverse counting training set by curating image-text examples where the image depicts multiple objects and its caption is verified to express their count. We then finetune a pretrained VLM by formulating counting as a discriminative task – for each example, we create" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "a counterfactual caption by swapping the spelled number associated with the object count with a different randomly selected number. The model's objective is then to associate the image correctly with its true count caption, discriminating it from the counterfactual one." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.469 + ], + "angle": 0, + "content": "To evaluate counting capabilities, we introduce \"Count-Bench\" - a carefully curated object counting benchmark, consisting of 540 diverse, high quality image-text examples. We evaluate our method on two prominent contrastive VLMs: CLIP [38] and BASIC [36], and demonstrate a significant improvement in accuracy in the task of zero-shot count classification over baseline models. Importantly, we achieve this while maintaining the original knowledge learned by the VLM, as demonstrated by an extensive evaluation of our model on standard zero-shot downstream tasks. The quantitative understanding of our model is further evident by our text-to-image retrieval results (e.g., Fig. 1(a)), as well as by the relevancy maps of our model, which demonstrate that the model correctly attends to all visible objects whose count is specified in the text (e.g., Fig. 1(b)). Finally, we train a large-scale text-to-image generative model [43] which incorporates our counting training set and counting-aware CLIP text encoder. The generated images from this model exhibit higher fidelity to the number of objects specified in the input prompts (Fig. 8)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.47, + 0.782, + 0.483 + ], + "angle": 0, + "content": "To summarize, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.533 + ], + "angle": 0, + "content": "1. A novel training framework for tackling the task of vision-language counting - an important limitation of current VLMs." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.537, + 0.891, + 0.566 + ], + "angle": 0, + "content": "2. A new benchmark, \"CountBench\", carefully filtered and validated for evaluating VLMs on the counting task." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.569, + 0.891, + 0.627 + ], + "angle": 0, + "content": "3. We apply our method to the widely-adopted VLMs CLIP and BASIC, demonstrating significant improvement on the counting task, while maintaining general (non-counting) performance on common benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.63, + 0.891, + 0.69 + ], + "angle": 0, + "content": "4. We utilize our counting-aware VLMs for downstream tasks including image retrieval and text-to-image generation, demonstrating more reliable results when the text prompt contains a specific number of objects." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.637, + 0.725 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Contrastive vision-language models: Vision-language models have demonstrated impressive success in vision and multimodal tasks [2, 11, 36, 38, 46, 54, 37]. In this work, we focus on contrastive VLMs, such as CLIP [38] and BASIC [36], as they are widely used both for downstream tasks and as backbones for generative models [40, 43]. Both are trained with a contrastive objective, where matching text-image pairs should have a low cosine distance, and non-matching texts and images should be far apart. The representations computed by CLIP have proven to be very effective in vision and multimodal tasks, due to their zero-shot" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3171" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.068, + 0.091, + 0.333, + 0.107 + ], + "angle": 0, + "content": "(a) Obtaining Counting Training Set:" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.109, + 0.472, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.069, + 0.207, + 0.269, + 0.222 + ], + "angle": 0, + "content": "(b) Teaching CLIP to count:" + }, + { + "type": "image", + "bbox": [ + 0.075, + 0.223, + 0.48, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.457, + 0.473, + 0.699 + ], + "angle": 0, + "content": "Figure 2: Method overview (a) We create a text-image counting training set in which each caption expresses the number of objects depicted in the corresponding image. This is done by using an off-the-shelf object detector to automatically identify text-image examples in which the text count matches the number of visible objects in the image (see Sec. 3.1). (b) We finetune a pre-trained CLIP model using our counting subset (a), through a dedicated contrastive objective \\( L_{count} \\), used in addition to the original (general) contrastive objective (\\( L_{clip} \\)). Specifically, given a text-image example from our counting subset, we automatically create a counterfactual prompt by replacing the true object count in the original caption with an incorrect count; \\( L_{count} \\) encourages the model to embed the image close to its original caption embedding (expressing the true object count) and far from its counterfactual count (see Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.793 + ], + "angle": 0, + "content": "capabilities and semantic nature, and have been used as a prominent component in numerous tasks and methods. BASIC [36] scaled up the size of the model, batch size and dataset, improving zero-shot accuracy on common benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Subitizing: \"Subitizing\" [24] is the ability of people to instantly recognize the number of objects at a glance - without actually counting them. It has been shown [23] that most humans can subitize only a small number of objects (up to \\(\\sim 4\\)). This cognitive phenomenon is reflected in the numbers people tend to specify in captions (\\(\\leqslant 4\\)), and sub-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.168 + ], + "angle": 0, + "content": "sequentially in the counting abilities of VLMs, which struggle to count larger numbers. Subitizing was also referred to in computer vision [50, 56]. These papers aim to predict a very small number (\\( \\leqslant 4 \\)) of salient objects in an image, and are pure image-based (do not involve any text or VLMs)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.186, + 0.892, + 0.307 + ], + "angle": 0, + "content": "Counting in vision-language models: While demonstrating impressive recognition capabilities, large VLMs such as CLIP[38] and BASIC [36] are known to only partially capture the meaning of the text. Thus, they fail to understand the number of objects in an image [3, 36, 38]. Paiss et al. [34] demonstrated that CLIP attends to only a small subset of its input, mainly the nouns, and often ignores adjectives, numbers and prepositions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Counting has remained a stand-alone task under the domain of visual question answering (VQA), tackled with specifically designed architectures and techniques [1, 31, 57]. Our work defers from these prior efforts in several key aspects: (i) While previous efforts are restricted to VQA architectures and problem formulation, our goal is to improve the quantitative understanding of general-purpose contrastive VLMs (e.g., CLIP and BASIC), used in various vision and multimodal tasks where counting-aware solutions are not currently available. (ii) Our work can enhance the zero-shot counting capabilities of VLMs to unrestricted objects, unlike prior methods that are trained on specific domains, which can be problematic for new domains where no counting labels are available." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.539, + 0.59, + 0.555 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Our goal is to teach a pre-trained VLM (e.g., CLIP) to count, i.e., to improve its quantitative textual and visual understanding. Our framework, illustrated in Fig. 2, consists of two main stages. We first automatically create a counting training set, comprising clean and diverse images along with corresponding captions that describe the number of visible objects in the scene. We then leverage this dataset to finetune the VLM through a designated count-based contrastive loss that is used in tandem with the original generic image-text objective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "More specifically, our key idea is to automatically generate counterfactual examples by swapping the true object count in the caption with a different random number. Our new counting loss encourages the model to embed an image close to its true count, as expressed by the original caption, while pushing it away from the embedding of the counterfactual count prompt. As the only difference between the correct caption and its counterfactual counterpart is a single word—the spelled number of objects—the model has to distinguish between the correct and incorrect count in order to succeed in its training task. Next, we describe our dataset creation and finetuning paradigm in detail." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3172" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.43, + 0.108 + ], + "angle": 0, + "content": "3.1. Creating an image-text counting train set" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.473, + 0.293 + ], + "angle": 0, + "content": "A naive approach for obtaining an image-text counting dataset is to filter a large-scale dataset for examples in which the caption contains a number. However, this approach results in a highly noisy dataset, since the numbers in the captions often refer to other numerical attributes that are unrelated to object counts. Such numerical attributes are age (\"A 7 year old girl\"), time (\"The time is 9 o'clock\"), addresses, model numbers, etc (see examples in supplementary). To ensure that the numbers in the captions indeed specify object counts, we employ several stages of automatic filtering in our data pipeline (Fig. 2 (a)):" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.47, + 0.395 + ], + "angle": 0, + "content": "First, we filter out all examples whose caption does not contain a spelled number \\(\\in \\{\\text{\"two\",\\ldots,\"ten\"}\\}\\). We do so, as we observed that non-spelled numbers, or numbers higher than ten, mostly appear in conjunction with a measure of time, (e.g. dates) or addresses, rather than numbers of objects present in the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.473, + 0.603 + ], + "angle": 0, + "content": "In the second stage, we verify that the spelled numbers serve as object counters, and that the counted objects are visible and detectable in the images. For example, for the caption \"A photo of three dogs\", we verify that the image indeed depicts three visible dogs, no more, and no less. This count verification is achieved automatically by first applying an off-the-shelf object detector [22], and counting the number of detections per object. We assume that the caption refers to the most prevalent object in the image. Thus, we retain only examples for which the number specified in the caption aligns with the number of instances of the maximally-detected object. We denote by \\( C \\) our automatically filtered train set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Naturally, the filtered data \\( C \\) is imbalanced. The number of examples that pass our filtering drops significantly as the count increases, e.g., the number of \"ten\" image-text pairs is around \\( 1000 \\times \\) smaller than \"two\". Training with such imbalanced data creates a bias—the loss can be reduced by classifying frequent numbers as the correct caption and rare numbers as counterfactual, regardless of the image content. Therefore, balancing the data is of essence. Due to scarcity of examples depicting more than six objects, we choose to balance the numbers \"two\" – \"six\" separately from the higher numbers \"seven\" – \"ten\". For each of the numbers \"two\" – \"six\", we sample around \\( 37K \\) samples, while for \"seven\" – \"ten\", we use all the samples passed by our filter. There are approximately \\( 7K \\) samples for \"seven\" down to around \\( 1.5K \\) samples for \"ten\". We found this approach to provide us with a diverse and relatively balanced training dataset, yet more sophisticated methods could be considered in the future. From this point on, \\( C \\) will denote our filtered and balanced numbered training set." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.722, + 0.108 + ], + "angle": 0, + "content": "3.2. Teaching CLIP to count" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.114, + 0.894, + 0.189 + ], + "angle": 0, + "content": "Our goal is to improve the quantitative understanding of a pre-trained VLM (e.g., CLIP), while preserving its real-world knowledge, as reflected by its zero-shot capabilities on commonly-evaluated benchmarks. Therefore, we use a combination of two loss functions:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.202, + 0.893, + 0.218 + ], + "angle": 0, + "content": "\\[\nL = L _ {C L I P} + \\lambda L _ {\\text {c o u n t}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.273 + ], + "angle": 0, + "content": "where \\( L_{CLIP} \\) is the regular contrastive loss of CLIP, \\( L_{count} \\) is our counting-designated loss (described below), and \\( \\lambda \\) is a hyperparameter used to weight the two losses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.893, + 0.486 + ], + "angle": 0, + "content": "We finetune the model on two training sets: (i) A very large dataset collected from the Web that contains general in-the-wild images and captions. (ii) Our filtered numbered training set \\(C\\), described in Sec. 3.1, which contains samples where object counts are spelled out in the captions. While \\(L_{CLIP}\\) is calculated on all samples, the counting loss \\(L_{count}\\) is calculated only on samples from \\(C\\). For each image-text pair \\((i_k, t_k) \\in C\\), a counterfactual caption \\(t_k^{CF}\\) is automatically created by swapping the number in the caption \\(t_k\\) with a different random number (e.g., the caption \"five dogs\" can be counterfactualized with \"eight dogs\"). At each training step, the triplets \\((i_k, t_k, t_k^{CF})_{k=1}^N\\) are then fed to CLIP's text and image encoders to obtain their embeddings \\((ei_k, et_k, et_k^{CF})_{k=1}^N\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.894, + 0.531 + ], + "angle": 0, + "content": "Then, a contrastive loss \\( L_{count} \\) is computed to enforce that the similarity score of the image is high with the original caption and low with the counterfactual caption:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.553, + 0.892, + 0.605 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c o u n t}} = - \\frac {1}{N} \\sum_ {k = 1} ^ {N} \\log \\frac {\\exp \\left(e i _ {k} \\cdot e t _ {k}\\right)}{\\exp \\left(e i _ {k} \\cdot e t _ {k}\\right) + \\exp \\left(e i _ {k} \\cdot e t _ {k} ^ {C F}\\right)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.681 + ], + "angle": 0, + "content": "Since the original ground truth caption and counterfactual caption differ only by the number of objects specified in them, this loss encourages the model to learn the relationship between the specified spelled number and the number of the objects it refers to." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.892, + 0.727 + ], + "angle": 0, + "content": "We use the negative samples only in the counting objective \\( L_{count} \\), instead of adding them to the batch for the contrastive loss \\( L_{CLIP} \\) in order to better weight their effect." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.74, + 0.882, + 0.756 + ], + "angle": 0, + "content": "4. \"CountBench\" - an evaluation benchmark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.902 + ], + "angle": 0, + "content": "We introduce a new object-counting evaluation benchmark called CountBench. This benchmark was automatically curated (and manually verified) from the publicly available LAION-400M image-text dataset [45]. CountBench contains a total of 540 images containing between 2 and 10 instances of a particular object, where their corresponding captions reflect this number. This benchmark is used only for testing and evaluation. CountBench has no overlap with our training set \\( C \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3173" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.293, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.254, + 0.273, + 0.266 + ], + "angle": 0, + "content": "(a) Baseline CLIP (L/14)" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.088, + 0.473, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.254, + 0.428, + 0.266 + ], + "angle": 0, + "content": "(b) Ours (L/14)" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.274, + 0.47, + 0.351 + ], + "angle": 0, + "content": "Figure 3: Confusion matrices on CountBench. Classification accuracy on CountBench, broken down into confusion matrices for the public CLIP ViT-L/14 (a), and our counting-aware CLIP ViT-L/14 model (b), demonstrating clear quantitative superiority of our model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.538 + ], + "angle": 0, + "content": "The images in CountBench were obtained by running our automatic filtering method described in Sec. 3.1 on the entire LAION-400M dataset. The filtering pipeline includes applying an object detector, whose benefits are twofold: (i) it facilitates verification of the object count specified in the caption, and (ii) it ensures that the counted objects are visible and detectable. The latter is important since benchmarks that do not disentangle the problem at hand from other challenges (e.g. occlusions and low resolution) may not reflect the performance on the actual task, since the difficulty can often stem from unrelated factors [14]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.706 + ], + "angle": 0, + "content": "This filtering produced over 158K images for the number \"two\", but only around 100 images for \"ten\", demonstrating again the severe number imbalance we encountered with our training set. After automatically balancing each number to 100-200 samples each, the entire dataset was manually verified to contain only pairs in which the spelled number in the caption matches the number of clearly visible objects in the image. The dataset was rebalanced after this stage, ending up with 60 image-text pairs per number \\(\\in\\{\\text{\"two\",\"..},\\text{\"ten\"}\\}\\), 540 in total. Samples from the dataset can be seen in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.47, + 0.751 + ], + "angle": 0, + "content": "We use the CountBench benchmark to evaluate the counting abilities of the models trained with our method in Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.21, + 0.787 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We thoroughly evaluate our method, both quantitatively and qualitatively, on object counting-related tasks using our CountBench benchmark. We further validate that the performance of our counting-aware models on a variety of general zero-shot classification benchmarks is retained [8, 13, 15, 16, 19, 20, 21, 26, 33, 42, 55]. To gain a better understanding of our models, we show visualizations" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.607, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.162, + 0.594, + 0.181 + ], + "angle": 0, + "content": "\"two zebras in Cape Town\"" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.09, + 0.748, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.162, + 0.74, + 0.181 + ], + "angle": 0, + "content": "\"three purebred chihuahuaas running...\"" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.09, + 0.892, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.755, + 0.162, + 0.887, + 0.181 + ], + "angle": 0, + "content": "\"...four antique pink and purple armchairs\"" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.183, + 0.631, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.256, + 0.62, + 0.275 + ], + "angle": 0, + "content": "\"GT's five favourite Champagne...\"" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.183, + 0.777, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.256, + 0.771, + 0.275 + ], + "angle": 0, + "content": "\"vintage silver plate tablespoons...set of six\"" + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.183, + 0.892, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.256, + 0.891, + 0.275 + ], + "angle": 0, + "content": "\"seven Chewy Peanut Butter Cookies...\"" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.277, + 0.633, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.351, + 0.63, + 0.372 + ], + "angle": 0, + "content": "...eight colorful bright shiny red apples..." + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.277, + 0.779, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.351, + 0.765, + 0.371 + ], + "angle": 0, + "content": "\"The view of the nine leftmost moai...\"" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.277, + 0.892, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.351, + 0.887, + 0.37 + ], + "angle": 0, + "content": "\"All ten Christmas Poinsettia...\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.384, + 0.893, + 0.459 + ], + "angle": 0, + "content": "Figure 4: CountBench benchmark. Sample images and their corresponding captions from our new CountBench object counting benchmark. This benchmark was automatically curated (and manually verified) from the publicly-available LAION-400M dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.572 + ], + "angle": 0, + "content": "of text-image relevancy maps, along with per-word relevancy scores, demonstrating that our model indeed attends to the number of objects in the image and text. Finally, we apply our model to text-to-image retrieval and generation, producing specific numbers of objects more reliably than baseline models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.588, + 0.756, + 0.604 + ], + "angle": 0, + "content": "5.1. Zero-shot counting accuracy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.748 + ], + "angle": 0, + "content": "We evaluate our models and baselines on CountBench on the task of classifying the number of objects in an image in a zero-shot manner. For each image in CountBench we augment the existing caption with eight other possible captions by replacing the number in its caption with all the numbers \\(\\in \\{\\text{two}, \\ldots, \\text{ten}\\}\\), and calculate the similarity score between the image and each of the nine captions. The number in the caption that obtains highest similarity score with the image is considered the predicted number." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Table 1 reports the results of this evaluation on two prominent contrastive VLMs: CLIP-B/32 and BASIC-S. We report both the counting accuracy (selection of the correct number) and the mean deviation of the models' predictions from the correct numbers. For each of the architectures, we compare our model (configuration \\( E \\)) with two baseline configurations: (A) the official baseline model, and (B) the baseline model finetuned on our general text-image dataset used in our implementation, with the standard contrastive loss. Comparing the performance of these config" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3174" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.089, + 0.089, + 0.885, + 0.189 + ], + "angle": 0, + "content": "
CLIP-B/32BASIC-S
A Official BaselineB Internal BaselineC Ours (w/o Lcount)D Ours (Naive Filtering)E OursA Official BaselineB Internal BaselineC Ours (w/o Lcount)D Ours (Naive Filtering)E Ours
Accuracy ↑31.6732.9444.2649.8175.9317.9722.7530.5928.6869.02
Mean deviation from the correct number ↓1.531.440.971.280.492.132.021.291.870.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.195, + 0.895, + 0.271 + ], + "angle": 0, + "content": "Table 1: Quantitative counting results. Top-1 zero-shot accuracy and the mean absolute distance between the predicted numbers and the true numbers on CountBench. We compare several configurations: (A) The official CLIP [38] and BASIC [36] models. (B) The official baselines finetuned on our internal curated data. (C) Models trained with our filtered counting set, without \\( L_{count} \\) (D) Models finetuned with \\( L_{count} \\) on a naively filtered counting set (E) Our method, which is significantly superior to all other configurations, both in accuracy and deviation from correct number." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.291, + 0.462, + 0.445 + ], + "angle": 0, + "content": "
CLIP-B/32BASIC-S
Official BaselineInternal BaselineOursOfficial BaselineInternal BaselineOurs
ImageNet62.9364.9764.0659.7061.9661.18
CIFAR1063.9161.0060.6576.2284.6984.05
CIFAR10033.1032.4933.5645.3556.8055.89
Caltech10175.9982.5082.3678.1681.0381.05
EuroSAT45.2341.6637.6928.3945.8245.97
Food10183.0880.7280.5377.0877.8077.06
ImageNetA31.8530.8529.8117.6522.5521.68
ImageNetR69.3870.1770.3067.1167.6866.95
ImageNetV255.6556.5656.6252.2254.3553.60
Oxford Pets87.3587.7487.4180.6285.1584.87
Oxford Flowers66.1465.7367.3964.7466.4065.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.51 + ], + "angle": 0, + "content": "Table 2: Zero-shot accuracy on common benchmarks. We compare the zero-shot accuracy of our method and baselines on a variety of popular benchmarks. As seen, our method preserves the performance of the original model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.47, + 0.577 + ], + "angle": 0, + "content": "urations allows us to quantify the effect of using our own large-scale text-image dataset, which differs from the original unpublished data the models were trained on." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.657 + ], + "angle": 0, + "content": "As can be seen, our method \\((E)\\) achieves significantly superior counting accuracy compared to the baselines \\((A, B)\\), attaining \\(2 - 3 \\times\\) higher counting accuracy and more than \\(3 \\times\\) lower mean deviation from the correct number. Results for different categories in CountBench are reported in App. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Tab. 1 also contains an ablation of the two components of our method: filtering a counting training set and finetuning with an additional loss \\( L_{count} \\). Models with configuration \\( C \\) are finetuned on the filtered subset with no counting loss. The large gap in accuracy on CountBench between configurations \\( C \\) and \\( E \\) shows the importance of our loss for the improvement in counting abilities. Models with configuration \\( D \\) are finetuned with the counting loss \\( L_{count} \\) on an alternative counting subset, which consists of all the samples that contain spelled numbers \\( \\in \\{ \"two\",.., \"ten\" \\} \\) without additional filtering. The significant difference in counting accuracy between configurations \\( D \\) and \\( E \\) demonstrates the importance of our restrictive filtering pipeline, as both configurations are finetuned with \\( L_{count} \\) over the samples from a dedicated counting training set. As can be seen in Tab. 1, while the naively filtered data does improve performance" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.286, + 0.636, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.286, + 0.752, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.286, + 0.868, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.352, + 0.733, + 0.369 + ], + "angle": 0, + "content": "\"Two ducks\" 0.23 0.77" + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.352, + 0.851, + 0.369 + ], + "angle": 0, + "content": "\"Two ducks\" 0.45 0.55" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.373, + 0.636, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.373, + 0.752, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.373, + 0.868, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.428, + 0.75, + 0.445 + ], + "angle": 0, + "content": "\"Three pumpkins\" 0.16 0.84" + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.428, + 0.865, + 0.445 + ], + "angle": 0, + "content": "\"Three pumpkins\" 0.45 0.55" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.449, + 0.636, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.449, + 0.752, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.449, + 0.868, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.502, + 0.73, + 0.52 + ], + "angle": 0, + "content": "Four cards 0.30 0.70" + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.502, + 0.852, + 0.52 + ], + "angle": 0, + "content": "\"Four cards\" 0.43 0.57" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.523, + 0.894, + 0.629 + ], + "angle": 0, + "content": "Figure 5: Relevancy maps of both image and text. Visualization of the relevancy maps of both image and text, which represent, for each patch in the image and token in the text, how important it is to the prediction. Using our counting-aware CLIP model, the relevancy of the number (e.g., \"four\") in the text is increased. In addition, the model focuses on areas in the image that are relevant for counting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.734 + ], + "angle": 0, + "content": "over a baseline trained without a dedicated counting subset, the obtained results are still significantly lower than those produced by our model. We attribute this gap in performance to mislabeled training samples in the naively filtered data, which are absent from our counting training set \\( C \\) due to our filtering pipeline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Confusion matrices for the counting evaluation described above are shown in Fig. 3. For this experiment, we compare a public CLIP-L/14 model against our counting-aware version of it. As can be seen, our counting-aware CLIP model is significantly superior to the baseline across all numbers. It is worth noting that the baseline CLIP performs reasonably well for numbers within the subitizing range \\((2 - 4)\\), but fails on larger numbers. This supports our claim that curated image captions with numbers beyond this range are rare, since specifying them requires manually counting the objects." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3175" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.082, + 0.468, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.195, + 0.472, + 0.287 + ], + "angle": 0, + "content": "Figure 6: Relevancy maps for similarity between the image and different numbers. We compare the relevancy map of the input image with text prompts of several numbers (i.e. four to six) for both the baseline CLIP and our counting-aware CLIP. Our model focuses on the five stars when calculating the similarity with the prompt \"five\"." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.301, + 0.333, + 0.318 + ], + "angle": 0, + "content": "5.2. Count-based image retrieval" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.47, + 0.445 + ], + "angle": 0, + "content": "We consider the task of text-to-image retrieval where the text explicitly describes the desired count of objects. To obtain a diverse dataset that consists of varied numbers of objects, yet facilitates retrieval in reasonable time, we split the public LAION-400M dataset [45] into coarse categorical subsets by filtering samples where the caption contains a certain word (e.g., \"dogs\", \"animals\", \"cars\"), and perform retrieval on each of these subsets separately." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.473, + 0.719 + ], + "angle": 0, + "content": "For each category, we use the caption \"a photo of \" where number \\(\\in\\) {\"two\",..,\"ten\"} (e.g. \"a photo of six dogs\"). For each caption, we retrieve the three images in the dataset that are predicted by the model to be most similar to the caption. Note that since there are no ground truth labels for the counts of objects, we present qualitative results. Fig. 7 shows the retrieved images using the original CLIP model and our counting-aware CLIP model. An extended figure can be found in the supplementary. As can be seen, when the requested number is larger than three, the images retrieved by the baseline model often depict arbitrary numbers of objects. Additionally, the baseline often retrieves the same images for several different requested numbers. This further implies that the baseline model mostly focuses on the existence of the described object in the image, and ignores the number in the caption. In contrast, our results depict accurate object counts in most cases." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.446, + 0.743 + ], + "angle": 0, + "content": "5.3. Performance on non-counting classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To verify that our counting-aware models preserve the powerful image-text representation capabilities of the original models, we evaluate the zero-shot performance of our models on a variety of common classification benchmarks. Table 2 reports the zero-shot accuracy of our counting aware models against the baselines (corresponding to configurations \\( A \\), \\( B \\) in Tab. 1). As can be seen, our models maintain similar overall accuracy. Additionally, comparing the official baseline and the internal baseline indicates that finetuning the models on our general text-image datasets leads to" + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.082, + 0.788, + 0.098 + ], + "angle": 0, + "content": "\"A photo of {} shirts.\"" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.103, + 0.891, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.512, + 0.894, + 0.603 + ], + "angle": 0, + "content": "Figure 7: Top-3 count-based image retrieval Text-to-image retrieval results for different counts of shirts, ordered from left to right according to their similarity score (descending). Images that match the caption are marked with \\(\\checkmark\\) and images that do not match it are marked with X. Extended results are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.893, + 0.668 + ], + "angle": 0, + "content": "only a slight shift in the accuracy of the models on common benchmarks. An additional evaluation on the task of general zero-shot image retrieval is reported in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.756, + 0.697 + ], + "angle": 0, + "content": "5.4. Relevancy map visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To gain a better understanding of what our model learns, we use the explainability method of Chefer et al. [10] to visualize relevancy maps, which indicate the importance of different parts of the text and image to the model's similarity score prediction. Fig. 5 displays the normalized relevancy maps of several image-text pairs. Examining the relevancy maps of the text, it is apparent that the relevancy score of the spelled number in the caption is significantly higher for our model than the baseline model, which suggests that our model concentrates more on the mentioned number than the original one. Additionally, examining the relevancy maps of the images, it is evident that our model focuses on all pertinent objects in the image, whereas the original model" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3176" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.089, + 0.368, + 0.107 + ], + "angle": 0, + "content": "\"{} colorful tulips\"" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.107, + 0.171, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.107, + 0.258, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.107, + 0.344, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.107, + 0.4, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.107, + 0.482, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.164, + 0.149, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.206, + 0.13, + 0.214 + ], + "angle": 0, + "content": "\"six\"" + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.164, + 0.233, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.206, + 0.213, + 0.215 + ], + "angle": 0, + "content": "\"seven\"" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.163, + 0.303, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.206, + 0.293, + 0.215 + ], + "angle": 0, + "content": "\"eight\"" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.163, + 0.395, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.206, + 0.366, + 0.215 + ], + "angle": 0, + "content": "\"nine\"" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.163, + 0.483, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.206, + 0.457, + 0.215 + ], + "angle": 0, + "content": "\"ten\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.092, + 0.743, + 0.108 + ], + "angle": 0, + "content": "{}lamps" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.108, + 0.555, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.109, + 0.608, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.109, + 0.683, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.108, + 0.736, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.108, + 0.825, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.826, + 0.109, + 0.88, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.164, + 0.553, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.206, + 0.542, + 0.215 + ], + "angle": 0, + "content": "\"six\"" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.164, + 0.677, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.206, + 0.642, + 0.215 + ], + "angle": 0, + "content": "\"seven\"" + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.164, + 0.75, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.206, + 0.734, + 0.215 + ], + "angle": 0, + "content": "\"eight\"" + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.164, + 0.804, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.206, + 0.796, + 0.215 + ], + "angle": 0, + "content": "\"nine\"" + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.164, + 0.882, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.824, + 0.206, + 0.856, + 0.215 + ], + "angle": 0, + "content": "\"ten\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.074, + 0.223, + 0.894, + 0.269 + ], + "angle": 0, + "content": "Figure 8: Generated samples with Imagen using our counting-aware CLIP as backbone. The Imagen model benefits from the counting-aware representations produced by the our CLIP model, and is able to generate images that accurately follow the amounts specifies in the captions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.468, + 0.303 + ], + "angle": 0, + "content": "primarily identifies a single instance of the described object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.468, + 0.439 + ], + "angle": 0, + "content": "To verify that our model does not simply attend to all objects that appear in the image, we examined the relevancy maps in Fig. 6 using negative prompts (i.e. \"four\" when there are five elements in the image). Our model focuses only on relevant objects when the correct number is used, unlike the baseline CLIP that highlights all object types in the image. This demonstrates that our model learns to associate the number in the caption with the suitable number of objects, and does not exploit shortcuts or undesired content." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.448, + 0.404, + 0.464 + ], + "angle": 0, + "content": "5.5. Count-based text-to-image generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.468, + 0.667 + ], + "angle": 0, + "content": "Text-to-image generative models that rely on CLIP for representations, are known to inherit its limitations [43, 52], thus fail to reliably produce specific counts of objects. In order to demonstrate the effectiveness of our counting-aware model for downstream image generation tasks, we train an Imagen [43] model conditioned on the textual embeddings of a pretrained CLIP-L/14, and another Imagen model conditioned on our counting-aware version of this model. To compare our model and the baseline, we synthesize 12 samples for each prompt in the counting category of the Draw-Bench benchmark [43]. For each sample, we check whether or not it contains the requested number of objects, as stated in its prompt." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.468, + 0.727 + ], + "angle": 0, + "content": "The Imagen model trained with the baseline CLIP achieves a binary accuracy of \\(24.12\\%\\), while the Imagen trained with out counting-aware model reaches \\(40.35\\%\\) accuracy. See the supplementary for additional evaluations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Performance on non-counting image generation: To verify that the generation quality is preserved for non-counting captions, we conducted a user study. We presented 52 participants with 20 pairs of images, each containing one image generated with an Imagen model trained with the baseline CLIP, and one with our counting-aware CLIP as backbone. The same random prompt from COCO [28] and seed were used for each pair. Participants were asked which image is better in terms of quality and realism. \\(54.7\\%\\) of ratings favored our images, suggesting that our model does not degrade generation quality on general text captions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.288, + 0.625, + 0.302 + ], + "angle": 0, + "content": "5.6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.567 + ], + "angle": 0, + "content": "Our method is limited by the insufficient existence of training data with images containing multiple instances of an object, along with a corresponding caption that correctly spells out this information. The effect of this data scarcity on our method increases with larger numbers (7, 8, etc.), as people tend to use \"a group of\" or \"many\" for large numbers of objects, instead of gruelingly counting them. Furthermore, many of the correct training pairs with higher numbers that do exist, contain relatively simplistic 2D collections of objects, as opposed to objects in a real-world scene. This is also the reason why our current method teaches CLIP to count only up to ten. Mitigating the above-mentioned data limitations, and generalizing to counting beyond 10, can possibly be achieved by concatenating multiple images with smaller numbers of objects. This will result in new training examples with more than ten natural-looking objects, and is part of our future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.627, + 0.604 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We present the first method to enhance VLMs with counting capabilities, while maintaining their overall performance on common benchmarks. This is an essential step towards enabling more accurate text-based image retrieval, classification and generation. A new counting-based contrastive loss is used to finetune a pre-trained VLM in tandem with its original objective. Our counting loss is deployed over automatically-created counterfactual examples, each consisting of an image and a caption containing an incorrect object count. We demonstrate a significant improvement over state-of-the-art baseline models in multiple tasks (classification, retrieval and generation) on several different datasets. Furthermore, we introduce CountBench - a new image-text counting benchmark for evaluating VLMs' object counting capabilities. This new benchmark contains in-the-wild images and captions, where the object counts are specified in the captions. Finally, our approach is not specific to counting, and can be generalized to other compositional concepts that VLMs fail to learn." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3177" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In AAAI Conference on Artificial Intelligence, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.295 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. ArXiv, abs/2204.14198, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.297, + 0.468, + 0.338 + ], + "angle": 0, + "content": "[3] Jong Wook Kim Gretchen Krueger Sandhini Agarwal Alec Radford, Ilya Sutskever. Clip: Connecting text and images. https://openai.com/research/clip, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.468, + 0.395 + ], + "angle": 0, + "content": "[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18187-18197, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.468, + 0.476 + ], + "angle": 0, + "content": "[5] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.468, + 0.521 + ], + "angle": 0, + "content": "[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. ArXiv, abs/2204.02491, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.468, + 0.564 + ], + "angle": 0, + "content": "[7] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[8] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.468, + 0.647 + ], + "angle": 0, + "content": "[9] Hila Chefer, Sagie Benaim, Roni Paiss, and Lior Wolf. Image-based clip-guided essence transfer. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.468, + 0.717 + ], + "angle": 0, + "content": "[10] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 387-396, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[11] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiiovanni, Piotr Padlewski, Daniel M. Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model. ArXiv, abs/2209.06794, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[12] Katherine Crowson. Vqgan+clip. https://colab.research.google.com/drive/1L8oL-vLJXVcRzCFbPwOoMkPKJ8-aYdPN, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[14] Anuj Diwan, Layne Berry, Eunsol Choi, David F. Harwath, and Kyle Mahowald. Why is winoground hard? investigating failures in visuolinguistic compositionality. ArXiv, abs/2211.00768, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.249 + ], + "angle": 0, + "content": "[15] Raveen Doon, Tarun Kumar Rawat, and Shweta Gautam. Cifar-10 classification using deep convolutional neural network. 2018 IEEE Punecon, pages 1–5, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.252, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[16] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. 2004 Conference on Computer Vision and Pattern Recognition Workshop, pages 178–178, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.324, + 0.892, + 0.365 + ], + "angle": 0, + "content": "[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada. ACM Transactions on Graphics (TOG), 41:1 - 13, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.368, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[18] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arxiv:2307.10373, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.412, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[19] Patrick Helber, Benjamin Bischke, Andreas R. Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12:2217-2226, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.483, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadayath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, October 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[21] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. CVPR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.626, + 0.892, + 0.695 + ], + "angle": 0, + "content": "[22] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.698, + 0.892, + 0.724 + ], + "angle": 0, + "content": "[23] William Stanley Jevons. The power of numerical discrimination. Nature, 3:281-282, 1871. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.728, + 0.892, + 0.768 + ], + "angle": 0, + "content": "[24] E. L. Kaufman and Miles W. Lord. The discrimination of visual number. The American journal of psychology, 62 4:498-525, 1949. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.772, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[25] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.83, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. ArXiv, abs/2201.03546, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3178" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[29] Xingchao Liu, Chengyue Gong, Lemeng Wu, Shujian Zhang, Haoran Su, and Qiang Liu. Fusedream: Training-free text-to-image generation with improved clip+gan space optimization. ArXiv, abs/2112.01573, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[30] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.292 + ], + "angle": 0, + "content": "[31] Duy-Kien Nguyen, Vedanuj Goswami, and Xinlei Chen. Movie: Revisiting modulated convolutions. 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[32] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.469, + 0.419 + ], + "angle": 0, + "content": "[33] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.461 + ], + "angle": 0, + "content": "[34] Roni Paiss, Hila Chefer, and Lior Wolf. No token left behind: Explainability-aided image classification and generation. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.464, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[35] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[36] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, Mingxing Tan, and Quoc V. Le. Combined Scaling for Open-Vocabulary Image Classification. arXiv preprint arXiv:2111.10050, Nov. 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[37] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6967-6977, June 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[42] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 5389-5400. PMLR, 09-15 Jun 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[43] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[45] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[46] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15617-15629, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[47] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[48] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[49] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Bachmann, Amit H. Bermano, Daniel Cohen-Or, Amir Roshan Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Trans. Graph., 41:86:1-86:11, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[50] Rijnder Wever and Tom F.H. Runia. Subitizing with variational autoencoders. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, September 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[51] Yonghui Wu and David Fleet. How ai creates photorealistic images from text. https://blog.google/technology/research/how-ai-creates-photorealistic-images-from-text, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yin" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3179" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.16 + ], + "angle": 0, + "content": "fei Yang, Burcu Karagol Ayan, Benton C. Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. ArXiv, abs/2206.10789, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[53] Nir Zabari and Yedid Hoshen. Semantic segmentation in the-wild without seeing any segmentation examples. ArXiv, abs/2112.03185, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.288 + ], + "angle": 0, + "content": "[54] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18123-18133, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[55] Hui Zhang, Shenglong Zhou, Geoffrey Y. Li, and Naihua Xiu. 0/1 deep neural networks via block coordinate descent. ArXiv, abs/2206.09379, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[56] Jianming Zhang, Shugao Ma, Mehrnoosh Sameki, Stan Sclaroff, Margrit Betke, Zhe L. Lin, Xiaohui Shen, Brian L. Price, and Radomir Mech. Salient object subitizing. International Journal of Computer Vision, 124:169-186, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[57] Y. Zhang, Jonathon S. Hare, and Adam Prügel-Bennett. Learning to count objects in natural images for visual question answering. ArXiv, abs/1802.05766, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[58] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3180" + } + ] +] \ No newline at end of file diff --git a/2023/Teaching CLIP to Count to Ten/56d47f2c-86af-49f0-9e72-8decf5e81c8a_origin.pdf b/2023/Teaching CLIP to Count to Ten/56d47f2c-86af-49f0-9e72-8decf5e81c8a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..df185f6eb623ab4907cde27d3ab00c9c2d865486 --- /dev/null +++ b/2023/Teaching CLIP to Count to Ten/56d47f2c-86af-49f0-9e72-8decf5e81c8a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a431c21724254853948dfb0857af6c8766f83c0ff27454819a047dbc45bd4bd +size 5457841 diff --git a/2023/Teaching CLIP to Count to Ten/full.md b/2023/Teaching CLIP to Count to Ten/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3487ecf0c9dd416d47d89f96790d17f057aaa7e2 --- /dev/null +++ b/2023/Teaching CLIP to Count to Ten/full.md @@ -0,0 +1,385 @@ +# Teaching CLIP to Count to Ten + +Roni Paiss $^{1,2}$ + +Inbar Mosseri1 + +Ariel Ephrat1 + +Michal Irani1,3 + +Omer Tov1 + +Shiran Zada + +Tali Dekel1,3 + +3Weizmann Institute of Science + +1Google Research + +![](images/91eee95232937f0e252058f7c8c5886e2d7a2d041b876888df891b7a006fd360.jpg) +(a) Image Retrieval Results (from LAION): + +![](images/2f8bfc496bef39364a209a44028010b02b8a5719d381c46c36f5b74ed665813f.jpg) +Figure 1: Inability of CLIP to count. We show that CLIP is insensitive to the number of objects in an image, and demonstrate the effectiveness of our Counting-Aware CLIP. (a) Image retrieval results using text captions of the form “a photo of ”, with different numbers and types of objects. Images that match the caption are marked with $\checkmark$ ; images that do not are marked with $\times$ . Our Counting-Aware model retrieves images that depict the requested number of objects, while the original CLIP often retrieves images that contain the wrong number of objects, or images where the number is explicitly written in the image (e.g. “nine hearts” - the image contains the number “9”, but has 11 hearts). (b) Attention maps demonstrating that our model attends to all matching object instances in the image, unlike the original CLIP. + +![](images/1c98c803b7445f5f5ac68ae7b8587f31b5a883e8793b620c6a23af52beb17b6d.jpg) +(b) Relevancy Maps: + +# Abstract + +Large vision-language models (VLMs), such as CLIP, learn rich joint image-text representations, facilitating advances in numerous downstream tasks, including zero-shot classification and text-to-image generation. Nevertheless, existing VLMs exhibit a prominent well-documented limitation - they fail to encapsulate compositional concepts such as counting. We introduce a simple yet effective method to improve the quantitative understanding of VLMs, while maintaining their overall performance on common benchmarks. Specifically, we propose a new counting-contrastive loss used to finetune a pre-trained VLM in tandem with its original objective. Our counting loss is deployed over automatically-created counterfactual examples, each consisting of an image and a caption containing an incorrect object count. For example, an image depicting three dogs is paired with the caption "Six dogs playing in the yard" as a negative example. Our loss encourages discrimination + +between the correct caption and its counterfactual variant which serves as a hard negative example. To the best of our knowledge, this work is the first to extend CLIP's capabilities to object counting. Furthermore, we introduce "CountBench" - a new image-text counting benchmark for evaluating object counting capabilities. We demonstrate a significant improvement over state-of-the-art baseline models on this task. Finally, we leverage our counting-aware CLIP model for image retrieval and text-conditioned image generation, demonstrating that our model can produce specific counts of objects more reliably than existing ones. + +# 1. Introduction + +Since the advent of CLIP [38], training large vision-language models (VLMs) has become a prominent paradigm for representation learning in computer vision. By observing huge corpora of paired images and captions crawled from the Web, these models learn powerful + +and rich joint image-text embedding spaces, which have been employed in numerous visual tasks, including classification [58, 59], segmentation [27, 53], motion generation [47], image captioning [30, 48], text-to-image generation [12, 29, 32, 41, 44] and image or video editing [4, 6, 9, 17, 25, 35, 49, 18, 7]. Recently, VLMs have also been a key component in text-to-image generative models [5, 39, 41, 43], which rely on their textual representations to encapsulate the semantic meaning of the input text. + +Despite their power, prominent VLMs, such as CLIP [38] and BASIC [36], are known to possess a weak understanding of the number of objects present in an image [3, 36, 38]. This is demonstrated in Fig. 1, where, when given a caption of the template "a photo of ", CLIP often fails to retrieve images that correctly match the described number. Downstream applications that rely on VLM-based representations inherit these limitations, e.g., image generation models struggle to reliably produce specific counts of objects [43, 51, 52]. + +In this work, we introduce a novel method that enhances the quantitative understanding of large-scale VLMs by encouraging them to produce representations that are sensitive to the number of objects in the image and text. + +We hypothesize that the reason existing VLMs fail to learn the concept of counting is severalfold: (i) Captions that accurately specify the number of objects become extremely rare in the data as the number of objects increases. For example, we found that for more than five objects, captions typically contain a general form of quantity, e.g., "a group of..." or "many...", rather than an accurate count. This can be attributed to the fact that people cannot instantly identify numerical quantities larger than four without explicitly counting the objects [24]. (ii) Numbers in the caption often refer to attributes that are NOT related to counting - e.g. age, time, address, temperature, model number ("This is an iPhone 5"), etc. (iii) The task of counting (associating the visible number of objects in an image with the number in the caption), is not explicitly enforced in current VLM training objectives. Therefore, current VLMs are able to count reasonably well only up to two or three (for which there are sufficient image-caption examples). + +We thus suggest to mitigate each of these problems by: (i) Creating suitable training data in which the captions contain accurate numbers of objects. (ii) Designing a training objective whereby understanding object counts is critical for discriminating between the correctly associated caption and incorrect ones. + +More specifically, as illustrated in Fig. 2, we automatically create a clean and diverse counting training set by curating image-text examples where the image depicts multiple objects and its caption is verified to express their count. We then finetune a pretrained VLM by formulating counting as a discriminative task – for each example, we create + +a counterfactual caption by swapping the spelled number associated with the object count with a different randomly selected number. The model's objective is then to associate the image correctly with its true count caption, discriminating it from the counterfactual one. + +To evaluate counting capabilities, we introduce "Count-Bench" - a carefully curated object counting benchmark, consisting of 540 diverse, high quality image-text examples. We evaluate our method on two prominent contrastive VLMs: CLIP [38] and BASIC [36], and demonstrate a significant improvement in accuracy in the task of zero-shot count classification over baseline models. Importantly, we achieve this while maintaining the original knowledge learned by the VLM, as demonstrated by an extensive evaluation of our model on standard zero-shot downstream tasks. The quantitative understanding of our model is further evident by our text-to-image retrieval results (e.g., Fig. 1(a)), as well as by the relevancy maps of our model, which demonstrate that the model correctly attends to all visible objects whose count is specified in the text (e.g., Fig. 1(b)). Finally, we train a large-scale text-to-image generative model [43] which incorporates our counting training set and counting-aware CLIP text encoder. The generated images from this model exhibit higher fidelity to the number of objects specified in the input prompts (Fig. 8). + +To summarize, our main contributions are: + +1. A novel training framework for tackling the task of vision-language counting - an important limitation of current VLMs. +2. A new benchmark, "CountBench", carefully filtered and validated for evaluating VLMs on the counting task. +3. We apply our method to the widely-adopted VLMs CLIP and BASIC, demonstrating significant improvement on the counting task, while maintaining general (non-counting) performance on common benchmarks. +4. We utilize our counting-aware VLMs for downstream tasks including image retrieval and text-to-image generation, demonstrating more reliable results when the text prompt contains a specific number of objects. + +# 2. Related work + +Contrastive vision-language models: Vision-language models have demonstrated impressive success in vision and multimodal tasks [2, 11, 36, 38, 46, 54, 37]. In this work, we focus on contrastive VLMs, such as CLIP [38] and BASIC [36], as they are widely used both for downstream tasks and as backbones for generative models [40, 43]. Both are trained with a contrastive objective, where matching text-image pairs should have a low cosine distance, and non-matching texts and images should be far apart. The representations computed by CLIP have proven to be very effective in vision and multimodal tasks, due to their zero-shot + +![](images/8a31b465668d00df40532343923c99b3a06dffdeb2d866729316ad01775add10.jpg) +(a) Obtaining Counting Training Set: + +![](images/73c8c96307071d976e9816dda85057c3279c8dc726282ab2e619184fa84eaa57.jpg) +(b) Teaching CLIP to count: +Figure 2: Method overview (a) We create a text-image counting training set in which each caption expresses the number of objects depicted in the corresponding image. This is done by using an off-the-shelf object detector to automatically identify text-image examples in which the text count matches the number of visible objects in the image (see Sec. 3.1). (b) We finetune a pre-trained CLIP model using our counting subset (a), through a dedicated contrastive objective $L_{count}$ , used in addition to the original (general) contrastive objective ( $L_{clip}$ ). Specifically, given a text-image example from our counting subset, we automatically create a counterfactual prompt by replacing the true object count in the original caption with an incorrect count; $L_{count}$ encourages the model to embed the image close to its original caption embedding (expressing the true object count) and far from its counterfactual count (see Sec. 3.2). + +capabilities and semantic nature, and have been used as a prominent component in numerous tasks and methods. BASIC [36] scaled up the size of the model, batch size and dataset, improving zero-shot accuracy on common benchmarks. + +Subitizing: "Subitizing" [24] is the ability of people to instantly recognize the number of objects at a glance - without actually counting them. It has been shown [23] that most humans can subitize only a small number of objects (up to $\sim 4$ ). This cognitive phenomenon is reflected in the numbers people tend to specify in captions ( $\leqslant 4$ ), and sub- + +sequentially in the counting abilities of VLMs, which struggle to count larger numbers. Subitizing was also referred to in computer vision [50, 56]. These papers aim to predict a very small number ( $\leqslant 4$ ) of salient objects in an image, and are pure image-based (do not involve any text or VLMs). + +Counting in vision-language models: While demonstrating impressive recognition capabilities, large VLMs such as CLIP[38] and BASIC [36] are known to only partially capture the meaning of the text. Thus, they fail to understand the number of objects in an image [3, 36, 38]. Paiss et al. [34] demonstrated that CLIP attends to only a small subset of its input, mainly the nouns, and often ignores adjectives, numbers and prepositions. + +Counting has remained a stand-alone task under the domain of visual question answering (VQA), tackled with specifically designed architectures and techniques [1, 31, 57]. Our work defers from these prior efforts in several key aspects: (i) While previous efforts are restricted to VQA architectures and problem formulation, our goal is to improve the quantitative understanding of general-purpose contrastive VLMs (e.g., CLIP and BASIC), used in various vision and multimodal tasks where counting-aware solutions are not currently available. (ii) Our work can enhance the zero-shot counting capabilities of VLMs to unrestricted objects, unlike prior methods that are trained on specific domains, which can be problematic for new domains where no counting labels are available. + +# 3. Method + +Our goal is to teach a pre-trained VLM (e.g., CLIP) to count, i.e., to improve its quantitative textual and visual understanding. Our framework, illustrated in Fig. 2, consists of two main stages. We first automatically create a counting training set, comprising clean and diverse images along with corresponding captions that describe the number of visible objects in the scene. We then leverage this dataset to finetune the VLM through a designated count-based contrastive loss that is used in tandem with the original generic image-text objective. + +More specifically, our key idea is to automatically generate counterfactual examples by swapping the true object count in the caption with a different random number. Our new counting loss encourages the model to embed an image close to its true count, as expressed by the original caption, while pushing it away from the embedding of the counterfactual count prompt. As the only difference between the correct caption and its counterfactual counterpart is a single word—the spelled number of objects—the model has to distinguish between the correct and incorrect count in order to succeed in its training task. Next, we describe our dataset creation and finetuning paradigm in detail. + +# 3.1. Creating an image-text counting train set + +A naive approach for obtaining an image-text counting dataset is to filter a large-scale dataset for examples in which the caption contains a number. However, this approach results in a highly noisy dataset, since the numbers in the captions often refer to other numerical attributes that are unrelated to object counts. Such numerical attributes are age ("A 7 year old girl"), time ("The time is 9 o'clock"), addresses, model numbers, etc (see examples in supplementary). To ensure that the numbers in the captions indeed specify object counts, we employ several stages of automatic filtering in our data pipeline (Fig. 2 (a)): + +First, we filter out all examples whose caption does not contain a spelled number $\in \{\text{"two",\ldots,"ten"}\}$ . We do so, as we observed that non-spelled numbers, or numbers higher than ten, mostly appear in conjunction with a measure of time, (e.g. dates) or addresses, rather than numbers of objects present in the image. + +In the second stage, we verify that the spelled numbers serve as object counters, and that the counted objects are visible and detectable in the images. For example, for the caption "A photo of three dogs", we verify that the image indeed depicts three visible dogs, no more, and no less. This count verification is achieved automatically by first applying an off-the-shelf object detector [22], and counting the number of detections per object. We assume that the caption refers to the most prevalent object in the image. Thus, we retain only examples for which the number specified in the caption aligns with the number of instances of the maximally-detected object. We denote by $C$ our automatically filtered train set. + +Naturally, the filtered data $C$ is imbalanced. The number of examples that pass our filtering drops significantly as the count increases, e.g., the number of "ten" image-text pairs is around $1000 \times$ smaller than "two". Training with such imbalanced data creates a bias—the loss can be reduced by classifying frequent numbers as the correct caption and rare numbers as counterfactual, regardless of the image content. Therefore, balancing the data is of essence. Due to scarcity of examples depicting more than six objects, we choose to balance the numbers "two" – "six" separately from the higher numbers "seven" – "ten". For each of the numbers "two" – "six", we sample around $37K$ samples, while for "seven" – "ten", we use all the samples passed by our filter. There are approximately $7K$ samples for "seven" down to around $1.5K$ samples for "ten". We found this approach to provide us with a diverse and relatively balanced training dataset, yet more sophisticated methods could be considered in the future. From this point on, $C$ will denote our filtered and balanced numbered training set. + +# 3.2. Teaching CLIP to count + +Our goal is to improve the quantitative understanding of a pre-trained VLM (e.g., CLIP), while preserving its real-world knowledge, as reflected by its zero-shot capabilities on commonly-evaluated benchmarks. Therefore, we use a combination of two loss functions: + +$$ +L = L _ {C L I P} + \lambda L _ {\text {c o u n t}} \tag {1} +$$ + +where $L_{CLIP}$ is the regular contrastive loss of CLIP, $L_{count}$ is our counting-designated loss (described below), and $\lambda$ is a hyperparameter used to weight the two losses. + +We finetune the model on two training sets: (i) A very large dataset collected from the Web that contains general in-the-wild images and captions. (ii) Our filtered numbered training set $C$ , described in Sec. 3.1, which contains samples where object counts are spelled out in the captions. While $L_{CLIP}$ is calculated on all samples, the counting loss $L_{count}$ is calculated only on samples from $C$ . For each image-text pair $(i_k, t_k) \in C$ , a counterfactual caption $t_k^{CF}$ is automatically created by swapping the number in the caption $t_k$ with a different random number (e.g., the caption "five dogs" can be counterfactualized with "eight dogs"). At each training step, the triplets $(i_k, t_k, t_k^{CF})_{k=1}^N$ are then fed to CLIP's text and image encoders to obtain their embeddings $(ei_k, et_k, et_k^{CF})_{k=1}^N$ . + +Then, a contrastive loss $L_{count}$ is computed to enforce that the similarity score of the image is high with the original caption and low with the counterfactual caption: + +$$ +L _ {\text {c o u n t}} = - \frac {1}{N} \sum_ {k = 1} ^ {N} \log \frac {\exp \left(e i _ {k} \cdot e t _ {k}\right)}{\exp \left(e i _ {k} \cdot e t _ {k}\right) + \exp \left(e i _ {k} \cdot e t _ {k} ^ {C F}\right)} \tag {2} +$$ + +Since the original ground truth caption and counterfactual caption differ only by the number of objects specified in them, this loss encourages the model to learn the relationship between the specified spelled number and the number of the objects it refers to. + +We use the negative samples only in the counting objective $L_{count}$ , instead of adding them to the batch for the contrastive loss $L_{CLIP}$ in order to better weight their effect. + +# 4. "CountBench" - an evaluation benchmark + +We introduce a new object-counting evaluation benchmark called CountBench. This benchmark was automatically curated (and manually verified) from the publicly available LAION-400M image-text dataset [45]. CountBench contains a total of 540 images containing between 2 and 10 instances of a particular object, where their corresponding captions reflect this number. This benchmark is used only for testing and evaluation. CountBench has no overlap with our training set $C$ . + +![](images/60a53d01d6a9ce3c2b1c101e54229284346907e7b53a99bf0eed46556aee0dd5.jpg) +(a) Baseline CLIP (L/14) + +![](images/f72c4513de1134111a33303371c9c4125b9deb7b71eb4fb0718f6d9ca0f9e3f4.jpg) +(b) Ours (L/14) + +The images in CountBench were obtained by running our automatic filtering method described in Sec. 3.1 on the entire LAION-400M dataset. The filtering pipeline includes applying an object detector, whose benefits are twofold: (i) it facilitates verification of the object count specified in the caption, and (ii) it ensures that the counted objects are visible and detectable. The latter is important since benchmarks that do not disentangle the problem at hand from other challenges (e.g. occlusions and low resolution) may not reflect the performance on the actual task, since the difficulty can often stem from unrelated factors [14]. + +This filtering produced over 158K images for the number "two", but only around 100 images for "ten", demonstrating again the severe number imbalance we encountered with our training set. After automatically balancing each number to 100-200 samples each, the entire dataset was manually verified to contain only pairs in which the spelled number in the caption matches the number of clearly visible objects in the image. The dataset was rebalanced after this stage, ending up with 60 image-text pairs per number $\in\{\text{"two","..},\text{"ten"}\}$ , 540 in total. Samples from the dataset can be seen in Fig. 4. + +We use the CountBench benchmark to evaluate the counting abilities of the models trained with our method in Sec. 5. + +# 5. Experiments + +We thoroughly evaluate our method, both quantitatively and qualitatively, on object counting-related tasks using our CountBench benchmark. We further validate that the performance of our counting-aware models on a variety of general zero-shot classification benchmarks is retained [8, 13, 15, 16, 19, 20, 21, 26, 33, 42, 55]. To gain a better understanding of our models, we show visualizations + +![](images/809322e4d93c9c1c9fcc88d8ca69afa6090ee5a3399ef0e8c40779bbf8074e6f.jpg) +"two zebras in Cape Town" + +![](images/b4bf294d4124d4f9a8c295c122d52bf63f60ce1bc63a3fc69efe0af171d6ffa0.jpg) +"three purebred chihuahuaas running..." + +![](images/e5eee1eaa287978d9ca669b958af5e975956b258f8c511dbff96408a952f7c1e.jpg) +"...four antique pink and purple armchairs" + +![](images/f7020c2a1501ee4efb6c80d611fcdf289fb86300e156a6925d8088aaa5f1f2d9.jpg) +"GT's five favourite Champagne..." + +![](images/fac855746b08ea777dc8fff3a873c7d6181ce5873289b6cbc2b3a02c18710ceb.jpg) +"vintage silver plate tablespoons...set of six" +"seven Chewy Peanut Butter Cookies..." + +![](images/af29610451d12236218eff88d75dd5dcdee5352eceee7d27533794583d1e246a.jpg) + +![](images/2edc89da46c157df84fecdb8ce32e2025531e54496d17b28ba7b6476a7e015a1.jpg) +Figure 3: Confusion matrices on CountBench. Classification accuracy on CountBench, broken down into confusion matrices for the public CLIP ViT-L/14 (a), and our counting-aware CLIP ViT-L/14 model (b), demonstrating clear quantitative superiority of our model. +...eight colorful bright shiny red apples... +Figure 4: CountBench benchmark. Sample images and their corresponding captions from our new CountBench object counting benchmark. This benchmark was automatically curated (and manually verified) from the publicly-available LAION-400M dataset. + +![](images/2b4d7081091cf46a15d607ee3a58790cd8c805c541c9f8571f26a7e17e6715a9.jpg) +"The view of the nine leftmost moai..." + +![](images/ad0fcb7059a51d822e0dc10016f65fdb018648ff4457989345d3075f4c6191f1.jpg) +"All ten Christmas Poinsettia..." + +of text-image relevancy maps, along with per-word relevancy scores, demonstrating that our model indeed attends to the number of objects in the image and text. Finally, we apply our model to text-to-image retrieval and generation, producing specific numbers of objects more reliably than baseline models. + +# 5.1. Zero-shot counting accuracy + +We evaluate our models and baselines on CountBench on the task of classifying the number of objects in an image in a zero-shot manner. For each image in CountBench we augment the existing caption with eight other possible captions by replacing the number in its caption with all the numbers $\in \{\text{two}, \ldots, \text{ten}\}$ , and calculate the similarity score between the image and each of the nine captions. The number in the caption that obtains highest similarity score with the image is considered the predicted number. + +Table 1 reports the results of this evaluation on two prominent contrastive VLMs: CLIP-B/32 and BASIC-S. We report both the counting accuracy (selection of the correct number) and the mean deviation of the models' predictions from the correct numbers. For each of the architectures, we compare our model (configuration $E$ ) with two baseline configurations: (A) the official baseline model, and (B) the baseline model finetuned on our general text-image dataset used in our implementation, with the standard contrastive loss. Comparing the performance of these config + +
CLIP-B/32BASIC-S
A Official BaselineB Internal BaselineC Ours (w/o Lcount)D Ours (Naive Filtering)E OursA Official BaselineB Internal BaselineC Ours (w/o Lcount)D Ours (Naive Filtering)E Ours
Accuracy ↑31.6732.9444.2649.8175.9317.9722.7530.5928.6869.02
Mean deviation from the correct number ↓1.531.440.971.280.492.132.021.291.870.64
+ +Table 1: Quantitative counting results. Top-1 zero-shot accuracy and the mean absolute distance between the predicted numbers and the true numbers on CountBench. We compare several configurations: (A) The official CLIP [38] and BASIC [36] models. (B) The official baselines finetuned on our internal curated data. (C) Models trained with our filtered counting set, without $L_{count}$ (D) Models finetuned with $L_{count}$ on a naively filtered counting set (E) Our method, which is significantly superior to all other configurations, both in accuracy and deviation from correct number. + +
CLIP-B/32BASIC-S
Official BaselineInternal BaselineOursOfficial BaselineInternal BaselineOurs
ImageNet62.9364.9764.0659.7061.9661.18
CIFAR1063.9161.0060.6576.2284.6984.05
CIFAR10033.1032.4933.5645.3556.8055.89
Caltech10175.9982.5082.3678.1681.0381.05
EuroSAT45.2341.6637.6928.3945.8245.97
Food10183.0880.7280.5377.0877.8077.06
ImageNetA31.8530.8529.8117.6522.5521.68
ImageNetR69.3870.1770.3067.1167.6866.95
ImageNetV255.6556.5656.6252.2254.3553.60
Oxford Pets87.3587.7487.4180.6285.1584.87
Oxford Flowers66.1465.7367.3964.7466.4065.90
+ +Table 2: Zero-shot accuracy on common benchmarks. We compare the zero-shot accuracy of our method and baselines on a variety of popular benchmarks. As seen, our method preserves the performance of the original model. + +urations allows us to quantify the effect of using our own large-scale text-image dataset, which differs from the original unpublished data the models were trained on. + +As can be seen, our method $(E)$ achieves significantly superior counting accuracy compared to the baselines $(A, B)$ , attaining $2 - 3 \times$ higher counting accuracy and more than $3 \times$ lower mean deviation from the correct number. Results for different categories in CountBench are reported in App. 1. + +Tab. 1 also contains an ablation of the two components of our method: filtering a counting training set and finetuning with an additional loss $L_{count}$ . Models with configuration $C$ are finetuned on the filtered subset with no counting loss. The large gap in accuracy on CountBench between configurations $C$ and $E$ shows the importance of our loss for the improvement in counting abilities. Models with configuration $D$ are finetuned with the counting loss $L_{count}$ on an alternative counting subset, which consists of all the samples that contain spelled numbers $\in \{ "two",.., "ten" \}$ without additional filtering. The significant difference in counting accuracy between configurations $D$ and $E$ demonstrates the importance of our restrictive filtering pipeline, as both configurations are finetuned with $L_{count}$ over the samples from a dedicated counting training set. As can be seen in Tab. 1, while the naively filtered data does improve performance + +![](images/88ed71bec5f49597fb8e758243758382837c7068e11280ad99421afd496216da.jpg) + +![](images/d77c3f1fea7e20d39b76477ea44e25ebdc61b08744524187df25655b93fdfe5a.jpg) +"Two ducks" 0.23 0.77 + +![](images/9bc28b3af8480c287b33546b2bce9017a574af90c90c41f5dcfa86257fcc9e08.jpg) +"Two ducks" 0.45 0.55 + +![](images/3cf24cb2e1961fc90d2f81a01ace2126be093af29398a0f837802ff4f38369cb.jpg) + +![](images/69547da3ae2ee280ea58edeb7e3ba2b83426ef2c89e51ff0f73d63fdecef3fa6.jpg) +"Three pumpkins" 0.16 0.84 + +![](images/25f91edf186465f58417bbebbe5a24a8956fc030015c019bc7f59f7d92ec3542.jpg) +"Three pumpkins" 0.45 0.55 + +![](images/4b208bdb5d1566e2206f404ce99325a0cdb6408779e278fa188ad8e2e83af8eb.jpg) +Figure 5: Relevancy maps of both image and text. Visualization of the relevancy maps of both image and text, which represent, for each patch in the image and token in the text, how important it is to the prediction. Using our counting-aware CLIP model, the relevancy of the number (e.g., "four") in the text is increased. In addition, the model focuses on areas in the image that are relevant for counting. + +![](images/36002ba95075fead18bba537240aefdb39b32682285019ba140700f92dc5c5fe.jpg) +Four cards 0.30 0.70 + +![](images/a86317d09edd9c60f83620a0061d575f796add80a7392b018526cbf4903c2bb1.jpg) +"Four cards" 0.43 0.57 + +over a baseline trained without a dedicated counting subset, the obtained results are still significantly lower than those produced by our model. We attribute this gap in performance to mislabeled training samples in the naively filtered data, which are absent from our counting training set $C$ due to our filtering pipeline. + +Confusion matrices for the counting evaluation described above are shown in Fig. 3. For this experiment, we compare a public CLIP-L/14 model against our counting-aware version of it. As can be seen, our counting-aware CLIP model is significantly superior to the baseline across all numbers. It is worth noting that the baseline CLIP performs reasonably well for numbers within the subitizing range $(2 - 4)$ , but fails on larger numbers. This supports our claim that curated image captions with numbers beyond this range are rare, since specifying them requires manually counting the objects. + +![](images/9b288a39429169ee78a73b8e0159f6b682a0ec79bb79996a0d89b3ccf90dcfdd.jpg) +Figure 6: Relevancy maps for similarity between the image and different numbers. We compare the relevancy map of the input image with text prompts of several numbers (i.e. four to six) for both the baseline CLIP and our counting-aware CLIP. Our model focuses on the five stars when calculating the similarity with the prompt "five". + +# 5.2. Count-based image retrieval + +We consider the task of text-to-image retrieval where the text explicitly describes the desired count of objects. To obtain a diverse dataset that consists of varied numbers of objects, yet facilitates retrieval in reasonable time, we split the public LAION-400M dataset [45] into coarse categorical subsets by filtering samples where the caption contains a certain word (e.g., "dogs", "animals", "cars"), and perform retrieval on each of these subsets separately. + +For each category, we use the caption "a photo of " where number $\in$ {"two",..,"ten"} (e.g. "a photo of six dogs"). For each caption, we retrieve the three images in the dataset that are predicted by the model to be most similar to the caption. Note that since there are no ground truth labels for the counts of objects, we present qualitative results. Fig. 7 shows the retrieved images using the original CLIP model and our counting-aware CLIP model. An extended figure can be found in the supplementary. As can be seen, when the requested number is larger than three, the images retrieved by the baseline model often depict arbitrary numbers of objects. Additionally, the baseline often retrieves the same images for several different requested numbers. This further implies that the baseline model mostly focuses on the existence of the described object in the image, and ignores the number in the caption. In contrast, our results depict accurate object counts in most cases. + +# 5.3. Performance on non-counting classification + +To verify that our counting-aware models preserve the powerful image-text representation capabilities of the original models, we evaluate the zero-shot performance of our models on a variety of common classification benchmarks. Table 2 reports the zero-shot accuracy of our counting aware models against the baselines (corresponding to configurations $A$ , $B$ in Tab. 1). As can be seen, our models maintain similar overall accuracy. Additionally, comparing the official baseline and the internal baseline indicates that finetuning the models on our general text-image datasets leads to + +![](images/6c953c2ed96e8d820c8bb4adc5cafd6d49100ff1e5e582ea8f5575099330d4fb.jpg) +"A photo of {} shirts." +Figure 7: Top-3 count-based image retrieval Text-to-image retrieval results for different counts of shirts, ordered from left to right according to their similarity score (descending). Images that match the caption are marked with $\checkmark$ and images that do not match it are marked with X. Extended results are in the supplementary. + +only a slight shift in the accuracy of the models on common benchmarks. An additional evaluation on the task of general zero-shot image retrieval is reported in the supplementary. + +# 5.4. Relevancy map visualization + +To gain a better understanding of what our model learns, we use the explainability method of Chefer et al. [10] to visualize relevancy maps, which indicate the importance of different parts of the text and image to the model's similarity score prediction. Fig. 5 displays the normalized relevancy maps of several image-text pairs. Examining the relevancy maps of the text, it is apparent that the relevancy score of the spelled number in the caption is significantly higher for our model than the baseline model, which suggests that our model concentrates more on the mentioned number than the original one. Additionally, examining the relevancy maps of the images, it is evident that our model focuses on all pertinent objects in the image, whereas the original model + +![](images/2c97bf8ffdabe56d3ed450c20f7de34cfc2f33c62c97b2998b11bb126d4da6f2.jpg) + +![](images/e0001ba9b2ed6a87b7e609a22d9c4f2d9b8a8cc28900ae6f10c140e3d6258482.jpg) +"{} colorful tulips" + +![](images/ebe34c920450ea7c10831de000a7d5cd2520856fb71157c21d005f885dd837b1.jpg) + +![](images/426d15c5a0c64de4dfbb85d6b1148ea3d66c789081fcc85f07ff3d4f821308b4.jpg) + +![](images/e91967d31ccb05b26080542fe688696ae515cfc785cca4405ceb17c35b79b63e.jpg) + +![](images/0bc615c021b687a209861375174b46f94e9ec0e774b8d380d775da1d81298f0a.jpg) +"six" +Figure 8: Generated samples with Imagen using our counting-aware CLIP as backbone. The Imagen model benefits from the counting-aware representations produced by the our CLIP model, and is able to generate images that accurately follow the amounts specifies in the captions. + +![](images/db909db848b4f13445c4831d1595874f0455457fc232f3e9a674f867c06d46da.jpg) +"seven" + +![](images/ce72b3daa2f8576c510945ef95efe9da054c57d935a7ca5c39596912c2139328.jpg) +"eight" + +![](images/65c1382adcc99112df7c077cde483b17a94438615e241eb15b235c073f30de74.jpg) +"nine" + +![](images/e87c9cdbe4c81ff5d3134626a484e351b6117e00a05e29131994a2e1b2eac164.jpg) +"ten" + +![](images/32b94397b9daf17b2b617470b432e3770b095bff07386d0e07abb93d2fbaa4aa.jpg) + +![](images/4af404bbe7405baa102568802635fffcb01ef2b2f6a3fd907aed01a080d827d5.jpg) + +![](images/f83638fe5c5d627631142fccc209b8cc667401c5bfb2a8364a57b21b95b6af94.jpg) + +![](images/765b6d2321f0d93e25a9406b8d8994b7613ac292dc4423f35c35fdeebc84c074.jpg) +{}lamps + +![](images/180a81614f695a9d10633dc1bd3e2e071ed5ce9e975be81d420b52be09d7fe01.jpg) + +![](images/a450210894305ea521f3a9b8f81268d9ec90ce5a60e102f3e0ad8ba5bc765ad4.jpg) + +![](images/af9d53653afef40bc2ef1f6f8484ca738f5ea8277b792ff9f8b91f76dcfa6ec9.jpg) +"six" + +![](images/394468bffc7083cbb7d1b2dda123a7751ae18eca03374945b5a6c83026b3cf42.jpg) +"seven" + +![](images/be9fd6733a0a700645d1e2f720f0e9e5cc9539b8b034f9b7e0285521049cd0ad.jpg) +"eight" + +![](images/52030ed1d203acc98d76eee8221b8a8e2b112fea4a1e55d4373d67e830d5d514.jpg) +"nine" + +![](images/aae2c2ac76cf3623d30f5c59e8730f7564d25830c9b608b9bcce0e398e7cb0d7.jpg) +"ten" + +primarily identifies a single instance of the described object. + +To verify that our model does not simply attend to all objects that appear in the image, we examined the relevancy maps in Fig. 6 using negative prompts (i.e. "four" when there are five elements in the image). Our model focuses only on relevant objects when the correct number is used, unlike the baseline CLIP that highlights all object types in the image. This demonstrates that our model learns to associate the number in the caption with the suitable number of objects, and does not exploit shortcuts or undesired content. + +# 5.5. Count-based text-to-image generation + +Text-to-image generative models that rely on CLIP for representations, are known to inherit its limitations [43, 52], thus fail to reliably produce specific counts of objects. In order to demonstrate the effectiveness of our counting-aware model for downstream image generation tasks, we train an Imagen [43] model conditioned on the textual embeddings of a pretrained CLIP-L/14, and another Imagen model conditioned on our counting-aware version of this model. To compare our model and the baseline, we synthesize 12 samples for each prompt in the counting category of the Draw-Bench benchmark [43]. For each sample, we check whether or not it contains the requested number of objects, as stated in its prompt. + +The Imagen model trained with the baseline CLIP achieves a binary accuracy of $24.12\%$ , while the Imagen trained with out counting-aware model reaches $40.35\%$ accuracy. See the supplementary for additional evaluations. + +Performance on non-counting image generation: To verify that the generation quality is preserved for non-counting captions, we conducted a user study. We presented 52 participants with 20 pairs of images, each containing one image generated with an Imagen model trained with the baseline CLIP, and one with our counting-aware CLIP as backbone. The same random prompt from COCO [28] and seed were used for each pair. Participants were asked which image is better in terms of quality and realism. $54.7\%$ of ratings favored our images, suggesting that our model does not degrade generation quality on general text captions. + +# 5.6. Limitations + +Our method is limited by the insufficient existence of training data with images containing multiple instances of an object, along with a corresponding caption that correctly spells out this information. The effect of this data scarcity on our method increases with larger numbers (7, 8, etc.), as people tend to use "a group of" or "many" for large numbers of objects, instead of gruelingly counting them. Furthermore, many of the correct training pairs with higher numbers that do exist, contain relatively simplistic 2D collections of objects, as opposed to objects in a real-world scene. This is also the reason why our current method teaches CLIP to count only up to ten. Mitigating the above-mentioned data limitations, and generalizing to counting beyond 10, can possibly be achieved by concatenating multiple images with smaller numbers of objects. This will result in new training examples with more than ten natural-looking objects, and is part of our future work. + +# 6. Conclusions + +We present the first method to enhance VLMs with counting capabilities, while maintaining their overall performance on common benchmarks. This is an essential step towards enabling more accurate text-based image retrieval, classification and generation. A new counting-based contrastive loss is used to finetune a pre-trained VLM in tandem with its original objective. Our counting loss is deployed over automatically-created counterfactual examples, each consisting of an image and a caption containing an incorrect object count. We demonstrate a significant improvement over state-of-the-art baseline models in multiple tasks (classification, retrieval and generation) on several different datasets. Furthermore, we introduce CountBench - a new image-text counting benchmark for evaluating VLMs' object counting capabilities. This new benchmark contains in-the-wild images and captions, where the object counts are specified in the captions. Finally, our approach is not specific to counting, and can be generalized to other compositional concepts that VLMs fail to learn. + +# References + +[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In AAAI Conference on Artificial Intelligence, 2018. 3 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. ArXiv, abs/2204.14198, 2022. 2 +[3] Jong Wook Kim Gretchen Krueger Sandhini Agarwal Alec Radford, Ilya Sutskever. Clip: Connecting text and images. https://openai.com/research/clip, 2021. 2, 3 +[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18187-18197, 2022. 2 +[5] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2 +[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. ArXiv, abs/2204.02491, 2022. 2 +[7] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 2 +[8] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, 2014. 5 +[9] Hila Chefer, Sagie Benaim, Roni Paiss, and Lior Wolf. Image-based clip-guided essence transfer. In ECCV, 2022. 2 +[10] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 387-396, 2021. 7 +[11] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiiovanni, Piotr Padlewski, Daniel M. Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model. ArXiv, abs/2209.06794, 2022. 2 +[12] Katherine Crowson. Vqgan+clip. https://colab.research.google.com/drive/1L8oL-vLJXVcRzCFbPwOoMkPKJ8-aYdPN, 2021. 2 + +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 5 +[14] Anuj Diwan, Layne Berry, Eunsol Choi, David F. Harwath, and Kyle Mahowald. Why is winoground hard? investigating failures in visuolinguistic compositionality. ArXiv, abs/2211.00768, 2022. 5 +[15] Raveen Doon, Tarun Kumar Rawat, and Shweta Gautam. Cifar-10 classification using deep convolutional neural network. 2018 IEEE Punecon, pages 1–5, 2018. 5 +[16] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. 2004 Conference on Computer Vision and Pattern Recognition Workshop, pages 178–178, 2004. 5 +[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada. ACM Transactions on Graphics (TOG), 41:1 - 13, 2022. 2 +[18] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arxiv:2307.10373, 2023. 2 +[19] Patrick Helber, Benjamin Bischke, Andreas R. Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12:2217-2226, 2019. 5 +[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadayath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, October 2021. 5 +[21] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. CVPR, 2021. 5 +[22] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 4 +[23] William Stanley Jevons. The power of numerical discrimination. Nature, 3:281-282, 1871. 3 +[24] E. L. Kaufman and Miles W. Lord. The discrimination of visual number. The American journal of psychology, 62 4:498-525, 1949. 2, 3 +[25] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2 +[26] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5 +[27] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. ArXiv, abs/2201.03546, 2022. 2 + +[28] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 8 +[29] Xingchao Liu, Chengyue Gong, Lemeng Wu, Shujian Zhang, Haoran Su, and Qiang Liu. Fusedream: Training-free text-to-image generation with improved clip+gan space optimization. ArXiv, abs/2112.01573, 2021. 2 +[30] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 2 +[31] Duy-Kien Nguyen, Vedanuj Goswami, and Xinlei Chen. Movie: Revisiting modulated convolutions. 2021. 3 +[32] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2 +[33] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5 +[34] Roni Paiss, Hila Chefer, and Lior Wolf. No token left behind: Explainability-aided image classification and generation. In ECCV, 2022. 3 +[35] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2 +[36] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, Mingxing Tan, and Quoc V. Le. Combined Scaling for Open-Vocabulary Image Classification. arXiv preprint arXiv:2111.10050, Nov. 2021. 2, 3, 6 +[37] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6967-6977, June 2023. 2 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 6 +[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 2 + +[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 2 +[42] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 5389-5400. PMLR, 09-15 Jun 2019. 5 +[43] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, 2022. 2, 8 +[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 2 +[45] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 4, 7 +[46] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15617-15629, 2021. 2 +[47] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022. 2 +[48] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021. 2 +[49] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Bachmann, Amit H. Bermano, Daniel Cohen-Or, Amir Roshan Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Trans. Graph., 41:86:1-86:11, 2022. 2 +[50] Rijnder Wever and Tom F.H. Runia. Subitizing with variational autoencoders. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, September 2018. 3 +[51] Yonghui Wu and David Fleet. How ai creates photorealistic images from text. https://blog.google/technology/research/how-ai-creates-photorealistic-images-from-text, 2022. 2 +[52] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yin + +fei Yang, Burcu Karagol Ayan, Benton C. Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. ArXiv, abs/2206.10789, 2022. 2, 8 +[53] Nir Zabari and Yedid Hoshen. Semantic segmentation in the-wild without seeing any segmentation examples. ArXiv, abs/2112.03185, 2021. 2 +[54] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18123-18133, June 2022. 2 +[55] Hui Zhang, Shenglong Zhou, Geoffrey Y. Li, and Naihua Xiu. 0/1 deep neural networks via block coordinate descent. ArXiv, abs/2206.09379, 2022. 5 +[56] Jianming Zhang, Shugao Ma, Mehrnoosh Sameki, Stan Sclaroff, Margrit Betke, Zhe L. Lin, Xiaohui Shen, Brian L. Price, and Radomir Mech. Salient object subitizing. International Journal of Computer Vision, 124:169-186, 2015. 3 +[57] Y. Zhang, Jonathon S. Hare, and Adam Prügel-Bennett. Learning to count objects in natural images for visual question answering. ArXiv, abs/1802.05766, 2018. 3 +[58] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022. 2 \ No newline at end of file diff --git a/2023/Teaching CLIP to Count to Ten/images.zip b/2023/Teaching CLIP to Count to Ten/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c33de83a998d0227629bafe2aaf28a71fcaf47be --- /dev/null +++ b/2023/Teaching CLIP to Count to Ten/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6aa28761a1be8f5fc371cd1d6b9606fdf6f80a56be9f031f435cdef9c9923057 +size 682879 diff --git a/2023/Teaching CLIP to Count to Ten/layout.json b/2023/Teaching CLIP to Count to Ten/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..611f4187a75b1afc4b430ae2e16cb1112232e42e --- /dev/null +++ b/2023/Teaching CLIP to Count to Ten/layout.json @@ -0,0 +1,10407 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 199, + 103, + 395, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 103, + 395, + 121 + ], + "spans": [ + { + "bbox": [ + 199, + 103, + 395, + 121 + ], + "type": "text", + "content": "Teaching CLIP to Count to Ten" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 147, + 216, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 147, + 216, + 162 + ], + "spans": [ + { + "bbox": [ + 151, + 147, + 216, + 162 + ], + "type": "text", + "content": "Roni Paiss" + }, + { + "bbox": [ + 151, + 147, + 216, + 162 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 168, + 163, + 241, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 163, + 241, + 175 + ], + "spans": [ + { + "bbox": [ + 168, + 163, + 241, + 175 + ], + "type": "text", + "content": "Inbar Mosseri1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 229, + 148, + 294, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 148, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 229, + 148, + 294, + 161 + ], + "type": "text", + "content": "Ariel Ephrat1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 267, + 163, + 339, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 163, + 339, + 175 + ], + "spans": [ + { + "bbox": [ + 267, + 163, + 339, + 175 + ], + "type": "text", + "content": "Michal Irani1,3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 148, + 362, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 148, + 362, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 148, + 362, + 161 + ], + "type": "text", + "content": "Omer Tov1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 362, + 148, + 452, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 148, + 452, + 161 + ], + "spans": [ + { + "bbox": [ + 362, + 148, + 452, + 161 + ], + "type": "text", + "content": "Shiran Zada" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 362, + 162, + 424, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 162, + 424, + 175 + ], + "spans": [ + { + "bbox": [ + 362, + 162, + 424, + 175 + ], + "type": "text", + "content": "Tali Dekel1,3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 362, + 176, + 517, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 176, + 517, + 189 + ], + "spans": [ + { + "bbox": [ + 362, + 176, + 517, + 189 + ], + "type": "text", + "content": "3Weizmann Institute of Science" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 76, + 175, + 166, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 175, + 166, + 190 + ], + "spans": [ + { + "bbox": [ + 76, + 175, + 166, + 190 + ], + "type": "text", + "content": "1Google Research" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 78, + 214, + 248, + 386 + ], + "blocks": [ + { + "bbox": [ + 73, + 198, + 268, + 211 + ], + "lines": [ + { + "bbox": [ + 73, + 198, + 268, + 211 + ], + "spans": [ + { + "bbox": [ + 73, + 198, + 268, + 211 + ], + "type": "text", + "content": "(a) Image Retrieval Results (from LAION):" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 214, + 248, + 386 + ], + "lines": [ + { + "bbox": [ + 78, + 214, + 248, + 386 + ], + "spans": [ + { + "bbox": [ + 78, + 214, + 248, + 386 + ], + "type": "image", + "image_path": "91eee95232937f0e252058f7c8c5886e2d7a2d041b876888df891b7a006fd360.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 256, + 214, + 411, + 385 + ], + "blocks": [ + { + "bbox": [ + 256, + 214, + 411, + 385 + ], + "lines": [ + { + "bbox": [ + 256, + 214, + 411, + 385 + ], + "spans": [ + { + "bbox": [ + 256, + 214, + 411, + 385 + ], + "type": "image", + "image_path": "2f8bfc496bef39364a209a44028010b02b8a5719d381c46c36f5b74ed665813f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 393, + 547, + 478 + ], + "lines": [ + { + "bbox": [ + 45, + 393, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 45, + 393, + 547, + 478 + ], + "type": "text", + "content": "Figure 1: Inability of CLIP to count. We show that CLIP is insensitive to the number of objects in an image, and demonstrate the effectiveness of our Counting-Aware CLIP. (a) Image retrieval results using text captions of the form “a photo of ”, with different numbers and types of objects. Images that match the caption are marked with " + }, + { + "bbox": [ + 45, + 393, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 45, + 393, + 547, + 478 + ], + "type": "text", + "content": "; images that do not are marked with " + }, + { + "bbox": [ + 45, + 393, + 547, + 478 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 45, + 393, + 547, + 478 + ], + "type": "text", + "content": ". Our Counting-Aware model retrieves images that depict the requested number of objects, while the original CLIP often retrieves images that contain the wrong number of objects, or images where the number is explicitly written in the image (e.g. “nine hearts” - the image contains the number “9”, but has 11 hearts). (b) Attention maps demonstrating that our model attends to all matching object instances in the image, unlike the original CLIP." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 424, + 224, + 522, + 363 + ], + "blocks": [ + { + "bbox": [ + 424, + 198, + 520, + 212 + ], + "lines": [ + { + "bbox": [ + 424, + 198, + 520, + 212 + ], + "spans": [ + { + "bbox": [ + 424, + 198, + 520, + 212 + ], + "type": "text", + "content": "(b) Relevancy Maps:" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 424, + 224, + 522, + 363 + ], + "lines": [ + { + "bbox": [ + 424, + 224, + 522, + 363 + ], + "spans": [ + { + "bbox": [ + 424, + 224, + 522, + 363 + ], + "type": "image", + "image_path": "1c98c803b7445f5f5ac68ae7b8587f31b5a883e8793b620c6a23af52beb17b6d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 143, + 487, + 192, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 487, + 192, + 499 + ], + "spans": [ + { + "bbox": [ + 143, + 487, + 192, + 499 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": "Large vision-language models (VLMs), such as CLIP, learn rich joint image-text representations, facilitating advances in numerous downstream tasks, including zero-shot classification and text-to-image generation. Nevertheless, existing VLMs exhibit a prominent well-documented limitation - they fail to encapsulate compositional concepts such as counting. We introduce a simple yet effective method to improve the quantitative understanding of VLMs, while maintaining their overall performance on common benchmarks. Specifically, we propose a new counting-contrastive loss used to finetune a pre-trained VLM in tandem with its original objective. Our counting loss is deployed over automatically-created counterfactual examples, each consisting of an image and a caption containing an incorrect object count. For example, an image depicting three dogs is paired with the caption \"Six dogs playing in the yard\" as a negative example. Our loss encourages discrimination" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 488, + 547, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 547, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 547, + 621 + ], + "type": "text", + "content": "between the correct caption and its counterfactual variant which serves as a hard negative example. To the best of our knowledge, this work is the first to extend CLIP's capabilities to object counting. Furthermore, we introduce \"CountBench\" - a new image-text counting benchmark for evaluating object counting capabilities. We demonstrate a significant improvement over state-of-the-art baseline models on this task. Finally, we leverage our counting-aware CLIP model for image retrieval and text-conditioned image generation, demonstrating that our model can produce specific counts of objects more reliably than existing ones." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 633, + 387, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 387, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 387, + 646 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Since the advent of CLIP [38], training large vision-language models (VLMs) has become a prominent paradigm for representation learning in computer vision. By observing huge corpora of paired images and captions crawled from the Web, these models learn powerful" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3170" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 179 + ], + "type": "text", + "content": "and rich joint image-text embedding spaces, which have been employed in numerous visual tasks, including classification [58, 59], segmentation [27, 53], motion generation [47], image captioning [30, 48], text-to-image generation [12, 29, 32, 41, 44] and image or video editing [4, 6, 9, 17, 25, 35, 49, 18, 7]. Recently, VLMs have also been a key component in text-to-image generative models [5, 39, 41, 43], which rely on their textual representations to encapsulate the semantic meaning of the input text." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 181, + 286, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 286, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 286, + 301 + ], + "type": "text", + "content": "Despite their power, prominent VLMs, such as CLIP [38] and BASIC [36], are known to possess a weak understanding of the number of objects present in an image [3, 36, 38]. This is demonstrated in Fig. 1, where, when given a caption of the template \"a photo of \", CLIP often fails to retrieve images that correctly match the described number. Downstream applications that rely on VLM-based representations inherit these limitations, e.g., image generation models struggle to reliably produce specific counts of objects [43, 51, 52]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 302, + 286, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 302, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 302, + 286, + 350 + ], + "type": "text", + "content": "In this work, we introduce a novel method that enhances the quantitative understanding of large-scale VLMs by encouraging them to produce representations that are sensitive to the number of objects in the image and text." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 351, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 286, + 567 + ], + "type": "text", + "content": "We hypothesize that the reason existing VLMs fail to learn the concept of counting is severalfold: (i) Captions that accurately specify the number of objects become extremely rare in the data as the number of objects increases. For example, we found that for more than five objects, captions typically contain a general form of quantity, e.g., \"a group of...\" or \"many...\", rather than an accurate count. This can be attributed to the fact that people cannot instantly identify numerical quantities larger than four without explicitly counting the objects [24]. (ii) Numbers in the caption often refer to attributes that are NOT related to counting - e.g. age, time, address, temperature, model number (\"This is an iPhone 5\"), etc. (iii) The task of counting (associating the visible number of objects in an image with the number in the caption), is not explicitly enforced in current VLM training objectives. Therefore, current VLMs are able to count reasonably well only up to two or three (for which there are sufficient image-caption examples)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 568, + 286, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 286, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 286, + 639 + ], + "type": "text", + "content": "We thus suggest to mitigate each of these problems by: (i) Creating suitable training data in which the captions contain accurate numbers of objects. (ii) Designing a training objective whereby understanding object counts is critical for discriminating between the correctly associated caption and incorrect ones." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 286, + 713 + ], + "type": "text", + "content": "More specifically, as illustrated in Fig. 2, we automatically create a clean and diverse counting training set by curating image-text examples where the image depicts multiple objects and its caption is verified to express their count. We then finetune a pretrained VLM by formulating counting as a discriminative task – for each example, we create" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "a counterfactual caption by swapping the spelled number associated with the object count with a different randomly selected number. The model's objective is then to associate the image correctly with its true count caption, discriminating it from the counterfactual one." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 133, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 371 + ], + "type": "text", + "content": "To evaluate counting capabilities, we introduce \"Count-Bench\" - a carefully curated object counting benchmark, consisting of 540 diverse, high quality image-text examples. We evaluate our method on two prominent contrastive VLMs: CLIP [38] and BASIC [36], and demonstrate a significant improvement in accuracy in the task of zero-shot count classification over baseline models. Importantly, we achieve this while maintaining the original knowledge learned by the VLM, as demonstrated by an extensive evaluation of our model on standard zero-shot downstream tasks. The quantitative understanding of our model is further evident by our text-to-image retrieval results (e.g., Fig. 1(a)), as well as by the relevancy maps of our model, which demonstrate that the model correctly attends to all visible objects whose count is specified in the text (e.g., Fig. 1(b)). Finally, we train a large-scale text-to-image generative model [43] which incorporates our counting training set and counting-aware CLIP text encoder. The generated images from this model exhibit higher fidelity to the number of objects specified in the input prompts (Fig. 8)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 372, + 478, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 478, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 478, + 382 + ], + "type": "text", + "content": "To summarize, our main contributions are:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 388, + 545, + 546 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 308, + 388, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 422 + ], + "type": "text", + "content": "1. A novel training framework for tackling the task of vision-language counting - an important limitation of current VLMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 425, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 425, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 425, + 545, + 448 + ], + "type": "text", + "content": "2. A new benchmark, \"CountBench\", carefully filtered and validated for evaluating VLMs on the counting task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 450, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 450, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 308, + 450, + 545, + 496 + ], + "type": "text", + "content": "3. We apply our method to the widely-adopted VLMs CLIP and BASIC, demonstrating significant improvement on the counting task, while maintaining general (non-counting) performance on common benchmarks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 498, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 498, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 308, + 498, + 545, + 546 + ], + "type": "text", + "content": "4. We utilize our counting-aware VLMs for downstream tasks including image retrieval and text-to-image generation, demonstrating more reliable results when the text prompt contains a specific number of objects." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 562, + 389, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 389, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 389, + 574 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "content": "Contrastive vision-language models: Vision-language models have demonstrated impressive success in vision and multimodal tasks [2, 11, 36, 38, 46, 54, 37]. In this work, we focus on contrastive VLMs, such as CLIP [38] and BASIC [36], as they are widely used both for downstream tasks and as backbones for generative models [40, 43]. Both are trained with a contrastive objective, where matching text-image pairs should have a low cosine distance, and non-matching texts and images should be far apart. The representations computed by CLIP have proven to be very effective in vision and multimodal tasks, due to their zero-shot" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3171" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 86, + 288, + 148 + ], + "blocks": [ + { + "bbox": [ + 41, + 72, + 203, + 84 + ], + "lines": [ + { + "bbox": [ + 41, + 72, + 203, + 84 + ], + "spans": [ + { + "bbox": [ + 41, + 72, + 203, + 84 + ], + "type": "text", + "content": "(a) Obtaining Counting Training Set:" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 86, + 288, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 86, + 288, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 86, + 288, + 148 + ], + "type": "image", + "image_path": "8a31b465668d00df40532343923c99b3a06dffdeb2d866729316ad01775add10.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 45, + 176, + 293, + 356 + ], + "blocks": [ + { + "bbox": [ + 42, + 163, + 164, + 175 + ], + "lines": [ + { + "bbox": [ + 42, + 163, + 164, + 175 + ], + "spans": [ + { + "bbox": [ + 42, + 163, + 164, + 175 + ], + "type": "text", + "content": "(b) Teaching CLIP to count:" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 176, + 293, + 356 + ], + "lines": [ + { + "bbox": [ + 45, + 176, + 293, + 356 + ], + "spans": [ + { + "bbox": [ + 45, + 176, + 293, + 356 + ], + "type": "image", + "image_path": "73c8c96307071d976e9816dda85057c3279c8dc726282ab2e619184fa84eaa57.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 361, + 289, + 553 + ], + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 553 + ], + "type": "text", + "content": "Figure 2: Method overview (a) We create a text-image counting training set in which each caption expresses the number of objects depicted in the corresponding image. This is done by using an off-the-shelf object detector to automatically identify text-image examples in which the text count matches the number of visible objects in the image (see Sec. 3.1). (b) We finetune a pre-trained CLIP model using our counting subset (a), through a dedicated contrastive objective " + }, + { + "bbox": [ + 46, + 361, + 289, + 553 + ], + "type": "inline_equation", + "content": "L_{count}" + }, + { + "bbox": [ + 46, + 361, + 289, + 553 + ], + "type": "text", + "content": ", used in addition to the original (general) contrastive objective (" + }, + { + "bbox": [ + 46, + 361, + 289, + 553 + ], + "type": "inline_equation", + "content": "L_{clip}" + }, + { + "bbox": [ + 46, + 361, + 289, + 553 + ], + "type": "text", + "content": "). Specifically, given a text-image example from our counting subset, we automatically create a counterfactual prompt by replacing the true object count in the original caption with an incorrect count; " + }, + { + "bbox": [ + 46, + 361, + 289, + 553 + ], + "type": "inline_equation", + "content": "L_{count}" + }, + { + "bbox": [ + 46, + 361, + 289, + 553 + ], + "type": "text", + "content": " encourages the model to embed the image close to its original caption embedding (expressing the true object count) and far from its counterfactual count (see Sec. 3.2)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 568, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 628 + ], + "type": "text", + "content": "capabilities and semantic nature, and have been used as a prominent component in numerous tasks and methods. BASIC [36] scaled up the size of the model, batch size and dataset, improving zero-shot accuracy on common benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Subitizing: \"Subitizing\" [24] is the ability of people to instantly recognize the number of objects at a glance - without actually counting them. It has been shown [23] that most humans can subitize only a small number of objects (up to " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\sim 4" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "). This cognitive phenomenon is reflected in the numbers people tend to specify in captions (" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\leqslant 4" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "), and sub-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "sequentially in the counting abilities of VLMs, which struggle to count larger numbers. Subitizing was also referred to in computer vision [50, 56]. These papers aim to predict a very small number (" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "inline_equation", + "content": "\\leqslant 4" + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": ") of salient objects in an image, and are pure image-based (do not involve any text or VLMs)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 147, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 545, + 243 + ], + "type": "text", + "content": "Counting in vision-language models: While demonstrating impressive recognition capabilities, large VLMs such as CLIP[38] and BASIC [36] are known to only partially capture the meaning of the text. Thus, they fail to understand the number of objects in an image [3, 36, 38]. Paiss et al. [34] demonstrated that CLIP attends to only a small subset of its input, mainly the nouns, and often ignores adjectives, numbers and prepositions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 244, + 546, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 546, + 411 + ], + "type": "text", + "content": "Counting has remained a stand-alone task under the domain of visual question answering (VQA), tackled with specifically designed architectures and techniques [1, 31, 57]. Our work defers from these prior efforts in several key aspects: (i) While previous efforts are restricted to VQA architectures and problem formulation, our goal is to improve the quantitative understanding of general-purpose contrastive VLMs (e.g., CLIP and BASIC), used in various vision and multimodal tasks where counting-aware solutions are not currently available. (ii) Our work can enhance the zero-shot counting capabilities of VLMs to unrestricted objects, unlike prior methods that are trained on specific domains, which can be problematic for new domains where no counting labels are available." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 426, + 361, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 361, + 439 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 361, + 439 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 568 + ], + "type": "text", + "content": "Our goal is to teach a pre-trained VLM (e.g., CLIP) to count, i.e., to improve its quantitative textual and visual understanding. Our framework, illustrated in Fig. 2, consists of two main stages. We first automatically create a counting training set, comprising clean and diverse images along with corresponding captions that describe the number of visible objects in the scene. We then leverage this dataset to finetune the VLM through a designated count-based contrastive loss that is used in tandem with the original generic image-text objective." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "More specifically, our key idea is to automatically generate counterfactual examples by swapping the true object count in the caption with a different random number. Our new counting loss encourages the model to embed an image close to its true count, as expressed by the original caption, while pushing it away from the embedding of the counterfactual count prompt. As the only difference between the correct caption and its counterfactual counterpart is a single word—the spelled number of objects—the model has to distinguish between the correct and incorrect count in order to succeed in its training task. Next, we describe our dataset creation and finetuning paradigm in detail." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3172" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 263, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 263, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 263, + 85 + ], + "type": "text", + "content": "3.1. Creating an image-text counting train set" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 99, + 289, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 99, + 289, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 99, + 289, + 232 + ], + "type": "text", + "content": "A naive approach for obtaining an image-text counting dataset is to filter a large-scale dataset for examples in which the caption contains a number. However, this approach results in a highly noisy dataset, since the numbers in the captions often refer to other numerical attributes that are unrelated to object counts. Such numerical attributes are age (\"A 7 year old girl\"), time (\"The time is 9 o'clock\"), addresses, model numbers, etc (see examples in supplementary). To ensure that the numbers in the captions indeed specify object counts, we employ several stages of automatic filtering in our data pipeline (Fig. 2 (a)):" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "text", + "content": "First, we filter out all examples whose caption does not contain a spelled number " + }, + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\in \\{\\text{\"two\",\\ldots,\"ten\"}\\}" + }, + { + "bbox": [ + 46, + 240, + 287, + 312 + ], + "type": "text", + "content": ". We do so, as we observed that non-spelled numbers, or numbers higher than ten, mostly appear in conjunction with a measure of time, (e.g. dates) or addresses, rather than numbers of objects present in the image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 321, + 289, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 289, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 289, + 477 + ], + "type": "text", + "content": "In the second stage, we verify that the spelled numbers serve as object counters, and that the counted objects are visible and detectable in the images. For example, for the caption \"A photo of three dogs\", we verify that the image indeed depicts three visible dogs, no more, and no less. This count verification is achieved automatically by first applying an off-the-shelf object detector [22], and counting the number of detections per object. We assume that the caption refers to the most prevalent object in the image. Thus, we retain only examples for which the number specified in the caption aligns with the number of instances of the maximally-detected object. We denote by " + }, + { + "bbox": [ + 46, + 321, + 289, + 477 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 321, + 289, + 477 + ], + "type": "text", + "content": " our automatically filtered train set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "Naturally, the filtered data " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " is imbalanced. The number of examples that pass our filtering drops significantly as the count increases, e.g., the number of \"ten\" image-text pairs is around " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "1000 \\times" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " smaller than \"two\". Training with such imbalanced data creates a bias—the loss can be reduced by classifying frequent numbers as the correct caption and rare numbers as counterfactual, regardless of the image content. Therefore, balancing the data is of essence. Due to scarcity of examples depicting more than six objects, we choose to balance the numbers \"two\" – \"six\" separately from the higher numbers \"seven\" – \"ten\". For each of the numbers \"two\" – \"six\", we sample around " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "37K" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " samples, while for \"seven\" – \"ten\", we use all the samples passed by our filter. There are approximately " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "7K" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " samples for \"seven\" down to around " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "1.5K" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " samples for \"ten\". We found this approach to provide us with a diverse and relatively balanced training dataset, yet more sophisticated methods could be considered in the future. From this point on, " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " will denote our filtered and balanced numbered training set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 441, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 441, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 441, + 85 + ], + "type": "text", + "content": "3.2. Teaching CLIP to count" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 90, + 547, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 90, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 305, + 90, + 547, + 149 + ], + "type": "text", + "content": "Our goal is to improve the quantitative understanding of a pre-trained VLM (e.g., CLIP), while preserving its real-world knowledge, as reflected by its zero-shot capabilities on commonly-evaluated benchmarks. Therefore, we use a combination of two loss functions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 376, + 159, + 546, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 159, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 376, + 159, + 546, + 172 + ], + "type": "interline_equation", + "content": "L = L _ {C L I P} + \\lambda L _ {\\text {c o u n t}} \\tag {1}", + "image_path": "d8fe0f8a53fb8343f25e625b2fea3cff74427905c4cbfa9192d0f0f18a28dceb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 180, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 216 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 180, + 545, + 216 + ], + "type": "inline_equation", + "content": "L_{CLIP}" + }, + { + "bbox": [ + 304, + 180, + 545, + 216 + ], + "type": "text", + "content": " is the regular contrastive loss of CLIP, " + }, + { + "bbox": [ + 304, + 180, + 545, + 216 + ], + "type": "inline_equation", + "content": "L_{count}" + }, + { + "bbox": [ + 304, + 180, + 545, + 216 + ], + "type": "text", + "content": " is our counting-designated loss (described below), and " + }, + { + "bbox": [ + 304, + 180, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 180, + 545, + 216 + ], + "type": "text", + "content": " is a hyperparameter used to weight the two losses." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "text", + "content": "We finetune the model on two training sets: (i) A very large dataset collected from the Web that contains general in-the-wild images and captions. (ii) Our filtered numbered training set " + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "text", + "content": ", described in Sec. 3.1, which contains samples where object counts are spelled out in the captions. While " + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "inline_equation", + "content": "L_{CLIP}" + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "text", + "content": " is calculated on all samples, the counting loss " + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "inline_equation", + "content": "L_{count}" + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "text", + "content": " is calculated only on samples from " + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "text", + "content": ". For each image-text pair " + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "inline_equation", + "content": "(i_k, t_k) \\in C" + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "text", + "content": ", a counterfactual caption " + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "inline_equation", + "content": "t_k^{CF}" + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "text", + "content": " is automatically created by swapping the number in the caption " + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "text", + "content": " with a different random number (e.g., the caption \"five dogs\" can be counterfactualized with \"eight dogs\"). At each training step, the triplets " + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "inline_equation", + "content": "(i_k, t_k, t_k^{CF})_{k=1}^N" + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "text", + "content": " are then fed to CLIP's text and image encoders to obtain their embeddings " + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "inline_equation", + "content": "(ei_k, et_k, et_k^{CF})_{k=1}^N" + }, + { + "bbox": [ + 304, + 217, + 546, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 384, + 547, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 420 + ], + "type": "text", + "content": "Then, a contrastive loss " + }, + { + "bbox": [ + 304, + 384, + 547, + 420 + ], + "type": "inline_equation", + "content": "L_{count}" + }, + { + "bbox": [ + 304, + 384, + 547, + 420 + ], + "type": "text", + "content": " is computed to enforce that the similarity score of the image is high with the original caption and low with the counterfactual caption:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 437, + 545, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 545, + 479 + ], + "type": "interline_equation", + "content": "L _ {\\text {c o u n t}} = - \\frac {1}{N} \\sum_ {k = 1} ^ {N} \\log \\frac {\\exp \\left(e i _ {k} \\cdot e t _ {k}\\right)}{\\exp \\left(e i _ {k} \\cdot e t _ {k}\\right) + \\exp \\left(e i _ {k} \\cdot e t _ {k} ^ {C F}\\right)} \\tag {2}", + "image_path": "a6368cdc19dad0a120897545ea9120edbe2f2d0ba17157501bda84fb613848af.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 479, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 539 + ], + "type": "text", + "content": "Since the original ground truth caption and counterfactual caption differ only by the number of objects specified in them, this loss encourages the model to learn the relationship between the specified spelled number and the number of the objects it refers to." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "type": "text", + "content": "We use the negative samples only in the counting objective " + }, + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "type": "inline_equation", + "content": "L_{count}" + }, + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "type": "text", + "content": ", instead of adding them to the batch for the contrastive loss " + }, + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "type": "inline_equation", + "content": "L_{CLIP}" + }, + { + "bbox": [ + 304, + 539, + 545, + 575 + ], + "type": "text", + "content": " in order to better weight their effect." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 586, + 539, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 539, + 598 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 539, + 598 + ], + "type": "text", + "content": "4. \"CountBench\" - an evaluation benchmark" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "We introduce a new object-counting evaluation benchmark called CountBench. This benchmark was automatically curated (and manually verified) from the publicly available LAION-400M image-text dataset [45]. CountBench contains a total of 540 images containing between 2 and 10 instances of a particular object, where their corresponding captions reflect this number. This benchmark is used only for testing and evaluation. CountBench has no overlap with our training set " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3173" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 179, + 197 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 179, + 197 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 179, + 197 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 179, + 197 + ], + "type": "image", + "image_path": "60a53d01d6a9ce3c2b1c101e54229284346907e7b53a99bf0eed46556aee0dd5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 201, + 167, + 210 + ], + "lines": [ + { + "bbox": [ + 85, + 201, + 167, + 210 + ], + "spans": [ + { + "bbox": [ + 85, + 201, + 167, + 210 + ], + "type": "text", + "content": "(a) Baseline CLIP (L/14)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 69, + 289, + 198 + ], + "blocks": [ + { + "bbox": [ + 181, + 69, + 289, + 198 + ], + "lines": [ + { + "bbox": [ + 181, + 69, + 289, + 198 + ], + "spans": [ + { + "bbox": [ + 181, + 69, + 289, + 198 + ], + "type": "image", + "image_path": "f72c4513de1134111a33303371c9c4125b9deb7b71eb4fb0718f6d9ca0f9e3f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 201, + 261, + 210 + ], + "lines": [ + { + "bbox": [ + 211, + 201, + 261, + 210 + ], + "spans": [ + { + "bbox": [ + 211, + 201, + 261, + 210 + ], + "type": "text", + "content": "(b) Ours (L/14)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 294, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 426 + ], + "type": "text", + "content": "The images in CountBench were obtained by running our automatic filtering method described in Sec. 3.1 on the entire LAION-400M dataset. The filtering pipeline includes applying an object detector, whose benefits are twofold: (i) it facilitates verification of the object count specified in the caption, and (ii) it ensures that the counted objects are visible and detectable. The latter is important since benchmarks that do not disentangle the problem at hand from other challenges (e.g. occlusions and low resolution) may not reflect the performance on the actual task, since the difficulty can often stem from unrelated factors [14]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 559 + ], + "type": "text", + "content": "This filtering produced over 158K images for the number \"two\", but only around 100 images for \"ten\", demonstrating again the severe number imbalance we encountered with our training set. After automatically balancing each number to 100-200 samples each, the entire dataset was manually verified to contain only pairs in which the spelled number in the caption matches the number of clearly visible objects in the image. The dataset was rebalanced after this stage, ending up with 60 image-text pairs per number " + }, + { + "bbox": [ + 46, + 426, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\in\\{\\text{\"two\",\"..},\\text{\"ten\"}\\}" + }, + { + "bbox": [ + 46, + 426, + 287, + 559 + ], + "type": "text", + "content": ", 540 in total. Samples from the dataset can be seen in Fig. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 559, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 594 + ], + "type": "text", + "content": "We use the CountBench benchmark to evaluate the counting abilities of the models trained with our method in Sec. 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 609, + 128, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 128, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 128, + 623 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "We thoroughly evaluate our method, both quantitatively and qualitatively, on object counting-related tasks using our CountBench benchmark. We further validate that the performance of our counting-aware models on a variety of general zero-shot classification benchmarks is retained [8, 13, 15, 16, 19, 20, 21, 26, 33, 42, 55]. To gain a better understanding of our models, we show visualizations" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 371, + 127 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 371, + 127 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 371, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 371, + 127 + ], + "type": "image", + "image_path": "809322e4d93c9c1c9fcc88d8ca69afa6090ee5a3399ef0e8c40779bbf8074e6f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 128, + 363, + 143 + ], + "lines": [ + { + "bbox": [ + 316, + 128, + 363, + 143 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 363, + 143 + ], + "type": "text", + "content": "\"two zebras in Cape Town\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 373, + 71, + 457, + 127 + ], + "blocks": [ + { + "bbox": [ + 373, + 71, + 457, + 127 + ], + "lines": [ + { + "bbox": [ + 373, + 71, + 457, + 127 + ], + "spans": [ + { + "bbox": [ + 373, + 71, + 457, + 127 + ], + "type": "image", + "image_path": "b4bf294d4124d4f9a8c295c122d52bf63f60ce1bc63a3fc69efe0af171d6ffa0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 128, + 452, + 143 + ], + "lines": [ + { + "bbox": [ + 377, + 128, + 452, + 143 + ], + "spans": [ + { + "bbox": [ + 377, + 128, + 452, + 143 + ], + "type": "text", + "content": "\"three purebred chihuahuaas running...\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 460, + 71, + 545, + 127 + ], + "blocks": [ + { + "bbox": [ + 460, + 71, + 545, + 127 + ], + "lines": [ + { + "bbox": [ + 460, + 71, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 460, + 71, + 545, + 127 + ], + "type": "image", + "image_path": "e5eee1eaa287978d9ca669b958af5e975956b258f8c511dbff96408a952f7c1e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 128, + 542, + 143 + ], + "lines": [ + { + "bbox": [ + 462, + 128, + 542, + 143 + ], + "spans": [ + { + "bbox": [ + 462, + 128, + 542, + 143 + ], + "type": "text", + "content": "\"...four antique pink and purple armchairs\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 144, + 386, + 201 + ], + "blocks": [ + { + "bbox": [ + 307, + 144, + 386, + 201 + ], + "lines": [ + { + "bbox": [ + 307, + 144, + 386, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 144, + 386, + 201 + ], + "type": "image", + "image_path": "f7020c2a1501ee4efb6c80d611fcdf289fb86300e156a6925d8088aaa5f1f2d9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 202, + 379, + 217 + ], + "lines": [ + { + "bbox": [ + 315, + 202, + 379, + 217 + ], + "spans": [ + { + "bbox": [ + 315, + 202, + 379, + 217 + ], + "type": "text", + "content": "\"GT's five favourite Champagne...\"" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 391, + 144, + 475, + 201 + ], + "blocks": [ + { + "bbox": [ + 391, + 144, + 475, + 201 + ], + "lines": [ + { + "bbox": [ + 391, + 144, + 475, + 201 + ], + "spans": [ + { + "bbox": [ + 391, + 144, + 475, + 201 + ], + "type": "image", + "image_path": "fac855746b08ea777dc8fff3a873c7d6181ce5873289b6cbc2b3a02c18710ceb.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 202, + 471, + 217 + ], + "lines": [ + { + "bbox": [ + 396, + 202, + 471, + 217 + ], + "spans": [ + { + "bbox": [ + 396, + 202, + 471, + 217 + ], + "type": "text", + "content": "\"vintage silver plate tablespoons...set of six\"" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 475, + 202, + 545, + 217 + ], + "lines": [ + { + "bbox": [ + 475, + 202, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 475, + 202, + 545, + 217 + ], + "type": "text", + "content": "\"seven Chewy Peanut Butter Cookies...\"" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 477, + 144, + 545, + 201 + ], + "blocks": [ + { + "bbox": [ + 477, + 144, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 477, + 144, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 477, + 144, + 545, + 201 + ], + "type": "image", + "image_path": "af29610451d12236218eff88d75dd5dcdee5352eceee7d27533794583d1e246a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 307, + 219, + 387, + 277 + ], + "blocks": [ + { + "bbox": [ + 45, + 217, + 287, + 277 + ], + "lines": [ + { + "bbox": [ + 45, + 217, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 45, + 217, + 287, + 277 + ], + "type": "text", + "content": "Figure 3: Confusion matrices on CountBench. Classification accuracy on CountBench, broken down into confusion matrices for the public CLIP ViT-L/14 (a), and our counting-aware CLIP ViT-L/14 model (b), demonstrating clear quantitative superiority of our model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 219, + 387, + 277 + ], + "lines": [ + { + "bbox": [ + 307, + 219, + 387, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 387, + 277 + ], + "type": "image", + "image_path": "2edc89da46c157df84fecdb8ce32e2025531e54496d17b28ba7b6476a7e015a1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 277, + 385, + 294 + ], + "lines": [ + { + "bbox": [ + 309, + 277, + 385, + 294 + ], + "spans": [ + { + "bbox": [ + 309, + 277, + 385, + 294 + ], + "type": "text", + "content": "...eight colorful bright shiny red apples..." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 304, + 546, + 363 + ], + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 363 + ], + "type": "text", + "content": "Figure 4: CountBench benchmark. Sample images and their corresponding captions from our new CountBench object counting benchmark. This benchmark was automatically curated (and manually verified) from the publicly-available LAION-400M dataset." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 391, + 219, + 476, + 277 + ], + "blocks": [ + { + "bbox": [ + 391, + 219, + 476, + 277 + ], + "lines": [ + { + "bbox": [ + 391, + 219, + 476, + 277 + ], + "spans": [ + { + "bbox": [ + 391, + 219, + 476, + 277 + ], + "type": "image", + "image_path": "2b4d7081091cf46a15d607ee3a58790cd8c805c541c9f8571f26a7e17e6715a9.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 277, + 468, + 293 + ], + "lines": [ + { + "bbox": [ + 397, + 277, + 468, + 293 + ], + "spans": [ + { + "bbox": [ + 397, + 277, + 468, + 293 + ], + "type": "text", + "content": "\"The view of the nine leftmost moai...\"" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 481, + 219, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 481, + 219, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 481, + 219, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 481, + 219, + 545, + 277 + ], + "type": "image", + "image_path": "ad0fcb7059a51d822e0dc10016f65fdb018648ff4457989345d3075f4c6191f1.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 277, + 542, + 293 + ], + "lines": [ + { + "bbox": [ + 481, + 277, + 542, + 293 + ], + "spans": [ + { + "bbox": [ + 481, + 277, + 542, + 293 + ], + "type": "text", + "content": "\"All ten Christmas Poinsettia...\"" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 381, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 453 + ], + "type": "text", + "content": "of text-image relevancy maps, along with per-word relevancy scores, demonstrating that our model indeed attends to the number of objects in the image and text. Finally, we apply our model to text-to-image retrieval and generation, producing specific numbers of objects more reliably than baseline models." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 465, + 462, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 462, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 462, + 478 + ], + "type": "text", + "content": "5.1. Zero-shot counting accuracy" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 485, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 592 + ], + "type": "text", + "content": "We evaluate our models and baselines on CountBench on the task of classifying the number of objects in an image in a zero-shot manner. For each image in CountBench we augment the existing caption with eight other possible captions by replacing the number in its caption with all the numbers " + }, + { + "bbox": [ + 304, + 485, + 545, + 592 + ], + "type": "inline_equation", + "content": "\\in \\{\\text{two}, \\ldots, \\text{ten}\\}" + }, + { + "bbox": [ + 304, + 485, + 545, + 592 + ], + "type": "text", + "content": ", and calculate the similarity score between the image and each of the nine captions. The number in the caption that obtains highest similarity score with the image is considered the predicted number." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Table 1 reports the results of this evaluation on two prominent contrastive VLMs: CLIP-B/32 and BASIC-S. We report both the counting accuracy (selection of the correct number) and the mean deviation of the models' predictions from the correct numbers. For each of the architectures, we compare our model (configuration " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ") with two baseline configurations: (A) the official baseline model, and (B) the baseline model finetuned on our general text-image dataset used in our implementation, with the standard contrastive loss. Comparing the performance of these config" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3174" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 70, + 541, + 149 + ], + "blocks": [ + { + "bbox": [ + 54, + 70, + 541, + 149 + ], + "lines": [ + { + "bbox": [ + 54, + 70, + 541, + 149 + ], + "spans": [ + { + "bbox": [ + 54, + 70, + 541, + 149 + ], + "type": "table", + "html": "
CLIP-B/32BASIC-S
A Official BaselineB Internal BaselineC Ours (w/o Lcount)D Ours (Naive Filtering)E OursA Official BaselineB Internal BaselineC Ours (w/o Lcount)D Ours (Naive Filtering)E Ours
Accuracy ↑31.6732.9444.2649.8175.9317.9722.7530.5928.6869.02
Mean deviation from the correct number ↓1.531.440.971.280.492.132.021.291.870.64
", + "image_path": "5487f7e23302be46c5055ac4c71d9b951458dfa3f05a34640b51a96d8bae6b3c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 230, + 282, + 352 + ], + "blocks": [ + { + "bbox": [ + 46, + 154, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 46, + 154, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 547, + 214 + ], + "type": "text", + "content": "Table 1: Quantitative counting results. Top-1 zero-shot accuracy and the mean absolute distance between the predicted numbers and the true numbers on CountBench. We compare several configurations: (A) The official CLIP [38] and BASIC [36] models. (B) The official baselines finetuned on our internal curated data. (C) Models trained with our filtered counting set, without " + }, + { + "bbox": [ + 46, + 154, + 547, + 214 + ], + "type": "inline_equation", + "content": "L_{count}" + }, + { + "bbox": [ + 46, + 154, + 547, + 214 + ], + "type": "text", + "content": " (D) Models finetuned with " + }, + { + "bbox": [ + 46, + 154, + 547, + 214 + ], + "type": "inline_equation", + "content": "L_{count}" + }, + { + "bbox": [ + 46, + 154, + 547, + 214 + ], + "type": "text", + "content": " on a naively filtered counting set (E) Our method, which is significantly superior to all other configurations, both in accuracy and deviation from correct number." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 230, + 282, + 352 + ], + "lines": [ + { + "bbox": [ + 51, + 230, + 282, + 352 + ], + "spans": [ + { + "bbox": [ + 51, + 230, + 282, + 352 + ], + "type": "table", + "html": "
CLIP-B/32BASIC-S
Official BaselineInternal BaselineOursOfficial BaselineInternal BaselineOurs
ImageNet62.9364.9764.0659.7061.9661.18
CIFAR1063.9161.0060.6576.2284.6984.05
CIFAR10033.1032.4933.5645.3556.8055.89
Caltech10175.9982.5082.3678.1681.0381.05
EuroSAT45.2341.6637.6928.3945.8245.97
Food10183.0880.7280.5377.0877.8077.06
ImageNetA31.8530.8529.8117.6522.5521.68
ImageNetR69.3870.1770.3067.1167.6866.95
ImageNetV255.6556.5656.6252.2254.3553.60
Oxford Pets87.3587.7487.4180.6285.1584.87
Oxford Flowers66.1465.7367.3964.7466.4065.90
", + "image_path": "3baf9b6a1f4af6a0e2678ef3ae3a1d6abefe220273dcf73af28db96d17c93d14.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 355, + 287, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 403 + ], + "type": "text", + "content": "Table 2: Zero-shot accuracy on common benchmarks. We compare the zero-shot accuracy of our method and baselines on a variety of popular benchmarks. As seen, our method preserves the performance of the original model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 421, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 287, + 456 + ], + "type": "text", + "content": "urations allows us to quantify the effect of using our own large-scale text-image dataset, which differs from the original unpublished data the models were trained on." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "type": "text", + "content": "As can be seen, our method " + }, + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "type": "inline_equation", + "content": "(E)" + }, + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "type": "text", + "content": " achieves significantly superior counting accuracy compared to the baselines " + }, + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "type": "inline_equation", + "content": "(A, B)" + }, + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "type": "text", + "content": ", attaining " + }, + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "type": "inline_equation", + "content": "2 - 3 \\times" + }, + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "type": "text", + "content": " higher counting accuracy and more than " + }, + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 460, + 287, + 520 + ], + "type": "text", + "content": " lower mean deviation from the correct number. Results for different categories in CountBench are reported in App. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": "Tab. 1 also contains an ablation of the two components of our method: filtering a counting training set and finetuning with an additional loss " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "L_{count}" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": ". Models with configuration " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " are finetuned on the filtered subset with no counting loss. The large gap in accuracy on CountBench between configurations " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " shows the importance of our loss for the improvement in counting abilities. Models with configuration " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " are finetuned with the counting loss " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "L_{count}" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " on an alternative counting subset, which consists of all the samples that contain spelled numbers " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\in \\{ \"two\",.., \"ten\" \\}" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " without additional filtering. The significant difference in counting accuracy between configurations " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " demonstrates the importance of our restrictive filtering pipeline, as both configurations are finetuned with " + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "inline_equation", + "content": "L_{count}" + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": " over the samples from a dedicated counting training set. As can be seen in Tab. 1, while the naively filtered data does improve performance" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 226, + 389, + 279 + ], + "blocks": [ + { + "bbox": [ + 318, + 226, + 389, + 279 + ], + "lines": [ + { + "bbox": [ + 318, + 226, + 389, + 279 + ], + "spans": [ + { + "bbox": [ + 318, + 226, + 389, + 279 + ], + "type": "image", + "image_path": "88ed71bec5f49597fb8e758243758382837c7068e11280ad99421afd496216da.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 390, + 226, + 460, + 278 + ], + "blocks": [ + { + "bbox": [ + 390, + 226, + 460, + 278 + ], + "lines": [ + { + "bbox": [ + 390, + 226, + 460, + 278 + ], + "spans": [ + { + "bbox": [ + 390, + 226, + 460, + 278 + ], + "type": "image", + "image_path": "d77c3f1fea7e20d39b76477ea44e25ebdc61b08744524187df25655b93fdfe5a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 278, + 448, + 292 + ], + "lines": [ + { + "bbox": [ + 402, + 278, + 448, + 292 + ], + "spans": [ + { + "bbox": [ + 402, + 278, + 448, + 292 + ], + "type": "text", + "content": "\"Two ducks\" 0.23 0.77" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 461, + 226, + 531, + 278 + ], + "blocks": [ + { + "bbox": [ + 461, + 226, + 531, + 278 + ], + "lines": [ + { + "bbox": [ + 461, + 226, + 531, + 278 + ], + "spans": [ + { + "bbox": [ + 461, + 226, + 531, + 278 + ], + "type": "image", + "image_path": "9bc28b3af8480c287b33546b2bce9017a574af90c90c41f5dcfa86257fcc9e08.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 278, + 520, + 292 + ], + "lines": [ + { + "bbox": [ + 473, + 278, + 520, + 292 + ], + "spans": [ + { + "bbox": [ + 473, + 278, + 520, + 292 + ], + "type": "text", + "content": "\"Two ducks\" 0.45 0.55" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 295, + 389, + 338 + ], + "blocks": [ + { + "bbox": [ + 318, + 295, + 389, + 338 + ], + "lines": [ + { + "bbox": [ + 318, + 295, + 389, + 338 + ], + "spans": [ + { + "bbox": [ + 318, + 295, + 389, + 338 + ], + "type": "image", + "image_path": "3cf24cb2e1961fc90d2f81a01ace2126be093af29398a0f837802ff4f38369cb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 390, + 295, + 460, + 338 + ], + "blocks": [ + { + "bbox": [ + 390, + 295, + 460, + 338 + ], + "lines": [ + { + "bbox": [ + 390, + 295, + 460, + 338 + ], + "spans": [ + { + "bbox": [ + 390, + 295, + 460, + 338 + ], + "type": "image", + "image_path": "69547da3ae2ee280ea58edeb7e3ba2b83426ef2c89e51ff0f73d63fdecef3fa6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 338, + 459, + 352 + ], + "lines": [ + { + "bbox": [ + 391, + 338, + 459, + 352 + ], + "spans": [ + { + "bbox": [ + 391, + 338, + 459, + 352 + ], + "type": "text", + "content": "\"Three pumpkins\" 0.16 0.84" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 461, + 295, + 531, + 338 + ], + "blocks": [ + { + "bbox": [ + 461, + 295, + 531, + 338 + ], + "lines": [ + { + "bbox": [ + 461, + 295, + 531, + 338 + ], + "spans": [ + { + "bbox": [ + 461, + 295, + 531, + 338 + ], + "type": "image", + "image_path": "25f91edf186465f58417bbebbe5a24a8956fc030015c019bc7f59f7d92ec3542.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 338, + 529, + 352 + ], + "lines": [ + { + "bbox": [ + 461, + 338, + 529, + 352 + ], + "spans": [ + { + "bbox": [ + 461, + 338, + 529, + 352 + ], + "type": "text", + "content": "\"Three pumpkins\" 0.45 0.55" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 319, + 355, + 389, + 397 + ], + "blocks": [ + { + "bbox": [ + 319, + 355, + 389, + 397 + ], + "lines": [ + { + "bbox": [ + 319, + 355, + 389, + 397 + ], + "spans": [ + { + "bbox": [ + 319, + 355, + 389, + 397 + ], + "type": "image", + "image_path": "4b208bdb5d1566e2206f404ce99325a0cdb6408779e278fa188ad8e2e83af8eb.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 414, + 547, + 498 + ], + "lines": [ + { + "bbox": [ + 304, + 414, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 547, + 498 + ], + "type": "text", + "content": "Figure 5: Relevancy maps of both image and text. Visualization of the relevancy maps of both image and text, which represent, for each patch in the image and token in the text, how important it is to the prediction. Using our counting-aware CLIP model, the relevancy of the number (e.g., \"four\") in the text is increased. In addition, the model focuses on areas in the image that are relevant for counting." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 390, + 355, + 460, + 397 + ], + "blocks": [ + { + "bbox": [ + 390, + 355, + 460, + 397 + ], + "lines": [ + { + "bbox": [ + 390, + 355, + 460, + 397 + ], + "spans": [ + { + "bbox": [ + 390, + 355, + 460, + 397 + ], + "type": "image", + "image_path": "36002ba95075fead18bba537240aefdb39b32682285019ba140700f92dc5c5fe.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 397, + 446, + 411 + ], + "lines": [ + { + "bbox": [ + 402, + 397, + 446, + 411 + ], + "spans": [ + { + "bbox": [ + 402, + 397, + 446, + 411 + ], + "type": "text", + "content": "Four cards 0.30 0.70" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 461, + 355, + 531, + 397 + ], + "blocks": [ + { + "bbox": [ + 461, + 355, + 531, + 397 + ], + "lines": [ + { + "bbox": [ + 461, + 355, + 531, + 397 + ], + "spans": [ + { + "bbox": [ + 461, + 355, + 531, + 397 + ], + "type": "image", + "image_path": "a86317d09edd9c60f83620a0061d575f796add80a7392b018526cbf4903c2bb1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 397, + 521, + 411 + ], + "lines": [ + { + "bbox": [ + 473, + 397, + 521, + 411 + ], + "spans": [ + { + "bbox": [ + 473, + 397, + 521, + 411 + ], + "type": "text", + "content": "\"Four cards\" 0.43 0.57" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 508, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 581 + ], + "type": "text", + "content": "over a baseline trained without a dedicated counting subset, the obtained results are still significantly lower than those produced by our model. We attribute this gap in performance to mislabeled training samples in the naively filtered data, which are absent from our counting training set " + }, + { + "bbox": [ + 304, + 508, + 545, + 581 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 508, + 545, + 581 + ], + "type": "text", + "content": " due to our filtering pipeline." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": "Confusion matrices for the counting evaluation described above are shown in Fig. 3. For this experiment, we compare a public CLIP-L/14 model against our counting-aware version of it. As can be seen, our counting-aware CLIP model is significantly superior to the baseline across all numbers. It is worth noting that the baseline CLIP performs reasonably well for numbers within the subitizing range " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "(2 - 4)" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": ", but fails on larger numbers. This supports our claim that curated image captions with numbers beyond this range are rare, since specifying them requires manually counting the objects." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3175" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 64, + 286, + 152 + ], + "blocks": [ + { + "bbox": [ + 50, + 64, + 286, + 152 + ], + "lines": [ + { + "bbox": [ + 50, + 64, + 286, + 152 + ], + "spans": [ + { + "bbox": [ + 50, + 64, + 286, + 152 + ], + "type": "image", + "image_path": "9b288a39429169ee78a73b8e0159f6b682a0ec79bb79996a0d89b3ccf90dcfdd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 154, + 288, + 227 + ], + "lines": [ + { + "bbox": [ + 46, + 154, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 288, + 227 + ], + "type": "text", + "content": "Figure 6: Relevancy maps for similarity between the image and different numbers. We compare the relevancy map of the input image with text prompts of several numbers (i.e. four to six) for both the baseline CLIP and our counting-aware CLIP. Our model focuses on the five stars when calculating the similarity with the prompt \"five\"." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 238, + 203, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 203, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 203, + 251 + ], + "type": "text", + "content": "5.2. Count-based image retrieval" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 256, + 287, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 352 + ], + "type": "text", + "content": "We consider the task of text-to-image retrieval where the text explicitly describes the desired count of objects. To obtain a diverse dataset that consists of varied numbers of objects, yet facilitates retrieval in reasonable time, we split the public LAION-400M dataset [45] into coarse categorical subsets by filtering samples where the caption contains a certain word (e.g., \"dogs\", \"animals\", \"cars\"), and perform retrieval on each of these subsets separately." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 353, + 289, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 289, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 289, + 569 + ], + "type": "text", + "content": "For each category, we use the caption \"a photo of \" where number " + }, + { + "bbox": [ + 46, + 353, + 289, + 569 + ], + "type": "inline_equation", + "content": "\\in" + }, + { + "bbox": [ + 46, + 353, + 289, + 569 + ], + "type": "text", + "content": " {\"two\",..,\"ten\"} (e.g. \"a photo of six dogs\"). For each caption, we retrieve the three images in the dataset that are predicted by the model to be most similar to the caption. Note that since there are no ground truth labels for the counts of objects, we present qualitative results. Fig. 7 shows the retrieved images using the original CLIP model and our counting-aware CLIP model. An extended figure can be found in the supplementary. As can be seen, when the requested number is larger than three, the images retrieved by the baseline model often depict arbitrary numbers of objects. Additionally, the baseline often retrieves the same images for several different requested numbers. This further implies that the baseline model mostly focuses on the existence of the described object in the image, and ignores the number in the caption. In contrast, our results depict accurate object counts in most cases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 575, + 272, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 272, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 272, + 588 + ], + "type": "text", + "content": "5.3. Performance on non-counting classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": "To verify that our counting-aware models preserve the powerful image-text representation capabilities of the original models, we evaluate the zero-shot performance of our models on a variety of common classification benchmarks. Table 2 reports the zero-shot accuracy of our counting aware models against the baselines (corresponding to configurations " + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": " in Tab. 1). As can be seen, our models maintain similar overall accuracy. Additionally, comparing the official baseline and the internal baseline indicates that finetuning the models on our general text-image datasets leads to" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 81, + 545, + 401 + ], + "blocks": [ + { + "bbox": [ + 376, + 64, + 482, + 77 + ], + "lines": [ + { + "bbox": [ + 376, + 64, + 482, + 77 + ], + "spans": [ + { + "bbox": [ + 376, + 64, + 482, + 77 + ], + "type": "text", + "content": "\"A photo of {} shirts.\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 81, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 310, + 81, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 310, + 81, + 545, + 401 + ], + "type": "image", + "image_path": "6c953c2ed96e8d820c8bb4adc5cafd6d49100ff1e5e582ea8f5575099330d4fb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 405, + 547, + 477 + ], + "lines": [ + { + "bbox": [ + 304, + 405, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 547, + 477 + ], + "type": "text", + "content": "Figure 7: Top-3 count-based image retrieval Text-to-image retrieval results for different counts of shirts, ordered from left to right according to their similarity score (descending). Images that match the caption are marked with " + }, + { + "bbox": [ + 304, + 405, + 547, + 477 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 304, + 405, + 547, + 477 + ], + "type": "text", + "content": " and images that do not match it are marked with X. Extended results are in the supplementary." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 491, + 546, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 546, + 529 + ], + "type": "text", + "content": "only a slight shift in the accuracy of the models on common benchmarks. An additional evaluation on the task of general zero-shot image retrieval is reported in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 538, + 462, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 462, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 462, + 552 + ], + "type": "text", + "content": "5.4. Relevancy map visualization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "To gain a better understanding of what our model learns, we use the explainability method of Chefer et al. [10] to visualize relevancy maps, which indicate the importance of different parts of the text and image to the model's similarity score prediction. Fig. 5 displays the normalized relevancy maps of several image-text pairs. Examining the relevancy maps of the text, it is apparent that the relevancy score of the spelled number in the caption is significantly higher for our model than the baseline model, which suggests that our model concentrates more on the mentioned number than the original one. Additionally, examining the relevancy maps of the images, it is evident that our model focuses on all pertinent objects in the image, whereas the original model" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3176" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 84, + 104, + 123 + ], + "blocks": [ + { + "bbox": [ + 53, + 84, + 104, + 123 + ], + "lines": [ + { + "bbox": [ + 53, + 84, + 104, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 84, + 104, + 123 + ], + "type": "image", + "image_path": "2c97bf8ffdabe56d3ed450c20f7de34cfc2f33c62c97b2998b11bb126d4da6f2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 106, + 84, + 157, + 123 + ], + "blocks": [ + { + "bbox": [ + 116, + 70, + 225, + 84 + ], + "lines": [ + { + "bbox": [ + 116, + 70, + 225, + 84 + ], + "spans": [ + { + "bbox": [ + 116, + 70, + 225, + 84 + ], + "type": "text", + "content": "\"{} colorful tulips\"" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 106, + 84, + 157, + 123 + ], + "lines": [ + { + "bbox": [ + 106, + 84, + 157, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 84, + 157, + 123 + ], + "type": "image", + "image_path": "e0001ba9b2ed6a87b7e609a22d9c4f2d9b8a8cc28900ae6f10c140e3d6258482.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 159, + 84, + 210, + 123 + ], + "blocks": [ + { + "bbox": [ + 159, + 84, + 210, + 123 + ], + "lines": [ + { + "bbox": [ + 159, + 84, + 210, + 123 + ], + "spans": [ + { + "bbox": [ + 159, + 84, + 210, + 123 + ], + "type": "image", + "image_path": "ebe34c920450ea7c10831de000a7d5cd2520856fb71157c21d005f885dd837b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 212, + 84, + 244, + 123 + ], + "blocks": [ + { + "bbox": [ + 212, + 84, + 244, + 123 + ], + "lines": [ + { + "bbox": [ + 212, + 84, + 244, + 123 + ], + "spans": [ + { + "bbox": [ + 212, + 84, + 244, + 123 + ], + "type": "image", + "image_path": "426d15c5a0c64de4dfbb85d6b1148ea3d66c789081fcc85f07ff3d4f821308b4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 246, + 84, + 294, + 123 + ], + "blocks": [ + { + "bbox": [ + 246, + 84, + 294, + 123 + ], + "lines": [ + { + "bbox": [ + 246, + 84, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 246, + 84, + 294, + 123 + ], + "type": "image", + "image_path": "e91967d31ccb05b26080542fe688696ae515cfc785cca4405ceb17c35b79b63e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 53, + 129, + 91, + 163 + ], + "blocks": [ + { + "bbox": [ + 53, + 129, + 91, + 163 + ], + "lines": [ + { + "bbox": [ + 53, + 129, + 91, + 163 + ], + "spans": [ + { + "bbox": [ + 53, + 129, + 91, + 163 + ], + "type": "image", + "image_path": "0bc615c021b687a209861375174b46f94e9ec0e774b8d380d775da1d81298f0a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 163, + 79, + 169 + ], + "lines": [ + { + "bbox": [ + 61, + 163, + 79, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 79, + 169 + ], + "type": "text", + "content": "\"six\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 176, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 45, + 176, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 45, + 176, + 547, + 213 + ], + "type": "text", + "content": "Figure 8: Generated samples with Imagen using our counting-aware CLIP as backbone. The Imagen model benefits from the counting-aware representations produced by the our CLIP model, and is able to generate images that accurately follow the amounts specifies in the captions." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 93, + 129, + 142, + 163 + ], + "blocks": [ + { + "bbox": [ + 93, + 129, + 142, + 163 + ], + "lines": [ + { + "bbox": [ + 93, + 129, + 142, + 163 + ], + "spans": [ + { + "bbox": [ + 93, + 129, + 142, + 163 + ], + "type": "image", + "image_path": "db909db848b4f13445c4831d1595874f0455457fc232f3e9a674f867c06d46da.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 163, + 130, + 170 + ], + "lines": [ + { + "bbox": [ + 100, + 163, + 130, + 170 + ], + "spans": [ + { + "bbox": [ + 100, + 163, + 130, + 170 + ], + "type": "text", + "content": "\"seven\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 143, + 129, + 185, + 163 + ], + "blocks": [ + { + "bbox": [ + 143, + 129, + 185, + 163 + ], + "lines": [ + { + "bbox": [ + 143, + 129, + 185, + 163 + ], + "spans": [ + { + "bbox": [ + 143, + 129, + 185, + 163 + ], + "type": "image", + "image_path": "ce72b3daa2f8576c510945ef95efe9da054c57d935a7ca5c39596912c2139328.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 163, + 179, + 170 + ], + "lines": [ + { + "bbox": [ + 151, + 163, + 179, + 170 + ], + "spans": [ + { + "bbox": [ + 151, + 163, + 179, + 170 + ], + "type": "text", + "content": "\"eight\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 187, + 129, + 241, + 162 + ], + "blocks": [ + { + "bbox": [ + 187, + 129, + 241, + 162 + ], + "lines": [ + { + "bbox": [ + 187, + 129, + 241, + 162 + ], + "spans": [ + { + "bbox": [ + 187, + 129, + 241, + 162 + ], + "type": "image", + "image_path": "65c1382adcc99112df7c077cde483b17a94438615e241eb15b235c073f30de74.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 163, + 223, + 170 + ], + "lines": [ + { + "bbox": [ + 200, + 163, + 223, + 170 + ], + "spans": [ + { + "bbox": [ + 200, + 163, + 223, + 170 + ], + "type": "text", + "content": "\"nine\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 243, + 129, + 295, + 162 + ], + "blocks": [ + { + "bbox": [ + 243, + 129, + 295, + 162 + ], + "lines": [ + { + "bbox": [ + 243, + 129, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 243, + 129, + 295, + 162 + ], + "type": "image", + "image_path": "e87c9cdbe4c81ff5d3134626a484e351b6117e00a05e29131994a2e1b2eac164.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 163, + 279, + 170 + ], + "lines": [ + { + "bbox": [ + 260, + 163, + 279, + 170 + ], + "spans": [ + { + "bbox": [ + 260, + 163, + 279, + 170 + ], + "type": "text", + "content": "\"ten\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 306, + 85, + 339, + 123 + ], + "blocks": [ + { + "bbox": [ + 306, + 85, + 339, + 123 + ], + "lines": [ + { + "bbox": [ + 306, + 85, + 339, + 123 + ], + "spans": [ + { + "bbox": [ + 306, + 85, + 339, + 123 + ], + "type": "image", + "image_path": "32b94397b9daf17b2b617470b432e3770b095bff07386d0e07abb93d2fbaa4aa.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 340, + 86, + 372, + 123 + ], + "blocks": [ + { + "bbox": [ + 340, + 86, + 372, + 123 + ], + "lines": [ + { + "bbox": [ + 340, + 86, + 372, + 123 + ], + "spans": [ + { + "bbox": [ + 340, + 86, + 372, + 123 + ], + "type": "image", + "image_path": "4af404bbe7405baa102568802635fffcb01ef2b2f6a3fd907aed01a080d827d5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 373, + 86, + 417, + 123 + ], + "blocks": [ + { + "bbox": [ + 373, + 86, + 417, + 123 + ], + "lines": [ + { + "bbox": [ + 373, + 86, + 417, + 123 + ], + "spans": [ + { + "bbox": [ + 373, + 86, + 417, + 123 + ], + "type": "image", + "image_path": "f83638fe5c5d627631142fccc209b8cc667401c5bfb2a8364a57b21b95b6af94.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 417, + 85, + 450, + 123 + ], + "blocks": [ + { + "bbox": [ + 392, + 72, + 454, + 85 + ], + "lines": [ + { + "bbox": [ + 392, + 72, + 454, + 85 + ], + "spans": [ + { + "bbox": [ + 392, + 72, + 454, + 85 + ], + "type": "text", + "content": "{}lamps" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 85, + 450, + 123 + ], + "lines": [ + { + "bbox": [ + 417, + 85, + 450, + 123 + ], + "spans": [ + { + "bbox": [ + 417, + 85, + 450, + 123 + ], + "type": "image", + "image_path": "765b6d2321f0d93e25a9406b8d8994b7613ac292dc4423f35c35fdeebc84c074.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 450, + 85, + 504, + 123 + ], + "blocks": [ + { + "bbox": [ + 450, + 85, + 504, + 123 + ], + "lines": [ + { + "bbox": [ + 450, + 85, + 504, + 123 + ], + "spans": [ + { + "bbox": [ + 450, + 85, + 504, + 123 + ], + "type": "image", + "image_path": "180a81614f695a9d10633dc1bd3e2e071ed5ce9e975be81d420b52be09d7fe01.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 505, + 86, + 538, + 123 + ], + "blocks": [ + { + "bbox": [ + 505, + 86, + 538, + 123 + ], + "lines": [ + { + "bbox": [ + 505, + 86, + 538, + 123 + ], + "spans": [ + { + "bbox": [ + 505, + 86, + 538, + 123 + ], + "type": "image", + "image_path": "a450210894305ea521f3a9b8f81268d9ec90ce5a60e102f3e0ad8ba5bc765ad4.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 306, + 129, + 338, + 163 + ], + "blocks": [ + { + "bbox": [ + 306, + 129, + 338, + 163 + ], + "lines": [ + { + "bbox": [ + 306, + 129, + 338, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 129, + 338, + 163 + ], + "type": "image", + "image_path": "af9d53653afef40bc2ef1f6f8484ca738f5ea8277b792ff9f8b91f76dcfa6ec9.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 163, + 331, + 170 + ], + "lines": [ + { + "bbox": [ + 312, + 163, + 331, + 170 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 331, + 170 + ], + "type": "text", + "content": "\"six\"" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 339, + 129, + 414, + 162 + ], + "blocks": [ + { + "bbox": [ + 339, + 129, + 414, + 162 + ], + "lines": [ + { + "bbox": [ + 339, + 129, + 414, + 162 + ], + "spans": [ + { + "bbox": [ + 339, + 129, + 414, + 162 + ], + "type": "image", + "image_path": "394468bffc7083cbb7d1b2dda123a7751ae18eca03374945b5a6c83026b3cf42.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 163, + 392, + 170 + ], + "lines": [ + { + "bbox": [ + 362, + 163, + 392, + 170 + ], + "spans": [ + { + "bbox": [ + 362, + 163, + 392, + 170 + ], + "type": "text", + "content": "\"seven\"" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 414, + 129, + 459, + 162 + ], + "blocks": [ + { + "bbox": [ + 414, + 129, + 459, + 162 + ], + "lines": [ + { + "bbox": [ + 414, + 129, + 459, + 162 + ], + "spans": [ + { + "bbox": [ + 414, + 129, + 459, + 162 + ], + "type": "image", + "image_path": "be9fd6733a0a700645d1e2f720f0e9e5cc9539b8b034f9b7e0285521049cd0ad.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 163, + 449, + 170 + ], + "lines": [ + { + "bbox": [ + 422, + 163, + 449, + 170 + ], + "spans": [ + { + "bbox": [ + 422, + 163, + 449, + 170 + ], + "type": "text", + "content": "\"eight\"" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 459, + 129, + 492, + 162 + ], + "blocks": [ + { + "bbox": [ + 459, + 129, + 492, + 162 + ], + "lines": [ + { + "bbox": [ + 459, + 129, + 492, + 162 + ], + "spans": [ + { + "bbox": [ + 459, + 129, + 492, + 162 + ], + "type": "image", + "image_path": "52030ed1d203acc98d76eee8221b8a8e2b112fea4a1e55d4373d67e830d5d514.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 163, + 487, + 170 + ], + "lines": [ + { + "bbox": [ + 463, + 163, + 487, + 170 + ], + "spans": [ + { + "bbox": [ + 463, + 163, + 487, + 170 + ], + "type": "text", + "content": "\"nine\"" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 492, + 129, + 539, + 162 + ], + "blocks": [ + { + "bbox": [ + 492, + 129, + 539, + 162 + ], + "lines": [ + { + "bbox": [ + 492, + 129, + 539, + 162 + ], + "spans": [ + { + "bbox": [ + 492, + 129, + 539, + 162 + ], + "type": "image", + "image_path": "aae2c2ac76cf3623d30f5c59e8730f7564d25830c9b608b9bcce0e398e7cb0d7.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 163, + 523, + 170 + ], + "lines": [ + { + "bbox": [ + 504, + 163, + 523, + 170 + ], + "spans": [ + { + "bbox": [ + 504, + 163, + 523, + 170 + ], + "type": "text", + "content": "\"ten\"" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 228, + 286, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 286, + 239 + ], + "type": "text", + "content": "primarily identifies a single instance of the described object." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 46, + 240, + 286, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 286, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 286, + 347 + ], + "type": "text", + "content": "To verify that our model does not simply attend to all objects that appear in the image, we examined the relevancy maps in Fig. 6 using negative prompts (i.e. \"four\" when there are five elements in the image). Our model focuses only on relevant objects when the correct number is used, unlike the baseline CLIP that highlights all object types in the image. This demonstrates that our model learns to associate the number in the caption with the suitable number of objects, and does not exploit shortcuts or undesired content." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 47, + 354, + 247, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 247, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 247, + 367 + ], + "type": "text", + "content": "5.5. Count-based text-to-image generation" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 46, + 373, + 286, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 286, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 286, + 528 + ], + "type": "text", + "content": "Text-to-image generative models that rely on CLIP for representations, are known to inherit its limitations [43, 52], thus fail to reliably produce specific counts of objects. In order to demonstrate the effectiveness of our counting-aware model for downstream image generation tasks, we train an Imagen [43] model conditioned on the textual embeddings of a pretrained CLIP-L/14, and another Imagen model conditioned on our counting-aware version of this model. To compare our model and the baseline, we synthesize 12 samples for each prompt in the counting category of the Draw-Bench benchmark [43]. For each sample, we check whether or not it contains the requested number of objects, as stated in its prompt." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 46, + 529, + 286, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 286, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 286, + 575 + ], + "type": "text", + "content": "The Imagen model trained with the baseline CLIP achieves a binary accuracy of " + }, + { + "bbox": [ + 46, + 529, + 286, + 575 + ], + "type": "inline_equation", + "content": "24.12\\%" + }, + { + "bbox": [ + 46, + 529, + 286, + 575 + ], + "type": "text", + "content": ", while the Imagen trained with out counting-aware model reaches " + }, + { + "bbox": [ + 46, + 529, + 286, + 575 + ], + "type": "inline_equation", + "content": "40.35\\%" + }, + { + "bbox": [ + 46, + 529, + 286, + 575 + ], + "type": "text", + "content": " accuracy. See the supplementary for additional evaluations." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "Performance on non-counting image generation: To verify that the generation quality is preserved for non-counting captions, we conducted a user study. We presented 52 participants with 20 pairs of images, each containing one image generated with an Imagen model trained with the baseline CLIP, and one with our counting-aware CLIP as backbone. The same random prompt from COCO [28] and seed were used for each pair. Participants were asked which image is better in terms of quality and realism. " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "54.7\\%" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " of ratings favored our images, suggesting that our model does not degrade generation quality on general text captions." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 306, + 228, + 382, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 382, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 382, + 239 + ], + "type": "text", + "content": "5.6. Limitations" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 246, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 449 + ], + "type": "text", + "content": "Our method is limited by the insufficient existence of training data with images containing multiple instances of an object, along with a corresponding caption that correctly spells out this information. The effect of this data scarcity on our method increases with larger numbers (7, 8, etc.), as people tend to use \"a group of\" or \"many\" for large numbers of objects, instead of gruelingly counting them. Furthermore, many of the correct training pairs with higher numbers that do exist, contain relatively simplistic 2D collections of objects, as opposed to objects in a real-world scene. This is also the reason why our current method teaches CLIP to count only up to ten. Mitigating the above-mentioned data limitations, and generalizing to counting beyond 10, can possibly be achieved by concatenating multiple images with smaller numbers of objects. This will result in new training examples with more than ten natural-looking objects, and is part of our future work." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 306, + 466, + 383, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 383, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 383, + 478 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "We present the first method to enhance VLMs with counting capabilities, while maintaining their overall performance on common benchmarks. This is an essential step towards enabling more accurate text-based image retrieval, classification and generation. A new counting-based contrastive loss is used to finetune a pre-trained VLM in tandem with its original objective. Our counting loss is deployed over automatically-created counterfactual examples, each consisting of an image and a caption containing an incorrect object count. We demonstrate a significant improvement over state-of-the-art baseline models in multiple tasks (classification, retrieval and generation) on several different datasets. Furthermore, we introduce CountBench - a new image-text counting benchmark for evaluating VLMs' object counting capabilities. This new benchmark contains in-the-wild images and captions, where the object counts are specified in the captions. Finally, our approach is not specific to counting, and can be generalized to other compositional concepts that VLMs fail to learn." + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3177" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In AAAI Conference on Artificial Intelligence, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 233 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 233 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. ArXiv, abs/2204.14198, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 235, + 286, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 235, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 235, + 286, + 267 + ], + "type": "text", + "content": "[3] Jong Wook Kim Gretchen Krueger Sandhini Agarwal Alec Radford, Ilya Sutskever. Clip: Connecting text and images. https://openai.com/research/clip, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 286, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 286, + 312 + ], + "type": "text", + "content": "[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18187-18197, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 286, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 286, + 376 + ], + "type": "text", + "content": "[5] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 286, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 412 + ], + "type": "text", + "content": "[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. ArXiv, abs/2204.02491, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 414, + 286, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 286, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 286, + 446 + ], + "type": "text", + "content": "[7] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 286, + 479 + ], + "type": "text", + "content": "[8] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, 2014. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 481, + 286, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 286, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 286, + 512 + ], + "type": "text", + "content": "[9] Hila Chefer, Sagie Benaim, Roni Paiss, and Lior Wolf. Image-based clip-guided essence transfer. In ECCV, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 286, + 567 + ], + "type": "text", + "content": "[10] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 387-396, 2021. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 286, + 678 + ], + "type": "text", + "content": "[11] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiiovanni, Piotr Padlewski, Daniel M. Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model. ArXiv, abs/2209.06794, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "text", + "content": "[12] Katherine Crowson. Vqgan+clip. https://colab.research.google.com/drive/1L8oL-vLJXVcRzCFbPwOoMkPKJ8-aYdPN, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "text", + "content": "[14] Anuj Diwan, Layne Berry, Eunsol Choi, David F. Harwath, and Kyle Mahowald. Why is winoground hard? investigating failures in visuolinguistic compositionality. ArXiv, abs/2211.00768, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 164, + 545, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 197 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 197 + ], + "type": "text", + "content": "[15] Raveen Doon, Tarun Kumar Rawat, and Shweta Gautam. Cifar-10 classification using deep convolutional neural network. 2018 IEEE Punecon, pages 1–5, 2018. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 199, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 199, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 308, + 199, + 545, + 253 + ], + "type": "text", + "content": "[16] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. 2004 Conference on Computer Vision and Pattern Recognition Workshop, pages 178–178, 2004. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 256, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 256, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 308, + 256, + 545, + 289 + ], + "type": "text", + "content": "[17] Rinon Gal, Or Patashnik, Haggai Maron, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada. ACM Transactions on Graphics (TOG), 41:1 - 13, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 291, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 291, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 308, + 291, + 545, + 323 + ], + "type": "text", + "content": "[18] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arxiv:2307.10373, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 326, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 326, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 545, + 380 + ], + "type": "text", + "content": "[19] Patrick Helber, Benjamin Bischke, Andreas R. Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12:2217-2226, 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 382, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 545, + 458 + ], + "type": "text", + "content": "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadayath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, October 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 492 + ], + "type": "text", + "content": "[21] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. CVPR, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 495, + 545, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 495, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 308, + 495, + 545, + 550 + ], + "type": "text", + "content": "[22] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 552, + 545, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 552, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 308, + 552, + 545, + 573 + ], + "type": "text", + "content": "[23] William Stanley Jevons. The power of numerical discrimination. Nature, 3:281-282, 1871. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 576, + 545, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 576, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 308, + 576, + 545, + 608 + ], + "type": "text", + "content": "[24] E. L. Kaufman and Miles W. Lord. The discrimination of visual number. The American journal of psychology, 62 4:498-525, 1949. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "type": "text", + "content": "[25] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 657, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 657, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 657, + 545, + 677 + ], + "type": "text", + "content": "[26] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[27] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. ArXiv, abs/2201.03546, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3178" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[29] Xingchao Liu, Chengyue Gong, Lemeng Wu, Shujian Zhang, Haoran Su, and Qiang Liu. Fusedream: Training-free text-to-image generation with improved clip+gan space optimization. ArXiv, abs/2112.01573, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[30] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 231 + ], + "type": "text", + "content": "[31] Duy-Kien Nguyen, Vedanuj Goswami, and Xinlei Chen. Movie: Revisiting modulated convolutions. 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "type": "text", + "content": "[32] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "type": "text", + "content": "[33] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "type": "text", + "content": "[34] Roni Paiss, Hila Chefer, and Lior Wolf. No token left behind: Explainability-aided image classification and generation. In ECCV, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 367, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 367, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 367, + 287, + 421 + ], + "type": "text", + "content": "[35] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085–2094, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 423, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 287, + 488 + ], + "type": "text", + "content": "[36] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, Mingxing Tan, and Quoc V. Le. Combined Scaling for Open-Vocabulary Image Classification. arXiv preprint arXiv:2111.10050, Nov. 2021. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 566 + ], + "type": "text", + "content": "[37] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6967-6977, June 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 634 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 193 + ], + "type": "text", + "content": "[42] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do ImageNet classifiers generalize to ImageNet? In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 5389-5400. PMLR, 09-15 Jun 2019. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 194, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 270 + ], + "type": "text", + "content": "[43] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, 2022. 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 272, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 348 + ], + "type": "text", + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 349, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 403 + ], + "type": "text", + "content": "[45] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 469 + ], + "type": "text", + "content": "[46] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15617-15629, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 471, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 513 + ], + "type": "text", + "content": "[47] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "text", + "content": "[48] Yoad Tewel, Yoav Shalev, Idan Schwartz, and Lior Wolf. Zero-shot image-to-text generation for visual-semantic arithmetic. arXiv preprint arXiv:2111.14447, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "text", + "content": "[49] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Bachmann, Amit H. Bermano, Daniel Cohen-Or, Amir Roshan Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Trans. Graph., 41:86:1-86:11, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 602, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 645 + ], + "type": "text", + "content": "[50] Rijnder Wever and Tom F.H. Runia. Subitizing with variational autoencoders. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, September 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 690 + ], + "type": "text", + "content": "[51] Yonghui Wu and David Fleet. How ai creates photorealistic images from text. https://blog.google/technology/research/how-ai-creates-photorealistic-images-from-text, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[52] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yin" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3179" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 420 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 126 + ], + "type": "text", + "content": "fei Yang, Burcu Karagol Ayan, Benton C. Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. ArXiv, abs/2206.10789, 2022. 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 162 + ], + "type": "text", + "content": "[53] Nir Zabari and Yedid Hoshen. Semantic segmentation in the-wild without seeing any segmentation examples. ArXiv, abs/2112.03185, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 228 + ], + "type": "text", + "content": "[54] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18123-18133, June 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "text", + "content": "[55] Hui Zhang, Shenglong Zhou, Geoffrey Y. Li, and Naihua Xiu. 0/1 deep neural networks via block coordinate descent. ArXiv, abs/2206.09379, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[56] Jianming Zhang, Shugao Ma, Mehrnoosh Sameki, Stan Sclaroff, Margrit Betke, Zhe L. Lin, Xiaohui Shen, Brian L. Price, and Radomir Mech. Salient object subitizing. International Journal of Computer Vision, 124:169-186, 2015. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "text", + "content": "[57] Y. Zhang, Jonathon S. Hare, and Adam Prügel-Bennett. Learning to count objects in natural images for visual question answering. ArXiv, abs/1802.05766, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 386 + ], + "type": "text", + "content": "[58] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "type": "text", + "content": "[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 2022. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3180" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/4a5fd117-e0c8-457c-9985-c705ba827466_content_list.json b/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/4a5fd117-e0c8-457c-9985-c705ba827466_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..616ba528bbb5b44dd81602165f50d2c5cd287829 --- /dev/null +++ b/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/4a5fd117-e0c8-457c-9985-c705ba827466_content_list.json @@ -0,0 +1,1823 @@ +[ + { + "type": "text", + "text": "Tem-adapter: Adapting Image-Text Pretraining for Video Question Answer", + "text_level": 1, + "bbox": [ + 104, + 130, + 867, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guangyi Chen $^{1,2*}$ , Xiao Liu $^{5*}$ , Guangrun Wang $^{4}$ , Kun Zhang $^{1,2}$ , Philip H.S. Torr $^{4}$ , Xiao-Ping Zhang $^{3,6}$ , Yansong Tang $^{3\\dagger}$", + "bbox": [ + 161, + 178, + 807, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carnegie Mellon University, Pittsburgh PA, USA", + "bbox": [ + 287, + 215, + 683, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE", + "bbox": [ + 186, + 233, + 785, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Shenzhen International Graduate School, Tsinghua University, China", + "bbox": [ + 207, + 251, + 761, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ University of Oxford, UK", + "bbox": [ + 377, + 268, + 591, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Eindhoven University of Technology, NL", + "bbox": [ + 318, + 286, + 651, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ Toronto Metropolitan University, Canada", + "bbox": [ + 316, + 304, + 651, + 321 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f247dfaba628aee3bd2515697887805abb6092559e7ba458bd4a9b56d2b9de1.jpg", + "image_caption": [ + "Figure 1: Illustration of the domain gaps between the pre-trained CLIP and the downstream VideoQA. CLIP is trained to align visual and textual domains, while VideoQA requires understanding the temporal dynamics and complex semantics of videos." + ], + "image_footnote": [], + "bbox": [ + 81, + 354, + 888, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 558, + 313, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-language pre-trained models have shown remarkable success in guiding video question-answering (VideoQA) tasks. However, due to the length of video sequences, training large-scale video-based models incurs considerably higher costs than training image-based ones. This motivates us to leverage the knowledge from image-based pretraining, despite the obvious gaps between image and video domains. To bridge these gaps, in this paper, we propose Tem-Adapter, which enables the learning of temporal dynamics and complex semantics by a visual Temporal Aligner and a textual Semantic Aligner. Unlike conventional pretrained knowledge adaptation methods that only concentrate on the downstream task objective, the Temporal Aligner introduces an extra language-guided autoregressive task aimed at facilitating the learning of temporal dependencies, with the objective of predicting future states based on historical clues and language guidance that describes event progression. Besides, to reduce the semantic gap and adapt the textual representation for better event description, we introduce a Semantic Aligner that first designs a template to fuse", + "bbox": [ + 75, + 598, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "question and answer pairs as event descriptions and then learns a Transformer decoder with the whole video sequence as guidance for refinement. We evaluate Tem-Adapter and different pre-train transferring methods on two VideoQA benchmarks, and the significant performance improvement demonstrates the effectiveness of our method.", + "bbox": [ + 498, + 560, + 893, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 671, + 632, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Question Answering (VideoQA) is a task that aims to answer natural language questions based on the information available in observed videos. It has attracted considerable attention recently due to its promise to develop interactive AI systems capable of communicating with the dynamic visual environment using natural language. Despite significant advancements in recent years, VideoQA remains a challenging problem that requires models to comprehensively understand and dynamically align the semantics of both the visual world and natural language.", + "bbox": [ + 496, + 696, + 895, + 849 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Our code can be found at: https://github.com/XLiu443/Tem-adapter", + "bbox": [ + 514, + 862, + 880, + 876 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 519, + 876, + 627, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 519, + 887, + 643, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13945", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, a variety of methods [79, 78, 82, 34, 74, 63] have demonstrated impressive results in utilizing the largescale vision-language pre-trained (VLP) model to enhance downstream VideoQA tasks. These methods pre-train models by aligning the video and language domains and then apply them to VideoQA tasks via fine-tuning or even zero-shot learning. Nevertheless, pre-training large-scale models necessitate a large number of video-text pairs, e.g., more than 10 million videos, and entails expensive computational costs. This motivates us to explore cheaper and lighter alternative pre-trained models.", + "bbox": [ + 80, + 90, + 468, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Using image-VLP models is one potential option as they also align the semantics of vision and language domains. Compared to video-based models, image-based models offer two significant advantages. Firstly, training image-based models is less expensive when considering an equal number of data pairs. Secondly, image-VLP models have made significant progress in recent years, with some of them widely available and used, such as CLIP [56]. However, as depicted in Figure 1, domain gaps persist between image-based pretraining and VideoQA tasks since image-based models are unable to learn the temporal dynamics and corresponding complex event semantics of the visual world.", + "bbox": [ + 80, + 260, + 468, + 440 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenge posed by the domain gaps, this paper proposes Tem-adapter , an adapter network that leverages the interaction between visual and textual modalities to learn temporal dynamics and complex semantics. Unlike existing methods that directly adapt visual and textual features with the objective of the downstream tasks, such as cross-modal matching in VideoQA, Tem-Adapter introduces an additional language-guided autoregressive task to facilitate learning of temporal dynamics, which enables visual representations to predict future states based on historical information and language guidance. Besides, we introduce cross-modal interactions to further reduce the semantic gap and refine textual representation.", + "bbox": [ + 80, + 444, + 468, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach consists of a visual Temporal Aligner and a textual Semantic Aligner. The Temporal Aligner uses a Transformer encoder to learn the temporal relations and refine visual features. To optimize it, we build an auto-regression model by a Transformer decoder to generate the future state with historical information. We also incorporate the language information for visual refinement by adding textual embeddings as the condition memory of the Transformer decoder. This auto-regression model is supervised by a reconstruction loss with ground truth future frames to encourage the learning of temporal dependencies. In addition to the visual branch, we introduce a Semantic Aligner for better event description by reducing the semantic gap between the original texts crawled from the net and question-answer pairs in VideoQA. First, we design templates to fuse textual questions and answers to generate declarative sentences to reduce the domain gap between training and downstream", + "bbox": [ + 80, + 643, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "languages. Then, a Transformer decoder is employed to refine textual embeddings by incorporating the entire video sequence as a memory condition for video-text interactions.", + "bbox": [ + 503, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct experiments on two public VideoQA benchmarks including SUTD-TrafficQA [76] and MSR-VTTMC [75], and evaluate different categories of methods that transfer the pre-trained knowledge to downstream tasks, such as other adapter models [12], prompt learning [89, 21], and fully or partial finetuning. The significant improvement demonstrates the effectiveness of Tem-adapter to adapt the image-language pre-trained model for VideoQA tasks.", + "bbox": [ + 503, + 136, + 890, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 271, + 638, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Question Answering. Compared with Image-based Question Answering, Video QA is more challenging due to the complex temporal dependency, which requires models to dynamically align the visual observation and natural language. Some previous methods [67, 3, 48, 81, 28, 65, 77, 27, 26, 4, 7, 6, 8, 41, 84, 91, 10, 11, 36, 13, 17, 22, 23, 51, 62, 72] learn the models with the given manually annotated datasets. Recently, more and more methods [79, 78, 82, 34, 74, 63, 29] propose to utilize the representation ability of large-scale pre-trained vision-language models to solve the VideoQA task. For example, VQA-T [78] pre-trains the vision language model with randomly sampled video-question-answer triplets in the HowTo100M [47] dataset. BiLM [79] pre-trains frozen bidirectional language models with some adapter layers on the WebVid10M [1] dataset and achieve excellent zero-shot VideoQA performance on the downstream tasks. However, pretraining the large-scale model is expensive, which motivates us to find a cheaper alternative manner. In this paper, we propose to use existing image-language pre-trained models, such as CLIP [56], to provide the alignment between the visual and textual domains. The most related method is ATP [2], which also applies CLIP as the pre-trained model to extract frame and language embeddings, and learns an atemporal probe to select the most discriminative frames to match the video and QA pairs. Different from ATP selecting frames to transfer the video to images, we propose Tem-adapter to learn the temporal dynamics and reduce the gap between the pre-trained image domain and downstream video domain.", + "bbox": [ + 503, + 296, + 890, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Language Pre-training. The goal of VLP is to help downstream tasks by pre-training the model on large-scale vision-text pairs [56, 37, 20]. It always crawls a huge amount of vision-text pairs from the web and trains the model in a self-supervised manner, such as reconstruction [40, 15, 9, 30], contrastive matching [56, 20, 18], or the combination of both two [38, 71, 25]. Different from image-VLP, video-VLP has more diverse proxy tasks, such as frame/language mask [66, 90], video-language matching [46], video/language generation [73, 45], and", + "bbox": [ + 503, + 750, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13946", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e23089f8115c8ff6203662a03a9be750cba9d115f2975625c50bc694c461adc1.jpg", + "image_caption": [ + "(a) Finetuning" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 464, + 145 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9deb2796159aac63423441de2909d2c1dd92490dbda499f47f2d7db1034284fc.jpg", + "image_caption": [ + "(b) Prompt learning" + ], + "image_footnote": [], + "bbox": [ + 81, + 160, + 464, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c07ebb8cbbf5e36fad87359dfba7e353bbf371faa1d7b93cc86de6a2270b9e5d.jpg", + "image_caption": [ + "(c) Adapter" + ], + "image_footnote": [], + "bbox": [ + 81, + 227, + 465, + 286 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/41862b9e82b6f391f70318d06e9508bad8cd59f288245db8eeaedba90993c1f3.jpg", + "image_caption": [ + "(d) Tem-Adapter", + "Figure 2: Various methods for pre-trained knowledge adaptation in VideoQA. (a) Fine-tuning updates pretrained parameters with downstream training. (b) Prompt learning freezes the pre-trained model and introduces learnable textual tokens for adaptation. (c) Adapter-based methods add new adapter layers to transfer knowledge. (d) Our approach, Tem-adapter , leverages the visual-textual interactions to learn temporal dynamics to reduce the gap between image-based pre-training and video-based tasks." + ], + "image_footnote": [], + "bbox": [ + 81, + 301, + 465, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "frame/language order prediction [33, 39]. Recently, VLP models have shown great potential for other downstream applications, such as zero-shot and few-shot visual recognition [56, 12, 89, 88, 86], segmentation and detection [60, 87], image generation [50, 59, 52], and action understanding [70, 68]. In this paper, we use the CLIP model as our pre-trained model.", + "bbox": [ + 75, + 547, + 468, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-trained Knowledge Transfer. In light of the remarkable success of the large-scale pre-trained model, many methods have been proposed to solve the problem that how to efficiently adapt the pre-trained knowledge to the downstream tasks. Early methods finetune all parameters of the pre-trained model with the downstream objectives [14, 53]. However, with the model scale increasing, the cost of finetuning the full model is extremely expensive. Thus, an increasing number of methods focus on knowledge adaptation with light parameters. A straightforward way is to partially finetune the model parameters [49]. Besides, prompt learning methods are proposed recently to reformulate the downstream tasks as the pre-trained tasks and freeze all model parameters [43, 54, 58, 55]. The previous prompt-based methods [54, 55] design the template of prompt using the prior of natural human language. Then, some methods [24, 64]", + "bbox": [ + 75, + 657, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "propose to automatically optimize the prompt template, and [42, 69, 44, 89, 5] further learn the prompts in the continuous embedding space where the discrete word constraint is relaxed. In addition, some methods freeze the model and train some extra adapter layers and also achieve the supersizing performance [12, 85, 1]. As shown in Figure 2, we illustrate the difference between Tem-adapter and other categories of methods transferring the image-based VLP model to the VideoQA task. Compared with other pre-train adaptation methods, our Tem-adapter proposes an autoregression task to reduce the temporal gap (tem-align) and introduce the cross-domain interactions (sem-align) to reduce the semantic gap between the pre-training and downstream tasks.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 500, + 305, + 607, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first briefly introduce the VideoQA task and then present our approach, Tem-adapter , which introduces the cross-modal interactions to learn temporal dynamics and complex semantics. Finally, we give the details of our implementation.", + "bbox": [ + 496, + 332, + 893, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 500, + 421, + 681, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of VideoQA is to find an answer $\\hat{\\mathbf{a}}$ from the answer pool, given the conditions of question $\\mathbf{q}$ and observed video data $\\mathbf{v}$ as below:", + "bbox": [ + 496, + 446, + 892, + 491 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {a}} = \\underset {\\mathbf {a} \\in \\mathcal {A}} {\\arg \\max } \\mathcal {S} (\\mathbf {a} | \\mathbf {v}, \\mathbf {q}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 497, + 892, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $S$ is a match function to generate the score of each answer candidate given the question and video content.", + "bbox": [ + 500, + 526, + 890, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Tem-Adapter", + "text_level": 1, + "bbox": [ + 500, + 570, + 638, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we propose Tem-adapter to learn temporal dynamics and complex semantics by the interactions of visual and textual domains to adapt pre-trained Image-VLP models to downstream VideoQA tasks. Figure 3 shows an overview of Tem-Adapter's framework. In particular, Tem-adapter consists of two components, including a visual Temporal Aligner and a textual Semantic Aligner. Details for each component are as follows.", + "bbox": [ + 496, + 595, + 893, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Aligner. The goal of the Temporal Aligner is to reduce the gap between image-based pre-training and video-based downstream tasks. To achieve this goal, we introduce an auxiliary language-guided autoregressive task to facilitate the learning of temporal dynamics.", + "bbox": [ + 496, + 717, + 893, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, given a video $\\mathbf{v}$ , we apply the frozen CLIP [56] image encoder $f$ to extract the feature of each frame as $f(\\mathbf{v}) = \\{f(\\mathbf{v}_t)|_{t = 1}^T\\} \\in \\mathcal{R}^{T\\times C}$ , where $T$ is the frame number of videos and $C$ is the feature dimension of CLIP. Then we refine these features with our Temporal Aligner model which is built by a transformer encoder $f_{e}$ with parameters $\\phi$ to learn the temporal relations of frames.", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13947", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/221a9583e07cb8316e2249cea0dbeafaa5408829cc2a561d13a9d45d1e13d6a9.jpg", + "image_caption": [ + "Figure 3: The overall framework of Tem-adapter. Our Tem-adapter consists of two modules, Semantic Aligner (orange) and Temporal Aligner (blue). Semantic Aligner takes the text embedding of template-based QA pair as the input, and applies a Transformer decoder to incorporate cross-modal interactions. The Temporal Aligner introduces a language-guided autoregressive task to learn temporal dynamics. The training is guided by classification and reconstruction losses." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 872, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The refined features can be represented as:", + "bbox": [ + 76, + 359, + 359, + 376 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} = f _ {e} (f (\\mathbf {v}); \\phi), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 387, + 470, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which are with the same dimension $\\mathcal{R}^{T\\times C}$", + "bbox": [ + 75, + 415, + 366, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To guide the Temporal Aligner to learn temporal dynamics, we further build an autoregressive sequential model, following GPT [57], which generates the feature of the current frame with previous frames and guidance from textual information that describes event progression. Mathematically, it can be formulated as:", + "bbox": [ + 75, + 431, + 470, + 521 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} (\\mathbf {v} _ {i}) = f _ {d} \\left(f \\left(\\mathbf {v} _ {1: i - 1}\\right) | \\mathbf {f}, \\mathbf {g}; \\psi\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 526, + 468, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f(\\mathbf{v}_{1:i-1})$ is the input of the transformer decoder, which is achieved by applying an attention mask to focus on the history clues. $\\mathbf{f}, \\mathbf{g}$ are used as the guidance information, where $\\mathbf{f}$ denotes latent visual embeddings obtained by Temporal Aligner and $\\mathbf{g}$ (see equation 5) is the textual embeddings describing temporal event. $\\mathbf{f}$ and $\\mathbf{g}$ are first fused by adding and then serve as the memory key and value of the Transformer decoder. We apply a reconstruction loss to encourage the learning of temporal dependencies which is formulated as the distance between the results of the auto-regression model $\\hat{f}(\\mathbf{v}_i)$ and the ground-truth $f(\\mathbf{v}_i)$ :", + "bbox": [ + 75, + 549, + 472, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {M S E} = \\sum_ {i = 1} ^ {n} | | f (\\mathbf {v} _ {i}) - \\hat {f} (\\mathbf {v} _ {i}) | | _ {2} ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 720, + 468, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Aligner. In addition to the visual branch, we introduce a Semantic Aligner to adapt the textual representation for better event description and reduce the semantic gaps. We first design templates to align the question-answer pairs in VideoQA to declarative sentences, since the training data of CLIP is crawled from the web which always describes the image in a declarative manner. Then, we apply a Transformer decoder with the video sequence as the memory for video-text interactions.", + "bbox": [ + 75, + 763, + 472, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a question $\\mathbf{q}$ and a corresponding answer $\\mathbf{a}$ , we first apply the grammar parsing on the question $\\mathbf{q}$ to obtain the subject-verb-object structure, and then we modify the order of the sentence to obtain a declarative sentence and leave the position of the interrogative pronouns as a mask '[']. Finally, we fill in the default '['] with the answer to update the QA pair as a refined event description. For instance, the question is 'Which might be the reason for this accident?', and the answer is 'Improper lane change'. We first obtain a declarative sentence with the default as '[ ] is the reason for this accident', and then we can fill in the answer to get 'Speeding vehicle is the reason for this accident.' In addition, for the general question, we first change it to a declarative sentence and then convert the sentence to negative form if the answer is No.", + "bbox": [ + 496, + 359, + 893, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the refined sentence $\\mathbf{q}\\mathbf{a}$ as the input, the frozen text encoder $g$ of the CLIP model outputs the textual feature as $g(\\mathbf{q}\\mathbf{a}) \\in \\mathcal{R}^C$ . To further refine the textual feature using the video content, we apply a transformer decoder with the visual feature sequence as the memory key and value and add a residual link to avoid overfitting. Formally, it can be formulated as follows:", + "bbox": [ + 496, + 587, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {g} = g (\\mathbf {q a}) + \\lambda g _ {d} (g (\\mathbf {q a}), f (\\mathbf {v}); \\theta), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 699, + 890, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{g}$ denotes the learned textual embedding, $\\lambda$ is a parameter of the residual link, and $\\theta$ is the parameters of the transformer decoder $g_{d}$ . Both $\\lambda$ and $\\theta$ can be learned automatically in the training process. With the cross-attention module in the transformer decoder, we encourage the model to find key information from the video to refine the textual embedding for better event description.", + "bbox": [ + 496, + 720, + 893, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-modal Matching. Given the refined visual and textual embeddings $\\mathbf{f}$ and $\\mathbf{g}$ , we instantiate the score function in equation 1 by cosine similarity as:", + "bbox": [ + 496, + 825, + 893, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS (\\mathbf {a} | \\mathbf {v}, \\mathbf {q}) = \\operatorname {s o f t m a x} (\\bar {\\mathbf {f}} ^ {T} \\mathbf {g}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 883, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13948", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{f}}$ is video representation using the average pool on all frames, $\\mathbf{g}$ is the language embedding with given answer question pair $(\\mathbf{a},\\mathbf{q})$ . Then, we can obtain the answer to the question with the highest matching score.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we apply a standard multi-choice classification Hinge loss [19] integrated with the auto-regression loss for training. The overall loss function can be described as:", + "bbox": [ + 76, + 152, + 468, + 198 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {H i n g e} (S (\\mathbf {a} | \\mathbf {v}, \\mathbf {q}), y) + \\gamma L _ {M S E}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 204, + 468, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ is a ratio to balance two loss functions, and $y$ is the ground truth label of the answers. Please note that our Tem-adapter consists of 3 learnable modules including the textual decoder $\\theta$ , the visual encoder $\\phi$ , and the visual decoder $\\psi$ . The Hinge loss is applied to optimize $\\theta$ and $\\phi$ , while the auto-regression loss $L_{MSE}$ is used to optimize all three modules. In the inference, only textual decoder $\\theta$ and visual encoder $\\phi$ are used.", + "bbox": [ + 75, + 226, + 470, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 359, + 294, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this subsection, we provide more implementation details of our approach for better reproduction.", + "bbox": [ + 75, + 383, + 472, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Architecture. Figure 3 shows the framework of Tem-adapter. In our experiments, we use the frozen model CLIP (ViT-B32) as the text and image encoders. The Semantic Aligner is a one-layer Transformer decoder with 16 heads. There is a linear layer before the Transformer decoder which maps text embeddings and image embeddings to a lower dimension (128). Another linear projection layer is added after the Transformer decoder to map the reduced dimension back to 512. The Temporal Aligner is a light transformer encoder-decoder architecture. Both the encoder and decoder contain only one layer. We set the number of heads to 16 in the experiments.", + "bbox": [ + 75, + 414, + 470, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference Details. In the inference (given 1 question, 1 video, and $k$ answers), we first apply the template to fuse the question and $k$ answers to obtain $k$ QA pairs (e.g. $k = 4$ ). Then we feed the QA pairs and video into CLIP textual and visual encoders to obtain the initial features. After that, we apply the learned semantic and temporal aligners to refine these features. Please note that the temporal auto-regression module is not used in the inference. Finally, we obtain 1 video feature with $k$ QA features and apply the cosine similarity as well as softmax to select the answer.", + "bbox": [ + 75, + 597, + 470, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hyper-parameters. We follow the same video frames sampling method with prior methods [76]. We uniformly select 8 frames over a long video sequence and then pick consecutive 16 frames which are centered on those 8 frames. Overall 128 frames are selected for each video. We also tried another sampling method: uniformly extracting 128 or 256 frames on a video sequence, and the final results do not support the sampling method very well. In the loss function, we set $\\gamma$ as 100 to encourage the model to learn the temporal dynamics. And the learnable parameter $\\lambda$ in the Semantic", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Aligner is set to be the same dimension (512) as the output embeddings of the CLIP model.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. We use the pre-trained CLIP model to extract text embeddings and image embeddings. The Adam [31] optimizer is utilized to optimize the training process. We set the initial learning rate to be $1e - 4$ with a batch size of 128. And we let the learning rate decay with a factor of 0.5 for every 10 epochs. Our model is trained for 50 epochs on a single Tesla V100 GPU and the training takes 8 hours to finish. Our model is implemented in PyTorch 1.8.0.", + "bbox": [ + 496, + 121, + 893, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 258, + 633, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first benchmark our Tem-adapter and different categories of methods that transfer the pretrained model into downstream tasks. Then we provide ablation studies to investigate the effect of each component.", + "bbox": [ + 496, + 285, + 893, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets & Experimental Protocols", + "text_level": 1, + "bbox": [ + 498, + 357, + 805, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SUTD-TrafficQA. SUTD-TrafficQA focuses on the question-answering task in the traffic situation, which requires the model to understand the traffic event and underlying causal relation. It consists of more than 10k videos with different traffic events and provides over 62,535 human-annotated QA pairs. Among them, 56,460 QA pairs are used for training and the rest 6,075 QA pairs are used for testing. This dataset highlights the cognitive capability of video event understanding models in complex traffic scenarios. It provides 6 challenging traffic-related reasoning tasks including \"Basic understanding\", \"Event forecasting\", \"Reverse reasoning\", \"Counterfactual inference\", \"Introspection\", \"Attribution\". All tasks are formulated as multiple-choice without limiting the number of candidate answers.", + "bbox": [ + 496, + 380, + 893, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MSR-VTT-MC. MSR-VTT is collected and released mainly for the text-to-video retrieval task. Following [34] and [2], we use the multi-choice test setting to benchmark the VideoQA methods. The MSR-VTT-MC test set consists of 2990 QA pairs. Each video corresponds to five answers, with the original caption as the correct answer. We use the standard protocol with prior works in our experiments, where 7010 train+val video sequences and 140,200 QA pairs are utilized for training. Each video only contains one caption. We assign this caption as the correct answer for our question-answering task. To formulate four negative answers during training, we randomly select four captions from other captions that do not belong to the current video.", + "bbox": [ + 496, + 593, + 893, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Baseline Methods", + "text_level": 1, + "bbox": [ + 498, + 801, + 671, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To investigate different categories of methods that transfer the pre-trained model into downstream tasks, such as finetuning, prompt learning, and adapter. We compare Tem-adapter with the other 10 baseline methods including Unsupervised CLIP [56], Unsupervised CLIP [56]", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13949", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4928a77f32bebceeff81172a5fa02e81593c65bf81d956a20d6928eece3da2ca.jpg", + "table_caption": [ + "Table 1: VideoQA accuracies of baseline methods on both SUTD-TrafficQA and MSR-VTT-MC. B: \"Basic understanding\", F: \"Forecasting task\", R: \"Reverse Reasoning\", C: \"Counterfactual inference\", I: \"Introspection\", A: \"Attribution\". (C) and (A) denote strategies for adding prompts. (C) and $(\\mathbf{C}^{*})$ denote training prompts with/without adapter heads." + ], + "table_footnote": [], + "table_body": "
MethodsSUTD-TrafficQAMSR-VTT-MC
BFRCIAAvgAcc
Unsupervised CLIP [56]25.620.134.030.822.828.826.574.1
CLIP [56] + Template31.836.029.971.822.133.432.374.1
Totally finetuning39.835.146.645.637.240.540.389.0
Partially finetuning41.637.844.650.033.141.741.788.5
LORA [16]38.738.736.737.934.538.138.385.4
CLIP-Adapter [12]35.832.035.442.333.132.134.883.4
Multi-layer Adapter [12]30.526.626.538.528.325.829.190.7
Prompt learning (C) [89]42.432.445.255.540.743.642.989.0
Prompt learning (C*) [89]40.333.241.046.534.938.439.790.8
Prompt learning (A) [21]41.731.540.148.433.141.441.188.0
Tem-adapter46.036.544.655.034.547.746.094.3
", + "bbox": [ + 125, + 143, + 843, + 371 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e3e70cee006617c67fac8a15a0831e8dc684a384ae6de4ac026c9e08a40a79dc.jpg", + "table_caption": [ + "Table 2: Ablation studies on the SUTD-TrafficQA and MSR-VTT-MC datasets. Tp, SA, TA, and Ar denote Template, Semantic Aligner, Temporal Aligner, and Auto-regression, respectively. No ablation studies (“\\”) concentrate on Tp, as it is not compatible with the MSR-VTT-MC dataset." + ], + "table_footnote": [], + "table_body": "
MethodsSUTD-TrafficQAMSR-VTT-MC
AvgAcc
CLIP-Adapter [12]34.883.4
w/o Tp43.9\\
w/o SA44.291.8
w/o TA43.691.6
w/o Ar44.992.6
w/o TA and SA32.374.1
Tem-adapter46.094.3
", + "bbox": [ + 84, + 464, + 460, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "+ Language template, Totally fine-tuning, Partially finetuning, LORA [16], CLIP-Adapter [12], Multi-layer CLIP-Adapter [12], Prompt learning (change words) [89] (with/without using adapter heads), and Prompt learning (add words) [21]. The detailed descriptions are shown in the supplementary materials.", + "bbox": [ + 73, + 638, + 470, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Benchmarking against Baseline Methods", + "text_level": 1, + "bbox": [ + 75, + 739, + 426, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we compare and discuss our approach's performance against the above-mentioned baselines on both SUTD-TrafficQA and MSR-VTT-MC datasets. The results of all methods on both datasets are summarized in Table 1. We can draw conclusions as follows from the results.", + "bbox": [ + 75, + 763, + 470, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP can obtain good zero-shot results. With our language template, the clip model can achieve $32.3\\%$ accuracy on the SUTD-TrafficQA dataset, which is almost close to the previous state-of-the-art performance $(37.1\\%)$ . It shows", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the strong representation ability of the CLIP model.", + "bbox": [ + 498, + 380, + 841, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "More parameters $\\neq$ Better performance. We have two groups of experiments that focus on the number of parameters, including totally/partially finetuning and single-layer/multi-layer Adapters. We obtain contradictory conclusions from different datasets. On SUTD-TrafficQA, the model with fewer parameters works better, while on MSR-VTT-MC, the model with more parameters has better performance. It is not surprising since the scale of training data is different. MSR-VTT-MC has more training data and can support training more parameters. We also compare the learnable parameters of adapter-based methods in Table 5. Tem-adapter achieve better performance with fewer parameters than Multi-layer Adapter. Compared with the prompt-learning-based method, our method needs more parameters, but its inference is not time-consuming.", + "bbox": [ + 496, + 397, + 893, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different categories of methods are comparable. Balancing the performance on both datasets, different categories of methods that transfer the image-based VLP model to VideoQA don't have clear performance differences. Adapter-based methods obtain poor results on SUTD-TrafficQA, but multi-layer Adapter works well on MSR-VTT-MC. Besides, by comparing the prompt learning methods with/without adapter heads, we found that the prompt and adapter can be joint learning to obtain better performance.", + "bbox": [ + 496, + 626, + 895, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-modal interaction and temporal dynamics matter. Compared with CLIP-Adapter (both single-layer and multi-layer ones), our Tem-adapter achieves more than $10\\%$ improvement on SUTD-TrafficQA. The key difference between Tem-adapter and other adapter-based methods is that we introduce temporal alignment to reduce the temporal gap and apply the cross-modal interaction to refine the representation and reduce the semantic gap. Tem-adapter outperforms other methods by a large mar", + "bbox": [ + 496, + 763, + 895, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13950", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3c28eba5066b0934583c6481c1dab079ca56c63bcfce0b78ecb1cfe6a0f83700.jpg", + "table_caption": [ + "Table 3: Comparison with the state-of-the-art methods on the SUTD-TrafficQA dataset." + ], + "table_footnote": [], + "table_body": "
MethodsAccuracyYear
Human95.4-
VIS+LSTM [61]29.92015
TVQA [35]35.22018
HCRN [32]36.52020
BERT-VQA [80]33.72020
Eclipse [76]37.12021
CLIP [56]32.32021
CLIP+HCRN[56, 32]41.92021
ATP [2]35.62022
Tem-adapter46.0
", + "bbox": [ + 101, + 128, + 446, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gin on both two datasets. For example, it obtains $3.1\\%$ improvement than Prompt learning (C) on SUTD-TrafficQA and $3.6\\%$ than Multi-layer Adapter on MSR-VTT-MC.", + "bbox": [ + 75, + 316, + 470, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison against State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 76, + 373, + 467, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SUTD-TrafficQA. On SUTD-TrafficQA, we compare our proposed method with state-of-the-art methods, including [32, 76, 61, 35, 80]. The results are summarized in Table 3. We can observe that our Tem-adapter outperforms other methods by a large margin. For example, compared with HCRN [32], Tem-adapter achieves almost $10\\%$ improvement. By using the CLIP model to extract visual features, HCRN+CLIP can obtain $41.9\\%$ accuracy, which is better than using the original HCRN. Besides, we show that the performance improvement is not only from the better representation ability of the CLIP model since our Tem-adapter further improve HCRN+CLIP with $+4.1\\%$ accuracy. In addition, we also significantly outperform the ATP method, which also uses the CLIP model. This significant performance improvement demonstrates the effectiveness of Tem-adapter.", + "bbox": [ + 75, + 397, + 470, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MSR-VTT-MC. On MSR-VTT-MC, we compare Tem-adapter with some recent state-of-the-art methods, including ActBERT[90], ClipBERT[34], MERLOT [83], VideoCLIP [66], CLIP [56], and ATP [2]. Results are shown in Table 4. First, we can find that the CLIP model benefits the performance on the MSR-VTT-MC dataset. All methods using the CLIP model, such as CLIP [56], ATP [2], and Tem-adapter achieve good performance. ATP [2] learns an atemporal probe to select a frame from a video sequence to obtain the visual embedding and achieve good performance on MSR-VTT-MC. Compared with it, Tem-adapter can further obtain $94.3\\%$ accuracy and $1.1\\%$ relative improvement.", + "bbox": [ + 75, + 638, + 470, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 845, + 243, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide ablation studies to investigate the effect of each component. Four main modules in our Tem-adapter", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/07eaf825934fabd225f5efc47a414372420e128fe08c38228ed86e97e6186fe3.jpg", + "table_caption": [ + "Table 4: Comparison with the state-of-the-art methods on the MSR-VTT-MC dataset." + ], + "table_footnote": [], + "table_body": "
MethodsAccuracyYear
VideoCLIP [66]92.12019
ActBERT[90]85.72020
ClipBERT[34]88.22021
MERLOT [83]90.92021
CLIP [56]74.12021
ATP [2]93.22022
Tem-adapter94.3
", + "bbox": [ + 542, + 128, + 852, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are evaluated, including Temporal Aligner, Semantic Aligner, language template, and auto-regression.", + "bbox": [ + 498, + 287, + 893, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 2, we provide the performance of different variants of Tem-adapter and a baseline adapter method on each task in the SUTD-TrafficQA and MSR-VTTMC datasets. The baseline method is CLIP-adapter, which also uses the language template and learns a linear adapter layer without cross-domain interaction. We observe that all variants of Tem-adapter can achieve significant improvement over the baseline method, which shows that the interaction across visual and textual is very important for video analysis. Besides, compared with Tem-adapter, we can observe a clear performance drop when we remove each component, which demonstrates all of them are indispensable. When both temporal and semantic aligners are removed, the performance drops dramatically.", + "bbox": [ + 496, + 318, + 893, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we also conducted an experiment to show whether the Temporal Aligner learns the temporal dynamic information. Table 6 shows the reconstruction quality of our methods with/without language guidance $(g)$ and with/without Temporal Aligner $(f)$ . We can observe that both the language guidance and Temporal Aligner assist in the learning of temporal dynamics.", + "bbox": [ + 496, + 531, + 893, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we also provide some parameter analysis experiments about our network design. And evaluate Tem-adapter on different image-text pre-train models. Please kindly refer to the supplementary materials for the detailed results and analysis.", + "bbox": [ + 496, + 638, + 893, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Visualization", + "text_level": 1, + "bbox": [ + 500, + 724, + 637, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we provide some qualitative results to analyze our approach. As shown in Figure 4, we show some visualization examples of VdideoQA on the SUTD-TrafficQA dataset. In the top row, we plot two positive examples. We can observe that our Tem-adapter learns the temporal dynamics with the cross-domain interactions to obtain a better understanding of video events. Besides, we also provide some failure cases to help explore the boundary of our approach. We find that Tem-adapter can not work well for situations that need complex reasoning, especially", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13951", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Q: What could possibly cause this accident?", + "bbox": [ + 125, + 93, + 419, + 108 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03e235e26b72a5879a70ea52c006d335d3753213ce85a1a893e4be646de2423c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 109, + 467, + 165 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ours: Sudden or extreme movement by a vehicle.", + "bbox": [ + 127, + 167, + 410, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CLIP-Adapter: Unfavorable weather.", + "bbox": [ + 127, + 180, + 339, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt learning: Violation of traffic rules.", + "bbox": [ + 127, + 193, + 367, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finetuning: Violation of traffic rules.", + "bbox": [ + 127, + 205, + 336, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q: Could any vehicles steer away to prevent the accident?", + "bbox": [ + 125, + 232, + 421, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1cd05c0e3725a624008bedcf98cbe31a252065b747e5c8e14ed03eb864520497.jpg", + "image_caption": [ + "Figure 4: Visualization of some examples of the VideoQA task from the SUTD-TrafficQA dataset. The top row shows some positive examples, where Tem-adapter learns temporal dynamics to understand the traffic event and give the correct answers. The bottom row shows some failed cases. Despite better matching between visual and textual domains, our approach fails on complex problems which require reasoning ability." + ], + "image_footnote": [], + "bbox": [ + 91, + 266, + 460, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ours: No, there was no space. X", + "bbox": [ + 148, + 323, + 346, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GT: Yes, there was enough space.", + "bbox": [ + 148, + 335, + 341, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q: Which factors contribute to the accident?", + "bbox": [ + 522, + 93, + 816, + 108 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/872773b7ceeb0ea45defc8048668de003dd580de6a1153bd856e6e2f932bed1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 109, + 870, + 164 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ours: Not paying attention to their surroundings.", + "bbox": [ + 542, + 167, + 816, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CLIP-Adapter: Bad road surfaces", + "bbox": [ + 542, + 180, + 733, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt learning: Bad road surfaces.", + "bbox": [ + 542, + 193, + 746, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finetuning: Speeding vehicles.", + "bbox": [ + 542, + 205, + 715, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q: Could the accident be prevented if a convex mirror is installed at the turning point?", + "bbox": [ + 522, + 233, + 834, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51fbfa3d51d96754fa0e286797bf9782353f3153f2c66544d1103bc396a8929e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 266, + 867, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ours: Yes, there would be no obstructed views.", + "bbox": [ + 540, + 323, + 810, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GT: No, that was not the main cause of the accident.", + "bbox": [ + 540, + 335, + 839, + 348 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3d4e1ca4ce98a3cab4ac840c115206c6484c89e892b25ef81285a63ab1eacc30.jpg", + "table_caption": [ + "Table 5: The comparisons of our method and other baseline methods on model parameters (M), inference time (ms/video), and accuracy in the SUTD-TrafficQA dataset." + ], + "table_footnote": [], + "table_body": "
MetricsLORA [16]Prompt learning (C) [89]Prompt learning (C*) [89]CLIP-Adapter [12]Multi-layer Adapter [12]Ours
Parameters1.443.550.771.054.853.62
Inference time5.375.335.175.135.145.32
Accuracy38.342.939.734.829.146.0
", + "bbox": [ + 76, + 476, + 898, + 537 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b79c3e408ee4475342ab5238d4af0d1be0974715d93a16e5d46c2fecadf95938.jpg", + "table_caption": [ + "Table 6: The comparison of reconstruction ability. We use peak signal-to-noise ratio (PSNR) as the metric." + ], + "table_footnote": [], + "table_body": "
Tem-adapterw/o gw/o f and g
PSNR29.128.727.4
", + "bbox": [ + 98, + 575, + 444, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for the underlying cause-effect relations. We leave it as a future work to enhance the reasoning ability of our model. Besides, we also provide more examples of our language templates in the supplementary materials.", + "bbox": [ + 75, + 631, + 468, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 707, + 194, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed Tem-adapter to adapt the pre-trained image-language model to the downstream VideoQA task. We jointly learn the Temporal Aligner and Semantic Aligner through cross-modal interactions. Tem-adapter introduces a language-guided autoregressive task to guide the learning of temporal dependency and thus reduce the temporal gap between image-based pre-train and video-based QA tasks. Besides, we utilize a rule-based template and video-based interaction to refine textual representation and reduce semantic gaps. We validate the effectiveness of our approach on four VideoQA benchmarks.", + "bbox": [ + 75, + 734, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Broader Impacts and Limitations. Understanding how the big model adaptation works on VideoQA is of great help in developing general foundation models and interactive AI. However, our work also has certain limitations. For example, since extracting video features online takes up too many computing resources, we extract their features offline and fix them. We leave it as future work to learn the adapter within the backbone network.", + "bbox": [ + 496, + 542, + 893, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 678, + 651, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62206153, Young Elite Scientists Sponsorship Program by CAST (No. 2023QNRC002), by the Tsinghua Shenzhen International Graduate School-Shenzhen Pengrui Endowed Professorship Scheme of Shenzhen Pengrui Foundation, by the UKRI grant: Turing AI Fellowship EP/W002981/1, by the National Institutes of Health (NIH) under Contract R01HL159805, by the NSF-Convergence Accelerator Track-D award #2134901, by a grant from Apple Inc., a grant from KDDI Research Inc, and generous gifts from Salesforce Inc., Microsoft Research, and Amazon Research. We would also like to thank the Royal Academy of Engineering and FiveAI.", + "bbox": [ + 496, + 703, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13952", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 2, 3", + "[2] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the\" video\" in video-language understanding. In $CVPR$ , pages 2917-2927, 2022. 2, 5, 7", + "[3] Santiago Castro, Mahmoud Azab, Jonathan Stroud, Cristina Noujaim, Ruoyao Wang, Jia Deng, and Rada Mihalcea. Lifeqa: A real-life dataset for video question answering. In LREC, pages 4352-4358, 2020. 2", + "[4] Aman Chadha, Gurneet Arora, and Navpreet Kaloty. iperceive: Applying common-sense reasoning to multi-modal dense video captioning and video question answering. arXiv preprint arXiv:2011.07735, 2020. 2", + "[5] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. In ICLR, 2023. 3", + "[6] Seongho Choi, Kyoung-Woon On, Yu-Jung Heo, Ahjeong Seo, Youwon Jang, Minsu Lee, and Byoung-Tak Zhang. Dramaqa: Character-centered video story understanding with hierarchical qa. In AAAI, volume 35, pages 1166-1174, 2021. 2", + "[7] Anthony Colas, Seokhwan Kim, Franck Dernoncourt, Siddhesh Gupte, Daisy Zhe Wang, and Doo Soon Kim. Tutorial: Question answering dataset for tutorial videos. arXiv preprint arXiv:1912.01046, 2019. 2", + "[8] Long Hoang Dang, Thao Minh Le, Vuong Le, and Truyen Tran. Object-centric representation learning for video question answering. In IJCNN, pages 1-8, 2021. 2", + "[9] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Zicheng Liu, Michael Zeng, et al. An empirical study of training end-to-end vision-and-language transformers. arXiv preprint arXiv:2111.02387, 2021. 2", + "[10] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In CVPR, pages 1999-2007, 2019. 2", + "[11] Jiyang Gao, Runzhou Ge, Kan Chen, and Ram Nevatia. Motion-appearance co-memory networks for video question answering. In CVPR, pages 6576–6585, 2018. 2", + "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2, 3, 6, 8", + "[13] Noa Garcia, Mayu Otani, Chenhui Chu, and Yuta Nakashima. Knowit vqa: Answering knowledge-based questions about videos. In AAAI, volume 34, pages 10826-10834, 2020. 2", + "[14] Ross Girshick. Fast r-cnn. In ICCV, pages 1440-1448, 2015. 3", + "[15] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. VIn bert: A recurrent vision-and-language bert for navigation. In CVPR, pages 1643–1653, 2021. 2" + ], + "bbox": [ + 76, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6, 8", + "[17] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware graph convolutional networks for video question answering. In AAAI, volume 34, pages 11021-11028, 2020. 2", + "[18] Aashi Jain, Mandy Guo, Krishna Srinivasan, Ting Chen, Sneha Kudugunta, Chao Jia, Yinfei Yang, and Jason Baldridge. Mural: multimodal, multitask retrieval across languages. arXiv preprint arXiv:2109.05125, 2021. 2", + "[19] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In CVPR, pages 2758–2766, 2017. 5", + "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 2", + "[21] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022. 2, 6", + "[22] Jianwen Jiang, Ziqiang Chen, Haojie Lin, Xibin Zhao, and Yue Gao. Divide and conquer: Question-guided spatiotemporal contextual attention for video question answering. In AAAI, volume 34, pages 11101-11108, 2020. 2", + "[23] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In AAAI, volume 34, pages 11109-11116, 2020. 2", + "[24] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? TACL, 8:423-438, 2020. 3", + "[25] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1780-1790, 2021. 2", + "[26] Hyoungun Kim, Zineng Tang, and Mohit Bansal. Dense-ecaption matching and frame-selection gating for temporal localization in videoqa. arXiv preprint arXiv:2005.06409, 2020.2", + "[27] Junyeong Kim, Minuk Ma, Trung Pham, Kyungsu Kim, and Chang D Yoo. Modality shifting attention network for multimodal video question answering. In CVPR, pages 10106-10115, 2020. 2", + "[28] Kyung-Min Kim, Min-Oh Heo, Seong-Ho Choi, and Byoung-Tak Zhang. Deepstory: Video story qa by deep embedded memory networks. arXiv preprint arXiv:1707.00836, 2017. 2", + "[29] Seonhoon Kim, Seohyeong Jeong, Eunbyul Kim, Inho Kang, and Nojun Kwak. Self-supervised pre-training and contrastive representation learning for multiple-choice video qa. In AAAI, volume 35, pages 13171-13179, 2021. 2", + "[30] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, pages 5583-5594, 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13953", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[32] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In CVPR, pages 9972-9981, 2020. 7", + "[33] Chenyi Lei, Shixian Luo, Yong Liu, Wanggui He, Jiamang Wang, Guoxin Wang, Haihong Tang, Chunyan Miao, and Houqiang Li. Understanding chinese video and language via contrastive multimodal pre-training. In ACM MM, pages 2567-2576, 2021. 3", + "[34] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, pages 7331-7341, 2021. 2, 5, 7", + "[35] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. arXiv preprint arXiv:1809.01696, 2018. 7", + "[36] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvqa+: Spatio-temporal grounding for video question answering. arXiv preprint arXiv:1904.11574, 2019. 2", + "[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 2", + "[38] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 34, 2021. 2", + "[39] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 3", + "[40] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2", + "[41] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond mnns: Positional self-attention with co-attention for video question answering. In AAAI, volume 33, pages 8658–8665, 2019. 2", + "[42] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3", + "[43] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint arXiv:2107.13586, 2021. 3", + "[44] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands, too. arXiv preprint arXiv:2103.10385, 2021. 3", + "[45] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 2", + "[47] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 2", + "[48] Jonghwan Mun, Paul Hongsuck Seo, Ilchae Jung, and Bohyung Han. Marioqa: Answering questions by watching gameplay videos. In ICCV, pages 2867-2875, 2017. 2", + "[49] Akihiro Nakamura and Tatsuya Harada. Revisiting fine-tuning for few-shot learning. arXiv preprint arXiv:1910.00216, 2019.3", + "[50] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3", + "[51] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Bridge to answer: Structure-aware graph interaction network for video question answering. In CVPR, pages 15526-15535, 2021. 2", + "[52] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, pages 2085–2094, 2021. 3", + "[53] Matthew E Peters, Sebastian Ruder, and Noah A Smith. To tune or not to tune? adapting pretrained representations to diverse tasks. arXiv preprint arXiv:1903.05987, 2019. 3", + "[54] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? arXiv preprint arXiv:1909.01066, 2019. 3", + "[55] Nina Poerner, Ulli Waltinger, and Hinrich Schütze. Bert is not a knowledge base (yet): Factual knowledge vs. name-based reasoning in unsupervised qa. arXiv preprint arXiv:1911.03681, 2019. 3", + "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6, 7", + "[57] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 4", + "[58] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3", + "[59] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.3", + "[60] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. arXiv preprint arXiv:2112.01518, 2021.3" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13954", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. Advances in neural information processing systems, 28, 2015. 7", + "[62] Arka Sadhu, Kan Chen, and Ram Nevatia. Video question answering with phrases via semantic roles. arXiv preprint arXiv:2104.03762, 2021. 2", + "[63] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In CVPR, pages 16877-16887, 2021. 2", + "[64] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3", + "[65] Xiaomeng Song, Yucheng Shi, Xin Chen, and Yahong Han. Explore multi-step reasoning in video question answering. In ACM MM, pages 239-247, 2018. 2", + "[66] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, pages 7464-7473, 2019. 2, 7", + "[67] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In CVPR, pages 4631-4640, 2016. 2", + "[68] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022.3", + "[69] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. NeurlPS, 34:200-212, 2021. 3", + "[70] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021.3", + "[71] Wenhui Wang, Hangbo Bao, Li Dong, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. arXiv preprint arXiv:2111.02358, 2021. 2", + "[72] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In CVPR, pages 9777–9786, 2021. 2", + "[73] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic video-language model pre-training for video understanding. arXiv preprint arXiv:2105.09996, 2021. 2", + "[74] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. EMNLP, 2021. 2", + "[75] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, pages 5288-5296, 2016. 2", + "[76] Li Xu, He Huang, and Jun Liu. Sutd-trafficqa: A question answering benchmark and an efficient network for video reasoning over traffic events. In CVPR, pages 9878–9888, 2021. 2, 5, 7" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[77] Hongyang Xue, Wenqing Chu, Zhou Zhao, and Deng Cai. A better way to attend: Attention with trees for video question answering. IEEE TIP, 27(11):5563-5574, 2018. 2", + "[78] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In ICCV, pages 1686-1697, 2021. 2", + "[79] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. NeurIPS, 2022. 2", + "[80] Zekun Yang, Noa Garcia, Chenhui Chu, Mayu Otani, Yuta Nakashima, and Haruo Takemura. Bert representations for video question answering. In WACV, pages 1556-1565, 2020. 7", + "[81] Yunan Ye, Zhou Zhao, Yimeng Li, Long Chen, Jun Xiao, and Yueting Zhuang. Video question answering via attribute-augmented attention network learning. In ACM SIGIR, pages 829-832, 2017. 2", + "[82] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In CVPR, pages 16375-16387, 2022. 2", + "[83] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. NeurIPS, 34:23634-23651, 2021. 7", + "[84] Zheng-Jun Zha, Jiawei Liu, Tianhao Yang, and Yongdong Zhang. Spatiotemporal-textual co-attention network for video question answering. ACM TOMM, 15(2s):1-18, 2019. 2", + "[85] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. In CVPR, 2022. 3", + "[86] Renrui Zhang, Longtian Qiu, Wei Zhang, and Ziyao Zeng. Vt-clip: Enhancing vision-language models with visual-guided texts. arXiv preprint arXiv:2112.02399, 2021. 3", + "[87] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021. 3", + "[88] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 3", + "[89] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 2, 3, 6, 8", + "[90] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, pages 8746-8755, 2020, 2, 7", + "[91] Yueting Zhuang, Dejing Xu, Xin Yan, Wenzhuo Cheng, Zhou Zhao, Shiliang Pu, and Jun Xiao. Multichannel attention refinement for video question answering. ACM TOMM, 16(1s):1-23, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 842 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13955", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/4a5fd117-e0c8-457c-9985-c705ba827466_model.json b/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/4a5fd117-e0c8-457c-9985-c705ba827466_model.json new file mode 100644 index 0000000000000000000000000000000000000000..de3c4a31027d6fd97b20a87a0861cf9be0007345 --- /dev/null +++ b/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/4a5fd117-e0c8-457c-9985-c705ba827466_model.json @@ -0,0 +1,2818 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.131, + 0.868, + 0.155 + ], + "angle": 0, + "content": "Tem-adapter: Adapting Image-Text Pretraining for Video Question Answer" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.179, + 0.808, + 0.217 + ], + "angle": 0, + "content": "Guangyi Chen\\(^{1,2*}\\), Xiao Liu\\(^{5*}\\), Guangrun Wang\\(^{4}\\), Kun Zhang\\(^{1,2}\\), Philip H.S. Torr\\(^{4}\\), Xiao-Ping Zhang\\(^{3,6}\\), Yansong Tang\\(^{3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.217, + 0.684, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)Carnegie Mellon University, Pittsburgh PA, USA" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.234, + 0.786, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\)Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.252, + 0.762, + 0.27 + ], + "angle": 0, + "content": "3Shenzhen International Graduate School, Tsinghua University, China" + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.27, + 0.593, + 0.287 + ], + "angle": 0, + "content": "\\(^{4}\\)University of Oxford, UK" + }, + { + "type": "text", + "bbox": [ + 0.319, + 0.287, + 0.653, + 0.305 + ], + "angle": 0, + "content": "\\(^{5}\\)Eindhoven University of Technology, NL" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.305, + 0.652, + 0.323 + ], + "angle": 0, + "content": "\\(^{6}\\)Toronto Metropolitan University, Canada" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.355, + 0.889, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.515, + 0.893, + 0.547 + ], + "angle": 0, + "content": "Figure 1: Illustration of the domain gaps between the pre-trained CLIP and the downstream VideoQA. CLIP is trained to align visual and textual domains, while VideoQA requires understanding the temporal dynamics and complex semantics of videos." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.559, + 0.314, + 0.575 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.474, + 0.903 + ], + "angle": 0, + "content": "Video-language pre-trained models have shown remarkable success in guiding video question-answering (VideoQA) tasks. However, due to the length of video sequences, training large-scale video-based models incurs considerably higher costs than training image-based ones. This motivates us to leverage the knowledge from image-based pretraining, despite the obvious gaps between image and video domains. To bridge these gaps, in this paper, we propose Tem-Adapter, which enables the learning of temporal dynamics and complex semantics by a visual Temporal Aligner and a textual Semantic Aligner. Unlike conventional pretrained knowledge adaptation methods that only concentrate on the downstream task objective, the Temporal Aligner introduces an extra language-guided autoregressive task aimed at facilitating the learning of temporal dependencies, with the objective of predicting future states based on historical clues and language guidance that describes event progression. Besides, to reduce the semantic gap and adapt the textual representation for better event description, we introduce a Semantic Aligner that first designs a template to fuse" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.561, + 0.895, + 0.651 + ], + "angle": 0, + "content": "question and answer pairs as event descriptions and then learns a Transformer decoder with the whole video sequence as guidance for refinement. We evaluate Tem-Adapter and different pre-train transferring methods on two VideoQA benchmarks, and the significant performance improvement demonstrates the effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.672, + 0.633, + 0.687 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.896, + 0.85 + ], + "angle": 0, + "content": "Video Question Answering (VideoQA) is a task that aims to answer natural language questions based on the information available in observed videos. It has attracted considerable attention recently due to its promise to develop interactive AI systems capable of communicating with the dynamic visual environment using natural language. Despite significant advancements in recent years, VideoQA remains a challenging problem that requires models to comprehensively understand and dynamically align the semantics of both the visual world and natural language." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.863, + 0.882, + 0.877 + ], + "angle": 0, + "content": "1Our code can be found at: https://github.com/XLiu443/Tem-adapter" + }, + { + "type": "page_footnote", + "bbox": [ + 0.521, + 0.877, + 0.629, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.521, + 0.888, + 0.645, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.863, + 0.882, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13945" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.47, + 0.257 + ], + "angle": 0, + "content": "Recently, a variety of methods [79, 78, 82, 34, 74, 63] have demonstrated impressive results in utilizing the largescale vision-language pre-trained (VLP) model to enhance downstream VideoQA tasks. These methods pre-train models by aligning the video and language domains and then apply them to VideoQA tasks via fine-tuning or even zero-shot learning. Nevertheless, pre-training large-scale models necessitate a large number of video-text pairs, e.g., more than 10 million videos, and entails expensive computational costs. This motivates us to explore cheaper and lighter alternative pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.261, + 0.47, + 0.441 + ], + "angle": 0, + "content": "Using image-VLP models is one potential option as they also align the semantics of vision and language domains. Compared to video-based models, image-based models offer two significant advantages. Firstly, training image-based models is less expensive when considering an equal number of data pairs. Secondly, image-VLP models have made significant progress in recent years, with some of them widely available and used, such as CLIP [56]. However, as depicted in Figure 1, domain gaps persist between image-based pretraining and VideoQA tasks since image-based models are unable to learn the temporal dynamics and corresponding complex event semantics of the visual world." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.445, + 0.47, + 0.641 + ], + "angle": 0, + "content": "To address the challenge posed by the domain gaps, this paper proposes Tem-adapter , an adapter network that leverages the interaction between visual and textual modalities to learn temporal dynamics and complex semantics. Unlike existing methods that directly adapt visual and textual features with the objective of the downstream tasks, such as cross-modal matching in VideoQA, Tem-Adapter introduces an additional language-guided autoregressive task to facilitate learning of temporal dynamics, which enables visual representations to predict future states based on historical information and language guidance. Besides, we introduce cross-modal interactions to further reduce the semantic gap and refine textual representation." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.645, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Our approach consists of a visual Temporal Aligner and a textual Semantic Aligner. The Temporal Aligner uses a Transformer encoder to learn the temporal relations and refine visual features. To optimize it, we build an auto-regression model by a Transformer decoder to generate the future state with historical information. We also incorporate the language information for visual refinement by adding textual embeddings as the condition memory of the Transformer decoder. This auto-regression model is supervised by a reconstruction loss with ground truth future frames to encourage the learning of temporal dependencies. In addition to the visual branch, we introduce a Semantic Aligner for better event description by reducing the semantic gap between the original texts crawled from the net and question-answer pairs in VideoQA. First, we design templates to fuse textual questions and answers to generate declarative sentences to reduce the domain gap between training and downstream" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "languages. Then, a Transformer decoder is employed to refine textual embeddings by incorporating the entire video sequence as a memory condition for video-text interactions." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.257 + ], + "angle": 0, + "content": "We conduct experiments on two public VideoQA benchmarks including SUTD-TrafficQA [76] and MSR-VTTMC [75], and evaluate different categories of methods that transfer the pre-trained knowledge to downstream tasks, such as other adapter models [12], prompt learning [89, 21], and fully or partial finetuning. The significant improvement demonstrates the effectiveness of Tem-adapter to adapt the image-language pre-trained model for VideoQA tasks." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.272, + 0.64, + 0.287 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.297, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Video Question Answering. Compared with Image-based Question Answering, Video QA is more challenging due to the complex temporal dependency, which requires models to dynamically align the visual observation and natural language. Some previous methods [67, 3, 48, 81, 28, 65, 77, 27, 26, 4, 7, 6, 8, 41, 84, 91, 10, 11, 36, 13, 17, 22, 23, 51, 62, 72] learn the models with the given manually annotated datasets. Recently, more and more methods [79, 78, 82, 34, 74, 63, 29] propose to utilize the representation ability of large-scale pre-trained vision-language models to solve the VideoQA task. For example, VQA-T [78] pre-trains the vision language model with randomly sampled video-question-answer triplets in the HowTo100M [47] dataset. BiLM [79] pre-trains frozen bidirectional language models with some adapter layers on the WebVid10M [1] dataset and achieve excellent zero-shot VideoQA performance on the downstream tasks. However, pretraining the large-scale model is expensive, which motivates us to find a cheaper alternative manner. In this paper, we propose to use existing image-language pre-trained models, such as CLIP [56], to provide the alignment between the visual and textual domains. The most related method is ATP [2], which also applies CLIP as the pre-trained model to extract frame and language embeddings, and learns an atemporal probe to select the most discriminative frames to match the video and QA pairs. Different from ATP selecting frames to transfer the video to images, we propose Tem-adapter to learn the temporal dynamics and reduce the gap between the pre-trained image domain and downstream video domain." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.75, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Vision Language Pre-training. The goal of VLP is to help downstream tasks by pre-training the model on large-scale vision-text pairs [56, 37, 20]. It always crawls a huge amount of vision-text pairs from the web and trains the model in a self-supervised manner, such as reconstruction [40, 15, 9, 30], contrastive matching [56, 20, 18], or the combination of both two [38, 71, 25]. Different from image-VLP, video-VLP has more diverse proxy tasks, such as frame/language mask [66, 90], video-language matching [46], video/language generation [73, 45], and" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13946" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.465, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.148, + 0.323, + 0.16 + ], + "angle": 0, + "content": "(a) Finetuning" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.161, + 0.465, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.215, + 0.338, + 0.227 + ], + "angle": 0, + "content": "(b) Prompt learning" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.228, + 0.466, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.289, + 0.316, + 0.301 + ], + "angle": 0, + "content": "(c) Adapter" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.302, + 0.466, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.363, + 0.33, + 0.375 + ], + "angle": 0, + "content": "(d) Tem-Adapter" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.387, + 0.473, + 0.522 + ], + "angle": 0, + "content": "Figure 2: Various methods for pre-trained knowledge adaptation in VideoQA. (a) Fine-tuning updates pretrained parameters with downstream training. (b) Prompt learning freezes the pre-trained model and introduces learnable textual tokens for adaptation. (c) Adapter-based methods add new adapter layers to transfer knowledge. (d) Our approach, Tem-adapter , leverages the visual-textual interactions to learn temporal dynamics to reduce the gap between image-based pre-training and video-based tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.47, + 0.655 + ], + "angle": 0, + "content": "frame/language order prediction [33, 39]. Recently, VLP models have shown great potential for other downstream applications, such as zero-shot and few-shot visual recognition [56, 12, 89, 88, 86], segmentation and detection [60, 87], image generation [50, 59, 52], and action understanding [70, 68]. In this paper, we use the CLIP model as our pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Pre-trained Knowledge Transfer. In light of the remarkable success of the large-scale pre-trained model, many methods have been proposed to solve the problem that how to efficiently adapt the pre-trained knowledge to the downstream tasks. Early methods finetune all parameters of the pre-trained model with the downstream objectives [14, 53]. However, with the model scale increasing, the cost of finetuning the full model is extremely expensive. Thus, an increasing number of methods focus on knowledge adaptation with light parameters. A straightforward way is to partially finetune the model parameters [49]. Besides, prompt learning methods are proposed recently to reformulate the downstream tasks as the pre-trained tasks and freeze all model parameters [43, 54, 58, 55]. The previous prompt-based methods [54, 55] design the template of prompt using the prior of natural human language. Then, some methods [24, 64]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.289 + ], + "angle": 0, + "content": "propose to automatically optimize the prompt template, and [42, 69, 44, 89, 5] further learn the prompts in the continuous embedding space where the discrete word constraint is relaxed. In addition, some methods freeze the model and train some extra adapter layers and also achieve the supersizing performance [12, 85, 1]. As shown in Figure 2, we illustrate the difference between Tem-adapter and other categories of methods transferring the image-based VLP model to the VideoQA task. Compared with other pre-train adaptation methods, our Tem-adapter proposes an autoregression task to reduce the temporal gap (tem-align) and introduce the cross-domain interactions (sem-align) to reduce the semantic gap between the pre-training and downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.306, + 0.609, + 0.323 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.894, + 0.409 + ], + "angle": 0, + "content": "In this section, we first briefly introduce the VideoQA task and then present our approach, Tem-adapter , which introduces the cross-modal interactions to learn temporal dynamics and complex semantics. Finally, we give the details of our implementation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.422, + 0.683, + 0.437 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.893, + 0.492 + ], + "angle": 0, + "content": "The goal of VideoQA is to find an answer \\(\\hat{\\mathbf{a}}\\) from the answer pool, given the conditions of question \\(\\mathbf{q}\\) and observed video data \\(\\mathbf{v}\\) as below:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.498, + 0.893, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {a}} = \\underset {\\mathbf {a} \\in \\mathcal {A}} {\\arg \\max } \\mathcal {S} (\\mathbf {a} | \\mathbf {v}, \\mathbf {q}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.527, + 0.892, + 0.558 + ], + "angle": 0, + "content": "where \\( S \\) is a match function to generate the score of each answer candidate given the question and video content." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.571, + 0.64, + 0.587 + ], + "angle": 0, + "content": "3.2. Tem-Adapter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.895, + 0.717 + ], + "angle": 0, + "content": "Here, we propose Tem-adapter to learn temporal dynamics and complex semantics by the interactions of visual and textual domains to adapt pre-trained Image-VLP models to downstream VideoQA tasks. Figure 3 shows an overview of Tem-Adapter's framework. In particular, Tem-adapter consists of two components, including a visual Temporal Aligner and a textual Semantic Aligner. Details for each component are as follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.894, + 0.794 + ], + "angle": 0, + "content": "Temporal Aligner. The goal of the Temporal Aligner is to reduce the gap between image-based pre-training and video-based downstream tasks. To achieve this goal, we introduce an auxiliary language-guided autoregressive task to facilitate the learning of temporal dynamics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Specifically, given a video \\(\\mathbf{v}\\), we apply the frozen CLIP [56] image encoder \\(f\\) to extract the feature of each frame as \\(f(\\mathbf{v}) = \\{f(\\mathbf{v}_t)|_{t = 1}^T\\} \\in \\mathcal{R}^{T\\times C}\\), where \\(T\\) is the frame number of videos and \\(C\\) is the feature dimension of CLIP. Then we refine these features with our Temporal Aligner model which is built by a transformer encoder \\(f_{e}\\) with parameters \\(\\phi\\) to learn the temporal relations of frames." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13947" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.873, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.28, + 0.894, + 0.342 + ], + "angle": 0, + "content": "Figure 3: The overall framework of Tem-adapter. Our Tem-adapter consists of two modules, Semantic Aligner (orange) and Temporal Aligner (blue). Semantic Aligner takes the text embedding of template-based QA pair as the input, and applies a Transformer decoder to incorporate cross-modal interactions. The Temporal Aligner introduces a language-guided autoregressive task to learn temporal dynamics. The training is guided by classification and reconstruction losses." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.361, + 0.361, + 0.377 + ], + "angle": 0, + "content": "The refined features can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.388, + 0.471, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} = f _ {e} (f (\\mathbf {v}); \\phi), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.367, + 0.431 + ], + "angle": 0, + "content": "which are with the same dimension \\(\\mathcal{R}^{T\\times C}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.471, + 0.522 + ], + "angle": 0, + "content": "To guide the Temporal Aligner to learn temporal dynamics, we further build an autoregressive sequential model, following GPT [57], which generates the feature of the current frame with previous frames and guidance from textual information that describes event progression. Mathematically, it can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.527, + 0.47, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\hat {f} (\\mathbf {v} _ {i}) = f _ {d} \\left(f \\left(\\mathbf {v} _ {1: i - 1}\\right) | \\mathbf {f}, \\mathbf {g}; \\psi\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.473, + 0.718 + ], + "angle": 0, + "content": "where \\( f(\\mathbf{v}_{1:i-1}) \\) is the input of the transformer decoder, which is achieved by applying an attention mask to focus on the history clues. \\( \\mathbf{f}, \\mathbf{g} \\) are used as the guidance information, where \\( \\mathbf{f} \\) denotes latent visual embeddings obtained by Temporal Aligner and \\( \\mathbf{g} \\) (see equation 5) is the textual embeddings describing temporal event. \\( \\mathbf{f} \\) and \\( \\mathbf{g} \\) are first fused by adding and then serve as the memory key and value of the Transformer decoder. We apply a reconstruction loss to encourage the learning of temporal dependencies which is formulated as the distance between the results of the auto-regression model \\( \\hat{f}(\\mathbf{v}_i) \\) and the ground-truth \\( f(\\mathbf{v}_i) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.722, + 0.47, + 0.761 + ], + "angle": 0, + "content": "\\[\nL _ {M S E} = \\sum_ {i = 1} ^ {n} | | f (\\mathbf {v} _ {i}) - \\hat {f} (\\mathbf {v} _ {i}) | | _ {2} ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Semantic Aligner. In addition to the visual branch, we introduce a Semantic Aligner to adapt the textual representation for better event description and reduce the semantic gaps. We first design templates to align the question-answer pairs in VideoQA to declarative sentences, since the training data of CLIP is crawled from the web which always describes the image in a declarative manner. Then, we apply a Transformer decoder with the video sequence as the memory for video-text interactions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.895, + 0.587 + ], + "angle": 0, + "content": "Given a question \\(\\mathbf{q}\\) and a corresponding answer \\(\\mathbf{a}\\), we first apply the grammar parsing on the question \\(\\mathbf{q}\\) to obtain the subject-verb-object structure, and then we modify the order of the sentence to obtain a declarative sentence and leave the position of the interrogative pronouns as a mask '[']. Finally, we fill in the default '['] with the answer to update the QA pair as a refined event description. For instance, the question is 'Which might be the reason for this accident?', and the answer is 'Improper lane change'. We first obtain a declarative sentence with the default as '[ ] is the reason for this accident', and then we can fill in the answer to get 'Speeding vehicle is the reason for this accident.' In addition, for the general question, we first change it to a declarative sentence and then convert the sentence to negative form if the answer is No." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.892, + 0.693 + ], + "angle": 0, + "content": "With the refined sentence \\( \\mathbf{q}\\mathbf{a} \\) as the input, the frozen text encoder \\( g \\) of the CLIP model outputs the textual feature as \\( g(\\mathbf{q}\\mathbf{a}) \\in \\mathcal{R}^C \\). To further refine the textual feature using the video content, we apply a transformer decoder with the visual feature sequence as the memory key and value and add a residual link to avoid overfitting. Formally, it can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.7, + 0.892, + 0.718 + ], + "angle": 0, + "content": "\\[\n\\mathbf {g} = g (\\mathbf {q a}) + \\lambda g _ {d} (g (\\mathbf {q a}), f (\\mathbf {v}); \\theta), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.722, + 0.895, + 0.827 + ], + "angle": 0, + "content": "where \\(\\mathbf{g}\\) denotes the learned textual embedding, \\(\\lambda\\) is a parameter of the residual link, and \\(\\theta\\) is the parameters of the transformer decoder \\(g_{d}\\). Both \\(\\lambda\\) and \\(\\theta\\) can be learned automatically in the training process. With the cross-attention module in the transformer decoder, we encourage the model to find key information from the video to refine the textual embedding for better event description." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.827, + 0.894, + 0.873 + ], + "angle": 0, + "content": "Cross-modal Matching. Given the refined visual and textual embeddings \\(\\mathbf{f}\\) and \\(\\mathbf{g}\\), we instantiate the score function in equation 1 by cosine similarity as:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.884, + 0.893, + 0.903 + ], + "angle": 0, + "content": "\\[\nS (\\mathbf {a} | \\mathbf {v}, \\mathbf {q}) = \\operatorname {s o f t m a x} (\\bar {\\mathbf {f}} ^ {T} \\mathbf {g}), \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13948" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.152 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{f}}\\) is video representation using the average pool on all frames, \\(\\mathbf{g}\\) is the language embedding with given answer question pair \\((\\mathbf{a},\\mathbf{q})\\). Then, we can obtain the answer to the question with the highest matching score." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.153, + 0.47, + 0.199 + ], + "angle": 0, + "content": "Then we apply a standard multi-choice classification Hinge loss [19] integrated with the auto-regression loss for training. The overall loss function can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.205, + 0.47, + 0.222 + ], + "angle": 0, + "content": "\\[\nL = L _ {H i n g e} (S (\\mathbf {a} | \\mathbf {v}, \\mathbf {q}), y) + \\gamma L _ {M S E}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.471, + 0.348 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a ratio to balance two loss functions, and \\(y\\) is the ground truth label of the answers. Please note that our Tem-adapter consists of 3 learnable modules including the textual decoder \\(\\theta\\), the visual encoder \\(\\phi\\), and the visual decoder \\(\\psi\\). The Hinge loss is applied to optimize \\(\\theta\\) and \\(\\phi\\), while the auto-regression loss \\(L_{MSE}\\) is used to optimize all three modules. In the inference, only textual decoder \\(\\theta\\) and visual encoder \\(\\phi\\) are used." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.36, + 0.295, + 0.377 + ], + "angle": 0, + "content": "3.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.473, + 0.414 + ], + "angle": 0, + "content": "In this subsection, we provide more implementation details of our approach for better reproduction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.471, + 0.597 + ], + "angle": 0, + "content": "Network Architecture. Figure 3 shows the framework of Tem-adapter. In our experiments, we use the frozen model CLIP (ViT-B32) as the text and image encoders. The Semantic Aligner is a one-layer Transformer decoder with 16 heads. There is a linear layer before the Transformer decoder which maps text embeddings and image embeddings to a lower dimension (128). Another linear projection layer is added after the Transformer decoder to map the reduced dimension back to 512. The Temporal Aligner is a light transformer encoder-decoder architecture. Both the encoder and decoder contain only one layer. We set the number of heads to 16 in the experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.471, + 0.749 + ], + "angle": 0, + "content": "Inference Details. In the inference (given 1 question, 1 video, and \\( k \\) answers), we first apply the template to fuse the question and \\( k \\) answers to obtain \\( k \\) QA pairs (e.g. \\( k = 4 \\)). Then we feed the QA pairs and video into CLIP textual and visual encoders to obtain the initial features. After that, we apply the learned semantic and temporal aligners to refine these features. Please note that the temporal auto-regression module is not used in the inference. Finally, we obtain 1 video feature with \\( k \\) QA features and apply the cosine similarity as well as softmax to select the answer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Hyper-parameters. We follow the same video frames sampling method with prior methods [76]. We uniformly select 8 frames over a long video sequence and then pick consecutive 16 frames which are centered on those 8 frames. Overall 128 frames are selected for each video. We also tried another sampling method: uniformly extracting 128 or 256 frames on a video sequence, and the final results do not support the sampling method very well. In the loss function, we set \\(\\gamma\\) as 100 to encourage the model to learn the temporal dynamics. And the learnable parameter \\(\\lambda\\) in the Semantic" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Aligner is set to be the same dimension (512) as the output embeddings of the CLIP model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.894, + 0.244 + ], + "angle": 0, + "content": "Training Details. We use the pre-trained CLIP model to extract text embeddings and image embeddings. The Adam [31] optimizer is utilized to optimize the training process. We set the initial learning rate to be \\( 1e - 4 \\) with a batch size of 128. And we let the learning rate decay with a factor of 0.5 for every 10 epochs. Our model is trained for 50 epochs on a single Tesla V100 GPU and the training takes 8 hours to finish. Our model is implemented in PyTorch 1.8.0." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.26, + 0.634, + 0.277 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.286, + 0.895, + 0.348 + ], + "angle": 0, + "content": "In this section, we first benchmark our Tem-adapter and different categories of methods that transfer the pretrained model into downstream tasks. Then we provide ablation studies to investigate the effect of each component." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.358, + 0.807, + 0.374 + ], + "angle": 0, + "content": "4.1. Datasets & Experimental Protocols" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.895, + 0.593 + ], + "angle": 0, + "content": "SUTD-TrafficQA. SUTD-TrafficQA focuses on the question-answering task in the traffic situation, which requires the model to understand the traffic event and underlying causal relation. It consists of more than 10k videos with different traffic events and provides over 62,535 human-annotated QA pairs. Among them, 56,460 QA pairs are used for training and the rest 6,075 QA pairs are used for testing. This dataset highlights the cognitive capability of video event understanding models in complex traffic scenarios. It provides 6 challenging traffic-related reasoning tasks including \"Basic understanding\", \"Event forecasting\", \"Reverse reasoning\", \"Counterfactual inference\", \"Introspection\", \"Attribution\". All tasks are formulated as multiple-choice without limiting the number of candidate answers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.895, + 0.791 + ], + "angle": 0, + "content": "MSR-VTT-MC. MSR-VTT is collected and released mainly for the text-to-video retrieval task. Following [34] and [2], we use the multi-choice test setting to benchmark the VideoQA methods. The MSR-VTT-MC test set consists of 2990 QA pairs. Each video corresponds to five answers, with the original caption as the correct answer. We use the standard protocol with prior works in our experiments, where 7010 train+val video sequences and 140,200 QA pairs are utilized for training. Each video only contains one caption. We assign this caption as the correct answer for our question-answering task. To formulate four negative answers during training, we randomly select four captions from other captions that do not belong to the current video." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.802, + 0.673, + 0.817 + ], + "angle": 0, + "content": "4.2. Baseline Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To investigate different categories of methods that transfer the pre-trained model into downstream tasks, such as finetuning, prompt learning, and adapter. We compare Tem-adapter with the other 10 baseline methods including Unsupervised CLIP [56], Unsupervised CLIP [56]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13949" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.138 + ], + "angle": 0, + "content": "Table 1: VideoQA accuracies of baseline methods on both SUTD-TrafficQA and MSR-VTT-MC. B: \"Basic understanding\", F: \"Forecasting task\", R: \"Reverse Reasoning\", C: \"Counterfactual inference\", I: \"Introspection\", A: \"Attribution\". (C) and (A) denote strategies for adding prompts. (C) and \\((\\mathbf{C}^{*})\\) denote training prompts with/without adapter heads." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.145, + 0.844, + 0.372 + ], + "angle": 0, + "content": "
MethodsSUTD-TrafficQAMSR-VTT-MC
BFRCIAAvgAcc
Unsupervised CLIP [56]25.620.134.030.822.828.826.574.1
CLIP [56] + Template31.836.029.971.822.133.432.374.1
Totally finetuning39.835.146.645.637.240.540.389.0
Partially finetuning41.637.844.650.033.141.741.788.5
LORA [16]38.738.736.737.934.538.138.385.4
CLIP-Adapter [12]35.832.035.442.333.132.134.883.4
Multi-layer Adapter [12]30.526.626.538.528.325.829.190.7
Prompt learning (C) [89]42.432.445.255.540.743.642.989.0
Prompt learning (C*) [89]40.333.241.046.534.938.439.790.8
Prompt learning (A) [21]41.731.540.148.433.141.441.188.0
Tem-adapter46.036.544.655.034.547.746.094.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.378, + 0.472, + 0.455 + ], + "angle": 0, + "content": "Table 2: Ablation studies on the SUTD-TrafficQA and MSR-VTT-MC datasets. Tp, SA, TA, and Ar denote Template, Semantic Aligner, Temporal Aligner, and Auto-regression, respectively. No ablation studies (“\\”) concentrate on Tp, as it is not compatible with the MSR-VTT-MC dataset." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.465, + 0.462, + 0.624 + ], + "angle": 0, + "content": "
MethodsSUTD-TrafficQAMSR-VTT-MC
AvgAcc
CLIP-Adapter [12]34.883.4
w/o Tp43.9\\
w/o SA44.291.8
w/o TA43.691.6
w/o Ar44.992.6
w/o TA and SA32.374.1
Tem-adapter46.094.3
" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.639, + 0.472, + 0.732 + ], + "angle": 0, + "content": "+ Language template, Totally fine-tuning, Partially finetuning, LORA [16], CLIP-Adapter [12], Multi-layer CLIP-Adapter [12], Prompt learning (change words) [89] (with/without using adapter heads), and Prompt learning (add words) [21]. The detailed descriptions are shown in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.741, + 0.427, + 0.758 + ], + "angle": 0, + "content": "4.3. Benchmarking against Baseline Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.84 + ], + "angle": 0, + "content": "In this subsection, we compare and discuss our approach's performance against the above-mentioned baselines on both SUTD-TrafficQA and MSR-VTT-MC datasets. The results of all methods on both datasets are summarized in Table 1. We can draw conclusions as follows from the results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "CLIP can obtain good zero-shot results. With our language template, the clip model can achieve \\(32.3\\%\\) accuracy on the SUTD-TrafficQA dataset, which is almost close to the previous state-of-the-art performance \\((37.1\\%)\\). It shows" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.381, + 0.842, + 0.396 + ], + "angle": 0, + "content": "the strong representation ability of the CLIP model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.895, + 0.626 + ], + "angle": 0, + "content": "More parameters \\(\\neq\\) Better performance. We have two groups of experiments that focus on the number of parameters, including totally/partially finetuning and single-layer/multi-layer Adapters. We obtain contradictory conclusions from different datasets. On SUTD-TrafficQA, the model with fewer parameters works better, while on MSR-VTT-MC, the model with more parameters has better performance. It is not surprising since the scale of training data is different. MSR-VTT-MC has more training data and can support training more parameters. We also compare the learnable parameters of adapter-based methods in Table 5. Tem-adapter achieve better performance with fewer parameters than Multi-layer Adapter. Compared with the prompt-learning-based method, our method needs more parameters, but its inference is not time-consuming." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.896, + 0.764 + ], + "angle": 0, + "content": "Different categories of methods are comparable. Balancing the performance on both datasets, different categories of methods that transfer the image-based VLP model to VideoQA don't have clear performance differences. Adapter-based methods obtain poor results on SUTD-TrafficQA, but multi-layer Adapter works well on MSR-VTT-MC. Besides, by comparing the prompt learning methods with/without adapter heads, we found that the prompt and adapter can be joint learning to obtain better performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Cross-modal interaction and temporal dynamics matter. Compared with CLIP-Adapter (both single-layer and multi-layer ones), our Tem-adapter achieves more than \\(10\\%\\) improvement on SUTD-TrafficQA. The key difference between Tem-adapter and other adapter-based methods is that we introduce temporal alignment to reduce the temporal gap and apply the cross-modal interaction to refine the representation and reduce the semantic gap. Tem-adapter outperforms other methods by a large mar" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "13950" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.12 + ], + "angle": 0, + "content": "Table 3: Comparison with the state-of-the-art methods on the SUTD-TrafficQA dataset." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.13, + 0.447, + 0.307 + ], + "angle": 0, + "content": "
MethodsAccuracyYear
Human95.4-
VIS+LSTM [61]29.92015
TVQA [35]35.22018
HCRN [32]36.52020
BERT-VQA [80]33.72020
Eclipse [76]37.12021
CLIP [56]32.32021
CLIP+HCRN[56, 32]41.92021
ATP [2]35.62022
Tem-adapter46.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.471, + 0.365 + ], + "angle": 0, + "content": "gin on both two datasets. For example, it obtains \\(3.1\\%\\) improvement than Prompt learning (C) on SUTD-TrafficQA and \\(3.6\\%\\) than Multi-layer Adapter on MSR-VTT-MC." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.374, + 0.468, + 0.391 + ], + "angle": 0, + "content": "4.4. Comparison against State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.472, + 0.638 + ], + "angle": 0, + "content": "SUTD-TrafficQA. On SUTD-TrafficQA, we compare our proposed method with state-of-the-art methods, including [32, 76, 61, 35, 80]. The results are summarized in Table 3. We can observe that our Tem-adapter outperforms other methods by a large margin. For example, compared with HCRN [32], Tem-adapter achieves almost \\(10\\%\\) improvement. By using the CLIP model to extract visual features, HCRN+CLIP can obtain \\(41.9\\%\\) accuracy, which is better than using the original HCRN. Besides, we show that the performance improvement is not only from the better representation ability of the CLIP model since our Tem-adapter further improve HCRN+CLIP with \\(+4.1\\%\\) accuracy. In addition, we also significantly outperform the ATP method, which also uses the CLIP model. This significant performance improvement demonstrates the effectiveness of Tem-adapter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.472, + 0.838 + ], + "angle": 0, + "content": "MSR-VTT-MC. On MSR-VTT-MC, we compare Tem-adapter with some recent state-of-the-art methods, including ActBERT[90], ClipBERT[34], MERLOT [83], VideoCLIP [66], CLIP [56], and ATP [2]. Results are shown in Table 4. First, we can find that the CLIP model benefits the performance on the MSR-VTT-MC dataset. All methods using the CLIP model, such as CLIP [56], ATP [2], and Tem-adapter achieve good performance. ATP [2] learns an atemporal probe to select a frame from a video sequence to obtain the visual embedding and achieve good performance on MSR-VTT-MC. Compared with it, Tem-adapter can further obtain \\(94.3\\%\\) accuracy and \\(1.1\\%\\) relative improvement." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.245, + 0.862 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "We provide ablation studies to investigate the effect of each component. Four main modules in our Tem-adapter" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.894, + 0.12 + ], + "angle": 0, + "content": "Table 4: Comparison with the state-of-the-art methods on the MSR-VTT-MC dataset." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.13, + 0.853, + 0.273 + ], + "angle": 0, + "content": "
MethodsAccuracyYear
VideoCLIP [66]92.12019
ActBERT[90]85.72020
ClipBERT[34]88.22021
MERLOT [83]90.92021
CLIP [56]74.12021
ATP [2]93.22022
Tem-adapter94.3
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.288, + 0.894, + 0.318 + ], + "angle": 0, + "content": "are evaluated, including Temporal Aligner, Semantic Aligner, language template, and auto-regression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.895, + 0.531 + ], + "angle": 0, + "content": "As shown in Table 2, we provide the performance of different variants of Tem-adapter and a baseline adapter method on each task in the SUTD-TrafficQA and MSR-VTTMC datasets. The baseline method is CLIP-adapter, which also uses the language template and learns a linear adapter layer without cross-domain interaction. We observe that all variants of Tem-adapter can achieve significant improvement over the baseline method, which shows that the interaction across visual and textual is very important for video analysis. Besides, compared with Tem-adapter, we can observe a clear performance drop when we remove each component, which demonstrates all of them are indispensable. When both temporal and semantic aligners are removed, the performance drops dramatically." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.894, + 0.637 + ], + "angle": 0, + "content": "In addition, we also conducted an experiment to show whether the Temporal Aligner learns the temporal dynamic information. Table 6 shows the reconstruction quality of our methods with/without language guidance \\((g)\\) and with/without Temporal Aligner \\((f)\\). We can observe that both the language guidance and Temporal Aligner assist in the learning of temporal dynamics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.895, + 0.714 + ], + "angle": 0, + "content": "Furthermore, we also provide some parameter analysis experiments about our network design. And evaluate Tem-adapter on different image-text pre-train models. Please kindly refer to the supplementary materials for the detailed results and analysis." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.638, + 0.74 + ], + "angle": 0, + "content": "4.6. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this subsection, we provide some qualitative results to analyze our approach. As shown in Figure 4, we show some visualization examples of VdideoQA on the SUTD-TrafficQA dataset. In the top row, we plot two positive examples. We can observe that our Tem-adapter learns the temporal dynamics with the cross-domain interactions to obtain a better understanding of video events. Besides, we also provide some failure cases to help explore the boundary of our approach. We find that Tem-adapter can not work well for situations that need complex reasoning, especially" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13951" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.127, + 0.094, + 0.421, + 0.109 + ], + "angle": 0, + "content": "Q: What could possibly cause this accident?" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.11, + 0.468, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.168, + 0.411, + 0.181 + ], + "angle": 0, + "content": "Ours: Sudden or extreme movement by a vehicle." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.181, + 0.34, + 0.194 + ], + "angle": 0, + "content": "CLIP-Adapter: Unfavorable weather." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.194, + 0.368, + 0.207 + ], + "angle": 0, + "content": "Prompt learning: Violation of traffic rules." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.207, + 0.338, + 0.221 + ], + "angle": 0, + "content": "Finetuning: Violation of traffic rules." + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.233, + 0.422, + 0.262 + ], + "angle": 0, + "content": "Q: Could any vehicles steer away to prevent the accident?" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.267, + 0.462, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.324, + 0.347, + 0.337 + ], + "angle": 0, + "content": "Ours: No, there was no space. X" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.337, + 0.343, + 0.351 + ], + "angle": 0, + "content": "GT: Yes, there was enough space." + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.094, + 0.818, + 0.109 + ], + "angle": 0, + "content": "Q: Which factors contribute to the accident?" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.111, + 0.871, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.543, + 0.168, + 0.818, + 0.181 + ], + "angle": 0, + "content": "Ours: Not paying attention to their surroundings." + }, + { + "type": "text", + "bbox": [ + 0.543, + 0.181, + 0.735, + 0.193 + ], + "angle": 0, + "content": "CLIP-Adapter: Bad road surfaces" + }, + { + "type": "text", + "bbox": [ + 0.543, + 0.194, + 0.747, + 0.207 + ], + "angle": 0, + "content": "Prompt learning: Bad road surfaces." + }, + { + "type": "text", + "bbox": [ + 0.543, + 0.207, + 0.717, + 0.221 + ], + "angle": 0, + "content": "Finetuning: Speeding vehicles." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.234, + 0.835, + 0.265 + ], + "angle": 0, + "content": "Q: Could the accident be prevented if a convex mirror is installed at the turning point?" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.267, + 0.869, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.324, + 0.812, + 0.336 + ], + "angle": 0, + "content": "Ours: Yes, there would be no obstructed views." + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.337, + 0.841, + 0.349 + ], + "angle": 0, + "content": "GT: No, that was not the main cause of the accident." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.372, + 0.892, + 0.434 + ], + "angle": 0, + "content": "Figure 4: Visualization of some examples of the VideoQA task from the SUTD-TrafficQA dataset. The top row shows some positive examples, where Tem-adapter learns temporal dynamics to understand the traffic event and give the correct answers. The bottom row shows some failed cases. Despite better matching between visual and textual domains, our approach fails on complex problems which require reasoning ability." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.442, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Table 5: The comparisons of our method and other baseline methods on model parameters (M), inference time (ms/video), and accuracy in the SUTD-TrafficQA dataset." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.477, + 0.9, + 0.538 + ], + "angle": 0, + "content": "
MetricsLORA [16]Prompt learning (C) [89]Prompt learning (C*) [89]CLIP-Adapter [12]Multi-layer Adapter [12]Ours
Parameters1.443.550.771.054.853.62
Inference time5.375.335.175.135.145.32
Accuracy38.342.939.734.829.146.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.573 + ], + "angle": 0, + "content": "Table 6: The comparison of reconstruction ability. We use peak signal-to-noise ratio (PSNR) as the metric." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.577, + 0.446, + 0.623 + ], + "angle": 0, + "content": "
Tem-adapterw/o gw/o f and g
PSNR29.128.727.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.47, + 0.692 + ], + "angle": 0, + "content": "for the underlying cause-effect relations. We leave it as a future work to enhance the reasoning ability of our model. Besides, we also provide more examples of our language templates in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.708, + 0.196, + 0.724 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.902 + ], + "angle": 0, + "content": "In this paper, we proposed Tem-adapter to adapt the pre-trained image-language model to the downstream VideoQA task. We jointly learn the Temporal Aligner and Semantic Aligner through cross-modal interactions. Tem-adapter introduces a language-guided autoregressive task to guide the learning of temporal dependency and thus reduce the temporal gap between image-based pre-train and video-based QA tasks. Besides, we utilize a rule-based template and video-based interaction to refine textual representation and reduce semantic gaps. We validate the effectiveness of our approach on four VideoQA benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.895, + 0.665 + ], + "angle": 0, + "content": "Broader Impacts and Limitations. Understanding how the big model adaptation works on VideoQA is of great help in developing general foundation models and interactive AI. However, our work also has certain limitations. For example, since extracting video features online takes up too many computing resources, we extract their features offline and fix them. We leave it as future work to learn the adapter within the backbone network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.652, + 0.696 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.902 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62206153, Young Elite Scientists Sponsorship Program by CAST (No. 2023QNRC002), by the Tsinghua Shenzhen International Graduate School-Shenzhen Pengrui Endowed Professorship Scheme of Shenzhen Pengrui Foundation, by the UKRI grant: Turing AI Fellowship EP/W002981/1, by the National Institutes of Health (NIH) under Contract R01HL159805, by the NSF-Convergence Accelerator Track-D award #2134901, by a grant from Apple Inc., a grant from KDDI Research Inc, and generous gifts from Salesforce Inc., Microsoft Research, and Amazon Research. We would also like to thank the Royal Academy of Engineering and FiveAI." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13952" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.157 + ], + "angle": 0, + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the\" video\" in video-language understanding. In \\(CVPR\\), pages 2917-2927, 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[3] Santiago Castro, Mahmoud Azab, Jonathan Stroud, Cristina Noujaim, Ruoyao Wang, Jia Deng, and Rada Mihalcea. Lifeqa: A real-life dataset for video question answering. In LREC, pages 4352-4358, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.472, + 0.326 + ], + "angle": 0, + "content": "[4] Aman Chadha, Gurneet Arora, and Navpreet Kaloty. iperceive: Applying common-sense reasoning to multi-modal dense video captioning and video question answering. arXiv preprint arXiv:2011.07735, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.472, + 0.381 + ], + "angle": 0, + "content": "[5] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.381, + 0.472, + 0.45 + ], + "angle": 0, + "content": "[6] Seongho Choi, Kyoung-Woon On, Yu-Jung Heo, Ahjeong Seo, Youwon Jang, Minsu Lee, and Byoung-Tak Zhang. Dramaqa: Character-centered video story understanding with hierarchical qa. In AAAI, volume 35, pages 1166-1174, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.451, + 0.472, + 0.507 + ], + "angle": 0, + "content": "[7] Anthony Colas, Seokhwan Kim, Franck Dernoncourt, Siddhesh Gupte, Daisy Zhe Wang, and Doo Soon Kim. Tutorial: Question answering dataset for tutorial videos. arXiv preprint arXiv:1912.01046, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.472, + 0.55 + ], + "angle": 0, + "content": "[8] Long Hoang Dang, Thao Minh Le, Vuong Le, and Truyen Tran. Object-centric representation learning for video question answering. In IJCNN, pages 1-8, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.472, + 0.619 + ], + "angle": 0, + "content": "[9] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Zicheng Liu, Michael Zeng, et al. An empirical study of training end-to-end vision-and-language transformers. arXiv preprint arXiv:2111.02387, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.62, + 0.472, + 0.676 + ], + "angle": 0, + "content": "[10] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In CVPR, pages 1999-2007, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.676, + 0.472, + 0.718 + ], + "angle": 0, + "content": "[11] Jiyang Gao, Runzhou Ge, Kan Chen, and Ram Nevatia. Motion-appearance co-memory networks for video question answering. In CVPR, pages 6576–6585, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.718, + 0.472, + 0.774 + ], + "angle": 0, + "content": "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.774, + 0.472, + 0.816 + ], + "angle": 0, + "content": "[13] Noa Garcia, Mayu Otani, Chenhui Chu, and Yuta Nakashima. Knowit vqa: Answering knowledge-based questions about videos. In AAAI, volume 34, pages 10826-10834, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.816, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[14] Ross Girshick. Fast r-cnn. In ICCV, pages 1440-1448, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.845, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[15] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. VIn bert: A recurrent vision-and-language bert for navigation. In CVPR, pages 1643–1653, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.207 + ], + "angle": 0, + "content": "[17] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware graph convolutional networks for video question answering. In AAAI, volume 34, pages 11021-11028, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.894, + 0.264 + ], + "angle": 0, + "content": "[18] Aashi Jain, Mandy Guo, Krishna Srinivasan, Ting Chen, Sneha Kudugunta, Chao Jia, Yinfei Yang, and Jason Baldridge. Mural: multimodal, multitask retrieval across languages. arXiv preprint arXiv:2109.05125, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.894, + 0.32 + ], + "angle": 0, + "content": "[19] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In CVPR, pages 2758–2766, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.323, + 0.894, + 0.392 + ], + "angle": 0, + "content": "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.394, + 0.894, + 0.438 + ], + "angle": 0, + "content": "[21] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.438, + 0.894, + 0.495 + ], + "angle": 0, + "content": "[22] Jianwen Jiang, Ziqiang Chen, Haojie Lin, Xibin Zhao, and Yue Gao. Divide and conquer: Question-guided spatiotemporal contextual attention for video question answering. In AAAI, volume 34, pages 11101-11108, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.496, + 0.894, + 0.539 + ], + "angle": 0, + "content": "[23] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In AAAI, volume 34, pages 11109-11116, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.54, + 0.894, + 0.582 + ], + "angle": 0, + "content": "[24] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? TACL, 8:423-438, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.584, + 0.894, + 0.64 + ], + "angle": 0, + "content": "[25] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1780-1790, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.641, + 0.894, + 0.696 + ], + "angle": 0, + "content": "[26] Hyoungun Kim, Zineng Tang, and Mohit Bansal. Dense-ecaption matching and frame-selection gating for temporal localization in videoqa. arXiv preprint arXiv:2005.06409, 2020.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.699, + 0.894, + 0.755 + ], + "angle": 0, + "content": "[27] Junyeong Kim, Minuk Ma, Trung Pham, Kyungsu Kim, and Chang D Yoo. Modality shifting attention network for multimodal video question answering. In CVPR, pages 10106-10115, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.757, + 0.894, + 0.8 + ], + "angle": 0, + "content": "[28] Kyung-Min Kim, Min-Oh Heo, Seong-Ho Choi, and Byoung-Tak Zhang. Deepstory: Video story qa by deep embedded memory networks. arXiv preprint arXiv:1707.00836, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.8, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[29] Seonhoon Kim, Seohyeong Jeong, Eunbyul Kim, Inho Kang, and Nojun Kwak. Self-supervised pre-training and contrastive representation learning for multiple-choice video qa. In AAAI, volume 35, pages 13171-13179, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[30] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, pages 5583-5594, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13953" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[32] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In CVPR, pages 9972-9981, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[33] Chenyi Lei, Shixian Luo, Yong Liu, Wanggui He, Jiamang Wang, Guoxin Wang, Haihong Tang, Chunyan Miao, and Houqiang Li. Understanding chinese video and language via contrastive multimodal pre-training. In ACM MM, pages 2567-2576, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[34] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, pages 7331-7341, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[35] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. arXiv preprint arXiv:1809.01696, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[36] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvqa+: Spatio-temporal grounding for video question answering. arXiv preprint arXiv:1904.11574, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[38] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[39] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[40] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[41] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond mnns: Positional self-attention with co-attention for video question answering. In AAAI, volume 33, pages 8658–8665, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[42] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[43] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint arXiv:2107.13586, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[44] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands, too. arXiv preprint arXiv:2103.10385, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[45] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[46] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[47] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[48] Jonghwan Mun, Paul Hongsuck Seo, Ilchae Jung, and Bohyung Han. Marioqa: Answering questions by watching gameplay videos. In ICCV, pages 2867-2875, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[49] Akihiro Nakamura and Tatsuya Harada. Revisiting fine-tuning for few-shot learning. arXiv preprint arXiv:1910.00216, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[50] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[51] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Bridge to answer: Structure-aware graph interaction network for video question answering. In CVPR, pages 15526-15535, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[52] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, pages 2085–2094, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[53] Matthew E Peters, Sebastian Ruder, and Noah A Smith. To tune or not to tune? adapting pretrained representations to diverse tasks. arXiv preprint arXiv:1903.05987, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[54] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? arXiv preprint arXiv:1909.01066, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[55] Nina Poerner, Ulli Waltinger, and Hinrich Schütze. Bert is not a knowledge base (yet): Factual knowledge vs. name-based reasoning in unsupervised qa. arXiv preprint arXiv:1911.03681, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[57] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[58] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[59] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[60] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. arXiv preprint arXiv:2112.01518, 2021.3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13954" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[61] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. Advances in neural information processing systems, 28, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.134, + 0.471, + 0.176 + ], + "angle": 0, + "content": "[62] Arka Sadhu, Kan Chen, and Ram Nevatia. Video question answering with phrases via semantic roles. arXiv preprint arXiv:2104.03762, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.177, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[63] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In CVPR, pages 16877-16887, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.218, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[64] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.471, + 0.316 + ], + "angle": 0, + "content": "[65] Xiaomeng Song, Yucheng Shi, Xin Chen, and Yahong Han. Explore multi-step reasoning in video question answering. In ACM MM, pages 239-247, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.316, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[66] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, pages 7464-7473, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.371, + 0.471, + 0.427 + ], + "angle": 0, + "content": "[67] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In CVPR, pages 4631-4640, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.427, + 0.471, + 0.481 + ], + "angle": 0, + "content": "[68] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.483, + 0.471, + 0.537 + ], + "angle": 0, + "content": "[69] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. NeurlPS, 34:200-212, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.538, + 0.471, + 0.58 + ], + "angle": 0, + "content": "[70] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.58, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[71] Wenhui Wang, Hangbo Bao, Li Dong, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. arXiv preprint arXiv:2111.02358, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.636, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[72] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In CVPR, pages 9777–9786, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.679, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[73] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic video-language model pre-training for video understanding. arXiv preprint arXiv:2105.09996, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.748, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[74] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. EMNLP, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.804, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[75] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, pages 5288-5296, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[76] Li Xu, He Huang, and Jun Liu. Sutd-trafficqa: A question answering benchmark and an efficient network for video reasoning over traffic events. In CVPR, pages 9878–9888, 2021. 2, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[77] Hongyang Xue, Wenqing Chu, Zhou Zhao, and Deng Cai. A better way to attend: Attention with trees for video question answering. IEEE TIP, 27(11):5563-5574, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[78] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In ICCV, pages 1686-1697, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[79] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[80] Zekun Yang, Noa Garcia, Chenhui Chu, Mayu Otani, Yuta Nakashima, and Haruo Takemura. Bert representations for video question answering. In WACV, pages 1556-1565, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[81] Yunan Ye, Zhou Zhao, Yimeng Li, Long Chen, Jun Xiao, and Yueting Zhuang. Video question answering via attribute-augmented attention network learning. In ACM SIGIR, pages 829-832, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[82] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In CVPR, pages 16375-16387, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[83] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. NeurIPS, 34:23634-23651, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.474, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[84] Zheng-Jun Zha, Jiawei Liu, Tianhao Yang, and Yongdong Zhang. Spatiotemporal-textual co-attention network for video question answering. ACM TOMM, 15(2s):1-18, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[85] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[86] Renrui Zhang, Longtian Qiu, Wei Zhang, and Ziyao Zeng. Vt-clip: Enhancing vision-language models with visual-guided texts. arXiv preprint arXiv:2112.02399, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[87] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[88] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[89] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[90] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, pages 8746-8755, 2020, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[91] Yueting Zhuang, Dejing Xu, Xin Yan, Wenzhuo Cheng, Zhou Zhao, Shiliang Pu, and Jun Xiao. Multichannel attention refinement for video question answering. ACM TOMM, 16(1s):1-23, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.843 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13955" + } + ] +] \ No newline at end of file diff --git a/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/4a5fd117-e0c8-457c-9985-c705ba827466_origin.pdf b/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/4a5fd117-e0c8-457c-9985-c705ba827466_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c4fe441ca7aef29b6cfd337ca383f3d90c767635 --- /dev/null +++ b/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/4a5fd117-e0c8-457c-9985-c705ba827466_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:392016261741f7e7e0d7f93092ad8fc475dc0f067fe7069f0cae7d19b3a9f8db +size 1107533 diff --git a/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/full.md b/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..882459b4cb0e7ee21f3b27006c19e5808796b72a --- /dev/null +++ b/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/full.md @@ -0,0 +1,382 @@ +# Tem-adapter: Adapting Image-Text Pretraining for Video Question Answer + +Guangyi Chen $^{1,2*}$ , Xiao Liu $^{5*}$ , Guangrun Wang $^{4}$ , Kun Zhang $^{1,2}$ , Philip H.S. Torr $^{4}$ , Xiao-Ping Zhang $^{3,6}$ , Yansong Tang $^{3\dagger}$ + +$^{1}$ Carnegie Mellon University, Pittsburgh PA, USA + +$^{2}$ Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE + +3Shenzhen International Graduate School, Tsinghua University, China + +$^{4}$ University of Oxford, UK + +$^{5}$ Eindhoven University of Technology, NL + +$^{6}$ Toronto Metropolitan University, Canada + +![](images/8f247dfaba628aee3bd2515697887805abb6092559e7ba458bd4a9b56d2b9de1.jpg) +Figure 1: Illustration of the domain gaps between the pre-trained CLIP and the downstream VideoQA. CLIP is trained to align visual and textual domains, while VideoQA requires understanding the temporal dynamics and complex semantics of videos. + +# Abstract + +Video-language pre-trained models have shown remarkable success in guiding video question-answering (VideoQA) tasks. However, due to the length of video sequences, training large-scale video-based models incurs considerably higher costs than training image-based ones. This motivates us to leverage the knowledge from image-based pretraining, despite the obvious gaps between image and video domains. To bridge these gaps, in this paper, we propose Tem-Adapter, which enables the learning of temporal dynamics and complex semantics by a visual Temporal Aligner and a textual Semantic Aligner. Unlike conventional pretrained knowledge adaptation methods that only concentrate on the downstream task objective, the Temporal Aligner introduces an extra language-guided autoregressive task aimed at facilitating the learning of temporal dependencies, with the objective of predicting future states based on historical clues and language guidance that describes event progression. Besides, to reduce the semantic gap and adapt the textual representation for better event description, we introduce a Semantic Aligner that first designs a template to fuse + +question and answer pairs as event descriptions and then learns a Transformer decoder with the whole video sequence as guidance for refinement. We evaluate Tem-Adapter and different pre-train transferring methods on two VideoQA benchmarks, and the significant performance improvement demonstrates the effectiveness of our method. + +# 1. Introduction + +Video Question Answering (VideoQA) is a task that aims to answer natural language questions based on the information available in observed videos. It has attracted considerable attention recently due to its promise to develop interactive AI systems capable of communicating with the dynamic visual environment using natural language. Despite significant advancements in recent years, VideoQA remains a challenging problem that requires models to comprehensively understand and dynamically align the semantics of both the visual world and natural language. + +Recently, a variety of methods [79, 78, 82, 34, 74, 63] have demonstrated impressive results in utilizing the largescale vision-language pre-trained (VLP) model to enhance downstream VideoQA tasks. These methods pre-train models by aligning the video and language domains and then apply them to VideoQA tasks via fine-tuning or even zero-shot learning. Nevertheless, pre-training large-scale models necessitate a large number of video-text pairs, e.g., more than 10 million videos, and entails expensive computational costs. This motivates us to explore cheaper and lighter alternative pre-trained models. + +Using image-VLP models is one potential option as they also align the semantics of vision and language domains. Compared to video-based models, image-based models offer two significant advantages. Firstly, training image-based models is less expensive when considering an equal number of data pairs. Secondly, image-VLP models have made significant progress in recent years, with some of them widely available and used, such as CLIP [56]. However, as depicted in Figure 1, domain gaps persist between image-based pretraining and VideoQA tasks since image-based models are unable to learn the temporal dynamics and corresponding complex event semantics of the visual world. + +To address the challenge posed by the domain gaps, this paper proposes Tem-adapter , an adapter network that leverages the interaction between visual and textual modalities to learn temporal dynamics and complex semantics. Unlike existing methods that directly adapt visual and textual features with the objective of the downstream tasks, such as cross-modal matching in VideoQA, Tem-Adapter introduces an additional language-guided autoregressive task to facilitate learning of temporal dynamics, which enables visual representations to predict future states based on historical information and language guidance. Besides, we introduce cross-modal interactions to further reduce the semantic gap and refine textual representation. + +Our approach consists of a visual Temporal Aligner and a textual Semantic Aligner. The Temporal Aligner uses a Transformer encoder to learn the temporal relations and refine visual features. To optimize it, we build an auto-regression model by a Transformer decoder to generate the future state with historical information. We also incorporate the language information for visual refinement by adding textual embeddings as the condition memory of the Transformer decoder. This auto-regression model is supervised by a reconstruction loss with ground truth future frames to encourage the learning of temporal dependencies. In addition to the visual branch, we introduce a Semantic Aligner for better event description by reducing the semantic gap between the original texts crawled from the net and question-answer pairs in VideoQA. First, we design templates to fuse textual questions and answers to generate declarative sentences to reduce the domain gap between training and downstream + +languages. Then, a Transformer decoder is employed to refine textual embeddings by incorporating the entire video sequence as a memory condition for video-text interactions. + +We conduct experiments on two public VideoQA benchmarks including SUTD-TrafficQA [76] and MSR-VTTMC [75], and evaluate different categories of methods that transfer the pre-trained knowledge to downstream tasks, such as other adapter models [12], prompt learning [89, 21], and fully or partial finetuning. The significant improvement demonstrates the effectiveness of Tem-adapter to adapt the image-language pre-trained model for VideoQA tasks. + +# 2. Related Work + +Video Question Answering. Compared with Image-based Question Answering, Video QA is more challenging due to the complex temporal dependency, which requires models to dynamically align the visual observation and natural language. Some previous methods [67, 3, 48, 81, 28, 65, 77, 27, 26, 4, 7, 6, 8, 41, 84, 91, 10, 11, 36, 13, 17, 22, 23, 51, 62, 72] learn the models with the given manually annotated datasets. Recently, more and more methods [79, 78, 82, 34, 74, 63, 29] propose to utilize the representation ability of large-scale pre-trained vision-language models to solve the VideoQA task. For example, VQA-T [78] pre-trains the vision language model with randomly sampled video-question-answer triplets in the HowTo100M [47] dataset. BiLM [79] pre-trains frozen bidirectional language models with some adapter layers on the WebVid10M [1] dataset and achieve excellent zero-shot VideoQA performance on the downstream tasks. However, pretraining the large-scale model is expensive, which motivates us to find a cheaper alternative manner. In this paper, we propose to use existing image-language pre-trained models, such as CLIP [56], to provide the alignment between the visual and textual domains. The most related method is ATP [2], which also applies CLIP as the pre-trained model to extract frame and language embeddings, and learns an atemporal probe to select the most discriminative frames to match the video and QA pairs. Different from ATP selecting frames to transfer the video to images, we propose Tem-adapter to learn the temporal dynamics and reduce the gap between the pre-trained image domain and downstream video domain. + +Vision Language Pre-training. The goal of VLP is to help downstream tasks by pre-training the model on large-scale vision-text pairs [56, 37, 20]. It always crawls a huge amount of vision-text pairs from the web and trains the model in a self-supervised manner, such as reconstruction [40, 15, 9, 30], contrastive matching [56, 20, 18], or the combination of both two [38, 71, 25]. Different from image-VLP, video-VLP has more diverse proxy tasks, such as frame/language mask [66, 90], video-language matching [46], video/language generation [73, 45], and + +![](images/e23089f8115c8ff6203662a03a9be750cba9d115f2975625c50bc694c461adc1.jpg) +(a) Finetuning + +![](images/9deb2796159aac63423441de2909d2c1dd92490dbda499f47f2d7db1034284fc.jpg) +(b) Prompt learning + +![](images/c07ebb8cbbf5e36fad87359dfba7e353bbf371faa1d7b93cc86de6a2270b9e5d.jpg) +(c) Adapter + +![](images/41862b9e82b6f391f70318d06e9508bad8cd59f288245db8eeaedba90993c1f3.jpg) +(d) Tem-Adapter +Figure 2: Various methods for pre-trained knowledge adaptation in VideoQA. (a) Fine-tuning updates pretrained parameters with downstream training. (b) Prompt learning freezes the pre-trained model and introduces learnable textual tokens for adaptation. (c) Adapter-based methods add new adapter layers to transfer knowledge. (d) Our approach, Tem-adapter , leverages the visual-textual interactions to learn temporal dynamics to reduce the gap between image-based pre-training and video-based tasks. + +frame/language order prediction [33, 39]. Recently, VLP models have shown great potential for other downstream applications, such as zero-shot and few-shot visual recognition [56, 12, 89, 88, 86], segmentation and detection [60, 87], image generation [50, 59, 52], and action understanding [70, 68]. In this paper, we use the CLIP model as our pre-trained model. + +Pre-trained Knowledge Transfer. In light of the remarkable success of the large-scale pre-trained model, many methods have been proposed to solve the problem that how to efficiently adapt the pre-trained knowledge to the downstream tasks. Early methods finetune all parameters of the pre-trained model with the downstream objectives [14, 53]. However, with the model scale increasing, the cost of finetuning the full model is extremely expensive. Thus, an increasing number of methods focus on knowledge adaptation with light parameters. A straightforward way is to partially finetune the model parameters [49]. Besides, prompt learning methods are proposed recently to reformulate the downstream tasks as the pre-trained tasks and freeze all model parameters [43, 54, 58, 55]. The previous prompt-based methods [54, 55] design the template of prompt using the prior of natural human language. Then, some methods [24, 64] + +propose to automatically optimize the prompt template, and [42, 69, 44, 89, 5] further learn the prompts in the continuous embedding space where the discrete word constraint is relaxed. In addition, some methods freeze the model and train some extra adapter layers and also achieve the supersizing performance [12, 85, 1]. As shown in Figure 2, we illustrate the difference between Tem-adapter and other categories of methods transferring the image-based VLP model to the VideoQA task. Compared with other pre-train adaptation methods, our Tem-adapter proposes an autoregression task to reduce the temporal gap (tem-align) and introduce the cross-domain interactions (sem-align) to reduce the semantic gap between the pre-training and downstream tasks. + +# 3. Approach + +In this section, we first briefly introduce the VideoQA task and then present our approach, Tem-adapter , which introduces the cross-modal interactions to learn temporal dynamics and complex semantics. Finally, we give the details of our implementation. + +# 3.1. Problem Definition + +The goal of VideoQA is to find an answer $\hat{\mathbf{a}}$ from the answer pool, given the conditions of question $\mathbf{q}$ and observed video data $\mathbf{v}$ as below: + +$$ +\hat {\mathbf {a}} = \underset {\mathbf {a} \in \mathcal {A}} {\arg \max } \mathcal {S} (\mathbf {a} | \mathbf {v}, \mathbf {q}), \tag {1} +$$ + +where $S$ is a match function to generate the score of each answer candidate given the question and video content. + +# 3.2. Tem-Adapter + +Here, we propose Tem-adapter to learn temporal dynamics and complex semantics by the interactions of visual and textual domains to adapt pre-trained Image-VLP models to downstream VideoQA tasks. Figure 3 shows an overview of Tem-Adapter's framework. In particular, Tem-adapter consists of two components, including a visual Temporal Aligner and a textual Semantic Aligner. Details for each component are as follows. + +Temporal Aligner. The goal of the Temporal Aligner is to reduce the gap between image-based pre-training and video-based downstream tasks. To achieve this goal, we introduce an auxiliary language-guided autoregressive task to facilitate the learning of temporal dynamics. + +Specifically, given a video $\mathbf{v}$ , we apply the frozen CLIP [56] image encoder $f$ to extract the feature of each frame as $f(\mathbf{v}) = \{f(\mathbf{v}_t)|_{t = 1}^T\} \in \mathcal{R}^{T\times C}$ , where $T$ is the frame number of videos and $C$ is the feature dimension of CLIP. Then we refine these features with our Temporal Aligner model which is built by a transformer encoder $f_{e}$ with parameters $\phi$ to learn the temporal relations of frames. + +![](images/221a9583e07cb8316e2249cea0dbeafaa5408829cc2a561d13a9d45d1e13d6a9.jpg) +Figure 3: The overall framework of Tem-adapter. Our Tem-adapter consists of two modules, Semantic Aligner (orange) and Temporal Aligner (blue). Semantic Aligner takes the text embedding of template-based QA pair as the input, and applies a Transformer decoder to incorporate cross-modal interactions. The Temporal Aligner introduces a language-guided autoregressive task to learn temporal dynamics. The training is guided by classification and reconstruction losses. + +The refined features can be represented as: + +$$ +\mathbf {f} = f _ {e} (f (\mathbf {v}); \phi), \tag {2} +$$ + +which are with the same dimension $\mathcal{R}^{T\times C}$ + +To guide the Temporal Aligner to learn temporal dynamics, we further build an autoregressive sequential model, following GPT [57], which generates the feature of the current frame with previous frames and guidance from textual information that describes event progression. Mathematically, it can be formulated as: + +$$ +\hat {f} (\mathbf {v} _ {i}) = f _ {d} \left(f \left(\mathbf {v} _ {1: i - 1}\right) | \mathbf {f}, \mathbf {g}; \psi\right), \tag {3} +$$ + +where $f(\mathbf{v}_{1:i-1})$ is the input of the transformer decoder, which is achieved by applying an attention mask to focus on the history clues. $\mathbf{f}, \mathbf{g}$ are used as the guidance information, where $\mathbf{f}$ denotes latent visual embeddings obtained by Temporal Aligner and $\mathbf{g}$ (see equation 5) is the textual embeddings describing temporal event. $\mathbf{f}$ and $\mathbf{g}$ are first fused by adding and then serve as the memory key and value of the Transformer decoder. We apply a reconstruction loss to encourage the learning of temporal dependencies which is formulated as the distance between the results of the auto-regression model $\hat{f}(\mathbf{v}_i)$ and the ground-truth $f(\mathbf{v}_i)$ : + +$$ +L _ {M S E} = \sum_ {i = 1} ^ {n} | | f (\mathbf {v} _ {i}) - \hat {f} (\mathbf {v} _ {i}) | | _ {2} ^ {2}. \tag {4} +$$ + +Semantic Aligner. In addition to the visual branch, we introduce a Semantic Aligner to adapt the textual representation for better event description and reduce the semantic gaps. We first design templates to align the question-answer pairs in VideoQA to declarative sentences, since the training data of CLIP is crawled from the web which always describes the image in a declarative manner. Then, we apply a Transformer decoder with the video sequence as the memory for video-text interactions. + +Given a question $\mathbf{q}$ and a corresponding answer $\mathbf{a}$ , we first apply the grammar parsing on the question $\mathbf{q}$ to obtain the subject-verb-object structure, and then we modify the order of the sentence to obtain a declarative sentence and leave the position of the interrogative pronouns as a mask '[']. Finally, we fill in the default '['] with the answer to update the QA pair as a refined event description. For instance, the question is 'Which might be the reason for this accident?', and the answer is 'Improper lane change'. We first obtain a declarative sentence with the default as '[ ] is the reason for this accident', and then we can fill in the answer to get 'Speeding vehicle is the reason for this accident.' In addition, for the general question, we first change it to a declarative sentence and then convert the sentence to negative form if the answer is No. + +With the refined sentence $\mathbf{q}\mathbf{a}$ as the input, the frozen text encoder $g$ of the CLIP model outputs the textual feature as $g(\mathbf{q}\mathbf{a}) \in \mathcal{R}^C$ . To further refine the textual feature using the video content, we apply a transformer decoder with the visual feature sequence as the memory key and value and add a residual link to avoid overfitting. Formally, it can be formulated as follows: + +$$ +\mathbf {g} = g (\mathbf {q a}) + \lambda g _ {d} (g (\mathbf {q a}), f (\mathbf {v}); \theta), \tag {5} +$$ + +where $\mathbf{g}$ denotes the learned textual embedding, $\lambda$ is a parameter of the residual link, and $\theta$ is the parameters of the transformer decoder $g_{d}$ . Both $\lambda$ and $\theta$ can be learned automatically in the training process. With the cross-attention module in the transformer decoder, we encourage the model to find key information from the video to refine the textual embedding for better event description. + +Cross-modal Matching. Given the refined visual and textual embeddings $\mathbf{f}$ and $\mathbf{g}$ , we instantiate the score function in equation 1 by cosine similarity as: + +$$ +S (\mathbf {a} | \mathbf {v}, \mathbf {q}) = \operatorname {s o f t m a x} (\bar {\mathbf {f}} ^ {T} \mathbf {g}), \tag {6} +$$ + +where $\bar{\mathbf{f}}$ is video representation using the average pool on all frames, $\mathbf{g}$ is the language embedding with given answer question pair $(\mathbf{a},\mathbf{q})$ . Then, we can obtain the answer to the question with the highest matching score. + +Then we apply a standard multi-choice classification Hinge loss [19] integrated with the auto-regression loss for training. The overall loss function can be described as: + +$$ +L = L _ {H i n g e} (S (\mathbf {a} | \mathbf {v}, \mathbf {q}), y) + \gamma L _ {M S E}, \tag {7} +$$ + +where $\gamma$ is a ratio to balance two loss functions, and $y$ is the ground truth label of the answers. Please note that our Tem-adapter consists of 3 learnable modules including the textual decoder $\theta$ , the visual encoder $\phi$ , and the visual decoder $\psi$ . The Hinge loss is applied to optimize $\theta$ and $\phi$ , while the auto-regression loss $L_{MSE}$ is used to optimize all three modules. In the inference, only textual decoder $\theta$ and visual encoder $\phi$ are used. + +# 3.3. Implementation Details + +In this subsection, we provide more implementation details of our approach for better reproduction. + +Network Architecture. Figure 3 shows the framework of Tem-adapter. In our experiments, we use the frozen model CLIP (ViT-B32) as the text and image encoders. The Semantic Aligner is a one-layer Transformer decoder with 16 heads. There is a linear layer before the Transformer decoder which maps text embeddings and image embeddings to a lower dimension (128). Another linear projection layer is added after the Transformer decoder to map the reduced dimension back to 512. The Temporal Aligner is a light transformer encoder-decoder architecture. Both the encoder and decoder contain only one layer. We set the number of heads to 16 in the experiments. + +Inference Details. In the inference (given 1 question, 1 video, and $k$ answers), we first apply the template to fuse the question and $k$ answers to obtain $k$ QA pairs (e.g. $k = 4$ ). Then we feed the QA pairs and video into CLIP textual and visual encoders to obtain the initial features. After that, we apply the learned semantic and temporal aligners to refine these features. Please note that the temporal auto-regression module is not used in the inference. Finally, we obtain 1 video feature with $k$ QA features and apply the cosine similarity as well as softmax to select the answer. + +Hyper-parameters. We follow the same video frames sampling method with prior methods [76]. We uniformly select 8 frames over a long video sequence and then pick consecutive 16 frames which are centered on those 8 frames. Overall 128 frames are selected for each video. We also tried another sampling method: uniformly extracting 128 or 256 frames on a video sequence, and the final results do not support the sampling method very well. In the loss function, we set $\gamma$ as 100 to encourage the model to learn the temporal dynamics. And the learnable parameter $\lambda$ in the Semantic + +Aligner is set to be the same dimension (512) as the output embeddings of the CLIP model. + +Training Details. We use the pre-trained CLIP model to extract text embeddings and image embeddings. The Adam [31] optimizer is utilized to optimize the training process. We set the initial learning rate to be $1e - 4$ with a batch size of 128. And we let the learning rate decay with a factor of 0.5 for every 10 epochs. Our model is trained for 50 epochs on a single Tesla V100 GPU and the training takes 8 hours to finish. Our model is implemented in PyTorch 1.8.0. + +# 4. Experiments + +In this section, we first benchmark our Tem-adapter and different categories of methods that transfer the pretrained model into downstream tasks. Then we provide ablation studies to investigate the effect of each component. + +# 4.1. Datasets & Experimental Protocols + +SUTD-TrafficQA. SUTD-TrafficQA focuses on the question-answering task in the traffic situation, which requires the model to understand the traffic event and underlying causal relation. It consists of more than 10k videos with different traffic events and provides over 62,535 human-annotated QA pairs. Among them, 56,460 QA pairs are used for training and the rest 6,075 QA pairs are used for testing. This dataset highlights the cognitive capability of video event understanding models in complex traffic scenarios. It provides 6 challenging traffic-related reasoning tasks including "Basic understanding", "Event forecasting", "Reverse reasoning", "Counterfactual inference", "Introspection", "Attribution". All tasks are formulated as multiple-choice without limiting the number of candidate answers. + +MSR-VTT-MC. MSR-VTT is collected and released mainly for the text-to-video retrieval task. Following [34] and [2], we use the multi-choice test setting to benchmark the VideoQA methods. The MSR-VTT-MC test set consists of 2990 QA pairs. Each video corresponds to five answers, with the original caption as the correct answer. We use the standard protocol with prior works in our experiments, where 7010 train+val video sequences and 140,200 QA pairs are utilized for training. Each video only contains one caption. We assign this caption as the correct answer for our question-answering task. To formulate four negative answers during training, we randomly select four captions from other captions that do not belong to the current video. + +# 4.2. Baseline Methods + +To investigate different categories of methods that transfer the pre-trained model into downstream tasks, such as finetuning, prompt learning, and adapter. We compare Tem-adapter with the other 10 baseline methods including Unsupervised CLIP [56], Unsupervised CLIP [56] + +Table 1: VideoQA accuracies of baseline methods on both SUTD-TrafficQA and MSR-VTT-MC. B: "Basic understanding", F: "Forecasting task", R: "Reverse Reasoning", C: "Counterfactual inference", I: "Introspection", A: "Attribution". (C) and (A) denote strategies for adding prompts. (C) and $(\mathbf{C}^{*})$ denote training prompts with/without adapter heads. + +
MethodsSUTD-TrafficQAMSR-VTT-MC
BFRCIAAvgAcc
Unsupervised CLIP [56]25.620.134.030.822.828.826.574.1
CLIP [56] + Template31.836.029.971.822.133.432.374.1
Totally finetuning39.835.146.645.637.240.540.389.0
Partially finetuning41.637.844.650.033.141.741.788.5
LORA [16]38.738.736.737.934.538.138.385.4
CLIP-Adapter [12]35.832.035.442.333.132.134.883.4
Multi-layer Adapter [12]30.526.626.538.528.325.829.190.7
Prompt learning (C) [89]42.432.445.255.540.743.642.989.0
Prompt learning (C*) [89]40.333.241.046.534.938.439.790.8
Prompt learning (A) [21]41.731.540.148.433.141.441.188.0
Tem-adapter46.036.544.655.034.547.746.094.3
+ +Table 2: Ablation studies on the SUTD-TrafficQA and MSR-VTT-MC datasets. Tp, SA, TA, and Ar denote Template, Semantic Aligner, Temporal Aligner, and Auto-regression, respectively. No ablation studies (“\”) concentrate on Tp, as it is not compatible with the MSR-VTT-MC dataset. + +
MethodsSUTD-TrafficQAMSR-VTT-MC
AvgAcc
CLIP-Adapter [12]34.883.4
w/o Tp43.9\
w/o SA44.291.8
w/o TA43.691.6
w/o Ar44.992.6
w/o TA and SA32.374.1
Tem-adapter46.094.3
+ ++ Language template, Totally fine-tuning, Partially finetuning, LORA [16], CLIP-Adapter [12], Multi-layer CLIP-Adapter [12], Prompt learning (change words) [89] (with/without using adapter heads), and Prompt learning (add words) [21]. The detailed descriptions are shown in the supplementary materials. + +# 4.3. Benchmarking against Baseline Methods + +In this subsection, we compare and discuss our approach's performance against the above-mentioned baselines on both SUTD-TrafficQA and MSR-VTT-MC datasets. The results of all methods on both datasets are summarized in Table 1. We can draw conclusions as follows from the results. + +CLIP can obtain good zero-shot results. With our language template, the clip model can achieve $32.3\%$ accuracy on the SUTD-TrafficQA dataset, which is almost close to the previous state-of-the-art performance $(37.1\%)$ . It shows + +the strong representation ability of the CLIP model. + +More parameters $\neq$ Better performance. We have two groups of experiments that focus on the number of parameters, including totally/partially finetuning and single-layer/multi-layer Adapters. We obtain contradictory conclusions from different datasets. On SUTD-TrafficQA, the model with fewer parameters works better, while on MSR-VTT-MC, the model with more parameters has better performance. It is not surprising since the scale of training data is different. MSR-VTT-MC has more training data and can support training more parameters. We also compare the learnable parameters of adapter-based methods in Table 5. Tem-adapter achieve better performance with fewer parameters than Multi-layer Adapter. Compared with the prompt-learning-based method, our method needs more parameters, but its inference is not time-consuming. + +Different categories of methods are comparable. Balancing the performance on both datasets, different categories of methods that transfer the image-based VLP model to VideoQA don't have clear performance differences. Adapter-based methods obtain poor results on SUTD-TrafficQA, but multi-layer Adapter works well on MSR-VTT-MC. Besides, by comparing the prompt learning methods with/without adapter heads, we found that the prompt and adapter can be joint learning to obtain better performance. + +Cross-modal interaction and temporal dynamics matter. Compared with CLIP-Adapter (both single-layer and multi-layer ones), our Tem-adapter achieves more than $10\%$ improvement on SUTD-TrafficQA. The key difference between Tem-adapter and other adapter-based methods is that we introduce temporal alignment to reduce the temporal gap and apply the cross-modal interaction to refine the representation and reduce the semantic gap. Tem-adapter outperforms other methods by a large mar + +Table 3: Comparison with the state-of-the-art methods on the SUTD-TrafficQA dataset. + +
MethodsAccuracyYear
Human95.4-
VIS+LSTM [61]29.92015
TVQA [35]35.22018
HCRN [32]36.52020
BERT-VQA [80]33.72020
Eclipse [76]37.12021
CLIP [56]32.32021
CLIP+HCRN[56, 32]41.92021
ATP [2]35.62022
Tem-adapter46.0
+ +gin on both two datasets. For example, it obtains $3.1\%$ improvement than Prompt learning (C) on SUTD-TrafficQA and $3.6\%$ than Multi-layer Adapter on MSR-VTT-MC. + +# 4.4. Comparison against State-of-the-Art Methods + +SUTD-TrafficQA. On SUTD-TrafficQA, we compare our proposed method with state-of-the-art methods, including [32, 76, 61, 35, 80]. The results are summarized in Table 3. We can observe that our Tem-adapter outperforms other methods by a large margin. For example, compared with HCRN [32], Tem-adapter achieves almost $10\%$ improvement. By using the CLIP model to extract visual features, HCRN+CLIP can obtain $41.9\%$ accuracy, which is better than using the original HCRN. Besides, we show that the performance improvement is not only from the better representation ability of the CLIP model since our Tem-adapter further improve HCRN+CLIP with $+4.1\%$ accuracy. In addition, we also significantly outperform the ATP method, which also uses the CLIP model. This significant performance improvement demonstrates the effectiveness of Tem-adapter. + +MSR-VTT-MC. On MSR-VTT-MC, we compare Tem-adapter with some recent state-of-the-art methods, including ActBERT[90], ClipBERT[34], MERLOT [83], VideoCLIP [66], CLIP [56], and ATP [2]. Results are shown in Table 4. First, we can find that the CLIP model benefits the performance on the MSR-VTT-MC dataset. All methods using the CLIP model, such as CLIP [56], ATP [2], and Tem-adapter achieve good performance. ATP [2] learns an atemporal probe to select a frame from a video sequence to obtain the visual embedding and achieve good performance on MSR-VTT-MC. Compared with it, Tem-adapter can further obtain $94.3\%$ accuracy and $1.1\%$ relative improvement. + +# 4.5. Ablation Studies + +We provide ablation studies to investigate the effect of each component. Four main modules in our Tem-adapter + +Table 4: Comparison with the state-of-the-art methods on the MSR-VTT-MC dataset. + +
MethodsAccuracyYear
VideoCLIP [66]92.12019
ActBERT[90]85.72020
ClipBERT[34]88.22021
MERLOT [83]90.92021
CLIP [56]74.12021
ATP [2]93.22022
Tem-adapter94.3
+ +are evaluated, including Temporal Aligner, Semantic Aligner, language template, and auto-regression. + +As shown in Table 2, we provide the performance of different variants of Tem-adapter and a baseline adapter method on each task in the SUTD-TrafficQA and MSR-VTTMC datasets. The baseline method is CLIP-adapter, which also uses the language template and learns a linear adapter layer without cross-domain interaction. We observe that all variants of Tem-adapter can achieve significant improvement over the baseline method, which shows that the interaction across visual and textual is very important for video analysis. Besides, compared with Tem-adapter, we can observe a clear performance drop when we remove each component, which demonstrates all of them are indispensable. When both temporal and semantic aligners are removed, the performance drops dramatically. + +In addition, we also conducted an experiment to show whether the Temporal Aligner learns the temporal dynamic information. Table 6 shows the reconstruction quality of our methods with/without language guidance $(g)$ and with/without Temporal Aligner $(f)$ . We can observe that both the language guidance and Temporal Aligner assist in the learning of temporal dynamics. + +Furthermore, we also provide some parameter analysis experiments about our network design. And evaluate Tem-adapter on different image-text pre-train models. Please kindly refer to the supplementary materials for the detailed results and analysis. + +# 4.6. Visualization + +In this subsection, we provide some qualitative results to analyze our approach. As shown in Figure 4, we show some visualization examples of VdideoQA on the SUTD-TrafficQA dataset. In the top row, we plot two positive examples. We can observe that our Tem-adapter learns the temporal dynamics with the cross-domain interactions to obtain a better understanding of video events. Besides, we also provide some failure cases to help explore the boundary of our approach. We find that Tem-adapter can not work well for situations that need complex reasoning, especially + +Q: What could possibly cause this accident? + +![](images/03e235e26b72a5879a70ea52c006d335d3753213ce85a1a893e4be646de2423c.jpg) + +Ours: Sudden or extreme movement by a vehicle. + +CLIP-Adapter: Unfavorable weather. + +Prompt learning: Violation of traffic rules. + +Finetuning: Violation of traffic rules. + +Q: Could any vehicles steer away to prevent the accident? + +![](images/1cd05c0e3725a624008bedcf98cbe31a252065b747e5c8e14ed03eb864520497.jpg) +Figure 4: Visualization of some examples of the VideoQA task from the SUTD-TrafficQA dataset. The top row shows some positive examples, where Tem-adapter learns temporal dynamics to understand the traffic event and give the correct answers. The bottom row shows some failed cases. Despite better matching between visual and textual domains, our approach fails on complex problems which require reasoning ability. + +Ours: No, there was no space. X + +GT: Yes, there was enough space. + +Q: Which factors contribute to the accident? + +![](images/872773b7ceeb0ea45defc8048668de003dd580de6a1153bd856e6e2f932bed1a.jpg) + +Ours: Not paying attention to their surroundings. + +CLIP-Adapter: Bad road surfaces + +Prompt learning: Bad road surfaces. + +Finetuning: Speeding vehicles. + +Q: Could the accident be prevented if a convex mirror is installed at the turning point? + +![](images/51fbfa3d51d96754fa0e286797bf9782353f3153f2c66544d1103bc396a8929e.jpg) + +Ours: Yes, there would be no obstructed views. + +GT: No, that was not the main cause of the accident. + +Table 5: The comparisons of our method and other baseline methods on model parameters (M), inference time (ms/video), and accuracy in the SUTD-TrafficQA dataset. + +
MetricsLORA [16]Prompt learning (C) [89]Prompt learning (C*) [89]CLIP-Adapter [12]Multi-layer Adapter [12]Ours
Parameters1.443.550.771.054.853.62
Inference time5.375.335.175.135.145.32
Accuracy38.342.939.734.829.146.0
+ +Table 6: The comparison of reconstruction ability. We use peak signal-to-noise ratio (PSNR) as the metric. + +
Tem-adapterw/o gw/o f and g
PSNR29.128.727.4
+ +for the underlying cause-effect relations. We leave it as a future work to enhance the reasoning ability of our model. Besides, we also provide more examples of our language templates in the supplementary materials. + +# 5. Conclusion + +In this paper, we proposed Tem-adapter to adapt the pre-trained image-language model to the downstream VideoQA task. We jointly learn the Temporal Aligner and Semantic Aligner through cross-modal interactions. Tem-adapter introduces a language-guided autoregressive task to guide the learning of temporal dependency and thus reduce the temporal gap between image-based pre-train and video-based QA tasks. Besides, we utilize a rule-based template and video-based interaction to refine textual representation and reduce semantic gaps. We validate the effectiveness of our approach on four VideoQA benchmarks. + +Broader Impacts and Limitations. Understanding how the big model adaptation works on VideoQA is of great help in developing general foundation models and interactive AI. However, our work also has certain limitations. For example, since extracting video features online takes up too many computing resources, we extract their features offline and fix them. We leave it as future work to learn the adapter within the backbone network. + +# Acknowledgment + +This work was supported in part by the National Natural Science Foundation of China under Grant 62206153, Young Elite Scientists Sponsorship Program by CAST (No. 2023QNRC002), by the Tsinghua Shenzhen International Graduate School-Shenzhen Pengrui Endowed Professorship Scheme of Shenzhen Pengrui Foundation, by the UKRI grant: Turing AI Fellowship EP/W002981/1, by the National Institutes of Health (NIH) under Contract R01HL159805, by the NSF-Convergence Accelerator Track-D award #2134901, by a grant from Apple Inc., a grant from KDDI Research Inc, and generous gifts from Salesforce Inc., Microsoft Research, and Amazon Research. We would also like to thank the Royal Academy of Engineering and FiveAI. + +# References + +[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 2, 3 +[2] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the" video" in video-language understanding. In $CVPR$ , pages 2917-2927, 2022. 2, 5, 7 +[3] Santiago Castro, Mahmoud Azab, Jonathan Stroud, Cristina Noujaim, Ruoyao Wang, Jia Deng, and Rada Mihalcea. Lifeqa: A real-life dataset for video question answering. In LREC, pages 4352-4358, 2020. 2 +[4] Aman Chadha, Gurneet Arora, and Navpreet Kaloty. iperceive: Applying common-sense reasoning to multi-modal dense video captioning and video question answering. arXiv preprint arXiv:2011.07735, 2020. 2 +[5] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. In ICLR, 2023. 3 +[6] Seongho Choi, Kyoung-Woon On, Yu-Jung Heo, Ahjeong Seo, Youwon Jang, Minsu Lee, and Byoung-Tak Zhang. Dramaqa: Character-centered video story understanding with hierarchical qa. In AAAI, volume 35, pages 1166-1174, 2021. 2 +[7] Anthony Colas, Seokhwan Kim, Franck Dernoncourt, Siddhesh Gupte, Daisy Zhe Wang, and Doo Soon Kim. Tutorial: Question answering dataset for tutorial videos. arXiv preprint arXiv:1912.01046, 2019. 2 +[8] Long Hoang Dang, Thao Minh Le, Vuong Le, and Truyen Tran. Object-centric representation learning for video question answering. In IJCNN, pages 1-8, 2021. 2 +[9] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Zicheng Liu, Michael Zeng, et al. An empirical study of training end-to-end vision-and-language transformers. arXiv preprint arXiv:2111.02387, 2021. 2 +[10] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In CVPR, pages 1999-2007, 2019. 2 +[11] Jiyang Gao, Runzhou Ge, Kan Chen, and Ram Nevatia. Motion-appearance co-memory networks for video question answering. In CVPR, pages 6576–6585, 2018. 2 +[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2, 3, 6, 8 +[13] Noa Garcia, Mayu Otani, Chenhui Chu, and Yuta Nakashima. Knowit vqa: Answering knowledge-based questions about videos. In AAAI, volume 34, pages 10826-10834, 2020. 2 +[14] Ross Girshick. Fast r-cnn. In ICCV, pages 1440-1448, 2015. 3 +[15] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. VIn bert: A recurrent vision-and-language bert for navigation. In CVPR, pages 1643–1653, 2021. 2 + +[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6, 8 +[17] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware graph convolutional networks for video question answering. In AAAI, volume 34, pages 11021-11028, 2020. 2 +[18] Aashi Jain, Mandy Guo, Krishna Srinivasan, Ting Chen, Sneha Kudugunta, Chao Jia, Yinfei Yang, and Jason Baldridge. Mural: multimodal, multitask retrieval across languages. arXiv preprint arXiv:2109.05125, 2021. 2 +[19] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In CVPR, pages 2758–2766, 2017. 5 +[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 2 +[21] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022. 2, 6 +[22] Jianwen Jiang, Ziqiang Chen, Haojie Lin, Xibin Zhao, and Yue Gao. Divide and conquer: Question-guided spatiotemporal contextual attention for video question answering. In AAAI, volume 34, pages 11101-11108, 2020. 2 +[23] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In AAAI, volume 34, pages 11109-11116, 2020. 2 +[24] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? TACL, 8:423-438, 2020. 3 +[25] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1780-1790, 2021. 2 +[26] Hyoungun Kim, Zineng Tang, and Mohit Bansal. Dense-ecaption matching and frame-selection gating for temporal localization in videoqa. arXiv preprint arXiv:2005.06409, 2020.2 +[27] Junyeong Kim, Minuk Ma, Trung Pham, Kyungsu Kim, and Chang D Yoo. Modality shifting attention network for multimodal video question answering. In CVPR, pages 10106-10115, 2020. 2 +[28] Kyung-Min Kim, Min-Oh Heo, Seong-Ho Choi, and Byoung-Tak Zhang. Deepstory: Video story qa by deep embedded memory networks. arXiv preprint arXiv:1707.00836, 2017. 2 +[29] Seonhoon Kim, Seohyeong Jeong, Eunbyul Kim, Inho Kang, and Nojun Kwak. Self-supervised pre-training and contrastive representation learning for multiple-choice video qa. In AAAI, volume 35, pages 13171-13179, 2021. 2 +[30] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, pages 5583-5594, 2021. 2 + +[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[32] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In CVPR, pages 9972-9981, 2020. 7 +[33] Chenyi Lei, Shixian Luo, Yong Liu, Wanggui He, Jiamang Wang, Guoxin Wang, Haihong Tang, Chunyan Miao, and Houqiang Li. Understanding chinese video and language via contrastive multimodal pre-training. In ACM MM, pages 2567-2576, 2021. 3 +[34] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, pages 7331-7341, 2021. 2, 5, 7 +[35] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. arXiv preprint arXiv:1809.01696, 2018. 7 +[36] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvqa+: Spatio-temporal grounding for video question answering. arXiv preprint arXiv:1904.11574, 2019. 2 +[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 2 +[38] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 34, 2021. 2 +[39] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 3 +[40] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2 +[41] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond mnns: Positional self-attention with co-attention for video question answering. In AAAI, volume 33, pages 8658–8665, 2019. 2 +[42] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3 +[43] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint arXiv:2107.13586, 2021. 3 +[44] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands, too. arXiv preprint arXiv:2103.10385, 2021. 3 +[45] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2 + +[46] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 2 +[47] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 2 +[48] Jonghwan Mun, Paul Hongsuck Seo, Ilchae Jung, and Bohyung Han. Marioqa: Answering questions by watching gameplay videos. In ICCV, pages 2867-2875, 2017. 2 +[49] Akihiro Nakamura and Tatsuya Harada. Revisiting fine-tuning for few-shot learning. arXiv preprint arXiv:1910.00216, 2019.3 +[50] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3 +[51] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Bridge to answer: Structure-aware graph interaction network for video question answering. In CVPR, pages 15526-15535, 2021. 2 +[52] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, pages 2085–2094, 2021. 3 +[53] Matthew E Peters, Sebastian Ruder, and Noah A Smith. To tune or not to tune? adapting pretrained representations to diverse tasks. arXiv preprint arXiv:1903.05987, 2019. 3 +[54] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? arXiv preprint arXiv:1909.01066, 2019. 3 +[55] Nina Poerner, Ulli Waltinger, and Hinrich Schütze. Bert is not a knowledge base (yet): Factual knowledge vs. name-based reasoning in unsupervised qa. arXiv preprint arXiv:1911.03681, 2019. 3 +[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6, 7 +[57] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 4 +[58] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3 +[59] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.3 +[60] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. arXiv preprint arXiv:2112.01518, 2021.3 + +[61] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. Advances in neural information processing systems, 28, 2015. 7 +[62] Arka Sadhu, Kan Chen, and Ram Nevatia. Video question answering with phrases via semantic roles. arXiv preprint arXiv:2104.03762, 2021. 2 +[63] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In CVPR, pages 16877-16887, 2021. 2 +[64] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3 +[65] Xiaomeng Song, Yucheng Shi, Xin Chen, and Yahong Han. Explore multi-step reasoning in video question answering. In ACM MM, pages 239-247, 2018. 2 +[66] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, pages 7464-7473, 2019. 2, 7 +[67] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In CVPR, pages 4631-4640, 2016. 2 +[68] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022.3 +[69] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. NeurlPS, 34:200-212, 2021. 3 +[70] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021.3 +[71] Wenhui Wang, Hangbo Bao, Li Dong, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. arXiv preprint arXiv:2111.02358, 2021. 2 +[72] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In CVPR, pages 9777–9786, 2021. 2 +[73] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic video-language model pre-training for video understanding. arXiv preprint arXiv:2105.09996, 2021. 2 +[74] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. EMNLP, 2021. 2 +[75] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, pages 5288-5296, 2016. 2 +[76] Li Xu, He Huang, and Jun Liu. Sutd-trafficqa: A question answering benchmark and an efficient network for video reasoning over traffic events. In CVPR, pages 9878–9888, 2021. 2, 5, 7 + +[77] Hongyang Xue, Wenqing Chu, Zhou Zhao, and Deng Cai. A better way to attend: Attention with trees for video question answering. IEEE TIP, 27(11):5563-5574, 2018. 2 +[78] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In ICCV, pages 1686-1697, 2021. 2 +[79] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. NeurIPS, 2022. 2 +[80] Zekun Yang, Noa Garcia, Chenhui Chu, Mayu Otani, Yuta Nakashima, and Haruo Takemura. Bert representations for video question answering. In WACV, pages 1556-1565, 2020. 7 +[81] Yunan Ye, Zhou Zhao, Yimeng Li, Long Chen, Jun Xiao, and Yueting Zhuang. Video question answering via attribute-augmented attention network learning. In ACM SIGIR, pages 829-832, 2017. 2 +[82] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In CVPR, pages 16375-16387, 2022. 2 +[83] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. NeurIPS, 34:23634-23651, 2021. 7 +[84] Zheng-Jun Zha, Jiawei Liu, Tianhao Yang, and Yongdong Zhang. Spatiotemporal-textual co-attention network for video question answering. ACM TOMM, 15(2s):1-18, 2019. 2 +[85] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. In CVPR, 2022. 3 +[86] Renrui Zhang, Longtian Qiu, Wei Zhang, and Ziyao Zeng. Vt-clip: Enhancing vision-language models with visual-guided texts. arXiv preprint arXiv:2112.02399, 2021. 3 +[87] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021. 3 +[88] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 3 +[89] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 2, 3, 6, 8 +[90] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, pages 8746-8755, 2020, 2, 7 +[91] Yueting Zhuang, Dejing Xu, Xin Yan, Wenzhuo Cheng, Zhou Zhao, Shiliang Pu, and Jun Xiao. Multichannel attention refinement for video question answering. ACM TOMM, 16(1s):1-23, 2020. 2 \ No newline at end of file diff --git a/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/images.zip b/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7d8828c120a0b56309cb3d49e221f367d7b63457 --- /dev/null +++ b/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92574b6ee82873fe1452193c74c172b044a9a92bb9d2219cef2f36bd74003e23 +size 508525 diff --git a/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/layout.json b/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d3d6d1af786f11e0e4c0881211625f0178d581cb --- /dev/null +++ b/2023/Tem-Adapter_ Adapting Image-Text Pretraining for Video Question Answer/layout.json @@ -0,0 +1,10137 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 531, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 531, + 122 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 531, + 122 + ], + "type": "text", + "content": "Tem-adapter: Adapting Image-Text Pretraining for Video Question Answer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "spans": [ + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "text", + "content": "Guangyi Chen" + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "text", + "content": ", Xiao Liu" + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "inline_equation", + "content": "^{5*}" + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "text", + "content": ", Guangrun Wang" + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "text", + "content": ", Kun Zhang" + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "text", + "content": ", Philip H.S. Torr" + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "text", + "content": ", Xiao-Ping Zhang" + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "inline_equation", + "content": "^{3,6}" + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "text", + "content": ", Yansong Tang" + }, + { + "bbox": [ + 99, + 141, + 494, + 171 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 176, + 171, + 418, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 171, + 418, + 185 + ], + "spans": [ + { + "bbox": [ + 176, + 171, + 418, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 176, + 171, + 418, + 185 + ], + "type": "text", + "content": "Carnegie Mellon University, Pittsburgh PA, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 185, + 481, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 185, + 481, + 199 + ], + "spans": [ + { + "bbox": [ + 114, + 185, + 481, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 114, + 185, + 481, + 199 + ], + "type": "text", + "content": "Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 199, + 466, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 199, + 466, + 213 + ], + "spans": [ + { + "bbox": [ + 127, + 199, + 466, + 213 + ], + "type": "text", + "content": "3Shenzhen International Graduate School, Tsinghua University, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 231, + 213, + 362, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 213, + 362, + 227 + ], + "spans": [ + { + "bbox": [ + 231, + 213, + 362, + 227 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 231, + 213, + 362, + 227 + ], + "type": "text", + "content": "University of Oxford, UK" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 195, + 227, + 399, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 227, + 399, + 241 + ], + "spans": [ + { + "bbox": [ + 195, + 227, + 399, + 241 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 195, + 227, + 399, + 241 + ], + "type": "text", + "content": "Eindhoven University of Technology, NL" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 194, + 241, + 399, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 241, + 399, + 255 + ], + "spans": [ + { + "bbox": [ + 194, + 241, + 399, + 255 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 194, + 241, + 399, + 255 + ], + "type": "text", + "content": "Toronto Metropolitan University, Canada" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 50, + 281, + 544, + 401 + ], + "blocks": [ + { + "bbox": [ + 50, + 281, + 544, + 401 + ], + "lines": [ + { + "bbox": [ + 50, + 281, + 544, + 401 + ], + "spans": [ + { + "bbox": [ + 50, + 281, + 544, + 401 + ], + "type": "image", + "image_path": "8f247dfaba628aee3bd2515697887805abb6092559e7ba458bd4a9b56d2b9de1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 407, + 546, + 433 + ], + "lines": [ + { + "bbox": [ + 46, + 407, + 546, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 546, + 433 + ], + "type": "text", + "content": "Figure 1: Illustration of the domain gaps between the pre-trained CLIP and the downstream VideoQA. CLIP is trained to align visual and textual domains, while VideoQA requires understanding the temporal dynamics and complex semantics of videos." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 442, + 192, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 442, + 192, + 455 + ], + "spans": [ + { + "bbox": [ + 143, + 442, + 192, + 455 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "content": "Video-language pre-trained models have shown remarkable success in guiding video question-answering (VideoQA) tasks. However, due to the length of video sequences, training large-scale video-based models incurs considerably higher costs than training image-based ones. This motivates us to leverage the knowledge from image-based pretraining, despite the obvious gaps between image and video domains. To bridge these gaps, in this paper, we propose Tem-Adapter, which enables the learning of temporal dynamics and complex semantics by a visual Temporal Aligner and a textual Semantic Aligner. Unlike conventional pretrained knowledge adaptation methods that only concentrate on the downstream task objective, the Temporal Aligner introduces an extra language-guided autoregressive task aimed at facilitating the learning of temporal dependencies, with the objective of predicting future states based on historical clues and language guidance that describes event progression. Besides, to reduce the semantic gap and adapt the textual representation for better event description, we introduce a Semantic Aligner that first designs a template to fuse" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 444, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 444, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 444, + 547, + 515 + ], + "type": "text", + "content": "question and answer pairs as event descriptions and then learns a Transformer decoder with the whole video sequence as guidance for refinement. We evaluate Tem-Adapter and different pre-train transferring methods on two VideoQA benchmarks, and the significant performance improvement demonstrates the effectiveness of our method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 532, + 387, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 532, + 387, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 532, + 387, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 552, + 548, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 548, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 548, + 673 + ], + "type": "text", + "content": "Video Question Answering (VideoQA) is a task that aims to answer natural language questions based on the information available in observed videos. It has attracted considerable attention recently due to its promise to develop interactive AI systems capable of communicating with the dynamic visual environment using natural language. Despite significant advancements in recent years, VideoQA remains a challenging problem that requires models to comprehensively understand and dynamically align the semantics of both the visual world and natural language." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 315, + 683, + 539, + 694 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 683, + 539, + 694 + ], + "spans": [ + { + "bbox": [ + 315, + 683, + 539, + 694 + ], + "type": "text", + "content": "1Our code can be found at: https://github.com/XLiu443/Tem-adapter" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 318, + 694, + 384, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 694, + 384, + 703 + ], + "spans": [ + { + "bbox": [ + 318, + 694, + 384, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 703, + 394, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 703, + 394, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 703, + 394, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13945" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 287, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 287, + 203 + ], + "type": "text", + "content": "Recently, a variety of methods [79, 78, 82, 34, 74, 63] have demonstrated impressive results in utilizing the largescale vision-language pre-trained (VLP) model to enhance downstream VideoQA tasks. These methods pre-train models by aligning the video and language domains and then apply them to VideoQA tasks via fine-tuning or even zero-shot learning. Nevertheless, pre-training large-scale models necessitate a large number of video-text pairs, e.g., more than 10 million videos, and entails expensive computational costs. This motivates us to explore cheaper and lighter alternative pre-trained models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 206, + 287, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 206, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 49, + 206, + 287, + 349 + ], + "type": "text", + "content": "Using image-VLP models is one potential option as they also align the semantics of vision and language domains. Compared to video-based models, image-based models offer two significant advantages. Firstly, training image-based models is less expensive when considering an equal number of data pairs. Secondly, image-VLP models have made significant progress in recent years, with some of them widely available and used, such as CLIP [56]. However, as depicted in Figure 1, domain gaps persist between image-based pretraining and VideoQA tasks since image-based models are unable to learn the temporal dynamics and corresponding complex event semantics of the visual world." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 352, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 352, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 49, + 352, + 287, + 507 + ], + "type": "text", + "content": "To address the challenge posed by the domain gaps, this paper proposes Tem-adapter , an adapter network that leverages the interaction between visual and textual modalities to learn temporal dynamics and complex semantics. Unlike existing methods that directly adapt visual and textual features with the objective of the downstream tasks, such as cross-modal matching in VideoQA, Tem-Adapter introduces an additional language-guided autoregressive task to facilitate learning of temporal dynamics, which enables visual representations to predict future states based on historical information and language guidance. Besides, we introduce cross-modal interactions to further reduce the semantic gap and refine textual representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 510, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 510, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 510, + 287, + 712 + ], + "type": "text", + "content": "Our approach consists of a visual Temporal Aligner and a textual Semantic Aligner. The Temporal Aligner uses a Transformer encoder to learn the temporal relations and refine visual features. To optimize it, we build an auto-regression model by a Transformer decoder to generate the future state with historical information. We also incorporate the language information for visual refinement by adding textual embeddings as the condition memory of the Transformer decoder. This auto-regression model is supervised by a reconstruction loss with ground truth future frames to encourage the learning of temporal dependencies. In addition to the visual branch, we introduce a Semantic Aligner for better event description by reducing the semantic gap between the original texts crawled from the net and question-answer pairs in VideoQA. First, we design templates to fuse textual questions and answers to generate declarative sentences to reduce the domain gap between training and downstream" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 545, + 108 + ], + "type": "text", + "content": "languages. Then, a Transformer decoder is employed to refine textual embeddings by incorporating the entire video sequence as a memory condition for video-text interactions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 108, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 203 + ], + "type": "text", + "content": "We conduct experiments on two public VideoQA benchmarks including SUTD-TrafficQA [76] and MSR-VTTMC [75], and evaluate different categories of methods that transfer the pre-trained knowledge to downstream tasks, such as other adapter models [12], prompt learning [89, 21], and fully or partial finetuning. The significant improvement demonstrates the effectiveness of Tem-adapter to adapt the image-language pre-trained model for VideoQA tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 215, + 391, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 215, + 391, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 215, + 391, + 227 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 235, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 235, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 235, + 545, + 592 + ], + "type": "text", + "content": "Video Question Answering. Compared with Image-based Question Answering, Video QA is more challenging due to the complex temporal dependency, which requires models to dynamically align the visual observation and natural language. Some previous methods [67, 3, 48, 81, 28, 65, 77, 27, 26, 4, 7, 6, 8, 41, 84, 91, 10, 11, 36, 13, 17, 22, 23, 51, 62, 72] learn the models with the given manually annotated datasets. Recently, more and more methods [79, 78, 82, 34, 74, 63, 29] propose to utilize the representation ability of large-scale pre-trained vision-language models to solve the VideoQA task. For example, VQA-T [78] pre-trains the vision language model with randomly sampled video-question-answer triplets in the HowTo100M [47] dataset. BiLM [79] pre-trains frozen bidirectional language models with some adapter layers on the WebVid10M [1] dataset and achieve excellent zero-shot VideoQA performance on the downstream tasks. However, pretraining the large-scale model is expensive, which motivates us to find a cheaper alternative manner. In this paper, we propose to use existing image-language pre-trained models, such as CLIP [56], to provide the alignment between the visual and textual domains. The most related method is ATP [2], which also applies CLIP as the pre-trained model to extract frame and language embeddings, and learns an atemporal probe to select the most discriminative frames to match the video and QA pairs. Different from ATP selecting frames to transfer the video to images, we propose Tem-adapter to learn the temporal dynamics and reduce the gap between the pre-trained image domain and downstream video domain." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 594, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 594, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 594, + 545, + 712 + ], + "type": "text", + "content": "Vision Language Pre-training. The goal of VLP is to help downstream tasks by pre-training the model on large-scale vision-text pairs [56, 37, 20]. It always crawls a huge amount of vision-text pairs from the web and trains the model in a self-supervised manner, such as reconstruction [40, 15, 9, 30], contrastive matching [56, 20, 18], or the combination of both two [38, 71, 25]. Different from image-VLP, video-VLP has more diverse proxy tasks, such as frame/language mask [66, 90], video-language matching [46], video/language generation [73, 45], and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "13946" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 284, + 115 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 284, + 115 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 284, + 115 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 284, + 115 + ], + "type": "image", + "image_path": "e23089f8115c8ff6203662a03a9be750cba9d115f2975625c50bc694c461adc1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 117, + 197, + 126 + ], + "lines": [ + { + "bbox": [ + 150, + 117, + 197, + 126 + ], + "spans": [ + { + "bbox": [ + 150, + 117, + 197, + 126 + ], + "type": "text", + "content": "(a) Finetuning" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 127, + 284, + 168 + ], + "blocks": [ + { + "bbox": [ + 50, + 127, + 284, + 168 + ], + "lines": [ + { + "bbox": [ + 50, + 127, + 284, + 168 + ], + "spans": [ + { + "bbox": [ + 50, + 127, + 284, + 168 + ], + "type": "image", + "image_path": "9deb2796159aac63423441de2909d2c1dd92490dbda499f47f2d7db1034284fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 170, + 206, + 179 + ], + "lines": [ + { + "bbox": [ + 141, + 170, + 206, + 179 + ], + "spans": [ + { + "bbox": [ + 141, + 170, + 206, + 179 + ], + "type": "text", + "content": "(b) Prompt learning" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 180, + 285, + 227 + ], + "blocks": [ + { + "bbox": [ + 50, + 180, + 285, + 227 + ], + "lines": [ + { + "bbox": [ + 50, + 180, + 285, + 227 + ], + "spans": [ + { + "bbox": [ + 50, + 180, + 285, + 227 + ], + "type": "image", + "image_path": "c07ebb8cbbf5e36fad87359dfba7e353bbf371faa1d7b93cc86de6a2270b9e5d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 228, + 193, + 238 + ], + "lines": [ + { + "bbox": [ + 155, + 228, + 193, + 238 + ], + "spans": [ + { + "bbox": [ + 155, + 228, + 193, + 238 + ], + "type": "text", + "content": "(c) Adapter" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 239, + 285, + 285 + ], + "blocks": [ + { + "bbox": [ + 50, + 239, + 285, + 285 + ], + "lines": [ + { + "bbox": [ + 50, + 239, + 285, + 285 + ], + "spans": [ + { + "bbox": [ + 50, + 239, + 285, + 285 + ], + "type": "image", + "image_path": "41862b9e82b6f391f70318d06e9508bad8cd59f288245db8eeaedba90993c1f3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 287, + 201, + 297 + ], + "lines": [ + { + "bbox": [ + 146, + 287, + 201, + 297 + ], + "spans": [ + { + "bbox": [ + 146, + 287, + 201, + 297 + ], + "type": "text", + "content": "(d) Tem-Adapter" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 306, + 289, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 306, + 289, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 289, + 413 + ], + "type": "text", + "content": "Figure 2: Various methods for pre-trained knowledge adaptation in VideoQA. (a) Fine-tuning updates pretrained parameters with downstream training. (b) Prompt learning freezes the pre-trained model and introduces learnable textual tokens for adaptation. (c) Adapter-based methods add new adapter layers to transfer knowledge. (d) Our approach, Tem-adapter , leverages the visual-textual interactions to learn temporal dynamics to reduce the gap between image-based pre-training and video-based tasks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "text", + "content": "frame/language order prediction [33, 39]. Recently, VLP models have shown great potential for other downstream applications, such as zero-shot and few-shot visual recognition [56, 12, 89, 88, 86], segmentation and detection [60, 87], image generation [50, 59, 52], and action understanding [70, 68]. In this paper, we use the CLIP model as our pre-trained model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": "Pre-trained Knowledge Transfer. In light of the remarkable success of the large-scale pre-trained model, many methods have been proposed to solve the problem that how to efficiently adapt the pre-trained knowledge to the downstream tasks. Early methods finetune all parameters of the pre-trained model with the downstream objectives [14, 53]. However, with the model scale increasing, the cost of finetuning the full model is extremely expensive. Thus, an increasing number of methods focus on knowledge adaptation with light parameters. A straightforward way is to partially finetune the model parameters [49]. Besides, prompt learning methods are proposed recently to reformulate the downstream tasks as the pre-trained tasks and freeze all model parameters [43, 54, 58, 55]. The previous prompt-based methods [54, 55] design the template of prompt using the prior of natural human language. Then, some methods [24, 64]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "propose to automatically optimize the prompt template, and [42, 69, 44, 89, 5] further learn the prompts in the continuous embedding space where the discrete word constraint is relaxed. In addition, some methods freeze the model and train some extra adapter layers and also achieve the supersizing performance [12, 85, 1]. As shown in Figure 2, we illustrate the difference between Tem-adapter and other categories of methods transferring the image-based VLP model to the VideoQA task. Compared with other pre-train adaptation methods, our Tem-adapter proposes an autoregression task to reduce the temporal gap (tem-align) and introduce the cross-domain interactions (sem-align) to reduce the semantic gap between the pre-training and downstream tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 242, + 372, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 372, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 372, + 255 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 263, + 547, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 547, + 323 + ], + "type": "text", + "content": "In this section, we first briefly introduce the VideoQA task and then present our approach, Tem-adapter , which introduces the cross-modal interactions to learn temporal dynamics and complex semantics. Finally, we give the details of our implementation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 334, + 417, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 334, + 417, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 417, + 346 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 354, + 546, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 389 + ], + "type": "text", + "content": "The goal of VideoQA is to find an answer " + }, + { + "bbox": [ + 304, + 354, + 546, + 389 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}" + }, + { + "bbox": [ + 304, + 354, + 546, + 389 + ], + "type": "text", + "content": " from the answer pool, given the conditions of question " + }, + { + "bbox": [ + 304, + 354, + 546, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 304, + 354, + 546, + 389 + ], + "type": "text", + "content": " and observed video data " + }, + { + "bbox": [ + 304, + 354, + 546, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 354, + 546, + 389 + ], + "type": "text", + "content": " as below:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 394, + 546, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 394, + 546, + 414 + ], + "spans": [ + { + "bbox": [ + 375, + 394, + 546, + 414 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {a}} = \\underset {\\mathbf {a} \\in \\mathcal {A}} {\\arg \\max } \\mathcal {S} (\\mathbf {a} | \\mathbf {v}, \\mathbf {q}), \\tag {1}", + "image_path": "1f5ead87b41cdec9200735f598f22081df3f48634e0f9947ae7cfdaf4bfc77e7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 417, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 417, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 417, + 545, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 417, + 545, + 441 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 306, + 417, + 545, + 441 + ], + "type": "text", + "content": " is a match function to generate the score of each answer candidate given the question and video content." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 452, + 391, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 452, + 391, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 452, + 391, + 464 + ], + "type": "text", + "content": "3.2. Tem-Adapter" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 472, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 567 + ], + "type": "text", + "content": "Here, we propose Tem-adapter to learn temporal dynamics and complex semantics by the interactions of visual and textual domains to adapt pre-trained Image-VLP models to downstream VideoQA tasks. Figure 3 shows an overview of Tem-Adapter's framework. In particular, Tem-adapter consists of two components, including a visual Temporal Aligner and a textual Semantic Aligner. Details for each component are as follows." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 547, + 628 + ], + "type": "text", + "content": "Temporal Aligner. The goal of the Temporal Aligner is to reduce the gap between image-based pre-training and video-based downstream tasks. To achieve this goal, we introduce an auxiliary language-guided autoregressive task to facilitate the learning of temporal dynamics." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "Specifically, given a video " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": ", we apply the frozen CLIP [56] image encoder " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": " to extract the feature of each frame as " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "f(\\mathbf{v}) = \\{f(\\mathbf{v}_t)|_{t = 1}^T\\} \\in \\mathcal{R}^{T\\times C}" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": " is the frame number of videos and " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": " is the feature dimension of CLIP. Then we refine these features with our Temporal Aligner model which is built by a transformer encoder " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": " to learn the temporal relations of frames." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13947" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 534, + 220 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 534, + 220 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 534, + 220 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 534, + 220 + ], + "type": "image", + "image_path": "221a9583e07cb8316e2249cea0dbeafaa5408829cc2a561d13a9d45d1e13d6a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 221, + 547, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 547, + 270 + ], + "type": "text", + "content": "Figure 3: The overall framework of Tem-adapter. Our Tem-adapter consists of two modules, Semantic Aligner (orange) and Temporal Aligner (blue). Semantic Aligner takes the text embedding of template-based QA pair as the input, and applies a Transformer decoder to incorporate cross-modal interactions. The Temporal Aligner introduces a language-guided autoregressive task to learn temporal dynamics. The training is guided by classification and reconstruction losses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 285, + 220, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 220, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 220, + 298 + ], + "type": "text", + "content": "The refined features can be represented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 307, + 288, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 307, + 288, + 321 + ], + "spans": [ + { + "bbox": [ + 131, + 307, + 288, + 321 + ], + "type": "interline_equation", + "content": "\\mathbf {f} = f _ {e} (f (\\mathbf {v}); \\phi), \\tag {2}", + "image_path": "c9b4d42ea79dc495d41bcba82d03df507cc9d4089301bc0942491692d2e38131.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 329, + 224, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 224, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 224, + 341 + ], + "type": "text", + "content": "which are with the same dimension " + }, + { + "bbox": [ + 46, + 329, + 224, + 341 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{T\\times C}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 342, + 288, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 288, + 413 + ], + "type": "text", + "content": "To guide the Temporal Aligner to learn temporal dynamics, we further build an autoregressive sequential model, following GPT [57], which generates the feature of the current frame with previous frames and guidance from textual information that describes event progression. Mathematically, it can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 417, + 287, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 417, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 102, + 417, + 287, + 432 + ], + "type": "interline_equation", + "content": "\\hat {f} (\\mathbf {v} _ {i}) = f _ {d} \\left(f \\left(\\mathbf {v} _ {1: i - 1}\\right) | \\mathbf {f}, \\mathbf {g}; \\psi\\right), \\tag {3}", + "image_path": "a1e505c606a556517c14934a2960e640ff8e722ee49416fa262a2111f46d061d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "inline_equation", + "content": "f(\\mathbf{v}_{1:i-1})" + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "text", + "content": " is the input of the transformer decoder, which is achieved by applying an attention mask to focus on the history clues. " + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{f}, \\mathbf{g}" + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "text", + "content": " are used as the guidance information, where " + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "text", + "content": " denotes latent visual embeddings obtained by Temporal Aligner and " + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "text", + "content": " (see equation 5) is the textual embeddings describing temporal event. " + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "text", + "content": " are first fused by adding and then serve as the memory key and value of the Transformer decoder. We apply a reconstruction loss to encourage the learning of temporal dependencies which is formulated as the distance between the results of the auto-regression model " + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "inline_equation", + "content": "\\hat{f}(\\mathbf{v}_i)" + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "text", + "content": " and the ground-truth " + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "inline_equation", + "content": "f(\\mathbf{v}_i)" + }, + { + "bbox": [ + 46, + 435, + 289, + 568 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 571, + 287, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 571, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 99, + 571, + 287, + 602 + ], + "type": "interline_equation", + "content": "L _ {M S E} = \\sum_ {i = 1} ^ {n} | | f (\\mathbf {v} _ {i}) - \\hat {f} (\\mathbf {v} _ {i}) | | _ {2} ^ {2}. \\tag {4}", + "image_path": "90816d12e05a7051fe3f01cd9d19d74fdfed99944295739379274e1fb870969b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 713 + ], + "type": "text", + "content": "Semantic Aligner. In addition to the visual branch, we introduce a Semantic Aligner to adapt the textual representation for better event description and reduce the semantic gaps. We first design templates to align the question-answer pairs in VideoQA to declarative sentences, since the training data of CLIP is crawled from the web which always describes the image in a declarative manner. Then, we apply a Transformer decoder with the video sequence as the memory for video-text interactions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 285, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 547, + 464 + ], + "type": "text", + "content": "Given a question " + }, + { + "bbox": [ + 304, + 285, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 304, + 285, + 547, + 464 + ], + "type": "text", + "content": " and a corresponding answer " + }, + { + "bbox": [ + 304, + 285, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 304, + 285, + 547, + 464 + ], + "type": "text", + "content": ", we first apply the grammar parsing on the question " + }, + { + "bbox": [ + 304, + 285, + 547, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 304, + 285, + 547, + 464 + ], + "type": "text", + "content": " to obtain the subject-verb-object structure, and then we modify the order of the sentence to obtain a declarative sentence and leave the position of the interrogative pronouns as a mask '[']. Finally, we fill in the default '['] with the answer to update the QA pair as a refined event description. For instance, the question is 'Which might be the reason for this accident?', and the answer is 'Improper lane change'. We first obtain a declarative sentence with the default as '[ ] is the reason for this accident', and then we can fill in the answer to get 'Speeding vehicle is the reason for this accident.' In addition, for the general question, we first change it to a declarative sentence and then convert the sentence to negative form if the answer is No." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 465, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 545, + 548 + ], + "type": "text", + "content": "With the refined sentence " + }, + { + "bbox": [ + 304, + 465, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{q}\\mathbf{a}" + }, + { + "bbox": [ + 304, + 465, + 545, + 548 + ], + "type": "text", + "content": " as the input, the frozen text encoder " + }, + { + "bbox": [ + 304, + 465, + 545, + 548 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 465, + 545, + 548 + ], + "type": "text", + "content": " of the CLIP model outputs the textual feature as " + }, + { + "bbox": [ + 304, + 465, + 545, + 548 + ], + "type": "inline_equation", + "content": "g(\\mathbf{q}\\mathbf{a}) \\in \\mathcal{R}^C" + }, + { + "bbox": [ + 304, + 465, + 545, + 548 + ], + "type": "text", + "content": ". To further refine the textual feature using the video content, we apply a transformer decoder with the visual feature sequence as the memory key and value and add a residual link to avoid overfitting. Formally, it can be formulated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 554, + 545, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 554, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 353, + 554, + 545, + 568 + ], + "type": "interline_equation", + "content": "\\mathbf {g} = g (\\mathbf {q a}) + \\lambda g _ {d} (g (\\mathbf {q a}), f (\\mathbf {v}); \\theta), \\tag {5}", + "image_path": "b42a1e4b059a459f3179984552de94ddab8c385434c3701b9d4949957427fd3b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "text", + "content": " denotes the learned textual embedding, " + }, + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "text", + "content": " is a parameter of the residual link, and " + }, + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "text", + "content": " is the parameters of the transformer decoder " + }, + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "inline_equation", + "content": "g_{d}" + }, + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "text", + "content": ". Both " + }, + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 571, + 547, + 654 + ], + "type": "text", + "content": " can be learned automatically in the training process. With the cross-attention module in the transformer decoder, we encourage the model to find key information from the video to refine the textual embedding for better event description." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 691 + ], + "type": "text", + "content": "Cross-modal Matching. Given the refined visual and textual embeddings " + }, + { + "bbox": [ + 304, + 654, + 547, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 304, + 654, + 547, + 691 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 547, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 304, + 654, + 547, + 691 + ], + "type": "text", + "content": ", we instantiate the score function in equation 1 by cosine similarity as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 367, + 700, + 546, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 700, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 367, + 700, + 546, + 715 + ], + "type": "interline_equation", + "content": "S (\\mathbf {a} | \\mathbf {v}, \\mathbf {q}) = \\operatorname {s o f t m a x} (\\bar {\\mathbf {f}} ^ {T} \\mathbf {g}), \\tag {6}", + "image_path": "1b52401a29bafecdd172392108172d29f150bcd6c02ed21c9355f442598c79b8.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13948" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{f}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " is video representation using the average pool on all frames, " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " is the language embedding with given answer question pair " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "(\\mathbf{a},\\mathbf{q})" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ". Then, we can obtain the answer to the question with the highest matching score." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 121, + 287, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 287, + 157 + ], + "type": "text", + "content": "Then we apply a standard multi-choice classification Hinge loss [19] integrated with the auto-regression loss for training. The overall loss function can be described as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 162, + 287, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 162, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 88, + 162, + 287, + 175 + ], + "type": "interline_equation", + "content": "L = L _ {H i n g e} (S (\\mathbf {a} | \\mathbf {v}, \\mathbf {q}), y) + \\gamma L _ {M S E}, \\tag {7}", + "image_path": "639d10e78fc467af887be4b72d5352d06c88142c3e583d1827f97718647baadc.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "text", + "content": " is a ratio to balance two loss functions, and " + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "text", + "content": " is the ground truth label of the answers. Please note that our Tem-adapter consists of 3 learnable modules including the textual decoder " + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "text", + "content": ", the visual encoder " + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "text", + "content": ", and the visual decoder " + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "text", + "content": ". The Hinge loss is applied to optimize " + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "text", + "content": ", while the auto-regression loss " + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "inline_equation", + "content": "L_{MSE}" + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "text", + "content": " is used to optimize all three modules. In the inference, only textual decoder " + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "text", + "content": " and visual encoder " + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 179, + 288, + 275 + ], + "type": "text", + "content": " are used." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 285, + 180, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 180, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 180, + 298 + ], + "type": "text", + "content": "3.3. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 304, + 289, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 289, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 289, + 327 + ], + "type": "text", + "content": "In this subsection, we provide more implementation details of our approach for better reproduction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 328, + 288, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 288, + 472 + ], + "type": "text", + "content": "Network Architecture. Figure 3 shows the framework of Tem-adapter. In our experiments, we use the frozen model CLIP (ViT-B32) as the text and image encoders. The Semantic Aligner is a one-layer Transformer decoder with 16 heads. There is a linear layer before the Transformer decoder which maps text embeddings and image embeddings to a lower dimension (128). Another linear projection layer is added after the Transformer decoder to map the reduced dimension back to 512. The Temporal Aligner is a light transformer encoder-decoder architecture. Both the encoder and decoder contain only one layer. We set the number of heads to 16 in the experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 473, + 288, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 288, + 593 + ], + "type": "text", + "content": "Inference Details. In the inference (given 1 question, 1 video, and " + }, + { + "bbox": [ + 46, + 473, + 288, + 593 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 473, + 288, + 593 + ], + "type": "text", + "content": " answers), we first apply the template to fuse the question and " + }, + { + "bbox": [ + 46, + 473, + 288, + 593 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 473, + 288, + 593 + ], + "type": "text", + "content": " answers to obtain " + }, + { + "bbox": [ + 46, + 473, + 288, + 593 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 473, + 288, + 593 + ], + "type": "text", + "content": " QA pairs (e.g. " + }, + { + "bbox": [ + 46, + 473, + 288, + 593 + ], + "type": "inline_equation", + "content": "k = 4" + }, + { + "bbox": [ + 46, + 473, + 288, + 593 + ], + "type": "text", + "content": "). Then we feed the QA pairs and video into CLIP textual and visual encoders to obtain the initial features. After that, we apply the learned semantic and temporal aligners to refine these features. Please note that the temporal auto-regression module is not used in the inference. Finally, we obtain 1 video feature with " + }, + { + "bbox": [ + 46, + 473, + 288, + 593 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 473, + 288, + 593 + ], + "type": "text", + "content": " QA features and apply the cosine similarity as well as softmax to select the answer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Hyper-parameters. We follow the same video frames sampling method with prior methods [76]. We uniformly select 8 frames over a long video sequence and then pick consecutive 16 frames which are centered on those 8 frames. Overall 128 frames are selected for each video. We also tried another sampling method: uniformly extracting 128 or 256 frames on a video sequence, and the final results do not support the sampling method very well. In the loss function, we set " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " as 100 to encourage the model to learn the temporal dynamics. And the learnable parameter " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " in the Semantic" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "Aligner is set to be the same dimension (512) as the output embeddings of the CLIP model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 96, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 193 + ], + "type": "text", + "content": "Training Details. We use the pre-trained CLIP model to extract text embeddings and image embeddings. The Adam [31] optimizer is utilized to optimize the training process. We set the initial learning rate to be " + }, + { + "bbox": [ + 304, + 96, + 547, + 193 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 304, + 96, + 547, + 193 + ], + "type": "text", + "content": " with a batch size of 128. And we let the learning rate decay with a factor of 0.5 for every 10 epochs. Our model is trained for 50 epochs on a single Tesla V100 GPU and the training takes 8 hours to finish. Our model is implemented in PyTorch 1.8.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 205, + 388, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 388, + 219 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 388, + 219 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 226, + 547, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 226, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 547, + 275 + ], + "type": "text", + "content": "In this section, we first benchmark our Tem-adapter and different categories of methods that transfer the pretrained model into downstream tasks. Then we provide ablation studies to investigate the effect of each component." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 283, + 493, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 283, + 493, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 283, + 493, + 296 + ], + "type": "text", + "content": "4.1. Datasets & Experimental Protocols" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 301, + 547, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 547, + 469 + ], + "type": "text", + "content": "SUTD-TrafficQA. SUTD-TrafficQA focuses on the question-answering task in the traffic situation, which requires the model to understand the traffic event and underlying causal relation. It consists of more than 10k videos with different traffic events and provides over 62,535 human-annotated QA pairs. Among them, 56,460 QA pairs are used for training and the rest 6,075 QA pairs are used for testing. This dataset highlights the cognitive capability of video event understanding models in complex traffic scenarios. It provides 6 challenging traffic-related reasoning tasks including \"Basic understanding\", \"Event forecasting\", \"Reverse reasoning\", \"Counterfactual inference\", \"Introspection\", \"Attribution\". All tasks are formulated as multiple-choice without limiting the number of candidate answers." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 470, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 547, + 626 + ], + "type": "text", + "content": "MSR-VTT-MC. MSR-VTT is collected and released mainly for the text-to-video retrieval task. Following [34] and [2], we use the multi-choice test setting to benchmark the VideoQA methods. The MSR-VTT-MC test set consists of 2990 QA pairs. Each video corresponds to five answers, with the original caption as the correct answer. We use the standard protocol with prior works in our experiments, where 7010 train+val video sequences and 140,200 QA pairs are utilized for training. Each video only contains one caption. We assign this caption as the correct answer for our question-answering task. To formulate four negative answers during training, we randomly select four captions from other captions that do not belong to the current video." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 635, + 411, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 635, + 411, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 635, + 411, + 647 + ], + "type": "text", + "content": "4.2. Baseline Methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "To investigate different categories of methods that transfer the pre-trained model into downstream tasks, such as finetuning, prompt learning, and adapter. We compare Tem-adapter with the other 10 baseline methods including Unsupervised CLIP [56], Unsupervised CLIP [56]" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13949" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 114, + 516, + 294 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": "Table 1: VideoQA accuracies of baseline methods on both SUTD-TrafficQA and MSR-VTT-MC. B: \"Basic understanding\", F: \"Forecasting task\", R: \"Reverse Reasoning\", C: \"Counterfactual inference\", I: \"Introspection\", A: \"Attribution\". (C) and (A) denote strategies for adding prompts. (C) and " + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "inline_equation", + "content": "(\\mathbf{C}^{*})" + }, + { + "bbox": [ + 46, + 70, + 547, + 109 + ], + "type": "text", + "content": " denote training prompts with/without adapter heads." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 114, + 516, + 294 + ], + "lines": [ + { + "bbox": [ + 77, + 114, + 516, + 294 + ], + "spans": [ + { + "bbox": [ + 77, + 114, + 516, + 294 + ], + "type": "table", + "html": "
MethodsSUTD-TrafficQAMSR-VTT-MC
BFRCIAAvgAcc
Unsupervised CLIP [56]25.620.134.030.822.828.826.574.1
CLIP [56] + Template31.836.029.971.822.133.432.374.1
Totally finetuning39.835.146.645.637.240.540.389.0
Partially finetuning41.637.844.650.033.141.741.788.5
LORA [16]38.738.736.737.934.538.138.385.4
CLIP-Adapter [12]35.832.035.442.333.132.134.883.4
Multi-layer Adapter [12]30.526.626.538.528.325.829.190.7
Prompt learning (C) [89]42.432.445.255.540.743.642.989.0
Prompt learning (C*) [89]40.333.241.046.534.938.439.790.8
Prompt learning (A) [21]41.731.540.148.433.141.441.188.0
Tem-adapter46.036.544.655.034.547.746.094.3
", + "image_path": "4928a77f32bebceeff81172a5fa02e81593c65bf81d956a20d6928eece3da2ca.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 52, + 368, + 282, + 494 + ], + "blocks": [ + { + "bbox": [ + 46, + 299, + 288, + 360 + ], + "lines": [ + { + "bbox": [ + 46, + 299, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 288, + 360 + ], + "type": "text", + "content": "Table 2: Ablation studies on the SUTD-TrafficQA and MSR-VTT-MC datasets. Tp, SA, TA, and Ar denote Template, Semantic Aligner, Temporal Aligner, and Auto-regression, respectively. No ablation studies (“\\”) concentrate on Tp, as it is not compatible with the MSR-VTT-MC dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 368, + 282, + 494 + ], + "lines": [ + { + "bbox": [ + 52, + 368, + 282, + 494 + ], + "spans": [ + { + "bbox": [ + 52, + 368, + 282, + 494 + ], + "type": "table", + "html": "
MethodsSUTD-TrafficQAMSR-VTT-MC
AvgAcc
CLIP-Adapter [12]34.883.4
w/o Tp43.9\\
w/o SA44.291.8
w/o TA43.691.6
w/o Ar44.992.6
w/o TA and SA32.374.1
Tem-adapter46.094.3
", + "image_path": "e3e70cee006617c67fac8a15a0831e8dc684a384ae6de4ac026c9e08a40a79dc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 506, + 288, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 506, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 45, + 506, + 288, + 579 + ], + "type": "text", + "content": "+ Language template, Totally fine-tuning, Partially finetuning, LORA [16], CLIP-Adapter [12], Multi-layer CLIP-Adapter [12], Prompt learning (change words) [89] (with/without using adapter heads), and Prompt learning (add words) [21]. The detailed descriptions are shown in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 586, + 261, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 261, + 600 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 261, + 600 + ], + "type": "text", + "content": "4.3. Benchmarking against Baseline Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 665 + ], + "type": "text", + "content": "In this subsection, we compare and discuss our approach's performance against the above-mentioned baselines on both SUTD-TrafficQA and MSR-VTT-MC datasets. The results of all methods on both datasets are summarized in Table 1. We can draw conclusions as follows from the results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "CLIP can obtain good zero-shot results. With our language template, the clip model can achieve " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "32.3\\%" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " accuracy on the SUTD-TrafficQA dataset, which is almost close to the previous state-of-the-art performance " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "(37.1\\%)" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": ". It shows" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 301, + 515, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 301, + 515, + 313 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 515, + 313 + ], + "type": "text", + "content": "the strong representation ability of the CLIP model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 315, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 547, + 495 + ], + "type": "text", + "content": "More parameters " + }, + { + "bbox": [ + 304, + 315, + 547, + 495 + ], + "type": "inline_equation", + "content": "\\neq" + }, + { + "bbox": [ + 304, + 315, + 547, + 495 + ], + "type": "text", + "content": " Better performance. We have two groups of experiments that focus on the number of parameters, including totally/partially finetuning and single-layer/multi-layer Adapters. We obtain contradictory conclusions from different datasets. On SUTD-TrafficQA, the model with fewer parameters works better, while on MSR-VTT-MC, the model with more parameters has better performance. It is not surprising since the scale of training data is different. MSR-VTT-MC has more training data and can support training more parameters. We also compare the learnable parameters of adapter-based methods in Table 5. Tem-adapter achieve better performance with fewer parameters than Multi-layer Adapter. Compared with the prompt-learning-based method, our method needs more parameters, but its inference is not time-consuming." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 496, + 548, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 548, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 548, + 605 + ], + "type": "text", + "content": "Different categories of methods are comparable. Balancing the performance on both datasets, different categories of methods that transfer the image-based VLP model to VideoQA don't have clear performance differences. Adapter-based methods obtain poor results on SUTD-TrafficQA, but multi-layer Adapter works well on MSR-VTT-MC. Besides, by comparing the prompt learning methods with/without adapter heads, we found that the prompt and adapter can be joint learning to obtain better performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": "Cross-modal interaction and temporal dynamics matter. Compared with CLIP-Adapter (both single-layer and multi-layer ones), our Tem-adapter achieves more than " + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 605, + 548, + 715 + ], + "type": "text", + "content": " improvement on SUTD-TrafficQA. The key difference between Tem-adapter and other adapter-based methods is that we introduce temporal alignment to reduce the temporal gap and apply the cross-modal interaction to refine the representation and reduce the semantic gap. Tem-adapter outperforms other methods by a large mar" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13950" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 102, + 273, + 243 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 95 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 95 + ], + "type": "text", + "content": "Table 3: Comparison with the state-of-the-art methods on the SUTD-TrafficQA dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 102, + 273, + 243 + ], + "lines": [ + { + "bbox": [ + 62, + 102, + 273, + 243 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 273, + 243 + ], + "type": "table", + "html": "
MethodsAccuracyYear
Human95.4-
VIS+LSTM [61]29.92015
TVQA [35]35.22018
HCRN [32]36.52020
BERT-VQA [80]33.72020
Eclipse [76]37.12021
CLIP [56]32.32021
CLIP+HCRN[56, 32]41.92021
ATP [2]35.62022
Tem-adapter46.0
", + "image_path": "3c28eba5066b0934583c6481c1dab079ca56c63bcfce0b78ecb1cfe6a0f83700.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 289 + ], + "type": "text", + "content": "gin on both two datasets. For example, it obtains " + }, + { + "bbox": [ + 46, + 251, + 288, + 289 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 46, + 251, + 288, + 289 + ], + "type": "text", + "content": " improvement than Prompt learning (C) on SUTD-TrafficQA and " + }, + { + "bbox": [ + 46, + 251, + 288, + 289 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 46, + 251, + 288, + 289 + ], + "type": "text", + "content": " than Multi-layer Adapter on MSR-VTT-MC." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 296, + 286, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 296, + 286, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 296, + 286, + 309 + ], + "type": "text", + "content": "4.4. Comparison against State-of-the-Art Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 315, + 288, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 288, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 288, + 505 + ], + "type": "text", + "content": "SUTD-TrafficQA. On SUTD-TrafficQA, we compare our proposed method with state-of-the-art methods, including [32, 76, 61, 35, 80]. The results are summarized in Table 3. We can observe that our Tem-adapter outperforms other methods by a large margin. For example, compared with HCRN [32], Tem-adapter achieves almost " + }, + { + "bbox": [ + 46, + 315, + 288, + 505 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 315, + 288, + 505 + ], + "type": "text", + "content": " improvement. By using the CLIP model to extract visual features, HCRN+CLIP can obtain " + }, + { + "bbox": [ + 46, + 315, + 288, + 505 + ], + "type": "inline_equation", + "content": "41.9\\%" + }, + { + "bbox": [ + 46, + 315, + 288, + 505 + ], + "type": "text", + "content": " accuracy, which is better than using the original HCRN. Besides, we show that the performance improvement is not only from the better representation ability of the CLIP model since our Tem-adapter further improve HCRN+CLIP with " + }, + { + "bbox": [ + 46, + 315, + 288, + 505 + ], + "type": "inline_equation", + "content": "+4.1\\%" + }, + { + "bbox": [ + 46, + 315, + 288, + 505 + ], + "type": "text", + "content": " accuracy. In addition, we also significantly outperform the ATP method, which also uses the CLIP model. This significant performance improvement demonstrates the effectiveness of Tem-adapter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 506, + 288, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 288, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 288, + 663 + ], + "type": "text", + "content": "MSR-VTT-MC. On MSR-VTT-MC, we compare Tem-adapter with some recent state-of-the-art methods, including ActBERT[90], ClipBERT[34], MERLOT [83], VideoCLIP [66], CLIP [56], and ATP [2]. Results are shown in Table 4. First, we can find that the CLIP model benefits the performance on the MSR-VTT-MC dataset. All methods using the CLIP model, such as CLIP [56], ATP [2], and Tem-adapter achieve good performance. ATP [2] learns an atemporal probe to select a frame from a video sequence to obtain the visual embedding and achieve good performance on MSR-VTT-MC. Compared with it, Tem-adapter can further obtain " + }, + { + "bbox": [ + 46, + 506, + 288, + 663 + ], + "type": "inline_equation", + "content": "94.3\\%" + }, + { + "bbox": [ + 46, + 506, + 288, + 663 + ], + "type": "text", + "content": " accuracy and " + }, + { + "bbox": [ + 46, + 506, + 288, + 663 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 46, + 506, + 288, + 663 + ], + "type": "text", + "content": " relative improvement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 670, + 149, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 149, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 149, + 682 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "We provide ablation studies to investigate the effect of each component. Four main modules in our Tem-adapter" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 332, + 102, + 522, + 216 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 4: Comparison with the state-of-the-art methods on the MSR-VTT-MC dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 102, + 522, + 216 + ], + "lines": [ + { + "bbox": [ + 332, + 102, + 522, + 216 + ], + "spans": [ + { + "bbox": [ + 332, + 102, + 522, + 216 + ], + "type": "table", + "html": "
MethodsAccuracyYear
VideoCLIP [66]92.12019
ActBERT[90]85.72020
ClipBERT[34]88.22021
MERLOT [83]90.92021
CLIP [56]74.12021
ATP [2]93.22022
Tem-adapter94.3
", + "image_path": "07eaf825934fabd225f5efc47a414372420e128fe08c38228ed86e97e6186fe3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 228, + 547, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 228, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 547, + 251 + ], + "type": "text", + "content": "are evaluated, including Temporal Aligner, Semantic Aligner, language template, and auto-regression." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 252, + 547, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 547, + 420 + ], + "type": "text", + "content": "As shown in Table 2, we provide the performance of different variants of Tem-adapter and a baseline adapter method on each task in the SUTD-TrafficQA and MSR-VTTMC datasets. The baseline method is CLIP-adapter, which also uses the language template and learns a linear adapter layer without cross-domain interaction. We observe that all variants of Tem-adapter can achieve significant improvement over the baseline method, which shows that the interaction across visual and textual is very important for video analysis. Besides, compared with Tem-adapter, we can observe a clear performance drop when we remove each component, which demonstrates all of them are indispensable. When both temporal and semantic aligners are removed, the performance drops dramatically." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 421, + 547, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 547, + 504 + ], + "type": "text", + "content": "In addition, we also conducted an experiment to show whether the Temporal Aligner learns the temporal dynamic information. Table 6 shows the reconstruction quality of our methods with/without language guidance " + }, + { + "bbox": [ + 304, + 421, + 547, + 504 + ], + "type": "inline_equation", + "content": "(g)" + }, + { + "bbox": [ + 304, + 421, + 547, + 504 + ], + "type": "text", + "content": " and with/without Temporal Aligner " + }, + { + "bbox": [ + 304, + 421, + 547, + 504 + ], + "type": "inline_equation", + "content": "(f)" + }, + { + "bbox": [ + 304, + 421, + 547, + 504 + ], + "type": "text", + "content": ". We can observe that both the language guidance and Temporal Aligner assist in the learning of temporal dynamics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 506, + 547, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 565 + ], + "type": "text", + "content": "Furthermore, we also provide some parameter analysis experiments about our network design. And evaluate Tem-adapter on different image-text pre-train models. Please kindly refer to the supplementary materials for the detailed results and analysis." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 574, + 390, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 390, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 390, + 586 + ], + "type": "text", + "content": "4.6. Visualization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "In this subsection, we provide some qualitative results to analyze our approach. As shown in Figure 4, we show some visualization examples of VdideoQA on the SUTD-TrafficQA dataset. In the top row, we plot two positive examples. We can observe that our Tem-adapter learns the temporal dynamics with the cross-domain interactions to obtain a better understanding of video events. Besides, we also provide some failure cases to help explore the boundary of our approach. We find that Tem-adapter can not work well for situations that need complex reasoning, especially" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13951" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 74, + 257, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 74, + 257, + 86 + ], + "spans": [ + { + "bbox": [ + 77, + 74, + 257, + 86 + ], + "type": "text", + "content": "Q: What could possibly cause this accident?" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 87, + 286, + 131 + ], + "blocks": [ + { + "bbox": [ + 59, + 87, + 286, + 131 + ], + "lines": [ + { + "bbox": [ + 59, + 87, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 59, + 87, + 286, + 131 + ], + "type": "image", + "image_path": "03e235e26b72a5879a70ea52c006d335d3753213ce85a1a893e4be646de2423c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 78, + 133, + 251, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 133, + 251, + 143 + ], + "spans": [ + { + "bbox": [ + 78, + 133, + 251, + 143 + ], + "type": "text", + "content": "Ours: Sudden or extreme movement by a vehicle." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 143, + 208, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 143, + 208, + 153 + ], + "spans": [ + { + "bbox": [ + 78, + 143, + 208, + 153 + ], + "type": "text", + "content": "CLIP-Adapter: Unfavorable weather." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 153, + 225, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 153, + 225, + 163 + ], + "spans": [ + { + "bbox": [ + 78, + 153, + 225, + 163 + ], + "type": "text", + "content": "Prompt learning: Violation of traffic rules." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 163, + 206, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 163, + 206, + 175 + ], + "spans": [ + { + "bbox": [ + 78, + 163, + 206, + 175 + ], + "type": "text", + "content": "Finetuning: Violation of traffic rules." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 184, + 258, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 184, + 258, + 207 + ], + "spans": [ + { + "bbox": [ + 77, + 184, + 258, + 207 + ], + "type": "text", + "content": "Q: Could any vehicles steer away to prevent the accident?" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 56, + 211, + 282, + 255 + ], + "blocks": [ + { + "bbox": [ + 56, + 211, + 282, + 255 + ], + "lines": [ + { + "bbox": [ + 56, + 211, + 282, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 282, + 255 + ], + "type": "image", + "image_path": "1cd05c0e3725a624008bedcf98cbe31a252065b747e5c8e14ed03eb864520497.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 294, + 545, + 343 + ], + "lines": [ + { + "bbox": [ + 46, + 294, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 545, + 343 + ], + "type": "text", + "content": "Figure 4: Visualization of some examples of the VideoQA task from the SUTD-TrafficQA dataset. The top row shows some positive examples, where Tem-adapter learns temporal dynamics to understand the traffic event and give the correct answers. The bottom row shows some failed cases. Despite better matching between visual and textual domains, our approach fails on complex problems which require reasoning ability." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 256, + 212, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 256, + 212, + 266 + ], + "spans": [ + { + "bbox": [ + 91, + 256, + 212, + 266 + ], + "type": "text", + "content": "Ours: No, there was no space. X" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 266, + 209, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 266, + 209, + 277 + ], + "spans": [ + { + "bbox": [ + 91, + 266, + 209, + 277 + ], + "type": "text", + "content": "GT: Yes, there was enough space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 74, + 500, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 74, + 500, + 86 + ], + "spans": [ + { + "bbox": [ + 320, + 74, + 500, + 86 + ], + "type": "text", + "content": "Q: Which factors contribute to the accident?" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 87, + 533, + 130 + ], + "blocks": [ + { + "bbox": [ + 308, + 87, + 533, + 130 + ], + "lines": [ + { + "bbox": [ + 308, + 87, + 533, + 130 + ], + "spans": [ + { + "bbox": [ + 308, + 87, + 533, + 130 + ], + "type": "image", + "image_path": "872773b7ceeb0ea45defc8048668de003dd580de6a1153bd856e6e2f932bed1a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 332, + 133, + 500, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 133, + 500, + 143 + ], + "spans": [ + { + "bbox": [ + 332, + 133, + 500, + 143 + ], + "type": "text", + "content": "Ours: Not paying attention to their surroundings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 332, + 143, + 449, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 143, + 449, + 152 + ], + "spans": [ + { + "bbox": [ + 332, + 143, + 449, + 152 + ], + "type": "text", + "content": "CLIP-Adapter: Bad road surfaces" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 332, + 153, + 457, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 153, + 457, + 163 + ], + "spans": [ + { + "bbox": [ + 332, + 153, + 457, + 163 + ], + "type": "text", + "content": "Prompt learning: Bad road surfaces." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 332, + 163, + 438, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 163, + 438, + 175 + ], + "spans": [ + { + "bbox": [ + 332, + 163, + 438, + 175 + ], + "type": "text", + "content": "Finetuning: Speeding vehicles." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 185, + 511, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 185, + 511, + 209 + ], + "spans": [ + { + "bbox": [ + 320, + 185, + 511, + 209 + ], + "type": "text", + "content": "Q: Could the accident be prevented if a convex mirror is installed at the turning point?" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 211, + 531, + 254 + ], + "blocks": [ + { + "bbox": [ + 307, + 211, + 531, + 254 + ], + "lines": [ + { + "bbox": [ + 307, + 211, + 531, + 254 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 531, + 254 + ], + "type": "image", + "image_path": "51fbfa3d51d96754fa0e286797bf9782353f3153f2c66544d1103bc396a8929e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 331, + 256, + 496, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 256, + 496, + 266 + ], + "spans": [ + { + "bbox": [ + 331, + 256, + 496, + 266 + ], + "type": "text", + "content": "Ours: Yes, there would be no obstructed views." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 331, + 266, + 514, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 266, + 514, + 276 + ], + "spans": [ + { + "bbox": [ + 331, + 266, + 514, + 276 + ], + "type": "text", + "content": "GT: No, that was not the main cause of the accident." + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 47, + 377, + 550, + 426 + ], + "blocks": [ + { + "bbox": [ + 46, + 350, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 350, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 545, + 373 + ], + "type": "text", + "content": "Table 5: The comparisons of our method and other baseline methods on model parameters (M), inference time (ms/video), and accuracy in the SUTD-TrafficQA dataset." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 377, + 550, + 426 + ], + "lines": [ + { + "bbox": [ + 47, + 377, + 550, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 377, + 550, + 426 + ], + "type": "table", + "html": "
MetricsLORA [16]Prompt learning (C) [89]Prompt learning (C*) [89]CLIP-Adapter [12]Multi-layer Adapter [12]Ours
Parameters1.443.550.771.054.853.62
Inference time5.375.335.175.135.145.32
Accuracy38.342.939.734.829.146.0
", + "image_path": "3d4e1ca4ce98a3cab4ac840c115206c6484c89e892b25ef81285a63ab1eacc30.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 60, + 456, + 272, + 493 + ], + "blocks": [ + { + "bbox": [ + 46, + 429, + 287, + 453 + ], + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 453 + ], + "type": "text", + "content": "Table 6: The comparison of reconstruction ability. We use peak signal-to-noise ratio (PSNR) as the metric." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 456, + 272, + 493 + ], + "lines": [ + { + "bbox": [ + 60, + 456, + 272, + 493 + ], + "spans": [ + { + "bbox": [ + 60, + 456, + 272, + 493 + ], + "type": "table", + "html": "
Tem-adapterw/o gw/o f and g
PSNR29.128.727.4
", + "image_path": "b79c3e408ee4475342ab5238d4af0d1be0974715d93a16e5d46c2fecadf95938.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 500, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 287, + 548 + ], + "type": "text", + "content": "for the underlying cause-effect relations. We leave it as a future work to enhance the reasoning ability of our model. Besides, we also provide more examples of our language templates in the supplementary materials." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 47, + 560, + 119, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 119, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 119, + 573 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": "In this paper, we proposed Tem-adapter to adapt the pre-trained image-language model to the downstream VideoQA task. We jointly learn the Temporal Aligner and Semantic Aligner through cross-modal interactions. Tem-adapter introduces a language-guided autoregressive task to guide the learning of temporal dependency and thus reduce the temporal gap between image-based pre-train and video-based QA tasks. Besides, we utilize a rule-based template and video-based interaction to refine textual representation and reduce semantic gaps. We validate the effectiveness of our approach on four VideoQA benchmarks." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 430, + 547, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 526 + ], + "type": "text", + "content": "Broader Impacts and Limitations. Understanding how the big model adaptation works on VideoQA is of great help in developing general foundation models and interactive AI. However, our work also has certain limitations. For example, since extracting video features online takes up too many computing resources, we extract their features offline and fix them. We leave it as future work to learn the adapter within the backbone network." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 537, + 399, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 399, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 399, + 551 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 557, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 714 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62206153, Young Elite Scientists Sponsorship Program by CAST (No. 2023QNRC002), by the Tsinghua Shenzhen International Graduate School-Shenzhen Pengrui Endowed Professorship Scheme of Shenzhen Pengrui Foundation, by the UKRI grant: Turing AI Fellowship EP/W002981/1, by the National Institutes of Health (NIH) under Contract R01HL159805, by the NSF-Convergence Accelerator Track-D award #2134901, by a grant from Apple Inc., a grant from KDDI Research Inc, and generous gifts from Salesforce Inc., Microsoft Research, and Amazon Research. We would also like to thank the Royal Academy of Engineering and FiveAI." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13952" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 124 + ], + "type": "text", + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "text", + "content": "[2] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the\" video\" in video-language understanding. In " + }, + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "inline_equation", + "content": "CVPR" + }, + { + "bbox": [ + 53, + 124, + 288, + 168 + ], + "type": "text", + "content": ", pages 2917-2927, 2022. 2, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "text", + "content": "[3] Santiago Castro, Mahmoud Azab, Jonathan Stroud, Cristina Noujaim, Ruoyao Wang, Jia Deng, and Rada Mihalcea. Lifeqa: A real-life dataset for video question answering. In LREC, pages 4352-4358, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 258 + ], + "type": "text", + "content": "[4] Aman Chadha, Gurneet Arora, and Navpreet Kaloty. iperceive: Applying common-sense reasoning to multi-modal dense video captioning and video question answering. arXiv preprint arXiv:2011.07735, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 301 + ], + "type": "text", + "content": "[5] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Plot: Prompt learning with optimal transport for vision-language models. In ICLR, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 301, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 301, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 301, + 288, + 356 + ], + "type": "text", + "content": "[6] Seongho Choi, Kyoung-Woon On, Yu-Jung Heo, Ahjeong Seo, Youwon Jang, Minsu Lee, and Byoung-Tak Zhang. Dramaqa: Character-centered video story understanding with hierarchical qa. In AAAI, volume 35, pages 1166-1174, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 401 + ], + "type": "text", + "content": "[7] Anthony Colas, Seokhwan Kim, Franck Dernoncourt, Siddhesh Gupte, Daisy Zhe Wang, and Doo Soon Kim. Tutorial: Question answering dataset for tutorial videos. arXiv preprint arXiv:1912.01046, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "type": "text", + "content": "[8] Long Hoang Dang, Thao Minh Le, Vuong Le, and Truyen Tran. Object-centric representation learning for video question answering. In IJCNN, pages 1-8, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 435, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 288, + 490 + ], + "type": "text", + "content": "[9] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Zicheng Liu, Michael Zeng, et al. An empirical study of training end-to-end vision-and-language transformers. arXiv preprint arXiv:2111.02387, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 491, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 288, + 535 + ], + "type": "text", + "content": "[10] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In CVPR, pages 1999-2007, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 535, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 288, + 568 + ], + "type": "text", + "content": "[11] Jiyang Gao, Runzhou Ge, Kan Chen, and Ram Nevatia. Motion-appearance co-memory networks for video question answering. In CVPR, pages 6576–6585, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 568, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 288, + 613 + ], + "type": "text", + "content": "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint arXiv:2110.04544, 2021. 2, 3, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 613, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 288, + 646 + ], + "type": "text", + "content": "[13] Noa Garcia, Mayu Otani, Chenhui Chu, and Yuta Nakashima. Knowit vqa: Answering knowledge-based questions about videos. In AAAI, volume 34, pages 10826-10834, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 646, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 288, + 667 + ], + "type": "text", + "content": "[14] Ross Girshick. Fast r-cnn. In ICCV, pages 1440-1448, 2015. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 288, + 712 + ], + "type": "text", + "content": "[15] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. VIn bert: A recurrent vision-and-language bert for navigation. In CVPR, pages 1643–1653, 2021. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 547, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 163 + ], + "type": "text", + "content": "[17] Deng Huang, Peihao Chen, Runhao Zeng, Qing Du, Mingkui Tan, and Chuang Gan. Location-aware graph convolutional networks for video question answering. In AAAI, volume 34, pages 11021-11028, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 547, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 209 + ], + "type": "text", + "content": "[18] Aashi Jain, Mandy Guo, Krishna Srinivasan, Ting Chen, Sneha Kudugunta, Chao Jia, Yinfei Yang, and Jason Baldridge. Mural: multimodal, multitask retrieval across languages. arXiv preprint arXiv:2109.05125, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 209, + 547, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 547, + 253 + ], + "type": "text", + "content": "[19] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In CVPR, pages 2758–2766, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 255, + 547, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 547, + 310 + ], + "type": "text", + "content": "[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 312, + 547, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 312, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 547, + 346 + ], + "type": "text", + "content": "[21] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. arXiv preprint arXiv:2203.12119, 2022. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 346, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 346, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 346, + 547, + 392 + ], + "type": "text", + "content": "[22] Jianwen Jiang, Ziqiang Chen, Haojie Lin, Xibin Zhao, and Yue Gao. Divide and conquer: Question-guided spatiotemporal contextual attention for video question answering. In AAAI, volume 34, pages 11101-11108, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 392, + 547, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 392, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 392, + 547, + 426 + ], + "type": "text", + "content": "[23] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In AAAI, volume 34, pages 11109-11116, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 427, + 547, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 427, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 427, + 547, + 460 + ], + "type": "text", + "content": "[24] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? TACL, 8:423-438, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 462, + 547, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 462, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 462, + 547, + 506 + ], + "type": "text", + "content": "[25] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1780-1790, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 507, + 547, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 507, + 547, + 551 + ], + "spans": [ + { + "bbox": [ + 307, + 507, + 547, + 551 + ], + "type": "text", + "content": "[26] Hyoungun Kim, Zineng Tang, and Mohit Bansal. Dense-ecaption matching and frame-selection gating for temporal localization in videoqa. arXiv preprint arXiv:2005.06409, 2020.2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 553, + 547, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 553, + 547, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 553, + 547, + 597 + ], + "type": "text", + "content": "[27] Junyeong Kim, Minuk Ma, Trung Pham, Kyungsu Kim, and Chang D Yoo. Modality shifting attention network for multimodal video question answering. In CVPR, pages 10106-10115, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 599, + 547, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 599, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 599, + 547, + 633 + ], + "type": "text", + "content": "[28] Kyung-Min Kim, Min-Oh Heo, Seong-Ho Choi, and Byoung-Tak Zhang. Deepstory: Video story qa by deep embedded memory networks. arXiv preprint arXiv:1707.00836, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 633, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 633, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 633, + 547, + 678 + ], + "type": "text", + "content": "[29] Seonhoon Kim, Seohyeong Jeong, Eunbyul Kim, Inho Kang, and Nojun Kwak. Self-supervised pre-training and contrastive representation learning for multiple-choice video qa. In AAAI, volume 35, pages 13171-13179, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 679, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 547, + 714 + ], + "type": "text", + "content": "[30] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, pages 5583-5594, 2021. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13953" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "text", + "content": "[32] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In CVPR, pages 9972-9981, 2020. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "text", + "content": "[33] Chenyi Lei, Shixian Luo, Yong Liu, Wanggui He, Jiamang Wang, Guoxin Wang, Haihong Tang, Chunyan Miao, and Houqiang Li. Understanding chinese video and language via contrastive multimodal pre-training. In ACM MM, pages 2567-2576, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "type": "text", + "content": "[34] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In CVPR, pages 7331-7341, 2021. 2, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 271 + ], + "type": "text", + "content": "[35] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. Tvqa: Localized, compositional video question answering. arXiv preprint arXiv:1809.01696, 2018. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "text", + "content": "[36] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvqa+: Spatio-temporal grounding for video question answering. arXiv preprint arXiv:1904.11574, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 348 + ], + "type": "text", + "content": "[37] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "text", + "content": "[38] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 34, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "text", + "content": "[39] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[40] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 482, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 535 + ], + "type": "text", + "content": "[41] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond mnns: Positional self-attention with co-attention for video question answering. In AAAI, volume 33, pages 8658–8665, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[42] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "text", + "content": "[43] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint arXiv:2107.13586, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 658 + ], + "type": "text", + "content": "[44] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands, too. arXiv preprint arXiv:2103.10385, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[45] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[46] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In CVPR, pages 9879-9889, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 171 + ], + "type": "text", + "content": "[47] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, pages 2630-2640, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 207 + ], + "type": "text", + "content": "[48] Jonghwan Mun, Paul Hongsuck Seo, Ilchae Jung, and Bohyung Han. Marioqa: Answering questions by watching gameplay videos. In ICCV, pages 2867-2875, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "type": "text", + "content": "[49] Akihiro Nakamura and Tatsuya Harada. Revisiting fine-tuning for few-shot learning. arXiv preprint arXiv:1910.00216, 2019.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "text", + "content": "[50] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 297, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 331 + ], + "type": "text", + "content": "[51] Jungin Park, Jiyoung Lee, and Kwanghoon Sohn. Bridge to answer: Structure-aware graph interaction network for video question answering. In CVPR, pages 15526-15535, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "type": "text", + "content": "[52] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, pages 2085–2094, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[53] Matthew E Peters, Sebastian Ruder, and Noah A Smith. To tune or not to tune? adapting pretrained representations to diverse tasks. arXiv preprint arXiv:1903.05987, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[54] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and Sebastian Riedel. Language models as knowledge bases? arXiv preprint arXiv:1909.01066, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "text", + "content": "[55] Nina Poerner, Ulli Waltinger, and Hinrich Schütze. Bert is not a knowledge base (yet): Factual knowledge vs. name-based reasoning in unsupervised qa. arXiv preprint arXiv:1911.03681, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 489, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 544 + ], + "type": "text", + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[57] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "type": "text", + "content": "[58] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 656 + ], + "type": "text", + "content": "[59] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[60] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. arXiv preprint arXiv:2112.01518, 2021.3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13954" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[61] Mengye Ren, Ryan Kiros, and Richard Zemel. Exploring models and data for image question answering. Advances in neural information processing systems, 28, 2015. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 106, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 106, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 288, + 139 + ], + "type": "text", + "content": "[62] Arka Sadhu, Kan Chen, and Ram Nevatia. Video question answering with phrases via semantic roles. arXiv preprint arXiv:2104.03762, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 140, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 288, + 172 + ], + "type": "text", + "content": "[63] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In CVPR, pages 16877-16887, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 172, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 172, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 49, + 172, + 288, + 216 + ], + "type": "text", + "content": "[64] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 217, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 288, + 250 + ], + "type": "text", + "content": "[65] Xiaomeng Song, Yucheng Shi, Xin Chen, and Yahong Han. Explore multi-step reasoning in video question answering. In ACM MM, pages 239-247, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 250, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 250, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 49, + 250, + 288, + 293 + ], + "type": "text", + "content": "[66] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, pages 7464-7473, 2019. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 293, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 293, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 49, + 293, + 288, + 338 + ], + "type": "text", + "content": "[67] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In CVPR, pages 4631-4640, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 338, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 338, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 338, + 288, + 380 + ], + "type": "text", + "content": "[68] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 382, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 382, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 49, + 382, + 288, + 425 + ], + "type": "text", + "content": "[69] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. NeurlPS, 34:200-212, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 426, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 426, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 49, + 426, + 288, + 459 + ], + "type": "text", + "content": "[70] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 459, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 459, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 49, + 459, + 288, + 502 + ], + "type": "text", + "content": "[71] Wenhui Wang, Hangbo Bao, Li Dong, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. arXiv preprint arXiv:2111.02358, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 503, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 503, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 49, + 503, + 288, + 536 + ], + "type": "text", + "content": "[72] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In CVPR, pages 9777–9786, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 537, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 537, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 537, + 288, + 591 + ], + "type": "text", + "content": "[73] Hu Xu, Gargi Ghosh, Po-Yao Huang, Prahal Arora, Masoumeh Aminzadeh, Christoph Feichtenhofer, Florian Metze, and Luke Zettlemoyer. Vlm: Task-agnostic video-language model pre-training for video understanding. arXiv preprint arXiv:2105.09996, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 592, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 592, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 49, + 592, + 288, + 635 + ], + "type": "text", + "content": "[74] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. EMNLP, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 636, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 636, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 49, + 636, + 288, + 669 + ], + "type": "text", + "content": "[75] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, pages 5288-5296, 2016. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 288, + 712 + ], + "type": "text", + "content": "[76] Li Xu, He Huang, and Jun Liu. Sutd-trafficqa: A question answering benchmark and an efficient network for video reasoning over traffic events. In CVPR, pages 9878–9888, 2021. 2, 5, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 667 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[77] Hongyang Xue, Wenqing Chu, Zhou Zhao, and Deng Cai. A better way to attend: Attention with trees for video question answering. IEEE TIP, 27(11):5563-5574, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[78] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In ICCV, pages 1686-1697, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 151, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 185 + ], + "type": "text", + "content": "[79] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "text", + "content": "[80] Zekun Yang, Noa Garcia, Chenhui Chu, Mayu Otani, Yuta Nakashima, and Haruo Takemura. Bert representations for video question answering. In WACV, pages 1556-1565, 2020. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 229, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 545, + 274 + ], + "type": "text", + "content": "[81] Yunan Ye, Zhou Zhao, Yimeng Li, Long Chen, Jun Xiao, and Yueting Zhuang. Video question answering via attribute-augmented attention network learning. In ACM SIGIR, pages 829-832, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 275, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 330 + ], + "type": "text", + "content": "[82] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In CVPR, pages 16375-16387, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "text", + "content": "[83] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: Multimodal neural script knowledge models. NeurIPS, 34:23634-23651, 2021. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 375, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 545, + 409 + ], + "type": "text", + "content": "[84] Zheng-Jun Zha, Jiawei Liu, Tianhao Yang, and Yongdong Zhang. Spatiotemporal-textual co-attention network for video question answering. ACM TOMM, 15(2s):1-18, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "type": "text", + "content": "[85] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. In CVPR, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 454, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 545, + 487 + ], + "type": "text", + "content": "[86] Renrui Zhang, Longtian Qiu, Wei Zhang, and Ziyao Zeng. Vt-clip: Enhancing vision-language models with visual-guided texts. arXiv preprint arXiv:2112.02399, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 488, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 520 + ], + "type": "text", + "content": "[87] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint arXiv:2112.01071, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "text", + "content": "[88] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 555, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 589 + ], + "type": "text", + "content": "[89] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 590, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 622 + ], + "type": "text", + "content": "[90] Linchao Zhu and Yi Yang. Actbert: Learning global-local video-text representations. In CVPR, pages 8746-8755, 2020, 2, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[91] Yueting Zhuang, Dejing Xu, Xin Yan, Wenzhuo Cheng, Zhou Zhao, Shiliang Pu, and Jun Xiao. Multichannel attention refinement for video question answering. ACM TOMM, 16(1s):1-23, 2020. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13955" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/203e5250-94d9-4833-ba8d-b9c548251baa_content_list.json b/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/203e5250-94d9-4833-ba8d-b9c548251baa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f6787075e4eb5efb8ce7f19dd5cc0ef0e63c4541 --- /dev/null +++ b/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/203e5250-94d9-4833-ba8d-b9c548251baa_content_list.json @@ -0,0 +1,1862 @@ +[ + { + "type": "text", + "text": "Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction", + "text_level": 1, + "bbox": [ + 179, + 130, + 794, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hatef Otroshi Shahreza1,2 and Sébastien Marcel1,3", + "bbox": [ + 285, + 202, + 683, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Idiap Research Institute, Martigny, Switzerland", + "bbox": [ + 294, + 224, + 676, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland", + "bbox": [ + 186, + 243, + 784, + 260 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Université de Lausanne (UNIL), Lausanne, Switzerland", + "bbox": [ + 261, + 261, + 709, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hatef.otoreshi,sebastien.marcel}@idiap.ch", + "bbox": [ + 303, + 282, + 665, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/035c2bb9e1ceeb4117f61cbc86a31e5035a4f7fd586f85e37b75393285451c54.jpg", + "image_caption": [ + "Original" + ], + "image_footnote": [], + "bbox": [ + 101, + 318, + 212, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cecc708ec233d620428171b9e000ebf21ccaf531a092acf9d983fc467f701f54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 318, + 325, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ecc70f9ac35c83befec4b24257dafcf95775a163cae8da28ad74d8cc975475ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 319, + 437, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7727bb2f6da68c3858ae4978efe8b4b6bf8ef6b88c91ce352a598038377dec9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 319, + 550, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d2ec5ae4f1bbe82a80c41e6fbdc8eba56146f333aa1ea41ac9fcc2a86820835.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 552, + 319, + 661, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7c94f7064769039f3dbac5dd309f1f60960fb6f03a4c26d87f8e5bf38329dbb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 319, + 774, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/349efb905bf85293b34164fcbee27fab5a57ab20de571a4d08f86b6c01f0fac0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 319, + 888, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a75ab9e6fdf7ff4bdfd80452416c75e18bb050109458d3fbea5500d29b8e0d3.jpg", + "image_caption": [ + "Frontal Reconst." + ], + "image_footnote": [], + "bbox": [ + 101, + 407, + 210, + 492 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/181fc30daa5b5d459e824179e0c1aead580a31165d5fe7185fd661739e613fd1.jpg", + "image_caption": [ + "0.674" + ], + "image_footnote": [], + "bbox": [ + 214, + 407, + 323, + 492 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01aaeec64460585df091b5029d76fee21ded327899cf99a42e46e72b336fe15d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 407, + 437, + 492 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c9df9e6cc6054fed9b6a42a50849d994005c041ccdbee0108b9316701ac9f351.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 407, + 549, + 492 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d662cba4694d056de21d1b95807f5ff3742cc7f41d6bad0d8b83c24d99568b14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 552, + 407, + 661, + 492 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e8bfa93a90e7700d64587a8680a2c84118a9779a036e1e61712d19a0c4314dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 407, + 774, + 492 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62db6d5e3e9c9ce1ff792919efb9aeebccba7ac3365a5163a7db481c2ad93048.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 407, + 888, + 492 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d831fc20456ae67f432db5e803c609ea7b426e05b9ccf179539b88ae262503c1.jpg", + "image_caption": [ + "3D Reconst.", + "0.706" + ], + "image_footnote": [], + "bbox": [ + 101, + 512, + 210, + 593 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d1266d882582bc90a7a3818750f5473eebcf87aeb0aae729916650ce16ba9b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 214, + 513, + 321, + 593 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f886986be8b1e674a87979042bc046fc91daeb81de1a09e77cca9b0c672fa800.jpg", + "image_caption": [ + "0.705" + ], + "image_footnote": [], + "bbox": [ + 326, + 513, + 429, + 593 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4745ef3c430bd2398cdcabe4fcafedb0f6bbbf53a6675e563fec03a91057f7f6.jpg", + "image_caption": [ + "0.679" + ], + "image_footnote": [], + "bbox": [ + 439, + 513, + 544, + 593 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/25807260325bf9ec372cb3cc2cc2c7628f18c5a2c929ca7350429094e965f8a3.jpg", + "image_caption": [ + "0.663" + ], + "image_footnote": [], + "bbox": [ + 558, + 512, + 661, + 590 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d8dc0f040695b7a4c8a2e11c99d46f312d81b13878c00570d2be3f908884f67.jpg", + "image_caption": [ + "0.733" + ], + "image_footnote": [], + "bbox": [ + 666, + 512, + 774, + 590 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2af1d830e444c8df0aad2aab90167008b473d189f05594a80d95724162c1b0a7.jpg", + "image_caption": [ + "0.782", + "Figure 1: Sample face images from the FFHQ dataset (first row) as well as their corresponding 3D (third row) and frontal 2D reconstruction (second row) from facial templates in the whitebox template inversion attack against ArcFace. Values show the cosine similarity between the templates of the original and frontal reconstructed face images." + ], + "image_footnote": [], + "bbox": [ + 784, + 512, + 888, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 655, + 313, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face recognition systems are increasingly being used in different applications. In such systems, some features (also known as embeddings or templates) are extracted from each face image. Then, the extracted templates are stored in the system's database during the enrollment stage and are later used for recognition. In this paper, we focus on template inversion attacks against face recognition systems and introduce a novel method (dubbed GaFaR) to reconstruct 3D face from facial templates. To this end, we use a geometry-aware generator network based on generative neural radiance fields (GNeRF), and learn a mapping from facial templates to the intermediate latent space of the generator network. We train our network with a semi-supervised learning approach using real and synthetic images simul", + "bbox": [ + 73, + 689, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "taneously. For the real training data, we use a Generative Adversarial Network (GAN) based framework to learn the distribution of the latent space. For the synthetic training data, where we have the true latent code, we directly train in the latent space of the generator network. In addition, during the inference stage, we also propose optimization on the camera parameters to generate face images to improve the success attack rate (up to $17.14\\%$ in our experiments). We evaluate the performance of our method in the whitebox and blackbox attacks against state-of-the-art face recognition models on the LFW and MOBIO datasets. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates. The project page is available at: https://www.idiap.ch/paper/gafar", + "bbox": [ + 496, + 656, + 893, + 869 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19662", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 89, + 207, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Automatic face recognition (FR) has become a well-known biometric authentication tool which has been widely used in different applications, including smartphone locks, border controls, etc. In such systems, generally, some features (also referred to as facial embeddings or facial templates) are extracted from users and stored in the system's database during the enrollment stage. In the recognition stage, either verification or identification, similar features are extracted from the user and are compared with the features in the system's database. Therefore, facial features play the main role in automatic face recognition systems and convey important information about users.", + "bbox": [ + 75, + 114, + 467, + 295 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Template inversion (TI) attack against FR systems refers to an attempt by an adversary to reconstruct face images from the templates stored in the system's database. Then, the adversary may find sensitive information about the enrolled users and also can use the reconstructed face image to impersonate and enter the system. Therefore, compared to most attacks on FR systems which threaten the security of the system [19, 3, 22, 18, 35, 34, 10], TI attack jeopardizes both security and privacy of users, and thus requires further study. In this paper, we introduce a novel method (called geometry-aware face reconstruction, shortly GaFaR) for the TI attack in FR systems to reconstruct a 3D face from facial features (extracted from 2D FR models). The 3D face reconstruction provides further information than 2D face reconstruction, and in particular can be used to generate face image from any pose to improve the attack to the FR system. To our knowledge, this is the first paper on 3D face reconstruction from facial templates.", + "bbox": [ + 75, + 296, + 467, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, neural radiance fields (NeRF) [38] absorbed considerable attention in computer vision society due to remarkable results in the novel-view synthesis problem. Built upon NeRF, generative NeRF (GNeRF) methods such as [37, 44, 56, 6, 8, 41, 42, 20, 55, 7, 12, 48, 43] combine conditional NeRF with generative models for geometry-aware image generation tasks. In these methods, a generative model, such as a generative adversarial network (GAN), is used to embed the shape and appearance of an object in a latent space. Then, the latent code of the GAN, along with the camera parameters, are fed to a NeRF for the rendering process. Among GNeRF methods, there are several works, such as [8, 41, 42, 20, 55, 7, 12, 48], for geometry-aware 3D face generation, which can generate face images from different views.", + "bbox": [ + 75, + 568, + 467, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our proposed TI method, we use a geometry-aware face generator network based on GNeRF, and learn a new mapping from facial templates to the intermediate latent space of the generator network. We train our network using real and synthesized images simultaneously with a semisupervised learning approach. For real training data where we do not have the corresponding latent code (i.e., unsuper", + "bbox": [ + 75, + 795, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vised), we use a GAN-based framework to learn the distribution of the latent space. For the synthetic training data where we have true values of latent codes (i.e., supervised), we directly learn the latent space. Because we have 3D reconstructed face, we can use any arbitrary pose to inject the the FR system for the attack. Therefore, during the inference stage, we also use optimization on the camera parameters to improve the attack against the FR system. We evaluate the performance of our proposed method in the whitebox (i.e., where the adversary knows the internal functioning and parameters of the FR model) and blackbox (i.e., where the adversary does not have information about the internal functioning of the FR model) attacks against state-of-the-art (SOTA) FR systems. For evaluation, we consider FR systems with the LFW [23] and MOBIO [36] datasets and evaluate our attack if the reconstructed face can be used to enter the system. Fig. 1 illustrates sample face images from FFHQ [26] dataset and their 3D reconstruction from ArcFace [11] templates using our attack.", + "bbox": [ + 496, + 90, + 890, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To elaborate on the contributions of this paper, we list them hereunder:", + "bbox": [ + 500, + 378, + 890, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel template inversion method to reconstruct 3D faces from the facial templates of a face recognition system. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates.", + "- We use a geometry-aware generator network based on GNeRF, and learn a mapping from facial templates to the intermediate latent space of the generator network. We train our mapping network using a GAN-based framework and a semi-supervised learning approach using real and synthetic face images.", + "- We use optimization on the camera parameters in the geometry-aware generator network during the inference stage. The optimization of camera parameters can find a pose that improves the success attack rate." + ], + "bbox": [ + 517, + 422, + 890, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The remainder of the paper is organized as follows. First, we review the related works in the literature in Sec. 2. Then, we present our proposed method in Sec. 3. Next, we describe our experiments in Sec. 4 and discuss limitations and ethical considerations in Sec. 5. Finally, the paper is concluded in Sec. 6.", + "bbox": [ + 496, + 691, + 890, + 781 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 797, + 648, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generally, the methods in the literature for reconstructing face from facial templates generate 2D face images. In [57], authors proposed two methods based on optimization and learning to reconstruct 2D low-resolution face images from facial templates in the whitebox scenario. In the", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19663", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/a51079498b56fde9a955061d0962f1c4a27e8e798db5eb3510e915f9569bebc6.jpg", + "table_caption": [ + "Table 1: Comparison with related works." + ], + "table_footnote": [ + "*The method is based on the whitebox attack, and is also applied in the blackbox scenario by removing a loss term that required the FR model.", + "The method is based on the whitebox attack, and is extended to the blackbox with knowledge distillation of the FR model.", + "$\\ddagger$ The method is based on the whitebox attack, and is extended to blackbox using a different FR model." + ], + "table_body": "
Ref.2D/3DResolutionWhitebox/BlackboxMethod BasisAvailable code
[57]2Dlowwhitebox1) optimization\n2) learningX
[9]2Dlowboth*learningX
[33]2Dlowblackboxlearning✓
[16]2Dlowboth†learningX
[49]2Dlowboth†learningX
[1]2Dlowblackboxlearning + opt.X
[14]2Dhighblackboxlearning✓
[51]2Dhighblackboxoptimization✓
[15]2DhighblackboxoptimizationX
[Ours]3Dhighboth‡learning✓
", + "bbox": [ + 81, + 108, + 475, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "optimization-based method, they used a gradient ascent-based algorithm to reconstruct the face image, and optimized the generated image to minimize a multi-term loss function, including the $\\ell_2$ distance between target templates and templates of the reconstructed face image. In addition, they applied total-variation and Laplacian pyramid gradient normalization [5] on the reconstructed image to generate a smooth image. In their learning-based method, they used a deconvolutional neural network to reconstruct face images and trained it with the same loss function that they used in their optimization-based method.", + "bbox": [ + 75, + 364, + 468, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In [9], a multi-layer perceptron (MLP) is trained to find facial landmark coordinates (optimized using the mean squared error) from the given template. They also trained a convolutional neural network (CNN) to generate face texture (optimized using mean absolute error) from the target template. Then, they used a differentiable warping to combine estimated landmarks (from MLP) and textures (from CNN) and reconstruct low-resolution face images. In the whitebox scenario, they further optimized MLP and CNN by minimizing the distance between templates of reconstructed and original face images. However, in the blackbox scenario, they trained MLP and CNN separately, and used warping in the inference only.", + "bbox": [ + 75, + 534, + 468, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In [33], two new deconvolutional networks are proposed, called NbBlock-A and NbBlock-B, and are trained with pixel loss $(\\ell_{1}$ norm of reconstruction error) and perceptual loss (distance of middle layers of VGG-19[46] when given the reconstructed and original face images) to reconstruct low-resolution face images in the blackbox scenario. In [16] and [49], a same bijection-learning-based method is used to train GAN models with PO-GAN [25] and TransGAN [24] structure, respectively. The method is based on whitebox attack and for training the GAN model, authors also minimized the distance between target templates and templates", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "extracted from the reconstructed face images using the FR model. In the blackbox attack, they proposed to use the distillation of knowledge to train a student network that mimics the target FR model and used the student network in their method. However, they did not report any detail (and no published source code) about the training of the student network (e.g., network structure, etc.).", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In [1], a 3-step method is proposed to reconstruct low-resolution face images in the blacbox scenario. First, they trained a GAN for general face generation. In the second step, they trained a MLP to map target templates to embeddings of a known FR model. In the final step, they found a latent code in the input of their generator (of their GAN) which generates a face image that maximizes two terms; the discriminator score (for being a real face image) and cosine similarity between the mapped embedding and the embeddings extracted by the known FR model.", + "bbox": [ + 496, + 198, + 890, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to most works in the literature which generate low-resolution face images, recently there have been few works to generate high-resolution 2D face images. In [51], a grid-search optimization using the simulated annealing [50] approach is used on the latent vector (i.e., input noise) of StyleGAN2 [28] to find latent codes that can generate face images which have templates similar to the target templates. However, their proposed method is computationally expensive, and they reported their evaluation on only 20 face images. In [15], a similar optimization to [51] on the latent vector of StyleGAN2 [28] is considered, but it is solved using the standard genetic algorithm [47]. In contrast to [51, 15] which are based on optimization, in [14], a learning-based method is proposed in the blackbox scenario. The authors proposed to generate some face images using StyleGAN2 [28] and extract the templates using the FR model. Then, they trained a MLP to map facial templates to the input latent codes of StyleGAN2 [28].", + "bbox": [ + 496, + 348, + 892, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tab. 1 compares our proposed method with the previous works in the literature. To our knowledge, our proposed method is the first work on 3D face reconstruction from facial templates (extracted from 2D face recognition models). Furthermore, unlike most works in the literature, our method generates high-resolution face images. Last but not least, our method can be used in both whitebox and black-box attacks against FR systems.", + "bbox": [ + 496, + 619, + 890, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 753, + 671, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we consider a threat model as described in Sec. 3.1 and use the proposed 3D face reconstruction method (dubbed GaFaR) described in Sec. 3.2. After training and in the inference stage, we use an optimization on the camera parameters as described in Sec. 3.3 to generate a 2D face image with a pose that can lead to a higher success attack rate (SAR) and inject the face image into the system. Fig. 2 depicts the block-diagram of the proposed method.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19664", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Threat Model", + "text_level": 1, + "bbox": [ + 76, + 90, + 220, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider an attack against FR systems, where the adversary gains access to the system's database and aims to generate a 3D reconstruction of enrolled users. To this end, the adversary trains a network to reconstruct a 3D face given a target facial template. Then, the adversary can use the 3D reconstruction to impersonate users into the system (i.e., using a 3D face mask, etc.). For simplicity, we consider that the adversary can inject a 2D face image (from the 3D reconstructed face) into the system. Hence, the adversary uses an appropriate pose to generate a 2D face image for the attack. According to the adversary's knowledge of the FR model, we consider both whitebox and blackbox attacks against FR systems. In the whitebox scenario, we assume that the adversary has complete knowledge of the FR model from which the templates are leaked. However, in the blackbox scenario, we assume that the adversary does not know the internal functioning of the FR model and can only use the FR model to generate facial template for any face image. In addition, we assume that in the blackbox scenario, the adversary has access to another FR model and knows its internal parameters and functioning.", + "bbox": [ + 75, + 114, + 473, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. 3D Face Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 449, + 295, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To reconstruct 3D faces from facial templates, we consider a geometry-aware face generator network based on GNeRF, such as EG3D [7], and learn a mapping $M_{\\mathrm{GaFaR}}: \\mathcal{T} \\to \\mathcal{W}$ from the facial templates $t \\in \\mathcal{T}$ to the intermediate latent space $\\mathcal{W}$ of the GNeRF model. Then, we use the remainder of GNeRF model $G(.)$ to synthesize an image $\\hat{I} = G(\\hat{w}, c)$ from an arbitrary view using the mapped intermediate latent vector $\\hat{w}$ and camera parameters $c$ . We consider a network with two fully-connected layers with the Leaky ReLU activation functions for our mapping network $M_{\\mathrm{GaFaR}}$ .", + "bbox": [ + 75, + 474, + 468, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train the mapping network $M_{\\mathrm{GaFaR}}$ with a semi-supervised approach using synthetic and real data. For the real data, we consider a set of real face images $\\{I_{\\mathrm{real},i}\\}_{i = 0}^{N}$ and extract the facial template $t_{\\mathrm{real},i} = F(I_{\\mathrm{real},i})$ from each face image $I_{\\mathrm{real},i}$ using the FR model $F(.)$ . As we do not have the true value of intermediate latent space $\\mathcal{W}$ of the GNeRF model to generate the real face image, we consider training our mapping network using this data as unsupervised learning. On the other hand, for the synthetic data, we use the pre-trained geometry-aware face generator network to generate a set of random face images $\\{I_{\\mathrm{syn},i}\\}_{i = 0}^{M}$ . Therefore, as opposed to real data, we have the true value of intermediate latent space $\\boldsymbol {w}\\in \\mathcal{W}$ to generate the same synthetic face image. Hence, we consider training our mapping network using the synthetic data as supervised learning. We train our mapping network simultaneously using real and synthetic training data as follows:", + "bbox": [ + 75, + 643, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unsupervised learning using real data To train our mapping network $M_{\\mathrm{GaFaR}}(.)$ with the real data, we use a GAN-based framework based on Wasserstein GAN (WGAN) [2] algorithm to learn the distribution of intermediate latent space $\\mathcal{W}$ of the GNeRF model. In this framework, our mapping network $M_{\\mathrm{GaFaR}}$ generates a latent code $\\hat{\\pmb{w}} = M_{\\mathrm{GaFaR}}([n,t])$ using the facial template $\\pmb{t}$ and a random vector $\\pmb{n} \\in \\mathcal{N}$ . We can also generate real latent codes $\\pmb{w} = M(\\pmb{z}) \\in \\mathcal{W}$ for training our GAN using the GNeRF mapping function $M$ and a random vector $\\pmb{z} \\in \\mathcal{Z}$ . Then, we can use a critic network $C(.)$ to score the latent codes generated by our mapping $M_{\\mathrm{GaFaR}}(.)$ and GNeRF mapping $M(.)$ . Therefore, we can train the critic network $C(.)$ and our mapping $M_{\\mathrm{GaFaR}}$ using the following loss functions:", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C} ^ {\\mathrm {W G A N}} = \\mathbb {E} _ {\\boldsymbol {w} \\sim M (\\boldsymbol {z})} [ C (\\boldsymbol {w}) ] - \\mathbb {E} _ {\\hat {\\boldsymbol {w}} \\sim M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, t ])} [ C (\\hat {\\boldsymbol {w}}) ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 327, + 890, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M _ {\\mathrm {G a F a R}}} ^ {\\mathrm {W G A N}} = \\mathbb {E} _ {\\hat {\\boldsymbol {w}} \\sim M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, \\boldsymbol {t} ])} [ C (\\hat {\\boldsymbol {w}}) ] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 363, + 890, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the WGAN training, we use the generated face image $\\hat{\\pmb{I}} = G(M_{\\mathrm{GaFaR}}([n,t]),c)$ to optimize our mapping network using the following loss function:", + "bbox": [ + 496, + 393, + 890, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e a l}} ^ {\\text {r e c}} = \\mathcal {L} ^ {\\text {P i x e l}} + \\mathcal {L} ^ {\\text {I D}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 448, + 890, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}^{\\mathrm{Pixel}}$ and $\\mathcal{L}^{\\mathrm{ID}}$ are pixel loss and ID loss, respectively, and are defined as:", + "bbox": [ + 496, + 477, + 890, + 507 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {P i x e l}} = \\mathbb {E} _ {\\hat {\\boldsymbol {w}} \\sim M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, \\boldsymbol {t} ])} \\left[ \\| \\boldsymbol {I} - G (\\hat {\\boldsymbol {w}}, \\boldsymbol {c}) \\| _ {2} ^ {2} \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 518, + 890, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {I D}} = \\mathbb {E} _ {\\hat {\\boldsymbol {w}} \\sim M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, \\boldsymbol {t} ])} [ \\| F _ {\\text {l o s s}} (\\boldsymbol {I}) - F _ {\\text {l o s s}} (G (\\hat {\\boldsymbol {w}}, \\boldsymbol {c})) \\| _ {2} ^ {2} ] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 540, + 890, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pixel loss minimizes the pixel-level reconstruction error and ID loss helps the network to generate face images with a similar facial templates extracted by $F_{\\mathrm{loss}}$ . To calculate the ID loss in Eq. (5), we use the same FR model as the one by which leaked templates are calculated (i.e., $F$ ) in the whitebox scenario. However, in the blackbox scenario, we use another FR model that we assume the adversary has access to. The critic network is a fully-connected network with three layers and Leaky ReLU activation function.", + "bbox": [ + 496, + 584, + 890, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Supervised learning using synthetic data To train our mapping network $M_{\\mathrm{GaFaR}}(.)$ with synthetic data, we can directly learn the GNeRF intermediate latent code $\\pmb {w} = M(z)$ . In addition to directly learning $\\pmb{w}$ , we use the generated face image to train our mapping network by minimizing the following multi-term loss function:", + "bbox": [ + 496, + 738, + 890, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s y n}} ^ {\\text {r e c}} = \\mathcal {L} ^ {w} + \\mathcal {L} ^ {\\text {P i x e l}} + \\mathcal {L} ^ {\\text {I D}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 838, + 890, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}^{\\mathrm{Pixel}}$ and $\\mathcal{L}^{\\mathrm{ID}}$ are pixel loss and ID loss, respectively. Furthermore, $\\mathcal{L}^w$ is $w$ -loss to directly learn the latent space", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19665", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9f43988b9d9f5cd9ef97505a35d855d9e3c18c0d20c345fd25302e2beff1e390.jpg", + "image_caption": [ + "Training (Semi-supervised Learning)" + ], + "image_footnote": [], + "bbox": [ + 91, + 109, + 803, + 276 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2c97c30b32e73f35907f3fc56a205f1323af9a9debc6756cec69c57e410bcc82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 282, + 870, + 419 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/78af3af4a1872df784c28a6ce55670a117fdd4278f577e1dac6a8a9e90e4b7b5.jpg", + "image_caption": [ + "Inference", + "Figure 2: Block-diagram of the proposed method: In the training process, a semi-supervised approach is used to learn our mapping $M_{\\mathrm{GaFaR}}$ (depicted with a green block) from the facial templates to the intermediate latent space of the GNeRF model using real data (unsupervised, where we don't have the corresponding latent code) and synthetic data (supervised, where we have the corresponding latent code $\\boldsymbol{w}$ ), simultaneously. During the inference stage, the leaked template $t$ is given to our mapping network to find corresponding latent code $\\hat{\\boldsymbol{w}} = M_{\\mathrm{GaFaR}}([n, t])$ . Then, $\\hat{\\boldsymbol{w}}$ along with camera parameters $c$ are given to the generator and renderer of GNeRF $G$ to generate a reconstructed face image $\\hat{I} = G(\\hat{\\boldsymbol{w}}, c)$ . Finally, an optimization (either grid search or continuous optimization) is applied on two of the camera parameters, $\\theta$ and $\\psi$ , from $c$ , to find the best pose that minimizes the distance between leaked template $t$ and the template of reconstructed face image." + ], + "image_footnote": [], + "bbox": [ + 112, + 455, + 890, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by minimizing mean squared error between $\\pmb{w}$ and $\\hat{\\pmb{w}} = M_{\\mathrm{GaFaR}}([n,t])$ as follows:", + "bbox": [ + 76, + 763, + 470, + 795 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {w} = \\mathbb {E} _ {\\boldsymbol {w} \\sim M (\\boldsymbol {z})} [ \\| \\boldsymbol {w} - M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, \\boldsymbol {t} ]) \\| _ {2} ^ {2} ] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 804, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Camera Parameters Optimization", + "text_level": 1, + "bbox": [ + 76, + 832, + 375, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the inference stage, after generating a 3D reconstruction of face using our method described in Sec. 3.2, we optimize the camera parameters to find a pose that increases", + "bbox": [ + 76, + 854, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the SAR. Among different camera parameters $c$ , the parameters which correspond to the camera rotation can change the pose of the generated face image. Note that by varying the camera rotations, we aim to change the pitch and yaw rotations of the 3D reconstructed face and do not modify the roll rotation. In fact, the roll rotation of the face will be eliminated through face alignment as a preprocess prior to feature extraction in the FR system. To optimize camera parameters, we consider two different approaches as follows:", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19666", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Grid Search (GS) For the grid search, we consider predefined steps to change the camera pitch $\\theta \\in \\Theta$ and yaw $\\psi \\in \\Psi$ rotations of camera parameters $c$ . We generate the 2D face image for all values of camera rotation steps ( $\\theta_{\\mathrm{step}}$ and $\\psi_{\\mathrm{step}}$ ) and find the facial template for each generated image. Then, we select the face image $\\hat{\\pmb{I}} = G(M_{\\mathrm{GaFaR}}([n,t]),c)$ that has a template $\\hat{\\pmb{t}} = F(\\hat{\\pmb{I}})$ with the minimum mean squared error with the target template $\\pmb{t}$ . Note that the grid search can be applied in both whitebox and blackbox scenarios using the FR model $F$ .", + "bbox": [ + 76, + 90, + 472, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Continuous Optimization (CO) In our continuous optimization approach, we start from the frontal image and use Adam [30] optimizer to solve the following minimization using $\\hat{\\boldsymbol{w}} = M_{\\mathrm{GaFaR}}([\\boldsymbol{n},\\boldsymbol{t}])$", + "bbox": [ + 76, + 262, + 470, + 324 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta , \\psi} \\| F (G (\\hat {w}, c)), t \\| _ {2} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 330, + 468, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this optimization, we find the parameters $\\theta$ and $\\psi$ for camera rotation that lead to a face image with templates close to the target template $t$ . In contrast to the grid search approach, the continuous optimization can be applied only in the whitebox scenario.", + "bbox": [ + 75, + 366, + 468, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 455, + 209, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 479, + 266, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Face Recognition models In our experiments, we consider different SOTA FR models including ArcFace [11], ElasticFace [4] as well as four different FR models with SOTA backbones from FaceX-Zoo [53], including AttentionNet [52], HRNet [54], RepVGG [13] and Swin [32]. The recognition performance of these models are reported in the supplementary material.", + "bbox": [ + 75, + 503, + 468, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets All the mentioned FR models are trained on the MS-Celeb1M dataset [21]. To train the face reconstruction network, we assume that adversary does not have any knowledge of the dataset used in training FR model, and uses a different dataset. To this end, we consider the Flickr-Faces-HQ (FFHQ) dataset [27], which consists of 70,000 high-quality images crawled from internet (with no identity label), for training our face reconstruction network. We randomly split FFHQ into train $(90\\%)$ and test $(10\\%)$ set.", + "bbox": [ + 75, + 625, + 468, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the vulnerability of FR models, we use two other datasets, including Labeled Faces in the Wild (LFW) [23] and MOBIO [36] datasets. The LFW dataset consists of 13,233 face images of 5,749 people collected from internet, where 1,680 people have two or more images. The MOBIO dataset includes face images of 150 people captured using mobile devices in 12 sessions for each person1.", + "bbox": [ + 75, + 760, + 470, + 881 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f24a996c9116d06dcf1064e78bab90b4b34d8ae47041e82063a5d6a8f834af9.jpg", + "image_caption": [ + "rnnnne" + ], + "image_footnote": [], + "bbox": [ + 532, + 87, + 617, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c707859e073481a45313f586e169e175f0760c39725f52398a9790b8ce0bb33f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 87, + 705, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb4b809289f0232e9642d9ba5ef6b2b1c1505cf40bb9de299b37e6d986b38c49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 87, + 795, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e9f98b45067148fedd9851fde2af3188b7cb5eeffe1223495452bb80f76a66b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 88, + 885, + 154 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/10f20f965b6c96984ed8ed638edb789860477b0768a37fa3266f4d4c9c936255.jpg", + "image_caption": [ + "Reconnst", + "0.695", + "Figure 3: Sample face images from the FFHQ dataset (first row) and their corresponding frontal (second row) reconstructed face images using our method in the blackbox attack against ElasticFace using ArcFace for $F_{\\mathrm{loss}}$ . The values show the cosine similarity between templates of original and frontal reconstructed face images." + ], + "image_footnote": [], + "bbox": [ + 531, + 156, + 616, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/50def58c439c1660b2904b3edf74d78dcd9102a8ee6f86f3e39473f9bc45a74f.jpg", + "image_caption": [ + "0.721" + ], + "image_footnote": [], + "bbox": [ + 620, + 156, + 705, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7541912d6a86061a538840da7c897aa8959125de47b879f4b2eba1c74cac88a0.jpg", + "image_caption": [ + "0.699" + ], + "image_footnote": [], + "bbox": [ + 710, + 156, + 795, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e984cb56fba50a0634111494671e4142c7a5b44dfc3107a7a020d93fe9f80500.jpg", + "image_caption": [ + "0.732" + ], + "image_footnote": [], + "bbox": [ + 800, + 156, + 883, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocol For evaluation with each of the LFW and MOBIO datasets, we build a separate FR system, and use the reference templates (i.e., enrolled in the system's database) as input to our face reconstruction method. Then, we inject the reconstructed face image as a query to the system and evaluate the adversary's success attack rate (SAR) to enter the FR system. We should note that for each image, we use only one reconstructed face image as a query to the FR system and evaluate the SAR.", + "bbox": [ + 496, + 359, + 890, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details and Source Code In our experiments, we use the grid search (in both whitebox and black-box attacks) and continuous optimization (in whitebox attacks only) to optimize camera parameters, as described in Sec. 3.3, in the inference stage. For grid search, we consider $\\psi \\in [-45^{\\circ}, +45^{\\circ}]$ and $\\theta \\in [-30^{\\circ}, +30^{\\circ}]$ for a $11 \\times 11$ grid with step sizes of $\\psi_{\\mathrm{step}} = 9^{\\circ}$ and $\\theta_{\\mathrm{step}} = 6^{\\circ}$ . In the continuous optimization approach, we use 121 iterations of Adam optimizer [30] with the learning rate of $10^{-2}$ . An ablation study on these hyperparameters and the corresponding execution times are reported in the supplementary material. All our models (for different FR models and different scenarios) are trained in 15 epochs (each takes around 2 days) on a system equipped with a single NVIDIA RTX 3090 GPU. For GNeRF, we used a pretrained model of EG3D [7] in our experiments and generate $512 \\times 512$ resolution face images. The input noise $z$ to GNeRF model has 512 dimensions and the noise $n$ in the input of our mapping network $M_{\\mathrm{GaFaR}}$ has 16 dimensions. For unsupervised learning using real face images (FFHQ), we assume that camera parameters for real face images as frontal view. The source code of our experiments are publicly available2.", + "bbox": [ + 496, + 523, + 892, + 858 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1The results on AgeDB [39] dataset is also reported in the supplimen-", + "bbox": [ + 94, + 886, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "tary material.", + "bbox": [ + 500, + 875, + 573, + 887 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2 https://www.idiap.ch/paper/gafar", + "bbox": [ + 519, + 887, + 699, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19667", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/eb3bfd3ec68da89d04f4c1ab3c7c11e30cb89a89aad0209e318baa3c3eaec2a7.jpg", + "table_caption": [ + "Table 2: Evaluation of blackbox attack against SOTA FR models at systems' FMR=10 $^{-3}$ on the LFW and MOBIO datasets in terms of success attack rate (SAR). For attacks using our proposed method, we use ArcFace and also ElasticFace as $F_{\\mathrm{loss}}$ to calculate the ID loss in Eq. (5). The values are in percentage." + ], + "table_footnote": [], + "table_body": "
LFWMOBIO
ArcFaceElsFaceAtt.NetHRNetRepVGGSwinArcFaceEls.FaceAtt.NetHRNetRepVGGSwin
NBNetA-M [33]4.3210.901.241.601.133.8202.380000
NBNetA-P [33]16.8326.980.661.445.729.704.7616.190.48014.297.14
NBNetB-M [33]10.9821.443.224.473.2111.231.903.803.337.143.338.57
NBNetB-P [33]40.2658.1616.2918.4215.2440.7615.2443.8131.9026.6723.8144.29
Dong et al. [14]13.2112.613.904.073.2212.383.338.1010.486.679.053.33
Vendrow and Vendrow [51]57.7053.0321.1218.859.6246.8429.0543.8127.1426.6720.9545.24
[Floss= Els.Face] GaFaR51.78-18.0711.6811.6347.1547.62-54.2943.3345.7170.48
[Floss= Els.Face] GaFaR + GS61.56-23.5617.2115.8254.0864.76-63.8155.2358.5782.38
[Floss= ArcFace] GaFaR-74.5433.5937.8025.4067.11-84.7672.3876.6772.8689.05
[Floss= ArcFace] GaFaR + GS-78.6738.4243.2729.8470.82-86.6280.0083.8073.3393.33
", + "bbox": [ + 78, + 140, + 890, + 294 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a42d7031d371c8b232bcb1f75d469457df974aeecf539a87287500d10c18ee61.jpg", + "table_caption": [ + "Table 3: Evaluation of whitebox attack against SOTA FR models at systems' FMR = 10 $^{-3}$ on the LFW and MOBIO datasets in terms of success attack rate (SAR). The values are in percentage." + ], + "table_footnote": [], + "table_body": "
LFWMOBIO
ArcFaceElsFaceAtt.NetHRNetRepVGGSwinArcFaceEls.FaceAtt.NetHRNetRepVGGSwin
GaFaR79.8462.3227.0031.8717.3374.0882.8681.4364.2971.4353.8194.76
GaFaR + GS82.5268.7132.4237.4819.8376.0685.2382.3870.4781.4259.5294.29
GaFaR + CO84.2571.6934.5339.5320.4477.3689.0584.2975.7181.4362.8695.71
", + "bbox": [ + 78, + 340, + 879, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Analyze", + "text_level": 1, + "bbox": [ + 76, + 439, + 176, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Blackbox Scenario To evaluate the proposed method in the blackbox attack against SOTA FR models, we consider ArcFace and also ElasticFace as $F_{\\mathrm{loss}}$ to calculate the ID loss in Eq. (5). Tab. 2 compares the performance of the proposed method with blackbox face reconstruction methods in the literature in terms of SAR for systems configured at false match rate (FMR) of $10^{-3}$ . Similar results for $\\mathrm{FMR} = 10^{-2}$ are reported in the supplementary material. As this table shows, the frontal reconstruction by our method (i.e., GaFaR) achieves superior performance than previous methods in the literature. Furthermore, camera parameter optimization (i.e., GaFaR+GS) improves the performance of our attack up to $17.14\\%$ compared to our frontal face reconstruction (i.e., GaFaR). Comparing the use of ArcFace and ElasticFace as $F_{\\mathrm{loss}}$ , the performance of attacks with the ArcFace model is better. This is due to the fact that ArcFace has a higher recognition accuracy than ElasticFace. Fig. 3 shows sample face images reconstructed from ElasticFace templates in the blackbox attack (using ArcFace for $F_{\\mathrm{loss}}$ ).", + "bbox": [ + 75, + 463, + 473, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Whitebox Scenario Tab. 3 reports the performance of the proposed method in the whitebox attack $^3$ in terms of SAR for FR systems configured at false match rate (FMR) of $10^{-3}$ . Similar results for $\\mathrm{FMR} = 10^{-2}$ are reported in the supplementary material. According to this table, all", + "bbox": [ + 75, + 787, + 470, + 864 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df1e69460daef7d142695376886637ae42161f20d045057589ed9e39f2fc87ad.jpg", + "image_caption": [ + "Figure 4: Sample face images from the FFHQ dataset (first row) and their corresponding frontal (second row) and camera parameters sub-grid (third row) reconstructed face images using our method in the whitebox attack against ArcFace. The values show the cosine similarity between templates of original and frontal reconstructed face images." + ], + "image_footnote": [], + "bbox": [ + 500, + 436, + 890, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "these FR models are highly vulnerable to our attack. Furthermore, the camera parameter optimization improves the performance of GaFaR. Comparing grid search with contin", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3The whitebox attack methods reported in Tab. 1 do not have available source code, and we could not reproduce their results.", + "bbox": [ + 76, + 875, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19668", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0a67217bd880fb6f4229176abcb09bc7db1752f1d715878f2cdf139528651861.jpg", + "table_caption": [ + "Table 4: Ablation study on the proposed semi-supervised learning approach and evaluation of loss terms in the white-box attack against ArcFace model in terms of success attack rate (SAR) on the LFW and MOBIO datasets. The SAR values are in percentage and for an attack without any camera parameter optimization (GS/CO)." + ], + "table_footnote": [], + "table_body": "
approachLoss FunctionsLFWMOBIO
FMR=10-2FMR=10-3FMR=10-2FMR=10-3
supervisedLrecsyn = Lw + LPixel + LID83.8069.46790.9682.38
Lrecsyn = Lw + LPixel31.7513.9243.818.57
Lrecsyn = Lw + LID0.860.3000
Lrecsyn = Lw33.6915.4332.389.52
unsupervisedLrecreal = LPixel + LID[without WGAN]0.440.1500
Lrecreal = LPixel + LID[with WGAN]67.7245.7670.4831.90
Lrecreal = LID[with WGAN]54.5130.8352.8619.52
Lrecreal = LPixel[with WGAN]2.210.4000
semi-supervisedEqs. (1) to (3) and (6)89.2779.8495.7182.86
", + "bbox": [ + 81, + 183, + 467, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "uous optimization, results show that with the same number of iterations continuous optimization achieves better performance. In addition, comparing these results with recognition performance of FR models (available in the supplementary material), we conclude that models with higher recognition accuracy are more vulnerable to our attack. Comparing results in Tab. 2 and Tab. 3, we observe that when we use ArcFace as $F_{\\mathrm{loss}}$ in our loss function, blackbox attacks achieve better results than whitebox attacks for most cases, which can be explained considering the superior recognition performance of ArcFace than other FR models. Fig. 4 shows sample face images reconstructed from ArcFace templates in the whitebox attack. This figure also presents a grid of reconstructed face images with different poses4.", + "bbox": [ + 75, + 375, + 472, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study We evaluate the effect of our semi-supervised learning approach in our proposed method compared to fully supervised learning and fully unsupervised learning approach. In each case, we evaluate the effect of each loss function too. Furthermore, for the fully unsupervised learning approach, we evaluate the effect of GAN learning which we used in our method. Tab. 4 reports our ablation study in whitebox attack against ArcFace model on the LFW and MOBIO datasets in terms of SAR at system's FMR of $10^{-2}$ and $10^{-3}$ . As these results show, the proposed semi-supervised approach achieves a better performance than fully supervised learning and fully unsupervised learning approaches. Furthermore, each of our loss terms has an important effect on the performance. In particular, using WGAN for real data, where we don't have the true value of intermediate latent codes, helps the network to", + "bbox": [ + 75, + 608, + 472, + 851 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8798e7321e78a0f6137cc0e68b20383a95ebb3729cf9aad6c5a7c0d4a4a670d3.jpg", + "image_caption": [ + "rannnnnne" + ], + "image_footnote": [], + "bbox": [ + 532, + 88, + 617, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7337d7108f4ee44f027eccc3aa8c24f62312e0e442758f71d5d60127df74f917.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 88, + 705, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f57540adc2b0420978fcf657260012b7b3cf1d24128dd36543888a965188075d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 88, + 795, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b927cab3b27e2f2aacfd11caaa983a6a24706d20e95730cfee34611914f8213.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 88, + 885, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/803cab6322cd82db265021e32e77709b1e57ecf566313ce1c4e2519526820914.jpg", + "image_caption": [ + "Reconstrn", + "0.160", + "Figure 5: Sample failure cases from the LFW dataset (first row) and their corresponding frontal (second row) reconstructed face images using our method in the whitebox attack against ArcFace. The values below each image show the cosine similarity between templates of original and reconstructed face images." + ], + "image_footnote": [], + "bbox": [ + 531, + 156, + 616, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/50d2376b518d512e60618ff6a319cb0ffb8b76af926b318f5acdb9405c358d47.jpg", + "image_caption": [ + "0.150" + ], + "image_footnote": [], + "bbox": [ + 620, + 156, + 704, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04e7586916a2ad3d75191f701c0fe01cf55719767c7e17c01bb03ee9ea1308e4.jpg", + "image_caption": [ + "0.170" + ], + "image_footnote": [], + "bbox": [ + 710, + 156, + 794, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03a9aa130a829ca641bcf4f43d5a220c03e43fecb9d6986266616663472ad07c.jpg", + "image_caption": [ + "0.160" + ], + "image_footnote": [], + "bbox": [ + 800, + 156, + 883, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "learn the distribution of GNeRF intermediate latent space $\\mathcal{W}$ . Otherwise, the generated latent code by the mapping network will be out of distribution, and therefore generator part of GNeRF will not generate face-like images.", + "bbox": [ + 496, + 359, + 890, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 433, + 612, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations In spite of the considerable success attack rate of our method in our experiments, the reconstructed face images from some facial templates fail to attack the system. Fig. 5 illustrates sample failed cases in the whitebox attack against ArcFace. As these sample images show, the failed reconstructions reflect a bias in the generated faces for dark skin and also old people. Such a bias in final results can be caused by inherent biases of each of the datasets used to train the FR model, GNeRF model, and our face reconstruction model. In addition to bias, comparing attacks against different FR models in Tabs. 2 and 3 shows that FR models with worse recognition performance have lower SAR and more failed cases.", + "bbox": [ + 496, + 459, + 890, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ethical considerations This work is proposing a new method for a TI attack against FR models. The 3D reconstructed face can be used to generate 3D face masks or 2D printed photographs for presentation attacks against FR models. In addition to the security threats, the reconstructed face images can reveal important privacy-sensitive information of enrolled users, such as age, gender, ethnicity, etc. We do not condone using our work with the intent of attack to real FR systems. As a matter of fact, the main motivation for this work is to show such a vulnerability in the FR systems, and to encourage the scientific community to develop and propose the next generation of safe and protected FR systems. Along the same lines, data protection regulations, such as the European Union General Data Protection Regulation (EU-GDPR) [17], consider biometric data as sensi", + "bbox": [ + 496, + 672, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "4Samples of reconstructed face images with the camera parameters grid used in our grid search optimization are presented in the supplementary material.", + "bbox": [ + 76, + 862, + 470, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19669", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tive information, and put legal obligations to protect them. To this end and to mitigate such threats, several biometric template protection schemes are also proposed in the literature [40, 29, 31, 45]. We should also note that the project on which the work has been conducted has passed an Internal Ethical Review Board (IRB).", + "bbox": [ + 75, + 90, + 472, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 195, + 194, + 210 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we proposed a new method (dubbed GaFaR) to generate 3D face reconstruction from facial templates for a TI attack against FR models. We used a geometry-aware face generation network based on GNeRF and trained a mapping from facial templates to the intermediate latent space of the GNeRF model using a semi-supervised learning approach. To train our model, we used synthetic and real face images. For the synthetic training data, we had the latent code of each face image and could train our mapping with supervised learning. For the real training data, we used a GAN-based framework to learn the distribution of latent space. In the inference stage, we used optimization on the camera parameters to find the pose which increases the success attack rate. We evaluated our method on the whitebox and blackbox attacks against SOTA FR models on the LFW and MOBIO datasets. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates (extracted from 2D face recognition models).", + "bbox": [ + 75, + 220, + 472, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 76, + 522, + 228, + 540 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is based upon work supported by the H2020 TReSPAsS-ETN Marie Skłodowska-Curie early training network (grant agreement 860813).", + "bbox": [ + 76, + 547, + 468, + 594 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 607, + 174, + 623 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Muku Akasaka, Soshi Maeda, Yuya Sato, Masakatsu Nishigaki, and Tetsushi Ohki. Model-free template reconstruction attack with feature converter. In 2022 International Conference of the Biometrics Special Interest Group (BIOSIG), pages 1-5. IEEE, 2022.", + "[2] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Proceedings of the International Conference on Machine Learning (ICML), pages 214-223. PMLR, 2017.", + "[3] Battista Biggio, Paolo Russu, Luca Didaci, Fabio Roli, et al. Adversarial biometric recognition: A review on biometric system security from the adversarial machine-learning perspective. IEEE Signal Processing Magazine, 32(5):31-41, 2015.", + "[4] Fadi Boutros, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Elasticface: Elastic margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1578-1587, 2022." + ], + "bbox": [ + 84, + 632, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[5] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in Computer Vision, pages 671-679. Elsevier, 1987.", + "[6] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3981-3990, 2022.", + "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, 2022.", + "[8] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5799-5809, 2021.", + "[9] Forrester Cole, David Belanger, Dilip Krishnan, Aaron Sarna, Inbar Mosseri, and William T Freeman. Synthesizing normalized faces from facial identity features. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3703-3712, 2017.", + "[10] Debayan Deb, Jianbang Zhang, and Anil K Jain. Advfaces: Adversarial face synthesis. In Proceedings of the 2020 IEEE International Joint Conference on Biometrics (IJCB), pages 1-10. IEEE, 2020.", + "[11] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019.", + "[12] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10673-10683, 2022.", + "[13] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13733-13742, 2021.", + "[14] Xingbo Dong, Zhe Jin, Zhenhua Guo, and Andrew Beng Jin Teoh. Towards generating high definition face images from deep templates. In 2021 International Conference of the Biometrics Special Interest Group (BIOSIG), pages 1-11. IEEE, 2021.", + "[15] Xingbo Dong, Zhihui Miao, Lan Ma, Jiajun Shen, Zhe Jin, Zhenhua Guo, and Andrew Beng Jin Teoh. Reconstruct face from features using gan generator as a distribution constraint. arXiv preprint arXiv:2206.04295, 2022.", + "[16] Chi Nhan Duong, Thanh-Dat Truong, Khoa Luu, Kha Gia Quach, Hung Bui, and Kaushik Roy. Vec2face: Unveil human faces from their blackbox features in face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6132-6141, 2020." + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19670", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] European Council. Regulation of the european parliament and of the council on the protection of individuals with regard to the processing of personal data and on the free movement of such data (general data protection regulation), April 2016.", + "[18] Javier Galbally, Sébastien Marcel, and Julian Fierrez. Biometric antispoofing methods: A survey in face recognition. IEEE Access, 2:1530-1552, 2014.", + "[19] Javier Galbally, Chris McCool, Julian Fierrez, Sebastien Marcel, and Javier Ortega-Garcia. On the vulnerability of face verification systems to hill-climbing attacks. Pattern Recognition, 43(3):1027-1038, 2010.", + "[20] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In Proceedings of the 10th International Conference on Learning Representations (ICLR), pages 1-25, 2022.", + "[21] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 87-102. Springer, 2016.", + "[22] Abdenour Hadid, Nicholas Evans, Sebastien Marcel, and Julian Fierrez. Biometrics systems under spoofing attack: an evaluation methodology and lessons learned. IEEE Signal Processing Magazine, 32(5):20-30, 2015.", + "[23] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, October 2007.", + "[24] Yifan Jiang, Shiyu Chang, and Zhangyang Wang. Transgan: Two pure transformers can make one strong gan, and that can scale up. Advances in Neural Information Processing Systems, 34:14745-14758, 2021.", + "[25] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017.", + "[26] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019.", + "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019.", + "[28] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8110-8119, 2020.", + "[29] Prabhjot Kaur, Nitin Kumar, and Maheep Singh. Biometric cryptosystems: a comprehensive survey. Multimedia Tools and Applications, pages 1-56, 2022.", + "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Learning Representations (ICLR), San Diego, California., USA, May 2015.", + "[31] Nitin Kumar et al. Cancelable biometrics: a comprehensive survey. Artificial Intelligence Review, 53(5):3403-3446, 2020.", + "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 10012-10022, 2021.", + "[33] Guangcan Mai, Kai Cao, Pong C Yuen, and Anil K Jain. On the reconstruction of face images from deep face templates. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(5):1188-1202, 2018.", + "[34] Sébastien Marcel, Julian Fierrez, and Nicholas Evans. Handbook of Biometric Anti-Spoofing: Presentation Attack Detection and Vulnerability Assessment. Springer Nature, 2023.", + "[35] Sébastien Marcel, Mark S Nixon, Julian Fierrez, and Nicholas Evans. Handbook of biometric anti-spoofing: Presentation attack detection, volume 2. Springer, 2019.", + "[36] Chris McCool, Roy Wallace, Mitchell McLaren, Laurent El Shafey, and Sébastien Marcel. Session variability modelling for face authentication. IET Biometrics, 2(3):117-129, Sept. 2013.", + "[37] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6351-6361, 2021.", + "[38] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405-421, 2020.", + "[39] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 51-59, 2017.", + "[40] Karthik Nandakumar and Anil K Jain. Biometric template protection: Bridging the performance gap between theory and practice. IEEE Signal Processing Magazine, 32(5):88-100, 2015.", + "[41] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11453-11464, 2021.", + "[42] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022.", + "[43] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagan, and Andrea Tagliasacchi. Lolnerf: Learn from one look." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19671", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1558-1567, 2022.", + "[44] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020.", + "[45] Hatef Otroshi Shahreza, Christian Rathgeb, Daile Osorio-Roig, Vedrana Krivokuca Hahn, Sébastien Marcel, and Christoph Busch. Hybrid protection of biometric templates by combining homomorphic encryption and cancelable biometrics. In Proceedings of the 2022 IEEE International Joint Conference on Biometrics (IJCB), pages 1-10. IEEE, 2022.", + "[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.", + "[47] Mandavilli Srinivas and Lalit M Patnaik. Genetic algorithms: A survey. Computer, 27(6):17-26, 1994.", + "[48] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. arXiv preprint arXiv:2205.15517, 2022.", + "[49] Thanh-Dat Truong, Chi Nhan Duong, Ngan Le, Marios Savvides, and Khoa Luu. Vec2face-v2: Unveil human faces from their blackbox features via attention-based network in face recognition, 2022.", + "[50] Peter JM Van Laarhoven and Emile HL Aarts. Simulated annealing. In Simulated annealing: Theory and applications, pages 7-15. Springer, 1987.", + "[51] Edward Vendrow and Joshua Vendrow. Realistic face reconstruction from deep embeddings. In NeurIPS 2021 Workshop Privacy in Machine Learning, 2021.", + "[52] Fei Wang, Mengqing Jiang, Chen Qian, Shuo Yang, Cheng Li, Honggang Zhang, Xiaogang Wang, and Xiaou Tang. Residual attention network for image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3156-3164, 2017.", + "[53] Jun Wang, Yinglu Liu, Yibo Hu, Hailin Shi, and Tao Mei. Facex-zoo: A pytorch toolbox for face recognition. In Proceedings of the 29th ACM International Conference on Multimedia, 2021.", + "[54] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020.", + "[55] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18430-18439, 2022.", + "[56] Jichao Zhang, Enver Sangineto, Hao Tang, Aliaksandr Siarohin, Zhun Zhong, Nicu Sebe, and Wei Wang. 3d-aware semantic-guided generative model for human synthesis. arXiv preprint arXiv:2112.01422, 2021.", + "[57] Andrey Zhmoginov and Mark Sandler. Inverting face embeddings with convolutional neural networks. arXiv preprint arXiv:1606.04189, 2016." + ], + "bbox": [ + 78, + 90, + 468, + 895 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19672", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/203e5250-94d9-4833-ba8d-b9c548251baa_model.json b/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/203e5250-94d9-4833-ba8d-b9c548251baa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..13370fb3cdd99987218d12073b733287f60c03fa --- /dev/null +++ b/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/203e5250-94d9-4833-ba8d-b9c548251baa_model.json @@ -0,0 +1,2730 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.795, + 0.177 + ], + "angle": 0, + "content": "Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.203, + 0.684, + 0.221 + ], + "angle": 0, + "content": "Hatef Otroshi Shahreza1,2 and Sébastien Marcel1,3" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.225, + 0.677, + 0.243 + ], + "angle": 0, + "content": "1Idiap Research Institute, Martigny, Switzerland" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.244, + 0.785, + 0.261 + ], + "angle": 0, + "content": "\\(^{2}\\)École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.262, + 0.71, + 0.279 + ], + "angle": 0, + "content": "3Université de Lausanne (UNIL), Lausanne, Switzerland" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.283, + 0.666, + 0.298 + ], + "angle": 0, + "content": "{hatef.otoreshi,sebastien.marcel}@idiap.ch" + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.346, + 0.097, + 0.386 + ], + "angle": 270, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.319, + 0.213, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.319, + 0.326, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.32, + 0.438, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.32, + 0.551, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.32, + 0.663, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.32, + 0.776, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.32, + 0.89, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.413, + 0.097, + 0.49 + ], + "angle": 270, + "content": "Frontal Reconst." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.408, + 0.212, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.215, + 0.408, + 0.324, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.408, + 0.438, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.408, + 0.55, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.408, + 0.663, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.408, + 0.776, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.408, + 0.89, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.523, + 0.097, + 0.58 + ], + "angle": 270, + "content": "3D Reconst." + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.5, + 0.175, + 0.511 + ], + "angle": 0, + "content": "0.706" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.513, + 0.212, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.5, + 0.288, + 0.511 + ], + "angle": 0, + "content": "0.674" + }, + { + "type": "image", + "bbox": [ + 0.215, + 0.514, + 0.323, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.5, + 0.4, + 0.511 + ], + "angle": 0, + "content": "0.705" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.514, + 0.431, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.5, + 0.513, + 0.511 + ], + "angle": 0, + "content": "0.679" + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.514, + 0.545, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.5, + 0.626, + 0.511 + ], + "angle": 0, + "content": "0.663" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.513, + 0.663, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.5, + 0.738, + 0.511 + ], + "angle": 0, + "content": "0.733" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.513, + 0.776, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.5, + 0.852, + 0.511 + ], + "angle": 0, + "content": "0.782" + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.513, + 0.89, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.599, + 0.893, + 0.645 + ], + "angle": 0, + "content": "Figure 1: Sample face images from the FFHQ dataset (first row) as well as their corresponding 3D (third row) and frontal 2D reconstruction (second row) from facial templates in the whitebox template inversion attack against ArcFace. Values show the cosine similarity between the templates of the original and frontal reconstructed face images." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.656, + 0.314, + 0.672 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.69, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Face recognition systems are increasingly being used in different applications. In such systems, some features (also known as embeddings or templates) are extracted from each face image. Then, the extracted templates are stored in the system's database during the enrollment stage and are later used for recognition. In this paper, we focus on template inversion attacks against face recognition systems and introduce a novel method (dubbed GaFaR) to reconstruct 3D face from facial templates. To this end, we use a geometry-aware generator network based on generative neural radiance fields (GNeRF), and learn a mapping from facial templates to the intermediate latent space of the generator network. We train our network with a semi-supervised learning approach using real and synthetic images simul" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.895, + 0.871 + ], + "angle": 0, + "content": "taneously. For the real training data, we use a Generative Adversarial Network (GAN) based framework to learn the distribution of the latent space. For the synthetic training data, where we have the true latent code, we directly train in the latent space of the generator network. In addition, during the inference stage, we also propose optimization on the camera parameters to generate face images to improve the success attack rate (up to \\(17.14\\%\\) in our experiments). We evaluate the performance of our method in the whitebox and blackbox attacks against state-of-the-art face recognition models on the LFW and MOBIO datasets. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates. The project page is available at: https://www.idiap.ch/paper/gafar" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19662" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.208, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.468, + 0.296 + ], + "angle": 0, + "content": "Automatic face recognition (FR) has become a well-known biometric authentication tool which has been widely used in different applications, including smartphone locks, border controls, etc. In such systems, generally, some features (also referred to as facial embeddings or facial templates) are extracted from users and stored in the system's database during the enrollment stage. In the recognition stage, either verification or identification, similar features are extracted from the user and are compared with the features in the system's database. Therefore, facial features play the main role in automatic face recognition systems and convey important information about users." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.468, + 0.568 + ], + "angle": 0, + "content": "Template inversion (TI) attack against FR systems refers to an attempt by an adversary to reconstruct face images from the templates stored in the system's database. Then, the adversary may find sensitive information about the enrolled users and also can use the reconstructed face image to impersonate and enter the system. Therefore, compared to most attacks on FR systems which threaten the security of the system [19, 3, 22, 18, 35, 34, 10], TI attack jeopardizes both security and privacy of users, and thus requires further study. In this paper, we introduce a novel method (called geometry-aware face reconstruction, shortly GaFaR) for the TI attack in FR systems to reconstruct a 3D face from facial features (extracted from 2D FR models). The 3D face reconstruction provides further information than 2D face reconstruction, and in particular can be used to generate face image from any pose to improve the attack to the FR system. To our knowledge, this is the first paper on 3D face reconstruction from facial templates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.468, + 0.794 + ], + "angle": 0, + "content": "Recently, neural radiance fields (NeRF) [38] absorbed considerable attention in computer vision society due to remarkable results in the novel-view synthesis problem. Built upon NeRF, generative NeRF (GNeRF) methods such as [37, 44, 56, 6, 8, 41, 42, 20, 55, 7, 12, 48, 43] combine conditional NeRF with generative models for geometry-aware image generation tasks. In these methods, a generative model, such as a generative adversarial network (GAN), is used to embed the shape and appearance of an object in a latent space. Then, the latent code of the GAN, along with the camera parameters, are fed to a NeRF for the rendering process. Among GNeRF methods, there are several works, such as [8, 41, 42, 20, 55, 7, 12, 48], for geometry-aware 3D face generation, which can generate face images from different views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In our proposed TI method, we use a geometry-aware face generator network based on GNeRF, and learn a new mapping from facial templates to the intermediate latent space of the generator network. We train our network using real and synthesized images simultaneously with a semisupervised learning approach. For real training data where we do not have the corresponding latent code (i.e., unsuper" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.379 + ], + "angle": 0, + "content": "vised), we use a GAN-based framework to learn the distribution of the latent space. For the synthetic training data where we have true values of latent codes (i.e., supervised), we directly learn the latent space. Because we have 3D reconstructed face, we can use any arbitrary pose to inject the the FR system for the attack. Therefore, during the inference stage, we also use optimization on the camera parameters to improve the attack against the FR system. We evaluate the performance of our proposed method in the whitebox (i.e., where the adversary knows the internal functioning and parameters of the FR model) and blackbox (i.e., where the adversary does not have information about the internal functioning of the FR model) attacks against state-of-the-art (SOTA) FR systems. For evaluation, we consider FR systems with the LFW [23] and MOBIO [36] datasets and evaluate our attack if the reconstructed face can be used to enter the system. Fig. 1 illustrates sample face images from FFHQ [26] dataset and their 3D reconstruction from ArcFace [11] templates using our attack." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.38, + 0.892, + 0.409 + ], + "angle": 0, + "content": "To elaborate on the contributions of this paper, we list them hereunder:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.423, + 0.892, + 0.499 + ], + "angle": 0, + "content": "- We propose a novel template inversion method to reconstruct 3D faces from the facial templates of a face recognition system. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.513, + 0.892, + 0.605 + ], + "angle": 0, + "content": "- We use a geometry-aware generator network based on GNeRF, and learn a mapping from facial templates to the intermediate latent space of the generator network. We train our mapping network using a GAN-based framework and a semi-supervised learning approach using real and synthetic face images." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.618, + 0.892, + 0.678 + ], + "angle": 0, + "content": "- We use optimization on the camera parameters in the geometry-aware generator network during the inference stage. The optimization of camera parameters can find a pose that improves the success attack rate." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.423, + 0.892, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.892, + 0.782 + ], + "angle": 0, + "content": "The remainder of the paper is organized as follows. First, we review the related works in the literature in Sec. 2. Then, we present our proposed method in Sec. 3. Next, we describe our experiments in Sec. 4 and discuss limitations and ethical considerations in Sec. 5. Finally, the paper is concluded in Sec. 6." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.799, + 0.65, + 0.815 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Generally, the methods in the literature for reconstructing face from facial templates generate 2D face images. In [57], authors proposed two methods based on optimization and learning to reconstruct 2D low-resolution face images from facial templates in the whitebox scenario. In the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19663" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.138, + 0.09, + 0.409, + 0.105 + ], + "angle": 0, + "content": "Table 1: Comparison with related works." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.109, + 0.477, + 0.276 + ], + "angle": 0, + "content": "
Ref.2D/3DResolutionWhitebox/BlackboxMethod BasisAvailable code
[57]2Dlowwhitebox1) optimization\n2) learningX
[9]2Dlowboth*learningX
[33]2Dlowblackboxlearning✓
[16]2Dlowboth†learningX
[49]2Dlowboth†learningX
[1]2Dlowblackboxlearning + opt.X
[14]2Dhighblackboxlearning✓
[51]2Dhighblackboxoptimization✓
[15]2DhighblackboxoptimizationX
[Ours]3Dhighboth‡learning✓
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.08, + 0.279, + 0.47, + 0.3 + ], + "angle": 0, + "content": "*The method is based on the whitebox attack, and is also applied in the blackbox scenario by removing a loss term that required the FR model." + }, + { + "type": "table_footnote", + "bbox": [ + 0.08, + 0.301, + 0.469, + 0.323 + ], + "angle": 0, + "content": "The method is based on the whitebox attack, and is extended to the blackbox with knowledge distillation of the FR model." + }, + { + "type": "table_footnote", + "bbox": [ + 0.08, + 0.324, + 0.469, + 0.345 + ], + "angle": 0, + "content": "\\(\\ddagger\\) The method is based on the whitebox attack, and is extended to blackbox using a different FR model." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.279, + 0.47, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.47, + 0.531 + ], + "angle": 0, + "content": "optimization-based method, they used a gradient ascent-based algorithm to reconstruct the face image, and optimized the generated image to minimize a multi-term loss function, including the \\(\\ell_2\\) distance between target templates and templates of the reconstructed face image. In addition, they applied total-variation and Laplacian pyramid gradient normalization [5] on the reconstructed image to generate a smooth image. In their learning-based method, they used a deconvolutional neural network to reconstruct face images and trained it with the same loss function that they used in their optimization-based method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.732 + ], + "angle": 0, + "content": "In [9], a multi-layer perceptron (MLP) is trained to find facial landmark coordinates (optimized using the mean squared error) from the given template. They also trained a convolutional neural network (CNN) to generate face texture (optimized using mean absolute error) from the target template. Then, they used a differentiable warping to combine estimated landmarks (from MLP) and textures (from CNN) and reconstruct low-resolution face images. In the whitebox scenario, they further optimized MLP and CNN by minimizing the distance between templates of reconstructed and original face images. However, in the blackbox scenario, they trained MLP and CNN separately, and used warping in the inference only." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In [33], two new deconvolutional networks are proposed, called NbBlock-A and NbBlock-B, and are trained with pixel loss \\((\\ell_{1}\\) norm of reconstruction error) and perceptual loss (distance of middle layers of VGG-19[46] when given the reconstructed and original face images) to reconstruct low-resolution face images in the blackbox scenario. In [16] and [49], a same bijection-learning-based method is used to train GAN models with PO-GAN [25] and TransGAN [24] structure, respectively. The method is based on whitebox attack and for training the GAN model, authors also minimized the distance between target templates and templates" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "extracted from the reconstructed face images using the FR model. In the blackbox attack, they proposed to use the distillation of knowledge to train a student network that mimics the target FR model and used the student network in their method. However, they did not report any detail (and no published source code) about the training of the student network (e.g., network structure, etc.)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.892, + 0.348 + ], + "angle": 0, + "content": "In [1], a 3-step method is proposed to reconstruct low-resolution face images in the blacbox scenario. First, they trained a GAN for general face generation. In the second step, they trained a MLP to map target templates to embeddings of a known FR model. In the final step, they found a latent code in the input of their generator (of their GAN) which generates a face image that maximizes two terms; the discriminator score (for being a real face image) and cosine similarity between the mapped embedding and the embeddings extracted by the known FR model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.893, + 0.62 + ], + "angle": 0, + "content": "In contrast to most works in the literature which generate low-resolution face images, recently there have been few works to generate high-resolution 2D face images. In [51], a grid-search optimization using the simulated annealing [50] approach is used on the latent vector (i.e., input noise) of StyleGAN2 [28] to find latent codes that can generate face images which have templates similar to the target templates. However, their proposed method is computationally expensive, and they reported their evaluation on only 20 face images. In [15], a similar optimization to [51] on the latent vector of StyleGAN2 [28] is considered, but it is solved using the standard genetic algorithm [47]. In contrast to [51, 15] which are based on optimization, in [14], a learning-based method is proposed in the blackbox scenario. The authors proposed to generate some face images using StyleGAN2 [28] and extract the templates using the FR model. Then, they trained a MLP to map facial templates to the input latent codes of StyleGAN2 [28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.741 + ], + "angle": 0, + "content": "Tab. 1 compares our proposed method with the previous works in the literature. To our knowledge, our proposed method is the first work on 3D face reconstruction from facial templates (extracted from 2D face recognition models). Furthermore, unlike most works in the literature, our method generates high-resolution face images. Last but not least, our method can be used in both whitebox and black-box attacks against FR systems." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.673, + 0.772 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we consider a threat model as described in Sec. 3.1 and use the proposed 3D face reconstruction method (dubbed GaFaR) described in Sec. 3.2. After training and in the inference stage, we use an optimization on the camera parameters as described in Sec. 3.3 to generate a 2D face image with a pose that can lead to a higher success attack rate (SAR) and inject the face image into the system. Fig. 2 depicts the block-diagram of the proposed method." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19664" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.221, + 0.106 + ], + "angle": 0, + "content": "3.1. Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.474, + 0.435 + ], + "angle": 0, + "content": "We consider an attack against FR systems, where the adversary gains access to the system's database and aims to generate a 3D reconstruction of enrolled users. To this end, the adversary trains a network to reconstruct a 3D face given a target facial template. Then, the adversary can use the 3D reconstruction to impersonate users into the system (i.e., using a 3D face mask, etc.). For simplicity, we consider that the adversary can inject a 2D face image (from the 3D reconstructed face) into the system. Hence, the adversary uses an appropriate pose to generate a 2D face image for the attack. According to the adversary's knowledge of the FR model, we consider both whitebox and blackbox attacks against FR systems. In the whitebox scenario, we assume that the adversary has complete knowledge of the FR model from which the templates are leaked. However, in the blackbox scenario, we assume that the adversary does not know the internal functioning of the FR model and can only use the FR model to generate facial template for any face image. In addition, we assume that in the blackbox scenario, the adversary has access to another FR model and knows its internal parameters and functioning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.45, + 0.297, + 0.465 + ], + "angle": 0, + "content": "3.2. 3D Face Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.47, + 0.643 + ], + "angle": 0, + "content": "To reconstruct 3D faces from facial templates, we consider a geometry-aware face generator network based on GNeRF, such as EG3D [7], and learn a mapping \\( M_{\\mathrm{GaFaR}}: \\mathcal{T} \\to \\mathcal{W} \\) from the facial templates \\( t \\in \\mathcal{T} \\) to the intermediate latent space \\( \\mathcal{W} \\) of the GNeRF model. Then, we use the remainder of GNeRF model \\( G(.) \\) to synthesize an image \\( \\hat{I} = G(\\hat{w}, c) \\) from an arbitrary view using the mapped intermediate latent vector \\( \\hat{w} \\) and camera parameters \\( c \\). We consider a network with two fully-connected layers with the Leaky ReLU activation functions for our mapping network \\( M_{\\mathrm{GaFaR}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We train the mapping network \\(M_{\\mathrm{GaFaR}}\\) with a semi-supervised approach using synthetic and real data. For the real data, we consider a set of real face images \\(\\{I_{\\mathrm{real},i}\\}_{i = 0}^{N}\\) and extract the facial template \\(t_{\\mathrm{real},i} = F(I_{\\mathrm{real},i})\\) from each face image \\(I_{\\mathrm{real},i}\\) using the FR model \\(F(.)\\). As we do not have the true value of intermediate latent space \\(\\mathcal{W}\\) of the GNeRF model to generate the real face image, we consider training our mapping network using this data as unsupervised learning. On the other hand, for the synthetic data, we use the pre-trained geometry-aware face generator network to generate a set of random face images \\(\\{I_{\\mathrm{syn},i}\\}_{i = 0}^{M}\\). Therefore, as opposed to real data, we have the true value of intermediate latent space \\(\\boldsymbol {w}\\in \\mathcal{W}\\) to generate the same synthetic face image. Hence, we consider training our mapping network using the synthetic data as supervised learning. We train our mapping network simultaneously using real and synthetic training data as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Unsupervised learning using real data To train our mapping network \\(M_{\\mathrm{GaFaR}}(.)\\) with the real data, we use a GAN-based framework based on Wasserstein GAN (WGAN) [2] algorithm to learn the distribution of intermediate latent space \\(\\mathcal{W}\\) of the GNeRF model. In this framework, our mapping network \\(M_{\\mathrm{GaFaR}}\\) generates a latent code \\(\\hat{\\pmb{w}} = M_{\\mathrm{GaFaR}}([n,t])\\) using the facial template \\(\\pmb{t}\\) and a random vector \\(\\pmb{n} \\in \\mathcal{N}\\). We can also generate real latent codes \\(\\pmb{w} = M(\\pmb{z}) \\in \\mathcal{W}\\) for training our GAN using the GNeRF mapping function \\(M\\) and a random vector \\(\\pmb{z} \\in \\mathcal{Z}\\). Then, we can use a critic network \\(C(.)\\) to score the latent codes generated by our mapping \\(M_{\\mathrm{GaFaR}}(.)\\) and GNeRF mapping \\(M(.)\\). Therefore, we can train the critic network \\(C(.)\\) and our mapping \\(M_{\\mathrm{GaFaR}}\\) using the following loss functions:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.328, + 0.891, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C} ^ {\\mathrm {W G A N}} = \\mathbb {E} _ {\\boldsymbol {w} \\sim M (\\boldsymbol {z})} [ C (\\boldsymbol {w}) ] - \\mathbb {E} _ {\\hat {\\boldsymbol {w}} \\sim M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, t ])} [ C (\\hat {\\boldsymbol {w}}) ] \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.364, + 0.892, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M _ {\\mathrm {G a F a R}}} ^ {\\mathrm {W G A N}} = \\mathbb {E} _ {\\hat {\\boldsymbol {w}} \\sim M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, \\boldsymbol {t} ])} [ C (\\hat {\\boldsymbol {w}}) ] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.44 + ], + "angle": 0, + "content": "In addition to the WGAN training, we use the generated face image \\(\\hat{\\pmb{I}} = G(M_{\\mathrm{GaFaR}}([n,t]),c)\\) to optimize our mapping network using the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.449, + 0.892, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e a l}} ^ {\\text {r e c}} = \\mathcal {L} ^ {\\text {P i x e l}} + \\mathcal {L} ^ {\\text {I D}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.508 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}^{\\mathrm{Pixel}}\\) and \\(\\mathcal{L}^{\\mathrm{ID}}\\) are pixel loss and ID loss, respectively, and are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.519, + 0.891, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {P i x e l}} = \\mathbb {E} _ {\\hat {\\boldsymbol {w}} \\sim M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, \\boldsymbol {t} ])} \\left[ \\| \\boldsymbol {I} - G (\\hat {\\boldsymbol {w}}, \\boldsymbol {c}) \\| _ {2} ^ {2} \\right] \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.541, + 0.892, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {I D}} = \\mathbb {E} _ {\\hat {\\boldsymbol {w}} \\sim M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, \\boldsymbol {t} ])} [ \\| F _ {\\text {l o s s}} (\\boldsymbol {I}) - F _ {\\text {l o s s}} (G (\\hat {\\boldsymbol {w}}, \\boldsymbol {c})) \\| _ {2} ^ {2} ] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.721 + ], + "angle": 0, + "content": "The pixel loss minimizes the pixel-level reconstruction error and ID loss helps the network to generate face images with a similar facial templates extracted by \\( F_{\\mathrm{loss}} \\). To calculate the ID loss in Eq. (5), we use the same FR model as the one by which leaked templates are calculated (i.e., \\( F \\)) in the whitebox scenario. However, in the blackbox scenario, we use another FR model that we assume the adversary has access to. The critic network is a fully-connected network with three layers and Leaky ReLU activation function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.739, + 0.892, + 0.83 + ], + "angle": 0, + "content": "Supervised learning using synthetic data To train our mapping network \\(M_{\\mathrm{GaFaR}}(.)\\) with synthetic data, we can directly learn the GNeRF intermediate latent code \\(\\pmb {w} = M(z)\\). In addition to directly learning \\(\\pmb{w}\\), we use the generated face image to train our mapping network by minimizing the following multi-term loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.839, + 0.892, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s y n}} ^ {\\text {r e c}} = \\mathcal {L} ^ {w} + \\mathcal {L} ^ {\\text {P i x e l}} + \\mathcal {L} ^ {\\text {I D}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.87, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}^{\\mathrm{Pixel}}\\) and \\(\\mathcal{L}^{\\mathrm{ID}}\\) are pixel loss and ID loss, respectively. Furthermore, \\(\\mathcal{L}^w\\) is \\(w\\)-loss to directly learn the latent space" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19665" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.093, + 0.35, + 0.107 + ], + "angle": 0, + "content": "Training (Semi-supervised Learning)" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.111, + 0.805, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.283, + 0.871, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.445, + 0.203, + 0.457 + ], + "angle": 0, + "content": "Inference" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.457, + 0.891, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.618, + 0.895, + 0.741 + ], + "angle": 0, + "content": "Figure 2: Block-diagram of the proposed method: In the training process, a semi-supervised approach is used to learn our mapping \\( M_{\\mathrm{GaFaR}} \\) (depicted with a green block) from the facial templates to the intermediate latent space of the GNeRF model using real data (unsupervised, where we don't have the corresponding latent code) and synthetic data (supervised, where we have the corresponding latent code \\( \\boldsymbol{w} \\)), simultaneously. During the inference stage, the leaked template \\( t \\) is given to our mapping network to find corresponding latent code \\( \\hat{\\boldsymbol{w}} = M_{\\mathrm{GaFaR}}([n, t]) \\). Then, \\( \\hat{\\boldsymbol{w}} \\) along with camera parameters \\( c \\) are given to the generator and renderer of GNeRF \\( G \\) to generate a reconstructed face image \\( \\hat{I} = G(\\hat{\\boldsymbol{w}}, c) \\). Finally, an optimization (either grid search or continuous optimization) is applied on two of the camera parameters, \\( \\theta \\) and \\( \\psi \\), from \\( c \\), to find the best pose that minimizes the distance between leaked template \\( t \\) and the template of reconstructed face image." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.471, + 0.796 + ], + "angle": 0, + "content": "by minimizing mean squared error between \\(\\pmb{w}\\) and \\(\\hat{\\pmb{w}} = M_{\\mathrm{GaFaR}}([n,t])\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.805, + 0.47, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {w} = \\mathbb {E} _ {\\boldsymbol {w} \\sim M (\\boldsymbol {z})} [ \\| \\boldsymbol {w} - M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, \\boldsymbol {t} ]) \\| _ {2} ^ {2} ] \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.833, + 0.376, + 0.849 + ], + "angle": 0, + "content": "3.3. Camera Parameters Optimization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "During the inference stage, after generating a 3D reconstruction of face using our method described in Sec. 3.2, we optimize the camera parameters to find a pose that increases" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "the SAR. Among different camera parameters \\( c \\), the parameters which correspond to the camera rotation can change the pose of the generated face image. Note that by varying the camera rotations, we aim to change the pitch and yaw rotations of the 3D reconstructed face and do not modify the roll rotation. In fact, the roll rotation of the face will be eliminated through face alignment as a preprocess prior to feature extraction in the FR system. To optimize camera parameters, we consider two different approaches as follows:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19666" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.246 + ], + "angle": 0, + "content": "Grid Search (GS) For the grid search, we consider predefined steps to change the camera pitch \\(\\theta \\in \\Theta\\) and yaw \\(\\psi \\in \\Psi\\) rotations of camera parameters \\(c\\). We generate the 2D face image for all values of camera rotation steps (\\(\\theta_{\\mathrm{step}}\\) and \\(\\psi_{\\mathrm{step}}\\)) and find the facial template for each generated image. Then, we select the face image \\(\\hat{\\pmb{I}} = G(M_{\\mathrm{GaFaR}}([n,t]),c)\\) that has a template \\(\\hat{\\pmb{t}} = F(\\hat{\\pmb{I}})\\) with the minimum mean squared error with the target template \\(\\pmb{t}\\). Note that the grid search can be applied in both whitebox and blackbox scenarios using the FR model \\(F\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.263, + 0.472, + 0.325 + ], + "angle": 0, + "content": "Continuous Optimization (CO) In our continuous optimization approach, we start from the frontal image and use Adam [30] optimizer to solve the following minimization using \\(\\hat{\\boldsymbol{w}} = M_{\\mathrm{GaFaR}}([\\boldsymbol{n},\\boldsymbol{t}])\\)" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.332, + 0.47, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta , \\psi} \\| F (G (\\hat {w}, c)), t \\| _ {2} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.47, + 0.442 + ], + "angle": 0, + "content": "In this optimization, we find the parameters \\(\\theta\\) and \\(\\psi\\) for camera rotation that lead to a face image with templates close to the target template \\(t\\). In contrast to the grid search approach, the continuous optimization can be applied only in the whitebox scenario." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.456, + 0.21, + 0.472 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.48, + 0.267, + 0.497 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Face Recognition models In our experiments, we consider different SOTA FR models including ArcFace [11], ElasticFace [4] as well as four different FR models with SOTA backbones from FaceX-Zoo [53], including AttentionNet [52], HRNet [54], RepVGG [13] and Swin [32]. The recognition performance of these models are reported in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.761 + ], + "angle": 0, + "content": "Datasets All the mentioned FR models are trained on the MS-Celeb1M dataset [21]. To train the face reconstruction network, we assume that adversary does not have any knowledge of the dataset used in training FR model, and uses a different dataset. To this end, we consider the Flickr-Faces-HQ (FFHQ) dataset [27], which consists of 70,000 high-quality images crawled from internet (with no identity label), for training our face reconstruction network. We randomly split FFHQ into train \\((90\\%)\\) and test \\((10\\%)\\) set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.471, + 0.882 + ], + "angle": 0, + "content": "To evaluate the vulnerability of FR models, we use two other datasets, including Labeled Faces in the Wild (LFW) [23] and MOBIO [36] datasets. The LFW dataset consists of 13,233 face images of 5,749 people collected from internet, where 1,680 people have two or more images. The MOBIO dataset includes face images of 150 people captured using mobile devices in 12 sessions for each person1." + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.103, + 0.523, + 0.144 + ], + "angle": 0, + "content": "rnnnne" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.088, + 0.618, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.088, + 0.707, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.088, + 0.797, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.089, + 0.886, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.175, + 0.521, + 0.215 + ], + "angle": 0, + "content": "Reconnst" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.157, + 0.617, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.157, + 0.706, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.157, + 0.796, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.157, + 0.885, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.228, + 0.591, + 0.24 + ], + "angle": 0, + "content": "0.695" + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.228, + 0.68, + 0.24 + ], + "angle": 0, + "content": "0.721" + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.228, + 0.77, + 0.24 + ], + "angle": 0, + "content": "0.699" + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.228, + 0.86, + 0.24 + ], + "angle": 0, + "content": "0.732" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.246, + 0.892, + 0.338 + ], + "angle": 0, + "content": "Figure 3: Sample face images from the FFHQ dataset (first row) and their corresponding frontal (second row) reconstructed face images using our method in the blackbox attack against ElasticFace using ArcFace for \\( F_{\\mathrm{loss}} \\). The values show the cosine similarity between templates of original and frontal reconstructed face images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.497 + ], + "angle": 0, + "content": "Evaluation Protocol For evaluation with each of the LFW and MOBIO datasets, we build a separate FR system, and use the reference templates (i.e., enrolled in the system's database) as input to our face reconstruction method. Then, we inject the reconstructed face image as a query to the system and evaluate the adversary's success attack rate (SAR) to enter the FR system. We should note that for each image, we use only one reconstructed face image as a query to the FR system and evaluate the SAR." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.893, + 0.859 + ], + "angle": 0, + "content": "Implementation Details and Source Code In our experiments, we use the grid search (in both whitebox and black-box attacks) and continuous optimization (in whitebox attacks only) to optimize camera parameters, as described in Sec. 3.3, in the inference stage. For grid search, we consider \\(\\psi \\in [-45^{\\circ}, +45^{\\circ}]\\) and \\(\\theta \\in [-30^{\\circ}, +30^{\\circ}]\\) for a \\(11 \\times 11\\) grid with step sizes of \\(\\psi_{\\mathrm{step}} = 9^{\\circ}\\) and \\(\\theta_{\\mathrm{step}} = 6^{\\circ}\\). In the continuous optimization approach, we use 121 iterations of Adam optimizer [30] with the learning rate of \\(10^{-2}\\). An ablation study on these hyperparameters and the corresponding execution times are reported in the supplementary material. All our models (for different FR models and different scenarios) are trained in 15 epochs (each takes around 2 days) on a system equipped with a single NVIDIA RTX 3090 GPU. For GNeRF, we used a pretrained model of EG3D [7] in our experiments and generate \\(512 \\times 512\\) resolution face images. The input noise \\(z\\) to GNeRF model has 512 dimensions and the noise \\(n\\) in the input of our mapping network \\(M_{\\mathrm{GaFaR}}\\) has 16 dimensions. For unsupervised learning using real face images (FFHQ), we assume that camera parameters for real face images as frontal view. The source code of our experiments are publicly available2." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.468, + 0.901 + ], + "angle": 0, + "content": "1The results on AgeDB [39] dataset is also reported in the supplimen-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.501, + 0.876, + 0.575, + 0.888 + ], + "angle": 0, + "content": "tary material." + }, + { + "type": "page_footnote", + "bbox": [ + 0.52, + 0.888, + 0.7, + 0.901 + ], + "angle": 0, + "content": "2 https://www.idiap.ch/paper/gafar" + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.876, + 0.7, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19667" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.137 + ], + "angle": 0, + "content": "Table 2: Evaluation of blackbox attack against SOTA FR models at systems' FMR=10\\(^{-3}\\) on the LFW and MOBIO datasets in terms of success attack rate (SAR). For attacks using our proposed method, we use ArcFace and also ElasticFace as \\(F_{\\mathrm{loss}}\\) to calculate the ID loss in Eq. (5). The values are in percentage." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.141, + 0.891, + 0.295 + ], + "angle": 0, + "content": "
LFWMOBIO
ArcFaceElsFaceAtt.NetHRNetRepVGGSwinArcFaceEls.FaceAtt.NetHRNetRepVGGSwin
NBNetA-M [33]4.3210.901.241.601.133.8202.380000
NBNetA-P [33]16.8326.980.661.445.729.704.7616.190.48014.297.14
NBNetB-M [33]10.9821.443.224.473.2111.231.903.803.337.143.338.57
NBNetB-P [33]40.2658.1616.2918.4215.2440.7615.2443.8131.9026.6723.8144.29
Dong et al. [14]13.2112.613.904.073.2212.383.338.1010.486.679.053.33
Vendrow and Vendrow [51]57.7053.0321.1218.859.6246.8429.0543.8127.1426.6720.9545.24
[Floss= Els.Face] GaFaR51.78-18.0711.6811.6347.1547.62-54.2943.3345.7170.48
[Floss= Els.Face] GaFaR + GS61.56-23.5617.2115.8254.0864.76-63.8155.2358.5782.38
[Floss= ArcFace] GaFaR-74.5433.5937.8025.4067.11-84.7672.3876.6772.8689.05
[Floss= ArcFace] GaFaR + GS-78.6738.4243.2729.8470.82-86.6280.0083.8073.3393.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.306, + 0.895, + 0.339 + ], + "angle": 0, + "content": "Table 3: Evaluation of whitebox attack against SOTA FR models at systems' FMR = 10\\(^{-3}\\) on the LFW and MOBIO datasets in terms of success attack rate (SAR). The values are in percentage." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.342, + 0.88, + 0.419 + ], + "angle": 0, + "content": "
LFWMOBIO
ArcFaceElsFaceAtt.NetHRNetRepVGGSwinArcFaceEls.FaceAtt.NetHRNetRepVGGSwin
GaFaR79.8462.3227.0031.8717.3374.0882.8681.4364.2971.4353.8194.76
GaFaR + GS82.5268.7132.4237.4819.8376.0685.2382.3870.4781.4259.5294.29
GaFaR + CO84.2571.6934.5339.5320.4477.3689.0584.2975.7181.4362.8695.71
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.44, + 0.178, + 0.456 + ], + "angle": 0, + "content": "4.2. Analyze" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.474, + 0.768 + ], + "angle": 0, + "content": "Blackbox Scenario To evaluate the proposed method in the blackbox attack against SOTA FR models, we consider ArcFace and also ElasticFace as \\( F_{\\mathrm{loss}} \\) to calculate the ID loss in Eq. (5). Tab. 2 compares the performance of the proposed method with blackbox face reconstruction methods in the literature in terms of SAR for systems configured at false match rate (FMR) of \\( 10^{-3} \\). Similar results for \\( \\mathrm{FMR} = 10^{-2} \\) are reported in the supplementary material. As this table shows, the frontal reconstruction by our method (i.e., GaFaR) achieves superior performance than previous methods in the literature. Furthermore, camera parameter optimization (i.e., GaFaR+GS) improves the performance of our attack up to \\( 17.14\\% \\) compared to our frontal face reconstruction (i.e., GaFaR). Comparing the use of ArcFace and ElasticFace as \\( F_{\\mathrm{loss}} \\), the performance of attacks with the ArcFace model is better. This is due to the fact that ArcFace has a higher recognition accuracy than ElasticFace. Fig. 3 shows sample face images reconstructed from ElasticFace templates in the blackbox attack (using ArcFace for \\( F_{\\mathrm{loss}} \\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.471, + 0.866 + ], + "angle": 0, + "content": "Whitebox Scenario Tab. 3 reports the performance of the proposed method in the whitebox attack\\(^3\\) in terms of SAR for FR systems configured at false match rate (FMR) of \\(10^{-3}\\). Similar results for \\(\\mathrm{FMR} = 10^{-2}\\) are reported in the supplementary material. According to this table, all" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.437, + 0.892, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.743, + 0.895, + 0.837 + ], + "angle": 0, + "content": "Figure 4: Sample face images from the FFHQ dataset (first row) and their corresponding frontal (second row) and camera parameters sub-grid (third row) reconstructed face images using our method in the whitebox attack against ArcFace. The values show the cosine similarity between templates of original and frontal reconstructed face images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "these FR models are highly vulnerable to our attack. Furthermore, the camera parameter optimization improves the performance of GaFaR. Comparing grid search with contin" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.471, + 0.901 + ], + "angle": 0, + "content": "3The whitebox attack methods reported in Tab. 1 do not have available source code, and we could not reproduce their results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19668" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.473, + 0.18 + ], + "angle": 0, + "content": "Table 4: Ablation study on the proposed semi-supervised learning approach and evaluation of loss terms in the white-box attack against ArcFace model in terms of success attack rate (SAR) on the LFW and MOBIO datasets. The SAR values are in percentage and for an attack without any camera parameter optimization (GS/CO)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.184, + 0.468, + 0.35 + ], + "angle": 0, + "content": "
approachLoss FunctionsLFWMOBIO
FMR=10-2FMR=10-3FMR=10-2FMR=10-3
supervisedLrecsyn = Lw + LPixel + LID83.8069.46790.9682.38
Lrecsyn = Lw + LPixel31.7513.9243.818.57
Lrecsyn = Lw + LID0.860.3000
Lrecsyn = Lw33.6915.4332.389.52
unsupervisedLrecreal = LPixel + LID[without WGAN]0.440.1500
Lrecreal = LPixel + LID[with WGAN]67.7245.7670.4831.90
Lrecreal = LID[with WGAN]54.5130.8352.8619.52
Lrecreal = LPixel[with WGAN]2.210.4000
semi-supervisedEqs. (1) to (3) and (6)89.2779.8495.7182.86
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.473, + 0.588 + ], + "angle": 0, + "content": "uous optimization, results show that with the same number of iterations continuous optimization achieves better performance. In addition, comparing these results with recognition performance of FR models (available in the supplementary material), we conclude that models with higher recognition accuracy are more vulnerable to our attack. Comparing results in Tab. 2 and Tab. 3, we observe that when we use ArcFace as \\( F_{\\mathrm{loss}} \\) in our loss function, blackbox attacks achieve better results than whitebox attacks for most cases, which can be explained considering the superior recognition performance of ArcFace than other FR models. Fig. 4 shows sample face images reconstructed from ArcFace templates in the whitebox attack. This figure also presents a grid of reconstructed face images with different poses4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.473, + 0.852 + ], + "angle": 0, + "content": "Ablation Study We evaluate the effect of our semi-supervised learning approach in our proposed method compared to fully supervised learning and fully unsupervised learning approach. In each case, we evaluate the effect of each loss function too. Furthermore, for the fully unsupervised learning approach, we evaluate the effect of GAN learning which we used in our method. Tab. 4 reports our ablation study in whitebox attack against ArcFace model on the LFW and MOBIO datasets in terms of SAR at system's FMR of \\(10^{-2}\\) and \\(10^{-3}\\). As these results show, the proposed semi-supervised approach achieves a better performance than fully supervised learning and fully unsupervised learning approaches. Furthermore, each of our loss terms has an important effect on the performance. In particular, using WGAN for real data, where we don't have the true value of intermediate latent codes, helps the network to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.863, + 0.472, + 0.9 + ], + "angle": 0, + "content": "4Samples of reconstructed face images with the camera parameters grid used in our grid search optimization are presented in the supplementary material." + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.103, + 0.523, + 0.144 + ], + "angle": 0, + "content": "rannnnnne" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.089, + 0.618, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.089, + 0.706, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.089, + 0.796, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.089, + 0.886, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.175, + 0.522, + 0.216 + ], + "angle": 0, + "content": "Reconstrn" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.157, + 0.617, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.157, + 0.705, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.157, + 0.795, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.157, + 0.885, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.228, + 0.591, + 0.241 + ], + "angle": 0, + "content": "0.160" + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.228, + 0.681, + 0.241 + ], + "angle": 0, + "content": "0.150" + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.228, + 0.771, + 0.241 + ], + "angle": 0, + "content": "0.170" + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.228, + 0.86, + 0.241 + ], + "angle": 0, + "content": "0.160" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.25, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Figure 5: Sample failure cases from the LFW dataset (first row) and their corresponding frontal (second row) reconstructed face images using our method in the whitebox attack against ArcFace. The values below each image show the cosine similarity between templates of original and reconstructed face images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.422 + ], + "angle": 0, + "content": "learn the distribution of GNeRF intermediate latent space \\(\\mathcal{W}\\). Otherwise, the generated latent code by the mapping network will be out of distribution, and therefore generator part of GNeRF will not generate face-like images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.434, + 0.614, + 0.451 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Limitations In spite of the considerable success attack rate of our method in our experiments, the reconstructed face images from some facial templates fail to attack the system. Fig. 5 illustrates sample failed cases in the whitebox attack against ArcFace. As these sample images show, the failed reconstructions reflect a bias in the generated faces for dark skin and also old people. Such a bias in final results can be caused by inherent biases of each of the datasets used to train the FR model, GNeRF model, and our face reconstruction model. In addition to bias, comparing attacks against different FR models in Tabs. 2 and 3 shows that FR models with worse recognition performance have lower SAR and more failed cases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ethical considerations This work is proposing a new method for a TI attack against FR models. The 3D reconstructed face can be used to generate 3D face masks or 2D printed photographs for presentation attacks against FR models. In addition to the security threats, the reconstructed face images can reveal important privacy-sensitive information of enrolled users, such as age, gender, ethnicity, etc. We do not condone using our work with the intent of attack to real FR systems. As a matter of fact, the main motivation for this work is to show such a vulnerability in the FR systems, and to encourage the scientific community to develop and propose the next generation of safe and protected FR systems. Along the same lines, data protection regulations, such as the European Union General Data Protection Regulation (EU-GDPR) [17], consider biometric data as sensi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19669" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.182 + ], + "angle": 0, + "content": "tive information, and put legal obligations to protect them. To this end and to mitigate such threats, several biometric template protection schemes are also proposed in the literature [40, 29, 31, 45]. We should also note that the project on which the work has been conducted has passed an Internal Ethical Review Board (IRB)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.196, + 0.196, + 0.212 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.473, + 0.51 + ], + "angle": 0, + "content": "In this paper, we proposed a new method (dubbed GaFaR) to generate 3D face reconstruction from facial templates for a TI attack against FR models. We used a geometry-aware face generation network based on GNeRF and trained a mapping from facial templates to the intermediate latent space of the GNeRF model using a semi-supervised learning approach. To train our model, we used synthetic and real face images. For the synthetic training data, we had the latent code of each face image and could train our mapping with supervised learning. For the real training data, we used a GAN-based framework to learn the distribution of latent space. In the inference stage, we used optimization on the camera parameters to find the pose which increases the success attack rate. We evaluated our method on the whitebox and blackbox attacks against SOTA FR models on the LFW and MOBIO datasets. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates (extracted from 2D face recognition models)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.523, + 0.229, + 0.541 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.549, + 0.47, + 0.595 + ], + "angle": 0, + "content": "This research is based upon work supported by the H2020 TReSPAsS-ETN Marie Skłodowska-Curie early training network (grant agreement 860813)." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.608, + 0.175, + 0.624 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.633, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[1] Muku Akasaka, Soshi Maeda, Yuya Sato, Masakatsu Nishigaki, and Tetsushi Ohki. Model-free template reconstruction attack with feature converter. In 2022 International Conference of the Biometrics Special Interest Group (BIOSIG), pages 1-5. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.704, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[2] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Proceedings of the International Conference on Machine Learning (ICML), pages 214-223. PMLR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.761, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[3] Battista Biggio, Paolo Russu, Luca Didaci, Fabio Roli, et al. Adversarial biometric recognition: A review on biometric system security from the adversarial machine-learning perspective. IEEE Signal Processing Magazine, 32(5):31-41, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.831, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[4] Fadi Boutros, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Elasticface: Elastic margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1578-1587, 2022." + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.633, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[5] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in Computer Vision, pages 671-679. Elsevier, 1987." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.136, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[6] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3981-3990, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.205, + 0.895, + 0.299 + ], + "angle": 0, + "content": "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.302, + 0.895, + 0.37 + ], + "angle": 0, + "content": "[8] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5799-5809, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.372, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[9] Forrester Cole, David Belanger, Dilip Krishnan, Aaron Sarna, Inbar Mosseri, and William T Freeman. Synthesizing normalized faces from facial identity features. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3703-3712, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[10] Debayan Deb, Jianbang Zhang, and Anil K Jain. Advfaces: Adversarial face synthesis. In Proceedings of the 2020 IEEE International Joint Conference on Biometrics (IJCB), pages 1-10. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[11] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.62 + ], + "angle": 0, + "content": "[12] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10673-10683, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[13] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13733-13742, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.693, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[14] Xingbo Dong, Zhe Jin, Zhenhua Guo, and Andrew Beng Jin Teoh. Towards generating high definition face images from deep templates. In 2021 International Conference of the Biometrics Special Interest Group (BIOSIG), pages 1-11. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[15] Xingbo Dong, Zhihui Miao, Lan Ma, Jiajun Shen, Zhe Jin, Zhenhua Guo, and Andrew Beng Jin Teoh. Reconstruct face from features using gan generator as a distribution constraint. arXiv preprint arXiv:2206.04295, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[16] Chi Nhan Duong, Thanh-Dat Truong, Khoa Luu, Kha Gia Quach, Hung Bui, and Kaushik Roy. Vec2face: Unveil human faces from their blackbox features in face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6132-6141, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19670" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[17] European Council. Regulation of the european parliament and of the council on the protection of individuals with regard to the processing of personal data and on the free movement of such data (general data protection regulation), April 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.161, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[18] Javier Galbally, Sébastien Marcel, and Julian Fierrez. Biometric antispoofing methods: A survey in face recognition. IEEE Access, 2:1530-1552, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[19] Javier Galbally, Chris McCool, Julian Fierrez, Sebastien Marcel, and Javier Ortega-Garcia. On the vulnerability of face verification systems to hill-climbing attacks. Pattern Recognition, 43(3):1027-1038, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.47, + 0.329 + ], + "angle": 0, + "content": "[20] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In Proceedings of the 10th International Conference on Learning Representations (ICLR), pages 1-25, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[21] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 87-102. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[22] Abdenour Hadid, Nicholas Evans, Sebastien Marcel, and Julian Fierrez. Biometrics systems under spoofing attack: an evaluation methodology and lessons learned. IEEE Signal Processing Magazine, 32(5):20-30, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.523 + ], + "angle": 0, + "content": "[23] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, October 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.579 + ], + "angle": 0, + "content": "[24] Yifan Jiang, Shiyu Chang, and Zhangyang Wang. Transgan: Two pure transformers can make one strong gan, and that can scale up. Advances in Neural Information Processing Systems, 34:14745-14758, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[25] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.69 + ], + "angle": 0, + "content": "[26] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[28] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8110-8119, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[29] Prabhjot Kaur, Nitin Kumar, and Maheep Singh. Biometric cryptosystems: a comprehensive survey. Multimedia Tools and Applications, pages 1-56, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conference on Learning Representations (ICLR), San Diego, California., USA, May 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[31] Nitin Kumar et al. Cancelable biometrics: a comprehensive survey. Artificial Intelligence Review, 53(5):3403-3446, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 10012-10022, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[33] Guangcan Mai, Kai Cao, Pong C Yuen, and Anil K Jain. On the reconstruction of face images from deep face templates. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(5):1188-1202, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.334 + ], + "angle": 0, + "content": "[34] Sébastien Marcel, Julian Fierrez, and Nicholas Evans. Handbook of Biometric Anti-Spoofing: Presentation Attack Detection and Vulnerability Assessment. Springer Nature, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.336, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[35] Sébastien Marcel, Mark S Nixon, Julian Fierrez, and Nicholas Evans. Handbook of biometric anti-spoofing: Presentation attack detection, volume 2. Springer, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.379, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[36] Chris McCool, Roy Wallace, Mitchell McLaren, Laurent El Shafey, and Sébastien Marcel. Session variability modelling for face authentication. IET Biometrics, 2(3):117-129, Sept. 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[37] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6351-6361, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.506, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[38] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405-421, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[39] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 51-59, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[40] Karthik Nandakumar and Anil K Jain. Biometric template protection: Bridging the performance gap between theory and practice. IEEE Signal Processing Magazine, 32(5):88-100, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[41] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11453-11464, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[42] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[43] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagan, and Andrea Tagliasacchi. Lolnerf: Learn from one look." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19671" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1558-1567, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[44] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[45] Hatef Otroshi Shahreza, Christian Rathgeb, Daile Osorio-Roig, Vedrana Krivokuca Hahn, Sébastien Marcel, and Christoph Busch. Hybrid protection of biometric templates by combining homomorphic encryption and cancelable biometrics. In Proceedings of the 2022 IEEE International Joint Conference on Biometrics (IJCB), pages 1-10. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.273, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.315, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[47] Mandavilli Srinivas and Lalit M Patnaik. Genetic algorithms: A survey. Computer, 27(6):17-26, 1994." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.469, + 0.397 + ], + "angle": 0, + "content": "[48] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. arXiv preprint arXiv:2205.15517, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.398, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[49] Thanh-Dat Truong, Chi Nhan Duong, Ngan Le, Marios Savvides, and Khoa Luu. Vec2face-v2: Unveil human faces from their blackbox features via attention-based network in face recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.453, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[50] Peter JM Van Laarhoven and Emile HL Aarts. Simulated annealing. In Simulated annealing: Theory and applications, pages 7-15. Springer, 1987." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.495, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[51] Edward Vendrow and Joshua Vendrow. Realistic face reconstruction from deep embeddings. In NeurIPS 2021 Workshop Privacy in Machine Learning, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.537, + 0.469, + 0.606 + ], + "angle": 0, + "content": "[52] Fei Wang, Mengqing Jiang, Chen Qian, Shuo Yang, Cheng Li, Honggang Zhang, Xiaogang Wang, and Xiaou Tang. Residual attention network for image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3156-3164, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.469, + 0.66 + ], + "angle": 0, + "content": "[53] Jun Wang, Yinglu Liu, Yibo Hu, Hailin Shi, and Tao Mei. Facex-zoo: A pytorch toolbox for face recognition. In Proceedings of the 29th ACM International Conference on Multimedia, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[54] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[55] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18430-18439, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.8, + 0.469, + 0.855 + ], + "angle": 0, + "content": "[56] Jichao Zhang, Enver Sangineto, Hao Tang, Aliaksandr Siarohin, Zhun Zhong, Nicu Sebe, and Wei Wang. 3d-aware semantic-guided generative model for human synthesis. arXiv preprint arXiv:2112.01422, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.855, + 0.469, + 0.896 + ], + "angle": 0, + "content": "[57] Andrey Zhmoginov and Mark Sandler. Inverting face embeddings with convolutional neural networks. arXiv preprint arXiv:1606.04189, 2016." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19672" + } + ] +] \ No newline at end of file diff --git a/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/203e5250-94d9-4833-ba8d-b9c548251baa_origin.pdf b/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/203e5250-94d9-4833-ba8d-b9c548251baa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bb885693a6b15bc0873ac78fe2f4f6747aec02f6 --- /dev/null +++ b/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/203e5250-94d9-4833-ba8d-b9c548251baa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d4755fefa3e321c4d6c0434b00ccf335e42522d26b38795aa384a38019417a1 +size 8653832 diff --git a/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/full.md b/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a2d741beab6336a557ad87b63dd3aa0ea42d1d89 --- /dev/null +++ b/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/full.md @@ -0,0 +1,369 @@ +# Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction + +Hatef Otroshi Shahreza1,2 and Sébastien Marcel1,3 + +1Idiap Research Institute, Martigny, Switzerland + +$^{2}$ École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland + +3Université de Lausanne (UNIL), Lausanne, Switzerland + +{hatef.otoreshi,sebastien.marcel}@idiap.ch + +![](images/035c2bb9e1ceeb4117f61cbc86a31e5035a4f7fd586f85e37b75393285451c54.jpg) +Original + +![](images/cecc708ec233d620428171b9e000ebf21ccaf531a092acf9d983fc467f701f54.jpg) + +![](images/ecc70f9ac35c83befec4b24257dafcf95775a163cae8da28ad74d8cc975475ea.jpg) + +![](images/d7727bb2f6da68c3858ae4978efe8b4b6bf8ef6b88c91ce352a598038377dec9.jpg) + +![](images/1d2ec5ae4f1bbe82a80c41e6fbdc8eba56146f333aa1ea41ac9fcc2a86820835.jpg) + +![](images/7c94f7064769039f3dbac5dd309f1f60960fb6f03a4c26d87f8e5bf38329dbb0.jpg) + +![](images/349efb905bf85293b34164fcbee27fab5a57ab20de571a4d08f86b6c01f0fac0.jpg) + +![](images/0a75ab9e6fdf7ff4bdfd80452416c75e18bb050109458d3fbea5500d29b8e0d3.jpg) +Frontal Reconst. + +![](images/181fc30daa5b5d459e824179e0c1aead580a31165d5fe7185fd661739e613fd1.jpg) +0.674 + +![](images/01aaeec64460585df091b5029d76fee21ded327899cf99a42e46e72b336fe15d.jpg) + +![](images/c9df9e6cc6054fed9b6a42a50849d994005c041ccdbee0108b9316701ac9f351.jpg) + +![](images/d662cba4694d056de21d1b95807f5ff3742cc7f41d6bad0d8b83c24d99568b14.jpg) + +![](images/4e8bfa93a90e7700d64587a8680a2c84118a9779a036e1e61712d19a0c4314dd.jpg) + +![](images/62db6d5e3e9c9ce1ff792919efb9aeebccba7ac3365a5163a7db481c2ad93048.jpg) + +![](images/d831fc20456ae67f432db5e803c609ea7b426e05b9ccf179539b88ae262503c1.jpg) +3D Reconst. +0.706 + +![](images/6d1266d882582bc90a7a3818750f5473eebcf87aeb0aae729916650ce16ba9b3.jpg) + +![](images/f886986be8b1e674a87979042bc046fc91daeb81de1a09e77cca9b0c672fa800.jpg) +0.705 + +![](images/4745ef3c430bd2398cdcabe4fcafedb0f6bbbf53a6675e563fec03a91057f7f6.jpg) +0.679 + +![](images/25807260325bf9ec372cb3cc2cc2c7628f18c5a2c929ca7350429094e965f8a3.jpg) +0.663 + +![](images/7d8dc0f040695b7a4c8a2e11c99d46f312d81b13878c00570d2be3f908884f67.jpg) +0.733 + +![](images/2af1d830e444c8df0aad2aab90167008b473d189f05594a80d95724162c1b0a7.jpg) +0.782 +Figure 1: Sample face images from the FFHQ dataset (first row) as well as their corresponding 3D (third row) and frontal 2D reconstruction (second row) from facial templates in the whitebox template inversion attack against ArcFace. Values show the cosine similarity between the templates of the original and frontal reconstructed face images. + +# Abstract + +Face recognition systems are increasingly being used in different applications. In such systems, some features (also known as embeddings or templates) are extracted from each face image. Then, the extracted templates are stored in the system's database during the enrollment stage and are later used for recognition. In this paper, we focus on template inversion attacks against face recognition systems and introduce a novel method (dubbed GaFaR) to reconstruct 3D face from facial templates. To this end, we use a geometry-aware generator network based on generative neural radiance fields (GNeRF), and learn a mapping from facial templates to the intermediate latent space of the generator network. We train our network with a semi-supervised learning approach using real and synthetic images simul + +taneously. For the real training data, we use a Generative Adversarial Network (GAN) based framework to learn the distribution of the latent space. For the synthetic training data, where we have the true latent code, we directly train in the latent space of the generator network. In addition, during the inference stage, we also propose optimization on the camera parameters to generate face images to improve the success attack rate (up to $17.14\%$ in our experiments). We evaluate the performance of our method in the whitebox and blackbox attacks against state-of-the-art face recognition models on the LFW and MOBIO datasets. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates. The project page is available at: https://www.idiap.ch/paper/gafar + +# 1. Introduction + +Automatic face recognition (FR) has become a well-known biometric authentication tool which has been widely used in different applications, including smartphone locks, border controls, etc. In such systems, generally, some features (also referred to as facial embeddings or facial templates) are extracted from users and stored in the system's database during the enrollment stage. In the recognition stage, either verification or identification, similar features are extracted from the user and are compared with the features in the system's database. Therefore, facial features play the main role in automatic face recognition systems and convey important information about users. + +Template inversion (TI) attack against FR systems refers to an attempt by an adversary to reconstruct face images from the templates stored in the system's database. Then, the adversary may find sensitive information about the enrolled users and also can use the reconstructed face image to impersonate and enter the system. Therefore, compared to most attacks on FR systems which threaten the security of the system [19, 3, 22, 18, 35, 34, 10], TI attack jeopardizes both security and privacy of users, and thus requires further study. In this paper, we introduce a novel method (called geometry-aware face reconstruction, shortly GaFaR) for the TI attack in FR systems to reconstruct a 3D face from facial features (extracted from 2D FR models). The 3D face reconstruction provides further information than 2D face reconstruction, and in particular can be used to generate face image from any pose to improve the attack to the FR system. To our knowledge, this is the first paper on 3D face reconstruction from facial templates. + +Recently, neural radiance fields (NeRF) [38] absorbed considerable attention in computer vision society due to remarkable results in the novel-view synthesis problem. Built upon NeRF, generative NeRF (GNeRF) methods such as [37, 44, 56, 6, 8, 41, 42, 20, 55, 7, 12, 48, 43] combine conditional NeRF with generative models for geometry-aware image generation tasks. In these methods, a generative model, such as a generative adversarial network (GAN), is used to embed the shape and appearance of an object in a latent space. Then, the latent code of the GAN, along with the camera parameters, are fed to a NeRF for the rendering process. Among GNeRF methods, there are several works, such as [8, 41, 42, 20, 55, 7, 12, 48], for geometry-aware 3D face generation, which can generate face images from different views. + +In our proposed TI method, we use a geometry-aware face generator network based on GNeRF, and learn a new mapping from facial templates to the intermediate latent space of the generator network. We train our network using real and synthesized images simultaneously with a semisupervised learning approach. For real training data where we do not have the corresponding latent code (i.e., unsuper + +vised), we use a GAN-based framework to learn the distribution of the latent space. For the synthetic training data where we have true values of latent codes (i.e., supervised), we directly learn the latent space. Because we have 3D reconstructed face, we can use any arbitrary pose to inject the the FR system for the attack. Therefore, during the inference stage, we also use optimization on the camera parameters to improve the attack against the FR system. We evaluate the performance of our proposed method in the whitebox (i.e., where the adversary knows the internal functioning and parameters of the FR model) and blackbox (i.e., where the adversary does not have information about the internal functioning of the FR model) attacks against state-of-the-art (SOTA) FR systems. For evaluation, we consider FR systems with the LFW [23] and MOBIO [36] datasets and evaluate our attack if the reconstructed face can be used to enter the system. Fig. 1 illustrates sample face images from FFHQ [26] dataset and their 3D reconstruction from ArcFace [11] templates using our attack. + +To elaborate on the contributions of this paper, we list them hereunder: + +- We propose a novel template inversion method to reconstruct 3D faces from the facial templates of a face recognition system. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates. +- We use a geometry-aware generator network based on GNeRF, and learn a mapping from facial templates to the intermediate latent space of the generator network. We train our mapping network using a GAN-based framework and a semi-supervised learning approach using real and synthetic face images. +- We use optimization on the camera parameters in the geometry-aware generator network during the inference stage. The optimization of camera parameters can find a pose that improves the success attack rate. + +The remainder of the paper is organized as follows. First, we review the related works in the literature in Sec. 2. Then, we present our proposed method in Sec. 3. Next, we describe our experiments in Sec. 4 and discuss limitations and ethical considerations in Sec. 5. Finally, the paper is concluded in Sec. 6. + +# 2. Related Works + +Generally, the methods in the literature for reconstructing face from facial templates generate 2D face images. In [57], authors proposed two methods based on optimization and learning to reconstruct 2D low-resolution face images from facial templates in the whitebox scenario. In the + +Table 1: Comparison with related works. + +
Ref.2D/3DResolutionWhitebox/BlackboxMethod BasisAvailable code
[57]2Dlowwhitebox1) optimization +2) learningX
[9]2Dlowboth*learningX
[33]2Dlowblackboxlearning✓
[16]2Dlowboth†learningX
[49]2Dlowboth†learningX
[1]2Dlowblackboxlearning + opt.X
[14]2Dhighblackboxlearning✓
[51]2Dhighblackboxoptimization✓
[15]2DhighblackboxoptimizationX
[Ours]3Dhighboth‡learning✓
+ +*The method is based on the whitebox attack, and is also applied in the blackbox scenario by removing a loss term that required the FR model. +The method is based on the whitebox attack, and is extended to the blackbox with knowledge distillation of the FR model. +$\ddagger$ The method is based on the whitebox attack, and is extended to blackbox using a different FR model. + +optimization-based method, they used a gradient ascent-based algorithm to reconstruct the face image, and optimized the generated image to minimize a multi-term loss function, including the $\ell_2$ distance between target templates and templates of the reconstructed face image. In addition, they applied total-variation and Laplacian pyramid gradient normalization [5] on the reconstructed image to generate a smooth image. In their learning-based method, they used a deconvolutional neural network to reconstruct face images and trained it with the same loss function that they used in their optimization-based method. + +In [9], a multi-layer perceptron (MLP) is trained to find facial landmark coordinates (optimized using the mean squared error) from the given template. They also trained a convolutional neural network (CNN) to generate face texture (optimized using mean absolute error) from the target template. Then, they used a differentiable warping to combine estimated landmarks (from MLP) and textures (from CNN) and reconstruct low-resolution face images. In the whitebox scenario, they further optimized MLP and CNN by minimizing the distance between templates of reconstructed and original face images. However, in the blackbox scenario, they trained MLP and CNN separately, and used warping in the inference only. + +In [33], two new deconvolutional networks are proposed, called NbBlock-A and NbBlock-B, and are trained with pixel loss $(\ell_{1}$ norm of reconstruction error) and perceptual loss (distance of middle layers of VGG-19[46] when given the reconstructed and original face images) to reconstruct low-resolution face images in the blackbox scenario. In [16] and [49], a same bijection-learning-based method is used to train GAN models with PO-GAN [25] and TransGAN [24] structure, respectively. The method is based on whitebox attack and for training the GAN model, authors also minimized the distance between target templates and templates + +extracted from the reconstructed face images using the FR model. In the blackbox attack, they proposed to use the distillation of knowledge to train a student network that mimics the target FR model and used the student network in their method. However, they did not report any detail (and no published source code) about the training of the student network (e.g., network structure, etc.). + +In [1], a 3-step method is proposed to reconstruct low-resolution face images in the blacbox scenario. First, they trained a GAN for general face generation. In the second step, they trained a MLP to map target templates to embeddings of a known FR model. In the final step, they found a latent code in the input of their generator (of their GAN) which generates a face image that maximizes two terms; the discriminator score (for being a real face image) and cosine similarity between the mapped embedding and the embeddings extracted by the known FR model. + +In contrast to most works in the literature which generate low-resolution face images, recently there have been few works to generate high-resolution 2D face images. In [51], a grid-search optimization using the simulated annealing [50] approach is used on the latent vector (i.e., input noise) of StyleGAN2 [28] to find latent codes that can generate face images which have templates similar to the target templates. However, their proposed method is computationally expensive, and they reported their evaluation on only 20 face images. In [15], a similar optimization to [51] on the latent vector of StyleGAN2 [28] is considered, but it is solved using the standard genetic algorithm [47]. In contrast to [51, 15] which are based on optimization, in [14], a learning-based method is proposed in the blackbox scenario. The authors proposed to generate some face images using StyleGAN2 [28] and extract the templates using the FR model. Then, they trained a MLP to map facial templates to the input latent codes of StyleGAN2 [28]. + +Tab. 1 compares our proposed method with the previous works in the literature. To our knowledge, our proposed method is the first work on 3D face reconstruction from facial templates (extracted from 2D face recognition models). Furthermore, unlike most works in the literature, our method generates high-resolution face images. Last but not least, our method can be used in both whitebox and black-box attacks against FR systems. + +# 3. Proposed Method + +In this paper, we consider a threat model as described in Sec. 3.1 and use the proposed 3D face reconstruction method (dubbed GaFaR) described in Sec. 3.2. After training and in the inference stage, we use an optimization on the camera parameters as described in Sec. 3.3 to generate a 2D face image with a pose that can lead to a higher success attack rate (SAR) and inject the face image into the system. Fig. 2 depicts the block-diagram of the proposed method. + +# 3.1. Threat Model + +We consider an attack against FR systems, where the adversary gains access to the system's database and aims to generate a 3D reconstruction of enrolled users. To this end, the adversary trains a network to reconstruct a 3D face given a target facial template. Then, the adversary can use the 3D reconstruction to impersonate users into the system (i.e., using a 3D face mask, etc.). For simplicity, we consider that the adversary can inject a 2D face image (from the 3D reconstructed face) into the system. Hence, the adversary uses an appropriate pose to generate a 2D face image for the attack. According to the adversary's knowledge of the FR model, we consider both whitebox and blackbox attacks against FR systems. In the whitebox scenario, we assume that the adversary has complete knowledge of the FR model from which the templates are leaked. However, in the blackbox scenario, we assume that the adversary does not know the internal functioning of the FR model and can only use the FR model to generate facial template for any face image. In addition, we assume that in the blackbox scenario, the adversary has access to another FR model and knows its internal parameters and functioning. + +# 3.2. 3D Face Reconstruction + +To reconstruct 3D faces from facial templates, we consider a geometry-aware face generator network based on GNeRF, such as EG3D [7], and learn a mapping $M_{\mathrm{GaFaR}}: \mathcal{T} \to \mathcal{W}$ from the facial templates $t \in \mathcal{T}$ to the intermediate latent space $\mathcal{W}$ of the GNeRF model. Then, we use the remainder of GNeRF model $G(.)$ to synthesize an image $\hat{I} = G(\hat{w}, c)$ from an arbitrary view using the mapped intermediate latent vector $\hat{w}$ and camera parameters $c$ . We consider a network with two fully-connected layers with the Leaky ReLU activation functions for our mapping network $M_{\mathrm{GaFaR}}$ . + +We train the mapping network $M_{\mathrm{GaFaR}}$ with a semi-supervised approach using synthetic and real data. For the real data, we consider a set of real face images $\{I_{\mathrm{real},i}\}_{i = 0}^{N}$ and extract the facial template $t_{\mathrm{real},i} = F(I_{\mathrm{real},i})$ from each face image $I_{\mathrm{real},i}$ using the FR model $F(.)$ . As we do not have the true value of intermediate latent space $\mathcal{W}$ of the GNeRF model to generate the real face image, we consider training our mapping network using this data as unsupervised learning. On the other hand, for the synthetic data, we use the pre-trained geometry-aware face generator network to generate a set of random face images $\{I_{\mathrm{syn},i}\}_{i = 0}^{M}$ . Therefore, as opposed to real data, we have the true value of intermediate latent space $\boldsymbol {w}\in \mathcal{W}$ to generate the same synthetic face image. Hence, we consider training our mapping network using the synthetic data as supervised learning. We train our mapping network simultaneously using real and synthetic training data as follows: + +Unsupervised learning using real data To train our mapping network $M_{\mathrm{GaFaR}}(.)$ with the real data, we use a GAN-based framework based on Wasserstein GAN (WGAN) [2] algorithm to learn the distribution of intermediate latent space $\mathcal{W}$ of the GNeRF model. In this framework, our mapping network $M_{\mathrm{GaFaR}}$ generates a latent code $\hat{\pmb{w}} = M_{\mathrm{GaFaR}}([n,t])$ using the facial template $\pmb{t}$ and a random vector $\pmb{n} \in \mathcal{N}$ . We can also generate real latent codes $\pmb{w} = M(\pmb{z}) \in \mathcal{W}$ for training our GAN using the GNeRF mapping function $M$ and a random vector $\pmb{z} \in \mathcal{Z}$ . Then, we can use a critic network $C(.)$ to score the latent codes generated by our mapping $M_{\mathrm{GaFaR}}(.)$ and GNeRF mapping $M(.)$ . Therefore, we can train the critic network $C(.)$ and our mapping $M_{\mathrm{GaFaR}}$ using the following loss functions: + +$$ +\mathcal {L} _ {C} ^ {\mathrm {W G A N}} = \mathbb {E} _ {\boldsymbol {w} \sim M (\boldsymbol {z})} [ C (\boldsymbol {w}) ] - \mathbb {E} _ {\hat {\boldsymbol {w}} \sim M _ {\mathrm {G a F a R}} ([ \boldsymbol {n}, t ])} [ C (\hat {\boldsymbol {w}}) ] \tag {1} +$$ + +$$ +\mathcal {L} _ {M _ {\mathrm {G a F a R}}} ^ {\mathrm {W G A N}} = \mathbb {E} _ {\hat {\boldsymbol {w}} \sim M _ {\mathrm {G a F a R}} ([ \boldsymbol {n}, \boldsymbol {t} ])} [ C (\hat {\boldsymbol {w}}) ] \tag {2} +$$ + +In addition to the WGAN training, we use the generated face image $\hat{\pmb{I}} = G(M_{\mathrm{GaFaR}}([n,t]),c)$ to optimize our mapping network using the following loss function: + +$$ +\mathcal {L} _ {\text {r e a l}} ^ {\text {r e c}} = \mathcal {L} ^ {\text {P i x e l}} + \mathcal {L} ^ {\text {I D}}, \tag {3} +$$ + +where $\mathcal{L}^{\mathrm{Pixel}}$ and $\mathcal{L}^{\mathrm{ID}}$ are pixel loss and ID loss, respectively, and are defined as: + +$$ +\mathcal {L} ^ {\text {P i x e l}} = \mathbb {E} _ {\hat {\boldsymbol {w}} \sim M _ {\mathrm {G a F a R}} ([ \boldsymbol {n}, \boldsymbol {t} ])} \left[ \| \boldsymbol {I} - G (\hat {\boldsymbol {w}}, \boldsymbol {c}) \| _ {2} ^ {2} \right] \tag {4} +$$ + +$$ +\mathcal {L} ^ {\mathrm {I D}} = \mathbb {E} _ {\hat {\boldsymbol {w}} \sim M _ {\mathrm {G a F a R}} ([ \boldsymbol {n}, \boldsymbol {t} ])} [ \| F _ {\text {l o s s}} (\boldsymbol {I}) - F _ {\text {l o s s}} (G (\hat {\boldsymbol {w}}, \boldsymbol {c})) \| _ {2} ^ {2} ] \tag {5} +$$ + +The pixel loss minimizes the pixel-level reconstruction error and ID loss helps the network to generate face images with a similar facial templates extracted by $F_{\mathrm{loss}}$ . To calculate the ID loss in Eq. (5), we use the same FR model as the one by which leaked templates are calculated (i.e., $F$ ) in the whitebox scenario. However, in the blackbox scenario, we use another FR model that we assume the adversary has access to. The critic network is a fully-connected network with three layers and Leaky ReLU activation function. + +Supervised learning using synthetic data To train our mapping network $M_{\mathrm{GaFaR}}(.)$ with synthetic data, we can directly learn the GNeRF intermediate latent code $\pmb {w} = M(z)$ . In addition to directly learning $\pmb{w}$ , we use the generated face image to train our mapping network by minimizing the following multi-term loss function: + +$$ +\mathcal {L} _ {\text {s y n}} ^ {\text {r e c}} = \mathcal {L} ^ {w} + \mathcal {L} ^ {\text {P i x e l}} + \mathcal {L} ^ {\text {I D}}, \tag {6} +$$ + +where $\mathcal{L}^{\mathrm{Pixel}}$ and $\mathcal{L}^{\mathrm{ID}}$ are pixel loss and ID loss, respectively. Furthermore, $\mathcal{L}^w$ is $w$ -loss to directly learn the latent space + +![](images/9f43988b9d9f5cd9ef97505a35d855d9e3c18c0d20c345fd25302e2beff1e390.jpg) +Training (Semi-supervised Learning) + +![](images/2c97c30b32e73f35907f3fc56a205f1323af9a9debc6756cec69c57e410bcc82.jpg) + +![](images/78af3af4a1872df784c28a6ce55670a117fdd4278f577e1dac6a8a9e90e4b7b5.jpg) +Inference +Figure 2: Block-diagram of the proposed method: In the training process, a semi-supervised approach is used to learn our mapping $M_{\mathrm{GaFaR}}$ (depicted with a green block) from the facial templates to the intermediate latent space of the GNeRF model using real data (unsupervised, where we don't have the corresponding latent code) and synthetic data (supervised, where we have the corresponding latent code $\boldsymbol{w}$ ), simultaneously. During the inference stage, the leaked template $t$ is given to our mapping network to find corresponding latent code $\hat{\boldsymbol{w}} = M_{\mathrm{GaFaR}}([n, t])$ . Then, $\hat{\boldsymbol{w}}$ along with camera parameters $c$ are given to the generator and renderer of GNeRF $G$ to generate a reconstructed face image $\hat{I} = G(\hat{\boldsymbol{w}}, c)$ . Finally, an optimization (either grid search or continuous optimization) is applied on two of the camera parameters, $\theta$ and $\psi$ , from $c$ , to find the best pose that minimizes the distance between leaked template $t$ and the template of reconstructed face image. + +by minimizing mean squared error between $\pmb{w}$ and $\hat{\pmb{w}} = M_{\mathrm{GaFaR}}([n,t])$ as follows: + +$$ +\mathcal {L} ^ {w} = \mathbb {E} _ {\boldsymbol {w} \sim M (\boldsymbol {z})} [ \| \boldsymbol {w} - M _ {\mathrm {G a F a R}} ([ \boldsymbol {n}, \boldsymbol {t} ]) \| _ {2} ^ {2} ] \tag {7} +$$ + +# 3.3. Camera Parameters Optimization + +During the inference stage, after generating a 3D reconstruction of face using our method described in Sec. 3.2, we optimize the camera parameters to find a pose that increases + +the SAR. Among different camera parameters $c$ , the parameters which correspond to the camera rotation can change the pose of the generated face image. Note that by varying the camera rotations, we aim to change the pitch and yaw rotations of the 3D reconstructed face and do not modify the roll rotation. In fact, the roll rotation of the face will be eliminated through face alignment as a preprocess prior to feature extraction in the FR system. To optimize camera parameters, we consider two different approaches as follows: + +Grid Search (GS) For the grid search, we consider predefined steps to change the camera pitch $\theta \in \Theta$ and yaw $\psi \in \Psi$ rotations of camera parameters $c$ . We generate the 2D face image for all values of camera rotation steps ( $\theta_{\mathrm{step}}$ and $\psi_{\mathrm{step}}$ ) and find the facial template for each generated image. Then, we select the face image $\hat{\pmb{I}} = G(M_{\mathrm{GaFaR}}([n,t]),c)$ that has a template $\hat{\pmb{t}} = F(\hat{\pmb{I}})$ with the minimum mean squared error with the target template $\pmb{t}$ . Note that the grid search can be applied in both whitebox and blackbox scenarios using the FR model $F$ . + +Continuous Optimization (CO) In our continuous optimization approach, we start from the frontal image and use Adam [30] optimizer to solve the following minimization using $\hat{\boldsymbol{w}} = M_{\mathrm{GaFaR}}([\boldsymbol{n},\boldsymbol{t}])$ + +$$ +\min _ {\theta , \psi} \| F (G (\hat {w}, c)), t \| _ {2} ^ {2}, \tag {8} +$$ + +In this optimization, we find the parameters $\theta$ and $\psi$ for camera rotation that lead to a face image with templates close to the target template $t$ . In contrast to the grid search approach, the continuous optimization can be applied only in the whitebox scenario. + +# 4. Experiments + +# 4.1. Experimental Setup + +Face Recognition models In our experiments, we consider different SOTA FR models including ArcFace [11], ElasticFace [4] as well as four different FR models with SOTA backbones from FaceX-Zoo [53], including AttentionNet [52], HRNet [54], RepVGG [13] and Swin [32]. The recognition performance of these models are reported in the supplementary material. + +Datasets All the mentioned FR models are trained on the MS-Celeb1M dataset [21]. To train the face reconstruction network, we assume that adversary does not have any knowledge of the dataset used in training FR model, and uses a different dataset. To this end, we consider the Flickr-Faces-HQ (FFHQ) dataset [27], which consists of 70,000 high-quality images crawled from internet (with no identity label), for training our face reconstruction network. We randomly split FFHQ into train $(90\%)$ and test $(10\%)$ set. + +To evaluate the vulnerability of FR models, we use two other datasets, including Labeled Faces in the Wild (LFW) [23] and MOBIO [36] datasets. The LFW dataset consists of 13,233 face images of 5,749 people collected from internet, where 1,680 people have two or more images. The MOBIO dataset includes face images of 150 people captured using mobile devices in 12 sessions for each person1. + +![](images/1f24a996c9116d06dcf1064e78bab90b4b34d8ae47041e82063a5d6a8f834af9.jpg) +rnnnne + +![](images/c707859e073481a45313f586e169e175f0760c39725f52398a9790b8ce0bb33f.jpg) + +![](images/fb4b809289f0232e9642d9ba5ef6b2b1c1505cf40bb9de299b37e6d986b38c49.jpg) + +![](images/e9f98b45067148fedd9851fde2af3188b7cb5eeffe1223495452bb80f76a66b6.jpg) + +![](images/10f20f965b6c96984ed8ed638edb789860477b0768a37fa3266f4d4c9c936255.jpg) +Reconnst +0.695 +Figure 3: Sample face images from the FFHQ dataset (first row) and their corresponding frontal (second row) reconstructed face images using our method in the blackbox attack against ElasticFace using ArcFace for $F_{\mathrm{loss}}$ . The values show the cosine similarity between templates of original and frontal reconstructed face images. + +![](images/50def58c439c1660b2904b3edf74d78dcd9102a8ee6f86f3e39473f9bc45a74f.jpg) +0.721 + +![](images/7541912d6a86061a538840da7c897aa8959125de47b879f4b2eba1c74cac88a0.jpg) +0.699 + +![](images/e984cb56fba50a0634111494671e4142c7a5b44dfc3107a7a020d93fe9f80500.jpg) +0.732 + +Evaluation Protocol For evaluation with each of the LFW and MOBIO datasets, we build a separate FR system, and use the reference templates (i.e., enrolled in the system's database) as input to our face reconstruction method. Then, we inject the reconstructed face image as a query to the system and evaluate the adversary's success attack rate (SAR) to enter the FR system. We should note that for each image, we use only one reconstructed face image as a query to the FR system and evaluate the SAR. + +Implementation Details and Source Code In our experiments, we use the grid search (in both whitebox and black-box attacks) and continuous optimization (in whitebox attacks only) to optimize camera parameters, as described in Sec. 3.3, in the inference stage. For grid search, we consider $\psi \in [-45^{\circ}, +45^{\circ}]$ and $\theta \in [-30^{\circ}, +30^{\circ}]$ for a $11 \times 11$ grid with step sizes of $\psi_{\mathrm{step}} = 9^{\circ}$ and $\theta_{\mathrm{step}} = 6^{\circ}$ . In the continuous optimization approach, we use 121 iterations of Adam optimizer [30] with the learning rate of $10^{-2}$ . An ablation study on these hyperparameters and the corresponding execution times are reported in the supplementary material. All our models (for different FR models and different scenarios) are trained in 15 epochs (each takes around 2 days) on a system equipped with a single NVIDIA RTX 3090 GPU. For GNeRF, we used a pretrained model of EG3D [7] in our experiments and generate $512 \times 512$ resolution face images. The input noise $z$ to GNeRF model has 512 dimensions and the noise $n$ in the input of our mapping network $M_{\mathrm{GaFaR}}$ has 16 dimensions. For unsupervised learning using real face images (FFHQ), we assume that camera parameters for real face images as frontal view. The source code of our experiments are publicly available2. + +Table 2: Evaluation of blackbox attack against SOTA FR models at systems' FMR=10 $^{-3}$ on the LFW and MOBIO datasets in terms of success attack rate (SAR). For attacks using our proposed method, we use ArcFace and also ElasticFace as $F_{\mathrm{loss}}$ to calculate the ID loss in Eq. (5). The values are in percentage. + +
LFWMOBIO
ArcFaceElsFaceAtt.NetHRNetRepVGGSwinArcFaceEls.FaceAtt.NetHRNetRepVGGSwin
NBNetA-M [33]4.3210.901.241.601.133.8202.380000
NBNetA-P [33]16.8326.980.661.445.729.704.7616.190.48014.297.14
NBNetB-M [33]10.9821.443.224.473.2111.231.903.803.337.143.338.57
NBNetB-P [33]40.2658.1616.2918.4215.2440.7615.2443.8131.9026.6723.8144.29
Dong et al. [14]13.2112.613.904.073.2212.383.338.1010.486.679.053.33
Vendrow and Vendrow [51]57.7053.0321.1218.859.6246.8429.0543.8127.1426.6720.9545.24
[Floss= Els.Face] GaFaR51.78-18.0711.6811.6347.1547.62-54.2943.3345.7170.48
[Floss= Els.Face] GaFaR + GS61.56-23.5617.2115.8254.0864.76-63.8155.2358.5782.38
[Floss= ArcFace] GaFaR-74.5433.5937.8025.4067.11-84.7672.3876.6772.8689.05
[Floss= ArcFace] GaFaR + GS-78.6738.4243.2729.8470.82-86.6280.0083.8073.3393.33
+ +Table 3: Evaluation of whitebox attack against SOTA FR models at systems' FMR = 10 $^{-3}$ on the LFW and MOBIO datasets in terms of success attack rate (SAR). The values are in percentage. + +
LFWMOBIO
ArcFaceElsFaceAtt.NetHRNetRepVGGSwinArcFaceEls.FaceAtt.NetHRNetRepVGGSwin
GaFaR79.8462.3227.0031.8717.3374.0882.8681.4364.2971.4353.8194.76
GaFaR + GS82.5268.7132.4237.4819.8376.0685.2382.3870.4781.4259.5294.29
GaFaR + CO84.2571.6934.5339.5320.4477.3689.0584.2975.7181.4362.8695.71
+ +# 4.2. Analyze + +Blackbox Scenario To evaluate the proposed method in the blackbox attack against SOTA FR models, we consider ArcFace and also ElasticFace as $F_{\mathrm{loss}}$ to calculate the ID loss in Eq. (5). Tab. 2 compares the performance of the proposed method with blackbox face reconstruction methods in the literature in terms of SAR for systems configured at false match rate (FMR) of $10^{-3}$ . Similar results for $\mathrm{FMR} = 10^{-2}$ are reported in the supplementary material. As this table shows, the frontal reconstruction by our method (i.e., GaFaR) achieves superior performance than previous methods in the literature. Furthermore, camera parameter optimization (i.e., GaFaR+GS) improves the performance of our attack up to $17.14\%$ compared to our frontal face reconstruction (i.e., GaFaR). Comparing the use of ArcFace and ElasticFace as $F_{\mathrm{loss}}$ , the performance of attacks with the ArcFace model is better. This is due to the fact that ArcFace has a higher recognition accuracy than ElasticFace. Fig. 3 shows sample face images reconstructed from ElasticFace templates in the blackbox attack (using ArcFace for $F_{\mathrm{loss}}$ ). + +Whitebox Scenario Tab. 3 reports the performance of the proposed method in the whitebox attack $^3$ in terms of SAR for FR systems configured at false match rate (FMR) of $10^{-3}$ . Similar results for $\mathrm{FMR} = 10^{-2}$ are reported in the supplementary material. According to this table, all + +![](images/df1e69460daef7d142695376886637ae42161f20d045057589ed9e39f2fc87ad.jpg) +Figure 4: Sample face images from the FFHQ dataset (first row) and their corresponding frontal (second row) and camera parameters sub-grid (third row) reconstructed face images using our method in the whitebox attack against ArcFace. The values show the cosine similarity between templates of original and frontal reconstructed face images. + +these FR models are highly vulnerable to our attack. Furthermore, the camera parameter optimization improves the performance of GaFaR. Comparing grid search with contin + +Table 4: Ablation study on the proposed semi-supervised learning approach and evaluation of loss terms in the white-box attack against ArcFace model in terms of success attack rate (SAR) on the LFW and MOBIO datasets. The SAR values are in percentage and for an attack without any camera parameter optimization (GS/CO). + +
approachLoss FunctionsLFWMOBIO
FMR=10-2FMR=10-3FMR=10-2FMR=10-3
supervisedLrecsyn = Lw + LPixel + LID83.8069.46790.9682.38
Lrecsyn = Lw + LPixel31.7513.9243.818.57
Lrecsyn = Lw + LID0.860.3000
Lrecsyn = Lw33.6915.4332.389.52
unsupervisedLrecreal = LPixel + LID[without WGAN]0.440.1500
Lrecreal = LPixel + LID[with WGAN]67.7245.7670.4831.90
Lrecreal = LID[with WGAN]54.5130.8352.8619.52
Lrecreal = LPixel[with WGAN]2.210.4000
semi-supervisedEqs. (1) to (3) and (6)89.2779.8495.7182.86
+ +uous optimization, results show that with the same number of iterations continuous optimization achieves better performance. In addition, comparing these results with recognition performance of FR models (available in the supplementary material), we conclude that models with higher recognition accuracy are more vulnerable to our attack. Comparing results in Tab. 2 and Tab. 3, we observe that when we use ArcFace as $F_{\mathrm{loss}}$ in our loss function, blackbox attacks achieve better results than whitebox attacks for most cases, which can be explained considering the superior recognition performance of ArcFace than other FR models. Fig. 4 shows sample face images reconstructed from ArcFace templates in the whitebox attack. This figure also presents a grid of reconstructed face images with different poses4. + +Ablation Study We evaluate the effect of our semi-supervised learning approach in our proposed method compared to fully supervised learning and fully unsupervised learning approach. In each case, we evaluate the effect of each loss function too. Furthermore, for the fully unsupervised learning approach, we evaluate the effect of GAN learning which we used in our method. Tab. 4 reports our ablation study in whitebox attack against ArcFace model on the LFW and MOBIO datasets in terms of SAR at system's FMR of $10^{-2}$ and $10^{-3}$ . As these results show, the proposed semi-supervised approach achieves a better performance than fully supervised learning and fully unsupervised learning approaches. Furthermore, each of our loss terms has an important effect on the performance. In particular, using WGAN for real data, where we don't have the true value of intermediate latent codes, helps the network to + +![](images/8798e7321e78a0f6137cc0e68b20383a95ebb3729cf9aad6c5a7c0d4a4a670d3.jpg) +rannnnnne + +![](images/7337d7108f4ee44f027eccc3aa8c24f62312e0e442758f71d5d60127df74f917.jpg) + +![](images/f57540adc2b0420978fcf657260012b7b3cf1d24128dd36543888a965188075d.jpg) + +![](images/3b927cab3b27e2f2aacfd11caaa983a6a24706d20e95730cfee34611914f8213.jpg) + +![](images/803cab6322cd82db265021e32e77709b1e57ecf566313ce1c4e2519526820914.jpg) +Reconstrn +0.160 +Figure 5: Sample failure cases from the LFW dataset (first row) and their corresponding frontal (second row) reconstructed face images using our method in the whitebox attack against ArcFace. The values below each image show the cosine similarity between templates of original and reconstructed face images. + +![](images/50d2376b518d512e60618ff6a319cb0ffb8b76af926b318f5acdb9405c358d47.jpg) +0.150 + +![](images/04e7586916a2ad3d75191f701c0fe01cf55719767c7e17c01bb03ee9ea1308e4.jpg) +0.170 + +![](images/03a9aa130a829ca641bcf4f43d5a220c03e43fecb9d6986266616663472ad07c.jpg) +0.160 + +learn the distribution of GNeRF intermediate latent space $\mathcal{W}$ . Otherwise, the generated latent code by the mapping network will be out of distribution, and therefore generator part of GNeRF will not generate face-like images. + +# 5. Discussion + +Limitations In spite of the considerable success attack rate of our method in our experiments, the reconstructed face images from some facial templates fail to attack the system. Fig. 5 illustrates sample failed cases in the whitebox attack against ArcFace. As these sample images show, the failed reconstructions reflect a bias in the generated faces for dark skin and also old people. Such a bias in final results can be caused by inherent biases of each of the datasets used to train the FR model, GNeRF model, and our face reconstruction model. In addition to bias, comparing attacks against different FR models in Tabs. 2 and 3 shows that FR models with worse recognition performance have lower SAR and more failed cases. + +Ethical considerations This work is proposing a new method for a TI attack against FR models. The 3D reconstructed face can be used to generate 3D face masks or 2D printed photographs for presentation attacks against FR models. In addition to the security threats, the reconstructed face images can reveal important privacy-sensitive information of enrolled users, such as age, gender, ethnicity, etc. We do not condone using our work with the intent of attack to real FR systems. As a matter of fact, the main motivation for this work is to show such a vulnerability in the FR systems, and to encourage the scientific community to develop and propose the next generation of safe and protected FR systems. Along the same lines, data protection regulations, such as the European Union General Data Protection Regulation (EU-GDPR) [17], consider biometric data as sensi + +tive information, and put legal obligations to protect them. To this end and to mitigate such threats, several biometric template protection schemes are also proposed in the literature [40, 29, 31, 45]. We should also note that the project on which the work has been conducted has passed an Internal Ethical Review Board (IRB). + +# 6. Conclusion + +In this paper, we proposed a new method (dubbed GaFaR) to generate 3D face reconstruction from facial templates for a TI attack against FR models. We used a geometry-aware face generation network based on GNeRF and trained a mapping from facial templates to the intermediate latent space of the GNeRF model using a semi-supervised learning approach. To train our model, we used synthetic and real face images. For the synthetic training data, we had the latent code of each face image and could train our mapping with supervised learning. For the real training data, we used a GAN-based framework to learn the distribution of latent space. In the inference stage, we used optimization on the camera parameters to find the pose which increases the success attack rate. We evaluated our method on the whitebox and blackbox attacks against SOTA FR models on the LFW and MOBIO datasets. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates (extracted from 2D face recognition models). + +# Acknowledgment + +This research is based upon work supported by the H2020 TReSPAsS-ETN Marie Skłodowska-Curie early training network (grant agreement 860813). + +# References + +[1] Muku Akasaka, Soshi Maeda, Yuya Sato, Masakatsu Nishigaki, and Tetsushi Ohki. Model-free template reconstruction attack with feature converter. In 2022 International Conference of the Biometrics Special Interest Group (BIOSIG), pages 1-5. IEEE, 2022. +[2] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Proceedings of the International Conference on Machine Learning (ICML), pages 214-223. PMLR, 2017. +[3] Battista Biggio, Paolo Russu, Luca Didaci, Fabio Roli, et al. Adversarial biometric recognition: A review on biometric system security from the adversarial machine-learning perspective. IEEE Signal Processing Magazine, 32(5):31-41, 2015. +[4] Fadi Boutros, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Elasticface: Elastic margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1578-1587, 2022. + +[5] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in Computer Vision, pages 671-679. Elsevier, 1987. +[6] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3981-3990, 2022. +[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, 2022. +[8] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5799-5809, 2021. +[9] Forrester Cole, David Belanger, Dilip Krishnan, Aaron Sarna, Inbar Mosseri, and William T Freeman. Synthesizing normalized faces from facial identity features. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3703-3712, 2017. +[10] Debayan Deb, Jianbang Zhang, and Anil K Jain. Advfaces: Adversarial face synthesis. In Proceedings of the 2020 IEEE International Joint Conference on Biometrics (IJCB), pages 1-10. IEEE, 2020. +[11] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. +[12] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10673-10683, 2022. +[13] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13733-13742, 2021. +[14] Xingbo Dong, Zhe Jin, Zhenhua Guo, and Andrew Beng Jin Teoh. Towards generating high definition face images from deep templates. In 2021 International Conference of the Biometrics Special Interest Group (BIOSIG), pages 1-11. IEEE, 2021. +[15] Xingbo Dong, Zhihui Miao, Lan Ma, Jiajun Shen, Zhe Jin, Zhenhua Guo, and Andrew Beng Jin Teoh. Reconstruct face from features using gan generator as a distribution constraint. arXiv preprint arXiv:2206.04295, 2022. +[16] Chi Nhan Duong, Thanh-Dat Truong, Khoa Luu, Kha Gia Quach, Hung Bui, and Kaushik Roy. Vec2face: Unveil human faces from their blackbox features in face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6132-6141, 2020. + +[17] European Council. Regulation of the european parliament and of the council on the protection of individuals with regard to the processing of personal data and on the free movement of such data (general data protection regulation), April 2016. +[18] Javier Galbally, Sébastien Marcel, and Julian Fierrez. Biometric antispoofing methods: A survey in face recognition. IEEE Access, 2:1530-1552, 2014. +[19] Javier Galbally, Chris McCool, Julian Fierrez, Sebastien Marcel, and Javier Ortega-Garcia. On the vulnerability of face verification systems to hill-climbing attacks. Pattern Recognition, 43(3):1027-1038, 2010. +[20] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In Proceedings of the 10th International Conference on Learning Representations (ICLR), pages 1-25, 2022. +[21] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 87-102. Springer, 2016. +[22] Abdenour Hadid, Nicholas Evans, Sebastien Marcel, and Julian Fierrez. Biometrics systems under spoofing attack: an evaluation methodology and lessons learned. IEEE Signal Processing Magazine, 32(5):20-30, 2015. +[23] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, October 2007. +[24] Yifan Jiang, Shiyu Chang, and Zhangyang Wang. Transgan: Two pure transformers can make one strong gan, and that can scale up. Advances in Neural Information Processing Systems, 34:14745-14758, 2021. +[25] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. +[26] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019. +[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019. +[28] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8110-8119, 2020. +[29] Prabhjot Kaur, Nitin Kumar, and Maheep Singh. Biometric cryptosystems: a comprehensive survey. Multimedia Tools and Applications, pages 1-56, 2022. +[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International + +Conference on Learning Representations (ICLR), San Diego, California., USA, May 2015. +[31] Nitin Kumar et al. Cancelable biometrics: a comprehensive survey. Artificial Intelligence Review, 53(5):3403-3446, 2020. +[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 10012-10022, 2021. +[33] Guangcan Mai, Kai Cao, Pong C Yuen, and Anil K Jain. On the reconstruction of face images from deep face templates. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(5):1188-1202, 2018. +[34] Sébastien Marcel, Julian Fierrez, and Nicholas Evans. Handbook of Biometric Anti-Spoofing: Presentation Attack Detection and Vulnerability Assessment. Springer Nature, 2023. +[35] Sébastien Marcel, Mark S Nixon, Julian Fierrez, and Nicholas Evans. Handbook of biometric anti-spoofing: Presentation attack detection, volume 2. Springer, 2019. +[36] Chris McCool, Roy Wallace, Mitchell McLaren, Laurent El Shafey, and Sébastien Marcel. Session variability modelling for face authentication. IET Biometrics, 2(3):117-129, Sept. 2013. +[37] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6351-6361, 2021. +[38] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405-421, 2020. +[39] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 51-59, 2017. +[40] Karthik Nandakumar and Anil K Jain. Biometric template protection: Bridging the performance gap between theory and practice. IEEE Signal Processing Magazine, 32(5):88-100, 2015. +[41] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11453-11464, 2021. +[42] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022. +[43] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagan, and Andrea Tagliasacchi. Lolnerf: Learn from one look. + +In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1558-1567, 2022. +[44] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. +[45] Hatef Otroshi Shahreza, Christian Rathgeb, Daile Osorio-Roig, Vedrana Krivokuca Hahn, Sébastien Marcel, and Christoph Busch. Hybrid protection of biometric templates by combining homomorphic encryption and cancelable biometrics. In Proceedings of the 2022 IEEE International Joint Conference on Biometrics (IJCB), pages 1-10. IEEE, 2022. +[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. +[47] Mandavilli Srinivas and Lalit M Patnaik. Genetic algorithms: A survey. Computer, 27(6):17-26, 1994. +[48] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. arXiv preprint arXiv:2205.15517, 2022. +[49] Thanh-Dat Truong, Chi Nhan Duong, Ngan Le, Marios Savvides, and Khoa Luu. Vec2face-v2: Unveil human faces from their blackbox features via attention-based network in face recognition, 2022. +[50] Peter JM Van Laarhoven and Emile HL Aarts. Simulated annealing. In Simulated annealing: Theory and applications, pages 7-15. Springer, 1987. +[51] Edward Vendrow and Joshua Vendrow. Realistic face reconstruction from deep embeddings. In NeurIPS 2021 Workshop Privacy in Machine Learning, 2021. +[52] Fei Wang, Mengqing Jiang, Chen Qian, Shuo Yang, Cheng Li, Honggang Zhang, Xiaogang Wang, and Xiaou Tang. Residual attention network for image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3156-3164, 2017. +[53] Jun Wang, Yinglu Liu, Yibo Hu, Hailin Shi, and Tao Mei. Facex-zoo: A pytorch toolbox for face recognition. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. +[54] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. +[55] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18430-18439, 2022. +[56] Jichao Zhang, Enver Sangineto, Hao Tang, Aliaksandr Siarohin, Zhun Zhong, Nicu Sebe, and Wei Wang. 3d-aware semantic-guided generative model for human synthesis. arXiv preprint arXiv:2112.01422, 2021. +[57] Andrey Zhmoginov and Mark Sandler. Inverting face embeddings with convolutional neural networks. arXiv preprint arXiv:1606.04189, 2016. \ No newline at end of file diff --git a/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/images.zip b/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..df1a9a19b40b52f48a6e82d9c8d8edbffb3e7a59 --- /dev/null +++ b/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57d5577b5946d1e57c4bbaa45532a553105d38c741c5dd01a06c0fb6c2c7e782 +size 687841 diff --git a/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/layout.json b/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..64c3b9c9c28b4457120bc94fec45f7f4e27f409f --- /dev/null +++ b/2023/Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction/layout.json @@ -0,0 +1,10667 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 486, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 486, + 140 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 486, + 140 + ], + "type": "text", + "content": "Template Inversion Attack against Face Recognition Systems using 3D Face Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 160, + 418, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 160, + 418, + 175 + ], + "spans": [ + { + "bbox": [ + 175, + 160, + 418, + 175 + ], + "type": "text", + "content": "Hatef Otroshi Shahreza1,2 and Sébastien Marcel1,3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 180, + 178, + 414, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 178, + 414, + 192 + ], + "spans": [ + { + "bbox": [ + 180, + 178, + 414, + 192 + ], + "type": "text", + "content": "1Idiap Research Institute, Martigny, Switzerland" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 193, + 480, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 193, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 114, + 193, + 480, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 114, + 193, + 480, + 206 + ], + "type": "text", + "content": "École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 160, + 207, + 434, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 207, + 434, + 220 + ], + "spans": [ + { + "bbox": [ + 160, + 207, + 434, + 220 + ], + "type": "text", + "content": "3Université de Lausanne (UNIL), Lausanne, Switzerland" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 186, + 224, + 407, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 224, + 407, + 236 + ], + "spans": [ + { + "bbox": [ + 186, + 224, + 407, + 236 + ], + "type": "text", + "content": "{hatef.otoreshi,sebastien.marcel}@idiap.ch" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 62, + 252, + 130, + 320 + ], + "blocks": [ + { + "bbox": [ + 47, + 274, + 59, + 305 + ], + "lines": [ + { + "bbox": [ + 47, + 274, + 59, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 59, + 305 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 8, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 252, + 130, + 320 + ], + "lines": [ + { + "bbox": [ + 62, + 252, + 130, + 320 + ], + "spans": [ + { + "bbox": [ + 62, + 252, + 130, + 320 + ], + "type": "image", + "image_path": "035c2bb9e1ceeb4117f61cbc86a31e5035a4f7fd586f85e37b75393285451c54.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 132, + 252, + 199, + 321 + ], + "blocks": [ + { + "bbox": [ + 132, + 252, + 199, + 321 + ], + "lines": [ + { + "bbox": [ + 132, + 252, + 199, + 321 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 199, + 321 + ], + "type": "image", + "image_path": "cecc708ec233d620428171b9e000ebf21ccaf531a092acf9d983fc467f701f54.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 200, + 253, + 268, + 321 + ], + "blocks": [ + { + "bbox": [ + 200, + 253, + 268, + 321 + ], + "lines": [ + { + "bbox": [ + 200, + 253, + 268, + 321 + ], + "spans": [ + { + "bbox": [ + 200, + 253, + 268, + 321 + ], + "type": "image", + "image_path": "ecc70f9ac35c83befec4b24257dafcf95775a163cae8da28ad74d8cc975475ea.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 269, + 253, + 337, + 321 + ], + "blocks": [ + { + "bbox": [ + 269, + 253, + 337, + 321 + ], + "lines": [ + { + "bbox": [ + 269, + 253, + 337, + 321 + ], + "spans": [ + { + "bbox": [ + 269, + 253, + 337, + 321 + ], + "type": "image", + "image_path": "d7727bb2f6da68c3858ae4978efe8b4b6bf8ef6b88c91ce352a598038377dec9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 338, + 253, + 405, + 321 + ], + "blocks": [ + { + "bbox": [ + 338, + 253, + 405, + 321 + ], + "lines": [ + { + "bbox": [ + 338, + 253, + 405, + 321 + ], + "spans": [ + { + "bbox": [ + 338, + 253, + 405, + 321 + ], + "type": "image", + "image_path": "1d2ec5ae4f1bbe82a80c41e6fbdc8eba56146f333aa1ea41ac9fcc2a86820835.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 407, + 253, + 474, + 321 + ], + "blocks": [ + { + "bbox": [ + 407, + 253, + 474, + 321 + ], + "lines": [ + { + "bbox": [ + 407, + 253, + 474, + 321 + ], + "spans": [ + { + "bbox": [ + 407, + 253, + 474, + 321 + ], + "type": "image", + "image_path": "7c94f7064769039f3dbac5dd309f1f60960fb6f03a4c26d87f8e5bf38329dbb0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 476, + 253, + 544, + 321 + ], + "blocks": [ + { + "bbox": [ + 476, + 253, + 544, + 321 + ], + "lines": [ + { + "bbox": [ + 476, + 253, + 544, + 321 + ], + "spans": [ + { + "bbox": [ + 476, + 253, + 544, + 321 + ], + "type": "image", + "image_path": "349efb905bf85293b34164fcbee27fab5a57ab20de571a4d08f86b6c01f0fac0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 62, + 323, + 129, + 390 + ], + "blocks": [ + { + "bbox": [ + 47, + 327, + 59, + 388 + ], + "lines": [ + { + "bbox": [ + 47, + 327, + 59, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 59, + 388 + ], + "type": "text", + "content": "Frontal Reconst." + } + ] + } + ], + "index": 16, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 323, + 129, + 390 + ], + "lines": [ + { + "bbox": [ + 62, + 323, + 129, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 129, + 390 + ], + "type": "image", + "image_path": "0a75ab9e6fdf7ff4bdfd80452416c75e18bb050109458d3fbea5500d29b8e0d3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 131, + 323, + 198, + 390 + ], + "blocks": [ + { + "bbox": [ + 131, + 323, + 198, + 390 + ], + "lines": [ + { + "bbox": [ + 131, + 323, + 198, + 390 + ], + "spans": [ + { + "bbox": [ + 131, + 323, + 198, + 390 + ], + "type": "image", + "image_path": "181fc30daa5b5d459e824179e0c1aead580a31165d5fe7185fd661739e613fd1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 396, + 176, + 404 + ], + "lines": [ + { + "bbox": [ + 153, + 396, + 176, + 404 + ], + "spans": [ + { + "bbox": [ + 153, + 396, + 176, + 404 + ], + "type": "text", + "content": "0.674" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 200, + 323, + 268, + 390 + ], + "blocks": [ + { + "bbox": [ + 200, + 323, + 268, + 390 + ], + "lines": [ + { + "bbox": [ + 200, + 323, + 268, + 390 + ], + "spans": [ + { + "bbox": [ + 200, + 323, + 268, + 390 + ], + "type": "image", + "image_path": "01aaeec64460585df091b5029d76fee21ded327899cf99a42e46e72b336fe15d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 269, + 323, + 336, + 390 + ], + "blocks": [ + { + "bbox": [ + 269, + 323, + 336, + 390 + ], + "lines": [ + { + "bbox": [ + 269, + 323, + 336, + 390 + ], + "spans": [ + { + "bbox": [ + 269, + 323, + 336, + 390 + ], + "type": "image", + "image_path": "c9df9e6cc6054fed9b6a42a50849d994005c041ccdbee0108b9316701ac9f351.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 338, + 323, + 405, + 390 + ], + "blocks": [ + { + "bbox": [ + 338, + 323, + 405, + 390 + ], + "lines": [ + { + "bbox": [ + 338, + 323, + 405, + 390 + ], + "spans": [ + { + "bbox": [ + 338, + 323, + 405, + 390 + ], + "type": "image", + "image_path": "d662cba4694d056de21d1b95807f5ff3742cc7f41d6bad0d8b83c24d99568b14.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 407, + 323, + 474, + 390 + ], + "blocks": [ + { + "bbox": [ + 407, + 323, + 474, + 390 + ], + "lines": [ + { + "bbox": [ + 407, + 323, + 474, + 390 + ], + "spans": [ + { + "bbox": [ + 407, + 323, + 474, + 390 + ], + "type": "image", + "image_path": "4e8bfa93a90e7700d64587a8680a2c84118a9779a036e1e61712d19a0c4314dd.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 476, + 323, + 544, + 390 + ], + "blocks": [ + { + "bbox": [ + 476, + 323, + 544, + 390 + ], + "lines": [ + { + "bbox": [ + 476, + 323, + 544, + 390 + ], + "spans": [ + { + "bbox": [ + 476, + 323, + 544, + 390 + ], + "type": "image", + "image_path": "62db6d5e3e9c9ce1ff792919efb9aeebccba7ac3365a5163a7db481c2ad93048.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 62, + 406, + 129, + 470 + ], + "blocks": [ + { + "bbox": [ + 47, + 414, + 59, + 459 + ], + "lines": [ + { + "bbox": [ + 47, + 414, + 59, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 414, + 59, + 459 + ], + "type": "text", + "content": "3D Reconst." + } + ] + } + ], + "index": 24, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 396, + 107, + 404 + ], + "lines": [ + { + "bbox": [ + 85, + 396, + 107, + 404 + ], + "spans": [ + { + "bbox": [ + 85, + 396, + 107, + 404 + ], + "type": "text", + "content": "0.706" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 406, + 129, + 470 + ], + "lines": [ + { + "bbox": [ + 62, + 406, + 129, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 129, + 470 + ], + "type": "image", + "image_path": "d831fc20456ae67f432db5e803c609ea7b426e05b9ccf179539b88ae262503c1.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 131, + 407, + 197, + 470 + ], + "blocks": [ + { + "bbox": [ + 131, + 407, + 197, + 470 + ], + "lines": [ + { + "bbox": [ + 131, + 407, + 197, + 470 + ], + "spans": [ + { + "bbox": [ + 131, + 407, + 197, + 470 + ], + "type": "image", + "image_path": "6d1266d882582bc90a7a3818750f5473eebcf87aeb0aae729916650ce16ba9b3.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 200, + 407, + 263, + 470 + ], + "blocks": [ + { + "bbox": [ + 222, + 396, + 244, + 404 + ], + "lines": [ + { + "bbox": [ + 222, + 396, + 244, + 404 + ], + "spans": [ + { + "bbox": [ + 222, + 396, + 244, + 404 + ], + "type": "text", + "content": "0.705" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 200, + 407, + 263, + 470 + ], + "lines": [ + { + "bbox": [ + 200, + 407, + 263, + 470 + ], + "spans": [ + { + "bbox": [ + 200, + 407, + 263, + 470 + ], + "type": "image", + "image_path": "f886986be8b1e674a87979042bc046fc91daeb81de1a09e77cca9b0c672fa800.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 269, + 407, + 333, + 470 + ], + "blocks": [ + { + "bbox": [ + 291, + 396, + 313, + 404 + ], + "lines": [ + { + "bbox": [ + 291, + 396, + 313, + 404 + ], + "spans": [ + { + "bbox": [ + 291, + 396, + 313, + 404 + ], + "type": "text", + "content": "0.679" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 269, + 407, + 333, + 470 + ], + "lines": [ + { + "bbox": [ + 269, + 407, + 333, + 470 + ], + "spans": [ + { + "bbox": [ + 269, + 407, + 333, + 470 + ], + "type": "image", + "image_path": "4745ef3c430bd2398cdcabe4fcafedb0f6bbbf53a6675e563fec03a91057f7f6.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 342, + 406, + 405, + 468 + ], + "blocks": [ + { + "bbox": [ + 360, + 396, + 383, + 404 + ], + "lines": [ + { + "bbox": [ + 360, + 396, + 383, + 404 + ], + "spans": [ + { + "bbox": [ + 360, + 396, + 383, + 404 + ], + "type": "text", + "content": "0.663" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 406, + 405, + 468 + ], + "lines": [ + { + "bbox": [ + 342, + 406, + 405, + 468 + ], + "spans": [ + { + "bbox": [ + 342, + 406, + 405, + 468 + ], + "type": "image", + "image_path": "25807260325bf9ec372cb3cc2cc2c7628f18c5a2c929ca7350429094e965f8a3.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 408, + 406, + 474, + 468 + ], + "blocks": [ + { + "bbox": [ + 429, + 396, + 451, + 404 + ], + "lines": [ + { + "bbox": [ + 429, + 396, + 451, + 404 + ], + "spans": [ + { + "bbox": [ + 429, + 396, + 451, + 404 + ], + "type": "text", + "content": "0.733" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 408, + 406, + 474, + 468 + ], + "lines": [ + { + "bbox": [ + 408, + 406, + 474, + 468 + ], + "spans": [ + { + "bbox": [ + 408, + 406, + 474, + 468 + ], + "type": "image", + "image_path": "7d8dc0f040695b7a4c8a2e11c99d46f312d81b13878c00570d2be3f908884f67.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 480, + 406, + 544, + 471 + ], + "blocks": [ + { + "bbox": [ + 498, + 396, + 521, + 404 + ], + "lines": [ + { + "bbox": [ + 498, + 396, + 521, + 404 + ], + "spans": [ + { + "bbox": [ + 498, + 396, + 521, + 404 + ], + "type": "text", + "content": "0.782" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 480, + 406, + 544, + 471 + ], + "lines": [ + { + "bbox": [ + 480, + 406, + 544, + 471 + ], + "spans": [ + { + "bbox": [ + 480, + 406, + 544, + 471 + ], + "type": "image", + "image_path": "2af1d830e444c8df0aad2aab90167008b473d189f05594a80d95724162c1b0a7.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 474, + 546, + 510 + ], + "lines": [ + { + "bbox": [ + 45, + 474, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 45, + 474, + 546, + 510 + ], + "type": "text", + "content": "Figure 1: Sample face images from the FFHQ dataset (first row) as well as their corresponding 3D (third row) and frontal 2D reconstruction (second row) from facial templates in the whitebox template inversion attack against ArcFace. Values show the cosine similarity between the templates of the original and frontal reconstructed face images." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "bbox": [ + 143, + 519, + 192, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 519, + 192, + 532 + ], + "spans": [ + { + "bbox": [ + 143, + 519, + 192, + 532 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 45, + 546, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 546, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 546, + 289, + 715 + ], + "type": "text", + "content": "Face recognition systems are increasingly being used in different applications. In such systems, some features (also known as embeddings or templates) are extracted from each face image. Then, the extracted templates are stored in the system's database during the enrollment stage and are later used for recognition. In this paper, we focus on template inversion attacks against face recognition systems and introduce a novel method (dubbed GaFaR) to reconstruct 3D face from facial templates. To this end, we use a geometry-aware generator network based on generative neural radiance fields (GNeRF), and learn a mapping from facial templates to the intermediate latent space of the generator network. We train our network with a semi-supervised learning approach using real and synthetic images simul" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 520, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 689 + ], + "type": "text", + "content": "taneously. For the real training data, we use a Generative Adversarial Network (GAN) based framework to learn the distribution of the latent space. For the synthetic training data, where we have the true latent code, we directly train in the latent space of the generator network. In addition, during the inference stage, we also propose optimization on the camera parameters to generate face images to improve the success attack rate (up to " + }, + { + "bbox": [ + 304, + 520, + 547, + 689 + ], + "type": "inline_equation", + "content": "17.14\\%" + }, + { + "bbox": [ + 304, + 520, + 547, + 689 + ], + "type": "text", + "content": " in our experiments). We evaluate the performance of our method in the whitebox and blackbox attacks against state-of-the-art face recognition models on the LFW and MOBIO datasets. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates. The project page is available at: https://www.idiap.ch/paper/gafar" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19662" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 127, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 286, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 286, + 234 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 286, + 234 + ], + "type": "text", + "content": "Automatic face recognition (FR) has become a well-known biometric authentication tool which has been widely used in different applications, including smartphone locks, border controls, etc. In such systems, generally, some features (also referred to as facial embeddings or facial templates) are extracted from users and stored in the system's database during the enrollment stage. In the recognition stage, either verification or identification, similar features are extracted from the user and are compared with the features in the system's database. Therefore, facial features play the main role in automatic face recognition systems and convey important information about users." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 235, + 286, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 286, + 449 + ], + "type": "text", + "content": "Template inversion (TI) attack against FR systems refers to an attempt by an adversary to reconstruct face images from the templates stored in the system's database. Then, the adversary may find sensitive information about the enrolled users and also can use the reconstructed face image to impersonate and enter the system. Therefore, compared to most attacks on FR systems which threaten the security of the system [19, 3, 22, 18, 35, 34, 10], TI attack jeopardizes both security and privacy of users, and thus requires further study. In this paper, we introduce a novel method (called geometry-aware face reconstruction, shortly GaFaR) for the TI attack in FR systems to reconstruct a 3D face from facial features (extracted from 2D FR models). The 3D face reconstruction provides further information than 2D face reconstruction, and in particular can be used to generate face image from any pose to improve the attack to the FR system. To our knowledge, this is the first paper on 3D face reconstruction from facial templates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 450, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 286, + 628 + ], + "type": "text", + "content": "Recently, neural radiance fields (NeRF) [38] absorbed considerable attention in computer vision society due to remarkable results in the novel-view synthesis problem. Built upon NeRF, generative NeRF (GNeRF) methods such as [37, 44, 56, 6, 8, 41, 42, 20, 55, 7, 12, 48, 43] combine conditional NeRF with generative models for geometry-aware image generation tasks. In these methods, a generative model, such as a generative adversarial network (GAN), is used to embed the shape and appearance of an object in a latent space. Then, the latent code of the GAN, along with the camera parameters, are fed to a NeRF for the rendering process. Among GNeRF methods, there are several works, such as [8, 41, 42, 20, 55, 7, 12, 48], for geometry-aware 3D face generation, which can generate face images from different views." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "content": "In our proposed TI method, we use a geometry-aware face generator network based on GNeRF, and learn a new mapping from facial templates to the intermediate latent space of the generator network. We train our network using real and synthesized images simultaneously with a semisupervised learning approach. For real training data where we do not have the corresponding latent code (i.e., unsuper" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": "vised), we use a GAN-based framework to learn the distribution of the latent space. For the synthetic training data where we have true values of latent codes (i.e., supervised), we directly learn the latent space. Because we have 3D reconstructed face, we can use any arbitrary pose to inject the the FR system for the attack. Therefore, during the inference stage, we also use optimization on the camera parameters to improve the attack against the FR system. We evaluate the performance of our proposed method in the whitebox (i.e., where the adversary knows the internal functioning and parameters of the FR model) and blackbox (i.e., where the adversary does not have information about the internal functioning of the FR model) attacks against state-of-the-art (SOTA) FR systems. For evaluation, we consider FR systems with the LFW [23] and MOBIO [36] datasets and evaluate our attack if the reconstructed face can be used to enter the system. Fig. 1 illustrates sample face images from FFHQ [26] dataset and their 3D reconstruction from ArcFace [11] templates using our attack." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 300, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 545, + 323 + ], + "type": "text", + "content": "To elaborate on the contributions of this paper, we list them hereunder:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 335, + 545, + 536 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 335, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 335, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 317, + 335, + 545, + 395 + ], + "type": "text", + "content": "- We propose a novel template inversion method to reconstruct 3D faces from the facial templates of a face recognition system. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 406, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 406, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 317, + 406, + 545, + 479 + ], + "type": "text", + "content": "- We use a geometry-aware generator network based on GNeRF, and learn a mapping from facial templates to the intermediate latent space of the generator network. We train our mapping network using a GAN-based framework and a semi-supervised learning approach using real and synthetic face images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 489, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 545, + 536 + ], + "type": "text", + "content": "- We use optimization on the camera parameters in the geometry-aware generator network during the inference stage. The optimization of camera parameters can find a pose that improves the success attack rate." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": "The remainder of the paper is organized as follows. First, we review the related works in the literature in Sec. 2. Then, we present our proposed method in Sec. 3. Next, we describe our experiments in Sec. 4 and discuss limitations and ethical considerations in Sec. 5. Finally, the paper is concluded in Sec. 6." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 632, + 397, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 397, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 397, + 645 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Generally, the methods in the literature for reconstructing face from facial templates generate 2D face images. In [57], authors proposed two methods based on optimization and learning to reconstruct 2D low-resolution face images from facial templates in the whitebox scenario. In the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19663" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 86, + 291, + 218 + ], + "blocks": [ + { + "bbox": [ + 84, + 71, + 250, + 83 + ], + "lines": [ + { + "bbox": [ + 84, + 71, + 250, + 83 + ], + "spans": [ + { + "bbox": [ + 84, + 71, + 250, + 83 + ], + "type": "text", + "content": "Table 1: Comparison with related works." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 86, + 291, + 218 + ], + "lines": [ + { + "bbox": [ + 50, + 86, + 291, + 218 + ], + "spans": [ + { + "bbox": [ + 50, + 86, + 291, + 218 + ], + "type": "table", + "html": "
Ref.2D/3DResolutionWhitebox/BlackboxMethod BasisAvailable code
[57]2Dlowwhitebox1) optimization\n2) learningX
[9]2Dlowboth*learningX
[33]2Dlowblackboxlearning✓
[16]2Dlowboth†learningX
[49]2Dlowboth†learningX
[1]2Dlowblackboxlearning + opt.X
[14]2Dhighblackboxlearning✓
[51]2Dhighblackboxoptimization✓
[15]2DhighblackboxoptimizationX
[Ours]3Dhighboth‡learning✓
", + "image_path": "a51079498b56fde9a955061d0962f1c4a27e8e798db5eb3510e915f9569bebc6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 48, + 220, + 287, + 237 + ], + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 237 + ], + "type": "text", + "content": "*The method is based on the whitebox attack, and is also applied in the blackbox scenario by removing a loss term that required the FR model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 48, + 238, + 287, + 255 + ], + "lines": [ + { + "bbox": [ + 48, + 238, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 287, + 255 + ], + "type": "text", + "content": "The method is based on the whitebox attack, and is extended to the blackbox with knowledge distillation of the FR model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 48, + 256, + 287, + 273 + ], + "lines": [ + { + "bbox": [ + 48, + 256, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 287, + 273 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 48, + 256, + 287, + 273 + ], + "type": "text", + "content": " The method is based on the whitebox attack, and is extended to blackbox using a different FR model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 420 + ], + "type": "text", + "content": "optimization-based method, they used a gradient ascent-based algorithm to reconstruct the face image, and optimized the generated image to minimize a multi-term loss function, including the " + }, + { + "bbox": [ + 46, + 289, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 289, + 287, + 420 + ], + "type": "text", + "content": " distance between target templates and templates of the reconstructed face image. In addition, they applied total-variation and Laplacian pyramid gradient normalization [5] on the reconstructed image to generate a smooth image. In their learning-based method, they used a deconvolutional neural network to reconstruct face images and trained it with the same loss function that they used in their optimization-based method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 423, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 579 + ], + "type": "text", + "content": "In [9], a multi-layer perceptron (MLP) is trained to find facial landmark coordinates (optimized using the mean squared error) from the given template. They also trained a convolutional neural network (CNN) to generate face texture (optimized using mean absolute error) from the target template. Then, they used a differentiable warping to combine estimated landmarks (from MLP) and textures (from CNN) and reconstruct low-resolution face images. In the whitebox scenario, they further optimized MLP and CNN by minimizing the distance between templates of reconstructed and original face images. However, in the blackbox scenario, they trained MLP and CNN separately, and used warping in the inference only." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "In [33], two new deconvolutional networks are proposed, called NbBlock-A and NbBlock-B, and are trained with pixel loss " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\ell_{1}" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " norm of reconstruction error) and perceptual loss (distance of middle layers of VGG-19[46] when given the reconstructed and original face images) to reconstruct low-resolution face images in the blackbox scenario. In [16] and [49], a same bijection-learning-based method is used to train GAN models with PO-GAN [25] and TransGAN [24] structure, respectively. The method is based on whitebox attack and for training the GAN model, authors also minimized the distance between target templates and templates" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "extracted from the reconstructed face images using the FR model. In the blackbox attack, they proposed to use the distillation of knowledge to train a student network that mimics the target FR model and used the student network in their method. However, they did not report any detail (and no published source code) about the training of the student network (e.g., network structure, etc.)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 157, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 545, + 275 + ], + "type": "text", + "content": "In [1], a 3-step method is proposed to reconstruct low-resolution face images in the blacbox scenario. First, they trained a GAN for general face generation. In the second step, they trained a MLP to map target templates to embeddings of a known FR model. In the final step, they found a latent code in the input of their generator (of their GAN) which generates a face image that maximizes two terms; the discriminator score (for being a real face image) and cosine similarity between the mapped embedding and the embeddings extracted by the known FR model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 276, + 546, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 546, + 491 + ], + "type": "text", + "content": "In contrast to most works in the literature which generate low-resolution face images, recently there have been few works to generate high-resolution 2D face images. In [51], a grid-search optimization using the simulated annealing [50] approach is used on the latent vector (i.e., input noise) of StyleGAN2 [28] to find latent codes that can generate face images which have templates similar to the target templates. However, their proposed method is computationally expensive, and they reported their evaluation on only 20 face images. In [15], a similar optimization to [51] on the latent vector of StyleGAN2 [28] is considered, but it is solved using the standard genetic algorithm [47]. In contrast to [51, 15] which are based on optimization, in [14], a learning-based method is proposed in the blackbox scenario. The authors proposed to generate some face images using StyleGAN2 [28] and extract the templates using the FR model. Then, they trained a MLP to map facial templates to the input latent codes of StyleGAN2 [28]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 491, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 586 + ], + "type": "text", + "content": "Tab. 1 compares our proposed method with the previous works in the literature. To our knowledge, our proposed method is the first work on 3D face reconstruction from facial templates (extracted from 2D face recognition models). Furthermore, unlike most works in the literature, our method generates high-resolution face images. Last but not least, our method can be used in both whitebox and black-box attacks against FR systems." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 597, + 411, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 411, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 411, + 611 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we consider a threat model as described in Sec. 3.1 and use the proposed 3D face reconstruction method (dubbed GaFaR) described in Sec. 3.2. After training and in the inference stage, we use an optimization on the camera parameters as described in Sec. 3.3 to generate a 2D face image with a pose that can lead to a higher success attack rate (SAR) and inject the face image into the system. Fig. 2 depicts the block-diagram of the proposed method." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19664" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 135, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 135, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 135, + 83 + ], + "type": "text", + "content": "3.1. Threat Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 290, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 290, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 290, + 344 + ], + "type": "text", + "content": "We consider an attack against FR systems, where the adversary gains access to the system's database and aims to generate a 3D reconstruction of enrolled users. To this end, the adversary trains a network to reconstruct a 3D face given a target facial template. Then, the adversary can use the 3D reconstruction to impersonate users into the system (i.e., using a 3D face mask, etc.). For simplicity, we consider that the adversary can inject a 2D face image (from the 3D reconstructed face) into the system. Hence, the adversary uses an appropriate pose to generate a 2D face image for the attack. According to the adversary's knowledge of the FR model, we consider both whitebox and blackbox attacks against FR systems. In the whitebox scenario, we assume that the adversary has complete knowledge of the FR model from which the templates are leaked. However, in the blackbox scenario, we assume that the adversary does not know the internal functioning of the FR model and can only use the FR model to generate facial template for any face image. In addition, we assume that in the blackbox scenario, the adversary has access to another FR model and knows its internal parameters and functioning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 356, + 181, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 181, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 181, + 368 + ], + "type": "text", + "content": "3.2. 3D Face Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": "To reconstruct 3D faces from facial templates, we consider a geometry-aware face generator network based on GNeRF, such as EG3D [7], and learn a mapping " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{GaFaR}}: \\mathcal{T} \\to \\mathcal{W}" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": " from the facial templates " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "t \\in \\mathcal{T}" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": " to the intermediate latent space " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": " of the GNeRF model. Then, we use the remainder of GNeRF model " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "G(.)" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": " to synthesize an image " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\hat{I} = G(\\hat{w}, c)" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": " from an arbitrary view using the mapped intermediate latent vector " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": " and camera parameters " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": ". We consider a network with two fully-connected layers with the Leaky ReLU activation functions for our mapping network " + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{GaFaR}}" + }, + { + "bbox": [ + 46, + 376, + 287, + 509 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": "We train the mapping network " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{GaFaR}}" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " with a semi-supervised approach using synthetic and real data. For the real data, we consider a set of real face images " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\{I_{\\mathrm{real},i}\\}_{i = 0}^{N}" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " and extract the facial template " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{real},i} = F(I_{\\mathrm{real},i})" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " from each face image " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{real},i}" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " using the FR model " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "F(.)" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": ". As we do not have the true value of intermediate latent space " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " of the GNeRF model to generate the real face image, we consider training our mapping network using this data as unsupervised learning. On the other hand, for the synthetic data, we use the pre-trained geometry-aware face generator network to generate a set of random face images " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\{I_{\\mathrm{syn},i}\\}_{i = 0}^{M}" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": ". Therefore, as opposed to real data, we have the true value of intermediate latent space " + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol {w}\\in \\mathcal{W}" + }, + { + "bbox": [ + 46, + 510, + 289, + 714 + ], + "type": "text", + "content": " to generate the same synthetic face image. Hence, we consider training our mapping network using the synthetic data as supervised learning. We train our mapping network simultaneously using real and synthetic training data as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "Unsupervised learning using real data To train our mapping network " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{GaFaR}}(.)" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " with the real data, we use a GAN-based framework based on Wasserstein GAN (WGAN) [2] algorithm to learn the distribution of intermediate latent space " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " of the GNeRF model. In this framework, our mapping network " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{GaFaR}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " generates a latent code " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{w}} = M_{\\mathrm{GaFaR}}([n,t])" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " using the facial template " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " and a random vector " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\pmb{n} \\in \\mathcal{N}" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": ". We can also generate real latent codes " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\pmb{w} = M(\\pmb{z}) \\in \\mathcal{W}" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " for training our GAN using the GNeRF mapping function " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " and a random vector " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\pmb{z} \\in \\mathcal{Z}" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": ". Then, we can use a critic network " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "C(.)" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " to score the latent codes generated by our mapping " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{GaFaR}}(.)" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " and GNeRF mapping " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "M(.)" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": ". Therefore, we can train the critic network " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "C(.)" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " and our mapping " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{GaFaR}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " using the following loss functions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 259, + 545, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 259, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 315, + 259, + 545, + 285 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C} ^ {\\mathrm {W G A N}} = \\mathbb {E} _ {\\boldsymbol {w} \\sim M (\\boldsymbol {z})} [ C (\\boldsymbol {w}) ] - \\mathbb {E} _ {\\hat {\\boldsymbol {w}} \\sim M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, t ])} [ C (\\hat {\\boldsymbol {w}}) ] \\tag {1}", + "image_path": "6e86e27905b3fd6982639ec21b0074d272c160e8f350417ccc0012e4a0945897.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 288, + 545, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 288, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 315, + 288, + 545, + 304 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M _ {\\mathrm {G a F a R}}} ^ {\\mathrm {W G A N}} = \\mathbb {E} _ {\\hat {\\boldsymbol {w}} \\sim M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, \\boldsymbol {t} ])} [ C (\\hat {\\boldsymbol {w}}) ] \\tag {2}", + "image_path": "fef76466eaad77df522a89c4b951a4c3110b2b604644e41629a44f8227591f89.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 312, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 348 + ], + "type": "text", + "content": "In addition to the WGAN training, we use the generated face image " + }, + { + "bbox": [ + 304, + 312, + 545, + 348 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{I}} = G(M_{\\mathrm{GaFaR}}([n,t]),c)" + }, + { + "bbox": [ + 304, + 312, + 545, + 348 + ], + "type": "text", + "content": " to optimize our mapping network using the following loss function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 382, + 355, + 545, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 355, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 382, + 355, + 545, + 370 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e a l}} ^ {\\text {r e c}} = \\mathcal {L} ^ {\\text {P i x e l}} + \\mathcal {L} ^ {\\text {I D}}, \\tag {3}", + "image_path": "9e5476f789b9fee8bac2fa7bada54141ef295e4fcfaded0bd989f2cb4ee2254d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 378, + 545, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 378, + 545, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{Pixel}}" + }, + { + "bbox": [ + 304, + 378, + 545, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 378, + 545, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{ID}}" + }, + { + "bbox": [ + 304, + 378, + 545, + 402 + ], + "type": "text", + "content": " are pixel loss and ID loss, respectively, and are defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 411, + 545, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 545, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {P i x e l}} = \\mathbb {E} _ {\\hat {\\boldsymbol {w}} \\sim M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, \\boldsymbol {t} ])} \\left[ \\| \\boldsymbol {I} - G (\\hat {\\boldsymbol {w}}, \\boldsymbol {c}) \\| _ {2} ^ {2} \\right] \\tag {4}", + "image_path": "c522c61927e91aa6422c4120ce0cbec0da11b4b08e4fa3c346676863b0e5563b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 428, + 545, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 428, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 428, + 545, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {I D}} = \\mathbb {E} _ {\\hat {\\boldsymbol {w}} \\sim M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, \\boldsymbol {t} ])} [ \\| F _ {\\text {l o s s}} (\\boldsymbol {I}) - F _ {\\text {l o s s}} (G (\\hat {\\boldsymbol {w}}, \\boldsymbol {c})) \\| _ {2} ^ {2} ] \\tag {5}", + "image_path": "795894b8d7cf8f9d61b99fcb74b9205d9d0f8226dbecc9c1044c11d8012fcd8a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 463, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 571 + ], + "type": "text", + "content": "The pixel loss minimizes the pixel-level reconstruction error and ID loss helps the network to generate face images with a similar facial templates extracted by " + }, + { + "bbox": [ + 304, + 463, + 545, + 571 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{loss}}" + }, + { + "bbox": [ + 304, + 463, + 545, + 571 + ], + "type": "text", + "content": ". To calculate the ID loss in Eq. (5), we use the same FR model as the one by which leaked templates are calculated (i.e., " + }, + { + "bbox": [ + 304, + 463, + 545, + 571 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 463, + 545, + 571 + ], + "type": "text", + "content": ") in the whitebox scenario. However, in the blackbox scenario, we use another FR model that we assume the adversary has access to. The critic network is a fully-connected network with three layers and Leaky ReLU activation function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 585, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 585, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 585, + 545, + 657 + ], + "type": "text", + "content": "Supervised learning using synthetic data To train our mapping network " + }, + { + "bbox": [ + 304, + 585, + 545, + 657 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{GaFaR}}(.)" + }, + { + "bbox": [ + 304, + 585, + 545, + 657 + ], + "type": "text", + "content": " with synthetic data, we can directly learn the GNeRF intermediate latent code " + }, + { + "bbox": [ + 304, + 585, + 545, + 657 + ], + "type": "inline_equation", + "content": "\\pmb {w} = M(z)" + }, + { + "bbox": [ + 304, + 585, + 545, + 657 + ], + "type": "text", + "content": ". In addition to directly learning " + }, + { + "bbox": [ + 304, + 585, + 545, + 657 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 304, + 585, + 545, + 657 + ], + "type": "text", + "content": ", we use the generated face image to train our mapping network by minimizing the following multi-term loss function:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 370, + 664, + 545, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 664, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 370, + 664, + 545, + 680 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s y n}} ^ {\\text {r e c}} = \\mathcal {L} ^ {w} + \\mathcal {L} ^ {\\text {P i x e l}} + \\mathcal {L} ^ {\\text {I D}}, \\tag {6}", + "image_path": "ef8fb9d5d641afccd9bc0bcdd3b1c801483a2a314d21b0b3f3d8c156c8b322bf.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{Pixel}}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{ID}}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " are pixel loss and ID loss, respectively. Furthermore, " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^w" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "-loss to directly learn the latent space" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19665" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 87, + 492, + 219 + ], + "blocks": [ + { + "bbox": [ + 82, + 73, + 214, + 84 + ], + "lines": [ + { + "bbox": [ + 82, + 73, + 214, + 84 + ], + "spans": [ + { + "bbox": [ + 82, + 73, + 214, + 84 + ], + "type": "text", + "content": "Training (Semi-supervised Learning)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 87, + 492, + 219 + ], + "lines": [ + { + "bbox": [ + 56, + 87, + 492, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 492, + 219 + ], + "type": "image", + "image_path": "9f43988b9d9f5cd9ef97505a35d855d9e3c18c0d20c345fd25302e2beff1e390.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 55, + 224, + 533, + 332 + ], + "blocks": [ + { + "bbox": [ + 55, + 224, + 533, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 533, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 533, + 332 + ], + "type": "image", + "image_path": "2c97c30b32e73f35907f3fc56a205f1323af9a9debc6756cec69c57e410bcc82.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 69, + 361, + 545, + 475 + ], + "blocks": [ + { + "bbox": [ + 88, + 352, + 124, + 361 + ], + "lines": [ + { + "bbox": [ + 88, + 352, + 124, + 361 + ], + "spans": [ + { + "bbox": [ + 88, + 352, + 124, + 361 + ], + "type": "text", + "content": "Inference" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 361, + 545, + 475 + ], + "lines": [ + { + "bbox": [ + 69, + 361, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 69, + 361, + 545, + 475 + ], + "type": "image", + "image_path": "78af3af4a1872df784c28a6ce55670a117fdd4278f577e1dac6a8a9e90e4b7b5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "lines": [ + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "text", + "content": "Figure 2: Block-diagram of the proposed method: In the training process, a semi-supervised approach is used to learn our mapping " + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{GaFaR}}" + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "text", + "content": " (depicted with a green block) from the facial templates to the intermediate latent space of the GNeRF model using real data (unsupervised, where we don't have the corresponding latent code) and synthetic data (supervised, where we have the corresponding latent code " + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}" + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "text", + "content": "), simultaneously. During the inference stage, the leaked template " + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "text", + "content": " is given to our mapping network to find corresponding latent code " + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{w}} = M_{\\mathrm{GaFaR}}([n, t])" + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{w}}" + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "text", + "content": " along with camera parameters " + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "text", + "content": " are given to the generator and renderer of GNeRF " + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "text", + "content": " to generate a reconstructed face image " + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\hat{I} = G(\\hat{\\boldsymbol{w}}, c)" + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "text", + "content": ". Finally, an optimization (either grid search or continuous optimization) is applied on two of the camera parameters, " + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "text", + "content": ", from " + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "text", + "content": ", to find the best pose that minimizes the distance between leaked template " + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 489, + 547, + 586 + ], + "type": "text", + "content": " and the template of reconstructed face image." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 605, + 288, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 288, + 630 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 288, + 630 + ], + "type": "text", + "content": "by minimizing mean squared error between " + }, + { + "bbox": [ + 47, + 605, + 288, + 630 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 47, + 605, + 288, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 605, + 288, + 630 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{w}} = M_{\\mathrm{GaFaR}}([n,t])" + }, + { + "bbox": [ + 47, + 605, + 288, + 630 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 637, + 287, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 637, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 82, + 637, + 287, + 653 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {w} = \\mathbb {E} _ {\\boldsymbol {w} \\sim M (\\boldsymbol {z})} [ \\| \\boldsymbol {w} - M _ {\\mathrm {G a F a R}} ([ \\boldsymbol {n}, \\boldsymbol {t} ]) \\| _ {2} ^ {2} ] \\tag {7}", + "image_path": "445219bd8682dc82b976eb462e5a6f9d82ebc41ab3d1a050a6d70c78a706bcf1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 659, + 230, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 230, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 230, + 672 + ], + "type": "text", + "content": "3.3. Camera Parameters Optimization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 714 + ], + "type": "text", + "content": "During the inference stage, after generating a 3D reconstruction of face using our method described in Sec. 3.2, we optimize the camera parameters to find a pose that increases" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "the SAR. Among different camera parameters " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", the parameters which correspond to the camera rotation can change the pose of the generated face image. Note that by varying the camera rotations, we aim to change the pitch and yaw rotations of the 3D reconstructed face and do not modify the roll rotation. In fact, the roll rotation of the face will be eliminated through face alignment as a preprocess prior to feature extraction in the FR system. To optimize camera parameters, we consider two different approaches as follows:" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "19666" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "text", + "content": "Grid Search (GS) For the grid search, we consider predefined steps to change the camera pitch " + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\theta \\in \\Theta" + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "text", + "content": " and yaw " + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\psi \\in \\Psi" + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "text", + "content": " rotations of camera parameters " + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "text", + "content": ". We generate the 2D face image for all values of camera rotation steps (" + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{step}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathrm{step}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "text", + "content": ") and find the facial template for each generated image. Then, we select the face image " + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{I}} = G(M_{\\mathrm{GaFaR}}([n,t]),c)" + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "text", + "content": " that has a template " + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{t}} = F(\\hat{\\pmb{I}})" + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "text", + "content": " with the minimum mean squared error with the target template " + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "text", + "content": ". Note that the grid search can be applied in both whitebox and blackbox scenarios using the FR model " + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 72, + 289, + 194 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 208, + 288, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 208, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 208, + 288, + 257 + ], + "type": "text", + "content": "Continuous Optimization (CO) In our continuous optimization approach, we start from the frontal image and use Adam [30] optimizer to solve the following minimization using " + }, + { + "bbox": [ + 47, + 208, + 288, + 257 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{w}} = M_{\\mathrm{GaFaR}}([\\boldsymbol{n},\\boldsymbol{t}])" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 117, + 262, + 287, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 262, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 117, + 262, + 287, + 283 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta , \\psi} \\| F (G (\\hat {w}, c)), t \\| _ {2} ^ {2}, \\tag {8}", + "image_path": "3383ad7d5aa2f5366d4b442234689fc7c396e1eeb14e1d6049c7022f30dfcf19.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 290, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 287, + 350 + ], + "type": "text", + "content": "In this optimization, we find the parameters " + }, + { + "bbox": [ + 46, + 290, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 290, + 287, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 290, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 290, + 287, + 350 + ], + "type": "text", + "content": " for camera rotation that lead to a face image with templates close to the target template " + }, + { + "bbox": [ + 46, + 290, + 287, + 350 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 290, + 287, + 350 + ], + "type": "text", + "content": ". In contrast to the grid search approach, the continuous optimization can be applied only in the whitebox scenario." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 361, + 128, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 128, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 128, + 373 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 380, + 163, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 163, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 163, + 393 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "content": "Face Recognition models In our experiments, we consider different SOTA FR models including ArcFace [11], ElasticFace [4] as well as four different FR models with SOTA backbones from FaceX-Zoo [53], including AttentionNet [52], HRNet [54], RepVGG [13] and Swin [32]. The recognition performance of these models are reported in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 495, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 602 + ], + "type": "text", + "content": "Datasets All the mentioned FR models are trained on the MS-Celeb1M dataset [21]. To train the face reconstruction network, we assume that adversary does not have any knowledge of the dataset used in training FR model, and uses a different dataset. To this end, we consider the Flickr-Faces-HQ (FFHQ) dataset [27], which consists of 70,000 high-quality images crawled from internet (with no identity label), for training our face reconstruction network. We randomly split FFHQ into train " + }, + { + "bbox": [ + 46, + 495, + 287, + 602 + ], + "type": "inline_equation", + "content": "(90\\%)" + }, + { + "bbox": [ + 46, + 495, + 287, + 602 + ], + "type": "text", + "content": " and test " + }, + { + "bbox": [ + 46, + 495, + 287, + 602 + ], + "type": "inline_equation", + "content": "(10\\%)" + }, + { + "bbox": [ + 46, + 495, + 287, + 602 + ], + "type": "text", + "content": " set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 602, + 288, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 288, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 288, + 698 + ], + "type": "text", + "content": "To evaluate the vulnerability of FR models, we use two other datasets, including Labeled Faces in the Wild (LFW) [23] and MOBIO [36] datasets. The LFW dataset consists of 13,233 face images of 5,749 people collected from internet, where 1,680 people have two or more images. The MOBIO dataset includes face images of 150 people captured using mobile devices in 12 sessions for each person1." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 326, + 69, + 378, + 122 + ], + "blocks": [ + { + "bbox": [ + 309, + 81, + 320, + 114 + ], + "lines": [ + { + "bbox": [ + 309, + 81, + 320, + 114 + ], + "spans": [ + { + "bbox": [ + 309, + 81, + 320, + 114 + ], + "type": "text", + "content": "rnnnne" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 69, + 378, + 122 + ], + "lines": [ + { + "bbox": [ + 326, + 69, + 378, + 122 + ], + "spans": [ + { + "bbox": [ + 326, + 69, + 378, + 122 + ], + "type": "image", + "image_path": "1f24a996c9116d06dcf1064e78bab90b4b34d8ae47041e82063a5d6a8f834af9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 381, + 69, + 432, + 122 + ], + "blocks": [ + { + "bbox": [ + 381, + 69, + 432, + 122 + ], + "lines": [ + { + "bbox": [ + 381, + 69, + 432, + 122 + ], + "spans": [ + { + "bbox": [ + 381, + 69, + 432, + 122 + ], + "type": "image", + "image_path": "c707859e073481a45313f586e169e175f0760c39725f52398a9790b8ce0bb33f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 435, + 69, + 487, + 122 + ], + "blocks": [ + { + "bbox": [ + 435, + 69, + 487, + 122 + ], + "lines": [ + { + "bbox": [ + 435, + 69, + 487, + 122 + ], + "spans": [ + { + "bbox": [ + 435, + 69, + 487, + 122 + ], + "type": "image", + "image_path": "fb4b809289f0232e9642d9ba5ef6b2b1c1505cf40bb9de299b37e6d986b38c49.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 490, + 70, + 542, + 122 + ], + "blocks": [ + { + "bbox": [ + 490, + 70, + 542, + 122 + ], + "lines": [ + { + "bbox": [ + 490, + 70, + 542, + 122 + ], + "spans": [ + { + "bbox": [ + 490, + 70, + 542, + 122 + ], + "type": "image", + "image_path": "e9f98b45067148fedd9851fde2af3188b7cb5eeffe1223495452bb80f76a66b6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 325, + 124, + 377, + 175 + ], + "blocks": [ + { + "bbox": [ + 310, + 138, + 318, + 170 + ], + "lines": [ + { + "bbox": [ + 310, + 138, + 318, + 170 + ], + "spans": [ + { + "bbox": [ + 310, + 138, + 318, + 170 + ], + "type": "text", + "content": "Reconnst" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 124, + 377, + 175 + ], + "lines": [ + { + "bbox": [ + 325, + 124, + 377, + 175 + ], + "spans": [ + { + "bbox": [ + 325, + 124, + 377, + 175 + ], + "type": "image", + "image_path": "10f20f965b6c96984ed8ed638edb789860477b0768a37fa3266f4d4c9c936255.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 180, + 361, + 190 + ], + "lines": [ + { + "bbox": [ + 339, + 180, + 361, + 190 + ], + "spans": [ + { + "bbox": [ + 339, + 180, + 361, + 190 + ], + "type": "text", + "content": "0.695" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 194, + 545, + 267 + ], + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 267 + ], + "type": "text", + "content": "Figure 3: Sample face images from the FFHQ dataset (first row) and their corresponding frontal (second row) reconstructed face images using our method in the blackbox attack against ElasticFace using ArcFace for " + }, + { + "bbox": [ + 304, + 194, + 545, + 267 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{loss}}" + }, + { + "bbox": [ + 304, + 194, + 545, + 267 + ], + "type": "text", + "content": ". The values show the cosine similarity between templates of original and frontal reconstructed face images." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 380, + 124, + 432, + 175 + ], + "blocks": [ + { + "bbox": [ + 380, + 124, + 432, + 175 + ], + "lines": [ + { + "bbox": [ + 380, + 124, + 432, + 175 + ], + "spans": [ + { + "bbox": [ + 380, + 124, + 432, + 175 + ], + "type": "image", + "image_path": "50def58c439c1660b2904b3edf74d78dcd9102a8ee6f86f3e39473f9bc45a74f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 180, + 416, + 190 + ], + "lines": [ + { + "bbox": [ + 395, + 180, + 416, + 190 + ], + "spans": [ + { + "bbox": [ + 395, + 180, + 416, + 190 + ], + "type": "text", + "content": "0.721" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 435, + 124, + 487, + 175 + ], + "blocks": [ + { + "bbox": [ + 435, + 124, + 487, + 175 + ], + "lines": [ + { + "bbox": [ + 435, + 124, + 487, + 175 + ], + "spans": [ + { + "bbox": [ + 435, + 124, + 487, + 175 + ], + "type": "image", + "image_path": "7541912d6a86061a538840da7c897aa8959125de47b879f4b2eba1c74cac88a0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 180, + 471, + 190 + ], + "lines": [ + { + "bbox": [ + 449, + 180, + 471, + 190 + ], + "spans": [ + { + "bbox": [ + 449, + 180, + 471, + 190 + ], + "type": "text", + "content": "0.699" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 490, + 124, + 541, + 175 + ], + "blocks": [ + { + "bbox": [ + 490, + 124, + 541, + 175 + ], + "lines": [ + { + "bbox": [ + 490, + 124, + 541, + 175 + ], + "spans": [ + { + "bbox": [ + 490, + 124, + 541, + 175 + ], + "type": "image", + "image_path": "e984cb56fba50a0634111494671e4142c7a5b44dfc3107a7a020d93fe9f80500.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 180, + 526, + 190 + ], + "lines": [ + { + "bbox": [ + 504, + 180, + 526, + 190 + ], + "spans": [ + { + "bbox": [ + 504, + 180, + 526, + 190 + ], + "type": "text", + "content": "0.732" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 285, + 545, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 393 + ], + "type": "text", + "content": "Evaluation Protocol For evaluation with each of the LFW and MOBIO datasets, we build a separate FR system, and use the reference templates (i.e., enrolled in the system's database) as input to our face reconstruction method. Then, we inject the reconstructed face image as a query to the system and evaluate the adversary's success attack rate (SAR) to enter the FR system. We should note that for each image, we use only one reconstructed face image as a query to the FR system and evaluate the SAR." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "text", + "content": "Implementation Details and Source Code In our experiments, we use the grid search (in both whitebox and black-box attacks) and continuous optimization (in whitebox attacks only) to optimize camera parameters, as described in Sec. 3.3, in the inference stage. For grid search, we consider " + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "inline_equation", + "content": "\\psi \\in [-45^{\\circ}, +45^{\\circ}]" + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "inline_equation", + "content": "\\theta \\in [-30^{\\circ}, +30^{\\circ}]" + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "text", + "content": " for a " + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "inline_equation", + "content": "11 \\times 11" + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "text", + "content": " grid with step sizes of " + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathrm{step}} = 9^{\\circ}" + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{step}} = 6^{\\circ}" + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "text", + "content": ". In the continuous optimization approach, we use 121 iterations of Adam optimizer [30] with the learning rate of " + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "text", + "content": ". An ablation study on these hyperparameters and the corresponding execution times are reported in the supplementary material. All our models (for different FR models and different scenarios) are trained in 15 epochs (each takes around 2 days) on a system equipped with a single NVIDIA RTX 3090 GPU. For GNeRF, we used a pretrained model of EG3D [7] in our experiments and generate " + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "text", + "content": " resolution face images. The input noise " + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "text", + "content": " to GNeRF model has 512 dimensions and the noise " + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "text", + "content": " in the input of our mapping network " + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{GaFaR}}" + }, + { + "bbox": [ + 304, + 415, + 546, + 680 + ], + "type": "text", + "content": " has 16 dimensions. For unsupervised learning using real face images (FFHQ), we assume that camera parameters for real face images as frontal view. The source code of our experiments are publicly available2." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 286, + 713 + ], + "type": "text", + "content": "1The results on AgeDB [39] dataset is also reported in the supplimen-" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 693, + 351, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 351, + 703 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 351, + 703 + ], + "type": "text", + "content": "tary material." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 703, + 428, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 703, + 428, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 703, + 428, + 713 + ], + "type": "text", + "content": "2 https://www.idiap.ch/paper/gafar" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19667" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 111, + 545, + 233 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 108 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 108 + ], + "type": "text", + "content": "Table 2: Evaluation of blackbox attack against SOTA FR models at systems' FMR=10" + }, + { + "bbox": [ + 46, + 70, + 547, + 108 + ], + "type": "inline_equation", + "content": "^{-3}" + }, + { + "bbox": [ + 46, + 70, + 547, + 108 + ], + "type": "text", + "content": " on the LFW and MOBIO datasets in terms of success attack rate (SAR). For attacks using our proposed method, we use ArcFace and also ElasticFace as " + }, + { + "bbox": [ + 46, + 70, + 547, + 108 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{loss}}" + }, + { + "bbox": [ + 46, + 70, + 547, + 108 + ], + "type": "text", + "content": " to calculate the ID loss in Eq. (5). The values are in percentage." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 111, + 545, + 233 + ], + "lines": [ + { + "bbox": [ + 48, + 111, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 111, + 545, + 233 + ], + "type": "table", + "html": "
LFWMOBIO
ArcFaceElsFaceAtt.NetHRNetRepVGGSwinArcFaceEls.FaceAtt.NetHRNetRepVGGSwin
NBNetA-M [33]4.3210.901.241.601.133.8202.380000
NBNetA-P [33]16.8326.980.661.445.729.704.7616.190.48014.297.14
NBNetB-M [33]10.9821.443.224.473.2111.231.903.803.337.143.338.57
NBNetB-P [33]40.2658.1616.2918.4215.2440.7615.2443.8131.9026.6723.8144.29
Dong et al. [14]13.2112.613.904.073.2212.383.338.1010.486.679.053.33
Vendrow and Vendrow [51]57.7053.0321.1218.859.6246.8429.0543.8127.1426.6720.9545.24
[Floss= Els.Face] GaFaR51.78-18.0711.6811.6347.1547.62-54.2943.3345.7170.48
[Floss= Els.Face] GaFaR + GS61.56-23.5617.2115.8254.0864.76-63.8155.2358.5782.38
[Floss= ArcFace] GaFaR-74.5433.5937.8025.4067.11-84.7672.3876.6772.8689.05
[Floss= ArcFace] GaFaR + GS-78.6738.4243.2729.8470.82-86.6280.0083.8073.3393.33
", + "image_path": "eb3bfd3ec68da89d04f4c1ab3c7c11e30cb89a89aad0209e318baa3c3eaec2a7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 270, + 538, + 331 + ], + "blocks": [ + { + "bbox": [ + 46, + 242, + 547, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 242, + 547, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 547, + 268 + ], + "type": "text", + "content": "Table 3: Evaluation of whitebox attack against SOTA FR models at systems' FMR = 10" + }, + { + "bbox": [ + 46, + 242, + 547, + 268 + ], + "type": "inline_equation", + "content": "^{-3}" + }, + { + "bbox": [ + 46, + 242, + 547, + 268 + ], + "type": "text", + "content": " on the LFW and MOBIO datasets in terms of success attack rate (SAR). The values are in percentage." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 270, + 538, + 331 + ], + "lines": [ + { + "bbox": [ + 48, + 270, + 538, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 270, + 538, + 331 + ], + "type": "table", + "html": "
LFWMOBIO
ArcFaceElsFaceAtt.NetHRNetRepVGGSwinArcFaceEls.FaceAtt.NetHRNetRepVGGSwin
GaFaR79.8462.3227.0031.8717.3374.0882.8681.4364.2971.4353.8194.76
GaFaR + GS82.5268.7132.4237.4819.8376.0685.2382.3870.4781.4259.5294.29
GaFaR + CO84.2571.6934.5339.5320.4477.3689.0584.2975.7181.4362.8695.71
", + "image_path": "a42d7031d371c8b232bcb1f75d469457df974aeecf539a87287500d10c18ee61.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 348, + 108, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 108, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 108, + 361 + ], + "type": "text", + "content": "4.2. Analyze" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "text", + "content": "Blackbox Scenario To evaluate the proposed method in the blackbox attack against SOTA FR models, we consider ArcFace and also ElasticFace as " + }, + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{loss}}" + }, + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "text", + "content": " to calculate the ID loss in Eq. (5). Tab. 2 compares the performance of the proposed method with blackbox face reconstruction methods in the literature in terms of SAR for systems configured at false match rate (FMR) of " + }, + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "text", + "content": ". Similar results for " + }, + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "inline_equation", + "content": "\\mathrm{FMR} = 10^{-2}" + }, + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "text", + "content": " are reported in the supplementary material. As this table shows, the frontal reconstruction by our method (i.e., GaFaR) achieves superior performance than previous methods in the literature. Furthermore, camera parameter optimization (i.e., GaFaR+GS) improves the performance of our attack up to " + }, + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "inline_equation", + "content": "17.14\\%" + }, + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "text", + "content": " compared to our frontal face reconstruction (i.e., GaFaR). Comparing the use of ArcFace and ElasticFace as " + }, + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{loss}}" + }, + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "text", + "content": ", the performance of attacks with the ArcFace model is better. This is due to the fact that ArcFace has a higher recognition accuracy than ElasticFace. Fig. 3 shows sample face images reconstructed from ElasticFace templates in the blackbox attack (using ArcFace for " + }, + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{loss}}" + }, + { + "bbox": [ + 46, + 367, + 290, + 608 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 624, + 288, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 288, + 685 + ], + "type": "text", + "content": "Whitebox Scenario Tab. 3 reports the performance of the proposed method in the whitebox attack" + }, + { + "bbox": [ + 46, + 624, + 288, + 685 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 46, + 624, + 288, + 685 + ], + "type": "text", + "content": " in terms of SAR for FR systems configured at false match rate (FMR) of " + }, + { + "bbox": [ + 46, + 624, + 288, + 685 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 624, + 288, + 685 + ], + "type": "text", + "content": ". Similar results for " + }, + { + "bbox": [ + 46, + 624, + 288, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{FMR} = 10^{-2}" + }, + { + "bbox": [ + 46, + 624, + 288, + 685 + ], + "type": "text", + "content": " are reported in the supplementary material. According to this table, all" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 346, + 545, + 582 + ], + "blocks": [ + { + "bbox": [ + 306, + 346, + 545, + 582 + ], + "lines": [ + { + "bbox": [ + 306, + 346, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 545, + 582 + ], + "type": "image", + "image_path": "df1e69460daef7d142695376886637ae42161f20d045057589ed9e39f2fc87ad.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 588, + 547, + 662 + ], + "lines": [ + { + "bbox": [ + 304, + 588, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 547, + 662 + ], + "type": "text", + "content": "Figure 4: Sample face images from the FFHQ dataset (first row) and their corresponding frontal (second row) and camera parameters sub-grid (third row) reconstructed face images using our method in the whitebox attack against ArcFace. The values show the cosine similarity between templates of original and frontal reconstructed face images." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "these FR models are highly vulnerable to our attack. Furthermore, the camera parameter optimization improves the performance of GaFaR. Comparing grid search with contin" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 288, + 713 + ], + "type": "text", + "content": "3The whitebox attack methods reported in Tab. 1 do not have available source code, and we could not reproduce their results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19668" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 145, + 286, + 277 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 142 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 142 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 142 + ], + "type": "text", + "content": "Table 4: Ablation study on the proposed semi-supervised learning approach and evaluation of loss terms in the white-box attack against ArcFace model in terms of success attack rate (SAR) on the LFW and MOBIO datasets. The SAR values are in percentage and for an attack without any camera parameter optimization (GS/CO)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 145, + 286, + 277 + ], + "lines": [ + { + "bbox": [ + 50, + 145, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 50, + 145, + 286, + 277 + ], + "type": "table", + "html": "
approachLoss FunctionsLFWMOBIO
FMR=10-2FMR=10-3FMR=10-2FMR=10-3
supervisedLrecsyn = Lw + LPixel + LID83.8069.46790.9682.38
Lrecsyn = Lw + LPixel31.7513.9243.818.57
Lrecsyn = Lw + LID0.860.3000
Lrecsyn = Lw33.6915.4332.389.52
unsupervisedLrecreal = LPixel + LID[without WGAN]0.440.1500
Lrecreal = LPixel + LID[with WGAN]67.7245.7670.4831.90
Lrecreal = LID[with WGAN]54.5130.8352.8619.52
Lrecreal = LPixel[with WGAN]2.210.4000
semi-supervisedEqs. (1) to (3) and (6)89.2779.8495.7182.86
", + "image_path": "0a67217bd880fb6f4229176abcb09bc7db1752f1d715878f2cdf139528651861.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 297, + 289, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 289, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 289, + 465 + ], + "type": "text", + "content": "uous optimization, results show that with the same number of iterations continuous optimization achieves better performance. In addition, comparing these results with recognition performance of FR models (available in the supplementary material), we conclude that models with higher recognition accuracy are more vulnerable to our attack. Comparing results in Tab. 2 and Tab. 3, we observe that when we use ArcFace as " + }, + { + "bbox": [ + 46, + 297, + 289, + 465 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{loss}}" + }, + { + "bbox": [ + 46, + 297, + 289, + 465 + ], + "type": "text", + "content": " in our loss function, blackbox attacks achieve better results than whitebox attacks for most cases, which can be explained considering the superior recognition performance of ArcFace than other FR models. Fig. 4 shows sample face images reconstructed from ArcFace templates in the whitebox attack. This figure also presents a grid of reconstructed face images with different poses4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 482, + 289, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 289, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 289, + 674 + ], + "type": "text", + "content": "Ablation Study We evaluate the effect of our semi-supervised learning approach in our proposed method compared to fully supervised learning and fully unsupervised learning approach. In each case, we evaluate the effect of each loss function too. Furthermore, for the fully unsupervised learning approach, we evaluate the effect of GAN learning which we used in our method. Tab. 4 reports our ablation study in whitebox attack against ArcFace model on the LFW and MOBIO datasets in terms of SAR at system's FMR of " + }, + { + "bbox": [ + 46, + 482, + 289, + 674 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 46, + 482, + 289, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 482, + 289, + 674 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 482, + 289, + 674 + ], + "type": "text", + "content": ". As these results show, the proposed semi-supervised approach achieves a better performance than fully supervised learning and fully unsupervised learning approaches. Furthermore, each of our loss terms has an important effect on the performance. In particular, using WGAN for real data, where we don't have the true value of intermediate latent codes, helps the network to" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 378, + 122 + ], + "blocks": [ + { + "bbox": [ + 310, + 81, + 320, + 114 + ], + "lines": [ + { + "bbox": [ + 310, + 81, + 320, + 114 + ], + "spans": [ + { + "bbox": [ + 310, + 81, + 320, + 114 + ], + "type": "text", + "content": "rannnnnne" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 70, + 378, + 122 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 378, + 122 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 378, + 122 + ], + "type": "image", + "image_path": "8798e7321e78a0f6137cc0e68b20383a95ebb3729cf9aad6c5a7c0d4a4a670d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 381, + 70, + 432, + 121 + ], + "blocks": [ + { + "bbox": [ + 381, + 70, + 432, + 121 + ], + "lines": [ + { + "bbox": [ + 381, + 70, + 432, + 121 + ], + "spans": [ + { + "bbox": [ + 381, + 70, + 432, + 121 + ], + "type": "image", + "image_path": "7337d7108f4ee44f027eccc3aa8c24f62312e0e442758f71d5d60127df74f917.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 435, + 70, + 487, + 121 + ], + "blocks": [ + { + "bbox": [ + 435, + 70, + 487, + 121 + ], + "lines": [ + { + "bbox": [ + 435, + 70, + 487, + 121 + ], + "spans": [ + { + "bbox": [ + 435, + 70, + 487, + 121 + ], + "type": "image", + "image_path": "f57540adc2b0420978fcf657260012b7b3cf1d24128dd36543888a965188075d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 490, + 70, + 542, + 122 + ], + "blocks": [ + { + "bbox": [ + 490, + 70, + 542, + 122 + ], + "lines": [ + { + "bbox": [ + 490, + 70, + 542, + 122 + ], + "spans": [ + { + "bbox": [ + 490, + 70, + 542, + 122 + ], + "type": "image", + "image_path": "3b927cab3b27e2f2aacfd11caaa983a6a24706d20e95730cfee34611914f8213.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 325, + 124, + 377, + 175 + ], + "blocks": [ + { + "bbox": [ + 310, + 138, + 319, + 171 + ], + "lines": [ + { + "bbox": [ + 310, + 138, + 319, + 171 + ], + "spans": [ + { + "bbox": [ + 310, + 138, + 319, + 171 + ], + "type": "text", + "content": "Reconstrn" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 124, + 377, + 175 + ], + "lines": [ + { + "bbox": [ + 325, + 124, + 377, + 175 + ], + "spans": [ + { + "bbox": [ + 325, + 124, + 377, + 175 + ], + "type": "image", + "image_path": "803cab6322cd82db265021e32e77709b1e57ecf566313ce1c4e2519526820914.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 180, + 361, + 190 + ], + "lines": [ + { + "bbox": [ + 339, + 180, + 361, + 190 + ], + "spans": [ + { + "bbox": [ + 339, + 180, + 361, + 190 + ], + "type": "text", + "content": "0.160" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 198, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 269 + ], + "type": "text", + "content": "Figure 5: Sample failure cases from the LFW dataset (first row) and their corresponding frontal (second row) reconstructed face images using our method in the whitebox attack against ArcFace. The values below each image show the cosine similarity between templates of original and reconstructed face images." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 380, + 124, + 431, + 173 + ], + "blocks": [ + { + "bbox": [ + 380, + 124, + 431, + 173 + ], + "lines": [ + { + "bbox": [ + 380, + 124, + 431, + 173 + ], + "spans": [ + { + "bbox": [ + 380, + 124, + 431, + 173 + ], + "type": "image", + "image_path": "50d2376b518d512e60618ff6a319cb0ffb8b76af926b318f5acdb9405c358d47.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 180, + 416, + 190 + ], + "lines": [ + { + "bbox": [ + 395, + 180, + 416, + 190 + ], + "spans": [ + { + "bbox": [ + 395, + 180, + 416, + 190 + ], + "type": "text", + "content": "0.150" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 435, + 124, + 486, + 175 + ], + "blocks": [ + { + "bbox": [ + 435, + 124, + 486, + 175 + ], + "lines": [ + { + "bbox": [ + 435, + 124, + 486, + 175 + ], + "spans": [ + { + "bbox": [ + 435, + 124, + 486, + 175 + ], + "type": "image", + "image_path": "04e7586916a2ad3d75191f701c0fe01cf55719767c7e17c01bb03ee9ea1308e4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 180, + 471, + 190 + ], + "lines": [ + { + "bbox": [ + 449, + 180, + 471, + 190 + ], + "spans": [ + { + "bbox": [ + 449, + 180, + 471, + 190 + ], + "type": "text", + "content": "0.170" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 490, + 124, + 541, + 175 + ], + "blocks": [ + { + "bbox": [ + 490, + 124, + 541, + 175 + ], + "lines": [ + { + "bbox": [ + 490, + 124, + 541, + 175 + ], + "spans": [ + { + "bbox": [ + 490, + 124, + 541, + 175 + ], + "type": "image", + "image_path": "03a9aa130a829ca641bcf4f43d5a220c03e43fecb9d6986266616663472ad07c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 180, + 526, + 190 + ], + "lines": [ + { + "bbox": [ + 504, + 180, + 526, + 190 + ], + "spans": [ + { + "bbox": [ + 504, + 180, + 526, + 190 + ], + "type": "text", + "content": "0.160" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 285, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 334 + ], + "type": "text", + "content": "learn the distribution of GNeRF intermediate latent space " + }, + { + "bbox": [ + 304, + 285, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 285, + 545, + 334 + ], + "type": "text", + "content": ". Otherwise, the generated latent code by the mapping network will be out of distribution, and therefore generator part of GNeRF will not generate face-like images." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 343, + 375, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 343, + 375, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 343, + 375, + 357 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 364, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 519 + ], + "type": "text", + "content": "Limitations In spite of the considerable success attack rate of our method in our experiments, the reconstructed face images from some facial templates fail to attack the system. Fig. 5 illustrates sample failed cases in the whitebox attack against ArcFace. As these sample images show, the failed reconstructions reflect a bias in the generated faces for dark skin and also old people. Such a bias in final results can be caused by inherent biases of each of the datasets used to train the FR model, GNeRF model, and our face reconstruction model. In addition to bias, comparing attacks against different FR models in Tabs. 2 and 3 shows that FR models with worse recognition performance have lower SAR and more failed cases." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 713 + ], + "type": "text", + "content": "Ethical considerations This work is proposing a new method for a TI attack against FR models. The 3D reconstructed face can be used to generate 3D face masks or 2D printed photographs for presentation attacks against FR models. In addition to the security threats, the reconstructed face images can reveal important privacy-sensitive information of enrolled users, such as age, gender, ethnicity, etc. We do not condone using our work with the intent of attack to real FR systems. As a matter of fact, the main motivation for this work is to show such a vulnerability in the FR systems, and to encourage the scientific community to develop and propose the next generation of safe and protected FR systems. Along the same lines, data protection regulations, such as the European Union General Data Protection Regulation (EU-GDPR) [17], consider biometric data as sensi" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 683, + 288, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 288, + 712 + ], + "type": "text", + "content": "4Samples of reconstructed face images with the camera parameters grid used in our grid search optimization are presented in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19669" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": "tive information, and put legal obligations to protect them. To this end and to mitigate such threats, several biometric template protection schemes are also proposed in the literature [40, 29, 31, 45]. We should also note that the project on which the work has been conducted has passed an Internal Ethical Review Board (IRB)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 155, + 119, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 155, + 119, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 155, + 119, + 167 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 289, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 289, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 289, + 403 + ], + "type": "text", + "content": "In this paper, we proposed a new method (dubbed GaFaR) to generate 3D face reconstruction from facial templates for a TI attack against FR models. We used a geometry-aware face generation network based on GNeRF and trained a mapping from facial templates to the intermediate latent space of the GNeRF model using a semi-supervised learning approach. To train our model, we used synthetic and real face images. For the synthetic training data, we had the latent code of each face image and could train our mapping with supervised learning. For the real training data, we used a GAN-based framework to learn the distribution of latent space. In the inference stage, we used optimization on the camera parameters to find the pose which increases the success attack rate. We evaluated our method on the whitebox and blackbox attacks against SOTA FR models on the LFW and MOBIO datasets. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates (extracted from 2D face recognition models)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 414, + 140, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 414, + 140, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 414, + 140, + 428 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 434, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 287, + 471 + ], + "type": "text", + "content": "This research is based upon work supported by the H2020 TReSPAsS-ETN Marie Skłodowska-Curie early training network (grant agreement 860813)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 481, + 107, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 107, + 494 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 107, + 494 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 501, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 52, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 52, + 501, + 287, + 555 + ], + "type": "text", + "content": "[1] Muku Akasaka, Soshi Maeda, Yuya Sato, Masakatsu Nishigaki, and Tetsushi Ohki. Model-free template reconstruction attack with feature converter. In 2022 International Conference of the Biometrics Special Interest Group (BIOSIG), pages 1-5. IEEE, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 557, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 557, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 52, + 557, + 287, + 601 + ], + "type": "text", + "content": "[2] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Proceedings of the International Conference on Machine Learning (ICML), pages 214-223. PMLR, 2017." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 602, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 602, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 52, + 602, + 287, + 656 + ], + "type": "text", + "content": "[3] Battista Biggio, Paolo Russu, Luca Didaci, Fabio Roli, et al. Adversarial biometric recognition: A review on biometric system security from the adversarial machine-learning perspective. IEEE Signal Processing Magazine, 32(5):31-41, 2015." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 52, + 658, + 287, + 712 + ], + "type": "text", + "content": "[4] Fadi Boutros, Naser Damer, Florian Kirchbuchner, and Arjan Kuijper. Elasticface: Elastic margin loss for deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1578-1587, 2022." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 311, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 311, + 73, + 545, + 106 + ], + "type": "text", + "content": "[5] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in Computer Vision, pages 671-679. Elsevier, 1987." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 107, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 107, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 311, + 107, + 547, + 161 + ], + "type": "text", + "content": "[6] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3981-3990, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 162, + 547, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 162, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 311, + 162, + 547, + 236 + ], + "type": "text", + "content": "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16123-16133, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 239, + 547, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 239, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 311, + 239, + 547, + 293 + ], + "type": "text", + "content": "[8] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5799-5809, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 294, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 294, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 311, + 294, + 545, + 348 + ], + "type": "text", + "content": "[9] Forrester Cole, David Belanger, Dilip Krishnan, Aaron Sarna, Inbar Mosseri, and William T Freeman. Synthesizing normalized faces from facial identity features. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3703-3712, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "text", + "content": "[10] Debayan Deb, Jianbang Zhang, and Anil K Jain. Advfaces: Adversarial face synthesis. In Proceedings of the 2020 IEEE International Joint Conference on Biometrics (IJCB), pages 1-10. IEEE, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 393, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 436 + ], + "type": "text", + "content": "[11] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "type": "text", + "content": "[12] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10673-10683, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 493, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 547 + ], + "type": "text", + "content": "[13] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13733-13742, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 548, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 601 + ], + "type": "text", + "content": "[14] Xingbo Dong, Zhe Jin, Zhenhua Guo, and Andrew Beng Jin Teoh. Towards generating high definition face images from deep templates. In 2021 International Conference of the Biometrics Special Interest Group (BIOSIG), pages 1-11. IEEE, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 603, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 646 + ], + "type": "text", + "content": "[15] Xingbo Dong, Zhihui Miao, Lan Ma, Jiajun Shen, Zhe Jin, Zhenhua Guo, and Andrew Beng Jin Teoh. Reconstruct face from features using gan generator as a distribution constraint. arXiv preprint arXiv:2206.04295, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[16] Chi Nhan Duong, Thanh-Dat Truong, Khoa Luu, Kha Gia Quach, Hung Bui, and Kaushik Roy. Vec2face: Unveil human faces from their blackbox features in face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6132-6141, 2020." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19670" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[17] European Council. Regulation of the european parliament and of the council on the protection of individuals with regard to the processing of personal data and on the free movement of such data (general data protection regulation), April 2016." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 127, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 127, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 127, + 287, + 160 + ], + "type": "text", + "content": "[18] Javier Galbally, Sébastien Marcel, and Julian Fierrez. Biometric antispoofing methods: A survey in face recognition. IEEE Access, 2:1530-1552, 2014." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "text", + "content": "[19] Javier Galbally, Chris McCool, Julian Fierrez, Sebastien Marcel, and Javier Ortega-Garcia. On the vulnerability of face verification systems to hill-climbing attacks. Pattern Recognition, 43(3):1027-1038, 2010." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 260 + ], + "type": "text", + "content": "[20] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In Proceedings of the 10th International Conference on Learning Representations (ICLR), pages 1-25, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 316 + ], + "type": "text", + "content": "[21] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 87-102. Springer, 2016." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "text", + "content": "[22] Abdenour Hadid, Nicholas Evans, Sebastien Marcel, and Julian Fierrez. Biometrics systems under spoofing attack: an evaluation methodology and lessons learned. IEEE Signal Processing Magazine, 32(5):20-30, 2015." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 414 + ], + "type": "text", + "content": "[23] Gary B. Huang, Manu Ramesh, Tamara Berg, and Erik Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, October 2007." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 458 + ], + "type": "text", + "content": "[24] Yifan Jiang, Shiyu Chang, and Zhangyang Wang. Transgan: Two pure transformers can make one strong gan, and that can scale up. Advances in Neural Information Processing Systems, 34:14745-14758, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 491 + ], + "type": "text", + "content": "[25] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 546 + ], + "type": "text", + "content": "[26] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "text", + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 658 + ], + "type": "text", + "content": "[28] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8110-8119, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "text", + "content": "[29] Prabhjot Kaur, Nitin Kumar, and Maheep Singh. Biometric cryptosystems: a comprehensive survey. Multimedia Tools and Applications, pages 1-56, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Conference on Learning Representations (ICLR), San Diego, California., USA, May 2015." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "type": "text", + "content": "[31] Nitin Kumar et al. Cancelable biometrics: a comprehensive survey. Artificial Intelligence Review, 53(5):3403-3446, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 130, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 185 + ], + "type": "text", + "content": "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 10012-10022, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "text", + "content": "[33] Guangcan Mai, Kai Cao, Pong C Yuen, and Anil K Jain. On the reconstruction of face images from deep face templates. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(5):1188-1202, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 232, + 545, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 264 + ], + "type": "text", + "content": "[34] Sébastien Marcel, Julian Fierrez, and Nicholas Evans. Handbook of Biometric Anti-Spoofing: Presentation Attack Detection and Vulnerability Assessment. Springer Nature, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 266, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 298 + ], + "type": "text", + "content": "[35] Sébastien Marcel, Mark S Nixon, Julian Fierrez, and Nicholas Evans. Handbook of biometric anti-spoofing: Presentation attack detection, volume 2. Springer, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 300, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 300, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 300, + 545, + 342 + ], + "type": "text", + "content": "[36] Chris McCool, Roy Wallace, Mitchell McLaren, Laurent El Shafey, and Sébastien Marcel. Session variability modelling for face authentication. IET Biometrics, 2(3):117-129, Sept. 2013." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 344, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 399 + ], + "type": "text", + "content": "[37] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6351-6361, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 400, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 545, + 455 + ], + "type": "text", + "content": "[38] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405-421, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 456, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 521 + ], + "type": "text", + "content": "[39] Stylianos Moschoglou, Athanasios Papaioannou, Christos Sagonas, Jiankang Deng, Irene Kotsia, and Stefanos Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 51-59, 2017." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "text", + "content": "[40] Karthik Nandakumar and Anil K Jain. Biometric template protection: Bridging the performance gap between theory and practice. IEEE Signal Processing Magazine, 32(5):88-100, 2015." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 568, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 622 + ], + "type": "text", + "content": "[41] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11453-11464, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "text", + "content": "[42] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13503–13513, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[43] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagan, and Andrea Tagliasacchi. Lolnerf: Learn from one look." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19671" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 709 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1558-1567, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "text", + "content": "[44] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 215 + ], + "type": "text", + "content": "[45] Hatef Otroshi Shahreza, Christian Rathgeb, Daile Osorio-Roig, Vedrana Krivokuca Hahn, Sébastien Marcel, and Christoph Busch. Hybrid protection of biometric templates by combining homomorphic encryption and cancelable biometrics. In Proceedings of the 2022 IEEE International Joint Conference on Biometrics (IJCB), pages 1-10. IEEE, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 216, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 216, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 216, + 287, + 248 + ], + "type": "text", + "content": "[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 249, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 249, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 249, + 287, + 270 + ], + "type": "text", + "content": "[47] Mandavilli Srinivas and Lalit M Patnaik. Genetic algorithms: A survey. Computer, 27(6):17-26, 1994." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 271, + 287, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 287, + 314 + ], + "type": "text", + "content": "[48] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. arXiv preprint arXiv:2205.15517, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 315, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 315, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 315, + 287, + 358 + ], + "type": "text", + "content": "[49] Thanh-Dat Truong, Chi Nhan Duong, Ngan Le, Marios Savvides, and Khoa Luu. Vec2face-v2: Unveil human faces from their blackbox features via attention-based network in face recognition, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 358, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 358, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 358, + 287, + 392 + ], + "type": "text", + "content": "[50] Peter JM Van Laarhoven and Emile HL Aarts. Simulated annealing. In Simulated annealing: Theory and applications, pages 7-15. Springer, 1987." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 392, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 392, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 287, + 424 + ], + "type": "text", + "content": "[51] Edward Vendrow and Joshua Vendrow. Realistic face reconstruction from deep embeddings. In NeurIPS 2021 Workshop Privacy in Machine Learning, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 425, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 425, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 425, + 287, + 479 + ], + "type": "text", + "content": "[52] Fei Wang, Mengqing Jiang, Chen Qian, Shuo Yang, Cheng Li, Honggang Zhang, Xiaogang Wang, and Xiaou Tang. Residual attention network for image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3156-3164, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 479, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 522 + ], + "type": "text", + "content": "[53] Jun Wang, Yinglu Liu, Yibo Hu, Hailin Shi, and Tao Mei. Facex-zoo: A pytorch toolbox for face recognition. In Proceedings of the 29th ACM International Conference on Multimedia, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "text", + "content": "[54] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "type": "text", + "content": "[55] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18430-18439, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 633, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 633, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 633, + 287, + 677 + ], + "type": "text", + "content": "[56] Jichao Zhang, Enver Sangineto, Hao Tang, Aliaksandr Siarohin, Zhun Zhong, Nicu Sebe, and Wei Wang. 3d-aware semantic-guided generative model for human synthesis. arXiv preprint arXiv:2112.01422, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 677, + 287, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 677, + 287, + 709 + ], + "spans": [ + { + "bbox": [ + 48, + 677, + 287, + 709 + ], + "type": "text", + "content": "[57] Andrey Zhmoginov and Mark Sandler. Inverting face embeddings with convolutional neural networks. arXiv preprint arXiv:1606.04189, 2016." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "19672" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/b5ce36c9-1639-4be6-bce9-ead12d7207f0_content_list.json b/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/b5ce36c9-1639-4be6-bce9-ead12d7207f0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..79eebe97a8ad02b3359913755c89196759ead446 --- /dev/null +++ b/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/b5ce36c9-1639-4be6-bce9-ead12d7207f0_content_list.json @@ -0,0 +1,1733 @@ +[ + { + "type": "text", + "text": "Template-guided Hierarchical Feature Restoration for Anomaly Detection", + "text_level": 1, + "bbox": [ + 112, + 130, + 856, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hewei Guo $^{1*‡}$ Liping Ren $^{2*‡}$ Jingjing Fu $^{3†}$ Yuwang Wang $^{2†}$ Zhizheng Zhang $^{3}$ \nCuiling Lan $^{3}$ Haoqian Wang $^{2}$ Xinwen Hou $^{1}$ $^{1}$ Institute of Automation, Chinese Academy of Sciences \n $^{2}$ Tsinghua University ${}^{3}$ Microsoft Research Asia \n{guohewei2020, xinwen.hou}@ia.ac.cn \n{rlp20@mails., wang-yuwang@mail., wanghaoqian@}tsinghua.edu.cn \n{jifu, zhizhhang, culan}@microsoft.com", + "bbox": [ + 161, + 178, + 807, + 305 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 339, + 313, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Targeting for detecting anomalies of various sizes for complicated normal patterns, we propose a Template-guided Hierarchical Feature Restoration method, which introduces two key techniques, bottleneck compression and template-guided compensation, for anomaly-free feature restoration. Specially, our framework compresses hierarchical features of an image by bottleneck structure to preserve the most crucial features shared among normal samples. We design template-guided compensation to restore the distorted features towards anomaly-free features. Particularly, we choose the most similar normal sample as the template, and leverage hierarchical features from the template to compensate the distorted features. The bottleneck could partially filter out anomaly features, while the compensation further converts the reminding anomaly features towards normal with template guidance. Finally, anomalies are detected in terms of the cosine distance between the pre-trained features of an inference image and the corresponding restored anomaly-free features. Experimental results demonstrate the effectiveness of our approach, which achieves the state-of-the-art performance on the MVtec LOCO AD dataset.", + "bbox": [ + 75, + 371, + 473, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 734, + 209, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anomaly detection is typically treated as an out-of-distribution detection problem, which learns the distribution from normal samples during training and detects outliers as anomalies during inference. Existing anomaly detection methods have achieved promising results on the MVTec AD benchmark [4], which is mainly composed of images with", + "bbox": [ + 75, + 760, + 468, + 852 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a81c492f87a3228e3d087ecb576de674f8ce82e3cef71c9f826b51477f6b3bd2.jpg", + "image_caption": [ + "Figure 1: Anomaly detection on different categories. From top to bottom: breakfast box, pushpins, splicing connectors in MVtec LOCO AD [3], and cable, transistor in MVtec AD [4]. From left to right: normal image, anomaly image, anomaly maps of PatchCore [28], RD4AD [11] and our method, and ground truth. The red color corresponds to high anomaly score, whereas the blue represents low anomaly score. Best view in color." + ], + "image_footnote": [], + "bbox": [ + 501, + 335, + 887, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "simple normal patterns, i.e. the material surface is uniform, and background is clean, and only one object exists. However, in many practical applications, normal samples are composed of multiple objects placed by rule, such as PCBs with multiple electronic components, printed surfaces with diverse contents and products with parts assembly. Figure 1 illustrates several examples with complicated structures and predefined normal layout patterns. For example, a normal breakfast box contains oatmeal, nuts, oranges, and peach, which are carefully arranged following a certain rule. Its complicated normal pattern creates difficulties for the existing methods to formulate its normal distribution, and consequently leads to degradation on detection accuracy. More", + "bbox": [ + 496, + 703, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 862, + 205, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Work done during an internship at Microsoft Research Asia.", + "bbox": [ + 94, + 887, + 419, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6447", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "over, the accurate anomaly localization is critical in practical industrial scenarios for evaluating the impact of defects and identifying their underlying causes. Therefore, to enhance the usability of anomaly detection methods in various practical scenarios, it is crucial to resolve the challenges associated with complex normal patterns and the requirements for accurate localization.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Substantial anomaly detection methods can be categorized as the embedding-based paradigm, as shown in Figure 2 (a). These methods identify anomalies based on the patch-level dissimilarity between the input embedding feature extracted from a pre-trained network and the normal feature distribution, through such as kNN [9, 25, 28], Mahalanobis distance [10, 26] or normalizing flows [30, 14, 44]. These methods rely on patch-wise feature comparison and prefer to detect local anomalies rather than global anomalies. Once the normal patterns become complicated, the corresponding normal distribution becomes challenging to be formulated, which in turn harms the detection performance.", + "bbox": [ + 75, + 199, + 468, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From the aspect of reconstruction quality divergence, numerous reconstruction-based methods [13, 39, 16] are proposed to expect perfect reconstruction on normal regions and poor reconstruction on anomaly regions, as shown in Figure 2 (b), where the samples could be images or features. The reconstruction-based paradigm is highly adaptable to normal samples with complicated distributions, making it suitable for anomaly detection with rich diversity. However, they suffer from a significant drawback that in certain cases they may fail to detect defects, where the anomaly regions are reconstructed with similar quality as the normal regions. Meanwhile, reconstruction-based methods often incorporate autoencoder (AE) that involves several down-sampling operations. Image details are susceptible to being lost due to feature compression operations, which leads to blurry outputs with large reconstruction error even for normal samples [16], and inevitably causes inaccurate identification of defect locations.", + "bbox": [ + 75, + 383, + 470, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenges of the complicated normal data and the growing demand on accurate defect localization, we propose Template-guided Hierarchical Feature Restoration (THFR) network for anomaly detection, which restores anomaly-free features. The basic idea of our method is shown in Figure 2 (c), the image features are first compressed with bottleneck structure to filter out anomaly features at different levels, and then the distorted features are compensated with template embedding features, which is retrieved from template bank through image-level nearest neighbor search. The bottleneck structure in our approach is designed to retain the essential features that are common to normal samples. To achieve this, we propose global bottleneck and local bottleneck that respectively preserve normal semantic features and normal detailed features. We establish a template bank using multi-level embedding fea", + "bbox": [ + 75, + 657, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tures of normal samples, and retrieve the most similar sample as the template based on the cosine similarity with the input feature. Additionally, we introduce a template-guided compensation module that leverages relation representation between input features and template features to restore anomaly-free features.", + "bbox": [ + 496, + 90, + 890, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thanks to the template guidance and hierarchical compensation design, our method achieves state-of-the-art performance on MVTec LOCO AD [3], a dataset containing normal patterns of rich diversity, and also achieves competitive performance on MVTec AD [4], which is commonly used as anomaly detection benchmark.", + "bbox": [ + 496, + 181, + 890, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are threefold:", + "bbox": [ + 517, + 272, + 848, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new framework to tackle the problem of anomaly detection upon data of rich diversity, in which the anomaly-free features with complicated distribution can be restored from the anomaly features with the guidance of template features.", + "- We design bottleneck compression to retain normal features while partially filtering out anomaly features, and propose template-guided compensation that leverage template embedding features to restore the distorted features towards normal.", + "- Extensive experiments on the standard benchmarks demonstrate the outstanding performance of the proposed method, especially for localization." + ], + "bbox": [ + 517, + 296, + 890, + 511 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 523, + 637, + 540 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised anomaly detection methods are trained on normal samples, while testing on both normal and anomaly samples. Classical anomaly detection methods [36, 38, 32, 42] treat the task as one class classification and focus on defining a compact closed one-class distribution.", + "bbox": [ + 496, + 550, + 890, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Embedding-based methods Embedding-based methods use deep neural networks pre-trained on a large dataset to extract features from an image for anomaly detection and localization. Schirrmeister et al. [34] show that large natural-image datasets such as ImageNet [12] can extract more powerful features than a small application specific dataset. SPADE [9] uses a pre-trained network with multiscale pyramid pooling and segments the anomalies region via among input image features and kNN normal features. PatchCore [28] also uses nearest neighbor to detect anomaly and meanwhile leverages greedy coreset subsampling to lighten memory bank. Different from methods using memory bank, PaDiM [10] abandons slow kNN algorithm and uses Mahalanobis distance metric as an anomaly score. There are follow-up methods [50, 17] to continue to improve the effectiveness of normal features based on Mahalanobis distance. DifferNet [30], CFLOW [14] and Fast-", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6448", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8a4535b285ba8b219b7b7b5628b029e80606ca3db1c2c3cbe218674d009cf947.jpg", + "image_caption": [ + "(a) Embedding-based paradigm [28]" + ], + "image_footnote": [], + "bbox": [ + 81, + 126, + 302, + 233 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d8501ee0d7a0c1abb66f498ff7d6c8ee949eb7591cc41fe6d2d1747bc41d448e.jpg", + "image_caption": [ + "(b) Reconstruction-based paradigm [16]", + "Figure 2: Different paradigms for anomaly detection. (a) Embedding-based methods detect anomaly based on patch-level pre-trained feature similarity. (b) Reconstruction-based methods detect anomaly relying on the assumption that anomalies cannot be reconstructed with good quality. (c) THFR method detects anomaly by comparison between the original and restored anomaly-free features. It utilizes a combination of global and local bottlenecks to filter out anomaly features, known as the G-L bottleneck, and restores distorted features with template features retrieved by image-level nearest neighbor search." + ], + "image_footnote": [], + "bbox": [ + 321, + 154, + 537, + 233 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/09d4515cc7598d8f16ac328505680434151da043803c404a45b517b224a2daa9.jpg", + "image_caption": [ + "(c) Restoration-based THFR method (ours)" + ], + "image_footnote": [], + "bbox": [ + 558, + 101, + 875, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Flow [44] utilize normalizing flows to explicitly approximate the data density, and detect anomaly samples based on their assigned likelihood. It is challenging for embedding-based methods to model the normal distribution when the patterns are complicated.", + "bbox": [ + 75, + 364, + 468, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reconstruction-based methods Reconstruction-based methods are proposed based on an assumption that a model trained on normal data only, cannot represent or reconstruct the anomalies accurately [51, 6]. They typically reconstruct samples from the manifold of the training data, using generative adversarial network (GAN) [20], autoencoder (AE) [29], or variational autoencoder (VAE) [19]. DAGAN [37] trains autoencoder with adversarial losses as the anomaly score of the image. RIAD [47], InTra [23], and SCADN [41] design inpainting frameworks and train models on masked normal data to recover the unseen regions using context for anomaly detection. Deng et al. [11] propose a feature reconstruction network, in which the decoder reconstructs multi-level pre-trained features. Ristea et al. [27] integrate the reconstruction-based functionality into a generic self-supervised block to improve the anomaly detection performance.", + "bbox": [ + 75, + 462, + 468, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nevertheless, due to the powerful representation capability of CNNs, considerable anomalies are reconstructed as well as the normal samples, leading to hypothesis failure. Several memory augmented networks [13, 16, 22] are proposed to enlarge the reconstruction quality gap between normal and anomaly regions. However, these learnable memory based methods focus on patch-wise normal feature representation and thus fail to tackle layout anomalies. Meanwhile, the down-sampling operations are commonly used for crucial feature extraction in reconstruction-based methods, causing the loss of detailed information, which harms the pixel-level anomaly localization accuracy.", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several works [2, 16] adopt skip-connection to assist the reconstruction of details, but anomaly information is also passed to the decoder, resulting in defect leakage.", + "bbox": [ + 496, + 364, + 890, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method aims to restore anomaly-free features from anomaly features instead of recognizing the reconstruction quality gap. For more effective restoration, we filter anomalies via bottleneck and recover the normal pattern and details through template-guided compensation. Note that our method uses the image-level pre-trained features of normal samples as templates to provide both global normal pattern priors and detailed normal feature priors for restoration.", + "bbox": [ + 496, + 410, + 890, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distillation-based methods have recently been successfully applied to anomaly detection, where anomalies are identified based on feature differences between a teacher network pre-trained on a large dataset and a student network trained only on normal samples using knowledge distillation [7, 5, 33, 3, 11, 40, 31]. Additionally, some researchers have attempted to convert unsupervised anomaly detection into a supervised learning task by augmenting normal samples with pseudo-anomalies [24, 21, 46, 35]. However, these approaches are prone to bias towards pseudo outliers and fail to detect a large variety of anomaly types. To improve the embedding feature quality, Zou et al. [52] propose SPD which leverage pseudo anomalies as negative sample for contrastive self-supervised pre-training instead of supervised pre-trained with classification.", + "bbox": [ + 496, + 531, + 890, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 768, + 589, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview architecture of our proposed method is shown in Figure 3. Our method first obtains the multi-level embedding features of a given image using a backbone pretrained on ImageNet [12] as feature extractor. In the training stage, these embedding features are treated as the targets of feature restoration. In order to filter out the anomalies of various sizes, the embedding features are compressed by", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6449", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bd4eca6fa9a4868d597b07f46d9b23c9b1191832f941413896b0997a2520f31e.jpg", + "image_caption": [ + "Figure 3: Overview architecture of proposed Template-guided Hierarchical Feature Restoration (THFR) framework. (a) THFR consists of one pre-trained backbone encoder, bottleneck, and a restoration decoder. Given a test image, THFR uses ImageNet pre-trained backbone encoder to extract features for restoration target. THFR compresses input features through bottleneck, and the compressed features are compensated by template features retrieved from the template bank using image-level nearest neighbor search (image NNS). Finally, anomaly detection is performed in terms of the cosine distance between restored anomaly-free features and input features. (b) Detailed network design of global bottleneck and local bottleneck." + ], + "image_footnote": [], + "bbox": [ + 76, + 92, + 651, + 396 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/246e2e40bfa543104735597004ed9e8b3c9d29d900e3b6402ca56b5cd7f8c506.jpg", + "image_caption": [ + "(b) G-L bottleneck" + ], + "image_footnote": [], + "bbox": [ + 707, + 102, + 893, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "global bottleneck (GBN) and local bottleneck (LBN).", + "bbox": [ + 75, + 523, + 431, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The compression along spatial and channel dimensions inevitably causes information loss on features. To recover the normal pattern with compensation for the lost details, we establish a template bank to record image-level normal embedding features and retrieve the most similar normal feature from the template bank through an image-level nearest neighbor search module called image NNS, and treat the retrieved result as template to guide feature restoration.", + "bbox": [ + 75, + 541, + 468, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Bottleneck", + "text_level": 1, + "bbox": [ + 76, + 678, + 194, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We design global bottleneck and local bottleneck that hierarchically compress the embedding features to retain the crucial normal features at multiple levels, and thereby filter out anomaly features caused by anomalies of various sizes. The previous reconstruction-based works [13, 16, 11] use a fixed-scale loose bottleneck, which works on small-scale anomalies, but underperforms when anomalies are of large sizes. In our network, uncompressed template embedding features could provide rich hierarchical normal features while compensating the compressed features, so we could employ the bottleneck of different tightness to effectively filter anomaly features and obtain compact normal representation.", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a given image, its multi-level pyramid features $I_{k} \\in \\mathbb{R}^{H_{k} \\times W_{k} \\times C_{k}}$ are extracted by pre-trained ResNet [15] and fed into bottleneck structure for hierarchical compression, where $H_{k} \\times W_{k}$ denotes the spatial dimension, $C_{k}$ is the number of channels and $k$ indicates the layer index with maximum value equal to $K$ . The input feature maps are distinct between global bottleneck and local bottleneck. Specifically, the high-level feature $I_{K}$ is fed to global bottleneck, while $[I_{1}, \\dots, I_{K}]$ for local bottleneck.", + "bbox": [ + 496, + 523, + 893, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global bottleneck Considerable large-size anomalies result from missing or misplaced normal components [3], which appear to be globally abnormal but locally normal, creating difficulties for existing anomaly detection methods. We propose global bottleneck to preserve the most representative normal semantic features and filter out the anomaly semantic features. As shown in Figure 3 (b), global bottleneck reforms the deep-level feature $I_{K}$ by a ResBlock and then compress the features by global maxpooling, producing a feature vector $Z_{GBN} \\in \\mathbb{R}^{1 \\times 1 \\times 2C_K}$ . Global max-pooling extremely compresses the feature spatially to eliminate the anomaly features. It could alleviate the attention on object misalignment and instead focuses on global pattern representation. Finally, we use $2 \\times 2$ deconvolution [48]", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6450", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with stride 2 to upsample the compressed feature $Z_{GBN}$ to a $8 \\times 8$ spatial feature map for succeeding feature restoration.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Local bottleneck Besides the compression on semantic features, we introduce a trainable local bottleneck to compress the multi-level features to preserve the local normal features and filter out the local anomaly features. The design of the local bottleneck follows the work [11], as shown in Figure 3 (b). To align multi-level features, the shallow features are downsampled through $3 \\times 3$ convolution layers with stride of 2. We concatenate multi-level representations and fuse them by a ResBlock to obtain compressed feature $Z_{LBN} \\in \\mathbb{R}^{\\frac{H_K}{2} \\times \\frac{W_K}{2} \\times 2C_K}$ , which is fed into detail compensation module for fine-grained feature restoration.", + "bbox": [ + 76, + 140, + 467, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Template-guided compensation", + "text_level": 1, + "bbox": [ + 76, + 319, + 352, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the essential normal features are kept during compression, the information loss on normal features is unavoidable and needs to be compensated. We leverage multilevel normal embedding features as template to guide feature restoration. Since there is no one-to-one correspondence between the positional information of input and template features, we leverage the relationship between the input and template features to guide the restoration process. We fuse the input and template by stacking these two embeddings and the relation feature, and then passing them through multiple $1 \\times 1$ convolution layers. The fused feature is fed to a series of ResBlocks for further restoration.", + "bbox": [ + 75, + 342, + 467, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Template bank We construct template bank using image-level pre-trained features extracted from $N$ normal samples in the all train set as $\\mathcal{B} = \\{B_1^1,\\dots ,B_k^i,\\dots ,B_K^N\\}$ , where $B_{k}^{i}$ represents the $k$ -th layer template feature of the $i$ -th normal sample in template bank. During inference, We could reduce template bank using coreset subsampling method [1, 8] to reduce inference time and memory usage.", + "bbox": [ + 76, + 542, + 467, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image-level nearest neighbor search Given an input sample, we take the deepest pre-trained $K$ -th layer feature $I_K$ as the query to retrieve its correlated template. Image NNS obtains the template with index $t$ by randomly selecting from the template candidates which are $n$ most similar templates to increase the robustness during the training process. The template selection process could be formulated as follows:", + "bbox": [ + 76, + 667, + 467, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nt = \\operatorname {r a n d o m} \\left(\\underset {\\mathcal {S} \\subset \\{1, \\dots , N \\}, | \\mathcal {S} | = n} {\\arg \\min }\\right. \\sum_ {i \\in \\mathcal {S}} d \\left(I _ {K}, B _ {K} ^ {i}\\right)), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 797, + 467, + 828 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d(\\cdot)$ denotes the image-level distance between input feature query $I_{K}$ and template feature key $B_{K}^{i}$ by flattening them to vectors to compute cosine distance. $S$ is a subset of $\\{1,\\dots ,N\\}$ denotes the indexes of $n$ template candidates,", + "bbox": [ + 76, + 839, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2938306d15959934a1dd5be2fa74b42f75b4bfe64c0fb8c4912e55e9d52b2276.jpg", + "image_caption": [ + "Figure 4: Template-guided compensation process. We leverage normal template feature to guide anomaly-free feature restoration with the relation representation between the input feature and the template feature." + ], + "image_footnote": [], + "bbox": [ + 513, + 89, + 880, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which are the top-n nearest neighbors of the input feature $I_K$ . Note that, the $n$ set to 1 during inference. Compared with the traditional point-by-point search strategy, image NNS not only ensures that the reference is completely normal but also improves search efficiency. The multi-level embedding features of the template sample are used as references in the corresponding feature level, which are denoted as $T$ for simplicity.", + "bbox": [ + 498, + 473, + 890, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Relation representation We use pairwise relation [49] between input feature and template feature as additional information to restore normal feature. Specifically, we use $r_{I^i \\to T^j}$ to represent the relation from the $i$ -th feature point of input feature to the $j$ -th feature point of the template feature. The pairwise relation can be defined as a dot-product affinity in the embedding spaces as:", + "bbox": [ + 498, + 616, + 890, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {I ^ {i} \\rightarrow T ^ {j}} = \\theta \\left(I ^ {i}\\right) ^ {\\top} \\phi \\left(T ^ {j}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 733, + 890, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta$ and $\\phi$ are two embedding functions implemented by a $1\\times 1$ spatial convolutional layer. Similarly, we can get the affinity from $j$ -th feature point of template feature to $i$ -th feature point of input feature as $r_{T^j\\rightarrow I^i}$ . We use the pair $(r_{I^i\\to T^j},r_{T^j\\to I^i})$ to describe the bi-directional relation between $i$ -th feature point of input feature to $j$ -th feature point of template feature. Then, we represent the relation matrix $\\mathbf{R}\\in \\mathbb{R}^{m\\times m}$ among all the nodes, where $m$ is the resolution of the feature. The bi-directional", + "bbox": [ + 498, + 763, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6451", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pairwise relation vector of $i$ -th feature point is $\\mathbf{r}_i = [r_{I^i \\to T\\{1, \\dots, m\\}}, r_{T^i \\to I\\{1, \\dots, m\\}}]$ . For example, as shown in Figure 4, the third row and the third column of the relation matrix, i.e. $\\mathbf{r}_3 = [r_{I^3 \\to T\\{1, \\dots, 9\\}}, r_{T^3 \\to I\\{1, \\dots, 9\\}}]$ . In this relation vector, the first half channels indicate the relation that regards the input feature point as query and template feature point as key, while the second half indicates the relation of reverse operation. The bi-directional pairwise relation vectors of all nodes form the relation feature.", + "bbox": [ + 76, + 90, + 472, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Feature compensation We leverage normal template features to guide anomaly-free feature restoration. As shown in Figure 4, the input compressed feature with channel size $C$ is projected to input embedding $E_{i} \\in \\mathbb{R}^{H \\times W \\times C_{i}}$ , and the template feature with channel size $C$ is projected to template embedding $E_{t} \\in \\mathbb{R}^{H \\times W \\times C_{t}}$ . Here, $C_{i}$ represents the channel size of input embedding, and $C_{t}$ represents the channel size of template embedding, while $H$ and $W$ represent the height and width of the feature maps, respectively. The values of $C_{i}$ and $C_{t}$ can be adjusted with respect to the feature compression method used. For global bottleneck compression, $C_{i}$ is set to the original channel size $C$ to provide detailed information, while $C_{t}$ is set to 1 to generate normal layout guidance for high-level semantic feature compensation. For local bottleneck compression, $C_{i}$ is set to 1 to provide spatial layout guidance, and $C_{t}$ is set to $C$ to supply detail feature for low-level feature compensation. Then, we concatenate the input embedding $E_{i}$ , the template embedding $E_{t}$ and the relation feature, and passing them through multiple $1 \\times 1$ convolution layers to get the restored feature. In order to explicitly demonstrate the effectiveness of our method, we visualize the restored feature using a visualization network [43] in experiment.", + "bbox": [ + 76, + 244, + 472, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Loss", + "text_level": 1, + "bbox": [ + 76, + 599, + 150, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the cosine distance between the pre-trained feature and restored feature as restoration loss, as it is capable to evaluate the feature differences at multi-levels regardless of feature dimension differences among each level. Mathematically, for input feature $I$ and restored feature $\\hat{I}$ , we calculate their feature cosine distance along the channel axis as restoration loss:", + "bbox": [ + 76, + 623, + 468, + 728 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {k} = 1 - \\frac {I _ {k} ^ {\\top} \\cdot \\hat {I} _ {k}}{\\| I _ {k} \\| \\| \\hat {I} _ {k} \\|}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 734, + 468, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{k}$ denotes the restoration loss of $k$ -th level.", + "bbox": [ + 76, + 777, + 415, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Anomaly map", + "text_level": 1, + "bbox": [ + 76, + 801, + 223, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the testing phase, we detect anomalies in terms of the cosine distance between the pre-trained feature of an inference image and the corresponding restored anomaly-free feature. The restored feature tends to be close to input image feature for the normal regions, and departs on", + "bbox": [ + 76, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "anomalies. Mathematically, we calculate pixel-wise cosine distance to obtain a 2D anomaly map in $k$ -th level:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM _ {k} (h, w) = 1 - \\frac {\\left(I _ {k} (h , w)\\right) ^ {\\top} \\cdot \\hat {I} _ {k} (h , w)}{\\| I _ {k} (h , w) \\| \\| \\hat {I} _ {k} (h , w) \\|}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 133, + 890, + 170 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $(h,w)$ denotes the position of the feature vector, and $M_{k}\\in \\mathbb{R}^{H_{k}\\times W_{k}}$ . The final anomaly map is calculated as:", + "bbox": [ + 500, + 180, + 893, + 212 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM = \\sum_ {k = 1} ^ {K} U _ {k} \\left(M _ {k}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 223, + 890, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $U_{k}(\\cdot)$ denotes the upsampling operation of $k$ -th anomaly map. In order to remove the noises in the score map, we smooth anomaly map $M$ by a Gaussian filter.", + "bbox": [ + 498, + 270, + 890, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 330, + 633, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 500, + 354, + 687, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets We evaluate the proposed method on two industrial anomaly detection datasets: MVTec LOCO AD [3] and MVTec AD [4]. MVTec LOCO AD consists of 5 real-world sub-datasets for anomaly detection, including 1772 images for training, 304 for validation and 1568 for testing. The normal samples in MVTec LOCO AD often have complicated patterns, and the anomaly samples appear in form of logical or structural anomalies [3]. MVTec AD dataset consists of 15 real-world sub-datasets, with 5 categories of textures and 10 categories of objects. It contains 3629 training images and 1725 test images. Note that all the training images of the two datasets are normal, while some of test images are normal and others are anomalies.", + "bbox": [ + 498, + 378, + 890, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details In our experiments, anomaly detection and localization are performed on one category at a time. We resize input images to $256 \\times 256$ for data preprocessing and use pre-trained WideResNet50 [45] on ImageNet as feature extractor, with the deepest feature level $K$ set to 3. To train our THFR network, we utilize Adam optimizer [18] and set the learning rate to 0.005. Each model is trained for 200 epochs with a batch size of 16 on NVIDIA Tesla V100 GPU. We use Gaussian filter with $\\sigma = 2$ to smooth anomaly map on MVtec LOCO AD, while $\\sigma = 4$ on MVtec AD. For training sample pair selection, we use $n = 1$ for all categories in MVtec LOCO AD and $n = 6$ for all categories in MVtec AD.", + "bbox": [ + 498, + 594, + 890, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics To fairly compare with other methods, we follow the common evaluation metrics of each dataset. For MVTec AD [4], we take area under the receiver operating characteristic (AUROC) as the image-level and pixel-level evaluation metric, and per-region-overlap (PRO) for pixel-level evaluation metric, which can better evaluate", + "bbox": [ + 498, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6452", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/02d2838b8c536fcf8482e84a69c7f60e9c4bfd585f02464e3fe426da10a3e3b5.jpg", + "table_caption": [ + "Table 1: Pixel-level anomaly localization accuracy on MVtec LOCO AD dataset (sPRO) [3]. Remarkably, our approach achieves state-of-the-art performance in four categories and the mean of all categories. Best and second-best scores are bolded and underlined." + ], + "table_footnote": [], + "table_body": "
MethodBreakfastScrew BagPushpinsConnectorsJuice BottleMean
S-T [5]49.660.252.369.881.162.6
DR/EM [46]49.949.049.367.380.059.1
CFLOW [14]48.660.255.171.683.663.8
RD4AD [11]42.257.461.471.385.163.5
PatchCore [28]51.659.853.575.783.264.7
GCAD [3]50.255.873.979.891.070.1
THFR (ours)58.361.576.384.889.674.1
", + "bbox": [ + 196, + 145, + 776, + 272 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8ecda353d3a6ae27d571e0e64704edd30a0a1d7c8a39e5a647ba5f2073ad94c.jpg", + "table_caption": [ + "Table 2: Image-level anomaly detection accuracy on MVtec LOCO AD dataset (AUROC) [3]. Remarkably, our approach achieves the state-of-the-art performance on four categories and the mean of all categories. Best and second-best scores are bolded and underlined." + ], + "table_footnote": [], + "table_body": "
MethodBreakfastScrew BagPushpinsConnectorsJuice BottleMean
S-T [5]-----77.3
DR/EM [46]75.772.776.082.593.980.1
CFLOW [14]77.473.076.082.695.380.8
RD4AD [11]68.774.975.984.494.879.7
PatchCore [28]77.173.374.186.094.681.0
GCAD [3]-----83.3
THFR (ours)78.073.788.392.797.186.0
", + "bbox": [ + 196, + 340, + 774, + 468 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f7254c7548d64413cb1c8dde10cea07a7cd87cabf7c46b0704fcf17bf0549505.jpg", + "table_caption": [ + "Table 3: Anomaly detection and localization results on MVTec AD dataset [4], including image-level and pixel-level AUROC, and pixel-level PRO. Best and second-best scores are bolded and underlined. * To ensure fair comparison with previous studies, CFLOW [14] is evaluated using input images with resolution of 256x256 pixels." + ], + "table_footnote": [], + "table_body": "
MethodImagePixelPixel(PRO)
SPADE [9]85.596.091.7
PaDiM [10]95.397.592.1
S-T [5]--91.4
DR/EM [46]98.097.3-
CFLOW * [14]96.897.992.7
RD4AD [11]98.597.893.9
PatchCore [28]99.198.193.4
THFR (ours)99.298.295.0
", + "bbox": [ + 101, + 592, + 446, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the performance of small size anomalies through weighting ground-truth regions of different size equally. Following the protocol mentioned in [5], we evaluate the PRO value for a large number of increasing thresholds until an average per-pixel false-positive rate of $30\\%$ for the entire dataset is reached. For MVTec LOCO AD [3], we use AUROC to measure image-level performance, and use saturated-per-region-overlaps (sPRO) [3] with per-pixel false-positive rate of $5\\%$ to evaluate pixel-level performance.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main results", + "text_level": 1, + "bbox": [ + 500, + 491, + 633, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the proposed model with several deep learning-based methods for anomaly detection as baselines, including embedding-based methods [9, 10, 14, 28], reconstruction-based methods [11, 46], distillation-based methods [5, 11, 3], and data-augmented methods [46].", + "bbox": [ + 498, + 515, + 893, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MVTec LOCO AD Anomaly detection results on MVTec LOCO AD [3] are presented in Table 1 and Table 2. In comparison with the SOTA method [3], our approach obtain an absolute AUROC gain of $2.7\\%$ for the image-level detection, and an remarkable absolute sPRO gain of $4.0\\%$ for the pixel-level localization. Our approach achieves new state-of-the-art performance on both detection and localization results, and surpasses other methods by a large margin. More detailed results will be provided in the supplementary materials.", + "bbox": [ + 498, + 609, + 893, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MVTec AD Quantitative comparison results on MVTec AD [4] between baselines and our approach are summarized in Table 3. For overall categories, our method produces comparable results with the other advanced methods. Our approach excels in achieving high accuracy on both image-level detection and pixel-level localization at the same time. More detailed results will be presented in the supplementary materials.", + "bbox": [ + 498, + 779, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6453", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6788f70a6555af00e53aaf5356ef19037be6aee330666c6d2979f16778fa657.jpg", + "image_caption": [ + "Normal" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 166, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1fc883d1f989afc10215983b9779d4b486ae0c83d6f5d90bd4ac1349b3de38ac.jpg", + "image_caption": [ + "Anomaly" + ], + "image_footnote": [], + "bbox": [ + 169, + 88, + 256, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/773b7c428cf547c1ee7372aae07448f5ffdb1abe0661d8690b8c7e39ac02ed68.jpg", + "image_caption": [ + "Recon" + ], + "image_footnote": [], + "bbox": [ + 259, + 88, + 349, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d95ad2598904e39c77050c4b0d06e3782152a292bb38cf3918b4371a2ee1fb7b.jpg", + "image_caption": [ + "GBN", + "Figure 5: Visualization of restored features from different restoration networks. From left to right: normal image, anomaly image, reconstructed images from the pre-trained embedding features (Recon), restored features of restoration network only with GBN, only with LBN, with G-L bottleneck, and with our template-guided hierarchical feature restoration framework, anomaly map, and ground truth. The visualization results show how restoration networks progressively achieve anomaly-free restoration with G-L bottleneck and hierarchical feature compensation." + ], + "image_footnote": [], + "bbox": [ + 351, + 88, + 437, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/855c6f8ba47840794975389d75c06801867ea1761992514e9ecbf4ffca75d9ce.jpg", + "image_caption": [ + "LBN" + ], + "image_footnote": [], + "bbox": [ + 439, + 88, + 527, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74643412a8c200c4fa0f76a4629944bcbb67f4f1390a42324a0d3e476b479fcd.jpg", + "image_caption": [ + "G-L bottleneck" + ], + "image_footnote": [], + "bbox": [ + 529, + 88, + 619, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ec141cc2a6a517b9e3c999fcbec9de14bbe10b77aca8f6adfc474f1967180905.jpg", + "image_caption": [ + "THFR" + ], + "image_footnote": [], + "bbox": [ + 620, + 88, + 709, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eeb24bdd9f397180733332fd91b550dd895f5fddaf6ec045b76cf737d080bd53.jpg", + "image_caption": [ + "Anomaly map" + ], + "image_footnote": [], + "bbox": [ + 710, + 88, + 799, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ec064cf12a89045a325b84b3da3e29b8a081c6906d92320f3dc211c6bf47510.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 802, + 88, + 890, + 214 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 339, + 225, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To explicitly demonstrate the effectiveness of our method, we visualize the restored features using a visualization network. We investigate the effectiveness of bottleneck design and template-guided compensation, and study the influence of subsampling within the template bank.", + "bbox": [ + 75, + 363, + 468, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of restored features As shown in Figure 5, we visually compared the pre-trained feature with the restored feature acquired from the network utilizing a visualization network [43]. We observe that the restoration networks employing solely local bottleneck (LBN) cannot effectively filter out anomalies that violate complex normal rules, such as missing connector heads and absent labels on the juice bottle. On the other hand, restoration networks with only global bottleneck (GBN) can filter out rule-breaking anomalies, but it encounters challenges in accurately restoring intricate details, such as the recovery of oatmeal and banana slices in the breakfast box. The restoration network with G-L bottleneck also has anomaly feature leakage after bottleneck compression, manifesting as anomalies like the irregular coloration of the juice. In contrast, our THFR network recovers anomaly-free feature through both semantic level correction and detailed feature compensation, resulting in an effective restoration to normal of various types of anomaly areas.", + "bbox": [ + 75, + 458, + 468, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of bottleneck Table 4 shows anomaly detection performance of local bottleneck and global bottleneck on MVTec LOCO AD dataset. The dataset has a rich diversity of data and anomalies, and the combination of local bottleneck and global bottleneck yields substantial enhancements in detection and localization results. Figure 6 illustrates that the global bottleneck primarily concentrates on filtering large-size anomalies, while the local bottleneck excels in filtering small-size anomalies.", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f0fe87b349db9bd226af66d5f2c28ced2cebf0a14d49da7becb6ad2d85325f5.jpg", + "image_caption": [ + "Normal" + ], + "image_footnote": [], + "bbox": [ + 501, + 337, + 576, + 398 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ebfe71ef8140a0b7c97158f49a7ba37a532b4cbbe034b2b689c2ca18ae3048ea.jpg", + "image_caption": [ + "Anomaly" + ], + "image_footnote": [], + "bbox": [ + 578, + 337, + 651, + 397 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/faf4945f46ba28fcde9da1765cf9df1eceeb1222058b752cd2ea06087a3998f1.jpg", + "image_caption": [ + "GBN" + ], + "image_footnote": [], + "bbox": [ + 653, + 337, + 730, + 397 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f02cc670a86be81c5ec6808f72f3f5ff7651d365ec4d5844b37b89db3a06cda3.jpg", + "image_caption": [ + "LBN" + ], + "image_footnote": [], + "bbox": [ + 730, + 337, + 805, + 397 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/232d52900d0adab7f8ea03f067916c8e27e8c37e5f7d5c500b948036c0d296e6.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 807, + 337, + 883, + 397 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74e7d20e35118fe25a4ab49b325aa4104fc38b50e274c183d7bc4b7ac9fac548.jpg", + "image_caption": [ + "Figure 6: Visualization of anomaly detection results with different bottlenecks.", + "Normal", + "Figure 7: Visualization of anomaly detection results with or without template-guided compensation. More precise localization results are obtained with template-guided compensation." + ], + "image_footnote": [], + "bbox": [ + 501, + 458, + 576, + 565 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e656beda0a91fed597cce3ac0741a91e907961decfb5defbb47ef1c8b711d4a1.jpg", + "image_caption": [ + "Anomaly" + ], + "image_footnote": [], + "bbox": [ + 578, + 458, + 653, + 565 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a029d4afb646fb8a6d2d22fc47d7367ab15a06cf7b2c00423afeadcba6f5cf6.jpg", + "image_caption": [ + "W/O comp." + ], + "image_footnote": [], + "bbox": [ + 653, + 458, + 730, + 565 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3511ca2481eea3975bdcd5dff2f760096af3872c6e0a12e3dfd6a7e406b1d462.jpg", + "image_caption": [ + "W/ comp." + ], + "image_footnote": [], + "bbox": [ + 730, + 458, + 805, + 565 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5d22b4ca3a2ac4c78cfe3f563e4f065a174c8700993a1aa8e7eec4058ef61cc.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 807, + 458, + 883, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "G-L bottleneck vs. different defect size To access the ability of different bottleneck designs to detect anomalies of different sizes, we categorize defect images of LOCO [3] into 6 groups based on their defect sizes, and compare the performance of the restoration networks with LBN only, GBN only, and a combination of the two on these 6 groups in terms of anomaly localization. As shown in Figure 8, the accuracy of anomaly localization varies with the size of defects. LBN performs well on small defects, while GBN exhibits superior performance on large defects. The best performance is achieved by using a combination of LBN and GBN across all groups.", + "bbox": [ + 496, + 670, + 890, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of template-guided compensation The effectiveness of our compensation design, with or without com", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6454", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/327c5000445a398255a8688ef4fb028a21738eddfc162e660b3910ddd1e97d8f.jpg", + "image_caption": [ + "Figure 8: Performance comparison of the restoration networks with distinct bottleneck settings on varying defect sizes in terms of anomaly localization. We divide anomaly samples into 6 groups based on defect sizes, and show the percentage of each group with a bar chart." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 464, + 238 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "pensation, is shown in Table 5 and Figure 7. Our compensation design improves detection result by $2.8\\%$ and localization result by $2.9\\%$ over the baseline on MVTec LOCO AD, and improves detection result by $0.6\\%$ , localization result by $0.4\\%$ and localization (PRO) result by $1.1\\%$ over the baseline on MVTec AD. By incorporating compensation design, it is possible to enhance the precision of localization and achieve more accurate coverage of abnormal areas. We also investigated the impact of relation design in our compensation module. The relation design improves detection result by $6.3\\%$ and localization result by $5.4\\%$ over the baseline on MVTec LOCO AD, and improves detection result by $0.5\\%$ , localization result by $0.1\\%$ and localization (PRO) result by $0.4\\%$ over the baseline on MVTec AD.", + "bbox": [ + 75, + 352, + 468, + 564 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Impact of template bank subsampling The inference time and memory usage are highly relevant to the size of the template bank. We adopt coreset subsampling [1, 8] to find a subset that best describes the template bank during the testing phase. In Table 6, we compare performance of template banks with different subsampling ratios, and observe that subsampling has little impact on performance of MVTec AD. Although subsampling slightly reduces the mean detection accuracy of MVTec LOCO AD, its detection accuracy surpasses peers by a remarkable margin.", + "bbox": [ + 75, + 580, + 467, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.4. Complexity analysis", + "text_level": 1, + "bbox": [ + 76, + 741, + 266, + 758 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We compare inference speed and memory usage with different methods in Table 7. The inference time is measured with Intel Xeon E5-2698v4. Because the backbones are identical (i.e. WideResNet50 [45]), we only report complexity on the additional models. We use image NNS and template bank subsampling to improve searching efficiency and reduce memory cost. Compared with peer works, our method with tempalte bank subsampling to $10\\%$ obtain competitive performance on complexity and accuracy.", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1e64db0a0022bd75a1588cddb2c8319073689232402385f83cf83ea8aa0d64e9.jpg", + "table_caption": [ + "Table 4: The impact of bottleneck design on MVtec LOCO AD Dataset. Best scores are bolded." + ], + "table_footnote": [], + "table_body": "
Metrics\\BottleneckGBNLBNG-L bottleneck
Image-level79.780.386.0
Pixel-level55.666.874.1
", + "bbox": [ + 504, + 130, + 887, + 181 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/ca48f368d56afdcbb4ffa974afc355aef358439d57b9669a9c0b82ea8d77cc73.jpg", + "table_caption": [ + "Table 5: Ablation study on template compensation and relation representation. Best scores are bolded." + ], + "table_footnote": [], + "table_body": "
DatasetLOCO [3]AD [4]
w/ temp. w/ rela.ImagePixelImagePixelPixel(PRO)
83.271.298.597.893.9
✓79.768.798.698.194.6
✓86.074.199.298.295.0
", + "bbox": [ + 501, + 238, + 888, + 319 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/859b7f154f800bde695f0a0f21859bc73dc9c272daa1e41055975bf71e6a15fd.jpg", + "table_caption": [ + "Table 6: Anomaly detection performance with template bank subsampling at different ratios." + ], + "table_footnote": [], + "table_body": "
Dataset\\Ratio10%25%50%100%
LOCO [3]Image85.686.086.086.0
Pixel73.774.074.074.1
AD [4]Image99.199.299.299.2
Pixel98.198.298.298.2
Pixel(PRO)95.095.095.095.0
", + "bbox": [ + 501, + 376, + 890, + 470 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/68e1030cfbac625a4763f9f682a2b142da337201d0b8045368dda66c9289ee59.jpg", + "table_caption": [ + "Table 7: The comparison of pre-trained based approaches in terms of inference time (s), memory usage (MB), and performance (image/pixel/pixel-PRO) on MVTec AD [4]." + ], + "table_footnote": [], + "table_body": "
MethodInf. timeMemoryPerformance
CFLOW [14]0.178947(96.8/97.9/92.7)
RD4AD [11]0.079352(98.5/97.8/93.9)
PatchCore(100%) [28]0.1491015(99.1/98.0/93.3)
PatchCore(10%) [28]0.113102(99.0/98.1/93.5)
Ours(100%)0.1301130(99.2/98.2/95.0)
Ours(10%)0.099448(99.1/98.1/95.0)
", + "bbox": [ + 501, + 544, + 890, + 652 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 676, + 617, + 693 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In the practical industrial scenarios, it is challenging to identify anomalies from the complicated normal patterns, and meanwhile precisely localize the anomalous regions of various sizes. We propose a novel template-guided hierarchical feature restoration network for anomaly detection. The proposed method recovers anomaly-free features from anomaly features by using bottleneck to filter anomaly features, and compensating compressed features with template retrieved from template bank. The hierarchical design benefits detection upon anomalies of various sizes. Experimental results demonstrate the effectiveness of our approach, especially on localization. Our method achieves state-of-the-art performance on MVTec LOCO AD dataset.", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6455", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pankaj Agarwal, Sariel Har, Peled Kasturi, and R Varadarajan. Geometric approximation via coresets. Combinatorial and Computational Geometry, 52, 11 2004. 5, 9", + "[2] Samet Akçay, Amir Atapour-Abarghouei, and Toby P. Breckon. Skip-ganomaly: Skip connected and adversarially trained encoder-decoder anomaly detection. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1–8, 2019. 3", + "[3] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130(4):947-969, 2022. 1, 2, 3, 4, 6, 7, 8, 9", + "[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad - a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 1, 2, 6, 7, 9", + "[5] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3, 7", + "[6] Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv preprint arXiv:1807.02011, 2018. 3", + "[7] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5008-5017, June 2021. 3", + "[8] Kenneth L. Clarkson. Coresets, sparse greedy approximation, and the frank-wolfe algorithm. ACM Trans. Algorithms, 6(4), sep 2010. 5, 9", + "[9] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 2, 7", + "[10] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: A patch distribution modeling framework for anomaly detection and localization. In Alberto Del Bimbo, Rita Cucchiara, Stan Sclaroff, Giovanni Maria Farinella, Tao Mei, Marco Bertini, Hugo Jair Escalante, and Roberto Vezzani, editors, Pattern Recognition. ICPR International Workshops and Challenges, pages 475-489, Cham, 2021. Springer International Publishing. 2, 7", + "[11] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9737-9746, June 2022. 1, 3, 4, 5, 7, 9", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2, 3, 4", + "[14] Denis Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 98-107, January 2022. 2, 7, 9", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 4", + "[16] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8791-8800, October 2021. 2, 3, 4", + "[17] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 303–319, Cham, 2022. Springer Nature Switzerland. 2", + "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[19] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[20] Yann LeCun et al. Generalization and network design strategies. _Connectionism in perspective_, 19(143-155):18, 1989. 3", + "[21] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9664-9674, June 2021. 3", + "[22] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3", + "[23] Jonathan Pirnay and Keng Chai. Inpainting transformer for anomaly detection. In Stan Sclaroff, Cosimo Distante, Marco Leo, Giovanni M. Farinella, and Federico Tombari, editors, Image Analysis and Processing - ICIAP 2022, pages 394-406, Cham, 2022. Springer International Publishing. 3", + "[24] Masoud Pourreza, Bahram Mohammadi, Mostafa Khaki, Samir Bouindour, Hichem Snoussi, and Mohammad Sabokrou. G2d: Generate to detect anomaly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2003-2012, January 2021. 3" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6456", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2806-2814, June 2021. 2", + "[26] Oliver Rippel, Patrick Mertens, and Dorit Merhof. Modeling the distribution of normal data in pre-trained deep features for anomaly detection. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6726-6733, 2021. 2", + "[27] Nicolae-Cătălin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13576-13586, June 2022. 3", + "[28] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14318-14328, June 2022. 1, 2, 3, 7, 9", + "[29] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Structuring autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, Oct 2019. 3", + "[30] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1907-1916, January 2021. 2", + "[31] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmetric student-teacher networks for industrial anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2592–2602, January 2023. 3", + "[32] Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel Müller, and Marius Kloft. Deep one-class classification. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 4393-4402. PMLR, 10-15 Jul 2018. 2", + "[33] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H. Rohban, and Hamid R. Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14902-14912, June 2021. 3", + "[34] Robin Schirrmeister, Yuxuan Zhou, Tonio Ball, and Dan Zhang. Understanding anomaly detection with deep invertible networks through hierarchies of distributions and features. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21038-21049. Curran Associates, Inc., 2020. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Hannah M. Schlüter, Jeremy Tan, Benjamin Hou, and Bernhard Kainz. Natural synthetic anomalies for self-supervised anomaly detection and localization. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022, pages 474-489, Cham, 2022. Springer Nature Switzerland. 3", + "[36] Bernhard Scholkopf, John C. Platt, John Shawe-Taylor, Alex J. Smola, and Robert C. Williamson. Estimating the Support of a High-Dimensional Distribution. Neural Computation, 13(7):1443-1471, 07 2001. 2", + "[37] Ta-Wei Tang, Wei-Han Kuo, Jauh-Hsiang Lan, Chien-Fang Ding, Hakiem Hsu, and Hong-Tsu Young. Anomaly detection neural network with dual auto-encoders gan and its industrial inspection applications. Sensors, 20(12), 2020. 3", + "[38] David MJ Tax and Robert PW Duin. Support vector data description. Machine learning, 54(1):45-66, 2004. 2", + "[39] Shashanka Venkataramanan, Kuan-Chuan Peng, Rajat Vikram Singh, and Abhijit Mahalanobis. Attention guided anomaly localization in images. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 485-503, Cham, 2020. Springer International Publishing. 2", + "[40] Shinji Yamada, Satoshi Kamiya, and Kazuhiro Hotta. Reconstructed student-teacher and discriminative networks for anomaly detection. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2725–2732, 2022. 3", + "[41] Xudong Yan, Huaidong Zhang, Xuemiao Xu, Xiaowei Hu, and Pheng-Ann Heng. Learning semantic context from normal samples for unsupervised anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence, 35(4):3110-3118, May 2021. 3", + "[42] Jihun Yi and Sungroh Yoon. Patch svdd: Patch-level svdd for anomaly detection and segmentation. In Proceedings of the Asian Conference on Computer Vision (ACCV), November 2020. 2", + "[43] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 6, 8", + "[44] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint arXiv:2111.07677, 2021. 2, 3", + "[45] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Edwin R. Hancock Richard C. Wilson and William A. P. Smith, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 87.1-87.12. BMVA Press, September 2016. 6, 9", + "[46] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem - a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8330-8339, October 2021. 3, 7", + "[47] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Reconstruction by inpainting for visual anomaly detection. Pattern Recognition, 112:107706, 2021. 3" + ], + "bbox": [ + 501, + 92, + 892, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6457", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Matthew D. Zeiler, Dilip Krishnan, Graham W. Taylor, and Rob Fergus. Deconvolutional networks. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 2528-2535, 2010. 4", + "[49] Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, Xin Jin, and Zhibo Chen. Relation-aware global attention for person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5", + "[50] Ye Zheng, Xiang Wang, Rui Deng, Tianpeng Bao, Rui Zhao, and Liwei Wu. Focus your distribution: Coarse-to-fine noncontrastive learning for anomaly detection and localization. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2022. 2", + "[51] Bo Zong, Qi Song, Martin Renqiang Min, Wei Cheng, Cristian Lumezanu, Daeki Cho, and Haifeng Chen. Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In International Conference on Learning Representations, 2018. 3", + "[52] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 392–408, Cham, 2022. Springer Nature Switzerland. 3" + ], + "bbox": [ + 78, + 90, + 468, + 455 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "6458", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/b5ce36c9-1639-4be6-bce9-ead12d7207f0_model.json b/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/b5ce36c9-1639-4be6-bce9-ead12d7207f0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9af30a8231c6c6c0b722210fa073e8ec1e92bdbc --- /dev/null +++ b/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/b5ce36c9-1639-4be6-bce9-ead12d7207f0_model.json @@ -0,0 +1,2556 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.131, + 0.857, + 0.154 + ], + "angle": 0, + "content": "Template-guided Hierarchical Feature Restoration for Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.179, + 0.808, + 0.306 + ], + "angle": 0, + "content": "Hewei Guo\\(^{1*‡}\\) Liping Ren\\(^{2*‡}\\) Jingjing Fu\\(^{3†}\\) Yuwang Wang\\(^{2†}\\) Zhizheng Zhang\\(^{3}\\) \nCuiling Lan\\(^{3}\\) Haoqian Wang\\(^{2}\\) Xinwen Hou\\(^{1}\\) \n\\(^{1}\\)Institute of Automation, Chinese Academy of Sciences \n\\(^{2}\\)Tsinghua University \\({}^{3}\\)Microsoft Research Asia \n{guohewei2020, xinwen.hou}@ia.ac.cn \n{rlp20@mails., wang-yuwang@mail., wanghaoqian@}tsinghua.edu.cn \n{jifu, zhizhhang, culan}@microsoft.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.34, + 0.314, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.474, + 0.705 + ], + "angle": 0, + "content": "Targeting for detecting anomalies of various sizes for complicated normal patterns, we propose a Template-guided Hierarchical Feature Restoration method, which introduces two key techniques, bottleneck compression and template-guided compensation, for anomaly-free feature restoration. Specially, our framework compresses hierarchical features of an image by bottleneck structure to preserve the most crucial features shared among normal samples. We design template-guided compensation to restore the distorted features towards anomaly-free features. Particularly, we choose the most similar normal sample as the template, and leverage hierarchical features from the template to compensate the distorted features. The bottleneck could partially filter out anomaly features, while the compensation further converts the reminding anomaly features towards normal with template guidance. Finally, anomalies are detected in terms of the cosine distance between the pre-trained features of an inference image and the corresponding restored anomaly-free features. Experimental results demonstrate the effectiveness of our approach, which achieves the state-of-the-art performance on the MVtec LOCO AD dataset." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.735, + 0.21, + 0.751 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.761, + 0.47, + 0.853 + ], + "angle": 0, + "content": "Anomaly detection is typically treated as an out-of-distribution detection problem, which learns the distribution from normal samples during training and detects outliers as anomalies during inference. Existing anomaly detection methods have achieved promising results on the MVTec AD benchmark [4], which is mainly composed of images with" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.337, + 0.888, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Figure 1: Anomaly detection on different categories. From top to bottom: breakfast box, pushpins, splicing connectors in MVtec LOCO AD [3], and cable, transistor in MVtec AD [4]. From left to right: normal image, anomaly image, anomaly maps of PatchCore [28], RD4AD [11] and our method, and ground truth. The red color corresponds to high anomaly score, whereas the blue represents low anomaly score. Best view in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.903 + ], + "angle": 0, + "content": "simple normal patterns, i.e. the material surface is uniform, and background is clean, and only one object exists. However, in many practical applications, normal samples are composed of multiple objects placed by rule, such as PCBs with multiple electronic components, printed surfaces with diverse contents and products with parts assembly. Figure 1 illustrates several examples with complicated structures and predefined normal layout patterns. For example, a normal breakfast box contains oatmeal, nuts, oranges, and peach, which are carefully arranged following a certain rule. Its complicated normal pattern creates difficulties for the existing methods to formulate its normal distribution, and consequently leads to degradation on detection accuracy. More" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.863, + 0.207, + 0.876 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.876, + 0.223, + 0.888 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.42, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\)Work done during an internship at Microsoft Research Asia." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.863, + 0.42, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6447" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "over, the accurate anomaly localization is critical in practical industrial scenarios for evaluating the impact of defects and identifying their underlying causes. Therefore, to enhance the usability of anomaly detection methods in various practical scenarios, it is crucial to resolve the challenges associated with complex normal patterns and the requirements for accurate localization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.47, + 0.382 + ], + "angle": 0, + "content": "Substantial anomaly detection methods can be categorized as the embedding-based paradigm, as shown in Figure 2 (a). These methods identify anomalies based on the patch-level dissimilarity between the input embedding feature extracted from a pre-trained network and the normal feature distribution, through such as kNN [9, 25, 28], Mahalanobis distance [10, 26] or normalizing flows [30, 14, 44]. These methods rely on patch-wise feature comparison and prefer to detect local anomalies rather than global anomalies. Once the normal patterns become complicated, the corresponding normal distribution becomes challenging to be formulated, which in turn harms the detection performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.471, + 0.656 + ], + "angle": 0, + "content": "From the aspect of reconstruction quality divergence, numerous reconstruction-based methods [13, 39, 16] are proposed to expect perfect reconstruction on normal regions and poor reconstruction on anomaly regions, as shown in Figure 2 (b), where the samples could be images or features. The reconstruction-based paradigm is highly adaptable to normal samples with complicated distributions, making it suitable for anomaly detection with rich diversity. However, they suffer from a significant drawback that in certain cases they may fail to detect defects, where the anomaly regions are reconstructed with similar quality as the normal regions. Meanwhile, reconstruction-based methods often incorporate autoencoder (AE) that involves several down-sampling operations. Image details are susceptible to being lost due to feature compression operations, which leads to blurry outputs with large reconstruction error even for normal samples [16], and inevitably causes inaccurate identification of defect locations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To address the challenges of the complicated normal data and the growing demand on accurate defect localization, we propose Template-guided Hierarchical Feature Restoration (THFR) network for anomaly detection, which restores anomaly-free features. The basic idea of our method is shown in Figure 2 (c), the image features are first compressed with bottleneck structure to filter out anomaly features at different levels, and then the distorted features are compensated with template embedding features, which is retrieved from template bank through image-level nearest neighbor search. The bottleneck structure in our approach is designed to retain the essential features that are common to normal samples. To achieve this, we propose global bottleneck and local bottleneck that respectively preserve normal semantic features and normal detailed features. We establish a template bank using multi-level embedding fea" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.181 + ], + "angle": 0, + "content": "tures of normal samples, and retrieve the most similar sample as the template based on the cosine similarity with the input feature. Additionally, we introduce a template-guided compensation module that leverages relation representation between input features and template features to restore anomaly-free features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.182, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Thanks to the template guidance and hierarchical compensation design, our method achieves state-of-the-art performance on MVTec LOCO AD [3], a dataset containing normal patterns of rich diversity, and also achieves competitive performance on MVTec AD [4], which is commonly used as anomaly detection benchmark." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.273, + 0.849, + 0.287 + ], + "angle": 0, + "content": "In summary, our main contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.297, + 0.892, + 0.372 + ], + "angle": 0, + "content": "- We propose a new framework to tackle the problem of anomaly detection upon data of rich diversity, in which the anomaly-free features with complicated distribution can be restored from the anomaly features with the guidance of template features." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.382, + 0.892, + 0.456 + ], + "angle": 0, + "content": "- We design bottleneck compression to retain normal features while partially filtering out anomaly features, and propose template-guided compensation that leverage template embedding features to restore the distorted features towards normal." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.467, + 0.892, + 0.512 + ], + "angle": 0, + "content": "- Extensive experiments on the standard benchmarks demonstrate the outstanding performance of the proposed method, especially for localization." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.297, + 0.892, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.525, + 0.638, + 0.541 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.626 + ], + "angle": 0, + "content": "Unsupervised anomaly detection methods are trained on normal samples, while testing on both normal and anomaly samples. Classical anomaly detection methods [36, 38, 32, 42] treat the task as one class classification and focus on defining a compact closed one-class distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Embedding-based methods Embedding-based methods use deep neural networks pre-trained on a large dataset to extract features from an image for anomaly detection and localization. Schirrmeister et al. [34] show that large natural-image datasets such as ImageNet [12] can extract more powerful features than a small application specific dataset. SPADE [9] uses a pre-trained network with multiscale pyramid pooling and segments the anomalies region via among input image features and kNN normal features. PatchCore [28] also uses nearest neighbor to detect anomaly and meanwhile leverages greedy coreset subsampling to lighten memory bank. Different from methods using memory bank, PaDiM [10] abandons slow kNN algorithm and uses Mahalanobis distance metric as an anomaly score. There are follow-up methods [50, 17] to continue to improve the effectiveness of normal features based on Mahalanobis distance. DifferNet [30], CFLOW [14] and Fast-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6448" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.127, + 0.303, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.238, + 0.288, + 0.251 + ], + "angle": 0, + "content": "(a) Embedding-based paradigm [28]" + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.155, + 0.538, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.238, + 0.536, + 0.251 + ], + "angle": 0, + "content": "(b) Reconstruction-based paradigm [16]" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.102, + 0.876, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.238, + 0.838, + 0.251 + ], + "angle": 0, + "content": "(c) Restoration-based THFR method (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.892, + 0.339 + ], + "angle": 0, + "content": "Figure 2: Different paradigms for anomaly detection. (a) Embedding-based methods detect anomaly based on patch-level pre-trained feature similarity. (b) Reconstruction-based methods detect anomaly relying on the assumption that anomalies cannot be reconstructed with good quality. (c) THFR method detects anomaly by comparison between the original and restored anomaly-free features. It utilizes a combination of global and local bottlenecks to filter out anomaly features, known as the G-L bottleneck, and restores distorted features with template features retrieved by image-level nearest neighbor search." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.469, + 0.442 + ], + "angle": 0, + "content": "Flow [44] utilize normalizing flows to explicitly approximate the data density, and detect anomaly samples based on their assigned likelihood. It is challenging for embedding-based methods to model the normal distribution when the patterns are complicated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.469, + 0.719 + ], + "angle": 0, + "content": "Reconstruction-based methods Reconstruction-based methods are proposed based on an assumption that a model trained on normal data only, cannot represent or reconstruct the anomalies accurately [51, 6]. They typically reconstruct samples from the manifold of the training data, using generative adversarial network (GAN) [20], autoencoder (AE) [29], or variational autoencoder (VAE) [19]. DAGAN [37] trains autoencoder with adversarial losses as the anomaly score of the image. RIAD [47], InTra [23], and SCADN [41] design inpainting frameworks and train models on masked normal data to recover the unseen regions using context for anomaly detection. Deng et al. [11] propose a feature reconstruction network, in which the decoder reconstructs multi-level pre-trained features. Ristea et al. [27] integrate the reconstruction-based functionality into a generic self-supervised block to improve the anomaly detection performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Nevertheless, due to the powerful representation capability of CNNs, considerable anomalies are reconstructed as well as the normal samples, leading to hypothesis failure. Several memory augmented networks [13, 16, 22] are proposed to enlarge the reconstruction quality gap between normal and anomaly regions. However, these learnable memory based methods focus on patch-wise normal feature representation and thus fail to tackle layout anomalies. Meanwhile, the down-sampling operations are commonly used for crucial feature extraction in reconstruction-based methods, causing the loss of detailed information, which harms the pixel-level anomaly localization accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Several works [2, 16] adopt skip-connection to assist the reconstruction of details, but anomaly information is also passed to the decoder, resulting in defect leakage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.531 + ], + "angle": 0, + "content": "Our method aims to restore anomaly-free features from anomaly features instead of recognizing the reconstruction quality gap. For more effective restoration, we filter anomalies via bottleneck and recover the normal pattern and details through template-guided compensation. Note that our method uses the image-level pre-trained features of normal samples as templates to provide both global normal pattern priors and detailed normal feature priors for restoration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.758 + ], + "angle": 0, + "content": "Distillation-based methods have recently been successfully applied to anomaly detection, where anomalies are identified based on feature differences between a teacher network pre-trained on a large dataset and a student network trained only on normal samples using knowledge distillation [7, 5, 33, 3, 11, 40, 31]. Additionally, some researchers have attempted to convert unsupervised anomaly detection into a supervised learning task by augmenting normal samples with pseudo-anomalies [24, 21, 46, 35]. However, these approaches are prone to bias towards pseudo outliers and fail to detect a large variety of anomaly types. To improve the embedding feature quality, Zou et al. [52] propose SPD which leverage pseudo anomalies as negative sample for contrastive self-supervised pre-training instead of supervised pre-trained with classification." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.591, + 0.785 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The overview architecture of our proposed method is shown in Figure 3. Our method first obtains the multi-level embedding features of a given image using a backbone pretrained on ImageNet [12] as feature extractor. In the training stage, these embedding features are treated as the targets of feature restoration. In order to filter out the anomalies of various sizes, the embedding features are compressed by" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6449" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.093, + 0.652, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.103, + 0.895, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.381, + 0.851, + 0.394 + ], + "angle": 0, + "content": "(b) G-L bottleneck" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.406, + 0.892, + 0.499 + ], + "angle": 0, + "content": "Figure 3: Overview architecture of proposed Template-guided Hierarchical Feature Restoration (THFR) framework. (a) THFR consists of one pre-trained backbone encoder, bottleneck, and a restoration decoder. Given a test image, THFR uses ImageNet pre-trained backbone encoder to extract features for restoration target. THFR compresses input features through bottleneck, and the compressed features are compensated by template features retrieved from the template bank using image-level nearest neighbor search (image NNS). Finally, anomaly detection is performed in terms of the cosine distance between restored anomaly-free features and input features. (b) Detailed network design of global bottleneck and local bottleneck." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.433, + 0.539 + ], + "angle": 0, + "content": "global bottleneck (GBN) and local bottleneck (LBN)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.664 + ], + "angle": 0, + "content": "The compression along spatial and channel dimensions inevitably causes information loss on features. To recover the normal pattern with compensation for the lost details, we establish a template bank to record image-level normal embedding features and retrieve the most similar normal feature from the template bank through an image-level nearest neighbor search module called image NNS, and treat the retrieved result as template to guide feature restoration." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.196, + 0.695 + ], + "angle": 0, + "content": "3.1. Bottleneck" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We design global bottleneck and local bottleneck that hierarchically compress the embedding features to retain the crucial normal features at multiple levels, and thereby filter out anomaly features caused by anomalies of various sizes. The previous reconstruction-based works [13, 16, 11] use a fixed-scale loose bottleneck, which works on small-scale anomalies, but underperforms when anomalies are of large sizes. In our network, uncompressed template embedding features could provide rich hierarchical normal features while compensating the compressed features, so we could employ the bottleneck of different tightness to effectively filter anomaly features and obtain compact normal representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.895, + 0.661 + ], + "angle": 0, + "content": "For a given image, its multi-level pyramid features \\( I_{k} \\in \\mathbb{R}^{H_{k} \\times W_{k} \\times C_{k}} \\) are extracted by pre-trained ResNet [15] and fed into bottleneck structure for hierarchical compression, where \\( H_{k} \\times W_{k} \\) denotes the spatial dimension, \\( C_{k} \\) is the number of channels and \\( k \\) indicates the layer index with maximum value equal to \\( K \\). The input feature maps are distinct between global bottleneck and local bottleneck. Specifically, the high-level feature \\( I_{K} \\) is fed to global bottleneck, while \\( [I_{1}, \\dots, I_{K}] \\) for local bottleneck." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Global bottleneck Considerable large-size anomalies result from missing or misplaced normal components [3], which appear to be globally abnormal but locally normal, creating difficulties for existing anomaly detection methods. We propose global bottleneck to preserve the most representative normal semantic features and filter out the anomaly semantic features. As shown in Figure 3 (b), global bottleneck reforms the deep-level feature \\( I_{K} \\) by a ResBlock and then compress the features by global maxpooling, producing a feature vector \\( Z_{GBN} \\in \\mathbb{R}^{1 \\times 1 \\times 2C_K} \\). Global max-pooling extremely compresses the feature spatially to eliminate the anomaly features. It could alleviate the attention on object misalignment and instead focuses on global pattern representation. Finally, we use \\( 2 \\times 2 \\) deconvolution [48]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6450" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "with stride 2 to upsample the compressed feature \\( Z_{GBN} \\) to a \\( 8 \\times 8 \\) spatial feature map for succeeding feature restoration." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.141, + 0.468, + 0.31 + ], + "angle": 0, + "content": "Local bottleneck Besides the compression on semantic features, we introduce a trainable local bottleneck to compress the multi-level features to preserve the local normal features and filter out the local anomaly features. The design of the local bottleneck follows the work [11], as shown in Figure 3 (b). To align multi-level features, the shallow features are downsampled through \\(3 \\times 3\\) convolution layers with stride of 2. We concatenate multi-level representations and fuse them by a ResBlock to obtain compressed feature \\(Z_{LBN} \\in \\mathbb{R}^{\\frac{H_K}{2} \\times \\frac{W_K}{2} \\times 2C_K}\\), which is fed into detail compensation module for fine-grained feature restoration." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.32, + 0.353, + 0.336 + ], + "angle": 0, + "content": "3.2. Template-guided compensation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.468, + 0.523 + ], + "angle": 0, + "content": "Although the essential normal features are kept during compression, the information loss on normal features is unavoidable and needs to be compensated. We leverage multilevel normal embedding features as template to guide feature restoration. Since there is no one-to-one correspondence between the positional information of input and template features, we leverage the relationship between the input and template features to guide the restoration process. We fuse the input and template by stacking these two embeddings and the relation feature, and then passing them through multiple \\(1 \\times 1\\) convolution layers. The fused feature is fed to a series of ResBlocks for further restoration." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.544, + 0.468, + 0.65 + ], + "angle": 0, + "content": "Template bank We construct template bank using image-level pre-trained features extracted from \\(N\\) normal samples in the all train set as \\(\\mathcal{B} = \\{B_1^1,\\dots ,B_k^i,\\dots ,B_K^N\\}\\), where \\(B_{k}^{i}\\) represents the \\(k\\)-th layer template feature of the \\(i\\)-th normal sample in template bank. During inference, We could reduce template bank using coreset subsampling method [1, 8] to reduce inference time and memory usage." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.669, + 0.468, + 0.788 + ], + "angle": 0, + "content": "Image-level nearest neighbor search Given an input sample, we take the deepest pre-trained \\(K\\)-th layer feature \\(I_K\\) as the query to retrieve its correlated template. Image NNS obtains the template with index \\(t\\) by randomly selecting from the template candidates which are \\(n\\) most similar templates to increase the robustness during the training process. The template selection process could be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.799, + 0.468, + 0.829 + ], + "angle": 0, + "content": "\\[\nt = \\operatorname {r a n d o m} \\left(\\underset {\\mathcal {S} \\subset \\{1, \\dots , N \\}, | \\mathcal {S} | = n} {\\arg \\min }\\right. \\sum_ {i \\in \\mathcal {S}} d \\left(I _ {K}, B _ {K} ^ {i}\\right)), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "where \\(d(\\cdot)\\) denotes the image-level distance between input feature query \\(I_{K}\\) and template feature key \\(B_{K}^{i}\\) by flattening them to vectors to compute cosine distance. \\(S\\) is a subset of \\(\\{1,\\dots ,N\\}\\) denotes the indexes of \\(n\\) template candidates," + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.09, + 0.882, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.385, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Figure 4: Template-guided compensation process. We leverage normal template feature to guide anomaly-free feature restoration with the relation representation between the input feature and the template feature." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.474, + 0.892, + 0.595 + ], + "angle": 0, + "content": "which are the top-n nearest neighbors of the input feature \\( I_K \\). Note that, the \\( n \\) set to 1 during inference. Compared with the traditional point-by-point search strategy, image NNS not only ensures that the reference is completely normal but also improves search efficiency. The multi-level embedding features of the template sample are used as references in the corresponding feature level, which are denoted as \\( T \\) for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.617, + 0.892, + 0.722 + ], + "angle": 0, + "content": "Relation representation We use pairwise relation [49] between input feature and template feature as additional information to restore normal feature. Specifically, we use \\( r_{I^i \\to T^j} \\) to represent the relation from the \\( i \\)-th feature point of input feature to the \\( j \\)-th feature point of the template feature. The pairwise relation can be defined as a dot-product affinity in the embedding spaces as:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.734, + 0.891, + 0.753 + ], + "angle": 0, + "content": "\\[\nr _ {I ^ {i} \\rightarrow T ^ {j}} = \\theta \\left(I ^ {i}\\right) ^ {\\top} \\phi \\left(T ^ {j}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.9 + ], + "angle": 0, + "content": "where \\(\\theta\\) and \\(\\phi\\) are two embedding functions implemented by a \\(1\\times 1\\) spatial convolutional layer. Similarly, we can get the affinity from \\(j\\)-th feature point of template feature to \\(i\\)-th feature point of input feature as \\(r_{T^j\\rightarrow I^i}\\). We use the pair \\((r_{I^i\\to T^j},r_{T^j\\to I^i})\\) to describe the bi-directional relation between \\(i\\)-th feature point of input feature to \\(j\\)-th feature point of template feature. Then, we represent the relation matrix \\(\\mathbf{R}\\in \\mathbb{R}^{m\\times m}\\) among all the nodes, where \\(m\\) is the resolution of the feature. The bi-directional" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6451" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "pairwise relation vector of \\(i\\)-th feature point is \\(\\mathbf{r}_i = [r_{I^i \\to T\\{1, \\dots, m\\}}, r_{T^i \\to I\\{1, \\dots, m\\}}]\\). For example, as shown in Figure 4, the third row and the third column of the relation matrix, i.e. \\(\\mathbf{r}_3 = [r_{I^3 \\to T\\{1, \\dots, 9\\}}, r_{T^3 \\to I\\{1, \\dots, 9\\}}]\\). In this relation vector, the first half channels indicate the relation that regards the input feature point as query and template feature point as key, while the second half indicates the relation of reverse operation. The bi-directional pairwise relation vectors of all nodes form the relation feature." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.246, + 0.473, + 0.594 + ], + "angle": 0, + "content": "Feature compensation We leverage normal template features to guide anomaly-free feature restoration. As shown in Figure 4, the input compressed feature with channel size \\( C \\) is projected to input embedding \\( E_{i} \\in \\mathbb{R}^{H \\times W \\times C_{i}} \\), and the template feature with channel size \\( C \\) is projected to template embedding \\( E_{t} \\in \\mathbb{R}^{H \\times W \\times C_{t}} \\). Here, \\( C_{i} \\) represents the channel size of input embedding, and \\( C_{t} \\) represents the channel size of template embedding, while \\( H \\) and \\( W \\) represent the height and width of the feature maps, respectively. The values of \\( C_{i} \\) and \\( C_{t} \\) can be adjusted with respect to the feature compression method used. For global bottleneck compression, \\( C_{i} \\) is set to the original channel size \\( C \\) to provide detailed information, while \\( C_{t} \\) is set to 1 to generate normal layout guidance for high-level semantic feature compensation. For local bottleneck compression, \\( C_{i} \\) is set to 1 to provide spatial layout guidance, and \\( C_{t} \\) is set to \\( C \\) to supply detail feature for low-level feature compensation. Then, we concatenate the input embedding \\( E_{i} \\), the template embedding \\( E_{t} \\) and the relation feature, and passing them through multiple \\( 1 \\times 1 \\) convolution layers to get the restored feature. In order to explicitly demonstrate the effectiveness of our method, we visualize the restored feature using a visualization network [43] in experiment." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.601, + 0.151, + 0.615 + ], + "angle": 0, + "content": "3.3. Loss" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.624, + 0.47, + 0.729 + ], + "angle": 0, + "content": "We use the cosine distance between the pre-trained feature and restored feature as restoration loss, as it is capable to evaluate the feature differences at multi-levels regardless of feature dimension differences among each level. Mathematically, for input feature \\( I \\) and restored feature \\( \\hat{I} \\), we calculate their feature cosine distance along the channel axis as restoration loss:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.735, + 0.47, + 0.773 + ], + "angle": 0, + "content": "\\[\nL _ {k} = 1 - \\frac {I _ {k} ^ {\\top} \\cdot \\hat {I} _ {k}}{\\| I _ {k} \\| \\| \\hat {I} _ {k} \\|}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.779, + 0.416, + 0.794 + ], + "angle": 0, + "content": "where \\(L_{k}\\) denotes the restoration loss of \\(k\\)-th level." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.224, + 0.818 + ], + "angle": 0, + "content": "3.4. Anomaly map" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In the testing phase, we detect anomalies in terms of the cosine distance between the pre-trained feature of an inference image and the corresponding restored anomaly-free feature. The restored feature tends to be close to input image feature for the normal regions, and departs on" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "anomalies. Mathematically, we calculate pixel-wise cosine distance to obtain a 2D anomaly map in \\(k\\)-th level:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.134, + 0.892, + 0.171 + ], + "angle": 0, + "content": "\\[\nM _ {k} (h, w) = 1 - \\frac {\\left(I _ {k} (h , w)\\right) ^ {\\top} \\cdot \\hat {I} _ {k} (h , w)}{\\| I _ {k} (h , w) \\| \\| \\hat {I} _ {k} (h , w) \\|}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.181, + 0.894, + 0.213 + ], + "angle": 0, + "content": "where \\((h,w)\\) denotes the position of the feature vector, and \\(M_{k}\\in \\mathbb{R}^{H_{k}\\times W_{k}}\\). The final anomaly map is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.224, + 0.892, + 0.265 + ], + "angle": 0, + "content": "\\[\nM = \\sum_ {k = 1} ^ {K} U _ {k} \\left(M _ {k}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.271, + 0.892, + 0.318 + ], + "angle": 0, + "content": "where \\( U_{k}(\\cdot) \\) denotes the upsampling operation of \\( k \\)-th anomaly map. In order to remove the noises in the score map, we smooth anomaly map \\( M \\) by a Gaussian filter." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.331, + 0.634, + 0.348 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.356, + 0.688, + 0.373 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.379, + 0.892, + 0.576 + ], + "angle": 0, + "content": "Datasets We evaluate the proposed method on two industrial anomaly detection datasets: MVTec LOCO AD [3] and MVTec AD [4]. MVTec LOCO AD consists of 5 real-world sub-datasets for anomaly detection, including 1772 images for training, 304 for validation and 1568 for testing. The normal samples in MVTec LOCO AD often have complicated patterns, and the anomaly samples appear in form of logical or structural anomalies [3]. MVTec AD dataset consists of 15 real-world sub-datasets, with 5 categories of textures and 10 categories of objects. It contains 3629 training images and 1725 test images. Note that all the training images of the two datasets are normal, while some of test images are normal and others are anomalies." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.595, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Implementation details In our experiments, anomaly detection and localization are performed on one category at a time. We resize input images to \\(256 \\times 256\\) for data preprocessing and use pre-trained WideResNet50 [45] on ImageNet as feature extractor, with the deepest feature level \\(K\\) set to 3. To train our THFR network, we utilize Adam optimizer [18] and set the learning rate to 0.005. Each model is trained for 200 epochs with a batch size of 16 on NVIDIA Tesla V100 GPU. We use Gaussian filter with \\(\\sigma = 2\\) to smooth anomaly map on MVtec LOCO AD, while \\(\\sigma = 4\\) on MVtec AD. For training sample pair selection, we use \\(n = 1\\) for all categories in MVtec LOCO AD and \\(n = 6\\) for all categories in MVtec AD." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation metrics To fairly compare with other methods, we follow the common evaluation metrics of each dataset. For MVTec AD [4], we take area under the receiver operating characteristic (AUROC) as the image-level and pixel-level evaluation metric, and per-region-overlap (PRO) for pixel-level evaluation metric, which can better evaluate" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6452" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.135 + ], + "angle": 0, + "content": "Table 1: Pixel-level anomaly localization accuracy on MVtec LOCO AD dataset (sPRO) [3]. Remarkably, our approach achieves state-of-the-art performance in four categories and the mean of all categories. Best and second-best scores are bolded and underlined." + }, + { + "type": "table", + "bbox": [ + 0.197, + 0.146, + 0.777, + 0.273 + ], + "angle": 0, + "content": "
MethodBreakfastScrew BagPushpinsConnectorsJuice BottleMean
S-T [5]49.660.252.369.881.162.6
DR/EM [46]49.949.049.367.380.059.1
CFLOW [14]48.660.255.171.683.663.8
RD4AD [11]42.257.461.471.385.163.5
PatchCore [28]51.659.853.575.783.264.7
GCAD [3]50.255.873.979.891.070.1
THFR (ours)58.361.576.384.889.674.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.285, + 0.895, + 0.332 + ], + "angle": 0, + "content": "Table 2: Image-level anomaly detection accuracy on MVtec LOCO AD dataset (AUROC) [3]. Remarkably, our approach achieves the state-of-the-art performance on four categories and the mean of all categories. Best and second-best scores are bolded and underlined." + }, + { + "type": "table", + "bbox": [ + 0.197, + 0.342, + 0.776, + 0.469 + ], + "angle": 0, + "content": "
MethodBreakfastScrew BagPushpinsConnectorsJuice BottleMean
S-T [5]-----77.3
DR/EM [46]75.772.776.082.593.980.1
CFLOW [14]77.473.076.082.695.380.8
RD4AD [11]68.774.975.984.494.879.7
PatchCore [28]77.173.374.186.094.681.0
GCAD [3]-----83.3
THFR (ours)78.073.788.392.797.186.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.491, + 0.472, + 0.584 + ], + "angle": 0, + "content": "Table 3: Anomaly detection and localization results on MVTec AD dataset [4], including image-level and pixel-level AUROC, and pixel-level PRO. Best and second-best scores are bolded and underlined. * To ensure fair comparison with previous studies, CFLOW [14] is evaluated using input images with resolution of 256x256 pixels." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.593, + 0.447, + 0.735 + ], + "angle": 0, + "content": "
MethodImagePixelPixel(PRO)
SPADE [9]85.596.091.7
PaDiM [10]95.397.592.1
S-T [5]--91.4
DR/EM [46]98.097.3-
CFLOW * [14]96.897.992.7
RD4AD [11]98.597.893.9
PatchCore [28]99.198.193.4
THFR (ours)99.298.295.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "the performance of small size anomalies through weighting ground-truth regions of different size equally. Following the protocol mentioned in [5], we evaluate the PRO value for a large number of increasing thresholds until an average per-pixel false-positive rate of \\(30\\%\\) for the entire dataset is reached. For MVTec LOCO AD [3], we use AUROC to measure image-level performance, and use saturated-per-region-overlaps (sPRO) [3] with per-pixel false-positive rate of \\(5\\%\\) to evaluate pixel-level performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.492, + 0.634, + 0.508 + ], + "angle": 0, + "content": "4.2. Main results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.516, + 0.895, + 0.593 + ], + "angle": 0, + "content": "We compare the proposed model with several deep learning-based methods for anomaly detection as baselines, including embedding-based methods [9, 10, 14, 28], reconstruction-based methods [11, 46], distillation-based methods [5, 11, 3], and data-augmented methods [46]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.61, + 0.895, + 0.761 + ], + "angle": 0, + "content": "MVTec LOCO AD Anomaly detection results on MVTec LOCO AD [3] are presented in Table 1 and Table 2. In comparison with the SOTA method [3], our approach obtain an absolute AUROC gain of \\(2.7\\%\\) for the image-level detection, and an remarkable absolute sPRO gain of \\(4.0\\%\\) for the pixel-level localization. Our approach achieves new state-of-the-art performance on both detection and localization results, and surpasses other methods by a large margin. More detailed results will be provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "MVTec AD Quantitative comparison results on MVTec AD [4] between baselines and our approach are summarized in Table 3. For overall categories, our method produces comparable results with the other advanced methods. Our approach excels in achieving high accuracy on both image-level detection and pixel-level localization at the same time. More detailed results will be presented in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "6453" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.168, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.217, + 0.143, + 0.226 + ], + "angle": 0, + "content": "Normal" + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.089, + 0.258, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.193, + 0.217, + 0.237, + 0.227 + ], + "angle": 0, + "content": "Anomaly" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.089, + 0.35, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.217, + 0.321, + 0.226 + ], + "angle": 0, + "content": "Recon" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.089, + 0.439, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.217, + 0.409, + 0.226 + ], + "angle": 0, + "content": "GBN" + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.089, + 0.528, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.217, + 0.498, + 0.226 + ], + "angle": 0, + "content": "LBN" + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.089, + 0.62, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.217, + 0.612, + 0.226 + ], + "angle": 0, + "content": "G-L bottleneck" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.089, + 0.71, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.217, + 0.682, + 0.226 + ], + "angle": 0, + "content": "THFR" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.089, + 0.8, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.217, + 0.79, + 0.227 + ], + "angle": 0, + "content": "Anomaly map" + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.089, + 0.892, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.217, + 0.855, + 0.226 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.894, + 0.314 + ], + "angle": 0, + "content": "Figure 5: Visualization of restored features from different restoration networks. From left to right: normal image, anomaly image, reconstructed images from the pre-trained embedding features (Recon), restored features of restoration network only with GBN, only with LBN, with G-L bottleneck, and with our template-guided hierarchical feature restoration framework, anomaly map, and ground truth. The visualization results show how restoration networks progressively achieve anomaly-free restoration with G-L bottleneck and hierarchical feature compensation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.34, + 0.227, + 0.356 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.469, + 0.44 + ], + "angle": 0, + "content": "To explicitly demonstrate the effectiveness of our method, we visualize the restored features using a visualization network. We investigate the effectiveness of bottleneck design and template-guided compensation, and study the influence of subsampling within the template bank." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.746 + ], + "angle": 0, + "content": "Visualization of restored features As shown in Figure 5, we visually compared the pre-trained feature with the restored feature acquired from the network utilizing a visualization network [43]. We observe that the restoration networks employing solely local bottleneck (LBN) cannot effectively filter out anomalies that violate complex normal rules, such as missing connector heads and absent labels on the juice bottle. On the other hand, restoration networks with only global bottleneck (GBN) can filter out rule-breaking anomalies, but it encounters challenges in accurately restoring intricate details, such as the recovery of oatmeal and banana slices in the breakfast box. The restoration network with G-L bottleneck also has anomaly feature leakage after bottleneck compression, manifesting as anomalies like the irregular coloration of the juice. In contrast, our THFR network recovers anomaly-free feature through both semantic level correction and detailed feature compensation, resulting in an effective restoration to normal of various types of anomaly areas." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Impact of bottleneck Table 4 shows anomaly detection performance of local bottleneck and global bottleneck on MVTec LOCO AD dataset. The dataset has a rich diversity of data and anomalies, and the combination of local bottleneck and global bottleneck yields substantial enhancements in detection and localization results. Figure 6 illustrates that the global bottleneck primarily concentrates on filtering large-size anomalies, while the local bottleneck excels in filtering small-size anomalies." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.338, + 0.578, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.4, + 0.559, + 0.41 + ], + "angle": 0, + "content": "Normal" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.338, + 0.653, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.4, + 0.639, + 0.41 + ], + "angle": 0, + "content": "Anomaly" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.338, + 0.731, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.4, + 0.705, + 0.409 + ], + "angle": 0, + "content": "GBN" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.338, + 0.807, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.4, + 0.781, + 0.409 + ], + "angle": 0, + "content": "LBN" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.338, + 0.885, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.836, + 0.4, + 0.854, + 0.409 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.419, + 0.892, + 0.448 + ], + "angle": 0, + "content": "Figure 6: Visualization of anomaly detection results with different bottlenecks." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.459, + 0.578, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.568, + 0.559, + 0.578 + ], + "angle": 0, + "content": "Normal" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.459, + 0.655, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.568, + 0.639, + 0.578 + ], + "angle": 0, + "content": "Anomaly" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.459, + 0.731, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.568, + 0.72, + 0.578 + ], + "angle": 0, + "content": "W/O comp." + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.459, + 0.807, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.568, + 0.792, + 0.578 + ], + "angle": 0, + "content": "W/ comp." + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.459, + 0.884, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.836, + 0.568, + 0.854, + 0.578 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.588, + 0.892, + 0.648 + ], + "angle": 0, + "content": "Figure 7: Visualization of anomaly detection results with or without template-guided compensation. More precise localization results are obtained with template-guided compensation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.853 + ], + "angle": 0, + "content": "G-L bottleneck vs. different defect size To access the ability of different bottleneck designs to detect anomalies of different sizes, we categorize defect images of LOCO [3] into 6 groups based on their defect sizes, and compare the performance of the restoration networks with LBN only, GBN only, and a combination of the two on these 6 groups in terms of anomaly localization. As shown in Figure 8, the accuracy of anomaly localization varies with the size of defects. LBN performs well on small defects, while GBN exhibits superior performance on large defects. The best performance is achieved by using a combination of LBN and GBN across all groups." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Impact of template-guided compensation The effectiveness of our compensation design, with or without com" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6454" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.465, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.252, + 0.47, + 0.328 + ], + "angle": 0, + "content": "Figure 8: Performance comparison of the restoration networks with distinct bottleneck settings on varying defect sizes in terms of anomaly localization. We divide anomaly samples into 6 groups based on defect sizes, and show the percentage of each group with a bar chart." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.565 + ], + "angle": 0, + "content": "pensation, is shown in Table 5 and Figure 7. Our compensation design improves detection result by \\(2.8\\%\\) and localization result by \\(2.9\\%\\) over the baseline on MVTec LOCO AD, and improves detection result by \\(0.6\\%\\), localization result by \\(0.4\\%\\) and localization (PRO) result by \\(1.1\\%\\) over the baseline on MVTec AD. By incorporating compensation design, it is possible to enhance the precision of localization and achieve more accurate coverage of abnormal areas. We also investigated the impact of relation design in our compensation module. The relation design improves detection result by \\(6.3\\%\\) and localization result by \\(5.4\\%\\) over the baseline on MVTec LOCO AD, and improves detection result by \\(0.5\\%\\), localization result by \\(0.1\\%\\) and localization (PRO) result by \\(0.4\\%\\) over the baseline on MVTec AD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.468, + 0.734 + ], + "angle": 0, + "content": "Impact of template bank subsampling The inference time and memory usage are highly relevant to the size of the template bank. We adopt coreset subsampling [1, 8] to find a subset that best describes the template bank during the testing phase. In Table 6, we compare performance of template banks with different subsampling ratios, and observe that subsampling has little impact on performance of MVTec AD. Although subsampling slightly reduces the mean detection accuracy of MVTec LOCO AD, its detection accuracy surpasses peers by a remarkable margin." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.267, + 0.759 + ], + "angle": 0, + "content": "4.4. Complexity analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We compare inference speed and memory usage with different methods in Table 7. The inference time is measured with Intel Xeon E5-2698v4. Because the backbones are identical (i.e. WideResNet50 [45]), we only report complexity on the additional models. We use image NNS and template bank subsampling to improve searching efficiency and reduce memory cost. Compared with peer works, our method with tempalte bank subsampling to \\(10\\%\\) obtain competitive performance on complexity and accuracy." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.09, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 4: The impact of bottleneck design on MVtec LOCO AD Dataset. Best scores are bolded." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.131, + 0.888, + 0.182 + ], + "angle": 0, + "content": "
Metrics\\BottleneckGBNLBNG-L bottleneck
Image-level79.780.386.0
Pixel-level55.666.874.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.197, + 0.89, + 0.227 + ], + "angle": 0, + "content": "Table 5: Ablation study on template compensation and relation representation. Best scores are bolded." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.239, + 0.89, + 0.32 + ], + "angle": 0, + "content": "
DatasetLOCO [3]AD [4]
w/ temp. w/ rela.ImagePixelImagePixelPixel(PRO)
83.271.298.597.893.9
✓79.768.798.698.194.6
✓86.074.199.298.295.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.335, + 0.892, + 0.365 + ], + "angle": 0, + "content": "Table 6: Anomaly detection performance with template bank subsampling at different ratios." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.377, + 0.891, + 0.472 + ], + "angle": 0, + "content": "
Dataset\\Ratio10%25%50%100%
LOCO [3]Image85.686.086.086.0
Pixel73.774.074.074.1
AD [4]Image99.199.299.299.2
Pixel98.198.298.298.2
Pixel(PRO)95.095.095.095.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.488, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Table 7: The comparison of pre-trained based approaches in terms of inference time (s), memory usage (MB), and performance (image/pixel/pixel-PRO) on MVTec AD [4]." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.545, + 0.891, + 0.653 + ], + "angle": 0, + "content": "
MethodInf. timeMemoryPerformance
CFLOW [14]0.178947(96.8/97.9/92.7)
RD4AD [11]0.079352(98.5/97.8/93.9)
PatchCore(100%) [28]0.1491015(99.1/98.0/93.3)
PatchCore(10%) [28]0.113102(99.0/98.1/93.5)
Ours(100%)0.1301130(99.2/98.2/95.0)
Ours(10%)0.099448(99.1/98.1/95.0)
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.618, + 0.694 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the practical industrial scenarios, it is challenging to identify anomalies from the complicated normal patterns, and meanwhile precisely localize the anomalous regions of various sizes. We propose a novel template-guided hierarchical feature restoration network for anomaly detection. The proposed method recovers anomaly-free features from anomaly features by using bottleneck to filter anomaly features, and compensating compressed features with template retrieved from template bank. The hierarchical design benefits detection upon anomalies of various sizes. Experimental results demonstrate the effectiveness of our approach, especially on localization. Our method achieves state-of-the-art performance on MVTec LOCO AD dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6455" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.158 + ], + "angle": 0, + "content": "[1] Pankaj Agarwal, Sariel Har, Peled Kasturi, and R Varadarajan. Geometric approximation via coresets. Combinatorial and Computational Geometry, 52, 11 2004. 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Samet Akçay, Amir Atapour-Abarghouei, and Toby P. Breckon. Skip-ganomaly: Skip connected and adversarially trained encoder-decoder anomaly detection. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1–8, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130(4):947-969, 2022. 1, 2, 3, 4, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad - a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 1, 2, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.471, + 0.444 + ], + "angle": 0, + "content": "[5] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.446, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[6] Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv preprint arXiv:1807.02011, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.503, + 0.471, + 0.559 + ], + "angle": 0, + "content": "[7] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5008-5017, June 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.561, + 0.47, + 0.602 + ], + "angle": 0, + "content": "[8] Kenneth L. Clarkson. Coresets, sparse greedy approximation, and the frank-wolfe algorithm. ACM Trans. Algorithms, 6(4), sep 2010. 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.604, + 0.471, + 0.645 + ], + "angle": 0, + "content": "[9] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[10] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: A patch distribution modeling framework for anomaly detection and localization. In Alberto Del Bimbo, Rita Cucchiara, Stan Sclaroff, Giovanni Maria Farinella, Tao Mei, Marco Bertini, Hugo Jair Escalante, and Roberto Vezzani, editors, Pattern Recognition. ICPR International Workshops and Challenges, pages 475-489, Cham, 2021. Springer International Publishing. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[11] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9737-9746, June 2022. 1, 3, 4, 5, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[13] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.895, + 0.274 + ], + "angle": 0, + "content": "[14] Denis Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 98-107, January 2022. 2, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.894, + 0.331 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.895, + 0.415 + ], + "angle": 0, + "content": "[16] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8791-8800, October 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.895, + 0.501 + ], + "angle": 0, + "content": "[17] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 303–319, Cham, 2022. Springer Nature Switzerland. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.895, + 0.544 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.895, + 0.573 + ], + "angle": 0, + "content": "[19] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.894, + 0.615 + ], + "angle": 0, + "content": "[20] Yann LeCun et al. Generalization and network design strategies. _Connectionism in perspective_, 19(143-155):18, 1989. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.895, + 0.688 + ], + "angle": 0, + "content": "[21] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9664-9674, June 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.895, + 0.745 + ], + "angle": 0, + "content": "[22] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.895, + 0.816 + ], + "angle": 0, + "content": "[23] Jonathan Pirnay and Keng Chai. Inpainting transformer for anomaly detection. In Stan Sclaroff, Cosimo Distante, Marco Leo, Giovanni M. Farinella, and Federico Tombari, editors, Image Analysis and Processing - ICIAP 2022, pages 394-406, Cham, 2022. Springer International Publishing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.895, + 0.899 + ], + "angle": 0, + "content": "[24] Masoud Pourreza, Bahram Mohammadi, Mostafa Khaki, Samir Bouindour, Hichem Snoussi, and Mohammad Sabokrou. G2d: Generate to detect anomaly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2003-2012, January 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6456" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[25] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2806-2814, June 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[26] Oliver Rippel, Patrick Mertens, and Dorit Merhof. Modeling the distribution of normal data in pre-trained deep features for anomaly detection. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6726-6733, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.472, + 0.331 + ], + "angle": 0, + "content": "[27] Nicolae-Cătălin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13576-13586, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.472, + 0.417 + ], + "angle": 0, + "content": "[28] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14318-14328, June 2022. 1, 2, 3, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.472, + 0.475 + ], + "angle": 0, + "content": "[29] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Structuring autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, Oct 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.472, + 0.547 + ], + "angle": 0, + "content": "[30] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1907-1916, January 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.549, + 0.472, + 0.619 + ], + "angle": 0, + "content": "[31] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmetric student-teacher networks for industrial anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2592–2602, January 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.62, + 0.472, + 0.717 + ], + "angle": 0, + "content": "[32] Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel Müller, and Marius Kloft. Deep one-class classification. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 4393-4402. PMLR, 10-15 Jul 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.472, + 0.802 + ], + "angle": 0, + "content": "[33] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H. Rohban, and Hamid R. Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14902-14912, June 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[34] Robin Schirrmeister, Yuxuan Zhou, Tonio Ball, and Dan Zhang. Understanding anomaly detection with deep invertible networks through hierarchies of distributions and features. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21038-21049. Curran Associates, Inc., 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.176 + ], + "angle": 0, + "content": "[35] Hannah M. Schlüter, Jeremy Tan, Benjamin Hou, and Bernhard Kainz. Natural synthetic anomalies for self-supervised anomaly detection and localization. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022, pages 474-489, Cham, 2022. Springer Nature Switzerland. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.893, + 0.232 + ], + "angle": 0, + "content": "[36] Bernhard Scholkopf, John C. Platt, John Shawe-Taylor, Alex J. Smola, and Robert C. Williamson. Estimating the Support of a High-Dimensional Distribution. Neural Computation, 13(7):1443-1471, 07 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.893, + 0.288 + ], + "angle": 0, + "content": "[37] Ta-Wei Tang, Wei-Han Kuo, Jauh-Hsiang Lan, Chien-Fang Ding, Hakiem Hsu, and Hong-Tsu Young. Anomaly detection neural network with dual auto-encoders gan and its industrial inspection applications. Sensors, 20(12), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.893, + 0.316 + ], + "angle": 0, + "content": "[38] David MJ Tax and Robert PW Duin. Support vector data description. Machine learning, 54(1):45-66, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.893, + 0.399 + ], + "angle": 0, + "content": "[39] Shashanka Venkataramanan, Kuan-Chuan Peng, Rajat Vikram Singh, and Abhijit Mahalanobis. Attention guided anomaly localization in images. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 485-503, Cham, 2020. Springer International Publishing. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.893, + 0.468 + ], + "angle": 0, + "content": "[40] Shinji Yamada, Satoshi Kamiya, and Kazuhiro Hotta. Reconstructed student-teacher and discriminative networks for anomaly detection. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2725–2732, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.893, + 0.538 + ], + "angle": 0, + "content": "[41] Xudong Yan, Huaidong Zhang, Xuemiao Xu, Xiaowei Hu, and Pheng-Ann Heng. Learning semantic context from normal samples for unsupervised anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence, 35(4):3110-3118, May 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.893, + 0.593 + ], + "angle": 0, + "content": "[42] Jihun Yi and Sungroh Yoon. Patch svdd: Patch-level svdd for anomaly detection and segmentation. In Proceedings of the Asian Conference on Computer Vision (ACCV), November 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.893, + 0.663 + ], + "angle": 0, + "content": "[43] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.893, + 0.72 + ], + "angle": 0, + "content": "[44] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint arXiv:2111.07677, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.893, + 0.789 + ], + "angle": 0, + "content": "[45] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Edwin R. Hancock Richard C. Wilson and William A. P. Smith, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 87.1-87.12. BMVA Press, September 2016. 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[46] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem - a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8330-8339, October 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[47] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Reconstruction by inpainting for visual anomaly detection. Pattern Recognition, 112:107706, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6457" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[48] Matthew D. Zeiler, Dilip Krishnan, Graham W. Taylor, and Rob Fergus. Deconvolutional networks. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 2528-2535, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[49] Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, Xin Jin, and Zhibo Chen. Relation-aware global attention for person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[50] Ye Zheng, Xiang Wang, Rui Deng, Tianpeng Bao, Rui Zhao, and Liwei Wu. Focus your distribution: Coarse-to-fine noncontrastive learning for anomaly detection and localization. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.358 + ], + "angle": 0, + "content": "[51] Bo Zong, Qi Song, Martin Renqiang Min, Wei Cheng, Cristian Lumezanu, Daeki Cho, and Haifeng Chen. Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.47, + 0.456 + ], + "angle": 0, + "content": "[52] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 392–408, Cham, 2022. Springer Nature Switzerland. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "6458" + } + ] +] \ No newline at end of file diff --git a/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/b5ce36c9-1639-4be6-bce9-ead12d7207f0_origin.pdf b/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/b5ce36c9-1639-4be6-bce9-ead12d7207f0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e2513169f13d37c663cfc03ba590c4e77d75f224 --- /dev/null +++ b/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/b5ce36c9-1639-4be6-bce9-ead12d7207f0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6540db532e7f433e1cf345f0657e8a5fc85633fa8155769b6a932b08c1701847 +size 6333021 diff --git a/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/full.md b/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4fd5a3da1db133efdaf41031b09aa442ea865b15 --- /dev/null +++ b/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/full.md @@ -0,0 +1,346 @@ +# Template-guided Hierarchical Feature Restoration for Anomaly Detection + +Hewei Guo $^{1*‡}$ Liping Ren $^{2*‡}$ Jingjing Fu $^{3†}$ Yuwang Wang $^{2†}$ Zhizheng Zhang $^{3}$ +Cuiling Lan $^{3}$ Haoqian Wang $^{2}$ Xinwen Hou $^{1}$ $^{1}$ Institute of Automation, Chinese Academy of Sciences + $^{2}$ Tsinghua University ${}^{3}$ Microsoft Research Asia +{guohewei2020, xinwen.hou}@ia.ac.cn +{rlp20@mails., wang-yuwang@mail., wanghaoqian@}tsinghua.edu.cn +{jifu, zhizhhang, culan}@microsoft.com + +# Abstract + +Targeting for detecting anomalies of various sizes for complicated normal patterns, we propose a Template-guided Hierarchical Feature Restoration method, which introduces two key techniques, bottleneck compression and template-guided compensation, for anomaly-free feature restoration. Specially, our framework compresses hierarchical features of an image by bottleneck structure to preserve the most crucial features shared among normal samples. We design template-guided compensation to restore the distorted features towards anomaly-free features. Particularly, we choose the most similar normal sample as the template, and leverage hierarchical features from the template to compensate the distorted features. The bottleneck could partially filter out anomaly features, while the compensation further converts the reminding anomaly features towards normal with template guidance. Finally, anomalies are detected in terms of the cosine distance between the pre-trained features of an inference image and the corresponding restored anomaly-free features. Experimental results demonstrate the effectiveness of our approach, which achieves the state-of-the-art performance on the MVtec LOCO AD dataset. + +# 1. Introduction + +Anomaly detection is typically treated as an out-of-distribution detection problem, which learns the distribution from normal samples during training and detects outliers as anomalies during inference. Existing anomaly detection methods have achieved promising results on the MVTec AD benchmark [4], which is mainly composed of images with + +![](images/a81c492f87a3228e3d087ecb576de674f8ce82e3cef71c9f826b51477f6b3bd2.jpg) +Figure 1: Anomaly detection on different categories. From top to bottom: breakfast box, pushpins, splicing connectors in MVtec LOCO AD [3], and cable, transistor in MVtec AD [4]. From left to right: normal image, anomaly image, anomaly maps of PatchCore [28], RD4AD [11] and our method, and ground truth. The red color corresponds to high anomaly score, whereas the blue represents low anomaly score. Best view in color. + +simple normal patterns, i.e. the material surface is uniform, and background is clean, and only one object exists. However, in many practical applications, normal samples are composed of multiple objects placed by rule, such as PCBs with multiple electronic components, printed surfaces with diverse contents and products with parts assembly. Figure 1 illustrates several examples with complicated structures and predefined normal layout patterns. For example, a normal breakfast box contains oatmeal, nuts, oranges, and peach, which are carefully arranged following a certain rule. Its complicated normal pattern creates difficulties for the existing methods to formulate its normal distribution, and consequently leads to degradation on detection accuracy. More + +over, the accurate anomaly localization is critical in practical industrial scenarios for evaluating the impact of defects and identifying their underlying causes. Therefore, to enhance the usability of anomaly detection methods in various practical scenarios, it is crucial to resolve the challenges associated with complex normal patterns and the requirements for accurate localization. + +Substantial anomaly detection methods can be categorized as the embedding-based paradigm, as shown in Figure 2 (a). These methods identify anomalies based on the patch-level dissimilarity between the input embedding feature extracted from a pre-trained network and the normal feature distribution, through such as kNN [9, 25, 28], Mahalanobis distance [10, 26] or normalizing flows [30, 14, 44]. These methods rely on patch-wise feature comparison and prefer to detect local anomalies rather than global anomalies. Once the normal patterns become complicated, the corresponding normal distribution becomes challenging to be formulated, which in turn harms the detection performance. + +From the aspect of reconstruction quality divergence, numerous reconstruction-based methods [13, 39, 16] are proposed to expect perfect reconstruction on normal regions and poor reconstruction on anomaly regions, as shown in Figure 2 (b), where the samples could be images or features. The reconstruction-based paradigm is highly adaptable to normal samples with complicated distributions, making it suitable for anomaly detection with rich diversity. However, they suffer from a significant drawback that in certain cases they may fail to detect defects, where the anomaly regions are reconstructed with similar quality as the normal regions. Meanwhile, reconstruction-based methods often incorporate autoencoder (AE) that involves several down-sampling operations. Image details are susceptible to being lost due to feature compression operations, which leads to blurry outputs with large reconstruction error even for normal samples [16], and inevitably causes inaccurate identification of defect locations. + +To address the challenges of the complicated normal data and the growing demand on accurate defect localization, we propose Template-guided Hierarchical Feature Restoration (THFR) network for anomaly detection, which restores anomaly-free features. The basic idea of our method is shown in Figure 2 (c), the image features are first compressed with bottleneck structure to filter out anomaly features at different levels, and then the distorted features are compensated with template embedding features, which is retrieved from template bank through image-level nearest neighbor search. The bottleneck structure in our approach is designed to retain the essential features that are common to normal samples. To achieve this, we propose global bottleneck and local bottleneck that respectively preserve normal semantic features and normal detailed features. We establish a template bank using multi-level embedding fea + +tures of normal samples, and retrieve the most similar sample as the template based on the cosine similarity with the input feature. Additionally, we introduce a template-guided compensation module that leverages relation representation between input features and template features to restore anomaly-free features. + +Thanks to the template guidance and hierarchical compensation design, our method achieves state-of-the-art performance on MVTec LOCO AD [3], a dataset containing normal patterns of rich diversity, and also achieves competitive performance on MVTec AD [4], which is commonly used as anomaly detection benchmark. + +In summary, our main contributions are threefold: + +- We propose a new framework to tackle the problem of anomaly detection upon data of rich diversity, in which the anomaly-free features with complicated distribution can be restored from the anomaly features with the guidance of template features. +- We design bottleneck compression to retain normal features while partially filtering out anomaly features, and propose template-guided compensation that leverage template embedding features to restore the distorted features towards normal. +- Extensive experiments on the standard benchmarks demonstrate the outstanding performance of the proposed method, especially for localization. + +# 2. Related work + +Unsupervised anomaly detection methods are trained on normal samples, while testing on both normal and anomaly samples. Classical anomaly detection methods [36, 38, 32, 42] treat the task as one class classification and focus on defining a compact closed one-class distribution. + +Embedding-based methods Embedding-based methods use deep neural networks pre-trained on a large dataset to extract features from an image for anomaly detection and localization. Schirrmeister et al. [34] show that large natural-image datasets such as ImageNet [12] can extract more powerful features than a small application specific dataset. SPADE [9] uses a pre-trained network with multiscale pyramid pooling and segments the anomalies region via among input image features and kNN normal features. PatchCore [28] also uses nearest neighbor to detect anomaly and meanwhile leverages greedy coreset subsampling to lighten memory bank. Different from methods using memory bank, PaDiM [10] abandons slow kNN algorithm and uses Mahalanobis distance metric as an anomaly score. There are follow-up methods [50, 17] to continue to improve the effectiveness of normal features based on Mahalanobis distance. DifferNet [30], CFLOW [14] and Fast- + +![](images/8a4535b285ba8b219b7b7b5628b029e80606ca3db1c2c3cbe218674d009cf947.jpg) +(a) Embedding-based paradigm [28] + +![](images/d8501ee0d7a0c1abb66f498ff7d6c8ee949eb7591cc41fe6d2d1747bc41d448e.jpg) +(b) Reconstruction-based paradigm [16] +Figure 2: Different paradigms for anomaly detection. (a) Embedding-based methods detect anomaly based on patch-level pre-trained feature similarity. (b) Reconstruction-based methods detect anomaly relying on the assumption that anomalies cannot be reconstructed with good quality. (c) THFR method detects anomaly by comparison between the original and restored anomaly-free features. It utilizes a combination of global and local bottlenecks to filter out anomaly features, known as the G-L bottleneck, and restores distorted features with template features retrieved by image-level nearest neighbor search. + +![](images/09d4515cc7598d8f16ac328505680434151da043803c404a45b517b224a2daa9.jpg) +(c) Restoration-based THFR method (ours) + +Flow [44] utilize normalizing flows to explicitly approximate the data density, and detect anomaly samples based on their assigned likelihood. It is challenging for embedding-based methods to model the normal distribution when the patterns are complicated. + +Reconstruction-based methods Reconstruction-based methods are proposed based on an assumption that a model trained on normal data only, cannot represent or reconstruct the anomalies accurately [51, 6]. They typically reconstruct samples from the manifold of the training data, using generative adversarial network (GAN) [20], autoencoder (AE) [29], or variational autoencoder (VAE) [19]. DAGAN [37] trains autoencoder with adversarial losses as the anomaly score of the image. RIAD [47], InTra [23], and SCADN [41] design inpainting frameworks and train models on masked normal data to recover the unseen regions using context for anomaly detection. Deng et al. [11] propose a feature reconstruction network, in which the decoder reconstructs multi-level pre-trained features. Ristea et al. [27] integrate the reconstruction-based functionality into a generic self-supervised block to improve the anomaly detection performance. + +Nevertheless, due to the powerful representation capability of CNNs, considerable anomalies are reconstructed as well as the normal samples, leading to hypothesis failure. Several memory augmented networks [13, 16, 22] are proposed to enlarge the reconstruction quality gap between normal and anomaly regions. However, these learnable memory based methods focus on patch-wise normal feature representation and thus fail to tackle layout anomalies. Meanwhile, the down-sampling operations are commonly used for crucial feature extraction in reconstruction-based methods, causing the loss of detailed information, which harms the pixel-level anomaly localization accuracy. + +Several works [2, 16] adopt skip-connection to assist the reconstruction of details, but anomaly information is also passed to the decoder, resulting in defect leakage. + +Our method aims to restore anomaly-free features from anomaly features instead of recognizing the reconstruction quality gap. For more effective restoration, we filter anomalies via bottleneck and recover the normal pattern and details through template-guided compensation. Note that our method uses the image-level pre-trained features of normal samples as templates to provide both global normal pattern priors and detailed normal feature priors for restoration. + +Distillation-based methods have recently been successfully applied to anomaly detection, where anomalies are identified based on feature differences between a teacher network pre-trained on a large dataset and a student network trained only on normal samples using knowledge distillation [7, 5, 33, 3, 11, 40, 31]. Additionally, some researchers have attempted to convert unsupervised anomaly detection into a supervised learning task by augmenting normal samples with pseudo-anomalies [24, 21, 46, 35]. However, these approaches are prone to bias towards pseudo outliers and fail to detect a large variety of anomaly types. To improve the embedding feature quality, Zou et al. [52] propose SPD which leverage pseudo anomalies as negative sample for contrastive self-supervised pre-training instead of supervised pre-trained with classification. + +# 3. Method + +The overview architecture of our proposed method is shown in Figure 3. Our method first obtains the multi-level embedding features of a given image using a backbone pretrained on ImageNet [12] as feature extractor. In the training stage, these embedding features are treated as the targets of feature restoration. In order to filter out the anomalies of various sizes, the embedding features are compressed by + +![](images/bd4eca6fa9a4868d597b07f46d9b23c9b1191832f941413896b0997a2520f31e.jpg) +Figure 3: Overview architecture of proposed Template-guided Hierarchical Feature Restoration (THFR) framework. (a) THFR consists of one pre-trained backbone encoder, bottleneck, and a restoration decoder. Given a test image, THFR uses ImageNet pre-trained backbone encoder to extract features for restoration target. THFR compresses input features through bottleneck, and the compressed features are compensated by template features retrieved from the template bank using image-level nearest neighbor search (image NNS). Finally, anomaly detection is performed in terms of the cosine distance between restored anomaly-free features and input features. (b) Detailed network design of global bottleneck and local bottleneck. + +![](images/246e2e40bfa543104735597004ed9e8b3c9d29d900e3b6402ca56b5cd7f8c506.jpg) +(b) G-L bottleneck + +global bottleneck (GBN) and local bottleneck (LBN). + +The compression along spatial and channel dimensions inevitably causes information loss on features. To recover the normal pattern with compensation for the lost details, we establish a template bank to record image-level normal embedding features and retrieve the most similar normal feature from the template bank through an image-level nearest neighbor search module called image NNS, and treat the retrieved result as template to guide feature restoration. + +# 3.1. Bottleneck + +We design global bottleneck and local bottleneck that hierarchically compress the embedding features to retain the crucial normal features at multiple levels, and thereby filter out anomaly features caused by anomalies of various sizes. The previous reconstruction-based works [13, 16, 11] use a fixed-scale loose bottleneck, which works on small-scale anomalies, but underperforms when anomalies are of large sizes. In our network, uncompressed template embedding features could provide rich hierarchical normal features while compensating the compressed features, so we could employ the bottleneck of different tightness to effectively filter anomaly features and obtain compact normal representation. + +For a given image, its multi-level pyramid features $I_{k} \in \mathbb{R}^{H_{k} \times W_{k} \times C_{k}}$ are extracted by pre-trained ResNet [15] and fed into bottleneck structure for hierarchical compression, where $H_{k} \times W_{k}$ denotes the spatial dimension, $C_{k}$ is the number of channels and $k$ indicates the layer index with maximum value equal to $K$ . The input feature maps are distinct between global bottleneck and local bottleneck. Specifically, the high-level feature $I_{K}$ is fed to global bottleneck, while $[I_{1}, \dots, I_{K}]$ for local bottleneck. + +Global bottleneck Considerable large-size anomalies result from missing or misplaced normal components [3], which appear to be globally abnormal but locally normal, creating difficulties for existing anomaly detection methods. We propose global bottleneck to preserve the most representative normal semantic features and filter out the anomaly semantic features. As shown in Figure 3 (b), global bottleneck reforms the deep-level feature $I_{K}$ by a ResBlock and then compress the features by global maxpooling, producing a feature vector $Z_{GBN} \in \mathbb{R}^{1 \times 1 \times 2C_K}$ . Global max-pooling extremely compresses the feature spatially to eliminate the anomaly features. It could alleviate the attention on object misalignment and instead focuses on global pattern representation. Finally, we use $2 \times 2$ deconvolution [48] + +with stride 2 to upsample the compressed feature $Z_{GBN}$ to a $8 \times 8$ spatial feature map for succeeding feature restoration. + +Local bottleneck Besides the compression on semantic features, we introduce a trainable local bottleneck to compress the multi-level features to preserve the local normal features and filter out the local anomaly features. The design of the local bottleneck follows the work [11], as shown in Figure 3 (b). To align multi-level features, the shallow features are downsampled through $3 \times 3$ convolution layers with stride of 2. We concatenate multi-level representations and fuse them by a ResBlock to obtain compressed feature $Z_{LBN} \in \mathbb{R}^{\frac{H_K}{2} \times \frac{W_K}{2} \times 2C_K}$ , which is fed into detail compensation module for fine-grained feature restoration. + +# 3.2. Template-guided compensation + +Although the essential normal features are kept during compression, the information loss on normal features is unavoidable and needs to be compensated. We leverage multilevel normal embedding features as template to guide feature restoration. Since there is no one-to-one correspondence between the positional information of input and template features, we leverage the relationship between the input and template features to guide the restoration process. We fuse the input and template by stacking these two embeddings and the relation feature, and then passing them through multiple $1 \times 1$ convolution layers. The fused feature is fed to a series of ResBlocks for further restoration. + +Template bank We construct template bank using image-level pre-trained features extracted from $N$ normal samples in the all train set as $\mathcal{B} = \{B_1^1,\dots ,B_k^i,\dots ,B_K^N\}$ , where $B_{k}^{i}$ represents the $k$ -th layer template feature of the $i$ -th normal sample in template bank. During inference, We could reduce template bank using coreset subsampling method [1, 8] to reduce inference time and memory usage. + +Image-level nearest neighbor search Given an input sample, we take the deepest pre-trained $K$ -th layer feature $I_K$ as the query to retrieve its correlated template. Image NNS obtains the template with index $t$ by randomly selecting from the template candidates which are $n$ most similar templates to increase the robustness during the training process. The template selection process could be formulated as follows: + +$$ +t = \operatorname {r a n d o m} \left(\underset {\mathcal {S} \subset \{1, \dots , N \}, | \mathcal {S} | = n} {\arg \min }\right. \sum_ {i \in \mathcal {S}} d \left(I _ {K}, B _ {K} ^ {i}\right)), \tag {1} +$$ + +where $d(\cdot)$ denotes the image-level distance between input feature query $I_{K}$ and template feature key $B_{K}^{i}$ by flattening them to vectors to compute cosine distance. $S$ is a subset of $\{1,\dots ,N\}$ denotes the indexes of $n$ template candidates, + +![](images/2938306d15959934a1dd5be2fa74b42f75b4bfe64c0fb8c4912e55e9d52b2276.jpg) +Figure 4: Template-guided compensation process. We leverage normal template feature to guide anomaly-free feature restoration with the relation representation between the input feature and the template feature. + +which are the top-n nearest neighbors of the input feature $I_K$ . Note that, the $n$ set to 1 during inference. Compared with the traditional point-by-point search strategy, image NNS not only ensures that the reference is completely normal but also improves search efficiency. The multi-level embedding features of the template sample are used as references in the corresponding feature level, which are denoted as $T$ for simplicity. + +Relation representation We use pairwise relation [49] between input feature and template feature as additional information to restore normal feature. Specifically, we use $r_{I^i \to T^j}$ to represent the relation from the $i$ -th feature point of input feature to the $j$ -th feature point of the template feature. The pairwise relation can be defined as a dot-product affinity in the embedding spaces as: + +$$ +r _ {I ^ {i} \rightarrow T ^ {j}} = \theta \left(I ^ {i}\right) ^ {\top} \phi \left(T ^ {j}\right) \tag {2} +$$ + +where $\theta$ and $\phi$ are two embedding functions implemented by a $1\times 1$ spatial convolutional layer. Similarly, we can get the affinity from $j$ -th feature point of template feature to $i$ -th feature point of input feature as $r_{T^j\rightarrow I^i}$ . We use the pair $(r_{I^i\to T^j},r_{T^j\to I^i})$ to describe the bi-directional relation between $i$ -th feature point of input feature to $j$ -th feature point of template feature. Then, we represent the relation matrix $\mathbf{R}\in \mathbb{R}^{m\times m}$ among all the nodes, where $m$ is the resolution of the feature. The bi-directional + +pairwise relation vector of $i$ -th feature point is $\mathbf{r}_i = [r_{I^i \to T\{1, \dots, m\}}, r_{T^i \to I\{1, \dots, m\}}]$ . For example, as shown in Figure 4, the third row and the third column of the relation matrix, i.e. $\mathbf{r}_3 = [r_{I^3 \to T\{1, \dots, 9\}}, r_{T^3 \to I\{1, \dots, 9\}}]$ . In this relation vector, the first half channels indicate the relation that regards the input feature point as query and template feature point as key, while the second half indicates the relation of reverse operation. The bi-directional pairwise relation vectors of all nodes form the relation feature. + +Feature compensation We leverage normal template features to guide anomaly-free feature restoration. As shown in Figure 4, the input compressed feature with channel size $C$ is projected to input embedding $E_{i} \in \mathbb{R}^{H \times W \times C_{i}}$ , and the template feature with channel size $C$ is projected to template embedding $E_{t} \in \mathbb{R}^{H \times W \times C_{t}}$ . Here, $C_{i}$ represents the channel size of input embedding, and $C_{t}$ represents the channel size of template embedding, while $H$ and $W$ represent the height and width of the feature maps, respectively. The values of $C_{i}$ and $C_{t}$ can be adjusted with respect to the feature compression method used. For global bottleneck compression, $C_{i}$ is set to the original channel size $C$ to provide detailed information, while $C_{t}$ is set to 1 to generate normal layout guidance for high-level semantic feature compensation. For local bottleneck compression, $C_{i}$ is set to 1 to provide spatial layout guidance, and $C_{t}$ is set to $C$ to supply detail feature for low-level feature compensation. Then, we concatenate the input embedding $E_{i}$ , the template embedding $E_{t}$ and the relation feature, and passing them through multiple $1 \times 1$ convolution layers to get the restored feature. In order to explicitly demonstrate the effectiveness of our method, we visualize the restored feature using a visualization network [43] in experiment. + +# 3.3. Loss + +We use the cosine distance between the pre-trained feature and restored feature as restoration loss, as it is capable to evaluate the feature differences at multi-levels regardless of feature dimension differences among each level. Mathematically, for input feature $I$ and restored feature $\hat{I}$ , we calculate their feature cosine distance along the channel axis as restoration loss: + +$$ +L _ {k} = 1 - \frac {I _ {k} ^ {\top} \cdot \hat {I} _ {k}}{\| I _ {k} \| \| \hat {I} _ {k} \|}, \tag {3} +$$ + +where $L_{k}$ denotes the restoration loss of $k$ -th level. + +# 3.4. Anomaly map + +In the testing phase, we detect anomalies in terms of the cosine distance between the pre-trained feature of an inference image and the corresponding restored anomaly-free feature. The restored feature tends to be close to input image feature for the normal regions, and departs on + +anomalies. Mathematically, we calculate pixel-wise cosine distance to obtain a 2D anomaly map in $k$ -th level: + +$$ +M _ {k} (h, w) = 1 - \frac {\left(I _ {k} (h , w)\right) ^ {\top} \cdot \hat {I} _ {k} (h , w)}{\| I _ {k} (h , w) \| \| \hat {I} _ {k} (h , w) \|}, \tag {4} +$$ + +where $(h,w)$ denotes the position of the feature vector, and $M_{k}\in \mathbb{R}^{H_{k}\times W_{k}}$ . The final anomaly map is calculated as: + +$$ +M = \sum_ {k = 1} ^ {K} U _ {k} \left(M _ {k}\right), \tag {5} +$$ + +where $U_{k}(\cdot)$ denotes the upsampling operation of $k$ -th anomaly map. In order to remove the noises in the score map, we smooth anomaly map $M$ by a Gaussian filter. + +# 4. Experiments + +# 4.1. Experimental setup + +Datasets We evaluate the proposed method on two industrial anomaly detection datasets: MVTec LOCO AD [3] and MVTec AD [4]. MVTec LOCO AD consists of 5 real-world sub-datasets for anomaly detection, including 1772 images for training, 304 for validation and 1568 for testing. The normal samples in MVTec LOCO AD often have complicated patterns, and the anomaly samples appear in form of logical or structural anomalies [3]. MVTec AD dataset consists of 15 real-world sub-datasets, with 5 categories of textures and 10 categories of objects. It contains 3629 training images and 1725 test images. Note that all the training images of the two datasets are normal, while some of test images are normal and others are anomalies. + +Implementation details In our experiments, anomaly detection and localization are performed on one category at a time. We resize input images to $256 \times 256$ for data preprocessing and use pre-trained WideResNet50 [45] on ImageNet as feature extractor, with the deepest feature level $K$ set to 3. To train our THFR network, we utilize Adam optimizer [18] and set the learning rate to 0.005. Each model is trained for 200 epochs with a batch size of 16 on NVIDIA Tesla V100 GPU. We use Gaussian filter with $\sigma = 2$ to smooth anomaly map on MVtec LOCO AD, while $\sigma = 4$ on MVtec AD. For training sample pair selection, we use $n = 1$ for all categories in MVtec LOCO AD and $n = 6$ for all categories in MVtec AD. + +Evaluation metrics To fairly compare with other methods, we follow the common evaluation metrics of each dataset. For MVTec AD [4], we take area under the receiver operating characteristic (AUROC) as the image-level and pixel-level evaluation metric, and per-region-overlap (PRO) for pixel-level evaluation metric, which can better evaluate + +Table 1: Pixel-level anomaly localization accuracy on MVtec LOCO AD dataset (sPRO) [3]. Remarkably, our approach achieves state-of-the-art performance in four categories and the mean of all categories. Best and second-best scores are bolded and underlined. + +
MethodBreakfastScrew BagPushpinsConnectorsJuice BottleMean
S-T [5]49.660.252.369.881.162.6
DR/EM [46]49.949.049.367.380.059.1
CFLOW [14]48.660.255.171.683.663.8
RD4AD [11]42.257.461.471.385.163.5
PatchCore [28]51.659.853.575.783.264.7
GCAD [3]50.255.873.979.891.070.1
THFR (ours)58.361.576.384.889.674.1
+ +Table 2: Image-level anomaly detection accuracy on MVtec LOCO AD dataset (AUROC) [3]. Remarkably, our approach achieves the state-of-the-art performance on four categories and the mean of all categories. Best and second-best scores are bolded and underlined. + +
MethodBreakfastScrew BagPushpinsConnectorsJuice BottleMean
S-T [5]-----77.3
DR/EM [46]75.772.776.082.593.980.1
CFLOW [14]77.473.076.082.695.380.8
RD4AD [11]68.774.975.984.494.879.7
PatchCore [28]77.173.374.186.094.681.0
GCAD [3]-----83.3
THFR (ours)78.073.788.392.797.186.0
+ +Table 3: Anomaly detection and localization results on MVTec AD dataset [4], including image-level and pixel-level AUROC, and pixel-level PRO. Best and second-best scores are bolded and underlined. * To ensure fair comparison with previous studies, CFLOW [14] is evaluated using input images with resolution of 256x256 pixels. + +
MethodImagePixelPixel(PRO)
SPADE [9]85.596.091.7
PaDiM [10]95.397.592.1
S-T [5]--91.4
DR/EM [46]98.097.3-
CFLOW * [14]96.897.992.7
RD4AD [11]98.597.893.9
PatchCore [28]99.198.193.4
THFR (ours)99.298.295.0
+ +the performance of small size anomalies through weighting ground-truth regions of different size equally. Following the protocol mentioned in [5], we evaluate the PRO value for a large number of increasing thresholds until an average per-pixel false-positive rate of $30\%$ for the entire dataset is reached. For MVTec LOCO AD [3], we use AUROC to measure image-level performance, and use saturated-per-region-overlaps (sPRO) [3] with per-pixel false-positive rate of $5\%$ to evaluate pixel-level performance. + +# 4.2. Main results + +We compare the proposed model with several deep learning-based methods for anomaly detection as baselines, including embedding-based methods [9, 10, 14, 28], reconstruction-based methods [11, 46], distillation-based methods [5, 11, 3], and data-augmented methods [46]. + +MVTec LOCO AD Anomaly detection results on MVTec LOCO AD [3] are presented in Table 1 and Table 2. In comparison with the SOTA method [3], our approach obtain an absolute AUROC gain of $2.7\%$ for the image-level detection, and an remarkable absolute sPRO gain of $4.0\%$ for the pixel-level localization. Our approach achieves new state-of-the-art performance on both detection and localization results, and surpasses other methods by a large margin. More detailed results will be provided in the supplementary materials. + +MVTec AD Quantitative comparison results on MVTec AD [4] between baselines and our approach are summarized in Table 3. For overall categories, our method produces comparable results with the other advanced methods. Our approach excels in achieving high accuracy on both image-level detection and pixel-level localization at the same time. More detailed results will be presented in the supplementary materials. + +![](images/c6788f70a6555af00e53aaf5356ef19037be6aee330666c6d2979f16778fa657.jpg) +Normal + +![](images/1fc883d1f989afc10215983b9779d4b486ae0c83d6f5d90bd4ac1349b3de38ac.jpg) +Anomaly + +![](images/773b7c428cf547c1ee7372aae07448f5ffdb1abe0661d8690b8c7e39ac02ed68.jpg) +Recon + +![](images/d95ad2598904e39c77050c4b0d06e3782152a292bb38cf3918b4371a2ee1fb7b.jpg) +GBN +Figure 5: Visualization of restored features from different restoration networks. From left to right: normal image, anomaly image, reconstructed images from the pre-trained embedding features (Recon), restored features of restoration network only with GBN, only with LBN, with G-L bottleneck, and with our template-guided hierarchical feature restoration framework, anomaly map, and ground truth. The visualization results show how restoration networks progressively achieve anomaly-free restoration with G-L bottleneck and hierarchical feature compensation. + +![](images/855c6f8ba47840794975389d75c06801867ea1761992514e9ecbf4ffca75d9ce.jpg) +LBN + +![](images/74643412a8c200c4fa0f76a4629944bcbb67f4f1390a42324a0d3e476b479fcd.jpg) +G-L bottleneck + +![](images/ec141cc2a6a517b9e3c999fcbec9de14bbe10b77aca8f6adfc474f1967180905.jpg) +THFR + +![](images/eeb24bdd9f397180733332fd91b550dd895f5fddaf6ec045b76cf737d080bd53.jpg) +Anomaly map + +![](images/7ec064cf12a89045a325b84b3da3e29b8a081c6906d92320f3dc211c6bf47510.jpg) +GT + +# 4.3. Ablation study + +To explicitly demonstrate the effectiveness of our method, we visualize the restored features using a visualization network. We investigate the effectiveness of bottleneck design and template-guided compensation, and study the influence of subsampling within the template bank. + +Visualization of restored features As shown in Figure 5, we visually compared the pre-trained feature with the restored feature acquired from the network utilizing a visualization network [43]. We observe that the restoration networks employing solely local bottleneck (LBN) cannot effectively filter out anomalies that violate complex normal rules, such as missing connector heads and absent labels on the juice bottle. On the other hand, restoration networks with only global bottleneck (GBN) can filter out rule-breaking anomalies, but it encounters challenges in accurately restoring intricate details, such as the recovery of oatmeal and banana slices in the breakfast box. The restoration network with G-L bottleneck also has anomaly feature leakage after bottleneck compression, manifesting as anomalies like the irregular coloration of the juice. In contrast, our THFR network recovers anomaly-free feature through both semantic level correction and detailed feature compensation, resulting in an effective restoration to normal of various types of anomaly areas. + +Impact of bottleneck Table 4 shows anomaly detection performance of local bottleneck and global bottleneck on MVTec LOCO AD dataset. The dataset has a rich diversity of data and anomalies, and the combination of local bottleneck and global bottleneck yields substantial enhancements in detection and localization results. Figure 6 illustrates that the global bottleneck primarily concentrates on filtering large-size anomalies, while the local bottleneck excels in filtering small-size anomalies. + +![](images/0f0fe87b349db9bd226af66d5f2c28ced2cebf0a14d49da7becb6ad2d85325f5.jpg) +Normal + +![](images/ebfe71ef8140a0b7c97158f49a7ba37a532b4cbbe034b2b689c2ca18ae3048ea.jpg) +Anomaly + +![](images/faf4945f46ba28fcde9da1765cf9df1eceeb1222058b752cd2ea06087a3998f1.jpg) +GBN + +![](images/f02cc670a86be81c5ec6808f72f3f5ff7651d365ec4d5844b37b89db3a06cda3.jpg) +LBN + +![](images/232d52900d0adab7f8ea03f067916c8e27e8c37e5f7d5c500b948036c0d296e6.jpg) +GT + +![](images/74e7d20e35118fe25a4ab49b325aa4104fc38b50e274c183d7bc4b7ac9fac548.jpg) +Figure 6: Visualization of anomaly detection results with different bottlenecks. +Normal +Figure 7: Visualization of anomaly detection results with or without template-guided compensation. More precise localization results are obtained with template-guided compensation. + +![](images/e656beda0a91fed597cce3ac0741a91e907961decfb5defbb47ef1c8b711d4a1.jpg) +Anomaly + +![](images/9a029d4afb646fb8a6d2d22fc47d7367ab15a06cf7b2c00423afeadcba6f5cf6.jpg) +W/O comp. + +![](images/3511ca2481eea3975bdcd5dff2f760096af3872c6e0a12e3dfd6a7e406b1d462.jpg) +W/ comp. + +![](images/c5d22b4ca3a2ac4c78cfe3f563e4f065a174c8700993a1aa8e7eec4058ef61cc.jpg) +GT + +G-L bottleneck vs. different defect size To access the ability of different bottleneck designs to detect anomalies of different sizes, we categorize defect images of LOCO [3] into 6 groups based on their defect sizes, and compare the performance of the restoration networks with LBN only, GBN only, and a combination of the two on these 6 groups in terms of anomaly localization. As shown in Figure 8, the accuracy of anomaly localization varies with the size of defects. LBN performs well on small defects, while GBN exhibits superior performance on large defects. The best performance is achieved by using a combination of LBN and GBN across all groups. + +Impact of template-guided compensation The effectiveness of our compensation design, with or without com + +![](images/327c5000445a398255a8688ef4fb028a21738eddfc162e660b3910ddd1e97d8f.jpg) +Figure 8: Performance comparison of the restoration networks with distinct bottleneck settings on varying defect sizes in terms of anomaly localization. We divide anomaly samples into 6 groups based on defect sizes, and show the percentage of each group with a bar chart. + +pensation, is shown in Table 5 and Figure 7. Our compensation design improves detection result by $2.8\%$ and localization result by $2.9\%$ over the baseline on MVTec LOCO AD, and improves detection result by $0.6\%$ , localization result by $0.4\%$ and localization (PRO) result by $1.1\%$ over the baseline on MVTec AD. By incorporating compensation design, it is possible to enhance the precision of localization and achieve more accurate coverage of abnormal areas. We also investigated the impact of relation design in our compensation module. The relation design improves detection result by $6.3\%$ and localization result by $5.4\%$ over the baseline on MVTec LOCO AD, and improves detection result by $0.5\%$ , localization result by $0.1\%$ and localization (PRO) result by $0.4\%$ over the baseline on MVTec AD. + +Impact of template bank subsampling The inference time and memory usage are highly relevant to the size of the template bank. We adopt coreset subsampling [1, 8] to find a subset that best describes the template bank during the testing phase. In Table 6, we compare performance of template banks with different subsampling ratios, and observe that subsampling has little impact on performance of MVTec AD. Although subsampling slightly reduces the mean detection accuracy of MVTec LOCO AD, its detection accuracy surpasses peers by a remarkable margin. + +# 4.4. Complexity analysis + +We compare inference speed and memory usage with different methods in Table 7. The inference time is measured with Intel Xeon E5-2698v4. Because the backbones are identical (i.e. WideResNet50 [45]), we only report complexity on the additional models. We use image NNS and template bank subsampling to improve searching efficiency and reduce memory cost. Compared with peer works, our method with tempalte bank subsampling to $10\%$ obtain competitive performance on complexity and accuracy. + +Table 4: The impact of bottleneck design on MVtec LOCO AD Dataset. Best scores are bolded. + +
Metrics\BottleneckGBNLBNG-L bottleneck
Image-level79.780.386.0
Pixel-level55.666.874.1
+ +Table 5: Ablation study on template compensation and relation representation. Best scores are bolded. + +
DatasetLOCO [3]AD [4]
w/ temp. w/ rela.ImagePixelImagePixelPixel(PRO)
83.271.298.597.893.9
✓79.768.798.698.194.6
✓86.074.199.298.295.0
+ +Table 6: Anomaly detection performance with template bank subsampling at different ratios. + +
Dataset\Ratio10%25%50%100%
LOCO [3]Image85.686.086.086.0
Pixel73.774.074.074.1
AD [4]Image99.199.299.299.2
Pixel98.198.298.298.2
Pixel(PRO)95.095.095.095.0
+ +Table 7: The comparison of pre-trained based approaches in terms of inference time (s), memory usage (MB), and performance (image/pixel/pixel-PRO) on MVTec AD [4]. + +
MethodInf. timeMemoryPerformance
CFLOW [14]0.178947(96.8/97.9/92.7)
RD4AD [11]0.079352(98.5/97.8/93.9)
PatchCore(100%) [28]0.1491015(99.1/98.0/93.3)
PatchCore(10%) [28]0.113102(99.0/98.1/93.5)
Ours(100%)0.1301130(99.2/98.2/95.0)
Ours(10%)0.099448(99.1/98.1/95.0)
+ +# 5. Conclusion + +In the practical industrial scenarios, it is challenging to identify anomalies from the complicated normal patterns, and meanwhile precisely localize the anomalous regions of various sizes. We propose a novel template-guided hierarchical feature restoration network for anomaly detection. The proposed method recovers anomaly-free features from anomaly features by using bottleneck to filter anomaly features, and compensating compressed features with template retrieved from template bank. The hierarchical design benefits detection upon anomalies of various sizes. Experimental results demonstrate the effectiveness of our approach, especially on localization. Our method achieves state-of-the-art performance on MVTec LOCO AD dataset. + +# References + +[1] Pankaj Agarwal, Sariel Har, Peled Kasturi, and R Varadarajan. Geometric approximation via coresets. Combinatorial and Computational Geometry, 52, 11 2004. 5, 9 +[2] Samet Akçay, Amir Atapour-Abarghouei, and Toby P. Breckon. Skip-ganomaly: Skip connected and adversarially trained encoder-decoder anomaly detection. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1–8, 2019. 3 +[3] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130(4):947-969, 2022. 1, 2, 3, 4, 6, 7, 8, 9 +[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad - a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 1, 2, 6, 7, 9 +[5] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3, 7 +[6] Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv preprint arXiv:1807.02011, 2018. 3 +[7] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5008-5017, June 2021. 3 +[8] Kenneth L. Clarkson. Coresets, sparse greedy approximation, and the frank-wolfe algorithm. ACM Trans. Algorithms, 6(4), sep 2010. 5, 9 +[9] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 2, 7 +[10] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: A patch distribution modeling framework for anomaly detection and localization. In Alberto Del Bimbo, Rita Cucchiara, Stan Sclaroff, Giovanni Maria Farinella, Tao Mei, Marco Bertini, Hugo Jair Escalante, and Roberto Vezzani, editors, Pattern Recognition. ICPR International Workshops and Challenges, pages 475-489, Cham, 2021. Springer International Publishing. 2, 7 +[11] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9737-9746, June 2022. 1, 3, 4, 5, 7, 9 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 3 + +[13] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2, 3, 4 +[14] Denis Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 98-107, January 2022. 2, 7, 9 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 4 +[16] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8791-8800, October 2021. 2, 3, 4 +[17] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 303–319, Cham, 2022. Springer Nature Switzerland. 2 +[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[19] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[20] Yann LeCun et al. Generalization and network design strategies. _Connectionism in perspective_, 19(143-155):18, 1989. 3 +[21] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9664-9674, June 2021. 3 +[22] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3 +[23] Jonathan Pirnay and Keng Chai. Inpainting transformer for anomaly detection. In Stan Sclaroff, Cosimo Distante, Marco Leo, Giovanni M. Farinella, and Federico Tombari, editors, Image Analysis and Processing - ICIAP 2022, pages 394-406, Cham, 2022. Springer International Publishing. 3 +[24] Masoud Pourreza, Bahram Mohammadi, Mostafa Khaki, Samir Bouindour, Hichem Snoussi, and Mohammad Sabokrou. G2d: Generate to detect anomaly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2003-2012, January 2021. 3 + +[25] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2806-2814, June 2021. 2 +[26] Oliver Rippel, Patrick Mertens, and Dorit Merhof. Modeling the distribution of normal data in pre-trained deep features for anomaly detection. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6726-6733, 2021. 2 +[27] Nicolae-Cătălin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13576-13586, June 2022. 3 +[28] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14318-14328, June 2022. 1, 2, 3, 7, 9 +[29] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Structuring autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, Oct 2019. 3 +[30] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1907-1916, January 2021. 2 +[31] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmetric student-teacher networks for industrial anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2592–2602, January 2023. 3 +[32] Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel Müller, and Marius Kloft. Deep one-class classification. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 4393-4402. PMLR, 10-15 Jul 2018. 2 +[33] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H. Rohban, and Hamid R. Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14902-14912, June 2021. 3 +[34] Robin Schirrmeister, Yuxuan Zhou, Tonio Ball, and Dan Zhang. Understanding anomaly detection with deep invertible networks through hierarchies of distributions and features. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21038-21049. Curran Associates, Inc., 2020. 2 + +[35] Hannah M. Schlüter, Jeremy Tan, Benjamin Hou, and Bernhard Kainz. Natural synthetic anomalies for self-supervised anomaly detection and localization. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022, pages 474-489, Cham, 2022. Springer Nature Switzerland. 3 +[36] Bernhard Scholkopf, John C. Platt, John Shawe-Taylor, Alex J. Smola, and Robert C. Williamson. Estimating the Support of a High-Dimensional Distribution. Neural Computation, 13(7):1443-1471, 07 2001. 2 +[37] Ta-Wei Tang, Wei-Han Kuo, Jauh-Hsiang Lan, Chien-Fang Ding, Hakiem Hsu, and Hong-Tsu Young. Anomaly detection neural network with dual auto-encoders gan and its industrial inspection applications. Sensors, 20(12), 2020. 3 +[38] David MJ Tax and Robert PW Duin. Support vector data description. Machine learning, 54(1):45-66, 2004. 2 +[39] Shashanka Venkataramanan, Kuan-Chuan Peng, Rajat Vikram Singh, and Abhijit Mahalanobis. Attention guided anomaly localization in images. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 485-503, Cham, 2020. Springer International Publishing. 2 +[40] Shinji Yamada, Satoshi Kamiya, and Kazuhiro Hotta. Reconstructed student-teacher and discriminative networks for anomaly detection. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2725–2732, 2022. 3 +[41] Xudong Yan, Huaidong Zhang, Xuemiao Xu, Xiaowei Hu, and Pheng-Ann Heng. Learning semantic context from normal samples for unsupervised anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence, 35(4):3110-3118, May 2021. 3 +[42] Jihun Yi and Sungroh Yoon. Patch svdd: Patch-level svdd for anomaly detection and segmentation. In Proceedings of the Asian Conference on Computer Vision (ACCV), November 2020. 2 +[43] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 6, 8 +[44] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint arXiv:2111.07677, 2021. 2, 3 +[45] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Edwin R. Hancock Richard C. Wilson and William A. P. Smith, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 87.1-87.12. BMVA Press, September 2016. 6, 9 +[46] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem - a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8330-8339, October 2021. 3, 7 +[47] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Reconstruction by inpainting for visual anomaly detection. Pattern Recognition, 112:107706, 2021. 3 + +[48] Matthew D. Zeiler, Dilip Krishnan, Graham W. Taylor, and Rob Fergus. Deconvolutional networks. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 2528-2535, 2010. 4 +[49] Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, Xin Jin, and Zhibo Chen. Relation-aware global attention for person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5 +[50] Ye Zheng, Xiang Wang, Rui Deng, Tianpeng Bao, Rui Zhao, and Liwei Wu. Focus your distribution: Coarse-to-fine noncontrastive learning for anomaly detection and localization. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2022. 2 +[51] Bo Zong, Qi Song, Martin Renqiang Min, Wei Cheng, Cristian Lumezanu, Daeki Cho, and Haifeng Chen. Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In International Conference on Learning Representations, 2018. 3 +[52] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 392–408, Cham, 2022. Springer Nature Switzerland. 3 \ No newline at end of file diff --git a/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/images.zip b/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..603004c6f81aba37361b9801b9afb99eec40378e --- /dev/null +++ b/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a64e626cbc058ca497579129034720a54d8e0da0a06a67e6a7c0a4bb803742a +size 621047 diff --git a/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/layout.json b/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c38b08108ab2e24434b28377a388a4e7484dce18 --- /dev/null +++ b/2023/Template-guided Hierarchical Feature Restoration for Anomaly Detection/layout.json @@ -0,0 +1,10221 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 524, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 524, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 524, + 121 + ], + "type": "text", + "content": "Template-guided Hierarchical Feature Restoration for Anomaly Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "spans": [ + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "text", + "content": "Hewei Guo" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "inline_equation", + "content": "^{1*‡}" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "text", + "content": " Liping Ren" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "inline_equation", + "content": "^{2*‡}" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "text", + "content": " Jingjing Fu" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "inline_equation", + "content": "^{3†}" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "text", + "content": " Yuwang Wang" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "inline_equation", + "content": "^{2†}" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "text", + "content": " Zhizheng Zhang" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "text", + "content": " \nCuiling Lan" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "text", + "content": " Haoqian Wang" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "text", + "content": " Xinwen Hou" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "text", + "content": "Institute of Automation, Chinese Academy of Sciences \n" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 99, + 141, + 494, + 242 + ], + "type": "text", + "content": "Microsoft Research Asia \n{guohewei2020, xinwen.hou}@ia.ac.cn \n{rlp20@mails., wang-yuwang@mail., wanghaoqian@}tsinghua.edu.cn \n{jifu, zhizhhang, culan}@microsoft.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "spans": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 294, + 290, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 290, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 290, + 558 + ], + "type": "text", + "content": "Targeting for detecting anomalies of various sizes for complicated normal patterns, we propose a Template-guided Hierarchical Feature Restoration method, which introduces two key techniques, bottleneck compression and template-guided compensation, for anomaly-free feature restoration. Specially, our framework compresses hierarchical features of an image by bottleneck structure to preserve the most crucial features shared among normal samples. We design template-guided compensation to restore the distorted features towards anomaly-free features. Particularly, we choose the most similar normal sample as the template, and leverage hierarchical features from the template to compensate the distorted features. The bottleneck could partially filter out anomaly features, while the compensation further converts the reminding anomaly features towards normal with template guidance. Finally, anomalies are detected in terms of the cosine distance between the pre-trained features of an inference image and the corresponding restored anomaly-free features. Experimental results demonstrate the effectiveness of our approach, which achieves the state-of-the-art performance on the MVtec LOCO AD dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 582, + 128, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 128, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 128, + 594 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 602, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 602, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 602, + 287, + 675 + ], + "type": "text", + "content": "Anomaly detection is typically treated as an out-of-distribution detection problem, which learns the distribution from normal samples during training and detects outliers as anomalies during inference. Existing anomaly detection methods have achieved promising results on the MVTec AD benchmark [4], which is mainly composed of images with" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 266, + 543, + 428 + ], + "blocks": [ + { + "bbox": [ + 307, + 266, + 543, + 428 + ], + "lines": [ + { + "bbox": [ + 307, + 266, + 543, + 428 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 543, + 428 + ], + "type": "image", + "image_path": "a81c492f87a3228e3d087ecb576de674f8ce82e3cef71c9f826b51477f6b3bd2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": "Figure 1: Anomaly detection on different categories. From top to bottom: breakfast box, pushpins, splicing connectors in MVtec LOCO AD [3], and cable, transistor in MVtec AD [4]. From left to right: normal image, anomaly image, anomaly maps of PatchCore [28], RD4AD [11] and our method, and ground truth. The red color corresponds to high anomaly score, whereas the blue represents low anomaly score. Best view in color." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 715 + ], + "type": "text", + "content": "simple normal patterns, i.e. the material surface is uniform, and background is clean, and only one object exists. However, in many practical applications, normal samples are composed of multiple objects placed by rule, such as PCBs with multiple electronic components, printed surfaces with diverse contents and products with parts assembly. Figure 1 illustrates several examples with complicated structures and predefined normal layout patterns. For example, a normal breakfast box contains oatmeal, nuts, oranges, and peach, which are carefully arranged following a certain rule. Its complicated normal pattern creates difficulties for the existing methods to formulate its normal distribution, and consequently leads to degradation on detection accuracy. More" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 683, + 126, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 126, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 126, + 693 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 59, + 693, + 136, + 703 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 58, + 703, + 257, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 257, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 257, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 58, + 703, + 257, + 712 + ], + "type": "text", + "content": "Work done during an internship at Microsoft Research Asia." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6447" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "over, the accurate anomaly localization is critical in practical industrial scenarios for evaluating the impact of defects and identifying their underlying causes. Therefore, to enhance the usability of anomaly detection methods in various practical scenarios, it is crucial to resolve the challenges associated with complex normal patterns and the requirements for accurate localization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 287, + 302 + ], + "type": "text", + "content": "Substantial anomaly detection methods can be categorized as the embedding-based paradigm, as shown in Figure 2 (a). These methods identify anomalies based on the patch-level dissimilarity between the input embedding feature extracted from a pre-trained network and the normal feature distribution, through such as kNN [9, 25, 28], Mahalanobis distance [10, 26] or normalizing flows [30, 14, 44]. These methods rely on patch-wise feature comparison and prefer to detect local anomalies rather than global anomalies. Once the normal patterns become complicated, the corresponding normal distribution becomes challenging to be formulated, which in turn harms the detection performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 304, + 288, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 519 + ], + "type": "text", + "content": "From the aspect of reconstruction quality divergence, numerous reconstruction-based methods [13, 39, 16] are proposed to expect perfect reconstruction on normal regions and poor reconstruction on anomaly regions, as shown in Figure 2 (b), where the samples could be images or features. The reconstruction-based paradigm is highly adaptable to normal samples with complicated distributions, making it suitable for anomaly detection with rich diversity. However, they suffer from a significant drawback that in certain cases they may fail to detect defects, where the anomaly regions are reconstructed with similar quality as the normal regions. Meanwhile, reconstruction-based methods often incorporate autoencoder (AE) that involves several down-sampling operations. Image details are susceptible to being lost due to feature compression operations, which leads to blurry outputs with large reconstruction error even for normal samples [16], and inevitably causes inaccurate identification of defect locations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": "To address the challenges of the complicated normal data and the growing demand on accurate defect localization, we propose Template-guided Hierarchical Feature Restoration (THFR) network for anomaly detection, which restores anomaly-free features. The basic idea of our method is shown in Figure 2 (c), the image features are first compressed with bottleneck structure to filter out anomaly features at different levels, and then the distorted features are compensated with template embedding features, which is retrieved from template bank through image-level nearest neighbor search. The bottleneck structure in our approach is designed to retain the essential features that are common to normal samples. To achieve this, we propose global bottleneck and local bottleneck that respectively preserve normal semantic features and normal detailed features. We establish a template bank using multi-level embedding fea" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 143 + ], + "type": "text", + "content": "tures of normal samples, and retrieve the most similar sample as the template based on the cosine similarity with the input feature. Additionally, we introduce a template-guided compensation module that leverages relation representation between input features and template features to restore anomaly-free features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 215 + ], + "type": "text", + "content": "Thanks to the template guidance and hierarchical compensation design, our method achieves state-of-the-art performance on MVTec LOCO AD [3], a dataset containing normal patterns of rich diversity, and also achieves competitive performance on MVTec AD [4], which is commonly used as anomaly detection benchmark." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 216, + 519, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 216, + 519, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 216, + 519, + 227 + ], + "type": "text", + "content": "In summary, our main contributions are threefold:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 235, + 545, + 405 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 235, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 235, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 317, + 235, + 545, + 294 + ], + "type": "text", + "content": "- We propose a new framework to tackle the problem of anomaly detection upon data of rich diversity, in which the anomaly-free features with complicated distribution can be restored from the anomaly features with the guidance of template features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 302, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 302, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 317, + 302, + 545, + 361 + ], + "type": "text", + "content": "- We design bottleneck compression to retain normal features while partially filtering out anomaly features, and propose template-guided compensation that leverage template embedding features to restore the distorted features towards normal." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 369, + 545, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 369, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 317, + 369, + 545, + 405 + ], + "type": "text", + "content": "- Extensive experiments on the standard benchmarks demonstrate the outstanding performance of the proposed method, especially for localization." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 415, + 390, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 390, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 390, + 428 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 436, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 495 + ], + "type": "text", + "content": "Unsupervised anomaly detection methods are trained on normal samples, while testing on both normal and anomaly samples. Classical anomaly detection methods [36, 38, 32, 42] treat the task as one class classification and focus on defining a compact closed one-class distribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Embedding-based methods Embedding-based methods use deep neural networks pre-trained on a large dataset to extract features from an image for anomaly detection and localization. Schirrmeister et al. [34] show that large natural-image datasets such as ImageNet [12] can extract more powerful features than a small application specific dataset. SPADE [9] uses a pre-trained network with multiscale pyramid pooling and segments the anomalies region via among input image features and kNN normal features. PatchCore [28] also uses nearest neighbor to detect anomaly and meanwhile leverages greedy coreset subsampling to lighten memory bank. Different from methods using memory bank, PaDiM [10] abandons slow kNN algorithm and uses Mahalanobis distance metric as an anomaly score. There are follow-up methods [50, 17] to continue to improve the effectiveness of normal features based on Mahalanobis distance. DifferNet [30], CFLOW [14] and Fast-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6448" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 100, + 185, + 185 + ], + "blocks": [ + { + "bbox": [ + 50, + 100, + 185, + 185 + ], + "lines": [ + { + "bbox": [ + 50, + 100, + 185, + 185 + ], + "spans": [ + { + "bbox": [ + 50, + 100, + 185, + 185 + ], + "type": "image", + "image_path": "8a4535b285ba8b219b7b7b5628b029e80606ca3db1c2c3cbe218674d009cf947.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 188, + 176, + 198 + ], + "lines": [ + { + "bbox": [ + 58, + 188, + 176, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 188, + 176, + 198 + ], + "type": "text", + "content": "(a) Embedding-based paradigm [28]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 197, + 122, + 329, + 185 + ], + "blocks": [ + { + "bbox": [ + 197, + 122, + 329, + 185 + ], + "lines": [ + { + "bbox": [ + 197, + 122, + 329, + 185 + ], + "spans": [ + { + "bbox": [ + 197, + 122, + 329, + 185 + ], + "type": "image", + "image_path": "d8501ee0d7a0c1abb66f498ff7d6c8ee949eb7591cc41fe6d2d1747bc41d448e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 188, + 328, + 198 + ], + "lines": [ + { + "bbox": [ + 198, + 188, + 328, + 198 + ], + "spans": [ + { + "bbox": [ + 198, + 188, + 328, + 198 + ], + "type": "text", + "content": "(b) Reconstruction-based paradigm [16]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 208, + 545, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 545, + 268 + ], + "type": "text", + "content": "Figure 2: Different paradigms for anomaly detection. (a) Embedding-based methods detect anomaly based on patch-level pre-trained feature similarity. (b) Reconstruction-based methods detect anomaly relying on the assumption that anomalies cannot be reconstructed with good quality. (c) THFR method detects anomaly by comparison between the original and restored anomaly-free features. It utilizes a combination of global and local bottlenecks to filter out anomaly features, known as the G-L bottleneck, and restores distorted features with template features retrieved by image-level nearest neighbor search." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 342, + 80, + 536, + 185 + ], + "blocks": [ + { + "bbox": [ + 342, + 80, + 536, + 185 + ], + "lines": [ + { + "bbox": [ + 342, + 80, + 536, + 185 + ], + "spans": [ + { + "bbox": [ + 342, + 80, + 536, + 185 + ], + "type": "image", + "image_path": "09d4515cc7598d8f16ac328505680434151da043803c404a45b517b224a2daa9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 188, + 512, + 198 + ], + "lines": [ + { + "bbox": [ + 373, + 188, + 512, + 198 + ], + "spans": [ + { + "bbox": [ + 373, + 188, + 512, + 198 + ], + "type": "text", + "content": "(c) Restoration-based THFR method (ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 289, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 350 + ], + "type": "text", + "content": "Flow [44] utilize normalizing flows to explicitly approximate the data density, and detect anomaly samples based on their assigned likelihood. It is challenging for embedding-based methods to model the normal distribution when the patterns are complicated." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 366, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 569 + ], + "type": "text", + "content": "Reconstruction-based methods Reconstruction-based methods are proposed based on an assumption that a model trained on normal data only, cannot represent or reconstruct the anomalies accurately [51, 6]. They typically reconstruct samples from the manifold of the training data, using generative adversarial network (GAN) [20], autoencoder (AE) [29], or variational autoencoder (VAE) [19]. DAGAN [37] trains autoencoder with adversarial losses as the anomaly score of the image. RIAD [47], InTra [23], and SCADN [41] design inpainting frameworks and train models on masked normal data to recover the unseen regions using context for anomaly detection. Deng et al. [11] propose a feature reconstruction network, in which the decoder reconstructs multi-level pre-trained features. Ristea et al. [27] integrate the reconstruction-based functionality into a generic self-supervised block to improve the anomaly detection performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Nevertheless, due to the powerful representation capability of CNNs, considerable anomalies are reconstructed as well as the normal samples, leading to hypothesis failure. Several memory augmented networks [13, 16, 22] are proposed to enlarge the reconstruction quality gap between normal and anomaly regions. However, these learnable memory based methods focus on patch-wise normal feature representation and thus fail to tackle layout anomalies. Meanwhile, the down-sampling operations are commonly used for crucial feature extraction in reconstruction-based methods, causing the loss of detailed information, which harms the pixel-level anomaly localization accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 289, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 324 + ], + "type": "text", + "content": "Several works [2, 16] adopt skip-connection to assist the reconstruction of details, but anomaly information is also passed to the decoder, resulting in defect leakage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 325, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 420 + ], + "type": "text", + "content": "Our method aims to restore anomaly-free features from anomaly features instead of recognizing the reconstruction quality gap. For more effective restoration, we filter anomalies via bottleneck and recover the normal pattern and details through template-guided compensation. Note that our method uses the image-level pre-trained features of normal samples as templates to provide both global normal pattern priors and detailed normal feature priors for restoration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 421, + 545, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 600 + ], + "type": "text", + "content": "Distillation-based methods have recently been successfully applied to anomaly detection, where anomalies are identified based on feature differences between a teacher network pre-trained on a large dataset and a student network trained only on normal samples using knowledge distillation [7, 5, 33, 3, 11, 40, 31]. Additionally, some researchers have attempted to convert unsupervised anomaly detection into a supervised learning task by augmenting normal samples with pseudo-anomalies [24, 21, 46, 35]. However, these approaches are prone to bias towards pseudo outliers and fail to detect a large variety of anomaly types. To improve the embedding feature quality, Zou et al. [52] propose SPD which leverage pseudo anomalies as negative sample for contrastive self-supervised pre-training instead of supervised pre-trained with classification." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 609, + 361, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 361, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 361, + 621 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "The overview architecture of our proposed method is shown in Figure 3. Our method first obtains the multi-level embedding features of a given image using a backbone pretrained on ImageNet [12] as feature extractor. In the training stage, these embedding features are treated as the targets of feature restoration. In order to filter out the anomalies of various sizes, the embedding features are compressed by" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6449" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 73, + 399, + 314 + ], + "blocks": [ + { + "bbox": [ + 47, + 73, + 399, + 314 + ], + "lines": [ + { + "bbox": [ + 47, + 73, + 399, + 314 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 399, + 314 + ], + "type": "image", + "image_path": "bd4eca6fa9a4868d597b07f46d9b23c9b1191832f941413896b0997a2520f31e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 321, + 545, + 395 + ], + "lines": [ + { + "bbox": [ + 45, + 321, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 45, + 321, + 545, + 395 + ], + "type": "text", + "content": "Figure 3: Overview architecture of proposed Template-guided Hierarchical Feature Restoration (THFR) framework. (a) THFR consists of one pre-trained backbone encoder, bottleneck, and a restoration decoder. Given a test image, THFR uses ImageNet pre-trained backbone encoder to extract features for restoration target. THFR compresses input features through bottleneck, and the compressed features are compensated by template features retrieved from the template bank using image-level nearest neighbor search (image NNS). Finally, anomaly detection is performed in terms of the cosine distance between restored anomaly-free features and input features. (b) Detailed network design of global bottleneck and local bottleneck." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 433, + 81, + 547, + 296 + ], + "blocks": [ + { + "bbox": [ + 433, + 81, + 547, + 296 + ], + "lines": [ + { + "bbox": [ + 433, + 81, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 433, + 81, + 547, + 296 + ], + "type": "image", + "image_path": "246e2e40bfa543104735597004ed9e8b3c9d29d900e3b6402ca56b5cd7f8c506.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 301, + 520, + 312 + ], + "lines": [ + { + "bbox": [ + 459, + 301, + 520, + 312 + ], + "spans": [ + { + "bbox": [ + 459, + 301, + 520, + 312 + ], + "type": "text", + "content": "(b) G-L bottleneck" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 415, + 264, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 264, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 264, + 426 + ], + "type": "text", + "content": "global bottleneck (GBN) and local bottleneck (LBN)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 429, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 525 + ], + "type": "text", + "content": "The compression along spatial and channel dimensions inevitably causes information loss on features. To recover the normal pattern with compensation for the lost details, we establish a template bank to record image-level normal embedding features and retrieve the most similar normal feature from the template bank through an image-level nearest neighbor search module called image NNS, and treat the retrieved result as template to guide feature restoration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 537, + 119, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 119, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 119, + 550 + ], + "type": "text", + "content": "3.1. Bottleneck" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "We design global bottleneck and local bottleneck that hierarchically compress the embedding features to retain the crucial normal features at multiple levels, and thereby filter out anomaly features caused by anomalies of various sizes. The previous reconstruction-based works [13, 16, 11] use a fixed-scale loose bottleneck, which works on small-scale anomalies, but underperforms when anomalies are of large sizes. In our network, uncompressed template embedding features could provide rich hierarchical normal features while compensating the compressed features, so we could employ the bottleneck of different tightness to effectively filter anomaly features and obtain compact normal representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "text", + "content": "For a given image, its multi-level pyramid features " + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "inline_equation", + "content": "I_{k} \\in \\mathbb{R}^{H_{k} \\times W_{k} \\times C_{k}}" + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "text", + "content": " are extracted by pre-trained ResNet [15] and fed into bottleneck structure for hierarchical compression, where " + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "inline_equation", + "content": "H_{k} \\times W_{k}" + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "text", + "content": " denotes the spatial dimension, " + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "inline_equation", + "content": "C_{k}" + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "text", + "content": " is the number of channels and " + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "text", + "content": " indicates the layer index with maximum value equal to " + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "text", + "content": ". The input feature maps are distinct between global bottleneck and local bottleneck. Specifically, the high-level feature " + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "inline_equation", + "content": "I_{K}" + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "text", + "content": " is fed to global bottleneck, while " + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "inline_equation", + "content": "[I_{1}, \\dots, I_{K}]" + }, + { + "bbox": [ + 304, + 415, + 547, + 523 + ], + "type": "text", + "content": " for local bottleneck." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "Global bottleneck Considerable large-size anomalies result from missing or misplaced normal components [3], which appear to be globally abnormal but locally normal, creating difficulties for existing anomaly detection methods. We propose global bottleneck to preserve the most representative normal semantic features and filter out the anomaly semantic features. As shown in Figure 3 (b), global bottleneck reforms the deep-level feature " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "I_{K}" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " by a ResBlock and then compress the features by global maxpooling, producing a feature vector " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "Z_{GBN} \\in \\mathbb{R}^{1 \\times 1 \\times 2C_K}" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": ". Global max-pooling extremely compresses the feature spatially to eliminate the anomaly features. It could alleviate the attention on object misalignment and instead focuses on global pattern representation. Finally, we use " + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": " deconvolution [48]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6450" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "with stride 2 to upsample the compressed feature " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "Z_{GBN}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " to a " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " spatial feature map for succeeding feature restoration." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 111, + 286, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 111, + 286, + 245 + ], + "spans": [ + { + "bbox": [ + 47, + 111, + 286, + 245 + ], + "type": "text", + "content": "Local bottleneck Besides the compression on semantic features, we introduce a trainable local bottleneck to compress the multi-level features to preserve the local normal features and filter out the local anomaly features. The design of the local bottleneck follows the work [11], as shown in Figure 3 (b). To align multi-level features, the shallow features are downsampled through " + }, + { + "bbox": [ + 47, + 111, + 286, + 245 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 47, + 111, + 286, + 245 + ], + "type": "text", + "content": " convolution layers with stride of 2. We concatenate multi-level representations and fuse them by a ResBlock to obtain compressed feature " + }, + { + "bbox": [ + 47, + 111, + 286, + 245 + ], + "type": "inline_equation", + "content": "Z_{LBN} \\in \\mathbb{R}^{\\frac{H_K}{2} \\times \\frac{W_K}{2} \\times 2C_K}" + }, + { + "bbox": [ + 47, + 111, + 286, + 245 + ], + "type": "text", + "content": ", which is fed into detail compensation module for fine-grained feature restoration." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 253, + 216, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 216, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 216, + 266 + ], + "type": "text", + "content": "3.2. Template-guided compensation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 271, + 286, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 286, + 414 + ], + "type": "text", + "content": "Although the essential normal features are kept during compression, the information loss on normal features is unavoidable and needs to be compensated. We leverage multilevel normal embedding features as template to guide feature restoration. Since there is no one-to-one correspondence between the positional information of input and template features, we leverage the relationship between the input and template features to guide the restoration process. We fuse the input and template by stacking these two embeddings and the relation feature, and then passing them through multiple " + }, + { + "bbox": [ + 46, + 271, + 286, + 414 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 271, + 286, + 414 + ], + "type": "text", + "content": " convolution layers. The fused feature is fed to a series of ResBlocks for further restoration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 430, + 286, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 286, + 514 + ], + "type": "text", + "content": "Template bank We construct template bank using image-level pre-trained features extracted from " + }, + { + "bbox": [ + 47, + 430, + 286, + 514 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 430, + 286, + 514 + ], + "type": "text", + "content": " normal samples in the all train set as " + }, + { + "bbox": [ + 47, + 430, + 286, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{B_1^1,\\dots ,B_k^i,\\dots ,B_K^N\\}" + }, + { + "bbox": [ + 47, + 430, + 286, + 514 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 430, + 286, + 514 + ], + "type": "inline_equation", + "content": "B_{k}^{i}" + }, + { + "bbox": [ + 47, + 430, + 286, + 514 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 47, + 430, + 286, + 514 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 430, + 286, + 514 + ], + "type": "text", + "content": "-th layer template feature of the " + }, + { + "bbox": [ + 47, + 430, + 286, + 514 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 430, + 286, + 514 + ], + "type": "text", + "content": "-th normal sample in template bank. During inference, We could reduce template bank using coreset subsampling method [1, 8] to reduce inference time and memory usage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 529, + 286, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 286, + 624 + ], + "type": "text", + "content": "Image-level nearest neighbor search Given an input sample, we take the deepest pre-trained " + }, + { + "bbox": [ + 47, + 529, + 286, + 624 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 529, + 286, + 624 + ], + "type": "text", + "content": "-th layer feature " + }, + { + "bbox": [ + 47, + 529, + 286, + 624 + ], + "type": "inline_equation", + "content": "I_K" + }, + { + "bbox": [ + 47, + 529, + 286, + 624 + ], + "type": "text", + "content": " as the query to retrieve its correlated template. Image NNS obtains the template with index " + }, + { + "bbox": [ + 47, + 529, + 286, + 624 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 529, + 286, + 624 + ], + "type": "text", + "content": " by randomly selecting from the template candidates which are " + }, + { + "bbox": [ + 47, + 529, + 286, + 624 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 529, + 286, + 624 + ], + "type": "text", + "content": " most similar templates to increase the robustness during the training process. The template selection process could be formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 632, + 286, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 632, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 59, + 632, + 286, + 656 + ], + "type": "interline_equation", + "content": "t = \\operatorname {r a n d o m} \\left(\\underset {\\mathcal {S} \\subset \\{1, \\dots , N \\}, | \\mathcal {S} | = n} {\\arg \\min }\\right. \\sum_ {i \\in \\mathcal {S}} d \\left(I _ {K}, B _ {K} ^ {i}\\right)), \\tag {1}", + "image_path": "7ff09eb94deed5b11bcfdd6d82fd1b32c04ef0f2777784c2f017ddb107ac4e49.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": " denotes the image-level distance between input feature query " + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "I_{K}" + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": " and template feature key " + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "B_{K}^{i}" + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": " by flattening them to vectors to compute cosine distance. " + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": " is a subset of " + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\{1,\\dots ,N\\}" + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": " denotes the indexes of " + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 665, + 286, + 713 + ], + "type": "text", + "content": " template candidates," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 314, + 71, + 539, + 297 + ], + "blocks": [ + { + "bbox": [ + 314, + 71, + 539, + 297 + ], + "lines": [ + { + "bbox": [ + 314, + 71, + 539, + 297 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 539, + 297 + ], + "type": "image", + "image_path": "2938306d15959934a1dd5be2fa74b42f75b4bfe64c0fb8c4912e55e9d52b2276.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 304, + 545, + 353 + ], + "lines": [ + { + "bbox": [ + 305, + 304, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 545, + 353 + ], + "type": "text", + "content": "Figure 4: Template-guided compensation process. We leverage normal template feature to guide anomaly-free feature restoration with the relation representation between the input feature and the template feature." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 375, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 375, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 305, + 375, + 545, + 471 + ], + "type": "text", + "content": "which are the top-n nearest neighbors of the input feature " + }, + { + "bbox": [ + 305, + 375, + 545, + 471 + ], + "type": "inline_equation", + "content": "I_K" + }, + { + "bbox": [ + 305, + 375, + 545, + 471 + ], + "type": "text", + "content": ". Note that, the " + }, + { + "bbox": [ + 305, + 375, + 545, + 471 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 375, + 545, + 471 + ], + "type": "text", + "content": " set to 1 during inference. Compared with the traditional point-by-point search strategy, image NNS not only ensures that the reference is completely normal but also improves search efficiency. The multi-level embedding features of the template sample are used as references in the corresponding feature level, which are denoted as " + }, + { + "bbox": [ + 305, + 375, + 545, + 471 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 375, + 545, + 471 + ], + "type": "text", + "content": " for simplicity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 488, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 488, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 305, + 488, + 545, + 571 + ], + "type": "text", + "content": "Relation representation We use pairwise relation [49] between input feature and template feature as additional information to restore normal feature. Specifically, we use " + }, + { + "bbox": [ + 305, + 488, + 545, + 571 + ], + "type": "inline_equation", + "content": "r_{I^i \\to T^j}" + }, + { + "bbox": [ + 305, + 488, + 545, + 571 + ], + "type": "text", + "content": " to represent the relation from the " + }, + { + "bbox": [ + 305, + 488, + 545, + 571 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 488, + 545, + 571 + ], + "type": "text", + "content": "-th feature point of input feature to the " + }, + { + "bbox": [ + 305, + 488, + 545, + 571 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 488, + 545, + 571 + ], + "type": "text", + "content": "-th feature point of the template feature. The pairwise relation can be defined as a dot-product affinity in the embedding spaces as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 376, + 581, + 545, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 581, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 376, + 581, + 545, + 596 + ], + "type": "interline_equation", + "content": "r _ {I ^ {i} \\rightarrow T ^ {j}} = \\theta \\left(I ^ {i}\\right) ^ {\\top} \\phi \\left(T ^ {j}\\right) \\tag {2}", + "image_path": "c39f18fb60dc8eb479bb7bbeace3c1d9c3e573d08f4e35cd94e5751e19525961.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "text", + "content": " are two embedding functions implemented by a " + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "text", + "content": " spatial convolutional layer. Similarly, we can get the affinity from " + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "text", + "content": "-th feature point of template feature to " + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "text", + "content": "-th feature point of input feature as " + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "r_{T^j\\rightarrow I^i}" + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "text", + "content": ". We use the pair " + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "(r_{I^i\\to T^j},r_{T^j\\to I^i})" + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "text", + "content": " to describe the bi-directional relation between " + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "text", + "content": "-th feature point of input feature to " + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "text", + "content": "-th feature point of template feature. Then, we represent the relation matrix " + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{R}\\in \\mathbb{R}^{m\\times m}" + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "text", + "content": " among all the nodes, where " + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 605, + 545, + 712 + ], + "type": "text", + "content": " is the resolution of the feature. The bi-directional" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6451" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": "pairwise relation vector of " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": "-th feature point is " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i = [r_{I^i \\to T\\{1, \\dots, m\\}}, r_{T^i \\to I\\{1, \\dots, m\\}}]" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": ". For example, as shown in Figure 4, the third row and the third column of the relation matrix, i.e. " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_3 = [r_{I^3 \\to T\\{1, \\dots, 9\\}}, r_{T^3 \\to I\\{1, \\dots, 9\\}}]" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": ". In this relation vector, the first half channels indicate the relation that regards the input feature point as query and template feature point as key, while the second half indicates the relation of reverse operation. The bi-directional pairwise relation vectors of all nodes form the relation feature." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": "Feature compensation We leverage normal template features to guide anomaly-free feature restoration. As shown in Figure 4, the input compressed feature with channel size " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " is projected to input embedding " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "E_{i} \\in \\mathbb{R}^{H \\times W \\times C_{i}}" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": ", and the template feature with channel size " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " is projected to template embedding " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "E_{t} \\in \\mathbb{R}^{H \\times W \\times C_{t}}" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "C_{i}" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " represents the channel size of input embedding, and " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "C_{t}" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " represents the channel size of template embedding, while " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " represent the height and width of the feature maps, respectively. The values of " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "C_{i}" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "C_{t}" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " can be adjusted with respect to the feature compression method used. For global bottleneck compression, " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "C_{i}" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " is set to the original channel size " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " to provide detailed information, while " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "C_{t}" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " is set to 1 to generate normal layout guidance for high-level semantic feature compensation. For local bottleneck compression, " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "C_{i}" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " is set to 1 to provide spatial layout guidance, and " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "C_{t}" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " to supply detail feature for low-level feature compensation. Then, we concatenate the input embedding " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": ", the template embedding " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " and the relation feature, and passing them through multiple " + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 47, + 194, + 289, + 470 + ], + "type": "text", + "content": " convolution layers to get the restored feature. In order to explicitly demonstrate the effectiveness of our method, we visualize the restored feature using a visualization network [43] in experiment." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 475, + 92, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 92, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 92, + 487 + ], + "type": "text", + "content": "3.3. Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 494, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 287, + 577 + ], + "type": "text", + "content": "We use the cosine distance between the pre-trained feature and restored feature as restoration loss, as it is capable to evaluate the feature differences at multi-levels regardless of feature dimension differences among each level. Mathematically, for input feature " + }, + { + "bbox": [ + 47, + 494, + 287, + 577 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 494, + 287, + 577 + ], + "type": "text", + "content": " and restored feature " + }, + { + "bbox": [ + 47, + 494, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 47, + 494, + 287, + 577 + ], + "type": "text", + "content": ", we calculate their feature cosine distance along the channel axis as restoration loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 582, + 287, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 582, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 123, + 582, + 287, + 612 + ], + "type": "interline_equation", + "content": "L _ {k} = 1 - \\frac {I _ {k} ^ {\\top} \\cdot \\hat {I} _ {k}}{\\| I _ {k} \\| \\| \\hat {I} _ {k} \\|}, \\tag {3}", + "image_path": "21832557f1cbfc4c30fd799d6bd3fce63519ad2d398629062407d12dd62ae653.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 616, + 254, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 254, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 254, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 616, + 254, + 628 + ], + "type": "inline_equation", + "content": "L_{k}" + }, + { + "bbox": [ + 47, + 616, + 254, + 628 + ], + "type": "text", + "content": " denotes the restoration loss of " + }, + { + "bbox": [ + 47, + 616, + 254, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 616, + 254, + 628 + ], + "type": "text", + "content": "-th level." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 635, + 137, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 137, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 137, + 647 + ], + "type": "text", + "content": "3.4. Anomaly map" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": "In the testing phase, we detect anomalies in terms of the cosine distance between the pre-trained feature of an inference image and the corresponding restored anomaly-free feature. The restored feature tends to be close to input image feature for the normal regions, and departs on" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "anomalies. Mathematically, we calculate pixel-wise cosine distance to obtain a 2D anomaly map in " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "-th level:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 341, + 106, + 545, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 106, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 341, + 106, + 545, + 135 + ], + "type": "interline_equation", + "content": "M _ {k} (h, w) = 1 - \\frac {\\left(I _ {k} (h , w)\\right) ^ {\\top} \\cdot \\hat {I} _ {k} (h , w)}{\\| I _ {k} (h , w) \\| \\| \\hat {I} _ {k} (h , w) \\|}, \\tag {4}", + "image_path": "3adb50495b8bbe157c1476a435ea0119519563d0f378212939932bad23ab4fa6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 143, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 143, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 143, + 547, + 168 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 143, + 547, + 168 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 306, + 143, + 547, + 168 + ], + "type": "text", + "content": " denotes the position of the feature vector, and " + }, + { + "bbox": [ + 306, + 143, + 547, + 168 + ], + "type": "inline_equation", + "content": "M_{k}\\in \\mathbb{R}^{H_{k}\\times W_{k}}" + }, + { + "bbox": [ + 306, + 143, + 547, + 168 + ], + "type": "text", + "content": ". The final anomaly map is calculated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 386, + 177, + 545, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 177, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 386, + 177, + 545, + 209 + ], + "type": "interline_equation", + "content": "M = \\sum_ {k = 1} ^ {K} U _ {k} \\left(M _ {k}\\right), \\tag {5}", + "image_path": "b2612c89ca9a797cef9453e6b08498d93bc0011cf4ff9bb68806f03e89f95f88.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 214, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 214, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 214, + 545, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 214, + 545, + 251 + ], + "type": "inline_equation", + "content": "U_{k}(\\cdot)" + }, + { + "bbox": [ + 305, + 214, + 545, + 251 + ], + "type": "text", + "content": " denotes the upsampling operation of " + }, + { + "bbox": [ + 305, + 214, + 545, + 251 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 214, + 545, + 251 + ], + "type": "text", + "content": "-th anomaly map. In order to remove the noises in the score map, we smooth anomaly map " + }, + { + "bbox": [ + 305, + 214, + 545, + 251 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 214, + 545, + 251 + ], + "type": "text", + "content": " by a Gaussian filter." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 262, + 388, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 388, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 388, + 275 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 281, + 421, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 421, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 421, + 295 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 300, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 456 + ], + "type": "text", + "content": "Datasets We evaluate the proposed method on two industrial anomaly detection datasets: MVTec LOCO AD [3] and MVTec AD [4]. MVTec LOCO AD consists of 5 real-world sub-datasets for anomaly detection, including 1772 images for training, 304 for validation and 1568 for testing. The normal samples in MVTec LOCO AD often have complicated patterns, and the anomaly samples appear in form of logical or structural anomalies [3]. MVTec AD dataset consists of 15 real-world sub-datasets, with 5 categories of textures and 10 categories of objects. It contains 3629 training images and 1725 test images. Note that all the training images of the two datasets are normal, while some of test images are normal and others are anomalies." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "text", + "content": "Implementation details In our experiments, anomaly detection and localization are performed on one category at a time. We resize input images to " + }, + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "text", + "content": " for data preprocessing and use pre-trained WideResNet50 [45] on ImageNet as feature extractor, with the deepest feature level " + }, + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "text", + "content": " set to 3. To train our THFR network, we utilize Adam optimizer [18] and set the learning rate to 0.005. Each model is trained for 200 epochs with a batch size of 16 on NVIDIA Tesla V100 GPU. We use Gaussian filter with " + }, + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\sigma = 2" + }, + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "text", + "content": " to smooth anomaly map on MVtec LOCO AD, while " + }, + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\sigma = 4" + }, + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "text", + "content": " on MVtec AD. For training sample pair selection, we use " + }, + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "inline_equation", + "content": "n = 1" + }, + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "text", + "content": " for all categories in MVtec LOCO AD and " + }, + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "inline_equation", + "content": "n = 6" + }, + { + "bbox": [ + 305, + 471, + 545, + 626 + ], + "type": "text", + "content": " for all categories in MVtec AD." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "text", + "content": "Evaluation metrics To fairly compare with other methods, we follow the common evaluation metrics of each dataset. For MVTec AD [4], we take area under the receiver operating characteristic (AUROC) as the image-level and pixel-level evaluation metric, and per-region-overlap (PRO) for pixel-level evaluation metric, which can better evaluate" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6452" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 120, + 115, + 475, + 216 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "type": "text", + "content": "Table 1: Pixel-level anomaly localization accuracy on MVtec LOCO AD dataset (sPRO) [3]. Remarkably, our approach achieves state-of-the-art performance in four categories and the mean of all categories. Best and second-best scores are bolded and underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 120, + 115, + 475, + 216 + ], + "lines": [ + { + "bbox": [ + 120, + 115, + 475, + 216 + ], + "spans": [ + { + "bbox": [ + 120, + 115, + 475, + 216 + ], + "type": "table", + "html": "
MethodBreakfastScrew BagPushpinsConnectorsJuice BottleMean
S-T [5]49.660.252.369.881.162.6
DR/EM [46]49.949.049.367.380.059.1
CFLOW [14]48.660.255.171.683.663.8
RD4AD [11]42.257.461.471.385.163.5
PatchCore [28]51.659.853.575.783.264.7
GCAD [3]50.255.873.979.891.070.1
THFR (ours)58.361.576.384.889.674.1
", + "image_path": "02d2838b8c536fcf8482e84a69c7f60e9c4bfd585f02464e3fe426da10a3e3b5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 120, + 270, + 474, + 371 + ], + "blocks": [ + { + "bbox": [ + 46, + 225, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 547, + 262 + ], + "type": "text", + "content": "Table 2: Image-level anomaly detection accuracy on MVtec LOCO AD dataset (AUROC) [3]. Remarkably, our approach achieves the state-of-the-art performance on four categories and the mean of all categories. Best and second-best scores are bolded and underlined." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 120, + 270, + 474, + 371 + ], + "lines": [ + { + "bbox": [ + 120, + 270, + 474, + 371 + ], + "spans": [ + { + "bbox": [ + 120, + 270, + 474, + 371 + ], + "type": "table", + "html": "
MethodBreakfastScrew BagPushpinsConnectorsJuice BottleMean
S-T [5]-----77.3
DR/EM [46]75.772.776.082.593.980.1
CFLOW [14]77.473.076.082.695.380.8
RD4AD [11]68.774.975.984.494.879.7
PatchCore [28]77.173.374.186.094.681.0
GCAD [3]-----83.3
THFR (ours)78.073.788.392.797.186.0
", + "image_path": "f8ecda353d3a6ae27d571e0e64704edd30a0a1d7c8a39e5a647ba5f2073ad94c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 62, + 469, + 273, + 582 + ], + "blocks": [ + { + "bbox": [ + 46, + 388, + 288, + 462 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 288, + 462 + ], + "type": "text", + "content": "Table 3: Anomaly detection and localization results on MVTec AD dataset [4], including image-level and pixel-level AUROC, and pixel-level PRO. Best and second-best scores are bolded and underlined. * To ensure fair comparison with previous studies, CFLOW [14] is evaluated using input images with resolution of 256x256 pixels." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 469, + 273, + 582 + ], + "lines": [ + { + "bbox": [ + 62, + 469, + 273, + 582 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 273, + 582 + ], + "type": "table", + "html": "
MethodImagePixelPixel(PRO)
SPADE [9]85.596.091.7
PaDiM [10]95.397.592.1
S-T [5]--91.4
DR/EM [46]98.097.3-
CFLOW * [14]96.897.992.7
RD4AD [11]98.597.893.9
PatchCore [28]99.198.193.4
THFR (ours)99.298.295.0
", + "image_path": "f7254c7548d64413cb1c8dde10cea07a7cd87cabf7c46b0704fcf17bf0549505.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "the performance of small size anomalies through weighting ground-truth regions of different size equally. Following the protocol mentioned in [5], we evaluate the PRO value for a large number of increasing thresholds until an average per-pixel false-positive rate of " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " for the entire dataset is reached. For MVTec LOCO AD [3], we use AUROC to measure image-level performance, and use saturated-per-region-overlaps (sPRO) [3] with per-pixel false-positive rate of " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " to evaluate pixel-level performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 389, + 388, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 389, + 388, + 402 + ], + "spans": [ + { + "bbox": [ + 306, + 389, + 388, + 402 + ], + "type": "text", + "content": "4.2. Main results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 408, + 547, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 408, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 408, + 547, + 469 + ], + "type": "text", + "content": "We compare the proposed model with several deep learning-based methods for anomaly detection as baselines, including embedding-based methods [9, 10, 14, 28], reconstruction-based methods [11, 46], distillation-based methods [5, 11, 3], and data-augmented methods [46]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 483, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 547, + 602 + ], + "type": "text", + "content": "MVTec LOCO AD Anomaly detection results on MVTec LOCO AD [3] are presented in Table 1 and Table 2. In comparison with the SOTA method [3], our approach obtain an absolute AUROC gain of " + }, + { + "bbox": [ + 305, + 483, + 547, + 602 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 305, + 483, + 547, + 602 + ], + "type": "text", + "content": " for the image-level detection, and an remarkable absolute sPRO gain of " + }, + { + "bbox": [ + 305, + 483, + 547, + 602 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 305, + 483, + 547, + 602 + ], + "type": "text", + "content": " for the pixel-level localization. Our approach achieves new state-of-the-art performance on both detection and localization results, and surpasses other methods by a large margin. More detailed results will be provided in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 547, + 713 + ], + "type": "text", + "content": "MVTec AD Quantitative comparison results on MVTec AD [4] between baselines and our approach are summarized in Table 3. For overall categories, our method produces comparable results with the other advanced methods. Our approach excels in achieving high accuracy on both image-level detection and pixel-level localization at the same time. More detailed results will be presented in the supplementary materials." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "6453" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 102, + 170 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 102, + 170 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 102, + 170 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 102, + 170 + ], + "type": "image", + "image_path": "c6788f70a6555af00e53aaf5356ef19037be6aee330666c6d2979f16778fa657.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 171, + 87, + 178 + ], + "lines": [ + { + "bbox": [ + 64, + 171, + 87, + 178 + ], + "spans": [ + { + "bbox": [ + 64, + 171, + 87, + 178 + ], + "type": "text", + "content": "Normal" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 104, + 70, + 157, + 170 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 157, + 170 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 157, + 170 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 157, + 170 + ], + "type": "image", + "image_path": "1fc883d1f989afc10215983b9779d4b486ae0c83d6f5d90bd4ac1349b3de38ac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 171, + 145, + 179 + ], + "lines": [ + { + "bbox": [ + 118, + 171, + 145, + 179 + ], + "spans": [ + { + "bbox": [ + 118, + 171, + 145, + 179 + ], + "type": "text", + "content": "Anomaly" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 159, + 70, + 214, + 170 + ], + "blocks": [ + { + "bbox": [ + 159, + 70, + 214, + 170 + ], + "lines": [ + { + "bbox": [ + 159, + 70, + 214, + 170 + ], + "spans": [ + { + "bbox": [ + 159, + 70, + 214, + 170 + ], + "type": "image", + "image_path": "773b7c428cf547c1ee7372aae07448f5ffdb1abe0661d8690b8c7e39ac02ed68.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 171, + 196, + 178 + ], + "lines": [ + { + "bbox": [ + 176, + 171, + 196, + 178 + ], + "spans": [ + { + "bbox": [ + 176, + 171, + 196, + 178 + ], + "type": "text", + "content": "Recon" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 215, + 70, + 268, + 170 + ], + "blocks": [ + { + "bbox": [ + 215, + 70, + 268, + 170 + ], + "lines": [ + { + "bbox": [ + 215, + 70, + 268, + 170 + ], + "spans": [ + { + "bbox": [ + 215, + 70, + 268, + 170 + ], + "type": "image", + "image_path": "d95ad2598904e39c77050c4b0d06e3782152a292bb38cf3918b4371a2ee1fb7b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 171, + 250, + 178 + ], + "lines": [ + { + "bbox": [ + 234, + 171, + 250, + 178 + ], + "spans": [ + { + "bbox": [ + 234, + 171, + 250, + 178 + ], + "type": "text", + "content": "GBN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 188, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 547, + 248 + ], + "type": "text", + "content": "Figure 5: Visualization of restored features from different restoration networks. From left to right: normal image, anomaly image, reconstructed images from the pre-trained embedding features (Recon), restored features of restoration network only with GBN, only with LBN, with G-L bottleneck, and with our template-guided hierarchical feature restoration framework, anomaly map, and ground truth. The visualization results show how restoration networks progressively achieve anomaly-free restoration with G-L bottleneck and hierarchical feature compensation." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 269, + 70, + 323, + 170 + ], + "blocks": [ + { + "bbox": [ + 269, + 70, + 323, + 170 + ], + "lines": [ + { + "bbox": [ + 269, + 70, + 323, + 170 + ], + "spans": [ + { + "bbox": [ + 269, + 70, + 323, + 170 + ], + "type": "image", + "image_path": "855c6f8ba47840794975389d75c06801867ea1761992514e9ecbf4ffca75d9ce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 171, + 304, + 178 + ], + "lines": [ + { + "bbox": [ + 288, + 171, + 304, + 178 + ], + "spans": [ + { + "bbox": [ + 288, + 171, + 304, + 178 + ], + "type": "text", + "content": "LBN" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 324, + 70, + 379, + 170 + ], + "blocks": [ + { + "bbox": [ + 324, + 70, + 379, + 170 + ], + "lines": [ + { + "bbox": [ + 324, + 70, + 379, + 170 + ], + "spans": [ + { + "bbox": [ + 324, + 70, + 379, + 170 + ], + "type": "image", + "image_path": "74643412a8c200c4fa0f76a4629944bcbb67f4f1390a42324a0d3e476b479fcd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 330, + 171, + 374, + 178 + ], + "lines": [ + { + "bbox": [ + 330, + 171, + 374, + 178 + ], + "spans": [ + { + "bbox": [ + 330, + 171, + 374, + 178 + ], + "type": "text", + "content": "G-L bottleneck" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 380, + 70, + 434, + 170 + ], + "blocks": [ + { + "bbox": [ + 380, + 70, + 434, + 170 + ], + "lines": [ + { + "bbox": [ + 380, + 70, + 434, + 170 + ], + "spans": [ + { + "bbox": [ + 380, + 70, + 434, + 170 + ], + "type": "image", + "image_path": "ec141cc2a6a517b9e3c999fcbec9de14bbe10b77aca8f6adfc474f1967180905.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 171, + 417, + 178 + ], + "lines": [ + { + "bbox": [ + 397, + 171, + 417, + 178 + ], + "spans": [ + { + "bbox": [ + 397, + 171, + 417, + 178 + ], + "type": "text", + "content": "THFR" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 435, + 70, + 489, + 170 + ], + "blocks": [ + { + "bbox": [ + 435, + 70, + 489, + 170 + ], + "lines": [ + { + "bbox": [ + 435, + 70, + 489, + 170 + ], + "spans": [ + { + "bbox": [ + 435, + 70, + 489, + 170 + ], + "type": "image", + "image_path": "eeb24bdd9f397180733332fd91b550dd895f5fddaf6ec045b76cf737d080bd53.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 171, + 483, + 179 + ], + "lines": [ + { + "bbox": [ + 442, + 171, + 483, + 179 + ], + "spans": [ + { + "bbox": [ + 442, + 171, + 483, + 179 + ], + "type": "text", + "content": "Anomaly map" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 491, + 70, + 545, + 170 + ], + "blocks": [ + { + "bbox": [ + 491, + 70, + 545, + 170 + ], + "lines": [ + { + "bbox": [ + 491, + 70, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 491, + 70, + 545, + 170 + ], + "type": "image", + "image_path": "7ec064cf12a89045a325b84b3da3e29b8a081c6906d92320f3dc211c6bf47510.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 171, + 523, + 178 + ], + "lines": [ + { + "bbox": [ + 512, + 171, + 523, + 178 + ], + "spans": [ + { + "bbox": [ + 512, + 171, + 523, + 178 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 269, + 138, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 138, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 138, + 281 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 288, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 287, + 348 + ], + "type": "text", + "content": "To explicitly demonstrate the effectiveness of our method, we visualize the restored features using a visualization network. We investigate the effectiveness of bottleneck design and template-guided compensation, and study the influence of subsampling within the template bank." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 363, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 590 + ], + "type": "text", + "content": "Visualization of restored features As shown in Figure 5, we visually compared the pre-trained feature with the restored feature acquired from the network utilizing a visualization network [43]. We observe that the restoration networks employing solely local bottleneck (LBN) cannot effectively filter out anomalies that violate complex normal rules, such as missing connector heads and absent labels on the juice bottle. On the other hand, restoration networks with only global bottleneck (GBN) can filter out rule-breaking anomalies, but it encounters challenges in accurately restoring intricate details, such as the recovery of oatmeal and banana slices in the breakfast box. The restoration network with G-L bottleneck also has anomaly feature leakage after bottleneck compression, manifesting as anomalies like the irregular coloration of the juice. In contrast, our THFR network recovers anomaly-free feature through both semantic level correction and detailed feature compensation, resulting in an effective restoration to normal of various types of anomaly areas." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Impact of bottleneck Table 4 shows anomaly detection performance of local bottleneck and global bottleneck on MVTec LOCO AD dataset. The dataset has a rich diversity of data and anomalies, and the combination of local bottleneck and global bottleneck yields substantial enhancements in detection and localization results. Figure 6 illustrates that the global bottleneck primarily concentrates on filtering large-size anomalies, while the local bottleneck excels in filtering small-size anomalies." + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 307, + 267, + 353, + 316 + ], + "blocks": [ + { + "bbox": [ + 307, + 267, + 353, + 316 + ], + "lines": [ + { + "bbox": [ + 307, + 267, + 353, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 267, + 353, + 316 + ], + "type": "image", + "image_path": "0f0fe87b349db9bd226af66d5f2c28ced2cebf0a14d49da7becb6ad2d85325f5.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 316, + 342, + 324 + ], + "lines": [ + { + "bbox": [ + 319, + 316, + 342, + 324 + ], + "spans": [ + { + "bbox": [ + 319, + 316, + 342, + 324 + ], + "type": "text", + "content": "Normal" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 354, + 267, + 399, + 315 + ], + "blocks": [ + { + "bbox": [ + 354, + 267, + 399, + 315 + ], + "lines": [ + { + "bbox": [ + 354, + 267, + 399, + 315 + ], + "spans": [ + { + "bbox": [ + 354, + 267, + 399, + 315 + ], + "type": "image", + "image_path": "ebfe71ef8140a0b7c97158f49a7ba37a532b4cbbe034b2b689c2ca18ae3048ea.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 316, + 391, + 324 + ], + "lines": [ + { + "bbox": [ + 364, + 316, + 391, + 324 + ], + "spans": [ + { + "bbox": [ + 364, + 316, + 391, + 324 + ], + "type": "text", + "content": "Anomaly" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 400, + 267, + 447, + 315 + ], + "blocks": [ + { + "bbox": [ + 400, + 267, + 447, + 315 + ], + "lines": [ + { + "bbox": [ + 400, + 267, + 447, + 315 + ], + "spans": [ + { + "bbox": [ + 400, + 267, + 447, + 315 + ], + "type": "image", + "image_path": "faf4945f46ba28fcde9da1765cf9df1eceeb1222058b752cd2ea06087a3998f1.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 316, + 431, + 323 + ], + "lines": [ + { + "bbox": [ + 416, + 316, + 431, + 323 + ], + "spans": [ + { + "bbox": [ + 416, + 316, + 431, + 323 + ], + "type": "text", + "content": "GBN" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 447, + 267, + 493, + 315 + ], + "blocks": [ + { + "bbox": [ + 447, + 267, + 493, + 315 + ], + "lines": [ + { + "bbox": [ + 447, + 267, + 493, + 315 + ], + "spans": [ + { + "bbox": [ + 447, + 267, + 493, + 315 + ], + "type": "image", + "image_path": "f02cc670a86be81c5ec6808f72f3f5ff7651d365ec4d5844b37b89db3a06cda3.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 316, + 477, + 323 + ], + "lines": [ + { + "bbox": [ + 462, + 316, + 477, + 323 + ], + "spans": [ + { + "bbox": [ + 462, + 316, + 477, + 323 + ], + "type": "text", + "content": "LBN" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 494, + 267, + 541, + 315 + ], + "blocks": [ + { + "bbox": [ + 494, + 267, + 541, + 315 + ], + "lines": [ + { + "bbox": [ + 494, + 267, + 541, + 315 + ], + "spans": [ + { + "bbox": [ + 494, + 267, + 541, + 315 + ], + "type": "image", + "image_path": "232d52900d0adab7f8ea03f067916c8e27e8c37e5f7d5c500b948036c0d296e6.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 316, + 522, + 323 + ], + "lines": [ + { + "bbox": [ + 511, + 316, + 522, + 323 + ], + "spans": [ + { + "bbox": [ + 511, + 316, + 522, + 323 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 307, + 363, + 353, + 448 + ], + "blocks": [ + { + "bbox": [ + 306, + 331, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 306, + 331, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 545, + 354 + ], + "type": "text", + "content": "Figure 6: Visualization of anomaly detection results with different bottlenecks." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 363, + 353, + 448 + ], + "lines": [ + { + "bbox": [ + 307, + 363, + 353, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 353, + 448 + ], + "type": "image", + "image_path": "74e7d20e35118fe25a4ab49b325aa4104fc38b50e274c183d7bc4b7ac9fac548.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 449, + 342, + 457 + ], + "lines": [ + { + "bbox": [ + 319, + 449, + 342, + 457 + ], + "spans": [ + { + "bbox": [ + 319, + 449, + 342, + 457 + ], + "type": "text", + "content": "Normal" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 465, + 545, + 513 + ], + "lines": [ + { + "bbox": [ + 306, + 465, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 545, + 513 + ], + "type": "text", + "content": "Figure 7: Visualization of anomaly detection results with or without template-guided compensation. More precise localization results are obtained with template-guided compensation." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 354, + 363, + 400, + 448 + ], + "blocks": [ + { + "bbox": [ + 354, + 363, + 400, + 448 + ], + "lines": [ + { + "bbox": [ + 354, + 363, + 400, + 448 + ], + "spans": [ + { + "bbox": [ + 354, + 363, + 400, + 448 + ], + "type": "image", + "image_path": "e656beda0a91fed597cce3ac0741a91e907961decfb5defbb47ef1c8b711d4a1.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 449, + 391, + 457 + ], + "lines": [ + { + "bbox": [ + 364, + 449, + 391, + 457 + ], + "spans": [ + { + "bbox": [ + 364, + 449, + 391, + 457 + ], + "type": "text", + "content": "Anomaly" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 400, + 363, + 447, + 448 + ], + "blocks": [ + { + "bbox": [ + 400, + 363, + 447, + 448 + ], + "lines": [ + { + "bbox": [ + 400, + 363, + 447, + 448 + ], + "spans": [ + { + "bbox": [ + 400, + 363, + 447, + 448 + ], + "type": "image", + "image_path": "9a029d4afb646fb8a6d2d22fc47d7367ab15a06cf7b2c00423afeadcba6f5cf6.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 449, + 440, + 457 + ], + "lines": [ + { + "bbox": [ + 406, + 449, + 440, + 457 + ], + "spans": [ + { + "bbox": [ + 406, + 449, + 440, + 457 + ], + "type": "text", + "content": "W/O comp." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 447, + 363, + 493, + 448 + ], + "blocks": [ + { + "bbox": [ + 447, + 363, + 493, + 448 + ], + "lines": [ + { + "bbox": [ + 447, + 363, + 493, + 448 + ], + "spans": [ + { + "bbox": [ + 447, + 363, + 493, + 448 + ], + "type": "image", + "image_path": "3511ca2481eea3975bdcd5dff2f760096af3872c6e0a12e3dfd6a7e406b1d462.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 449, + 484, + 457 + ], + "lines": [ + { + "bbox": [ + 456, + 449, + 484, + 457 + ], + "spans": [ + { + "bbox": [ + 456, + 449, + 484, + 457 + ], + "type": "text", + "content": "W/ comp." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 494, + 363, + 541, + 448 + ], + "blocks": [ + { + "bbox": [ + 494, + 363, + 541, + 448 + ], + "lines": [ + { + "bbox": [ + 494, + 363, + 541, + 448 + ], + "spans": [ + { + "bbox": [ + 494, + 363, + 541, + 448 + ], + "type": "image", + "image_path": "c5d22b4ca3a2ac4c78cfe3f563e4f065a174c8700993a1aa8e7eec4058ef61cc.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 449, + 522, + 457 + ], + "lines": [ + { + "bbox": [ + 511, + 449, + 522, + 457 + ], + "spans": [ + { + "bbox": [ + 511, + 449, + 522, + 457 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 531, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 675 + ], + "type": "text", + "content": "G-L bottleneck vs. different defect size To access the ability of different bottleneck designs to detect anomalies of different sizes, we categorize defect images of LOCO [3] into 6 groups based on their defect sizes, and compare the performance of the restoration networks with LBN only, GBN only, and a combination of the two on these 6 groups in terms of anomaly localization. As shown in Figure 8, the accuracy of anomaly localization varies with the size of defects. LBN performs well on small defects, while GBN exhibits superior performance on large defects. The best performance is achieved by using a combination of LBN and GBN across all groups." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Impact of template-guided compensation The effectiveness of our compensation design, with or without com" + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6454" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 284, + 189 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 284, + 189 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 284, + 189 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 284, + 189 + ], + "type": "image", + "image_path": "327c5000445a398255a8688ef4fb028a21738eddfc162e660b3910ddd1e97d8f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 199, + 287, + 259 + ], + "lines": [ + { + "bbox": [ + 46, + 199, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 287, + 259 + ], + "type": "text", + "content": "Figure 8: Performance comparison of the restoration networks with distinct bottleneck settings on varying defect sizes in terms of anomaly localization. We divide anomaly samples into 6 groups based on defect sizes, and show the percentage of each group with a bar chart." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "text", + "content": "pensation, is shown in Table 5 and Figure 7. Our compensation design improves detection result by " + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "text", + "content": " and localization result by " + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "text", + "content": " over the baseline on MVTec LOCO AD, and improves detection result by " + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "text", + "content": ", localization result by " + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "text", + "content": " and localization (PRO) result by " + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "text", + "content": " over the baseline on MVTec AD. By incorporating compensation design, it is possible to enhance the precision of localization and achieve more accurate coverage of abnormal areas. We also investigated the impact of relation design in our compensation module. The relation design improves detection result by " + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "text", + "content": " and localization result by " + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "text", + "content": " over the baseline on MVTec LOCO AD, and improves detection result by " + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "text", + "content": ", localization result by " + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "text", + "content": " and localization (PRO) result by " + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 46, + 279, + 287, + 447 + ], + "type": "text", + "content": " over the baseline on MVTec AD." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 581 + ], + "type": "text", + "content": "Impact of template bank subsampling The inference time and memory usage are highly relevant to the size of the template bank. We adopt coreset subsampling [1, 8] to find a subset that best describes the template bank during the testing phase. In Table 6, we compare performance of template banks with different subsampling ratios, and observe that subsampling has little impact on performance of MVTec AD. Although subsampling slightly reduces the mean detection accuracy of MVTec LOCO AD, its detection accuracy surpasses peers by a remarkable margin." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 587, + 163, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 163, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 163, + 601 + ], + "type": "text", + "content": "4.4. Complexity analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "We compare inference speed and memory usage with different methods in Table 7. The inference time is measured with Intel Xeon E5-2698v4. Because the backbones are identical (i.e. WideResNet50 [45]), we only report complexity on the additional models. We use image NNS and template bank subsampling to improve searching efficiency and reduce memory cost. Compared with peer works, our method with tempalte bank subsampling to " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": " obtain competitive performance on complexity and accuracy." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 103, + 543, + 144 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 545, + 94 + ], + "type": "text", + "content": "Table 4: The impact of bottleneck design on MVtec LOCO AD Dataset. Best scores are bolded." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 103, + 543, + 144 + ], + "lines": [ + { + "bbox": [ + 309, + 103, + 543, + 144 + ], + "spans": [ + { + "bbox": [ + 309, + 103, + 543, + 144 + ], + "type": "table", + "html": "
Metrics\\BottleneckGBNLBNG-L bottleneck
Image-level79.780.386.0
Pixel-level55.666.874.1
", + "image_path": "1e64db0a0022bd75a1588cddb2c8319073689232402385f83cf83ea8aa0d64e9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 307, + 189, + 544, + 253 + ], + "blocks": [ + { + "bbox": [ + 306, + 156, + 544, + 179 + ], + "lines": [ + { + "bbox": [ + 306, + 156, + 544, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 544, + 179 + ], + "type": "text", + "content": "Table 5: Ablation study on template compensation and relation representation. Best scores are bolded." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 189, + 544, + 253 + ], + "lines": [ + { + "bbox": [ + 307, + 189, + 544, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 189, + 544, + 253 + ], + "type": "table", + "html": "
DatasetLOCO [3]AD [4]
w/ temp. w/ rela.ImagePixelImagePixelPixel(PRO)
83.271.298.597.893.9
✓79.768.798.698.194.6
✓86.074.199.298.295.0
", + "image_path": "ca48f368d56afdcbb4ffa974afc355aef358439d57b9669a9c0b82ea8d77cc73.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 298, + 545, + 373 + ], + "blocks": [ + { + "bbox": [ + 306, + 265, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 306, + 265, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 265, + 545, + 289 + ], + "type": "text", + "content": "Table 6: Anomaly detection performance with template bank subsampling at different ratios." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 298, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 373 + ], + "type": "table", + "html": "
Dataset\\Ratio10%25%50%100%
LOCO [3]Image85.686.086.086.0
Pixel73.774.074.074.1
AD [4]Image99.199.299.299.2
Pixel98.198.298.298.2
Pixel(PRO)95.095.095.095.0
", + "image_path": "859b7f154f800bde695f0a0f21859bc73dc9c272daa1e41055975bf71e6a15fd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 431, + 545, + 517 + ], + "blocks": [ + { + "bbox": [ + 305, + 386, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 305, + 386, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 545, + 422 + ], + "type": "text", + "content": "Table 7: The comparison of pre-trained based approaches in terms of inference time (s), memory usage (MB), and performance (image/pixel/pixel-PRO) on MVTec AD [4]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 431, + 545, + 517 + ], + "lines": [ + { + "bbox": [ + 307, + 431, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 431, + 545, + 517 + ], + "type": "table", + "html": "
MethodInf. timeMemoryPerformance
CFLOW [14]0.178947(96.8/97.9/92.7)
RD4AD [11]0.079352(98.5/97.8/93.9)
PatchCore(100%) [28]0.1491015(99.1/98.0/93.3)
PatchCore(10%) [28]0.113102(99.0/98.1/93.5)
Ours(100%)0.1301130(99.2/98.2/95.0)
Ours(10%)0.099448(99.1/98.1/95.0)
", + "image_path": "68e1030cfbac625a4763f9f682a2b142da337201d0b8045368dda66c9289ee59.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 536, + 378, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 378, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 378, + 549 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "In the practical industrial scenarios, it is challenging to identify anomalies from the complicated normal patterns, and meanwhile precisely localize the anomalous regions of various sizes. We propose a novel template-guided hierarchical feature restoration network for anomaly detection. The proposed method recovers anomaly-free features from anomaly features by using bottleneck to filter anomaly features, and compensating compressed features with template retrieved from template bank. The hierarchical design benefits detection upon anomalies of various sizes. Experimental results demonstrate the effectiveness of our approach, especially on localization. Our method achieves state-of-the-art performance on MVTec LOCO AD dataset." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6455" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "text", + "content": "[1] Pankaj Agarwal, Sariel Har, Peled Kasturi, and R Varadarajan. Geometric approximation via coresets. Combinatorial and Computational Geometry, 52, 11 2004. 5, 9" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 180 + ], + "type": "text", + "content": "[2] Samet Akçay, Amir Atapour-Abarghouei, and Toby P. Breckon. Skip-ganomaly: Skip connected and adversarially trained encoder-decoder anomaly detection. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1–8, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 237 + ], + "type": "text", + "content": "[3] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130(4):947-969, 2022. 1, 2, 3, 4, 6, 7, 8, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "text", + "content": "[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad - a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 1, 2, 6, 7, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 296, + 288, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 288, + 351 + ], + "type": "text", + "content": "[5] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 353, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 287, + 396 + ], + "type": "text", + "content": "[6] Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv preprint arXiv:1807.02011, 2018. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 398, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 398, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 53, + 398, + 288, + 442 + ], + "type": "text", + "content": "[7] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5008-5017, June 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 444, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 444, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 53, + 444, + 287, + 476 + ], + "type": "text", + "content": "[8] Kenneth L. Clarkson. Coresets, sparse greedy approximation, and the frank-wolfe algorithm. ACM Trans. Algorithms, 6(4), sep 2010. 5, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 478, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 288, + 510 + ], + "type": "text", + "content": "[9] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 513, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 610 + ], + "type": "text", + "content": "[10] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: A patch distribution modeling framework for anomaly detection and localization. In Alberto Del Bimbo, Rita Cucchiara, Stan Sclaroff, Giovanni Maria Farinella, Tao Mei, Marco Bertini, Hugo Jair Escalante, and Roberto Vezzani, editors, Pattern Recognition. ICPR International Workshops and Challenges, pages 475-489, Cham, 2021. Springer International Publishing. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "text", + "content": "[11] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9737-9746, June 2022. 1, 3, 4, 5, 7, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 148 + ], + "type": "text", + "content": "[13] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 2, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 152, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 217 + ], + "type": "text", + "content": "[14] Denis Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 98-107, January 2022. 2, 7, 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 219, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 547, + 262 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 264, + 547, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 547, + 328 + ], + "type": "text", + "content": "[16] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8791-8800, October 2021. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 331, + 547, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 547, + 396 + ], + "type": "text", + "content": "[17] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 303–319, Cham, 2022. Springer Nature Switzerland. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 399, + 547, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 547, + 430 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 433, + 547, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 547, + 453 + ], + "type": "text", + "content": "[19] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 456, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 547, + 487 + ], + "type": "text", + "content": "[20] Yann LeCun et al. Generalization and network design strategies. _Connectionism in perspective_, 19(143-155):18, 1989. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 490, + 547, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 547, + 544 + ], + "type": "text", + "content": "[21] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9664-9674, June 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 546, + 547, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 547, + 590 + ], + "type": "text", + "content": "[22] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 592, + 547, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 547, + 646 + ], + "type": "text", + "content": "[23] Jonathan Pirnay and Keng Chai. Inpainting transformer for anomaly detection. In Stan Sclaroff, Cosimo Distante, Marco Leo, Giovanni M. Farinella, and Federico Tombari, editors, Image Analysis and Processing - ICIAP 2022, pages 394-406, Cham, 2022. Springer International Publishing. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "type": "text", + "content": "[24] Masoud Pourreza, Bahram Mohammadi, Mostafa Khaki, Samir Bouindour, Hichem Snoussi, and Mohammad Sabokrou. G2d: Generate to detect anomaly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2003-2012, January 2021. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6456" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 128 + ], + "type": "text", + "content": "[25] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2806-2814, June 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "text", + "content": "[26] Oliver Rippel, Patrick Mertens, and Dorit Merhof. Modeling the distribution of normal data in pre-trained deep features for anomaly detection. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6726-6733, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 262 + ], + "type": "text", + "content": "[27] Nicolae-Cătălin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13576-13586, June 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 264, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 330 + ], + "type": "text", + "content": "[28] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14318-14328, June 2022. 1, 2, 3, 7, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "type": "text", + "content": "[29] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Structuring autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, Oct 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 377, + 288, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 288, + 433 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 288, + 433 + ], + "type": "text", + "content": "[30] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1907-1916, January 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 434, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 490 + ], + "type": "text", + "content": "[31] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmetric student-teacher networks for industrial anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2592–2602, January 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 491, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 567 + ], + "type": "text", + "content": "[32] Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel Müller, and Marius Kloft. Deep one-class classification. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 4393-4402. PMLR, 10-15 Jul 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 635 + ], + "type": "text", + "content": "[33] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H. Rohban, and Hamid R. Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14902-14912, June 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 636, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 713 + ], + "type": "text", + "content": "[34] Robin Schirrmeister, Yuxuan Zhou, Tonio Ball, and Dan Zhang. Understanding anomaly detection with deep invertible networks through hierarchies of distributions and features. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 21038-21049. Curran Associates, Inc., 2020. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 139 + ], + "type": "text", + "content": "[35] Hannah M. Schlüter, Jeremy Tan, Benjamin Hou, and Bernhard Kainz. Natural synthetic anomalies for self-supervised anomaly detection and localization. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022, pages 474-489, Cham, 2022. Springer Nature Switzerland. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 140, + 546, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 546, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 546, + 183 + ], + "type": "text", + "content": "[36] Bernhard Scholkopf, John C. Platt, John Shawe-Taylor, Alex J. Smola, and Robert C. Williamson. Estimating the Support of a High-Dimensional Distribution. Neural Computation, 13(7):1443-1471, 07 2001. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 184, + 546, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 546, + 228 + ], + "type": "text", + "content": "[37] Ta-Wei Tang, Wei-Han Kuo, Jauh-Hsiang Lan, Chien-Fang Ding, Hakiem Hsu, and Hong-Tsu Young. Anomaly detection neural network with dual auto-encoders gan and its industrial inspection applications. Sensors, 20(12), 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 228, + 546, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 546, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 546, + 250 + ], + "type": "text", + "content": "[38] David MJ Tax and Robert PW Duin. Support vector data description. Machine learning, 54(1):45-66, 2004. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 251, + 546, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 546, + 316 + ], + "type": "text", + "content": "[39] Shashanka Venkataramanan, Kuan-Chuan Peng, Rajat Vikram Singh, and Abhijit Mahalanobis. Attention guided anomaly localization in images. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 485-503, Cham, 2020. Springer International Publishing. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 316, + 546, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 546, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 546, + 370 + ], + "type": "text", + "content": "[40] Shinji Yamada, Satoshi Kamiya, and Kazuhiro Hotta. Reconstructed student-teacher and discriminative networks for anomaly detection. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2725–2732, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 372, + 546, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 546, + 426 + ], + "type": "text", + "content": "[41] Xudong Yan, Huaidong Zhang, Xuemiao Xu, Xiaowei Hu, and Pheng-Ann Heng. Learning semantic context from normal samples for unsupervised anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence, 35(4):3110-3118, May 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 426, + 546, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 546, + 469 + ], + "type": "text", + "content": "[42] Jihun Yi and Sungroh Yoon. Patch svdd: Patch-level svdd for anomaly detection and segmentation. In Proceedings of the Asian Conference on Computer Vision (ACCV), November 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 471, + 546, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 546, + 525 + ], + "type": "text", + "content": "[43] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 526, + 546, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 546, + 570 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 546, + 570 + ], + "type": "text", + "content": "[44] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint arXiv:2111.07677, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 571, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 546, + 624 + ], + "type": "text", + "content": "[45] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Edwin R. Hancock Richard C. Wilson and William A. P. Smith, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 87.1-87.12. BMVA Press, September 2016. 6, 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 625, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 546, + 679 + ], + "type": "text", + "content": "[46] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem - a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8330-8339, October 2021. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 681, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 546, + 713 + ], + "type": "text", + "content": "[47] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Reconstruction by inpainting for visual anomaly detection. Pattern Recognition, 112:107706, 2021. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6457" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 361 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[48] Matthew D. Zeiler, Dilip Krishnan, Graham W. Taylor, and Rob Fergus. Deconvolutional networks. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 2528-2535, 2010. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 171 + ], + "type": "text", + "content": "[49] Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, Xin Jin, and Zhibo Chen. Relation-aware global attention for person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "text", + "content": "[50] Ye Zheng, Xiang Wang, Rui Deng, Tianpeng Bao, Rui Zhao, and Liwei Wu. Focus your distribution: Coarse-to-fine noncontrastive learning for anomaly detection and localization. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 283 + ], + "type": "text", + "content": "[51] Bo Zong, Qi Song, Martin Renqiang Min, Wei Cheng, Cristian Lumezanu, Daeki Cho, and Haifeng Chen. Deep autoencoding gaussian mixture model for unsupervised anomaly detection. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 361 + ], + "type": "text", + "content": "[52] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In Shai Avidan, Gabriel Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision – ECCV 2022, pages 392–408, Cham, 2022. Springer Nature Switzerland. 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6458" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/3b08deeb-33eb-47b0-94a4-e0af89436d2d_content_list.json b/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/3b08deeb-33eb-47b0-94a4-e0af89436d2d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f15e796547f94704a4afa87c40d12ed21bbaa732 --- /dev/null +++ b/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/3b08deeb-33eb-47b0-94a4-e0af89436d2d_content_list.json @@ -0,0 +1,1585 @@ +[ + { + "type": "text", + "text": "Temporal Collection and Distribution for Referring Video Object Segmentation", + "text_level": 1, + "bbox": [ + 84, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiajin Tang Ge Zheng Sibei Yang", + "bbox": [ + 330, + 180, + 640, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Information Science and Technology, ShanghaiTech University", + "bbox": [ + 192, + 205, + 774, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{tangjj,zhengge,yangsb}@shanghaiitech.edu.cn", + "bbox": [ + 292, + 224, + 671, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://toneyaya.github.io/tempcd", + "bbox": [ + 336, + 244, + 633, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 292, + 313, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring video object segmentation aims to segment a referent throughout a video sequence according to a natural language expression. It requires aligning the natural language expression with the objects' motions and their dynamic associations at the global video level but segmenting objects at the frame level. To achieve this goal, we propose to simultaneously maintain a global referent token and a sequence of object queries, where the former is responsible for capturing video-level referent according to the language expression, while the latter serves to better locate and segment objects with each frame. Furthermore, to explicitly capture object motions and spatial-temporal cross-modal reasoning over objects, we propose a novel temporal collection-distribution mechanism for interacting between the global referent token and object queries. Specifically, the temporal collection mechanism collects global information for the referent token from object queries to the temporal motions to the language expression. In turn, the temporal distribution first distributes the referent token to the referent sequence across all frames and then performs efficient cross-frame reasoning between the referent sequence and object queries in every frame. Experimental results show that our method outperforms state-of-the-art methods on all benchmarks consistently and significantly.", + "bbox": [ + 76, + 325, + 473, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring video object segmentation task (RVOS) aims to segment the target referent throughout a video sequence given a natural language expression [22]. It has attracted increasing attention from the academic community, as it is a basis for assessing a comprehensive understanding of visual, temporal and linguistic information. Meanwhile, ROVS benefits various downstream interactive applications such as language-driven human-robot interaction [39], video editing [25] and video surveillance [44].", + "bbox": [ + 75, + 741, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bafe92976a35acedb3b8ce976d1db92ce74f7a8af21b70493eb0c521e1411e84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 292, + 883, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4f611c80f4a4201d2996c73827234247737fe60153475524c6b58d4fcb4293db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 366, + 883, + 439 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/931cffb550ef7d9911fd981ff68b348b47676c55072ab0ee144a76fc4002f7e6.jpg", + "image_caption": [ + "Figure 1. The comparison of (a) previous best-performing ReferFormer, (b) video-level decoding, and (c) ours referent-guided dynamic temporal decoding." + ], + "image_footnote": [], + "bbox": [ + 504, + 441, + 880, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compared to referring image segmentation [12,62] that segments the target object in a single image mainly according to appearances and spatial and semantic relations, ROVS requires locating referents with temporal consistency according to object motions and dynamic associations among objects across frames.", + "bbox": [ + 496, + 611, + 890, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous works mainly follow a single-stage bottom-up manner [9, 15, 38, 47, 48] or a two-stage top-down fashion [28]. The bottom-up methods early align the vision and language at the local patch level, failing to explicitly model objects and their relations [28, 53]. Although top-down approaches explicitly extract object tracklets and select object tracklets matched with the language expression, their two-stage pipeline is complex and less efficient [53]. Recently, transformer-based frameworks [1, 53] have been proposed to use object queries to capture objects and their associations in an end-to-end manner. They generate object queries and utilize the transformer decoder to search objects for these object queries in every frame, as shown in", + "bbox": [ + 496, + 704, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Sibei Yang is the corresponding author.", + "bbox": [ + 94, + 886, + 310, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15466", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2a7980f87e9de4412ff13fde2c04a413ae106c5127220478fc3632414e609da.jpg", + "image_caption": [ + "Figure 2. Comparison of temporal modeling for previous works, naive attempt, and ours. Previous works (MTTR and ReferFormer) lack temporal modeling when decoding. Besides, object motion and interaction with other objects are crucial for referring video object segmentation. However, simply modeling temporal interaction with globally visible video-level decoding cannot achieve satisfactory results. Note that our alternating interaction between the global referent and object queries can dynamically capture spatial-temporal cross-modal reasoning over objects guided by the global referent information." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 890, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Figure 1a and b.", + "text_level": 1, + "bbox": [ + 76, + 364, + 187, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although these end-to-end transformer-based frameworks have achieved impressive segmentation results, they still have limitations needed to be improved. First, their object queries are frame-level and independently responsible for the search for objects in each frame, which fails to capture temporal object motions and effective object tracking, as shown in Figure 2a and b. Therefore, we observe that they often fail to obtain temporally consistent predictions for target referents across an entire video, as shown in the left example in Figure 2. Second, their interactions between object queries are also solely performed per frame, leading to failure to model object-level spatial-temporal associations between objects. Therefore, they cannot correctly ground expressions that require cross-modal spatial-temporal reasoning over multiple objects. For example (see the right example in Figure 2), they mislocate the left camel as the target object because they lack to model the relation between the left camel in the current frame and the camel walking in the subsequent frame.", + "bbox": [ + 75, + 383, + 473, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim to address these limitations by empowering the decoder with the capability of temporal modeling. A straightforward approach to decode objects with temporal information is to convert the frame-level decoding into video-level by inputting a sequence of object queries to search for objects in a sequence of frames like VisTR [49] for video instance segmentation, as shown in Figure 1b and 2e. However, this simple attempt fails to achieve satisfactory results while significantly increasing the computational cost and even underperforming existing query-based transformer frameworks. We further analyze the reasons: (1) The alignment between the natural language expression and the referent relies on overall objects' motions and temporal associations in the entire video. For example, the two camels can be distinguished only if the walking motion of", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the target camel is identified based on the entire video and aligned with the language expression. In contrast, the above naive attempt prematurely aligns language with fine-grained frame-level objects so that its attention is distracted from focusing accurately on the overall motions and relations of objects at the video level. (2) However, the precise localization and segmentation of target objects should go back to and rely more on every single frame because objects with different motions may cause them to have very different spatial locations in different frames. A similar observation is discussed by SeqFormer [54] for video instance segmentation, which suggests processing attention with each frame independently.", + "bbox": [ + 496, + 364, + 893, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, to address the challenge of aligning the language expression with objects' motions and temporal associations at the global video level but segmenting objects at the local frame level, we propose to maintain both local object queries and a global referent token. Global referent token captures the video-level referent information according to the language expression, while local object queries locate and segment objects with each frame. Furthermore, the object queries and the referent token are interacted to achieve spatial-temporal object information exchange through our well-designed temporal collection and distribution mechanisms, as shown in Figure 2f. Specifically, the temporal collection collects the referent information from object queries with spatial and visual attributes to temporal object motions to language semantics. In turn, the temporal distribution first dynamically distributes the referent token to every frame to extract the referent sequence. Then, the referent sequence interacts with object queries within each frame to achieve efficient spatial-temporal reasoning over objects. Note that the temporal collection and distribution alternately propagate information between object queries and the referent token to update each other.", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15467", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we propose a Temporal Collection and Distribution network (TempCD) which integrates our novel temporal collection and distribution mechanisms into the query-based transformer framework. Without using sequence matching or post-processing during inference like the previous methods [1,53], Our TempCD can directly predict the segmentation result of every frame based on object queries in the frame and the referent token.", + "bbox": [ + 76, + 90, + 472, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows,", + "bbox": [ + 96, + 213, + 434, + 227 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce to maintain a global referent token and a sequence of local object queries parallelly to bridge the gap between video-level object alignment with language and frame-level object segmentation.", + "- We propose a novel collection-distribution mechanism for interacting between the referent token and object queries to capture spatial-temporal cross-modal reasoning over objects.", + "- We present an end-to-end Temporal Query Collection and Distribution (TempCD) for RVOS, which can directly predict the segmentation referent of every frame without any post-processing and sequence matching.", + "- Experiments on Ref-Youtu-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB-Sentences datasets show that TempCD outperforms state-of-the-art methods on all benchmarks consistently and significantly." + ], + "bbox": [ + 94, + 239, + 467, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 496, + 218, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Referring Image Segmentation (RIS) involves segmenting objects in images based on natural language expressions. Compared to Referring Video Object Segmentation (RVOS), RIS operates on individual images without temporal information. Previous works [3, 4, 26, 29, 42, 45, 57-59] focus on the joint modeling of vision and language. Various methods [6, 13, 14, 16, 20, 21, 35, 60, 64] are explored successively, such as using fusion operators like concatenation, ConvLSTM, attention mechanisms [46], and GNN [10] to obtain multimodal semantic feature maps. In addition, some studies attempt to decouple different components or key semantics from language, through explicit two-stage approaches [51, 66] or implicit attention modules [43, 62, 64, 65], to achieve a more fine-grained understanding. Recent models explore novel fusion techniques, including early fusion [8, 63], Linguistic Seed Encoder [23], and contrastive learning [50], to improve the performance of RIS. However, for RVOS, it requires not only the multimodal integration of language and image frames but also temporal modeling at the video level.", + "bbox": [ + 76, + 522, + 472, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Referring Video Object Segmentation (RVOS) aims to segment the target object described in natural language from videos. Previous works in RVOS has primarily used two frameworks: bottom-up and top-down. The top-down framework [28] directly models the motion informa", + "bbox": [ + 76, + 825, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion between segmentation masks of consecutive frames. However, this two-stage approach, which includes complex computational costs, is limited in terms of video input length. For the bottom-up framework, some approaches [9, 15, 38, 47, 48] directly apply RIS methods to construct multi-modal feature maps for referring segmentation of keyframes. They rely solely on replacing the traditional image backbone with a 3D temporal backbone for temporal modeling, which limits the performance of multiframes segmentation. To address this issue, URVOS [41] extends previous methods by building a memory bank that propagates language referent object information in the temporal dimension. Furthermore, URVOS introduces the RefYoutube-VOS [41] dataset, which provides segmentation annotations for each frame that require more efficient temporal modeling. In addition, recent works [7, 52, 68] adopt language guided fusion between temporal features and visual features to obtain more efficient temporal modeling. Recently, query-based methods [1, 53] with a transformer encoder-decoder frameworks [2, 5, 69] attempt to capture object-level information. They achieve higher performance by constructing a multimodal feature map of visual and linguistic information, or by using language-initialized queries with independently decoding with queries for each frame.", + "bbox": [ + 496, + 90, + 890, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Modeling in Video Instance Segmentation (VIS). VIS aims to simultaneously track, segment and classify interest instances in a video. Similarly to RVOS, VIS requires obtaining instance information from every frame in the video and maintaining the temporal consistency of instances. Recent works focus on spatial and temporal modeling based on instance-level information. VisTR [49] employs an encoder-decoder transformer based on DETR, with a concatenated sequence of instance queries as the input to enable spatial-temporal interactions. The following IFC [17] and TeViT [61] explored a more efficient approach to temporal modeling in the encoder by introducing supplementary tokens for interaction across frames, with the decoder aligned with VisTR. Other following works [24, 56] take advantage of memories or queries from several previous frames to help instance segmentation in the current frame, extending to a online fashion. Recently, SeqFormer [54] proposes that the acquisition of frame-level instance information such as position needs to go back to each frame, i.e., the decoding process should be done independently for each frame. While these methods provide potential solutions for temporal modeling, the direct application to RVOS is not feasible, because of inconsistent object semantics over frames.", + "bbox": [ + 496, + 454, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 829, + 591, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The framework of our proposed TempCD is shown in Figure 3. We introduce a sequence of object queries to capture object information in every frame independently and", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15468", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/494aa9f57b9ad145567289aeec76eac8c793e0eeba4ed37fe906091c9e192b24.jpg", + "image_caption": [ + "Figure 3. Overall framework of TempCD. TempCD simultaneously maintains a global referent token and a sequence of object queries, and they interact via the proposed collection-distribution mechanism. Specifically, the temporal collection collects global information for the referent token from object queries to the temporal motions to the language expression. In turn, the temporal distribution first distributes the referent token to the referent sequence and then performs efficient cross-frame reasoning between the referent sequence and object queries." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 851, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "maintain a referent token to capture the global referent information for aligning with language expression. The referent token and object queries are alternately updated to stepwise identify the target object and better segment the target object in every frame through the proposed temporal collection and distribution mechanisms.", + "bbox": [ + 75, + 462, + 470, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we first introduce encoders and the definition of object queries in Section 3.1. Then, we present the temporal collection that collects the referent information for the referent token from object queries to temporal object motions to the language expression (see Section 3.2). Next, we introduce the temporal distribution mechanism that distributes the global referent information to the referent sequence across frames to object queries in each frame (see Section 3.3). The collection-distribution mechanism explicitly captures object motions and spatial-temporal cross-modal reasoning over objects. Finally, we introduce segmentation heads and loss functions in Section 3.4. Note that as we explicitly maintain the referent token, we can directly identify the referent object in each frame without requiring sequence matching as in previous queries-based methods.", + "bbox": [ + 75, + 551, + 468, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Encoders and Query Definition", + "text_level": 1, + "bbox": [ + 76, + 786, + 352, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language Encoder. Following previous work [53], we adopt RoBERTa [32] as our language encoder. Given a language expression with $L$ words, we extract its sentence feature $\\mathcal{F}_s \\in \\mathbb{R}^C$ , where $C$ denotes channel dimensions. In addition, we obtain the output feature representation of each word before the last pooling layer as $\\mathcal{F}_w \\in \\mathbb{R}^{L \\times C}$ .", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Backbone and Encoder. Following previous works [1, 52, 53, 61], we use ResNet-50 [11] as our visual backbone. We also follow recent work [1, 53] to evaluate our model using the temporal visual backbone network Video-Swin [33] in experiments. For an input video $V \\in \\mathbb{R}^{H \\times W \\times 3 \\times T}$ with $T$ frames, we first extract its visual feature maps and further construct multimodal feature maps $\\mathcal{V} = \\{v_i\\}_{i=1}^T$ by fusing the visual feature maps with the word features of the expression $\\mathcal{F}_w$ following [53].", + "bbox": [ + 496, + 462, + 890, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Query Definition. Inspired by query-based detection and segmentation frameworks [2, 5, 69], we leverage object queries to represent object-level information. For the video with $\\mathcal{T}$ frames, we introduce a sequence of $\\mathcal{T} \\times N$ object queries to represent the objects in every frame and follow [53] to use the language feature $\\mathcal{F}_s$ to initialize the object queries as $\\mathcal{Q} \\in \\mathbb{R}^{\\mathcal{T} \\times N \\times C}$ , where the number of objects and the dimension of channels are $N$ and $C$ , respectively.", + "bbox": [ + 496, + 598, + 890, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Referent Token Definition. In addition to adopting the object queries to represent objects in every frame, we define a referent token $\\mathcal{Q}^g\\in \\mathbb{R}^C$ to represent the global information of the target object. As the language expression naturally indicates the referent information, we thus use the language feature $\\mathcal{F}_s$ to initialize the $\\mathcal{Q}^g$ .", + "bbox": [ + 496, + 734, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Temporal Collection from Object Queries", + "text_level": 1, + "bbox": [ + 500, + 832, + 856, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The temporal collection mechanism collects object motions from frames and aggregates object motions consistent with the language expression into the global referent token.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15469", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first perform motion collection across object queries (see Section 3.2.1) and then update the global referent token from the object motions and language expression (see Section 3.2.2).", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1 Motion Collection across Object Queries", + "text_level": 1, + "bbox": [ + 76, + 167, + 413, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For simplicity of demonstration, we denote the object queries output from the $(l - 1)$ -th layer as $\\mathcal{Q}_{l - 1}$ and input them into the $l$ -th layer of TempCD. Note that we input the object queries $\\mathcal{Q}$ defined in Section 3.1 into the first layer of TempCD. We utilize $\\mathcal{Q}_{l - 1}$ to capture object information in each frame and integrate them to obtain temporal information such as object motion and action. Specifically, for $i$ -th frame, we feed its object queries $q_{i} = \\mathcal{Q}_{l - 1}^{i} \\in \\mathbb{R}^{N \\times C}$ and the multimodal features $v_{i}$ into the DETR transformer decoder [2,69] to locate objects, where $\\mathcal{Q}_{l - 1}^{i}$ means the object queries of $i$ -th frame. The computation proceeds as follows,", + "bbox": [ + 75, + 191, + 470, + 359 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {Q}} _ {l} = \\left\\{\\operatorname {D E T R} \\left(q _ {i}, v _ {i}\\right) \\right\\} _ {i = 1} ^ {\\mathcal {T}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 371, + 468, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{DETR}(\\cdot ,\\cdot)$ represents the DETR transformer decoder, and $\\hat{\\mathcal{Q}}_l$ denotes the object queries output by the decoder.", + "bbox": [ + 75, + 393, + 468, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, to aggregate object motions across the video sequence, we combine objects information from queries of each frame. Specifically, we generate temporal weights $G \\in \\mathbb{R}^{T \\times N \\times 1}$ for each frame given the object queries $\\hat{\\mathcal{Q}}_l$ . This is done by employing a nonlinear layer followed by a softmax operator. Then, we use the temporal weights $G$ to integrate the frame-level information and obtain object motions $M \\in \\mathbb{R}^{N \\times C}$ through a gating mechanism. The integration is computed as follows,", + "bbox": [ + 75, + 439, + 468, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG = \\operatorname {s o f t m a x} (\\hat {\\mathcal {Q}} _ {t} W),\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 580, + 348, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM = \\sum_ {i = 1} ^ {\\mathcal {T}} \\left(G _ {i} \\odot \\hat {\\mathcal {Q}} _ {l} ^ {i}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 602, + 468, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W \\in \\mathbb{R}^{C \\times 1}$ denotes the learnable weights of the linear layer, $G_{i} \\in \\mathbb{R}^{N \\times 1}$ is the weight for $i$ -th frames in $G$ , $\\mathrm{softmax}(\\cdot)$ represents the softmax operation along the temporal dimension, and $\\odot$ is the element-wise product.", + "bbox": [ + 75, + 650, + 468, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Global Referent Token Collection", + "text_level": 1, + "bbox": [ + 76, + 729, + 362, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we update the global referent token $\\mathcal{Q}_{l-1}^g \\in \\mathbb{R}^C$ at the $(l-1)$ -th layer to referent token $\\mathcal{Q}_l^g \\in \\mathbb{R}^C$ by selecting the object motions $M \\in \\mathbb{R}^{N \\times C}$ consistent with the language feature $\\mathcal{F}_s \\in \\mathbb{R}^C$ . Note that we input the global referent token $\\mathcal{Q}^g$ defined in Section 3.1 into the first layer of TempCD. We first calculate the cosine similarity between the language feature $\\mathcal{F}_s$ and the object motions $M$ . We obtain the similarity scores $S = \\{s_j\\}_{j=1}^N$ as follows,", + "bbox": [ + 75, + 752, + 468, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {j} = \\cos \\left(\\mathcal {F} _ {s}, m _ {j}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 883, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot ,\\cdot)$ means the cosine similarity function and $m_j$ denotes the $j$ -th object's motion feature in $M$ . Next, we update the global referent token $\\mathcal{Q}_g$ as the motion feature of the object with the highest similarity score to the global referent token. To ensure the selection process is differentiable, we implement this selection through the Gumbel Softmax operator [18, 36]. The calculation is as follows:", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\text {g u m b e l}} = \\operatorname {s o f t m a x} \\left(\\left(S + G _ {s}\\right) / \\tau\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 208, + 784, + 224 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\text {o n e h o t}} = \\text {o n e h o t} \\left(\\arg \\max _ {N} \\left(S _ {\\text {g u m b e l}}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 227, + 890, + 243 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\text {r e f e r e n t}} = S _ {\\text {o n e h o t}} - \\mathrm {s g} \\left[ S _ {\\text {g u m b e l}} \\right] + S _ {\\text {g u m b e l}},\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 244, + 831, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\operatorname{softmax}(\\cdot)$ represents the softmax operation along dimension $N$ , $G_{s} \\in \\mathbb{R}^{N \\times 1}$ is sampled from the Gumbel(0,1) distribution, $\\tau$ is the temperature coefficient hyperparameter, and $\\mathrm{sg}$ is the stop gradient operation. Additionally, $\\operatorname{argmax}_{N}$ represents choosing the index with the highest similarity score, and onehot represents setting the highest-scoring item to 1 based on the selected index and setting the others to 0. The one-hot vector $S_{\\text{referent}} \\in \\mathbb{R}^{N \\times 1}$ indicates the selection from the object motions. The global referent token $\\mathcal{Q}^g \\in \\mathbb{R}^C$ is updated as follows:", + "bbox": [ + 496, + 273, + 890, + 425 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} _ {l} ^ {g} = \\operatorname {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(S _ {\\text {r e f e r e n t}} ^ {\\top} M\\right) + Q _ {l - 1} ^ {g}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 434, + 890, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{MLP}(\\cdot)$ denotes a three-layer linear layer with activation function, and LayerNorm $(\\cdot)$ is the same as the LayerNorm in standard Transformer.", + "bbox": [ + 496, + 463, + 890, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We obtain the updated global referent token $\\mathcal{Q}^g$ that contains information on the referent object aligned with the language expression. Next, we feed it into the Temporal Distribution module to assist in distributing the referent information into each frame.", + "bbox": [ + 496, + 508, + 890, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Temporal Distribution from Referent Token", + "text_level": 1, + "bbox": [ + 500, + 594, + 870, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The temporal distribution aims to propagate the referent information to all frames and perform dynamically spatial-temporal reasoning over objects based on the referent information. First, we distribute the global referent token to each frame independently to extract the referent sequence across all frames (see Section 3.3.1). Next, we perform the efficient cross-frame reasoning between the referent sequence and object queries in every frame (see Section 3.3.2).", + "bbox": [ + 496, + 617, + 890, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Global Referent Token Distribution", + "text_level": 1, + "bbox": [ + 500, + 756, + 800, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the global referent token $\\mathcal{Q}_l^g$ , we distribute the referent object information to every frame and further obtain the referent sequence across all frames, denoted by $\\mathcal{R}_l = \\{r_i\\}_{i=1}^T$ , which refers to the target object in each frame.", + "bbox": [ + 496, + 779, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we first compute the cosine similarity scores between the object queries $\\hat{\\mathcal{Q}}_l = \\{q_i\\}_{i=1}^T$ and the global referent token $\\mathcal{Q}_l^g$ . Then we obtain the index $d_i$ of the", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15470", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "object query with the highest similarity score to the global referent token for every frame $i$ independently. The computation is as follows,", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i, j} = \\cos \\left(q _ {i, j}, Q _ {l} ^ {g}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 147, + 468, + 171 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd _ {i} = \\operatorname {a r g m a x} _ {j} \\left\\{s _ {i, j} \\right\\},\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 167, + 344, + 184 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $q_{i,j}$ is the $j$ -th object query in the $i$ -th frame, and $s_{i,j}$ is the corresponding similarity score. For every frame $i$ , we select the object query $d_i$ with the highest similarity score and fuse it with the global referent token $\\mathcal{Q}_l^g$ to generate the referent sequence $\\mathcal{R}_l = \\{r_i\\}_{i=1}^T$ as follows,", + "bbox": [ + 75, + 188, + 468, + 266 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {l} = \\left\\{\\mathrm {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(q _ {i, d _ {i}} + \\mathcal {Q} _ {l} ^ {g}\\right)\\right) \\right\\} _ {i = 1} ^ {\\mathcal {T}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 273, + 468, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We end up with referent sequence $\\mathcal{R}_l\\in \\mathbb{R}^{T\\times C}$ across all frames, which is used as crucial objects to interact with object queries in every frame independently.", + "bbox": [ + 75, + 303, + 468, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.2 Spatial-temporal Reasoning over Object Queries", + "text_level": 1, + "bbox": [ + 76, + 367, + 468, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We leverage the referent sequence $\\mathcal{R}_l\\in \\mathbb{R}^{T\\times C}$ , which encodes the referent object information across all frames, to interact with the object queries to achieve the cross-frame temporal reasoning over multiple objects and further distribute referent information to all object queries in all frames.", + "bbox": [ + 75, + 388, + 468, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, for each object query $q_{i,j}$ at the $i$ -th frame, we concatenate it with the objects of the referent sequence in all other frames to construct the cross-object query sequence $C_{i,j} = \\{r_1,..,r_{i - 1},q_{i,j},r_{i + 1},..,r_{\\mathcal{T}}\\} \\in \\mathbb{R}^{\\mathcal{T}\\times C}$ . The object-level interaction across frames is implemented by multi-head self attention, which is computed as follows,", + "bbox": [ + 75, + 481, + 468, + 571 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} _ {i, j} = \\operatorname {M H S A} \\left(C _ {i, j}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 585, + 468, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{MHSA}(\\cdot)$ denotes the multi-head self-attention mechanism. The $j$ -th updated object queries in $C_{i,j}^{\\hat{}}$ for every frame $i$ are concatenated as $\\mathcal{Q}_l^f \\in \\mathbb{R}^{\\mathcal{T} \\times N \\times C}$ to input to next decoder layer. Note that our cross-frame interaction only requires interaction with referent sequences to object queries in each frame, which achieves efficient cross-frame interaction over multiple objects.", + "bbox": [ + 75, + 609, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Prediction Heads and Loss", + "text_level": 1, + "bbox": [ + 76, + 726, + 316, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We integrate the referent sequence and object queries in each decoder layer to the prediction heads for predicting results. To simplify the presentation, we use the final layer for details.", + "bbox": [ + 75, + 750, + 468, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We extract the referent sequence $\\mathcal{R} \\in \\mathbb{R}^{T \\times C}$ and combine them with the object queries $Q_{l}$ to obtain the output referent sequence $\\hat{\\mathcal{R}}$ for prediction. Specifically, we use a gating mechanism to integrate the information from $\\hat{C}$ into video-level features, which is similar to Eq 2. Next, we calculate cosine similarity between each sequence $C_{j,i}$ and the", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "language sentence feature $\\mathcal{F}_s$ , which is similar to Eq 3, to obtain the score for each object $q_{i,j}$ . We select the highest-scoring object $q_i^r$ in each frame and combine it with $\\mathcal{R}$ to obtain the final referent sequence $\\hat{\\mathcal{R}} \\in \\mathbb{R}^{T \\times C}$ of the referred object in each frame:", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {R}} = \\left\\{\\mathrm {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(r _ {i} + q _ {i} ^ {r}\\right)\\right) \\right\\}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 179, + 890, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we feed $\\hat{\\mathcal{R}}$ into the segmentation head to predict the segmentation results. Additionally, similar to [53] and [1], we use a box prediction head to regress the box coordinates for each query. We utilize Dice [37] loss and Focal [30] loss as the segmentation loss and GIOU [40] loss and L1 loss as the box loss.", + "bbox": [ + 496, + 208, + 890, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 498, + 313, + 624, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on four benchmark datasets that are publicly available: Ref-Youtu-VOS [41], Ref-Davis-2017 [22], A2D-Sentences [9], and JHMDB-Sentences [9]. Ref-Youtu-VOS consists of 3978 videos and approximately 15K language descriptions. Ref-Davis-2017 contains approximately 90 videos. A2D-Sentences and JHMDB-Sentences [19, 55], originally focused on action recognition, have been expanded to incorporate language expression annotations. This expansion has yielded a total of 3782 videos for A2D-Sentences and 928 videos for JHMDB-Sentences, each accompanied by corresponding language descriptions.", + "bbox": [ + 496, + 338, + 890, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Following [53], we pre-train our model on RefCOCO dataset [67]. We utilize a text encoder derived from RoBERTa [32], coupled with either ResNet50 [11] or Video-Swin serving [33] as our visual backbone. We train our model for 6 epochs with an initial learning rate of 1e-4 and the AdamW [34] optimizer. Consistent with prior works [1, 41, 53], we use the $\\mathcal{I}$ , $\\mathcal{F}$ , and $\\mathcal{J} \\& \\mathcal{F}$ metrics for evaluation on Ref-Youtu-VOS and Ref-Davis-2017 datasets, and Overall IoU and Mean IoU as evaluation metrics on A2D-Sentences and JHMDB-Sentences datasets.", + "bbox": [ + 496, + 520, + 890, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 498, + 696, + 872, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1 and Table 2, we compare TempCD with state-of-the-art methods on four benchmarks. TempCD consistently outperforms state-of-the-art methods on all datasets.", + "bbox": [ + 496, + 719, + 890, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison of Ref-Youtube-VOS and Ref-Davis-2017 Datasets. Results for the Ref-Youtu-VOS and Ref-Davis-2017 datasets are presented in Table 1. With a standard ResNet-50 [11] visual backbone, our model exhibits improvements of $2.7\\%$ , $4\\%$ , and $3.4\\%$ for the $\\mathcal{I}, \\mathcal{F}$ , and $\\mathcal{J} \\& \\mathcal{F}$ metrics, respectively, on the Ref-YoutuVOS dataset. Employing the advanced temporal visual backbone, VideoSwin-B [33], our approach consistently outperforms the", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15471", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a84d4845037750c995256ff0f09d6ae81703b401571e17c14f099eab43a482d0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneMethodRef-Youtube-VOSRef-Davis17
JFJ&FJFJ&F
ResNet-50CMSA [65]33.336.534.932.237.234.7
CMSA+RNN [65]34.838.136.436.943.540.2
URVOS [41]45.349.247.247.356.051.5
MLRL [52]48.451.049.750.155.452.7
LBDT [7]48.250.649.4--54.5
ReferFormer [53]54.856.555.655.861.358.5
Ours57.560.559.057.362.760.0
Video-Swin-TMTTR [1]54.056.655.3---
ReferFormer [53]58.060.959.4---
Ours60.564.062.359.365.062.2
Video-Swin-BReferFormer [53]61.364.662.958.164.161.1
Ours63.668.065.861.667.664.6
", + "bbox": [ + 160, + 89, + 810, + 291 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/04f742db8df15990ece20e0225ab6711af0123fc6ce9033186a5223e7d2ea33b.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-art models for RVOS on Ref-Youtu-VOS and Ref-Davis-2017 datasets." + ], + "table_footnote": [], + "table_body": "
BackboneMethodA2D-SJHMDB-S
oIoUmIoUoIoUmIoU
VGG16Hu et al. [12]47.435.054.652.8
I3DGavrilyuk et al. [9]53.642.154.154.2
ACAN [48]60.149.075.656.4
CMPC-V [31]65.357.361.661.7
Resnet-50ClawCraneNet [27]63.159.964.465.6
MMMMTBVS [68]67.355.861.961.3
LBDT [7]70.462.164.565.8
Ours76.668.670.669.6
", + "bbox": [ + 89, + 329, + 457, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison with state-of-the-art models for RVOS on A2D-Sentences (A2D-S) and JHMDB-Sentences (JHMDB-S).", + "bbox": [ + 76, + 502, + 468, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "previous state-of-the-art model [53] on the Ref-Davis-2017 dataset, achieving a $3.5\\%$ increase across the aforementioned metrics. Note that evaluations on the RefDavis-2017 dataset are performed using models trained on the RefYoutube-VOS dataset.", + "bbox": [ + 75, + 546, + 468, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we present a thorough evaluation of the effectiveness of our TempCD through subsequent comparisons: (1) We compare our model to MTTR [1], a pioneering method that adopts a query-based framework for RVOS, despite lacking temporal modeling in both the encoder and decoder stages. Our TempCD surpasses MTTR, achieving a $7\\%$ improvement in the $\\mathcal{J}\\& \\mathcal{F}$ metric on the Ref-YoutuVOS dataset. This outcome underscores the heightened efficacy of our introduced temporal decoding modules. (2) Compared with LBDT [7], our TempCD achieves a significant $9.6\\%$ enhancement in the $\\mathcal{J}\\& \\mathcal{F}$ metric. This result suggests that, unlike [7, 41, 52] that integrate temporal and visual features via a bottom-up framework, our proposed TempCD achieves more efficient temporal modeling.", + "bbox": [ + 75, + 626, + 468, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on A2D-Sentences and JHMDB-Sentences. As presented in Table 2, our approach yields mean improvements of $6.2\\%$ in Overall IoU and $5.2\\%$ in Mean IoU, surpassing the predominant state-of-the-art method [7] for", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "these two datasets. In contrast to Ref-Youtu-VOS, these datasets predominantly comprise segmentation annotations for keyframes that encapsulate actions. When compared to prior studies [9, 15, 38, 47, 48] that adopt temporal encoders to process video temporal information, our method consistently attains enhancements across these two datasets. This highlights the significance of explicitly capturing and integrating video-level temporal context into individual frames.", + "bbox": [ + 496, + 332, + 890, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 465, + 653, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the impact of modules in our model, we evaluate six variants of our TempCD on Ref-Youtu-VOS dataset. The results are shown in Table 3. All experiments are performed with Video-Swin-B as the visual backbone.", + "bbox": [ + 496, + 489, + 890, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. (1) Our baseline model employs $\\mathcal{T}$ sets of frame-level object queries (\"Local Query\"). These queries are concatenated sequentially and fed into the decoder, following a similar approach to VisTR [49]. The derived results (53.7% $\\mathcal{J}$ , 56.3% $\\mathcal{F}$ , and 55.0% $\\mathcal{J} \\& \\mathcal{F}$ ) suggest a misalignment between the natural language representations and the referent semantics at the individual frame level. (2) Another subsequent baseline is established by integrating a singular shared set of video-level object queries across all frames (\"Global Query\"). The global queries are performed an inner product with the feature map $v$ of each frame, yielding segmentation outcomes. While this approach intrinsically aligns with language representations and achieves consistency between frames, it struggles to adapt to frame-level variability, resulting in a performance decline relative to the first baseline.", + "bbox": [ + 496, + 551, + 892, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Query Collection and Distribution. (3) A natural enhancement for baseline (1) entails integrating cross-frame temporal attention on a referent sequence, which is selected from local queries by the similarity with language. This modification leads to a performance improvement of $3.4\\%$ . (4) Another endeavor to enhance the baseline (1) involves the introduction of an additional set of global queries. These", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15472", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67f0786b2f7e9b6f23e9232133078953265aa0ec4c670a7e8395adff828d2075.jpg", + "image_caption": [ + "Figure 4. Visualization for segmentation results of previous best-performing ReferFormer [53] and ours." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 890, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f499ba1123f5edf0bf3665ac67d62f8bff8c3048d7b48338bfd6749060e82275.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodJFJ&F
1Local Query53.756.355.0
2Global Query52.955.554.2
31+Cross-Frame56.259.157.6
41+Motion Collection58.161.559.8
53+460.764.462.6
65+Gumble-Softmax61.465.463.4
7Ours63.668.065.8
", + "bbox": [ + 84, + 292, + 460, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation Study on Ref-Youtu-VOS dataset.", + "bbox": [ + 109, + 429, + 433, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "queries collect motion information from local queries and serve to distribute global context to local queries. Specifically, we utilize a Top-K strategy to select global information based on the similarity score with language. This facilitates implicit frame level interaction and, significantly, fosters semantic alignment between global queries and language expressions. This alignment contributes to a marked performance gain of $5.6\\%$ . (5) We further incorporate both (3)'s explicit cross-frame temporal modeling and (4)'s motion collection and distribution, which delivers a $2.8\\%$ improvement. (6) We replace the Top-K selection, which does not allow gradient calculation, with Gumbel Softmax, improving $\\mathcal{J} \\& \\mathcal{F}$ by $0.8\\%$ . (7) Our complete model further enhances the vanilla cross-frame temporal attention mechanism. Instead of solely updating the referent sequence via temporal interaction, local object queries are also extended to incorporate temporal context information from referent sequence, facilitating cross-frame interaction over multiple objects. The full decoder of our proposed method achieves the performance of $65.8\\%$ in terms of $\\mathcal{J} \\& \\mathcal{F}$ .", + "bbox": [ + 75, + 455, + 472, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Visualization", + "text_level": 1, + "bbox": [ + 76, + 770, + 215, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4 visualizes several qualitative results. The referring expression in (a) describes the motion of a white duck, a pivotal feature that distinguishes it from similar ducks. Our proposed TempCD effectively captures the motion of the referent which is aligned with the language expression and enables accurate localization and segmentation of the specific duck. In contrast, Referformer fails to", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "locate the referent object, primarily attributed to a semantic incongruity between the expression and the frame-level queries. In the instance of (b), the deer's corresponding action appears in certain frames but lacks uniform presence. Our approach successfully captures its global semantics including motions and aligns it seamlessly with the referring expression, bridging the gap between local semantics and the referring expression via the Collection and Distribution mechanisms. Conversely, ReferFormer identifies the referent accurately in specific frames that correspond to the specified action, yet exhibits errors in others due to a noticeable lack of temporal context. All these observations collectively underscore the value of temporal interaction in fostering refined segmentation across frames. Besides, (c) illustrates that our method can still precisely locate and segment referent objects in complex scenes. The second frame presents a scenario of occlusion, wherein the referent is obscured by another bicycle. Our approach detects the lack of correspondence between visible entities in a frame and the specified referent sequence, facilitated by cross-frame interactions between the visible objects and those delineated by the referent sequence.", + "bbox": [ + 496, + 295, + 893, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 654, + 619, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes an end-to-end Temporal Query Collection and Distribution (TempCD) network for referring video object segmentation, which maintains object queries and the referent token and achieves alternating interaction between them via the proposed novel temporal collection-distribution mechanism.", + "bbox": [ + 496, + 684, + 890, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment: This work was supported by the National Natural Science Foundation of China (No.62206174), Shanghai Pujiang Program (No.21PJ1410900), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (ShangHAI), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University), and Shanghai Engineering Research Center of Intelligent Vision and Imaging.", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15473", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4985-4995, 2022. 1, 3, 4, 6, 7", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3, 4, 5", + "[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7454-7463, 2019. 3", + "[4] Yi-Wen Chen, Yi-Hsuan Tsai, Tiantian Wang, Yen-Yu Lin, and Ming-Hsuan Yang. Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748, 2019. 3", + "[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 3, 4", + "[6] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 3", + "[7] Zihan Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Jizhong Han, and Si Liu. Language-bridged spatial-temporal interaction for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4964–4973, 2022. 3, 7", + "[8] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 3", + "[9] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5958-5966, 2018. 1, 3, 6, 7", + "[10] Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. In Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., volume 2, pages 729-734. IEEE, 2005. 3", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 6", + "[12] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 1, 7" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4424-4433, 2020. 3", + "[14] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10488-10497, 2020. 3", + "[15] Tianrui Hui, Shaofei Huang, Si Liu, Zihan Ding, Guanbin Li, Wenguan Wang, Jizhong Han, and Fei Wang. Collaborative spatial-temporal modeling for language-queried video actor segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4187–4196, 2021. 1, 3, 7", + "[16] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 59-75. Springer, 2020. 3", + "[17] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. Advances in Neural Information Processing Systems, 34:13352-13363, 2021. 3", + "[18] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5", + "[19] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 6", + "[20] Yang Jiao, Zequn Jie, Weixin Luo, Jingjing Chen, Yu-Gang Jiang, Xiaolin Wei, and Lin Ma. Two-stage visual cues enhancement network for referring image segmentation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1331-1340, 2021. 3", + "[21] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tieniu Tan. Locate then segment: A strong pipeline for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9858-9867, 2021. 3", + "[22] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 1, 6", + "[23] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18154, 2022. 3", + "[24] Rajat Koner, Tanveer Hannan, Suprosanna Shit, Sahand Sharifzadeh, Matthias Schubert, Thomas Seidl, and Volker" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15474", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tresp. Instanceformer: An online video instance segmentation framework. arXiv preprint arXiv:2208.10547, 2022. 3", + "[25] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip HS Torr. Manigan: Text-guided image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7880-7889, 2020. 1", + "[26] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2018. 3", + "[27] Chen Liang, Yu Wu, Yawei Luo, and Yi Yang. Clawranenet: Leveraging object-level relation for text-based video segmentation. arXiv preprint arXiv:2103.10702, 2021. 7", + "[28] Chen Liang, Yu Wu, Tianfei Zhou, Wenguan Wang, Zongxin Yang, Yunchao Wei, and Yi Yang. Rethinking cross-modal interaction from a top-down perspective for referring video object segmentation. arXiv preprint arXiv:2106.01061, 2021. 1, 3", + "[29] Liang Lin, Pengxiang Yan, Xiaogqian Xu, Sibei Yang, Kun Zeng, and Guanbin Li. Structured attention network for referring image segmentation. IEEE Transactions on Multimedia, 24:1922-1932, 2021. 3", + "[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6", + "[31] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4761-4775, 2021. 7", + "[32] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4, 6", + "[33] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. arXiv preprint arXiv:2106.13230, 2021. 4, 6", + "[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[35] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020. 3", + "[36] Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016. 5", + "[37] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Ke Ning, Lingxi Xie, Fei Wu, and Qi Tian. Polar relative positional encoding for video-language segmentation. In *IJ-CAI*, volume 9, page 10, 2020. 1, 3, 7", + "[39] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 1", + "[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6", + "[41] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 3, 6, 7", + "[42] Cheng Shi and Sibei Yang. Spatial and visual perspectivetaking via view rotation and relation reasoning for embodied reference understanding. In European Conference on Computer Vision, pages 201-218. Springer, 2022. 3", + "[43] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 38-54, 2018. 3", + "[44] G Sreenu and Saleem Durai. Intelligent video surveillance: a review through deep learning techniques for crowd analysis. Journal of Big Data, 6(1):1-27, 2019. 1", + "[45] Jiajin Tang, Ge Zheng, Cheng Shi, and Sibei Yang. Contrastive grouping with transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23570-23580, 2023. 3", + "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[47] Hao Wang, Cheng Deng, Fan Ma, and Yi Yang. Context modulated dynamic networks for actor and action video segmentation with language queries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12152-12159, 2020. 1, 3, 7", + "[48] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3939-3948, 2019. 1, 3, 7", + "[49] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8741-8750, 2021. 2, 3, 7", + "[50] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip" + ], + "bbox": [ + 501, + 90, + 893, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15475", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "driven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 3", + "[51] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasescut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 3", + "[52] Dongming Wu, Xingping Dong, Ling Shao, and Jianbing Shen. Multi-level representation learning with semantic alignment for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4996-5005, 2022. 3, 4, 7", + "[53] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 1, 3, 4, 6, 7, 8", + "[54] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. arXiv preprint arXiv:2112.08275, 2021. 2, 3", + "[55] Chenliang Xu, Shao-Hang Hsieh, Caiming Xiong, and Jason J Corso. Can humans fly? action understanding with multiple classes of actors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2264-2273, 2015. 6", + "[56] Zhujun Xu and Damien Vivet. Instance sequence queries for video instance segmentation with transformers. Sensors, 21(13):4507, 2021. 3", + "[57] Sibei Yang, Guanbin Li, and Yizhou Yu. Cross-modal relationship inference for grounding referring expressions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4145-4154, 2019. 3", + "[58] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4644-4653, 2019. 3", + "[59] Sibei Yang, Guanbin Li, and Yizhou Yu. Propagating over phrase relations for one-stage visual grounding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 589-605. Springer, 2020. 3", + "[60] Sibei Yang, Guanbin Li, and Yizhou Yu. Relationship-embedded representation learning for grounding referring expressions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2765-2779, 2020. 3", + "[61] Shusheng Yang, Xinggang Wang, Yu Li, Yuxin Fang, Jiemin Fang, Wenyu Liu, Xun Zhao, and Ying Shan. Temporally efficient vision transformer for video instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2885-2895, 2022. 3, 4", + "[62] Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11266-11275, 2021. 1, 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 3", + "[64] Linwei Ye, Zhi Liu, and Yang Wang. Dual convolutional LSTM network for referring image segmentation. IEEE Transactions on Multimedia, 22(12):3224-3235, 2020. 3", + "[65] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 3, 7", + "[66] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1307-1315, 2018. 3", + "[67] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6", + "[68] Wangbo Zhao, Kai Wang, Xiangxiang Chu, Fuzhao Xue, Xinchao Wang, and Yang You. Modeling motion with multimodal features for text-based video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11737-11746, 2022. 3, 7", + "[69] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 4, 5" + ], + "bbox": [ + 501, + 92, + 890, + 555 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15476", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/3b08deeb-33eb-47b0-94a4-e0af89436d2d_model.json b/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/3b08deeb-33eb-47b0-94a4-e0af89436d2d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3affea00cc96108a3f4a9ad7f909d75fdb7d7766 --- /dev/null +++ b/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/3b08deeb-33eb-47b0-94a4-e0af89436d2d_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.086, + 0.131, + 0.885, + 0.154 + ], + "angle": 0, + "content": "Temporal Collection and Distribution for Referring Video Object Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.181, + 0.641, + 0.201 + ], + "angle": 0, + "content": "Jiajin Tang Ge Zheng Sibei Yang" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.206, + 0.775, + 0.224 + ], + "angle": 0, + "content": "School of Information Science and Technology, ShanghaiTech University" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.226, + 0.673, + 0.242 + ], + "angle": 0, + "content": "{tangjj,zhengge,yangsb}@shanghaiitech.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.245, + 0.634, + 0.259 + ], + "angle": 0, + "content": "https://toneyaya.github.io/tempcd" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.294, + 0.314, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.326, + 0.474, + 0.691 + ], + "angle": 0, + "content": "Referring video object segmentation aims to segment a referent throughout a video sequence according to a natural language expression. It requires aligning the natural language expression with the objects' motions and their dynamic associations at the global video level but segmenting objects at the frame level. To achieve this goal, we propose to simultaneously maintain a global referent token and a sequence of object queries, where the former is responsible for capturing video-level referent according to the language expression, while the latter serves to better locate and segment objects with each frame. Furthermore, to explicitly capture object motions and spatial-temporal cross-modal reasoning over objects, we propose a novel temporal collection-distribution mechanism for interacting between the global referent token and object queries. Specifically, the temporal collection mechanism collects global information for the referent token from object queries to the temporal motions to the language expression. In turn, the temporal distribution first distributes the referent token to the referent sequence across all frames and then performs efficient cross-frame reasoning between the referent sequence and object queries in every frame. Experimental results show that our method outperforms state-of-the-art methods on all benchmarks consistently and significantly." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.717, + 0.21, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Referring video object segmentation task (RVOS) aims to segment the target referent throughout a video sequence given a natural language expression [22]. It has attracted increasing attention from the academic community, as it is a basis for assessing a comprehensive understanding of visual, temporal and linguistic information. Meanwhile, ROVS benefits various downstream interactive applications such as language-driven human-robot interaction [39], video editing [25] and video surveillance [44]." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.293, + 0.885, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.367, + 0.884, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.442, + 0.882, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.542, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Figure 1. The comparison of (a) previous best-performing ReferFormer, (b) video-level decoding, and (c) ours referent-guided dynamic temporal decoding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Compared to referring image segmentation [12,62] that segments the target object in a single image mainly according to appearances and spatial and semantic relations, ROVS requires locating referents with temporal consistency according to object motions and dynamic associations among objects across frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Previous works mainly follow a single-stage bottom-up manner [9, 15, 38, 47, 48] or a two-stage top-down fashion [28]. The bottom-up methods early align the vision and language at the local patch level, failing to explicitly model objects and their relations [28, 53]. Although top-down approaches explicitly extract object tracklets and select object tracklets matched with the language expression, their two-stage pipeline is complex and less efficient [53]. Recently, transformer-based frameworks [1, 53] have been proposed to use object queries to capture objects and their associations in an end-to-end manner. They generate object queries and utilize the transformer decoder to search objects for these object queries in every frame, as shown in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.311, + 0.901 + ], + "angle": 0, + "content": "†Sibei Yang is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15466" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.892, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.279, + 0.895, + 0.35 + ], + "angle": 0, + "content": "Figure 2. Comparison of temporal modeling for previous works, naive attempt, and ours. Previous works (MTTR and ReferFormer) lack temporal modeling when decoding. Besides, object motion and interaction with other objects are crucial for referring video object segmentation. However, simply modeling temporal interaction with globally visible video-level decoding cannot achieve satisfactory results. Note that our alternating interaction between the global referent and object queries can dynamically capture spatial-temporal cross-modal reasoning over objects guided by the global referent information." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.366, + 0.189, + 0.381 + ], + "angle": 0, + "content": "Figure 1a and b." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.474, + 0.672 + ], + "angle": 0, + "content": "Although these end-to-end transformer-based frameworks have achieved impressive segmentation results, they still have limitations needed to be improved. First, their object queries are frame-level and independently responsible for the search for objects in each frame, which fails to capture temporal object motions and effective object tracking, as shown in Figure 2a and b. Therefore, we observe that they often fail to obtain temporally consistent predictions for target referents across an entire video, as shown in the left example in Figure 2. Second, their interactions between object queries are also solely performed per frame, leading to failure to model object-level spatial-temporal associations between objects. Therefore, they cannot correctly ground expressions that require cross-modal spatial-temporal reasoning over multiple objects. For example (see the right example in Figure 2), they mislocate the left camel as the target object because they lack to model the relation between the left camel in the current frame and the camel walking in the subsequent frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this paper, we aim to address these limitations by empowering the decoder with the capability of temporal modeling. A straightforward approach to decode objects with temporal information is to convert the frame-level decoding into video-level by inputting a sequence of object queries to search for objects in a sequence of frames like VisTR [49] for video instance segmentation, as shown in Figure 1b and 2e. However, this simple attempt fails to achieve satisfactory results while significantly increasing the computational cost and even underperforming existing query-based transformer frameworks. We further analyze the reasons: (1) The alignment between the natural language expression and the referent relies on overall objects' motions and temporal associations in the entire video. For example, the two camels can be distinguished only if the walking motion of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.895, + 0.563 + ], + "angle": 0, + "content": "the target camel is identified based on the entire video and aligned with the language expression. In contrast, the above naive attempt prematurely aligns language with fine-grained frame-level objects so that its attention is distracted from focusing accurately on the overall motions and relations of objects at the video level. (2) However, the precise localization and segmentation of target objects should go back to and rely more on every single frame because objects with different motions may cause them to have very different spatial locations in different frames. A similar observation is discussed by SeqFormer [54] for video instance segmentation, which suggests processing attention with each frame independently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Therefore, to address the challenge of aligning the language expression with objects' motions and temporal associations at the global video level but segmenting objects at the local frame level, we propose to maintain both local object queries and a global referent token. Global referent token captures the video-level referent information according to the language expression, while local object queries locate and segment objects with each frame. Furthermore, the object queries and the referent token are interacted to achieve spatial-temporal object information exchange through our well-designed temporal collection and distribution mechanisms, as shown in Figure 2f. Specifically, the temporal collection collects the referent information from object queries with spatial and visual attributes to temporal object motions to language semantics. In turn, the temporal distribution first dynamically distributes the referent token to every frame to extract the referent sequence. Then, the referent sequence interacts with object queries within each frame to achieve efficient spatial-temporal reasoning over objects. Note that the temporal collection and distribution alternately propagate information between object queries and the referent token to update each other." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15467" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "Finally, we propose a Temporal Collection and Distribution network (TempCD) which integrates our novel temporal collection and distribution mechanisms into the query-based transformer framework. Without using sequence matching or post-processing during inference like the previous methods [1,53], Our TempCD can directly predict the segmentation result of every frame based on object queries in the frame and the referent token." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.214, + 0.435, + 0.228 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows," + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.24, + 0.468, + 0.3 + ], + "angle": 0, + "content": "- We introduce to maintain a global referent token and a sequence of local object queries parallelly to bridge the gap between video-level object alignment with language and frame-level object segmentation." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.301, + 0.468, + 0.36 + ], + "angle": 0, + "content": "- We propose a novel collection-distribution mechanism for interacting between the referent token and object queries to capture spatial-temporal cross-modal reasoning over objects." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.361, + 0.468, + 0.42 + ], + "angle": 0, + "content": "- We present an end-to-end Temporal Query Collection and Distribution (TempCD) for RVOS, which can directly predict the segmentation referent of every frame without any post-processing and sequence matching." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.421, + 0.468, + 0.481 + ], + "angle": 0, + "content": "- Experiments on Ref-Youtu-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB-Sentences datasets show that TempCD outperforms state-of-the-art methods on all benchmarks consistently and significantly." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.24, + 0.468, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.497, + 0.22, + 0.513 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.523, + 0.473, + 0.825 + ], + "angle": 0, + "content": "Referring Image Segmentation (RIS) involves segmenting objects in images based on natural language expressions. Compared to Referring Video Object Segmentation (RVOS), RIS operates on individual images without temporal information. Previous works [3, 4, 26, 29, 42, 45, 57-59] focus on the joint modeling of vision and language. Various methods [6, 13, 14, 16, 20, 21, 35, 60, 64] are explored successively, such as using fusion operators like concatenation, ConvLSTM, attention mechanisms [46], and GNN [10] to obtain multimodal semantic feature maps. In addition, some studies attempt to decouple different components or key semantics from language, through explicit two-stage approaches [51, 66] or implicit attention modules [43, 62, 64, 65], to achieve a more fine-grained understanding. Recent models explore novel fusion techniques, including early fusion [8, 63], Linguistic Seed Encoder [23], and contrastive learning [50], to improve the performance of RIS. However, for RVOS, it requires not only the multimodal integration of language and image frames but also temporal modeling at the video level." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Referring Video Object Segmentation (RVOS) aims to segment the target object described in natural language from videos. Previous works in RVOS has primarily used two frameworks: bottom-up and top-down. The top-down framework [28] directly models the motion informa" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.453 + ], + "angle": 0, + "content": "tion between segmentation masks of consecutive frames. However, this two-stage approach, which includes complex computational costs, is limited in terms of video input length. For the bottom-up framework, some approaches [9, 15, 38, 47, 48] directly apply RIS methods to construct multi-modal feature maps for referring segmentation of keyframes. They rely solely on replacing the traditional image backbone with a 3D temporal backbone for temporal modeling, which limits the performance of multiframes segmentation. To address this issue, URVOS [41] extends previous methods by building a memory bank that propagates language referent object information in the temporal dimension. Furthermore, URVOS introduces the RefYoutube-VOS [41] dataset, which provides segmentation annotations for each frame that require more efficient temporal modeling. In addition, recent works [7, 52, 68] adopt language guided fusion between temporal features and visual features to obtain more efficient temporal modeling. Recently, query-based methods [1, 53] with a transformer encoder-decoder frameworks [2, 5, 69] attempt to capture object-level information. They achieve higher performance by constructing a multimodal feature map of visual and linguistic information, or by using language-initialized queries with independently decoding with queries for each frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.817 + ], + "angle": 0, + "content": "Temporal Modeling in Video Instance Segmentation (VIS). VIS aims to simultaneously track, segment and classify interest instances in a video. Similarly to RVOS, VIS requires obtaining instance information from every frame in the video and maintaining the temporal consistency of instances. Recent works focus on spatial and temporal modeling based on instance-level information. VisTR [49] employs an encoder-decoder transformer based on DETR, with a concatenated sequence of instance queries as the input to enable spatial-temporal interactions. The following IFC [17] and TeViT [61] explored a more efficient approach to temporal modeling in the encoder by introducing supplementary tokens for interaction across frames, with the decoder aligned with VisTR. Other following works [24, 56] take advantage of memories or queries from several previous frames to help instance segmentation in the current frame, extending to a online fashion. Recently, SeqFormer [54] proposes that the acquisition of frame-level instance information such as position needs to go back to each frame, i.e., the decoding process should be done independently for each frame. While these methods provide potential solutions for temporal modeling, the direct application to RVOS is not feasible, because of inconsistent object semantics over frames." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.593, + 0.846 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The framework of our proposed TempCD is shown in Figure 3. We introduce a sequence of object queries to capture object information in every frame independently and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15468" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.088, + 0.852, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.389, + 0.895, + 0.448 + ], + "angle": 0, + "content": "Figure 3. Overall framework of TempCD. TempCD simultaneously maintains a global referent token and a sequence of object queries, and they interact via the proposed collection-distribution mechanism. Specifically, the temporal collection collects global information for the referent token from object queries to the temporal motions to the language expression. In turn, the temporal distribution first distributes the referent token to the referent sequence and then performs efficient cross-frame reasoning between the referent sequence and object queries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.471, + 0.553 + ], + "angle": 0, + "content": "maintain a referent token to capture the global referent information for aligning with language expression. The referent token and object queries are alternately updated to stepwise identify the target object and better segment the target object in every frame through the proposed temporal collection and distribution mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.782 + ], + "angle": 0, + "content": "Specifically, we first introduce encoders and the definition of object queries in Section 3.1. Then, we present the temporal collection that collects the referent information for the referent token from object queries to temporal object motions to the language expression (see Section 3.2). Next, we introduce the temporal distribution mechanism that distributes the global referent information to the referent sequence across frames to object queries in each frame (see Section 3.3). The collection-distribution mechanism explicitly captures object motions and spatial-temporal cross-modal reasoning over objects. Finally, we introduce segmentation heads and loss functions in Section 3.4. Note that as we explicitly maintain the referent token, we can directly identify the referent object in each frame without requiring sequence matching as in previous queries-based methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.353, + 0.803 + ], + "angle": 0, + "content": "3.1. Encoders and Query Definition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Language Encoder. Following previous work [53], we adopt RoBERTa [32] as our language encoder. Given a language expression with \\( L \\) words, we extract its sentence feature \\( \\mathcal{F}_s \\in \\mathbb{R}^C \\), where \\( C \\) denotes channel dimensions. In addition, we obtain the output feature representation of each word before the last pooling layer as \\( \\mathcal{F}_w \\in \\mathbb{R}^{L \\times C} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.599 + ], + "angle": 0, + "content": "Visual Backbone and Encoder. Following previous works [1, 52, 53, 61], we use ResNet-50 [11] as our visual backbone. We also follow recent work [1, 53] to evaluate our model using the temporal visual backbone network Video-Swin [33] in experiments. For an input video \\( V \\in \\mathbb{R}^{H \\times W \\times 3 \\times T} \\) with \\( T \\) frames, we first extract its visual feature maps and further construct multimodal feature maps \\( \\mathcal{V} = \\{v_i\\}_{i=1}^T \\) by fusing the visual feature maps with the word features of the expression \\( \\mathcal{F}_w \\) following [53]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Object Query Definition. Inspired by query-based detection and segmentation frameworks [2, 5, 69], we leverage object queries to represent object-level information. For the video with \\(\\mathcal{T}\\) frames, we introduce a sequence of \\(\\mathcal{T} \\times N\\) object queries to represent the objects in every frame and follow [53] to use the language feature \\(\\mathcal{F}_s\\) to initialize the object queries as \\(\\mathcal{Q} \\in \\mathbb{R}^{\\mathcal{T} \\times N \\times C}\\), where the number of objects and the dimension of channels are \\(N\\) and \\(C\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Referent Token Definition. In addition to adopting the object queries to represent objects in every frame, we define a referent token \\(\\mathcal{Q}^g\\in \\mathbb{R}^C\\) to represent the global information of the target object. As the language expression naturally indicates the referent information, we thus use the language feature \\(\\mathcal{F}_s\\) to initialize the \\(\\mathcal{Q}^g\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.857, + 0.85 + ], + "angle": 0, + "content": "3.2. Temporal Collection from Object Queries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The temporal collection mechanism collects object motions from frames and aggregates object motions consistent with the language expression into the global referent token." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15469" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "We first perform motion collection across object queries (see Section 3.2.1) and then update the global referent token from the object motions and language expression (see Section 3.2.2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.169, + 0.414, + 0.185 + ], + "angle": 0, + "content": "3.2.1 Motion Collection across Object Queries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.193, + 0.471, + 0.36 + ], + "angle": 0, + "content": "For simplicity of demonstration, we denote the object queries output from the \\((l - 1)\\)-th layer as \\(\\mathcal{Q}_{l - 1}\\) and input them into the \\(l\\)-th layer of TempCD. Note that we input the object queries \\(\\mathcal{Q}\\) defined in Section 3.1 into the first layer of TempCD. We utilize \\(\\mathcal{Q}_{l - 1}\\) to capture object information in each frame and integrate them to obtain temporal information such as object motion and action. Specifically, for \\(i\\)-th frame, we feed its object queries \\(q_{i} = \\mathcal{Q}_{l - 1}^{i} \\in \\mathbb{R}^{N \\times C}\\) and the multimodal features \\(v_{i}\\) into the DETR transformer decoder [2,69] to locate objects, where \\(\\mathcal{Q}_{l - 1}^{i}\\) means the object queries of \\(i\\)-th frame. The computation proceeds as follows," + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.372, + 0.469, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {Q}} _ {l} = \\left\\{\\operatorname {D E T R} \\left(q _ {i}, v _ {i}\\right) \\right\\} _ {i = 1} ^ {\\mathcal {T}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.469, + 0.438 + ], + "angle": 0, + "content": "where \\(\\mathrm{DETR}(\\cdot ,\\cdot)\\) represents the DETR transformer decoder, and \\(\\hat{\\mathcal{Q}}_l\\) denotes the object queries output by the decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.576 + ], + "angle": 0, + "content": "Next, to aggregate object motions across the video sequence, we combine objects information from queries of each frame. Specifically, we generate temporal weights \\( G \\in \\mathbb{R}^{T \\times N \\times 1} \\) for each frame given the object queries \\( \\hat{\\mathcal{Q}}_l \\). This is done by employing a nonlinear layer followed by a softmax operator. Then, we use the temporal weights \\( G \\) to integrate the frame-level information and obtain object motions \\( M \\in \\mathbb{R}^{N \\times C} \\) through a gating mechanism. The integration is computed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.582, + 0.349, + 0.6 + ], + "angle": 0, + "content": "\\[\nG = \\operatorname {s o f t m a x} (\\hat {\\mathcal {Q}} _ {t} W),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.603, + 0.469, + 0.645 + ], + "angle": 0, + "content": "\\[\nM = \\sum_ {i = 1} ^ {\\mathcal {T}} \\left(G _ {i} \\odot \\hat {\\mathcal {Q}} _ {l} ^ {i}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.469, + 0.713 + ], + "angle": 0, + "content": "where \\(W \\in \\mathbb{R}^{C \\times 1}\\) denotes the learnable weights of the linear layer, \\(G_{i} \\in \\mathbb{R}^{N \\times 1}\\) is the weight for \\(i\\)-th frames in \\(G\\), \\(\\mathrm{softmax}(\\cdot)\\) represents the softmax operation along the temporal dimension, and \\(\\odot\\) is the element-wise product." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.731, + 0.364, + 0.744 + ], + "angle": 0, + "content": "3.2.2 Global Referent Token Collection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.753, + 0.469, + 0.877 + ], + "angle": 0, + "content": "Next, we update the global referent token \\(\\mathcal{Q}_{l-1}^g \\in \\mathbb{R}^C\\) at the \\((l-1)\\)-th layer to referent token \\(\\mathcal{Q}_l^g \\in \\mathbb{R}^C\\) by selecting the object motions \\(M \\in \\mathbb{R}^{N \\times C}\\) consistent with the language feature \\(\\mathcal{F}_s \\in \\mathbb{R}^C\\). Note that we input the global referent token \\(\\mathcal{Q}^g\\) defined in Section 3.1 into the first layer of TempCD. We first calculate the cosine similarity between the language feature \\(\\mathcal{F}_s\\) and the object motions \\(M\\). We obtain the similarity scores \\(S = \\{s_j\\}_{j=1}^N\\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.885, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\ns _ {j} = \\cos \\left(\\mathcal {F} _ {s}, m _ {j}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot ,\\cdot)\\) means the cosine similarity function and \\(m_j\\) denotes the \\(j\\)-th object's motion feature in \\(M\\). Next, we update the global referent token \\(\\mathcal{Q}_g\\) as the motion feature of the object with the highest similarity score to the global referent token. To ensure the selection process is differentiable, we implement this selection through the Gumbel Softmax operator [18, 36]. The calculation is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.209, + 0.785, + 0.226 + ], + "angle": 0, + "content": "\\[\nS _ {\\text {g u m b e l}} = \\operatorname {s o f t m a x} \\left(\\left(S + G _ {s}\\right) / \\tau\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.228, + 0.891, + 0.244 + ], + "angle": 0, + "content": "\\[\nS _ {\\text {o n e h o t}} = \\text {o n e h o t} \\left(\\arg \\max _ {N} \\left(S _ {\\text {g u m b e l}}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.246, + 0.832, + 0.263 + ], + "angle": 0, + "content": "\\[\nS _ {\\text {r e f e r e n t}} = S _ {\\text {o n e h o t}} - \\mathrm {s g} \\left[ S _ {\\text {g u m b e l}} \\right] + S _ {\\text {g u m b e l}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.892, + 0.426 + ], + "angle": 0, + "content": "where \\(\\operatorname{softmax}(\\cdot)\\) represents the softmax operation along dimension \\(N\\), \\(G_{s} \\in \\mathbb{R}^{N \\times 1}\\) is sampled from the Gumbel(0,1) distribution, \\(\\tau\\) is the temperature coefficient hyperparameter, and \\(\\mathrm{sg}\\) is the stop gradient operation. Additionally, \\(\\operatorname{argmax}_{N}\\) represents choosing the index with the highest similarity score, and onehot represents setting the highest-scoring item to 1 based on the selected index and setting the others to 0. The one-hot vector \\(S_{\\text{referent}} \\in \\mathbb{R}^{N \\times 1}\\) indicates the selection from the object motions. The global referent token \\(\\mathcal{Q}^g \\in \\mathbb{R}^C\\) is updated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.435, + 0.891, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} _ {l} ^ {g} = \\operatorname {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(S _ {\\text {r e f e r e n t}} ^ {\\top} M\\right) + Q _ {l - 1} ^ {g}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.891, + 0.509 + ], + "angle": 0, + "content": "where \\(\\mathrm{MLP}(\\cdot)\\) denotes a three-layer linear layer with activation function, and LayerNorm \\((\\cdot)\\) is the same as the LayerNorm in standard Transformer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.891, + 0.584 + ], + "angle": 0, + "content": "We obtain the updated global referent token \\(\\mathcal{Q}^g\\) that contains information on the referent object aligned with the language expression. Next, we feed it into the Temporal Distribution module to assist in distributing the referent information into each frame." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.595, + 0.872, + 0.611 + ], + "angle": 0, + "content": "3.3. Temporal Distribution from Referent Token" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.739 + ], + "angle": 0, + "content": "The temporal distribution aims to propagate the referent information to all frames and perform dynamically spatial-temporal reasoning over objects based on the referent information. First, we distribute the global referent token to each frame independently to extract the referent sequence across all frames (see Section 3.3.1). Next, we perform the efficient cross-frame reasoning between the referent sequence and object queries in every frame (see Section 3.3.2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.802, + 0.771 + ], + "angle": 0, + "content": "3.3.1 Global Referent Token Distribution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.891, + 0.854 + ], + "angle": 0, + "content": "Given the global referent token \\(\\mathcal{Q}_l^g\\), we distribute the referent object information to every frame and further obtain the referent sequence across all frames, denoted by \\(\\mathcal{R}_l = \\{r_i\\}_{i=1}^T\\), which refers to the target object in each frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Specifically, we first compute the cosine similarity scores between the object queries \\(\\hat{\\mathcal{Q}}_l = \\{q_i\\}_{i=1}^T\\) and the global referent token \\(\\mathcal{Q}_l^g\\). Then we obtain the index \\(d_i\\) of the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15470" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "object query with the highest similarity score to the global referent token for every frame \\(i\\) independently. The computation is as follows," + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.148, + 0.469, + 0.172 + ], + "angle": 0, + "content": "\\[\ns _ {i, j} = \\cos \\left(q _ {i, j}, Q _ {l} ^ {g}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.168, + 0.345, + 0.185 + ], + "angle": 0, + "content": "\\[\nd _ {i} = \\operatorname {a r g m a x} _ {j} \\left\\{s _ {i, j} \\right\\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.189, + 0.469, + 0.267 + ], + "angle": 0, + "content": "where \\(q_{i,j}\\) is the \\(j\\)-th object query in the \\(i\\)-th frame, and \\(s_{i,j}\\) is the corresponding similarity score. For every frame \\(i\\), we select the object query \\(d_i\\) with the highest similarity score and fuse it with the global referent token \\(\\mathcal{Q}_l^g\\) to generate the referent sequence \\(\\mathcal{R}_l = \\{r_i\\}_{i=1}^T\\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.275, + 0.469, + 0.294 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {l} = \\left\\{\\mathrm {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(q _ {i, d _ {i}} + \\mathcal {Q} _ {l} ^ {g}\\right)\\right) \\right\\} _ {i = 1} ^ {\\mathcal {T}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.469, + 0.351 + ], + "angle": 0, + "content": "We end up with referent sequence \\(\\mathcal{R}_l\\in \\mathbb{R}^{T\\times C}\\) across all frames, which is used as crucial objects to interact with object queries in every frame independently." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.368, + 0.469, + 0.383 + ], + "angle": 0, + "content": "3.3.2 Spatial-temporal Reasoning over Object Queries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.469, + 0.48 + ], + "angle": 0, + "content": "We leverage the referent sequence \\(\\mathcal{R}_l\\in \\mathbb{R}^{T\\times C}\\), which encodes the referent object information across all frames, to interact with the object queries to achieve the cross-frame temporal reasoning over multiple objects and further distribute referent information to all object queries in all frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.469, + 0.572 + ], + "angle": 0, + "content": "Specifically, for each object query \\(q_{i,j}\\) at the \\(i\\)-th frame, we concatenate it with the objects of the referent sequence in all other frames to construct the cross-object query sequence \\(C_{i,j} = \\{r_1,..,r_{i - 1},q_{i,j},r_{i + 1},..,r_{\\mathcal{T}}\\} \\in \\mathbb{R}^{\\mathcal{T}\\times C}\\). The object-level interaction across frames is implemented by multi-head self attention, which is computed as follows," + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.587, + 0.469, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\hat {C} _ {i, j} = \\operatorname {M H S A} \\left(C _ {i, j}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.469, + 0.717 + ], + "angle": 0, + "content": "where \\(\\mathrm{MHSA}(\\cdot)\\) denotes the multi-head self-attention mechanism. The \\(j\\)-th updated object queries in \\(C_{i,j}^{\\hat{}}\\) for every frame \\(i\\) are concatenated as \\(\\mathcal{Q}_l^f \\in \\mathbb{R}^{\\mathcal{T} \\times N \\times C}\\) to input to next decoder layer. Note that our cross-frame interaction only requires interaction with referent sequences to object queries in each frame, which achieves efficient cross-frame interaction over multiple objects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.318, + 0.74 + ], + "angle": 0, + "content": "3.4. Prediction Heads and Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.809 + ], + "angle": 0, + "content": "We integrate the referent sequence and object queries in each decoder layer to the prediction heads for predicting results. To simplify the presentation, we use the final layer for details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We extract the referent sequence \\(\\mathcal{R} \\in \\mathbb{R}^{T \\times C}\\) and combine them with the object queries \\(Q_{l}\\) to obtain the output referent sequence \\(\\hat{\\mathcal{R}}\\) for prediction. Specifically, we use a gating mechanism to integrate the information from \\(\\hat{C}\\) into video-level features, which is similar to Eq 2. Next, we calculate cosine similarity between each sequence \\(C_{j,i}\\) and the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "language sentence feature \\(\\mathcal{F}_s\\), which is similar to Eq 3, to obtain the score for each object \\(q_{i,j}\\). We select the highest-scoring object \\(q_i^r\\) in each frame and combine it with \\(\\mathcal{R}\\) to obtain the final referent sequence \\(\\hat{\\mathcal{R}} \\in \\mathbb{R}^{T \\times C}\\) of the referred object in each frame:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.18, + 0.892, + 0.198 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {R}} = \\left\\{\\mathrm {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(r _ {i} + q _ {i} ^ {r}\\right)\\right) \\right\\}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.892, + 0.299 + ], + "angle": 0, + "content": "Finally, we feed \\(\\hat{\\mathcal{R}}\\) into the segmentation head to predict the segmentation results. Additionally, similar to [53] and [1], we use a box prediction head to regress the box coordinates for each query. We utilize Dice [37] loss and Focal [30] loss as the segmentation loss and GIOU [40] loss and L1 loss as the box loss." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.314, + 0.625, + 0.331 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.52 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on four benchmark datasets that are publicly available: Ref-Youtu-VOS [41], Ref-Davis-2017 [22], A2D-Sentences [9], and JHMDB-Sentences [9]. Ref-Youtu-VOS consists of 3978 videos and approximately 15K language descriptions. Ref-Davis-2017 contains approximately 90 videos. A2D-Sentences and JHMDB-Sentences [19, 55], originally focused on action recognition, have been expanded to incorporate language expression annotations. This expansion has yielded a total of 3782 videos for A2D-Sentences and 928 videos for JHMDB-Sentences, each accompanied by corresponding language descriptions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.686 + ], + "angle": 0, + "content": "Implementation Details. Following [53], we pre-train our model on RefCOCO dataset [67]. We utilize a text encoder derived from RoBERTa [32], coupled with either ResNet50 [11] or Video-Swin serving [33] as our visual backbone. We train our model for 6 epochs with an initial learning rate of 1e-4 and the AdamW [34] optimizer. Consistent with prior works [1, 41, 53], we use the \\(\\mathcal{I}\\), \\(\\mathcal{F}\\), and \\(\\mathcal{J} \\& \\mathcal{F}\\) metrics for evaluation on Ref-Youtu-VOS and Ref-Davis-2017 datasets, and Overall IoU and Mean IoU as evaluation metrics on A2D-Sentences and JHMDB-Sentences datasets." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.697, + 0.874, + 0.712 + ], + "angle": 0, + "content": "4.1. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.779 + ], + "angle": 0, + "content": "As shown in Table 1 and Table 2, we compare TempCD with state-of-the-art methods on four benchmarks. TempCD consistently outperforms state-of-the-art methods on all datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison of Ref-Youtube-VOS and Ref-Davis-2017 Datasets. Results for the Ref-Youtu-VOS and Ref-Davis-2017 datasets are presented in Table 1. With a standard ResNet-50 [11] visual backbone, our model exhibits improvements of \\(2.7\\%\\), \\(4\\%\\), and \\(3.4\\%\\) for the \\(\\mathcal{I}, \\mathcal{F}\\), and \\(\\mathcal{J} \\& \\mathcal{F}\\) metrics, respectively, on the Ref-YoutuVOS dataset. Employing the advanced temporal visual backbone, VideoSwin-B [33], our approach consistently outperforms the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15471" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.161, + 0.09, + 0.812, + 0.292 + ], + "angle": 0, + "content": "
BackboneMethodRef-Youtube-VOSRef-Davis17
JFJ&FJFJ&F
ResNet-50CMSA [65]33.336.534.932.237.234.7
CMSA+RNN [65]34.838.136.436.943.540.2
URVOS [41]45.349.247.247.356.051.5
MLRL [52]48.451.049.750.155.452.7
LBDT [7]48.250.649.4--54.5
ReferFormer [53]54.856.555.655.861.358.5
Ours57.560.559.057.362.760.0
Video-Swin-TMTTR [1]54.056.655.3---
ReferFormer [53]58.060.959.4---
Ours60.564.062.359.365.062.2
Video-Swin-BReferFormer [53]61.364.662.958.164.161.1
Ours63.668.065.861.667.664.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.153, + 0.301, + 0.816, + 0.316 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art models for RVOS on Ref-Youtu-VOS and Ref-Davis-2017 datasets." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.33, + 0.459, + 0.493 + ], + "angle": 0, + "content": "
BackboneMethodA2D-SJHMDB-S
oIoUmIoUoIoUmIoU
VGG16Hu et al. [12]47.435.054.652.8
I3DGavrilyuk et al. [9]53.642.154.154.2
ACAN [48]60.149.075.656.4
CMPC-V [31]65.357.361.661.7
Resnet-50ClawCraneNet [27]63.159.964.465.6
MMMMTBVS [68]67.355.861.961.3
LBDT [7]70.462.164.565.8
Ours76.668.670.669.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.503, + 0.47, + 0.531 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-art models for RVOS on A2D-Sentences (A2D-S) and JHMDB-Sentences (JHMDB-S)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.469, + 0.623 + ], + "angle": 0, + "content": "previous state-of-the-art model [53] on the Ref-Davis-2017 dataset, achieving a \\(3.5\\%\\) increase across the aforementioned metrics. Note that evaluations on the RefDavis-2017 dataset are performed using models trained on the RefYoutube-VOS dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.469, + 0.838 + ], + "angle": 0, + "content": "Furthermore, we present a thorough evaluation of the effectiveness of our TempCD through subsequent comparisons: (1) We compare our model to MTTR [1], a pioneering method that adopts a query-based framework for RVOS, despite lacking temporal modeling in both the encoder and decoder stages. Our TempCD surpasses MTTR, achieving a \\(7\\%\\) improvement in the \\(\\mathcal{J}\\& \\mathcal{F}\\) metric on the Ref-YoutuVOS dataset. This outcome underscores the heightened efficacy of our introduced temporal decoding modules. (2) Compared with LBDT [7], our TempCD achieves a significant \\(9.6\\%\\) enhancement in the \\(\\mathcal{J}\\& \\mathcal{F}\\) metric. This result suggests that, unlike [7, 41, 52] that integrate temporal and visual features via a bottom-up framework, our proposed TempCD achieves more efficient temporal modeling." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Comparison on A2D-Sentences and JHMDB-Sentences. As presented in Table 2, our approach yields mean improvements of \\(6.2\\%\\) in Overall IoU and \\(5.2\\%\\) in Mean IoU, surpassing the predominant state-of-the-art method [7] for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.456 + ], + "angle": 0, + "content": "these two datasets. In contrast to Ref-Youtu-VOS, these datasets predominantly comprise segmentation annotations for keyframes that encapsulate actions. When compared to prior studies [9, 15, 38, 47, 48] that adopt temporal encoders to process video temporal information, our method consistently attains enhancements across these two datasets. This highlights the significance of explicitly capturing and integrating video-level temporal context into individual frames." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.467, + 0.654, + 0.483 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.551 + ], + "angle": 0, + "content": "To validate the impact of modules in our model, we evaluate six variants of our TempCD on Ref-Youtu-VOS dataset. The results are shown in Table 3. All experiments are performed with Video-Swin-B as the visual backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.793 + ], + "angle": 0, + "content": "Baselines. (1) Our baseline model employs \\(\\mathcal{T}\\) sets of frame-level object queries (\"Local Query\"). These queries are concatenated sequentially and fed into the decoder, following a similar approach to VisTR [49]. The derived results (53.7% \\(\\mathcal{J}\\), 56.3% \\(\\mathcal{F}\\), and 55.0% \\(\\mathcal{J} \\& \\mathcal{F}\\)) suggest a misalignment between the natural language representations and the referent semantics at the individual frame level. (2) Another subsequent baseline is established by integrating a singular shared set of video-level object queries across all frames (\"Global Query\"). The global queries are performed an inner product with the feature map \\(v\\) of each frame, yielding segmentation outcomes. While this approach intrinsically aligns with language representations and achieves consistency between frames, it struggles to adapt to frame-level variability, resulting in a performance decline relative to the first baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Query Collection and Distribution. (3) A natural enhancement for baseline (1) entails integrating cross-frame temporal attention on a referent sequence, which is selected from local queries by the similarity with language. This modification leads to a performance improvement of \\(3.4\\%\\). (4) Another endeavor to enhance the baseline (1) involves the introduction of an additional set of global queries. These" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15472" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.089, + 0.891, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.264, + 0.794, + 0.279 + ], + "angle": 0, + "content": "Figure 4. Visualization for segmentation results of previous best-performing ReferFormer [53] and ours." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.293, + 0.462, + 0.42 + ], + "angle": 0, + "content": "
MethodJFJ&F
1Local Query53.756.355.0
2Global Query52.955.554.2
31+Cross-Frame56.259.157.6
41+Motion Collection58.161.559.8
53+460.764.462.6
65+Gumble-Softmax61.465.463.4
7Ours63.668.065.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.43, + 0.434, + 0.444 + ], + "angle": 0, + "content": "Table 3. Ablation Study on Ref-Youtu-VOS dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.473, + 0.76 + ], + "angle": 0, + "content": "queries collect motion information from local queries and serve to distribute global context to local queries. Specifically, we utilize a Top-K strategy to select global information based on the similarity score with language. This facilitates implicit frame level interaction and, significantly, fosters semantic alignment between global queries and language expressions. This alignment contributes to a marked performance gain of \\(5.6\\%\\). (5) We further incorporate both (3)'s explicit cross-frame temporal modeling and (4)'s motion collection and distribution, which delivers a \\(2.8\\%\\) improvement. (6) We replace the Top-K selection, which does not allow gradient calculation, with Gumbel Softmax, improving \\(\\mathcal{J} \\& \\mathcal{F}\\) by \\(0.8\\%\\). (7) Our complete model further enhances the vanilla cross-frame temporal attention mechanism. Instead of solely updating the referent sequence via temporal interaction, local object queries are also extended to incorporate temporal context information from referent sequence, facilitating cross-frame interaction over multiple objects. The full decoder of our proposed method achieves the performance of \\(65.8\\%\\) in terms of \\(\\mathcal{J} \\& \\mathcal{F}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.216, + 0.787 + ], + "angle": 0, + "content": "4.3. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Figure 4 visualizes several qualitative results. The referring expression in (a) describes the motion of a white duck, a pivotal feature that distinguishes it from similar ducks. Our proposed TempCD effectively captures the motion of the referent which is aligned with the language expression and enables accurate localization and segmentation of the specific duck. In contrast, Referformer fails to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.895, + 0.629 + ], + "angle": 0, + "content": "locate the referent object, primarily attributed to a semantic incongruity between the expression and the frame-level queries. In the instance of (b), the deer's corresponding action appears in certain frames but lacks uniform presence. Our approach successfully captures its global semantics including motions and aligns it seamlessly with the referring expression, bridging the gap between local semantics and the referring expression via the Collection and Distribution mechanisms. Conversely, ReferFormer identifies the referent accurately in specific frames that correspond to the specified action, yet exhibits errors in others due to a noticeable lack of temporal context. All these observations collectively underscore the value of temporal interaction in fostering refined segmentation across frames. Besides, (c) illustrates that our method can still precisely locate and segment referent objects in complex scenes. The second frame presents a scenario of occlusion, wherein the referent is obscured by another bicycle. Our approach detects the lack of correspondence between visible entities in a frame and the specified referent sequence, facilitated by cross-frame interactions between the visible objects and those delineated by the referent sequence." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.655, + 0.62, + 0.672 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.775 + ], + "angle": 0, + "content": "This paper proposes an end-to-end Temporal Query Collection and Distribution (TempCD) network for referring video object segmentation, which maintains object queries and the referent token and achieves alternating interaction between them via the proposed novel temporal collection-distribution mechanism." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Acknowledgment: This work was supported by the National Natural Science Foundation of China (No.62206174), Shanghai Pujiang Program (No.21PJ1410900), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (ShangHAI), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University), and Shanghai Engineering Research Center of Intelligent Vision and Imaging." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15473" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4985-4995, 2022. 1, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7454-7463, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.394 + ], + "angle": 0, + "content": "[4] Yi-Wen Chen, Yi-Hsuan Tsai, Tiantian Wang, Yen-Yu Lin, and Ming-Hsuan Yang. Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.45 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.52 + ], + "angle": 0, + "content": "[6] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[7] Zihan Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Jizhong Han, and Si Liu. Language-bridged spatial-temporal interaction for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4964–4973, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.471, + 0.662 + ], + "angle": 0, + "content": "[8] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.664, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[9] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5958-5966, 2018. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[10] Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. In Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., volume 2, pages 729-734. IEEE, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.833, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 1, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[13] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4424-4433, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[14] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10488-10497, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.894, + 0.33 + ], + "angle": 0, + "content": "[15] Tianrui Hui, Shaofei Huang, Si Liu, Zihan Ding, Guanbin Li, Wenguan Wang, Jizhong Han, and Fei Wang. Collaborative spatial-temporal modeling for language-queried video actor segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4187–4196, 2021. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.894, + 0.416 + ], + "angle": 0, + "content": "[16] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 59-75. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.893, + 0.473 + ], + "angle": 0, + "content": "[17] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. Advances in Neural Information Processing Systems, 34:13352-13363, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.893, + 0.516 + ], + "angle": 0, + "content": "[18] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.893, + 0.574 + ], + "angle": 0, + "content": "[19] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.893, + 0.645 + ], + "angle": 0, + "content": "[20] Yang Jiao, Zequn Jie, Weixin Luo, Jingjing Chen, Yu-Gang Jiang, Xiaolin Wei, and Lin Ma. Two-stage visual cues enhancement network for referring image segmentation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1331-1340, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.893, + 0.715 + ], + "angle": 0, + "content": "[21] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tieniu Tan. Locate then segment: A strong pipeline for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9858-9867, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.894, + 0.799 + ], + "angle": 0, + "content": "[22] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[23] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18154, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[24] Rajat Koner, Tanveer Hannan, Suprosanna Shit, Sahand Sharifzadeh, Matthias Schubert, Thomas Seidl, and Volker" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15474" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "Tresp. Instanceformer: An online video instance segmentation framework. arXiv preprint arXiv:2208.10547, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[25] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip HS Torr. Manigan: Text-guided image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7880-7889, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.47, + 0.26 + ], + "angle": 0, + "content": "[26] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[27] Chen Liang, Yu Wu, Yawei Luo, and Yi Yang. Clawranenet: Leveraging object-level relation for text-based video segmentation. arXiv preprint arXiv:2103.10702, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[28] Chen Liang, Yu Wu, Tianfei Zhou, Wenguan Wang, Zongxin Yang, Yunchao Wei, and Yi Yang. Rethinking cross-modal interaction from a top-down perspective for referring video object segmentation. arXiv preprint arXiv:2106.01061, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.47, + 0.426 + ], + "angle": 0, + "content": "[29] Liang Lin, Pengxiang Yan, Xiaogqian Xu, Sibei Yang, Kun Zeng, and Guanbin Li. Structured attention network for referring image segmentation. IEEE Transactions on Multimedia, 24:1922-1932, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.47, + 0.482 + ], + "angle": 0, + "content": "[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.482, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[31] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4761-4775, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.47, + 0.621 + ], + "angle": 0, + "content": "[32] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.664 + ], + "angle": 0, + "content": "[33] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. arXiv preprint arXiv:2106.13230, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.692 + ], + "angle": 0, + "content": "[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[35] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[36] Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[37] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.134 + ], + "angle": 0, + "content": "[38] Ke Ning, Lingxi Xie, Fei Wu, and Qi Tian. Polar relative positional encoding for video-language segmentation. In *IJ-CAI*, volume 9, page 10, 2020. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[39] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.893, + 0.3 + ], + "angle": 0, + "content": "[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.301, + 0.893, + 0.385 + ], + "angle": 0, + "content": "[41] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.893, + 0.441 + ], + "angle": 0, + "content": "[42] Cheng Shi and Sibei Yang. Spatial and visual perspectivetaking via view rotation and relation reasoning for embodied reference understanding. In European Conference on Computer Vision, pages 201-218. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.893, + 0.496 + ], + "angle": 0, + "content": "[43] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 38-54, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.496, + 0.893, + 0.538 + ], + "angle": 0, + "content": "[44] G Sreenu and Saleem Durai. Intelligent video surveillance: a review through deep learning techniques for crowd analysis. Journal of Big Data, 6(1):1-27, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.893, + 0.607 + ], + "angle": 0, + "content": "[45] Jiajin Tang, Ge Zheng, Cheng Shi, and Sibei Yang. Contrastive grouping with transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23570-23580, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.608, + 0.893, + 0.664 + ], + "angle": 0, + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.893, + 0.733 + ], + "angle": 0, + "content": "[47] Hao Wang, Cheng Deng, Fan Ma, and Yi Yang. Context modulated dynamic networks for actor and action video segmentation with language queries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12152-12159, 2020. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.893, + 0.803 + ], + "angle": 0, + "content": "[48] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3939-3948, 2019. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.893, + 0.873 + ], + "angle": 0, + "content": "[49] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8741-8750, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.893, + 0.903 + ], + "angle": 0, + "content": "[50] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15475" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "driven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[51] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasescut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[52] Dongming Wu, Xingping Dong, Ling Shao, and Jianbing Shen. Multi-level representation learning with semantic alignment for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4996-5005, 2022. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[53] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 1, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[54] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. arXiv preprint arXiv:2112.08275, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[55] Chenliang Xu, Shao-Hang Hsieh, Caiming Xiong, and Jason J Corso. Can humans fly? action understanding with multiple classes of actors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2264-2273, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[56] Zhujun Xu and Damien Vivet. Instance sequence queries for video instance segmentation with transformers. Sensors, 21(13):4507, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.567 + ], + "angle": 0, + "content": "[57] Sibei Yang, Guanbin Li, and Yizhou Yu. Cross-modal relationship inference for grounding referring expressions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4145-4154, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.471, + 0.622 + ], + "angle": 0, + "content": "[58] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4644-4653, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.471, + 0.692 + ], + "angle": 0, + "content": "[59] Sibei Yang, Guanbin Li, and Yizhou Yu. Propagating over phrase relations for one-stage visual grounding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 589-605. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.748 + ], + "angle": 0, + "content": "[60] Sibei Yang, Guanbin Li, and Yizhou Yu. Relationship-embedded representation learning for grounding referring expressions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2765-2779, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.471, + 0.829 + ], + "angle": 0, + "content": "[61] Shusheng Yang, Xinggang Wang, Yu Li, Yuxin Fang, Jiemin Fang, Wenyu Liu, Xun Zhao, and Ying Shan. Temporally efficient vision transformer for video instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2885-2895, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[62] Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11266-11275, 2021. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[63] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[64] Linwei Ye, Zhi Liu, and Yang Wang. Dual convolutional LSTM network for referring image segmentation. IEEE Transactions on Multimedia, 22(12):3224-3235, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[65] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.276, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[66] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1307-1315, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.346, + 0.892, + 0.429 + ], + "angle": 0, + "content": "[67] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.43, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[68] Wangbo Zhao, Kai Wang, Xiangxiang Chu, Fuzhao Xue, Xinchao Wang, and Yang You. Modeling motion with multimodal features for text-based video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11737-11746, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.501, + 0.892, + 0.556 + ], + "angle": 0, + "content": "[69] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15476" + } + ] +] \ No newline at end of file diff --git a/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/3b08deeb-33eb-47b0-94a4-e0af89436d2d_origin.pdf b/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/3b08deeb-33eb-47b0-94a4-e0af89436d2d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9c5a60f1bb0046963ad1d8bd632bbc91524f7bee --- /dev/null +++ b/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/3b08deeb-33eb-47b0-94a4-e0af89436d2d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71a473d1bfc88bab67a89a0fd16e32b0051d22cafa64c7005acd687cbd31b5dc +size 9072784 diff --git a/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/full.md b/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8ed5ee461a90dce5f6c79910f9fc606b1f2cb041 --- /dev/null +++ b/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/full.md @@ -0,0 +1,331 @@ +# Temporal Collection and Distribution for Referring Video Object Segmentation + +Jiajin Tang Ge Zheng Sibei Yang + +School of Information Science and Technology, ShanghaiTech University + +{tangjj,zhengge,yangsb}@shanghaiitech.edu.cn + +https://toneyaya.github.io/tempcd + +# Abstract + +Referring video object segmentation aims to segment a referent throughout a video sequence according to a natural language expression. It requires aligning the natural language expression with the objects' motions and their dynamic associations at the global video level but segmenting objects at the frame level. To achieve this goal, we propose to simultaneously maintain a global referent token and a sequence of object queries, where the former is responsible for capturing video-level referent according to the language expression, while the latter serves to better locate and segment objects with each frame. Furthermore, to explicitly capture object motions and spatial-temporal cross-modal reasoning over objects, we propose a novel temporal collection-distribution mechanism for interacting between the global referent token and object queries. Specifically, the temporal collection mechanism collects global information for the referent token from object queries to the temporal motions to the language expression. In turn, the temporal distribution first distributes the referent token to the referent sequence across all frames and then performs efficient cross-frame reasoning between the referent sequence and object queries in every frame. Experimental results show that our method outperforms state-of-the-art methods on all benchmarks consistently and significantly. + +# 1. Introduction + +Referring video object segmentation task (RVOS) aims to segment the target referent throughout a video sequence given a natural language expression [22]. It has attracted increasing attention from the academic community, as it is a basis for assessing a comprehensive understanding of visual, temporal and linguistic information. Meanwhile, ROVS benefits various downstream interactive applications such as language-driven human-robot interaction [39], video editing [25] and video surveillance [44]. + +![](images/bafe92976a35acedb3b8ce976d1db92ce74f7a8af21b70493eb0c521e1411e84.jpg) + +![](images/4f611c80f4a4201d2996c73827234247737fe60153475524c6b58d4fcb4293db.jpg) + +![](images/931cffb550ef7d9911fd981ff68b348b47676c55072ab0ee144a76fc4002f7e6.jpg) +Figure 1. The comparison of (a) previous best-performing ReferFormer, (b) video-level decoding, and (c) ours referent-guided dynamic temporal decoding. + +Compared to referring image segmentation [12,62] that segments the target object in a single image mainly according to appearances and spatial and semantic relations, ROVS requires locating referents with temporal consistency according to object motions and dynamic associations among objects across frames. + +Previous works mainly follow a single-stage bottom-up manner [9, 15, 38, 47, 48] or a two-stage top-down fashion [28]. The bottom-up methods early align the vision and language at the local patch level, failing to explicitly model objects and their relations [28, 53]. Although top-down approaches explicitly extract object tracklets and select object tracklets matched with the language expression, their two-stage pipeline is complex and less efficient [53]. Recently, transformer-based frameworks [1, 53] have been proposed to use object queries to capture objects and their associations in an end-to-end manner. They generate object queries and utilize the transformer decoder to search objects for these object queries in every frame, as shown in + +![](images/d2a7980f87e9de4412ff13fde2c04a413ae106c5127220478fc3632414e609da.jpg) +Figure 2. Comparison of temporal modeling for previous works, naive attempt, and ours. Previous works (MTTR and ReferFormer) lack temporal modeling when decoding. Besides, object motion and interaction with other objects are crucial for referring video object segmentation. However, simply modeling temporal interaction with globally visible video-level decoding cannot achieve satisfactory results. Note that our alternating interaction between the global referent and object queries can dynamically capture spatial-temporal cross-modal reasoning over objects guided by the global referent information. + +# Figure 1a and b. + +Although these end-to-end transformer-based frameworks have achieved impressive segmentation results, they still have limitations needed to be improved. First, their object queries are frame-level and independently responsible for the search for objects in each frame, which fails to capture temporal object motions and effective object tracking, as shown in Figure 2a and b. Therefore, we observe that they often fail to obtain temporally consistent predictions for target referents across an entire video, as shown in the left example in Figure 2. Second, their interactions between object queries are also solely performed per frame, leading to failure to model object-level spatial-temporal associations between objects. Therefore, they cannot correctly ground expressions that require cross-modal spatial-temporal reasoning over multiple objects. For example (see the right example in Figure 2), they mislocate the left camel as the target object because they lack to model the relation between the left camel in the current frame and the camel walking in the subsequent frame. + +In this paper, we aim to address these limitations by empowering the decoder with the capability of temporal modeling. A straightforward approach to decode objects with temporal information is to convert the frame-level decoding into video-level by inputting a sequence of object queries to search for objects in a sequence of frames like VisTR [49] for video instance segmentation, as shown in Figure 1b and 2e. However, this simple attempt fails to achieve satisfactory results while significantly increasing the computational cost and even underperforming existing query-based transformer frameworks. We further analyze the reasons: (1) The alignment between the natural language expression and the referent relies on overall objects' motions and temporal associations in the entire video. For example, the two camels can be distinguished only if the walking motion of + +the target camel is identified based on the entire video and aligned with the language expression. In contrast, the above naive attempt prematurely aligns language with fine-grained frame-level objects so that its attention is distracted from focusing accurately on the overall motions and relations of objects at the video level. (2) However, the precise localization and segmentation of target objects should go back to and rely more on every single frame because objects with different motions may cause them to have very different spatial locations in different frames. A similar observation is discussed by SeqFormer [54] for video instance segmentation, which suggests processing attention with each frame independently. + +Therefore, to address the challenge of aligning the language expression with objects' motions and temporal associations at the global video level but segmenting objects at the local frame level, we propose to maintain both local object queries and a global referent token. Global referent token captures the video-level referent information according to the language expression, while local object queries locate and segment objects with each frame. Furthermore, the object queries and the referent token are interacted to achieve spatial-temporal object information exchange through our well-designed temporal collection and distribution mechanisms, as shown in Figure 2f. Specifically, the temporal collection collects the referent information from object queries with spatial and visual attributes to temporal object motions to language semantics. In turn, the temporal distribution first dynamically distributes the referent token to every frame to extract the referent sequence. Then, the referent sequence interacts with object queries within each frame to achieve efficient spatial-temporal reasoning over objects. Note that the temporal collection and distribution alternately propagate information between object queries and the referent token to update each other. + +Finally, we propose a Temporal Collection and Distribution network (TempCD) which integrates our novel temporal collection and distribution mechanisms into the query-based transformer framework. Without using sequence matching or post-processing during inference like the previous methods [1,53], Our TempCD can directly predict the segmentation result of every frame based on object queries in the frame and the referent token. + +In summary, our main contributions are as follows, + +- We introduce to maintain a global referent token and a sequence of local object queries parallelly to bridge the gap between video-level object alignment with language and frame-level object segmentation. +- We propose a novel collection-distribution mechanism for interacting between the referent token and object queries to capture spatial-temporal cross-modal reasoning over objects. +- We present an end-to-end Temporal Query Collection and Distribution (TempCD) for RVOS, which can directly predict the segmentation referent of every frame without any post-processing and sequence matching. +- Experiments on Ref-Youtu-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB-Sentences datasets show that TempCD outperforms state-of-the-art methods on all benchmarks consistently and significantly. + +# 2. Related Work + +Referring Image Segmentation (RIS) involves segmenting objects in images based on natural language expressions. Compared to Referring Video Object Segmentation (RVOS), RIS operates on individual images without temporal information. Previous works [3, 4, 26, 29, 42, 45, 57-59] focus on the joint modeling of vision and language. Various methods [6, 13, 14, 16, 20, 21, 35, 60, 64] are explored successively, such as using fusion operators like concatenation, ConvLSTM, attention mechanisms [46], and GNN [10] to obtain multimodal semantic feature maps. In addition, some studies attempt to decouple different components or key semantics from language, through explicit two-stage approaches [51, 66] or implicit attention modules [43, 62, 64, 65], to achieve a more fine-grained understanding. Recent models explore novel fusion techniques, including early fusion [8, 63], Linguistic Seed Encoder [23], and contrastive learning [50], to improve the performance of RIS. However, for RVOS, it requires not only the multimodal integration of language and image frames but also temporal modeling at the video level. + +Referring Video Object Segmentation (RVOS) aims to segment the target object described in natural language from videos. Previous works in RVOS has primarily used two frameworks: bottom-up and top-down. The top-down framework [28] directly models the motion informa + +tion between segmentation masks of consecutive frames. However, this two-stage approach, which includes complex computational costs, is limited in terms of video input length. For the bottom-up framework, some approaches [9, 15, 38, 47, 48] directly apply RIS methods to construct multi-modal feature maps for referring segmentation of keyframes. They rely solely on replacing the traditional image backbone with a 3D temporal backbone for temporal modeling, which limits the performance of multiframes segmentation. To address this issue, URVOS [41] extends previous methods by building a memory bank that propagates language referent object information in the temporal dimension. Furthermore, URVOS introduces the RefYoutube-VOS [41] dataset, which provides segmentation annotations for each frame that require more efficient temporal modeling. In addition, recent works [7, 52, 68] adopt language guided fusion between temporal features and visual features to obtain more efficient temporal modeling. Recently, query-based methods [1, 53] with a transformer encoder-decoder frameworks [2, 5, 69] attempt to capture object-level information. They achieve higher performance by constructing a multimodal feature map of visual and linguistic information, or by using language-initialized queries with independently decoding with queries for each frame. + +Temporal Modeling in Video Instance Segmentation (VIS). VIS aims to simultaneously track, segment and classify interest instances in a video. Similarly to RVOS, VIS requires obtaining instance information from every frame in the video and maintaining the temporal consistency of instances. Recent works focus on spatial and temporal modeling based on instance-level information. VisTR [49] employs an encoder-decoder transformer based on DETR, with a concatenated sequence of instance queries as the input to enable spatial-temporal interactions. The following IFC [17] and TeViT [61] explored a more efficient approach to temporal modeling in the encoder by introducing supplementary tokens for interaction across frames, with the decoder aligned with VisTR. Other following works [24, 56] take advantage of memories or queries from several previous frames to help instance segmentation in the current frame, extending to a online fashion. Recently, SeqFormer [54] proposes that the acquisition of frame-level instance information such as position needs to go back to each frame, i.e., the decoding process should be done independently for each frame. While these methods provide potential solutions for temporal modeling, the direct application to RVOS is not feasible, because of inconsistent object semantics over frames. + +# 3. Method + +The framework of our proposed TempCD is shown in Figure 3. We introduce a sequence of object queries to capture object information in every frame independently and + +![](images/494aa9f57b9ad145567289aeec76eac8c793e0eeba4ed37fe906091c9e192b24.jpg) +Figure 3. Overall framework of TempCD. TempCD simultaneously maintains a global referent token and a sequence of object queries, and they interact via the proposed collection-distribution mechanism. Specifically, the temporal collection collects global information for the referent token from object queries to the temporal motions to the language expression. In turn, the temporal distribution first distributes the referent token to the referent sequence and then performs efficient cross-frame reasoning between the referent sequence and object queries. + +maintain a referent token to capture the global referent information for aligning with language expression. The referent token and object queries are alternately updated to stepwise identify the target object and better segment the target object in every frame through the proposed temporal collection and distribution mechanisms. + +Specifically, we first introduce encoders and the definition of object queries in Section 3.1. Then, we present the temporal collection that collects the referent information for the referent token from object queries to temporal object motions to the language expression (see Section 3.2). Next, we introduce the temporal distribution mechanism that distributes the global referent information to the referent sequence across frames to object queries in each frame (see Section 3.3). The collection-distribution mechanism explicitly captures object motions and spatial-temporal cross-modal reasoning over objects. Finally, we introduce segmentation heads and loss functions in Section 3.4. Note that as we explicitly maintain the referent token, we can directly identify the referent object in each frame without requiring sequence matching as in previous queries-based methods. + +# 3.1. Encoders and Query Definition + +Language Encoder. Following previous work [53], we adopt RoBERTa [32] as our language encoder. Given a language expression with $L$ words, we extract its sentence feature $\mathcal{F}_s \in \mathbb{R}^C$ , where $C$ denotes channel dimensions. In addition, we obtain the output feature representation of each word before the last pooling layer as $\mathcal{F}_w \in \mathbb{R}^{L \times C}$ . + +Visual Backbone and Encoder. Following previous works [1, 52, 53, 61], we use ResNet-50 [11] as our visual backbone. We also follow recent work [1, 53] to evaluate our model using the temporal visual backbone network Video-Swin [33] in experiments. For an input video $V \in \mathbb{R}^{H \times W \times 3 \times T}$ with $T$ frames, we first extract its visual feature maps and further construct multimodal feature maps $\mathcal{V} = \{v_i\}_{i=1}^T$ by fusing the visual feature maps with the word features of the expression $\mathcal{F}_w$ following [53]. + +Object Query Definition. Inspired by query-based detection and segmentation frameworks [2, 5, 69], we leverage object queries to represent object-level information. For the video with $\mathcal{T}$ frames, we introduce a sequence of $\mathcal{T} \times N$ object queries to represent the objects in every frame and follow [53] to use the language feature $\mathcal{F}_s$ to initialize the object queries as $\mathcal{Q} \in \mathbb{R}^{\mathcal{T} \times N \times C}$ , where the number of objects and the dimension of channels are $N$ and $C$ , respectively. + +Referent Token Definition. In addition to adopting the object queries to represent objects in every frame, we define a referent token $\mathcal{Q}^g\in \mathbb{R}^C$ to represent the global information of the target object. As the language expression naturally indicates the referent information, we thus use the language feature $\mathcal{F}_s$ to initialize the $\mathcal{Q}^g$ . + +# 3.2. Temporal Collection from Object Queries + +The temporal collection mechanism collects object motions from frames and aggregates object motions consistent with the language expression into the global referent token. + +We first perform motion collection across object queries (see Section 3.2.1) and then update the global referent token from the object motions and language expression (see Section 3.2.2). + +# 3.2.1 Motion Collection across Object Queries + +For simplicity of demonstration, we denote the object queries output from the $(l - 1)$ -th layer as $\mathcal{Q}_{l - 1}$ and input them into the $l$ -th layer of TempCD. Note that we input the object queries $\mathcal{Q}$ defined in Section 3.1 into the first layer of TempCD. We utilize $\mathcal{Q}_{l - 1}$ to capture object information in each frame and integrate them to obtain temporal information such as object motion and action. Specifically, for $i$ -th frame, we feed its object queries $q_{i} = \mathcal{Q}_{l - 1}^{i} \in \mathbb{R}^{N \times C}$ and the multimodal features $v_{i}$ into the DETR transformer decoder [2,69] to locate objects, where $\mathcal{Q}_{l - 1}^{i}$ means the object queries of $i$ -th frame. The computation proceeds as follows, + +$$ +\hat {\mathcal {Q}} _ {l} = \left\{\operatorname {D E T R} \left(q _ {i}, v _ {i}\right) \right\} _ {i = 1} ^ {\mathcal {T}}, \tag {1} +$$ + +where $\mathrm{DETR}(\cdot ,\cdot)$ represents the DETR transformer decoder, and $\hat{\mathcal{Q}}_l$ denotes the object queries output by the decoder. + +Next, to aggregate object motions across the video sequence, we combine objects information from queries of each frame. Specifically, we generate temporal weights $G \in \mathbb{R}^{T \times N \times 1}$ for each frame given the object queries $\hat{\mathcal{Q}}_l$ . This is done by employing a nonlinear layer followed by a softmax operator. Then, we use the temporal weights $G$ to integrate the frame-level information and obtain object motions $M \in \mathbb{R}^{N \times C}$ through a gating mechanism. The integration is computed as follows, + +$$ +G = \operatorname {s o f t m a x} (\hat {\mathcal {Q}} _ {t} W), +$$ + +$$ +M = \sum_ {i = 1} ^ {\mathcal {T}} \left(G _ {i} \odot \hat {\mathcal {Q}} _ {l} ^ {i}\right), \tag {2} +$$ + +where $W \in \mathbb{R}^{C \times 1}$ denotes the learnable weights of the linear layer, $G_{i} \in \mathbb{R}^{N \times 1}$ is the weight for $i$ -th frames in $G$ , $\mathrm{softmax}(\cdot)$ represents the softmax operation along the temporal dimension, and $\odot$ is the element-wise product. + +# 3.2.2 Global Referent Token Collection + +Next, we update the global referent token $\mathcal{Q}_{l-1}^g \in \mathbb{R}^C$ at the $(l-1)$ -th layer to referent token $\mathcal{Q}_l^g \in \mathbb{R}^C$ by selecting the object motions $M \in \mathbb{R}^{N \times C}$ consistent with the language feature $\mathcal{F}_s \in \mathbb{R}^C$ . Note that we input the global referent token $\mathcal{Q}^g$ defined in Section 3.1 into the first layer of TempCD. We first calculate the cosine similarity between the language feature $\mathcal{F}_s$ and the object motions $M$ . We obtain the similarity scores $S = \{s_j\}_{j=1}^N$ as follows, + +$$ +s _ {j} = \cos \left(\mathcal {F} _ {s}, m _ {j}\right), \tag {3} +$$ + +where $\cos (\cdot ,\cdot)$ means the cosine similarity function and $m_j$ denotes the $j$ -th object's motion feature in $M$ . Next, we update the global referent token $\mathcal{Q}_g$ as the motion feature of the object with the highest similarity score to the global referent token. To ensure the selection process is differentiable, we implement this selection through the Gumbel Softmax operator [18, 36]. The calculation is as follows: + +$$ +S _ {\text {g u m b e l}} = \operatorname {s o f t m a x} \left(\left(S + G _ {s}\right) / \tau\right), +$$ + +$$ +S _ {\text {o n e h o t}} = \text {o n e h o t} \left(\arg \max _ {N} \left(S _ {\text {g u m b e l}}\right)\right), \tag {4} +$$ + +$$ +S _ {\text {r e f e r e n t}} = S _ {\text {o n e h o t}} - \mathrm {s g} \left[ S _ {\text {g u m b e l}} \right] + S _ {\text {g u m b e l}}, +$$ + +where $\operatorname{softmax}(\cdot)$ represents the softmax operation along dimension $N$ , $G_{s} \in \mathbb{R}^{N \times 1}$ is sampled from the Gumbel(0,1) distribution, $\tau$ is the temperature coefficient hyperparameter, and $\mathrm{sg}$ is the stop gradient operation. Additionally, $\operatorname{argmax}_{N}$ represents choosing the index with the highest similarity score, and onehot represents setting the highest-scoring item to 1 based on the selected index and setting the others to 0. The one-hot vector $S_{\text{referent}} \in \mathbb{R}^{N \times 1}$ indicates the selection from the object motions. The global referent token $\mathcal{Q}^g \in \mathbb{R}^C$ is updated as follows: + +$$ +\mathcal {Q} _ {l} ^ {g} = \operatorname {M L P} \left(\operatorname {L a y e r N o r m} \left(S _ {\text {r e f e r e n t}} ^ {\top} M\right) + Q _ {l - 1} ^ {g}\right), \tag {5} +$$ + +where $\mathrm{MLP}(\cdot)$ denotes a three-layer linear layer with activation function, and LayerNorm $(\cdot)$ is the same as the LayerNorm in standard Transformer. + +We obtain the updated global referent token $\mathcal{Q}^g$ that contains information on the referent object aligned with the language expression. Next, we feed it into the Temporal Distribution module to assist in distributing the referent information into each frame. + +# 3.3. Temporal Distribution from Referent Token + +The temporal distribution aims to propagate the referent information to all frames and perform dynamically spatial-temporal reasoning over objects based on the referent information. First, we distribute the global referent token to each frame independently to extract the referent sequence across all frames (see Section 3.3.1). Next, we perform the efficient cross-frame reasoning between the referent sequence and object queries in every frame (see Section 3.3.2). + +# 3.3.1 Global Referent Token Distribution + +Given the global referent token $\mathcal{Q}_l^g$ , we distribute the referent object information to every frame and further obtain the referent sequence across all frames, denoted by $\mathcal{R}_l = \{r_i\}_{i=1}^T$ , which refers to the target object in each frame. + +Specifically, we first compute the cosine similarity scores between the object queries $\hat{\mathcal{Q}}_l = \{q_i\}_{i=1}^T$ and the global referent token $\mathcal{Q}_l^g$ . Then we obtain the index $d_i$ of the + +object query with the highest similarity score to the global referent token for every frame $i$ independently. The computation is as follows, + +$$ +s _ {i, j} = \cos \left(q _ {i, j}, Q _ {l} ^ {g}\right), \tag {6} +$$ + +$$ +d _ {i} = \operatorname {a r g m a x} _ {j} \left\{s _ {i, j} \right\}, +$$ + +where $q_{i,j}$ is the $j$ -th object query in the $i$ -th frame, and $s_{i,j}$ is the corresponding similarity score. For every frame $i$ , we select the object query $d_i$ with the highest similarity score and fuse it with the global referent token $\mathcal{Q}_l^g$ to generate the referent sequence $\mathcal{R}_l = \{r_i\}_{i=1}^T$ as follows, + +$$ +\mathcal {R} _ {l} = \left\{\mathrm {M L P} \left(\operatorname {L a y e r N o r m} \left(q _ {i, d _ {i}} + \mathcal {Q} _ {l} ^ {g}\right)\right) \right\} _ {i = 1} ^ {\mathcal {T}}. \tag {7} +$$ + +We end up with referent sequence $\mathcal{R}_l\in \mathbb{R}^{T\times C}$ across all frames, which is used as crucial objects to interact with object queries in every frame independently. + +# 3.3.2 Spatial-temporal Reasoning over Object Queries + +We leverage the referent sequence $\mathcal{R}_l\in \mathbb{R}^{T\times C}$ , which encodes the referent object information across all frames, to interact with the object queries to achieve the cross-frame temporal reasoning over multiple objects and further distribute referent information to all object queries in all frames. + +Specifically, for each object query $q_{i,j}$ at the $i$ -th frame, we concatenate it with the objects of the referent sequence in all other frames to construct the cross-object query sequence $C_{i,j} = \{r_1,..,r_{i - 1},q_{i,j},r_{i + 1},..,r_{\mathcal{T}}\} \in \mathbb{R}^{\mathcal{T}\times C}$ . The object-level interaction across frames is implemented by multi-head self attention, which is computed as follows, + +$$ +\hat {C} _ {i, j} = \operatorname {M H S A} \left(C _ {i, j}\right), \tag {8} +$$ + +where $\mathrm{MHSA}(\cdot)$ denotes the multi-head self-attention mechanism. The $j$ -th updated object queries in $C_{i,j}^{\hat{}}$ for every frame $i$ are concatenated as $\mathcal{Q}_l^f \in \mathbb{R}^{\mathcal{T} \times N \times C}$ to input to next decoder layer. Note that our cross-frame interaction only requires interaction with referent sequences to object queries in each frame, which achieves efficient cross-frame interaction over multiple objects. + +# 3.4. Prediction Heads and Loss + +We integrate the referent sequence and object queries in each decoder layer to the prediction heads for predicting results. To simplify the presentation, we use the final layer for details. + +We extract the referent sequence $\mathcal{R} \in \mathbb{R}^{T \times C}$ and combine them with the object queries $Q_{l}$ to obtain the output referent sequence $\hat{\mathcal{R}}$ for prediction. Specifically, we use a gating mechanism to integrate the information from $\hat{C}$ into video-level features, which is similar to Eq 2. Next, we calculate cosine similarity between each sequence $C_{j,i}$ and the + +language sentence feature $\mathcal{F}_s$ , which is similar to Eq 3, to obtain the score for each object $q_{i,j}$ . We select the highest-scoring object $q_i^r$ in each frame and combine it with $\mathcal{R}$ to obtain the final referent sequence $\hat{\mathcal{R}} \in \mathbb{R}^{T \times C}$ of the referred object in each frame: + +$$ +\hat {\mathcal {R}} = \left\{\mathrm {M L P} \left(\operatorname {L a y e r N o r m} \left(r _ {i} + q _ {i} ^ {r}\right)\right) \right\}. \tag {9} +$$ + +Finally, we feed $\hat{\mathcal{R}}$ into the segmentation head to predict the segmentation results. Additionally, similar to [53] and [1], we use a box prediction head to regress the box coordinates for each query. We utilize Dice [37] loss and Focal [30] loss as the segmentation loss and GIOU [40] loss and L1 loss as the box loss. + +# 4. Experiment + +Datasets. We conduct experiments on four benchmark datasets that are publicly available: Ref-Youtu-VOS [41], Ref-Davis-2017 [22], A2D-Sentences [9], and JHMDB-Sentences [9]. Ref-Youtu-VOS consists of 3978 videos and approximately 15K language descriptions. Ref-Davis-2017 contains approximately 90 videos. A2D-Sentences and JHMDB-Sentences [19, 55], originally focused on action recognition, have been expanded to incorporate language expression annotations. This expansion has yielded a total of 3782 videos for A2D-Sentences and 928 videos for JHMDB-Sentences, each accompanied by corresponding language descriptions. + +Implementation Details. Following [53], we pre-train our model on RefCOCO dataset [67]. We utilize a text encoder derived from RoBERTa [32], coupled with either ResNet50 [11] or Video-Swin serving [33] as our visual backbone. We train our model for 6 epochs with an initial learning rate of 1e-4 and the AdamW [34] optimizer. Consistent with prior works [1, 41, 53], we use the $\mathcal{I}$ , $\mathcal{F}$ , and $\mathcal{J} \& \mathcal{F}$ metrics for evaluation on Ref-Youtu-VOS and Ref-Davis-2017 datasets, and Overall IoU and Mean IoU as evaluation metrics on A2D-Sentences and JHMDB-Sentences datasets. + +# 4.1. Comparison with State-of-the-Art Methods + +As shown in Table 1 and Table 2, we compare TempCD with state-of-the-art methods on four benchmarks. TempCD consistently outperforms state-of-the-art methods on all datasets. + +Comparison of Ref-Youtube-VOS and Ref-Davis-2017 Datasets. Results for the Ref-Youtu-VOS and Ref-Davis-2017 datasets are presented in Table 1. With a standard ResNet-50 [11] visual backbone, our model exhibits improvements of $2.7\%$ , $4\%$ , and $3.4\%$ for the $\mathcal{I}, \mathcal{F}$ , and $\mathcal{J} \& \mathcal{F}$ metrics, respectively, on the Ref-YoutuVOS dataset. Employing the advanced temporal visual backbone, VideoSwin-B [33], our approach consistently outperforms the + +
BackboneMethodRef-Youtube-VOSRef-Davis17
JFJ&FJFJ&F
ResNet-50CMSA [65]33.336.534.932.237.234.7
CMSA+RNN [65]34.838.136.436.943.540.2
URVOS [41]45.349.247.247.356.051.5
MLRL [52]48.451.049.750.155.452.7
LBDT [7]48.250.649.4--54.5
ReferFormer [53]54.856.555.655.861.358.5
Ours57.560.559.057.362.760.0
Video-Swin-TMTTR [1]54.056.655.3---
ReferFormer [53]58.060.959.4---
Ours60.564.062.359.365.062.2
Video-Swin-BReferFormer [53]61.364.662.958.164.161.1
Ours63.668.065.861.667.664.6
+ +Table 1. Comparison with state-of-the-art models for RVOS on Ref-Youtu-VOS and Ref-Davis-2017 datasets. + +
BackboneMethodA2D-SJHMDB-S
oIoUmIoUoIoUmIoU
VGG16Hu et al. [12]47.435.054.652.8
I3DGavrilyuk et al. [9]53.642.154.154.2
ACAN [48]60.149.075.656.4
CMPC-V [31]65.357.361.661.7
Resnet-50ClawCraneNet [27]63.159.964.465.6
MMMMTBVS [68]67.355.861.961.3
LBDT [7]70.462.164.565.8
Ours76.668.670.669.6
+ +Table 2. Comparison with state-of-the-art models for RVOS on A2D-Sentences (A2D-S) and JHMDB-Sentences (JHMDB-S). + +previous state-of-the-art model [53] on the Ref-Davis-2017 dataset, achieving a $3.5\%$ increase across the aforementioned metrics. Note that evaluations on the RefDavis-2017 dataset are performed using models trained on the RefYoutube-VOS dataset. + +Furthermore, we present a thorough evaluation of the effectiveness of our TempCD through subsequent comparisons: (1) We compare our model to MTTR [1], a pioneering method that adopts a query-based framework for RVOS, despite lacking temporal modeling in both the encoder and decoder stages. Our TempCD surpasses MTTR, achieving a $7\%$ improvement in the $\mathcal{J}\& \mathcal{F}$ metric on the Ref-YoutuVOS dataset. This outcome underscores the heightened efficacy of our introduced temporal decoding modules. (2) Compared with LBDT [7], our TempCD achieves a significant $9.6\%$ enhancement in the $\mathcal{J}\& \mathcal{F}$ metric. This result suggests that, unlike [7, 41, 52] that integrate temporal and visual features via a bottom-up framework, our proposed TempCD achieves more efficient temporal modeling. + +Comparison on A2D-Sentences and JHMDB-Sentences. As presented in Table 2, our approach yields mean improvements of $6.2\%$ in Overall IoU and $5.2\%$ in Mean IoU, surpassing the predominant state-of-the-art method [7] for + +these two datasets. In contrast to Ref-Youtu-VOS, these datasets predominantly comprise segmentation annotations for keyframes that encapsulate actions. When compared to prior studies [9, 15, 38, 47, 48] that adopt temporal encoders to process video temporal information, our method consistently attains enhancements across these two datasets. This highlights the significance of explicitly capturing and integrating video-level temporal context into individual frames. + +# 4.2. Ablation Study + +To validate the impact of modules in our model, we evaluate six variants of our TempCD on Ref-Youtu-VOS dataset. The results are shown in Table 3. All experiments are performed with Video-Swin-B as the visual backbone. + +Baselines. (1) Our baseline model employs $\mathcal{T}$ sets of frame-level object queries ("Local Query"). These queries are concatenated sequentially and fed into the decoder, following a similar approach to VisTR [49]. The derived results (53.7% $\mathcal{J}$ , 56.3% $\mathcal{F}$ , and 55.0% $\mathcal{J} \& \mathcal{F}$ ) suggest a misalignment between the natural language representations and the referent semantics at the individual frame level. (2) Another subsequent baseline is established by integrating a singular shared set of video-level object queries across all frames ("Global Query"). The global queries are performed an inner product with the feature map $v$ of each frame, yielding segmentation outcomes. While this approach intrinsically aligns with language representations and achieves consistency between frames, it struggles to adapt to frame-level variability, resulting in a performance decline relative to the first baseline. + +Query Collection and Distribution. (3) A natural enhancement for baseline (1) entails integrating cross-frame temporal attention on a referent sequence, which is selected from local queries by the similarity with language. This modification leads to a performance improvement of $3.4\%$ . (4) Another endeavor to enhance the baseline (1) involves the introduction of an additional set of global queries. These + +![](images/67f0786b2f7e9b6f23e9232133078953265aa0ec4c670a7e8395adff828d2075.jpg) +Figure 4. Visualization for segmentation results of previous best-performing ReferFormer [53] and ours. + +
MethodJFJ&F
1Local Query53.756.355.0
2Global Query52.955.554.2
31+Cross-Frame56.259.157.6
41+Motion Collection58.161.559.8
53+460.764.462.6
65+Gumble-Softmax61.465.463.4
7Ours63.668.065.8
+ +Table 3. Ablation Study on Ref-Youtu-VOS dataset. + +queries collect motion information from local queries and serve to distribute global context to local queries. Specifically, we utilize a Top-K strategy to select global information based on the similarity score with language. This facilitates implicit frame level interaction and, significantly, fosters semantic alignment between global queries and language expressions. This alignment contributes to a marked performance gain of $5.6\%$ . (5) We further incorporate both (3)'s explicit cross-frame temporal modeling and (4)'s motion collection and distribution, which delivers a $2.8\%$ improvement. (6) We replace the Top-K selection, which does not allow gradient calculation, with Gumbel Softmax, improving $\mathcal{J} \& \mathcal{F}$ by $0.8\%$ . (7) Our complete model further enhances the vanilla cross-frame temporal attention mechanism. Instead of solely updating the referent sequence via temporal interaction, local object queries are also extended to incorporate temporal context information from referent sequence, facilitating cross-frame interaction over multiple objects. The full decoder of our proposed method achieves the performance of $65.8\%$ in terms of $\mathcal{J} \& \mathcal{F}$ . + +# 4.3. Visualization + +Figure 4 visualizes several qualitative results. The referring expression in (a) describes the motion of a white duck, a pivotal feature that distinguishes it from similar ducks. Our proposed TempCD effectively captures the motion of the referent which is aligned with the language expression and enables accurate localization and segmentation of the specific duck. In contrast, Referformer fails to + +locate the referent object, primarily attributed to a semantic incongruity between the expression and the frame-level queries. In the instance of (b), the deer's corresponding action appears in certain frames but lacks uniform presence. Our approach successfully captures its global semantics including motions and aligns it seamlessly with the referring expression, bridging the gap between local semantics and the referring expression via the Collection and Distribution mechanisms. Conversely, ReferFormer identifies the referent accurately in specific frames that correspond to the specified action, yet exhibits errors in others due to a noticeable lack of temporal context. All these observations collectively underscore the value of temporal interaction in fostering refined segmentation across frames. Besides, (c) illustrates that our method can still precisely locate and segment referent objects in complex scenes. The second frame presents a scenario of occlusion, wherein the referent is obscured by another bicycle. Our approach detects the lack of correspondence between visible entities in a frame and the specified referent sequence, facilitated by cross-frame interactions between the visible objects and those delineated by the referent sequence. + +# 5. Conclusion + +This paper proposes an end-to-end Temporal Query Collection and Distribution (TempCD) network for referring video object segmentation, which maintains object queries and the referent token and achieves alternating interaction between them via the proposed novel temporal collection-distribution mechanism. + +Acknowledgment: This work was supported by the National Natural Science Foundation of China (No.62206174), Shanghai Pujiang Program (No.21PJ1410900), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (ShangHAI), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University), and Shanghai Engineering Research Center of Intelligent Vision and Imaging. + +# References + +[1] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4985-4995, 2022. 1, 3, 4, 6, 7 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3, 4, 5 +[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7454-7463, 2019. 3 +[4] Yi-Wen Chen, Yi-Hsuan Tsai, Tiantian Wang, Yen-Yu Lin, and Ming-Hsuan Yang. Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748, 2019. 3 +[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 3, 4 +[6] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 3 +[7] Zihan Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Jizhong Han, and Si Liu. Language-bridged spatial-temporal interaction for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4964–4973, 2022. 3, 7 +[8] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 3 +[9] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5958-5966, 2018. 1, 3, 6, 7 +[10] Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. In Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., volume 2, pages 729-734. IEEE, 2005. 3 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 6 +[12] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 1, 7 + +[13] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4424-4433, 2020. 3 +[14] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10488-10497, 2020. 3 +[15] Tianrui Hui, Shaofei Huang, Si Liu, Zihan Ding, Guanbin Li, Wenguan Wang, Jizhong Han, and Fei Wang. Collaborative spatial-temporal modeling for language-queried video actor segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4187–4196, 2021. 1, 3, 7 +[16] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 59-75. Springer, 2020. 3 +[17] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. Advances in Neural Information Processing Systems, 34:13352-13363, 2021. 3 +[18] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5 +[19] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 6 +[20] Yang Jiao, Zequn Jie, Weixin Luo, Jingjing Chen, Yu-Gang Jiang, Xiaolin Wei, and Lin Ma. Two-stage visual cues enhancement network for referring image segmentation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1331-1340, 2021. 3 +[21] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tieniu Tan. Locate then segment: A strong pipeline for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9858-9867, 2021. 3 +[22] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 1, 6 +[23] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18154, 2022. 3 +[24] Rajat Koner, Tanveer Hannan, Suprosanna Shit, Sahand Sharifzadeh, Matthias Schubert, Thomas Seidl, and Volker + +Tresp. Instanceformer: An online video instance segmentation framework. arXiv preprint arXiv:2208.10547, 2022. 3 +[25] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip HS Torr. Manigan: Text-guided image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7880-7889, 2020. 1 +[26] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2018. 3 +[27] Chen Liang, Yu Wu, Yawei Luo, and Yi Yang. Clawranenet: Leveraging object-level relation for text-based video segmentation. arXiv preprint arXiv:2103.10702, 2021. 7 +[28] Chen Liang, Yu Wu, Tianfei Zhou, Wenguan Wang, Zongxin Yang, Yunchao Wei, and Yi Yang. Rethinking cross-modal interaction from a top-down perspective for referring video object segmentation. arXiv preprint arXiv:2106.01061, 2021. 1, 3 +[29] Liang Lin, Pengxiang Yan, Xiaogqian Xu, Sibei Yang, Kun Zeng, and Guanbin Li. Structured attention network for referring image segmentation. IEEE Transactions on Multimedia, 24:1922-1932, 2021. 3 +[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6 +[31] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4761-4775, 2021. 7 +[32] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4, 6 +[33] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. arXiv preprint arXiv:2106.13230, 2021. 4, 6 +[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[35] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020. 3 +[36] Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016. 5 +[37] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6 + +[38] Ke Ning, Lingxi Xie, Fei Wu, and Qi Tian. Polar relative positional encoding for video-language segmentation. In *IJ-CAI*, volume 9, page 10, 2020. 1, 3, 7 +[39] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 1 +[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6 +[41] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 3, 6, 7 +[42] Cheng Shi and Sibei Yang. Spatial and visual perspectivetaking via view rotation and relation reasoning for embodied reference understanding. In European Conference on Computer Vision, pages 201-218. Springer, 2022. 3 +[43] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 38-54, 2018. 3 +[44] G Sreenu and Saleem Durai. Intelligent video surveillance: a review through deep learning techniques for crowd analysis. Journal of Big Data, 6(1):1-27, 2019. 1 +[45] Jiajin Tang, Ge Zheng, Cheng Shi, and Sibei Yang. Contrastive grouping with transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23570-23580, 2023. 3 +[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[47] Hao Wang, Cheng Deng, Fan Ma, and Yi Yang. Context modulated dynamic networks for actor and action video segmentation with language queries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12152-12159, 2020. 1, 3, 7 +[48] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3939-3948, 2019. 1, 3, 7 +[49] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8741-8750, 2021. 2, 3, 7 +[50] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip + +driven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 3 +[51] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasescut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 3 +[52] Dongming Wu, Xingping Dong, Ling Shao, and Jianbing Shen. Multi-level representation learning with semantic alignment for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4996-5005, 2022. 3, 4, 7 +[53] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 1, 3, 4, 6, 7, 8 +[54] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. arXiv preprint arXiv:2112.08275, 2021. 2, 3 +[55] Chenliang Xu, Shao-Hang Hsieh, Caiming Xiong, and Jason J Corso. Can humans fly? action understanding with multiple classes of actors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2264-2273, 2015. 6 +[56] Zhujun Xu and Damien Vivet. Instance sequence queries for video instance segmentation with transformers. Sensors, 21(13):4507, 2021. 3 +[57] Sibei Yang, Guanbin Li, and Yizhou Yu. Cross-modal relationship inference for grounding referring expressions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4145-4154, 2019. 3 +[58] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4644-4653, 2019. 3 +[59] Sibei Yang, Guanbin Li, and Yizhou Yu. Propagating over phrase relations for one-stage visual grounding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 589-605. Springer, 2020. 3 +[60] Sibei Yang, Guanbin Li, and Yizhou Yu. Relationship-embedded representation learning for grounding referring expressions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2765-2779, 2020. 3 +[61] Shusheng Yang, Xinggang Wang, Yu Li, Yuxin Fang, Jiemin Fang, Wenyu Liu, Xun Zhao, and Ying Shan. Temporally efficient vision transformer for video instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2885-2895, 2022. 3, 4 +[62] Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11266-11275, 2021. 1, 3 + +[63] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 3 +[64] Linwei Ye, Zhi Liu, and Yang Wang. Dual convolutional LSTM network for referring image segmentation. IEEE Transactions on Multimedia, 22(12):3224-3235, 2020. 3 +[65] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 3, 7 +[66] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1307-1315, 2018. 3 +[67] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6 +[68] Wangbo Zhao, Kai Wang, Xiangxiang Chu, Fuzhao Xue, Xinchao Wang, and Yang You. Modeling motion with multimodal features for text-based video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11737-11746, 2022. 3, 7 +[69] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 4, 5 \ No newline at end of file diff --git a/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/images.zip b/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0debc890bcc0698e554ac73f97a9cdea5df92fe0 --- /dev/null +++ b/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9d773f7147a9f6b3f024f1d9a3fba567a2bc34df4b4d5ec30b13fef9b1c4fb6 +size 555498 diff --git a/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/layout.json b/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..791693314f3e7a0e1d99c6b0f7eba4527d053c6b --- /dev/null +++ b/2023/Temporal Collection and Distribution for Referring Video Object Segmentation/layout.json @@ -0,0 +1,9568 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "text", + "content": "Temporal Collection and Distribution for Referring Video Object Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 202, + 143, + 392, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 143, + 392, + 159 + ], + "spans": [ + { + "bbox": [ + 202, + 143, + 392, + 159 + ], + "type": "text", + "content": "Jiajin Tang Ge Zheng Sibei Yang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 163, + 474, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 163, + 474, + 177 + ], + "spans": [ + { + "bbox": [ + 118, + 163, + 474, + 177 + ], + "type": "text", + "content": "School of Information Science and Technology, ShanghaiTech University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 179, + 178, + 411, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 178, + 411, + 191 + ], + "spans": [ + { + "bbox": [ + 179, + 178, + 411, + 191 + ], + "type": "text", + "content": "{tangjj,zhengge,yangsb}@shanghaiitech.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 206, + 194, + 388, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 194, + 388, + 205 + ], + "spans": [ + { + "bbox": [ + 206, + 194, + 388, + 205 + ], + "type": "text", + "content": "https://toneyaya.github.io/tempcd" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 232, + 192, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 232, + 192, + 245 + ], + "spans": [ + { + "bbox": [ + 143, + 232, + 192, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 258, + 290, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 290, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 290, + 547 + ], + "type": "text", + "content": "Referring video object segmentation aims to segment a referent throughout a video sequence according to a natural language expression. It requires aligning the natural language expression with the objects' motions and their dynamic associations at the global video level but segmenting objects at the frame level. To achieve this goal, we propose to simultaneously maintain a global referent token and a sequence of object queries, where the former is responsible for capturing video-level referent according to the language expression, while the latter serves to better locate and segment objects with each frame. Furthermore, to explicitly capture object motions and spatial-temporal cross-modal reasoning over objects, we propose a novel temporal collection-distribution mechanism for interacting between the global referent token and object queries. Specifically, the temporal collection mechanism collects global information for the referent token from object queries to the temporal motions to the language expression. In turn, the temporal distribution first distributes the referent token to the referent sequence across all frames and then performs efficient cross-frame reasoning between the referent sequence and object queries in every frame. Experimental results show that our method outperforms state-of-the-art methods on all benchmarks consistently and significantly." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "type": "text", + "content": "Referring video object segmentation task (RVOS) aims to segment the target referent throughout a video sequence given a natural language expression [22]. It has attracted increasing attention from the academic community, as it is a basis for assessing a comprehensive understanding of visual, temporal and linguistic information. Meanwhile, ROVS benefits various downstream interactive applications such as language-driven human-robot interaction [39], video editing [25] and video surveillance [44]." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 232, + 541, + 289 + ], + "blocks": [ + { + "bbox": [ + 309, + 232, + 541, + 289 + ], + "lines": [ + { + "bbox": [ + 309, + 232, + 541, + 289 + ], + "spans": [ + { + "bbox": [ + 309, + 232, + 541, + 289 + ], + "type": "image", + "image_path": "bafe92976a35acedb3b8ce976d1db92ce74f7a8af21b70493eb0c521e1411e84.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 290, + 541, + 348 + ], + "blocks": [ + { + "bbox": [ + 309, + 290, + 541, + 348 + ], + "lines": [ + { + "bbox": [ + 309, + 290, + 541, + 348 + ], + "spans": [ + { + "bbox": [ + 309, + 290, + 541, + 348 + ], + "type": "image", + "image_path": "4f611c80f4a4201d2996c73827234247737fe60153475524c6b58d4fcb4293db.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 350, + 539, + 427 + ], + "blocks": [ + { + "bbox": [ + 309, + 350, + 539, + 427 + ], + "lines": [ + { + "bbox": [ + 309, + 350, + 539, + 427 + ], + "spans": [ + { + "bbox": [ + 309, + 350, + 539, + 427 + ], + "type": "image", + "image_path": "931cffb550ef7d9911fd981ff68b348b47676c55072ab0ee144a76fc4002f7e6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 429, + 545, + 461 + ], + "lines": [ + { + "bbox": [ + 305, + 429, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 429, + 545, + 461 + ], + "type": "text", + "content": "Figure 1. The comparison of (a) previous best-performing ReferFormer, (b) video-level decoding, and (c) ours referent-guided dynamic temporal decoding." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 484, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 556 + ], + "type": "text", + "content": "Compared to referring image segmentation [12,62] that segments the target object in a single image mainly according to appearances and spatial and semantic relations, ROVS requires locating referents with temporal consistency according to object motions and dynamic associations among objects across frames." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 715 + ], + "type": "text", + "content": "Previous works mainly follow a single-stage bottom-up manner [9, 15, 38, 47, 48] or a two-stage top-down fashion [28]. The bottom-up methods early align the vision and language at the local patch level, failing to explicitly model objects and their relations [28, 53]. Although top-down approaches explicitly extract object tracklets and select object tracklets matched with the language expression, their two-stage pipeline is complex and less efficient [53]. Recently, transformer-based frameworks [1, 53] have been proposed to use object queries to capture objects and their associations in an end-to-end manner. They generate object queries and utilize the transformer decoder to search objects for these object queries in every frame, as shown in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 190, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 190, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 190, + 713 + ], + "type": "text", + "content": "†Sibei Yang is the corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15466" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 545, + 217 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 217 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 217 + ], + "type": "image", + "image_path": "d2a7980f87e9de4412ff13fde2c04a413ae106c5127220478fc3632414e609da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 277 + ], + "type": "text", + "content": "Figure 2. Comparison of temporal modeling for previous works, naive attempt, and ours. Previous works (MTTR and ReferFormer) lack temporal modeling when decoding. Besides, object motion and interaction with other objects are crucial for referring video object segmentation. However, simply modeling temporal interaction with globally visible video-level decoding cannot achieve satisfactory results. Note that our alternating interaction between the global referent and object queries can dynamically capture spatial-temporal cross-modal reasoning over objects guided by the global referent information." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 289, + 115, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 115, + 301 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 115, + 301 + ], + "type": "text", + "content": "Figure 1a and b." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 290, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 290, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 290, + 532 + ], + "type": "text", + "content": "Although these end-to-end transformer-based frameworks have achieved impressive segmentation results, they still have limitations needed to be improved. First, their object queries are frame-level and independently responsible for the search for objects in each frame, which fails to capture temporal object motions and effective object tracking, as shown in Figure 2a and b. Therefore, we observe that they often fail to obtain temporally consistent predictions for target referents across an entire video, as shown in the left example in Figure 2. Second, their interactions between object queries are also solely performed per frame, leading to failure to model object-level spatial-temporal associations between objects. Therefore, they cannot correctly ground expressions that require cross-modal spatial-temporal reasoning over multiple objects. For example (see the right example in Figure 2), they mislocate the left camel as the target object because they lack to model the relation between the left camel in the current frame and the camel walking in the subsequent frame." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "In this paper, we aim to address these limitations by empowering the decoder with the capability of temporal modeling. A straightforward approach to decode objects with temporal information is to convert the frame-level decoding into video-level by inputting a sequence of object queries to search for objects in a sequence of frames like VisTR [49] for video instance segmentation, as shown in Figure 1b and 2e. However, this simple attempt fails to achieve satisfactory results while significantly increasing the computational cost and even underperforming existing query-based transformer frameworks. We further analyze the reasons: (1) The alignment between the natural language expression and the referent relies on overall objects' motions and temporal associations in the entire video. For example, the two camels can be distinguished only if the walking motion of" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 289, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 445 + ], + "type": "text", + "content": "the target camel is identified based on the entire video and aligned with the language expression. In contrast, the above naive attempt prematurely aligns language with fine-grained frame-level objects so that its attention is distracted from focusing accurately on the overall motions and relations of objects at the video level. (2) However, the precise localization and segmentation of target objects should go back to and rely more on every single frame because objects with different motions may cause them to have very different spatial locations in different frames. A similar observation is discussed by SeqFormer [54] for video instance segmentation, which suggests processing attention with each frame independently." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "Therefore, to address the challenge of aligning the language expression with objects' motions and temporal associations at the global video level but segmenting objects at the local frame level, we propose to maintain both local object queries and a global referent token. Global referent token captures the video-level referent information according to the language expression, while local object queries locate and segment objects with each frame. Furthermore, the object queries and the referent token are interacted to achieve spatial-temporal object information exchange through our well-designed temporal collection and distribution mechanisms, as shown in Figure 2f. Specifically, the temporal collection collects the referent information from object queries with spatial and visual attributes to temporal object motions to language semantics. In turn, the temporal distribution first dynamically distributes the referent token to every frame to extract the referent sequence. Then, the referent sequence interacts with object queries within each frame to achieve efficient spatial-temporal reasoning over objects. Note that the temporal collection and distribution alternately propagate information between object queries and the referent token to update each other." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15467" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 167 + ], + "type": "text", + "content": "Finally, we propose a Temporal Collection and Distribution network (TempCD) which integrates our novel temporal collection and distribution mechanisms into the query-based transformer framework. Without using sequence matching or post-processing during inference like the previous methods [1,53], Our TempCD can directly predict the segmentation result of every frame based on object queries in the frame and the referent token." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 169, + 266, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 169, + 266, + 180 + ], + "spans": [ + { + "bbox": [ + 59, + 169, + 266, + 180 + ], + "type": "text", + "content": "In summary, our main contributions are as follows," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 190, + 286, + 380 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 190, + 286, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 190, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 58, + 190, + 286, + 237 + ], + "type": "text", + "content": "- We introduce to maintain a global referent token and a sequence of local object queries parallelly to bridge the gap between video-level object alignment with language and frame-level object segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 238, + 286, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 238, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 58, + 238, + 286, + 285 + ], + "type": "text", + "content": "- We propose a novel collection-distribution mechanism for interacting between the referent token and object queries to capture spatial-temporal cross-modal reasoning over objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 285, + 286, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 285, + 286, + 332 + ], + "spans": [ + { + "bbox": [ + 58, + 285, + 286, + 332 + ], + "type": "text", + "content": "- We present an end-to-end Temporal Query Collection and Distribution (TempCD) for RVOS, which can directly predict the segmentation referent of every frame without any post-processing and sequence matching." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 333, + 286, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 333, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 58, + 333, + 286, + 380 + ], + "type": "text", + "content": "- Experiments on Ref-Youtu-VOS, Ref-DAVIS17, A2D-Sentences and JHMDB-Sentences datasets show that TempCD outperforms state-of-the-art methods on all benchmarks consistently and significantly." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 393, + 134, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 134, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 134, + 406 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 414, + 289, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 414, + 289, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 414, + 289, + 653 + ], + "type": "text", + "content": "Referring Image Segmentation (RIS) involves segmenting objects in images based on natural language expressions. Compared to Referring Video Object Segmentation (RVOS), RIS operates on individual images without temporal information. Previous works [3, 4, 26, 29, 42, 45, 57-59] focus on the joint modeling of vision and language. Various methods [6, 13, 14, 16, 20, 21, 35, 60, 64] are explored successively, such as using fusion operators like concatenation, ConvLSTM, attention mechanisms [46], and GNN [10] to obtain multimodal semantic feature maps. In addition, some studies attempt to decouple different components or key semantics from language, through explicit two-stage approaches [51, 66] or implicit attention modules [43, 62, 64, 65], to achieve a more fine-grained understanding. Recent models explore novel fusion techniques, including early fusion [8, 63], Linguistic Seed Encoder [23], and contrastive learning [50], to improve the performance of RIS. However, for RVOS, it requires not only the multimodal integration of language and image frames but also temporal modeling at the video level." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 289, + 715 + ], + "type": "text", + "content": "Referring Video Object Segmentation (RVOS) aims to segment the target object described in natural language from videos. Previous works in RVOS has primarily used two frameworks: bottom-up and top-down. The top-down framework [28] directly models the motion informa" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 358 + ], + "type": "text", + "content": "tion between segmentation masks of consecutive frames. However, this two-stage approach, which includes complex computational costs, is limited in terms of video input length. For the bottom-up framework, some approaches [9, 15, 38, 47, 48] directly apply RIS methods to construct multi-modal feature maps for referring segmentation of keyframes. They rely solely on replacing the traditional image backbone with a 3D temporal backbone for temporal modeling, which limits the performance of multiframes segmentation. To address this issue, URVOS [41] extends previous methods by building a memory bank that propagates language referent object information in the temporal dimension. Furthermore, URVOS introduces the RefYoutube-VOS [41] dataset, which provides segmentation annotations for each frame that require more efficient temporal modeling. In addition, recent works [7, 52, 68] adopt language guided fusion between temporal features and visual features to obtain more efficient temporal modeling. Recently, query-based methods [1, 53] with a transformer encoder-decoder frameworks [2, 5, 69] attempt to capture object-level information. They achieve higher performance by constructing a multimodal feature map of visual and linguistic information, or by using language-initialized queries with independently decoding with queries for each frame." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 360, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 647 + ], + "type": "text", + "content": "Temporal Modeling in Video Instance Segmentation (VIS). VIS aims to simultaneously track, segment and classify interest instances in a video. Similarly to RVOS, VIS requires obtaining instance information from every frame in the video and maintaining the temporal consistency of instances. Recent works focus on spatial and temporal modeling based on instance-level information. VisTR [49] employs an encoder-decoder transformer based on DETR, with a concatenated sequence of instance queries as the input to enable spatial-temporal interactions. The following IFC [17] and TeViT [61] explored a more efficient approach to temporal modeling in the encoder by introducing supplementary tokens for interaction across frames, with the decoder aligned with VisTR. Other following works [24, 56] take advantage of memories or queries from several previous frames to help instance segmentation in the current frame, extending to a online fashion. Recently, SeqFormer [54] proposes that the acquisition of frame-level instance information such as position needs to go back to each frame, i.e., the decoding process should be done independently for each frame. While these methods provide potential solutions for temporal modeling, the direct application to RVOS is not feasible, because of inconsistent object semantics over frames." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 657, + 362, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 362, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 362, + 670 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "The framework of our proposed TempCD is shown in Figure 3. We introduce a sequence of object queries to capture object information in every frame independently and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15468" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 521, + 305 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 521, + 305 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 521, + 305 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 521, + 305 + ], + "type": "image", + "image_path": "494aa9f57b9ad145567289aeec76eac8c793e0eeba4ed37fe906091c9e192b24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 308, + 547, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 547, + 354 + ], + "type": "text", + "content": "Figure 3. Overall framework of TempCD. TempCD simultaneously maintains a global referent token and a sequence of object queries, and they interact via the proposed collection-distribution mechanism. Specifically, the temporal collection collects global information for the referent token from object queries to the temporal motions to the language expression. In turn, the temporal distribution first distributes the referent token to the referent sequence and then performs efficient cross-frame reasoning between the referent sequence and object queries." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 366, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 288, + 437 + ], + "type": "text", + "content": "maintain a referent token to capture the global referent information for aligning with language expression. The referent token and object queries are alternately updated to stepwise identify the target object and better segment the target object in every frame through the proposed temporal collection and distribution mechanisms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 287, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 619 + ], + "type": "text", + "content": "Specifically, we first introduce encoders and the definition of object queries in Section 3.1. Then, we present the temporal collection that collects the referent information for the referent token from object queries to temporal object motions to the language expression (see Section 3.2). Next, we introduce the temporal distribution mechanism that distributes the global referent information to the referent sequence across frames to object queries in each frame (see Section 3.3). The collection-distribution mechanism explicitly captures object motions and spatial-temporal cross-modal reasoning over objects. Finally, we introduce segmentation heads and loss functions in Section 3.4. Note that as we explicitly maintain the referent token, we can directly identify the referent object in each frame without requiring sequence matching as in previous queries-based methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 623, + 216, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 216, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 216, + 635 + ], + "type": "text", + "content": "3.1. Encoders and Query Definition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Language Encoder. Following previous work [53], we adopt RoBERTa [32] as our language encoder. Given a language expression with " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " words, we extract its sentence feature " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " denotes channel dimensions. In addition, we obtain the output feature representation of each word before the last pooling layer as " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w \\in \\mathbb{R}^{L \\times C}" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "text", + "content": "Visual Backbone and Encoder. Following previous works [1, 52, 53, 61], we use ResNet-50 [11] as our visual backbone. We also follow recent work [1, 53] to evaluate our model using the temporal visual backbone network Video-Swin [33] in experiments. For an input video " + }, + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{H \\times W \\times 3 \\times T}" + }, + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "text", + "content": " frames, we first extract its visual feature maps and further construct multimodal feature maps " + }, + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{v_i\\}_{i=1}^T" + }, + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "text", + "content": " by fusing the visual feature maps with the word features of the expression " + }, + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_w" + }, + { + "bbox": [ + 304, + 366, + 545, + 474 + ], + "type": "text", + "content": " following [53]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "text", + "content": "Object Query Definition. Inspired by query-based detection and segmentation frameworks [2, 5, 69], we leverage object queries to represent object-level information. For the video with " + }, + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "text", + "content": " frames, we introduce a sequence of " + }, + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\times N" + }, + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "text", + "content": " object queries to represent the objects in every frame and follow [53] to use the language feature " + }, + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "text", + "content": " to initialize the object queries as " + }, + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} \\in \\mathbb{R}^{\\mathcal{T} \\times N \\times C}" + }, + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "text", + "content": ", where the number of objects and the dimension of channels are " + }, + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 474, + 545, + 581 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "content": "Referent Token Definition. In addition to adopting the object queries to represent objects in every frame, we define a referent token " + }, + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^g\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "content": " to represent the global information of the target object. As the language expression naturally indicates the referent information, we thus use the language feature " + }, + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "content": " to initialize the " + }, + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^g" + }, + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 659, + 524, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 524, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 524, + 673 + ], + "type": "text", + "content": "3.2. Temporal Collection from Object Queries" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "The temporal collection mechanism collects object motions from frames and aggregates object motions consistent with the language expression into the global referent token." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15469" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "We first perform motion collection across object queries (see Section 3.2.1) and then update the global referent token from the object motions and language expression (see Section 3.2.2)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 133, + 253, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 253, + 146 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 253, + 146 + ], + "type": "text", + "content": "3.2.1 Motion Collection across Object Queries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "text", + "content": "For simplicity of demonstration, we denote the object queries output from the " + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "inline_equation", + "content": "(l - 1)" + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "text", + "content": "-th layer as " + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{l - 1}" + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "text", + "content": " and input them into the " + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "text", + "content": "-th layer of TempCD. Note that we input the object queries " + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "text", + "content": " defined in Section 3.1 into the first layer of TempCD. We utilize " + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{l - 1}" + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "text", + "content": " to capture object information in each frame and integrate them to obtain temporal information such as object motion and action. Specifically, for " + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "text", + "content": "-th frame, we feed its object queries " + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "inline_equation", + "content": "q_{i} = \\mathcal{Q}_{l - 1}^{i} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "text", + "content": " and the multimodal features " + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "text", + "content": " into the DETR transformer decoder [2,69] to locate objects, where " + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{l - 1}^{i}" + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "text", + "content": " means the object queries of " + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 152, + 288, + 285 + ], + "type": "text", + "content": "-th frame. The computation proceeds as follows," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 294, + 287, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 294, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 107, + 294, + 287, + 310 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {Q}} _ {l} = \\left\\{\\operatorname {D E T R} \\left(q _ {i}, v _ {i}\\right) \\right\\} _ {i = 1} ^ {\\mathcal {T}}, \\tag {1}", + "image_path": "1a878d256b436d8009379e0aa003165d0c1e38ad4875111620703fd18b3ab48a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{DETR}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "text", + "content": " represents the DETR transformer decoder, and " + }, + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Q}}_l" + }, + { + "bbox": [ + 46, + 312, + 287, + 346 + ], + "type": "text", + "content": " denotes the object queries output by the decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "type": "text", + "content": "Next, to aggregate object motions across the video sequence, we combine objects information from queries of each frame. Specifically, we generate temporal weights " + }, + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{T \\times N \\times 1}" + }, + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "type": "text", + "content": " for each frame given the object queries " + }, + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Q}}_l" + }, + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "type": "text", + "content": ". This is done by employing a nonlinear layer followed by a softmax operator. Then, we use the temporal weights " + }, + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "type": "text", + "content": " to integrate the frame-level information and obtain object motions " + }, + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "type": "text", + "content": " through a gating mechanism. The integration is computed as follows," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 120, + 460, + 213, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 460, + 213, + 475 + ], + "spans": [ + { + "bbox": [ + 120, + 460, + 213, + 475 + ], + "type": "interline_equation", + "content": "G = \\operatorname {s o f t m a x} (\\hat {\\mathcal {Q}} _ {t} W),", + "image_path": "1118f11410e958e0c7ef0b009606421c689d35359f01980411436d2c4eb1d467.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 122, + 477, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 477, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 122, + 477, + 287, + 510 + ], + "type": "interline_equation", + "content": "M = \\sum_ {i = 1} ^ {\\mathcal {T}} \\left(G _ {i} \\odot \\hat {\\mathcal {Q}} _ {l} ^ {i}\\right), \\tag {2}", + "image_path": "ac10219929f8761ccae0ebdc6682d615c2b86171398c246d5266d7df324aee0a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{C \\times 1}" + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "text", + "content": " denotes the learnable weights of the linear layer, " + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "inline_equation", + "content": "G_{i} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "text", + "content": " is the weight for " + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "text", + "content": "-th frames in " + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\mathrm{softmax}(\\cdot)" + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "text", + "content": " represents the softmax operation along the temporal dimension, and " + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 515, + 287, + 564 + ], + "type": "text", + "content": " is the element-wise product." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 578, + 222, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 222, + 589 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 222, + 589 + ], + "type": "text", + "content": "3.2.2 Global Referent Token Collection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "text", + "content": "Next, we update the global referent token " + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{l-1}^g \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "inline_equation", + "content": "(l-1)" + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "text", + "content": "-th layer to referent token " + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_l^g \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "text", + "content": " by selecting the object motions " + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "text", + "content": " consistent with the language feature " + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "text", + "content": ". Note that we input the global referent token " + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^g" + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "text", + "content": " defined in Section 3.1 into the first layer of TempCD. We first calculate the cosine similarity between the language feature " + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "text", + "content": " and the object motions " + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "text", + "content": ". We obtain the similarity scores " + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "inline_equation", + "content": "S = \\{s_j\\}_{j=1}^N" + }, + { + "bbox": [ + 46, + 596, + 287, + 694 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 127, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "s _ {j} = \\cos \\left(\\mathcal {F} _ {s}, m _ {j}\\right), \\tag {3}", + "image_path": "498d5d4342ba6cc89e5b910bde5a26b430b5e8266920cbf617413ce28ff79d0e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " means the cosine similarity function and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "-th object's motion feature in " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". Next, we update the global referent token " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_g" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " as the motion feature of the object with the highest similarity score to the global referent token. To ensure the selection process is differentiable, we implement this selection through the Gumbel Softmax operator [18, 36]. The calculation is as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 341, + 165, + 480, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 165, + 480, + 178 + ], + "spans": [ + { + "bbox": [ + 341, + 165, + 480, + 178 + ], + "type": "interline_equation", + "content": "S _ {\\text {g u m b e l}} = \\operatorname {s o f t m a x} \\left(\\left(S + G _ {s}\\right) / \\tau\\right),", + "image_path": "8586ac3e62e4455a2b922615691a6a7e384772587cdfdc2f4905ecd8d791142b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 342, + 180, + 545, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 180, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 342, + 180, + 545, + 193 + ], + "type": "interline_equation", + "content": "S _ {\\text {o n e h o t}} = \\text {o n e h o t} \\left(\\arg \\max _ {N} \\left(S _ {\\text {g u m b e l}}\\right)\\right), \\tag {4}", + "image_path": "2d5604d98d78984a336eed4996d98cf01d66356183d7491ef0446fddaf1bd445.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 342, + 194, + 509, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 194, + 509, + 208 + ], + "spans": [ + { + "bbox": [ + 342, + 194, + 509, + 208 + ], + "type": "interline_equation", + "content": "S _ {\\text {r e f e r e n t}} = S _ {\\text {o n e h o t}} - \\mathrm {s g} \\left[ S _ {\\text {g u m b e l}} \\right] + S _ {\\text {g u m b e l}},", + "image_path": "1e577ca4126a612c103623a88bf8074d270973986adc5cf61dfec18969796c47.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\operatorname{softmax}(\\cdot)" + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "content": " represents the softmax operation along dimension " + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "inline_equation", + "content": "G_{s} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "content": " is sampled from the Gumbel(0,1) distribution, " + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "content": " is the temperature coefficient hyperparameter, and " + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}" + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "content": " is the stop gradient operation. Additionally, " + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\operatorname{argmax}_{N}" + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "content": " represents choosing the index with the highest similarity score, and onehot represents setting the highest-scoring item to 1 based on the selected index and setting the others to 0. The one-hot vector " + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "inline_equation", + "content": "S_{\\text{referent}} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "content": " indicates the selection from the object motions. The global referent token " + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^g \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "content": " is updated as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 330, + 344, + 545, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 344, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 330, + 344, + 545, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} _ {l} ^ {g} = \\operatorname {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(S _ {\\text {r e f e r e n t}} ^ {\\top} M\\right) + Q _ {l - 1} ^ {g}\\right), \\tag {5}", + "image_path": "fb2b24485f246e2cfe988f75c1572da1c139dd5190caf40369fe194e28949b33.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 367, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 545, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 367, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}(\\cdot)" + }, + { + "bbox": [ + 304, + 367, + 545, + 403 + ], + "type": "text", + "content": " denotes a three-layer linear layer with activation function, and LayerNorm " + }, + { + "bbox": [ + 304, + 367, + 545, + 403 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 304, + 367, + 545, + 403 + ], + "type": "text", + "content": " is the same as the LayerNorm in standard Transformer." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 403, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 462 + ], + "type": "text", + "content": "We obtain the updated global referent token " + }, + { + "bbox": [ + 304, + 403, + 545, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^g" + }, + { + "bbox": [ + 304, + 403, + 545, + 462 + ], + "type": "text", + "content": " that contains information on the referent object aligned with the language expression. Next, we feed it into the Temporal Distribution module to assist in distributing the referent information into each frame." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 471, + 533, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 533, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 533, + 483 + ], + "type": "text", + "content": "3.3. Temporal Distribution from Referent Token" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 489, + 545, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 545, + 585 + ], + "type": "text", + "content": "The temporal distribution aims to propagate the referent information to all frames and perform dynamically spatial-temporal reasoning over objects based on the referent information. First, we distribute the global referent token to each frame independently to extract the referent sequence across all frames (see Section 3.3.1). Next, we perform the efficient cross-frame reasoning between the referent sequence and object queries in every frame (see Section 3.3.2)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 599, + 490, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 490, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 490, + 610 + ], + "type": "text", + "content": "3.3.1 Global Referent Token Distribution" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 617, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 676 + ], + "type": "text", + "content": "Given the global referent token " + }, + { + "bbox": [ + 304, + 617, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_l^g" + }, + { + "bbox": [ + 304, + 617, + 545, + 676 + ], + "type": "text", + "content": ", we distribute the referent object information to every frame and further obtain the referent sequence across all frames, denoted by " + }, + { + "bbox": [ + 304, + 617, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_l = \\{r_i\\}_{i=1}^T" + }, + { + "bbox": [ + 304, + 617, + 545, + 676 + ], + "type": "text", + "content": ", which refers to the target object in each frame." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Specifically, we first compute the cosine similarity scores between the object queries " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Q}}_l = \\{q_i\\}_{i=1}^T" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " and the global referent token " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_l^g" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": ". Then we obtain the index " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " of the" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15470" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "object query with the highest similarity score to the global referent token for every frame " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": " independently. The computation is as follows," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 122, + 117, + 287, + 136 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 117, + 287, + 136 + ], + "spans": [ + { + "bbox": [ + 122, + 117, + 287, + 136 + ], + "type": "interline_equation", + "content": "s _ {i, j} = \\cos \\left(q _ {i, j}, Q _ {l} ^ {g}\\right), \\tag {6}", + "image_path": "d7482584d14f43a3109c41395d7a04c0d054f16ad8020213cbaf38298f05b351.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 123, + 133, + 211, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 133, + 211, + 146 + ], + "spans": [ + { + "bbox": [ + 123, + 133, + 211, + 146 + ], + "type": "interline_equation", + "content": "d _ {i} = \\operatorname {a r g m a x} _ {j} \\left\\{s _ {i, j} \\right\\},", + "image_path": "6d3f8b0a361d4364215cd95fc464fd03541250dfe54830be7c5d7c12f4902613.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "inline_equation", + "content": "q_{i,j}" + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "text", + "content": "-th object query in the " + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "text", + "content": "-th frame, and " + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "inline_equation", + "content": "s_{i,j}" + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "text", + "content": " is the corresponding similarity score. For every frame " + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "text", + "content": ", we select the object query " + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "text", + "content": " with the highest similarity score and fuse it with the global referent token " + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_l^g" + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "text", + "content": " to generate the referent sequence " + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_l = \\{r_i\\}_{i=1}^T" + }, + { + "bbox": [ + 46, + 149, + 287, + 211 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 217, + 287, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 217, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 75, + 217, + 287, + 232 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {l} = \\left\\{\\mathrm {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(q _ {i, d _ {i}} + \\mathcal {Q} _ {l} ^ {g}\\right)\\right) \\right\\} _ {i = 1} ^ {\\mathcal {T}}. \\tag {7}", + "image_path": "34890c39cff0a6546712b640ec55745ee3dbddc9b2e1ac728e28e00d1111f658.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 240, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 277 + ], + "type": "text", + "content": "We end up with referent sequence " + }, + { + "bbox": [ + 46, + 240, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_l\\in \\mathbb{R}^{T\\times C}" + }, + { + "bbox": [ + 46, + 240, + 287, + 277 + ], + "type": "text", + "content": " across all frames, which is used as crucial objects to interact with object queries in every frame independently." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 291, + 287, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 287, + 303 + ], + "type": "text", + "content": "3.3.2 Spatial-temporal Reasoning over Object Queries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 308, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 380 + ], + "type": "text", + "content": "We leverage the referent sequence " + }, + { + "bbox": [ + 46, + 308, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_l\\in \\mathbb{R}^{T\\times C}" + }, + { + "bbox": [ + 46, + 308, + 287, + 380 + ], + "type": "text", + "content": ", which encodes the referent object information across all frames, to interact with the object queries to achieve the cross-frame temporal reasoning over multiple objects and further distribute referent information to all object queries in all frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "text", + "content": "Specifically, for each object query " + }, + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "inline_equation", + "content": "q_{i,j}" + }, + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "text", + "content": "-th frame, we concatenate it with the objects of the referent sequence in all other frames to construct the cross-object query sequence " + }, + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "inline_equation", + "content": "C_{i,j} = \\{r_1,..,r_{i - 1},q_{i,j},r_{i + 1},..,r_{\\mathcal{T}}\\} \\in \\mathbb{R}^{\\mathcal{T}\\times C}" + }, + { + "bbox": [ + 46, + 381, + 287, + 453 + ], + "type": "text", + "content": ". The object-level interaction across frames is implemented by multi-head self attention, which is computed as follows," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 121, + 464, + 287, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 464, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 121, + 464, + 287, + 479 + ], + "type": "interline_equation", + "content": "\\hat {C} _ {i, j} = \\operatorname {M H S A} \\left(C _ {i, j}\\right), \\tag {8}", + "image_path": "dffd459c2df069b0ad0497ed9a66bf1db167a6d5a66d8e63373e2d7766d2466f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{MHSA}(\\cdot)" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": " denotes the multi-head self-attention mechanism. The " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": "-th updated object queries in " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "C_{i,j}^{\\hat{}}" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": " for every frame " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": " are concatenated as " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_l^f \\in \\mathbb{R}^{\\mathcal{T} \\times N \\times C}" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": " to input to next decoder layer. Note that our cross-frame interaction only requires interaction with referent sequences to object queries in each frame, which achieves efficient cross-frame interaction over multiple objects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 575, + 194, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 194, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 194, + 586 + ], + "type": "text", + "content": "3.4. Prediction Heads and Loss" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 594, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 640 + ], + "type": "text", + "content": "We integrate the referent sequence and object queries in each decoder layer to the prediction heads for predicting results. To simplify the presentation, we use the final layer for details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "We extract the referent sequence " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R} \\in \\mathbb{R}^{T \\times C}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " and combine them with the object queries " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "Q_{l}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " to obtain the output referent sequence " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " for prediction. Specifically, we use a gating mechanism to integrate the information from " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " into video-level features, which is similar to Eq 2. Next, we calculate cosine similarity between each sequence " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "C_{j,i}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " and the" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "language sentence feature " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", which is similar to Eq 3, to obtain the score for each object " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "q_{i,j}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ". We select the highest-scoring object " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "q_i^r" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " in each frame and combine it with " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " to obtain the final referent sequence " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}} \\in \\mathbb{R}^{T \\times C}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " of the referred object in each frame:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 348, + 142, + 545, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 142, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 348, + 142, + 545, + 156 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {R}} = \\left\\{\\mathrm {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(r _ {i} + q _ {i} ^ {r}\\right)\\right) \\right\\}. \\tag {9}", + "image_path": "537b13e26b0ff84304beaac0379549c975d2f8ed96884446b98abf023f86dbfd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 165, + 545, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 545, + 236 + ], + "type": "text", + "content": "Finally, we feed " + }, + { + "bbox": [ + 304, + 165, + 545, + 236 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}}" + }, + { + "bbox": [ + 304, + 165, + 545, + 236 + ], + "type": "text", + "content": " into the segmentation head to predict the segmentation results. Additionally, similar to [53] and [1], we use a box prediction head to regress the box coordinates for each query. We utilize Dice [37] loss and Focal [30] loss as the segmentation loss and GIOU [40] loss and L1 loss as the box loss." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 248, + 382, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 248, + 382, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 248, + 382, + 262 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 268, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 411 + ], + "type": "text", + "content": "Datasets. We conduct experiments on four benchmark datasets that are publicly available: Ref-Youtu-VOS [41], Ref-Davis-2017 [22], A2D-Sentences [9], and JHMDB-Sentences [9]. Ref-Youtu-VOS consists of 3978 videos and approximately 15K language descriptions. Ref-Davis-2017 contains approximately 90 videos. A2D-Sentences and JHMDB-Sentences [19, 55], originally focused on action recognition, have been expanded to incorporate language expression annotations. This expansion has yielded a total of 3782 videos for A2D-Sentences and 928 videos for JHMDB-Sentences, each accompanied by corresponding language descriptions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 412, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 543 + ], + "type": "text", + "content": "Implementation Details. Following [53], we pre-train our model on RefCOCO dataset [67]. We utilize a text encoder derived from RoBERTa [32], coupled with either ResNet50 [11] or Video-Swin serving [33] as our visual backbone. We train our model for 6 epochs with an initial learning rate of 1e-4 and the AdamW [34] optimizer. Consistent with prior works [1, 41, 53], we use the " + }, + { + "bbox": [ + 304, + 412, + 545, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 412, + 545, + 543 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 412, + 545, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 412, + 545, + 543 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 412, + 545, + 543 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 304, + 412, + 545, + 543 + ], + "type": "text", + "content": " metrics for evaluation on Ref-Youtu-VOS and Ref-Davis-2017 datasets, and Overall IoU and Mean IoU as evaluation metrics on A2D-Sentences and JHMDB-Sentences datasets." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 552, + 534, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 552, + 534, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 552, + 534, + 563 + ], + "type": "text", + "content": "4.1. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 570, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 616 + ], + "type": "text", + "content": "As shown in Table 1 and Table 2, we compare TempCD with state-of-the-art methods on four benchmarks. TempCD consistently outperforms state-of-the-art methods on all datasets." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Comparison of Ref-Youtube-VOS and Ref-Davis-2017 Datasets. Results for the Ref-Youtu-VOS and Ref-Davis-2017 datasets are presented in Table 1. With a standard ResNet-50 [11] visual backbone, our model exhibits improvements of " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}, \\mathcal{F}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " metrics, respectively, on the Ref-YoutuVOS dataset. Employing the advanced temporal visual backbone, VideoSwin-B [33], our approach consistently outperforms the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15471" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 98, + 71, + 496, + 231 + ], + "blocks": [ + { + "bbox": [ + 98, + 71, + 496, + 231 + ], + "lines": [ + { + "bbox": [ + 98, + 71, + 496, + 231 + ], + "spans": [ + { + "bbox": [ + 98, + 71, + 496, + 231 + ], + "type": "table", + "html": "
BackboneMethodRef-Youtube-VOSRef-Davis17
JFJ&FJFJ&F
ResNet-50CMSA [65]33.336.534.932.237.234.7
CMSA+RNN [65]34.838.136.436.943.540.2
URVOS [41]45.349.247.247.356.051.5
MLRL [52]48.451.049.750.155.452.7
LBDT [7]48.250.649.4--54.5
ReferFormer [53]54.856.555.655.861.358.5
Ours57.560.559.057.362.760.0
Video-Swin-TMTTR [1]54.056.655.3---
ReferFormer [53]58.060.959.4---
Ours60.564.062.359.365.062.2
Video-Swin-BReferFormer [53]61.364.662.958.164.161.1
Ours63.668.065.861.667.664.6
", + "image_path": "a84d4845037750c995256ff0f09d6ae81703b401571e17c14f099eab43a482d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 261, + 280, + 390 + ], + "blocks": [ + { + "bbox": [ + 93, + 238, + 499, + 250 + ], + "lines": [ + { + "bbox": [ + 93, + 238, + 499, + 250 + ], + "spans": [ + { + "bbox": [ + 93, + 238, + 499, + 250 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art models for RVOS on Ref-Youtu-VOS and Ref-Davis-2017 datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 261, + 280, + 390 + ], + "lines": [ + { + "bbox": [ + 55, + 261, + 280, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 280, + 390 + ], + "type": "table", + "html": "
BackboneMethodA2D-SJHMDB-S
oIoUmIoUoIoUmIoU
VGG16Hu et al. [12]47.435.054.652.8
I3DGavrilyuk et al. [9]53.642.154.154.2
ACAN [48]60.149.075.656.4
CMPC-V [31]65.357.361.661.7
Resnet-50ClawCraneNet [27]63.159.964.465.6
MMMMTBVS [68]67.355.861.961.3
LBDT [7]70.462.164.565.8
Ours76.668.670.669.6
", + "image_path": "04f742db8df15990ece20e0225ab6711af0123fc6ce9033186a5223e7d2ea33b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 398, + 287, + 420 + ], + "lines": [ + { + "bbox": [ + 47, + 398, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 398, + 287, + 420 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-art models for RVOS on A2D-Sentences (A2D-S) and JHMDB-Sentences (JHMDB-S)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 433, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 493 + ], + "type": "text", + "content": "previous state-of-the-art model [53] on the Ref-Davis-2017 dataset, achieving a " + }, + { + "bbox": [ + 46, + 433, + 287, + 493 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 46, + 433, + 287, + 493 + ], + "type": "text", + "content": " increase across the aforementioned metrics. Note that evaluations on the RefDavis-2017 dataset are performed using models trained on the RefYoutube-VOS dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 496, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 663 + ], + "type": "text", + "content": "Furthermore, we present a thorough evaluation of the effectiveness of our TempCD through subsequent comparisons: (1) We compare our model to MTTR [1], a pioneering method that adopts a query-based framework for RVOS, despite lacking temporal modeling in both the encoder and decoder stages. Our TempCD surpasses MTTR, achieving a " + }, + { + "bbox": [ + 46, + 496, + 287, + 663 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 46, + 496, + 287, + 663 + ], + "type": "text", + "content": " improvement in the " + }, + { + "bbox": [ + 46, + 496, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 46, + 496, + 287, + 663 + ], + "type": "text", + "content": " metric on the Ref-YoutuVOS dataset. This outcome underscores the heightened efficacy of our introduced temporal decoding modules. (2) Compared with LBDT [7], our TempCD achieves a significant " + }, + { + "bbox": [ + 46, + 496, + 287, + 663 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 46, + 496, + 287, + 663 + ], + "type": "text", + "content": " enhancement in the " + }, + { + "bbox": [ + 46, + 496, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 46, + 496, + 287, + 663 + ], + "type": "text", + "content": " metric. This result suggests that, unlike [7, 41, 52] that integrate temporal and visual features via a bottom-up framework, our proposed TempCD achieves more efficient temporal modeling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "Comparison on A2D-Sentences and JHMDB-Sentences. As presented in Table 2, our approach yields mean improvements of " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " in Overall IoU and " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " in Mean IoU, surpassing the predominant state-of-the-art method [7] for" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 263, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 361 + ], + "type": "text", + "content": "these two datasets. In contrast to Ref-Youtu-VOS, these datasets predominantly comprise segmentation annotations for keyframes that encapsulate actions. When compared to prior studies [9, 15, 38, 47, 48] that adopt temporal encoders to process video temporal information, our method consistently attains enhancements across these two datasets. This highlights the significance of explicitly capturing and integrating video-level temporal context into individual frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 369, + 400, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 400, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 400, + 382 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 388, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 436 + ], + "type": "text", + "content": "To validate the impact of modules in our model, we evaluate six variants of our TempCD on Ref-Youtu-VOS dataset. The results are shown in Table 3. All experiments are performed with Video-Swin-B as the visual backbone." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 437, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 628 + ], + "type": "text", + "content": "Baselines. (1) Our baseline model employs " + }, + { + "bbox": [ + 304, + 437, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 437, + 546, + 628 + ], + "type": "text", + "content": " sets of frame-level object queries (\"Local Query\"). These queries are concatenated sequentially and fed into the decoder, following a similar approach to VisTR [49]. The derived results (53.7% " + }, + { + "bbox": [ + 304, + 437, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 304, + 437, + 546, + 628 + ], + "type": "text", + "content": ", 56.3% " + }, + { + "bbox": [ + 304, + 437, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 437, + 546, + 628 + ], + "type": "text", + "content": ", and 55.0% " + }, + { + "bbox": [ + 304, + 437, + 546, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 304, + 437, + 546, + 628 + ], + "type": "text", + "content": ") suggest a misalignment between the natural language representations and the referent semantics at the individual frame level. (2) Another subsequent baseline is established by integrating a singular shared set of video-level object queries across all frames (\"Global Query\"). The global queries are performed an inner product with the feature map " + }, + { + "bbox": [ + 304, + 437, + 546, + 628 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 437, + 546, + 628 + ], + "type": "text", + "content": " of each frame, yielding segmentation outcomes. While this approach intrinsically aligns with language representations and achieves consistency between frames, it struggles to adapt to frame-level variability, resulting in a performance decline relative to the first baseline." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Query Collection and Distribution. (3) A natural enhancement for baseline (1) entails integrating cross-frame temporal attention on a referent sequence, which is selected from local queries by the similarity with language. This modification leads to a performance improvement of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". (4) Another endeavor to enhance the baseline (1) involves the introduction of an additional set of global queries. These" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15472" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 205 + ], + "type": "image", + "image_path": "67f0786b2f7e9b6f23e9232133078953265aa0ec4c670a7e8395adff828d2075.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 209, + 485, + 220 + ], + "lines": [ + { + "bbox": [ + 107, + 209, + 485, + 220 + ], + "spans": [ + { + "bbox": [ + 107, + 209, + 485, + 220 + ], + "type": "text", + "content": "Figure 4. Visualization for segmentation results of previous best-performing ReferFormer [53] and ours." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 232, + 282, + 332 + ], + "blocks": [ + { + "bbox": [ + 52, + 232, + 282, + 332 + ], + "lines": [ + { + "bbox": [ + 52, + 232, + 282, + 332 + ], + "spans": [ + { + "bbox": [ + 52, + 232, + 282, + 332 + ], + "type": "table", + "html": "
MethodJFJ&F
1Local Query53.756.355.0
2Global Query52.955.554.2
31+Cross-Frame56.259.157.6
41+Motion Collection58.161.559.8
53+460.764.462.6
65+Gumble-Softmax61.465.463.4
7Ours63.668.065.8
", + "image_path": "f499ba1123f5edf0bf3665ac67d62f8bff8c3048d7b48338bfd6749060e82275.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 340, + 265, + 351 + ], + "lines": [ + { + "bbox": [ + 67, + 340, + 265, + 351 + ], + "spans": [ + { + "bbox": [ + 67, + 340, + 265, + 351 + ], + "type": "text", + "content": "Table 3. Ablation Study on Ref-Youtu-VOS dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": "queries collect motion information from local queries and serve to distribute global context to local queries. Specifically, we utilize a Top-K strategy to select global information based on the similarity score with language. This facilitates implicit frame level interaction and, significantly, fosters semantic alignment between global queries and language expressions. This alignment contributes to a marked performance gain of " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": ". (5) We further incorporate both (3)'s explicit cross-frame temporal modeling and (4)'s motion collection and distribution, which delivers a " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " improvement. (6) We replace the Top-K selection, which does not allow gradient calculation, with Gumbel Softmax, improving " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": ". (7) Our complete model further enhances the vanilla cross-frame temporal attention mechanism. Instead of solely updating the referent sequence via temporal interaction, local object queries are also extended to incorporate temporal context information from referent sequence, facilitating cross-frame interaction over multiple objects. The full decoder of our proposed method achieves the performance of " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "65.8\\%" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " in terms of " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 610, + 132, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 132, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 132, + 623 + ], + "type": "text", + "content": "4.3. Visualization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Figure 4 visualizes several qualitative results. The referring expression in (a) describes the motion of a white duck, a pivotal feature that distinguishes it from similar ducks. Our proposed TempCD effectively captures the motion of the referent which is aligned with the language expression and enables accurate localization and segmentation of the specific duck. In contrast, Referformer fails to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 234, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 547, + 498 + ], + "type": "text", + "content": "locate the referent object, primarily attributed to a semantic incongruity between the expression and the frame-level queries. In the instance of (b), the deer's corresponding action appears in certain frames but lacks uniform presence. Our approach successfully captures its global semantics including motions and aligns it seamlessly with the referring expression, bridging the gap between local semantics and the referring expression via the Collection and Distribution mechanisms. Conversely, ReferFormer identifies the referent accurately in specific frames that correspond to the specified action, yet exhibits errors in others due to a noticeable lack of temporal context. All these observations collectively underscore the value of temporal interaction in fostering refined segmentation across frames. Besides, (c) illustrates that our method can still precisely locate and segment referent objects in complex scenes. The second frame presents a scenario of occlusion, wherein the referent is obscured by another bicycle. Our approach detects the lack of correspondence between visible entities in a frame and the specified referent sequence, facilitated by cross-frame interactions between the visible objects and those delineated by the referent sequence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 518, + 379, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 518, + 379, + 532 + ], + "spans": [ + { + "bbox": [ + 306, + 518, + 379, + 532 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 542, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 613 + ], + "type": "text", + "content": "This paper proposes an end-to-end Temporal Query Collection and Distribution (TempCD) network for referring video object segmentation, which maintains object queries and the referent token and achieves alternating interaction between them via the proposed novel temporal collection-distribution mechanism." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Acknowledgment: This work was supported by the National Natural Science Foundation of China (No.62206174), Shanghai Pujiang Program (No.21PJ1410900), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (ShangHAI), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University), and Shanghai Engineering Research Center of Intelligent Vision and Imaging." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15473" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-end referring video object segmentation with multimodal transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4985-4995, 2022. 1, 3, 4, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 213 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 3, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 214, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 288, + 267 + ], + "type": "text", + "content": "[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7454-7463, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "text", + "content": "[4] Yi-Wen Chen, Yi-Hsuan Tsai, Tiantian Wang, Yen-Yu Lin, and Ming-Hsuan Yang. Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 288, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 356 + ], + "type": "text", + "content": "[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 287, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 411 + ], + "type": "text", + "content": "[6] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "text", + "content": "[7] Zihan Ding, Tianrui Hui, Junshi Huang, Xiaoming Wei, Jizhong Han, and Si Liu. Language-bridged spatial-temporal interaction for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4964–4973, 2022. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 469, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 288, + 524 + ], + "type": "text", + "content": "[8] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15506–15515, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 525, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 288, + 568 + ], + "type": "text", + "content": "[9] Kirill Gavrilyuk, Amir Ghodrati, Zhenyang Li, and Cees GM Snoek. Actor and action video segmentation from a sentence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5958-5966, 2018. 1, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "text", + "content": "[10] Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. In Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., volume 2, pages 729-734. IEEE, 2005. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 615, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 615, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 615, + 287, + 658 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 659, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 659, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 659, + 287, + 713 + ], + "type": "text", + "content": "[12] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 108-124. Springer, 2016. 1, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[13] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4424-4433, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "type": "text", + "content": "[14] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10488-10497, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 197, + 547, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 261 + ], + "type": "text", + "content": "[15] Tianrui Hui, Shaofei Huang, Si Liu, Zihan Ding, Guanbin Li, Wenguan Wang, Jizhong Han, and Fei Wang. Collaborative spatial-temporal modeling for language-queried video actor segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4187–4196, 2021. 1, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 264, + 547, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 547, + 329 + ], + "type": "text", + "content": "[16] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 59-75. Springer, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 331, + 546, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 546, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 546, + 374 + ], + "type": "text", + "content": "[17] Sukjun Hwang, Miran Heo, Seoung Wug Oh, and Seon Joo Kim. Video instance segmentation using inter-frame communication transformers. Advances in Neural Information Processing Systems, 34:13352-13363, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 376, + 546, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 546, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 546, + 408 + ], + "type": "text", + "content": "[18] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 411, + 546, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 546, + 454 + ], + "type": "text", + "content": "[19] Hueihan Jhuang, Juergen Gall, Silvia Zuffi, Cordelia Schmid, and Michael J Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision, pages 3192-3199, 2013. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 456, + 546, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 546, + 510 + ], + "type": "text", + "content": "[20] Yang Jiao, Zequn Jie, Weixin Luo, Jingjing Chen, Yu-Gang Jiang, Xiaolin Wei, and Lin Ma. Two-stage visual cues enhancement network for referring image segmentation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1331-1340, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 512, + 546, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 546, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 546, + 566 + ], + "type": "text", + "content": "[21] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tieniu Tan. Locate then segment: A strong pipeline for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9858-9867, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 568, + 547, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 547, + 632 + ], + "type": "text", + "content": "[22] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part IV 14, pages 123–141. Springer, 2019. 1, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 635, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 547, + 689 + ], + "type": "text", + "content": "[23] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18154, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 692, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 546, + 713 + ], + "type": "text", + "content": "[24] Rajat Koner, Tanveer Hannan, Suprosanna Shit, Sahand Sharifzadeh, Matthias Schubert, Thomas Seidl, and Volker" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15474" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "Tresp. Instanceformer: An online video instance segmentation framework. arXiv preprint arXiv:2208.10547, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "text", + "content": "[25] Bowen Li, Xiaojuan Qi, Thomas Lukasiewicz, and Philip HS Torr. Manigan: Text-guided image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7880-7889, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "type": "text", + "content": "[26] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 239 + ], + "type": "text", + "content": "[27] Chen Liang, Yu Wu, Yawei Luo, and Yi Yang. Clawranenet: Leveraging object-level relation for text-based video segmentation. arXiv preprint arXiv:2103.10702, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "text", + "content": "[28] Chen Liang, Yu Wu, Tianfei Zhou, Wenguan Wang, Zongxin Yang, Yunchao Wei, and Yi Yang. Rethinking cross-modal interaction from a top-down perspective for referring video object segmentation. arXiv preprint arXiv:2106.01061, 2021. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 293, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 337 + ], + "type": "text", + "content": "[29] Liang Lin, Pengxiang Yan, Xiaogqian Xu, Sibei Yang, Kun Zeng, and Guanbin Li. Structured attention network for referring image segmentation. IEEE Transactions on Multimedia, 24:1922-1932, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 337, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 381 + ], + "type": "text", + "content": "[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 381, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 381, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 381, + 287, + 435 + ], + "type": "text", + "content": "[31] Si Liu, Tianrui Hui, Shaofei Huang, Yunchao Wei, Bo Li, and Guanbin Li. Cross-modal progressive comprehension for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4761-4775, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "text", + "content": "[32] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle-moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 525 + ], + "type": "text", + "content": "[33] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. arXiv preprint arXiv:2106.13230, 2021. 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 548 + ], + "type": "text", + "content": "[34] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "text", + "content": "[35] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[36] Chris J Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables. arXiv preprint arXiv:1611.00712, 2016. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[37] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "text", + "content": "[38] Ke Ning, Lingxi Xie, Fei Wu, and Qi Tian. Polar relative positional encoding for video-language segmentation. In *IJ-CAI*, volume 9, page 10, 2020. 1, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 106, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 547, + 172 + ], + "type": "text", + "content": "[39] Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and Anton van den Hengel. Reverie: Remote embodied visual referring expression in real indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9982-9991, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 546, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 546, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 546, + 237 + ], + "type": "text", + "content": "[40] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 238, + 546, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 238, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 238, + 546, + 304 + ], + "type": "text", + "content": "[41] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Urvos: Unified referring video object segmentation network with a large-scale benchmark. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16, pages 208–223. Springer, 2020. 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 304, + 546, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 546, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 546, + 349 + ], + "type": "text", + "content": "[42] Cheng Shi and Sibei Yang. Spatial and visual perspectivetaking via view rotation and relation reasoning for embodied reference understanding. In European Conference on Computer Vision, pages 201-218. Springer, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 349, + 546, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 546, + 392 + ], + "type": "text", + "content": "[43] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 38-54, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 392, + 546, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 392, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 392, + 546, + 426 + ], + "type": "text", + "content": "[44] G Sreenu and Saleem Durai. Intelligent video surveillance: a review through deep learning techniques for crowd analysis. Journal of Big Data, 6(1):1-27, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 426, + 546, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 546, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 546, + 480 + ], + "type": "text", + "content": "[45] Jiajin Tang, Ge Zheng, Cheng Shi, and Sibei Yang. Contrastive grouping with transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23570-23580, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 481, + 546, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 481, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 481, + 546, + 525 + ], + "type": "text", + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 525, + 546, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 546, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 546, + 580 + ], + "type": "text", + "content": "[47] Hao Wang, Cheng Deng, Fan Ma, and Yi Yang. Context modulated dynamic networks for actor and action video segmentation with language queries. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12152-12159, 2020. 1, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 580, + 546, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 546, + 635 + ], + "type": "text", + "content": "[48] Hao Wang, Cheng Deng, Junchi Yan, and Dacheng Tao. Asymmetric cross-guided attention network for actor and action video segmentation from natural language query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3939-3948, 2019. 1, 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 546, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 546, + 691 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 546, + 691 + ], + "type": "text", + "content": "[49] Yuqing Wang, Zhaoliang Xu, Xinlong Wang, Chunhua Shen, Baoshan Cheng, Hao Shen, and Huaxia Xia. End-to-end video instance segmentation with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8741-8750, 2021. 2, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 546, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 546, + 715 + ], + "type": "text", + "content": "[50] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15475" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "text", + "content": "driven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 160 + ], + "type": "text", + "content": "[51] Chenyun Wu, Zhe Lin, Scott Cohen, Trung Bui, and Subhransu Maji. Phrasescut: Language-based image segmentation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10216-10225, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 216 + ], + "type": "text", + "content": "[52] Dongming Wu, Xingping Dong, Ling Shao, and Jianbing Shen. Multi-level representation learning with semantic alignment for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4996-5005, 2022. 3, 4, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 270 + ], + "type": "text", + "content": "[53] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4984, 2022. 1, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 271, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 288, + 315 + ], + "type": "text", + "content": "[54] Junfeng Wu, Yi Jiang, Wenqing Zhang, Xiang Bai, and Song Bai. Seqformer: a frustratingly simple model for video instance segmentation. arXiv preprint arXiv:2112.08275, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "type": "text", + "content": "[55] Chenliang Xu, Shao-Hang Hsieh, Caiming Xiong, and Jason J Corso. Can humans fly? action understanding with multiple classes of actors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2264-2273, 2015. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 403 + ], + "type": "text", + "content": "[56] Zhujun Xu and Damien Vivet. Instance sequence queries for video instance segmentation with transformers. Sensors, 21(13):4507, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 449 + ], + "type": "text", + "content": "[57] Sibei Yang, Guanbin Li, and Yizhou Yu. Cross-modal relationship inference for grounding referring expressions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4145-4154, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 288, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 492 + ], + "type": "text", + "content": "[58] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4644-4653, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 493, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 288, + 548 + ], + "type": "text", + "content": "[59] Sibei Yang, Guanbin Li, and Yizhou Yu. Propagating over phrase relations for one-stage visual grounding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 589-605. Springer, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "type": "text", + "content": "[60] Sibei Yang, Guanbin Li, and Yizhou Yu. Relationship-embedded representation learning for grounding referring expressions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2765-2779, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 593, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 288, + 656 + ], + "type": "text", + "content": "[61] Shusheng Yang, Xinggang Wang, Yu Li, Yuxin Fang, Jiemin Fang, Wenyu Liu, Xun Zhao, and Ying Shan. Temporally efficient vision transformer for video instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2885-2895, 2022. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[62] Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11266-11275, 2021. 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 440 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[63] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[64] Linwei Ye, Zhi Liu, and Yang Wang. Dual convolutional LSTM network for referring image segmentation. IEEE Transactions on Multimedia, 22(12):3224-3235, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 217 + ], + "type": "text", + "content": "[65] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10502–10511, 2019. 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 218, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 218, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 218, + 545, + 274 + ], + "type": "text", + "content": "[66] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1307-1315, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 274, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 339 + ], + "type": "text", + "content": "[67] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 340, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 340, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 340, + 545, + 396 + ], + "type": "text", + "content": "[68] Wangbo Zhao, Kai Wang, Xiangxiang Chu, Fuzhao Xue, Xinchao Wang, and Yang You. Modeling motion with multimodal features for text-based video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11737-11746, 2022. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 396, + 545, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 396, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 308, + 396, + 545, + 440 + ], + "type": "text", + "content": "[69] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3, 4, 5" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15476" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/553f0662-4b8d-4936-ab47-ad291ff03dd2_content_list.json b/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/553f0662-4b8d-4936-ab47-ad291ff03dd2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..efd1a3330d0f410d9598363628133e489cb51f3b --- /dev/null +++ b/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/553f0662-4b8d-4936-ab47-ad291ff03dd2_content_list.json @@ -0,0 +1,1555 @@ +[ + { + "type": "text", + "text": "Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction", + "text_level": 1, + "bbox": [ + 94, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhuofan Zong $^{1*}$ Dongzhi Jiang $^{1,2*}$ Guanglu Song $^{1}$ Zeyue Xue $^{3}$ Jingyong Su $^{4}$ Hongsheng Li $^{2,5,6\\dagger}$ Yu Liu $^{1\\dagger}$ SenseTime Research 2CUHK MMLAB 3HKU 4HITSZ CPII 6Shanghai AI Laboratory", + "bbox": [ + 212, + 202, + 754, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zongzhuofan, jdzcarr7, liuyuisanai}@gmail.com hsli@ee.cuhk.edu.hk", + "bbox": [ + 197, + 277, + 771, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a new paradigm, named Historical Object Prediction (HoP) for multi-view 3D detection to leverage temporal information more effectively. The HoP approach is straightforward: given the current timestamp $t$ , we generate a pseudo Bird's-Eye View (BEV) feature of timestamp $t-k$ from its adjacent frames and utilize this feature to predict the object set at timestamp $t-k$ . Our approach is motivated by the observation that enforcing the detector to capture both the spatial location and temporal motion of objects occurring at historical timestamps can lead to more accurate BEV feature learning. First, we elaborately design short-term and long-term temporal decoders, which can generate the pseudo BEV feature for timestamp $t-k$ without the involvement of its corresponding camera images. Second, an additional object decoder is flexibly attached to predict the object targets using the generated pseudo BEV feature. Note that we only perform HoP during training, thus the proposed method does not introduce extra overheads during inference. As a plug-and-play approach, HoP can be easily incorporated into state-of-the-art BEV detection frameworks, including BEVFormer and BEVDet series. Furthermore, the auxiliary HoP approach is complementary to prevalent temporal modeling methods, leading to significant performance gains. Extensive experiments are conducted to evaluate the effectiveness of the proposed HoP on the nuScenes dataset. We choose the representative methods, including BEVFormer and BEVDet4D-Depth to evaluate our method. Surprisingly, HoP achieves $68.5\\%$ NDS and $62.4\\%$ mAP with ViT-L on nuScenes test, outperforming all the 3D object detectors on the leaderboard. Codes are available at https://github.com/Sense-X/HoP.", + "bbox": [ + 75, + 359, + 473, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79d6afd5fdef55323ed71adc3b5bd4ce75cc18dc927d8a2bb56fee23b52bb689.jpg", + "image_caption": [ + "Figure 1: Incorporating HoP into 3D object detector with other temporal fusion methods [8, 7, 32, 12, 24]. Our HoP is plug-and-play and complementary to them." + ], + "image_footnote": [], + "bbox": [ + 504, + 327, + 898, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 614, + 632, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camera-only 3D detection from multi-view images is a challenging task for autonomous driving and has received increasing attention recently. Introducing the Bird's-Eye View (BEV) representation [25, 14, 8, 30, 18, 2] with the temporal features aggregation has become the superior design manner for camera-only 3D perception with multi-camera images. Based on this, a series of detectors [7, 15, 12, 24, 31, 19, 17] delve into the elaborated temporal information fusion methods and achieve significant breakthroughs. They consider the regions in 3D space and aggregate the image features corresponding to these hypothesis locations from multiple timestamps. These outstanding advances reveal the great potential of temporal information in camera-only 3D detection.", + "bbox": [ + 496, + 642, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Beyond these temporal fusion mechanisms, in this paper, we propose a new paradigm for leveraging temporal information to enhance the temporal modeling ability of", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Work done during internship at SenseTime Research. † Corresponding authors.", + "bbox": [ + 76, + 875, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3781", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the 3D object detector. It's a temporal-based auxiliary task only adopted during training stage, namely Historical Object Prediction (HoP). Our approach is motivated by the observation that enforcing the detector to capture both the spatial location and temporal motion of objects occurring at historical timestamps can lead to more accurate BEV feature learning.", + "bbox": [ + 78, + 90, + 467, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, given the current timestamp $t$ , we generate a pseudo BEV feature of timestamp $t - k$ from its adjacent frames and utilize this feature to predict the object set at $t - k$ . First, we elaborately design short-term and long-term temporal decoders, which can generate the pseudo BEV feature for timestamp $t - k$ without the involvement of its corresponding camera images. Thanks to the marginal temporal difference between two adjacent frames, the short-term temporal decoder process two adjacent frames of timestamp $t - k$ to provide spatial semantics of objects. The long-term decoder captures long-term motion of the whole frame sequence and contributes to better object motion estimation, which is complementary to spatial localization of the short-term branch. Subsequently, an additional object decoder is flexibly attached to predict the object targets using the generated pseudo BEV feature. Note that we only perform HoP during training, thus the proposed method does not introduce extra overheads during inference. As a plug-and-play approach, HoP can be flexibly incorporated into the state-of-the-art BEV detection frameworks, including BEVFormer [15] and BEVDet series [8, 7, 14]. Furthermore, the auxiliary HoP approach is complementary to prevalent temporal modeling methods, leading to significant gains.", + "bbox": [ + 80, + 198, + 467, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments are conducted to evaluate the effectiveness of the proposed HoP on the nuScenes dataset [1]. We choose the representative methods, including BEVFormer and BEVDet4D-Depth [7, 14] to evaluate our method. To be specific, we obtain $55.8\\%$ NDS with ResNet-101-DCN and $60.3\\%$ NDS with VoVNet-99 when evaluating HoP on nuScenes val. Surprisingly, HoP achieves $68.5\\%$ NDS and $62.4\\%$ mAP with ViT-L [4] on nuScenes test, surpassing all the 3D object detectors on the leaderboard by a large margin.", + "bbox": [ + 80, + 545, + 467, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In conclusion, our contributions can be summarized as follows:", + "bbox": [ + 80, + 696, + 467, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a novel temporal enhanced training scheme, namely Historical Object Prediction (HoP), to encourage more accurate BEV feature learning. HoP can force the model to capture the spatial semantics and temporal motion of objects in the historical frame during training.", + "bbox": [ + 96, + 737, + 467, + 828 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We design a temporal decoder that consists of a short-term decoder and a long-term decoder to provide reliable spatial localization and accurate motion estimation of objects.", + "bbox": [ + 96, + 839, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We equip the competitive 3D object detector baselines with our approach and yield significant improvements on the nuScenes dataset. Surprisingly, HoP with ViT-L achieves $68.5\\%$ NDS and $62.4\\%$ mAP on nuScenes test, establishing the new state-of-the-art performance.", + "bbox": [ + 522, + 90, + 888, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 503, + 181, + 648, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-view 3D Object Detection", + "text_level": 1, + "bbox": [ + 503, + 205, + 777, + 222 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Modern multi-view methods for 3D object detection could be mainly categorized into two branches, LSS-based [25] and query-based methods. We will introduce how these two methods differ in feature aggregation and leave the temporal modeling part in the next section.", + "bbox": [ + 503, + 229, + 888, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "BEVDet [8] is a representative method in LSS-based methods. Following the Lift-Splat-Shoot paradigm, the method first explicitly estimates depth for every image pixel, then lifts the 2D features to 3D voxels according to the depth, and finally splats 3D features to BEV features and conducts object detection on it. BEVDepth [14] further improves the view transformation module with explicit depth supervision.", + "bbox": [ + 503, + 306, + 888, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Query-based methods typically employ learnable queries to aggregate 2D image features by attention [27, 37] mechanism. DETR3D [30] utilizes object queries for predicting 3D positions and projects them back to 2D coordinates to obtain the corresponding features. Graph-DETR3D [2] and Sparse4D [17] enhance it respectively with a learnable 3D graph and sparse 4D sampling. PETR [18] encodes 3D position into image features and therefore directly query with global 2D features. BEVFormer [15] leverages grid-shaped BEV queries to interact with 2D features by deformable attention; on its basis, BEVFormerv2 [32] introduces perspective supervision. Besides, PolarFormer and Ego3RT [22, 9] advocates modeling BEV queries in polar coordinates to fit the real-world scenario.", + "bbox": [ + 503, + 428, + 888, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Temporal Modeling for Multi-view 3D Object Detection", + "text_level": 1, + "bbox": [ + 503, + 648, + 888, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The motion information in autonomous driving scenes has been increasingly explored to utilize the temporal cues for improving detection performance in recent 3D perception frameworks. BEVFormer [15] proposes the temporal self-attention mechanism to dynamically fuse the previous BEV features by deformable attention [37] in an RNN manner. BEVDet4D [7] introduces the temporal modeling to lift BEVDet [8] to spatial-temporal 4D space. The 3D position embedding (3D PE) in PETR [18] is extended by PETRv2 [19] with the temporal alignment. BEVStereo [12] and STS [31] both leverage temporal views for constructing multi-view stereo by an effective temporal stereo method. SOLOFusion [24] fully exploits the synergy of short-term and long-term temporal information that is highly comple", + "bbox": [ + 503, + 689, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3782", + "bbox": [ + 483, + 946, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/96b0d700a01eee3b95c2722bc1a9465ce4b1bb3630d51fb6e32d0da7d3b79612.jpg", + "image_caption": [ + "Figure 2: Framework of our Historical Object Prediction (HoP). The auxiliary branches are discarded during evaluation. The symbol © denotes feature concatenation and fusion." + ], + "image_footnote": [], + "bbox": [ + 122, + 90, + 848, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mentary as well as achieving state-of-the-art performance. BEVFormerV2 [32] uses the bidirectional temporal encoder to utilize both history and future BEV features, improving the performance by a large margin. In contrast with aforementioned works, we provide a novel temporal enhanced training paradigm to improve the temporal modeling without additional inference costs.", + "bbox": [ + 75, + 391, + 472, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 508, + 169, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 76, + 534, + 272, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, HoP can be easily incorporated into many multi-view 3D object detectors, e.g., BEVFormer and BEVDet4D-Depth. We transform the extrinsic parameters of previous frames to the coordinate system of the current ego to perform temporal alignment following BEVDet.", + "bbox": [ + 75, + 556, + 468, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Historical Object Prediction", + "text_level": 1, + "bbox": [ + 76, + 642, + 330, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pipeline. The architecture of our Historical Object Prediction (HoP) is composed of a temporal decoder $\\mathcal{T}$ and an object decoder $\\hat{\\mathcal{D}}$ . Given the BEV feature sequence $\\{\\mathbf{B}_{t-N}, \\cdots, \\mathbf{B}_t\\}$ that consists of $N$ historical BEV features and the current BEV feature, we denote the corresponding 3D ground truths as $\\{\\mathbf{G}_{t-N}, \\cdots, \\mathbf{G}_t\\}$ . If the BEV feature at timestamp $t-k$ is discarded, we use the remaining BEV feature sequence $\\{\\mathbf{B}_{t-N}, \\cdots, \\mathbf{B}_t\\} - \\{\\mathbf{B}_{t-k}\\}$ (denoted as $\\mathbf{B}^{rem}$ ) to predict the 3D objects at timestamp $t-k$ . To be specific, we adopt $\\mathcal{T}$ and $\\hat{\\mathcal{D}}$ to transform the remaining BEV features to obtain the 3D predictions $\\hat{\\mathbf{P}}_{t-k}$ :", + "bbox": [ + 75, + 664, + 468, + 832 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {P}} _ {t - k} = \\hat {\\mathcal {D}} \\left(\\mathcal {T} \\left(\\left\\{\\mathbf {B} _ {t - N}, \\dots , \\mathbf {B} _ {t} \\right\\} - \\left\\{\\mathbf {B} _ {t - k} \\right\\}\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 840, + 468, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The temporal decoder $\\mathcal{T}$ that models both short-term and long-term information is developed to reconstruct the BEV", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "feature at timestamp $t - k$ . The object decoder $\\hat{\\mathcal{D}}$ is attached to generate the predictions $\\hat{\\mathbf{P}}_{t - k}$ using the pseudo BEV feature. Considering the ego-motion and temporal alignment, we transform 3D coordinates of $\\mathbf{G}_{t - k}$ into the coordinate system of current frame $t$ (denoted as $\\hat{\\mathbf{G}}_{t - k}$ ). The final training objective is to minimize the differences between $\\hat{\\mathbf{P}}_{t - k}$ and $\\hat{\\mathbf{G}}_{t - k}$ .", + "bbox": [ + 496, + 390, + 890, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The input to the temporal decoder is the remaining set of BEV features consisting of (i) short-term BEV features $\\{\\mathbf{B}_{t - k - 1},\\mathbf{B}_{t - k + 1}\\}$ , and (ii) long-term BEV features $\\{\\mathbf{B}_{t - N},\\dots ,\\mathbf{B}_t\\} -\\{\\mathbf{B}_{t - k}\\}$ . We first add temporal positional embeddings to BEV features in this full set. Our temporal decoder reconstructs the feature $\\hat{\\mathbf{B}}_{t - k}$ by utilizing the spatial semantics and temporal cues of the BEV feature sequence. More specifically, we present a temporal decoder that consists of two separate branches with their special expertise to capture short-term and long-term information. The learned strong spatial-temporal representation can help the network better estimate the object location in frame at timestamp $t - k$ .", + "bbox": [ + 496, + 497, + 892, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Short-term temporal decoder. Thanks to the high temporal correlation between adjacent frames, the short-term temporal decoder only operates on the adjacent BEV feature set $\\{\\mathbf{B}_{t - k - 1},\\mathbf{B}_{t - k + 1}\\}$ (denoted as $\\mathbf{B}^{adj}$ ), building a detailed spatial representation in BEV space. We first define a grid-shaped learnable short-term BEV queries $\\mathbf{Q}_{t - k}^{short}\\in$ $\\mathbb{R}^{H\\times W\\times C}$ , where $H$ and $W$ refer to the spatial shape of the BEV plane. The short-term BEV queries aggregate the spatial information and model the short-term motion from $\\mathbf{B}_{t - k - 1}$ and $\\mathbf{B}_{t - k + 1}$ through the short-term temporal attention, which can be formulated as:", + "bbox": [ + 496, + 693, + 893, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {B}} _ {t - k} ^ {s h o r t} = \\sum_ {\\mathbf {V} \\in \\mathbf {B} ^ {a d j}} \\operatorname {D e f o r m A t t n} \\left(\\mathbf {Q} _ {t - k, p} ^ {s h o r t}, p, \\mathbf {V}\\right), \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 869, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3783", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p$ is the spatial index in the BEV plane, DeformAttn( $q, p, x$ ) refers to the deformable attention [37] with query $q$ , reference point $p$ and input features $x$ . We further feed $\\hat{\\mathbf{B}}_{t-k}^{short}$ into a feed-forward network [27] and obtain the output of this short-term branch. However, it is still challenging to precisely construct the temporal relations of the same objects between $\\mathbf{B}_{t-k}$ and other BEV features with only two adjacent frames.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Long-term temporal decoder. The long-term temporal processes the whole remaining BEV set to perceive the motion clues over long-term history, which increases the localization potential [24] and contributes to more accurate localization. Therefore, these two branches in the temporal decoder are complementary to each other. As for long-term motion, it is intuitive that we only focus on the spatial motion of the same objects in the bird's eye view, ignoring the height information of objects. For most 3D object detectors with 2D BEV features, e.g., BEVFormer and BEVDet, the height information has been flattened to the feature embeddings. Accordingly, we first employ a channel reduction operation to the input set $\\mathbf{B}^{rem}$ to prune the height information and achieve better training efficiency. Given the learnable long-term BEV queries $\\mathbf{Q}_{t - k}^{long}\\in \\mathbb{R}^{H\\times W\\times C / r}$ and reduction layer with parameter $\\mathbf{W}^r\\in \\mathbb{R}^{C\\times C / r}$ , we can capture the long-term dependencies as:", + "bbox": [ + 75, + 212, + 473, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {B}} _ {t - k} ^ {l o n g} = \\sum_ {\\mathbf {V} \\in \\mathbf {B} ^ {r e m}} \\mathrm {D e f o r m A t t n} (\\mathbf {Q} _ {t - k, p} ^ {l o n g}, p, \\mathbf {V W} ^ {r}), (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 479, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $r$ denotes the reduction ratio and is set as 4 by default. After the feed-forward network, the long-term decoder output the BEV features. Finally, we concatenate the short-term and long-term BEV features and perform feature fusion by $3 \\times 3$ convolution.", + "bbox": [ + 75, + 522, + 468, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object decoder. The reconstructed BEV feature $\\hat{\\mathbf{B}}_{t - k}$ is further processed by a lightweight object decoder. The decoder is used to generate 3D predictions $\\hat{\\mathbf{P}}_{t - k}$ upon the BEV feature and thus its implementation is flexible. Variants of BEV detection heads are an obvious choice and we will consider multiple instantiations in our experiments. After obtaining $\\hat{\\mathbf{P}}_{t - k}$ , we should align the coordinates between learning targets $\\mathbf{G}_{t - k}$ and the final predictions $\\hat{\\mathbf{P}}_{t - k}$ . For clarity, we first denote the ego coordinate as $e(t)$ at frame $t$ . In our implementation, we transform the extrinsic parameters of previous frames to the coordinate system of the current ego, thus BEV features and predictions at different timestamps share the same $e(t)$ . To simplify the learning targets, 3D coordinates of $\\mathbf{G}_{t - k}$ should be transformed into $e(t)$ . Considering the ego motion, we convert $\\mathbf{G}_{t - k}$ to $\\hat{\\mathbf{G}}_{t - k}$ by ego transformation matrix as:", + "bbox": [ + 75, + 597, + 473, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {G}} _ {t - k} = \\mathbf {T} _ {e (t - k)} ^ {e (t)} \\mathbf {G} _ {t - k}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 849, + 468, + 872 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{T}_{e(t - k)}^{e(t)}$ is the transformation matrix from the source", + "bbox": [ + 76, + 882, + 472, + 904 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f4e9e4ac013edac8e17ea90fa39be6f633033e7d11c401c255118b8721ae347e.jpg", + "image_caption": [ + "Figure 3: Illustration of the temporal multi-stage decoder. Object queries are enhanced by their counterparts derived from the last frame, forming a recurrent connection." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 880, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "coordinate system $e(t - k)$ to the target coordinate system $e(t)$ .", + "bbox": [ + 496, + 285, + 890, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion. The act of discarding a frame in HoP might share similarity with masked image modeling works, like MAE [5]. However, there are several points where HoP differs greatly from those works. First, HoP is not a self-supervised learning method as it requires 3D ground-truth boxes for training. HoP leverages short-term and long-term branches to extract diverse information from objects with different velocities to predict 3D targets in the discarded frame. In contrast, masked image modeling methods generally choose pixels or features as reconstruction targets, but predicting features fails to bring significant gains as HoP in Table 6c. Moreover, MAE is adopted for backbone pretraining. HoP is designed for 3D BEV detection to enhance temporal learning by predicting objects in the discarded frame. In fact, HoP is compatible with MAE since we can use a MAE pretrained ViT as the backbone.", + "bbox": [ + 496, + 315, + 892, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Historical Temporal Query Fusion", + "text_level": 1, + "bbox": [ + 498, + 566, + 802, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Apart from the HoP method that exploits temporal information in BEV feature level, we also explore query-level temporal modeling in query-based methods, e.g., BEV-Former, by fusing historical object queries into current queries. This fusion step provides current queries with an initialization of history perception, forming a refining process.", + "bbox": [ + 496, + 589, + 890, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The BEV feature sequence $\\{\\mathbf{B}_{t - N},\\dots ,\\mathbf{B}_t\\}$ is inferred in the time order. For the timestamp $t - k$ we perform detection with the pre-defined object queries $\\mathbf{O}$ as follows:", + "bbox": [ + 496, + 695, + 890, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {O}} _ {t - k} = \\mathcal {D} (\\mathbf {B} _ {t - k}, \\mathbf {O}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 750, + 890, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ is the BEV decoder and $\\bar{\\mathbf{O}}_{t - k}$ is its output.", + "bbox": [ + 498, + 779, + 846, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each object query learns to specialize in certain areas in the BEV plane, so the output queries contain rich semantic information from the regions. When detecting on the current frame, we refine these output historical queries instead of starting from the pre-defined ones to make the learning process easier. For example, if an object query detects a still object in the last frame, its historical counterpart could help", + "bbox": [ + 496, + 796, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3784", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/f80f1116b140f7b8e24da335820c2bbae4f4a07b4ab94cb406347b0a473fb6fb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneQuery-basedLiDAREpochNDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVDepth [14]R101-DCN‡✓90†0.5380.418-----
UVTR [13]R101-DCN‡✓240.4830.3790.7310.2670.3500.5100.200
BEVFormer [15]R101-DCN‡✓240.5170.4160.6730.2740.3720.3940.198
PETRv2 [19]R101-DCN‡✓240.5240.4210.6810.2670.3570.3770.186
PolarFormer [9]R101-DCN‡✓240.5280.4320.6480.2700.3480.4090.201
Sparse4D [17]R101-DCN‡✓240.5410.4360.6330.2790.3630.3170.177
HoP-BEVFormerR101-DCN‡✓240.5580.4540.5650.2650.3270.3370.194
", + "bbox": [ + 78, + 88, + 897, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Comparison of recent works on the nuScenes val set. $\\dagger$ indicates methods with CBGS which will elongate 1 epoch into 4.5 epochs. $\\ddagger$ notes the backbone is initialized from FCOS3D [28] backbone. LiDAR: trained with LiDAR supervision.", + "bbox": [ + 75, + 247, + 893, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "it quickly locate the object without searching on its area in the BEV plane again. On the other hand, historical information from a moving object also contributes to localizing and detecting its velocity.", + "bbox": [ + 75, + 304, + 467, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 3, to infer on the BEV feature $\\mathbf{B}_{t - k}$ we first collect historical object queries $\\mathbf{O}_{t - k}^{his}$ as follows:", + "bbox": [ + 76, + 364, + 467, + 396 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} _ {t - k} ^ {h i s} = \\operatorname {M L P} \\left(\\bar {\\mathbf {O}} _ {t - k - 1}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 402, + 468, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where MLP is a multi-layer perceptron with two linear layers and an expansion ratio of 4, which is used to fit the history query into the current timestamp. Note that we only use historical queries from the last timestamp because we hypothesize that queries from more distant history possess little information overlap with currently visible objects.", + "bbox": [ + 75, + 429, + 467, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we merge them with the pre-defined object queries $\\mathbf{O}$ as Equation 7,", + "bbox": [ + 76, + 521, + 467, + 550 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {O}} _ {t - k} = \\mathcal {D} \\left(\\mathbf {B} _ {t - k}, \\mathbf {O} + \\mathbf {O} _ {t - k} ^ {h i s}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 556, + 468, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the first shown-up objects in the current frame, the merge process takes both historical queries and predefined queries into account.", + "bbox": [ + 75, + 583, + 467, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 641, + 209, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and Metrics", + "text_level": 1, + "bbox": [ + 76, + 665, + 267, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We validate our method on the popular nuScenes[1] dataset. The nuScenes dataset contains 1000 driving scenes in total, which are split into 700, 150, and 150 respectively for training, validation, and testing. Each scene lasts for 20 seconds and annotations of 3D boxes are supplied every 0.5s in the keyframe. There are 6 cameras surrounding the vehicle, providing images covering the whole 360-degree FOV. As for evaluation, we adopt the standard metrics for nuScenes detection task. The final nuScenes detection score (NDS) is a weighted sum of mean Average Precision (mAP) and five True Positive metrics including Average Translation Error (ATE), Average Scale Error (ASE), Average Orientation Error (AOE), Average Velocity Error (AVE), and Average Attribute Error (AAE).", + "bbox": [ + 75, + 688, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 303, + 718, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We mainly implement our method on the optimized version of BEVFormer [15] (denoted as BEVFormer-opt), where we generate the current BEV feature with resolution $200 \\times 200$ by 6 BEV spatial encoder layers and fuse the current and 4 previous BEV features by a BEV temporal encoder. All models are trained with a mini-batch of 8, an initial learning rate of 0.0002, and a weight decay of 0.01 for 24 epochs without CBGS [36]. Unless otherwise specified, we use ResNet-50 [6] pre-trained on COCO [16] as the image backbone, and the image size is processed to $1408 \\times 512$ . Both image and BEV augmentation are adopted following BEVDet [8]. Besides, we also adopt the BEVDet4D-Depth [7] as baseline models and follow the original settings.", + "bbox": [ + 496, + 329, + 890, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 500, + 554, + 638, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To compare with previous state-of-the-art 3D object detection methods, we adopt ResNet101-DCN [6, 3] initialized from FCOS3D [28] and VoVNet-99 [10, 11] initialized from DD3D [23] checkpoint. We also enlarge the input resolution to $1600 \\times 640$ and use 1500 object queries. The historical temporal query fusion is employed in Table 1 and 2.", + "bbox": [ + 496, + 579, + 890, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We report the results of our HoP with the ResNet101-DCN backbone on nuScenes val in Table 1. Surprisingly, our method outperforms BEVDepth with CBGS strategy and depth supervision from LIDAR over $+2.0\\%$ NDS and $+3.6\\%$ mAP under comparable complexity levels. Besides, HoP achieves the best performance compared with other competitive query-based object detectors, e.g., PETRv2 [19] and PolarFormer [9]. We also compared HoP with other state-of-the-art algorithms on nuScenes test, and still obtain the best performance of $61.2\\%$ NDS and $52.8\\%$ mAP.", + "bbox": [ + 496, + 671, + 890, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To show the scalability of our model to stronger backbones, we apply HoP to BEVDet4D-Depth with ViT-L [4] as the backbone. The ViT-L we employ is initialized with the Co-DETR [38] pre-trained on the 2D detection bench", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3785", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d8232adb826dea4cae958e3e7b6cd051531b25c04c725b148492d786ae574eaf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneQuery-basedLiDAREpochNDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVDepth* [14]V2-99✓\\(90^{\\dagger}\\)0.6000.5030.4450.2450.3780.3200.126
DETR3D [30]V2-99✓240.4790.4120.6410.2550.3940.8450.133
UVTR [13]V2-99✓240.5510.4720.5770.2530.3910.5080.123
BEVFormer [15]V2-99✓240.5690.4810.5820.2560.3750.3780.126
PETRv2 [19]V2-99✓240.5820.4900.5610.2430.3610.3430.120
Sparse4D [17]V2-99✓240.5950.5110.5330.2630.3690.3170.124
HoP-BEVFormerV2-99✓240.6030.5170.5010.2450.3460.3620.105
HoP-BEVFormert†V2-99✓240.6120.5280.4910.2420.3320.3430.109
", + "bbox": [ + 78, + 88, + 908, + 257 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7563c01f94516d95ccfa1500b1b0400e2775e6328d58eefb9228859fa6e99651.jpg", + "table_caption": [ + "Table 2: Comparison of recent works on the nuScenes test set. * indicates test-time augmentation. † means we do not detach 1 historical frame. VoVNet-99 (V2-99) [10] was pre-trained on the depth estimation task with extra data [23]." + ], + "table_footnote": [], + "table_body": "
MethodBackboneTTANDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
SOLOFusion [24]ConvNeXt-B [21]0.6190.5400.4530.2570.3760.2760.148
BEVFormerV2 [32]InternImage-XL [29]0.6480.5800.4480.2620.3420.2380.128
BEVDet-Gamma [7]Swin-B [20]✓0.6640.5860.3750.2430.3770.1740.123
HoP-BEVDet4D-DepthViT-L [4]0.6850.6240.3670.2490.3540.1710.131
", + "bbox": [ + 96, + 306, + 870, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3: Comparisons on the nuScenes camera-only 3D detection leaderboard. TTA denotes test-time augmentation.", + "bbox": [ + 101, + 409, + 864, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mark Objects365 [26]. We use 8 frames into the future to provide additional information from the future. To reduce the training time, we first train the ViT-L using FCOS3D on the nuScenes trainval for 36 epochs. Then we only train HoP-BEVDet4D-Depth for 5 epochs with CBGS. Without test-time augmentation, HoP pushes the current best result to a new record of $68.5\\%$ NDS and $62.4\\%$ mAP on nuScenes test.", + "bbox": [ + 75, + 448, + 470, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 585, + 240, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unless stated otherwise, we conduct the ablation experiments on BEVFormer-opt with a ResNet-50 backbone. We follow the original settings when adopting BEVDet4D as the baseline.", + "bbox": [ + 75, + 611, + 468, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of each component. We perform a component-wise ablation to thoroughly analyze the effect of each component in Table 4. Surprisingly, incorporating the historical temporal prediction into the optimized BEVFormer brings significant performance gains $(+1.8\\%)$ NDS, $+1.4\\%$ mAP) over the baseline. The overall improvements hold when we apply the training strategy to BEVDet4D-Depth, achieving $+1.6\\%$ NDS and $+1.4\\%$ mAP improvements, respectively. Alternatively, the temporal information provided by historical queries also contributes to the performance improvements for BEVFormer, where a $+1.2\\%$ NDS gain is observed. Without detaching the gradients of historical BEV features in the last two layers of BEV encoder, the model can be further improved to $53.9\\%$ NDS. These results prove the effectiveness and generalization of our HoP.", + "bbox": [ + 75, + 674, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optimized BEVFormer baseline. Table 5 illustrates the critical designs of our optimized BEVFormer baseline. We can observe setting the dropout rate within the transformer as 0 significantly improves the performance and look forward twice [34] is an essential factor to improve mAP. Besides, we replace the heavy backbone with the ResNet50 and decrease the input resolution to $1408 \\times 512$ to improve the training efficiency of BEVFormer. At the same time, we also introduce step learning rate decay and data augmentation, which includes BEV augmentation, random flipping, cropping, and rotation, to achieve better performance without inference costs increase. Decoupling the BEV temporal encoder from the spatial encoder and using 4 previous frames, is crucial in the later promotion process. Combining these aforementioned modifications on BEVFormer results in our optimized baseline BEVFormer-opt. Finally, we achieve comparable performance with the original BEVFormer-R101-DCN while running much faster and requiring fewer parameters.", + "bbox": [ + 496, + 448, + 890, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Temporal decoder design. Table 6a illustrates the key designs of our temporal decoder for prediction targets prediction. The performance is always improved when using the structure of one individual decoder alone, reaching $52.9\\%$ NDS and $52.7\\%$ NDS with the short-term and the long-term decoder, respectively. We achieve the best performance with $53.1\\%$ NDS and $42.8\\%$ mAP when both short-term and long-term decoders are adopted, demonstrating the complementarity between these two branches.", + "bbox": [ + 496, + 741, + 890, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object decoder design. The object decoder detects the 3D", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3786", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d1b7dcc5e7867cb89fef11fdc7fb83278c791e06e9a3ba00aa16ec85f56300c0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodHoPHQ FusionNDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVFormer-opt0.5130.4140.6550.2740.4470.3680.194
BEVFormer-opt✓0.5250.4250.6600.2740.3950.3570.188
BEVFormer-opt✓0.5310.4280.6380.2690.3520.3690.185
BEVFormer-opt†✓0.5390.4350.6290.2680.3420.3600.184
BEVFormer-opt†✓✓0.5440.4390.6070.2650.3540.3400.193
BEVDet4D-Depth0.4930.3850.6320.2830.5810.2890.212
BEVDet4D-Depth✓0.5090.3990.6080.2720.5410.2810.205
", + "bbox": [ + 101, + 88, + 866, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f4ea1197608a1b9f8c8a2fd4c31a12563b9a932ca3237908f688b51aaaaffc0e.jpg", + "image_caption": [ + "Figure 4: Visualization result of BEVFormer-opt and HoP-BEVFormer on nuScenes val set. The predicted boxes are marked in yellow, while the ground truth boxes are marked in blue in the LIDAR_TOP figure." + ], + "image_footnote": [], + "bbox": [ + 81, + 262, + 883, + 412 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b918675bba87c0ab11a3b0fdec598d93e4d5fbc6ca94b2052907444ea7fa67ba.jpg", + "table_caption": [ + "Table 4: Component-wise ablations. † indicates we do not detach historical BEV features of last two layers in the BEV encoder. “HQ Fusion” refers to historical temporal query fusion." + ], + "table_footnote": [], + "table_body": "
MethodNDSmAP#params (M)Training time (sec/iter)
BEVFormer-R101-DCN0.5170.41678.62.7
+ Dropout 00.5290.42478.62.7
+ Look forward twice0.5310.43078.62.7
+ Smaller backbone R500.4870.38857.72.0
+ Step LR decay0.4930.39157.72.0
+ Data augmentation0.4990.40457.72.0
+ Decoupled BEV encoder0.5210.42055.52.2
+ Smaller resolution0.5130.41455.51.6
BEVFormer-opt-R500.5130.41455.51.6
", + "bbox": [ + 81, + 458, + 465, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5: Step-by-step optimizing the BEVFormer baseline.", + "bbox": [ + 94, + 614, + 450, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "targets with the reconstructed features, thus it can be flexibly designed, as studied in Table 6b. We find these detection heads can bring consistent gains. For example, ATSS [35] is extended to 3D object detection and brings $1.6\\%$ NDS and $1.2\\%$ mAP gains. We choose CenterPoint [33], which achieves the best performance as the default object decoder.", + "bbox": [ + 75, + 641, + 467, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prediction target. We also study the design that only predicts pseudo BEV features at timestamp $t - k$ supervised by $\\mathbf{B}_{t - k}$ generated by the BEV encoder in a self-distillation manner. As presented in Table 6c, the feature supervision signals only slightly improve the baseline while the 3D objects supervision boosts the performance by a large margin. In a summary, explicit supervision from 3D objects in the BEV space is crucial to HoP.", + "bbox": [ + 75, + 733, + 467, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Choice of prediction index $k$ . The selection of prediction index $k$ is a critical design of our method and is ablated in Table 6d. We first study whether the model can benefit from", + "bbox": [ + 75, + 854, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "predicting the 3D objects in the future frame, e.g., frame $t + 1$ ( $k = -1$ ). Interestingly, this future 3D object prediction also brings non-trivial gains but performs inferior to history prediction. The relatively weaker gains can be explained by the difference between a history prediction task and a future prediction task: it is much more challenging to predict 3D objects in the future frame at timestamp $t + 1$ with only history frames from timestamp $t - N$ to $t$ . Predicting objects in frame $t - k$ with both history frames from timestamp $t - N$ to $t - k - 1$ and future frames from timestamp $t - k + 1$ to $t$ is much easier. A challenging auxiliary task can disturb the learning progress and slightly decrease the performance gains. When predicting the objects of history frame $t - k$ , we do not detach the feature maps of frame $t - k + 1$ to allow the gradients back propagate through the both short-term and long-term temporal decoder. Therefore, we observe the method imposes a significant increase in GPU memory for $k > 1$ since the features at timestamp $t$ and $t - k + 1$ are not detached. Accordingly, we choose $k$ as 1 since it achieves the best trade-offs between accuracy and training costs.", + "bbox": [ + 496, + 460, + 890, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training time. We present the training speed (seconds per training iteration) under different settings in Table 6e. This training speed is evaluated with 8 NVIDIA A100 GPUs. Compared with the baseline, HoP introduces an additional temporal decoder and object decoder during training, leading to an inevitable increase in training time. Thanks to the lightweight design of our temporal decoder and object decoder, we only observe a $20\\%$ increase (from 1.6 seconds to 1.9 seconds) in training time for BEVFormer baseline.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3787", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e527cfff8c3741392057d9689c1114f39335a8c410c4358e80ca23bd89265acc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CaseNDSmAP
None0.5130.414
Short-term0.5280.426
Long-term0.5260.425
Both0.5310.428
", + "bbox": [ + 125, + 98, + 287, + 162 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/48c65fe156436b8e9a2b054e22f8c2f4cb85313d19900c6da38733e07f0f852a.jpg", + "table_caption": [ + "(a) Temporal decoder design. Both shortterm and long-term can improve accuracy." + ], + "table_footnote": [], + "table_body": "
kNDSmAPMemory
None0.5130.4141×
-10.5230.4201.6×
10.5310.4281.3×
20.5340.4302.3×
30.5360.4322.3×
", + "bbox": [ + 98, + 209, + 313, + 287 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4ac19c40d1b1024ff36d9397a0301b16da0f464d8b6f840122e8ee143e5d28d0.jpg", + "table_caption": [ + "(d) Prediction index. Predicting the objects in frame $t - 1$ achieves the best trade-offs." + ], + "table_footnote": [], + "table_body": "
CaseNDSmAP
None0.5130.414
Transformer0.5210.421
ATSS0.5290.426
CenterPoint0.5310.428
", + "bbox": [ + 401, + 98, + 570, + 162 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/32c30931f1f9a8a17d2c34fe3d8a6c5bcb92a5ced2db0db291cfe3eb3a454e2a.jpg", + "table_caption": [ + "(b) Object decoder design. The decoder can be flexibly implemented." + ], + "table_footnote": [], + "table_body": "
CaseNDSSpeed
BEVFormer-opt0.5131.6
HoP-BEVFormer-opt0.5311.9
HoP-BEVFormer-opt†0.5392.1
BEVDet4D-Depth0.4934.5
HoP-BEVDet4D-Depth0.5094.7
", + "bbox": [ + 369, + 209, + 602, + 287 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0cf4b5a66ec1e278618d26aa564e06a1443e63051c8c6c2eefb15d4a2e4a97fd.jpg", + "table_caption": [ + "(e) Training speed. HoP slightly increases the training time." + ], + "table_footnote": [], + "table_body": "
CaseNDSmAP
None0.5130.414
BEV feature0.5180.418
3D objects0.5310.428
", + "bbox": [ + 679, + 98, + 846, + 152 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4c8fda3f78cb10e9bc21ee91f72f66888343166467115f228433ab65130e79ef.jpg", + "table_caption": [ + "(c) Prediction target. The explicit supervi sions of 3D targets are critical." + ], + "table_footnote": [], + "table_body": "
Connection formNDSmAP
None0.5130.414
Recurrent0.5250.425
Fully-Connected0.5210.422
Dense0.5230.421
", + "bbox": [ + 668, + 209, + 859, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(f) Historical query collection form. All three connection forms bring improvements.", + "bbox": [ + 638, + 290, + 887, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6: Ablation experiments with ResNet-50 on nuScenes val. Default settings are marked in gray. † indicates we do not detach historical BEV features of last two layers in the BEV encoder.", + "bbox": [ + 75, + 325, + 892, + 356 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae303011ebde8430125259ca2f1227b96ccb8cef332aacf0109c74c40d01b22f.jpg", + "image_caption": [ + "Figure 5: Visualization results of object decoder with three variants of the temporal decoder." + ], + "image_footnote": [], + "bbox": [ + 161, + 363, + 354, + 561 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6514fbf7550fecb62afd9a9c70ade661fff1fce9e8d66c5e00628f6ca649dfa3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 363, + 581, + 560 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb087d68a1ee64af9cdd70bb09cc8c3d1c55fde84ad833f50d05614c62822541.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 363, + 810, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When the gradients of historical BEV features are not detached in the last two layers of BEV encoder, the training time is increased by $30\\%$ . As for BEVDet4D-Depth, the BEV feature resolution and dimension is smaller than BEV-Former. Therefore, the additional training costs brought by HoP are negligible (from 4.5 seconds to 4.7 seconds).", + "bbox": [ + 75, + 590, + 468, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Historical query collection form. We explore different methods to obtain historical object queries $\\mathbf{O}_{t - k}^{his}$ in Table 6f. The Recurrent method only utilizes historical queries from the last frame, which is described in Equation 6; the Fully-Connected method collects all the historical object queries $\\{\\bar{\\mathbf{O}}_{t - N},\\dots,\\bar{\\mathbf{O}}_{t - 1}\\}$ but only uses them for the current frame $t$ ; the Dense method extends the Fully-Connected method with using all historical queries on every timestamp. These three variants can bring consistent improvement, proving the insensitivity of the collection form; the recurrent form obtains the best result, where mAP increased by $1.2\\%$ and NDS increased by $1.3\\%$ . The results also prove our hypothesis that historical queries from the previous frame provide a more reliable initialization.", + "bbox": [ + 75, + 688, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 500, + 590, + 635, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4 shows the detection result of the baseline method BEVFormer-opt and our proposed HoP-BEVFormer. The predictions are marked in yellow, while the ground truth boxes are marked in blue in the LiDAR_TOP figure. We observe that HoP-BEVFormer successfully detects small objects or occluded objects thanks to the superior ability to aggregate temporal information.", + "bbox": [ + 496, + 621, + 890, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also visualize the predictions of the object decoder with three variants of temporal decoder $\\mathcal{T}$ : only long-term, only short-term, and both of them in Figure 5. The decoder with only a short-term branch can detect accurate attributes of objects, e.g., height and length, but fails to localize them precisely. Besides, results in the second row show that the short-term branch contributes to detecting more foreground objects because of its detailed spatial semantics. On the contrary, the decoder with only long-term modeling performs precise localization due to the long-term motion while struggling to detect the size of the object, e.g., height.", + "bbox": [ + 496, + 734, + 892, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3788", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When both long-term and short-term branches are adopted, we achieve the best result by combining their advantages.", + "bbox": [ + 76, + 90, + 468, + 122 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 135, + 194, + 150 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose a new paradigm, named Historical Object Prediction (HoP) for multi-view 3D detection to leverage temporal information more effectively. The HoP approach is straightforward: given the current timestamp $t$ , we generate a pseudo BEV feature of timestamp $t - k$ from its adjacent frames and utilize this feature to predict the object set at timestamp $t - k$ . First, we elaborately design short-term and long-term temporal decoders, which can generate the pseudo BEV feature for timestamp $t - k$ without the involvement of its corresponding camera images. Second, an additional object decoder is flexibly attached to predict the object targets using the generated pseudo BEV feature. As a plug-and-play approach, HoP can be easily incorporated into state-of-the-art BEV detection frameworks, including BEVFormer and BEVDet series. Extensive experiments are conducted to evaluate the effectiveness of the proposed HoP on the nuScenes dataset. Surprisingly, HoP achieves $68.5\\%$ NDS and $62.4\\%$ mAP on nuScenes test, outperforming all the 3D object detectors on the leaderboard by a significant margin.", + "bbox": [ + 76, + 161, + 472, + 464 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 76, + 477, + 236, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project is funded in part by National Key RD Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK.", + "bbox": [ + 76, + 502, + 468, + 594 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 607, + 173, + 622 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020.", + "[2] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Graph-detr3d: rethinking overlapping regions for multi-view 3d object detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5999-6008, 2022.", + "[3] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017.", + "[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans" + ], + "bbox": [ + 86, + 632, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[5] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, June 2022.", + "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[7] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022.", + "[8] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021.", + "[9] Yanqin Jiang, Li Zhang, Zhenwei Miao, Xiatian Zhu, Jin Gao, Weiming Hu, and Yu-Gang Jiang. Polarformer: Multicamera 3d object detection with polar transformers. arXiv preprint arXiv:2206.15398, 2022.", + "[10] Youngwan Lee, Joong-won Hwang, Sangrok Lee, Yuseok Bae, and Jongyoul Park. An energy andgpu-computation efficient backbone network for real-time object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019.", + "[11] Youngwan Lee and Jongyoul Park. Centermask: Realtime anchor-free instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13906-13915, 2020.", + "[12] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo. arXiv preprint arXiv:2209.10248, 2022.", + "[13] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. arXiv preprint arXiv:2206.00630, 2022.", + "[14] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022.", + "[15] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 1-18. Springer, 2022.", + "[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014.", + "[17] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detec" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3789", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022.", + "[18] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 531-548. Springer, 2022.", + "[19] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022.", + "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021.", + "[21] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022.", + "[22] Jiachen Lu, Zheyuan Zhou, Xiatian Zhu, Hang Xu, and Li Zhang. Learning ego 3d representation as ray tracing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI, pages 129-144. Springer, 2022.", + "[23] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021.", + "[24] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022.", + "[25] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020.", + "[26] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019.", + "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "[28] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021.", + "[29] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022." + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022.", + "[31] Zengran Wang, Chen Min, Zheng Ge, Yinhao Li, Zeming Li, Hongyu Yang, and Di Huang. Sts: Surround-view temporal stereo for multi-view 3d detection. arXiv preprint arXiv:2208.10145, 2022.", + "[32] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. arXiv preprint arXiv:2211.10439, 2022.", + "[33] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021.", + "[34] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on Learning Representations, 2022.", + "[35] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759-9768, 2020.", + "[36] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019.", + "[37] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.", + "[38] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. arXiv preprint arXiv:2211.12860, 2022." + ], + "bbox": [ + 501, + 92, + 890, + 641 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3790", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/553f0662-4b8d-4936-ab47-ad291ff03dd2_model.json b/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/553f0662-4b8d-4936-ab47-ad291ff03dd2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5a47b730cab92413706631167b8dbcd2733bf5b2 --- /dev/null +++ b/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/553f0662-4b8d-4936-ab47-ad291ff03dd2_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.203, + 0.756, + 0.277 + ], + "angle": 0, + "content": "Zhuofan Zong\\(^{1*}\\) Dongzhi Jiang\\(^{1,2*}\\) Guanglu Song\\(^{1}\\) Zeyue Xue\\(^{3}\\) Jingyong Su\\(^{4}\\) Hongsheng Li\\(^{2,5,6\\dagger}\\) Yu Liu\\(^{1\\dagger}\\) SenseTime Research 2CUHK MMLAB 3HKU 4HITSZ CPII 6Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.278, + 0.772, + 0.293 + ], + "angle": 0, + "content": "{zongzhuofan, jdzcarr7, liuyuisanai}@gmail.com hsli@ee.cuhk.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.474, + 0.844 + ], + "angle": 0, + "content": "In this paper, we propose a new paradigm, named Historical Object Prediction (HoP) for multi-view 3D detection to leverage temporal information more effectively. The HoP approach is straightforward: given the current timestamp \\( t \\), we generate a pseudo Bird's-Eye View (BEV) feature of timestamp \\( t-k \\) from its adjacent frames and utilize this feature to predict the object set at timestamp \\( t-k \\). Our approach is motivated by the observation that enforcing the detector to capture both the spatial location and temporal motion of objects occurring at historical timestamps can lead to more accurate BEV feature learning. First, we elaborately design short-term and long-term temporal decoders, which can generate the pseudo BEV feature for timestamp \\( t-k \\) without the involvement of its corresponding camera images. Second, an additional object decoder is flexibly attached to predict the object targets using the generated pseudo BEV feature. Note that we only perform HoP during training, thus the proposed method does not introduce extra overheads during inference. As a plug-and-play approach, HoP can be easily incorporated into state-of-the-art BEV detection frameworks, including BEVFormer and BEVDet series. Furthermore, the auxiliary HoP approach is complementary to prevalent temporal modeling methods, leading to significant performance gains. Extensive experiments are conducted to evaluate the effectiveness of the proposed HoP on the nuScenes dataset. We choose the representative methods, including BEVFormer and BEVDet4D-Depth to evaluate our method. Surprisingly, HoP achieves \\(68.5\\%\\) NDS and \\(62.4\\%\\) mAP with ViT-L on nuScenes test, outperforming all the 3D object detectors on the leaderboard. Codes are available at https://github.com/Sense-X/HoP." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.328, + 0.9, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.55, + 0.895, + 0.596 + ], + "angle": 0, + "content": "Figure 1: Incorporating HoP into 3D object detector with other temporal fusion methods [8, 7, 32, 12, 24]. Our HoP is plug-and-play and complementary to them." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.616, + 0.633, + 0.632 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Camera-only 3D detection from multi-view images is a challenging task for autonomous driving and has received increasing attention recently. Introducing the Bird's-Eye View (BEV) representation [25, 14, 8, 30, 18, 2] with the temporal features aggregation has become the superior design manner for camera-only 3D perception with multi-camera images. Based on this, a series of detectors [7, 15, 12, 24, 31, 19, 17] delve into the elaborated temporal information fusion methods and achieve significant breakthroughs. They consider the regions in 3D space and aggregate the image features corresponding to these hypothesis locations from multiple timestamps. These outstanding advances reveal the great potential of temporal information in camera-only 3D detection." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Beyond these temporal fusion mechanisms, in this paper, we propose a new paradigm for leveraging temporal information to enhance the temporal modeling ability of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.471, + 0.902 + ], + "angle": 0, + "content": "*Equal contribution. Work done during internship at SenseTime Research. † Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3781" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.197 + ], + "angle": 0, + "content": "the 3D object detector. It's a temporal-based auxiliary task only adopted during training stage, namely Historical Object Prediction (HoP). Our approach is motivated by the observation that enforcing the detector to capture both the spatial location and temporal motion of objects occurring at historical timestamps can lead to more accurate BEV feature learning." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.199, + 0.468, + 0.544 + ], + "angle": 0, + "content": "Specifically, given the current timestamp \\( t \\), we generate a pseudo BEV feature of timestamp \\( t - k \\) from its adjacent frames and utilize this feature to predict the object set at \\( t - k \\). First, we elaborately design short-term and long-term temporal decoders, which can generate the pseudo BEV feature for timestamp \\( t - k \\) without the involvement of its corresponding camera images. Thanks to the marginal temporal difference between two adjacent frames, the short-term temporal decoder process two adjacent frames of timestamp \\( t - k \\) to provide spatial semantics of objects. The long-term decoder captures long-term motion of the whole frame sequence and contributes to better object motion estimation, which is complementary to spatial localization of the short-term branch. Subsequently, an additional object decoder is flexibly attached to predict the object targets using the generated pseudo BEV feature. Note that we only perform HoP during training, thus the proposed method does not introduce extra overheads during inference. As a plug-and-play approach, HoP can be flexibly incorporated into the state-of-the-art BEV detection frameworks, including BEVFormer [15] and BEVDet series [8, 7, 14]. Furthermore, the auxiliary HoP approach is complementary to prevalent temporal modeling methods, leading to significant gains." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.546, + 0.468, + 0.696 + ], + "angle": 0, + "content": "Extensive experiments are conducted to evaluate the effectiveness of the proposed HoP on the nuScenes dataset [1]. We choose the representative methods, including BEVFormer and BEVDet4D-Depth [7, 14] to evaluate our method. To be specific, we obtain \\(55.8\\%\\) NDS with ResNet-101-DCN and \\(60.3\\%\\) NDS with VoVNet-99 when evaluating HoP on nuScenes val. Surprisingly, HoP achieves \\(68.5\\%\\) NDS and \\(62.4\\%\\) mAP with ViT-L [4] on nuScenes test, surpassing all the 3D object detectors on the leaderboard by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.697, + 0.468, + 0.725 + ], + "angle": 0, + "content": "In conclusion, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.738, + 0.468, + 0.829 + ], + "angle": 0, + "content": "- We propose a novel temporal enhanced training scheme, namely Historical Object Prediction (HoP), to encourage more accurate BEV feature learning. HoP can force the model to capture the spatial semantics and temporal motion of objects in the historical frame during training." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- We design a temporal decoder that consists of a short-term decoder and a long-term decoder to provide reliable spatial localization and accurate motion estimation of objects." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.092, + 0.89, + 0.166 + ], + "angle": 0, + "content": "- We equip the competitive 3D object detector baselines with our approach and yield significant improvements on the nuScenes dataset. Surprisingly, HoP with ViT-L achieves \\(68.5\\%\\) NDS and \\(62.4\\%\\) mAP on nuScenes test, establishing the new state-of-the-art performance." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.182, + 0.649, + 0.198 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.207, + 0.779, + 0.223 + ], + "angle": 0, + "content": "2.1. Multi-view 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.231, + 0.89, + 0.306 + ], + "angle": 0, + "content": "Modern multi-view methods for 3D object detection could be mainly categorized into two branches, LSS-based [25] and query-based methods. We will introduce how these two methods differ in feature aggregation and leave the temporal modeling part in the next section." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.308, + 0.89, + 0.427 + ], + "angle": 0, + "content": "BEVDet [8] is a representative method in LSS-based methods. Following the Lift-Splat-Shoot paradigm, the method first explicitly estimates depth for every image pixel, then lifts the 2D features to 3D voxels according to the depth, and finally splats 3D features to BEV features and conducts object detection on it. BEVDepth [14] further improves the view transformation module with explicit depth supervision." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.429, + 0.89, + 0.638 + ], + "angle": 0, + "content": "Query-based methods typically employ learnable queries to aggregate 2D image features by attention [27, 37] mechanism. DETR3D [30] utilizes object queries for predicting 3D positions and projects them back to 2D coordinates to obtain the corresponding features. Graph-DETR3D [2] and Sparse4D [17] enhance it respectively with a learnable 3D graph and sparse 4D sampling. PETR [18] encodes 3D position into image features and therefore directly query with global 2D features. BEVFormer [15] leverages grid-shaped BEV queries to interact with 2D features by deformable attention; on its basis, BEVFormerv2 [32] introduces perspective supervision. Besides, PolarFormer and Ego3RT [22, 9] advocates modeling BEV queries in polar coordinates to fit the real-world scenario." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.65, + 0.89, + 0.68 + ], + "angle": 0, + "content": "2.2. Temporal Modeling for Multi-view 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.69, + 0.89, + 0.9 + ], + "angle": 0, + "content": "The motion information in autonomous driving scenes has been increasingly explored to utilize the temporal cues for improving detection performance in recent 3D perception frameworks. BEVFormer [15] proposes the temporal self-attention mechanism to dynamically fuse the previous BEV features by deformable attention [37] in an RNN manner. BEVDet4D [7] introduces the temporal modeling to lift BEVDet [8] to spatial-temporal 4D space. The 3D position embedding (3D PE) in PETR [18] is extended by PETRv2 [19] with the temporal alignment. BEVStereo [12] and STS [31] both leverage temporal views for constructing multi-view stereo by an effective temporal stereo method. SOLOFusion [24] fully exploits the synergy of short-term and long-term temporal information that is highly comple" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.515, + 0.957 + ], + "angle": 0, + "content": "3782" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.092, + 0.849, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.892, + 0.371 + ], + "angle": 0, + "content": "Figure 2: Framework of our Historical Object Prediction (HoP). The auxiliary branches are discarded during evaluation. The symbol © denotes feature concatenation and fusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.473, + 0.496 + ], + "angle": 0, + "content": "mentary as well as achieving state-of-the-art performance. BEVFormerV2 [32] uses the bidirectional temporal encoder to utilize both history and future BEV features, improving the performance by a large margin. In contrast with aforementioned works, we provide a novel temporal enhanced training paradigm to improve the temporal modeling without additional inference costs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.51, + 0.17, + 0.525 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.274, + 0.55 + ], + "angle": 0, + "content": "3.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.47, + 0.635 + ], + "angle": 0, + "content": "As illustrated in Figure 2, HoP can be easily incorporated into many multi-view 3D object detectors, e.g., BEVFormer and BEVDet4D-Depth. We transform the extrinsic parameters of previous frames to the coordinate system of the current ego to perform temporal alignment following BEVDet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.643, + 0.331, + 0.659 + ], + "angle": 0, + "content": "3.2. Historical Object Prediction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.47, + 0.833 + ], + "angle": 0, + "content": "Pipeline. The architecture of our Historical Object Prediction (HoP) is composed of a temporal decoder \\(\\mathcal{T}\\) and an object decoder \\(\\hat{\\mathcal{D}}\\). Given the BEV feature sequence \\(\\{\\mathbf{B}_{t-N}, \\cdots, \\mathbf{B}_t\\}\\) that consists of \\(N\\) historical BEV features and the current BEV feature, we denote the corresponding 3D ground truths as \\(\\{\\mathbf{G}_{t-N}, \\cdots, \\mathbf{G}_t\\}\\). If the BEV feature at timestamp \\(t-k\\) is discarded, we use the remaining BEV feature sequence \\(\\{\\mathbf{B}_{t-N}, \\cdots, \\mathbf{B}_t\\} - \\{\\mathbf{B}_{t-k}\\}\\) (denoted as \\(\\mathbf{B}^{rem}\\)) to predict the 3D objects at timestamp \\(t-k\\). To be specific, we adopt \\(\\mathcal{T}\\) and \\(\\hat{\\mathcal{D}}\\) to transform the remaining BEV features to obtain the 3D predictions \\(\\hat{\\mathbf{P}}_{t-k}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.842, + 0.469, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {P}} _ {t - k} = \\hat {\\mathcal {D}} \\left(\\mathcal {T} \\left(\\left\\{\\mathbf {B} _ {t - N}, \\dots , \\mathbf {B} _ {t} \\right\\} - \\left\\{\\mathbf {B} _ {t - k} \\right\\}\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The temporal decoder \\(\\mathcal{T}\\) that models both short-term and long-term information is developed to reconstruct the BEV" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.892, + 0.496 + ], + "angle": 0, + "content": "feature at timestamp \\( t - k \\). The object decoder \\( \\hat{\\mathcal{D}} \\) is attached to generate the predictions \\( \\hat{\\mathbf{P}}_{t - k} \\) using the pseudo BEV feature. Considering the ego-motion and temporal alignment, we transform 3D coordinates of \\( \\mathbf{G}_{t - k} \\) into the coordinate system of current frame \\( t \\) (denoted as \\( \\hat{\\mathbf{G}}_{t - k} \\)). The final training objective is to minimize the differences between \\( \\hat{\\mathbf{P}}_{t - k} \\) and \\( \\hat{\\mathbf{G}}_{t - k} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.893, + 0.693 + ], + "angle": 0, + "content": "The input to the temporal decoder is the remaining set of BEV features consisting of (i) short-term BEV features \\(\\{\\mathbf{B}_{t - k - 1},\\mathbf{B}_{t - k + 1}\\}\\), and (ii) long-term BEV features \\(\\{\\mathbf{B}_{t - N},\\dots ,\\mathbf{B}_t\\} -\\{\\mathbf{B}_{t - k}\\}\\). We first add temporal positional embeddings to BEV features in this full set. Our temporal decoder reconstructs the feature \\(\\hat{\\mathbf{B}}_{t - k}\\) by utilizing the spatial semantics and temporal cues of the BEV feature sequence. More specifically, we present a temporal decoder that consists of two separate branches with their special expertise to capture short-term and long-term information. The learned strong spatial-temporal representation can help the network better estimate the object location in frame at timestamp \\(t - k\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.895, + 0.86 + ], + "angle": 0, + "content": "Short-term temporal decoder. Thanks to the high temporal correlation between adjacent frames, the short-term temporal decoder only operates on the adjacent BEV feature set \\(\\{\\mathbf{B}_{t - k - 1},\\mathbf{B}_{t - k + 1}\\}\\) (denoted as \\(\\mathbf{B}^{adj}\\)), building a detailed spatial representation in BEV space. We first define a grid-shaped learnable short-term BEV queries \\(\\mathbf{Q}_{t - k}^{short}\\in\\) \\(\\mathbb{R}^{H\\times W\\times C}\\), where \\(H\\) and \\(W\\) refer to the spatial shape of the BEV plane. The short-term BEV queries aggregate the spatial information and model the short-term motion from \\(\\mathbf{B}_{t - k - 1}\\) and \\(\\mathbf{B}_{t - k + 1}\\) through the short-term temporal attention, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.871, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {B}} _ {t - k} ^ {s h o r t} = \\sum_ {\\mathbf {V} \\in \\mathbf {B} ^ {a d j}} \\operatorname {D e f o r m A t t n} \\left(\\mathbf {Q} _ {t - k, p} ^ {s h o r t}, p, \\mathbf {V}\\right), \\quad (2)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3783" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "where \\( p \\) is the spatial index in the BEV plane, DeformAttn(\\( q, p, x \\)) refers to the deformable attention [37] with query \\( q \\), reference point \\( p \\) and input features \\( x \\). We further feed \\( \\hat{\\mathbf{B}}_{t-k}^{short} \\) into a feed-forward network [27] and obtain the output of this short-term branch. However, it is still challenging to precisely construct the temporal relations of the same objects between \\( \\mathbf{B}_{t-k} \\) and other BEV features with only two adjacent frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.474, + 0.472 + ], + "angle": 0, + "content": "Long-term temporal decoder. The long-term temporal processes the whole remaining BEV set to perceive the motion clues over long-term history, which increases the localization potential [24] and contributes to more accurate localization. Therefore, these two branches in the temporal decoder are complementary to each other. As for long-term motion, it is intuitive that we only focus on the spatial motion of the same objects in the bird's eye view, ignoring the height information of objects. For most 3D object detectors with 2D BEV features, e.g., BEVFormer and BEVDet, the height information has been flattened to the feature embeddings. Accordingly, we first employ a channel reduction operation to the input set \\(\\mathbf{B}^{rem}\\) to prune the height information and achieve better training efficiency. Given the learnable long-term BEV queries \\(\\mathbf{Q}_{t - k}^{long}\\in \\mathbb{R}^{H\\times W\\times C / r}\\) and reduction layer with parameter \\(\\mathbf{W}^r\\in \\mathbb{R}^{C\\times C / r}\\), we can capture the long-term dependencies as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.481, + 0.47, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {B}} _ {t - k} ^ {l o n g} = \\sum_ {\\mathbf {V} \\in \\mathbf {B} ^ {r e m}} \\mathrm {D e f o r m A t t n} (\\mathbf {Q} _ {t - k, p} ^ {l o n g}, p, \\mathbf {V W} ^ {r}), (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.597 + ], + "angle": 0, + "content": "where \\( r \\) denotes the reduction ratio and is set as 4 by default. After the feed-forward network, the long-term decoder output the BEV features. Finally, we concatenate the short-term and long-term BEV features and perform feature fusion by \\( 3 \\times 3 \\) convolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.474, + 0.841 + ], + "angle": 0, + "content": "Object decoder. The reconstructed BEV feature \\(\\hat{\\mathbf{B}}_{t - k}\\) is further processed by a lightweight object decoder. The decoder is used to generate 3D predictions \\(\\hat{\\mathbf{P}}_{t - k}\\) upon the BEV feature and thus its implementation is flexible. Variants of BEV detection heads are an obvious choice and we will consider multiple instantiations in our experiments. After obtaining \\(\\hat{\\mathbf{P}}_{t - k}\\), we should align the coordinates between learning targets \\(\\mathbf{G}_{t - k}\\) and the final predictions \\(\\hat{\\mathbf{P}}_{t - k}\\). For clarity, we first denote the ego coordinate as \\(e(t)\\) at frame \\(t\\). In our implementation, we transform the extrinsic parameters of previous frames to the coordinate system of the current ego, thus BEV features and predictions at different timestamps share the same \\(e(t)\\). To simplify the learning targets, 3D coordinates of \\(\\mathbf{G}_{t - k}\\) should be transformed into \\(e(t)\\). Considering the ego motion, we convert \\(\\mathbf{G}_{t - k}\\) to \\(\\hat{\\mathbf{G}}_{t - k}\\) by ego transformation matrix as:" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.85, + 0.47, + 0.873 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {G}} _ {t - k} = \\mathbf {T} _ {e (t - k)} ^ {e (t)} \\mathbf {G} _ {t - k}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.883, + 0.473, + 0.905 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}_{e(t - k)}^{e(t)}\\) is the transformation matrix from the source" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.088, + 0.882, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.214, + 0.895, + 0.261 + ], + "angle": 0, + "content": "Figure 3: Illustration of the temporal multi-stage decoder. Object queries are enhanced by their counterparts derived from the last frame, forming a recurrent connection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.286, + 0.892, + 0.315 + ], + "angle": 0, + "content": "coordinate system \\( e(t - k) \\) to the target coordinate system \\( e(t) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.316, + 0.893, + 0.558 + ], + "angle": 0, + "content": "Discussion. The act of discarding a frame in HoP might share similarity with masked image modeling works, like MAE [5]. However, there are several points where HoP differs greatly from those works. First, HoP is not a self-supervised learning method as it requires 3D ground-truth boxes for training. HoP leverages short-term and long-term branches to extract diverse information from objects with different velocities to predict 3D targets in the discarded frame. In contrast, masked image modeling methods generally choose pixels or features as reconstruction targets, but predicting features fails to bring significant gains as HoP in Table 6c. Moreover, MAE is adopted for backbone pretraining. HoP is designed for 3D BEV detection to enhance temporal learning by predicting objects in the discarded frame. In fact, HoP is compatible with MAE since we can use a MAE pretrained ViT as the backbone." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.567, + 0.803, + 0.584 + ], + "angle": 0, + "content": "3.3. Historical Temporal Query Fusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.695 + ], + "angle": 0, + "content": "Apart from the HoP method that exploits temporal information in BEV feature level, we also explore query-level temporal modeling in query-based methods, e.g., BEV-Former, by fusing historical object queries into current queries. This fusion step provides current queries with an initialization of history perception, forming a refining process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.696, + 0.892, + 0.743 + ], + "angle": 0, + "content": "The BEV feature sequence \\(\\{\\mathbf{B}_{t - N},\\dots ,\\mathbf{B}_t\\}\\) is inferred in the time order. For the timestamp \\(t - k\\) we perform detection with the pre-defined object queries \\(\\mathbf{O}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.751, + 0.892, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {O}} _ {t - k} = \\mathcal {D} (\\mathbf {B} _ {t - k}, \\mathbf {O}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.848, + 0.796 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) is the BEV decoder and \\(\\bar{\\mathbf{O}}_{t - k}\\) is its output." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Each object query learns to specialize in certain areas in the BEV plane, so the output queries contain rich semantic information from the regions. When detecting on the current frame, we refine these output historical queries instead of starting from the pre-defined ones to make the learning process easier. For example, if an object query detects a still object in the last frame, its historical counterpart could help" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3784" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.898, + 0.239 + ], + "angle": 0, + "content": "
MethodBackboneQuery-basedLiDAREpochNDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVDepth [14]R101-DCN‡✓90†0.5380.418-----
UVTR [13]R101-DCN‡✓240.4830.3790.7310.2670.3500.5100.200
BEVFormer [15]R101-DCN‡✓240.5170.4160.6730.2740.3720.3940.198
PETRv2 [19]R101-DCN‡✓240.5240.4210.6810.2670.3570.3770.186
PolarFormer [9]R101-DCN‡✓240.5280.4320.6480.2700.3480.4090.201
Sparse4D [17]R101-DCN‡✓240.5410.4360.6330.2790.3630.3170.177
HoP-BEVFormerR101-DCN‡✓240.5580.4540.5650.2650.3270.3370.194
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.248, + 0.895, + 0.279 + ], + "angle": 0, + "content": "Table 1: Comparison of recent works on the nuScenes val set. \\(\\dagger\\) indicates methods with CBGS which will elongate 1 epoch into 4.5 epochs. \\(\\ddagger\\) notes the backbone is initialized from FCOS3D [28] backbone. LiDAR: trained with LiDAR supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.468, + 0.365 + ], + "angle": 0, + "content": "it quickly locate the object without searching on its area in the BEV plane again. On the other hand, historical information from a moving object also contributes to localizing and detecting its velocity." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.365, + 0.468, + 0.397 + ], + "angle": 0, + "content": "As shown in Figure 3, to infer on the BEV feature \\(\\mathbf{B}_{t - k}\\) we first collect historical object queries \\(\\mathbf{O}_{t - k}^{his}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.404, + 0.469, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} _ {t - k} ^ {h i s} = \\operatorname {M L P} \\left(\\bar {\\mathbf {O}} _ {t - k - 1}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.468, + 0.521 + ], + "angle": 0, + "content": "where MLP is a multi-layer perceptron with two linear layers and an expansion ratio of 4, which is used to fit the history query into the current timestamp. Note that we only use historical queries from the last timestamp because we hypothesize that queries from more distant history possess little information overlap with currently visible objects." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.522, + 0.468, + 0.551 + ], + "angle": 0, + "content": "Then we merge them with the pre-defined object queries \\(\\mathbf{O}\\) as Equation 7," + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.558, + 0.469, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {O}} _ {t - k} = \\mathcal {D} \\left(\\mathbf {B} _ {t - k}, \\mathbf {O} + \\mathbf {O} _ {t - k} ^ {h i s}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.468, + 0.63 + ], + "angle": 0, + "content": "Considering the first shown-up objects in the current frame, the merge process takes both historical queries and predefined queries into account." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.642, + 0.21, + 0.658 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.268, + 0.681 + ], + "angle": 0, + "content": "4.1. Dataset and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We validate our method on the popular nuScenes[1] dataset. The nuScenes dataset contains 1000 driving scenes in total, which are split into 700, 150, and 150 respectively for training, validation, and testing. Each scene lasts for 20 seconds and annotations of 3D boxes are supplied every 0.5s in the keyframe. There are 6 cameras surrounding the vehicle, providing images covering the whole 360-degree FOV. As for evaluation, we adopt the standard metrics for nuScenes detection task. The final nuScenes detection score (NDS) is a weighted sum of mean Average Precision (mAP) and five True Positive metrics including Average Translation Error (ATE), Average Scale Error (ASE), Average Orientation Error (AOE), Average Velocity Error (AVE), and Average Attribute Error (AAE)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.304, + 0.719, + 0.32 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.33, + 0.892, + 0.542 + ], + "angle": 0, + "content": "We mainly implement our method on the optimized version of BEVFormer [15] (denoted as BEVFormer-opt), where we generate the current BEV feature with resolution \\(200 \\times 200\\) by 6 BEV spatial encoder layers and fuse the current and 4 previous BEV features by a BEV temporal encoder. All models are trained with a mini-batch of 8, an initial learning rate of 0.0002, and a weight decay of 0.01 for 24 epochs without CBGS [36]. Unless otherwise specified, we use ResNet-50 [6] pre-trained on COCO [16] as the image backbone, and the image size is processed to \\(1408 \\times 512\\). Both image and BEV augmentation are adopted following BEVDet [8]. Besides, we also adopt the BEVDet4D-Depth [7] as baseline models and follow the original settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.64, + 0.57 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.671 + ], + "angle": 0, + "content": "To compare with previous state-of-the-art 3D object detection methods, we adopt ResNet101-DCN [6, 3] initialized from FCOS3D [28] and VoVNet-99 [10, 11] initialized from DD3D [23] checkpoint. We also enlarge the input resolution to \\(1600 \\times 640\\) and use 1500 object queries. The historical temporal query fusion is employed in Table 1 and 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.837 + ], + "angle": 0, + "content": "We report the results of our HoP with the ResNet101-DCN backbone on nuScenes val in Table 1. Surprisingly, our method outperforms BEVDepth with CBGS strategy and depth supervision from LIDAR over \\(+2.0\\%\\) NDS and \\(+3.6\\%\\) mAP under comparable complexity levels. Besides, HoP achieves the best performance compared with other competitive query-based object detectors, e.g., PETRv2 [19] and PolarFormer [9]. We also compared HoP with other state-of-the-art algorithms on nuScenes test, and still obtain the best performance of \\(61.2\\%\\) NDS and \\(52.8\\%\\) mAP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To show the scalability of our model to stronger backbones, we apply HoP to BEVDet4D-Depth with ViT-L [4] as the backbone. The ViT-L we employ is initialized with the Co-DETR [38] pre-trained on the 2D detection bench" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3785" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.91, + 0.258 + ], + "angle": 0, + "content": "
MethodBackboneQuery-basedLiDAREpochNDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVDepth* [14]V2-99✓\\(90^{\\dagger}\\)0.6000.5030.4450.2450.3780.3200.126
DETR3D [30]V2-99✓240.4790.4120.6410.2550.3940.8450.133
UVTR [13]V2-99✓240.5510.4720.5770.2530.3910.5080.123
BEVFormer [15]V2-99✓240.5690.4810.5820.2560.3750.3780.126
PETRv2 [19]V2-99✓240.5820.4900.5610.2430.3610.3430.120
Sparse4D [17]V2-99✓240.5950.5110.5330.2630.3690.3170.124
HoP-BEVFormerV2-99✓240.6030.5170.5010.2450.3460.3620.105
HoP-BEVFormert†V2-99✓240.6120.5280.4910.2420.3320.3430.109
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.266, + 0.892, + 0.297 + ], + "angle": 0, + "content": "Table 2: Comparison of recent works on the nuScenes test set. * indicates test-time augmentation. † means we do not detach 1 historical frame. VoVNet-99 (V2-99) [10] was pre-trained on the depth estimation task with extra data [23]." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.307, + 0.872, + 0.404 + ], + "angle": 0, + "content": "
MethodBackboneTTANDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
SOLOFusion [24]ConvNeXt-B [21]0.6190.5400.4530.2570.3760.2760.148
BEVFormerV2 [32]InternImage-XL [29]0.6480.5800.4480.2620.3420.2380.128
BEVDet-Gamma [7]Swin-B [20]✓0.6640.5860.3750.2430.3770.1740.123
HoP-BEVDet4D-DepthViT-L [4]0.6850.6240.3670.2490.3540.1710.131
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.41, + 0.866, + 0.426 + ], + "angle": 0, + "content": "Table 3: Comparisons on the nuScenes camera-only 3D detection leaderboard. TTA denotes test-time augmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.471, + 0.569 + ], + "angle": 0, + "content": "mark Objects365 [26]. We use 8 frames into the future to provide additional information from the future. To reduce the training time, we first train the ViT-L using FCOS3D on the nuScenes trainval for 36 epochs. Then we only train HoP-BEVDet4D-Depth for 5 epochs with CBGS. Without test-time augmentation, HoP pushes the current best result to a new record of \\(68.5\\%\\) NDS and \\(62.4\\%\\) mAP on nuScenes test." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.586, + 0.241, + 0.601 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.671 + ], + "angle": 0, + "content": "Unless stated otherwise, we conduct the ablation experiments on BEVFormer-opt with a ResNet-50 backbone. We follow the original settings when adopting BEVDet4D as the baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of each component. We perform a component-wise ablation to thoroughly analyze the effect of each component in Table 4. Surprisingly, incorporating the historical temporal prediction into the optimized BEVFormer brings significant performance gains \\((+1.8\\%)\\) NDS, \\(+1.4\\%\\) mAP) over the baseline. The overall improvements hold when we apply the training strategy to BEVDet4D-Depth, achieving \\(+1.6\\%\\) NDS and \\(+1.4\\%\\) mAP improvements, respectively. Alternatively, the temporal information provided by historical queries also contributes to the performance improvements for BEVFormer, where a \\(+1.2\\%\\) NDS gain is observed. Without detaching the gradients of historical BEV features in the last two layers of BEV encoder, the model can be further improved to \\(53.9\\%\\) NDS. These results prove the effectiveness and generalization of our HoP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.736 + ], + "angle": 0, + "content": "Optimized BEVFormer baseline. Table 5 illustrates the critical designs of our optimized BEVFormer baseline. We can observe setting the dropout rate within the transformer as 0 significantly improves the performance and look forward twice [34] is an essential factor to improve mAP. Besides, we replace the heavy backbone with the ResNet50 and decrease the input resolution to \\(1408 \\times 512\\) to improve the training efficiency of BEVFormer. At the same time, we also introduce step learning rate decay and data augmentation, which includes BEV augmentation, random flipping, cropping, and rotation, to achieve better performance without inference costs increase. Decoupling the BEV temporal encoder from the spatial encoder and using 4 previous frames, is crucial in the later promotion process. Combining these aforementioned modifications on BEVFormer results in our optimized baseline BEVFormer-opt. Finally, we achieve comparable performance with the original BEVFormer-R101-DCN while running much faster and requiring fewer parameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.742, + 0.892, + 0.879 + ], + "angle": 0, + "content": "Temporal decoder design. Table 6a illustrates the key designs of our temporal decoder for prediction targets prediction. The performance is always improved when using the structure of one individual decoder alone, reaching \\(52.9\\%\\) NDS and \\(52.7\\%\\) NDS with the short-term and the long-term decoder, respectively. We achieve the best performance with \\(53.1\\%\\) NDS and \\(42.8\\%\\) mAP when both short-term and long-term decoders are adopted, demonstrating the complementarity between these two branches." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Object decoder design. The object decoder detects the 3D" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3786" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.867, + 0.228 + ], + "angle": 0, + "content": "
MethodHoPHQ FusionNDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVFormer-opt0.5130.4140.6550.2740.4470.3680.194
BEVFormer-opt✓0.5250.4250.6600.2740.3950.3570.188
BEVFormer-opt✓0.5310.4280.6380.2690.3520.3690.185
BEVFormer-opt†✓0.5390.4350.6290.2680.3420.3600.184
BEVFormer-opt†✓✓0.5440.4390.6070.2650.3540.3400.193
BEVDet4D-Depth0.4930.3850.6320.2830.5810.2890.212
BEVDet4D-Depth✓0.5090.3990.6080.2720.5410.2810.205
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.234, + 0.892, + 0.261 + ], + "angle": 0, + "content": "Table 4: Component-wise ablations. † indicates we do not detach historical BEV features of last two layers in the BEV encoder. “HQ Fusion” refers to historical temporal query fusion." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.263, + 0.885, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.426, + 0.893, + 0.456 + ], + "angle": 0, + "content": "Figure 4: Visualization result of BEVFormer-opt and HoP-BEVFormer on nuScenes val set. The predicted boxes are marked in yellow, while the ground truth boxes are marked in blue in the LIDAR_TOP figure." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.459, + 0.466, + 0.61 + ], + "angle": 0, + "content": "
MethodNDSmAP#params (M)Training time (sec/iter)
BEVFormer-R101-DCN0.5170.41678.62.7
+ Dropout 00.5290.42478.62.7
+ Look forward twice0.5310.43078.62.7
+ Smaller backbone R500.4870.38857.72.0
+ Step LR decay0.4930.39157.72.0
+ Data augmentation0.4990.40457.72.0
+ Decoupled BEV encoder0.5210.42055.52.2
+ Smaller resolution0.5130.41455.51.6
BEVFormer-opt-R500.5130.41455.51.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.615, + 0.451, + 0.63 + ], + "angle": 0, + "content": "Table 5: Step-by-step optimizing the BEVFormer baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.468, + 0.732 + ], + "angle": 0, + "content": "targets with the reconstructed features, thus it can be flexibly designed, as studied in Table 6b. We find these detection heads can bring consistent gains. For example, ATSS [35] is extended to 3D object detection and brings \\(1.6\\%\\) NDS and \\(1.2\\%\\) mAP gains. We choose CenterPoint [33], which achieves the best performance as the default object decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.468, + 0.853 + ], + "angle": 0, + "content": "Prediction target. We also study the design that only predicts pseudo BEV features at timestamp \\( t - k \\) supervised by \\( \\mathbf{B}_{t - k} \\) generated by the BEV encoder in a self-distillation manner. As presented in Table 6c, the feature supervision signals only slightly improve the baseline while the 3D objects supervision boosts the performance by a large margin. In a summary, explicit supervision from 3D objects in the BEV space is crucial to HoP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Choice of prediction index \\( k \\). The selection of prediction index \\( k \\) is a critical design of our method and is ablated in Table 6d. We first study whether the model can benefit from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.763 + ], + "angle": 0, + "content": "predicting the 3D objects in the future frame, e.g., frame \\( t + 1 \\) (\\( k = -1 \\)). Interestingly, this future 3D object prediction also brings non-trivial gains but performs inferior to history prediction. The relatively weaker gains can be explained by the difference between a history prediction task and a future prediction task: it is much more challenging to predict 3D objects in the future frame at timestamp \\( t + 1 \\) with only history frames from timestamp \\( t - N \\) to \\( t \\). Predicting objects in frame \\( t - k \\) with both history frames from timestamp \\( t - N \\) to \\( t - k - 1 \\) and future frames from timestamp \\( t - k + 1 \\) to \\( t \\) is much easier. A challenging auxiliary task can disturb the learning progress and slightly decrease the performance gains. When predicting the objects of history frame \\( t - k \\), we do not detach the feature maps of frame \\( t - k + 1 \\) to allow the gradients back propagate through the both short-term and long-term temporal decoder. Therefore, we observe the method imposes a significant increase in GPU memory for \\( k > 1 \\) since the features at timestamp \\( t \\) and \\( t - k + 1 \\) are not detached. Accordingly, we choose \\( k \\) as 1 since it achieves the best trade-offs between accuracy and training costs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training time. We present the training speed (seconds per training iteration) under different settings in Table 6e. This training speed is evaluated with 8 NVIDIA A100 GPUs. Compared with the baseline, HoP introduces an additional temporal decoder and object decoder during training, leading to an inevitable increase in training time. Thanks to the lightweight design of our temporal decoder and object decoder, we only observe a \\(20\\%\\) increase (from 1.6 seconds to 1.9 seconds) in training time for BEVFormer baseline." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3787" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.099, + 0.288, + 0.164 + ], + "angle": 0, + "content": "
CaseNDSmAP
None0.5130.414
Short-term0.5280.426
Long-term0.5260.425
Both0.5310.428
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.168, + 0.331, + 0.194 + ], + "angle": 0, + "content": "(a) Temporal decoder design. Both shortterm and long-term can improve accuracy." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.21, + 0.315, + 0.288 + ], + "angle": 0, + "content": "
kNDSmAPMemory
None0.5130.4141×
-10.5230.4201.6×
10.5310.4281.3×
20.5340.4302.3×
30.5360.4322.3×
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.291, + 0.331, + 0.316 + ], + "angle": 0, + "content": "(d) Prediction index. Predicting the objects in frame \\( t - 1 \\) achieves the best trade-offs." + }, + { + "type": "table", + "bbox": [ + 0.402, + 0.099, + 0.571, + 0.164 + ], + "angle": 0, + "content": "
CaseNDSmAP
None0.5130.414
Transformer0.5210.421
ATSS0.5290.426
CenterPoint0.5310.428
" + }, + { + "type": "table_caption", + "bbox": [ + 0.361, + 0.168, + 0.608, + 0.194 + ], + "angle": 0, + "content": "(b) Object decoder design. The decoder can be flexibly implemented." + }, + { + "type": "table", + "bbox": [ + 0.37, + 0.21, + 0.603, + 0.288 + ], + "angle": 0, + "content": "
CaseNDSSpeed
BEVFormer-opt0.5131.6
HoP-BEVFormer-opt0.5311.9
HoP-BEVFormer-opt†0.5392.1
BEVDet4D-Depth0.4934.5
HoP-BEVDet4D-Depth0.5094.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.361, + 0.291, + 0.608, + 0.317 + ], + "angle": 0, + "content": "(e) Training speed. HoP slightly increases the training time." + }, + { + "type": "table", + "bbox": [ + 0.681, + 0.099, + 0.848, + 0.153 + ], + "angle": 0, + "content": "
CaseNDSmAP
None0.5130.414
BEV feature0.5180.418
3D objects0.5310.428
" + }, + { + "type": "table_caption", + "bbox": [ + 0.64, + 0.168, + 0.887, + 0.194 + ], + "angle": 0, + "content": "(c) Prediction target. The explicit supervi sions of 3D targets are critical." + }, + { + "type": "table", + "bbox": [ + 0.669, + 0.21, + 0.86, + 0.275 + ], + "angle": 0, + "content": "
Connection formNDSmAP
None0.5130.414
Recurrent0.5250.425
Fully-Connected0.5210.422
Dense0.5230.421
" + }, + { + "type": "table_caption", + "bbox": [ + 0.64, + 0.291, + 0.888, + 0.317 + ], + "angle": 0, + "content": "(f) Historical query collection form. All three connection forms bring improvements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.893, + 0.357 + ], + "angle": 0, + "content": "Table 6: Ablation experiments with ResNet-50 on nuScenes val. Default settings are marked in gray. † indicates we do not detach historical BEV features of last two layers in the BEV encoder." + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.364, + 0.355, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.364, + 0.583, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.364, + 0.811, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.572, + 0.788, + 0.587 + ], + "angle": 0, + "content": "Figure 5: Visualization results of object decoder with three variants of the temporal decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.47, + 0.684 + ], + "angle": 0, + "content": "When the gradients of historical BEV features are not detached in the last two layers of BEV encoder, the training time is increased by \\(30\\%\\). As for BEVDet4D-Depth, the BEV feature resolution and dimension is smaller than BEV-Former. Therefore, the additional training costs brought by HoP are negligible (from 4.5 seconds to 4.7 seconds)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Historical query collection form. We explore different methods to obtain historical object queries \\(\\mathbf{O}_{t - k}^{his}\\) in Table 6f. The Recurrent method only utilizes historical queries from the last frame, which is described in Equation 6; the Fully-Connected method collects all the historical object queries \\(\\{\\bar{\\mathbf{O}}_{t - N},\\dots,\\bar{\\mathbf{O}}_{t - 1}\\}\\) but only uses them for the current frame \\(t\\); the Dense method extends the Fully-Connected method with using all historical queries on every timestamp. These three variants can bring consistent improvement, proving the insensitivity of the collection form; the recurrent form obtains the best result, where mAP increased by \\(1.2\\%\\) and NDS increased by \\(1.3\\%\\). The results also prove our hypothesis that historical queries from the previous frame provide a more reliable initialization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.592, + 0.637, + 0.607 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.892, + 0.729 + ], + "angle": 0, + "content": "Figure 4 shows the detection result of the baseline method BEVFormer-opt and our proposed HoP-BEVFormer. The predictions are marked in yellow, while the ground truth boxes are marked in blue in the LiDAR_TOP figure. We observe that HoP-BEVFormer successfully detects small objects or occluded objects thanks to the superior ability to aggregate temporal information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We also visualize the predictions of the object decoder with three variants of temporal decoder \\(\\mathcal{T}\\): only long-term, only short-term, and both of them in Figure 5. The decoder with only a short-term branch can detect accurate attributes of objects, e.g., height and length, but fails to localize them precisely. Besides, results in the second row show that the short-term branch contributes to detecting more foreground objects because of its detailed spatial semantics. On the contrary, the decoder with only long-term modeling performs precise localization due to the long-term motion while struggling to detect the size of the object, e.g., height." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3788" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "When both long-term and short-term branches are adopted, we achieve the best result by combining their advantages." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.136, + 0.196, + 0.151 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.162, + 0.473, + 0.465 + ], + "angle": 0, + "content": "In this paper, we propose a new paradigm, named Historical Object Prediction (HoP) for multi-view 3D detection to leverage temporal information more effectively. The HoP approach is straightforward: given the current timestamp \\( t \\), we generate a pseudo BEV feature of timestamp \\( t - k \\) from its adjacent frames and utilize this feature to predict the object set at timestamp \\( t - k \\). First, we elaborately design short-term and long-term temporal decoders, which can generate the pseudo BEV feature for timestamp \\( t - k \\) without the involvement of its corresponding camera images. Second, an additional object decoder is flexibly attached to predict the object targets using the generated pseudo BEV feature. As a plug-and-play approach, HoP can be easily incorporated into state-of-the-art BEV detection frameworks, including BEVFormer and BEVDet series. Extensive experiments are conducted to evaluate the effectiveness of the proposed HoP on the nuScenes dataset. Surprisingly, HoP achieves \\( 68.5\\% \\) NDS and \\( 62.4\\% \\) mAP on nuScenes test, outperforming all the 3D object detectors on the leaderboard by a significant margin." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.478, + 0.238, + 0.495 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.503, + 0.47, + 0.595 + ], + "angle": 0, + "content": "This project is funded in part by National Key RD Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.608, + 0.174, + 0.623 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.633, + 0.472, + 0.716 + ], + "angle": 0, + "content": "[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.718, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[2] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Graph-detr3d: rethinking overlapping regions for multi-view 3d object detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5999-6008, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.789, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[3] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.633, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[5] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.192, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.25, + 0.894, + 0.29 + ], + "angle": 0, + "content": "[7] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.293, + 0.892, + 0.334 + ], + "angle": 0, + "content": "[8] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.336, + 0.892, + 0.391 + ], + "angle": 0, + "content": "[9] Yanqin Jiang, Li Zhang, Zhenwei Miao, Xiatian Zhu, Jin Gao, Weiming Hu, and Yu-Gang Jiang. Polarformer: Multicamera 3d object detection with polar transformers. arXiv preprint arXiv:2206.15398, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.393, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[10] Youngwan Lee, Joong-won Hwang, Sangrok Lee, Yuseok Bae, and Jongyoul Park. An energy andgpu-computation efficient backbone network for real-time object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.894, + 0.518 + ], + "angle": 0, + "content": "[11] Youngwan Lee and Jongyoul Park. Centermask: Realtime anchor-free instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13906-13915, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[12] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo. arXiv preprint arXiv:2209.10248, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[13] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. arXiv preprint arXiv:2206.00630, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[14] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.691, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[15] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 1-18. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[17] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detec" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3789" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "tion with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[18] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 531-548. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[19] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.469, + 0.383 + ], + "angle": 0, + "content": "[21] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[22] Jiachen Lu, Zheyuan Zhou, Xiatian Zhu, Hang Xu, and Li Zhang. Learning ego 3d representation as ray tracing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI, pages 129-144. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[23] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[24] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[25] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[26] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.831 + ], + "angle": 0, + "content": "[28] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[29] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[30] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[31] Zengran Wang, Chen Min, Zheng Ge, Yinhao Li, Zeming Li, Hongyu Yang, and Di Huang. Sts: Surround-view temporal stereo for multi-view 3d detection. arXiv preprint arXiv:2208.10145, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[32] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. arXiv preprint arXiv:2211.10439, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[33] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[34] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.486 + ], + "angle": 0, + "content": "[35] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759-9768, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.488, + 0.892, + 0.542 + ], + "angle": 0, + "content": "[36] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.544, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[37] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[38] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. arXiv preprint arXiv:2211.12860, 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3790" + } + ] +] \ No newline at end of file diff --git a/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/553f0662-4b8d-4936-ab47-ad291ff03dd2_origin.pdf b/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/553f0662-4b8d-4936-ab47-ad291ff03dd2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..963168ce8ecc7af03d21cb7e4a51fcfeb2208743 --- /dev/null +++ b/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/553f0662-4b8d-4936-ab47-ad291ff03dd2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e63da0dcd247f9f3ce812f9929d622b41c30c0963c0086bab7178844458f575d +size 9223949 diff --git a/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/full.md b/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a74a1210fecbd58e6e7b8867a40e55806fb8bc63 --- /dev/null +++ b/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/full.md @@ -0,0 +1,298 @@ +# Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction + +Zhuofan Zong $^{1*}$ Dongzhi Jiang $^{1,2*}$ Guanglu Song $^{1}$ Zeyue Xue $^{3}$ Jingyong Su $^{4}$ Hongsheng Li $^{2,5,6\dagger}$ Yu Liu $^{1\dagger}$ SenseTime Research 2CUHK MMLAB 3HKU 4HITSZ CPII 6Shanghai AI Laboratory + +{zongzhuofan, jdzcarr7, liuyuisanai}@gmail.com hsli@ee.cuhk.edu.hk + +# Abstract + +In this paper, we propose a new paradigm, named Historical Object Prediction (HoP) for multi-view 3D detection to leverage temporal information more effectively. The HoP approach is straightforward: given the current timestamp $t$ , we generate a pseudo Bird's-Eye View (BEV) feature of timestamp $t-k$ from its adjacent frames and utilize this feature to predict the object set at timestamp $t-k$ . Our approach is motivated by the observation that enforcing the detector to capture both the spatial location and temporal motion of objects occurring at historical timestamps can lead to more accurate BEV feature learning. First, we elaborately design short-term and long-term temporal decoders, which can generate the pseudo BEV feature for timestamp $t-k$ without the involvement of its corresponding camera images. Second, an additional object decoder is flexibly attached to predict the object targets using the generated pseudo BEV feature. Note that we only perform HoP during training, thus the proposed method does not introduce extra overheads during inference. As a plug-and-play approach, HoP can be easily incorporated into state-of-the-art BEV detection frameworks, including BEVFormer and BEVDet series. Furthermore, the auxiliary HoP approach is complementary to prevalent temporal modeling methods, leading to significant performance gains. Extensive experiments are conducted to evaluate the effectiveness of the proposed HoP on the nuScenes dataset. We choose the representative methods, including BEVFormer and BEVDet4D-Depth to evaluate our method. Surprisingly, HoP achieves $68.5\%$ NDS and $62.4\%$ mAP with ViT-L on nuScenes test, outperforming all the 3D object detectors on the leaderboard. Codes are available at https://github.com/Sense-X/HoP. + +![](images/79d6afd5fdef55323ed71adc3b5bd4ce75cc18dc927d8a2bb56fee23b52bb689.jpg) +Figure 1: Incorporating HoP into 3D object detector with other temporal fusion methods [8, 7, 32, 12, 24]. Our HoP is plug-and-play and complementary to them. + +# 1. Introduction + +Camera-only 3D detection from multi-view images is a challenging task for autonomous driving and has received increasing attention recently. Introducing the Bird's-Eye View (BEV) representation [25, 14, 8, 30, 18, 2] with the temporal features aggregation has become the superior design manner for camera-only 3D perception with multi-camera images. Based on this, a series of detectors [7, 15, 12, 24, 31, 19, 17] delve into the elaborated temporal information fusion methods and achieve significant breakthroughs. They consider the regions in 3D space and aggregate the image features corresponding to these hypothesis locations from multiple timestamps. These outstanding advances reveal the great potential of temporal information in camera-only 3D detection. + +Beyond these temporal fusion mechanisms, in this paper, we propose a new paradigm for leveraging temporal information to enhance the temporal modeling ability of + +the 3D object detector. It's a temporal-based auxiliary task only adopted during training stage, namely Historical Object Prediction (HoP). Our approach is motivated by the observation that enforcing the detector to capture both the spatial location and temporal motion of objects occurring at historical timestamps can lead to more accurate BEV feature learning. + +Specifically, given the current timestamp $t$ , we generate a pseudo BEV feature of timestamp $t - k$ from its adjacent frames and utilize this feature to predict the object set at $t - k$ . First, we elaborately design short-term and long-term temporal decoders, which can generate the pseudo BEV feature for timestamp $t - k$ without the involvement of its corresponding camera images. Thanks to the marginal temporal difference between two adjacent frames, the short-term temporal decoder process two adjacent frames of timestamp $t - k$ to provide spatial semantics of objects. The long-term decoder captures long-term motion of the whole frame sequence and contributes to better object motion estimation, which is complementary to spatial localization of the short-term branch. Subsequently, an additional object decoder is flexibly attached to predict the object targets using the generated pseudo BEV feature. Note that we only perform HoP during training, thus the proposed method does not introduce extra overheads during inference. As a plug-and-play approach, HoP can be flexibly incorporated into the state-of-the-art BEV detection frameworks, including BEVFormer [15] and BEVDet series [8, 7, 14]. Furthermore, the auxiliary HoP approach is complementary to prevalent temporal modeling methods, leading to significant gains. + +Extensive experiments are conducted to evaluate the effectiveness of the proposed HoP on the nuScenes dataset [1]. We choose the representative methods, including BEVFormer and BEVDet4D-Depth [7, 14] to evaluate our method. To be specific, we obtain $55.8\%$ NDS with ResNet-101-DCN and $60.3\%$ NDS with VoVNet-99 when evaluating HoP on nuScenes val. Surprisingly, HoP achieves $68.5\%$ NDS and $62.4\%$ mAP with ViT-L [4] on nuScenes test, surpassing all the 3D object detectors on the leaderboard by a large margin. + +In conclusion, our contributions can be summarized as follows: + +- We propose a novel temporal enhanced training scheme, namely Historical Object Prediction (HoP), to encourage more accurate BEV feature learning. HoP can force the model to capture the spatial semantics and temporal motion of objects in the historical frame during training. + +- We design a temporal decoder that consists of a short-term decoder and a long-term decoder to provide reliable spatial localization and accurate motion estimation of objects. + +- We equip the competitive 3D object detector baselines with our approach and yield significant improvements on the nuScenes dataset. Surprisingly, HoP with ViT-L achieves $68.5\%$ NDS and $62.4\%$ mAP on nuScenes test, establishing the new state-of-the-art performance. + +# 2. Related Works + +# 2.1. Multi-view 3D Object Detection + +Modern multi-view methods for 3D object detection could be mainly categorized into two branches, LSS-based [25] and query-based methods. We will introduce how these two methods differ in feature aggregation and leave the temporal modeling part in the next section. + +BEVDet [8] is a representative method in LSS-based methods. Following the Lift-Splat-Shoot paradigm, the method first explicitly estimates depth for every image pixel, then lifts the 2D features to 3D voxels according to the depth, and finally splats 3D features to BEV features and conducts object detection on it. BEVDepth [14] further improves the view transformation module with explicit depth supervision. + +Query-based methods typically employ learnable queries to aggregate 2D image features by attention [27, 37] mechanism. DETR3D [30] utilizes object queries for predicting 3D positions and projects them back to 2D coordinates to obtain the corresponding features. Graph-DETR3D [2] and Sparse4D [17] enhance it respectively with a learnable 3D graph and sparse 4D sampling. PETR [18] encodes 3D position into image features and therefore directly query with global 2D features. BEVFormer [15] leverages grid-shaped BEV queries to interact with 2D features by deformable attention; on its basis, BEVFormerv2 [32] introduces perspective supervision. Besides, PolarFormer and Ego3RT [22, 9] advocates modeling BEV queries in polar coordinates to fit the real-world scenario. + +# 2.2. Temporal Modeling for Multi-view 3D Object Detection + +The motion information in autonomous driving scenes has been increasingly explored to utilize the temporal cues for improving detection performance in recent 3D perception frameworks. BEVFormer [15] proposes the temporal self-attention mechanism to dynamically fuse the previous BEV features by deformable attention [37] in an RNN manner. BEVDet4D [7] introduces the temporal modeling to lift BEVDet [8] to spatial-temporal 4D space. The 3D position embedding (3D PE) in PETR [18] is extended by PETRv2 [19] with the temporal alignment. BEVStereo [12] and STS [31] both leverage temporal views for constructing multi-view stereo by an effective temporal stereo method. SOLOFusion [24] fully exploits the synergy of short-term and long-term temporal information that is highly comple + +![](images/96b0d700a01eee3b95c2722bc1a9465ce4b1bb3630d51fb6e32d0da7d3b79612.jpg) +Figure 2: Framework of our Historical Object Prediction (HoP). The auxiliary branches are discarded during evaluation. The symbol © denotes feature concatenation and fusion. + +mentary as well as achieving state-of-the-art performance. BEVFormerV2 [32] uses the bidirectional temporal encoder to utilize both history and future BEV features, improving the performance by a large margin. In contrast with aforementioned works, we provide a novel temporal enhanced training paradigm to improve the temporal modeling without additional inference costs. + +# 3. Method + +# 3.1. Overall Architecture + +As illustrated in Figure 2, HoP can be easily incorporated into many multi-view 3D object detectors, e.g., BEVFormer and BEVDet4D-Depth. We transform the extrinsic parameters of previous frames to the coordinate system of the current ego to perform temporal alignment following BEVDet. + +# 3.2. Historical Object Prediction + +Pipeline. The architecture of our Historical Object Prediction (HoP) is composed of a temporal decoder $\mathcal{T}$ and an object decoder $\hat{\mathcal{D}}$ . Given the BEV feature sequence $\{\mathbf{B}_{t-N}, \cdots, \mathbf{B}_t\}$ that consists of $N$ historical BEV features and the current BEV feature, we denote the corresponding 3D ground truths as $\{\mathbf{G}_{t-N}, \cdots, \mathbf{G}_t\}$ . If the BEV feature at timestamp $t-k$ is discarded, we use the remaining BEV feature sequence $\{\mathbf{B}_{t-N}, \cdots, \mathbf{B}_t\} - \{\mathbf{B}_{t-k}\}$ (denoted as $\mathbf{B}^{rem}$ ) to predict the 3D objects at timestamp $t-k$ . To be specific, we adopt $\mathcal{T}$ and $\hat{\mathcal{D}}$ to transform the remaining BEV features to obtain the 3D predictions $\hat{\mathbf{P}}_{t-k}$ : + +$$ +\hat {\mathbf {P}} _ {t - k} = \hat {\mathcal {D}} \left(\mathcal {T} \left(\left\{\mathbf {B} _ {t - N}, \dots , \mathbf {B} _ {t} \right\} - \left\{\mathbf {B} _ {t - k} \right\}\right)\right). \tag {1} +$$ + +The temporal decoder $\mathcal{T}$ that models both short-term and long-term information is developed to reconstruct the BEV + +feature at timestamp $t - k$ . The object decoder $\hat{\mathcal{D}}$ is attached to generate the predictions $\hat{\mathbf{P}}_{t - k}$ using the pseudo BEV feature. Considering the ego-motion and temporal alignment, we transform 3D coordinates of $\mathbf{G}_{t - k}$ into the coordinate system of current frame $t$ (denoted as $\hat{\mathbf{G}}_{t - k}$ ). The final training objective is to minimize the differences between $\hat{\mathbf{P}}_{t - k}$ and $\hat{\mathbf{G}}_{t - k}$ . + +The input to the temporal decoder is the remaining set of BEV features consisting of (i) short-term BEV features $\{\mathbf{B}_{t - k - 1},\mathbf{B}_{t - k + 1}\}$ , and (ii) long-term BEV features $\{\mathbf{B}_{t - N},\dots ,\mathbf{B}_t\} -\{\mathbf{B}_{t - k}\}$ . We first add temporal positional embeddings to BEV features in this full set. Our temporal decoder reconstructs the feature $\hat{\mathbf{B}}_{t - k}$ by utilizing the spatial semantics and temporal cues of the BEV feature sequence. More specifically, we present a temporal decoder that consists of two separate branches with their special expertise to capture short-term and long-term information. The learned strong spatial-temporal representation can help the network better estimate the object location in frame at timestamp $t - k$ . + +Short-term temporal decoder. Thanks to the high temporal correlation between adjacent frames, the short-term temporal decoder only operates on the adjacent BEV feature set $\{\mathbf{B}_{t - k - 1},\mathbf{B}_{t - k + 1}\}$ (denoted as $\mathbf{B}^{adj}$ ), building a detailed spatial representation in BEV space. We first define a grid-shaped learnable short-term BEV queries $\mathbf{Q}_{t - k}^{short}\in$ $\mathbb{R}^{H\times W\times C}$ , where $H$ and $W$ refer to the spatial shape of the BEV plane. The short-term BEV queries aggregate the spatial information and model the short-term motion from $\mathbf{B}_{t - k - 1}$ and $\mathbf{B}_{t - k + 1}$ through the short-term temporal attention, which can be formulated as: + +$$ +\hat {\mathbf {B}} _ {t - k} ^ {s h o r t} = \sum_ {\mathbf {V} \in \mathbf {B} ^ {a d j}} \operatorname {D e f o r m A t t n} \left(\mathbf {Q} _ {t - k, p} ^ {s h o r t}, p, \mathbf {V}\right), \quad (2) +$$ + +where $p$ is the spatial index in the BEV plane, DeformAttn( $q, p, x$ ) refers to the deformable attention [37] with query $q$ , reference point $p$ and input features $x$ . We further feed $\hat{\mathbf{B}}_{t-k}^{short}$ into a feed-forward network [27] and obtain the output of this short-term branch. However, it is still challenging to precisely construct the temporal relations of the same objects between $\mathbf{B}_{t-k}$ and other BEV features with only two adjacent frames. + +Long-term temporal decoder. The long-term temporal processes the whole remaining BEV set to perceive the motion clues over long-term history, which increases the localization potential [24] and contributes to more accurate localization. Therefore, these two branches in the temporal decoder are complementary to each other. As for long-term motion, it is intuitive that we only focus on the spatial motion of the same objects in the bird's eye view, ignoring the height information of objects. For most 3D object detectors with 2D BEV features, e.g., BEVFormer and BEVDet, the height information has been flattened to the feature embeddings. Accordingly, we first employ a channel reduction operation to the input set $\mathbf{B}^{rem}$ to prune the height information and achieve better training efficiency. Given the learnable long-term BEV queries $\mathbf{Q}_{t - k}^{long}\in \mathbb{R}^{H\times W\times C / r}$ and reduction layer with parameter $\mathbf{W}^r\in \mathbb{R}^{C\times C / r}$ , we can capture the long-term dependencies as: + +$$ +\hat {\mathbf {B}} _ {t - k} ^ {l o n g} = \sum_ {\mathbf {V} \in \mathbf {B} ^ {r e m}} \mathrm {D e f o r m A t t n} (\mathbf {Q} _ {t - k, p} ^ {l o n g}, p, \mathbf {V W} ^ {r}), (3) +$$ + +where $r$ denotes the reduction ratio and is set as 4 by default. After the feed-forward network, the long-term decoder output the BEV features. Finally, we concatenate the short-term and long-term BEV features and perform feature fusion by $3 \times 3$ convolution. + +Object decoder. The reconstructed BEV feature $\hat{\mathbf{B}}_{t - k}$ is further processed by a lightweight object decoder. The decoder is used to generate 3D predictions $\hat{\mathbf{P}}_{t - k}$ upon the BEV feature and thus its implementation is flexible. Variants of BEV detection heads are an obvious choice and we will consider multiple instantiations in our experiments. After obtaining $\hat{\mathbf{P}}_{t - k}$ , we should align the coordinates between learning targets $\mathbf{G}_{t - k}$ and the final predictions $\hat{\mathbf{P}}_{t - k}$ . For clarity, we first denote the ego coordinate as $e(t)$ at frame $t$ . In our implementation, we transform the extrinsic parameters of previous frames to the coordinate system of the current ego, thus BEV features and predictions at different timestamps share the same $e(t)$ . To simplify the learning targets, 3D coordinates of $\mathbf{G}_{t - k}$ should be transformed into $e(t)$ . Considering the ego motion, we convert $\mathbf{G}_{t - k}$ to $\hat{\mathbf{G}}_{t - k}$ by ego transformation matrix as: + +$$ +\hat {\mathbf {G}} _ {t - k} = \mathbf {T} _ {e (t - k)} ^ {e (t)} \mathbf {G} _ {t - k}, \tag {4} +$$ + +where $\mathbf{T}_{e(t - k)}^{e(t)}$ is the transformation matrix from the source + +![](images/f4e9e4ac013edac8e17ea90fa39be6f633033e7d11c401c255118b8721ae347e.jpg) +Figure 3: Illustration of the temporal multi-stage decoder. Object queries are enhanced by their counterparts derived from the last frame, forming a recurrent connection. + +coordinate system $e(t - k)$ to the target coordinate system $e(t)$ . + +Discussion. The act of discarding a frame in HoP might share similarity with masked image modeling works, like MAE [5]. However, there are several points where HoP differs greatly from those works. First, HoP is not a self-supervised learning method as it requires 3D ground-truth boxes for training. HoP leverages short-term and long-term branches to extract diverse information from objects with different velocities to predict 3D targets in the discarded frame. In contrast, masked image modeling methods generally choose pixels or features as reconstruction targets, but predicting features fails to bring significant gains as HoP in Table 6c. Moreover, MAE is adopted for backbone pretraining. HoP is designed for 3D BEV detection to enhance temporal learning by predicting objects in the discarded frame. In fact, HoP is compatible with MAE since we can use a MAE pretrained ViT as the backbone. + +# 3.3. Historical Temporal Query Fusion + +Apart from the HoP method that exploits temporal information in BEV feature level, we also explore query-level temporal modeling in query-based methods, e.g., BEV-Former, by fusing historical object queries into current queries. This fusion step provides current queries with an initialization of history perception, forming a refining process. + +The BEV feature sequence $\{\mathbf{B}_{t - N},\dots ,\mathbf{B}_t\}$ is inferred in the time order. For the timestamp $t - k$ we perform detection with the pre-defined object queries $\mathbf{O}$ as follows: + +$$ +\bar {\mathbf {O}} _ {t - k} = \mathcal {D} (\mathbf {B} _ {t - k}, \mathbf {O}), \tag {5} +$$ + +where $\mathcal{D}$ is the BEV decoder and $\bar{\mathbf{O}}_{t - k}$ is its output. + +Each object query learns to specialize in certain areas in the BEV plane, so the output queries contain rich semantic information from the regions. When detecting on the current frame, we refine these output historical queries instead of starting from the pre-defined ones to make the learning process easier. For example, if an object query detects a still object in the last frame, its historical counterpart could help + +
MethodBackboneQuery-basedLiDAREpochNDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVDepth [14]R101-DCN‡✓90†0.5380.418-----
UVTR [13]R101-DCN‡✓240.4830.3790.7310.2670.3500.5100.200
BEVFormer [15]R101-DCN‡✓240.5170.4160.6730.2740.3720.3940.198
PETRv2 [19]R101-DCN‡✓240.5240.4210.6810.2670.3570.3770.186
PolarFormer [9]R101-DCN‡✓240.5280.4320.6480.2700.3480.4090.201
Sparse4D [17]R101-DCN‡✓240.5410.4360.6330.2790.3630.3170.177
HoP-BEVFormerR101-DCN‡✓240.5580.4540.5650.2650.3270.3370.194
+ +Table 1: Comparison of recent works on the nuScenes val set. $\dagger$ indicates methods with CBGS which will elongate 1 epoch into 4.5 epochs. $\ddagger$ notes the backbone is initialized from FCOS3D [28] backbone. LiDAR: trained with LiDAR supervision. + +it quickly locate the object without searching on its area in the BEV plane again. On the other hand, historical information from a moving object also contributes to localizing and detecting its velocity. + +As shown in Figure 3, to infer on the BEV feature $\mathbf{B}_{t - k}$ we first collect historical object queries $\mathbf{O}_{t - k}^{his}$ as follows: + +$$ +\mathbf {O} _ {t - k} ^ {h i s} = \operatorname {M L P} \left(\bar {\mathbf {O}} _ {t - k - 1}\right), \tag {6} +$$ + +where MLP is a multi-layer perceptron with two linear layers and an expansion ratio of 4, which is used to fit the history query into the current timestamp. Note that we only use historical queries from the last timestamp because we hypothesize that queries from more distant history possess little information overlap with currently visible objects. + +Then we merge them with the pre-defined object queries $\mathbf{O}$ as Equation 7, + +$$ +\bar {\mathbf {O}} _ {t - k} = \mathcal {D} \left(\mathbf {B} _ {t - k}, \mathbf {O} + \mathbf {O} _ {t - k} ^ {h i s}\right). \tag {7} +$$ + +Considering the first shown-up objects in the current frame, the merge process takes both historical queries and predefined queries into account. + +# 4. Experiments + +# 4.1. Dataset and Metrics + +We validate our method on the popular nuScenes[1] dataset. The nuScenes dataset contains 1000 driving scenes in total, which are split into 700, 150, and 150 respectively for training, validation, and testing. Each scene lasts for 20 seconds and annotations of 3D boxes are supplied every 0.5s in the keyframe. There are 6 cameras surrounding the vehicle, providing images covering the whole 360-degree FOV. As for evaluation, we adopt the standard metrics for nuScenes detection task. The final nuScenes detection score (NDS) is a weighted sum of mean Average Precision (mAP) and five True Positive metrics including Average Translation Error (ATE), Average Scale Error (ASE), Average Orientation Error (AOE), Average Velocity Error (AVE), and Average Attribute Error (AAE). + +# 4.2. Implementation Details + +We mainly implement our method on the optimized version of BEVFormer [15] (denoted as BEVFormer-opt), where we generate the current BEV feature with resolution $200 \times 200$ by 6 BEV spatial encoder layers and fuse the current and 4 previous BEV features by a BEV temporal encoder. All models are trained with a mini-batch of 8, an initial learning rate of 0.0002, and a weight decay of 0.01 for 24 epochs without CBGS [36]. Unless otherwise specified, we use ResNet-50 [6] pre-trained on COCO [16] as the image backbone, and the image size is processed to $1408 \times 512$ . Both image and BEV augmentation are adopted following BEVDet [8]. Besides, we also adopt the BEVDet4D-Depth [7] as baseline models and follow the original settings. + +# 4.3. Main Results + +To compare with previous state-of-the-art 3D object detection methods, we adopt ResNet101-DCN [6, 3] initialized from FCOS3D [28] and VoVNet-99 [10, 11] initialized from DD3D [23] checkpoint. We also enlarge the input resolution to $1600 \times 640$ and use 1500 object queries. The historical temporal query fusion is employed in Table 1 and 2. + +We report the results of our HoP with the ResNet101-DCN backbone on nuScenes val in Table 1. Surprisingly, our method outperforms BEVDepth with CBGS strategy and depth supervision from LIDAR over $+2.0\%$ NDS and $+3.6\%$ mAP under comparable complexity levels. Besides, HoP achieves the best performance compared with other competitive query-based object detectors, e.g., PETRv2 [19] and PolarFormer [9]. We also compared HoP with other state-of-the-art algorithms on nuScenes test, and still obtain the best performance of $61.2\%$ NDS and $52.8\%$ mAP. + +To show the scalability of our model to stronger backbones, we apply HoP to BEVDet4D-Depth with ViT-L [4] as the backbone. The ViT-L we employ is initialized with the Co-DETR [38] pre-trained on the 2D detection bench + +
MethodBackboneQuery-basedLiDAREpochNDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVDepth* [14]V2-99✓\(90^{\dagger}\)0.6000.5030.4450.2450.3780.3200.126
DETR3D [30]V2-99✓240.4790.4120.6410.2550.3940.8450.133
UVTR [13]V2-99✓240.5510.4720.5770.2530.3910.5080.123
BEVFormer [15]V2-99✓240.5690.4810.5820.2560.3750.3780.126
PETRv2 [19]V2-99✓240.5820.4900.5610.2430.3610.3430.120
Sparse4D [17]V2-99✓240.5950.5110.5330.2630.3690.3170.124
HoP-BEVFormerV2-99✓240.6030.5170.5010.2450.3460.3620.105
HoP-BEVFormert†V2-99✓240.6120.5280.4910.2420.3320.3430.109
+ +Table 2: Comparison of recent works on the nuScenes test set. * indicates test-time augmentation. † means we do not detach 1 historical frame. VoVNet-99 (V2-99) [10] was pre-trained on the depth estimation task with extra data [23]. + +
MethodBackboneTTANDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
SOLOFusion [24]ConvNeXt-B [21]0.6190.5400.4530.2570.3760.2760.148
BEVFormerV2 [32]InternImage-XL [29]0.6480.5800.4480.2620.3420.2380.128
BEVDet-Gamma [7]Swin-B [20]✓0.6640.5860.3750.2430.3770.1740.123
HoP-BEVDet4D-DepthViT-L [4]0.6850.6240.3670.2490.3540.1710.131
+ +Table 3: Comparisons on the nuScenes camera-only 3D detection leaderboard. TTA denotes test-time augmentation. + +mark Objects365 [26]. We use 8 frames into the future to provide additional information from the future. To reduce the training time, we first train the ViT-L using FCOS3D on the nuScenes trainval for 36 epochs. Then we only train HoP-BEVDet4D-Depth for 5 epochs with CBGS. Without test-time augmentation, HoP pushes the current best result to a new record of $68.5\%$ NDS and $62.4\%$ mAP on nuScenes test. + +# 4.4. Ablation Studies + +Unless stated otherwise, we conduct the ablation experiments on BEVFormer-opt with a ResNet-50 backbone. We follow the original settings when adopting BEVDet4D as the baseline. + +Effectiveness of each component. We perform a component-wise ablation to thoroughly analyze the effect of each component in Table 4. Surprisingly, incorporating the historical temporal prediction into the optimized BEVFormer brings significant performance gains $(+1.8\%)$ NDS, $+1.4\%$ mAP) over the baseline. The overall improvements hold when we apply the training strategy to BEVDet4D-Depth, achieving $+1.6\%$ NDS and $+1.4\%$ mAP improvements, respectively. Alternatively, the temporal information provided by historical queries also contributes to the performance improvements for BEVFormer, where a $+1.2\%$ NDS gain is observed. Without detaching the gradients of historical BEV features in the last two layers of BEV encoder, the model can be further improved to $53.9\%$ NDS. These results prove the effectiveness and generalization of our HoP. + +Optimized BEVFormer baseline. Table 5 illustrates the critical designs of our optimized BEVFormer baseline. We can observe setting the dropout rate within the transformer as 0 significantly improves the performance and look forward twice [34] is an essential factor to improve mAP. Besides, we replace the heavy backbone with the ResNet50 and decrease the input resolution to $1408 \times 512$ to improve the training efficiency of BEVFormer. At the same time, we also introduce step learning rate decay and data augmentation, which includes BEV augmentation, random flipping, cropping, and rotation, to achieve better performance without inference costs increase. Decoupling the BEV temporal encoder from the spatial encoder and using 4 previous frames, is crucial in the later promotion process. Combining these aforementioned modifications on BEVFormer results in our optimized baseline BEVFormer-opt. Finally, we achieve comparable performance with the original BEVFormer-R101-DCN while running much faster and requiring fewer parameters. + +Temporal decoder design. Table 6a illustrates the key designs of our temporal decoder for prediction targets prediction. The performance is always improved when using the structure of one individual decoder alone, reaching $52.9\%$ NDS and $52.7\%$ NDS with the short-term and the long-term decoder, respectively. We achieve the best performance with $53.1\%$ NDS and $42.8\%$ mAP when both short-term and long-term decoders are adopted, demonstrating the complementarity between these two branches. + +Object decoder design. The object decoder detects the 3D + +
MethodHoPHQ FusionNDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVFormer-opt0.5130.4140.6550.2740.4470.3680.194
BEVFormer-opt✓0.5250.4250.6600.2740.3950.3570.188
BEVFormer-opt✓0.5310.4280.6380.2690.3520.3690.185
BEVFormer-opt†✓0.5390.4350.6290.2680.3420.3600.184
BEVFormer-opt†✓✓0.5440.4390.6070.2650.3540.3400.193
BEVDet4D-Depth0.4930.3850.6320.2830.5810.2890.212
BEVDet4D-Depth✓0.5090.3990.6080.2720.5410.2810.205
+ +![](images/f4ea1197608a1b9f8c8a2fd4c31a12563b9a932ca3237908f688b51aaaaffc0e.jpg) +Figure 4: Visualization result of BEVFormer-opt and HoP-BEVFormer on nuScenes val set. The predicted boxes are marked in yellow, while the ground truth boxes are marked in blue in the LIDAR_TOP figure. + +Table 4: Component-wise ablations. † indicates we do not detach historical BEV features of last two layers in the BEV encoder. “HQ Fusion” refers to historical temporal query fusion. + +
MethodNDSmAP#params (M)Training time (sec/iter)
BEVFormer-R101-DCN0.5170.41678.62.7
+ Dropout 00.5290.42478.62.7
+ Look forward twice0.5310.43078.62.7
+ Smaller backbone R500.4870.38857.72.0
+ Step LR decay0.4930.39157.72.0
+ Data augmentation0.4990.40457.72.0
+ Decoupled BEV encoder0.5210.42055.52.2
+ Smaller resolution0.5130.41455.51.6
BEVFormer-opt-R500.5130.41455.51.6
+ +Table 5: Step-by-step optimizing the BEVFormer baseline. + +targets with the reconstructed features, thus it can be flexibly designed, as studied in Table 6b. We find these detection heads can bring consistent gains. For example, ATSS [35] is extended to 3D object detection and brings $1.6\%$ NDS and $1.2\%$ mAP gains. We choose CenterPoint [33], which achieves the best performance as the default object decoder. + +Prediction target. We also study the design that only predicts pseudo BEV features at timestamp $t - k$ supervised by $\mathbf{B}_{t - k}$ generated by the BEV encoder in a self-distillation manner. As presented in Table 6c, the feature supervision signals only slightly improve the baseline while the 3D objects supervision boosts the performance by a large margin. In a summary, explicit supervision from 3D objects in the BEV space is crucial to HoP. + +Choice of prediction index $k$ . The selection of prediction index $k$ is a critical design of our method and is ablated in Table 6d. We first study whether the model can benefit from + +predicting the 3D objects in the future frame, e.g., frame $t + 1$ ( $k = -1$ ). Interestingly, this future 3D object prediction also brings non-trivial gains but performs inferior to history prediction. The relatively weaker gains can be explained by the difference between a history prediction task and a future prediction task: it is much more challenging to predict 3D objects in the future frame at timestamp $t + 1$ with only history frames from timestamp $t - N$ to $t$ . Predicting objects in frame $t - k$ with both history frames from timestamp $t - N$ to $t - k - 1$ and future frames from timestamp $t - k + 1$ to $t$ is much easier. A challenging auxiliary task can disturb the learning progress and slightly decrease the performance gains. When predicting the objects of history frame $t - k$ , we do not detach the feature maps of frame $t - k + 1$ to allow the gradients back propagate through the both short-term and long-term temporal decoder. Therefore, we observe the method imposes a significant increase in GPU memory for $k > 1$ since the features at timestamp $t$ and $t - k + 1$ are not detached. Accordingly, we choose $k$ as 1 since it achieves the best trade-offs between accuracy and training costs. + +Training time. We present the training speed (seconds per training iteration) under different settings in Table 6e. This training speed is evaluated with 8 NVIDIA A100 GPUs. Compared with the baseline, HoP introduces an additional temporal decoder and object decoder during training, leading to an inevitable increase in training time. Thanks to the lightweight design of our temporal decoder and object decoder, we only observe a $20\%$ increase (from 1.6 seconds to 1.9 seconds) in training time for BEVFormer baseline. + +
CaseNDSmAP
None0.5130.414
Short-term0.5280.426
Long-term0.5260.425
Both0.5310.428
+ +(a) Temporal decoder design. Both shortterm and long-term can improve accuracy. + +
kNDSmAPMemory
None0.5130.4141×
-10.5230.4201.6×
10.5310.4281.3×
20.5340.4302.3×
30.5360.4322.3×
+ +(d) Prediction index. Predicting the objects in frame $t - 1$ achieves the best trade-offs. + +
CaseNDSmAP
None0.5130.414
Transformer0.5210.421
ATSS0.5290.426
CenterPoint0.5310.428
+ +(b) Object decoder design. The decoder can be flexibly implemented. + +
CaseNDSSpeed
BEVFormer-opt0.5131.6
HoP-BEVFormer-opt0.5311.9
HoP-BEVFormer-opt†0.5392.1
BEVDet4D-Depth0.4934.5
HoP-BEVDet4D-Depth0.5094.7
+ +(e) Training speed. HoP slightly increases the training time. + +
CaseNDSmAP
None0.5130.414
BEV feature0.5180.418
3D objects0.5310.428
+ +(c) Prediction target. The explicit supervi sions of 3D targets are critical. + +
Connection formNDSmAP
None0.5130.414
Recurrent0.5250.425
Fully-Connected0.5210.422
Dense0.5230.421
+ +(f) Historical query collection form. All three connection forms bring improvements. + +Table 6: Ablation experiments with ResNet-50 on nuScenes val. Default settings are marked in gray. † indicates we do not detach historical BEV features of last two layers in the BEV encoder. + +![](images/ae303011ebde8430125259ca2f1227b96ccb8cef332aacf0109c74c40d01b22f.jpg) +Figure 5: Visualization results of object decoder with three variants of the temporal decoder. + +![](images/6514fbf7550fecb62afd9a9c70ade661fff1fce9e8d66c5e00628f6ca649dfa3.jpg) + +![](images/eb087d68a1ee64af9cdd70bb09cc8c3d1c55fde84ad833f50d05614c62822541.jpg) + +When the gradients of historical BEV features are not detached in the last two layers of BEV encoder, the training time is increased by $30\%$ . As for BEVDet4D-Depth, the BEV feature resolution and dimension is smaller than BEV-Former. Therefore, the additional training costs brought by HoP are negligible (from 4.5 seconds to 4.7 seconds). + +Historical query collection form. We explore different methods to obtain historical object queries $\mathbf{O}_{t - k}^{his}$ in Table 6f. The Recurrent method only utilizes historical queries from the last frame, which is described in Equation 6; the Fully-Connected method collects all the historical object queries $\{\bar{\mathbf{O}}_{t - N},\dots,\bar{\mathbf{O}}_{t - 1}\}$ but only uses them for the current frame $t$ ; the Dense method extends the Fully-Connected method with using all historical queries on every timestamp. These three variants can bring consistent improvement, proving the insensitivity of the collection form; the recurrent form obtains the best result, where mAP increased by $1.2\%$ and NDS increased by $1.3\%$ . The results also prove our hypothesis that historical queries from the previous frame provide a more reliable initialization. + +# 4.5. Visualization + +Figure 4 shows the detection result of the baseline method BEVFormer-opt and our proposed HoP-BEVFormer. The predictions are marked in yellow, while the ground truth boxes are marked in blue in the LiDAR_TOP figure. We observe that HoP-BEVFormer successfully detects small objects or occluded objects thanks to the superior ability to aggregate temporal information. + +We also visualize the predictions of the object decoder with three variants of temporal decoder $\mathcal{T}$ : only long-term, only short-term, and both of them in Figure 5. The decoder with only a short-term branch can detect accurate attributes of objects, e.g., height and length, but fails to localize them precisely. Besides, results in the second row show that the short-term branch contributes to detecting more foreground objects because of its detailed spatial semantics. On the contrary, the decoder with only long-term modeling performs precise localization due to the long-term motion while struggling to detect the size of the object, e.g., height. + +When both long-term and short-term branches are adopted, we achieve the best result by combining their advantages. + +# 5. Conclusion + +In this paper, we propose a new paradigm, named Historical Object Prediction (HoP) for multi-view 3D detection to leverage temporal information more effectively. The HoP approach is straightforward: given the current timestamp $t$ , we generate a pseudo BEV feature of timestamp $t - k$ from its adjacent frames and utilize this feature to predict the object set at timestamp $t - k$ . First, we elaborately design short-term and long-term temporal decoders, which can generate the pseudo BEV feature for timestamp $t - k$ without the involvement of its corresponding camera images. Second, an additional object decoder is flexibly attached to predict the object targets using the generated pseudo BEV feature. As a plug-and-play approach, HoP can be easily incorporated into state-of-the-art BEV detection frameworks, including BEVFormer and BEVDet series. Extensive experiments are conducted to evaluate the effectiveness of the proposed HoP on the nuScenes dataset. Surprisingly, HoP achieves $68.5\%$ NDS and $62.4\%$ mAP on nuScenes test, outperforming all the 3D object detectors on the leaderboard by a significant margin. + +# Acknowledgement + +This project is funded in part by National Key RD Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. + +# References + +[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. +[2] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Graph-detr3d: rethinking overlapping regions for multi-view 3d object detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5999-6008, 2022. +[3] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. +[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans + +formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[5] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, June 2022. +[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[7] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. +[8] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. +[9] Yanqin Jiang, Li Zhang, Zhenwei Miao, Xiatian Zhu, Jin Gao, Weiming Hu, and Yu-Gang Jiang. Polarformer: Multicamera 3d object detection with polar transformers. arXiv preprint arXiv:2206.15398, 2022. +[10] Youngwan Lee, Joong-won Hwang, Sangrok Lee, Yuseok Bae, and Jongyoul Park. An energy andgpu-computation efficient backbone network for real-time object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019. +[11] Youngwan Lee and Jongyoul Park. Centermask: Realtime anchor-free instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13906-13915, 2020. +[12] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo. arXiv preprint arXiv:2209.10248, 2022. +[13] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. arXiv preprint arXiv:2206.00630, 2022. +[14] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. +[15] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 1-18. Springer, 2022. +[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. +[17] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detec + +tion with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. +[18] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 531-548. Springer, 2022. +[19] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022. +[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. +[21] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022. +[22] Jiachen Lu, Zheyuan Zhou, Xiatian Zhu, Hang Xu, and Li Zhang. Learning ego 3d representation as ray tracing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI, pages 129-144. Springer, 2022. +[23] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021. +[24] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022. +[25] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. +[26] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. +[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[28] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021. +[29] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022. + +[30] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022. +[31] Zengran Wang, Chen Min, Zheng Ge, Yinhao Li, Zeming Li, Hongyu Yang, and Di Huang. Sts: Surround-view temporal stereo for multi-view 3d detection. arXiv preprint arXiv:2208.10145, 2022. +[32] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. arXiv preprint arXiv:2211.10439, 2022. +[33] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. +[34] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on Learning Representations, 2022. +[35] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759-9768, 2020. +[36] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019. +[37] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. +[38] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. arXiv preprint arXiv:2211.12860, 2022. \ No newline at end of file diff --git a/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/images.zip b/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..095397d3ba465d8f71fa623e4f819dd791550bef --- /dev/null +++ b/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6cd024b2add3e8a1d699ddfbbeadf09d6deca5b4a482415073ca26eb624dedb3 +size 675873 diff --git a/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/layout.json b/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6378105a3e5bff141f771800dae8322126eaf994 --- /dev/null +++ b/2023/Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction/layout.json @@ -0,0 +1,9224 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 536, + 140 + ], + "type": "text", + "content": "Temporal Enhanced Training of Multi-view 3D Object Detector via Historical Object Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "spans": [ + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "text", + "content": "Zhuofan Zong" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "text", + "content": " Dongzhi Jiang" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "text", + "content": " Guanglu Song" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "text", + "content": " Zeyue Xue" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "text", + "content": " Jingyong Su" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "text", + "content": " Hongsheng Li" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "inline_equation", + "content": "^{2,5,6\\dagger}" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "text", + "content": " Yu Liu" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 130, + 160, + 462, + 219 + ], + "type": "text", + "content": " SenseTime Research 2CUHK MMLAB 3HKU 4HITSZ CPII 6Shanghai AI Laboratory" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 220, + 472, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 220, + 472, + 232 + ], + "spans": [ + { + "bbox": [ + 121, + 220, + 472, + 232 + ], + "type": "text", + "content": "{zongzhuofan, jdzcarr7, liuyuisanai}@gmail.com hsli@ee.cuhk.edu.hk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "text", + "content": "In this paper, we propose a new paradigm, named Historical Object Prediction (HoP) for multi-view 3D detection to leverage temporal information more effectively. The HoP approach is straightforward: given the current timestamp " + }, + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "text", + "content": ", we generate a pseudo Bird's-Eye View (BEV) feature of timestamp " + }, + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "inline_equation", + "content": "t-k" + }, + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "text", + "content": " from its adjacent frames and utilize this feature to predict the object set at timestamp " + }, + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "inline_equation", + "content": "t-k" + }, + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "text", + "content": ". Our approach is motivated by the observation that enforcing the detector to capture both the spatial location and temporal motion of objects occurring at historical timestamps can lead to more accurate BEV feature learning. First, we elaborately design short-term and long-term temporal decoders, which can generate the pseudo BEV feature for timestamp " + }, + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "inline_equation", + "content": "t-k" + }, + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "text", + "content": " without the involvement of its corresponding camera images. Second, an additional object decoder is flexibly attached to predict the object targets using the generated pseudo BEV feature. Note that we only perform HoP during training, thus the proposed method does not introduce extra overheads during inference. As a plug-and-play approach, HoP can be easily incorporated into state-of-the-art BEV detection frameworks, including BEVFormer and BEVDet series. Furthermore, the auxiliary HoP approach is complementary to prevalent temporal modeling methods, leading to significant performance gains. Extensive experiments are conducted to evaluate the effectiveness of the proposed HoP on the nuScenes dataset. We choose the representative methods, including BEVFormer and BEVDet4D-Depth to evaluate our method. Surprisingly, HoP achieves " + }, + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "inline_equation", + "content": "68.5\\%" + }, + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "text", + "content": " NDS and " + }, + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 46, + 285, + 290, + 668 + ], + "type": "text", + "content": " mAP with ViT-L on nuScenes test, outperforming all the 3D object detectors on the leaderboard. Codes are available at https://github.com/Sense-X/HoP." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 259, + 550, + 424 + ], + "blocks": [ + { + "bbox": [ + 309, + 259, + 550, + 424 + ], + "lines": [ + { + "bbox": [ + 309, + 259, + 550, + 424 + ], + "spans": [ + { + "bbox": [ + 309, + 259, + 550, + 424 + ], + "type": "image", + "image_path": "79d6afd5fdef55323ed71adc3b5bd4ce75cc18dc927d8a2bb56fee23b52bb689.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 435, + 547, + 472 + ], + "lines": [ + { + "bbox": [ + 305, + 435, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 435, + 547, + 472 + ], + "type": "text", + "content": "Figure 1: Incorporating HoP into 3D object detector with other temporal fusion methods [8, 7, 32, 12, 24]. Our HoP is plug-and-play and complementary to them." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 487, + 387, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 387, + 500 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 387, + 500 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 509, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 676 + ], + "type": "text", + "content": "Camera-only 3D detection from multi-view images is a challenging task for autonomous driving and has received increasing attention recently. Introducing the Bird's-Eye View (BEV) representation [25, 14, 8, 30, 18, 2] with the temporal features aggregation has become the superior design manner for camera-only 3D perception with multi-camera images. Based on this, a series of detectors [7, 15, 12, 24, 31, 19, 17] delve into the elaborated temporal information fusion methods and achieve significant breakthroughs. They consider the regions in 3D space and aggregate the image features corresponding to these hypothesis locations from multiple timestamps. These outstanding advances reveal the great potential of temporal information in camera-only 3D detection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Beyond these temporal fusion mechanisms, in this paper, we propose a new paradigm for leveraging temporal information to enhance the temporal modeling ability of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 288, + 714 + ], + "type": "text", + "content": "*Equal contribution. Work done during internship at SenseTime Research. † Corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3781" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 156 + ], + "type": "text", + "content": "the 3D object detector. It's a temporal-based auxiliary task only adopted during training stage, namely Historical Object Prediction (HoP). Our approach is motivated by the observation that enforcing the detector to capture both the spatial location and temporal motion of objects occurring at historical timestamps can lead to more accurate BEV feature learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 157, + 286, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 157, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 157, + 286, + 430 + ], + "type": "text", + "content": "Specifically, given the current timestamp " + }, + { + "bbox": [ + 49, + 157, + 286, + 430 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 49, + 157, + 286, + 430 + ], + "type": "text", + "content": ", we generate a pseudo BEV feature of timestamp " + }, + { + "bbox": [ + 49, + 157, + 286, + 430 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 49, + 157, + 286, + 430 + ], + "type": "text", + "content": " from its adjacent frames and utilize this feature to predict the object set at " + }, + { + "bbox": [ + 49, + 157, + 286, + 430 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 49, + 157, + 286, + 430 + ], + "type": "text", + "content": ". First, we elaborately design short-term and long-term temporal decoders, which can generate the pseudo BEV feature for timestamp " + }, + { + "bbox": [ + 49, + 157, + 286, + 430 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 49, + 157, + 286, + 430 + ], + "type": "text", + "content": " without the involvement of its corresponding camera images. Thanks to the marginal temporal difference between two adjacent frames, the short-term temporal decoder process two adjacent frames of timestamp " + }, + { + "bbox": [ + 49, + 157, + 286, + 430 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 49, + 157, + 286, + 430 + ], + "type": "text", + "content": " to provide spatial semantics of objects. The long-term decoder captures long-term motion of the whole frame sequence and contributes to better object motion estimation, which is complementary to spatial localization of the short-term branch. Subsequently, an additional object decoder is flexibly attached to predict the object targets using the generated pseudo BEV feature. Note that we only perform HoP during training, thus the proposed method does not introduce extra overheads during inference. As a plug-and-play approach, HoP can be flexibly incorporated into the state-of-the-art BEV detection frameworks, including BEVFormer [15] and BEVDet series [8, 7, 14]. Furthermore, the auxiliary HoP approach is complementary to prevalent temporal modeling methods, leading to significant gains." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 432, + 286, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 432, + 286, + 551 + ], + "spans": [ + { + "bbox": [ + 49, + 432, + 286, + 551 + ], + "type": "text", + "content": "Extensive experiments are conducted to evaluate the effectiveness of the proposed HoP on the nuScenes dataset [1]. We choose the representative methods, including BEVFormer and BEVDet4D-Depth [7, 14] to evaluate our method. To be specific, we obtain " + }, + { + "bbox": [ + 49, + 432, + 286, + 551 + ], + "type": "inline_equation", + "content": "55.8\\%" + }, + { + "bbox": [ + 49, + 432, + 286, + 551 + ], + "type": "text", + "content": " NDS with ResNet-101-DCN and " + }, + { + "bbox": [ + 49, + 432, + 286, + 551 + ], + "type": "inline_equation", + "content": "60.3\\%" + }, + { + "bbox": [ + 49, + 432, + 286, + 551 + ], + "type": "text", + "content": " NDS with VoVNet-99 when evaluating HoP on nuScenes val. Surprisingly, HoP achieves " + }, + { + "bbox": [ + 49, + 432, + 286, + 551 + ], + "type": "inline_equation", + "content": "68.5\\%" + }, + { + "bbox": [ + 49, + 432, + 286, + 551 + ], + "type": "text", + "content": " NDS and " + }, + { + "bbox": [ + 49, + 432, + 286, + 551 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 49, + 432, + 286, + 551 + ], + "type": "text", + "content": " mAP with ViT-L [4] on nuScenes test, surpassing all the 3D object detectors on the leaderboard by a large margin." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 552, + 286, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 552, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 49, + 552, + 286, + 574 + ], + "type": "text", + "content": "In conclusion, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 584, + 286, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 584, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 59, + 584, + 286, + 656 + ], + "type": "text", + "content": "- We propose a novel temporal enhanced training scheme, namely Historical Object Prediction (HoP), to encourage more accurate BEV feature learning. HoP can force the model to capture the spatial semantics and temporal motion of objects in the historical frame during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 665, + 286, + 713 + ], + "type": "text", + "content": "- We design a temporal decoder that consists of a short-term decoder and a long-term decoder to provide reliable spatial localization and accurate motion estimation of objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 320, + 72, + 544, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 72, + 544, + 131 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 544, + 131 + ], + "type": "text", + "content": "- We equip the competitive 3D object detector baselines with our approach and yield significant improvements on the nuScenes dataset. Surprisingly, HoP with ViT-L achieves " + }, + { + "bbox": [ + 320, + 72, + 544, + 131 + ], + "type": "inline_equation", + "content": "68.5\\%" + }, + { + "bbox": [ + 320, + 72, + 544, + 131 + ], + "type": "text", + "content": " NDS and " + }, + { + "bbox": [ + 320, + 72, + 544, + 131 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 320, + 72, + 544, + 131 + ], + "type": "text", + "content": " mAP on nuScenes test, establishing the new state-of-the-art performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 144, + 397, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 144, + 397, + 156 + ], + "spans": [ + { + "bbox": [ + 308, + 144, + 397, + 156 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 163, + 476, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 476, + 176 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 476, + 176 + ], + "type": "text", + "content": "2.1. Multi-view 3D Object Detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 182, + 544, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 182, + 544, + 242 + ], + "spans": [ + { + "bbox": [ + 308, + 182, + 544, + 242 + ], + "type": "text", + "content": "Modern multi-view methods for 3D object detection could be mainly categorized into two branches, LSS-based [25] and query-based methods. We will introduce how these two methods differ in feature aggregation and leave the temporal modeling part in the next section." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 243, + 544, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 544, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 544, + 338 + ], + "type": "text", + "content": "BEVDet [8] is a representative method in LSS-based methods. Following the Lift-Splat-Shoot paradigm, the method first explicitly estimates depth for every image pixel, then lifts the 2D features to 3D voxels according to the depth, and finally splats 3D features to BEV features and conducts object detection on it. BEVDepth [14] further improves the view transformation module with explicit depth supervision." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 339, + 544, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 339, + 544, + 505 + ], + "spans": [ + { + "bbox": [ + 308, + 339, + 544, + 505 + ], + "type": "text", + "content": "Query-based methods typically employ learnable queries to aggregate 2D image features by attention [27, 37] mechanism. DETR3D [30] utilizes object queries for predicting 3D positions and projects them back to 2D coordinates to obtain the corresponding features. Graph-DETR3D [2] and Sparse4D [17] enhance it respectively with a learnable 3D graph and sparse 4D sampling. PETR [18] encodes 3D position into image features and therefore directly query with global 2D features. BEVFormer [15] leverages grid-shaped BEV queries to interact with 2D features by deformable attention; on its basis, BEVFormerv2 [32] introduces perspective supervision. Besides, PolarFormer and Ego3RT [22, 9] advocates modeling BEV queries in polar coordinates to fit the real-world scenario." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 308, + 514, + 544, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 544, + 538 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 544, + 538 + ], + "type": "text", + "content": "2.2. Temporal Modeling for Multi-view 3D Object Detection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 546, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 544, + 712 + ], + "type": "text", + "content": "The motion information in autonomous driving scenes has been increasingly explored to utilize the temporal cues for improving detection performance in recent 3D perception frameworks. BEVFormer [15] proposes the temporal self-attention mechanism to dynamically fuse the previous BEV features by deformable attention [37] in an RNN manner. BEVDet4D [7] introduces the temporal modeling to lift BEVDet [8] to spatial-temporal 4D space. The 3D position embedding (3D PE) in PETR [18] is extended by PETRv2 [19] with the temporal alignment. BEVStereo [12] and STS [31] both leverage temporal views for constructing multi-view stereo by an effective temporal stereo method. SOLOFusion [24] fully exploits the synergy of short-term and long-term temporal information that is highly comple" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "text", + "content": "3782" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 72, + 519, + 259 + ], + "blocks": [ + { + "bbox": [ + 75, + 72, + 519, + 259 + ], + "lines": [ + { + "bbox": [ + 75, + 72, + 519, + 259 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 519, + 259 + ], + "type": "image", + "image_path": "96b0d700a01eee3b95c2722bc1a9465ce4b1bb3630d51fb6e32d0da7d3b79612.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 270, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 545, + 293 + ], + "type": "text", + "content": "Figure 2: Framework of our Historical Object Prediction (HoP). The auxiliary branches are discarded during evaluation. The symbol © denotes feature concatenation and fusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 310, + 289, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 289, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 289, + 392 + ], + "type": "text", + "content": "mentary as well as achieving state-of-the-art performance. BEVFormerV2 [32] uses the bidirectional temporal encoder to utilize both history and future BEV features, improving the performance by a large margin. In contrast with aforementioned works, we provide a novel temporal enhanced training paradigm to improve the temporal modeling without additional inference costs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 403, + 104, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 104, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 104, + 415 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 423, + 167, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 167, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 167, + 435 + ], + "type": "text", + "content": "3.1. Overall Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 441, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 502 + ], + "type": "text", + "content": "As illustrated in Figure 2, HoP can be easily incorporated into many multi-view 3D object detectors, e.g., BEVFormer and BEVDet4D-Depth. We transform the extrinsic parameters of previous frames to the coordinate system of the current ego to perform temporal alignment following BEVDet." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 509, + 202, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 202, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 202, + 521 + ], + "type": "text", + "content": "3.2. Historical Object Prediction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "content": "Pipeline. The architecture of our Historical Object Prediction (HoP) is composed of a temporal decoder " + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "content": " and an object decoder " + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "content": ". Given the BEV feature sequence " + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{B}_{t-N}, \\cdots, \\mathbf{B}_t\\}" + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "content": " that consists of " + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "content": " historical BEV features and the current BEV feature, we denote the corresponding 3D ground truths as " + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{G}_{t-N}, \\cdots, \\mathbf{G}_t\\}" + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "content": ". If the BEV feature at timestamp " + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "inline_equation", + "content": "t-k" + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "content": " is discarded, we use the remaining BEV feature sequence " + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{B}_{t-N}, \\cdots, \\mathbf{B}_t\\} - \\{\\mathbf{B}_{t-k}\\}" + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "content": " (denoted as " + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^{rem}" + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "content": ") to predict the 3D objects at timestamp " + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "inline_equation", + "content": "t-k" + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "content": ". To be specific, we adopt " + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}" + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "content": " to transform the remaining BEV features to obtain the 3D predictions " + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_{t-k}" + }, + { + "bbox": [ + 46, + 526, + 287, + 659 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 72, + 666, + 287, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 666, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 72, + 666, + 287, + 681 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {P}} _ {t - k} = \\hat {\\mathcal {D}} \\left(\\mathcal {T} \\left(\\left\\{\\mathbf {B} _ {t - N}, \\dots , \\mathbf {B} _ {t} \\right\\} - \\left\\{\\mathbf {B} _ {t - k} \\right\\}\\right)\\right). \\tag {1}", + "image_path": "1b4142e0d37363dca3db1363fefb18c108402b684afb5a30794224c3d1b932c3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "The temporal decoder " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " that models both short-term and long-term information is developed to reconstruct the BEV" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "text", + "content": "feature at timestamp " + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "text", + "content": ". The object decoder " + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}" + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "text", + "content": " is attached to generate the predictions " + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_{t - k}" + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "text", + "content": " using the pseudo BEV feature. Considering the ego-motion and temporal alignment, we transform 3D coordinates of " + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{t - k}" + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "text", + "content": " into the coordinate system of current frame " + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "text", + "content": " (denoted as " + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{G}}_{t - k}" + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "text", + "content": "). The final training objective is to minimize the differences between " + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_{t - k}" + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{G}}_{t - k}" + }, + { + "bbox": [ + 304, + 309, + 545, + 392 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 394, + 546, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 546, + 548 + ], + "type": "text", + "content": "The input to the temporal decoder is the remaining set of BEV features consisting of (i) short-term BEV features " + }, + { + "bbox": [ + 304, + 394, + 546, + 548 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{B}_{t - k - 1},\\mathbf{B}_{t - k + 1}\\}" + }, + { + "bbox": [ + 304, + 394, + 546, + 548 + ], + "type": "text", + "content": ", and (ii) long-term BEV features " + }, + { + "bbox": [ + 304, + 394, + 546, + 548 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{B}_{t - N},\\dots ,\\mathbf{B}_t\\} -\\{\\mathbf{B}_{t - k}\\}" + }, + { + "bbox": [ + 304, + 394, + 546, + 548 + ], + "type": "text", + "content": ". We first add temporal positional embeddings to BEV features in this full set. Our temporal decoder reconstructs the feature " + }, + { + "bbox": [ + 304, + 394, + 546, + 548 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{B}}_{t - k}" + }, + { + "bbox": [ + 304, + 394, + 546, + 548 + ], + "type": "text", + "content": " by utilizing the spatial semantics and temporal cues of the BEV feature sequence. More specifically, we present a temporal decoder that consists of two separate branches with their special expertise to capture short-term and long-term information. The learned strong spatial-temporal representation can help the network better estimate the object location in frame at timestamp " + }, + { + "bbox": [ + 304, + 394, + 546, + 548 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 304, + 394, + 546, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "text", + "content": "Short-term temporal decoder. Thanks to the high temporal correlation between adjacent frames, the short-term temporal decoder only operates on the adjacent BEV feature set " + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{B}_{t - k - 1},\\mathbf{B}_{t - k + 1}\\}" + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "text", + "content": " (denoted as " + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^{adj}" + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "text", + "content": "), building a detailed spatial representation in BEV space. We first define a grid-shaped learnable short-term BEV queries " + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{t - k}^{short}\\in" + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "text", + "content": " refer to the spatial shape of the BEV plane. The short-term BEV queries aggregate the spatial information and model the short-term motion from " + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{t - k - 1}" + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{t - k + 1}" + }, + { + "bbox": [ + 304, + 549, + 547, + 681 + ], + "type": "text", + "content": " through the short-term temporal attention, which can be formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 329, + 689, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 329, + 689, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {B}} _ {t - k} ^ {s h o r t} = \\sum_ {\\mathbf {V} \\in \\mathbf {B} ^ {a d j}} \\operatorname {D e f o r m A t t n} \\left(\\mathbf {Q} _ {t - k, p} ^ {s h o r t}, p, \\mathbf {V}\\right), \\quad (2)", + "image_path": "a13c7ec6277f5bace4b7bf332d314fc851e9983a0c3a9a55d1314ba69a4b3190.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3783" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " is the spatial index in the BEV plane, DeformAttn(" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "q, p, x" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": ") refers to the deformable attention [37] with query " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": ", reference point " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " and input features " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": ". We further feed " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{B}}_{t-k}^{short}" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " into a feed-forward network [27] and obtain the output of this short-term branch. However, it is still challenging to precisely construct the temporal relations of the same objects between " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{t-k}" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " and other BEV features with only two adjacent frames." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 290, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 290, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 290, + 373 + ], + "type": "text", + "content": "Long-term temporal decoder. The long-term temporal processes the whole remaining BEV set to perceive the motion clues over long-term history, which increases the localization potential [24] and contributes to more accurate localization. Therefore, these two branches in the temporal decoder are complementary to each other. As for long-term motion, it is intuitive that we only focus on the spatial motion of the same objects in the bird's eye view, ignoring the height information of objects. For most 3D object detectors with 2D BEV features, e.g., BEVFormer and BEVDet, the height information has been flattened to the feature embeddings. Accordingly, we first employ a channel reduction operation to the input set " + }, + { + "bbox": [ + 46, + 168, + 290, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^{rem}" + }, + { + "bbox": [ + 46, + 168, + 290, + 373 + ], + "type": "text", + "content": " to prune the height information and achieve better training efficiency. Given the learnable long-term BEV queries " + }, + { + "bbox": [ + 46, + 168, + 290, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{t - k}^{long}\\in \\mathbb{R}^{H\\times W\\times C / r}" + }, + { + "bbox": [ + 46, + 168, + 290, + 373 + ], + "type": "text", + "content": " and reduction layer with parameter " + }, + { + "bbox": [ + 46, + 168, + 290, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^r\\in \\mathbb{R}^{C\\times C / r}" + }, + { + "bbox": [ + 46, + 168, + 290, + 373 + ], + "type": "text", + "content": ", we can capture the long-term dependencies as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 380, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 380, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 58, + 380, + 287, + 407 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {B}} _ {t - k} ^ {l o n g} = \\sum_ {\\mathbf {V} \\in \\mathbf {B} ^ {r e m}} \\mathrm {D e f o r m A t t n} (\\mathbf {Q} _ {t - k, p} ^ {l o n g}, p, \\mathbf {V W} ^ {r}), (3)", + "image_path": "e9759760deb01039d742122e96a2053dd3781a61561cfde52ca7a2b708a31450.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 414, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 414, + 287, + 472 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 414, + 287, + 472 + ], + "type": "text", + "content": " denotes the reduction ratio and is set as 4 by default. After the feed-forward network, the long-term decoder output the BEV features. Finally, we concatenate the short-term and long-term BEV features and perform feature fusion by " + }, + { + "bbox": [ + 46, + 414, + 287, + 472 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 414, + 287, + 472 + ], + "type": "text", + "content": " convolution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "content": "Object decoder. The reconstructed BEV feature " + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{B}}_{t - k}" + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "content": " is further processed by a lightweight object decoder. The decoder is used to generate 3D predictions " + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_{t - k}" + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "content": " upon the BEV feature and thus its implementation is flexible. Variants of BEV detection heads are an obvious choice and we will consider multiple instantiations in our experiments. After obtaining " + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_{t - k}" + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "content": ", we should align the coordinates between learning targets " + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{t - k}" + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "content": " and the final predictions " + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}_{t - k}" + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "content": ". For clarity, we first denote the ego coordinate as " + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "inline_equation", + "content": "e(t)" + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "content": " at frame " + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "content": ". In our implementation, we transform the extrinsic parameters of previous frames to the coordinate system of the current ego, thus BEV features and predictions at different timestamps share the same " + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "inline_equation", + "content": "e(t)" + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "content": ". To simplify the learning targets, 3D coordinates of " + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{t - k}" + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "content": " should be transformed into " + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "inline_equation", + "content": "e(t)" + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "content": ". Considering the ego motion, we convert " + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{t - k}" + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{G}}_{t - k}" + }, + { + "bbox": [ + 46, + 473, + 290, + 666 + ], + "type": "text", + "content": " by ego transformation matrix as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 673, + 287, + 691 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 673, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 118, + 673, + 287, + 691 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {G}} _ {t - k} = \\mathbf {T} _ {e (t - k)} ^ {e (t)} \\mathbf {G} _ {t - k}, \\tag {4}", + "image_path": "0f09416a3cf227acaab930e9011647d9ead25349293e4229e1be503848417a08.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 699, + 289, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 699, + 289, + 716 + ], + "spans": [ + { + "bbox": [ + 47, + 699, + 289, + 716 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 699, + 289, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{e(t - k)}^{e(t)}" + }, + { + "bbox": [ + 47, + 699, + 289, + 716 + ], + "type": "text", + "content": " is the transformation matrix from the source" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 539, + 162 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 539, + 162 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 539, + 162 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 539, + 162 + ], + "type": "image", + "image_path": "f4e9e4ac013edac8e17ea90fa39be6f633033e7d11c401c255118b8721ae347e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 169, + 547, + 206 + ], + "lines": [ + { + "bbox": [ + 304, + 169, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 547, + 206 + ], + "type": "text", + "content": "Figure 3: Illustration of the temporal multi-stage decoder. Object queries are enhanced by their counterparts derived from the last frame, forming a recurrent connection." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 226, + 545, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 226, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 545, + 249 + ], + "type": "text", + "content": "coordinate system " + }, + { + "bbox": [ + 304, + 226, + 545, + 249 + ], + "type": "inline_equation", + "content": "e(t - k)" + }, + { + "bbox": [ + 304, + 226, + 545, + 249 + ], + "type": "text", + "content": " to the target coordinate system " + }, + { + "bbox": [ + 304, + 226, + 545, + 249 + ], + "type": "inline_equation", + "content": "e(t)" + }, + { + "bbox": [ + 304, + 226, + 545, + 249 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 250, + 546, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 250, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 250, + 546, + 441 + ], + "type": "text", + "content": "Discussion. The act of discarding a frame in HoP might share similarity with masked image modeling works, like MAE [5]. However, there are several points where HoP differs greatly from those works. First, HoP is not a self-supervised learning method as it requires 3D ground-truth boxes for training. HoP leverages short-term and long-term branches to extract diverse information from objects with different velocities to predict 3D targets in the discarded frame. In contrast, masked image modeling methods generally choose pixels or features as reconstruction targets, but predicting features fails to bring significant gains as HoP in Table 6c. Moreover, MAE is adopted for backbone pretraining. HoP is designed for 3D BEV detection to enhance temporal learning by predicting objects in the discarded frame. In fact, HoP is compatible with MAE since we can use a MAE pretrained ViT as the backbone." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 449, + 491, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 491, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 491, + 462 + ], + "type": "text", + "content": "3.3. Historical Temporal Query Fusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 467, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 550 + ], + "type": "text", + "content": "Apart from the HoP method that exploits temporal information in BEV feature level, we also explore query-level temporal modeling in query-based methods, e.g., BEV-Former, by fusing historical object queries into current queries. This fusion step provides current queries with an initialization of history perception, forming a refining process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 551, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 551, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 551, + 545, + 588 + ], + "type": "text", + "content": "The BEV feature sequence " + }, + { + "bbox": [ + 304, + 551, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{B}_{t - N},\\dots ,\\mathbf{B}_t\\}" + }, + { + "bbox": [ + 304, + 551, + 545, + 588 + ], + "type": "text", + "content": " is inferred in the time order. For the timestamp " + }, + { + "bbox": [ + 304, + 551, + 545, + 588 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 304, + 551, + 545, + 588 + ], + "type": "text", + "content": " we perform detection with the pre-defined object queries " + }, + { + "bbox": [ + 304, + 551, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 304, + 551, + 545, + 588 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 378, + 594, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 594, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 378, + 594, + 545, + 609 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {O}} _ {t - k} = \\mathcal {D} (\\mathbf {B} _ {t - k}, \\mathbf {O}), \\tag {5}", + "image_path": "f4551f4b5d37b1db819cf077c087225a876bca2d54e514f7d9e403c1e8e8a520.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 617, + 518, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 518, + 630 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 518, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 617, + 518, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 617, + 518, + 630 + ], + "type": "text", + "content": " is the BEV decoder and " + }, + { + "bbox": [ + 305, + 617, + 518, + 630 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{O}}_{t - k}" + }, + { + "bbox": [ + 305, + 617, + 518, + 630 + ], + "type": "text", + "content": " is its output." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 631, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 631, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 547, + 715 + ], + "type": "text", + "content": "Each object query learns to specialize in certain areas in the BEV plane, so the output queries contain rich semantic information from the regions. When detecting on the current frame, we refine these output historical queries instead of starting from the pre-defined ones to make the learning process easier. For example, if an object query detects a still object in the last frame, its historical counterpart could help" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3784" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 549, + 189 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 549, + 189 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 549, + 189 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 549, + 189 + ], + "type": "table", + "html": "
MethodBackboneQuery-basedLiDAREpochNDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVDepth [14]R101-DCN‡✓90†0.5380.418-----
UVTR [13]R101-DCN‡✓240.4830.3790.7310.2670.3500.5100.200
BEVFormer [15]R101-DCN‡✓240.5170.4160.6730.2740.3720.3940.198
PETRv2 [19]R101-DCN‡✓240.5240.4210.6810.2670.3570.3770.186
PolarFormer [9]R101-DCN‡✓240.5280.4320.6480.2700.3480.4090.201
Sparse4D [17]R101-DCN‡✓240.5410.4360.6330.2790.3630.3170.177
HoP-BEVFormerR101-DCN‡✓240.5580.4540.5650.2650.3270.3370.194
", + "image_path": "f80f1116b140f7b8e24da335820c2bbae4f4a07b4ab94cb406347b0a473fb6fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 196, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 196, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 547, + 220 + ], + "type": "text", + "content": "Table 1: Comparison of recent works on the nuScenes val set. " + }, + { + "bbox": [ + 46, + 196, + 547, + 220 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 196, + 547, + 220 + ], + "type": "text", + "content": " indicates methods with CBGS which will elongate 1 epoch into 4.5 epochs. " + }, + { + "bbox": [ + 46, + 196, + 547, + 220 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 196, + 547, + 220 + ], + "type": "text", + "content": " notes the backbone is initialized from FCOS3D [28] backbone. LiDAR: trained with LiDAR supervision." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 241, + 286, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 286, + 289 + ], + "type": "text", + "content": "it quickly locate the object without searching on its area in the BEV plane again. On the other hand, historical information from a moving object also contributes to localizing and detecting its velocity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 289, + 286, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 286, + 314 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 286, + 314 + ], + "type": "text", + "content": "As shown in Figure 3, to infer on the BEV feature " + }, + { + "bbox": [ + 47, + 289, + 286, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{t - k}" + }, + { + "bbox": [ + 47, + 289, + 286, + 314 + ], + "type": "text", + "content": " we first collect historical object queries " + }, + { + "bbox": [ + 47, + 289, + 286, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{t - k}^{his}" + }, + { + "bbox": [ + 47, + 289, + 286, + 314 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 319, + 287, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 319, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 113, + 319, + 287, + 335 + ], + "type": "interline_equation", + "content": "\\mathbf {O} _ {t - k} ^ {h i s} = \\operatorname {M L P} \\left(\\bar {\\mathbf {O}} _ {t - k - 1}\\right), \\tag {6}", + "image_path": "4f808ec96532149099974e33741a3adc15b5e4eddcf356ef2218e3a03fab6e60.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 340, + 286, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 286, + 412 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 286, + 412 + ], + "type": "text", + "content": "where MLP is a multi-layer perceptron with two linear layers and an expansion ratio of 4, which is used to fit the history query into the current timestamp. Note that we only use historical queries from the last timestamp because we hypothesize that queries from more distant history possess little information overlap with currently visible objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 413, + 286, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 413, + 286, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 286, + 436 + ], + "type": "text", + "content": "Then we merge them with the pre-defined object queries " + }, + { + "bbox": [ + 47, + 413, + 286, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 47, + 413, + 286, + 436 + ], + "type": "text", + "content": " as Equation 7," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 441, + 287, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 441, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 102, + 441, + 287, + 456 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {O}} _ {t - k} = \\mathcal {D} \\left(\\mathbf {B} _ {t - k}, \\mathbf {O} + \\mathbf {O} _ {t - k} ^ {h i s}\\right). \\tag {7}", + "image_path": "157356ca11c808f3ee530f48b46269db370f9276f5ca7251ab62e37864a41613.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 462, + 286, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 286, + 498 + ], + "type": "text", + "content": "Considering the first shown-up objects in the current frame, the merge process takes both historical queries and predefined queries into account." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 508, + 128, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 128, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 128, + 521 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 527, + 164, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 164, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 164, + 539 + ], + "type": "text", + "content": "4.1. Dataset and Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": "We validate our method on the popular nuScenes[1] dataset. The nuScenes dataset contains 1000 driving scenes in total, which are split into 700, 150, and 150 respectively for training, validation, and testing. Each scene lasts for 20 seconds and annotations of 3D boxes are supplied every 0.5s in the keyframe. There are 6 cameras surrounding the vehicle, providing images covering the whole 360-degree FOV. As for evaluation, we adopt the standard metrics for nuScenes detection task. The final nuScenes detection score (NDS) is a weighted sum of mean Average Precision (mAP) and five True Positive metrics including Average Translation Error (ATE), Average Scale Error (ASE), Average Orientation Error (AOE), Average Velocity Error (AVE), and Average Attribute Error (AAE)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 240, + 440, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 240, + 440, + 253 + ], + "spans": [ + { + "bbox": [ + 306, + 240, + 440, + 253 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 261, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 545, + 429 + ], + "type": "text", + "content": "We mainly implement our method on the optimized version of BEVFormer [15] (denoted as BEVFormer-opt), where we generate the current BEV feature with resolution " + }, + { + "bbox": [ + 304, + 261, + 545, + 429 + ], + "type": "inline_equation", + "content": "200 \\times 200" + }, + { + "bbox": [ + 304, + 261, + 545, + 429 + ], + "type": "text", + "content": " by 6 BEV spatial encoder layers and fuse the current and 4 previous BEV features by a BEV temporal encoder. All models are trained with a mini-batch of 8, an initial learning rate of 0.0002, and a weight decay of 0.01 for 24 epochs without CBGS [36]. Unless otherwise specified, we use ResNet-50 [6] pre-trained on COCO [16] as the image backbone, and the image size is processed to " + }, + { + "bbox": [ + 304, + 261, + 545, + 429 + ], + "type": "inline_equation", + "content": "1408 \\times 512" + }, + { + "bbox": [ + 304, + 261, + 545, + 429 + ], + "type": "text", + "content": ". Both image and BEV augmentation are adopted following BEVDet [8]. Besides, we also adopt the BEVDet4D-Depth [7] as baseline models and follow the original settings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 439, + 391, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 391, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 391, + 451 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 459, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 531 + ], + "type": "text", + "content": "To compare with previous state-of-the-art 3D object detection methods, we adopt ResNet101-DCN [6, 3] initialized from FCOS3D [28] and VoVNet-99 [10, 11] initialized from DD3D [23] checkpoint. We also enlarge the input resolution to " + }, + { + "bbox": [ + 304, + 459, + 545, + 531 + ], + "type": "inline_equation", + "content": "1600 \\times 640" + }, + { + "bbox": [ + 304, + 459, + 545, + 531 + ], + "type": "text", + "content": " and use 1500 object queries. The historical temporal query fusion is employed in Table 1 and 2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 532, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 662 + ], + "type": "text", + "content": "We report the results of our HoP with the ResNet101-DCN backbone on nuScenes val in Table 1. Surprisingly, our method outperforms BEVDepth with CBGS strategy and depth supervision from LIDAR over " + }, + { + "bbox": [ + 304, + 532, + 545, + 662 + ], + "type": "inline_equation", + "content": "+2.0\\%" + }, + { + "bbox": [ + 304, + 532, + 545, + 662 + ], + "type": "text", + "content": " NDS and " + }, + { + "bbox": [ + 304, + 532, + 545, + 662 + ], + "type": "inline_equation", + "content": "+3.6\\%" + }, + { + "bbox": [ + 304, + 532, + 545, + 662 + ], + "type": "text", + "content": " mAP under comparable complexity levels. Besides, HoP achieves the best performance compared with other competitive query-based object detectors, e.g., PETRv2 [19] and PolarFormer [9]. We also compared HoP with other state-of-the-art algorithms on nuScenes test, and still obtain the best performance of " + }, + { + "bbox": [ + 304, + 532, + 545, + 662 + ], + "type": "inline_equation", + "content": "61.2\\%" + }, + { + "bbox": [ + 304, + 532, + 545, + 662 + ], + "type": "text", + "content": " NDS and " + }, + { + "bbox": [ + 304, + 532, + 545, + 662 + ], + "type": "inline_equation", + "content": "52.8\\%" + }, + { + "bbox": [ + 304, + 532, + 545, + 662 + ], + "type": "text", + "content": " mAP." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "To show the scalability of our model to stronger backbones, we apply HoP to BEVDet4D-Depth with ViT-L [4] as the backbone. The ViT-L we employ is initialized with the Co-DETR [38] pre-trained on the 2D detection bench" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3785" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 556, + 204 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 556, + 204 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 556, + 204 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 556, + 204 + ], + "type": "table", + "html": "
MethodBackboneQuery-basedLiDAREpochNDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVDepth* [14]V2-99✓\\(90^{\\dagger}\\)0.6000.5030.4450.2450.3780.3200.126
DETR3D [30]V2-99✓240.4790.4120.6410.2550.3940.8450.133
UVTR [13]V2-99✓240.5510.4720.5770.2530.3910.5080.123
BEVFormer [15]V2-99✓240.5690.4810.5820.2560.3750.3780.126
PETRv2 [19]V2-99✓240.5820.4900.5610.2430.3610.3430.120
Sparse4D [17]V2-99✓240.5950.5110.5330.2630.3690.3170.124
HoP-BEVFormerV2-99✓240.6030.5170.5010.2450.3460.3620.105
HoP-BEVFormert†V2-99✓240.6120.5280.4910.2420.3320.3430.109
", + "image_path": "d8232adb826dea4cae958e3e7b6cd051531b25c04c725b148492d786ae574eaf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 243, + 533, + 319 + ], + "blocks": [ + { + "bbox": [ + 47, + 210, + 545, + 235 + ], + "lines": [ + { + "bbox": [ + 47, + 210, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 210, + 545, + 235 + ], + "type": "text", + "content": "Table 2: Comparison of recent works on the nuScenes test set. * indicates test-time augmentation. † means we do not detach 1 historical frame. VoVNet-99 (V2-99) [10] was pre-trained on the depth estimation task with extra data [23]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 243, + 533, + 319 + ], + "lines": [ + { + "bbox": [ + 59, + 243, + 533, + 319 + ], + "spans": [ + { + "bbox": [ + 59, + 243, + 533, + 319 + ], + "type": "table", + "html": "
MethodBackboneTTANDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
SOLOFusion [24]ConvNeXt-B [21]0.6190.5400.4530.2570.3760.2760.148
BEVFormerV2 [32]InternImage-XL [29]0.6480.5800.4480.2620.3420.2380.128
BEVDet-Gamma [7]Swin-B [20]✓0.6640.5860.3750.2430.3770.1740.123
HoP-BEVDet4D-DepthViT-L [4]0.6850.6240.3670.2490.3540.1710.131
", + "image_path": "7563c01f94516d95ccfa1500b1b0400e2775e6328d58eefb9228859fa6e99651.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 324, + 529, + 337 + ], + "lines": [ + { + "bbox": [ + 62, + 324, + 529, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 529, + 337 + ], + "type": "text", + "content": "Table 3: Comparisons on the nuScenes camera-only 3D detection leaderboard. TTA denotes test-time augmentation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 355, + 288, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 288, + 450 + ], + "type": "text", + "content": "mark Objects365 [26]. We use 8 frames into the future to provide additional information from the future. To reduce the training time, we first train the ViT-L using FCOS3D on the nuScenes trainval for 36 epochs. Then we only train HoP-BEVDet4D-Depth for 5 epochs with CBGS. Without test-time augmentation, HoP pushes the current best result to a new record of " + }, + { + "bbox": [ + 46, + 355, + 288, + 450 + ], + "type": "inline_equation", + "content": "68.5\\%" + }, + { + "bbox": [ + 46, + 355, + 288, + 450 + ], + "type": "text", + "content": " NDS and " + }, + { + "bbox": [ + 46, + 355, + 288, + 450 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 46, + 355, + 288, + 450 + ], + "type": "text", + "content": " mAP on nuScenes test." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 464, + 147, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 147, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 147, + 475 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 484, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 531 + ], + "type": "text", + "content": "Unless stated otherwise, we conduct the ablation experiments on BEVFormer-opt with a ResNet-50 backbone. We follow the original settings when adopting BEVDet4D as the baseline." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": "Effectiveness of each component. We perform a component-wise ablation to thoroughly analyze the effect of each component in Table 4. Surprisingly, incorporating the historical temporal prediction into the optimized BEVFormer brings significant performance gains " + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "(+1.8\\%)" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": " NDS, " + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "+1.4\\%" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": " mAP) over the baseline. The overall improvements hold when we apply the training strategy to BEVDet4D-Depth, achieving " + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "+1.6\\%" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": " NDS and " + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "+1.4\\%" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": " mAP improvements, respectively. Alternatively, the temporal information provided by historical queries also contributes to the performance improvements for BEVFormer, where a " + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "+1.2\\%" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": " NDS gain is observed. Without detaching the gradients of historical BEV features in the last two layers of BEV encoder, the model can be further improved to " + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "53.9\\%" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": " NDS. These results prove the effectiveness and generalization of our HoP." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 355, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 545, + 582 + ], + "type": "text", + "content": "Optimized BEVFormer baseline. Table 5 illustrates the critical designs of our optimized BEVFormer baseline. We can observe setting the dropout rate within the transformer as 0 significantly improves the performance and look forward twice [34] is an essential factor to improve mAP. Besides, we replace the heavy backbone with the ResNet50 and decrease the input resolution to " + }, + { + "bbox": [ + 304, + 355, + 545, + 582 + ], + "type": "inline_equation", + "content": "1408 \\times 512" + }, + { + "bbox": [ + 304, + 355, + 545, + 582 + ], + "type": "text", + "content": " to improve the training efficiency of BEVFormer. At the same time, we also introduce step learning rate decay and data augmentation, which includes BEV augmentation, random flipping, cropping, and rotation, to achieve better performance without inference costs increase. Decoupling the BEV temporal encoder from the spatial encoder and using 4 previous frames, is crucial in the later promotion process. Combining these aforementioned modifications on BEVFormer results in our optimized baseline BEVFormer-opt. Finally, we achieve comparable performance with the original BEVFormer-R101-DCN while running much faster and requiring fewer parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "type": "text", + "content": "Temporal decoder design. Table 6a illustrates the key designs of our temporal decoder for prediction targets prediction. The performance is always improved when using the structure of one individual decoder alone, reaching " + }, + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "type": "inline_equation", + "content": "52.9\\%" + }, + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "type": "text", + "content": " NDS and " + }, + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "type": "inline_equation", + "content": "52.7\\%" + }, + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "type": "text", + "content": " NDS with the short-term and the long-term decoder, respectively. We achieve the best performance with " + }, + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "type": "inline_equation", + "content": "53.1\\%" + }, + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "type": "text", + "content": " NDS and " + }, + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "type": "inline_equation", + "content": "42.8\\%" + }, + { + "bbox": [ + 304, + 587, + 545, + 696 + ], + "type": "text", + "content": " mAP when both short-term and long-term decoders are adopted, demonstrating the complementarity between these two branches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Object decoder design. The object decoder detects the 3D" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3786" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 530, + 180 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 530, + 180 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 530, + 180 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 530, + 180 + ], + "type": "table", + "html": "
MethodHoPHQ FusionNDS↑mAP↑mATE↓mASE↓mAOE↓mAVE↓mAAE↓
BEVFormer-opt0.5130.4140.6550.2740.4470.3680.194
BEVFormer-opt✓0.5250.4250.6600.2740.3950.3570.188
BEVFormer-opt✓0.5310.4280.6380.2690.3520.3690.185
BEVFormer-opt†✓0.5390.4350.6290.2680.3420.3600.184
BEVFormer-opt†✓✓0.5440.4390.6070.2650.3540.3400.193
BEVDet4D-Depth0.4930.3850.6320.2830.5810.2890.212
BEVDet4D-Depth✓0.5090.3990.6080.2720.5410.2810.205
", + "image_path": "d1b7dcc5e7867cb89fef11fdc7fb83278c791e06e9a3ba00aa16ec85f56300c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 208, + 541, + 327 + ], + "blocks": [ + { + "bbox": [ + 50, + 208, + 541, + 327 + ], + "lines": [ + { + "bbox": [ + 50, + 208, + 541, + 327 + ], + "spans": [ + { + "bbox": [ + 50, + 208, + 541, + 327 + ], + "type": "image", + "image_path": "f4ea1197608a1b9f8c8a2fd4c31a12563b9a932ca3237908f688b51aaaaffc0e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 337, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 337, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 546, + 361 + ], + "type": "text", + "content": "Figure 4: Visualization result of BEVFormer-opt and HoP-BEVFormer on nuScenes val set. The predicted boxes are marked in yellow, while the ground truth boxes are marked in blue in the LIDAR_TOP figure." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 363, + 285, + 483 + ], + "blocks": [ + { + "bbox": [ + 47, + 185, + 545, + 206 + ], + "lines": [ + { + "bbox": [ + 47, + 185, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 47, + 185, + 545, + 206 + ], + "type": "text", + "content": "Table 4: Component-wise ablations. † indicates we do not detach historical BEV features of last two layers in the BEV encoder. “HQ Fusion” refers to historical temporal query fusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 363, + 285, + 483 + ], + "lines": [ + { + "bbox": [ + 50, + 363, + 285, + 483 + ], + "spans": [ + { + "bbox": [ + 50, + 363, + 285, + 483 + ], + "type": "table", + "html": "
MethodNDSmAP#params (M)Training time (sec/iter)
BEVFormer-R101-DCN0.5170.41678.62.7
+ Dropout 00.5290.42478.62.7
+ Look forward twice0.5310.43078.62.7
+ Smaller backbone R500.4870.38857.72.0
+ Step LR decay0.4930.39157.72.0
+ Data augmentation0.4990.40457.72.0
+ Decoupled BEV encoder0.5210.42055.52.2
+ Smaller resolution0.5130.41455.51.6
BEVFormer-opt-R500.5130.41455.51.6
", + "image_path": "b918675bba87c0ab11a3b0fdec598d93e4d5fbc6ca94b2052907444ea7fa67ba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 487, + 276, + 498 + ], + "lines": [ + { + "bbox": [ + 58, + 487, + 276, + 498 + ], + "spans": [ + { + "bbox": [ + 58, + 487, + 276, + 498 + ], + "type": "text", + "content": "Table 5: Step-by-step optimizing the BEVFormer baseline." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 508, + 286, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 286, + 579 + ], + "type": "text", + "content": "targets with the reconstructed features, thus it can be flexibly designed, as studied in Table 6b. We find these detection heads can bring consistent gains. For example, ATSS [35] is extended to 3D object detection and brings " + }, + { + "bbox": [ + 46, + 508, + 286, + 579 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 46, + 508, + 286, + 579 + ], + "type": "text", + "content": " NDS and " + }, + { + "bbox": [ + 46, + 508, + 286, + 579 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 508, + 286, + 579 + ], + "type": "text", + "content": " mAP gains. We choose CenterPoint [33], which achieves the best performance as the default object decoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 581, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 286, + 675 + ], + "type": "text", + "content": "Prediction target. We also study the design that only predicts pseudo BEV features at timestamp " + }, + { + "bbox": [ + 46, + 581, + 286, + 675 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 46, + 581, + 286, + 675 + ], + "type": "text", + "content": " supervised by " + }, + { + "bbox": [ + 46, + 581, + 286, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{t - k}" + }, + { + "bbox": [ + 46, + 581, + 286, + 675 + ], + "type": "text", + "content": " generated by the BEV encoder in a self-distillation manner. As presented in Table 6c, the feature supervision signals only slightly improve the baseline while the 3D objects supervision boosts the performance by a large margin. In a summary, explicit supervision from 3D objects in the BEV space is crucial to HoP." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "text", + "content": "Choice of prediction index " + }, + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "text", + "content": ". The selection of prediction index " + }, + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 677, + 286, + 712 + ], + "type": "text", + "content": " is a critical design of our method and is ablated in Table 6d. We first study whether the model can benefit from" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": "predicting the 3D objects in the future frame, e.g., frame " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "k = -1" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": "). Interestingly, this future 3D object prediction also brings non-trivial gains but performs inferior to history prediction. The relatively weaker gains can be explained by the difference between a history prediction task and a future prediction task: it is much more challenging to predict 3D objects in the future frame at timestamp " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": " with only history frames from timestamp " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "t - N" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": ". Predicting objects in frame " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": " with both history frames from timestamp " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "t - N" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "t - k - 1" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": " and future frames from timestamp " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "t - k + 1" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": " is much easier. A challenging auxiliary task can disturb the learning progress and slightly decrease the performance gains. When predicting the objects of history frame " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": ", we do not detach the feature maps of frame " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "t - k + 1" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": " to allow the gradients back propagate through the both short-term and long-term temporal decoder. Therefore, we observe the method imposes a significant increase in GPU memory for " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "k > 1" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": " since the features at timestamp " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "t - k + 1" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": " are not detached. Accordingly, we choose " + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 365, + 545, + 604 + ], + "type": "text", + "content": " as 1 since it achieves the best trade-offs between accuracy and training costs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Training time. We present the training speed (seconds per training iteration) under different settings in Table 6e. This training speed is evaluated with 8 NVIDIA A100 GPUs. Compared with the baseline, HoP introduces an additional temporal decoder and object decoder during training, leading to an inevitable increase in training time. Thanks to the lightweight design of our temporal decoder and object decoder, we only observe a " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " increase (from 1.6 seconds to 1.9 seconds) in training time for BEVFormer baseline." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3787" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 78, + 176, + 129 + ], + "blocks": [ + { + "bbox": [ + 77, + 78, + 176, + 129 + ], + "lines": [ + { + "bbox": [ + 77, + 78, + 176, + 129 + ], + "spans": [ + { + "bbox": [ + 77, + 78, + 176, + 129 + ], + "type": "table", + "html": "
CaseNDSmAP
None0.5130.414
Short-term0.5280.426
Long-term0.5260.425
Both0.5310.428
", + "image_path": "e527cfff8c3741392057d9689c1114f39335a8c410c4358e80ca23bd89265acc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 166, + 192, + 228 + ], + "blocks": [ + { + "bbox": [ + 49, + 133, + 202, + 153 + ], + "lines": [ + { + "bbox": [ + 49, + 133, + 202, + 153 + ], + "spans": [ + { + "bbox": [ + 49, + 133, + 202, + 153 + ], + "type": "text", + "content": "(a) Temporal decoder design. Both shortterm and long-term can improve accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 166, + 192, + 228 + ], + "lines": [ + { + "bbox": [ + 60, + 166, + 192, + 228 + ], + "spans": [ + { + "bbox": [ + 60, + 166, + 192, + 228 + ], + "type": "table", + "html": "
kNDSmAPMemory
None0.5130.4141×
-10.5230.4201.6×
10.5310.4281.3×
20.5340.4302.3×
30.5360.4322.3×
", + "image_path": "48c65fe156436b8e9a2b054e22f8c2f4cb85313d19900c6da38733e07f0f852a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 246, + 78, + 349, + 129 + ], + "blocks": [ + { + "bbox": [ + 49, + 230, + 202, + 250 + ], + "lines": [ + { + "bbox": [ + 49, + 230, + 202, + 250 + ], + "spans": [ + { + "bbox": [ + 49, + 230, + 202, + 250 + ], + "type": "text", + "content": "(d) Prediction index. Predicting the objects in frame " + }, + { + "bbox": [ + 49, + 230, + 202, + 250 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 49, + 230, + 202, + 250 + ], + "type": "text", + "content": " achieves the best trade-offs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 246, + 78, + 349, + 129 + ], + "lines": [ + { + "bbox": [ + 246, + 78, + 349, + 129 + ], + "spans": [ + { + "bbox": [ + 246, + 78, + 349, + 129 + ], + "type": "table", + "html": "
CaseNDSmAP
None0.5130.414
Transformer0.5210.421
ATSS0.5290.426
CenterPoint0.5310.428
", + "image_path": "4ac19c40d1b1024ff36d9397a0301b16da0f464d8b6f840122e8ee143e5d28d0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 226, + 166, + 369, + 228 + ], + "blocks": [ + { + "bbox": [ + 220, + 133, + 372, + 153 + ], + "lines": [ + { + "bbox": [ + 220, + 133, + 372, + 153 + ], + "spans": [ + { + "bbox": [ + 220, + 133, + 372, + 153 + ], + "type": "text", + "content": "(b) Object decoder design. The decoder can be flexibly implemented." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 226, + 166, + 369, + 228 + ], + "lines": [ + { + "bbox": [ + 226, + 166, + 369, + 228 + ], + "spans": [ + { + "bbox": [ + 226, + 166, + 369, + 228 + ], + "type": "table", + "html": "
CaseNDSSpeed
BEVFormer-opt0.5131.6
HoP-BEVFormer-opt0.5311.9
HoP-BEVFormer-opt†0.5392.1
BEVDet4D-Depth0.4934.5
HoP-BEVDet4D-Depth0.5094.7
", + "image_path": "32c30931f1f9a8a17d2c34fe3d8a6c5bcb92a5ced2db0db291cfe3eb3a454e2a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 416, + 78, + 518, + 121 + ], + "blocks": [ + { + "bbox": [ + 220, + 230, + 372, + 251 + ], + "lines": [ + { + "bbox": [ + 220, + 230, + 372, + 251 + ], + "spans": [ + { + "bbox": [ + 220, + 230, + 372, + 251 + ], + "type": "text", + "content": "(e) Training speed. HoP slightly increases the training time." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 416, + 78, + 518, + 121 + ], + "lines": [ + { + "bbox": [ + 416, + 78, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 416, + 78, + 518, + 121 + ], + "type": "table", + "html": "
CaseNDSmAP
None0.5130.414
BEV feature0.5180.418
3D objects0.5310.428
", + "image_path": "0cf4b5a66ec1e278618d26aa564e06a1443e63051c8c6c2eefb15d4a2e4a97fd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 409, + 166, + 526, + 217 + ], + "blocks": [ + { + "bbox": [ + 391, + 133, + 542, + 153 + ], + "lines": [ + { + "bbox": [ + 391, + 133, + 542, + 153 + ], + "spans": [ + { + "bbox": [ + 391, + 133, + 542, + 153 + ], + "type": "text", + "content": "(c) Prediction target. The explicit supervi sions of 3D targets are critical." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 409, + 166, + 526, + 217 + ], + "lines": [ + { + "bbox": [ + 409, + 166, + 526, + 217 + ], + "spans": [ + { + "bbox": [ + 409, + 166, + 526, + 217 + ], + "type": "table", + "html": "
Connection formNDSmAP
None0.5130.414
Recurrent0.5250.425
Fully-Connected0.5210.422
Dense0.5230.421
", + "image_path": "4c8fda3f78cb10e9bc21ee91f72f66888343166467115f228433ab65130e79ef.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 391, + 230, + 543, + 251 + ], + "lines": [ + { + "bbox": [ + 391, + 230, + 543, + 251 + ], + "spans": [ + { + "bbox": [ + 391, + 230, + 543, + 251 + ], + "type": "text", + "content": "(f) Historical query collection form. All three connection forms bring improvements." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 258, + 546, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 546, + 282 + ], + "type": "text", + "content": "Table 6: Ablation experiments with ResNet-50 on nuScenes val. Default settings are marked in gray. † indicates we do not detach historical BEV features of last two layers in the BEV encoder." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 99, + 288, + 217, + 445 + ], + "blocks": [ + { + "bbox": [ + 99, + 288, + 217, + 445 + ], + "lines": [ + { + "bbox": [ + 99, + 288, + 217, + 445 + ], + "spans": [ + { + "bbox": [ + 99, + 288, + 217, + 445 + ], + "type": "image", + "image_path": "ae303011ebde8430125259ca2f1227b96ccb8cef332aacf0109c74c40d01b22f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 453, + 482, + 464 + ], + "lines": [ + { + "bbox": [ + 110, + 453, + 482, + 464 + ], + "spans": [ + { + "bbox": [ + 110, + 453, + 482, + 464 + ], + "type": "text", + "content": "Figure 5: Visualization results of object decoder with three variants of the temporal decoder." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 236, + 288, + 356, + 444 + ], + "blocks": [ + { + "bbox": [ + 236, + 288, + 356, + 444 + ], + "lines": [ + { + "bbox": [ + 236, + 288, + 356, + 444 + ], + "spans": [ + { + "bbox": [ + 236, + 288, + 356, + 444 + ], + "type": "image", + "image_path": "6514fbf7550fecb62afd9a9c70ade661fff1fce9e8d66c5e00628f6ca649dfa3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 375, + 288, + 496, + 444 + ], + "blocks": [ + { + "bbox": [ + 375, + 288, + 496, + 444 + ], + "lines": [ + { + "bbox": [ + 375, + 288, + 496, + 444 + ], + "spans": [ + { + "bbox": [ + 375, + 288, + 496, + 444 + ], + "type": "image", + "image_path": "eb087d68a1ee64af9cdd70bb09cc8c3d1c55fde84ad833f50d05614c62822541.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 468, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 541 + ], + "type": "text", + "content": "When the gradients of historical BEV features are not detached in the last two layers of BEV encoder, the training time is increased by " + }, + { + "bbox": [ + 46, + 468, + 287, + 541 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 468, + 287, + 541 + ], + "type": "text", + "content": ". As for BEVDet4D-Depth, the BEV feature resolution and dimension is smaller than BEV-Former. Therefore, the additional training costs brought by HoP are negligible (from 4.5 seconds to 4.7 seconds)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "text", + "content": "Historical query collection form. We explore different methods to obtain historical object queries " + }, + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{t - k}^{his}" + }, + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "text", + "content": " in Table 6f. The Recurrent method only utilizes historical queries from the last frame, which is described in Equation 6; the Fully-Connected method collects all the historical object queries " + }, + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\mathbf{O}}_{t - N},\\dots,\\bar{\\mathbf{O}}_{t - 1}\\}" + }, + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "text", + "content": " but only uses them for the current frame " + }, + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "text", + "content": "; the Dense method extends the Fully-Connected method with using all historical queries on every timestamp. These three variants can bring consistent improvement, proving the insensitivity of the collection form; the recurrent form obtains the best result, where mAP increased by " + }, + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "text", + "content": " and NDS increased by " + }, + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 46, + 545, + 289, + 714 + ], + "type": "text", + "content": ". The results also prove our hypothesis that historical queries from the previous frame provide a more reliable initialization." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 468, + 389, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 389, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 389, + 480 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 492, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 545, + 577 + ], + "type": "text", + "content": "Figure 4 shows the detection result of the baseline method BEVFormer-opt and our proposed HoP-BEVFormer. The predictions are marked in yellow, while the ground truth boxes are marked in blue in the LiDAR_TOP figure. We observe that HoP-BEVFormer successfully detects small objects or occluded objects thanks to the superior ability to aggregate temporal information." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": "We also visualize the predictions of the object decoder with three variants of temporal decoder " + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 582, + 546, + 714 + ], + "type": "text", + "content": ": only long-term, only short-term, and both of them in Figure 5. The decoder with only a short-term branch can detect accurate attributes of objects, e.g., height and length, but fails to localize them precisely. Besides, results in the second row show that the short-term branch contributes to detecting more foreground objects because of its detailed spatial semantics. On the contrary, the decoder with only long-term modeling performs precise localization due to the long-term motion while struggling to detect the size of the object, e.g., height." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3788" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": "When both long-term and short-term branches are adopted, we achieve the best result by combining their advantages." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 107, + 119, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 107, + 119, + 119 + ], + "spans": [ + { + "bbox": [ + 47, + 107, + 119, + 119 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "text", + "content": "In this paper, we propose a new paradigm, named Historical Object Prediction (HoP) for multi-view 3D detection to leverage temporal information more effectively. The HoP approach is straightforward: given the current timestamp " + }, + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "text", + "content": ", we generate a pseudo BEV feature of timestamp " + }, + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "text", + "content": " from its adjacent frames and utilize this feature to predict the object set at timestamp " + }, + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "text", + "content": ". First, we elaborately design short-term and long-term temporal decoders, which can generate the pseudo BEV feature for timestamp " + }, + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "text", + "content": " without the involvement of its corresponding camera images. Second, an additional object decoder is flexibly attached to predict the object targets using the generated pseudo BEV feature. As a plug-and-play approach, HoP can be easily incorporated into state-of-the-art BEV detection frameworks, including BEVFormer and BEVDet series. Extensive experiments are conducted to evaluate the effectiveness of the proposed HoP on the nuScenes dataset. Surprisingly, HoP achieves " + }, + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "inline_equation", + "content": "68.5\\%" + }, + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "text", + "content": " NDS and " + }, + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 47, + 128, + 289, + 368 + ], + "type": "text", + "content": " mAP on nuScenes test, outperforming all the 3D object detectors on the leaderboard by a significant margin." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 378, + 145, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 145, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 145, + 392 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 398, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 398, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 398, + 287, + 471 + ], + "type": "text", + "content": "This project is funded in part by National Key RD Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 481, + 106, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 106, + 493 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 106, + 493 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 501, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 53, + 501, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 501, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 501, + 288, + 567 + ], + "type": "text", + "content": "[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 568, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 568, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 53, + 568, + 287, + 623 + ], + "type": "text", + "content": "[2] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Graph-detr3d: rethinking overlapping regions for multi-view 3d object detection. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5999-6008, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 287, + 668 + ], + "type": "text", + "content": "[3] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 287, + 713 + ], + "type": "text", + "content": "[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 547, + 150 + ], + "type": "text", + "content": "[5] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, June 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 152, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 547, + 196 + ], + "type": "text", + "content": "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 198, + 547, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 198, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 547, + 229 + ], + "type": "text", + "content": "[7] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 232, + 545, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 232, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 312, + 232, + 545, + 264 + ], + "type": "text", + "content": "[8] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 266, + 545, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 266, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 312, + 266, + 545, + 309 + ], + "type": "text", + "content": "[9] Yanqin Jiang, Li Zhang, Zhenwei Miao, Xiatian Zhu, Jin Gao, Weiming Hu, and Yu-Gang Jiang. Polarformer: Multicamera 3d object detection with polar transformers. arXiv preprint arXiv:2206.15398, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 311, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 311, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 311, + 545, + 365 + ], + "type": "text", + "content": "[10] Youngwan Lee, Joong-won Hwang, Sangrok Lee, Yuseok Bae, and Jongyoul Park. An energy andgpu-computation efficient backbone network for real-time object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0-0, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 366, + 547, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 547, + 410 + ], + "type": "text", + "content": "[11] Youngwan Lee and Jongyoul Park. Centermask: Realtime anchor-free instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13906-13915, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "type": "text", + "content": "[12] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with dynamic temporal stereo. arXiv preprint arXiv:2209.10248, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "type": "text", + "content": "[13] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. arXiv preprint arXiv:2206.00630, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 501, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 545 + ], + "type": "text", + "content": "[14] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 547, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 547, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 547, + 545, + 623 + ], + "type": "text", + "content": "[15] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part IX, pages 1-18. Springer, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 624, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 690 + ], + "type": "text", + "content": "[16] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[17] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detec" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3789" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "tion with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[18] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 531-548. Springer, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "type": "text", + "content": "[19] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petrv2: A unified framework for 3d perception from multi-camera images. arXiv preprint arXiv:2206.01256, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 249 + ], + "type": "text", + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 251, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 287, + 303 + ], + "type": "text", + "content": "[21] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11976-11986, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 360 + ], + "type": "text", + "content": "[22] Jiachen Lu, Zheyuan Zhou, Xiatian Zhu, Hang Xu, and Li Zhang. Learning ego 3d representation as ray tracing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVI, pages 129-144. Springer, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 414 + ], + "type": "text", + "content": "[23] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3142-3152, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "text", + "content": "[24] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multi-view 3d object detection. arXiv preprint arXiv:2210.02443, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[25] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 515, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 569 + ], + "type": "text", + "content": "[26] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 613 + ], + "type": "text", + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "text", + "content": "[28] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 913-922, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[29] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 508 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[30] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, pages 180–191. PMLR, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "text", + "content": "[31] Zengran Wang, Chen Min, Zheng Ge, Yinhao Li, Zeming Li, Hongyu Yang, and Di Huang. Sts: Surround-view temporal stereo for multi-view 3d detection. arXiv preprint arXiv:2208.10145, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "text", + "content": "[32] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. arXiv preprint arXiv:2211.10439, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 273 + ], + "type": "text", + "content": "[33] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 274, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 328 + ], + "type": "text", + "content": "[34] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 331, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 384 + ], + "type": "text", + "content": "[35] Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, and Stan Z Li. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9759-9768, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 386, + 545, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 386, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 307, + 386, + 545, + 429 + ], + "type": "text", + "content": "[36] Benjin Zhu, Zhengkai Jiang, Xiangxin Zhou, Zeming Li, and Gang Yu. Class-balanced grouping and sampling for point cloud 3d object detection. arXiv preprint arXiv:1908.09492, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 430, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 545, + 473 + ], + "type": "text", + "content": "[37] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 475, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 508 + ], + "type": "text", + "content": "[38] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. arXiv preprint arXiv:2211.12860, 2022." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3790" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/6f1f26ae-7b19-4337-978a-c5b559913901_content_list.json b/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/6f1f26ae-7b19-4337-978a-c5b559913901_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..656543f9d6851d7117368cff9c01c4dc6942ae73 --- /dev/null +++ b/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/6f1f26ae-7b19-4337-978a-c5b559913901_content_list.json @@ -0,0 +1,1693 @@ +[ + { + "type": "text", + "text": "Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold: Learning with Event-Driven Backpropagation", + "text_level": 1, + "bbox": [ + 102, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenjie Wei $^{1}$ , Malu Zhang $^{1*}$ , Hong Qu $^{1}$ , Ammar Belatreche $^{2}$ , Jian Zhang $^{3}$ , and Hong Chen $^{3}$", + "bbox": [ + 122, + 214, + 844, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Electronic Science and Technology of China, 2Northumbria University, 3Tsinghua University", + "bbox": [ + 246, + 244, + 720, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{maluzhang,hongqu}@uestc.edu.cn, ammar.belatreche@northumbria.ac.uk, hongchen@tsinghua.edu.cn", + "bbox": [ + 83, + 284, + 883, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 344, + 313, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spiking Neural Networks (SNNs) offer a highly promising computing paradigm due to their biological plausibility, exceptional spatiotemporal information processing capability and low power consumption. As a temporal encoding scheme for SNNs, Time-To-First-Spike (TTFS) encodes information using the timing of a single spike, which allows spiking neurons to transmit information through sparse spike trains and results in lower power consumption and higher computational efficiency compared to traditional rate-based encoding counterparts. However, despite the advantages of the TTFS encoding scheme, the effective and efficient training of TTFS-based deep SNNs remains a significant and open research problem. In this work, we first examine the factors underlying the limitations of applying existing TTFS-based learning algorithms to deep SNNs. Specifically, we investigate issues related to over-sparsity of spikes and the complexity of finding the 'causal set'. We then propose a simple yet efficient dynamic firing threshold (DFT) mechanism for spiking neurons to address these issues. Building upon the proposed DFT mechanism, we further introduce a novel direct training algorithm for TTFS-based deep SNNs, called DTA-TTFS. This method utilizes event-driven processing and spike timing to enable efficient learning of deep SNNs. The proposed training method was validated on the image classification task and experimental results clearly demonstrate that our proposed method achieves state-of-the-art accuracy in comparison to existing TTFS-based learning algorithms, while maintaining high levels of sparsity and energy efficiency on neuromorphic inference accelerator.", + "bbox": [ + 73, + 382, + 472, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 344, + 630, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning approaches have demonstrated remarkable performance in various intelligent applications [18, 27, 24, 23] and innovative simulations [62, 61]. However, these achievements come at a significant cost of energy consumption, severely limiting the deployment of artificial neural networks (ANNs) on resource-limited platforms [43, 9, 3, 50]. Brain-inspired spiking neural networks (SNNs) have emerged as a promising and energy-efficient alternative to ANNs [22, 31, 57]. SNNs have garnered attention for their distinctive features of sparse-asynchronous spikes and event-driven computing, which have played a pivotal role in driving the development of neuromorphic computing platforms such as ANP-I [54], TrueNorth [1], Loihi [8], and Tianjic [39].", + "bbox": [ + 496, + 369, + 890, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the inherent energy efficiency of SNNs, their extensive application to real-world practical scenarios has been limited by the absence of efficient and scalable training algorithms [52, 35, 30, 14, 20]. Traditional training techniques for ANNs, such as error backpropagation (BP) and standard GPU-accelerated training packages, cannot be leveraged directly by SNNs due to the complex temporal dynamics and the non-differentiability of spiking neurons' activity [33, 6, 56]. To address these challenges, various solutions have been proposed, broadly categorized into three distinct groups.", + "bbox": [ + 496, + 582, + 892, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The first category of solutions proposed to address the lack of scalable training algorithms for SNNs is ANN-to-SNN conversion method [40, 41, 49, 45]. These methods involve training an ANN first and then converting it to its SNN counterpart. The resulting SNN may achieve comparable performance to pre-trained ANNs through manual and careful setting of the SNN parameters such as the firing threshold [28] and the chosen spiking neuron model [44]. However, such conversion often results in precision loss due to approximation and requires longer time steps for lossless", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10552", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "conversion [19]. Moreover, ANN-to-SNN conversion methods often encode the activation values of artificial neurons using the firing rate of spiking neurons [11, 4, 5], which fails to fully leverage the timing information of spikes. Furthermore, the runtime computation on neuromorphic hardware can often be highly energy-intensive.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The second category of the proposed SNN training solutions is RNN-like method, commonly known as surrogate gradient learning [47, 48]. In these surrogate gradient-based SNNs, information is conveyed through spikes during feedforward computation, while during backpropagation, gradient information is calculated by treating the neurons' membrane potential as a differentiable signal at the current time step [42, 34]. To address the non-differentiability of spikes in the direct training of SNNs, surrogate functions are employed to approximate the derivative of the spike generation function during the backward process [53, 29]. However, while these methods have successfully tackled the challenges associated with the non-differentiability of spiking neurons, they necessitate updating synaptic weights at each time step, regardless of whether there is a spike emission or not [47, 15]. Consequently, this results in substantial energy and memory demands as large intermediate activation values necessitate storage to facilitate gradient computation.", + "bbox": [ + 75, + 183, + 470, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike the above-mentioned methods, the third category of learning algorithms in SNNs is spike-driven method [2, 33, 58, 60, 38, 6], which operate in a strictly event-driven manner during both forward and backward computations. SpikeProp [2] and its various extensions [33, 6, 25] are typical examples of this category. However, these algorithms remain limited to shallow network structures and relatively simple classification tasks. To overcome this limitation, Zhang et al. [58] and Zhou et al. [60] extended the spikedriven algorithm to more complex network structures. Despite achieving competitive accuracy on large datasets [60], the identification of the 'causal set' in this approach requires substantial computational resources. To improve the training efficiency, Park et al. [38] proposed a surrogate deep neural network (DNN) model that is subsequently converted into a TTFS-based SNN. However, this conversion method is prone to conversion errors, potentially impairing the temporal learning capability of SNNs. Recently, several multi-spike-based spike-driven algorithms have been developed [59, 65]. While Zhang et al. [59] and Zhu et al. [65] achieve competitive performance on CIFAR utilizing the rate coding scheme, we focus on developing an energy-efficient TTFS-based spike-driven algorithm.", + "bbox": [ + 75, + 455, + 470, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce an effective and efficient spike-driven learning algorithm that facilitates the training", + "bbox": [ + 76, + 805, + 468, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of high-performance TTFS-based deep SNNs. The main contributions of this work are outlined as follows:", + "bbox": [ + 500, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We provide a comprehensive analysis of the main shortcomings of existing methods in achieving high performance in TTFS-based deep SNNs. Specifically, we examine issues such as the over-sparsity of spikes and the complexity of determining the 'causal set'.", + "- We propose a simple yet efficient dynamic firing threshold (DFT) mechanism for spiking neurons that can effectively address the aforementioned issues.", + "- Building upon the proposed DFT, we further introduce a novel direct training algorithm for TTFS-based deep SNNs, which we refer to as DTA-TTFS. In this training method, the timing of a single spike is considered the basic information carrier and the learning process is performed strictly in an event-driven manner.", + "- The proposed method is validated through extensive experiments on benchmark image classification tasks and achieves state-of-the-art accuracy compared to existing TTFS-based learning algorithms. Furthermore, we demonstrate the ultra-low power capability of the SNN with DTA-TTFS on a recently developed neuromorphic accelerator." + ], + "bbox": [ + 517, + 132, + 890, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries and Problem Analysis", + "text_level": 1, + "bbox": [ + 500, + 497, + 825, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first briefly review the TTFS coding and the employed spiking neuron model. Subsequently, we provide a comprehensive analysis of the primary shortcomings in existing TTFS-based deep SNN methods.", + "bbox": [ + 498, + 523, + 890, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 593, + 638, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "TTFS coding scheme: There are two primary coding schemes utilized to represent information in SNNs: rate coding and temporal coding. The rate coding method represents information based on the average firing rate of a spiking neuron. However, this encoding scheme neglects the temporal information carried by spikes, leading to higher spike frequency and increased energy consumption. Conversely, temporal coding leverages the precise timing of spikes to encode information, resulting in lower power consumption and higher computational efficiency compared to the rate-based coding method.", + "bbox": [ + 496, + 617, + 890, + 782 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a temporal coding method, TTFS encodes information by the timing of the first spike. As depicted in Fig. 1, the higher the activation value, the earlier the first spike is emitted. The relationship between the input activation value $a_{i}$ and the encoded spike time $t_i$ can be expressed as", + "bbox": [ + 496, + 784, + 890, + 859 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nt _ {i} = \\left(1 - \\frac {a _ {i}}{a _ {\\text {m a x}}}\\right) \\times T _ {w}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 869, + 890, + 904 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Assume a postsynaptic neuron fire a first spike at $t_{out}$ , only a subset of input spikes had arrived before $t_{out}$ and contributed to the output spike. This set of spikes, $C = \\{i : t_i < t_{out}\\}$ , is called the causal set in this work.", + "bbox": [ + 75, + 849, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10553", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2c6d616ed5816f58b20b1c4c29eb8d0000ecdf289ee7efde9b362c28d5c5ef92.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 297, + 193 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3c7b1045b9577e1c511f3fe5260a99ef67a796ad09755afd0d1ba86d2ca99205.jpg", + "image_caption": [ + "(b)", + "Figure 1. (a) TTFS coding scheme. The larger intensity value corresponds to the earlier spike. (b) ReL-PSP spiking neuron model. Neuron $j$ accumulates the linear PSP and fires a spike when the membrane potential crosses the firing threshold $\\vartheta$ ." + ], + "image_footnote": [], + "bbox": [ + 300, + 88, + 454, + 190 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $a_{max}$ is the maximum input activation value, $T_w$ is the length of the permissible spike time window. As the TTFS coding scheme requires only a single spike to convey information, it effectively diminishes the number of spikes and significantly decreases the energy requirements. Consequently, it achieves higher sparsity and energy efficiency, making it a powerful coding scheme that maximizes the energy efficiency of SNNs.", + "bbox": [ + 75, + 296, + 468, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spiking neuron model: The Rectified Linear Postsynaptic Potential (ReL-PSP) spiking neuron [58] is chosen for our method due to its simplicity and energy efficiency. As shown in Fig. 1(b), the ReL-PSP neuron model utilizes a rectified linear kernel function to shape the neuron's PSP, and its dynamics is defined as", + "bbox": [ + 75, + 417, + 468, + 507 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV _ {j} ^ {l} (t) = \\sum_ {i = 1} ^ {N} w _ {i j} ^ {l} K \\left(t - t _ {i} ^ {l - 1}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 518, + 468, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $V_{j}^{l}(t)$ is the membrane potential of neuron $j$ in layer $l$ , and $w_{ij}^{l}$ is the synaptic weight between neuron $j$ and its presynaptic neuron $i$ . $N$ is the number of neurons in the previous layer $l - 1$ , $t_{i}^{l - 1}$ is the spike time of input neuron $i$ , and $K\\left(t - t_{i}^{l - 1}\\right)$ is the PSP function that is defined as", + "bbox": [ + 76, + 569, + 468, + 650 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nK \\left(t - t _ {i} ^ {l - 1}\\right) = \\left\\{ \\begin{array}{l l} t - t _ {i} ^ {l - 1}, & \\text {i f} \\quad t > t _ {i} ^ {l - 1}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 659, + 468, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The neuron $j$ emits a spike when $V_{j}^{l}(t)$ reaches the firing threshold $\\vartheta$ . Mathematically, the spike generation function $F$ is defined as", + "bbox": [ + 76, + 710, + 468, + 756 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nt _ {j} ^ {l} = F \\{t | V _ {j} ^ {l} (t) \\geq \\vartheta , t \\geq 0 \\}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 765, + 468, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The ReL-PSP neuron has proven its efficacy in tackling the main challenges encountered during the training of deep SNNs, such as the non-differentiability of the spike function, gradient explosion, and the dead neuron issue. Despite the successful performance on the MNIST dataset, its performance is not ideal on deeper network structures and more complex datasets, such as CIFAR-10 and CIFAR-100.", + "bbox": [ + 75, + 794, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the subsequent section, we will conduct a comprehensive analysis to identify the underlying reasons for this limitation.", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Problem Analysis", + "text_level": 1, + "bbox": [ + 500, + 145, + 669, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Over-sparsity of spikes: Sparse neural networks offer several advantages such as efficient variable-size representation, information decomposition, and energy efficiency. However, over-sparsity may diminish the model's effective capacity and adversely affect its performance [17]. This issue becomes more serious in TTFS-based deep SNNs, owing to their firing mechanism, asynchronous transmission, and TTFS-based decision strategy.", + "bbox": [ + 496, + 167, + 890, + 289 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/64ae6d96b091219cd27021032a4b4cf4a6b1e7cbb361c424d6de03cab5378082.jpg", + "image_caption": [ + "(a)", + "Figure 2. The over-sparsity problem caused by asynchronous transmission and TTFS-based decision strategy. (a) The presynaptic (gray) neuron fires later than the postsynaptic (blue) neuron $j$ , making it ineffective in contributing to the firing of the postsynaptic neuron. As a result, this presynaptic neuron becomes invalid. (b) After learning, the postsynaptic (blue) neuron fires at an earlier time, leading to the appearance of additional invalid neurons." + ], + "image_footnote": [], + "bbox": [ + 535, + 301, + 687, + 407 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ca111c35d49f8fac874315a4faf27d2eab53d643ab7c43b09565d777bf2422fc.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 714, + 303, + 857, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to the firing mechanism of spiking neurons, a neuron emits one spike only when its membrane potential reaches the firing threshold. This leads to fewer activated spiking neurons compared to ANNs. In addition, the asynchronous transmission of SNNs further diminishes the number of valid activated neurons. As shown in Fig. 2(a), when the presynaptic neuron $i$ fires after the postsynaptic neuron $j$ , the spike of neuron $i$ becomes invalid for neuron $j$ . Furthermore, TTFS-based SNNs make decisions based on the earliest spike, and the synaptic weights are trained to make the target neuron fire as early as possible. As shown in Fig. 2(b), after learning, the firing time of the postsynaptic neuron $j$ may become earlier, thereby further exacerbating the sparsity induced by asynchronous transmission.", + "bbox": [ + 496, + 553, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In event-driven learning algorithms of SNNs, error backpropagation depends on the timing of spikes. Therefore, if a spiking neuron does not fire or does not contribute to the postsynaptic neuron, it will not participate in learning. The issue of over-sparsity not only limits the information representation capability of SNNs but also leads to learning failure. Consequently, the achievement of a harmonious balance between sparsity and performance in the training of TTFS-based deep SNNs remains an open question.", + "bbox": [ + 496, + 765, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10554", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The complexity of finding the 'causal set': Due to the sparsity and asynchronicity of SNNs, not all presynaptic neurons $\\{i|i = 1,\\dots ,N\\}$ contribute to their postsynaptic neuron $j$ , where the contributory neurons are defined as the 'causal set' of the postsynaptic neuron $j$ , denoted as $C_j = \\{i|t_i < t_j\\}$ [33, 60]. To leverage standard GPU-accelerated ANN training packages, it is crucial to identify the 'causal set' of each neuron during the feed-forward computation process [33, 60]. However, this process is highly complex and time-consuming. For example, in order to find the 'causal set', Zhou et al. [60] proposed an approach where they first sort all of this neuron input spikes $\\{t_i|i = 1,\\dots N\\}$ in ascending order and subsequently consider each individual spike $t_i$ for inclusion in the causal set. This approach, however, is computationally expensive and results in inefficient training even with the implementation of a space-for-time strategy [60].", + "bbox": [ + 76, + 90, + 472, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite the success of algorithms with this approach in utilizing standard GPU-accelerated ANN training packages, the training process involves a significant amount of sorting and iteration. This results in higher demand for training resources and increased latency, especially as the network depth increases. Consequently, the process of finding the 'causal set' may become a bottleneck for training deep SNNs. In the next section, we propose a dynamic firing threshold mechanism for spiking neurons to avoid this resource-intensive process.", + "bbox": [ + 75, + 347, + 472, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 511, + 169, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first introduce the dynamic firing threshold (DFT) for spiking neurons and discuss how it effectively addresses the above issues. Then, based on the DFT, we further propose a direct training algorithm for TTFS-based SNN, referred to as DTA-TTFS.", + "bbox": [ + 75, + 537, + 468, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. DFT-based Spiking Neuron Model", + "text_level": 1, + "bbox": [ + 76, + 621, + 375, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As demonstrated in Fig. 3(a), the proposed DFT is a layer-dependent and time-varying function, which is defined as", + "bbox": [ + 75, + 643, + 468, + 688 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\vartheta^ {l} (t) = \\left\\{ \\begin{array}{l l} T _ {w} (l + 1) - t, & \\text {i f} \\quad T _ {w} l \\leq t \\leq T _ {w} (l + 1), \\\\ + \\infty , & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 696, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\vartheta^l (t)$ represents the firing threshold of the neurons in the $l$ -th layer and $T_{w}$ denotes the length of the permissible spike time window. In this work, $T_{w}$ is set to 1 (unless otherwise stated) and the input layer is regarded as the 0-th layer. Eq. 5 reveals that the firing threshold is infinite outside the interval $t\\in [T_wl,T_w(l + 1)]$ , and neurons in the $l$ -th layer can only generate a spike within this time interval. Therefore, as shown in Fig. 4, the generated spikes are non-overlapping across different layers, which we refer to as the layer-dependent firing feature.", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the membrane potential dynamics in Eq. 2 and the proposed DFT in Eq. 5, the output of neuron $j$ can be divided into three categories:", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Spikes at $t_j^l = T_{wl}$ : Since the firing threshold is infinity when $t < T_{wl}$ , the neuron cannot fire a spike before $t = T_{wl}$ even though the membrane potential is strong (i.e., $V_j^l(t) \\geq \\vartheta^l(T_{wl})$ and $t < T_{wl}$ ). As shown in Fig. 3(b), the earliest spike time for DFT-based spiking neurons in the $l$ -th layer is $t_j^l = T_{wl}$ .", + "- Spikes within $(T_{w}l, T_{w}(l + 1))$ : The neuron generates a spike when its membrane potential exceeds the threshold, i.e. $V_{j}^{l}(t) \\geq \\vartheta^{l}(t)$ . The firing threshold $\\vartheta^{l}$ in this time interval is a linear function, which decays linearly from 1 to 0. As depicted in Fig. 3(c), by combining the membrane potential in Eq. 2 and the firing threshold in Eq. 5, the precise spiking time $t_{j}^{l}$ can be calculated as" + ], + "bbox": [ + 517, + 147, + 890, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {j} ^ {l} = \\frac {T _ {w} (l + 1) + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l} t _ {i} ^ {l - 1}}{1 + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 371, + 890, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Nonspike: The neuron will not fire a spike if its membrane potential is below the dynamic firing threshold (i.e., $V_{j}^{l}(t) < \\vartheta^{l}(t)$ ). Fig. 3(d) shows three typical scenarios of nonspike.", + "bbox": [ + 517, + 448, + 890, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, the spike time of a DFT-based neuron $j$ in the $l$ -th layer can be represented as", + "bbox": [ + 496, + 520, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {j} ^ {l} = \\left\\{ \\begin{array}{l l} T _ {w} l, & \\text {i f} \\quad V _ {j} ^ {l} \\left(T _ {w} l\\right) \\geq \\vartheta^ {l} \\left(T _ {w} l\\right), \\\\ \\frac {T _ {w} (l + 1) + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l} t _ {i} ^ {l - 1}}{1 + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l}}, & \\text {i f} \\quad T _ {w} l < t _ {j} ^ {l} \\leq T _ {w} (l + 1), \\\\ \\text {n o n s p i k e}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 564, + 890, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following, we will analyze how the proposed DFT effectively addresses issues of over-sparsity of spikes and the complexity of finding the 'causal set'.", + "bbox": [ + 496, + 657, + 890, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1) Over-sparsity of spikes: As described in Eq. 5, the DFT is a piecewise function, whose linearly decreased part and layer-dependent firing feature effectively address the over-sparsity of spikes.", + "bbox": [ + 496, + 703, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The linearly decreased part of the DFT is purposefully designed to tackle the over-sparsity arising from the spiking neuron firing mechanism. By gradually decreasing the firing threshold from 1 to 0, the neuron $j$ remains active as long as its membrane potential is non-negative at $T_{w}(l + 1)$ . As a result, the proposed DFT effectively reduces the number of inactive neurons. Moreover, the linear decay design enhances the discriminative nature of information from active neurons.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10555", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fd574486f30506d416cadb62ae3654d1cc072c345b6c02a4414e5a44c49e9808.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 81, + 93, + 281, + 193 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/41bdde42027fe031e85074cac6e17050278ef41be38365c8fa51d1ad476e9245.jpg", + "image_caption": [ + "(b)", + "Figure 3. Proposed dynamic firing threshold and output of the DFT-based spiking neurons. (a) Dynamic firing threshold function(DFT). (b) Spikes at $t_j^l = T_w l$ . (c) Spikes within $(T_w l, T_w (l + 1)]$ . (d) Nonspike. This case is further divided into three sub-cases, which are shown by the curves of different colors." + ], + "image_footnote": [], + "bbox": [ + 292, + 92, + 485, + 193 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/91f457e988f999826c4da082ffc2f2be246ea95f46742ab2c645c0e015ae782b.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 495, + 90, + 686, + 193 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/133938a2ea83c75d7176e3ad1904a5a37ff3d5db422aa2dfa2d77e561d052ba8.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 699, + 88, + 887, + 193 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a954087f8b2fee71906adadcf04891013d7f9a7db6a9add4ef3c32991f1ebf05.jpg", + "image_caption": [ + "Figure 4. The propagation pipeline of the SNN with DFT." + ], + "image_footnote": [], + "bbox": [ + 94, + 276, + 450, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The layer-dependent firing feature enables SNN with non-overlapping spike time windows. This effectively resolves the over-sparsity arising from asynchronous transmission and TTFS-based decision strategy. Fig. 4 illustrates that the spike time window of postsynaptic neurons occurs later than that of presynaptic neurons. Consequently, when the presynaptic neuron fires a spike, its postsynaptic neuron absolutely receives it. Furthermore, the non-overlapping spike window provides output layer neurons with a predetermined earliest spike time, preventing them from being forced forward indefinitely during the training process.", + "bbox": [ + 75, + 430, + 467, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2) The complexity of finding the 'causal set': The layer-dependent firing feature of DFT circumvents the need for SNNs to identify the 'causal set', resulting in efficient training. This efficiency is further improved by the output analytic equation of the DFT-based neuron.", + "bbox": [ + 75, + 597, + 467, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As depicted in Fig. 4, the layer-dependent firing feature ensures that in the $l$ -th layer of the SNN utilizing DFT, the spike time is controlled within the interval $[T_w l, T_w (l + 1)]$ . Therefore, we can easily learn that neurons in the $(l - 1)$ -th layer must fire earlier than neurons in the $l$ -th layer. In other words, all active presynaptic neurons must remain in the 'causal set' of their postsynaptic neurons. Therefore, the DFT mechanism prevents the need for resource-intensive sorting and iteration process, enhancing training efficiency and facilitating its extension to deeper network architecture.", + "bbox": [ + 75, + 672, + 467, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, the employed ReL-PSP spiking neuron and the DFT provide an analytical relation between input spikes and output spikes, as described in Eq.7. Therefore, the dynamics of SNNs within the training loop need not be simulated using long time steps. This allows the utilization", + "bbox": [ + 75, + 824, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of standard GPU-accelerated training packages of ANNs, which further leads to an efficient training process.", + "bbox": [ + 498, + 280, + 888, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. DTA-TTFS Learning Algorithm", + "text_level": 1, + "bbox": [ + 500, + 327, + 782, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consider a task with $n$ categories, where each output neuron corresponds to a specific category. In this task, the target category is represented by the output neuron that fires the earliest. To assess the effectiveness of the model, we adopt the widely utilized cross-entropy loss function $L$ , which aims to maximize the output value of the target neuron. In order to ensure the desired neuron fires earliest and others fire as late as possible, we utilize the negative spike times of the output layer [58, 6], i.e., $-t$ , for evaluation. Therefore, the loss function is given by", + "bbox": [ + 496, + 353, + 890, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL (\\mathbf {t}, d) = - \\log \\frac {\\exp (- \\mathbf {t} [ d ])}{\\sum_ {i = 1} ^ {n} \\exp (- \\mathbf {t} [ i ])}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 518, + 890, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{t}$ is the vector of spike times in the output layer, and $d$ is the desired category index.", + "bbox": [ + 498, + 566, + 888, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to minimize the $L(\\mathbf{t},d)$ , the DTA-TTFS algorithm adjusts spike times by modifying the synaptic weights during gradient backpropagation. For driving the gradient backpropagation, it is necessary to calculate derivatives of the output spike time $t_j^l$ with respect to the synaptic weight $w_{ij}^{l}$ and the input spike time $t_i^{l - 1}$ . According to Eq. 7, we can easily get these two derivatives,", + "bbox": [ + 496, + 599, + 890, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial t _ {j} ^ {l}}{\\partial w _ {i j} ^ {l}} = \\left\\{ \\begin{array}{l l} \\frac {t _ {i} ^ {l - 1} - t _ {j} ^ {l}}{1 + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l}}, & \\text {i f} \\quad T _ {w} l < t _ {j} ^ {l} < T _ {w} (l + 1), \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 723, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial t _ {j} ^ {l}}{\\partial t _ {i} ^ {l - 1}} = \\left\\{ \\begin{array}{l l} \\frac {w _ {i j} ^ {l}}{1 + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l}}, & \\text {i f} \\quad T _ {w} l < t _ {j} ^ {l} < T _ {w} (l + 1), \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 801, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following Eq. 9 and Eq. 10, gradient backpropagation can be directly applied.", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10556", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/30e36d01e61ae234b1feb76572f7972f7da67c114fc25c4ee754fdfae2833b55.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsModelsNetwork ArchitectureNeural CodingMethodAccuracySparsity
MNISTMostafa 2017 [33]MLP¹TFTSDT97.55%0.51
Zhang et al. 2020 [58]CNN¹TFTSDT99.4%0.6614
Zhou et al. 2021 [60]CNN²TFTSDT99.33%0.94
DTA-TTFSCNN¹TFTSDT99.4%0.3913
CIFAR-10Wu et al. 2022 [46]VGG11Rateconv91.24%no
Park et al. 2020 [37]VGG16TFTSconv91.43%0.2459
Zhou et al. 2021 [60]VGG16TFTSDT92.68%0.62
Park et al. 2021 [38]VGG16TFTSDT91.90%0.1746
DTA-TTFSVGG11TFTSDT91.17%0.4387
DTA-TTFSVGG16TFTSDT93.05%0.2561
CIFAR-100Park et al. 2020 [37]VGG16TFTSconv68.79%0.2994
Park et al. 2021 [38]VGG16TFTSDT65.98%0.2780
DTA-TTFSVGG16TFTSDT69.66%0.2845
", + "bbox": [ + 155, + 88, + 816, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of accuracy and sparsity between the DTA-TTFS method and other related works. DT: direct training, conv: ANN-to-SNN conversion. MLP1: 784-800-10. CNN1: 784-16C5-P2-32C5-P2-800-128-10. CNN2: 784-32C5-16C5-10.", + "bbox": [ + 75, + 335, + 890, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 386, + 209, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first evaluate the performance of the DTA-TTFS algorithm. Then, we carry out validation studies to illustrate the impact of the DFT. Finally, we employ theoretical analysis along with the neuromorphic hardware simulator to evaluate the energy consumption during the inference process.", + "bbox": [ + 75, + 411, + 470, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Details", + "text_level": 1, + "bbox": [ + 76, + 508, + 166, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on three image classification benchmark datasets: MNIST $^2$ , CIFAR-10, and CIFAR $100^3$ . All experiments are conducted using the PyTorch library, which allows for efficient training on multi-GPU machines with accelerated computations and optimized memory usage. For the MNIST dataset, we employ a network structure with 784-16C5-P2-32C5-P2-800-128-10 and train it for 150 epochs. As for CIFAR datasets, we adopt augmentation techniques [7, 10] and train the VGG11 and VGG16 network structures without Batch Normalization (BN) for 300 epochs. Moreover, in order to maintain the initial parameter domain at a good level, the SNN is initialized with the parameters from the pre-trained identical ANN when training CIFAR datasets. During the training process, we adopt the Adam optimizer and implement milestones learning rate decay, with an initial learning rate $1e^{-3}$ for MNIST and $1e^{-4}$ for CIFAR datasets.", + "bbox": [ + 75, + 532, + 468, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Related Works", + "text_level": 1, + "bbox": [ + 76, + 797, + 366, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 presents a comparison of the accuracies and sparsity achieved by DTA-TTFS and other related works. Experimental results demonstrate that the DTA-TTFS algo", + "bbox": [ + 75, + 821, + 468, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rithm exhibits excellent performance, with an accuracy of $99.4\\%$ on MNIST, $93.05\\%$ on CIFAR-10, and $69.66\\%$ on CIFAR-100. Obviously, our method outperforms all previous direct learning algorithms of TTFS-based SNNs. Moreover, our method still maintains competitive sparsity when achieving high performance. The sparsity is calculated by $S / N$ averaged on the entire testing set, where $S$ represents the total number of spikes and $N$ denotes the total number of neurons. Specifically, the DTA-TTFS achieves sparsity of 0.3913 on MNIST, 0.2561 on CIFAR-10, and 0.2845 on CIFAR-100. For the MNIST dataset, when employing the identical network structure as described in [58], our method achieves the same accuracy while significantly reducing the sparsity from 0.6614 to 0.3913. For CIFAR datasets, our method surpasses related works in terms of accuracy while maintaining a competitive sparsity. In summary, the DTA-TTFS algorithm achieves a harmonious balance between accuracy and sparsity.", + "bbox": [ + 500, + 387, + 893, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Validation Study", + "text_level": 1, + "bbox": [ + 500, + 676, + 665, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct a series of validation experiments to confirm the functionality of the proposed DFT. These experiments aim to display two aspects. Firstly, the DFT provides SNNs with the layer-dependent firing feature. Additionally, it effectively addresses the over-sparsity of spikes. Validation experiments choose the well-trained VGG11 structure on the CIFAR-10 dataset.", + "bbox": [ + 496, + 700, + 890, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to show the layer-dependent firing feature of the SNN with DFT, we record the spike activity in each layer for the first sample of the testing set. For the purpose of comparison, we also provide the spike activity of the SNN without DFT. The results are visualized in Fig. 5, elucidating the evident regulation of spike activity in each layer", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2 http://yann.lecun.com/exdb/mnist/", + "bbox": [ + 94, + 875, + 285, + 887 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3https://www.cs.toronto.edu/ kriz/cifar.html", + "bbox": [ + 96, + 887, + 326, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10557", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b460326f15f624006660d7611722d9c9bd15fe4c2efe06e757da78131de887b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 85, + 454, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/765d6467c54f485876cd3f38e9fffb4844a792cbd0afc4092506dc4dd11c8723.jpg", + "image_caption": [ + "Figure 5. Spike activity in each layer of SNN with and without DFT. The vertical axis represents the neuron index scaled to [0,1]." + ], + "image_footnote": [], + "bbox": [ + 76, + 262, + 455, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "within a permitted time window for the SNN incorporating DFT. In contrast, the spike activity in the SNN without DFT exhibits a disorderly and chaotic pattern.", + "bbox": [ + 75, + 488, + 468, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to verify the efficacy of the DFT in addressing the over-sparsity of spikes, we count the number of inactive neurons per layer averaged on the whole testing set. This experiment is conducted on three models: the ANN, the SNN with DFT, and the SNN without DFT [58]. As shown in Fig. 6, the SNN without DFT displays significant sparsity, however, with only $10\\%$ accuracy. Conversely, the SNN with DFT exhibits a comparable amount of inactive neurons to that of ANN. As a result, the DFT achieves a harmonious trade-off between sparse representation and information representation ability.", + "bbox": [ + 75, + 534, + 470, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Energy Estimation", + "text_level": 1, + "bbox": [ + 76, + 713, + 256, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1 Theoretical Analysis", + "text_level": 1, + "bbox": [ + 76, + 738, + 274, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Energy consumption is a critical metric to evaluate the efficiency of SNNs. To prove the energy efficiency of the proposed DTA-TTFS, we employ the theoretical analysis presented in [37]. This method analyzes energy consumption based on multiple factors, including latency, spike counts, as well as data related to neuromorphic architecture. More specifically, the estimated energy is defined as", + "bbox": [ + 75, + 763, + 468, + 869 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {T o t a l}} = E _ {\\text {S t a t i c}} \\times \\text {L a t e n c y} + E _ {\\text {D y n a m i c}} \\times \\text {S p i k e s}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 885, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ff63c4767d4b2ffdcdbefa9152c1788863df93f74c0ee37c4aa662e05619d35.jpg", + "image_caption": [ + "Figure 6. Percentage of inactive neurons per layer of the ANN, the SNN with DFT, and the SNN without DFT." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 890, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where Latency is the required time steps, Spikes is the total number of spikes, $E_{Static}$ and $E_{Dynamic}$ are static and dynamic energy coefficients depending on the neuromorphic architecture. We choose two neuromorphic architectures, TrueNorth [32] and SpiNNaker [16], to conduct analysis. The energy parameters $(E_{Static}, E_{Dynamic})$ are set to (0.6, 0.4) for TrueNorth and (0.36, 0.64) for SpiNNaker. For comparison, we analyze various coding schemes, such as rate coding [19], phase coding [26], burst coding [36], and TTFS coding [37, 38]. Theoretical analysis is performed on CIFAR-10 with the well-trained VGG16 structure, and we normalized estimated energy consumption based on the rate coding. As indicated in Tab. 2, under the same conditions, the DTA-TTFS outperforms other methods in terms of energy efficiency. It is worth noting that, despite the fewer spikes achieved by [37, 38], their energy consumption remains higher than ours due to their demand for long latency.", + "bbox": [ + 496, + 316, + 892, + 574 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e5ec04af19f1b0e6b8c42b33f3fbe53269532b134b67ee067364029f214060bc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Neural CodingTime StepSpike (106)Acc. (%)Normalized Energy
TrueNorthSpiNNaker
Rate [19]5122.61293.3911
Phase [26]150035.19691.217.14769.6785
Burst [36]11256.9291.412.37812.4865
TTFS [37]6800.06991.430.80740.4950
TTFS [38]5440.06791.90.64780.3989
DTA-TTFS1600.07393.050.19870.1304
", + "bbox": [ + 503, + 585, + 890, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Theoretical analysis of energy consumption.", + "bbox": [ + 537, + 718, + 854, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2 Hardware Simulation", + "text_level": 1, + "bbox": [ + 500, + 768, + 705, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to further verify the energy efficiency and hardware friendliness of our method, we map the SNN trained by the DTA-TTFS algorithm on a recently developed neuromorphic hardware simulator, called configurable asynchronous neuromorphic hardware simulator (CanMore) [55]. CanMore estimates hardware performance through system-level modeling and simulation. Fig. 7 illustrates the main", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10558", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5274cbf7e6d8854be059704150bc939ca36f94baf27acc47f8c38e3507dca2c4.jpg", + "image_caption": [ + "Figure 7. The framework of CanMore [55], including neuromorphic cores comprised of neurons and SRAM, and NoC routers consisting of buffers, route computation units, switch allocator, and crossbar switch." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 472, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "framework of the CanMore, including neuromorphic cores comprised of neurons and SRAM, and network-on-chip (NoC) routers consisting of buffers, route computation unit, switch allocator and crossbar switch. With the system-level asynchronous circuit modeling, CanMore is able to simulate various neuromorphic hardware architectures by changing the parameters and configuration files, estimate latency and energy consumption, and analyze spike messages routing information such as routing hops, congestion latency, routing directions, and so on.", + "bbox": [ + 75, + 332, + 468, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In [13], a deep ANN is mapped to TrueNorth neuromorphic hardware. To the best of our knowledge, this is the only work that verifies the energy efficiency of SNNs on the CIFAR-10 dataset, while the majority of existing works focus on smaller datasets, such as MNIST [12, 58]. Here, we train an SNN on the CIFAR-10 dataset with the proposed DTA-TTFS method. After learning, the learned weights are transferred to CanMore for accelerating the inference operations. Specifically, our network structure is mapped by partitioning each layer into equally sized groups along the feature dimension, with each group being mapped to a neuromorphic core with a maximum of 1024 neurons. In addition, we compress the network by quantizing the weights to 4 bits. The mesh topology and XY routing algorithm are adopted in the simulated architecture. The energy consumption per frame during the inference process is calculated by", + "bbox": [ + 75, + 484, + 468, + 726 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nE _ {T o t a l} = E _ {S O P} \\times \\text {S p i k e s} + E _ {H o p} \\times \\text {H o p s} + E _ {\\text {S t a t i c}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 737, + 468, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $E_{SOP}$ is the dynamic energy consumption per synaptic operation (SOP), which is calculated with the real technology parameters of the CMOS 22nm process in the simulation. In our work, $E_{SOP}$ is $20\\mathrm{pJ}$ . $E_{Hop}$ is the dynamic energy consumption per hop between the mesh routers on the NoC, which is $100\\mathrm{pJ}$ according to [21]. $E_{Static}$ is static energy consumption, which is $3.52\\mu \\mathrm{J}/\\mathrm{frame}$ in CMOS 22nm process. The average Spikes and Hops per frame during the inference process are 71M and 45.4M re", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/30ea7d1644ab9eca5b950198984c2e5bbe5fab515b084f649c5cfa991da6616a.jpg", + "image_caption": [ + "Figure 8. The accuracy and energy consumption of Esser et al. [13] and DTA-TTFS on CIFAR-10 dataset." + ], + "image_footnote": [], + "bbox": [ + 506, + 85, + 686, + 267 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e88121b7bffd7f5e17f2a83957cdeecfd52034d77883c573bf47f33d6df05453.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 88, + 883, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "spectively, given by CanMore. Therefore, we obtain the energy consumption $E_{Total}$ during the inference process $50.36\\mu \\mathrm{J / frame}$ .", + "bbox": [ + 496, + 321, + 890, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The accuracy and energy consumption are reported in Fig. 8. our work achieves a classification accuracy of $90.67\\%$ on CIFAR-10, surpassing the accuracy reported in [13] by $7.26\\%$ . In addition, the energy consumption of our work is $50.36~\\mu \\mathrm{J} / \\mathrm{frame}$ , which is $69.2\\%$ reduced compared with the energy consumption $163.65\\mu \\mathrm{J} / \\mathrm{frame}$ in [13]. These results confirm that the proposed DTA-TTFS algorithm is effective and energy-efficiency.", + "bbox": [ + 496, + 366, + 892, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 502, + 617, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we analyze the shortcomings of existing methods in achieving high-performance TTFS-based deep SNNs. In order to overcome these issues, we propose an efficient DFT mechanism for spiking neurons and introduce the DTA-TTFS algorithm for TTFS-based SNNs. We evaluate our method on the image classification task, and the results demonstrate that our method achieves a fine trade-off between accuracy and sparsity. Additionally, we estimate the energy consumption of our method through theoretical analysis and the neuromorphic hardware accelerator CanMore. Both theoretical analysis and hardware verification confirm the energy efficiency of the DTA-TTFS algorithm. Overall, our method not only attains high performance but also is quite hardware friendly. In our future work, we intend to leverage DTA-TTFS to train more advanced network models, including SNN-based Attention [64, 51] and Transformer architectures [63].", + "bbox": [ + 496, + 527, + 892, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 799, + 651, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Science Foundation of China under Grants 62236007 and 61976043 and 62106038, and in part by the Sichuan Science and Technology Program under Grants 2022YFG0313 and 2023YFG0259.", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10559", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34(10):1537-1557, 2015.", + "[2] Sander M Bohte, Joost N Kok, and Johannes A La Poutre. Spikeprop: backpropagation for networks of spiking neurons. In *ESANN*, volume 48, pages 419–424. Bruges, 2000.", + "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020.", + "[4] Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Optimized potential initialization for low-latency spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11-20, 2022.", + "[5] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347, 2023.", + "[6] Iulia-Maria Comsa, Krzysztof Potempa, Luca Versari, Thomas Fischbacher, Andrea Gesmundo, and Jyrki Alakuijala. Temporal coding in spiking neural networks with alpha synaptic function: learning with backpropagation. IEEE transactions on neural networks and learning systems, 33(10):5939-5952, 2021.", + "[7] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019.", + "[8] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 38(1):82-99, 2018.", + "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.", + "[10] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017.", + "[11] Peter U Diehl, Daniel Neil, Jonathan Binas, Matthew Cook, Shih-Chii Liu, and Michael Pfeiffer. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2015.", + "[12] Steve K Esser, Rathinakumar Appuswamy, Paul Merolla, John V Arthur, and Dharmendra S Modha. Backpropagation for energy-efficient neuromorphic computing. Advances in neural information processing systems, 28, 2015." + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Steven K Esser, Paul A Merolla, John V Arthur, Andrew S Cassidy, Rathinakumar Appuswamy, Alexander Andreopoulos, David J Berg, Jeffrey L McKinstry, Timothy Melano, Davis R Barch, et al. From the cover: Convolutional networks for fast, energy-efficient neuromorphic computing. Proceedings of the National Academy of Sciences of the United States of America, 113(41):11441, 2016.", + "[14] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021.", + "[15] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothée Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2661-2671, 2021.", + "[16] Steve B Furber, Francesco Galluppi, Steve Temple, and Luis A Plana. The spinnaker project. Proceedings of the IEEE, 102(5):652-665, 2014.", + "[17] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 315-323. JMLR Workshop and Conference Proceedings, 2011.", + "[18] Alex Graves and Navdeep Jaitly. Towards end-to-end speech recognition with recurrent neural networks. In International conference on machine learning, pages 1764-1772. PMLR, 2014.", + "[19] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13558–13567, 2020.", + "[20] Yangfan Hu, Huajin Tang, and Gang Pan. Spiking deep residual networks. IEEE Transactions on Neural Networks and Learning Systems, 2021.", + "[21] Masashi Imai, Thiem Van Chu, Kenji Kise, and Tomohiro Yoneda. The synchronous vs. asynchronous noc routers: an apple-to/apple comparison between synchronous and transition signaling asynchronous designs. In 2016 Tenth IEEE/ACM International Symposium on Networks-on-Chip (NOCS), pages 1-8. IEEE, 2016.", + "[22] Eugene M Izhikevich. Simple model of spiking neurons. IEEE Transactions on neural networks, 14(6):1569-1572, 2003.", + "[23] Yeying Jin, Aashish Sharma, and Robby T Tan. Dshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5027-5036, 2021.", + "[24] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In European Conference on Computer Vision, pages 404-421. Springer, 2022.", + "[25] S. R. Kheradpisheh and T. Masquelier. Temporal backpropagation for spiking neural networks with one spike per neuron. International Journal of Neural Systems, 30(6), 2020." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10560", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Jaehyun Kim, Heesu Kim, Subin Huh, Jinho Lee, and Kiyoung Choi. Deep neural networks with weighted spikes. Neurocomputing, 311:373-386, 2018.", + "[27] Jinyu Li et al. Recent advances in end-to-end automatic speech recognition. APSIPA Transactions on Signal and Information Processing, 11(1), 2022.", + "[28] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International conference on machine learning, pages 6316-6325. PMLR, 2021.", + "[29] Yuhang Li, Yufei Guo, Shanghang Zhang, Shikuang Deng, Yongqing Hai, and Shi Gu. Differentiable spike: Rethinking gradient-descent for training spiking neural networks. Advances in Neural Information Processing Systems, 34:23426-23439, 2021.", + "[30] Yuguo Liu, Wenyu Chen, Hanwen Liu, Yun Zhang, Malu Zhang, and Hong Qu. Biologically plausible sparse temporal word representations. IEEE Transactions on Neural Networks and Learning Systems, 2023.", + "[31] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10(9):1659-1671, 1997.", + "[32] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668-673, 2014.", + "[33] Hesham Mostafa. Supervised learning based on temporal coding in spiking neural networks. IEEE transactions on neural networks and learning systems, 29(7):3227-3235, 2017.", + "[34] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 36(6):51-63, 2019.", + "[35] Zihan Pan, Malu Zhang, Jibin Wu, Jiadong Wang, and Haizhou Li. Multi-tone phase coding of interaural time difference for sound source localization with spiking neural networks. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:2656-2670, 2021.", + "[36] Seongsik Park, Seijoon Kim, Hyeokjun Choe, and Sungroh Yoon. Fast and efficient information transmission with burst spikes in deep spiking neural networks. In Proceedings of the 56th Annual Design Automation Conference 2019, pages 1-6, 2019.", + "[37] Seongsik Park, Seijoon Kim, Byunggook Na, and Sungroh Yoon. T2fsnn: Deep spiking neural networks with time-to-first-spike coding. In 2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2020.", + "[38] Seongsik Park and Sungroh Yoon. Training energy-efficient deep spiking neural networks with time-to-first-spike coding. arXiv preprint arXiv:2106.02568, 2021.", + "[39] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with" + ], + "bbox": [ + 78, + 90, + 467, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019.", + "[40] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, and Michael Pfeiffer. Theory and tools for the conversion of analog to spiking convolutional neural networks. arXiv preprint arXiv:1612.04052, 2016.", + "[41] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, Michael Pfeiffer, and Shih-Chii Liu. Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in neuroscience, 11:682, 2017.", + "[42] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in neural information processing systems, 31, 2018.", + "[43] V. Sze, Y. H. Chen, T. J. Yang, and J. S. Emer. Efficient processing of deep neural networks: A tutorial and survey. Proceedings of the IEEE, 105(12), 2017.", + "[44] Yuchen Wang, Malu Zhang, Yi Chen, and Hong Qu. Signed neuron with memory: Towards simple, accurate and high-efficient ann-snn conversion. In International Joint Conference on Artificial Intelligence, 2022.", + "[45] Jibin Wu, Yansong Chua, Malu Zhang, Guoqi Li, Haizhou Li, and Kay Chen Tan. A tandem learning rule for effective training and rapid inference of deep spiking neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2021.", + "[46] Jibin Wu, Chenglin Xu, Xiao Han, Daquan Zhou, Malu Zhang, Haizhou Li, and Kay Chen Tan. Progressive tandem learning for pattern recognition with deep spiking neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7824-7840, 2021.", + "[47] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018.", + "[48] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 1311-1318, 2019.", + "[49] Qi Xu, Yaxin Li, Jiangrong Shen, Jian K Liu, Huajin Tang, and Gang Pan. Constructing deep spiking neural networks from artificial neural networks with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7886-7895, 2023.", + "[50] Qi Xu, Yaxin Li, Jiangrong Shen, Pingping Zhang, Jian K Liu, Huajin Tang, and Gang Pan. Hierarchical spiking-based model for efficient image classification with enhanced feature extraction and encoding. IEEE Transactions on Neural Networks and Learning Systems, 2022.", + "[51] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence, 2023.", + "[52] Qiang Yu, Huajin Tang, Kay Chen Tan, and Haizhou Li. Rapid feedforward computation by temporal encoding and learning with spiking neurons. IEEE transactions on neural networks and learning systems, 24(10):1539-1552, 2013." + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10561", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Friedemann Zenke and Tim P Vogels. The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks. Neural computation, 33(4):899-925, 2021.", + "[54] Jilin Zhang, Dexuan Huo, Jian Zhang, Chunqi Qian, Qi Liu, Liyang Pan, Zhihua Wang, Ning Qiao, Kea-Tiong Tang, and Hong Chen. 22.6 anp-i: A 28nm 1.5 pj/sop asynchronous spiking neural network processor enabling sub-o. 1 $\\mu$ j/sample on-chip learning for edge-ai applications. In 2023 IEEE International Solid-State Circuits Conference (ISSCC), pages 21–23. IEEE, 2023.", + "[55] Jian Zhang, Jilin Zhang, Dexuan Huo, and Hong Chen. Anas: Asynchronous neuromorphic hardware architecture search based on a system-level simulator. in proc. 60th acm/ieee design automation conference (dac), 2023. In Proc. 60th ACM/IEEE Design Automation Conference (DAC), 2023, 2023.", + "[56] Malu Zhang, Xiaoling Luo, Yi Chen, Jibin Wu, Ammar Belatreche, Zihan Pan, Hong Qu, and Haizhou Li. An efficient threshold-driven aggregate-label learning algorithm for multimodal information processing. IEEE Journal of Selected Topics in Signal Processing, 14(3):592-602, 2020.", + "[57] Malu Zhang, Hong Qu, Ammar Belatreche, Yi Chen, and Zhang Yi. A highly effective and robust membrane potential-driven supervised learning method for spiking neurons. IEEE transactions on neural networks and learning systems, 30(1):123-137, 2018.", + "[58] Malu Zhang, Jiadong Wang, Jibin Wu, Ammar Belatreche, Burin Amornpaisannon, Zhixuan Zhang, Venkata Pavan Kumar Miriyala, Hong Qu, Yansong Chua, Trevor E Carlson, et al. Rectified linear postsynaptic potential function for backpropagation in deep spiking neural networks. IEEE transactions on neural networks and learning systems, 33(5):1947-1958, 2021.", + "[59] Wenrui Zhang and Peng Li. Temporal spike sequence learning via backpropagation for deep spiking neural networks. Advances in Neural Information Processing Systems, 33:12022-12033, 2020.", + "[60] Shibo Zhou, Xiaohua Li, Ying Chen, Sanjeev T Chandrasekaran, and Arindam Sanyal. Temporal-coded deep spiking neural network with easy training and robust performance. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 11143-11151, 2021.", + "[61] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, pages 3-11. Springer, 2018.", + "[62] Zongwei Zhou, Jae Shin, Lei Zhang, Suryakanth Gurudu, Michael Gotway, and Jianming Liang. Fine-tuning convolutional neural networks for biomedical image analysis: actively and incrementally. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7340-7351, 2017." + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, YAN Shuicheng, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations, 2022.", + "[64] Rui-Jie Zhu, Qihang Zhao, Tianjing Zhang, Haoyu Deng, Yule Duan, Malu Zhang, and Liang-Jian Deng. Tcjasnn: Temporal-channel joint attention for spiking neural networks. arXiv preprint arXiv:2206.10177, 2022.", + "[65] Yaoyu Zhu, Zhaofei Yu, Wei Fang, Xiaodong Xie, Tiejun Huang, and Timothée Masquelier. Training spiking neural networks with event-driven backpropagation. In 36th Conference on Neural Information Processing Systems (NeurIPS 2022), 2022." + ], + "bbox": [ + 501, + 92, + 890, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10562", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/6f1f26ae-7b19-4337-978a-c5b559913901_model.json b/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/6f1f26ae-7b19-4337-978a-c5b559913901_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fde5afe95a2380c37d719729c5d803324692e4d6 --- /dev/null +++ b/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/6f1f26ae-7b19-4337-978a-c5b559913901_model.json @@ -0,0 +1,2510 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.13, + 0.868, + 0.177 + ], + "angle": 0, + "content": "Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold: Learning with Event-Driven Backpropagation" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.215, + 0.846, + 0.235 + ], + "angle": 0, + "content": "Wenjie Wei\\(^{1}\\), Malu Zhang\\(^{1*}\\), Hong Qu\\(^{1}\\), Ammar Belatreche\\(^{2}\\), Jian Zhang\\(^{3}\\), and Hong Chen\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.246, + 0.721, + 0.282 + ], + "angle": 0, + "content": "1University of Electronic Science and Technology of China, 2Northumbria University, 3Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.285, + 0.885, + 0.3 + ], + "angle": 0, + "content": "{maluzhang,hongqu}@uestc.edu.cn, ammar.belatreche@northumbria.ac.uk, hongchen@tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.345, + 0.314, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.383, + 0.473, + 0.837 + ], + "angle": 0, + "content": "Spiking Neural Networks (SNNs) offer a highly promising computing paradigm due to their biological plausibility, exceptional spatiotemporal information processing capability and low power consumption. As a temporal encoding scheme for SNNs, Time-To-First-Spike (TTFS) encodes information using the timing of a single spike, which allows spiking neurons to transmit information through sparse spike trains and results in lower power consumption and higher computational efficiency compared to traditional rate-based encoding counterparts. However, despite the advantages of the TTFS encoding scheme, the effective and efficient training of TTFS-based deep SNNs remains a significant and open research problem. In this work, we first examine the factors underlying the limitations of applying existing TTFS-based learning algorithms to deep SNNs. Specifically, we investigate issues related to over-sparsity of spikes and the complexity of finding the 'causal set'. We then propose a simple yet efficient dynamic firing threshold (DFT) mechanism for spiking neurons to address these issues. Building upon the proposed DFT mechanism, we further introduce a novel direct training algorithm for TTFS-based deep SNNs, called DTA-TTFS. This method utilizes event-driven processing and spike timing to enable efficient learning of deep SNNs. The proposed training method was validated on the image classification task and experimental results clearly demonstrate that our proposed method achieves state-of-the-art accuracy in comparison to existing TTFS-based learning algorithms, while maintaining high levels of sparsity and energy efficiency on neuromorphic inference accelerator." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.345, + 0.632, + 0.361 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Deep learning approaches have demonstrated remarkable performance in various intelligent applications [18, 27, 24, 23] and innovative simulations [62, 61]. However, these achievements come at a significant cost of energy consumption, severely limiting the deployment of artificial neural networks (ANNs) on resource-limited platforms [43, 9, 3, 50]. Brain-inspired spiking neural networks (SNNs) have emerged as a promising and energy-efficient alternative to ANNs [22, 31, 57]. SNNs have garnered attention for their distinctive features of sparse-asynchronous spikes and event-driven computing, which have played a pivotal role in driving the development of neuromorphic computing platforms such as ANP-I [54], TrueNorth [1], Loihi [8], and Tianjic [39]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.893, + 0.749 + ], + "angle": 0, + "content": "Despite the inherent energy efficiency of SNNs, their extensive application to real-world practical scenarios has been limited by the absence of efficient and scalable training algorithms [52, 35, 30, 14, 20]. Traditional training techniques for ANNs, such as error backpropagation (BP) and standard GPU-accelerated training packages, cannot be leveraged directly by SNNs due to the complex temporal dynamics and the non-differentiability of spiking neurons' activity [33, 6, 56]. To address these challenges, various solutions have been proposed, broadly categorized into three distinct groups." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The first category of solutions proposed to address the lack of scalable training algorithms for SNNs is ANN-to-SNN conversion method [40, 41, 49, 45]. These methods involve training an ANN first and then converting it to its SNN counterpart. The resulting SNN may achieve comparable performance to pre-trained ANNs through manual and careful setting of the SNN parameters such as the firing threshold [28] and the chosen spiking neuron model [44]. However, such conversion often results in precision loss due to approximation and requires longer time steps for lossless" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10552" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "conversion [19]. Moreover, ANN-to-SNN conversion methods often encode the activation values of artificial neurons using the firing rate of spiking neurons [11, 4, 5], which fails to fully leverage the timing information of spikes. Furthermore, the runtime computation on neuromorphic hardware can often be highly energy-intensive." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.471, + 0.456 + ], + "angle": 0, + "content": "The second category of the proposed SNN training solutions is RNN-like method, commonly known as surrogate gradient learning [47, 48]. In these surrogate gradient-based SNNs, information is conveyed through spikes during feedforward computation, while during backpropagation, gradient information is calculated by treating the neurons' membrane potential as a differentiable signal at the current time step [42, 34]. To address the non-differentiability of spikes in the direct training of SNNs, surrogate functions are employed to approximate the derivative of the spike generation function during the backward process [53, 29]. However, while these methods have successfully tackled the challenges associated with the non-differentiability of spiking neurons, they necessitate updating synaptic weights at each time step, regardless of whether there is a spike emission or not [47, 15]. Consequently, this results in substantial energy and memory demands as large intermediate activation values necessitate storage to facilitate gradient computation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.471, + 0.805 + ], + "angle": 0, + "content": "Unlike the above-mentioned methods, the third category of learning algorithms in SNNs is spike-driven method [2, 33, 58, 60, 38, 6], which operate in a strictly event-driven manner during both forward and backward computations. SpikeProp [2] and its various extensions [33, 6, 25] are typical examples of this category. However, these algorithms remain limited to shallow network structures and relatively simple classification tasks. To overcome this limitation, Zhang et al. [58] and Zhou et al. [60] extended the spikedriven algorithm to more complex network structures. Despite achieving competitive accuracy on large datasets [60], the identification of the 'causal set' in this approach requires substantial computational resources. To improve the training efficiency, Park et al. [38] proposed a surrogate deep neural network (DNN) model that is subsequently converted into a TTFS-based SNN. However, this conversion method is prone to conversion errors, potentially impairing the temporal learning capability of SNNs. Recently, several multi-spike-based spike-driven algorithms have been developed [59, 65]. While Zhang et al. [59] and Zhu et al. [65] achieve competitive performance on CIFAR utilizing the rate coding scheme, we focus on developing an energy-efficient TTFS-based spike-driven algorithm." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.806, + 0.47, + 0.837 + ], + "angle": 0, + "content": "In this paper, we introduce an effective and efficient spike-driven learning algorithm that facilitates the training" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "of high-performance TTFS-based deep SNNs. The main contributions of this work are outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.133, + 0.892, + 0.209 + ], + "angle": 0, + "content": "- We provide a comprehensive analysis of the main shortcomings of existing methods in achieving high performance in TTFS-based deep SNNs. Specifically, we examine issues such as the over-sparsity of spikes and the complexity of determining the 'causal set'." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.22, + 0.892, + 0.265 + ], + "angle": 0, + "content": "- We propose a simple yet efficient dynamic firing threshold (DFT) mechanism for spiking neurons that can effectively address the aforementioned issues." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.276, + 0.892, + 0.367 + ], + "angle": 0, + "content": "- Building upon the proposed DFT, we further introduce a novel direct training algorithm for TTFS-based deep SNNs, which we refer to as DTA-TTFS. In this training method, the timing of a single spike is considered the basic information carrier and the learning process is performed strictly in an event-driven manner." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.378, + 0.892, + 0.483 + ], + "angle": 0, + "content": "- The proposed method is validated through extensive experiments on benchmark image classification tasks and achieves state-of-the-art accuracy compared to existing TTFS-based learning algorithms. Furthermore, we demonstrate the ultra-low power capability of the SNN with DTA-TTFS on a recently developed neuromorphic accelerator." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.133, + 0.892, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.826, + 0.515 + ], + "angle": 0, + "content": "2. Preliminaries and Problem Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.524, + 0.892, + 0.584 + ], + "angle": 0, + "content": "In this section, we first briefly review the TTFS coding and the employed spiking neuron model. Subsequently, we provide a comprehensive analysis of the primary shortcomings in existing TTFS-based deep SNN methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.594, + 0.64, + 0.609 + ], + "angle": 0, + "content": "2.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.784 + ], + "angle": 0, + "content": "TTFS coding scheme: There are two primary coding schemes utilized to represent information in SNNs: rate coding and temporal coding. The rate coding method represents information based on the average firing rate of a spiking neuron. However, this encoding scheme neglects the temporal information carried by spikes, leading to higher spike frequency and increased energy consumption. Conversely, temporal coding leverages the precise timing of spikes to encode information, resulting in lower power consumption and higher computational efficiency compared to the rate-based coding method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.785, + 0.892, + 0.86 + ], + "angle": 0, + "content": "As a temporal coding method, TTFS encodes information by the timing of the first spike. As depicted in Fig. 1, the higher the activation value, the earlier the first spike is emitted. The relationship between the input activation value \\(a_{i}\\) and the encoded spike time \\(t_i\\) can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.871, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\nt _ {i} = \\left(1 - \\frac {a _ {i}}{a _ {\\text {m a x}}}\\right) \\times T _ {w}, \\tag {1}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.851, + 0.47, + 0.9 + ], + "angle": 0, + "content": "1Assume a postsynaptic neuron fire a first spike at \\( t_{out} \\), only a subset of input spikes had arrived before \\( t_{out} \\) and contributed to the output spike. This set of spikes, \\( C = \\{i : t_i < t_{out}\\} \\), is called the causal set in this work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10553" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.299, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.203, + 0.19, + 0.214 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.089, + 0.455, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.202, + 0.391, + 0.214 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.217, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Figure 1. (a) TTFS coding scheme. The larger intensity value corresponds to the earlier spike. (b) ReL-PSP spiking neuron model. Neuron \\( j \\) accumulates the linear PSP and fires a spike when the membrane potential crosses the firing threshold \\( \\vartheta \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.469, + 0.417 + ], + "angle": 0, + "content": "where \\( a_{max} \\) is the maximum input activation value, \\( T_w \\) is the length of the permissible spike time window. As the TTFS coding scheme requires only a single spike to convey information, it effectively diminishes the number of spikes and significantly decreases the energy requirements. Consequently, it achieves higher sparsity and energy efficiency, making it a powerful coding scheme that maximizes the energy efficiency of SNNs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.469, + 0.508 + ], + "angle": 0, + "content": "Spiking neuron model: The Rectified Linear Postsynaptic Potential (ReL-PSP) spiking neuron [58] is chosen for our method due to its simplicity and energy efficiency. As shown in Fig. 1(b), the ReL-PSP neuron model utilizes a rectified linear kernel function to shape the neuron's PSP, and its dynamics is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.519, + 0.469, + 0.559 + ], + "angle": 0, + "content": "\\[\nV _ {j} ^ {l} (t) = \\sum_ {i = 1} ^ {N} w _ {i j} ^ {l} K \\left(t - t _ {i} ^ {l - 1}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.57, + 0.469, + 0.651 + ], + "angle": 0, + "content": "where \\( V_{j}^{l}(t) \\) is the membrane potential of neuron \\( j \\) in layer \\( l \\), and \\( w_{ij}^{l} \\) is the synaptic weight between neuron \\( j \\) and its presynaptic neuron \\( i \\). \\( N \\) is the number of neurons in the previous layer \\( l - 1 \\), \\( t_{i}^{l - 1} \\) is the spike time of input neuron \\( i \\), and \\( K\\left(t - t_{i}^{l - 1}\\right) \\) is the PSP function that is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.66, + 0.469, + 0.701 + ], + "angle": 0, + "content": "\\[\nK \\left(t - t _ {i} ^ {l - 1}\\right) = \\left\\{ \\begin{array}{l l} t - t _ {i} ^ {l - 1}, & \\text {i f} \\quad t > t _ {i} ^ {l - 1}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.711, + 0.469, + 0.757 + ], + "angle": 0, + "content": "The neuron \\( j \\) emits a spike when \\( V_{j}^{l}(t) \\) reaches the firing threshold \\( \\vartheta \\). Mathematically, the spike generation function \\( F \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.766, + 0.469, + 0.786 + ], + "angle": 0, + "content": "\\[\nt _ {j} ^ {l} = F \\{t | V _ {j} ^ {l} (t) \\geq \\vartheta , t \\geq 0 \\}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.9 + ], + "angle": 0, + "content": "The ReL-PSP neuron has proven its efficacy in tackling the main challenges encountered during the training of deep SNNs, such as the non-differentiability of the spike function, gradient explosion, and the dead neuron issue. Despite the successful performance on the MNIST dataset, its performance is not ideal on deeper network structures and more complex datasets, such as CIFAR-10 and CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "In the subsequent section, we will conduct a comprehensive analysis to identify the underlying reasons for this limitation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.146, + 0.671, + 0.163 + ], + "angle": 0, + "content": "2.2. Problem Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.169, + 0.892, + 0.29 + ], + "angle": 0, + "content": "Over-sparsity of spikes: Sparse neural networks offer several advantages such as efficient variable-size representation, information decomposition, and energy efficiency. However, over-sparsity may diminish the model's effective capacity and adversely affect its performance [17]. This issue becomes more serious in TTFS-based deep SNNs, owing to their firing mechanism, asynchronous transmission, and TTFS-based decision strategy." + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.302, + 0.689, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.414, + 0.617, + 0.426 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.304, + 0.859, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.414, + 0.792, + 0.426 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.54 + ], + "angle": 0, + "content": "Figure 2. The over-sparsity problem caused by asynchronous transmission and TTFS-based decision strategy. (a) The presynaptic (gray) neuron fires later than the postsynaptic (blue) neuron \\( j \\), making it ineffective in contributing to the firing of the postsynaptic neuron. As a result, this presynaptic neuron becomes invalid. (b) After learning, the postsynaptic (blue) neuron fires at an earlier time, leading to the appearance of additional invalid neurons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.765 + ], + "angle": 0, + "content": "According to the firing mechanism of spiking neurons, a neuron emits one spike only when its membrane potential reaches the firing threshold. This leads to fewer activated spiking neurons compared to ANNs. In addition, the asynchronous transmission of SNNs further diminishes the number of valid activated neurons. As shown in Fig. 2(a), when the presynaptic neuron \\(i\\) fires after the postsynaptic neuron \\(j\\), the spike of neuron \\(i\\) becomes invalid for neuron \\(j\\). Furthermore, TTFS-based SNNs make decisions based on the earliest spike, and the synaptic weights are trained to make the target neuron fire as early as possible. As shown in Fig. 2(b), after learning, the firing time of the postsynaptic neuron \\(j\\) may become earlier, thereby further exacerbating the sparsity induced by asynchronous transmission." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In event-driven learning algorithms of SNNs, error backpropagation depends on the timing of spikes. Therefore, if a spiking neuron does not fire or does not contribute to the postsynaptic neuron, it will not participate in learning. The issue of over-sparsity not only limits the information representation capability of SNNs but also leads to learning failure. Consequently, the achievement of a harmonious balance between sparsity and performance in the training of TTFS-based deep SNNs remains an open question." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10554" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "The complexity of finding the 'causal set': Due to the sparsity and asynchronicity of SNNs, not all presynaptic neurons \\(\\{i|i = 1,\\dots ,N\\}\\) contribute to their postsynaptic neuron \\(j\\), where the contributory neurons are defined as the 'causal set' of the postsynaptic neuron \\(j\\), denoted as \\(C_j = \\{i|t_i < t_j\\}\\) [33, 60]. To leverage standard GPU-accelerated ANN training packages, it is crucial to identify the 'causal set' of each neuron during the feed-forward computation process [33, 60]. However, this process is highly complex and time-consuming. For example, in order to find the 'causal set', Zhou et al. [60] proposed an approach where they first sort all of this neuron input spikes \\(\\{t_i|i = 1,\\dots N\\}\\) in ascending order and subsequently consider each individual spike \\(t_i\\) for inclusion in the causal set. This approach, however, is computationally expensive and results in inefficient training even with the implementation of a space-for-time strategy [60]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.501 + ], + "angle": 0, + "content": "Despite the success of algorithms with this approach in utilizing standard GPU-accelerated ANN training packages, the training process involves a significant amount of sorting and iteration. This results in higher demand for training resources and increased latency, especially as the network depth increases. Consequently, the process of finding the 'causal set' may become a bottleneck for training deep SNNs. In the next section, we propose a dynamic firing threshold mechanism for spiking neurons to avoid this resource-intensive process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.512, + 0.17, + 0.527 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.613 + ], + "angle": 0, + "content": "In this section, we first introduce the dynamic firing threshold (DFT) for spiking neurons and discuss how it effectively addresses the above issues. Then, based on the DFT, we further propose a direct training algorithm for TTFS-based SNN, referred to as DTA-TTFS." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.622, + 0.377, + 0.639 + ], + "angle": 0, + "content": "3.1. DFT-based Spiking Neuron Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.689 + ], + "angle": 0, + "content": "As demonstrated in Fig. 3(a), the proposed DFT is a layer-dependent and time-varying function, which is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.697, + 0.47, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\vartheta^ {l} (t) = \\left\\{ \\begin{array}{l l} T _ {w} (l + 1) - t, & \\text {i f} \\quad T _ {w} l \\leq t \\leq T _ {w} (l + 1), \\\\ + \\infty , & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\(\\vartheta^l (t)\\) represents the firing threshold of the neurons in the \\(l\\)-th layer and \\(T_{w}\\) denotes the length of the permissible spike time window. In this work, \\(T_{w}\\) is set to 1 (unless otherwise stated) and the input layer is regarded as the 0-th layer. Eq. 5 reveals that the firing threshold is infinite outside the interval \\(t\\in [T_wl,T_w(l + 1)]\\), and neurons in the \\(l\\)-th layer can only generate a spike within this time interval. Therefore, as shown in Fig. 4, the generated spikes are non-overlapping across different layers, which we refer to as the layer-dependent firing feature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "Based on the membrane potential dynamics in Eq. 2 and the proposed DFT in Eq. 5, the output of neuron \\( j \\) can be divided into three categories:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.148, + 0.892, + 0.243 + ], + "angle": 0, + "content": "- Spikes at \\( t_j^l = T_{wl} \\): Since the firing threshold is infinity when \\( t < T_{wl} \\), the neuron cannot fire a spike before \\( t = T_{wl} \\) even though the membrane potential is strong (i.e., \\( V_j^l(t) \\geq \\vartheta^l(T_{wl}) \\) and \\( t < T_{wl} \\)). As shown in Fig. 3(b), the earliest spike time for DFT-based spiking neurons in the \\( l \\)-th layer is \\( t_j^l = T_{wl} \\)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.252, + 0.892, + 0.361 + ], + "angle": 0, + "content": "- Spikes within \\((T_{w}l, T_{w}(l + 1))\\): The neuron generates a spike when its membrane potential exceeds the threshold, i.e. \\(V_{j}^{l}(t) \\geq \\vartheta^{l}(t)\\). The firing threshold \\(\\vartheta^{l}\\) in this time interval is a linear function, which decays linearly from 1 to 0. As depicted in Fig. 3(c), by combining the membrane potential in Eq. 2 and the firing threshold in Eq. 5, the precise spiking time \\(t_{j}^{l}\\) can be calculated as" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.148, + 0.892, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.372, + 0.892, + 0.431 + ], + "angle": 0, + "content": "\\[\nt _ {j} ^ {l} = \\frac {T _ {w} (l + 1) + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l} t _ {i} ^ {l - 1}}{1 + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.449, + 0.892, + 0.51 + ], + "angle": 0, + "content": "- Nonspike: The neuron will not fire a spike if its membrane potential is below the dynamic firing threshold (i.e., \\( V_{j}^{l}(t) < \\vartheta^{l}(t) \\)). Fig. 3(d) shows three typical scenarios of nonspike." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Overall, the spike time of a DFT-based neuron \\( j \\) in the \\( l \\)-th layer can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.565, + 0.892, + 0.657 + ], + "angle": 0, + "content": "\\[\nt _ {j} ^ {l} = \\left\\{ \\begin{array}{l l} T _ {w} l, & \\text {i f} \\quad V _ {j} ^ {l} \\left(T _ {w} l\\right) \\geq \\vartheta^ {l} \\left(T _ {w} l\\right), \\\\ \\frac {T _ {w} (l + 1) + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l} t _ {i} ^ {l - 1}}{1 + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l}}, & \\text {i f} \\quad T _ {w} l < t _ {j} ^ {l} \\leq T _ {w} (l + 1), \\\\ \\text {n o n s p i k e}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.703 + ], + "angle": 0, + "content": "In the following, we will analyze how the proposed DFT effectively addresses issues of over-sparsity of spikes and the complexity of finding the 'causal set'." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.765 + ], + "angle": 0, + "content": "1) Over-sparsity of spikes: As described in Eq. 5, the DFT is a piecewise function, whose linearly decreased part and layer-dependent firing feature effectively address the over-sparsity of spikes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The linearly decreased part of the DFT is purposefully designed to tackle the over-sparsity arising from the spiking neuron firing mechanism. By gradually decreasing the firing threshold from 1 to 0, the neuron \\( j \\) remains active as long as its membrane potential is non-negative at \\( T_{w}(l + 1) \\). As a result, the proposed DFT effectively reduces the number of inactive neurons. Moreover, the linear decay design enhances the discriminative nature of information from active neurons." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10555" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.094, + 0.282, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.199, + 0.188, + 0.21 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.093, + 0.486, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.199, + 0.396, + 0.21 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.092, + 0.687, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.199, + 0.597, + 0.21 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.089, + 0.888, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.199, + 0.797, + 0.21 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.215, + 0.894, + 0.256 + ], + "angle": 0, + "content": "Figure 3. Proposed dynamic firing threshold and output of the DFT-based spiking neurons. (a) Dynamic firing threshold function(DFT). (b) Spikes at \\( t_j^l = T_w l \\). (c) Spikes within \\( (T_w l, T_w (l + 1)] \\). (d) Nonspike. This case is further divided into three sub-cases, which are shown by the curves of different colors." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.277, + 0.452, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.391, + 0.444, + 0.405 + ], + "angle": 0, + "content": "Figure 4. The propagation pipeline of the SNN with DFT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.468, + 0.598 + ], + "angle": 0, + "content": "The layer-dependent firing feature enables SNN with non-overlapping spike time windows. This effectively resolves the over-sparsity arising from asynchronous transmission and TTFS-based decision strategy. Fig. 4 illustrates that the spike time window of postsynaptic neurons occurs later than that of presynaptic neurons. Consequently, when the presynaptic neuron fires a spike, its postsynaptic neuron absolutely receives it. Furthermore, the non-overlapping spike window provides output layer neurons with a predetermined earliest spike time, preventing them from being forced forward indefinitely during the training process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.468, + 0.673 + ], + "angle": 0, + "content": "2) The complexity of finding the 'causal set': The layer-dependent firing feature of DFT circumvents the need for SNNs to identify the 'causal set', resulting in efficient training. This efficiency is further improved by the output analytic equation of the DFT-based neuron." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.468, + 0.825 + ], + "angle": 0, + "content": "As depicted in Fig. 4, the layer-dependent firing feature ensures that in the \\(l\\)-th layer of the SNN utilizing DFT, the spike time is controlled within the interval \\([T_w l, T_w (l + 1)]\\). Therefore, we can easily learn that neurons in the \\((l - 1)\\)-th layer must fire earlier than neurons in the \\(l\\)-th layer. In other words, all active presynaptic neurons must remain in the 'causal set' of their postsynaptic neurons. Therefore, the DFT mechanism prevents the need for resource-intensive sorting and iteration process, enhancing training efficiency and facilitating its extension to deeper network architecture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In addition, the employed ReL-PSP spiking neuron and the DFT provide an analytical relation between input spikes and output spikes, as described in Eq.7. Therefore, the dynamics of SNNs within the training loop need not be simulated using long time steps. This allows the utilization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.281, + 0.89, + 0.312 + ], + "angle": 0, + "content": "of standard GPU-accelerated training packages of ANNs, which further leads to an efficient training process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.328, + 0.783, + 0.344 + ], + "angle": 0, + "content": "3.2. DTA-TTFS Learning Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.505 + ], + "angle": 0, + "content": "Consider a task with \\( n \\) categories, where each output neuron corresponds to a specific category. In this task, the target category is represented by the output neuron that fires the earliest. To assess the effectiveness of the model, we adopt the widely utilized cross-entropy loss function \\( L \\), which aims to maximize the output value of the target neuron. In order to ensure the desired neuron fires earliest and others fire as late as possible, we utilize the negative spike times of the output layer [58, 6], i.e., \\( -t \\), for evaluation. Therefore, the loss function is given by" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.52, + 0.891, + 0.554 + ], + "angle": 0, + "content": "\\[\nL (\\mathbf {t}, d) = - \\log \\frac {\\exp (- \\mathbf {t} [ d ])}{\\sum_ {i = 1} ^ {n} \\exp (- \\mathbf {t} [ i ])}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.568, + 0.89, + 0.598 + ], + "angle": 0, + "content": "where \\(\\mathbf{t}\\) is the vector of spike times in the output layer, and \\(d\\) is the desired category index." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.709 + ], + "angle": 0, + "content": "In order to minimize the \\(L(\\mathbf{t},d)\\), the DTA-TTFS algorithm adjusts spike times by modifying the synaptic weights during gradient backpropagation. For driving the gradient backpropagation, it is necessary to calculate derivatives of the output spike time \\(t_j^l\\) with respect to the synaptic weight \\(w_{ij}^{l}\\) and the input spike time \\(t_i^{l - 1}\\). According to Eq. 7, we can easily get these two derivatives," + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.724, + 0.891, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial t _ {j} ^ {l}}{\\partial w _ {i j} ^ {l}} = \\left\\{ \\begin{array}{l l} \\frac {t _ {i} ^ {l - 1} - t _ {j} ^ {l}}{1 + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l}}, & \\text {i f} \\quad T _ {w} l < t _ {j} ^ {l} < T _ {w} (l + 1), \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.802, + 0.891, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial t _ {j} ^ {l}}{\\partial t _ {i} ^ {l - 1}} = \\left\\{ \\begin{array}{l l} \\frac {w _ {i j} ^ {l}}{1 + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l}}, & \\text {i f} \\quad T _ {w} l < t _ {j} ^ {l} < T _ {w} (l + 1), \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Following Eq. 9 and Eq. 10, gradient backpropagation can be directly applied." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10556" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.156, + 0.089, + 0.817, + 0.328 + ], + "angle": 0, + "content": "
DatasetsModelsNetwork ArchitectureNeural CodingMethodAccuracySparsity
MNISTMostafa 2017 [33]MLP¹TFTSDT97.55%0.51
Zhang et al. 2020 [58]CNN¹TFTSDT99.4%0.6614
Zhou et al. 2021 [60]CNN²TFTSDT99.33%0.94
DTA-TTFSCNN¹TFTSDT99.4%0.3913
CIFAR-10Wu et al. 2022 [46]VGG11Rateconv91.24%no
Park et al. 2020 [37]VGG16TFTSconv91.43%0.2459
Zhou et al. 2021 [60]VGG16TFTSDT92.68%0.62
Park et al. 2021 [38]VGG16TFTSDT91.90%0.1746
DTA-TTFSVGG11TFTSDT91.17%0.4387
DTA-TTFSVGG16TFTSDT93.05%0.2561
CIFAR-100Park et al. 2020 [37]VGG16TFTSconv68.79%0.2994
Park et al. 2021 [38]VGG16TFTSDT65.98%0.2780
DTA-TTFSVGG16TFTSDT69.66%0.2845
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.337, + 0.892, + 0.365 + ], + "angle": 0, + "content": "Table 1. Comparison of accuracy and sparsity between the DTA-TTFS method and other related works. DT: direct training, conv: ANN-to-SNN conversion. MLP1: 784-800-10. CNN1: 784-16C5-P2-32C5-P2-800-128-10. CNN2: 784-32C5-16C5-10." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.387, + 0.21, + 0.403 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.471, + 0.504 + ], + "angle": 0, + "content": "In this section, we first evaluate the performance of the DTA-TTFS algorithm. Then, we carry out validation studies to illustrate the impact of the DFT. Finally, we employ theoretical analysis along with the neuromorphic hardware simulator to evaluate the energy consumption during the inference process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.51, + 0.168, + 0.525 + ], + "angle": 0, + "content": "4.1. Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.79 + ], + "angle": 0, + "content": "We conduct experiments on three image classification benchmark datasets: MNIST\\(^2\\), CIFAR-10, and CIFAR\\(100^3\\). All experiments are conducted using the PyTorch library, which allows for efficient training on multi-GPU machines with accelerated computations and optimized memory usage. For the MNIST dataset, we employ a network structure with 784-16C5-P2-32C5-P2-800-128-10 and train it for 150 epochs. As for CIFAR datasets, we adopt augmentation techniques [7, 10] and train the VGG11 and VGG16 network structures without Batch Normalization (BN) for 300 epochs. Moreover, in order to maintain the initial parameter domain at a good level, the SNN is initialized with the parameters from the pre-trained identical ANN when training CIFAR datasets. During the training process, we adopt the Adam optimizer and implement milestones learning rate decay, with an initial learning rate \\(1e^{-3}\\) for MNIST and \\(1e^{-4}\\) for CIFAR datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.799, + 0.367, + 0.815 + ], + "angle": 0, + "content": "4.2. Comparison with Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.822, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Tab. 1 presents a comparison of the accuracies and sparsity achieved by DTA-TTFS and other related works. Experimental results demonstrate that the DTA-TTFS algo" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.388, + 0.895, + 0.66 + ], + "angle": 0, + "content": "rithm exhibits excellent performance, with an accuracy of \\(99.4\\%\\) on MNIST, \\(93.05\\%\\) on CIFAR-10, and \\(69.66\\%\\) on CIFAR-100. Obviously, our method outperforms all previous direct learning algorithms of TTFS-based SNNs. Moreover, our method still maintains competitive sparsity when achieving high performance. The sparsity is calculated by \\(S / N\\) averaged on the entire testing set, where \\(S\\) represents the total number of spikes and \\(N\\) denotes the total number of neurons. Specifically, the DTA-TTFS achieves sparsity of 0.3913 on MNIST, 0.2561 on CIFAR-10, and 0.2845 on CIFAR-100. For the MNIST dataset, when employing the identical network structure as described in [58], our method achieves the same accuracy while significantly reducing the sparsity from 0.6614 to 0.3913. For CIFAR datasets, our method surpasses related works in terms of accuracy while maintaining a competitive sparsity. In summary, the DTA-TTFS algorithm achieves a harmonious balance between accuracy and sparsity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.677, + 0.666, + 0.693 + ], + "angle": 0, + "content": "4.3. Validation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.807 + ], + "angle": 0, + "content": "We conduct a series of validation experiments to confirm the functionality of the proposed DFT. These experiments aim to display two aspects. Firstly, the DFT provides SNNs with the layer-dependent firing feature. Additionally, it effectively addresses the over-sparsity of spikes. Validation experiments choose the well-trained VGG11 structure on the CIFAR-10 dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In order to show the layer-dependent firing feature of the SNN with DFT, we record the spike activity in each layer for the first sample of the testing set. For the purpose of comparison, we also provide the spike activity of the SNN without DFT. The results are visualized in Fig. 5, elucidating the evident regulation of spike activity in each layer" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.286, + 0.888 + ], + "angle": 0, + "content": "2 http://yann.lecun.com/exdb/mnist/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.327, + 0.9 + ], + "angle": 0, + "content": "3https://www.cs.toronto.edu/ kriz/cifar.html" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.327, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10557" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.455, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.263, + 0.457, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.433, + 0.47, + 0.461 + ], + "angle": 0, + "content": "Figure 5. Spike activity in each layer of SNN with and without DFT. The vertical axis represents the neuron index scaled to [0,1]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.534 + ], + "angle": 0, + "content": "within a permitted time window for the SNN incorporating DFT. In contrast, the spike activity in the SNN without DFT exhibits a disorderly and chaotic pattern." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.471, + 0.702 + ], + "angle": 0, + "content": "In order to verify the efficacy of the DFT in addressing the over-sparsity of spikes, we count the number of inactive neurons per layer averaged on the whole testing set. This experiment is conducted on three models: the ANN, the SNN with DFT, and the SNN without DFT [58]. As shown in Fig. 6, the SNN without DFT displays significant sparsity, however, with only \\(10\\%\\) accuracy. Conversely, the SNN with DFT exhibits a comparable amount of inactive neurons to that of ANN. As a result, the DFT achieves a harmonious trade-off between sparse representation and information representation ability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.714, + 0.258, + 0.73 + ], + "angle": 0, + "content": "4.4. Energy Estimation" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.739, + 0.275, + 0.755 + ], + "angle": 0, + "content": "4.4.1 Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.871 + ], + "angle": 0, + "content": "Energy consumption is a critical metric to evaluate the efficiency of SNNs. To prove the energy efficiency of the proposed DTA-TTFS, we employ the theoretical analysis presented in [37]. This method analyzes energy consumption based on multiple factors, including latency, spike counts, as well as data related to neuromorphic architecture. More specifically, the estimated energy is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.886, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {T o t a l}} = E _ {\\text {S t a t i c}} \\times \\text {L a t e n c y} + E _ {\\text {D y n a m i c}} \\times \\text {S p i k e s}, \\tag {11}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.892, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.263, + 0.892, + 0.292 + ], + "angle": 0, + "content": "Figure 6. Percentage of inactive neurons per layer of the ANN, the SNN with DFT, and the SNN without DFT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.893, + 0.575 + ], + "angle": 0, + "content": "where Latency is the required time steps, Spikes is the total number of spikes, \\( E_{Static} \\) and \\( E_{Dynamic} \\) are static and dynamic energy coefficients depending on the neuromorphic architecture. We choose two neuromorphic architectures, TrueNorth [32] and SpiNNaker [16], to conduct analysis. The energy parameters \\( (E_{Static}, E_{Dynamic}) \\) are set to (0.6, 0.4) for TrueNorth and (0.36, 0.64) for SpiNNaker. For comparison, we analyze various coding schemes, such as rate coding [19], phase coding [26], burst coding [36], and TTFS coding [37, 38]. Theoretical analysis is performed on CIFAR-10 with the well-trained VGG16 structure, and we normalized estimated energy consumption based on the rate coding. As indicated in Tab. 2, under the same conditions, the DTA-TTFS outperforms other methods in terms of energy efficiency. It is worth noting that, despite the fewer spikes achieved by [37, 38], their energy consumption remains higher than ours due to their demand for long latency." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.587, + 0.891, + 0.715 + ], + "angle": 0, + "content": "
Neural CodingTime StepSpike (106)Acc. (%)Normalized Energy
TrueNorthSpiNNaker
Rate [19]5122.61293.3911
Phase [26]150035.19691.217.14769.6785
Burst [36]11256.9291.412.37812.4865
TTFS [37]6800.06991.430.80740.4950
TTFS [38]5440.06791.90.64780.3989
DTA-TTFS1600.07393.050.19870.1304
" + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.719, + 0.855, + 0.734 + ], + "angle": 0, + "content": "Table 2. Theoretical analysis of energy consumption." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.707, + 0.785 + ], + "angle": 0, + "content": "4.4.2 Hardware Simulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In order to further verify the energy efficiency and hardware friendliness of our method, we map the SNN trained by the DTA-TTFS algorithm on a recently developed neuromorphic hardware simulator, called configurable asynchronous neuromorphic hardware simulator (CanMore) [55]. CanMore estimates hardware performance through system-level modeling and simulation. Fig. 7 illustrates the main" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10558" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.473, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.47, + 0.31 + ], + "angle": 0, + "content": "Figure 7. The framework of CanMore [55], including neuromorphic cores comprised of neurons and SRAM, and NoC routers consisting of buffers, route computation units, switch allocator, and crossbar switch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.469, + 0.484 + ], + "angle": 0, + "content": "framework of the CanMore, including neuromorphic cores comprised of neurons and SRAM, and network-on-chip (NoC) routers consisting of buffers, route computation unit, switch allocator and crossbar switch. With the system-level asynchronous circuit modeling, CanMore is able to simulate various neuromorphic hardware architectures by changing the parameters and configuration files, estimate latency and energy consumption, and analyze spike messages routing information such as routing hops, congestion latency, routing directions, and so on." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.727 + ], + "angle": 0, + "content": "In [13], a deep ANN is mapped to TrueNorth neuromorphic hardware. To the best of our knowledge, this is the only work that verifies the energy efficiency of SNNs on the CIFAR-10 dataset, while the majority of existing works focus on smaller datasets, such as MNIST [12, 58]. Here, we train an SNN on the CIFAR-10 dataset with the proposed DTA-TTFS method. After learning, the learned weights are transferred to CanMore for accelerating the inference operations. Specifically, our network structure is mapped by partitioning each layer into equally sized groups along the feature dimension, with each group being mapped to a neuromorphic core with a maximum of 1024 neurons. In addition, we compress the network by quantizing the weights to 4 bits. The mesh topology and XY routing algorithm are adopted in the simulated architecture. The energy consumption per frame during the inference process is calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.738, + 0.469, + 0.755 + ], + "angle": 0, + "content": "\\[\nE _ {T o t a l} = E _ {S O P} \\times \\text {S p i k e s} + E _ {H o p} \\times \\text {H o p s} + E _ {\\text {S t a t i c}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\( E_{SOP} \\) is the dynamic energy consumption per synaptic operation (SOP), which is calculated with the real technology parameters of the CMOS 22nm process in the simulation. In our work, \\( E_{SOP} \\) is \\( 20\\mathrm{pJ} \\). \\( E_{Hop} \\) is the dynamic energy consumption per hop between the mesh routers on the NoC, which is \\( 100\\mathrm{pJ} \\) according to [21]. \\( E_{Static} \\) is static energy consumption, which is \\( 3.52\\mu \\mathrm{J}/\\mathrm{frame} \\) in CMOS 22nm process. The average Spikes and Hops per frame during the inference process are 71M and 45.4M re" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.087, + 0.687, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.089, + 0.885, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.27, + 0.892, + 0.298 + ], + "angle": 0, + "content": "Figure 8. The accuracy and energy consumption of Esser et al. [13] and DTA-TTFS on CIFAR-10 dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.367 + ], + "angle": 0, + "content": "spectively, given by CanMore. Therefore, we obtain the energy consumption \\( E_{Total} \\) during the inference process \\( 50.36\\mu \\mathrm{J / frame} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.893, + 0.488 + ], + "angle": 0, + "content": "The accuracy and energy consumption are reported in Fig. 8. our work achieves a classification accuracy of \\(90.67\\%\\) on CIFAR-10, surpassing the accuracy reported in [13] by \\(7.26\\%\\). In addition, the energy consumption of our work is \\(50.36~\\mu \\mathrm{J} / \\mathrm{frame}\\), which is \\(69.2\\%\\) reduced compared with the energy consumption \\(163.65\\mu \\mathrm{J} / \\mathrm{frame}\\) in [13]. These results confirm that the proposed DTA-TTFS algorithm is effective and energy-efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.503, + 0.619, + 0.518 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.893, + 0.786 + ], + "angle": 0, + "content": "In this study, we analyze the shortcomings of existing methods in achieving high-performance TTFS-based deep SNNs. In order to overcome these issues, we propose an efficient DFT mechanism for spiking neurons and introduce the DTA-TTFS algorithm for TTFS-based SNNs. We evaluate our method on the image classification task, and the results demonstrate that our method achieves a fine trade-off between accuracy and sparsity. Additionally, we estimate the energy consumption of our method through theoretical analysis and the neuromorphic hardware accelerator CanMore. Both theoretical analysis and hardware verification confirm the energy efficiency of the DTA-TTFS algorithm. Overall, our method not only attains high performance but also is quite hardware friendly. In our future work, we intend to leverage DTA-TTFS to train more advanced network models, including SNN-based Attention [64, 51] and Transformer architectures [63]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.652, + 0.817 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported by the National Science Foundation of China under Grants 62236007 and 61976043 and 62106038, and in part by the Sichuan Science and Technology Program under Grants 2022YFG0313 and 2023YFG0259." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10559" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[1] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34(10):1537-1557, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.257 + ], + "angle": 0, + "content": "[2] Sander M Bohte, Joost N Kok, and Johannes A La Poutre. Spikeprop: backpropagation for networks of spiking neurons. In *ESANN*, volume 48, pages 419–424. Bruges, 2000." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.468, + 0.327 + ], + "angle": 0, + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.468, + 0.385 + ], + "angle": 0, + "content": "[4] Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Optimized potential initialization for low-latency spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11-20, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.468, + 0.442 + ], + "angle": 0, + "content": "[5] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.468, + 0.527 + ], + "angle": 0, + "content": "[6] Iulia-Maria Comsa, Krzysztof Potempa, Luca Versari, Thomas Fischbacher, Andrea Gesmundo, and Jyrki Alakuijala. Temporal coding in spiking neural networks with alpha synaptic function: learning with backpropagation. IEEE transactions on neural networks and learning systems, 33(10):5939-5952, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.468, + 0.599 + ], + "angle": 0, + "content": "[7] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.468, + 0.67 + ], + "angle": 0, + "content": "[8] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 38(1):82-99, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.672, + 0.468, + 0.728 + ], + "angle": 0, + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.468, + 0.772 + ], + "angle": 0, + "content": "[10] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[11] Peter U Diehl, Daniel Neil, Jonathan Binas, Matthew Cook, Shih-Chii Liu, and Michael Pfeiffer. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[12] Steve K Esser, Rathinakumar Appuswamy, Paul Merolla, John V Arthur, and Dharmendra S Modha. Backpropagation for energy-efficient neuromorphic computing. Advances in neural information processing systems, 28, 2015." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[13] Steven K Esser, Paul A Merolla, John V Arthur, Andrew S Cassidy, Rathinakumar Appuswamy, Alexander Andreopoulos, David J Berg, Jeffrey L McKinstry, Timothy Melano, Davis R Barch, et al. From the cover: Convolutional networks for fast, energy-efficient neuromorphic computing. Proceedings of the National Academy of Sciences of the United States of America, 113(41):11441, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[14] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[15] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothée Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2661-2671, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[16] Steve B Furber, Francesco Galluppi, Steve Temple, and Luis A Plana. The spinnaker project. Proceedings of the IEEE, 102(5):652-665, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[17] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 315-323. JMLR Workshop and Conference Proceedings, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[18] Alex Graves and Navdeep Jaitly. Towards end-to-end speech recognition with recurrent neural networks. In International conference on machine learning, pages 1764-1772. PMLR, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[19] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13558–13567, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[20] Yangfan Hu, Huajin Tang, and Gang Pan. Spiking deep residual networks. IEEE Transactions on Neural Networks and Learning Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[21] Masashi Imai, Thiem Van Chu, Kenji Kise, and Tomohiro Yoneda. The synchronous vs. asynchronous noc routers: an apple-to/apple comparison between synchronous and transition signaling asynchronous designs. In 2016 Tenth IEEE/ACM International Symposium on Networks-on-Chip (NOCS), pages 1-8. IEEE, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.732 + ], + "angle": 0, + "content": "[22] Eugene M Izhikevich. Simple model of spiking neurons. IEEE Transactions on neural networks, 14(6):1569-1572, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[23] Yeying Jin, Aashish Sharma, and Robby T Tan. Dshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5027-5036, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[24] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In European Conference on Computer Vision, pages 404-421. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] S. R. Kheradpisheh and T. Masquelier. Temporal backpropagation for spiking neural networks with one spike per neuron. International Journal of Neural Systems, 30(6), 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10560" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[26] Jaehyun Kim, Heesu Kim, Subin Huh, Jinho Lee, and Kiyoung Choi. Deep neural networks with weighted spikes. Neurocomputing, 311:373-386, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[27] Jinyu Li et al. Recent advances in end-to-end automatic speech recognition. APSIPA Transactions on Signal and Information Processing, 11(1), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.468, + 0.234 + ], + "angle": 0, + "content": "[28] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International conference on machine learning, pages 6316-6325. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.468, + 0.303 + ], + "angle": 0, + "content": "[29] Yuhang Li, Yufei Guo, Shanghang Zhang, Shikuang Deng, Yongqing Hai, and Shi Gu. Differentiable spike: Rethinking gradient-descent for training spiking neural networks. Advances in Neural Information Processing Systems, 34:23426-23439, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.307, + 0.468, + 0.361 + ], + "angle": 0, + "content": "[30] Yuguo Liu, Wenyu Chen, Hanwen Liu, Yun Zhang, Malu Zhang, and Hong Qu. Biologically plausible sparse temporal word representations. IEEE Transactions on Neural Networks and Learning Systems, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.468, + 0.404 + ], + "angle": 0, + "content": "[31] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10(9):1659-1671, 1997." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.406, + 0.468, + 0.488 + ], + "angle": 0, + "content": "[32] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668-673, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.491, + 0.468, + 0.544 + ], + "angle": 0, + "content": "[33] Hesham Mostafa. Supervised learning based on temporal coding in spiking neural networks. IEEE transactions on neural networks and learning systems, 29(7):3227-3235, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.548, + 0.468, + 0.616 + ], + "angle": 0, + "content": "[34] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 36(6):51-63, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.619, + 0.468, + 0.687 + ], + "angle": 0, + "content": "[35] Zihan Pan, Malu Zhang, Jibin Wu, Jiadong Wang, and Haizhou Li. Multi-tone phase coding of interaural time difference for sound source localization with spiking neural networks. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:2656-2670, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.689, + 0.468, + 0.757 + ], + "angle": 0, + "content": "[36] Seongsik Park, Seijoon Kim, Hyeokjun Choe, and Sungroh Yoon. Fast and efficient information transmission with burst spikes in deep spiking neural networks. In Proceedings of the 56th Annual Design Automation Conference 2019, pages 1-6, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.76, + 0.468, + 0.815 + ], + "angle": 0, + "content": "[37] Seongsik Park, Seijoon Kim, Byunggook Na, and Sungroh Yoon. T2fsnn: Deep spiking neural networks with time-to-first-spike coding. In 2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[38] Seongsik Park and Sungroh Yoon. Training energy-efficient deep spiking neural networks with time-to-first-spike coding. arXiv preprint arXiv:2106.02568, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[39] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[40] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, and Michael Pfeiffer. Theory and tools for the conversion of analog to spiking convolutional neural networks. arXiv preprint arXiv:1612.04052, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[41] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, Michael Pfeiffer, and Shih-Chii Liu. Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in neuroscience, 11:682, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[42] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[43] V. Sze, Y. H. Chen, T. J. Yang, and J. S. Emer. Efficient processing of deep neural networks: A tutorial and survey. Proceedings of the IEEE, 105(12), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.336, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[44] Yuchen Wang, Malu Zhang, Yi Chen, and Hong Qu. Signed neuron with memory: Towards simple, accurate and high-efficient ann-snn conversion. In International Joint Conference on Artificial Intelligence, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.393, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[45] Jibin Wu, Yansong Chua, Malu Zhang, Guoqi Li, Haizhou Li, and Kay Chen Tan. A tandem learning rule for effective training and rapid inference of deep spiking neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[46] Jibin Wu, Chenglin Xu, Xiao Han, Daquan Zhou, Malu Zhang, Haizhou Li, and Kay Chen Tan. Progressive tandem learning for pattern recognition with deep spiking neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7824-7840, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[47] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[48] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 1311-1318, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.649, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[49] Qi Xu, Yaxin Li, Jiangrong Shen, Jian K Liu, Huajin Tang, and Gang Pan. Constructing deep spiking neural networks from artificial neural networks with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7886-7895, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[50] Qi Xu, Yaxin Li, Jiangrong Shen, Pingping Zhang, Jian K Liu, Huajin Tang, and Gang Pan. Hierarchical spiking-based model for efficient image classification with enhanced feature extraction and encoding. IEEE Transactions on Neural Networks and Learning Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[51] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Qiang Yu, Huajin Tang, Kay Chen Tan, and Haizhou Li. Rapid feedforward computation by temporal encoding and learning with spiking neurons. IEEE transactions on neural networks and learning systems, 24(10):1539-1552, 2013." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10561" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[53] Friedemann Zenke and Tim P Vogels. The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks. Neural computation, 33(4):899-925, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[54] Jilin Zhang, Dexuan Huo, Jian Zhang, Chunqi Qian, Qi Liu, Liyang Pan, Zhihua Wang, Ning Qiao, Kea-Tiong Tang, and Hong Chen. 22.6 anp-i: A 28nm 1.5 pj/sop asynchronous spiking neural network processor enabling sub-o. 1 \\(\\mu\\)j/sample on-chip learning for edge-ai applications. In 2023 IEEE International Solid-State Circuits Conference (ISSCC), pages 21–23. IEEE, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.468, + 0.327 + ], + "angle": 0, + "content": "[55] Jian Zhang, Jilin Zhang, Dexuan Huo, and Hong Chen. Anas: Asynchronous neuromorphic hardware architecture search based on a system-level simulator. in proc. 60th acm/ieee design automation conference (dac), 2023. In Proc. 60th ACM/IEEE Design Automation Conference (DAC), 2023, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.468, + 0.399 + ], + "angle": 0, + "content": "[56] Malu Zhang, Xiaoling Luo, Yi Chen, Jibin Wu, Ammar Belatreche, Zihan Pan, Hong Qu, and Haizhou Li. An efficient threshold-driven aggregate-label learning algorithm for multimodal information processing. IEEE Journal of Selected Topics in Signal Processing, 14(3):592-602, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.468, + 0.468 + ], + "angle": 0, + "content": "[57] Malu Zhang, Hong Qu, Ammar Belatreche, Yi Chen, and Zhang Yi. A highly effective and robust membrane potential-driven supervised learning method for spiking neurons. IEEE transactions on neural networks and learning systems, 30(1):123-137, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.468, + 0.565 + ], + "angle": 0, + "content": "[58] Malu Zhang, Jiadong Wang, Jibin Wu, Ammar Belatreche, Burin Amornpaisannon, Zhixuan Zhang, Venkata Pavan Kumar Miriyala, Hong Qu, Yansong Chua, Trevor E Carlson, et al. Rectified linear postsynaptic potential function for backpropagation in deep spiking neural networks. IEEE transactions on neural networks and learning systems, 33(5):1947-1958, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.468, + 0.621 + ], + "angle": 0, + "content": "[59] Wenrui Zhang and Peng Li. Temporal spike sequence learning via backpropagation for deep spiking neural networks. Advances in Neural Information Processing Systems, 33:12022-12033, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.624, + 0.468, + 0.692 + ], + "angle": 0, + "content": "[60] Shibo Zhou, Xiaohua Li, Ying Chen, Sanjeev T Chandrasekaran, and Arindam Sanyal. Temporal-coded deep spiking neural network with easy training and robust performance. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 11143-11151, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.693, + 0.468, + 0.817 + ], + "angle": 0, + "content": "[61] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, pages 3-11. Springer, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[62] Zongwei Zhou, Jae Shin, Lei Zhang, Suryakanth Gurudu, Michael Gotway, and Jianming Liang. Fine-tuning convolutional neural networks for biomedical image analysis: actively and incrementally. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7340-7351, 2017." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[63] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, YAN Shuicheng, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[64] Rui-Jie Zhu, Qihang Zhao, Tianjing Zhang, Haoyu Deng, Yule Duan, Malu Zhang, and Liang-Jian Deng. Tcjasnn: Temporal-channel joint attention for spiking neural networks. arXiv preprint arXiv:2206.10177, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[65] Yaoyu Zhu, Zhaofei Yu, Wei Fang, Xiaodong Xie, Tiejun Huang, and Timothée Masquelier. Training spiking neural networks with event-driven backpropagation. In 36th Conference on Neural Information Processing Systems (NeurIPS 2022), 2022." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10562" + } + ] +] \ No newline at end of file diff --git a/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/6f1f26ae-7b19-4337-978a-c5b559913901_origin.pdf b/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/6f1f26ae-7b19-4337-978a-c5b559913901_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..208be99047deca151397fab2071cf86e885beb5f --- /dev/null +++ b/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/6f1f26ae-7b19-4337-978a-c5b559913901_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa614e5bc87265e810b06e8134b38af53fed6357168cdab4e14997580e35de89 +size 2400733 diff --git a/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/full.md b/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dfc31408d5589ec91d7bcde82fa2e1ea173d6c30 --- /dev/null +++ b/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/full.md @@ -0,0 +1,343 @@ +# Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold: Learning with Event-Driven Backpropagation + +Wenjie Wei $^{1}$ , Malu Zhang $^{1*}$ , Hong Qu $^{1}$ , Ammar Belatreche $^{2}$ , Jian Zhang $^{3}$ , and Hong Chen $^{3}$ + +1University of Electronic Science and Technology of China, 2Northumbria University, 3Tsinghua University + +{maluzhang,hongqu}@uestc.edu.cn, ammar.belatreche@northumbria.ac.uk, hongchen@tsinghua.edu.cn + +# Abstract + +Spiking Neural Networks (SNNs) offer a highly promising computing paradigm due to their biological plausibility, exceptional spatiotemporal information processing capability and low power consumption. As a temporal encoding scheme for SNNs, Time-To-First-Spike (TTFS) encodes information using the timing of a single spike, which allows spiking neurons to transmit information through sparse spike trains and results in lower power consumption and higher computational efficiency compared to traditional rate-based encoding counterparts. However, despite the advantages of the TTFS encoding scheme, the effective and efficient training of TTFS-based deep SNNs remains a significant and open research problem. In this work, we first examine the factors underlying the limitations of applying existing TTFS-based learning algorithms to deep SNNs. Specifically, we investigate issues related to over-sparsity of spikes and the complexity of finding the 'causal set'. We then propose a simple yet efficient dynamic firing threshold (DFT) mechanism for spiking neurons to address these issues. Building upon the proposed DFT mechanism, we further introduce a novel direct training algorithm for TTFS-based deep SNNs, called DTA-TTFS. This method utilizes event-driven processing and spike timing to enable efficient learning of deep SNNs. The proposed training method was validated on the image classification task and experimental results clearly demonstrate that our proposed method achieves state-of-the-art accuracy in comparison to existing TTFS-based learning algorithms, while maintaining high levels of sparsity and energy efficiency on neuromorphic inference accelerator. + +# 1. Introduction + +Deep learning approaches have demonstrated remarkable performance in various intelligent applications [18, 27, 24, 23] and innovative simulations [62, 61]. However, these achievements come at a significant cost of energy consumption, severely limiting the deployment of artificial neural networks (ANNs) on resource-limited platforms [43, 9, 3, 50]. Brain-inspired spiking neural networks (SNNs) have emerged as a promising and energy-efficient alternative to ANNs [22, 31, 57]. SNNs have garnered attention for their distinctive features of sparse-asynchronous spikes and event-driven computing, which have played a pivotal role in driving the development of neuromorphic computing platforms such as ANP-I [54], TrueNorth [1], Loihi [8], and Tianjic [39]. + +Despite the inherent energy efficiency of SNNs, their extensive application to real-world practical scenarios has been limited by the absence of efficient and scalable training algorithms [52, 35, 30, 14, 20]. Traditional training techniques for ANNs, such as error backpropagation (BP) and standard GPU-accelerated training packages, cannot be leveraged directly by SNNs due to the complex temporal dynamics and the non-differentiability of spiking neurons' activity [33, 6, 56]. To address these challenges, various solutions have been proposed, broadly categorized into three distinct groups. + +The first category of solutions proposed to address the lack of scalable training algorithms for SNNs is ANN-to-SNN conversion method [40, 41, 49, 45]. These methods involve training an ANN first and then converting it to its SNN counterpart. The resulting SNN may achieve comparable performance to pre-trained ANNs through manual and careful setting of the SNN parameters such as the firing threshold [28] and the chosen spiking neuron model [44]. However, such conversion often results in precision loss due to approximation and requires longer time steps for lossless + +conversion [19]. Moreover, ANN-to-SNN conversion methods often encode the activation values of artificial neurons using the firing rate of spiking neurons [11, 4, 5], which fails to fully leverage the timing information of spikes. Furthermore, the runtime computation on neuromorphic hardware can often be highly energy-intensive. + +The second category of the proposed SNN training solutions is RNN-like method, commonly known as surrogate gradient learning [47, 48]. In these surrogate gradient-based SNNs, information is conveyed through spikes during feedforward computation, while during backpropagation, gradient information is calculated by treating the neurons' membrane potential as a differentiable signal at the current time step [42, 34]. To address the non-differentiability of spikes in the direct training of SNNs, surrogate functions are employed to approximate the derivative of the spike generation function during the backward process [53, 29]. However, while these methods have successfully tackled the challenges associated with the non-differentiability of spiking neurons, they necessitate updating synaptic weights at each time step, regardless of whether there is a spike emission or not [47, 15]. Consequently, this results in substantial energy and memory demands as large intermediate activation values necessitate storage to facilitate gradient computation. + +Unlike the above-mentioned methods, the third category of learning algorithms in SNNs is spike-driven method [2, 33, 58, 60, 38, 6], which operate in a strictly event-driven manner during both forward and backward computations. SpikeProp [2] and its various extensions [33, 6, 25] are typical examples of this category. However, these algorithms remain limited to shallow network structures and relatively simple classification tasks. To overcome this limitation, Zhang et al. [58] and Zhou et al. [60] extended the spikedriven algorithm to more complex network structures. Despite achieving competitive accuracy on large datasets [60], the identification of the 'causal set' in this approach requires substantial computational resources. To improve the training efficiency, Park et al. [38] proposed a surrogate deep neural network (DNN) model that is subsequently converted into a TTFS-based SNN. However, this conversion method is prone to conversion errors, potentially impairing the temporal learning capability of SNNs. Recently, several multi-spike-based spike-driven algorithms have been developed [59, 65]. While Zhang et al. [59] and Zhu et al. [65] achieve competitive performance on CIFAR utilizing the rate coding scheme, we focus on developing an energy-efficient TTFS-based spike-driven algorithm. + +In this paper, we introduce an effective and efficient spike-driven learning algorithm that facilitates the training + +of high-performance TTFS-based deep SNNs. The main contributions of this work are outlined as follows: + +- We provide a comprehensive analysis of the main shortcomings of existing methods in achieving high performance in TTFS-based deep SNNs. Specifically, we examine issues such as the over-sparsity of spikes and the complexity of determining the 'causal set'. +- We propose a simple yet efficient dynamic firing threshold (DFT) mechanism for spiking neurons that can effectively address the aforementioned issues. +- Building upon the proposed DFT, we further introduce a novel direct training algorithm for TTFS-based deep SNNs, which we refer to as DTA-TTFS. In this training method, the timing of a single spike is considered the basic information carrier and the learning process is performed strictly in an event-driven manner. +- The proposed method is validated through extensive experiments on benchmark image classification tasks and achieves state-of-the-art accuracy compared to existing TTFS-based learning algorithms. Furthermore, we demonstrate the ultra-low power capability of the SNN with DTA-TTFS on a recently developed neuromorphic accelerator. + +# 2. Preliminaries and Problem Analysis + +In this section, we first briefly review the TTFS coding and the employed spiking neuron model. Subsequently, we provide a comprehensive analysis of the primary shortcomings in existing TTFS-based deep SNN methods. + +# 2.1. Preliminaries + +TTFS coding scheme: There are two primary coding schemes utilized to represent information in SNNs: rate coding and temporal coding. The rate coding method represents information based on the average firing rate of a spiking neuron. However, this encoding scheme neglects the temporal information carried by spikes, leading to higher spike frequency and increased energy consumption. Conversely, temporal coding leverages the precise timing of spikes to encode information, resulting in lower power consumption and higher computational efficiency compared to the rate-based coding method. + +As a temporal coding method, TTFS encodes information by the timing of the first spike. As depicted in Fig. 1, the higher the activation value, the earlier the first spike is emitted. The relationship between the input activation value $a_{i}$ and the encoded spike time $t_i$ can be expressed as + +$$ +t _ {i} = \left(1 - \frac {a _ {i}}{a _ {\text {m a x}}}\right) \times T _ {w}, \tag {1} +$$ + +![](images/2c6d616ed5816f58b20b1c4c29eb8d0000ecdf289ee7efde9b362c28d5c5ef92.jpg) +(a) + +![](images/3c7b1045b9577e1c511f3fe5260a99ef67a796ad09755afd0d1ba86d2ca99205.jpg) +(b) +Figure 1. (a) TTFS coding scheme. The larger intensity value corresponds to the earlier spike. (b) ReL-PSP spiking neuron model. Neuron $j$ accumulates the linear PSP and fires a spike when the membrane potential crosses the firing threshold $\vartheta$ . + +where $a_{max}$ is the maximum input activation value, $T_w$ is the length of the permissible spike time window. As the TTFS coding scheme requires only a single spike to convey information, it effectively diminishes the number of spikes and significantly decreases the energy requirements. Consequently, it achieves higher sparsity and energy efficiency, making it a powerful coding scheme that maximizes the energy efficiency of SNNs. + +Spiking neuron model: The Rectified Linear Postsynaptic Potential (ReL-PSP) spiking neuron [58] is chosen for our method due to its simplicity and energy efficiency. As shown in Fig. 1(b), the ReL-PSP neuron model utilizes a rectified linear kernel function to shape the neuron's PSP, and its dynamics is defined as + +$$ +V _ {j} ^ {l} (t) = \sum_ {i = 1} ^ {N} w _ {i j} ^ {l} K \left(t - t _ {i} ^ {l - 1}\right), \tag {2} +$$ + +where $V_{j}^{l}(t)$ is the membrane potential of neuron $j$ in layer $l$ , and $w_{ij}^{l}$ is the synaptic weight between neuron $j$ and its presynaptic neuron $i$ . $N$ is the number of neurons in the previous layer $l - 1$ , $t_{i}^{l - 1}$ is the spike time of input neuron $i$ , and $K\left(t - t_{i}^{l - 1}\right)$ is the PSP function that is defined as + +$$ +K \left(t - t _ {i} ^ {l - 1}\right) = \left\{ \begin{array}{l l} t - t _ {i} ^ {l - 1}, & \text {i f} \quad t > t _ {i} ^ {l - 1}, \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {3} +$$ + +The neuron $j$ emits a spike when $V_{j}^{l}(t)$ reaches the firing threshold $\vartheta$ . Mathematically, the spike generation function $F$ is defined as + +$$ +t _ {j} ^ {l} = F \{t | V _ {j} ^ {l} (t) \geq \vartheta , t \geq 0 \}. \tag {4} +$$ + +The ReL-PSP neuron has proven its efficacy in tackling the main challenges encountered during the training of deep SNNs, such as the non-differentiability of the spike function, gradient explosion, and the dead neuron issue. Despite the successful performance on the MNIST dataset, its performance is not ideal on deeper network structures and more complex datasets, such as CIFAR-10 and CIFAR-100. + +In the subsequent section, we will conduct a comprehensive analysis to identify the underlying reasons for this limitation. + +# 2.2. Problem Analysis + +Over-sparsity of spikes: Sparse neural networks offer several advantages such as efficient variable-size representation, information decomposition, and energy efficiency. However, over-sparsity may diminish the model's effective capacity and adversely affect its performance [17]. This issue becomes more serious in TTFS-based deep SNNs, owing to their firing mechanism, asynchronous transmission, and TTFS-based decision strategy. + +![](images/64ae6d96b091219cd27021032a4b4cf4a6b1e7cbb361c424d6de03cab5378082.jpg) +(a) +Figure 2. The over-sparsity problem caused by asynchronous transmission and TTFS-based decision strategy. (a) The presynaptic (gray) neuron fires later than the postsynaptic (blue) neuron $j$ , making it ineffective in contributing to the firing of the postsynaptic neuron. As a result, this presynaptic neuron becomes invalid. (b) After learning, the postsynaptic (blue) neuron fires at an earlier time, leading to the appearance of additional invalid neurons. + +![](images/ca111c35d49f8fac874315a4faf27d2eab53d643ab7c43b09565d777bf2422fc.jpg) +(b) + +According to the firing mechanism of spiking neurons, a neuron emits one spike only when its membrane potential reaches the firing threshold. This leads to fewer activated spiking neurons compared to ANNs. In addition, the asynchronous transmission of SNNs further diminishes the number of valid activated neurons. As shown in Fig. 2(a), when the presynaptic neuron $i$ fires after the postsynaptic neuron $j$ , the spike of neuron $i$ becomes invalid for neuron $j$ . Furthermore, TTFS-based SNNs make decisions based on the earliest spike, and the synaptic weights are trained to make the target neuron fire as early as possible. As shown in Fig. 2(b), after learning, the firing time of the postsynaptic neuron $j$ may become earlier, thereby further exacerbating the sparsity induced by asynchronous transmission. + +In event-driven learning algorithms of SNNs, error backpropagation depends on the timing of spikes. Therefore, if a spiking neuron does not fire or does not contribute to the postsynaptic neuron, it will not participate in learning. The issue of over-sparsity not only limits the information representation capability of SNNs but also leads to learning failure. Consequently, the achievement of a harmonious balance between sparsity and performance in the training of TTFS-based deep SNNs remains an open question. + +The complexity of finding the 'causal set': Due to the sparsity and asynchronicity of SNNs, not all presynaptic neurons $\{i|i = 1,\dots ,N\}$ contribute to their postsynaptic neuron $j$ , where the contributory neurons are defined as the 'causal set' of the postsynaptic neuron $j$ , denoted as $C_j = \{i|t_i < t_j\}$ [33, 60]. To leverage standard GPU-accelerated ANN training packages, it is crucial to identify the 'causal set' of each neuron during the feed-forward computation process [33, 60]. However, this process is highly complex and time-consuming. For example, in order to find the 'causal set', Zhou et al. [60] proposed an approach where they first sort all of this neuron input spikes $\{t_i|i = 1,\dots N\}$ in ascending order and subsequently consider each individual spike $t_i$ for inclusion in the causal set. This approach, however, is computationally expensive and results in inefficient training even with the implementation of a space-for-time strategy [60]. + +Despite the success of algorithms with this approach in utilizing standard GPU-accelerated ANN training packages, the training process involves a significant amount of sorting and iteration. This results in higher demand for training resources and increased latency, especially as the network depth increases. Consequently, the process of finding the 'causal set' may become a bottleneck for training deep SNNs. In the next section, we propose a dynamic firing threshold mechanism for spiking neurons to avoid this resource-intensive process. + +# 3. Method + +In this section, we first introduce the dynamic firing threshold (DFT) for spiking neurons and discuss how it effectively addresses the above issues. Then, based on the DFT, we further propose a direct training algorithm for TTFS-based SNN, referred to as DTA-TTFS. + +# 3.1. DFT-based Spiking Neuron Model + +As demonstrated in Fig. 3(a), the proposed DFT is a layer-dependent and time-varying function, which is defined as + +$$ +\vartheta^ {l} (t) = \left\{ \begin{array}{l l} T _ {w} (l + 1) - t, & \text {i f} \quad T _ {w} l \leq t \leq T _ {w} (l + 1), \\ + \infty , & \text {o t h e r w i s e}, \end{array} \right. \tag {5} +$$ + +where $\vartheta^l (t)$ represents the firing threshold of the neurons in the $l$ -th layer and $T_{w}$ denotes the length of the permissible spike time window. In this work, $T_{w}$ is set to 1 (unless otherwise stated) and the input layer is regarded as the 0-th layer. Eq. 5 reveals that the firing threshold is infinite outside the interval $t\in [T_wl,T_w(l + 1)]$ , and neurons in the $l$ -th layer can only generate a spike within this time interval. Therefore, as shown in Fig. 4, the generated spikes are non-overlapping across different layers, which we refer to as the layer-dependent firing feature. + +Based on the membrane potential dynamics in Eq. 2 and the proposed DFT in Eq. 5, the output of neuron $j$ can be divided into three categories: + +- Spikes at $t_j^l = T_{wl}$ : Since the firing threshold is infinity when $t < T_{wl}$ , the neuron cannot fire a spike before $t = T_{wl}$ even though the membrane potential is strong (i.e., $V_j^l(t) \geq \vartheta^l(T_{wl})$ and $t < T_{wl}$ ). As shown in Fig. 3(b), the earliest spike time for DFT-based spiking neurons in the $l$ -th layer is $t_j^l = T_{wl}$ . +- Spikes within $(T_{w}l, T_{w}(l + 1))$ : The neuron generates a spike when its membrane potential exceeds the threshold, i.e. $V_{j}^{l}(t) \geq \vartheta^{l}(t)$ . The firing threshold $\vartheta^{l}$ in this time interval is a linear function, which decays linearly from 1 to 0. As depicted in Fig. 3(c), by combining the membrane potential in Eq. 2 and the firing threshold in Eq. 5, the precise spiking time $t_{j}^{l}$ can be calculated as + +$$ +t _ {j} ^ {l} = \frac {T _ {w} (l + 1) + \sum_ {i \in C _ {j}} w _ {i j} ^ {l} t _ {i} ^ {l - 1}}{1 + \sum_ {i \in C _ {j}} w _ {i j} ^ {l}}. \tag {6} +$$ + +- Nonspike: The neuron will not fire a spike if its membrane potential is below the dynamic firing threshold (i.e., $V_{j}^{l}(t) < \vartheta^{l}(t)$ ). Fig. 3(d) shows three typical scenarios of nonspike. + +Overall, the spike time of a DFT-based neuron $j$ in the $l$ -th layer can be represented as + +$$ +t _ {j} ^ {l} = \left\{ \begin{array}{l l} T _ {w} l, & \text {i f} \quad V _ {j} ^ {l} \left(T _ {w} l\right) \geq \vartheta^ {l} \left(T _ {w} l\right), \\ \frac {T _ {w} (l + 1) + \sum_ {i \in C _ {j}} w _ {i j} ^ {l} t _ {i} ^ {l - 1}}{1 + \sum_ {i \in C _ {j}} w _ {i j} ^ {l}}, & \text {i f} \quad T _ {w} l < t _ {j} ^ {l} \leq T _ {w} (l + 1), \\ \text {n o n s p i k e}, & \text {o t h e r w i s e .} \end{array} \right. \tag {7} +$$ + +In the following, we will analyze how the proposed DFT effectively addresses issues of over-sparsity of spikes and the complexity of finding the 'causal set'. + +1) Over-sparsity of spikes: As described in Eq. 5, the DFT is a piecewise function, whose linearly decreased part and layer-dependent firing feature effectively address the over-sparsity of spikes. + +The linearly decreased part of the DFT is purposefully designed to tackle the over-sparsity arising from the spiking neuron firing mechanism. By gradually decreasing the firing threshold from 1 to 0, the neuron $j$ remains active as long as its membrane potential is non-negative at $T_{w}(l + 1)$ . As a result, the proposed DFT effectively reduces the number of inactive neurons. Moreover, the linear decay design enhances the discriminative nature of information from active neurons. + +![](images/fd574486f30506d416cadb62ae3654d1cc072c345b6c02a4414e5a44c49e9808.jpg) +(a) + +![](images/41bdde42027fe031e85074cac6e17050278ef41be38365c8fa51d1ad476e9245.jpg) +(b) +Figure 3. Proposed dynamic firing threshold and output of the DFT-based spiking neurons. (a) Dynamic firing threshold function(DFT). (b) Spikes at $t_j^l = T_w l$ . (c) Spikes within $(T_w l, T_w (l + 1)]$ . (d) Nonspike. This case is further divided into three sub-cases, which are shown by the curves of different colors. + +![](images/91f457e988f999826c4da082ffc2f2be246ea95f46742ab2c645c0e015ae782b.jpg) +(c) + +![](images/133938a2ea83c75d7176e3ad1904a5a37ff3d5db422aa2dfa2d77e561d052ba8.jpg) +(d) + +![](images/a954087f8b2fee71906adadcf04891013d7f9a7db6a9add4ef3c32991f1ebf05.jpg) +Figure 4. The propagation pipeline of the SNN with DFT. + +The layer-dependent firing feature enables SNN with non-overlapping spike time windows. This effectively resolves the over-sparsity arising from asynchronous transmission and TTFS-based decision strategy. Fig. 4 illustrates that the spike time window of postsynaptic neurons occurs later than that of presynaptic neurons. Consequently, when the presynaptic neuron fires a spike, its postsynaptic neuron absolutely receives it. Furthermore, the non-overlapping spike window provides output layer neurons with a predetermined earliest spike time, preventing them from being forced forward indefinitely during the training process. + +2) The complexity of finding the 'causal set': The layer-dependent firing feature of DFT circumvents the need for SNNs to identify the 'causal set', resulting in efficient training. This efficiency is further improved by the output analytic equation of the DFT-based neuron. + +As depicted in Fig. 4, the layer-dependent firing feature ensures that in the $l$ -th layer of the SNN utilizing DFT, the spike time is controlled within the interval $[T_w l, T_w (l + 1)]$ . Therefore, we can easily learn that neurons in the $(l - 1)$ -th layer must fire earlier than neurons in the $l$ -th layer. In other words, all active presynaptic neurons must remain in the 'causal set' of their postsynaptic neurons. Therefore, the DFT mechanism prevents the need for resource-intensive sorting and iteration process, enhancing training efficiency and facilitating its extension to deeper network architecture. + +In addition, the employed ReL-PSP spiking neuron and the DFT provide an analytical relation between input spikes and output spikes, as described in Eq.7. Therefore, the dynamics of SNNs within the training loop need not be simulated using long time steps. This allows the utilization + +of standard GPU-accelerated training packages of ANNs, which further leads to an efficient training process. + +# 3.2. DTA-TTFS Learning Algorithm + +Consider a task with $n$ categories, where each output neuron corresponds to a specific category. In this task, the target category is represented by the output neuron that fires the earliest. To assess the effectiveness of the model, we adopt the widely utilized cross-entropy loss function $L$ , which aims to maximize the output value of the target neuron. In order to ensure the desired neuron fires earliest and others fire as late as possible, we utilize the negative spike times of the output layer [58, 6], i.e., $-t$ , for evaluation. Therefore, the loss function is given by + +$$ +L (\mathbf {t}, d) = - \log \frac {\exp (- \mathbf {t} [ d ])}{\sum_ {i = 1} ^ {n} \exp (- \mathbf {t} [ i ])}, \tag {8} +$$ + +where $\mathbf{t}$ is the vector of spike times in the output layer, and $d$ is the desired category index. + +In order to minimize the $L(\mathbf{t},d)$ , the DTA-TTFS algorithm adjusts spike times by modifying the synaptic weights during gradient backpropagation. For driving the gradient backpropagation, it is necessary to calculate derivatives of the output spike time $t_j^l$ with respect to the synaptic weight $w_{ij}^{l}$ and the input spike time $t_i^{l - 1}$ . According to Eq. 7, we can easily get these two derivatives, + +$$ +\frac {\partial t _ {j} ^ {l}}{\partial w _ {i j} ^ {l}} = \left\{ \begin{array}{l l} \frac {t _ {i} ^ {l - 1} - t _ {j} ^ {l}}{1 + \sum_ {i \in C _ {j}} w _ {i j} ^ {l}}, & \text {i f} \quad T _ {w} l < t _ {j} ^ {l} < T _ {w} (l + 1), \\ 0, & \text {o t h e r w i s e}, \end{array} \right. \tag {9} +$$ + +$$ +\frac {\partial t _ {j} ^ {l}}{\partial t _ {i} ^ {l - 1}} = \left\{ \begin{array}{l l} \frac {w _ {i j} ^ {l}}{1 + \sum_ {i \in C _ {j}} w _ {i j} ^ {l}}, & \text {i f} \quad T _ {w} l < t _ {j} ^ {l} < T _ {w} (l + 1), \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {10} +$$ + +Following Eq. 9 and Eq. 10, gradient backpropagation can be directly applied. + +
DatasetsModelsNetwork ArchitectureNeural CodingMethodAccuracySparsity
MNISTMostafa 2017 [33]MLP¹TFTSDT97.55%0.51
Zhang et al. 2020 [58]CNN¹TFTSDT99.4%0.6614
Zhou et al. 2021 [60]CNN²TFTSDT99.33%0.94
DTA-TTFSCNN¹TFTSDT99.4%0.3913
CIFAR-10Wu et al. 2022 [46]VGG11Rateconv91.24%no
Park et al. 2020 [37]VGG16TFTSconv91.43%0.2459
Zhou et al. 2021 [60]VGG16TFTSDT92.68%0.62
Park et al. 2021 [38]VGG16TFTSDT91.90%0.1746
DTA-TTFSVGG11TFTSDT91.17%0.4387
DTA-TTFSVGG16TFTSDT93.05%0.2561
CIFAR-100Park et al. 2020 [37]VGG16TFTSconv68.79%0.2994
Park et al. 2021 [38]VGG16TFTSDT65.98%0.2780
DTA-TTFSVGG16TFTSDT69.66%0.2845
+ +Table 1. Comparison of accuracy and sparsity between the DTA-TTFS method and other related works. DT: direct training, conv: ANN-to-SNN conversion. MLP1: 784-800-10. CNN1: 784-16C5-P2-32C5-P2-800-128-10. CNN2: 784-32C5-16C5-10. + +# 4. Experiments + +In this section, we first evaluate the performance of the DTA-TTFS algorithm. Then, we carry out validation studies to illustrate the impact of the DFT. Finally, we employ theoretical analysis along with the neuromorphic hardware simulator to evaluate the energy consumption during the inference process. + +# 4.1. Details + +We conduct experiments on three image classification benchmark datasets: MNIST $^2$ , CIFAR-10, and CIFAR $100^3$ . All experiments are conducted using the PyTorch library, which allows for efficient training on multi-GPU machines with accelerated computations and optimized memory usage. For the MNIST dataset, we employ a network structure with 784-16C5-P2-32C5-P2-800-128-10 and train it for 150 epochs. As for CIFAR datasets, we adopt augmentation techniques [7, 10] and train the VGG11 and VGG16 network structures without Batch Normalization (BN) for 300 epochs. Moreover, in order to maintain the initial parameter domain at a good level, the SNN is initialized with the parameters from the pre-trained identical ANN when training CIFAR datasets. During the training process, we adopt the Adam optimizer and implement milestones learning rate decay, with an initial learning rate $1e^{-3}$ for MNIST and $1e^{-4}$ for CIFAR datasets. + +# 4.2. Comparison with Related Works + +Tab. 1 presents a comparison of the accuracies and sparsity achieved by DTA-TTFS and other related works. Experimental results demonstrate that the DTA-TTFS algo + +rithm exhibits excellent performance, with an accuracy of $99.4\%$ on MNIST, $93.05\%$ on CIFAR-10, and $69.66\%$ on CIFAR-100. Obviously, our method outperforms all previous direct learning algorithms of TTFS-based SNNs. Moreover, our method still maintains competitive sparsity when achieving high performance. The sparsity is calculated by $S / N$ averaged on the entire testing set, where $S$ represents the total number of spikes and $N$ denotes the total number of neurons. Specifically, the DTA-TTFS achieves sparsity of 0.3913 on MNIST, 0.2561 on CIFAR-10, and 0.2845 on CIFAR-100. For the MNIST dataset, when employing the identical network structure as described in [58], our method achieves the same accuracy while significantly reducing the sparsity from 0.6614 to 0.3913. For CIFAR datasets, our method surpasses related works in terms of accuracy while maintaining a competitive sparsity. In summary, the DTA-TTFS algorithm achieves a harmonious balance between accuracy and sparsity. + +# 4.3. Validation Study + +We conduct a series of validation experiments to confirm the functionality of the proposed DFT. These experiments aim to display two aspects. Firstly, the DFT provides SNNs with the layer-dependent firing feature. Additionally, it effectively addresses the over-sparsity of spikes. Validation experiments choose the well-trained VGG11 structure on the CIFAR-10 dataset. + +In order to show the layer-dependent firing feature of the SNN with DFT, we record the spike activity in each layer for the first sample of the testing set. For the purpose of comparison, we also provide the spike activity of the SNN without DFT. The results are visualized in Fig. 5, elucidating the evident regulation of spike activity in each layer + +![](images/b460326f15f624006660d7611722d9c9bd15fe4c2efe06e757da78131de887b8.jpg) + +![](images/765d6467c54f485876cd3f38e9fffb4844a792cbd0afc4092506dc4dd11c8723.jpg) +Figure 5. Spike activity in each layer of SNN with and without DFT. The vertical axis represents the neuron index scaled to [0,1]. + +within a permitted time window for the SNN incorporating DFT. In contrast, the spike activity in the SNN without DFT exhibits a disorderly and chaotic pattern. + +In order to verify the efficacy of the DFT in addressing the over-sparsity of spikes, we count the number of inactive neurons per layer averaged on the whole testing set. This experiment is conducted on three models: the ANN, the SNN with DFT, and the SNN without DFT [58]. As shown in Fig. 6, the SNN without DFT displays significant sparsity, however, with only $10\%$ accuracy. Conversely, the SNN with DFT exhibits a comparable amount of inactive neurons to that of ANN. As a result, the DFT achieves a harmonious trade-off between sparse representation and information representation ability. + +# 4.4. Energy Estimation + +# 4.4.1 Theoretical Analysis + +Energy consumption is a critical metric to evaluate the efficiency of SNNs. To prove the energy efficiency of the proposed DTA-TTFS, we employ the theoretical analysis presented in [37]. This method analyzes energy consumption based on multiple factors, including latency, spike counts, as well as data related to neuromorphic architecture. More specifically, the estimated energy is defined as + +$$ +E _ {\text {T o t a l}} = E _ {\text {S t a t i c}} \times \text {L a t e n c y} + E _ {\text {D y n a m i c}} \times \text {S p i k e s}, \tag {11} +$$ + +![](images/9ff63c4767d4b2ffdcdbefa9152c1788863df93f74c0ee37c4aa662e05619d35.jpg) +Figure 6. Percentage of inactive neurons per layer of the ANN, the SNN with DFT, and the SNN without DFT. + +where Latency is the required time steps, Spikes is the total number of spikes, $E_{Static}$ and $E_{Dynamic}$ are static and dynamic energy coefficients depending on the neuromorphic architecture. We choose two neuromorphic architectures, TrueNorth [32] and SpiNNaker [16], to conduct analysis. The energy parameters $(E_{Static}, E_{Dynamic})$ are set to (0.6, 0.4) for TrueNorth and (0.36, 0.64) for SpiNNaker. For comparison, we analyze various coding schemes, such as rate coding [19], phase coding [26], burst coding [36], and TTFS coding [37, 38]. Theoretical analysis is performed on CIFAR-10 with the well-trained VGG16 structure, and we normalized estimated energy consumption based on the rate coding. As indicated in Tab. 2, under the same conditions, the DTA-TTFS outperforms other methods in terms of energy efficiency. It is worth noting that, despite the fewer spikes achieved by [37, 38], their energy consumption remains higher than ours due to their demand for long latency. + +
Neural CodingTime StepSpike (106)Acc. (%)Normalized Energy
TrueNorthSpiNNaker
Rate [19]5122.61293.3911
Phase [26]150035.19691.217.14769.6785
Burst [36]11256.9291.412.37812.4865
TTFS [37]6800.06991.430.80740.4950
TTFS [38]5440.06791.90.64780.3989
DTA-TTFS1600.07393.050.19870.1304
+ +Table 2. Theoretical analysis of energy consumption. + +# 4.4.2 Hardware Simulation + +In order to further verify the energy efficiency and hardware friendliness of our method, we map the SNN trained by the DTA-TTFS algorithm on a recently developed neuromorphic hardware simulator, called configurable asynchronous neuromorphic hardware simulator (CanMore) [55]. CanMore estimates hardware performance through system-level modeling and simulation. Fig. 7 illustrates the main + +![](images/5274cbf7e6d8854be059704150bc939ca36f94baf27acc47f8c38e3507dca2c4.jpg) +Figure 7. The framework of CanMore [55], including neuromorphic cores comprised of neurons and SRAM, and NoC routers consisting of buffers, route computation units, switch allocator, and crossbar switch. + +framework of the CanMore, including neuromorphic cores comprised of neurons and SRAM, and network-on-chip (NoC) routers consisting of buffers, route computation unit, switch allocator and crossbar switch. With the system-level asynchronous circuit modeling, CanMore is able to simulate various neuromorphic hardware architectures by changing the parameters and configuration files, estimate latency and energy consumption, and analyze spike messages routing information such as routing hops, congestion latency, routing directions, and so on. + +In [13], a deep ANN is mapped to TrueNorth neuromorphic hardware. To the best of our knowledge, this is the only work that verifies the energy efficiency of SNNs on the CIFAR-10 dataset, while the majority of existing works focus on smaller datasets, such as MNIST [12, 58]. Here, we train an SNN on the CIFAR-10 dataset with the proposed DTA-TTFS method. After learning, the learned weights are transferred to CanMore for accelerating the inference operations. Specifically, our network structure is mapped by partitioning each layer into equally sized groups along the feature dimension, with each group being mapped to a neuromorphic core with a maximum of 1024 neurons. In addition, we compress the network by quantizing the weights to 4 bits. The mesh topology and XY routing algorithm are adopted in the simulated architecture. The energy consumption per frame during the inference process is calculated by + +$$ +E _ {T o t a l} = E _ {S O P} \times \text {S p i k e s} + E _ {H o p} \times \text {H o p s} + E _ {\text {S t a t i c}}, \tag {12} +$$ + +where $E_{SOP}$ is the dynamic energy consumption per synaptic operation (SOP), which is calculated with the real technology parameters of the CMOS 22nm process in the simulation. In our work, $E_{SOP}$ is $20\mathrm{pJ}$ . $E_{Hop}$ is the dynamic energy consumption per hop between the mesh routers on the NoC, which is $100\mathrm{pJ}$ according to [21]. $E_{Static}$ is static energy consumption, which is $3.52\mu \mathrm{J}/\mathrm{frame}$ in CMOS 22nm process. The average Spikes and Hops per frame during the inference process are 71M and 45.4M re + +![](images/30ea7d1644ab9eca5b950198984c2e5bbe5fab515b084f649c5cfa991da6616a.jpg) +Figure 8. The accuracy and energy consumption of Esser et al. [13] and DTA-TTFS on CIFAR-10 dataset. + +![](images/e88121b7bffd7f5e17f2a83957cdeecfd52034d77883c573bf47f33d6df05453.jpg) + +spectively, given by CanMore. Therefore, we obtain the energy consumption $E_{Total}$ during the inference process $50.36\mu \mathrm{J / frame}$ . + +The accuracy and energy consumption are reported in Fig. 8. our work achieves a classification accuracy of $90.67\%$ on CIFAR-10, surpassing the accuracy reported in [13] by $7.26\%$ . In addition, the energy consumption of our work is $50.36~\mu \mathrm{J} / \mathrm{frame}$ , which is $69.2\%$ reduced compared with the energy consumption $163.65\mu \mathrm{J} / \mathrm{frame}$ in [13]. These results confirm that the proposed DTA-TTFS algorithm is effective and energy-efficiency. + +# 5. Conclusion + +In this study, we analyze the shortcomings of existing methods in achieving high-performance TTFS-based deep SNNs. In order to overcome these issues, we propose an efficient DFT mechanism for spiking neurons and introduce the DTA-TTFS algorithm for TTFS-based SNNs. We evaluate our method on the image classification task, and the results demonstrate that our method achieves a fine trade-off between accuracy and sparsity. Additionally, we estimate the energy consumption of our method through theoretical analysis and the neuromorphic hardware accelerator CanMore. Both theoretical analysis and hardware verification confirm the energy efficiency of the DTA-TTFS algorithm. Overall, our method not only attains high performance but also is quite hardware friendly. In our future work, we intend to leverage DTA-TTFS to train more advanced network models, including SNN-based Attention [64, 51] and Transformer architectures [63]. + +# Acknowledgment + +This work was supported by the National Science Foundation of China under Grants 62236007 and 61976043 and 62106038, and in part by the Sichuan Science and Technology Program under Grants 2022YFG0313 and 2023YFG0259. + +# References + +[1] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34(10):1537-1557, 2015. +[2] Sander M Bohte, Joost N Kok, and Johannes A La Poutre. Spikeprop: backpropagation for networks of spiking neurons. In *ESANN*, volume 48, pages 419–424. Bruges, 2000. +[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. +[4] Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Optimized potential initialization for low-latency spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11-20, 2022. +[5] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347, 2023. +[6] Iulia-Maria Comsa, Krzysztof Potempa, Luca Versari, Thomas Fischbacher, Andrea Gesmundo, and Jyrki Alakuijala. Temporal coding in spiking neural networks with alpha synaptic function: learning with backpropagation. IEEE transactions on neural networks and learning systems, 33(10):5939-5952, 2021. +[7] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019. +[8] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 38(1):82-99, 2018. +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. +[10] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. +[11] Peter U Diehl, Daniel Neil, Jonathan Binas, Matthew Cook, Shih-Chii Liu, and Michael Pfeiffer. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2015. +[12] Steve K Esser, Rathinakumar Appuswamy, Paul Merolla, John V Arthur, and Dharmendra S Modha. Backpropagation for energy-efficient neuromorphic computing. Advances in neural information processing systems, 28, 2015. + +[13] Steven K Esser, Paul A Merolla, John V Arthur, Andrew S Cassidy, Rathinakumar Appuswamy, Alexander Andreopoulos, David J Berg, Jeffrey L McKinstry, Timothy Melano, Davis R Barch, et al. From the cover: Convolutional networks for fast, energy-efficient neuromorphic computing. Proceedings of the National Academy of Sciences of the United States of America, 113(41):11441, 2016. +[14] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. +[15] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothée Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2661-2671, 2021. +[16] Steve B Furber, Francesco Galluppi, Steve Temple, and Luis A Plana. The spinnaker project. Proceedings of the IEEE, 102(5):652-665, 2014. +[17] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 315-323. JMLR Workshop and Conference Proceedings, 2011. +[18] Alex Graves and Navdeep Jaitly. Towards end-to-end speech recognition with recurrent neural networks. In International conference on machine learning, pages 1764-1772. PMLR, 2014. +[19] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13558–13567, 2020. +[20] Yangfan Hu, Huajin Tang, and Gang Pan. Spiking deep residual networks. IEEE Transactions on Neural Networks and Learning Systems, 2021. +[21] Masashi Imai, Thiem Van Chu, Kenji Kise, and Tomohiro Yoneda. The synchronous vs. asynchronous noc routers: an apple-to/apple comparison between synchronous and transition signaling asynchronous designs. In 2016 Tenth IEEE/ACM International Symposium on Networks-on-Chip (NOCS), pages 1-8. IEEE, 2016. +[22] Eugene M Izhikevich. Simple model of spiking neurons. IEEE Transactions on neural networks, 14(6):1569-1572, 2003. +[23] Yeying Jin, Aashish Sharma, and Robby T Tan. Dshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5027-5036, 2021. +[24] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In European Conference on Computer Vision, pages 404-421. Springer, 2022. +[25] S. R. Kheradpisheh and T. Masquelier. Temporal backpropagation for spiking neural networks with one spike per neuron. International Journal of Neural Systems, 30(6), 2020. + +[26] Jaehyun Kim, Heesu Kim, Subin Huh, Jinho Lee, and Kiyoung Choi. Deep neural networks with weighted spikes. Neurocomputing, 311:373-386, 2018. +[27] Jinyu Li et al. Recent advances in end-to-end automatic speech recognition. APSIPA Transactions on Signal and Information Processing, 11(1), 2022. +[28] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International conference on machine learning, pages 6316-6325. PMLR, 2021. +[29] Yuhang Li, Yufei Guo, Shanghang Zhang, Shikuang Deng, Yongqing Hai, and Shi Gu. Differentiable spike: Rethinking gradient-descent for training spiking neural networks. Advances in Neural Information Processing Systems, 34:23426-23439, 2021. +[30] Yuguo Liu, Wenyu Chen, Hanwen Liu, Yun Zhang, Malu Zhang, and Hong Qu. Biologically plausible sparse temporal word representations. IEEE Transactions on Neural Networks and Learning Systems, 2023. +[31] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10(9):1659-1671, 1997. +[32] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668-673, 2014. +[33] Hesham Mostafa. Supervised learning based on temporal coding in spiking neural networks. IEEE transactions on neural networks and learning systems, 29(7):3227-3235, 2017. +[34] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 36(6):51-63, 2019. +[35] Zihan Pan, Malu Zhang, Jibin Wu, Jiadong Wang, and Haizhou Li. Multi-tone phase coding of interaural time difference for sound source localization with spiking neural networks. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:2656-2670, 2021. +[36] Seongsik Park, Seijoon Kim, Hyeokjun Choe, and Sungroh Yoon. Fast and efficient information transmission with burst spikes in deep spiking neural networks. In Proceedings of the 56th Annual Design Automation Conference 2019, pages 1-6, 2019. +[37] Seongsik Park, Seijoon Kim, Byunggook Na, and Sungroh Yoon. T2fsnn: Deep spiking neural networks with time-to-first-spike coding. In 2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2020. +[38] Seongsik Park and Sungroh Yoon. Training energy-efficient deep spiking neural networks with time-to-first-spike coding. arXiv preprint arXiv:2106.02568, 2021. +[39] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with + +hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019. +[40] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, and Michael Pfeiffer. Theory and tools for the conversion of analog to spiking convolutional neural networks. arXiv preprint arXiv:1612.04052, 2016. +[41] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, Michael Pfeiffer, and Shih-Chii Liu. Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in neuroscience, 11:682, 2017. +[42] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in neural information processing systems, 31, 2018. +[43] V. Sze, Y. H. Chen, T. J. Yang, and J. S. Emer. Efficient processing of deep neural networks: A tutorial and survey. Proceedings of the IEEE, 105(12), 2017. +[44] Yuchen Wang, Malu Zhang, Yi Chen, and Hong Qu. Signed neuron with memory: Towards simple, accurate and high-efficient ann-snn conversion. In International Joint Conference on Artificial Intelligence, 2022. +[45] Jibin Wu, Yansong Chua, Malu Zhang, Guoqi Li, Haizhou Li, and Kay Chen Tan. A tandem learning rule for effective training and rapid inference of deep spiking neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2021. +[46] Jibin Wu, Chenglin Xu, Xiao Han, Daquan Zhou, Malu Zhang, Haizhou Li, and Kay Chen Tan. Progressive tandem learning for pattern recognition with deep spiking neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7824-7840, 2021. +[47] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018. +[48] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 1311-1318, 2019. +[49] Qi Xu, Yaxin Li, Jiangrong Shen, Jian K Liu, Huajin Tang, and Gang Pan. Constructing deep spiking neural networks from artificial neural networks with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7886-7895, 2023. +[50] Qi Xu, Yaxin Li, Jiangrong Shen, Pingping Zhang, Jian K Liu, Huajin Tang, and Gang Pan. Hierarchical spiking-based model for efficient image classification with enhanced feature extraction and encoding. IEEE Transactions on Neural Networks and Learning Systems, 2022. +[51] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence, 2023. +[52] Qiang Yu, Huajin Tang, Kay Chen Tan, and Haizhou Li. Rapid feedforward computation by temporal encoding and learning with spiking neurons. IEEE transactions on neural networks and learning systems, 24(10):1539-1552, 2013. + +[53] Friedemann Zenke and Tim P Vogels. The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks. Neural computation, 33(4):899-925, 2021. +[54] Jilin Zhang, Dexuan Huo, Jian Zhang, Chunqi Qian, Qi Liu, Liyang Pan, Zhihua Wang, Ning Qiao, Kea-Tiong Tang, and Hong Chen. 22.6 anp-i: A 28nm 1.5 pj/sop asynchronous spiking neural network processor enabling sub-o. 1 $\mu$ j/sample on-chip learning for edge-ai applications. In 2023 IEEE International Solid-State Circuits Conference (ISSCC), pages 21–23. IEEE, 2023. +[55] Jian Zhang, Jilin Zhang, Dexuan Huo, and Hong Chen. Anas: Asynchronous neuromorphic hardware architecture search based on a system-level simulator. in proc. 60th acm/ieee design automation conference (dac), 2023. In Proc. 60th ACM/IEEE Design Automation Conference (DAC), 2023, 2023. +[56] Malu Zhang, Xiaoling Luo, Yi Chen, Jibin Wu, Ammar Belatreche, Zihan Pan, Hong Qu, and Haizhou Li. An efficient threshold-driven aggregate-label learning algorithm for multimodal information processing. IEEE Journal of Selected Topics in Signal Processing, 14(3):592-602, 2020. +[57] Malu Zhang, Hong Qu, Ammar Belatreche, Yi Chen, and Zhang Yi. A highly effective and robust membrane potential-driven supervised learning method for spiking neurons. IEEE transactions on neural networks and learning systems, 30(1):123-137, 2018. +[58] Malu Zhang, Jiadong Wang, Jibin Wu, Ammar Belatreche, Burin Amornpaisannon, Zhixuan Zhang, Venkata Pavan Kumar Miriyala, Hong Qu, Yansong Chua, Trevor E Carlson, et al. Rectified linear postsynaptic potential function for backpropagation in deep spiking neural networks. IEEE transactions on neural networks and learning systems, 33(5):1947-1958, 2021. +[59] Wenrui Zhang and Peng Li. Temporal spike sequence learning via backpropagation for deep spiking neural networks. Advances in Neural Information Processing Systems, 33:12022-12033, 2020. +[60] Shibo Zhou, Xiaohua Li, Ying Chen, Sanjeev T Chandrasekaran, and Arindam Sanyal. Temporal-coded deep spiking neural network with easy training and robust performance. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 11143-11151, 2021. +[61] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, pages 3-11. Springer, 2018. +[62] Zongwei Zhou, Jae Shin, Lei Zhang, Suryakanth Gurudu, Michael Gotway, and Jianming Liang. Fine-tuning convolutional neural networks for biomedical image analysis: actively and incrementally. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7340-7351, 2017. + +[63] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, YAN Shuicheng, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations, 2022. +[64] Rui-Jie Zhu, Qihang Zhao, Tianjing Zhang, Haoyu Deng, Yule Duan, Malu Zhang, and Liang-Jian Deng. Tcjasnn: Temporal-channel joint attention for spiking neural networks. arXiv preprint arXiv:2206.10177, 2022. +[65] Yaoyu Zhu, Zhaofei Yu, Wei Fang, Xiaodong Xie, Tiejun Huang, and Timothée Masquelier. Training spiking neural networks with event-driven backpropagation. In 36th Conference on Neural Information Processing Systems (NeurIPS 2022), 2022. \ No newline at end of file diff --git a/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/images.zip b/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8ef030a552c6d0d307b9eff4a5a240de50721955 --- /dev/null +++ b/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc23c745309fb9a4f785777646bd5aacd6d7e06835d1735dc23ec87254e01153 +size 504537 diff --git a/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/layout.json b/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ee1996a42820d7f18235c901518092adece3d30d --- /dev/null +++ b/2023/Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold_ Learning with Event-Driven Backpropagation/layout.json @@ -0,0 +1,9788 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 531, + 140 + ], + "type": "text", + "content": "Temporal-Coded Spiking Neural Networks with Dynamic Firing Threshold: Learning with Event-Driven Backpropagation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "spans": [ + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "type": "text", + "content": "Wenjie Wei" + }, + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "type": "text", + "content": ", Malu Zhang" + }, + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "type": "text", + "content": ", Hong Qu" + }, + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "type": "text", + "content": ", Ammar Belatreche" + }, + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "type": "text", + "content": ", Jian Zhang" + }, + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "type": "text", + "content": ", and Hong Chen" + }, + { + "bbox": [ + 75, + 170, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 194, + 441, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 194, + 441, + 223 + ], + "spans": [ + { + "bbox": [ + 151, + 194, + 441, + 223 + ], + "type": "text", + "content": "1University of Electronic Science and Technology of China, 2Northumbria University, 3Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 225, + 541, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 225, + 541, + 237 + ], + "spans": [ + { + "bbox": [ + 51, + 225, + 541, + 237 + ], + "type": "text", + "content": "{maluzhang,hongqu}@uestc.edu.cn, ammar.belatreche@northumbria.ac.uk, hongchen@tsinghua.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 273, + 192, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 303, + 289, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 303, + 289, + 662 + ], + "spans": [ + { + "bbox": [ + 45, + 303, + 289, + 662 + ], + "type": "text", + "content": "Spiking Neural Networks (SNNs) offer a highly promising computing paradigm due to their biological plausibility, exceptional spatiotemporal information processing capability and low power consumption. As a temporal encoding scheme for SNNs, Time-To-First-Spike (TTFS) encodes information using the timing of a single spike, which allows spiking neurons to transmit information through sparse spike trains and results in lower power consumption and higher computational efficiency compared to traditional rate-based encoding counterparts. However, despite the advantages of the TTFS encoding scheme, the effective and efficient training of TTFS-based deep SNNs remains a significant and open research problem. In this work, we first examine the factors underlying the limitations of applying existing TTFS-based learning algorithms to deep SNNs. Specifically, we investigate issues related to over-sparsity of spikes and the complexity of finding the 'causal set'. We then propose a simple yet efficient dynamic firing threshold (DFT) mechanism for spiking neurons to address these issues. Building upon the proposed DFT mechanism, we further introduce a novel direct training algorithm for TTFS-based deep SNNs, called DTA-TTFS. This method utilizes event-driven processing and spike timing to enable efficient learning of deep SNNs. The proposed training method was validated on the image classification task and experimental results clearly demonstrate that our proposed method achieves state-of-the-art accuracy in comparison to existing TTFS-based learning algorithms, while maintaining high levels of sparsity and energy efficiency on neuromorphic inference accelerator." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 273, + 386, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 273, + 386, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 273, + 386, + 285 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 293, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 461 + ], + "type": "text", + "content": "Deep learning approaches have demonstrated remarkable performance in various intelligent applications [18, 27, 24, 23] and innovative simulations [62, 61]. However, these achievements come at a significant cost of energy consumption, severely limiting the deployment of artificial neural networks (ANNs) on resource-limited platforms [43, 9, 3, 50]. Brain-inspired spiking neural networks (SNNs) have emerged as a promising and energy-efficient alternative to ANNs [22, 31, 57]. SNNs have garnered attention for their distinctive features of sparse-asynchronous spikes and event-driven computing, which have played a pivotal role in driving the development of neuromorphic computing platforms such as ANP-I [54], TrueNorth [1], Loihi [8], and Tianjic [39]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 461, + 546, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 546, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 546, + 593 + ], + "type": "text", + "content": "Despite the inherent energy efficiency of SNNs, their extensive application to real-world practical scenarios has been limited by the absence of efficient and scalable training algorithms [52, 35, 30, 14, 20]. Traditional training techniques for ANNs, such as error backpropagation (BP) and standard GPU-accelerated training packages, cannot be leveraged directly by SNNs due to the complex temporal dynamics and the non-differentiability of spiking neurons' activity [33, 6, 56]. To address these challenges, various solutions have been proposed, broadly categorized into three distinct groups." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "The first category of solutions proposed to address the lack of scalable training algorithms for SNNs is ANN-to-SNN conversion method [40, 41, 49, 45]. These methods involve training an ANN first and then converting it to its SNN counterpart. The resulting SNN may achieve comparable performance to pre-trained ANNs through manual and careful setting of the SNN parameters such as the firing threshold [28] and the chosen spiking neuron model [44]. However, such conversion often results in precision loss due to approximation and requires longer time steps for lossless" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10552" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "conversion [19]. Moreover, ANN-to-SNN conversion methods often encode the activation values of artificial neurons using the firing rate of spiking neurons [11, 4, 5], which fails to fully leverage the timing information of spikes. Furthermore, the runtime computation on neuromorphic hardware can often be highly energy-intensive." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 288, + 361 + ], + "type": "text", + "content": "The second category of the proposed SNN training solutions is RNN-like method, commonly known as surrogate gradient learning [47, 48]. In these surrogate gradient-based SNNs, information is conveyed through spikes during feedforward computation, while during backpropagation, gradient information is calculated by treating the neurons' membrane potential as a differentiable signal at the current time step [42, 34]. To address the non-differentiability of spikes in the direct training of SNNs, surrogate functions are employed to approximate the derivative of the spike generation function during the backward process [53, 29]. However, while these methods have successfully tackled the challenges associated with the non-differentiability of spiking neurons, they necessitate updating synaptic weights at each time step, regardless of whether there is a spike emission or not [47, 15]. Consequently, this results in substantial energy and memory demands as large intermediate activation values necessitate storage to facilitate gradient computation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 361, + 288, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 288, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 288, + 637 + ], + "type": "text", + "content": "Unlike the above-mentioned methods, the third category of learning algorithms in SNNs is spike-driven method [2, 33, 58, 60, 38, 6], which operate in a strictly event-driven manner during both forward and backward computations. SpikeProp [2] and its various extensions [33, 6, 25] are typical examples of this category. However, these algorithms remain limited to shallow network structures and relatively simple classification tasks. To overcome this limitation, Zhang et al. [58] and Zhou et al. [60] extended the spikedriven algorithm to more complex network structures. Despite achieving competitive accuracy on large datasets [60], the identification of the 'causal set' in this approach requires substantial computational resources. To improve the training efficiency, Park et al. [38] proposed a surrogate deep neural network (DNN) model that is subsequently converted into a TTFS-based SNN. However, this conversion method is prone to conversion errors, potentially impairing the temporal learning capability of SNNs. Recently, several multi-spike-based spike-driven algorithms have been developed [59, 65]. While Zhang et al. [59] and Zhu et al. [65] achieve competitive performance on CIFAR utilizing the rate coding scheme, we focus on developing an energy-efficient TTFS-based spike-driven algorithm." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 638, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 287, + 662 + ], + "type": "text", + "content": "In this paper, we introduce an effective and efficient spike-driven learning algorithm that facilitates the training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "content": "of high-performance TTFS-based deep SNNs. The main contributions of this work are outlined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 105, + 545, + 382 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 105, + 545, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 105, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 317, + 105, + 545, + 165 + ], + "type": "text", + "content": "- We provide a comprehensive analysis of the main shortcomings of existing methods in achieving high performance in TTFS-based deep SNNs. Specifically, we examine issues such as the over-sparsity of spikes and the complexity of determining the 'causal set'." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 174, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 545, + 209 + ], + "type": "text", + "content": "- We propose a simple yet efficient dynamic firing threshold (DFT) mechanism for spiking neurons that can effectively address the aforementioned issues." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 218, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 218, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 317, + 218, + 545, + 290 + ], + "type": "text", + "content": "- Building upon the proposed DFT, we further introduce a novel direct training algorithm for TTFS-based deep SNNs, which we refer to as DTA-TTFS. In this training method, the timing of a single spike is considered the basic information carrier and the learning process is performed strictly in an event-driven manner." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 299, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 545, + 382 + ], + "type": "text", + "content": "- The proposed method is validated through extensive experiments on benchmark image classification tasks and achieves state-of-the-art accuracy compared to existing TTFS-based learning algorithms. Furthermore, we demonstrate the ultra-low power capability of the SNN with DTA-TTFS on a recently developed neuromorphic accelerator." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 394, + 505, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 505, + 407 + ], + "type": "text", + "content": "2. Preliminaries and Problem Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 415, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 545, + 462 + ], + "type": "text", + "content": "In this section, we first briefly review the TTFS coding and the employed spiking neuron model. Subsequently, we provide a comprehensive analysis of the primary shortcomings in existing TTFS-based deep SNN methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 470, + 391, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 470, + 391, + 482 + ], + "spans": [ + { + "bbox": [ + 306, + 470, + 391, + 482 + ], + "type": "text", + "content": "2.1. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 489, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 545, + 620 + ], + "type": "text", + "content": "TTFS coding scheme: There are two primary coding schemes utilized to represent information in SNNs: rate coding and temporal coding. The rate coding method represents information based on the average firing rate of a spiking neuron. However, this encoding scheme neglects the temporal information carried by spikes, leading to higher spike frequency and increased energy consumption. Conversely, temporal coding leverages the precise timing of spikes to encode information, resulting in lower power consumption and higher computational efficiency compared to the rate-based coding method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 621, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 621, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 621, + 545, + 681 + ], + "type": "text", + "content": "As a temporal coding method, TTFS encodes information by the timing of the first spike. As depicted in Fig. 1, the higher the activation value, the earlier the first spike is emitted. The relationship between the input activation value " + }, + { + "bbox": [ + 304, + 621, + 545, + 681 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 304, + 621, + 545, + 681 + ], + "type": "text", + "content": " and the encoded spike time " + }, + { + "bbox": [ + 304, + 621, + 545, + 681 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 621, + 545, + 681 + ], + "type": "text", + "content": " can be expressed as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 372, + 689, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 689, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 372, + 689, + 545, + 716 + ], + "type": "interline_equation", + "content": "t _ {i} = \\left(1 - \\frac {a _ {i}}{a _ {\\text {m a x}}}\\right) \\times T _ {w}, \\tag {1}", + "image_path": "4eab239508f99103a8e205d6cec79a6deeebcf2f95e57686097046b1638624f3.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "text", + "content": "1Assume a postsynaptic neuron fire a first spike at " + }, + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "inline_equation", + "content": "t_{out}" + }, + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "text", + "content": ", only a subset of input spikes had arrived before " + }, + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "inline_equation", + "content": "t_{out}" + }, + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "text", + "content": " and contributed to the output spike. This set of spikes, " + }, + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "inline_equation", + "content": "C = \\{i : t_i < t_{out}\\}" + }, + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "text", + "content": ", is called the causal set in this work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10553" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 182, + 153 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 182, + 153 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 182, + 153 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 182, + 153 + ], + "type": "image", + "image_path": "2c6d616ed5816f58b20b1c4c29eb8d0000ecdf289ee7efde9b362c28d5c5ef92.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 160, + 116, + 169 + ], + "lines": [ + { + "bbox": [ + 104, + 160, + 116, + 169 + ], + "spans": [ + { + "bbox": [ + 104, + 160, + 116, + 169 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 70, + 278, + 151 + ], + "blocks": [ + { + "bbox": [ + 184, + 70, + 278, + 151 + ], + "lines": [ + { + "bbox": [ + 184, + 70, + 278, + 151 + ], + "spans": [ + { + "bbox": [ + 184, + 70, + 278, + 151 + ], + "type": "image", + "image_path": "3c7b1045b9577e1c511f3fe5260a99ef67a796ad09755afd0d1ba86d2ca99205.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 159, + 239, + 169 + ], + "lines": [ + { + "bbox": [ + 228, + 159, + 239, + 169 + ], + "spans": [ + { + "bbox": [ + 228, + 159, + 239, + 169 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": "Figure 1. (a) TTFS coding scheme. The larger intensity value corresponds to the earlier spike. (b) ReL-PSP spiking neuron model. Neuron " + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": " accumulates the linear PSP and fires a spike when the membrane potential crosses the firing threshold " + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\vartheta" + }, + { + "bbox": [ + 46, + 171, + 287, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 235, + 287, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 287, + 330 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 235, + 287, + 330 + ], + "type": "inline_equation", + "content": "a_{max}" + }, + { + "bbox": [ + 46, + 235, + 287, + 330 + ], + "type": "text", + "content": " is the maximum input activation value, " + }, + { + "bbox": [ + 46, + 235, + 287, + 330 + ], + "type": "inline_equation", + "content": "T_w" + }, + { + "bbox": [ + 46, + 235, + 287, + 330 + ], + "type": "text", + "content": " is the length of the permissible spike time window. As the TTFS coding scheme requires only a single spike to convey information, it effectively diminishes the number of spikes and significantly decreases the energy requirements. Consequently, it achieves higher sparsity and energy efficiency, making it a powerful coding scheme that maximizes the energy efficiency of SNNs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 331, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 402 + ], + "type": "text", + "content": "Spiking neuron model: The Rectified Linear Postsynaptic Potential (ReL-PSP) spiking neuron [58] is chosen for our method due to its simplicity and energy efficiency. As shown in Fig. 1(b), the ReL-PSP neuron model utilizes a rectified linear kernel function to shape the neuron's PSP, and its dynamics is defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 411, + 287, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 411, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 102, + 411, + 287, + 442 + ], + "type": "interline_equation", + "content": "V _ {j} ^ {l} (t) = \\sum_ {i = 1} ^ {N} w _ {i j} ^ {l} K \\left(t - t _ {i} ^ {l - 1}\\right), \\tag {2}", + "image_path": "ca1ee814fd149cbfc0aa43318563117ba83afe1cdbf7121b6ae3c98b4e59655e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "inline_equation", + "content": "V_{j}^{l}(t)" + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "text", + "content": " is the membrane potential of neuron " + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "text", + "content": " in layer " + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "inline_equation", + "content": "w_{ij}^{l}" + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "text", + "content": " is the synaptic weight between neuron " + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "text", + "content": " and its presynaptic neuron " + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "text", + "content": " is the number of neurons in the previous layer " + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "inline_equation", + "content": "l - 1" + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "inline_equation", + "content": "t_{i}^{l - 1}" + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "text", + "content": " is the spike time of input neuron " + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "inline_equation", + "content": "K\\left(t - t_{i}^{l - 1}\\right)" + }, + { + "bbox": [ + 47, + 451, + 287, + 515 + ], + "type": "text", + "content": " is the PSP function that is defined as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 78, + 522, + 287, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 522, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 78, + 522, + 287, + 555 + ], + "type": "interline_equation", + "content": "K \\left(t - t _ {i} ^ {l - 1}\\right) = \\left\\{ \\begin{array}{l l} t - t _ {i} ^ {l - 1}, & \\text {i f} \\quad t > t _ {i} ^ {l - 1}, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {3}", + "image_path": "06e6583b707a898844e6ac76e197795d4ef50bd5be4a60753110d6201d3be7ec.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 563, + 287, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 287, + 599 + ], + "type": "text", + "content": "The neuron " + }, + { + "bbox": [ + 47, + 563, + 287, + 599 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 563, + 287, + 599 + ], + "type": "text", + "content": " emits a spike when " + }, + { + "bbox": [ + 47, + 563, + 287, + 599 + ], + "type": "inline_equation", + "content": "V_{j}^{l}(t)" + }, + { + "bbox": [ + 47, + 563, + 287, + 599 + ], + "type": "text", + "content": " reaches the firing threshold " + }, + { + "bbox": [ + 47, + 563, + 287, + 599 + ], + "type": "inline_equation", + "content": "\\vartheta" + }, + { + "bbox": [ + 47, + 563, + 287, + 599 + ], + "type": "text", + "content": ". Mathematically, the spike generation function " + }, + { + "bbox": [ + 47, + 563, + 287, + 599 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 563, + 287, + 599 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 606, + 287, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 606, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 107, + 606, + 287, + 622 + ], + "type": "interline_equation", + "content": "t _ {j} ^ {l} = F \\{t | V _ {j} ^ {l} (t) \\geq \\vartheta , t \\geq 0 \\}. \\tag {4}", + "image_path": "7ec4400beb670802b3deb7bbe223b9df5c155aab6959cbcae18a94bc1bd63b8e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "content": "The ReL-PSP neuron has proven its efficacy in tackling the main challenges encountered during the training of deep SNNs, such as the non-differentiability of the spike function, gradient explosion, and the dead neuron issue. Despite the successful performance on the MNIST dataset, its performance is not ideal on deeper network structures and more complex datasets, such as CIFAR-10 and CIFAR-100." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "In the subsequent section, we will conduct a comprehensive analysis to identify the underlying reasons for this limitation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 115, + 410, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 115, + 410, + 129 + ], + "spans": [ + { + "bbox": [ + 306, + 115, + 410, + 129 + ], + "type": "text", + "content": "2.2. Problem Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 133, + 545, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 229 + ], + "type": "text", + "content": "Over-sparsity of spikes: Sparse neural networks offer several advantages such as efficient variable-size representation, information decomposition, and energy efficiency. However, over-sparsity may diminish the model's effective capacity and adversely affect its performance [17]. This issue becomes more serious in TTFS-based deep SNNs, owing to their firing mechanism, asynchronous transmission, and TTFS-based decision strategy." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 328, + 239, + 421, + 323 + ], + "blocks": [ + { + "bbox": [ + 328, + 239, + 421, + 323 + ], + "lines": [ + { + "bbox": [ + 328, + 239, + 421, + 323 + ], + "spans": [ + { + "bbox": [ + 328, + 239, + 421, + 323 + ], + "type": "image", + "image_path": "64ae6d96b091219cd27021032a4b4cf4a6b1e7cbb361c424d6de03cab5378082.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 327, + 377, + 337 + ], + "lines": [ + { + "bbox": [ + 366, + 327, + 377, + 337 + ], + "spans": [ + { + "bbox": [ + 366, + 327, + 377, + 337 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 350, + 545, + 427 + ], + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 427 + ], + "type": "text", + "content": "Figure 2. The over-sparsity problem caused by asynchronous transmission and TTFS-based decision strategy. (a) The presynaptic (gray) neuron fires later than the postsynaptic (blue) neuron " + }, + { + "bbox": [ + 304, + 350, + 545, + 427 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 350, + 545, + 427 + ], + "type": "text", + "content": ", making it ineffective in contributing to the firing of the postsynaptic neuron. As a result, this presynaptic neuron becomes invalid. (b) After learning, the postsynaptic (blue) neuron fires at an earlier time, leading to the appearance of additional invalid neurons." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 437, + 240, + 525, + 323 + ], + "blocks": [ + { + "bbox": [ + 437, + 240, + 525, + 323 + ], + "lines": [ + { + "bbox": [ + 437, + 240, + 525, + 323 + ], + "spans": [ + { + "bbox": [ + 437, + 240, + 525, + 323 + ], + "type": "image", + "image_path": "ca111c35d49f8fac874315a4faf27d2eab53d643ab7c43b09565d777bf2422fc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 327, + 484, + 337 + ], + "lines": [ + { + "bbox": [ + 473, + 327, + 484, + 337 + ], + "spans": [ + { + "bbox": [ + 473, + 327, + 484, + 337 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "text", + "content": "According to the firing mechanism of spiking neurons, a neuron emits one spike only when its membrane potential reaches the firing threshold. This leads to fewer activated spiking neurons compared to ANNs. In addition, the asynchronous transmission of SNNs further diminishes the number of valid activated neurons. As shown in Fig. 2(a), when the presynaptic neuron " + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "text", + "content": " fires after the postsynaptic neuron " + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "text", + "content": ", the spike of neuron " + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "text", + "content": " becomes invalid for neuron " + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "text", + "content": ". Furthermore, TTFS-based SNNs make decisions based on the earliest spike, and the synaptic weights are trained to make the target neuron fire as early as possible. As shown in Fig. 2(b), after learning, the firing time of the postsynaptic neuron " + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "text", + "content": " may become earlier, thereby further exacerbating the sparsity induced by asynchronous transmission." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": "In event-driven learning algorithms of SNNs, error backpropagation depends on the timing of spikes. Therefore, if a spiking neuron does not fire or does not contribute to the postsynaptic neuron, it will not participate in learning. The issue of over-sparsity not only limits the information representation capability of SNNs but also leads to learning failure. Consequently, the achievement of a harmonious balance between sparsity and performance in the training of TTFS-based deep SNNs remains an open question." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10554" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": "The complexity of finding the 'causal set': Due to the sparsity and asynchronicity of SNNs, not all presynaptic neurons " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\{i|i = 1,\\dots ,N\\}" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " contribute to their postsynaptic neuron " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": ", where the contributory neurons are defined as the 'causal set' of the postsynaptic neuron " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "C_j = \\{i|t_i < t_j\\}" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " [33, 60]. To leverage standard GPU-accelerated ANN training packages, it is crucial to identify the 'causal set' of each neuron during the feed-forward computation process [33, 60]. However, this process is highly complex and time-consuming. For example, in order to find the 'causal set', Zhou et al. [60] proposed an approach where they first sort all of this neuron input spikes " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\{t_i|i = 1,\\dots N\\}" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " in ascending order and subsequently consider each individual spike " + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 47, + 72, + 289, + 275 + ], + "type": "text", + "content": " for inclusion in the causal set. This approach, however, is computationally expensive and results in inefficient training even with the implementation of a space-for-time strategy [60]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 396 + ], + "type": "text", + "content": "Despite the success of algorithms with this approach in utilizing standard GPU-accelerated ANN training packages, the training process involves a significant amount of sorting and iteration. This results in higher demand for training resources and increased latency, especially as the network depth increases. Consequently, the process of finding the 'causal set' may become a bottleneck for training deep SNNs. In the next section, we propose a dynamic firing threshold mechanism for spiking neurons to avoid this resource-intensive process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 405, + 104, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 104, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 104, + 417 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 485 + ], + "type": "text", + "content": "In this section, we first introduce the dynamic firing threshold (DFT) for spiking neurons and discuss how it effectively addresses the above issues. Then, based on the DFT, we further propose a direct training algorithm for TTFS-based SNN, referred to as DTA-TTFS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 492, + 230, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 492, + 230, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 492, + 230, + 506 + ], + "type": "text", + "content": "3.1. DFT-based Spiking Neuron Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 545 + ], + "type": "text", + "content": "As demonstrated in Fig. 3(a), the proposed DFT is a layer-dependent and time-varying function, which is defined as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 287, + 594 + ], + "type": "interline_equation", + "content": "\\vartheta^ {l} (t) = \\left\\{ \\begin{array}{l l} T _ {w} (l + 1) - t, & \\text {i f} \\quad T _ {w} l \\leq t \\leq T _ {w} (l + 1), \\\\ + \\infty , & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {5}", + "image_path": "01d4476056078eab33bdd5f77eef65c7ef8385f1fb9b69ef5375855cc412a47d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\vartheta^l (t)" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " represents the firing threshold of the neurons in the " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "-th layer and " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "T_{w}" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " denotes the length of the permissible spike time window. In this work, " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "T_{w}" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " is set to 1 (unless otherwise stated) and the input layer is regarded as the 0-th layer. Eq. 5 reveals that the firing threshold is infinite outside the interval " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "t\\in [T_wl,T_w(l + 1)]" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": ", and neurons in the " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "-th layer can only generate a spike within this time interval. Therefore, as shown in Fig. 4, the generated spikes are non-overlapping across different layers, which we refer to as the layer-dependent firing feature." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "Based on the membrane potential dynamics in Eq. 2 and the proposed DFT in Eq. 5, the output of neuron " + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": " can be divided into three categories:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 117, + 545, + 285 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "text", + "content": "- Spikes at " + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "inline_equation", + "content": "t_j^l = T_{wl}" + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "text", + "content": ": Since the firing threshold is infinity when " + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "inline_equation", + "content": "t < T_{wl}" + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "text", + "content": ", the neuron cannot fire a spike before " + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "inline_equation", + "content": "t = T_{wl}" + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "text", + "content": " even though the membrane potential is strong (i.e., " + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "inline_equation", + "content": "V_j^l(t) \\geq \\vartheta^l(T_{wl})" + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "inline_equation", + "content": "t < T_{wl}" + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "text", + "content": "). As shown in Fig. 3(b), the earliest spike time for DFT-based spiking neurons in the " + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "text", + "content": "-th layer is " + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "inline_equation", + "content": "t_j^l = T_{wl}" + }, + { + "bbox": [ + 317, + 117, + 545, + 192 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 199, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 199, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 199, + 545, + 285 + ], + "type": "text", + "content": "- Spikes within " + }, + { + "bbox": [ + 317, + 199, + 545, + 285 + ], + "type": "inline_equation", + "content": "(T_{w}l, T_{w}(l + 1))" + }, + { + "bbox": [ + 317, + 199, + 545, + 285 + ], + "type": "text", + "content": ": The neuron generates a spike when its membrane potential exceeds the threshold, i.e. " + }, + { + "bbox": [ + 317, + 199, + 545, + 285 + ], + "type": "inline_equation", + "content": "V_{j}^{l}(t) \\geq \\vartheta^{l}(t)" + }, + { + "bbox": [ + 317, + 199, + 545, + 285 + ], + "type": "text", + "content": ". The firing threshold " + }, + { + "bbox": [ + 317, + 199, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\vartheta^{l}" + }, + { + "bbox": [ + 317, + 199, + 545, + 285 + ], + "type": "text", + "content": " in this time interval is a linear function, which decays linearly from 1 to 0. As depicted in Fig. 3(c), by combining the membrane potential in Eq. 2 and the firing threshold in Eq. 5, the precise spiking time " + }, + { + "bbox": [ + 317, + 199, + 545, + 285 + ], + "type": "inline_equation", + "content": "t_{j}^{l}" + }, + { + "bbox": [ + 317, + 199, + 545, + 285 + ], + "type": "text", + "content": " can be calculated as" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 369, + 294, + 545, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 294, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 369, + 294, + 545, + 341 + ], + "type": "interline_equation", + "content": "t _ {j} ^ {l} = \\frac {T _ {w} (l + 1) + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l} t _ {i} ^ {l - 1}}{1 + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l}}. \\tag {6}", + "image_path": "f4b7cbdb125b7329be43c09019a6b061ac3bbe1d5e9d84a6f98ebd79b7a25996.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 355, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 545, + 403 + ], + "type": "text", + "content": "- Nonspike: The neuron will not fire a spike if its membrane potential is below the dynamic firing threshold (i.e., " + }, + { + "bbox": [ + 317, + 355, + 545, + 403 + ], + "type": "inline_equation", + "content": "V_{j}^{l}(t) < \\vartheta^{l}(t)" + }, + { + "bbox": [ + 317, + 355, + 545, + 403 + ], + "type": "text", + "content": "). Fig. 3(d) shows three typical scenarios of nonspike." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 412, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 437 + ], + "type": "text", + "content": "Overall, the spike time of a DFT-based neuron " + }, + { + "bbox": [ + 304, + 412, + 545, + 437 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 412, + 545, + 437 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 412, + 545, + 437 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 412, + 545, + 437 + ], + "type": "text", + "content": "-th layer can be represented as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 447, + 545, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 447, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 305, + 447, + 545, + 520 + ], + "type": "interline_equation", + "content": "t _ {j} ^ {l} = \\left\\{ \\begin{array}{l l} T _ {w} l, & \\text {i f} \\quad V _ {j} ^ {l} \\left(T _ {w} l\\right) \\geq \\vartheta^ {l} \\left(T _ {w} l\\right), \\\\ \\frac {T _ {w} (l + 1) + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l} t _ {i} ^ {l - 1}}{1 + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l}}, & \\text {i f} \\quad T _ {w} l < t _ {j} ^ {l} \\leq T _ {w} (l + 1), \\\\ \\text {n o n s p i k e}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}", + "image_path": "8e6d56908ae120ddaa7ad66cd3bf1daaa61205fc78b5ee9b7c9b6442690f042e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 521, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 556 + ], + "type": "text", + "content": "In the following, we will analyze how the proposed DFT effectively addresses issues of over-sparsity of spikes and the complexity of finding the 'causal set'." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 605 + ], + "type": "text", + "content": "1) Over-sparsity of spikes: As described in Eq. 5, the DFT is a piecewise function, whose linearly decreased part and layer-dependent firing feature effectively address the over-sparsity of spikes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "The linearly decreased part of the DFT is purposefully designed to tackle the over-sparsity arising from the spiking neuron firing mechanism. By gradually decreasing the firing threshold from 1 to 0, the neuron " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " remains active as long as its membrane potential is non-negative at " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "T_{w}(l + 1)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". As a result, the proposed DFT effectively reduces the number of inactive neurons. Moreover, the linear decay design enhances the discriminative nature of information from active neurons." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10555" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 172, + 153 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 172, + 153 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 172, + 153 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 172, + 153 + ], + "type": "image", + "image_path": "fd574486f30506d416cadb62ae3654d1cc072c345b6c02a4414e5a44c49e9808.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 157, + 115, + 166 + ], + "lines": [ + { + "bbox": [ + 104, + 157, + 115, + 166 + ], + "spans": [ + { + "bbox": [ + 104, + 157, + 115, + 166 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 73, + 297, + 153 + ], + "blocks": [ + { + "bbox": [ + 179, + 73, + 297, + 153 + ], + "lines": [ + { + "bbox": [ + 179, + 73, + 297, + 153 + ], + "spans": [ + { + "bbox": [ + 179, + 73, + 297, + 153 + ], + "type": "image", + "image_path": "41bdde42027fe031e85074cac6e17050278ef41be38365c8fa51d1ad476e9245.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 157, + 242, + 166 + ], + "lines": [ + { + "bbox": [ + 231, + 157, + 242, + 166 + ], + "spans": [ + { + "bbox": [ + 231, + 157, + 242, + 166 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 170, + 547, + 202 + ], + "lines": [ + { + "bbox": [ + 46, + 170, + 547, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 547, + 202 + ], + "type": "text", + "content": "Figure 3. Proposed dynamic firing threshold and output of the DFT-based spiking neurons. (a) Dynamic firing threshold function(DFT). (b) Spikes at " + }, + { + "bbox": [ + 46, + 170, + 547, + 202 + ], + "type": "inline_equation", + "content": "t_j^l = T_w l" + }, + { + "bbox": [ + 46, + 170, + 547, + 202 + ], + "type": "text", + "content": ". (c) Spikes within " + }, + { + "bbox": [ + 46, + 170, + 547, + 202 + ], + "type": "inline_equation", + "content": "(T_w l, T_w (l + 1)]" + }, + { + "bbox": [ + 46, + 170, + 547, + 202 + ], + "type": "text", + "content": ". (d) Nonspike. This case is further divided into three sub-cases, which are shown by the curves of different colors." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 303, + 72, + 420, + 153 + ], + "blocks": [ + { + "bbox": [ + 303, + 72, + 420, + 153 + ], + "lines": [ + { + "bbox": [ + 303, + 72, + 420, + 153 + ], + "spans": [ + { + "bbox": [ + 303, + 72, + 420, + 153 + ], + "type": "image", + "image_path": "91f457e988f999826c4da082ffc2f2be246ea95f46742ab2c645c0e015ae782b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 157, + 365, + 166 + ], + "lines": [ + { + "bbox": [ + 355, + 157, + 365, + 166 + ], + "spans": [ + { + "bbox": [ + 355, + 157, + 365, + 166 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 428, + 70, + 543, + 153 + ], + "blocks": [ + { + "bbox": [ + 428, + 70, + 543, + 153 + ], + "lines": [ + { + "bbox": [ + 428, + 70, + 543, + 153 + ], + "spans": [ + { + "bbox": [ + 428, + 70, + 543, + 153 + ], + "type": "image", + "image_path": "133938a2ea83c75d7176e3ad1904a5a37ff3d5db422aa2dfa2d77e561d052ba8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 157, + 487, + 166 + ], + "lines": [ + { + "bbox": [ + 477, + 157, + 487, + 166 + ], + "spans": [ + { + "bbox": [ + 477, + 157, + 487, + 166 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 219, + 276, + 307 + ], + "blocks": [ + { + "bbox": [ + 58, + 219, + 276, + 307 + ], + "lines": [ + { + "bbox": [ + 58, + 219, + 276, + 307 + ], + "spans": [ + { + "bbox": [ + 58, + 219, + 276, + 307 + ], + "type": "image", + "image_path": "a954087f8b2fee71906adadcf04891013d7f9a7db6a9add4ef3c32991f1ebf05.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 309, + 271, + 320 + ], + "lines": [ + { + "bbox": [ + 61, + 309, + 271, + 320 + ], + "spans": [ + { + "bbox": [ + 61, + 309, + 271, + 320 + ], + "type": "text", + "content": "Figure 4. The propagation pipeline of the SNN with DFT." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 341, + 286, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 286, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 286, + 473 + ], + "type": "text", + "content": "The layer-dependent firing feature enables SNN with non-overlapping spike time windows. This effectively resolves the over-sparsity arising from asynchronous transmission and TTFS-based decision strategy. Fig. 4 illustrates that the spike time window of postsynaptic neurons occurs later than that of presynaptic neurons. Consequently, when the presynaptic neuron fires a spike, its postsynaptic neuron absolutely receives it. Furthermore, the non-overlapping spike window provides output layer neurons with a predetermined earliest spike time, preventing them from being forced forward indefinitely during the training process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 473, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 286, + 533 + ], + "type": "text", + "content": "2) The complexity of finding the 'causal set': The layer-dependent firing feature of DFT circumvents the need for SNNs to identify the 'causal set', resulting in efficient training. This efficiency is further improved by the output analytic equation of the DFT-based neuron." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 533, + 286, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 653 + ], + "type": "text", + "content": "As depicted in Fig. 4, the layer-dependent firing feature ensures that in the " + }, + { + "bbox": [ + 46, + 533, + 286, + 653 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 533, + 286, + 653 + ], + "type": "text", + "content": "-th layer of the SNN utilizing DFT, the spike time is controlled within the interval " + }, + { + "bbox": [ + 46, + 533, + 286, + 653 + ], + "type": "inline_equation", + "content": "[T_w l, T_w (l + 1)]" + }, + { + "bbox": [ + 46, + 533, + 286, + 653 + ], + "type": "text", + "content": ". Therefore, we can easily learn that neurons in the " + }, + { + "bbox": [ + 46, + 533, + 286, + 653 + ], + "type": "inline_equation", + "content": "(l - 1)" + }, + { + "bbox": [ + 46, + 533, + 286, + 653 + ], + "type": "text", + "content": "-th layer must fire earlier than neurons in the " + }, + { + "bbox": [ + 46, + 533, + 286, + 653 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 533, + 286, + 653 + ], + "type": "text", + "content": "-th layer. In other words, all active presynaptic neurons must remain in the 'causal set' of their postsynaptic neurons. Therefore, the DFT mechanism prevents the need for resource-intensive sorting and iteration process, enhancing training efficiency and facilitating its extension to deeper network architecture." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "In addition, the employed ReL-PSP spiking neuron and the DFT provide an analytical relation between input spikes and output spikes, as described in Eq.7. Therefore, the dynamics of SNNs within the training loop need not be simulated using long time steps. This allows the utilization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 222, + 544, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 222, + 544, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 222, + 544, + 247 + ], + "type": "text", + "content": "of standard GPU-accelerated training packages of ANNs, which further leads to an efficient training process." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 259, + 479, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 479, + 272 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 479, + 272 + ], + "type": "text", + "content": "3.2. DTA-TTFS Learning Algorithm" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "text", + "content": "Consider a task with " + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "text", + "content": " categories, where each output neuron corresponds to a specific category. In this task, the target category is represented by the output neuron that fires the earliest. To assess the effectiveness of the model, we adopt the widely utilized cross-entropy loss function " + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "text", + "content": ", which aims to maximize the output value of the target neuron. In order to ensure the desired neuron fires earliest and others fire as late as possible, we utilize the negative spike times of the output layer [58, 6], i.e., " + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "inline_equation", + "content": "-t" + }, + { + "bbox": [ + 304, + 280, + 545, + 399 + ], + "type": "text", + "content": ", for evaluation. Therefore, the loss function is given by" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 353, + 411, + 545, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 411, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 353, + 411, + 545, + 438 + ], + "type": "interline_equation", + "content": "L (\\mathbf {t}, d) = - \\log \\frac {\\exp (- \\mathbf {t} [ d ])}{\\sum_ {i = 1} ^ {n} \\exp (- \\mathbf {t} [ i ])}, \\tag {8}", + "image_path": "1a7e51102ae8d19ceea8dd5f8386994e81692b622245578deaa281fe1058fa16.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 449, + 544, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 544, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 544, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 449, + 544, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 305, + 449, + 544, + 473 + ], + "type": "text", + "content": " is the vector of spike times in the output layer, and " + }, + { + "bbox": [ + 305, + 449, + 544, + 473 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 305, + 449, + 544, + 473 + ], + "type": "text", + "content": " is the desired category index." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 475, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 561 + ], + "type": "text", + "content": "In order to minimize the " + }, + { + "bbox": [ + 304, + 475, + 545, + 561 + ], + "type": "inline_equation", + "content": "L(\\mathbf{t},d)" + }, + { + "bbox": [ + 304, + 475, + 545, + 561 + ], + "type": "text", + "content": ", the DTA-TTFS algorithm adjusts spike times by modifying the synaptic weights during gradient backpropagation. For driving the gradient backpropagation, it is necessary to calculate derivatives of the output spike time " + }, + { + "bbox": [ + 304, + 475, + 545, + 561 + ], + "type": "inline_equation", + "content": "t_j^l" + }, + { + "bbox": [ + 304, + 475, + 545, + 561 + ], + "type": "text", + "content": " with respect to the synaptic weight " + }, + { + "bbox": [ + 304, + 475, + 545, + 561 + ], + "type": "inline_equation", + "content": "w_{ij}^{l}" + }, + { + "bbox": [ + 304, + 475, + 545, + 561 + ], + "type": "text", + "content": " and the input spike time " + }, + { + "bbox": [ + 304, + 475, + 545, + 561 + ], + "type": "inline_equation", + "content": "t_i^{l - 1}" + }, + { + "bbox": [ + 304, + 475, + 545, + 561 + ], + "type": "text", + "content": ". According to Eq. 7, we can easily get these two derivatives," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 573, + 545, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 573, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 312, + 573, + 545, + 617 + ], + "type": "interline_equation", + "content": "\\frac {\\partial t _ {j} ^ {l}}{\\partial w _ {i j} ^ {l}} = \\left\\{ \\begin{array}{l l} \\frac {t _ {i} ^ {l - 1} - t _ {j} ^ {l}}{1 + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l}}, & \\text {i f} \\quad T _ {w} l < t _ {j} ^ {l} < T _ {w} (l + 1), \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {9}", + "image_path": "b4cc4e5188a65a2a5cfba30aaee507bb0d95c3c6102bfc9253a2b0086dd63876.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 635, + 545, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 545, + 689 + ], + "type": "interline_equation", + "content": "\\frac {\\partial t _ {j} ^ {l}}{\\partial t _ {i} ^ {l - 1}} = \\left\\{ \\begin{array}{l l} \\frac {w _ {i j} ^ {l}}{1 + \\sum_ {i \\in C _ {j}} w _ {i j} ^ {l}}, & \\text {i f} \\quad T _ {w} l < t _ {j} ^ {l} < T _ {w} (l + 1), \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {10}", + "image_path": "0d1963dcebd0c098b4561f7f84a5e1757ab1ba6a177699c7bb5e7145393b43e4.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Following Eq. 9 and Eq. 10, gradient backpropagation can be directly applied." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10556" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 70, + 500, + 259 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 500, + 259 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 500, + 259 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 500, + 259 + ], + "type": "table", + "html": "
DatasetsModelsNetwork ArchitectureNeural CodingMethodAccuracySparsity
MNISTMostafa 2017 [33]MLP¹TFTSDT97.55%0.51
Zhang et al. 2020 [58]CNN¹TFTSDT99.4%0.6614
Zhou et al. 2021 [60]CNN²TFTSDT99.33%0.94
DTA-TTFSCNN¹TFTSDT99.4%0.3913
CIFAR-10Wu et al. 2022 [46]VGG11Rateconv91.24%no
Park et al. 2020 [37]VGG16TFTSconv91.43%0.2459
Zhou et al. 2021 [60]VGG16TFTSDT92.68%0.62
Park et al. 2021 [38]VGG16TFTSDT91.90%0.1746
DTA-TTFSVGG11TFTSDT91.17%0.4387
DTA-TTFSVGG16TFTSDT93.05%0.2561
CIFAR-100Park et al. 2020 [37]VGG16TFTSconv68.79%0.2994
Park et al. 2021 [38]VGG16TFTSDT65.98%0.2780
DTA-TTFSVGG16TFTSDT69.66%0.2845
", + "image_path": "30e36d01e61ae234b1feb76572f7972f7da67c114fc25c4ee754fdfae2833b55.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 545, + 289 + ], + "type": "text", + "content": "Table 1. Comparison of accuracy and sparsity between the DTA-TTFS method and other related works. DT: direct training, conv: ANN-to-SNN conversion. MLP1: 784-800-10. CNN1: 784-16C5-P2-32C5-P2-800-128-10. CNN2: 784-32C5-16C5-10." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 306, + 128, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 306, + 128, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 128, + 319 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 326, + 288, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 288, + 399 + ], + "type": "text", + "content": "In this section, we first evaluate the performance of the DTA-TTFS algorithm. Then, we carry out validation studies to illustrate the impact of the DFT. Finally, we employ theoretical analysis along with the neuromorphic hardware simulator to evaluate the energy consumption during the inference process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 403, + 102, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 102, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 102, + 415 + ], + "type": "text", + "content": "4.1. Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "type": "text", + "content": "We conduct experiments on three image classification benchmark datasets: MNIST" + }, + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "type": "text", + "content": ", CIFAR-10, and CIFAR" + }, + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "type": "inline_equation", + "content": "100^3" + }, + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "type": "text", + "content": ". All experiments are conducted using the PyTorch library, which allows for efficient training on multi-GPU machines with accelerated computations and optimized memory usage. For the MNIST dataset, we employ a network structure with 784-16C5-P2-32C5-P2-800-128-10 and train it for 150 epochs. As for CIFAR datasets, we adopt augmentation techniques [7, 10] and train the VGG11 and VGG16 network structures without Batch Normalization (BN) for 300 epochs. Moreover, in order to maintain the initial parameter domain at a good level, the SNN is initialized with the parameters from the pre-trained identical ANN when training CIFAR datasets. During the training process, we adopt the Adam optimizer and implement milestones learning rate decay, with an initial learning rate " + }, + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "type": "text", + "content": " for MNIST and " + }, + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 46, + 422, + 287, + 625 + ], + "type": "text", + "content": " for CIFAR datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 632, + 224, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 224, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 224, + 645 + ], + "type": "text", + "content": "4.2. Comparison with Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 651, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 651, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 287, + 687 + ], + "type": "text", + "content": "Tab. 1 presents a comparison of the accuracies and sparsity achieved by DTA-TTFS and other related works. Experimental results demonstrate that the DTA-TTFS algo" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "text", + "content": "rithm exhibits excellent performance, with an accuracy of " + }, + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "inline_equation", + "content": "99.4\\%" + }, + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "text", + "content": " on MNIST, " + }, + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "inline_equation", + "content": "93.05\\%" + }, + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "text", + "content": " on CIFAR-10, and " + }, + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "inline_equation", + "content": "69.66\\%" + }, + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "text", + "content": " on CIFAR-100. Obviously, our method outperforms all previous direct learning algorithms of TTFS-based SNNs. Moreover, our method still maintains competitive sparsity when achieving high performance. The sparsity is calculated by " + }, + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "inline_equation", + "content": "S / N" + }, + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "text", + "content": " averaged on the entire testing set, where " + }, + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "text", + "content": " represents the total number of spikes and " + }, + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 307, + 547, + 522 + ], + "type": "text", + "content": " denotes the total number of neurons. Specifically, the DTA-TTFS achieves sparsity of 0.3913 on MNIST, 0.2561 on CIFAR-10, and 0.2845 on CIFAR-100. For the MNIST dataset, when employing the identical network structure as described in [58], our method achieves the same accuracy while significantly reducing the sparsity from 0.6614 to 0.3913. For CIFAR datasets, our method surpasses related works in terms of accuracy while maintaining a competitive sparsity. In summary, the DTA-TTFS algorithm achieves a harmonious balance between accuracy and sparsity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 536, + 407, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 407, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 407, + 548 + ], + "type": "text", + "content": "4.3. Validation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 639 + ], + "type": "text", + "content": "We conduct a series of validation experiments to confirm the functionality of the proposed DFT. These experiments aim to display two aspects. Firstly, the DFT provides SNNs with the layer-dependent firing feature. Additionally, it effectively addresses the over-sparsity of spikes. Validation experiments choose the well-trained VGG11 structure on the CIFAR-10 dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "In order to show the layer-dependent firing feature of the SNN with DFT, we record the spike activity in each layer for the first sample of the testing set. For the purpose of comparison, we also provide the spike activity of the SNN without DFT. The results are visualized in Fig. 5, elucidating the evident regulation of spike activity in each layer" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 693, + 175, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 175, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 175, + 703 + ], + "type": "text", + "content": "2 http://yann.lecun.com/exdb/mnist/" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 703, + 200, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 200, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 200, + 712 + ], + "type": "text", + "content": "3https://www.cs.toronto.edu/ kriz/cifar.html" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10557" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 278, + 201 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 278, + 201 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 278, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 278, + 201 + ], + "type": "image", + "image_path": "b460326f15f624006660d7611722d9c9bd15fe4c2efe06e757da78131de887b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 208, + 279, + 340 + ], + "blocks": [ + { + "bbox": [ + 47, + 208, + 279, + 340 + ], + "lines": [ + { + "bbox": [ + 47, + 208, + 279, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 208, + 279, + 340 + ], + "type": "image", + "image_path": "765d6467c54f485876cd3f38e9fffb4844a792cbd0afc4092506dc4dd11c8723.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 342, + 287, + 365 + ], + "lines": [ + { + "bbox": [ + 47, + 342, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 287, + 365 + ], + "type": "text", + "content": "Figure 5. Spike activity in each layer of SNN with and without DFT. The vertical axis represents the neuron index scaled to [0,1]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 387, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 422 + ], + "type": "text", + "content": "within a permitted time window for the SNN incorporating DFT. In contrast, the spike activity in the SNN without DFT exhibits a disorderly and chaotic pattern." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 423, + 288, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 288, + 555 + ], + "type": "text", + "content": "In order to verify the efficacy of the DFT in addressing the over-sparsity of spikes, we count the number of inactive neurons per layer averaged on the whole testing set. This experiment is conducted on three models: the ANN, the SNN with DFT, and the SNN without DFT [58]. As shown in Fig. 6, the SNN without DFT displays significant sparsity, however, with only " + }, + { + "bbox": [ + 46, + 423, + 288, + 555 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 423, + 288, + 555 + ], + "type": "text", + "content": " accuracy. Conversely, the SNN with DFT exhibits a comparable amount of inactive neurons to that of ANN. As a result, the DFT achieves a harmonious trade-off between sparse representation and information representation ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 565, + 157, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 565, + 157, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 565, + 157, + 578 + ], + "type": "text", + "content": "4.4. Energy Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 585, + 168, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 168, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 168, + 597 + ], + "type": "text", + "content": "4.4.1 Theoretical Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "text", + "content": "Energy consumption is a critical metric to evaluate the efficiency of SNNs. To prove the energy efficiency of the proposed DTA-TTFS, we employ the theoretical analysis presented in [37]. This method analyzes energy consumption based on multiple factors, including latency, spike counts, as well as data related to neuromorphic architecture. More specifically, the estimated energy is defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 701, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 701, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 52, + 701, + 287, + 715 + ], + "type": "interline_equation", + "content": "E _ {\\text {T o t a l}} = E _ {\\text {S t a t i c}} \\times \\text {L a t e n c y} + E _ {\\text {D y n a m i c}} \\times \\text {S p i k e s}, \\tag {11}", + "image_path": "78b8f13a4f11ce7b2971608506df12ad885cdea013574b833c9e5c3eb3455bf7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 545, + 205 + ], + "type": "image", + "image_path": "9ff63c4767d4b2ffdcdbefa9152c1788863df93f74c0ee37c4aa662e05619d35.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 208, + 545, + 231 + ], + "lines": [ + { + "bbox": [ + 306, + 208, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 306, + 208, + 545, + 231 + ], + "type": "text", + "content": "Figure 6. Percentage of inactive neurons per layer of the ANN, the SNN with DFT, and the SNN without DFT." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 251, + 546, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 546, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 546, + 455 + ], + "type": "text", + "content": "where Latency is the required time steps, Spikes is the total number of spikes, " + }, + { + "bbox": [ + 304, + 251, + 546, + 455 + ], + "type": "inline_equation", + "content": "E_{Static}" + }, + { + "bbox": [ + 304, + 251, + 546, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 251, + 546, + 455 + ], + "type": "inline_equation", + "content": "E_{Dynamic}" + }, + { + "bbox": [ + 304, + 251, + 546, + 455 + ], + "type": "text", + "content": " are static and dynamic energy coefficients depending on the neuromorphic architecture. We choose two neuromorphic architectures, TrueNorth [32] and SpiNNaker [16], to conduct analysis. The energy parameters " + }, + { + "bbox": [ + 304, + 251, + 546, + 455 + ], + "type": "inline_equation", + "content": "(E_{Static}, E_{Dynamic})" + }, + { + "bbox": [ + 304, + 251, + 546, + 455 + ], + "type": "text", + "content": " are set to (0.6, 0.4) for TrueNorth and (0.36, 0.64) for SpiNNaker. For comparison, we analyze various coding schemes, such as rate coding [19], phase coding [26], burst coding [36], and TTFS coding [37, 38]. Theoretical analysis is performed on CIFAR-10 with the well-trained VGG16 structure, and we normalized estimated energy consumption based on the rate coding. As indicated in Tab. 2, under the same conditions, the DTA-TTFS outperforms other methods in terms of energy efficiency. It is worth noting that, despite the fewer spikes achieved by [37, 38], their energy consumption remains higher than ours due to their demand for long latency." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 308, + 464, + 545, + 566 + ], + "blocks": [ + { + "bbox": [ + 308, + 464, + 545, + 566 + ], + "lines": [ + { + "bbox": [ + 308, + 464, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 464, + 545, + 566 + ], + "type": "table", + "html": "
Neural CodingTime StepSpike (106)Acc. (%)Normalized Energy
TrueNorthSpiNNaker
Rate [19]5122.61293.3911
Phase [26]150035.19691.217.14769.6785
Burst [36]11256.9291.412.37812.4865
TTFS [37]6800.06991.430.80740.4950
TTFS [38]5440.06791.90.64780.3989
DTA-TTFS1600.07393.050.19870.1304
", + "image_path": "e5ec04af19f1b0e6b8c42b33f3fbe53269532b134b67ee067364029f214060bc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 329, + 569, + 523, + 581 + ], + "lines": [ + { + "bbox": [ + 329, + 569, + 523, + 581 + ], + "spans": [ + { + "bbox": [ + 329, + 569, + 523, + 581 + ], + "type": "text", + "content": "Table 2. Theoretical analysis of energy consumption." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 609, + 432, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 432, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 432, + 621 + ], + "type": "text", + "content": "4.4.2 Hardware Simulation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "In order to further verify the energy efficiency and hardware friendliness of our method, we map the SNN trained by the DTA-TTFS algorithm on a recently developed neuromorphic hardware simulator, called configurable asynchronous neuromorphic hardware simulator (CanMore) [55]. CanMore estimates hardware performance through system-level modeling and simulation. Fig. 7 illustrates the main" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10558" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 289, + 198 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 289, + 198 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 289, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 289, + 198 + ], + "type": "image", + "image_path": "5274cbf7e6d8854be059704150bc939ca36f94baf27acc47f8c38e3507dca2c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 287, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 245 + ], + "type": "text", + "content": "Figure 7. The framework of CanMore [55], including neuromorphic cores comprised of neurons and SRAM, and NoC routers consisting of buffers, route computation units, switch allocator, and crossbar switch." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 263, + 287, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 383 + ], + "type": "text", + "content": "framework of the CanMore, including neuromorphic cores comprised of neurons and SRAM, and network-on-chip (NoC) routers consisting of buffers, route computation unit, switch allocator and crossbar switch. With the system-level asynchronous circuit modeling, CanMore is able to simulate various neuromorphic hardware architectures by changing the parameters and configuration files, estimate latency and energy consumption, and analyze spike messages routing information such as routing hops, congestion latency, routing directions, and so on." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 575 + ], + "type": "text", + "content": "In [13], a deep ANN is mapped to TrueNorth neuromorphic hardware. To the best of our knowledge, this is the only work that verifies the energy efficiency of SNNs on the CIFAR-10 dataset, while the majority of existing works focus on smaller datasets, such as MNIST [12, 58]. Here, we train an SNN on the CIFAR-10 dataset with the proposed DTA-TTFS method. After learning, the learned weights are transferred to CanMore for accelerating the inference operations. Specifically, our network structure is mapped by partitioning each layer into equally sized groups along the feature dimension, with each group being mapped to a neuromorphic core with a maximum of 1024 neurons. In addition, we compress the network by quantizing the weights to 4 bits. The mesh topology and XY routing algorithm are adopted in the simulated architecture. The energy consumption per frame during the inference process is calculated by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 584, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 584, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 52, + 584, + 287, + 597 + ], + "type": "interline_equation", + "content": "E _ {T o t a l} = E _ {S O P} \\times \\text {S p i k e s} + E _ {H o p} \\times \\text {H o p s} + E _ {\\text {S t a t i c}}, \\tag {12}", + "image_path": "14b43972e8615ce6d5d89d10214d500a6078d863f110871be5301fdd08a14da9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "E_{SOP}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " is the dynamic energy consumption per synaptic operation (SOP), which is calculated with the real technology parameters of the CMOS 22nm process in the simulation. In our work, " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "E_{SOP}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "20\\mathrm{pJ}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "E_{Hop}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " is the dynamic energy consumption per hop between the mesh routers on the NoC, which is " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "100\\mathrm{pJ}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " according to [21]. " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "E_{Static}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " is static energy consumption, which is " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "3.52\\mu \\mathrm{J}/\\mathrm{frame}" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": " in CMOS 22nm process. The average Spikes and Hops per frame during the inference process are 71M and 45.4M re" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 68, + 420, + 212 + ], + "blocks": [ + { + "bbox": [ + 310, + 68, + 420, + 212 + ], + "lines": [ + { + "bbox": [ + 310, + 68, + 420, + 212 + ], + "spans": [ + { + "bbox": [ + 310, + 68, + 420, + 212 + ], + "type": "image", + "image_path": "30ea7d1644ab9eca5b950198984c2e5bbe5fab515b084f649c5cfa991da6616a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 213, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 306, + 213, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 545, + 236 + ], + "type": "text", + "content": "Figure 8. The accuracy and energy consumption of Esser et al. [13] and DTA-TTFS on CIFAR-10 dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 541, + 212 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 541, + 212 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 541, + 212 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 541, + 212 + ], + "type": "image", + "image_path": "e88121b7bffd7f5e17f2a83957cdeecfd52034d77883c573bf47f33d6df05453.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 255, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 290 + ], + "type": "text", + "content": "spectively, given by CanMore. Therefore, we obtain the energy consumption " + }, + { + "bbox": [ + 304, + 255, + 545, + 290 + ], + "type": "inline_equation", + "content": "E_{Total}" + }, + { + "bbox": [ + 304, + 255, + 545, + 290 + ], + "type": "text", + "content": " during the inference process " + }, + { + "bbox": [ + 304, + 255, + 545, + 290 + ], + "type": "inline_equation", + "content": "50.36\\mu \\mathrm{J / frame}" + }, + { + "bbox": [ + 304, + 255, + 545, + 290 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 290, + 546, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 546, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 546, + 386 + ], + "type": "text", + "content": "The accuracy and energy consumption are reported in Fig. 8. our work achieves a classification accuracy of " + }, + { + "bbox": [ + 304, + 290, + 546, + 386 + ], + "type": "inline_equation", + "content": "90.67\\%" + }, + { + "bbox": [ + 304, + 290, + 546, + 386 + ], + "type": "text", + "content": " on CIFAR-10, surpassing the accuracy reported in [13] by " + }, + { + "bbox": [ + 304, + 290, + 546, + 386 + ], + "type": "inline_equation", + "content": "7.26\\%" + }, + { + "bbox": [ + 304, + 290, + 546, + 386 + ], + "type": "text", + "content": ". In addition, the energy consumption of our work is " + }, + { + "bbox": [ + 304, + 290, + 546, + 386 + ], + "type": "inline_equation", + "content": "50.36~\\mu \\mathrm{J} / \\mathrm{frame}" + }, + { + "bbox": [ + 304, + 290, + 546, + 386 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 304, + 290, + 546, + 386 + ], + "type": "inline_equation", + "content": "69.2\\%" + }, + { + "bbox": [ + 304, + 290, + 546, + 386 + ], + "type": "text", + "content": " reduced compared with the energy consumption " + }, + { + "bbox": [ + 304, + 290, + 546, + 386 + ], + "type": "inline_equation", + "content": "163.65\\mu \\mathrm{J} / \\mathrm{frame}" + }, + { + "bbox": [ + 304, + 290, + 546, + 386 + ], + "type": "text", + "content": " in [13]. These results confirm that the proposed DTA-TTFS algorithm is effective and energy-efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 398, + 378, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 398, + 378, + 410 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 378, + 410 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 418, + 546, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 546, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 546, + 622 + ], + "type": "text", + "content": "In this study, we analyze the shortcomings of existing methods in achieving high-performance TTFS-based deep SNNs. In order to overcome these issues, we propose an efficient DFT mechanism for spiking neurons and introduce the DTA-TTFS algorithm for TTFS-based SNNs. We evaluate our method on the image classification task, and the results demonstrate that our method achieves a fine trade-off between accuracy and sparsity. Additionally, we estimate the energy consumption of our method through theoretical analysis and the neuromorphic hardware accelerator CanMore. Both theoretical analysis and hardware verification confirm the energy efficiency of the DTA-TTFS algorithm. Overall, our method not only attains high performance but also is quite hardware friendly. In our future work, we intend to leverage DTA-TTFS to train more advanced network models, including SNN-based Attention [64, 51] and Transformer architectures [63]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 633, + 399, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 399, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 399, + 647 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "This work was supported by the National Science Foundation of China under Grants 62236007 and 61976043 and 62106038, and in part by the Sichuan Science and Technology Program under Grants 2022YFG0313 and 2023YFG0259." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10559" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 168 + ], + "type": "text", + "content": "[1] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34(10):1537-1557, 2015." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "type": "text", + "content": "[2] Sander M Bohte, Joost N Kok, and Johannes A La Poutre. Spikeprop: backpropagation for networks of spiking neurons. In *ESANN*, volume 48, pages 419–424. Bruges, 2000." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 286, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 286, + 258 + ], + "type": "text", + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 261, + 286, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 286, + 304 + ], + "type": "text", + "content": "[4] Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Optimized potential initialization for low-latency spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11-20, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 307, + 286, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 286, + 350 + ], + "type": "text", + "content": "[5] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 352, + 286, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 286, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 286, + 417 + ], + "type": "text", + "content": "[6] Iulia-Maria Comsa, Krzysztof Potempa, Luca Versari, Thomas Fischbacher, Andrea Gesmundo, and Jyrki Alakuijala. Temporal coding in spiking neural networks with alpha synaptic function: learning with backpropagation. IEEE transactions on neural networks and learning systems, 33(10):5939-5952, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 419, + 286, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 286, + 474 + ], + "type": "text", + "content": "[7] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 475, + 286, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 286, + 530 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 286, + 530 + ], + "type": "text", + "content": "[8] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 38(1):82-99, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 532, + 286, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 532, + 286, + 576 + ], + "spans": [ + { + "bbox": [ + 53, + 532, + 286, + 576 + ], + "type": "text", + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 286, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 286, + 611 + ], + "type": "text", + "content": "[10] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 667 + ], + "type": "text", + "content": "[11] Peter U Diehl, Daniel Neil, Jonathan Binas, Matthew Cook, Shih-Chii Liu, and Michael Pfeiffer. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "text", + "content": "[12] Steve K Esser, Rathinakumar Appuswamy, Paul Merolla, John V Arthur, and Dharmendra S Modha. Backpropagation for energy-efficient neuromorphic computing. Advances in neural information processing systems, 28, 2015." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "type": "text", + "content": "[13] Steven K Esser, Paul A Merolla, John V Arthur, Andrew S Cassidy, Rathinakumar Appuswamy, Alexander Andreopoulos, David J Berg, Jeffrey L McKinstry, Timothy Melano, Davis R Barch, et al. From the cover: Convolutional networks for fast, energy-efficient neuromorphic computing. Proceedings of the National Academy of Sciences of the United States of America, 113(41):11441, 2016." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "text", + "content": "[14] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "type": "text", + "content": "[15] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothée Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2661-2671, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 250, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 282 + ], + "type": "text", + "content": "[16] Steve B Furber, Francesco Galluppi, Steve Temple, and Luis A Plana. The spinnaker project. Proceedings of the IEEE, 102(5):652-665, 2014." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 283, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 337 + ], + "type": "text", + "content": "[17] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 315-323. JMLR Workshop and Conference Proceedings, 2011." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 380 + ], + "type": "text", + "content": "[18] Alex Graves and Navdeep Jaitly. Towards end-to-end speech recognition with recurrent neural networks. In International conference on machine learning, pages 1764-1772. PMLR, 2014." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 382, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 446 + ], + "type": "text", + "content": "[19] Bing Han, Gopalakrishnan Srinivasan, and Kaushik Roy. Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13558–13567, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 449, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 481 + ], + "type": "text", + "content": "[20] Yangfan Hu, Huajin Tang, and Gang Pan. Spiking deep residual networks. IEEE Transactions on Neural Networks and Learning Systems, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 482, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 547 + ], + "type": "text", + "content": "[21] Masashi Imai, Thiem Van Chu, Kenji Kise, and Tomohiro Yoneda. The synchronous vs. asynchronous noc routers: an apple-to/apple comparison between synchronous and transition signaling asynchronous designs. In 2016 Tenth IEEE/ACM International Symposium on Networks-on-Chip (NOCS), pages 1-8. IEEE, 2016." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 548, + 545, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 579 + ], + "type": "text", + "content": "[22] Eugene M Izhikevich. Simple model of spiking neurons. IEEE Transactions on neural networks, 14(6):1569-1572, 2003." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 581, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 545, + 635 + ], + "type": "text", + "content": "[23] Yeying Jin, Aashish Sharma, and Robby T Tan. Dshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5027-5036, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 636, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 680 + ], + "type": "text", + "content": "[24] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In European Conference on Computer Vision, pages 404-421. Springer, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[25] S. R. Kheradpisheh and T. Masquelier. Temporal backpropagation for spiking neural networks with one spike per neuron. International Journal of Neural Systems, 30(6), 2020." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10560" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[26] Jaehyun Kim, Heesu Kim, Subin Huh, Jinho Lee, and Kiyoung Choi. Deep neural networks with weighted spikes. Neurocomputing, 311:373-386, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "text", + "content": "[27] Jinyu Li et al. Recent advances in end-to-end automatic speech recognition. APSIPA Transactions on Signal and Information Processing, 11(1), 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 286, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 286, + 185 + ], + "type": "text", + "content": "[28] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International conference on machine learning, pages 6316-6325. PMLR, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 239 + ], + "type": "text", + "content": "[29] Yuhang Li, Yufei Guo, Shanghang Zhang, Shikuang Deng, Yongqing Hai, and Shi Gu. Differentiable spike: Rethinking gradient-descent for training spiking neural networks. Advances in Neural Information Processing Systems, 34:23426-23439, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 243, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 243, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 286, + 285 + ], + "type": "text", + "content": "[30] Yuguo Liu, Wenyu Chen, Hanwen Liu, Yun Zhang, Malu Zhang, and Hong Qu. Biologically plausible sparse temporal word representations. IEEE Transactions on Neural Networks and Learning Systems, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 287, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 287, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 287, + 286, + 319 + ], + "type": "text", + "content": "[31] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10(9):1659-1671, 1997." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 321, + 286, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 321, + 286, + 386 + ], + "spans": [ + { + "bbox": [ + 49, + 321, + 286, + 386 + ], + "type": "text", + "content": "[32] Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668-673, 2014." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 388, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 286, + 430 + ], + "type": "text", + "content": "[33] Hesham Mostafa. Supervised learning based on temporal coding in spiking neural networks. IEEE transactions on neural networks and learning systems, 29(7):3227-3235, 2017." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 434, + 286, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 286, + 487 + ], + "type": "text", + "content": "[34] Emre O Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, 36(6):51-63, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 490, + 286, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 490, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 490, + 286, + 544 + ], + "type": "text", + "content": "[35] Zihan Pan, Malu Zhang, Jibin Wu, Jiadong Wang, and Haizhou Li. Multi-tone phase coding of interaural time difference for sound source localization with spiking neural networks. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:2656-2670, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 545, + 286, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 545, + 286, + 599 + ], + "spans": [ + { + "bbox": [ + 49, + 545, + 286, + 599 + ], + "type": "text", + "content": "[36] Seongsik Park, Seijoon Kim, Hyeokjun Choe, and Sungroh Yoon. Fast and efficient information transmission with burst spikes in deep spiking neural networks. In Proceedings of the 56th Annual Design Automation Conference 2019, pages 1-6, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 601, + 286, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 286, + 645 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 286, + 645 + ], + "type": "text", + "content": "[37] Seongsik Park, Seijoon Kim, Byunggook Na, and Sungroh Yoon. T2fsnn: Deep spiking neural networks with time-to-first-spike coding. In 2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 647, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 286, + 678 + ], + "type": "text", + "content": "[38] Seongsik Park and Sungroh Yoon. Training energy-efficient deep spiking neural networks with time-to-first-spike coding. arXiv preprint arXiv:2106.02568, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 681, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 286, + 713 + ], + "type": "text", + "content": "[39] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 97, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 139 + ], + "type": "text", + "content": "[40] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, and Michael Pfeiffer. Theory and tools for the conversion of analog to spiking convolutional neural networks. arXiv preprint arXiv:1612.04052, 2016." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 141, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 194 + ], + "type": "text", + "content": "[41] Bodo Rueckauer, Iulia-Alexandra Lungu, Yuhuang Hu, Michael Pfeiffer, and Shih-Chii Liu. Conversion of continuous-valued deep networks to efficient event-driven networks for image classification. Frontiers in neuroscience, 11:682, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 198, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 230 + ], + "type": "text", + "content": "[42] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 263 + ], + "type": "text", + "content": "[43] V. Sze, Y. H. Chen, T. J. Yang, and J. S. Emer. Efficient processing of deep neural networks: A tutorial and survey. Proceedings of the IEEE, 105(12), 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 266, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 266, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 545, + 308 + ], + "type": "text", + "content": "[44] Yuchen Wang, Malu Zhang, Yi Chen, and Hong Qu. Signed neuron with memory: Towards simple, accurate and high-efficient ann-snn conversion. In International Joint Conference on Artificial Intelligence, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 311, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 311, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 311, + 545, + 364 + ], + "type": "text", + "content": "[45] Jibin Wu, Yansong Chua, Malu Zhang, Guoqi Li, Haizhou Li, and Kay Chen Tan. A tandem learning rule for effective training and rapid inference of deep spiking neural networks. IEEE Transactions on Neural Networks and Learning Systems, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 366, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 421 + ], + "type": "text", + "content": "[46] Jibin Wu, Chenglin Xu, Xiao Han, Daquan Zhou, Malu Zhang, Haizhou Li, and Kay Chen Tan. Progressive tandem learning for pattern recognition with deep spiking neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7824-7840, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "text", + "content": "[47] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "text", + "content": "[48] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 1311-1318, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 514, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 567 + ], + "type": "text", + "content": "[49] Qi Xu, Yaxin Li, Jiangrong Shen, Jian K Liu, Huajin Tang, and Gang Pan. Constructing deep spiking neural networks from artificial neural networks with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7886-7895, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 623 + ], + "type": "text", + "content": "[50] Qi Xu, Yaxin Li, Jiangrong Shen, Pingping Zhang, Jian K Liu, Huajin Tang, and Gang Pan. Hierarchical spiking-based model for efficient image classification with enhanced feature extraction and encoding. IEEE Transactions on Neural Networks and Learning Systems, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "text", + "content": "[51] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence, 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[52] Qiang Yu, Huajin Tang, Kay Chen Tan, and Haizhou Li. Rapid feedforward computation by temporal encoding and learning with spiking neurons. IEEE transactions on neural networks and learning systems, 24(10):1539-1552, 2013." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10561" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[53] Friedemann Zenke and Tim P Vogels. The remarkable robustness of surrogate gradient learning for instilling complex function in spiking neural networks. Neural computation, 33(4):899-925, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 194 + ], + "type": "text", + "content": "[54] Jilin Zhang, Dexuan Huo, Jian Zhang, Chunqi Qian, Qi Liu, Liyang Pan, Zhihua Wang, Ning Qiao, Kea-Tiong Tang, and Hong Chen. 22.6 anp-i: A 28nm 1.5 pj/sop asynchronous spiking neural network processor enabling sub-o. 1 " + }, + { + "bbox": [ + 48, + 118, + 287, + 194 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 48, + 118, + 287, + 194 + ], + "type": "text", + "content": "j/sample on-chip learning for edge-ai applications. In 2023 IEEE International Solid-State Circuits Conference (ISSCC), pages 21–23. IEEE, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 195, + 286, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 286, + 258 + ], + "type": "text", + "content": "[55] Jian Zhang, Jilin Zhang, Dexuan Huo, and Hong Chen. Anas: Asynchronous neuromorphic hardware architecture search based on a system-level simulator. in proc. 60th acm/ieee design automation conference (dac), 2023. In Proc. 60th ACM/IEEE Design Automation Conference (DAC), 2023, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 261, + 286, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 286, + 316 + ], + "type": "text", + "content": "[56] Malu Zhang, Xiaoling Luo, Yi Chen, Jibin Wu, Ammar Belatreche, Zihan Pan, Hong Qu, and Haizhou Li. An efficient threshold-driven aggregate-label learning algorithm for multimodal information processing. IEEE Journal of Selected Topics in Signal Processing, 14(3):592-602, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 317, + 286, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 286, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 286, + 370 + ], + "type": "text", + "content": "[57] Malu Zhang, Hong Qu, Ammar Belatreche, Yi Chen, and Zhang Yi. A highly effective and robust membrane potential-driven supervised learning method for spiking neurons. IEEE transactions on neural networks and learning systems, 30(1):123-137, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 372, + 286, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 286, + 447 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 286, + 447 + ], + "type": "text", + "content": "[58] Malu Zhang, Jiadong Wang, Jibin Wu, Ammar Belatreche, Burin Amornpaisannon, Zhixuan Zhang, Venkata Pavan Kumar Miriyala, Hong Qu, Yansong Chua, Trevor E Carlson, et al. Rectified linear postsynaptic potential function for backpropagation in deep spiking neural networks. IEEE transactions on neural networks and learning systems, 33(5):1947-1958, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 449, + 286, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 286, + 491 + ], + "type": "text", + "content": "[59] Wenrui Zhang and Peng Li. Temporal spike sequence learning via backpropagation for deep spiking neural networks. Advances in Neural Information Processing Systems, 33:12022-12033, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 494, + 286, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 494, + 286, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 494, + 286, + 548 + ], + "type": "text", + "content": "[60] Shibo Zhou, Xiaohua Li, Ying Chen, Sanjeev T Chandrasekaran, and Arindam Sanyal. Temporal-coded deep spiking neural network with easy training and robust performance. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 11143-11151, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 548, + 286, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 286, + 647 + ], + "type": "text", + "content": "[61] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, pages 3-11. Springer, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 647, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 712 + ], + "type": "text", + "content": "[62] Zongwei Zhou, Jae Shin, Lei Zhang, Suryakanth Gurudu, Michael Gotway, and Jianming Liang. Fine-tuning convolutional neural networks for biomedical image analysis: actively and incrementally. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7340-7351, 2017." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 228 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[63] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, YAN Shuicheng, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. In The Eleventh International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[64] Rui-Jie Zhu, Qihang Zhao, Tianjing Zhang, Haoyu Deng, Yule Duan, Malu Zhang, and Liang-Jian Deng. Tcjasnn: Temporal-channel joint attention for spiking neural networks. arXiv preprint arXiv:2206.10177, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "text", + "content": "[65] Yaoyu Zhu, Zhaofei Yu, Wei Fang, Xiaodong Xie, Tiejun Huang, and Timothée Masquelier. Training spiking neural networks with event-driven backpropagation. In 36th Conference on Neural Information Processing Systems (NeurIPS 2022), 2022." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10562" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Test Time Adaptation for Blind Image Quality Assessment/4f23b86f-f4db-4257-aee1-6b77c4f73657_content_list.json b/2023/Test Time Adaptation for Blind Image Quality Assessment/4f23b86f-f4db-4257-aee1-6b77c4f73657_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..399867e1b8f1f89bb08ef7eb0056c57363d7dba7 --- /dev/null +++ b/2023/Test Time Adaptation for Blind Image Quality Assessment/4f23b86f-f4db-4257-aee1-6b77c4f73657_content_list.json @@ -0,0 +1,1516 @@ +[ + { + "type": "text", + "text": "Test Time Adaptation for Blind Image Quality Assessment", + "text_level": 1, + "bbox": [ + 191, + 130, + 779, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Subhadeep Roy* Shankhanil Mitra* Soma Biswas Rajiv Soundararajan Indian Institute of Science, Bengaluru, India", + "bbox": [ + 181, + 180, + 785, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "subhadeeproy2000@gmail.com, {shankhanilm, somabiswas, rajivs}@iisc.ac.in", + "bbox": [ + 166, + 219, + 800, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While the design of blind image quality assessment (IQA) algorithms has improved significantly, the distribution shift between the training and testing scenarios often leads to a poor performance of these methods at inference time. This motivates the study of test time adaptation (TTA) techniques to improve their performance at inference time. Existing auxiliary tasks and loss functions used for TTA may not be relevant for quality-aware adaptation of the pre-trained model. In this work, we introduce two novel quality-relevant auxiliary tasks at the batch and sample levels to enable TTA for blind IQA. In particular, we introduce a group contrastive loss at the batch level and a relative rank loss at the sample level to make the model quality aware and adapt to the target data. Our experiments reveal that even using a small batch of images from the test distribution helps achieve significant improvement in performance by updating the batch normalization statistics of the source model.", + "bbox": [ + 75, + 300, + 473, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 584, + 209, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The problem of image quality assessment (IQA) is extremely important in diverse image capture, processing, and sharing applications. However, a reference image is often not available for quality assessment. No reference (NR) or blind IQA primarily deals with the question of predicting image quality without using a reference image. Such NR IQA algorithms are often designed using machine learning approaches. More recently, deep learning based approaches have been extremely successful in achieving impressive performance. However, IQA applications are quite diverse and deal with several different distortions and distributional shifts. IQA models often have poor generalization ability and find it difficult to perform well under such shifts.", + "bbox": [ + 75, + 609, + 468, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Test time adaptation (TTA) has emerged as an important approach to address distributional shifts at test time [8]. It has been shown that by modifying a few global parameters of the model using a suitable loss that does not require the", + "bbox": [ + 75, + 806, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ground truth, one can significantly improve the performance of the model on the test data. Further, source-free adaptation, where the source data on which the original model was trained is not available while updating the model, is a realistic setting. While such approaches have been studied extensively in image classification literature [31, 33], there is hardly any literature on TTA for IQA.", + "bbox": [ + 500, + 270, + 890, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are multiple challenges in designing TTA for IQA. Typical losses used for TTA, such as entropy minimization [33], are not applicable for IQA. For example, IQA is often studied in the regression context. This makes it difficult to extend models based on class confidences [24] or class prototypes [10] for classification to IQA. Also, the relevance of other self-supervised tasks such as rotation prediction [14], context prediction [5], colorization [15], noise prediction [1], feature clustering [3] for adapting IQA models is not clear. While contrastive learning has also been employed for TTA [20], such a framework is not explicitly based on contrasting image quality, and its relevance is also not clear.", + "bbox": [ + 496, + 378, + 892, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our main contribution is in the design of auxiliary tasks to enable TTA for IQA. We start with a source model trained on a large IQA dataset and fine-tune the model on individual batches of test samples. The first task we introduce for adaptation is based on contrasting groups of low and high quality images in a batch. Thus, we exploit the initial knowledge of the source model and try to adapt it by enforcing quality relationships among the batch of samples. Such a group contrastive (GC) learning approach fits naturally to our setting to account for any errors on individual samples that the source model may be prone to.", + "bbox": [ + 496, + 579, + 893, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast to the GC learning that depends on the batch, our second auxiliary task is an image specific task based on distorted augmentations of different types. Here, our goal is to enable the model to rank the image quality of further distorted versions of each test sample. We explore the role of different distortion types to leverage the maximum benefit of this task. While GC learning is more effective when samples in a batch are diverse in quality, the rank order based learning is more effective when the quality of the images is not extremely poor. Thus, a combination of the tasks", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors contributed equally to this work\n§https://github.com/subhadeeproy2000/TTA-IQA", + "bbox": [ + 91, + 875, + 442, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16742", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "helps overcome the shortcoming of both tasks and leads to an overall superior performance.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We study the TTA problem under different settings of source and target datasets for multiple state of the art IQA models. Our results show significant improvements of the source model and the importance of TTA for IQA. We summarize our main contributions as follows:", + "bbox": [ + 76, + 122, + 468, + 198 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose source-free test time adaptation techniques in the context of blind image quality assessment to mitigate distribution shifts between train and test data.", + "- We formulate two novel quality-aware self-supervised auxiliary tasks to adapt to the test data distribution. While group contrastive learning helps capture quality discriminative information among several images within a batch, rank ordering helps maintain the quality order between two different distorted versions of the same image.", + "- We show that our TTA method can significantly improve the performance of four different quality-aware source models, each on four different test IQA databases." + ], + "bbox": [ + 94, + 218, + 467, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 515, + 217, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Test Time Adaptation", + "text_level": 1, + "bbox": [ + 76, + 541, + 279, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One of the first pieces of work on TTA [31] introduces a joint training framework using a loss for the main task and a self-supervised auxiliary task loss. The choice of a relevant auxiliary task is a challenging part of the design. Sun et al. [31] use rotation prediction as a pre-text task for image classification applications. Researchers also explore simpler tasks that are highly correlated with the main task, such as feature alignment with the batch and a simple contrastive loss framework [20] for adaptation. However, these methods require the source data to train the source model all over again to enable TTA.", + "bbox": [ + 75, + 566, + 467, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "TENT [33] studies source-free TTA, where entropy minimization-based adaptation even outperforms some methods that use the source data. Moreover, only the batch normalization parameters of the model are adapted in TENT. There are several works on batch norm statistics adaptation [25, 28] to improve the robustness of the model at test time. SHOT [16] presents a clustering-based pseudolabeling method to align features from the target domain to the source domain using an information maximization loss. While a plethora of methods exists as above, the auxiliary tasks in these methods are not relevant for IQA.", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. No-Reference Image Quality Assessment", + "text_level": 1, + "bbox": [ + 500, + 90, + 846, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most classical NR IQA algorithms are mainly based on natural scene statistics (NSS) [19, 22, 38]. In [23], the naturalness of the distorted image in the wavelet domain is modeled based on NSS. Saad et al. [27] also design the NSS model in the discrete cosine transform (DCT) domain. CORNIA [35], and HOSA [34] are among the earliest codebook learning based methods to predict quality. With the emergence of deep learning and the availability of large subject-rated IQA databases, various general-purpose NR IQA methods have been designed based on convolutional neural network (CNN) architectures [2, 18, 21, 36]. Some of these methods require end-to-end training [2, 21] of deep neural networks (DNN), while others [7, 12, 30, 39] are based on updating pre-trained models with some modifications.", + "bbox": [ + 496, + 113, + 890, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, Zhang et al. [39] propose a method to train bilinear DNNs to simultaneously model both authentic and synthetic distortions. Su et al. [30] design a self-adaptive hyper network to provide weights for the quality prediction module parameters and handle various types of distortions and content in the images. Several methods [7, 12] explore transformer-based architectures along with a CNN to capture dependencies between local and global features. MetaIQA [40] proposes meta-learning on synthetic distortions by using a shared quality prior knowledge model to adapt to any kind of distortion.", + "bbox": [ + 496, + 340, + 890, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All the existing methods assume that the train and test data come from the same distribution. If there is a distribution shift across different databases, we need adaptation of the pre-trained model to learn about the target distributional information.", + "bbox": [ + 496, + 507, + 890, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 595, + 633, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a novel self-supervised Test Time Adaptation technique for Image Quality Assessment (TTA-IQA) to adapt pre-trained quality models and mitigate distribution shifts between the source and target data. We consider source-free TTA, where we only have access to the pre-trained quality-aware models and no access to the source training data. When a batch of test data $D = \\{x_{j}\\}_{j=1}^{n}$ arrives, we adapt the model using the batch without knowledge of corresponding ground truth $\\{y_{j}\\}_{j=1}^{n}$ .", + "bbox": [ + 496, + 621, + 890, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Approach", + "text_level": 1, + "bbox": [ + 500, + 771, + 612, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Let the model trained on the source data be $f_{\\theta}$ where $\\theta = (\\theta_e, \\theta_c)$ corresponds to the parameters of the network, and $\\theta_e$ and $\\theta_c$ correspond to the parameters of the feature extractor and regression layers. Thus, for an input image $x$ , we denote $f_{\\theta}(x) = f_{\\theta_c}(f_{\\theta_e}(x))$ . Since our goal is to learn the distribution shift between train and test data, we only update the parameters of the feature extractor, $\\theta_e$ , to", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16743", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c9a8509225d28991a43f53554069fc95fb969f3955e955998951bd882f811d93.jpg", + "image_caption": [ + "Figure 1. Block diagram of a general architecture for test time adaptation. At test time training, the normalization layers of the feature extractor are adapted by optimizing the combination of rank and GC loss. At inference time, we predict the quality scores of test images using the updated feature extractor and pre-trained quality regressor." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 869, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "align the features between the train and test distributions in a lower dimensional space.", + "bbox": [ + 75, + 323, + 467, + 353 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key challenge in TTA is the choice of a self-supervised auxiliary task that is highly correlated with the main task of IQA. However, relying too much on the auxiliary task can affect the performance of the main task. To prevent loss of learnt information induced by the auxiliary task [4], we first project the features to a lower dimensional space using a non-linear projection head $f_{\\theta_s}$ , parameterized by $\\theta_s$ . The auxiliary task drives the adaptation of the feature extractor through the projection head. Thus our model now has three parts parametrized by $(\\theta_e, \\theta_s, \\theta_c)$ to resemble a Y-shape as shown in Figure 1.", + "bbox": [ + 75, + 353, + 467, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When a batch of test instances $D$ arrives, we extract features from the feature extractor, followed by the projection head, and update the set of parameters $(\\theta_e,\\theta_s)$ by optimizing a self-supervised objective function $\\mathcal{L}_s(D)$ . Updating all the model parameters of the feature extractor $\\theta_{e}$ can cause the model to diverge too much from training, and the performance can drop drastically. Inspired by prior work on test time entropy minimization [33] and improving robustness for test data [28], we only update the linear and lower-dimensional feature modulation parameters. In a neural network, normalization layers satisfy these properties. So we only adapt the batch normalization layers by updating the affine parameters to mitigate distributional shifts. Thus we update these parameters by optimizing the auxiliary task loss given by", + "bbox": [ + 75, + 518, + 467, + 746 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {e} ^ {*}, \\theta_ {s} ^ {*} = \\underset {\\left(\\theta_ {e}, \\theta_ {s}\\right)} {\\operatorname {a r g m i n}} \\mathcal {L} _ {s} (D) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 757, + 467, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After optimizing the above loss function, we use the updated feature extractor parameters $\\theta_e^*$ and pre-trained quality regressor parameters $\\theta_c$ to predict the quality for a batch of test data. Since the distribution across batches can vary significantly, we ignore the earlier updated model and start the TTA based on source model weights for new test instants. Thus the updated target model always only depends", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on the incoming test data.", + "bbox": [ + 500, + 324, + 671, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Self-Supervised Auxiliary Tasks", + "text_level": 1, + "bbox": [ + 500, + 348, + 779, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to carefully choose self-supervised auxiliary tasks that capture quality-aware distributional information to adapt the feature encoder. We formulate two novel and complementary self-supervised learning techniques which help to learn the distribution shift between train and test data. These self-supervised objective functions are - 1) Group Contrastive Loss and 2) Rank Loss. While the GC loss works well when there is a reasonable separation of quality among the samples in a batch, the rank loss works better even when the quality of the batch samples is similar. On the other hand, the rank loss is meaningful only when the quality of the input image is not extremely low, while the GC loss is independent of the quality of a given image. Thus, a combination of the two losses renders our TTA extremely effective across various scenarios.", + "bbox": [ + 496, + 371, + 890, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Group Contrastive Loss", + "text_level": 1, + "bbox": [ + 500, + 618, + 720, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While contrastive learning has been used for TTA of deep image classification, its direct application does not appear relevant to the task of IQA. Thus, we introduce group contrastive (GC) learning as an auxiliary task for TTA of IQA models. In particular, we make two groups of images from a single batch of $N$ images based on the pseudo-labels given by the pre-trained source model. We sort the images in ascending order as $x_{(1)}, x_{(2)}, \\ldots, x_{(N)}$ based on the pseudolabels, where $x_{(i)}, i = 1, 2, \\ldots, N$ , corresponds to the $i^{th}$ lowest quality image in the batch. We then segregate images with high quality scores (say, top $p$ fraction of the data in a batch) and include them in a group of higher-quality images. Similarly, we separate out lower quality images (say, the lowest $p$ fraction of the images in a batch) together to form another group. Here we assume that $pN$ and $(1 - p)N$ are integers for simplicity; else, they can be rounded off to the nearest integers.", + "bbox": [ + 496, + 642, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16744", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f7d8fac2d877a1359ceeb87ce57c4b7638faa7a5b704c0d452944e44000910ca.jpg", + "image_caption": [ + "Figure 2. The overview framework of group contrastive loss for test time adaptation. For a batch of images, two different groups are formed based on pseudo-labels given by the source model. The group contrastive loss tries to minimize the distance between features extracted from images belonging to the same group while maximizing the distance between features from different groups." + ], + "image_footnote": [], + "bbox": [ + 76, + 89, + 893, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The premise behind our loss for GC learning is that images from the same quality group should give similar feature representations in a lower dimensional space while features of images from different groups are separated out. Thus image pairs from the same group act as positive pairs, and image pairs from different groups act as negative pairs. By separating out these two groups, the model adapts itself by better separating the intermediate quality samples.", + "bbox": [ + 75, + 387, + 468, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let a positive pair $x_{(i)}$ and $x_{(j)}$ come from the same group, i.e., either both $i, j \\leq pN$ or $i, j > (1 - p)N$ . We use a modified NT-Xent contrastive loss [4] as our objective function. For a pair of images coming from the lower quality group where $i, j \\leq pN$ and $i \\neq j$ , the GC loss is defined as", + "bbox": [ + 75, + 508, + 470, + 598 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i, j} ^ {g c} = - \\log \\frac {\\exp (\\operatorname* {s i m} (z _ {(i)} , z _ {(j)}) / \\tau)}{\\sum_ {k > (1 - p) N} ^ {N} \\exp (\\operatorname* {s i m} (z _ {(i)} , z _ {(k)}) / \\tau)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 606, + 468, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_{(i)} = f_{\\theta_s}(f_{\\theta_e}(x_{(i)}))$ represents the feature at the output of the projection head for sample $x_{(i)}$ , sim refers to the cosine similarity between two features. and $\\mathbf{z}_{(k)}$ is the feature extracted from an image from the higher quality group with $k > (1 - p)N$ . Also, $\\tau$ represents temperature scaling parameter. While we define the loss above when $i,j\\leq pN$ , we can define a similar loss when $i,j > (1 - p)N$ . For all pairs of images within the same group, we obtain the GC loss and add them together to obtain the final loss, $\\mathcal{L}^{gc}$ as", + "bbox": [ + 75, + 657, + 472, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal{L}^{gc} = \\sum_{i = 1}^{pN}\\sum_{\\substack{j = 1\\\\ j\\neq i}}^{pN}\\mathcal{L}_{i,j}^{gc} + \\sum_{i = (1 - p)N + 1}^{N}\\sum_{\\substack{j = (1 - p)N + 1\\\\ j\\neq i}}^{N}\\mathcal{L}_{i,j}^{gc}. \\tag{3}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 820, + 468, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A block diagram of the GC loss is shown in Figure 2.", + "bbox": [ + 76, + 885, + 431, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Rank Loss", + "text_level": 1, + "bbox": [ + 500, + 387, + 627, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a sample of images from test data, the rank loss helps adapt the features by capturing quality-aware distributional information at the sample level. We introduce a ranking objective to learn the quality orders between two distorted versions of the test images that are quality discriminable. We distort each test image $x_{i}$ , $i = 1,2,\\dots ,N$ , in the minibatch of size $N$ with two different degrees of degradation of a given distortion type. Let $x_{i}^{high}$ and $x_{i}^{low}$ denote the higher and lower distorted image, respectively. The degradation types include synthetic distortions such as blur, compression, and noise. We choose degradation levels randomly from two sets of parameters sufficiently farther apart. Ideally, the distance between the features extracted from the test image and the highly distorted image should be greater than that between the features extracted from the test image and the lower distorted image. Our rank loss tries to capture this order and adapt the model to the test data.", + "bbox": [ + 496, + 426, + 890, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We project the triplet of images $(x_{i}, x_{i}^{high}, x_{i}^{low})$ in the feature space through the feature extractor and the projection head. We measure the distance between the distortion-augmented image and the original test data using the Euclidean distance in the projected feature space. Mathematically, let $\\mathbf{z}_{i} = f_{\\theta_{s}}(f_{\\theta_{e}}(x_{i}))$ be the feature extracted for test sample $x_{i}$ . Similarly we define $\\mathbf{z}_{i}^{high}$ and $\\mathbf{z}_{i}^{low}$ . Let $d_{i}^{high}$ denote the distance between $\\mathbf{z}_{i}$ and $\\mathbf{z}_{i}^{high}$ . Similarly, $d_{i}^{low}$ denotes the distance between $\\mathbf{z}_{i}$ and $\\mathbf{z}_{i}^{low}$ . The target model is now fine-tuned to obtain the correct ranking between the distances by achieving $d_{i}^{high} \\geq d_{i}^{low}$ shown in Figure 3.", + "bbox": [ + 496, + 690, + 892, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The probability of achieving this order is estimated by passing the difference in distances through a sigmoid func", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16745", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cfff0b4fd362c208c3eae5f7b14560808024eac16a0865be0cd580daca1a0a59.jpg", + "image_caption": [ + "Figure 3. The rank loss framework for test time adaptation. We distort the test image with two different degrees of degradation. We then project the triplet of images in the features space through the shared feature extractor and projection head. Distances calculated between the features extracted from each distorted image and the original test image are used to compare their rank order." + ], + "image_footnote": [], + "bbox": [ + 117, + 85, + 841, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion as", + "bbox": [ + 76, + 435, + 127, + 449 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i} = \\Pr \\left(d _ {i} ^ {\\text {h i g h}} \\geq d _ {i} ^ {\\text {l o w}}\\right) = \\frac {\\exp \\left(d _ {i} ^ {\\text {h i g h}} - d _ {i} ^ {\\text {l o w}}\\right)}{1 + \\exp \\left(d _ {i} ^ {\\text {h i g h}} - d _ {i} ^ {\\text {l o w}}\\right)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 458, + 468, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use binary cross-entropy loss between true label $\\bar{P}_i = 1$ and the predicted probability $P_{i}$ to obtain the rank loss as", + "bbox": [ + 76, + 523, + 468, + 555 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} ^ {r} (\\bar {P} _ {i}, P _ {i}) = - \\bar {P} _ {i} \\log P _ {i} - (1 - \\bar {P} _ {i}) \\log (1 - P _ {i}). (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 564, + 468, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a batch of size $N$ , the overall rank loss is given as", + "bbox": [ + 76, + 594, + 434, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {r} = \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {i} ^ {r}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 619, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One of the challenges in the rank loss is selecting the distortion type for every image. The original test image may already be distorted, which poses difficulties in deciding the distortion type that can help TTA. For example, if the test image is extremely blurred and we blur it again with two different levels, both images may be visually indistinguishable, thereby limiting the extent to which adaptation can help. We exploit the source model's knowledge to overcome this limitation in choosing the distortion type. We predict the quality scores of the pair $(x_{i}^{high}, x_{i}^{low})$ for each distortion type using the source model. We hypothesize that the pair with the maximal difference in predicted quality is sufficiently different in visual quality. Thus choosing such a pair in Equation (5) can help adapt the model.", + "bbox": [ + 75, + 672, + 470, + 885 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our overall self-supervised TTA objective function is a", + "bbox": [ + 96, + 885, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "combination of both the rank and GC loss given as", + "bbox": [ + 500, + 435, + 836, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\mathcal {L} ^ {g c} + \\lambda \\mathcal {L} ^ {r}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 455, + 890, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyper-parameter used to combine the losses.", + "bbox": [ + 500, + 479, + 877, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 505, + 633, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Quality Models, Datasets and Metrics", + "text_level": 1, + "bbox": [ + 500, + 530, + 825, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate TTA on four popular IQA databases using four different quality-aware models. In particular, we consider state-of-the-art deep IQA models such as TReS [7], MUSIQ [12], HyperIQA [30], and MetaIQA [40]. These models contain a ResNet backbone with batch normalization layers, which we model as the feature extractor $f_{\\theta_e}$ and only update its batch normalization parameters. We model the rest of the network as the quality regressor part, $f_{\\theta_c}$ . TReS and MUSIQ use transformers as a part of their architecture, which we include as a part of the quality regressor in all our main experiments. We also explore the adaptation of transformers by including them as part of the feature extractor and updating their layer normalization statistics in the supplementary material.", + "bbox": [ + 496, + 553, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We project the features extracted from the last layer of ResNet through a 256-dimensional fully connected (FC) layer corresponding to the self-supervised projection head $f_{\\theta_s}$ . These lower dimensional features are used to adapt the model at test time. Three IQA models, TReS, MUSIQ, and HyperIQA are trained on the LIVEFB database [37] containing 39,810 images. MetaIQA is trained on two synthetically distorted databases, TID2013 [26] and KADID-10k [17].", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16746", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2d5e46c9492bc68414c538c8d0fce700fa3a6ecab7594bbb4c9d21a5e9db9bf0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneDatabaseKonIQ-10kPIPALCID2013LIVE-IQA
MethodSROCCPLCCSROCCPLCCSROCCPLCCSROCCPLCC
TReSBaseline0.65200.69550.38450.40780.52720.64630.54350.4450
Rotation0.65060.68050.40610.41140.57060.66510.58660.5311
TTA-IQA0.65780.70740.42780.42040.60320.67100.67220.5963
MUSIQBaseline0.63040.68020.31900.34140.51730.60320.25960.3351
Rotation0.65770.71540.36650.36930.54870.61640.35120.3976
TTA-IQA0.66930.72300.37430.37310.54990.62200.36490.4031
HyperIQABaseline0.58610.63130.30370.33040.48950.61230.51430.4377
Rotation0.60330.65360.32680.34820.49020.61500.62680.5469
TTA-IQA0.59600.64950.36530.37670.50390.59880.62180.5438
MetaIQABaseline0.51620.44600.32870.29550.72130.68170.73230.6732
Rotation0.58230.53110.33530.30420.71770.67450.72710.6868
TTA-IQA0.58380.54280.40730.35100.78090.73990.79990.7726
", + "bbox": [ + 120, + 88, + 846, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of TTA-IQA with popular NR IQA methods and one popular auxiliary task - rotation prediction on authentically and synthetically distorted datasets. Bold entries imply the best performance for every individual quality-aware model on respective datasets.", + "bbox": [ + 75, + 306, + 893, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We choose challenging databases to evaluate the generalization capability of our TTA-IQA method. The test datasets are described as follows:", + "bbox": [ + 75, + 361, + 470, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "KonIQ-10k [9] is a popular in-the-wild authentically distorted database consisting of 10,073 quality-scored images.", + "bbox": [ + 75, + 407, + 470, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PIPAL [11] is a large-scale IQA database for evaluating perceptual image restoration. It contains 29k distorted images, including 19 different GAN-based algorithms.", + "bbox": [ + 75, + 452, + 468, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CID2013 [32] consists of 480 images in six image sets captured by 79 imaging devices.", + "bbox": [ + 75, + 500, + 468, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LIVE-IQA [29] consists of 779 synthetically distorted images from 29 reference images.", + "bbox": [ + 75, + 530, + 468, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our results using Spearman's rank-order correlation coefficient (SROCC) and Pearson's linear correlation coefficient (PLCC).", + "bbox": [ + 75, + 561, + 468, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 619, + 294, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement our setup in PyTorch and conduct all the experiments with an 16 GB NVIDIA RTX A4000 GPU. During TTA, we randomly select a patch of size $224 \\times 224$ from the input image and apply quality preserving augmentations such as horizontal flip and vertical flip before passing it through the network for TTA. We use the ADAM [13] optimizer with a learning rate of 0.001 and set the number of iteration as 3. All the experiments were run on five different seeds using a batch size of 8, and the final results were obtained by averaging.", + "bbox": [ + 75, + 642, + 468, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With regard to the self-supervised auxiliary tasks, we use a combination of rank loss and GC loss as our objective function using $\\lambda = 1$ given in Equation (7). We choose $p = 0.25$ to determine the groups. We calculate the GC loss using $\\tau = 1$ . For the rank loss, the test image is synthetically blurred using a Gaussian blur filter of size $5 \\times 5$ with two sets of standard deviations $\\sigma$ , for the Gaussian", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "blur kernel. We keep $\\sigma \\in [40,80]$ for highly distorted images and $\\sigma \\in [1,20]$ for less blurred images. For compression distortions, we specify the quality factor in [80, 95] for lower compression and [30, 60] for higher compression rates. Similarly, we add zero mean Gaussian white noise to the test image with variance in [0.05, 0.1] for higher distortion and in [0.005, 0.01] for lower distortion.", + "bbox": [ + 496, + 359, + 890, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Performance Evaluation", + "text_level": 1, + "bbox": [ + 498, + 479, + 723, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows the performance of our method on all four test datasets using all four quality-aware models. In addition to the source model, we also compare with rotation prediction [31] as the auxiliary task during TTA. A comparison with such a task helps understand the role of quality-aware losses for the TTA of IQA models.", + "bbox": [ + 496, + 503, + 890, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observe that TTA-IQA using the combination of rank and GC loss outperforms the source models in all the cases. Note that the PIPAL dataset has a huge distribution shift from the authentically distorted LIVEFB dataset on which three of the models were trained. While most source models perform very poorly on PIPAL, we achieve $10\\% -20\\%$ improvement over the source models. On the KonIQ-10k dataset, TTA-IQA gives around $1\\% -10\\%$ improvement over the source model. As both KonIQ-10k and LIVEFB are authentically distorted datasets, the distribution shift is reasonably small, leading to smaller improvements using adaptation. CID2013 is also an authentically distorted dataset and gives similar improvements of around $2\\% -13\\%$ over the source models. Our experiments on the LIVE-IQA database provide a significantly greater improvement of $10\\% -33\\%$ owing to the shift from authentic to synthetic distortions for three models.", + "bbox": [ + 496, + 597, + 892, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also observe from Table 1 that our approach outperforms the rotation prediction task in most cases. Other TTA ideas, such as TENT [33] and training using masked autoen", + "bbox": [ + 496, + 854, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16747", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4113a24e2f0307a7b176fb35d62c0450d445aa0a72eaa6fbdd66d9d3584444ec.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneDatabaseKonIQ-10kPIPALCID2013LIVE-IQA
RankGCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCC
TReS✓×0.65620.69890.41710.42040.60160.67360.67050.5908
×✓0.65160.69460.46660.51830.53660.64930.71600.6290
✓✓0.65780.70740.42780.42040.60320.67100.67220.5963
MUSIQ✓×0.65490.71490.37680.37180.52160.60340.36340.4024
×✓0.66110.71760.35850.36420.54460.61590.32990.3914
✓✓0.66930.72300.37430.37310.54990.62200.36490.4031
HyperIQA✓×0.59280.64550.36160.37320.50390.59910.55050.5050
×✓0.60940.65670.33330.35520.51200.62550.63310.5592
✓✓0.59600.64950.36530.37670.50390.59880.62180.5438
MetaIQA✓×0.55800.51180.39920.34370.75790.72820.78940.7534
×✓0.54140.46360.37100.32870.78610.70670.75660.6927
✓✓0.58380.54280.40730.35100.78090.73990.79990.7726
", + "bbox": [ + 130, + 88, + 836, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study results on authentically and synthetically distorted datasets using popular quality aware source models. Bold entries imply the best performance among all three settings.", + "bbox": [ + 75, + 306, + 893, + 338 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b2a67dd12412c5190c97d77c2b4e37b7602a5953485e000dae19d17651968b99.jpg", + "image_caption": [ + "Figure 4. Scatter plot for predicting quality of similar quality images from LIVE-IQA dataset for the source model and TTA using only rank loss or GC loss. We mark two sets of images having similar human opinion score (DMOS) by ellipses in each case." + ], + "image_footnote": [], + "bbox": [ + 76, + 356, + 320, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7bd95081d74e2ce9ecfbb5a0049f21149d987219de81d803708f0ced53b53fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 356, + 594, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87fd3f40ea4958c2c0f6c22ded6f839fe1920dd2d1bd37a605b6ef5cd8f59904.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 356, + 867, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "coders [6] are not applicable for the IQA task. In particular, the notion of entropy in regression tasks is not clear. On the other hand, masked reconstructions tend to lead to quality degradation. Thus, we do not compare with them.", + "bbox": [ + 75, + 551, + 468, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study & Other Experiments", + "text_level": 1, + "bbox": [ + 76, + 619, + 401, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Need for Rank Loss and GC Loss for TTA. We perform an ablation study with respect to the two losses in Table 2. We see from the results that the rank loss and the GC loss individually always improve on the source models. Further, the combination of the rank and GC loss provides an even better performance over the individual losses in most evaluation scenarios. Even in scenarios, where the combination loss achieves the second-best performance, its performance is very close to the best. For the rest of the experiments in this section, we present results on all four datasets using the TReS method alone.", + "bbox": [ + 75, + 643, + 468, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now discuss scenarios where the GC loss is particularly more useful than the rank loss. If the test image is extremely distorted, none of the three kinds of distortion types can create much difference in the perceptual quality of the two degraded versions. To understand this further, we consider highly distorted images (corresponding to a dif", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ferential mean opinion score (DMOS) greater than 70) of different distortion types in the LIVE-IQA dataset. We apply TTA only for these images using different losses. The SROCC performance for these images in Figure 5 reveal that the rank loss is not very effective. However, the GC loss works well in such scenarios.", + "bbox": [ + 496, + 551, + 892, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conversely, we also discuss scenarios where the rank loss is more useful than the GC loss. To illustrate this idea, we select multiple images having similar quality with DMOS in [28, 44] from the LIVE-IQA dataset. We adapt the source model for this set of images using both the GC and rank losses. From Figure 4, we observe that the rank loss leads to much better adaptation in terms of the resulting model correlating with the ground truth scores. Intuitively, when the input images have a similar quality, the pseudolabels given by the source model may be noisy, leading to an inaccurate grouping of the images into the lower and higher quality groups. This can adversely affect the GC loss. So the rank loss is more effective than the GC loss for test batches with similar quality images.", + "bbox": [ + 496, + 643, + 893, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Selecting Multiple Distortion Types in the Rank Loss. We experiment with the impact of different kinds of distortion types while incorporating the rank loss.", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16748", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca0e444cd1e1d3fbe99e30372c6eb592b42876d877fcb1c4894a6afaa4f4b806.jpg", + "image_caption": [ + "Results for Highly Distorted Images", + "Figure 5. TTA-IQA on different distortion-specific images from LIVE-IQA dataset with respect to individual losses" + ], + "image_footnote": [], + "bbox": [ + 93, + 116, + 460, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In particular, we consider three scenarios: one where only one of the distortion types is used in the rank loss, another where all the three types are used to obtain three loss terms which are summed together to update the model, and a third, where only one of the distortion types is chosen based on the discussion in Section 3.2.2. We observe from Table 3 that choosing the distortion type with the maximal difference in quality gives the best performance for the rank loss. This experiment validates our hypothesis that only a rank loss that can discriminate between the degraded image pairs is useful for TTA.", + "bbox": [ + 75, + 316, + 470, + 483 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6461869ff6e1f20e98da5f14d21fe74c81db965b37da6dadbe7aeab558cec652.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatabaseBlurCompNoiseAllBest
KonIQ-10k0.6560.6560.6670.6150.656
PIPAL0.3840.4150.3890.4150.417
CID20130.5270.6090.6060.5600.609
LIVE-IQA0.5780.6050.6180.6340.671
", + "bbox": [ + 78, + 494, + 464, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Selection of Group Size by Varying $p$ . We explore different values of $p$ for constructing the GC loss. For a batch size of 8, possible choices of the group size are 2, 3, 4 corresponding to $p = 0.25, 0.375, 0.5$ respectively. From Table 4, we observe that the performances are roughly similar for different values of $p$ with a slightly superior performance for $p = 0.25$ . We note that as $p$ increases, the groups are larger and probably less discriminative in terms of quality, leading to a slightly poorer performance.", + "bbox": [ + 75, + 609, + 468, + 744 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/89dcf9545d45cb7c08e04645c4673f040deee72e8357c5d87073cbbe69250f63.jpg", + "table_caption": [ + "Table 3. Impact of selecting multiple distortion types on rank loss" + ], + "table_footnote": [], + "table_body": "
Databasep=0.25p=0.375p=0.5
KonIQ-10k0.65160.64970.6467
PIPAL0.46660.46020.4616
CID20130.53660.54860.5536
LIVE-IQA0.71600.71710.6937
", + "bbox": [ + 120, + 758, + 423, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Selection of Number of Groups for GC Loss. In our formulation, we define the GC loss only between two con", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trastive groups. In principle, we can extend our idea of GC loss to multiple groups. In particular, we can cluster the images in a batch into multiple groups and apply the GC loss between every pair of groups and sum the loss terms. Thus, images coming from the same group act as positive pairs, and images from two different groups are considered as negative pairs. From Table 5, we observe that the number of groups $G$ does not impact the performance much.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d9b1104a07845e4bcea16e540c9dfd6c28d97d9b8eaf3c43d77b8c7838e48b9f.jpg", + "table_caption": [ + "Table 4. Impact of varying group size using $p$ for GC loss" + ], + "table_footnote": [], + "table_body": "
DatabaseG=2G=3G=4
KonIQ-10k0.65160.65240.6509
PIPAL0.46660.46680.4663
CID20130.53660.53270.5388
LIVE-IQA0.71600.70110.6883
", + "bbox": [ + 547, + 224, + 841, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Impact of varying number of groups $G$ for GC loss", + "bbox": [ + 500, + 306, + 854, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Choice of Number of Iterations for Learning Auxiliary Task. We also examine the effect of the number of iterations of parameter updates during TTA. Figure 6 shows that the performance on CID2013 dataset is fairly robust until 4 iterations. Beyond that, the model overfits the auxiliary task and leads to poor performance at test time.", + "bbox": [ + 496, + 337, + 890, + 429 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e8585276d93fbbe16b41ecc7e3ba270b2b147debf9baea6eb4de19e5d629d7ed.jpg", + "image_caption": [ + "Figure 6. Effect of increasing number of iterations" + ], + "image_footnote": [], + "bbox": [ + 532, + 446, + 864, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 647, + 617, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Test time adaptation has become very popular due to its simplicity and the lack of need for end-to-end training. Our work is perhaps one of the first attempts to design the method of TTA in the context of blind IQA. While most IQA methods focus on making the models robust enough to perform well for cross-database experiments, our TTA-IQA method can outperform existing state-of-the-art methods because of the adaptation at test time. We formulate novel self-supervised auxiliary tasks using the rank and group contrastive losses, which can learn quality-aware information from the test data. While primarily explored TTA for IQA, it would be interesting to understand the role of TTA for video quality assessment as well.", + "bbox": [ + 496, + 672, + 890, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement: This work was supported in part by a grant from Kotak IISc AI-ML Centre and Department of", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16749", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Science and Technology, Government of India, under grant CRG/2020/003516.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 133, + 173, + 148 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Piotr Bojanowski and Armand Joulin. Unsupervised learning by predicting noise. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 517-526. PMLR, 06-11 Aug 2017. 1", + "[2] Sebastian Bosse, Dominique Maniry, Klaus-Robert Müller, Thomas Wiegand, and Wojciech Samek. Deep neural networks for no-reference and full-reference image quality assessment. IEEE Transactions on Image Processing, 27(1):206-219, 2018. 2", + "[3] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 139–156, Cham, 2018. Springer International Publishing. 1", + "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org, 2020. 3, 4", + "[5] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), December 2015. 1", + "[6] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A. Efros. Test-time training with masked autoencoders. arXiv preprint arXiv:2209.07522, 2022. 7", + "[7] S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3209-3218, 2022. 2, 5", + "[8] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 1", + "[9] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. Trans. Img. Proc., 29:4041-4056, jan 2020. 6", + "[10] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 2427–2440. Curran Associates, Inc., 2021. 1", + "[11] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S. Ren, and Dong Chao. Pipal: A large-scale image quality assessment dataset for perceptual image restoration. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020," + ], + "bbox": [ + 78, + 157, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pages 633-651, Cham, 2020. Springer International Publishing. 6", + "[12] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5148-5157, October 2021. 2, 5", + "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[14] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), Vancouver, Canada, Apr. 2018. 1", + "[15] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Colorization as a proxy task for visual understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1", + "[16] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? Source hypothesis transfer for unsupervised domain adaptation. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 6028-6039. PMLR, 13-18 Jul 2020. 2", + "[17] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1–3, 2019. 5", + "[18] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. 2", + "[19] Yutao Liu, Ke Gu, Yongbing Zhang, Xiu Li, Guangtao Zhai, Debin Zhao, and Wen Gao. Unsupervised blind image quality evaluation via statistical measurements of structure, naturalness, and perception. IEEE Transactions on Circuits and Systems for Video Technology, 30(4):929-943, 2020. 2", + "[20] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 21808-21820. Curran Associates, Inc., 2021. 1, 2", + "[21] Kede Ma, Wentao Liu, Kai Zhang, Zhengfang Duanmu, Zhou Wang, and Wangmeng Zuo. End-to-end blind image quality assessment using deep neural networks. IEEE Transactions on Image Processing, 27(3):1202-1213, 2018. 2", + "[22] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21(12):4695-4708, 2012. 2", + "[23] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE Transactions on Image Processing, 20(12):3350-3364, 2011. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16750", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Chaithanya Kumar Mummadi, Robin Hutmacher, Kilian Rambach, Evgeny Levinkov, Thomas Brox, and Jan Hendrik Metzen. Test-time adaptation to distribution shift by confidence maximization and input transformation, 2021. 1", + "[25] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 2", + "[26] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, and C.-C. Jay Kuo. Image database tid2013: Peculiarities, results and perspectives. Signal Processing: Image Communication, 30:57-77, 2015. 5", + "[27] Michele A. Saad, Alan C. Bovik, and Christophe Charrier. Blind image quality assessment: A natural scene statistics approach in the dct domain. IEEE Transactions on Image Processing, 21(8):3339-3352, 2012. 2", + "[28] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 11539-11551. Curran Associates, Inc., 2020. 2, 3", + "[29] H.R. Sheikh, M.F. Sabir, and A.C. Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on Image Processing, 15(11):3440-3451, 2006. 6", + "[30] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 5", + "[31] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 9229-9248. PMLR, 13-18 Jul 2020. 1, 2, 6", + "[32] Toni Virtanen, Mikko Nuutinen, Mikko Vaahteranoksa, Pirkko Oittinen, and Jukka Häkkinen. Cid2013: A database for evaluating no-reference image quality assessment algorithms. IEEE Transactions on Image Processing, 24(1):390-402, 2015. 6", + "[33] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 1, 2, 3, 6", + "[34] Jingtao Xu, Peng Ye, Qiaohong Li, Haiqing Du, Yong Liu, and David Doermann. Blind image quality assessment based on high order statistics aggregation. IEEE Transactions on Image Processing, 25(9):4444-4457, 2016. 2", + "[35] Peng Ye, Jayant Kumar, Le Kang, and David Doermann. Unsupervised feature learning framework for no-reference" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image quality assessment. In 2012 IEEE conference on computer vision and pattern recognition, pages 1098-1105. IEEE, 2012. 2", + "[36] Peng Ye, Jayant Kumar, Le Kang, and David Doermann. Unsupervised feature learning framework for no-reference image quality assessment. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1098-1105, 2012. 2", + "[37] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5", + "[38] Guangtao Zhai, Xiongkuo Min, and Ning Liu. Free-energy principle inspired visual quality assessment: An overview. Digital Signal Processing, 91:11-20, 2019. 2", + "[39] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2020. 2", + "[40] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 5" + ], + "bbox": [ + 501, + 92, + 892, + 470 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16751", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Test Time Adaptation for Blind Image Quality Assessment/4f23b86f-f4db-4257-aee1-6b77c4f73657_model.json b/2023/Test Time Adaptation for Blind Image Quality Assessment/4f23b86f-f4db-4257-aee1-6b77c4f73657_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c2f4fe70e02cc2188b3b23b1b10305521e8015ab --- /dev/null +++ b/2023/Test Time Adaptation for Blind Image Quality Assessment/4f23b86f-f4db-4257-aee1-6b77c4f73657_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.192, + 0.131, + 0.78, + 0.153 + ], + "angle": 0, + "content": "Test Time Adaptation for Blind Image Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.181, + 0.786, + 0.217 + ], + "angle": 0, + "content": "Subhadeep Roy* Shankhanil Mitra* Soma Biswas Rajiv Soundararajan Indian Institute of Science, Bengaluru, India" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.22, + 0.802, + 0.234 + ], + "angle": 0, + "content": "subhadeeproy2000@gmail.com, {shankhanilm, somabiswas, rajivs}@iisc.ac.in" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.558 + ], + "angle": 0, + "content": "While the design of blind image quality assessment (IQA) algorithms has improved significantly, the distribution shift between the training and testing scenarios often leads to a poor performance of these methods at inference time. This motivates the study of test time adaptation (TTA) techniques to improve their performance at inference time. Existing auxiliary tasks and loss functions used for TTA may not be relevant for quality-aware adaptation of the pre-trained model. In this work, we introduce two novel quality-relevant auxiliary tasks at the batch and sample levels to enable TTA for blind IQA. In particular, we introduce a group contrastive loss at the batch level and a relative rank loss at the sample level to make the model quality aware and adapt to the target data. Our experiments reveal that even using a small batch of images from the test distribution helps achieve significant improvement in performance by updating the batch normalization statistics of the source model." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.585, + 0.21, + 0.601 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.807 + ], + "angle": 0, + "content": "The problem of image quality assessment (IQA) is extremely important in diverse image capture, processing, and sharing applications. However, a reference image is often not available for quality assessment. No reference (NR) or blind IQA primarily deals with the question of predicting image quality without using a reference image. Such NR IQA algorithms are often designed using machine learning approaches. More recently, deep learning based approaches have been extremely successful in achieving impressive performance. However, IQA applications are quite diverse and deal with several different distortions and distributional shifts. IQA models often have poor generalization ability and find it difficult to perform well under such shifts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.807, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Test time adaptation (TTA) has emerged as an important approach to address distributional shifts at test time [8]. It has been shown that by modifying a few global parameters of the model using a suitable loss that does not require the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.271, + 0.892, + 0.376 + ], + "angle": 0, + "content": "ground truth, one can significantly improve the performance of the model on the test data. Further, source-free adaptation, where the source data on which the original model was trained is not available while updating the model, is a realistic setting. While such approaches have been studied extensively in image classification literature [31, 33], there is hardly any literature on TTA for IQA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.893, + 0.576 + ], + "angle": 0, + "content": "There are multiple challenges in designing TTA for IQA. Typical losses used for TTA, such as entropy minimization [33], are not applicable for IQA. For example, IQA is often studied in the regression context. This makes it difficult to extend models based on class confidences [24] or class prototypes [10] for classification to IQA. Also, the relevance of other self-supervised tasks such as rotation prediction [14], context prediction [5], colorization [15], noise prediction [1], feature clustering [3] for adapting IQA models is not clear. While contrastive learning has also been employed for TTA [20], such a framework is not explicitly based on contrasting image quality, and its relevance is also not clear." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.895, + 0.746 + ], + "angle": 0, + "content": "Our main contribution is in the design of auxiliary tasks to enable TTA for IQA. We start with a source model trained on a large IQA dataset and fine-tune the model on individual batches of test samples. The first task we introduce for adaptation is based on contrasting groups of low and high quality images in a batch. Thus, we exploit the initial knowledge of the source model and try to adapt it by enforcing quality relationships among the batch of samples. Such a group contrastive (GC) learning approach fits naturally to our setting to account for any errors on individual samples that the source model may be prone to." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In contrast to the GC learning that depends on the batch, our second auxiliary task is an image specific task based on distorted augmentations of different types. Here, our goal is to enable the model to rank the image quality of further distorted versions of each test sample. We explore the role of different distortion types to leverage the maximum benefit of this task. While GC learning is more effective when samples in a batch are diverse in quality, the rank order based learning is more effective when the quality of the images is not extremely poor. Thus, a combination of the tasks" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.875, + 0.444, + 0.901 + ], + "angle": 0, + "content": "*Authors contributed equally to this work\n§https://github.com/subhadeeproy2000/TTA-IQA" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16742" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "helps overcome the shortcoming of both tasks and leads to an overall superior performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.123, + 0.47, + 0.199 + ], + "angle": 0, + "content": "We study the TTA problem under different settings of source and target datasets for multiple state of the art IQA models. Our results show significant improvements of the source model and the importance of TTA for IQA. We summarize our main contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.219, + 0.468, + 0.277 + ], + "angle": 0, + "content": "- We propose source-free test time adaptation techniques in the context of blind image quality assessment to mitigate distribution shifts between train and test data." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.296, + 0.468, + 0.4 + ], + "angle": 0, + "content": "- We formulate two novel quality-aware self-supervised auxiliary tasks to adapt to the test data distribution. While group contrastive learning helps capture quality discriminative information among several images within a batch, rank ordering helps maintain the quality order between two different distorted versions of the same image." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.419, + 0.467, + 0.476 + ], + "angle": 0, + "content": "- We show that our TTA method can significantly improve the performance of four different quality-aware source models, each on four different test IQA databases." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.219, + 0.468, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.516, + 0.218, + 0.531 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.542, + 0.28, + 0.559 + ], + "angle": 0, + "content": "2.1. Test Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.468, + 0.732 + ], + "angle": 0, + "content": "One of the first pieces of work on TTA [31] introduces a joint training framework using a loss for the main task and a self-supervised auxiliary task loss. The choice of a relevant auxiliary task is a challenging part of the design. Sun et al. [31] use rotation prediction as a pre-text task for image classification applications. Researchers also explore simpler tasks that are highly correlated with the main task, such as feature alignment with the batch and a simple contrastive loss framework [20] for adaptation. However, these methods require the source data to train the source model all over again to enable TTA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "TENT [33] studies source-free TTA, where entropy minimization-based adaptation even outperforms some methods that use the source data. Moreover, only the batch normalization parameters of the model are adapted in TENT. There are several works on batch norm statistics adaptation [25, 28] to improve the robustness of the model at test time. SHOT [16] presents a clustering-based pseudolabeling method to align features from the target domain to the source domain using an information maximization loss. While a plethora of methods exists as above, the auxiliary tasks in these methods are not relevant for IQA." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.848, + 0.107 + ], + "angle": 0, + "content": "2.2. No-Reference Image Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.339 + ], + "angle": 0, + "content": "Most classical NR IQA algorithms are mainly based on natural scene statistics (NSS) [19, 22, 38]. In [23], the naturalness of the distorted image in the wavelet domain is modeled based on NSS. Saad et al. [27] also design the NSS model in the discrete cosine transform (DCT) domain. CORNIA [35], and HOSA [34] are among the earliest codebook learning based methods to predict quality. With the emergence of deep learning and the availability of large subject-rated IQA databases, various general-purpose NR IQA methods have been designed based on convolutional neural network (CNN) architectures [2, 18, 21, 36]. Some of these methods require end-to-end training [2, 21] of deep neural networks (DNN), while others [7, 12, 30, 39] are based on updating pre-trained models with some modifications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.892, + 0.506 + ], + "angle": 0, + "content": "More recently, Zhang et al. [39] propose a method to train bilinear DNNs to simultaneously model both authentic and synthetic distortions. Su et al. [30] design a self-adaptive hyper network to provide weights for the quality prediction module parameters and handle various types of distortions and content in the images. Several methods [7, 12] explore transformer-based architectures along with a CNN to capture dependencies between local and global features. MetaIQA [40] proposes meta-learning on synthetic distortions by using a shared quality prior knowledge model to adapt to any kind of distortion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.581 + ], + "angle": 0, + "content": "All the existing methods assume that the train and test data come from the same distribution. If there is a distribution shift across different databases, we need adaptation of the pre-trained model to learn about the target distributional information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.597, + 0.634, + 0.614 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.892, + 0.764 + ], + "angle": 0, + "content": "We propose a novel self-supervised Test Time Adaptation technique for Image Quality Assessment (TTA-IQA) to adapt pre-trained quality models and mitigate distribution shifts between the source and target data. We consider source-free TTA, where we only have access to the pre-trained quality-aware models and no access to the source training data. When a batch of test data \\( D = \\{x_{j}\\}_{j=1}^{n} \\) arrives, we adapt the model using the batch without knowledge of corresponding ground truth \\( \\{y_{j}\\}_{j=1}^{n} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.613, + 0.789 + ], + "angle": 0, + "content": "3.1. Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Let the model trained on the source data be \\( f_{\\theta} \\) where \\( \\theta = (\\theta_e, \\theta_c) \\) corresponds to the parameters of the network, and \\( \\theta_e \\) and \\( \\theta_c \\) correspond to the parameters of the feature extractor and regression layers. Thus, for an input image \\( x \\), we denote \\( f_{\\theta}(x) = f_{\\theta_c}(f_{\\theta_e}(x)) \\). Since our goal is to learn the distribution shift between train and test data, we only update the parameters of the feature extractor, \\( \\theta_e \\), to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16743" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.87, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.244, + 0.894, + 0.288 + ], + "angle": 0, + "content": "Figure 1. Block diagram of a general architecture for test time adaptation. At test time training, the normalization layers of the feature extractor are adapted by optimizing the combination of rank and GC loss. At inference time, we predict the quality scores of test images using the updated feature extractor and pre-trained quality regressor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.468, + 0.354 + ], + "angle": 0, + "content": "align the features between the train and test distributions in a lower dimensional space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.468, + 0.519 + ], + "angle": 0, + "content": "The key challenge in TTA is the choice of a self-supervised auxiliary task that is highly correlated with the main task of IQA. However, relying too much on the auxiliary task can affect the performance of the main task. To prevent loss of learnt information induced by the auxiliary task [4], we first project the features to a lower dimensional space using a non-linear projection head \\( f_{\\theta_s} \\), parameterized by \\( \\theta_s \\). The auxiliary task drives the adaptation of the feature extractor through the projection head. Thus our model now has three parts parametrized by \\( (\\theta_e, \\theta_s, \\theta_c) \\) to resemble a Y-shape as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.468, + 0.747 + ], + "angle": 0, + "content": "When a batch of test instances \\(D\\) arrives, we extract features from the feature extractor, followed by the projection head, and update the set of parameters \\((\\theta_e,\\theta_s)\\) by optimizing a self-supervised objective function \\(\\mathcal{L}_s(D)\\). Updating all the model parameters of the feature extractor \\(\\theta_{e}\\) can cause the model to diverge too much from training, and the performance can drop drastically. Inspired by prior work on test time entropy minimization [33] and improving robustness for test data [28], we only update the linear and lower-dimensional feature modulation parameters. In a neural network, normalization layers satisfy these properties. So we only adapt the batch normalization layers by updating the affine parameters to mitigate distributional shifts. Thus we update these parameters by optimizing the auxiliary task loss given by" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.758, + 0.468, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\theta_ {e} ^ {*}, \\theta_ {s} ^ {*} = \\underset {\\left(\\theta_ {e}, \\theta_ {s}\\right)} {\\operatorname {a r g m i n}} \\mathcal {L} _ {s} (D) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "After optimizing the above loss function, we use the updated feature extractor parameters \\(\\theta_e^*\\) and pre-trained quality regressor parameters \\(\\theta_c\\) to predict the quality for a batch of test data. Since the distribution across batches can vary significantly, we ignore the earlier updated model and start the TTA based on source model weights for new test instants. Thus the updated target model always only depends" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.325, + 0.673, + 0.339 + ], + "angle": 0, + "content": "on the incoming test data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.349, + 0.78, + 0.366 + ], + "angle": 0, + "content": "3.2. Self-Supervised Auxiliary Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.599 + ], + "angle": 0, + "content": "Our goal is to carefully choose self-supervised auxiliary tasks that capture quality-aware distributional information to adapt the feature encoder. We formulate two novel and complementary self-supervised learning techniques which help to learn the distribution shift between train and test data. These self-supervised objective functions are - 1) Group Contrastive Loss and 2) Rank Loss. While the GC loss works well when there is a reasonable separation of quality among the samples in a batch, the rank loss works better even when the quality of the batch samples is similar. On the other hand, the rank loss is meaningful only when the quality of the input image is not extremely low, while the GC loss is independent of the quality of a given image. Thus, a combination of the two losses renders our TTA extremely effective across various scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.722, + 0.634 + ], + "angle": 0, + "content": "3.2.1 Group Contrastive Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While contrastive learning has been used for TTA of deep image classification, its direct application does not appear relevant to the task of IQA. Thus, we introduce group contrastive (GC) learning as an auxiliary task for TTA of IQA models. In particular, we make two groups of images from a single batch of \\(N\\) images based on the pseudo-labels given by the pre-trained source model. We sort the images in ascending order as \\(x_{(1)}, x_{(2)}, \\ldots, x_{(N)}\\) based on the pseudolabels, where \\(x_{(i)}, i = 1, 2, \\ldots, N\\), corresponds to the \\(i^{th}\\) lowest quality image in the batch. We then segregate images with high quality scores (say, top \\(p\\) fraction of the data in a batch) and include them in a group of higher-quality images. Similarly, we separate out lower quality images (say, the lowest \\(p\\) fraction of the images in a batch) together to form another group. Here we assume that \\(pN\\) and \\((1 - p)N\\) are integers for simplicity; else, they can be rounded off to the nearest integers." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16744" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.09, + 0.895, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.308, + 0.895, + 0.352 + ], + "angle": 0, + "content": "Figure 2. The overview framework of group contrastive loss for test time adaptation. For a batch of images, two different groups are formed based on pseudo-labels given by the source model. The group contrastive loss tries to minimize the distance between features extracted from images belonging to the same group while maximizing the distance between features from different groups." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.47, + 0.509 + ], + "angle": 0, + "content": "The premise behind our loss for GC learning is that images from the same quality group should give similar feature representations in a lower dimensional space while features of images from different groups are separated out. Thus image pairs from the same group act as positive pairs, and image pairs from different groups act as negative pairs. By separating out these two groups, the model adapts itself by better separating the intermediate quality samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.471, + 0.599 + ], + "angle": 0, + "content": "Let a positive pair \\( x_{(i)} \\) and \\( x_{(j)} \\) come from the same group, i.e., either both \\( i, j \\leq pN \\) or \\( i, j > (1 - p)N \\). We use a modified NT-Xent contrastive loss [4] as our objective function. For a pair of images coming from the lower quality group where \\( i, j \\leq pN \\) and \\( i \\neq j \\), the GC loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.607, + 0.47, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i, j} ^ {g c} = - \\log \\frac {\\exp (\\operatorname* {s i m} (z _ {(i)} , z _ {(j)}) / \\tau)}{\\sum_ {k > (1 - p) N} ^ {N} \\exp (\\operatorname* {s i m} (z _ {(i)} , z _ {(k)}) / \\tau)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.81 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_{(i)} = f_{\\theta_s}(f_{\\theta_e}(x_{(i)}))\\) represents the feature at the output of the projection head for sample \\(x_{(i)}\\), sim refers to the cosine similarity between two features. and \\(\\mathbf{z}_{(k)}\\) is the feature extracted from an image from the higher quality group with \\(k > (1 - p)N\\). Also, \\(\\tau\\) represents temperature scaling parameter. While we define the loss above when \\(i,j\\leq pN\\), we can define a similar loss when \\(i,j > (1 - p)N\\). For all pairs of images within the same group, we obtain the GC loss and add them together to obtain the final loss, \\(\\mathcal{L}^{gc}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.821, + 0.47, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\mathcal{L}^{gc} = \\sum_{i = 1}^{pN}\\sum_{\\substack{j = 1\\\\ j\\neq i}}^{pN}\\mathcal{L}_{i,j}^{gc} + \\sum_{i = (1 - p)N + 1}^{N}\\sum_{\\substack{j = (1 - p)N + 1\\\\ j\\neq i}}^{N}\\mathcal{L}_{i,j}^{gc}. \\tag{3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.433, + 0.902 + ], + "angle": 0, + "content": "A block diagram of the GC loss is shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.388, + 0.629, + 0.402 + ], + "angle": 0, + "content": "3.2.2 Rank Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.684 + ], + "angle": 0, + "content": "Given a sample of images from test data, the rank loss helps adapt the features by capturing quality-aware distributional information at the sample level. We introduce a ranking objective to learn the quality orders between two distorted versions of the test images that are quality discriminable. We distort each test image \\( x_{i} \\), \\( i = 1,2,\\dots ,N \\), in the minibatch of size \\( N \\) with two different degrees of degradation of a given distortion type. Let \\( x_{i}^{high} \\) and \\( x_{i}^{low} \\) denote the higher and lower distorted image, respectively. The degradation types include synthetic distortions such as blur, compression, and noise. We choose degradation levels randomly from two sets of parameters sufficiently farther apart. Ideally, the distance between the features extracted from the test image and the highly distorted image should be greater than that between the features extracted from the test image and the lower distorted image. Our rank loss tries to capture this order and adapt the model to the test data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.691, + 0.893, + 0.865 + ], + "angle": 0, + "content": "We project the triplet of images \\((x_{i}, x_{i}^{high}, x_{i}^{low})\\) in the feature space through the feature extractor and the projection head. We measure the distance between the distortion-augmented image and the original test data using the Euclidean distance in the projected feature space. Mathematically, let \\(\\mathbf{z}_{i} = f_{\\theta_{s}}(f_{\\theta_{e}}(x_{i}))\\) be the feature extracted for test sample \\(x_{i}\\). Similarly we define \\(\\mathbf{z}_{i}^{high}\\) and \\(\\mathbf{z}_{i}^{low}\\). Let \\(d_{i}^{high}\\) denote the distance between \\(\\mathbf{z}_{i}\\) and \\(\\mathbf{z}_{i}^{high}\\). Similarly, \\(d_{i}^{low}\\) denotes the distance between \\(\\mathbf{z}_{i}\\) and \\(\\mathbf{z}_{i}^{low}\\). The target model is now fine-tuned to obtain the correct ranking between the distances by achieving \\(d_{i}^{high} \\geq d_{i}^{low}\\) shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The probability of achieving this order is estimated by passing the difference in distances through a sigmoid func" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16745" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.087, + 0.843, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.366, + 0.895, + 0.41 + ], + "angle": 0, + "content": "Figure 3. The rank loss framework for test time adaptation. We distort the test image with two different degrees of degradation. We then project the triplet of images in the features space through the shared feature extractor and projection head. Distances calculated between the features extracted from each distorted image and the original test image are used to compare their rank order." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.436, + 0.129, + 0.45 + ], + "angle": 0, + "content": "tion as" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.459, + 0.47, + 0.524 + ], + "angle": 0, + "content": "\\[\nP _ {i} = \\Pr \\left(d _ {i} ^ {\\text {h i g h}} \\geq d _ {i} ^ {\\text {l o w}}\\right) = \\frac {\\exp \\left(d _ {i} ^ {\\text {h i g h}} - d _ {i} ^ {\\text {l o w}}\\right)}{1 + \\exp \\left(d _ {i} ^ {\\text {h i g h}} - d _ {i} ^ {\\text {l o w}}\\right)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.524, + 0.47, + 0.556 + ], + "angle": 0, + "content": "We use binary cross-entropy loss between true label \\(\\bar{P}_i = 1\\) and the predicted probability \\(P_{i}\\) to obtain the rank loss as" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.565, + 0.47, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} ^ {r} (\\bar {P} _ {i}, P _ {i}) = - \\bar {P} _ {i} \\log P _ {i} - (1 - \\bar {P} _ {i}) \\log (1 - P _ {i}). (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.595, + 0.435, + 0.611 + ], + "angle": 0, + "content": "For a batch of size \\(N\\), the overall rank loss is given as" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.621, + 0.47, + 0.664 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {r} = \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {i} ^ {r}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.886 + ], + "angle": 0, + "content": "One of the challenges in the rank loss is selecting the distortion type for every image. The original test image may already be distorted, which poses difficulties in deciding the distortion type that can help TTA. For example, if the test image is extremely blurred and we blur it again with two different levels, both images may be visually indistinguishable, thereby limiting the extent to which adaptation can help. We exploit the source model's knowledge to overcome this limitation in choosing the distortion type. We predict the quality scores of the pair \\((x_{i}^{high}, x_{i}^{low})\\) for each distortion type using the source model. We hypothesize that the pair with the maximal difference in predicted quality is sufficiently different in visual quality. Thus choosing such a pair in Equation (5) can help adapt the model." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our overall self-supervised TTA objective function is a" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.436, + 0.837, + 0.451 + ], + "angle": 0, + "content": "combination of both the rank and GC loss given as" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.457, + 0.892, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\mathcal {L} ^ {g c} + \\lambda \\mathcal {L} ^ {r}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.48, + 0.878, + 0.495 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyper-parameter used to combine the losses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.506, + 0.634, + 0.523 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.826, + 0.547 + ], + "angle": 0, + "content": "4.1. Quality Models, Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.765 + ], + "angle": 0, + "content": "We evaluate TTA on four popular IQA databases using four different quality-aware models. In particular, we consider state-of-the-art deep IQA models such as TReS [7], MUSIQ [12], HyperIQA [30], and MetaIQA [40]. These models contain a ResNet backbone with batch normalization layers, which we model as the feature extractor \\( f_{\\theta_e} \\) and only update its batch normalization parameters. We model the rest of the network as the quality regressor part, \\( f_{\\theta_c} \\). TReS and MUSIQ use transformers as a part of their architecture, which we include as a part of the quality regressor in all our main experiments. We also explore the adaptation of transformers by including them as part of the feature extractor and updating their layer normalization statistics in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We project the features extracted from the last layer of ResNet through a 256-dimensional fully connected (FC) layer corresponding to the self-supervised projection head \\( f_{\\theta_s} \\). These lower dimensional features are used to adapt the model at test time. Three IQA models, TReS, MUSIQ, and HyperIQA are trained on the LIVEFB database [37] containing 39,810 images. MetaIQA is trained on two synthetically distorted databases, TID2013 [26] and KADID-10k [17]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16746" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.089, + 0.847, + 0.306 + ], + "angle": 0, + "content": "
BackboneDatabaseKonIQ-10kPIPALCID2013LIVE-IQA
MethodSROCCPLCCSROCCPLCCSROCCPLCCSROCCPLCC
TReSBaseline0.65200.69550.38450.40780.52720.64630.54350.4450
Rotation0.65060.68050.40610.41140.57060.66510.58660.5311
TTA-IQA0.65780.70740.42780.42040.60320.67100.67220.5963
MUSIQBaseline0.63040.68020.31900.34140.51730.60320.25960.3351
Rotation0.65770.71540.36650.36930.54870.61640.35120.3976
TTA-IQA0.66930.72300.37430.37310.54990.62200.36490.4031
HyperIQABaseline0.58610.63130.30370.33040.48950.61230.51430.4377
Rotation0.60330.65360.32680.34820.49020.61500.62680.5469
TTA-IQA0.59600.64950.36530.37670.50390.59880.62180.5438
MetaIQABaseline0.51620.44600.32870.29550.72130.68170.73230.6732
Rotation0.58230.53110.33530.30420.71770.67450.72710.6868
TTA-IQA0.58380.54280.40730.35100.78090.73990.79990.7726
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.308, + 0.895, + 0.339 + ], + "angle": 0, + "content": "Table 1. Comparison of TTA-IQA with popular NR IQA methods and one popular auxiliary task - rotation prediction on authentically and synthetically distorted datasets. Bold entries imply the best performance for every individual quality-aware model on respective datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.471, + 0.406 + ], + "angle": 0, + "content": "We choose challenging databases to evaluate the generalization capability of our TTA-IQA method. The test datasets are described as follows:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.471, + 0.452 + ], + "angle": 0, + "content": "KonIQ-10k [9] is a popular in-the-wild authentically distorted database consisting of 10,073 quality-scored images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.47, + 0.499 + ], + "angle": 0, + "content": "PIPAL [11] is a large-scale IQA database for evaluating perceptual image restoration. It contains 29k distorted images, including 19 different GAN-based algorithms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.469, + 0.53 + ], + "angle": 0, + "content": "CID2013 [32] consists of 480 images in six image sets captured by 79 imaging devices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.469, + 0.561 + ], + "angle": 0, + "content": "LIVE-IQA [29] consists of 779 synthetically distorted images from 29 reference images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.469, + 0.607 + ], + "angle": 0, + "content": "We evaluate our results using Spearman's rank-order correlation coefficient (SROCC) and Pearson's linear correlation coefficient (PLCC)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.295, + 0.636 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.795 + ], + "angle": 0, + "content": "We implement our setup in PyTorch and conduct all the experiments with an 16 GB NVIDIA RTX A4000 GPU. During TTA, we randomly select a patch of size \\(224 \\times 224\\) from the input image and apply quality preserving augmentations such as horizontal flip and vertical flip before passing it through the network for TTA. We use the ADAM [13] optimizer with a learning rate of 0.001 and set the number of iteration as 3. All the experiments were run on five different seeds using a batch size of 8, and the final results were obtained by averaging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "With regard to the self-supervised auxiliary tasks, we use a combination of rank loss and GC loss as our objective function using \\(\\lambda = 1\\) given in Equation (7). We choose \\(p = 0.25\\) to determine the groups. We calculate the GC loss using \\(\\tau = 1\\). For the rank loss, the test image is synthetically blurred using a Gaussian blur filter of size \\(5 \\times 5\\) with two sets of standard deviations \\(\\sigma\\), for the Gaussian" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.468 + ], + "angle": 0, + "content": "blur kernel. We keep \\(\\sigma \\in [40,80]\\) for highly distorted images and \\(\\sigma \\in [1,20]\\) for less blurred images. For compression distortions, we specify the quality factor in [80, 95] for lower compression and [30, 60] for higher compression rates. Similarly, we add zero mean Gaussian white noise to the test image with variance in [0.05, 0.1] for higher distortion and in [0.005, 0.01] for lower distortion." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.481, + 0.725, + 0.495 + ], + "angle": 0, + "content": "4.3. Performance Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Table 1 shows the performance of our method on all four test datasets using all four quality-aware models. In addition to the source model, we also compare with rotation prediction [31] as the auxiliary task during TTA. A comparison with such a task helps understand the role of quality-aware losses for the TTA of IQA models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.893, + 0.853 + ], + "angle": 0, + "content": "We observe that TTA-IQA using the combination of rank and GC loss outperforms the source models in all the cases. Note that the PIPAL dataset has a huge distribution shift from the authentically distorted LIVEFB dataset on which three of the models were trained. While most source models perform very poorly on PIPAL, we achieve \\(10\\% -20\\%\\) improvement over the source models. On the KonIQ-10k dataset, TTA-IQA gives around \\(1\\% -10\\%\\) improvement over the source model. As both KonIQ-10k and LIVEFB are authentically distorted datasets, the distribution shift is reasonably small, leading to smaller improvements using adaptation. CID2013 is also an authentically distorted dataset and gives similar improvements of around \\(2\\% -13\\%\\) over the source models. Our experiments on the LIVE-IQA database provide a significantly greater improvement of \\(10\\% -33\\%\\) owing to the shift from authentic to synthetic distortions for three models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.902 + ], + "angle": 0, + "content": "We also observe from Table 1 that our approach outperforms the rotation prediction task in most cases. Other TTA ideas, such as TENT [33] and training using masked autoen" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16747" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.132, + 0.089, + 0.837, + 0.306 + ], + "angle": 0, + "content": "
BackboneDatabaseKonIQ-10kPIPALCID2013LIVE-IQA
RankGCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCC
TReS✓×0.65620.69890.41710.42040.60160.67360.67050.5908
×✓0.65160.69460.46660.51830.53660.64930.71600.6290
✓✓0.65780.70740.42780.42040.60320.67100.67220.5963
MUSIQ✓×0.65490.71490.37680.37180.52160.60340.36340.4024
×✓0.66110.71760.35850.36420.54460.61590.32990.3914
✓✓0.66930.72300.37430.37310.54990.62200.36490.4031
HyperIQA✓×0.59280.64550.36160.37320.50390.59910.55050.5050
×✓0.60940.65670.33330.35520.51200.62550.63310.5592
✓✓0.59600.64950.36530.37670.50390.59880.62180.5438
MetaIQA✓×0.55800.51180.39920.34370.75790.72820.78940.7534
×✓0.54140.46360.37100.32870.78610.70670.75660.6927
✓✓0.58380.54280.40730.35100.78090.73990.79990.7726
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.308, + 0.895, + 0.339 + ], + "angle": 0, + "content": "Table 2. Ablation study results on authentically and synthetically distorted datasets using popular quality aware source models. Bold entries imply the best performance among all three settings." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.357, + 0.321, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.357, + 0.596, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.357, + 0.868, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.5, + 0.893, + 0.528 + ], + "angle": 0, + "content": "Figure 4. Scatter plot for predicting quality of similar quality images from LIVE-IQA dataset for the source model and TTA using only rank loss or GC loss. We mark two sets of images having similar human opinion score (DMOS) by ellipses in each case." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.613 + ], + "angle": 0, + "content": "coders [6] are not applicable for the IQA task. In particular, the notion of entropy in regression tasks is not clear. On the other hand, masked reconstructions tend to lead to quality degradation. Thus, we do not compare with them." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.403, + 0.637 + ], + "angle": 0, + "content": "4.4. Ablation Study & Other Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.81 + ], + "angle": 0, + "content": "Need for Rank Loss and GC Loss for TTA. We perform an ablation study with respect to the two losses in Table 2. We see from the results that the rank loss and the GC loss individually always improve on the source models. Further, the combination of the rank and GC loss provides an even better performance over the individual losses in most evaluation scenarios. Even in scenarios, where the combination loss achieves the second-best performance, its performance is very close to the best. For the rest of the experiments in this section, we present results on all four datasets using the TReS method alone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We now discuss scenarios where the GC loss is particularly more useful than the rank loss. If the test image is extremely distorted, none of the three kinds of distortion types can create much difference in the perceptual quality of the two degraded versions. To understand this further, we consider highly distorted images (corresponding to a dif" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.893, + 0.642 + ], + "angle": 0, + "content": "ferential mean opinion score (DMOS) greater than 70) of different distortion types in the LIVE-IQA dataset. We apply TTA only for these images using different losses. The SROCC performance for these images in Figure 5 reveal that the rank loss is not very effective. However, the GC loss works well in such scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Conversely, we also discuss scenarios where the rank loss is more useful than the GC loss. To illustrate this idea, we select multiple images having similar quality with DMOS in [28, 44] from the LIVE-IQA dataset. We adapt the source model for this set of images using both the GC and rank losses. From Figure 4, we observe that the rank loss leads to much better adaptation in terms of the resulting model correlating with the ground truth scores. Intuitively, when the input images have a similar quality, the pseudolabels given by the source model may be noisy, leading to an inaccurate grouping of the images into the lower and higher quality groups. This can adversely affect the GC loss. So the rank loss is more effective than the GC loss for test batches with similar quality images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Effect of Selecting Multiple Distortion Types in the Rank Loss. We experiment with the impact of different kinds of distortion types while incorporating the rank loss." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16748" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.098, + 0.413, + 0.116 + ], + "angle": 0, + "content": "Results for Highly Distorted Images" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.117, + 0.462, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.47, + 0.294 + ], + "angle": 0, + "content": "Figure 5. TTA-IQA on different distortion-specific images from LIVE-IQA dataset with respect to individual losses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.471, + 0.484 + ], + "angle": 0, + "content": "In particular, we consider three scenarios: one where only one of the distortion types is used in the rank loss, another where all the three types are used to obtain three loss terms which are summed together to update the model, and a third, where only one of the distortion types is chosen based on the discussion in Section 3.2.2. We observe from Table 3 that choosing the distortion type with the maximal difference in quality gives the best performance for the rank loss. This experiment validates our hypothesis that only a rank loss that can discriminate between the degraded image pairs is useful for TTA." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.495, + 0.465, + 0.576 + ], + "angle": 0, + "content": "
DatabaseBlurCompNoiseAllBest
KonIQ-10k0.6560.6560.6670.6150.656
PIPAL0.3840.4150.3890.4150.417
CID20130.5270.6090.6060.5600.609
LIVE-IQA0.5780.6050.6180.6340.671
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.577, + 0.467, + 0.592 + ], + "angle": 0, + "content": "Table 3. Impact of selecting multiple distortion types on rank loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.746 + ], + "angle": 0, + "content": "Selection of Group Size by Varying \\( p \\). We explore different values of \\( p \\) for constructing the GC loss. For a batch size of 8, possible choices of the group size are 2, 3, 4 corresponding to \\( p = 0.25, 0.375, 0.5 \\) respectively. From Table 4, we observe that the performances are roughly similar for different values of \\( p \\) with a slightly superior performance for \\( p = 0.25 \\). We note that as \\( p \\) increases, the groups are larger and probably less discriminative in terms of quality, leading to a slightly poorer performance." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.759, + 0.424, + 0.839 + ], + "angle": 0, + "content": "
Databasep=0.25p=0.375p=0.5
KonIQ-10k0.65160.64970.6467
PIPAL0.46660.46020.4616
CID20130.53660.54860.5536
LIVE-IQA0.71600.71710.6937
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.841, + 0.444, + 0.856 + ], + "angle": 0, + "content": "Table 4. Impact of varying group size using \\( p \\) for GC loss" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Selection of Number of Groups for GC Loss. In our formulation, we define the GC loss only between two con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "trastive groups. In principle, we can extend our idea of GC loss to multiple groups. In particular, we can cluster the images in a batch into multiple groups and apply the GC loss between every pair of groups and sum the loss terms. Thus, images coming from the same group act as positive pairs, and images from two different groups are considered as negative pairs. From Table 5, we observe that the number of groups \\( G \\) does not impact the performance much." + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.226, + 0.842, + 0.306 + ], + "angle": 0, + "content": "
DatabaseG=2G=3G=4
KonIQ-10k0.65160.65240.6509
PIPAL0.46660.46680.4663
CID20130.53660.53270.5388
LIVE-IQA0.71600.70110.6883
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.308, + 0.855, + 0.322 + ], + "angle": 0, + "content": "Table 5. Impact of varying number of groups \\( G \\) for GC loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.43 + ], + "angle": 0, + "content": "Choice of Number of Iterations for Learning Auxiliary Task. We also examine the effect of the number of iterations of parameter updates during TTA. Figure 6 shows that the performance on CID2013 dataset is fairly robust until 4 iterations. Beyond that, the model overfits the auxiliary task and leads to poor performance at test time." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.448, + 0.865, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.611, + 0.846, + 0.625 + ], + "angle": 0, + "content": "Figure 6. Effect of increasing number of iterations" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.648, + 0.619, + 0.664 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Test time adaptation has become very popular due to its simplicity and the lack of need for end-to-end training. Our work is perhaps one of the first attempts to design the method of TTA in the context of blind IQA. While most IQA methods focus on making the models robust enough to perform well for cross-database experiments, our TTA-IQA method can outperform existing state-of-the-art methods because of the adaptation at test time. We formulate novel self-supervised auxiliary tasks using the rank and group contrastive losses, which can learn quality-aware information from the test data. While primarily explored TTA for IQA, it would be interesting to understand the role of TTA for video quality assessment as well." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement: This work was supported in part by a grant from Kotak IISc AI-ML Centre and Department of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16749" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "Science and Technology, Government of India, under grant CRG/2020/003516." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.134, + 0.174, + 0.15 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[1] Piotr Bojanowski and Armand Joulin. Unsupervised learning by predicting noise. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 517-526. PMLR, 06-11 Aug 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.472, + 0.312 + ], + "angle": 0, + "content": "[2] Sebastian Bosse, Dominique Maniry, Klaus-Robert Müller, Thomas Wiegand, and Wojciech Samek. Deep neural networks for no-reference and full-reference image quality assessment. IEEE Transactions on Image Processing, 27(1):206-219, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.472, + 0.396 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 139–156, Cham, 2018. Springer International Publishing. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.398, + 0.472, + 0.466 + ], + "angle": 0, + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.468, + 0.472, + 0.523 + ], + "angle": 0, + "content": "[5] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), December 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.524, + 0.472, + 0.564 + ], + "angle": 0, + "content": "[6] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A. Efros. Test-time training with masked autoencoders. arXiv preprint arXiv:2209.07522, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.566, + 0.472, + 0.635 + ], + "angle": 0, + "content": "[7] S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3209-3218, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.636, + 0.472, + 0.691 + ], + "angle": 0, + "content": "[8] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.692, + 0.472, + 0.748 + ], + "angle": 0, + "content": "[9] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. Trans. Img. Proc., 29:4041-4056, jan 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[10] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 2427–2440. Curran Associates, Inc., 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[11] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S. Ren, and Dong Chao. Pipal: A large-scale image quality assessment dataset for perceptual image restoration. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.159, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "pages 633-651, Cham, 2020. Springer International Publishing. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[12] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5148-5157, October 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[14] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), Vancouver, Canada, Apr. 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[15] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Colorization as a proxy task for visual understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[16] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? Source hypothesis transfer for unsupervised domain adaptation. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 6028-6039. PMLR, 13-18 Jul 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[17] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1–3, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[18] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[19] Yutao Liu, Ke Gu, Yongbing Zhang, Xiu Li, Guangtao Zhai, Debin Zhao, and Wen Gao. Unsupervised blind image quality evaluation via statistical measurements of structure, naturalness, and perception. IEEE Transactions on Circuits and Systems for Video Technology, 30(4):929-943, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.636, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[20] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 21808-21820. Curran Associates, Inc., 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[21] Kede Ma, Wentao Liu, Kai Zhang, Zhengfang Duanmu, Zhou Wang, and Wangmeng Zuo. End-to-end blind image quality assessment using deep neural networks. IEEE Transactions on Image Processing, 27(3):1202-1213, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[22] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21(12):4695-4708, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[23] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE Transactions on Image Processing, 20(12):3350-3364, 2011. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16750" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[24] Chaithanya Kumar Mummadi, Robin Hutmacher, Kilian Rambach, Evgeny Levinkov, Thomas Brox, and Jan Hendrik Metzen. Test-time adaptation to distribution shift by confidence maximization and input transformation, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[25] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[26] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, and C.-C. Jay Kuo. Image database tid2013: Peculiarities, results and perspectives. Signal Processing: Image Communication, 30:57-77, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.36 + ], + "angle": 0, + "content": "[27] Michele A. Saad, Alan C. Bovik, and Christophe Charrier. Blind image quality assessment: A natural scene statistics approach in the dct domain. IEEE Transactions on Image Processing, 21(8):3339-3352, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.459 + ], + "angle": 0, + "content": "[28] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 11539-11551. Curran Associates, Inc., 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[29] H.R. Sheikh, M.F. Sabir, and A.C. Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on Image Processing, 15(11):3440-3451, 2006. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[30] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[31] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 9229-9248. PMLR, 13-18 Jul 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.756 + ], + "angle": 0, + "content": "[32] Toni Virtanen, Mikko Nuutinen, Mikko Vaahteranoksa, Pirkko Oittinen, and Jukka Häkkinen. Cid2013: A database for evaluating no-reference image quality assessment algorithms. IEEE Transactions on Image Processing, 24(1):390-402, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[33] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[34] Jingtao Xu, Peng Ye, Qiaohong Li, Haiqing Du, Yong Liu, and David Doermann. Blind image quality assessment based on high order statistics aggregation. IEEE Transactions on Image Processing, 25(9):4444-4457, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[35] Peng Ye, Jayant Kumar, Le Kang, and David Doermann. Unsupervised feature learning framework for no-reference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "image quality assessment. In 2012 IEEE conference on computer vision and pattern recognition, pages 1098-1105. IEEE, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[36] Peng Ye, Jayant Kumar, Le Kang, and David Doermann. Unsupervised feature learning framework for no-reference image quality assessment. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1098-1105, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[37] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[38] Guangtao Zhai, Xiongkuo Min, and Ning Liu. Free-energy principle inspired visual quality assessment: An overview. Digital Signal Processing, 91:11-20, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.333, + 0.892, + 0.4 + ], + "angle": 0, + "content": "[39] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.893, + 0.472 + ], + "angle": 0, + "content": "[40] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16751" + } + ] +] \ No newline at end of file diff --git a/2023/Test Time Adaptation for Blind Image Quality Assessment/4f23b86f-f4db-4257-aee1-6b77c4f73657_origin.pdf b/2023/Test Time Adaptation for Blind Image Quality Assessment/4f23b86f-f4db-4257-aee1-6b77c4f73657_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..76986b946e17a375897b71ebfdda0752d141b3b9 --- /dev/null +++ b/2023/Test Time Adaptation for Blind Image Quality Assessment/4f23b86f-f4db-4257-aee1-6b77c4f73657_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6d4329fec78e5b4f88a5ee3549d576c96762724ac74bd2a04495f45e422fb9c +size 3025265 diff --git a/2023/Test Time Adaptation for Blind Image Quality Assessment/full.md b/2023/Test Time Adaptation for Blind Image Quality Assessment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..df8e1ae14b7e1e3128c52a8823a5d0e93f8a7a1f --- /dev/null +++ b/2023/Test Time Adaptation for Blind Image Quality Assessment/full.md @@ -0,0 +1,290 @@ +# Test Time Adaptation for Blind Image Quality Assessment + +Subhadeep Roy* Shankhanil Mitra* Soma Biswas Rajiv Soundararajan Indian Institute of Science, Bengaluru, India + +subhadeeproy2000@gmail.com, {shankhanilm, somabiswas, rajivs}@iisc.ac.in + +# Abstract + +While the design of blind image quality assessment (IQA) algorithms has improved significantly, the distribution shift between the training and testing scenarios often leads to a poor performance of these methods at inference time. This motivates the study of test time adaptation (TTA) techniques to improve their performance at inference time. Existing auxiliary tasks and loss functions used for TTA may not be relevant for quality-aware adaptation of the pre-trained model. In this work, we introduce two novel quality-relevant auxiliary tasks at the batch and sample levels to enable TTA for blind IQA. In particular, we introduce a group contrastive loss at the batch level and a relative rank loss at the sample level to make the model quality aware and adapt to the target data. Our experiments reveal that even using a small batch of images from the test distribution helps achieve significant improvement in performance by updating the batch normalization statistics of the source model. + +# 1. Introduction + +The problem of image quality assessment (IQA) is extremely important in diverse image capture, processing, and sharing applications. However, a reference image is often not available for quality assessment. No reference (NR) or blind IQA primarily deals with the question of predicting image quality without using a reference image. Such NR IQA algorithms are often designed using machine learning approaches. More recently, deep learning based approaches have been extremely successful in achieving impressive performance. However, IQA applications are quite diverse and deal with several different distortions and distributional shifts. IQA models often have poor generalization ability and find it difficult to perform well under such shifts. + +Test time adaptation (TTA) has emerged as an important approach to address distributional shifts at test time [8]. It has been shown that by modifying a few global parameters of the model using a suitable loss that does not require the + +ground truth, one can significantly improve the performance of the model on the test data. Further, source-free adaptation, where the source data on which the original model was trained is not available while updating the model, is a realistic setting. While such approaches have been studied extensively in image classification literature [31, 33], there is hardly any literature on TTA for IQA. + +There are multiple challenges in designing TTA for IQA. Typical losses used for TTA, such as entropy minimization [33], are not applicable for IQA. For example, IQA is often studied in the regression context. This makes it difficult to extend models based on class confidences [24] or class prototypes [10] for classification to IQA. Also, the relevance of other self-supervised tasks such as rotation prediction [14], context prediction [5], colorization [15], noise prediction [1], feature clustering [3] for adapting IQA models is not clear. While contrastive learning has also been employed for TTA [20], such a framework is not explicitly based on contrasting image quality, and its relevance is also not clear. + +Our main contribution is in the design of auxiliary tasks to enable TTA for IQA. We start with a source model trained on a large IQA dataset and fine-tune the model on individual batches of test samples. The first task we introduce for adaptation is based on contrasting groups of low and high quality images in a batch. Thus, we exploit the initial knowledge of the source model and try to adapt it by enforcing quality relationships among the batch of samples. Such a group contrastive (GC) learning approach fits naturally to our setting to account for any errors on individual samples that the source model may be prone to. + +In contrast to the GC learning that depends on the batch, our second auxiliary task is an image specific task based on distorted augmentations of different types. Here, our goal is to enable the model to rank the image quality of further distorted versions of each test sample. We explore the role of different distortion types to leverage the maximum benefit of this task. While GC learning is more effective when samples in a batch are diverse in quality, the rank order based learning is more effective when the quality of the images is not extremely poor. Thus, a combination of the tasks + +helps overcome the shortcoming of both tasks and leads to an overall superior performance. + +We study the TTA problem under different settings of source and target datasets for multiple state of the art IQA models. Our results show significant improvements of the source model and the importance of TTA for IQA. We summarize our main contributions as follows: + +- We propose source-free test time adaptation techniques in the context of blind image quality assessment to mitigate distribution shifts between train and test data. +- We formulate two novel quality-aware self-supervised auxiliary tasks to adapt to the test data distribution. While group contrastive learning helps capture quality discriminative information among several images within a batch, rank ordering helps maintain the quality order between two different distorted versions of the same image. +- We show that our TTA method can significantly improve the performance of four different quality-aware source models, each on four different test IQA databases. + +# 2. Related Work + +# 2.1. Test Time Adaptation + +One of the first pieces of work on TTA [31] introduces a joint training framework using a loss for the main task and a self-supervised auxiliary task loss. The choice of a relevant auxiliary task is a challenging part of the design. Sun et al. [31] use rotation prediction as a pre-text task for image classification applications. Researchers also explore simpler tasks that are highly correlated with the main task, such as feature alignment with the batch and a simple contrastive loss framework [20] for adaptation. However, these methods require the source data to train the source model all over again to enable TTA. + +TENT [33] studies source-free TTA, where entropy minimization-based adaptation even outperforms some methods that use the source data. Moreover, only the batch normalization parameters of the model are adapted in TENT. There are several works on batch norm statistics adaptation [25, 28] to improve the robustness of the model at test time. SHOT [16] presents a clustering-based pseudolabeling method to align features from the target domain to the source domain using an information maximization loss. While a plethora of methods exists as above, the auxiliary tasks in these methods are not relevant for IQA. + +# 2.2. No-Reference Image Quality Assessment + +Most classical NR IQA algorithms are mainly based on natural scene statistics (NSS) [19, 22, 38]. In [23], the naturalness of the distorted image in the wavelet domain is modeled based on NSS. Saad et al. [27] also design the NSS model in the discrete cosine transform (DCT) domain. CORNIA [35], and HOSA [34] are among the earliest codebook learning based methods to predict quality. With the emergence of deep learning and the availability of large subject-rated IQA databases, various general-purpose NR IQA methods have been designed based on convolutional neural network (CNN) architectures [2, 18, 21, 36]. Some of these methods require end-to-end training [2, 21] of deep neural networks (DNN), while others [7, 12, 30, 39] are based on updating pre-trained models with some modifications. + +More recently, Zhang et al. [39] propose a method to train bilinear DNNs to simultaneously model both authentic and synthetic distortions. Su et al. [30] design a self-adaptive hyper network to provide weights for the quality prediction module parameters and handle various types of distortions and content in the images. Several methods [7, 12] explore transformer-based architectures along with a CNN to capture dependencies between local and global features. MetaIQA [40] proposes meta-learning on synthetic distortions by using a shared quality prior knowledge model to adapt to any kind of distortion. + +All the existing methods assume that the train and test data come from the same distribution. If there is a distribution shift across different databases, we need adaptation of the pre-trained model to learn about the target distributional information. + +# 3. Methodology + +We propose a novel self-supervised Test Time Adaptation technique for Image Quality Assessment (TTA-IQA) to adapt pre-trained quality models and mitigate distribution shifts between the source and target data. We consider source-free TTA, where we only have access to the pre-trained quality-aware models and no access to the source training data. When a batch of test data $D = \{x_{j}\}_{j=1}^{n}$ arrives, we adapt the model using the batch without knowledge of corresponding ground truth $\{y_{j}\}_{j=1}^{n}$ . + +# 3.1. Approach + +Let the model trained on the source data be $f_{\theta}$ where $\theta = (\theta_e, \theta_c)$ corresponds to the parameters of the network, and $\theta_e$ and $\theta_c$ correspond to the parameters of the feature extractor and regression layers. Thus, for an input image $x$ , we denote $f_{\theta}(x) = f_{\theta_c}(f_{\theta_e}(x))$ . Since our goal is to learn the distribution shift between train and test data, we only update the parameters of the feature extractor, $\theta_e$ , to + +![](images/c9a8509225d28991a43f53554069fc95fb969f3955e955998951bd882f811d93.jpg) +Figure 1. Block diagram of a general architecture for test time adaptation. At test time training, the normalization layers of the feature extractor are adapted by optimizing the combination of rank and GC loss. At inference time, we predict the quality scores of test images using the updated feature extractor and pre-trained quality regressor. + +align the features between the train and test distributions in a lower dimensional space. + +The key challenge in TTA is the choice of a self-supervised auxiliary task that is highly correlated with the main task of IQA. However, relying too much on the auxiliary task can affect the performance of the main task. To prevent loss of learnt information induced by the auxiliary task [4], we first project the features to a lower dimensional space using a non-linear projection head $f_{\theta_s}$ , parameterized by $\theta_s$ . The auxiliary task drives the adaptation of the feature extractor through the projection head. Thus our model now has three parts parametrized by $(\theta_e, \theta_s, \theta_c)$ to resemble a Y-shape as shown in Figure 1. + +When a batch of test instances $D$ arrives, we extract features from the feature extractor, followed by the projection head, and update the set of parameters $(\theta_e,\theta_s)$ by optimizing a self-supervised objective function $\mathcal{L}_s(D)$ . Updating all the model parameters of the feature extractor $\theta_{e}$ can cause the model to diverge too much from training, and the performance can drop drastically. Inspired by prior work on test time entropy minimization [33] and improving robustness for test data [28], we only update the linear and lower-dimensional feature modulation parameters. In a neural network, normalization layers satisfy these properties. So we only adapt the batch normalization layers by updating the affine parameters to mitigate distributional shifts. Thus we update these parameters by optimizing the auxiliary task loss given by + +$$ +\theta_ {e} ^ {*}, \theta_ {s} ^ {*} = \underset {\left(\theta_ {e}, \theta_ {s}\right)} {\operatorname {a r g m i n}} \mathcal {L} _ {s} (D) \tag {1} +$$ + +After optimizing the above loss function, we use the updated feature extractor parameters $\theta_e^*$ and pre-trained quality regressor parameters $\theta_c$ to predict the quality for a batch of test data. Since the distribution across batches can vary significantly, we ignore the earlier updated model and start the TTA based on source model weights for new test instants. Thus the updated target model always only depends + +on the incoming test data. + +# 3.2. Self-Supervised Auxiliary Tasks + +Our goal is to carefully choose self-supervised auxiliary tasks that capture quality-aware distributional information to adapt the feature encoder. We formulate two novel and complementary self-supervised learning techniques which help to learn the distribution shift between train and test data. These self-supervised objective functions are - 1) Group Contrastive Loss and 2) Rank Loss. While the GC loss works well when there is a reasonable separation of quality among the samples in a batch, the rank loss works better even when the quality of the batch samples is similar. On the other hand, the rank loss is meaningful only when the quality of the input image is not extremely low, while the GC loss is independent of the quality of a given image. Thus, a combination of the two losses renders our TTA extremely effective across various scenarios. + +# 3.2.1 Group Contrastive Loss + +While contrastive learning has been used for TTA of deep image classification, its direct application does not appear relevant to the task of IQA. Thus, we introduce group contrastive (GC) learning as an auxiliary task for TTA of IQA models. In particular, we make two groups of images from a single batch of $N$ images based on the pseudo-labels given by the pre-trained source model. We sort the images in ascending order as $x_{(1)}, x_{(2)}, \ldots, x_{(N)}$ based on the pseudolabels, where $x_{(i)}, i = 1, 2, \ldots, N$ , corresponds to the $i^{th}$ lowest quality image in the batch. We then segregate images with high quality scores (say, top $p$ fraction of the data in a batch) and include them in a group of higher-quality images. Similarly, we separate out lower quality images (say, the lowest $p$ fraction of the images in a batch) together to form another group. Here we assume that $pN$ and $(1 - p)N$ are integers for simplicity; else, they can be rounded off to the nearest integers. + +![](images/f7d8fac2d877a1359ceeb87ce57c4b7638faa7a5b704c0d452944e44000910ca.jpg) +Figure 2. The overview framework of group contrastive loss for test time adaptation. For a batch of images, two different groups are formed based on pseudo-labels given by the source model. The group contrastive loss tries to minimize the distance between features extracted from images belonging to the same group while maximizing the distance between features from different groups. + +The premise behind our loss for GC learning is that images from the same quality group should give similar feature representations in a lower dimensional space while features of images from different groups are separated out. Thus image pairs from the same group act as positive pairs, and image pairs from different groups act as negative pairs. By separating out these two groups, the model adapts itself by better separating the intermediate quality samples. + +Let a positive pair $x_{(i)}$ and $x_{(j)}$ come from the same group, i.e., either both $i, j \leq pN$ or $i, j > (1 - p)N$ . We use a modified NT-Xent contrastive loss [4] as our objective function. For a pair of images coming from the lower quality group where $i, j \leq pN$ and $i \neq j$ , the GC loss is defined as + +$$ +\mathcal {L} _ {i, j} ^ {g c} = - \log \frac {\exp (\operatorname* {s i m} (z _ {(i)} , z _ {(j)}) / \tau)}{\sum_ {k > (1 - p) N} ^ {N} \exp (\operatorname* {s i m} (z _ {(i)} , z _ {(k)}) / \tau)}, \tag {2} +$$ + +where $\mathbf{z}_{(i)} = f_{\theta_s}(f_{\theta_e}(x_{(i)}))$ represents the feature at the output of the projection head for sample $x_{(i)}$ , sim refers to the cosine similarity between two features. and $\mathbf{z}_{(k)}$ is the feature extracted from an image from the higher quality group with $k > (1 - p)N$ . Also, $\tau$ represents temperature scaling parameter. While we define the loss above when $i,j\leq pN$ , we can define a similar loss when $i,j > (1 - p)N$ . For all pairs of images within the same group, we obtain the GC loss and add them together to obtain the final loss, $\mathcal{L}^{gc}$ as + +$$ +\mathcal{L}^{gc} = \sum_{i = 1}^{pN}\sum_{\substack{j = 1\\ j\neq i}}^{pN}\mathcal{L}_{i,j}^{gc} + \sum_{i = (1 - p)N + 1}^{N}\sum_{\substack{j = (1 - p)N + 1\\ j\neq i}}^{N}\mathcal{L}_{i,j}^{gc}. \tag{3} +$$ + +A block diagram of the GC loss is shown in Figure 2. + +# 3.2.2 Rank Loss + +Given a sample of images from test data, the rank loss helps adapt the features by capturing quality-aware distributional information at the sample level. We introduce a ranking objective to learn the quality orders between two distorted versions of the test images that are quality discriminable. We distort each test image $x_{i}$ , $i = 1,2,\dots ,N$ , in the minibatch of size $N$ with two different degrees of degradation of a given distortion type. Let $x_{i}^{high}$ and $x_{i}^{low}$ denote the higher and lower distorted image, respectively. The degradation types include synthetic distortions such as blur, compression, and noise. We choose degradation levels randomly from two sets of parameters sufficiently farther apart. Ideally, the distance between the features extracted from the test image and the highly distorted image should be greater than that between the features extracted from the test image and the lower distorted image. Our rank loss tries to capture this order and adapt the model to the test data. + +We project the triplet of images $(x_{i}, x_{i}^{high}, x_{i}^{low})$ in the feature space through the feature extractor and the projection head. We measure the distance between the distortion-augmented image and the original test data using the Euclidean distance in the projected feature space. Mathematically, let $\mathbf{z}_{i} = f_{\theta_{s}}(f_{\theta_{e}}(x_{i}))$ be the feature extracted for test sample $x_{i}$ . Similarly we define $\mathbf{z}_{i}^{high}$ and $\mathbf{z}_{i}^{low}$ . Let $d_{i}^{high}$ denote the distance between $\mathbf{z}_{i}$ and $\mathbf{z}_{i}^{high}$ . Similarly, $d_{i}^{low}$ denotes the distance between $\mathbf{z}_{i}$ and $\mathbf{z}_{i}^{low}$ . The target model is now fine-tuned to obtain the correct ranking between the distances by achieving $d_{i}^{high} \geq d_{i}^{low}$ shown in Figure 3. + +The probability of achieving this order is estimated by passing the difference in distances through a sigmoid func + +![](images/cfff0b4fd362c208c3eae5f7b14560808024eac16a0865be0cd580daca1a0a59.jpg) +Figure 3. The rank loss framework for test time adaptation. We distort the test image with two different degrees of degradation. We then project the triplet of images in the features space through the shared feature extractor and projection head. Distances calculated between the features extracted from each distorted image and the original test image are used to compare their rank order. + +tion as + +$$ +P _ {i} = \Pr \left(d _ {i} ^ {\text {h i g h}} \geq d _ {i} ^ {\text {l o w}}\right) = \frac {\exp \left(d _ {i} ^ {\text {h i g h}} - d _ {i} ^ {\text {l o w}}\right)}{1 + \exp \left(d _ {i} ^ {\text {h i g h}} - d _ {i} ^ {\text {l o w}}\right)}. \tag {4} +$$ + +We use binary cross-entropy loss between true label $\bar{P}_i = 1$ and the predicted probability $P_{i}$ to obtain the rank loss as + +$$ +\mathcal {L} _ {i} ^ {r} (\bar {P} _ {i}, P _ {i}) = - \bar {P} _ {i} \log P _ {i} - (1 - \bar {P} _ {i}) \log (1 - P _ {i}). (5) +$$ + +For a batch of size $N$ , the overall rank loss is given as + +$$ +\mathcal {L} ^ {r} = \sum_ {i = 1} ^ {N} \mathcal {L} _ {i} ^ {r}. \tag {6} +$$ + +One of the challenges in the rank loss is selecting the distortion type for every image. The original test image may already be distorted, which poses difficulties in deciding the distortion type that can help TTA. For example, if the test image is extremely blurred and we blur it again with two different levels, both images may be visually indistinguishable, thereby limiting the extent to which adaptation can help. We exploit the source model's knowledge to overcome this limitation in choosing the distortion type. We predict the quality scores of the pair $(x_{i}^{high}, x_{i}^{low})$ for each distortion type using the source model. We hypothesize that the pair with the maximal difference in predicted quality is sufficiently different in visual quality. Thus choosing such a pair in Equation (5) can help adapt the model. + +Our overall self-supervised TTA objective function is a + +combination of both the rank and GC loss given as + +$$ +\mathcal {L} _ {s} = \mathcal {L} ^ {g c} + \lambda \mathcal {L} ^ {r}, \tag {7} +$$ + +where $\lambda$ is a hyper-parameter used to combine the losses. + +# 4. Experiments + +# 4.1. Quality Models, Datasets and Metrics + +We evaluate TTA on four popular IQA databases using four different quality-aware models. In particular, we consider state-of-the-art deep IQA models such as TReS [7], MUSIQ [12], HyperIQA [30], and MetaIQA [40]. These models contain a ResNet backbone with batch normalization layers, which we model as the feature extractor $f_{\theta_e}$ and only update its batch normalization parameters. We model the rest of the network as the quality regressor part, $f_{\theta_c}$ . TReS and MUSIQ use transformers as a part of their architecture, which we include as a part of the quality regressor in all our main experiments. We also explore the adaptation of transformers by including them as part of the feature extractor and updating their layer normalization statistics in the supplementary material. + +We project the features extracted from the last layer of ResNet through a 256-dimensional fully connected (FC) layer corresponding to the self-supervised projection head $f_{\theta_s}$ . These lower dimensional features are used to adapt the model at test time. Three IQA models, TReS, MUSIQ, and HyperIQA are trained on the LIVEFB database [37] containing 39,810 images. MetaIQA is trained on two synthetically distorted databases, TID2013 [26] and KADID-10k [17]. + +
BackboneDatabaseKonIQ-10kPIPALCID2013LIVE-IQA
MethodSROCCPLCCSROCCPLCCSROCCPLCCSROCCPLCC
TReSBaseline0.65200.69550.38450.40780.52720.64630.54350.4450
Rotation0.65060.68050.40610.41140.57060.66510.58660.5311
TTA-IQA0.65780.70740.42780.42040.60320.67100.67220.5963
MUSIQBaseline0.63040.68020.31900.34140.51730.60320.25960.3351
Rotation0.65770.71540.36650.36930.54870.61640.35120.3976
TTA-IQA0.66930.72300.37430.37310.54990.62200.36490.4031
HyperIQABaseline0.58610.63130.30370.33040.48950.61230.51430.4377
Rotation0.60330.65360.32680.34820.49020.61500.62680.5469
TTA-IQA0.59600.64950.36530.37670.50390.59880.62180.5438
MetaIQABaseline0.51620.44600.32870.29550.72130.68170.73230.6732
Rotation0.58230.53110.33530.30420.71770.67450.72710.6868
TTA-IQA0.58380.54280.40730.35100.78090.73990.79990.7726
+ +Table 1. Comparison of TTA-IQA with popular NR IQA methods and one popular auxiliary task - rotation prediction on authentically and synthetically distorted datasets. Bold entries imply the best performance for every individual quality-aware model on respective datasets. + +We choose challenging databases to evaluate the generalization capability of our TTA-IQA method. The test datasets are described as follows: + +KonIQ-10k [9] is a popular in-the-wild authentically distorted database consisting of 10,073 quality-scored images. + +PIPAL [11] is a large-scale IQA database for evaluating perceptual image restoration. It contains 29k distorted images, including 19 different GAN-based algorithms. + +CID2013 [32] consists of 480 images in six image sets captured by 79 imaging devices. + +LIVE-IQA [29] consists of 779 synthetically distorted images from 29 reference images. + +We evaluate our results using Spearman's rank-order correlation coefficient (SROCC) and Pearson's linear correlation coefficient (PLCC). + +# 4.2. Implementation Details + +We implement our setup in PyTorch and conduct all the experiments with an 16 GB NVIDIA RTX A4000 GPU. During TTA, we randomly select a patch of size $224 \times 224$ from the input image and apply quality preserving augmentations such as horizontal flip and vertical flip before passing it through the network for TTA. We use the ADAM [13] optimizer with a learning rate of 0.001 and set the number of iteration as 3. All the experiments were run on five different seeds using a batch size of 8, and the final results were obtained by averaging. + +With regard to the self-supervised auxiliary tasks, we use a combination of rank loss and GC loss as our objective function using $\lambda = 1$ given in Equation (7). We choose $p = 0.25$ to determine the groups. We calculate the GC loss using $\tau = 1$ . For the rank loss, the test image is synthetically blurred using a Gaussian blur filter of size $5 \times 5$ with two sets of standard deviations $\sigma$ , for the Gaussian + +blur kernel. We keep $\sigma \in [40,80]$ for highly distorted images and $\sigma \in [1,20]$ for less blurred images. For compression distortions, we specify the quality factor in [80, 95] for lower compression and [30, 60] for higher compression rates. Similarly, we add zero mean Gaussian white noise to the test image with variance in [0.05, 0.1] for higher distortion and in [0.005, 0.01] for lower distortion. + +# 4.3. Performance Evaluation + +Table 1 shows the performance of our method on all four test datasets using all four quality-aware models. In addition to the source model, we also compare with rotation prediction [31] as the auxiliary task during TTA. A comparison with such a task helps understand the role of quality-aware losses for the TTA of IQA models. + +We observe that TTA-IQA using the combination of rank and GC loss outperforms the source models in all the cases. Note that the PIPAL dataset has a huge distribution shift from the authentically distorted LIVEFB dataset on which three of the models were trained. While most source models perform very poorly on PIPAL, we achieve $10\% -20\%$ improvement over the source models. On the KonIQ-10k dataset, TTA-IQA gives around $1\% -10\%$ improvement over the source model. As both KonIQ-10k and LIVEFB are authentically distorted datasets, the distribution shift is reasonably small, leading to smaller improvements using adaptation. CID2013 is also an authentically distorted dataset and gives similar improvements of around $2\% -13\%$ over the source models. Our experiments on the LIVE-IQA database provide a significantly greater improvement of $10\% -33\%$ owing to the shift from authentic to synthetic distortions for three models. + +We also observe from Table 1 that our approach outperforms the rotation prediction task in most cases. Other TTA ideas, such as TENT [33] and training using masked autoen + +
BackboneDatabaseKonIQ-10kPIPALCID2013LIVE-IQA
RankGCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCC
TReS✓×0.65620.69890.41710.42040.60160.67360.67050.5908
×✓0.65160.69460.46660.51830.53660.64930.71600.6290
✓✓0.65780.70740.42780.42040.60320.67100.67220.5963
MUSIQ✓×0.65490.71490.37680.37180.52160.60340.36340.4024
×✓0.66110.71760.35850.36420.54460.61590.32990.3914
✓✓0.66930.72300.37430.37310.54990.62200.36490.4031
HyperIQA✓×0.59280.64550.36160.37320.50390.59910.55050.5050
×✓0.60940.65670.33330.35520.51200.62550.63310.5592
✓✓0.59600.64950.36530.37670.50390.59880.62180.5438
MetaIQA✓×0.55800.51180.39920.34370.75790.72820.78940.7534
×✓0.54140.46360.37100.32870.78610.70670.75660.6927
✓✓0.58380.54280.40730.35100.78090.73990.79990.7726
+ +Table 2. Ablation study results on authentically and synthetically distorted datasets using popular quality aware source models. Bold entries imply the best performance among all three settings. + +![](images/b2a67dd12412c5190c97d77c2b4e37b7602a5953485e000dae19d17651968b99.jpg) +Figure 4. Scatter plot for predicting quality of similar quality images from LIVE-IQA dataset for the source model and TTA using only rank loss or GC loss. We mark two sets of images having similar human opinion score (DMOS) by ellipses in each case. + +![](images/b7bd95081d74e2ce9ecfbb5a0049f21149d987219de81d803708f0ced53b53fa.jpg) + +![](images/87fd3f40ea4958c2c0f6c22ded6f839fe1920dd2d1bd37a605b6ef5cd8f59904.jpg) + +coders [6] are not applicable for the IQA task. In particular, the notion of entropy in regression tasks is not clear. On the other hand, masked reconstructions tend to lead to quality degradation. Thus, we do not compare with them. + +# 4.4. Ablation Study & Other Experiments + +Need for Rank Loss and GC Loss for TTA. We perform an ablation study with respect to the two losses in Table 2. We see from the results that the rank loss and the GC loss individually always improve on the source models. Further, the combination of the rank and GC loss provides an even better performance over the individual losses in most evaluation scenarios. Even in scenarios, where the combination loss achieves the second-best performance, its performance is very close to the best. For the rest of the experiments in this section, we present results on all four datasets using the TReS method alone. + +We now discuss scenarios where the GC loss is particularly more useful than the rank loss. If the test image is extremely distorted, none of the three kinds of distortion types can create much difference in the perceptual quality of the two degraded versions. To understand this further, we consider highly distorted images (corresponding to a dif + +ferential mean opinion score (DMOS) greater than 70) of different distortion types in the LIVE-IQA dataset. We apply TTA only for these images using different losses. The SROCC performance for these images in Figure 5 reveal that the rank loss is not very effective. However, the GC loss works well in such scenarios. + +Conversely, we also discuss scenarios where the rank loss is more useful than the GC loss. To illustrate this idea, we select multiple images having similar quality with DMOS in [28, 44] from the LIVE-IQA dataset. We adapt the source model for this set of images using both the GC and rank losses. From Figure 4, we observe that the rank loss leads to much better adaptation in terms of the resulting model correlating with the ground truth scores. Intuitively, when the input images have a similar quality, the pseudolabels given by the source model may be noisy, leading to an inaccurate grouping of the images into the lower and higher quality groups. This can adversely affect the GC loss. So the rank loss is more effective than the GC loss for test batches with similar quality images. + +Effect of Selecting Multiple Distortion Types in the Rank Loss. We experiment with the impact of different kinds of distortion types while incorporating the rank loss. + +![](images/ca0e444cd1e1d3fbe99e30372c6eb592b42876d877fcb1c4894a6afaa4f4b806.jpg) +Results for Highly Distorted Images +Figure 5. TTA-IQA on different distortion-specific images from LIVE-IQA dataset with respect to individual losses + +In particular, we consider three scenarios: one where only one of the distortion types is used in the rank loss, another where all the three types are used to obtain three loss terms which are summed together to update the model, and a third, where only one of the distortion types is chosen based on the discussion in Section 3.2.2. We observe from Table 3 that choosing the distortion type with the maximal difference in quality gives the best performance for the rank loss. This experiment validates our hypothesis that only a rank loss that can discriminate between the degraded image pairs is useful for TTA. + +
DatabaseBlurCompNoiseAllBest
KonIQ-10k0.6560.6560.6670.6150.656
PIPAL0.3840.4150.3890.4150.417
CID20130.5270.6090.6060.5600.609
LIVE-IQA0.5780.6050.6180.6340.671
+ +Selection of Group Size by Varying $p$ . We explore different values of $p$ for constructing the GC loss. For a batch size of 8, possible choices of the group size are 2, 3, 4 corresponding to $p = 0.25, 0.375, 0.5$ respectively. From Table 4, we observe that the performances are roughly similar for different values of $p$ with a slightly superior performance for $p = 0.25$ . We note that as $p$ increases, the groups are larger and probably less discriminative in terms of quality, leading to a slightly poorer performance. + +Table 3. Impact of selecting multiple distortion types on rank loss + +
Databasep=0.25p=0.375p=0.5
KonIQ-10k0.65160.64970.6467
PIPAL0.46660.46020.4616
CID20130.53660.54860.5536
LIVE-IQA0.71600.71710.6937
+ +Selection of Number of Groups for GC Loss. In our formulation, we define the GC loss only between two con + +trastive groups. In principle, we can extend our idea of GC loss to multiple groups. In particular, we can cluster the images in a batch into multiple groups and apply the GC loss between every pair of groups and sum the loss terms. Thus, images coming from the same group act as positive pairs, and images from two different groups are considered as negative pairs. From Table 5, we observe that the number of groups $G$ does not impact the performance much. + +Table 4. Impact of varying group size using $p$ for GC loss + +
DatabaseG=2G=3G=4
KonIQ-10k0.65160.65240.6509
PIPAL0.46660.46680.4663
CID20130.53660.53270.5388
LIVE-IQA0.71600.70110.6883
+ +Table 5. Impact of varying number of groups $G$ for GC loss + +Choice of Number of Iterations for Learning Auxiliary Task. We also examine the effect of the number of iterations of parameter updates during TTA. Figure 6 shows that the performance on CID2013 dataset is fairly robust until 4 iterations. Beyond that, the model overfits the auxiliary task and leads to poor performance at test time. + +![](images/e8585276d93fbbe16b41ecc7e3ba270b2b147debf9baea6eb4de19e5d629d7ed.jpg) +Figure 6. Effect of increasing number of iterations + +# 5. Conclusion + +Test time adaptation has become very popular due to its simplicity and the lack of need for end-to-end training. Our work is perhaps one of the first attempts to design the method of TTA in the context of blind IQA. While most IQA methods focus on making the models robust enough to perform well for cross-database experiments, our TTA-IQA method can outperform existing state-of-the-art methods because of the adaptation at test time. We formulate novel self-supervised auxiliary tasks using the rank and group contrastive losses, which can learn quality-aware information from the test data. While primarily explored TTA for IQA, it would be interesting to understand the role of TTA for video quality assessment as well. + +Acknowledgement: This work was supported in part by a grant from Kotak IISc AI-ML Centre and Department of + +Science and Technology, Government of India, under grant CRG/2020/003516. + +# References + +[1] Piotr Bojanowski and Armand Joulin. Unsupervised learning by predicting noise. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 517-526. PMLR, 06-11 Aug 2017. 1 +[2] Sebastian Bosse, Dominique Maniry, Klaus-Robert Müller, Thomas Wiegand, and Wojciech Samek. Deep neural networks for no-reference and full-reference image quality assessment. IEEE Transactions on Image Processing, 27(1):206-219, 2018. 2 +[3] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 139–156, Cham, 2018. Springer International Publishing. 1 +[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org, 2020. 3, 4 +[5] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), December 2015. 1 +[6] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A. Efros. Test-time training with masked autoencoders. arXiv preprint arXiv:2209.07522, 2022. 7 +[7] S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3209-3218, 2022. 2, 5 +[8] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 1 +[9] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. Trans. Img. Proc., 29:4041-4056, jan 2020. 6 +[10] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 2427–2440. Curran Associates, Inc., 2021. 1 +[11] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S. Ren, and Dong Chao. Pipal: A large-scale image quality assessment dataset for perceptual image restoration. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, + +pages 633-651, Cham, 2020. Springer International Publishing. 6 +[12] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5148-5157, October 2021. 2, 5 +[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[14] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), Vancouver, Canada, Apr. 2018. 1 +[15] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Colorization as a proxy task for visual understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1 +[16] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? Source hypothesis transfer for unsupervised domain adaptation. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 6028-6039. PMLR, 13-18 Jul 2020. 2 +[17] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1–3, 2019. 5 +[18] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. 2 +[19] Yutao Liu, Ke Gu, Yongbing Zhang, Xiu Li, Guangtao Zhai, Debin Zhao, and Wen Gao. Unsupervised blind image quality evaluation via statistical measurements of structure, naturalness, and perception. IEEE Transactions on Circuits and Systems for Video Technology, 30(4):929-943, 2020. 2 +[20] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 21808-21820. Curran Associates, Inc., 2021. 1, 2 +[21] Kede Ma, Wentao Liu, Kai Zhang, Zhengfang Duanmu, Zhou Wang, and Wangmeng Zuo. End-to-end blind image quality assessment using deep neural networks. IEEE Transactions on Image Processing, 27(3):1202-1213, 2018. 2 +[22] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21(12):4695-4708, 2012. 2 +[23] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE Transactions on Image Processing, 20(12):3350-3364, 2011. 2 + +[24] Chaithanya Kumar Mummadi, Robin Hutmacher, Kilian Rambach, Evgeny Levinkov, Thomas Brox, and Jan Hendrik Metzen. Test-time adaptation to distribution shift by confidence maximization and input transformation, 2021. 1 +[25] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 2 +[26] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, and C.-C. Jay Kuo. Image database tid2013: Peculiarities, results and perspectives. Signal Processing: Image Communication, 30:57-77, 2015. 5 +[27] Michele A. Saad, Alan C. Bovik, and Christophe Charrier. Blind image quality assessment: A natural scene statistics approach in the dct domain. IEEE Transactions on Image Processing, 21(8):3339-3352, 2012. 2 +[28] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 11539-11551. Curran Associates, Inc., 2020. 2, 3 +[29] H.R. Sheikh, M.F. Sabir, and A.C. Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on Image Processing, 15(11):3440-3451, 2006. 6 +[30] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 5 +[31] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 9229-9248. PMLR, 13-18 Jul 2020. 1, 2, 6 +[32] Toni Virtanen, Mikko Nuutinen, Mikko Vaahteranoksa, Pirkko Oittinen, and Jukka Häkkinen. Cid2013: A database for evaluating no-reference image quality assessment algorithms. IEEE Transactions on Image Processing, 24(1):390-402, 2015. 6 +[33] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 1, 2, 3, 6 +[34] Jingtao Xu, Peng Ye, Qiaohong Li, Haiqing Du, Yong Liu, and David Doermann. Blind image quality assessment based on high order statistics aggregation. IEEE Transactions on Image Processing, 25(9):4444-4457, 2016. 2 +[35] Peng Ye, Jayant Kumar, Le Kang, and David Doermann. Unsupervised feature learning framework for no-reference + +image quality assessment. In 2012 IEEE conference on computer vision and pattern recognition, pages 1098-1105. IEEE, 2012. 2 +[36] Peng Ye, Jayant Kumar, Le Kang, and David Doermann. Unsupervised feature learning framework for no-reference image quality assessment. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1098-1105, 2012. 2 +[37] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5 +[38] Guangtao Zhai, Xiongkuo Min, and Ning Liu. Free-energy principle inspired visual quality assessment: An overview. Digital Signal Processing, 91:11-20, 2019. 2 +[39] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2020. 2 +[40] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 5 \ No newline at end of file diff --git a/2023/Test Time Adaptation for Blind Image Quality Assessment/images.zip b/2023/Test Time Adaptation for Blind Image Quality Assessment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dc6b4d895ceb9cd2537bda69e3e04d589fd83aec --- /dev/null +++ b/2023/Test Time Adaptation for Blind Image Quality Assessment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cce26cbc53c264b834f1b46301bc4212512fb0f81fffcca5a6129657e851884d +size 616395 diff --git a/2023/Test Time Adaptation for Blind Image Quality Assessment/layout.json b/2023/Test Time Adaptation for Blind Image Quality Assessment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..59a4b847eb250fa4bdb646812af68575c0be86d6 --- /dev/null +++ b/2023/Test Time Adaptation for Blind Image Quality Assessment/layout.json @@ -0,0 +1,7849 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 117, + 103, + 477, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 103, + 477, + 121 + ], + "spans": [ + { + "bbox": [ + 117, + 103, + 477, + 121 + ], + "type": "text", + "content": "Test Time Adaptation for Blind Image Quality Assessment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 143, + 481, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 143, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 111, + 143, + 481, + 171 + ], + "type": "text", + "content": "Subhadeep Roy* Shankhanil Mitra* Soma Biswas Rajiv Soundararajan Indian Institute of Science, Bengaluru, India" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 174, + 490, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 174, + 490, + 185 + ], + "spans": [ + { + "bbox": [ + 102, + 174, + 490, + 185 + ], + "type": "text", + "content": "subhadeeproy2000@gmail.com, {shankhanilm, somabiswas, rajivs}@iisc.ac.in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 441 + ], + "type": "text", + "content": "While the design of blind image quality assessment (IQA) algorithms has improved significantly, the distribution shift between the training and testing scenarios often leads to a poor performance of these methods at inference time. This motivates the study of test time adaptation (TTA) techniques to improve their performance at inference time. Existing auxiliary tasks and loss functions used for TTA may not be relevant for quality-aware adaptation of the pre-trained model. In this work, we introduce two novel quality-relevant auxiliary tasks at the batch and sample levels to enable TTA for blind IQA. In particular, we introduce a group contrastive loss at the batch level and a relative rank loss at the sample level to make the model quality aware and adapt to the target data. Our experiments reveal that even using a small batch of images from the test distribution helps achieve significant improvement in performance by updating the batch normalization statistics of the source model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 463, + 128, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 128, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 128, + 475 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 483, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 639 + ], + "type": "text", + "content": "The problem of image quality assessment (IQA) is extremely important in diverse image capture, processing, and sharing applications. However, a reference image is often not available for quality assessment. No reference (NR) or blind IQA primarily deals with the question of predicting image quality without using a reference image. Such NR IQA algorithms are often designed using machine learning approaches. More recently, deep learning based approaches have been extremely successful in achieving impressive performance. However, IQA applications are quite diverse and deal with several different distortions and distributional shifts. IQA models often have poor generalization ability and find it difficult to perform well under such shifts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "type": "text", + "content": "Test time adaptation (TTA) has emerged as an important approach to address distributional shifts at test time [8]. It has been shown that by modifying a few global parameters of the model using a suitable loss that does not require the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 214, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 214, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 306, + 214, + 545, + 297 + ], + "type": "text", + "content": "ground truth, one can significantly improve the performance of the model on the test data. Further, source-free adaptation, where the source data on which the original model was trained is not available while updating the model, is a realistic setting. While such approaches have been studied extensively in image classification literature [31, 33], there is hardly any literature on TTA for IQA." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 300, + 546, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 546, + 456 + ], + "type": "text", + "content": "There are multiple challenges in designing TTA for IQA. Typical losses used for TTA, such as entropy minimization [33], are not applicable for IQA. For example, IQA is often studied in the regression context. This makes it difficult to extend models based on class confidences [24] or class prototypes [10] for classification to IQA. Also, the relevance of other self-supervised tasks such as rotation prediction [14], context prediction [5], colorization [15], noise prediction [1], feature clustering [3] for adapting IQA models is not clear. While contrastive learning has also been employed for TTA [20], such a framework is not explicitly based on contrasting image quality, and its relevance is also not clear." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 459, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 547, + 590 + ], + "type": "text", + "content": "Our main contribution is in the design of auxiliary tasks to enable TTA for IQA. We start with a source model trained on a large IQA dataset and fine-tune the model on individual batches of test samples. The first task we introduce for adaptation is based on contrasting groups of low and high quality images in a batch. Thus, we exploit the initial knowledge of the source model and try to adapt it by enforcing quality relationships among the batch of samples. Such a group contrastive (GC) learning approach fits naturally to our setting to account for any errors on individual samples that the source model may be prone to." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "In contrast to the GC learning that depends on the batch, our second auxiliary task is an image specific task based on distorted augmentations of different types. Here, our goal is to enable the model to rank the image quality of further distorted versions of each test sample. We explore the role of different distortion types to leverage the maximum benefit of this task. While GC learning is more effective when samples in a batch are diverse in quality, the rank order based learning is more effective when the quality of the images is not extremely poor. Thus, a combination of the tasks" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 693, + 271, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 693, + 271, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 693, + 271, + 713 + ], + "type": "text", + "content": "*Authors contributed equally to this work\n§https://github.com/subhadeeproy2000/TTA-IQA" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16742" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "helps overcome the shortcoming of both tasks and leads to an overall superior performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 97, + 287, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 97, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 47, + 97, + 287, + 157 + ], + "type": "text", + "content": "We study the TTA problem under different settings of source and target datasets for multiple state of the art IQA models. Our results show significant improvements of the source model and the importance of TTA for IQA. We summarize our main contributions as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 173, + 286, + 376 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 58, + 173, + 286, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 173, + 286, + 219 + ], + "spans": [ + { + "bbox": [ + 58, + 173, + 286, + 219 + ], + "type": "text", + "content": "- We propose source-free test time adaptation techniques in the context of blind image quality assessment to mitigate distribution shifts between train and test data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 234, + 286, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 234, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 58, + 234, + 286, + 316 + ], + "type": "text", + "content": "- We formulate two novel quality-aware self-supervised auxiliary tasks to adapt to the test data distribution. While group contrastive learning helps capture quality discriminative information among several images within a batch, rank ordering helps maintain the quality order between two different distorted versions of the same image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 331, + 285, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 331, + 285, + 376 + ], + "spans": [ + { + "bbox": [ + 58, + 331, + 285, + 376 + ], + "type": "text", + "content": "- We show that our TTA method can significantly improve the performance of four different quality-aware source models, each on four different test IQA databases." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 408, + 133, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 133, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 133, + 420 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 429, + 171, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 171, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 171, + 442 + ], + "type": "text", + "content": "2.1. Test Time Adaptation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 449, + 286, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 286, + 579 + ], + "type": "text", + "content": "One of the first pieces of work on TTA [31] introduces a joint training framework using a loss for the main task and a self-supervised auxiliary task loss. The choice of a relevant auxiliary task is a challenging part of the design. Sun et al. [31] use rotation prediction as a pre-text task for image classification applications. Researchers also explore simpler tasks that are highly correlated with the main task, such as feature alignment with the batch and a simple contrastive loss framework [20] for adaptation. However, these methods require the source data to train the source model all over again to enable TTA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "TENT [33] studies source-free TTA, where entropy minimization-based adaptation even outperforms some methods that use the source data. Moreover, only the batch normalization parameters of the model are adapted in TENT. There are several works on batch norm statistics adaptation [25, 28] to improve the robustness of the model at test time. SHOT [16] presents a clustering-based pseudolabeling method to align features from the target domain to the source domain using an information maximization loss. While a plethora of methods exists as above, the auxiliary tasks in these methods are not relevant for IQA." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 518, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 518, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 518, + 84 + ], + "type": "text", + "content": "2.2. No-Reference Image Quality Assessment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 90, + 545, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 268 + ], + "type": "text", + "content": "Most classical NR IQA algorithms are mainly based on natural scene statistics (NSS) [19, 22, 38]. In [23], the naturalness of the distorted image in the wavelet domain is modeled based on NSS. Saad et al. [27] also design the NSS model in the discrete cosine transform (DCT) domain. CORNIA [35], and HOSA [34] are among the earliest codebook learning based methods to predict quality. With the emergence of deep learning and the availability of large subject-rated IQA databases, various general-purpose NR IQA methods have been designed based on convolutional neural network (CNN) architectures [2, 18, 21, 36]. Some of these methods require end-to-end training [2, 21] of deep neural networks (DNN), while others [7, 12, 30, 39] are based on updating pre-trained models with some modifications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 270, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 400 + ], + "type": "text", + "content": "More recently, Zhang et al. [39] propose a method to train bilinear DNNs to simultaneously model both authentic and synthetic distortions. Su et al. [30] design a self-adaptive hyper network to provide weights for the quality prediction module parameters and handle various types of distortions and content in the images. Several methods [7, 12] explore transformer-based architectures along with a CNN to capture dependencies between local and global features. MetaIQA [40] proposes meta-learning on synthetic distortions by using a shared quality prior knowledge model to adapt to any kind of distortion." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 402, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 460 + ], + "type": "text", + "content": "All the existing methods assume that the train and test data come from the same distribution. If there is a distribution shift across different databases, we need adaptation of the pre-trained model to learn about the target distributional information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 472, + 388, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 388, + 486 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 388, + 486 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 492, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 545, + 605 + ], + "type": "text", + "content": "We propose a novel self-supervised Test Time Adaptation technique for Image Quality Assessment (TTA-IQA) to adapt pre-trained quality models and mitigate distribution shifts between the source and target data. We consider source-free TTA, where we only have access to the pre-trained quality-aware models and no access to the source training data. When a batch of test data " + }, + { + "bbox": [ + 304, + 492, + 545, + 605 + ], + "type": "inline_equation", + "content": "D = \\{x_{j}\\}_{j=1}^{n}" + }, + { + "bbox": [ + 304, + 492, + 545, + 605 + ], + "type": "text", + "content": " arrives, we adapt the model using the batch without knowledge of corresponding ground truth " + }, + { + "bbox": [ + 304, + 492, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\{y_{j}\\}_{j=1}^{n}" + }, + { + "bbox": [ + 304, + 492, + 545, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 611, + 375, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 375, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 375, + 624 + ], + "type": "text", + "content": "3.1. Approach" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Let the model trained on the source data be " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta = (\\theta_e, \\theta_c)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " corresponds to the parameters of the network, and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_e" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_c" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " correspond to the parameters of the feature extractor and regression layers. Thus, for an input image " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", we denote " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x) = f_{\\theta_c}(f_{\\theta_e}(x))" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". Since our goal is to learn the distribution shift between train and test data, we only update the parameters of the feature extractor, " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_e" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", to" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16743" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 532, + 190 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 532, + 190 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 532, + 190 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 532, + 190 + ], + "type": "image", + "image_path": "c9a8509225d28991a43f53554069fc95fb969f3955e955998951bd882f811d93.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 193, + 547, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 547, + 228 + ], + "type": "text", + "content": "Figure 1. Block diagram of a general architecture for test time adaptation. At test time training, the normalization layers of the feature extractor are adapted by optimizing the combination of rank and GC loss. At inference time, we predict the quality scores of test images using the updated feature extractor and pre-trained quality regressor." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 256, + 286, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 286, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 286, + 280 + ], + "type": "text", + "content": "align the features between the train and test distributions in a lower dimensional space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 286, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 286, + 411 + ], + "type": "text", + "content": "The key challenge in TTA is the choice of a self-supervised auxiliary task that is highly correlated with the main task of IQA. However, relying too much on the auxiliary task can affect the performance of the main task. To prevent loss of learnt information induced by the auxiliary task [4], we first project the features to a lower dimensional space using a non-linear projection head " + }, + { + "bbox": [ + 46, + 280, + 286, + 411 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}" + }, + { + "bbox": [ + 46, + 280, + 286, + 411 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 46, + 280, + 286, + 411 + ], + "type": "inline_equation", + "content": "\\theta_s" + }, + { + "bbox": [ + 46, + 280, + 286, + 411 + ], + "type": "text", + "content": ". The auxiliary task drives the adaptation of the feature extractor through the projection head. Thus our model now has three parts parametrized by " + }, + { + "bbox": [ + 46, + 280, + 286, + 411 + ], + "type": "inline_equation", + "content": "(\\theta_e, \\theta_s, \\theta_c)" + }, + { + "bbox": [ + 46, + 280, + 286, + 411 + ], + "type": "text", + "content": " to resemble a Y-shape as shown in Figure 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 286, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 286, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 286, + 591 + ], + "type": "text", + "content": "When a batch of test instances " + }, + { + "bbox": [ + 46, + 411, + 286, + 591 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 411, + 286, + 591 + ], + "type": "text", + "content": " arrives, we extract features from the feature extractor, followed by the projection head, and update the set of parameters " + }, + { + "bbox": [ + 46, + 411, + 286, + 591 + ], + "type": "inline_equation", + "content": "(\\theta_e,\\theta_s)" + }, + { + "bbox": [ + 46, + 411, + 286, + 591 + ], + "type": "text", + "content": " by optimizing a self-supervised objective function " + }, + { + "bbox": [ + 46, + 411, + 286, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s(D)" + }, + { + "bbox": [ + 46, + 411, + 286, + 591 + ], + "type": "text", + "content": ". Updating all the model parameters of the feature extractor " + }, + { + "bbox": [ + 46, + 411, + 286, + 591 + ], + "type": "inline_equation", + "content": "\\theta_{e}" + }, + { + "bbox": [ + 46, + 411, + 286, + 591 + ], + "type": "text", + "content": " can cause the model to diverge too much from training, and the performance can drop drastically. Inspired by prior work on test time entropy minimization [33] and improving robustness for test data [28], we only update the linear and lower-dimensional feature modulation parameters. In a neural network, normalization layers satisfy these properties. So we only adapt the batch normalization layers by updating the affine parameters to mitigate distributional shifts. Thus we update these parameters by optimizing the auxiliary task loss given by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 600, + 286, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 600, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 117, + 600, + 286, + 622 + ], + "type": "interline_equation", + "content": "\\theta_ {e} ^ {*}, \\theta_ {s} ^ {*} = \\underset {\\left(\\theta_ {e}, \\theta_ {s}\\right)} {\\operatorname {a r g m i n}} \\mathcal {L} _ {s} (D) \\tag {1}", + "image_path": "acdb097b1f046adc8ff33c0c4c192c8ac6032b512532e4088ed72935054e4a5a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "After optimizing the above loss function, we use the updated feature extractor parameters " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\theta_e^*" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": " and pre-trained quality regressor parameters " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\theta_c" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": " to predict the quality for a batch of test data. Since the distribution across batches can vary significantly, we ignore the earlier updated model and start the TTA based on source model weights for new test instants. Thus the updated target model always only depends" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 257, + 411, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 257, + 411, + 268 + ], + "spans": [ + { + "bbox": [ + 306, + 257, + 411, + 268 + ], + "type": "text", + "content": "on the incoming test data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 276, + 477, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 276, + 477, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 276, + 477, + 289 + ], + "type": "text", + "content": "3.2. Self-Supervised Auxiliary Tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 294, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 474 + ], + "type": "text", + "content": "Our goal is to carefully choose self-supervised auxiliary tasks that capture quality-aware distributional information to adapt the feature encoder. We formulate two novel and complementary self-supervised learning techniques which help to learn the distribution shift between train and test data. These self-supervised objective functions are - 1) Group Contrastive Loss and 2) Rank Loss. While the GC loss works well when there is a reasonable separation of quality among the samples in a batch, the rank loss works better even when the quality of the batch samples is similar. On the other hand, the rank loss is meaningful only when the quality of the input image is not extremely low, while the GC loss is independent of the quality of a given image. Thus, a combination of the two losses renders our TTA extremely effective across various scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 490, + 441, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 441, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 441, + 502 + ], + "type": "text", + "content": "3.2.1 Group Contrastive Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "text", + "content": "While contrastive learning has been used for TTA of deep image classification, its direct application does not appear relevant to the task of IQA. Thus, we introduce group contrastive (GC) learning as an auxiliary task for TTA of IQA models. In particular, we make two groups of images from a single batch of " + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "text", + "content": " images based on the pseudo-labels given by the pre-trained source model. We sort the images in ascending order as " + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{(1)}, x_{(2)}, \\ldots, x_{(N)}" + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "text", + "content": " based on the pseudolabels, where " + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{(i)}, i = 1, 2, \\ldots, N" + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "text", + "content": ", corresponds to the " + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "text", + "content": " lowest quality image in the batch. We then segregate images with high quality scores (say, top " + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "text", + "content": " fraction of the data in a batch) and include them in a group of higher-quality images. Similarly, we separate out lower quality images (say, the lowest " + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "text", + "content": " fraction of the images in a batch) together to form another group. Here we assume that " + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "inline_equation", + "content": "pN" + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "inline_equation", + "content": "(1 - p)N" + }, + { + "bbox": [ + 304, + 509, + 545, + 713 + ], + "type": "text", + "content": " are integers for simplicity; else, they can be rounded off to the nearest integers." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16744" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 71, + 547, + 240 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 547, + 240 + ], + "type": "image", + "image_path": "f7d8fac2d877a1359ceeb87ce57c4b7638faa7a5b704c0d452944e44000910ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 278 + ], + "type": "text", + "content": "Figure 2. The overview framework of group contrastive loss for test time adaptation. For a batch of images, two different groups are formed based on pseudo-labels given by the source model. The group contrastive loss tries to minimize the distance between features extracted from images belonging to the same group while maximizing the distance between features from different groups." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 307, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 403 + ], + "type": "text", + "content": "The premise behind our loss for GC learning is that images from the same quality group should give similar feature representations in a lower dimensional space while features of images from different groups are separated out. Thus image pairs from the same group act as positive pairs, and image pairs from different groups act as negative pairs. By separating out these two groups, the model adapts itself by better separating the intermediate quality samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "text", + "content": "Let a positive pair " + }, + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "inline_equation", + "content": "x_{(i)}" + }, + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "inline_equation", + "content": "x_{(j)}" + }, + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "text", + "content": " come from the same group, i.e., either both " + }, + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "inline_equation", + "content": "i, j \\leq pN" + }, + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "inline_equation", + "content": "i, j > (1 - p)N" + }, + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "text", + "content": ". We use a modified NT-Xent contrastive loss [4] as our objective function. For a pair of images coming from the lower quality group where " + }, + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "inline_equation", + "content": "i, j \\leq pN" + }, + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 46, + 403, + 288, + 474 + ], + "type": "text", + "content": ", the GC loss is defined as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 480, + 287, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 513 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i, j} ^ {g c} = - \\log \\frac {\\exp (\\operatorname* {s i m} (z _ {(i)} , z _ {(j)}) / \\tau)}{\\sum_ {k > (1 - p) N} ^ {N} \\exp (\\operatorname* {s i m} (z _ {(i)} , z _ {(k)}) / \\tau)}, \\tag {2}", + "image_path": "936b529cd197e8da91d57487b2162f766aa918c77063a21ea42081c0a835668a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{(i)} = f_{\\theta_s}(f_{\\theta_e}(x_{(i)}))" + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "text", + "content": " represents the feature at the output of the projection head for sample " + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "inline_equation", + "content": "x_{(i)}" + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "text", + "content": ", sim refers to the cosine similarity between two features. and " + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{(k)}" + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "text", + "content": " is the feature extracted from an image from the higher quality group with " + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "inline_equation", + "content": "k > (1 - p)N" + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "text", + "content": ". Also, " + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "text", + "content": " represents temperature scaling parameter. While we define the loss above when " + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "inline_equation", + "content": "i,j\\leq pN" + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "text", + "content": ", we can define a similar loss when " + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "inline_equation", + "content": "i,j > (1 - p)N" + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "text", + "content": ". For all pairs of images within the same group, we obtain the GC loss and add them together to obtain the final loss, " + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{gc}" + }, + { + "bbox": [ + 46, + 521, + 289, + 641 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 650, + 287, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 650, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 56, + 650, + 287, + 693 + ], + "type": "interline_equation", + "content": "\\mathcal{L}^{gc} = \\sum_{i = 1}^{pN}\\sum_{\\substack{j = 1\\\\ j\\neq i}}^{pN}\\mathcal{L}_{i,j}^{gc} + \\sum_{i = (1 - p)N + 1}^{N}\\sum_{\\substack{j = (1 - p)N + 1\\\\ j\\neq i}}^{N}\\mathcal{L}_{i,j}^{gc}. \\tag{3}", + "image_path": "d4d44c1ff540a59da144c70a8ba6d0d16fc1a048f127b122db0e1815d8a8b63f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 264, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 264, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 264, + 714 + ], + "type": "text", + "content": "A block diagram of the GC loss is shown in Figure 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 307, + 384, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 307, + 384, + 318 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 384, + 318 + ], + "type": "text", + "content": "3.2.2 Rank Loss" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 338, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 541 + ], + "type": "text", + "content": "Given a sample of images from test data, the rank loss helps adapt the features by capturing quality-aware distributional information at the sample level. We introduce a ranking objective to learn the quality orders between two distorted versions of the test images that are quality discriminable. We distort each test image " + }, + { + "bbox": [ + 304, + 338, + 545, + 541 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 338, + 545, + 541 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 338, + 545, + 541 + ], + "type": "inline_equation", + "content": "i = 1,2,\\dots ,N" + }, + { + "bbox": [ + 304, + 338, + 545, + 541 + ], + "type": "text", + "content": ", in the minibatch of size " + }, + { + "bbox": [ + 304, + 338, + 545, + 541 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 338, + 545, + 541 + ], + "type": "text", + "content": " with two different degrees of degradation of a given distortion type. Let " + }, + { + "bbox": [ + 304, + 338, + 545, + 541 + ], + "type": "inline_equation", + "content": "x_{i}^{high}" + }, + { + "bbox": [ + 304, + 338, + 545, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 338, + 545, + 541 + ], + "type": "inline_equation", + "content": "x_{i}^{low}" + }, + { + "bbox": [ + 304, + 338, + 545, + 541 + ], + "type": "text", + "content": " denote the higher and lower distorted image, respectively. The degradation types include synthetic distortions such as blur, compression, and noise. We choose degradation levels randomly from two sets of parameters sufficiently farther apart. Ideally, the distance between the features extracted from the test image and the highly distorted image should be greater than that between the features extracted from the test image and the lower distorted image. Our rank loss tries to capture this order and adapt the model to the test data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "content": "We project the triplet of images " + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "inline_equation", + "content": "(x_{i}, x_{i}^{high}, x_{i}^{low})" + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "content": " in the feature space through the feature extractor and the projection head. We measure the distance between the distortion-augmented image and the original test data using the Euclidean distance in the projected feature space. Mathematically, let " + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i} = f_{\\theta_{s}}(f_{\\theta_{e}}(x_{i}))" + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "content": " be the feature extracted for test sample " + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "content": ". Similarly we define " + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i}^{high}" + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i}^{low}" + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "inline_equation", + "content": "d_{i}^{high}" + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "content": " denote the distance between " + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i}" + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i}^{high}" + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "content": ". Similarly, " + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "inline_equation", + "content": "d_{i}^{low}" + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "content": " denotes the distance between " + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i}" + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i}^{low}" + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "content": ". The target model is now fine-tuned to obtain the correct ranking between the distances by achieving " + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "inline_equation", + "content": "d_{i}^{high} \\geq d_{i}^{low}" + }, + { + "bbox": [ + 304, + 547, + 546, + 685 + ], + "type": "text", + "content": " shown in Figure 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "The probability of achieving this order is estimated by passing the difference in distances through a sigmoid func" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16745" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 68, + 515, + 286 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 515, + 286 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 515, + 286 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 515, + 286 + ], + "type": "image", + "image_path": "cfff0b4fd362c208c3eae5f7b14560808024eac16a0865be0cd580daca1a0a59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 289, + 547, + 324 + ], + "lines": [ + { + "bbox": [ + 46, + 289, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 547, + 324 + ], + "type": "text", + "content": "Figure 3. The rank loss framework for test time adaptation. We distort the test image with two different degrees of degradation. We then project the triplet of images in the features space through the shared feature extractor and projection head. Distances calculated between the features extracted from each distorted image and the original test image are used to compare their rank order." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 345, + 78, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 345, + 78, + 356 + ], + "spans": [ + { + "bbox": [ + 47, + 345, + 78, + 356 + ], + "type": "text", + "content": "tion as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 363, + 287, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 363, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 363, + 287, + 415 + ], + "type": "interline_equation", + "content": "P _ {i} = \\Pr \\left(d _ {i} ^ {\\text {h i g h}} \\geq d _ {i} ^ {\\text {l o w}}\\right) = \\frac {\\exp \\left(d _ {i} ^ {\\text {h i g h}} - d _ {i} ^ {\\text {l o w}}\\right)}{1 + \\exp \\left(d _ {i} ^ {\\text {h i g h}} - d _ {i} ^ {\\text {l o w}}\\right)}. \\tag {4}", + "image_path": "b3beccf2b2756e19cddbd2ff13cdc453f10158769d5e1e04ca521c4f51bfd8c0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "text", + "content": "We use binary cross-entropy loss between true label " + }, + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\bar{P}_i = 1" + }, + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "text", + "content": " and the predicted probability " + }, + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 47, + 415, + 287, + 440 + ], + "type": "text", + "content": " to obtain the rank loss as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 447, + 287, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 447, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 447, + 287, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} ^ {r} (\\bar {P} _ {i}, P _ {i}) = - \\bar {P} _ {i} \\log P _ {i} - (1 - \\bar {P} _ {i}) \\log (1 - P _ {i}). (5)", + "image_path": "a2995690f60c5fcbc160cd406aacdc657214c1868eb4ef50eb428d337a13b259.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 471, + 266, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 471, + 266, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 266, + 483 + ], + "type": "text", + "content": "For a batch of size " + }, + { + "bbox": [ + 47, + 471, + 266, + 483 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 471, + 266, + 483 + ], + "type": "text", + "content": ", the overall rank loss is given as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 491, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 491, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 138, + 491, + 287, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {r} = \\sum_ {i = 1} ^ {N} \\mathcal {L} _ {i} ^ {r}. \\tag {6}", + "image_path": "d1d404602f6c92c55e9801c5133fb93eba71334e99b13dabc7d2d839ca4df370.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "content": "One of the challenges in the rank loss is selecting the distortion type for every image. The original test image may already be distorted, which poses difficulties in deciding the distortion type that can help TTA. For example, if the test image is extremely blurred and we blur it again with two different levels, both images may be visually indistinguishable, thereby limiting the extent to which adaptation can help. We exploit the source model's knowledge to overcome this limitation in choosing the distortion type. We predict the quality scores of the pair " + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "inline_equation", + "content": "(x_{i}^{high}, x_{i}^{low})" + }, + { + "bbox": [ + 46, + 533, + 288, + 701 + ], + "type": "text", + "content": " for each distortion type using the source model. We hypothesize that the pair with the maximal difference in predicted quality is sufficiently different in visual quality. Thus choosing such a pair in Equation (5) can help adapt the model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "content": "Our overall self-supervised TTA objective function is a" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 345, + 512, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 345, + 512, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 512, + 357 + ], + "type": "text", + "content": "combination of both the rank and GC loss given as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 389, + 361, + 545, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 361, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 389, + 361, + 545, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\mathcal {L} ^ {g c} + \\lambda \\mathcal {L} ^ {r}, \\tag {7}", + "image_path": "69c079383ae76a64643e21174e104bdfbd256b12963b474c13af1e97e7721817.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 380, + 537, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 537, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 537, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 380, + 537, + 392 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 380, + 537, + 392 + ], + "type": "text", + "content": " is a hyper-parameter used to combine the losses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 400, + 388, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 400, + 388, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 400, + 388, + 414 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 420, + 505, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 505, + 433 + ], + "type": "text", + "content": "4.1. Quality Models, Datasets and Metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "text", + "content": "We evaluate TTA on four popular IQA databases using four different quality-aware models. In particular, we consider state-of-the-art deep IQA models such as TReS [7], MUSIQ [12], HyperIQA [30], and MetaIQA [40]. These models contain a ResNet backbone with batch normalization layers, which we model as the feature extractor " + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "inline_equation", + "content": "f_{\\theta_e}" + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "text", + "content": " and only update its batch normalization parameters. We model the rest of the network as the quality regressor part, " + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "inline_equation", + "content": "f_{\\theta_c}" + }, + { + "bbox": [ + 304, + 438, + 545, + 605 + ], + "type": "text", + "content": ". TReS and MUSIQ use transformers as a part of their architecture, which we include as a part of the quality regressor in all our main experiments. We also explore the adaptation of transformers by including them as part of the feature extractor and updating their layer normalization statistics in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "We project the features extracted from the last layer of ResNet through a 256-dimensional fully connected (FC) layer corresponding to the self-supervised projection head " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": ". These lower dimensional features are used to adapt the model at test time. Three IQA models, TReS, MUSIQ, and HyperIQA are trained on the LIVEFB database [37] containing 39,810 images. MetaIQA is trained on two synthetically distorted databases, TID2013 [26] and KADID-10k [17]." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16746" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 70, + 518, + 242 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 518, + 242 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 518, + 242 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 518, + 242 + ], + "type": "table", + "html": "
BackboneDatabaseKonIQ-10kPIPALCID2013LIVE-IQA
MethodSROCCPLCCSROCCPLCCSROCCPLCCSROCCPLCC
TReSBaseline0.65200.69550.38450.40780.52720.64630.54350.4450
Rotation0.65060.68050.40610.41140.57060.66510.58660.5311
TTA-IQA0.65780.70740.42780.42040.60320.67100.67220.5963
MUSIQBaseline0.63040.68020.31900.34140.51730.60320.25960.3351
Rotation0.65770.71540.36650.36930.54870.61640.35120.3976
TTA-IQA0.66930.72300.37430.37310.54990.62200.36490.4031
HyperIQABaseline0.58610.63130.30370.33040.48950.61230.51430.4377
Rotation0.60330.65360.32680.34820.49020.61500.62680.5469
TTA-IQA0.59600.64950.36530.37670.50390.59880.62180.5438
MetaIQABaseline0.51620.44600.32870.29550.72130.68170.73230.6732
Rotation0.58230.53110.33530.30420.71770.67450.72710.6868
TTA-IQA0.58380.54280.40730.35100.78090.73990.79990.7726
", + "image_path": "2d5e46c9492bc68414c538c8d0fce700fa3a6ecab7594bbb4c9d21a5e9db9bf0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 547, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 268 + ], + "type": "text", + "content": "Table 1. Comparison of TTA-IQA with popular NR IQA methods and one popular auxiliary task - rotation prediction on authentically and synthetically distorted datasets. Bold entries imply the best performance for every individual quality-aware model on respective datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 286, + 288, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 288, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 288, + 321 + ], + "type": "text", + "content": "We choose challenging databases to evaluate the generalization capability of our TTA-IQA method. The test datasets are described as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 323, + 288, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 357 + ], + "type": "text", + "content": "KonIQ-10k [9] is a popular in-the-wild authentically distorted database consisting of 10,073 quality-scored images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 358, + 287, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 395 + ], + "type": "text", + "content": "PIPAL [11] is a large-scale IQA database for evaluating perceptual image restoration. It contains 29k distorted images, including 19 different GAN-based algorithms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 396, + 287, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 419 + ], + "type": "text", + "content": "CID2013 [32] consists of 480 images in six image sets captured by 79 imaging devices." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 420, + 287, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 444 + ], + "type": "text", + "content": "LIVE-IQA [29] consists of 779 synthetically distorted images from 29 reference images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 445, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 480 + ], + "type": "text", + "content": "We evaluate our results using Spearman's rank-order correlation coefficient (SROCC) and Pearson's linear correlation coefficient (PLCC)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 491, + 180, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 180, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 180, + 503 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 509, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 629 + ], + "type": "text", + "content": "We implement our setup in PyTorch and conduct all the experiments with an 16 GB NVIDIA RTX A4000 GPU. During TTA, we randomly select a patch of size " + }, + { + "bbox": [ + 46, + 509, + 287, + 629 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 509, + 287, + 629 + ], + "type": "text", + "content": " from the input image and apply quality preserving augmentations such as horizontal flip and vertical flip before passing it through the network for TTA. We use the ADAM [13] optimizer with a learning rate of 0.001 and set the number of iteration as 3. All the experiments were run on five different seeds using a batch size of 8, and the final results were obtained by averaging." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "With regard to the self-supervised auxiliary tasks, we use a combination of rank loss and GC loss as our objective function using " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " given in Equation (7). We choose " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "p = 0.25" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " to determine the groups. We calculate the GC loss using " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\tau = 1" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": ". For the rank loss, the test image is synthetically blurred using a Gaussian blur filter of size " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": " with two sets of standard deviations " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": ", for the Gaussian" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 285, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 370 + ], + "type": "text", + "content": "blur kernel. We keep " + }, + { + "bbox": [ + 304, + 285, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\sigma \\in [40,80]" + }, + { + "bbox": [ + 304, + 285, + 545, + 370 + ], + "type": "text", + "content": " for highly distorted images and " + }, + { + "bbox": [ + 304, + 285, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\sigma \\in [1,20]" + }, + { + "bbox": [ + 304, + 285, + 545, + 370 + ], + "type": "text", + "content": " for less blurred images. For compression distortions, we specify the quality factor in [80, 95] for lower compression and [30, 60] for higher compression rates. Similarly, we add zero mean Gaussian white noise to the test image with variance in [0.05, 0.1] for higher distortion and in [0.005, 0.01] for lower distortion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 380, + 443, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 443, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 443, + 392 + ], + "type": "text", + "content": "4.3. Performance Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": "Table 1 shows the performance of our method on all four test datasets using all four quality-aware models. In addition to the source model, we also compare with rotation prediction [31] as the auxiliary task during TTA. A comparison with such a task helps understand the role of quality-aware losses for the TTA of IQA models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 473, + 546, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 546, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 546, + 675 + ], + "type": "text", + "content": "We observe that TTA-IQA using the combination of rank and GC loss outperforms the source models in all the cases. Note that the PIPAL dataset has a huge distribution shift from the authentically distorted LIVEFB dataset on which three of the models were trained. While most source models perform very poorly on PIPAL, we achieve " + }, + { + "bbox": [ + 304, + 473, + 546, + 675 + ], + "type": "inline_equation", + "content": "10\\% -20\\%" + }, + { + "bbox": [ + 304, + 473, + 546, + 675 + ], + "type": "text", + "content": " improvement over the source models. On the KonIQ-10k dataset, TTA-IQA gives around " + }, + { + "bbox": [ + 304, + 473, + 546, + 675 + ], + "type": "inline_equation", + "content": "1\\% -10\\%" + }, + { + "bbox": [ + 304, + 473, + 546, + 675 + ], + "type": "text", + "content": " improvement over the source model. As both KonIQ-10k and LIVEFB are authentically distorted datasets, the distribution shift is reasonably small, leading to smaller improvements using adaptation. CID2013 is also an authentically distorted dataset and gives similar improvements of around " + }, + { + "bbox": [ + 304, + 473, + 546, + 675 + ], + "type": "inline_equation", + "content": "2\\% -13\\%" + }, + { + "bbox": [ + 304, + 473, + 546, + 675 + ], + "type": "text", + "content": " over the source models. Our experiments on the LIVE-IQA database provide a significantly greater improvement of " + }, + { + "bbox": [ + 304, + 473, + 546, + 675 + ], + "type": "inline_equation", + "content": "10\\% -33\\%" + }, + { + "bbox": [ + 304, + 473, + 546, + 675 + ], + "type": "text", + "content": " owing to the shift from authentic to synthetic distortions for three models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": "We also observe from Table 1 that our approach outperforms the rotation prediction task in most cases. Other TTA ideas, such as TENT [33] and training using masked autoen" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16747" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 512, + 242 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 512, + 242 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 512, + 242 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 512, + 242 + ], + "type": "table", + "html": "
BackboneDatabaseKonIQ-10kPIPALCID2013LIVE-IQA
RankGCSRCCPLCCSRCCPLCCSRCCPLCCSRCCPLCC
TReS✓×0.65620.69890.41710.42040.60160.67360.67050.5908
×✓0.65160.69460.46660.51830.53660.64930.71600.6290
✓✓0.65780.70740.42780.42040.60320.67100.67220.5963
MUSIQ✓×0.65490.71490.37680.37180.52160.60340.36340.4024
×✓0.66110.71760.35850.36420.54460.61590.32990.3914
✓✓0.66930.72300.37430.37310.54990.62200.36490.4031
HyperIQA✓×0.59280.64550.36160.37320.50390.59910.55050.5050
×✓0.60940.65670.33330.35520.51200.62550.63310.5592
✓✓0.59600.64950.36530.37670.50390.59880.62180.5438
MetaIQA✓×0.55800.51180.39920.34370.75790.72820.78940.7534
×✓0.54140.46360.37100.32870.78610.70670.75660.6927
✓✓0.58380.54280.40730.35100.78090.73990.79990.7726
", + "image_path": "4113a24e2f0307a7b176fb35d62c0450d445aa0a72eaa6fbdd66d9d3584444ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 547, + 268 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 268 + ], + "type": "text", + "content": "Table 2. Ablation study results on authentically and synthetically distorted datasets using popular quality aware source models. Bold entries imply the best performance among all three settings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 47, + 282, + 196, + 396 + ], + "blocks": [ + { + "bbox": [ + 47, + 282, + 196, + 396 + ], + "lines": [ + { + "bbox": [ + 47, + 282, + 196, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 196, + 396 + ], + "type": "image", + "image_path": "b2a67dd12412c5190c97d77c2b4e37b7602a5953485e000dae19d17651968b99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 396, + 546, + 418 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 546, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 546, + 418 + ], + "type": "text", + "content": "Figure 4. Scatter plot for predicting quality of similar quality images from LIVE-IQA dataset for the source model and TTA using only rank loss or GC loss. We mark two sets of images having similar human opinion score (DMOS) by ellipses in each case." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 214, + 282, + 364, + 396 + ], + "blocks": [ + { + "bbox": [ + 214, + 282, + 364, + 396 + ], + "lines": [ + { + "bbox": [ + 214, + 282, + 364, + 396 + ], + "spans": [ + { + "bbox": [ + 214, + 282, + 364, + 396 + ], + "type": "image", + "image_path": "b7bd95081d74e2ce9ecfbb5a0049f21149d987219de81d803708f0ced53b53fa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 388, + 282, + 531, + 396 + ], + "blocks": [ + { + "bbox": [ + 388, + 282, + 531, + 396 + ], + "lines": [ + { + "bbox": [ + 388, + 282, + 531, + 396 + ], + "spans": [ + { + "bbox": [ + 388, + 282, + 531, + 396 + ], + "type": "image", + "image_path": "87fd3f40ea4958c2c0f6c22ded6f839fe1920dd2d1bd37a605b6ef5cd8f59904.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 485 + ], + "type": "text", + "content": "coders [6] are not applicable for the IQA task. In particular, the notion of entropy in regression tasks is not clear. On the other hand, masked reconstructions tend to lead to quality degradation. Thus, we do not compare with them." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 491, + 246, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 246, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 246, + 504 + ], + "type": "text", + "content": "4.4. Ablation Study & Other Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 641 + ], + "type": "text", + "content": "Need for Rank Loss and GC Loss for TTA. We perform an ablation study with respect to the two losses in Table 2. We see from the results that the rank loss and the GC loss individually always improve on the source models. Further, the combination of the rank and GC loss provides an even better performance over the individual losses in most evaluation scenarios. Even in scenarios, where the combination loss achieves the second-best performance, its performance is very close to the best. For the rest of the experiments in this section, we present results on all four datasets using the TReS method alone." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "We now discuss scenarios where the GC loss is particularly more useful than the rank loss. If the test image is extremely distorted, none of the three kinds of distortion types can create much difference in the perceptual quality of the two degraded versions. To understand this further, we consider highly distorted images (corresponding to a dif" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 437, + 546, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 508 + ], + "type": "text", + "content": "ferential mean opinion score (DMOS) greater than 70) of different distortion types in the LIVE-IQA dataset. We apply TTA only for these images using different losses. The SROCC performance for these images in Figure 5 reveal that the rank loss is not very effective. However, the GC loss works well in such scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 676 + ], + "type": "text", + "content": "Conversely, we also discuss scenarios where the rank loss is more useful than the GC loss. To illustrate this idea, we select multiple images having similar quality with DMOS in [28, 44] from the LIVE-IQA dataset. We adapt the source model for this set of images using both the GC and rank losses. From Figure 4, we observe that the rank loss leads to much better adaptation in terms of the resulting model correlating with the ground truth scores. Intuitively, when the input images have a similar quality, the pseudolabels given by the source model may be noisy, leading to an inaccurate grouping of the images into the lower and higher quality groups. This can adversely affect the GC loss. So the rank loss is more effective than the GC loss for test batches with similar quality images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Effect of Selecting Multiple Distortion Types in the Rank Loss. We experiment with the impact of different kinds of distortion types while incorporating the rank loss." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16748" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 92, + 282, + 203 + ], + "blocks": [ + { + "bbox": [ + 83, + 77, + 252, + 91 + ], + "lines": [ + { + "bbox": [ + 83, + 77, + 252, + 91 + ], + "spans": [ + { + "bbox": [ + 83, + 77, + 252, + 91 + ], + "type": "text", + "content": "Results for Highly Distorted Images" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 92, + 282, + 203 + ], + "lines": [ + { + "bbox": [ + 57, + 92, + 282, + 203 + ], + "spans": [ + { + "bbox": [ + 57, + 92, + 282, + 203 + ], + "type": "image", + "image_path": "ca0e444cd1e1d3fbe99e30372c6eb592b42876d877fcb1c4894a6afaa4f4b806.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 287, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 232 + ], + "type": "text", + "content": "Figure 5. TTA-IQA on different distortion-specific images from LIVE-IQA dataset with respect to individual losses" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 383 + ], + "type": "text", + "content": "In particular, we consider three scenarios: one where only one of the distortion types is used in the rank loss, another where all the three types are used to obtain three loss terms which are summed together to update the model, and a third, where only one of the distortion types is chosen based on the discussion in Section 3.2.2. We observe from Table 3 that choosing the distortion type with the maximal difference in quality gives the best performance for the rank loss. This experiment validates our hypothesis that only a rank loss that can discriminate between the degraded image pairs is useful for TTA." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 392, + 284, + 456 + ], + "blocks": [ + { + "bbox": [ + 48, + 392, + 284, + 456 + ], + "lines": [ + { + "bbox": [ + 48, + 392, + 284, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 284, + 456 + ], + "type": "table", + "html": "
DatabaseBlurCompNoiseAllBest
KonIQ-10k0.6560.6560.6670.6150.656
PIPAL0.3840.4150.3890.4150.417
CID20130.5270.6090.6060.5600.609
LIVE-IQA0.5780.6050.6180.6340.671
", + "image_path": "6461869ff6e1f20e98da5f14d21fe74c81db965b37da6dadbe7aeab558cec652.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "content": "Selection of Group Size by Varying " + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "content": ". We explore different values of " + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "content": " for constructing the GC loss. For a batch size of 8, possible choices of the group size are 2, 3, 4 corresponding to " + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "inline_equation", + "content": "p = 0.25, 0.375, 0.5" + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "content": " respectively. From Table 4, we observe that the performances are roughly similar for different values of " + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "content": " with a slightly superior performance for " + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "inline_equation", + "content": "p = 0.25" + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "content": ". We note that as " + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 483, + 287, + 590 + ], + "type": "text", + "content": " increases, the groups are larger and probably less discriminative in terms of quality, leading to a slightly poorer performance." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 74, + 601, + 259, + 664 + ], + "blocks": [ + { + "bbox": [ + 47, + 456, + 285, + 468 + ], + "lines": [ + { + "bbox": [ + 47, + 456, + 285, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 285, + 468 + ], + "type": "text", + "content": "Table 3. Impact of selecting multiple distortion types on rank loss" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 601, + 259, + 664 + ], + "lines": [ + { + "bbox": [ + 74, + 601, + 259, + 664 + ], + "spans": [ + { + "bbox": [ + 74, + 601, + 259, + 664 + ], + "type": "table", + "html": "
Databasep=0.25p=0.375p=0.5
KonIQ-10k0.65160.64970.6467
PIPAL0.46660.46020.4616
CID20130.53660.54860.5536
LIVE-IQA0.71600.71710.6937
", + "image_path": "89dcf9545d45cb7c08e04645c4673f040deee72e8357c5d87073cbbe69250f63.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Selection of Number of Groups for GC Loss. In our formulation, we define the GC loss only between two con" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "trastive groups. In principle, we can extend our idea of GC loss to multiple groups. In particular, we can cluster the images in a batch into multiple groups and apply the GC loss between every pair of groups and sum the loss terms. Thus, images coming from the same group act as positive pairs, and images from two different groups are considered as negative pairs. From Table 5, we observe that the number of groups " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " does not impact the performance much." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 335, + 178, + 515, + 242 + ], + "blocks": [ + { + "bbox": [ + 62, + 666, + 271, + 677 + ], + "lines": [ + { + "bbox": [ + 62, + 666, + 271, + 677 + ], + "spans": [ + { + "bbox": [ + 62, + 666, + 271, + 677 + ], + "type": "text", + "content": "Table 4. Impact of varying group size using " + }, + { + "bbox": [ + 62, + 666, + 271, + 677 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 62, + 666, + 271, + 677 + ], + "type": "text", + "content": " for GC loss" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 178, + 515, + 242 + ], + "lines": [ + { + "bbox": [ + 335, + 178, + 515, + 242 + ], + "spans": [ + { + "bbox": [ + 335, + 178, + 515, + 242 + ], + "type": "table", + "html": "
DatabaseG=2G=3G=4
KonIQ-10k0.65160.65240.6509
PIPAL0.46660.46680.4663
CID20130.53660.53270.5388
LIVE-IQA0.71600.70110.6883
", + "image_path": "d9b1104a07845e4bcea16e540c9dfd6c28d97d9b8eaf3c43d77b8c7838e48b9f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 243, + 523, + 255 + ], + "lines": [ + { + "bbox": [ + 306, + 243, + 523, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 523, + 255 + ], + "type": "text", + "content": "Table 5. Impact of varying number of groups " + }, + { + "bbox": [ + 306, + 243, + 523, + 255 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 306, + 243, + 523, + 255 + ], + "type": "text", + "content": " for GC loss" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 267, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 340 + ], + "type": "text", + "content": "Choice of Number of Iterations for Learning Auxiliary Task. We also examine the effect of the number of iterations of parameter updates during TTA. Figure 6 shows that the performance on CID2013 dataset is fairly robust until 4 iterations. Beyond that, the model overfits the auxiliary task and leads to poor performance at test time." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 326, + 354, + 529, + 473 + ], + "blocks": [ + { + "bbox": [ + 326, + 354, + 529, + 473 + ], + "lines": [ + { + "bbox": [ + 326, + 354, + 529, + 473 + ], + "spans": [ + { + "bbox": [ + 326, + 354, + 529, + 473 + ], + "type": "image", + "image_path": "e8585276d93fbbe16b41ecc7e3ba270b2b147debf9baea6eb4de19e5d629d7ed.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 483, + 517, + 495 + ], + "lines": [ + { + "bbox": [ + 334, + 483, + 517, + 495 + ], + "spans": [ + { + "bbox": [ + 334, + 483, + 517, + 495 + ], + "type": "text", + "content": "Figure 6. Effect of increasing number of iterations" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 513, + 378, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 513, + 378, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 513, + 378, + 525 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 533, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 689 + ], + "type": "text", + "content": "Test time adaptation has become very popular due to its simplicity and the lack of need for end-to-end training. Our work is perhaps one of the first attempts to design the method of TTA in the context of blind IQA. While most IQA methods focus on making the models robust enough to perform well for cross-database experiments, our TTA-IQA method can outperform existing state-of-the-art methods because of the adaptation at test time. We formulate novel self-supervised auxiliary tasks using the rank and group contrastive losses, which can learn quality-aware information from the test data. While primarily explored TTA for IQA, it would be interesting to understand the role of TTA for video quality assessment as well." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement: This work was supported in part by a grant from Kotak IISc AI-ML Centre and Department of" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16749" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "Science and Technology, Government of India, under grant CRG/2020/003516." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 106, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 106, + 118 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 106, + 118 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 125, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 125, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 192 + ], + "type": "text", + "content": "[1] Piotr Bojanowski and Armand Joulin. Unsupervised learning by predicting noise. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 517-526. PMLR, 06-11 Aug 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 192, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 288, + 247 + ], + "type": "text", + "content": "[2] Sebastian Bosse, Dominique Maniry, Klaus-Robert Müller, Thomas Wiegand, and Wojciech Samek. Deep neural networks for no-reference and full-reference image quality assessment. IEEE Transactions on Image Processing, 27(1):206-219, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 313 + ], + "type": "text", + "content": "[3] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 139–156, Cham, 2018. Springer International Publishing. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 315, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 288, + 369 + ], + "type": "text", + "content": "[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org, 2020. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 370, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 370, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 370, + 288, + 414 + ], + "type": "text", + "content": "[5] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), December 2015. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 415, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 288, + 446 + ], + "type": "text", + "content": "[6] Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A. Efros. Test-time training with masked autoencoders. arXiv preprint arXiv:2209.07522, 2022. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 448, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 448, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 448, + 288, + 502 + ], + "type": "text", + "content": "[7] S Alireza Golestaneh, Saba Dadsetan, and Kris M Kitani. No-reference image quality assessment via transformers, relative ranking, and self-consistency. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3209-3218, 2022. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 503, + 288, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 503, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 53, + 503, + 288, + 547 + ], + "type": "text", + "content": "[8] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 548, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 288, + 592 + ], + "type": "text", + "content": "[9] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. Trans. Img. Proc., 29:4041-4056, jan 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 657 + ], + "type": "text", + "content": "[10] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 2427–2440. Curran Associates, Inc., 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[11] Gu Jinjin, Cai Haoming, Chen Haoyu, Ye Xiaoxing, Jimmy S. Ren, and Dong Chao. Pipal: A large-scale image quality assessment dataset for perceptual image restoration. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "pages 633-651, Cham, 2020. Springer International Publishing. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "text", + "content": "[12] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5148-5157, October 2021. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 150, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 183 + ], + "type": "text", + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "text", + "content": "[14] Nikos Komodakis and Spyros Gidaris. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations (ICLR), Vancouver, Canada, Apr. 2018. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "text", + "content": "[15] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Colorization as a proxy task for visual understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 283, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 358 + ], + "type": "text", + "content": "[16] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? Source hypothesis transfer for unsupervised domain adaptation. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 6028-6039. PMLR, 13-18 Jul 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 360, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 404 + ], + "type": "text", + "content": "[17] Hanhe Lin, Vlad Hosu, and Dietmar Saupe. Kadid-10k: A large-scale artificially distorted iqa database. In 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), pages 1–3, 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "text", + "content": "[18] Xialei Liu, Joost van de Weijer, and Andrew D. Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Oct 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 503 + ], + "type": "text", + "content": "[19] Yutao Liu, Ke Gu, Yongbing Zhang, Xiu Li, Guangtao Zhai, Debin Zhao, and Wen Gao. Unsupervised blind image quality evaluation via statistical measurements of structure, naturalness, and perception. IEEE Transactions on Circuits and Systems for Video Technology, 30(4):929-943, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 503, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 503, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 503, + 545, + 580 + ], + "type": "text", + "content": "[20] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 21808-21820. Curran Associates, Inc., 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 581, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 624 + ], + "type": "text", + "content": "[21] Kede Ma, Wentao Liu, Kai Zhang, Zhengfang Duanmu, Zhou Wang, and Wangmeng Zuo. End-to-end blind image quality assessment using deep neural networks. IEEE Transactions on Image Processing, 27(3):1202-1213, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "text", + "content": "[22] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21(12):4695-4708, 2012. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[23] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE Transactions on Image Processing, 20(12):3350-3364, 2011. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16750" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[24] Chaithanya Kumar Mummadi, Robin Hutmacher, Kilian Rambach, Evgeny Levinkov, Thomas Brox, and Jan Hendrik Metzen. Test-time adaptation to distribution shift by confidence maximization and input transformation, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "text", + "content": "[25] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "text", + "content": "[26] Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola, Benoit Vozel, Kacem Chehdi, Marco Carli, Federica Battisti, and C.-C. Jay Kuo. Image database tid2013: Peculiarities, results and perspectives. Signal Processing: Image Communication, 30:57-77, 2015. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "text", + "content": "[27] Michele A. Saad, Alan C. Bovik, and Christophe Charrier. Blind image quality assessment: A natural scene statistics approach in the dct domain. IEEE Transactions on Image Processing, 21(8):3339-3352, 2012. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 363 + ], + "type": "text", + "content": "[28] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 11539-11551. Curran Associates, Inc., 2020. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "type": "text", + "content": "[29] H.R. Sheikh, M.F. Sabir, and A.C. Bovik. A statistical evaluation of recent full reference image quality assessment algorithms. IEEE Transactions on Image Processing, 15(11):3440-3451, 2006. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 464 + ], + "type": "text", + "content": "[30] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 466, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 543 + ], + "type": "text", + "content": "[31] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 9229-9248. PMLR, 13-18 Jul 2020. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 544, + 287, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 598 + ], + "type": "text", + "content": "[32] Toni Virtanen, Mikko Nuutinen, Mikko Vaahteranoksa, Pirkko Oittinen, and Jukka Häkkinen. Cid2013: A database for evaluating no-reference image quality assessment algorithms. IEEE Transactions on Image Processing, 24(1):390-402, 2015. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 644 + ], + "type": "text", + "content": "[33] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[34] Jingtao Xu, Peng Ye, Qiaohong Li, Haiqing Du, Yong Liu, and David Doermann. Blind image quality assessment based on high order statistics aggregation. IEEE Transactions on Image Processing, 25(9):4444-4457, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[35] Peng Ye, Jayant Kumar, Le Kang, and David Doermann. Unsupervised feature learning framework for no-reference" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 373 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "image quality assessment. In 2012 IEEE conference on computer vision and pattern recognition, pages 1098-1105. IEEE, 2012. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[36] Peng Ye, Jayant Kumar, Le Kang, and David Doermann. Unsupervised feature learning framework for no-reference image quality assessment. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1098-1105, 2012. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 163, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 228 + ], + "type": "text", + "content": "[37] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 229, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 545, + 262 + ], + "type": "text", + "content": "[38] Guangtao Zhai, Xiongkuo Min, and Ning Liu. Free-energy principle inspired visual quality assessment: An overview. Digital Signal Processing, 91:11-20, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 263, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 263, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 263, + 545, + 316 + ], + "type": "text", + "content": "[39] Weixia Zhang, Kede Ma, Jia Yan, Dexiang Deng, and Zhou Wang. Blind image quality assessment using a deep bilinear convolutional neural network. IEEE Transactions on Circuits and Systems for Video Technology, 30(1):36-47, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 319, + 546, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 546, + 373 + ], + "type": "text", + "content": "[40] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16751" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Test-time Personalizable Forecasting of 3D Human Poses/205eafc7-d672-437b-9b70-81a2995fa72d_content_list.json b/2023/Test-time Personalizable Forecasting of 3D Human Poses/205eafc7-d672-437b-9b70-81a2995fa72d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..67874655fdf1e342ebb602352b8b19aa8700c037 --- /dev/null +++ b/2023/Test-time Personalizable Forecasting of 3D Human Poses/205eafc7-d672-437b-9b70-81a2995fa72d_content_list.json @@ -0,0 +1,1912 @@ +[ + { + "type": "text", + "text": "Test-time Personalizable Forecasting of 3D Human Poses", + "text_level": 1, + "bbox": [ + 197, + 130, + 772, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qiongjie Cui $^{1}$ Huaijiang Sun $^{1*}$ \nJianfeng Lu $^{1}$ Weiqing Li $^{1}$ Bin Li $^{2}$ Hongwei Yi $^{3}$ , Haofan Wang $^{4}$ $^{1}$ Nanjing University of Science and Technology, Nanjing, China \n $^{2}$ Tianjin AiForward Science and Technology Co., Ltd., China \n $^{3}$ Max Planck Institute for Intelligent Systems, Tubingen, Germany ${ }^{4}$ Xiaohongshu Inc \ncuiqiongjie@njust.edu.cn sunhuaijiang@njust.edu.cn", + "bbox": [ + 140, + 179, + 831, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current motion forecasting approaches typically train a deep end-to-end model from the source domain data, and then apply it directly to target subjects. Despite promising results, they remain non-optimal, due to privacy considerations, the test person and his/her natural properties (e.g., behavioral trait) are typically unseen in training. In this case, the source pre-trained model has a low ability to adapt to these out-of-source characteristics, resulting in an unreliable prediction. To tackle this issue, we propose a novel helper-predictor test-time personalization approach (H/P-TTP), which allows for a generalizable representation of out-of-source subjects to gain more realistic predictions. Concretely, the helper is preceded by explicit and implicit augmenters, where the former yields noisy sequences to improve robustness, while the latter is to generate novel-domain data with an adversarial learning paradigm. Then, the domain-generalizable learning is achieved where the helper can extract cross-subject invariant-knowledge to update the predictor. At test time, given a new person, the predictor is able to be further optimized to empower personalized capabilities to the specific properties. Extensive experiments show that with H/P-TTP, the existing models are significantly improved for various unseen subjects. The project page is available at https://sites.google.com/view/hp-ttp.", + "bbox": [ + 76, + 347, + 473, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 744, + 209, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Forecasting the high-fidelity future human poses, conditioned on a given historical sequence, has attracted increasing attention in recent years. In a variety of 3D vision-based applications, such as autonomous driving and robot navigation, it offers tremendous potential, especially for tasks that call for seamless interaction with humans [11, 9, 29].", + "bbox": [ + 76, + 767, + 468, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed a proliferation in the ex", + "bbox": [ + 96, + 859, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1186abb990347563455d35c0471caaede26b0fa002a3c508ce4cf29ec8d7d050.jpg", + "image_caption": [ + "Figure 1. Starting from off-the-shelf networks, we first train a baseline model [26] on the source data, which is further optimized and personalized to the target person during the test phase. Here, we present an illustration of the pose sequence of various subjects, in which each 3D feature embedding is created by t-SNE. We notice that the motion patterns of test subjects fail to match the source one. Moreover, our prediction skeleton (blue) is closer to the ground truth (red) against the vanilla baseline [26]." + ], + "image_footnote": [], + "bbox": [ + 506, + 315, + 890, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ploration of this emerging field [16, 5, 2]. In particular, ongoing efforts are being devoted to deep learning approaches, which typically train a generic model from large-scale datasets, and then apply it to new test subjects and samples during the deployment phase. We notice that these methods have become mainstream and are now rapidly evolving [49, 58, 24, 32, 3, 33].", + "bbox": [ + 496, + 642, + 893, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite notable successes, the existing approaches are sub-optimal, because the generic model cannot be adapted to unseen personalized subjects [48, 21, 52]. To be precise, due to data limitations or privacy concerns, unseen test persons, as well as their individual properties (e.g., age, gender, or behavioral trait), are typically disjointed with the training. It may lead to a large distribution gap between the target and source domain [19, 42, 50], as shown in Figure 1. For example, for the H3.6M dataset [22], a widely-adopted data splitting strategy takes subject-5 (S5) as the test and", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "274", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the rest as the training [10, 34, 25]. Here, the inherent properties (height, and body proportion) of S5 are significantly different from the other subjects; hence, the pre-trained parameters may not be specific or personalized for the S5. We also note that, out-of-source persons are inevitable in the deployment, which poses a major bottleneck to current predictive algorithms [3].", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve it, a novel helper-predictor test-time personalization framework (H/P-TTP) is proposed. Starting with ready-to-use networks, our approach can attain tailored parameters according to the characteristics of various unseen subjects. Concretely, the H/P-TTP advances the recent success of knowledge distillation into test-time adaptation by augmenting source instances to out-of-source distributions, enabling domain-generalizable learning. The helper involves an identical architecture with the predictor, except for the explicit and implicit augmenters. To improve the robustness, by adding noise to the samples, the explicit augmenter extracts task-relevant information and ignores irrelevant ones. In contrast, the implicit augmenter is a trainable motion-style transformer to achieve the augmentations with new properties, which follows the adversarial training paradigm, with the goal of maximizing the discrepancy with the source subject, while ensuring semantic proximity. With the max-min optimization, the helper would be able to observe and cope with the novel-styled samples to gain the cross-subject invariant context, and distill the knowledge to update the predictor.", + "bbox": [ + 75, + 198, + 467, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During testing, the model can be further updated to adapt to an unseen test person. We also observe that due to psychological or environmental factors, even for the same person, his/her motion characteristics may vary. More intuitively, for test data with large domain gaps, the predictor needs greater updates, and conversely, a mild personalization is expected. For this purpose, an adaptive learning rate strategy is proposed to dynamically consider the demand for test-time personalization. Owing to the adaptive test-time learning rate, our T/P-TTP can adapt to the intrinsic attributes of an unseen test subject, as well as to his/her specific motion dynamics, as shown in Figure 2.", + "bbox": [ + 75, + 513, + 467, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows: 1) We notice the issue: the existing approaches are ill-adapted to the out-of-source subjects, and propose a novel H/P-TTP model to address it. 2) We also propose an adaptive learning rate strategy to further consider the extent to which the model needs to be personalized. 3) Our H/P-TTP is integrated into many existing motion forecasting approaches, and substantially boosts the performance on unseen subjects.", + "bbox": [ + 75, + 696, + 467, + 815 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 830, + 217, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Pose Forecasting. Deep end-to-end networks are highly sought-after for solving this issue [5, 41]. In particular, due to the intrinsic temporal nature of 3D human", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "motion, it is typically considered a sequence-to-sequence problem, where various variants of RNNs have been developed [35, 36, 15, 17]. Despite the promising development, RNNs suffer from error accumulation, convergence to an undesired static pose [24, 27, 25].", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle these drawbacks, feed-forward networks have been recently explored as a potential alternative [25, 39, 34, 33]. Particularly, with better interpretability, GCNs are introduced to extract the local semantic connectivity of the 3D human skeleton [34, 10, 8, 11, 24, 27, 36, 26, 54]. Transformers have also been introduced to capture the long-range dependencies of motion sequences [32].", + "bbox": [ + 498, + 169, + 890, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We have noticed that, current deep learning-based approaches have emerged as the prevailing solution with promising results. However, an obvious challenge still exists, namely, the source pre-trained model cannot be personalized for the specific properties of unseen target persons.", + "bbox": [ + 498, + 277, + 890, + 352 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain Generalization (DG). It aims to generalize the learned model from source domain datasets to unseen target domains. Recent advances reveal that data augmentation is an attractive strategy for DG, with the increase in the diversity of training and test data [20, 1, 56, 59]. For this purpose, a variety of works are proposed, e.g., adversarial attack [57, 60], image perturbation [60], and domain randomization [46, 47], where the main heuristic is to learn cross-domain invariant feature. Particularly, [55] introduces a novel-style augmentation in teacher-student frameworks, in which novel synthesized data allows for domain-invariant context to be transferred to the student network. In contrast to the standard DG, which trains a more generalizable model, our work considers going to improve the off-the-shelf network to adapt to the specific characteristics of unseen new subjects and their motion sequence.", + "bbox": [ + 498, + 354, + 890, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test-time Adaptation (TTA). Our test-time personalization is relevant to the TTA, which is expected to improve the generalization ability of out-of-distribution test samples. TTA falls into source-free adaptation, and before making decisions, allows for fine-tuning the pre-trained model to adapt to the specific test samples during inference, without accessing the source domain data. For example, according to the test data, prediction-time batch normalization (PTBN) [38] can recalibrate the BN statistics. [51] utilizes test-time entropy minimization (TENT) to update the trainable parameters in BN layers. Test-time training (TTT) [44] adjusts the feature extractor in the test phase via the update of the self-supervised auxiliary task of image rotation prediction. $\\mathrm{TTT}++$ [30] exploits contrastive learning to align the first- and second-order statistics of the test samples with the source domain data, improving the TTT further. [7] introduces an auxiliary image reconstruction to the primary deblurring task, and then resorts to meta-learning to fine-tune the primary branch to adapt to test samples. Considering the changing application environment, [53] integrates", + "bbox": [ + 498, + 599, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "275", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/03f95f96bf887cf52742262d5dfc3dcf1c16863a61ccde499de8c53293750a9d.jpg", + "image_caption": [ + "(a) H/P Domain-Generalizable Learning" + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 439, + 296 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3c8a8c443fe97fc86ac4f0976275b2556c1a40fded8a9072110c42dd1bfc4526.jpg", + "image_caption": [ + "(b) H/P Test-time Personalizable Forecasting", + "Figure 2. Illustration of our approach. It mainly includes a parameter-free explicit augmenter $\\mathcal{A}^{(E)}$ and a learnable implicit augmenter $\\mathcal{A}_{\\phi}^{(I)}$ , as well as the helper $\\mathcal{H}_{\\psi}$ and predictor $\\mathcal{P}_{\\theta}$ . $\\mathcal{A}^{(E)}$ is to attain the noisy samples to improve the robustness and assist $\\mathcal{A}_{\\phi}^{(I)}$ to generate the novel-domain samples. (a) H/P domain-generalizable learning involves two steps: The implicit augmenter is first trained with the adversarial learning paradigm, to make the diverse novel-domain samples, and reserve the semantics; $\\mathcal{P}_{\\theta}$ is trained to minimize the prediction loss $\\mathcal{L}_{pred}$ , while $\\mathcal{H}_{\\psi}$ is expected to achieve the consistent output with the predictor to distill the domain-invariant knowledge. We set the learning rate to $\\beta$ to obtain the base model $\\{\\stackrel{\\circ}{\\psi}, \\stackrel{\\circ}{\\theta}\\}$ . (b) H/P Test-time Personalizable Forecasting is to further optimize the base helper/predictor to attain the personalized parameters for a new subject. We also analyze the demand of test-time personalization of a specific sequence, which is achieved by dynamically adjusting the test learning rate." + ], + "image_footnote": [], + "bbox": [ + 442, + 89, + 883, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the teacher-student framework into continue test-time adaptation (CoTTA) to distill domain-generalizable knowledge and improve the robustness of the unseen domain.", + "bbox": [ + 75, + 436, + 467, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our inspiration comes partially from the above literature, with the following important distinctions: (1) It can be integrated with any existing deep end-to-end model. (2) Due to the meaningful novel-domain augmentation, high interpretability is provided. (3) It is possible to identify the demand to which the test sample needs to be adapted.", + "bbox": [ + 75, + 481, + 467, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 76, + 583, + 266, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an observed sequence $\\mathbf{x}_{1:T} = [x_1, x_2, \\dots, x_T]$ of a person across a time horizon $T$ , our goal is to predict his/her future actions $\\mathbf{y}_{1:\\Delta T} = [y_1, y_2, \\dots, y_{\\Delta T}]$ , in which each frame records the 3D coordinates of $N$ joints. Typically, current approaches [35, 34, 11, 10] learn a generic mapping $f: \\mathbf{x}_{1:T} \\rightarrow \\mathbf{y}_{1:\\Delta T}$ . As an extension, our H/P-TTP intends to boost their generalization capability to make personalizable predictions of unseen test subjects/motions.", + "bbox": [ + 75, + 604, + 467, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Novel Data Augmentation", + "text_level": 1, + "bbox": [ + 76, + 729, + 310, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our H/P-TTP consists of a predictor and a helper with an identical structure, parameterized by $\\mathcal{P}_{\\theta}$ and $\\mathcal{H}_{\\psi}$ , derived from any existing networks or newly-designed as well. In contrast to the predictor, the helper, outfitted with the implicit and explicit augmenters, w.r.t. $\\mathcal{A}_{\\phi}^{(I)}$ and $\\mathcal{A}^{(E)}$ , is dedicated to generating the augmented data.", + "bbox": [ + 75, + 747, + 467, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implicit Augmenter. It is introduced to attain a generalizable feature, which falls into the max-min adversarial learning scope. Intuitively, the implicit augmenter $\\mathcal{A}_{\\phi}^{(T)}$ is expected to generate diverse novel-domain augmentations", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with a variety of attributes (e.g., ages, genders, statures) for the source sequence. After the adversarial learning is finished, the predictor allows for the distillation of domain-generalized representations from the helper, thus improving its personalization ability for unseen subject sequences.", + "bbox": [ + 496, + 436, + 890, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We note that, $\\mathcal{A}_{\\phi}^{(I)}$ can be interpreted as a specially designed motion-style migration network, which is used to implicitly diversify the sequence into its younger/older or taller/lower views. It is associated with the literature on diverse human motion generation [3, 2, 12]. Therefore, we set up $\\mathcal{A}_{\\phi}^{(I)}$ with a similar network as [33], where the significant extraordinary is that it is trained with the max-min optimization and to yield novel-styled counterparts [55].", + "bbox": [ + 496, + 511, + 890, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an observed sequence $\\mathbf{x}$ , the purpose of $\\mathcal{A}_{\\phi}^{(I)}$ is to produce stylized augmentations $\\tilde{\\mathbf{x}}^{(I)} = \\mathcal{A}_{\\phi}^{(I)}(\\mathbf{x})$ yet retaining its semantic (action label). To stimulate the augmentations with sufficiently distinct appearances, we propose to maximize the discrepancy $\\mathcal{L}_{dis}$ between the original observed motion and $H$ augmented ones $\\{\\tilde{\\mathbf{x}}_h^{(I)}\\}_{h=1}^H$ :", + "bbox": [ + 496, + 633, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\mathcal {A} ^ {(I)}} \\mathcal {L} _ {d i s} \\left(\\tilde {\\mathbf {x}} ^ {(I)}, \\mathbf {x}\\right) = \\sum_ {h = 1} ^ {H} \\left\\| \\frac {\\tilde {\\mathbf {x}} _ {h} ^ {(I)}}{\\left\\| \\tilde {\\mathbf {x}} _ {h} ^ {(I)} \\right\\| _ {2}} - \\frac {\\mathbf {x}}{\\left\\| \\mathbf {x} \\right\\| _ {2}} \\right\\| _ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 733, + 890, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To motivate diversity across the implicit augmentations, consistent with [58, 33], the diversity-promoting prior loss $\\mathcal{L}_{div}$ with a normalizing factor $\\delta = 100$ is adopted:", + "bbox": [ + 496, + 776, + 890, + 819 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i v} = \\frac {2}{H (H - 1)} \\sum_ {j = 1} ^ {H} \\sum_ {h = j + 1} ^ {H} \\exp \\left(- \\frac {\\left\\| \\tilde {\\mathbf {x}} _ {j} ^ {(I)} - \\tilde {\\mathbf {x}} _ {h} ^ {(I)} \\right\\| _ {1}}{\\delta}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 824, + 890, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, to preserve the semantic proximity, we also introduce an action classifier to make the category label of", + "bbox": [ + 496, + 871, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "276", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the novel-styled sequence consistent with the observed one. Given the correct label $p$ in the $C$ -sized label space of a sequence and the resulting one $\\tilde{p}_h^{(I)}$ of $h$ -th implicit augmentation, it is trained with cross-entropy loss $\\mathcal{L}_{cls}$ :", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = - \\frac {1}{C H} \\sum_ {h = 1} ^ {H} p \\log \\tilde {p} _ {h} ^ {(I)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 154, + 468, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, the implicit augmenter $\\mathcal{A}_{\\phi}^{(I)}$ is optimized to maximize the discrepancy, and minimize the diversity and classification error, which is expected to generate diverse augmentations as much as possible, while maintaining the semantic information with the original one [24]. After observing the new-domain augmentations, the helper can extract the domain-generalizable representation to optimize the predictor to personalize according to input samples.", + "bbox": [ + 75, + 195, + 468, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Explicit Augmenter. To incorporate the additional information into the implicit augmenter, we introduce a nontrainable explicit one. In instance learning, each image is regarded as an instance, and we wish to train the network so that the representations of different augmented views of the same instance are as close as possible to each other. It provides more rich data for training, and helps learn transformation invariant representations [37, 40]. For natural images, random rotation, flipping, and noise injection are the most common augmentation strategies [14, 23, 42].", + "bbox": [ + 75, + 316, + 468, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides, [8] develops a multi-task framework, in which the auxiliary branch is designed as motion reconstruction to provide meaningful cues for human pose prediction.", + "bbox": [ + 75, + 469, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by the above-mentioned work, we, therefore, design the explicit augmenter $\\mathcal{A}^{(E)}$ to produce the noisy counterparts. Formally, given a distribution sampled from the observed sequences, $\\mathbf{x} \\sim q(\\mathbf{x})$ , it is injected with several Gaussian noises with a series of variances of a fixed interval, producing $K$ augmentations with different noise levels:", + "bbox": [ + 75, + 513, + 468, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} \\left(\\tilde {\\mathbf {x}} _ {k} ^ {(E)} \\mid \\mathbf {x}\\right) = \\mathcal {N} \\left(\\tilde {\\mathbf {x}} _ {k} ^ {(E)}; \\mathbf {x}, \\sigma^ {2} \\mathbf {I}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 609, + 468, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{\\tilde{\\mathbf{x}}_k^{(E)}\\}_{k = 1}^K$ is the resulting noisy counterparts, which, together with the novel-domain augmentation $\\{\\tilde{\\mathbf{x}}_h^{(I)}\\}_{h = 1}^H$ from the learnable $\\mathcal{A}_{\\phi}^{(I)}$ , is used to train the helper.", + "bbox": [ + 75, + 637, + 467, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we note that, the non-parametric explicit augmenter provides the following advantages: (1) The noise counterparts are supplied, facilitating the helper to explore new patterns in the novel-domain augmented data. (2) It eliminates an overly optimistic assumption: the trainable implicit augmenter will necessarily yield novel-domain samples. (3) More robustness to noise is also included.", + "bbox": [ + 75, + 691, + 468, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. H/P Domain-Generalizable Learning", + "text_level": 1, + "bbox": [ + 76, + 804, + 397, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The objective of our helper is to improve the generalization of the predictor for unseen domain samples by observing the new-domain augmentations and noise counterparts, as shown in Figure 2 (left). The $\\mathcal{A}^{(E)}$ is parameter-free, while $\\mathcal{A}_{\\phi}^{(I)}$ is trained to solve the following problem:", + "bbox": [ + 75, + 823, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 H/P Domain-Generalizable Learning", + "text_level": 1, + "bbox": [ + 500, + 90, + 834, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Require: implicit augmenter $\\mathcal{A}_{\\phi}^{(I)}$ ; helper $\\mathcal{H}_{\\psi}$ and predictor $\\mathcal{P}_{\\theta}$ with the identical structure; learning rates $\\alpha$ , $\\beta$ ; batch size $B$ ; hyperparameters $\\lambda_{dis}$ , $\\lambda_{div}$ , $\\lambda_{cls}$", + "bbox": [ + 500, + 109, + 890, + 156 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: the learned parameter $\\psi, \\theta, \\phi$ ;", + "bbox": [ + 500, + 156, + 756, + 171 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: randomly initialize $\\psi$ , $\\theta$ and $\\phi$ ;", + "2: while not converge do", + "3: sample a training mini-batch $\\{\\mathbf{x}_b,\\mathbf{y}_b\\}_{b = 1}^B$", + "4: for each $b$ do $\\triangleright$ update implicit augmenter", + "5: $\\phi \\gets \\phi -\\alpha \\nabla_{\\phi}(-\\lambda_{dis}\\mathcal{L}_{dis} + \\lambda_{div}\\mathcal{L}_{div} + \\lambda_{cls}\\mathcal{L}_{cls})$", + "6: end for", + "7: end while", + "8: while not converge do", + "9: sample a training mini-batch $\\{\\mathbf{x}_b,\\mathbf{y}_b\\}_{b = 1}^B$ update the helper and predictor", + "10: $\\psi \\gets \\psi -\\sum_{b = 1}^{B}\\beta \\nabla_{\\psi}\\mathcal{L}_{pred}(\\tilde{\\bar{\\mathbf{y}}}_b^{(\\mathcal{H})},\\hat{\\mathbf{y}}_b^{(\\mathcal{P})});$", + "11: $\\theta \\gets \\theta -\\sum_{b = 1}^{B}\\beta \\nabla_{\\theta}\\mathcal{L}_{pred}(\\hat{\\mathbf{y}}_b^{(\\mathcal{P})},\\mathbf{y}_b)$", + "12: end while" + ], + "bbox": [ + 501, + 172, + 890, + 369 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\phi \\in \\Phi} \\max \\lambda_ {d i s} \\mathcal {L} _ {d i s} + \\lambda_ {d i v} \\mathcal {L} _ {d i v} + \\lambda_ {c l s} \\mathcal {L} _ {c l s}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 377, + 890, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{dis} = 0.5$ , $\\lambda_{div} = 0.3$ and $\\lambda_{cls} = 0.2$ are the trade-off parameters. With a learning rate $\\alpha = 0.001$ , the implicit augmenter $\\mathcal{A}_{\\phi}^{(I)}$ is optimized:", + "bbox": [ + 498, + 405, + 890, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi \\leftarrow \\phi - \\alpha \\nabla_ {\\phi} \\left(- \\lambda_ {d i s} \\mathcal {L} _ {d i s} + \\lambda_ {d i v} \\mathcal {L} _ {d i v} + \\lambda_ {c l s} \\mathcal {L} _ {c l s}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 459, + 890, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then train both the helper $\\mathcal{H}_{\\psi}$ and predictor $\\mathcal{P}_{\\theta}$ to achieve the H/P domain-generalizable learning. The optimization of $\\mathcal{H}_{\\psi}$ is achieved using the gradient descent:", + "bbox": [ + 498, + 481, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\psi \\leftarrow \\psi - \\beta \\nabla_ {\\psi} \\mathcal {L} _ {p r e d} \\left(\\tilde {\\mathbf {y}} ^ {(\\mathcal {H})}, \\hat {\\mathbf {y}} ^ {(\\mathcal {P})}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 532, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{y}}^{(\\mathcal{P})}$ is the predictor's result, $\\beta = 0.001$ is the learning rate. The average outcome $\\tilde{\\bar{\\xi}}^{(\\mathcal{H})}$ is yielded over a total of $H + K$ augmentations $\\{\\tilde{\\mathbf{x}}_i\\}_{i=1}^{H+K}$ fed into the helper $\\mathcal{H}_{\\psi}$ . Following [10, 8], the prediction loss $\\mathcal{L}_{pred}$ is denoted as $L_{2}$ distance. The update of the predictor $\\mathcal{P}_{\\theta}$ can be achieved:", + "bbox": [ + 498, + 558, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta \\leftarrow \\theta - \\beta \\nabla_ {\\theta} \\mathcal {L} _ {p r e d} (\\hat {\\mathbf {y}} ^ {(\\mathcal {P})}, \\mathbf {y}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 637, + 890, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{y}$ is the future ground-truth (gt) action.", + "bbox": [ + 498, + 657, + 800, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The optimization of Eq. 7 and Eq. 8 ensures that the helper and predictor have consistent outcomes that are as close to the actual future sequence as possible. The domain-generalizable knowledge from the helper is progressively transferred to the predictor. The overall H/P domain-generalizable learning is summarized in Algorithm 1. Then, the base model, w.r.t $\\{\\ddot{\\psi},\\dot{\\theta}\\}$ , is achieved.", + "bbox": [ + 496, + 672, + 890, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. H/P Test-time Personalizable Forecasting", + "text_level": 1, + "bbox": [ + 500, + 787, + 851, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During inference, for a new person, we consider further fine-tuning the base model to adapt to his/her specific properties. However, in practical applications, even for the same subject, his/her motion patterns may differ due to external factors, such as environmental changes. This requires varying degrees of test-time personalization (TTP) to adapt.", + "bbox": [ + 496, + 810, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "277", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For this purpose, we propose to adjust the learning rate of test-time personalization. Our intuition is that, for the current step, its previous samples and their correct future sequences are observed deterministically. We are able to know how the association of the helper/predictor feature leads to the difference between the predicted and true values. Therefore, for the next sample, at test-time, calculating the distinction between the features of the helper and predictor, we can also gain insight into the relation of their expected outputs; if it is strong, then the helper can guide the predictor well, and a lower learning rate should be assigned to update the predictor, otherwise, a larger one.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Memory queue. For simplicity, the base helper $\\mathcal{H}_{\\hat{\\psi}}$ and predictor $\\mathcal{P}_{\\hat{\\theta}}$ are decomposed: feature extractor and generator, w.r.t $\\hat{\\psi} = \\{\\hat{\\psi}_1,\\hat{\\psi}_2\\}$ and $\\dot{\\theta} = \\{\\dot{\\theta}_1,\\dot{\\theta}_2\\}$ . We denote $h$ and $p$ as the helper and predictor. The outcome of the feature extractor and generator are denoted as $\\mathbf{f}$ and $\\mathbf{o}$ . For a historical sample $\\mathbf{x}^{(m)}$ , $\\mathbf{f}_{\\{h\\leftrightarrow p\\}}^{(m)}$ and $\\mathbf{o}_{\\{h\\leftrightarrow p\\}}^{(m)}$ are the feature and outcome differences of $h / p$ branches, calculated by a distance metric $\\mathcal{D}$ (i.e., euclidean distance), informally expressed $\\leftrightarrow$ . $\\mathbf{o}_{\\{p\\leftrightarrow gt\\}}^{(m)}$ calculates that the difference between the predictor's result and ground truth, indicating, from real history views, whether the predictor is well enough.", + "bbox": [ + 75, + 272, + 472, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We build a $M$ -elements memory queue $\\mathbb{M} = \\{(\\mathbf{f}_{\\{h \\leftrightarrow p\\}}^{(m)}, \\mathbf{o}_{\\{p \\leftrightarrow gt\\}}^{(m)})\\}_{m=1}^{M}$ , denoted as:", + "bbox": [ + 76, + 445, + 472, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\rightarrow \\underbrace {\\left(\\mathbf {f} _ {\\{h \\leftrightarrow p \\}} ^ {(1)} , \\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(1)}\\right) , \\dots , \\left(\\mathbf {f} _ {\\{h \\leftrightarrow p \\}} ^ {(M)} , \\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(M)}\\right)} _ {\\mathbb {M}}, \\rightarrow \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 488, + 472, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It records, in terms of the real history, the relation of the $h / p$ features and the difference between the predicted results and the ground truth (GT). In other words, it measures whether the helper is able to distill domain-generalizable information to the predictor for previous samples (or whether the model has been calibrated well). We note that $\\mathbb{M}$ is updated by First In First Out (FIFO).", + "bbox": [ + 75, + 527, + 472, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a new arriving sample $\\mathbf{x}^{(0)}$ , we calculate the feature difference $\\mathbf{f}_{\\{h\\leftrightarrow p\\}}^{(0)}$ and the outcome difference $\\mathbf{o}_{\\{h\\leftrightarrow p\\}}^{(0)}$ . To measure its demand to be personalized, $\\mathbf{f}_{\\{h\\leftrightarrow p\\}}^{(0)}$ is taken as the query, and retrieve $D$ most similar ones $\\{\\mathbf{f}_{\\{h\\leftrightarrow p\\}}^{(d)}\\}_{d = 1}^D$ from $\\mathbb{M}$ , with their corresponding $\\{\\mathbf{o}_{\\{p\\leftrightarrow gt\\}}^{(d)}\\}_{d = 1}^D$ as the support set $\\mathbb{D}$ :", + "bbox": [ + 75, + 633, + 472, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underbrace {\\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(1)} , \\dots \\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(d)} , \\dots \\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(D)}} _ {\\mathbb {D}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 747, + 468, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This retrieval process is implemented by $D$ -Nearest Neighbor algorithm.", + "bbox": [ + 75, + 786, + 468, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dynamic learning rate. We then average the values in $\\mathbb{D}$ as the reference $\\mathbf{r}$ . The prediction discrepancy (PD) of $\\mathbf{x}^{(0)}$ is derived:", + "bbox": [ + 76, + 816, + 470, + 858 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P D} = \\frac {1}{2} \\left(\\mathcal {L} _ {K L} \\left(\\mathbf {r} \\| \\mathbf {o} _ {\\{h \\leftrightarrow p \\}} ^ {(0)}\\right) + \\mathcal {L} _ {K L} \\left(\\mathbf {o} _ {\\{h \\leftrightarrow p \\}} ^ {(0)} \\| \\mathbf {r}\\right)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 859, + 468, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{KL}$ is the KL divergence of two distributions.", + "bbox": [ + 81, + 885, + 437, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 2 H/P Test-time Personalizable Forecasting", + "text_level": 1, + "bbox": [ + 500, + 90, + 864, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: test sample $\\mathbf{x}^{(0)}$ ; base predictor $\\dot{\\theta} = \\{\\ddot{\\theta}_1,\\ddot{\\theta}_2\\}$ ; base helper $\\dot{\\psi} = \\{\\bar{\\psi}_1,\\bar{\\psi}_2\\}$ ; memory queue $\\mathbb{M}$ ; learning rate $\\beta$ ; momentum $\\mu$ ; number of nearest neighbors $D$ ;", + "bbox": [ + 500, + 108, + 890, + 156 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: final prediction $\\hat{\\mathbf{y}}^{*}$", + "bbox": [ + 500, + 156, + 691, + 171 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: augment $\\mathbf{x}^{(0)}$ to $H + K$ augmentations $\\{\\tilde{\\mathbf{x}}_i^{(0)}\\}_{i = 1}^{H + K}$ and make $\\tilde{\\bar{\\mathbf{x}}}^{(0)}$ equal to their average;", + "2: calculate $\\mathbf{f}_{h\\leftrightarrow p}^{(0)} = \\mathcal{D}(\\mathcal{P}_{\\theta_1}(\\mathbf{x}^{(0)}),\\mathcal{H}_{\\psi_1}(\\bar{\\tilde{\\mathbf{x}}}^{(0)}))$ $\\mathbf{o}_{h\\leftrightarrow p}^{(0)} = \\mathcal{D}(\\mathcal{P}_{\\theta}(\\mathbf{x}^{(0)}),\\mathcal{H}_{\\psi}(\\bar{\\tilde{\\mathbf{x}}}^{(0)}));$", + "3: retrieve the support set $\\mathbb{D}$ from $\\mathbb{M}$ using $D$ -neighbors nearest with the query $\\mathbf{f}_{h\\leftrightarrow p}^{(0)}$ ;", + "4: ensemble the values in $\\mathbb{D}$ as the reference $\\mathbf{r}$ ;", + "5: calculate the prediction discrepancy $\\mathcal{L}_{PD}$ using Eq. 11;", + "6: obtain the adjusted learning rate $\\gamma$ using Eq. 12;", + "7: update the predictor $\\mathcal{P}_{\\theta}:\\theta^{*}\\gets \\dot{\\theta}$ using Eq. 13;", + "8: update the helper $\\mathcal{H}_{\\psi}:\\psi^{*}\\gets \\bar{\\psi}$ , using Eq. 14;", + "9: return $\\hat{\\mathbf{y}}^{*} = \\mathcal{P}_{\\theta^{*}}(\\mathbf{x})$" + ], + "bbox": [ + 500, + 171, + 890, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the next sample in continuous flow, we have known the actual future sequence $\\mathbf{y}^{(0)}$ of the previous sample $\\mathbf{x}^{(0)}$ .", + "bbox": [ + 500, + 369, + 890, + 398 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $\\mathbf{o}_{p\\leftrightarrow qt}^{(0)} = \\mathcal{D}(\\mathcal{P}_{\\theta}(\\mathbf{x}^{(0)}),\\mathbf{y}^{(0)})$ ; $\\triangleright$ update memory queue", + "2: M.pop() && M.push( $\\mathbf{f}_{h \\leftrightarrow p}^{(0)}, \\mathbf{o}_{p \\leftrightarrow gt}^{(0)}$ );" + ], + "bbox": [ + 501, + 398, + 890, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If the value of $\\mathcal{L}_{PD}$ is small, the predictor's outcome coincides with the real history (the helper can guide the predictor well); otherwise, it is not consistent (a major update is needed). Based on this consideration, we adjust the learning rate to make a dynamic TTP. At test time, we obtain the adjusted learning rate $\\gamma$ as:", + "bbox": [ + 498, + 449, + 890, + 541 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma = \\beta \\cdot \\mathcal {L} _ {P D}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 547, + 890, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our TTP begins with the base model, $\\mathcal{P}_{\\hat{\\theta}}$ and $\\mathcal{H}_{\\hat{\\psi}}$ , obtained from the H/P novel-domain generalizable training stage. Then, we attain the predictor's personalized parameter to boost the generalization ability for the specific subject $\\mathbf{x}^{(0)}$ using a single gradient descent step:", + "bbox": [ + 498, + 568, + 890, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\stackrel {\\circ} {\\theta} \\leftarrow \\stackrel {\\circ} {\\theta} - \\gamma \\cdot \\nabla_ {\\hat {\\theta}} \\mathcal {L} _ {p r e d} \\left(\\hat {\\mathbf {y}} ^ {(0)}, \\bar {\\tilde {\\mathbf {y}}} ^ {(0)}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 648, + 890, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{y}}^{(0)} = \\mathcal{P}_{\\hat{\\theta}}(\\mathbf{x})$ is the immediate result, and the pseudo label $\\bar{\\tilde{\\mathbf{y}}}^{(0)}$ is the average prediction of the base helper over $H + K$ augmentations. Our model can be regarded as a special teacher-student framework, and therefore, after the predictor's update, we use exponential moving average (EMA) [53] with $\\mu = 0.95$ momentum to optimize the helper:", + "bbox": [ + 498, + 674, + 890, + 768 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\stackrel {\\circ} {\\psi} \\leftarrow \\mu \\cdot \\stackrel {\\circ} {\\psi} + (1 - \\mu) \\cdot \\stackrel {\\circ} {\\theta}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 771, + 890, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When the TTP is completed, we re-write the personalized parameters $\\{\\hat{\\theta},\\hat{\\psi}\\}$ as $\\{\\theta^{*},\\psi^{*}\\}$ , and perform a forward pass to yield the final prediction $\\hat{\\mathbf{y}}^{*} = \\mathcal{P}_{\\theta^{*}}(\\mathbf{x})$ . Note that, after $P = 72$ test-time personalizations, the personalized parameters $\\{\\theta^{*},\\psi^{*}\\}$ are rolled back to the base model $\\{\\hat{\\theta},\\hat{\\psi}\\}$ to avoid catastrophic forgetting. We summarize the H/P test-time personalizable forecasting in Algorithm 2.", + "bbox": [ + 496, + 795, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "278", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 89, + 209, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 112, + 294, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, our H/P-TTP begins with an off-the-shelf network, and includes two major phases: (1) H/P novel-domain generalizable training, (2) test-time personalizable forecasting. For the former, we first train the implicit augmenter $\\mathcal{A}_{\\phi}^{(I)}$ using Eq. 6. Both implicit and explicit augmenters yield $H = K = 8$ augmentations. The variances of Gaussian noise injected into the explicit augmenter are set to $\\sigma^2 = [10,15,20,25,30,35,40,45]$ in millimeter. Moreover, the classifier is set to a 3-layers MLP with 512 hidden units to ensure a similar action label between the augmentations and the original sample. Once $\\mathcal{A}_{\\phi}^{(I)}$ is trained, it is fixed in the following process. We exploit Eq. 7 and Eq. 8 to train the predictor $\\mathcal{P}_{\\theta}$ and helper $\\mathcal{H}_{\\psi}$ . The predictor/helper are optimized using AdamW optimizer with the weight decay factor (1e-2) on mini-batches of size 32, to achieve the base model. The size of the memory queue $\\mathbb{M}$ is set to $M = 36$ , and the support set $\\mathbb{D}$ is $D = 12$ . Specifically, we divide the baseline networks by the penultimate layer, where the previous layers are treated as the feature extractor, and it and its following layers are the generator.", + "bbox": [ + 75, + 133, + 472, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Datasets", + "text_level": 1, + "bbox": [ + 76, + 448, + 179, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our H/P-TTP is evaluated on 3 benchmark datasets.", + "bbox": [ + 94, + 468, + 439, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset-1: H3.6M [22] is the most widely-used dataset for pose prediction, containing $\\approx 3.6$ million frames of 15 action categories performed by 7 human subjects. Consistent with [34, 6], all sequences are downsampled to 25 fps, and represented as a $N = 17$ joint skeleton.", + "bbox": [ + 75, + 484, + 468, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset-2: GRAB [45] is a newly-introduced benchmark, including $\\approx 1.6$ million poses of 29 actions from 10 human subjects. Compared with H3.6M, the pose sequences in GRAB are more diverse and involve interaction with the physical world, which is, therefore, more complex and challenging. We down-sample the sequences to 30 fps with 25 joints in each skeleton [13].", + "bbox": [ + 75, + 563, + 468, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset-3: HumanEva-I [43] consists of 3 human subjects, performing 6 pre-defined actions. All sequence is down-sampled to 30 fps, represented by a 15-joint skeleton. The results are reported in the supplementary material.", + "bbox": [ + 75, + 670, + 468, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We note that, for H3.6M [22], the length of the observed and predicted sequence is 25 frames, while for both GRAB [45] and HumanEva-I [43], it is 30.", + "bbox": [ + 75, + 732, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Experimental Setups", + "text_level": 1, + "bbox": [ + 76, + 789, + 274, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use 2 main setups to analyze our model (1.x and 2.x as stated in Table 1). As in [34, 32, 10, 31, 13], the former is to evaluate the general prediction ability on common data splitting. The latter is newly-designed to evaluate the personalizable ability of unseen subjects, where $S_{x}$ is the test subjects, and the disjointed ones $S_{x}^{-}$ is the training.", + "bbox": [ + 75, + 811, + 470, + 905 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/581e40941f262d875ea987a0b87d2e58dce17b2030f323042e7599af3c200ee3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Purposegeneral predictive ability (1.x)predictive ability for unseen subjects (2.x)
Setup(1.1)(1.2)(2.1)(2.2)(2.3)
DatasetH3.6M [22]GRAB[45]H3.6M[22]GRAB[45]HumanEva-I [43]
TestS5S10SxSxSx
TrainingS-5S-10S-S-S-
", + "bbox": [ + 503, + 88, + 890, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Experimental setups.", + "bbox": [ + 602, + 166, + 781, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Baselines", + "text_level": 1, + "bbox": [ + 500, + 180, + 607, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We exploit 6 approaches, emerged in recent years as our baselines, including ResSup [35], LTD [34], HisRep [32], MSR [11], PGBIG [31], and SPGSN [26]. They are separately integrated into our H/P-TTP framework to activate the prediction performance for unseen targets. [35] is RNN-based, [32] is attention-based, and the rest are based on GCNs. For the general predictive ability, we use the common data splitting, as the setup-1.x in Table 1, and the setup2.x in Table 1 is used to endow the baselines with the personalizable ability for unseen subjects. We note that our ultimate goal is to equip the H/P-TTP with each baseline to improve the prediction ability, which is therefore renamed, e.g., PGBIG [31] + H/P-TTP, SPGSN [26] + H/P-TTP.", + "bbox": [ + 496, + 199, + 890, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Protocols", + "text_level": 1, + "bbox": [ + 500, + 402, + 609, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To comprehensively investigate the performance of our H/P-TTP, the following 4 protocols are exploited.", + "bbox": [ + 500, + 420, + 890, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Protocol-1: The widely-applied mean per-joint position error (MPJPE) [22, 34, 6, 31] is first used:", + "bbox": [ + 500, + 452, + 890, + 481 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathrm {M P J P E}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\| \\boldsymbol {p} _ {n} - \\hat {\\boldsymbol {p}} _ {n} ^ {*} \\| _ {2}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 484, + 890, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{p}}_n^*$ is the 3D position of $n$ -th joint in a final predicted frame $\\hat{\\pmb{y}}_t^*$ , $\\pmb{p}_n$ is the gt. $N$ is the number of joints.", + "bbox": [ + 500, + 518, + 890, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Protocol-2: We observe that compared with the GT, some predicted pose coordinates have a slight global offset. To eliminate the error independent of poses, we supplement the Procrustes aligned MPJPE (P-MPJPE), which calculates the MPJPE after aligning the predicted pose to the gt pose by a rigid transformation called Procrustes Analysis (PA) [28]. For a frame $t$ in a skeleton layout $S$ , it is defined as:", + "bbox": [ + 496, + 549, + 890, + 655 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathrm {P - M P J P E}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\| m _ {\\mathrm {P A}, \\mathrm {S}} ^ {t} (n) - m _ {\\mathrm {g t}, \\mathrm {S}} ^ {t} (n) \\| _ {2}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 657, + 890, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $m_{\\mathrm{PA,S}}^t (n)$ is the coordinate of $n$ -th joint after alignment using PA, $m_{\\mathrm{glS}}^{t}(n)$ is the corresponding gt.", + "bbox": [ + 496, + 691, + 890, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Protocol-3: Both MPJPE and P-MPJPE are hard to quantify the performance bound. Therefore, Percentage of Correct 3D Keypoint (PCK) is also introduced, which statistics the proportion of predicted joints with MPJPE errors smaller than a pre-defined threshold. Motivated by the current 3D human pose estimation [4, 18], we use the threshold of $150\\mathrm{mm}$ , re-written as PCK@150mm.", + "bbox": [ + 496, + 723, + 890, + 827 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Protocol-4: In addition, we also use the mean per-bone length error (MPBLE) to evaluate the bone-length difference. The MPBLE is defined as: -", + "bbox": [ + 496, + 829, + 890, + 872 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathrm {M P B L E}} = \\frac {1}{N - 1} \\sum_ {n = 1} ^ {N - 1} \\| l _ {n} - \\hat {l} _ {n} ^ {*} \\| _ {1}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 875, + 890, + 904 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "279", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9d65fb181fc66469f5590b2acec3a04a34d14c70c0ee3d7713c9fd6cfce1848d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodResSupResSup +H/P-TTPLTDLTD +H/P-TTPHisRepHisRep +H/P-TTPMSR [11]MSR +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
Time (ms)400100040010004001000400100040010004001000400100040010004001000400100040010004001000
MPJPE [mm] ↓115.2130.5107.4125.663.5114.359.6109.858.3112.156.8109.462.9114.260.8108.658.5110.356.3105.158.3109.655.6103.7
P-MPJPE [mm] ↓96.6111.089.2107.153.1101.547.394.652.6100.146.294.855.7101.351.293.751.297.249.795.651.896.749.490.2
PCK@150mm [%]↑57.550.359.253.770.466.072.569.771.265.573.769.475.570.180.372.077.369.682.471.580.171.285.474.6
MPBLE [mm] ↓33.645.730.543.625.334.023.532.626.432.422.031.127.235.224.733.223.029.319.427.319.325.516.323.8
", + "bbox": [ + 81, + 88, + 883, + 178 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/600feca13f0c49e9d7533bef36e34559c83318ce929306b06a6db9b60f7bfc46.jpg", + "table_caption": [ + "Table 2. Average short-term (400ms) & long-term (1000ms) prediction performance (on H3.6M [22]) using the common setup-1.1. We observe that, in general, with the H/P-TTP, the baselines achieve better results than the trivial ones. It reveals that $S_{5}$ includes the different motion properties from the other subjects, while our H/T-TTP can adapt them with the personalized parameters." + ], + "table_footnote": [], + "table_body": "
MethodMSRMSR +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
Time (ms)500100050010005001000500100050010005001000
MPJPE ↓96.3178.688.2147.392.2157.284.4141.291.3144.585.5136.4
P-MPJPE ↓82.3131.378.8122.879.8127.566.2114.277.2129.064.9117.3
PCK@150mm↑75.365.678.367.675.866.479.069.177.067.881.170.4
MPBLE ↓21.328.520.227.020.328.019.526.219.226.618.024.7
", + "bbox": [ + 78, + 233, + 478, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Average short- (500ms) & long-term (1000ms) performance (on GRAB [45]) using the setup-1.2.", + "bbox": [ + 76, + 324, + 470, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $l_{n}$ and $\\hat{l}_n^*$ are the bone length of $n$ -th bone in the gt and predicted pose, respectively. $N - 1$ is the total number of bones in a skeleton. Smaller MPBLEs tend to show a more stable visualization of the predicted frames.", + "bbox": [ + 76, + 354, + 470, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Results Analysis", + "text_level": 1, + "bbox": [ + 76, + 424, + 238, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "General prediction ability is first evaluated under the condition of common data splitting [10, 8] with all 6 baselines, as the setups-1.x in Table 1. Specifically, for the H3.6M, Table 2 compares the average result of the 4 protocols over the samples of $S_{5}$ with the experimental setup-1.1. Moreover, under the setup-1.2, Table 3 reports the comparisons on the GRAB dataset [45]. From the results, we observe that, for the used 2 datasets, once equipped with the H/P-TTP, all baselines coincidentally achieve better performance. The reason is that for the common data splitting of H3.6M and GRAB, the test subject (subject-5 or subject-10) is disjointed from the training ones. Stated in a different way, the motion properties of the target subject-5 (S5) are unknown in training (out-of-source), which is not considered by the trivial baselines, thus making the prediction task more challenging. By contrast, with the domain-generalizable learning and test-time personalization, our H/P-TTP can attain the personalized parameters to adapt to the dynamic properties of the new test subject (S5).", + "bbox": [ + 75, + 446, + 470, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prediction ability for unseen subjects is then analyzed using the newly-designed data splitting. Because of data limitations and privacy concerns, the statures, ages, or genders of the target subjects $S_{x}$ in the deployment environment are usually evident in the training stage, resulting in the distribution gap and out-of-source motion properties (e.g., motion style, rhythm). To further evaluate the personalization ability of our H/P-TTP for unseen subjects, we design the setup-2.1&2.2 in Table 1 for H3.6M and GRAB datasets, respectively. Due to their better performance, only both PGBIG and SPGSN are selected as the competitors.", + "bbox": [ + 75, + 734, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The comparison results are reported in Table 4 and Table 5, in which each subject $S_{x}$ is tested separately and averaged over all its samples. From the results, we observe that, with H/P-TTP, the PGBIG [31] + H/P-TTP and SPGSN [26] + H/P-TTP are calibrated under almost all target subjects. It evidences that the motion properties of out-of-source unseen subjects are indeed adapted.", + "bbox": [ + 496, + 234, + 890, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also present the qualitative comparison of the SoTA SPGSN [26], and SPGSN [26] $+\\mathrm{H} / \\mathrm{P}$ -TTP in Figure 3, under camera-takepicture-1 of unseen $S_{7}$ from the GRAB dataset.", + "bbox": [ + 496, + 340, + 890, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 396, + 663, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we investigate the impact of several key components. The ablation experiments report the average (at 1000ms) under the setup-2.1 in Table 1 on H3.6M dataset.", + "bbox": [ + 496, + 419, + 890, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intuitively, the (1) test-time personalization (TTP) forecasting process can further optimize the base model to adapt to unseen target subjects. It is confirmed in Table 6. We also evaluate the impact of the (2) implicit and explicit augmenter $(\\mathcal{A}^{(E)}$ and $\\mathcal{A}_{\\phi}^{(I)})$ , and (3) dynamic learning rate. As shown in Table 6, both $\\mathcal{A}^{(E)}$ and $\\mathcal{A}_{\\phi}^{(I)}$ provide the positive effect, while $\\mathcal{A}_{\\phi}^{(I)}$ is greater. Moreover, TTP is indeed helpful, and especially with the dynamic learning rate, the performance is further improved.", + "bbox": [ + 496, + 465, + 890, + 612 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(4) Impact of the initial learning rate $\\beta$ is also investigated in Table 7(left), which is essential to H/P domain-generalizable learning and TTP. We see that $\\beta$ is insensitive to the performance. We suggest that it is because it can be calibrated by the dynamic learning rate at test time.", + "(5) Impact of the number of augmentations of the implicit and explicit augmenter is shown in Table 7(left), and $H = K = 8$ achieves the better prediction.", + "(6) Number $(u = [0,1,2,3,4])$ of gradient descent of TTP is investigated in Table 7(right). We observe that, with the dynamic learning rate, our TTP only needs to perform a single update. More updates bring no further benefits.", + "(7) Impact of the variances of Gaussian noise injection to the explicit augmenter is shown in Table 8(left), where the middle one attains the best performance.", + "(8) Impact of the sizes of $\\mathbb{M}$ and $\\mathbb{D}$ . As reported in Table 8(right), the better result is balanced on the condition of $M = 36$ and $D = 12$ . Too small leads to catastrophic forgetting, while too large increases the memory burden." + ], + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "280", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9971ce0c05e1bc58e98975d67bb355e7c36325a64c194e726b8dd223b8385a33.jpg", + "image_caption": [ + "Figure 3. Predicted pose visualization, where the top is vanilla SPGSN [26] and the bottom is SPGSN $[26] + \\mathrm{H} / \\mathrm{P}$ -TTP. The underlying green denotes the gt, and the red/blue is the prediction. The contrasting segments are highlighted in the purple box. We also enlarge some important details in the yellow ellipse. It is clear that with our H/P-TTP, the predicted poses are more accurate." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 885, + 180 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/358087b91bf385fd4d7c14691f2fe2217d5cec6830d62bfb9f463285afdc1d26.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MPJPE [mm] ↓PMPJPE [mm] ↓PCK@150mm [%] ↑MPBLE [mm] ↓
unseen subjectsPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
S1116.6109.3111.4109.697.794.595.491.967.269.367.972.429.625.826.025.2
S6103.698.298.695.284.980.981.878.561.863.964.567.026.222.923.122.3
S7107.0102.1110.4103.888.984.592.986.364.266.364.969.427.423.924.123.3
S896.592.1100.395.478.877.883.474.267.871.968.573.024.020.921.120.3
S9105.3101.2104.5100.187.083.590.386.863.265.363.968.426.623.323.522.7
S11120.8115.5117.0112.8101.597.599.095.570.272.370.975.430.026.726.926.1
Average108.3103.1107.0102.889.886.590.585.565.768.266.470.927.323.924.123.3
", + "bbox": [ + 78, + 237, + 893, + 347 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1e00478fc3018556c7a2fecbf5131237e8cdede9d08271c278ffe85956effb9a.jpg", + "table_caption": [ + "Table 4. Average performance comparison (of both SoTA PGBIG [31] and SPGSN [26], on H3.6M dataset [22]) of the end predicted pose (1000ms) over samples of each unseen subject $S_{x}$ with $x = [1,6,7,8,9,11]$ , and the corresponding average over all unseen subjects." + ], + "table_footnote": [], + "table_body": "
MPJPE [mm] ↓PMPJPE [mm] ↓PCK@150mm [%] ↑MPBLE [mm] ↓
unseen subjectsPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
S1177.5152.0176.3151.5146.8136.3149.6137.063.865.364.168.628.425.327.125.0
S2182.5164.3179.3156.0154.3150.4150.8146.458.060.660.463.731.528.629.126.6
S3162.3143.2157.2143.6144.1134.2140.7136.764.266.066.370.426.825.225.723.8
S4167.0149.1160.3145.7148.5133.9138.7128.063.769.367.268.927.725.426.024.1
S5142.1126.0142.8127.5121.4109.5125.6104.965.067.370.372.722.220.322.720.4
S6167.2150.4163.7144.2147.7139.3139.1126.862.068.365.367.627.824.326.224.9
S7140.4122.8138.7116.5123.4111.2115.2107.064.867.268.371.022.420.520.918.6
S8133.9112.7130.4111.9118.5106.5110.7106.266.970.270.173.220.818.819.217.2
S9153.6120.0154.6121.3135.7107.1134.8105.462.263.066.370.225.122.825.223.1
Average158.6137.7155.9135.5137.8125.4133.9124.163.466.466.569.625.923.524.722.6
", + "bbox": [ + 78, + 381, + 893, + 522 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4e19ef76752f0fee538e5f2bf101ae13f8fc866c50b1be0cd6fb616742fe049c.jpg", + "table_caption": [ + "Table 5. Average performance comparison (of both SoTA PGBIG [31] and SPGSN [26], on GRAB dataset [45]) of the end predicted pose (1000ms) over samples of each unseen subject $S_{x}$ with $x = [1,2,3,4,5,6,7,8,9]$ , and the average over all unseen subjects." + ], + "table_footnote": [], + "table_body": "
A(E)A(1)ψTTPdynamic learning rateMPJPE [mm] ↓
SPGSN [26]✓✓✓109.5
✓✓✓104.2
✓✓114.3
✓✓✓107.4
✓✓✓✓102.8
", + "bbox": [ + 78, + 565, + 467, + 630 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f75c743ae7d3ecc95fbb4edc84d014a2f2b4365d3f454b5f50ede4099b9f34be.jpg", + "table_caption": [ + "Table 6. Impact of both augmenters $(\\mathcal{A}^{(E)}$ and $\\mathcal{A}^{(I)})$ , TTP and dynamic learning rate on the SPGSN [26] performance." + ], + "table_footnote": [], + "table_body": "
βHKMPJPE [mm] ↓uMPJPE [mm] ↓
SPGSN [26]+H/P-TTP0.000588103.4SPGSN [26]+H/P-TTP1102.8
0.00166109.22103.4
88102.8
1212107.43104.8
1616105.5
0.001588103.1
", + "bbox": [ + 81, + 669, + 462, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 789, + 194, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel helper-predictor test-time personalization (H/P-TTP) framework for 3D human pose forecasting with unseen target subjects. It can be easily integrated into existing deep end-to-end approaches to enable the personalizability of specific target test samples. With the dynamic test-time learning rate, the H/P-TTP fur", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/70e7ef663f8a4e8927a569c95fa7c260cf0dd2f78701cde418e4289cc36b773f.jpg", + "table_caption": [ + "Table 7. Impact of the initial learning rate $\\beta$ ,augmentation numbers (H and K) of both ${\\mathcal{A}}_{\\psi }^{\\left( I\\right) }$ and ${\\mathcal{A}}^{\\left( E\\right) }$ ,and the update numbers $u$ of the TTP process on MPJPE of the SPGSN [26]+H/P-TTP." + ], + "table_footnote": [], + "table_body": "
σ2 [mm]MPJPE [mm] ↓MDMPJPE [mm] ↓
SPGSN [26] +HPTP[5,10,15,20,25,30,35,40]103.1186112.4
9109.5
[10,15,20,25,30,35,40,45]102.8369104.5
12102.8
[15,20,25,30,35,40,45,50]103.54812107.4
18106.8
", + "bbox": [ + 501, + 565, + 888, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Impact of the variance levels of noise injected into $\\mathcal{A}^{(E)}$ , the size of the memory queue and support set.", + "bbox": [ + 500, + 640, + 890, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ther considers the degree to which a specific person's motion pattern needs to be personalized. On several benchmarks, experiments demonstrate that the proposed H/P-TTP significantly boosts their prediction performance. Therefore, we reasonably conclude that our model is more practical for unseen test subjects in real-world applications.", + "bbox": [ + 496, + 684, + 890, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 787, + 666, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China (62176125), in part by the Jiangsu Funding Program for Excellent Postdoctoral Talent (2022ZB269), in part by the Natural Science Foundation of Jiangsu Province (BK20220939), and in part by the China Postdoctoral Science Foundation (2022M721629).", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "281", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peshal Agarwal, Danda Pani Paudel, Jan-Nico Zaech, and Luc Van Gool. Unsupervised Robust Domain Adaptation without Source Data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2009–2018, 2022. 2", + "[2] Sadegh Aliakbarian, Fatemeh Saleh, Lars Petersson, Stephen Gould, and Mathieu Salzmann. Contextually plausible and diverse 3d human motion prediction. In ICCV, pages 11333-11342, 2021. 1, 3", + "[3] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A Stochastic Conditioning Scheme for Diverse Human Motion Prediction. In CVPR, pages 5223-5232, 2020. 1, 2, 3", + "[4] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d Human Pose Estimation: New Benchmark and State of The Art Analysis. In CVPR, pages 3686-3693, 2014. 6", + "[5] Emad Barsoum, John Kender, and Zicheng Liu. HP-GAN: Probabilistic 3D Human Motion Prediction via GAN. In CVPR, pages 1418-1427, 2018. 1, 2", + "[6] Yujun Cai, Lin Huang, Yiwei Wang, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Xu Yang, Yiheng Zhu, Xiaohui Shen, et al. Learning Progressive Joint Propagation for Human Motion Prediction. In ECCV, pages 226-242. Springer, 2020. 6", + "[7] Zhixiang Chi, Yang Wang, Yuanhao Yu, and Jin Tang. Test-Time Fast Adaptation for Dynamic Scene Deblurring via Meta-Auxiliary Learning. In CVPR, pages 9137-9146, 2021. 2", + "[8] Qiongjie Cui and Huaijiang Sun. Towards Accurate 3D Human Motion Prediction From Incomplete Observations. In CVPR, pages 4801-4810, June 2021. 2, 4, 7", + "[9] Qiongjie Cui, Huaijiang Sun, Yupeng Li, and Yue Kong. Efficient human motion recovery using bidirectional attention network. Neural Computing and Applications, 32:10127-10142, 2020. 1", + "[10] Qiongjie Cui, Huaijiang Sun, and Fei Yang. Learning Dynamic Relationships for 3D Human Motion Prediction. In CVPR, pages 6519-6527, 2020. 2, 3, 4, 6, 7", + "[11] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. MSR-GCN: Multi-Scale Residual Graph Convolution Networks for Human Motion Prediction. In ICCV, pages 11467-11476, 2021. 1, 2, 3, 6, 7", + "[12] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Diverse Human Motion Prediction via Gumbel-Softmax Sampling from an Auxiliary Space. In ACM MM, pages 5162–5171, 2022. 3", + "[13] Christian Diller, Thomas A. Funkhouser, and Angela Dai. Forecasting Characteristic 3D Poses of Human Actions. CVPR, pages 15893–15902, 2022. 6", + "[14] Zeyu Feng, Chang Xu, and Dacheng Tao. Self-supervised Representation Learning by Rotation Feature Decoupling. In CVPR, pages 10364–10374, 2019. 4", + "[15] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent Network Models for Human Dynamics. In ICCV, pages 4346-4354, 2015. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Anand Gopalakrishnan, Ankur Mali, Dan Kifer, C. Lee Giles, and Alexander Ororbia. A Neural Temporal Model for Human Motion Prediction. In CVPR, 2019. 1", + "[17] Liang-Yan Gui, Yu-Xiong Wang, Xiaodan Liang, and José M. F. Moura. Adversarial Geometry-Aware Human Motion Prediction. In ECCV, pages 786-803, 2018. 2", + "[18] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Gerard Pons-Moll, and Christian Theobalt. In The Wild Human Pose Estimation using Explicit 2D Features and Intermediate 3D Representations. In CVPR, pages 10905-10914, 2019. 6", + "[19] Miao Hao, Yizhuo Li, Zonglin Di, Nitesh B Gundavarapu, and Xiaolong Wang. Test-Time Personalization with a Transformer for Human Pose Estimation. In NeurIPS, 2021. 1", + "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The Many Faces of Robustness: A Critical Analysis of Out-of-distribution Generalization. In ICCV, pages 8340-8349, 2021. 2", + "[21] Minhao Hu, Tao Song, Yujun Gu, Xiangde Luo, Jieneng Chen, Yinan Chen, Ya Zhang, and Shaoting Zhang. Fully Test-Time Adaptation for Image Segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 251-260. Springer, 2021. 1", + "[22] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large Scale Datasets and Predictive Methods for 3D Human Sensing in Natural Environments. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 36:1325-1339, 2014. 1, 6, 7, 8", + "[23] Nikos Komodakis and Spyros Gidaris. Unsupervised Representation Learning by Predicting Image Rotations. In ICLR, 2018. 4", + "[24] Bin Li, Jian Tian, Zhongfei Zhang, Hailin Feng, and Xi Li. Multitask Non-Autoregressive Model for Human Motion Prediction. IEEE Transactions on Image Processing, 2020. 1, 2, 4", + "[25] Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional Sequence to Sequence Model for Human Dynamics. In CVPR, pages 5226-5234, 2018. 2", + "[26] Maosen Li, Siheng Chen, Zijing Zhang, Lingxi Xie, Qi Tian, and Ya Zhang. Skeleton-Parted Graph Scattering Networks for 3D Human Motion Prediction. arXiv preprint arXiv:2208.00368, 2022. 1, 2, 6, 7, 8", + "[27] Maosen Li, Siheng Chen, Yangheng Zhao, Ya Zhang, Yanfeng Wang, and Qi Tian. Dynamic Multiscale Graph Neural Networks for 3D Skeleton Based Human Motion Prediction. In CVPR, pages 214-223, 2020. 2", + "[28] Yang Li, Kan Li, Shuai Jiang, Ziyue Zhang, Congzhen-tao Huang, and Richard Yi Da Xu. Geometry-driven self-supervised method for 3d human pose estimation. In AAAI, volume 34, pages 11442-11449, 2020. 6", + "[29] Mengyuan Liu, Fanyang Meng, and Yongsheng Liang. Generalized pose decoupled network for unsupervised 3d skeleton sequence-based action representation learning. Cyborg and Bionic Systems, 2022:0002, 2022. 1", + "[30] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "282", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "When Does Self-supervised Test-time Training Fail or Thrive? NeurIPS, 34:21808-21820, 2021. 2", + "[31] Tiezheng Ma, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Progressively Generating Better Initial Guesses Towards Next Stages for High-Quality Human Motion Prediction. In CVPR, pages 6437-6446, 2022. 6, 7, 8", + "[32] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. History Repeats Itself: Human Motion Prediction via Motion Attention. In ECCV, pages 474-489, 2020. 1, 2, 6", + "[33] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Generating Smooth Pose Sequences for Diverse Human Motion Prediction. In ICCV, 2021. 1, 2, 3", + "[34] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning Trajectory Dependencies for Human Motion Prediction. In ICCV, pages 9489-9497, 2019. 2, 3, 6", + "[35] Julieta Martinez, Michael J Black, and Javier Romero. On Human Motion Prediction using Recurrent Neural Networks. In CVPR, pages 2891-2900, 2017. 2, 3, 6", + "[36] Angel Martínez-González, Michael Villamizar, and Jean-Marc Odobez. Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive Transformers. In ICCV, pages 2276-2284, 2021. 2", + "[37] Ishan Misra and Laurens van der Maaten. Self-supervised Learning of Predict-invariant Representations. In CVPR, pages 6707-6717, 2020. 4", + "[38] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating Prediction-time Batch Normalization for Robustness under Covariate Shift. arXiv preprint arXiv:2006.10963, 2020. 2", + "[39] Dario Pavllo, Christoph Feichtenhofer, Michael Auli, and David Grangier. *Quaternion: A Quaternion-based Recurrent Model for Human Motion*. International Journal of Computer Vision, (128):855-872, 2020. 2", + "[40] Senthil Purushwalkam and Abhinav Gupta. Demystifying Contrastive Self-supervised Learning: Invariances, Augmentations and Dataset Biases. NeurIPS, 33:3407-3418, 2020. 4", + "[41] Alejandro Hernandez Ruiz, Juergen Gall, and Francesc Moreno-Noguer. Human Motion Prediction via Spatio-Temporal Inpainting. In CVPR, pages 7134-7143, 2018. 2", + "[42] Chajin Shin, Taeoh Kim, Sangjin Lee, and Sangyoun Leey. Test-Time Adaptation for Out-Of-Distributed Image Inpainting. In ICIP, 2021. 1, 4", + "[43] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International Journal of Computer Vision, 87(1):4-27, 2010. 6", + "[44] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time Training with Self-supervision for Generalization under Distribution Shifts. In ICML, pages 9229-9248. PMLR, 2020. 2", + "[45] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A Dataset of Whole-Body Human Grasping of Objects. In ECCV, 2020. 6, 7, 8" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain Randomization for Transferring Deep Neural Networks from Simulation to The Real World. In IROS, pages 23-30, 2017. 2", + "[47] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In CVPRW, pages 969-977, 2018. 2", + "[48] Thomas Varsavsky, Mauricio Orbes-Arteaga, Carole H Sudre, Mark S Graham, Parashkev Nachev, and M Jorge Cardoso. Test-time Unsupervised Domain Adaptation. In MIC-CAI, pages 428–436. Springer, 2020. 1", + "[49] Borui Wang, Ehsan Adeli, Hsu-kuang Chiu, De-An Huang, and Juan Carlos Niebles. Imitation Learning for Human Pose Prediction. In ICCV, pages 7124-7133, 2019. 1", + "[50] Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, and Trevor Darrell. On-target adaptation. arXiv preprint arXiv:2109.01087, 2021. 1", + "[51] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully Test-time Adaptation By Entropy Minimization. arXiv preprint arXiv:2006.10726, 2020. 2", + "[52] Qin Wang, Dengxin Dai, Lukas Hoyer, Luc Van Gool, and Olga Fink. Domain adaptive semantic segmentation with self-supervised depth estimation. In ICCV, pages 8515-8525, 2021. 1", + "[53] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual Test-time Domain Adaptation. In CVPR, pages 7201-7211, 2022. 2, 5", + "[54] Sirui Xu, Y-X Wang, and L-Y Gui. Diverse Human Motion Prediction Guided by Multi-level Spatial-temporal Anchors. In ECCV, 2022. 2", + "[55] Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, and Yu-Chiang Frank Wang. Adversarial Teacher-student Representation Learning for Domain Generalization. NeurIPS, 34:19448-19460, 2021. 2, 3", + "[56] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A Fourier Perspective on Model Robustness in Computer Vision. NeurIPS, 32, 2019. 2", + "[57] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial Examples: Attacks and Defenses for Deep Learning. IEEE transactions on neural networks and learning systems, 30(9):2805-2824, 2019. 2", + "[58] Ye Yuan and Kris Kitani. Dlow: Diversifying Latent Flows for Diverse Human Motion Prediction. In ECCV, pages 346-364, 2020. 1, 3", + "[59] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test Time Robustness via Adaptation and Augmentation. arXiv preprint arXiv:2110.09506, 2021. 2", + "[60] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Deep Domain-adversarial Image Generation for Domain Generalisation. In AAAI, volume 34, pages 13025-13032, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 854 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "283", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Test-time Personalizable Forecasting of 3D Human Poses/205eafc7-d672-437b-9b70-81a2995fa72d_model.json b/2023/Test-time Personalizable Forecasting of 3D Human Poses/205eafc7-d672-437b-9b70-81a2995fa72d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b38e845fa0762a0a715b272fc297122e6f5c6a8b --- /dev/null +++ b/2023/Test-time Personalizable Forecasting of 3D Human Poses/205eafc7-d672-437b-9b70-81a2995fa72d_model.json @@ -0,0 +1,2827 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.199, + 0.131, + 0.773, + 0.154 + ], + "angle": 0, + "content": "Test-time Personalizable Forecasting of 3D Human Poses" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.18, + 0.832, + 0.291 + ], + "angle": 0, + "content": "Qiongjie Cui\\(^{1}\\) Huaijiang Sun\\(^{1*}\\) \nJianfeng Lu\\(^{1}\\) Weiqing Li\\(^{1}\\) Bin Li\\(^{2}\\) Hongwei Yi\\(^{3}\\), Haofan Wang\\(^{4}\\) \n\\(^{1}\\)Nanjing University of Science and Technology, Nanjing, China \n\\(^{2}\\)Tianjin AiForward Science and Technology Co., Ltd., China \n\\(^{3}\\)Max Planck Institute for Intelligent Systems, Tubingen, Germany \\({ }^{4}\\)Xiaohongshu Inc \ncuiqiongjie@njust.edu.cn sunhuaijiang@njust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.348, + 0.474, + 0.726 + ], + "angle": 0, + "content": "Current motion forecasting approaches typically train a deep end-to-end model from the source domain data, and then apply it directly to target subjects. Despite promising results, they remain non-optimal, due to privacy considerations, the test person and his/her natural properties (e.g., behavioral trait) are typically unseen in training. In this case, the source pre-trained model has a low ability to adapt to these out-of-source characteristics, resulting in an unreliable prediction. To tackle this issue, we propose a novel helper-predictor test-time personalization approach (H/P-TTP), which allows for a generalizable representation of out-of-source subjects to gain more realistic predictions. Concretely, the helper is preceded by explicit and implicit augmenters, where the former yields noisy sequences to improve robustness, while the latter is to generate novel-domain data with an adversarial learning paradigm. Then, the domain-generalizable learning is achieved where the helper can extract cross-subject invariant-knowledge to update the predictor. At test time, given a new person, the predictor is able to be further optimized to empower personalized capabilities to the specific properties. Extensive experiments show that with H/P-TTP, the existing models are significantly improved for various unseen subjects. The project page is available at https://sites.google.com/view/hp-ttp." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.745, + 0.21, + 0.76 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.768, + 0.47, + 0.858 + ], + "angle": 0, + "content": "Forecasting the high-fidelity future human poses, conditioned on a given historical sequence, has attracted increasing attention in recent years. In a variety of 3D vision-based applications, such as autonomous driving and robot navigation, it offers tremendous potential, especially for tasks that call for seamless interaction with humans [11, 9, 29]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.86, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Recent years have witnessed a proliferation in the ex" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.316, + 0.891, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.524, + 0.895, + 0.636 + ], + "angle": 0, + "content": "Figure 1. Starting from off-the-shelf networks, we first train a baseline model [26] on the source data, which is further optimized and personalized to the target person during the test phase. Here, we present an illustration of the pose sequence of various subjects, in which each 3D feature embedding is created by t-SNE. We notice that the motion patterns of test subjects fail to match the source one. Moreover, our prediction skeleton (blue) is closer to the ground truth (red) against the vanilla baseline [26]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.894, + 0.748 + ], + "angle": 0, + "content": "ploration of this emerging field [16, 5, 2]. In particular, ongoing efforts are being devoted to deep learning approaches, which typically train a generic model from large-scale datasets, and then apply it to new test subjects and samples during the deployment phase. We notice that these methods have become mainstream and are now rapidly evolving [49, 58, 24, 32, 3, 33]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Despite notable successes, the existing approaches are sub-optimal, because the generic model cannot be adapted to unseen personalized subjects [48, 21, 52]. To be precise, due to data limitations or privacy concerns, unseen test persons, as well as their individual properties (e.g., age, gender, or behavioral trait), are typically disjointed with the training. It may lead to a large distribution gap between the target and source domain [19, 42, 50], as shown in Figure 1. For example, for the H3.6M dataset [22], a widely-adopted data splitting strategy takes subject-5 (S5) as the test and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "274" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.198 + ], + "angle": 0, + "content": "the rest as the training [10, 34, 25]. Here, the inherent properties (height, and body proportion) of S5 are significantly different from the other subjects; hence, the pre-trained parameters may not be specific or personalized for the S5. We also note that, out-of-source persons are inevitable in the deployment, which poses a major bottleneck to current predictive algorithms [3]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.468, + 0.514 + ], + "angle": 0, + "content": "To solve it, a novel helper-predictor test-time personalization framework (H/P-TTP) is proposed. Starting with ready-to-use networks, our approach can attain tailored parameters according to the characteristics of various unseen subjects. Concretely, the H/P-TTP advances the recent success of knowledge distillation into test-time adaptation by augmenting source instances to out-of-source distributions, enabling domain-generalizable learning. The helper involves an identical architecture with the predictor, except for the explicit and implicit augmenters. To improve the robustness, by adding noise to the samples, the explicit augmenter extracts task-relevant information and ignores irrelevant ones. In contrast, the implicit augmenter is a trainable motion-style transformer to achieve the augmentations with new properties, which follows the adversarial training paradigm, with the goal of maximizing the discrepancy with the source subject, while ensuring semantic proximity. With the max-min optimization, the helper would be able to observe and cope with the novel-styled samples to gain the cross-subject invariant context, and distill the knowledge to update the predictor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.468, + 0.695 + ], + "angle": 0, + "content": "During testing, the model can be further updated to adapt to an unseen test person. We also observe that due to psychological or environmental factors, even for the same person, his/her motion characteristics may vary. More intuitively, for test data with large domain gaps, the predictor needs greater updates, and conversely, a mild personalization is expected. For this purpose, an adaptive learning rate strategy is proposed to dynamically consider the demand for test-time personalization. Owing to the adaptive test-time learning rate, our T/P-TTP can adapt to the intrinsic attributes of an unseen test subject, as well as to his/her specific motion dynamics, as shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.468, + 0.816 + ], + "angle": 0, + "content": "Our contributions are as follows: 1) We notice the issue: the existing approaches are ill-adapted to the out-of-source subjects, and propose a novel H/P-TTP model to address it. 2) We also propose an adaptive learning rate strategy to further consider the extent to which the model needs to be personalized. 3) Our H/P-TTP is integrated into many existing motion forecasting approaches, and substantially boosts the performance on unseen subjects." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.831, + 0.218, + 0.845 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Human Pose Forecasting. Deep end-to-end networks are highly sought-after for solving this issue [5, 41]. In particular, due to the intrinsic temporal nature of 3D human" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "motion, it is typically considered a sequence-to-sequence problem, where various variants of RNNs have been developed [35, 36, 15, 17]. Despite the promising development, RNNs suffer from error accumulation, convergence to an undesired static pose [24, 27, 25]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.17, + 0.892, + 0.275 + ], + "angle": 0, + "content": "To tackle these drawbacks, feed-forward networks have been recently explored as a potential alternative [25, 39, 34, 33]. Particularly, with better interpretability, GCNs are introduced to extract the local semantic connectivity of the 3D human skeleton [34, 10, 8, 11, 24, 27, 36, 26, 54]. Transformers have also been introduced to capture the long-range dependencies of motion sequences [32]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.278, + 0.892, + 0.353 + ], + "angle": 0, + "content": "We have noticed that, current deep learning-based approaches have emerged as the prevailing solution with promising results. However, an obvious challenge still exists, namely, the source pre-trained model cannot be personalized for the specific properties of unseen target persons." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.356, + 0.892, + 0.597 + ], + "angle": 0, + "content": "Domain Generalization (DG). It aims to generalize the learned model from source domain datasets to unseen target domains. Recent advances reveal that data augmentation is an attractive strategy for DG, with the increase in the diversity of training and test data [20, 1, 56, 59]. For this purpose, a variety of works are proposed, e.g., adversarial attack [57, 60], image perturbation [60], and domain randomization [46, 47], where the main heuristic is to learn cross-domain invariant feature. Particularly, [55] introduces a novel-style augmentation in teacher-student frameworks, in which novel synthesized data allows for domain-invariant context to be transferred to the student network. In contrast to the standard DG, which trains a more generalizable model, our work considers going to improve the off-the-shelf network to adapt to the specific characteristics of unseen new subjects and their motion sequence." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.6, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Test-time Adaptation (TTA). Our test-time personalization is relevant to the TTA, which is expected to improve the generalization ability of out-of-distribution test samples. TTA falls into source-free adaptation, and before making decisions, allows for fine-tuning the pre-trained model to adapt to the specific test samples during inference, without accessing the source domain data. For example, according to the test data, prediction-time batch normalization (PTBN) [38] can recalibrate the BN statistics. [51] utilizes test-time entropy minimization (TENT) to update the trainable parameters in BN layers. Test-time training (TTT) [44] adjusts the feature extractor in the test phase via the update of the self-supervised auxiliary task of image rotation prediction. \\(\\mathrm{TTT}++\\) [30] exploits contrastive learning to align the first- and second-order statistics of the test samples with the source domain data, improving the TTT further. [7] introduces an auxiliary image reconstruction to the primary deblurring task, and then resorts to meta-learning to fine-tune the primary branch to adapt to test samples. Considering the changing application environment, [53] integrates" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "275" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.441, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.298, + 0.345, + 0.308 + ], + "angle": 0, + "content": "(a) H/P Domain-Generalizable Learning" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.09, + 0.885, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.297, + 0.745, + 0.308 + ], + "angle": 0, + "content": "(b) H/P Test-time Personalizable Forecasting" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.311, + 0.892, + 0.425 + ], + "angle": 0, + "content": "Figure 2. Illustration of our approach. It mainly includes a parameter-free explicit augmenter \\(\\mathcal{A}^{(E)}\\) and a learnable implicit augmenter \\(\\mathcal{A}_{\\phi}^{(I)}\\), as well as the helper \\(\\mathcal{H}_{\\psi}\\) and predictor \\(\\mathcal{P}_{\\theta}\\). \\(\\mathcal{A}^{(E)}\\) is to attain the noisy samples to improve the robustness and assist \\(\\mathcal{A}_{\\phi}^{(I)}\\) to generate the novel-domain samples. (a) H/P domain-generalizable learning involves two steps: The implicit augmenter is first trained with the adversarial learning paradigm, to make the diverse novel-domain samples, and reserve the semantics; \\(\\mathcal{P}_{\\theta}\\) is trained to minimize the prediction loss \\(\\mathcal{L}_{pred}\\), while \\(\\mathcal{H}_{\\psi}\\) is expected to achieve the consistent output with the predictor to distill the domain-invariant knowledge. We set the learning rate to \\(\\beta\\) to obtain the base model \\(\\{\\stackrel{\\circ}{\\psi}, \\stackrel{\\circ}{\\theta}\\}\\). (b) H/P Test-time Personalizable Forecasting is to further optimize the base helper/predictor to attain the personalized parameters for a new subject. We also analyze the demand of test-time personalization of a specific sequence, which is achieved by dynamically adjusting the test learning rate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.468, + 0.481 + ], + "angle": 0, + "content": "the teacher-student framework into continue test-time adaptation (CoTTA) to distill domain-generalizable knowledge and improve the robustness of the unseen domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.468, + 0.573 + ], + "angle": 0, + "content": "Our inspiration comes partially from the above literature, with the following important distinctions: (1) It can be integrated with any existing deep end-to-end model. (2) Due to the meaningful novel-domain augmentation, high interpretability is provided. (3) It is possible to identify the demand to which the test sample needs to be adapted." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.584, + 0.267, + 0.601 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.468, + 0.727 + ], + "angle": 0, + "content": "Given an observed sequence \\(\\mathbf{x}_{1:T} = [x_1, x_2, \\dots, x_T]\\) of a person across a time horizon \\(T\\), our goal is to predict his/her future actions \\(\\mathbf{y}_{1:\\Delta T} = [y_1, y_2, \\dots, y_{\\Delta T}]\\), in which each frame records the 3D coordinates of \\(N\\) joints. Typically, current approaches [35, 34, 11, 10] learn a generic mapping \\(f: \\mathbf{x}_{1:T} \\rightarrow \\mathbf{y}_{1:\\Delta T}\\). As an extension, our H/P-TTP intends to boost their generalization capability to make personalizable predictions of unseen test subjects/motions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.731, + 0.312, + 0.747 + ], + "angle": 0, + "content": "3.1. Novel Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.468, + 0.84 + ], + "angle": 0, + "content": "Our H/P-TTP consists of a predictor and a helper with an identical structure, parameterized by \\(\\mathcal{P}_{\\theta}\\) and \\(\\mathcal{H}_{\\psi}\\), derived from any existing networks or newly-designed as well. In contrast to the predictor, the helper, outfitted with the implicit and explicit augmenters, w.r.t. \\(\\mathcal{A}_{\\phi}^{(I)}\\) and \\(\\mathcal{A}^{(E)}\\), is dedicated to generating the augmented data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Implicit Augmenter. It is introduced to attain a generalizable feature, which falls into the max-min adversarial learning scope. Intuitively, the implicit augmenter \\(\\mathcal{A}_{\\phi}^{(T)}\\) is expected to generate diverse novel-domain augmentations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.512 + ], + "angle": 0, + "content": "with a variety of attributes (e.g., ages, genders, statures) for the source sequence. After the adversarial learning is finished, the predictor allows for the distillation of domain-generalized representations from the helper, thus improving its personalization ability for unseen subject sequences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.892, + 0.635 + ], + "angle": 0, + "content": "We note that, \\(\\mathcal{A}_{\\phi}^{(I)}\\) can be interpreted as a specially designed motion-style migration network, which is used to implicitly diversify the sequence into its younger/older or taller/lower views. It is associated with the literature on diverse human motion generation [3, 2, 12]. Therefore, we set up \\(\\mathcal{A}_{\\phi}^{(I)}\\) with a similar network as [33], where the significant extraordinary is that it is trained with the max-min optimization and to yield novel-styled counterparts [55]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Given an observed sequence \\(\\mathbf{x}\\), the purpose of \\(\\mathcal{A}_{\\phi}^{(I)}\\) is to produce stylized augmentations \\(\\tilde{\\mathbf{x}}^{(I)} = \\mathcal{A}_{\\phi}^{(I)}(\\mathbf{x})\\) yet retaining its semantic (action label). To stimulate the augmentations with sufficiently distinct appearances, we propose to maximize the discrepancy \\(\\mathcal{L}_{dis}\\) between the original observed motion and \\(H\\) augmented ones \\(\\{\\tilde{\\mathbf{x}}_h^{(I)}\\}_{h=1}^H\\):" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.734, + 0.892, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\mathcal {A} ^ {(I)}} \\mathcal {L} _ {d i s} \\left(\\tilde {\\mathbf {x}} ^ {(I)}, \\mathbf {x}\\right) = \\sum_ {h = 1} ^ {H} \\left\\| \\frac {\\tilde {\\mathbf {x}} _ {h} ^ {(I)}}{\\left\\| \\tilde {\\mathbf {x}} _ {h} ^ {(I)} \\right\\| _ {2}} - \\frac {\\mathbf {x}}{\\left\\| \\mathbf {x} \\right\\| _ {2}} \\right\\| _ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.777, + 0.891, + 0.82 + ], + "angle": 0, + "content": "To motivate diversity across the implicit augmentations, consistent with [58, 33], the diversity-promoting prior loss \\(\\mathcal{L}_{div}\\) with a normalizing factor \\(\\delta = 100\\) is adopted:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.825, + 0.891, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i v} = \\frac {2}{H (H - 1)} \\sum_ {j = 1} ^ {H} \\sum_ {h = j + 1} ^ {H} \\exp \\left(- \\frac {\\left\\| \\tilde {\\mathbf {x}} _ {j} ^ {(I)} - \\tilde {\\mathbf {x}} _ {h} ^ {(I)} \\right\\| _ {1}}{\\delta}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.872, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In addition, to preserve the semantic proximity, we also introduce an action classifier to make the category label of" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "276" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "the novel-styled sequence consistent with the observed one. Given the correct label \\( p \\) in the \\( C \\)-sized label space of a sequence and the resulting one \\( \\tilde{p}_h^{(I)} \\) of \\( h \\)-th implicit augmentation, it is trained with cross-entropy loss \\( \\mathcal{L}_{cls} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.155, + 0.469, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = - \\frac {1}{C H} \\sum_ {h = 1} ^ {H} p \\log \\tilde {p} _ {h} ^ {(I)}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.196, + 0.469, + 0.318 + ], + "angle": 0, + "content": "Then, the implicit augmenter \\(\\mathcal{A}_{\\phi}^{(I)}\\) is optimized to maximize the discrepancy, and minimize the diversity and classification error, which is expected to generate diverse augmentations as much as possible, while maintaining the semantic information with the original one [24]. After observing the new-domain augmentations, the helper can extract the domain-generalizable representation to optimize the predictor to personalize according to input samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.469, + 0.469 + ], + "angle": 0, + "content": "Explicit Augmenter. To incorporate the additional information into the implicit augmenter, we introduce a nontrainable explicit one. In instance learning, each image is regarded as an instance, and we wish to train the network so that the representations of different augmented views of the same instance are as close as possible to each other. It provides more rich data for training, and helps learn transformation invariant representations [37, 40]. For natural images, random rotation, flipping, and noise injection are the most common augmentation strategies [14, 23, 42]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.469, + 0.515 + ], + "angle": 0, + "content": "Besides, [8] develops a multi-task framework, in which the auxiliary branch is designed as motion reconstruction to provide meaningful cues for human pose prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.469, + 0.606 + ], + "angle": 0, + "content": "Motivated by the above-mentioned work, we, therefore, design the explicit augmenter \\( \\mathcal{A}^{(E)} \\) to produce the noisy counterparts. Formally, given a distribution sampled from the observed sequences, \\( \\mathbf{x} \\sim q(\\mathbf{x}) \\), it is injected with several Gaussian noises with a series of variances of a fixed interval, producing \\( K \\) augmentations with different noise levels:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.611, + 0.469, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} \\left(\\tilde {\\mathbf {x}} _ {k} ^ {(E)} \\mid \\mathbf {x}\\right) = \\mathcal {N} \\left(\\tilde {\\mathbf {x}} _ {k} ^ {(E)}; \\mathbf {x}, \\sigma^ {2} \\mathbf {I}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.468, + 0.692 + ], + "angle": 0, + "content": "where \\(\\{\\tilde{\\mathbf{x}}_k^{(E)}\\}_{k = 1}^K\\) is the resulting noisy counterparts, which, together with the novel-domain augmentation \\(\\{\\tilde{\\mathbf{x}}_h^{(I)}\\}_{h = 1}^H\\) from the learnable \\(\\mathcal{A}_{\\phi}^{(I)}\\), is used to train the helper." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.469, + 0.797 + ], + "angle": 0, + "content": "Here, we note that, the non-parametric explicit augmenter provides the following advantages: (1) The noise counterparts are supplied, facilitating the helper to explore new patterns in the novel-domain augmented data. (2) It eliminates an overly optimistic assumption: the trainable implicit augmenter will necessarily yield novel-domain samples. (3) More robustness to noise is also included." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.805, + 0.398, + 0.822 + ], + "angle": 0, + "content": "3.2. H/P Domain-Generalizable Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.824, + 0.469, + 0.903 + ], + "angle": 0, + "content": "The objective of our helper is to improve the generalization of the predictor for unseen domain samples by observing the new-domain augmentations and noise counterparts, as shown in Figure 2 (left). The \\(\\mathcal{A}^{(E)}\\) is parameter-free, while \\(\\mathcal{A}_{\\phi}^{(I)}\\) is trained to solve the following problem:" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.092, + 0.835, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 H/P Domain-Generalizable Learning" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.11, + 0.892, + 0.157 + ], + "angle": 0, + "content": "Require: implicit augmenter \\(\\mathcal{A}_{\\phi}^{(I)}\\); helper \\(\\mathcal{H}_{\\psi}\\) and predictor \\(\\mathcal{P}_{\\theta}\\) with the identical structure; learning rates \\(\\alpha\\), \\(\\beta\\); batch size \\(B\\); hyperparameters \\(\\lambda_{dis}\\), \\(\\lambda_{div}\\), \\(\\lambda_{cls}\\)" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.157, + 0.757, + 0.172 + ], + "angle": 0, + "content": "Output: the learned parameter \\(\\psi, \\theta, \\phi\\);" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.173, + 0.725, + 0.187 + ], + "angle": 0, + "content": "1: randomly initialize \\(\\psi\\), \\(\\theta\\) and \\(\\phi\\);" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.187, + 0.671, + 0.201 + ], + "angle": 0, + "content": "2: while not converge do" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.202, + 0.812, + 0.218 + ], + "angle": 0, + "content": "3: sample a training mini-batch \\(\\{\\mathbf{x}_b,\\mathbf{y}_b\\}_{b = 1}^B\\)" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.218, + 0.891, + 0.233 + ], + "angle": 0, + "content": "4: for each \\(b\\) do \\(\\triangleright\\) update implicit augmenter" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.233, + 0.891, + 0.248 + ], + "angle": 0, + "content": "5: \\(\\phi \\gets \\phi -\\alpha \\nabla_{\\phi}(-\\lambda_{dis}\\mathcal{L}_{dis} + \\lambda_{div}\\mathcal{L}_{div} + \\lambda_{cls}\\mathcal{L}_{cls})\\)" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.248, + 0.59, + 0.261 + ], + "angle": 0, + "content": "6: end for" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.263, + 0.59, + 0.275 + ], + "angle": 0, + "content": "7: end while" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.278, + 0.671, + 0.291 + ], + "angle": 0, + "content": "8: while not converge do" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.292, + 0.811, + 0.323 + ], + "angle": 0, + "content": "9: sample a training mini-batch \\(\\{\\mathbf{x}_b,\\mathbf{y}_b\\}_{b = 1}^B\\) update the helper and predictor" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.323, + 0.812, + 0.341 + ], + "angle": 0, + "content": "10: \\(\\psi \\gets \\psi -\\sum_{b = 1}^{B}\\beta \\nabla_{\\psi}\\mathcal{L}_{pred}(\\tilde{\\bar{\\mathbf{y}}}_b^{(\\mathcal{H})},\\hat{\\mathbf{y}}_b^{(\\mathcal{P})});\\)" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.341, + 0.779, + 0.359 + ], + "angle": 0, + "content": "11: \\(\\theta \\gets \\theta -\\sum_{b = 1}^{B}\\beta \\nabla_{\\theta}\\mathcal{L}_{pred}(\\hat{\\mathbf{y}}_b^{(\\mathcal{P})},\\mathbf{y}_b)\\)" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.359, + 0.599, + 0.371 + ], + "angle": 0, + "content": "12: end while" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.173, + 0.891, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.378, + 0.891, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\phi \\in \\Phi} \\max \\lambda_ {d i s} \\mathcal {L} _ {d i s} + \\lambda_ {d i v} \\mathcal {L} _ {d i v} + \\lambda_ {c l s} \\mathcal {L} _ {c l s}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.406, + 0.892, + 0.455 + ], + "angle": 0, + "content": "where \\(\\lambda_{dis} = 0.5\\), \\(\\lambda_{div} = 0.3\\) and \\(\\lambda_{cls} = 0.2\\) are the trade-off parameters. With a learning rate \\(\\alpha = 0.001\\), the implicit augmenter \\(\\mathcal{A}_{\\phi}^{(I)}\\) is optimized:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.46, + 0.891, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\phi \\leftarrow \\phi - \\alpha \\nabla_ {\\phi} \\left(- \\lambda_ {d i s} \\mathcal {L} _ {d i s} + \\lambda_ {d i v} \\mathcal {L} _ {d i v} + \\lambda_ {c l s} \\mathcal {L} _ {c l s}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.482, + 0.892, + 0.528 + ], + "angle": 0, + "content": "We then train both the helper \\(\\mathcal{H}_{\\psi}\\) and predictor \\(\\mathcal{P}_{\\theta}\\) to achieve the H/P domain-generalizable learning. The optimization of \\(\\mathcal{H}_{\\psi}\\) is achieved using the gradient descent:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.533, + 0.891, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\psi \\leftarrow \\psi - \\beta \\nabla_ {\\psi} \\mathcal {L} _ {p r e d} \\left(\\tilde {\\mathbf {y}} ^ {(\\mathcal {H})}, \\hat {\\mathbf {y}} ^ {(\\mathcal {P})}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.559, + 0.892, + 0.636 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{y}}^{(\\mathcal{P})}\\) is the predictor's result, \\(\\beta = 0.001\\) is the learning rate. The average outcome \\(\\tilde{\\bar{\\xi}}^{(\\mathcal{H})}\\) is yielded over a total of \\(H + K\\) augmentations \\(\\{\\tilde{\\mathbf{x}}_i\\}_{i=1}^{H+K}\\) fed into the helper \\(\\mathcal{H}_{\\psi}\\). Following [10, 8], the prediction loss \\(\\mathcal{L}_{pred}\\) is denoted as \\(L_{2}\\) distance. The update of the predictor \\(\\mathcal{P}_{\\theta}\\) can be achieved:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.638, + 0.891, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\theta \\leftarrow \\theta - \\beta \\nabla_ {\\theta} \\mathcal {L} _ {p r e d} (\\hat {\\mathbf {y}} ^ {(\\mathcal {P})}, \\mathbf {y}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.659, + 0.801, + 0.674 + ], + "angle": 0, + "content": "where \\(\\mathbf{y}\\) is the future ground-truth (gt) action." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.78 + ], + "angle": 0, + "content": "The optimization of Eq. 7 and Eq. 8 ensures that the helper and predictor have consistent outcomes that are as close to the actual future sequence as possible. The domain-generalizable knowledge from the helper is progressively transferred to the predictor. The overall H/P domain-generalizable learning is summarized in Algorithm 1. Then, the base model, w.r.t \\(\\{\\ddot{\\psi},\\dot{\\theta}\\}\\), is achieved." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.788, + 0.852, + 0.805 + ], + "angle": 0, + "content": "3.3. H/P Test-time Personalizable Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.902 + ], + "angle": 0, + "content": "During inference, for a new person, we consider further fine-tuning the base model to adapt to his/her specific properties. However, in practical applications, even for the same subject, his/her motion patterns may differ due to external factors, such as environmental changes. This requires varying degrees of test-time personalization (TTP) to adapt." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "277" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "For this purpose, we propose to adjust the learning rate of test-time personalization. Our intuition is that, for the current step, its previous samples and their correct future sequences are observed deterministically. We are able to know how the association of the helper/predictor feature leads to the difference between the predicted and true values. Therefore, for the next sample, at test-time, calculating the distinction between the features of the helper and predictor, we can also gain insight into the relation of their expected outputs; if it is strong, then the helper can guide the predictor well, and a lower learning rate should be assigned to update the predictor, otherwise, a larger one." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.446 + ], + "angle": 0, + "content": "Memory queue. For simplicity, the base helper \\(\\mathcal{H}_{\\hat{\\psi}}\\) and predictor \\(\\mathcal{P}_{\\hat{\\theta}}\\) are decomposed: feature extractor and generator, w.r.t \\(\\hat{\\psi} = \\{\\hat{\\psi}_1,\\hat{\\psi}_2\\}\\) and \\(\\dot{\\theta} = \\{\\dot{\\theta}_1,\\dot{\\theta}_2\\}\\). We denote \\(h\\) and \\(p\\) as the helper and predictor. The outcome of the feature extractor and generator are denoted as \\(\\mathbf{f}\\) and \\(\\mathbf{o}\\). For a historical sample \\(\\mathbf{x}^{(m)}\\), \\(\\mathbf{f}_{\\{h\\leftrightarrow p\\}}^{(m)}\\) and \\(\\mathbf{o}_{\\{h\\leftrightarrow p\\}}^{(m)}\\) are the feature and outcome differences of \\(h / p\\) branches, calculated by a distance metric \\(\\mathcal{D}\\) (i.e., euclidean distance), informally expressed \\(\\leftrightarrow\\). \\(\\mathbf{o}_{\\{p\\leftrightarrow gt\\}}^{(m)}\\) calculates that the difference between the predictor's result and ground truth, indicating, from real history views, whether the predictor is well enough." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.446, + 0.473, + 0.482 + ], + "angle": 0, + "content": "We build a \\(M\\)-elements memory queue \\(\\mathbb{M} = \\{(\\mathbf{f}_{\\{h \\leftrightarrow p\\}}^{(m)}, \\mathbf{o}_{\\{p \\leftrightarrow gt\\}}^{(m)})\\}_{m=1}^{M}\\), denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.489, + 0.473, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\rightarrow \\underbrace {\\left(\\mathbf {f} _ {\\{h \\leftrightarrow p \\}} ^ {(1)} , \\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(1)}\\right) , \\dots , \\left(\\mathbf {f} _ {\\{h \\leftrightarrow p \\}} ^ {(M)} , \\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(M)}\\right)} _ {\\mathbb {M}}, \\rightarrow \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.473, + 0.634 + ], + "angle": 0, + "content": "It records, in terms of the real history, the relation of the \\( h / p \\) features and the difference between the predicted results and the ground truth (GT). In other words, it measures whether the helper is able to distill domain-generalizable information to the predictor for previous samples (or whether the model has been calibrated well). We note that \\( \\mathbb{M} \\) is updated by First In First Out (FIFO)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.635, + 0.473, + 0.746 + ], + "angle": 0, + "content": "Given a new arriving sample \\(\\mathbf{x}^{(0)}\\), we calculate the feature difference \\(\\mathbf{f}_{\\{h\\leftrightarrow p\\}}^{(0)}\\) and the outcome difference \\(\\mathbf{o}_{\\{h\\leftrightarrow p\\}}^{(0)}\\). To measure its demand to be personalized, \\(\\mathbf{f}_{\\{h\\leftrightarrow p\\}}^{(0)}\\) is taken as the query, and retrieve \\(D\\) most similar ones \\(\\{\\mathbf{f}_{\\{h\\leftrightarrow p\\}}^{(d)}\\}_{d = 1}^D\\) from \\(\\mathbb{M}\\), with their corresponding \\(\\{\\mathbf{o}_{\\{p\\leftrightarrow gt\\}}^{(d)}\\}_{d = 1}^D\\) as the support set \\(\\mathbb{D}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.748, + 0.47, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\underbrace {\\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(1)} , \\dots \\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(d)} , \\dots \\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(D)}} _ {\\mathbb {D}} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.47, + 0.816 + ], + "angle": 0, + "content": "This retrieval process is implemented by \\(D\\)-Nearest Neighbor algorithm." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.817, + 0.471, + 0.859 + ], + "angle": 0, + "content": "Dynamic learning rate. We then average the values in \\(\\mathbb{D}\\) as the reference \\(\\mathbf{r}\\). The prediction discrepancy (PD) of \\(\\mathbf{x}^{(0)}\\) is derived:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.86, + 0.47, + 0.885 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P D} = \\frac {1}{2} \\left(\\mathcal {L} _ {K L} \\left(\\mathbf {r} \\| \\mathbf {o} _ {\\{h \\leftrightarrow p \\}} ^ {(0)}\\right) + \\mathcal {L} _ {K L} \\left(\\mathbf {o} _ {\\{h \\leftrightarrow p \\}} ^ {(0)} \\| \\mathbf {r}\\right)\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.886, + 0.438, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{KL}\\) is the KL divergence of two distributions." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.091, + 0.865, + 0.107 + ], + "angle": 0, + "content": "Algorithm 2 H/P Test-time Personalizable Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.109, + 0.892, + 0.157 + ], + "angle": 0, + "content": "Require: test sample \\(\\mathbf{x}^{(0)}\\); base predictor \\(\\dot{\\theta} = \\{\\ddot{\\theta}_1,\\ddot{\\theta}_2\\}\\); base helper \\(\\dot{\\psi} = \\{\\bar{\\psi}_1,\\bar{\\psi}_2\\}\\); memory queue \\(\\mathbb{M}\\); learning rate \\(\\beta\\); momentum \\(\\mu\\); number of nearest neighbors \\(D\\);" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.157, + 0.692, + 0.172 + ], + "angle": 0, + "content": "Output: final prediction \\(\\hat{\\mathbf{y}}^{*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.172, + 0.892, + 0.204 + ], + "angle": 0, + "content": "1: augment \\(\\mathbf{x}^{(0)}\\) to \\(H + K\\) augmentations \\(\\{\\tilde{\\mathbf{x}}_i^{(0)}\\}_{i = 1}^{H + K}\\) and make \\(\\tilde{\\bar{\\mathbf{x}}}^{(0)}\\) equal to their average;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.204, + 0.892, + 0.24 + ], + "angle": 0, + "content": "2: calculate \\(\\mathbf{f}_{h\\leftrightarrow p}^{(0)} = \\mathcal{D}(\\mathcal{P}_{\\theta_1}(\\mathbf{x}^{(0)}),\\mathcal{H}_{\\psi_1}(\\bar{\\tilde{\\mathbf{x}}}^{(0)}))\\) \\(\\mathbf{o}_{h\\leftrightarrow p}^{(0)} = \\mathcal{D}(\\mathcal{P}_{\\theta}(\\mathbf{x}^{(0)}),\\mathcal{H}_{\\psi}(\\bar{\\tilde{\\mathbf{x}}}^{(0)}));\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.24, + 0.892, + 0.273 + ], + "angle": 0, + "content": "3: retrieve the support set \\(\\mathbb{D}\\) from \\(\\mathbb{M}\\) using \\(D\\)-neighbors nearest with the query \\(\\mathbf{f}_{h\\leftrightarrow p}^{(0)}\\);" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.272, + 0.816, + 0.286 + ], + "angle": 0, + "content": "4: ensemble the values in \\(\\mathbb{D}\\) as the reference \\(\\mathbf{r}\\);" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.287, + 0.892, + 0.302 + ], + "angle": 0, + "content": "5: calculate the prediction discrepancy \\(\\mathcal{L}_{PD}\\) using Eq. 11;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.302, + 0.844, + 0.317 + ], + "angle": 0, + "content": "6: obtain the adjusted learning rate \\(\\gamma\\) using Eq. 12;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.317, + 0.841, + 0.332 + ], + "angle": 0, + "content": "7: update the predictor \\(\\mathcal{P}_{\\theta}:\\theta^{*}\\gets \\dot{\\theta}\\) using Eq. 13;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.332, + 0.836, + 0.347 + ], + "angle": 0, + "content": "8: update the helper \\(\\mathcal{H}_{\\psi}:\\psi^{*}\\gets \\bar{\\psi}\\) , using Eq. 14;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.347, + 0.658, + 0.362 + ], + "angle": 0, + "content": "9: return \\(\\hat{\\mathbf{y}}^{*} = \\mathcal{P}_{\\theta^{*}}(\\mathbf{x})\\)" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.172, + 0.892, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.371, + 0.892, + 0.399 + ], + "angle": 0, + "content": "For the next sample in continuous flow, we have known the actual future sequence \\(\\mathbf{y}^{(0)}\\) of the previous sample \\(\\mathbf{x}^{(0)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.399, + 0.892, + 0.417 + ], + "angle": 0, + "content": "1: \\(\\mathbf{o}_{p\\leftrightarrow qt}^{(0)} = \\mathcal{D}(\\mathcal{P}_{\\theta}(\\mathbf{x}^{(0)}),\\mathbf{y}^{(0)})\\) ; \\(\\triangleright\\) update memory queue" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.417, + 0.763, + 0.438 + ], + "angle": 0, + "content": "2: M.pop() && M.push(\\(\\mathbf{f}_{h \\leftrightarrow p}^{(0)}, \\mathbf{o}_{p \\leftrightarrow gt}^{(0)}\\));" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.399, + 0.892, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.45, + 0.892, + 0.542 + ], + "angle": 0, + "content": "If the value of \\(\\mathcal{L}_{PD}\\) is small, the predictor's outcome coincides with the real history (the helper can guide the predictor well); otherwise, it is not consistent (a major update is needed). Based on this consideration, we adjust the learning rate to make a dynamic TTP. At test time, we obtain the adjusted learning rate \\(\\gamma\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.548, + 0.892, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\gamma = \\beta \\cdot \\mathcal {L} _ {P D}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.569, + 0.892, + 0.645 + ], + "angle": 0, + "content": "Our TTP begins with the base model, \\(\\mathcal{P}_{\\hat{\\theta}}\\) and \\(\\mathcal{H}_{\\hat{\\psi}}\\), obtained from the H/P novel-domain generalizable training stage. Then, we attain the predictor's personalized parameter to boost the generalization ability for the specific subject \\(\\mathbf{x}^{(0)}\\) using a single gradient descent step:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.649, + 0.892, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\stackrel {\\circ} {\\theta} \\leftarrow \\stackrel {\\circ} {\\theta} - \\gamma \\cdot \\nabla_ {\\hat {\\theta}} \\mathcal {L} _ {p r e d} \\left(\\hat {\\mathbf {y}} ^ {(0)}, \\bar {\\tilde {\\mathbf {y}}} ^ {(0)}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.675, + 0.892, + 0.77 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{y}}^{(0)} = \\mathcal{P}_{\\hat{\\theta}}(\\mathbf{x})\\) is the immediate result, and the pseudo label \\(\\bar{\\tilde{\\mathbf{y}}}^{(0)}\\) is the average prediction of the base helper over \\(H + K\\) augmentations. Our model can be regarded as a special teacher-student framework, and therefore, after the predictor's update, we use exponential moving average (EMA) [53] with \\(\\mu = 0.95\\) momentum to optimize the helper:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.772, + 0.892, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\stackrel {\\circ} {\\psi} \\leftarrow \\mu \\cdot \\stackrel {\\circ} {\\psi} + (1 - \\mu) \\cdot \\stackrel {\\circ} {\\theta}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.903 + ], + "angle": 0, + "content": "When the TTP is completed, we re-write the personalized parameters \\(\\{\\hat{\\theta},\\hat{\\psi}\\}\\) as \\(\\{\\theta^{*},\\psi^{*}\\}\\), and perform a forward pass to yield the final prediction \\(\\hat{\\mathbf{y}}^{*} = \\mathcal{P}_{\\theta^{*}}(\\mathbf{x})\\). Note that, after \\(P = 72\\) test-time personalizations, the personalized parameters \\(\\{\\theta^{*},\\psi^{*}\\}\\) are rolled back to the base model \\(\\{\\hat{\\theta},\\hat{\\psi}\\}\\) to avoid catastrophic forgetting. We summarize the H/P test-time personalizable forecasting in Algorithm 2." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "278" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.21, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.113, + 0.295, + 0.129 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.134, + 0.473, + 0.44 + ], + "angle": 0, + "content": "As illustrated in Figure 2, our H/P-TTP begins with an off-the-shelf network, and includes two major phases: (1) H/P novel-domain generalizable training, (2) test-time personalizable forecasting. For the former, we first train the implicit augmenter \\(\\mathcal{A}_{\\phi}^{(I)}\\) using Eq. 6. Both implicit and explicit augmenters yield \\(H = K = 8\\) augmentations. The variances of Gaussian noise injected into the explicit augmenter are set to \\(\\sigma^2 = [10,15,20,25,30,35,40,45]\\) in millimeter. Moreover, the classifier is set to a 3-layers MLP with 512 hidden units to ensure a similar action label between the augmentations and the original sample. Once \\(\\mathcal{A}_{\\phi}^{(I)}\\) is trained, it is fixed in the following process. We exploit Eq. 7 and Eq. 8 to train the predictor \\(\\mathcal{P}_{\\theta}\\) and helper \\(\\mathcal{H}_{\\psi}\\). The predictor/helper are optimized using AdamW optimizer with the weight decay factor (1e-2) on mini-batches of size 32, to achieve the base model. The size of the memory queue \\(\\mathbb{M}\\) is set to \\(M = 36\\), and the support set \\(\\mathbb{D}\\) is \\(D = 12\\). Specifically, we divide the baseline networks by the penultimate layer, where the previous layers are treated as the feature extractor, and it and its following layers are the generator." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.449, + 0.18, + 0.463 + ], + "angle": 0, + "content": "4.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.469, + 0.441, + 0.484 + ], + "angle": 0, + "content": "Our H/P-TTP is evaluated on 3 benchmark datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.47, + 0.563 + ], + "angle": 0, + "content": "Dataset-1: H3.6M [22] is the most widely-used dataset for pose prediction, containing \\(\\approx 3.6\\) million frames of 15 action categories performed by 7 human subjects. Consistent with [34, 6], all sequences are downsampled to 25 fps, and represented as a \\(N = 17\\) joint skeleton." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.67 + ], + "angle": 0, + "content": "Dataset-2: GRAB [45] is a newly-introduced benchmark, including \\(\\approx 1.6\\) million poses of 29 actions from 10 human subjects. Compared with H3.6M, the pose sequences in GRAB are more diverse and involve interaction with the physical world, which is, therefore, more complex and challenging. We down-sample the sequences to 30 fps with 25 joints in each skeleton [13]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Dataset-3: HumanEva-I [43] consists of 3 human subjects, performing 6 pre-defined actions. All sequence is down-sampled to 30 fps, represented by a 15-joint skeleton. The results are reported in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.78 + ], + "angle": 0, + "content": "We note that, for H3.6M [22], the length of the observed and predicted sequence is 25 frames, while for both GRAB [45] and HumanEva-I [43], it is 30." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.79, + 0.275, + 0.807 + ], + "angle": 0, + "content": "4.3. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.812, + 0.471, + 0.906 + ], + "angle": 0, + "content": "We use 2 main setups to analyze our model (1.x and 2.x as stated in Table 1). As in [34, 32, 10, 31, 13], the former is to evaluate the general prediction ability on common data splitting. The latter is newly-designed to evaluate the personalizable ability of unseen subjects, where \\( S_{x} \\) is the test subjects, and the disjointed ones \\( S_{x}^{-} \\)is the training." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.089, + 0.891, + 0.164 + ], + "angle": 0, + "content": "
Purposegeneral predictive ability (1.x)predictive ability for unseen subjects (2.x)
Setup(1.1)(1.2)(2.1)(2.2)(2.3)
DatasetH3.6M [22]GRAB[45]H3.6M[22]GRAB[45]HumanEva-I [43]
TestS5S10SxSxSx
TrainingS-5S-10S-S-S-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.604, + 0.167, + 0.782, + 0.18 + ], + "angle": 0, + "content": "Table 1. Experimental setups." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.181, + 0.609, + 0.196 + ], + "angle": 0, + "content": "4.4. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.397 + ], + "angle": 0, + "content": "We exploit 6 approaches, emerged in recent years as our baselines, including ResSup [35], LTD [34], HisRep [32], MSR [11], PGBIG [31], and SPGSN [26]. They are separately integrated into our H/P-TTP framework to activate the prediction performance for unseen targets. [35] is RNN-based, [32] is attention-based, and the rest are based on GCNs. For the general predictive ability, we use the common data splitting, as the setup-1.x in Table 1, and the setup2.x in Table 1 is used to endow the baselines with the personalizable ability for unseen subjects. We note that our ultimate goal is to equip the H/P-TTP with each baseline to improve the prediction ability, which is therefore renamed, e.g., PGBIG [31] + H/P-TTP, SPGSN [26] + H/P-TTP." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.403, + 0.61, + 0.417 + ], + "angle": 0, + "content": "4.5. Protocols" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.421, + 0.892, + 0.452 + ], + "angle": 0, + "content": "To comprehensively investigate the performance of our H/P-TTP, the following 4 protocols are exploited." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.453, + 0.892, + 0.482 + ], + "angle": 0, + "content": "Protocol-1: The widely-applied mean per-joint position error (MPJPE) [22, 34, 6, 31] is first used:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.485, + 0.892, + 0.514 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathrm {M P J P E}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\| \\boldsymbol {p} _ {n} - \\hat {\\boldsymbol {p}} _ {n} ^ {*} \\| _ {2}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.519, + 0.892, + 0.55 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{p}}_n^*\\) is the 3D position of \\(n\\)-th joint in a final predicted frame \\(\\hat{\\pmb{y}}_t^*\\), \\(\\pmb{p}_n\\) is the gt. \\(N\\) is the number of joints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Protocol-2: We observe that compared with the GT, some predicted pose coordinates have a slight global offset. To eliminate the error independent of poses, we supplement the Procrustes aligned MPJPE (P-MPJPE), which calculates the MPJPE after aligning the predicted pose to the gt pose by a rigid transformation called Procrustes Analysis (PA) [28]. For a frame \\(t\\) in a skeleton layout \\(S\\), it is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.659, + 0.892, + 0.687 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathrm {P - M P J P E}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\| m _ {\\mathrm {P A}, \\mathrm {S}} ^ {t} (n) - m _ {\\mathrm {g t}, \\mathrm {S}} ^ {t} (n) \\| _ {2}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.891, + 0.723 + ], + "angle": 0, + "content": "where \\(m_{\\mathrm{PA,S}}^t (n)\\) is the coordinate of \\(n\\) -th joint after alignment using PA, \\(m_{\\mathrm{glS}}^{t}(n)\\) is the corresponding gt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.724, + 0.892, + 0.828 + ], + "angle": 0, + "content": "Protocol-3: Both MPJPE and P-MPJPE are hard to quantify the performance bound. Therefore, Percentage of Correct 3D Keypoint (PCK) is also introduced, which statistics the proportion of predicted joints with MPJPE errors smaller than a pre-defined threshold. Motivated by the current 3D human pose estimation [4, 18], we use the threshold of \\(150\\mathrm{mm}\\), re-written as PCK@150mm." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.83, + 0.892, + 0.873 + ], + "angle": 0, + "content": "Protocol-4: In addition, we also use the mean per-bone length error (MPBLE) to evaluate the bone-length difference. The MPBLE is defined as: -" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.875, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathrm {M P B L E}} = \\frac {1}{N - 1} \\sum_ {n = 1} ^ {N - 1} \\| l _ {n} - \\hat {l} _ {n} ^ {*} \\| _ {1}, \\tag {17}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "279" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.885, + 0.179 + ], + "angle": 0, + "content": "
MethodResSupResSup +H/P-TTPLTDLTD +H/P-TTPHisRepHisRep +H/P-TTPMSR [11]MSR +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
Time (ms)400100040010004001000400100040010004001000400100040010004001000400100040010004001000
MPJPE [mm] ↓115.2130.5107.4125.663.5114.359.6109.858.3112.156.8109.462.9114.260.8108.658.5110.356.3105.158.3109.655.6103.7
P-MPJPE [mm] ↓96.6111.089.2107.153.1101.547.394.652.6100.146.294.855.7101.351.293.751.297.249.795.651.896.749.490.2
PCK@150mm [%]↑57.550.359.253.770.466.072.569.771.265.573.769.475.570.180.372.077.369.682.471.580.171.285.474.6
MPBLE [mm] ↓33.645.730.543.625.334.023.532.626.432.422.031.127.235.224.733.223.029.319.427.319.325.516.323.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.182, + 0.894, + 0.226 + ], + "angle": 0, + "content": "Table 2. Average short-term (400ms) & long-term (1000ms) prediction performance (on H3.6M [22]) using the common setup-1.1. We observe that, in general, with the H/P-TTP, the baselines achieve better results than the trivial ones. It reveals that \\( S_{5} \\) includes the different motion properties from the other subjects, while our H/T-TTP can adapt them with the personalized parameters." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.234, + 0.479, + 0.325 + ], + "angle": 0, + "content": "
MethodMSRMSR +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
Time (ms)500100050010005001000500100050010005001000
MPJPE ↓96.3178.688.2147.392.2157.284.4141.291.3144.585.5136.4
P-MPJPE ↓82.3131.378.8122.879.8127.566.2114.277.2129.064.9117.3
PCK@150mm↑75.365.678.367.675.866.479.069.177.067.881.170.4
MPBLE ↓21.328.520.227.020.328.019.526.219.226.618.024.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.325, + 0.471, + 0.352 + ], + "angle": 0, + "content": "Table 3. Average short- (500ms) & long-term (1000ms) performance (on GRAB [45]) using the setup-1.2." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.356, + 0.471, + 0.418 + ], + "angle": 0, + "content": "where \\( l_{n} \\) and \\( \\hat{l}_n^* \\) are the bone length of \\( n \\)-th bone in the gt and predicted pose, respectively. \\( N - 1 \\) is the total number of bones in a skeleton. Smaller MPBLEs tend to show a more stable visualization of the predicted frames." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.24, + 0.44 + ], + "angle": 0, + "content": "4.6. Results Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.471, + 0.735 + ], + "angle": 0, + "content": "General prediction ability is first evaluated under the condition of common data splitting [10, 8] with all 6 baselines, as the setups-1.x in Table 1. Specifically, for the H3.6M, Table 2 compares the average result of the 4 protocols over the samples of \\( S_{5} \\) with the experimental setup-1.1. Moreover, under the setup-1.2, Table 3 reports the comparisons on the GRAB dataset [45]. From the results, we observe that, for the used 2 datasets, once equipped with the H/P-TTP, all baselines coincidentally achieve better performance. The reason is that for the common data splitting of H3.6M and GRAB, the test subject (subject-5 or subject-10) is disjointed from the training ones. Stated in a different way, the motion properties of the target subject-5 (S5) are unknown in training (out-of-source), which is not considered by the trivial baselines, thus making the prediction task more challenging. By contrast, with the domain-generalizable learning and test-time personalization, our H/P-TTP can attain the personalized parameters to adapt to the dynamic properties of the new test subject (S5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Prediction ability for unseen subjects is then analyzed using the newly-designed data splitting. Because of data limitations and privacy concerns, the statures, ages, or genders of the target subjects \\( S_{x} \\) in the deployment environment are usually evident in the training stage, resulting in the distribution gap and out-of-source motion properties (e.g., motion style, rhythm). To further evaluate the personalization ability of our H/P-TTP for unseen subjects, we design the setup-2.1&2.2 in Table 1 for H3.6M and GRAB datasets, respectively. Due to their better performance, only both PGBIG and SPGSN are selected as the competitors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.235, + 0.892, + 0.341 + ], + "angle": 0, + "content": "The comparison results are reported in Table 4 and Table 5, in which each subject \\( S_{x} \\) is tested separately and averaged over all its samples. From the results, we observe that, with H/P-TTP, the PGBIG [31] + H/P-TTP and SPGSN [26] + H/P-TTP are calibrated under almost all target subjects. It evidences that the motion properties of out-of-source unseen subjects are indeed adapted." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.892, + 0.387 + ], + "angle": 0, + "content": "We also present the qualitative comparison of the SoTA SPGSN [26], and SPGSN [26] \\(+\\mathrm{H} / \\mathrm{P}\\) -TTP in Figure 3, under camera-takepicture-1 of unseen \\(S_{7}\\) from the GRAB dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.397, + 0.665, + 0.412 + ], + "angle": 0, + "content": "4.7. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.892, + 0.465 + ], + "angle": 0, + "content": "Next, we investigate the impact of several key components. The ablation experiments report the average (at 1000ms) under the setup-2.1 in Table 1 on H3.6M dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.892, + 0.613 + ], + "angle": 0, + "content": "Intuitively, the (1) test-time personalization (TTP) forecasting process can further optimize the base model to adapt to unseen target subjects. It is confirmed in Table 6. We also evaluate the impact of the (2) implicit and explicit augmenter \\((\\mathcal{A}^{(E)}\\) and \\(\\mathcal{A}_{\\phi}^{(I)})\\), and (3) dynamic learning rate. As shown in Table 6, both \\(\\mathcal{A}^{(E)}\\) and \\(\\mathcal{A}_{\\phi}^{(I)}\\) provide the positive effect, while \\(\\mathcal{A}_{\\phi}^{(I)}\\) is greater. Moreover, TTP is indeed helpful, and especially with the dynamic learning rate, the performance is further improved." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.688 + ], + "angle": 0, + "content": "(4) Impact of the initial learning rate \\(\\beta\\) is also investigated in Table 7(left), which is essential to H/P domain-generalizable learning and TTP. We see that \\(\\beta\\) is insensitive to the performance. We suggest that it is because it can be calibrated by the dynamic learning rate at test time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.734 + ], + "angle": 0, + "content": "(5) Impact of the number of augmentations of the implicit and explicit augmenter is shown in Table 7(left), and \\( H = K = 8 \\) achieves the better prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.795 + ], + "angle": 0, + "content": "(6) Number \\((u = [0,1,2,3,4])\\) of gradient descent of TTP is investigated in Table 7(right). We observe that, with the dynamic learning rate, our TTP only needs to perform a single update. More updates bring no further benefits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.84 + ], + "angle": 0, + "content": "(7) Impact of the variances of Gaussian noise injection to the explicit augmenter is shown in Table 8(left), where the middle one attains the best performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "(8) Impact of the sizes of \\(\\mathbb{M}\\) and \\(\\mathbb{D}\\). As reported in Table 8(right), the better result is balanced on the condition of \\(M = 36\\) and \\(D = 12\\). Too small leads to catastrophic forgetting, while too large increases the memory burden." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "280" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.887, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.185, + 0.895, + 0.227 + ], + "angle": 0, + "content": "Figure 3. Predicted pose visualization, where the top is vanilla SPGSN [26] and the bottom is SPGSN \\([26] + \\mathrm{H} / \\mathrm{P}\\) -TTP. The underlying green denotes the gt, and the red/blue is the prediction. The contrasting segments are highlighted in the purple box. We also enlarge some important details in the yellow ellipse. It is clear that with our H/P-TTP, the predicted poses are more accurate." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.238, + 0.895, + 0.348 + ], + "angle": 0, + "content": "
MPJPE [mm] ↓PMPJPE [mm] ↓PCK@150mm [%] ↑MPBLE [mm] ↓
unseen subjectsPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
S1116.6109.3111.4109.697.794.595.491.967.269.367.972.429.625.826.025.2
S6103.698.298.695.284.980.981.878.561.863.964.567.026.222.923.122.3
S7107.0102.1110.4103.888.984.592.986.364.266.364.969.427.423.924.123.3
S896.592.1100.395.478.877.883.474.267.871.968.573.024.020.921.120.3
S9105.3101.2104.5100.187.083.590.386.863.265.363.968.426.623.323.522.7
S11120.8115.5117.0112.8101.597.599.095.570.272.370.975.430.026.726.926.1
Average108.3103.1107.0102.889.886.590.585.565.768.266.470.927.323.924.123.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.348, + 0.892, + 0.377 + ], + "angle": 0, + "content": "Table 4. Average performance comparison (of both SoTA PGBIG [31] and SPGSN [26], on H3.6M dataset [22]) of the end predicted pose (1000ms) over samples of each unseen subject \\( S_{x} \\) with \\( x = [1,6,7,8,9,11] \\), and the corresponding average over all unseen subjects." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.382, + 0.895, + 0.523 + ], + "angle": 0, + "content": "
MPJPE [mm] ↓PMPJPE [mm] ↓PCK@150mm [%] ↑MPBLE [mm] ↓
unseen subjectsPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
S1177.5152.0176.3151.5146.8136.3149.6137.063.865.364.168.628.425.327.125.0
S2182.5164.3179.3156.0154.3150.4150.8146.458.060.660.463.731.528.629.126.6
S3162.3143.2157.2143.6144.1134.2140.7136.764.266.066.370.426.825.225.723.8
S4167.0149.1160.3145.7148.5133.9138.7128.063.769.367.268.927.725.426.024.1
S5142.1126.0142.8127.5121.4109.5125.6104.965.067.370.372.722.220.322.720.4
S6167.2150.4163.7144.2147.7139.3139.1126.862.068.365.367.627.824.326.224.9
S7140.4122.8138.7116.5123.4111.2115.2107.064.867.268.371.022.420.520.918.6
S8133.9112.7130.4111.9118.5106.5110.7106.266.970.270.173.220.818.819.217.2
S9153.6120.0154.6121.3135.7107.1134.8105.462.263.066.370.225.122.825.223.1
Average158.6137.7155.9135.5137.8125.4133.9124.163.466.466.569.625.923.524.722.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.523, + 0.892, + 0.551 + ], + "angle": 0, + "content": "Table 5. Average performance comparison (of both SoTA PGBIG [31] and SPGSN [26], on GRAB dataset [45]) of the end predicted pose (1000ms) over samples of each unseen subject \\( S_{x} \\) with \\( x = [1,2,3,4,5,6,7,8,9] \\), and the average over all unseen subjects." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.566, + 0.468, + 0.631 + ], + "angle": 0, + "content": "
A(E)A(1)ψTTPdynamic learning rateMPJPE [mm] ↓
SPGSN [26]✓✓✓109.5
✓✓✓104.2
✓✓114.3
✓✓✓107.4
✓✓✓✓102.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.632, + 0.469, + 0.661 + ], + "angle": 0, + "content": "Table 6. Impact of both augmenters \\((\\mathcal{A}^{(E)}\\) and \\(\\mathcal{A}^{(I)})\\), TTP and dynamic learning rate on the SPGSN [26] performance." + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.67, + 0.464, + 0.743 + ], + "angle": 0, + "content": "
βHKMPJPE [mm] ↓uMPJPE [mm] ↓
SPGSN [26]+H/P-TTP0.000588103.4SPGSN [26]+H/P-TTP1102.8
0.00166109.22103.4
88102.8
1212107.43104.8
1616105.5
0.001588103.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.743, + 0.468, + 0.786 + ], + "angle": 0, + "content": "Table 7. Impact of the initial learning rate \\( \\beta \\) ,augmentation numbers (H and K) of both \\( {\\mathcal{A}}_{\\psi }^{\\left( I\\right) } \\) and \\( {\\mathcal{A}}^{\\left( E\\right) } \\) ,and the update numbers \\( u \\) of the TTP process on MPJPE of the SPGSN [26]+H/P-TTP." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.79, + 0.196, + 0.807 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a novel helper-predictor test-time personalization (H/P-TTP) framework for 3D human pose forecasting with unseen target subjects. It can be easily integrated into existing deep end-to-end approaches to enable the personalizability of specific target test samples. With the dynamic test-time learning rate, the H/P-TTP fur" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.566, + 0.889, + 0.641 + ], + "angle": 0, + "content": "
σ2 [mm]MPJPE [mm] ↓MDMPJPE [mm] ↓
SPGSN [26] +HPTP[5,10,15,20,25,30,35,40]103.1186112.4
9109.5
[10,15,20,25,30,35,40,45]102.8369104.5
12102.8
[15,20,25,30,35,40,45,50]103.54812107.4
18106.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.641, + 0.892, + 0.671 + ], + "angle": 0, + "content": "Table 8. Impact of the variance levels of noise injected into \\(\\mathcal{A}^{(E)}\\), the size of the memory queue and support set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.777 + ], + "angle": 0, + "content": "ther considers the degree to which a specific person's motion pattern needs to be personalized. On several benchmarks, experiments demonstrate that the proposed H/P-TTP significantly boosts their prediction performance. Therefore, we reasonably conclude that our model is more practical for unseen test subjects in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.788, + 0.667, + 0.805 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China (62176125), in part by the Jiangsu Funding Program for Excellent Postdoctoral Talent (2022ZB269), in part by the Natural Science Foundation of Jiangsu Province (BK20220939), and in part by the China Postdoctoral Science Foundation (2022M721629)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "281" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Peshal Agarwal, Danda Pani Paudel, Jan-Nico Zaech, and Luc Van Gool. Unsupervised Robust Domain Adaptation without Source Data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2009–2018, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Sadegh Aliakbarian, Fatemeh Saleh, Lars Petersson, Stephen Gould, and Mathieu Salzmann. Contextually plausible and diverse 3d human motion prediction. In ICCV, pages 11333-11342, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[3] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A Stochastic Conditioning Scheme for Diverse Human Motion Prediction. In CVPR, pages 5223-5232, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.352 + ], + "angle": 0, + "content": "[4] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d Human Pose Estimation: New Benchmark and State of The Art Analysis. In CVPR, pages 3686-3693, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[5] Emad Barsoum, John Kender, and Zicheng Liu. HP-GAN: Probabilistic 3D Human Motion Prediction via GAN. In CVPR, pages 1418-1427, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.47, + 0.464 + ], + "angle": 0, + "content": "[6] Yujun Cai, Lin Huang, Yiwei Wang, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Xu Yang, Yiheng Zhu, Xiaohui Shen, et al. Learning Progressive Joint Propagation for Human Motion Prediction. In ECCV, pages 226-242. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.47, + 0.52 + ], + "angle": 0, + "content": "[7] Zhixiang Chi, Yang Wang, Yuanhao Yu, and Jin Tang. Test-Time Fast Adaptation for Dynamic Scene Deblurring via Meta-Auxiliary Learning. In CVPR, pages 9137-9146, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[8] Qiongjie Cui and Huaijiang Sun. Towards Accurate 3D Human Motion Prediction From Incomplete Observations. In CVPR, pages 4801-4810, June 2021. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.619 + ], + "angle": 0, + "content": "[9] Qiongjie Cui, Huaijiang Sun, Yupeng Li, and Yue Kong. Efficient human motion recovery using bidirectional attention network. Neural Computing and Applications, 32:10127-10142, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.621, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[10] Qiongjie Cui, Huaijiang Sun, and Fei Yang. Learning Dynamic Relationships for 3D Human Motion Prediction. In CVPR, pages 6519-6527, 2020. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.663, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[11] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. MSR-GCN: Multi-Scale Residual Graph Convolution Networks for Human Motion Prediction. In ICCV, pages 11467-11476, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[12] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Diverse Human Motion Prediction via Gumbel-Softmax Sampling from an Auxiliary Space. In ACM MM, pages 5162–5171, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[13] Christian Diller, Thomas A. Funkhouser, and Angela Dai. Forecasting Characteristic 3D Poses of Human Actions. CVPR, pages 15893–15902, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[14] Zeyu Feng, Chang Xu, and Dacheng Tao. Self-supervised Representation Learning by Rotation Feature Decoupling. In CVPR, pages 10364–10374, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent Network Models for Human Dynamics. In ICCV, pages 4346-4354, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Anand Gopalakrishnan, Ankur Mali, Dan Kifer, C. Lee Giles, and Alexander Ororbia. A Neural Temporal Model for Human Motion Prediction. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[17] Liang-Yan Gui, Yu-Xiong Wang, Xiaodan Liang, and José M. F. Moura. Adversarial Geometry-Aware Human Motion Prediction. In ECCV, pages 786-803, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[18] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Gerard Pons-Moll, and Christian Theobalt. In The Wild Human Pose Estimation using Explicit 2D Features and Intermediate 3D Representations. In CVPR, pages 10905-10914, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[19] Miao Hao, Yizhuo Li, Zonglin Di, Nitesh B Gundavarapu, and Xiaolong Wang. Test-Time Personalization with a Transformer for Human Pose Estimation. In NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The Many Faces of Robustness: A Critical Analysis of Out-of-distribution Generalization. In ICCV, pages 8340-8349, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[21] Minhao Hu, Tao Song, Yujun Gu, Xiangde Luo, Jieneng Chen, Yinan Chen, Ya Zhang, and Shaoting Zhang. Fully Test-Time Adaptation for Image Segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 251-260. Springer, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[22] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large Scale Datasets and Predictive Methods for 3D Human Sensing in Natural Environments. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 36:1325-1339, 2014. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[23] Nikos Komodakis and Spyros Gidaris. Unsupervised Representation Learning by Predicting Image Rotations. In ICLR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.6 + ], + "angle": 0, + "content": "[24] Bin Li, Jian Tian, Zhongfei Zhang, Hailin Feng, and Xi Li. Multitask Non-Autoregressive Model for Human Motion Prediction. IEEE Transactions on Image Processing, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[25] Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional Sequence to Sequence Model for Human Dynamics. In CVPR, pages 5226-5234, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[26] Maosen Li, Siheng Chen, Zijing Zhang, Lingxi Xie, Qi Tian, and Ya Zhang. Skeleton-Parted Graph Scattering Networks for 3D Human Motion Prediction. arXiv preprint arXiv:2208.00368, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[27] Maosen Li, Siheng Chen, Yangheng Zhao, Ya Zhang, Yanfeng Wang, and Qi Tian. Dynamic Multiscale Graph Neural Networks for 3D Skeleton Based Human Motion Prediction. In CVPR, pages 214-223, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[28] Yang Li, Kan Li, Shuai Jiang, Ziyue Zhang, Congzhen-tao Huang, and Richard Yi Da Xu. Geometry-driven self-supervised method for 3d human pose estimation. In AAAI, volume 34, pages 11442-11449, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[29] Mengyuan Liu, Fanyang Meng, and Yongsheng Liang. Generalized pose decoupled network for unsupervised 3d skeleton sequence-based action representation learning. Cyborg and Bionic Systems, 2022:0002, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "282" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "When Does Self-supervised Test-time Training Fail or Thrive? NeurIPS, 34:21808-21820, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[31] Tiezheng Ma, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Progressively Generating Better Initial Guesses Towards Next Stages for High-Quality Human Motion Prediction. In CVPR, pages 6437-6446, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[32] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. History Repeats Itself: Human Motion Prediction via Motion Attention. In ECCV, pages 474-489, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[33] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Generating Smooth Pose Sequences for Diverse Human Motion Prediction. In ICCV, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[34] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning Trajectory Dependencies for Human Motion Prediction. In ICCV, pages 9489-9497, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.469, + 0.362 + ], + "angle": 0, + "content": "[35] Julieta Martinez, Michael J Black, and Javier Romero. On Human Motion Prediction using Recurrent Neural Networks. In CVPR, pages 2891-2900, 2017. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[36] Angel Martínez-González, Michael Villamizar, and Jean-Marc Odobez. Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive Transformers. In ICCV, pages 2276-2284, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.462 + ], + "angle": 0, + "content": "[37] Ishan Misra and Laurens van der Maaten. Self-supervised Learning of Predict-invariant Representations. In CVPR, pages 6707-6717, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[38] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating Prediction-time Batch Normalization for Robustness under Covariate Shift. arXiv preprint arXiv:2006.10963, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[39] Dario Pavllo, Christoph Feichtenhofer, Michael Auli, and David Grangier. *Quaternion: A Quaternion-based Recurrent Model for Human Motion*. International Journal of Computer Vision, (128):855-872, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[40] Senthil Purushwalkam and Abhinav Gupta. Demystifying Contrastive Self-supervised Learning: Invariances, Augmentations and Dataset Biases. NeurIPS, 33:3407-3418, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.688 + ], + "angle": 0, + "content": "[41] Alejandro Hernandez Ruiz, Juergen Gall, and Francesc Moreno-Noguer. Human Motion Prediction via Spatio-Temporal Inpainting. In CVPR, pages 7134-7143, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[42] Chajin Shin, Taeoh Kim, Sangjin Lee, and Sangyoun Leey. Test-Time Adaptation for Out-Of-Distributed Image Inpainting. In ICIP, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[43] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International Journal of Computer Vision, 87(1):4-27, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[44] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time Training with Self-supervision for Generalization under Distribution Shifts. In ICML, pages 9229-9248. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[45] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A Dataset of Whole-Body Human Grasping of Objects. In ECCV, 2020. 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[46] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain Randomization for Transferring Deep Neural Networks from Simulation to The Real World. In IROS, pages 23-30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[47] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In CVPRW, pages 969-977, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[48] Thomas Varsavsky, Mauricio Orbes-Arteaga, Carole H Sudre, Mark S Graham, Parashkev Nachev, and M Jorge Cardoso. Test-time Unsupervised Domain Adaptation. In MIC-CAI, pages 428–436. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[49] Borui Wang, Ehsan Adeli, Hsu-kuang Chiu, De-An Huang, and Juan Carlos Niebles. Imitation Learning for Human Pose Prediction. In ICCV, pages 7124-7133, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[50] Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, and Trevor Darrell. On-target adaptation. arXiv preprint arXiv:2109.01087, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[51] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully Test-time Adaptation By Entropy Minimization. arXiv preprint arXiv:2006.10726, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[52] Qin Wang, Dengxin Dai, Lukas Hoyer, Luc Van Gool, and Olga Fink. Domain adaptive semantic segmentation with self-supervised depth estimation. In ICCV, pages 8515-8525, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.475, + 0.892, + 0.515 + ], + "angle": 0, + "content": "[53] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual Test-time Domain Adaptation. In CVPR, pages 7201-7211, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[54] Sirui Xu, Y-X Wang, and L-Y Gui. Diverse Human Motion Prediction Guided by Multi-level Spatial-temporal Anchors. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[55] Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, and Yu-Chiang Frank Wang. Adversarial Teacher-student Representation Learning for Domain Generalization. NeurIPS, 34:19448-19460, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[56] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A Fourier Perspective on Model Robustness in Computer Vision. NeurIPS, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[57] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial Examples: Attacks and Defenses for Deep Learning. IEEE transactions on neural networks and learning systems, 30(9):2805-2824, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[58] Ye Yuan and Kris Kitani. Dlow: Diversifying Latent Flows for Diverse Human Motion Prediction. In ECCV, pages 346-364, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[59] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test Time Robustness via Adaptation and Augmentation. arXiv preprint arXiv:2110.09506, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[60] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Deep Domain-adversarial Image Generation for Domain Generalisation. In AAAI, volume 34, pages 13025-13032, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.856 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "283" + } + ] +] \ No newline at end of file diff --git a/2023/Test-time Personalizable Forecasting of 3D Human Poses/205eafc7-d672-437b-9b70-81a2995fa72d_origin.pdf b/2023/Test-time Personalizable Forecasting of 3D Human Poses/205eafc7-d672-437b-9b70-81a2995fa72d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..033159a43b97e047f4ee31ba77e77b0c1a524885 --- /dev/null +++ b/2023/Test-time Personalizable Forecasting of 3D Human Poses/205eafc7-d672-437b-9b70-81a2995fa72d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb599c3e8d50a7243093123d06a561b2fa044e4d29348271dddb8a905cb88880 +size 4573129 diff --git a/2023/Test-time Personalizable Forecasting of 3D Human Poses/full.md b/2023/Test-time Personalizable Forecasting of 3D Human Poses/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a7c6d741d146bf4b65beb49d4b80bcdf33dd8857 --- /dev/null +++ b/2023/Test-time Personalizable Forecasting of 3D Human Poses/full.md @@ -0,0 +1,419 @@ +# Test-time Personalizable Forecasting of 3D Human Poses + +Qiongjie Cui $^{1}$ Huaijiang Sun $^{1*}$ +Jianfeng Lu $^{1}$ Weiqing Li $^{1}$ Bin Li $^{2}$ Hongwei Yi $^{3}$ , Haofan Wang $^{4}$ $^{1}$ Nanjing University of Science and Technology, Nanjing, China + $^{2}$ Tianjin AiForward Science and Technology Co., Ltd., China + $^{3}$ Max Planck Institute for Intelligent Systems, Tubingen, Germany ${ }^{4}$ Xiaohongshu Inc +cuiqiongjie@njust.edu.cn sunhuaijiang@njust.edu.cn + +# Abstract + +Current motion forecasting approaches typically train a deep end-to-end model from the source domain data, and then apply it directly to target subjects. Despite promising results, they remain non-optimal, due to privacy considerations, the test person and his/her natural properties (e.g., behavioral trait) are typically unseen in training. In this case, the source pre-trained model has a low ability to adapt to these out-of-source characteristics, resulting in an unreliable prediction. To tackle this issue, we propose a novel helper-predictor test-time personalization approach (H/P-TTP), which allows for a generalizable representation of out-of-source subjects to gain more realistic predictions. Concretely, the helper is preceded by explicit and implicit augmenters, where the former yields noisy sequences to improve robustness, while the latter is to generate novel-domain data with an adversarial learning paradigm. Then, the domain-generalizable learning is achieved where the helper can extract cross-subject invariant-knowledge to update the predictor. At test time, given a new person, the predictor is able to be further optimized to empower personalized capabilities to the specific properties. Extensive experiments show that with H/P-TTP, the existing models are significantly improved for various unseen subjects. The project page is available at https://sites.google.com/view/hp-ttp. + +# 1. Introduction + +Forecasting the high-fidelity future human poses, conditioned on a given historical sequence, has attracted increasing attention in recent years. In a variety of 3D vision-based applications, such as autonomous driving and robot navigation, it offers tremendous potential, especially for tasks that call for seamless interaction with humans [11, 9, 29]. + +Recent years have witnessed a proliferation in the ex + +![](images/1186abb990347563455d35c0471caaede26b0fa002a3c508ce4cf29ec8d7d050.jpg) +Figure 1. Starting from off-the-shelf networks, we first train a baseline model [26] on the source data, which is further optimized and personalized to the target person during the test phase. Here, we present an illustration of the pose sequence of various subjects, in which each 3D feature embedding is created by t-SNE. We notice that the motion patterns of test subjects fail to match the source one. Moreover, our prediction skeleton (blue) is closer to the ground truth (red) against the vanilla baseline [26]. + +ploration of this emerging field [16, 5, 2]. In particular, ongoing efforts are being devoted to deep learning approaches, which typically train a generic model from large-scale datasets, and then apply it to new test subjects and samples during the deployment phase. We notice that these methods have become mainstream and are now rapidly evolving [49, 58, 24, 32, 3, 33]. + +Despite notable successes, the existing approaches are sub-optimal, because the generic model cannot be adapted to unseen personalized subjects [48, 21, 52]. To be precise, due to data limitations or privacy concerns, unseen test persons, as well as their individual properties (e.g., age, gender, or behavioral trait), are typically disjointed with the training. It may lead to a large distribution gap between the target and source domain [19, 42, 50], as shown in Figure 1. For example, for the H3.6M dataset [22], a widely-adopted data splitting strategy takes subject-5 (S5) as the test and + +the rest as the training [10, 34, 25]. Here, the inherent properties (height, and body proportion) of S5 are significantly different from the other subjects; hence, the pre-trained parameters may not be specific or personalized for the S5. We also note that, out-of-source persons are inevitable in the deployment, which poses a major bottleneck to current predictive algorithms [3]. + +To solve it, a novel helper-predictor test-time personalization framework (H/P-TTP) is proposed. Starting with ready-to-use networks, our approach can attain tailored parameters according to the characteristics of various unseen subjects. Concretely, the H/P-TTP advances the recent success of knowledge distillation into test-time adaptation by augmenting source instances to out-of-source distributions, enabling domain-generalizable learning. The helper involves an identical architecture with the predictor, except for the explicit and implicit augmenters. To improve the robustness, by adding noise to the samples, the explicit augmenter extracts task-relevant information and ignores irrelevant ones. In contrast, the implicit augmenter is a trainable motion-style transformer to achieve the augmentations with new properties, which follows the adversarial training paradigm, with the goal of maximizing the discrepancy with the source subject, while ensuring semantic proximity. With the max-min optimization, the helper would be able to observe and cope with the novel-styled samples to gain the cross-subject invariant context, and distill the knowledge to update the predictor. + +During testing, the model can be further updated to adapt to an unseen test person. We also observe that due to psychological or environmental factors, even for the same person, his/her motion characteristics may vary. More intuitively, for test data with large domain gaps, the predictor needs greater updates, and conversely, a mild personalization is expected. For this purpose, an adaptive learning rate strategy is proposed to dynamically consider the demand for test-time personalization. Owing to the adaptive test-time learning rate, our T/P-TTP can adapt to the intrinsic attributes of an unseen test subject, as well as to his/her specific motion dynamics, as shown in Figure 2. + +Our contributions are as follows: 1) We notice the issue: the existing approaches are ill-adapted to the out-of-source subjects, and propose a novel H/P-TTP model to address it. 2) We also propose an adaptive learning rate strategy to further consider the extent to which the model needs to be personalized. 3) Our H/P-TTP is integrated into many existing motion forecasting approaches, and substantially boosts the performance on unseen subjects. + +# 2. Related Work + +Human Pose Forecasting. Deep end-to-end networks are highly sought-after for solving this issue [5, 41]. In particular, due to the intrinsic temporal nature of 3D human + +motion, it is typically considered a sequence-to-sequence problem, where various variants of RNNs have been developed [35, 36, 15, 17]. Despite the promising development, RNNs suffer from error accumulation, convergence to an undesired static pose [24, 27, 25]. + +To tackle these drawbacks, feed-forward networks have been recently explored as a potential alternative [25, 39, 34, 33]. Particularly, with better interpretability, GCNs are introduced to extract the local semantic connectivity of the 3D human skeleton [34, 10, 8, 11, 24, 27, 36, 26, 54]. Transformers have also been introduced to capture the long-range dependencies of motion sequences [32]. + +We have noticed that, current deep learning-based approaches have emerged as the prevailing solution with promising results. However, an obvious challenge still exists, namely, the source pre-trained model cannot be personalized for the specific properties of unseen target persons. + +Domain Generalization (DG). It aims to generalize the learned model from source domain datasets to unseen target domains. Recent advances reveal that data augmentation is an attractive strategy for DG, with the increase in the diversity of training and test data [20, 1, 56, 59]. For this purpose, a variety of works are proposed, e.g., adversarial attack [57, 60], image perturbation [60], and domain randomization [46, 47], where the main heuristic is to learn cross-domain invariant feature. Particularly, [55] introduces a novel-style augmentation in teacher-student frameworks, in which novel synthesized data allows for domain-invariant context to be transferred to the student network. In contrast to the standard DG, which trains a more generalizable model, our work considers going to improve the off-the-shelf network to adapt to the specific characteristics of unseen new subjects and their motion sequence. + +Test-time Adaptation (TTA). Our test-time personalization is relevant to the TTA, which is expected to improve the generalization ability of out-of-distribution test samples. TTA falls into source-free adaptation, and before making decisions, allows for fine-tuning the pre-trained model to adapt to the specific test samples during inference, without accessing the source domain data. For example, according to the test data, prediction-time batch normalization (PTBN) [38] can recalibrate the BN statistics. [51] utilizes test-time entropy minimization (TENT) to update the trainable parameters in BN layers. Test-time training (TTT) [44] adjusts the feature extractor in the test phase via the update of the self-supervised auxiliary task of image rotation prediction. $\mathrm{TTT}++$ [30] exploits contrastive learning to align the first- and second-order statistics of the test samples with the source domain data, improving the TTT further. [7] introduces an auxiliary image reconstruction to the primary deblurring task, and then resorts to meta-learning to fine-tune the primary branch to adapt to test samples. Considering the changing application environment, [53] integrates + +![](images/03f95f96bf887cf52742262d5dfc3dcf1c16863a61ccde499de8c53293750a9d.jpg) +(a) H/P Domain-Generalizable Learning + +![](images/3c8a8c443fe97fc86ac4f0976275b2556c1a40fded8a9072110c42dd1bfc4526.jpg) +(b) H/P Test-time Personalizable Forecasting +Figure 2. Illustration of our approach. It mainly includes a parameter-free explicit augmenter $\mathcal{A}^{(E)}$ and a learnable implicit augmenter $\mathcal{A}_{\phi}^{(I)}$ , as well as the helper $\mathcal{H}_{\psi}$ and predictor $\mathcal{P}_{\theta}$ . $\mathcal{A}^{(E)}$ is to attain the noisy samples to improve the robustness and assist $\mathcal{A}_{\phi}^{(I)}$ to generate the novel-domain samples. (a) H/P domain-generalizable learning involves two steps: The implicit augmenter is first trained with the adversarial learning paradigm, to make the diverse novel-domain samples, and reserve the semantics; $\mathcal{P}_{\theta}$ is trained to minimize the prediction loss $\mathcal{L}_{pred}$ , while $\mathcal{H}_{\psi}$ is expected to achieve the consistent output with the predictor to distill the domain-invariant knowledge. We set the learning rate to $\beta$ to obtain the base model $\{\stackrel{\circ}{\psi}, \stackrel{\circ}{\theta}\}$ . (b) H/P Test-time Personalizable Forecasting is to further optimize the base helper/predictor to attain the personalized parameters for a new subject. We also analyze the demand of test-time personalization of a specific sequence, which is achieved by dynamically adjusting the test learning rate. + +the teacher-student framework into continue test-time adaptation (CoTTA) to distill domain-generalizable knowledge and improve the robustness of the unseen domain. + +Our inspiration comes partially from the above literature, with the following important distinctions: (1) It can be integrated with any existing deep end-to-end model. (2) Due to the meaningful novel-domain augmentation, high interpretability is provided. (3) It is possible to identify the demand to which the test sample needs to be adapted. + +# 3. Proposed Approach + +Given an observed sequence $\mathbf{x}_{1:T} = [x_1, x_2, \dots, x_T]$ of a person across a time horizon $T$ , our goal is to predict his/her future actions $\mathbf{y}_{1:\Delta T} = [y_1, y_2, \dots, y_{\Delta T}]$ , in which each frame records the 3D coordinates of $N$ joints. Typically, current approaches [35, 34, 11, 10] learn a generic mapping $f: \mathbf{x}_{1:T} \rightarrow \mathbf{y}_{1:\Delta T}$ . As an extension, our H/P-TTP intends to boost their generalization capability to make personalizable predictions of unseen test subjects/motions. + +# 3.1. Novel Data Augmentation + +Our H/P-TTP consists of a predictor and a helper with an identical structure, parameterized by $\mathcal{P}_{\theta}$ and $\mathcal{H}_{\psi}$ , derived from any existing networks or newly-designed as well. In contrast to the predictor, the helper, outfitted with the implicit and explicit augmenters, w.r.t. $\mathcal{A}_{\phi}^{(I)}$ and $\mathcal{A}^{(E)}$ , is dedicated to generating the augmented data. + +Implicit Augmenter. It is introduced to attain a generalizable feature, which falls into the max-min adversarial learning scope. Intuitively, the implicit augmenter $\mathcal{A}_{\phi}^{(T)}$ is expected to generate diverse novel-domain augmentations + +with a variety of attributes (e.g., ages, genders, statures) for the source sequence. After the adversarial learning is finished, the predictor allows for the distillation of domain-generalized representations from the helper, thus improving its personalization ability for unseen subject sequences. + +We note that, $\mathcal{A}_{\phi}^{(I)}$ can be interpreted as a specially designed motion-style migration network, which is used to implicitly diversify the sequence into its younger/older or taller/lower views. It is associated with the literature on diverse human motion generation [3, 2, 12]. Therefore, we set up $\mathcal{A}_{\phi}^{(I)}$ with a similar network as [33], where the significant extraordinary is that it is trained with the max-min optimization and to yield novel-styled counterparts [55]. + +Given an observed sequence $\mathbf{x}$ , the purpose of $\mathcal{A}_{\phi}^{(I)}$ is to produce stylized augmentations $\tilde{\mathbf{x}}^{(I)} = \mathcal{A}_{\phi}^{(I)}(\mathbf{x})$ yet retaining its semantic (action label). To stimulate the augmentations with sufficiently distinct appearances, we propose to maximize the discrepancy $\mathcal{L}_{dis}$ between the original observed motion and $H$ augmented ones $\{\tilde{\mathbf{x}}_h^{(I)}\}_{h=1}^H$ : + +$$ +\max _ {\mathcal {A} ^ {(I)}} \mathcal {L} _ {d i s} \left(\tilde {\mathbf {x}} ^ {(I)}, \mathbf {x}\right) = \sum_ {h = 1} ^ {H} \left\| \frac {\tilde {\mathbf {x}} _ {h} ^ {(I)}}{\left\| \tilde {\mathbf {x}} _ {h} ^ {(I)} \right\| _ {2}} - \frac {\mathbf {x}}{\left\| \mathbf {x} \right\| _ {2}} \right\| _ {2}. \tag {1} +$$ + +To motivate diversity across the implicit augmentations, consistent with [58, 33], the diversity-promoting prior loss $\mathcal{L}_{div}$ with a normalizing factor $\delta = 100$ is adopted: + +$$ +\mathcal {L} _ {d i v} = \frac {2}{H (H - 1)} \sum_ {j = 1} ^ {H} \sum_ {h = j + 1} ^ {H} \exp \left(- \frac {\left\| \tilde {\mathbf {x}} _ {j} ^ {(I)} - \tilde {\mathbf {x}} _ {h} ^ {(I)} \right\| _ {1}}{\delta}\right). \tag {2} +$$ + +In addition, to preserve the semantic proximity, we also introduce an action classifier to make the category label of + +the novel-styled sequence consistent with the observed one. Given the correct label $p$ in the $C$ -sized label space of a sequence and the resulting one $\tilde{p}_h^{(I)}$ of $h$ -th implicit augmentation, it is trained with cross-entropy loss $\mathcal{L}_{cls}$ : + +$$ +\mathcal {L} _ {c l s} = - \frac {1}{C H} \sum_ {h = 1} ^ {H} p \log \tilde {p} _ {h} ^ {(I)}. \tag {3} +$$ + +Then, the implicit augmenter $\mathcal{A}_{\phi}^{(I)}$ is optimized to maximize the discrepancy, and minimize the diversity and classification error, which is expected to generate diverse augmentations as much as possible, while maintaining the semantic information with the original one [24]. After observing the new-domain augmentations, the helper can extract the domain-generalizable representation to optimize the predictor to personalize according to input samples. + +Explicit Augmenter. To incorporate the additional information into the implicit augmenter, we introduce a nontrainable explicit one. In instance learning, each image is regarded as an instance, and we wish to train the network so that the representations of different augmented views of the same instance are as close as possible to each other. It provides more rich data for training, and helps learn transformation invariant representations [37, 40]. For natural images, random rotation, flipping, and noise injection are the most common augmentation strategies [14, 23, 42]. + +Besides, [8] develops a multi-task framework, in which the auxiliary branch is designed as motion reconstruction to provide meaningful cues for human pose prediction. + +Motivated by the above-mentioned work, we, therefore, design the explicit augmenter $\mathcal{A}^{(E)}$ to produce the noisy counterparts. Formally, given a distribution sampled from the observed sequences, $\mathbf{x} \sim q(\mathbf{x})$ , it is injected with several Gaussian noises with a series of variances of a fixed interval, producing $K$ augmentations with different noise levels: + +$$ +\boldsymbol {q} \left(\tilde {\mathbf {x}} _ {k} ^ {(E)} \mid \mathbf {x}\right) = \mathcal {N} \left(\tilde {\mathbf {x}} _ {k} ^ {(E)}; \mathbf {x}, \sigma^ {2} \mathbf {I}\right), \tag {4} +$$ + +where $\{\tilde{\mathbf{x}}_k^{(E)}\}_{k = 1}^K$ is the resulting noisy counterparts, which, together with the novel-domain augmentation $\{\tilde{\mathbf{x}}_h^{(I)}\}_{h = 1}^H$ from the learnable $\mathcal{A}_{\phi}^{(I)}$ , is used to train the helper. + +Here, we note that, the non-parametric explicit augmenter provides the following advantages: (1) The noise counterparts are supplied, facilitating the helper to explore new patterns in the novel-domain augmented data. (2) It eliminates an overly optimistic assumption: the trainable implicit augmenter will necessarily yield novel-domain samples. (3) More robustness to noise is also included. + +# 3.2. H/P Domain-Generalizable Learning + +The objective of our helper is to improve the generalization of the predictor for unseen domain samples by observing the new-domain augmentations and noise counterparts, as shown in Figure 2 (left). The $\mathcal{A}^{(E)}$ is parameter-free, while $\mathcal{A}_{\phi}^{(I)}$ is trained to solve the following problem: + +# Algorithm 1 H/P Domain-Generalizable Learning + +Require: implicit augmenter $\mathcal{A}_{\phi}^{(I)}$ ; helper $\mathcal{H}_{\psi}$ and predictor $\mathcal{P}_{\theta}$ with the identical structure; learning rates $\alpha$ , $\beta$ ; batch size $B$ ; hyperparameters $\lambda_{dis}$ , $\lambda_{div}$ , $\lambda_{cls}$ + +Output: the learned parameter $\psi, \theta, \phi$ ; + +1: randomly initialize $\psi$ , $\theta$ and $\phi$ ; +2: while not converge do +3: sample a training mini-batch $\{\mathbf{x}_b,\mathbf{y}_b\}_{b = 1}^B$ +4: for each $b$ do $\triangleright$ update implicit augmenter +5: $\phi \gets \phi -\alpha \nabla_{\phi}(-\lambda_{dis}\mathcal{L}_{dis} + \lambda_{div}\mathcal{L}_{div} + \lambda_{cls}\mathcal{L}_{cls})$ +6: end for +7: end while +8: while not converge do +9: sample a training mini-batch $\{\mathbf{x}_b,\mathbf{y}_b\}_{b = 1}^B$ update the helper and predictor +10: $\psi \gets \psi -\sum_{b = 1}^{B}\beta \nabla_{\psi}\mathcal{L}_{pred}(\tilde{\bar{\mathbf{y}}}_b^{(\mathcal{H})},\hat{\mathbf{y}}_b^{(\mathcal{P})});$ +11: $\theta \gets \theta -\sum_{b = 1}^{B}\beta \nabla_{\theta}\mathcal{L}_{pred}(\hat{\mathbf{y}}_b^{(\mathcal{P})},\mathbf{y}_b)$ +12: end while + +$$ +\arg \min _ {\phi \in \Phi} \max \lambda_ {d i s} \mathcal {L} _ {d i s} + \lambda_ {d i v} \mathcal {L} _ {d i v} + \lambda_ {c l s} \mathcal {L} _ {c l s}, \tag {5} +$$ + +where $\lambda_{dis} = 0.5$ , $\lambda_{div} = 0.3$ and $\lambda_{cls} = 0.2$ are the trade-off parameters. With a learning rate $\alpha = 0.001$ , the implicit augmenter $\mathcal{A}_{\phi}^{(I)}$ is optimized: + +$$ +\phi \leftarrow \phi - \alpha \nabla_ {\phi} \left(- \lambda_ {d i s} \mathcal {L} _ {d i s} + \lambda_ {d i v} \mathcal {L} _ {d i v} + \lambda_ {c l s} \mathcal {L} _ {c l s}\right). \tag {6} +$$ + +We then train both the helper $\mathcal{H}_{\psi}$ and predictor $\mathcal{P}_{\theta}$ to achieve the H/P domain-generalizable learning. The optimization of $\mathcal{H}_{\psi}$ is achieved using the gradient descent: + +$$ +\psi \leftarrow \psi - \beta \nabla_ {\psi} \mathcal {L} _ {p r e d} \left(\tilde {\mathbf {y}} ^ {(\mathcal {H})}, \hat {\mathbf {y}} ^ {(\mathcal {P})}\right), \tag {7} +$$ + +where $\hat{\mathbf{y}}^{(\mathcal{P})}$ is the predictor's result, $\beta = 0.001$ is the learning rate. The average outcome $\tilde{\bar{\xi}}^{(\mathcal{H})}$ is yielded over a total of $H + K$ augmentations $\{\tilde{\mathbf{x}}_i\}_{i=1}^{H+K}$ fed into the helper $\mathcal{H}_{\psi}$ . Following [10, 8], the prediction loss $\mathcal{L}_{pred}$ is denoted as $L_{2}$ distance. The update of the predictor $\mathcal{P}_{\theta}$ can be achieved: + +$$ +\theta \leftarrow \theta - \beta \nabla_ {\theta} \mathcal {L} _ {p r e d} (\hat {\mathbf {y}} ^ {(\mathcal {P})}, \mathbf {y}), \tag {8} +$$ + +where $\mathbf{y}$ is the future ground-truth (gt) action. + +The optimization of Eq. 7 and Eq. 8 ensures that the helper and predictor have consistent outcomes that are as close to the actual future sequence as possible. The domain-generalizable knowledge from the helper is progressively transferred to the predictor. The overall H/P domain-generalizable learning is summarized in Algorithm 1. Then, the base model, w.r.t $\{\ddot{\psi},\dot{\theta}\}$ , is achieved. + +# 3.3. H/P Test-time Personalizable Forecasting + +During inference, for a new person, we consider further fine-tuning the base model to adapt to his/her specific properties. However, in practical applications, even for the same subject, his/her motion patterns may differ due to external factors, such as environmental changes. This requires varying degrees of test-time personalization (TTP) to adapt. + +For this purpose, we propose to adjust the learning rate of test-time personalization. Our intuition is that, for the current step, its previous samples and their correct future sequences are observed deterministically. We are able to know how the association of the helper/predictor feature leads to the difference between the predicted and true values. Therefore, for the next sample, at test-time, calculating the distinction between the features of the helper and predictor, we can also gain insight into the relation of their expected outputs; if it is strong, then the helper can guide the predictor well, and a lower learning rate should be assigned to update the predictor, otherwise, a larger one. + +Memory queue. For simplicity, the base helper $\mathcal{H}_{\hat{\psi}}$ and predictor $\mathcal{P}_{\hat{\theta}}$ are decomposed: feature extractor and generator, w.r.t $\hat{\psi} = \{\hat{\psi}_1,\hat{\psi}_2\}$ and $\dot{\theta} = \{\dot{\theta}_1,\dot{\theta}_2\}$ . We denote $h$ and $p$ as the helper and predictor. The outcome of the feature extractor and generator are denoted as $\mathbf{f}$ and $\mathbf{o}$ . For a historical sample $\mathbf{x}^{(m)}$ , $\mathbf{f}_{\{h\leftrightarrow p\}}^{(m)}$ and $\mathbf{o}_{\{h\leftrightarrow p\}}^{(m)}$ are the feature and outcome differences of $h / p$ branches, calculated by a distance metric $\mathcal{D}$ (i.e., euclidean distance), informally expressed $\leftrightarrow$ . $\mathbf{o}_{\{p\leftrightarrow gt\}}^{(m)}$ calculates that the difference between the predictor's result and ground truth, indicating, from real history views, whether the predictor is well enough. + +We build a $M$ -elements memory queue $\mathbb{M} = \{(\mathbf{f}_{\{h \leftrightarrow p\}}^{(m)}, \mathbf{o}_{\{p \leftrightarrow gt\}}^{(m)})\}_{m=1}^{M}$ , denoted as: + +$$ +\rightarrow \underbrace {\left(\mathbf {f} _ {\{h \leftrightarrow p \}} ^ {(1)} , \mathbf {o} _ {\{p \leftrightarrow g t \}} ^ {(1)}\right) , \dots , \left(\mathbf {f} _ {\{h \leftrightarrow p \}} ^ {(M)} , \mathbf {o} _ {\{p \leftrightarrow g t \}} ^ {(M)}\right)} _ {\mathbb {M}}, \rightarrow \tag {9} +$$ + +It records, in terms of the real history, the relation of the $h / p$ features and the difference between the predicted results and the ground truth (GT). In other words, it measures whether the helper is able to distill domain-generalizable information to the predictor for previous samples (or whether the model has been calibrated well). We note that $\mathbb{M}$ is updated by First In First Out (FIFO). + +Given a new arriving sample $\mathbf{x}^{(0)}$ , we calculate the feature difference $\mathbf{f}_{\{h\leftrightarrow p\}}^{(0)}$ and the outcome difference $\mathbf{o}_{\{h\leftrightarrow p\}}^{(0)}$ . To measure its demand to be personalized, $\mathbf{f}_{\{h\leftrightarrow p\}}^{(0)}$ is taken as the query, and retrieve $D$ most similar ones $\{\mathbf{f}_{\{h\leftrightarrow p\}}^{(d)}\}_{d = 1}^D$ from $\mathbb{M}$ , with their corresponding $\{\mathbf{o}_{\{p\leftrightarrow gt\}}^{(d)}\}_{d = 1}^D$ as the support set $\mathbb{D}$ : + +$$ +\underbrace {\mathbf {o} _ {\{p \leftrightarrow g t \}} ^ {(1)} , \dots \mathbf {o} _ {\{p \leftrightarrow g t \}} ^ {(d)} , \dots \mathbf {o} _ {\{p \leftrightarrow g t \}} ^ {(D)}} _ {\mathbb {D}} \tag {10} +$$ + +This retrieval process is implemented by $D$ -Nearest Neighbor algorithm. + +Dynamic learning rate. We then average the values in $\mathbb{D}$ as the reference $\mathbf{r}$ . The prediction discrepancy (PD) of $\mathbf{x}^{(0)}$ is derived: + +$$ +\mathcal {L} _ {P D} = \frac {1}{2} \left(\mathcal {L} _ {K L} \left(\mathbf {r} \| \mathbf {o} _ {\{h \leftrightarrow p \}} ^ {(0)}\right) + \mathcal {L} _ {K L} \left(\mathbf {o} _ {\{h \leftrightarrow p \}} ^ {(0)} \| \mathbf {r}\right)\right), \tag {11} +$$ + +where $\mathcal{L}_{KL}$ is the KL divergence of two distributions. + +# Algorithm 2 H/P Test-time Personalizable Forecasting + +Require: test sample $\mathbf{x}^{(0)}$ ; base predictor $\dot{\theta} = \{\ddot{\theta}_1,\ddot{\theta}_2\}$ ; base helper $\dot{\psi} = \{\bar{\psi}_1,\bar{\psi}_2\}$ ; memory queue $\mathbb{M}$ ; learning rate $\beta$ ; momentum $\mu$ ; number of nearest neighbors $D$ ; + +Output: final prediction $\hat{\mathbf{y}}^{*}$ + +1: augment $\mathbf{x}^{(0)}$ to $H + K$ augmentations $\{\tilde{\mathbf{x}}_i^{(0)}\}_{i = 1}^{H + K}$ and make $\tilde{\bar{\mathbf{x}}}^{(0)}$ equal to their average; +2: calculate $\mathbf{f}_{h\leftrightarrow p}^{(0)} = \mathcal{D}(\mathcal{P}_{\theta_1}(\mathbf{x}^{(0)}),\mathcal{H}_{\psi_1}(\bar{\tilde{\mathbf{x}}}^{(0)}))$ $\mathbf{o}_{h\leftrightarrow p}^{(0)} = \mathcal{D}(\mathcal{P}_{\theta}(\mathbf{x}^{(0)}),\mathcal{H}_{\psi}(\bar{\tilde{\mathbf{x}}}^{(0)}));$ +3: retrieve the support set $\mathbb{D}$ from $\mathbb{M}$ using $D$ -neighbors nearest with the query $\mathbf{f}_{h\leftrightarrow p}^{(0)}$ ; +4: ensemble the values in $\mathbb{D}$ as the reference $\mathbf{r}$ ; +5: calculate the prediction discrepancy $\mathcal{L}_{PD}$ using Eq. 11; +6: obtain the adjusted learning rate $\gamma$ using Eq. 12; +7: update the predictor $\mathcal{P}_{\theta}:\theta^{*}\gets \dot{\theta}$ using Eq. 13; +8: update the helper $\mathcal{H}_{\psi}:\psi^{*}\gets \bar{\psi}$ , using Eq. 14; +9: return $\hat{\mathbf{y}}^{*} = \mathcal{P}_{\theta^{*}}(\mathbf{x})$ + +For the next sample in continuous flow, we have known the actual future sequence $\mathbf{y}^{(0)}$ of the previous sample $\mathbf{x}^{(0)}$ . + +1: $\mathbf{o}_{p\leftrightarrow qt}^{(0)} = \mathcal{D}(\mathcal{P}_{\theta}(\mathbf{x}^{(0)}),\mathbf{y}^{(0)})$ ; $\triangleright$ update memory queue +2: M.pop() && M.push( $\mathbf{f}_{h \leftrightarrow p}^{(0)}, \mathbf{o}_{p \leftrightarrow gt}^{(0)}$ ); + +If the value of $\mathcal{L}_{PD}$ is small, the predictor's outcome coincides with the real history (the helper can guide the predictor well); otherwise, it is not consistent (a major update is needed). Based on this consideration, we adjust the learning rate to make a dynamic TTP. At test time, we obtain the adjusted learning rate $\gamma$ as: + +$$ +\gamma = \beta \cdot \mathcal {L} _ {P D}. \tag {12} +$$ + +Our TTP begins with the base model, $\mathcal{P}_{\hat{\theta}}$ and $\mathcal{H}_{\hat{\psi}}$ , obtained from the H/P novel-domain generalizable training stage. Then, we attain the predictor's personalized parameter to boost the generalization ability for the specific subject $\mathbf{x}^{(0)}$ using a single gradient descent step: + +$$ +\stackrel {\circ} {\theta} \leftarrow \stackrel {\circ} {\theta} - \gamma \cdot \nabla_ {\hat {\theta}} \mathcal {L} _ {p r e d} \left(\hat {\mathbf {y}} ^ {(0)}, \bar {\tilde {\mathbf {y}}} ^ {(0)}\right), \tag {13} +$$ + +where $\hat{\mathbf{y}}^{(0)} = \mathcal{P}_{\hat{\theta}}(\mathbf{x})$ is the immediate result, and the pseudo label $\bar{\tilde{\mathbf{y}}}^{(0)}$ is the average prediction of the base helper over $H + K$ augmentations. Our model can be regarded as a special teacher-student framework, and therefore, after the predictor's update, we use exponential moving average (EMA) [53] with $\mu = 0.95$ momentum to optimize the helper: + +$$ +\stackrel {\circ} {\psi} \leftarrow \mu \cdot \stackrel {\circ} {\psi} + (1 - \mu) \cdot \stackrel {\circ} {\theta}. \tag {14} +$$ + +When the TTP is completed, we re-write the personalized parameters $\{\hat{\theta},\hat{\psi}\}$ as $\{\theta^{*},\psi^{*}\}$ , and perform a forward pass to yield the final prediction $\hat{\mathbf{y}}^{*} = \mathcal{P}_{\theta^{*}}(\mathbf{x})$ . Note that, after $P = 72$ test-time personalizations, the personalized parameters $\{\theta^{*},\psi^{*}\}$ are rolled back to the base model $\{\hat{\theta},\hat{\psi}\}$ to avoid catastrophic forgetting. We summarize the H/P test-time personalizable forecasting in Algorithm 2. + +# 4. Experiments + +# 4.1. Implementation Details + +As illustrated in Figure 2, our H/P-TTP begins with an off-the-shelf network, and includes two major phases: (1) H/P novel-domain generalizable training, (2) test-time personalizable forecasting. For the former, we first train the implicit augmenter $\mathcal{A}_{\phi}^{(I)}$ using Eq. 6. Both implicit and explicit augmenters yield $H = K = 8$ augmentations. The variances of Gaussian noise injected into the explicit augmenter are set to $\sigma^2 = [10,15,20,25,30,35,40,45]$ in millimeter. Moreover, the classifier is set to a 3-layers MLP with 512 hidden units to ensure a similar action label between the augmentations and the original sample. Once $\mathcal{A}_{\phi}^{(I)}$ is trained, it is fixed in the following process. We exploit Eq. 7 and Eq. 8 to train the predictor $\mathcal{P}_{\theta}$ and helper $\mathcal{H}_{\psi}$ . The predictor/helper are optimized using AdamW optimizer with the weight decay factor (1e-2) on mini-batches of size 32, to achieve the base model. The size of the memory queue $\mathbb{M}$ is set to $M = 36$ , and the support set $\mathbb{D}$ is $D = 12$ . Specifically, we divide the baseline networks by the penultimate layer, where the previous layers are treated as the feature extractor, and it and its following layers are the generator. + +# 4.2. Datasets + +Our H/P-TTP is evaluated on 3 benchmark datasets. + +Dataset-1: H3.6M [22] is the most widely-used dataset for pose prediction, containing $\approx 3.6$ million frames of 15 action categories performed by 7 human subjects. Consistent with [34, 6], all sequences are downsampled to 25 fps, and represented as a $N = 17$ joint skeleton. + +Dataset-2: GRAB [45] is a newly-introduced benchmark, including $\approx 1.6$ million poses of 29 actions from 10 human subjects. Compared with H3.6M, the pose sequences in GRAB are more diverse and involve interaction with the physical world, which is, therefore, more complex and challenging. We down-sample the sequences to 30 fps with 25 joints in each skeleton [13]. + +Dataset-3: HumanEva-I [43] consists of 3 human subjects, performing 6 pre-defined actions. All sequence is down-sampled to 30 fps, represented by a 15-joint skeleton. The results are reported in the supplementary material. + +We note that, for H3.6M [22], the length of the observed and predicted sequence is 25 frames, while for both GRAB [45] and HumanEva-I [43], it is 30. + +# 4.3. Experimental Setups + +We use 2 main setups to analyze our model (1.x and 2.x as stated in Table 1). As in [34, 32, 10, 31, 13], the former is to evaluate the general prediction ability on common data splitting. The latter is newly-designed to evaluate the personalizable ability of unseen subjects, where $S_{x}$ is the test subjects, and the disjointed ones $S_{x}^{-}$ is the training. + +
Purposegeneral predictive ability (1.x)predictive ability for unseen subjects (2.x)
Setup(1.1)(1.2)(2.1)(2.2)(2.3)
DatasetH3.6M [22]GRAB[45]H3.6M[22]GRAB[45]HumanEva-I [43]
TestS5S10SxSxSx
TrainingS-5S-10S-S-S-
+ +Table 1. Experimental setups. + +# 4.4. Baselines + +We exploit 6 approaches, emerged in recent years as our baselines, including ResSup [35], LTD [34], HisRep [32], MSR [11], PGBIG [31], and SPGSN [26]. They are separately integrated into our H/P-TTP framework to activate the prediction performance for unseen targets. [35] is RNN-based, [32] is attention-based, and the rest are based on GCNs. For the general predictive ability, we use the common data splitting, as the setup-1.x in Table 1, and the setup2.x in Table 1 is used to endow the baselines with the personalizable ability for unseen subjects. We note that our ultimate goal is to equip the H/P-TTP with each baseline to improve the prediction ability, which is therefore renamed, e.g., PGBIG [31] + H/P-TTP, SPGSN [26] + H/P-TTP. + +# 4.5. Protocols + +To comprehensively investigate the performance of our H/P-TTP, the following 4 protocols are exploited. + +Protocol-1: The widely-applied mean per-joint position error (MPJPE) [22, 34, 6, 31] is first used: + +$$ +E _ {\mathrm {M P J P E}} = \frac {1}{N} \sum_ {n = 1} ^ {N} \| \boldsymbol {p} _ {n} - \hat {\boldsymbol {p}} _ {n} ^ {*} \| _ {2}, \tag {15} +$$ + +where $\hat{\pmb{p}}_n^*$ is the 3D position of $n$ -th joint in a final predicted frame $\hat{\pmb{y}}_t^*$ , $\pmb{p}_n$ is the gt. $N$ is the number of joints. + +Protocol-2: We observe that compared with the GT, some predicted pose coordinates have a slight global offset. To eliminate the error independent of poses, we supplement the Procrustes aligned MPJPE (P-MPJPE), which calculates the MPJPE after aligning the predicted pose to the gt pose by a rigid transformation called Procrustes Analysis (PA) [28]. For a frame $t$ in a skeleton layout $S$ , it is defined as: + +$$ +E _ {\mathrm {P - M P J P E}} = \frac {1}{N} \sum_ {n = 1} ^ {N} \| m _ {\mathrm {P A}, \mathrm {S}} ^ {t} (n) - m _ {\mathrm {g t}, \mathrm {S}} ^ {t} (n) \| _ {2}, \tag {16} +$$ + +where $m_{\mathrm{PA,S}}^t (n)$ is the coordinate of $n$ -th joint after alignment using PA, $m_{\mathrm{glS}}^{t}(n)$ is the corresponding gt. + +Protocol-3: Both MPJPE and P-MPJPE are hard to quantify the performance bound. Therefore, Percentage of Correct 3D Keypoint (PCK) is also introduced, which statistics the proportion of predicted joints with MPJPE errors smaller than a pre-defined threshold. Motivated by the current 3D human pose estimation [4, 18], we use the threshold of $150\mathrm{mm}$ , re-written as PCK@150mm. + +Protocol-4: In addition, we also use the mean per-bone length error (MPBLE) to evaluate the bone-length difference. The MPBLE is defined as: - + +$$ +E _ {\mathrm {M P B L E}} = \frac {1}{N - 1} \sum_ {n = 1} ^ {N - 1} \| l _ {n} - \hat {l} _ {n} ^ {*} \| _ {1}, \tag {17} +$$ + +
MethodResSupResSup +H/P-TTPLTDLTD +H/P-TTPHisRepHisRep +H/P-TTPMSR [11]MSR +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
Time (ms)400100040010004001000400100040010004001000400100040010004001000400100040010004001000
MPJPE [mm] ↓115.2130.5107.4125.663.5114.359.6109.858.3112.156.8109.462.9114.260.8108.658.5110.356.3105.158.3109.655.6103.7
P-MPJPE [mm] ↓96.6111.089.2107.153.1101.547.394.652.6100.146.294.855.7101.351.293.751.297.249.795.651.896.749.490.2
PCK@150mm [%]↑57.550.359.253.770.466.072.569.771.265.573.769.475.570.180.372.077.369.682.471.580.171.285.474.6
MPBLE [mm] ↓33.645.730.543.625.334.023.532.626.432.422.031.127.235.224.733.223.029.319.427.319.325.516.323.8
+ +Table 2. Average short-term (400ms) & long-term (1000ms) prediction performance (on H3.6M [22]) using the common setup-1.1. We observe that, in general, with the H/P-TTP, the baselines achieve better results than the trivial ones. It reveals that $S_{5}$ includes the different motion properties from the other subjects, while our H/T-TTP can adapt them with the personalized parameters. + +
MethodMSRMSR +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
Time (ms)500100050010005001000500100050010005001000
MPJPE ↓96.3178.688.2147.392.2157.284.4141.291.3144.585.5136.4
P-MPJPE ↓82.3131.378.8122.879.8127.566.2114.277.2129.064.9117.3
PCK@150mm↑75.365.678.367.675.866.479.069.177.067.881.170.4
MPBLE ↓21.328.520.227.020.328.019.526.219.226.618.024.7
+ +Table 3. Average short- (500ms) & long-term (1000ms) performance (on GRAB [45]) using the setup-1.2. + +where $l_{n}$ and $\hat{l}_n^*$ are the bone length of $n$ -th bone in the gt and predicted pose, respectively. $N - 1$ is the total number of bones in a skeleton. Smaller MPBLEs tend to show a more stable visualization of the predicted frames. + +# 4.6. Results Analysis + +General prediction ability is first evaluated under the condition of common data splitting [10, 8] with all 6 baselines, as the setups-1.x in Table 1. Specifically, for the H3.6M, Table 2 compares the average result of the 4 protocols over the samples of $S_{5}$ with the experimental setup-1.1. Moreover, under the setup-1.2, Table 3 reports the comparisons on the GRAB dataset [45]. From the results, we observe that, for the used 2 datasets, once equipped with the H/P-TTP, all baselines coincidentally achieve better performance. The reason is that for the common data splitting of H3.6M and GRAB, the test subject (subject-5 or subject-10) is disjointed from the training ones. Stated in a different way, the motion properties of the target subject-5 (S5) are unknown in training (out-of-source), which is not considered by the trivial baselines, thus making the prediction task more challenging. By contrast, with the domain-generalizable learning and test-time personalization, our H/P-TTP can attain the personalized parameters to adapt to the dynamic properties of the new test subject (S5). + +Prediction ability for unseen subjects is then analyzed using the newly-designed data splitting. Because of data limitations and privacy concerns, the statures, ages, or genders of the target subjects $S_{x}$ in the deployment environment are usually evident in the training stage, resulting in the distribution gap and out-of-source motion properties (e.g., motion style, rhythm). To further evaluate the personalization ability of our H/P-TTP for unseen subjects, we design the setup-2.1&2.2 in Table 1 for H3.6M and GRAB datasets, respectively. Due to their better performance, only both PGBIG and SPGSN are selected as the competitors. + +The comparison results are reported in Table 4 and Table 5, in which each subject $S_{x}$ is tested separately and averaged over all its samples. From the results, we observe that, with H/P-TTP, the PGBIG [31] + H/P-TTP and SPGSN [26] + H/P-TTP are calibrated under almost all target subjects. It evidences that the motion properties of out-of-source unseen subjects are indeed adapted. + +We also present the qualitative comparison of the SoTA SPGSN [26], and SPGSN [26] $+\mathrm{H} / \mathrm{P}$ -TTP in Figure 3, under camera-takepicture-1 of unseen $S_{7}$ from the GRAB dataset. + +# 4.7. Ablation Studies + +Next, we investigate the impact of several key components. The ablation experiments report the average (at 1000ms) under the setup-2.1 in Table 1 on H3.6M dataset. + +Intuitively, the (1) test-time personalization (TTP) forecasting process can further optimize the base model to adapt to unseen target subjects. It is confirmed in Table 6. We also evaluate the impact of the (2) implicit and explicit augmenter $(\mathcal{A}^{(E)}$ and $\mathcal{A}_{\phi}^{(I)})$ , and (3) dynamic learning rate. As shown in Table 6, both $\mathcal{A}^{(E)}$ and $\mathcal{A}_{\phi}^{(I)}$ provide the positive effect, while $\mathcal{A}_{\phi}^{(I)}$ is greater. Moreover, TTP is indeed helpful, and especially with the dynamic learning rate, the performance is further improved. + +(4) Impact of the initial learning rate $\beta$ is also investigated in Table 7(left), which is essential to H/P domain-generalizable learning and TTP. We see that $\beta$ is insensitive to the performance. We suggest that it is because it can be calibrated by the dynamic learning rate at test time. +(5) Impact of the number of augmentations of the implicit and explicit augmenter is shown in Table 7(left), and $H = K = 8$ achieves the better prediction. +(6) Number $(u = [0,1,2,3,4])$ of gradient descent of TTP is investigated in Table 7(right). We observe that, with the dynamic learning rate, our TTP only needs to perform a single update. More updates bring no further benefits. +(7) Impact of the variances of Gaussian noise injection to the explicit augmenter is shown in Table 8(left), where the middle one attains the best performance. +(8) Impact of the sizes of $\mathbb{M}$ and $\mathbb{D}$ . As reported in Table 8(right), the better result is balanced on the condition of $M = 36$ and $D = 12$ . Too small leads to catastrophic forgetting, while too large increases the memory burden. + +![](images/9971ce0c05e1bc58e98975d67bb355e7c36325a64c194e726b8dd223b8385a33.jpg) +Figure 3. Predicted pose visualization, where the top is vanilla SPGSN [26] and the bottom is SPGSN $[26] + \mathrm{H} / \mathrm{P}$ -TTP. The underlying green denotes the gt, and the red/blue is the prediction. The contrasting segments are highlighted in the purple box. We also enlarge some important details in the yellow ellipse. It is clear that with our H/P-TTP, the predicted poses are more accurate. + +
MPJPE [mm] ↓PMPJPE [mm] ↓PCK@150mm [%] ↑MPBLE [mm] ↓
unseen subjectsPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
S1116.6109.3111.4109.697.794.595.491.967.269.367.972.429.625.826.025.2
S6103.698.298.695.284.980.981.878.561.863.964.567.026.222.923.122.3
S7107.0102.1110.4103.888.984.592.986.364.266.364.969.427.423.924.123.3
S896.592.1100.395.478.877.883.474.267.871.968.573.024.020.921.120.3
S9105.3101.2104.5100.187.083.590.386.863.265.363.968.426.623.323.522.7
S11120.8115.5117.0112.8101.597.599.095.570.272.370.975.430.026.726.926.1
Average108.3103.1107.0102.889.886.590.585.565.768.266.470.927.323.924.123.3
+ +Table 4. Average performance comparison (of both SoTA PGBIG [31] and SPGSN [26], on H3.6M dataset [22]) of the end predicted pose (1000ms) over samples of each unseen subject $S_{x}$ with $x = [1,6,7,8,9,11]$ , and the corresponding average over all unseen subjects. + +
MPJPE [mm] ↓PMPJPE [mm] ↓PCK@150mm [%] ↑MPBLE [mm] ↓
unseen subjectsPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
S1177.5152.0176.3151.5146.8136.3149.6137.063.865.364.168.628.425.327.125.0
S2182.5164.3179.3156.0154.3150.4150.8146.458.060.660.463.731.528.629.126.6
S3162.3143.2157.2143.6144.1134.2140.7136.764.266.066.370.426.825.225.723.8
S4167.0149.1160.3145.7148.5133.9138.7128.063.769.367.268.927.725.426.024.1
S5142.1126.0142.8127.5121.4109.5125.6104.965.067.370.372.722.220.322.720.4
S6167.2150.4163.7144.2147.7139.3139.1126.862.068.365.367.627.824.326.224.9
S7140.4122.8138.7116.5123.4111.2115.2107.064.867.268.371.022.420.520.918.6
S8133.9112.7130.4111.9118.5106.5110.7106.266.970.270.173.220.818.819.217.2
S9153.6120.0154.6121.3135.7107.1134.8105.462.263.066.370.225.122.825.223.1
Average158.6137.7155.9135.5137.8125.4133.9124.163.466.466.569.625.923.524.722.6
+ +Table 5. Average performance comparison (of both SoTA PGBIG [31] and SPGSN [26], on GRAB dataset [45]) of the end predicted pose (1000ms) over samples of each unseen subject $S_{x}$ with $x = [1,2,3,4,5,6,7,8,9]$ , and the average over all unseen subjects. + +
A(E)A(1)ψTTPdynamic learning rateMPJPE [mm] ↓
SPGSN [26]✓✓✓109.5
✓✓✓104.2
✓✓114.3
✓✓✓107.4
✓✓✓✓102.8
+ +Table 6. Impact of both augmenters $(\mathcal{A}^{(E)}$ and $\mathcal{A}^{(I)})$ , TTP and dynamic learning rate on the SPGSN [26] performance. + +
βHKMPJPE [mm] ↓uMPJPE [mm] ↓
SPGSN [26]+H/P-TTP0.000588103.4SPGSN [26]+H/P-TTP1102.8
0.00166109.22103.4
88102.8
1212107.43104.8
1616105.5
0.001588103.1
+ +# 5. Conclusion + +In this work, we propose a novel helper-predictor test-time personalization (H/P-TTP) framework for 3D human pose forecasting with unseen target subjects. It can be easily integrated into existing deep end-to-end approaches to enable the personalizability of specific target test samples. With the dynamic test-time learning rate, the H/P-TTP fur + +Table 7. Impact of the initial learning rate $\beta$ ,augmentation numbers (H and K) of both ${\mathcal{A}}_{\psi }^{\left( I\right) }$ and ${\mathcal{A}}^{\left( E\right) }$ ,and the update numbers $u$ of the TTP process on MPJPE of the SPGSN [26]+H/P-TTP. + +
σ2 [mm]MPJPE [mm] ↓MDMPJPE [mm] ↓
SPGSN [26] +HPTP[5,10,15,20,25,30,35,40]103.1186112.4
9109.5
[10,15,20,25,30,35,40,45]102.8369104.5
12102.8
[15,20,25,30,35,40,45,50]103.54812107.4
18106.8
+ +Table 8. Impact of the variance levels of noise injected into $\mathcal{A}^{(E)}$ , the size of the memory queue and support set. + +ther considers the degree to which a specific person's motion pattern needs to be personalized. On several benchmarks, experiments demonstrate that the proposed H/P-TTP significantly boosts their prediction performance. Therefore, we reasonably conclude that our model is more practical for unseen test subjects in real-world applications. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China (62176125), in part by the Jiangsu Funding Program for Excellent Postdoctoral Talent (2022ZB269), in part by the Natural Science Foundation of Jiangsu Province (BK20220939), and in part by the China Postdoctoral Science Foundation (2022M721629). + +# References + +[1] Peshal Agarwal, Danda Pani Paudel, Jan-Nico Zaech, and Luc Van Gool. Unsupervised Robust Domain Adaptation without Source Data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2009–2018, 2022. 2 +[2] Sadegh Aliakbarian, Fatemeh Saleh, Lars Petersson, Stephen Gould, and Mathieu Salzmann. Contextually plausible and diverse 3d human motion prediction. In ICCV, pages 11333-11342, 2021. 1, 3 +[3] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A Stochastic Conditioning Scheme for Diverse Human Motion Prediction. In CVPR, pages 5223-5232, 2020. 1, 2, 3 +[4] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d Human Pose Estimation: New Benchmark and State of The Art Analysis. In CVPR, pages 3686-3693, 2014. 6 +[5] Emad Barsoum, John Kender, and Zicheng Liu. HP-GAN: Probabilistic 3D Human Motion Prediction via GAN. In CVPR, pages 1418-1427, 2018. 1, 2 +[6] Yujun Cai, Lin Huang, Yiwei Wang, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Xu Yang, Yiheng Zhu, Xiaohui Shen, et al. Learning Progressive Joint Propagation for Human Motion Prediction. In ECCV, pages 226-242. Springer, 2020. 6 +[7] Zhixiang Chi, Yang Wang, Yuanhao Yu, and Jin Tang. Test-Time Fast Adaptation for Dynamic Scene Deblurring via Meta-Auxiliary Learning. In CVPR, pages 9137-9146, 2021. 2 +[8] Qiongjie Cui and Huaijiang Sun. Towards Accurate 3D Human Motion Prediction From Incomplete Observations. In CVPR, pages 4801-4810, June 2021. 2, 4, 7 +[9] Qiongjie Cui, Huaijiang Sun, Yupeng Li, and Yue Kong. Efficient human motion recovery using bidirectional attention network. Neural Computing and Applications, 32:10127-10142, 2020. 1 +[10] Qiongjie Cui, Huaijiang Sun, and Fei Yang. Learning Dynamic Relationships for 3D Human Motion Prediction. In CVPR, pages 6519-6527, 2020. 2, 3, 4, 6, 7 +[11] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. MSR-GCN: Multi-Scale Residual Graph Convolution Networks for Human Motion Prediction. In ICCV, pages 11467-11476, 2021. 1, 2, 3, 6, 7 +[12] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Diverse Human Motion Prediction via Gumbel-Softmax Sampling from an Auxiliary Space. In ACM MM, pages 5162–5171, 2022. 3 +[13] Christian Diller, Thomas A. Funkhouser, and Angela Dai. Forecasting Characteristic 3D Poses of Human Actions. CVPR, pages 15893–15902, 2022. 6 +[14] Zeyu Feng, Chang Xu, and Dacheng Tao. Self-supervised Representation Learning by Rotation Feature Decoupling. In CVPR, pages 10364–10374, 2019. 4 +[15] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent Network Models for Human Dynamics. In ICCV, pages 4346-4354, 2015. 2 + +[16] Anand Gopalakrishnan, Ankur Mali, Dan Kifer, C. Lee Giles, and Alexander Ororbia. A Neural Temporal Model for Human Motion Prediction. In CVPR, 2019. 1 +[17] Liang-Yan Gui, Yu-Xiong Wang, Xiaodan Liang, and José M. F. Moura. Adversarial Geometry-Aware Human Motion Prediction. In ECCV, pages 786-803, 2018. 2 +[18] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Gerard Pons-Moll, and Christian Theobalt. In The Wild Human Pose Estimation using Explicit 2D Features and Intermediate 3D Representations. In CVPR, pages 10905-10914, 2019. 6 +[19] Miao Hao, Yizhuo Li, Zonglin Di, Nitesh B Gundavarapu, and Xiaolong Wang. Test-Time Personalization with a Transformer for Human Pose Estimation. In NeurIPS, 2021. 1 +[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The Many Faces of Robustness: A Critical Analysis of Out-of-distribution Generalization. In ICCV, pages 8340-8349, 2021. 2 +[21] Minhao Hu, Tao Song, Yujun Gu, Xiangde Luo, Jieneng Chen, Yinan Chen, Ya Zhang, and Shaoting Zhang. Fully Test-Time Adaptation for Image Segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 251-260. Springer, 2021. 1 +[22] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large Scale Datasets and Predictive Methods for 3D Human Sensing in Natural Environments. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 36:1325-1339, 2014. 1, 6, 7, 8 +[23] Nikos Komodakis and Spyros Gidaris. Unsupervised Representation Learning by Predicting Image Rotations. In ICLR, 2018. 4 +[24] Bin Li, Jian Tian, Zhongfei Zhang, Hailin Feng, and Xi Li. Multitask Non-Autoregressive Model for Human Motion Prediction. IEEE Transactions on Image Processing, 2020. 1, 2, 4 +[25] Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional Sequence to Sequence Model for Human Dynamics. In CVPR, pages 5226-5234, 2018. 2 +[26] Maosen Li, Siheng Chen, Zijing Zhang, Lingxi Xie, Qi Tian, and Ya Zhang. Skeleton-Parted Graph Scattering Networks for 3D Human Motion Prediction. arXiv preprint arXiv:2208.00368, 2022. 1, 2, 6, 7, 8 +[27] Maosen Li, Siheng Chen, Yangheng Zhao, Ya Zhang, Yanfeng Wang, and Qi Tian. Dynamic Multiscale Graph Neural Networks for 3D Skeleton Based Human Motion Prediction. In CVPR, pages 214-223, 2020. 2 +[28] Yang Li, Kan Li, Shuai Jiang, Ziyue Zhang, Congzhen-tao Huang, and Richard Yi Da Xu. Geometry-driven self-supervised method for 3d human pose estimation. In AAAI, volume 34, pages 11442-11449, 2020. 6 +[29] Mengyuan Liu, Fanyang Meng, and Yongsheng Liang. Generalized pose decoupled network for unsupervised 3d skeleton sequence-based action representation learning. Cyborg and Bionic Systems, 2022:0002, 2022. 1 +[30] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++. + +When Does Self-supervised Test-time Training Fail or Thrive? NeurIPS, 34:21808-21820, 2021. 2 +[31] Tiezheng Ma, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Progressively Generating Better Initial Guesses Towards Next Stages for High-Quality Human Motion Prediction. In CVPR, pages 6437-6446, 2022. 6, 7, 8 +[32] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. History Repeats Itself: Human Motion Prediction via Motion Attention. In ECCV, pages 474-489, 2020. 1, 2, 6 +[33] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Generating Smooth Pose Sequences for Diverse Human Motion Prediction. In ICCV, 2021. 1, 2, 3 +[34] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning Trajectory Dependencies for Human Motion Prediction. In ICCV, pages 9489-9497, 2019. 2, 3, 6 +[35] Julieta Martinez, Michael J Black, and Javier Romero. On Human Motion Prediction using Recurrent Neural Networks. In CVPR, pages 2891-2900, 2017. 2, 3, 6 +[36] Angel Martínez-González, Michael Villamizar, and Jean-Marc Odobez. Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive Transformers. In ICCV, pages 2276-2284, 2021. 2 +[37] Ishan Misra and Laurens van der Maaten. Self-supervised Learning of Predict-invariant Representations. In CVPR, pages 6707-6717, 2020. 4 +[38] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating Prediction-time Batch Normalization for Robustness under Covariate Shift. arXiv preprint arXiv:2006.10963, 2020. 2 +[39] Dario Pavllo, Christoph Feichtenhofer, Michael Auli, and David Grangier. *Quaternion: A Quaternion-based Recurrent Model for Human Motion*. International Journal of Computer Vision, (128):855-872, 2020. 2 +[40] Senthil Purushwalkam and Abhinav Gupta. Demystifying Contrastive Self-supervised Learning: Invariances, Augmentations and Dataset Biases. NeurIPS, 33:3407-3418, 2020. 4 +[41] Alejandro Hernandez Ruiz, Juergen Gall, and Francesc Moreno-Noguer. Human Motion Prediction via Spatio-Temporal Inpainting. In CVPR, pages 7134-7143, 2018. 2 +[42] Chajin Shin, Taeoh Kim, Sangjin Lee, and Sangyoun Leey. Test-Time Adaptation for Out-Of-Distributed Image Inpainting. In ICIP, 2021. 1, 4 +[43] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International Journal of Computer Vision, 87(1):4-27, 2010. 6 +[44] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time Training with Self-supervision for Generalization under Distribution Shifts. In ICML, pages 9229-9248. PMLR, 2020. 2 +[45] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A Dataset of Whole-Body Human Grasping of Objects. In ECCV, 2020. 6, 7, 8 + +[46] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain Randomization for Transferring Deep Neural Networks from Simulation to The Real World. In IROS, pages 23-30, 2017. 2 +[47] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In CVPRW, pages 969-977, 2018. 2 +[48] Thomas Varsavsky, Mauricio Orbes-Arteaga, Carole H Sudre, Mark S Graham, Parashkev Nachev, and M Jorge Cardoso. Test-time Unsupervised Domain Adaptation. In MIC-CAI, pages 428–436. Springer, 2020. 1 +[49] Borui Wang, Ehsan Adeli, Hsu-kuang Chiu, De-An Huang, and Juan Carlos Niebles. Imitation Learning for Human Pose Prediction. In ICCV, pages 7124-7133, 2019. 1 +[50] Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, and Trevor Darrell. On-target adaptation. arXiv preprint arXiv:2109.01087, 2021. 1 +[51] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully Test-time Adaptation By Entropy Minimization. arXiv preprint arXiv:2006.10726, 2020. 2 +[52] Qin Wang, Dengxin Dai, Lukas Hoyer, Luc Van Gool, and Olga Fink. Domain adaptive semantic segmentation with self-supervised depth estimation. In ICCV, pages 8515-8525, 2021. 1 +[53] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual Test-time Domain Adaptation. In CVPR, pages 7201-7211, 2022. 2, 5 +[54] Sirui Xu, Y-X Wang, and L-Y Gui. Diverse Human Motion Prediction Guided by Multi-level Spatial-temporal Anchors. In ECCV, 2022. 2 +[55] Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, and Yu-Chiang Frank Wang. Adversarial Teacher-student Representation Learning for Domain Generalization. NeurIPS, 34:19448-19460, 2021. 2, 3 +[56] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A Fourier Perspective on Model Robustness in Computer Vision. NeurIPS, 32, 2019. 2 +[57] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial Examples: Attacks and Defenses for Deep Learning. IEEE transactions on neural networks and learning systems, 30(9):2805-2824, 2019. 2 +[58] Ye Yuan and Kris Kitani. Dlow: Diversifying Latent Flows for Diverse Human Motion Prediction. In ECCV, pages 346-364, 2020. 1, 3 +[59] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test Time Robustness via Adaptation and Augmentation. arXiv preprint arXiv:2110.09506, 2021. 2 +[60] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Deep Domain-adversarial Image Generation for Domain Generalisation. In AAAI, volume 34, pages 13025-13032, 2020. 2 \ No newline at end of file diff --git a/2023/Test-time Personalizable Forecasting of 3D Human Poses/images.zip b/2023/Test-time Personalizable Forecasting of 3D Human Poses/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..901b7eab1bacb17bdb5a7eae5acaae3a426f1926 --- /dev/null +++ b/2023/Test-time Personalizable Forecasting of 3D Human Poses/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a7112c7db1a7a5970d8be2cb5be6fa2c14f2f9336848969f3b0d401265434fb +size 622406 diff --git a/2023/Test-time Personalizable Forecasting of 3D Human Poses/layout.json b/2023/Test-time Personalizable Forecasting of 3D Human Poses/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8e2e5a4644727e324339d1bf24b73430b9b209f8 --- /dev/null +++ b/2023/Test-time Personalizable Forecasting of 3D Human Poses/layout.json @@ -0,0 +1,13148 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 103, + 473, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 103, + 473, + 121 + ], + "spans": [ + { + "bbox": [ + 121, + 103, + 473, + 121 + ], + "type": "text", + "content": "Test-time Personalizable Forecasting of 3D Human Poses" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "spans": [ + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "text", + "content": "Qiongjie Cui" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "text", + "content": " Huaijiang Sun" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "text", + "content": " \nJianfeng Lu" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "text", + "content": " Weiqing Li" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "text", + "content": " Bin Li" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "text", + "content": " Hongwei Yi" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "text", + "content": ", Haofan Wang" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "text", + "content": "Nanjing University of Science and Technology, Nanjing, China \n" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "text", + "content": "Tianjin AiForward Science and Technology Co., Ltd., China \n" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "text", + "content": "Max Planck Institute for Intelligent Systems, Tubingen, Germany " + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "inline_equation", + "content": "{ }^{4}" + }, + { + "bbox": [ + 86, + 142, + 509, + 230 + ], + "type": "text", + "content": "Xiaohongshu Inc \ncuiqiongjie@njust.edu.cn sunhuaijiang@njust.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 275, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 290, + 574 + ], + "type": "text", + "content": "Current motion forecasting approaches typically train a deep end-to-end model from the source domain data, and then apply it directly to target subjects. Despite promising results, they remain non-optimal, due to privacy considerations, the test person and his/her natural properties (e.g., behavioral trait) are typically unseen in training. In this case, the source pre-trained model has a low ability to adapt to these out-of-source characteristics, resulting in an unreliable prediction. To tackle this issue, we propose a novel helper-predictor test-time personalization approach (H/P-TTP), which allows for a generalizable representation of out-of-source subjects to gain more realistic predictions. Concretely, the helper is preceded by explicit and implicit augmenters, where the former yields noisy sequences to improve robustness, while the latter is to generate novel-domain data with an adversarial learning paradigm. Then, the domain-generalizable learning is achieved where the helper can extract cross-subject invariant-knowledge to update the predictor. At test time, given a new person, the predictor is able to be further optimized to empower personalized capabilities to the specific properties. Extensive experiments show that with H/P-TTP, the existing models are significantly improved for various unseen subjects. The project page is available at https://sites.google.com/view/hp-ttp." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 590, + 128, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 128, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 128, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 608, + 287, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 287, + 679 + ], + "type": "text", + "content": "Forecasting the high-fidelity future human poses, conditioned on a given historical sequence, has attracted increasing attention in recent years. In a variety of 3D vision-based applications, such as autonomous driving and robot navigation, it offers tremendous potential, especially for tasks that call for seamless interaction with humans [11, 9, 29]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 681, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 681, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 59, + 681, + 287, + 693 + ], + "type": "text", + "content": "Recent years have witnessed a proliferation in the ex" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 250, + 545, + 411 + ], + "blocks": [ + { + "bbox": [ + 310, + 250, + 545, + 411 + ], + "lines": [ + { + "bbox": [ + 310, + 250, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 310, + 250, + 545, + 411 + ], + "type": "image", + "image_path": "1186abb990347563455d35c0471caaede26b0fa002a3c508ce4cf29ec8d7d050.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 415, + 547, + 503 + ], + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 503 + ], + "type": "text", + "content": "Figure 1. Starting from off-the-shelf networks, we first train a baseline model [26] on the source data, which is further optimized and personalized to the target person during the test phase. Here, we present an illustration of the pose sequence of various subjects, in which each 3D feature embedding is created by t-SNE. We notice that the motion patterns of test subjects fail to match the source one. Moreover, our prediction skeleton (blue) is closer to the ground truth (red) against the vanilla baseline [26]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 509, + 547, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 592 + ], + "type": "text", + "content": "ploration of this emerging field [16, 5, 2]. In particular, ongoing efforts are being devoted to deep learning approaches, which typically train a generic model from large-scale datasets, and then apply it to new test subjects and samples during the deployment phase. We notice that these methods have become mainstream and are now rapidly evolving [49, 58, 24, 32, 3, 33]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Despite notable successes, the existing approaches are sub-optimal, because the generic model cannot be adapted to unseen personalized subjects [48, 21, 52]. To be precise, due to data limitations or privacy concerns, unseen test persons, as well as their individual properties (e.g., age, gender, or behavioral trait), are typically disjointed with the training. It may lead to a large distribution gap between the target and source domain [19, 42, 50], as shown in Figure 1. For example, for the H3.6M dataset [22], a widely-adopted data splitting strategy takes subject-5 (S5) as the test and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "274" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "the rest as the training [10, 34, 25]. Here, the inherent properties (height, and body proportion) of S5 are significantly different from the other subjects; hence, the pre-trained parameters may not be specific or personalized for the S5. We also note that, out-of-source persons are inevitable in the deployment, which poses a major bottleneck to current predictive algorithms [3]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 286, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 286, + 407 + ], + "type": "text", + "content": "To solve it, a novel helper-predictor test-time personalization framework (H/P-TTP) is proposed. Starting with ready-to-use networks, our approach can attain tailored parameters according to the characteristics of various unseen subjects. Concretely, the H/P-TTP advances the recent success of knowledge distillation into test-time adaptation by augmenting source instances to out-of-source distributions, enabling domain-generalizable learning. The helper involves an identical architecture with the predictor, except for the explicit and implicit augmenters. To improve the robustness, by adding noise to the samples, the explicit augmenter extracts task-relevant information and ignores irrelevant ones. In contrast, the implicit augmenter is a trainable motion-style transformer to achieve the augmentations with new properties, which follows the adversarial training paradigm, with the goal of maximizing the discrepancy with the source subject, while ensuring semantic proximity. With the max-min optimization, the helper would be able to observe and cope with the novel-styled samples to gain the cross-subject invariant context, and distill the knowledge to update the predictor." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 407, + 286, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 286, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 286, + 550 + ], + "type": "text", + "content": "During testing, the model can be further updated to adapt to an unseen test person. We also observe that due to psychological or environmental factors, even for the same person, his/her motion characteristics may vary. More intuitively, for test data with large domain gaps, the predictor needs greater updates, and conversely, a mild personalization is expected. For this purpose, an adaptive learning rate strategy is proposed to dynamically consider the demand for test-time personalization. Owing to the adaptive test-time learning rate, our T/P-TTP can adapt to the intrinsic attributes of an unseen test subject, as well as to his/her specific motion dynamics, as shown in Figure 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 552, + 286, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 286, + 646 + ], + "type": "text", + "content": "Our contributions are as follows: 1) We notice the issue: the existing approaches are ill-adapted to the out-of-source subjects, and propose a novel H/P-TTP model to address it. 2) We also propose an adaptive learning rate strategy to further consider the extent to which the model needs to be personalized. 3) Our H/P-TTP is integrated into many existing motion forecasting approaches, and substantially boosts the performance on unseen subjects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 658, + 133, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 133, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 133, + 669 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Human Pose Forecasting. Deep end-to-end networks are highly sought-after for solving this issue [5, 41]. In particular, due to the intrinsic temporal nature of 3D human" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "motion, it is typically considered a sequence-to-sequence problem, where various variants of RNNs have been developed [35, 36, 15, 17]. Despite the promising development, RNNs suffer from error accumulation, convergence to an undesired static pose [24, 27, 25]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 134, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 134, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 134, + 545, + 217 + ], + "type": "text", + "content": "To tackle these drawbacks, feed-forward networks have been recently explored as a potential alternative [25, 39, 34, 33]. Particularly, with better interpretability, GCNs are introduced to extract the local semantic connectivity of the 3D human skeleton [34, 10, 8, 11, 24, 27, 36, 26, 54]. Transformers have also been introduced to capture the long-range dependencies of motion sequences [32]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 220, + 545, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 220, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 545, + 279 + ], + "type": "text", + "content": "We have noticed that, current deep learning-based approaches have emerged as the prevailing solution with promising results. However, an obvious challenge still exists, namely, the source pre-trained model cannot be personalized for the specific properties of unseen target persons." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 281, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 472 + ], + "type": "text", + "content": "Domain Generalization (DG). It aims to generalize the learned model from source domain datasets to unseen target domains. Recent advances reveal that data augmentation is an attractive strategy for DG, with the increase in the diversity of training and test data [20, 1, 56, 59]. For this purpose, a variety of works are proposed, e.g., adversarial attack [57, 60], image perturbation [60], and domain randomization [46, 47], where the main heuristic is to learn cross-domain invariant feature. Particularly, [55] introduces a novel-style augmentation in teacher-student frameworks, in which novel synthesized data allows for domain-invariant context to be transferred to the student network. In contrast to the standard DG, which trains a more generalizable model, our work considers going to improve the off-the-shelf network to adapt to the specific characteristics of unseen new subjects and their motion sequence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 475, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 475, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 475, + 545, + 713 + ], + "type": "text", + "content": "Test-time Adaptation (TTA). Our test-time personalization is relevant to the TTA, which is expected to improve the generalization ability of out-of-distribution test samples. TTA falls into source-free adaptation, and before making decisions, allows for fine-tuning the pre-trained model to adapt to the specific test samples during inference, without accessing the source domain data. For example, according to the test data, prediction-time batch normalization (PTBN) [38] can recalibrate the BN statistics. [51] utilizes test-time entropy minimization (TENT) to update the trainable parameters in BN layers. Test-time training (TTT) [44] adjusts the feature extractor in the test phase via the update of the self-supervised auxiliary task of image rotation prediction. " + }, + { + "bbox": [ + 305, + 475, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{TTT}++" + }, + { + "bbox": [ + 305, + 475, + 545, + 713 + ], + "type": "text", + "content": " [30] exploits contrastive learning to align the first- and second-order statistics of the test samples with the source domain data, improving the TTT further. [7] introduces an auxiliary image reconstruction to the primary deblurring task, and then resorts to meta-learning to fine-tune the primary branch to adapt to test samples. Considering the changing application environment, [53] integrates" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "275" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 269, + 235 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 269, + 235 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 269, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 269, + 235 + ], + "type": "image", + "image_path": "03f95f96bf887cf52742262d5dfc3dcf1c16863a61ccde499de8c53293750a9d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 236, + 211, + 243 + ], + "lines": [ + { + "bbox": [ + 110, + 236, + 211, + 243 + ], + "spans": [ + { + "bbox": [ + 110, + 236, + 211, + 243 + ], + "type": "text", + "content": "(a) H/P Domain-Generalizable Learning" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 271, + 71, + 541, + 235 + ], + "blocks": [ + { + "bbox": [ + 271, + 71, + 541, + 235 + ], + "lines": [ + { + "bbox": [ + 271, + 71, + 541, + 235 + ], + "spans": [ + { + "bbox": [ + 271, + 71, + 541, + 235 + ], + "type": "image", + "image_path": "3c8a8c443fe97fc86ac4f0976275b2556c1a40fded8a9072110c42dd1bfc4526.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 235, + 455, + 243 + ], + "lines": [ + { + "bbox": [ + 343, + 235, + 455, + 243 + ], + "spans": [ + { + "bbox": [ + 343, + 235, + 455, + 243 + ], + "type": "text", + "content": "(b) H/P Test-time Personalizable Forecasting" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "lines": [ + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "text", + "content": "Figure 2. Illustration of our approach. It mainly includes a parameter-free explicit augmenter " + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{(E)}" + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "text", + "content": " and a learnable implicit augmenter " + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "text", + "content": ", as well as the helper " + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\psi}" + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "text", + "content": " and predictor " + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\theta}" + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{(E)}" + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "text", + "content": " is to attain the noisy samples to improve the robustness and assist " + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "text", + "content": " to generate the novel-domain samples. (a) H/P domain-generalizable learning involves two steps: The implicit augmenter is first trained with the adversarial learning paradigm, to make the diverse novel-domain samples, and reserve the semantics; " + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\theta}" + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "text", + "content": " is trained to minimize the prediction loss " + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pred}" + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\psi}" + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "text", + "content": " is expected to achieve the consistent output with the predictor to distill the domain-invariant knowledge. We set the learning rate to " + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "text", + "content": " to obtain the base model " + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "inline_equation", + "content": "\\{\\stackrel{\\circ}{\\psi}, \\stackrel{\\circ}{\\theta}\\}" + }, + { + "bbox": [ + 46, + 246, + 545, + 336 + ], + "type": "text", + "content": ". (b) H/P Test-time Personalizable Forecasting is to further optimize the base helper/predictor to attain the personalized parameters for a new subject. We also analyze the demand of test-time personalization of a specific sequence, which is achieved by dynamically adjusting the test learning rate." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 286, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 286, + 380 + ], + "type": "text", + "content": "the teacher-student framework into continue test-time adaptation (CoTTA) to distill domain-generalizable knowledge and improve the robustness of the unseen domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 286, + 453 + ], + "type": "text", + "content": "Our inspiration comes partially from the above literature, with the following important distinctions: (1) It can be integrated with any existing deep end-to-end model. (2) Due to the meaningful novel-domain augmentation, high interpretability is provided. (3) It is possible to identify the demand to which the test sample needs to be adapted." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 462, + 163, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 163, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 163, + 475 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 479, + 286, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 286, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 286, + 575 + ], + "type": "text", + "content": "Given an observed sequence " + }, + { + "bbox": [ + 46, + 479, + 286, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{1:T} = [x_1, x_2, \\dots, x_T]" + }, + { + "bbox": [ + 46, + 479, + 286, + 575 + ], + "type": "text", + "content": " of a person across a time horizon " + }, + { + "bbox": [ + 46, + 479, + 286, + 575 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 479, + 286, + 575 + ], + "type": "text", + "content": ", our goal is to predict his/her future actions " + }, + { + "bbox": [ + 46, + 479, + 286, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{1:\\Delta T} = [y_1, y_2, \\dots, y_{\\Delta T}]" + }, + { + "bbox": [ + 46, + 479, + 286, + 575 + ], + "type": "text", + "content": ", in which each frame records the 3D coordinates of " + }, + { + "bbox": [ + 46, + 479, + 286, + 575 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 479, + 286, + 575 + ], + "type": "text", + "content": " joints. Typically, current approaches [35, 34, 11, 10] learn a generic mapping " + }, + { + "bbox": [ + 46, + 479, + 286, + 575 + ], + "type": "inline_equation", + "content": "f: \\mathbf{x}_{1:T} \\rightarrow \\mathbf{y}_{1:\\Delta T}" + }, + { + "bbox": [ + 46, + 479, + 286, + 575 + ], + "type": "text", + "content": ". As an extension, our H/P-TTP intends to boost their generalization capability to make personalizable predictions of unseen test subjects/motions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 578, + 190, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 190, + 591 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 190, + 591 + ], + "type": "text", + "content": "3.1. Novel Data Augmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 592, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 286, + 665 + ], + "type": "text", + "content": "Our H/P-TTP consists of a predictor and a helper with an identical structure, parameterized by " + }, + { + "bbox": [ + 46, + 592, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\theta}" + }, + { + "bbox": [ + 46, + 592, + 286, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 592, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\psi}" + }, + { + "bbox": [ + 46, + 592, + 286, + 665 + ], + "type": "text", + "content": ", derived from any existing networks or newly-designed as well. In contrast to the predictor, the helper, outfitted with the implicit and explicit augmenters, w.r.t. " + }, + { + "bbox": [ + 46, + 592, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 46, + 592, + 286, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 592, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{(E)}" + }, + { + "bbox": [ + 46, + 592, + 286, + 665 + ], + "type": "text", + "content": ", is dedicated to generating the augmented data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Implicit Augmenter. It is introduced to attain a generalizable feature, which falls into the max-min adversarial learning scope. Intuitively, the implicit augmenter " + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(T)}" + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": " is expected to generate diverse novel-domain augmentations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 346, + 545, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 405 + ], + "type": "text", + "content": "with a variety of attributes (e.g., ages, genders, statures) for the source sequence. After the adversarial learning is finished, the predictor allows for the distillation of domain-generalized representations from the helper, thus improving its personalization ability for unseen subject sequences." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 405, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 545, + 502 + ], + "type": "text", + "content": "We note that, " + }, + { + "bbox": [ + 304, + 405, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 304, + 405, + 545, + 502 + ], + "type": "text", + "content": " can be interpreted as a specially designed motion-style migration network, which is used to implicitly diversify the sequence into its younger/older or taller/lower views. It is associated with the literature on diverse human motion generation [3, 2, 12]. Therefore, we set up " + }, + { + "bbox": [ + 304, + 405, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 304, + 405, + 545, + 502 + ], + "type": "text", + "content": " with a similar network as [33], where the significant extraordinary is that it is trained with the max-min optimization and to yield novel-styled counterparts [55]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "text", + "content": "Given an observed sequence " + }, + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "text", + "content": ", the purpose of " + }, + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "text", + "content": " is to produce stylized augmentations " + }, + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}^{(I)} = \\mathcal{A}_{\\phi}^{(I)}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "text", + "content": " yet retaining its semantic (action label). To stimulate the augmentations with sufficiently distinct appearances, we propose to maximize the discrepancy " + }, + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "text", + "content": " between the original observed motion and " + }, + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "text", + "content": " augmented ones " + }, + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{x}}_h^{(I)}\\}_{h=1}^H" + }, + { + "bbox": [ + 304, + 502, + 545, + 579 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 337, + 581, + 545, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 581, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 337, + 581, + 545, + 611 + ], + "type": "interline_equation", + "content": "\\max _ {\\mathcal {A} ^ {(I)}} \\mathcal {L} _ {d i s} \\left(\\tilde {\\mathbf {x}} ^ {(I)}, \\mathbf {x}\\right) = \\sum_ {h = 1} ^ {H} \\left\\| \\frac {\\tilde {\\mathbf {x}} _ {h} ^ {(I)}}{\\left\\| \\tilde {\\mathbf {x}} _ {h} ^ {(I)} \\right\\| _ {2}} - \\frac {\\mathbf {x}}{\\left\\| \\mathbf {x} \\right\\| _ {2}} \\right\\| _ {2}. \\tag {1}", + "image_path": "1492bb54db0e56cbb9cae18fc638a16c61a26bd02a95e9c46b285d7d5a23d8fc.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 615, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 615, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 615, + 545, + 649 + ], + "type": "text", + "content": "To motivate diversity across the implicit augmentations, consistent with [58, 33], the diversity-promoting prior loss " + }, + { + "bbox": [ + 304, + 615, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{div}" + }, + { + "bbox": [ + 304, + 615, + 545, + 649 + ], + "type": "text", + "content": " with a normalizing factor " + }, + { + "bbox": [ + 304, + 615, + 545, + 649 + ], + "type": "inline_equation", + "content": "\\delta = 100" + }, + { + "bbox": [ + 304, + 615, + 545, + 649 + ], + "type": "text", + "content": " is adopted:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 653, + 545, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 653, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 314, + 653, + 545, + 687 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i v} = \\frac {2}{H (H - 1)} \\sum_ {j = 1} ^ {H} \\sum_ {h = j + 1} ^ {H} \\exp \\left(- \\frac {\\left\\| \\tilde {\\mathbf {x}} _ {j} ^ {(I)} - \\tilde {\\mathbf {x}} _ {h} ^ {(I)} \\right\\| _ {1}}{\\delta}\\right). \\tag {2}", + "image_path": "2b398989e5be83b3b29e0b23fbefcbd071aa6e81b1e9c7acad907bd9734a9fae.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 690, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 690, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 690, + 545, + 713 + ], + "type": "text", + "content": "In addition, to preserve the semantic proximity, we also introduce an action classifier to make the category label of" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "276" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "the novel-styled sequence consistent with the observed one. Given the correct label " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "-sized label space of a sequence and the resulting one " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\tilde{p}_h^{(I)}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "-th implicit augmentation, it is trained with cross-entropy loss " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 122, + 287, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 122, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 111, + 122, + 287, + 152 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = - \\frac {1}{C H} \\sum_ {h = 1} ^ {H} p \\log \\tilde {p} _ {h} ^ {(I)}. \\tag {3}", + "image_path": "03c48367ccd878c19340621b2f75f3a6fb84077f12add49a855d6361ac0de066.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 155, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 155, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 155, + 287, + 251 + ], + "type": "text", + "content": "Then, the implicit augmenter " + }, + { + "bbox": [ + 46, + 155, + 287, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 46, + 155, + 287, + 251 + ], + "type": "text", + "content": " is optimized to maximize the discrepancy, and minimize the diversity and classification error, which is expected to generate diverse augmentations as much as possible, while maintaining the semantic information with the original one [24]. After observing the new-domain augmentations, the helper can extract the domain-generalizable representation to optimize the predictor to personalize according to input samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 287, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 371 + ], + "type": "text", + "content": "Explicit Augmenter. To incorporate the additional information into the implicit augmenter, we introduce a nontrainable explicit one. In instance learning, each image is regarded as an instance, and we wish to train the network so that the representations of different augmented views of the same instance are as close as possible to each other. It provides more rich data for training, and helps learn transformation invariant representations [37, 40]. For natural images, random rotation, flipping, and noise injection are the most common augmentation strategies [14, 23, 42]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 372, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 407 + ], + "type": "text", + "content": "Besides, [8] develops a multi-task framework, in which the auxiliary branch is designed as motion reconstruction to provide meaningful cues for human pose prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "content": "Motivated by the above-mentioned work, we, therefore, design the explicit augmenter " + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{(E)}" + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "content": " to produce the noisy counterparts. Formally, given a distribution sampled from the observed sequences, " + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\sim q(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "content": ", it is injected with several Gaussian noises with a series of variances of a fixed interval, producing " + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 407, + 287, + 479 + ], + "type": "text", + "content": " augmentations with different noise levels:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 483, + 287, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 483, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 287, + 500 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} \\left(\\tilde {\\mathbf {x}} _ {k} ^ {(E)} \\mid \\mathbf {x}\\right) = \\mathcal {N} \\left(\\tilde {\\mathbf {x}} _ {k} ^ {(E)}; \\mathbf {x}, \\sigma^ {2} \\mathbf {I}\\right), \\tag {4}", + "image_path": "b932a6b4e787eb9f4df119bc889b89f164555ae0d275132752238fad8b31bd99.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 505, + 286, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 286, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 286, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 505, + 286, + 548 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{x}}_k^{(E)}\\}_{k = 1}^K" + }, + { + "bbox": [ + 46, + 505, + 286, + 548 + ], + "type": "text", + "content": " is the resulting noisy counterparts, which, together with the novel-domain augmentation " + }, + { + "bbox": [ + 46, + 505, + 286, + 548 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{x}}_h^{(I)}\\}_{h = 1}^H" + }, + { + "bbox": [ + 46, + 505, + 286, + 548 + ], + "type": "text", + "content": " from the learnable " + }, + { + "bbox": [ + 46, + 505, + 286, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 46, + 505, + 286, + 548 + ], + "type": "text", + "content": ", is used to train the helper." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 548, + 287, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 631 + ], + "type": "text", + "content": "Here, we note that, the non-parametric explicit augmenter provides the following advantages: (1) The noise counterparts are supplied, facilitating the helper to explore new patterns in the novel-domain augmented data. (2) It eliminates an overly optimistic assumption: the trainable implicit augmenter will necessarily yield novel-domain samples. (3) More robustness to noise is also included." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 637, + 243, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 637, + 243, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 637, + 243, + 651 + ], + "type": "text", + "content": "3.2. H/P Domain-Generalizable Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 652, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 652, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 652, + 287, + 715 + ], + "type": "text", + "content": "The objective of our helper is to improve the generalization of the predictor for unseen domain samples by observing the new-domain augmentations and noise counterparts, as shown in Figure 2 (left). The " + }, + { + "bbox": [ + 46, + 652, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{(E)}" + }, + { + "bbox": [ + 46, + 652, + 287, + 715 + ], + "type": "text", + "content": " is parameter-free, while " + }, + { + "bbox": [ + 46, + 652, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 46, + 652, + 287, + 715 + ], + "type": "text", + "content": " is trained to solve the following problem:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 511, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 511, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 511, + 84 + ], + "type": "text", + "content": "Algorithm 1 H/P Domain-Generalizable Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "spans": [ + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "text", + "content": "Require: implicit augmenter " + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "text", + "content": "; helper " + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\psi}" + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "text", + "content": " and predictor " + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\theta}" + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "text", + "content": " with the identical structure; learning rates " + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "text", + "content": "; batch size " + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "text", + "content": "; hyperparameters " + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\lambda_{dis}" + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\lambda_{div}" + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 87, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 124, + 463, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 124, + 463, + 136 + ], + "spans": [ + { + "bbox": [ + 306, + 124, + 463, + 136 + ], + "type": "text", + "content": "Output: the learned parameter " + }, + { + "bbox": [ + 306, + 124, + 463, + 136 + ], + "type": "inline_equation", + "content": "\\psi, \\theta, \\phi" + }, + { + "bbox": [ + 306, + 124, + 463, + 136 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 137, + 545, + 293 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 137, + 443, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 137, + 443, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 137, + 443, + 148 + ], + "type": "text", + "content": "1: randomly initialize " + }, + { + "bbox": [ + 307, + 137, + 443, + 148 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 307, + 137, + 443, + 148 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 307, + 137, + 443, + 148 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 307, + 137, + 443, + 148 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 307, + 137, + 443, + 148 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 307, + 137, + 443, + 148 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 148, + 410, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 410, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 410, + 159 + ], + "type": "text", + "content": "2: while not converge do" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 159, + 496, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 159, + 496, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 496, + 172 + ], + "type": "text", + "content": "3: sample a training mini-batch " + }, + { + "bbox": [ + 307, + 159, + 496, + 172 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_b,\\mathbf{y}_b\\}_{b = 1}^B" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 184 + ], + "type": "text", + "content": "4: for each " + }, + { + "bbox": [ + 307, + 172, + 545, + 184 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 307, + 172, + 545, + 184 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 307, + 172, + 545, + 184 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 307, + 172, + 545, + 184 + ], + "type": "text", + "content": " update implicit augmenter" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 184, + 545, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 196 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 307, + 184, + 545, + 196 + ], + "type": "inline_equation", + "content": "\\phi \\gets \\phi -\\alpha \\nabla_{\\phi}(-\\lambda_{dis}\\mathcal{L}_{dis} + \\lambda_{div}\\mathcal{L}_{div} + \\lambda_{cls}\\mathcal{L}_{cls})" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 196, + 361, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 361, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 361, + 206 + ], + "type": "text", + "content": "6: end for" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 208, + 361, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 361, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 361, + 217 + ], + "type": "text", + "content": "7: end while" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 220, + 410, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 410, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 410, + 230 + ], + "type": "text", + "content": "8: while not converge do" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 231, + 496, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 496, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 496, + 255 + ], + "type": "text", + "content": "9: sample a training mini-batch " + }, + { + "bbox": [ + 307, + 231, + 496, + 255 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_b,\\mathbf{y}_b\\}_{b = 1}^B" + }, + { + "bbox": [ + 307, + 231, + 496, + 255 + ], + "type": "text", + "content": " update the helper and predictor" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 255, + 496, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 496, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 496, + 270 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 307, + 255, + 496, + 270 + ], + "type": "inline_equation", + "content": "\\psi \\gets \\psi -\\sum_{b = 1}^{B}\\beta \\nabla_{\\psi}\\mathcal{L}_{pred}(\\tilde{\\bar{\\mathbf{y}}}_b^{(\\mathcal{H})},\\hat{\\mathbf{y}}_b^{(\\mathcal{P})});" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 270, + 476, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 270, + 476, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 270, + 476, + 284 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 307, + 270, + 476, + 284 + ], + "type": "inline_equation", + "content": "\\theta \\gets \\theta -\\sum_{b = 1}^{B}\\beta \\nabla_{\\theta}\\mathcal{L}_{pred}(\\hat{\\mathbf{y}}_b^{(\\mathcal{P})},\\mathbf{y}_b)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 284, + 366, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 284, + 366, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 366, + 293 + ], + "type": "text", + "content": "12: end while" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 330, + 299, + 545, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 299, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 330, + 299, + 545, + 317 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\phi \\in \\Phi} \\max \\lambda_ {d i s} \\mathcal {L} _ {d i s} + \\lambda_ {d i v} \\mathcal {L} _ {d i v} + \\lambda_ {c l s} \\mathcal {L} _ {c l s}, \\tag {5}", + "image_path": "5f861215056a28881bdbfcd231dcc00900bec0359093e7731521e5f1a5122041.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 321, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 321, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 545, + 360 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 321, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\lambda_{dis} = 0.5" + }, + { + "bbox": [ + 305, + 321, + 545, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 321, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\lambda_{div} = 0.3" + }, + { + "bbox": [ + 305, + 321, + 545, + 360 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 321, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\lambda_{cls} = 0.2" + }, + { + "bbox": [ + 305, + 321, + 545, + 360 + ], + "type": "text", + "content": " are the trade-off parameters. With a learning rate " + }, + { + "bbox": [ + 305, + 321, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\alpha = 0.001" + }, + { + "bbox": [ + 305, + 321, + 545, + 360 + ], + "type": "text", + "content": ", the implicit augmenter " + }, + { + "bbox": [ + 305, + 321, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 305, + 321, + 545, + 360 + ], + "type": "text", + "content": " is optimized:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 364, + 545, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 364, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 315, + 364, + 545, + 377 + ], + "type": "interline_equation", + "content": "\\phi \\leftarrow \\phi - \\alpha \\nabla_ {\\phi} \\left(- \\lambda_ {d i s} \\mathcal {L} _ {d i s} + \\lambda_ {d i v} \\mathcal {L} _ {d i v} + \\lambda_ {c l s} \\mathcal {L} _ {c l s}\\right). \\tag {6}", + "image_path": "fc7c698830cb997d94208076eb9621517a5903e76b4be109332108c3a1ea31b5.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 381, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 381, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 381, + 545, + 418 + ], + "type": "text", + "content": "We then train both the helper " + }, + { + "bbox": [ + 305, + 381, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\psi}" + }, + { + "bbox": [ + 305, + 381, + 545, + 418 + ], + "type": "text", + "content": " and predictor " + }, + { + "bbox": [ + 305, + 381, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\theta}" + }, + { + "bbox": [ + 305, + 381, + 545, + 418 + ], + "type": "text", + "content": " to achieve the H/P domain-generalizable learning. The optimization of " + }, + { + "bbox": [ + 305, + 381, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\psi}" + }, + { + "bbox": [ + 305, + 381, + 545, + 418 + ], + "type": "text", + "content": " is achieved using the gradient descent:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 356, + 422, + 545, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 422, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 356, + 422, + 545, + 437 + ], + "type": "interline_equation", + "content": "\\psi \\leftarrow \\psi - \\beta \\nabla_ {\\psi} \\mathcal {L} _ {p r e d} \\left(\\tilde {\\mathbf {y}} ^ {(\\mathcal {H})}, \\hat {\\mathbf {y}} ^ {(\\mathcal {P})}\\right), \\tag {7}", + "image_path": "49285ee8f14ec5eda99e851f221f401fb389d496037197022e32b570a53d0609.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}^{(\\mathcal{P})}" + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "text", + "content": " is the predictor's result, " + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\beta = 0.001" + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "text", + "content": " is the learning rate. The average outcome " + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\tilde{\\bar{\\xi}}^{(\\mathcal{H})}" + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "text", + "content": " is yielded over a total of " + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "inline_equation", + "content": "H + K" + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "text", + "content": " augmentations " + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{x}}_i\\}_{i=1}^{H+K}" + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "text", + "content": " fed into the helper " + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\psi}" + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "text", + "content": ". Following [10, 8], the prediction loss " + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pred}" + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "text", + "content": " distance. The update of the predictor " + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\theta}" + }, + { + "bbox": [ + 305, + 442, + 545, + 503 + ], + "type": "text", + "content": " can be achieved:" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 365, + 505, + 545, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 505, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 365, + 505, + 545, + 520 + ], + "type": "interline_equation", + "content": "\\theta \\leftarrow \\theta - \\beta \\nabla_ {\\theta} \\mathcal {L} _ {p r e d} (\\hat {\\mathbf {y}} ^ {(\\mathcal {P})}, \\mathbf {y}), \\tag {8}", + "image_path": "a2b76a56a3e5dcc9e805b8ab7a77479dd99dae116dad74b5b8cda4e4ac361566.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 305, + 521, + 490, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 490, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 490, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 521, + 490, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 305, + 521, + 490, + 533 + ], + "type": "text", + "content": " is the future ground-truth (gt) action." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": "The optimization of Eq. 7 and Eq. 8 ensures that the helper and predictor have consistent outcomes that are as close to the actual future sequence as possible. The domain-generalizable knowledge from the helper is progressively transferred to the predictor. The overall H/P domain-generalizable learning is summarized in Algorithm 1. Then, the base model, w.r.t " + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\{\\ddot{\\psi},\\dot{\\theta}\\}" + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": ", is achieved." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 624, + 521, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 624, + 521, + 637 + ], + "spans": [ + { + "bbox": [ + 306, + 624, + 521, + 637 + ], + "type": "text", + "content": "3.3. H/P Test-time Personalizable Forecasting" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "content": "During inference, for a new person, we consider further fine-tuning the base model to adapt to his/her specific properties. However, in practical applications, even for the same subject, his/her motion patterns may differ due to external factors, such as environmental changes. This requires varying degrees of test-time personalization (TTP) to adapt." + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "277" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "For this purpose, we propose to adjust the learning rate of test-time personalization. Our intuition is that, for the current step, its previous samples and their correct future sequences are observed deterministically. We are able to know how the association of the helper/predictor feature leads to the difference between the predicted and true values. Therefore, for the next sample, at test-time, calculating the distinction between the features of the helper and predictor, we can also gain insight into the relation of their expected outputs; if it is strong, then the helper can guide the predictor well, and a lower learning rate should be assigned to update the predictor, otherwise, a larger one." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": "Memory queue. For simplicity, the base helper " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\hat{\\psi}}" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": " and predictor " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\hat{\\theta}}" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": " are decomposed: feature extractor and generator, w.r.t " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\hat{\\psi} = \\{\\hat{\\psi}_1,\\hat{\\psi}_2\\}" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\dot{\\theta} = \\{\\dot{\\theta}_1,\\dot{\\theta}_2\\}" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": ". We denote " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": " as the helper and predictor. The outcome of the feature extractor and generator are denoted as " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": ". For a historical sample " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(m)}" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\{h\\leftrightarrow p\\}}^{(m)}" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_{\\{h\\leftrightarrow p\\}}^{(m)}" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": " are the feature and outcome differences of " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "h / p" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": " branches, calculated by a distance metric " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": " (i.e., euclidean distance), informally expressed " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_{\\{p\\leftrightarrow gt\\}}^{(m)}" + }, + { + "bbox": [ + 46, + 216, + 289, + 353 + ], + "type": "text", + "content": " calculates that the difference between the predictor's result and ground truth, indicating, from real history views, whether the predictor is well enough." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 353, + 289, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 353, + 289, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 289, + 381 + ], + "type": "text", + "content": "We build a " + }, + { + "bbox": [ + 47, + 353, + 289, + 381 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 353, + 289, + 381 + ], + "type": "text", + "content": "-elements memory queue " + }, + { + "bbox": [ + 47, + 353, + 289, + 381 + ], + "type": "inline_equation", + "content": "\\mathbb{M} = \\{(\\mathbf{f}_{\\{h \\leftrightarrow p\\}}^{(m)}, \\mathbf{o}_{\\{p \\leftrightarrow gt\\}}^{(m)})\\}_{m=1}^{M}" + }, + { + "bbox": [ + 47, + 353, + 289, + 381 + ], + "type": "text", + "content": ", denoted as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 387, + 289, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 387, + 289, + 417 + ], + "spans": [ + { + "bbox": [ + 75, + 387, + 289, + 417 + ], + "type": "interline_equation", + "content": "\\rightarrow \\underbrace {\\left(\\mathbf {f} _ {\\{h \\leftrightarrow p \\}} ^ {(1)} , \\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(1)}\\right) , \\dots , \\left(\\mathbf {f} _ {\\{h \\leftrightarrow p \\}} ^ {(M)} , \\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(M)}\\right)} _ {\\mathbb {M}}, \\rightarrow \\tag {9}", + "image_path": "068c37e0bd59a801dadadc510f08dd71b735ea21284a291028acfecac0c76f47.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 418, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 289, + 502 + ], + "type": "text", + "content": "It records, in terms of the real history, the relation of the " + }, + { + "bbox": [ + 46, + 418, + 289, + 502 + ], + "type": "inline_equation", + "content": "h / p" + }, + { + "bbox": [ + 46, + 418, + 289, + 502 + ], + "type": "text", + "content": " features and the difference between the predicted results and the ground truth (GT). In other words, it measures whether the helper is able to distill domain-generalizable information to the predictor for previous samples (or whether the model has been calibrated well). We note that " + }, + { + "bbox": [ + 46, + 418, + 289, + 502 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 46, + 418, + 289, + 502 + ], + "type": "text", + "content": " is updated by First In First Out (FIFO)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "text", + "content": "Given a new arriving sample " + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(0)}" + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "text", + "content": ", we calculate the feature difference " + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\{h\\leftrightarrow p\\}}^{(0)}" + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "text", + "content": " and the outcome difference " + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_{\\{h\\leftrightarrow p\\}}^{(0)}" + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "text", + "content": ". To measure its demand to be personalized, " + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\{h\\leftrightarrow p\\}}^{(0)}" + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "text", + "content": " is taken as the query, and retrieve " + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "text", + "content": " most similar ones " + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_{\\{h\\leftrightarrow p\\}}^{(d)}\\}_{d = 1}^D" + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "text", + "content": ", with their corresponding " + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{o}_{\\{p\\leftrightarrow gt\\}}^{(d)}\\}_{d = 1}^D" + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "text", + "content": " as the support set " + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 46, + 502, + 289, + 590 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 592, + 287, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 592, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 102, + 592, + 287, + 623 + ], + "type": "interline_equation", + "content": "\\underbrace {\\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(1)} , \\dots \\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(d)} , \\dots \\mathbf {o} _ {\\{p \\leftrightarrow g t \\}} ^ {(D)}} _ {\\mathbb {D}} \\tag {10}", + "image_path": "349ab48f3865a7f3a38bca8125124683e80bf6bea8db70516e718c38bd3b959a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 623, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 287, + 646 + ], + "type": "text", + "content": "This retrieval process is implemented by " + }, + { + "bbox": [ + 46, + 623, + 287, + 646 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 623, + 287, + 646 + ], + "type": "text", + "content": "-Nearest Neighbor algorithm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 647, + 288, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 288, + 680 + ], + "type": "text", + "content": "Dynamic learning rate. We then average the values in " + }, + { + "bbox": [ + 47, + 647, + 288, + 680 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 47, + 647, + 288, + 680 + ], + "type": "text", + "content": " as the reference " + }, + { + "bbox": [ + 47, + 647, + 288, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 47, + 647, + 288, + 680 + ], + "type": "text", + "content": ". The prediction discrepancy (PD) of " + }, + { + "bbox": [ + 47, + 647, + 288, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(0)}" + }, + { + "bbox": [ + 47, + 647, + 288, + 680 + ], + "type": "text", + "content": " is derived:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 681, + 287, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 681, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 61, + 681, + 287, + 700 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P D} = \\frac {1}{2} \\left(\\mathcal {L} _ {K L} \\left(\\mathbf {r} \\| \\mathbf {o} _ {\\{h \\leftrightarrow p \\}} ^ {(0)}\\right) + \\mathcal {L} _ {K L} \\left(\\mathbf {o} _ {\\{h \\leftrightarrow p \\}} ^ {(0)} \\| \\mathbf {r}\\right)\\right), \\tag {11}", + "image_path": "036d8f57c28e35ceafb14f7c4fa64e5f83da3637a6b82242f3b0bd8ee3242f4a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 50, + 701, + 268, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 701, + 268, + 713 + ], + "spans": [ + { + "bbox": [ + 50, + 701, + 268, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 50, + 701, + 268, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}" + }, + { + "bbox": [ + 50, + 701, + 268, + 713 + ], + "type": "text", + "content": " is the KL divergence of two distributions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 529, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 529, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 529, + 84 + ], + "type": "text", + "content": "Algorithm 2 H/P Test-time Personalizable Forecasting" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "spans": [ + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "text", + "content": "Require: test sample " + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(0)}" + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "text", + "content": "; base predictor " + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\dot{\\theta} = \\{\\ddot{\\theta}_1,\\ddot{\\theta}_2\\}" + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "text", + "content": "; base helper " + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\dot{\\psi} = \\{\\bar{\\psi}_1,\\bar{\\psi}_2\\}" + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "text", + "content": "; memory queue " + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "text", + "content": "; learning rate " + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "text", + "content": "; momentum " + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "text", + "content": "; number of nearest neighbors " + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 306, + 86, + 545, + 124 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 124, + 423, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 124, + 423, + 136 + ], + "spans": [ + { + "bbox": [ + 306, + 124, + 423, + 136 + ], + "type": "text", + "content": "Output: final prediction " + }, + { + "bbox": [ + 306, + 124, + 423, + 136 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}^{*}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 136, + 545, + 286 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 306, + 136, + 545, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 136, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 306, + 136, + 545, + 161 + ], + "type": "text", + "content": "1: augment " + }, + { + "bbox": [ + 306, + 136, + 545, + 161 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(0)}" + }, + { + "bbox": [ + 306, + 136, + 545, + 161 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 136, + 545, + 161 + ], + "type": "inline_equation", + "content": "H + K" + }, + { + "bbox": [ + 306, + 136, + 545, + 161 + ], + "type": "text", + "content": " augmentations " + }, + { + "bbox": [ + 306, + 136, + 545, + 161 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{x}}_i^{(0)}\\}_{i = 1}^{H + K}" + }, + { + "bbox": [ + 306, + 136, + 545, + 161 + ], + "type": "text", + "content": " and make " + }, + { + "bbox": [ + 306, + 136, + 545, + 161 + ], + "type": "inline_equation", + "content": "\\tilde{\\bar{\\mathbf{x}}}^{(0)}" + }, + { + "bbox": [ + 306, + 136, + 545, + 161 + ], + "type": "text", + "content": " equal to their average;" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 161, + 545, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 161, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 161, + 545, + 190 + ], + "type": "text", + "content": "2: calculate " + }, + { + "bbox": [ + 306, + 161, + 545, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{h\\leftrightarrow p}^{(0)} = \\mathcal{D}(\\mathcal{P}_{\\theta_1}(\\mathbf{x}^{(0)}),\\mathcal{H}_{\\psi_1}(\\bar{\\tilde{\\mathbf{x}}}^{(0)}))" + }, + { + "bbox": [ + 306, + 161, + 545, + 190 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_{h\\leftrightarrow p}^{(0)} = \\mathcal{D}(\\mathcal{P}_{\\theta}(\\mathbf{x}^{(0)}),\\mathcal{H}_{\\psi}(\\bar{\\tilde{\\mathbf{x}}}^{(0)}));" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 190, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 545, + 216 + ], + "type": "text", + "content": "3: retrieve the support set " + }, + { + "bbox": [ + 306, + 190, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 306, + 190, + 545, + 216 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 306, + 190, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 306, + 190, + 545, + 216 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 306, + 190, + 545, + 216 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 306, + 190, + 545, + 216 + ], + "type": "text", + "content": "-neighbors nearest with the query " + }, + { + "bbox": [ + 306, + 190, + 545, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{h\\leftrightarrow p}^{(0)}" + }, + { + "bbox": [ + 306, + 190, + 545, + 216 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 215, + 499, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 215, + 499, + 226 + ], + "spans": [ + { + "bbox": [ + 306, + 215, + 499, + 226 + ], + "type": "text", + "content": "4: ensemble the values in " + }, + { + "bbox": [ + 306, + 215, + 499, + 226 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 306, + 215, + 499, + 226 + ], + "type": "text", + "content": " as the reference " + }, + { + "bbox": [ + 306, + 215, + 499, + 226 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 306, + 215, + 499, + 226 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 227, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 227, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 227, + 545, + 239 + ], + "type": "text", + "content": "5: calculate the prediction discrepancy " + }, + { + "bbox": [ + 306, + 227, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PD}" + }, + { + "bbox": [ + 306, + 227, + 545, + 239 + ], + "type": "text", + "content": " using Eq. 11;" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 239, + 516, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 516, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 516, + 251 + ], + "type": "text", + "content": "6: obtain the adjusted learning rate " + }, + { + "bbox": [ + 306, + 239, + 516, + 251 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 306, + 239, + 516, + 251 + ], + "type": "text", + "content": " using Eq. 12;" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 251, + 514, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 514, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 514, + 262 + ], + "type": "text", + "content": "7: update the predictor " + }, + { + "bbox": [ + 306, + 251, + 514, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\theta}:\\theta^{*}\\gets \\dot{\\theta}" + }, + { + "bbox": [ + 306, + 251, + 514, + 262 + ], + "type": "text", + "content": " using Eq. 13;" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 262, + 511, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 511, + 274 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 511, + 274 + ], + "type": "text", + "content": "8: update the helper " + }, + { + "bbox": [ + 306, + 262, + 511, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\psi}:\\psi^{*}\\gets \\bar{\\psi}" + }, + { + "bbox": [ + 306, + 262, + 511, + 274 + ], + "type": "text", + "content": " , using Eq. 14;" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 274, + 402, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 402, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 402, + 286 + ], + "type": "text", + "content": "9: return " + }, + { + "bbox": [ + 306, + 274, + 402, + 286 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}^{*} = \\mathcal{P}_{\\theta^{*}}(\\mathbf{x})" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "type": "text", + "content": "For the next sample in continuous flow, we have known the actual future sequence " + }, + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{(0)}" + }, + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "type": "text", + "content": " of the previous sample " + }, + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(0)}" + }, + { + "bbox": [ + 306, + 293, + 545, + 316 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 316, + 545, + 346 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 316, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 330 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 307, + 316, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_{p\\leftrightarrow qt}^{(0)} = \\mathcal{D}(\\mathcal{P}_{\\theta}(\\mathbf{x}^{(0)}),\\mathbf{y}^{(0)})" + }, + { + "bbox": [ + 307, + 316, + 545, + 330 + ], + "type": "text", + "content": " ; " + }, + { + "bbox": [ + 307, + 316, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 307, + 316, + 545, + 330 + ], + "type": "text", + "content": " update memory queue" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 330, + 466, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 330, + 466, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 330, + 466, + 346 + ], + "type": "text", + "content": "2: M.pop() && M.push(" + }, + { + "bbox": [ + 307, + 330, + 466, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{h \\leftrightarrow p}^{(0)}, \\mathbf{o}_{p \\leftrightarrow gt}^{(0)}" + }, + { + "bbox": [ + 307, + 330, + 466, + 346 + ], + "type": "text", + "content": ");" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 356, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 356, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 545, + 429 + ], + "type": "text", + "content": "If the value of " + }, + { + "bbox": [ + 305, + 356, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{PD}" + }, + { + "bbox": [ + 305, + 356, + 545, + 429 + ], + "type": "text", + "content": " is small, the predictor's outcome coincides with the real history (the helper can guide the predictor well); otherwise, it is not consistent (a major update is needed). Based on this consideration, we adjust the learning rate to make a dynamic TTP. At test time, we obtain the adjusted learning rate " + }, + { + "bbox": [ + 305, + 356, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 356, + 545, + 429 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 397, + 434, + 545, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 434, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 397, + 434, + 545, + 445 + ], + "type": "interline_equation", + "content": "\\gamma = \\beta \\cdot \\mathcal {L} _ {P D}. \\tag {12}", + "image_path": "06ef5055ca06247a017e1aa4b16e6e0a23c6db9c7cc3870a31aa95e60c31a9bc.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "text", + "content": "Our TTP begins with the base model, " + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\hat{\\theta}}" + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\hat{\\psi}}" + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "text", + "content": ", obtained from the H/P novel-domain generalizable training stage. Then, we attain the predictor's personalized parameter to boost the generalization ability for the specific subject " + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{(0)}" + }, + { + "bbox": [ + 305, + 450, + 545, + 510 + ], + "type": "text", + "content": " using a single gradient descent step:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 357, + 514, + 545, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 514, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 357, + 514, + 545, + 530 + ], + "type": "interline_equation", + "content": "\\stackrel {\\circ} {\\theta} \\leftarrow \\stackrel {\\circ} {\\theta} - \\gamma \\cdot \\nabla_ {\\hat {\\theta}} \\mathcal {L} _ {p r e d} \\left(\\hat {\\mathbf {y}} ^ {(0)}, \\bar {\\tilde {\\mathbf {y}}} ^ {(0)}\\right), \\tag {13}", + "image_path": "ef097f67bcb448c86fdbd378666fb9222862f8ea08c54319eb71eb3ff88f52b7.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 305, + 534, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 534, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 305, + 534, + 545, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 534, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}^{(0)} = \\mathcal{P}_{\\hat{\\theta}}(\\mathbf{x})" + }, + { + "bbox": [ + 305, + 534, + 545, + 609 + ], + "type": "text", + "content": " is the immediate result, and the pseudo label " + }, + { + "bbox": [ + 305, + 534, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\bar{\\tilde{\\mathbf{y}}}^{(0)}" + }, + { + "bbox": [ + 305, + 534, + 545, + 609 + ], + "type": "text", + "content": " is the average prediction of the base helper over " + }, + { + "bbox": [ + 305, + 534, + 545, + 609 + ], + "type": "inline_equation", + "content": "H + K" + }, + { + "bbox": [ + 305, + 534, + 545, + 609 + ], + "type": "text", + "content": " augmentations. Our model can be regarded as a special teacher-student framework, and therefore, after the predictor's update, we use exponential moving average (EMA) [53] with " + }, + { + "bbox": [ + 305, + 534, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\mu = 0.95" + }, + { + "bbox": [ + 305, + 534, + 545, + 609 + ], + "type": "text", + "content": " momentum to optimize the helper:" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 374, + 611, + 545, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 611, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 374, + 611, + 545, + 626 + ], + "type": "interline_equation", + "content": "\\stackrel {\\circ} {\\psi} \\leftarrow \\mu \\cdot \\stackrel {\\circ} {\\psi} + (1 - \\mu) \\cdot \\stackrel {\\circ} {\\theta}. \\tag {14}", + "image_path": "4dde57ece3beed0871d0fa4bc4e443a9cc1127e141d5aa421727a1e9e294ee8e.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": "When the TTP is completed, we re-write the personalized parameters " + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\theta},\\hat{\\psi}\\}" + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\{\\theta^{*},\\psi^{*}\\}" + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": ", and perform a forward pass to yield the final prediction " + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}^{*} = \\mathcal{P}_{\\theta^{*}}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": ". Note that, after " + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "inline_equation", + "content": "P = 72" + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": " test-time personalizations, the personalized parameters " + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\{\\theta^{*},\\psi^{*}\\}" + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": " are rolled back to the base model " + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\theta},\\hat{\\psi}\\}" + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": " to avoid catastrophic forgetting. We summarize the H/P test-time personalizable forecasting in Algorithm 2." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "278" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 89, + 180, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 89, + 180, + 102 + ], + "spans": [ + { + "bbox": [ + 47, + 89, + 180, + 102 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "text", + "content": "As illustrated in Figure 2, our H/P-TTP begins with an off-the-shelf network, and includes two major phases: (1) H/P novel-domain generalizable training, (2) test-time personalizable forecasting. For the former, we first train the implicit augmenter " + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "text", + "content": " using Eq. 6. Both implicit and explicit augmenters yield " + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "inline_equation", + "content": "H = K = 8" + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "text", + "content": " augmentations. The variances of Gaussian noise injected into the explicit augmenter are set to " + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "inline_equation", + "content": "\\sigma^2 = [10,15,20,25,30,35,40,45]" + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "text", + "content": " in millimeter. Moreover, the classifier is set to a 3-layers MLP with 512 hidden units to ensure a similar action label between the augmentations and the original sample. Once " + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "text", + "content": " is trained, it is fixed in the following process. We exploit Eq. 7 and Eq. 8 to train the predictor " + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\theta}" + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "text", + "content": " and helper " + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\psi}" + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "text", + "content": ". The predictor/helper are optimized using AdamW optimizer with the weight decay factor (1e-2) on mini-batches of size 32, to achieve the base model. The size of the memory queue " + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "inline_equation", + "content": "M = 36" + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "text", + "content": ", and the support set " + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "inline_equation", + "content": "D = 12" + }, + { + "bbox": [ + 46, + 106, + 289, + 348 + ], + "type": "text", + "content": ". Specifically, we divide the baseline networks by the penultimate layer, where the previous layers are treated as the feature extractor, and it and its following layers are the generator." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 355, + 110, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 110, + 366 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 110, + 366 + ], + "type": "text", + "content": "4.2. Datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 371, + 269, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 371, + 269, + 383 + ], + "spans": [ + { + "bbox": [ + 58, + 371, + 269, + 383 + ], + "type": "text", + "content": "Our H/P-TTP is evaluated on 3 benchmark datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 384, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 445 + ], + "type": "text", + "content": "Dataset-1: H3.6M [22] is the most widely-used dataset for pose prediction, containing " + }, + { + "bbox": [ + 46, + 384, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\approx 3.6" + }, + { + "bbox": [ + 46, + 384, + 287, + 445 + ], + "type": "text", + "content": " million frames of 15 action categories performed by 7 human subjects. Consistent with [34, 6], all sequences are downsampled to 25 fps, and represented as a " + }, + { + "bbox": [ + 46, + 384, + 287, + 445 + ], + "type": "inline_equation", + "content": "N = 17" + }, + { + "bbox": [ + 46, + 384, + 287, + 445 + ], + "type": "text", + "content": " joint skeleton." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 446, + 287, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 530 + ], + "type": "text", + "content": "Dataset-2: GRAB [45] is a newly-introduced benchmark, including " + }, + { + "bbox": [ + 46, + 446, + 287, + 530 + ], + "type": "inline_equation", + "content": "\\approx 1.6" + }, + { + "bbox": [ + 46, + 446, + 287, + 530 + ], + "type": "text", + "content": " million poses of 29 actions from 10 human subjects. Compared with H3.6M, the pose sequences in GRAB are more diverse and involve interaction with the physical world, which is, therefore, more complex and challenging. We down-sample the sequences to 30 fps with 25 joints in each skeleton [13]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 579 + ], + "type": "text", + "content": "Dataset-3: HumanEva-I [43] consists of 3 human subjects, performing 6 pre-defined actions. All sequence is down-sampled to 30 fps, represented by a 15-joint skeleton. The results are reported in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 580, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 617 + ], + "type": "text", + "content": "We note that, for H3.6M [22], the length of the observed and predicted sequence is 25 frames, while for both GRAB [45] and HumanEva-I [43], it is 30." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 625, + 168, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 168, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 168, + 639 + ], + "type": "text", + "content": "4.3. Experimental Setups" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "text", + "content": "We use 2 main setups to analyze our model (1.x and 2.x as stated in Table 1). As in [34, 32, 10, 31, 13], the former is to evaluate the general prediction ability on common data splitting. The latter is newly-designed to evaluate the personalizable ability of unseen subjects, where " + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "inline_equation", + "content": "S_{x}" + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "text", + "content": " is the test subjects, and the disjointed ones " + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "inline_equation", + "content": "S_{x}^{-}" + }, + { + "bbox": [ + 46, + 643, + 288, + 717 + ], + "type": "text", + "content": "is the training." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 70, + 545, + 129 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 545, + 129 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 545, + 129 + ], + "type": "table", + "html": "
Purposegeneral predictive ability (1.x)predictive ability for unseen subjects (2.x)
Setup(1.1)(1.2)(2.1)(2.2)(2.3)
DatasetH3.6M [22]GRAB[45]H3.6M[22]GRAB[45]HumanEva-I [43]
TestS5S10SxSxSx
TrainingS-5S-10S-S-S-
", + "image_path": "581e40941f262d875ea987a0b87d2e58dce17b2030f323042e7599af3c200ee3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 369, + 132, + 478, + 142 + ], + "lines": [ + { + "bbox": [ + 369, + 132, + 478, + 142 + ], + "spans": [ + { + "bbox": [ + 369, + 132, + 478, + 142 + ], + "type": "text", + "content": "Table 1. Experimental setups." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 143, + 372, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 143, + 372, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 143, + 372, + 155 + ], + "type": "text", + "content": "4.4. Baselines" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 158, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 314 + ], + "type": "text", + "content": "We exploit 6 approaches, emerged in recent years as our baselines, including ResSup [35], LTD [34], HisRep [32], MSR [11], PGBIG [31], and SPGSN [26]. They are separately integrated into our H/P-TTP framework to activate the prediction performance for unseen targets. [35] is RNN-based, [32] is attention-based, and the rest are based on GCNs. For the general predictive ability, we use the common data splitting, as the setup-1.x in Table 1, and the setup2.x in Table 1 is used to endow the baselines with the personalizable ability for unseen subjects. We note that our ultimate goal is to equip the H/P-TTP with each baseline to improve the prediction ability, which is therefore renamed, e.g., PGBIG [31] + H/P-TTP, SPGSN [26] + H/P-TTP." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 319, + 373, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 373, + 330 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 373, + 330 + ], + "type": "text", + "content": "4.5. Protocols" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 333, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 545, + 357 + ], + "type": "text", + "content": "To comprehensively investigate the performance of our H/P-TTP, the following 4 protocols are exploited." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 358, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 545, + 381 + ], + "type": "text", + "content": "Protocol-1: The widely-applied mean per-joint position error (MPJPE) [22, 34, 6, 31] is first used:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 354, + 384, + 545, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 384, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 354, + 384, + 545, + 407 + ], + "type": "interline_equation", + "content": "E _ {\\mathrm {M P J P E}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\| \\boldsymbol {p} _ {n} - \\hat {\\boldsymbol {p}} _ {n} ^ {*} \\| _ {2}, \\tag {15}", + "image_path": "83a45fed01e4191107e125a31d5a1209bba1d49a88182e70c9dc1da79f00d4a8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 411, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 545, + 435 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 411, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}_n^*" + }, + { + "bbox": [ + 306, + 411, + 545, + 435 + ], + "type": "text", + "content": " is the 3D position of " + }, + { + "bbox": [ + 306, + 411, + 545, + 435 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 306, + 411, + 545, + 435 + ], + "type": "text", + "content": "-th joint in a final predicted frame " + }, + { + "bbox": [ + 306, + 411, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_t^*" + }, + { + "bbox": [ + 306, + 411, + 545, + 435 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 411, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\pmb{p}_n" + }, + { + "bbox": [ + 306, + 411, + 545, + 435 + ], + "type": "text", + "content": " is the gt. " + }, + { + "bbox": [ + 306, + 411, + 545, + 435 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 411, + 545, + 435 + ], + "type": "text", + "content": " is the number of joints." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 435, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 519 + ], + "type": "text", + "content": "Protocol-2: We observe that compared with the GT, some predicted pose coordinates have a slight global offset. To eliminate the error independent of poses, we supplement the Procrustes aligned MPJPE (P-MPJPE), which calculates the MPJPE after aligning the predicted pose to the gt pose by a rigid transformation called Procrustes Analysis (PA) [28]. For a frame " + }, + { + "bbox": [ + 304, + 435, + 545, + 519 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 435, + 545, + 519 + ], + "type": "text", + "content": " in a skeleton layout " + }, + { + "bbox": [ + 304, + 435, + 545, + 519 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 435, + 545, + 519 + ], + "type": "text", + "content": ", it is defined as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 521, + 545, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 521, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 318, + 521, + 545, + 544 + ], + "type": "interline_equation", + "content": "E _ {\\mathrm {P - M P J P E}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\| m _ {\\mathrm {P A}, \\mathrm {S}} ^ {t} (n) - m _ {\\mathrm {g t}, \\mathrm {S}} ^ {t} (n) \\| _ {2}, \\tag {16}", + "image_path": "f802a4db83bdf554b2c697a23433191c50e46d01247529485c684d233557cefe.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 548, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 548, + 545, + 572 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{PA,S}}^t (n)" + }, + { + "bbox": [ + 304, + 548, + 545, + 572 + ], + "type": "text", + "content": " is the coordinate of " + }, + { + "bbox": [ + 304, + 548, + 545, + 572 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 548, + 545, + 572 + ], + "type": "text", + "content": " -th joint after alignment using PA, " + }, + { + "bbox": [ + 304, + 548, + 545, + 572 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{glS}}^{t}(n)" + }, + { + "bbox": [ + 304, + 548, + 545, + 572 + ], + "type": "text", + "content": " is the corresponding gt." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 573, + 545, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 573, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 573, + 545, + 655 + ], + "type": "text", + "content": "Protocol-3: Both MPJPE and P-MPJPE are hard to quantify the performance bound. Therefore, Percentage of Correct 3D Keypoint (PCK) is also introduced, which statistics the proportion of predicted joints with MPJPE errors smaller than a pre-defined threshold. Motivated by the current 3D human pose estimation [4, 18], we use the threshold of " + }, + { + "bbox": [ + 304, + 573, + 545, + 655 + ], + "type": "inline_equation", + "content": "150\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 573, + 545, + 655 + ], + "type": "text", + "content": ", re-written as PCK@150mm." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 657, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 657, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 657, + 545, + 691 + ], + "type": "text", + "content": "Protocol-4: In addition, we also use the mean per-bone length error (MPBLE) to evaluate the bone-length difference. The MPBLE is defined as: -" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 347, + 693, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 693, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 347, + 693, + 545, + 716 + ], + "type": "interline_equation", + "content": "E _ {\\mathrm {M P B L E}} = \\frac {1}{N - 1} \\sum_ {n = 1} ^ {N - 1} \\| l _ {n} - \\hat {l} _ {n} ^ {*} \\| _ {1}, \\tag {17}", + "image_path": "74a9931733b54da2867de24944976914618a0fef918c5430936814148d9ceff3.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "279" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 541, + 141 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 541, + 141 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 541, + 141 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 541, + 141 + ], + "type": "table", + "html": "
MethodResSupResSup +H/P-TTPLTDLTD +H/P-TTPHisRepHisRep +H/P-TTPMSR [11]MSR +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
Time (ms)400100040010004001000400100040010004001000400100040010004001000400100040010004001000
MPJPE [mm] ↓115.2130.5107.4125.663.5114.359.6109.858.3112.156.8109.462.9114.260.8108.658.5110.356.3105.158.3109.655.6103.7
P-MPJPE [mm] ↓96.6111.089.2107.153.1101.547.394.652.6100.146.294.855.7101.351.293.751.297.249.795.651.896.749.490.2
PCK@150mm [%]↑57.550.359.253.770.466.072.569.771.265.573.769.475.570.180.372.077.369.682.471.580.171.285.474.6
MPBLE [mm] ↓33.645.730.543.625.334.023.532.626.432.422.031.127.235.224.733.223.029.319.427.319.325.516.323.8
", + "image_path": "9d65fb181fc66469f5590b2acec3a04a34d14c70c0ee3d7713c9fd6cfce1848d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 185, + 293, + 257 + ], + "blocks": [ + { + "bbox": [ + 47, + 144, + 547, + 178 + ], + "lines": [ + { + "bbox": [ + 47, + 144, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 547, + 178 + ], + "type": "text", + "content": "Table 2. Average short-term (400ms) & long-term (1000ms) prediction performance (on H3.6M [22]) using the common setup-1.1. We observe that, in general, with the H/P-TTP, the baselines achieve better results than the trivial ones. It reveals that " + }, + { + "bbox": [ + 47, + 144, + 547, + 178 + ], + "type": "inline_equation", + "content": "S_{5}" + }, + { + "bbox": [ + 47, + 144, + 547, + 178 + ], + "type": "text", + "content": " includes the different motion properties from the other subjects, while our H/T-TTP can adapt them with the personalized parameters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 185, + 293, + 257 + ], + "lines": [ + { + "bbox": [ + 48, + 185, + 293, + 257 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 293, + 257 + ], + "type": "table", + "html": "
MethodMSRMSR +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
Time (ms)500100050010005001000500100050010005001000
MPJPE ↓96.3178.688.2147.392.2157.284.4141.291.3144.585.5136.4
P-MPJPE ↓82.3131.378.8122.879.8127.566.2114.277.2129.064.9117.3
PCK@150mm↑75.365.678.367.675.866.479.069.177.067.881.170.4
MPBLE ↓21.328.520.227.020.328.019.526.219.226.618.024.7
", + "image_path": "600feca13f0c49e9d7533bef36e34559c83318ce929306b06a6db9b60f7bfc46.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 257, + 288, + 278 + ], + "lines": [ + { + "bbox": [ + 47, + 257, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 288, + 278 + ], + "type": "text", + "content": "Table 3. Average short- (500ms) & long-term (1000ms) performance (on GRAB [45]) using the setup-1.2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 281, + 288, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 288, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 281, + 288, + 331 + ], + "type": "inline_equation", + "content": "l_{n}" + }, + { + "bbox": [ + 47, + 281, + 288, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 281, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\hat{l}_n^*" + }, + { + "bbox": [ + 47, + 281, + 288, + 331 + ], + "type": "text", + "content": " are the bone length of " + }, + { + "bbox": [ + 47, + 281, + 288, + 331 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 281, + 288, + 331 + ], + "type": "text", + "content": "-th bone in the gt and predicted pose, respectively. " + }, + { + "bbox": [ + 47, + 281, + 288, + 331 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 47, + 281, + 288, + 331 + ], + "type": "text", + "content": " is the total number of bones in a skeleton. Smaller MPBLEs tend to show a more stable visualization of the predicted frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 336, + 146, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 146, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 146, + 348 + ], + "type": "text", + "content": "4.6. Results Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 354, + 288, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 288, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 288, + 582 + ], + "type": "text", + "content": "General prediction ability is first evaluated under the condition of common data splitting [10, 8] with all 6 baselines, as the setups-1.x in Table 1. Specifically, for the H3.6M, Table 2 compares the average result of the 4 protocols over the samples of " + }, + { + "bbox": [ + 46, + 354, + 288, + 582 + ], + "type": "inline_equation", + "content": "S_{5}" + }, + { + "bbox": [ + 46, + 354, + 288, + 582 + ], + "type": "text", + "content": " with the experimental setup-1.1. Moreover, under the setup-1.2, Table 3 reports the comparisons on the GRAB dataset [45]. From the results, we observe that, for the used 2 datasets, once equipped with the H/P-TTP, all baselines coincidentally achieve better performance. The reason is that for the common data splitting of H3.6M and GRAB, the test subject (subject-5 or subject-10) is disjointed from the training ones. Stated in a different way, the motion properties of the target subject-5 (S5) are unknown in training (out-of-source), which is not considered by the trivial baselines, thus making the prediction task more challenging. By contrast, with the domain-generalizable learning and test-time personalization, our H/P-TTP can attain the personalized parameters to adapt to the dynamic properties of the new test subject (S5)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "Prediction ability for unseen subjects is then analyzed using the newly-designed data splitting. Because of data limitations and privacy concerns, the statures, ages, or genders of the target subjects " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "S_{x}" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " in the deployment environment are usually evident in the training stage, resulting in the distribution gap and out-of-source motion properties (e.g., motion style, rhythm). To further evaluate the personalization ability of our H/P-TTP for unseen subjects, we design the setup-2.1&2.2 in Table 1 for H3.6M and GRAB datasets, respectively. Due to their better performance, only both PGBIG and SPGSN are selected as the competitors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 186, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 270 + ], + "type": "text", + "content": "The comparison results are reported in Table 4 and Table 5, in which each subject " + }, + { + "bbox": [ + 304, + 186, + 545, + 270 + ], + "type": "inline_equation", + "content": "S_{x}" + }, + { + "bbox": [ + 304, + 186, + 545, + 270 + ], + "type": "text", + "content": " is tested separately and averaged over all its samples. From the results, we observe that, with H/P-TTP, the PGBIG [31] + H/P-TTP and SPGSN [26] + H/P-TTP are calibrated under almost all target subjects. It evidences that the motion properties of out-of-source unseen subjects are indeed adapted." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 270, + 545, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 306 + ], + "type": "text", + "content": "We also present the qualitative comparison of the SoTA SPGSN [26], and SPGSN [26] " + }, + { + "bbox": [ + 304, + 270, + 545, + 306 + ], + "type": "inline_equation", + "content": "+\\mathrm{H} / \\mathrm{P}" + }, + { + "bbox": [ + 304, + 270, + 545, + 306 + ], + "type": "text", + "content": " -TTP in Figure 3, under camera-takepicture-1 of unseen " + }, + { + "bbox": [ + 304, + 270, + 545, + 306 + ], + "type": "inline_equation", + "content": "S_{7}" + }, + { + "bbox": [ + 304, + 270, + 545, + 306 + ], + "type": "text", + "content": " from the GRAB dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 314, + 406, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 314, + 406, + 326 + ], + "spans": [ + { + "bbox": [ + 306, + 314, + 406, + 326 + ], + "type": "text", + "content": "4.7. Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 332, + 545, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 545, + 368 + ], + "type": "text", + "content": "Next, we investigate the impact of several key components. The ablation experiments report the average (at 1000ms) under the setup-2.1 in Table 1 on H3.6M dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 369, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 485 + ], + "type": "text", + "content": "Intuitively, the (1) test-time personalization (TTP) forecasting process can further optimize the base model to adapt to unseen target subjects. It is confirmed in Table 6. We also evaluate the impact of the (2) implicit and explicit augmenter " + }, + { + "bbox": [ + 304, + 369, + 545, + 485 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}^{(E)}" + }, + { + "bbox": [ + 304, + 369, + 545, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 369, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)})" + }, + { + "bbox": [ + 304, + 369, + 545, + 485 + ], + "type": "text", + "content": ", and (3) dynamic learning rate. As shown in Table 6, both " + }, + { + "bbox": [ + 304, + 369, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{(E)}" + }, + { + "bbox": [ + 304, + 369, + 545, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 369, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 304, + 369, + 545, + 485 + ], + "type": "text", + "content": " provide the positive effect, while " + }, + { + "bbox": [ + 304, + 369, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\phi}^{(I)}" + }, + { + "bbox": [ + 304, + 369, + 545, + 485 + ], + "type": "text", + "content": " is greater. Moreover, TTP is indeed helpful, and especially with the dynamic learning rate, the performance is further improved." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 304, + 486, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 544 + ], + "type": "text", + "content": "(4) Impact of the initial learning rate " + }, + { + "bbox": [ + 304, + 486, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 486, + 545, + 544 + ], + "type": "text", + "content": " is also investigated in Table 7(left), which is essential to H/P domain-generalizable learning and TTP. We see that " + }, + { + "bbox": [ + 304, + 486, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 486, + 545, + 544 + ], + "type": "text", + "content": " is insensitive to the performance. We suggest that it is because it can be calibrated by the dynamic learning rate at test time." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "text", + "content": "(5) Impact of the number of augmentations of the implicit and explicit augmenter is shown in Table 7(left), and " + }, + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "inline_equation", + "content": "H = K = 8" + }, + { + "bbox": [ + 304, + 545, + 545, + 581 + ], + "type": "text", + "content": " achieves the better prediction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 629 + ], + "type": "text", + "content": "(6) Number " + }, + { + "bbox": [ + 304, + 582, + 545, + 629 + ], + "type": "inline_equation", + "content": "(u = [0,1,2,3,4])" + }, + { + "bbox": [ + 304, + 582, + 545, + 629 + ], + "type": "text", + "content": " of gradient descent of TTP is investigated in Table 7(right). We observe that, with the dynamic learning rate, our TTP only needs to perform a single update. More updates bring no further benefits." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "text", + "content": "(7) Impact of the variances of Gaussian noise injection to the explicit augmenter is shown in Table 8(left), where the middle one attains the best performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "(8) Impact of the sizes of " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{D}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ". As reported in Table 8(right), the better result is balanced on the condition of " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "M = 36" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "D = 12" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ". Too small leads to catastrophic forgetting, while too large increases the memory burden." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "280" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 542, + 143 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 542, + 143 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 542, + 143 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 542, + 143 + ], + "type": "image", + "image_path": "9971ce0c05e1bc58e98975d67bb355e7c36325a64c194e726b8dd223b8385a33.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 146, + 547, + 179 + ], + "lines": [ + { + "bbox": [ + 46, + 146, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 547, + 179 + ], + "type": "text", + "content": "Figure 3. Predicted pose visualization, where the top is vanilla SPGSN [26] and the bottom is SPGSN " + }, + { + "bbox": [ + 46, + 146, + 547, + 179 + ], + "type": "inline_equation", + "content": "[26] + \\mathrm{H} / \\mathrm{P}" + }, + { + "bbox": [ + 46, + 146, + 547, + 179 + ], + "type": "text", + "content": " -TTP. The underlying green denotes the gt, and the red/blue is the prediction. The contrasting segments are highlighted in the purple box. We also enlarge some important details in the yellow ellipse. It is clear that with our H/P-TTP, the predicted poses are more accurate." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 188, + 547, + 275 + ], + "blocks": [ + { + "bbox": [ + 48, + 188, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 48, + 188, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 547, + 275 + ], + "type": "table", + "html": "
MPJPE [mm] ↓PMPJPE [mm] ↓PCK@150mm [%] ↑MPBLE [mm] ↓
unseen subjectsPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
S1116.6109.3111.4109.697.794.595.491.967.269.367.972.429.625.826.025.2
S6103.698.298.695.284.980.981.878.561.863.964.567.026.222.923.122.3
S7107.0102.1110.4103.888.984.592.986.364.266.364.969.427.423.924.123.3
S896.592.1100.395.478.877.883.474.267.871.968.573.024.020.921.120.3
S9105.3101.2104.5100.187.083.590.386.863.265.363.968.426.623.323.522.7
S11120.8115.5117.0112.8101.597.599.095.570.272.370.975.430.026.726.926.1
Average108.3103.1107.0102.889.886.590.585.565.768.266.470.927.323.924.123.3
", + "image_path": "358087b91bf385fd4d7c14691f2fe2217d5cec6830d62bfb9f463285afdc1d26.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 302, + 547, + 414 + ], + "blocks": [ + { + "bbox": [ + 46, + 275, + 545, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 545, + 298 + ], + "type": "text", + "content": "Table 4. Average performance comparison (of both SoTA PGBIG [31] and SPGSN [26], on H3.6M dataset [22]) of the end predicted pose (1000ms) over samples of each unseen subject " + }, + { + "bbox": [ + 46, + 275, + 545, + 298 + ], + "type": "inline_equation", + "content": "S_{x}" + }, + { + "bbox": [ + 46, + 275, + 545, + 298 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 275, + 545, + 298 + ], + "type": "inline_equation", + "content": "x = [1,6,7,8,9,11]" + }, + { + "bbox": [ + 46, + 275, + 545, + 298 + ], + "type": "text", + "content": ", and the corresponding average over all unseen subjects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 302, + 547, + 414 + ], + "lines": [ + { + "bbox": [ + 48, + 302, + 547, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 302, + 547, + 414 + ], + "type": "table", + "html": "
MPJPE [mm] ↓PMPJPE [mm] ↓PCK@150mm [%] ↑MPBLE [mm] ↓
unseen subjectsPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTPPGBIGPGBIG +H/P-TTPSPGSNSPGSN +H/P-TTP
S1177.5152.0176.3151.5146.8136.3149.6137.063.865.364.168.628.425.327.125.0
S2182.5164.3179.3156.0154.3150.4150.8146.458.060.660.463.731.528.629.126.6
S3162.3143.2157.2143.6144.1134.2140.7136.764.266.066.370.426.825.225.723.8
S4167.0149.1160.3145.7148.5133.9138.7128.063.769.367.268.927.725.426.024.1
S5142.1126.0142.8127.5121.4109.5125.6104.965.067.370.372.722.220.322.720.4
S6167.2150.4163.7144.2147.7139.3139.1126.862.068.365.367.627.824.326.224.9
S7140.4122.8138.7116.5123.4111.2115.2107.064.867.268.371.022.420.520.918.6
S8133.9112.7130.4111.9118.5106.5110.7106.266.970.270.173.220.818.819.217.2
S9153.6120.0154.6121.3135.7107.1134.8105.462.263.066.370.225.122.825.223.1
Average158.6137.7155.9135.5137.8125.4133.9124.163.466.466.569.625.923.524.722.6
", + "image_path": "1e00478fc3018556c7a2fecbf5131237e8cdede9d08271c278ffe85956effb9a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 448, + 286, + 499 + ], + "blocks": [ + { + "bbox": [ + 46, + 414, + 545, + 436 + ], + "lines": [ + { + "bbox": [ + 46, + 414, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 545, + 436 + ], + "type": "text", + "content": "Table 5. Average performance comparison (of both SoTA PGBIG [31] and SPGSN [26], on GRAB dataset [45]) of the end predicted pose (1000ms) over samples of each unseen subject " + }, + { + "bbox": [ + 46, + 414, + 545, + 436 + ], + "type": "inline_equation", + "content": "S_{x}" + }, + { + "bbox": [ + 46, + 414, + 545, + 436 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 414, + 545, + 436 + ], + "type": "inline_equation", + "content": "x = [1,2,3,4,5,6,7,8,9]" + }, + { + "bbox": [ + 46, + 414, + 545, + 436 + ], + "type": "text", + "content": ", and the average over all unseen subjects." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 448, + 286, + 499 + ], + "lines": [ + { + "bbox": [ + 48, + 448, + 286, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 286, + 499 + ], + "type": "table", + "html": "
A(E)A(1)ψTTPdynamic learning rateMPJPE [mm] ↓
SPGSN [26]✓✓✓109.5
✓✓✓104.2
✓✓114.3
✓✓✓107.4
✓✓✓✓102.8
", + "image_path": "4e19ef76752f0fee538e5f2bf101ae13f8fc866c50b1be0cd6fb616742fe049c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 50, + 530, + 283, + 588 + ], + "blocks": [ + { + "bbox": [ + 46, + 500, + 287, + 523 + ], + "lines": [ + { + "bbox": [ + 46, + 500, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 287, + 523 + ], + "type": "text", + "content": "Table 6. Impact of both augmenters " + }, + { + "bbox": [ + 46, + 500, + 287, + 523 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}^{(E)}" + }, + { + "bbox": [ + 46, + 500, + 287, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 500, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{(I)})" + }, + { + "bbox": [ + 46, + 500, + 287, + 523 + ], + "type": "text", + "content": ", TTP and dynamic learning rate on the SPGSN [26] performance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 530, + 283, + 588 + ], + "lines": [ + { + "bbox": [ + 50, + 530, + 283, + 588 + ], + "spans": [ + { + "bbox": [ + 50, + 530, + 283, + 588 + ], + "type": "table", + "html": "
βHKMPJPE [mm] ↓uMPJPE [mm] ↓
SPGSN [26]+H/P-TTP0.000588103.4SPGSN [26]+H/P-TTP1102.8
0.00166109.22103.4
88102.8
1212107.43104.8
1616105.5
0.001588103.1
", + "image_path": "f75c743ae7d3ecc95fbb4edc84d014a2f2b4365d3f454b5f50ede4099b9f34be.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 625, + 119, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 119, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 119, + 639 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "In this work, we propose a novel helper-predictor test-time personalization (H/P-TTP) framework for 3D human pose forecasting with unseen target subjects. It can be easily integrated into existing deep end-to-end approaches to enable the personalizability of specific target test samples. With the dynamic test-time learning rate, the H/P-TTP fur" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 448, + 544, + 507 + ], + "blocks": [ + { + "bbox": [ + 46, + 588, + 286, + 622 + ], + "lines": [ + { + "bbox": [ + 46, + 588, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 286, + 622 + ], + "type": "text", + "content": "Table 7. Impact of the initial learning rate " + }, + { + "bbox": [ + 46, + 588, + 286, + 622 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 588, + 286, + 622 + ], + "type": "text", + "content": " ,augmentation numbers (H and K) of both " + }, + { + "bbox": [ + 46, + 588, + 286, + 622 + ], + "type": "inline_equation", + "content": "{\\mathcal{A}}_{\\psi }^{\\left( I\\right) }" + }, + { + "bbox": [ + 46, + 588, + 286, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 588, + 286, + 622 + ], + "type": "inline_equation", + "content": "{\\mathcal{A}}^{\\left( E\\right) }" + }, + { + "bbox": [ + 46, + 588, + 286, + 622 + ], + "type": "text", + "content": " ,and the update numbers " + }, + { + "bbox": [ + 46, + 588, + 286, + 622 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 588, + 286, + 622 + ], + "type": "text", + "content": " of the TTP process on MPJPE of the SPGSN [26]+H/P-TTP." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 448, + 544, + 507 + ], + "lines": [ + { + "bbox": [ + 307, + 448, + 544, + 507 + ], + "spans": [ + { + "bbox": [ + 307, + 448, + 544, + 507 + ], + "type": "table", + "html": "
σ2 [mm]MPJPE [mm] ↓MDMPJPE [mm] ↓
SPGSN [26] +HPTP[5,10,15,20,25,30,35,40]103.1186112.4
9109.5
[10,15,20,25,30,35,40,45]102.8369104.5
12102.8
[15,20,25,30,35,40,45,50]103.54812107.4
18106.8
", + "image_path": "70e7ef663f8a4e8927a569c95fa7c260cf0dd2f78701cde418e4289cc36b773f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 507, + 545, + 531 + ], + "lines": [ + { + "bbox": [ + 306, + 507, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 306, + 507, + 545, + 531 + ], + "type": "text", + "content": "Table 8. Impact of the variance levels of noise injected into " + }, + { + "bbox": [ + 306, + 507, + 545, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{(E)}" + }, + { + "bbox": [ + 306, + 507, + 545, + 531 + ], + "type": "text", + "content": ", the size of the memory queue and support set." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 542, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 615 + ], + "type": "text", + "content": "ther considers the degree to which a specific person's motion pattern needs to be personalized. On several benchmarks, experiments demonstrate that the proposed H/P-TTP significantly boosts their prediction performance. Therefore, we reasonably conclude that our model is more practical for unseen test subjects in real-world applications." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 624, + 408, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 624, + 408, + 637 + ], + "spans": [ + { + "bbox": [ + 306, + 624, + 408, + 637 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China (62176125), in part by the Jiangsu Funding Program for Excellent Postdoctoral Talent (2022ZB269), in part by the Natural Science Foundation of Jiangsu Province (BK20220939), and in part by the China Postdoctoral Science Foundation (2022M721629)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "281" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Peshal Agarwal, Danda Pani Paudel, Jan-Nico Zaech, and Luc Van Gool. Unsupervised Robust Domain Adaptation without Source Data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2009–2018, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "type": "text", + "content": "[2] Sadegh Aliakbarian, Fatemeh Saleh, Lars Petersson, Stephen Gould, and Mathieu Salzmann. Contextually plausible and diverse 3d human motion prediction. In ICCV, pages 11333-11342, 2021. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 235 + ], + "type": "text", + "content": "[3] Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. A Stochastic Conditioning Scheme for Diverse Human Motion Prediction. In CVPR, pages 5223-5232, 2020. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 278 + ], + "type": "text", + "content": "[4] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d Human Pose Estimation: New Benchmark and State of The Art Analysis. In CVPR, pages 3686-3693, 2014. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 312 + ], + "type": "text", + "content": "[5] Emad Barsoum, John Kender, and Zicheng Liu. HP-GAN: Probabilistic 3D Human Motion Prediction via GAN. In CVPR, pages 1418-1427, 2018. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 367 + ], + "type": "text", + "content": "[6] Yujun Cai, Lin Huang, Yiwei Wang, Tat-Jen Cham, Jianfei Cai, Junsong Yuan, Jun Liu, Xu Yang, Yiheng Zhu, Xiaohui Shen, et al. Learning Progressive Joint Propagation for Human Motion Prediction. In ECCV, pages 226-242. Springer, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 411 + ], + "type": "text", + "content": "[7] Zhixiang Chi, Yang Wang, Yuanhao Yu, and Jin Tang. Test-Time Fast Adaptation for Dynamic Scene Deblurring via Meta-Auxiliary Learning. In CVPR, pages 9137-9146, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 413, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 287, + 446 + ], + "type": "text", + "content": "[8] Qiongjie Cui and Huaijiang Sun. Towards Accurate 3D Human Motion Prediction From Incomplete Observations. In CVPR, pages 4801-4810, June 2021. 2, 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 490 + ], + "type": "text", + "content": "[9] Qiongjie Cui, Huaijiang Sun, Yupeng Li, and Yue Kong. Efficient human motion recovery using bidirectional attention network. Neural Computing and Applications, 32:10127-10142, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 524 + ], + "type": "text", + "content": "[10] Qiongjie Cui, Huaijiang Sun, and Fei Yang. Learning Dynamic Relationships for 3D Human Motion Prediction. In CVPR, pages 6519-6527, 2020. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[11] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. MSR-GCN: Multi-Scale Residual Graph Convolution Networks for Human Motion Prediction. In ICCV, pages 11467-11476, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "text", + "content": "[12] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Diverse Human Motion Prediction via Gumbel-Softmax Sampling from an Auxiliary Space. In ACM MM, pages 5162–5171, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "text", + "content": "[13] Christian Diller, Thomas A. Funkhouser, and Angela Dai. Forecasting Characteristic 3D Poses of Human Actions. CVPR, pages 15893–15902, 2022. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[14] Zeyu Feng, Chang Xu, and Dacheng Tao. Self-supervised Representation Learning by Rotation Feature Decoupling. In CVPR, pages 10364–10374, 2019. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[15] Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent Network Models for Human Dynamics. In ICCV, pages 4346-4354, 2015. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Anand Gopalakrishnan, Ankur Mali, Dan Kifer, C. Lee Giles, and Alexander Ororbia. A Neural Temporal Model for Human Motion Prediction. In CVPR, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[17] Liang-Yan Gui, Yu-Xiong Wang, Xiaodan Liang, and José M. F. Moura. Adversarial Geometry-Aware Human Motion Prediction. In ECCV, pages 786-803, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[18] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Gerard Pons-Moll, and Christian Theobalt. In The Wild Human Pose Estimation using Explicit 2D Features and Intermediate 3D Representations. In CVPR, pages 10905-10914, 2019. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "text", + "content": "[19] Miao Hao, Yizhuo Li, Zonglin Di, Nitesh B Gundavarapu, and Xiaolong Wang. Test-Time Personalization with a Transformer for Human Pose Estimation. In NeurIPS, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 220, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 275 + ], + "type": "text", + "content": "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The Many Faces of Robustness: A Critical Analysis of Out-of-distribution Generalization. In ICCV, pages 8340-8349, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 276, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 341 + ], + "type": "text", + "content": "[21] Minhao Hu, Tao Song, Yujun Gu, Xiangde Luo, Jieneng Chen, Yinan Chen, Ya Zhang, and Shaoting Zhang. Fully Test-Time Adaptation for Image Segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 251-260. Springer, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 342, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 397 + ], + "type": "text", + "content": "[22] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large Scale Datasets and Predictive Methods for 3D Human Sensing in Natural Environments. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 36:1325-1339, 2014. 1, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 430 + ], + "type": "text", + "content": "[23] Nikos Komodakis and Spyros Gidaris. Unsupervised Representation Learning by Predicting Image Rotations. In ICLR, 2018. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "type": "text", + "content": "[24] Bin Li, Jian Tian, Zhongfei Zhang, Hailin Feng, and Xi Li. Multitask Non-Autoregressive Model for Human Motion Prediction. IEEE Transactions on Image Processing, 2020. 1, 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 545, + 510 + ], + "type": "text", + "content": "[25] Chen Li, Zhen Zhang, Wee Sun Lee, and Gim Hee Lee. Convolutional Sequence to Sequence Model for Human Dynamics. In CVPR, pages 5226-5234, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "text", + "content": "[26] Maosen Li, Siheng Chen, Zijing Zhang, Lingxi Xie, Qi Tian, and Ya Zhang. Skeleton-Parted Graph Scattering Networks for 3D Human Motion Prediction. arXiv preprint arXiv:2208.00368, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 555, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 600 + ], + "type": "text", + "content": "[27] Maosen Li, Siheng Chen, Yangheng Zhao, Ya Zhang, Yanfeng Wang, and Qi Tian. Dynamic Multiscale Graph Neural Networks for 3D Skeleton Based Human Motion Prediction. In CVPR, pages 214-223, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "text", + "content": "[28] Yang Li, Kan Li, Shuai Jiang, Ziyue Zhang, Congzhen-tao Huang, and Richard Yi Da Xu. Geometry-driven self-supervised method for 3d human pose estimation. In AAAI, volume 34, pages 11442-11449, 2020. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[29] Mengyuan Liu, Fanyang Meng, and Yongsheng Liang. Generalized pose decoupled network for unsupervised 3d skeleton sequence-based action representation learning. Cyborg and Bionic Systems, 2022:0002, 2022. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[30] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "282" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "When Does Self-supervised Test-time Training Fail or Thrive? NeurIPS, 34:21808-21820, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[31] Tiezheng Ma, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Progressively Generating Better Initial Guesses Towards Next Stages for High-Quality Human Motion Prediction. In CVPR, pages 6437-6446, 2022. 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[32] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. History Repeats Itself: Human Motion Prediction via Motion Attention. In ECCV, pages 474-489, 2020. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 217 + ], + "type": "text", + "content": "[33] Wei Mao, Miaomiao Liu, and Mathieu Salzmann. Generating Smooth Pose Sequences for Diverse Human Motion Prediction. In ICCV, 2021. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 252 + ], + "type": "text", + "content": "[34] Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning Trajectory Dependencies for Human Motion Prediction. In ICCV, pages 9489-9497, 2019. 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 286 + ], + "type": "text", + "content": "[35] Julieta Martinez, Michael J Black, and Javier Romero. On Human Motion Prediction using Recurrent Neural Networks. In CVPR, pages 2891-2900, 2017. 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "type": "text", + "content": "[36] Angel Martínez-González, Michael Villamizar, and Jean-Marc Odobez. Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive Transformers. In ICCV, pages 2276-2284, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "type": "text", + "content": "[37] Ishan Misra and Laurens van der Maaten. Self-supervised Learning of Predict-invariant Representations. In CVPR, pages 6707-6717, 2020. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 366, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 287, + 421 + ], + "type": "text", + "content": "[38] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating Prediction-time Batch Normalization for Robustness under Covariate Shift. arXiv preprint arXiv:2006.10963, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[39] Dario Pavllo, Christoph Feichtenhofer, Michael Auli, and David Grangier. *Quaternion: A Quaternion-based Recurrent Model for Human Motion*. International Journal of Computer Vision, (128):855-872, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 468, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 509 + ], + "type": "text", + "content": "[40] Senthil Purushwalkam and Abhinav Gupta. Demystifying Contrastive Self-supervised Learning: Invariances, Augmentations and Dataset Biases. NeurIPS, 33:3407-3418, 2020. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "text", + "content": "[41] Alejandro Hernandez Ruiz, Juergen Gall, and Francesc Moreno-Noguer. Human Motion Prediction via Spatio-Temporal Inpainting. In CVPR, pages 7134-7143, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "text", + "content": "[42] Chajin Shin, Taeoh Kim, Sangjin Lee, and Sangyoun Leey. Test-Time Adaptation for Out-Of-Distributed Image Inpainting. In ICIP, 2021. 1, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "type": "text", + "content": "[43] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International Journal of Computer Vision, 87(1):4-27, 2010. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[44] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time Training with Self-supervision for Generalization under Distribution Shifts. In ICML, pages 9229-9248. PMLR, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[45] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A Dataset of Whole-Body Human Grasping of Objects. In ECCV, 2020. 6, 7, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 677 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[46] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain Randomization for Transferring Deep Neural Networks from Simulation to The Real World. In IROS, pages 23-30, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[47] Jonathan Tremblay, Aayush Prakash, David Acuna, Mark Brophy, Varun Jampani, Cem Anil, Thang To, Eric Cameracci, Shaad Boochoon, and Stan Birchfield. Training deep networks with synthetic data: Bridging the reality gap by domain randomization. In CVPRW, pages 969-977, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "text", + "content": "[48] Thomas Varsavsky, Mauricio Orbes-Arteaga, Carole H Sudre, Mark S Graham, Parashkev Nachev, and M Jorge Cardoso. Test-time Unsupervised Domain Adaptation. In MIC-CAI, pages 428–436. Springer, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "text", + "content": "[49] Borui Wang, Ehsan Adeli, Hsu-kuang Chiu, De-An Huang, and Juan Carlos Niebles. Imitation Learning for Human Pose Prediction. In ICCV, pages 7124-7133, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "text", + "content": "[50] Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, and Trevor Darrell. On-target adaptation. arXiv preprint arXiv:2109.01087, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "text", + "content": "[51] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully Test-time Adaptation By Entropy Minimization. arXiv preprint arXiv:2006.10726, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "text", + "content": "[52] Qin Wang, Dengxin Dai, Lukas Hoyer, Luc Van Gool, and Olga Fink. Domain adaptive semantic segmentation with self-supervised depth estimation. In ICCV, pages 8515-8525, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 376, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 407 + ], + "type": "text", + "content": "[53] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual Test-time Domain Adaptation. In CVPR, pages 7201-7211, 2022. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 410, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 441 + ], + "type": "text", + "content": "[54] Sirui Xu, Y-X Wang, and L-Y Gui. Diverse Human Motion Prediction Guided by Multi-level Spatial-temporal Anchors. In ECCV, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 444, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 486 + ], + "type": "text", + "content": "[55] Fu-En Yang, Yuan-Chia Cheng, Zu-Yun Shiau, and Yu-Chiang Frank Wang. Adversarial Teacher-student Representation Learning for Domain Generalization. NeurIPS, 34:19448-19460, 2021. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "type": "text", + "content": "[56] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A Fourier Perspective on Model Robustness in Computer Vision. NeurIPS, 32, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "text", + "content": "[57] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial Examples: Attacks and Defenses for Deep Learning. IEEE transactions on neural networks and learning systems, 30(9):2805-2824, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "type": "text", + "content": "[58] Ye Yuan and Kris Kitani. Dlow: Diversifying Latent Flows for Diverse Human Motion Prediction. In ECCV, pages 346-364, 2020. 1, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "text", + "content": "[59] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test Time Robustness via Adaptation and Augmentation. arXiv preprint arXiv:2110.09506, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 677 + ], + "type": "text", + "content": "[60] Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. Deep Domain-adversarial Image Generation for Domain Generalisation. In AAAI, volume 34, pages 13025-13032, 2020. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "283" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/8c410ac7-5eb0-4f17-a4bb-daadb2446edc_content_list.json b/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/8c410ac7-5eb0-4f17-a4bb-daadb2446edc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..75da5668771ef01c35caaa2ca57f43d06eed29f2 --- /dev/null +++ b/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/8c410ac7-5eb0-4f17-a4bb-daadb2446edc_content_list.json @@ -0,0 +1,1697 @@ +[ + { + "type": "text", + "text": "Tetra-NeRF: Representing Neural Radiance Fields Using Tetrahedra", + "text_level": 1, + "bbox": [ + 138, + 130, + 831, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jonas Kulhanek \nCzech Technical University in Prague \njonas.kulhanek@cvut.cz", + "bbox": [ + 158, + 181, + 460, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Torsten Sattler \nCzech Technical University in Prague \ntorsten.sattler@cvut.cz", + "bbox": [ + 506, + 181, + 807, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d718ca298af7923b76dae45526cd7fc1c5c03f76746559ba4400da2025a3561.jpg", + "image_caption": [ + "3) barycetric interpolation of features 4) shallow MLP predicts density and colour", + "Figure 1. The input to Tetra-NeRF is a point cloud which is triangulated to get a set of tetrahedra used to represent the radiance field. Rays are sampled, and the field is queried. Barycentric interpolation is used to interpolate tetrahedra vertices, and the resulting features are passed through a shallow MLP to get the density and colours for volumetric rendering." + ], + "image_footnote": [], + "bbox": [ + 181, + 247, + 598, + 417 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 484, + 313, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRFs) are a very recent and very popular approach for the problems of novel view synthesis and 3D reconstruction. A popular scene representation used by NeRFs is to combine a uniform, voxel-based subdivision of the scene with an MLP. Based on the observation that a (sparse) point cloud of the scene is often available, this paper proposes to use an adaptive representation based on tetrahedra obtained by Delaunay triangulation instead of uniform subdivision or point-based representations. We show that such a representation enables efficient training and leads to state-of-the-art results. Our approach elegantly combines concepts from 3D geometry processing, triangle-based rendering, and modern neural radiance fields. Compared to voxel-based representations, ours provides more detail around parts of the scene likely to be close to the surface. Compared to point-based representations, our approach achieves better performance. The source code is publicly available at: https://jkulhanek.com/tetra-nerf.", + "bbox": [ + 75, + 516, + 473, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 814, + 209, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing 3D scenes from images and rendering photo-realistic novel views is a key problem in computer vision. Recently, NeRFs [3,4,39] became dominant in the field for their superior photo-realistic results. Originally,", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NeRFs used MLPs to represent the 3D scene as an implicit function. Given a set of posed images, NeRF randomly samples a batch of pixels, casts rays from the pixels into the 3D space, queries the implicit function at randomly sampled distances along the rays, and aggregates the sampled values using volumetric rendering [37, 39]. While the visual results of such methods are of high quality, the problem is that querying large MLPs at millions of points is costly. Also, once the network is trained, it is difficult to make any changes to the represented radiance field as everything is baked into the MLPs parameters, and any change has a non-local effect. Since then, there have been a lot of proposed alternatives to the large MLP field representation [9, 17, 32, 41, 57, 70, 74]. These methods combine an MLP with a voxel feature grid [41, 57], or in some cases represent the radiance field directly as a tensor [9, 10, 17]. When querying these representations, first, the containing voxel is found, and the features stored at the eight corner points of the voxel are trilinearly interpolated. The result is either passed through a shallow MLP [9, 10, 41, 57] or is used directly as the density and colour [17, 32, 57].", + "bbox": [ + 496, + 486, + 892, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Having a dense tensor represent the entire scene is very inefficient, as we only need to represent a small space around surfaces. Therefore, different methods propose different ways of tackling the issue. Instant-NGP [41], for example, uses a hash grid instead of a dense tensor, where it relies on optimisation to resolve the hash collisions. How-", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18458", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ever, similarly to MLPs, any change to the stored hashmap influences the field in many places. A more common direction to addressing the issue is by directly using a sparse tensor representation [9, 17]. These methods start with a low-resolution grid and, at predefined steps, subsample the representation, increasing the resolution. These approaches tend to require a careful setting of hyperparameters, such as the scene bounding box and the subdivision steps, in order for the methods to work well.", + "bbox": [ + 75, + 90, + 472, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Because many of these methods use traditional structure from motion (SfM) [54, 55] methods to generate the initial poses for the captured images, we can reuse the original reconstruction in the scene representation. Inspired by classical surface reconstruction methods [22, 23, 29, 30], we represent the scene as a dense triangulation of the input point cloud, where the scene is a set of non-overlapping tetrahedra whose union is the convex hull of the original point cloud [14]. When querying such a representation, we find to which tetrahedron the query point belongs and perform barycentric linear interpolation of the features stored in the vertices of the tetrahedron. This very simple representation can be thought of as the direct extension of the classical triangle-rendering pipelines used in graphics [40, 43, 45]. The representation avoids problems with the sparsity of the input point cloud as the tetrahedra fully cover the scene, resulting in a continuous rather than discrete representation.", + "bbox": [ + 75, + 227, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper makes the following contributions: (1) We propose a novel radiance field representation which is initialised from a sparse or dense point cloud. This representation is naturally denser in the proximity of surfaces and, therefore, provides a higher resolution in these regions. (2) The proposed representation is evaluated on multiple synthetic and real-world datasets and is compared with a state-of-the-art point-cloud-based representation - Point-NeRF [70]. The presented results show that our method clearly outperforms this baseline. We further demonstrate the effectiveness of our adaptive representation by comparing it with a voxel-based representation that uses the same number of trainable parameters. (3) We make the source code and model checkpoints publicly available.", + "bbox": [ + 75, + 484, + 472, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 708, + 215, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-view reconstruction. The problem of multi-view reconstruction has been studied extensively and tackled with a variety of structure from motion (SfM) [54, 61, 63], and multi-view stereo (MVS) [12, 18, 55, 71] methods. These methods usually output the scene represented as a point cloud [54, 55]. In most rendering approaches, the point cloud is converted into a mesh [24, 35], and novel views are rendered by reprojecting observed images into each novel viewpoint and blending them together using ei-", + "bbox": [ + 75, + 738, + 468, + 875 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ther heuristically-defined [8, 13, 69] or learned [20, 50, 51, 73] blending weights.", + "bbox": [ + 500, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the process of getting the meshes is usually quite noisy, and the resulting meshes tend to have inaccurate geometry in regions with fine details or complex materials. Instead of using noisy meshes, point-based neural rendering methods [1, 26, 38, 53] perform splatting of neural features and use 2D convolutions to render them. In contrast to these methods, our approach operates and aggregates features directly in 3D and does not suffer from the noise in the point cloud or the reconstructed mesh.", + "bbox": [ + 496, + 125, + 893, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural radiance fields. Recently, NeRFs [3,4,34,39,77] have gained a lot of attention thanks to their high-quality rendering performance. The original NeRF method [39] was extended to better handle aliasing artefacts in [3], to better represent unbounded scenes in [4,48,77], or to handle real-world captured images [36,58]. The training of the large MLPs used in these methods can be quite slow, and there has been a lot of effort on speeding up either the training [9,17,41] or the rendering [21,47,48,74] sometimes at the cost of larger storage requirements. Other approaches tackled different aspects of NeRFs like view-dependent artefacts [62], relighting [5,7], or proposed generative models [46,64]. Also, a popular research direction is making the models generalize across different scenes [11,28,49,66,75]. A large area of research is dedicated to the surface reconstruction and, instead of using the radiance fields, represents the scene implicitly by modelling the signed distance function (SDF) [52,65,67,72,76]. Unlike those approaches, we focus only on the radiance field representation and consider these methods orthogonal to ours.", + "bbox": [ + 496, + 268, + 892, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although there are some methods that train radiance fields while fine-tuning the poses [31, 60] or without known cameras [6, 68], most methods need camera poses for the reconstruction. SfM, e.g., COLMAP [54, 55], is typically used for estimating the poses, which also produces a (sparse) point cloud. Our approach makes use of this by-product of the pose recovery process instead of only using the poses themselves.", + "bbox": [ + 496, + 574, + 893, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Field representations. When a single MLP is used to represent the entire scene, everything is baked into a single set of parameters which cannot be easily modified, because any change to the parameters has a non-local effect, i.e., it changes the scene at multiple unrelated places. To overcome this problem, others have experimented with different representations of the radiance fields [9, 10, 17, 32, 41, 44, 57, 70]. A common practice is to represent the scene as a shallow MLP and an efficiently encoded voxel grid [9, 10, 32, 41, 44]. The encoded voxel grid can also be used represent the radiance field directly [17, 57]. The voxel grid can be encoded as a sparse tensor [17, 32, 57], a factorisation of the 4D tensor [9, 10], or a hashmap [41]. When", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1https://github.com/jkulhanek/tetra-nerf", + "bbox": [ + 99, + 886, + 418, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18459", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7abd8e15c57f18ac8cd3c30e5301bc371167faebce01af3348e4d354d3156e33.jpg", + "image_caption": [ + "(a) dense point cloud" + ], + "image_footnote": [], + "bbox": [ + 93, + 90, + 272, + 175 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7863d3ad9175ea27038d1cfdc437f43a530d2fdd5cf19f167a5d9461cf3284e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 92, + 442, + 169 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5f74a44cddba3fcb15c66ff5be32b246eb46c7aa4a1dd90de9b54feda4e1b0a2.jpg", + "image_caption": [ + "(c) sparse point cloud" + ], + "image_footnote": [], + "bbox": [ + 93, + 191, + 272, + 277 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/326c524dd6484d7b5f579d4b15d08a3f0aa3570b24e5bfc89431dc848ea7bda9.jpg", + "image_caption": [ + "(b) dense tetrahedra slice", + "(d) sparse tetrahedra slice", + "Figure 2. Input point cloud and a slice through the triangulated tetrahedra. Note that smaller tetrahedra are created closer to the surface of the scene, i.e., regions close to the surface are represented with a finer resolution.", + "Fig. 2). In the following, we explain how this adaptive subdivision of the scene can be used instead of voxels for volume rendering and neural rendering." + ], + "image_footnote": [], + "bbox": [ + 274, + 191, + 454, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "these structures are queried, trilinear interpolation is used to combine the feature vectors stored in the containing voxel corners. Unfortunately, the hashes [41] and hierarchical representations [10] have the same non-local effect problem, and the rest of the approaches rely on subsequent upsampling of the field and can be overly complicated. Also, feature vectors cannot be placed arbitrarily in the 3D space as they must lie on the grid with a fixed resolution. In contrast, our approach is much more flexible as it stores the feature vectors freely in 3D space.", + "bbox": [ + 75, + 358, + 468, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, Point-NeRF [70] represents the scene as a point cloud of features which are queried using $k$ -nearest neighbours search. However, when the point cloud contains sparse regions, the rays do not intersect any neighbourhoods of any points and the pixels stay empty without the ability to optimise. Therefore, Point-NeRF [70] relies on gradually adding more points during training and increasing the scene complexity. Since we use the triangulation of the point cloud rather than the discrete point cloud itself, our representation is continuous and does not suffer from empty regions. Therefore, we do not have to add any points during training.", + "bbox": [ + 75, + 510, + 468, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 707, + 168, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A common strategy in the literature is to represent the scene explicitly through a voxel volume. In contrast to this uniform subdivision, we investigate using an adaptive subdivision of the scene. In many scenarios, an approximation of the scene geometry is either given, e.g., when using SfM to compute the input camera poses, or can be computed, e.g., via MVS or single-view depth predictions [79]. This allows us to compute an adaptive subdivision of the scene via Delaunay triangulation [14] of such a point cloud. This results in a set of non-overlapping tetrahedra, where smaller tetrahedra are created close to the surface of the scene (c.f.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7898f45a2777e4512c769d742cb35004c7be94bd2da66bdbad9dec6738eb60b0.jpg", + "image_caption": [ + "(a) trilinear interpolation", + "Figure 3. Trilinear and barycentric interpolation. Trilinear interpolation is a weighted combination of the eight voxel corners. Barycentric interpolation weights the four vertices of the tetrahedron vertices based on the barycentric coordinates [16]." + ], + "image_footnote": [], + "bbox": [ + 535, + 90, + 714, + 199 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e7480085e9cc46610d1b010b85b34b8598fda893d58e624a4e3789cb494ef03e.jpg", + "image_caption": [ + "(b) barycentric interpolation" + ], + "image_footnote": [], + "bbox": [ + 718, + 90, + 856, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 342, + 638, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRFs) [39] represent a scene through an implicit function $F(\\mathbf{x}, \\mathbf{d})$ , often modelled via a neural network, that returns a colour value and a volume density prediction for a given 3D point $\\mathbf{x}$ in the scene observed from a viewing direction $\\mathbf{d}$ . Volume rendering [37] is used to synthesise novel views: for each pixel in a virtual view, we project a ray from the camera plane into the scene and sample the radiance field to obtain the colour and density values $\\mathbf{c}_i$ and $\\sigma_i$ at distances $t_i$ along the ray, where $i = 1 \\dots N$ . The individual samples are then combined to predict the colour $\\mathbf{C}$ for the pixel:", + "bbox": [ + 496, + 366, + 892, + 532 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} = \\sum_ {i = 1} ^ {N} T _ {i} \\left(1 - \\exp \\left(\\sigma_ {i} \\delta_ {i}\\right)\\right) \\mathbf {c} _ {i}, \\text {w h e r e} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 545, + 890, + 594 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} = \\exp \\left(- \\sum_ {i = 1} ^ {N - 1} \\sigma_ {i} \\delta_ {i}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 589, + 730, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and $\\delta_{i} = t_{i} - t_{i - 1}$ is the distance between adjacent samples. This process is fully differentiable. Thus, after computing the mean squared error (MSE) between the predicted and ground truth colours associated with each ray, we can backpropagate the gradients to the radiance field.", + "bbox": [ + 496, + 643, + 890, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Voxel-based feature fields such as NSVF [32] represent the scene as a voxel grid and an MLP. Each grid point of the voxel grid is assigned a trainable vector. There are eight feature vectors associated with a single voxel, but the vectors are shared between neighbouring voxels. For each query point sampled along the ray, its corresponding voxel is found first. A feature for the point is then computed via trilinear interpolation of the features of the voxel grid points based on the position of the query point (c.f. Fig. 3, left). The resulting feature vector is passed through an MLP to predict the density and appearance vector. The appearance vector is combined with the ray direction and", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18460", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "passed through a second MLP in order to compute a view-dependent colour.", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Tetrahedra fields", + "text_level": 1, + "bbox": [ + 76, + 128, + 243, + 142 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a set of points in 3D space, we build the tetrahedra structure by triangulating the points. We apply the Delaunay triangulation [14] to obtain a set of non-overlapping tetrahedra whose union is the convex hull of the original points. Fig. 2 shows example tetrahedra obtained by triangulating dense and sparse COLMAP-reconstructed point clouds [54, 55]. Note that the resulting representation is adaptive as it uses a higher resolution (smaller tetrahedra) closer to the surface and larger tetrahedra to model regions farther away from the surface.", + "bbox": [ + 75, + 151, + 468, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We associate all vertices of the tetrahedra with trainable vectors. As in the voxel grid case, vertices, and thus vectors, are shared between adjacent tetrahedra. The resulting tetrahedra field can be queried in the same way a voxel grid representation is queried: for each query point, we first find the tetrahedron containing the point. Instead of the trilinear interpolation used for voxel volumes, we use the barycentric interpolation [16] to compute a feature vector for the query point from the four feature vectors stored at the tetrahedron's vertices (c.f. Fig. 3). To this end, we compute the query point $\\mathbf{x}$ 's barycentric coordinates $\\lambda$ , which express the point's 3D coordinates as a unique weighted combination of the 3D coordinates of the tetrahedron's vertices. In particular, the weight for a vertex is the volume of the tetrahedron constructed from the query point and the face opposite to the vertex divided by the volume of the full tetrahedron:", + "bbox": [ + 75, + 301, + 468, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\lambda} = \\left(\\frac {V _ {x 2 3 4}}{V _ {1 2 3 4}}, \\frac {V _ {1 x 3 4}}{V _ {1 2 3 4}}, \\frac {V _ {1 2 x 4}}{V _ {1 2 3 4}}, \\frac {V _ {1 2 3 x}}{V _ {1 2 3 4}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 561, + 468, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V_{1234}$ is the volume of the full tetrahedron and $V_{x234}$ , $V_{1x34}, \\ldots$ , are volumes of tetrahedra with $1^{\\mathrm{st}}$ , $2^{\\mathrm{nd}}$ , ..., vertex replaced by $\\mathbf{x}$ . The same weights $\\lambda$ are applied to the feature vectors of the vertices to obtain the query feature.", + "bbox": [ + 76, + 598, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The interpolated features are used as the input to a small MLP in order to predict density and colour at the query point. We first pass the barycentric-interpolated features through a three-layer MLP to compute the density and appearance features. We then concatenate the appearance features with the ray direction vector encoded using Fourier features [39, 59] and pass the result through a single linear layer to get the raw RGB colour value.", + "bbox": [ + 75, + 659, + 468, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, to map the raw density values $\\bar{\\sigma}_i$ returned by the network to the volume density $\\sigma_{i}$ required by volume rendering, we apply the softplus activation function [3]. For the RGB colour value, we use the sigmoid function [17].", + "bbox": [ + 75, + 779, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Efficiently querying a tetrahedra field", + "text_level": 1, + "bbox": [ + 76, + 847, + 403, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Determining the corresponding voxel for a given query point can be done highly efficiently via hashing [42]. In", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "contrast, finding the corresponding tetrahedron for a query point is more complex. This in turn can significantly impact rendering, and thus training, efficiency [17,41,57].", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to efficiently look up the corresponding tetrahedra, we exploit that we are not considering isolated points, but points sampled from a ray. We compute the tetrahedra that are intersected by the ray, allowing us to march through the tetrahedra rather than computing them individually per point. The relevant tetrahedra can be found efficiently using acceleration structures for fast ray-triangle intersection computations, e.g., via NVidia's OptiX library [43]: We first compute the intersections between the rays from a synthetic view and all faces of the tetrahedra. For each ray, we take the first 512 intersected triangles², and determine the corresponding tetrahedra. The tetrahedra can be ordered based on the intersections along the ray, allowing us to easily march through the tetrahedra to determine which tetrahedron to use for a given query point.", + "bbox": [ + 496, + 136, + 890, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A side benefit of computing ray-triangle intersections is that we can simplify computing the barycentric coordinates: For each intersection, we compute the 2D barycentric coordinates w.r.t. the triangle. We then obtain the 3D barycentric coordinates w.r.t. the associate tetrahedron by simply adding zero for the vertex opposite to the triangle. For a query point inside a tetrahedron, we can compute its tetrahedron barycentric coordinates by linearly interpolating between the barycentric coordinates of the two intersections of the ray and the tetrahedron.", + "bbox": [ + 496, + 363, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Coarse and fine sampling", + "text_level": 1, + "bbox": [ + 500, + 521, + 730, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow the common practice of having a two-stage sampling procedure [9, 39]. In the coarse stage, we sample uniformly along the ray. In the fine stage, we use the density weights from the coarse sampling stage to bias the sampling towards sampling closer to the potential surface. Following [39], we use the stratified uniform sampling for the coarse stage. The stratified uniform sampling splits the ray into equally long intervals and samples uniformly in each interval. Unlike NeRF [39], we limit the sampling to the space occupied by tetrahedra. In the fine sampling stage, we use the same network as in the coarse sampling stage.", + "bbox": [ + 496, + 545, + 890, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the fine sampling stage, we take the accumulated weights $w_{i}$ from the coarse sampling:", + "bbox": [ + 498, + 710, + 890, + 741 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {w} _ {i} = \\left(1 - \\exp \\left(- \\sigma_ {i} \\delta_ {i}\\right)\\right) \\exp \\left(- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {j} \\delta_ {j}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 750, + 890, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These weights are the coefficients used in Equation 1 as multipliers for the colours [39]. We obtain $w_{i}$ by normalizing $\\bar{w}_{i}$ . Following [39], we sample set of fine samples using", + "bbox": [ + 498, + 797, + 890, + 844 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2For efficiency, we only consider a fixed number of triangles. As discussed later on, this can degrade results in larger scenes / scenes with a fine-grained tetrahedralisation, where more intersections are needed. Naturally, more triangles can be considered at the cost of longer run-times.", + "bbox": [ + 498, + 849, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18461", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "weight $w_{i}$ . To render the final colour, we merge the dense and fine samples and use all in the rendering equation [9].", + "bbox": [ + 76, + 90, + 468, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 133, + 209, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare Tetra-NeRF to relevant methods on the commonly used synthetic Blender [39], the real-world Tanks and Temples [25], and the challenging object-centric Mip-NeRF 360 [4] datasets. To show its effectiveness, we compare Tetra-NeRF to a dense-grid representation and evaluate it with reduced quality of the input point cloud. We start by describing the exact hyperparameters used.", + "bbox": [ + 75, + 159, + 468, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 273, + 290, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generating point cloud & triangulation. Given a set of posed images, we use the COLMAP reconstruction pipeline [54, 55] to get the point cloud used in our tetrahedra field representation. We then reduce the size of the resulting point cloud such that if the number of points is larger than $10^{6}$ , we subsample $10^{6}$ points randomly. For the Blender dataset experiments, where the number of points is smaller, we add more randomly generated points. In that case, the number of random points is half the number of original points. The reason for adding the points is that with a low number of points, some pixels on edges may not intersect any tetrahedra, potentially producing artefacts on the edges. Each added point is sampled as follows: we sample a random point $x_{0}$ from the original point cloud, sample a random normal vector $n$ , and a number $\\alpha \\sim \\mathcal{N}(\\bar{d}, \\bar{d}^{2})$ , where $\\bar{d}$ is the average spacing of the original point cloud, i.e. the average distance between each point and its six closest neighbours. We then add the point $x = x_{0} + \\alpha n$ .", + "bbox": [ + 75, + 296, + 470, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initialization. Given the processed point cloud, we apply the Delaunay triangulation [14] to get a set of tetrahedra. To this end, we use the CGAL library [2], which in our experiments runs in the order of milliseconds. Following [41], we initialise the features of size 64 at the vertices of the tetrahedra with small values sampled uniformly in the range $-10^{-4}$ to $10^{-4}$ . However, to allow the model to reuse the information contained in the point cloud, we set the first four dimensions of the feature field to the RGBA colour values (rescaled to interval [0, 1]) stored at the associated points in the point cloud. The alpha value of all original points is one, whereas all randomly sampled points have an alpha value of zero. For the MLP, we follow the common practice of using the Kaiming uniform initialisation [19]. The hidden sizes in all MLPs are 128.", + "bbox": [ + 75, + 569, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. During training, we sample batches of 4,096 random rays from random training images. We use volumetric rendering to predict the colour of each ray. The gradients are computed by backpropagating the MSE loss between the predicted colour and the ground truth colour value. We use the RAdam optimizer [33] and decay the learning rate exponentially from $10^{-3}$ to $10^{-4}$ in 300k steps. The train-", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0b7aded02f932d813f11ed50c31f54d24f4c0743941531ff638f697c43e5e8af.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓
NeRF [39]31.000.9470.081
NSVF [32]31.770.953-
mip-NeRF [3]34.510.9610.043
instant-NGP [41]33.18--
Plenoxels [17]31.710.9580.049
Point-NeRFcol [70]31.770.9730.062
Point-NeRFmvs [70]33.310.9780.049
Tetra-NeRF32.530.9820.041
", + "bbox": [ + 524, + 88, + 867, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Results on the Blender dataset [39] averaged over all scenes in the dataset. Even though we use the same input point cloud as Point-NeRF $^{col}$ , we outperform it greatly. We perform on par with Point-NeRF $^{mvs}$ even though it uses many more points and densifies the point cloud during training. We highlight the best, second, and third values.", + "bbox": [ + 498, + 224, + 893, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing code is built on top of the Nerfstudio framework [60], and the tetrahedra field is implemented in CUDA and uses the OptiX library [43]. We train on a single NVIDIA A100 GPU and the training speed ranges from 15k rays per second to 30k rays per second. The speed depends on how well-structured the triangulation is, how many vertices there are, and if there is empty space around the object. The full training with 300k iterations takes between 11 and 24 hours, depending on the scene complexity. However, good results are typically obtained much earlier, e.g., in 100k iterations.", + "bbox": [ + 496, + 328, + 890, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 500, + 492, + 591, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Blender dataset [39] results. We compare to relevant baselines on the standard Blender dataset. We used the same split and evaluation procedure as in the original NeRF paper [39] and the same SSIM implementation as Point-NeRF [70]. In order to ensure a fair comparison with Point-NeRF [70] when COLMAP points were used, we use the exact same COLMAP reconstruction as Point-NeRF. We report the PSNR, SSIM, and LPIPS (VGG) [78] metrics. Tab. 1 shows averaged results, Fig. 4 shows qualitative results, and the results for individual scenes are given in Supp. Mat.", + "bbox": [ + 496, + 521, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When we use the exact same COLMAP points as Point-NeRF (row $\\text{Point-NeRF}^{col}$ ), we outperform it in all three metrics. We score comparably with $\\text{Point-NeRF}^{mvs}$ , even though it starts from a much denser and higher-quality initial point cloud, which it generates from a jointly trained model. Also note that both of these Point-NeRF configurations grow the point cloud during training and, therefore, the complexity of the scene representation grows. For us, the points are fixed, and the number of parameters stays the same. We also outperform Plenoxels [17], which uses a sparse grid. Note that same as Point-NeRF, Plenoxels also gradually increases the representation complexity by subdividing the grid resolution at predefined training epochs. Even though both Mip-NeRF and instant-NGP outperform", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18462", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6bfebca5a0ce9a87375b85378673e4954bfba8d593be0a01e5e7e69b39c99b10.jpg", + "image_caption": [ + "Figure 4. Qualitative results on the Blender dataset. We compare with Point-NeRF $^{col}$ as we use the same input point cloud. In the top-row picture, we can see that Tetra-NeRF is able to represent fine details well on the mic scene. On the drums scene ( $2^{\\text{nd}}$ row), both methods struggle with shiny materials, but our method performs slightly better. $3^{\\text{rd}}$ row: Tetra-NeRF can render thin ropes. Bottom row: accumulation is nonzero in areas close to surface." + ], + "image_footnote": [], + "bbox": [ + 83, + 93, + 467, + 404 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/abc0b7ef8c4c9fdbb9facec634b10180697b99e9f215bf439c5c925821114fb7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
hotdogship
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
Point-NeRFstatic29.910.97819.350.905
Tetra-NeRF33.310.98931.130.994
", + "bbox": [ + 94, + 527, + 449, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "our approach in terms of PSNR, our method is slightly better in terms of SSIM, and on par with Mip-NeRF in terms of LPIPS.", + "bbox": [ + 75, + 671, + 468, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To analyze the tetrahedra field representation, we compare it with the point cloud field representation used in Point-NeRF. In Table 2, we compare our method to Point-NeRF when we disable point cloud growth and pruning. We show the results on two scenes from the Blender dataset [39] which were selected in the Point-NeRF paper. With this setup, we vastly outperform Point-NeRF in all metrics. The reason is that Point-NeRF requires the point cloud to be dense such that all rays have a chance to intersect a neighbourhood of a point. Since we use a continuous representation (tetrahedra field) rather than a discrete one, we can achieve good results even for sparser point clouds.", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a0f6006fce6cb015a0c296a9650831deb6ac66ff2fcd5039259d67e2e6fa1efa.jpg", + "image_caption": [ + "Figure 5. Comparison with dense grid field representation. Tetra-NeRF on the left, dense grid on the right. We used a comparable number of parameters for both methods (the dense grid having slightly more parameters). Due to the adaptive nature of tetrahedra fields, Tetra-NeRF produces significantly better rendering as it is able to focus on relevant parts of the scene." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 694, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1eb0eafcf005a95649a4bacc914c609ccbd5dbdc50537d3010005cc416078e94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 90, + 885, + 193 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fd041eee811d758b315593b07a68d139ffe3d18fc98a239dfd4b63827d022a56.jpg", + "table_caption": [ + "Table 2. Comparison with Point-NeRF with disabled point cloud growing and pruning shows that Point-NeRF performs significantly worse in all measured metrics because it struggles to handle the sparse point clouds." + ], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓
NV [34]23.700.8480.260
NeRF [39]25.780.8640.198
NSVF [32]28.400.9000.153
Point-NeRF [70]*28.350.9420.090
Tetra-NeRF28.900.9570.059
", + "bbox": [ + 532, + 290, + 859, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Result on the Tanks and Temples dataset [25] as processed by NSVF [32] averaged over all scenes. We show the PSNR, SSIM, and LPIPS (Alex) [78] metrics. We highlight the best, second, and third values. *Note, that Point-NeRF [70] results differ from the paper as they were recomputed with the resolution used in other methods.", + "bbox": [ + 498, + 387, + 893, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with the dense grid representation. In order to show the utility of the adaptive tetrahedra field representation, we compare it to a dense grid representation. Similarly to NSVF [32], we split the 3D scene bounding box into equally-sized voxels. When querying the field, we find the voxel to which the query point belongs and perform trilinear interpolation of the eight corners of the voxel. We choose the grid resolution such that the number of grid points is the lowest cube number larger than the number of points of the original point cloud. This ensures a fair comparison as the baseline uses a comparable (but larger) number of features than our approach. All other hyperparameters are kept the same as for Tetra-NeRF. We evaluate both approaches on two scenes from the Blender dataset [39]. The dense grid representation only scores PSNR 18.81 and 18.91 on the lego and ship scenes, respectively, whereas Tetra-NeRF scores PSNR 33.79 and 30.69, respectively. The results can be seen in Figure 5. Note that in this experiment, we only trained for 100,000 iterations to save computation time. From the numbers and the figure, we can clearly see that the dense grid resolution is not sufficient to reconstruct the scene in enough detail. Because Tetra-NeRF uses an adaptive subdivision, which is more detailed around the surface, we are better able to focus on relevant scene parts. With a similar number of trainable parameters, we thus obtain better results.", + "bbox": [ + 496, + 487, + 893, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tanks and Temples [25] dataset. To be able to compare", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18463", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/de87efc8bcf436104c7f063ae6d9ba29f1a8b57a0da7a3ea0fa8834ca562da6e.jpg", + "image_caption": [ + "Figure 6. Results on Tanks and Temples dataset. In the top row, we can see that we are able to represent the rim of the wheel better than Point-NeRF. Similarly, we can represent the tyre better in the third row. Finally, in the bottom, Point-NeRF fails on the wall and the roof, whereas Tetra-NeRF can render these parts well." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 465, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with Point-NeRF [70] on real-world data, we have evaluated Tetra-NeRF on the Tanks and Temples [25] dataset. We use the same setup as in NSVF [32], where the object is masked. We report the usual metrics: PSNR, SSIM, and LPIPS (Alex).3 We used the dense COLMAP reconstruction to get the point cloud used in the tetrahedra field. Quantitative results are shown in Table 3, qualitative results are presented in Figure 6, and the per-scene results are given in the Supp. Mat. Note that the results originally reported in the Point-NeRF paper [70] were evaluated with a resolution different from the compared methods. Therefore, to ensure a fair comparison, we have recomputed the metrics in the publicly available Point-NeRF's predictions with the full resolution of $1920 \\times 1080$ as used in NSVF [32]. In the public dataset, one of the scenes had corrupted camera parameters and we had to reconstruct the poses again for the Ignatius scene. With the reconstructed poses, our method performs slightly worse on that scene compared to others.", + "bbox": [ + 75, + 496, + 468, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results show that our method outperforms the baselines in all compared metrics. This indicates that even though Point-NeRF relies on the ability to grow the point cloud density during training, this is not needed when using a continuous instead of a discrete representation.", + "bbox": [ + 75, + 768, + 467, + 844 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sparse and dense point cloud comparison. Previ", + "text_level": 1, + "bbox": [ + 76, + 849, + 467, + 864 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78cf97762c2780e26f58d53f1195666cb06cf01dae15f4bfacb7457fd8eef17f.jpg", + "image_caption": [ + "(a) sparse point cloud" + ], + "image_footnote": [], + "bbox": [ + 535, + 90, + 694, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e000c812427b2302432dbe438fc19fdeb2fe078553e8afabaa66464968020dd6.jpg", + "image_caption": [ + "(b) ground-truth image" + ], + "image_footnote": [], + "bbox": [ + 697, + 90, + 856, + 171 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f48f77d9e1fdf500161ab6215fe29131aed9c880763c6e16f8321274b73c075d.jpg", + "image_caption": [ + "(c) sparse model", + "Figure 7. Sparse vs. dense model comparison detail. Even though the area in the bottom contains few points, we are still able to reconstruct at least low-frequency data." + ], + "image_footnote": [], + "bbox": [ + 535, + 186, + 694, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/959c1c2a59ebfcd4ef5ce2ac8f5fe7f84af82e40cf39d11ff1abd733d0a4372c.jpg", + "image_caption": [ + "(d) dense model" + ], + "image_footnote": [], + "bbox": [ + 697, + 186, + 856, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ous experiments have used dense COLMAP reconstructions. However, dense point clouds may not be required to achieve good reconstructions. In this section, we compare models trained on dense reconstructions and sparse reconstructions. Table 5 compares PSNR, and SSIM metrics on two synthetic scenes from the Blender dataset [39] and two real-world scenes from the Mip-NeRF 360 dataset [4]. We also show the number of vertices of the tetrahedra, which is the size of the input point cloud, including the random points added at the beginning. Note that in this experiment, we only trained for 100,000 iterations to save computation time.", + "bbox": [ + 496, + 334, + 890, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the results, we can see that for the Blender dataset [39] case, the dense model is much better. This is to be expected since the number of points obtained from the sparse reconstruction on the Blender dataset is very small and our MLP does not have enough capacity to represent fine details. However, even in regions with zero point coverage, we are still able to provide at least some low-frequency data (c.f. Fig. 7). On real-world scenes, the sparse point cloud model is almost on par with the dense one. The sparse reconstructions on the real-world dataset provide many more points compared to the synthetic dataset case, and the coverage is sufficient. From these results, we conclude that on real-world data the sparse model can be sufficient to achieve good performance.", + "bbox": [ + 496, + 516, + 892, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mip-NeRF 360 [4] dataset. We have further evaluated our method on the Mip-NeRF 360 [4] dataset. In order to ensure a fair comparison with Mip-NeRF 360 [4], we trained and evaluated on four times downsized images for the outdoor scenes and two times downsized images for the indoor scenes. The quantitative results are presented in Table 4, and the qualitative results are shown in Figure 8.", + "bbox": [ + 496, + 733, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the results, we can see that Tetra-NeRF is able to outperform both the vanilla NeRF [39] and Mip-NeRF [3]. We also outperform Stable View Synthesis [51] in terms of PSNR and score comparably in terms of LPIPS. This is pos", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3We always choose the type of LPIPS (Alex [27]/VGG [56]) such that we can compare with more methods as some only evaluate using one type.", + "bbox": [ + 76, + 875, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18464", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3bab8b28c2be22a4afc566a04b267096097993072367f7ec88f0a7c318a8a196.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OutdoorIndoorMean
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [15, 39]21.460.4580.51526.840.7900.37023.850.6050.451
mip-NeRF [3]21.690.4710.50526.980.7980.36124.040.6160.441
NeRF++ [77]22.760.5480.42728.050.8360.30925.110.6760.375
Deep Blending [20]21.540.5240.36426.390.8440.26123.700.6660.318
Point-Based Neural Rendering [26]21.660.6120.30226.280.8870.19123.710.7340.253
Stable View Synthesis [51]23.010.6620.25328.220.9070.16025.330.7710.211
mip-NeRF 360 [4]23.720.6870.28229.430.9110.18126.260.7860.237
Tetra-NeRF23.170.5860.29830.210.8810.10326.300.7170.211
", + "bbox": [ + 83, + 87, + 885, + 229 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5d0ce67c4cb03a1ab29de2471ad17c62c82b8cda1a763bbe79eac4c36303d945.jpg", + "table_caption": [ + "Table 4. Mip-NeRF 360 dataset results. We show the PSNR, SSIM, and LPIPS (Alex) [78] on two categories of Mip-NeRF 360 [4] scenes: outdoor, and indoor. On the outdoor scenes, we outperform the Stable View Synthesis [51] and are comparable to Mip-NeRF 360 [4], even though our method does not implement the improvements suggested in NeRF++ [77] and Mip-NeRF 360 [4], and is more comparable to vanilla NeRF [39]. We highlight the best, second, and third values." + ], + "table_footnote": [], + "table_body": "
sparsedense
PSNR/SSIM#pointsPSNR/SSIM#points
Blender/lego29.77/0.95925,78433.79/0.985302,781
Blender/ship26.90/0.8997,15230.69/0.942321,861
360/bonsai28.12/0.902413,22628.34/0.9021,000,000
360/garden24.79/0.806227,53225.41/0.8381,000,000
", + "bbox": [ + 88, + 305, + 459, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Dense and sparse point cloud comparison. We present results on two scenes from the Blender [39] and Mip-NeRF 360 [4] datasets. We also show the number of vertices of the tetrahedra field – including the randomly sampled points.", + "bbox": [ + 75, + 407, + 470, + 464 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/056b494f4cb937613adcb130f764578d6efc865838862e0ddf1f93f8eb5543f9.jpg", + "image_caption": [ + "Figure 8. Mip-NeRF 360 results. $1^{\\text{st}}$ row: we show the results on the garden scene. $2^{\\text{nd}}$ : we can see that Tetra-NeRF has some problems rendering grass. $3^{\\text{rd}}$ : Tetra-NeRF is able to represent the delicate texture of the tablecloth well. $4^{\\text{th}}$ : Tetra-NeRF is able to recover fine geometry." + ], + "image_footnote": [], + "bbox": [ + 84, + 472, + 462, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sible because outdoor scenes contain a lot of high-resolution geometries, such as leaves and Stable View Synthesis is not able to overcome noise in the approximate geometry. Tetra-", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "NeRF does not suffer from these problems as we aggregate features in 3D along rays rather than on the surface. Mip-NeRF 360 scores comparably to Tetra-NeRF in terms of PSNR and LPIPS, but it outperforms Tetra-NeRF in terms of SSIM. However, Mip-NeRF 360 implements some tricks designed to boost its performance. On the other hand, our approach is based on vanilla NeRF, and we believe that the performance can be boosted similarly.", + "bbox": [ + 496, + 308, + 893, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Varying the number of input points We conducted a study on the effect of the input point cloud size on the reconstruction quality. We consider both the sparse and the dense point clouds and analyse the performance as we subsample the points or add more points randomly. We evaluate on the family and the truck scenes from the Tanks and Temples [25] (tt) dataset, and the room and the garden scenes from the Mip-NeRF 360 [4] dataset. The sizes of sparse point clouds were roughly 16K, 30K, 113K, and 139K respectively and the sizes of dense point clouds were larger than 5M for all scenes. When the size of the original point cloud was larger than the desired size, we selected points uniformly at random, and when smaller, we added new random points as described Section 4.1. We include detailed results in the Supp. Mat. In order to save computational resources, we only train the method for 100k iterations. The results are visualised in Figure 9a.", + "bbox": [ + 496, + 436, + 892, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As expected, the performance improves with the number of points used, as it leads to a finer subdivision of the scene around the surface. We can also observe (e.g., from tt/family) that the sparse reconstruction leads to a better reconstruction quality at the same point cloud size. The likely explanation is that sampling uniformly at random will be biased towards regions with high density and some regions may be missed or covered extremely sparsely.", + "bbox": [ + 496, + 696, + 893, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Speed of convergence Figures 9b, 9c show the rate of convergence on the ship scene from the Blender dataset and aggregated results on the Tanks and Temples [25], and Mip-NeRF 360 (indoor/outdoor) [4] datasets. Similar to Point-NeRF, good results can be achieved quite early in the train", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18465", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9861e61b062a099c2b7fec2c1f74dc9daa6a5bff69fdcfdf20a10c42835c1acf.jpg", + "image_caption": [ + "Figure 9. a) shows the PSNR with different sizes of input point clouds. The solid and dashed lines represent the dense and coarse COLMAP reconstructions. The following scenes were evaluated (best to worst): tt/family, 360/room, tt/truck, and 360/garden from Tanks and Temples [25] (tt) and Mip-NeRF 360 [4] (360-indoor/360-outdoor) datasets. As expected, the quality increases with the number of points. b) visualizes the training speed (PSNR at different training iterations) on the blender/ship scene [39]. It shows Tetra-NeRF training times for various timesteps and compares the rate of convergence with Point-NeRF [70] and Mip-NeRF [3]. Similarly, c) shows the rate of convergence aggregated over Tanks and Temples [25] (tt) and Mip-NeRF 360 [4] (360-indoor/360-outdoor) datasets. It shows the average PSNR, and the highlighted area represents the standard deviation." + ], + "image_footnote": [], + "bbox": [ + 84, + 95, + 348, + 250 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a8fcbdbe0348c4c19d2d907b1d33dd173ab1363e6056d909064cc7f302ad70d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 92, + 614, + 250 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0cc67eab62bbbc84549638a939872ce4ef7f3874a9f6b44330922790930cc799.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 98, + 880, + 250 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/af0ad92993a5cb97e75e153e92119f9c5f9af71eb9f578ec6a1c00da16e49b38.jpg", + "image_caption": [ + "(a) Failure case 1: low point cloud density on the ground" + ], + "image_footnote": [], + "bbox": [ + 84, + 363, + 274, + 459 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/31f34bbd148a667e04de4f51d7754f63b3ea14fc42569a08ca1b9bbf99ff7342.jpg", + "image_caption": [ + "(b) Failure case 2: too many intersected tetrahedra", + "Figure 10. Failure cases. Left: barn scene from [25]. Right: tree hill scene from [4]." + ], + "image_footnote": [], + "bbox": [ + 279, + 362, + 472, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ing process. E.g., on the blender/ship scene, in 24 minutes, our method achieves performance similar to Mip-NeRF [3] after several hours of training. The efficiency of the implementation is a large factor in the speed of different methods. E.g., in Instant-NGP [41], the authors invested a significant effort in optimizing GPU memory accesses, and the result is highly optimised efficient code (a main contribution of [41]). There are several places where our implementation can easily be made more efficient. E.g., changing the memory layout (row-major vs. column-major) of the tetrahedra field leads to a $10\\%$ speedup.", + "bbox": [ + 75, + 525, + 467, + 691 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 76, + 707, + 199, + 722 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "One drawback of our approach is that the quality of different regions of the rendered scene depends on the density of the point cloud in these regions. If the original point cloud was constructed using 2D feature matching, there can be regions with a very low number of 3D points. We show such an example in Fig. 10a. There are few 3D points on the ground, resulting in blurry renderings for that region.", + "bbox": [ + 75, + 733, + 467, + 839 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Another issue is that the current implementation has a limit on the number of intersected tetrahedra per ray. For large or badly structured scenes, where rays intersect many tetrahedra, this can limit the reconstruction quality. Fig. 10b", + "bbox": [ + 75, + 840, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "shows such a case on a Mip-NeRF 360 scene [4]. This problem can be addressed by increasing the limit on the number of visited tetrahedra at the cost of higher memory requirements and run-time. Alternatively, coarse-to-fine schemes that start with a coarser tetrahedralisation and prune the space could potentially handle this issue.", + "bbox": [ + 496, + 361, + 890, + 452 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 472, + 617, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper proposes a novel radiance field representation that, compared to standard voxel-based representations, is easily able to adapt to 3D geometry priors given in the form of a (sparse) point cloud. Our approach elegantly combines concepts from geometry processing (Delaunay triangulation) and triangle-based rendering (ray-triangle intersections) with modern neural rendering approaches. The representation has a naturally higher resolution in the space near surfaces, and the input point cloud provides a straightforward way to initialise the radiance field. Compared to Point-NeRF, a state-of-the-art point cloud-based radiance field representation which uses the same input as our approach, Tetra-NeRF shows clearly better results. Our method performs comparably to state-of-the-art MLP-based methods. The results demonstrate that Tetra-NeRF is an interesting alternative to existing radiance field representations that is worth further investigation. Interesting research directions include adaptive refinement and pruning of the tetrahedralisation and exploiting the fact that the surface of the scene is likely close to some of the triangles in the scene.", + "bbox": [ + 496, + 500, + 890, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported by the Czech Science Foundation (GACR) EXPRO (grant no. 23-07973X), the Grant Agency of the Czech Technical University in Prague (grant no. SGS22/112/OHK3/2T/13), and by the Ministry of Education, Youth and Sports of the Czech Republic through the e-INFRA CZ (ID:90254).", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18466", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In ECCV, 2020. 2", + "[2] Pierre Alliez, Simon Giraudot, Clément Jamin, Florent Lafarge, Quentin Merigot, Jocelyn Meyron, Laurent Saboret, Nader Salman, Shihao Wu, and Necip Fazil Yildiran. Point set processing. In CGAL User and Reference Manual. CGAL Editorial Board, 5.5.1 edition, 2022. 5", + "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. In ICCV, 2021. 1, 2, 4, 5, 7, 8, 9", + "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1, 2, 5, 7, 8, 9", + "[5] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv:2008.03824, 2020. 2", + "[6] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. NoPe-NeRF: Optimising neural radiance field with no pose prior. arXiv:2212.07388, 2022.2", + "[7] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. NeRD: Neural reflectance decomposition from image collections. In ICCV, 2021. 2", + "[8] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001. 2", + "[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial radiance fields. In ECCV, 2022. 1, 2, 4, 5", + "[10] Anpei Chen, Zexiang Xu, Xinyue Wei, Siyu Tang, Hao Su, and Andreas Geiger. Factor fields: A unified framework for neural fields and beyond. arXiv:2302.01226, 2023. 1, 2, 3", + "[11] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 2", + "[12] Shuo Cheng, Zexiang Xu, Shilin Zhu, Zhuwen Li, Li Erran Li, Ravi Ramamoorthi, and Hao Su. Deep stereo using adaptive thin volume representation with uncertainty awareness. In CVPR, 2020. 2", + "[13] Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2", + "[14] B Delaunay, S Vide, A Lamémoire, and V De Georges. Bulletin de l'académie des sciences de l'URSS. Classe des sciences mathématiques et naturelles, 6:793-800, 1934. 2, 3, 4, 5" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Boyang Deng, Jonathan T. Barron, and Pratul P. Srinivasan. JaxNeRF: an efficient JAX implementation of NeRF, 2020. 8", + "[16] Michael S Floater. Generalized barycentric coordinates and applications. Acta Numerica, 24:161-214, 2015. 3, 4", + "[17] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2, 4, 5", + "[18] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32:1362-1376, 2009. 2", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5", + "[20] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. SIGGRAPH, ToG, 37(6):1-15, 2018. 2, 8", + "[21] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In ICCV, 2021. 2", + "[22] Vu Hoang Hiep, Renaud Keriven, Patrick Labatut, and Jean-Philippe Pons. Towards high-resolution large-scale multiview stereo. In CVPR, 2009. 2", + "[23] Michal Jancosek and Tomas Pajdla. Multi-view reconstruction preserving weakly-supported surfaces. In CVPR, 2011. 2", + "[24] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In SGP, 2006. 2", + "[25] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. SIGGRAPH, ToG, 2017. 5, 6, 7, 8, 9", + "[26] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, volume 40, pages 29-43. Wiley Online Library, 2021. 2, 8", + "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 2012. 7", + "[28] Jonás Kulhánek, Erik Derner, Torsten Sattler, and Robert Babuška. ViewFormer: Nerf-free neural rendering from few images using transformers. In ECCV, 2022. 2", + "[29] Patrick Labatut, Jean-Philippe Pons, and Renaud Keriven. Efficient multi-view reconstruction of large-scale scenes using interest points, Delaunay triangulation and graph cuts. In ICCV, 2007. 2", + "[30] Patrick Labatut, J-P Pons, and Renaud Keriven. Robust and efficient surface reconstruction from range data. In Computer graphics forum, volume 28, pages 2275-2290. Wiley Online Library, 2009. 2", + "[31] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. BARF: Bundle-adjusting neural radiance fields. In ICCV, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18467", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 1, 2, 3, 5, 6, 7", + "[33] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond. In International Conference on Learning Representations, 2020. 5", + "[34] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: learning dynamic renderable volumes from images. SIGGRAPH, ToG, 38(4):1-14, 2019. 2, 6", + "[35] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3D surface construction algorithm. SIGGRAPH, ToG, 1987. 2", + "[36] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 2", + "[37] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 1, 3", + "[38] Moustafa Meshry, Dan B. Goldman, Sameh Khamis, Hugues Hoppe, Rohit Pandey, Noah Snavely, and Ricardo MartinBrualla. Neural rerendering in the wild. In $CVPR$ , June 2019. 2", + "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 4, 5, 6, 7, 8, 9", + "[40] Tomas Möller. A fast triangle-triangle intersection test. Journal of Graphics Tools, 2(2):25-30, 1997. 2", + "[41] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. SIGGRAPH, ToG, 2022. 1, 2, 3, 4, 5, 9", + "[42] M. Nießner, M. Zollhöfer, S. Izadi, and M. Stamminger. Real-time 3D reconstruction at scale using voxel hashing. SIGGRAPH, ToG, 2013. 4", + "[43] Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith Morley, Austin Robison, et al. OptiX: a general purpose ray tracing engine. ACM Transactions on Graphics (ToG), 29(4):1-13, 2010. 2, 4, 5", + "[44] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, 2020. 2", + "[45] Bui Tuong Phong. Illumination for computer generated pictures. Communications of the ACM, 18(6):311-317, 1975. 2", + "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. arXiv:2209.14988, 2022. 2", + "[47] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding up neural radiance fields with thousands of tiny mlp's. In ICCV, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T Barron, and Peter Hedman. MERF: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. arXiv:2302.12249, 2023. 2", + "[49] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In ICCV, 2021. 2", + "[50] Gernot Riegler and Vladlen Koltun. Free view synthesis. In ECCV, 2020. 2", + "[51] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In CVPR, 2021. 2, 7, 8", + "[52] Radu Alexandru Rosu and Sven Behnke. Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In CVPR, 2023. 2", + "[53] Darius Rückert, Linus Franke, and Marc Stamminger. ADOP: Approximate differentiable one-pixel point rendering. SIGGRAPH, ToG, 41(4):1-14, 2022. 2", + "[54] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 2, 4, 5", + "[55] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 2, 4, 5", + "[56] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv:1409.1556, 2014. 7", + "[57] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 1, 2, 4", + "[58] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-NeRF: Scalable large scene neural view synthesis. In CVPR, 2022. 2", + "[59] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 4", + "[60] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, et al. Nerfstudio: A modular framework for neural radiance field development. arXiv:2302.04264, 2023. 2, 5", + "[61] Chengzhou Tang and Ping Tan. BA-net: Dense bundle adjustment networks. In International Conference on Learning Representations, 2019. 2", + "[62] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 2", + "[63] Sudheendra Vijayanarasimhan, Susanna Ricco, Cordelia Schmid, Rahul Sukthankar, and Katerina Fragkiadaki. SfM-Net: Learning of structure and motion from video. arXiv:1704.07804, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18468", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-NeRF: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2022. 2", + "[65] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34:27171-27183, 2021. 2", + "[66] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snively, and Thomas Funkhouser. IBRNet: Learning multi-view image-based rendering. In CVPR, 2021. 2", + "[67] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. NeuS2: Fast learning of neural implicit surfaces for multi-view reconstruction. arXiv:2212.05231, 2022. 2", + "[68] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF-: Neural radiance fields without known camera parameters. arXiv:2102.07064, 2021. 2", + "[69] Daniel N Wood, Daniel I Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H Salesin, and Werner Stuetzle. Surface light fields for 3D photography. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, 2000. 2", + "[70] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-NeRF: Point-based neural radiance fields. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 9", + "[71] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multiview stereo. In ECCV, 2018. 2", + "[72] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2", + "[73] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. BakedSDF: Meshing neural SDFs for real-time view synthesis. arXiv:2302.14859, 2023. 2", + "[74] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 1, 2", + "[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2", + "[76] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. In Advances in Neural Information Processing Systems, 2022. 2", + "[77] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. NeRF++: Analyzing and improving neural radiance fields. arXiv:2010.07492, 2020. 2, 8", + "[78] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, June 2018. 5, 6, 8" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 11 + }, + { + "type": "ref_text", + "text": "[79] Chaoqiang Zhao, Qiyu Sun, Chongzhen Zhang, Yang Tang, and Feng Qian. Monocular depth estimation based on deep learning: An overview. Science China Technological Sciences, 63(9):1612-1627, 2020. 3", + "bbox": [ + 501, + 90, + 892, + 147 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "18469", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/8c410ac7-5eb0-4f17-a4bb-daadb2446edc_model.json b/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/8c410ac7-5eb0-4f17-a4bb-daadb2446edc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2b7c3bdb26a39c3890313cfc2bcce730ee25c54c --- /dev/null +++ b/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/8c410ac7-5eb0-4f17-a4bb-daadb2446edc_model.json @@ -0,0 +1,2655 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.131, + 0.833, + 0.154 + ], + "angle": 0, + "content": "Tetra-NeRF: Representing Neural Radiance Fields Using Tetrahedra" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.182, + 0.462, + 0.234 + ], + "angle": 0, + "content": "Jonas Kulhanek \nCzech Technical University in Prague \njonas.kulhanek@cvut.cz" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.182, + 0.808, + 0.234 + ], + "angle": 0, + "content": "Torsten Sattler \nCzech Technical University in Prague \ntorsten.sattler@cvut.cz" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.248, + 0.599, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.261, + 0.802, + 0.426 + ], + "angle": 0, + "content": "3) barycetric interpolation of features 4) shallow MLP predicts density and colour" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.427, + 0.894, + 0.47 + ], + "angle": 0, + "content": "Figure 1. The input to Tetra-NeRF is a point cloud which is triangulated to get a set of tetrahedra used to represent the radiance field. Rays are sampled, and the field is queried. Barycentric interpolation is used to interpolate tetrahedra vertices, and the resulting features are passed through a shallow MLP to get the density and colours for volumetric rendering." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.485, + 0.314, + 0.501 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.474, + 0.79 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRFs) are a very recent and very popular approach for the problems of novel view synthesis and 3D reconstruction. A popular scene representation used by NeRFs is to combine a uniform, voxel-based subdivision of the scene with an MLP. Based on the observation that a (sparse) point cloud of the scene is often available, this paper proposes to use an adaptive representation based on tetrahedra obtained by Delaunay triangulation instead of uniform subdivision or point-based representations. We show that such a representation enables efficient training and leads to state-of-the-art results. Our approach elegantly combines concepts from 3D geometry processing, triangle-based rendering, and modern neural radiance fields. Compared to voxel-based representations, ours provides more detail around parts of the scene likely to be close to the surface. Compared to point-based representations, our approach achieves better performance. The source code is publicly available at: https://jkulhanek.com/tetra-nerf." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.815, + 0.21, + 0.831 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Reconstructing 3D scenes from images and rendering photo-realistic novel views is a key problem in computer vision. Recently, NeRFs [3,4,39] became dominant in the field for their superior photo-realistic results. Originally," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.893, + 0.804 + ], + "angle": 0, + "content": "NeRFs used MLPs to represent the 3D scene as an implicit function. Given a set of posed images, NeRF randomly samples a batch of pixels, casts rays from the pixels into the 3D space, queries the implicit function at randomly sampled distances along the rays, and aggregates the sampled values using volumetric rendering [37, 39]. While the visual results of such methods are of high quality, the problem is that querying large MLPs at millions of points is costly. Also, once the network is trained, it is difficult to make any changes to the represented radiance field as everything is baked into the MLPs parameters, and any change has a non-local effect. Since then, there have been a lot of proposed alternatives to the large MLP field representation [9, 17, 32, 41, 57, 70, 74]. These methods combine an MLP with a voxel feature grid [41, 57], or in some cases represent the radiance field directly as a tensor [9, 10, 17]. When querying these representations, first, the containing voxel is found, and the features stored at the eight corner points of the voxel are trilinearly interpolated. The result is either passed through a shallow MLP [9, 10, 41, 57] or is used directly as the density and colour [17, 32, 57]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Having a dense tensor represent the entire scene is very inefficient, as we only need to represent a small space around surfaces. Therefore, different methods propose different ways of tackling the issue. Instant-NGP [41], for example, uses a hash grid instead of a dense tensor, where it relies on optimisation to resolve the hash collisions. How-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18458" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.226 + ], + "angle": 0, + "content": "ever, similarly to MLPs, any change to the stored hashmap influences the field in many places. A more common direction to addressing the issue is by directly using a sparse tensor representation [9, 17]. These methods start with a low-resolution grid and, at predefined steps, subsample the representation, increasing the resolution. These approaches tend to require a careful setting of hyperparameters, such as the scene bounding box and the subdivision steps, in order for the methods to work well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.485 + ], + "angle": 0, + "content": "Because many of these methods use traditional structure from motion (SfM) [54, 55] methods to generate the initial poses for the captured images, we can reuse the original reconstruction in the scene representation. Inspired by classical surface reconstruction methods [22, 23, 29, 30], we represent the scene as a dense triangulation of the input point cloud, where the scene is a set of non-overlapping tetrahedra whose union is the convex hull of the original point cloud [14]. When querying such a representation, we find to which tetrahedron the query point belongs and perform barycentric linear interpolation of the features stored in the vertices of the tetrahedron. This very simple representation can be thought of as the direct extension of the classical triangle-rendering pipelines used in graphics [40, 43, 45]. The representation avoids problems with the sparsity of the input point cloud as the tetrahedra fully cover the scene, resulting in a continuous rather than discrete representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.473, + 0.696 + ], + "angle": 0, + "content": "This paper makes the following contributions: (1) We propose a novel radiance field representation which is initialised from a sparse or dense point cloud. This representation is naturally denser in the proximity of surfaces and, therefore, provides a higher resolution in these regions. (2) The proposed representation is evaluated on multiple synthetic and real-world datasets and is compared with a state-of-the-art point-cloud-based representation - Point-NeRF [70]. The presented results show that our method clearly outperforms this baseline. We further demonstrate the effectiveness of our adaptive representation by comparing it with a voxel-based representation that uses the same number of trainable parameters. (3) We make the source code and model checkpoints publicly available." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.709, + 0.216, + 0.725 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Multi-view reconstruction. The problem of multi-view reconstruction has been studied extensively and tackled with a variety of structure from motion (SfM) [54, 61, 63], and multi-view stereo (MVS) [12, 18, 55, 71] methods. These methods usually output the scene represented as a point cloud [54, 55]. In most rendering approaches, the point cloud is converted into a mesh [24, 35], and novel views are rendered by reprojecting observed images into each novel viewpoint and blending them together using ei-" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "ther heuristically-defined [8, 13, 69] or learned [20, 50, 51, 73] blending weights." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.125, + 0.894, + 0.261 + ], + "angle": 0, + "content": "However, the process of getting the meshes is usually quite noisy, and the resulting meshes tend to have inaccurate geometry in regions with fine details or complex materials. Instead of using noisy meshes, point-based neural rendering methods [1, 26, 38, 53] perform splatting of neural features and use 2D convolutions to render them. In contrast to these methods, our approach operates and aggregates features directly in 3D and does not suffer from the noise in the point cloud or the reconstructed mesh." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.269, + 0.893, + 0.572 + ], + "angle": 0, + "content": "Neural radiance fields. Recently, NeRFs [3,4,34,39,77] have gained a lot of attention thanks to their high-quality rendering performance. The original NeRF method [39] was extended to better handle aliasing artefacts in [3], to better represent unbounded scenes in [4,48,77], or to handle real-world captured images [36,58]. The training of the large MLPs used in these methods can be quite slow, and there has been a lot of effort on speeding up either the training [9,17,41] or the rendering [21,47,48,74] sometimes at the cost of larger storage requirements. Other approaches tackled different aspects of NeRFs like view-dependent artefacts [62], relighting [5,7], or proposed generative models [46,64]. Also, a popular research direction is making the models generalize across different scenes [11,28,49,66,75]. A large area of research is dedicated to the surface reconstruction and, instead of using the radiance fields, represents the scene implicitly by modelling the signed distance function (SDF) [52,65,67,72,76]. Unlike those approaches, we focus only on the radiance field representation and consider these methods orthogonal to ours." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.894, + 0.697 + ], + "angle": 0, + "content": "Although there are some methods that train radiance fields while fine-tuning the poses [31, 60] or without known cameras [6, 68], most methods need camera poses for the reconstruction. SfM, e.g., COLMAP [54, 55], is typically used for estimating the poses, which also produces a (sparse) point cloud. Our approach makes use of this by-product of the pose recovery process instead of only using the poses themselves." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Field representations. When a single MLP is used to represent the entire scene, everything is baked into a single set of parameters which cannot be easily modified, because any change to the parameters has a non-local effect, i.e., it changes the scene at multiple unrelated places. To overcome this problem, others have experimented with different representations of the radiance fields [9, 10, 17, 32, 41, 44, 57, 70]. A common practice is to represent the scene as a shallow MLP and an efficiently encoded voxel grid [9, 10, 32, 41, 44]. The encoded voxel grid can also be used represent the radiance field directly [17, 57]. The voxel grid can be encoded as a sparse tensor [17, 32, 57], a factorisation of the 4D tensor [9, 10], or a hashmap [41]. When" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.887, + 0.419, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/jkulhanek/tetra-nerf" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18459" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.091, + 0.273, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.178, + 0.24, + 0.19 + ], + "angle": 0, + "content": "(a) dense point cloud" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.093, + 0.444, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.178, + 0.432, + 0.19 + ], + "angle": 0, + "content": "(b) dense tetrahedra slice" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.192, + 0.273, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.28, + 0.241, + 0.291 + ], + "angle": 0, + "content": "(c) sparse point cloud" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.192, + 0.455, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.28, + 0.434, + 0.291 + ], + "angle": 0, + "content": "(d) sparse tetrahedra slice" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.47, + 0.346 + ], + "angle": 0, + "content": "Figure 2. Input point cloud and a slice through the triangulated tetrahedra. Note that smaller tetrahedra are created closer to the surface of the scene, i.e., regions close to the surface are represented with a finer resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.47, + 0.51 + ], + "angle": 0, + "content": "these structures are queried, trilinear interpolation is used to combine the feature vectors stored in the containing voxel corners. Unfortunately, the hashes [41] and hierarchical representations [10] have the same non-local effect problem, and the rest of the approaches rely on subsequent upsampling of the field and can be overly complicated. Also, feature vectors cannot be placed arbitrarily in the 3D space as they must lie on the grid with a fixed resolution. In contrast, our approach is much more flexible as it stores the feature vectors freely in 3D space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.692 + ], + "angle": 0, + "content": "Finally, Point-NeRF [70] represents the scene as a point cloud of features which are queried using \\( k \\)-nearest neighbours search. However, when the point cloud contains sparse regions, the rays do not intersect any neighbourhoods of any points and the pixels stay empty without the ability to optimise. Therefore, Point-NeRF [70] relies on gradually adding more points during training and increasing the scene complexity. Since we use the triangulation of the point cloud rather than the discrete point cloud itself, our representation is continuous and does not suffer from empty regions. Therefore, we do not have to add any points during training." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.708, + 0.169, + 0.724 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "A common strategy in the literature is to represent the scene explicitly through a voxel volume. In contrast to this uniform subdivision, we investigate using an adaptive subdivision of the scene. In many scenarios, an approximation of the scene geometry is either given, e.g., when using SfM to compute the input camera poses, or can be computed, e.g., via MVS or single-view depth predictions [79]. This allows us to compute an adaptive subdivision of the scene via Delaunay triangulation [14] of such a point cloud. This results in a set of non-overlapping tetrahedra, where smaller tetrahedra are created close to the surface of the scene (c.f." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.091, + 0.715, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.201, + 0.691, + 0.213 + ], + "angle": 0, + "content": "(a) trilinear interpolation" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.092, + 0.857, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.2, + 0.871, + 0.213 + ], + "angle": 0, + "content": "(b) barycentric interpolation" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.216, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Figure 3. Trilinear and barycentric interpolation. Trilinear interpolation is a weighted combination of the eight voxel corners. Barycentric interpolation weights the four vertices of the tetrahedron vertices based on the barycentric coordinates [16]." + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.285, + 0.892, + 0.332 + ], + "angle": 0, + "content": "Fig. 2). In the following, we explain how this adaptive subdivision of the scene can be used instead of voxels for volume rendering and neural rendering." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.343, + 0.64, + 0.358 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.367, + 0.893, + 0.534 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRFs) [39] represent a scene through an implicit function \\( F(\\mathbf{x}, \\mathbf{d}) \\), often modelled via a neural network, that returns a colour value and a volume density prediction for a given 3D point \\( \\mathbf{x} \\) in the scene observed from a viewing direction \\( \\mathbf{d} \\). Volume rendering [37] is used to synthesise novel views: for each pixel in a virtual view, we project a ray from the camera plane into the scene and sample the radiance field to obtain the colour and density values \\( \\mathbf{c}_i \\) and \\( \\sigma_i \\) at distances \\( t_i \\) along the ray, where \\( i = 1 \\dots N \\). The individual samples are then combined to predict the colour \\( \\mathbf{C} \\) for the pixel:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.546, + 0.892, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} = \\sum_ {i = 1} ^ {N} T _ {i} \\left(1 - \\exp \\left(\\sigma_ {i} \\delta_ {i}\\right)\\right) \\mathbf {c} _ {i}, \\text {w h e r e} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.59, + 0.732, + 0.632 + ], + "angle": 0, + "content": "\\[\nT _ {i} = \\exp \\left(- \\sum_ {i = 1} ^ {N - 1} \\sigma_ {i} \\delta_ {i}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.719 + ], + "angle": 0, + "content": "and \\(\\delta_{i} = t_{i} - t_{i - 1}\\) is the distance between adjacent samples. This process is fully differentiable. Thus, after computing the mean squared error (MSE) between the predicted and ground truth colours associated with each ray, we can backpropagate the gradients to the radiance field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Voxel-based feature fields such as NSVF [32] represent the scene as a voxel grid and an MLP. Each grid point of the voxel grid is assigned a trainable vector. There are eight feature vectors associated with a single voxel, but the vectors are shared between neighbouring voxels. For each query point sampled along the ray, its corresponding voxel is found first. A feature for the point is then computed via trilinear interpolation of the features of the voxel grid points based on the position of the query point (c.f. Fig. 3, left). The resulting feature vector is passed through an MLP to predict the density and appearance vector. The appearance vector is combined with the ray direction and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18460" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "passed through a second MLP in order to compute a view-dependent colour." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.129, + 0.245, + 0.143 + ], + "angle": 0, + "content": "3.2. Tetrahedra fields" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.469, + 0.302 + ], + "angle": 0, + "content": "Given a set of points in 3D space, we build the tetrahedra structure by triangulating the points. We apply the Delaunay triangulation [14] to obtain a set of non-overlapping tetrahedra whose union is the convex hull of the original points. Fig. 2 shows example tetrahedra obtained by triangulating dense and sparse COLMAP-reconstructed point clouds [54, 55]. Note that the resulting representation is adaptive as it uses a higher resolution (smaller tetrahedra) closer to the surface and larger tetrahedra to model regions farther away from the surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.469, + 0.558 + ], + "angle": 0, + "content": "We associate all vertices of the tetrahedra with trainable vectors. As in the voxel grid case, vertices, and thus vectors, are shared between adjacent tetrahedra. The resulting tetrahedra field can be queried in the same way a voxel grid representation is queried: for each query point, we first find the tetrahedron containing the point. Instead of the trilinear interpolation used for voxel volumes, we use the barycentric interpolation [16] to compute a feature vector for the query point from the four feature vectors stored at the tetrahedron's vertices (c.f. Fig. 3). To this end, we compute the query point \\(\\mathbf{x}\\)'s barycentric coordinates \\(\\lambda\\), which express the point's 3D coordinates as a unique weighted combination of the 3D coordinates of the tetrahedron's vertices. In particular, the weight for a vertex is the volume of the tetrahedron constructed from the query point and the face opposite to the vertex divided by the volume of the full tetrahedron:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.562, + 0.469, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\lambda} = \\left(\\frac {V _ {x 2 3 4}}{V _ {1 2 3 4}}, \\frac {V _ {1 x 3 4}}{V _ {1 2 3 4}}, \\frac {V _ {1 2 x 4}}{V _ {1 2 3 4}}, \\frac {V _ {1 2 3 x}}{V _ {1 2 3 4}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.599, + 0.469, + 0.659 + ], + "angle": 0, + "content": "where \\( V_{1234} \\) is the volume of the full tetrahedron and \\( V_{x234} \\), \\( V_{1x34}, \\ldots \\), are volumes of tetrahedra with \\( 1^{\\mathrm{st}} \\), \\( 2^{\\mathrm{nd}} \\), ..., vertex replaced by \\( \\mathbf{x} \\). The same weights \\( \\lambda \\) are applied to the feature vectors of the vertices to obtain the query feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.78 + ], + "angle": 0, + "content": "The interpolated features are used as the input to a small MLP in order to predict density and colour at the query point. We first pass the barycentric-interpolated features through a three-layer MLP to compute the density and appearance features. We then concatenate the appearance features with the ray direction vector encoded using Fourier features [39, 59] and pass the result through a single linear layer to get the raw RGB colour value." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.841 + ], + "angle": 0, + "content": "Finally, to map the raw density values \\(\\bar{\\sigma}_i\\) returned by the network to the volume density \\(\\sigma_{i}\\) required by volume rendering, we apply the softplus activation function [3]. For the RGB colour value, we use the sigmoid function [17]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.405, + 0.864 + ], + "angle": 0, + "content": "3.3. Efficiently querying a tetrahedra field" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Determining the corresponding voxel for a given query point can be done highly efficiently via hashing [42]. In" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "contrast, finding the corresponding tetrahedron for a query point is more complex. This in turn can significantly impact rendering, and thus training, efficiency [17,41,57]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.363 + ], + "angle": 0, + "content": "In order to efficiently look up the corresponding tetrahedra, we exploit that we are not considering isolated points, but points sampled from a ray. We compute the tetrahedra that are intersected by the ray, allowing us to march through the tetrahedra rather than computing them individually per point. The relevant tetrahedra can be found efficiently using acceleration structures for fast ray-triangle intersection computations, e.g., via NVidia's OptiX library [43]: We first compute the intersections between the rays from a synthetic view and all faces of the tetrahedra. For each ray, we take the first 512 intersected triangles², and determine the corresponding tetrahedra. The tetrahedra can be ordered based on the intersections along the ray, allowing us to easily march through the tetrahedra to determine which tetrahedron to use for a given query point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.514 + ], + "angle": 0, + "content": "A side benefit of computing ray-triangle intersections is that we can simplify computing the barycentric coordinates: For each intersection, we compute the 2D barycentric coordinates w.r.t. the triangle. We then obtain the 3D barycentric coordinates w.r.t. the associate tetrahedron by simply adding zero for the vertex opposite to the triangle. For a query point inside a tetrahedron, we can compute its tetrahedron barycentric coordinates by linearly interpolating between the barycentric coordinates of the two intersections of the ray and the tetrahedron." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.522, + 0.731, + 0.538 + ], + "angle": 0, + "content": "3.4. Coarse and fine sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.711 + ], + "angle": 0, + "content": "We follow the common practice of having a two-stage sampling procedure [9, 39]. In the coarse stage, we sample uniformly along the ray. In the fine stage, we use the density weights from the coarse sampling stage to bias the sampling towards sampling closer to the potential surface. Following [39], we use the stratified uniform sampling for the coarse stage. The stratified uniform sampling splits the ray into equally long intervals and samples uniformly in each interval. Unlike NeRF [39], we limit the sampling to the space occupied by tetrahedra. In the fine sampling stage, we use the same network as in the coarse sampling stage." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.712, + 0.892, + 0.742 + ], + "angle": 0, + "content": "For the fine sampling stage, we take the accumulated weights \\( w_{i} \\) from the coarse sampling:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.75, + 0.892, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\bar {w} _ {i} = \\left(1 - \\exp \\left(- \\sigma_ {i} \\delta_ {i}\\right)\\right) \\exp \\left(- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {j} \\delta_ {j}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.799, + 0.892, + 0.845 + ], + "angle": 0, + "content": "These weights are the coefficients used in Equation 1 as multipliers for the colours [39]. We obtain \\( w_{i} \\) by normalizing \\( \\bar{w}_{i} \\). Following [39], we sample set of fine samples using" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.851, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2For efficiency, we only consider a fixed number of triangles. As discussed later on, this can degrade results in larger scenes / scenes with a fine-grained tetrahedralisation, where more intersections are needed. Naturally, more triangles can be considered at the cost of longer run-times." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18461" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "weight \\( w_{i} \\). To render the final colour, we merge the dense and fine samples and use all in the rendering equation [9]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.134, + 0.21, + 0.151 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.16, + 0.47, + 0.266 + ], + "angle": 0, + "content": "We compare Tetra-NeRF to relevant methods on the commonly used synthetic Blender [39], the real-world Tanks and Temples [25], and the challenging object-centric Mip-NeRF 360 [4] datasets. To show its effectiveness, we compare Tetra-NeRF to a dense-grid representation and evaluate it with reduced quality of the input point cloud. We start by describing the exact hyperparameters used." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.274, + 0.292, + 0.29 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.471, + 0.569 + ], + "angle": 0, + "content": "Generating point cloud & triangulation. Given a set of posed images, we use the COLMAP reconstruction pipeline [54, 55] to get the point cloud used in our tetrahedra field representation. We then reduce the size of the resulting point cloud such that if the number of points is larger than \\(10^{6}\\), we subsample \\(10^{6}\\) points randomly. For the Blender dataset experiments, where the number of points is smaller, we add more randomly generated points. In that case, the number of random points is half the number of original points. The reason for adding the points is that with a low number of points, some pixels on edges may not intersect any tetrahedra, potentially producing artefacts on the edges. Each added point is sampled as follows: we sample a random point \\(x_{0}\\) from the original point cloud, sample a random normal vector \\(n\\), and a number \\(\\alpha \\sim \\mathcal{N}(\\bar{d}, \\bar{d}^{2})\\), where \\(\\bar{d}\\) is the average spacing of the original point cloud, i.e. the average distance between each point and its six closest neighbours. We then add the point \\(x = x_{0} + \\alpha n\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.794 + ], + "angle": 0, + "content": "Initialization. Given the processed point cloud, we apply the Delaunay triangulation [14] to get a set of tetrahedra. To this end, we use the CGAL library [2], which in our experiments runs in the order of milliseconds. Following [41], we initialise the features of size 64 at the vertices of the tetrahedra with small values sampled uniformly in the range \\(-10^{-4}\\) to \\(10^{-4}\\). However, to allow the model to reuse the information contained in the point cloud, we set the first four dimensions of the feature field to the RGBA colour values (rescaled to interval [0, 1]) stored at the associated points in the point cloud. The alpha value of all original points is one, whereas all randomly sampled points have an alpha value of zero. For the MLP, we follow the common practice of using the Kaiming uniform initialisation [19]. The hidden sizes in all MLPs are 128." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Training. During training, we sample batches of 4,096 random rays from random training images. We use volumetric rendering to predict the colour of each ray. The gradients are computed by backpropagating the MSE loss between the predicted colour and the ground truth colour value. We use the RAdam optimizer [33] and decay the learning rate exponentially from \\(10^{-3}\\) to \\(10^{-4}\\) in 300k steps. The train-" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.089, + 0.868, + 0.219 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓
NeRF [39]31.000.9470.081
NSVF [32]31.770.953-
mip-NeRF [3]34.510.9610.043
instant-NGP [41]33.18--
Plenoxels [17]31.710.9580.049
Point-NeRFcol [70]31.770.9730.062
Point-NeRFmvs [70]33.310.9780.049
Tetra-NeRF32.530.9820.041
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.226, + 0.894, + 0.314 + ], + "angle": 0, + "content": "Table 1. Results on the Blender dataset [39] averaged over all scenes in the dataset. Even though we use the same input point cloud as Point-NeRF\\(^{col}\\), we outperform it greatly. We perform on par with Point-NeRF\\(^{mvs}\\) even though it uses many more points and densifies the point cloud during training. We highlight the best, second, and third values." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.329, + 0.892, + 0.482 + ], + "angle": 0, + "content": "ing code is built on top of the Nerfstudio framework [60], and the tetrahedra field is implemented in CUDA and uses the OptiX library [43]. We train on a single NVIDIA A100 GPU and the training speed ranges from 15k rays per second to 30k rays per second. The speed depends on how well-structured the triangulation is, how many vertices there are, and if there is empty space around the object. The full training with 300k iterations takes between 11 and 24 hours, depending on the scene complexity. However, good results are typically obtained much earlier, e.g., in 100k iterations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.493, + 0.593, + 0.508 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.687 + ], + "angle": 0, + "content": "Blender dataset [39] results. We compare to relevant baselines on the standard Blender dataset. We used the same split and evaluation procedure as in the original NeRF paper [39] and the same SSIM implementation as Point-NeRF [70]. In order to ensure a fair comparison with Point-NeRF [70] when COLMAP points were used, we use the exact same COLMAP reconstruction as Point-NeRF. We report the PSNR, SSIM, and LPIPS (VGG) [78] metrics. Tab. 1 shows averaged results, Fig. 4 shows qualitative results, and the results for individual scenes are given in Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "When we use the exact same COLMAP points as Point-NeRF (row \\( \\text{Point-NeRF}^{col} \\)), we outperform it in all three metrics. We score comparably with \\( \\text{Point-NeRF}^{mvs} \\), even though it starts from a much denser and higher-quality initial point cloud, which it generates from a jointly trained model. Also note that both of these Point-NeRF configurations grow the point cloud during training and, therefore, the complexity of the scene representation grows. For us, the points are fixed, and the number of parameters stays the same. We also outperform Plenoxels [17], which uses a sparse grid. Note that same as Point-NeRF, Plenoxels also gradually increases the representation complexity by subdividing the grid resolution at predefined training epochs. Even though both Mip-NeRF and instant-NGP outperform" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18462" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.094, + 0.468, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.406, + 0.473, + 0.518 + ], + "angle": 0, + "content": "Figure 4. Qualitative results on the Blender dataset. We compare with Point-NeRF\\(^{col}\\) as we use the same input point cloud. In the top-row picture, we can see that Tetra-NeRF is able to represent fine details well on the mic scene. On the drums scene (\\(2^{\\text{nd}}\\) row), both methods struggle with shiny materials, but our method performs slightly better. \\(3^{\\text{rd}}\\) row: Tetra-NeRF can render thin ropes. Bottom row: accumulation is nonzero in areas close to surface." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.529, + 0.45, + 0.589 + ], + "angle": 0, + "content": "
hotdogship
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
Point-NeRFstatic29.910.97819.350.905
Tetra-NeRF33.310.98931.130.994
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.598, + 0.472, + 0.655 + ], + "angle": 0, + "content": "Table 2. Comparison with Point-NeRF with disabled point cloud growing and pruning shows that Point-NeRF performs significantly worse in all measured metrics because it struggles to handle the sparse point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.47, + 0.718 + ], + "angle": 0, + "content": "our approach in terms of PSNR, our method is slightly better in terms of SSIM, and on par with Mip-NeRF in terms of LPIPS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To analyze the tetrahedra field representation, we compare it with the point cloud field representation used in Point-NeRF. In Table 2, we compare our method to Point-NeRF when we disable point cloud growth and pruning. We show the results on two scenes from the Blender dataset [39] which were selected in the Point-NeRF paper. With this setup, we vastly outperform Point-NeRF in all metrics. The reason is that Point-NeRF requires the point cloud to be dense such that all rays have a chance to intersect a neighbourhood of a point. Since we use a continuous representation (tetrahedra field) rather than a discrete one, we can achieve good results even for sparser point clouds." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.696, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.092, + 0.886, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.197, + 0.895, + 0.28 + ], + "angle": 0, + "content": "Figure 5. Comparison with dense grid field representation. Tetra-NeRF on the left, dense grid on the right. We used a comparable number of parameters for both methods (the dense grid having slightly more parameters). Due to the adaptive nature of tetrahedra fields, Tetra-NeRF produces significantly better rendering as it is able to focus on relevant parts of the scene." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.291, + 0.86, + 0.38 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓
NV [34]23.700.8480.260
NeRF [39]25.780.8640.198
NSVF [32]28.400.9000.153
Point-NeRF [70]*28.350.9420.090
Tetra-NeRF28.900.9570.059
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.388, + 0.895, + 0.473 + ], + "angle": 0, + "content": "Table 3. Result on the Tanks and Temples dataset [25] as processed by NSVF [32] averaged over all scenes. We show the PSNR, SSIM, and LPIPS (Alex) [78] metrics. We highlight the best, second, and third values. *Note, that Point-NeRF [70] results differ from the paper as they were recomputed with the resolution used in other methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.895, + 0.879 + ], + "angle": 0, + "content": "Comparison with the dense grid representation. In order to show the utility of the adaptive tetrahedra field representation, we compare it to a dense grid representation. Similarly to NSVF [32], we split the 3D scene bounding box into equally-sized voxels. When querying the field, we find the voxel to which the query point belongs and perform trilinear interpolation of the eight corners of the voxel. We choose the grid resolution such that the number of grid points is the lowest cube number larger than the number of points of the original point cloud. This ensures a fair comparison as the baseline uses a comparable (but larger) number of features than our approach. All other hyperparameters are kept the same as for Tetra-NeRF. We evaluate both approaches on two scenes from the Blender dataset [39]. The dense grid representation only scores PSNR 18.81 and 18.91 on the lego and ship scenes, respectively, whereas Tetra-NeRF scores PSNR 33.79 and 30.69, respectively. The results can be seen in Figure 5. Note that in this experiment, we only trained for 100,000 iterations to save computation time. From the numbers and the figure, we can clearly see that the dense grid resolution is not sufficient to reconstruct the scene in enough detail. Because Tetra-NeRF uses an adaptive subdivision, which is more detailed around the surface, we are better able to focus on relevant scene parts. With a similar number of trainable parameters, we thus obtain better results." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Tanks and Temples [25] dataset. To be able to compare" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18463" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.466, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.488 + ], + "angle": 0, + "content": "Figure 6. Results on Tanks and Temples dataset. In the top row, we can see that we are able to represent the rim of the wheel better than Point-NeRF. Similarly, we can represent the tyre better in the third row. Finally, in the bottom, Point-NeRF fails on the wall and the roof, whereas Tetra-NeRF can render these parts well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.47, + 0.768 + ], + "angle": 0, + "content": "with Point-NeRF [70] on real-world data, we have evaluated Tetra-NeRF on the Tanks and Temples [25] dataset. We use the same setup as in NSVF [32], where the object is masked. We report the usual metrics: PSNR, SSIM, and LPIPS (Alex).3 We used the dense COLMAP reconstruction to get the point cloud used in the tetrahedra field. Quantitative results are shown in Table 3, qualitative results are presented in Figure 6, and the per-scene results are given in the Supp. Mat. Note that the results originally reported in the Point-NeRF paper [70] were evaluated with a resolution different from the compared methods. Therefore, to ensure a fair comparison, we have recomputed the metrics in the publicly available Point-NeRF's predictions with the full resolution of \\(1920 \\times 1080\\) as used in NSVF [32]. In the public dataset, one of the scenes had corrupted camera parameters and we had to reconstruct the poses again for the Ignatius scene. With the reconstructed poses, our method performs slightly worse on that scene compared to others." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.468, + 0.845 + ], + "angle": 0, + "content": "The results show that our method outperforms the baselines in all compared metrics. This indicates that even though Point-NeRF relies on the ability to grow the point cloud density during training, this is not needed when using a continuous instead of a discrete representation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.85, + 0.468, + 0.865 + ], + "angle": 0, + "content": "Sparse and dense point cloud comparison. Previ" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": "3We always choose the type of LPIPS (Alex [27]/VGG [56]) such that we can compare with more methods as some only evaluate using one type." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.091, + 0.695, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.173, + 0.673, + 0.184 + ], + "angle": 0, + "content": "(a) sparse point cloud" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.092, + 0.857, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.173, + 0.839, + 0.184 + ], + "angle": 0, + "content": "(b) ground-truth image" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.187, + 0.695, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.268, + 0.66, + 0.278 + ], + "angle": 0, + "content": "(c) sparse model" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.187, + 0.857, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.268, + 0.822, + 0.278 + ], + "angle": 0, + "content": "(d) dense model" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.279, + 0.892, + 0.32 + ], + "angle": 0, + "content": "Figure 7. Sparse vs. dense model comparison detail. Even though the area in the bottom contains few points, we are still able to reconstruct at least low-frequency data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.514 + ], + "angle": 0, + "content": "ous experiments have used dense COLMAP reconstructions. However, dense point clouds may not be required to achieve good reconstructions. In this section, we compare models trained on dense reconstructions and sparse reconstructions. Table 5 compares PSNR, and SSIM metrics on two synthetic scenes from the Blender dataset [39] and two real-world scenes from the Mip-NeRF 360 dataset [4]. We also show the number of vertices of the tetrahedra, which is the size of the input point cloud, including the random points added at the beginning. Note that in this experiment, we only trained for 100,000 iterations to save computation time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.893, + 0.729 + ], + "angle": 0, + "content": "From the results, we can see that for the Blender dataset [39] case, the dense model is much better. This is to be expected since the number of points obtained from the sparse reconstruction on the Blender dataset is very small and our MLP does not have enough capacity to represent fine details. However, even in regions with zero point coverage, we are still able to provide at least some low-frequency data (c.f. Fig. 7). On real-world scenes, the sparse point cloud model is almost on par with the dense one. The sparse reconstructions on the real-world dataset provide many more points compared to the synthetic dataset case, and the coverage is sufficient. From these results, we conclude that on real-world data the sparse model can be sufficient to achieve good performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Mip-NeRF 360 [4] dataset. We have further evaluated our method on the Mip-NeRF 360 [4] dataset. In order to ensure a fair comparison with Mip-NeRF 360 [4], we trained and evaluated on four times downsized images for the outdoor scenes and two times downsized images for the indoor scenes. The quantitative results are presented in Table 4, and the qualitative results are shown in Figure 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "From the results, we can see that Tetra-NeRF is able to outperform both the vanilla NeRF [39] and Mip-NeRF [3]. We also outperform Stable View Synthesis [51] in terms of PSNR and score comparably in terms of LPIPS. This is pos" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18464" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.088, + 0.887, + 0.231 + ], + "angle": 0, + "content": "
OutdoorIndoorMean
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [15, 39]21.460.4580.51526.840.7900.37023.850.6050.451
mip-NeRF [3]21.690.4710.50526.980.7980.36124.040.6160.441
NeRF++ [77]22.760.5480.42728.050.8360.30925.110.6760.375
Deep Blending [20]21.540.5240.36426.390.8440.26123.700.6660.318
Point-Based Neural Rendering [26]21.660.6120.30226.280.8870.19123.710.7340.253
Stable View Synthesis [51]23.010.6620.25328.220.9070.16025.330.7710.211
mip-NeRF 360 [4]23.720.6870.28229.430.9110.18126.260.7860.237
Tetra-NeRF23.170.5860.29830.210.8810.10326.300.7170.211
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.24, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Table 4. Mip-NeRF 360 dataset results. We show the PSNR, SSIM, and LPIPS (Alex) [78] on two categories of Mip-NeRF 360 [4] scenes: outdoor, and indoor. On the outdoor scenes, we outperform the Stable View Synthesis [51] and are comparable to Mip-NeRF 360 [4], even though our method does not implement the improvements suggested in NeRF++ [77] and Mip-NeRF 360 [4], and is more comparable to vanilla NeRF [39]. We highlight the best, second, and third values." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.306, + 0.46, + 0.394 + ], + "angle": 0, + "content": "
sparsedense
PSNR/SSIM#pointsPSNR/SSIM#points
Blender/lego29.77/0.95925,78433.79/0.985302,781
Blender/ship26.90/0.8997,15230.69/0.942321,861
360/bonsai28.12/0.902413,22628.34/0.9021,000,000
360/garden24.79/0.806227,53225.41/0.8381,000,000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.408, + 0.471, + 0.465 + ], + "angle": 0, + "content": "Table 5. Dense and sparse point cloud comparison. We present results on two scenes from the Blender [39] and Mip-NeRF 360 [4] datasets. We also show the number of vertices of the tetrahedra field – including the randomly sampled points." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.473, + 0.464, + 0.773 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.775, + 0.471, + 0.847 + ], + "angle": 0, + "content": "Figure 8. Mip-NeRF 360 results. \\(1^{\\text{st}}\\) row: we show the results on the garden scene. \\(2^{\\text{nd}}\\): we can see that Tetra-NeRF has some problems rendering grass. \\(3^{\\text{rd}}\\): Tetra-NeRF is able to represent the delicate texture of the tablecloth well. \\(4^{\\text{th}}\\): Tetra-NeRF is able to recover fine geometry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "sible because outdoor scenes contain a lot of high-resolution geometries, such as leaves and Stable View Synthesis is not able to overcome noise in the approximate geometry. Tetra-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.895, + 0.432 + ], + "angle": 0, + "content": "NeRF does not suffer from these problems as we aggregate features in 3D along rays rather than on the surface. Mip-NeRF 360 scores comparably to Tetra-NeRF in terms of PSNR and LPIPS, but it outperforms Tetra-NeRF in terms of SSIM. However, Mip-NeRF 360 implements some tricks designed to boost its performance. On the other hand, our approach is based on vanilla NeRF, and we believe that the performance can be boosted similarly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.893, + 0.695 + ], + "angle": 0, + "content": "Varying the number of input points We conducted a study on the effect of the input point cloud size on the reconstruction quality. We consider both the sparse and the dense point clouds and analyse the performance as we subsample the points or add more points randomly. We evaluate on the family and the truck scenes from the Tanks and Temples [25] (tt) dataset, and the room and the garden scenes from the Mip-NeRF 360 [4] dataset. The sizes of sparse point clouds were roughly 16K, 30K, 113K, and 139K respectively and the sizes of dense point clouds were larger than 5M for all scenes. When the size of the original point cloud was larger than the desired size, we selected points uniformly at random, and when smaller, we added new random points as described Section 4.1. We include detailed results in the Supp. Mat. In order to save computational resources, we only train the method for 100k iterations. The results are visualised in Figure 9a." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.895, + 0.82 + ], + "angle": 0, + "content": "As expected, the performance improves with the number of points used, as it leads to a finer subdivision of the scene around the surface. We can also observe (e.g., from tt/family) that the sparse reconstruction leads to a better reconstruction quality at the same point cloud size. The likely explanation is that sampling uniformly at random will be biased towards regions with high density and some regions may be missed or covered extremely sparsely." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Speed of convergence Figures 9b, 9c show the rate of convergence on the ship scene from the Blender dataset and aggregated results on the Tanks and Temples [25], and Mip-NeRF 360 (indoor/outdoor) [4] datasets. Similar to Point-NeRF, good results can be achieved quite early in the train" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18465" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.085, + 0.096, + 0.349, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.093, + 0.616, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.099, + 0.882, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.892, + 0.352 + ], + "angle": 0, + "content": "Figure 9. a) shows the PSNR with different sizes of input point clouds. The solid and dashed lines represent the dense and coarse COLMAP reconstructions. The following scenes were evaluated (best to worst): tt/family, 360/room, tt/truck, and 360/garden from Tanks and Temples [25] (tt) and Mip-NeRF 360 [4] (360-indoor/360-outdoor) datasets. As expected, the quality increases with the number of points. b) visualizes the training speed (PSNR at different training iterations) on the blender/ship scene [39]. It shows Tetra-NeRF training times for various timesteps and compares the rate of convergence with Point-NeRF [70] and Mip-NeRF [3]. Similarly, c) shows the rate of convergence aggregated over Tanks and Temples [25] (tt) and Mip-NeRF 360 [4] (360-indoor/360-outdoor) datasets. It shows the average PSNR, and the highlighted area represents the standard deviation." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.364, + 0.276, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.462, + 0.272, + 0.485 + ], + "angle": 0, + "content": "(a) Failure case 1: low point cloud density on the ground" + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.363, + 0.473, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.463, + 0.453, + 0.485 + ], + "angle": 0, + "content": "(b) Failure case 2: too many intersected tetrahedra" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.485, + 0.468, + 0.512 + ], + "angle": 0, + "content": "Figure 10. Failure cases. Left: barn scene from [25]. Right: tree hill scene from [4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.468, + 0.692 + ], + "angle": 0, + "content": "ing process. E.g., on the blender/ship scene, in 24 minutes, our method achieves performance similar to Mip-NeRF [3] after several hours of training. The efficiency of the implementation is a large factor in the speed of different methods. E.g., in Instant-NGP [41], the authors invested a significant effort in optimizing GPU memory accesses, and the result is highly optimised efficient code (a main contribution of [41]). There are several places where our implementation can easily be made more efficient. E.g., changing the memory layout (row-major vs. column-major) of the tetrahedra field leads to a \\(10\\%\\) speedup." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.708, + 0.2, + 0.723 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.468, + 0.84 + ], + "angle": 0, + "content": "One drawback of our approach is that the quality of different regions of the rendered scene depends on the density of the point cloud in these regions. If the original point cloud was constructed using 2D feature matching, there can be regions with a very low number of 3D points. We show such an example in Fig. 10a. There are few 3D points on the ground, resulting in blurry renderings for that region." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Another issue is that the current implementation has a limit on the number of intersected tetrahedra per ray. For large or badly structured scenes, where rays intersect many tetrahedra, this can limit the reconstruction quality. Fig. 10b" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.453 + ], + "angle": 0, + "content": "shows such a case on a Mip-NeRF 360 scene [4]. This problem can be addressed by increasing the limit on the number of visited tetrahedra at the cost of higher memory requirements and run-time. Alternatively, coarse-to-fine schemes that start with a coarser tetrahedralisation and prune the space could potentially handle this issue." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.473, + 0.619, + 0.489 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.803 + ], + "angle": 0, + "content": "This paper proposes a novel radiance field representation that, compared to standard voxel-based representations, is easily able to adapt to 3D geometry priors given in the form of a (sparse) point cloud. Our approach elegantly combines concepts from geometry processing (Delaunay triangulation) and triangle-based rendering (ray-triangle intersections) with modern neural rendering approaches. The representation has a naturally higher resolution in the space near surfaces, and the input point cloud provides a straightforward way to initialise the radiance field. Compared to Point-NeRF, a state-of-the-art point cloud-based radiance field representation which uses the same input as our approach, Tetra-NeRF shows clearly better results. Our method performs comparably to state-of-the-art MLP-based methods. The results demonstrate that Tetra-NeRF is an interesting alternative to existing radiance field representations that is worth further investigation. Interesting research directions include adaptive refinement and pruning of the tetrahedralisation and exploiting the fact that the surface of the scene is likely close to some of the triangles in the scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements This work was supported by the Czech Science Foundation (GACR) EXPRO (grant no. 23-07973X), the Grant Agency of the Czech Technical University in Prague (grant no. SGS22/112/OHK3/2T/13), and by the Ministry of Education, Youth and Sports of the Czech Republic through the e-INFRA CZ (ID:90254)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18466" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.468, + 0.225 + ], + "angle": 0, + "content": "[2] Pierre Alliez, Simon Giraudot, Clément Jamin, Florent Lafarge, Quentin Merigot, Jocelyn Meyron, Laurent Saboret, Nader Salman, Shihao Wu, and Necip Fazil Yildiran. Point set processing. In CGAL User and Reference Manual. CGAL Editorial Board, 5.5.1 edition, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.468, + 0.295 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. In ICCV, 2021. 1, 2, 4, 5, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.468, + 0.352 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1, 2, 5, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.468, + 0.422 + ], + "angle": 0, + "content": "[5] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv:2008.03824, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.468, + 0.477 + ], + "angle": 0, + "content": "[6] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. NoPe-NeRF: Optimising neural radiance field with no pose prior. arXiv:2212.07388, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.468, + 0.522 + ], + "angle": 0, + "content": "[7] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. NeRD: Neural reflectance decomposition from image collections. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.468, + 0.578 + ], + "angle": 0, + "content": "[8] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.468, + 0.619 + ], + "angle": 0, + "content": "[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial radiance fields. In ECCV, 2022. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.468, + 0.662 + ], + "angle": 0, + "content": "[10] Anpei Chen, Zexiang Xu, Xinyue Wei, Siyu Tang, Hao Su, and Andreas Geiger. Factor fields: A unified framework for neural fields and beyond. arXiv:2302.01226, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.468, + 0.718 + ], + "angle": 0, + "content": "[11] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[12] Shuo Cheng, Zexiang Xu, Shilin Zhu, Zhuwen Li, Li Erran Li, Ravi Ramamoorthi, and Hao Su. Deep stereo using adaptive thin volume representation with uncertainty awareness. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[13] Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[14] B Delaunay, S Vide, A Lamémoire, and V De Georges. Bulletin de l'académie des sciences de l'URSS. Classe des sciences mathématiques et naturelles, 6:793-800, 1934. 2, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[15] Boyang Deng, Jonathan T. Barron, and Pratul P. Srinivasan. JaxNeRF: an efficient JAX implementation of NeRF, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[16] Michael S Floater. Generalized barycentric coordinates and applications. Acta Numerica, 24:161-214, 2015. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[17] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.265 + ], + "angle": 0, + "content": "[18] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32:1362-1376, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.268, + 0.892, + 0.308 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.311, + 0.892, + 0.365 + ], + "angle": 0, + "content": "[20] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. SIGGRAPH, ToG, 37(6):1-15, 2018. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.368, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[21] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[22] Vu Hoang Hiep, Renaud Keriven, Patrick Labatut, and Jean-Philippe Pons. Towards high-resolution large-scale multiview stereo. In CVPR, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[23] Michal Jancosek and Tomas Pajdla. Multi-view reconstruction preserving weakly-supported surfaces. In CVPR, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.513, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[24] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In SGP, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.543, + 0.892, + 0.583 + ], + "angle": 0, + "content": "[25] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. SIGGRAPH, ToG, 2017. 5, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.586, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[26] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, volume 40, pages 29-43. Wiley Online Library, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.892, + 0.698 + ], + "angle": 0, + "content": "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[28] Jonás Kulhánek, Erik Derner, Torsten Sattler, and Robert Babuška. ViewFormer: Nerf-free neural rendering from few images using transformers. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[29] Patrick Labatut, Jean-Philippe Pons, and Renaud Keriven. Efficient multi-view reconstruction of large-scale scenes using interest points, Delaunay triangulation and graph cuts. In ICCV, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[30] Patrick Labatut, J-P Pons, and Renaud Keriven. Robust and efficient surface reconstruction from range data. In Computer graphics forum, volume 28, pages 2275-2290. Wiley Online Library, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[31] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. BARF: Bundle-adjusting neural radiance fields. In ICCV, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18467" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[32] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[33] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond. In International Conference on Learning Representations, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.47, + 0.263 + ], + "angle": 0, + "content": "[34] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: learning dynamic renderable volumes from images. SIGGRAPH, ToG, 38(4):1-14, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.469, + 0.307 + ], + "angle": 0, + "content": "[35] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3D surface construction algorithm. SIGGRAPH, ToG, 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.469, + 0.364 + ], + "angle": 0, + "content": "[36] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.367, + 0.469, + 0.407 + ], + "angle": 0, + "content": "[37] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.41, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[38] Moustafa Meshry, Dan B. Goldman, Sameh Khamis, Hugues Hoppe, Rohit Pandey, Noah Snavely, and Ricardo MartinBrualla. Neural rerendering in the wild. In \\(CVPR\\), June 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.468, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 4, 5, 6, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.553 + ], + "angle": 0, + "content": "[40] Tomas Möller. A fast triangle-triangle intersection test. Journal of Graphics Tools, 2(2):25-30, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.469, + 0.609 + ], + "angle": 0, + "content": "[41] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. SIGGRAPH, ToG, 2022. 1, 2, 3, 4, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.469, + 0.654 + ], + "angle": 0, + "content": "[42] M. Nießner, M. Zollhöfer, S. Izadi, and M. Stamminger. Real-time 3D reconstruction at scale using voxel hashing. SIGGRAPH, ToG, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.726 + ], + "angle": 0, + "content": "[43] Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith Morley, Austin Robison, et al. OptiX: a general purpose ray tracing engine. ACM Transactions on Graphics (ToG), 29(4):1-13, 2010. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.728, + 0.469, + 0.769 + ], + "angle": 0, + "content": "[44] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[45] Bui Tuong Phong. Illumination for computer generated pictures. Communications of the ACM, 18(6):311-317, 1975. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. arXiv:2209.14988, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[47] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding up neural radiance fields with thousands of tiny mlp's. In ICCV, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[48] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T Barron, and Peter Hedman. MERF: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. arXiv:2302.12249, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[49] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[50] Gernot Riegler and Vladlen Koltun. Free view synthesis. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.252, + 0.892, + 0.278 + ], + "angle": 0, + "content": "[51] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In CVPR, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.282, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[52] Radu Alexandru Rosu and Sven Behnke. Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.326, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[53] Darius Rückert, Linus Franke, and Marc Stamminger. ADOP: Approximate differentiable one-pixel point rendering. SIGGRAPH, ToG, 41(4):1-14, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.369, + 0.892, + 0.396 + ], + "angle": 0, + "content": "[54] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.399, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[55] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.443, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[56] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv:1409.1556, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.487, + 0.892, + 0.527 + ], + "angle": 0, + "content": "[57] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.53, + 0.892, + 0.585 + ], + "angle": 0, + "content": "[58] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-NeRF: Scalable large scene neural view synthesis. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.588, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[59] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[60] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, et al. Nerfstudio: A modular framework for neural radiance field development. arXiv:2302.04264, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[61] Chengzhou Tang and Ping Tan. BA-net: Dense bundle adjustment networks. In International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[62] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[63] Sudheendra Vijayanarasimhan, Susanna Ricco, Cordelia Schmid, Rahul Sukthankar, and Katerina Fragkiadaki. SfM-Net: Learning of structure and motion from video. arXiv:1704.07804, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18468" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[64] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-NeRF: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[65] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34:27171-27183, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.274 + ], + "angle": 0, + "content": "[66] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snively, and Thomas Funkhouser. IBRNet: Learning multi-view image-based rendering. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[67] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. NeuS2: Fast learning of neural implicit surfaces for multi-view reconstruction. arXiv:2212.05231, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[68] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF-: Neural radiance fields without known camera parameters. arXiv:2102.07064, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.471, + 0.445 + ], + "angle": 0, + "content": "[69] Daniel N Wood, Daniel I Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H Salesin, and Werner Stuetzle. Surface light fields for 3D photography. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[70] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-NeRF: Point-based neural radiance fields. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.544 + ], + "angle": 0, + "content": "[71] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multiview stereo. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[72] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[73] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. BakedSDF: Meshing neural SDFs for real-time view synthesis. arXiv:2302.14859, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[74] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[76] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[77] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. NeRF++: Analyzing and improving neural radiance fields. arXiv:2010.07492, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[78] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, June 2018. 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.148 + ], + "angle": 0, + "content": "[79] Chaoqiang Zhao, Qiyu Sun, Chongzhen Zhang, Yang Tang, and Feng Qian. Monocular depth estimation based on deep learning: An overview. Science China Technological Sciences, 63(9):1612-1627, 2020. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18469" + } + ] +] \ No newline at end of file diff --git a/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/8c410ac7-5eb0-4f17-a4bb-daadb2446edc_origin.pdf b/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/8c410ac7-5eb0-4f17-a4bb-daadb2446edc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9a9f07585763b6b2b0836b6cbcf8d6cd70dd1374 --- /dev/null +++ b/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/8c410ac7-5eb0-4f17-a4bb-daadb2446edc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:577f9331989fdd85180b247d038d0d230620d2dabe9b34e6893f1594185c5a29 +size 6348438 diff --git a/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/full.md b/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/full.md new file mode 100644 index 0000000000000000000000000000000000000000..51d6756ffc71ef6f466ad0afeae1baceb3fa6193 --- /dev/null +++ b/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/full.md @@ -0,0 +1,348 @@ +# Tetra-NeRF: Representing Neural Radiance Fields Using Tetrahedra + +Jonas Kulhanek +Czech Technical University in Prague +jonas.kulhanek@cvut.cz + +Torsten Sattler +Czech Technical University in Prague +torsten.sattler@cvut.cz + +![](images/5d718ca298af7923b76dae45526cd7fc1c5c03f76746559ba4400da2025a3561.jpg) +3) barycetric interpolation of features 4) shallow MLP predicts density and colour +Figure 1. The input to Tetra-NeRF is a point cloud which is triangulated to get a set of tetrahedra used to represent the radiance field. Rays are sampled, and the field is queried. Barycentric interpolation is used to interpolate tetrahedra vertices, and the resulting features are passed through a shallow MLP to get the density and colours for volumetric rendering. + +# Abstract + +Neural Radiance Fields (NeRFs) are a very recent and very popular approach for the problems of novel view synthesis and 3D reconstruction. A popular scene representation used by NeRFs is to combine a uniform, voxel-based subdivision of the scene with an MLP. Based on the observation that a (sparse) point cloud of the scene is often available, this paper proposes to use an adaptive representation based on tetrahedra obtained by Delaunay triangulation instead of uniform subdivision or point-based representations. We show that such a representation enables efficient training and leads to state-of-the-art results. Our approach elegantly combines concepts from 3D geometry processing, triangle-based rendering, and modern neural radiance fields. Compared to voxel-based representations, ours provides more detail around parts of the scene likely to be close to the surface. Compared to point-based representations, our approach achieves better performance. The source code is publicly available at: https://jkulhanek.com/tetra-nerf. + +# 1. Introduction + +Reconstructing 3D scenes from images and rendering photo-realistic novel views is a key problem in computer vision. Recently, NeRFs [3,4,39] became dominant in the field for their superior photo-realistic results. Originally, + +NeRFs used MLPs to represent the 3D scene as an implicit function. Given a set of posed images, NeRF randomly samples a batch of pixels, casts rays from the pixels into the 3D space, queries the implicit function at randomly sampled distances along the rays, and aggregates the sampled values using volumetric rendering [37, 39]. While the visual results of such methods are of high quality, the problem is that querying large MLPs at millions of points is costly. Also, once the network is trained, it is difficult to make any changes to the represented radiance field as everything is baked into the MLPs parameters, and any change has a non-local effect. Since then, there have been a lot of proposed alternatives to the large MLP field representation [9, 17, 32, 41, 57, 70, 74]. These methods combine an MLP with a voxel feature grid [41, 57], or in some cases represent the radiance field directly as a tensor [9, 10, 17]. When querying these representations, first, the containing voxel is found, and the features stored at the eight corner points of the voxel are trilinearly interpolated. The result is either passed through a shallow MLP [9, 10, 41, 57] or is used directly as the density and colour [17, 32, 57]. + +Having a dense tensor represent the entire scene is very inefficient, as we only need to represent a small space around surfaces. Therefore, different methods propose different ways of tackling the issue. Instant-NGP [41], for example, uses a hash grid instead of a dense tensor, where it relies on optimisation to resolve the hash collisions. How- + +ever, similarly to MLPs, any change to the stored hashmap influences the field in many places. A more common direction to addressing the issue is by directly using a sparse tensor representation [9, 17]. These methods start with a low-resolution grid and, at predefined steps, subsample the representation, increasing the resolution. These approaches tend to require a careful setting of hyperparameters, such as the scene bounding box and the subdivision steps, in order for the methods to work well. + +Because many of these methods use traditional structure from motion (SfM) [54, 55] methods to generate the initial poses for the captured images, we can reuse the original reconstruction in the scene representation. Inspired by classical surface reconstruction methods [22, 23, 29, 30], we represent the scene as a dense triangulation of the input point cloud, where the scene is a set of non-overlapping tetrahedra whose union is the convex hull of the original point cloud [14]. When querying such a representation, we find to which tetrahedron the query point belongs and perform barycentric linear interpolation of the features stored in the vertices of the tetrahedron. This very simple representation can be thought of as the direct extension of the classical triangle-rendering pipelines used in graphics [40, 43, 45]. The representation avoids problems with the sparsity of the input point cloud as the tetrahedra fully cover the scene, resulting in a continuous rather than discrete representation. + +This paper makes the following contributions: (1) We propose a novel radiance field representation which is initialised from a sparse or dense point cloud. This representation is naturally denser in the proximity of surfaces and, therefore, provides a higher resolution in these regions. (2) The proposed representation is evaluated on multiple synthetic and real-world datasets and is compared with a state-of-the-art point-cloud-based representation - Point-NeRF [70]. The presented results show that our method clearly outperforms this baseline. We further demonstrate the effectiveness of our adaptive representation by comparing it with a voxel-based representation that uses the same number of trainable parameters. (3) We make the source code and model checkpoints publicly available. + +# 2. Related work + +Multi-view reconstruction. The problem of multi-view reconstruction has been studied extensively and tackled with a variety of structure from motion (SfM) [54, 61, 63], and multi-view stereo (MVS) [12, 18, 55, 71] methods. These methods usually output the scene represented as a point cloud [54, 55]. In most rendering approaches, the point cloud is converted into a mesh [24, 35], and novel views are rendered by reprojecting observed images into each novel viewpoint and blending them together using ei- + +ther heuristically-defined [8, 13, 69] or learned [20, 50, 51, 73] blending weights. + +However, the process of getting the meshes is usually quite noisy, and the resulting meshes tend to have inaccurate geometry in regions with fine details or complex materials. Instead of using noisy meshes, point-based neural rendering methods [1, 26, 38, 53] perform splatting of neural features and use 2D convolutions to render them. In contrast to these methods, our approach operates and aggregates features directly in 3D and does not suffer from the noise in the point cloud or the reconstructed mesh. + +Neural radiance fields. Recently, NeRFs [3,4,34,39,77] have gained a lot of attention thanks to their high-quality rendering performance. The original NeRF method [39] was extended to better handle aliasing artefacts in [3], to better represent unbounded scenes in [4,48,77], or to handle real-world captured images [36,58]. The training of the large MLPs used in these methods can be quite slow, and there has been a lot of effort on speeding up either the training [9,17,41] or the rendering [21,47,48,74] sometimes at the cost of larger storage requirements. Other approaches tackled different aspects of NeRFs like view-dependent artefacts [62], relighting [5,7], or proposed generative models [46,64]. Also, a popular research direction is making the models generalize across different scenes [11,28,49,66,75]. A large area of research is dedicated to the surface reconstruction and, instead of using the radiance fields, represents the scene implicitly by modelling the signed distance function (SDF) [52,65,67,72,76]. Unlike those approaches, we focus only on the radiance field representation and consider these methods orthogonal to ours. + +Although there are some methods that train radiance fields while fine-tuning the poses [31, 60] or without known cameras [6, 68], most methods need camera poses for the reconstruction. SfM, e.g., COLMAP [54, 55], is typically used for estimating the poses, which also produces a (sparse) point cloud. Our approach makes use of this by-product of the pose recovery process instead of only using the poses themselves. + +Field representations. When a single MLP is used to represent the entire scene, everything is baked into a single set of parameters which cannot be easily modified, because any change to the parameters has a non-local effect, i.e., it changes the scene at multiple unrelated places. To overcome this problem, others have experimented with different representations of the radiance fields [9, 10, 17, 32, 41, 44, 57, 70]. A common practice is to represent the scene as a shallow MLP and an efficiently encoded voxel grid [9, 10, 32, 41, 44]. The encoded voxel grid can also be used represent the radiance field directly [17, 57]. The voxel grid can be encoded as a sparse tensor [17, 32, 57], a factorisation of the 4D tensor [9, 10], or a hashmap [41]. When + +![](images/7abd8e15c57f18ac8cd3c30e5301bc371167faebce01af3348e4d354d3156e33.jpg) +(a) dense point cloud + +![](images/7863d3ad9175ea27038d1cfdc437f43a530d2fdd5cf19f167a5d9461cf3284e2.jpg) + +![](images/5f74a44cddba3fcb15c66ff5be32b246eb46c7aa4a1dd90de9b54feda4e1b0a2.jpg) +(c) sparse point cloud + +![](images/326c524dd6484d7b5f579d4b15d08a3f0aa3570b24e5bfc89431dc848ea7bda9.jpg) +(b) dense tetrahedra slice +(d) sparse tetrahedra slice +Figure 2. Input point cloud and a slice through the triangulated tetrahedra. Note that smaller tetrahedra are created closer to the surface of the scene, i.e., regions close to the surface are represented with a finer resolution. +Fig. 2). In the following, we explain how this adaptive subdivision of the scene can be used instead of voxels for volume rendering and neural rendering. + +these structures are queried, trilinear interpolation is used to combine the feature vectors stored in the containing voxel corners. Unfortunately, the hashes [41] and hierarchical representations [10] have the same non-local effect problem, and the rest of the approaches rely on subsequent upsampling of the field and can be overly complicated. Also, feature vectors cannot be placed arbitrarily in the 3D space as they must lie on the grid with a fixed resolution. In contrast, our approach is much more flexible as it stores the feature vectors freely in 3D space. + +Finally, Point-NeRF [70] represents the scene as a point cloud of features which are queried using $k$ -nearest neighbours search. However, when the point cloud contains sparse regions, the rays do not intersect any neighbourhoods of any points and the pixels stay empty without the ability to optimise. Therefore, Point-NeRF [70] relies on gradually adding more points during training and increasing the scene complexity. Since we use the triangulation of the point cloud rather than the discrete point cloud itself, our representation is continuous and does not suffer from empty regions. Therefore, we do not have to add any points during training. + +# 3. Method + +A common strategy in the literature is to represent the scene explicitly through a voxel volume. In contrast to this uniform subdivision, we investigate using an adaptive subdivision of the scene. In many scenarios, an approximation of the scene geometry is either given, e.g., when using SfM to compute the input camera poses, or can be computed, e.g., via MVS or single-view depth predictions [79]. This allows us to compute an adaptive subdivision of the scene via Delaunay triangulation [14] of such a point cloud. This results in a set of non-overlapping tetrahedra, where smaller tetrahedra are created close to the surface of the scene (c.f. + +![](images/7898f45a2777e4512c769d742cb35004c7be94bd2da66bdbad9dec6738eb60b0.jpg) +(a) trilinear interpolation +Figure 3. Trilinear and barycentric interpolation. Trilinear interpolation is a weighted combination of the eight voxel corners. Barycentric interpolation weights the four vertices of the tetrahedron vertices based on the barycentric coordinates [16]. + +![](images/e7480085e9cc46610d1b010b85b34b8598fda893d58e624a4e3789cb494ef03e.jpg) +(b) barycentric interpolation + +# 3.1. Preliminaries + +Neural Radiance Fields (NeRFs) [39] represent a scene through an implicit function $F(\mathbf{x}, \mathbf{d})$ , often modelled via a neural network, that returns a colour value and a volume density prediction for a given 3D point $\mathbf{x}$ in the scene observed from a viewing direction $\mathbf{d}$ . Volume rendering [37] is used to synthesise novel views: for each pixel in a virtual view, we project a ray from the camera plane into the scene and sample the radiance field to obtain the colour and density values $\mathbf{c}_i$ and $\sigma_i$ at distances $t_i$ along the ray, where $i = 1 \dots N$ . The individual samples are then combined to predict the colour $\mathbf{C}$ for the pixel: + +$$ +\mathbf {C} = \sum_ {i = 1} ^ {N} T _ {i} \left(1 - \exp \left(\sigma_ {i} \delta_ {i}\right)\right) \mathbf {c} _ {i}, \text {w h e r e} \tag {1} +$$ + +$$ +T _ {i} = \exp \left(- \sum_ {i = 1} ^ {N - 1} \sigma_ {i} \delta_ {i}\right), +$$ + +and $\delta_{i} = t_{i} - t_{i - 1}$ is the distance between adjacent samples. This process is fully differentiable. Thus, after computing the mean squared error (MSE) between the predicted and ground truth colours associated with each ray, we can backpropagate the gradients to the radiance field. + +Voxel-based feature fields such as NSVF [32] represent the scene as a voxel grid and an MLP. Each grid point of the voxel grid is assigned a trainable vector. There are eight feature vectors associated with a single voxel, but the vectors are shared between neighbouring voxels. For each query point sampled along the ray, its corresponding voxel is found first. A feature for the point is then computed via trilinear interpolation of the features of the voxel grid points based on the position of the query point (c.f. Fig. 3, left). The resulting feature vector is passed through an MLP to predict the density and appearance vector. The appearance vector is combined with the ray direction and + +passed through a second MLP in order to compute a view-dependent colour. + +# 3.2. Tetrahedra fields + +Given a set of points in 3D space, we build the tetrahedra structure by triangulating the points. We apply the Delaunay triangulation [14] to obtain a set of non-overlapping tetrahedra whose union is the convex hull of the original points. Fig. 2 shows example tetrahedra obtained by triangulating dense and sparse COLMAP-reconstructed point clouds [54, 55]. Note that the resulting representation is adaptive as it uses a higher resolution (smaller tetrahedra) closer to the surface and larger tetrahedra to model regions farther away from the surface. + +We associate all vertices of the tetrahedra with trainable vectors. As in the voxel grid case, vertices, and thus vectors, are shared between adjacent tetrahedra. The resulting tetrahedra field can be queried in the same way a voxel grid representation is queried: for each query point, we first find the tetrahedron containing the point. Instead of the trilinear interpolation used for voxel volumes, we use the barycentric interpolation [16] to compute a feature vector for the query point from the four feature vectors stored at the tetrahedron's vertices (c.f. Fig. 3). To this end, we compute the query point $\mathbf{x}$ 's barycentric coordinates $\lambda$ , which express the point's 3D coordinates as a unique weighted combination of the 3D coordinates of the tetrahedron's vertices. In particular, the weight for a vertex is the volume of the tetrahedron constructed from the query point and the face opposite to the vertex divided by the volume of the full tetrahedron: + +$$ +\boldsymbol {\lambda} = \left(\frac {V _ {x 2 3 4}}{V _ {1 2 3 4}}, \frac {V _ {1 x 3 4}}{V _ {1 2 3 4}}, \frac {V _ {1 2 x 4}}{V _ {1 2 3 4}}, \frac {V _ {1 2 3 x}}{V _ {1 2 3 4}}\right), \tag {2} +$$ + +where $V_{1234}$ is the volume of the full tetrahedron and $V_{x234}$ , $V_{1x34}, \ldots$ , are volumes of tetrahedra with $1^{\mathrm{st}}$ , $2^{\mathrm{nd}}$ , ..., vertex replaced by $\mathbf{x}$ . The same weights $\lambda$ are applied to the feature vectors of the vertices to obtain the query feature. + +The interpolated features are used as the input to a small MLP in order to predict density and colour at the query point. We first pass the barycentric-interpolated features through a three-layer MLP to compute the density and appearance features. We then concatenate the appearance features with the ray direction vector encoded using Fourier features [39, 59] and pass the result through a single linear layer to get the raw RGB colour value. + +Finally, to map the raw density values $\bar{\sigma}_i$ returned by the network to the volume density $\sigma_{i}$ required by volume rendering, we apply the softplus activation function [3]. For the RGB colour value, we use the sigmoid function [17]. + +# 3.3. Efficiently querying a tetrahedra field + +Determining the corresponding voxel for a given query point can be done highly efficiently via hashing [42]. In + +contrast, finding the corresponding tetrahedron for a query point is more complex. This in turn can significantly impact rendering, and thus training, efficiency [17,41,57]. + +In order to efficiently look up the corresponding tetrahedra, we exploit that we are not considering isolated points, but points sampled from a ray. We compute the tetrahedra that are intersected by the ray, allowing us to march through the tetrahedra rather than computing them individually per point. The relevant tetrahedra can be found efficiently using acceleration structures for fast ray-triangle intersection computations, e.g., via NVidia's OptiX library [43]: We first compute the intersections between the rays from a synthetic view and all faces of the tetrahedra. For each ray, we take the first 512 intersected triangles², and determine the corresponding tetrahedra. The tetrahedra can be ordered based on the intersections along the ray, allowing us to easily march through the tetrahedra to determine which tetrahedron to use for a given query point. + +A side benefit of computing ray-triangle intersections is that we can simplify computing the barycentric coordinates: For each intersection, we compute the 2D barycentric coordinates w.r.t. the triangle. We then obtain the 3D barycentric coordinates w.r.t. the associate tetrahedron by simply adding zero for the vertex opposite to the triangle. For a query point inside a tetrahedron, we can compute its tetrahedron barycentric coordinates by linearly interpolating between the barycentric coordinates of the two intersections of the ray and the tetrahedron. + +# 3.4. Coarse and fine sampling + +We follow the common practice of having a two-stage sampling procedure [9, 39]. In the coarse stage, we sample uniformly along the ray. In the fine stage, we use the density weights from the coarse sampling stage to bias the sampling towards sampling closer to the potential surface. Following [39], we use the stratified uniform sampling for the coarse stage. The stratified uniform sampling splits the ray into equally long intervals and samples uniformly in each interval. Unlike NeRF [39], we limit the sampling to the space occupied by tetrahedra. In the fine sampling stage, we use the same network as in the coarse sampling stage. + +For the fine sampling stage, we take the accumulated weights $w_{i}$ from the coarse sampling: + +$$ +\bar {w} _ {i} = \left(1 - \exp \left(- \sigma_ {i} \delta_ {i}\right)\right) \exp \left(- \sum_ {j = 1} ^ {i - 1} \sigma_ {j} \delta_ {j}\right). \tag {3} +$$ + +These weights are the coefficients used in Equation 1 as multipliers for the colours [39]. We obtain $w_{i}$ by normalizing $\bar{w}_{i}$ . Following [39], we sample set of fine samples using + +weight $w_{i}$ . To render the final colour, we merge the dense and fine samples and use all in the rendering equation [9]. + +# 4. Experiments + +We compare Tetra-NeRF to relevant methods on the commonly used synthetic Blender [39], the real-world Tanks and Temples [25], and the challenging object-centric Mip-NeRF 360 [4] datasets. To show its effectiveness, we compare Tetra-NeRF to a dense-grid representation and evaluate it with reduced quality of the input point cloud. We start by describing the exact hyperparameters used. + +# 4.1. Implementation details + +Generating point cloud & triangulation. Given a set of posed images, we use the COLMAP reconstruction pipeline [54, 55] to get the point cloud used in our tetrahedra field representation. We then reduce the size of the resulting point cloud such that if the number of points is larger than $10^{6}$ , we subsample $10^{6}$ points randomly. For the Blender dataset experiments, where the number of points is smaller, we add more randomly generated points. In that case, the number of random points is half the number of original points. The reason for adding the points is that with a low number of points, some pixels on edges may not intersect any tetrahedra, potentially producing artefacts on the edges. Each added point is sampled as follows: we sample a random point $x_{0}$ from the original point cloud, sample a random normal vector $n$ , and a number $\alpha \sim \mathcal{N}(\bar{d}, \bar{d}^{2})$ , where $\bar{d}$ is the average spacing of the original point cloud, i.e. the average distance between each point and its six closest neighbours. We then add the point $x = x_{0} + \alpha n$ . + +Initialization. Given the processed point cloud, we apply the Delaunay triangulation [14] to get a set of tetrahedra. To this end, we use the CGAL library [2], which in our experiments runs in the order of milliseconds. Following [41], we initialise the features of size 64 at the vertices of the tetrahedra with small values sampled uniformly in the range $-10^{-4}$ to $10^{-4}$ . However, to allow the model to reuse the information contained in the point cloud, we set the first four dimensions of the feature field to the RGBA colour values (rescaled to interval [0, 1]) stored at the associated points in the point cloud. The alpha value of all original points is one, whereas all randomly sampled points have an alpha value of zero. For the MLP, we follow the common practice of using the Kaiming uniform initialisation [19]. The hidden sizes in all MLPs are 128. + +Training. During training, we sample batches of 4,096 random rays from random training images. We use volumetric rendering to predict the colour of each ray. The gradients are computed by backpropagating the MSE loss between the predicted colour and the ground truth colour value. We use the RAdam optimizer [33] and decay the learning rate exponentially from $10^{-3}$ to $10^{-4}$ in 300k steps. The train- + +
PSNR ↑SSIM ↑LPIPS ↓
NeRF [39]31.000.9470.081
NSVF [32]31.770.953-
mip-NeRF [3]34.510.9610.043
instant-NGP [41]33.18--
Plenoxels [17]31.710.9580.049
Point-NeRFcol [70]31.770.9730.062
Point-NeRFmvs [70]33.310.9780.049
Tetra-NeRF32.530.9820.041
+ +Table 1. Results on the Blender dataset [39] averaged over all scenes in the dataset. Even though we use the same input point cloud as Point-NeRF $^{col}$ , we outperform it greatly. We perform on par with Point-NeRF $^{mvs}$ even though it uses many more points and densifies the point cloud during training. We highlight the best, second, and third values. + +ing code is built on top of the Nerfstudio framework [60], and the tetrahedra field is implemented in CUDA and uses the OptiX library [43]. We train on a single NVIDIA A100 GPU and the training speed ranges from 15k rays per second to 30k rays per second. The speed depends on how well-structured the triangulation is, how many vertices there are, and if there is empty space around the object. The full training with 300k iterations takes between 11 and 24 hours, depending on the scene complexity. However, good results are typically obtained much earlier, e.g., in 100k iterations. + +# 4.2. Results + +Blender dataset [39] results. We compare to relevant baselines on the standard Blender dataset. We used the same split and evaluation procedure as in the original NeRF paper [39] and the same SSIM implementation as Point-NeRF [70]. In order to ensure a fair comparison with Point-NeRF [70] when COLMAP points were used, we use the exact same COLMAP reconstruction as Point-NeRF. We report the PSNR, SSIM, and LPIPS (VGG) [78] metrics. Tab. 1 shows averaged results, Fig. 4 shows qualitative results, and the results for individual scenes are given in Supp. Mat. + +When we use the exact same COLMAP points as Point-NeRF (row $\text{Point-NeRF}^{col}$ ), we outperform it in all three metrics. We score comparably with $\text{Point-NeRF}^{mvs}$ , even though it starts from a much denser and higher-quality initial point cloud, which it generates from a jointly trained model. Also note that both of these Point-NeRF configurations grow the point cloud during training and, therefore, the complexity of the scene representation grows. For us, the points are fixed, and the number of parameters stays the same. We also outperform Plenoxels [17], which uses a sparse grid. Note that same as Point-NeRF, Plenoxels also gradually increases the representation complexity by subdividing the grid resolution at predefined training epochs. Even though both Mip-NeRF and instant-NGP outperform + +![](images/6bfebca5a0ce9a87375b85378673e4954bfba8d593be0a01e5e7e69b39c99b10.jpg) +Figure 4. Qualitative results on the Blender dataset. We compare with Point-NeRF $^{col}$ as we use the same input point cloud. In the top-row picture, we can see that Tetra-NeRF is able to represent fine details well on the mic scene. On the drums scene ( $2^{\text{nd}}$ row), both methods struggle with shiny materials, but our method performs slightly better. $3^{\text{rd}}$ row: Tetra-NeRF can render thin ropes. Bottom row: accumulation is nonzero in areas close to surface. + +
hotdogship
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
Point-NeRFstatic29.910.97819.350.905
Tetra-NeRF33.310.98931.130.994
+ +our approach in terms of PSNR, our method is slightly better in terms of SSIM, and on par with Mip-NeRF in terms of LPIPS. + +To analyze the tetrahedra field representation, we compare it with the point cloud field representation used in Point-NeRF. In Table 2, we compare our method to Point-NeRF when we disable point cloud growth and pruning. We show the results on two scenes from the Blender dataset [39] which were selected in the Point-NeRF paper. With this setup, we vastly outperform Point-NeRF in all metrics. The reason is that Point-NeRF requires the point cloud to be dense such that all rays have a chance to intersect a neighbourhood of a point. Since we use a continuous representation (tetrahedra field) rather than a discrete one, we can achieve good results even for sparser point clouds. + +![](images/a0f6006fce6cb015a0c296a9650831deb6ac66ff2fcd5039259d67e2e6fa1efa.jpg) +Figure 5. Comparison with dense grid field representation. Tetra-NeRF on the left, dense grid on the right. We used a comparable number of parameters for both methods (the dense grid having slightly more parameters). Due to the adaptive nature of tetrahedra fields, Tetra-NeRF produces significantly better rendering as it is able to focus on relevant parts of the scene. + +![](images/1eb0eafcf005a95649a4bacc914c609ccbd5dbdc50537d3010005cc416078e94.jpg) + +Table 2. Comparison with Point-NeRF with disabled point cloud growing and pruning shows that Point-NeRF performs significantly worse in all measured metrics because it struggles to handle the sparse point clouds. + +
PSNR ↑SSIM ↑LPIPS ↓
NV [34]23.700.8480.260
NeRF [39]25.780.8640.198
NSVF [32]28.400.9000.153
Point-NeRF [70]*28.350.9420.090
Tetra-NeRF28.900.9570.059
+ +Table 3. Result on the Tanks and Temples dataset [25] as processed by NSVF [32] averaged over all scenes. We show the PSNR, SSIM, and LPIPS (Alex) [78] metrics. We highlight the best, second, and third values. *Note, that Point-NeRF [70] results differ from the paper as they were recomputed with the resolution used in other methods. + +Comparison with the dense grid representation. In order to show the utility of the adaptive tetrahedra field representation, we compare it to a dense grid representation. Similarly to NSVF [32], we split the 3D scene bounding box into equally-sized voxels. When querying the field, we find the voxel to which the query point belongs and perform trilinear interpolation of the eight corners of the voxel. We choose the grid resolution such that the number of grid points is the lowest cube number larger than the number of points of the original point cloud. This ensures a fair comparison as the baseline uses a comparable (but larger) number of features than our approach. All other hyperparameters are kept the same as for Tetra-NeRF. We evaluate both approaches on two scenes from the Blender dataset [39]. The dense grid representation only scores PSNR 18.81 and 18.91 on the lego and ship scenes, respectively, whereas Tetra-NeRF scores PSNR 33.79 and 30.69, respectively. The results can be seen in Figure 5. Note that in this experiment, we only trained for 100,000 iterations to save computation time. From the numbers and the figure, we can clearly see that the dense grid resolution is not sufficient to reconstruct the scene in enough detail. Because Tetra-NeRF uses an adaptive subdivision, which is more detailed around the surface, we are better able to focus on relevant scene parts. With a similar number of trainable parameters, we thus obtain better results. + +Tanks and Temples [25] dataset. To be able to compare + +![](images/de87efc8bcf436104c7f063ae6d9ba29f1a8b57a0da7a3ea0fa8834ca562da6e.jpg) +Figure 6. Results on Tanks and Temples dataset. In the top row, we can see that we are able to represent the rim of the wheel better than Point-NeRF. Similarly, we can represent the tyre better in the third row. Finally, in the bottom, Point-NeRF fails on the wall and the roof, whereas Tetra-NeRF can render these parts well. + +with Point-NeRF [70] on real-world data, we have evaluated Tetra-NeRF on the Tanks and Temples [25] dataset. We use the same setup as in NSVF [32], where the object is masked. We report the usual metrics: PSNR, SSIM, and LPIPS (Alex).3 We used the dense COLMAP reconstruction to get the point cloud used in the tetrahedra field. Quantitative results are shown in Table 3, qualitative results are presented in Figure 6, and the per-scene results are given in the Supp. Mat. Note that the results originally reported in the Point-NeRF paper [70] were evaluated with a resolution different from the compared methods. Therefore, to ensure a fair comparison, we have recomputed the metrics in the publicly available Point-NeRF's predictions with the full resolution of $1920 \times 1080$ as used in NSVF [32]. In the public dataset, one of the scenes had corrupted camera parameters and we had to reconstruct the poses again for the Ignatius scene. With the reconstructed poses, our method performs slightly worse on that scene compared to others. + +The results show that our method outperforms the baselines in all compared metrics. This indicates that even though Point-NeRF relies on the ability to grow the point cloud density during training, this is not needed when using a continuous instead of a discrete representation. + +# Sparse and dense point cloud comparison. Previ + +![](images/78cf97762c2780e26f58d53f1195666cb06cf01dae15f4bfacb7457fd8eef17f.jpg) +(a) sparse point cloud + +![](images/e000c812427b2302432dbe438fc19fdeb2fe078553e8afabaa66464968020dd6.jpg) +(b) ground-truth image + +![](images/f48f77d9e1fdf500161ab6215fe29131aed9c880763c6e16f8321274b73c075d.jpg) +(c) sparse model +Figure 7. Sparse vs. dense model comparison detail. Even though the area in the bottom contains few points, we are still able to reconstruct at least low-frequency data. + +![](images/959c1c2a59ebfcd4ef5ce2ac8f5fe7f84af82e40cf39d11ff1abd733d0a4372c.jpg) +(d) dense model + +ous experiments have used dense COLMAP reconstructions. However, dense point clouds may not be required to achieve good reconstructions. In this section, we compare models trained on dense reconstructions and sparse reconstructions. Table 5 compares PSNR, and SSIM metrics on two synthetic scenes from the Blender dataset [39] and two real-world scenes from the Mip-NeRF 360 dataset [4]. We also show the number of vertices of the tetrahedra, which is the size of the input point cloud, including the random points added at the beginning. Note that in this experiment, we only trained for 100,000 iterations to save computation time. + +From the results, we can see that for the Blender dataset [39] case, the dense model is much better. This is to be expected since the number of points obtained from the sparse reconstruction on the Blender dataset is very small and our MLP does not have enough capacity to represent fine details. However, even in regions with zero point coverage, we are still able to provide at least some low-frequency data (c.f. Fig. 7). On real-world scenes, the sparse point cloud model is almost on par with the dense one. The sparse reconstructions on the real-world dataset provide many more points compared to the synthetic dataset case, and the coverage is sufficient. From these results, we conclude that on real-world data the sparse model can be sufficient to achieve good performance. + +Mip-NeRF 360 [4] dataset. We have further evaluated our method on the Mip-NeRF 360 [4] dataset. In order to ensure a fair comparison with Mip-NeRF 360 [4], we trained and evaluated on four times downsized images for the outdoor scenes and two times downsized images for the indoor scenes. The quantitative results are presented in Table 4, and the qualitative results are shown in Figure 8. + +From the results, we can see that Tetra-NeRF is able to outperform both the vanilla NeRF [39] and Mip-NeRF [3]. We also outperform Stable View Synthesis [51] in terms of PSNR and score comparably in terms of LPIPS. This is pos + +
OutdoorIndoorMean
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [15, 39]21.460.4580.51526.840.7900.37023.850.6050.451
mip-NeRF [3]21.690.4710.50526.980.7980.36124.040.6160.441
NeRF++ [77]22.760.5480.42728.050.8360.30925.110.6760.375
Deep Blending [20]21.540.5240.36426.390.8440.26123.700.6660.318
Point-Based Neural Rendering [26]21.660.6120.30226.280.8870.19123.710.7340.253
Stable View Synthesis [51]23.010.6620.25328.220.9070.16025.330.7710.211
mip-NeRF 360 [4]23.720.6870.28229.430.9110.18126.260.7860.237
Tetra-NeRF23.170.5860.29830.210.8810.10326.300.7170.211
+ +Table 4. Mip-NeRF 360 dataset results. We show the PSNR, SSIM, and LPIPS (Alex) [78] on two categories of Mip-NeRF 360 [4] scenes: outdoor, and indoor. On the outdoor scenes, we outperform the Stable View Synthesis [51] and are comparable to Mip-NeRF 360 [4], even though our method does not implement the improvements suggested in NeRF++ [77] and Mip-NeRF 360 [4], and is more comparable to vanilla NeRF [39]. We highlight the best, second, and third values. + +
sparsedense
PSNR/SSIM#pointsPSNR/SSIM#points
Blender/lego29.77/0.95925,78433.79/0.985302,781
Blender/ship26.90/0.8997,15230.69/0.942321,861
360/bonsai28.12/0.902413,22628.34/0.9021,000,000
360/garden24.79/0.806227,53225.41/0.8381,000,000
+ +Table 5. Dense and sparse point cloud comparison. We present results on two scenes from the Blender [39] and Mip-NeRF 360 [4] datasets. We also show the number of vertices of the tetrahedra field – including the randomly sampled points. + +![](images/056b494f4cb937613adcb130f764578d6efc865838862e0ddf1f93f8eb5543f9.jpg) +Figure 8. Mip-NeRF 360 results. $1^{\text{st}}$ row: we show the results on the garden scene. $2^{\text{nd}}$ : we can see that Tetra-NeRF has some problems rendering grass. $3^{\text{rd}}$ : Tetra-NeRF is able to represent the delicate texture of the tablecloth well. $4^{\text{th}}$ : Tetra-NeRF is able to recover fine geometry. + +sible because outdoor scenes contain a lot of high-resolution geometries, such as leaves and Stable View Synthesis is not able to overcome noise in the approximate geometry. Tetra- + +NeRF does not suffer from these problems as we aggregate features in 3D along rays rather than on the surface. Mip-NeRF 360 scores comparably to Tetra-NeRF in terms of PSNR and LPIPS, but it outperforms Tetra-NeRF in terms of SSIM. However, Mip-NeRF 360 implements some tricks designed to boost its performance. On the other hand, our approach is based on vanilla NeRF, and we believe that the performance can be boosted similarly. + +Varying the number of input points We conducted a study on the effect of the input point cloud size on the reconstruction quality. We consider both the sparse and the dense point clouds and analyse the performance as we subsample the points or add more points randomly. We evaluate on the family and the truck scenes from the Tanks and Temples [25] (tt) dataset, and the room and the garden scenes from the Mip-NeRF 360 [4] dataset. The sizes of sparse point clouds were roughly 16K, 30K, 113K, and 139K respectively and the sizes of dense point clouds were larger than 5M for all scenes. When the size of the original point cloud was larger than the desired size, we selected points uniformly at random, and when smaller, we added new random points as described Section 4.1. We include detailed results in the Supp. Mat. In order to save computational resources, we only train the method for 100k iterations. The results are visualised in Figure 9a. + +As expected, the performance improves with the number of points used, as it leads to a finer subdivision of the scene around the surface. We can also observe (e.g., from tt/family) that the sparse reconstruction leads to a better reconstruction quality at the same point cloud size. The likely explanation is that sampling uniformly at random will be biased towards regions with high density and some regions may be missed or covered extremely sparsely. + +Speed of convergence Figures 9b, 9c show the rate of convergence on the ship scene from the Blender dataset and aggregated results on the Tanks and Temples [25], and Mip-NeRF 360 (indoor/outdoor) [4] datasets. Similar to Point-NeRF, good results can be achieved quite early in the train + +![](images/9861e61b062a099c2b7fec2c1f74dc9daa6a5bff69fdcfdf20a10c42835c1acf.jpg) +Figure 9. a) shows the PSNR with different sizes of input point clouds. The solid and dashed lines represent the dense and coarse COLMAP reconstructions. The following scenes were evaluated (best to worst): tt/family, 360/room, tt/truck, and 360/garden from Tanks and Temples [25] (tt) and Mip-NeRF 360 [4] (360-indoor/360-outdoor) datasets. As expected, the quality increases with the number of points. b) visualizes the training speed (PSNR at different training iterations) on the blender/ship scene [39]. It shows Tetra-NeRF training times for various timesteps and compares the rate of convergence with Point-NeRF [70] and Mip-NeRF [3]. Similarly, c) shows the rate of convergence aggregated over Tanks and Temples [25] (tt) and Mip-NeRF 360 [4] (360-indoor/360-outdoor) datasets. It shows the average PSNR, and the highlighted area represents the standard deviation. + +![](images/a8fcbdbe0348c4c19d2d907b1d33dd173ab1363e6056d909064cc7f302ad70d9.jpg) + +![](images/0cc67eab62bbbc84549638a939872ce4ef7f3874a9f6b44330922790930cc799.jpg) + +![](images/af0ad92993a5cb97e75e153e92119f9c5f9af71eb9f578ec6a1c00da16e49b38.jpg) +(a) Failure case 1: low point cloud density on the ground + +![](images/31f34bbd148a667e04de4f51d7754f63b3ea14fc42569a08ca1b9bbf99ff7342.jpg) +(b) Failure case 2: too many intersected tetrahedra +Figure 10. Failure cases. Left: barn scene from [25]. Right: tree hill scene from [4]. + +ing process. E.g., on the blender/ship scene, in 24 minutes, our method achieves performance similar to Mip-NeRF [3] after several hours of training. The efficiency of the implementation is a large factor in the speed of different methods. E.g., in Instant-NGP [41], the authors invested a significant effort in optimizing GPU memory accesses, and the result is highly optimised efficient code (a main contribution of [41]). There are several places where our implementation can easily be made more efficient. E.g., changing the memory layout (row-major vs. column-major) of the tetrahedra field leads to a $10\%$ speedup. + +# 5. Limitations + +One drawback of our approach is that the quality of different regions of the rendered scene depends on the density of the point cloud in these regions. If the original point cloud was constructed using 2D feature matching, there can be regions with a very low number of 3D points. We show such an example in Fig. 10a. There are few 3D points on the ground, resulting in blurry renderings for that region. + +Another issue is that the current implementation has a limit on the number of intersected tetrahedra per ray. For large or badly structured scenes, where rays intersect many tetrahedra, this can limit the reconstruction quality. Fig. 10b + +shows such a case on a Mip-NeRF 360 scene [4]. This problem can be addressed by increasing the limit on the number of visited tetrahedra at the cost of higher memory requirements and run-time. Alternatively, coarse-to-fine schemes that start with a coarser tetrahedralisation and prune the space could potentially handle this issue. + +# 6. Conclusion + +This paper proposes a novel radiance field representation that, compared to standard voxel-based representations, is easily able to adapt to 3D geometry priors given in the form of a (sparse) point cloud. Our approach elegantly combines concepts from geometry processing (Delaunay triangulation) and triangle-based rendering (ray-triangle intersections) with modern neural rendering approaches. The representation has a naturally higher resolution in the space near surfaces, and the input point cloud provides a straightforward way to initialise the radiance field. Compared to Point-NeRF, a state-of-the-art point cloud-based radiance field representation which uses the same input as our approach, Tetra-NeRF shows clearly better results. Our method performs comparably to state-of-the-art MLP-based methods. The results demonstrate that Tetra-NeRF is an interesting alternative to existing radiance field representations that is worth further investigation. Interesting research directions include adaptive refinement and pruning of the tetrahedralisation and exploiting the fact that the surface of the scene is likely close to some of the triangles in the scene. + +Acknowledgements This work was supported by the Czech Science Foundation (GACR) EXPRO (grant no. 23-07973X), the Grant Agency of the Czech Technical University in Prague (grant no. SGS22/112/OHK3/2T/13), and by the Ministry of Education, Youth and Sports of the Czech Republic through the e-INFRA CZ (ID:90254). + +# References + +[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In ECCV, 2020. 2 +[2] Pierre Alliez, Simon Giraudot, Clément Jamin, Florent Lafarge, Quentin Merigot, Jocelyn Meyron, Laurent Saboret, Nader Salman, Shihao Wu, and Necip Fazil Yildiran. Point set processing. In CGAL User and Reference Manual. CGAL Editorial Board, 5.5.1 edition, 2022. 5 +[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. In ICCV, 2021. 1, 2, 4, 5, 7, 8, 9 +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1, 2, 5, 7, 8, 9 +[5] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv:2008.03824, 2020. 2 +[6] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. NoPe-NeRF: Optimising neural radiance field with no pose prior. arXiv:2212.07388, 2022.2 +[7] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. NeRD: Neural reflectance decomposition from image collections. In ICCV, 2021. 2 +[8] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001. 2 +[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial radiance fields. In ECCV, 2022. 1, 2, 4, 5 +[10] Anpei Chen, Zexiang Xu, Xinyue Wei, Siyu Tang, Hao Su, and Andreas Geiger. Factor fields: A unified framework for neural fields and beyond. arXiv:2302.01226, 2023. 1, 2, 3 +[11] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 2 +[12] Shuo Cheng, Zexiang Xu, Shilin Zhu, Zhuwen Li, Li Erran Li, Ravi Ramamoorthi, and Hao Su. Deep stereo using adaptive thin volume representation with uncertainty awareness. In CVPR, 2020. 2 +[13] Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2 +[14] B Delaunay, S Vide, A Lamémoire, and V De Georges. Bulletin de l'académie des sciences de l'URSS. Classe des sciences mathématiques et naturelles, 6:793-800, 1934. 2, 3, 4, 5 + +[15] Boyang Deng, Jonathan T. Barron, and Pratul P. Srinivasan. JaxNeRF: an efficient JAX implementation of NeRF, 2020. 8 +[16] Michael S Floater. Generalized barycentric coordinates and applications. Acta Numerica, 24:161-214, 2015. 3, 4 +[17] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2, 4, 5 +[18] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32:1362-1376, 2009. 2 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5 +[20] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. SIGGRAPH, ToG, 37(6):1-15, 2018. 2, 8 +[21] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In ICCV, 2021. 2 +[22] Vu Hoang Hiep, Renaud Keriven, Patrick Labatut, and Jean-Philippe Pons. Towards high-resolution large-scale multiview stereo. In CVPR, 2009. 2 +[23] Michal Jancosek and Tomas Pajdla. Multi-view reconstruction preserving weakly-supported surfaces. In CVPR, 2011. 2 +[24] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In SGP, 2006. 2 +[25] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. SIGGRAPH, ToG, 2017. 5, 6, 7, 8, 9 +[26] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, volume 40, pages 29-43. Wiley Online Library, 2021. 2, 8 +[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 2012. 7 +[28] Jonás Kulhánek, Erik Derner, Torsten Sattler, and Robert Babuška. ViewFormer: Nerf-free neural rendering from few images using transformers. In ECCV, 2022. 2 +[29] Patrick Labatut, Jean-Philippe Pons, and Renaud Keriven. Efficient multi-view reconstruction of large-scale scenes using interest points, Delaunay triangulation and graph cuts. In ICCV, 2007. 2 +[30] Patrick Labatut, J-P Pons, and Renaud Keriven. Robust and efficient surface reconstruction from range data. In Computer graphics forum, volume 28, pages 2275-2290. Wiley Online Library, 2009. 2 +[31] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. BARF: Bundle-adjusting neural radiance fields. In ICCV, 2021. 2 + +[32] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 1, 2, 3, 5, 6, 7 +[33] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond. In International Conference on Learning Representations, 2020. 5 +[34] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: learning dynamic renderable volumes from images. SIGGRAPH, ToG, 38(4):1-14, 2019. 2, 6 +[35] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3D surface construction algorithm. SIGGRAPH, ToG, 1987. 2 +[36] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 2 +[37] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 1, 3 +[38] Moustafa Meshry, Dan B. Goldman, Sameh Khamis, Hugues Hoppe, Rohit Pandey, Noah Snavely, and Ricardo MartinBrualla. Neural rerendering in the wild. In $CVPR$ , June 2019. 2 +[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 4, 5, 6, 7, 8, 9 +[40] Tomas Möller. A fast triangle-triangle intersection test. Journal of Graphics Tools, 2(2):25-30, 1997. 2 +[41] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. SIGGRAPH, ToG, 2022. 1, 2, 3, 4, 5, 9 +[42] M. Nießner, M. Zollhöfer, S. Izadi, and M. Stamminger. Real-time 3D reconstruction at scale using voxel hashing. SIGGRAPH, ToG, 2013. 4 +[43] Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith Morley, Austin Robison, et al. OptiX: a general purpose ray tracing engine. ACM Transactions on Graphics (ToG), 29(4):1-13, 2010. 2, 4, 5 +[44] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, 2020. 2 +[45] Bui Tuong Phong. Illumination for computer generated pictures. Communications of the ACM, 18(6):311-317, 1975. 2 +[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. arXiv:2209.14988, 2022. 2 +[47] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding up neural radiance fields with thousands of tiny mlp's. In ICCV, 2021. 2 + +[48] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T Barron, and Peter Hedman. MERF: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. arXiv:2302.12249, 2023. 2 +[49] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In ICCV, 2021. 2 +[50] Gernot Riegler and Vladlen Koltun. Free view synthesis. In ECCV, 2020. 2 +[51] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In CVPR, 2021. 2, 7, 8 +[52] Radu Alexandru Rosu and Sven Behnke. Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In CVPR, 2023. 2 +[53] Darius Rückert, Linus Franke, and Marc Stamminger. ADOP: Approximate differentiable one-pixel point rendering. SIGGRAPH, ToG, 41(4):1-14, 2022. 2 +[54] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 2, 4, 5 +[55] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 2, 4, 5 +[56] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv:1409.1556, 2014. 7 +[57] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 1, 2, 4 +[58] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-NeRF: Scalable large scene neural view synthesis. In CVPR, 2022. 2 +[59] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 4 +[60] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, et al. Nerfstudio: A modular framework for neural radiance field development. arXiv:2302.04264, 2023. 2, 5 +[61] Chengzhou Tang and Ping Tan. BA-net: Dense bundle adjustment networks. In International Conference on Learning Representations, 2019. 2 +[62] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 2 +[63] Sudheendra Vijayanarasimhan, Susanna Ricco, Cordelia Schmid, Rahul Sukthankar, and Katerina Fragkiadaki. SfM-Net: Learning of structure and motion from video. arXiv:1704.07804, 2017. 2 + +[64] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-NeRF: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2022. 2 +[65] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34:27171-27183, 2021. 2 +[66] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snively, and Thomas Funkhouser. IBRNet: Learning multi-view image-based rendering. In CVPR, 2021. 2 +[67] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. NeuS2: Fast learning of neural implicit surfaces for multi-view reconstruction. arXiv:2212.05231, 2022. 2 +[68] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF-: Neural radiance fields without known camera parameters. arXiv:2102.07064, 2021. 2 +[69] Daniel N Wood, Daniel I Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H Salesin, and Werner Stuetzle. Surface light fields for 3D photography. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, 2000. 2 +[70] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-NeRF: Point-based neural radiance fields. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 9 +[71] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multiview stereo. In ECCV, 2018. 2 +[72] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2 +[73] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. BakedSDF: Meshing neural SDFs for real-time view synthesis. arXiv:2302.14859, 2023. 2 +[74] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 1, 2 +[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2 +[76] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. In Advances in Neural Information Processing Systems, 2022. 2 +[77] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. NeRF++: Analyzing and improving neural radiance fields. arXiv:2010.07492, 2020. 2, 8 +[78] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, June 2018. 5, 6, 8 + +[79] Chaoqiang Zhao, Qiyu Sun, Chongzhen Zhang, Yang Tang, and Feng Qian. Monocular depth estimation based on deep learning: An overview. Science China Technological Sciences, 63(9):1612-1627, 2020. 3 \ No newline at end of file diff --git a/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/images.zip b/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..13bb2e5a69a028f154a9f9cd7aa161f0e0042d1d --- /dev/null +++ b/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9a74d36b97e4e4873734464396de22af25e31ca899a5227d5acd700321fca8e +size 734289 diff --git a/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/layout.json b/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ec5c168b3048f83a58dabbaae73962be542a2878 --- /dev/null +++ b/2023/Tetra-NeRF_ Representing Neural Radiance Fields Using Tetrahedra/layout.json @@ -0,0 +1,9251 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 509, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 509, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 509, + 121 + ], + "type": "text", + "content": "Tetra-NeRF: Representing Neural Radiance Fields Using Tetrahedra" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 144, + 282, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 144, + 282, + 185 + ], + "spans": [ + { + "bbox": [ + 97, + 144, + 282, + 185 + ], + "type": "text", + "content": "Jonas Kulhanek \nCzech Technical University in Prague \njonas.kulhanek@cvut.cz" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 310, + 144, + 494, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 144, + 494, + 185 + ], + "spans": [ + { + "bbox": [ + 310, + 144, + 494, + 185 + ], + "type": "text", + "content": "Torsten Sattler \nCzech Technical University in Prague \ntorsten.sattler@cvut.cz" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 111, + 196, + 366, + 331 + ], + "blocks": [ + { + "bbox": [ + 111, + 196, + 366, + 331 + ], + "lines": [ + { + "bbox": [ + 111, + 196, + 366, + 331 + ], + "spans": [ + { + "bbox": [ + 111, + 196, + 366, + 331 + ], + "type": "image", + "image_path": "5d718ca298af7923b76dae45526cd7fc1c5c03f76746559ba4400da2025a3561.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 206, + 490, + 337 + ], + "lines": [ + { + "bbox": [ + 360, + 206, + 490, + 337 + ], + "spans": [ + { + "bbox": [ + 360, + 206, + 490, + 337 + ], + "type": "text", + "content": "3) barycetric interpolation of features 4) shallow MLP predicts density and colour" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 338, + 547, + 372 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 547, + 372 + ], + "type": "text", + "content": "Figure 1. The input to Tetra-NeRF is a point cloud which is triangulated to get a set of tetrahedra used to represent the radiance field. Rays are sampled, and the field is queried. Barycentric interpolation is used to interpolate tetrahedra vertices, and the resulting features are passed through a shallow MLP to get the density and colours for volumetric rendering." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 384, + 192, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 384, + 192, + 396 + ], + "spans": [ + { + "bbox": [ + 143, + 384, + 192, + 396 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 409, + 290, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 290, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 290, + 625 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRFs) are a very recent and very popular approach for the problems of novel view synthesis and 3D reconstruction. A popular scene representation used by NeRFs is to combine a uniform, voxel-based subdivision of the scene with an MLP. Based on the observation that a (sparse) point cloud of the scene is often available, this paper proposes to use an adaptive representation based on tetrahedra obtained by Delaunay triangulation instead of uniform subdivision or point-based representations. We show that such a representation enables efficient training and leads to state-of-the-art results. Our approach elegantly combines concepts from 3D geometry processing, triangle-based rendering, and modern neural radiance fields. Compared to voxel-based representations, ours provides more detail around parts of the scene likely to be close to the surface. Compared to point-based representations, our approach achieves better performance. The source code is publicly available at: https://jkulhanek.com/tetra-nerf." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 645, + 128, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 128, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 128, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Reconstructing 3D scenes from images and rendering photo-realistic novel views is a key problem in computer vision. Recently, NeRFs [3,4,39] became dominant in the field for their superior photo-realistic results. Originally," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 385, + 546, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 546, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 546, + 636 + ], + "type": "text", + "content": "NeRFs used MLPs to represent the 3D scene as an implicit function. Given a set of posed images, NeRF randomly samples a batch of pixels, casts rays from the pixels into the 3D space, queries the implicit function at randomly sampled distances along the rays, and aggregates the sampled values using volumetric rendering [37, 39]. While the visual results of such methods are of high quality, the problem is that querying large MLPs at millions of points is costly. Also, once the network is trained, it is difficult to make any changes to the represented radiance field as everything is baked into the MLPs parameters, and any change has a non-local effect. Since then, there have been a lot of proposed alternatives to the large MLP field representation [9, 17, 32, 41, 57, 70, 74]. These methods combine an MLP with a voxel feature grid [41, 57], or in some cases represent the radiance field directly as a tensor [9, 10, 17]. When querying these representations, first, the containing voxel is found, and the features stored at the eight corner points of the voxel are trilinearly interpolated. The result is either passed through a shallow MLP [9, 10, 41, 57] or is used directly as the density and colour [17, 32, 57]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Having a dense tensor represent the entire scene is very inefficient, as we only need to represent a small space around surfaces. Therefore, different methods propose different ways of tackling the issue. Instant-NGP [41], for example, uses a hash grid instead of a dense tensor, where it relies on optimisation to resolve the hash collisions. How-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18458" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": "ever, similarly to MLPs, any change to the stored hashmap influences the field in many places. A more common direction to addressing the issue is by directly using a sparse tensor representation [9, 17]. These methods start with a low-resolution grid and, at predefined steps, subsample the representation, increasing the resolution. These approaches tend to require a careful setting of hyperparameters, such as the scene bounding box and the subdivision steps, in order for the methods to work well." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 384 + ], + "type": "text", + "content": "Because many of these methods use traditional structure from motion (SfM) [54, 55] methods to generate the initial poses for the captured images, we can reuse the original reconstruction in the scene representation. Inspired by classical surface reconstruction methods [22, 23, 29, 30], we represent the scene as a dense triangulation of the input point cloud, where the scene is a set of non-overlapping tetrahedra whose union is the convex hull of the original point cloud [14]. When querying such a representation, we find to which tetrahedron the query point belongs and perform barycentric linear interpolation of the features stored in the vertices of the tetrahedron. This very simple representation can be thought of as the direct extension of the classical triangle-rendering pipelines used in graphics [40, 43, 45]. The representation avoids problems with the sparsity of the input point cloud as the tetrahedra fully cover the scene, resulting in a continuous rather than discrete representation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 384, + 289, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 289, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 289, + 551 + ], + "type": "text", + "content": "This paper makes the following contributions: (1) We propose a novel radiance field representation which is initialised from a sparse or dense point cloud. This representation is naturally denser in the proximity of surfaces and, therefore, provides a higher resolution in these regions. (2) The proposed representation is evaluated on multiple synthetic and real-world datasets and is compared with a state-of-the-art point-cloud-based representation - Point-NeRF [70]. The presented results show that our method clearly outperforms this baseline. We further demonstrate the effectiveness of our adaptive representation by comparing it with a voxel-based representation that uses the same number of trainable parameters. (3) We make the source code and model checkpoints publicly available." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 561, + 132, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 132, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 132, + 574 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 287, + 693 + ], + "type": "text", + "content": "Multi-view reconstruction. The problem of multi-view reconstruction has been studied extensively and tackled with a variety of structure from motion (SfM) [54, 61, 63], and multi-view stereo (MVS) [12, 18, 55, 71] methods. These methods usually output the scene represented as a point cloud [54, 55]. In most rendering approaches, the point cloud is converted into a mesh [24, 35], and novel views are rendered by reprojecting observed images into each novel viewpoint and blending them together using ei-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "ther heuristically-defined [8, 13, 69] or learned [20, 50, 51, 73] blending weights." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 99, + 547, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 99, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 304, + 99, + 547, + 206 + ], + "type": "text", + "content": "However, the process of getting the meshes is usually quite noisy, and the resulting meshes tend to have inaccurate geometry in regions with fine details or complex materials. Instead of using noisy meshes, point-based neural rendering methods [1, 26, 38, 53] perform splatting of neural features and use 2D convolutions to render them. In contrast to these methods, our approach operates and aggregates features directly in 3D and does not suffer from the noise in the point cloud or the reconstructed mesh." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 213, + 546, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 546, + 453 + ], + "type": "text", + "content": "Neural radiance fields. Recently, NeRFs [3,4,34,39,77] have gained a lot of attention thanks to their high-quality rendering performance. The original NeRF method [39] was extended to better handle aliasing artefacts in [3], to better represent unbounded scenes in [4,48,77], or to handle real-world captured images [36,58]. The training of the large MLPs used in these methods can be quite slow, and there has been a lot of effort on speeding up either the training [9,17,41] or the rendering [21,47,48,74] sometimes at the cost of larger storage requirements. Other approaches tackled different aspects of NeRFs like view-dependent artefacts [62], relighting [5,7], or proposed generative models [46,64]. Also, a popular research direction is making the models generalize across different scenes [11,28,49,66,75]. A large area of research is dedicated to the surface reconstruction and, instead of using the radiance fields, represents the scene implicitly by modelling the signed distance function (SDF) [52,65,67,72,76]. Unlike those approaches, we focus only on the radiance field representation and consider these methods orthogonal to ours." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 455, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 547, + 552 + ], + "type": "text", + "content": "Although there are some methods that train radiance fields while fine-tuning the poses [31, 60] or without known cameras [6, 68], most methods need camera poses for the reconstruction. SfM, e.g., COLMAP [54, 55], is typically used for estimating the poses, which also produces a (sparse) point cloud. Our approach makes use of this by-product of the pose recovery process instead of only using the poses themselves." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "Field representations. When a single MLP is used to represent the entire scene, everything is baked into a single set of parameters which cannot be easily modified, because any change to the parameters has a non-local effect, i.e., it changes the scene at multiple unrelated places. To overcome this problem, others have experimented with different representations of the radiance fields [9, 10, 17, 32, 41, 44, 57, 70]. A common practice is to represent the scene as a shallow MLP and an efficiently encoded voxel grid [9, 10, 32, 41, 44]. The encoded voxel grid can also be used represent the radiance field directly [17, 57]. The voxel grid can be encoded as a sparse tensor [17, 32, 57], a factorisation of the 4D tensor [9, 10], or a hashmap [41]. When" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 61, + 702, + 256, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 256, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 256, + 712 + ], + "type": "text", + "content": "1https://github.com/jkulhanek/tetra-nerf" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18459" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 72, + 167, + 139 + ], + "blocks": [ + { + "bbox": [ + 57, + 72, + 167, + 139 + ], + "lines": [ + { + "bbox": [ + 57, + 72, + 167, + 139 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 167, + 139 + ], + "type": "image", + "image_path": "7abd8e15c57f18ac8cd3c30e5301bc371167faebce01af3348e4d354d3156e33.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 140, + 146, + 150 + ], + "lines": [ + { + "bbox": [ + 78, + 140, + 146, + 150 + ], + "spans": [ + { + "bbox": [ + 78, + 140, + 146, + 150 + ], + "type": "text", + "content": "(a) dense point cloud" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 73, + 271, + 134 + ], + "blocks": [ + { + "bbox": [ + 175, + 73, + 271, + 134 + ], + "lines": [ + { + "bbox": [ + 175, + 73, + 271, + 134 + ], + "spans": [ + { + "bbox": [ + 175, + 73, + 271, + 134 + ], + "type": "image", + "image_path": "7863d3ad9175ea27038d1cfdc437f43a530d2fdd5cf19f167a5d9461cf3284e2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 152, + 167, + 220 + ], + "blocks": [ + { + "bbox": [ + 57, + 152, + 167, + 220 + ], + "lines": [ + { + "bbox": [ + 57, + 152, + 167, + 220 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 167, + 220 + ], + "type": "image", + "image_path": "5f74a44cddba3fcb15c66ff5be32b246eb46c7aa4a1dd90de9b54feda4e1b0a2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 221, + 147, + 230 + ], + "lines": [ + { + "bbox": [ + 76, + 221, + 147, + 230 + ], + "spans": [ + { + "bbox": [ + 76, + 221, + 147, + 230 + ], + "type": "text", + "content": "(c) sparse point cloud" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 152, + 278, + 221 + ], + "blocks": [ + { + "bbox": [ + 182, + 140, + 264, + 150 + ], + "lines": [ + { + "bbox": [ + 182, + 140, + 264, + 150 + ], + "spans": [ + { + "bbox": [ + 182, + 140, + 264, + 150 + ], + "type": "text", + "content": "(b) dense tetrahedra slice" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 152, + 278, + 221 + ], + "lines": [ + { + "bbox": [ + 168, + 152, + 278, + 221 + ], + "spans": [ + { + "bbox": [ + 168, + 152, + 278, + 221 + ], + "type": "image", + "image_path": "326c524dd6484d7b5f579d4b15d08a3f0aa3570b24e5bfc89431dc848ea7bda9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 221, + 265, + 230 + ], + "lines": [ + { + "bbox": [ + 181, + 221, + 265, + 230 + ], + "spans": [ + { + "bbox": [ + 181, + 221, + 265, + 230 + ], + "type": "text", + "content": "(d) sparse tetrahedra slice" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 231, + 287, + 274 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 274 + ], + "type": "text", + "content": "Figure 2. Input point cloud and a slice through the triangulated tetrahedra. Note that smaller tetrahedra are created closer to the surface of the scene, i.e., regions close to the surface are represented with a finer resolution." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 225, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 304, + 225, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 225, + 545, + 262 + ], + "type": "text", + "content": "Fig. 2). In the following, we explain how this adaptive subdivision of the scene can be used instead of voxels for volume rendering and neural rendering." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 284, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 403 + ], + "type": "text", + "content": "these structures are queried, trilinear interpolation is used to combine the feature vectors stored in the containing voxel corners. Unfortunately, the hashes [41] and hierarchical representations [10] have the same non-local effect problem, and the rest of the approaches rely on subsequent upsampling of the field and can be overly complicated. Also, feature vectors cannot be placed arbitrarily in the 3D space as they must lie on the grid with a fixed resolution. In contrast, our approach is much more flexible as it stores the feature vectors freely in 3D space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 404, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 548 + ], + "type": "text", + "content": "Finally, Point-NeRF [70] represents the scene as a point cloud of features which are queried using " + }, + { + "bbox": [ + 46, + 404, + 287, + 548 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 404, + 287, + 548 + ], + "type": "text", + "content": "-nearest neighbours search. However, when the point cloud contains sparse regions, the rays do not intersect any neighbourhoods of any points and the pixels stay empty without the ability to optimise. Therefore, Point-NeRF [70] relies on gradually adding more points during training and increasing the scene complexity. Since we use the triangulation of the point cloud rather than the discrete point cloud itself, our representation is continuous and does not suffer from empty regions. Therefore, we do not have to add any points during training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 560, + 103, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 103, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 103, + 573 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "A common strategy in the literature is to represent the scene explicitly through a voxel volume. In contrast to this uniform subdivision, we investigate using an adaptive subdivision of the scene. In many scenarios, an approximation of the scene geometry is either given, e.g., when using SfM to compute the input camera poses, or can be computed, e.g., via MVS or single-view depth predictions [79]. This allows us to compute an adaptive subdivision of the scene via Delaunay triangulation [14] of such a point cloud. This results in a set of non-overlapping tetrahedra, where smaller tetrahedra are created close to the surface of the scene (c.f." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 437, + 158 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 437, + 158 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 437, + 158 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 437, + 158 + ], + "type": "image", + "image_path": "7898f45a2777e4512c769d742cb35004c7be94bd2da66bdbad9dec6738eb60b0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 159, + 422, + 168 + ], + "lines": [ + { + "bbox": [ + 342, + 159, + 422, + 168 + ], + "spans": [ + { + "bbox": [ + 342, + 159, + 422, + 168 + ], + "type": "text", + "content": "(a) trilinear interpolation" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 171, + 545, + 215 + ], + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 215 + ], + "type": "text", + "content": "Figure 3. Trilinear and barycentric interpolation. Trilinear interpolation is a weighted combination of the eight voxel corners. Barycentric interpolation weights the four vertices of the tetrahedron vertices based on the barycentric coordinates [16]." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 440, + 72, + 524, + 158 + ], + "blocks": [ + { + "bbox": [ + 440, + 72, + 524, + 158 + ], + "lines": [ + { + "bbox": [ + 440, + 72, + 524, + 158 + ], + "spans": [ + { + "bbox": [ + 440, + 72, + 524, + 158 + ], + "type": "image", + "image_path": "e7480085e9cc46610d1b010b85b34b8598fda893d58e624a4e3789cb494ef03e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 158, + 533, + 168 + ], + "lines": [ + { + "bbox": [ + 441, + 158, + 533, + 168 + ], + "spans": [ + { + "bbox": [ + 441, + 158, + 533, + 168 + ], + "type": "text", + "content": "(b) barycentric interpolation" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 271, + 391, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 271, + 391, + 283 + ], + "spans": [ + { + "bbox": [ + 306, + 271, + 391, + 283 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRFs) [39] represent a scene through an implicit function " + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x}, \\mathbf{d})" + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "text", + "content": ", often modelled via a neural network, that returns a colour value and a volume density prediction for a given 3D point " + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "text", + "content": " in the scene observed from a viewing direction " + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "text", + "content": ". Volume rendering [37] is used to synthesise novel views: for each pixel in a virtual view, we project a ray from the camera plane into the scene and sample the radiance field to obtain the colour and density values " + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\sigma_i" + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "text", + "content": " at distances " + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "text", + "content": " along the ray, where " + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "inline_equation", + "content": "i = 1 \\dots N" + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "text", + "content": ". The individual samples are then combined to predict the colour " + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 304, + 290, + 546, + 422 + ], + "type": "text", + "content": " for the pixel:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 347, + 432, + 545, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 432, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 347, + 432, + 545, + 471 + ], + "type": "interline_equation", + "content": "\\mathbf {C} = \\sum_ {i = 1} ^ {N} T _ {i} \\left(1 - \\exp \\left(\\sigma_ {i} \\delta_ {i}\\right)\\right) \\mathbf {c} _ {i}, \\text {w h e r e} \\tag {1}", + "image_path": "55a4b87dcf398d6e57d6480487cf7f0a55c8d8ac048918f5bfbb739a1603004f.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 347, + 467, + 447, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 467, + 447, + 500 + ], + "spans": [ + { + "bbox": [ + 347, + 467, + 447, + 500 + ], + "type": "interline_equation", + "content": "T _ {i} = \\exp \\left(- \\sum_ {i = 1} ^ {N - 1} \\sigma_ {i} \\delta_ {i}\\right),", + "image_path": "d0e6d597ed446aa62efa055ea2cd14843ef09a364fdc61a64271facfc3b07d52.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 510, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 569 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 304, + 510, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\delta_{i} = t_{i} - t_{i - 1}" + }, + { + "bbox": [ + 304, + 510, + 545, + 569 + ], + "type": "text", + "content": " is the distance between adjacent samples. This process is fully differentiable. Thus, after computing the mean squared error (MSE) between the predicted and ground truth colours associated with each ray, we can backpropagate the gradients to the radiance field." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Voxel-based feature fields such as NSVF [32] represent the scene as a voxel grid and an MLP. Each grid point of the voxel grid is assigned a trainable vector. There are eight feature vectors associated with a single voxel, but the vectors are shared between neighbouring voxels. For each query point sampled along the ray, its corresponding voxel is found first. A feature for the point is then computed via trilinear interpolation of the features of the voxel grid points based on the position of the query point (c.f. Fig. 3, left). The resulting feature vector is passed through an MLP to predict the density and appearance vector. The appearance vector is combined with the ray direction and" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18460" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "passed through a second MLP in order to compute a view-dependent colour." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 102, + 149, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 149, + 113 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 149, + 113 + ], + "type": "text", + "content": "3.2. Tetrahedra fields" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 120, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 287, + 239 + ], + "type": "text", + "content": "Given a set of points in 3D space, we build the tetrahedra structure by triangulating the points. We apply the Delaunay triangulation [14] to obtain a set of non-overlapping tetrahedra whose union is the convex hull of the original points. Fig. 2 shows example tetrahedra obtained by triangulating dense and sparse COLMAP-reconstructed point clouds [54, 55]. Note that the resulting representation is adaptive as it uses a higher resolution (smaller tetrahedra) closer to the surface and larger tetrahedra to model regions farther away from the surface." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 239, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 287, + 441 + ], + "type": "text", + "content": "We associate all vertices of the tetrahedra with trainable vectors. As in the voxel grid case, vertices, and thus vectors, are shared between adjacent tetrahedra. The resulting tetrahedra field can be queried in the same way a voxel grid representation is queried: for each query point, we first find the tetrahedron containing the point. Instead of the trilinear interpolation used for voxel volumes, we use the barycentric interpolation [16] to compute a feature vector for the query point from the four feature vectors stored at the tetrahedron's vertices (c.f. Fig. 3). To this end, we compute the query point " + }, + { + "bbox": [ + 46, + 239, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 239, + 287, + 441 + ], + "type": "text", + "content": "'s barycentric coordinates " + }, + { + "bbox": [ + 46, + 239, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 239, + 287, + 441 + ], + "type": "text", + "content": ", which express the point's 3D coordinates as a unique weighted combination of the 3D coordinates of the tetrahedron's vertices. In particular, the weight for a vertex is the volume of the tetrahedron constructed from the query point and the face opposite to the vertex divided by the volume of the full tetrahedron:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 445, + 287, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 445, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 88, + 445, + 287, + 472 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\lambda} = \\left(\\frac {V _ {x 2 3 4}}{V _ {1 2 3 4}}, \\frac {V _ {1 x 3 4}}{V _ {1 2 3 4}}, \\frac {V _ {1 2 x 4}}{V _ {1 2 3 4}}, \\frac {V _ {1 2 3 x}}{V _ {1 2 3 4}}\\right), \\tag {2}", + "image_path": "42236560ea625345933187248086637a0ccda94771352ba82a510a68998cf70e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "inline_equation", + "content": "V_{1234}" + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "text", + "content": " is the volume of the full tetrahedron and " + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "inline_equation", + "content": "V_{x234}" + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "inline_equation", + "content": "V_{1x34}, \\ldots" + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "text", + "content": ", are volumes of tetrahedra with " + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "text", + "content": ", ..., vertex replaced by " + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "text", + "content": ". The same weights " + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 474, + 287, + 521 + ], + "type": "text", + "content": " are applied to the feature vectors of the vertices to obtain the query feature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "text", + "content": "The interpolated features are used as the input to a small MLP in order to predict density and colour at the query point. We first pass the barycentric-interpolated features through a three-layer MLP to compute the density and appearance features. We then concatenate the appearance features with the ray direction vector encoded using Fourier features [39, 59] and pass the result through a single linear layer to get the raw RGB colour value." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": "Finally, to map the raw density values " + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\bar{\\sigma}_i" + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": " returned by the network to the volume density " + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": " required by volume rendering, we apply the softplus activation function [3]. For the RGB colour value, we use the sigmoid function [17]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 671, + 247, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 247, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 247, + 684 + ], + "type": "text", + "content": "3.3. Efficiently querying a tetrahedra field" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Determining the corresponding voxel for a given query point can be done highly efficiently via hashing [42]. In" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "contrast, finding the corresponding tetrahedron for a query point is more complex. This in turn can significantly impact rendering, and thus training, efficiency [17,41,57]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 108, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 287 + ], + "type": "text", + "content": "In order to efficiently look up the corresponding tetrahedra, we exploit that we are not considering isolated points, but points sampled from a ray. We compute the tetrahedra that are intersected by the ray, allowing us to march through the tetrahedra rather than computing them individually per point. The relevant tetrahedra can be found efficiently using acceleration structures for fast ray-triangle intersection computations, e.g., via NVidia's OptiX library [43]: We first compute the intersections between the rays from a synthetic view and all faces of the tetrahedra. For each ray, we take the first 512 intersected triangles², and determine the corresponding tetrahedra. The tetrahedra can be ordered based on the intersections along the ray, allowing us to easily march through the tetrahedra to determine which tetrahedron to use for a given query point." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 288, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 407 + ], + "type": "text", + "content": "A side benefit of computing ray-triangle intersections is that we can simplify computing the barycentric coordinates: For each intersection, we compute the 2D barycentric coordinates w.r.t. the triangle. We then obtain the 3D barycentric coordinates w.r.t. the associate tetrahedron by simply adding zero for the vertex opposite to the triangle. For a query point inside a tetrahedron, we can compute its tetrahedron barycentric coordinates by linearly interpolating between the barycentric coordinates of the two intersections of the ray and the tetrahedron." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 413, + 447, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 413, + 447, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 413, + 447, + 426 + ], + "type": "text", + "content": "3.4. Coarse and fine sampling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 432, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 563 + ], + "type": "text", + "content": "We follow the common practice of having a two-stage sampling procedure [9, 39]. In the coarse stage, we sample uniformly along the ray. In the fine stage, we use the density weights from the coarse sampling stage to bias the sampling towards sampling closer to the potential surface. Following [39], we use the stratified uniform sampling for the coarse stage. The stratified uniform sampling splits the ray into equally long intervals and samples uniformly in each interval. Unlike NeRF [39], we limit the sampling to the space occupied by tetrahedra. In the fine sampling stage, we use the same network as in the coarse sampling stage." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 563, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 545, + 587 + ], + "type": "text", + "content": "For the fine sampling stage, we take the accumulated weights " + }, + { + "bbox": [ + 305, + 563, + 545, + 587 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 305, + 563, + 545, + 587 + ], + "type": "text", + "content": " from the coarse sampling:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 332, + 594, + 545, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 594, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 332, + 594, + 545, + 626 + ], + "type": "interline_equation", + "content": "\\bar {w} _ {i} = \\left(1 - \\exp \\left(- \\sigma_ {i} \\delta_ {i}\\right)\\right) \\exp \\left(- \\sum_ {j = 1} ^ {i - 1} \\sigma_ {j} \\delta_ {j}\\right). \\tag {3}", + "image_path": "a822a53c8d45a1aec77ceebdb5d782a8d163dcbddc1c5a7e0b928adcc2cfa9fd.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "text", + "content": "These weights are the coefficients used in Equation 1 as multipliers for the colours [39]. We obtain " + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "text", + "content": " by normalizing " + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "inline_equation", + "content": "\\bar{w}_{i}" + }, + { + "bbox": [ + 305, + 632, + 545, + 669 + ], + "type": "text", + "content": ". Following [39], we sample set of fine samples using" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 673, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 673, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 673, + 545, + 713 + ], + "type": "text", + "content": "2For efficiency, we only consider a fixed number of triangles. As discussed later on, this can degrade results in larger scenes / scenes with a fine-grained tetrahedralisation, where more intersections are needed. Naturally, more triangles can be considered at the cost of longer run-times." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18461" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": "weight " + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": ". To render the final colour, we merge the dense and fine samples and use all in the rendering equation [9]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 106, + 128, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 128, + 119 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 128, + 119 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 126, + 287, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 126, + 287, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 126, + 287, + 210 + ], + "type": "text", + "content": "We compare Tetra-NeRF to relevant methods on the commonly used synthetic Blender [39], the real-world Tanks and Temples [25], and the challenging object-centric Mip-NeRF 360 [4] datasets. To show its effectiveness, we compare Tetra-NeRF to a dense-grid representation and evaluate it with reduced quality of the input point cloud. We start by describing the exact hyperparameters used." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 217, + 178, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 178, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 178, + 229 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "text", + "content": "Generating point cloud & triangulation. Given a set of posed images, we use the COLMAP reconstruction pipeline [54, 55] to get the point cloud used in our tetrahedra field representation. We then reduce the size of the resulting point cloud such that if the number of points is larger than " + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "inline_equation", + "content": "10^{6}" + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "text", + "content": ", we subsample " + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "inline_equation", + "content": "10^{6}" + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "text", + "content": " points randomly. For the Blender dataset experiments, where the number of points is smaller, we add more randomly generated points. In that case, the number of random points is half the number of original points. The reason for adding the points is that with a low number of points, some pixels on edges may not intersect any tetrahedra, potentially producing artefacts on the edges. Each added point is sampled as follows: we sample a random point " + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "text", + "content": " from the original point cloud, sample a random normal vector " + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "text", + "content": ", and a number " + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "inline_equation", + "content": "\\alpha \\sim \\mathcal{N}(\\bar{d}, \\bar{d}^{2})" + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "inline_equation", + "content": "\\bar{d}" + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "text", + "content": " is the average spacing of the original point cloud, i.e. the average distance between each point and its six closest neighbours. We then add the point " + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "inline_equation", + "content": "x = x_{0} + \\alpha n" + }, + { + "bbox": [ + 46, + 235, + 288, + 450 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 451, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 628 + ], + "type": "text", + "content": "Initialization. Given the processed point cloud, we apply the Delaunay triangulation [14] to get a set of tetrahedra. To this end, we use the CGAL library [2], which in our experiments runs in the order of milliseconds. Following [41], we initialise the features of size 64 at the vertices of the tetrahedra with small values sampled uniformly in the range " + }, + { + "bbox": [ + 46, + 451, + 287, + 628 + ], + "type": "inline_equation", + "content": "-10^{-4}" + }, + { + "bbox": [ + 46, + 451, + 287, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 451, + 287, + 628 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 451, + 287, + 628 + ], + "type": "text", + "content": ". However, to allow the model to reuse the information contained in the point cloud, we set the first four dimensions of the feature field to the RGBA colour values (rescaled to interval [0, 1]) stored at the associated points in the point cloud. The alpha value of all original points is one, whereas all randomly sampled points have an alpha value of zero. For the MLP, we follow the common practice of using the Kaiming uniform initialisation [19]. The hidden sizes in all MLPs are 128." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "Training. During training, we sample batches of 4,096 random rays from random training images. We use volumetric rendering to predict the colour of each ray. The gradients are computed by backpropagating the MSE loss between the predicted colour and the ground truth colour value. We use the RAdam optimizer [33] and decay the learning rate exponentially from " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": " in 300k steps. The train-" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 321, + 70, + 531, + 173 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 531, + 173 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 531, + 173 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 531, + 173 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓
NeRF [39]31.000.9470.081
NSVF [32]31.770.953-
mip-NeRF [3]34.510.9610.043
instant-NGP [41]33.18--
Plenoxels [17]31.710.9580.049
Point-NeRFcol [70]31.770.9730.062
Point-NeRFmvs [70]33.310.9780.049
Tetra-NeRF32.530.9820.041
", + "image_path": "0b7aded02f932d813f11ed50c31f54d24f4c0743941531ff638f697c43e5e8af.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 178, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 547, + 248 + ], + "type": "text", + "content": "Table 1. Results on the Blender dataset [39] averaged over all scenes in the dataset. Even though we use the same input point cloud as Point-NeRF" + }, + { + "bbox": [ + 305, + 178, + 547, + 248 + ], + "type": "inline_equation", + "content": "^{col}" + }, + { + "bbox": [ + 305, + 178, + 547, + 248 + ], + "type": "text", + "content": ", we outperform it greatly. We perform on par with Point-NeRF" + }, + { + "bbox": [ + 305, + 178, + 547, + 248 + ], + "type": "inline_equation", + "content": "^{mvs}" + }, + { + "bbox": [ + 305, + 178, + 547, + 248 + ], + "type": "text", + "content": " even though it uses many more points and densifies the point cloud during training. We highlight the best, second, and third values." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 260, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 545, + 381 + ], + "type": "text", + "content": "ing code is built on top of the Nerfstudio framework [60], and the tetrahedra field is implemented in CUDA and uses the OptiX library [43]. We train on a single NVIDIA A100 GPU and the training speed ranges from 15k rays per second to 30k rays per second. The speed depends on how well-structured the triangulation is, how many vertices there are, and if there is empty space around the object. The full training with 300k iterations takes between 11 and 24 hours, depending on the scene complexity. However, good results are typically obtained much earlier, e.g., in 100k iterations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 390, + 362, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 390, + 362, + 402 + ], + "spans": [ + { + "bbox": [ + 306, + 390, + 362, + 402 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 544 + ], + "type": "text", + "content": "Blender dataset [39] results. We compare to relevant baselines on the standard Blender dataset. We used the same split and evaluation procedure as in the original NeRF paper [39] and the same SSIM implementation as Point-NeRF [70]. In order to ensure a fair comparison with Point-NeRF [70] when COLMAP points were used, we use the exact same COLMAP reconstruction as Point-NeRF. We report the PSNR, SSIM, and LPIPS (VGG) [78] metrics. Tab. 1 shows averaged results, Fig. 4 shows qualitative results, and the results for individual scenes are given in Supp. Mat." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "When we use the exact same COLMAP points as Point-NeRF (row " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\text{Point-NeRF}^{col}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "), we outperform it in all three metrics. We score comparably with " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\text{Point-NeRF}^{mvs}" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": ", even though it starts from a much denser and higher-quality initial point cloud, which it generates from a jointly trained model. Also note that both of these Point-NeRF configurations grow the point cloud during training and, therefore, the complexity of the scene representation grows. For us, the points are fixed, and the number of parameters stays the same. We also outperform Plenoxels [17], which uses a sparse grid. Note that same as Point-NeRF, Plenoxels also gradually increases the representation complexity by subdividing the grid resolution at predefined training epochs. Even though both Mip-NeRF and instant-NGP outperform" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18462" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 74, + 286, + 320 + ], + "blocks": [ + { + "bbox": [ + 51, + 74, + 286, + 320 + ], + "lines": [ + { + "bbox": [ + 51, + 74, + 286, + 320 + ], + "spans": [ + { + "bbox": [ + 51, + 74, + 286, + 320 + ], + "type": "image", + "image_path": "6bfebca5a0ce9a87375b85378673e4954bfba8d593be0a01e5e7e69b39c99b10.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 321, + 289, + 410 + ], + "lines": [ + { + "bbox": [ + 46, + 321, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 289, + 410 + ], + "type": "text", + "content": "Figure 4. Qualitative results on the Blender dataset. We compare with Point-NeRF" + }, + { + "bbox": [ + 46, + 321, + 289, + 410 + ], + "type": "inline_equation", + "content": "^{col}" + }, + { + "bbox": [ + 46, + 321, + 289, + 410 + ], + "type": "text", + "content": " as we use the same input point cloud. In the top-row picture, we can see that Tetra-NeRF is able to represent fine details well on the mic scene. On the drums scene (" + }, + { + "bbox": [ + 46, + 321, + 289, + 410 + ], + "type": "inline_equation", + "content": "2^{\\text{nd}}" + }, + { + "bbox": [ + 46, + 321, + 289, + 410 + ], + "type": "text", + "content": " row), both methods struggle with shiny materials, but our method performs slightly better. " + }, + { + "bbox": [ + 46, + 321, + 289, + 410 + ], + "type": "inline_equation", + "content": "3^{\\text{rd}}" + }, + { + "bbox": [ + 46, + 321, + 289, + 410 + ], + "type": "text", + "content": " row: Tetra-NeRF can render thin ropes. Bottom row: accumulation is nonzero in areas close to surface." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 418, + 275, + 466 + ], + "blocks": [ + { + "bbox": [ + 58, + 418, + 275, + 466 + ], + "lines": [ + { + "bbox": [ + 58, + 418, + 275, + 466 + ], + "spans": [ + { + "bbox": [ + 58, + 418, + 275, + 466 + ], + "type": "table", + "html": "
hotdogship
PSNR ↑SSIM ↑PSNR ↑SSIM ↑
Point-NeRFstatic29.910.97819.350.905
Tetra-NeRF33.310.98931.130.994
", + "image_path": "abc0b7ef8c4c9fdbb9facec634b10180697b99e9f215bf439c5c925821114fb7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 532, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 568 + ], + "type": "text", + "content": "our approach in terms of PSNR, our method is slightly better in terms of SSIM, and on par with Mip-NeRF in terms of LPIPS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "To analyze the tetrahedra field representation, we compare it with the point cloud field representation used in Point-NeRF. In Table 2, we compare our method to Point-NeRF when we disable point cloud growth and pruning. We show the results on two scenes from the Blender dataset [39] which were selected in the Point-NeRF paper. With this setup, we vastly outperform Point-NeRF in all metrics. The reason is that Point-NeRF requires the point cloud to be dense such that all rays have a chance to intersect a neighbourhood of a point. Since we use a continuous representation (tetrahedra field) rather than a discrete one, we can achieve good results even for sparser point clouds." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 425, + 153 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 425, + 153 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 425, + 153 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 425, + 153 + ], + "type": "image", + "image_path": "a0f6006fce6cb015a0c296a9650831deb6ac66ff2fcd5039259d67e2e6fa1efa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 156, + 547, + 221 + ], + "lines": [ + { + "bbox": [ + 305, + 156, + 547, + 221 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 547, + 221 + ], + "type": "text", + "content": "Figure 5. Comparison with dense grid field representation. Tetra-NeRF on the left, dense grid on the right. We used a comparable number of parameters for both methods (the dense grid having slightly more parameters). Due to the adaptive nature of tetrahedra fields, Tetra-NeRF produces significantly better rendering as it is able to focus on relevant parts of the scene." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 430, + 72, + 542, + 153 + ], + "blocks": [ + { + "bbox": [ + 430, + 72, + 542, + 153 + ], + "lines": [ + { + "bbox": [ + 430, + 72, + 542, + 153 + ], + "spans": [ + { + "bbox": [ + 430, + 72, + 542, + 153 + ], + "type": "image", + "image_path": "1eb0eafcf005a95649a4bacc914c609ccbd5dbdc50537d3010005cc416078e94.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 326, + 230, + 526, + 300 + ], + "blocks": [ + { + "bbox": [ + 46, + 473, + 288, + 518 + ], + "lines": [ + { + "bbox": [ + 46, + 473, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 288, + 518 + ], + "type": "text", + "content": "Table 2. Comparison with Point-NeRF with disabled point cloud growing and pruning shows that Point-NeRF performs significantly worse in all measured metrics because it struggles to handle the sparse point clouds." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 230, + 526, + 300 + ], + "lines": [ + { + "bbox": [ + 326, + 230, + 526, + 300 + ], + "spans": [ + { + "bbox": [ + 326, + 230, + 526, + 300 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓
NV [34]23.700.8480.260
NeRF [39]25.780.8640.198
NSVF [32]28.400.9000.153
Point-NeRF [70]*28.350.9420.090
Tetra-NeRF28.900.9570.059
", + "image_path": "fd041eee811d758b315593b07a68d139ffe3d18fc98a239dfd4b63827d022a56.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 307, + 547, + 374 + ], + "lines": [ + { + "bbox": [ + 305, + 307, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 547, + 374 + ], + "type": "text", + "content": "Table 3. Result on the Tanks and Temples dataset [25] as processed by NSVF [32] averaged over all scenes. We show the PSNR, SSIM, and LPIPS (Alex) [78] metrics. We highlight the best, second, and third values. *Note, that Point-NeRF [70] results differ from the paper as they were recomputed with the resolution used in other methods." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 386, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 547, + 696 + ], + "type": "text", + "content": "Comparison with the dense grid representation. In order to show the utility of the adaptive tetrahedra field representation, we compare it to a dense grid representation. Similarly to NSVF [32], we split the 3D scene bounding box into equally-sized voxels. When querying the field, we find the voxel to which the query point belongs and perform trilinear interpolation of the eight corners of the voxel. We choose the grid resolution such that the number of grid points is the lowest cube number larger than the number of points of the original point cloud. This ensures a fair comparison as the baseline uses a comparable (but larger) number of features than our approach. All other hyperparameters are kept the same as for Tetra-NeRF. We evaluate both approaches on two scenes from the Blender dataset [39]. The dense grid representation only scores PSNR 18.81 and 18.91 on the lego and ship scenes, respectively, whereas Tetra-NeRF scores PSNR 33.79 and 30.69, respectively. The results can be seen in Figure 5. Note that in this experiment, we only trained for 100,000 iterations to save computation time. From the numbers and the figure, we can clearly see that the dense grid resolution is not sufficient to reconstruct the scene in enough detail. Because Tetra-NeRF uses an adaptive subdivision, which is more detailed around the surface, we are better able to focus on relevant scene parts. With a similar number of trainable parameters, we thus obtain better results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Tanks and Temples [25] dataset. To be able to compare" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18463" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 285, + 328 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 285, + 328 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 285, + 328 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 285, + 328 + ], + "type": "image", + "image_path": "de87efc8bcf436104c7f063ae6d9ba29f1a8b57a0da7a3ea0fa8834ca562da6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 331, + 287, + 386 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 386 + ], + "type": "text", + "content": "Figure 6. Results on Tanks and Temples dataset. In the top row, we can see that we are able to represent the rim of the wheel better than Point-NeRF. Similarly, we can represent the tyre better in the third row. Finally, in the bottom, Point-NeRF fails on the wall and the roof, whereas Tetra-NeRF can render these parts well." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 393, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 608 + ], + "type": "text", + "content": "with Point-NeRF [70] on real-world data, we have evaluated Tetra-NeRF on the Tanks and Temples [25] dataset. We use the same setup as in NSVF [32], where the object is masked. We report the usual metrics: PSNR, SSIM, and LPIPS (Alex).3 We used the dense COLMAP reconstruction to get the point cloud used in the tetrahedra field. Quantitative results are shown in Table 3, qualitative results are presented in Figure 6, and the per-scene results are given in the Supp. Mat. Note that the results originally reported in the Point-NeRF paper [70] were evaluated with a resolution different from the compared methods. Therefore, to ensure a fair comparison, we have recomputed the metrics in the publicly available Point-NeRF's predictions with the full resolution of " + }, + { + "bbox": [ + 46, + 393, + 287, + 608 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 46, + 393, + 287, + 608 + ], + "type": "text", + "content": " as used in NSVF [32]. In the public dataset, one of the scenes had corrupted camera parameters and we had to reconstruct the poses again for the Ignatius scene. With the reconstructed poses, our method performs slightly worse on that scene compared to others." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 609, + 286, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 286, + 669 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 286, + 669 + ], + "type": "text", + "content": "The results show that our method outperforms the baselines in all compared metrics. This indicates that even though Point-NeRF relies on the ability to grow the point cloud density during training, this is not needed when using a continuous instead of a discrete representation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 673, + 286, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 286, + 685 + ], + "type": "text", + "content": "Sparse and dense point cloud comparison. Previ" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 425, + 137 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 425, + 137 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 425, + 137 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 425, + 137 + ], + "type": "image", + "image_path": "78cf97762c2780e26f58d53f1195666cb06cf01dae15f4bfacb7457fd8eef17f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 137, + 411, + 145 + ], + "lines": [ + { + "bbox": [ + 341, + 137, + 411, + 145 + ], + "spans": [ + { + "bbox": [ + 341, + 137, + 411, + 145 + ], + "type": "text", + "content": "(a) sparse point cloud" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 524, + 136 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 524, + 136 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 524, + 136 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 524, + 136 + ], + "type": "image", + "image_path": "e000c812427b2302432dbe438fc19fdeb2fe078553e8afabaa66464968020dd6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 439, + 137, + 513, + 145 + ], + "lines": [ + { + "bbox": [ + 439, + 137, + 513, + 145 + ], + "spans": [ + { + "bbox": [ + 439, + 137, + 513, + 145 + ], + "type": "text", + "content": "(b) ground-truth image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 148, + 425, + 211 + ], + "blocks": [ + { + "bbox": [ + 328, + 148, + 425, + 211 + ], + "lines": [ + { + "bbox": [ + 328, + 148, + 425, + 211 + ], + "spans": [ + { + "bbox": [ + 328, + 148, + 425, + 211 + ], + "type": "image", + "image_path": "f48f77d9e1fdf500161ab6215fe29131aed9c880763c6e16f8321274b73c075d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 212, + 403, + 220 + ], + "lines": [ + { + "bbox": [ + 349, + 212, + 403, + 220 + ], + "spans": [ + { + "bbox": [ + 349, + 212, + 403, + 220 + ], + "type": "text", + "content": "(c) sparse model" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 220, + 545, + 253 + ], + "lines": [ + { + "bbox": [ + 305, + 220, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 545, + 253 + ], + "type": "text", + "content": "Figure 7. Sparse vs. dense model comparison detail. Even though the area in the bottom contains few points, we are still able to reconstruct at least low-frequency data." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 148, + 524, + 211 + ], + "blocks": [ + { + "bbox": [ + 427, + 148, + 524, + 211 + ], + "lines": [ + { + "bbox": [ + 427, + 148, + 524, + 211 + ], + "spans": [ + { + "bbox": [ + 427, + 148, + 524, + 211 + ], + "type": "image", + "image_path": "959c1c2a59ebfcd4ef5ce2ac8f5fe7f84af82e40cf39d11ff1abd733d0a4372c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 212, + 503, + 220 + ], + "lines": [ + { + "bbox": [ + 449, + 212, + 503, + 220 + ], + "spans": [ + { + "bbox": [ + 449, + 212, + 503, + 220 + ], + "type": "text", + "content": "(d) dense model" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 265, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 407 + ], + "type": "text", + "content": "ous experiments have used dense COLMAP reconstructions. However, dense point clouds may not be required to achieve good reconstructions. In this section, we compare models trained on dense reconstructions and sparse reconstructions. Table 5 compares PSNR, and SSIM metrics on two synthetic scenes from the Blender dataset [39] and two real-world scenes from the Mip-NeRF 360 dataset [4]. We also show the number of vertices of the tetrahedra, which is the size of the input point cloud, including the random points added at the beginning. Note that in this experiment, we only trained for 100,000 iterations to save computation time." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 409, + 546, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 546, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 546, + 577 + ], + "type": "text", + "content": "From the results, we can see that for the Blender dataset [39] case, the dense model is much better. This is to be expected since the number of points obtained from the sparse reconstruction on the Blender dataset is very small and our MLP does not have enough capacity to represent fine details. However, even in regions with zero point coverage, we are still able to provide at least some low-frequency data (c.f. Fig. 7). On real-world scenes, the sparse point cloud model is almost on par with the dense one. The sparse reconstructions on the real-world dataset provide many more points compared to the synthetic dataset case, and the coverage is sufficient. From these results, we conclude that on real-world data the sparse model can be sufficient to achieve good performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "text", + "content": "Mip-NeRF 360 [4] dataset. We have further evaluated our method on the Mip-NeRF 360 [4] dataset. In order to ensure a fair comparison with Mip-NeRF 360 [4], we trained and evaluated on four times downsized images for the outdoor scenes and two times downsized images for the indoor scenes. The quantitative results are presented in Table 4, and the qualitative results are shown in Figure 8." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "From the results, we can see that Tetra-NeRF is able to outperform both the vanilla NeRF [39] and Mip-NeRF [3]. We also outperform Stable View Synthesis [51] in terms of PSNR and score comparably in terms of LPIPS. This is pos" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "text", + "content": "3We always choose the type of LPIPS (Alex [27]/VGG [56]) such that we can compare with more methods as some only evaluate using one type." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18464" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 69, + 542, + 182 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 542, + 182 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 542, + 182 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 542, + 182 + ], + "type": "table", + "html": "
OutdoorIndoorMean
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
NeRF [15, 39]21.460.4580.51526.840.7900.37023.850.6050.451
mip-NeRF [3]21.690.4710.50526.980.7980.36124.040.6160.441
NeRF++ [77]22.760.5480.42728.050.8360.30925.110.6760.375
Deep Blending [20]21.540.5240.36426.390.8440.26123.700.6660.318
Point-Based Neural Rendering [26]21.660.6120.30226.280.8870.19123.710.7340.253
Stable View Synthesis [51]23.010.6620.25328.220.9070.16025.330.7710.211
mip-NeRF 360 [4]23.720.6870.28229.430.9110.18126.260.7860.237
Tetra-NeRF23.170.5860.29830.210.8810.10326.300.7170.211
", + "image_path": "3bab8b28c2be22a4afc566a04b267096097993072367f7ec88f0a7c318a8a196.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 54, + 242, + 281, + 312 + ], + "blocks": [ + { + "bbox": [ + 46, + 190, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 237 + ], + "type": "text", + "content": "Table 4. Mip-NeRF 360 dataset results. We show the PSNR, SSIM, and LPIPS (Alex) [78] on two categories of Mip-NeRF 360 [4] scenes: outdoor, and indoor. On the outdoor scenes, we outperform the Stable View Synthesis [51] and are comparable to Mip-NeRF 360 [4], even though our method does not implement the improvements suggested in NeRF++ [77] and Mip-NeRF 360 [4], and is more comparable to vanilla NeRF [39]. We highlight the best, second, and third values." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 242, + 281, + 312 + ], + "lines": [ + { + "bbox": [ + 54, + 242, + 281, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 242, + 281, + 312 + ], + "type": "table", + "html": "
sparsedense
PSNR/SSIM#pointsPSNR/SSIM#points
Blender/lego29.77/0.95925,78433.79/0.985302,781
Blender/ship26.90/0.8997,15230.69/0.942321,861
360/bonsai28.12/0.902413,22628.34/0.9021,000,000
360/garden24.79/0.806227,53225.41/0.8381,000,000
", + "image_path": "5d0ce67c4cb03a1ab29de2471ad17c62c82b8cda1a763bbe79eac4c36303d945.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 323, + 288, + 368 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 368 + ], + "type": "text", + "content": "Table 5. Dense and sparse point cloud comparison. We present results on two scenes from the Blender [39] and Mip-NeRF 360 [4] datasets. We also show the number of vertices of the tetrahedra field – including the randomly sampled points." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 52, + 374, + 283, + 612 + ], + "blocks": [ + { + "bbox": [ + 52, + 374, + 283, + 612 + ], + "lines": [ + { + "bbox": [ + 52, + 374, + 283, + 612 + ], + "spans": [ + { + "bbox": [ + 52, + 374, + 283, + 612 + ], + "type": "image", + "image_path": "056b494f4cb937613adcb130f764578d6efc865838862e0ddf1f93f8eb5543f9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 613, + 288, + 670 + ], + "lines": [ + { + "bbox": [ + 46, + 613, + 288, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 288, + 670 + ], + "type": "text", + "content": "Figure 8. Mip-NeRF 360 results. " + }, + { + "bbox": [ + 46, + 613, + 288, + 670 + ], + "type": "inline_equation", + "content": "1^{\\text{st}}" + }, + { + "bbox": [ + 46, + 613, + 288, + 670 + ], + "type": "text", + "content": " row: we show the results on the garden scene. " + }, + { + "bbox": [ + 46, + 613, + 288, + 670 + ], + "type": "inline_equation", + "content": "2^{\\text{nd}}" + }, + { + "bbox": [ + 46, + 613, + 288, + 670 + ], + "type": "text", + "content": ": we can see that Tetra-NeRF has some problems rendering grass. " + }, + { + "bbox": [ + 46, + 613, + 288, + 670 + ], + "type": "inline_equation", + "content": "3^{\\text{rd}}" + }, + { + "bbox": [ + 46, + 613, + 288, + 670 + ], + "type": "text", + "content": ": Tetra-NeRF is able to represent the delicate texture of the tablecloth well. " + }, + { + "bbox": [ + 46, + 613, + 288, + 670 + ], + "type": "inline_equation", + "content": "4^{\\text{th}}" + }, + { + "bbox": [ + 46, + 613, + 288, + 670 + ], + "type": "text", + "content": ": Tetra-NeRF is able to recover fine geometry." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "sible because outdoor scenes contain a lot of high-resolution geometries, such as leaves and Stable View Synthesis is not able to overcome noise in the approximate geometry. Tetra-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 244, + 547, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 547, + 342 + ], + "type": "text", + "content": "NeRF does not suffer from these problems as we aggregate features in 3D along rays rather than on the surface. Mip-NeRF 360 scores comparably to Tetra-NeRF in terms of PSNR and LPIPS, but it outperforms Tetra-NeRF in terms of SSIM. However, Mip-NeRF 360 implements some tricks designed to boost its performance. On the other hand, our approach is based on vanilla NeRF, and we believe that the performance can be boosted similarly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 346, + 546, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 546, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 546, + 550 + ], + "type": "text", + "content": "Varying the number of input points We conducted a study on the effect of the input point cloud size on the reconstruction quality. We consider both the sparse and the dense point clouds and analyse the performance as we subsample the points or add more points randomly. We evaluate on the family and the truck scenes from the Tanks and Temples [25] (tt) dataset, and the room and the garden scenes from the Mip-NeRF 360 [4] dataset. The sizes of sparse point clouds were roughly 16K, 30K, 113K, and 139K respectively and the sizes of dense point clouds were larger than 5M for all scenes. When the size of the original point cloud was larger than the desired size, we selected points uniformly at random, and when smaller, we added new random points as described Section 4.1. We include detailed results in the Supp. Mat. In order to save computational resources, we only train the method for 100k iterations. The results are visualised in Figure 9a." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 552, + 547, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 547, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 547, + 649 + ], + "type": "text", + "content": "As expected, the performance improves with the number of points used, as it leads to a finer subdivision of the scene around the surface. We can also observe (e.g., from tt/family) that the sparse reconstruction leads to a better reconstruction quality at the same point cloud size. The likely explanation is that sampling uniformly at random will be biased towards regions with high density and some regions may be missed or covered extremely sparsely." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Speed of convergence Figures 9b, 9c show the rate of convergence on the ship scene from the Blender dataset and aggregated results on the Tanks and Temples [25], and Mip-NeRF 360 (indoor/outdoor) [4] datasets. Similar to Point-NeRF, good results can be achieved quite early in the train" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18465" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 76, + 213, + 198 + ], + "blocks": [ + { + "bbox": [ + 52, + 76, + 213, + 198 + ], + "lines": [ + { + "bbox": [ + 52, + 76, + 213, + 198 + ], + "spans": [ + { + "bbox": [ + 52, + 76, + 213, + 198 + ], + "type": "image", + "image_path": "9861e61b062a099c2b7fec2c1f74dc9daa6a5bff69fdcfdf20a10c42835c1acf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 545, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 545, + 278 + ], + "type": "text", + "content": "Figure 9. a) shows the PSNR with different sizes of input point clouds. The solid and dashed lines represent the dense and coarse COLMAP reconstructions. The following scenes were evaluated (best to worst): tt/family, 360/room, tt/truck, and 360/garden from Tanks and Temples [25] (tt) and Mip-NeRF 360 [4] (360-indoor/360-outdoor) datasets. As expected, the quality increases with the number of points. b) visualizes the training speed (PSNR at different training iterations) on the blender/ship scene [39]. It shows Tetra-NeRF training times for various timesteps and compares the rate of convergence with Point-NeRF [70] and Mip-NeRF [3]. Similarly, c) shows the rate of convergence aggregated over Tanks and Temples [25] (tt) and Mip-NeRF 360 [4] (360-indoor/360-outdoor) datasets. It shows the average PSNR, and the highlighted area represents the standard deviation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 215, + 73, + 376, + 198 + ], + "blocks": [ + { + "bbox": [ + 215, + 73, + 376, + 198 + ], + "lines": [ + { + "bbox": [ + 215, + 73, + 376, + 198 + ], + "spans": [ + { + "bbox": [ + 215, + 73, + 376, + 198 + ], + "type": "image", + "image_path": "a8fcbdbe0348c4c19d2d907b1d33dd173ab1363e6056d909064cc7f302ad70d9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 379, + 78, + 539, + 198 + ], + "blocks": [ + { + "bbox": [ + 379, + 78, + 539, + 198 + ], + "lines": [ + { + "bbox": [ + 379, + 78, + 539, + 198 + ], + "spans": [ + { + "bbox": [ + 379, + 78, + 539, + 198 + ], + "type": "image", + "image_path": "0cc67eab62bbbc84549638a939872ce4ef7f3874a9f6b44330922790930cc799.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 52, + 288, + 168, + 364 + ], + "blocks": [ + { + "bbox": [ + 52, + 288, + 168, + 364 + ], + "lines": [ + { + "bbox": [ + 52, + 288, + 168, + 364 + ], + "spans": [ + { + "bbox": [ + 52, + 288, + 168, + 364 + ], + "type": "image", + "image_path": "af0ad92993a5cb97e75e153e92119f9c5f9af71eb9f578ec6a1c00da16e49b38.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 365, + 166, + 384 + ], + "lines": [ + { + "bbox": [ + 53, + 365, + 166, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 166, + 384 + ], + "type": "text", + "content": "(a) Failure case 1: low point cloud density on the ground" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 171, + 287, + 289, + 365 + ], + "blocks": [ + { + "bbox": [ + 171, + 287, + 289, + 365 + ], + "lines": [ + { + "bbox": [ + 171, + 287, + 289, + 365 + ], + "spans": [ + { + "bbox": [ + 171, + 287, + 289, + 365 + ], + "type": "image", + "image_path": "31f34bbd148a667e04de4f51d7754f63b3ea14fc42569a08ca1b9bbf99ff7342.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 366, + 277, + 384 + ], + "lines": [ + { + "bbox": [ + 183, + 366, + 277, + 384 + ], + "spans": [ + { + "bbox": [ + 183, + 366, + 277, + 384 + ], + "type": "text", + "content": "(b) Failure case 2: too many intersected tetrahedra" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 384, + 286, + 405 + ], + "lines": [ + { + "bbox": [ + 47, + 384, + 286, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 286, + 405 + ], + "type": "text", + "content": "Figure 10. Failure cases. Left: barn scene from [25]. Right: tree hill scene from [4]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 416, + 286, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 286, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 286, + 548 + ], + "type": "text", + "content": "ing process. E.g., on the blender/ship scene, in 24 minutes, our method achieves performance similar to Mip-NeRF [3] after several hours of training. The efficiency of the implementation is a large factor in the speed of different methods. E.g., in Instant-NGP [41], the authors invested a significant effort in optimizing GPU memory accesses, and the result is highly optimised efficient code (a main contribution of [41]). There are several places where our implementation can easily be made more efficient. E.g., changing the memory layout (row-major vs. column-major) of the tetrahedra field leads to a " + }, + { + "bbox": [ + 46, + 416, + 286, + 548 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 416, + 286, + 548 + ], + "type": "text", + "content": " speedup." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 560, + 122, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 122, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 122, + 572 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 286, + 665 + ], + "type": "text", + "content": "One drawback of our approach is that the quality of different regions of the rendered scene depends on the density of the point cloud in these regions. If the original point cloud was constructed using 2D feature matching, there can be regions with a very low number of 3D points. We show such an example in Fig. 10a. There are few 3D points on the ground, resulting in blurry renderings for that region." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 666, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 286, + 713 + ], + "type": "text", + "content": "Another issue is that the current implementation has a limit on the number of intersected tetrahedra per ray. For large or badly structured scenes, where rays intersect many tetrahedra, this can limit the reconstruction quality. Fig. 10b" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 286, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 358 + ], + "type": "text", + "content": "shows such a case on a Mip-NeRF 360 scene [4]. This problem can be addressed by increasing the limit on the number of visited tetrahedra at the cost of higher memory requirements and run-time. Alternatively, coarse-to-fine schemes that start with a coarser tetrahedralisation and prune the space could potentially handle this issue." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 374, + 378, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 374, + 378, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 374, + 378, + 387 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 396, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 635 + ], + "type": "text", + "content": "This paper proposes a novel radiance field representation that, compared to standard voxel-based representations, is easily able to adapt to 3D geometry priors given in the form of a (sparse) point cloud. Our approach elegantly combines concepts from geometry processing (Delaunay triangulation) and triangle-based rendering (ray-triangle intersections) with modern neural rendering approaches. The representation has a naturally higher resolution in the space near surfaces, and the input point cloud provides a straightforward way to initialise the radiance field. Compared to Point-NeRF, a state-of-the-art point cloud-based radiance field representation which uses the same input as our approach, Tetra-NeRF shows clearly better results. Our method performs comparably to state-of-the-art MLP-based methods. The results demonstrate that Tetra-NeRF is an interesting alternative to existing radiance field representations that is worth further investigation. Interesting research directions include adaptive refinement and pruning of the tetrahedralisation and exploiting the fact that the surface of the scene is likely close to some of the triangles in the scene." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements This work was supported by the Czech Science Foundation (GACR) EXPRO (grant no. 23-07973X), the Grant Agency of the Czech Technical University in Prague (grant no. SGS22/112/OHK3/2T/13), and by the Ministry of Education, Youth and Sports of the Czech Republic through the e-INFRA CZ (ID:90254)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18466" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In ECCV, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 178 + ], + "type": "text", + "content": "[2] Pierre Alliez, Simon Giraudot, Clément Jamin, Florent Lafarge, Quentin Merigot, Jocelyn Meyron, Laurent Saboret, Nader Salman, Shihao Wu, and Necip Fazil Yildiran. Point set processing. In CGAL User and Reference Manual. CGAL Editorial Board, 5.5.1 edition, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 286, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 286, + 233 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 286, + 233 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. In ICCV, 2021. 1, 2, 4, 5, 7, 8, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 286, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 286, + 278 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 1, 2, 5, 7, 8, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 286, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 286, + 334 + ], + "type": "text", + "content": "[5] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv:2008.03824, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 286, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 286, + 377 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 286, + 377 + ], + "type": "text", + "content": "[6] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. NoPe-NeRF: Optimising neural radiance field with no pose prior. arXiv:2212.07388, 2022.2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 286, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 413 + ], + "type": "text", + "content": "[7] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. NeRD: Neural reflectance decomposition from image collections. In ICCV, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 286, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 286, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 286, + 457 + ], + "type": "text", + "content": "[8] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 458, + 286, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 286, + 490 + ], + "type": "text", + "content": "[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. TensoRF: Tensorial radiance fields. In ECCV, 2022. 1, 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 286, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 286, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 286, + 524 + ], + "type": "text", + "content": "[10] Anpei Chen, Zexiang Xu, Xinyue Wei, Siyu Tang, Hao Su, and Andreas Geiger. Factor fields: A unified framework for neural fields and beyond. arXiv:2302.01226, 2023. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 286, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 286, + 568 + ], + "type": "text", + "content": "[11] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 286, + 612 + ], + "type": "text", + "content": "[12] Shuo Cheng, Zexiang Xu, Shilin Zhu, Zhuwen Li, Li Erran Li, Ravi Ramamoorthi, and Hao Su. Deep stereo using adaptive thin volume representation with uncertainty awareness. In CVPR, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 668 + ], + "type": "text", + "content": "[13] Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 712 + ], + "type": "text", + "content": "[14] B Delaunay, S Vide, A Lamémoire, and V De Georges. Bulletin de l'académie des sciences de l'URSS. Classe des sciences mathématiques et naturelles, 6:793-800, 1934. 2, 3, 4, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[15] Boyang Deng, Jonathan T. Barron, and Pratul P. Srinivasan. JaxNeRF: an efficient JAX implementation of NeRF, 2020. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 129 + ], + "type": "text", + "content": "[16] Michael S Floater. Generalized barycentric coordinates and applications. Acta Numerica, 24:161-214, 2015. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 131, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 131, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 545, + 174 + ], + "type": "text", + "content": "[17] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 177, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 209 + ], + "type": "text", + "content": "[18] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32:1362-1376, 2009. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 212, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 212, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 545, + 243 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 246, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 246, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 246, + 545, + 289 + ], + "type": "text", + "content": "[20] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. SIGGRAPH, ToG, 37(6):1-15, 2018. 2, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 291, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 291, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 545, + 334 + ], + "type": "text", + "content": "[21] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In ICCV, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 337, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 545, + 369 + ], + "type": "text", + "content": "[22] Vu Hoang Hiep, Renaud Keriven, Patrick Labatut, and Jean-Philippe Pons. Towards high-resolution large-scale multiview stereo. In CVPR, 2009. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 403 + ], + "type": "text", + "content": "[23] Michal Jancosek and Tomas Pajdla. Multi-view reconstruction preserving weakly-supported surfaces. In CVPR, 2011. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 406, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 545, + 426 + ], + "type": "text", + "content": "[24] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In SGP, 2006. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 430, + 545, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 545, + 461 + ], + "type": "text", + "content": "[25] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. SIGGRAPH, ToG, 2017. 5, 6, 7, 8, 9" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 464, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 507 + ], + "type": "text", + "content": "[26] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with perview optimization. In Computer Graphics Forum, volume 40, pages 29-43. Wiley Online Library, 2021. 2, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 510, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 552 + ], + "type": "text", + "content": "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems, 25, 2012. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 555, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 587 + ], + "type": "text", + "content": "[28] Jonás Kulhánek, Erik Derner, Torsten Sattler, and Robert Babuška. ViewFormer: Nerf-free neural rendering from few images using transformers. In ECCV, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 632 + ], + "type": "text", + "content": "[29] Patrick Labatut, Jean-Philippe Pons, and Renaud Keriven. Efficient multi-view reconstruction of large-scale scenes using interest points, Delaunay triangulation and graph cuts. In ICCV, 2007. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[30] Patrick Labatut, J-P Pons, and Renaud Keriven. Robust and efficient surface reconstruction from range data. In Computer graphics forum, volume 28, pages 2275-2290. Wiley Online Library, 2009. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[31] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. BARF: Bundle-adjusting neural radiance fields. In ICCV, 2021. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18467" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[32] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "type": "text", + "content": "[33] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond. In International Conference on Learning Representations, 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "type": "text", + "content": "[34] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: learning dynamic renderable volumes from images. SIGGRAPH, ToG, 38(4):1-14, 2019. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 210, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 287, + 243 + ], + "type": "text", + "content": "[35] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3D surface construction algorithm. SIGGRAPH, ToG, 1987. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 244, + 287, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 244, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 287, + 288 + ], + "type": "text", + "content": "[36] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 290, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 290, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 290, + 287, + 322 + ], + "type": "text", + "content": "[37] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 324, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 324, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 48, + 324, + 287, + 367 + ], + "type": "text", + "content": "[38] Moustafa Meshry, Dan B. Goldman, Sameh Khamis, Hugues Hoppe, Rohit Pandey, Noah Snavely, and Ricardo MartinBrualla. Neural rerendering in the wild. In " + }, + { + "bbox": [ + 48, + 324, + 287, + 367 + ], + "type": "inline_equation", + "content": "CVPR" + }, + { + "bbox": [ + 48, + 324, + 287, + 367 + ], + "type": "text", + "content": ", June 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 370, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 370, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 370, + 287, + 414 + ], + "type": "text", + "content": "[39] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 3, 4, 5, 6, 7, 8, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 437 + ], + "type": "text", + "content": "[40] Tomas Möller. A fast triangle-triangle intersection test. Journal of Graphics Tools, 2(2):25-30, 1997. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 440, + 287, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 287, + 482 + ], + "type": "text", + "content": "[41] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. SIGGRAPH, ToG, 2022. 1, 2, 3, 4, 5, 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 485, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 517 + ], + "type": "text", + "content": "[42] M. Nießner, M. Zollhöfer, S. Izadi, and M. Stamminger. Real-time 3D reconstruction at scale using voxel hashing. SIGGRAPH, ToG, 2013. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 520, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 574 + ], + "type": "text", + "content": "[43] Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith Morley, Austin Robison, et al. OptiX: a general purpose ray tracing engine. ACM Transactions on Graphics (ToG), 29(4):1-13, 2010. 2, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 576, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 576, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 576, + 287, + 609 + ], + "type": "text", + "content": "[44] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "text", + "content": "[45] Bui Tuong Phong. Illumination for computer generated pictures. Communications of the ACM, 18(6):311-317, 1975. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "text", + "content": "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. arXiv:2209.14988, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[47] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. KiloNeRF: Speeding up neural radiance fields with thousands of tiny mlp's. In ICCV, 2021. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[48] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T Barron, and Peter Hedman. MERF: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. arXiv:2302.12249, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "type": "text", + "content": "[49] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In ICCV, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 196 + ], + "type": "text", + "content": "[50] Gernot Riegler and Vladlen Koltun. Free view synthesis. In ECCV, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 199, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 199, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 199, + 545, + 220 + ], + "type": "text", + "content": "[51] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In CVPR, 2021. 2, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 223, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 223, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 223, + 545, + 255 + ], + "type": "text", + "content": "[52] Radu Alexandru Rosu and Sven Behnke. Permutosdf: Fast multi-view reconstruction with implicit surfaces using permutohedral lattices. In CVPR, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 258, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 258, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 258, + 545, + 289 + ], + "type": "text", + "content": "[53] Darius Rückert, Linus Franke, and Marc Stamminger. ADOP: Approximate differentiable one-pixel point rendering. SIGGRAPH, ToG, 41(4):1-14, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 292, + 545, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 292, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 307, + 292, + 545, + 313 + ], + "type": "text", + "content": "[54] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 2, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "text", + "content": "[55] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 2, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 382 + ], + "type": "text", + "content": "[56] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv:1409.1556, 2014. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 385, + 545, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 385, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 307, + 385, + 545, + 417 + ], + "type": "text", + "content": "[57] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 1, 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 419, + 545, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 419, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 545, + 463 + ], + "type": "text", + "content": "[58] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-NeRF: Scalable large scene neural view synthesis. In CVPR, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 465, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 465, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 465, + 545, + 529 + ], + "type": "text", + "content": "[59] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 533, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 586 + ], + "type": "text", + "content": "[60] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, et al. Nerfstudio: A modular framework for neural radiance field development. arXiv:2302.04264, 2023. 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 589, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 621 + ], + "type": "text", + "content": "[61] Chengzhou Tang and Ping Tan. BA-net: Dense bundle adjustment networks. In International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 666 + ], + "type": "text", + "content": "[62] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[63] Sudheendra Vijayanarasimhan, Susanna Ricco, Cordelia Schmid, Rahul Sukthankar, and Katerina Fragkiadaki. SfM-Net: Learning of structure and motion from video. arXiv:1704.07804, 2017. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18468" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[64] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-NeRF: Text-and-image driven manipulation of neural radiance fields. In CVPR, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[65] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34:27171-27183, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[66] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snively, and Thomas Funkhouser. IBRNet: Learning multi-view image-based rendering. In CVPR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "text", + "content": "[67] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. NeuS2: Fast learning of neural implicit surfaces for multi-view reconstruction. arXiv:2212.05231, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "text", + "content": "[68] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF-: Neural radiance fields without known camera parameters. arXiv:2102.07064, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 352 + ], + "type": "text", + "content": "[69] Daniel N Wood, Daniel I Azuma, Ken Aldinger, Brian Curless, Tom Duchamp, David H Salesin, and Werner Stuetzle. Surface light fields for 3D photography. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, 2000. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "text", + "content": "[70] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-NeRF: Point-based neural radiance fields. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 430 + ], + "type": "text", + "content": "[71] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multiview stereo. In ECCV, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 433, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 465 + ], + "type": "text", + "content": "[72] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[73] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. BakedSDF: Meshing neural SDFs for real-time view synthesis. arXiv:2302.14859, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "type": "text", + "content": "[74] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "text", + "content": "[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 633 + ], + "type": "text", + "content": "[76] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[77] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. NeRF++: Analyzing and improving neural radiance fields. arXiv:2010.07492, 2020. 2, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[78] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, June 2018. 5, 6, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "type": "text", + "content": "[79] Chaoqiang Zhao, Qiyu Sun, Chongzhen Zhang, Yang Tang, and Feng Qian. Monocular depth estimation based on deep learning: An overview. Science China Technological Sciences, 63(9):1612-1627, 2020. 3" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18469" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/698eb5a8-8d74-4b19-b9a7-1fe068b4d1fd_content_list.json b/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/698eb5a8-8d74-4b19-b9a7-1fe068b4d1fd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ad7ed15d7cbd71588b9e76ec12e2d85523f84c64 --- /dev/null +++ b/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/698eb5a8-8d74-4b19-b9a7-1fe068b4d1fd_content_list.json @@ -0,0 +1,1954 @@ +[ + { + "type": "text", + "text": "TexFusion: Synthesizing 3D Textures with Text-Guided Image Diffusion Models", + "text_level": 1, + "bbox": [ + 140, + 89, + 828, + 133 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianshi Cao $^{1,2,3}$", + "bbox": [ + 98, + 160, + 232, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Karsten Kreis", + "bbox": [ + 261, + 161, + 385, + 179 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sanja Fidler $^{1,2,3}$", + "bbox": [ + 413, + 161, + 542, + 179 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nicholas Sharp $^{1,*}$", + "bbox": [ + 573, + 161, + 712, + 180 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kangxue Yin $^{1,*}$", + "bbox": [ + 743, + 161, + 867, + 179 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ NVIDIA,", + "bbox": [ + 310, + 180, + 380, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Toronto,", + "bbox": [ + 395, + 181, + 539, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Vector Institute", + "bbox": [ + 553, + 181, + 658, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{tianshic, kkreis, sfidler, nsharp, kangxuey}@nvidia.com", + "bbox": [ + 232, + 200, + 727, + 215 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2be179c4a496587e538c93d3ec1f159a45848594f67b3190a66c184d9434b3db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 174, + 247, + 331, + 343 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee7570178093fd5858103cc7570b20f724afff240872004e423356b2fbffdde4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 258, + 468, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/740797af04e6edcb45341dbbc02b0a26e86ca97f6ad3853b2fe850f064762177.jpg", + "image_caption": [ + "\"Crocodile skin handbag\"" + ], + "image_footnote": [], + "bbox": [ + 470, + 258, + 560, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41f09a7a345570b2c8ab89b5f6c8376f9f8691b3c7df5d547e30e740731d7119.jpg", + "image_caption": [ + "\"White Mercedes Benz SUV\"" + ], + "image_footnote": [], + "bbox": [ + 571, + 258, + 656, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9331011979d2a6b5de089c668e779b29776a486c708bef2efd0f19252c643358.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 273, + 797, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80ccf636a0e14e45a7d3f86fbe0e0260c4d4b6cc0c8df82cd2605f65436a4ba8.jpg", + "image_caption": [ + "\"White humanoid robot, movie poster, main character of a science fiction movie\"", + "\"Oil painting of a bald, middle aged banker with pointed moustache\"" + ], + "image_footnote": [], + "bbox": [ + 173, + 362, + 349, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b67e6e0e19d4a3408ade25aa37a0ee05cd0f93e3952a4c9d33a7f208b38d9cc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 363, + 449, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0b7ebf7c5ea315d5d0c2ac36573d8f1f493a3909a219c0670f512e1a9cf74dd4.jpg", + "image_caption": [ + "\"Portrait photo of Abraham Lincoln, full color\"" + ], + "image_footnote": [], + "bbox": [ + 468, + 358, + 558, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b2e1084442a5131dd593d45319288578e40cee64bcbe020c8a08ea4619b49045.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 354, + 669, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0ca7c7502335d034f196e2b5cd00e73bc6c3432995f6d26a5cef59671bd9caa8.jpg", + "image_caption": [ + "\"Railroad worker wearing high-vis vest\"", + "Figure 1: Text-conditioned 3D texturing results with TexFusion." + ], + "image_footnote": [], + "bbox": [ + 678, + 356, + 746, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 522, + 313, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present TexFusion (Texture Diffusion), a new method to synthesize textures for given 3D geometries, using large-scale text-guided image diffusion models. In contrast to recent works that leverage 2D text-to-image diffusion models to distill 3D objects using a slow and fragile optimization process, TexFusion introduces a new 3D-consistent generation technique specifically designed for texture synthesis that employs regular diffusion model sampling on different 2D rendered views. Specifically, we leverage latent diffusion models, apply the diffusion model's denoiser on a set of 2D renders of the 3D object, and aggregate the different denoising predictions on a shared latent texture map. Final output RGB textures are produced by optimizing an intermediate neural color field on the decodings of 2D renders of the latent texture. We thoroughly validate TexFusion and show that we can efficiently generate diverse, high quality and globally coherent textures. We achieve state-of-the-art text-guided texture synthesis performance using only image diffusion models, while avoiding the pitfalls of previous distillation-based methods. The text-conditioning offers detailed control and we also do not rely on any ground truth 3D textures for training. This makes our method versatile", + "bbox": [ + 76, + 546, + 472, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and applicable to a broad range of geometry and texture types. We hope that TexFusion will advance AI-based texturing of 3D assets for applications in virtual reality, game design, simulation, and more.", + "bbox": [ + 500, + 508, + 890, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 577, + 632, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the past decade, deep learning-based 3D object generation has been studied extensively [1, 4, 18, 23, 25, 27, 32, 33, 37, 39, 44, 45, 50, 52, 53, 55, 58, 69, 71, 72, 78, 82, 85, 87, 89, 91-93, 95, 97, 100, 101], due to the demand for high-quality 3D assets in 3D applications such as VR/AR, simulation, digital twins, etc. While many prior works on 3D synthesis focus on the geometric components of the assets, textures are studied much less, despite the fact that they are important components of realistic 3D assets which assign colors and materials to meshes to make the rendering vivid. Recent advances in text-conditioned image diffusion models trained on internet-scale data [2, 54, 61, 63, 66] have unlocked the capability to generate images with stunning visual detail and practically unlimited diversity. These high-performance diffusion models have also been used as image priors to synthesize 3D objects with textures using textual guidance [40, 48, 59, 84].", + "bbox": [ + 496, + 597, + 890, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim to perform text-driven high-quality 3D texture synthesis for given meshes, by leveraging the information about the appearance of textures carried by the", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 101, + 887, + 215, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4169", + "bbox": [ + 482, + 925, + 514, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image prior of a pre-trained text-to-image diffusion model. The main workhorse in current text-to-3D methods that leverage 2D image diffusion models is Score Distillation Sampling (SDS) [59]. SDS is used to distill, or optimize, a 3D representation such that its renders are encouraged to be high-likelihood under the image prior. Methods utilizing SDS also share two common limitations in that: 1). a high classifier-free guidance weight [22] is required for the optimization to converge, resulting in high color saturation and low generation diversity; 2). a lengthy optimization process is needed for every sample.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we present Texture Diffusion, or TexFusion for short. TexFusion is a sampler for sampling surface textures from image diffusion models. Specifically, we use latent diffusion models that efficiently autoencode images into a latent space and generate images in that space [63, 81]. TexFusion leverages latent diffusion trajectories in multiple object views, encoded by a shared latent texture map. Rendering of the shared latent texture map are provided as input to the denoiser of the latent diffusion model, and the output of every denoising step is projected back to the shared texture map in a 3D-consistent manner. To transform the generated latent textures into RGB textures, we optimize a neural color field on the outputs of the latent diffusion model's decoder applied to different views of the object. We use the publicly available latent diffusion model Stable-Diffusion with depth conditioning [63] (SD2-depth) as our diffusion backbone. Compared to methods relying on SDS, TexFusion produces textures with more natural tone, stronger view consistency, and is significantly faster to sample (3 minutes vs. 30 minutes reported by previous works).", + "bbox": [ + 75, + 258, + 472, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We qualitatively and quantitatively validate TexFusion on various texture generation tasks. We find that TexFusion generates high quality, globally coherent and detailed textures that are well-aligned with the text prompts used for conditioning (e.g. Fig. 1). Since we leverage powerful text-to-image diffusion models for texture sampling, we can generate highly diverse textures and are not limited to single categories or restricted by explicit 3D textures as training data, which are limitations of previous works [7, 11, 18, 57, 73]. In summary, our main contribution is a novel method for 3D texture generation from 2D image diffusion models, that is view-consistent, avoids over-saturation and achieves state-of-the-art text-driven texture synthesis performance.", + "bbox": [ + 75, + 575, + 472, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 787, + 218, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Classic Computer Graphics Techniques Early work on texture generation focused on tiling exemplar patterns across a surface, often with an explicit direction field for local orientation [34, 36, 80, 86]. See [88] for a survey. This research established the value of texture image representations and the challenges of global coherence, both central", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to this work. However, modern learning-based priors have proven necessary to go beyond simple patterns and synthesize plausible shape-specific texture details.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Texture Synthesis with 3D Priors Textures are defined on the meshed surface of 3D objects, which is an irregular representation. To enable 2D texture generation on 3D meshes, AUV-Net [11] learns an aligned UV space for a set of 3D shapes in a given class, mapping 3D texture synthesis to a 2D domain. Texturify [73] trains a 3D StyleGAN in the quad-parameterized surface domain on a set of textured 3D shapes in a given class. Different from AUV-NET or Texturify, which embed the geometric prior into a UV map or mesh parameterization, EG3D [7] and GET3D [18] directly train 3D StyleGANs to generate geometry and texture jointly, where the textures are implicit texture fields [57]. Other works also represent 3D texture by vertex colors [49], voxels [9], cube mapping [90], etc. In contrast to TexFusion, these works mostly don't offer text conditioning, often work only on single object categories or require textured 3D shapes for training, which limits their broad applicability.", + "bbox": [ + 496, + 150, + 893, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models Our approach directly builds on diffusion models [5, 12, 94], which have recently emerged as new state-of-the-art generative models. In particular, they have demonstrated outstanding performance in image generation [14-16, 21, 56, 81], outperforming GANs, and led to breakthroughs in text-to-image synthesis [2, 54, 61, 63, 66]. They have also been successfully used for a variety of image editing and processing tasks [17, 19, 30, 38, 43, 47, 64, 65, 67, 70, 77]. Moreover 3D object generation has been addressed with diffusion models, too, for instance leveraging point clouds [45, 55, 97, 101], meshes [27], or neural fields [4, 52, 72, 85] as 3D representations. However, these works focus on geometry generation and do not specifically tackle 3D texture synthesis.", + "bbox": [ + 496, + 421, + 893, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Distilling 3D Objects from 2D Image Diffusion Models Recently, large-scale 2D text-to-image diffusion models have been leveraged to distill individual 3D objects as neural radiance fields using Score Distillation Sampling (SDS) [13, 40, 48, 59, 84]. In SDS, the radiance field is rendered from different directions into 2D images and it is optimized such that each render has a high probability under the text-to-image diffusion model while conditioning on a text prompt. DreamFusion [59] pioneered the approach, Magic3D [40] proposed a coarse-to-fine strategy improving quality, and Latent-NeRF [48] performs distillation in latent space leveraging a latent diffusion model [63]. These approaches do not specifically target texture generation, which is the focus of this work. More deeply, a crucial drawback of this line of work is that SDS typically requires strong guidance [22] to condition the diffusion model, which can hurt quality and diversity. Moreover, SDS's", + "bbox": [ + 496, + 643, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4170", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/62a8b985cad542cd68bad8c5a63f5d957d8a5bf86481af1571019c27b918e07d.jpg", + "image_caption": [ + "Distill with Neural Color Field" + ], + "image_footnote": [], + "bbox": [ + 78, + 78, + 676, + 271 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5e6ac964bdbdf3700fd301f627c54e9cd45b942d2e2821452810a02e111b4ffd.jpg", + "image_caption": [ + "Figure 2: Overview of TexFusion. TexFusion takes a text prompt and mesh geometry as input and produces a UV parameterized texture image matching the prompt and mesh. Key to TexFusion is the Sequential Interlaced Multiview Sampler (SIMS) - SIMS performs denoising diffusion iterations in multiple camera views, yet the trajectories are aggregated through a latent texture map after every denoising step. SIMS produces a set of 3D consistent latent images (TexFusion uses Stable Diffusion [63] as text-to-image diffusion backbone), which are decoded and fused into a texture map via optimizing an intermediate neural color field." + ], + "image_footnote": [], + "bbox": [ + 687, + 97, + 893, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "iterative optimization process makes synthesis very slow. In contrast, our approach avoids SDS entirely and leverages regular diffusion model sampling in a new, 3D-consistent manner. Earlier works also leverage CLIP [60] for 3D object or texture distillation [10, 26, 31, 49], but this performs usually worse than using diffusion models instead.", + "bbox": [ + 75, + 357, + 470, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concurrent Work Concurrently with this work, TEX-Ture [62] proposes a related approach. Like this work, TEX-Ture performs multiview denoising on a texture map representation. However, TEX-Ture runs an entire generative denoising process in each camera view in sequence, conditioning on the previous views and projecting to the texture map only after full denoising. In contrast, in TexFusion we propose to interleave texture aggregation with denoising steps in different camera views, simultaneously generating the entire output. This insight significantly reduces view inconsistencies and improves quality in TexFusion compared to TEX-Ture, as validated in Sec. 5. MultiDiffusion [3] concurrently introduces a method for panorama generation and other controlled image generation tasks, leveraging relatively lower-resolution image diffusion models. Algorithmically, this approach of aggregating different denoising predictions from different image crops is closely related to TexFusion's aggregation from different camera views. However, MultiDiffusion only tackles image synthesis, and is not concerned with any 3D or texture generation at all.", + "bbox": [ + 75, + 464, + 472, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 76, + 797, + 205, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models Diffusion models [20, 74, 76] model a data distribution $p_{\\mathrm{data}}(\\mathbf{x})$ via iterative denoising, and are trained with denoising score matching [20, 24, 46, 74, 76, 83]. Given samples $\\mathbf{x} \\sim p_{\\mathrm{data}}$ and $\\epsilon \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ , a denoiser model $\\epsilon_{\\theta}$ parameterized with learnable parameters", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\theta$ receives diffused inputs $\\mathbf{x}_t(\\epsilon, t, \\mathbf{x})$ and is optimized by minimizing the denoising score matching objective", + "bbox": [ + 498, + 357, + 890, + 388 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\mathbf {x} \\sim p _ {\\text {d a t a}}, t \\sim p _ {t}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}; \\mathbf {c}, t) \\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 396, + 890, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{c}$ is optional conditioning information, such as a text prompt, and $p_t$ is a uniform distribution over the diffusion time $t$ . The model effectively learns to predict the noise $\\epsilon$ that was used to perturb the data (other formulations are possible [28, 68]). Letting $\\alpha_{t}$ define a noise schedule, parameterized via a diffusion-time $t$ , we construct $\\mathbf{x}_t$ as $\\mathbf{x}_t = \\sqrt{\\alpha_t}\\mathbf{x} + \\sqrt{1 - \\alpha_t}\\epsilon$ , $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ ; this particular formulation corresponds to a variance-preserving schedule [76]. The forward diffusion as well as the reverse generation process in diffusion models can be described in a continuous-time framework [76], but in practice a fixed discretization can be used [20]. The maximum diffusion time is generally chosen such that the input data is entirely perturbed into Gaussian random noise and an iterative generative denoising process can be initialized from such Gaussian noise to synthesize novel data.", + "bbox": [ + 496, + 422, + 892, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-free guidance [22] can be used for improved conditioning. By randomly dropping out the conditioning c during training, we can learn both a conditional and an unconditional model at the same time, and their predictions can be combined to achieve stronger conditioning.", + "bbox": [ + 496, + 665, + 890, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We perform iterative diffusion model sampling via the Denoising Diffusion Implicit Models (DDIM) scheme [75]:", + "bbox": [ + 498, + 739, + 890, + 772 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {x} _ {i - 1} = \\sqrt {\\alpha_ {i - 1}} \\left(\\frac {\\boldsymbol {x} _ {i} - \\sqrt {1 - \\alpha_ {i}} \\boldsymbol {\\epsilon} _ {\\theta} ^ {(t _ {i})} (\\boldsymbol {x} _ {i})}{\\sqrt {\\alpha_ {i}}}\\right) \\tag {2} \\\\ + \\sqrt {1 - \\alpha_ {i - 1} - \\sigma_ {t _ {i}} ^ {2}} \\cdot \\pmb {\\epsilon} _ {\\theta} ^ {(t _ {i})} (\\pmb {x} _ {i}) + \\sigma_ {t _ {i}} \\pmb {\\epsilon} _ {t _ {i}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 792, + 890, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $\\epsilon_{t_i}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and $\\sigma_{t_i}$ is a variance hyperparameter. We express obtaining $\\pmb{x}_{i - 1}$ via DDIM sampling as $x_{i - 1}\\sim$", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4171", + "bbox": [ + 480, + 924, + 513, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\pmb{f}_{\\theta}^{(t_i)}(\\pmb{x}_{i-1}|\\pmb{x}_i)$ . See Supp. Material for more details.", + "bbox": [ + 76, + 89, + 426, + 108 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent Diffusion Models (LDMs) and Stable Diffusion Instead of directly operating on pixels, LDMs [63] utilize an encoder $\\mathcal{E}$ and a decoder $\\mathcal{D}$ for translation between images $\\xi$ and latents $\\mathbf{x} \\in X$ of a lower spatial dimension. The diffusion process is then defined over the distribution of $X$ . Stable Diffusion is an LDM trained on the LAION-5B image-text dataset. In addition to text-conditioning, Stable Diffusion 2.0 permits depth conditioning with a depth map $D$ (SD2-depth). This allows detailed control of the shape and configuration of objects in its synthesized images.", + "bbox": [ + 76, + 127, + 472, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Rendering and Geometry Representation** In principle, our method applies to any geometry representation for which a textured surface can be rendered; in practice, our experiments use a surface mesh $\\mathcal{M} = (\\mathcal{V},\\mathcal{F})$ , with vertices $\\mathcal{V} = \\{v_i\\}$ , $v_{i}\\in \\mathbb{R}^{3}$ and triangular faces $\\mathcal{F} = \\{f_i\\}$ where each $f_{i}$ is a triplet of vertices. Textures are defined in 2D image space in an injective UV parameterization of $\\mathcal{M}$ , $UV:p\\in \\mathcal{M}\\mapsto (u,v)\\in [0,1]^2$ . If needed, this parameterization can be automatically constructed via tools such as XATLAS [6, 96]. We encode textures as multi-channel images discretized at pixels in UV space $\\mathbf{z}\\in \\mathbb{R}^{(H\\times W,C)}$ , notationally collapsing the spatial dimension for simplicity.", + "bbox": [ + 76, + 300, + 468, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We denote the rendering function as $\\mathcal{R}(\\mathbf{z};\\mathcal{M},C):\\mathbf{z}\\mapsto$ $\\mathbf{x}$ $\\mathbf{x}\\in \\mathbb{R}^{(h\\times w,C)}$ , which takes as input a mesh $\\mathcal{M}$ , camera $C$ , and texture $\\mathbf{z}$ , and produces as output a rendered image. The inverse of this function $\\mathcal{R}^{-1}(\\mathbf{x};\\mathcal{M},C):\\mathbf{x}\\mapsto \\mathbf{z}$ projects values from camera image-space onto the UV texture map. Notationally, we often omit the dependence on $\\mathcal{M}$ and $C$ for brevity. In this work, we do not model any lighting or shading effects, images are formed by directly projecting textures into screen space (and then decoding, for latent textures).", + "bbox": [ + 75, + 482, + 470, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Texture Sampling with 2D Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 648, + 468, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a mesh geometry $\\mathcal{M}$ , a text prompt $y$ , and a geometry (depth) conditioned image diffusion model $\\theta$ , how could one sample a complete surface texture? Assuming access to the rendering $\\mathcal{R}$ and inverse rendering $\\mathcal{R}^{-1}$ functions defined above, perhaps the most naive approach is to compute a set of $\\{C_1,..,C_N\\}$ camera views that envelopes the surface, render the depth map $d_{n}$ in each view, sample images from the image diffusion model with depth conditioning, and then back-project these images to the mesh surface (e.g. as done in [29]). However, image diffusion models in each view have no information about the generated results in other views, thus there is no coherency in the contents generated in each view. As an alternative, one may define a canonical order of camera views, and autoregressively condition image sampling in the subsequent views on", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "previously sampled regions (as done in [48, 62]). However, for most geometries of interest (i.e. not a plane), a single camera can not observe the entirety of the geometry. Consequently, images synthesized early in the sequence could produce errors that are not reconcilable with the geometry that is observed later in the sequence (see Fig. 3). Thus, it is desirable for the image sampling distribution $p(\\cdot | d_i, y)$ in each camera view to be conditioned on that of every other camera view.", + "bbox": [ + 496, + 90, + 893, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7c2b3b8da3a5357c3094c2ff57a1b9983a0efdb4c0ea5354ba408ab5edd0c9a2.jpg", + "image_caption": [ + "Figure 3: Illustration of irreconcilable mistakes in early views impacting denoised results in later views (images sampled from TEXTure). While the highlighted area appears natural in the first view, it does not match the geometry when viewed from a different angle, thereby creating poorly denoised results when the second view is inpainted." + ], + "image_footnote": [], + "bbox": [ + 535, + 234, + 857, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Sequential Interlaced Multiview Sampler", + "text_level": 1, + "bbox": [ + 498, + 435, + 852, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Leveraging the sequential nature of the denoising process, we can interlace the synchronization of content across views with the denoising steps within each view to achieve coherency over the entire shape. Suppose that at step $i$ of the denoising process, we have a set of partially denoised images $\\{\\mathbf{x}_{i,n}\\}_{n = 1}^{N} = \\{\\mathbf{x}_{i,1}\\dots \\mathbf{x}_{i,N}\\}$ . Our goal is to sample $\\{\\mathbf{x}_{i - 1,n}\\}_{n = 1}^{N} = \\mathbf{x}_{i - 1,1}\\dots \\mathbf{x}_{i - 1,N}$ that is 3D-consistent, i.e., two pixels in $\\mathbf{x}_{i - 1,a},\\mathbf{x}_{i - 1,b}$ that project to the same point in 3D should have the same value. Taking inspiration from the autoregressive modeling literature, we sample the joint distribution via first decomposing it into a product of conditionals, and approximating each term in the product by using the diffusion model to denoise the renders of a dynamically updated latent texture map. Specifically, we first initialize an initial latent texture map $\\mathbf{z}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , ( $T$ demarks the first step of the diffusion trajectory). Then, suppose that we have a 3D consistent latent texture map $\\mathbf{z}_i$ , we decompose the joint distribution as follows (conditioning on depth and prompt omitted for space):", + "bbox": [ + 496, + 458, + 893, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p _ {\\theta} \\left(\\left\\{\\mathbf {x} _ {i - 1, j} \\right\\} _ {j = 1} ^ {N} \\mid \\mathbf {z} _ {i}\\right) = p _ {\\theta} \\left(\\mathbf {x} _ {i - 1, 1} \\mid \\mathbf {z} _ {i}\\right) \\times \\\\ \\prod_ {n = 2} ^ {N} p _ {\\theta} \\left(\\mathbf {x} _ {i - 1, n} \\mid \\left\\{\\mathbf {x} _ {i - 1, j} \\right\\} _ {j = 1} ^ {n - 1}, \\mathbf {z} _ {i}\\right) \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 753, + 890, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can compute the first term by first rendering $\\mathbf{z}_i$ into $\\mathbf{x}_{i,1}^{\\prime} = \\mathcal{R}(\\mathbf{z}_i;C_1)$ . Eqn. 2 can now be applied to $\\mathbf{x}_{i,1}^{\\prime}$ to draw a latent image at the next time step:", + "bbox": [ + 498, + 827, + 893, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {i - 1, 1} \\sim f _ {\\theta} ^ {(t _ {i})} \\left(\\mathbf {x} _ {i - 1, 1} \\mid \\mathbf {x} _ {i, 1} ^ {\\prime} = \\mathcal {R} \\left(\\mathbf {z} _ {i}; C _ {1}\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 882, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4172", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Later terms in Eq. 3 additionally depend on the result of previous denoising steps. We again model this dependency through the latent texture map. For each view starting at $n = 1$ , we inverse render $\\mathbf{x}_{i - 1,n}$ into texture space to obtain $\\mathbf{z}_{i - 1,n}^{\\prime} = \\mathcal{R}^{-1}(\\mathbf{x}_{i - 1,n})$ , and update the pixels of $\\mathbf{z}_{i - 1,n - 1}$ that are newly visible in $\\mathbf{z}_{i - 1,n}^{\\prime}$ to obtain $\\mathbf{z}_{i - 1,n}$ (See Sec. 4.2.2 for details). Then, in the $n + 1$ iteration, since $\\mathbf{z}_{i - 1,n}$ contain regions at two noise scales (unseen regions are at noise scale $\\sigma_{i}$ , while visited regions are at noise scale $\\sigma_{i - 1}$ ), we add appropriate 3D consistent noise to the visited regions of $\\mathbf{z}_{i - 1,n}$ to match the noise scale of step $i$ before rendering it as input to $f_{\\theta}$ . Letting $M_{i,n}$ represent the mask for visited regions and $\\epsilon_{i}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , we can write the sampling procedure for $\\mathbf{x}_{i - 1,n}$ as:", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {z} _ {i, n} = M _ {i, n} \\odot \\left(\\sqrt {\\frac {\\alpha_ {i - 1}}{\\alpha_ {i}}} \\mathbf {z} _ {i - 1, n - 1} + \\sigma_ {i} \\epsilon_ {i}\\right) \\\\ + \\left(\\mathbf {1} - M _ {i, n}\\right) \\odot \\mathbf {z} _ {i} \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 316, + 468, + 369 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {i - 1, n} \\sim f _ {\\theta} ^ {(t _ {i})} (\\mathbf {x} _ {i - 1, n} | \\mathbf {x} _ {i, n} ^ {\\prime} = \\mathcal {R} (\\mathbf {z} _ {i, n}; C _ {n}))\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 373, + 411, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By iteratively applying Eqn. 5, we obtain a sequence of 3D consistent images $\\{\\mathbf{x}_{i - 1,n}\\}_{n = 1}^{N}$ and a texture map $\\mathbf{z}_{i - 1,n}$ that has been aggregated from these images. We can then decrement $i$ and repeat this process in the next time step.", + "bbox": [ + 75, + 404, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We name this approach Sequential Interlaced Multiview Sampler (SIMS). SIMS communicates the denoising direction of previous views to the current view and resolves overlaps between views by the order of aggregation. It ameliorates inconsistent predictions while circumventing performance degradation due to the averaging of latent predictions during parallel aggregation. In the single-view case, SIMS is equivalent to standard DDIM sampling. A complete algorithm for SIMS can be found in the appendix.", + "bbox": [ + 75, + 465, + 468, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. The TexFusion Algorithm", + "text_level": 1, + "bbox": [ + 76, + 614, + 312, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Sec. 4.1, we have presented a generic algorithm for sampling 3D consistent multi-view images and texture maps using 2D diffusion models. We now present a concrete algorithm, TexFusion, that uses SIMS to texture 3D meshes using SD2-depth [63] as the diffusion model. An illustrative overview of TexFusion can be found in Fig. 2.", + "bbox": [ + 75, + 638, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As SD2-depth is a latent diffusion model, we apply SIMS in the latent space: $\\mathbf{x}$ and $\\mathbf{z}$ represent latent images and latent texture maps respectively, and Section 4.2.3 will describe a final distillation post-process to color space. In this section, we address several challenges specific to using LDMs with SIMS, and detail our design choices to tackle these challenges. We find that a canonical set of cameras works well for most objects, but cameras are tailored to specific objects to improve resolution and ameliorate occlusion. We further illustrate a technique for obtaining high-quality results by operating SIMS in a cascade of increasing resolutions.", + "bbox": [ + 75, + 729, + 470, + 910 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 (Inverse) Rendering of Latent Textures", + "text_level": 1, + "bbox": [ + 498, + 90, + 825, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use NVDIFFRAST [35] to efficiently render textured meshes via rasterization, as described in Sec. 3. Our implementation sets the rendered image size $h = w = 64$ to match Stable Diffusion's UNet, with latent vectors of dimension $D = 4$ . The texture image dimensions $H$ and $W$ are chosen based on the surface area of the mesh relative to its diameter (detailed in Sec. 4.2.4).", + "bbox": [ + 496, + 114, + 890, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each non-background pixel $s$ in a rendered image $\\mathbf{x}_{i,..}^{\\prime}$ , rasterization gives a corresponding location on the texture image via the UV map $(u,v) = UV(p)$ , and we retrieve the value at the nearest texture pixel. This texture value is latent, and no shading or lighting is applied. In other settings, texture data is often rendered with bilinear filtering and mipmaps to improve image quality and reduce aliasing, but in this setting, we found it essential to avoid such techniques. We experimentally ablate texturing approaches in the supplementary. The issue is that early in the diffusion process, i.e. when $t_i\\ll 0$ , $\\mathbf{z}_i$ and $\\mathbf{x}_{i,..}^{\\prime}$ are dominated by noise, but interpolation and mipmapping change the variance in the pixels $\\mathbf{x}_{i,..}$ , thereby moving $\\mathbf{x}_{i,..}^{\\prime}$ out of the training distribution of $\\epsilon_{\\theta}^{(t_i)}$ . Instead, we retrieve only the nearest texture pixels for diffusion, and resolve aliasing and interpolation via a simple distillation post-process (Sec. 4.2.3). For each background pixel in rendered image $\\mathbf{x}_{i,..}^{\\prime}$ , we apply a Gaussian noise of standard deviation $\\sigma_{i}$ such that the background matches the marginal distribution at diffusion step $t_i$ . This ensures that the diffusion model $f_{\\theta}^{(t_i)}$ focuses solely on the foreground pixels of $\\mathbf{x}_{i,..}^{\\prime}$ .", + "bbox": [ + 496, + 222, + 892, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that in this setting a rendered image is simply a selection of pixels from the texture maps, and thus we can leverage backpropagation to easily implement the inverse rendering function $\\mathcal{R}^{-1}$ . Additionally, forward rasterization yields a depth map of the scene, which we use as a conditioning input to the diffusion model.", + "bbox": [ + 496, + 542, + 893, + 632 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d5c6ef76b5bab196f4581d3541c42694133e12b8572ae7ea659f5f4f80c147e1.jpg", + "image_caption": [ + "Figure 4: Left: Depth map for conditioning SD2-depth, and quality image computed from screen space derivatives. Right: Output of SD2 decoder in two views using renders of the latent texture map as input. Note how the horizontal line across the doors changes appearance from one view to another." + ], + "image_footnote": [], + "bbox": [ + 535, + 646, + 854, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Aggregating Latent Textures.", + "text_level": 1, + "bbox": [ + 498, + 815, + 758, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recall from Sec. 4.1, before iterating through cameras $\\{C_1,\\dots,C_N\\}$ , we first initialize a latent texture map $\\mathbf{z}_{i - 1,0} = \\mathbf{z}_i$ and render latent image $\\mathbf{x}_{i,1}^{\\prime} = \\mathcal{R}(\\mathbf{z}_{i - 1,0})$ . Then, in each step (iterating through cameras), we ob", + "bbox": [ + 498, + 839, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4173", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tain $\\mathbf{x}_{i - 1,n}$ from $f_{\\theta}^{(t_i)}(\\mathbf{x}_{i,n}^{\\prime})$ , and inverse render it to get $\\mathbf{z}_{i - 1,n}^{\\prime} = \\mathcal{R}^{-1}(\\mathbf{x}_{i - 1,n})$ . Finally, we need to aggregate the partial texture map $\\mathbf{z}_{i - 1,n}^{\\prime}$ with $\\mathbf{z}_{i - 1,n - 1}$ to obtain a partially updated texture map $\\mathbf{z}_{i - 1,n}$ . We perform this aggregation step based on a simple heuristic that the value of each pixel $(u,v)$ in $\\mathbf{z}$ should be determined by the camera that has the most \"direct and up-close\" view to its corresponding point on the mesh. We measure view quality using image-space derivatives - the amount of change in UV coordinates per infinitesimal change in the image coordinate. This quantity is commonly used for determining mipmapping resolution and anti-aliasing, and it can be efficiently computed by nvdiffrast when rasterizing $\\mathbf{z}_{i,n - 1}$ . For each pixel location $(p,q)$ in $\\mathbf{x}_{i - 1,n}$ , we compute the negative Jacobian magnitude as $-|\\frac{\\partial u}{\\partial p}\\cdot \\frac{\\partial v}{\\partial q} -\\frac{\\partial u}{\\partial q}\\cdot \\frac{\\partial v}{\\partial p} |,$ and inverse render it to the texture space, which we denote as $Q_{i,n}$ . Higher values of $Q_{i,n}(u,v)$ means that camera $n$ has a better view of $(u,v)$ (see Fig.4). In addition, we compute $M_{i,n} = \\mathcal{R}^{-1}(\\boldsymbol {I})$ , such that $M_{i,n}(u,v)$ represents the number of pixels in $\\mathbf{x}_{i,n}^{\\prime}$ that received value from $\\mathbf{z}_i(u,v)$", + "bbox": [ + 76, + 89, + 472, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While iterating through cameras 1 through $N$ , we maintain mask $M_{i}$ (initialized to 0), which is used to track which pixels of $\\mathbf{z}_{i,n}$ have been seen by cameras up to $n$ , and quality buffer $Q_{i}$ (initialized to $-\\inf$ ), which is used to track the highest non-zero value of $Q_{i,n}$ at each $(u,v)$ . The value at $(u,v)$ of $\\mathbf{z}_{i-1,n}$ is determined as:", + "bbox": [ + 76, + 396, + 470, + 488 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i - 1, n} (u, v) = \\left\\{ \\begin{array}{l l} \\frac {\\mathbf {z} _ {i - 1 , n} ^ {\\prime} (u , v)}{M _ {i , n} (u , v)} & M _ {i, n} (u, v) > 0, \\text {a n d} \\\\ \\mathbf {z} _ {i - 1, n - 1} (u, v) & Q _ {i, n} (u, v) > Q _ {i} (u, v) \\\\ \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 500, + 467, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$M_{i}$ and $Q_{i}$ are then updated pixelwise as $M_{i} = \\min (M_{i,n} + M_{i},1)$ and $Q_{i} = \\max (Q_{i,n},Q_{i})$ (min and max are applied element-wise). We use Eqn. 6 in conjunction with Eqn. 5 in SIMS to perform denoising of images in the sequence of camera views.", + "bbox": [ + 76, + 566, + 468, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further note that a side effect of using nearest pixel filtering during texture sampling in SIMS is aliasing. When the image space derivative Jacobian is much higher than 1 (low quality), neighboring pixels in screen space will have gaps when mapped to uv space. This results in incomplete filling of the latent texture map. Conversely, when the image space derivative Jacobian magnitude is much smaller than 1 (high quality), multiple screen space pixels $(p,q)$ will map to the same $(u,v)$ , creating uniform blocks in $\\mathbf{x}$ . Similar to interpolated latents, these blocks are particularly detrimental early during the diffusion process, as they cannot be correctly denoised by $\\epsilon_{\\theta}$ which is expecting high spatial variance. Our quality-based aggregation overwrites low-quality regions in the latent texture when higher quality views are available. Thus we can set the resolution of the texture map such that the image space derivatives in each view do not fall below 1 (i.e. $\\max Q_{i,n} < -1$ ) to prevent", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "aliasing of the second type, and rely on cameras with better views to fix aliasing of the first type.", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3 From Latent to RGB Textures", + "text_level": 1, + "bbox": [ + 498, + 127, + 764, + 142 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "So far, we have described how to use SIMS to produce a set of 3D consistent latent images and latent texture maps, but have yet to describe how to translate this into a RGB texture map that can be rendered for viewing. Towards this, we experimented with multiple approaches and found performing multi-view distillation of decoded latent images $\\{\\mathbf{x}_{0,n}\\}_{n=1}^{N}$ with a neural color field to be most performant. Specifically, we use the decoder $\\mathcal{D}$ of Stable Diffusion to decode latent multi-view images $\\{\\mathbf{x}_{0,n}\\}_{n=1}^{N}$ into RGB multi-view images $\\{\\xi_{n} = \\mathcal{D}(\\mathbf{x}_{0,n})\\}_{n=1}^{N}$ . Notably, decoding with $\\mathcal{D}$ introduce inconsistencies such that $\\{\\xi_{n}\\}_{n=1}^{N}$ is not 3D consistent even when $\\{\\mathbf{x}_{0,n}\\}_{n=1}^{N}$ is 3D consistent (see Fig. 4 for example). In stable Diffusion, each latent vector (pixel of $\\mathbf{x}$ ) needs to carry the information of a $8 \\times 8$ patch of RGB pixels. Thus, the value of the latent vector encodes both color values and spatial patterns. We therefore cannot expect their decoded results to be equivariant to perspective transformations. To address this problem, we leverage a neural color field optimized with appearance-based losses to smooth out inconsistencies. Since we know the ground-truth camera poses, we can directly obtain the 3D spatial coordinates $\\{xyz_{n}\\}_{n=1}^{N}$ of all pixels of $\\{\\xi_{n}\\}_{n=1}^{N}$ by projecting pixels to the mesh surface. Background pixels that do not intersect any surface are discarded. Following [40], we use a multi-resolution hash encoding based on instantNGP [51] along with a shallow MLP $f_{\\phi}$ to parameterize a function from 3D spatial coordinates to RGB values for each sample $rgb = f_{\\phi}(hash(xyz))$ . We then distill multi-view images $\\{\\xi_{n}\\}_{n=1}^{N}$ into this parametric function via optimization. Since our goal is to export $\\phi$ into a texture map, we do not use any view-dependent parameterization. To reconcile inconsistencies, we use both a standard $L2$ loss and a VGG-based perceptual loss, applied between $f_{\\phi}$ and $\\{\\xi_{n}\\}_{n=1}^{N}$ , to train $\\phi$ . We use Adam with a learning rate of 0.01, and optimization of $\\phi$ converges within 100 iterations. After optimization, we compute the spatial coordinate of the centers of each pixel in a high-resolution texture map, and query $f_{\\phi}$ to predict RGB values for the texture map.", + "bbox": [ + 496, + 151, + 893, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.4 Geometry Processing, Camera, and Multi-resolution Refinement", + "text_level": 1, + "bbox": [ + 498, + 739, + 890, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We normalize $\\mathcal{M}$ such that it fits inside a cube of side length 1, and center it at the origin. Perspective cameras are placed facing the origin, and their FOV is adjusted to fit the object. Detailed parameters can be found in the appendix. As the diffusion model relies on context captured in $\\mathbf{x}_i$ , to perform denoising, camera views that are too small w.r.t the size of the object often result in content drift - the texture in distant areas of the mesh can have content that is seman", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4174", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tically inconsistent. This can be seen as our version of the Janus face problem known to Dreamfusion and similar approaches [40, 59, 84]. Our solution is to perform two rounds of TexFusion at different resolution scales to obtain high-resolution textures while retaining semantic consistency.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we first run TexFusion using cameras that cover the entire object, and a low-resolution latent texture map that is suitable for these cameras. We do not run view distillation in this first round as we are only interested in the latent texture map. We denote the denoised latent texture map from this step as $\\mathbf{z}_{0,lr}$ . We then use a second set of cameras with a narrower field of view; these cameras are also more numerous to still cover the full object. We determine a new texture map resolution using the square of the ratio of the tangent of the old FOV over the new FOV - corresponding to the relative change in surface area covered by the camera before and after the change in FOV. We then up-sample $\\mathbf{z}_{0,lr}$ with nearest neighbor filtering to this higher resolution, and stochastically encode it to a partially noised state (e.g. $T = 500$ in the diffusion model time schedule). The second round of TexFusion uses these new cameras, and initializes SIMS with the partially noised latent texture map. The multi-view images produced by the second round of SIMS is used to produce the final output texture map via neural color field distillation.", + "bbox": [ + 75, + 167, + 472, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 482, + 209, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We apply TexFusion on various geometries and text prompts to evaluate its ability to produce high quality, natural, and 3D consistent textures. We focus our experimental comparisons on TEXTure [62], a text-driven texture generation method that also uses SD2-depth. We choose TEXTure as the baseline because (1) it represents the current state-of-the-art for language-conditioned texture generation, (2) it uses the same SD2-depth model, which can be prompted for a wide variety of content, and (3) it is concurrent to our work. In the supplementary materials, we further compare TexFusion to SDS-based text-driven texture distillation [48, 59], leveraging the Stable Diffusion model and show that TexFusion achieves superior results in terms of quality and speed.", + "bbox": [ + 75, + 508, + 470, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset We collect 35 meshes of various content, and write 2-4 prompts for each mesh. In total, we evaluate each texture synthesis method on 86 mesh-text pairs. More details of this dataset are in the supplementary materials.", + "bbox": [ + 75, + 731, + 468, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 76, + 801, + 305, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We visualize textures produced by TexFusion on multiple geometries and prompts, and compare to state-of-the-art baselines in Figure 6. We render each object in 3 surrounding views to allow better examination of 3D consistency of the produced texture. We use Blender's Cycles renderer", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with a studio light-setup. Textures are applied as base color to a diffuse material. Additional visualizations, including videos showing 360 pans of all objects presented in the paper, and renders of our results using only texture color, can be found in the supplementary materials.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In terms of local 3D consistency (consistency in neighborhoods on the surface of the mesh), textures produced by TexFusion are locally consistent - there are no visible seam lines or stitching artifacts. In contrast, we often find severe artifacts when viewing the top and back sides of TEXTure's outputs. These artifacts are most noticeable when a clean color is expected, such as when texturing vehicles. In terms of global consistency (semantic coherency of the entire texture, e.g. exactly 2 eyes and 1 nose to a face), TEXTure performs poorly and suffers from problems similar to Dream-Fusion's Janus face problem [59]: as the 2D diffusion model captures context of the object through its own camera view, it is not aware of the appearance of the opposite side of the object. This problem is ameliorated in TexFusion due to frequent communication between views in SIMS.", + "bbox": [ + 496, + 167, + 892, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "There are noticeably more baked specular highlights and shadows in textures generated by TEXTure. These effects are physically inaccurate as they conflict with the lighting effects simulated by the renderer. In contrast, TexFusion produces textures that are smoothly illuminated. We hypothesize that this is due to interlacing aggregations in SIMS, which removes view-dependent effects as they arise in the sampling process. We provide additional visualizations of TexFusion in Fig. 5 to showcase how text prompting is effective at controlling the generation of textures, thereby producing varying appearances using the same mesh.", + "bbox": [ + 496, + 395, + 892, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Runtime TexFusion takes approximately 3 minutes on a machine with a single GPU to sample one texture. We are slightly faster than TEXTure (5 min.) [62], since we only need to optimize a color field once after sampling is", + "bbox": [ + 496, + 585, + 890, + 647 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2c4ddf721294c7501817252ac0d1318e1bbd0ee17be264ff308a64f8ffa5ed5.jpg", + "image_caption": [ + "Figure 5: More TexFusion text-conditioned texturing results." + ], + "image_footnote": [], + "bbox": [ + 524, + 667, + 861, + 887 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4175", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fae4b983bc70a652edb2ecabf272f2b8cd3230dec047727c9935b6d93110a90d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 93, + 890, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29f8503ae5077a06264fb3f0e19107c2cf68b29efd8d9cd279da2fa35f71d8d2.jpg", + "image_caption": [ + "Figure 6: Visual comparison of textures generated by TEXTure [62] and TexFusion." + ], + "image_footnote": [], + "bbox": [ + 169, + 292, + 815, + 431 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e3b62d547b19e4cfc370a233e49ca2123929e8b08426dc675c15ed019a8b5f86.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID (↓)User study (%)
Natural Color (↑)More Detailed (↑)Less Artifacts (↑)Align with Prompt (↑)
TEXTure79.4724.4265.1231.4043.02
TexFusion59.7875.5834.8868.6056.98
", + "bbox": [ + 81, + 450, + 470, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1: Quantitative and qualitative comparisons between TEX-Ture and TexFusion. FID is computed w.r.t. a set of images synthesized by SD2-depth.", + "bbox": [ + 76, + 532, + 468, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "complete. We are also an order of magnitude faster than methods that rely on SDS loss which report upwards of 40 minutes [40, 48, 59]. Additional runtime comparisons are in the appendix.", + "bbox": [ + 76, + 590, + 468, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 76, + 661, + 316, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It is inherently difficult to quantitatively evaluate the quality of text and geometry-conditioned texture generation as one must take alignment with the conditioning inputs into account. We rely on both automated and human evaluation to gauge the quality of the synthesized textures.", + "bbox": [ + 75, + 684, + 468, + 761 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FID Since SD2-depth can produce high-quality images that match the structure of the conditioning geometry, we sample it to create a proxy ground truth set. Furthermore, as both TEXTure and TexFusion use SD2-depth as 2D prior, samples from it serve as an upper bound to the quality and diversity of the appearance of textures: We render the depth map of all meshes in 8 canonical views, and condition SD2-depth on both these depth maps and our text", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "prompts to generate a ground truth set. We also render textures generated by both methods in these same 8 views, and compute the Fréchet Inception Distance (FID) between images rendered from the textures and the ground truth set. FID measures the overlap between two sets of images based on their similarity in a learned feature space. We white out the background regions in rendered images and ground truth images to focus the comparisons to the textured object. We present the results in Tab. 1. FID scores indicate that textures synthesized by TexFusion are closer in appearance to the ground truth set than those from TEXTure. The absolute FID values are likely high due to the relatively small set of images in each dataset.", + "bbox": [ + 496, + 453, + 893, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User study We conducted a user study to measure the overall quality of the generated textures. We use all mesh and text prompts in our dataset, and render each textured mesh into a video showing the results from a $360^{\\circ}$ rotating view. We present videos from TEXTure and TexFusion side by side in random left-right order, to avoid bias. The user study was conducted on Amazon Mechanical Turk. Each participant is asked to select the preferred result under four criteria: natural color (non-saturation), details, cleanliness, and alignment with the text prompt. To avoid randomness in the answers, we let 3 participants answer the same question and determine the choice by max-voting, a question screenshot is provided in the supplement. We provide the user study results in Tab. 1. Although not deemed as detailed as TEXTure, our results are overall preferred by the users in terms of better aligning with the provided text prompt, and", + "bbox": [ + 496, + 659, + 892, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4176", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8fabd9c349e734537082503eedec140df2c7aba5d26cf9c1ac6e33a0da769ab6.jpg", + "image_caption": [ + "\"red backpack\"" + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 176, + 159 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ab5e0688b22c38345ddbb3a105cc3367b15da5e8e4c1056f4c9345b053452440.jpg", + "image_caption": [ + "\"Portrait of a humanoid robot, futuristic, science fiction\"" + ], + "image_footnote": [], + "bbox": [ + 189, + 95, + 259, + 156 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0221b34bce87f77084fce0a3f15edb6f11a4b3273ec4371173a5f464fead88ca.jpg", + "image_caption": [ + "\"Medieval celtic Iouse, stone bricks, wooden roof\"" + ], + "image_footnote": [], + "bbox": [ + 287, + 89, + 359, + 155 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/780bba85ba986bbae9197888854e1c8eeb585f8134b439a4048be95443f9a1ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 104, + 450, + 152 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/21f3936efeae468b7e7e07f9cc89d4d70ee5347dd9be49c0c0d51be4d453f95b.jpg", + "image_caption": [ + "\"camper bag, camouflage\"" + ], + "image_footnote": [], + "bbox": [ + 106, + 186, + 176, + 255 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d99f9e40c816035f068e488dd2283d0b354039b1e9ae53e62f151cf6133393f9.jpg", + "image_caption": [ + "\"Brown rabbit\"" + ], + "image_footnote": [], + "bbox": [ + 187, + 196, + 267, + 250 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/f1d973a6d898dbf7da172e05271a43ef4ba51b399f52a44b5de39bffa6e3716a.jpg", + "image_caption": [ + "\"Crocodile skin handbag\"", + "\"Blonde girl with green eyes, hair in tied a bun, DSLR portrait photo.\"" + ], + "image_footnote": [], + "bbox": [ + 290, + 189, + 357, + 253 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9d49a98b80f4182b08cfee95b73e69b89ed8554776369f675d90b41f6c204a26.jpg", + "image_caption": [ + "\"Leather lounge chair\"", + "Figure 7: TexFusion + non-stochastic DDIM sampling $(\\sigma_{t_i} = 0)$ . This setting emphasizes textural details, such as the leather (bottom, left), roof shingles (top, right)" + ], + "image_footnote": [], + "bbox": [ + 370, + 195, + 457, + 257 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "having more natural color, and fewer artifacts/flaws.", + "bbox": [ + 76, + 369, + 419, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.3. Improving Texture Details", + "text_level": 1, + "bbox": [ + 76, + 396, + 316, + 412 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Texfusion can produce more texture details with adjusted hyperparameters and diffusion backbones. First, we find that using the non-stochastic version of DDIM $(\\eta = 0)$ adds more materialistic details to the textures on smooth/low-poly geometries. We showcase some examples with particularly large improvements in Fig. 7. Second, we explore the use of ControlNet[98], which can be easily substituted as the LDM backbone for our method without any additional changes. We find that its high-resolution depth conditioning allows TexFusion to capture fine-grained geometric details in the input mesh. In Fig. 8, we further compare TexFusion + ControlNet in \"normal mode\" (apply classifier-free guidance to text prompt only) and ControlNet's \"guess mode\" (apply classifier-free guidance to both text and depth) on meshes with fine geometric details. TexFusion produces high-contrast textures with the appearance of strong lighting under \"guess mode\", and realistic textures with smooth lighting in \"normal mode\". These modifications improve details at a cost of reduced robustness. The non-stochastic DDIM setting may create artifacts when a smooth and clean texture is desired. On the other hand, the increase in depth resolution makes TexFusion+ControlNet susceptible to capturing face boundaries on low-poly meshes. We provide visualizations of these failure cases in Fig. 9. Nonetheless, these modifications offer further dimensions of control that can be explored by practitioners.", + "bbox": [ + 75, + 420, + 472, + 813 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 76, + 829, + 333, + 844 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We presented TexFusion, a novel approach to text-driven texture synthesis for given 3D geometries, using only large-scale text-guided image diffusion models. TexFusion lever", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8131710eb84cee6fdd39c315df4f2a089d0175148a010ede602cde4daaaa5859.jpg", + "image_caption": [ + "TexFusion + ControlNet in \"normal\" mode", + "\"Portrait of greek-egyptian deity hermanubis, lapis skin and gold clothing\"", + "Figure 8: Left: TexFusion + ControlNet in \"normal mode\"; right: TexFusion + ControlNet in \"guess mode\"." + ], + "image_footnote": [], + "bbox": [ + 521, + 112, + 681, + 247 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/342524332e5f027f4c6422333f0c6a22293ac48ccb4a9a180ea6d72f42b3ca4f.jpg", + "image_caption": [ + "TexFusion + ControlNet in \"guess\" mode" + ], + "image_footnote": [], + "bbox": [ + 694, + 111, + 846, + 247 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/18cfcce3440166573610ceeeaf21b6025f2c0152d1abc84a53ec672a0419e388.jpg", + "image_caption": [ + "Deterministic DDIM: Artifacts in smooth areas", + "Figure 9: Failure modes exhibited by (left) TexFusion + non-stochastic DDIM sampling and (right) TexFusion + ControlNet." + ], + "image_footnote": [], + "bbox": [ + 527, + 329, + 697, + 393 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ea750aea5dd5f9ba5149e88a15d5419e000fc683e1514c0df820f9b952159cb8.jpg", + "image_caption": [ + "TexFusion+ControlNet: Face boundaries on low poly meshes" + ], + "image_footnote": [], + "bbox": [ + 715, + 329, + 771, + 392 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4f29b5525b3af72c1dc596a52f30a2b77334ba439db004cbbaa2d781d7e7616c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 329, + 870, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ages the latent diffusion model Stable Diffusion and relies on a new 3D-consistent diffusion model sampling scheme that runs a 2D denoiser network in different views of the object and aggregates the predictions in a latent texture map after each denoising iteration. We find that our method can efficiently generate diverse, high-quality, and globally coherent textures and offers detailed control through text conditioning. Limitations of our approach include a not-yet-ideal sharpness and that texture generation is not real-time. Future work can address these issues; for instance, it would be interesting to leverage the recent literature on faster diffusion model samplers [15, 41, 42, 99].", + "bbox": [ + 496, + 449, + 890, + 630 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Concurrent Works Since Submission Since the time of writing this paper, new works has appeared in the text to texture space [8, 79]. They improve upon TEXTure in aspects such as adding a refinement procedure to automatically fix low quality areas from the initial texturing process [8], and using images sampled from a text-to-image model and per-prompt finetuning to provide stronger conditioning [79]. TexFusion is methodologically distinct from these methods which are based-on TEXTure. Improvements proposed in these works could be combined with TexFusion in future work.", + "bbox": [ + 496, + 648, + 890, + 815 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements We would like to thank Jun Gao for the helpful discussions during the project. Tianshi Cao acknowledges additional income from Vector Scholarships in Artificial Intelligence, which are not in direct support of this work.", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4177", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In ICML, 2018. 1", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1, 2", + "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 3", + "[4] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh Susskind. Gaudi: A neural architect for immersive 3d scene generation. arXiv preprint arXiv:2207.13751, 2022. 1, 2", + "[5] Hanqun Cao, Cheng Tan, Zhangyang Gao, Guangyong Chen, Pheng-Ann Heng, and Stan Z. Li. A survey on generative diffusion model. arXiv preprint arXiv:2209.02646, 2022. 2", + "[6] Ignacio Castaño. thekla交通大学 In github.com/Thekla/thekla_italas, 2015. 4", + "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2", + "[8] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 9", + "[9] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 100–116. Springer, 2019. 2", + "[10] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3", + "[11] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In The Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[12] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. arXiv preprint arXiv:2209.04747, 2022. 2", + "[13] Congyue Deng, Chiyu \"Max\" Jiang, Charles R. Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, and Dragomir Anguelov. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. arXiv" + ], + "bbox": [ + 86, + 114, + 470, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "preprint arXiv:2212.03267, 2022. 2", + "[14] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In Advances in Neural Information Processing Systems, 2021. 2", + "[15] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. GENIE: Higher-Order Denoising Diffusion Solvers. In Advances in Neural Information Processing Systems, 2022. 9", + "[16] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Score-based generative modeling with critically-damped Langevin diffusion. In International Conference on Learning Representations (ICLR), 2022. 2", + "[17] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2", + "[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems. 1, 2", + "[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, 2020. 3", + "[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. arXiv preprint arXiv:2106.15282, 2021. 2", + "[22] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2, 3", + "[23] Wenlong Huang, Brian Lai, Weijian Xu, and Zhuowen Tu. 3d volumetric modeling with introspective neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33(01), pages 8481-8488, 2019. 1", + "[24] Aapo Hyvarinen. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6:695-709, 2005. 3", + "[25] Moritz Ibing, Isaak Lim, and Leif P. Kobbelt. 3d shape generation with grid-based implicit functions. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1", + "[26] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022. 3", + "[27] Nikolai Kalischek, Torben Peters, Jan D Wegner, and Konrad Schindler. Tetrahedral diffusion models for 3d shape generation. arXiv preprint arXiv:2211.13220, 2022. 1, 2", + "[28] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, 2022. 3", + "[29] Carson Katri. Dream textures. In github.com/carson-katri/dream-textures, 2022. 4", + "[30] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming" + ], + "bbox": [ + 509, + 93, + 890, + 887 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4178", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Song. Denoising diffusion restoration models. arXiv preprint arXiv:2201.11793, 2022. 2", + "[31] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Popa Tiberiu. Clip-mesh: Generating textured meshes from text using pretrained image-text models. December 2022. 3", + "[32] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. 2020. 1", + "[33] Wei-Jan Ko, Hui-Yu Huang, Yu-Liang Kuo, Chen-Yi Chiu, Li-Heng Wang, and Wei-Chen Chiu. Rpg: Learning recursive point cloud generation. arXiv preprint arXiv:2105.14322, 2021. 1", + "[34] Johannes Kopf, Chi-Wing Fu, Daniel Cohen-Or, Oliver Deussen, Dani Lischinski, and Tien-Tsin Wong. Solid texture synthesis from 2d exemplars. ACM Trans. Graph., 26(3):2-es, jul 2007. 2", + "[35] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5", + "[36] Sylvain Lefebvre and Hugues Hoppe. Appearance-space texture synthesis. ACM Trans. Graph., 25(3):541-548, jul 2006. 2", + "[37] Chun-Liang Li, Manzil Zaheer, Yang Zhang, Barnabas Poczos, and Ruslan Salakhutdinov. Point cloud gan. arXiv preprint arXiv:1810.05795, 2018. 1", + "[38] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2", + "[39] Ruihui Li, Xianzhi Li, Ke-Hei Hui, and Chi-Wing Fu. SPGAN:sphere-guided 3d shape generation and manipulation. ACM Transactions on Graphics (Proc. SIGGRAPH), 40(4), 2021. 1", + "[40] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 1, 2, 6, 7, 8", + "[41] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 9", + "[42] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv:2206.00927, 2022. 9", + "[43] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. arXiv preprint arXiv:2201.09865, 2022. 2", + "[44] A. Luo, T. Li, W. Zhang, and T. Lee. Surfgen: Adversarial 3d shape synthesis with explicit surface discriminators. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1", + "[45] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern" + ], + "bbox": [ + 86, + 92, + 470, + 890 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Recognition (CVPR), 2021. 1, 2", + "[46] Siwei Lyu. Interpretation and generalization of score matching. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, UAI '09, page 359-366, Arlington, Virginia, USA, 2009. AUAI Press. 3", + "[47] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2", + "[48] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. 1, 2, 4, 7, 8", + "[49] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2, 3", + "[50] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 1", + "[51] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 6", + "[52] Gimin Nam, Mariem Khlifi, Andrew Rodriguez, Alberto Tono, Linqi Zhou, and Paul Guerrero. 3d-ldm: Neural implicit 3d shape generation with latent diffusion models. arXiv preprint arXiv:2212.00842, 2022. 1, 2", + "[53] Charlie Nash, Yaroslav Ganin, S. M. Ali Eslami, and Peter W. Battaglia. *Polygen: An autoregressive generative model of 3d meshes.* ICML, 2020. 1", + "[54] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2", + "[55] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1, 2", + "[56] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, 2021. 2", + "[57] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 2", + "[58] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1", + "[59] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 1, 2, 7, 8" + ], + "bbox": [ + 509, + 92, + 890, + 890 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4179", + "bbox": [ + 482, + 925, + 514, + 936 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR, 2021. 3", + "[61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2", + "[62] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 3, 4, 7, 8", + "[63] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 1, 2, 3, 4, 5", + "[64] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image dissusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2", + "[65] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. arXiv preprint arXiv:2111.05826, 2021. 2", + "[66] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2", + "[67] Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. arXiv preprint arXiv:2104.07636, 2021. 2", + "[68] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations (ICLR), 2022. 3", + "[69] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. arXiv preprint arXiv:2110.02624, 2021. 1", + "[70] Hiroshi Sasaki, Chris G. Willcocks, and Toby P. Breckon. UNIT-DDPM: Unpaired image translation with denoising diffusion probabilistic models. arXiv preprint arXiv:2104.05358, 2021. 2", + "[71] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3859-3868, 2019. 1", + "[72] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary" + ], + "bbox": [ + 86, + 90, + 470, + 893 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. arXiv preprint arXiv:2211.16677, 2022. 1, 2", + "[73] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In Shai Avidan, Gabriel J. Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, volume 13663 of Lecture Notes in Computer Science, pages 72-88. Springer, 2022. 2", + "[74] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015. 3", + "[75] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3", + "[76] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3", + "[77] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. arXiv preprint arXiv:2203.08382, 2022. 2", + "[78] Yongbin Sun, Yue Wang, Ziwei Liu, Joshua E Siegel, and Sanjay E Sarma. Pointgrow: Autoregressively learned point cloud generation with self-attention. In Winter Conference on Applications of Computer Vision, 2020. 1", + "[79] Zhibin Tang and Tiantong He. Text-guided high-definition consistency texture model. ArXiv, abs/2305.05901, 2023. 9", + "[80] Greg Turk. Texture synthesis on surfaces. SIGGRAPH '01, page 347-354, New York, NY, USA, 2001. Association for Computing Machinery. 2", + "[81] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In Neural Information Processing Systems (NeurIPS), 2021. 2", + "[82] Diego Valsesia, Giulia Fracastoro, and Enrico Magli. Learning localized generative models for 3d point clouds via graph convolution. In International Conference on Learning Representations (ICLR) 2019, 2019. 1", + "[83] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 23(7):1661-1674, 2011. 3", + "[84] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022. 1, 2, 7", + "[85] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jian-min Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135, 2022. 1, 2", + "[86] Li-Yi Wei and Marc Levoy. Texture synthesis over arbitrary manifold surfaces. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '01, page 355-360, New York, NY, USA, 2001. Association for Computing Machinery. 2" + ], + "bbox": [ + 509, + 92, + 890, + 892 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "4180", + "bbox": [ + 482, + 925, + 514, + 936 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[87] Cheng Wen, Baosheng Yu, and Dacheng Tao. Learning progressive point embeddings for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10266-10275, 2021. 1", + "[88] Li-Yi Wie, Sylvain Lefebvre, Vivek Kwatra, and Greg Turk. State of the Art in Example-based Texture Synthesis. In M. Pauly and G. Greiner, editors, Eurographics 2009 - State of the Art Reports. The Eurographics Association, 2009. 2", + "[89] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 1", + "[90] Fanbo Xiang, Zexiang Xu, Milos Hasan, Yannick Hold-Geoffroy, Kalyan Sunkavalli, and Hao Su. Neutex: Neural texture mapping for volumetric neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7119-7128, 2021. 2", + "[91] Jianwen Xie, Yifei Xu, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Zhu Song-Chun, and Ying Nian Wu. Generative pointnet: Deep energy-based learning on unordered point sets for 3d generation, reconstruction and classification. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1", + "[92] Jianwen Xie, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Zhu Song-Chun, and Ying Nian Wu. Learning descriptor networks for 3d shape synthesis and analysis. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018.", + "[93] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. PointFlow: 3D point cloud generation with continuous normalizing flows. 2019. 1", + "[94] Ling Yang, Zhilong Zhang, and Shenda Hong. Diffusion models: A comprehensive survey of methods and applications. arXiv preprint arXiv:2209.00796, 2022. 2", + "[95] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations. In Proceedings of International Conference on Computer Vision (ICCV), 2021. 1", + "[96] Jonathan Young. xatlas. In github.com/jpcy/xatlas, 2016. 4", + "[97] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2", + "[98] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 9", + "[99] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv:2204.13902, 2022. 9", + "[100] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yi-nan Zhang, Antonio Torralba, and Sanja Fidler. Image {gan}s meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In International Conference on Learning Representations, 2021. 1", + "[101] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation" + ], + "bbox": [ + 78, + 90, + 470, + 891 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2", + "bbox": [ + 539, + 90, + 890, + 133 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "4181", + "bbox": [ + 482, + 925, + 513, + 936 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/698eb5a8-8d74-4b19-b9a7-1fe068b4d1fd_model.json b/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/698eb5a8-8d74-4b19-b9a7-1fe068b4d1fd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d8d5db9d0322fff6b625d09ccd14bec11c0a87d2 --- /dev/null +++ b/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/698eb5a8-8d74-4b19-b9a7-1fe068b4d1fd_model.json @@ -0,0 +1,3196 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.142, + 0.09, + 0.83, + 0.134 + ], + "angle": 0, + "content": "TexFusion: Synthesizing 3D Textures with Text-Guided Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.161, + 0.233, + 0.178 + ], + "angle": 0, + "content": "Tianshi Cao \\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.162, + 0.386, + 0.18 + ], + "angle": 0, + "content": "Karsten Kreis" + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.162, + 0.543, + 0.18 + ], + "angle": 0, + "content": "Sanja Fidler\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.574, + 0.162, + 0.714, + 0.181 + ], + "angle": 0, + "content": "Nicholas Sharp\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.745, + 0.162, + 0.869, + 0.18 + ], + "angle": 0, + "content": "Kangxue Yin\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.181, + 0.381, + 0.197 + ], + "angle": 0, + "content": "\\(^{1}\\) NVIDIA," + }, + { + "type": "text", + "bbox": [ + 0.396, + 0.182, + 0.54, + 0.197 + ], + "angle": 0, + "content": "\\(^{2}\\) University of Toronto," + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.182, + 0.659, + 0.196 + ], + "angle": 0, + "content": "3 Vector Institute" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.201, + 0.728, + 0.216 + ], + "angle": 0, + "content": "{tianshic, kkreis, sfidler, nsharp, kangxuey}@nvidia.com" + }, + { + "type": "image", + "bbox": [ + 0.175, + 0.248, + 0.332, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.345, + 0.347, + 0.363 + ], + "angle": 0, + "content": "\"White humanoid robot, movie poster, main character of a science fiction movie\"" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.26, + 0.47, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.346, + 0.48, + 0.356 + ], + "angle": 0, + "content": "\"Crocodile skin handbag\"" + }, + { + "type": "image", + "bbox": [ + 0.471, + 0.26, + 0.562, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.26, + 0.658, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.274, + 0.798, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.345, + 0.687, + 0.353 + ], + "angle": 0, + "content": "\"White Mercedes Benz SUV\"" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.363, + 0.351, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.45, + 0.331, + 0.468 + ], + "angle": 0, + "content": "\"Oil painting of a bald, middle aged banker with pointed moustache\"" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.364, + 0.45, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.458, + 0.551, + 0.468 + ], + "angle": 0, + "content": "\"Portrait photo of Abraham Lincoln, full color\"" + }, + { + "type": "image", + "bbox": [ + 0.469, + 0.359, + 0.56, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.356, + 0.671, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.357, + 0.747, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.457, + 0.737, + 0.466 + ], + "angle": 0, + "content": "\"Railroad worker wearing high-vis vest\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.481, + 0.678, + 0.495 + ], + "angle": 0, + "content": "Figure 1: Text-conditioned 3D texturing results with TexFusion." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.523, + 0.314, + 0.538 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.547, + 0.473, + 0.88 + ], + "angle": 0, + "content": "We present TexFusion (Texture Diffusion), a new method to synthesize textures for given 3D geometries, using large-scale text-guided image diffusion models. In contrast to recent works that leverage 2D text-to-image diffusion models to distill 3D objects using a slow and fragile optimization process, TexFusion introduces a new 3D-consistent generation technique specifically designed for texture synthesis that employs regular diffusion model sampling on different 2D rendered views. Specifically, we leverage latent diffusion models, apply the diffusion model's denoiser on a set of 2D renders of the 3D object, and aggregate the different denoising predictions on a shared latent texture map. Final output RGB textures are produced by optimizing an intermediate neural color field on the decodings of 2D renders of the latent texture. We thoroughly validate TexFusion and show that we can efficiently generate diverse, high quality and globally coherent textures. We achieve state-of-the-art text-guided texture synthesis performance using only image diffusion models, while avoiding the pitfalls of previous distillation-based methods. The text-conditioning offers detailed control and we also do not rely on any ground truth 3D textures for training. This makes our method versatile" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.509, + 0.892, + 0.57 + ], + "angle": 0, + "content": "and applicable to a broad range of geometry and texture types. We hope that TexFusion will advance AI-based texturing of 3D assets for applications in virtual reality, game design, simulation, and more." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.578, + 0.633, + 0.593 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.853 + ], + "angle": 0, + "content": "In the past decade, deep learning-based 3D object generation has been studied extensively [1, 4, 18, 23, 25, 27, 32, 33, 37, 39, 44, 45, 50, 52, 53, 55, 58, 69, 71, 72, 78, 82, 85, 87, 89, 91-93, 95, 97, 100, 101], due to the demand for high-quality 3D assets in 3D applications such as VR/AR, simulation, digital twins, etc. While many prior works on 3D synthesis focus on the geometric components of the assets, textures are studied much less, despite the fact that they are important components of realistic 3D assets which assign colors and materials to meshes to make the rendering vivid. Recent advances in text-conditioned image diffusion models trained on internet-scale data [2, 54, 61, 63, 66] have unlocked the capability to generate images with stunning visual detail and practically unlimited diversity. These high-performance diffusion models have also been used as image priors to synthesize 3D objects with textures using textual guidance [40, 48, 59, 84]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we aim to perform text-driven high-quality 3D texture synthesis for given meshes, by leveraging the information about the appearance of textures carried by the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.216, + 0.9 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.926, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4169" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "image prior of a pre-trained text-to-image diffusion model. The main workhorse in current text-to-3D methods that leverage 2D image diffusion models is Score Distillation Sampling (SDS) [59]. SDS is used to distill, or optimize, a 3D representation such that its renders are encouraged to be high-likelihood under the image prior. Methods utilizing SDS also share two common limitations in that: 1). a high classifier-free guidance weight [22] is required for the optimization to converge, resulting in high color saturation and low generation diversity; 2). a lengthy optimization process is needed for every sample." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.473, + 0.575 + ], + "angle": 0, + "content": "To address the above issues, we present Texture Diffusion, or TexFusion for short. TexFusion is a sampler for sampling surface textures from image diffusion models. Specifically, we use latent diffusion models that efficiently autoencode images into a latent space and generate images in that space [63, 81]. TexFusion leverages latent diffusion trajectories in multiple object views, encoded by a shared latent texture map. Rendering of the shared latent texture map are provided as input to the denoiser of the latent diffusion model, and the output of every denoising step is projected back to the shared texture map in a 3D-consistent manner. To transform the generated latent textures into RGB textures, we optimize a neural color field on the outputs of the latent diffusion model's decoder applied to different views of the object. We use the publicly available latent diffusion model Stable-Diffusion with depth conditioning [63] (SD2-depth) as our diffusion backbone. Compared to methods relying on SDS, TexFusion produces textures with more natural tone, stronger view consistency, and is significantly faster to sample (3 minutes vs. 30 minutes reported by previous works)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.473, + 0.775 + ], + "angle": 0, + "content": "We qualitatively and quantitatively validate TexFusion on various texture generation tasks. We find that TexFusion generates high quality, globally coherent and detailed textures that are well-aligned with the text prompts used for conditioning (e.g. Fig. 1). Since we leverage powerful text-to-image diffusion models for texture sampling, we can generate highly diverse textures and are not limited to single categories or restricted by explicit 3D textures as training data, which are limitations of previous works [7, 11, 18, 57, 73]. In summary, our main contribution is a novel method for 3D texture generation from 2D image diffusion models, that is view-consistent, avoids over-saturation and achieves state-of-the-art text-driven texture synthesis performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.788, + 0.22, + 0.804 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Classic Computer Graphics Techniques Early work on texture generation focused on tiling exemplar patterns across a surface, often with an explicit direction field for local orientation [34, 36, 80, 86]. See [88] for a survey. This research established the value of texture image representations and the challenges of global coherence, both central" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "to this work. However, modern learning-based priors have proven necessary to go beyond simple patterns and synthesize plausible shape-specific texture details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.151, + 0.895, + 0.41 + ], + "angle": 0, + "content": "Texture Synthesis with 3D Priors Textures are defined on the meshed surface of 3D objects, which is an irregular representation. To enable 2D texture generation on 3D meshes, AUV-Net [11] learns an aligned UV space for a set of 3D shapes in a given class, mapping 3D texture synthesis to a 2D domain. Texturify [73] trains a 3D StyleGAN in the quad-parameterized surface domain on a set of textured 3D shapes in a given class. Different from AUV-NET or Texturify, which embed the geometric prior into a UV map or mesh parameterization, EG3D [7] and GET3D [18] directly train 3D StyleGANs to generate geometry and texture jointly, where the textures are implicit texture fields [57]. Other works also represent 3D texture by vertex colors [49], voxels [9], cube mapping [90], etc. In contrast to TexFusion, these works mostly don't offer text conditioning, often work only on single object categories or require textured 3D shapes for training, which limits their broad applicability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.895, + 0.634 + ], + "angle": 0, + "content": "Diffusion Models Our approach directly builds on diffusion models [5, 12, 94], which have recently emerged as new state-of-the-art generative models. In particular, they have demonstrated outstanding performance in image generation [14-16, 21, 56, 81], outperforming GANs, and led to breakthroughs in text-to-image synthesis [2, 54, 61, 63, 66]. They have also been successfully used for a variety of image editing and processing tasks [17, 19, 30, 38, 43, 47, 64, 65, 67, 70, 77]. Moreover 3D object generation has been addressed with diffusion models, too, for instance leveraging point clouds [45, 55, 97, 101], meshes [27], or neural fields [4, 52, 72, 85] as 3D representations. However, these works focus on geometry generation and do not specifically tackle 3D texture synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Distilling 3D Objects from 2D Image Diffusion Models Recently, large-scale 2D text-to-image diffusion models have been leveraged to distill individual 3D objects as neural radiance fields using Score Distillation Sampling (SDS) [13, 40, 48, 59, 84]. In SDS, the radiance field is rendered from different directions into 2D images and it is optimized such that each render has a high probability under the text-to-image diffusion model while conditioning on a text prompt. DreamFusion [59] pioneered the approach, Magic3D [40] proposed a coarse-to-fine strategy improving quality, and Latent-NeRF [48] performs distillation in latent space leveraging a latent diffusion model [63]. These approaches do not specifically target texture generation, which is the focus of this work. More deeply, a crucial drawback of this line of work is that SDS typically requires strong guidance [22] to condition the diffusion model, which can hurt quality and diversity. Moreover, SDS's" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.518, + 0.938 + ], + "angle": 0, + "content": "4170" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.079, + 0.677, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.082, + 0.894, + 0.097 + ], + "angle": 0, + "content": "Distill with Neural Color Field" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.098, + 0.894, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.276, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Figure 2: Overview of TexFusion. TexFusion takes a text prompt and mesh geometry as input and produces a UV parameterized texture image matching the prompt and mesh. Key to TexFusion is the Sequential Interlaced Multiview Sampler (SIMS) - SIMS performs denoising diffusion iterations in multiple camera views, yet the trajectories are aggregated through a latent texture map after every denoising step. SIMS produces a set of 3D consistent latent images (TexFusion uses Stable Diffusion [63] as text-to-image diffusion backbone), which are decoded and fused into a texture map via optimizing an intermediate neural color field." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.471, + 0.45 + ], + "angle": 0, + "content": "iterative optimization process makes synthesis very slow. In contrast, our approach avoids SDS entirely and leverages regular diffusion model sampling in a new, 3D-consistent manner. Earlier works also leverage CLIP [60] for 3D object or texture distillation [10, 26, 31, 49], but this performs usually worse than using diffusion models instead." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.473, + 0.784 + ], + "angle": 0, + "content": "Concurrent Work Concurrently with this work, TEX-Ture [62] proposes a related approach. Like this work, TEX-Ture performs multiview denoising on a texture map representation. However, TEX-Ture runs an entire generative denoising process in each camera view in sequence, conditioning on the previous views and projecting to the texture map only after full denoising. In contrast, in TexFusion we propose to interleave texture aggregation with denoising steps in different camera views, simultaneously generating the entire output. This insight significantly reduces view inconsistencies and improves quality in TexFusion compared to TEX-Ture, as validated in Sec. 5. MultiDiffusion [3] concurrently introduces a method for panorama generation and other controlled image generation tasks, leveraging relatively lower-resolution image diffusion models. Algorithmically, this approach of aggregating different denoising predictions from different image crops is closely related to TexFusion's aggregation from different camera views. However, MultiDiffusion only tackles image synthesis, and is not concerned with any 3D or texture generation at all." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.799, + 0.206, + 0.816 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Diffusion Models Diffusion models [20, 74, 76] model a data distribution \\( p_{\\mathrm{data}}(\\mathbf{x}) \\) via iterative denoising, and are trained with denoising score matching [20, 24, 46, 74, 76, 83]. Given samples \\( \\mathbf{x} \\sim p_{\\mathrm{data}} \\) and \\( \\epsilon \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I}) \\), a denoiser model \\( \\epsilon_{\\theta} \\) parameterized with learnable parameters" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.358, + 0.892, + 0.39 + ], + "angle": 0, + "content": "\\(\\theta\\) receives diffused inputs \\(\\mathbf{x}_t(\\epsilon, t, \\mathbf{x})\\) and is optimized by minimizing the denoising score matching objective" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.397, + 0.892, + 0.417 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\mathbf {x} \\sim p _ {\\text {d a t a}}, t \\sim p _ {t}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}; \\mathbf {c}, t) \\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.893, + 0.665 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}\\) is optional conditioning information, such as a text prompt, and \\(p_t\\) is a uniform distribution over the diffusion time \\(t\\). The model effectively learns to predict the noise \\(\\epsilon\\) that was used to perturb the data (other formulations are possible [28, 68]). Letting \\(\\alpha_{t}\\) define a noise schedule, parameterized via a diffusion-time \\(t\\), we construct \\(\\mathbf{x}_t\\) as \\(\\mathbf{x}_t = \\sqrt{\\alpha_t}\\mathbf{x} + \\sqrt{1 - \\alpha_t}\\epsilon\\), \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\); this particular formulation corresponds to a variance-preserving schedule [76]. The forward diffusion as well as the reverse generation process in diffusion models can be described in a continuous-time framework [76], but in practice a fixed discretization can be used [20]. The maximum diffusion time is generally chosen such that the input data is entirely perturbed into Gaussian random noise and an iterative generative denoising process can be initialized from such Gaussian noise to synthesize novel data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.892, + 0.741 + ], + "angle": 0, + "content": "Classifier-free guidance [22] can be used for improved conditioning. By randomly dropping out the conditioning c during training, we can learn both a conditional and an unconditional model at the same time, and their predictions can be combined to achieve stronger conditioning." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.741, + 0.892, + 0.773 + ], + "angle": 0, + "content": "We perform iterative diffusion model sampling via the Denoising Diffusion Implicit Models (DDIM) scheme [75]:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.793, + 0.892, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {x} _ {i - 1} = \\sqrt {\\alpha_ {i - 1}} \\left(\\frac {\\boldsymbol {x} _ {i} - \\sqrt {1 - \\alpha_ {i}} \\boldsymbol {\\epsilon} _ {\\theta} ^ {(t _ {i})} (\\boldsymbol {x} _ {i})}{\\sqrt {\\alpha_ {i}}}\\right) \\tag {2} \\\\ + \\sqrt {1 - \\alpha_ {i - 1} - \\sigma_ {t _ {i}} ^ {2}} \\cdot \\pmb {\\epsilon} _ {\\theta} ^ {(t _ {i})} (\\pmb {x} _ {i}) + \\sigma_ {t _ {i}} \\pmb {\\epsilon} _ {t _ {i}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "with \\(\\epsilon_{t_i}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) and \\(\\sigma_{t_i}\\) is a variance hyperparameter. We express obtaining \\(\\pmb{x}_{i - 1}\\) via DDIM sampling as \\(x_{i - 1}\\sim\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.514, + 0.937 + ], + "angle": 0, + "content": "4171" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.09, + 0.427, + 0.109 + ], + "angle": 0, + "content": "\\(\\pmb{f}_{\\theta}^{(t_i)}(\\pmb{x}_{i-1}|\\pmb{x}_i)\\). See Supp. Material for more details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.128, + 0.473, + 0.281 + ], + "angle": 0, + "content": "Latent Diffusion Models (LDMs) and Stable Diffusion Instead of directly operating on pixels, LDMs [63] utilize an encoder \\(\\mathcal{E}\\) and a decoder \\(\\mathcal{D}\\) for translation between images \\(\\xi\\) and latents \\(\\mathbf{x} \\in X\\) of a lower spatial dimension. The diffusion process is then defined over the distribution of \\(X\\). Stable Diffusion is an LDM trained on the LAION-5B image-text dataset. In addition to text-conditioning, Stable Diffusion 2.0 permits depth conditioning with a depth map \\(D\\) (SD2-depth). This allows detailed control of the shape and configuration of objects in its synthesized images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.301, + 0.47, + 0.481 + ], + "angle": 0, + "content": "**Rendering and Geometry Representation** In principle, our method applies to any geometry representation for which a textured surface can be rendered; in practice, our experiments use a surface mesh \\(\\mathcal{M} = (\\mathcal{V},\\mathcal{F})\\), with vertices \\(\\mathcal{V} = \\{v_i\\}\\), \\(v_{i}\\in \\mathbb{R}^{3}\\) and triangular faces \\(\\mathcal{F} = \\{f_i\\}\\) where each \\(f_{i}\\) is a triplet of vertices. Textures are defined in 2D image space in an injective UV parameterization of \\(\\mathcal{M}\\), \\(UV:p\\in \\mathcal{M}\\mapsto (u,v)\\in [0,1]^2\\). If needed, this parameterization can be automatically constructed via tools such as XATLAS [6, 96]. We encode textures as multi-channel images discretized at pixels in UV space \\(\\mathbf{z}\\in \\mathbb{R}^{(H\\times W,C)}\\), notationally collapsing the spatial dimension for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.471, + 0.633 + ], + "angle": 0, + "content": "We denote the rendering function as \\(\\mathcal{R}(\\mathbf{z};\\mathcal{M},C):\\mathbf{z}\\mapsto\\) \\(\\mathbf{x}\\) \\(\\mathbf{x}\\in \\mathbb{R}^{(h\\times w,C)}\\), which takes as input a mesh \\(\\mathcal{M}\\), camera \\(C\\), and texture \\(\\mathbf{z}\\), and produces as output a rendered image. The inverse of this function \\(\\mathcal{R}^{-1}(\\mathbf{x};\\mathcal{M},C):\\mathbf{x}\\mapsto \\mathbf{z}\\) projects values from camera image-space onto the UV texture map. Notationally, we often omit the dependence on \\(\\mathcal{M}\\) and \\(C\\) for brevity. In this work, we do not model any lighting or shading effects, images are formed by directly projecting textures into screen space (and then decoding, for latent textures)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.47, + 0.667 + ], + "angle": 0, + "content": "4. Texture Sampling with 2D Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Given a mesh geometry \\(\\mathcal{M}\\), a text prompt \\(y\\), and a geometry (depth) conditioned image diffusion model \\(\\theta\\), how could one sample a complete surface texture? Assuming access to the rendering \\(\\mathcal{R}\\) and inverse rendering \\(\\mathcal{R}^{-1}\\) functions defined above, perhaps the most naive approach is to compute a set of \\(\\{C_1,..,C_N\\}\\) camera views that envelopes the surface, render the depth map \\(d_{n}\\) in each view, sample images from the image diffusion model with depth conditioning, and then back-project these images to the mesh surface (e.g. as done in [29]). However, image diffusion models in each view have no information about the generated results in other views, thus there is no coherency in the contents generated in each view. As an alternative, one may define a canonical order of camera views, and autoregressively condition image sampling in the subsequent views on" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.227 + ], + "angle": 0, + "content": "previously sampled regions (as done in [48, 62]). However, for most geometries of interest (i.e. not a plane), a single camera can not observe the entirety of the geometry. Consequently, images synthesized early in the sequence could produce errors that are not reconcilable with the geometry that is observed later in the sequence (see Fig. 3). Thus, it is desirable for the image sampling distribution \\( p(\\cdot | d_i, y) \\) in each camera view to be conditioned on that of every other camera view." + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.236, + 0.859, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.339, + 0.895, + 0.424 + ], + "angle": 0, + "content": "Figure 3: Illustration of irreconcilable mistakes in early views impacting denoised results in later views (images sampled from TEXTure). While the highlighted area appears natural in the first view, it does not match the geometry when viewed from a different angle, thereby creating poorly denoised results when the second view is inpainted." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.436, + 0.854, + 0.453 + ], + "angle": 0, + "content": "4.1. Sequential Interlaced Multiview Sampler" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.895, + 0.746 + ], + "angle": 0, + "content": "Leveraging the sequential nature of the denoising process, we can interlace the synchronization of content across views with the denoising steps within each view to achieve coherency over the entire shape. Suppose that at step \\(i\\) of the denoising process, we have a set of partially denoised images \\(\\{\\mathbf{x}_{i,n}\\}_{n = 1}^{N} = \\{\\mathbf{x}_{i,1}\\dots \\mathbf{x}_{i,N}\\}\\). Our goal is to sample \\(\\{\\mathbf{x}_{i - 1,n}\\}_{n = 1}^{N} = \\mathbf{x}_{i - 1,1}\\dots \\mathbf{x}_{i - 1,N}\\) that is 3D-consistent, i.e., two pixels in \\(\\mathbf{x}_{i - 1,a},\\mathbf{x}_{i - 1,b}\\) that project to the same point in 3D should have the same value. Taking inspiration from the autoregressive modeling literature, we sample the joint distribution via first decomposing it into a product of conditionals, and approximating each term in the product by using the diffusion model to denoise the renders of a dynamically updated latent texture map. Specifically, we first initialize an initial latent texture map \\(\\mathbf{z}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), (\\(T\\) demarks the first step of the diffusion trajectory). Then, suppose that we have a 3D consistent latent texture map \\(\\mathbf{z}_i\\), we decompose the joint distribution as follows (conditioning on depth and prompt omitted for space):" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.755, + 0.892, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p _ {\\theta} \\left(\\left\\{\\mathbf {x} _ {i - 1, j} \\right\\} _ {j = 1} ^ {N} \\mid \\mathbf {z} _ {i}\\right) = p _ {\\theta} \\left(\\mathbf {x} _ {i - 1, 1} \\mid \\mathbf {z} _ {i}\\right) \\times \\\\ \\prod_ {n = 2} ^ {N} p _ {\\theta} \\left(\\mathbf {x} _ {i - 1, n} \\mid \\left\\{\\mathbf {x} _ {i - 1, j} \\right\\} _ {j = 1} ^ {n - 1}, \\mathbf {z} _ {i}\\right) \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.828, + 0.895, + 0.873 + ], + "angle": 0, + "content": "We can compute the first term by first rendering \\(\\mathbf{z}_i\\) into \\(\\mathbf{x}_{i,1}^{\\prime} = \\mathcal{R}(\\mathbf{z}_i;C_1)\\). Eqn. 2 can now be applied to \\(\\mathbf{x}_{i,1}^{\\prime}\\) to draw a latent image at the next time step:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.883, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {i - 1, 1} \\sim f _ {\\theta} ^ {(t _ {i})} \\left(\\mathbf {x} _ {i - 1, 1} \\mid \\mathbf {x} _ {i, 1} ^ {\\prime} = \\mathcal {R} \\left(\\mathbf {z} _ {i}; C _ {1}\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4172" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "Later terms in Eq. 3 additionally depend on the result of previous denoising steps. We again model this dependency through the latent texture map. For each view starting at \\( n = 1 \\), we inverse render \\( \\mathbf{x}_{i - 1,n} \\) into texture space to obtain \\( \\mathbf{z}_{i - 1,n}^{\\prime} = \\mathcal{R}^{-1}(\\mathbf{x}_{i - 1,n}) \\), and update the pixels of \\( \\mathbf{z}_{i - 1,n - 1} \\) that are newly visible in \\( \\mathbf{z}_{i - 1,n}^{\\prime} \\) to obtain \\( \\mathbf{z}_{i - 1,n} \\) (See Sec. 4.2.2 for details). Then, in the \\( n + 1 \\) iteration, since \\( \\mathbf{z}_{i - 1,n} \\) contain regions at two noise scales (unseen regions are at noise scale \\( \\sigma_{i} \\), while visited regions are at noise scale \\( \\sigma_{i - 1} \\)), we add appropriate 3D consistent noise to the visited regions of \\( \\mathbf{z}_{i - 1,n} \\) to match the noise scale of step \\( i \\) before rendering it as input to \\( f_{\\theta} \\). Letting \\( M_{i,n} \\) represent the mask for visited regions and \\( \\epsilon_{i}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\), we can write the sampling procedure for \\( \\mathbf{x}_{i - 1,n} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.317, + 0.469, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {z} _ {i, n} = M _ {i, n} \\odot \\left(\\sqrt {\\frac {\\alpha_ {i - 1}}{\\alpha_ {i}}} \\mathbf {z} _ {i - 1, n - 1} + \\sigma_ {i} \\epsilon_ {i}\\right) \\\\ + \\left(\\mathbf {1} - M _ {i, n}\\right) \\odot \\mathbf {z} _ {i} \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.374, + 0.412, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {i - 1, n} \\sim f _ {\\theta} ^ {(t _ {i})} (\\mathbf {x} _ {i - 1, n} | \\mathbf {x} _ {i, n} ^ {\\prime} = \\mathcal {R} (\\mathbf {z} _ {i, n}; C _ {n}))\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.405, + 0.47, + 0.466 + ], + "angle": 0, + "content": "By iteratively applying Eqn. 5, we obtain a sequence of 3D consistent images \\(\\{\\mathbf{x}_{i - 1,n}\\}_{n = 1}^{N}\\) and a texture map \\(\\mathbf{z}_{i - 1,n}\\) that has been aggregated from these images. We can then decrement \\(i\\) and repeat this process in the next time step." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.603 + ], + "angle": 0, + "content": "We name this approach Sequential Interlaced Multiview Sampler (SIMS). SIMS communicates the denoising direction of previous views to the current view and resolves overlaps between views by the order of aggregation. It ameliorates inconsistent predictions while circumventing performance degradation due to the averaging of latent predictions during parallel aggregation. In the single-view case, SIMS is equivalent to standard DDIM sampling. A complete algorithm for SIMS can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.615, + 0.313, + 0.631 + ], + "angle": 0, + "content": "4.2. The TexFusion Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.47, + 0.73 + ], + "angle": 0, + "content": "In Sec. 4.1, we have presented a generic algorithm for sampling 3D consistent multi-view images and texture maps using 2D diffusion models. We now present a concrete algorithm, TexFusion, that uses SIMS to texture 3D meshes using SD2-depth [63] as the diffusion model. An illustrative overview of TexFusion can be found in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.471, + 0.911 + ], + "angle": 0, + "content": "As SD2-depth is a latent diffusion model, we apply SIMS in the latent space: \\(\\mathbf{x}\\) and \\(\\mathbf{z}\\) represent latent images and latent texture maps respectively, and Section 4.2.3 will describe a final distillation post-process to color space. In this section, we address several challenges specific to using LDMs with SIMS, and detail our design choices to tackle these challenges. We find that a canonical set of cameras works well for most objects, but cameras are tailored to specific objects to improve resolution and ameliorate occlusion. We further illustrate a technique for obtaining high-quality results by operating SIMS in a cascade of increasing resolutions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.092, + 0.826, + 0.107 + ], + "angle": 0, + "content": "4.2.1 (Inverse) Rendering of Latent Textures" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.221 + ], + "angle": 0, + "content": "We use NVDIFFRAST [35] to efficiently render textured meshes via rasterization, as described in Sec. 3. Our implementation sets the rendered image size \\( h = w = 64 \\) to match Stable Diffusion's UNet, with latent vectors of dimension \\( D = 4 \\). The texture image dimensions \\( H \\) and \\( W \\) are chosen based on the surface area of the mesh relative to its diameter (detailed in Sec. 4.2.4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.223, + 0.893, + 0.542 + ], + "angle": 0, + "content": "For each non-background pixel \\(s\\) in a rendered image \\(\\mathbf{x}_{i,..}^{\\prime}\\), rasterization gives a corresponding location on the texture image via the UV map \\((u,v) = UV(p)\\), and we retrieve the value at the nearest texture pixel. This texture value is latent, and no shading or lighting is applied. In other settings, texture data is often rendered with bilinear filtering and mipmaps to improve image quality and reduce aliasing, but in this setting, we found it essential to avoid such techniques. We experimentally ablate texturing approaches in the supplementary. The issue is that early in the diffusion process, i.e. when \\(t_i\\ll 0\\), \\(\\mathbf{z}_i\\) and \\(\\mathbf{x}_{i,..}^{\\prime}\\) are dominated by noise, but interpolation and mipmapping change the variance in the pixels \\(\\mathbf{x}_{i,..}\\), thereby moving \\(\\mathbf{x}_{i,..}^{\\prime}\\) out of the training distribution of \\(\\epsilon_{\\theta}^{(t_i)}\\). Instead, we retrieve only the nearest texture pixels for diffusion, and resolve aliasing and interpolation via a simple distillation post-process (Sec. 4.2.3). For each background pixel in rendered image \\(\\mathbf{x}_{i,..}^{\\prime}\\), we apply a Gaussian noise of standard deviation \\(\\sigma_{i}\\) such that the background matches the marginal distribution at diffusion step \\(t_i\\). This ensures that the diffusion model \\(f_{\\theta}^{(t_i)}\\) focuses solely on the foreground pixels of \\(\\mathbf{x}_{i,..}^{\\prime}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.894, + 0.633 + ], + "angle": 0, + "content": "Note that in this setting a rendered image is simply a selection of pixels from the texture maps, and thus we can leverage backpropagation to easily implement the inverse rendering function \\(\\mathcal{R}^{-1}\\). Additionally, forward rasterization yields a depth map of the scene, which we use as a conditioning input to the diffusion model." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.647, + 0.855, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.721, + 0.894, + 0.791 + ], + "angle": 0, + "content": "Figure 4: Left: Depth map for conditioning SD2-depth, and quality image computed from screen space derivatives. Right: Output of SD2 decoder in two views using renders of the latent texture map as input. Note how the horizontal line across the doors changes appearance from one view to another." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.816, + 0.759, + 0.831 + ], + "angle": 0, + "content": "4.2.2 Aggregating Latent Textures." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Recall from Sec. 4.1, before iterating through cameras \\(\\{C_1,\\dots,C_N\\}\\), we first initialize a latent texture map \\(\\mathbf{z}_{i - 1,0} = \\mathbf{z}_i\\) and render latent image \\(\\mathbf{x}_{i,1}^{\\prime} = \\mathcal{R}(\\mathbf{z}_{i - 1,0})\\). Then, in each step (iterating through cameras), we ob" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4173" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.473, + 0.395 + ], + "angle": 0, + "content": "tain \\(\\mathbf{x}_{i - 1,n}\\) from \\(f_{\\theta}^{(t_i)}(\\mathbf{x}_{i,n}^{\\prime})\\) , and inverse render it to get \\(\\mathbf{z}_{i - 1,n}^{\\prime} = \\mathcal{R}^{-1}(\\mathbf{x}_{i - 1,n})\\) . Finally, we need to aggregate the partial texture map \\(\\mathbf{z}_{i - 1,n}^{\\prime}\\) with \\(\\mathbf{z}_{i - 1,n - 1}\\) to obtain a partially updated texture map \\(\\mathbf{z}_{i - 1,n}\\) . We perform this aggregation step based on a simple heuristic that the value of each pixel \\((u,v)\\) in \\(\\mathbf{z}\\) should be determined by the camera that has the most \"direct and up-close\" view to its corresponding point on the mesh. We measure view quality using image-space derivatives - the amount of change in UV coordinates per infinitesimal change in the image coordinate. This quantity is commonly used for determining mipmapping resolution and anti-aliasing, and it can be efficiently computed by nvdiffrast when rasterizing \\(\\mathbf{z}_{i,n - 1}\\) . For each pixel location \\((p,q)\\) in \\(\\mathbf{x}_{i - 1,n}\\) , we compute the negative Jacobian magnitude as \\(-|\\frac{\\partial u}{\\partial p}\\cdot \\frac{\\partial v}{\\partial q} -\\frac{\\partial u}{\\partial q}\\cdot \\frac{\\partial v}{\\partial p} |,\\) and inverse render it to the texture space, which we denote as \\(Q_{i,n}\\) . Higher values of \\(Q_{i,n}(u,v)\\) means that camera \\(n\\) has a better view of \\((u,v)\\) (see Fig.4). In addition, we compute \\(M_{i,n} = \\mathcal{R}^{-1}(\\boldsymbol {I})\\) , such that \\(M_{i,n}(u,v)\\) represents the number of pixels in \\(\\mathbf{x}_{i,n}^{\\prime}\\) that received value from \\(\\mathbf{z}_i(u,v)\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.397, + 0.472, + 0.489 + ], + "angle": 0, + "content": "While iterating through cameras 1 through \\(N\\), we maintain mask \\(M_{i}\\) (initialized to 0), which is used to track which pixels of \\(\\mathbf{z}_{i,n}\\) have been seen by cameras up to \\(n\\), and quality buffer \\(Q_{i}\\) (initialized to \\(-\\inf\\)), which is used to track the highest non-zero value of \\(Q_{i,n}\\) at each \\((u,v)\\). The value at \\((u,v)\\) of \\(\\mathbf{z}_{i-1,n}\\) is determined as:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.5, + 0.468, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i - 1, n} (u, v) = \\left\\{ \\begin{array}{l l} \\frac {\\mathbf {z} _ {i - 1 , n} ^ {\\prime} (u , v)}{M _ {i , n} (u , v)} & M _ {i, n} (u, v) > 0, \\text {a n d} \\\\ \\mathbf {z} _ {i - 1, n - 1} (u, v) & Q _ {i, n} (u, v) > Q _ {i} (u, v) \\\\ \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.568, + 0.47, + 0.643 + ], + "angle": 0, + "content": "\\(M_{i}\\) and \\(Q_{i}\\) are then updated pixelwise as \\(M_{i} = \\min (M_{i,n} + M_{i},1)\\) and \\(Q_{i} = \\max (Q_{i,n},Q_{i})\\) (min and max are applied element-wise). We use Eqn. 6 in conjunction with Eqn. 5 in SIMS to perform denoising of images in the sequence of camera views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We further note that a side effect of using nearest pixel filtering during texture sampling in SIMS is aliasing. When the image space derivative Jacobian is much higher than 1 (low quality), neighboring pixels in screen space will have gaps when mapped to uv space. This results in incomplete filling of the latent texture map. Conversely, when the image space derivative Jacobian magnitude is much smaller than 1 (high quality), multiple screen space pixels \\((p,q)\\) will map to the same \\((u,v)\\), creating uniform blocks in \\(\\mathbf{x}\\). Similar to interpolated latents, these blocks are particularly detrimental early during the diffusion process, as they cannot be correctly denoised by \\(\\epsilon_{\\theta}\\) which is expecting high spatial variance. Our quality-based aggregation overwrites low-quality regions in the latent texture when higher quality views are available. Thus we can set the resolution of the texture map such that the image space derivatives in each view do not fall below 1 (i.e. \\(\\max Q_{i,n} < -1\\)) to prevent" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "aliasing of the second type, and rely on cameras with better views to fix aliasing of the first type." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.128, + 0.766, + 0.143 + ], + "angle": 0, + "content": "4.2.3 From Latent to RGB Textures" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.895, + 0.729 + ], + "angle": 0, + "content": "So far, we have described how to use SIMS to produce a set of 3D consistent latent images and latent texture maps, but have yet to describe how to translate this into a RGB texture map that can be rendered for viewing. Towards this, we experimented with multiple approaches and found performing multi-view distillation of decoded latent images \\(\\{\\mathbf{x}_{0,n}\\}_{n=1}^{N}\\) with a neural color field to be most performant. Specifically, we use the decoder \\(\\mathcal{D}\\) of Stable Diffusion to decode latent multi-view images \\(\\{\\mathbf{x}_{0,n}\\}_{n=1}^{N}\\) into RGB multi-view images \\(\\{\\xi_{n} = \\mathcal{D}(\\mathbf{x}_{0,n})\\}_{n=1}^{N}\\). Notably, decoding with \\(\\mathcal{D}\\) introduce inconsistencies such that \\(\\{\\xi_{n}\\}_{n=1}^{N}\\) is not 3D consistent even when \\(\\{\\mathbf{x}_{0,n}\\}_{n=1}^{N}\\) is 3D consistent (see Fig. 4 for example). In stable Diffusion, each latent vector (pixel of \\(\\mathbf{x}\\)) needs to carry the information of a \\(8 \\times 8\\) patch of RGB pixels. Thus, the value of the latent vector encodes both color values and spatial patterns. We therefore cannot expect their decoded results to be equivariant to perspective transformations. To address this problem, we leverage a neural color field optimized with appearance-based losses to smooth out inconsistencies. Since we know the ground-truth camera poses, we can directly obtain the 3D spatial coordinates \\(\\{xyz_{n}\\}_{n=1}^{N}\\) of all pixels of \\(\\{\\xi_{n}\\}_{n=1}^{N}\\) by projecting pixels to the mesh surface. Background pixels that do not intersect any surface are discarded. Following [40], we use a multi-resolution hash encoding based on instantNGP [51] along with a shallow MLP \\(f_{\\phi}\\) to parameterize a function from 3D spatial coordinates to RGB values for each sample \\(rgb = f_{\\phi}(hash(xyz))\\). We then distill multi-view images \\(\\{\\xi_{n}\\}_{n=1}^{N}\\) into this parametric function via optimization. Since our goal is to export \\(\\phi\\) into a texture map, we do not use any view-dependent parameterization. To reconcile inconsistencies, we use both a standard \\(L2\\) loss and a VGG-based perceptual loss, applied between \\(f_{\\phi}\\) and \\(\\{\\xi_{n}\\}_{n=1}^{N}\\), to train \\(\\phi\\). We use Adam with a learning rate of 0.01, and optimization of \\(\\phi\\) converges within 100 iterations. After optimization, we compute the spatial coordinate of the centers of each pixel in a high-resolution texture map, and query \\(f_{\\phi}\\) to predict RGB values for the texture map." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.741, + 0.892, + 0.771 + ], + "angle": 0, + "content": "4.2.4 Geometry Processing, Camera, and Multi-resolution Refinement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We normalize \\(\\mathcal{M}\\) such that it fits inside a cube of side length 1, and center it at the origin. Perspective cameras are placed facing the origin, and their FOV is adjusted to fit the object. Detailed parameters can be found in the appendix. As the diffusion model relies on context captured in \\(\\mathbf{x}_i\\), to perform denoising, camera views that are too small w.r.t the size of the object often result in content drift - the texture in distant areas of the mesh can have content that is seman" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4174" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "tically inconsistent. This can be seen as our version of the Janus face problem known to Dreamfusion and similar approaches [40, 59, 84]. Our solution is to perform two rounds of TexFusion at different resolution scales to obtain high-resolution textures while retaining semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.473, + 0.469 + ], + "angle": 0, + "content": "Specifically, we first run TexFusion using cameras that cover the entire object, and a low-resolution latent texture map that is suitable for these cameras. We do not run view distillation in this first round as we are only interested in the latent texture map. We denote the denoised latent texture map from this step as \\(\\mathbf{z}_{0,lr}\\). We then use a second set of cameras with a narrower field of view; these cameras are also more numerous to still cover the full object. We determine a new texture map resolution using the square of the ratio of the tangent of the old FOV over the new FOV - corresponding to the relative change in surface area covered by the camera before and after the change in FOV. We then up-sample \\(\\mathbf{z}_{0,lr}\\) with nearest neighbor filtering to this higher resolution, and stochastically encode it to a partially noised state (e.g. \\(T = 500\\) in the diffusion model time schedule). The second round of TexFusion uses these new cameras, and initializes SIMS with the partially noised latent texture map. The multi-view images produced by the second round of SIMS is used to produce the final output texture map via neural color field distillation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.483, + 0.21, + 0.5 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.471, + 0.721 + ], + "angle": 0, + "content": "We apply TexFusion on various geometries and text prompts to evaluate its ability to produce high quality, natural, and 3D consistent textures. We focus our experimental comparisons on TEXTure [62], a text-driven texture generation method that also uses SD2-depth. We choose TEXTure as the baseline because (1) it represents the current state-of-the-art for language-conditioned texture generation, (2) it uses the same SD2-depth model, which can be prompted for a wide variety of content, and (3) it is concurrent to our work. In the supplementary materials, we further compare TexFusion to SDS-based text-driven texture distillation [48, 59], leveraging the Stable Diffusion model and show that TexFusion achieves superior results in terms of quality and speed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Dataset We collect 35 meshes of various content, and write 2-4 prompts for each mesh. In total, we evaluate each texture synthesis method on 86 mesh-text pairs. More details of this dataset are in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.307, + 0.819 + ], + "angle": 0, + "content": "5.1. Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We visualize textures produced by TexFusion on multiple geometries and prompts, and compare to state-of-the-art baselines in Figure 6. We render each object in 3 surrounding views to allow better examination of 3D consistency of the produced texture. We use Blender's Cycles renderer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "with a studio light-setup. Textures are applied as base color to a diffuse material. Additional visualizations, including videos showing 360 pans of all objects presented in the paper, and renders of our results using only texture color, can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.394 + ], + "angle": 0, + "content": "In terms of local 3D consistency (consistency in neighborhoods on the surface of the mesh), textures produced by TexFusion are locally consistent - there are no visible seam lines or stitching artifacts. In contrast, we often find severe artifacts when viewing the top and back sides of TEXTure's outputs. These artifacts are most noticeable when a clean color is expected, such as when texturing vehicles. In terms of global consistency (semantic coherency of the entire texture, e.g. exactly 2 eyes and 1 nose to a face), TEXTure performs poorly and suffers from problems similar to Dream-Fusion's Janus face problem [59]: as the 2D diffusion model captures context of the object through its own camera view, it is not aware of the appearance of the opposite side of the object. This problem is ameliorated in TexFusion due to frequent communication between views in SIMS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.893, + 0.563 + ], + "angle": 0, + "content": "There are noticeably more baked specular highlights and shadows in textures generated by TEXTure. These effects are physically inaccurate as they conflict with the lighting effects simulated by the renderer. In contrast, TexFusion produces textures that are smoothly illuminated. We hypothesize that this is due to interlacing aggregations in SIMS, which removes view-dependent effects as they arise in the sampling process. We provide additional visualizations of TexFusion in Fig. 5 to showcase how text prompting is effective at controlling the generation of textures, thereby producing varying appearances using the same mesh." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.648 + ], + "angle": 0, + "content": "Runtime TexFusion takes approximately 3 minutes on a machine with a single GPU to sample one texture. We are slightly faster than TEXTure (5 min.) [62], since we only need to optimize a color field once after sampling is" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.669, + 0.862, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.893, + 0.877, + 0.908 + ], + "angle": 0, + "content": "Figure 5: More TexFusion text-conditioned texturing results." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4175" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.094, + 0.892, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.294, + 0.816, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.435, + 0.735, + 0.45 + ], + "angle": 0, + "content": "Figure 6: Visual comparison of textures generated by TEXTure [62] and TexFusion." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.452, + 0.472, + 0.522 + ], + "angle": 0, + "content": "
MethodFID (↓)User study (%)
Natural Color (↑)More Detailed (↑)Less Artifacts (↑)Align with Prompt (↑)
TEXTure79.4724.4265.1231.4043.02
TexFusion59.7875.5834.8868.6056.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.533, + 0.47, + 0.576 + ], + "angle": 0, + "content": "Table 1: Quantitative and qualitative comparisons between TEX-Ture and TexFusion. FID is computed w.r.t. a set of images synthesized by SD2-depth." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.592, + 0.47, + 0.653 + ], + "angle": 0, + "content": "complete. We are also an order of magnitude faster than methods that rely on SDS loss which report upwards of 40 minutes [40, 48, 59]. Additional runtime comparisons are in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.662, + 0.317, + 0.679 + ], + "angle": 0, + "content": "5.2. Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.47, + 0.762 + ], + "angle": 0, + "content": "It is inherently difficult to quantitatively evaluate the quality of text and geometry-conditioned texture generation as one must take alignment with the conditioning inputs into account. We rely on both automated and human evaluation to gauge the quality of the synthesized textures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "FID Since SD2-depth can produce high-quality images that match the structure of the conditioning geometry, we sample it to create a proxy ground truth set. Furthermore, as both TEXTure and TexFusion use SD2-depth as 2D prior, samples from it serve as an upper bound to the quality and diversity of the appearance of textures: We render the depth map of all meshes in 8 canonical views, and condition SD2-depth on both these depth maps and our text" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.454, + 0.895, + 0.651 + ], + "angle": 0, + "content": "prompts to generate a ground truth set. We also render textures generated by both methods in these same 8 views, and compute the Fréchet Inception Distance (FID) between images rendered from the textures and the ground truth set. FID measures the overlap between two sets of images based on their similarity in a learned feature space. We white out the background regions in rendered images and ground truth images to focus the comparisons to the textured object. We present the results in Tab. 1. FID scores indicate that textures synthesized by TexFusion are closer in appearance to the ground truth set than those from TEXTure. The absolute FID values are likely high due to the relatively small set of images in each dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.902 + ], + "angle": 0, + "content": "User study We conducted a user study to measure the overall quality of the generated textures. We use all mesh and text prompts in our dataset, and render each textured mesh into a video showing the results from a \\(360^{\\circ}\\) rotating view. We present videos from TEXTure and TexFusion side by side in random left-right order, to avoid bias. The user study was conducted on Amazon Mechanical Turk. Each participant is asked to select the preferred result under four criteria: natural color (non-saturation), details, cleanliness, and alignment with the text prompt. To avoid randomness in the answers, we let 3 participants answer the same question and determine the choice by max-voting, a question screenshot is provided in the supplement. We provide the user study results in Tab. 1. Although not deemed as detailed as TEXTure, our results are overall preferred by the users in terms of better aligning with the provided text prompt, and" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4176" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.177, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.168, + 0.17, + 0.177 + ], + "angle": 0, + "content": "\"red backpack\"" + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.096, + 0.26, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.168, + 0.263, + 0.177 + ], + "angle": 0, + "content": "\"Brown rabbit\"" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.09, + 0.36, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.16, + 0.367, + 0.186 + ], + "angle": 0, + "content": "\"Portrait of a humanoid robot, futuristic, science fiction\"" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.105, + 0.451, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.16, + 0.455, + 0.187 + ], + "angle": 0, + "content": "\"Medieval celtic Iouse, stone bricks, wooden roof\"" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.188, + 0.177, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.264, + 0.165, + 0.283 + ], + "angle": 0, + "content": "\"camper bag, camouflage\"" + }, + { + "type": "image", + "bbox": [ + 0.189, + 0.197, + 0.269, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.193, + 0.263, + 0.264, + 0.283 + ], + "angle": 0, + "content": "\"Crocodile skin handbag\"" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.19, + 0.358, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.261, + 0.364, + 0.286 + ], + "angle": 0, + "content": "\"Blonde girl with green eyes, hair in tied a bun, DSLR portrait photo.\"" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.196, + 0.459, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.264, + 0.448, + 0.282 + ], + "angle": 0, + "content": "\"Leather lounge chair\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.47, + 0.342 + ], + "angle": 0, + "content": "Figure 7: TexFusion + non-stochastic DDIM sampling \\((\\sigma_{t_i} = 0)\\). This setting emphasizes textural details, such as the leather (bottom, left), roof shingles (top, right)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.37, + 0.421, + 0.385 + ], + "angle": 0, + "content": "having more natural color, and fewer artifacts/flaws." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.397, + 0.317, + 0.414 + ], + "angle": 0, + "content": "5.3. Improving Texture Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.473, + 0.814 + ], + "angle": 0, + "content": "Texfusion can produce more texture details with adjusted hyperparameters and diffusion backbones. First, we find that using the non-stochastic version of DDIM \\((\\eta = 0)\\) adds more materialistic details to the textures on smooth/low-poly geometries. We showcase some examples with particularly large improvements in Fig. 7. Second, we explore the use of ControlNet[98], which can be easily substituted as the LDM backbone for our method without any additional changes. We find that its high-resolution depth conditioning allows TexFusion to capture fine-grained geometric details in the input mesh. In Fig. 8, we further compare TexFusion + ControlNet in \"normal mode\" (apply classifier-free guidance to text prompt only) and ControlNet's \"guess mode\" (apply classifier-free guidance to both text and depth) on meshes with fine geometric details. TexFusion produces high-contrast textures with the appearance of strong lighting under \"guess mode\", and realistic textures with smooth lighting in \"normal mode\". These modifications improve details at a cost of reduced robustness. The non-stochastic DDIM setting may create artifacts when a smooth and clean texture is desired. On the other hand, the increase in depth resolution makes TexFusion+ControlNet susceptible to capturing face boundaries on low-poly meshes. We provide visualizations of these failure cases in Fig. 9. Nonetheless, these modifications offer further dimensions of control that can be explored by practitioners." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.334, + 0.845 + ], + "angle": 0, + "content": "6. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We presented TexFusion, a novel approach to text-driven texture synthesis for given 3D geometries, using only large-scale text-guided image diffusion models. TexFusion lever" + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.102, + 0.686, + 0.111 + ], + "angle": 0, + "content": "TexFusion + ControlNet in \"normal\" mode" + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.102, + 0.857, + 0.111 + ], + "angle": 0, + "content": "TexFusion + ControlNet in \"guess\" mode" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.113, + 0.682, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.112, + 0.848, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.253, + 0.835, + 0.264 + ], + "angle": 0, + "content": "\"Portrait of greek-egyptian deity hermanubis, lapis skin and gold clothing\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.275, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Figure 8: Left: TexFusion + ControlNet in \"normal mode\"; right: TexFusion + ControlNet in \"guess mode\"." + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.313, + 0.673, + 0.331 + ], + "angle": 0, + "content": "Deterministic DDIM: Artifacts in smooth areas" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.33, + 0.699, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.313, + 0.868, + 0.332 + ], + "angle": 0, + "content": "TexFusion+ControlNet: Face boundaries on low poly meshes" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.33, + 0.772, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.33, + 0.872, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.405, + 0.891, + 0.433 + ], + "angle": 0, + "content": "Figure 9: Failure modes exhibited by (left) TexFusion + non-stochastic DDIM sampling and (right) TexFusion + ControlNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.892, + 0.631 + ], + "angle": 0, + "content": "ages the latent diffusion model Stable Diffusion and relies on a new 3D-consistent diffusion model sampling scheme that runs a 2D denoiser network in different views of the object and aggregates the predictions in a latent texture map after each denoising iteration. We find that our method can efficiently generate diverse, high-quality, and globally coherent textures and offers detailed control through text conditioning. Limitations of our approach include a not-yet-ideal sharpness and that texture generation is not real-time. Future work can address these issues; for instance, it would be interesting to leverage the recent literature on faster diffusion model samplers [15, 41, 42, 99]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.816 + ], + "angle": 0, + "content": "Concurrent Works Since Submission Since the time of writing this paper, new works has appeared in the text to texture space [8, 79]. They improve upon TEXTure in aspects such as adding a refinement procedure to automatically fix low quality areas from the initial texturing process [8], and using images sampled from a text-to-image model and per-prompt finetuning to provide stronger conditioning [79]. TexFusion is methodologically distinct from these methods which are based-on TEXTure. Improvements proposed in these works could be combined with TexFusion in future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements We would like to thank Jun Gao for the helpful discussions during the project. Tianshi Cao acknowledges additional income from Vector Scholarships in Artificial Intelligence, which are not in direct support of this work." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4177" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In ICML, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.157, + 0.471, + 0.237 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.238, + 0.471, + 0.278 + ], + "angle": 0, + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.279, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[4] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh Susskind. Gaudi: A neural architect for immersive 3d scene generation. arXiv preprint arXiv:2207.13751, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.36, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[5] Hanqun Cao, Cheng Tan, Zhangyang Gao, Guangyong Chen, Pheng-Ann Heng, and Stan Z. Li. A survey on generative diffusion model. arXiv preprint arXiv:2209.02646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.414, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[6] Ignacio Castaño. thekla交通大学 In github.com/Thekla/thekla_italas, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.441, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.536, + 0.471, + 0.59 + ], + "angle": 0, + "content": "[8] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.591, + 0.471, + 0.686 + ], + "angle": 0, + "content": "[9] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 100–116. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.687, + 0.471, + 0.739 + ], + "angle": 0, + "content": "[10] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.74, + 0.471, + 0.794 + ], + "angle": 0, + "content": "[11] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In The Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.795, + 0.471, + 0.834 + ], + "angle": 0, + "content": "[12] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. arXiv preprint arXiv:2209.04747, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.835, + 0.471, + 0.89 + ], + "angle": 0, + "content": "[13] Congyue Deng, Chiyu \"Max\" Jiang, Charles R. Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, and Dragomir Anguelov. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. arXiv" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.471, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.094, + 0.758, + 0.106 + ], + "angle": 0, + "content": "preprint arXiv:2212.03267, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.107, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[14] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.147, + 0.892, + 0.186 + ], + "angle": 0, + "content": "[15] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. GENIE: Higher-Order Denoising Diffusion Solvers. In Advances in Neural Information Processing Systems, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.187, + 0.892, + 0.24 + ], + "angle": 0, + "content": "[16] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Score-based generative modeling with critically-damped Langevin diffusion. In International Conference on Learning Representations (ICLR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.241, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[17] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.309, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.377, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.431, + 0.892, + 0.47 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.471, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. arXiv preprint arXiv:2106.15282, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.526, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[22] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.566, + 0.892, + 0.619 + ], + "angle": 0, + "content": "[23] Wenlong Huang, Brian Lai, Weijian Xu, and Zhuowen Tu. 3d volumetric modeling with introspective neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33(01), pages 8481-8488, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.62, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[24] Aapo Hyvarinen. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6:695-709, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.66, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[25] Moritz Ibing, Isaak Lim, and Leif P. Kobbelt. 3d shape generation with grid-based implicit functions. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.714, + 0.892, + 0.753 + ], + "angle": 0, + "content": "[26] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.755, + 0.892, + 0.794 + ], + "angle": 0, + "content": "[27] Nikolai Kalischek, Torben Peters, Jan D Wegner, and Konrad Schindler. Tetrahedral diffusion models for 3d shape generation. arXiv preprint arXiv:2211.13220, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.795, + 0.892, + 0.847 + ], + "angle": 0, + "content": "[28] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.849, + 0.892, + 0.874 + ], + "angle": 0, + "content": "[29] Carson Katri. Dream textures. In github.com/carson-katri/dream-textures, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.875, + 0.892, + 0.888 + ], + "angle": 0, + "content": "[30] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.094, + 0.892, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "4178" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Song. Denoising diffusion restoration models. arXiv preprint arXiv:2201.11793, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.12, + 0.47, + 0.173 + ], + "angle": 0, + "content": "[31] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Popa Tiberiu. Clip-mesh: Generating textured meshes from text using pretrained image-text models. December 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.173, + 0.471, + 0.214 + ], + "angle": 0, + "content": "[32] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.214, + 0.471, + 0.268 + ], + "angle": 0, + "content": "[33] Wei-Jan Ko, Hui-Yu Huang, Yu-Liang Kuo, Chen-Yi Chiu, Li-Heng Wang, and Wei-Chen Chiu. Rpg: Learning recursive point cloud generation. arXiv preprint arXiv:2105.14322, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.268, + 0.471, + 0.322 + ], + "angle": 0, + "content": "[34] Johannes Kopf, Chi-Wing Fu, Daniel Cohen-Or, Oliver Deussen, Dani Lischinski, and Tien-Tsin Wong. Solid texture synthesis from 2d exemplars. ACM Trans. Graph., 26(3):2-es, jul 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.322, + 0.471, + 0.377 + ], + "angle": 0, + "content": "[35] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.377, + 0.471, + 0.416 + ], + "angle": 0, + "content": "[36] Sylvain Lefebvre and Hugues Hoppe. Appearance-space texture synthesis. ACM Trans. Graph., 25(3):541-548, jul 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.416, + 0.471, + 0.457 + ], + "angle": 0, + "content": "[37] Chun-Liang Li, Manzil Zaheer, Yang Zhang, Barnabas Poczos, and Ruslan Salakhutdinov. Point cloud gan. arXiv preprint arXiv:1810.05795, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.457, + 0.471, + 0.512 + ], + "angle": 0, + "content": "[38] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.511, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[39] Ruihui Li, Xianzhi Li, Ke-Hei Hui, and Chi-Wing Fu. SPGAN:sphere-guided 3d shape generation and manipulation. ACM Transactions on Graphics (Proc. SIGGRAPH), 40(4), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.565, + 0.471, + 0.632 + ], + "angle": 0, + "content": "[40] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.632, + 0.471, + 0.685 + ], + "angle": 0, + "content": "[41] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.686, + 0.471, + 0.741 + ], + "angle": 0, + "content": "[42] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv:2206.00927, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.741, + 0.471, + 0.795 + ], + "angle": 0, + "content": "[43] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. arXiv preprint arXiv:2201.09865, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.795, + 0.471, + 0.849 + ], + "angle": 0, + "content": "[44] A. Luo, T. Li, W. Zhang, and T. Lee. Surfgen: Adversarial 3d shape synthesis with explicit surface discriminators. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.849, + 0.471, + 0.891 + ], + "angle": 0, + "content": "[45] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.471, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.735, + 0.106 + ], + "angle": 0, + "content": "Recognition (CVPR), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.106, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[46] Siwei Lyu. Interpretation and generalization of score matching. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, UAI '09, page 359-366, Arlington, Virginia, USA, 2009. AUAI Press. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.16, + 0.892, + 0.214 + ], + "angle": 0, + "content": "[47] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.214, + 0.892, + 0.268 + ], + "angle": 0, + "content": "[48] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. 1, 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.268, + 0.892, + 0.335 + ], + "angle": 0, + "content": "[49] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.335, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[50] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.39, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[51] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.444, + 0.892, + 0.499 + ], + "angle": 0, + "content": "[52] Gimin Nam, Mariem Khlifi, Andrew Rodriguez, Alberto Tono, Linqi Zhou, and Paul Guerrero. 3d-ldm: Neural implicit 3d shape generation with latent diffusion models. arXiv preprint arXiv:2212.00842, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.499, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[53] Charlie Nash, Yaroslav Ganin, S. M. Ali Eslami, and Peter W. Battaglia. *Polygen: An autoregressive generative model of 3d meshes.* ICML, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.538, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[54] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.607, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[55] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.66, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[56] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.702, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[57] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.769, + 0.892, + 0.849 + ], + "angle": 0, + "content": "[58] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.85, + 0.892, + 0.891 + ], + "angle": 0, + "content": "[59] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 1, 2, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.926, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4179" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.229 + ], + "angle": 0, + "content": "[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.471, + 0.284 + ], + "angle": 0, + "content": "[61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[62] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.405 + ], + "angle": 0, + "content": "[63] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.407, + 0.471, + 0.461 + ], + "angle": 0, + "content": "[64] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image dissusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.462, + 0.471, + 0.514 + ], + "angle": 0, + "content": "[65] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. arXiv preprint arXiv:2111.05826, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.471, + 0.61 + ], + "angle": 0, + "content": "[66] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.611, + 0.471, + 0.664 + ], + "angle": 0, + "content": "[67] Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. arXiv preprint arXiv:2104.07636, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.665, + 0.471, + 0.705 + ], + "angle": 0, + "content": "[68] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations (ICLR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.706, + 0.471, + 0.759 + ], + "angle": 0, + "content": "[69] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. arXiv preprint arXiv:2110.02624, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.76, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[70] Hiroshi Sasaki, Chris G. Willcocks, and Toby P. Breckon. UNIT-DDPM: Unpaired image translation with denoising diffusion probabilistic models. arXiv preprint arXiv:2104.05358, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.814, + 0.471, + 0.881 + ], + "angle": 0, + "content": "[71] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3859-3868, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.882, + 0.471, + 0.895 + ], + "angle": 0, + "content": "[72] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.471, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. arXiv preprint arXiv:2211.16677, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.134, + 0.892, + 0.243 + ], + "angle": 0, + "content": "[73] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In Shai Avidan, Gabriel J. Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, volume 13663 of Lecture Notes in Computer Science, pages 72-88. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.244, + 0.892, + 0.297 + ], + "angle": 0, + "content": "[74] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.298, + 0.892, + 0.338 + ], + "angle": 0, + "content": "[75] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.338, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[76] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.406, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[77] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. arXiv preprint arXiv:2203.08382, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.446, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[78] Yongbin Sun, Yue Wang, Ziwei Liu, Joshua E Siegel, and Sanjay E Sarma. Pointgrow: Autoregressively learned point cloud generation with self-attention. In Winter Conference on Applications of Computer Vision, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.5, + 0.892, + 0.526 + ], + "angle": 0, + "content": "[79] Zhibin Tang and Tiantong He. Text-guided high-definition consistency texture model. ArXiv, abs/2305.05901, 2023. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.527, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[80] Greg Turk. Texture synthesis on surfaces. SIGGRAPH '01, page 347-354, New York, NY, USA, 2001. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.567, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[81] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.608, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[82] Diego Valsesia, Giulia Fracastoro, and Enrico Magli. Learning localized generative models for 3d point clouds via graph convolution. In International Conference on Learning Representations (ICLR) 2019, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.661, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[83] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 23(7):1661-1674, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.701, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[84] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.755, + 0.892, + 0.823 + ], + "angle": 0, + "content": "[85] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jian-min Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.823, + 0.892, + 0.893 + ], + "angle": 0, + "content": "[86] Li-Yi Wei and Marc Levoy. Texture synthesis over arbitrary manifold surfaces. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '01, page 355-360, New York, NY, USA, 2001. Association for Computing Machinery. 2" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.926, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4180" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[87] Cheng Wen, Baosheng Yu, and Dacheng Tao. Learning progressive point embeddings for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10266-10275, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.472, + 0.215 + ], + "angle": 0, + "content": "[88] Li-Yi Wie, Sylvain Lefebvre, Vivek Kwatra, and Greg Turk. State of the Art in Example-based Texture Synthesis. In M. Pauly and G. Greiner, editors, Eurographics 2009 - State of the Art Reports. The Eurographics Association, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[89] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.283, + 0.47, + 0.351 + ], + "angle": 0, + "content": "[90] Fanbo Xiang, Zexiang Xu, Milos Hasan, Yannick Hold-Geoffroy, Kalyan Sunkavalli, and Hao Su. Neutex: Neural texture mapping for volumetric neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7119-7128, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.351, + 0.47, + 0.433 + ], + "angle": 0, + "content": "[91] Jianwen Xie, Yifei Xu, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Zhu Song-Chun, and Ying Nian Wu. Generative pointnet: Deep energy-based learning on unordered point sets for 3d generation, reconstruction and classification. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.433, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[92] Jianwen Xie, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Zhu Song-Chun, and Ying Nian Wu. Learning descriptor networks for 3d shape synthesis and analysis. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.47, + 0.554 + ], + "angle": 0, + "content": "[93] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. PointFlow: 3D point cloud generation with continuous normalizing flows. 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.554, + 0.47, + 0.595 + ], + "angle": 0, + "content": "[94] Ling Yang, Zhilong Zhang, and Shenda Hong. Diffusion models: A comprehensive survey of methods and applications. arXiv preprint arXiv:2209.00796, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.595, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[95] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations. In Proceedings of International Conference on Computer Vision (ICCV), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.662, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[96] Jonathan Young. xatlas. In github.com/jpcy/xatlas, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.676, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[97] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.743, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[98] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.771, + 0.47, + 0.809 + ], + "angle": 0, + "content": "[99] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv:2204.13902, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.81, + 0.47, + 0.878 + ], + "angle": 0, + "content": "[100] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yi-nan Zhang, Antonio Torralba, and Sanja Fidler. Image {gan}s meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.879, + 0.47, + 0.892 + ], + "angle": 0, + "content": "[101] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.926, + 0.514, + 0.937 + ], + "angle": 0, + "content": "4181" + } + ] +] \ No newline at end of file diff --git a/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/698eb5a8-8d74-4b19-b9a7-1fe068b4d1fd_origin.pdf b/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/698eb5a8-8d74-4b19-b9a7-1fe068b4d1fd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..55bb532ee7219e8c82859a9767245efa94109ad1 --- /dev/null +++ b/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/698eb5a8-8d74-4b19-b9a7-1fe068b4d1fd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64aa163c6eaa3926675ceb838eaff60e76c7f9a60bf22b0b78d1f51ebb01a266 +size 3632501 diff --git a/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/full.md b/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..853aa13c687d6857b632e609f40e9d298c88bb2a --- /dev/null +++ b/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/full.md @@ -0,0 +1,416 @@ +# TexFusion: Synthesizing 3D Textures with Text-Guided Image Diffusion Models + +Tianshi Cao $^{1,2,3}$ + +Karsten Kreis + +Sanja Fidler $^{1,2,3}$ + +Nicholas Sharp $^{1,*}$ + +Kangxue Yin $^{1,*}$ + +$^{1}$ NVIDIA, + +$^{2}$ University of Toronto, + +3 Vector Institute + +{tianshic, kkreis, sfidler, nsharp, kangxuey}@nvidia.com + +![](images/2be179c4a496587e538c93d3ec1f159a45848594f67b3190a66c184d9434b3db.jpg) + +![](images/ee7570178093fd5858103cc7570b20f724afff240872004e423356b2fbffdde4.jpg) + +![](images/740797af04e6edcb45341dbbc02b0a26e86ca97f6ad3853b2fe850f064762177.jpg) +"Crocodile skin handbag" + +![](images/41f09a7a345570b2c8ab89b5f6c8376f9f8691b3c7df5d547e30e740731d7119.jpg) +"White Mercedes Benz SUV" + +![](images/9331011979d2a6b5de089c668e779b29776a486c708bef2efd0f19252c643358.jpg) + +![](images/80ccf636a0e14e45a7d3f86fbe0e0260c4d4b6cc0c8df82cd2605f65436a4ba8.jpg) +"White humanoid robot, movie poster, main character of a science fiction movie" +"Oil painting of a bald, middle aged banker with pointed moustache" + +![](images/b67e6e0e19d4a3408ade25aa37a0ee05cd0f93e3952a4c9d33a7f208b38d9cc1.jpg) + +![](images/0b7ebf7c5ea315d5d0c2ac36573d8f1f493a3909a219c0670f512e1a9cf74dd4.jpg) +"Portrait photo of Abraham Lincoln, full color" + +![](images/b2e1084442a5131dd593d45319288578e40cee64bcbe020c8a08ea4619b49045.jpg) + +![](images/0ca7c7502335d034f196e2b5cd00e73bc6c3432995f6d26a5cef59671bd9caa8.jpg) +"Railroad worker wearing high-vis vest" +Figure 1: Text-conditioned 3D texturing results with TexFusion. + +# Abstract + +We present TexFusion (Texture Diffusion), a new method to synthesize textures for given 3D geometries, using large-scale text-guided image diffusion models. In contrast to recent works that leverage 2D text-to-image diffusion models to distill 3D objects using a slow and fragile optimization process, TexFusion introduces a new 3D-consistent generation technique specifically designed for texture synthesis that employs regular diffusion model sampling on different 2D rendered views. Specifically, we leverage latent diffusion models, apply the diffusion model's denoiser on a set of 2D renders of the 3D object, and aggregate the different denoising predictions on a shared latent texture map. Final output RGB textures are produced by optimizing an intermediate neural color field on the decodings of 2D renders of the latent texture. We thoroughly validate TexFusion and show that we can efficiently generate diverse, high quality and globally coherent textures. We achieve state-of-the-art text-guided texture synthesis performance using only image diffusion models, while avoiding the pitfalls of previous distillation-based methods. The text-conditioning offers detailed control and we also do not rely on any ground truth 3D textures for training. This makes our method versatile + +and applicable to a broad range of geometry and texture types. We hope that TexFusion will advance AI-based texturing of 3D assets for applications in virtual reality, game design, simulation, and more. + +# 1. Introduction + +In the past decade, deep learning-based 3D object generation has been studied extensively [1, 4, 18, 23, 25, 27, 32, 33, 37, 39, 44, 45, 50, 52, 53, 55, 58, 69, 71, 72, 78, 82, 85, 87, 89, 91-93, 95, 97, 100, 101], due to the demand for high-quality 3D assets in 3D applications such as VR/AR, simulation, digital twins, etc. While many prior works on 3D synthesis focus on the geometric components of the assets, textures are studied much less, despite the fact that they are important components of realistic 3D assets which assign colors and materials to meshes to make the rendering vivid. Recent advances in text-conditioned image diffusion models trained on internet-scale data [2, 54, 61, 63, 66] have unlocked the capability to generate images with stunning visual detail and practically unlimited diversity. These high-performance diffusion models have also been used as image priors to synthesize 3D objects with textures using textual guidance [40, 48, 59, 84]. + +In this paper, we aim to perform text-driven high-quality 3D texture synthesis for given meshes, by leveraging the information about the appearance of textures carried by the + +image prior of a pre-trained text-to-image diffusion model. The main workhorse in current text-to-3D methods that leverage 2D image diffusion models is Score Distillation Sampling (SDS) [59]. SDS is used to distill, or optimize, a 3D representation such that its renders are encouraged to be high-likelihood under the image prior. Methods utilizing SDS also share two common limitations in that: 1). a high classifier-free guidance weight [22] is required for the optimization to converge, resulting in high color saturation and low generation diversity; 2). a lengthy optimization process is needed for every sample. + +To address the above issues, we present Texture Diffusion, or TexFusion for short. TexFusion is a sampler for sampling surface textures from image diffusion models. Specifically, we use latent diffusion models that efficiently autoencode images into a latent space and generate images in that space [63, 81]. TexFusion leverages latent diffusion trajectories in multiple object views, encoded by a shared latent texture map. Rendering of the shared latent texture map are provided as input to the denoiser of the latent diffusion model, and the output of every denoising step is projected back to the shared texture map in a 3D-consistent manner. To transform the generated latent textures into RGB textures, we optimize a neural color field on the outputs of the latent diffusion model's decoder applied to different views of the object. We use the publicly available latent diffusion model Stable-Diffusion with depth conditioning [63] (SD2-depth) as our diffusion backbone. Compared to methods relying on SDS, TexFusion produces textures with more natural tone, stronger view consistency, and is significantly faster to sample (3 minutes vs. 30 minutes reported by previous works). + +We qualitatively and quantitatively validate TexFusion on various texture generation tasks. We find that TexFusion generates high quality, globally coherent and detailed textures that are well-aligned with the text prompts used for conditioning (e.g. Fig. 1). Since we leverage powerful text-to-image diffusion models for texture sampling, we can generate highly diverse textures and are not limited to single categories or restricted by explicit 3D textures as training data, which are limitations of previous works [7, 11, 18, 57, 73]. In summary, our main contribution is a novel method for 3D texture generation from 2D image diffusion models, that is view-consistent, avoids over-saturation and achieves state-of-the-art text-driven texture synthesis performance. + +# 2. Related Work + +Classic Computer Graphics Techniques Early work on texture generation focused on tiling exemplar patterns across a surface, often with an explicit direction field for local orientation [34, 36, 80, 86]. See [88] for a survey. This research established the value of texture image representations and the challenges of global coherence, both central + +to this work. However, modern learning-based priors have proven necessary to go beyond simple patterns and synthesize plausible shape-specific texture details. + +Texture Synthesis with 3D Priors Textures are defined on the meshed surface of 3D objects, which is an irregular representation. To enable 2D texture generation on 3D meshes, AUV-Net [11] learns an aligned UV space for a set of 3D shapes in a given class, mapping 3D texture synthesis to a 2D domain. Texturify [73] trains a 3D StyleGAN in the quad-parameterized surface domain on a set of textured 3D shapes in a given class. Different from AUV-NET or Texturify, which embed the geometric prior into a UV map or mesh parameterization, EG3D [7] and GET3D [18] directly train 3D StyleGANs to generate geometry and texture jointly, where the textures are implicit texture fields [57]. Other works also represent 3D texture by vertex colors [49], voxels [9], cube mapping [90], etc. In contrast to TexFusion, these works mostly don't offer text conditioning, often work only on single object categories or require textured 3D shapes for training, which limits their broad applicability. + +Diffusion Models Our approach directly builds on diffusion models [5, 12, 94], which have recently emerged as new state-of-the-art generative models. In particular, they have demonstrated outstanding performance in image generation [14-16, 21, 56, 81], outperforming GANs, and led to breakthroughs in text-to-image synthesis [2, 54, 61, 63, 66]. They have also been successfully used for a variety of image editing and processing tasks [17, 19, 30, 38, 43, 47, 64, 65, 67, 70, 77]. Moreover 3D object generation has been addressed with diffusion models, too, for instance leveraging point clouds [45, 55, 97, 101], meshes [27], or neural fields [4, 52, 72, 85] as 3D representations. However, these works focus on geometry generation and do not specifically tackle 3D texture synthesis. + +Distilling 3D Objects from 2D Image Diffusion Models Recently, large-scale 2D text-to-image diffusion models have been leveraged to distill individual 3D objects as neural radiance fields using Score Distillation Sampling (SDS) [13, 40, 48, 59, 84]. In SDS, the radiance field is rendered from different directions into 2D images and it is optimized such that each render has a high probability under the text-to-image diffusion model while conditioning on a text prompt. DreamFusion [59] pioneered the approach, Magic3D [40] proposed a coarse-to-fine strategy improving quality, and Latent-NeRF [48] performs distillation in latent space leveraging a latent diffusion model [63]. These approaches do not specifically target texture generation, which is the focus of this work. More deeply, a crucial drawback of this line of work is that SDS typically requires strong guidance [22] to condition the diffusion model, which can hurt quality and diversity. Moreover, SDS's + +![](images/62a8b985cad542cd68bad8c5a63f5d957d8a5bf86481af1571019c27b918e07d.jpg) +Distill with Neural Color Field + +![](images/5e6ac964bdbdf3700fd301f627c54e9cd45b942d2e2821452810a02e111b4ffd.jpg) +Figure 2: Overview of TexFusion. TexFusion takes a text prompt and mesh geometry as input and produces a UV parameterized texture image matching the prompt and mesh. Key to TexFusion is the Sequential Interlaced Multiview Sampler (SIMS) - SIMS performs denoising diffusion iterations in multiple camera views, yet the trajectories are aggregated through a latent texture map after every denoising step. SIMS produces a set of 3D consistent latent images (TexFusion uses Stable Diffusion [63] as text-to-image diffusion backbone), which are decoded and fused into a texture map via optimizing an intermediate neural color field. + +iterative optimization process makes synthesis very slow. In contrast, our approach avoids SDS entirely and leverages regular diffusion model sampling in a new, 3D-consistent manner. Earlier works also leverage CLIP [60] for 3D object or texture distillation [10, 26, 31, 49], but this performs usually worse than using diffusion models instead. + +Concurrent Work Concurrently with this work, TEX-Ture [62] proposes a related approach. Like this work, TEX-Ture performs multiview denoising on a texture map representation. However, TEX-Ture runs an entire generative denoising process in each camera view in sequence, conditioning on the previous views and projecting to the texture map only after full denoising. In contrast, in TexFusion we propose to interleave texture aggregation with denoising steps in different camera views, simultaneously generating the entire output. This insight significantly reduces view inconsistencies and improves quality in TexFusion compared to TEX-Ture, as validated in Sec. 5. MultiDiffusion [3] concurrently introduces a method for panorama generation and other controlled image generation tasks, leveraging relatively lower-resolution image diffusion models. Algorithmically, this approach of aggregating different denoising predictions from different image crops is closely related to TexFusion's aggregation from different camera views. However, MultiDiffusion only tackles image synthesis, and is not concerned with any 3D or texture generation at all. + +# 3. Background + +Diffusion Models Diffusion models [20, 74, 76] model a data distribution $p_{\mathrm{data}}(\mathbf{x})$ via iterative denoising, and are trained with denoising score matching [20, 24, 46, 74, 76, 83]. Given samples $\mathbf{x} \sim p_{\mathrm{data}}$ and $\epsilon \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ , a denoiser model $\epsilon_{\theta}$ parameterized with learnable parameters + +$\theta$ receives diffused inputs $\mathbf{x}_t(\epsilon, t, \mathbf{x})$ and is optimized by minimizing the denoising score matching objective + +$$ +\mathbb {E} _ {\mathbf {x} \sim p _ {\text {d a t a}}, t \sim p _ {t}, \boldsymbol {\epsilon} \sim \mathcal {N} (\mathbf {0}, \mathbf {I})} \left[ \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} (\mathbf {x} _ {t}; \mathbf {c}, t) \| _ {2} ^ {2} \right], \tag {1} +$$ + +where $\mathbf{c}$ is optional conditioning information, such as a text prompt, and $p_t$ is a uniform distribution over the diffusion time $t$ . The model effectively learns to predict the noise $\epsilon$ that was used to perturb the data (other formulations are possible [28, 68]). Letting $\alpha_{t}$ define a noise schedule, parameterized via a diffusion-time $t$ , we construct $\mathbf{x}_t$ as $\mathbf{x}_t = \sqrt{\alpha_t}\mathbf{x} + \sqrt{1 - \alpha_t}\epsilon$ , $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ ; this particular formulation corresponds to a variance-preserving schedule [76]. The forward diffusion as well as the reverse generation process in diffusion models can be described in a continuous-time framework [76], but in practice a fixed discretization can be used [20]. The maximum diffusion time is generally chosen such that the input data is entirely perturbed into Gaussian random noise and an iterative generative denoising process can be initialized from such Gaussian noise to synthesize novel data. + +Classifier-free guidance [22] can be used for improved conditioning. By randomly dropping out the conditioning c during training, we can learn both a conditional and an unconditional model at the same time, and their predictions can be combined to achieve stronger conditioning. + +We perform iterative diffusion model sampling via the Denoising Diffusion Implicit Models (DDIM) scheme [75]: + +$$ +\begin{array}{l} \boldsymbol {x} _ {i - 1} = \sqrt {\alpha_ {i - 1}} \left(\frac {\boldsymbol {x} _ {i} - \sqrt {1 - \alpha_ {i}} \boldsymbol {\epsilon} _ {\theta} ^ {(t _ {i})} (\boldsymbol {x} _ {i})}{\sqrt {\alpha_ {i}}}\right) \tag {2} \\ + \sqrt {1 - \alpha_ {i - 1} - \sigma_ {t _ {i}} ^ {2}} \cdot \pmb {\epsilon} _ {\theta} ^ {(t _ {i})} (\pmb {x} _ {i}) + \sigma_ {t _ {i}} \pmb {\epsilon} _ {t _ {i}} \\ \end{array} +$$ + +with $\epsilon_{t_i}\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ and $\sigma_{t_i}$ is a variance hyperparameter. We express obtaining $\pmb{x}_{i - 1}$ via DDIM sampling as $x_{i - 1}\sim$ + +$\pmb{f}_{\theta}^{(t_i)}(\pmb{x}_{i-1}|\pmb{x}_i)$ . See Supp. Material for more details. + +Latent Diffusion Models (LDMs) and Stable Diffusion Instead of directly operating on pixels, LDMs [63] utilize an encoder $\mathcal{E}$ and a decoder $\mathcal{D}$ for translation between images $\xi$ and latents $\mathbf{x} \in X$ of a lower spatial dimension. The diffusion process is then defined over the distribution of $X$ . Stable Diffusion is an LDM trained on the LAION-5B image-text dataset. In addition to text-conditioning, Stable Diffusion 2.0 permits depth conditioning with a depth map $D$ (SD2-depth). This allows detailed control of the shape and configuration of objects in its synthesized images. + +**Rendering and Geometry Representation** In principle, our method applies to any geometry representation for which a textured surface can be rendered; in practice, our experiments use a surface mesh $\mathcal{M} = (\mathcal{V},\mathcal{F})$ , with vertices $\mathcal{V} = \{v_i\}$ , $v_{i}\in \mathbb{R}^{3}$ and triangular faces $\mathcal{F} = \{f_i\}$ where each $f_{i}$ is a triplet of vertices. Textures are defined in 2D image space in an injective UV parameterization of $\mathcal{M}$ , $UV:p\in \mathcal{M}\mapsto (u,v)\in [0,1]^2$ . If needed, this parameterization can be automatically constructed via tools such as XATLAS [6, 96]. We encode textures as multi-channel images discretized at pixels in UV space $\mathbf{z}\in \mathbb{R}^{(H\times W,C)}$ , notationally collapsing the spatial dimension for simplicity. + +We denote the rendering function as $\mathcal{R}(\mathbf{z};\mathcal{M},C):\mathbf{z}\mapsto$ $\mathbf{x}$ $\mathbf{x}\in \mathbb{R}^{(h\times w,C)}$ , which takes as input a mesh $\mathcal{M}$ , camera $C$ , and texture $\mathbf{z}$ , and produces as output a rendered image. The inverse of this function $\mathcal{R}^{-1}(\mathbf{x};\mathcal{M},C):\mathbf{x}\mapsto \mathbf{z}$ projects values from camera image-space onto the UV texture map. Notationally, we often omit the dependence on $\mathcal{M}$ and $C$ for brevity. In this work, we do not model any lighting or shading effects, images are formed by directly projecting textures into screen space (and then decoding, for latent textures). + +# 4. Texture Sampling with 2D Diffusion Models + +Given a mesh geometry $\mathcal{M}$ , a text prompt $y$ , and a geometry (depth) conditioned image diffusion model $\theta$ , how could one sample a complete surface texture? Assuming access to the rendering $\mathcal{R}$ and inverse rendering $\mathcal{R}^{-1}$ functions defined above, perhaps the most naive approach is to compute a set of $\{C_1,..,C_N\}$ camera views that envelopes the surface, render the depth map $d_{n}$ in each view, sample images from the image diffusion model with depth conditioning, and then back-project these images to the mesh surface (e.g. as done in [29]). However, image diffusion models in each view have no information about the generated results in other views, thus there is no coherency in the contents generated in each view. As an alternative, one may define a canonical order of camera views, and autoregressively condition image sampling in the subsequent views on + +previously sampled regions (as done in [48, 62]). However, for most geometries of interest (i.e. not a plane), a single camera can not observe the entirety of the geometry. Consequently, images synthesized early in the sequence could produce errors that are not reconcilable with the geometry that is observed later in the sequence (see Fig. 3). Thus, it is desirable for the image sampling distribution $p(\cdot | d_i, y)$ in each camera view to be conditioned on that of every other camera view. + +![](images/7c2b3b8da3a5357c3094c2ff57a1b9983a0efdb4c0ea5354ba408ab5edd0c9a2.jpg) +Figure 3: Illustration of irreconcilable mistakes in early views impacting denoised results in later views (images sampled from TEXTure). While the highlighted area appears natural in the first view, it does not match the geometry when viewed from a different angle, thereby creating poorly denoised results when the second view is inpainted. + +# 4.1. Sequential Interlaced Multiview Sampler + +Leveraging the sequential nature of the denoising process, we can interlace the synchronization of content across views with the denoising steps within each view to achieve coherency over the entire shape. Suppose that at step $i$ of the denoising process, we have a set of partially denoised images $\{\mathbf{x}_{i,n}\}_{n = 1}^{N} = \{\mathbf{x}_{i,1}\dots \mathbf{x}_{i,N}\}$ . Our goal is to sample $\{\mathbf{x}_{i - 1,n}\}_{n = 1}^{N} = \mathbf{x}_{i - 1,1}\dots \mathbf{x}_{i - 1,N}$ that is 3D-consistent, i.e., two pixels in $\mathbf{x}_{i - 1,a},\mathbf{x}_{i - 1,b}$ that project to the same point in 3D should have the same value. Taking inspiration from the autoregressive modeling literature, we sample the joint distribution via first decomposing it into a product of conditionals, and approximating each term in the product by using the diffusion model to denoise the renders of a dynamically updated latent texture map. Specifically, we first initialize an initial latent texture map $\mathbf{z}_T\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , ( $T$ demarks the first step of the diffusion trajectory). Then, suppose that we have a 3D consistent latent texture map $\mathbf{z}_i$ , we decompose the joint distribution as follows (conditioning on depth and prompt omitted for space): + +$$ +\begin{array}{l} p _ {\theta} \left(\left\{\mathbf {x} _ {i - 1, j} \right\} _ {j = 1} ^ {N} \mid \mathbf {z} _ {i}\right) = p _ {\theta} \left(\mathbf {x} _ {i - 1, 1} \mid \mathbf {z} _ {i}\right) \times \\ \prod_ {n = 2} ^ {N} p _ {\theta} \left(\mathbf {x} _ {i - 1, n} \mid \left\{\mathbf {x} _ {i - 1, j} \right\} _ {j = 1} ^ {n - 1}, \mathbf {z} _ {i}\right) \tag {3} \\ \end{array} +$$ + +We can compute the first term by first rendering $\mathbf{z}_i$ into $\mathbf{x}_{i,1}^{\prime} = \mathcal{R}(\mathbf{z}_i;C_1)$ . Eqn. 2 can now be applied to $\mathbf{x}_{i,1}^{\prime}$ to draw a latent image at the next time step: + +$$ +\mathbf {x} _ {i - 1, 1} \sim f _ {\theta} ^ {(t _ {i})} \left(\mathbf {x} _ {i - 1, 1} \mid \mathbf {x} _ {i, 1} ^ {\prime} = \mathcal {R} \left(\mathbf {z} _ {i}; C _ {1}\right)\right). \tag {4} +$$ + +Later terms in Eq. 3 additionally depend on the result of previous denoising steps. We again model this dependency through the latent texture map. For each view starting at $n = 1$ , we inverse render $\mathbf{x}_{i - 1,n}$ into texture space to obtain $\mathbf{z}_{i - 1,n}^{\prime} = \mathcal{R}^{-1}(\mathbf{x}_{i - 1,n})$ , and update the pixels of $\mathbf{z}_{i - 1,n - 1}$ that are newly visible in $\mathbf{z}_{i - 1,n}^{\prime}$ to obtain $\mathbf{z}_{i - 1,n}$ (See Sec. 4.2.2 for details). Then, in the $n + 1$ iteration, since $\mathbf{z}_{i - 1,n}$ contain regions at two noise scales (unseen regions are at noise scale $\sigma_{i}$ , while visited regions are at noise scale $\sigma_{i - 1}$ ), we add appropriate 3D consistent noise to the visited regions of $\mathbf{z}_{i - 1,n}$ to match the noise scale of step $i$ before rendering it as input to $f_{\theta}$ . Letting $M_{i,n}$ represent the mask for visited regions and $\epsilon_{i}\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , we can write the sampling procedure for $\mathbf{x}_{i - 1,n}$ as: + +$$ +\begin{array}{l} \mathbf {z} _ {i, n} = M _ {i, n} \odot \left(\sqrt {\frac {\alpha_ {i - 1}}{\alpha_ {i}}} \mathbf {z} _ {i - 1, n - 1} + \sigma_ {i} \epsilon_ {i}\right) \\ + \left(\mathbf {1} - M _ {i, n}\right) \odot \mathbf {z} _ {i} \tag {5} \\ \end{array} +$$ + +$$ +\mathbf {x} _ {i - 1, n} \sim f _ {\theta} ^ {(t _ {i})} (\mathbf {x} _ {i - 1, n} | \mathbf {x} _ {i, n} ^ {\prime} = \mathcal {R} (\mathbf {z} _ {i, n}; C _ {n})) +$$ + +By iteratively applying Eqn. 5, we obtain a sequence of 3D consistent images $\{\mathbf{x}_{i - 1,n}\}_{n = 1}^{N}$ and a texture map $\mathbf{z}_{i - 1,n}$ that has been aggregated from these images. We can then decrement $i$ and repeat this process in the next time step. + +We name this approach Sequential Interlaced Multiview Sampler (SIMS). SIMS communicates the denoising direction of previous views to the current view and resolves overlaps between views by the order of aggregation. It ameliorates inconsistent predictions while circumventing performance degradation due to the averaging of latent predictions during parallel aggregation. In the single-view case, SIMS is equivalent to standard DDIM sampling. A complete algorithm for SIMS can be found in the appendix. + +# 4.2. The TexFusion Algorithm + +In Sec. 4.1, we have presented a generic algorithm for sampling 3D consistent multi-view images and texture maps using 2D diffusion models. We now present a concrete algorithm, TexFusion, that uses SIMS to texture 3D meshes using SD2-depth [63] as the diffusion model. An illustrative overview of TexFusion can be found in Fig. 2. + +As SD2-depth is a latent diffusion model, we apply SIMS in the latent space: $\mathbf{x}$ and $\mathbf{z}$ represent latent images and latent texture maps respectively, and Section 4.2.3 will describe a final distillation post-process to color space. In this section, we address several challenges specific to using LDMs with SIMS, and detail our design choices to tackle these challenges. We find that a canonical set of cameras works well for most objects, but cameras are tailored to specific objects to improve resolution and ameliorate occlusion. We further illustrate a technique for obtaining high-quality results by operating SIMS in a cascade of increasing resolutions. + +# 4.2.1 (Inverse) Rendering of Latent Textures + +We use NVDIFFRAST [35] to efficiently render textured meshes via rasterization, as described in Sec. 3. Our implementation sets the rendered image size $h = w = 64$ to match Stable Diffusion's UNet, with latent vectors of dimension $D = 4$ . The texture image dimensions $H$ and $W$ are chosen based on the surface area of the mesh relative to its diameter (detailed in Sec. 4.2.4). + +For each non-background pixel $s$ in a rendered image $\mathbf{x}_{i,..}^{\prime}$ , rasterization gives a corresponding location on the texture image via the UV map $(u,v) = UV(p)$ , and we retrieve the value at the nearest texture pixel. This texture value is latent, and no shading or lighting is applied. In other settings, texture data is often rendered with bilinear filtering and mipmaps to improve image quality and reduce aliasing, but in this setting, we found it essential to avoid such techniques. We experimentally ablate texturing approaches in the supplementary. The issue is that early in the diffusion process, i.e. when $t_i\ll 0$ , $\mathbf{z}_i$ and $\mathbf{x}_{i,..}^{\prime}$ are dominated by noise, but interpolation and mipmapping change the variance in the pixels $\mathbf{x}_{i,..}$ , thereby moving $\mathbf{x}_{i,..}^{\prime}$ out of the training distribution of $\epsilon_{\theta}^{(t_i)}$ . Instead, we retrieve only the nearest texture pixels for diffusion, and resolve aliasing and interpolation via a simple distillation post-process (Sec. 4.2.3). For each background pixel in rendered image $\mathbf{x}_{i,..}^{\prime}$ , we apply a Gaussian noise of standard deviation $\sigma_{i}$ such that the background matches the marginal distribution at diffusion step $t_i$ . This ensures that the diffusion model $f_{\theta}^{(t_i)}$ focuses solely on the foreground pixels of $\mathbf{x}_{i,..}^{\prime}$ . + +Note that in this setting a rendered image is simply a selection of pixels from the texture maps, and thus we can leverage backpropagation to easily implement the inverse rendering function $\mathcal{R}^{-1}$ . Additionally, forward rasterization yields a depth map of the scene, which we use as a conditioning input to the diffusion model. + +![](images/d5c6ef76b5bab196f4581d3541c42694133e12b8572ae7ea659f5f4f80c147e1.jpg) +Figure 4: Left: Depth map for conditioning SD2-depth, and quality image computed from screen space derivatives. Right: Output of SD2 decoder in two views using renders of the latent texture map as input. Note how the horizontal line across the doors changes appearance from one view to another. + +# 4.2.2 Aggregating Latent Textures. + +Recall from Sec. 4.1, before iterating through cameras $\{C_1,\dots,C_N\}$ , we first initialize a latent texture map $\mathbf{z}_{i - 1,0} = \mathbf{z}_i$ and render latent image $\mathbf{x}_{i,1}^{\prime} = \mathcal{R}(\mathbf{z}_{i - 1,0})$ . Then, in each step (iterating through cameras), we ob + +tain $\mathbf{x}_{i - 1,n}$ from $f_{\theta}^{(t_i)}(\mathbf{x}_{i,n}^{\prime})$ , and inverse render it to get $\mathbf{z}_{i - 1,n}^{\prime} = \mathcal{R}^{-1}(\mathbf{x}_{i - 1,n})$ . Finally, we need to aggregate the partial texture map $\mathbf{z}_{i - 1,n}^{\prime}$ with $\mathbf{z}_{i - 1,n - 1}$ to obtain a partially updated texture map $\mathbf{z}_{i - 1,n}$ . We perform this aggregation step based on a simple heuristic that the value of each pixel $(u,v)$ in $\mathbf{z}$ should be determined by the camera that has the most "direct and up-close" view to its corresponding point on the mesh. We measure view quality using image-space derivatives - the amount of change in UV coordinates per infinitesimal change in the image coordinate. This quantity is commonly used for determining mipmapping resolution and anti-aliasing, and it can be efficiently computed by nvdiffrast when rasterizing $\mathbf{z}_{i,n - 1}$ . For each pixel location $(p,q)$ in $\mathbf{x}_{i - 1,n}$ , we compute the negative Jacobian magnitude as $-|\frac{\partial u}{\partial p}\cdot \frac{\partial v}{\partial q} -\frac{\partial u}{\partial q}\cdot \frac{\partial v}{\partial p} |,$ and inverse render it to the texture space, which we denote as $Q_{i,n}$ . Higher values of $Q_{i,n}(u,v)$ means that camera $n$ has a better view of $(u,v)$ (see Fig.4). In addition, we compute $M_{i,n} = \mathcal{R}^{-1}(\boldsymbol {I})$ , such that $M_{i,n}(u,v)$ represents the number of pixels in $\mathbf{x}_{i,n}^{\prime}$ that received value from $\mathbf{z}_i(u,v)$ + +While iterating through cameras 1 through $N$ , we maintain mask $M_{i}$ (initialized to 0), which is used to track which pixels of $\mathbf{z}_{i,n}$ have been seen by cameras up to $n$ , and quality buffer $Q_{i}$ (initialized to $-\inf$ ), which is used to track the highest non-zero value of $Q_{i,n}$ at each $(u,v)$ . The value at $(u,v)$ of $\mathbf{z}_{i-1,n}$ is determined as: + +$$ +\mathbf {z} _ {i - 1, n} (u, v) = \left\{ \begin{array}{l l} \frac {\mathbf {z} _ {i - 1 , n} ^ {\prime} (u , v)}{M _ {i , n} (u , v)} & M _ {i, n} (u, v) > 0, \text {a n d} \\ \mathbf {z} _ {i - 1, n - 1} (u, v) & Q _ {i, n} (u, v) > Q _ {i} (u, v) \\ \end{array} \right. \tag {6} +$$ + +$M_{i}$ and $Q_{i}$ are then updated pixelwise as $M_{i} = \min (M_{i,n} + M_{i},1)$ and $Q_{i} = \max (Q_{i,n},Q_{i})$ (min and max are applied element-wise). We use Eqn. 6 in conjunction with Eqn. 5 in SIMS to perform denoising of images in the sequence of camera views. + +We further note that a side effect of using nearest pixel filtering during texture sampling in SIMS is aliasing. When the image space derivative Jacobian is much higher than 1 (low quality), neighboring pixels in screen space will have gaps when mapped to uv space. This results in incomplete filling of the latent texture map. Conversely, when the image space derivative Jacobian magnitude is much smaller than 1 (high quality), multiple screen space pixels $(p,q)$ will map to the same $(u,v)$ , creating uniform blocks in $\mathbf{x}$ . Similar to interpolated latents, these blocks are particularly detrimental early during the diffusion process, as they cannot be correctly denoised by $\epsilon_{\theta}$ which is expecting high spatial variance. Our quality-based aggregation overwrites low-quality regions in the latent texture when higher quality views are available. Thus we can set the resolution of the texture map such that the image space derivatives in each view do not fall below 1 (i.e. $\max Q_{i,n} < -1$ ) to prevent + +aliasing of the second type, and rely on cameras with better views to fix aliasing of the first type. + +# 4.2.3 From Latent to RGB Textures + +So far, we have described how to use SIMS to produce a set of 3D consistent latent images and latent texture maps, but have yet to describe how to translate this into a RGB texture map that can be rendered for viewing. Towards this, we experimented with multiple approaches and found performing multi-view distillation of decoded latent images $\{\mathbf{x}_{0,n}\}_{n=1}^{N}$ with a neural color field to be most performant. Specifically, we use the decoder $\mathcal{D}$ of Stable Diffusion to decode latent multi-view images $\{\mathbf{x}_{0,n}\}_{n=1}^{N}$ into RGB multi-view images $\{\xi_{n} = \mathcal{D}(\mathbf{x}_{0,n})\}_{n=1}^{N}$ . Notably, decoding with $\mathcal{D}$ introduce inconsistencies such that $\{\xi_{n}\}_{n=1}^{N}$ is not 3D consistent even when $\{\mathbf{x}_{0,n}\}_{n=1}^{N}$ is 3D consistent (see Fig. 4 for example). In stable Diffusion, each latent vector (pixel of $\mathbf{x}$ ) needs to carry the information of a $8 \times 8$ patch of RGB pixels. Thus, the value of the latent vector encodes both color values and spatial patterns. We therefore cannot expect their decoded results to be equivariant to perspective transformations. To address this problem, we leverage a neural color field optimized with appearance-based losses to smooth out inconsistencies. Since we know the ground-truth camera poses, we can directly obtain the 3D spatial coordinates $\{xyz_{n}\}_{n=1}^{N}$ of all pixels of $\{\xi_{n}\}_{n=1}^{N}$ by projecting pixels to the mesh surface. Background pixels that do not intersect any surface are discarded. Following [40], we use a multi-resolution hash encoding based on instantNGP [51] along with a shallow MLP $f_{\phi}$ to parameterize a function from 3D spatial coordinates to RGB values for each sample $rgb = f_{\phi}(hash(xyz))$ . We then distill multi-view images $\{\xi_{n}\}_{n=1}^{N}$ into this parametric function via optimization. Since our goal is to export $\phi$ into a texture map, we do not use any view-dependent parameterization. To reconcile inconsistencies, we use both a standard $L2$ loss and a VGG-based perceptual loss, applied between $f_{\phi}$ and $\{\xi_{n}\}_{n=1}^{N}$ , to train $\phi$ . We use Adam with a learning rate of 0.01, and optimization of $\phi$ converges within 100 iterations. After optimization, we compute the spatial coordinate of the centers of each pixel in a high-resolution texture map, and query $f_{\phi}$ to predict RGB values for the texture map. + +# 4.2.4 Geometry Processing, Camera, and Multi-resolution Refinement + +We normalize $\mathcal{M}$ such that it fits inside a cube of side length 1, and center it at the origin. Perspective cameras are placed facing the origin, and their FOV is adjusted to fit the object. Detailed parameters can be found in the appendix. As the diffusion model relies on context captured in $\mathbf{x}_i$ , to perform denoising, camera views that are too small w.r.t the size of the object often result in content drift - the texture in distant areas of the mesh can have content that is seman + +tically inconsistent. This can be seen as our version of the Janus face problem known to Dreamfusion and similar approaches [40, 59, 84]. Our solution is to perform two rounds of TexFusion at different resolution scales to obtain high-resolution textures while retaining semantic consistency. + +Specifically, we first run TexFusion using cameras that cover the entire object, and a low-resolution latent texture map that is suitable for these cameras. We do not run view distillation in this first round as we are only interested in the latent texture map. We denote the denoised latent texture map from this step as $\mathbf{z}_{0,lr}$ . We then use a second set of cameras with a narrower field of view; these cameras are also more numerous to still cover the full object. We determine a new texture map resolution using the square of the ratio of the tangent of the old FOV over the new FOV - corresponding to the relative change in surface area covered by the camera before and after the change in FOV. We then up-sample $\mathbf{z}_{0,lr}$ with nearest neighbor filtering to this higher resolution, and stochastically encode it to a partially noised state (e.g. $T = 500$ in the diffusion model time schedule). The second round of TexFusion uses these new cameras, and initializes SIMS with the partially noised latent texture map. The multi-view images produced by the second round of SIMS is used to produce the final output texture map via neural color field distillation. + +# 5. Experiments + +We apply TexFusion on various geometries and text prompts to evaluate its ability to produce high quality, natural, and 3D consistent textures. We focus our experimental comparisons on TEXTure [62], a text-driven texture generation method that also uses SD2-depth. We choose TEXTure as the baseline because (1) it represents the current state-of-the-art for language-conditioned texture generation, (2) it uses the same SD2-depth model, which can be prompted for a wide variety of content, and (3) it is concurrent to our work. In the supplementary materials, we further compare TexFusion to SDS-based text-driven texture distillation [48, 59], leveraging the Stable Diffusion model and show that TexFusion achieves superior results in terms of quality and speed. + +Dataset We collect 35 meshes of various content, and write 2-4 prompts for each mesh. In total, we evaluate each texture synthesis method on 86 mesh-text pairs. More details of this dataset are in the supplementary materials. + +# 5.1. Qualitative Comparisons + +We visualize textures produced by TexFusion on multiple geometries and prompts, and compare to state-of-the-art baselines in Figure 6. We render each object in 3 surrounding views to allow better examination of 3D consistency of the produced texture. We use Blender's Cycles renderer + +with a studio light-setup. Textures are applied as base color to a diffuse material. Additional visualizations, including videos showing 360 pans of all objects presented in the paper, and renders of our results using only texture color, can be found in the supplementary materials. + +In terms of local 3D consistency (consistency in neighborhoods on the surface of the mesh), textures produced by TexFusion are locally consistent - there are no visible seam lines or stitching artifacts. In contrast, we often find severe artifacts when viewing the top and back sides of TEXTure's outputs. These artifacts are most noticeable when a clean color is expected, such as when texturing vehicles. In terms of global consistency (semantic coherency of the entire texture, e.g. exactly 2 eyes and 1 nose to a face), TEXTure performs poorly and suffers from problems similar to Dream-Fusion's Janus face problem [59]: as the 2D diffusion model captures context of the object through its own camera view, it is not aware of the appearance of the opposite side of the object. This problem is ameliorated in TexFusion due to frequent communication between views in SIMS. + +There are noticeably more baked specular highlights and shadows in textures generated by TEXTure. These effects are physically inaccurate as they conflict with the lighting effects simulated by the renderer. In contrast, TexFusion produces textures that are smoothly illuminated. We hypothesize that this is due to interlacing aggregations in SIMS, which removes view-dependent effects as they arise in the sampling process. We provide additional visualizations of TexFusion in Fig. 5 to showcase how text prompting is effective at controlling the generation of textures, thereby producing varying appearances using the same mesh. + +Runtime TexFusion takes approximately 3 minutes on a machine with a single GPU to sample one texture. We are slightly faster than TEXTure (5 min.) [62], since we only need to optimize a color field once after sampling is + +![](images/c2c4ddf721294c7501817252ac0d1318e1bbd0ee17be264ff308a64f8ffa5ed5.jpg) +Figure 5: More TexFusion text-conditioned texturing results. + +![](images/fae4b983bc70a652edb2ecabf272f2b8cd3230dec047727c9935b6d93110a90d.jpg) + +![](images/29f8503ae5077a06264fb3f0e19107c2cf68b29efd8d9cd279da2fa35f71d8d2.jpg) +Figure 6: Visual comparison of textures generated by TEXTure [62] and TexFusion. + +
MethodFID (↓)User study (%)
Natural Color (↑)More Detailed (↑)Less Artifacts (↑)Align with Prompt (↑)
TEXTure79.4724.4265.1231.4043.02
TexFusion59.7875.5834.8868.6056.98
+ +Table 1: Quantitative and qualitative comparisons between TEX-Ture and TexFusion. FID is computed w.r.t. a set of images synthesized by SD2-depth. + +complete. We are also an order of magnitude faster than methods that rely on SDS loss which report upwards of 40 minutes [40, 48, 59]. Additional runtime comparisons are in the appendix. + +# 5.2. Quantitative Comparisons + +It is inherently difficult to quantitatively evaluate the quality of text and geometry-conditioned texture generation as one must take alignment with the conditioning inputs into account. We rely on both automated and human evaluation to gauge the quality of the synthesized textures. + +FID Since SD2-depth can produce high-quality images that match the structure of the conditioning geometry, we sample it to create a proxy ground truth set. Furthermore, as both TEXTure and TexFusion use SD2-depth as 2D prior, samples from it serve as an upper bound to the quality and diversity of the appearance of textures: We render the depth map of all meshes in 8 canonical views, and condition SD2-depth on both these depth maps and our text + +prompts to generate a ground truth set. We also render textures generated by both methods in these same 8 views, and compute the Fréchet Inception Distance (FID) between images rendered from the textures and the ground truth set. FID measures the overlap between two sets of images based on their similarity in a learned feature space. We white out the background regions in rendered images and ground truth images to focus the comparisons to the textured object. We present the results in Tab. 1. FID scores indicate that textures synthesized by TexFusion are closer in appearance to the ground truth set than those from TEXTure. The absolute FID values are likely high due to the relatively small set of images in each dataset. + +User study We conducted a user study to measure the overall quality of the generated textures. We use all mesh and text prompts in our dataset, and render each textured mesh into a video showing the results from a $360^{\circ}$ rotating view. We present videos from TEXTure and TexFusion side by side in random left-right order, to avoid bias. The user study was conducted on Amazon Mechanical Turk. Each participant is asked to select the preferred result under four criteria: natural color (non-saturation), details, cleanliness, and alignment with the text prompt. To avoid randomness in the answers, we let 3 participants answer the same question and determine the choice by max-voting, a question screenshot is provided in the supplement. We provide the user study results in Tab. 1. Although not deemed as detailed as TEXTure, our results are overall preferred by the users in terms of better aligning with the provided text prompt, and + +![](images/8fabd9c349e734537082503eedec140df2c7aba5d26cf9c1ac6e33a0da769ab6.jpg) +"red backpack" + +![](images/ab5e0688b22c38345ddbb3a105cc3367b15da5e8e4c1056f4c9345b053452440.jpg) +"Portrait of a humanoid robot, futuristic, science fiction" + +![](images/0221b34bce87f77084fce0a3f15edb6f11a4b3273ec4371173a5f464fead88ca.jpg) +"Medieval celtic Iouse, stone bricks, wooden roof" + +![](images/780bba85ba986bbae9197888854e1c8eeb585f8134b439a4048be95443f9a1ea.jpg) + +![](images/21f3936efeae468b7e7e07f9cc89d4d70ee5347dd9be49c0c0d51be4d453f95b.jpg) +"camper bag, camouflage" + +![](images/d99f9e40c816035f068e488dd2283d0b354039b1e9ae53e62f151cf6133393f9.jpg) +"Brown rabbit" + +![](images/f1d973a6d898dbf7da172e05271a43ef4ba51b399f52a44b5de39bffa6e3716a.jpg) +"Crocodile skin handbag" +"Blonde girl with green eyes, hair in tied a bun, DSLR portrait photo." + +![](images/9d49a98b80f4182b08cfee95b73e69b89ed8554776369f675d90b41f6c204a26.jpg) +"Leather lounge chair" +Figure 7: TexFusion + non-stochastic DDIM sampling $(\sigma_{t_i} = 0)$ . This setting emphasizes textural details, such as the leather (bottom, left), roof shingles (top, right) + +having more natural color, and fewer artifacts/flaws. + +# 5.3. Improving Texture Details + +Texfusion can produce more texture details with adjusted hyperparameters and diffusion backbones. First, we find that using the non-stochastic version of DDIM $(\eta = 0)$ adds more materialistic details to the textures on smooth/low-poly geometries. We showcase some examples with particularly large improvements in Fig. 7. Second, we explore the use of ControlNet[98], which can be easily substituted as the LDM backbone for our method without any additional changes. We find that its high-resolution depth conditioning allows TexFusion to capture fine-grained geometric details in the input mesh. In Fig. 8, we further compare TexFusion + ControlNet in "normal mode" (apply classifier-free guidance to text prompt only) and ControlNet's "guess mode" (apply classifier-free guidance to both text and depth) on meshes with fine geometric details. TexFusion produces high-contrast textures with the appearance of strong lighting under "guess mode", and realistic textures with smooth lighting in "normal mode". These modifications improve details at a cost of reduced robustness. The non-stochastic DDIM setting may create artifacts when a smooth and clean texture is desired. On the other hand, the increase in depth resolution makes TexFusion+ControlNet susceptible to capturing face boundaries on low-poly meshes. We provide visualizations of these failure cases in Fig. 9. Nonetheless, these modifications offer further dimensions of control that can be explored by practitioners. + +# 6. Conclusion and Limitations + +We presented TexFusion, a novel approach to text-driven texture synthesis for given 3D geometries, using only large-scale text-guided image diffusion models. TexFusion lever + +![](images/8131710eb84cee6fdd39c315df4f2a089d0175148a010ede602cde4daaaa5859.jpg) +TexFusion + ControlNet in "normal" mode +"Portrait of greek-egyptian deity hermanubis, lapis skin and gold clothing" +Figure 8: Left: TexFusion + ControlNet in "normal mode"; right: TexFusion + ControlNet in "guess mode". + +![](images/342524332e5f027f4c6422333f0c6a22293ac48ccb4a9a180ea6d72f42b3ca4f.jpg) +TexFusion + ControlNet in "guess" mode + +![](images/18cfcce3440166573610ceeeaf21b6025f2c0152d1abc84a53ec672a0419e388.jpg) +Deterministic DDIM: Artifacts in smooth areas +Figure 9: Failure modes exhibited by (left) TexFusion + non-stochastic DDIM sampling and (right) TexFusion + ControlNet. + +![](images/ea750aea5dd5f9ba5149e88a15d5419e000fc683e1514c0df820f9b952159cb8.jpg) +TexFusion+ControlNet: Face boundaries on low poly meshes + +![](images/4f29b5525b3af72c1dc596a52f30a2b77334ba439db004cbbaa2d781d7e7616c.jpg) + +ages the latent diffusion model Stable Diffusion and relies on a new 3D-consistent diffusion model sampling scheme that runs a 2D denoiser network in different views of the object and aggregates the predictions in a latent texture map after each denoising iteration. We find that our method can efficiently generate diverse, high-quality, and globally coherent textures and offers detailed control through text conditioning. Limitations of our approach include a not-yet-ideal sharpness and that texture generation is not real-time. Future work can address these issues; for instance, it would be interesting to leverage the recent literature on faster diffusion model samplers [15, 41, 42, 99]. + +Concurrent Works Since Submission Since the time of writing this paper, new works has appeared in the text to texture space [8, 79]. They improve upon TEXTure in aspects such as adding a refinement procedure to automatically fix low quality areas from the initial texturing process [8], and using images sampled from a text-to-image model and per-prompt finetuning to provide stronger conditioning [79]. TexFusion is methodologically distinct from these methods which are based-on TEXTure. Improvements proposed in these works could be combined with TexFusion in future work. + +Acknowledgements We would like to thank Jun Gao for the helpful discussions during the project. Tianshi Cao acknowledges additional income from Vector Scholarships in Artificial Intelligence, which are not in direct support of this work. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In ICML, 2018. 1 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1, 2 +[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 3 +[4] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh Susskind. Gaudi: A neural architect for immersive 3d scene generation. arXiv preprint arXiv:2207.13751, 2022. 1, 2 +[5] Hanqun Cao, Cheng Tan, Zhangyang Gao, Guangyong Chen, Pheng-Ann Heng, and Stan Z. Li. A survey on generative diffusion model. arXiv preprint arXiv:2209.02646, 2022. 2 +[6] Ignacio Castaño. thekla交通大学 In github.com/Thekla/thekla_italas, 2015. 4 +[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2 +[8] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 9 +[9] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 100–116. Springer, 2019. 2 +[10] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3 +[11] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In The Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[12] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. arXiv preprint arXiv:2209.04747, 2022. 2 +[13] Congyue Deng, Chiyu "Max" Jiang, Charles R. Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, and Dragomir Anguelov. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. arXiv + +preprint arXiv:2212.03267, 2022. 2 +[14] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In Advances in Neural Information Processing Systems, 2021. 2 +[15] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. GENIE: Higher-Order Denoising Diffusion Solvers. In Advances in Neural Information Processing Systems, 2022. 9 +[16] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Score-based generative modeling with critically-damped Langevin diffusion. In International Conference on Learning Representations (ICLR), 2022. 2 +[17] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems. 1, 2 +[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, 2020. 3 +[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. arXiv preprint arXiv:2106.15282, 2021. 2 +[22] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2, 3 +[23] Wenlong Huang, Brian Lai, Weijian Xu, and Zhuowen Tu. 3d volumetric modeling with introspective neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33(01), pages 8481-8488, 2019. 1 +[24] Aapo Hyvarinen. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6:695-709, 2005. 3 +[25] Moritz Ibing, Isaak Lim, and Leif P. Kobbelt. 3d shape generation with grid-based implicit functions. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1 +[26] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022. 3 +[27] Nikolai Kalischek, Torben Peters, Jan D Wegner, and Konrad Schindler. Tetrahedral diffusion models for 3d shape generation. arXiv preprint arXiv:2211.13220, 2022. 1, 2 +[28] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, 2022. 3 +[29] Carson Katri. Dream textures. In github.com/carson-katri/dream-textures, 2022. 4 +[30] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming + +Song. Denoising diffusion restoration models. arXiv preprint arXiv:2201.11793, 2022. 2 +[31] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Popa Tiberiu. Clip-mesh: Generating textured meshes from text using pretrained image-text models. December 2022. 3 +[32] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. 2020. 1 +[33] Wei-Jan Ko, Hui-Yu Huang, Yu-Liang Kuo, Chen-Yi Chiu, Li-Heng Wang, and Wei-Chen Chiu. Rpg: Learning recursive point cloud generation. arXiv preprint arXiv:2105.14322, 2021. 1 +[34] Johannes Kopf, Chi-Wing Fu, Daniel Cohen-Or, Oliver Deussen, Dani Lischinski, and Tien-Tsin Wong. Solid texture synthesis from 2d exemplars. ACM Trans. Graph., 26(3):2-es, jul 2007. 2 +[35] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5 +[36] Sylvain Lefebvre and Hugues Hoppe. Appearance-space texture synthesis. ACM Trans. Graph., 25(3):541-548, jul 2006. 2 +[37] Chun-Liang Li, Manzil Zaheer, Yang Zhang, Barnabas Poczos, and Ruslan Salakhutdinov. Point cloud gan. arXiv preprint arXiv:1810.05795, 2018. 1 +[38] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2 +[39] Ruihui Li, Xianzhi Li, Ke-Hei Hui, and Chi-Wing Fu. SPGAN:sphere-guided 3d shape generation and manipulation. ACM Transactions on Graphics (Proc. SIGGRAPH), 40(4), 2021. 1 +[40] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 1, 2, 6, 7, 8 +[41] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 9 +[42] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv:2206.00927, 2022. 9 +[43] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. arXiv preprint arXiv:2201.09865, 2022. 2 +[44] A. Luo, T. Li, W. Zhang, and T. Lee. Surfgen: Adversarial 3d shape synthesis with explicit surface discriminators. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1 +[45] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern + +Recognition (CVPR), 2021. 1, 2 +[46] Siwei Lyu. Interpretation and generalization of score matching. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, UAI '09, page 359-366, Arlington, Virginia, USA, 2009. AUAI Press. 3 +[47] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2 +[48] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. 1, 2, 4, 7, 8 +[49] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2, 3 +[50] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 1 +[51] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 6 +[52] Gimin Nam, Mariem Khlifi, Andrew Rodriguez, Alberto Tono, Linqi Zhou, and Paul Guerrero. 3d-ldm: Neural implicit 3d shape generation with latent diffusion models. arXiv preprint arXiv:2212.00842, 2022. 1, 2 +[53] Charlie Nash, Yaroslav Ganin, S. M. Ali Eslami, and Peter W. Battaglia. *Polygen: An autoregressive generative model of 3d meshes.* ICML, 2020. 1 +[54] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2 +[55] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1, 2 +[56] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, 2021. 2 +[57] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 2 +[58] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1 +[59] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 1, 2, 7, 8 + +[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR, 2021. 3 +[61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2 +[62] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 3, 4, 7, 8 +[63] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 1, 2, 3, 4, 5 +[64] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image dissusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2 +[65] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. arXiv preprint arXiv:2111.05826, 2021. 2 +[66] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2 +[67] Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. arXiv preprint arXiv:2104.07636, 2021. 2 +[68] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations (ICLR), 2022. 3 +[69] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. arXiv preprint arXiv:2110.02624, 2021. 1 +[70] Hiroshi Sasaki, Chris G. Willcocks, and Toby P. Breckon. UNIT-DDPM: Unpaired image translation with denoising diffusion probabilistic models. arXiv preprint arXiv:2104.05358, 2021. 2 +[71] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3859-3868, 2019. 1 +[72] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary + +Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. arXiv preprint arXiv:2211.16677, 2022. 1, 2 +[73] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In Shai Avidan, Gabriel J. Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, volume 13663 of Lecture Notes in Computer Science, pages 72-88. Springer, 2022. 2 +[74] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015. 3 +[75] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3 +[76] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3 +[77] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. arXiv preprint arXiv:2203.08382, 2022. 2 +[78] Yongbin Sun, Yue Wang, Ziwei Liu, Joshua E Siegel, and Sanjay E Sarma. Pointgrow: Autoregressively learned point cloud generation with self-attention. In Winter Conference on Applications of Computer Vision, 2020. 1 +[79] Zhibin Tang and Tiantong He. Text-guided high-definition consistency texture model. ArXiv, abs/2305.05901, 2023. 9 +[80] Greg Turk. Texture synthesis on surfaces. SIGGRAPH '01, page 347-354, New York, NY, USA, 2001. Association for Computing Machinery. 2 +[81] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In Neural Information Processing Systems (NeurIPS), 2021. 2 +[82] Diego Valsesia, Giulia Fracastoro, and Enrico Magli. Learning localized generative models for 3d point clouds via graph convolution. In International Conference on Learning Representations (ICLR) 2019, 2019. 1 +[83] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 23(7):1661-1674, 2011. 3 +[84] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022. 1, 2, 7 +[85] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jian-min Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135, 2022. 1, 2 +[86] Li-Yi Wei and Marc Levoy. Texture synthesis over arbitrary manifold surfaces. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '01, page 355-360, New York, NY, USA, 2001. Association for Computing Machinery. 2 + +[87] Cheng Wen, Baosheng Yu, and Dacheng Tao. Learning progressive point embeddings for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10266-10275, 2021. 1 +[88] Li-Yi Wie, Sylvain Lefebvre, Vivek Kwatra, and Greg Turk. State of the Art in Example-based Texture Synthesis. In M. Pauly and G. Greiner, editors, Eurographics 2009 - State of the Art Reports. The Eurographics Association, 2009. 2 +[89] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 1 +[90] Fanbo Xiang, Zexiang Xu, Milos Hasan, Yannick Hold-Geoffroy, Kalyan Sunkavalli, and Hao Su. Neutex: Neural texture mapping for volumetric neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7119-7128, 2021. 2 +[91] Jianwen Xie, Yifei Xu, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Zhu Song-Chun, and Ying Nian Wu. Generative pointnet: Deep energy-based learning on unordered point sets for 3d generation, reconstruction and classification. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1 +[92] Jianwen Xie, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Zhu Song-Chun, and Ying Nian Wu. Learning descriptor networks for 3d shape synthesis and analysis. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. +[93] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. PointFlow: 3D point cloud generation with continuous normalizing flows. 2019. 1 +[94] Ling Yang, Zhilong Zhang, and Shenda Hong. Diffusion models: A comprehensive survey of methods and applications. arXiv preprint arXiv:2209.00796, 2022. 2 +[95] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations. In Proceedings of International Conference on Computer Vision (ICCV), 2021. 1 +[96] Jonathan Young. xatlas. In github.com/jpcy/xatlas, 2016. 4 +[97] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2 +[98] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 9 +[99] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv:2204.13902, 2022. 9 +[100] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yi-nan Zhang, Antonio Torralba, and Sanja Fidler. Image {gan}s meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In International Conference on Learning Representations, 2021. 1 +[101] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation + +and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2 \ No newline at end of file diff --git a/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/images.zip b/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ec2dd0e2985e221e2fa87326a8ac0e09e4fda881 --- /dev/null +++ b/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:259df99b329c677f14029dac77b60f14af4428f18a97a4c8c27cf0427bafff2d +size 479451 diff --git a/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/layout.json b/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..647ff2f876ddcb22aff4aef76228765064cd2bdc --- /dev/null +++ b/2023/TexFusion_ Synthesizing 3D Textures with Text-Guided Image Diffusion Models/layout.json @@ -0,0 +1,13330 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 71, + 507, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 71, + 507, + 106 + ], + "spans": [ + { + "bbox": [ + 86, + 71, + 507, + 106 + ], + "type": "text", + "content": "TexFusion: Synthesizing 3D Textures with Text-Guided Image Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 127, + 142, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 127, + 142, + 140 + ], + "spans": [ + { + "bbox": [ + 60, + 127, + 142, + 140 + ], + "type": "text", + "content": "Tianshi Cao " + }, + { + "bbox": [ + 60, + 127, + 142, + 140 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 128, + 236, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 128, + 236, + 142 + ], + "spans": [ + { + "bbox": [ + 160, + 128, + 236, + 142 + ], + "type": "text", + "content": "Karsten Kreis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 253, + 128, + 332, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 128, + 332, + 142 + ], + "spans": [ + { + "bbox": [ + 253, + 128, + 332, + 142 + ], + "type": "text", + "content": "Sanja Fidler" + }, + { + "bbox": [ + 253, + 128, + 332, + 142 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 351, + 128, + 436, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 128, + 436, + 143 + ], + "spans": [ + { + "bbox": [ + 351, + 128, + 436, + 143 + ], + "type": "text", + "content": "Nicholas Sharp" + }, + { + "bbox": [ + 351, + 128, + 436, + 143 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 455, + 128, + 531, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 455, + 128, + 531, + 142 + ], + "spans": [ + { + "bbox": [ + 455, + 128, + 531, + 142 + ], + "type": "text", + "content": "Kangxue Yin" + }, + { + "bbox": [ + 455, + 128, + 531, + 142 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 190, + 143, + 233, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 143, + 233, + 156 + ], + "spans": [ + { + "bbox": [ + 190, + 143, + 233, + 156 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 190, + 143, + 233, + 156 + ], + "type": "text", + "content": " NVIDIA," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 242, + 144, + 330, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 144, + 330, + 156 + ], + "spans": [ + { + "bbox": [ + 242, + 144, + 330, + 156 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 242, + 144, + 330, + 156 + ], + "type": "text", + "content": " University of Toronto," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 144, + 403, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 144, + 403, + 155 + ], + "spans": [ + { + "bbox": [ + 339, + 144, + 403, + 155 + ], + "type": "text", + "content": "3 Vector Institute" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 159, + 445, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 159, + 445, + 171 + ], + "spans": [ + { + "bbox": [ + 142, + 159, + 445, + 171 + ], + "type": "text", + "content": "{tianshic, kkreis, sfidler, nsharp, kangxuey}@nvidia.com" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 107, + 196, + 203, + 272 + ], + "blocks": [ + { + "bbox": [ + 107, + 196, + 203, + 272 + ], + "lines": [ + { + "bbox": [ + 107, + 196, + 203, + 272 + ], + "spans": [ + { + "bbox": [ + 107, + 196, + 203, + 272 + ], + "type": "image", + "image_path": "2be179c4a496587e538c93d3ec1f159a45848594f67b3190a66c184d9434b3db.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 219, + 205, + 287, + 273 + ], + "blocks": [ + { + "bbox": [ + 219, + 205, + 287, + 273 + ], + "lines": [ + { + "bbox": [ + 219, + 205, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 219, + 205, + 287, + 273 + ], + "type": "image", + "image_path": "ee7570178093fd5858103cc7570b20f724afff240872004e423356b2fbffdde4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 288, + 205, + 343, + 274 + ], + "blocks": [ + { + "bbox": [ + 232, + 274, + 293, + 281 + ], + "lines": [ + { + "bbox": [ + 232, + 274, + 293, + 281 + ], + "spans": [ + { + "bbox": [ + 232, + 274, + 293, + 281 + ], + "type": "text", + "content": "\"Crocodile skin handbag\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 288, + 205, + 343, + 274 + ], + "lines": [ + { + "bbox": [ + 288, + 205, + 343, + 274 + ], + "spans": [ + { + "bbox": [ + 288, + 205, + 343, + 274 + ], + "type": "image", + "image_path": "740797af04e6edcb45341dbbc02b0a26e86ca97f6ad3853b2fe850f064762177.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 350, + 205, + 402, + 273 + ], + "blocks": [ + { + "bbox": [ + 350, + 205, + 402, + 273 + ], + "lines": [ + { + "bbox": [ + 350, + 205, + 402, + 273 + ], + "spans": [ + { + "bbox": [ + 350, + 205, + 402, + 273 + ], + "type": "image", + "image_path": "41f09a7a345570b2c8ab89b5f6c8376f9f8691b3c7df5d547e30e740731d7119.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 273, + 420, + 279 + ], + "lines": [ + { + "bbox": [ + 350, + 273, + 420, + 279 + ], + "spans": [ + { + "bbox": [ + 350, + 273, + 420, + 279 + ], + "type": "text", + "content": "\"White Mercedes Benz SUV\"" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 408, + 217, + 488, + 274 + ], + "blocks": [ + { + "bbox": [ + 408, + 217, + 488, + 274 + ], + "lines": [ + { + "bbox": [ + 408, + 217, + 488, + 274 + ], + "spans": [ + { + "bbox": [ + 408, + 217, + 488, + 274 + ], + "type": "image", + "image_path": "9331011979d2a6b5de089c668e779b29776a486c708bef2efd0f19252c643358.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 106, + 287, + 214, + 354 + ], + "blocks": [ + { + "bbox": [ + 107, + 273, + 212, + 287 + ], + "lines": [ + { + "bbox": [ + 107, + 273, + 212, + 287 + ], + "spans": [ + { + "bbox": [ + 107, + 273, + 212, + 287 + ], + "type": "text", + "content": "\"White humanoid robot, movie poster, main character of a science fiction movie\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 106, + 287, + 214, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 287, + 214, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 214, + 354 + ], + "type": "image", + "image_path": "80ccf636a0e14e45a7d3f86fbe0e0260c4d4b6cc0c8df82cd2605f65436a4ba8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 356, + 202, + 370 + ], + "lines": [ + { + "bbox": [ + 116, + 356, + 202, + 370 + ], + "spans": [ + { + "bbox": [ + 116, + 356, + 202, + 370 + ], + "type": "text", + "content": "\"Oil painting of a bald, middle aged banker with pointed moustache\"" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 227, + 288, + 275, + 350 + ], + "blocks": [ + { + "bbox": [ + 227, + 288, + 275, + 350 + ], + "lines": [ + { + "bbox": [ + 227, + 288, + 275, + 350 + ], + "spans": [ + { + "bbox": [ + 227, + 288, + 275, + 350 + ], + "type": "image", + "image_path": "b67e6e0e19d4a3408ade25aa37a0ee05cd0f93e3952a4c9d33a7f208b38d9cc1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 287, + 284, + 342, + 351 + ], + "blocks": [ + { + "bbox": [ + 227, + 362, + 337, + 370 + ], + "lines": [ + { + "bbox": [ + 227, + 362, + 337, + 370 + ], + "spans": [ + { + "bbox": [ + 227, + 362, + 337, + 370 + ], + "type": "text", + "content": "\"Portrait photo of Abraham Lincoln, full color\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 287, + 284, + 342, + 351 + ], + "lines": [ + { + "bbox": [ + 287, + 284, + 342, + 351 + ], + "spans": [ + { + "bbox": [ + 287, + 284, + 342, + 351 + ], + "type": "image", + "image_path": "0b7ebf7c5ea315d5d0c2ac36573d8f1f493a3909a219c0670f512e1a9cf74dd4.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 356, + 281, + 410, + 360 + ], + "blocks": [ + { + "bbox": [ + 356, + 281, + 410, + 360 + ], + "lines": [ + { + "bbox": [ + 356, + 281, + 410, + 360 + ], + "spans": [ + { + "bbox": [ + 356, + 281, + 410, + 360 + ], + "type": "image", + "image_path": "b2e1084442a5131dd593d45319288578e40cee64bcbe020c8a08ea4619b49045.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 415, + 282, + 457, + 361 + ], + "blocks": [ + { + "bbox": [ + 415, + 282, + 457, + 361 + ], + "lines": [ + { + "bbox": [ + 415, + 282, + 457, + 361 + ], + "spans": [ + { + "bbox": [ + 415, + 282, + 457, + 361 + ], + "type": "image", + "image_path": "0ca7c7502335d034f196e2b5cd00e73bc6c3432995f6d26a5cef59671bd9caa8.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 361, + 451, + 369 + ], + "lines": [ + { + "bbox": [ + 356, + 361, + 451, + 369 + ], + "spans": [ + { + "bbox": [ + 356, + 361, + 451, + 369 + ], + "type": "text", + "content": "\"Railroad worker wearing high-vis vest\"" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 177, + 380, + 414, + 392 + ], + "lines": [ + { + "bbox": [ + 177, + 380, + 414, + 392 + ], + "spans": [ + { + "bbox": [ + 177, + 380, + 414, + 392 + ], + "type": "text", + "content": "Figure 1: Text-conditioned 3D texturing results with TexFusion." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 143, + 414, + 192, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 414, + 192, + 426 + ], + "spans": [ + { + "bbox": [ + 143, + 414, + 192, + 426 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 47, + 433, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 433, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 433, + 289, + 696 + ], + "type": "text", + "content": "We present TexFusion (Texture Diffusion), a new method to synthesize textures for given 3D geometries, using large-scale text-guided image diffusion models. In contrast to recent works that leverage 2D text-to-image diffusion models to distill 3D objects using a slow and fragile optimization process, TexFusion introduces a new 3D-consistent generation technique specifically designed for texture synthesis that employs regular diffusion model sampling on different 2D rendered views. Specifically, we leverage latent diffusion models, apply the diffusion model's denoiser on a set of 2D renders of the 3D object, and aggregate the different denoising predictions on a shared latent texture map. Final output RGB textures are produced by optimizing an intermediate neural color field on the decodings of 2D renders of the latent texture. We thoroughly validate TexFusion and show that we can efficiently generate diverse, high quality and globally coherent textures. We achieve state-of-the-art text-guided texture synthesis performance using only image diffusion models, while avoiding the pitfalls of previous distillation-based methods. The text-conditioning offers detailed control and we also do not rely on any ground truth 3D textures for training. This makes our method versatile" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 403, + 545, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 545, + 451 + ], + "type": "text", + "content": "and applicable to a broad range of geometry and texture types. We hope that TexFusion will advance AI-based texturing of 3D assets for applications in virtual reality, game design, simulation, and more." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 457, + 387, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 457, + 387, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 387, + 469 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 473, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 675 + ], + "type": "text", + "content": "In the past decade, deep learning-based 3D object generation has been studied extensively [1, 4, 18, 23, 25, 27, 32, 33, 37, 39, 44, 45, 50, 52, 53, 55, 58, 69, 71, 72, 78, 82, 85, 87, 89, 91-93, 95, 97, 100, 101], due to the demand for high-quality 3D assets in 3D applications such as VR/AR, simulation, digital twins, etc. While many prior works on 3D synthesis focus on the geometric components of the assets, textures are studied much less, despite the fact that they are important components of realistic 3D assets which assign colors and materials to meshes to make the rendering vivid. Recent advances in text-conditioned image diffusion models trained on internet-scale data [2, 54, 61, 63, 66] have unlocked the capability to generate images with stunning visual detail and practically unlimited diversity. These high-performance diffusion models have also been used as image priors to synthesize 3D objects with textures using textual guidance [40, 48, 59, 84]." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we aim to perform text-driven high-quality 3D texture synthesis for given meshes, by leveraging the information about the appearance of textures carried by the" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 132, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 132, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 132, + 712 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "type": "text", + "content": "4169" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "image prior of a pre-trained text-to-image diffusion model. The main workhorse in current text-to-3D methods that leverage 2D image diffusion models is Score Distillation Sampling (SDS) [59]. SDS is used to distill, or optimize, a 3D representation such that its renders are encouraged to be high-likelihood under the image prior. Methods utilizing SDS also share two common limitations in that: 1). a high classifier-free guidance weight [22] is required for the optimization to converge, resulting in high color saturation and low generation diversity; 2). a lengthy optimization process is needed for every sample." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 205, + 289, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 455 + ], + "type": "text", + "content": "To address the above issues, we present Texture Diffusion, or TexFusion for short. TexFusion is a sampler for sampling surface textures from image diffusion models. Specifically, we use latent diffusion models that efficiently autoencode images into a latent space and generate images in that space [63, 81]. TexFusion leverages latent diffusion trajectories in multiple object views, encoded by a shared latent texture map. Rendering of the shared latent texture map are provided as input to the denoiser of the latent diffusion model, and the output of every denoising step is projected back to the shared texture map in a 3D-consistent manner. To transform the generated latent textures into RGB textures, we optimize a neural color field on the outputs of the latent diffusion model's decoder applied to different views of the object. We use the publicly available latent diffusion model Stable-Diffusion with depth conditioning [63] (SD2-depth) as our diffusion backbone. Compared to methods relying on SDS, TexFusion produces textures with more natural tone, stronger view consistency, and is significantly faster to sample (3 minutes vs. 30 minutes reported by previous works)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 456, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 289, + 613 + ], + "type": "text", + "content": "We qualitatively and quantitatively validate TexFusion on various texture generation tasks. We find that TexFusion generates high quality, globally coherent and detailed textures that are well-aligned with the text prompts used for conditioning (e.g. Fig. 1). Since we leverage powerful text-to-image diffusion models for texture sampling, we can generate highly diverse textures and are not limited to single categories or restricted by explicit 3D textures as training data, which are limitations of previous works [7, 11, 18, 57, 73]. In summary, our main contribution is a novel method for 3D texture generation from 2D image diffusion models, that is view-consistent, avoids over-saturation and achieves state-of-the-art text-driven texture synthesis performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 624, + 134, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 134, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 134, + 636 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Classic Computer Graphics Techniques Early work on texture generation focused on tiling exemplar patterns across a surface, often with an explicit direction field for local orientation [34, 36, 80, 86]. See [88] for a survey. This research established the value of texture image representations and the challenges of global coherence, both central" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "to this work. However, modern learning-based priors have proven necessary to go beyond simple patterns and synthesize plausible shape-specific texture details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 119, + 547, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 119, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 119, + 547, + 324 + ], + "type": "text", + "content": "Texture Synthesis with 3D Priors Textures are defined on the meshed surface of 3D objects, which is an irregular representation. To enable 2D texture generation on 3D meshes, AUV-Net [11] learns an aligned UV space for a set of 3D shapes in a given class, mapping 3D texture synthesis to a 2D domain. Texturify [73] trains a 3D StyleGAN in the quad-parameterized surface domain on a set of textured 3D shapes in a given class. Different from AUV-NET or Texturify, which embed the geometric prior into a UV map or mesh parameterization, EG3D [7] and GET3D [18] directly train 3D StyleGANs to generate geometry and texture jointly, where the textures are implicit texture fields [57]. Other works also represent 3D texture by vertex colors [49], voxels [9], cube mapping [90], etc. In contrast to TexFusion, these works mostly don't offer text conditioning, often work only on single object categories or require textured 3D shapes for training, which limits their broad applicability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 334, + 547, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 547, + 502 + ], + "type": "text", + "content": "Diffusion Models Our approach directly builds on diffusion models [5, 12, 94], which have recently emerged as new state-of-the-art generative models. In particular, they have demonstrated outstanding performance in image generation [14-16, 21, 56, 81], outperforming GANs, and led to breakthroughs in text-to-image synthesis [2, 54, 61, 63, 66]. They have also been successfully used for a variety of image editing and processing tasks [17, 19, 30, 38, 43, 47, 64, 65, 67, 70, 77]. Moreover 3D object generation has been addressed with diffusion models, too, for instance leveraging point clouds [45, 55, 97, 101], meshes [27], or neural fields [4, 52, 72, 85] as 3D representations. However, these works focus on geometry generation and do not specifically tackle 3D texture synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "type": "text", + "content": "Distilling 3D Objects from 2D Image Diffusion Models Recently, large-scale 2D text-to-image diffusion models have been leveraged to distill individual 3D objects as neural radiance fields using Score Distillation Sampling (SDS) [13, 40, 48, 59, 84]. In SDS, the radiance field is rendered from different directions into 2D images and it is optimized such that each render has a high probability under the text-to-image diffusion model while conditioning on a text prompt. DreamFusion [59] pioneered the approach, Magic3D [40] proposed a coarse-to-fine strategy improving quality, and Latent-NeRF [48] performs distillation in latent space leveraging a latent diffusion model [63]. These approaches do not specifically target texture generation, which is the focus of this work. More deeply, a crucial drawback of this line of work is that SDS typically requires strong guidance [22] to condition the diffusion model, which can hurt quality and diversity. Moreover, SDS's" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "4170" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 62, + 414, + 215 + ], + "blocks": [ + { + "bbox": [ + 48, + 62, + 414, + 215 + ], + "lines": [ + { + "bbox": [ + 48, + 62, + 414, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 62, + 414, + 215 + ], + "type": "image", + "image_path": "62a8b985cad542cd68bad8c5a63f5d957d8a5bf86481af1571019c27b918e07d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 64, + 547, + 76 + ], + "lines": [ + { + "bbox": [ + 415, + 64, + 547, + 76 + ], + "spans": [ + { + "bbox": [ + 415, + 64, + 547, + 76 + ], + "type": "text", + "content": "Distill with Neural Color Field" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 421, + 77, + 547, + 214 + ], + "blocks": [ + { + "bbox": [ + 421, + 77, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 421, + 77, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 421, + 77, + 547, + 214 + ], + "type": "image", + "image_path": "5e6ac964bdbdf3700fd301f627c54e9cd45b942d2e2821452810a02e111b4ffd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 218, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 218, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 547, + 275 + ], + "type": "text", + "content": "Figure 2: Overview of TexFusion. TexFusion takes a text prompt and mesh geometry as input and produces a UV parameterized texture image matching the prompt and mesh. Key to TexFusion is the Sequential Interlaced Multiview Sampler (SIMS) - SIMS performs denoising diffusion iterations in multiple camera views, yet the trajectories are aggregated through a latent texture map after every denoising step. SIMS produces a set of 3D consistent latent images (TexFusion uses Stable Diffusion [63] as text-to-image diffusion backbone), which are decoded and fused into a texture map via optimizing an intermediate neural color field." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 283, + 288, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 288, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 288, + 356 + ], + "type": "text", + "content": "iterative optimization process makes synthesis very slow. In contrast, our approach avoids SDS entirely and leverages regular diffusion model sampling in a new, 3D-consistent manner. Earlier works also leverage CLIP [60] for 3D object or texture distillation [10, 26, 31, 49], but this performs usually worse than using diffusion models instead." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 368, + 289, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 289, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 289, + 620 + ], + "type": "text", + "content": "Concurrent Work Concurrently with this work, TEX-Ture [62] proposes a related approach. Like this work, TEX-Ture performs multiview denoising on a texture map representation. However, TEX-Ture runs an entire generative denoising process in each camera view in sequence, conditioning on the previous views and projecting to the texture map only after full denoising. In contrast, in TexFusion we propose to interleave texture aggregation with denoising steps in different camera views, simultaneously generating the entire output. This insight significantly reduces view inconsistencies and improves quality in TexFusion compared to TEX-Ture, as validated in Sec. 5. MultiDiffusion [3] concurrently introduces a method for panorama generation and other controlled image generation tasks, leveraging relatively lower-resolution image diffusion models. Algorithmically, this approach of aggregating different denoising predictions from different image crops is closely related to TexFusion's aggregation from different camera views. However, MultiDiffusion only tackles image synthesis, and is not concerned with any 3D or texture generation at all." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 632, + 126, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 126, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 126, + 646 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Diffusion Models Diffusion models [20, 74, 76] model a data distribution " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " via iterative denoising, and are trained with denoising score matching [20, 24, 46, 74, 76, 83]. Given samples " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\sim p_{\\mathrm{data}}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ", a denoiser model " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " parameterized with learnable parameters" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 283, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 283, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 283, + 545, + 308 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 283, + 545, + 308 + ], + "type": "text", + "content": " receives diffused inputs " + }, + { + "bbox": [ + 305, + 283, + 545, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t(\\epsilon, t, \\mathbf{x})" + }, + { + "bbox": [ + 305, + 283, + 545, + 308 + ], + "type": "text", + "content": " and is optimized by minimizing the denoising score matching objective" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 334, + 314, + 545, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 314, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 334, + 314, + 545, + 330 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\mathbf {x} \\sim p _ {\\text {d a t a}}, t \\sim p _ {t}, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}; \\mathbf {c}, t) \\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "93ba5c63ea82bbfaadc4b683f406532f032acf8fe03b62ba118e9ef2d507b22d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "text", + "content": " is optional conditioning information, such as a text prompt, and " + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "text", + "content": " is a uniform distribution over the diffusion time " + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "text", + "content": ". The model effectively learns to predict the noise " + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "text", + "content": " that was used to perturb the data (other formulations are possible [28, 68]). Letting " + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "text", + "content": " define a noise schedule, parameterized via a diffusion-time " + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "text", + "content": ", we construct " + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\sqrt{\\alpha_t}\\mathbf{x} + \\sqrt{1 - \\alpha_t}\\epsilon" + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 304, + 335, + 546, + 526 + ], + "type": "text", + "content": "; this particular formulation corresponds to a variance-preserving schedule [76]. The forward diffusion as well as the reverse generation process in diffusion models can be described in a continuous-time framework [76], but in practice a fixed discretization can be used [20]. The maximum diffusion time is generally chosen such that the input data is entirely perturbed into Gaussian random noise and an iterative generative denoising process can be initialized from such Gaussian noise to synthesize novel data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 527, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 545, + 586 + ], + "type": "text", + "content": "Classifier-free guidance [22] can be used for improved conditioning. By randomly dropping out the conditioning c during training, we can learn both a conditional and an unconditional model at the same time, and their predictions can be combined to achieve stronger conditioning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 586, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 545, + 612 + ], + "type": "text", + "content": "We perform iterative diffusion model sampling via the Denoising Diffusion Implicit Models (DDIM) scheme [75]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 336, + 628, + 545, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 628, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 336, + 628, + 545, + 682 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {x} _ {i - 1} = \\sqrt {\\alpha_ {i - 1}} \\left(\\frac {\\boldsymbol {x} _ {i} - \\sqrt {1 - \\alpha_ {i}} \\boldsymbol {\\epsilon} _ {\\theta} ^ {(t _ {i})} (\\boldsymbol {x} _ {i})}{\\sqrt {\\alpha_ {i}}}\\right) \\tag {2} \\\\ + \\sqrt {1 - \\alpha_ {i - 1} - \\sigma_ {t _ {i}} ^ {2}} \\cdot \\pmb {\\epsilon} _ {\\theta} ^ {(t _ {i})} (\\pmb {x} _ {i}) + \\sigma_ {t _ {i}} \\pmb {\\epsilon} _ {t _ {i}} \\\\ \\end{array}", + "image_path": "ebcd1f3bc5c96751c67f74093847c5a8b38ceccd4e3984a793db352aeb08b00f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{t_i}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{t_i}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " is a variance hyperparameter. We express obtaining " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{i - 1}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " via DDIM sampling as " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "x_{i - 1}\\sim" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 314, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 314, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 314, + 742 + ], + "type": "text", + "content": "4171" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 261, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 261, + 86 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 261, + 86 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\theta}^{(t_i)}(\\pmb{x}_{i-1}|\\pmb{x}_i)" + }, + { + "bbox": [ + 47, + 71, + 261, + 86 + ], + "type": "text", + "content": ". See Supp. Material for more details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "text", + "content": "Latent Diffusion Models (LDMs) and Stable Diffusion Instead of directly operating on pixels, LDMs [63] utilize an encoder " + }, + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "text", + "content": " for translation between images " + }, + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "text", + "content": " and latents " + }, + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in X" + }, + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "text", + "content": " of a lower spatial dimension. The diffusion process is then defined over the distribution of " + }, + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "text", + "content": ". Stable Diffusion is an LDM trained on the LAION-5B image-text dataset. In addition to text-conditioning, Stable Diffusion 2.0 permits depth conditioning with a depth map " + }, + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 101, + 289, + 222 + ], + "type": "text", + "content": " (SD2-depth). This allows detailed control of the shape and configuration of objects in its synthesized images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "text", + "content": "**Rendering and Geometry Representation** In principle, our method applies to any geometry representation for which a textured surface can be rendered; in practice, our experiments use a surface mesh " + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = (\\mathcal{V},\\mathcal{F})" + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "text", + "content": ", with vertices " + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{v_i\\}" + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "inline_equation", + "content": "v_{i}\\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "text", + "content": " and triangular faces " + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{f_i\\}" + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "text", + "content": " where each " + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "text", + "content": " is a triplet of vertices. Textures are defined in 2D image space in an injective UV parameterization of " + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "inline_equation", + "content": "UV:p\\in \\mathcal{M}\\mapsto (u,v)\\in [0,1]^2" + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "text", + "content": ". If needed, this parameterization can be automatically constructed via tools such as XATLAS [6, 96]. We encode textures as multi-channel images discretized at pixels in UV space " + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{z}\\in \\mathbb{R}^{(H\\times W,C)}" + }, + { + "bbox": [ + 47, + 238, + 287, + 380 + ], + "type": "text", + "content": ", notationally collapsing the spatial dimension for simplicity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "text", + "content": "We denote the rendering function as " + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\mathbf{z};\\mathcal{M},C):\\mathbf{z}\\mapsto" + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^{(h\\times w,C)}" + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "text", + "content": ", which takes as input a mesh " + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "text", + "content": ", camera " + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "text", + "content": ", and texture " + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "text", + "content": ", and produces as output a rendered image. The inverse of this function " + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{-1}(\\mathbf{x};\\mathcal{M},C):\\mathbf{x}\\mapsto \\mathbf{z}" + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "text", + "content": " projects values from camera image-space onto the UV texture map. Notationally, we often omit the dependence on " + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 382, + 288, + 501 + ], + "type": "text", + "content": " for brevity. In this work, we do not model any lighting or shading effects, images are formed by directly projecting textures into screen space (and then decoding, for latent textures)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 514, + 287, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 287, + 528 + ], + "type": "text", + "content": "4. Texture Sampling with 2D Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "Given a mesh geometry " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": ", a text prompt " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": ", and a geometry (depth) conditioned image diffusion model " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": ", how could one sample a complete surface texture? Assuming access to the rendering " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " and inverse rendering " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{-1}" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " functions defined above, perhaps the most naive approach is to compute a set of " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{C_1,..,C_N\\}" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " camera views that envelopes the surface, render the depth map " + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "inline_equation", + "content": "d_{n}" + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": " in each view, sample images from the image diffusion model with depth conditioning, and then back-project these images to the mesh surface (e.g. as done in [29]). However, image diffusion models in each view have no information about the generated results in other views, thus there is no coherency in the contents generated in each view. As an alternative, one may define a canonical order of camera views, and autoregressively condition image sampling in the subsequent views on" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": "previously sampled regions (as done in [48, 62]). However, for most geometries of interest (i.e. not a plane), a single camera can not observe the entirety of the geometry. Consequently, images synthesized early in the sequence could produce errors that are not reconcilable with the geometry that is observed later in the sequence (see Fig. 3). Thus, it is desirable for the image sampling distribution " + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "inline_equation", + "content": "p(\\cdot | d_i, y)" + }, + { + "bbox": [ + 304, + 72, + 547, + 179 + ], + "type": "text", + "content": " in each camera view to be conditioned on that of every other camera view." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 186, + 525, + 264 + ], + "blocks": [ + { + "bbox": [ + 328, + 186, + 525, + 264 + ], + "lines": [ + { + "bbox": [ + 328, + 186, + 525, + 264 + ], + "spans": [ + { + "bbox": [ + 328, + 186, + 525, + 264 + ], + "type": "image", + "image_path": "7c2b3b8da3a5357c3094c2ff57a1b9983a0efdb4c0ea5354ba408ab5edd0c9a2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 268, + 547, + 335 + ], + "lines": [ + { + "bbox": [ + 304, + 268, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 547, + 335 + ], + "type": "text", + "content": "Figure 3: Illustration of irreconcilable mistakes in early views impacting denoised results in later views (images sampled from TEXTure). While the highlighted area appears natural in the first view, it does not match the geometry when viewed from a different angle, thereby creating poorly denoised results when the second view is inpainted." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 345, + 522, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 345, + 522, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 345, + 522, + 358 + ], + "type": "text", + "content": "4.1. Sequential Interlaced Multiview Sampler" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "text", + "content": "Leveraging the sequential nature of the denoising process, we can interlace the synchronization of content across views with the denoising steps within each view to achieve coherency over the entire shape. Suppose that at step " + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "text", + "content": " of the denoising process, we have a set of partially denoised images " + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{i,n}\\}_{n = 1}^{N} = \\{\\mathbf{x}_{i,1}\\dots \\mathbf{x}_{i,N}\\}" + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "text", + "content": ". Our goal is to sample " + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{i - 1,n}\\}_{n = 1}^{N} = \\mathbf{x}_{i - 1,1}\\dots \\mathbf{x}_{i - 1,N}" + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "text", + "content": " that is 3D-consistent, i.e., two pixels in " + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i - 1,a},\\mathbf{x}_{i - 1,b}" + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "text", + "content": " that project to the same point in 3D should have the same value. Taking inspiration from the autoregressive modeling literature, we sample the joint distribution via first decomposing it into a product of conditionals, and approximating each term in the product by using the diffusion model to denoise the renders of a dynamically updated latent texture map. Specifically, we first initialize an initial latent texture map " + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "text", + "content": ", (" + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "text", + "content": " demarks the first step of the diffusion trajectory). Then, suppose that we have a 3D consistent latent texture map " + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 304, + 363, + 547, + 590 + ], + "type": "text", + "content": ", we decompose the joint distribution as follows (conditioning on depth and prompt omitted for space):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 347, + 597, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 597, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 347, + 597, + 545, + 647 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p _ {\\theta} \\left(\\left\\{\\mathbf {x} _ {i - 1, j} \\right\\} _ {j = 1} ^ {N} \\mid \\mathbf {z} _ {i}\\right) = p _ {\\theta} \\left(\\mathbf {x} _ {i - 1, 1} \\mid \\mathbf {z} _ {i}\\right) \\times \\\\ \\prod_ {n = 2} ^ {N} p _ {\\theta} \\left(\\mathbf {x} _ {i - 1, n} \\mid \\left\\{\\mathbf {x} _ {i - 1, j} \\right\\} _ {j = 1} ^ {n - 1}, \\mathbf {z} _ {i}\\right) \\tag {3} \\\\ \\end{array}", + "image_path": "26f39e0a7e6beb7eb38d780272c4978594dd014b0f70ede62113d8cd26c83b82.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 655, + 547, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 655, + 547, + 691 + ], + "spans": [ + { + "bbox": [ + 305, + 655, + 547, + 691 + ], + "type": "text", + "content": "We can compute the first term by first rendering " + }, + { + "bbox": [ + 305, + 655, + 547, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 305, + 655, + 547, + 691 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 305, + 655, + 547, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,1}^{\\prime} = \\mathcal{R}(\\mathbf{z}_i;C_1)" + }, + { + "bbox": [ + 305, + 655, + 547, + 691 + ], + "type": "text", + "content": ". Eqn. 2 can now be applied to " + }, + { + "bbox": [ + 305, + 655, + 547, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,1}^{\\prime}" + }, + { + "bbox": [ + 305, + 655, + 547, + 691 + ], + "type": "text", + "content": " to draw a latent image at the next time step:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 340, + 699, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 699, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 340, + 699, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {i - 1, 1} \\sim f _ {\\theta} ^ {(t _ {i})} \\left(\\mathbf {x} _ {i - 1, 1} \\mid \\mathbf {x} _ {i, 1} ^ {\\prime} = \\mathcal {R} \\left(\\mathbf {z} _ {i}; C _ {1}\\right)\\right). \\tag {4}", + "image_path": "159689232b4e4bfe67343b66b5e64892a09ccd2133c65f15648b5c3b57087d71.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4172" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "Later terms in Eq. 3 additionally depend on the result of previous denoising steps. We again model this dependency through the latent texture map. For each view starting at " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "n = 1" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": ", we inverse render " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i - 1,n}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " into texture space to obtain " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,n}^{\\prime} = \\mathcal{R}^{-1}(\\mathbf{x}_{i - 1,n})" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": ", and update the pixels of " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,n - 1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " that are newly visible in " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,n}^{\\prime}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,n}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " (See Sec. 4.2.2 for details). Then, in the " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " iteration, since " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,n}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " contain regions at two noise scales (unseen regions are at noise scale " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": ", while visited regions are at noise scale " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\sigma_{i - 1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "), we add appropriate 3D consistent noise to the visited regions of " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,n}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " to match the noise scale of step " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " before rendering it as input to " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": ". Letting " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "M_{i,n}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " represent the mask for visited regions and " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\epsilon_{i}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": ", we can write the sampling procedure for " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i - 1,n}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 83, + 251, + 287, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 251, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 83, + 251, + 287, + 293 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {z} _ {i, n} = M _ {i, n} \\odot \\left(\\sqrt {\\frac {\\alpha_ {i - 1}}{\\alpha_ {i}}} \\mathbf {z} _ {i - 1, n - 1} + \\sigma_ {i} \\epsilon_ {i}\\right) \\\\ + \\left(\\mathbf {1} - M _ {i, n}\\right) \\odot \\mathbf {z} _ {i} \\tag {5} \\\\ \\end{array}", + "image_path": "651b5ae497b8362e89f03b9aaf0c46df6f7aa77391eb78f5c3031536df782d3a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 296, + 252, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 296, + 252, + 312 + ], + "spans": [ + { + "bbox": [ + 74, + 296, + 252, + 312 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {i - 1, n} \\sim f _ {\\theta} ^ {(t _ {i})} (\\mathbf {x} _ {i - 1, n} | \\mathbf {x} _ {i, n} ^ {\\prime} = \\mathcal {R} (\\mathbf {z} _ {i, n}; C _ {n}))", + "image_path": "7645277da6f712b837ddb9d805fa211b893133caa972f15b32a976a76f2d84e4.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "text", + "content": "By iteratively applying Eqn. 5, we obtain a sequence of 3D consistent images " + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{i - 1,n}\\}_{n = 1}^{N}" + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "text", + "content": " and a texture map " + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,n}" + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "text", + "content": " that has been aggregated from these images. We can then decrement " + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 320, + 287, + 369 + ], + "type": "text", + "content": " and repeat this process in the next time step." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 477 + ], + "type": "text", + "content": "We name this approach Sequential Interlaced Multiview Sampler (SIMS). SIMS communicates the denoising direction of previous views to the current view and resolves overlaps between views by the order of aggregation. It ameliorates inconsistent predictions while circumventing performance degradation due to the averaging of latent predictions during parallel aggregation. In the single-view case, SIMS is equivalent to standard DDIM sampling. A complete algorithm for SIMS can be found in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 487, + 191, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 191, + 499 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 191, + 499 + ], + "type": "text", + "content": "4.2. The TexFusion Algorithm" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 506, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 578 + ], + "type": "text", + "content": "In Sec. 4.1, we have presented a generic algorithm for sampling 3D consistent multi-view images and texture maps using 2D diffusion models. We now present a concrete algorithm, TexFusion, that uses SIMS to texture 3D meshes using SD2-depth [63] as the diffusion model. An illustrative overview of TexFusion can be found in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 578, + 288, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 721 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 721 + ], + "type": "text", + "content": "As SD2-depth is a latent diffusion model, we apply SIMS in the latent space: " + }, + { + "bbox": [ + 46, + 578, + 288, + 721 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 578, + 288, + 721 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 578, + 288, + 721 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 578, + 288, + 721 + ], + "type": "text", + "content": " represent latent images and latent texture maps respectively, and Section 4.2.3 will describe a final distillation post-process to color space. In this section, we address several challenges specific to using LDMs with SIMS, and detail our design choices to tackle these challenges. We find that a canonical set of cameras works well for most objects, but cameras are tailored to specific objects to improve resolution and ameliorate occlusion. We further illustrate a technique for obtaining high-quality results by operating SIMS in a cascade of increasing resolutions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 505, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 505, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 505, + 84 + ], + "type": "text", + "content": "4.2.1 (Inverse) Rendering of Latent Textures" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "text", + "content": "We use NVDIFFRAST [35] to efficiently render textured meshes via rasterization, as described in Sec. 3. Our implementation sets the rendered image size " + }, + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "inline_equation", + "content": "h = w = 64" + }, + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "text", + "content": " to match Stable Diffusion's UNet, with latent vectors of dimension " + }, + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "inline_equation", + "content": "D = 4" + }, + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "text", + "content": ". The texture image dimensions " + }, + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 91, + 545, + 175 + ], + "type": "text", + "content": " are chosen based on the surface area of the mesh relative to its diameter (detailed in Sec. 4.2.4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": "For each non-background pixel " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": " in a rendered image " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,..}^{\\prime}" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": ", rasterization gives a corresponding location on the texture image via the UV map " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "(u,v) = UV(p)" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": ", and we retrieve the value at the nearest texture pixel. This texture value is latent, and no shading or lighting is applied. In other settings, texture data is often rendered with bilinear filtering and mipmaps to improve image quality and reduce aliasing, but in this setting, we found it essential to avoid such techniques. We experimentally ablate texturing approaches in the supplementary. The issue is that early in the diffusion process, i.e. when " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "t_i\\ll 0" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,..}^{\\prime}" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": " are dominated by noise, but interpolation and mipmapping change the variance in the pixels " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,..}" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": ", thereby moving " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,..}^{\\prime}" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": " out of the training distribution of " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{(t_i)}" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": ". Instead, we retrieve only the nearest texture pixels for diffusion, and resolve aliasing and interpolation via a simple distillation post-process (Sec. 4.2.3). For each background pixel in rendered image " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,..}^{\\prime}" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": ", we apply a Gaussian noise of standard deviation " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": " such that the background matches the marginal distribution at diffusion step " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": ". This ensures that the diffusion model " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{(t_i)}" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": " focuses solely on the foreground pixels of " + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,..}^{\\prime}" + }, + { + "bbox": [ + 304, + 176, + 546, + 429 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 430, + 547, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 501 + ], + "type": "text", + "content": "Note that in this setting a rendered image is simply a selection of pixels from the texture maps, and thus we can leverage backpropagation to easily implement the inverse rendering function " + }, + { + "bbox": [ + 304, + 430, + 547, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{-1}" + }, + { + "bbox": [ + 304, + 430, + 547, + 501 + ], + "type": "text", + "content": ". Additionally, forward rasterization yields a depth map of the scene, which we use as a conditioning input to the diffusion model." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 328, + 512, + 523, + 559 + ], + "blocks": [ + { + "bbox": [ + 328, + 512, + 523, + 559 + ], + "lines": [ + { + "bbox": [ + 328, + 512, + 523, + 559 + ], + "spans": [ + { + "bbox": [ + 328, + 512, + 523, + 559 + ], + "type": "image", + "image_path": "d5c6ef76b5bab196f4581d3541c42694133e12b8572ae7ea659f5f4f80c147e1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 571, + 547, + 626 + ], + "lines": [ + { + "bbox": [ + 304, + 571, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 547, + 626 + ], + "type": "text", + "content": "Figure 4: Left: Depth map for conditioning SD2-depth, and quality image computed from screen space derivatives. Right: Output of SD2 decoder in two views using renders of the latent texture map as input. Note how the horizontal line across the doors changes appearance from one view to another." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 646, + 464, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 646, + 464, + 658 + ], + "spans": [ + { + "bbox": [ + 305, + 646, + 464, + 658 + ], + "type": "text", + "content": "4.2.2 Aggregating Latent Textures." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 546, + 713 + ], + "type": "text", + "content": "Recall from Sec. 4.1, before iterating through cameras " + }, + { + "bbox": [ + 305, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\{C_1,\\dots,C_N\\}" + }, + { + "bbox": [ + 305, + 665, + 546, + 713 + ], + "type": "text", + "content": ", we first initialize a latent texture map " + }, + { + "bbox": [ + 305, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,0} = \\mathbf{z}_i" + }, + { + "bbox": [ + 305, + 665, + 546, + 713 + ], + "type": "text", + "content": " and render latent image " + }, + { + "bbox": [ + 305, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,1}^{\\prime} = \\mathcal{R}(\\mathbf{z}_{i - 1,0})" + }, + { + "bbox": [ + 305, + 665, + 546, + 713 + ], + "type": "text", + "content": ". Then, in each step (iterating through cameras), we ob" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4173" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": "tain " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i - 1,n}" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{(t_i)}(\\mathbf{x}_{i,n}^{\\prime})" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " , and inverse render it to get " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,n}^{\\prime} = \\mathcal{R}^{-1}(\\mathbf{x}_{i - 1,n})" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " . Finally, we need to aggregate the partial texture map " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,n}^{\\prime}" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,n - 1}" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " to obtain a partially updated texture map " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i - 1,n}" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " . We perform this aggregation step based on a simple heuristic that the value of each pixel " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " should be determined by the camera that has the most \"direct and up-close\" view to its corresponding point on the mesh. We measure view quality using image-space derivatives - the amount of change in UV coordinates per infinitesimal change in the image coordinate. This quantity is commonly used for determining mipmapping resolution and anti-aliasing, and it can be efficiently computed by nvdiffrast when rasterizing " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,n - 1}" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " . For each pixel location " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "(p,q)" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i - 1,n}" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " , we compute the negative Jacobian magnitude as " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "-|\\frac{\\partial u}{\\partial p}\\cdot \\frac{\\partial v}{\\partial q} -\\frac{\\partial u}{\\partial q}\\cdot \\frac{\\partial v}{\\partial p} |," + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " and inverse render it to the texture space, which we denote as " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "Q_{i,n}" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " . Higher values of " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "Q_{i,n}(u,v)" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " means that camera " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " has a better view of " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " (see Fig.4). In addition, we compute " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "M_{i,n} = \\mathcal{R}^{-1}(\\boldsymbol {I})" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " , such that " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "M_{i,n}(u,v)" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " represents the number of pixels in " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,n}^{\\prime}" + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "text", + "content": " that received value from " + }, + { + "bbox": [ + 47, + 71, + 289, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i(u,v)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "text", + "content": "While iterating through cameras 1 through " + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "text", + "content": ", we maintain mask " + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "text", + "content": " (initialized to 0), which is used to track which pixels of " + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,n}" + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "text", + "content": " have been seen by cameras up to " + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "text", + "content": ", and quality buffer " + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "text", + "content": " (initialized to " + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "inline_equation", + "content": "-\\inf" + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "text", + "content": "), which is used to track the highest non-zero value of " + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "inline_equation", + "content": "Q_{i,n}" + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "text", + "content": " at each " + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "text", + "content": ". The value at " + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i-1,n}" + }, + { + "bbox": [ + 47, + 314, + 288, + 387 + ], + "type": "text", + "content": " is determined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 396, + 286, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 396, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 286, + 449 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i - 1, n} (u, v) = \\left\\{ \\begin{array}{l l} \\frac {\\mathbf {z} _ {i - 1 , n} ^ {\\prime} (u , v)}{M _ {i , n} (u , v)} & M _ {i, n} (u, v) > 0, \\text {a n d} \\\\ \\mathbf {z} _ {i - 1, n - 1} (u, v) & Q _ {i, n} (u, v) > Q _ {i} (u, v) \\\\ \\end{array} \\right. \\tag {6}", + "image_path": "00588be84f6e509637922a4b04a5f88601265c50e3747f4ad53935b9d11b1377.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 449, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 287, + 509 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 47, + 449, + 287, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 449, + 287, + 509 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 47, + 449, + 287, + 509 + ], + "type": "text", + "content": " are then updated pixelwise as " + }, + { + "bbox": [ + 47, + 449, + 287, + 509 + ], + "type": "inline_equation", + "content": "M_{i} = \\min (M_{i,n} + M_{i},1)" + }, + { + "bbox": [ + 47, + 449, + 287, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 449, + 287, + 509 + ], + "type": "inline_equation", + "content": "Q_{i} = \\max (Q_{i,n},Q_{i})" + }, + { + "bbox": [ + 47, + 449, + 287, + 509 + ], + "type": "text", + "content": " (min and max are applied element-wise). We use Eqn. 6 in conjunction with Eqn. 5 in SIMS to perform denoising of images in the sequence of camera views." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "We further note that a side effect of using nearest pixel filtering during texture sampling in SIMS is aliasing. When the image space derivative Jacobian is much higher than 1 (low quality), neighboring pixels in screen space will have gaps when mapped to uv space. This results in incomplete filling of the latent texture map. Conversely, when the image space derivative Jacobian magnitude is much smaller than 1 (high quality), multiple screen space pixels " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "(p,q)" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " will map to the same " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": ", creating uniform blocks in " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": ". Similar to interpolated latents, these blocks are particularly detrimental early during the diffusion process, as they cannot be correctly denoised by " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " which is expecting high spatial variance. Our quality-based aggregation overwrites low-quality regions in the latent texture when higher quality views are available. Thus we can set the resolution of the texture map such that the image space derivatives in each view do not fall below 1 (i.e. " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\max Q_{i,n} < -1" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": ") to prevent" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "aliasing of the second type, and rely on cameras with better views to fix aliasing of the first type." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 101, + 468, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 101, + 468, + 113 + ], + "spans": [ + { + "bbox": [ + 305, + 101, + 468, + 113 + ], + "type": "text", + "content": "4.2.3 From Latent to RGB Textures" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": "So far, we have described how to use SIMS to produce a set of 3D consistent latent images and latent texture maps, but have yet to describe how to translate this into a RGB texture map that can be rendered for viewing. Towards this, we experimented with multiple approaches and found performing multi-view distillation of decoded latent images " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{0,n}\\}_{n=1}^{N}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " with a neural color field to be most performant. Specifically, we use the decoder " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " of Stable Diffusion to decode latent multi-view images " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{0,n}\\}_{n=1}^{N}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " into RGB multi-view images " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\{\\xi_{n} = \\mathcal{D}(\\mathbf{x}_{0,n})\\}_{n=1}^{N}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": ". Notably, decoding with " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " introduce inconsistencies such that " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\{\\xi_{n}\\}_{n=1}^{N}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " is not 3D consistent even when " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{0,n}\\}_{n=1}^{N}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " is 3D consistent (see Fig. 4 for example). In stable Diffusion, each latent vector (pixel of " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": ") needs to carry the information of a " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " patch of RGB pixels. Thus, the value of the latent vector encodes both color values and spatial patterns. We therefore cannot expect their decoded results to be equivariant to perspective transformations. To address this problem, we leverage a neural color field optimized with appearance-based losses to smooth out inconsistencies. Since we know the ground-truth camera poses, we can directly obtain the 3D spatial coordinates " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\{xyz_{n}\\}_{n=1}^{N}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " of all pixels of " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\{\\xi_{n}\\}_{n=1}^{N}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " by projecting pixels to the mesh surface. Background pixels that do not intersect any surface are discarded. Following [40], we use a multi-resolution hash encoding based on instantNGP [51] along with a shallow MLP " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "f_{\\phi}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " to parameterize a function from 3D spatial coordinates to RGB values for each sample " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "rgb = f_{\\phi}(hash(xyz))" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": ". We then distill multi-view images " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\{\\xi_{n}\\}_{n=1}^{N}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " into this parametric function via optimization. Since our goal is to export " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " into a texture map, we do not use any view-dependent parameterization. To reconcile inconsistencies, we use both a standard " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " loss and a VGG-based perceptual loss, applied between " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "f_{\\phi}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\{\\xi_{n}\\}_{n=1}^{N}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": ", to train " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": ". We use Adam with a learning rate of 0.01, and optimization of " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " converges within 100 iterations. After optimization, we compute the spatial coordinate of the centers of each pixel in a high-resolution texture map, and query " + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "inline_equation", + "content": "f_{\\phi}" + }, + { + "bbox": [ + 304, + 120, + 547, + 577 + ], + "type": "text", + "content": " to predict RGB values for the texture map." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 586, + 545, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 586, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 305, + 586, + 545, + 610 + ], + "type": "text", + "content": "4.2.4 Geometry Processing, Camera, and Multi-resolution Refinement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "We normalize " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": " such that it fits inside a cube of side length 1, and center it at the origin. Perspective cameras are placed facing the origin, and their FOV is adjusted to fit the object. Detailed parameters can be found in the appendix. As the diffusion model relies on context captured in " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": ", to perform denoising, camera views that are too small w.r.t the size of the object often result in content drift - the texture in distant areas of the mesh can have content that is seman" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4174" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "tically inconsistent. This can be seen as our version of the Janus face problem known to Dreamfusion and similar approaches [40, 59, 84]. Our solution is to perform two rounds of TexFusion at different resolution scales to obtain high-resolution textures while retaining semantic consistency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 289, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 289, + 371 + ], + "type": "text", + "content": "Specifically, we first run TexFusion using cameras that cover the entire object, and a low-resolution latent texture map that is suitable for these cameras. We do not run view distillation in this first round as we are only interested in the latent texture map. We denote the denoised latent texture map from this step as " + }, + { + "bbox": [ + 46, + 133, + 289, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{0,lr}" + }, + { + "bbox": [ + 46, + 133, + 289, + 371 + ], + "type": "text", + "content": ". We then use a second set of cameras with a narrower field of view; these cameras are also more numerous to still cover the full object. We determine a new texture map resolution using the square of the ratio of the tangent of the old FOV over the new FOV - corresponding to the relative change in surface area covered by the camera before and after the change in FOV. We then up-sample " + }, + { + "bbox": [ + 46, + 133, + 289, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{0,lr}" + }, + { + "bbox": [ + 46, + 133, + 289, + 371 + ], + "type": "text", + "content": " with nearest neighbor filtering to this higher resolution, and stochastically encode it to a partially noised state (e.g. " + }, + { + "bbox": [ + 46, + 133, + 289, + 371 + ], + "type": "inline_equation", + "content": "T = 500" + }, + { + "bbox": [ + 46, + 133, + 289, + 371 + ], + "type": "text", + "content": " in the diffusion model time schedule). The second round of TexFusion uses these new cameras, and initializes SIMS with the partially noised latent texture map. The multi-view images produced by the second round of SIMS is used to produce the final output texture map via neural color field distillation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 382, + 128, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 128, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 128, + 396 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 403, + 288, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 288, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 288, + 571 + ], + "type": "text", + "content": "We apply TexFusion on various geometries and text prompts to evaluate its ability to produce high quality, natural, and 3D consistent textures. We focus our experimental comparisons on TEXTure [62], a text-driven texture generation method that also uses SD2-depth. We choose TEXTure as the baseline because (1) it represents the current state-of-the-art for language-conditioned texture generation, (2) it uses the same SD2-depth model, which can be prompted for a wide variety of content, and (3) it is concurrent to our work. In the supplementary materials, we further compare TexFusion to SDS-based text-driven texture distillation [48, 59], leveraging the Stable Diffusion model and show that TexFusion achieves superior results in terms of quality and speed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 628 + ], + "type": "text", + "content": "Dataset We collect 35 meshes of various content, and write 2-4 prompts for each mesh. In total, we evaluate each texture synthesis method on 86 mesh-text pairs. More details of this dataset are in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 635, + 187, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 187, + 648 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 187, + 648 + ], + "type": "text", + "content": "5.1. Qualitative Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "We visualize textures produced by TexFusion on multiple geometries and prompts, and compare to state-of-the-art baselines in Figure 6. We render each object in 3 surrounding views to allow better examination of 3D consistency of the produced texture. We use Blender's Cycles renderer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "with a studio light-setup. Textures are applied as base color to a diffuse material. Additional visualizations, including videos showing 360 pans of all objects presented in the paper, and renders of our results using only texture color, can be found in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 133, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 312 + ], + "type": "text", + "content": "In terms of local 3D consistency (consistency in neighborhoods on the surface of the mesh), textures produced by TexFusion are locally consistent - there are no visible seam lines or stitching artifacts. In contrast, we often find severe artifacts when viewing the top and back sides of TEXTure's outputs. These artifacts are most noticeable when a clean color is expected, such as when texturing vehicles. In terms of global consistency (semantic coherency of the entire texture, e.g. exactly 2 eyes and 1 nose to a face), TEXTure performs poorly and suffers from problems similar to Dream-Fusion's Janus face problem [59]: as the 2D diffusion model captures context of the object through its own camera view, it is not aware of the appearance of the opposite side of the object. This problem is ameliorated in TexFusion due to frequent communication between views in SIMS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 313, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 546, + 445 + ], + "type": "text", + "content": "There are noticeably more baked specular highlights and shadows in textures generated by TEXTure. These effects are physically inaccurate as they conflict with the lighting effects simulated by the renderer. In contrast, TexFusion produces textures that are smoothly illuminated. We hypothesize that this is due to interlacing aggregations in SIMS, which removes view-dependent effects as they arise in the sampling process. We provide additional visualizations of TexFusion in Fig. 5 to showcase how text prompting is effective at controlling the generation of textures, thereby producing varying appearances using the same mesh." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 464, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 513 + ], + "type": "text", + "content": "Runtime TexFusion takes approximately 3 minutes on a machine with a single GPU to sample one texture. We are slightly faster than TEXTure (5 min.) [62], since we only need to optimize a color field once after sampling is" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 529, + 527, + 703 + ], + "blocks": [ + { + "bbox": [ + 321, + 529, + 527, + 703 + ], + "lines": [ + { + "bbox": [ + 321, + 529, + 527, + 703 + ], + "spans": [ + { + "bbox": [ + 321, + 529, + 527, + 703 + ], + "type": "image", + "image_path": "c2c4ddf721294c7501817252ac0d1318e1bbd0ee17be264ff308a64f8ffa5ed5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 707, + 536, + 719 + ], + "lines": [ + { + "bbox": [ + 311, + 707, + 536, + 719 + ], + "spans": [ + { + "bbox": [ + 311, + 707, + 536, + 719 + ], + "type": "text", + "content": "Figure 5: More TexFusion text-conditioned texturing results." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4175" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 74, + 545, + 228 + ], + "blocks": [ + { + "bbox": [ + 56, + 74, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 56, + 74, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 74, + 545, + 228 + ], + "type": "image", + "image_path": "fae4b983bc70a652edb2ecabf272f2b8cd3230dec047727c9935b6d93110a90d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 104, + 232, + 499, + 342 + ], + "blocks": [ + { + "bbox": [ + 104, + 232, + 499, + 342 + ], + "lines": [ + { + "bbox": [ + 104, + 232, + 499, + 342 + ], + "spans": [ + { + "bbox": [ + 104, + 232, + 499, + 342 + ], + "type": "image", + "image_path": "29f8503ae5077a06264fb3f0e19107c2cf68b29efd8d9cd279da2fa35f71d8d2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 344, + 449, + 356 + ], + "lines": [ + { + "bbox": [ + 140, + 344, + 449, + 356 + ], + "spans": [ + { + "bbox": [ + 140, + 344, + 449, + 356 + ], + "type": "text", + "content": "Figure 6: Visual comparison of textures generated by TEXTure [62] and TexFusion." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 357, + 288, + 413 + ], + "blocks": [ + { + "bbox": [ + 50, + 357, + 288, + 413 + ], + "lines": [ + { + "bbox": [ + 50, + 357, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 50, + 357, + 288, + 413 + ], + "type": "table", + "html": "
MethodFID (↓)User study (%)
Natural Color (↑)More Detailed (↑)Less Artifacts (↑)Align with Prompt (↑)
TEXTure79.4724.4265.1231.4043.02
TexFusion59.7875.5834.8868.6056.98
", + "image_path": "e3b62d547b19e4cfc370a233e49ca2123929e8b08426dc675c15ed019a8b5f86.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 422, + 287, + 456 + ], + "lines": [ + { + "bbox": [ + 47, + 422, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 287, + 456 + ], + "type": "text", + "content": "Table 1: Quantitative and qualitative comparisons between TEX-Ture and TexFusion. FID is computed w.r.t. a set of images synthesized by SD2-depth." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 468, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 287, + 517 + ], + "type": "text", + "content": "complete. We are also an order of magnitude faster than methods that rely on SDS loss which report upwards of 40 minutes [40, 48, 59]. Additional runtime comparisons are in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 524, + 194, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 524, + 194, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 524, + 194, + 537 + ], + "type": "text", + "content": "5.2. Quantitative Comparisons" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 542, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 603 + ], + "type": "text", + "content": "It is inherently difficult to quantitatively evaluate the quality of text and geometry-conditioned texture generation as one must take alignment with the conditioning inputs into account. We rely on both automated and human evaluation to gauge the quality of the synthesized textures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "FID Since SD2-depth can produce high-quality images that match the structure of the conditioning geometry, we sample it to create a proxy ground truth set. Furthermore, as both TEXTure and TexFusion use SD2-depth as 2D prior, samples from it serve as an upper bound to the quality and diversity of the appearance of textures: We render the depth map of all meshes in 8 canonical views, and condition SD2-depth on both these depth maps and our text" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 359, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 359, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 547, + 515 + ], + "type": "text", + "content": "prompts to generate a ground truth set. We also render textures generated by both methods in these same 8 views, and compute the Fréchet Inception Distance (FID) between images rendered from the textures and the ground truth set. FID measures the overlap between two sets of images based on their similarity in a learned feature space. We white out the background regions in rendered images and ground truth images to focus the comparisons to the textured object. We present the results in Tab. 1. FID scores indicate that textures synthesized by TexFusion are closer in appearance to the ground truth set than those from TEXTure. The absolute FID values are likely high due to the relatively small set of images in each dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 522, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 714 + ], + "type": "text", + "content": "User study We conducted a user study to measure the overall quality of the generated textures. We use all mesh and text prompts in our dataset, and render each textured mesh into a video showing the results from a " + }, + { + "bbox": [ + 304, + 522, + 546, + 714 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 522, + 546, + 714 + ], + "type": "text", + "content": " rotating view. We present videos from TEXTure and TexFusion side by side in random left-right order, to avoid bias. The user study was conducted on Amazon Mechanical Turk. Each participant is asked to select the preferred result under four criteria: natural color (non-saturation), details, cleanliness, and alignment with the text prompt. To avoid randomness in the answers, we let 3 participants answer the same question and determine the choice by max-voting, a question screenshot is provided in the supplement. We provide the user study results in Tab. 1. Although not deemed as detailed as TEXTure, our results are overall preferred by the users in terms of better aligning with the provided text prompt, and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4176" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 108, + 126 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 108, + 126 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 108, + 126 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 108, + 126 + ], + "type": "image", + "image_path": "8fabd9c349e734537082503eedec140df2c7aba5d26cf9c1ac6e33a0da769ab6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 133, + 104, + 140 + ], + "lines": [ + { + "bbox": [ + 62, + 133, + 104, + 140 + ], + "spans": [ + { + "bbox": [ + 62, + 133, + 104, + 140 + ], + "type": "text", + "content": "\"red backpack\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 116, + 76, + 159, + 124 + ], + "blocks": [ + { + "bbox": [ + 116, + 76, + 159, + 124 + ], + "lines": [ + { + "bbox": [ + 116, + 76, + 159, + 124 + ], + "spans": [ + { + "bbox": [ + 116, + 76, + 159, + 124 + ], + "type": "image", + "image_path": "ab5e0688b22c38345ddbb3a105cc3367b15da5e8e4c1056f4c9345b053452440.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 126, + 224, + 147 + ], + "lines": [ + { + "bbox": [ + 167, + 126, + 224, + 147 + ], + "spans": [ + { + "bbox": [ + 167, + 126, + 224, + 147 + ], + "type": "text", + "content": "\"Portrait of a humanoid robot, futuristic, science fiction\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 71, + 220, + 123 + ], + "blocks": [ + { + "bbox": [ + 176, + 71, + 220, + 123 + ], + "lines": [ + { + "bbox": [ + 176, + 71, + 220, + 123 + ], + "spans": [ + { + "bbox": [ + 176, + 71, + 220, + 123 + ], + "type": "image", + "image_path": "0221b34bce87f77084fce0a3f15edb6f11a4b3273ec4371173a5f464fead88ca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 126, + 278, + 148 + ], + "lines": [ + { + "bbox": [ + 230, + 126, + 278, + 148 + ], + "spans": [ + { + "bbox": [ + 230, + 126, + 278, + 148 + ], + "type": "text", + "content": "\"Medieval celtic Iouse, stone bricks, wooden roof\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 223, + 83, + 276, + 121 + ], + "blocks": [ + { + "bbox": [ + 223, + 83, + 276, + 121 + ], + "lines": [ + { + "bbox": [ + 223, + 83, + 276, + 121 + ], + "spans": [ + { + "bbox": [ + 223, + 83, + 276, + 121 + ], + "type": "image", + "image_path": "780bba85ba986bbae9197888854e1c8eeb585f8134b439a4048be95443f9a1ea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 65, + 148, + 108, + 202 + ], + "blocks": [ + { + "bbox": [ + 65, + 148, + 108, + 202 + ], + "lines": [ + { + "bbox": [ + 65, + 148, + 108, + 202 + ], + "spans": [ + { + "bbox": [ + 65, + 148, + 108, + 202 + ], + "type": "image", + "image_path": "21f3936efeae468b7e7e07f9cc89d4d70ee5347dd9be49c0c0d51be4d453f95b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 209, + 100, + 224 + ], + "lines": [ + { + "bbox": [ + 65, + 209, + 100, + 224 + ], + "spans": [ + { + "bbox": [ + 65, + 209, + 100, + 224 + ], + "type": "text", + "content": "\"camper bag, camouflage\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 115, + 156, + 164, + 198 + ], + "blocks": [ + { + "bbox": [ + 118, + 133, + 160, + 140 + ], + "lines": [ + { + "bbox": [ + 118, + 133, + 160, + 140 + ], + "spans": [ + { + "bbox": [ + 118, + 133, + 160, + 140 + ], + "type": "text", + "content": "\"Brown rabbit\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 115, + 156, + 164, + 198 + ], + "lines": [ + { + "bbox": [ + 115, + 156, + 164, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 164, + 198 + ], + "type": "image", + "image_path": "d99f9e40c816035f068e488dd2283d0b354039b1e9ae53e62f151cf6133393f9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 178, + 150, + 219, + 201 + ], + "blocks": [ + { + "bbox": [ + 118, + 208, + 161, + 224 + ], + "lines": [ + { + "bbox": [ + 118, + 208, + 161, + 224 + ], + "spans": [ + { + "bbox": [ + 118, + 208, + 161, + 224 + ], + "type": "text", + "content": "\"Crocodile skin handbag\"" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 150, + 219, + 201 + ], + "lines": [ + { + "bbox": [ + 178, + 150, + 219, + 201 + ], + "spans": [ + { + "bbox": [ + 178, + 150, + 219, + 201 + ], + "type": "image", + "image_path": "f1d973a6d898dbf7da172e05271a43ef4ba51b399f52a44b5de39bffa6e3716a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 206, + 222, + 226 + ], + "lines": [ + { + "bbox": [ + 168, + 206, + 222, + 226 + ], + "spans": [ + { + "bbox": [ + 168, + 206, + 222, + 226 + ], + "type": "text", + "content": "\"Blonde girl with green eyes, hair in tied a bun, DSLR portrait photo.\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 227, + 155, + 280, + 204 + ], + "blocks": [ + { + "bbox": [ + 227, + 155, + 280, + 204 + ], + "lines": [ + { + "bbox": [ + 227, + 155, + 280, + 204 + ], + "spans": [ + { + "bbox": [ + 227, + 155, + 280, + 204 + ], + "type": "image", + "image_path": "9d49a98b80f4182b08cfee95b73e69b89ed8554776369f675d90b41f6c204a26.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 209, + 274, + 223 + ], + "lines": [ + { + "bbox": [ + 230, + 209, + 274, + 223 + ], + "spans": [ + { + "bbox": [ + 230, + 209, + 274, + 223 + ], + "type": "text", + "content": "\"Leather lounge chair\"" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 236, + 287, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 270 + ], + "type": "text", + "content": "Figure 7: TexFusion + non-stochastic DDIM sampling " + }, + { + "bbox": [ + 46, + 236, + 287, + 270 + ], + "type": "inline_equation", + "content": "(\\sigma_{t_i} = 0)" + }, + { + "bbox": [ + 46, + 236, + 287, + 270 + ], + "type": "text", + "content": ". This setting emphasizes textural details, such as the leather (bottom, left), roof shingles (top, right)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 293, + 257, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 257, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 257, + 304 + ], + "type": "text", + "content": "having more natural color, and fewer artifacts/flaws." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 314, + 194, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 194, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 194, + 327 + ], + "type": "text", + "content": "5.3. Improving Texture Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 333, + 289, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 289, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 289, + 644 + ], + "type": "text", + "content": "Texfusion can produce more texture details with adjusted hyperparameters and diffusion backbones. First, we find that using the non-stochastic version of DDIM " + }, + { + "bbox": [ + 46, + 333, + 289, + 644 + ], + "type": "inline_equation", + "content": "(\\eta = 0)" + }, + { + "bbox": [ + 46, + 333, + 289, + 644 + ], + "type": "text", + "content": " adds more materialistic details to the textures on smooth/low-poly geometries. We showcase some examples with particularly large improvements in Fig. 7. Second, we explore the use of ControlNet[98], which can be easily substituted as the LDM backbone for our method without any additional changes. We find that its high-resolution depth conditioning allows TexFusion to capture fine-grained geometric details in the input mesh. In Fig. 8, we further compare TexFusion + ControlNet in \"normal mode\" (apply classifier-free guidance to text prompt only) and ControlNet's \"guess mode\" (apply classifier-free guidance to both text and depth) on meshes with fine geometric details. TexFusion produces high-contrast textures with the appearance of strong lighting under \"guess mode\", and realistic textures with smooth lighting in \"normal mode\". These modifications improve details at a cost of reduced robustness. The non-stochastic DDIM setting may create artifacts when a smooth and clean texture is desired. On the other hand, the increase in depth resolution makes TexFusion+ControlNet susceptible to capturing face boundaries on low-poly meshes. We provide visualizations of these failure cases in Fig. 9. Nonetheless, these modifications offer further dimensions of control that can be explored by practitioners." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 657, + 204, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 204, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 204, + 669 + ], + "type": "text", + "content": "6. Conclusion and Limitations" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "We presented TexFusion, a novel approach to text-driven texture synthesis for given 3D geometries, using only large-scale text-guided image diffusion models. TexFusion lever" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 319, + 89, + 417, + 196 + ], + "blocks": [ + { + "bbox": [ + 318, + 80, + 419, + 87 + ], + "lines": [ + { + "bbox": [ + 318, + 80, + 419, + 87 + ], + "spans": [ + { + "bbox": [ + 318, + 80, + 419, + 87 + ], + "type": "text", + "content": "TexFusion + ControlNet in \"normal\" mode" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 89, + 417, + 196 + ], + "lines": [ + { + "bbox": [ + 319, + 89, + 417, + 196 + ], + "spans": [ + { + "bbox": [ + 319, + 89, + 417, + 196 + ], + "type": "image", + "image_path": "8131710eb84cee6fdd39c315df4f2a089d0175148a010ede602cde4daaaa5859.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 200, + 511, + 209 + ], + "lines": [ + { + "bbox": [ + 333, + 200, + 511, + 209 + ], + "spans": [ + { + "bbox": [ + 333, + 200, + 511, + 209 + ], + "type": "text", + "content": "\"Portrait of greek-egyptian deity hermanubis, lapis skin and gold clothing\"" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 217, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 305, + 217, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 545, + 240 + ], + "type": "text", + "content": "Figure 8: Left: TexFusion + ControlNet in \"normal mode\"; right: TexFusion + ControlNet in \"guess mode\"." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 425, + 88, + 518, + 196 + ], + "blocks": [ + { + "bbox": [ + 425, + 80, + 524, + 87 + ], + "lines": [ + { + "bbox": [ + 425, + 80, + 524, + 87 + ], + "spans": [ + { + "bbox": [ + 425, + 80, + 524, + 87 + ], + "type": "text", + "content": "TexFusion + ControlNet in \"guess\" mode" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 425, + 88, + 518, + 196 + ], + "lines": [ + { + "bbox": [ + 425, + 88, + 518, + 196 + ], + "spans": [ + { + "bbox": [ + 425, + 88, + 518, + 196 + ], + "type": "image", + "image_path": "342524332e5f027f4c6422333f0c6a22293ac48ccb4a9a180ea6d72f42b3ca4f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 323, + 261, + 427, + 312 + ], + "blocks": [ + { + "bbox": [ + 321, + 247, + 411, + 262 + ], + "lines": [ + { + "bbox": [ + 321, + 247, + 411, + 262 + ], + "spans": [ + { + "bbox": [ + 321, + 247, + 411, + 262 + ], + "type": "text", + "content": "Deterministic DDIM: Artifacts in smooth areas" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 261, + 427, + 312 + ], + "lines": [ + { + "bbox": [ + 323, + 261, + 427, + 312 + ], + "spans": [ + { + "bbox": [ + 323, + 261, + 427, + 312 + ], + "type": "image", + "image_path": "18cfcce3440166573610ceeeaf21b6025f2c0152d1abc84a53ec672a0419e388.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 320, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 306, + 320, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 545, + 342 + ], + "type": "text", + "content": "Figure 9: Failure modes exhibited by (left) TexFusion + non-stochastic DDIM sampling and (right) TexFusion + ControlNet." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 438, + 261, + 472, + 311 + ], + "blocks": [ + { + "bbox": [ + 429, + 247, + 531, + 262 + ], + "lines": [ + { + "bbox": [ + 429, + 247, + 531, + 262 + ], + "spans": [ + { + "bbox": [ + 429, + 247, + 531, + 262 + ], + "type": "text", + "content": "TexFusion+ControlNet: Face boundaries on low poly meshes" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 438, + 261, + 472, + 311 + ], + "lines": [ + { + "bbox": [ + 438, + 261, + 472, + 311 + ], + "spans": [ + { + "bbox": [ + 438, + 261, + 472, + 311 + ], + "type": "image", + "image_path": "ea750aea5dd5f9ba5149e88a15d5419e000fc683e1514c0df820f9b952159cb8.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 495, + 261, + 533, + 312 + ], + "blocks": [ + { + "bbox": [ + 495, + 261, + 533, + 312 + ], + "lines": [ + { + "bbox": [ + 495, + 261, + 533, + 312 + ], + "spans": [ + { + "bbox": [ + 495, + 261, + 533, + 312 + ], + "type": "image", + "image_path": "4f29b5525b3af72c1dc596a52f30a2b77334ba439db004cbbaa2d781d7e7616c.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 356, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 545, + 499 + ], + "type": "text", + "content": "ages the latent diffusion model Stable Diffusion and relies on a new 3D-consistent diffusion model sampling scheme that runs a 2D denoiser network in different views of the object and aggregates the predictions in a latent texture map after each denoising iteration. We find that our method can efficiently generate diverse, high-quality, and globally coherent textures and offers detailed control through text conditioning. Limitations of our approach include a not-yet-ideal sharpness and that texture generation is not real-time. Future work can address these issues; for instance, it would be interesting to leverage the recent literature on faster diffusion model samplers [15, 41, 42, 99]." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 514, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 646 + ], + "type": "text", + "content": "Concurrent Works Since Submission Since the time of writing this paper, new works has appeared in the text to texture space [8, 79]. They improve upon TEXTure in aspects such as adding a refinement procedure to automatically fix low quality areas from the initial texturing process [8], and using images sampled from a text-to-image model and per-prompt finetuning to provide stronger conditioning [79]. TexFusion is methodologically distinct from these methods which are based-on TEXTure. Improvements proposed in these works could be combined with TexFusion in future work." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements We would like to thank Jun Gao for the helpful discussions during the project. Tianshi Cao acknowledges additional income from Vector Scholarships in Artificial Intelligence, which are not in direct support of this work." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4177" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 288, + 704 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In ICML, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 124, + 288, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 124, + 288, + 187 + ], + "spans": [ + { + "bbox": [ + 58, + 124, + 288, + 187 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediff-i: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 188, + 288, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 188, + 288, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 188, + 288, + 220 + ], + "type": "text", + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. arXiv preprint arXiv:2302.08113, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 220, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 220, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 58, + 220, + 288, + 284 + ], + "type": "text", + "content": "[4] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh Susskind. Gaudi: A neural architect for immersive 3d scene generation. arXiv preprint arXiv:2207.13751, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 285, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 285, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 58, + 285, + 288, + 327 + ], + "type": "text", + "content": "[5] Hanqun Cao, Cheng Tan, Zhangyang Gao, Guangyong Chen, Pheng-Ann Heng, and Stan Z. Li. A survey on generative diffusion model. arXiv preprint arXiv:2209.02646, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 327, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 327, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 58, + 327, + 288, + 348 + ], + "type": "text", + "content": "[6] Ignacio Castaño. thekla交通大学 In github.com/Thekla/thekla_italas, 2015. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 349, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 349, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 58, + 349, + 288, + 423 + ], + "type": "text", + "content": "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 424, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 424, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 58, + 424, + 288, + 467 + ], + "type": "text", + "content": "[8] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 468, + 288, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 468, + 288, + 543 + ], + "spans": [ + { + "bbox": [ + 58, + 468, + 288, + 543 + ], + "type": "text", + "content": "[9] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 100–116. Springer, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 544, + 288, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 288, + 585 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 288, + 585 + ], + "type": "text", + "content": "[10] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 586, + 288, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 586, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 53, + 586, + 288, + 628 + ], + "type": "text", + "content": "[11] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In The Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 629, + 288, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 629, + 288, + 660 + ], + "spans": [ + { + "bbox": [ + 53, + 629, + 288, + 660 + ], + "type": "text", + "content": "[12] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. arXiv preprint arXiv:2209.04747, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 661, + 288, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 661, + 288, + 704 + ], + "spans": [ + { + "bbox": [ + 53, + 661, + 288, + 704 + ], + "type": "text", + "content": "[13] Congyue Deng, Chiyu \"Max\" Jiang, Charles R. Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, and Dragomir Anguelov. Nerdi: Single-view nerf synthesis with language-guided diffusion as general image priors. arXiv" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 74, + 545, + 703 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 331, + 74, + 463, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 74, + 463, + 83 + ], + "spans": [ + { + "bbox": [ + 331, + 74, + 463, + 83 + ], + "type": "text", + "content": "preprint arXiv:2212.03267, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 84, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 84, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 312, + 84, + 545, + 115 + ], + "type": "text", + "content": "[14] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 116, + 545, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 116, + 545, + 147 + ], + "spans": [ + { + "bbox": [ + 312, + 116, + 545, + 147 + ], + "type": "text", + "content": "[15] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. GENIE: Higher-Order Denoising Diffusion Solvers. In Advances in Neural Information Processing Systems, 2022. 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 148, + 545, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 148, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 312, + 148, + 545, + 190 + ], + "type": "text", + "content": "[16] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Score-based generative modeling with critically-damped Langevin diffusion. In International Conference on Learning Representations (ICLR), 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 190, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 190, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 190, + 545, + 243 + ], + "type": "text", + "content": "[17] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 244, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 244, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 312, + 244, + 545, + 298 + ], + "type": "text", + "content": "[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 298, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 298, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 312, + 298, + 545, + 340 + ], + "type": "text", + "content": "[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 341, + 545, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 341, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 312, + 341, + 545, + 372 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 373, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 373, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 373, + 545, + 415 + ], + "type": "text", + "content": "[21] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. arXiv preprint arXiv:2106.15282, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 416, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 416, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 312, + 416, + 545, + 447 + ], + "type": "text", + "content": "[22] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 448, + 545, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 448, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 312, + 448, + 545, + 490 + ], + "type": "text", + "content": "[23] Wenlong Huang, Brian Lai, Weijian Xu, and Zhuowen Tu. 3d volumetric modeling with introspective neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33(01), pages 8481-8488, 2019. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 491, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 491, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 312, + 491, + 545, + 521 + ], + "type": "text", + "content": "[24] Aapo Hyvarinen. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6:695-709, 2005. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 522, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 522, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 312, + 522, + 545, + 564 + ], + "type": "text", + "content": "[25] Moritz Ibing, Isaak Lim, and Leif P. Kobbelt. 3d shape generation with grid-based implicit functions. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 565, + 545, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 565, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 312, + 565, + 545, + 596 + ], + "type": "text", + "content": "[26] Ajay Jain, Ben Mildenhall, Jonathan T. Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 597, + 545, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 597, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 597, + 545, + 628 + ], + "type": "text", + "content": "[27] Nikolai Kalischek, Torben Peters, Jan D Wegner, and Konrad Schindler. Tetrahedral diffusion models for 3d shape generation. arXiv preprint arXiv:2211.13220, 2022. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 629, + 545, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 629, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 312, + 629, + 545, + 670 + ], + "type": "text", + "content": "[28] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 672, + 545, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 672, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 312, + 672, + 545, + 692 + ], + "type": "text", + "content": "[29] Carson Katri. Dream textures. In github.com/carson-katri/dream-textures, 2022. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 693, + 545, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 693, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 312, + 693, + 545, + 703 + ], + "type": "text", + "content": "[30] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "4178" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 705 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 287, + 95 + ], + "type": "text", + "content": "Song. Denoising diffusion restoration models. arXiv preprint arXiv:2201.11793, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 95, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 95, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 53, + 95, + 287, + 137 + ], + "type": "text", + "content": "[31] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Popa Tiberiu. Clip-mesh: Generating textured meshes from text using pretrained image-text models. December 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 137, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 137, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 54, + 137, + 288, + 169 + ], + "type": "text", + "content": "[32] Roman Klokov, Edmond Boyer, and Jakob Verbeek. Discrete point flow networks for efficient point cloud generation. 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 169, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 169, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 54, + 169, + 288, + 212 + ], + "type": "text", + "content": "[33] Wei-Jan Ko, Hui-Yu Huang, Yu-Liang Kuo, Chen-Yi Chiu, Li-Heng Wang, and Wei-Chen Chiu. Rpg: Learning recursive point cloud generation. arXiv preprint arXiv:2105.14322, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 212, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 212, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 54, + 212, + 288, + 255 + ], + "type": "text", + "content": "[34] Johannes Kopf, Chi-Wing Fu, Daniel Cohen-Or, Oliver Deussen, Dani Lischinski, and Tien-Tsin Wong. Solid texture synthesis from 2d exemplars. ACM Trans. Graph., 26(3):2-es, jul 2007. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 255, + 288, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 288, + 298 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 288, + 298 + ], + "type": "text", + "content": "[35] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 298, + 288, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 298, + 288, + 329 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 288, + 329 + ], + "type": "text", + "content": "[36] Sylvain Lefebvre and Hugues Hoppe. Appearance-space texture synthesis. ACM Trans. Graph., 25(3):541-548, jul 2006. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 329, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 288, + 361 + ], + "type": "text", + "content": "[37] Chun-Liang Li, Manzil Zaheer, Yang Zhang, Barnabas Poczos, and Ruslan Salakhutdinov. Point cloud gan. arXiv preprint arXiv:1810.05795, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 361, + 288, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 288, + 405 + ], + "type": "text", + "content": "[38] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 404, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 404, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 54, + 404, + 288, + 446 + ], + "type": "text", + "content": "[39] Ruihui Li, Xianzhi Li, Ke-Hei Hui, and Chi-Wing Fu. SPGAN:sphere-guided 3d shape generation and manipulation. ACM Transactions on Graphics (Proc. SIGGRAPH), 40(4), 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 447, + 288, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 447, + 288, + 500 + ], + "spans": [ + { + "bbox": [ + 54, + 447, + 288, + 500 + ], + "type": "text", + "content": "[40] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 500, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 500, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 500, + 288, + 542 + ], + "type": "text", + "content": "[41] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 543, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 543, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 54, + 543, + 288, + 586 + ], + "type": "text", + "content": "[42] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv:2206.00927, 2022. 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 586, + 288, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 586, + 288, + 629 + ], + "spans": [ + { + "bbox": [ + 54, + 586, + 288, + 629 + ], + "type": "text", + "content": "[43] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. arXiv preprint arXiv:2201.09865, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 629, + 288, + 672 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 288, + 672 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 288, + 672 + ], + "type": "text", + "content": "[44] A. Luo, T. Li, W. Zhang, and T. Lee. Surfgen: Adversarial 3d shape synthesis with explicit surface discriminators. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 672, + 288, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 672, + 288, + 705 + ], + "spans": [ + { + "bbox": [ + 54, + 672, + 288, + 705 + ], + "type": "text", + "content": "[45] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 705 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 331, + 73, + 449, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 449, + 83 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 449, + 83 + ], + "type": "text", + "content": "Recognition (CVPR), 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 83, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 83, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 312, + 83, + 545, + 126 + ], + "type": "text", + "content": "[46] Siwei Lyu. Interpretation and generalization of score matching. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, UAI '09, page 359-366, Arlington, Virginia, USA, 2009. AUAI Press. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 126, + 545, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 126, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 312, + 126, + 545, + 169 + ], + "type": "text", + "content": "[47] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 169, + 545, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 169, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 312, + 169, + 545, + 212 + ], + "type": "text", + "content": "[48] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. 1, 2, 4, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 212, + 545, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 212, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 312, + 212, + 545, + 265 + ], + "type": "text", + "content": "[49] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 312, + 265, + 545, + 308 + ], + "type": "text", + "content": "[50] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 312, + 308, + 545, + 352 + ], + "type": "text", + "content": "[51] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 351, + 545, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 351, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 312, + 351, + 545, + 395 + ], + "type": "text", + "content": "[52] Gimin Nam, Mariem Khlifi, Andrew Rodriguez, Alberto Tono, Linqi Zhou, and Paul Guerrero. 3d-ldm: Neural implicit 3d shape generation with latent diffusion models. arXiv preprint arXiv:2212.00842, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 395, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 395, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 312, + 395, + 545, + 426 + ], + "type": "text", + "content": "[53] Charlie Nash, Yaroslav Ganin, S. M. Ali Eslami, and Peter W. Battaglia. *Polygen: An autoregressive generative model of 3d meshes.* ICML, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 545, + 480 + ], + "type": "text", + "content": "[54] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 480, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 480, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 312, + 480, + 545, + 522 + ], + "type": "text", + "content": "[55] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 522, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 522, + 545, + 555 + ], + "type": "text", + "content": "[56] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 555, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 555, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 312, + 555, + 545, + 609 + ], + "type": "text", + "content": "[57] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 609, + 545, + 672 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 609, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 312, + 609, + 545, + 672 + ], + "type": "text", + "content": "[58] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 673, + 545, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 673, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 312, + 673, + 545, + 705 + ], + "type": "text", + "content": "[59] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "type": "text", + "content": "4179" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 708 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 181 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 181 + ], + "type": "text", + "content": "[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 182, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 224 + ], + "type": "text", + "content": "[61] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 225, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 288, + 257 + ], + "type": "text", + "content": "[62] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 3, 4, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 258, + 288, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 320 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 320 + ], + "type": "text", + "content": "[63] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, June 2022. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 322, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 322, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 53, + 322, + 288, + 365 + ], + "type": "text", + "content": "[64] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image dissusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 365, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 288, + 407 + ], + "type": "text", + "content": "[65] Chitwan Sahara, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. arXiv preprint arXiv:2111.05826, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 407, + 288, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 288, + 483 + ], + "type": "text", + "content": "[66] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 483, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 483, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 53, + 483, + 288, + 525 + ], + "type": "text", + "content": "[67] Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. arXiv preprint arXiv:2104.07636, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 526, + 288, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 526, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 53, + 526, + 288, + 558 + ], + "type": "text", + "content": "[68] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations (ICLR), 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 559, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 288, + 601 + ], + "type": "text", + "content": "[69] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. arXiv preprint arXiv:2110.02624, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 601, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 288, + 643 + ], + "type": "text", + "content": "[70] Hiroshi Sasaki, Chris G. Willcocks, and Toby P. Breckon. UNIT-DDPM: Unpaired image translation with denoising diffusion probabilistic models. arXiv preprint arXiv:2104.05358, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 644, + 288, + 697 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 644, + 288, + 697 + ], + "spans": [ + { + "bbox": [ + 53, + 644, + 288, + 697 + ], + "type": "text", + "content": "[71] Dong Wook Shu, Sung Woo Park, and Junseok Kwon. 3d point cloud generative adversarial network based on tree structured graph convolutions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3859-3868, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 698, + 288, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 698, + 288, + 708 + ], + "spans": [ + { + "bbox": [ + 53, + 698, + 288, + 708 + ], + "type": "text", + "content": "[72] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 707 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "type": "text", + "content": "Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. arXiv preprint arXiv:2211.16677, 2022. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 106, + 545, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 106, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 312, + 106, + 545, + 192 + ], + "type": "text", + "content": "[73] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In Shai Avidan, Gabriel J. Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part III, volume 13663 of Lecture Notes in Computer Science, pages 72-88. Springer, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 193, + 545, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 193, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 312, + 193, + 545, + 235 + ], + "type": "text", + "content": "[74] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 236, + 545, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 236, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 312, + 236, + 545, + 267 + ], + "type": "text", + "content": "[75] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 267, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 267, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 312, + 267, + 545, + 320 + ], + "type": "text", + "content": "[76] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 321, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 321, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 312, + 321, + 545, + 353 + ], + "type": "text", + "content": "[77] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. arXiv preprint arXiv:2203.08382, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 353, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 353, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 312, + 353, + 545, + 396 + ], + "type": "text", + "content": "[78] Yongbin Sun, Yue Wang, Ziwei Liu, Joshua E Siegel, and Sanjay E Sarma. Pointgrow: Autoregressively learned point cloud generation with self-attention. In Winter Conference on Applications of Computer Vision, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 396, + 545, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 396, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 312, + 396, + 545, + 416 + ], + "type": "text", + "content": "[79] Zhibin Tang and Tiantong He. Text-guided high-definition consistency texture model. ArXiv, abs/2305.05901, 2023. 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 417, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 417, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 312, + 417, + 545, + 449 + ], + "type": "text", + "content": "[80] Greg Turk. Texture synthesis on surfaces. SIGGRAPH '01, page 347-354, New York, NY, USA, 2001. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 545, + 480 + ], + "type": "text", + "content": "[81] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 481, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 481, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 312, + 481, + 545, + 522 + ], + "type": "text", + "content": "[82] Diego Valsesia, Giulia Fracastoro, and Enrico Magli. Learning localized generative models for 3d point clouds via graph convolution. In International Conference on Learning Representations (ICLR) 2019, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 523, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 523, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 312, + 523, + 545, + 554 + ], + "type": "text", + "content": "[83] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 23(7):1661-1674, 2011. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 555, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 555, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 312, + 555, + 545, + 597 + ], + "type": "text", + "content": "[84] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022. 1, 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 597, + 545, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 597, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 312, + 597, + 545, + 651 + ], + "type": "text", + "content": "[85] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jian-min Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3d digital avatars using diffusion. arXiv preprint arXiv:2212.06135, 2022. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 651, + 545, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 651, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 312, + 651, + 545, + 707 + ], + "type": "text", + "content": "[86] Li-Yi Wei and Marc Levoy. Texture synthesis over arbitrary manifold surfaces. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '01, page 355-360, New York, NY, USA, 2001. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "type": "text", + "content": "4180" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 706 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 126 + ], + "type": "text", + "content": "[87] Cheng Wen, Baosheng Yu, and Dacheng Tao. Learning progressive point embeddings for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10266-10275, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 126, + 288, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 170 + ], + "type": "text", + "content": "[88] Li-Yi Wie, Sylvain Lefebvre, Vivek Kwatra, and Greg Turk. State of the Art in Example-based Texture Synthesis. In M. Pauly and G. Greiner, editors, Eurographics 2009 - State of the Art Reports. The Eurographics Association, 2009. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "type": "text", + "content": "[89] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 224, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 277 + ], + "type": "text", + "content": "[90] Fanbo Xiang, Zexiang Xu, Milos Hasan, Yannick Hold-Geoffroy, Kalyan Sunkavalli, and Hao Su. Neutex: Neural texture mapping for volumetric neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7119-7128, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 277, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 287, + 342 + ], + "type": "text", + "content": "[91] Jianwen Xie, Yifei Xu, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Zhu Song-Chun, and Ying Nian Wu. Generative pointnet: Deep energy-based learning on unordered point sets for 3d generation, reconstruction and classification. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 342, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 287, + 396 + ], + "type": "text", + "content": "[92] Jianwen Xie, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Zhu Song-Chun, and Ying Nian Wu. Learning descriptor networks for 3d shape synthesis and analysis. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 396, + 287, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 438 + ], + "type": "text", + "content": "[93] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. PointFlow: 3D point cloud generation with continuous normalizing flows. 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 438, + 287, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 438, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 53, + 438, + 287, + 471 + ], + "type": "text", + "content": "[94] Ling Yang, Zhilong Zhang, and Shenda Hong. Diffusion models: A comprehensive survey of methods and applications. arXiv preprint arXiv:2209.00796, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 471, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 471, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 471, + 287, + 524 + ], + "type": "text", + "content": "[95] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations. In Proceedings of International Conference on Computer Vision (ICCV), 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 524, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 524, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 53, + 524, + 287, + 535 + ], + "type": "text", + "content": "[96] Jonathan Young. xatlas. In github.com/jpcy/xatlas, 2016. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 535, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 535, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 535, + 287, + 588 + ], + "type": "text", + "content": "[97] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 588, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 588, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 53, + 588, + 287, + 610 + ], + "type": "text", + "content": "[98] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 610, + 287, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 610, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 53, + 610, + 287, + 640 + ], + "type": "text", + "content": "[99] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv:2204.13902, 2022. 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 641, + 287, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 641, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 48, + 641, + 287, + 695 + ], + "type": "text", + "content": "[100] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yi-nan Zhang, Antonio Torralba, and Sanja Fidler. Image {gan}s meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 696, + 287, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 696, + 287, + 706 + ], + "spans": [ + { + "bbox": [ + 48, + 696, + 287, + 706 + ], + "type": "text", + "content": "[101] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 330, + 72, + 545, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 330, + 72, + 545, + 106 + ], + "type": "text", + "content": "and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 733, + 314, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 733, + 314, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 733, + 314, + 742 + ], + "type": "text", + "content": "4181" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/c18683e6-ab91-4cf4-bf6b-203837440b00_content_list.json b/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/c18683e6-ab91-4cf4-bf6b-203837440b00_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..90a86f4ee06ae1cba0add9cd58558d50bbec06cf --- /dev/null +++ b/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/c18683e6-ab91-4cf4-bf6b-203837440b00_content_list.json @@ -0,0 +1,1645 @@ +[ + { + "type": "text", + "text": "Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models", + "text_level": 1, + "bbox": [ + 186, + 130, + 782, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaewoong Lee* KAIST", + "bbox": [ + 143, + 204, + 269, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Republic of Korea", + "bbox": [ + 133, + 239, + 281, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hello3196@kaist.ac.kr", + "bbox": [ + 112, + 258, + 300, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sangwon Jang* Yonsei University", + "bbox": [ + 346, + 204, + 483, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Republic of Korea", + "bbox": [ + 341, + 239, + 488, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "agwmon@gmail.com", + "bbox": [ + 341, + 258, + 486, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaehyeong Jo, Jaehong Yoon KAIST", + "bbox": [ + 576, + 204, + 805, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Republic of Korea", + "bbox": [ + 617, + 239, + 764, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{harryjo97, jaehong.yoon}@kaist.ac.kr", + "bbox": [ + 529, + 258, + 856, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunj Kim, Jin-Hwa Kim, Jung-Woo Ha Naver AI Lab", + "bbox": [ + 207, + 281, + 527, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Republic of Korea", + "bbox": [ + 295, + 316, + 441, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yunji.kim, jinhwa.kim, jungwoo.ha}@navercorp.com", + "bbox": [ + 151, + 338, + 584, + 352 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sung Ju Hwang KAIST", + "bbox": [ + 660, + 282, + 787, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Republic of Korea", + "bbox": [ + 651, + 316, + 797, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sjhwang82@kaist.ac.kr", + "bbox": [ + 630, + 338, + 818, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 387, + 312, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Token-based masked generative models are gaining popularity for their fast inference time with parallel decoding. While recent token-based approaches achieve competitive performance to diffusion-based models, their generation performance is still suboptimal as they sample multiple tokens simultaneously without considering the dependence among them. We empirically investigate this problem and propose a learnable sampling model, Text-Conditioned Token Selection (TCTS), to select optimal tokens via localized supervision with text information. TCTS improves not only the image quality but also the semantic alignment of the generated images with the given texts. To further improve the image quality, we introduce a cohesive sampling strategy, Frequency Adaptive Sampling (FAS), to each group of tokens divided according to the self-attention maps. We validate the efficacy of TCTS combined with FAS with various generative tasks, demonstrating that it significantly outperforms the baselines in image-text alignment and image quality. Our text-conditioned sampling framework further reduces the original inference time by more than $50\\%$ without modifying the original generative model.", + "bbox": [ + 75, + 417, + 473, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 748, + 207, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the flood of generative AI systems for vision domains, text-conditional image generation [26, 33, 42] is coming to the fore in recent years. Although many recent works have achieved success in synthesizing high-quality images [19, 34] with plausible class-alignment in class-conditional cases [4, 8], text-to-image (T2I) generation is more challenging since generating visual outputs", + "bbox": [ + 75, + 773, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that are semantically aligned with input texts is a nontrivial problem. We can roughly categorize the works on text-to-image generation into transformer-based autoregressive (AR) [12, 30, 33] and diffusion-based [18, 29] approaches. Along with the advancement of language models, AR models using transformers have shown impressive performance in text-to-image generation. Despite their success, they suffer from the problem of unidirectional bias, which is undesirable for image generation, and crucially, the sampling process requires over 10 times as much time compared to existing models. Another line of work is diffusion-based methods that aim to generate images by iteratively denoising noisy samples. In particular, several continuous diffusion models [10, 32] have achieved outstanding performance and reduced computational cost. Yet, they require excessive sampling steps to obtain high-quality images during the inference.", + "bbox": [ + 500, + 388, + 893, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, a new family of generative models, called token-based diffusion model [6, 23, 41], has emerged as an alternative to tackle the problem of text-to-image generation. Token-based diffusion models quantize the latent features into tokens and apply categorical corruption process in discrete state spaces [1], while conventional diffusion models use Gaussian noise in continuous space. Among various discrete diffusion methods, mask-based diffusion, similar to the absorbing state diffusion used in [1], is mostly used. Compared to existing AR models, this token-based approach is advantageous for speeding up the generation process via simultaneously sampling multiple tokens. Despite the limitation on the reconstruction capacity, these token-based diffusion models significantly outperform the competitors in terms of FID scores, even with fewer sampling steps compared to the continuous diffusion models [5, 37].", + "bbox": [ + 496, + 636, + 895, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, sampling multiple tokens at once often leads", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 94, + 886, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23252", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfae35d70e7ecfe92b1099a9975ede949fd90117189502f79dc084a5d2b6d3e3.jpg", + "image_caption": [ + "A man sitting on top of a bench near the ocean.", + "A box containing an interesting variety of donuts." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 310, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c8d7819a9fcbcb251285c601da72f376bdb81f8d84ae948374c3df55d7da4e2a.jpg", + "image_caption": [ + "The train is rounding the bend of a track on the mountain side.", + "A man standing in a field with some cattle.", + "A single seagull standing on the coast with waves in the background.", + "Autumn leaves radiate a golden glow in afternoon sunlight." + ], + "image_footnote": [], + "bbox": [ + 398, + 88, + 627, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e973a1da5d3c0c869038661864c295908823cf4d612a703b514d61019eeeae8b.jpg", + "image_caption": [ + "Figure 1: Generated samples on MS-COCO dataset and evaluation graph of various sampling methods showing their trade-off. Uniform sampling is a fixed strategy with notably poor text alignment compared to other methods (FID-40K: 15.61, MID-L: 21.23). Random revoke sampling is a revocable strategy with improved text alignment (FID-40K: 16.81, MID-L: 26.98). Ours is TCTS combined with FAS, where both the image quality and the text alignment are significantly better compared to those of baselines (FID-40K: 13.6, MID-L: 29.5). Metrics are measured on all 40K images with their corresponding single caption. The classifier-free guidance scale was fixed at 5 for all sampling methods." + ], + "image_footnote": [], + "bbox": [ + 709, + 93, + 890, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to inconsistency throughout a generated image [37]. For each location, the generator outputs a probability distribution that are coherent with each other. However, there is no guarantee that every single token sampled from the distributions will perfectly align with one another. In other words, it is still possible for the generator to sample incompatible tokens regardless of the generator's capability, leading to potentially nonsensical outputs [24, 37]. This results in a trade-off between the sampling steps and generation quality. Reducing the sampling steps leads to faster generation but results in performance degradation due to a large number of simultaneously sampled tokens. Especially, this problem further stands out in the text-to-image generation tasks since the distribution of text-aligned images is more restricted than the distribution of unconditioned or class-conditioned images.", + "bbox": [ + 75, + 376, + 470, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose a novel sampling approach that refines the images during the diffusion process based on the text condition, which we refer to as the Text Conditioned Token Selection (TCTS). TCTS is a sampling strategy that can mask out and resample previously sampled tokens, which we refer to as revocable. To find the tokens that do not align with the given text condition, we propose to train a model that selects tokens to be masked out and re-generated, to be well-aligned with the given text. Combining this approach with the revocable sampling scheme, TCTS can generate high-quality images with improved text-alignment in even fewer sampling steps compared to the naive generative model, as shown in Figure 1. We further introduce Frequency Adaptive Sampling (FAS) to solve the over-simplification problem that occurs when applying revocable methods for relatively longer steps. FAS leverages the self-attention map of the images and applies a mixed sampling method, which we call persistent sampling to prevent the issue. We summarize our contributions as follows:", + "bbox": [ + 75, + 613, + 470, + 898 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We experimentally show that the revocable sampling strategies are crucial for the trade-off between the text-alignment and the image quality and provide in-depth analysis compared to previous fixed sampling methods.", + "- We propose a novel revocable sampling method based on a learnable token selection model that significantly improves the text alignment as well as the image quality even with fewer sampling steps, without the need of retraining the generator model.", + "- Moreover, we propose a novel sampling method based on the self-attention map that can be combined with TCTS to solve the over-simplification problem." + ], + "bbox": [ + 517, + 376, + 890, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 593, + 635, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various works have tackled text-to-image generation tasks, and the majority of them are based on Generative Adversarial Networks (GANs) [38, 40, 44, 45, 46, 47, 48]. However, computing directly from the pixel space of an image is challenging, and generating high-resolution images requires substantial computation due to the large number of pixels involved. Therefore, a two-stage approach is often used: models first tokenize the images into a sequence of codes, and then predict the tokenized codes.", + "bbox": [ + 496, + 618, + 893, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoregressive models Following the success of Transformer [39] and GPT [28], Autoregressive (AR) models [9, 30, 31, 42] proposed to tokenize images and pose the text-to-image generation problem as a problem of translating textual descriptions into image tokens, much like machine translation of one language into a different language. With the advancement of image tokenization and language models, this approach has achieved very high performance. However, even with the help of this tokenizing stage, there remains", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23253", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a large number of tokens to predict, resulting in a significant delay in the generation time due to their autoregressive nature. Additionally, errors are accumulated through the irrevocable process, with undesirable unidirectional bias.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent diffusion models Recently, diffusion models [18, 36] introduced a new paradigm for text-to-image generation [33]. They generate high-quality images by progressively denoising noise, but require iterative steps and a large amount of computation. Two-stage approaches [10, 32] propose to perform diffusion in latent space successfully reducing computation while maintaining performance. However, most of them still require about 50 inference steps to show plausible performance, and since denoising is applied globally to the entire image during the inference process, locally correcting the image requires complicated processes [32].", + "bbox": [ + 75, + 159, + 472, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Token-based diffusion models Token-based diffusion models [5, 11, 13, 23, 43] have impressive performance in generating complex scenes from the text. Unlike conventional diffusion models, where the Gaussian noise is applied in a continuous space, token-based diffusion models quantize latent features into tokens and apply a categorical corruption process in a discrete state space [1]. Among various discrete diffusion methods, mask-based diffusion is most commonly used due to its attractive properties; they have the advantage of sampling multiple tokens simultaneously, resulting in fewer sampling steps and faster generation time than the latent diffusion model. However, selecting multiple tokens at once leads to inconsistency throughout the image which is often called a joint distribution issue. Moreover, tokens are heavily influenced by the spatially adjacent tokens that have already been modified so that a misaligned token could disperse the errors throughout the image. To tackle this issue, some works [6, 23] modified the sampling strategy while other works [24] introduced new models to find the misaligned tokens.", + "bbox": [ + 75, + 334, + 472, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In particular, Lezama et al. [24] proposed a learnable model that removes the misaligned tokens at each step. While most existing works randomly select the tokens to discard, this work examines all sampled tokens. It outputs the score of each location where the score indicates whether the corresponding token is under the real distribution. After the predicting and sampling process of the generator, the model selects the locations according to the scores of the corresponding tokens.", + "bbox": [ + 75, + 637, + 470, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Text-conditioned token selection", + "text_level": 1, + "bbox": [ + 76, + 784, + 372, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The diffusion process of the masked image generative models starts with a completely masked image, and samples the tokens of $x_{t}$ for the whole image to predict $\\hat{x}_0$ while re-masking a portion of them to obtain the denoised sample $x_{t - 1}$ . Existing masked image generative models use uniform or confidence-based sampling strategy to determine", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9637f9b9ddc0ff5c094d15ece8ae9a5060a6945e44a638766701d5786753e0eb.jpg", + "image_caption": [ + "Figure 2: Reconstructed images using $\\hat{\\mathbf{x}}_0^{(\\mathrm{t})}$ each step during diffusion using the fixed and revocable method. Only the revocable method is able to edit the tokens to generate a plane according to the text \"A view of the end of an airplane in the sky over mountains.\"" + ], + "image_footnote": [], + "bbox": [ + 503, + 89, + 893, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the locations to re-mask: Uniform sampling [13, 23] randomly selects the locations of the tokens to keep among the predicted tokens at all locations, while confidence-based strategy, such as purity sampling [37], selects the locations based on the confidence score which is defined as follows:", + "bbox": [ + 496, + 349, + 893, + 425 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C S} (i, t) = \\max _ {j = 1, \\dots , K} p \\left(x _ {0} ^ {i} = j \\mid x _ {t} ^ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 450, + 893, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t$ is the step number, $K$ is the size of the codebook, and $i$ is the location of the token. Intuitively, a high confidence score means the generator is relatively certain that a token should be sampled at the corresponding location. By keeping these confident tokens while discarding the ambiguous ones, purity sampling shows good performance in practice [5].", + "bbox": [ + 496, + 488, + 890, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, these strategies are non-revocable, i.e., fixed, which means tokens that are once sampled cannot be revised afterward, even if they do not match the given text or do not align with other tokens. Once the misaligned tokens are fixed during the process, errors accumulate and spread throughout the image. We observe that fixed strategies negatively affect the text alignment as visualized in Figure 1.", + "bbox": [ + 496, + 579, + 893, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Random revoke sampling strategy", + "text_level": 1, + "bbox": [ + 500, + 696, + 797, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to address this issue, we first introduce a new sampling strategy called random revoke sampling and use it as a baseline for our final method. It is similar to uniform sampling in a way that they both determine the locations by sampling from a uniform distribution. However, while uniform sampling is a fixed strategy, random revoke sampling selects the locations from the whole image space regardless of the previously fixed locations, making the process reversible. It randomly selects a certain number of tokens to preserve for the next step and revokes the previously fixed ones. The revocable feature can better align the images to the given text information, especially in complex scenes.", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23254", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e551c84949edbed60b486d5b3747aee997496a7a56e952bf7507dd7fbd43db19.jpg", + "image_caption": [ + "Figure 3: Reconstructed images using $\\hat{\\mathbf{x}}_0^{(\\mathrm{t})}$ each step during diffusion using the revocable method with and without FAS. The backgrounds of the images in the top row are over-simplified while our proposed FAS prevents this, as shown in the bottom row. The text is \"This small bird is greyish in color with flecks of yellow on the back and breast, and a bit of white on the belly.\"" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 467, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To be specific, as shown in the top row of Figure 2, fixed strategies have limitations in recovering from errors or generating diverse images due to their inability to regenerate incorrect tokens or modify missing essential objects or attributes. In contrast, revocable strategies enhance the text alignment by providing an opportunity to correct invalidly generated tokens, as shown in the bottom row of Figure 2. Since revocable strategies have no constraints on which tokens to remove or fix, they have the advantage of mitigating error accumulation.", + "bbox": [ + 75, + 364, + 472, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, random revoke sampling suffers from two main issues that must be addressed. The first issue is the instability resulting from randomly discarding previously sampled tokens which worked as a given condition of the step. It can lower the quality of the outputs as in Figure 1.", + "bbox": [ + 75, + 518, + 470, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The second issue, namely the over-simplification, arises when it is used in the generation process involving longer steps. The low-frequency parts of the output images, such as backgrounds, tend to be over-simplified compared to fixed methods, which generate more realistic backgrounds. The over-simplified samples can be seen in Figure 3, and it has a significant impact on the FID metric as in Table 1 and 2. We observe that this is due to excessive repetition of resampling over a large number of steps, while a fixed schedule does not resample any tokens. More opportunities for resampling make the tokens converge towards safer and simpler patterns. As shown in the top row of Figure 3, the high-frequency areas with rich visual details do not become over-simplified even after multiple resampling processes, but the low-frequency areas become less diverse and more simplified in patterns quickly due to excessive resampling.", + "bbox": [ + 75, + 595, + 470, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To confirm this trade-off between text alignment and image quality, we define a persistent sampling as an interpolation between uniform sampling and random revoke sampling. In details, a persistent weight boosts the probability of keep", + "bbox": [ + 75, + 839, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Persistent Sampling", + "text_level": 1, + "bbox": [ + 500, + 92, + 723, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: $k_{t}$ : number of tokens sampled at step $t$ , $w$ : persistent weight, $N$ : number of all tokens", + "bbox": [ + 500, + 109, + 890, + 136 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $x_{T}\\gets [[\\mathrm{MASK}]]_{N}$", + "2: for $t = T, T - 1, \\dots, 1$ do", + "3: $\\hat{x}_0 = G_\\theta(x_t, c)$", + "4: $A_{t} = \\{i|x_{t}^{i} = [\\mathrm{MASK}]\\}$", + "5: $m = k_{T} + k_{T - 1} + \\ldots + k_{t}$", + "6: $\\mathcal{U}_t(I = i|i\\in A_t):\\mathcal{U}_t(I = i|i\\in A_t^C) = 1:w$", + "7: $\\rightarrow$ uniform distribution + persistent weight", + "8: $i_1, i_2, \\dots, i_m \\sim \\mathcal{U}(I) \\gets$ sample without replacement", + "9: $x_{t}\\gets [[\\mathrm{MASK}]]_{N}$", + "10: for $i = i_1, i_2, \\ldots, i_m$ do", + "11: $x_{t}^{i}\\gets \\hat{x}_{0}^{i}$", + "12: end for", + "13: end for", + "14: Return: Generated image $x_0$" + ], + "bbox": [ + 503, + 138, + 854, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing the tokens that were sampled in the previous step. While random revoke sampling assigns the same probability to all the locations, persistent sampling multiplies the persistent weight to the probability of previously kept locations. Further details are shown in Algorithm 1. If the persistent weight $w$ is 1, all the tokens would have the same probability to be fixed, such as random revoke sampling. If $w$ is sufficiently large, the previously sampled tokens would be kept until the end, such as uniform sampling. In the graph in Figure 1, a trade-off can be seen that as the sampling strategy gets closer to the random revoke sampling, text alignment gets better, while the image quality gets worse. We further analyze the effect of these sampling methods in Section 4.5.", + "bbox": [ + 496, + 357, + 893, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Text-conditioned token selection", + "text_level": 1, + "bbox": [ + 500, + 566, + 782, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a learnable model called Text-Conditioned Token Selection (TCTS). TCTS is trained to output a score map to detect tokens that are not under the real distribution for the given text condition, and selects the tokens that are not aligned with the text or other tokens in the image. By masking the selected tokens at each step, the diffusion network receives high-quality text-aligned tokens as the input to sample the next set of tokens which alleviates the error accumulation and joint distribution issue of the previous sampling strategies. Moreover, since TCTS is trained to discriminate the well-aligned tokens, it detours the instability problem as mentioned earlier, unlike other revocable methods. From the experimental results in Section 4, we observe that TCTS enhances the text alignment without compromising the image quality.", + "bbox": [ + 496, + 590, + 893, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Frequency adaptive sampling (FAS) The simplest way to solve the over-simplification problem in longer steps is to use the persistent sampling to alleviate the excessive resampling of tokens. However, as previously mentioned, this method", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23255", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/98df41975cbefd50a342c61b36778d24047b963351f2b819e3512fda5bcd092f.jpg", + "image_caption": [ + "Figure 4: Overall generation framework of proposed TCTS and FAS. After the generator predicts the tokens, TCTS exploits the text condition to detect misaligned tokens and outputs the score map. Meanwhile, FAS splits the tokens according to the frequency using the self-attention map from the generator, performing revocable sampling to high-frequency split and persistent sampling to low-frequency split. The adaptive sampling predicts $\\hat{x}_0$ and decide a few of the locations to mask according to $x_{t}$ . The token maps produced by FAS is combined with the score map to predict $x_{t - 1}$ . After iterative process, our model removes all the masks and generates $x_0$ ." + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 890, + 272 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d73635d14933b74865194f404f2123ba3c2f5282a570d3e37ac5bd8cdb3906b7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StepMethodMID-L ↑SOA-I ↑CLIP-S ↑FID-30K ↓
16Purity11.0272.380.247419.20
Uniform17.9474.800.250016.17
RR23.6078.790.252617.10
TCTS + FAS26.7279.520.255914.45
25Purity16.8475.210.248718.39
Uniform22.2777.080.252415.91
RR26.7781.100.254318.43
TCTS + FAS27.7980.870.256315.39
", + "bbox": [ + 86, + 373, + 470, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "may result in a slight compromise in text alignment and limit the opportunity to correct the image through resampling. To further improve this trade-off, we propose a new method called Frequency Adaptive Sampling (FAS), which can be applied to TCTS.", + "bbox": [ + 75, + 608, + 468, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FAS is a method that utilizes the generator's self-attention map to limit resampling only in the low-frequency areas of the image. In [21], it is known that the generator's self-attention layer contains frequency information of the image, which is also observed in token-based diffusion models. We utilized this to distinguish the frequency areas of the image without additional operations, and applied persistent sampling only to the low-frequency areas using persistent weight. As shown in the bottom row of Figure 3, revocable sampling along with FAS method allows repetitive resampling in areas that require rich visual details, while limiting the number of resampling in relatively simple areas preventing oversimplification. Detailed algorithm is in the supplementary A. The overall framework of our model, including both TCTS", + "bbox": [ + 75, + 688, + 472, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0091a053f9dc2711e2f3c3ad653082bcc459ebde48cae130c51d4b70ffb206c6.jpg", + "table_caption": [ + "Table 1: Quantitative evaluation of sampling methods on MS-COCO dataset. The ground truth MID on MS-COCO image-caption pairs is 54.73. The classifier-free guidance scale was fixed at 5 for all sampling methods." + ], + "table_footnote": [], + "table_body": "
StepMethodMID-L ↑CLIP-S ↑FID ↓
16Purity-24.210.241015.21
Uniform-25.600.240416.57
RR-25.030.237117.38
TCTS + FAS-19.880.24612.35
25Purity-21.260.238412.60
Uniform-23.040.239613.02
RR-23.290.236414.53
TCTS + FAS-18.310.240913.67
", + "bbox": [ + 511, + 373, + 890, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2: Quantitative evaluation of sampling methods on CUB dataset. The ground truth MID on CUB image-caption pairs is 15.85. We omit SOA in this table as CUB dataset consists of only one bird per photo, diminishing the metric's significance.", + "bbox": [ + 498, + 529, + 893, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and FAS, is illustrated in Figure 4.", + "bbox": [ + 500, + 604, + 728, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 633, + 633, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To validate the efficacy of our token selection framework, we modify the transformer from [37] to contain 310M parameters for MS-COCO dataset and 83M parameters for CUB dataset. We extract text features from CLIP ViT-B/32 [27] for a text-conditioned generation. In the training process of TCTS, we freeze the generator's parameters and independently train our model, making it applicable to other types of token-based diffusion models. We use the binary cross-entropy loss for the objective and validate our method and baselines on MS-COCO [25] and CUB [14] datasets by training them for 200 epochs. We leverage classifier-free guidance [20] by stochastic sampling of the guidance strength to improve the generation quality. We further provide the details of the architecture, hyperparameters and additional analysis on the use of classifier-free guidance in the supplementary A.", + "bbox": [ + 496, + 657, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23256", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a7f4c907cdc927124b284e3879e4e5d45c07bdf909b28cd927867df0033c9f8c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMID-L ↑MID-B ↑SOA-I ↑FID-30K ↓
GLIDE [26]1.031.00-32.08
AttnGAN [40]-65.20-8.9039.0129.15
DM-GAN [23]-44.663.5148.0322.90
DF-GAN [38]-58.75-15.21-31.75
VQ-Diffusion [41]-19.635.77-13.13
LAFITE [47]6.2635.1774.788.03
TCTS + FAS26.0238.9879.529.75
", + "bbox": [ + 88, + 88, + 470, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics Since FID [16] metric is known to be problematic for its inability to exactly represent the image quality [5] and unable to consider the text alignment, we additionally use Mutual Information Divergence (MID) [22] to evaluate the text-alignment of generated images which enables sample-wise evaluation and outperforms previous metrics in human Likert-scale judgment correlations. In particular, since MID responds more sensitively to images generated with foiled captions [35], it is more appropriate for analyzing text alignment with complex and challenging captions. Moreover, we use the CLIP score [15] and SOA-I [17], which are widely used in text-to-image synthesis. Note that the CLIP score is calculated with ViT-L/14 throughout the experiments.", + "bbox": [ + 75, + 335, + 473, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Text-to-image synthesis", + "text_level": 1, + "bbox": [ + 76, + 544, + 294, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 1 Left and Middle visualizes the generated images using improved VQ diffusion [37] with varying sampling methods. Generated examples from uniform token sampling are poorly aligned with the given texts, and often include erratic partitions of the objects since the predicted tokens selected at the same step cannot appropriately reflect the change of others (e.g., smashed donuts, and imperfect chair and seagull). Random revoke (RR) sampling, which can re-update previously sampled tokens, seems to improve alignment with text by editing out nonsensical regions and iteratively considering the entire scene at each generation step. Yet, RR selects tokens in a completely random manner without any constraints, which results in suboptimal generation quality and alignment with the texts (e.g., disappeared man and train). On the other hand, ours mitigates the problem from RR by selecting text-conditioned tokens at the sampling phase and successfully generates high-quality images which contain a clear semantic connection to the given text captions.", + "bbox": [ + 75, + 566, + 470, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This phenomenon can be quantitatively seen in Table 1. 16-step generation of random revoke sampling shows better text alignment than even 25-step generation of other fixed", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2d6ff5a6f95806e95523125a66d69388a363a7848df400f3a1e5963933a2d8c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 88, + 888, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c787d8bd112bda7abb8468484b3884afd8c9c3321e9dd3a244d7aa6f3f6d84b2.jpg", + "image_caption": [ + "Figure 5: Results of image refinement using TCTS. Top: Original samples, Bottom: Refined images for 8-steps with TCTS." + ], + "image_footnote": [], + "bbox": [ + 503, + 191, + 888, + 291 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0b9240200803624f4c6e49ba4914a286f0961dcc6999b3ff3baa62fe08184d6e.jpg", + "table_caption": [ + "Table 3: Comparison of the proposed method with recent text-to-image generative models on MS-COCO dataset. MID-L and MID-B are calculated with ViT-L/14 and ViT-B/32 each. The ground truth MID-L and MID-B on MS-COCO image-caption pairs is 54.73, 41.57 each. Ours are evaluated with 16-step setting and we adopt classifier-free guidance from 3 to 5." + ], + "table_footnote": [], + "table_body": "
MethodMID-L ↑FID ↓
Original12.1016.44
RR13.6417.18
TCTS15.6416.92
Original4.5419.13
Random masked6.0619.12
TCTS masked6.4718.64
", + "bbox": [ + 566, + 323, + 831, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4: Qualitative evaluation of the refined images. Top: Refinement with additional revision steps. Bottom: Refinement with masking lowest-scoring tokens.", + "bbox": [ + 498, + 444, + 893, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sampling methods. We can also observe the trade-off between CLIP score and FID by comparing each method to their 25-step versions. Our model outperforms other baselines in most of the metrics especially when comparing MID-L to fixed sampling methods such as Purity and Uniform. Since our model can make corrections during the generation process, it requires fewer steps to match the performance of our baseline sampling methods, see Section 4.5 for further analysis. We also present the results of our experiment on the CUB dataset with fewer parameters in Table 2, which demonstrates satisfactory performance. However, the fact that the CUB dataset contains only single object per image resulted in a tendency that is slightly different from what we mentioned earlier. In Table 3, we compared the performance of our method with other text-to-image generative models. Our model highly exceeds other models in text alignment metrics, especially in MID-L and closely approximate the MID-B value of the ground truth images. More samples generated by our model is in the supplementary B.", + "bbox": [ + 496, + 501, + 893, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Image refinement", + "text_level": 1, + "bbox": [ + 500, + 800, + 671, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The advantage of the masked image generative model is that it enables fast local refinement. We conduct image refinement in two separate methods, inspired by [23] and [24]. First, we use TCTS to apply additional revision steps to images generated with uniform sampling. The overall image", + "bbox": [ + 498, + 824, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23257", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "quality is improved by adding additional refinement steps as shown in Figure 5. In order to demonstrate the image refinement performance of our TCTS, we also measure the performance of the image refinement with random revoke sampling. While RR improves does improve sample quality, additional refinement steps increase the FID score, caused by a similar effect to increasing the generation step.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We additionally perform the experiment by masking $60\\%$ lowest-scoring tokens with TCTS, and generate the images with uniform sampling. We observe that all metrics of the refined images outperform those of the original images. In Table 4, we compare them with samples refined after randomly masking the tokens without TCTS. Due to the page limit, we describe further details of the two refinement methods and visualize more samples in the supplementary B.", + "bbox": [ + 75, + 196, + 473, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Mask-free object editing", + "text_level": 1, + "bbox": [ + 76, + 329, + 302, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since the masked image generative model is capable of local refinement, image editing without manual masking is possible simply by randomly masking a part of the image and resampling it with the new text condition. However, this method requires a low masking ratio and many resampling steps to maintain the overall structure of the image. This can result in significant changes to unnecessary parts or, as mentioned earlier, over-simplification issues. Additionally, even with many steps, editing large objects with small masks can be more challenging due to the significant influence of surrounding tokens rather than the new text condition.", + "bbox": [ + 75, + 352, + 472, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motivated by the operation of self-attention maps in frequency adaptive sampling, we leverage a cross-attention map corresponding to the word of the object that is to be changed, giving weights to resample tokens so that the corresponding locations can be resampled. Then, it enables efficient image editing with fewer steps and makes it easier to edit larger objects. Although it is similar to DiffEdit [7], ours operates in a more straightforward manner without additional computations, thanks to the capability of local refinement of the masked generative model. In Figure 6, we edit the object by adding $25\\%$ noise in only 10-steps. With our method, we can edit more quickly with fewer editing steps, which further does not suffer from the over-simplification problem. More samples are in the supplementary B.", + "bbox": [ + 75, + 518, + 473, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. High-resolution image synthesis", + "text_level": 1, + "bbox": [ + 76, + 741, + 359, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Continuous diffusion models that use encoders and decoders, such as masked image generative models, can generate larger images that are not in the training set. Similar to Bond-Taylor et al. [3], we divide the tokens into subsets according to the model input size, individually pass them through the model, and then spatially aggregate tokens to synthesize a high-resolution image. In this way, we are able to generate more realistic and high-resolution samples with the same TCTS model without additional training. Addition-", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/00548aacf89f2c11b8b5c60d676ea59b691698d44e43eef4bb1a7e34b5ee3b55.jpg", + "image_caption": [ + "A dog teddy bear looks back from the passenger seat of a vehicle." + ], + "image_footnote": [], + "bbox": [ + 503, + 89, + 890, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d3582282c7d74b2a290ae6b339220c34417c771de7e168d4e7236fc6552a4a3.jpg", + "image_caption": [ + "A wooden table topped with earrot cake donuts, broccoli." + ], + "image_footnote": [], + "bbox": [ + 503, + 217, + 890, + 328 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1739847b09f6a36a2b9c07eea36e656d0371110d99d7c27ce027fe26809ffac8.jpg", + "image_caption": [ + "Figure 6: Examples of mask-free editing samples with cross-attention map. The cross-attention map is multiplied to the score map of TCTS to perform weighted sampling.", + "A short train traveling through a mountainous landscape.", + "Figure 7: Examples of high resolution $(512^{2})$ samples. Left: Image generated with all mask tokens, Right: Image generated with TCTS as a super-resolution unit." + ], + "image_footnote": [], + "bbox": [ + 503, + 407, + 694, + 556 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fdbd26e656252abb8373c775bf4bdaf072318cefcd9143841c268fcc99d65d1a.jpg", + "image_caption": [ + "A family on a boat sailing down a river." + ], + "image_footnote": [], + "bbox": [ + 697, + 407, + 890, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ally, we propose a new method to generate high-resolution images only with low-resolution TCTS. To be specific, we first generate a small-size image with TCTS and upsample the token map in bicubic mode to the desired size. Then we divide the high-resolution token map into overlapping small-size sections and refine all the sections several times. Then, we can generate high-resolution images with the low-resolution TCTS where we visualize the generated samples in Figure 7.", + "bbox": [ + 496, + 654, + 893, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Further analysis", + "text_level": 1, + "bbox": [ + 500, + 801, + 663, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importance of early-stage sampling strategy As mentioned in Section 3.1, by controlling the persistent weight, we can interpolate the two baseline sampling methods: uniform sampling and random revoke sampling. We designed various experiments to find out the impact of the sampling", + "bbox": [ + 496, + 825, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23258", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/705f6d7abccccd96d76b6fe280dd3a88e26197ec30e98f22b99bb45a34e8752c.jpg", + "image_caption": [ + "(a) Real" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 176, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32adead52fac8ee739171cd1644ff8159fe44b54bd03b206702e9c3a6d6c20a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 89, + 370, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cae83e708f754a98ac1338a6c0d5f3cd96f6d9ae92e93c5b7ecae2d2047c8302.jpg", + "image_caption": [ + "road" + ], + "image_footnote": [], + "bbox": [ + 370, + 90, + 467, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9420161652a97a82d01ca9af46df72d2602ede70dac5d3351411f6d9736ecc57.jpg", + "image_caption": [ + "(b) Synthesized" + ], + "image_footnote": [], + "bbox": [ + 80, + 181, + 176, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/91ee9af1d6669cd7d757c429b30c13a9c71f772362e97f26955cabed26a68855.jpg", + "image_caption": [ + "A" + ], + "image_footnote": [], + "bbox": [ + 176, + 181, + 272, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/491864ab18b4ebd38582b792720a487d9230f20ab8522995f53bc343dbe67136.jpg", + "image_caption": [ + "boat" + ], + "image_footnote": [], + "bbox": [ + 272, + 181, + 369, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5ec268ce93fb9b89c4f4eb02f4b0f168cf3f26a38b6c0a8fc705989ee7d4069.jpg", + "image_caption": [ + "water", + "Figure 8: Visualization of the averaged cross-attention maps for each word from TCTS. Top: Image from COCO dataset: \"A large black bear walking down a dirty road\". Bottom: Synthesized image: \"A small boat in a narrow body of water\"." + ], + "image_footnote": [], + "bbox": [ + 370, + 181, + 467, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "method with respect to the time steps. We switched the sampling methods during the time steps, from uniform sampling to random revoke sampling (U2R) and vice versa (R2U). In the rightmost plot in Figure 1, we can observe that U2R shows better text-alignment performance than R2U.", + "bbox": [ + 75, + 351, + 468, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These results suggest that the early stage sampling, in which the masked ratio is high, substantially influences the text alignment of the final image, which is similarly observed in continuous diffusion models [2]. Generating the whole image and then processing it through a few revision steps like [23] might help the image quality but can not conspicuously enhance the essential text alignment. Therefore, in order to obtain an image that aligns well with the text, sampling must be done carefully in the early stage when not many tokens are generated. This analysis is also shown in the result that the revocable strategy exhibits better performance by aggressively resampling tokens several times. In particular, since weak generators with poor generation performance suffer more from the joint distribution issues, giving a sufficient number of recovery chances to discard tokens and draw new ones at the early stage is the key to the desired text alignment.", + "bbox": [ + 75, + 428, + 468, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text-conditioned sampling The experiments show that TCTS plays a crucial role in enhancing text alignment in image generation. Although we did not utilize any text-specific loss during training, the model still effectively processes the text condition and achieves optimal performance. To verify the accuracy of TCTS's attention towards text-related parts of the image, we visualized the average cross-attention layer for each word in Figure 8. The results demonstrate TCTS's ability to attend to each word in both real and synthesized images. By helping the selection of relevant tokens during the generating stage, TCTS can boost model performance without the requirement of additional complex text-related loss.", + "bbox": [ + 75, + 703, + 470, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8fcf5099d5ce038907d8c5b892883ab847a2f0df6eaeae68e0daea6c029ea4f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMID-L ↑SOA-I ↑CLIP-S ↑FID-30K ↓
RR26.7781.100.254318.43
RR + FAS26.3280.780.253916.01
RR + pw24.2379.140.253415.96
TCTS27.8280.930.256516.69
TCTS + FAS27.7980.870.256315.39
TCTS + pw25.1279.510.255615.12
", + "bbox": [ + 511, + 88, + 890, + 209 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3bcd286f051fc06beafbda80a3865908ca6b158230fff92cf5038cebbdff1c9a.jpg", + "table_caption": [ + "Table 5: Ablation study of using FAS and persistent weight in revocable methods on MS-COCO dataset. All models were evaluated on 25-step setting. The ground truth MID-L on MS-COCO image-caption pairs is 54.73." + ], + "table_footnote": [], + "table_body": "
StepModelMID-L ↑FID-30K ↓Time
50Uniform24.0416.84×1
RR26.5521.12
16TCTS + FAS26.7214.45×0.45
", + "bbox": [ + 511, + 277, + 890, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6: Inference time relative to MID/FID of our baselines and models on MS-COCO dataset. We expressed the inference time of our models as a multiple of other baselines. Our model outspeeds other baselines while performing better in both MID-L and FID-30K. Further analysis is in the supplementary C.", + "bbox": [ + 498, + 356, + 892, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study of FAS As mentioned in Section 3.2, FAS decides whether to apply persistent weight or not to each location. If the whole image is considered low-frequency, FAS would apply the weight to every location, which is the same as the persistent sampling. On the other hand, if the whole image is considered high-frequency, combining FAS would not apply the weight anywhere, which does not change the sampling strategy. To further analyze the effect of FAS, we evaluated it on two revocable methods: random revoke sampling and TCTS. Attaching FAS to a base sampling method can be regarded as a mixture of two methods: the base revocable sampling method and persistent sampling method. Table 5 clearly shows this relation of FAS and persistent weight. While FAS-attached methods do not show the best performance, the performance of them is always in close proximity to the better one of the two comparing methods of each. FAS-attached methods balances well between text-alignment and image quality without any additional training.", + "bbox": [ + 496, + 448, + 893, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference time Our TCTS and FAS compute additional operations at each generation step, resulting in marginally increased inference time compared to other token selection baselines. However, regarding the text alignment and the image quality, the relative inference time is considerably decreased. In Table 6, our model outperforms 50-step image generation of baseline sampling methods within only 16 steps in MID-L and FID-30K, which means that our model can synthesize better samples with only $\\times 0.45$ time.", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23259", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 89, + 194, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper examines which factors in the masked diffusion process impact the output images and cause the trade-off between image quality and text alignment. We empirically find that the revocable sampling significantly improves the text alignment yet degrades the quality of the generated images. To tackle the problem, we propose a simple token sampling strategy, coined text-conditioned token selection, that combines learnable and revocable methods, pushing the boundary of the trade-off between image quality and text alignment without the necessity to update the large-scale pre-trained generator. We find that collaborative sampling in a persistent and revocable manner surprisingly alleviates over-simplification issues in the generated backgrounds. Our proposed method can be utilized as an image refinement tool for various generative tasks, which is remarkably fast to generate high-quality images within much fewer steps.", + "bbox": [ + 75, + 114, + 472, + 357 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 372, + 264, + 390 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST)) and (No. RS-2022-00187238, Development of Large Korean Language Model Technology for Efficient Pre-training). This work was supported by KAIST-NAVER Hypercreative AI Center.", + "bbox": [ + 75, + 398, + 472, + 518 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 76, + 534, + 174, + 550 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. In Advances in Neural Information Processing Systems, pages 17981-17993. Curran Associates, Inc., 2021. 1, 3", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 8", + "[3] Sam Bond-Taylor, Peter Hessey, Hiroshi Sasaki, Toby P Breckon, and Chris G Willcocks. Unleashing transformers: parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vector-quantized codes. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXIII, pages 170–188. Springer, 2022. 7", + "[4] Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdzal, and Adriana Romero Soriano. Instance-conditioned gan. Advances in Neural Information Processing Systems, 34:27517-27529, 2021. 1", + "[5] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to" + ], + "bbox": [ + 84, + 560, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1, 3, 6", + "[6] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 1, 3", + "[7] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance, 2022. 7", + "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1", + "[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2", + "[10] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. arXiv preprint arXiv:2208.13753, 2022. 1, 3", + "[11] Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what happened: Unifying text-guided video completion via multimodal masked video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10681-10692, 2023. 3", + "[12] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 1", + "[13] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 3", + "[14] Xiangteng He and Yuxin Peng. Fine-grained visual-textual representation learning. IEEE Transactions on Circuits and Systems for Video Technology, 30(2):520-531, 2019. 5", + "[15] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528. Association for Computational Linguistics, 2021. 6", + "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[17] Tobias Hinz, Stefan Heinrich, and Stefan Wermter. Semantic object accuracy for generative text-to-image synthesis. IEEE transactions on pattern analysis and machine intelligence, 2020. 6", + "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 3" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23260", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:47-1, 2022. 1", + "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5", + "[21] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance, 2022. 5", + "[22] Jin-Hwa Kim, Yunji Kim, Jiyoung Lee, Kang Min Yoo, and Sang-Woo Lee. Mutual information divergence: A unified metric for multimodal generative models. arXiv preprint arXiv:2205.13445, 2022. 6", + "[23] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Draft-and-revise: Effective image generation with contextual rq-transformer. arXiv preprint arXiv:2206.04452, 2022. 1, 3, 6, 8", + "[24] José Lezama, Huiwen Chang, Lu Jiang, and Irfan Essa. Improved masked image generation with token-critic. In European Conference on Computer Vision, pages 70-86. Springer, 2022. 2, 3, 6", + "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5", + "[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 6", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5", + "[28] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2", + "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2", + "[31] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 3" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 3", + "[34] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 1", + "[35] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. Foil it! find one mismatch between image and language caption. arXiv preprint arXiv:1705.01359, 2017. 6", + "[36] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3", + "[37] Zhicong Tang, Shuyang Gu, Jianmin Bao, Chen Dong, and Fang Wen. Improved vector quantized diffusion models. arXiv preprint arXiv:2205.16007, 2022. 1, 2, 3, 5, 6", + "[38] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2020. 2, 6", + "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[40] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2, 6", + "[41] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 1, 6", + "[42] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 1, 2", + "[43] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 3", + "[44] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2", + "[45] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stack" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23261", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "gan++: Realistic image synthesis with stacked generative adversarial networks. IEEE transactions on pattern analysis and machine intelligence, 41(8):1947-1962, 2018. 2", + "[46] Yufan Zhou, Chunyuan Li, Changyou Chen, Jianfeng Gao, and Jinhui Xu. Lafite2: Few-shot text-to-image generation. arXiv preprint arXiv:2210.14124, 2022. 2", + "[47] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17907-17917, 2022. 2, 6", + "[48] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dmgan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5810, 2019. 2" + ], + "bbox": [ + 78, + 90, + 470, + 330 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23262", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/c18683e6-ab91-4cf4-bf6b-203837440b00_model.json b/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/c18683e6-ab91-4cf4-bf6b-203837440b00_model.json new file mode 100644 index 0000000000000000000000000000000000000000..27942bbdfedf7c93c850666c6933e29127bf27ee --- /dev/null +++ b/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/c18683e6-ab91-4cf4-bf6b-203837440b00_model.json @@ -0,0 +1,2499 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.784, + 0.177 + ], + "angle": 0, + "content": "Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.205, + 0.27, + 0.238 + ], + "angle": 0, + "content": "Jaewoong Lee* KAIST" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.24, + 0.282, + 0.257 + ], + "angle": 0, + "content": "Republic of Korea" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.26, + 0.302, + 0.274 + ], + "angle": 0, + "content": "hello3196@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.205, + 0.485, + 0.239 + ], + "angle": 0, + "content": "Sangwon Jang* Yonsei University" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.24, + 0.489, + 0.257 + ], + "angle": 0, + "content": "Republic of Korea" + }, + { + "type": "text", + "bbox": [ + 0.343, + 0.26, + 0.487, + 0.274 + ], + "angle": 0, + "content": "agwmon@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.578, + 0.205, + 0.807, + 0.238 + ], + "angle": 0, + "content": "Jaehyeong Jo, Jaehong Yoon KAIST" + }, + { + "type": "text", + "bbox": [ + 0.619, + 0.24, + 0.766, + 0.257 + ], + "angle": 0, + "content": "Republic of Korea" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.26, + 0.857, + 0.275 + ], + "angle": 0, + "content": "{harryjo97, jaehong.yoon}@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.282, + 0.528, + 0.316 + ], + "angle": 0, + "content": "Yunj Kim, Jin-Hwa Kim, Jung-Woo Ha Naver AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.318, + 0.442, + 0.335 + ], + "angle": 0, + "content": "Republic of Korea" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.339, + 0.586, + 0.353 + ], + "angle": 0, + "content": "{yunji.kim, jinhwa.kim, jungwoo.ha}@navercorp.com" + }, + { + "type": "text", + "bbox": [ + 0.661, + 0.283, + 0.789, + 0.316 + ], + "angle": 0, + "content": "Sung Ju Hwang KAIST" + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.318, + 0.799, + 0.336 + ], + "angle": 0, + "content": "Republic of Korea" + }, + { + "type": "text", + "bbox": [ + 0.632, + 0.339, + 0.819, + 0.352 + ], + "angle": 0, + "content": "sjhwang82@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.388, + 0.313, + 0.403 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.474, + 0.736 + ], + "angle": 0, + "content": "Token-based masked generative models are gaining popularity for their fast inference time with parallel decoding. While recent token-based approaches achieve competitive performance to diffusion-based models, their generation performance is still suboptimal as they sample multiple tokens simultaneously without considering the dependence among them. We empirically investigate this problem and propose a learnable sampling model, Text-Conditioned Token Selection (TCTS), to select optimal tokens via localized supervision with text information. TCTS improves not only the image quality but also the semantic alignment of the generated images with the given texts. To further improve the image quality, we introduce a cohesive sampling strategy, Frequency Adaptive Sampling (FAS), to each group of tokens divided according to the self-attention maps. We validate the efficacy of TCTS combined with FAS with various generative tasks, demonstrating that it significantly outperforms the baselines in image-text alignment and image quality. Our text-conditioned sampling framework further reduces the original inference time by more than \\(50\\%\\) without modifying the original generative model." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.749, + 0.208, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.774, + 0.471, + 0.88 + ], + "angle": 0, + "content": "In the flood of generative AI systems for vision domains, text-conditional image generation [26, 33, 42] is coming to the fore in recent years. Although many recent works have achieved success in synthesizing high-quality images [19, 34] with plausible class-alignment in class-conditional cases [4, 8], text-to-image (T2I) generation is more challenging since generating visual outputs" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.389, + 0.895, + 0.631 + ], + "angle": 0, + "content": "that are semantically aligned with input texts is a nontrivial problem. We can roughly categorize the works on text-to-image generation into transformer-based autoregressive (AR) [12, 30, 33] and diffusion-based [18, 29] approaches. Along with the advancement of language models, AR models using transformers have shown impressive performance in text-to-image generation. Despite their success, they suffer from the problem of unidirectional bias, which is undesirable for image generation, and crucially, the sampling process requires over 10 times as much time compared to existing models. Another line of work is diffusion-based methods that aim to generate images by iteratively denoising noisy samples. In particular, several continuous diffusion models [10, 32] have achieved outstanding performance and reduced computational cost. Yet, they require excessive sampling steps to obtain high-quality images during the inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.896, + 0.879 + ], + "angle": 0, + "content": "Recently, a new family of generative models, called token-based diffusion model [6, 23, 41], has emerged as an alternative to tackle the problem of text-to-image generation. Token-based diffusion models quantize the latent features into tokens and apply categorical corruption process in discrete state spaces [1], while conventional diffusion models use Gaussian noise in continuous space. Among various discrete diffusion methods, mask-based diffusion, similar to the absorbing state diffusion used in [1], is mostly used. Compared to existing AR models, this token-based approach is advantageous for speeding up the generation process via simultaneously sampling multiple tokens. Despite the limitation on the reconstruction capacity, these token-based diffusion models significantly outperform the competitors in terms of FID scores, even with fewer sampling steps compared to the continuous diffusion models [5, 37]." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, sampling multiple tokens at once often leads" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.206, + 0.9 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "23252" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.311, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.103, + 0.391, + 0.139 + ], + "angle": 0, + "content": "The train is rounding the bend of a track on the mountain side." + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.167, + 0.392, + 0.193 + ], + "angle": 0, + "content": "A man sitting on top of a bench near the ocean." + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.225, + 0.394, + 0.251 + ], + "angle": 0, + "content": "A box containing an interesting variety of donuts." + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.089, + 0.628, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.103, + 0.707, + 0.128 + ], + "angle": 0, + "content": "A man standing in a field with some cattle." + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.165, + 0.709, + 0.2 + ], + "angle": 0, + "content": "A single seagull standing on the coast with waves in the background." + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.224, + 0.713, + 0.25 + ], + "angle": 0, + "content": "Autumn leaves radiate a golden glow in afternoon sunlight." + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.094, + 0.892, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.282, + 0.894, + 0.365 + ], + "angle": 0, + "content": "Figure 1: Generated samples on MS-COCO dataset and evaluation graph of various sampling methods showing their trade-off. Uniform sampling is a fixed strategy with notably poor text alignment compared to other methods (FID-40K: 15.61, MID-L: 21.23). Random revoke sampling is a revocable strategy with improved text alignment (FID-40K: 16.81, MID-L: 26.98). Ours is TCTS combined with FAS, where both the image quality and the text alignment are significantly better compared to those of baselines (FID-40K: 13.6, MID-L: 29.5). Metrics are measured on all 40K images with their corresponding single caption. The classifier-free guidance scale was fixed at 5 for all sampling methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.471, + 0.603 + ], + "angle": 0, + "content": "to inconsistency throughout a generated image [37]. For each location, the generator outputs a probability distribution that are coherent with each other. However, there is no guarantee that every single token sampled from the distributions will perfectly align with one another. In other words, it is still possible for the generator to sample incompatible tokens regardless of the generator's capability, leading to potentially nonsensical outputs [24, 37]. This results in a trade-off between the sampling steps and generation quality. Reducing the sampling steps leads to faster generation but results in performance degradation due to a large number of simultaneously sampled tokens. Especially, this problem further stands out in the text-to-image generation tasks since the distribution of text-aligned images is more restricted than the distribution of unconditioned or class-conditioned images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.9 + ], + "angle": 0, + "content": "To address these limitations, we propose a novel sampling approach that refines the images during the diffusion process based on the text condition, which we refer to as the Text Conditioned Token Selection (TCTS). TCTS is a sampling strategy that can mask out and resample previously sampled tokens, which we refer to as revocable. To find the tokens that do not align with the given text condition, we propose to train a model that selects tokens to be masked out and re-generated, to be well-aligned with the given text. Combining this approach with the revocable sampling scheme, TCTS can generate high-quality images with improved text-alignment in even fewer sampling steps compared to the naive generative model, as shown in Figure 1. We further introduce Frequency Adaptive Sampling (FAS) to solve the over-simplification problem that occurs when applying revocable methods for relatively longer steps. FAS leverages the self-attention map of the images and applies a mixed sampling method, which we call persistent sampling to prevent the issue. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.377, + 0.892, + 0.437 + ], + "angle": 0, + "content": "- We experimentally show that the revocable sampling strategies are crucial for the trade-off between the text-alignment and the image quality and provide in-depth analysis compared to previous fixed sampling methods." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.448, + 0.892, + 0.523 + ], + "angle": 0, + "content": "- We propose a novel revocable sampling method based on a learnable token selection model that significantly improves the text alignment as well as the image quality even with fewer sampling steps, without the need of retraining the generator model." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.534, + 0.892, + 0.579 + ], + "angle": 0, + "content": "- Moreover, we propose a novel sampling method based on the self-attention map that can be combined with TCTS to solve the over-simplification problem." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.377, + 0.892, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.594, + 0.637, + 0.608 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.619, + 0.894, + 0.754 + ], + "angle": 0, + "content": "Various works have tackled text-to-image generation tasks, and the majority of them are based on Generative Adversarial Networks (GANs) [38, 40, 44, 45, 46, 47, 48]. However, computing directly from the pixel space of an image is challenging, and generating high-resolution images requires substantial computation due to the large number of pixels involved. Therefore, a two-stage approach is often used: models first tokenize the images into a sequence of codes, and then predict the tokenized codes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Autoregressive models Following the success of Transformer [39] and GPT [28], Autoregressive (AR) models [9, 30, 31, 42] proposed to tokenize images and pose the text-to-image generation problem as a problem of translating textual descriptions into image tokens, much like machine translation of one language into a different language. With the advancement of image tokenization and language models, this approach has achieved very high performance. However, even with the help of this tokenizing stage, there remains" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23253" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "a large number of tokens to predict, resulting in a significant delay in the generation time due to their autoregressive nature. Additionally, errors are accumulated through the irrevocable process, with undesirable unidirectional bias." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.16, + 0.473, + 0.327 + ], + "angle": 0, + "content": "Latent diffusion models Recently, diffusion models [18, 36] introduced a new paradigm for text-to-image generation [33]. They generate high-quality images by progressively denoising noise, but require iterative steps and a large amount of computation. Two-stage approaches [10, 32] propose to perform diffusion in latent space successfully reducing computation while maintaining performance. However, most of them still require about 50 inference steps to show plausible performance, and since denoising is applied globally to the entire image during the inference process, locally correcting the image requires complicated processes [32]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.473, + 0.636 + ], + "angle": 0, + "content": "Token-based diffusion models Token-based diffusion models [5, 11, 13, 23, 43] have impressive performance in generating complex scenes from the text. Unlike conventional diffusion models, where the Gaussian noise is applied in a continuous space, token-based diffusion models quantize latent features into tokens and apply a categorical corruption process in a discrete state space [1]. Among various discrete diffusion methods, mask-based diffusion is most commonly used due to its attractive properties; they have the advantage of sampling multiple tokens simultaneously, resulting in fewer sampling steps and faster generation time than the latent diffusion model. However, selecting multiple tokens at once leads to inconsistency throughout the image which is often called a joint distribution issue. Moreover, tokens are heavily influenced by the spatially adjacent tokens that have already been modified so that a misaligned token could disperse the errors throughout the image. To tackle this issue, some works [6, 23] modified the sampling strategy while other works [24] introduced new models to find the misaligned tokens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.472, + 0.775 + ], + "angle": 0, + "content": "In particular, Lezama et al. [24] proposed a learnable model that removes the misaligned tokens at each step. While most existing works randomly select the tokens to discard, this work examines all sampled tokens. It outputs the score of each location where the score indicates whether the corresponding token is under the real distribution. After the predicting and sampling process of the generator, the model selects the locations according to the scores of the corresponding tokens." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.373, + 0.801 + ], + "angle": 0, + "content": "3. Text-conditioned token selection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The diffusion process of the masked image generative models starts with a completely masked image, and samples the tokens of \\( x_{t} \\) for the whole image to predict \\( \\hat{x}_0 \\) while re-masking a portion of them to obtain the denoised sample \\( x_{t - 1} \\). Existing masked image generative models use uniform or confidence-based sampling strategy to determine" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.09, + 0.895, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.265, + 0.895, + 0.335 + ], + "angle": 0, + "content": "Figure 2: Reconstructed images using \\(\\hat{\\mathbf{x}}_0^{(\\mathrm{t})}\\) each step during diffusion using the fixed and revocable method. Only the revocable method is able to edit the tokens to generate a plane according to the text \"A view of the end of an airplane in the sky over mountains.\"" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.895, + 0.426 + ], + "angle": 0, + "content": "the locations to re-mask: Uniform sampling [13, 23] randomly selects the locations of the tokens to keep among the predicted tokens at all locations, while confidence-based strategy, such as purity sampling [37], selects the locations based on the confidence score which is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.452, + 0.894, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C S} (i, t) = \\max _ {j = 1, \\dots , K} p \\left(x _ {0} ^ {i} = j \\mid x _ {t} ^ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.579 + ], + "angle": 0, + "content": "where \\( t \\) is the step number, \\( K \\) is the size of the codebook, and \\( i \\) is the location of the token. Intuitively, a high confidence score means the generator is relatively certain that a token should be sampled at the corresponding location. By keeping these confident tokens while discarding the ambiguous ones, purity sampling shows good performance in practice [5]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.894, + 0.686 + ], + "angle": 0, + "content": "However, these strategies are non-revocable, i.e., fixed, which means tokens that are once sampled cannot be revised afterward, even if they do not match the given text or do not align with other tokens. Once the misaligned tokens are fixed during the process, errors accumulate and spread throughout the image. We observe that fixed strategies negatively affect the text alignment as visualized in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.799, + 0.713 + ], + "angle": 0, + "content": "3.1. Random revoke sampling strategy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In order to address this issue, we first introduce a new sampling strategy called random revoke sampling and use it as a baseline for our final method. It is similar to uniform sampling in a way that they both determine the locations by sampling from a uniform distribution. However, while uniform sampling is a fixed strategy, random revoke sampling selects the locations from the whole image space regardless of the previously fixed locations, making the process reversible. It randomly selects a certain number of tokens to preserve for the next step and revokes the previously fixed ones. The revocable feature can better align the images to the given text information, especially in complex scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23254" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.468, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.348 + ], + "angle": 0, + "content": "Figure 3: Reconstructed images using \\(\\hat{\\mathbf{x}}_0^{(\\mathrm{t})}\\) each step during diffusion using the revocable method with and without FAS. The backgrounds of the images in the top row are over-simplified while our proposed FAS prevents this, as shown in the bottom row. The text is \"This small bird is greyish in color with flecks of yellow on the back and breast, and a bit of white on the belly.\"" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.473, + 0.514 + ], + "angle": 0, + "content": "To be specific, as shown in the top row of Figure 2, fixed strategies have limitations in recovering from errors or generating diverse images due to their inability to regenerate incorrect tokens or modify missing essential objects or attributes. In contrast, revocable strategies enhance the text alignment by providing an opportunity to correct invalidly generated tokens, as shown in the bottom row of Figure 2. Since revocable strategies have no constraints on which tokens to remove or fix, they have the advantage of mitigating error accumulation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.472, + 0.595 + ], + "angle": 0, + "content": "However, random revoke sampling suffers from two main issues that must be addressed. The first issue is the instability resulting from randomly discarding previously sampled tokens which worked as a given condition of the step. It can lower the quality of the outputs as in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.472, + 0.839 + ], + "angle": 0, + "content": "The second issue, namely the over-simplification, arises when it is used in the generation process involving longer steps. The low-frequency parts of the output images, such as backgrounds, tend to be over-simplified compared to fixed methods, which generate more realistic backgrounds. The over-simplified samples can be seen in Figure 3, and it has a significant impact on the FID metric as in Table 1 and 2. We observe that this is due to excessive repetition of resampling over a large number of steps, while a fixed schedule does not resample any tokens. More opportunities for resampling make the tokens converge towards safer and simpler patterns. As shown in the top row of Figure 3, the high-frequency areas with rich visual details do not become over-simplified even after multiple resampling processes, but the low-frequency areas become less diverse and more simplified in patterns quickly due to excessive resampling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.902 + ], + "angle": 0, + "content": "To confirm this trade-off between text alignment and image quality, we define a persistent sampling as an interpolation between uniform sampling and random revoke sampling. In details, a persistent weight boosts the probability of keep" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.093, + 0.724, + 0.108 + ], + "angle": 0, + "content": "Algorithm 1 Persistent Sampling" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.11, + 0.892, + 0.137 + ], + "angle": 0, + "content": "Input: \\( k_{t} \\): number of tokens sampled at step \\( t \\), \\( w \\): persistent weight, \\( N \\): number of all tokens" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.14, + 0.644, + 0.154 + ], + "angle": 0, + "content": "1: \\(x_{T}\\gets [[\\mathrm{MASK}]]_{N}\\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.155, + 0.684, + 0.167 + ], + "angle": 0, + "content": "2: for \\( t = T, T - 1, \\dots, 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.168, + 0.639, + 0.181 + ], + "angle": 0, + "content": "3: \\(\\hat{x}_0 = G_\\theta(x_t, c)\\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.182, + 0.687, + 0.196 + ], + "angle": 0, + "content": "4: \\(A_{t} = \\{i|x_{t}^{i} = [\\mathrm{MASK}]\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.196, + 0.709, + 0.208 + ], + "angle": 0, + "content": "5: \\( m = k_{T} + k_{T - 1} + \\ldots + k_{t} \\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.209, + 0.822, + 0.223 + ], + "angle": 0, + "content": "6: \\(\\mathcal{U}_t(I = i|i\\in A_t):\\mathcal{U}_t(I = i|i\\in A_t^C) = 1:w\\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.224, + 0.798, + 0.237 + ], + "angle": 0, + "content": "7: \\(\\rightarrow\\) uniform distribution + persistent weight" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.238, + 0.856, + 0.251 + ], + "angle": 0, + "content": "8: \\(i_1, i_2, \\dots, i_m \\sim \\mathcal{U}(I) \\gets\\) sample without replacement" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.252, + 0.653, + 0.265 + ], + "angle": 0, + "content": "9: \\(x_{t}\\gets [[\\mathrm{MASK}]]_{N}\\)" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.265, + 0.685, + 0.278 + ], + "angle": 0, + "content": "10: for \\( i = i_1, i_2, \\ldots, i_m \\) do" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.279, + 0.612, + 0.292 + ], + "angle": 0, + "content": "11: \\(x_{t}^{i}\\gets \\hat{x}_{0}^{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.293, + 0.59, + 0.305 + ], + "angle": 0, + "content": "12: end for" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.307, + 0.575, + 0.319 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.32, + 0.702, + 0.334 + ], + "angle": 0, + "content": "14: Return: Generated image \\( x_0 \\)" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.14, + 0.856, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.895, + 0.555 + ], + "angle": 0, + "content": "ing the tokens that were sampled in the previous step. While random revoke sampling assigns the same probability to all the locations, persistent sampling multiplies the persistent weight to the probability of previously kept locations. Further details are shown in Algorithm 1. If the persistent weight \\( w \\) is 1, all the tokens would have the same probability to be fixed, such as random revoke sampling. If \\( w \\) is sufficiently large, the previously sampled tokens would be kept until the end, such as uniform sampling. In the graph in Figure 1, a trade-off can be seen that as the sampling strategy gets closer to the random revoke sampling, text alignment gets better, while the image quality gets worse. We further analyze the effect of these sampling methods in Section 4.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.567, + 0.783, + 0.582 + ], + "angle": 0, + "content": "3.2. Text-conditioned token selection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.895, + 0.819 + ], + "angle": 0, + "content": "We propose a learnable model called Text-Conditioned Token Selection (TCTS). TCTS is trained to output a score map to detect tokens that are not under the real distribution for the given text condition, and selects the tokens that are not aligned with the text or other tokens in the image. By masking the selected tokens at each step, the diffusion network receives high-quality text-aligned tokens as the input to sample the next set of tokens which alleviates the error accumulation and joint distribution issue of the previous sampling strategies. Moreover, since TCTS is trained to discriminate the well-aligned tokens, it detours the instability problem as mentioned earlier, unlike other revocable methods. From the experimental results in Section 4, we observe that TCTS enhances the text alignment without compromising the image quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Frequency adaptive sampling (FAS) The simplest way to solve the over-simplification problem in longer steps is to use the persistent sampling to alleviate the excessive resampling of tokens. However, as previously mentioned, this method" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23255" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.088, + 0.891, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.28, + 0.897, + 0.351 + ], + "angle": 0, + "content": "Figure 4: Overall generation framework of proposed TCTS and FAS. After the generator predicts the tokens, TCTS exploits the text condition to detect misaligned tokens and outputs the score map. Meanwhile, FAS splits the tokens according to the frequency using the self-attention map from the generator, performing revocable sampling to high-frequency split and persistent sampling to low-frequency split. The adaptive sampling predicts \\(\\hat{x}_0\\) and decide a few of the locations to mask according to \\(x_{t}\\). The token maps produced by FAS is combined with the score map to predict \\(x_{t - 1}\\). After iterative process, our model removes all the masks and generates \\(x_0\\)." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.374, + 0.472, + 0.527 + ], + "angle": 0, + "content": "
StepMethodMID-L ↑SOA-I ↑CLIP-S ↑FID-30K ↓
16Purity11.0272.380.247419.20
Uniform17.9474.800.250016.17
RR23.6078.790.252617.10
TCTS + FAS26.7279.520.255914.45
25Purity16.8475.210.248718.39
Uniform22.2777.080.252415.91
RR26.7781.100.254318.43
TCTS + FAS27.7980.870.256315.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.531, + 0.472, + 0.587 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation of sampling methods on MS-COCO dataset. The ground truth MID on MS-COCO image-caption pairs is 54.73. The classifier-free guidance scale was fixed at 5 for all sampling methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.685 + ], + "angle": 0, + "content": "may result in a slight compromise in text alignment and limit the opportunity to correct the image through resampling. To further improve this trade-off, we propose a new method called Frequency Adaptive Sampling (FAS), which can be applied to TCTS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.901 + ], + "angle": 0, + "content": "FAS is a method that utilizes the generator's self-attention map to limit resampling only in the low-frequency areas of the image. In [21], it is known that the generator's self-attention layer contains frequency information of the image, which is also observed in token-based diffusion models. We utilized this to distinguish the frequency areas of the image without additional operations, and applied persistent sampling only to the low-frequency areas using persistent weight. As shown in the bottom row of Figure 3, revocable sampling along with FAS method allows repetitive resampling in areas that require rich visual details, while limiting the number of resampling in relatively simple areas preventing oversimplification. Detailed algorithm is in the supplementary A. The overall framework of our model, including both TCTS" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.374, + 0.892, + 0.526 + ], + "angle": 0, + "content": "
StepMethodMID-L ↑CLIP-S ↑FID ↓
16Purity-24.210.241015.21
Uniform-25.600.240416.57
RR-25.030.237117.38
TCTS + FAS-19.880.24612.35
25Purity-21.260.238412.60
Uniform-23.040.239613.02
RR-23.290.236414.53
TCTS + FAS-18.310.240913.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.53, + 0.895, + 0.587 + ], + "angle": 0, + "content": "Table 2: Quantitative evaluation of sampling methods on CUB dataset. The ground truth MID on CUB image-caption pairs is 15.85. We omit SOA in this table as CUB dataset consists of only one bird per photo, diminishing the metric's significance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.605, + 0.73, + 0.62 + ], + "angle": 0, + "content": "and FAS, is illustrated in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.634, + 0.651 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To validate the efficacy of our token selection framework, we modify the transformer from [37] to contain 310M parameters for MS-COCO dataset and 83M parameters for CUB dataset. We extract text features from CLIP ViT-B/32 [27] for a text-conditioned generation. In the training process of TCTS, we freeze the generator's parameters and independently train our model, making it applicable to other types of token-based diffusion models. We use the binary cross-entropy loss for the objective and validate our method and baselines on MS-COCO [25] and CUB [14] datasets by training them for 200 epochs. We leverage classifier-free guidance [20] by stochastic sampling of the guidance strength to improve the generation quality. We further provide the details of the architecture, hyperparameters and additional analysis on the use of classifier-free guidance in the supplementary A." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23256" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.089, + 0.089, + 0.472, + 0.219 + ], + "angle": 0, + "content": "
ModelMID-L ↑MID-B ↑SOA-I ↑FID-30K ↓
GLIDE [26]1.031.00-32.08
AttnGAN [40]-65.20-8.9039.0129.15
DM-GAN [23]-44.663.5148.0322.90
DF-GAN [38]-58.75-15.21-31.75
VQ-Diffusion [41]-19.635.77-13.13
LAFITE [47]6.2635.1774.788.03
TCTS + FAS26.0238.9879.529.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.225, + 0.472, + 0.31 + ], + "angle": 0, + "content": "Table 3: Comparison of the proposed method with recent text-to-image generative models on MS-COCO dataset. MID-L and MID-B are calculated with ViT-L/14 and ViT-B/32 each. The ground truth MID-L and MID-B on MS-COCO image-caption pairs is 54.73, 41.57 each. Ours are evaluated with 16-step setting and we adopt classifier-free guidance from 3 to 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.474, + 0.534 + ], + "angle": 0, + "content": "Metrics Since FID [16] metric is known to be problematic for its inability to exactly represent the image quality [5] and unable to consider the text alignment, we additionally use Mutual Information Divergence (MID) [22] to evaluate the text-alignment of generated images which enables sample-wise evaluation and outperforms previous metrics in human Likert-scale judgment correlations. In particular, since MID responds more sensitively to images generated with foiled captions [35], it is more appropriate for analyzing text alignment with complex and challenging captions. Moreover, we use the CLIP score [15] and SOA-I [17], which are widely used in text-to-image synthesis. Note that the CLIP score is calculated with ViT-L/14 throughout the experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.295, + 0.56 + ], + "angle": 0, + "content": "4.1. Text-to-image synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.472, + 0.853 + ], + "angle": 0, + "content": "Figure 1 Left and Middle visualizes the generated images using improved VQ diffusion [37] with varying sampling methods. Generated examples from uniform token sampling are poorly aligned with the given texts, and often include erratic partitions of the objects since the predicted tokens selected at the same step cannot appropriately reflect the change of others (e.g., smashed donuts, and imperfect chair and seagull). Random revoke (RR) sampling, which can re-update previously sampled tokens, seems to improve alignment with text by editing out nonsensical regions and iteratively considering the entire scene at each generation step. Yet, RR selects tokens in a completely random manner without any constraints, which results in suboptimal generation quality and alignment with the texts (e.g., disappeared man and train). On the other hand, ours mitigates the problem from RR by selecting text-conditioned tokens at the sampling phase and successfully generates high-quality images which contain a clear semantic connection to the given text captions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.901 + ], + "angle": 0, + "content": "This phenomenon can be quantitatively seen in Table 1. 16-step generation of random revoke sampling shows better text alignment than even 25-step generation of other fixed" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.89, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.192, + 0.89, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.293, + 0.894, + 0.322 + ], + "angle": 0, + "content": "Figure 5: Results of image refinement using TCTS. Top: Original samples, Bottom: Refined images for 8-steps with TCTS." + }, + { + "type": "table", + "bbox": [ + 0.568, + 0.324, + 0.832, + 0.442 + ], + "angle": 0, + "content": "
MethodMID-L ↑FID ↓
Original12.1016.44
RR13.6417.18
TCTS15.6416.92
Original4.5419.13
Random masked6.0619.12
TCTS masked6.4718.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.445, + 0.894, + 0.487 + ], + "angle": 0, + "content": "Table 4: Qualitative evaluation of the refined images. Top: Refinement with additional revision steps. Bottom: Refinement with masking lowest-scoring tokens." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.895, + 0.79 + ], + "angle": 0, + "content": "sampling methods. We can also observe the trade-off between CLIP score and FID by comparing each method to their 25-step versions. Our model outperforms other baselines in most of the metrics especially when comparing MID-L to fixed sampling methods such as Purity and Uniform. Since our model can make corrections during the generation process, it requires fewer steps to match the performance of our baseline sampling methods, see Section 4.5 for further analysis. We also present the results of our experiment on the CUB dataset with fewer parameters in Table 2, which demonstrates satisfactory performance. However, the fact that the CUB dataset contains only single object per image resulted in a tendency that is slightly different from what we mentioned earlier. In Table 3, we compared the performance of our method with other text-to-image generative models. Our model highly exceeds other models in text alignment metrics, especially in MID-L and closely approximate the MID-B value of the ground truth images. More samples generated by our model is in the supplementary B." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.673, + 0.817 + ], + "angle": 0, + "content": "4.2. Image refinement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The advantage of the masked image generative model is that it enables fast local refinement. We conduct image refinement in two separate methods, inspired by [23] and [24]. First, we use TCTS to apply additional revision steps to images generated with uniform sampling. The overall image" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23257" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "quality is improved by adding additional refinement steps as shown in Figure 5. In order to demonstrate the image refinement performance of our TCTS, we also measure the performance of the image refinement with random revoke sampling. While RR improves does improve sample quality, additional refinement steps increase the FID score, caused by a similar effect to increasing the generation step." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.474, + 0.32 + ], + "angle": 0, + "content": "We additionally perform the experiment by masking \\(60\\%\\) lowest-scoring tokens with TCTS, and generate the images with uniform sampling. We observe that all metrics of the refined images outperform those of the original images. In Table 4, we compare them with samples refined after randomly masking the tokens without TCTS. Due to the page limit, we describe further details of the two refinement methods and visualize more samples in the supplementary B." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.33, + 0.303, + 0.346 + ], + "angle": 0, + "content": "4.3. Mask-free object editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.473, + 0.519 + ], + "angle": 0, + "content": "Since the masked image generative model is capable of local refinement, image editing without manual masking is possible simply by randomly masking a part of the image and resampling it with the new text condition. However, this method requires a low masking ratio and many resampling steps to maintain the overall structure of the image. This can result in significant changes to unnecessary parts or, as mentioned earlier, over-simplification issues. Additionally, even with many steps, editing large objects with small masks can be more challenging due to the significant influence of surrounding tokens rather than the new text condition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.474, + 0.732 + ], + "angle": 0, + "content": "Motivated by the operation of self-attention maps in frequency adaptive sampling, we leverage a cross-attention map corresponding to the word of the object that is to be changed, giving weights to resample tokens so that the corresponding locations can be resampled. Then, it enables efficient image editing with fewer steps and makes it easier to edit larger objects. Although it is similar to DiffEdit [7], ours operates in a more straightforward manner without additional computations, thanks to the capability of local refinement of the masked generative model. In Figure 6, we edit the object by adding \\(25\\%\\) noise in only 10-steps. With our method, we can edit more quickly with fewer editing steps, which further does not suffer from the over-simplification problem. More samples are in the supplementary B." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.36, + 0.759 + ], + "angle": 0, + "content": "4.4. High-resolution image synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Continuous diffusion models that use encoders and decoders, such as masked image generative models, can generate larger images that are not in the training set. Similar to Bond-Taylor et al. [3], we divide the tokens into subsets according to the model input size, individually pass them through the model, and then spatially aggregate tokens to synthesize a high-resolution image. In this way, we are able to generate more realistic and high-resolution samples with the same TCTS model without additional training. Addition-" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.09, + 0.892, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.203, + 0.884, + 0.216 + ], + "angle": 0, + "content": "A dog teddy bear looks back from the passenger seat of a vehicle." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.218, + 0.892, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.33, + 0.86, + 0.342 + ], + "angle": 0, + "content": "A wooden table topped with earrot cake donuts, broccoli." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.352, + 0.895, + 0.395 + ], + "angle": 0, + "content": "Figure 6: Examples of mask-free editing samples with cross-attention map. The cross-attention map is multiplied to the score map of TCTS to perform weighted sampling." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.409, + 0.695, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.559, + 0.688, + 0.585 + ], + "angle": 0, + "content": "A short train traveling through a mountainous landscape." + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.409, + 0.891, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.559, + 0.869, + 0.584 + ], + "angle": 0, + "content": "A family on a boat sailing down a river." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.594, + 0.895, + 0.638 + ], + "angle": 0, + "content": "Figure 7: Examples of high resolution \\((512^{2})\\) samples. Left: Image generated with all mask tokens, Right: Image generated with TCTS as a super-resolution unit." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.895, + 0.791 + ], + "angle": 0, + "content": "ally, we propose a new method to generate high-resolution images only with low-resolution TCTS. To be specific, we first generate a small-size image with TCTS and upsample the token map in bicubic mode to the desired size. Then we divide the high-resolution token map into overlapping small-size sections and refine all the sections several times. Then, we can generate high-resolution images with the low-resolution TCTS where we visualize the generated samples in Figure 7." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.664, + 0.818 + ], + "angle": 0, + "content": "4.5. Further analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Importance of early-stage sampling strategy As mentioned in Section 3.1, by controlling the persistent weight, we can interpolate the two baseline sampling methods: uniform sampling and random revoke sampling. We designed various experiments to find out the impact of the sampling" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23258" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.177, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.169, + 0.152, + 0.179 + ], + "angle": 0, + "content": "(a) Real" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.09, + 0.371, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.091, + 0.468, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.169, + 0.434, + 0.178 + ], + "angle": 0, + "content": "road" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.183, + 0.177, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.183, + 0.273, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.183, + 0.37, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.183, + 0.468, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.261, + 0.172, + 0.272 + ], + "angle": 0, + "content": "(b) Synthesized" + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.261, + 0.232, + 0.27 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.261, + 0.335, + 0.27 + ], + "angle": 0, + "content": "boat" + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.262, + 0.437, + 0.271 + ], + "angle": 0, + "content": "water" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.278, + 0.469, + 0.334 + ], + "angle": 0, + "content": "Figure 8: Visualization of the averaged cross-attention maps for each word from TCTS. Top: Image from COCO dataset: \"A large black bear walking down a dirty road\". Bottom: Synthesized image: \"A small boat in a narrow body of water\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.47, + 0.428 + ], + "angle": 0, + "content": "method with respect to the time steps. We switched the sampling methods during the time steps, from uniform sampling to random revoke sampling (U2R) and vice versa (R2U). In the rightmost plot in Figure 1, we can observe that U2R shows better text-alignment performance than R2U." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.685 + ], + "angle": 0, + "content": "These results suggest that the early stage sampling, in which the masked ratio is high, substantially influences the text alignment of the final image, which is similarly observed in continuous diffusion models [2]. Generating the whole image and then processing it through a few revision steps like [23] might help the image quality but can not conspicuously enhance the essential text alignment. Therefore, in order to obtain an image that aligns well with the text, sampling must be done carefully in the early stage when not many tokens are generated. This analysis is also shown in the result that the revocable strategy exhibits better performance by aggressively resampling tokens several times. In particular, since weak generators with poor generation performance suffer more from the joint distribution issues, giving a sufficient number of recovery chances to discard tokens and draw new ones at the early stage is the key to the desired text alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Text-conditioned sampling The experiments show that TCTS plays a crucial role in enhancing text alignment in image generation. Although we did not utilize any text-specific loss during training, the model still effectively processes the text condition and achieves optimal performance. To verify the accuracy of TCTS's attention towards text-related parts of the image, we visualized the average cross-attention layer for each word in Figure 8. The results demonstrate TCTS's ability to attend to each word in both real and synthesized images. By helping the selection of relevant tokens during the generating stage, TCTS can boost model performance without the requirement of additional complex text-related loss." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.089, + 0.892, + 0.21 + ], + "angle": 0, + "content": "
MethodMID-L ↑SOA-I ↑CLIP-S ↑FID-30K ↓
RR26.7781.100.254318.43
RR + FAS26.3280.780.253916.01
RR + pw24.2379.140.253415.96
TCTS27.8280.930.256516.69
TCTS + FAS27.7980.870.256315.39
TCTS + pw25.1279.510.255615.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.213, + 0.894, + 0.271 + ], + "angle": 0, + "content": "Table 5: Ablation study of using FAS and persistent weight in revocable methods on MS-COCO dataset. All models were evaluated on 25-step setting. The ground truth MID-L on MS-COCO image-caption pairs is 54.73." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.278, + 0.892, + 0.354 + ], + "angle": 0, + "content": "
StepModelMID-L ↑FID-30K ↓Time
50Uniform24.0416.84×1
RR26.5521.12
16TCTS + FAS26.7214.45×0.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.357, + 0.893, + 0.428 + ], + "angle": 0, + "content": "Table 6: Inference time relative to MID/FID of our baselines and models on MS-COCO dataset. We expressed the inference time of our models as a multiple of other baselines. Our model outspeeds other baselines while performing better in both MID-L and FID-30K. Further analysis is in the supplementary C." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.895, + 0.736 + ], + "angle": 0, + "content": "Ablation study of FAS As mentioned in Section 3.2, FAS decides whether to apply persistent weight or not to each location. If the whole image is considered low-frequency, FAS would apply the weight to every location, which is the same as the persistent sampling. On the other hand, if the whole image is considered high-frequency, combining FAS would not apply the weight anywhere, which does not change the sampling strategy. To further analyze the effect of FAS, we evaluated it on two revocable methods: random revoke sampling and TCTS. Attaching FAS to a base sampling method can be regarded as a mixture of two methods: the base revocable sampling method and persistent sampling method. Table 5 clearly shows this relation of FAS and persistent weight. While FAS-attached methods do not show the best performance, the performance of them is always in close proximity to the better one of the two comparing methods of each. FAS-attached methods balances well between text-alignment and image quality without any additional training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Inference time Our TCTS and FAS compute additional operations at each generation step, resulting in marginally increased inference time compared to other token selection baselines. However, regarding the text alignment and the image quality, the relative inference time is considerably decreased. In Table 6, our model outperforms 50-step image generation of baseline sampling methods within only 16 steps in MID-L and FID-30K, which means that our model can synthesize better samples with only \\(\\times 0.45\\) time." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23259" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.196, + 0.106 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.473, + 0.358 + ], + "angle": 0, + "content": "This paper examines which factors in the masked diffusion process impact the output images and cause the trade-off between image quality and text alignment. We empirically find that the revocable sampling significantly improves the text alignment yet degrades the quality of the generated images. To tackle the problem, we propose a simple token sampling strategy, coined text-conditioned token selection, that combines learnable and revocable methods, pushing the boundary of the trade-off between image quality and text alignment without the necessity to update the large-scale pre-trained generator. We find that collaborative sampling in a persistent and revocable manner surprisingly alleviates over-simplification issues in the generated backgrounds. Our proposed method can be utilized as an image refinement tool for various generative tasks, which is remarkably fast to generate high-quality images within much fewer steps." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.373, + 0.266, + 0.391 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.473, + 0.52 + ], + "angle": 0, + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST)) and (No. RS-2022-00187238, Development of Large Korean Language Model Technology for Efficient Pre-training). This work was supported by KAIST-NAVER Hypercreative AI Center." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.535, + 0.175, + 0.551 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.561, + 0.472, + 0.63 + ], + "angle": 0, + "content": "[1] Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. In Advances in Neural Information Processing Systems, pages 17981-17993. Curran Associates, Inc., 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.633, + 0.473, + 0.701 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.704, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[3] Sam Bond-Taylor, Peter Hessey, Hiroshi Sasaki, Toby P Breckon, and Chris G Willcocks. Unleashing transformers: parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vector-quantized codes. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXIII, pages 170–188. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.803, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[4] Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdzal, and Adriana Romero Soriano. Instance-conditioned gan. Advances in Neural Information Processing Systems, 34:27517-27529, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.86, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[5] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to" + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.561, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.123, + 0.895, + 0.19 + ], + "angle": 0, + "content": "[6] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[7] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.236, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.279, + 0.895, + 0.334 + ], + "angle": 0, + "content": "[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.336, + 0.892, + 0.391 + ], + "angle": 0, + "content": "[10] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. arXiv preprint arXiv:2208.13753, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.393, + 0.895, + 0.476 + ], + "angle": 0, + "content": "[11] Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what happened: Unifying text-guided video completion via multimodal masked video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10681-10692, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.478, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[12] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.895, + 0.603 + ], + "angle": 0, + "content": "[13] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[14] Xiangteng He and Yuxin Peng. Fine-grained visual-textual representation learning. IEEE Transactions on Circuits and Systems for Video Technology, 30(2):520-531, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.895, + 0.73 + ], + "angle": 0, + "content": "[15] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528. Association for Computational Linguistics, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.733, + 0.895, + 0.8 + ], + "angle": 0, + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[17] Tobias Hinz, Stefan Heinrich, and Stefan Wermter. Semantic object accuracy for generative text-to-image synthesis. IEEE transactions on pattern analysis and machine intelligence, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23260" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:47-1, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.471, + 0.22 + ], + "angle": 0, + "content": "[21] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.222, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[22] Jin-Hwa Kim, Yunji Kim, Jiyoung Lee, Kang Min Yoo, and Sang-Woo Lee. Mutual information divergence: A unified metric for multimodal generative models. arXiv preprint arXiv:2205.13445, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.279, + 0.471, + 0.333 + ], + "angle": 0, + "content": "[23] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Draft-and-revise: Effective image generation with contextual rq-transformer. arXiv preprint arXiv:2206.04452, 2022. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.335, + 0.471, + 0.39 + ], + "angle": 0, + "content": "[24] José Lezama, Huiwen Chang, Lu Jiang, and Irfan Essa. Improved masked image generation with token-critic. In European Conference on Computer Vision, pages 70-86. Springer, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.392, + 0.471, + 0.461 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.463, + 0.471, + 0.532 + ], + "angle": 0, + "content": "[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.534, + 0.471, + 0.616 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.618, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[28] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.471, + 0.715 + ], + "angle": 0, + "content": "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.718, + 0.471, + 0.785 + ], + "angle": 0, + "content": "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.788, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[31] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.174 + ], + "angle": 0, + "content": "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.176, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[34] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.894, + 0.273 + ], + "angle": 0, + "content": "[35] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. Foil it! find one mismatch between image and language caption. arXiv preprint arXiv:1705.01359, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.894, + 0.328 + ], + "angle": 0, + "content": "[36] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.894, + 0.371 + ], + "angle": 0, + "content": "[37] Zhicong Tang, Shuyang Gu, Jianmin Bao, Chen Dong, and Fang Wen. Improved vector quantized diffusion models. arXiv preprint arXiv:2205.16007, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.894, + 0.427 + ], + "angle": 0, + "content": "[38] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.894, + 0.482 + ], + "angle": 0, + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.894, + 0.565 + ], + "angle": 0, + "content": "[40] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.894, + 0.634 + ], + "angle": 0, + "content": "[41] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.894, + 0.704 + ], + "angle": 0, + "content": "[42] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[43] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[44] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[45] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stack" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23261" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "gan++: Realistic image synthesis with stacked generative adversarial networks. IEEE transactions on pattern analysis and machine intelligence, 41(8):1947-1962, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.177 + ], + "angle": 0, + "content": "[46] Yufan Zhou, Chunyuan Li, Changyou Chen, Jianfeng Gao, and Jinhui Xu. Lafite2: Few-shot text-to-image generation. arXiv preprint arXiv:2210.14124, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.472, + 0.26 + ], + "angle": 0, + "content": "[47] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17907-17917, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.472, + 0.331 + ], + "angle": 0, + "content": "[48] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dmgan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5810, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23262" + } + ] +] \ No newline at end of file diff --git a/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/c18683e6-ab91-4cf4-bf6b-203837440b00_origin.pdf b/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/c18683e6-ab91-4cf4-bf6b-203837440b00_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..90109e390515a4f05e2475622f1adb61bb1854ce --- /dev/null +++ b/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/c18683e6-ab91-4cf4-bf6b-203837440b00_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1370f13e7a4e34e59eb0c975211c92c46ce4f6dafd735cf23b2a77fd90d3933 +size 2039161 diff --git a/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/full.md b/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..071f21991de58b86a34a50f5b9fe53936b605c91 --- /dev/null +++ b/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/full.md @@ -0,0 +1,325 @@ +# Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models + +Jaewoong Lee* KAIST + +Republic of Korea + +hello3196@kaist.ac.kr + +Sangwon Jang* Yonsei University + +Republic of Korea + +agwmon@gmail.com + +Jaehyeong Jo, Jaehong Yoon KAIST + +Republic of Korea + +{harryjo97, jaehong.yoon}@kaist.ac.kr + +Yunj Kim, Jin-Hwa Kim, Jung-Woo Ha Naver AI Lab + +Republic of Korea + +{yunji.kim, jinhwa.kim, jungwoo.ha}@navercorp.com + +Sung Ju Hwang KAIST + +Republic of Korea + +sjhwang82@kaist.ac.kr + +# Abstract + +Token-based masked generative models are gaining popularity for their fast inference time with parallel decoding. While recent token-based approaches achieve competitive performance to diffusion-based models, their generation performance is still suboptimal as they sample multiple tokens simultaneously without considering the dependence among them. We empirically investigate this problem and propose a learnable sampling model, Text-Conditioned Token Selection (TCTS), to select optimal tokens via localized supervision with text information. TCTS improves not only the image quality but also the semantic alignment of the generated images with the given texts. To further improve the image quality, we introduce a cohesive sampling strategy, Frequency Adaptive Sampling (FAS), to each group of tokens divided according to the self-attention maps. We validate the efficacy of TCTS combined with FAS with various generative tasks, demonstrating that it significantly outperforms the baselines in image-text alignment and image quality. Our text-conditioned sampling framework further reduces the original inference time by more than $50\%$ without modifying the original generative model. + +# 1. Introduction + +In the flood of generative AI systems for vision domains, text-conditional image generation [26, 33, 42] is coming to the fore in recent years. Although many recent works have achieved success in synthesizing high-quality images [19, 34] with plausible class-alignment in class-conditional cases [4, 8], text-to-image (T2I) generation is more challenging since generating visual outputs + +that are semantically aligned with input texts is a nontrivial problem. We can roughly categorize the works on text-to-image generation into transformer-based autoregressive (AR) [12, 30, 33] and diffusion-based [18, 29] approaches. Along with the advancement of language models, AR models using transformers have shown impressive performance in text-to-image generation. Despite their success, they suffer from the problem of unidirectional bias, which is undesirable for image generation, and crucially, the sampling process requires over 10 times as much time compared to existing models. Another line of work is diffusion-based methods that aim to generate images by iteratively denoising noisy samples. In particular, several continuous diffusion models [10, 32] have achieved outstanding performance and reduced computational cost. Yet, they require excessive sampling steps to obtain high-quality images during the inference. + +Recently, a new family of generative models, called token-based diffusion model [6, 23, 41], has emerged as an alternative to tackle the problem of text-to-image generation. Token-based diffusion models quantize the latent features into tokens and apply categorical corruption process in discrete state spaces [1], while conventional diffusion models use Gaussian noise in continuous space. Among various discrete diffusion methods, mask-based diffusion, similar to the absorbing state diffusion used in [1], is mostly used. Compared to existing AR models, this token-based approach is advantageous for speeding up the generation process via simultaneously sampling multiple tokens. Despite the limitation on the reconstruction capacity, these token-based diffusion models significantly outperform the competitors in terms of FID scores, even with fewer sampling steps compared to the continuous diffusion models [5, 37]. + +However, sampling multiple tokens at once often leads + +![](images/cfae35d70e7ecfe92b1099a9975ede949fd90117189502f79dc084a5d2b6d3e3.jpg) +A man sitting on top of a bench near the ocean. +A box containing an interesting variety of donuts. + +![](images/c8d7819a9fcbcb251285c601da72f376bdb81f8d84ae948374c3df55d7da4e2a.jpg) +The train is rounding the bend of a track on the mountain side. +A man standing in a field with some cattle. +A single seagull standing on the coast with waves in the background. +Autumn leaves radiate a golden glow in afternoon sunlight. + +![](images/e973a1da5d3c0c869038661864c295908823cf4d612a703b514d61019eeeae8b.jpg) +Figure 1: Generated samples on MS-COCO dataset and evaluation graph of various sampling methods showing their trade-off. Uniform sampling is a fixed strategy with notably poor text alignment compared to other methods (FID-40K: 15.61, MID-L: 21.23). Random revoke sampling is a revocable strategy with improved text alignment (FID-40K: 16.81, MID-L: 26.98). Ours is TCTS combined with FAS, where both the image quality and the text alignment are significantly better compared to those of baselines (FID-40K: 13.6, MID-L: 29.5). Metrics are measured on all 40K images with their corresponding single caption. The classifier-free guidance scale was fixed at 5 for all sampling methods. + +to inconsistency throughout a generated image [37]. For each location, the generator outputs a probability distribution that are coherent with each other. However, there is no guarantee that every single token sampled from the distributions will perfectly align with one another. In other words, it is still possible for the generator to sample incompatible tokens regardless of the generator's capability, leading to potentially nonsensical outputs [24, 37]. This results in a trade-off between the sampling steps and generation quality. Reducing the sampling steps leads to faster generation but results in performance degradation due to a large number of simultaneously sampled tokens. Especially, this problem further stands out in the text-to-image generation tasks since the distribution of text-aligned images is more restricted than the distribution of unconditioned or class-conditioned images. + +To address these limitations, we propose a novel sampling approach that refines the images during the diffusion process based on the text condition, which we refer to as the Text Conditioned Token Selection (TCTS). TCTS is a sampling strategy that can mask out and resample previously sampled tokens, which we refer to as revocable. To find the tokens that do not align with the given text condition, we propose to train a model that selects tokens to be masked out and re-generated, to be well-aligned with the given text. Combining this approach with the revocable sampling scheme, TCTS can generate high-quality images with improved text-alignment in even fewer sampling steps compared to the naive generative model, as shown in Figure 1. We further introduce Frequency Adaptive Sampling (FAS) to solve the over-simplification problem that occurs when applying revocable methods for relatively longer steps. FAS leverages the self-attention map of the images and applies a mixed sampling method, which we call persistent sampling to prevent the issue. We summarize our contributions as follows: + +- We experimentally show that the revocable sampling strategies are crucial for the trade-off between the text-alignment and the image quality and provide in-depth analysis compared to previous fixed sampling methods. +- We propose a novel revocable sampling method based on a learnable token selection model that significantly improves the text alignment as well as the image quality even with fewer sampling steps, without the need of retraining the generator model. +- Moreover, we propose a novel sampling method based on the self-attention map that can be combined with TCTS to solve the over-simplification problem. + +# 2. Related work + +Various works have tackled text-to-image generation tasks, and the majority of them are based on Generative Adversarial Networks (GANs) [38, 40, 44, 45, 46, 47, 48]. However, computing directly from the pixel space of an image is challenging, and generating high-resolution images requires substantial computation due to the large number of pixels involved. Therefore, a two-stage approach is often used: models first tokenize the images into a sequence of codes, and then predict the tokenized codes. + +Autoregressive models Following the success of Transformer [39] and GPT [28], Autoregressive (AR) models [9, 30, 31, 42] proposed to tokenize images and pose the text-to-image generation problem as a problem of translating textual descriptions into image tokens, much like machine translation of one language into a different language. With the advancement of image tokenization and language models, this approach has achieved very high performance. However, even with the help of this tokenizing stage, there remains + +a large number of tokens to predict, resulting in a significant delay in the generation time due to their autoregressive nature. Additionally, errors are accumulated through the irrevocable process, with undesirable unidirectional bias. + +Latent diffusion models Recently, diffusion models [18, 36] introduced a new paradigm for text-to-image generation [33]. They generate high-quality images by progressively denoising noise, but require iterative steps and a large amount of computation. Two-stage approaches [10, 32] propose to perform diffusion in latent space successfully reducing computation while maintaining performance. However, most of them still require about 50 inference steps to show plausible performance, and since denoising is applied globally to the entire image during the inference process, locally correcting the image requires complicated processes [32]. + +Token-based diffusion models Token-based diffusion models [5, 11, 13, 23, 43] have impressive performance in generating complex scenes from the text. Unlike conventional diffusion models, where the Gaussian noise is applied in a continuous space, token-based diffusion models quantize latent features into tokens and apply a categorical corruption process in a discrete state space [1]. Among various discrete diffusion methods, mask-based diffusion is most commonly used due to its attractive properties; they have the advantage of sampling multiple tokens simultaneously, resulting in fewer sampling steps and faster generation time than the latent diffusion model. However, selecting multiple tokens at once leads to inconsistency throughout the image which is often called a joint distribution issue. Moreover, tokens are heavily influenced by the spatially adjacent tokens that have already been modified so that a misaligned token could disperse the errors throughout the image. To tackle this issue, some works [6, 23] modified the sampling strategy while other works [24] introduced new models to find the misaligned tokens. + +In particular, Lezama et al. [24] proposed a learnable model that removes the misaligned tokens at each step. While most existing works randomly select the tokens to discard, this work examines all sampled tokens. It outputs the score of each location where the score indicates whether the corresponding token is under the real distribution. After the predicting and sampling process of the generator, the model selects the locations according to the scores of the corresponding tokens. + +# 3. Text-conditioned token selection + +The diffusion process of the masked image generative models starts with a completely masked image, and samples the tokens of $x_{t}$ for the whole image to predict $\hat{x}_0$ while re-masking a portion of them to obtain the denoised sample $x_{t - 1}$ . Existing masked image generative models use uniform or confidence-based sampling strategy to determine + +![](images/9637f9b9ddc0ff5c094d15ece8ae9a5060a6945e44a638766701d5786753e0eb.jpg) +Figure 2: Reconstructed images using $\hat{\mathbf{x}}_0^{(\mathrm{t})}$ each step during diffusion using the fixed and revocable method. Only the revocable method is able to edit the tokens to generate a plane according to the text "A view of the end of an airplane in the sky over mountains." + +the locations to re-mask: Uniform sampling [13, 23] randomly selects the locations of the tokens to keep among the predicted tokens at all locations, while confidence-based strategy, such as purity sampling [37], selects the locations based on the confidence score which is defined as follows: + +$$ +\operatorname {C S} (i, t) = \max _ {j = 1, \dots , K} p \left(x _ {0} ^ {i} = j \mid x _ {t} ^ {i}\right), \tag {1} +$$ + +where $t$ is the step number, $K$ is the size of the codebook, and $i$ is the location of the token. Intuitively, a high confidence score means the generator is relatively certain that a token should be sampled at the corresponding location. By keeping these confident tokens while discarding the ambiguous ones, purity sampling shows good performance in practice [5]. + +However, these strategies are non-revocable, i.e., fixed, which means tokens that are once sampled cannot be revised afterward, even if they do not match the given text or do not align with other tokens. Once the misaligned tokens are fixed during the process, errors accumulate and spread throughout the image. We observe that fixed strategies negatively affect the text alignment as visualized in Figure 1. + +# 3.1. Random revoke sampling strategy + +In order to address this issue, we first introduce a new sampling strategy called random revoke sampling and use it as a baseline for our final method. It is similar to uniform sampling in a way that they both determine the locations by sampling from a uniform distribution. However, while uniform sampling is a fixed strategy, random revoke sampling selects the locations from the whole image space regardless of the previously fixed locations, making the process reversible. It randomly selects a certain number of tokens to preserve for the next step and revokes the previously fixed ones. The revocable feature can better align the images to the given text information, especially in complex scenes. + +![](images/e551c84949edbed60b486d5b3747aee997496a7a56e952bf7507dd7fbd43db19.jpg) +Figure 3: Reconstructed images using $\hat{\mathbf{x}}_0^{(\mathrm{t})}$ each step during diffusion using the revocable method with and without FAS. The backgrounds of the images in the top row are over-simplified while our proposed FAS prevents this, as shown in the bottom row. The text is "This small bird is greyish in color with flecks of yellow on the back and breast, and a bit of white on the belly." + +To be specific, as shown in the top row of Figure 2, fixed strategies have limitations in recovering from errors or generating diverse images due to their inability to regenerate incorrect tokens or modify missing essential objects or attributes. In contrast, revocable strategies enhance the text alignment by providing an opportunity to correct invalidly generated tokens, as shown in the bottom row of Figure 2. Since revocable strategies have no constraints on which tokens to remove or fix, they have the advantage of mitigating error accumulation. + +However, random revoke sampling suffers from two main issues that must be addressed. The first issue is the instability resulting from randomly discarding previously sampled tokens which worked as a given condition of the step. It can lower the quality of the outputs as in Figure 1. + +The second issue, namely the over-simplification, arises when it is used in the generation process involving longer steps. The low-frequency parts of the output images, such as backgrounds, tend to be over-simplified compared to fixed methods, which generate more realistic backgrounds. The over-simplified samples can be seen in Figure 3, and it has a significant impact on the FID metric as in Table 1 and 2. We observe that this is due to excessive repetition of resampling over a large number of steps, while a fixed schedule does not resample any tokens. More opportunities for resampling make the tokens converge towards safer and simpler patterns. As shown in the top row of Figure 3, the high-frequency areas with rich visual details do not become over-simplified even after multiple resampling processes, but the low-frequency areas become less diverse and more simplified in patterns quickly due to excessive resampling. + +To confirm this trade-off between text alignment and image quality, we define a persistent sampling as an interpolation between uniform sampling and random revoke sampling. In details, a persistent weight boosts the probability of keep + +# Algorithm 1 Persistent Sampling + +Input: $k_{t}$ : number of tokens sampled at step $t$ , $w$ : persistent weight, $N$ : number of all tokens + +1: $x_{T}\gets [[\mathrm{MASK}]]_{N}$ +2: for $t = T, T - 1, \dots, 1$ do +3: $\hat{x}_0 = G_\theta(x_t, c)$ +4: $A_{t} = \{i|x_{t}^{i} = [\mathrm{MASK}]\}$ +5: $m = k_{T} + k_{T - 1} + \ldots + k_{t}$ +6: $\mathcal{U}_t(I = i|i\in A_t):\mathcal{U}_t(I = i|i\in A_t^C) = 1:w$ +7: $\rightarrow$ uniform distribution + persistent weight +8: $i_1, i_2, \dots, i_m \sim \mathcal{U}(I) \gets$ sample without replacement +9: $x_{t}\gets [[\mathrm{MASK}]]_{N}$ +10: for $i = i_1, i_2, \ldots, i_m$ do +11: $x_{t}^{i}\gets \hat{x}_{0}^{i}$ +12: end for +13: end for +14: Return: Generated image $x_0$ + +ing the tokens that were sampled in the previous step. While random revoke sampling assigns the same probability to all the locations, persistent sampling multiplies the persistent weight to the probability of previously kept locations. Further details are shown in Algorithm 1. If the persistent weight $w$ is 1, all the tokens would have the same probability to be fixed, such as random revoke sampling. If $w$ is sufficiently large, the previously sampled tokens would be kept until the end, such as uniform sampling. In the graph in Figure 1, a trade-off can be seen that as the sampling strategy gets closer to the random revoke sampling, text alignment gets better, while the image quality gets worse. We further analyze the effect of these sampling methods in Section 4.5. + +# 3.2. Text-conditioned token selection + +We propose a learnable model called Text-Conditioned Token Selection (TCTS). TCTS is trained to output a score map to detect tokens that are not under the real distribution for the given text condition, and selects the tokens that are not aligned with the text or other tokens in the image. By masking the selected tokens at each step, the diffusion network receives high-quality text-aligned tokens as the input to sample the next set of tokens which alleviates the error accumulation and joint distribution issue of the previous sampling strategies. Moreover, since TCTS is trained to discriminate the well-aligned tokens, it detours the instability problem as mentioned earlier, unlike other revocable methods. From the experimental results in Section 4, we observe that TCTS enhances the text alignment without compromising the image quality. + +Frequency adaptive sampling (FAS) The simplest way to solve the over-simplification problem in longer steps is to use the persistent sampling to alleviate the excessive resampling of tokens. However, as previously mentioned, this method + +![](images/98df41975cbefd50a342c61b36778d24047b963351f2b819e3512fda5bcd092f.jpg) +Figure 4: Overall generation framework of proposed TCTS and FAS. After the generator predicts the tokens, TCTS exploits the text condition to detect misaligned tokens and outputs the score map. Meanwhile, FAS splits the tokens according to the frequency using the self-attention map from the generator, performing revocable sampling to high-frequency split and persistent sampling to low-frequency split. The adaptive sampling predicts $\hat{x}_0$ and decide a few of the locations to mask according to $x_{t}$ . The token maps produced by FAS is combined with the score map to predict $x_{t - 1}$ . After iterative process, our model removes all the masks and generates $x_0$ . + +
StepMethodMID-L ↑SOA-I ↑CLIP-S ↑FID-30K ↓
16Purity11.0272.380.247419.20
Uniform17.9474.800.250016.17
RR23.6078.790.252617.10
TCTS + FAS26.7279.520.255914.45
25Purity16.8475.210.248718.39
Uniform22.2777.080.252415.91
RR26.7781.100.254318.43
TCTS + FAS27.7980.870.256315.39
+ +may result in a slight compromise in text alignment and limit the opportunity to correct the image through resampling. To further improve this trade-off, we propose a new method called Frequency Adaptive Sampling (FAS), which can be applied to TCTS. + +FAS is a method that utilizes the generator's self-attention map to limit resampling only in the low-frequency areas of the image. In [21], it is known that the generator's self-attention layer contains frequency information of the image, which is also observed in token-based diffusion models. We utilized this to distinguish the frequency areas of the image without additional operations, and applied persistent sampling only to the low-frequency areas using persistent weight. As shown in the bottom row of Figure 3, revocable sampling along with FAS method allows repetitive resampling in areas that require rich visual details, while limiting the number of resampling in relatively simple areas preventing oversimplification. Detailed algorithm is in the supplementary A. The overall framework of our model, including both TCTS + +Table 1: Quantitative evaluation of sampling methods on MS-COCO dataset. The ground truth MID on MS-COCO image-caption pairs is 54.73. The classifier-free guidance scale was fixed at 5 for all sampling methods. + +
StepMethodMID-L ↑CLIP-S ↑FID ↓
16Purity-24.210.241015.21
Uniform-25.600.240416.57
RR-25.030.237117.38
TCTS + FAS-19.880.24612.35
25Purity-21.260.238412.60
Uniform-23.040.239613.02
RR-23.290.236414.53
TCTS + FAS-18.310.240913.67
+ +Table 2: Quantitative evaluation of sampling methods on CUB dataset. The ground truth MID on CUB image-caption pairs is 15.85. We omit SOA in this table as CUB dataset consists of only one bird per photo, diminishing the metric's significance. + +and FAS, is illustrated in Figure 4. + +# 4. Experiments + +To validate the efficacy of our token selection framework, we modify the transformer from [37] to contain 310M parameters for MS-COCO dataset and 83M parameters for CUB dataset. We extract text features from CLIP ViT-B/32 [27] for a text-conditioned generation. In the training process of TCTS, we freeze the generator's parameters and independently train our model, making it applicable to other types of token-based diffusion models. We use the binary cross-entropy loss for the objective and validate our method and baselines on MS-COCO [25] and CUB [14] datasets by training them for 200 epochs. We leverage classifier-free guidance [20] by stochastic sampling of the guidance strength to improve the generation quality. We further provide the details of the architecture, hyperparameters and additional analysis on the use of classifier-free guidance in the supplementary A. + +
ModelMID-L ↑MID-B ↑SOA-I ↑FID-30K ↓
GLIDE [26]1.031.00-32.08
AttnGAN [40]-65.20-8.9039.0129.15
DM-GAN [23]-44.663.5148.0322.90
DF-GAN [38]-58.75-15.21-31.75
VQ-Diffusion [41]-19.635.77-13.13
LAFITE [47]6.2635.1774.788.03
TCTS + FAS26.0238.9879.529.75
+ +Metrics Since FID [16] metric is known to be problematic for its inability to exactly represent the image quality [5] and unable to consider the text alignment, we additionally use Mutual Information Divergence (MID) [22] to evaluate the text-alignment of generated images which enables sample-wise evaluation and outperforms previous metrics in human Likert-scale judgment correlations. In particular, since MID responds more sensitively to images generated with foiled captions [35], it is more appropriate for analyzing text alignment with complex and challenging captions. Moreover, we use the CLIP score [15] and SOA-I [17], which are widely used in text-to-image synthesis. Note that the CLIP score is calculated with ViT-L/14 throughout the experiments. + +# 4.1. Text-to-image synthesis + +Figure 1 Left and Middle visualizes the generated images using improved VQ diffusion [37] with varying sampling methods. Generated examples from uniform token sampling are poorly aligned with the given texts, and often include erratic partitions of the objects since the predicted tokens selected at the same step cannot appropriately reflect the change of others (e.g., smashed donuts, and imperfect chair and seagull). Random revoke (RR) sampling, which can re-update previously sampled tokens, seems to improve alignment with text by editing out nonsensical regions and iteratively considering the entire scene at each generation step. Yet, RR selects tokens in a completely random manner without any constraints, which results in suboptimal generation quality and alignment with the texts (e.g., disappeared man and train). On the other hand, ours mitigates the problem from RR by selecting text-conditioned tokens at the sampling phase and successfully generates high-quality images which contain a clear semantic connection to the given text captions. + +This phenomenon can be quantitatively seen in Table 1. 16-step generation of random revoke sampling shows better text alignment than even 25-step generation of other fixed + +![](images/2d6ff5a6f95806e95523125a66d69388a363a7848df400f3a1e5963933a2d8c6.jpg) + +![](images/c787d8bd112bda7abb8468484b3884afd8c9c3321e9dd3a244d7aa6f3f6d84b2.jpg) +Figure 5: Results of image refinement using TCTS. Top: Original samples, Bottom: Refined images for 8-steps with TCTS. + +Table 3: Comparison of the proposed method with recent text-to-image generative models on MS-COCO dataset. MID-L and MID-B are calculated with ViT-L/14 and ViT-B/32 each. The ground truth MID-L and MID-B on MS-COCO image-caption pairs is 54.73, 41.57 each. Ours are evaluated with 16-step setting and we adopt classifier-free guidance from 3 to 5. + +
MethodMID-L ↑FID ↓
Original12.1016.44
RR13.6417.18
TCTS15.6416.92
Original4.5419.13
Random masked6.0619.12
TCTS masked6.4718.64
+ +Table 4: Qualitative evaluation of the refined images. Top: Refinement with additional revision steps. Bottom: Refinement with masking lowest-scoring tokens. + +sampling methods. We can also observe the trade-off between CLIP score and FID by comparing each method to their 25-step versions. Our model outperforms other baselines in most of the metrics especially when comparing MID-L to fixed sampling methods such as Purity and Uniform. Since our model can make corrections during the generation process, it requires fewer steps to match the performance of our baseline sampling methods, see Section 4.5 for further analysis. We also present the results of our experiment on the CUB dataset with fewer parameters in Table 2, which demonstrates satisfactory performance. However, the fact that the CUB dataset contains only single object per image resulted in a tendency that is slightly different from what we mentioned earlier. In Table 3, we compared the performance of our method with other text-to-image generative models. Our model highly exceeds other models in text alignment metrics, especially in MID-L and closely approximate the MID-B value of the ground truth images. More samples generated by our model is in the supplementary B. + +# 4.2. Image refinement + +The advantage of the masked image generative model is that it enables fast local refinement. We conduct image refinement in two separate methods, inspired by [23] and [24]. First, we use TCTS to apply additional revision steps to images generated with uniform sampling. The overall image + +quality is improved by adding additional refinement steps as shown in Figure 5. In order to demonstrate the image refinement performance of our TCTS, we also measure the performance of the image refinement with random revoke sampling. While RR improves does improve sample quality, additional refinement steps increase the FID score, caused by a similar effect to increasing the generation step. + +We additionally perform the experiment by masking $60\%$ lowest-scoring tokens with TCTS, and generate the images with uniform sampling. We observe that all metrics of the refined images outperform those of the original images. In Table 4, we compare them with samples refined after randomly masking the tokens without TCTS. Due to the page limit, we describe further details of the two refinement methods and visualize more samples in the supplementary B. + +# 4.3. Mask-free object editing + +Since the masked image generative model is capable of local refinement, image editing without manual masking is possible simply by randomly masking a part of the image and resampling it with the new text condition. However, this method requires a low masking ratio and many resampling steps to maintain the overall structure of the image. This can result in significant changes to unnecessary parts or, as mentioned earlier, over-simplification issues. Additionally, even with many steps, editing large objects with small masks can be more challenging due to the significant influence of surrounding tokens rather than the new text condition. + +Motivated by the operation of self-attention maps in frequency adaptive sampling, we leverage a cross-attention map corresponding to the word of the object that is to be changed, giving weights to resample tokens so that the corresponding locations can be resampled. Then, it enables efficient image editing with fewer steps and makes it easier to edit larger objects. Although it is similar to DiffEdit [7], ours operates in a more straightforward manner without additional computations, thanks to the capability of local refinement of the masked generative model. In Figure 6, we edit the object by adding $25\%$ noise in only 10-steps. With our method, we can edit more quickly with fewer editing steps, which further does not suffer from the over-simplification problem. More samples are in the supplementary B. + +# 4.4. High-resolution image synthesis + +Continuous diffusion models that use encoders and decoders, such as masked image generative models, can generate larger images that are not in the training set. Similar to Bond-Taylor et al. [3], we divide the tokens into subsets according to the model input size, individually pass them through the model, and then spatially aggregate tokens to synthesize a high-resolution image. In this way, we are able to generate more realistic and high-resolution samples with the same TCTS model without additional training. Addition- + +![](images/00548aacf89f2c11b8b5c60d676ea59b691698d44e43eef4bb1a7e34b5ee3b55.jpg) +A dog teddy bear looks back from the passenger seat of a vehicle. + +![](images/7d3582282c7d74b2a290ae6b339220c34417c771de7e168d4e7236fc6552a4a3.jpg) +A wooden table topped with earrot cake donuts, broccoli. + +![](images/1739847b09f6a36a2b9c07eea36e656d0371110d99d7c27ce027fe26809ffac8.jpg) +Figure 6: Examples of mask-free editing samples with cross-attention map. The cross-attention map is multiplied to the score map of TCTS to perform weighted sampling. +A short train traveling through a mountainous landscape. +Figure 7: Examples of high resolution $(512^{2})$ samples. Left: Image generated with all mask tokens, Right: Image generated with TCTS as a super-resolution unit. + +![](images/fdbd26e656252abb8373c775bf4bdaf072318cefcd9143841c268fcc99d65d1a.jpg) +A family on a boat sailing down a river. + +ally, we propose a new method to generate high-resolution images only with low-resolution TCTS. To be specific, we first generate a small-size image with TCTS and upsample the token map in bicubic mode to the desired size. Then we divide the high-resolution token map into overlapping small-size sections and refine all the sections several times. Then, we can generate high-resolution images with the low-resolution TCTS where we visualize the generated samples in Figure 7. + +# 4.5. Further analysis + +Importance of early-stage sampling strategy As mentioned in Section 3.1, by controlling the persistent weight, we can interpolate the two baseline sampling methods: uniform sampling and random revoke sampling. We designed various experiments to find out the impact of the sampling + +![](images/705f6d7abccccd96d76b6fe280dd3a88e26197ec30e98f22b99bb45a34e8752c.jpg) +(a) Real + +![](images/32adead52fac8ee739171cd1644ff8159fe44b54bd03b206702e9c3a6d6c20a7.jpg) + +![](images/cae83e708f754a98ac1338a6c0d5f3cd96f6d9ae92e93c5b7ecae2d2047c8302.jpg) +road + +![](images/9420161652a97a82d01ca9af46df72d2602ede70dac5d3351411f6d9736ecc57.jpg) +(b) Synthesized + +![](images/91ee9af1d6669cd7d757c429b30c13a9c71f772362e97f26955cabed26a68855.jpg) +A + +![](images/491864ab18b4ebd38582b792720a487d9230f20ab8522995f53bc343dbe67136.jpg) +boat + +![](images/b5ec268ce93fb9b89c4f4eb02f4b0f168cf3f26a38b6c0a8fc705989ee7d4069.jpg) +water +Figure 8: Visualization of the averaged cross-attention maps for each word from TCTS. Top: Image from COCO dataset: "A large black bear walking down a dirty road". Bottom: Synthesized image: "A small boat in a narrow body of water". + +method with respect to the time steps. We switched the sampling methods during the time steps, from uniform sampling to random revoke sampling (U2R) and vice versa (R2U). In the rightmost plot in Figure 1, we can observe that U2R shows better text-alignment performance than R2U. + +These results suggest that the early stage sampling, in which the masked ratio is high, substantially influences the text alignment of the final image, which is similarly observed in continuous diffusion models [2]. Generating the whole image and then processing it through a few revision steps like [23] might help the image quality but can not conspicuously enhance the essential text alignment. Therefore, in order to obtain an image that aligns well with the text, sampling must be done carefully in the early stage when not many tokens are generated. This analysis is also shown in the result that the revocable strategy exhibits better performance by aggressively resampling tokens several times. In particular, since weak generators with poor generation performance suffer more from the joint distribution issues, giving a sufficient number of recovery chances to discard tokens and draw new ones at the early stage is the key to the desired text alignment. + +Text-conditioned sampling The experiments show that TCTS plays a crucial role in enhancing text alignment in image generation. Although we did not utilize any text-specific loss during training, the model still effectively processes the text condition and achieves optimal performance. To verify the accuracy of TCTS's attention towards text-related parts of the image, we visualized the average cross-attention layer for each word in Figure 8. The results demonstrate TCTS's ability to attend to each word in both real and synthesized images. By helping the selection of relevant tokens during the generating stage, TCTS can boost model performance without the requirement of additional complex text-related loss. + +
MethodMID-L ↑SOA-I ↑CLIP-S ↑FID-30K ↓
RR26.7781.100.254318.43
RR + FAS26.3280.780.253916.01
RR + pw24.2379.140.253415.96
TCTS27.8280.930.256516.69
TCTS + FAS27.7980.870.256315.39
TCTS + pw25.1279.510.255615.12
+ +Table 5: Ablation study of using FAS and persistent weight in revocable methods on MS-COCO dataset. All models were evaluated on 25-step setting. The ground truth MID-L on MS-COCO image-caption pairs is 54.73. + +
StepModelMID-L ↑FID-30K ↓Time
50Uniform24.0416.84×1
RR26.5521.12
16TCTS + FAS26.7214.45×0.45
+ +Table 6: Inference time relative to MID/FID of our baselines and models on MS-COCO dataset. We expressed the inference time of our models as a multiple of other baselines. Our model outspeeds other baselines while performing better in both MID-L and FID-30K. Further analysis is in the supplementary C. + +Ablation study of FAS As mentioned in Section 3.2, FAS decides whether to apply persistent weight or not to each location. If the whole image is considered low-frequency, FAS would apply the weight to every location, which is the same as the persistent sampling. On the other hand, if the whole image is considered high-frequency, combining FAS would not apply the weight anywhere, which does not change the sampling strategy. To further analyze the effect of FAS, we evaluated it on two revocable methods: random revoke sampling and TCTS. Attaching FAS to a base sampling method can be regarded as a mixture of two methods: the base revocable sampling method and persistent sampling method. Table 5 clearly shows this relation of FAS and persistent weight. While FAS-attached methods do not show the best performance, the performance of them is always in close proximity to the better one of the two comparing methods of each. FAS-attached methods balances well between text-alignment and image quality without any additional training. + +Inference time Our TCTS and FAS compute additional operations at each generation step, resulting in marginally increased inference time compared to other token selection baselines. However, regarding the text alignment and the image quality, the relative inference time is considerably decreased. In Table 6, our model outperforms 50-step image generation of baseline sampling methods within only 16 steps in MID-L and FID-30K, which means that our model can synthesize better samples with only $\times 0.45$ time. + +# 5. Conclusion + +This paper examines which factors in the masked diffusion process impact the output images and cause the trade-off between image quality and text alignment. We empirically find that the revocable sampling significantly improves the text alignment yet degrades the quality of the generated images. To tackle the problem, we propose a simple token sampling strategy, coined text-conditioned token selection, that combines learnable and revocable methods, pushing the boundary of the trade-off between image quality and text alignment without the necessity to update the large-scale pre-trained generator. We find that collaborative sampling in a persistent and revocable manner surprisingly alleviates over-simplification issues in the generated backgrounds. Our proposed method can be utilized as an image refinement tool for various generative tasks, which is remarkably fast to generate high-quality images within much fewer steps. + +# 6. Acknowledgements + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST)) and (No. RS-2022-00187238, Development of Large Korean Language Model Technology for Efficient Pre-training). This work was supported by KAIST-NAVER Hypercreative AI Center. + +# References + +[1] Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. In Advances in Neural Information Processing Systems, pages 17981-17993. Curran Associates, Inc., 2021. 1, 3 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 8 +[3] Sam Bond-Taylor, Peter Hessey, Hiroshi Sasaki, Toby P Breckon, and Chris G Willcocks. Unleashing transformers: parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vector-quantized codes. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXIII, pages 170–188. Springer, 2022. 7 +[4] Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdzal, and Adriana Romero Soriano. Instance-conditioned gan. Advances in Neural Information Processing Systems, 34:27517-27529, 2021. 1 +[5] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to + +image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1, 3, 6 +[6] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 1, 3 +[7] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance, 2022. 7 +[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1 +[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2 +[10] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. arXiv preprint arXiv:2208.13753, 2022. 1, 3 +[11] Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what happened: Unifying text-guided video completion via multimodal masked video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10681-10692, 2023. 3 +[12] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 1 +[13] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 3 +[14] Xiangteng He and Yuxin Peng. Fine-grained visual-textual representation learning. IEEE Transactions on Circuits and Systems for Video Technology, 30(2):520-531, 2019. 5 +[15] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528. Association for Computational Linguistics, 2021. 6 +[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[17] Tobias Hinz, Stefan Heinrich, and Stefan Wermter. Semantic object accuracy for generative text-to-image synthesis. IEEE transactions on pattern analysis and machine intelligence, 2020. 6 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 3 + +[19] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:47-1, 2022. 1 +[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5 +[21] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance, 2022. 5 +[22] Jin-Hwa Kim, Yunji Kim, Jiyoung Lee, Kang Min Yoo, and Sang-Woo Lee. Mutual information divergence: A unified metric for multimodal generative models. arXiv preprint arXiv:2205.13445, 2022. 6 +[23] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Draft-and-revise: Effective image generation with contextual rq-transformer. arXiv preprint arXiv:2206.04452, 2022. 1, 3, 6, 8 +[24] José Lezama, Huiwen Chang, Lu Jiang, and Irfan Essa. Improved masked image generation with token-critic. In European Conference on Computer Vision, pages 70-86. Springer, 2022. 2, 3, 6 +[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5 +[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 6 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5 +[28] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2 +[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2 +[31] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 3 + +[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 3 +[34] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 1 +[35] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. Foil it! find one mismatch between image and language caption. arXiv preprint arXiv:1705.01359, 2017. 6 +[36] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3 +[37] Zhicong Tang, Shuyang Gu, Jianmin Bao, Chen Dong, and Fang Wen. Improved vector quantized diffusion models. arXiv preprint arXiv:2205.16007, 2022. 1, 2, 3, 5, 6 +[38] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2020. 2, 6 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[40] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2, 6 +[41] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 1, 6 +[42] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 1, 2 +[43] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 3 +[44] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2 +[45] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stack + +gan++: Realistic image synthesis with stacked generative adversarial networks. IEEE transactions on pattern analysis and machine intelligence, 41(8):1947-1962, 2018. 2 +[46] Yufan Zhou, Chunyuan Li, Changyou Chen, Jianfeng Gao, and Jinhui Xu. Lafite2: Few-shot text-to-image generation. arXiv preprint arXiv:2210.14124, 2022. 2 +[47] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17907-17917, 2022. 2, 6 +[48] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dmgan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5810, 2019. 2 \ No newline at end of file diff --git a/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/images.zip b/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c1e2f3578c074ae94cfe0bd81e582fad6fd32f06 --- /dev/null +++ b/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad9f8853c35d7469129f1e3306623b9c202707e59010cb826dbc9b8826a1d544 +size 550125 diff --git a/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/layout.json b/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b34648756338fb7584674bef0db52fb7985dd5a5 --- /dev/null +++ b/2023/Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models/layout.json @@ -0,0 +1,8346 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 479, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 479, + 140 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 479, + 140 + ], + "type": "text", + "content": "Text-Conditioned Sampling Framework for Text-to-Image Generation with Masked Generative Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 162, + 165, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 162, + 165, + 188 + ], + "spans": [ + { + "bbox": [ + 88, + 162, + 165, + 188 + ], + "type": "text", + "content": "Jaewoong Lee* KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 190, + 172, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 190, + 172, + 203 + ], + "spans": [ + { + "bbox": [ + 82, + 190, + 172, + 203 + ], + "type": "text", + "content": "Republic of Korea" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 205, + 184, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 205, + 184, + 217 + ], + "spans": [ + { + "bbox": [ + 69, + 205, + 184, + 217 + ], + "type": "text", + "content": "hello3196@kaist.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 212, + 162, + 296, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 162, + 296, + 189 + ], + "spans": [ + { + "bbox": [ + 212, + 162, + 296, + 189 + ], + "type": "text", + "content": "Sangwon Jang* Yonsei University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 190, + 299, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 190, + 299, + 203 + ], + "spans": [ + { + "bbox": [ + 209, + 190, + 299, + 203 + ], + "type": "text", + "content": "Republic of Korea" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 209, + 205, + 298, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 205, + 298, + 217 + ], + "spans": [ + { + "bbox": [ + 209, + 205, + 298, + 217 + ], + "type": "text", + "content": "agwmon@gmail.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 162, + 493, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 162, + 493, + 188 + ], + "spans": [ + { + "bbox": [ + 353, + 162, + 493, + 188 + ], + "type": "text", + "content": "Jaehyeong Jo, Jaehong Yoon KAIST" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 378, + 190, + 468, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 190, + 468, + 203 + ], + "spans": [ + { + "bbox": [ + 378, + 190, + 468, + 203 + ], + "type": "text", + "content": "Republic of Korea" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 205, + 524, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 205, + 524, + 217 + ], + "spans": [ + { + "bbox": [ + 324, + 205, + 524, + 217 + ], + "type": "text", + "content": "{harryjo97, jaehong.yoon}@kaist.ac.kr" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 223, + 323, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 223, + 323, + 250 + ], + "spans": [ + { + "bbox": [ + 127, + 223, + 323, + 250 + ], + "type": "text", + "content": "Yunj Kim, Jin-Hwa Kim, Jung-Woo Ha Naver AI Lab" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 181, + 251, + 270, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 251, + 270, + 265 + ], + "spans": [ + { + "bbox": [ + 181, + 251, + 270, + 265 + ], + "type": "text", + "content": "Republic of Korea" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 93, + 268, + 358, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 268, + 358, + 279 + ], + "spans": [ + { + "bbox": [ + 93, + 268, + 358, + 279 + ], + "type": "text", + "content": "{yunji.kim, jinhwa.kim, jungwoo.ha}@navercorp.com" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 404, + 224, + 482, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 224, + 482, + 250 + ], + "spans": [ + { + "bbox": [ + 404, + 224, + 482, + 250 + ], + "type": "text", + "content": "Sung Ju Hwang KAIST" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 399, + 251, + 488, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 251, + 488, + 266 + ], + "spans": [ + { + "bbox": [ + 399, + 251, + 488, + 266 + ], + "type": "text", + "content": "Republic of Korea" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 386, + 268, + 501, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 268, + 501, + 278 + ], + "spans": [ + { + "bbox": [ + 386, + 268, + 501, + 278 + ], + "type": "text", + "content": "sjhwang82@kaist.ac.kr" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 143, + 307, + 191, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 307, + 191, + 319 + ], + "spans": [ + { + "bbox": [ + 143, + 307, + 191, + 319 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 331, + 290, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 290, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 290, + 582 + ], + "type": "text", + "content": "Token-based masked generative models are gaining popularity for their fast inference time with parallel decoding. While recent token-based approaches achieve competitive performance to diffusion-based models, their generation performance is still suboptimal as they sample multiple tokens simultaneously without considering the dependence among them. We empirically investigate this problem and propose a learnable sampling model, Text-Conditioned Token Selection (TCTS), to select optimal tokens via localized supervision with text information. TCTS improves not only the image quality but also the semantic alignment of the generated images with the given texts. To further improve the image quality, we introduce a cohesive sampling strategy, Frequency Adaptive Sampling (FAS), to each group of tokens divided according to the self-attention maps. We validate the efficacy of TCTS combined with FAS with various generative tasks, demonstrating that it significantly outperforms the baselines in image-text alignment and image quality. Our text-conditioned sampling framework further reduces the original inference time by more than " + }, + { + "bbox": [ + 46, + 331, + 290, + 582 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 331, + 290, + 582 + ], + "type": "text", + "content": " without modifying the original generative model." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 593, + 127, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 593, + 127, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 593, + 127, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 613, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 288, + 696 + ], + "type": "text", + "content": "In the flood of generative AI systems for vision domains, text-conditional image generation [26, 33, 42] is coming to the fore in recent years. Although many recent works have achieved success in synthesizing high-quality images [19, 34] with plausible class-alignment in class-conditional cases [4, 8], text-to-image (T2I) generation is more challenging since generating visual outputs" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 308, + 547, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 547, + 499 + ], + "type": "text", + "content": "that are semantically aligned with input texts is a nontrivial problem. We can roughly categorize the works on text-to-image generation into transformer-based autoregressive (AR) [12, 30, 33] and diffusion-based [18, 29] approaches. Along with the advancement of language models, AR models using transformers have shown impressive performance in text-to-image generation. Despite their success, they suffer from the problem of unidirectional bias, which is undesirable for image generation, and crucially, the sampling process requires over 10 times as much time compared to existing models. Another line of work is diffusion-based methods that aim to generate images by iteratively denoising noisy samples. In particular, several continuous diffusion models [10, 32] have achieved outstanding performance and reduced computational cost. Yet, they require excessive sampling steps to obtain high-quality images during the inference." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 504, + 548, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 548, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 548, + 696 + ], + "type": "text", + "content": "Recently, a new family of generative models, called token-based diffusion model [6, 23, 41], has emerged as an alternative to tackle the problem of text-to-image generation. Token-based diffusion models quantize the latent features into tokens and apply categorical corruption process in discrete state spaces [1], while conventional diffusion models use Gaussian noise in continuous space. Among various discrete diffusion methods, mask-based diffusion, similar to the absorbing state diffusion used in [1], is mostly used. Compared to existing AR models, this token-based approach is advantageous for speeding up the generation process via simultaneously sampling multiple tokens. Despite the limitation on the reconstruction capacity, these token-based diffusion models significantly outperform the competitors in terms of FID scores, even with fewer sampling steps compared to the continuous diffusion models [5, 37]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "However, sampling multiple tokens at once often leads" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23252" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 190, + 220 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 190, + 220 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 190, + 220 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 190, + 220 + ], + "type": "image", + "image_path": "cfae35d70e7ecfe92b1099a9975ede949fd90117189502f79dc084a5d2b6d3e3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 132, + 239, + 152 + ], + "lines": [ + { + "bbox": [ + 190, + 132, + 239, + 152 + ], + "spans": [ + { + "bbox": [ + 190, + 132, + 239, + 152 + ], + "type": "text", + "content": "A man sitting on top of a bench near the ocean." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 190, + 178, + 241, + 198 + ], + "lines": [ + { + "bbox": [ + 190, + 178, + 241, + 198 + ], + "spans": [ + { + "bbox": [ + 190, + 178, + 241, + 198 + ], + "type": "text", + "content": "A box containing an interesting variety of donuts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 244, + 70, + 384, + 220 + ], + "blocks": [ + { + "bbox": [ + 192, + 81, + 239, + 110 + ], + "lines": [ + { + "bbox": [ + 192, + 81, + 239, + 110 + ], + "spans": [ + { + "bbox": [ + 192, + 81, + 239, + 110 + ], + "type": "text", + "content": "The train is rounding the bend of a track on the mountain side." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 244, + 70, + 384, + 220 + ], + "lines": [ + { + "bbox": [ + 244, + 70, + 384, + 220 + ], + "spans": [ + { + "bbox": [ + 244, + 70, + 384, + 220 + ], + "type": "image", + "image_path": "c8d7819a9fcbcb251285c601da72f376bdb81f8d84ae948374c3df55d7da4e2a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 81, + 432, + 101 + ], + "lines": [ + { + "bbox": [ + 385, + 81, + 432, + 101 + ], + "spans": [ + { + "bbox": [ + 385, + 81, + 432, + 101 + ], + "type": "text", + "content": "A man standing in a field with some cattle." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 385, + 130, + 433, + 158 + ], + "lines": [ + { + "bbox": [ + 385, + 130, + 433, + 158 + ], + "spans": [ + { + "bbox": [ + 385, + 130, + 433, + 158 + ], + "type": "text", + "content": "A single seagull standing on the coast with waves in the background." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 384, + 177, + 436, + 198 + ], + "lines": [ + { + "bbox": [ + 384, + 177, + 436, + 198 + ], + "spans": [ + { + "bbox": [ + 384, + 177, + 436, + 198 + ], + "type": "text", + "content": "Autumn leaves radiate a golden glow in afternoon sunlight." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 434, + 74, + 545, + 222 + ], + "blocks": [ + { + "bbox": [ + 434, + 74, + 545, + 222 + ], + "lines": [ + { + "bbox": [ + 434, + 74, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 434, + 74, + 545, + 222 + ], + "type": "image", + "image_path": "e973a1da5d3c0c869038661864c295908823cf4d612a703b514d61019eeeae8b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 223, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 223, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 547, + 289 + ], + "type": "text", + "content": "Figure 1: Generated samples on MS-COCO dataset and evaluation graph of various sampling methods showing their trade-off. Uniform sampling is a fixed strategy with notably poor text alignment compared to other methods (FID-40K: 15.61, MID-L: 21.23). Random revoke sampling is a revocable strategy with improved text alignment (FID-40K: 16.81, MID-L: 26.98). Ours is TCTS combined with FAS, where both the image quality and the text alignment are significantly better compared to those of baselines (FID-40K: 13.6, MID-L: 29.5). Metrics are measured on all 40K images with their corresponding single caption. The classifier-free guidance scale was fixed at 5 for all sampling methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 298, + 288, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 288, + 477 + ], + "type": "text", + "content": "to inconsistency throughout a generated image [37]. For each location, the generator outputs a probability distribution that are coherent with each other. However, there is no guarantee that every single token sampled from the distributions will perfectly align with one another. In other words, it is still possible for the generator to sample incompatible tokens regardless of the generator's capability, leading to potentially nonsensical outputs [24, 37]. This results in a trade-off between the sampling steps and generation quality. Reducing the sampling steps leads to faster generation but results in performance degradation due to a large number of simultaneously sampled tokens. Especially, this problem further stands out in the text-to-image generation tasks since the distribution of text-aligned images is more restricted than the distribution of unconditioned or class-conditioned images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 486, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 712 + ], + "type": "text", + "content": "To address these limitations, we propose a novel sampling approach that refines the images during the diffusion process based on the text condition, which we refer to as the Text Conditioned Token Selection (TCTS). TCTS is a sampling strategy that can mask out and resample previously sampled tokens, which we refer to as revocable. To find the tokens that do not align with the given text condition, we propose to train a model that selects tokens to be masked out and re-generated, to be well-aligned with the given text. Combining this approach with the revocable sampling scheme, TCTS can generate high-quality images with improved text-alignment in even fewer sampling steps compared to the naive generative model, as shown in Figure 1. We further introduce Frequency Adaptive Sampling (FAS) to solve the over-simplification problem that occurs when applying revocable methods for relatively longer steps. FAS leverages the self-attention map of the images and applies a mixed sampling method, which we call persistent sampling to prevent the issue. We summarize our contributions as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 298, + 545, + 458 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 317, + 298, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 545, + 346 + ], + "type": "text", + "content": "- We experimentally show that the revocable sampling strategies are crucial for the trade-off between the text-alignment and the image quality and provide in-depth analysis compared to previous fixed sampling methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 354, + 545, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 545, + 414 + ], + "type": "text", + "content": "- We propose a novel revocable sampling method based on a learnable token selection model that significantly improves the text alignment as well as the image quality even with fewer sampling steps, without the need of retraining the generator model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 422, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 545, + 458 + ], + "type": "text", + "content": "- Moreover, we propose a novel sampling method based on the self-attention map that can be combined with TCTS to solve the over-simplification problem." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 470, + 389, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 470, + 389, + 481 + ], + "spans": [ + { + "bbox": [ + 306, + 470, + 389, + 481 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 490, + 547, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 490, + 547, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 490, + 547, + 597 + ], + "type": "text", + "content": "Various works have tackled text-to-image generation tasks, and the majority of them are based on Generative Adversarial Networks (GANs) [38, 40, 44, 45, 46, 47, 48]. However, computing directly from the pixel space of an image is challenging, and generating high-resolution images requires substantial computation due to the large number of pixels involved. Therefore, a two-stage approach is often used: models first tokenize the images into a sequence of codes, and then predict the tokenized codes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "Autoregressive models Following the success of Transformer [39] and GPT [28], Autoregressive (AR) models [9, 30, 31, 42] proposed to tokenize images and pose the text-to-image generation problem as a problem of translating textual descriptions into image tokens, much like machine translation of one language into a different language. With the advancement of image tokenization and language models, this approach has achieved very high performance. However, even with the help of this tokenizing stage, there remains" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23253" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "a large number of tokens to predict, resulting in a significant delay in the generation time due to their autoregressive nature. Additionally, errors are accumulated through the irrevocable process, with undesirable unidirectional bias." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 126, + 289, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 126, + 289, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 126, + 289, + 258 + ], + "type": "text", + "content": "Latent diffusion models Recently, diffusion models [18, 36] introduced a new paradigm for text-to-image generation [33]. They generate high-quality images by progressively denoising noise, but require iterative steps and a large amount of computation. Two-stage approaches [10, 32] propose to perform diffusion in latent space successfully reducing computation while maintaining performance. However, most of them still require about 50 inference steps to show plausible performance, and since denoising is applied globally to the entire image during the inference process, locally correcting the image requires complicated processes [32]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 265, + 289, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 289, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 289, + 503 + ], + "type": "text", + "content": "Token-based diffusion models Token-based diffusion models [5, 11, 13, 23, 43] have impressive performance in generating complex scenes from the text. Unlike conventional diffusion models, where the Gaussian noise is applied in a continuous space, token-based diffusion models quantize latent features into tokens and apply a categorical corruption process in a discrete state space [1]. Among various discrete diffusion methods, mask-based diffusion is most commonly used due to its attractive properties; they have the advantage of sampling multiple tokens simultaneously, resulting in fewer sampling steps and faster generation time than the latent diffusion model. However, selecting multiple tokens at once leads to inconsistency throughout the image which is often called a joint distribution issue. Moreover, tokens are heavily influenced by the spatially adjacent tokens that have already been modified so that a misaligned token could disperse the errors throughout the image. To tackle this issue, some works [6, 23] modified the sampling strategy while other works [24] introduced new models to find the misaligned tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 505, + 288, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 288, + 613 + ], + "type": "text", + "content": "In particular, Lezama et al. [24] proposed a learnable model that removes the misaligned tokens at each step. While most existing works randomly select the tokens to discard, this work examines all sampled tokens. It outputs the score of each location where the score indicates whether the corresponding token is under the real distribution. After the predicting and sampling process of the generator, the model selects the locations according to the scores of the corresponding tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 621, + 228, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 228, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 228, + 634 + ], + "type": "text", + "content": "3. Text-conditioned token selection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "The diffusion process of the masked image generative models starts with a completely masked image, and samples the tokens of " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " for the whole image to predict " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " while re-masking a portion of them to obtain the denoised sample " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": ". Existing masked image generative models use uniform or confidence-based sampling strategy to determine" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 71, + 547, + 205 + ], + "blocks": [ + { + "bbox": [ + 308, + 71, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 547, + 205 + ], + "type": "image", + "image_path": "9637f9b9ddc0ff5c094d15ece8ae9a5060a6945e44a638766701d5786753e0eb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 209, + 547, + 265 + ], + "lines": [ + { + "bbox": [ + 304, + 209, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 547, + 265 + ], + "type": "text", + "content": "Figure 2: Reconstructed images using " + }, + { + "bbox": [ + 304, + 209, + 547, + 265 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^{(\\mathrm{t})}" + }, + { + "bbox": [ + 304, + 209, + 547, + 265 + ], + "type": "text", + "content": " each step during diffusion using the fixed and revocable method. Only the revocable method is able to edit the tokens to generate a plane according to the text \"A view of the end of an airplane in the sky over mountains.\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 277, + 547, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 547, + 337 + ], + "type": "text", + "content": "the locations to re-mask: Uniform sampling [13, 23] randomly selects the locations of the tokens to keep among the predicted tokens at all locations, while confidence-based strategy, such as purity sampling [37], selects the locations based on the confidence score which is defined as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 356, + 357, + 547, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 357, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 356, + 357, + 547, + 378 + ], + "type": "interline_equation", + "content": "\\operatorname {C S} (i, t) = \\max _ {j = 1, \\dots , K} p \\left(x _ {0} ^ {i} = j \\mid x _ {t} ^ {i}\\right), \\tag {1}", + "image_path": "c41bab07be0146c5b0993e0f304ed924737849672d572710f6868f6f4aa59a0f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 387, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 387, + 545, + 458 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 387, + 545, + 458 + ], + "type": "text", + "content": " is the step number, " + }, + { + "bbox": [ + 304, + 387, + 545, + 458 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 387, + 545, + 458 + ], + "type": "text", + "content": " is the size of the codebook, and " + }, + { + "bbox": [ + 304, + 387, + 545, + 458 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 387, + 545, + 458 + ], + "type": "text", + "content": " is the location of the token. Intuitively, a high confidence score means the generator is relatively certain that a token should be sampled at the corresponding location. By keeping these confident tokens while discarding the ambiguous ones, purity sampling shows good performance in practice [5]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 459, + 547, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 547, + 543 + ], + "type": "text", + "content": "However, these strategies are non-revocable, i.e., fixed, which means tokens that are once sampled cannot be revised afterward, even if they do not match the given text or do not align with other tokens. Once the misaligned tokens are fixed during the process, errors accumulate and spread throughout the image. We observe that fixed strategies negatively affect the text alignment as visualized in Figure 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 552, + 488, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 488, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 488, + 564 + ], + "type": "text", + "content": "3.1. Random revoke sampling strategy" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "In order to address this issue, we first introduce a new sampling strategy called random revoke sampling and use it as a baseline for our final method. It is similar to uniform sampling in a way that they both determine the locations by sampling from a uniform distribution. However, while uniform sampling is a fixed strategy, random revoke sampling selects the locations from the whole image space regardless of the previously fixed locations, making the process reversible. It randomly selects a certain number of tokens to preserve for the next step and revokes the previously fixed ones. The revocable feature can better align the images to the given text information, especially in complex scenes." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23254" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 286, + 204 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 286, + 204 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 286, + 204 + ], + "type": "image", + "image_path": "e551c84949edbed60b486d5b3747aee997496a7a56e952bf7507dd7fbd43db19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 207, + 289, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 275 + ], + "type": "text", + "content": "Figure 3: Reconstructed images using " + }, + { + "bbox": [ + 46, + 207, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^{(\\mathrm{t})}" + }, + { + "bbox": [ + 46, + 207, + 289, + 275 + ], + "type": "text", + "content": " each step during diffusion using the revocable method with and without FAS. The backgrounds of the images in the top row are over-simplified while our proposed FAS prevents this, as shown in the bottom row. The text is \"This small bird is greyish in color with flecks of yellow on the back and breast, and a bit of white on the belly.\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 289, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 407 + ], + "type": "text", + "content": "To be specific, as shown in the top row of Figure 2, fixed strategies have limitations in recovering from errors or generating diverse images due to their inability to regenerate incorrect tokens or modify missing essential objects or attributes. In contrast, revocable strategies enhance the text alignment by providing an opportunity to correct invalidly generated tokens, as shown in the bottom row of Figure 2. Since revocable strategies have no constraints on which tokens to remove or fix, they have the advantage of mitigating error accumulation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 471 + ], + "type": "text", + "content": "However, random revoke sampling suffers from two main issues that must be addressed. The first issue is the instability resulting from randomly discarding previously sampled tokens which worked as a given condition of the step. It can lower the quality of the outputs as in Figure 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 288, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 664 + ], + "type": "text", + "content": "The second issue, namely the over-simplification, arises when it is used in the generation process involving longer steps. The low-frequency parts of the output images, such as backgrounds, tend to be over-simplified compared to fixed methods, which generate more realistic backgrounds. The over-simplified samples can be seen in Figure 3, and it has a significant impact on the FID metric as in Table 1 and 2. We observe that this is due to excessive repetition of resampling over a large number of steps, while a fixed schedule does not resample any tokens. More opportunities for resampling make the tokens converge towards safer and simpler patterns. As shown in the top row of Figure 3, the high-frequency areas with rich visual details do not become over-simplified even after multiple resampling processes, but the low-frequency areas become less diverse and more simplified in patterns quickly due to excessive resampling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": "To confirm this trade-off between text alignment and image quality, we define a persistent sampling as an interpolation between uniform sampling and random revoke sampling. In details, a persistent weight boosts the probability of keep" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 73, + 443, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 73, + 443, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 443, + 85 + ], + "type": "text", + "content": "Algorithm 1 Persistent Sampling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 87, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 87, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 306, + 87, + 545, + 108 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 306, + 87, + 545, + 108 + ], + "type": "inline_equation", + "content": "k_{t}" + }, + { + "bbox": [ + 306, + 87, + 545, + 108 + ], + "type": "text", + "content": ": number of tokens sampled at step " + }, + { + "bbox": [ + 306, + 87, + 545, + 108 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 306, + 87, + 545, + 108 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 87, + 545, + 108 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 306, + 87, + 545, + 108 + ], + "type": "text", + "content": ": persistent weight, " + }, + { + "bbox": [ + 306, + 87, + 545, + 108 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 87, + 545, + 108 + ], + "type": "text", + "content": ": number of all tokens" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 110, + 523, + 264 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 311, + 110, + 394, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 110, + 394, + 121 + ], + "spans": [ + { + "bbox": [ + 311, + 110, + 394, + 121 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 311, + 110, + 394, + 121 + ], + "type": "inline_equation", + "content": "x_{T}\\gets [[\\mathrm{MASK}]]_{N}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 122, + 418, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 122, + 418, + 132 + ], + "spans": [ + { + "bbox": [ + 311, + 122, + 418, + 132 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 311, + 122, + 418, + 132 + ], + "type": "inline_equation", + "content": "t = T, T - 1, \\dots, 1" + }, + { + "bbox": [ + 311, + 122, + 418, + 132 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 133, + 391, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 133, + 391, + 143 + ], + "spans": [ + { + "bbox": [ + 311, + 133, + 391, + 143 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 311, + 133, + 391, + 143 + ], + "type": "inline_equation", + "content": "\\hat{x}_0 = G_\\theta(x_t, c)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 144, + 420, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 144, + 420, + 155 + ], + "spans": [ + { + "bbox": [ + 311, + 144, + 420, + 155 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 311, + 144, + 420, + 155 + ], + "type": "inline_equation", + "content": "A_{t} = \\{i|x_{t}^{i} = [\\mathrm{MASK}]\\}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 155, + 433, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 155, + 433, + 164 + ], + "spans": [ + { + "bbox": [ + 311, + 155, + 433, + 164 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 311, + 155, + 433, + 164 + ], + "type": "inline_equation", + "content": "m = k_{T} + k_{T - 1} + \\ldots + k_{t}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 165, + 503, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 165, + 503, + 176 + ], + "spans": [ + { + "bbox": [ + 311, + 165, + 503, + 176 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 311, + 165, + 503, + 176 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_t(I = i|i\\in A_t):\\mathcal{U}_t(I = i|i\\in A_t^C) = 1:w" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 177, + 488, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 177, + 488, + 187 + ], + "spans": [ + { + "bbox": [ + 311, + 177, + 488, + 187 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 311, + 177, + 488, + 187 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 311, + 177, + 488, + 187 + ], + "type": "text", + "content": " uniform distribution + persistent weight" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 188, + 523, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 188, + 523, + 198 + ], + "spans": [ + { + "bbox": [ + 311, + 188, + 523, + 198 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 311, + 188, + 523, + 198 + ], + "type": "inline_equation", + "content": "i_1, i_2, \\dots, i_m \\sim \\mathcal{U}(I) \\gets" + }, + { + "bbox": [ + 311, + 188, + 523, + 198 + ], + "type": "text", + "content": " sample without replacement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 199, + 399, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 199, + 399, + 209 + ], + "spans": [ + { + "bbox": [ + 311, + 199, + 399, + 209 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 311, + 199, + 399, + 209 + ], + "type": "inline_equation", + "content": "x_{t}\\gets [[\\mathrm{MASK}]]_{N}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 419, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 419, + 220 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 419, + 220 + ], + "type": "text", + "content": "10: for " + }, + { + "bbox": [ + 308, + 209, + 419, + 220 + ], + "type": "inline_equation", + "content": "i = i_1, i_2, \\ldots, i_m" + }, + { + "bbox": [ + 308, + 209, + 419, + 220 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 309, + 220, + 374, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 220, + 374, + 231 + ], + "spans": [ + { + "bbox": [ + 309, + 220, + 374, + 231 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 309, + 220, + 374, + 231 + ], + "type": "inline_equation", + "content": "x_{t}^{i}\\gets \\hat{x}_{0}^{i}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 309, + 232, + 361, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 232, + 361, + 241 + ], + "spans": [ + { + "bbox": [ + 309, + 232, + 361, + 241 + ], + "type": "text", + "content": "12: end for" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 309, + 243, + 351, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 243, + 351, + 252 + ], + "spans": [ + { + "bbox": [ + 309, + 243, + 351, + 252 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 253, + 429, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 429, + 264 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 429, + 264 + ], + "type": "text", + "content": "14: Return: Generated image " + }, + { + "bbox": [ + 308, + 253, + 429, + 264 + ], + "type": "inline_equation", + "content": "x_0" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 283, + 547, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 547, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 547, + 439 + ], + "type": "text", + "content": "ing the tokens that were sampled in the previous step. While random revoke sampling assigns the same probability to all the locations, persistent sampling multiplies the persistent weight to the probability of previously kept locations. Further details are shown in Algorithm 1. If the persistent weight " + }, + { + "bbox": [ + 304, + 283, + 547, + 439 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 283, + 547, + 439 + ], + "type": "text", + "content": " is 1, all the tokens would have the same probability to be fixed, such as random revoke sampling. If " + }, + { + "bbox": [ + 304, + 283, + 547, + 439 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 283, + 547, + 439 + ], + "type": "text", + "content": " is sufficiently large, the previously sampled tokens would be kept until the end, such as uniform sampling. In the graph in Figure 1, a trade-off can be seen that as the sampling strategy gets closer to the random revoke sampling, text alignment gets better, while the image quality gets worse. We further analyze the effect of these sampling methods in Section 4.5." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 449, + 479, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 479, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 479, + 460 + ], + "type": "text", + "content": "3.2. Text-conditioned token selection" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 468, + 547, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 547, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 547, + 648 + ], + "type": "text", + "content": "We propose a learnable model called Text-Conditioned Token Selection (TCTS). TCTS is trained to output a score map to detect tokens that are not under the real distribution for the given text condition, and selects the tokens that are not aligned with the text or other tokens in the image. By masking the selected tokens at each step, the diffusion network receives high-quality text-aligned tokens as the input to sample the next set of tokens which alleviates the error accumulation and joint distribution issue of the previous sampling strategies. Moreover, since TCTS is trained to discriminate the well-aligned tokens, it detours the instability problem as mentioned earlier, unlike other revocable methods. From the experimental results in Section 4, we observe that TCTS enhances the text alignment without compromising the image quality." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Frequency adaptive sampling (FAS) The simplest way to solve the over-simplification problem in longer steps is to use the persistent sampling to alleviate the excessive resampling of tokens. However, as previously mentioned, this method" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23255" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 545, + 216 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 545, + 216 + ], + "type": "image", + "image_path": "98df41975cbefd50a342c61b36778d24047b963351f2b819e3512fda5bcd092f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "type": "text", + "content": "Figure 4: Overall generation framework of proposed TCTS and FAS. After the generator predicts the tokens, TCTS exploits the text condition to detect misaligned tokens and outputs the score map. Meanwhile, FAS splits the tokens according to the frequency using the self-attention map from the generator, performing revocable sampling to high-frequency split and persistent sampling to low-frequency split. The adaptive sampling predicts " + }, + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "type": "text", + "content": " and decide a few of the locations to mask according to " + }, + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "type": "text", + "content": ". The token maps produced by FAS is combined with the score map to predict " + }, + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "type": "text", + "content": ". After iterative process, our model removes all the masks and generates " + }, + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 221, + 548, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 296, + 288, + 417 + ], + "blocks": [ + { + "bbox": [ + 53, + 296, + 288, + 417 + ], + "lines": [ + { + "bbox": [ + 53, + 296, + 288, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 288, + 417 + ], + "type": "table", + "html": "
StepMethodMID-L ↑SOA-I ↑CLIP-S ↑FID-30K ↓
16Purity11.0272.380.247419.20
Uniform17.9474.800.250016.17
RR23.6078.790.252617.10
TCTS + FAS26.7279.520.255914.45
25Purity16.8475.210.248718.39
Uniform22.2777.080.252415.91
RR26.7781.100.254318.43
TCTS + FAS27.7980.870.256315.39
", + "image_path": "d73635d14933b74865194f404f2123ba3c2f5282a570d3e37ac5bd8cdb3906b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 482, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 542 + ], + "type": "text", + "content": "may result in a slight compromise in text alignment and limit the opportunity to correct the image through resampling. To further improve this trade-off, we propose a new method called Frequency Adaptive Sampling (FAS), which can be applied to TCTS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": "FAS is a method that utilizes the generator's self-attention map to limit resampling only in the low-frequency areas of the image. In [21], it is known that the generator's self-attention layer contains frequency information of the image, which is also observed in token-based diffusion models. We utilized this to distinguish the frequency areas of the image without additional operations, and applied persistent sampling only to the low-frequency areas using persistent weight. As shown in the bottom row of Figure 3, revocable sampling along with FAS method allows repetitive resampling in areas that require rich visual details, while limiting the number of resampling in relatively simple areas preventing oversimplification. Detailed algorithm is in the supplementary A. The overall framework of our model, including both TCTS" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 313, + 296, + 545, + 416 + ], + "blocks": [ + { + "bbox": [ + 46, + 420, + 288, + 464 + ], + "lines": [ + { + "bbox": [ + 46, + 420, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 288, + 464 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation of sampling methods on MS-COCO dataset. The ground truth MID on MS-COCO image-caption pairs is 54.73. The classifier-free guidance scale was fixed at 5 for all sampling methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 296, + 545, + 416 + ], + "lines": [ + { + "bbox": [ + 313, + 296, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 545, + 416 + ], + "type": "table", + "html": "
StepMethodMID-L ↑CLIP-S ↑FID ↓
16Purity-24.210.241015.21
Uniform-25.600.240416.57
RR-25.030.237117.38
TCTS + FAS-19.880.24612.35
25Purity-21.260.238412.60
Uniform-23.040.239613.02
RR-23.290.236414.53
TCTS + FAS-18.310.240913.67
", + "image_path": "0091a053f9dc2711e2f3c3ad653082bcc459ebde48cae130c51d4b70ffb206c6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 419, + 547, + 464 + ], + "lines": [ + { + "bbox": [ + 305, + 419, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 419, + 547, + 464 + ], + "type": "text", + "content": "Table 2: Quantitative evaluation of sampling methods on CUB dataset. The ground truth MID on CUB image-caption pairs is 15.85. We omit SOA in this table as CUB dataset consists of only one bird per photo, diminishing the metric's significance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 479, + 446, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 446, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 446, + 491 + ], + "type": "text", + "content": "and FAS, is illustrated in Figure 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 502, + 388, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 388, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 388, + 515 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 713 + ], + "type": "text", + "content": "To validate the efficacy of our token selection framework, we modify the transformer from [37] to contain 310M parameters for MS-COCO dataset and 83M parameters for CUB dataset. We extract text features from CLIP ViT-B/32 [27] for a text-conditioned generation. In the training process of TCTS, we freeze the generator's parameters and independently train our model, making it applicable to other types of token-based diffusion models. We use the binary cross-entropy loss for the objective and validate our method and baselines on MS-COCO [25] and CUB [14] datasets by training them for 200 epochs. We leverage classifier-free guidance [20] by stochastic sampling of the guidance strength to improve the generation quality. We further provide the details of the architecture, hyperparameters and additional analysis on the use of classifier-free guidance in the supplementary A." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23256" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 54, + 70, + 288, + 173 + ], + "blocks": [ + { + "bbox": [ + 54, + 70, + 288, + 173 + ], + "lines": [ + { + "bbox": [ + 54, + 70, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 54, + 70, + 288, + 173 + ], + "type": "table", + "html": "
ModelMID-L ↑MID-B ↑SOA-I ↑FID-30K ↓
GLIDE [26]1.031.00-32.08
AttnGAN [40]-65.20-8.9039.0129.15
DM-GAN [23]-44.663.5148.0322.90
DF-GAN [38]-58.75-15.21-31.75
VQ-Diffusion [41]-19.635.77-13.13
LAFITE [47]6.2635.1774.788.03
TCTS + FAS26.0238.9879.529.75
", + "image_path": "a7f4c907cdc927124b284e3879e4e5d45c07bdf909b28cd927867df0033c9f8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 290, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 422 + ], + "type": "text", + "content": "Metrics Since FID [16] metric is known to be problematic for its inability to exactly represent the image quality [5] and unable to consider the text alignment, we additionally use Mutual Information Divergence (MID) [22] to evaluate the text-alignment of generated images which enables sample-wise evaluation and outperforms previous metrics in human Likert-scale judgment correlations. In particular, since MID responds more sensitively to images generated with foiled captions [35], it is more appropriate for analyzing text alignment with complex and challenging captions. Moreover, we use the CLIP score [15] and SOA-I [17], which are widely used in text-to-image synthesis. Note that the CLIP score is calculated with ViT-L/14 throughout the experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 431, + 180, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 180, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 180, + 443 + ], + "type": "text", + "content": "4.1. Text-to-image synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 449, + 288, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 675 + ], + "type": "text", + "content": "Figure 1 Left and Middle visualizes the generated images using improved VQ diffusion [37] with varying sampling methods. Generated examples from uniform token sampling are poorly aligned with the given texts, and often include erratic partitions of the objects since the predicted tokens selected at the same step cannot appropriately reflect the change of others (e.g., smashed donuts, and imperfect chair and seagull). Random revoke (RR) sampling, which can re-update previously sampled tokens, seems to improve alignment with text by editing out nonsensical regions and iteratively considering the entire scene at each generation step. Yet, RR selects tokens in a completely random manner without any constraints, which results in suboptimal generation quality and alignment with the texts (e.g., disappeared man and train). On the other hand, ours mitigates the problem from RR by selecting text-conditioned tokens at the sampling phase and successfully generates high-quality images which contain a clear semantic connection to the given text captions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "This phenomenon can be quantitatively seen in Table 1. 16-step generation of random revoke sampling shows better text alignment than even 25-step generation of other fixed" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 544, + 149 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 544, + 149 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 149 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 149 + ], + "type": "image", + "image_path": "2d6ff5a6f95806e95523125a66d69388a363a7848df400f3a1e5963933a2d8c6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 152, + 544, + 231 + ], + "blocks": [ + { + "bbox": [ + 308, + 152, + 544, + 231 + ], + "lines": [ + { + "bbox": [ + 308, + 152, + 544, + 231 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 544, + 231 + ], + "type": "image", + "image_path": "c787d8bd112bda7abb8468484b3884afd8c9c3321e9dd3a244d7aa6f3f6d84b2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 232, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 306, + 232, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 547, + 255 + ], + "type": "text", + "content": "Figure 5: Results of image refinement using TCTS. Top: Original samples, Bottom: Refined images for 8-steps with TCTS." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 347, + 256, + 509, + 350 + ], + "blocks": [ + { + "bbox": [ + 46, + 178, + 288, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 288, + 245 + ], + "type": "text", + "content": "Table 3: Comparison of the proposed method with recent text-to-image generative models on MS-COCO dataset. MID-L and MID-B are calculated with ViT-L/14 and ViT-B/32 each. The ground truth MID-L and MID-B on MS-COCO image-caption pairs is 54.73, 41.57 each. Ours are evaluated with 16-step setting and we adopt classifier-free guidance from 3 to 5." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 256, + 509, + 350 + ], + "lines": [ + { + "bbox": [ + 347, + 256, + 509, + 350 + ], + "spans": [ + { + "bbox": [ + 347, + 256, + 509, + 350 + ], + "type": "table", + "html": "
MethodMID-L ↑FID ↓
Original12.1016.44
RR13.6417.18
TCTS15.6416.92
Original4.5419.13
Random masked6.0619.12
TCTS masked6.4718.64
", + "image_path": "0b9240200803624f4c6e49ba4914a286f0961dcc6999b3ff3baa62fe08184d6e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 352, + 547, + 385 + ], + "lines": [ + { + "bbox": [ + 305, + 352, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 305, + 352, + 547, + 385 + ], + "type": "text", + "content": "Table 4: Qualitative evaluation of the refined images. Top: Refinement with additional revision steps. Bottom: Refinement with masking lowest-scoring tokens." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 397, + 547, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 547, + 625 + ], + "type": "text", + "content": "sampling methods. We can also observe the trade-off between CLIP score and FID by comparing each method to their 25-step versions. Our model outperforms other baselines in most of the metrics especially when comparing MID-L to fixed sampling methods such as Purity and Uniform. Since our model can make corrections during the generation process, it requires fewer steps to match the performance of our baseline sampling methods, see Section 4.5 for further analysis. We also present the results of our experiment on the CUB dataset with fewer parameters in Table 2, which demonstrates satisfactory performance. However, the fact that the CUB dataset contains only single object per image resulted in a tendency that is slightly different from what we mentioned earlier. In Table 3, we compared the performance of our method with other text-to-image generative models. Our model highly exceeds other models in text alignment metrics, especially in MID-L and closely approximate the MID-B value of the ground truth images. More samples generated by our model is in the supplementary B." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 634, + 411, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 411, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 411, + 647 + ], + "type": "text", + "content": "4.2. Image refinement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 547, + 715 + ], + "type": "text", + "content": "The advantage of the masked image generative model is that it enables fast local refinement. We conduct image refinement in two separate methods, inspired by [23] and [24]. First, we use TCTS to apply additional revision steps to images generated with uniform sampling. The overall image" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23257" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "quality is improved by adding additional refinement steps as shown in Figure 5. In order to demonstrate the image refinement performance of our TCTS, we also measure the performance of the image refinement with random revoke sampling. While RR improves does improve sample quality, additional refinement steps increase the FID score, caused by a similar effect to increasing the generation step." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 290, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 290, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 290, + 253 + ], + "type": "text", + "content": "We additionally perform the experiment by masking " + }, + { + "bbox": [ + 46, + 156, + 290, + 253 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 156, + 290, + 253 + ], + "type": "text", + "content": " lowest-scoring tokens with TCTS, and generate the images with uniform sampling. We observe that all metrics of the refined images outperform those of the original images. In Table 4, we compare them with samples refined after randomly masking the tokens without TCTS. Due to the page limit, we describe further details of the two refinement methods and visualize more samples in the supplementary B." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 261, + 185, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 261, + 185, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 261, + 185, + 274 + ], + "type": "text", + "content": "4.3. Mask-free object editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 279, + 289, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 289, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 289, + 411 + ], + "type": "text", + "content": "Since the masked image generative model is capable of local refinement, image editing without manual masking is possible simply by randomly masking a part of the image and resampling it with the new text condition. However, this method requires a low masking ratio and many resampling steps to maintain the overall structure of the image. This can result in significant changes to unnecessary parts or, as mentioned earlier, over-simplification issues. Additionally, even with many steps, editing large objects with small masks can be more challenging due to the significant influence of surrounding tokens rather than the new text condition." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 290, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 290, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 290, + 579 + ], + "type": "text", + "content": "Motivated by the operation of self-attention maps in frequency adaptive sampling, we leverage a cross-attention map corresponding to the word of the object that is to be changed, giving weights to resample tokens so that the corresponding locations can be resampled. Then, it enables efficient image editing with fewer steps and makes it easier to edit larger objects. Although it is similar to DiffEdit [7], ours operates in a more straightforward manner without additional computations, thanks to the capability of local refinement of the masked generative model. In Figure 6, we edit the object by adding " + }, + { + "bbox": [ + 46, + 411, + 290, + 579 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 411, + 290, + 579 + ], + "type": "text", + "content": " noise in only 10-steps. With our method, we can edit more quickly with fewer editing steps, which further does not suffer from the over-simplification problem. More samples are in the supplementary B." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 587, + 220, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 220, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 220, + 601 + ], + "type": "text", + "content": "4.4. High-resolution image synthesis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "Continuous diffusion models that use encoders and decoders, such as masked image generative models, can generate larger images that are not in the training set. Similar to Bond-Taylor et al. [3], we divide the tokens into subsets according to the model input size, individually pass them through the model, and then spatially aggregate tokens to synthesize a high-resolution image. In this way, we are able to generate more realistic and high-resolution samples with the same TCTS model without additional training. Addition-" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 71, + 545, + 159 + ], + "blocks": [ + { + "bbox": [ + 308, + 71, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 545, + 159 + ], + "type": "image", + "image_path": "00548aacf89f2c11b8b5c60d676ea59b691698d44e43eef4bb1a7e34b5ee3b55.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 160, + 541, + 171 + ], + "lines": [ + { + "bbox": [ + 311, + 160, + 541, + 171 + ], + "spans": [ + { + "bbox": [ + 311, + 160, + 541, + 171 + ], + "type": "text", + "content": "A dog teddy bear looks back from the passenger seat of a vehicle." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 172, + 545, + 260 + ], + "blocks": [ + { + "bbox": [ + 308, + 172, + 545, + 260 + ], + "lines": [ + { + "bbox": [ + 308, + 172, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 172, + 545, + 260 + ], + "type": "image", + "image_path": "7d3582282c7d74b2a290ae6b339220c34417c771de7e168d4e7236fc6552a4a3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 261, + 526, + 270 + ], + "lines": [ + { + "bbox": [ + 324, + 261, + 526, + 270 + ], + "spans": [ + { + "bbox": [ + 324, + 261, + 526, + 270 + ], + "type": "text", + "content": "A wooden table topped with earrot cake donuts, broccoli." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 323, + 425, + 441 + ], + "blocks": [ + { + "bbox": [ + 305, + 278, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 305, + 278, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 278, + 547, + 312 + ], + "type": "text", + "content": "Figure 6: Examples of mask-free editing samples with cross-attention map. The cross-attention map is multiplied to the score map of TCTS to perform weighted sampling." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 323, + 425, + 441 + ], + "lines": [ + { + "bbox": [ + 308, + 323, + 425, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 323, + 425, + 441 + ], + "type": "image", + "image_path": "1739847b09f6a36a2b9c07eea36e656d0371110d99d7c27ce027fe26809ffac8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 442, + 421, + 463 + ], + "lines": [ + { + "bbox": [ + 310, + 442, + 421, + 463 + ], + "spans": [ + { + "bbox": [ + 310, + 442, + 421, + 463 + ], + "type": "text", + "content": "A short train traveling through a mountainous landscape." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 470, + 547, + 505 + ], + "lines": [ + { + "bbox": [ + 305, + 470, + 547, + 505 + ], + "spans": [ + { + "bbox": [ + 305, + 470, + 547, + 505 + ], + "type": "text", + "content": "Figure 7: Examples of high resolution " + }, + { + "bbox": [ + 305, + 470, + 547, + 505 + ], + "type": "inline_equation", + "content": "(512^{2})" + }, + { + "bbox": [ + 305, + 470, + 547, + 505 + ], + "type": "text", + "content": " samples. Left: Image generated with all mask tokens, Right: Image generated with TCTS as a super-resolution unit." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 323, + 545, + 441 + ], + "blocks": [ + { + "bbox": [ + 427, + 323, + 545, + 441 + ], + "lines": [ + { + "bbox": [ + 427, + 323, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 427, + 323, + 545, + 441 + ], + "type": "image", + "image_path": "fdbd26e656252abb8373c775bf4bdaf072318cefcd9143841c268fcc99d65d1a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 442, + 531, + 462 + ], + "lines": [ + { + "bbox": [ + 440, + 442, + 531, + 462 + ], + "spans": [ + { + "bbox": [ + 440, + 442, + 531, + 462 + ], + "type": "text", + "content": "A family on a boat sailing down a river." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 518, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 626 + ], + "type": "text", + "content": "ally, we propose a new method to generate high-resolution images only with low-resolution TCTS. To be specific, we first generate a small-size image with TCTS and upsample the token map in bicubic mode to the desired size. Then we divide the high-resolution token map into overlapping small-size sections and refine all the sections several times. Then, we can generate high-resolution images with the low-resolution TCTS where we visualize the generated samples in Figure 7." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 635, + 406, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 406, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 406, + 647 + ], + "type": "text", + "content": "4.5. Further analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": "Importance of early-stage sampling strategy As mentioned in Section 3.1, by controlling the persistent weight, we can interpolate the two baseline sampling methods: uniform sampling and random revoke sampling. We designed various experiments to find out the impact of the sampling" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23258" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 108, + 131 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 108, + 131 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 108, + 131 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 108, + 131 + ], + "type": "image", + "image_path": "705f6d7abccccd96d76b6fe280dd3a88e26197ec30e98f22b99bb45a34e8752c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 133, + 93, + 141 + ], + "lines": [ + { + "bbox": [ + 66, + 133, + 93, + 141 + ], + "spans": [ + { + "bbox": [ + 66, + 133, + 93, + 141 + ], + "type": "text", + "content": "(a) Real" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 71, + 227, + 131 + ], + "blocks": [ + { + "bbox": [ + 108, + 71, + 227, + 131 + ], + "lines": [ + { + "bbox": [ + 108, + 71, + 227, + 131 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 227, + 131 + ], + "type": "image", + "image_path": "32adead52fac8ee739171cd1644ff8159fe44b54bd03b206702e9c3a6d6c20a7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 227, + 72, + 286, + 131 + ], + "blocks": [ + { + "bbox": [ + 227, + 72, + 286, + 131 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 286, + 131 + ], + "type": "image", + "image_path": "cae83e708f754a98ac1338a6c0d5f3cd96f6d9ae92e93c5b7ecae2d2047c8302.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 133, + 265, + 140 + ], + "lines": [ + { + "bbox": [ + 248, + 133, + 265, + 140 + ], + "spans": [ + { + "bbox": [ + 248, + 133, + 265, + 140 + ], + "type": "text", + "content": "road" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 49, + 144, + 108, + 205 + ], + "blocks": [ + { + "bbox": [ + 49, + 144, + 108, + 205 + ], + "lines": [ + { + "bbox": [ + 49, + 144, + 108, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 144, + 108, + 205 + ], + "type": "image", + "image_path": "9420161652a97a82d01ca9af46df72d2602ede70dac5d3351411f6d9736ecc57.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 206, + 105, + 215 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 105, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 105, + 215 + ], + "type": "text", + "content": "(b) Synthesized" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 108, + 144, + 167, + 204 + ], + "blocks": [ + { + "bbox": [ + 108, + 144, + 167, + 204 + ], + "lines": [ + { + "bbox": [ + 108, + 144, + 167, + 204 + ], + "spans": [ + { + "bbox": [ + 108, + 144, + 167, + 204 + ], + "type": "image", + "image_path": "91ee9af1d6669cd7d757c429b30c13a9c71f772362e97f26955cabed26a68855.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 206, + 141, + 213 + ], + "lines": [ + { + "bbox": [ + 135, + 206, + 141, + 213 + ], + "spans": [ + { + "bbox": [ + 135, + 206, + 141, + 213 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 167, + 144, + 226, + 204 + ], + "blocks": [ + { + "bbox": [ + 167, + 144, + 226, + 204 + ], + "lines": [ + { + "bbox": [ + 167, + 144, + 226, + 204 + ], + "spans": [ + { + "bbox": [ + 167, + 144, + 226, + 204 + ], + "type": "image", + "image_path": "491864ab18b4ebd38582b792720a487d9230f20ab8522995f53bc343dbe67136.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 206, + 205, + 213 + ], + "lines": [ + { + "bbox": [ + 188, + 206, + 205, + 213 + ], + "spans": [ + { + "bbox": [ + 188, + 206, + 205, + 213 + ], + "type": "text", + "content": "boat" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 227, + 144, + 286, + 205 + ], + "blocks": [ + { + "bbox": [ + 227, + 144, + 286, + 205 + ], + "lines": [ + { + "bbox": [ + 227, + 144, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 227, + 144, + 286, + 205 + ], + "type": "image", + "image_path": "b5ec268ce93fb9b89c4f4eb02f4b0f168cf3f26a38b6c0a8fc705989ee7d4069.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 207, + 267, + 214 + ], + "lines": [ + { + "bbox": [ + 247, + 207, + 267, + 214 + ], + "spans": [ + { + "bbox": [ + 247, + 207, + 267, + 214 + ], + "type": "text", + "content": "water" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 220, + 287, + 264 + ], + "lines": [ + { + "bbox": [ + 47, + 220, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 287, + 264 + ], + "type": "text", + "content": "Figure 8: Visualization of the averaged cross-attention maps for each word from TCTS. Top: Image from COCO dataset: \"A large black bear walking down a dirty road\". Bottom: Synthesized image: \"A small boat in a narrow body of water\"." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 278, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 287, + 338 + ], + "type": "text", + "content": "method with respect to the time steps. We switched the sampling methods during the time steps, from uniform sampling to random revoke sampling (U2R) and vice versa (R2U). In the rightmost plot in Figure 1, we can observe that U2R shows better text-alignment performance than R2U." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 339, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 542 + ], + "type": "text", + "content": "These results suggest that the early stage sampling, in which the masked ratio is high, substantially influences the text alignment of the final image, which is similarly observed in continuous diffusion models [2]. Generating the whole image and then processing it through a few revision steps like [23] might help the image quality but can not conspicuously enhance the essential text alignment. Therefore, in order to obtain an image that aligns well with the text, sampling must be done carefully in the early stage when not many tokens are generated. This analysis is also shown in the result that the revocable strategy exhibits better performance by aggressively resampling tokens several times. In particular, since weak generators with poor generation performance suffer more from the joint distribution issues, giving a sufficient number of recovery chances to discard tokens and draw new ones at the early stage is the key to the desired text alignment." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": "Text-conditioned sampling The experiments show that TCTS plays a crucial role in enhancing text alignment in image generation. Although we did not utilize any text-specific loss during training, the model still effectively processes the text condition and achieves optimal performance. To verify the accuracy of TCTS's attention towards text-related parts of the image, we visualized the average cross-attention layer for each word in Figure 8. The results demonstrate TCTS's ability to attend to each word in both real and synthesized images. By helping the selection of relevant tokens during the generating stage, TCTS can boost model performance without the requirement of additional complex text-related loss." + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 313, + 70, + 545, + 166 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 545, + 166 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 545, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 545, + 166 + ], + "type": "table", + "html": "
MethodMID-L ↑SOA-I ↑CLIP-S ↑FID-30K ↓
RR26.7781.100.254318.43
RR + FAS26.3280.780.253916.01
RR + pw24.2379.140.253415.96
TCTS27.8280.930.256516.69
TCTS + FAS27.7980.870.256315.39
TCTS + pw25.1279.510.255615.12
", + "image_path": "8fcf5099d5ce038907d8c5b892883ab847a2f0df6eaeae68e0daea6c029ea4f3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 313, + 220, + 545, + 280 + ], + "blocks": [ + { + "bbox": [ + 305, + 168, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 305, + 168, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 305, + 168, + 547, + 214 + ], + "type": "text", + "content": "Table 5: Ablation study of using FAS and persistent weight in revocable methods on MS-COCO dataset. All models were evaluated on 25-step setting. The ground truth MID-L on MS-COCO image-caption pairs is 54.73." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 220, + 545, + 280 + ], + "lines": [ + { + "bbox": [ + 313, + 220, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 545, + 280 + ], + "type": "table", + "html": "
StepModelMID-L ↑FID-30K ↓Time
50Uniform24.0416.84×1
RR26.5521.12
16TCTS + FAS26.7214.45×0.45
", + "image_path": "3bcd286f051fc06beafbda80a3865908ca6b158230fff92cf5038cebbdff1c9a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "lines": [ + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "text", + "content": "Table 6: Inference time relative to MID/FID of our baselines and models on MS-COCO dataset. We expressed the inference time of our models as a multiple of other baselines. Our model outspeeds other baselines while performing better in both MID-L and FID-30K. Further analysis is in the supplementary C." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 355, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 547, + 582 + ], + "type": "text", + "content": "Ablation study of FAS As mentioned in Section 3.2, FAS decides whether to apply persistent weight or not to each location. If the whole image is considered low-frequency, FAS would apply the weight to every location, which is the same as the persistent sampling. On the other hand, if the whole image is considered high-frequency, combining FAS would not apply the weight anywhere, which does not change the sampling strategy. To further analyze the effect of FAS, we evaluated it on two revocable methods: random revoke sampling and TCTS. Attaching FAS to a base sampling method can be regarded as a mixture of two methods: the base revocable sampling method and persistent sampling method. Table 5 clearly shows this relation of FAS and persistent weight. While FAS-attached methods do not show the best performance, the performance of them is always in close proximity to the better one of the two comparing methods of each. FAS-attached methods balances well between text-alignment and image quality without any additional training." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "Inference time Our TCTS and FAS compute additional operations at each generation step, resulting in marginally increased inference time compared to other token selection baselines. However, regarding the text alignment and the image quality, the relative inference time is considerably decreased. In Table 6, our model outperforms 50-step image generation of baseline sampling methods within only 16 steps in MID-L and FID-30K, which means that our model can synthesize better samples with only " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\times 0.45" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": " time." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23259" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 119, + 83 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 283 + ], + "type": "text", + "content": "This paper examines which factors in the masked diffusion process impact the output images and cause the trade-off between image quality and text alignment. We empirically find that the revocable sampling significantly improves the text alignment yet degrades the quality of the generated images. To tackle the problem, we propose a simple token sampling strategy, coined text-conditioned token selection, that combines learnable and revocable methods, pushing the boundary of the trade-off between image quality and text alignment without the necessity to update the large-scale pre-trained generator. We find that collaborative sampling in a persistent and revocable manner surprisingly alleviates over-simplification issues in the generated backgrounds. Our proposed method can be utilized as an image refinement tool for various generative tasks, which is remarkably fast to generate high-quality images within much fewer steps." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 295, + 162, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 295, + 162, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 162, + 309 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 289, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 411 + ], + "type": "text", + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST)) and (No. RS-2022-00187238, Development of Large Korean Language Model Technology for Efficient Pre-training). This work was supported by KAIST-NAVER Hypercreative AI Center." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 423, + 107, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 107, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 107, + 436 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 444, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 52, + 444, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 444, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 52, + 444, + 288, + 498 + ], + "type": "text", + "content": "[1] Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. In Advances in Neural Information Processing Systems, pages 17981-17993. Curran Associates, Inc., 2021. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 501, + 289, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 501, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 501, + 289, + 555 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 557, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 288, + 634 + ], + "type": "text", + "content": "[3] Sam Bond-Taylor, Peter Hessey, Hiroshi Sasaki, Toby P Breckon, and Chris G Willcocks. Unleashing transformers: parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vector-quantized codes. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXIII, pages 170–188. Springer, 2022. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 288, + 678 + ], + "type": "text", + "content": "[4] Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdzal, and Adriana Romero Soriano. Instance-conditioned gan. Advances in Neural Information Processing Systems, 34:27517-27529, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 288, + 713 + ], + "type": "text", + "content": "[5] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 97, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 97, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 97, + 547, + 150 + ], + "type": "text", + "content": "[6] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 545, + 185 + ], + "type": "text", + "content": "[7] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance, 2022. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 186, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 186, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 312, + 186, + 545, + 219 + ], + "type": "text", + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 220, + 547, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 220, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 312, + 220, + 547, + 264 + ], + "type": "text", + "content": "[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 266, + 545, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 266, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 545, + 309 + ], + "type": "text", + "content": "[10] Wan-Cyuan Fan, Yen-Chun Chen, DongDong Chen, Yu Cheng, Lu Yuan, and Yu-Chiang Frank Wang. Frido: Feature pyramid diffusion for complex scene image synthesis. arXiv preprint arXiv:2208.13753, 2022. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 311, + 547, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 311, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 311, + 547, + 376 + ], + "type": "text", + "content": "[11] Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. Tell me what happened: Unifying text-guided video completion via multimodal masked video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10681-10692, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 378, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 378, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 378, + 545, + 421 + ], + "type": "text", + "content": "[12] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 422, + 547, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 547, + 477 + ], + "type": "text", + "content": "[13] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 478, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 511 + ], + "type": "text", + "content": "[14] Xiangteng He and Yuxin Peng. Fine-grained visual-textual representation learning. IEEE Transactions on Circuits and Systems for Video Technology, 30(2):520-531, 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 513, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 547, + 578 + ], + "type": "text", + "content": "[15] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528. Association for Computational Linguistics, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 580, + 547, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 580, + 547, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 580, + 547, + 633 + ], + "type": "text", + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "text", + "content": "[17] Tobias Hinz, Stefan Heinrich, and Stefan Wermter. Semantic object accuracy for generative text-to-image synthesis. IEEE transactions on pattern analysis and machine intelligence, 2020. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 680, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 547, + 713 + ], + "type": "text", + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1, 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23260" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 116 + ], + "type": "text", + "content": "[19] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:47-1, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 140 + ], + "type": "text", + "content": "[20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 288, + 174 + ], + "type": "text", + "content": "[21] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 175, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 288, + 218 + ], + "type": "text", + "content": "[22] Jin-Hwa Kim, Yunji Kim, Jiyoung Lee, Kang Min Yoo, and Sang-Woo Lee. Mutual information divergence: A unified metric for multimodal generative models. arXiv preprint arXiv:2205.13445, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 288, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 288, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 288, + 263 + ], + "type": "text", + "content": "[23] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Draft-and-revise: Effective image generation with contextual rq-transformer. arXiv preprint arXiv:2206.04452, 2022. 1, 3, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 265, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 265, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 265, + 288, + 308 + ], + "type": "text", + "content": "[24] José Lezama, Huiwen Chang, Lu Jiang, and Irfan Essa. Improved masked image generation with token-critic. In European Conference on Computer Vision, pages 70-86. Springer, 2022. 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 310, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 310, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 49, + 310, + 288, + 365 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 366, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 366, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 49, + 366, + 288, + 421 + ], + "type": "text", + "content": "[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 422, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 422, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 49, + 422, + 288, + 487 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 489, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 489, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 489, + 288, + 521 + ], + "type": "text", + "content": "[28] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 523, + 288, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 288, + 566 + ], + "type": "text", + "content": "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 568, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 568, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 49, + 568, + 288, + 621 + ], + "type": "text", + "content": "[30] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 624, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 288, + 657 + ], + "type": "text", + "content": "[31] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 288, + 713 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 137 + ], + "type": "text", + "content": "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 139, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 139, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 139, + 547, + 172 + ], + "type": "text", + "content": "[34] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 173, + 547, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 547, + 216 + ], + "type": "text", + "content": "[35] Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich, Aurélie Herbelot, Moin Nabi, Enver Sangineto, and Raffaella Bernardi. Foil it! find one mismatch between image and language caption. arXiv preprint arXiv:1705.01359, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 547, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 547, + 259 + ], + "type": "text", + "content": "[36] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 547, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 547, + 293 + ], + "type": "text", + "content": "[37] Zhicong Tang, Shuyang Gu, Jianmin Bao, Chen Dong, and Fang Wen. Improved vector quantized diffusion models. arXiv preprint arXiv:2205.16007, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 294, + 547, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 547, + 338 + ], + "type": "text", + "content": "[38] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2020. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 547, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 547, + 381 + ], + "type": "text", + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 547, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 547, + 447 + ], + "type": "text", + "content": "[40] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 449, + 547, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 547, + 502 + ], + "type": "text", + "content": "[41] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 1, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 504, + 547, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 547, + 557 + ], + "type": "text", + "content": "[42] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 547, + 624 + ], + "type": "text", + "content": "[43] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 625, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 547, + 689 + ], + "type": "text", + "content": "[44] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "text", + "content": "[45] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stack" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23261" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 262 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "gan++: Realistic image synthesis with stacked generative adversarial networks. IEEE transactions on pattern analysis and machine intelligence, 41(8):1947-1962, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 140 + ], + "type": "text", + "content": "[46] Yufan Zhou, Chunyuan Li, Changyou Chen, Jianfeng Gao, and Jinhui Xu. Lafite2: Few-shot text-to-image generation. arXiv preprint arXiv:2210.14124, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 205 + ], + "type": "text", + "content": "[47] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17907-17917, 2022. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 262 + ], + "type": "text", + "content": "[48] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dmgan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5810, 2019. 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23262" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/e13295e4-a033-4d38-a3a5-88f3b3385e81_content_list.json b/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/e13295e4-a033-4d38-a3a5-88f3b3385e81_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..62780e1c641bbb90c7f00c16d71a7d3347b08cbb --- /dev/null +++ b/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/e13295e4-a033-4d38-a3a5-88f3b3385e81_content_list.json @@ -0,0 +1,1624 @@ +[ + { + "type": "text", + "text": "Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization", + "text_level": 1, + "bbox": [ + 129, + 128, + 841, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenhuan Liu $^{1,2}$ Liang Li $^{1*}$ Jiayu Xiao $^{1,2}$ Zheng-Jun Zha $^{3}$ Qingming Huang $^{1,2,4}$ $^{1}$ Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, CAS \n $^{2}$ University of Chinese Academy of Sciences \n $^{3}$ University of Science and Technology of China \n $^{4}$ Peng Cheng Laboratory", + "bbox": [ + 98, + 200, + 869, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhenhuan.liu, jiayu.xiao}@vip1.ict.ac.cn, liang.li@ict.ac.cn, zhazj@ustc.edu.cn, qmhuang@ucas.ac.cn", + "bbox": [ + 210, + 294, + 754, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 361, + 313, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Combined with the generative prior of pre-trained models and the flexibility of text, text-driven generative domain adaptation can generate images from a wide range of target domains. However, current methods still suffer from overfitting and the mode collapse problem. In this paper, we analyze the mode collapse from the geometric point of view and reveal its relationship to the Hessian matrix of generator. To alleviate it, we propose the spectral consistency regularization to preserve the diversity of source domain without restricting the semantic adaptation to target domain. We also design granularity adaptive regularization to flexibly control the balance between diversity and stylization for target model. We conduct experiments for broad target domains compared with state-of-the-art methods and extensive ablation studies. The experiments demonstrate the effectiveness of our method to preserve the diversity of source domain and generate high fidelity target images. Source code has been released in https://github.com/Victarry/Adaptation-SCR.", + "bbox": [ + 76, + 393, + 473, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 712, + 209, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative image modeling has developed significantly in recent years and is able to generate diverse high-resolution images even indistinguishable from real images. However, training such models requires intense computation resources and large datasets, which restricts the application scope of generative models. For some scenarios, collecting large datasets is impossible like paintings by specific artists. Benefiting from Vision-Language models learning from large image-text pairs, text can be leveraged as a de", + "bbox": [ + 76, + 738, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scription of abstract visual semantics to guide generative domain adaptation instead of a collection of image samples in target domain. As an expressive representation, text has shown great success in semantic image generation and manipulation recently [23, 20]. Based on the generative prior of pre-trained models and flexible text description of target domain, text-driven generative domain adaptation can generate more various images and have promising applications.", + "bbox": [ + 496, + 362, + 893, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To reduce the requirement of training samples, traditional methods propose to train generative models in the target domain with only limited samples by adapting pretrained models in the large-scale source domain which contains high-level semantic knowledge as a generative prior. These few-shot adaptation methods either finetune only a part of parameters within networks to preserve most source domain knowledge [15] or impose strong regularization on the generated images [30, 36]. However, these methods still require additional training samples of target domain and adversarial training process. As the number of samples drops, the image fidelity and diversity also hurt severely. Different from these methods, text-driven generative domain adaptation requires no image samples but texts to describe the target domain. Pioneer work [5] proposed to encourage the visual change between samples from target and source generators to align with semantic direction described by text in the CLIP [19] embedding space, which achieves generative adaptation for miscellaneous domains in short training time.", + "bbox": [ + 496, + 488, + 895, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The main challenge of text-driven GAN adaptation is the mode collapse problem due to the entanglement of intradomain semantics and inter-domain style in text representation. Besides the specified target style described by text, there also exists an unknown pattern for the semantics of images. This leads to a decrease of variations in generated images when the style effect is optimized to approach target domain. As shown in Figure 2, while the number of qua", + "bbox": [ + 496, + 779, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 94, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7019", + "bbox": [ + 482, + 944, + 516, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/515369480a1c248fea0c85684a4156da85c52a1ed05fa55b0c61b4b6e3569dda.jpg", + "image_caption": [ + "Original" + ], + "image_footnote": [], + "bbox": [ + 91, + 99, + 248, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ea0686e26676b153e03298514bfa2bbce6bfdf29f753b30c5228666289dc52e4.jpg", + "image_caption": [ + "Sketch" + ], + "image_footnote": [], + "bbox": [ + 248, + 99, + 401, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e940d699410540f4fa44b31774cac8f327547f6703362a3200867b143a59b04a.jpg", + "image_caption": [ + "Ukiyo-e", + "Figure 1. Text-driven generative domain adaptation with various text descriptions. The generated samples should both reflect characteristics of target domain from text and preserve the original identity." + ], + "image_footnote": [], + "bbox": [ + 403, + 99, + 557, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2a66b4c5f146b89da43351e1f4f483152dfd6a0b41c56c9a2f5892acea894292.jpg", + "image_caption": [ + "Van Gogh" + ], + "image_footnote": [], + "bbox": [ + 558, + 99, + 712, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/97d84f6d91337b724cb3295cf008a74389dd98076f1b6b2630abc45981073a59.jpg", + "image_caption": [ + "Werewolf" + ], + "image_footnote": [], + "bbox": [ + 712, + 99, + 867, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions increases, the generated samples tend to have similar patterns of mouth and eyes, which reduces most of the variations in the origin model. The main reason for the mode collapse problem is that the optimization process only cares about the distance of generated samples to target domains, and the intra-domain feature variations are easily ignored.", + "bbox": [ + 75, + 301, + 467, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above challenge, researchers [36] proposed to preserve the diversity of source domain through a within-domain consistency loss which keeps consistency between sample changes in source domain and target domain. However, this regularization is too strong to restrict the style effect of target generator close to source domain. The previous theorem about GAN latents analysis has shown that the Hessian matrix of generator reflects the variations of generator and can be used to explore meaningful directions from top eigenvectors. Inspired by this, we try to leverage the spectrum of Hessian matrix as a quantitative evaluation of model diversity in the adaptation problem. This disentangles the relative diversity between generated samples from absolute generative distribution and makes a general way to regularize diversity of generative model.", + "bbox": [ + 75, + 397, + 467, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose spectral consistency regularization to solve the problem of mode collapse in text-driven generative domain adaptation from the geometric point of view. First, we analyze the Hessian matrix of generator's manifold in the metric space by eigendecomposition. The eigenvalues of Hessian matrix are decreasing in the adaptation process, which is consistent with the mode collapse problem of visual observations. Second, we introduce the spectral consistency regularization on the Hessian matrix to prevent the latent space of generator from degrading. This regularization helps preserve intra-domain variations of source domain without restricting style effects of target generators. We further develop a stochastic method to regularize the spectrum of Hessian matrix without calculating the full matrix, which reduces the expensive computational cost. Finally, we design the granularity adaptive regularization considering the layer-decomposition characteristic of $\\mathrm{W + }$ space in StyleGAN.", + "bbox": [ + 75, + 628, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this paper are summarized below:", + "bbox": [ + 517, + 301, + 879, + 316 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We analyze the commonly occurred mode collapse problem in GAN adaptation from the geometric point of view and provide a quantitative evaluation of model diversity to reveal the reason of mode collapse.", + "2. We propose a spectral consistency regularization for text-driven GAN adaptation, which preserves diversity of original domain and generates high fidelity images of target domain. A granularity adaptive regularization is further designed to flexibly control balance between diversity and stylization for target model.", + "3. We conduct experiments and ablation studies for a wide range of target domains. The experiments show the effectiveness of our proposed spectral consistency regularization and its applications to downstream tasks like image editing and image-to-image translation." + ], + "bbox": [ + 511, + 325, + 890, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 588, + 638, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-driven Image Synthesis and Manipulation. Traditional methods approached text-driven image generation by training a conditional GAN[22]. Several following works have been proposed to improve generation quality either by multi-scale networks [33] or attention mechanism [31]. Recently, transformer-based auto-regressive generative models were introduced to view text-driven image synthesis as conditional sequence generation of visual tokens conditioning on text embeddings [3, 21, 32]. Diffusion models were also leveraged as the decoder for image generation, which achieves tremendous improvement for generating high quality images [23, 20].", + "bbox": [ + 496, + 613, + 890, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another kind of method is to leverage Contrastive Language-Image Pre-training (CLIP) [19] models as knowledge guidance for text based image generation. This is achieved by optimizing the latent code of pre-trained generator to close the distance between generated images and input text in the shared embedding space. The optimized latent codes of generator are either in the StyleGAN latent", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7020", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/58d312ec650c0316d4273c2e4caefcb18cf80c4a38ff0607f270bb8f414cf9c5.jpg", + "image_caption": [ + "Figure 2. The mode collapse problem in the photo-to-sketch GAN adaptation by StyleGAN-NADA [5]. The trace norm of Hessian matrix is gradually decreasing, which is consistent with the visual examples showing similar patterns with mouth and eyes." + ], + "image_footnote": [], + "bbox": [ + 83, + 93, + 861, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "space [18] or in VQGAN codebook space [1]. Some methods also defined the parametrized space by vector graphics such as Bezier curves [4].", + "bbox": [ + 75, + 324, + 468, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-shot GAN Adaptation. It aims to transfer pretrained generator to another domain when there are not enough samples to train from scratch. Its main challenge is the mode collapse problem because the generator is prone to overfit training samples in target domain and lose the diversity of the origin domain. There have been many methods to tackle this problem. They either froze most of the parameters of the pre-trained network[15] or embedded a small number of trainable parameters into the source model[16]. Recently, [17] proposed cross-domain distance consistency loss to preserve the relative similarities and differences between instances in the source domain. [30] introduced spatial structural consistency loss to align the spatial information between the synthesis image pairs of the source and target domains. These methods still require manually collected samples, and as the number of samples decreases, the mode collapse problem becomes more apparent.", + "bbox": [ + 75, + 371, + 468, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides, StyleGAN-NADA [5] further proposed to take advantage of the CLIP [19] model as knowledge guidance for GAN adaptation, and only natural language prompts are required without even a single image. Similarly, [36] used the image encoder of CLIP for one-shot adaptation. However, these methods still suffer from the mode collapse problem. In this paper, we propose the spectral consistency regularization to tackle the problem of mode collapse without hurting target generation performance.", + "bbox": [ + 75, + 628, + 468, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent Space Analysis of GANs. Many works have explored the latent space of pretrained generator for image manipulation. Some methods used supervised datasets to learn directions in the latent space for attribute editing [24] or semantic image editing [13]. Other works instead applied unsupervised methods to reveal the latent space. [26] decomposed the learned weights of the pre-trained network to identify semantically meaningful directions. [7] applied principal component analysis in the latent space. Recently,", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[29] proposed to analyze the latent space of generative models from geometric point of view. They found that the eigenvectors corresponding to the largest eigenvalues of the Hessian matrix for generator dominate interpretable variations. In this paper, we analyze the GAN adaptation problem in a similar way and propose to regularize target generator by the spectrum of Hessian matrix.", + "bbox": [ + 496, + 324, + 893, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 460, + 591, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Text-Driven Generative Domain Adaptation", + "text_level": 1, + "bbox": [ + 500, + 491, + 872, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-driven GAN adaptation aims to transfer a pretrained generator to target domain specified by the text description. To guide the GAN adaptation by text, pre-trained CLIP model is leveraged to measure the similarity between image and text. CLIP is a Vision-Language model trained on 400 million (image, text) pairs collected from the internet with contrastive loss [19]. One commonly used objective function for text-driven image manipulation is the global loss that optimizes the similarity between generated images and target text:", + "bbox": [ + 496, + 520, + 893, + 672 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {g l o b a l}} = D _ {\\text {C L I P}} (G (z), t _ {\\text {t a r g e t}}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 695, + 890, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D_{CLIP}$ is the cosine distance in the CLIP space, $t_{target}$ is the target text. However, this only applies to in-domain image manipulation combined with identity consistency regularization[18]. This regularization is too strong for cross-domain GAN adaptation with large domain gaps like human to werewolf. Direct optimization of the above global loss leads to adversarial solutions since adding pixel-level perturbations can fool the CLIP classifier in the absence of a generative prior favoring real-image manifold [5]. To overcome this limit, the directional loss is used to optimize the direction between source and target", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7021", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a7ed1545cd85801af19d50d9c595ebec69273f370eef39f6dcb24c78312d2d3f.jpg", + "image_caption": [ + "Figure 3. Illustration of our proposed method for text-driven generative domain adaptation. The training objective of our model consists of directional loss and spectral consistency loss. We feed the same latent code to source generator and target generator and generate a pair of source and target images. The directional loss encourages the direction between embeddings of the pair to align with the semantic direction of text description. The spectral consistency loss regularizes the trace norm of Hessian matrix of target generator to prevent the mode collapse problem." + ], + "image_footnote": [], + "bbox": [ + 89, + 92, + 885, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "domain:", + "bbox": [ + 76, + 426, + 137, + 440 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i r e c t i o n}} = 1 - \\frac {\\Delta I \\cdot \\Delta T}{| \\Delta I | | \\Delta T |} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 460, + 468, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta T = E _ {T} \\left(t _ {\\text {t a r g e t}}\\right) - E _ {T} \\left(t _ {\\text {s o u r c e}}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 496, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta I = E _ {I} \\left(G _ {\\text {t r a i n}} (z)\\right) - E _ {I} \\left(G _ {\\text {f r o z e n}} (z)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 516, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $E_{T}$ and $E_{I}$ are the text and image encoder of CLIP, $t_{source}$ and $t_{target}$ are the source and target class texts. The directional loss can prevent adversarial solutions. Since target generated images should have given direction to corresponding source images, generating a single adversarial instance is impossible. At training time, the same latent code is fed into source and target generator, then the target generator is optimized using the directional loss. However, the directional loss proposed by StyleGAN-NADA still suffers from the mode collapse problem, as shown in the images of Figure 2. So we propose the spectral consistency regularization derived from geometry analysis of GANs to resolve this problem.", + "bbox": [ + 75, + 556, + 472, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Geometry Analysis of GAN Adaptation", + "text_level": 1, + "bbox": [ + 76, + 780, + 415, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We denote the generative network as a mapping from latent code $z$ to a manifold in image space as $G(z)$ . Considering a squared distance function $d^2$ for two images, we express the local variations of $G(z)$ from moving towards direction $\\Delta z$ by second-order Taylor expansion. This is formulated as:", + "bbox": [ + 75, + 810, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {\\Delta z \\to 0} d ^ {2} (G (z), G (z + \\Delta z)) =\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 450, + 740, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd ^ {2} (G (z), G (z)) + \\frac {\\partial d ^ {2} (G (z) , G (z + \\Delta z))}{\\partial \\Delta z} \\cdot \\Delta z +\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 476, + 862, + 507 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta z ^ {T} \\cdot \\frac {\\partial^ {2} d ^ {2} (G (z) , G (z + \\Delta z))}{\\partial \\Delta z ^ {2}} \\cdot \\Delta z \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 508, + 890, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first two terms are zero since $d^{2}(G(z), G(z + \\Delta z))$ is local minima when $\\Delta z = 0$ . Denote the second derivatives as Hessian matrix $H(z)$ , and we have $d^{2}(G(z), G(z + \\Delta z)) = \\Delta z^{T}H(z)\\Delta z$ . For a normalized vector $\\Delta z$ , we can conclude that $\\sigma_{min} \\leq d^{2}(G(z), G(z + \\Delta z)) \\leq \\sigma_{max}$ where $\\sigma_{min}$ and $\\sigma_{max}$ are the smallest and largest eigenvalues of $H$ . Thus, we can use the trace norm of $H_{z}$ to reflect the statistics of diversity in generative models, which is the sum of all eigenvalues of $H$ . Especially, for a squared $L2$ distance function in metric space $\\phi$ , $d^{2}(z_{1}, z_{2})_{\\phi} = \\frac{1}{2}\\|\\phi(G(z_{1})) - \\phi(G(z_{2}))\\|^{2}$ , the Hessian matrix $H_{\\phi}(z_{0})$ is a simple transformation from the Jacobian $J_{\\phi}(z_{0})$ :", + "bbox": [ + 498, + 547, + 892, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} H _ {\\phi} \\left(\\boldsymbol {z} _ {0}\\right) = \\frac {\\partial^ {2}}{\\partial \\boldsymbol {z} ^ {2}} \\frac {1}{2} \\| \\phi \\left(\\boldsymbol {z} _ {0}\\right) - \\phi (\\boldsymbol {z}) \\| _ {2} ^ {2} (6) \\\\ = J _ {\\phi} \\left(\\boldsymbol {z} _ {0}\\right) ^ {T} J _ {\\phi} \\left(\\boldsymbol {z} _ {0}\\right) (7) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 739, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the top eigenvectors of $H_{\\phi}$ correspond to right singular vectors of the Jacobian $J_{\\phi}$ . Besides, the Hessian-vector product is related to Jacobian-vector product as follows:", + "bbox": [ + 496, + 801, + 892, + 848 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} ^ {T} H _ {\\phi} (\\boldsymbol {z} _ {0}) \\boldsymbol {v} = \\| J _ {\\phi} (\\boldsymbol {z} _ {0}) \\boldsymbol {v} \\| ^ {2}, \\quad J _ {\\phi} (\\boldsymbol {z}) = \\frac {\\partial \\phi (G (\\boldsymbol {z}))}{\\partial z} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 867, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7022", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To analyze the mode collapse problem, we calculate the statistics of Hessian trace from different samples of $z$ during adaptation process of the previous state-of-the-art method. The results of GAN adaptation from photo to sketch with StyleGAN-NADA [5] is shown in Figure 2. We find that as the iteration count increases, the mode collapse problem becomes more severe and the Hessian trace is also decreasing, which proves that the Hessian trace can reflect the diversity of generator. In the early stage of adaptation, the decrease of trace norm is mainly caused by style adaptation since there are no color variations for sketch domain. But for the late stage from 200th step to 400th step, the style effect changes little and the face attributes tend to have similar pattern. Since the target text only represents a fixed direction without variation, different samples are encouraged to approach the same fixed pattern. We also give a proof of the relationship between Hessian Trace and model diversity in the following section.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Relation between Hessian Trace and Diversity", + "text_level": 1, + "bbox": [ + 76, + 372, + 465, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the mapping from standard normal distribution $z \\sim \\mathcal{N}(0, I)$ to general multivariate normal distribution $y \\sim \\mathcal{N}(\\mu, \\Sigma)$ with the generator as a linear function $G(z) = A z + b$ , which has $\\mu = b, \\Sigma = A A^T$ . The linear assumption of generator holds when we consider a small neighborhood around $z$ , i.e. $\\lim \\Delta z \\to 0$ .", + "bbox": [ + 76, + 395, + 468, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The variance of $y$ can be computed as:", + "bbox": [ + 96, + 486, + 352, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} V a r (y) = E [ \\| y - E [ y ] \\| _ {2} ^ {2} ] \\\\ = E \\left[ \\sum_ {i = 1} ^ {n} \\left(y _ {i} - E \\left[ y _ {i} \\right]\\right) ^ {2} \\right] \\\\ = \\sum_ {i = 1} ^ {n} E \\left[ \\left(y _ {i} - E [ y _ {i} ]\\right) ^ {2} \\right] \\\\ = \\sum_ {i = 1} ^ {n} V a r (y _ {i}) \\\\ = \\sum_ {i = 1} ^ {n} \\Sigma_ {i i} \\\\ = \\operatorname {T r a c e} (\\Sigma) = \\operatorname {T r a c e} \\left(A A ^ {T}\\right) \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 512, + 467, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the other side, the $d^2 (G(z),G(z + \\Delta z))$ can be expanded as follows:", + "bbox": [ + 76, + 727, + 468, + 756 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} d ^ {2} (G (z), G (z + \\Delta z)) = \\| G (z) - G (z + \\Delta z) \\| _ {2} ^ {2} \\\\ = \\| A z + b - (A (z + \\Delta z) + b) \\| _ {2} ^ {2} \\\\ = \\| A \\Delta z \\| \\\\ = \\Delta z ^ {T} A ^ {T} A \\Delta z \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 768, + 467, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "such the Hessian Matrix $H$ of $\\Delta z$ with respect to $d^{2}(G(z), G(z + \\Delta z))$ equals to $A^{T}A$ , e.g. $H = A^{T}A$ . Combining above two equations with $Trace(AA^{T}) =$", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Trace $(A^TA)$ , we have $Var(y) = Trace(H)$ , which means that the Hessian Trace for every sample in target distribution reflects the variance and diversity of this distribution. If the Hessian Trace is small, the target distribution only spans a small region in space. This is consistent to the mode collapse problem in generative models.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Spectral Consistency Regularization", + "text_level": 1, + "bbox": [ + 500, + 189, + 815, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To prevent the target generator from mode collapse, we propose spectral consistency regularization to prevent the diversity of generator from degrading, which is calculated as:", + "bbox": [ + 496, + 212, + 890, + 271 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e g} = \\left\\| \\operatorname {T r a c e} \\left(H _ {s} (z)\\right) - \\operatorname {T r a c e} \\left(H _ {t} (z)\\right) \\right\\| \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 281, + 890, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H_{s}(z)$ and $H_{t}(z)$ are the Hessian matrix of the source generator and target generator evaluated with the same latent code $z$ . However, directly computing the Hessian matrix requires backpropagation $n$ times where $n$ is the dimension of feature vector in metric space. Instead, we use the Hutchinson's method for trace estimator [9] to compute a stochastic estimator of Hessian Trace, which is formulated as:", + "bbox": [ + 496, + 303, + 890, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {T r a c e} (H (z)) = \\mathbb {E} [ v ^ {T} H (z) v ] = \\mathbb {E} [ \\| J _ {\\phi} (z) v \\| ^ {2} ] \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 431, + 890, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $v \\sim \\mathrm{Rademacher}(\\frac{1}{2})$ , and the second transformation is derived from Equation 8. So the calculation of Hessian matrix is transformed into the calculation of Jacobian-Vector product.", + "bbox": [ + 496, + 455, + 890, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different from the within-domain consistency loss[36] which restricts target generated samples based on relative difference of source samples, the spectral consistency regularization only cares about the diversity of target model. This doesn't impose restriction on the direction of target adaptation, so our method can generate samples more consistent with target text without losing model diversity.", + "bbox": [ + 496, + 517, + 890, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training objective of text-driven GAN adaptation is a weighted combination of directional loss and spectral consistency regularization loss $\\mathcal{L} = \\mathcal{L}_{dir} + \\lambda \\mathcal{L}_{reg}$ . Since different target domains have their own characteristic, it is required to tune the hyperparameter $\\lambda$ for better performance. To prevent the exhausting hyperparameter searching, we propose an adaptive loss reweighting method to balance the influence of these two loss items. Specifically, the adaptive weight $\\lambda$ is calculated as $\\lambda_{spectral} \\frac{\\|\\nabla_{G_L} \\mathcal{L}_{dir}\\|}{\\|\\nabla_{G_L} \\mathcal{L}_{reg}\\|}$ , where $G_L$ denotes the last layer of generator and $\\lambda_{spectral}$ is a manually specified hyperparameter, typically 1.0 is an appropriate choice.", + "bbox": [ + 496, + 623, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Granularity Adaptive Regularization", + "text_level": 1, + "bbox": [ + 500, + 816, + 821, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For text-driven GAN adaptation problem, the adaptation granularity of different target domains varies from texture to structure. For example, the photo-to-sketch adaptation mainly focuses on the appearance and texture change,", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7023", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a68ea668e00bafe197ea5b25168a2e57373e444b35b178b656f32ba89d709e39.jpg", + "image_caption": [ + "Figure 4. Visual results of our method and state-of-the-art StyleGAN-NADA [5] for different GAN adaptations. The top row shows the source text and target text description. The left column presents samples generated by source generator. Our method not only generates samples described by target text with high fidelity, but also produces diverse and identity consistent images corresponding to source domain." + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 883, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "while the werewolf domain has more variations on semantic structure. Regularization of diversity in the whole granularity will restrict the adaptation performance. To alleviate this problem, we propose granularity adaptive regularization based on the disentangled characteristic of StyleGAN latent code.", + "bbox": [ + 75, + 463, + 468, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The latent codes injected into different layers in StyleGAN influence different granularities. The style code in low resolution represents high-level aspects such as pose and shape, that in middle resolution controls facial features and hairstyle, and that in high resolution influences color scheme and local textures. Specifically, we use the $W+$ space as input space for Jacobian matrix calculation, where each input latent code consists of 18 512-dimensional vectors so both $z$ and $v$ in Equation 4 are in $R^{18 \\times 512}$ . By masking $v$ with a mask vector $m \\in \\{0,1\\}^{18}$ for different layers, we can specify the granularity of variations involved in the regularization. The calculation of Hessian trace under mask is formulated as:", + "bbox": [ + 75, + 554, + 472, + 750 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {T r a c e} (H (z)) = \\mathbb {E} [ (v \\odot m) ^ {T} H (z) (v \\odot m) ] (13) \\\\ = \\mathbb {E} [ \\| J _ {\\phi} (z) (v \\odot m) \\| ^ {2} ]. (14) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 761, + 468, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous convention, we divide the style code for 18 layers into 3 groups, which are for coarse, middle and fine scale. The results of different mask strategies are shown in Figure 6.", + "bbox": [ + 76, + 809, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To explore the best mask strategy for different target domains, we propose to use an adaptive soft mask vector", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\{\\tilde{m} |\\tilde{m}\\in R^{18},\\| \\tilde{m}\\| = 1\\}$ for all layers. During training, the mask vector is optimized with respect to the overall training objective. To reduce the directional loss, the mask vector will assign less value to the latent code corresponding to the granularity that changes most, while other values will increase to preserve the diversity of source generator.", + "bbox": [ + 498, + 462, + 890, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 571, + 624, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we will show the qualitative and quantitative results of our method. We illustrate the generated results for a wide range of target domains from style and texture changes to shape and semantic modifications. We also compare the proposed spectral consistency regularization with other regularization methods for diversity preservation. Next, we perform an ablation study on our method to evaluate the effectiveness of each component. Finally, we demonstrate the applications of text-guidance GAN adaptation, including image-to-image translation and image editing. To verify the generalization of proposed spectral consistency regularization, we also apply it to one-shot GAN adaptation, which is shown in supplementary appendix.", + "bbox": [ + 496, + 597, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the StyleGANv2 [11] generator pretrained on FFHQ as the source generator. During domain adaptation, we optimize all the parameters of generator except for the mapping network and toRGB layers. We use the Adam [12] optimizer with learning rate 0.001. The $\\lambda_{\\mathrm{spectral}}$ is set to 1.0. For most target domains, only 300 iterations are required to achieve convergence. Following CLIP, we use 79", + "bbox": [ + 496, + 795, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7024", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d90133782b5c4ac0795191d31b2f9fd064804e7e8e0d46a518e9d9dbe4b8aa23.jpg", + "image_caption": [ + "Figure 5. Comparison results of our spectral consistency regularization with other regularization methods. Detailed explanations of these methods are in Appendix." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 883, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "manually designed prompts like \"a photo of a...\" with provided target domain description and feed them to text encoder of CLIP to get the embeddings of target domain.", + "bbox": [ + 75, + 522, + 468, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Comparison Results", + "text_level": 1, + "bbox": [ + 76, + 575, + 269, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison In Figure 4, we show the comparison results of our method with state-of-the-art model StyleGAN-NADA [5] for a wide range of target domains, which varies from texture changes like sketch and Mona Lisa paintings to geometric change like werewolf and Pixar style. The results demonstrate that our method not only generates highly stylistic images consistent with target text description for different target domains, but also produces images with diversity inherited from the pretrained source generator. Compared with StyleGAN-NADA which has obvious mode collapse problem like the mouth pattern in sketch and hair in werewolf, our model generates target images with better identity consistency. This proves that the spectral consistency regularization can preserve the diversity of source domain. In the supplementary appendix, we present additional visual results for the dogs and cars domain.", + "bbox": [ + 75, + 598, + 468, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also perform text-driven GAN adaptation experiments with other regularization methods, including the Selective Cross-modal Consistency (SCC) loss[35], Within-Domain Consistency (WDC) loss[36], the Mode", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Seeking (MS) loss[14], Perceptual Path Length (PPL) regularization[11] and Cross Domain Correspondence (CDC) regularization[17]. Detailed explanations of these regularization methods are in supplementary appendix. As shown in Figure 5, SCC, WDC and MS regularization impose too strong regularization to target generator and restrict the domain specific attributes for target domain. Suffering from mode collapse problem, the generation results of PPL and CDC exhibits the same pattern across different samples. In comparison, our method has a better balance between the diversity and stylization of target generator.", + "bbox": [ + 496, + 522, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison Besides our proposed trace norm of Hessian Matrix in Equation 3, we also leverage the Perceptual Path Length(PPL) [10] for quantitative diversity comparison. It measures the perceptually-based pairwise image distance [34] for a linear interpolation path in the latent path. The average PPL in the latent space $\\mathcal{Z}$ is:", + "bbox": [ + 496, + 688, + 890, + 779 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP P L _ {\\mathcal {Z}} = \\mathbb {E} \\left[ \\right. \\frac {1}{\\epsilon^ {2}} d \\left( \\right.G \\left(\\operatorname {s l e r p} \\left(\\mathbf {z} _ {1}, \\mathbf {z} _ {2}; t\\right)\\right), G \\left(\\operatorname {s l e r p} \\left(\\mathbf {z} _ {1}, \\mathbf {z} _ {2}; t + \\epsilon\\right)\\right)\\left. \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 483, + 785, + 924, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $d(.,.)$ evaluates the perceptual distance between two images, and slerp denotes spherical interpolation since the input latent is normalized. The perceptual path length estimates the diversity of the generator via finite differences, and our method estimates the diversity analytically.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7025", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/eb11f4a96ebacdf737e5a9da984d45147e86c688735ea815847fbc71a96b7627.jpg", + "table_caption": [ + "Table 1. Comparison results for diversity by PPL and Hessian trace between different regularization methods. The PPL and Hessian trace for the pretrained source generator is 419.22 and 0.309." + ], + "table_footnote": [], + "table_body": "
ResultsPhoto → SketchPhoto → Mona LisaHuman → WerewolfPhoto → Pixar
PPLTracePPLTracePPLTracePPLTrace
SCC547.340.526485.700.521428.450.343440.170.535
WDC378.250.090363.170.191311.340.116345.310.167
MS466.990.167331.060.233352.610.191377.880.358
PPL241.500.028209.090.061297.740.062300.190.063
CDC348.010.062259.920.079351.730.089299.490.111
NADA323.250.061281.590.098302.910.101343.540.112
Ours463.570.116321.430.140383.190.137353.800.181
", + "bbox": [ + 171, + 125, + 802, + 271 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0aac0a09d70cd591fbc2301822724f96557d945df9ecd5a2950df209ecadd8cc.jpg", + "table_caption": [ + "Table 2. User preference study. The numbers represent the percentage of users who prefer the images synthesized by our method." + ], + "table_footnote": [], + "table_body": "
Model ComparisonQualityStyleAttributes
Ours vs. NADA [5]86.4%59.4%91.2%
Ours vs. SCC [35]64.2%86.2%65.0%
Ours vs. MS [14]59.4%83.6%54.2%
Ours vs. WDC [36]62.2%89.8%49.8%
Ours vs. PPL [11]92.4%71.4%95.0%
Ours vs. CDC [17]79.6%60.6%87.8%
", + "bbox": [ + 76, + 330, + 444, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The quantitative results are shown in Table 1. The mentioned strong regularization methods including SCC, WDC and MS have large values of PPL and Hessian trace at the cost of restricting adaptation effects. Our method performs competitive with these methods in diversity metrics, but does much better in transferring target styles. Compared with CDC and PPL, our method better preserve the diversity of source domain reflected in PPL and Hessian Trace. For all different target domains, our method outperforms previous state-of-the-art model StyleGAN-NADA for both PPL and Hessian trace. This benefits from that our method can alleviate the mode collapse problem with spectral consistency regularization and generate more diverse images without restricting the style adaptation results.", + "bbox": [ + 75, + 474, + 468, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study We conducted a user study to compare with other regularization methods. Specifically, participants were requested to select the superior synthesized samples in relation to three measurements: (1) image quality, (2) style consistency with the target description (3) attribute consistency with the source image. Five hundred samples were randomly generated for each comparative analysis. The aggregated results are presented in Table 2. The users strongly favor our method across all three aspects. In contrast to other methods, our method demonstrates a better tradeoff between style effects and attribute preservation. The spectral consistency regularization can better preserve the diversity of source domain while not restricting the style effects for target domain.", + "bbox": [ + 75, + 688, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 306, + 653, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Granularity Adaptive Regularization. In Figure 6, we demonstrate generated results with different regularization strategies applied in $W+$ space in Section 3.4. Regularization to the coarse scale will preserve the structure of source image but the diversity of fine features like hair will lose. The global regularization performs similarly to coarse regularization since the coarse features dominate the diversity of generator. Only applying regularization to the fine scale will generate high-frequency textures and the structure characteristic like necks will collapse. In comparison, our proposed granularity adaptive regularization both preserves the diversity of source domain in all scales and matches the styles of target domain.", + "bbox": [ + 496, + 330, + 893, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Strength of regularization. In Figure 7, we show the generated samples with a linear interpolated loss weight $\\lambda_{\\text{spectral}}$ . We can observe that with increasing $\\lambda_{\\text{spectral}}$ , the generated samples maintain more diversity of source domain, and they also illustrate the most significant characteristic of target domain.", + "bbox": [ + 496, + 527, + 893, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Choices of metric space. We conduct experiments about the metric space of spectral consistency regularization with different feature encoder $\\phi(x)$ that evaluates the distance between image samples. Besides CLIP [19] image encoder in our method, we also leverage VGG [27] which was commonly used in style transfer [6], MoCo [8] of contrastive representation learning, ArcFace [2] for face recognition and the plain pixel space. As shown in Figure 8, compared to other feature encoders, the spectral consistency regularization with CLIP encoder shows best performance for preserving the identity of source image.", + "bbox": [ + 496, + 619, + 893, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Applications", + "text_level": 1, + "bbox": [ + 500, + 800, + 635, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image-to-Image Translation We combine the adapted generator in target domain with a GAN inversion encoder to implement image-to-image translation. Given a real-world image, we invert it to the latent code in $W$ space via an e4e encoder [28], which is then fed to target generator to", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7026", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5fc7ca3b9b5b5c3f958f58a3b89e95fc1e22683196a7469b763cb5e2c2c18612.jpg", + "image_caption": [ + "Source", + "Figure 6. Results of different regularization strategies for GAN adaptation from photo to Edvard Munch paintings." + ], + "image_footnote": [], + "bbox": [ + 80, + 114, + 205, + 210 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/619c9ffed9b5792082c689dc9dc300f741848aed72236b0181a60a37dfb0e119.jpg", + "image_caption": [ + "Global" + ], + "image_footnote": [], + "bbox": [ + 218, + 114, + 341, + 210 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9024a80d7daccfd1347ef37f24a1da860714a87d93857f2ea591094ab964d0f2.jpg", + "image_caption": [ + "Coarse" + ], + "image_footnote": [], + "bbox": [ + 354, + 114, + 478, + 210 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a4ea90735778d2df4a253071106db1e4d963cc09716d807c5ff78b4f12e7876b.jpg", + "image_caption": [ + "Middle" + ], + "image_footnote": [], + "bbox": [ + 491, + 114, + 616, + 210 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dc931f537612bbd32897324360cc40b8bcdf707a9284110a879c9977b75ef8ae.jpg", + "image_caption": [ + "Fine" + ], + "image_footnote": [], + "bbox": [ + 629, + 114, + 753, + 210 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dd71d7613cfe1393d543968c28d55e2f7bd77ceb464c37c972e422415e27db0d.jpg", + "image_caption": [ + "Adaptive" + ], + "image_footnote": [], + "bbox": [ + 766, + 114, + 890, + 210 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/35970361d29c69aed78b67ee96e0df1924c04fba09bbf3750eab2690da5b6332.jpg", + "image_caption": [ + "Figure 7. Generation results of linearly interpolated $\\lambda_{\\mathrm{spectral}}$ in regularization." + ], + "image_footnote": [], + "bbox": [ + 78, + 255, + 468, + 354 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a73e311b2e92cf70a3332c7cfe3833894b6ffd67784ccc44416b98c288562e96.jpg", + "image_caption": [ + "Figure 8. The generated results with spectral consistency regularization under the metric spaces defined by different encoders." + ], + "image_footnote": [], + "bbox": [ + 83, + 402, + 460, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "produce target image. As shown in Figure 9, our method can achieve high-quality image translation and preserve the identity of source image for different domains.", + "bbox": [ + 75, + 547, + 468, + 593 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Image Editing In Figure 10, we demonstrate the image editing results performed on the target domain. We leverage the meaningful directions found by InterfaceGAN [25] to edit target images. We can observe that the editing directions from source domain still apply to target domains, which proves that the target domain preserves the semantic distribution of source domain.", + "bbox": [ + 75, + 595, + 468, + 699 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0e54b7b20cfa93fc789ffa81ce845f9de3b9158774154d649c0ccb41659a7370.jpg", + "image_caption": [ + "Figure 9. Image translation of real-world images to different target domains. Each column shows a target domain, and the top row is the text description for target domain. The transferred images represent both target style and the identity of source image." + ], + "image_footnote": [], + "bbox": [ + 84, + 712, + 460, + 838 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/bf5abcc8f4d6fb9fb4919fb89e3edb59808e878cffd554a466fa09aa033e6f45.jpg", + "image_caption": [ + "Figure 10. Editing images in target domain for real-world images. The top row shows the edited attributes." + ], + "image_footnote": [], + "bbox": [ + 501, + 256, + 890, + 306 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 373, + 617, + 388 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose the spectral consistency regularization for text-driven generative domain adaptation. The key insight of our method is to build a quantitative diversity estimator to preserve the intra-domain diversity of source generator without restricting the adaptation of target style. We also introduce an adaptive regularization strategy for granularity-flexible adaptation. The experiments demonstrate our method greatly improves the generation results for a wide range of target domains.", + "bbox": [ + 496, + 400, + 890, + 536 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 554, + 660, + 570 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key R&D Program of China under Grant 2018AAA0102000, in part by National Natural Science Foundation of China: 62322211, 62236008, U21B2038, 62225207, U19B2038 and 61931008, and Youth Innovation Promotion Association of Chinese Academy of Sciences under Grant 2020108.", + "bbox": [ + 496, + 580, + 890, + 671 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 688, + 596, + 704 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Katherine Crowson, Stella Rose Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. ArXiv, abs/2204.08583, 2022. 3", + "[2] Jiankang Deng, J. Guo, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4685-4694, 2019. 8", + "[3] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868-12878, 2021. 2" + ], + "bbox": [ + 506, + 714, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7027", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[4] Kevin Frans, Lisa B. Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. ArXiv, abs/2106.14843, 2021. 3", + "[5] Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ArXiv, abs/2108.00946, 2021. 1, 3, 5, 6, 7, 8", + "[6] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 8", + "[7] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. ArXiv, abs/2004.02546, 2020. 3", + "[8] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 8", + "[9] Michael F. Hutchinson. A stochastic estimator of the trace of the influence matrix for laplacian smoothing splines. Communications in Statistics - Simulation and Computation, 18:1059-1076, 1989. 5", + "[10] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4396-4405, 2019. 7", + "[11] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8107-8116, 2020. 6, 7, 8", + "[12] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. 6", + "[13] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. ArXiv, abs/2111.03186, 2021. 3", + "[14] Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Siwei Ma, and Ming-Hsuan Yang. Mode seeking generative adversarial networks for diverse image synthesis. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1429-1437, 2019. 7, 8", + "[15] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze discriminator: A simple baseline for fine-tuning gans. ArXiv, abs/2002.10964, 2020. 1, 3", + "[16] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2750-2758, 2019. 3", + "[17] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A. Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10738-10747, 2021. 3, 7, 8", + "[18] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and D. Lischinski. Styleclip: Text-driven manipulation of" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "stylegan imagery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2065-2074, 2021. 3", + "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 8", + "[20] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1, 2", + "[21] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 2", + "[22] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 2", + "[23] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1, 2", + "[24] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9240-9249, 2020. 3", + "[25] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:2004-2018, 2022. 9", + "[26] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1532-1540, 2021. 3", + "[27] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2015. 8", + "[28] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40:1 - 14, 2021. 8", + "[29] Binxu Wang and Carlos R. Ponce. The geometry of deep generative image models and its applications. ArXiv, abs/2101.06006, 2021. 3", + "[30] Jiayu Xiao, Liang Li, Chaofei Wang, Zhengjun Zha, and Qingming Huang. Few shot generative model adaption via relaxed spatial structural alignment. ArXiv, abs/2203.04121, 2022. 1, 3", + "[31] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316-1324, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7028", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Benton C. Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. ArXiv, abs/2206.10789, 2022. 2", + "[33] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N. Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5908-5916, 2017. 2", + "[34] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 7", + "[35] Yabo Zhang, Mingshuai Yao, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, and Wangmeng Zuo. Towards diverse and faithful one-shot adaption of generative adversarial networks. ArXiv, abs/2207.08736, 2022. 7, 8", + "[36] Peihao Zhu, Rameen Abdal, John C. Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. ArXiv, abs/2110.08398, 2021. 1, 2, 3, 5, 7, 8" + ], + "bbox": [ + 78, + 90, + 468, + 455 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7029", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/e13295e4-a033-4d38-a3a5-88f3b3385e81_model.json b/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/e13295e4-a033-4d38-a3a5-88f3b3385e81_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0b54c1f8946f9e78df2e7b94d23fa3d8fe600eb0 --- /dev/null +++ b/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/e13295e4-a033-4d38-a3a5-88f3b3385e81_model.json @@ -0,0 +1,2169 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.13, + 0.842, + 0.176 + ], + "angle": 0, + "content": "Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.202, + 0.87, + 0.293 + ], + "angle": 0, + "content": "Zhenhuan Liu\\(^{1,2}\\) Liang Li\\(^{1*}\\) Jiayu Xiao\\(^{1,2}\\) Zheng-Jun Zha\\(^{3}\\) Qingming Huang\\(^{1,2,4}\\) \n\\(^{1}\\)Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, CAS \n\\(^{2}\\)University of Chinese Academy of Sciences \n\\(^{3}\\)University of Science and Technology of China \n\\(^{4}\\)Peng Cheng Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.295, + 0.756, + 0.327 + ], + "angle": 0, + "content": "{zhenhuan.liu, jiayu.xiao}@vip1.ict.ac.cn, liang.li@ict.ac.cn, zhazj@ustc.edu.cn, qmhuang@ucas.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.362, + 0.314, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.395, + 0.474, + 0.683 + ], + "angle": 0, + "content": "Combined with the generative prior of pre-trained models and the flexibility of text, text-driven generative domain adaptation can generate images from a wide range of target domains. However, current methods still suffer from overfitting and the mode collapse problem. In this paper, we analyze the mode collapse from the geometric point of view and reveal its relationship to the Hessian matrix of generator. To alleviate it, we propose the spectral consistency regularization to preserve the diversity of source domain without restricting the semantic adaptation to target domain. We also design granularity adaptive regularization to flexibly control the balance between diversity and stylization for target model. We conduct experiments for broad target domains compared with state-of-the-art methods and extensive ablation studies. The experiments demonstrate the effectiveness of our method to preserve the diversity of source domain and generate high fidelity target images. Source code has been released in https://github.com/Victarry/Adaptation-SCR." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.713, + 0.21, + 0.729 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.739, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Generative image modeling has developed significantly in recent years and is able to generate diverse high-resolution images even indistinguishable from real images. However, training such models requires intense computation resources and large datasets, which restricts the application scope of generative models. For some scenarios, collecting large datasets is impossible like paintings by specific artists. Benefiting from Vision-Language models learning from large image-text pairs, text can be leveraged as a de" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.895, + 0.485 + ], + "angle": 0, + "content": "scription of abstract visual semantics to guide generative domain adaptation instead of a collection of image samples in target domain. As an expressive representation, text has shown great success in semantic image generation and manipulation recently [23, 20]. Based on the generative prior of pre-trained models and flexible text description of target domain, text-driven generative domain adaptation can generate more various images and have promising applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.896, + 0.777 + ], + "angle": 0, + "content": "To reduce the requirement of training samples, traditional methods propose to train generative models in the target domain with only limited samples by adapting pretrained models in the large-scale source domain which contains high-level semantic knowledge as a generative prior. These few-shot adaptation methods either finetune only a part of parameters within networks to preserve most source domain knowledge [15] or impose strong regularization on the generated images [30, 36]. However, these methods still require additional training samples of target domain and adversarial training process. As the number of samples drops, the image fidelity and diversity also hurt severely. Different from these methods, text-driven generative domain adaptation requires no image samples but texts to describe the target domain. Pioneer work [5] proposed to encourage the visual change between samples from target and source generators to align with semantic direction described by text in the CLIP [19] embedding space, which achieves generative adaptation for miscellaneous domains in short training time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.897, + 0.902 + ], + "angle": 0, + "content": "The main challenge of text-driven GAN adaptation is the mode collapse problem due to the entanglement of intradomain semantics and inter-domain style in text representation. Besides the specified target style described by text, there also exists an unknown pattern for the semantics of images. This leads to a decrease of variations in generated images when the style effect is optimized to approach target domain. As shown in Figure 2, while the number of qua" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.225, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.958 + ], + "angle": 0, + "content": "7019" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.101, + 0.249, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.223, + 0.204, + 0.238 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.249, + 0.101, + 0.403, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.223, + 0.355, + 0.237 + ], + "angle": 0, + "content": "Sketch" + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.101, + 0.558, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.449, + 0.223, + 0.515, + 0.238 + ], + "angle": 0, + "content": "Ukiyo-e" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.101, + 0.714, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.223, + 0.684, + 0.238 + ], + "angle": 0, + "content": "Van Gogh" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.101, + 0.869, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.223, + 0.834, + 0.237 + ], + "angle": 0, + "content": "Werewolf" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.249, + 0.892, + 0.276 + ], + "angle": 0, + "content": "Figure 1. Text-driven generative domain adaptation with various text descriptions. The generated samples should both reflect characteristics of target domain from text and preserve the original identity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.468, + 0.392 + ], + "angle": 0, + "content": "tions increases, the generated samples tend to have similar patterns of mouth and eyes, which reduces most of the variations in the origin model. The main reason for the mode collapse problem is that the optimization process only cares about the distance of generated samples to target domains, and the intra-domain feature variations are easily ignored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.468, + 0.624 + ], + "angle": 0, + "content": "To address the above challenge, researchers [36] proposed to preserve the diversity of source domain through a within-domain consistency loss which keeps consistency between sample changes in source domain and target domain. However, this regularization is too strong to restrict the style effect of target generator close to source domain. The previous theorem about GAN latents analysis has shown that the Hessian matrix of generator reflects the variations of generator and can be used to explore meaningful directions from top eigenvectors. Inspired by this, we try to leverage the spectrum of Hessian matrix as a quantitative evaluation of model diversity in the adaptation problem. This disentangles the relative diversity between generated samples from absolute generative distribution and makes a general way to regularize diversity of generative model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose spectral consistency regularization to solve the problem of mode collapse in text-driven generative domain adaptation from the geometric point of view. First, we analyze the Hessian matrix of generator's manifold in the metric space by eigendecomposition. The eigenvalues of Hessian matrix are decreasing in the adaptation process, which is consistent with the mode collapse problem of visual observations. Second, we introduce the spectral consistency regularization on the Hessian matrix to prevent the latent space of generator from degrading. This regularization helps preserve intra-domain variations of source domain without restricting style effects of target generators. We further develop a stochastic method to regularize the spectrum of Hessian matrix without calculating the full matrix, which reduces the expensive computational cost. Finally, we design the granularity adaptive regularization considering the layer-decomposition characteristic of \\(\\mathrm{W + }\\) space in StyleGAN." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.302, + 0.88, + 0.317 + ], + "angle": 0, + "content": "The contributions of this paper are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.327, + 0.892, + 0.388 + ], + "angle": 0, + "content": "1. We analyze the commonly occurred mode collapse problem in GAN adaptation from the geometric point of view and provide a quantitative evaluation of model diversity to reveal the reason of mode collapse." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.398, + 0.892, + 0.488 + ], + "angle": 0, + "content": "2. We propose a spectral consistency regularization for text-driven GAN adaptation, which preserves diversity of original domain and generates high fidelity images of target domain. A granularity adaptive regularization is further designed to flexibly control balance between diversity and stylization for target model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.499, + 0.892, + 0.575 + ], + "angle": 0, + "content": "3. We conduct experiments and ablation studies for a wide range of target domains. The experiments show the effectiveness of our proposed spectral consistency regularization and its applications to downstream tasks like image editing and image-to-image translation." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.327, + 0.892, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.64, + 0.604 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Text-driven Image Synthesis and Manipulation. Traditional methods approached text-driven image generation by training a conditional GAN[22]. Several following works have been proposed to improve generation quality either by multi-scale networks [33] or attention mechanism [31]. Recently, transformer-based auto-regressive generative models were introduced to view text-driven image synthesis as conditional sequence generation of visual tokens conditioning on text embeddings [3, 21, 32]. Diffusion models were also leveraged as the decoder for image generation, which achieves tremendous improvement for generating high quality images [23, 20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Another kind of method is to leverage Contrastive Language-Image Pre-training (CLIP) [19] models as knowledge guidance for text based image generation. This is achieved by optimizing the latent code of pre-trained generator to close the distance between generated images and input text in the shared embedding space. The optimized latent codes of generator are either in the StyleGAN latent" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7020" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.094, + 0.862, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.272, + 0.895, + 0.302 + ], + "angle": 0, + "content": "Figure 2. The mode collapse problem in the photo-to-sketch GAN adaptation by StyleGAN-NADA [5]. The trace norm of Hessian matrix is gradually decreasing, which is consistent with the visual examples showing similar patterns with mouth and eyes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.47, + 0.37 + ], + "angle": 0, + "content": "space [18] or in VQGAN codebook space [1]. Some methods also defined the parametrized space by vector graphics such as Bezier curves [4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.628 + ], + "angle": 0, + "content": "Few-shot GAN Adaptation. It aims to transfer pretrained generator to another domain when there are not enough samples to train from scratch. Its main challenge is the mode collapse problem because the generator is prone to overfit training samples in target domain and lose the diversity of the origin domain. There have been many methods to tackle this problem. They either froze most of the parameters of the pre-trained network[15] or embedded a small number of trainable parameters into the source model[16]. Recently, [17] proposed cross-domain distance consistency loss to preserve the relative similarities and differences between instances in the source domain. [30] introduced spatial structural consistency loss to align the spatial information between the synthesis image pairs of the source and target domains. These methods still require manually collected samples, and as the number of samples decreases, the mode collapse problem becomes more apparent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Besides, StyleGAN-NADA [5] further proposed to take advantage of the CLIP [19] model as knowledge guidance for GAN adaptation, and only natural language prompts are required without even a single image. Similarly, [36] used the image encoder of CLIP for one-shot adaptation. However, these methods still suffer from the mode collapse problem. In this paper, we propose the spectral consistency regularization to tackle the problem of mode collapse without hurting target generation performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Latent Space Analysis of GANs. Many works have explored the latent space of pretrained generator for image manipulation. Some methods used supervised datasets to learn directions in the latent space for attribute editing [24] or semantic image editing [13]. Other works instead applied unsupervised methods to reveal the latent space. [26] decomposed the learned weights of the pre-trained network to identify semantically meaningful directions. [7] applied principal component analysis in the latent space. Recently," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.895, + 0.432 + ], + "angle": 0, + "content": "[29] proposed to analyze the latent space of generative models from geometric point of view. They found that the eigenvectors corresponding to the largest eigenvalues of the Hessian matrix for generator dominate interpretable variations. In this paper, we analyze the GAN adaptation problem in a similar way and propose to regularize target generator by the spectrum of Hessian matrix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.461, + 0.593, + 0.477 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.492, + 0.874, + 0.509 + ], + "angle": 0, + "content": "3.1. Text-Driven Generative Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.895, + 0.673 + ], + "angle": 0, + "content": "Text-driven GAN adaptation aims to transfer a pretrained generator to target domain specified by the text description. To guide the GAN adaptation by text, pre-trained CLIP model is leveraged to measure the similarity between image and text. CLIP is a Vision-Language model trained on 400 million (image, text) pairs collected from the internet with contrastive loss [19]. One commonly used objective function for text-driven image manipulation is the global loss that optimizes the similarity between generated images and target text:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.696, + 0.892, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {g l o b a l}} = D _ {\\text {C L I P}} (G (z), t _ {\\text {t a r g e t}}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\(D_{CLIP}\\) is the cosine distance in the CLIP space, \\(t_{target}\\) is the target text. However, this only applies to in-domain image manipulation combined with identity consistency regularization[18]. This regularization is too strong for cross-domain GAN adaptation with large domain gaps like human to werewolf. Direct optimization of the above global loss leads to adversarial solutions since adding pixel-level perturbations can fool the CLIP classifier in the absence of a generative prior favoring real-image manifold [5]. To overcome this limit, the directional loss is used to optimize the direction between source and target" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "7021" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.093, + 0.887, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.333, + 0.896, + 0.405 + ], + "angle": 0, + "content": "Figure 3. Illustration of our proposed method for text-driven generative domain adaptation. The training objective of our model consists of directional loss and spectral consistency loss. We feed the same latent code to source generator and target generator and generate a pair of source and target images. The directional loss encourages the direction between embeddings of the pair to align with the semantic direction of text description. The spectral consistency loss regularizes the trace norm of Hessian matrix of target generator to prevent the mode collapse problem." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.428, + 0.138, + 0.441 + ], + "angle": 0, + "content": "domain:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.462, + 0.47, + 0.496 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i r e c t i o n}} = 1 - \\frac {\\Delta I \\cdot \\Delta T}{| \\Delta I | | \\Delta T |} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.497, + 0.47, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\Delta T = E _ {T} \\left(t _ {\\text {t a r g e t}}\\right) - E _ {T} \\left(t _ {\\text {s o u r c e}}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.517, + 0.47, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\Delta I = E _ {I} \\left(G _ {\\text {t r a i n}} (z)\\right) - E _ {I} \\left(G _ {\\text {f r o z e n}} (z)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.473, + 0.754 + ], + "angle": 0, + "content": "where \\( E_{T} \\) and \\( E_{I} \\) are the text and image encoder of CLIP, \\( t_{source} \\) and \\( t_{target} \\) are the source and target class texts. The directional loss can prevent adversarial solutions. Since target generated images should have given direction to corresponding source images, generating a single adversarial instance is impossible. At training time, the same latent code is fed into source and target generator, then the target generator is optimized using the directional loss. However, the directional loss proposed by StyleGAN-NADA still suffers from the mode collapse problem, as shown in the images of Figure 2. So we propose the spectral consistency regularization derived from geometry analysis of GANs to resolve this problem." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.781, + 0.416, + 0.798 + ], + "angle": 0, + "content": "3.2. Geometry Analysis of GAN Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.9 + ], + "angle": 0, + "content": "We denote the generative network as a mapping from latent code \\( z \\) to a manifold in image space as \\( G(z) \\). Considering a squared distance function \\( d^2 \\) for two images, we express the local variations of \\( G(z) \\) from moving towards direction \\( \\Delta z \\) by second-order Taylor expansion. This is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.452, + 0.741, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\lim _ {\\Delta z \\to 0} d ^ {2} (G (z), G (z + \\Delta z)) =\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.477, + 0.863, + 0.508 + ], + "angle": 0, + "content": "\\[\nd ^ {2} (G (z), G (z)) + \\frac {\\partial d ^ {2} (G (z) , G (z + \\Delta z))}{\\partial \\Delta z} \\cdot \\Delta z +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.51, + 0.892, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\Delta z ^ {T} \\cdot \\frac {\\partial^ {2} d ^ {2} (G (z) , G (z + \\Delta z))}{\\partial \\Delta z ^ {2}} \\cdot \\Delta z \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.549, + 0.893, + 0.732 + ], + "angle": 0, + "content": "The first two terms are zero since \\(d^{2}(G(z), G(z + \\Delta z))\\) is local minima when \\(\\Delta z = 0\\). Denote the second derivatives as Hessian matrix \\(H(z)\\), and we have \\(d^{2}(G(z), G(z + \\Delta z)) = \\Delta z^{T}H(z)\\Delta z\\). For a normalized vector \\(\\Delta z\\), we can conclude that \\(\\sigma_{min} \\leq d^{2}(G(z), G(z + \\Delta z)) \\leq \\sigma_{max}\\) where \\(\\sigma_{min}\\) and \\(\\sigma_{max}\\) are the smallest and largest eigenvalues of \\(H\\). Thus, we can use the trace norm of \\(H_{z}\\) to reflect the statistics of diversity in generative models, which is the sum of all eigenvalues of \\(H\\). Especially, for a squared \\(L2\\) distance function in metric space \\(\\phi\\), \\(d^{2}(z_{1}, z_{2})_{\\phi} = \\frac{1}{2}\\|\\phi(G(z_{1})) - \\phi(G(z_{2}))\\|^{2}\\), the Hessian matrix \\(H_{\\phi}(z_{0})\\) is a simple transformation from the Jacobian \\(J_{\\phi}(z_{0})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.741, + 0.891, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} H _ {\\phi} \\left(\\boldsymbol {z} _ {0}\\right) = \\frac {\\partial^ {2}}{\\partial \\boldsymbol {z} ^ {2}} \\frac {1}{2} \\| \\phi \\left(\\boldsymbol {z} _ {0}\\right) - \\phi (\\boldsymbol {z}) \\| _ {2} ^ {2} (6) \\\\ = J _ {\\phi} \\left(\\boldsymbol {z} _ {0}\\right) ^ {T} J _ {\\phi} \\left(\\boldsymbol {z} _ {0}\\right) (7) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.802, + 0.893, + 0.849 + ], + "angle": 0, + "content": "where the top eigenvectors of \\(H_{\\phi}\\) correspond to right singular vectors of the Jacobian \\(J_{\\phi}\\). Besides, the Hessian-vector product is related to Jacobian-vector product as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.868, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} ^ {T} H _ {\\phi} (\\boldsymbol {z} _ {0}) \\boldsymbol {v} = \\| J _ {\\phi} (\\boldsymbol {z} _ {0}) \\boldsymbol {v} \\| ^ {2}, \\quad J _ {\\phi} (\\boldsymbol {z}) = \\frac {\\partial \\phi (G (\\boldsymbol {z}))}{\\partial z} \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7022" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "To analyze the mode collapse problem, we calculate the statistics of Hessian trace from different samples of \\( z \\) during adaptation process of the previous state-of-the-art method. The results of GAN adaptation from photo to sketch with StyleGAN-NADA [5] is shown in Figure 2. We find that as the iteration count increases, the mode collapse problem becomes more severe and the Hessian trace is also decreasing, which proves that the Hessian trace can reflect the diversity of generator. In the early stage of adaptation, the decrease of trace norm is mainly caused by style adaptation since there are no color variations for sketch domain. But for the late stage from 200th step to 400th step, the style effect changes little and the face attributes tend to have similar pattern. Since the target text only represents a fixed direction without variation, different samples are encouraged to approach the same fixed pattern. We also give a proof of the relationship between Hessian Trace and model diversity in the following section." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.373, + 0.466, + 0.389 + ], + "angle": 0, + "content": "3.3. Relation between Hessian Trace and Diversity" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.396, + 0.47, + 0.485 + ], + "angle": 0, + "content": "Considering the mapping from standard normal distribution \\(z \\sim \\mathcal{N}(0, I)\\) to general multivariate normal distribution \\(y \\sim \\mathcal{N}(\\mu, \\Sigma)\\) with the generator as a linear function \\(G(z) = A z + b\\), which has \\(\\mu = b, \\Sigma = A A^T\\). The linear assumption of generator holds when we consider a small neighborhood around \\(z\\), i.e. \\(\\lim \\Delta z \\to 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.487, + 0.353, + 0.501 + ], + "angle": 0, + "content": "The variance of \\(y\\) can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.513, + 0.468, + 0.717 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} V a r (y) = E [ \\| y - E [ y ] \\| _ {2} ^ {2} ] \\\\ = E \\left[ \\sum_ {i = 1} ^ {n} \\left(y _ {i} - E \\left[ y _ {i} \\right]\\right) ^ {2} \\right] \\\\ = \\sum_ {i = 1} ^ {n} E \\left[ \\left(y _ {i} - E [ y _ {i} ]\\right) ^ {2} \\right] \\\\ = \\sum_ {i = 1} ^ {n} V a r (y _ {i}) \\\\ = \\sum_ {i = 1} ^ {n} \\Sigma_ {i i} \\\\ = \\operatorname {T r a c e} (\\Sigma) = \\operatorname {T r a c e} \\left(A A ^ {T}\\right) \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.728, + 0.469, + 0.757 + ], + "angle": 0, + "content": "On the other side, the \\(d^2 (G(z),G(z + \\Delta z))\\) can be expanded as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.769, + 0.468, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} d ^ {2} (G (z), G (z + \\Delta z)) = \\| G (z) - G (z + \\Delta z) \\| _ {2} ^ {2} \\\\ = \\| A z + b - (A (z + \\Delta z) + b) \\| _ {2} ^ {2} \\\\ = \\| A \\Delta z \\| \\\\ = \\Delta z ^ {T} A ^ {T} A \\Delta z \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "such the Hessian Matrix \\(H\\) of \\(\\Delta z\\) with respect to \\(d^{2}(G(z), G(z + \\Delta z))\\) equals to \\(A^{T}A\\), e.g. \\(H = A^{T}A\\). Combining above two equations with \\(Trace(AA^{T}) =\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.182 + ], + "angle": 0, + "content": "Trace \\((A^TA)\\), we have \\(Var(y) = Trace(H)\\), which means that the Hessian Trace for every sample in target distribution reflects the variance and diversity of this distribution. If the Hessian Trace is small, the target distribution only spans a small region in space. This is consistent to the mode collapse problem in generative models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.19, + 0.816, + 0.206 + ], + "angle": 0, + "content": "3.4. Spectral Consistency Regularization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.272 + ], + "angle": 0, + "content": "To prevent the target generator from mode collapse, we propose spectral consistency regularization to prevent the diversity of generator from degrading, which is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.282, + 0.891, + 0.298 + ], + "angle": 0, + "content": "\\[\nL _ {r e g} = \\left\\| \\operatorname {T r a c e} \\left(H _ {s} (z)\\right) - \\operatorname {T r a c e} \\left(H _ {t} (z)\\right) \\right\\| \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.424 + ], + "angle": 0, + "content": "where \\( H_{s}(z) \\) and \\( H_{t}(z) \\) are the Hessian matrix of the source generator and target generator evaluated with the same latent code \\( z \\). However, directly computing the Hessian matrix requires backpropagation \\( n \\) times where \\( n \\) is the dimension of feature vector in metric space. Instead, we use the Hutchinson's method for trace estimator [9] to compute a stochastic estimator of Hessian Trace, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.432, + 0.891, + 0.449 + ], + "angle": 0, + "content": "\\[\n\\operatorname {T r a c e} (H (z)) = \\mathbb {E} [ v ^ {T} H (z) v ] = \\mathbb {E} [ \\| J _ {\\phi} (z) v \\| ^ {2} ] \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.891, + 0.517 + ], + "angle": 0, + "content": "Where \\( v \\sim \\mathrm{Rademacher}(\\frac{1}{2}) \\), and the second transformation is derived from Equation 8. So the calculation of Hessian matrix is transformed into the calculation of Jacobian-Vector product." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.623 + ], + "angle": 0, + "content": "Different from the within-domain consistency loss[36] which restricts target generated samples based on relative difference of source samples, the spectral consistency regularization only cares about the diversity of target model. This doesn't impose restriction on the direction of target adaptation, so our method can generate samples more consistent with target text without losing model diversity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.809 + ], + "angle": 0, + "content": "The training objective of text-driven GAN adaptation is a weighted combination of directional loss and spectral consistency regularization loss \\(\\mathcal{L} = \\mathcal{L}_{dir} + \\lambda \\mathcal{L}_{reg}\\). Since different target domains have their own characteristic, it is required to tune the hyperparameter \\(\\lambda\\) for better performance. To prevent the exhausting hyperparameter searching, we propose an adaptive loss reweighting method to balance the influence of these two loss items. Specifically, the adaptive weight \\(\\lambda\\) is calculated as \\(\\lambda_{spectral} \\frac{\\|\\nabla_{G_L} \\mathcal{L}_{dir}\\|}{\\|\\nabla_{G_L} \\mathcal{L}_{reg}\\|}\\), where \\(G_L\\) denotes the last layer of generator and \\(\\lambda_{spectral}\\) is a manually specified hyperparameter, typically 1.0 is an appropriate choice." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.822, + 0.834 + ], + "angle": 0, + "content": "3.5. Granularity Adaptive Regularization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For text-driven GAN adaptation problem, the adaptation granularity of different target domains varies from texture to structure. For example, the photo-to-sketch adaptation mainly focuses on the appearance and texture change," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7023" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.885, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.397, + 0.895, + 0.44 + ], + "angle": 0, + "content": "Figure 4. Visual results of our method and state-of-the-art StyleGAN-NADA [5] for different GAN adaptations. The top row shows the source text and target text description. The left column presents samples generated by source generator. Our method not only generates samples described by target text with high fidelity, but also produces diverse and identity consistent images corresponding to source domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.47, + 0.553 + ], + "angle": 0, + "content": "while the werewolf domain has more variations on semantic structure. Regularization of diversity in the whole granularity will restrict the adaptation performance. To alleviate this problem, we propose granularity adaptive regularization based on the disentangled characteristic of StyleGAN latent code." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.473, + 0.751 + ], + "angle": 0, + "content": "The latent codes injected into different layers in StyleGAN influence different granularities. The style code in low resolution represents high-level aspects such as pose and shape, that in middle resolution controls facial features and hairstyle, and that in high resolution influences color scheme and local textures. Specifically, we use the \\( W+ \\) space as input space for Jacobian matrix calculation, where each input latent code consists of 18 512-dimensional vectors so both \\( z \\) and \\( v \\) in Equation 4 are in \\( R^{18 \\times 512} \\). By masking \\( v \\) with a mask vector \\( m \\in \\{0,1\\}^{18} \\) for different layers, we can specify the granularity of variations involved in the regularization. The calculation of Hessian trace under mask is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.762, + 0.47, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {T r a c e} (H (z)) = \\mathbb {E} [ (v \\odot m) ^ {T} H (z) (v \\odot m) ] (13) \\\\ = \\mathbb {E} [ \\| J _ {\\phi} (z) (v \\odot m) \\| ^ {2} ]. (14) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Following previous convention, we divide the style code for 18 layers into 3 groups, which are for coarse, middle and fine scale. The results of different mask strategies are shown in Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To explore the best mask strategy for different target domains, we propose to use an adaptive soft mask vector" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.463, + 0.892, + 0.555 + ], + "angle": 0, + "content": "\\(\\{\\tilde{m} |\\tilde{m}\\in R^{18},\\| \\tilde{m}\\| = 1\\}\\) for all layers. During training, the mask vector is optimized with respect to the overall training objective. To reduce the directional loss, the mask vector will assign less value to the latent code corresponding to the granularity that changes most, while other values will increase to preserve the diversity of source generator." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.572, + 0.625, + 0.589 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.794 + ], + "angle": 0, + "content": "In this section, we will show the qualitative and quantitative results of our method. We illustrate the generated results for a wide range of target domains from style and texture changes to shape and semantic modifications. We also compare the proposed spectral consistency regularization with other regularization methods for diversity preservation. Next, we perform an ablation study on our method to evaluate the effectiveness of each component. Finally, we demonstrate the applications of text-guidance GAN adaptation, including image-to-image translation and image editing. To verify the generalization of proposed spectral consistency regularization, we also apply it to one-shot GAN adaptation, which is shown in supplementary appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We use the StyleGANv2 [11] generator pretrained on FFHQ as the source generator. During domain adaptation, we optimize all the parameters of generator except for the mapping network and toRGB layers. We use the Adam [12] optimizer with learning rate 0.001. The \\(\\lambda_{\\mathrm{spectral}}\\) is set to 1.0. For most target domains, only 300 iterations are required to achieve convergence. Following CLIP, we use 79" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7024" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.885, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.469, + 0.893, + 0.499 + ], + "angle": 0, + "content": "Figure 5. Comparison results of our spectral consistency regularization with other regularization methods. Detailed explanations of these methods are in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.569 + ], + "angle": 0, + "content": "manually designed prompts like \"a photo of a...\" with provided target domain description and feed them to text encoder of CLIP to get the embeddings of target domain." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.576, + 0.27, + 0.592 + ], + "angle": 0, + "content": "4.1. Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Qualitative Comparison In Figure 4, we show the comparison results of our method with state-of-the-art model StyleGAN-NADA [5] for a wide range of target domains, which varies from texture changes like sketch and Mona Lisa paintings to geometric change like werewolf and Pixar style. The results demonstrate that our method not only generates highly stylistic images consistent with target text description for different target domains, but also produces images with diversity inherited from the pretrained source generator. Compared with StyleGAN-NADA which has obvious mode collapse problem like the mouth pattern in sketch and hair in werewolf, our model generates target images with better identity consistency. This proves that the spectral consistency regularization can preserve the diversity of source domain. In the supplementary appendix, we present additional visual results for the dogs and cars domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We also perform text-driven GAN adaptation experiments with other regularization methods, including the Selective Cross-modal Consistency (SCC) loss[35], Within-Domain Consistency (WDC) loss[36], the Mode" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.688 + ], + "angle": 0, + "content": "Seeking (MS) loss[14], Perceptual Path Length (PPL) regularization[11] and Cross Domain Correspondence (CDC) regularization[17]. Detailed explanations of these regularization methods are in supplementary appendix. As shown in Figure 5, SCC, WDC and MS regularization impose too strong regularization to target generator and restrict the domain specific attributes for target domain. Suffering from mode collapse problem, the generation results of PPL and CDC exhibits the same pattern across different samples. In comparison, our method has a better balance between the diversity and stylization of target generator." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Quantitative Comparison Besides our proposed trace norm of Hessian Matrix in Equation 3, we also leverage the Perceptual Path Length(PPL) [10] for quantitative diversity comparison. It measures the perceptually-based pairwise image distance [34] for a linear interpolation path in the latent path. The average PPL in the latent space \\(\\mathcal{Z}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.484, + 0.786, + 0.925, + 0.82 + ], + "angle": 0, + "content": "\\[\nP P L _ {\\mathcal {Z}} = \\mathbb {E} \\left[ \\right. \\frac {1}{\\epsilon^ {2}} d \\left( \\right.G \\left(\\operatorname {s l e r p} \\left(\\mathbf {z} _ {1}, \\mathbf {z} _ {2}; t\\right)\\right), G \\left(\\operatorname {s l e r p} \\left(\\mathbf {z} _ {1}, \\mathbf {z} _ {2}; t + \\epsilon\\right)\\right)\\left. \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( d(.,.) \\) evaluates the perceptual distance between two images, and slerp denotes spherical interpolation since the input latent is normalized. The perceptual path length estimates the diversity of the generator via finite differences, and our method estimates the diversity analytically." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7025" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.087, + 0.895, + 0.117 + ], + "angle": 0, + "content": "Table 1. Comparison results for diversity by PPL and Hessian trace between different regularization methods. The PPL and Hessian trace for the pretrained source generator is 419.22 and 0.309." + }, + { + "type": "table", + "bbox": [ + 0.173, + 0.126, + 0.803, + 0.272 + ], + "angle": 0, + "content": "
ResultsPhoto → SketchPhoto → Mona LisaHuman → WerewolfPhoto → Pixar
PPLTracePPLTracePPLTracePPLTrace
SCC547.340.526485.700.521428.450.343440.170.535
WDC378.250.090363.170.191311.340.116345.310.167
MS466.990.167331.060.233352.610.191377.880.358
PPL241.500.028209.090.061297.740.062300.190.063
CDC348.010.062259.920.079351.730.089299.490.111
NADA323.250.061281.590.098302.910.101343.540.112
Ours463.570.116321.430.140383.190.137353.800.181
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.304, + 0.47, + 0.331 + ], + "angle": 0, + "content": "Table 2. User preference study. The numbers represent the percentage of users who prefer the images synthesized by our method." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.332, + 0.446, + 0.442 + ], + "angle": 0, + "content": "
Model ComparisonQualityStyleAttributes
Ours vs. NADA [5]86.4%59.4%91.2%
Ours vs. SCC [35]64.2%86.2%65.0%
Ours vs. MS [14]59.4%83.6%54.2%
Ours vs. WDC [36]62.2%89.8%49.8%
Ours vs. PPL [11]92.4%71.4%95.0%
Ours vs. CDC [17]79.6%60.6%87.8%
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.686 + ], + "angle": 0, + "content": "The quantitative results are shown in Table 1. The mentioned strong regularization methods including SCC, WDC and MS have large values of PPL and Hessian trace at the cost of restricting adaptation effects. Our method performs competitive with these methods in diversity metrics, but does much better in transferring target styles. Compared with CDC and PPL, our method better preserve the diversity of source domain reflected in PPL and Hessian Trace. For all different target domains, our method outperforms previous state-of-the-art model StyleGAN-NADA for both PPL and Hessian trace. This benefits from that our method can alleviate the mode collapse problem with spectral consistency regularization and generate more diverse images without restricting the style adaptation results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.903 + ], + "angle": 0, + "content": "User Study We conducted a user study to compare with other regularization methods. Specifically, participants were requested to select the superior synthesized samples in relation to three measurements: (1) image quality, (2) style consistency with the target description (3) attribute consistency with the source image. Five hundred samples were randomly generated for each comparative analysis. The aggregated results are presented in Table 2. The users strongly favor our method across all three aspects. In contrast to other methods, our method demonstrates a better tradeoff between style effects and attribute preservation. The spectral consistency regularization can better preserve the diversity of source domain while not restricting the style effects for target domain." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.307, + 0.655, + 0.324 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.895, + 0.528 + ], + "angle": 0, + "content": "Granularity Adaptive Regularization. In Figure 6, we demonstrate generated results with different regularization strategies applied in \\( W+ \\) space in Section 3.4. Regularization to the coarse scale will preserve the structure of source image but the diversity of fine features like hair will lose. The global regularization performs similarly to coarse regularization since the coarse features dominate the diversity of generator. Only applying regularization to the fine scale will generate high-frequency textures and the structure characteristic like necks will collapse. In comparison, our proposed granularity adaptive regularization both preserves the diversity of source domain in all scales and matches the styles of target domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.529, + 0.895, + 0.62 + ], + "angle": 0, + "content": "Strength of regularization. In Figure 7, we show the generated samples with a linear interpolated loss weight \\(\\lambda_{\\text{spectral}}\\). We can observe that with increasing \\(\\lambda_{\\text{spectral}}\\), the generated samples maintain more diversity of source domain, and they also illustrate the most significant characteristic of target domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.895, + 0.789 + ], + "angle": 0, + "content": "Choices of metric space. We conduct experiments about the metric space of spectral consistency regularization with different feature encoder \\(\\phi(x)\\) that evaluates the distance between image samples. Besides CLIP [19] image encoder in our method, we also leverage VGG [27] which was commonly used in style transfer [6], MoCo [8] of contrastive representation learning, ArcFace [2] for face recognition and the plain pixel space. As shown in Figure 8, compared to other feature encoders, the spectral consistency regularization with CLIP encoder shows best performance for preserving the identity of source image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.636, + 0.817 + ], + "angle": 0, + "content": "4.3. Applications" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Image-to-Image Translation We combine the adapted generator in target domain with a GAN inversion encoder to implement image-to-image translation. Given a real-world image, we invert it to the latent code in \\( W \\) space via an e4e encoder [28], which is then fed to target generator to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7026" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.091, + 0.172, + 0.105 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.116, + 0.207, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.092, + 0.31, + 0.104 + ], + "angle": 0, + "content": "Global" + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.116, + 0.343, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.092, + 0.449, + 0.104 + ], + "angle": 0, + "content": "Coarse" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.116, + 0.48, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.091, + 0.59, + 0.105 + ], + "angle": 0, + "content": "Middle" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.116, + 0.617, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.092, + 0.712, + 0.105 + ], + "angle": 0, + "content": "Fine" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.116, + 0.754, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.092, + 0.866, + 0.107 + ], + "angle": 0, + "content": "Adaptive" + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.116, + 0.892, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.22, + 0.821, + 0.235 + ], + "angle": 0, + "content": "Figure 6. Results of different regularization strategies for GAN adaptation from photo to Edvard Munch paintings." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.256, + 0.47, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.36, + 0.469, + 0.388 + ], + "angle": 0, + "content": "Figure 7. Generation results of linearly interpolated \\(\\lambda_{\\mathrm{spectral}}\\) in regularization." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.404, + 0.462, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.495, + 0.468, + 0.523 + ], + "angle": 0, + "content": "Figure 8. The generated results with spectral consistency regularization under the metric spaces defined by different encoders." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.469, + 0.594 + ], + "angle": 0, + "content": "produce target image. As shown in Figure 9, our method can achieve high-quality image translation and preserve the identity of source image for different domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.469, + 0.7 + ], + "angle": 0, + "content": "Image Editing In Figure 10, we demonstrate the image editing results performed on the target domain. We leverage the meaningful directions found by InterfaceGAN [25] to edit target images. We can observe that the editing directions from source domain still apply to target domains, which proves that the target domain preserves the semantic distribution of source domain." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.713, + 0.462, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.843, + 0.469, + 0.899 + ], + "angle": 0, + "content": "Figure 9. Image translation of real-world images to different target domains. Each column shows a target domain, and the top row is the text description for target domain. The transferred images represent both target style and the identity of source image." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.257, + 0.892, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.31, + 0.891, + 0.338 + ], + "angle": 0, + "content": "Figure 10. Editing images in target domain for real-world images. The top row shows the edited attributes." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.374, + 0.618, + 0.39 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.537 + ], + "angle": 0, + "content": "In this paper, we propose the spectral consistency regularization for text-driven generative domain adaptation. The key insight of our method is to build a quantitative diversity estimator to preserve the intra-domain diversity of source generator without restricting the adaptation of target style. We also introduce an adaptive regularization strategy for granularity-flexible adaptation. The experiments demonstrate our method greatly improves the generation results for a wide range of target domains." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.555, + 0.661, + 0.571 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.672 + ], + "angle": 0, + "content": "This work was supported in part by the National Key R&D Program of China under Grant 2018AAA0102000, in part by National Natural Science Foundation of China: 62322211, 62236008, U21B2038, 62225207, U19B2038 and 61931008, and Youth Innovation Promotion Association of Chinese Academy of Sciences under Grant 2020108." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.689, + 0.597, + 0.705 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.715, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[1] Katherine Crowson, Stella Rose Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. ArXiv, abs/2204.08583, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.787, + 0.894, + 0.843 + ], + "angle": 0, + "content": "[2] Jiankang Deng, J. Guo, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4685-4694, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.845, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[3] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868-12878, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.715, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7027" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[4] Kevin Frans, Lisa B. Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. ArXiv, abs/2106.14843, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.136, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[5] Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ArXiv, abs/2108.00946, 2021. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.192, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[6] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.249, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[7] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. ArXiv, abs/2004.02546, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.292, + 0.471, + 0.36 + ], + "angle": 0, + "content": "[8] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.362, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[9] Michael F. Hutchinson. A stochastic estimator of the trace of the influence matrix for laplacian smoothing splines. Communications in Statistics - Simulation and Computation, 18:1059-1076, 1989. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.419, + 0.471, + 0.475 + ], + "angle": 0, + "content": "[10] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4396-4405, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.476, + 0.471, + 0.545 + ], + "angle": 0, + "content": "[11] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8107-8116, 2020. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.547, + 0.471, + 0.574 + ], + "angle": 0, + "content": "[12] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.576, + 0.471, + 0.617 + ], + "angle": 0, + "content": "[13] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. ArXiv, abs/2111.03186, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.618, + 0.471, + 0.687 + ], + "angle": 0, + "content": "[14] Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Siwei Ma, and Ming-Hsuan Yang. Mode seeking generative adversarial networks for diverse image synthesis. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1429-1437, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.689, + 0.471, + 0.731 + ], + "angle": 0, + "content": "[15] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze discriminator: A simple baseline for fine-tuning gans. ArXiv, abs/2002.10964, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.732, + 0.471, + 0.787 + ], + "angle": 0, + "content": "[16] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2750-2758, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.471, + 0.87 + ], + "angle": 0, + "content": "[17] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A. Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10738-10747, 2021. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.872, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[18] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and D. Lischinski. Styleclip: Text-driven manipulation of" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "stylegan imagery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2065-2074, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[20] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[21] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.348 + ], + "angle": 0, + "content": "[22] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[23] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[24] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9240-9249, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[25] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:2004-2018, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[26] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1532-1540, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[27] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[28] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40:1 - 14, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[29] Binxu Wang and Carlos R. Ponce. The geometry of deep generative image models and its applications. ArXiv, abs/2101.06006, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[30] Jiayu Xiao, Liang Li, Chaofei Wang, Zhengjun Zha, and Qingming Huang. Few shot generative model adaption via relaxed spatial structural alignment. ArXiv, abs/2203.04121, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[31] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316-1324, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7028" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.187 + ], + "angle": 0, + "content": "[32] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Benton C. Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. ArXiv, abs/2206.10789, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[33] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N. Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5908-5916, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[34] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.346, + 0.47, + 0.4 + ], + "angle": 0, + "content": "[35] Yabo Zhang, Mingshuai Yao, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, and Wangmeng Zuo. Towards diverse and faithful one-shot adaption of generative adversarial networks. ArXiv, abs/2207.08736, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.402, + 0.47, + 0.457 + ], + "angle": 0, + "content": "[36] Peihao Zhu, Rameen Abdal, John C. Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. ArXiv, abs/2110.08398, 2021. 1, 2, 3, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7029" + } + ] +] \ No newline at end of file diff --git a/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/e13295e4-a033-4d38-a3a5-88f3b3385e81_origin.pdf b/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/e13295e4-a033-4d38-a3a5-88f3b3385e81_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8d7270e475e7e28ff86d5ccdcc6c94b21b03bf18 --- /dev/null +++ b/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/e13295e4-a033-4d38-a3a5-88f3b3385e81_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27600122bce064278615212fc78e28fa5f111a1755e6b3f5b902ac5d6dd5e123 +size 3213144 diff --git a/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/full.md b/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a15028eb9e1677eae09eb0e4f86901d08c915ee9 --- /dev/null +++ b/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/full.md @@ -0,0 +1,328 @@ +# Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization + +Zhenhuan Liu $^{1,2}$ Liang Li $^{1*}$ Jiayu Xiao $^{1,2}$ Zheng-Jun Zha $^{3}$ Qingming Huang $^{1,2,4}$ $^{1}$ Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, CAS + $^{2}$ University of Chinese Academy of Sciences + $^{3}$ University of Science and Technology of China + $^{4}$ Peng Cheng Laboratory + +{zhenhuan.liu, jiayu.xiao}@vip1.ict.ac.cn, liang.li@ict.ac.cn, zhazj@ustc.edu.cn, qmhuang@ucas.ac.cn + +# Abstract + +Combined with the generative prior of pre-trained models and the flexibility of text, text-driven generative domain adaptation can generate images from a wide range of target domains. However, current methods still suffer from overfitting and the mode collapse problem. In this paper, we analyze the mode collapse from the geometric point of view and reveal its relationship to the Hessian matrix of generator. To alleviate it, we propose the spectral consistency regularization to preserve the diversity of source domain without restricting the semantic adaptation to target domain. We also design granularity adaptive regularization to flexibly control the balance between diversity and stylization for target model. We conduct experiments for broad target domains compared with state-of-the-art methods and extensive ablation studies. The experiments demonstrate the effectiveness of our method to preserve the diversity of source domain and generate high fidelity target images. Source code has been released in https://github.com/Victarry/Adaptation-SCR. + +# 1. Introduction + +Generative image modeling has developed significantly in recent years and is able to generate diverse high-resolution images even indistinguishable from real images. However, training such models requires intense computation resources and large datasets, which restricts the application scope of generative models. For some scenarios, collecting large datasets is impossible like paintings by specific artists. Benefiting from Vision-Language models learning from large image-text pairs, text can be leveraged as a de + +scription of abstract visual semantics to guide generative domain adaptation instead of a collection of image samples in target domain. As an expressive representation, text has shown great success in semantic image generation and manipulation recently [23, 20]. Based on the generative prior of pre-trained models and flexible text description of target domain, text-driven generative domain adaptation can generate more various images and have promising applications. + +To reduce the requirement of training samples, traditional methods propose to train generative models in the target domain with only limited samples by adapting pretrained models in the large-scale source domain which contains high-level semantic knowledge as a generative prior. These few-shot adaptation methods either finetune only a part of parameters within networks to preserve most source domain knowledge [15] or impose strong regularization on the generated images [30, 36]. However, these methods still require additional training samples of target domain and adversarial training process. As the number of samples drops, the image fidelity and diversity also hurt severely. Different from these methods, text-driven generative domain adaptation requires no image samples but texts to describe the target domain. Pioneer work [5] proposed to encourage the visual change between samples from target and source generators to align with semantic direction described by text in the CLIP [19] embedding space, which achieves generative adaptation for miscellaneous domains in short training time. + +The main challenge of text-driven GAN adaptation is the mode collapse problem due to the entanglement of intradomain semantics and inter-domain style in text representation. Besides the specified target style described by text, there also exists an unknown pattern for the semantics of images. This leads to a decrease of variations in generated images when the style effect is optimized to approach target domain. As shown in Figure 2, while the number of qua + +![](images/515369480a1c248fea0c85684a4156da85c52a1ed05fa55b0c61b4b6e3569dda.jpg) +Original + +![](images/ea0686e26676b153e03298514bfa2bbce6bfdf29f753b30c5228666289dc52e4.jpg) +Sketch + +![](images/e940d699410540f4fa44b31774cac8f327547f6703362a3200867b143a59b04a.jpg) +Ukiyo-e +Figure 1. Text-driven generative domain adaptation with various text descriptions. The generated samples should both reflect characteristics of target domain from text and preserve the original identity. + +![](images/2a66b4c5f146b89da43351e1f4f483152dfd6a0b41c56c9a2f5892acea894292.jpg) +Van Gogh + +![](images/97d84f6d91337b724cb3295cf008a74389dd98076f1b6b2630abc45981073a59.jpg) +Werewolf + +tions increases, the generated samples tend to have similar patterns of mouth and eyes, which reduces most of the variations in the origin model. The main reason for the mode collapse problem is that the optimization process only cares about the distance of generated samples to target domains, and the intra-domain feature variations are easily ignored. + +To address the above challenge, researchers [36] proposed to preserve the diversity of source domain through a within-domain consistency loss which keeps consistency between sample changes in source domain and target domain. However, this regularization is too strong to restrict the style effect of target generator close to source domain. The previous theorem about GAN latents analysis has shown that the Hessian matrix of generator reflects the variations of generator and can be used to explore meaningful directions from top eigenvectors. Inspired by this, we try to leverage the spectrum of Hessian matrix as a quantitative evaluation of model diversity in the adaptation problem. This disentangles the relative diversity between generated samples from absolute generative distribution and makes a general way to regularize diversity of generative model. + +In this work, we propose spectral consistency regularization to solve the problem of mode collapse in text-driven generative domain adaptation from the geometric point of view. First, we analyze the Hessian matrix of generator's manifold in the metric space by eigendecomposition. The eigenvalues of Hessian matrix are decreasing in the adaptation process, which is consistent with the mode collapse problem of visual observations. Second, we introduce the spectral consistency regularization on the Hessian matrix to prevent the latent space of generator from degrading. This regularization helps preserve intra-domain variations of source domain without restricting style effects of target generators. We further develop a stochastic method to regularize the spectrum of Hessian matrix without calculating the full matrix, which reduces the expensive computational cost. Finally, we design the granularity adaptive regularization considering the layer-decomposition characteristic of $\mathrm{W + }$ space in StyleGAN. + +The contributions of this paper are summarized below: + +1. We analyze the commonly occurred mode collapse problem in GAN adaptation from the geometric point of view and provide a quantitative evaluation of model diversity to reveal the reason of mode collapse. +2. We propose a spectral consistency regularization for text-driven GAN adaptation, which preserves diversity of original domain and generates high fidelity images of target domain. A granularity adaptive regularization is further designed to flexibly control balance between diversity and stylization for target model. +3. We conduct experiments and ablation studies for a wide range of target domains. The experiments show the effectiveness of our proposed spectral consistency regularization and its applications to downstream tasks like image editing and image-to-image translation. + +# 2. Related Work + +Text-driven Image Synthesis and Manipulation. Traditional methods approached text-driven image generation by training a conditional GAN[22]. Several following works have been proposed to improve generation quality either by multi-scale networks [33] or attention mechanism [31]. Recently, transformer-based auto-regressive generative models were introduced to view text-driven image synthesis as conditional sequence generation of visual tokens conditioning on text embeddings [3, 21, 32]. Diffusion models were also leveraged as the decoder for image generation, which achieves tremendous improvement for generating high quality images [23, 20]. + +Another kind of method is to leverage Contrastive Language-Image Pre-training (CLIP) [19] models as knowledge guidance for text based image generation. This is achieved by optimizing the latent code of pre-trained generator to close the distance between generated images and input text in the shared embedding space. The optimized latent codes of generator are either in the StyleGAN latent + +![](images/58d312ec650c0316d4273c2e4caefcb18cf80c4a38ff0607f270bb8f414cf9c5.jpg) +Figure 2. The mode collapse problem in the photo-to-sketch GAN adaptation by StyleGAN-NADA [5]. The trace norm of Hessian matrix is gradually decreasing, which is consistent with the visual examples showing similar patterns with mouth and eyes. + +space [18] or in VQGAN codebook space [1]. Some methods also defined the parametrized space by vector graphics such as Bezier curves [4]. + +Few-shot GAN Adaptation. It aims to transfer pretrained generator to another domain when there are not enough samples to train from scratch. Its main challenge is the mode collapse problem because the generator is prone to overfit training samples in target domain and lose the diversity of the origin domain. There have been many methods to tackle this problem. They either froze most of the parameters of the pre-trained network[15] or embedded a small number of trainable parameters into the source model[16]. Recently, [17] proposed cross-domain distance consistency loss to preserve the relative similarities and differences between instances in the source domain. [30] introduced spatial structural consistency loss to align the spatial information between the synthesis image pairs of the source and target domains. These methods still require manually collected samples, and as the number of samples decreases, the mode collapse problem becomes more apparent. + +Besides, StyleGAN-NADA [5] further proposed to take advantage of the CLIP [19] model as knowledge guidance for GAN adaptation, and only natural language prompts are required without even a single image. Similarly, [36] used the image encoder of CLIP for one-shot adaptation. However, these methods still suffer from the mode collapse problem. In this paper, we propose the spectral consistency regularization to tackle the problem of mode collapse without hurting target generation performance. + +Latent Space Analysis of GANs. Many works have explored the latent space of pretrained generator for image manipulation. Some methods used supervised datasets to learn directions in the latent space for attribute editing [24] or semantic image editing [13]. Other works instead applied unsupervised methods to reveal the latent space. [26] decomposed the learned weights of the pre-trained network to identify semantically meaningful directions. [7] applied principal component analysis in the latent space. Recently, + +[29] proposed to analyze the latent space of generative models from geometric point of view. They found that the eigenvectors corresponding to the largest eigenvalues of the Hessian matrix for generator dominate interpretable variations. In this paper, we analyze the GAN adaptation problem in a similar way and propose to regularize target generator by the spectrum of Hessian matrix. + +# 3. Method + +# 3.1. Text-Driven Generative Domain Adaptation + +Text-driven GAN adaptation aims to transfer a pretrained generator to target domain specified by the text description. To guide the GAN adaptation by text, pre-trained CLIP model is leveraged to measure the similarity between image and text. CLIP is a Vision-Language model trained on 400 million (image, text) pairs collected from the internet with contrastive loss [19]. One commonly used objective function for text-driven image manipulation is the global loss that optimizes the similarity between generated images and target text: + +$$ +\mathcal {L} _ {\text {g l o b a l}} = D _ {\text {C L I P}} (G (z), t _ {\text {t a r g e t}}) \tag {1} +$$ + +where $D_{CLIP}$ is the cosine distance in the CLIP space, $t_{target}$ is the target text. However, this only applies to in-domain image manipulation combined with identity consistency regularization[18]. This regularization is too strong for cross-domain GAN adaptation with large domain gaps like human to werewolf. Direct optimization of the above global loss leads to adversarial solutions since adding pixel-level perturbations can fool the CLIP classifier in the absence of a generative prior favoring real-image manifold [5]. To overcome this limit, the directional loss is used to optimize the direction between source and target + +![](images/a7ed1545cd85801af19d50d9c595ebec69273f370eef39f6dcb24c78312d2d3f.jpg) +Figure 3. Illustration of our proposed method for text-driven generative domain adaptation. The training objective of our model consists of directional loss and spectral consistency loss. We feed the same latent code to source generator and target generator and generate a pair of source and target images. The directional loss encourages the direction between embeddings of the pair to align with the semantic direction of text description. The spectral consistency loss regularizes the trace norm of Hessian matrix of target generator to prevent the mode collapse problem. + +domain: + +$$ +\mathcal {L} _ {\text {d i r e c t i o n}} = 1 - \frac {\Delta I \cdot \Delta T}{| \Delta I | | \Delta T |} \tag {2} +$$ + +$$ +\Delta T = E _ {T} \left(t _ {\text {t a r g e t}}\right) - E _ {T} \left(t _ {\text {s o u r c e}}\right) \tag {3} +$$ + +$$ +\Delta I = E _ {I} \left(G _ {\text {t r a i n}} (z)\right) - E _ {I} \left(G _ {\text {f r o z e n}} (z)\right) \tag {4} +$$ + +where $E_{T}$ and $E_{I}$ are the text and image encoder of CLIP, $t_{source}$ and $t_{target}$ are the source and target class texts. The directional loss can prevent adversarial solutions. Since target generated images should have given direction to corresponding source images, generating a single adversarial instance is impossible. At training time, the same latent code is fed into source and target generator, then the target generator is optimized using the directional loss. However, the directional loss proposed by StyleGAN-NADA still suffers from the mode collapse problem, as shown in the images of Figure 2. So we propose the spectral consistency regularization derived from geometry analysis of GANs to resolve this problem. + +# 3.2. Geometry Analysis of GAN Adaptation + +We denote the generative network as a mapping from latent code $z$ to a manifold in image space as $G(z)$ . Considering a squared distance function $d^2$ for two images, we express the local variations of $G(z)$ from moving towards direction $\Delta z$ by second-order Taylor expansion. This is formulated as: + +$$ +\lim _ {\Delta z \to 0} d ^ {2} (G (z), G (z + \Delta z)) = +$$ + +$$ +d ^ {2} (G (z), G (z)) + \frac {\partial d ^ {2} (G (z) , G (z + \Delta z))}{\partial \Delta z} \cdot \Delta z + +$$ + +$$ +\Delta z ^ {T} \cdot \frac {\partial^ {2} d ^ {2} (G (z) , G (z + \Delta z))}{\partial \Delta z ^ {2}} \cdot \Delta z \tag {5} +$$ + +The first two terms are zero since $d^{2}(G(z), G(z + \Delta z))$ is local minima when $\Delta z = 0$ . Denote the second derivatives as Hessian matrix $H(z)$ , and we have $d^{2}(G(z), G(z + \Delta z)) = \Delta z^{T}H(z)\Delta z$ . For a normalized vector $\Delta z$ , we can conclude that $\sigma_{min} \leq d^{2}(G(z), G(z + \Delta z)) \leq \sigma_{max}$ where $\sigma_{min}$ and $\sigma_{max}$ are the smallest and largest eigenvalues of $H$ . Thus, we can use the trace norm of $H_{z}$ to reflect the statistics of diversity in generative models, which is the sum of all eigenvalues of $H$ . Especially, for a squared $L2$ distance function in metric space $\phi$ , $d^{2}(z_{1}, z_{2})_{\phi} = \frac{1}{2}\|\phi(G(z_{1})) - \phi(G(z_{2}))\|^{2}$ , the Hessian matrix $H_{\phi}(z_{0})$ is a simple transformation from the Jacobian $J_{\phi}(z_{0})$ : + +$$ +\begin{array}{l} H _ {\phi} \left(\boldsymbol {z} _ {0}\right) = \frac {\partial^ {2}}{\partial \boldsymbol {z} ^ {2}} \frac {1}{2} \| \phi \left(\boldsymbol {z} _ {0}\right) - \phi (\boldsymbol {z}) \| _ {2} ^ {2} (6) \\ = J _ {\phi} \left(\boldsymbol {z} _ {0}\right) ^ {T} J _ {\phi} \left(\boldsymbol {z} _ {0}\right) (7) \\ \end{array} +$$ + +where the top eigenvectors of $H_{\phi}$ correspond to right singular vectors of the Jacobian $J_{\phi}$ . Besides, the Hessian-vector product is related to Jacobian-vector product as follows: + +$$ +\boldsymbol {v} ^ {T} H _ {\phi} (\boldsymbol {z} _ {0}) \boldsymbol {v} = \| J _ {\phi} (\boldsymbol {z} _ {0}) \boldsymbol {v} \| ^ {2}, \quad J _ {\phi} (\boldsymbol {z}) = \frac {\partial \phi (G (\boldsymbol {z}))}{\partial z} \tag {8} +$$ + +To analyze the mode collapse problem, we calculate the statistics of Hessian trace from different samples of $z$ during adaptation process of the previous state-of-the-art method. The results of GAN adaptation from photo to sketch with StyleGAN-NADA [5] is shown in Figure 2. We find that as the iteration count increases, the mode collapse problem becomes more severe and the Hessian trace is also decreasing, which proves that the Hessian trace can reflect the diversity of generator. In the early stage of adaptation, the decrease of trace norm is mainly caused by style adaptation since there are no color variations for sketch domain. But for the late stage from 200th step to 400th step, the style effect changes little and the face attributes tend to have similar pattern. Since the target text only represents a fixed direction without variation, different samples are encouraged to approach the same fixed pattern. We also give a proof of the relationship between Hessian Trace and model diversity in the following section. + +# 3.3. Relation between Hessian Trace and Diversity + +Considering the mapping from standard normal distribution $z \sim \mathcal{N}(0, I)$ to general multivariate normal distribution $y \sim \mathcal{N}(\mu, \Sigma)$ with the generator as a linear function $G(z) = A z + b$ , which has $\mu = b, \Sigma = A A^T$ . The linear assumption of generator holds when we consider a small neighborhood around $z$ , i.e. $\lim \Delta z \to 0$ . + +The variance of $y$ can be computed as: + +$$ +\begin{array}{l} V a r (y) = E [ \| y - E [ y ] \| _ {2} ^ {2} ] \\ = E \left[ \sum_ {i = 1} ^ {n} \left(y _ {i} - E \left[ y _ {i} \right]\right) ^ {2} \right] \\ = \sum_ {i = 1} ^ {n} E \left[ \left(y _ {i} - E [ y _ {i} ]\right) ^ {2} \right] \\ = \sum_ {i = 1} ^ {n} V a r (y _ {i}) \\ = \sum_ {i = 1} ^ {n} \Sigma_ {i i} \\ = \operatorname {T r a c e} (\Sigma) = \operatorname {T r a c e} \left(A A ^ {T}\right) \tag {9} \\ \end{array} +$$ + +On the other side, the $d^2 (G(z),G(z + \Delta z))$ can be expanded as follows: + +$$ +\begin{array}{l} d ^ {2} (G (z), G (z + \Delta z)) = \| G (z) - G (z + \Delta z) \| _ {2} ^ {2} \\ = \| A z + b - (A (z + \Delta z) + b) \| _ {2} ^ {2} \\ = \| A \Delta z \| \\ = \Delta z ^ {T} A ^ {T} A \Delta z \tag {10} \\ \end{array} +$$ + +such the Hessian Matrix $H$ of $\Delta z$ with respect to $d^{2}(G(z), G(z + \Delta z))$ equals to $A^{T}A$ , e.g. $H = A^{T}A$ . Combining above two equations with $Trace(AA^{T}) =$ + +Trace $(A^TA)$ , we have $Var(y) = Trace(H)$ , which means that the Hessian Trace for every sample in target distribution reflects the variance and diversity of this distribution. If the Hessian Trace is small, the target distribution only spans a small region in space. This is consistent to the mode collapse problem in generative models. + +# 3.4. Spectral Consistency Regularization + +To prevent the target generator from mode collapse, we propose spectral consistency regularization to prevent the diversity of generator from degrading, which is calculated as: + +$$ +L _ {r e g} = \left\| \operatorname {T r a c e} \left(H _ {s} (z)\right) - \operatorname {T r a c e} \left(H _ {t} (z)\right) \right\| \tag {11} +$$ + +where $H_{s}(z)$ and $H_{t}(z)$ are the Hessian matrix of the source generator and target generator evaluated with the same latent code $z$ . However, directly computing the Hessian matrix requires backpropagation $n$ times where $n$ is the dimension of feature vector in metric space. Instead, we use the Hutchinson's method for trace estimator [9] to compute a stochastic estimator of Hessian Trace, which is formulated as: + +$$ +\operatorname {T r a c e} (H (z)) = \mathbb {E} [ v ^ {T} H (z) v ] = \mathbb {E} [ \| J _ {\phi} (z) v \| ^ {2} ] \tag {12} +$$ + +Where $v \sim \mathrm{Rademacher}(\frac{1}{2})$ , and the second transformation is derived from Equation 8. So the calculation of Hessian matrix is transformed into the calculation of Jacobian-Vector product. + +Different from the within-domain consistency loss[36] which restricts target generated samples based on relative difference of source samples, the spectral consistency regularization only cares about the diversity of target model. This doesn't impose restriction on the direction of target adaptation, so our method can generate samples more consistent with target text without losing model diversity. + +The training objective of text-driven GAN adaptation is a weighted combination of directional loss and spectral consistency regularization loss $\mathcal{L} = \mathcal{L}_{dir} + \lambda \mathcal{L}_{reg}$ . Since different target domains have their own characteristic, it is required to tune the hyperparameter $\lambda$ for better performance. To prevent the exhausting hyperparameter searching, we propose an adaptive loss reweighting method to balance the influence of these two loss items. Specifically, the adaptive weight $\lambda$ is calculated as $\lambda_{spectral} \frac{\|\nabla_{G_L} \mathcal{L}_{dir}\|}{\|\nabla_{G_L} \mathcal{L}_{reg}\|}$ , where $G_L$ denotes the last layer of generator and $\lambda_{spectral}$ is a manually specified hyperparameter, typically 1.0 is an appropriate choice. + +# 3.5. Granularity Adaptive Regularization + +For text-driven GAN adaptation problem, the adaptation granularity of different target domains varies from texture to structure. For example, the photo-to-sketch adaptation mainly focuses on the appearance and texture change, + +![](images/a68ea668e00bafe197ea5b25168a2e57373e444b35b178b656f32ba89d709e39.jpg) +Figure 4. Visual results of our method and state-of-the-art StyleGAN-NADA [5] for different GAN adaptations. The top row shows the source text and target text description. The left column presents samples generated by source generator. Our method not only generates samples described by target text with high fidelity, but also produces diverse and identity consistent images corresponding to source domain. + +while the werewolf domain has more variations on semantic structure. Regularization of diversity in the whole granularity will restrict the adaptation performance. To alleviate this problem, we propose granularity adaptive regularization based on the disentangled characteristic of StyleGAN latent code. + +The latent codes injected into different layers in StyleGAN influence different granularities. The style code in low resolution represents high-level aspects such as pose and shape, that in middle resolution controls facial features and hairstyle, and that in high resolution influences color scheme and local textures. Specifically, we use the $W+$ space as input space for Jacobian matrix calculation, where each input latent code consists of 18 512-dimensional vectors so both $z$ and $v$ in Equation 4 are in $R^{18 \times 512}$ . By masking $v$ with a mask vector $m \in \{0,1\}^{18}$ for different layers, we can specify the granularity of variations involved in the regularization. The calculation of Hessian trace under mask is formulated as: + +$$ +\begin{array}{l} \operatorname {T r a c e} (H (z)) = \mathbb {E} [ (v \odot m) ^ {T} H (z) (v \odot m) ] (13) \\ = \mathbb {E} [ \| J _ {\phi} (z) (v \odot m) \| ^ {2} ]. (14) \\ \end{array} +$$ + +Following previous convention, we divide the style code for 18 layers into 3 groups, which are for coarse, middle and fine scale. The results of different mask strategies are shown in Figure 6. + +To explore the best mask strategy for different target domains, we propose to use an adaptive soft mask vector + +$\{\tilde{m} |\tilde{m}\in R^{18},\| \tilde{m}\| = 1\}$ for all layers. During training, the mask vector is optimized with respect to the overall training objective. To reduce the directional loss, the mask vector will assign less value to the latent code corresponding to the granularity that changes most, while other values will increase to preserve the diversity of source generator. + +# 4. Experiment + +In this section, we will show the qualitative and quantitative results of our method. We illustrate the generated results for a wide range of target domains from style and texture changes to shape and semantic modifications. We also compare the proposed spectral consistency regularization with other regularization methods for diversity preservation. Next, we perform an ablation study on our method to evaluate the effectiveness of each component. Finally, we demonstrate the applications of text-guidance GAN adaptation, including image-to-image translation and image editing. To verify the generalization of proposed spectral consistency regularization, we also apply it to one-shot GAN adaptation, which is shown in supplementary appendix. + +We use the StyleGANv2 [11] generator pretrained on FFHQ as the source generator. During domain adaptation, we optimize all the parameters of generator except for the mapping network and toRGB layers. We use the Adam [12] optimizer with learning rate 0.001. The $\lambda_{\mathrm{spectral}}$ is set to 1.0. For most target domains, only 300 iterations are required to achieve convergence. Following CLIP, we use 79 + +![](images/d90133782b5c4ac0795191d31b2f9fd064804e7e8e0d46a518e9d9dbe4b8aa23.jpg) +Figure 5. Comparison results of our spectral consistency regularization with other regularization methods. Detailed explanations of these methods are in Appendix. + +manually designed prompts like "a photo of a..." with provided target domain description and feed them to text encoder of CLIP to get the embeddings of target domain. + +# 4.1. Comparison Results + +Qualitative Comparison In Figure 4, we show the comparison results of our method with state-of-the-art model StyleGAN-NADA [5] for a wide range of target domains, which varies from texture changes like sketch and Mona Lisa paintings to geometric change like werewolf and Pixar style. The results demonstrate that our method not only generates highly stylistic images consistent with target text description for different target domains, but also produces images with diversity inherited from the pretrained source generator. Compared with StyleGAN-NADA which has obvious mode collapse problem like the mouth pattern in sketch and hair in werewolf, our model generates target images with better identity consistency. This proves that the spectral consistency regularization can preserve the diversity of source domain. In the supplementary appendix, we present additional visual results for the dogs and cars domain. + +We also perform text-driven GAN adaptation experiments with other regularization methods, including the Selective Cross-modal Consistency (SCC) loss[35], Within-Domain Consistency (WDC) loss[36], the Mode + +Seeking (MS) loss[14], Perceptual Path Length (PPL) regularization[11] and Cross Domain Correspondence (CDC) regularization[17]. Detailed explanations of these regularization methods are in supplementary appendix. As shown in Figure 5, SCC, WDC and MS regularization impose too strong regularization to target generator and restrict the domain specific attributes for target domain. Suffering from mode collapse problem, the generation results of PPL and CDC exhibits the same pattern across different samples. In comparison, our method has a better balance between the diversity and stylization of target generator. + +Quantitative Comparison Besides our proposed trace norm of Hessian Matrix in Equation 3, we also leverage the Perceptual Path Length(PPL) [10] for quantitative diversity comparison. It measures the perceptually-based pairwise image distance [34] for a linear interpolation path in the latent path. The average PPL in the latent space $\mathcal{Z}$ is: + +$$ +P P L _ {\mathcal {Z}} = \mathbb {E} \left[ \right. \frac {1}{\epsilon^ {2}} d \left( \right.G \left(\operatorname {s l e r p} \left(\mathbf {z} _ {1}, \mathbf {z} _ {2}; t\right)\right), G \left(\operatorname {s l e r p} \left(\mathbf {z} _ {1}, \mathbf {z} _ {2}; t + \epsilon\right)\right)\left. \right] +$$ + +where $d(.,.)$ evaluates the perceptual distance between two images, and slerp denotes spherical interpolation since the input latent is normalized. The perceptual path length estimates the diversity of the generator via finite differences, and our method estimates the diversity analytically. + +Table 1. Comparison results for diversity by PPL and Hessian trace between different regularization methods. The PPL and Hessian trace for the pretrained source generator is 419.22 and 0.309. + +
ResultsPhoto → SketchPhoto → Mona LisaHuman → WerewolfPhoto → Pixar
PPLTracePPLTracePPLTracePPLTrace
SCC547.340.526485.700.521428.450.343440.170.535
WDC378.250.090363.170.191311.340.116345.310.167
MS466.990.167331.060.233352.610.191377.880.358
PPL241.500.028209.090.061297.740.062300.190.063
CDC348.010.062259.920.079351.730.089299.490.111
NADA323.250.061281.590.098302.910.101343.540.112
Ours463.570.116321.430.140383.190.137353.800.181
+ +Table 2. User preference study. The numbers represent the percentage of users who prefer the images synthesized by our method. + +
Model ComparisonQualityStyleAttributes
Ours vs. NADA [5]86.4%59.4%91.2%
Ours vs. SCC [35]64.2%86.2%65.0%
Ours vs. MS [14]59.4%83.6%54.2%
Ours vs. WDC [36]62.2%89.8%49.8%
Ours vs. PPL [11]92.4%71.4%95.0%
Ours vs. CDC [17]79.6%60.6%87.8%
+ +The quantitative results are shown in Table 1. The mentioned strong regularization methods including SCC, WDC and MS have large values of PPL and Hessian trace at the cost of restricting adaptation effects. Our method performs competitive with these methods in diversity metrics, but does much better in transferring target styles. Compared with CDC and PPL, our method better preserve the diversity of source domain reflected in PPL and Hessian Trace. For all different target domains, our method outperforms previous state-of-the-art model StyleGAN-NADA for both PPL and Hessian trace. This benefits from that our method can alleviate the mode collapse problem with spectral consistency regularization and generate more diverse images without restricting the style adaptation results. + +User Study We conducted a user study to compare with other regularization methods. Specifically, participants were requested to select the superior synthesized samples in relation to three measurements: (1) image quality, (2) style consistency with the target description (3) attribute consistency with the source image. Five hundred samples were randomly generated for each comparative analysis. The aggregated results are presented in Table 2. The users strongly favor our method across all three aspects. In contrast to other methods, our method demonstrates a better tradeoff between style effects and attribute preservation. The spectral consistency regularization can better preserve the diversity of source domain while not restricting the style effects for target domain. + +# 4.2. Ablation Study + +Granularity Adaptive Regularization. In Figure 6, we demonstrate generated results with different regularization strategies applied in $W+$ space in Section 3.4. Regularization to the coarse scale will preserve the structure of source image but the diversity of fine features like hair will lose. The global regularization performs similarly to coarse regularization since the coarse features dominate the diversity of generator. Only applying regularization to the fine scale will generate high-frequency textures and the structure characteristic like necks will collapse. In comparison, our proposed granularity adaptive regularization both preserves the diversity of source domain in all scales and matches the styles of target domain. + +Strength of regularization. In Figure 7, we show the generated samples with a linear interpolated loss weight $\lambda_{\text{spectral}}$ . We can observe that with increasing $\lambda_{\text{spectral}}$ , the generated samples maintain more diversity of source domain, and they also illustrate the most significant characteristic of target domain. + +Choices of metric space. We conduct experiments about the metric space of spectral consistency regularization with different feature encoder $\phi(x)$ that evaluates the distance between image samples. Besides CLIP [19] image encoder in our method, we also leverage VGG [27] which was commonly used in style transfer [6], MoCo [8] of contrastive representation learning, ArcFace [2] for face recognition and the plain pixel space. As shown in Figure 8, compared to other feature encoders, the spectral consistency regularization with CLIP encoder shows best performance for preserving the identity of source image. + +# 4.3. Applications + +Image-to-Image Translation We combine the adapted generator in target domain with a GAN inversion encoder to implement image-to-image translation. Given a real-world image, we invert it to the latent code in $W$ space via an e4e encoder [28], which is then fed to target generator to + +![](images/5fc7ca3b9b5b5c3f958f58a3b89e95fc1e22683196a7469b763cb5e2c2c18612.jpg) +Source +Figure 6. Results of different regularization strategies for GAN adaptation from photo to Edvard Munch paintings. + +![](images/619c9ffed9b5792082c689dc9dc300f741848aed72236b0181a60a37dfb0e119.jpg) +Global + +![](images/9024a80d7daccfd1347ef37f24a1da860714a87d93857f2ea591094ab964d0f2.jpg) +Coarse + +![](images/a4ea90735778d2df4a253071106db1e4d963cc09716d807c5ff78b4f12e7876b.jpg) +Middle + +![](images/dc931f537612bbd32897324360cc40b8bcdf707a9284110a879c9977b75ef8ae.jpg) +Fine + +![](images/dd71d7613cfe1393d543968c28d55e2f7bd77ceb464c37c972e422415e27db0d.jpg) +Adaptive + +![](images/35970361d29c69aed78b67ee96e0df1924c04fba09bbf3750eab2690da5b6332.jpg) +Figure 7. Generation results of linearly interpolated $\lambda_{\mathrm{spectral}}$ in regularization. + +![](images/a73e311b2e92cf70a3332c7cfe3833894b6ffd67784ccc44416b98c288562e96.jpg) +Figure 8. The generated results with spectral consistency regularization under the metric spaces defined by different encoders. + +produce target image. As shown in Figure 9, our method can achieve high-quality image translation and preserve the identity of source image for different domains. + +Image Editing In Figure 10, we demonstrate the image editing results performed on the target domain. We leverage the meaningful directions found by InterfaceGAN [25] to edit target images. We can observe that the editing directions from source domain still apply to target domains, which proves that the target domain preserves the semantic distribution of source domain. + +![](images/0e54b7b20cfa93fc789ffa81ce845f9de3b9158774154d649c0ccb41659a7370.jpg) +Figure 9. Image translation of real-world images to different target domains. Each column shows a target domain, and the top row is the text description for target domain. The transferred images represent both target style and the identity of source image. + +![](images/bf5abcc8f4d6fb9fb4919fb89e3edb59808e878cffd554a466fa09aa033e6f45.jpg) +Figure 10. Editing images in target domain for real-world images. The top row shows the edited attributes. + +# 5. Conclusion + +In this paper, we propose the spectral consistency regularization for text-driven generative domain adaptation. The key insight of our method is to build a quantitative diversity estimator to preserve the intra-domain diversity of source generator without restricting the adaptation of target style. We also introduce an adaptive regularization strategy for granularity-flexible adaptation. The experiments demonstrate our method greatly improves the generation results for a wide range of target domains. + +# Acknowledgement + +This work was supported in part by the National Key R&D Program of China under Grant 2018AAA0102000, in part by National Natural Science Foundation of China: 62322211, 62236008, U21B2038, 62225207, U19B2038 and 61931008, and Youth Innovation Promotion Association of Chinese Academy of Sciences under Grant 2020108. + +# References + +[1] Katherine Crowson, Stella Rose Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. ArXiv, abs/2204.08583, 2022. 3 +[2] Jiankang Deng, J. Guo, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4685-4694, 2019. 8 +[3] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868-12878, 2021. 2 + +[4] Kevin Frans, Lisa B. Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. ArXiv, abs/2106.14843, 2021. 3 +[5] Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ArXiv, abs/2108.00946, 2021. 1, 3, 5, 6, 7, 8 +[6] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 8 +[7] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. ArXiv, abs/2004.02546, 2020. 3 +[8] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 8 +[9] Michael F. Hutchinson. A stochastic estimator of the trace of the influence matrix for laplacian smoothing splines. Communications in Statistics - Simulation and Computation, 18:1059-1076, 1989. 5 +[10] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4396-4405, 2019. 7 +[11] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8107-8116, 2020. 6, 7, 8 +[12] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. 6 +[13] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. ArXiv, abs/2111.03186, 2021. 3 +[14] Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Siwei Ma, and Ming-Hsuan Yang. Mode seeking generative adversarial networks for diverse image synthesis. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1429-1437, 2019. 7, 8 +[15] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze discriminator: A simple baseline for fine-tuning gans. ArXiv, abs/2002.10964, 2020. 1, 3 +[16] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2750-2758, 2019. 3 +[17] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A. Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10738-10747, 2021. 3, 7, 8 +[18] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and D. Lischinski. Styleclip: Text-driven manipulation of + +stylegan imagery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2065-2074, 2021. 3 +[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 8 +[20] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1, 2 +[21] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 2 +[22] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 2 +[23] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1, 2 +[24] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9240-9249, 2020. 3 +[25] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:2004-2018, 2022. 9 +[26] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1532-1540, 2021. 3 +[27] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2015. 8 +[28] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40:1 - 14, 2021. 8 +[29] Binxu Wang and Carlos R. Ponce. The geometry of deep generative image models and its applications. ArXiv, abs/2101.06006, 2021. 3 +[30] Jiayu Xiao, Liang Li, Chaofei Wang, Zhengjun Zha, and Qingming Huang. Few shot generative model adaption via relaxed spatial structural alignment. ArXiv, abs/2203.04121, 2022. 1, 3 +[31] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316-1324, 2018. 2 + +[32] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Benton C. Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. ArXiv, abs/2206.10789, 2022. 2 +[33] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N. Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5908-5916, 2017. 2 +[34] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 7 +[35] Yabo Zhang, Mingshuai Yao, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, and Wangmeng Zuo. Towards diverse and faithful one-shot adaption of generative adversarial networks. ArXiv, abs/2207.08736, 2022. 7, 8 +[36] Peihao Zhu, Rameen Abdal, John C. Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. ArXiv, abs/2110.08398, 2021. 1, 2, 3, 5, 7, 8 \ No newline at end of file diff --git a/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/images.zip b/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c7fa12c9e6a4b27207fc0f068a2681a4ec5dae9d --- /dev/null +++ b/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:428432b00ebc4ac33ee34e5be0f0718f23a052b74d5d83e653d43cc344c620c8 +size 993603 diff --git a/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/layout.json b/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dedab14d0dd3ecdfb58892c6772b9e2dcc841af6 --- /dev/null +++ b/2023/Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization/layout.json @@ -0,0 +1,8386 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 79, + 102, + 515, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 102, + 515, + 139 + ], + "spans": [ + { + "bbox": [ + 79, + 102, + 515, + 139 + ], + "type": "text", + "content": "Text-Driven Generative Domain Adaptation with Spectral Consistency Regularization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "spans": [ + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "text", + "content": "Zhenhuan Liu" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "text", + "content": " Liang Li" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "text", + "content": " Jiayu Xiao" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "text", + "content": " Zheng-Jun Zha" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "text", + "content": " Qingming Huang" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "text", + "content": "Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, CAS \n" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences \n" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "text", + "content": "University of Science and Technology of China \n" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 60, + 159, + 532, + 232 + ], + "type": "text", + "content": "Peng Cheng Laboratory" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 233, + 462, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 233, + 462, + 258 + ], + "spans": [ + { + "bbox": [ + 129, + 233, + 462, + 258 + ], + "type": "text", + "content": "{zhenhuan.liu, jiayu.xiao}@vip1.ict.ac.cn, liang.li@ict.ac.cn, zhazj@ustc.edu.cn, qmhuang@ucas.ac.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "spans": [ + { + "bbox": [ + 143, + 286, + 192, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 312, + 290, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 290, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 290, + 540 + ], + "type": "text", + "content": "Combined with the generative prior of pre-trained models and the flexibility of text, text-driven generative domain adaptation can generate images from a wide range of target domains. However, current methods still suffer from overfitting and the mode collapse problem. In this paper, we analyze the mode collapse from the geometric point of view and reveal its relationship to the Hessian matrix of generator. To alleviate it, we propose the spectral consistency regularization to preserve the diversity of source domain without restricting the semantic adaptation to target domain. We also design granularity adaptive regularization to flexibly control the balance between diversity and stylization for target model. We conduct experiments for broad target domains compared with state-of-the-art methods and extensive ablation studies. The experiments demonstrate the effectiveness of our method to preserve the diversity of source domain and generate high fidelity target images. Source code has been released in https://github.com/Victarry/Adaptation-SCR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 128, + 577 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 585, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 287, + 693 + ], + "type": "text", + "content": "Generative image modeling has developed significantly in recent years and is able to generate diverse high-resolution images even indistinguishable from real images. However, training such models requires intense computation resources and large datasets, which restricts the application scope of generative models. For some scenarios, collecting large datasets is impossible like paintings by specific artists. Benefiting from Vision-Language models learning from large image-text pairs, text can be leveraged as a de" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 287, + 547, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 547, + 384 + ], + "type": "text", + "content": "scription of abstract visual semantics to guide generative domain adaptation instead of a collection of image samples in target domain. As an expressive representation, text has shown great success in semantic image generation and manipulation recently [23, 20]. Based on the generative prior of pre-trained models and flexible text description of target domain, text-driven generative domain adaptation can generate more various images and have promising applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 387, + 548, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 548, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 548, + 615 + ], + "type": "text", + "content": "To reduce the requirement of training samples, traditional methods propose to train generative models in the target domain with only limited samples by adapting pretrained models in the large-scale source domain which contains high-level semantic knowledge as a generative prior. These few-shot adaptation methods either finetune only a part of parameters within networks to preserve most source domain knowledge [15] or impose strong regularization on the generated images [30, 36]. However, these methods still require additional training samples of target domain and adversarial training process. As the number of samples drops, the image fidelity and diversity also hurt severely. Different from these methods, text-driven generative domain adaptation requires no image samples but texts to describe the target domain. Pioneer work [5] proposed to encourage the visual change between samples from target and source generators to align with semantic direction described by text in the CLIP [19] embedding space, which achieves generative adaptation for miscellaneous domains in short training time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "content": "The main challenge of text-driven GAN adaptation is the mode collapse problem due to the entanglement of intradomain semantics and inter-domain style in text representation. Besides the specified target style described by text, there also exists an unknown pattern for the semantics of images. This leads to a decrease of variations in generated images when the style effect is optimized to approach target domain. As shown in Figure 2, while the number of qua" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "text", + "content": "7019" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 79, + 152, + 175 + ], + "blocks": [ + { + "bbox": [ + 56, + 79, + 152, + 175 + ], + "lines": [ + { + "bbox": [ + 56, + 79, + 152, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 79, + 152, + 175 + ], + "type": "image", + "image_path": "515369480a1c248fea0c85684a4156da85c52a1ed05fa55b0c61b4b6e3569dda.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 176, + 124, + 188 + ], + "lines": [ + { + "bbox": [ + 83, + 176, + 124, + 188 + ], + "spans": [ + { + "bbox": [ + 83, + 176, + 124, + 188 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 152, + 79, + 246, + 175 + ], + "blocks": [ + { + "bbox": [ + 152, + 79, + 246, + 175 + ], + "lines": [ + { + "bbox": [ + 152, + 79, + 246, + 175 + ], + "spans": [ + { + "bbox": [ + 152, + 79, + 246, + 175 + ], + "type": "image", + "image_path": "ea0686e26676b153e03298514bfa2bbce6bfdf29f753b30c5228666289dc52e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 176, + 217, + 187 + ], + "lines": [ + { + "bbox": [ + 182, + 176, + 217, + 187 + ], + "spans": [ + { + "bbox": [ + 182, + 176, + 217, + 187 + ], + "type": "text", + "content": "Sketch" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 247, + 79, + 341, + 175 + ], + "blocks": [ + { + "bbox": [ + 247, + 79, + 341, + 175 + ], + "lines": [ + { + "bbox": [ + 247, + 79, + 341, + 175 + ], + "spans": [ + { + "bbox": [ + 247, + 79, + 341, + 175 + ], + "type": "image", + "image_path": "e940d699410540f4fa44b31774cac8f327547f6703362a3200867b143a59b04a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 176, + 315, + 188 + ], + "lines": [ + { + "bbox": [ + 274, + 176, + 315, + 188 + ], + "spans": [ + { + "bbox": [ + 274, + 176, + 315, + 188 + ], + "type": "text", + "content": "Ukiyo-e" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 197, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 46, + 197, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 545, + 218 + ], + "type": "text", + "content": "Figure 1. Text-driven generative domain adaptation with various text descriptions. The generated samples should both reflect characteristics of target domain from text and preserve the original identity." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 342, + 79, + 436, + 175 + ], + "blocks": [ + { + "bbox": [ + 342, + 79, + 436, + 175 + ], + "lines": [ + { + "bbox": [ + 342, + 79, + 436, + 175 + ], + "spans": [ + { + "bbox": [ + 342, + 79, + 436, + 175 + ], + "type": "image", + "image_path": "2a66b4c5f146b89da43351e1f4f483152dfd6a0b41c56c9a2f5892acea894292.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 176, + 418, + 188 + ], + "lines": [ + { + "bbox": [ + 371, + 176, + 418, + 188 + ], + "spans": [ + { + "bbox": [ + 371, + 176, + 418, + 188 + ], + "type": "text", + "content": "Van Gogh" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 436, + 79, + 531, + 175 + ], + "blocks": [ + { + "bbox": [ + 436, + 79, + 531, + 175 + ], + "lines": [ + { + "bbox": [ + 436, + 79, + 531, + 175 + ], + "spans": [ + { + "bbox": [ + 436, + 79, + 531, + 175 + ], + "type": "image", + "image_path": "97d84f6d91337b724cb3295cf008a74389dd98076f1b6b2630abc45981073a59.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 176, + 510, + 187 + ], + "lines": [ + { + "bbox": [ + 462, + 176, + 510, + 187 + ], + "spans": [ + { + "bbox": [ + 462, + 176, + 510, + 187 + ], + "type": "text", + "content": "Werewolf" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 239, + 286, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 286, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 286, + 310 + ], + "type": "text", + "content": "tions increases, the generated samples tend to have similar patterns of mouth and eyes, which reduces most of the variations in the origin model. The main reason for the mode collapse problem is that the optimization process only cares about the distance of generated samples to target domains, and the intra-domain feature variations are easily ignored." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 315, + 286, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 286, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 286, + 494 + ], + "type": "text", + "content": "To address the above challenge, researchers [36] proposed to preserve the diversity of source domain through a within-domain consistency loss which keeps consistency between sample changes in source domain and target domain. However, this regularization is too strong to restrict the style effect of target generator close to source domain. The previous theorem about GAN latents analysis has shown that the Hessian matrix of generator reflects the variations of generator and can be used to explore meaningful directions from top eigenvectors. Inspired by this, we try to leverage the spectrum of Hessian matrix as a quantitative evaluation of model diversity in the adaptation problem. This disentangles the relative diversity between generated samples from absolute generative distribution and makes a general way to regularize diversity of generative model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": "In this work, we propose spectral consistency regularization to solve the problem of mode collapse in text-driven generative domain adaptation from the geometric point of view. First, we analyze the Hessian matrix of generator's manifold in the metric space by eigendecomposition. The eigenvalues of Hessian matrix are decreasing in the adaptation process, which is consistent with the mode collapse problem of visual observations. Second, we introduce the spectral consistency regularization on the Hessian matrix to prevent the latent space of generator from degrading. This regularization helps preserve intra-domain variations of source domain without restricting style effects of target generators. We further develop a stochastic method to regularize the spectrum of Hessian matrix without calculating the full matrix, which reduces the expensive computational cost. Finally, we design the granularity adaptive regularization considering the layer-decomposition characteristic of " + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{W + }" + }, + { + "bbox": [ + 46, + 498, + 286, + 713 + ], + "type": "text", + "content": " space in StyleGAN." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 239, + 538, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 239, + 538, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 538, + 251 + ], + "type": "text", + "content": "The contributions of this paper are summarized below:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 258, + 545, + 455 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 258, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 545, + 307 + ], + "type": "text", + "content": "1. We analyze the commonly occurred mode collapse problem in GAN adaptation from the geometric point of view and provide a quantitative evaluation of model diversity to reveal the reason of mode collapse." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 315, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 545, + 386 + ], + "type": "text", + "content": "2. We propose a spectral consistency regularization for text-driven GAN adaptation, which preserves diversity of original domain and generates high fidelity images of target domain. A granularity adaptive regularization is further designed to flexibly control balance between diversity and stylization for target model." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 395, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 545, + 455 + ], + "type": "text", + "content": "3. We conduct experiments and ablation studies for a wide range of target domains. The experiments show the effectiveness of our proposed spectral consistency regularization and its applications to downstream tasks like image editing and image-to-image translation." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 466, + 391, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 391, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 391, + 478 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 486, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 629 + ], + "type": "text", + "content": "Text-driven Image Synthesis and Manipulation. Traditional methods approached text-driven image generation by training a conditional GAN[22]. Several following works have been proposed to improve generation quality either by multi-scale networks [33] or attention mechanism [31]. Recently, transformer-based auto-regressive generative models were introduced to view text-driven image synthesis as conditional sequence generation of visual tokens conditioning on text embeddings [3, 21, 32]. Diffusion models were also leveraged as the decoder for image generation, which achieves tremendous improvement for generating high quality images [23, 20]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Another kind of method is to leverage Contrastive Language-Image Pre-training (CLIP) [19] models as knowledge guidance for text based image generation. This is achieved by optimizing the latent code of pre-trained generator to close the distance between generated images and input text in the shared embedding space. The optimized latent codes of generator are either in the StyleGAN latent" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7020" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 74, + 527, + 210 + ], + "blocks": [ + { + "bbox": [ + 51, + 74, + 527, + 210 + ], + "lines": [ + { + "bbox": [ + 51, + 74, + 527, + 210 + ], + "spans": [ + { + "bbox": [ + 51, + 74, + 527, + 210 + ], + "type": "image", + "image_path": "58d312ec650c0316d4273c2e4caefcb18cf80c4a38ff0607f270bb8f414cf9c5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 215, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 215, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 547, + 239 + ], + "type": "text", + "content": "Figure 2. The mode collapse problem in the photo-to-sketch GAN adaptation by StyleGAN-NADA [5]. The trace norm of Hessian matrix is gradually decreasing, which is consistent with the visual examples showing similar patterns with mouth and eyes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 257, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 293 + ], + "type": "text", + "content": "space [18] or in VQGAN codebook space [1]. Some methods also defined the parametrized space by vector graphics such as Bezier curves [4]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 294, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 497 + ], + "type": "text", + "content": "Few-shot GAN Adaptation. It aims to transfer pretrained generator to another domain when there are not enough samples to train from scratch. Its main challenge is the mode collapse problem because the generator is prone to overfit training samples in target domain and lose the diversity of the origin domain. There have been many methods to tackle this problem. They either froze most of the parameters of the pre-trained network[15] or embedded a small number of trainable parameters into the source model[16]. Recently, [17] proposed cross-domain distance consistency loss to preserve the relative similarities and differences between instances in the source domain. [30] introduced spatial structural consistency loss to align the spatial information between the synthesis image pairs of the source and target domains. These methods still require manually collected samples, and as the number of samples decreases, the mode collapse problem becomes more apparent." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": "Besides, StyleGAN-NADA [5] further proposed to take advantage of the CLIP [19] model as knowledge guidance for GAN adaptation, and only natural language prompts are required without even a single image. Similarly, [36] used the image encoder of CLIP for one-shot adaptation. However, these methods still suffer from the mode collapse problem. In this paper, we propose the spectral consistency regularization to tackle the problem of mode collapse without hurting target generation performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Latent Space Analysis of GANs. Many works have explored the latent space of pretrained generator for image manipulation. Some methods used supervised datasets to learn directions in the latent space for attribute editing [24] or semantic image editing [13]. Other works instead applied unsupervised methods to reveal the latent space. [26] decomposed the learned weights of the pre-trained network to identify semantically meaningful directions. [7] applied principal component analysis in the latent space. Recently," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 257, + 547, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 547, + 342 + ], + "type": "text", + "content": "[29] proposed to analyze the latent space of generative models from geometric point of view. They found that the eigenvectors corresponding to the largest eigenvalues of the Hessian matrix for generator dominate interpretable variations. In this paper, we analyze the GAN adaptation problem in a similar way and propose to regularize target generator by the spectrum of Hessian matrix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 365, + 362, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 362, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 362, + 377 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 389, + 534, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 389, + 534, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 389, + 534, + 403 + ], + "type": "text", + "content": "3.1. Text-Driven Generative Domain Adaptation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 412, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 547, + 533 + ], + "type": "text", + "content": "Text-driven GAN adaptation aims to transfer a pretrained generator to target domain specified by the text description. To guide the GAN adaptation by text, pre-trained CLIP model is leveraged to measure the similarity between image and text. CLIP is a Vision-Language model trained on 400 million (image, text) pairs collected from the internet with contrastive loss [19]. One commonly used objective function for text-driven image manipulation is the global loss that optimizes the similarity between generated images and target text:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 358, + 551, + 545, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 551, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 358, + 551, + 545, + 564 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {g l o b a l}} = D _ {\\text {C L I P}} (G (z), t _ {\\text {t a r g e t}}) \\tag {1}", + "image_path": "cdae90194a1d88ecb68d70ff2b49b34800b35f5d57b9f27694b76b9580c18952.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "D_{CLIP}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " is the cosine distance in the CLIP space, " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "t_{target}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " is the target text. However, this only applies to in-domain image manipulation combined with identity consistency regularization[18]. This regularization is too strong for cross-domain GAN adaptation with large domain gaps like human to werewolf. Direct optimization of the above global loss leads to adversarial solutions since adding pixel-level perturbations can fool the CLIP classifier in the absence of a generative prior favoring real-image manifold [5]. To overcome this limit, the directional loss is used to optimize the direction between source and target" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "7021" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 73, + 542, + 257 + ], + "blocks": [ + { + "bbox": [ + 55, + 73, + 542, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 73, + 542, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 542, + 257 + ], + "type": "image", + "image_path": "a7ed1545cd85801af19d50d9c595ebec69273f370eef39f6dcb24c78312d2d3f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 263, + 548, + 320 + ], + "lines": [ + { + "bbox": [ + 46, + 263, + 548, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 548, + 320 + ], + "type": "text", + "content": "Figure 3. Illustration of our proposed method for text-driven generative domain adaptation. The training objective of our model consists of directional loss and spectral consistency loss. We feed the same latent code to source generator and target generator and generate a pair of source and target images. The directional loss encourages the direction between embeddings of the pair to align with the semantic direction of text description. The spectral consistency loss regularizes the trace norm of Hessian matrix of target generator to prevent the mode collapse problem." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 338, + 84, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 84, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 84, + 349 + ], + "type": "text", + "content": "domain:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 365, + 287, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 365, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 86, + 365, + 287, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i r e c t i o n}} = 1 - \\frac {\\Delta I \\cdot \\Delta T}{| \\Delta I | | \\Delta T |} \\tag {2}", + "image_path": "b895610d9e0ce3d3700d5597ff28322d3524981733e2163cb723a8d09b272979.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 393, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 393, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 87, + 393, + 287, + 407 + ], + "type": "interline_equation", + "content": "\\Delta T = E _ {T} \\left(t _ {\\text {t a r g e t}}\\right) - E _ {T} \\left(t _ {\\text {s o u r c e}}\\right) \\tag {3}", + "image_path": "fac92042466e991d0785a6f610ec2992c3d907ae3df578f095abaf72346832de.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 87, + 409, + 287, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 409, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 87, + 409, + 287, + 422 + ], + "type": "interline_equation", + "content": "\\Delta I = E _ {I} \\left(G _ {\\text {t r a i n}} (z)\\right) - E _ {I} \\left(G _ {\\text {f r o z e n}} (z)\\right) \\tag {4}", + "image_path": "726de5f68a26322c40ce9e48d010310d9ef18fcea6554b9caf0ce5becea3275e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 441, + 289, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 289, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 289, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 441, + 289, + 597 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 46, + 441, + 289, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 441, + 289, + 597 + ], + "type": "inline_equation", + "content": "E_{I}" + }, + { + "bbox": [ + 46, + 441, + 289, + 597 + ], + "type": "text", + "content": " are the text and image encoder of CLIP, " + }, + { + "bbox": [ + 46, + 441, + 289, + 597 + ], + "type": "inline_equation", + "content": "t_{source}" + }, + { + "bbox": [ + 46, + 441, + 289, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 441, + 289, + 597 + ], + "type": "inline_equation", + "content": "t_{target}" + }, + { + "bbox": [ + 46, + 441, + 289, + 597 + ], + "type": "text", + "content": " are the source and target class texts. The directional loss can prevent adversarial solutions. Since target generated images should have given direction to corresponding source images, generating a single adversarial instance is impossible. At training time, the same latent code is fed into source and target generator, then the target generator is optimized using the directional loss. However, the directional loss proposed by StyleGAN-NADA still suffers from the mode collapse problem, as shown in the images of Figure 2. So we propose the spectral consistency regularization derived from geometry analysis of GANs to resolve this problem." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 618, + 254, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 254, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 254, + 632 + ], + "type": "text", + "content": "3.2. Geometry Analysis of GAN Adaptation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": "We denote the generative network as a mapping from latent code " + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": " to a manifold in image space as " + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "inline_equation", + "content": "G(z)" + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": ". Considering a squared distance function " + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "inline_equation", + "content": "d^2" + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": " for two images, we express the local variations of " + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "inline_equation", + "content": "G(z)" + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": " from moving towards direction " + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\Delta z" + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": " by second-order Taylor expansion. This is formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 357, + 453, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 357, + 453, + 376 + ], + "spans": [ + { + "bbox": [ + 326, + 357, + 453, + 376 + ], + "type": "interline_equation", + "content": "\\lim _ {\\Delta z \\to 0} d ^ {2} (G (z), G (z + \\Delta z)) =", + "image_path": "09b79620887fc0fd4b21ad99a1bffb2d4020c63ae007fecdc3f8704aa3c48cdc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 377, + 528, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 377, + 528, + 402 + ], + "spans": [ + { + "bbox": [ + 326, + 377, + 528, + 402 + ], + "type": "interline_equation", + "content": "d ^ {2} (G (z), G (z)) + \\frac {\\partial d ^ {2} (G (z) , G (z + \\Delta z))}{\\partial \\Delta z} \\cdot \\Delta z +", + "image_path": "af332f4c2819bd05fea62c7a60f5779dc8e6fad3e185b9c00fc66b5a8c196c9a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 403, + 545, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 403, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 324, + 403, + 545, + 429 + ], + "type": "interline_equation", + "content": "\\Delta z ^ {T} \\cdot \\frac {\\partial^ {2} d ^ {2} (G (z) , G (z + \\Delta z))}{\\partial \\Delta z ^ {2}} \\cdot \\Delta z \\tag {5}", + "image_path": "787d0ec1a830c63cfcb7a98bbee739bf561f20d9cbe2f6f7898a8f0f998dd278.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": "The first two terms are zero since " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "d^{2}(G(z), G(z + \\Delta z))" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": " is local minima when " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "\\Delta z = 0" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": ". Denote the second derivatives as Hessian matrix " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "H(z)" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": ", and we have " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "d^{2}(G(z), G(z + \\Delta z)) = \\Delta z^{T}H(z)\\Delta z" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": ". For a normalized vector " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "\\Delta z" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": ", we can conclude that " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "\\sigma_{min} \\leq d^{2}(G(z), G(z + \\Delta z)) \\leq \\sigma_{max}" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "\\sigma_{min}" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "\\sigma_{max}" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": " are the smallest and largest eigenvalues of " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": ". Thus, we can use the trace norm of " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "H_{z}" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": " to reflect the statistics of diversity in generative models, which is the sum of all eigenvalues of " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": ". Especially, for a squared " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": " distance function in metric space " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "d^{2}(z_{1}, z_{2})_{\\phi} = \\frac{1}{2}\\|\\phi(G(z_{1})) - \\phi(G(z_{2}))\\|^{2}" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": ", the Hessian matrix " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "H_{\\phi}(z_{0})" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": " is a simple transformation from the Jacobian " + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "inline_equation", + "content": "J_{\\phi}(z_{0})" + }, + { + "bbox": [ + 305, + 434, + 546, + 579 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 586, + 545, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 586, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 351, + 586, + 545, + 628 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} H _ {\\phi} \\left(\\boldsymbol {z} _ {0}\\right) = \\frac {\\partial^ {2}}{\\partial \\boldsymbol {z} ^ {2}} \\frac {1}{2} \\| \\phi \\left(\\boldsymbol {z} _ {0}\\right) - \\phi (\\boldsymbol {z}) \\| _ {2} ^ {2} (6) \\\\ = J _ {\\phi} \\left(\\boldsymbol {z} _ {0}\\right) ^ {T} J _ {\\phi} \\left(\\boldsymbol {z} _ {0}\\right) (7) \\\\ \\end{array}", + "image_path": "fe524223e947cdc69ccd037e9710bbd763c15f03363f97ea8936a49a461518fa.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 635, + 546, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 546, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 546, + 672 + ], + "type": "text", + "content": "where the top eigenvectors of " + }, + { + "bbox": [ + 304, + 635, + 546, + 672 + ], + "type": "inline_equation", + "content": "H_{\\phi}" + }, + { + "bbox": [ + 304, + 635, + 546, + 672 + ], + "type": "text", + "content": " correspond to right singular vectors of the Jacobian " + }, + { + "bbox": [ + 304, + 635, + 546, + 672 + ], + "type": "inline_equation", + "content": "J_{\\phi}" + }, + { + "bbox": [ + 304, + 635, + 546, + 672 + ], + "type": "text", + "content": ". Besides, the Hessian-vector product is related to Jacobian-vector product as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 687, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 687, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 687, + 545, + 713 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} ^ {T} H _ {\\phi} (\\boldsymbol {z} _ {0}) \\boldsymbol {v} = \\| J _ {\\phi} (\\boldsymbol {z} _ {0}) \\boldsymbol {v} \\| ^ {2}, \\quad J _ {\\phi} (\\boldsymbol {z}) = \\frac {\\partial \\phi (G (\\boldsymbol {z}))}{\\partial z} \\tag {8}", + "image_path": "40aeebe5ec2a2468a5bbd09a698acb3116ac1c2c16de20b76c7eb89d548a751e.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7022" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "To analyze the mode collapse problem, we calculate the statistics of Hessian trace from different samples of " + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": " during adaptation process of the previous state-of-the-art method. The results of GAN adaptation from photo to sketch with StyleGAN-NADA [5] is shown in Figure 2. We find that as the iteration count increases, the mode collapse problem becomes more severe and the Hessian trace is also decreasing, which proves that the Hessian trace can reflect the diversity of generator. In the early stage of adaptation, the decrease of trace norm is mainly caused by style adaptation since there are no color variations for sketch domain. But for the late stage from 200th step to 400th step, the style effect changes little and the face attributes tend to have similar pattern. Since the target text only represents a fixed direction without variation, different samples are encouraged to approach the same fixed pattern. We also give a proof of the relationship between Hessian Trace and model diversity in the following section." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 295, + 285, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 295, + 285, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 285, + 308 + ], + "type": "text", + "content": "3.3. Relation between Hessian Trace and Diversity" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": "Considering the mapping from standard normal distribution " + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "inline_equation", + "content": "z \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": " to general multivariate normal distribution " + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "inline_equation", + "content": "y \\sim \\mathcal{N}(\\mu, \\Sigma)" + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": " with the generator as a linear function " + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "inline_equation", + "content": "G(z) = A z + b" + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": ", which has " + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\mu = b, \\Sigma = A A^T" + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": ". The linear assumption of generator holds when we consider a small neighborhood around " + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\lim \\Delta z \\to 0" + }, + { + "bbox": [ + 47, + 313, + 287, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 385, + 216, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 385, + 216, + 396 + ], + "spans": [ + { + "bbox": [ + 59, + 385, + 216, + 396 + ], + "type": "text", + "content": "The variance of " + }, + { + "bbox": [ + 59, + 385, + 216, + 396 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 59, + 385, + 216, + 396 + ], + "type": "text", + "content": " can be computed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 406, + 286, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 406, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 91, + 406, + 286, + 567 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} V a r (y) = E [ \\| y - E [ y ] \\| _ {2} ^ {2} ] \\\\ = E \\left[ \\sum_ {i = 1} ^ {n} \\left(y _ {i} - E \\left[ y _ {i} \\right]\\right) ^ {2} \\right] \\\\ = \\sum_ {i = 1} ^ {n} E \\left[ \\left(y _ {i} - E [ y _ {i} ]\\right) ^ {2} \\right] \\\\ = \\sum_ {i = 1} ^ {n} V a r (y _ {i}) \\\\ = \\sum_ {i = 1} ^ {n} \\Sigma_ {i i} \\\\ = \\operatorname {T r a c e} (\\Sigma) = \\operatorname {T r a c e} \\left(A A ^ {T}\\right) \\tag {9} \\\\ \\end{array}", + "image_path": "b3498d04564a6b5d252decce9a62db32bd3e12e48b93d028fc6cbd5b16951853.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 576, + 287, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 576, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 576, + 287, + 599 + ], + "type": "text", + "content": "On the other side, the " + }, + { + "bbox": [ + 47, + 576, + 287, + 599 + ], + "type": "inline_equation", + "content": "d^2 (G(z),G(z + \\Delta z))" + }, + { + "bbox": [ + 47, + 576, + 287, + 599 + ], + "type": "text", + "content": " can be expanded as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 609, + 286, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 609, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 609, + 286, + 668 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} d ^ {2} (G (z), G (z + \\Delta z)) = \\| G (z) - G (z + \\Delta z) \\| _ {2} ^ {2} \\\\ = \\| A z + b - (A (z + \\Delta z) + b) \\| _ {2} ^ {2} \\\\ = \\| A \\Delta z \\| \\\\ = \\Delta z ^ {T} A ^ {T} A \\Delta z \\tag {10} \\\\ \\end{array}", + "image_path": "5076b52bdb7c64876b33786979c83699b0655544246b5aacdb9356c99ddae543.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "such the Hessian Matrix " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\Delta z" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "d^{2}(G(z), G(z + \\Delta z))" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " equals to " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "A^{T}A" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ", e.g. " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "H = A^{T}A" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ". Combining above two equations with " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "Trace(AA^{T}) =" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "Trace " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "(A^TA)" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "Var(y) = Trace(H)" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ", which means that the Hessian Trace for every sample in target distribution reflects the variance and diversity of this distribution. If the Hessian Trace is small, the target distribution only spans a small region in space. This is consistent to the mode collapse problem in generative models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 150, + 499, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 150, + 499, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 150, + 499, + 163 + ], + "type": "text", + "content": "3.4. Spectral Consistency Regularization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 168, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 215 + ], + "type": "text", + "content": "To prevent the target generator from mode collapse, we propose spectral consistency regularization to prevent the diversity of generator from degrading, which is calculated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 223, + 545, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 223, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 331, + 223, + 545, + 236 + ], + "type": "interline_equation", + "content": "L _ {r e g} = \\left\\| \\operatorname {T r a c e} \\left(H _ {s} (z)\\right) - \\operatorname {T r a c e} \\left(H _ {t} (z)\\right) \\right\\| \\tag {11}", + "image_path": "f03c7b630fb51803d588e8936cf8d22763d14fe7fb59b6b3aa8a4f37f761df20.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 240, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 240, + 545, + 335 + ], + "type": "inline_equation", + "content": "H_{s}(z)" + }, + { + "bbox": [ + 304, + 240, + 545, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 240, + 545, + 335 + ], + "type": "inline_equation", + "content": "H_{t}(z)" + }, + { + "bbox": [ + 304, + 240, + 545, + 335 + ], + "type": "text", + "content": " are the Hessian matrix of the source generator and target generator evaluated with the same latent code " + }, + { + "bbox": [ + 304, + 240, + 545, + 335 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 240, + 545, + 335 + ], + "type": "text", + "content": ". However, directly computing the Hessian matrix requires backpropagation " + }, + { + "bbox": [ + 304, + 240, + 545, + 335 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 240, + 545, + 335 + ], + "type": "text", + "content": " times where " + }, + { + "bbox": [ + 304, + 240, + 545, + 335 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 240, + 545, + 335 + ], + "type": "text", + "content": " is the dimension of feature vector in metric space. Instead, we use the Hutchinson's method for trace estimator [9] to compute a stochastic estimator of Hessian Trace, which is formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 342, + 545, + 355 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 342, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 323, + 342, + 545, + 355 + ], + "type": "interline_equation", + "content": "\\operatorname {T r a c e} (H (z)) = \\mathbb {E} [ v ^ {T} H (z) v ] = \\mathbb {E} [ \\| J _ {\\phi} (z) v \\| ^ {2} ] \\tag {12}", + "image_path": "7e1cf7a07a8f2d5deb25dc2ec5dfb6eca2c19ea786b507ea7fc5098288696f41.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 361, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 409 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 304, + 361, + 545, + 409 + ], + "type": "inline_equation", + "content": "v \\sim \\mathrm{Rademacher}(\\frac{1}{2})" + }, + { + "bbox": [ + 304, + 361, + 545, + 409 + ], + "type": "text", + "content": ", and the second transformation is derived from Equation 8. So the calculation of Hessian matrix is transformed into the calculation of Jacobian-Vector product." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 410, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 493 + ], + "type": "text", + "content": "Different from the within-domain consistency loss[36] which restricts target generated samples based on relative difference of source samples, the spectral consistency regularization only cares about the diversity of target model. This doesn't impose restriction on the direction of target adaptation, so our method can generate samples more consistent with target text without losing model diversity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "text", + "content": "The training objective of text-driven GAN adaptation is a weighted combination of directional loss and spectral consistency regularization loss " + }, + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{dir} + \\lambda \\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "text", + "content": ". Since different target domains have their own characteristic, it is required to tune the hyperparameter " + }, + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "text", + "content": " for better performance. To prevent the exhausting hyperparameter searching, we propose an adaptive loss reweighting method to balance the influence of these two loss items. Specifically, the adaptive weight " + }, + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "text", + "content": " is calculated as " + }, + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\lambda_{spectral} \\frac{\\|\\nabla_{G_L} \\mathcal{L}_{dir}\\|}{\\|\\nabla_{G_L} \\mathcal{L}_{reg}\\|}" + }, + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "inline_equation", + "content": "G_L" + }, + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "text", + "content": " denotes the last layer of generator and " + }, + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\lambda_{spectral}" + }, + { + "bbox": [ + 304, + 494, + 545, + 640 + ], + "type": "text", + "content": " is a manually specified hyperparameter, typically 1.0 is an appropriate choice." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 647, + 503, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 503, + 660 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 503, + 660 + ], + "type": "text", + "content": "3.5. Granularity Adaptive Regularization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "For text-driven GAN adaptation problem, the adaptation granularity of different target domains varies from texture to structure. For example, the photo-to-sketch adaptation mainly focuses on the appearance and texture change," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7023" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 541, + 309 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 541, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 541, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 541, + 309 + ], + "type": "image", + "image_path": "a68ea668e00bafe197ea5b25168a2e57373e444b35b178b656f32ba89d709e39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 314, + 547, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 314, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 547, + 348 + ], + "type": "text", + "content": "Figure 4. Visual results of our method and state-of-the-art StyleGAN-NADA [5] for different GAN adaptations. The top row shows the source text and target text description. The left column presents samples generated by source generator. Our method not only generates samples described by target text with high fidelity, but also produces diverse and identity consistent images corresponding to source domain." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 367, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 287, + 437 + ], + "type": "text", + "content": "while the werewolf domain has more variations on semantic structure. Regularization of diversity in the whole granularity will restrict the adaptation performance. To alleviate this problem, we propose granularity adaptive regularization based on the disentangled characteristic of StyleGAN latent code." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "text", + "content": "The latent codes injected into different layers in StyleGAN influence different granularities. The style code in low resolution represents high-level aspects such as pose and shape, that in middle resolution controls facial features and hairstyle, and that in high resolution influences color scheme and local textures. Specifically, we use the " + }, + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "inline_equation", + "content": "W+" + }, + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "text", + "content": " space as input space for Jacobian matrix calculation, where each input latent code consists of 18 512-dimensional vectors so both " + }, + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "text", + "content": " in Equation 4 are in " + }, + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "inline_equation", + "content": "R^{18 \\times 512}" + }, + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "text", + "content": ". By masking " + }, + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "text", + "content": " with a mask vector " + }, + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "inline_equation", + "content": "m \\in \\{0,1\\}^{18}" + }, + { + "bbox": [ + 46, + 439, + 289, + 594 + ], + "type": "text", + "content": " for different layers, we can specify the granularity of variations involved in the regularization. The calculation of Hessian trace under mask is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 603, + 287, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 603, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 67, + 603, + 287, + 633 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {T r a c e} (H (z)) = \\mathbb {E} [ (v \\odot m) ^ {T} H (z) (v \\odot m) ] (13) \\\\ = \\mathbb {E} [ \\| J _ {\\phi} (z) (v \\odot m) \\| ^ {2} ]. (14) \\\\ \\end{array}", + "image_path": "e87cb52f9064dffeb30394b6dbd4ed2cfea735682a9473548394d425a3b557c0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 641, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 689 + ], + "type": "text", + "content": "Following previous convention, we divide the style code for 18 layers into 3 groups, which are for coarse, middle and fine scale. The results of different mask strategies are shown in Figure 6." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "To explore the best mask strategy for different target domains, we propose to use an adaptive soft mask vector" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\{\\tilde{m} |\\tilde{m}\\in R^{18},\\| \\tilde{m}\\| = 1\\}" + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "text", + "content": " for all layers. During training, the mask vector is optimized with respect to the overall training objective. To reduce the directional loss, the mask vector will assign less value to the latent code corresponding to the granularity that changes most, while other values will increase to preserve the diversity of source generator." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 453, + 382, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 382, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 382, + 466 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "type": "text", + "content": "In this section, we will show the qualitative and quantitative results of our method. We illustrate the generated results for a wide range of target domains from style and texture changes to shape and semantic modifications. We also compare the proposed spectral consistency regularization with other regularization methods for diversity preservation. Next, we perform an ablation study on our method to evaluate the effectiveness of each component. Finally, we demonstrate the applications of text-guidance GAN adaptation, including image-to-image translation and image editing. To verify the generalization of proposed spectral consistency regularization, we also apply it to one-shot GAN adaptation, which is shown in supplementary appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "text", + "content": "We use the StyleGANv2 [11] generator pretrained on FFHQ as the source generator. During domain adaptation, we optimize all the parameters of generator except for the mapping network and toRGB layers. We use the Adam [12] optimizer with learning rate 0.001. The " + }, + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{spectral}}" + }, + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "text", + "content": " is set to 1.0. For most target domains, only 300 iterations are required to achieve convergence. Following CLIP, we use 79" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7024" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 541, + 366 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 541, + 366 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 541, + 366 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 541, + 366 + ], + "type": "image", + "image_path": "d90133782b5c4ac0795191d31b2f9fd064804e7e8e0d46a518e9d9dbe4b8aa23.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 371, + 546, + 395 + ], + "lines": [ + { + "bbox": [ + 46, + 371, + 546, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 546, + 395 + ], + "type": "text", + "content": "Figure 5. Comparison results of our spectral consistency regularization with other regularization methods. Detailed explanations of these methods are in Appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 414, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 450 + ], + "type": "text", + "content": "manually designed prompts like \"a photo of a...\" with provided target domain description and feed them to text encoder of CLIP to get the embeddings of target domain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 456, + 165, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 165, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 165, + 468 + ], + "type": "text", + "content": "4.1. Comparison Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 665 + ], + "type": "text", + "content": "Qualitative Comparison In Figure 4, we show the comparison results of our method with state-of-the-art model StyleGAN-NADA [5] for a wide range of target domains, which varies from texture changes like sketch and Mona Lisa paintings to geometric change like werewolf and Pixar style. The results demonstrate that our method not only generates highly stylistic images consistent with target text description for different target domains, but also produces images with diversity inherited from the pretrained source generator. Compared with StyleGAN-NADA which has obvious mode collapse problem like the mouth pattern in sketch and hair in werewolf, our model generates target images with better identity consistency. This proves that the spectral consistency regularization can preserve the diversity of source domain. In the supplementary appendix, we present additional visual results for the dogs and cars domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "We also perform text-driven GAN adaptation experiments with other regularization methods, including the Selective Cross-modal Consistency (SCC) loss[35], Within-Domain Consistency (WDC) loss[36], the Mode" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 414, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 544 + ], + "type": "text", + "content": "Seeking (MS) loss[14], Perceptual Path Length (PPL) regularization[11] and Cross Domain Correspondence (CDC) regularization[17]. Detailed explanations of these regularization methods are in supplementary appendix. As shown in Figure 5, SCC, WDC and MS regularization impose too strong regularization to target generator and restrict the domain specific attributes for target domain. Suffering from mode collapse problem, the generation results of PPL and CDC exhibits the same pattern across different samples. In comparison, our method has a better balance between the diversity and stylization of target generator." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 545, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 617 + ], + "type": "text", + "content": "Quantitative Comparison Besides our proposed trace norm of Hessian Matrix in Equation 3, we also leverage the Perceptual Path Length(PPL) [10] for quantitative diversity comparison. It measures the perceptually-based pairwise image distance [34] for a linear interpolation path in the latent path. The average PPL in the latent space " + }, + { + "bbox": [ + 304, + 545, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 545, + 545, + 617 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 622, + 566, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 622, + 566, + 649 + ], + "spans": [ + { + "bbox": [ + 296, + 622, + 566, + 649 + ], + "type": "interline_equation", + "content": "P P L _ {\\mathcal {Z}} = \\mathbb {E} \\left[ \\right. \\frac {1}{\\epsilon^ {2}} d \\left( \\right.G \\left(\\operatorname {s l e r p} \\left(\\mathbf {z} _ {1}, \\mathbf {z} _ {2}; t\\right)\\right), G \\left(\\operatorname {s l e r p} \\left(\\mathbf {z} _ {1}, \\mathbf {z} _ {2}; t + \\epsilon\\right)\\right)\\left. \\right]", + "image_path": "bffaf1e4e6aa073f0cddb0a9820ef3e10beba238214ede8f894d59363ba62806.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "d(.,.)" + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": " evaluates the perceptual distance between two images, and slerp denotes spherical interpolation since the input latent is normalized. The perceptual path length estimates the diversity of the generator via finite differences, and our method estimates the diversity analytically." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7025" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 105, + 99, + 491, + 215 + ], + "blocks": [ + { + "bbox": [ + 46, + 68, + 547, + 92 + ], + "lines": [ + { + "bbox": [ + 46, + 68, + 547, + 92 + ], + "spans": [ + { + "bbox": [ + 46, + 68, + 547, + 92 + ], + "type": "text", + "content": "Table 1. Comparison results for diversity by PPL and Hessian trace between different regularization methods. The PPL and Hessian trace for the pretrained source generator is 419.22 and 0.309." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 105, + 99, + 491, + 215 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 491, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 491, + 215 + ], + "type": "table", + "html": "
ResultsPhoto → SketchPhoto → Mona LisaHuman → WerewolfPhoto → Pixar
PPLTracePPLTracePPLTracePPLTrace
SCC547.340.526485.700.521428.450.343440.170.535
WDC378.250.090363.170.191311.340.116345.310.167
MS466.990.167331.060.233352.610.191377.880.358
PPL241.500.028209.090.061297.740.062300.190.063
CDC348.010.062259.920.079351.730.089299.490.111
NADA323.250.061281.590.098302.910.101343.540.112
Ours463.570.116321.430.140383.190.137353.800.181
", + "image_path": "eb11f4a96ebacdf737e5a9da984d45147e86c688735ea815847fbc71a96b7627.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 47, + 262, + 272, + 350 + ], + "blocks": [ + { + "bbox": [ + 46, + 240, + 287, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 287, + 262 + ], + "type": "text", + "content": "Table 2. User preference study. The numbers represent the percentage of users who prefer the images synthesized by our method." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 262, + 272, + 350 + ], + "lines": [ + { + "bbox": [ + 47, + 262, + 272, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 272, + 350 + ], + "type": "table", + "html": "
Model ComparisonQualityStyleAttributes
Ours vs. NADA [5]86.4%59.4%91.2%
Ours vs. SCC [35]64.2%86.2%65.0%
Ours vs. MS [14]59.4%83.6%54.2%
Ours vs. WDC [36]62.2%89.8%49.8%
Ours vs. PPL [11]92.4%71.4%95.0%
Ours vs. CDC [17]79.6%60.6%87.8%
", + "image_path": "0aac0a09d70cd591fbc2301822724f96557d945df9ecd5a2950df209ecadd8cc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 543 + ], + "type": "text", + "content": "The quantitative results are shown in Table 1. The mentioned strong regularization methods including SCC, WDC and MS have large values of PPL and Hessian trace at the cost of restricting adaptation effects. Our method performs competitive with these methods in diversity metrics, but does much better in transferring target styles. Compared with CDC and PPL, our method better preserve the diversity of source domain reflected in PPL and Hessian Trace. For all different target domains, our method outperforms previous state-of-the-art model StyleGAN-NADA for both PPL and Hessian trace. This benefits from that our method can alleviate the mode collapse problem with spectral consistency regularization and generate more diverse images without restricting the style adaptation results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "content": "User Study We conducted a user study to compare with other regularization methods. Specifically, participants were requested to select the superior synthesized samples in relation to three measurements: (1) image quality, (2) style consistency with the target description (3) attribute consistency with the source image. Five hundred samples were randomly generated for each comparative analysis. The aggregated results are presented in Table 2. The users strongly favor our method across all three aspects. In contrast to other methods, our method demonstrates a better tradeoff between style effects and attribute preservation. The spectral consistency regularization can better preserve the diversity of source domain while not restricting the style effects for target domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 243, + 400, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 400, + 256 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 400, + 256 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 262, + 547, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 547, + 418 + ], + "type": "text", + "content": "Granularity Adaptive Regularization. In Figure 6, we demonstrate generated results with different regularization strategies applied in " + }, + { + "bbox": [ + 304, + 262, + 547, + 418 + ], + "type": "inline_equation", + "content": "W+" + }, + { + "bbox": [ + 304, + 262, + 547, + 418 + ], + "type": "text", + "content": " space in Section 3.4. Regularization to the coarse scale will preserve the structure of source image but the diversity of fine features like hair will lose. The global regularization performs similarly to coarse regularization since the coarse features dominate the diversity of generator. Only applying regularization to the fine scale will generate high-frequency textures and the structure characteristic like necks will collapse. In comparison, our proposed granularity adaptive regularization both preserves the diversity of source domain in all scales and matches the styles of target domain." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 418, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 491 + ], + "type": "text", + "content": "Strength of regularization. In Figure 7, we show the generated samples with a linear interpolated loss weight " + }, + { + "bbox": [ + 304, + 418, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{spectral}}" + }, + { + "bbox": [ + 304, + 418, + 547, + 491 + ], + "type": "text", + "content": ". We can observe that with increasing " + }, + { + "bbox": [ + 304, + 418, + 547, + 491 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{spectral}}" + }, + { + "bbox": [ + 304, + 418, + 547, + 491 + ], + "type": "text", + "content": ", the generated samples maintain more diversity of source domain, and they also illustrate the most significant characteristic of target domain." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 491, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 547, + 624 + ], + "type": "text", + "content": "Choices of metric space. We conduct experiments about the metric space of spectral consistency regularization with different feature encoder " + }, + { + "bbox": [ + 304, + 491, + 547, + 624 + ], + "type": "inline_equation", + "content": "\\phi(x)" + }, + { + "bbox": [ + 304, + 491, + 547, + 624 + ], + "type": "text", + "content": " that evaluates the distance between image samples. Besides CLIP [19] image encoder in our method, we also leverage VGG [27] which was commonly used in style transfer [6], MoCo [8] of contrastive representation learning, ArcFace [2] for face recognition and the plain pixel space. As shown in Figure 8, compared to other feature encoders, the spectral consistency regularization with CLIP encoder shows best performance for preserving the identity of source image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 634, + 389, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 389, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 389, + 647 + ], + "type": "text", + "content": "4.3. Applications" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Image-to-Image Translation We combine the adapted generator in target domain with a GAN inversion encoder to implement image-to-image translation. Given a real-world image, we invert it to the latent code in " + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": " space via an e4e encoder [28], which is then fed to target generator to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7026" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 91, + 126, + 167 + ], + "blocks": [ + { + "bbox": [ + 73, + 72, + 105, + 83 + ], + "lines": [ + { + "bbox": [ + 73, + 72, + 105, + 83 + ], + "spans": [ + { + "bbox": [ + 73, + 72, + 105, + 83 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 91, + 126, + 167 + ], + "lines": [ + { + "bbox": [ + 49, + 91, + 126, + 167 + ], + "spans": [ + { + "bbox": [ + 49, + 91, + 126, + 167 + ], + "type": "image", + "image_path": "5fc7ca3b9b5b5c3f958f58a3b89e95fc1e22683196a7469b763cb5e2c2c18612.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 174, + 502, + 186 + ], + "lines": [ + { + "bbox": [ + 89, + 174, + 502, + 186 + ], + "spans": [ + { + "bbox": [ + 89, + 174, + 502, + 186 + ], + "type": "text", + "content": "Figure 6. Results of different regularization strategies for GAN adaptation from photo to Edvard Munch paintings." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 134, + 91, + 209, + 167 + ], + "blocks": [ + { + "bbox": [ + 159, + 72, + 189, + 82 + ], + "lines": [ + { + "bbox": [ + 159, + 72, + 189, + 82 + ], + "spans": [ + { + "bbox": [ + 159, + 72, + 189, + 82 + ], + "type": "text", + "content": "Global" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 91, + 209, + 167 + ], + "lines": [ + { + "bbox": [ + 134, + 91, + 209, + 167 + ], + "spans": [ + { + "bbox": [ + 134, + 91, + 209, + 167 + ], + "type": "image", + "image_path": "619c9ffed9b5792082c689dc9dc300f741848aed72236b0181a60a37dfb0e119.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 217, + 91, + 293, + 167 + ], + "blocks": [ + { + "bbox": [ + 243, + 72, + 274, + 82 + ], + "lines": [ + { + "bbox": [ + 243, + 72, + 274, + 82 + ], + "spans": [ + { + "bbox": [ + 243, + 72, + 274, + 82 + ], + "type": "text", + "content": "Coarse" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 217, + 91, + 293, + 167 + ], + "lines": [ + { + "bbox": [ + 217, + 91, + 293, + 167 + ], + "spans": [ + { + "bbox": [ + 217, + 91, + 293, + 167 + ], + "type": "image", + "image_path": "9024a80d7daccfd1347ef37f24a1da860714a87d93857f2ea591094ab964d0f2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 301, + 91, + 377, + 167 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 361, + 83 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 361, + 83 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 361, + 83 + ], + "type": "text", + "content": "Middle" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 91, + 377, + 167 + ], + "lines": [ + { + "bbox": [ + 301, + 91, + 377, + 167 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 377, + 167 + ], + "type": "image", + "image_path": "a4ea90735778d2df4a253071106db1e4d963cc09716d807c5ff78b4f12e7876b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 385, + 91, + 461, + 167 + ], + "blocks": [ + { + "bbox": [ + 414, + 72, + 435, + 83 + ], + "lines": [ + { + "bbox": [ + 414, + 72, + 435, + 83 + ], + "spans": [ + { + "bbox": [ + 414, + 72, + 435, + 83 + ], + "type": "text", + "content": "Fine" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 385, + 91, + 461, + 167 + ], + "lines": [ + { + "bbox": [ + 385, + 91, + 461, + 167 + ], + "spans": [ + { + "bbox": [ + 385, + 91, + 461, + 167 + ], + "type": "image", + "image_path": "dc931f537612bbd32897324360cc40b8bcdf707a9284110a879c9977b75ef8ae.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 469, + 91, + 545, + 167 + ], + "blocks": [ + { + "bbox": [ + 490, + 72, + 529, + 84 + ], + "lines": [ + { + "bbox": [ + 490, + 72, + 529, + 84 + ], + "spans": [ + { + "bbox": [ + 490, + 72, + 529, + 84 + ], + "type": "text", + "content": "Adaptive" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 469, + 91, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 469, + 91, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 469, + 91, + 545, + 167 + ], + "type": "image", + "image_path": "dd71d7613cfe1393d543968c28d55e2f7bd77ceb464c37c972e422415e27db0d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 48, + 202, + 287, + 281 + ], + "blocks": [ + { + "bbox": [ + 48, + 202, + 287, + 281 + ], + "lines": [ + { + "bbox": [ + 48, + 202, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 202, + 287, + 281 + ], + "type": "image", + "image_path": "35970361d29c69aed78b67ee96e0df1924c04fba09bbf3750eab2690da5b6332.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 285, + 287, + 307 + ], + "lines": [ + { + "bbox": [ + 47, + 285, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 287, + 307 + ], + "type": "text", + "content": "Figure 7. Generation results of linearly interpolated " + }, + { + "bbox": [ + 47, + 285, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{spectral}}" + }, + { + "bbox": [ + 47, + 285, + 287, + 307 + ], + "type": "text", + "content": " in regularization." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 51, + 319, + 282, + 387 + ], + "blocks": [ + { + "bbox": [ + 51, + 319, + 282, + 387 + ], + "lines": [ + { + "bbox": [ + 51, + 319, + 282, + 387 + ], + "spans": [ + { + "bbox": [ + 51, + 319, + 282, + 387 + ], + "type": "image", + "image_path": "a73e311b2e92cf70a3332c7cfe3833894b6ffd67784ccc44416b98c288562e96.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 392, + 286, + 414 + ], + "lines": [ + { + "bbox": [ + 47, + 392, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 286, + 414 + ], + "type": "text", + "content": "Figure 8. The generated results with spectral consistency regularization under the metric spaces defined by different encoders." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 434, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 470 + ], + "type": "text", + "content": "produce target image. As shown in Figure 9, our method can achieve high-quality image translation and preserve the identity of source image for different domains." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 472, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 554 + ], + "type": "text", + "content": "Image Editing In Figure 10, we demonstrate the image editing results performed on the target domain. We leverage the meaningful directions found by InterfaceGAN [25] to edit target images. We can observe that the editing directions from source domain still apply to target domains, which proves that the target domain preserves the semantic distribution of source domain." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 52, + 564, + 282, + 664 + ], + "blocks": [ + { + "bbox": [ + 52, + 564, + 282, + 664 + ], + "lines": [ + { + "bbox": [ + 52, + 564, + 282, + 664 + ], + "spans": [ + { + "bbox": [ + 52, + 564, + 282, + 664 + ], + "type": "image", + "image_path": "0e54b7b20cfa93fc789ffa81ce845f9de3b9158774154d649c0ccb41659a7370.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 667, + 287, + 712 + ], + "lines": [ + { + "bbox": [ + 46, + 667, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 667, + 287, + 712 + ], + "type": "text", + "content": "Figure 9. Image translation of real-world images to different target domains. Each column shows a target domain, and the top row is the text description for target domain. The transferred images represent both target style and the identity of source image." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 307, + 203, + 545, + 243 + ], + "blocks": [ + { + "bbox": [ + 307, + 203, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 307, + 203, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 203, + 545, + 243 + ], + "type": "image", + "image_path": "bf5abcc8f4d6fb9fb4919fb89e3edb59808e878cffd554a466fa09aa033e6f45.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 245, + 545, + 267 + ], + "lines": [ + { + "bbox": [ + 306, + 245, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 306, + 245, + 545, + 267 + ], + "type": "text", + "content": "Figure 10. Editing images in target domain for real-world images. The top row shows the edited attributes." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 296, + 378, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 296, + 378, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 296, + 378, + 308 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 317, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 425 + ], + "type": "text", + "content": "In this paper, we propose the spectral consistency regularization for text-driven generative domain adaptation. The key insight of our method is to build a quantitative diversity estimator to preserve the intra-domain diversity of source generator without restricting the adaptation of target style. We also introduce an adaptive regularization strategy for granularity-flexible adaptation. The experiments demonstrate our method greatly improves the generation results for a wide range of target domains." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 439, + 404, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 404, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 404, + 452 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 460, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 532 + ], + "type": "text", + "content": "This work was supported in part by the National Key R&D Program of China under Grant 2018AAA0102000, in part by National Natural Science Foundation of China: 62322211, 62236008, U21B2038, 62225207, U19B2038 and 61931008, and Youth Innovation Promotion Association of Chinese Academy of Sciences under Grant 2020108." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 545, + 365, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 545, + 365, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 545, + 365, + 558 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 310, + 566, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 310, + 566, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 566, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 310, + 566, + 545, + 620 + ], + "type": "text", + "content": "[1] Katherine Crowson, Stella Rose Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. ArXiv, abs/2204.08583, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 311, + 623, + 547, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 623, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 311, + 623, + 547, + 667 + ], + "type": "text", + "content": "[2] Jiankang Deng, J. Guo, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4685-4694, 2019. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 311, + 669, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 669, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 311, + 669, + 547, + 713 + ], + "type": "text", + "content": "[3] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868-12878, 2021. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7027" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 106 + ], + "type": "text", + "content": "[4] Kevin Frans, Lisa B. Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. ArXiv, abs/2106.14843, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "type": "text", + "content": "[5] Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ArXiv, abs/2108.00946, 2021. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 152, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 152, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 53, + 152, + 288, + 196 + ], + "type": "text", + "content": "[6] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 197, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 288, + 229 + ], + "type": "text", + "content": "[7] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. ArXiv, abs/2004.02546, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 231, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 231, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 231, + 288, + 285 + ], + "type": "text", + "content": "[8] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726-9735, 2020. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 286, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 286, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 286, + 288, + 330 + ], + "type": "text", + "content": "[9] Michael F. Hutchinson. A stochastic estimator of the trace of the influence matrix for laplacian smoothing splines. Communications in Statistics - Simulation and Computation, 18:1059-1076, 1989. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 376 + ], + "type": "text", + "content": "[10] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4396-4405, 2019. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 431 + ], + "type": "text", + "content": "[11] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8107-8116, 2020. 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 433, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 288, + 454 + ], + "type": "text", + "content": "[12] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2015. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 488 + ], + "type": "text", + "content": "[13] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. ArXiv, abs/2111.03186, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 288, + 544 + ], + "type": "text", + "content": "[14] Qi Mao, Hsin-Ying Lee, Hung-Yu Tseng, Siwei Ma, and Ming-Hsuan Yang. Mode seeking generative adversarial networks for diverse image synthesis. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1429-1437, 2019. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 578 + ], + "type": "text", + "content": "[15] Sangwoo Mo, Minsu Cho, and Jinwoo Shin. Freeze discriminator: A simple baseline for fine-tuning gans. ArXiv, abs/2002.10964, 2020. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 579, + 288, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 623 + ], + "type": "text", + "content": "[16] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2750-2758, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "text", + "content": "[17] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A. Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10738-10747, 2021. 3, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[18] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and D. Lischinski. Styleclip: Text-driven manipulation of" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "stylegan imagery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2065-2074, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 162 + ], + "type": "text", + "content": "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[20] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[21] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 545, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 275 + ], + "type": "text", + "content": "[22] Scott E. Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 276, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 352 + ], + "type": "text", + "content": "[23] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "text", + "content": "[24] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9240-9249, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[25] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:2004-2018, 2022. 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 487 + ], + "type": "text", + "content": "[26] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1532-1540, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "type": "text", + "content": "[27] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2015. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "text", + "content": "[28] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 40:1 - 14, 2021. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "text", + "content": "[29] Binxu Wang and Carlos R. Ponce. The geometry of deep generative image models and its applications. ArXiv, abs/2101.06006, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 602, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 645 + ], + "type": "text", + "content": "[30] Jiayu Xiao, Liang Li, Chaofei Wang, Zhengjun Zha, and Qingming Huang. Few shot generative model adaption via relaxed spatial structural alignment. ArXiv, abs/2203.04121, 2022. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[31] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316-1324, 2018. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7028" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 361 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 148 + ], + "type": "text", + "content": "[32] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Benton C. Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. ArXiv, abs/2206.10789, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 150, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 215 + ], + "type": "text", + "content": "[33] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N. Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5908-5916, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "text", + "content": "[34] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 274, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 316 + ], + "type": "text", + "content": "[35] Yabo Zhang, Mingshuai Yao, Yuxiang Wei, Zhilong Ji, Jinfeng Bai, and Wangmeng Zuo. Towards diverse and faithful one-shot adaption of generative adversarial networks. ArXiv, abs/2207.08736, 2022. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 318, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 318, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 318, + 287, + 361 + ], + "type": "text", + "content": "[36] Peihao Zhu, Rameen Abdal, John C. Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. ArXiv, abs/2110.08398, 2021. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7029" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Text2Performer_ Text-Driven Human Video Generation/45c9b878-7941-4c19-aa7b-acd77ad00875_content_list.json b/2023/Text2Performer_ Text-Driven Human Video Generation/45c9b878-7941-4c19-aa7b-acd77ad00875_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..39bfcb69bec0e73ba46d163debd52029c8030304 --- /dev/null +++ b/2023/Text2Performer_ Text-Driven Human Video Generation/45c9b878-7941-4c19-aa7b-acd77ad00875_content_list.json @@ -0,0 +1,2015 @@ +[ + { + "type": "text", + "text": "Text2 Performer: Text-Driven Human Video Generation", + "text_level": 1, + "bbox": [ + 202, + 130, + 767, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuming Jiang $^{1}$ Shuai Yang $^{1}$ Tong Liang Koh $^{1}$ Wayne Wu $^{2}$ Chen Change Loy $^{1}$ Ziwei Liu $^{1\\boxtimes}$ $^{1}$ S-Lab, Nanyang Technological University $^{2}$ Shanghai AI Laboratory", + "bbox": [ + 86, + 179, + 901, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yuming002, shuai.yang, koht0029, ccloy, ziwei.liu}@ntu.edu.sg wuwenyan0503@gmail.com", + "bbox": [ + 111, + 219, + 879, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b424bc7500ed5243fbb519ce0bb57e99df79987f7d9da9fb46f01047fc84abdb.jpg", + "image_caption": [ + "Figure 1: High-resolution videos generated by Text2Performer. The videos are generated by taking the text descriptions as the only input. The generated videos contain diverse appearances and flexible motions. Identities are well maintained." + ], + "image_footnote": [], + "bbox": [ + 86, + 271, + 500, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aefea0d56de1b949c95bb51e5647b73136b96d1b309a1cfae66a865a1cb7a6ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 422, + 148, + 454 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This woman wears a dress. It has long sleeves, and it is of long length. The texture of it is graphic. She is swinging to the right.", + "bbox": [ + 158, + 421, + 475, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/53d11366d4394c29962c13e7b70dd849945205b7218de02f41d9fafade57343a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 271, + 883, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/994493edf4f6027469fc298979094b0160332d4e85434250aa08a58bbbb0f596.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 426, + 558, + 454 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The person is wearing a sleeveless dress. It is of short length. Its pattern is pure color. She turns right from the front to the side.", + "bbox": [ + 568, + 421, + 864, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 521, + 313, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-driven content creation has evolved to be a transformative technique that revolutionizes creativity. Here we study the task of text-driven human video generation, where a video sequence is synthesized from texts describing the appearance and motions of a target performer. Compared to general text-driven video generation, human-centric video generation requires maintaining the appearance of synthesized human while performing complex motions. In this work, we present Text2Performer to generate vivid human videos with articulated motions from texts. Text2Performer has two novel designs: 1) decomposed human representation and 2) diffusion-based motion sampler. First, we decompose the VQVAE latent space into human appearance and pose representation in an unsupervised manner by utilizing the nature of human videos. In this way, the appearance is well maintained along the generated frames. Then, we propose continuous VQ-diffuser to sample a sequence of pose embeddings. Unlike existing VQ-based methods that operate in the discrete space, continuous VQ-diffuser directly outputs the continuous pose embeddings for better motion modeling. Finally, motion-aware masking strategy is designed to mask the pose embeddings spatial-temporally to enhance the temporal coherence. Moreover,", + "bbox": [ + 73, + 553, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to facilitate the task of text-driven human video generation, we contribute a Fashion-Text2Video dataset with manually annotated action labels and text descriptions. Extensive experiments demonstrate that Text2Performer generates high-quality human videos (up to $512 \\times 256$ resolution) with diverse appearances and flexible motions. Our project page is https://yumingj.github.io/projects/Text2Performer.html", + "bbox": [ + 496, + 523, + 893, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 662, + 632, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since its emergence, text-guided image synthesis (e.g. DALLE [37, 38]) has attracted substantial attention. Recent works [9, 10, 14, 18, 40, 47] have demonstrated fascinating performance for the quality of synthesized images and their consistency with the texts. Beyond image generation, text-driven video generation [22, 24, 44, 53] is an advanced topic to explore. Existing works rely on large-scale datasets to drive large models. Although they have achieved surprising performance on general objects, when applied to the generation of some specific tasks, such as generating videos of garment presents for e-commerce websites, they fail to generate plausible results. Take the CogVideo [24] as an example. As shown in Fig. 2, the generated human objects contain incomplete human structures, and the temporal", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22747", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3d5a6148d0c6be5ae87c6f5da2697825f347aa3b6a8b31c19cc0f87c3e3d7b7f.jpg", + "image_caption": [ + "The person is wearing a sleeveless dress. It is of short length. Its pattern is pure color. She turns right from the front to the side.", + "Figure 2: Results of General Large Text-to-Video Models. We use the pretrained general large Text-to-Video Models [24] to generate videos using the same texts as the right example of Fig. 1. The result fails to generate complete human structures and maintains temporal consistency." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 464, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "consistency is poorly maintained. On the other hand, these methods need billions of training data, which hampers the application to those specific tasks (e.g. human video generation) without a large amount of paired data.", + "bbox": [ + 75, + 277, + 468, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, it is worthwhile to explore text-to-video generation in human video generation, which has numerous applications [31, 45]. In this paper, we focus on the task of text-driven human video generation. Compared to general text-to-video generation, text-driven human video generation poses several unique challenges: 1) The human structure is articulated. The joint movements of different body components form many complicated out-of-plane motions, e.g., rotations. 2) When performing complicated motions, the appearance of the synthesized human should remain the same. For example, the appearance of a target human after turning around should be consistent with that at the first beginning. In sum, to achieve high-fidelity human video generation, consistent human representation and complicated human motions should be carefully modeled.", + "bbox": [ + 75, + 339, + 468, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a novel text-driven human video generation framework Text2Performer to handle consistent human representations and complex out-of-plane motions. As shown in Fig. 1, given texts describing appearance and motions, Text2Performer is able to generate temporally consistent human videos with complete human structures and unified human appearances. Text2Performer is built upon VQVAE-based frameworks [4, 11, 18, 52]. In Text2Performer, thanks to the specific nature of human videos which shares the same objects across the frames within one video, VQVAE latent space can be decomposed into appearance and pose representations. With the decomposed VQ-space, videos are generated by sampling appearance representation and a sequence of pose representations separately. This decomposition contributes to the maintenance of human identity. Besides, it makes the motion modeling more tractable, as the motion sampler does not need to take the appearance information into consideration.", + "bbox": [ + 75, + 566, + 468, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To model complicated human motions, a novel continuous VQ-diffuser is proposed to sample a sequence of meaningful pose representations. The architecture of the continuous VQ-diffuser is transformer. The key difference to", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the previous transformer-based samplers [4, 11, 52] is that the continuous VQ-diffuser directly predicts the continuous pose embeddings rather than their indices in the codebook. After predicting continuous pose embeddings, we also make use of the rich embeddings stored in the codebook by retrieving the nearest embeddings of the predicted embeddings from the codebook. Predicting continuous embeddings alleviates the one-to-many prediction issue in previous discrete methods and the use of codebook constrains the prediction space. With this design, more temporally coherent human motions and more complete structures of human frames are sampled. In addition, we borrow the idea of diffusion models [4, 18, 23, 39, 26] to progressively predict the long sequence of the pose embeddings. We propose a motion-aware masking strategy to sample the pose embeddings of the first frame and last frame firstly. Then the pose embeddings of the intermediate frames are gradually diffused. The motion-aware masking strategy enhances the completeness of human structures and temporal coherence.", + "bbox": [ + 496, + 90, + 890, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To facilitate the task of text-guided human video generation, we propose the Fashion-Text2Video Dataset. It is built upon the FashionDataset [56], which consists of 600 human videos performing the fashion show. We manually segment the whole video into clips and label the motion types. Each clip is performing one motion. With the manually labeled motion labels, we then pair them with text descriptions.", + "bbox": [ + 496, + 378, + 890, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 517, + 484, + 821, + 497 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present and study the task of text-guided human video generation. Our proposed Text2Performer can be well trained and have generative abilities with only a small amount of data for training.", + "- We propose to decompose the VQ-space into appearance and pose representations. The decomposition is achieved by making use of the nature of human videos, i.e., the motions are performed under one identity (appearance) across the frames. The decomposition of VQ-space improves the appearance coherence among frames and eases motion modeling.", + "- We propose the continuous VQ-diffuser to predict continuous pose embeddings with the pose codebook. The final continuous pose embeddings are iteratively predicted with the guidance of the motion-aware masking scheme. These designs contribute to generated frames of high quality and temporal coherence.", + "- We construct the Fashion-Text2Video Dataset with human motion labels and text descriptions to facilitate the research on text-driven human video generation." + ], + "bbox": [ + 517, + 502, + 890, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 829, + 640, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Synthesis. Similar to image generation, GAN [5, 16, 29] and VQVAE [11, 52] are two common paradigms in the field of video synthesis [49]. MoCoGAN-HD [48] and", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22748", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/33a0196bd4e55031ba67b38e1a59442ce858dbb553d648b67471fbe60c8f35e9.jpg", + "image_caption": [ + "(a) Sampling from the Decomposed VQ-Space" + ], + "image_footnote": [], + "bbox": [ + 104, + 111, + 699, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a0584001eead1165443f0f08c553bef0faac85e80dc3a85e0a4b65ab56a3680c.jpg", + "image_caption": [ + "(b) Motion Sampling with Continuous VQ-Diffuser" + ], + "image_footnote": [], + "bbox": [ + 106, + 215, + 697, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Legend", + "text_level": 1, + "bbox": [ + 774, + 93, + 825, + 107 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/59bff63d518594dfd52fb66d2b0e01f7401284000dd76265eb95db0d94597556.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 113, + 763, + 128 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text", + "bbox": [ + 776, + 117, + 799, + 127 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/85aa731ed744aed458b8ab43f654a68ecc3fd7659e9f1696b078d5429e5ad47b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 137, + 763, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text encoder", + "bbox": [ + 776, + 143, + 836, + 152 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3ea013016efacdbdfdfd7dbf7482211f991e481534cbd09049ff6b867dec737b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 164, + 763, + 184 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VQ-VAE Decoder", + "bbox": [ + 776, + 167, + 861, + 179 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4d5875d527c7998959fbb4570f0f14ad216a78d262cc73728a255b1ab9d69fcc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 191, + 763, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sampler", + "bbox": [ + 774, + 195, + 815, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/59d663e827f68c8c2970620179d4086d557fbb69f2651f052fab6af0b73ad28f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 220, + 763, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text embedding", + "bbox": [ + 774, + 220, + 826, + 241 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f6d93eaeed2b6d810a58daebf6a992bacefca2868301507d37166e06b38d8158.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 250, + 763, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continuous embedding", + "bbox": [ + 774, + 256, + 830, + 276 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ed1d88999b8746df380350992fbd77d8f6fd9f592b41f808cdf834152a611baf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 284, + 763, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked continuous embedding", + "bbox": [ + 774, + 290, + 864, + 311 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1d54f708c0fbfdb0ebdbe1ecf157277077373dd002308c207fe9dcb7b18a3c23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 320, + 763, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Codebook", + "bbox": [ + 774, + 325, + 823, + 335 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/17cb31432597bf5f7005976c93b0969fb9e7fb2b6d59d33788f6a086476bf9de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 348, + 763, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Appearance-related", + "bbox": [ + 774, + 352, + 864, + 362 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b7b20605642c93df6f29d3b3f80b918202fcc6cfedc2e116593d56f839f8ab30.jpg", + "image_caption": [ + "Figure 3: Overview of Text2Performer with (a) Sampling from the Decomposed VQ-Space and (b) Motion Sampling with Continuous VQ-Diffuser. Given a text, we first sample the target appearance features $\\hat{f}_a$ and exemplar pose features $\\hat{f}_p^0$ conditioned on the language features $e_a$ extracted by a pretrained text model. The motion sequence $\\hat{\\mathbf{F}}_p$ is then sampled by our proposed continuous VQ-Diffuser. The continuous VQ-Diffuser takes the extracted language features $e_m$ and $\\hat{f}_p^0$ as inputs. The prediction of continuous motion sequences $\\hat{\\mathbf{F}}_p$ starts with fully masked pose features and predicts continuous pose embeddings $\\{\\bar{f}_p^1,\\dots,\\bar{f}_p^n\\}$ with Non-Causal Transformer. The nearest neighbor pose embeddings $\\{\\hat{f}_p^1,\\dots,\\hat{f}_p^n\\}$ are then retrieved from the pose codebook $\\mathcal{C}_p$ . Guided by a motion-aware masking strategy, the continuous VQ-Diffuser is iteratively applied until the whole motion sequence is unmasked. The final videos are generated by feeding the continuous pose features and appearance features into the decoder of VQVAE." + ], + "image_footnote": [], + "bbox": [ + 736, + 373, + 763, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion-related", + "bbox": [ + 774, + 376, + 844, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "StyleGAN-V [46] harness the powerful StyleGAN [29, 30] to generate videos. DI-GAN [55] models an implicit neural representation to generate videos. Brooks et al. [6] propose to regularize the temporal consistency via a temporal discriminator. These works focus on unconditional video generation. As for the VQVAE-based methods, VideoGPT [54] firstly extends the idea of VQVAE and autoregressive transformer to video generation. The following works introduce time-agnostic VQGAN with time-sensitive transformer [15] and masked sequence modeling [20] to further improve the performance. These VQVAE-based video synthesis frameworks accept conditions appended to the beginning of the token sequences. Our proposed Text2Performer is developed on VQVAE. The differences lie in that the designs of VQ-space decomposition and continuous sampling. Recently, there are some concurrent works [22, 24, 44, 53] for text-to-video generation. These methods are designed for general objects, while our Text2Performer has the design to separate human appearance and pose representations, which is specific to human video generation.", + "bbox": [ + 75, + 551, + 472, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human Content Generation and Manipulation. Existing works focus on human images, and they are mainly divided into two types: unconditional generation and conditional", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "generation. StyleGAN-Human [12] employs StyleGAN to synthesize high-fidelity human images. TryOnGAN [32] and HumanGAN [41] generate human images conditioned on given human poses. Text2Human [27] proposes to generate human images conditioned on texts and poses. We focus on human video generation. For human content manipulation, pose transfer [1, 7, 33, 34, 35] is a popular topic. Pose transfer deals with video data. The task is to transfer the poses from the reference videos to the source video. Chan et al. [7] take human poses as inputs and the desired video is generated by an image-to-image translation network. Our task differs in that we only take texts as inputs and exemplar images are not compulsory. Motion synthesis works [19, 36] generate human motions in form of 3D human representations (e.g., SMPL and skeleton space), while our method generates human motions at the image level.", + "bbox": [ + 496, + 551, + 893, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Text2 Performer", + "text_level": 1, + "bbox": [ + 500, + 813, + 658, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 3, our proposed Text2Performer synthesizes desired human videos by taking the texts as inputs $(\\pmb{t}_a$ for appearance and $t_m$ for motions). To ensure consistent human representation, we propose to decompose the", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22749", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/18d2a22c6a25d1438a3ae1f1bc680e2ec52eb9ca89b38847f73cfceafa6ab0c8.jpg", + "image_caption": [ + "Figure 4: Pipeline of Decomposed VQ-Space. Two different frames $I_0$ and $I_k$ of a video serve as the identity frame and the pose frame to provide appearance and pose features, respectively. We apply data augmentations to $I_k$ to erase its appearance information to avoid information leakage. Two codebooks are built to store the pose features and appearance features. The quantized features are finally fed into the decoder to reconstruct the pose frame $I_k$ ." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 472, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VQ-space into appearance representation $\\pmb{f}_a$ and pose representation $\\pmb{f}_p$ as shown in Fig. 4. With the decomposed VQ-space, we sample the human appearance features $\\hat{\\pmb{f}}_a$ and exemplar pose features $\\hat{\\pmb{f}}_p^0$ according to $\\pmb{t}_a$ . To model the motion dynamics, we propose continuous VQ-diffuser as the motion sampler. It progressively diffuses the masked sequence of pose features until the whole sequence is un-masked. The final video clip is generated by feeding the appearance and pose embeddings into the decoder.", + "bbox": [ + 75, + 422, + 472, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Decomposed VQ-Space", + "text_level": 1, + "bbox": [ + 76, + 573, + 294, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human video generation requires consistent human appearance, i.e., the face and clothing of the target person should remain unchanged while performing the motion. Vanilla VQVAE [52] encodes images into a unified feature representation and builds a codebook on it. However, such design is prone to generate drifted identities along the video frames. To overcome this problem, we propose to decompose the VQ-space into the appearance representation $\\pmb{f}_a$ and pose representation $\\pmb{f}_p$ in an unsupervised manner. With the decomposed VQ-space, we can separately sample $\\hat{\\pmb{f}}_a$ and a sequence of $\\hat{\\pmb{f}}_p$ to generate the desired video.", + "bbox": [ + 75, + 597, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In human video data, different frames in one video share the same human identity. We utilize this property to train the decomposed VQ-space [13, 25]. The pipeline is shown in Fig. 4. Given a video $\\mathbf{V} = \\{\\mathbf{I}_0, \\mathbf{I}_1, \\dots, \\mathbf{I}_n\\}$ , we use its first frame $\\mathbf{I}_0$ for appearance information and another randomly sampled frame $\\mathbf{I}_k$ for pose information. $\\mathbf{I}_0$ and $\\mathbf{I}_k$ are fed into two encoders $E_a$ and $E_p$ , respectively. The two branches share the encoder $E_s$ . To prevent $\\mathbf{I}_k$ from leaking appearance information, data augmentations (e.g. color jitter", + "bbox": [ + 75, + 765, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tering and Gaussian blur) are applied to $\\mathbf{I}_k$ before the pose branch. $f_{a}$ and $f_{p}$ are obtained as", + "bbox": [ + 496, + 90, + 890, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {a} = E _ {a} \\left(E _ {s} \\left(\\mathbf {I} _ {0}\\right)\\right), \\boldsymbol {f} _ {p} = E _ {p} \\left(E _ {s} \\left(\\mathbf {I} _ {k} ^ {a}\\right)\\right)), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 136, + 890, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{I}_k^a$ is the augmented $\\mathbf{I}_k$ . To make $f_{p}$ learn compact and necessary pose information, we make the spatial size of $f_{p}$ smaller. Given $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ , we empirically find that the optimal spatial sizes for $f_{a}$ and $f_{p}$ are $H / 16 \\times W / 16$ and $H / 64 \\times W / 64$ , respectively.", + "bbox": [ + 496, + 166, + 893, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining $\\pmb{f}_a$ and $\\pmb{f}_p$ , two codebooks $\\mathcal{C}_a$ and $\\mathcal{C}_p$ are built to store the appearance and pose embeddings. The quantized feature $\\hat{\\pmb{f}}$ given codebook $\\mathcal{C}$ is obtained by", + "bbox": [ + 496, + 244, + 893, + 290 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {f}} = \\operatorname {Q u a n t} (\\boldsymbol {f}) := \\left\\{\\underset {c _ {k} \\in \\mathcal {C}} {\\operatorname {a r g m i n}} \\| f _ {i j} - c _ {k} \\| _ {2} \\mid f _ {i j} \\in \\boldsymbol {f} \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 303, + 890, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the quantized $\\hat{f}_a$ and $\\hat{f}_p$ , we then feed them into decoder $D$ to reconstruct the target pose frame $\\hat{\\mathbf{I}}_k$ :", + "bbox": [ + 496, + 345, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {I}} _ {k} = D ([ \\hat {\\boldsymbol {f}} _ {a}, D _ {p} (\\hat {\\boldsymbol {f}} _ {p}) ]), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 392, + 890, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[\\cdot ]$ is the concatenation operation, and $D_{p}$ upsamples $\\hat{\\pmb{f}}_p$ to make it have the same resolution as $\\hat{\\pmb{f}}_a$ . The whole network (including the encoders, decoders, and codebooks) is trained using:", + "bbox": [ + 496, + 424, + 890, + 487 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\left\\| \\mathbf {I} _ {k} - \\hat {\\mathbf {I}} _ {k} \\right\\| _ {1} + \\left\\| \\operatorname {s g} (\\hat {\\boldsymbol {f}} _ {a}) - \\boldsymbol {f} _ {a} \\right\\| _ {2} ^ {2} + \\left\\| \\operatorname {s g} (\\boldsymbol {f} _ {a}) - \\hat {\\boldsymbol {f}} _ {a} \\right\\| _ {2} ^ {2} \\\\ + \\left\\| \\operatorname {s g} \\left(\\hat {\\boldsymbol {f}} _ {p}\\right) - \\boldsymbol {f} _ {p} \\right\\| _ {2} ^ {2} + \\left\\| \\operatorname {s g} \\left(\\boldsymbol {f} _ {p}\\right) - \\hat {\\boldsymbol {f}} _ {p} \\right\\| _ {2} ^ {2}, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 500, + 890, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{sg}(\\cdot)$ is the stop-gradient operation.", + "bbox": [ + 500, + 590, + 777, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the decomposed VQ-space, an additional sampler is trained to sample $\\hat{\\pmb{f}}_a$ and $\\hat{\\pmb{f}}_p^0$ . This sampler has the same design as previous methods [4, 27].", + "bbox": [ + 496, + 607, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Continuous VQ-Diffuser", + "text_level": 1, + "bbox": [ + 500, + 666, + 725, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt the absorbing diffusion transformer [2, 4, 8, 18] to sample the motion, a sequence of pose embeddings $\\hat{\\mathbf{F}}_p = \\{\\hat{\\pmb{f}}_p^1,\\hat{\\pmb{f}}_p^2,\\dots,\\hat{\\pmb{f}}_p^n\\}$ from the learned pose codebook $\\mathcal{C}_p$ Different from autoregressive models [11, 52] that make predictions in a fixed order, absorbing diffusion transformer predicts multiple codebook indices in parallel. The prediction of codebook indices starts from $\\mathbf{F}_p^0$ i.e., fully masked $\\mathbf{F}_p^m$ . The prediction at time step $t$ is represented as follows:", + "bbox": [ + 496, + 691, + 890, + 814 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} _ {p} ^ {t} \\sim q _ {\\theta} \\left(\\mathbf {F} _ {p} ^ {t} \\mid \\mathbf {F} _ {p} ^ {t - 1}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 829, + 890, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta$ denotes the parameters of the transformer sampler.", + "bbox": [ + 500, + 859, + 888, + 873 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To model human motions, we propose 1) continuous space sampling, and 2) motion-aware masking strategy.", + "bbox": [ + 500, + 875, + 890, + 906 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22750", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e8efb86c90da220bbeed4ccc74257ceae82e6ca45e18df6a2ad5e09055a4da72.jpg", + "image_caption": [ + "Figure 5: Comparison with Image Samplers. (a) Discrete VQ Sampler predicts discrete codebook indices, which are trained by cross-entropy loss. (b) Continuous Sampler operates on unconstrained continuous RGB spaces. (c) Our VQ-Diffuser predicts continuous embeddings but also utilizes rich embeddings stored in codebooks." + ], + "image_footnote": [], + "bbox": [ + 89, + 90, + 459, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1 Continuous Space Sampling", + "text_level": 1, + "bbox": [ + 76, + 375, + 328, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 5(a), previous VQVAE-based methods [4, 8, 11, 18, 52] sample tokens in a discrete space. They predict codebook indices to form quantized features, which are then fed into the decoder to generate images. However, sampling at the discrete space is hard to converge because of redundant codebooks, which causes the one-to-many mapping prediction problem. This would lead to meaningless predicted pose sequences.", + "bbox": [ + 75, + 398, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In continuous VQ-diffuser $S_{\\theta}$ , we propose to train the non-causal transformer in the continuous embedding space as shown in Fig. 5(c). $S_{\\theta}$ directly predicts continuous pose embeddings $\\{\\pmb{f}_p^k\\}$ . $S_{\\theta}$ is trained to predict the full continuous $\\bar{\\mathbf{F}}_p$ from the masked pose sequences $\\mathbf{F}_p^m$ conditioned on the motion description $\\pmb{t}_{m}$ and the initial pose feature $\\hat{\\pmb{f}}_p^0$ :", + "bbox": [ + 75, + 518, + 468, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {F}} _ {p} = S _ {\\theta} ([ T (\\boldsymbol {t} _ {m}), \\hat {\\boldsymbol {f}} _ {p} ^ {0}, \\mathbf {F} _ {p} ^ {m} ]), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 628, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T(\\cdot)$ is the pretrained text feature extractor.", + "bbox": [ + 76, + 657, + 406, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To utilize rich embeddings stored in $\\mathcal{C}_p$ to constrain the prediction space, we retrieve the nearest embedding of the predicted continuous $\\bar{\\mathbf{F}}_p$ from $\\mathcal{C}_p$ to obtain the final $\\hat{\\mathbf{F}}_p$ :", + "bbox": [ + 75, + 672, + 468, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} _ {p} = \\operatorname {N e a r e s t} \\left(\\bar {\\mathbf {F}} _ {p}\\right) := \\left\\{\\underset {c _ {k} \\in \\mathcal {C} _ {p}} {\\operatorname {a r g m i n}} \\left\\| \\bar {f} _ {p} - c _ {k} \\right\\| _ {2} \\right\\}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 729, + 468, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\hat{\\mathbf{F}}_p$ is then fed into $D$ to reconstruct the final video $\\hat{\\mathbf{V}}$ following Eq. (3). Thanks to the continuous operation, we can add losses at both the image level and embedding level. The loss function to train our continuous VQ-diffuser is", + "bbox": [ + 75, + 772, + 468, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\left\\| \\bar {\\mathbf {F}} _ {p} - \\mathbf {F} _ {p} \\right\\| _ {1} + \\left\\| \\operatorname {s g} \\left(\\hat {\\mathbf {F}} _ {p}\\right) - \\bar {\\mathbf {F}} _ {p} \\right\\| _ {2} ^ {2} \\tag {8} \\\\ + \\left\\| \\mathrm {s g} (\\bar {\\mathbf {F}} _ {p}) - \\hat {\\mathbf {F}} _ {p} \\right\\| _ {2} ^ {2} + L _ {\\mathrm {r e c}} (\\hat {\\mathbf {V}}, \\mathbf {V}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 844, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{\\mathrm{rec}}$ is composed of $L_{1}$ loss and perceptual loss [28]. It should be noted that stop-gradient operation is applied to make the nearest operation differentiable.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared to continuous samplers like PixelCNN [51] in Fig. 5(b), which directly predict the continuous RGB values, our continuous VQ-diffuser fully utilizes the rich contents stored in the codebook while restricting the prediction space. Therefore, it inherits the benefits of the discrete VQ sampler and continuous sampler at the same time. The pose sequences sampled by our continuous VQ-diffuser are temporally coherent and meaningful.", + "bbox": [ + 496, + 137, + 893, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Motion-Aware Masking Strategy", + "text_level": 1, + "bbox": [ + 498, + 282, + 782, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To generate plausible videos, the sampled motion sequence $\\hat{\\mathbf{F}}_p$ should be both temporally and spatially reasonable. In order to make the continuous VQ-diffuser $S_{\\theta}$ correctly condition on $\\pmb{t}_{m}$ as well as generate reasonable human poses for each frame, we design a motion-aware masking strategy to meticulously mask $\\mathbf{F}_p$ temporally and spatially.", + "bbox": [ + 496, + 309, + 890, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the temporal dimension, $S_{\\theta}$ first predicts pose embeddings of the first frame and the last frame based on $t_m$ and $\\hat{\\pmb{f}}_p^0$ . Then the prediction diffuses to the intermediate frames according to the given conditions and previously unmasked frames. Therefore, during the training, if the first frame and the last frame are masked, we will mask all frames to prevent the intermediate frames from providing information to help frame predictions at two ends. A higher probability is assigned to masking all frames to help $S_{\\theta}$ better learn prediction at the most challenging two ends.", + "bbox": [ + 496, + 401, + 892, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the spatial dimension, we propose to predict the first and the last frames with more than one diffusion step. Otherwise, the structure of the sampled human tends to be incomplete. To this end, during the training, we mask the first and last frames partially at the spatial dimension.", + "bbox": [ + 496, + 553, + 890, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Fashion-Text2Video Dataset", + "text_level": 1, + "bbox": [ + 498, + 646, + 759, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing human video datasets are either of low resolution [3, 17, 49] or lack of text (action labels) [35, 43]. To support the study of text-driven human video generation, we propose a new dataset, Fashion-Text2Video dataset. Among existing human video datasets, we select videos in Fashion Dataset [56] as the source data for following reasons: 1) High resolution ( $\\geq 1024 \\times 1024$ ); 2) Diverse appearance.", + "bbox": [ + 496, + 672, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Fashion Dataset contains a total of 600 videos. For each video, we manually annotate motion labels, clothing shapes (length of sleeves and length of clothing) and textures. Motion labels are classified into 22 classes, including standing, moving right-side, moving left-side, turning around, etc. The lengths of sleeves are classified into no-sleeve, three-point, medium and long sleeves. The length of clothing is the length of the upper part of the clothes.", + "bbox": [ + 496, + 779, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22751", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/99167ea0ac00e7f42b045d427db1ab51b51e6955def74f448a20ec5efa2e4307.jpg", + "image_caption": [ + "D" + ], + "image_footnote": [], + "bbox": [ + 143, + 95, + 490, + 213 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0174ab8dd49f5ae04bd0c732244d3d00b335d51989ec143cbbb10a8df3e10b79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 94, + 848, + 213 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/037189fa832080d58a81d45de27ac184010dcf8059096744f9e8640b80af74e5.jpg", + "image_caption": [ + "eannnnnne" + ], + "image_footnote": [], + "bbox": [ + 143, + 231, + 490, + 351 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/464395c7197fc7442f2a6b9960c9d2620ad976ba1838c9a3d011b879d106368b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 231, + 846, + 351 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b427eb8988c41894caf93822b06a571db34ca69249cb041e011bd1f93f5b5eca.jpg", + "image_caption": [ + "Teeerree" + ], + "image_footnote": [ + "This female is wearing a dress, with graphic pattern. It has long sleeves and it is of three-point length. This female turns right from the side to the front." + ], + "bbox": [ + 143, + 368, + 490, + 486 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/28428872fb9b3e08cbf8fffd349a9456444ceb5e8bd5067a78a3943a41577bbe.jpg", + "image_caption": [], + "image_footnote": [ + "The dress this person wears has long-sleeve and it is of three-point length. The texture of it is graphic. The lady turns right from the back to the side." + ], + "bbox": [ + 493, + 368, + 846, + 486 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f88311fa6a7a84f8247165d3f5c3ebbcc261bedec3f9f3e6e891d53fae3f27f9.jpg", + "image_caption": [ + "Figure 6: Qualitative Comparisons. Text2Performer achieves superior generation qualities compared with baselines." + ], + "image_footnote": [], + "bbox": [ + 132, + 507, + 161, + 532 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/038ae4e547bb16aae03d74152db51a30401e081b77a6e3f7495c678fffc4d9eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 507, + 532, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Clothing textures are mainly divided into pure color, floral, graphic, etc. Upon the above labels, we then generate texts for each video using some pre-defined templates. The texts include descriptions of the appearances of the dressed clothes and descriptions for each sub-motions. The proposed Fashion-Text2Video dataset can be applied to the research on Text2Video and Label2Video generation.", + "bbox": [ + 75, + 595, + 468, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Comparison Methods", + "text_level": 1, + "bbox": [ + 76, + 739, + 279, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MMVID [20] is a VQVAE-based method for multimodal video synthesis. We follow the original setting to train MMVID on our dataset. StyleGAN-V [46] is a state-of-the-art video generation method built on StyleGAN. Action labels are used as conditions instead of texts, which we find significantly surpasses the performance of texts for this method. CogVideo [24] is a large-scale pretrained model for Text-to-Video generation. We implement two versions of CogVideo, i.e., CogVideo-v1 and CogVideo-v2. Since", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the original CogVideo utilized the pretrained Text2Image dataset, we train CogVideo-v1 to use the exemplar image as an additional input to compensate for the lack of large-scale pretrained Text2Image models for human images. CogVideo-v2 is with original designs.", + "bbox": [ + 496, + 595, + 893, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 681, + 684, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diversity and Quality. 1) We adopt Fréchet Inception Distance (FID) [21] to evaluate the quality of generated human frames. 2) We evaluate the quality of generated videos using Fréchet Video Distance (FVD) and Kernel Video Distance (KVD) [50]. We generate 2,048 video clips to compute the metrics, with 20 frames for each clip.", + "bbox": [ + 496, + 703, + 890, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Identity Preservation. We use two metrics to evaluate the maintenance of identity along the generated frames. 1) We extract face features of each generated frames using FaceNet [42]. For each generated video clip, we calculate the $l_{2}$ distances between features of the frames and the first frame. 2) We extract features of each generated frames using an off-the-shelf ReID model, i.e., OSNet [42]. These", + "bbox": [ + 496, + 795, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22752", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9ce8254c3a7eadc5fcdae678c1011e465d711adad0ac2d2efe72a2faab3b66a8.jpg", + "table_caption": [ + "Table 1: Quantitative Comparisons." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓FVD ↓KVD ↓Face ↓ReLU ↑
MMVID [20]11.85303.0278.670.90470.9096
StyleGAN-V [46]29.68219.6318.770.86750.9568
CogVideo-v1 [24]39.47645.0389.531.05640.8148
CogVideo-v2 [24]51.76799.80112.241.06210.7960
Text2Performer9.60124.7817.960.85930.9382
", + "bbox": [ + 78, + 106, + 480, + 184 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc611437ae5bf137664f8200d8583da06ab4cbe371b7d358e821a21c6a2f07e6.jpg", + "image_caption": [ + "Figure 7: User Study. We achieve the highest scores." + ], + "image_footnote": [], + "bbox": [ + 81, + 200, + 460, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "features are related to identity features. We calculate the average cosine distances between extracted features of the frames and the first frame. The final metrics are obtained by averaging distances among 2,048 generated video clips.", + "bbox": [ + 75, + 351, + 467, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. Users are asked to give three separate scores (the highest score is 4) at three dimensions: 1) The consistency with the text for appearance, 2) The consistency with the text for desired motion, and 3) The overall quality of the generated video in terms of temporal coherence and its realism. A total of 20 users participate in the user study. Each user is presented with 30 videos generated by different methods.", + "bbox": [ + 75, + 412, + 467, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 76, + 542, + 305, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show two examples in Fig. 6. In the first example, MMVID and Text2Performer successfully generate desired human motions. However, frames generated by MMVID have drifted identities. StyleGAN-V fails to generate plausible frames as well as consistent motions. As for the second example, only Text2Performer generates corresponding motions. Artifacts appear in the video clips generated by MMVID and StyleGAN-V. The visual comparisons demonstrate the superiority of Text2Performer.", + "bbox": [ + 75, + 565, + 468, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 76, + 710, + 316, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative comparisons are shown in Table 1. In terms of FID, FVD and KVD, our proposed Text2Performer has significant improvements over baselines, which demonstrates the diversity and temporal coherence of the videos generated by our methods. Existing baselines generate videos from entangled human representations and thus suffer from temporal incoherence. As for the identity metrics, our method achieves the best performance on face distance and comparable performance on ReID distance with StyleGAN-V. Since StyleGAN-V is prone to generate videos with small motions, features tend to be similar,", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4991f26c8fd6fae6c84895b0df4c4ccb01be5d1e75e35f61b4fdbfc0c75a7253.jpg", + "image_caption": [ + "Figure 8: Ablation Studies on (a) Decomposed VQ-Space, (b) Spatial Resolution of Pose Branch, (c) Discrete VQ sampler, and (d) Usage of Codebook." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 887, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "leading to high ReID scores. In our method, thanks to the decomposed VQ-space and continuous VQ-diffuser, temporal and identity consistency can be better achieved. The result of the user study is shown in Fig. 7. Text2Performer achieves the highest scores on all three dimensions. Specifically, the result of the assessment on overall quality is consistent with the other quantitative metrics, which further validates the effectiveness of Text2Performer.", + "bbox": [ + 496, + 381, + 890, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 513, + 653, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Decomposed VQ-Space. The ablation model trains the continuous VQ-diffuser with a unified VQ-space, which fuses the identity and pose information within one feature. As shown in Fig. 8(a), compared to our results, sampling in the unified VQ-Space generates drifted identity and incomplete body structures. This leads to inferior quantitative metrics in Table 2. The unified VQ-Space requires the sampler to handle identity and pose at the same time, which poses burdens for the sampler. With the decomposed VQ-Space, the identity can be easily maintained by fixing the appearance features, which further helps learn motions.", + "bbox": [ + 496, + 537, + 890, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spatial Resolution of Pose Branch. In our design, the pose feature has $4 \\times$ smaller spatial resolution than the appearance feature. In Fig. 8(b), we show the results generated by a decomposed VQVAE where pose and appearance features have the same resolution. Compared to results of the full model, body structures in the side view are incomplete. The quantitative metrics in Table 2 and Fig. 8(b) demonstrate that the smaller pose feature is necessary as the larger feature contains redundant information to harm valid feature disentanglement, thus adding challenges to the sampler.", + "bbox": [ + 496, + 704, + 890, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Discrete VQ sampler in Decomposed VQ-Space. We train a variant of VQ-Diffuser to predict discrete codebook indices with the decomposed VQ-Space. In Fig. 8(c), the", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22753", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2a0daad2d1921ccb971194ab1f31de5c234de4837d77777e9bbae5af3d9e3f62.jpg", + "image_caption": [ + "Figure 9: High-Resolution Results. We show one result generated by a long text sequence. The identity is well maintained." + ], + "image_footnote": [], + "bbox": [ + 117, + 84, + 851, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00b29ec0f0117bcf53c4c297361a9f5801c1bb434f26af13320908fb3f53b79e.jpg", + "image_caption": [ + "Figure 10: Cross-Dataset Performance. We compare with models trained on large-scale data and domain-specific data." + ], + "image_footnote": [], + "bbox": [ + 119, + 311, + 383, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/788fb8a095a8fb5514ee54804ee9f828d6f40acb660717fa7cdd60922b1df37d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 311, + 612, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3dd2f4f78edbfed8b4ebce517b971cfa2fdd37f206bd81970d3a3200a9e91328.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 311, + 848, + 446 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a41f40d7a3b33d028b696698e3954c264bb037803f463a67020f3f1f41383c60.jpg", + "table_caption": [ + "Table 2: Quantitative Results on Ablation Studies." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓FVD ↓KVD ↓Face ↓ReID ↑
w/o decomp.20.52266.9735.520.89360.9213
same res.15.69202.0632.960.86490.9399
disc. sampler10.37308.8873.170.91060.9161
w/o codebook11.31154.9322.800.94370.9109
Full Model10.92134.4320.310.85200.9421
", + "bbox": [ + 78, + 508, + 460, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "discrete VQ Sampler can generate complete body structures but fails to generate video clips consistent with the given action control, i.e., turning right from the back to the side. Failure on conditioning on action controls does not harm the quality of each generated frame, and thus this ablation study achieves a comparable FID score as our method. However, our method significantly outperforms this variant on other metrics as shown in Table 2. In VQVAE, some codebook entries have similar contents but with different codebook indices. Therefore, treating the training of the sampler as a classification task makes the model hard to converge.", + "bbox": [ + 75, + 613, + 468, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Usage of Codebook in Continuous VQ-Space. In our design, after the prediction of embeddings, the nearest neighbor of embeddings will be retrieved from the codebook. We train an auto-encoder without codebook in the pose branch. The sampler is trained to sample continuous embeddings from the latent space of auto-encoder without the retrieval from codebook. As shown in Fig. 8(d), without the codebook, the sampler results in a disrupted human image at", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the side view. The quality of generated videos deteriorates as reflected in Table 2. This demonstrates that predicting compact and meaningful continuous embeddings with codebooks enhances the quality of synthesized frames.", + "bbox": [ + 496, + 491, + 890, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Additional Analysis", + "text_level": 1, + "bbox": [ + 500, + 568, + 687, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "High-Resolution Results. Text2Performer can generate long videos with high resolutions (512 × 256) as shown in Fig. 1 and Fig. 9. Figure 9 shows an example of high-resolution videos for a long text sequence.", + "bbox": [ + 496, + 594, + 890, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-Dataset Performance. We further verify the performance of models trained on different datasets. The results are presented in Fig. 10. Compared with CogVideo [24] pretrained on large-scale dataset, our Text2Performer and MMVID succeed to generate plausible specific human content. As for models trained on small-scale domain-specific data, our Text2Performer is comparable to MMVID on the simple iPer dataset [35], and has superior performance on the more challenging Fashion-Text2Video dataset, especially on the completeness of body structures across frames.", + "bbox": [ + 496, + 656, + 890, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ability to Generate Novel Appearance. To verify that the generated humans are with novel appearances, we retrieve top-2 nearest neighbour images of generated frames using the perceptual distance [28]. As shown in Fig. 11, the generated query frames have different appearances from the top-2 nearest images retrieved from the dataset.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22754", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1da77b9d9e51b4dc40eebfafd041f852e91d73c50992015c987306cd6e814b42.jpg", + "image_caption": [ + "Query Image Top-2 Nearest Images Query Image Top-2 Nearest Images" + ], + "image_footnote": [], + "bbox": [ + 89, + 90, + 269, + 188 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dd84e38c790c40f4ae63428be22bba459e273829e5dc3fa5a05c2010967ba48c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 90, + 457, + 188 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Query Image Top-2 Nearest Images", + "bbox": [ + 277, + 189, + 460, + 200 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2f2efce4631626861f257dada972706ce0ef378b11a0fdceead142717aa7fb87.jpg", + "image_caption": [ + "Figure 11: Nearest Neighbour Image Search. Top-2 nearest images of generated images are retrieved.", + "Figure 12: Results on In-the-Wild Dataset. In our current pipeline, We treat the background as a part of appearance information. The results contain some artifacts." + ], + "image_footnote": [], + "bbox": [ + 81, + 263, + 467, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Discussions", + "text_level": 1, + "bbox": [ + 76, + 385, + 196, + 400 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We propose a novel Text2Performer for text-driven human video generation. We decompose the VQ-space into appearance and pose representations. Continuous VQ-diffuser is then introduced to sample motions conditioned on texts. These designs empower Text2Performer to synthesize high-quality and high-resolution human videos.", + "bbox": [ + 75, + 412, + 468, + 502 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations and Future Work. In this paper, we focus on the generation of human appearance and motion with clean backgrounds. The current pipeline is designed for generating human videos without textural backgrounds. It can be extended to generate plausible textual backgrounds by treating the background as a part of appearance information at the training time. The generated results are shown in Fig. 12. There are some artifacts in the generated videos. Generating videos with textual backgrounds is important in real-world applications. In future work, textual backgrounds can be segmented first. The results can be then improved by generating the foreground and textual background separately and fusing them harmoniously using an additional module. In addition, the synthesized human videos are biased toward generating females with dresses. This is because the Fashion Dataset only contains videos of females with dresses. In future applications, more data can be involved in the training to alleviate the bias. Besides, the control modality can also be extended to various types, such as audio and music.", + "bbox": [ + 75, + 503, + 468, + 805 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Potential Negative Societal Impacts. The model may be applied to generate fake videos performing various actions. To alleviate the negative impacts, DeepFake detection methods can be applied to evaluate the realism of videos.", + "bbox": [ + 75, + 808, + 468, + 868 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collabora", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "tion Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This study is also supported by NTU NAP, MOE AcRF Tier 1 (2021-T1-001-088), and the Google PhD Fellowship.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 167, + 596, + 183 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Badour Albahar, Jingwan Lu, Jimei Yang, Zhixin Shu, Eli Shechtman, and Jia-Bin Huang. Pose with style: Detail-preserving pose-guided image synthesis with conditional stylegan. ACM Transactions on Graphics (TOG), 40(6):1-11, 2021. 3", + "[2] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. NeurIPS, 34:17981-17993, 2021. 4", + "[3] Guha Balakrishnan, Amy Zhao, Adrian V Dalca, Fredo Durand, and John Guttag. Synthesizing images of humans in unseen poses. In CVPR, pages 8340-8348, 2018. 5", + "[4] Sam Bond-Taylor, Peter Hessey, Hiroshi Sasaki, Toby P Breckon, and Chris G Willcocks. Unleashing transformers: parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vector-quantized codes. In ECCV, pages 170–188. Springer, 2022. 2, 4, 5", + "[5] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ICLR, 2019. 2", + "[6] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. NeurIPS, 35:31769-31781, 2022. 3", + "[7] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In ICCV, pages 5933-5942, 2019. 3", + "[8] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, pages 11315-11325, 2022. 4, 5", + "[9] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. NeurlPS, 34:19822-19835, 2021. 1", + "[10] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. NeurIPS, 35:16890-16902, 2022. 1", + "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2, 4, 5", + "[12] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, pages 1-19. Springer, 2022. 3", + "[13] Aviv Gabbay and Yedid Hoshen. Demystifying inter-class disentanglement. *ICLR*, 2020. 4" + ], + "bbox": [ + 501, + 194, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22755", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, pages 89-106. Springer, 2022. 1", + "[15] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and timesensitive transformer. In ECCV, pages 102-118. Springer, 2022. 3", + "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2", + "[17] Lena Gorelick, Moshe Blank, Eli Shechtman, Michal Irani, and Ronen Basri. Actions as space-time shapes. IEEE transactions on pattern analysis and machine intelligence, 29(12):2247-2253, 2007. 5", + "[18] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, pages 10696-10706, 2022. 1, 2, 4, 5", + "[19] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 3", + "[20] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and Sergey Tulyakov. Show me what and tell me how: Video synthesis via multimodal conditioning. In CVPR, pages 3615-3625, 2022. 3, 6, 7", + "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 6", + "[22] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 3", + "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2", + "[24] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. *ICLR*, 2023. 1, 2, 3, 6, 7, 8", + "[25] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. NeurIPS, 31, 2018. 4", + "[26] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In ICML, pages 9902-9915. PMLR, 2022. 2", + "[27] Yuming Jiang, Shuai Yang, Haonan Qju, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3, 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016. 5, 8", + "[29] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410, 2019. 2, 3", + "[30] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR., pages 8110-8119, 2020. 3", + "[31] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. NeurIPS, 32, 2019. 2", + "[32] Kathleen M Lewis, Srivatsan Varadharajan, and Ira Kemelmacher-Shlizerman. Tryongan: Body-aware try-on via layered interpolation. ACM Transactions on Graphics (TOG), 40(4):1-10, 2021. 3", + "[33] Lingjie Liu, Weipeng Xu, Marc Habermann, Michael Zollhöfer, Florian Bernard, Hyeongwoo Kim, Wenping Wang, and Christian Theobalt. Neural human video rendering by learning dynamic textures and rendering-to-video translation. IEEE Transactions on Visualization and Computer Graphics, PP:1-1, 05 2020. 3", + "[34] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 38(5):1-14, 2019. 3", + "[35] Wen Liu, Zhixin Piao, Jie Min, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis. In ICCV, pages 5904-5913, 2019. 3, 5, 8", + "[36] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In ICCV, pages 10985-10995, 2021. 3", + "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. PMLR, 2021. 1", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2", + "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 1", + "[41] Kripasindhu Sarkar, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Humangan: A generative model of human images. In 2021 International Conference on 3D Vision (3DV), pages 258–267. IEEE, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22756", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823, 2015. 6", + "[43] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. NeurIPS, 32, 2019. 5", + "[44] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. ICLR, 2023. 1, 3", + "[45] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In CVPR, pages 11050-11059, 2022. 2", + "[46] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, pages 3626-3636, 2022. 3, 6, 7", + "[47] Jianxin Sun, Qiyao Deng, Qi Li, Muyi Sun, Min Ren, and Zhenan Sun. Anyface: Free-style text-to-face synthesis and manipulation. In CVPR, pages 18687-18696, 2022. 1", + "[48] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. ICLR, 2021. 2", + "[49] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, pages 1526-1535, 2018. 2, 5", + "[50] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6", + "[51] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. NeurIPS, 29, 2016. 5", + "[52] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 2, 4, 5", + "[53] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. ICLR, 2023. 1, 3", + "[54] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3", + "[55] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. *ICLR*, 2022. 3", + "[56] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose-guided human video generation. arXiv preprint arXiv:1910.09139, 2019. 2, 5" + ], + "bbox": [ + 78, + 90, + 467, + 842 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22757", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Text2Performer_ Text-Driven Human Video Generation/45c9b878-7941-4c19-aa7b-acd77ad00875_model.json b/2023/Text2Performer_ Text-Driven Human Video Generation/45c9b878-7941-4c19-aa7b-acd77ad00875_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b5ab7782414c4d30274602a6f9f756a07b284bc9 --- /dev/null +++ b/2023/Text2Performer_ Text-Driven Human Video Generation/45c9b878-7941-4c19-aa7b-acd77ad00875_model.json @@ -0,0 +1,2730 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.204, + 0.131, + 0.768, + 0.151 + ], + "angle": 0, + "content": "Text2 Performer: Text-Driven Human Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.18, + 0.903, + 0.218 + ], + "angle": 0, + "content": "Yuming Jiang\\(^{1}\\) Shuai Yang\\(^{1}\\) Tong Liang Koh\\(^{1}\\) Wayne Wu\\(^{2}\\) Chen Change Loy\\(^{1}\\) Ziwei Liu\\(^{1\\boxtimes}\\) \n\\(^{1}\\)S-Lab, Nanyang Technological University \\(^{2}\\)Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.22, + 0.88, + 0.235 + ], + "angle": 0, + "content": "{yuming002, shuai.yang, koht0029, ccloy, ziwei.liu}@ntu.edu.sg wuwenyan0503@gmail.com" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.272, + 0.501, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.424, + 0.149, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.422, + 0.476, + 0.461 + ], + "angle": 0, + "content": "This woman wears a dress. It has long sleeves, and it is of long length. The texture of it is graphic. She is swinging to the right." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.272, + 0.885, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.428, + 0.56, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.569, + 0.422, + 0.866, + 0.461 + ], + "angle": 0, + "content": "The person is wearing a sleeveless dress. It is of short length. Its pattern is pure color. She turns right from the front to the side." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.479, + 0.893, + 0.51 + ], + "angle": 0, + "content": "Figure 1: High-resolution videos generated by Text2Performer. The videos are generated by taking the text descriptions as the only input. The generated videos contain diverse appearances and flexible motions. Identities are well maintained." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.522, + 0.314, + 0.538 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.554, + 0.474, + 0.903 + ], + "angle": 0, + "content": "Text-driven content creation has evolved to be a transformative technique that revolutionizes creativity. Here we study the task of text-driven human video generation, where a video sequence is synthesized from texts describing the appearance and motions of a target performer. Compared to general text-driven video generation, human-centric video generation requires maintaining the appearance of synthesized human while performing complex motions. In this work, we present Text2Performer to generate vivid human videos with articulated motions from texts. Text2Performer has two novel designs: 1) decomposed human representation and 2) diffusion-based motion sampler. First, we decompose the VQVAE latent space into human appearance and pose representation in an unsupervised manner by utilizing the nature of human videos. In this way, the appearance is well maintained along the generated frames. Then, we propose continuous VQ-diffuser to sample a sequence of pose embeddings. Unlike existing VQ-based methods that operate in the discrete space, continuous VQ-diffuser directly outputs the continuous pose embeddings for better motion modeling. Finally, motion-aware masking strategy is designed to mask the pose embeddings spatial-temporally to enhance the temporal coherence. Moreover," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.895, + 0.646 + ], + "angle": 0, + "content": "to facilitate the task of text-driven human video generation, we contribute a Fashion-Text2Video dataset with manually annotated action labels and text descriptions. Extensive experiments demonstrate that Text2Performer generates high-quality human videos (up to \\(512 \\times 256\\) resolution) with diverse appearances and flexible motions. Our project page is https://yumingj.github.io/projects/Text2Performer.html" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.663, + 0.633, + 0.679 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Since its emergence, text-guided image synthesis (e.g. DALLE [37, 38]) has attracted substantial attention. Recent works [9, 10, 14, 18, 40, 47] have demonstrated fascinating performance for the quality of synthesized images and their consistency with the texts. Beyond image generation, text-driven video generation [22, 24, 44, 53] is an advanced topic to explore. Existing works rely on large-scale datasets to drive large models. Although they have achieved surprising performance on general objects, when applied to the generation of some specific tasks, such as generating videos of garment presents for e-commerce websites, they fail to generate plausible results. Take the CogVideo [24] as an example. As shown in Fig. 2, the generated human objects contain incomplete human structures, and the temporal" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22747" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.091, + 0.465, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.155, + 0.461, + 0.18 + ], + "angle": 0, + "content": "The person is wearing a sleeveless dress. It is of short length. Its pattern is pure color. She turns right from the front to the side." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.185, + 0.47, + 0.262 + ], + "angle": 0, + "content": "Figure 2: Results of General Large Text-to-Video Models. We use the pretrained general large Text-to-Video Models [24] to generate videos using the same texts as the right example of Fig. 1. The result fails to generate complete human structures and maintains temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.469, + 0.339 + ], + "angle": 0, + "content": "consistency is poorly maintained. On the other hand, these methods need billions of training data, which hampers the application to those specific tasks (e.g. human video generation) without a large amount of paired data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.469, + 0.566 + ], + "angle": 0, + "content": "Therefore, it is worthwhile to explore text-to-video generation in human video generation, which has numerous applications [31, 45]. In this paper, we focus on the task of text-driven human video generation. Compared to general text-to-video generation, text-driven human video generation poses several unique challenges: 1) The human structure is articulated. The joint movements of different body components form many complicated out-of-plane motions, e.g., rotations. 2) When performing complicated motions, the appearance of the synthesized human should remain the same. For example, the appearance of a target human after turning around should be consistent with that at the first beginning. In sum, to achieve high-fidelity human video generation, consistent human representation and complicated human motions should be carefully modeled." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.469, + 0.839 + ], + "angle": 0, + "content": "We propose a novel text-driven human video generation framework Text2Performer to handle consistent human representations and complex out-of-plane motions. As shown in Fig. 1, given texts describing appearance and motions, Text2Performer is able to generate temporally consistent human videos with complete human structures and unified human appearances. Text2Performer is built upon VQVAE-based frameworks [4, 11, 18, 52]. In Text2Performer, thanks to the specific nature of human videos which shares the same objects across the frames within one video, VQVAE latent space can be decomposed into appearance and pose representations. With the decomposed VQ-space, videos are generated by sampling appearance representation and a sequence of pose representations separately. This decomposition contributes to the maintenance of human identity. Besides, it makes the motion modeling more tractable, as the motion sampler does not need to take the appearance information into consideration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To model complicated human motions, a novel continuous VQ-diffuser is proposed to sample a sequence of meaningful pose representations. The architecture of the continuous VQ-diffuser is transformer. The key difference to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.378 + ], + "angle": 0, + "content": "the previous transformer-based samplers [4, 11, 52] is that the continuous VQ-diffuser directly predicts the continuous pose embeddings rather than their indices in the codebook. After predicting continuous pose embeddings, we also make use of the rich embeddings stored in the codebook by retrieving the nearest embeddings of the predicted embeddings from the codebook. Predicting continuous embeddings alleviates the one-to-many prediction issue in previous discrete methods and the use of codebook constrains the prediction space. With this design, more temporally coherent human motions and more complete structures of human frames are sampled. In addition, we borrow the idea of diffusion models [4, 18, 23, 39, 26] to progressively predict the long sequence of the pose embeddings. We propose a motion-aware masking strategy to sample the pose embeddings of the first frame and last frame firstly. Then the pose embeddings of the intermediate frames are gradually diffused. The motion-aware masking strategy enhances the completeness of human structures and temporal coherence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.484 + ], + "angle": 0, + "content": "To facilitate the task of text-guided human video generation, we propose the Fashion-Text2Video Dataset. It is built upon the FashionDataset [56], which consists of 600 human videos performing the fashion show. We manually segment the whole video into clips and label the motion types. Each clip is performing one motion. With the manually labeled motion labels, we then pair them with text descriptions." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.485, + 0.822, + 0.498 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.503, + 0.892, + 0.563 + ], + "angle": 0, + "content": "- We present and study the task of text-guided human video generation. Our proposed Text2Performer can be well trained and have generative abilities with only a small amount of data for training." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.565, + 0.892, + 0.669 + ], + "angle": 0, + "content": "- We propose to decompose the VQ-space into appearance and pose representations. The decomposition is achieved by making use of the nature of human videos, i.e., the motions are performed under one identity (appearance) across the frames. The decomposition of VQ-space improves the appearance coherence among frames and eases motion modeling." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.671, + 0.892, + 0.761 + ], + "angle": 0, + "content": "- We propose the continuous VQ-diffuser to predict continuous pose embeddings with the pose codebook. The final continuous pose embeddings are iteratively predicted with the guidance of the motion-aware masking scheme. These designs contribute to generated frames of high quality and temporal coherence." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.763, + 0.892, + 0.808 + ], + "angle": 0, + "content": "- We construct the Fashion-Text2Video Dataset with human motion labels and text descriptions to facilitate the research on text-driven human video generation." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.503, + 0.892, + 0.808 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.641, + 0.846 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Video Synthesis. Similar to image generation, GAN [5, 16, 29] and VQVAE [11, 52] are two common paradigms in the field of video synthesis [49]. MoCoGAN-HD [48] and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22748" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.092, + 0.376, + 0.106 + ], + "angle": 0, + "content": "(a) Sampling from the Decomposed VQ-Space" + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.112, + 0.7, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.2, + 0.4, + 0.214 + ], + "angle": 0, + "content": "(b) Motion Sampling with Continuous VQ-Diffuser" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.217, + 0.699, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.775, + 0.094, + 0.826, + 0.108 + ], + "angle": 0, + "content": "Legend" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.114, + 0.764, + 0.129 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.777, + 0.118, + 0.8, + 0.128 + ], + "angle": 0, + "content": "Text" + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.138, + 0.764, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.777, + 0.144, + 0.837, + 0.154 + ], + "angle": 0, + "content": "Text encoder" + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.165, + 0.764, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.777, + 0.169, + 0.862, + 0.18 + ], + "angle": 0, + "content": "VQ-VAE Decoder" + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.192, + 0.764, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.776, + 0.196, + 0.816, + 0.206 + ], + "angle": 0, + "content": "Sampler" + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.221, + 0.764, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.776, + 0.222, + 0.828, + 0.242 + ], + "angle": 0, + "content": "Text embedding" + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.251, + 0.764, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.776, + 0.257, + 0.831, + 0.277 + ], + "angle": 0, + "content": "Continuous embedding" + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.285, + 0.764, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.776, + 0.291, + 0.865, + 0.312 + ], + "angle": 0, + "content": "Masked continuous embedding" + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.321, + 0.764, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.776, + 0.327, + 0.824, + 0.336 + ], + "angle": 0, + "content": "Codebook" + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.349, + 0.764, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.776, + 0.353, + 0.866, + 0.363 + ], + "angle": 0, + "content": "Appearance-related" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.374, + 0.764, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.776, + 0.377, + 0.845, + 0.388 + ], + "angle": 0, + "content": "Motion-related" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.393, + 0.895, + 0.54 + ], + "angle": 0, + "content": "Figure 3: Overview of Text2Performer with (a) Sampling from the Decomposed VQ-Space and (b) Motion Sampling with Continuous VQ-Diffuser. Given a text, we first sample the target appearance features \\(\\hat{f}_a\\) and exemplar pose features \\(\\hat{f}_p^0\\) conditioned on the language features \\(e_a\\) extracted by a pretrained text model. The motion sequence \\(\\hat{\\mathbf{F}}_p\\) is then sampled by our proposed continuous VQ-Diffuser. The continuous VQ-Diffuser takes the extracted language features \\(e_m\\) and \\(\\hat{f}_p^0\\) as inputs. The prediction of continuous motion sequences \\(\\hat{\\mathbf{F}}_p\\) starts with fully masked pose features and predicts continuous pose embeddings \\(\\{\\bar{f}_p^1,\\dots,\\bar{f}_p^n\\}\\) with Non-Causal Transformer. The nearest neighbor pose embeddings \\(\\{\\hat{f}_p^1,\\dots,\\hat{f}_p^n\\}\\) are then retrieved from the pose codebook \\(\\mathcal{C}_p\\). Guided by a motion-aware masking strategy, the continuous VQ-Diffuser is iteratively applied until the whole motion sequence is unmasked. The final videos are generated by feeding the continuous pose features and appearance features into the decoder of VQVAE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.473, + 0.856 + ], + "angle": 0, + "content": "StyleGAN-V [46] harness the powerful StyleGAN [29, 30] to generate videos. DI-GAN [55] models an implicit neural representation to generate videos. Brooks et al. [6] propose to regularize the temporal consistency via a temporal discriminator. These works focus on unconditional video generation. As for the VQVAE-based methods, VideoGPT [54] firstly extends the idea of VQVAE and autoregressive transformer to video generation. The following works introduce time-agnostic VQGAN with time-sensitive transformer [15] and masked sequence modeling [20] to further improve the performance. These VQVAE-based video synthesis frameworks accept conditions appended to the beginning of the token sequences. Our proposed Text2Performer is developed on VQVAE. The differences lie in that the designs of VQ-space decomposition and continuous sampling. Recently, there are some concurrent works [22, 24, 44, 53] for text-to-video generation. These methods are designed for general objects, while our Text2Performer has the design to separate human appearance and pose representations, which is specific to human video generation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Human Content Generation and Manipulation. Existing works focus on human images, and they are mainly divided into two types: unconditional generation and conditional" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.895, + 0.797 + ], + "angle": 0, + "content": "generation. StyleGAN-Human [12] employs StyleGAN to synthesize high-fidelity human images. TryOnGAN [32] and HumanGAN [41] generate human images conditioned on given human poses. Text2Human [27] proposes to generate human images conditioned on texts and poses. We focus on human video generation. For human content manipulation, pose transfer [1, 7, 33, 34, 35] is a popular topic. Pose transfer deals with video data. The task is to transfer the poses from the reference videos to the source video. Chan et al. [7] take human poses as inputs and the desired video is generated by an image-to-image translation network. Our task differs in that we only take texts as inputs and exemplar images are not compulsory. Motion synthesis works [19, 36] generate human motions in form of 3D human representations (e.g., SMPL and skeleton space), while our method generates human motions at the image level." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.814, + 0.659, + 0.83 + ], + "angle": 0, + "content": "3. Text2 Performer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 3, our proposed Text2Performer synthesizes desired human videos by taking the texts as inputs \\((\\pmb{t}_a\\) for appearance and \\(t_m\\) for motions). To ensure consistent human representation, we propose to decompose the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22749" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.473, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.396 + ], + "angle": 0, + "content": "Figure 4: Pipeline of Decomposed VQ-Space. Two different frames \\( I_0 \\) and \\( I_k \\) of a video serve as the identity frame and the pose frame to provide appearance and pose features, respectively. We apply data augmentations to \\( I_k \\) to erase its appearance information to avoid information leakage. Two codebooks are built to store the pose features and appearance features. The quantized features are finally fed into the decoder to reconstruct the pose frame \\( I_k \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.473, + 0.562 + ], + "angle": 0, + "content": "VQ-space into appearance representation \\( \\pmb{f}_a \\) and pose representation \\( \\pmb{f}_p \\) as shown in Fig. 4. With the decomposed VQ-space, we sample the human appearance features \\( \\hat{\\pmb{f}}_a \\) and exemplar pose features \\( \\hat{\\pmb{f}}_p^0 \\) according to \\( \\pmb{t}_a \\). To model the motion dynamics, we propose continuous VQ-diffuser as the motion sampler. It progressively diffuses the masked sequence of pose features until the whole sequence is un-masked. The final video clip is generated by feeding the appearance and pose embeddings into the decoder." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.295, + 0.59 + ], + "angle": 0, + "content": "3.1. Decomposed VQ-Space" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Human video generation requires consistent human appearance, i.e., the face and clothing of the target person should remain unchanged while performing the motion. Vanilla VQVAE [52] encodes images into a unified feature representation and builds a codebook on it. However, such design is prone to generate drifted identities along the video frames. To overcome this problem, we propose to decompose the VQ-space into the appearance representation \\( \\pmb{f}_a \\) and pose representation \\( \\pmb{f}_p \\) in an unsupervised manner. With the decomposed VQ-space, we can separately sample \\( \\hat{\\pmb{f}}_a \\) and a sequence of \\( \\hat{\\pmb{f}}_p \\) to generate the desired video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In human video data, different frames in one video share the same human identity. We utilize this property to train the decomposed VQ-space [13, 25]. The pipeline is shown in Fig. 4. Given a video \\(\\mathbf{V} = \\{\\mathbf{I}_0, \\mathbf{I}_1, \\dots, \\mathbf{I}_n\\}\\), we use its first frame \\(\\mathbf{I}_0\\) for appearance information and another randomly sampled frame \\(\\mathbf{I}_k\\) for pose information. \\(\\mathbf{I}_0\\) and \\(\\mathbf{I}_k\\) are fed into two encoders \\(E_a\\) and \\(E_p\\), respectively. The two branches share the encoder \\(E_s\\). To prevent \\(\\mathbf{I}_k\\) from leaking appearance information, data augmentations (e.g. color jitter" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "tering and Gaussian blur) are applied to \\(\\mathbf{I}_k\\) before the pose branch. \\(f_{a}\\) and \\(f_{p}\\) are obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.137, + 0.892, + 0.154 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {a} = E _ {a} \\left(E _ {s} \\left(\\mathbf {I} _ {0}\\right)\\right), \\boldsymbol {f} _ {p} = E _ {p} \\left(E _ {s} \\left(\\mathbf {I} _ {k} ^ {a}\\right)\\right)), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.167, + 0.894, + 0.243 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}_k^a\\) is the augmented \\(\\mathbf{I}_k\\). To make \\(f_{p}\\) learn compact and necessary pose information, we make the spatial size of \\(f_{p}\\) smaller. Given \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\), we empirically find that the optimal spatial sizes for \\(f_{a}\\) and \\(f_{p}\\) are \\(H / 16 \\times W / 16\\) and \\(H / 64 \\times W / 64\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.894, + 0.291 + ], + "angle": 0, + "content": "After obtaining \\( \\pmb{f}_a \\) and \\( \\pmb{f}_p \\), two codebooks \\( \\mathcal{C}_a \\) and \\( \\mathcal{C}_p \\) are built to store the appearance and pose embeddings. The quantized feature \\( \\hat{\\pmb{f}} \\) given codebook \\( \\mathcal{C} \\) is obtained by" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.304, + 0.892, + 0.33 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {f}} = \\operatorname {Q u a n t} (\\boldsymbol {f}) := \\left\\{\\underset {c _ {k} \\in \\mathcal {C}} {\\operatorname {a r g m i n}} \\| f _ {i j} - c _ {k} \\| _ {2} \\mid f _ {i j} \\in \\boldsymbol {f} \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.347, + 0.892, + 0.38 + ], + "angle": 0, + "content": "With the quantized \\(\\hat{f}_a\\) and \\(\\hat{f}_p\\), we then feed them into decoder \\(D\\) to reconstruct the target pose frame \\(\\hat{\\mathbf{I}}_k\\):" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.393, + 0.892, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {I}} _ {k} = D ([ \\hat {\\boldsymbol {f}} _ {a}, D _ {p} (\\hat {\\boldsymbol {f}} _ {p}) ]), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.488 + ], + "angle": 0, + "content": "where \\([\\cdot ]\\) is the concatenation operation, and \\(D_{p}\\) upsamples \\(\\hat{\\pmb{f}}_p\\) to make it have the same resolution as \\(\\hat{\\pmb{f}}_a\\). The whole network (including the encoders, decoders, and codebooks) is trained using:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.501, + 0.892, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\left\\| \\mathbf {I} _ {k} - \\hat {\\mathbf {I}} _ {k} \\right\\| _ {1} + \\left\\| \\operatorname {s g} (\\hat {\\boldsymbol {f}} _ {a}) - \\boldsymbol {f} _ {a} \\right\\| _ {2} ^ {2} + \\left\\| \\operatorname {s g} (\\boldsymbol {f} _ {a}) - \\hat {\\boldsymbol {f}} _ {a} \\right\\| _ {2} ^ {2} \\\\ + \\left\\| \\operatorname {s g} \\left(\\hat {\\boldsymbol {f}} _ {p}\\right) - \\boldsymbol {f} _ {p} \\right\\| _ {2} ^ {2} + \\left\\| \\operatorname {s g} \\left(\\boldsymbol {f} _ {p}\\right) - \\hat {\\boldsymbol {f}} _ {p} \\right\\| _ {2} ^ {2}, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.592, + 0.779, + 0.607 + ], + "angle": 0, + "content": "where \\(\\mathrm{sg}(\\cdot)\\) is the stop-gradient operation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.654 + ], + "angle": 0, + "content": "With the decomposed VQ-space, an additional sampler is trained to sample \\(\\hat{\\pmb{f}}_a\\) and \\(\\hat{\\pmb{f}}_p^0\\). This sampler has the same design as previous methods [4, 27]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.667, + 0.726, + 0.683 + ], + "angle": 0, + "content": "3.2. Continuous VQ-Diffuser" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.892, + 0.815 + ], + "angle": 0, + "content": "We adopt the absorbing diffusion transformer [2, 4, 8, 18] to sample the motion, a sequence of pose embeddings \\(\\hat{\\mathbf{F}}_p = \\{\\hat{\\pmb{f}}_p^1,\\hat{\\pmb{f}}_p^2,\\dots,\\hat{\\pmb{f}}_p^n\\}\\) from the learned pose codebook \\(\\mathcal{C}_p\\) Different from autoregressive models [11, 52] that make predictions in a fixed order, absorbing diffusion transformer predicts multiple codebook indices in parallel. The prediction of codebook indices starts from \\(\\mathbf{F}_p^0\\) i.e., fully masked \\(\\mathbf{F}_p^m\\) . The prediction at time step \\(t\\) is represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.83, + 0.892, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} _ {p} ^ {t} \\sim q _ {\\theta} \\left(\\mathbf {F} _ {p} ^ {t} \\mid \\mathbf {F} _ {p} ^ {t - 1}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.86, + 0.89, + 0.874 + ], + "angle": 0, + "content": "where \\(\\theta\\) denotes the parameters of the transformer sampler." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.876, + 0.892, + 0.907 + ], + "angle": 0, + "content": "To model human motions, we propose 1) continuous space sampling, and 2) motion-aware masking strategy." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22750" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.092, + 0.46, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.473, + 0.349 + ], + "angle": 0, + "content": "Figure 5: Comparison with Image Samplers. (a) Discrete VQ Sampler predicts discrete codebook indices, which are trained by cross-entropy loss. (b) Continuous Sampler operates on unconstrained continuous RGB spaces. (c) Our VQ-Diffuser predicts continuous embeddings but also utilizes rich embeddings stored in codebooks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.375, + 0.33, + 0.391 + ], + "angle": 0, + "content": "3.2.1 Continuous Space Sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.47, + 0.519 + ], + "angle": 0, + "content": "As shown in Fig. 5(a), previous VQVAE-based methods [4, 8, 11, 18, 52] sample tokens in a discrete space. They predict codebook indices to form quantized features, which are then fed into the decoder to generate images. However, sampling at the discrete space is hard to converge because of redundant codebooks, which causes the one-to-many mapping prediction problem. This would lead to meaningless predicted pose sequences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.618 + ], + "angle": 0, + "content": "In continuous VQ-diffuser \\(S_{\\theta}\\), we propose to train the non-causal transformer in the continuous embedding space as shown in Fig. 5(c). \\(S_{\\theta}\\) directly predicts continuous pose embeddings \\(\\{\\pmb{f}_p^k\\}\\). \\(S_{\\theta}\\) is trained to predict the full continuous \\(\\bar{\\mathbf{F}}_p\\) from the masked pose sequences \\(\\mathbf{F}_p^m\\) conditioned on the motion description \\(\\pmb{t}_{m}\\) and the initial pose feature \\(\\hat{\\pmb{f}}_p^0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.629, + 0.47, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {F}} _ {p} = S _ {\\theta} ([ T (\\boldsymbol {t} _ {m}), \\hat {\\boldsymbol {f}} _ {p} ^ {0}, \\mathbf {F} _ {p} ^ {m} ]), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.659, + 0.408, + 0.674 + ], + "angle": 0, + "content": "where \\(T(\\cdot)\\) is the pretrained text feature extractor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.721 + ], + "angle": 0, + "content": "To utilize rich embeddings stored in \\(\\mathcal{C}_p\\) to constrain the prediction space, we retrieve the nearest embedding of the predicted continuous \\(\\bar{\\mathbf{F}}_p\\) from \\(\\mathcal{C}_p\\) to obtain the final \\(\\hat{\\mathbf{F}}_p\\):" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.731, + 0.47, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} _ {p} = \\operatorname {N e a r e s t} \\left(\\bar {\\mathbf {F}} _ {p}\\right) := \\left\\{\\underset {c _ {k} \\in \\mathcal {C} _ {p}} {\\operatorname {a r g m i n}} \\left\\| \\bar {f} _ {p} - c _ {k} \\right\\| _ {2} \\right\\}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.773, + 0.47, + 0.835 + ], + "angle": 0, + "content": "\\(\\hat{\\mathbf{F}}_p\\) is then fed into \\(D\\) to reconstruct the final video \\(\\hat{\\mathbf{V}}\\) following Eq. (3). Thanks to the continuous operation, we can add losses at both the image level and embedding level. The loss function to train our continuous VQ-diffuser is" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.845, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\left\\| \\bar {\\mathbf {F}} _ {p} - \\mathbf {F} _ {p} \\right\\| _ {1} + \\left\\| \\operatorname {s g} \\left(\\hat {\\mathbf {F}} _ {p}\\right) - \\bar {\\mathbf {F}} _ {p} \\right\\| _ {2} ^ {2} \\tag {8} \\\\ + \\left\\| \\mathrm {s g} (\\bar {\\mathbf {F}} _ {p}) - \\hat {\\mathbf {F}} _ {p} \\right\\| _ {2} ^ {2} + L _ {\\mathrm {r e c}} (\\hat {\\mathbf {V}}, \\mathbf {V}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "where \\( L_{\\mathrm{rec}} \\) is composed of \\( L_{1} \\) loss and perceptual loss [28]. It should be noted that stop-gradient operation is applied to make the nearest operation differentiable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.894, + 0.26 + ], + "angle": 0, + "content": "Compared to continuous samplers like PixelCNN [51] in Fig. 5(b), which directly predict the continuous RGB values, our continuous VQ-diffuser fully utilizes the rich contents stored in the codebook while restricting the prediction space. Therefore, it inherits the benefits of the discrete VQ sampler and continuous sampler at the same time. The pose sequences sampled by our continuous VQ-diffuser are temporally coherent and meaningful." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.284, + 0.783, + 0.3 + ], + "angle": 0, + "content": "3.2.2 Motion-Aware Masking Strategy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.401 + ], + "angle": 0, + "content": "To generate plausible videos, the sampled motion sequence \\(\\hat{\\mathbf{F}}_p\\) should be both temporally and spatially reasonable. In order to make the continuous VQ-diffuser \\(S_{\\theta}\\) correctly condition on \\(\\pmb{t}_{m}\\) as well as generate reasonable human poses for each frame, we design a motion-aware masking strategy to meticulously mask \\(\\mathbf{F}_p\\) temporally and spatially." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.893, + 0.553 + ], + "angle": 0, + "content": "At the temporal dimension, \\( S_{\\theta} \\) first predicts pose embeddings of the first frame and the last frame based on \\( t_m \\) and \\( \\hat{\\pmb{f}}_p^0 \\). Then the prediction diffuses to the intermediate frames according to the given conditions and previously unmasked frames. Therefore, during the training, if the first frame and the last frame are masked, we will mask all frames to prevent the intermediate frames from providing information to help frame predictions at two ends. A higher probability is assigned to masking all frames to help \\( S_{\\theta} \\) better learn prediction at the most challenging two ends." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.63 + ], + "angle": 0, + "content": "At the spatial dimension, we propose to predict the first and the last frames with more than one diffusion step. Otherwise, the structure of the sampled human tends to be incomplete. To this end, during the training, we mask the first and last frames partially at the spatial dimension." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.647, + 0.76, + 0.663 + ], + "angle": 0, + "content": "4. Fashion-Text2Video Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Existing human video datasets are either of low resolution [3, 17, 49] or lack of text (action labels) [35, 43]. To support the study of text-driven human video generation, we propose a new dataset, Fashion-Text2Video dataset. Among existing human video datasets, we select videos in Fashion Dataset [56] as the source data for following reasons: 1) High resolution (\\(\\geq 1024 \\times 1024\\)); 2) Diverse appearance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The Fashion Dataset contains a total of 600 videos. For each video, we manually annotate motion labels, clothing shapes (length of sleeves and length of clothing) and textures. Motion labels are classified into 22 classes, including standing, moving right-side, moving left-side, turning around, etc. The lengths of sleeves are classified into no-sleeve, three-point, medium and long sleeves. The length of clothing is the length of the upper part of the clothes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22751" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.141, + 0.14, + 0.179 + ], + "angle": 0, + "content": "D" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.096, + 0.491, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.095, + 0.849, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.269, + 0.141, + 0.325 + ], + "angle": 0, + "content": "eannnnnne" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.232, + 0.491, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.232, + 0.848, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.398, + 0.141, + 0.466 + ], + "angle": 0, + "content": "Teeerree" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.369, + 0.491, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.369, + 0.848, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.508, + 0.163, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.171, + 0.504, + 0.476, + 0.542 + ], + "angle": 0, + "content": "This female is wearing a dress, with graphic pattern. It has long sleeves and it is of three-point length. This female turns right from the side to the front." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.508, + 0.534, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.542, + 0.505, + 0.835, + 0.543 + ], + "angle": 0, + "content": "The dress this person wears has long-sleeve and it is of three-point length. The texture of it is graphic. The lady turns right from the back to the side." + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.555, + 0.871, + 0.571 + ], + "angle": 0, + "content": "Figure 6: Qualitative Comparisons. Text2Performer achieves superior generation qualities compared with baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Clothing textures are mainly divided into pure color, floral, graphic, etc. Upon the above labels, we then generate texts for each video using some pre-defined templates. The texts include descriptions of the appearances of the dressed clothes and descriptions for each sub-motions. The proposed Fashion-Text2Video dataset can be applied to the research on Text2Video and Label2Video generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.717, + 0.21, + 0.734 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.281, + 0.758 + ], + "angle": 0, + "content": "5.1. Comparison Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "MMVID [20] is a VQVAE-based method for multimodal video synthesis. We follow the original setting to train MMVID on our dataset. StyleGAN-V [46] is a state-of-the-art video generation method built on StyleGAN. Action labels are used as conditions instead of texts, which we find significantly surpasses the performance of texts for this method. CogVideo [24] is a large-scale pretrained model for Text-to-Video generation. We implement two versions of CogVideo, i.e., CogVideo-v1 and CogVideo-v2. Since" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.895, + 0.675 + ], + "angle": 0, + "content": "the original CogVideo utilized the pretrained Text2Image dataset, we train CogVideo-v1 to use the exemplar image as an additional input to compensate for the lack of large-scale pretrained Text2Image models for human images. CogVideo-v2 is with original designs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.685, + 0.697 + ], + "angle": 0, + "content": "5.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.796 + ], + "angle": 0, + "content": "Diversity and Quality. 1) We adopt Fréchet Inception Distance (FID) [21] to evaluate the quality of generated human frames. 2) We evaluate the quality of generated videos using Fréchet Video Distance (FVD) and Kernel Video Distance (KVD) [50]. We generate 2,048 video clips to compute the metrics, with 20 frames for each clip." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Identity Preservation. We use two metrics to evaluate the maintenance of identity along the generated frames. 1) We extract face features of each generated frames using FaceNet [42]. For each generated video clip, we calculate the \\( l_{2} \\) distances between features of the frames and the first frame. 2) We extract features of each generated frames using an off-the-shelf ReID model, i.e., OSNet [42]. These" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22752" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.148, + 0.09, + 0.398, + 0.104 + ], + "angle": 0, + "content": "Table 1: Quantitative Comparisons." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.107, + 0.482, + 0.185 + ], + "angle": 0, + "content": "
MethodFID ↓FVD ↓KVD ↓Face ↓ReLU ↑
MMVID [20]11.85303.0278.670.90470.9096
StyleGAN-V [46]29.68219.6318.770.86750.9568
CogVideo-v1 [24]39.47645.0389.531.05640.8148
CogVideo-v2 [24]51.76799.80112.241.06210.7960
Text2Performer9.60124.7817.960.85930.9382
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.201, + 0.462, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.311, + 0.448, + 0.326 + ], + "angle": 0, + "content": "Figure 7: User Study. We achieve the highest scores." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.468, + 0.412 + ], + "angle": 0, + "content": "features are related to identity features. We calculate the average cosine distances between extracted features of the frames and the first frame. The final metrics are obtained by averaging distances among 2,048 generated video clips." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.468, + 0.533 + ], + "angle": 0, + "content": "User Study. Users are asked to give three separate scores (the highest score is 4) at three dimensions: 1) The consistency with the text for appearance, 2) The consistency with the text for desired motion, and 3) The overall quality of the generated video in terms of temporal coherence and its realism. A total of 20 users participate in the user study. Each user is presented with 30 videos generated by different methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.543, + 0.306, + 0.56 + ], + "angle": 0, + "content": "5.3. Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.469, + 0.703 + ], + "angle": 0, + "content": "We show two examples in Fig. 6. In the first example, MMVID and Text2Performer successfully generate desired human motions. However, frames generated by MMVID have drifted identities. StyleGAN-V fails to generate plausible frames as well as consistent motions. As for the second example, only Text2Performer generates corresponding motions. Artifacts appear in the video clips generated by MMVID and StyleGAN-V. The visual comparisons demonstrate the superiority of Text2Performer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.317, + 0.728 + ], + "angle": 0, + "content": "5.4. Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The quantitative comparisons are shown in Table 1. In terms of FID, FVD and KVD, our proposed Text2Performer has significant improvements over baselines, which demonstrates the diversity and temporal coherence of the videos generated by our methods. Existing baselines generate videos from entangled human representations and thus suffer from temporal incoherence. As for the identity metrics, our method achieves the best performance on face distance and comparable performance on ReID distance with StyleGAN-V. Since StyleGAN-V is prone to generate videos with small motions, features tend to be similar," + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.888, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.309, + 0.892, + 0.355 + ], + "angle": 0, + "content": "Figure 8: Ablation Studies on (a) Decomposed VQ-Space, (b) Spatial Resolution of Pose Branch, (c) Discrete VQ sampler, and (d) Usage of Codebook." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.892, + 0.503 + ], + "angle": 0, + "content": "leading to high ReID scores. In our method, thanks to the decomposed VQ-space and continuous VQ-diffuser, temporal and identity consistency can be better achieved. The result of the user study is shown in Fig. 7. Text2Performer achieves the highest scores on all three dimensions. Specifically, the result of the assessment on overall quality is consistent with the other quantitative metrics, which further validates the effectiveness of Text2Performer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.514, + 0.654, + 0.53 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Decomposed VQ-Space. The ablation model trains the continuous VQ-diffuser with a unified VQ-space, which fuses the identity and pose information within one feature. As shown in Fig. 8(a), compared to our results, sampling in the unified VQ-Space generates drifted identity and incomplete body structures. This leads to inferior quantitative metrics in Table 2. The unified VQ-Space requires the sampler to handle identity and pose at the same time, which poses burdens for the sampler. With the decomposed VQ-Space, the identity can be easily maintained by fixing the appearance features, which further helps learn motions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Spatial Resolution of Pose Branch. In our design, the pose feature has \\(4 \\times\\) smaller spatial resolution than the appearance feature. In Fig. 8(b), we show the results generated by a decomposed VQVAE where pose and appearance features have the same resolution. Compared to results of the full model, body structures in the side view are incomplete. The quantitative metrics in Table 2 and Fig. 8(b) demonstrate that the smaller pose feature is necessary as the larger feature contains redundant information to harm valid feature disentanglement, thus adding challenges to the sampler." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Discrete VQ sampler in Decomposed VQ-Space. We train a variant of VQ-Diffuser to predict discrete codebook indices with the decomposed VQ-Space. In Fig. 8(c), the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22753" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.085, + 0.852, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.283, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Figure 9: High-Resolution Results. We show one result generated by a long text sequence. The identity is well maintained." + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.312, + 0.385, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.312, + 0.614, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.312, + 0.849, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.45, + 0.879, + 0.466 + ], + "angle": 0, + "content": "Figure 10: Cross-Dataset Performance. We compare with models trained on large-scale data and domain-specific data." + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.49, + 0.449, + 0.505 + ], + "angle": 0, + "content": "Table 2: Quantitative Results on Ablation Studies." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.51, + 0.462, + 0.589 + ], + "angle": 0, + "content": "
MethodFID ↓FVD ↓KVD ↓Face ↓ReID ↑
w/o decomp.20.52266.9735.520.89360.9213
same res.15.69202.0632.960.86490.9399
disc. sampler10.37308.8873.170.91060.9161
w/o codebook11.31154.9322.800.94370.9109
Full Model10.92134.4320.310.85200.9421
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.78 + ], + "angle": 0, + "content": "discrete VQ Sampler can generate complete body structures but fails to generate video clips consistent with the given action control, i.e., turning right from the back to the side. Failure on conditioning on action controls does not harm the quality of each generated frame, and thus this ablation study achieves a comparable FID score as our method. However, our method significantly outperforms this variant on other metrics as shown in Table 2. In VQVAE, some codebook entries have similar contents but with different codebook indices. Therefore, treating the training of the sampler as a classification task makes the model hard to converge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Usage of Codebook in Continuous VQ-Space. In our design, after the prediction of embeddings, the nearest neighbor of embeddings will be retrieved from the codebook. We train an auto-encoder without codebook in the pose branch. The sampler is trained to sample continuous embeddings from the latent space of auto-encoder without the retrieval from codebook. As shown in Fig. 8(d), without the codebook, the sampler results in a disrupted human image at" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.553 + ], + "angle": 0, + "content": "the side view. The quality of generated videos deteriorates as reflected in Table 2. This demonstrates that predicting compact and meaningful continuous embeddings with codebooks enhances the quality of synthesized frames." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.569, + 0.688, + 0.585 + ], + "angle": 0, + "content": "5.6. Additional Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.656 + ], + "angle": 0, + "content": "High-Resolution Results. Text2Performer can generate long videos with high resolutions (512 × 256) as shown in Fig. 1 and Fig. 9. Figure 9 shows an example of high-resolution videos for a long text sequence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Cross-Dataset Performance. We further verify the performance of models trained on different datasets. The results are presented in Fig. 10. Compared with CogVideo [24] pretrained on large-scale dataset, our Text2Performer and MMVID succeed to generate plausible specific human content. As for models trained on small-scale domain-specific data, our Text2Performer is comparable to MMVID on the simple iPer dataset [35], and has superior performance on the more challenging Fashion-Text2Video dataset, especially on the completeness of body structures across frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ability to Generate Novel Appearance. To verify that the generated humans are with novel appearances, we retrieve top-2 nearest neighbour images of generated frames using the perceptual distance [28]. As shown in Fig. 11, the generated query frames have different appearances from the top-2 nearest images retrieved from the dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22754" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.091, + 0.27, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.19, + 0.461, + 0.201 + ], + "angle": 0, + "content": "Query Image Top-2 Nearest Images Query Image Top-2 Nearest Images" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.091, + 0.458, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.19, + 0.461, + 0.201 + ], + "angle": 0, + "content": "Query Image Top-2 Nearest Images" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.214, + 0.47, + 0.245 + ], + "angle": 0, + "content": "Figure 11: Nearest Neighbour Image Search. Top-2 nearest images of generated images are retrieved." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.264, + 0.468, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.311, + 0.47, + 0.356 + ], + "angle": 0, + "content": "Figure 12: Results on In-the-Wild Dataset. In our current pipeline, We treat the background as a part of appearance information. The results contain some artifacts." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.386, + 0.197, + 0.401 + ], + "angle": 0, + "content": "6. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.469, + 0.503 + ], + "angle": 0, + "content": "We propose a novel Text2Performer for text-driven human video generation. We decompose the VQ-space into appearance and pose representations. Continuous VQ-diffuser is then introduced to sample motions conditioned on texts. These designs empower Text2Performer to synthesize high-quality and high-resolution human videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.806 + ], + "angle": 0, + "content": "Limitations and Future Work. In this paper, we focus on the generation of human appearance and motion with clean backgrounds. The current pipeline is designed for generating human videos without textural backgrounds. It can be extended to generate plausible textual backgrounds by treating the background as a part of appearance information at the training time. The generated results are shown in Fig. 12. There are some artifacts in the generated videos. Generating videos with textual backgrounds is important in real-world applications. In future work, textual backgrounds can be segmented first. The results can be then improved by generating the foreground and textual background separately and fusing them harmoniously using an additional module. In addition, the synthesized human videos are biased toward generating females with dresses. This is because the Fashion Dataset only contains videos of females with dresses. In future applications, more data can be involved in the training to alleviate the bias. Besides, the control modality can also be extended to various types, such as audio and music." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.809, + 0.469, + 0.869 + ], + "angle": 0, + "content": "Potential Negative Societal Impacts. The model may be applied to generate fake videos performing various actions. To alleviate the negative impacts, DeepFake detection methods can be applied to evaluate the realism of videos." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collabora" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "tion Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This study is also supported by NTU NAP, MOE AcRF Tier 1 (2021-T1-001-088), and the Google PhD Fellowship." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.169, + 0.597, + 0.184 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.195, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[1] Badour Albahar, Jingwan Lu, Jimei Yang, Zhixin Shu, Eli Shechtman, and Jia-Bin Huang. Pose with style: Detail-preserving pose-guided image synthesis with conditional stylegan. ACM Transactions on Graphics (TOG), 40(6):1-11, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.266, + 0.894, + 0.32 + ], + "angle": 0, + "content": "[2] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. NeurIPS, 34:17981-17993, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.324, + 0.892, + 0.365 + ], + "angle": 0, + "content": "[3] Guha Balakrishnan, Amy Zhao, Adrian V Dalca, Fredo Durand, and John Guttag. Synthesizing images of humans in unseen poses. In CVPR, pages 8340-8348, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.368, + 0.892, + 0.45 + ], + "angle": 0, + "content": "[4] Sam Bond-Taylor, Peter Hessey, Hiroshi Sasaki, Toby P Breckon, and Chris G Willcocks. Unleashing transformers: parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vector-quantized codes. In ECCV, pages 170–188. Springer, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.453, + 0.892, + 0.494 + ], + "angle": 0, + "content": "[5] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.497, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[6] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. NeurIPS, 35:31769-31781, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.555, + 0.892, + 0.595 + ], + "angle": 0, + "content": "[7] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In ICCV, pages 5933-5942, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.599, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[8] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, pages 11315-11325, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.642, + 0.892, + 0.709 + ], + "angle": 0, + "content": "[9] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. NeurlPS, 34:19822-19835, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.713, + 0.892, + 0.768 + ], + "angle": 0, + "content": "[10] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. NeurIPS, 35:16890-16902, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.771, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.816, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[12] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, pages 1-19. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[13] Aviv Gabbay and Yedid Hoshen. Demystifying inter-class disentanglement. *ICLR*, 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.195, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22755" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[14] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, pages 89-106. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[15] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and timesensitive transformer. In ECCV, pages 102-118. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.47, + 0.277 + ], + "angle": 0, + "content": "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.28, + 0.469, + 0.335 + ], + "angle": 0, + "content": "[17] Lena Gorelick, Moshe Blank, Eli Shechtman, Michal Irani, and Ronen Basri. Actions as space-time shapes. IEEE transactions on pattern analysis and machine intelligence, 29(12):2247-2253, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.338, + 0.469, + 0.393 + ], + "angle": 0, + "content": "[18] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, pages 10696-10706, 2022. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.396, + 0.469, + 0.465 + ], + "angle": 0, + "content": "[19] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.468, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[20] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and Sergey Tulyakov. Show me what and tell me how: Video synthesis via multimodal conditioning. In CVPR, pages 3615-3625, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.597, + 0.469, + 0.666 + ], + "angle": 0, + "content": "[22] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.669, + 0.469, + 0.709 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.713, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[24] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. *ICLR*, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[25] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. NeurIPS, 31, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[26] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In ICML, pages 9902-9915. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[27] Yuming Jiang, Shuai Yang, Haonan Qju, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[28] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[29] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[30] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR., pages 8110-8119, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.238, + 0.892, + 0.28 + ], + "angle": 0, + "content": "[31] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. NeurIPS, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.283, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[32] Kathleen M Lewis, Srivatsan Varadharajan, and Ira Kemelmacher-Shlizerman. Tryongan: Body-aware try-on via layered interpolation. ACM Transactions on Graphics (TOG), 40(4):1-10, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.34, + 0.892, + 0.423 + ], + "angle": 0, + "content": "[33] Lingjie Liu, Weipeng Xu, Marc Habermann, Michael Zollhöfer, Florian Bernard, Hyeongwoo Kim, Wenping Wang, and Christian Theobalt. Neural human video rendering by learning dynamic textures and rendering-to-video translation. IEEE Transactions on Visualization and Computer Graphics, PP:1-1, 05 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[34] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 38(5):1-14, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[35] Wen Liu, Zhixin Piao, Jie Min, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis. In ICCV, pages 5904-5913, 2019. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.556, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[36] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In ICCV, pages 10985-10995, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.892, + 0.654 + ], + "angle": 0, + "content": "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.716, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[41] Kripasindhu Sarkar, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Humangan: A generative model of human images. In 2021 International Conference on 3D Vision (3DV), pages 258–267. IEEE, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22756" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[42] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[43] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. NeurIPS, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[44] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. ICLR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.29 + ], + "angle": 0, + "content": "[45] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In CVPR, pages 11050-11059, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[46] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, pages 3626-3636, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.348, + 0.468, + 0.389 + ], + "angle": 0, + "content": "[47] Jianxin Sun, Qiyao Deng, Qi Li, Muyi Sun, Min Ren, and Zhenan Sun. Anyface: Free-style text-to-face synthesis and manipulation. In CVPR, pages 18687-18696, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.468, + 0.445 + ], + "angle": 0, + "content": "[48] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.448, + 0.468, + 0.489 + ], + "angle": 0, + "content": "[49] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, pages 1526-1535, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.491, + 0.468, + 0.545 + ], + "angle": 0, + "content": "[50] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.548, + 0.468, + 0.588 + ], + "angle": 0, + "content": "[51] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. NeurIPS, 29, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.59, + 0.468, + 0.617 + ], + "angle": 0, + "content": "[52] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.619, + 0.468, + 0.687 + ], + "angle": 0, + "content": "[53] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. ICLR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.689, + 0.468, + 0.73 + ], + "angle": 0, + "content": "[54] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.733, + 0.468, + 0.786 + ], + "angle": 0, + "content": "[55] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. *ICLR*, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[56] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose-guided human video generation. arXiv preprint arXiv:1910.09139, 2019. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.843 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22757" + } + ] +] \ No newline at end of file diff --git a/2023/Text2Performer_ Text-Driven Human Video Generation/45c9b878-7941-4c19-aa7b-acd77ad00875_origin.pdf b/2023/Text2Performer_ Text-Driven Human Video Generation/45c9b878-7941-4c19-aa7b-acd77ad00875_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..317e5e41ac2618ae32b405546e30b1df898ee846 --- /dev/null +++ b/2023/Text2Performer_ Text-Driven Human Video Generation/45c9b878-7941-4c19-aa7b-acd77ad00875_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbcf24f90b6622f58c34b04893803f43299fca9f07bae0fc75ec103199a4e478 +size 2153993 diff --git a/2023/Text2Performer_ Text-Driven Human Video Generation/full.md b/2023/Text2Performer_ Text-Driven Human Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d992b9cf775b63125db0f0a3c0d2377ba80a564a --- /dev/null +++ b/2023/Text2Performer_ Text-Driven Human Video Generation/full.md @@ -0,0 +1,395 @@ +# Text2 Performer: Text-Driven Human Video Generation + +Yuming Jiang $^{1}$ Shuai Yang $^{1}$ Tong Liang Koh $^{1}$ Wayne Wu $^{2}$ Chen Change Loy $^{1}$ Ziwei Liu $^{1\boxtimes}$ $^{1}$ S-Lab, Nanyang Technological University $^{2}$ Shanghai AI Laboratory + +{yuming002, shuai.yang, koht0029, ccloy, ziwei.liu}@ntu.edu.sg wuwenyan0503@gmail.com + +![](images/b424bc7500ed5243fbb519ce0bb57e99df79987f7d9da9fb46f01047fc84abdb.jpg) +Figure 1: High-resolution videos generated by Text2Performer. The videos are generated by taking the text descriptions as the only input. The generated videos contain diverse appearances and flexible motions. Identities are well maintained. + +![](images/aefea0d56de1b949c95bb51e5647b73136b96d1b309a1cfae66a865a1cb7a6ec.jpg) + +This woman wears a dress. It has long sleeves, and it is of long length. The texture of it is graphic. She is swinging to the right. + +![](images/53d11366d4394c29962c13e7b70dd849945205b7218de02f41d9fafade57343a.jpg) + +![](images/994493edf4f6027469fc298979094b0160332d4e85434250aa08a58bbbb0f596.jpg) + +The person is wearing a sleeveless dress. It is of short length. Its pattern is pure color. She turns right from the front to the side. + +# Abstract + +Text-driven content creation has evolved to be a transformative technique that revolutionizes creativity. Here we study the task of text-driven human video generation, where a video sequence is synthesized from texts describing the appearance and motions of a target performer. Compared to general text-driven video generation, human-centric video generation requires maintaining the appearance of synthesized human while performing complex motions. In this work, we present Text2Performer to generate vivid human videos with articulated motions from texts. Text2Performer has two novel designs: 1) decomposed human representation and 2) diffusion-based motion sampler. First, we decompose the VQVAE latent space into human appearance and pose representation in an unsupervised manner by utilizing the nature of human videos. In this way, the appearance is well maintained along the generated frames. Then, we propose continuous VQ-diffuser to sample a sequence of pose embeddings. Unlike existing VQ-based methods that operate in the discrete space, continuous VQ-diffuser directly outputs the continuous pose embeddings for better motion modeling. Finally, motion-aware masking strategy is designed to mask the pose embeddings spatial-temporally to enhance the temporal coherence. Moreover, + +to facilitate the task of text-driven human video generation, we contribute a Fashion-Text2Video dataset with manually annotated action labels and text descriptions. Extensive experiments demonstrate that Text2Performer generates high-quality human videos (up to $512 \times 256$ resolution) with diverse appearances and flexible motions. Our project page is https://yumingj.github.io/projects/Text2Performer.html + +# 1. Introduction + +Since its emergence, text-guided image synthesis (e.g. DALLE [37, 38]) has attracted substantial attention. Recent works [9, 10, 14, 18, 40, 47] have demonstrated fascinating performance for the quality of synthesized images and their consistency with the texts. Beyond image generation, text-driven video generation [22, 24, 44, 53] is an advanced topic to explore. Existing works rely on large-scale datasets to drive large models. Although they have achieved surprising performance on general objects, when applied to the generation of some specific tasks, such as generating videos of garment presents for e-commerce websites, they fail to generate plausible results. Take the CogVideo [24] as an example. As shown in Fig. 2, the generated human objects contain incomplete human structures, and the temporal + +![](images/3d5a6148d0c6be5ae87c6f5da2697825f347aa3b6a8b31c19cc0f87c3e3d7b7f.jpg) +The person is wearing a sleeveless dress. It is of short length. Its pattern is pure color. She turns right from the front to the side. +Figure 2: Results of General Large Text-to-Video Models. We use the pretrained general large Text-to-Video Models [24] to generate videos using the same texts as the right example of Fig. 1. The result fails to generate complete human structures and maintains temporal consistency. + +consistency is poorly maintained. On the other hand, these methods need billions of training data, which hampers the application to those specific tasks (e.g. human video generation) without a large amount of paired data. + +Therefore, it is worthwhile to explore text-to-video generation in human video generation, which has numerous applications [31, 45]. In this paper, we focus on the task of text-driven human video generation. Compared to general text-to-video generation, text-driven human video generation poses several unique challenges: 1) The human structure is articulated. The joint movements of different body components form many complicated out-of-plane motions, e.g., rotations. 2) When performing complicated motions, the appearance of the synthesized human should remain the same. For example, the appearance of a target human after turning around should be consistent with that at the first beginning. In sum, to achieve high-fidelity human video generation, consistent human representation and complicated human motions should be carefully modeled. + +We propose a novel text-driven human video generation framework Text2Performer to handle consistent human representations and complex out-of-plane motions. As shown in Fig. 1, given texts describing appearance and motions, Text2Performer is able to generate temporally consistent human videos with complete human structures and unified human appearances. Text2Performer is built upon VQVAE-based frameworks [4, 11, 18, 52]. In Text2Performer, thanks to the specific nature of human videos which shares the same objects across the frames within one video, VQVAE latent space can be decomposed into appearance and pose representations. With the decomposed VQ-space, videos are generated by sampling appearance representation and a sequence of pose representations separately. This decomposition contributes to the maintenance of human identity. Besides, it makes the motion modeling more tractable, as the motion sampler does not need to take the appearance information into consideration. + +To model complicated human motions, a novel continuous VQ-diffuser is proposed to sample a sequence of meaningful pose representations. The architecture of the continuous VQ-diffuser is transformer. The key difference to + +the previous transformer-based samplers [4, 11, 52] is that the continuous VQ-diffuser directly predicts the continuous pose embeddings rather than their indices in the codebook. After predicting continuous pose embeddings, we also make use of the rich embeddings stored in the codebook by retrieving the nearest embeddings of the predicted embeddings from the codebook. Predicting continuous embeddings alleviates the one-to-many prediction issue in previous discrete methods and the use of codebook constrains the prediction space. With this design, more temporally coherent human motions and more complete structures of human frames are sampled. In addition, we borrow the idea of diffusion models [4, 18, 23, 39, 26] to progressively predict the long sequence of the pose embeddings. We propose a motion-aware masking strategy to sample the pose embeddings of the first frame and last frame firstly. Then the pose embeddings of the intermediate frames are gradually diffused. The motion-aware masking strategy enhances the completeness of human structures and temporal coherence. + +To facilitate the task of text-guided human video generation, we propose the Fashion-Text2Video Dataset. It is built upon the FashionDataset [56], which consists of 600 human videos performing the fashion show. We manually segment the whole video into clips and label the motion types. Each clip is performing one motion. With the manually labeled motion labels, we then pair them with text descriptions. + +Our contributions are summarized as follows: + +- We present and study the task of text-guided human video generation. Our proposed Text2Performer can be well trained and have generative abilities with only a small amount of data for training. +- We propose to decompose the VQ-space into appearance and pose representations. The decomposition is achieved by making use of the nature of human videos, i.e., the motions are performed under one identity (appearance) across the frames. The decomposition of VQ-space improves the appearance coherence among frames and eases motion modeling. +- We propose the continuous VQ-diffuser to predict continuous pose embeddings with the pose codebook. The final continuous pose embeddings are iteratively predicted with the guidance of the motion-aware masking scheme. These designs contribute to generated frames of high quality and temporal coherence. +- We construct the Fashion-Text2Video Dataset with human motion labels and text descriptions to facilitate the research on text-driven human video generation. + +# 2. Related Work + +Video Synthesis. Similar to image generation, GAN [5, 16, 29] and VQVAE [11, 52] are two common paradigms in the field of video synthesis [49]. MoCoGAN-HD [48] and + +![](images/33a0196bd4e55031ba67b38e1a59442ce858dbb553d648b67471fbe60c8f35e9.jpg) +(a) Sampling from the Decomposed VQ-Space + +![](images/a0584001eead1165443f0f08c553bef0faac85e80dc3a85e0a4b65ab56a3680c.jpg) +(b) Motion Sampling with Continuous VQ-Diffuser + +# Legend + +![](images/59bff63d518594dfd52fb66d2b0e01f7401284000dd76265eb95db0d94597556.jpg) + +Text + +![](images/85aa731ed744aed458b8ab43f654a68ecc3fd7659e9f1696b078d5429e5ad47b.jpg) + +Text encoder + +![](images/3ea013016efacdbdfdfd7dbf7482211f991e481534cbd09049ff6b867dec737b.jpg) + +VQ-VAE Decoder + +![](images/4d5875d527c7998959fbb4570f0f14ad216a78d262cc73728a255b1ab9d69fcc.jpg) + +Sampler + +![](images/59d663e827f68c8c2970620179d4086d557fbb69f2651f052fab6af0b73ad28f.jpg) + +Text embedding + +![](images/f6d93eaeed2b6d810a58daebf6a992bacefca2868301507d37166e06b38d8158.jpg) + +Continuous embedding + +![](images/ed1d88999b8746df380350992fbd77d8f6fd9f592b41f808cdf834152a611baf.jpg) + +Masked continuous embedding + +![](images/1d54f708c0fbfdb0ebdbe1ecf157277077373dd002308c207fe9dcb7b18a3c23.jpg) + +Codebook + +![](images/17cb31432597bf5f7005976c93b0969fb9e7fb2b6d59d33788f6a086476bf9de.jpg) + +Appearance-related + +![](images/b7b20605642c93df6f29d3b3f80b918202fcc6cfedc2e116593d56f839f8ab30.jpg) +Figure 3: Overview of Text2Performer with (a) Sampling from the Decomposed VQ-Space and (b) Motion Sampling with Continuous VQ-Diffuser. Given a text, we first sample the target appearance features $\hat{f}_a$ and exemplar pose features $\hat{f}_p^0$ conditioned on the language features $e_a$ extracted by a pretrained text model. The motion sequence $\hat{\mathbf{F}}_p$ is then sampled by our proposed continuous VQ-Diffuser. The continuous VQ-Diffuser takes the extracted language features $e_m$ and $\hat{f}_p^0$ as inputs. The prediction of continuous motion sequences $\hat{\mathbf{F}}_p$ starts with fully masked pose features and predicts continuous pose embeddings $\{\bar{f}_p^1,\dots,\bar{f}_p^n\}$ with Non-Causal Transformer. The nearest neighbor pose embeddings $\{\hat{f}_p^1,\dots,\hat{f}_p^n\}$ are then retrieved from the pose codebook $\mathcal{C}_p$ . Guided by a motion-aware masking strategy, the continuous VQ-Diffuser is iteratively applied until the whole motion sequence is unmasked. The final videos are generated by feeding the continuous pose features and appearance features into the decoder of VQVAE. + +Motion-related + +StyleGAN-V [46] harness the powerful StyleGAN [29, 30] to generate videos. DI-GAN [55] models an implicit neural representation to generate videos. Brooks et al. [6] propose to regularize the temporal consistency via a temporal discriminator. These works focus on unconditional video generation. As for the VQVAE-based methods, VideoGPT [54] firstly extends the idea of VQVAE and autoregressive transformer to video generation. The following works introduce time-agnostic VQGAN with time-sensitive transformer [15] and masked sequence modeling [20] to further improve the performance. These VQVAE-based video synthesis frameworks accept conditions appended to the beginning of the token sequences. Our proposed Text2Performer is developed on VQVAE. The differences lie in that the designs of VQ-space decomposition and continuous sampling. Recently, there are some concurrent works [22, 24, 44, 53] for text-to-video generation. These methods are designed for general objects, while our Text2Performer has the design to separate human appearance and pose representations, which is specific to human video generation. + +Human Content Generation and Manipulation. Existing works focus on human images, and they are mainly divided into two types: unconditional generation and conditional + +generation. StyleGAN-Human [12] employs StyleGAN to synthesize high-fidelity human images. TryOnGAN [32] and HumanGAN [41] generate human images conditioned on given human poses. Text2Human [27] proposes to generate human images conditioned on texts and poses. We focus on human video generation. For human content manipulation, pose transfer [1, 7, 33, 34, 35] is a popular topic. Pose transfer deals with video data. The task is to transfer the poses from the reference videos to the source video. Chan et al. [7] take human poses as inputs and the desired video is generated by an image-to-image translation network. Our task differs in that we only take texts as inputs and exemplar images are not compulsory. Motion synthesis works [19, 36] generate human motions in form of 3D human representations (e.g., SMPL and skeleton space), while our method generates human motions at the image level. + +# 3. Text2 Performer + +As shown in Fig. 3, our proposed Text2Performer synthesizes desired human videos by taking the texts as inputs $(\pmb{t}_a$ for appearance and $t_m$ for motions). To ensure consistent human representation, we propose to decompose the + +![](images/18d2a22c6a25d1438a3ae1f1bc680e2ec52eb9ca89b38847f73cfceafa6ab0c8.jpg) +Figure 4: Pipeline of Decomposed VQ-Space. Two different frames $I_0$ and $I_k$ of a video serve as the identity frame and the pose frame to provide appearance and pose features, respectively. We apply data augmentations to $I_k$ to erase its appearance information to avoid information leakage. Two codebooks are built to store the pose features and appearance features. The quantized features are finally fed into the decoder to reconstruct the pose frame $I_k$ . + +VQ-space into appearance representation $\pmb{f}_a$ and pose representation $\pmb{f}_p$ as shown in Fig. 4. With the decomposed VQ-space, we sample the human appearance features $\hat{\pmb{f}}_a$ and exemplar pose features $\hat{\pmb{f}}_p^0$ according to $\pmb{t}_a$ . To model the motion dynamics, we propose continuous VQ-diffuser as the motion sampler. It progressively diffuses the masked sequence of pose features until the whole sequence is un-masked. The final video clip is generated by feeding the appearance and pose embeddings into the decoder. + +# 3.1. Decomposed VQ-Space + +Human video generation requires consistent human appearance, i.e., the face and clothing of the target person should remain unchanged while performing the motion. Vanilla VQVAE [52] encodes images into a unified feature representation and builds a codebook on it. However, such design is prone to generate drifted identities along the video frames. To overcome this problem, we propose to decompose the VQ-space into the appearance representation $\pmb{f}_a$ and pose representation $\pmb{f}_p$ in an unsupervised manner. With the decomposed VQ-space, we can separately sample $\hat{\pmb{f}}_a$ and a sequence of $\hat{\pmb{f}}_p$ to generate the desired video. + +In human video data, different frames in one video share the same human identity. We utilize this property to train the decomposed VQ-space [13, 25]. The pipeline is shown in Fig. 4. Given a video $\mathbf{V} = \{\mathbf{I}_0, \mathbf{I}_1, \dots, \mathbf{I}_n\}$ , we use its first frame $\mathbf{I}_0$ for appearance information and another randomly sampled frame $\mathbf{I}_k$ for pose information. $\mathbf{I}_0$ and $\mathbf{I}_k$ are fed into two encoders $E_a$ and $E_p$ , respectively. The two branches share the encoder $E_s$ . To prevent $\mathbf{I}_k$ from leaking appearance information, data augmentations (e.g. color jitter + +tering and Gaussian blur) are applied to $\mathbf{I}_k$ before the pose branch. $f_{a}$ and $f_{p}$ are obtained as + +$$ +\boldsymbol {f} _ {a} = E _ {a} \left(E _ {s} \left(\mathbf {I} _ {0}\right)\right), \boldsymbol {f} _ {p} = E _ {p} \left(E _ {s} \left(\mathbf {I} _ {k} ^ {a}\right)\right)), \tag {1} +$$ + +where $\mathbf{I}_k^a$ is the augmented $\mathbf{I}_k$ . To make $f_{p}$ learn compact and necessary pose information, we make the spatial size of $f_{p}$ smaller. Given $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ , we empirically find that the optimal spatial sizes for $f_{a}$ and $f_{p}$ are $H / 16 \times W / 16$ and $H / 64 \times W / 64$ , respectively. + +After obtaining $\pmb{f}_a$ and $\pmb{f}_p$ , two codebooks $\mathcal{C}_a$ and $\mathcal{C}_p$ are built to store the appearance and pose embeddings. The quantized feature $\hat{\pmb{f}}$ given codebook $\mathcal{C}$ is obtained by + +$$ +\hat {\boldsymbol {f}} = \operatorname {Q u a n t} (\boldsymbol {f}) := \left\{\underset {c _ {k} \in \mathcal {C}} {\operatorname {a r g m i n}} \| f _ {i j} - c _ {k} \| _ {2} \mid f _ {i j} \in \boldsymbol {f} \right\}. \tag {2} +$$ + +With the quantized $\hat{f}_a$ and $\hat{f}_p$ , we then feed them into decoder $D$ to reconstruct the target pose frame $\hat{\mathbf{I}}_k$ : + +$$ +\hat {\mathbf {I}} _ {k} = D ([ \hat {\boldsymbol {f}} _ {a}, D _ {p} (\hat {\boldsymbol {f}} _ {p}) ]), \tag {3} +$$ + +where $[\cdot ]$ is the concatenation operation, and $D_{p}$ upsamples $\hat{\pmb{f}}_p$ to make it have the same resolution as $\hat{\pmb{f}}_a$ . The whole network (including the encoders, decoders, and codebooks) is trained using: + +$$ +\begin{array}{l} \mathcal {L} = \left\| \mathbf {I} _ {k} - \hat {\mathbf {I}} _ {k} \right\| _ {1} + \left\| \operatorname {s g} (\hat {\boldsymbol {f}} _ {a}) - \boldsymbol {f} _ {a} \right\| _ {2} ^ {2} + \left\| \operatorname {s g} (\boldsymbol {f} _ {a}) - \hat {\boldsymbol {f}} _ {a} \right\| _ {2} ^ {2} \\ + \left\| \operatorname {s g} \left(\hat {\boldsymbol {f}} _ {p}\right) - \boldsymbol {f} _ {p} \right\| _ {2} ^ {2} + \left\| \operatorname {s g} \left(\boldsymbol {f} _ {p}\right) - \hat {\boldsymbol {f}} _ {p} \right\| _ {2} ^ {2}, \tag {4} \\ \end{array} +$$ + +where $\mathrm{sg}(\cdot)$ is the stop-gradient operation. + +With the decomposed VQ-space, an additional sampler is trained to sample $\hat{\pmb{f}}_a$ and $\hat{\pmb{f}}_p^0$ . This sampler has the same design as previous methods [4, 27]. + +# 3.2. Continuous VQ-Diffuser + +We adopt the absorbing diffusion transformer [2, 4, 8, 18] to sample the motion, a sequence of pose embeddings $\hat{\mathbf{F}}_p = \{\hat{\pmb{f}}_p^1,\hat{\pmb{f}}_p^2,\dots,\hat{\pmb{f}}_p^n\}$ from the learned pose codebook $\mathcal{C}_p$ Different from autoregressive models [11, 52] that make predictions in a fixed order, absorbing diffusion transformer predicts multiple codebook indices in parallel. The prediction of codebook indices starts from $\mathbf{F}_p^0$ i.e., fully masked $\mathbf{F}_p^m$ . The prediction at time step $t$ is represented as follows: + +$$ +\hat {\mathbf {F}} _ {p} ^ {t} \sim q _ {\theta} \left(\mathbf {F} _ {p} ^ {t} \mid \mathbf {F} _ {p} ^ {t - 1}\right), \tag {5} +$$ + +where $\theta$ denotes the parameters of the transformer sampler. + +To model human motions, we propose 1) continuous space sampling, and 2) motion-aware masking strategy. + +![](images/e8efb86c90da220bbeed4ccc74257ceae82e6ca45e18df6a2ad5e09055a4da72.jpg) +Figure 5: Comparison with Image Samplers. (a) Discrete VQ Sampler predicts discrete codebook indices, which are trained by cross-entropy loss. (b) Continuous Sampler operates on unconstrained continuous RGB spaces. (c) Our VQ-Diffuser predicts continuous embeddings but also utilizes rich embeddings stored in codebooks. + +# 3.2.1 Continuous Space Sampling + +As shown in Fig. 5(a), previous VQVAE-based methods [4, 8, 11, 18, 52] sample tokens in a discrete space. They predict codebook indices to form quantized features, which are then fed into the decoder to generate images. However, sampling at the discrete space is hard to converge because of redundant codebooks, which causes the one-to-many mapping prediction problem. This would lead to meaningless predicted pose sequences. + +In continuous VQ-diffuser $S_{\theta}$ , we propose to train the non-causal transformer in the continuous embedding space as shown in Fig. 5(c). $S_{\theta}$ directly predicts continuous pose embeddings $\{\pmb{f}_p^k\}$ . $S_{\theta}$ is trained to predict the full continuous $\bar{\mathbf{F}}_p$ from the masked pose sequences $\mathbf{F}_p^m$ conditioned on the motion description $\pmb{t}_{m}$ and the initial pose feature $\hat{\pmb{f}}_p^0$ : + +$$ +\bar {\mathbf {F}} _ {p} = S _ {\theta} ([ T (\boldsymbol {t} _ {m}), \hat {\boldsymbol {f}} _ {p} ^ {0}, \mathbf {F} _ {p} ^ {m} ]), \tag {6} +$$ + +where $T(\cdot)$ is the pretrained text feature extractor. + +To utilize rich embeddings stored in $\mathcal{C}_p$ to constrain the prediction space, we retrieve the nearest embedding of the predicted continuous $\bar{\mathbf{F}}_p$ from $\mathcal{C}_p$ to obtain the final $\hat{\mathbf{F}}_p$ : + +$$ +\hat {\mathbf {F}} _ {p} = \operatorname {N e a r e s t} \left(\bar {\mathbf {F}} _ {p}\right) := \left\{\underset {c _ {k} \in \mathcal {C} _ {p}} {\operatorname {a r g m i n}} \left\| \bar {f} _ {p} - c _ {k} \right\| _ {2} \right\}. \tag {7} +$$ + +$\hat{\mathbf{F}}_p$ is then fed into $D$ to reconstruct the final video $\hat{\mathbf{V}}$ following Eq. (3). Thanks to the continuous operation, we can add losses at both the image level and embedding level. The loss function to train our continuous VQ-diffuser is + +$$ +\begin{array}{l} \mathcal {L} = \left\| \bar {\mathbf {F}} _ {p} - \mathbf {F} _ {p} \right\| _ {1} + \left\| \operatorname {s g} \left(\hat {\mathbf {F}} _ {p}\right) - \bar {\mathbf {F}} _ {p} \right\| _ {2} ^ {2} \tag {8} \\ + \left\| \mathrm {s g} (\bar {\mathbf {F}} _ {p}) - \hat {\mathbf {F}} _ {p} \right\| _ {2} ^ {2} + L _ {\mathrm {r e c}} (\hat {\mathbf {V}}, \mathbf {V}), \\ \end{array} +$$ + +where $L_{\mathrm{rec}}$ is composed of $L_{1}$ loss and perceptual loss [28]. It should be noted that stop-gradient operation is applied to make the nearest operation differentiable. + +Compared to continuous samplers like PixelCNN [51] in Fig. 5(b), which directly predict the continuous RGB values, our continuous VQ-diffuser fully utilizes the rich contents stored in the codebook while restricting the prediction space. Therefore, it inherits the benefits of the discrete VQ sampler and continuous sampler at the same time. The pose sequences sampled by our continuous VQ-diffuser are temporally coherent and meaningful. + +# 3.2.2 Motion-Aware Masking Strategy + +To generate plausible videos, the sampled motion sequence $\hat{\mathbf{F}}_p$ should be both temporally and spatially reasonable. In order to make the continuous VQ-diffuser $S_{\theta}$ correctly condition on $\pmb{t}_{m}$ as well as generate reasonable human poses for each frame, we design a motion-aware masking strategy to meticulously mask $\mathbf{F}_p$ temporally and spatially. + +At the temporal dimension, $S_{\theta}$ first predicts pose embeddings of the first frame and the last frame based on $t_m$ and $\hat{\pmb{f}}_p^0$ . Then the prediction diffuses to the intermediate frames according to the given conditions and previously unmasked frames. Therefore, during the training, if the first frame and the last frame are masked, we will mask all frames to prevent the intermediate frames from providing information to help frame predictions at two ends. A higher probability is assigned to masking all frames to help $S_{\theta}$ better learn prediction at the most challenging two ends. + +At the spatial dimension, we propose to predict the first and the last frames with more than one diffusion step. Otherwise, the structure of the sampled human tends to be incomplete. To this end, during the training, we mask the first and last frames partially at the spatial dimension. + +# 4. Fashion-Text2Video Dataset + +Existing human video datasets are either of low resolution [3, 17, 49] or lack of text (action labels) [35, 43]. To support the study of text-driven human video generation, we propose a new dataset, Fashion-Text2Video dataset. Among existing human video datasets, we select videos in Fashion Dataset [56] as the source data for following reasons: 1) High resolution ( $\geq 1024 \times 1024$ ); 2) Diverse appearance. + +The Fashion Dataset contains a total of 600 videos. For each video, we manually annotate motion labels, clothing shapes (length of sleeves and length of clothing) and textures. Motion labels are classified into 22 classes, including standing, moving right-side, moving left-side, turning around, etc. The lengths of sleeves are classified into no-sleeve, three-point, medium and long sleeves. The length of clothing is the length of the upper part of the clothes. + +![](images/99167ea0ac00e7f42b045d427db1ab51b51e6955def74f448a20ec5efa2e4307.jpg) +D + +![](images/0174ab8dd49f5ae04bd0c732244d3d00b335d51989ec143cbbb10a8df3e10b79.jpg) + +![](images/037189fa832080d58a81d45de27ac184010dcf8059096744f9e8640b80af74e5.jpg) +eannnnnne + +![](images/464395c7197fc7442f2a6b9960c9d2620ad976ba1838c9a3d011b879d106368b.jpg) + +Teeerree +![](images/b427eb8988c41894caf93822b06a571db34ca69249cb041e011bd1f93f5b5eca.jpg) +This female is wearing a dress, with graphic pattern. It has long sleeves and it is of three-point length. This female turns right from the side to the front. + +![](images/28428872fb9b3e08cbf8fffd349a9456444ceb5e8bd5067a78a3943a41577bbe.jpg) +The dress this person wears has long-sleeve and it is of three-point length. The texture of it is graphic. The lady turns right from the back to the side. + +![](images/f88311fa6a7a84f8247165d3f5c3ebbcc261bedec3f9f3e6e891d53fae3f27f9.jpg) +Figure 6: Qualitative Comparisons. Text2Performer achieves superior generation qualities compared with baselines. + +![](images/038ae4e547bb16aae03d74152db51a30401e081b77a6e3f7495c678fffc4d9eb.jpg) + +Clothing textures are mainly divided into pure color, floral, graphic, etc. Upon the above labels, we then generate texts for each video using some pre-defined templates. The texts include descriptions of the appearances of the dressed clothes and descriptions for each sub-motions. The proposed Fashion-Text2Video dataset can be applied to the research on Text2Video and Label2Video generation. + +# 5. Experiments + +# 5.1. Comparison Methods + +MMVID [20] is a VQVAE-based method for multimodal video synthesis. We follow the original setting to train MMVID on our dataset. StyleGAN-V [46] is a state-of-the-art video generation method built on StyleGAN. Action labels are used as conditions instead of texts, which we find significantly surpasses the performance of texts for this method. CogVideo [24] is a large-scale pretrained model for Text-to-Video generation. We implement two versions of CogVideo, i.e., CogVideo-v1 and CogVideo-v2. Since + +the original CogVideo utilized the pretrained Text2Image dataset, we train CogVideo-v1 to use the exemplar image as an additional input to compensate for the lack of large-scale pretrained Text2Image models for human images. CogVideo-v2 is with original designs. + +# 5.2. Evaluation Metrics + +Diversity and Quality. 1) We adopt Fréchet Inception Distance (FID) [21] to evaluate the quality of generated human frames. 2) We evaluate the quality of generated videos using Fréchet Video Distance (FVD) and Kernel Video Distance (KVD) [50]. We generate 2,048 video clips to compute the metrics, with 20 frames for each clip. + +Identity Preservation. We use two metrics to evaluate the maintenance of identity along the generated frames. 1) We extract face features of each generated frames using FaceNet [42]. For each generated video clip, we calculate the $l_{2}$ distances between features of the frames and the first frame. 2) We extract features of each generated frames using an off-the-shelf ReID model, i.e., OSNet [42]. These + +Table 1: Quantitative Comparisons. + +
MethodFID ↓FVD ↓KVD ↓Face ↓ReLU ↑
MMVID [20]11.85303.0278.670.90470.9096
StyleGAN-V [46]29.68219.6318.770.86750.9568
CogVideo-v1 [24]39.47645.0389.531.05640.8148
CogVideo-v2 [24]51.76799.80112.241.06210.7960
Text2Performer9.60124.7817.960.85930.9382
+ +![](images/fc611437ae5bf137664f8200d8583da06ab4cbe371b7d358e821a21c6a2f07e6.jpg) +Figure 7: User Study. We achieve the highest scores. + +features are related to identity features. We calculate the average cosine distances between extracted features of the frames and the first frame. The final metrics are obtained by averaging distances among 2,048 generated video clips. + +User Study. Users are asked to give three separate scores (the highest score is 4) at three dimensions: 1) The consistency with the text for appearance, 2) The consistency with the text for desired motion, and 3) The overall quality of the generated video in terms of temporal coherence and its realism. A total of 20 users participate in the user study. Each user is presented with 30 videos generated by different methods. + +# 5.3. Qualitative Comparisons + +We show two examples in Fig. 6. In the first example, MMVID and Text2Performer successfully generate desired human motions. However, frames generated by MMVID have drifted identities. StyleGAN-V fails to generate plausible frames as well as consistent motions. As for the second example, only Text2Performer generates corresponding motions. Artifacts appear in the video clips generated by MMVID and StyleGAN-V. The visual comparisons demonstrate the superiority of Text2Performer. + +# 5.4. Quantitative Comparisons + +The quantitative comparisons are shown in Table 1. In terms of FID, FVD and KVD, our proposed Text2Performer has significant improvements over baselines, which demonstrates the diversity and temporal coherence of the videos generated by our methods. Existing baselines generate videos from entangled human representations and thus suffer from temporal incoherence. As for the identity metrics, our method achieves the best performance on face distance and comparable performance on ReID distance with StyleGAN-V. Since StyleGAN-V is prone to generate videos with small motions, features tend to be similar, + +![](images/4991f26c8fd6fae6c84895b0df4c4ccb01be5d1e75e35f61b4fdbfc0c75a7253.jpg) +Figure 8: Ablation Studies on (a) Decomposed VQ-Space, (b) Spatial Resolution of Pose Branch, (c) Discrete VQ sampler, and (d) Usage of Codebook. + +leading to high ReID scores. In our method, thanks to the decomposed VQ-space and continuous VQ-diffuser, temporal and identity consistency can be better achieved. The result of the user study is shown in Fig. 7. Text2Performer achieves the highest scores on all three dimensions. Specifically, the result of the assessment on overall quality is consistent with the other quantitative metrics, which further validates the effectiveness of Text2Performer. + +# 5.5. Ablation Study + +Decomposed VQ-Space. The ablation model trains the continuous VQ-diffuser with a unified VQ-space, which fuses the identity and pose information within one feature. As shown in Fig. 8(a), compared to our results, sampling in the unified VQ-Space generates drifted identity and incomplete body structures. This leads to inferior quantitative metrics in Table 2. The unified VQ-Space requires the sampler to handle identity and pose at the same time, which poses burdens for the sampler. With the decomposed VQ-Space, the identity can be easily maintained by fixing the appearance features, which further helps learn motions. + +Spatial Resolution of Pose Branch. In our design, the pose feature has $4 \times$ smaller spatial resolution than the appearance feature. In Fig. 8(b), we show the results generated by a decomposed VQVAE where pose and appearance features have the same resolution. Compared to results of the full model, body structures in the side view are incomplete. The quantitative metrics in Table 2 and Fig. 8(b) demonstrate that the smaller pose feature is necessary as the larger feature contains redundant information to harm valid feature disentanglement, thus adding challenges to the sampler. + +Discrete VQ sampler in Decomposed VQ-Space. We train a variant of VQ-Diffuser to predict discrete codebook indices with the decomposed VQ-Space. In Fig. 8(c), the + +![](images/2a0daad2d1921ccb971194ab1f31de5c234de4837d77777e9bbae5af3d9e3f62.jpg) +Figure 9: High-Resolution Results. We show one result generated by a long text sequence. The identity is well maintained. + +![](images/00b29ec0f0117bcf53c4c297361a9f5801c1bb434f26af13320908fb3f53b79e.jpg) +Figure 10: Cross-Dataset Performance. We compare with models trained on large-scale data and domain-specific data. + +![](images/788fb8a095a8fb5514ee54804ee9f828d6f40acb660717fa7cdd60922b1df37d.jpg) + +![](images/3dd2f4f78edbfed8b4ebce517b971cfa2fdd37f206bd81970d3a3200a9e91328.jpg) + +Table 2: Quantitative Results on Ablation Studies. + +
MethodFID ↓FVD ↓KVD ↓Face ↓ReID ↑
w/o decomp.20.52266.9735.520.89360.9213
same res.15.69202.0632.960.86490.9399
disc. sampler10.37308.8873.170.91060.9161
w/o codebook11.31154.9322.800.94370.9109
Full Model10.92134.4320.310.85200.9421
+ +discrete VQ Sampler can generate complete body structures but fails to generate video clips consistent with the given action control, i.e., turning right from the back to the side. Failure on conditioning on action controls does not harm the quality of each generated frame, and thus this ablation study achieves a comparable FID score as our method. However, our method significantly outperforms this variant on other metrics as shown in Table 2. In VQVAE, some codebook entries have similar contents but with different codebook indices. Therefore, treating the training of the sampler as a classification task makes the model hard to converge. + +Usage of Codebook in Continuous VQ-Space. In our design, after the prediction of embeddings, the nearest neighbor of embeddings will be retrieved from the codebook. We train an auto-encoder without codebook in the pose branch. The sampler is trained to sample continuous embeddings from the latent space of auto-encoder without the retrieval from codebook. As shown in Fig. 8(d), without the codebook, the sampler results in a disrupted human image at + +the side view. The quality of generated videos deteriorates as reflected in Table 2. This demonstrates that predicting compact and meaningful continuous embeddings with codebooks enhances the quality of synthesized frames. + +# 5.6. Additional Analysis + +High-Resolution Results. Text2Performer can generate long videos with high resolutions (512 × 256) as shown in Fig. 1 and Fig. 9. Figure 9 shows an example of high-resolution videos for a long text sequence. + +Cross-Dataset Performance. We further verify the performance of models trained on different datasets. The results are presented in Fig. 10. Compared with CogVideo [24] pretrained on large-scale dataset, our Text2Performer and MMVID succeed to generate plausible specific human content. As for models trained on small-scale domain-specific data, our Text2Performer is comparable to MMVID on the simple iPer dataset [35], and has superior performance on the more challenging Fashion-Text2Video dataset, especially on the completeness of body structures across frames. + +Ability to Generate Novel Appearance. To verify that the generated humans are with novel appearances, we retrieve top-2 nearest neighbour images of generated frames using the perceptual distance [28]. As shown in Fig. 11, the generated query frames have different appearances from the top-2 nearest images retrieved from the dataset. + +![](images/1da77b9d9e51b4dc40eebfafd041f852e91d73c50992015c987306cd6e814b42.jpg) +Query Image Top-2 Nearest Images Query Image Top-2 Nearest Images + +![](images/dd84e38c790c40f4ae63428be22bba459e273829e5dc3fa5a05c2010967ba48c.jpg) + +Query Image Top-2 Nearest Images + +![](images/2f2efce4631626861f257dada972706ce0ef378b11a0fdceead142717aa7fb87.jpg) +Figure 11: Nearest Neighbour Image Search. Top-2 nearest images of generated images are retrieved. +Figure 12: Results on In-the-Wild Dataset. In our current pipeline, We treat the background as a part of appearance information. The results contain some artifacts. + +# 6. Discussions + +We propose a novel Text2Performer for text-driven human video generation. We decompose the VQ-space into appearance and pose representations. Continuous VQ-diffuser is then introduced to sample motions conditioned on texts. These designs empower Text2Performer to synthesize high-quality and high-resolution human videos. + +Limitations and Future Work. In this paper, we focus on the generation of human appearance and motion with clean backgrounds. The current pipeline is designed for generating human videos without textural backgrounds. It can be extended to generate plausible textual backgrounds by treating the background as a part of appearance information at the training time. The generated results are shown in Fig. 12. There are some artifacts in the generated videos. Generating videos with textual backgrounds is important in real-world applications. In future work, textual backgrounds can be segmented first. The results can be then improved by generating the foreground and textual background separately and fusing them harmoniously using an additional module. In addition, the synthesized human videos are biased toward generating females with dresses. This is because the Fashion Dataset only contains videos of females with dresses. In future applications, more data can be involved in the training to alleviate the bias. Besides, the control modality can also be extended to various types, such as audio and music. + +Potential Negative Societal Impacts. The model may be applied to generate fake videos performing various actions. To alleviate the negative impacts, DeepFake detection methods can be applied to evaluate the realism of videos. + +Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collabora + +tion Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This study is also supported by NTU NAP, MOE AcRF Tier 1 (2021-T1-001-088), and the Google PhD Fellowship. + +# References + +[1] Badour Albahar, Jingwan Lu, Jimei Yang, Zhixin Shu, Eli Shechtman, and Jia-Bin Huang. Pose with style: Detail-preserving pose-guided image synthesis with conditional stylegan. ACM Transactions on Graphics (TOG), 40(6):1-11, 2021. 3 +[2] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. NeurIPS, 34:17981-17993, 2021. 4 +[3] Guha Balakrishnan, Amy Zhao, Adrian V Dalca, Fredo Durand, and John Guttag. Synthesizing images of humans in unseen poses. In CVPR, pages 8340-8348, 2018. 5 +[4] Sam Bond-Taylor, Peter Hessey, Hiroshi Sasaki, Toby P Breckon, and Chris G Willcocks. Unleashing transformers: parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vector-quantized codes. In ECCV, pages 170–188. Springer, 2022. 2, 4, 5 +[5] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ICLR, 2019. 2 +[6] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. NeurIPS, 35:31769-31781, 2022. 3 +[7] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In ICCV, pages 5933-5942, 2019. 3 +[8] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, pages 11315-11325, 2022. 4, 5 +[9] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. NeurlPS, 34:19822-19835, 2021. 1 +[10] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. NeurIPS, 35:16890-16902, 2022. 1 +[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2, 4, 5 +[12] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, pages 1-19. Springer, 2022. 3 +[13] Aviv Gabbay and Yedid Hoshen. Demystifying inter-class disentanglement. *ICLR*, 2020. 4 + +[14] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, pages 89-106. Springer, 2022. 1 +[15] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and timesensitive transformer. In ECCV, pages 102-118. Springer, 2022. 3 +[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2 +[17] Lena Gorelick, Moshe Blank, Eli Shechtman, Michal Irani, and Ronen Basri. Actions as space-time shapes. IEEE transactions on pattern analysis and machine intelligence, 29(12):2247-2253, 2007. 5 +[18] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, pages 10696-10706, 2022. 1, 2, 4, 5 +[19] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 3 +[20] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and Sergey Tulyakov. Show me what and tell me how: Video synthesis via multimodal conditioning. In CVPR, pages 3615-3625, 2022. 3, 6, 7 +[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 6 +[22] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 3 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2 +[24] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. *ICLR*, 2023. 1, 2, 3, 6, 7, 8 +[25] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. NeurIPS, 31, 2018. 4 +[26] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In ICML, pages 9902-9915. PMLR, 2022. 2 +[27] Yuming Jiang, Shuai Yang, Haonan Qju, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3, 4 + +[28] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016. 5, 8 +[29] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410, 2019. 2, 3 +[30] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR., pages 8110-8119, 2020. 3 +[31] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. NeurIPS, 32, 2019. 2 +[32] Kathleen M Lewis, Srivatsan Varadharajan, and Ira Kemelmacher-Shlizerman. Tryongan: Body-aware try-on via layered interpolation. ACM Transactions on Graphics (TOG), 40(4):1-10, 2021. 3 +[33] Lingjie Liu, Weipeng Xu, Marc Habermann, Michael Zollhöfer, Florian Bernard, Hyeongwoo Kim, Wenping Wang, and Christian Theobalt. Neural human video rendering by learning dynamic textures and rendering-to-video translation. IEEE Transactions on Visualization and Computer Graphics, PP:1-1, 05 2020. 3 +[34] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 38(5):1-14, 2019. 3 +[35] Wen Liu, Zhixin Piao, Jie Min, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis. In ICCV, pages 5904-5913, 2019. 3, 5, 8 +[36] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In ICCV, pages 10985-10995, 2021. 3 +[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. PMLR, 2021. 1 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2 +[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 1 +[41] Kripasindhu Sarkar, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Humangan: A generative model of human images. In 2021 International Conference on 3D Vision (3DV), pages 258–267. IEEE, 2021. 3 + +[42] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823, 2015. 6 +[43] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. NeurIPS, 32, 2019. 5 +[44] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. ICLR, 2023. 1, 3 +[45] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In CVPR, pages 11050-11059, 2022. 2 +[46] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, pages 3626-3636, 2022. 3, 6, 7 +[47] Jianxin Sun, Qiyao Deng, Qi Li, Muyi Sun, Min Ren, and Zhenan Sun. Anyface: Free-style text-to-face synthesis and manipulation. In CVPR, pages 18687-18696, 2022. 1 +[48] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. ICLR, 2021. 2 +[49] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, pages 1526-1535, 2018. 2, 5 +[50] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6 +[51] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. NeurIPS, 29, 2016. 5 +[52] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 2, 4, 5 +[53] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. ICLR, 2023. 1, 3 +[54] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3 +[55] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. *ICLR*, 2022. 3 +[56] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose-guided human video generation. arXiv preprint arXiv:1910.09139, 2019. 2, 5 \ No newline at end of file diff --git a/2023/Text2Performer_ Text-Driven Human Video Generation/images.zip b/2023/Text2Performer_ Text-Driven Human Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2417b8e5d9023b46e5090925fe89a715857cfccd --- /dev/null +++ b/2023/Text2Performer_ Text-Driven Human Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5e7ccffd5e8d818a5cf8e5d4a06dca0240795a639755885b6bfd4c68513e882 +size 670014 diff --git a/2023/Text2Performer_ Text-Driven Human Video Generation/layout.json b/2023/Text2Performer_ Text-Driven Human Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ae1116ef9375fb286f7886f8b297e3f3c476c5db --- /dev/null +++ b/2023/Text2Performer_ Text-Driven Human Video Generation/layout.json @@ -0,0 +1,10736 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 470, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 470, + 119 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 470, + 119 + ], + "type": "text", + "content": "Text2 Performer: Text-Driven Human Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "text", + "content": "Yuming Jiang" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "text", + "content": " Shuai Yang" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "text", + "content": " Tong Liang Koh" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "text", + "content": " Wayne Wu" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "text", + "content": " Chen Change Loy" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "text", + "content": " Ziwei Liu" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University " + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 552, + 172 + ], + "type": "text", + "content": "Shanghai AI Laboratory" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 174, + 538, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 174, + 538, + 186 + ], + "spans": [ + { + "bbox": [ + 68, + 174, + 538, + 186 + ], + "type": "text", + "content": "{yuming002, shuai.yang, koht0029, ccloy, ziwei.liu}@ntu.edu.sg wuwenyan0503@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 53, + 215, + 306, + 329 + ], + "blocks": [ + { + "bbox": [ + 53, + 215, + 306, + 329 + ], + "lines": [ + { + "bbox": [ + 53, + 215, + 306, + 329 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 306, + 329 + ], + "type": "image", + "image_path": "b424bc7500ed5243fbb519ce0bb57e99df79987f7d9da9fb46f01047fc84abdb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 379, + 546, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 379, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 546, + 403 + ], + "type": "text", + "content": "Figure 1: High-resolution videos generated by Text2Performer. The videos are generated by taking the text descriptions as the only input. The generated videos contain diverse appearances and flexible motions. Identities are well maintained." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 67, + 335, + 91, + 360 + ], + "blocks": [ + { + "bbox": [ + 67, + 335, + 91, + 360 + ], + "lines": [ + { + "bbox": [ + 67, + 335, + 91, + 360 + ], + "spans": [ + { + "bbox": [ + 67, + 335, + 91, + 360 + ], + "type": "image", + "image_path": "aefea0d56de1b949c95bb51e5647b73136b96d1b309a1cfae66a865a1cb7a6ec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 97, + 334, + 291, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 334, + 291, + 365 + ], + "spans": [ + { + "bbox": [ + 97, + 334, + 291, + 365 + ], + "type": "text", + "content": "This woman wears a dress. It has long sleeves, and it is of long length. The texture of it is graphic. She is swinging to the right." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 215, + 541, + 329 + ], + "blocks": [ + { + "bbox": [ + 309, + 215, + 541, + 329 + ], + "lines": [ + { + "bbox": [ + 309, + 215, + 541, + 329 + ], + "spans": [ + { + "bbox": [ + 309, + 215, + 541, + 329 + ], + "type": "image", + "image_path": "53d11366d4394c29962c13e7b70dd849945205b7218de02f41d9fafade57343a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 338, + 342, + 360 + ], + "blocks": [ + { + "bbox": [ + 317, + 338, + 342, + 360 + ], + "lines": [ + { + "bbox": [ + 317, + 338, + 342, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 342, + 360 + ], + "type": "image", + "image_path": "994493edf4f6027469fc298979094b0160332d4e85434250aa08a58bbbb0f596.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 348, + 334, + 529, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 334, + 529, + 365 + ], + "spans": [ + { + "bbox": [ + 348, + 334, + 529, + 365 + ], + "type": "text", + "content": "The person is wearing a sleeveless dress. It is of short length. Its pattern is pure color. She turns right from the front to the side." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 413, + 192, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 413, + 192, + 426 + ], + "spans": [ + { + "bbox": [ + 143, + 413, + 192, + 426 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 438, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 438, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 438, + 290, + 715 + ], + "type": "text", + "content": "Text-driven content creation has evolved to be a transformative technique that revolutionizes creativity. Here we study the task of text-driven human video generation, where a video sequence is synthesized from texts describing the appearance and motions of a target performer. Compared to general text-driven video generation, human-centric video generation requires maintaining the appearance of synthesized human while performing complex motions. In this work, we present Text2Performer to generate vivid human videos with articulated motions from texts. Text2Performer has two novel designs: 1) decomposed human representation and 2) diffusion-based motion sampler. First, we decompose the VQVAE latent space into human appearance and pose representation in an unsupervised manner by utilizing the nature of human videos. In this way, the appearance is well maintained along the generated frames. Then, we propose continuous VQ-diffuser to sample a sequence of pose embeddings. Unlike existing VQ-based methods that operate in the discrete space, continuous VQ-diffuser directly outputs the continuous pose embeddings for better motion modeling. Finally, motion-aware masking strategy is designed to mask the pose embeddings spatial-temporally to enhance the temporal coherence. Moreover," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 415, + 547, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 511 + ], + "type": "text", + "content": "to facilitate the task of text-driven human video generation, we contribute a Fashion-Text2Video dataset with manually annotated action labels and text descriptions. Extensive experiments demonstrate that Text2Performer generates high-quality human videos (up to " + }, + { + "bbox": [ + 304, + 415, + 547, + 511 + ], + "type": "inline_equation", + "content": "512 \\times 256" + }, + { + "bbox": [ + 304, + 415, + 547, + 511 + ], + "type": "text", + "content": " resolution) with diverse appearances and flexible motions. Our project page is https://yumingj.github.io/projects/Text2Performer.html" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 525, + 387, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 387, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 387, + 537 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "Since its emergence, text-guided image synthesis (e.g. DALLE [37, 38]) has attracted substantial attention. Recent works [9, 10, 14, 18, 40, 47] have demonstrated fascinating performance for the quality of synthesized images and their consistency with the texts. Beyond image generation, text-driven video generation [22, 24, 44, 53] is an advanced topic to explore. Existing works rely on large-scale datasets to drive large models. Although they have achieved surprising performance on general objects, when applied to the generation of some specific tasks, such as generating videos of garment presents for e-commerce websites, they fail to generate plausible results. Take the CogVideo [24] as an example. As shown in Fig. 2, the generated human objects contain incomplete human structures, and the temporal" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22747" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 284, + 120 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 284, + 120 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 284, + 120 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 284, + 120 + ], + "type": "image", + "image_path": "3d5a6148d0c6be5ae87c6f5da2697825f347aa3b6a8b31c19cc0f87c3e3d7b7f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 122, + 282, + 142 + ], + "lines": [ + { + "bbox": [ + 52, + 122, + 282, + 142 + ], + "spans": [ + { + "bbox": [ + 52, + 122, + 282, + 142 + ], + "type": "text", + "content": "The person is wearing a sleeveless dress. It is of short length. Its pattern is pure color. She turns right from the front to the side." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 146, + 287, + 207 + ], + "lines": [ + { + "bbox": [ + 46, + 146, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 287, + 207 + ], + "type": "text", + "content": "Figure 2: Results of General Large Text-to-Video Models. We use the pretrained general large Text-to-Video Models [24] to generate videos using the same texts as the right example of Fig. 1. The result fails to generate complete human structures and maintains temporal consistency." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 287, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 268 + ], + "type": "text", + "content": "consistency is poorly maintained. On the other hand, these methods need billions of training data, which hampers the application to those specific tasks (e.g. human video generation) without a large amount of paired data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 269, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 448 + ], + "type": "text", + "content": "Therefore, it is worthwhile to explore text-to-video generation in human video generation, which has numerous applications [31, 45]. In this paper, we focus on the task of text-driven human video generation. Compared to general text-to-video generation, text-driven human video generation poses several unique challenges: 1) The human structure is articulated. The joint movements of different body components form many complicated out-of-plane motions, e.g., rotations. 2) When performing complicated motions, the appearance of the synthesized human should remain the same. For example, the appearance of a target human after turning around should be consistent with that at the first beginning. In sum, to achieve high-fidelity human video generation, consistent human representation and complicated human motions should be carefully modeled." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 664 + ], + "type": "text", + "content": "We propose a novel text-driven human video generation framework Text2Performer to handle consistent human representations and complex out-of-plane motions. As shown in Fig. 1, given texts describing appearance and motions, Text2Performer is able to generate temporally consistent human videos with complete human structures and unified human appearances. Text2Performer is built upon VQVAE-based frameworks [4, 11, 18, 52]. In Text2Performer, thanks to the specific nature of human videos which shares the same objects across the frames within one video, VQVAE latent space can be decomposed into appearance and pose representations. With the decomposed VQ-space, videos are generated by sampling appearance representation and a sequence of pose representations separately. This decomposition contributes to the maintenance of human identity. Besides, it makes the motion modeling more tractable, as the motion sampler does not need to take the appearance information into consideration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "To model complicated human motions, a novel continuous VQ-diffuser is proposed to sample a sequence of meaningful pose representations. The architecture of the continuous VQ-diffuser is transformer. The key difference to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "type": "text", + "content": "the previous transformer-based samplers [4, 11, 52] is that the continuous VQ-diffuser directly predicts the continuous pose embeddings rather than their indices in the codebook. After predicting continuous pose embeddings, we also make use of the rich embeddings stored in the codebook by retrieving the nearest embeddings of the predicted embeddings from the codebook. Predicting continuous embeddings alleviates the one-to-many prediction issue in previous discrete methods and the use of codebook constrains the prediction space. With this design, more temporally coherent human motions and more complete structures of human frames are sampled. In addition, we borrow the idea of diffusion models [4, 18, 23, 39, 26] to progressively predict the long sequence of the pose embeddings. We propose a motion-aware masking strategy to sample the pose embeddings of the first frame and last frame firstly. Then the pose embeddings of the intermediate frames are gradually diffused. The motion-aware masking strategy enhances the completeness of human structures and temporal coherence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 300, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 383 + ], + "type": "text", + "content": "To facilitate the task of text-guided human video generation, we propose the Fashion-Text2Video Dataset. It is built upon the FashionDataset [56], which consists of 600 human videos performing the fashion show. We manually segment the whole video into clips and label the motion types. Each clip is performing one motion. With the manually labeled motion labels, we then pair them with text descriptions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 384, + 503, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 384, + 503, + 394 + ], + "spans": [ + { + "bbox": [ + 317, + 384, + 503, + 394 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 398, + 545, + 639 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 317, + 398, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 398, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 317, + 398, + 545, + 445 + ], + "type": "text", + "content": "- We present and study the task of text-guided human video generation. Our proposed Text2Performer can be well trained and have generative abilities with only a small amount of data for training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 447, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 447, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 447, + 545, + 529 + ], + "type": "text", + "content": "- We propose to decompose the VQ-space into appearance and pose representations. The decomposition is achieved by making use of the nature of human videos, i.e., the motions are performed under one identity (appearance) across the frames. The decomposition of VQ-space improves the appearance coherence among frames and eases motion modeling." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 531, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 531, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 531, + 545, + 602 + ], + "type": "text", + "content": "- We propose the continuous VQ-diffuser to predict continuous pose embeddings with the pose codebook. The final continuous pose embeddings are iteratively predicted with the guidance of the motion-aware masking scheme. These designs contribute to generated frames of high quality and temporal coherence." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 604, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 545, + 639 + ], + "type": "text", + "content": "- We construct the Fashion-Text2Video Dataset with human motion labels and text descriptions to facilitate the research on text-driven human video generation." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 657, + 392, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 392, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 392, + 670 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Video Synthesis. Similar to image generation, GAN [5, 16, 29] and VQVAE [11, 52] are two common paradigms in the field of video synthesis [49]. MoCoGAN-HD [48] and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22748" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 88, + 428, + 148 + ], + "blocks": [ + { + "bbox": [ + 64, + 72, + 230, + 83 + ], + "lines": [ + { + "bbox": [ + 64, + 72, + 230, + 83 + ], + "spans": [ + { + "bbox": [ + 64, + 72, + 230, + 83 + ], + "type": "text", + "content": "(a) Sampling from the Decomposed VQ-Space" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 88, + 428, + 148 + ], + "lines": [ + { + "bbox": [ + 64, + 88, + 428, + 148 + ], + "spans": [ + { + "bbox": [ + 64, + 88, + 428, + 148 + ], + "type": "image", + "image_path": "33a0196bd4e55031ba67b38e1a59442ce858dbb553d648b67471fbe60c8f35e9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 65, + 171, + 427, + 305 + ], + "blocks": [ + { + "bbox": [ + 62, + 158, + 244, + 169 + ], + "lines": [ + { + "bbox": [ + 62, + 158, + 244, + 169 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 244, + 169 + ], + "type": "text", + "content": "(b) Motion Sampling with Continuous VQ-Diffuser" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 171, + 427, + 305 + ], + "lines": [ + { + "bbox": [ + 65, + 171, + 427, + 305 + ], + "spans": [ + { + "bbox": [ + 65, + 171, + 427, + 305 + ], + "type": "image", + "image_path": "a0584001eead1165443f0f08c553bef0faac85e80dc3a85e0a4b65ab56a3680c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 474, + 74, + 505, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 74, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 474, + 74, + 505, + 85 + ], + "type": "text", + "content": "Legend" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 451, + 90, + 467, + 102 + ], + "blocks": [ + { + "bbox": [ + 451, + 90, + 467, + 102 + ], + "lines": [ + { + "bbox": [ + 451, + 90, + 467, + 102 + ], + "spans": [ + { + "bbox": [ + 451, + 90, + 467, + 102 + ], + "type": "image", + "image_path": "59bff63d518594dfd52fb66d2b0e01f7401284000dd76265eb95db0d94597556.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 475, + 93, + 489, + 101 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 475, + 93, + 489, + 101 + ], + "spans": [ + { + "bbox": [ + 475, + 93, + 489, + 101 + ], + "type": "text", + "content": "Text" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 451, + 109, + 467, + 125 + ], + "blocks": [ + { + "bbox": [ + 451, + 109, + 467, + 125 + ], + "lines": [ + { + "bbox": [ + 451, + 109, + 467, + 125 + ], + "spans": [ + { + "bbox": [ + 451, + 109, + 467, + 125 + ], + "type": "image", + "image_path": "85aa731ed744aed458b8ab43f654a68ecc3fd7659e9f1696b078d5429e5ad47b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 475, + 114, + 512, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 475, + 114, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 475, + 114, + 512, + 121 + ], + "type": "text", + "content": "Text encoder" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 451, + 130, + 467, + 146 + ], + "blocks": [ + { + "bbox": [ + 451, + 130, + 467, + 146 + ], + "lines": [ + { + "bbox": [ + 451, + 130, + 467, + 146 + ], + "spans": [ + { + "bbox": [ + 451, + 130, + 467, + 146 + ], + "type": "image", + "image_path": "3ea013016efacdbdfdfd7dbf7482211f991e481534cbd09049ff6b867dec737b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 475, + 133, + 527, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 475, + 133, + 527, + 142 + ], + "spans": [ + { + "bbox": [ + 475, + 133, + 527, + 142 + ], + "type": "text", + "content": "VQ-VAE Decoder" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 451, + 152, + 467, + 167 + ], + "blocks": [ + { + "bbox": [ + 451, + 152, + 467, + 167 + ], + "lines": [ + { + "bbox": [ + 451, + 152, + 467, + 167 + ], + "spans": [ + { + "bbox": [ + 451, + 152, + 467, + 167 + ], + "type": "image", + "image_path": "4d5875d527c7998959fbb4570f0f14ad216a78d262cc73728a255b1ab9d69fcc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 474, + 155, + 499, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 155, + 499, + 163 + ], + "spans": [ + { + "bbox": [ + 474, + 155, + 499, + 163 + ], + "type": "text", + "content": "Sampler" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 451, + 175, + 467, + 190 + ], + "blocks": [ + { + "bbox": [ + 451, + 175, + 467, + 190 + ], + "lines": [ + { + "bbox": [ + 451, + 175, + 467, + 190 + ], + "spans": [ + { + "bbox": [ + 451, + 175, + 467, + 190 + ], + "type": "image", + "image_path": "59d663e827f68c8c2970620179d4086d557fbb69f2651f052fab6af0b73ad28f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 474, + 175, + 506, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 175, + 506, + 191 + ], + "spans": [ + { + "bbox": [ + 474, + 175, + 506, + 191 + ], + "type": "text", + "content": "Text embedding" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 451, + 198, + 467, + 220 + ], + "blocks": [ + { + "bbox": [ + 451, + 198, + 467, + 220 + ], + "lines": [ + { + "bbox": [ + 451, + 198, + 467, + 220 + ], + "spans": [ + { + "bbox": [ + 451, + 198, + 467, + 220 + ], + "type": "image", + "image_path": "f6d93eaeed2b6d810a58daebf6a992bacefca2868301507d37166e06b38d8158.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 474, + 203, + 508, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 203, + 508, + 219 + ], + "spans": [ + { + "bbox": [ + 474, + 203, + 508, + 219 + ], + "type": "text", + "content": "Continuous embedding" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 451, + 225, + 467, + 247 + ], + "blocks": [ + { + "bbox": [ + 451, + 225, + 467, + 247 + ], + "lines": [ + { + "bbox": [ + 451, + 225, + 467, + 247 + ], + "spans": [ + { + "bbox": [ + 451, + 225, + 467, + 247 + ], + "type": "image", + "image_path": "ed1d88999b8746df380350992fbd77d8f6fd9f592b41f808cdf834152a611baf.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 474, + 230, + 529, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 230, + 529, + 247 + ], + "spans": [ + { + "bbox": [ + 474, + 230, + 529, + 247 + ], + "type": "text", + "content": "Masked continuous embedding" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 451, + 254, + 467, + 270 + ], + "blocks": [ + { + "bbox": [ + 451, + 254, + 467, + 270 + ], + "lines": [ + { + "bbox": [ + 451, + 254, + 467, + 270 + ], + "spans": [ + { + "bbox": [ + 451, + 254, + 467, + 270 + ], + "type": "image", + "image_path": "1d54f708c0fbfdb0ebdbe1ecf157277077373dd002308c207fe9dcb7b18a3c23.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 474, + 258, + 504, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 258, + 504, + 266 + ], + "spans": [ + { + "bbox": [ + 474, + 258, + 504, + 266 + ], + "type": "text", + "content": "Codebook" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 451, + 276, + 467, + 291 + ], + "blocks": [ + { + "bbox": [ + 451, + 276, + 467, + 291 + ], + "lines": [ + { + "bbox": [ + 451, + 276, + 467, + 291 + ], + "spans": [ + { + "bbox": [ + 451, + 276, + 467, + 291 + ], + "type": "image", + "image_path": "17cb31432597bf5f7005976c93b0969fb9e7fb2b6d59d33788f6a086476bf9de.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 474, + 279, + 529, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 279, + 529, + 287 + ], + "spans": [ + { + "bbox": [ + 474, + 279, + 529, + 287 + ], + "type": "text", + "content": "Appearance-related" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 451, + 296, + 467, + 309 + ], + "blocks": [ + { + "bbox": [ + 451, + 296, + 467, + 309 + ], + "lines": [ + { + "bbox": [ + 451, + 296, + 467, + 309 + ], + "spans": [ + { + "bbox": [ + 451, + 296, + 467, + 309 + ], + "type": "image", + "image_path": "b7b20605642c93df6f29d3b3f80b918202fcc6cfedc2e116593d56f839f8ab30.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "lines": [ + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "text", + "content": "Figure 3: Overview of Text2Performer with (a) Sampling from the Decomposed VQ-Space and (b) Motion Sampling with Continuous VQ-Diffuser. Given a text, we first sample the target appearance features " + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "inline_equation", + "content": "\\hat{f}_a" + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "text", + "content": " and exemplar pose features " + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "inline_equation", + "content": "\\hat{f}_p^0" + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "text", + "content": " conditioned on the language features " + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "inline_equation", + "content": "e_a" + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "text", + "content": " extracted by a pretrained text model. The motion sequence " + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_p" + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "text", + "content": " is then sampled by our proposed continuous VQ-Diffuser. The continuous VQ-Diffuser takes the extracted language features " + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "inline_equation", + "content": "e_m" + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "inline_equation", + "content": "\\hat{f}_p^0" + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "text", + "content": " as inputs. The prediction of continuous motion sequences " + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_p" + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "text", + "content": " starts with fully masked pose features and predicts continuous pose embeddings " + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "inline_equation", + "content": "\\{\\bar{f}_p^1,\\dots,\\bar{f}_p^n\\}" + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "text", + "content": " with Non-Causal Transformer. The nearest neighbor pose embeddings " + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "inline_equation", + "content": "\\{\\hat{f}_p^1,\\dots,\\hat{f}_p^n\\}" + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "text", + "content": " are then retrieved from the pose codebook " + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_p" + }, + { + "bbox": [ + 46, + 311, + 547, + 427 + ], + "type": "text", + "content": ". Guided by a motion-aware masking strategy, the continuous VQ-Diffuser is iteratively applied until the whole motion sequence is unmasked. The final videos are generated by feeding the continuous pose features and appearance features into the decoder of VQVAE." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 474, + 298, + 517, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 298, + 517, + 307 + ], + "spans": [ + { + "bbox": [ + 474, + 298, + 517, + 307 + ], + "type": "text", + "content": "Motion-related" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 437, + 289, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 289, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 289, + 677 + ], + "type": "text", + "content": "StyleGAN-V [46] harness the powerful StyleGAN [29, 30] to generate videos. DI-GAN [55] models an implicit neural representation to generate videos. Brooks et al. [6] propose to regularize the temporal consistency via a temporal discriminator. These works focus on unconditional video generation. As for the VQVAE-based methods, VideoGPT [54] firstly extends the idea of VQVAE and autoregressive transformer to video generation. The following works introduce time-agnostic VQGAN with time-sensitive transformer [15] and masked sequence modeling [20] to further improve the performance. These VQVAE-based video synthesis frameworks accept conditions appended to the beginning of the token sequences. Our proposed Text2Performer is developed on VQVAE. The differences lie in that the designs of VQ-space decomposition and continuous sampling. Recently, there are some concurrent works [22, 24, 44, 53] for text-to-video generation. These methods are designed for general objects, while our Text2Performer has the design to separate human appearance and pose representations, which is specific to human video generation." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "Human Content Generation and Manipulation. Existing works focus on human images, and they are mainly divided into two types: unconditional generation and conditional" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 437, + 547, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 631 + ], + "type": "text", + "content": "generation. StyleGAN-Human [12] employs StyleGAN to synthesize high-fidelity human images. TryOnGAN [32] and HumanGAN [41] generate human images conditioned on given human poses. Text2Human [27] proposes to generate human images conditioned on texts and poses. We focus on human video generation. For human content manipulation, pose transfer [1, 7, 33, 34, 35] is a popular topic. Pose transfer deals with video data. The task is to transfer the poses from the reference videos to the source video. Chan et al. [7] take human poses as inputs and the desired video is generated by an image-to-image translation network. Our task differs in that we only take texts as inputs and exemplar images are not compulsory. Motion synthesis works [19, 36] generate human motions in form of 3D human representations (e.g., SMPL and skeleton space), while our method generates human motions at the image level." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 644, + 403, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 644, + 403, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 644, + 403, + 657 + ], + "type": "text", + "content": "3. Text2 Performer" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "As shown in Fig. 3, our proposed Text2Performer synthesizes desired human videos by taking the texts as inputs " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\pmb{t}_a" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " for appearance and " + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": " for motions). To ensure consistent human representation, we propose to decompose the" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22749" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 289, + 207 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 289, + 207 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 289, + 207 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 289, + 207 + ], + "type": "image", + "image_path": "18d2a22c6a25d1438a3ae1f1bc680e2ec52eb9ca89b38847f73cfceafa6ab0c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 289, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 313 + ], + "type": "text", + "content": "Figure 4: Pipeline of Decomposed VQ-Space. Two different frames " + }, + { + "bbox": [ + 46, + 217, + 289, + 313 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 46, + 217, + 289, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 289, + 313 + ], + "type": "inline_equation", + "content": "I_k" + }, + { + "bbox": [ + 46, + 217, + 289, + 313 + ], + "type": "text", + "content": " of a video serve as the identity frame and the pose frame to provide appearance and pose features, respectively. We apply data augmentations to " + }, + { + "bbox": [ + 46, + 217, + 289, + 313 + ], + "type": "inline_equation", + "content": "I_k" + }, + { + "bbox": [ + 46, + 217, + 289, + 313 + ], + "type": "text", + "content": " to erase its appearance information to avoid information leakage. Two codebooks are built to store the pose features and appearance features. The quantized features are finally fed into the decoder to reconstruct the pose frame " + }, + { + "bbox": [ + 46, + 217, + 289, + 313 + ], + "type": "inline_equation", + "content": "I_k" + }, + { + "bbox": [ + 46, + 217, + 289, + 313 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 335, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 445 + ], + "type": "text", + "content": "VQ-space into appearance representation " + }, + { + "bbox": [ + 46, + 335, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\pmb{f}_a" + }, + { + "bbox": [ + 46, + 335, + 289, + 445 + ], + "type": "text", + "content": " and pose representation " + }, + { + "bbox": [ + 46, + 335, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\pmb{f}_p" + }, + { + "bbox": [ + 46, + 335, + 289, + 445 + ], + "type": "text", + "content": " as shown in Fig. 4. With the decomposed VQ-space, we sample the human appearance features " + }, + { + "bbox": [ + 46, + 335, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_a" + }, + { + "bbox": [ + 46, + 335, + 289, + 445 + ], + "type": "text", + "content": " and exemplar pose features " + }, + { + "bbox": [ + 46, + 335, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_p^0" + }, + { + "bbox": [ + 46, + 335, + 289, + 445 + ], + "type": "text", + "content": " according to " + }, + { + "bbox": [ + 46, + 335, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\pmb{t}_a" + }, + { + "bbox": [ + 46, + 335, + 289, + 445 + ], + "type": "text", + "content": ". To model the motion dynamics, we propose continuous VQ-diffuser as the motion sampler. It progressively diffuses the masked sequence of pose features until the whole sequence is un-masked. The final video clip is generated by feeding the appearance and pose embeddings into the decoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 454, + 180, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 180, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 180, + 467 + ], + "type": "text", + "content": "3.1. Decomposed VQ-Space" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 605 + ], + "type": "text", + "content": "Human video generation requires consistent human appearance, i.e., the face and clothing of the target person should remain unchanged while performing the motion. Vanilla VQVAE [52] encodes images into a unified feature representation and builds a codebook on it. However, such design is prone to generate drifted identities along the video frames. To overcome this problem, we propose to decompose the VQ-space into the appearance representation " + }, + { + "bbox": [ + 46, + 473, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{f}_a" + }, + { + "bbox": [ + 46, + 473, + 287, + 605 + ], + "type": "text", + "content": " and pose representation " + }, + { + "bbox": [ + 46, + 473, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{f}_p" + }, + { + "bbox": [ + 46, + 473, + 287, + 605 + ], + "type": "text", + "content": " in an unsupervised manner. With the decomposed VQ-space, we can separately sample " + }, + { + "bbox": [ + 46, + 473, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_a" + }, + { + "bbox": [ + 46, + 473, + 287, + 605 + ], + "type": "text", + "content": " and a sequence of " + }, + { + "bbox": [ + 46, + 473, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_p" + }, + { + "bbox": [ + 46, + 473, + 287, + 605 + ], + "type": "text", + "content": " to generate the desired video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "content": "In human video data, different frames in one video share the same human identity. We utilize this property to train the decomposed VQ-space [13, 25]. The pipeline is shown in Fig. 4. Given a video " + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{\\mathbf{I}_0, \\mathbf{I}_1, \\dots, \\mathbf{I}_n\\}" + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "content": ", we use its first frame " + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_0" + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "content": " for appearance information and another randomly sampled frame " + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k" + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "content": " for pose information. " + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_0" + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k" + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "content": " are fed into two encoders " + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "inline_equation", + "content": "E_a" + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "inline_equation", + "content": "E_p" + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "content": ", respectively. The two branches share the encoder " + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "inline_equation", + "content": "E_s" + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "content": ". To prevent " + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k" + }, + { + "bbox": [ + 46, + 606, + 288, + 715 + ], + "type": "text", + "content": " from leaking appearance information, data augmentations (e.g. color jitter" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 97 + ], + "type": "text", + "content": "tering and Gaussian blur) are applied to " + }, + { + "bbox": [ + 304, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k" + }, + { + "bbox": [ + 304, + 72, + 545, + 97 + ], + "type": "text", + "content": " before the pose branch. " + }, + { + "bbox": [ + 304, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 304, + 72, + 545, + 97 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 304, + 72, + 545, + 97 + ], + "type": "text", + "content": " are obtained as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 347, + 108, + 545, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 108, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 347, + 108, + 545, + 121 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {a} = E _ {a} \\left(E _ {s} \\left(\\mathbf {I} _ {0}\\right)\\right), \\boldsymbol {f} _ {p} = E _ {p} \\left(E _ {s} \\left(\\mathbf {I} _ {k} ^ {a}\\right)\\right)), \\tag {1}", + "image_path": "81f2f819d78fa38ba1ca34d0344ba5ddf22b08939c12ccf506a6bb5dfbfa4b0d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k^a" + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "text", + "content": " is the augmented " + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_k" + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "text", + "content": ". To make " + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "text", + "content": " learn compact and necessary pose information, we make the spatial size of " + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "text", + "content": " smaller. Given " + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "text", + "content": ", we empirically find that the optimal spatial sizes for " + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "inline_equation", + "content": "H / 16 \\times W / 16" + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "inline_equation", + "content": "H / 64 \\times W / 64" + }, + { + "bbox": [ + 304, + 132, + 547, + 192 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "text", + "content": "After obtaining " + }, + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\pmb{f}_a" + }, + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\pmb{f}_p" + }, + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "text", + "content": ", two codebooks " + }, + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_a" + }, + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_p" + }, + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "text", + "content": " are built to store the appearance and pose embeddings. The quantized feature " + }, + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}" + }, + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "text", + "content": " given codebook " + }, + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 194, + 547, + 230 + ], + "type": "text", + "content": " is obtained by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 240, + 545, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 545, + 261 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {f}} = \\operatorname {Q u a n t} (\\boldsymbol {f}) := \\left\\{\\underset {c _ {k} \\in \\mathcal {C}} {\\operatorname {a r g m i n}} \\| f _ {i j} - c _ {k} \\| _ {2} \\mid f _ {i j} \\in \\boldsymbol {f} \\right\\}. \\tag {2}", + "image_path": "244de60abb0e814926c081c569217e85392e4c5689966288a99e96580c77f858.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 274, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 300 + ], + "type": "text", + "content": "With the quantized " + }, + { + "bbox": [ + 304, + 274, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\hat{f}_a" + }, + { + "bbox": [ + 304, + 274, + 545, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 274, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\hat{f}_p" + }, + { + "bbox": [ + 304, + 274, + 545, + 300 + ], + "type": "text", + "content": ", we then feed them into decoder " + }, + { + "bbox": [ + 304, + 274, + 545, + 300 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 274, + 545, + 300 + ], + "type": "text", + "content": " to reconstruct the target pose frame " + }, + { + "bbox": [ + 304, + 274, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_k" + }, + { + "bbox": [ + 304, + 274, + 545, + 300 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 377, + 311, + 545, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 311, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 377, + 311, + 545, + 326 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {I}} _ {k} = D ([ \\hat {\\boldsymbol {f}} _ {a}, D _ {p} (\\hat {\\boldsymbol {f}} _ {p}) ]), \\tag {3}", + "image_path": "677732a26ffbac756b57829e88dad88f0d37e79fc89277be869f944f2f72e5e0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 336, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 386 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 336, + 545, + 386 + ], + "type": "inline_equation", + "content": "[\\cdot ]" + }, + { + "bbox": [ + 304, + 336, + 545, + 386 + ], + "type": "text", + "content": " is the concatenation operation, and " + }, + { + "bbox": [ + 304, + 336, + 545, + 386 + ], + "type": "inline_equation", + "content": "D_{p}" + }, + { + "bbox": [ + 304, + 336, + 545, + 386 + ], + "type": "text", + "content": " upsamples " + }, + { + "bbox": [ + 304, + 336, + 545, + 386 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_p" + }, + { + "bbox": [ + 304, + 336, + 545, + 386 + ], + "type": "text", + "content": " to make it have the same resolution as " + }, + { + "bbox": [ + 304, + 336, + 545, + 386 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_a" + }, + { + "bbox": [ + 304, + 336, + 545, + 386 + ], + "type": "text", + "content": ". The whole network (including the encoders, decoders, and codebooks) is trained using:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 396, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 396, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 311, + 396, + 545, + 456 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\left\\| \\mathbf {I} _ {k} - \\hat {\\mathbf {I}} _ {k} \\right\\| _ {1} + \\left\\| \\operatorname {s g} (\\hat {\\boldsymbol {f}} _ {a}) - \\boldsymbol {f} _ {a} \\right\\| _ {2} ^ {2} + \\left\\| \\operatorname {s g} (\\boldsymbol {f} _ {a}) - \\hat {\\boldsymbol {f}} _ {a} \\right\\| _ {2} ^ {2} \\\\ + \\left\\| \\operatorname {s g} \\left(\\hat {\\boldsymbol {f}} _ {p}\\right) - \\boldsymbol {f} _ {p} \\right\\| _ {2} ^ {2} + \\left\\| \\operatorname {s g} \\left(\\boldsymbol {f} _ {p}\\right) - \\hat {\\boldsymbol {f}} _ {p} \\right\\| _ {2} ^ {2}, \\tag {4} \\\\ \\end{array}", + "image_path": "ecb6b484b2e5ec1c9632030bf2a6f3b585ad6f1ef51d61acdcd8d870bc63e0c3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 468, + 476, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 476, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 476, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 468, + 476, + 480 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}(\\cdot)" + }, + { + "bbox": [ + 306, + 468, + 476, + 480 + ], + "type": "text", + "content": " is the stop-gradient operation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 481, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 517 + ], + "type": "text", + "content": "With the decomposed VQ-space, an additional sampler is trained to sample " + }, + { + "bbox": [ + 304, + 481, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_a" + }, + { + "bbox": [ + 304, + 481, + 545, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 481, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_p^0" + }, + { + "bbox": [ + 304, + 481, + 545, + 517 + ], + "type": "text", + "content": ". This sampler has the same design as previous methods [4, 27]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 528, + 444, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 444, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 444, + 540 + ], + "type": "text", + "content": "3.2. Continuous VQ-Diffuser" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 548, + 545, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 645 + ], + "type": "text", + "content": "We adopt the absorbing diffusion transformer [2, 4, 8, 18] to sample the motion, a sequence of pose embeddings " + }, + { + "bbox": [ + 304, + 548, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_p = \\{\\hat{\\pmb{f}}_p^1,\\hat{\\pmb{f}}_p^2,\\dots,\\hat{\\pmb{f}}_p^n\\}" + }, + { + "bbox": [ + 304, + 548, + 545, + 645 + ], + "type": "text", + "content": " from the learned pose codebook " + }, + { + "bbox": [ + 304, + 548, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_p" + }, + { + "bbox": [ + 304, + 548, + 545, + 645 + ], + "type": "text", + "content": " Different from autoregressive models [11, 52] that make predictions in a fixed order, absorbing diffusion transformer predicts multiple codebook indices in parallel. The prediction of codebook indices starts from " + }, + { + "bbox": [ + 304, + 548, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p^0" + }, + { + "bbox": [ + 304, + 548, + 545, + 645 + ], + "type": "text", + "content": " i.e., fully masked " + }, + { + "bbox": [ + 304, + 548, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p^m" + }, + { + "bbox": [ + 304, + 548, + 545, + 645 + ], + "type": "text", + "content": " . The prediction at time step " + }, + { + "bbox": [ + 304, + 548, + 545, + 645 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 548, + 545, + 645 + ], + "type": "text", + "content": " is represented as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 383, + 657, + 545, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 657, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 383, + 657, + 545, + 673 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} _ {p} ^ {t} \\sim q _ {\\theta} \\left(\\mathbf {F} _ {p} ^ {t} \\mid \\mathbf {F} _ {p} ^ {t - 1}\\right), \\tag {5}", + "image_path": "bd211f0d4ca32c524730ae5a391f653f59e8302bd303b1b9ecd2de8842bf005c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 681, + 544, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 681, + 544, + 692 + ], + "spans": [ + { + "bbox": [ + 306, + 681, + 544, + 692 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 681, + 544, + 692 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 306, + 681, + 544, + 692 + ], + "type": "text", + "content": " denotes the parameters of the transformer sampler." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 693, + 545, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 718 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 718 + ], + "type": "text", + "content": "To model human motions, we propose 1) continuous space sampling, and 2) motion-aware masking strategy." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22750" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 72, + 281, + 194 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 281, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 281, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 281, + 194 + ], + "type": "image", + "image_path": "e8efb86c90da220bbeed4ccc74257ceae82e6ca45e18df6a2ad5e09055a4da72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 289, + 276 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 289, + 276 + ], + "type": "text", + "content": "Figure 5: Comparison with Image Samplers. (a) Discrete VQ Sampler predicts discrete codebook indices, which are trained by cross-entropy loss. (b) Continuous Sampler operates on unconstrained continuous RGB spaces. (c) Our VQ-Diffuser predicts continuous embeddings but also utilizes rich embeddings stored in codebooks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 297, + 201, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 201, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 201, + 309 + ], + "type": "text", + "content": "3.2.1 Continuous Space Sampling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 411 + ], + "type": "text", + "content": "As shown in Fig. 5(a), previous VQVAE-based methods [4, 8, 11, 18, 52] sample tokens in a discrete space. They predict codebook indices to form quantized features, which are then fed into the decoder to generate images. However, sampling at the discrete space is hard to converge because of redundant codebooks, which causes the one-to-many mapping prediction problem. This would lead to meaningless predicted pose sequences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "text", + "content": "In continuous VQ-diffuser " + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "text", + "content": ", we propose to train the non-causal transformer in the continuous embedding space as shown in Fig. 5(c). " + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "text", + "content": " directly predicts continuous pose embeddings " + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}_p^k\\}" + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "text", + "content": " is trained to predict the full continuous " + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_p" + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "text", + "content": " from the masked pose sequences " + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p^m" + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "text", + "content": " conditioned on the motion description " + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{m}" + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "text", + "content": " and the initial pose feature " + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_p^0" + }, + { + "bbox": [ + 46, + 411, + 287, + 489 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 498, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 498, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 108, + 498, + 287, + 514 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {F}} _ {p} = S _ {\\theta} ([ T (\\boldsymbol {t} _ {m}), \\hat {\\boldsymbol {f}} _ {p} ^ {0}, \\mathbf {F} _ {p} ^ {m} ]), \\tag {6}", + "image_path": "67e1d8ba1dddfec57c25be5703324bd310f9fdcfaf27011a7a7905d6c364fc58.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 521, + 249, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 249, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 249, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 521, + 249, + 533 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 47, + 521, + 249, + 533 + ], + "type": "text", + "content": " is the pretrained text feature extractor." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 571 + ], + "type": "text", + "content": "To utilize rich embeddings stored in " + }, + { + "bbox": [ + 46, + 533, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_p" + }, + { + "bbox": [ + 46, + 533, + 287, + 571 + ], + "type": "text", + "content": " to constrain the prediction space, we retrieve the nearest embedding of the predicted continuous " + }, + { + "bbox": [ + 46, + 533, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_p" + }, + { + "bbox": [ + 46, + 533, + 287, + 571 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 533, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_p" + }, + { + "bbox": [ + 46, + 533, + 287, + 571 + ], + "type": "text", + "content": " to obtain the final " + }, + { + "bbox": [ + 46, + 533, + 287, + 571 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_p" + }, + { + "bbox": [ + 46, + 533, + 287, + 571 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 73, + 578, + 287, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 578, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 73, + 578, + 287, + 602 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} _ {p} = \\operatorname {N e a r e s t} \\left(\\bar {\\mathbf {F}} _ {p}\\right) := \\left\\{\\underset {c _ {k} \\in \\mathcal {C} _ {p}} {\\operatorname {a r g m i n}} \\left\\| \\bar {f} _ {p} - c _ {k} \\right\\| _ {2} \\right\\}. \\tag {7}", + "image_path": "fa81668aff1d8b75e9cb41792ad3b5c829ab5282478b4f4eb52a6d4f3e74d6bf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 612, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 612, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 612, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_p" + }, + { + "bbox": [ + 46, + 612, + 287, + 661 + ], + "type": "text", + "content": " is then fed into " + }, + { + "bbox": [ + 46, + 612, + 287, + 661 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 612, + 287, + 661 + ], + "type": "text", + "content": " to reconstruct the final video " + }, + { + "bbox": [ + 46, + 612, + 287, + 661 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}" + }, + { + "bbox": [ + 46, + 612, + 287, + 661 + ], + "type": "text", + "content": " following Eq. (3). Thanks to the continuous operation, we can add losses at both the image level and embedding level. The loss function to train our continuous VQ-diffuser is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 669, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 669, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 89, + 669, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\left\\| \\bar {\\mathbf {F}} _ {p} - \\mathbf {F} _ {p} \\right\\| _ {1} + \\left\\| \\operatorname {s g} \\left(\\hat {\\mathbf {F}} _ {p}\\right) - \\bar {\\mathbf {F}} _ {p} \\right\\| _ {2} ^ {2} \\tag {8} \\\\ + \\left\\| \\mathrm {s g} (\\bar {\\mathbf {F}} _ {p}) - \\hat {\\mathbf {F}} _ {p} \\right\\| _ {2} ^ {2} + L _ {\\mathrm {r e c}} (\\hat {\\mathbf {V}}, \\mathbf {V}), \\\\ \\end{array}", + "image_path": "cc42c1382f278e562417365fffe73d6882cc1c100ee5736416e8de8b9fcfb815.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{rec}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": " is composed of " + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": " loss and perceptual loss [28]. It should be noted that stop-gradient operation is applied to make the nearest operation differentiable." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 109, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 547, + 205 + ], + "type": "text", + "content": "Compared to continuous samplers like PixelCNN [51] in Fig. 5(b), which directly predict the continuous RGB values, our continuous VQ-diffuser fully utilizes the rich contents stored in the codebook while restricting the prediction space. Therefore, it inherits the benefits of the discrete VQ sampler and continuous sampler at the same time. The pose sequences sampled by our continuous VQ-diffuser are temporally coherent and meaningful." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 224, + 479, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 479, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 479, + 237 + ], + "type": "text", + "content": "3.2.2 Motion-Aware Masking Strategy" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "type": "text", + "content": "To generate plausible videos, the sampled motion sequence " + }, + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_p" + }, + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "type": "text", + "content": " should be both temporally and spatially reasonable. In order to make the continuous VQ-diffuser " + }, + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "type": "text", + "content": " correctly condition on " + }, + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{m}" + }, + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "type": "text", + "content": " as well as generate reasonable human poses for each frame, we design a motion-aware masking strategy to meticulously mask " + }, + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p" + }, + { + "bbox": [ + 304, + 245, + 545, + 317 + ], + "type": "text", + "content": " temporally and spatially." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "text", + "content": "At the temporal dimension, " + }, + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "text", + "content": " first predicts pose embeddings of the first frame and the last frame based on " + }, + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}_p^0" + }, + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "text", + "content": ". Then the prediction diffuses to the intermediate frames according to the given conditions and previously unmasked frames. Therefore, during the training, if the first frame and the last frame are masked, we will mask all frames to prevent the intermediate frames from providing information to help frame predictions at two ends. A higher probability is assigned to masking all frames to help " + }, + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "inline_equation", + "content": "S_{\\theta}" + }, + { + "bbox": [ + 304, + 318, + 546, + 437 + ], + "type": "text", + "content": " better learn prediction at the most challenging two ends." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 438, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 498 + ], + "type": "text", + "content": "At the spatial dimension, we propose to predict the first and the last frames with more than one diffusion step. Otherwise, the structure of the sampled human tends to be incomplete. To this end, during the training, we mask the first and last frames partially at the spatial dimension." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 512, + 465, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 512, + 465, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 512, + 465, + 525 + ], + "type": "text", + "content": "4. Fashion-Text2Video Dataset" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": "Existing human video datasets are either of low resolution [3, 17, 49] or lack of text (action labels) [35, 43]. To support the study of text-driven human video generation, we propose a new dataset, Fashion-Text2Video dataset. Among existing human video datasets, we select videos in Fashion Dataset [56] as the source data for following reasons: 1) High resolution (" + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\geq 1024 \\times 1024" + }, + { + "bbox": [ + 304, + 533, + 545, + 617 + ], + "type": "text", + "content": "); 2) Diverse appearance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": "The Fashion Dataset contains a total of 600 videos. For each video, we manually annotate motion labels, clothing shapes (length of sleeves and length of clothing) and textures. Motion labels are classified into 22 classes, including standing, moving right-side, moving left-side, turning around, etc. The lengths of sleeves are classified into no-sleeve, three-point, medium and long sleeves. The length of clothing is the length of the upper part of the clothes." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22751" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 76, + 300, + 169 + ], + "blocks": [ + { + "bbox": [ + 75, + 111, + 85, + 141 + ], + "lines": [ + { + "bbox": [ + 75, + 111, + 85, + 141 + ], + "spans": [ + { + "bbox": [ + 75, + 111, + 85, + 141 + ], + "type": "text", + "content": "D" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 76, + 300, + 169 + ], + "lines": [ + { + "bbox": [ + 88, + 76, + 300, + 169 + ], + "spans": [ + { + "bbox": [ + 88, + 76, + 300, + 169 + ], + "type": "image", + "image_path": "99167ea0ac00e7f42b045d427db1ab51b51e6955def74f448a20ec5efa2e4307.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 302, + 75, + 519, + 169 + ], + "blocks": [ + { + "bbox": [ + 302, + 75, + 519, + 169 + ], + "lines": [ + { + "bbox": [ + 302, + 75, + 519, + 169 + ], + "spans": [ + { + "bbox": [ + 302, + 75, + 519, + 169 + ], + "type": "image", + "image_path": "0174ab8dd49f5ae04bd0c732244d3d00b335d51989ec143cbbb10a8df3e10b79.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 88, + 183, + 300, + 278 + ], + "blocks": [ + { + "bbox": [ + 75, + 213, + 86, + 257 + ], + "lines": [ + { + "bbox": [ + 75, + 213, + 86, + 257 + ], + "spans": [ + { + "bbox": [ + 75, + 213, + 86, + 257 + ], + "type": "text", + "content": "eannnnnne" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 183, + 300, + 278 + ], + "lines": [ + { + "bbox": [ + 88, + 183, + 300, + 278 + ], + "spans": [ + { + "bbox": [ + 88, + 183, + 300, + 278 + ], + "type": "image", + "image_path": "037189fa832080d58a81d45de27ac184010dcf8059096744f9e8640b80af74e5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 302, + 183, + 518, + 278 + ], + "blocks": [ + { + "bbox": [ + 302, + 183, + 518, + 278 + ], + "lines": [ + { + "bbox": [ + 302, + 183, + 518, + 278 + ], + "spans": [ + { + "bbox": [ + 302, + 183, + 518, + 278 + ], + "type": "image", + "image_path": "464395c7197fc7442f2a6b9960c9d2620ad976ba1838c9a3d011b879d106368b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 88, + 292, + 300, + 385 + ], + "blocks": [ + { + "bbox": [ + 76, + 315, + 86, + 369 + ], + "lines": [ + { + "bbox": [ + 76, + 315, + 86, + 369 + ], + "spans": [ + { + "bbox": [ + 76, + 315, + 86, + 369 + ], + "type": "text", + "content": "Teeerree" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 292, + 300, + 385 + ], + "lines": [ + { + "bbox": [ + 88, + 292, + 300, + 385 + ], + "spans": [ + { + "bbox": [ + 88, + 292, + 300, + 385 + ], + "type": "image", + "image_path": "b427eb8988c41894caf93822b06a571db34ca69249cb041e011bd1f93f5b5eca.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 399, + 291, + 429 + ], + "lines": [ + { + "bbox": [ + 104, + 399, + 291, + 429 + ], + "spans": [ + { + "bbox": [ + 104, + 399, + 291, + 429 + ], + "type": "text", + "content": "This female is wearing a dress, with graphic pattern. It has long sleeves and it is of three-point length. This female turns right from the side to the front." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 302, + 292, + 518, + 385 + ], + "blocks": [ + { + "bbox": [ + 302, + 292, + 518, + 385 + ], + "lines": [ + { + "bbox": [ + 302, + 292, + 518, + 385 + ], + "spans": [ + { + "bbox": [ + 302, + 292, + 518, + 385 + ], + "type": "image", + "image_path": "28428872fb9b3e08cbf8fffd349a9456444ceb5e8bd5067a78a3943a41577bbe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 399, + 511, + 430 + ], + "lines": [ + { + "bbox": [ + 331, + 399, + 511, + 430 + ], + "spans": [ + { + "bbox": [ + 331, + 399, + 511, + 430 + ], + "type": "text", + "content": "The dress this person wears has long-sleeve and it is of three-point length. The texture of it is graphic. The lady turns right from the back to the side." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 81, + 402, + 99, + 422 + ], + "blocks": [ + { + "bbox": [ + 81, + 402, + 99, + 422 + ], + "lines": [ + { + "bbox": [ + 81, + 402, + 99, + 422 + ], + "spans": [ + { + "bbox": [ + 81, + 402, + 99, + 422 + ], + "type": "image", + "image_path": "f88311fa6a7a84f8247165d3f5c3ebbcc261bedec3f9f3e6e891d53fae3f27f9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 439, + 533, + 452 + ], + "lines": [ + { + "bbox": [ + 59, + 439, + 533, + 452 + ], + "spans": [ + { + "bbox": [ + 59, + 439, + 533, + 452 + ], + "type": "text", + "content": "Figure 6: Qualitative Comparisons. Text2Performer achieves superior generation qualities compared with baselines." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 402, + 326, + 422 + ], + "blocks": [ + { + "bbox": [ + 307, + 402, + 326, + 422 + ], + "lines": [ + { + "bbox": [ + 307, + 402, + 326, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 402, + 326, + 422 + ], + "type": "image", + "image_path": "038ae4e547bb16aae03d74152db51a30401e081b77a6e3f7495c678fffc4d9eb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 472, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 557 + ], + "type": "text", + "content": "Clothing textures are mainly divided into pure color, floral, graphic, etc. Upon the above labels, we then generate texts for each video using some pre-defined templates. The texts include descriptions of the appearances of the dressed clothes and descriptions for each sub-motions. The proposed Fashion-Text2Video dataset can be applied to the research on Text2Video and Label2Video generation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 567, + 128, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 581 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 586, + 171, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 171, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 171, + 600 + ], + "type": "text", + "content": "5.1. Comparison Methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "MMVID [20] is a VQVAE-based method for multimodal video synthesis. We follow the original setting to train MMVID on our dataset. StyleGAN-V [46] is a state-of-the-art video generation method built on StyleGAN. Action labels are used as conditions instead of texts, which we find significantly surpasses the performance of texts for this method. CogVideo [24] is a large-scale pretrained model for Text-to-Video generation. We implement two versions of CogVideo, i.e., CogVideo-v1 and CogVideo-v2. Since" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 472, + 547, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 534 + ], + "type": "text", + "content": "the original CogVideo utilized the pretrained Text2Image dataset, we train CogVideo-v1 to use the exemplar image as an additional input to compensate for the lack of large-scale pretrained Text2Image models for human images. CogVideo-v2 is with original designs." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 540, + 419, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 419, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 419, + 552 + ], + "type": "text", + "content": "5.2. Evaluation Metrics" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 630 + ], + "type": "text", + "content": "Diversity and Quality. 1) We adopt Fréchet Inception Distance (FID) [21] to evaluate the quality of generated human frames. 2) We evaluate the quality of generated videos using Fréchet Video Distance (FVD) and Kernel Video Distance (KVD) [50]. We generate 2,048 video clips to compute the metrics, with 20 frames for each clip." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "text", + "content": "Identity Preservation. We use two metrics to evaluate the maintenance of identity along the generated frames. 1) We extract face features of each generated frames using FaceNet [42]. For each generated video clip, we calculate the " + }, + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 630, + 546, + 714 + ], + "type": "text", + "content": " distances between features of the frames and the first frame. 2) We extract features of each generated frames using an off-the-shelf ReID model, i.e., OSNet [42]. These" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22752" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 84, + 294, + 146 + ], + "blocks": [ + { + "bbox": [ + 90, + 71, + 243, + 82 + ], + "lines": [ + { + "bbox": [ + 90, + 71, + 243, + 82 + ], + "spans": [ + { + "bbox": [ + 90, + 71, + 243, + 82 + ], + "type": "text", + "content": "Table 1: Quantitative Comparisons." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 84, + 294, + 146 + ], + "lines": [ + { + "bbox": [ + 48, + 84, + 294, + 146 + ], + "spans": [ + { + "bbox": [ + 48, + 84, + 294, + 146 + ], + "type": "table", + "html": "
MethodFID ↓FVD ↓KVD ↓Face ↓ReLU ↑
MMVID [20]11.85303.0278.670.90470.9096
StyleGAN-V [46]29.68219.6318.770.86750.9568
CogVideo-v1 [24]39.47645.0389.531.05640.8148
CogVideo-v2 [24]51.76799.80112.241.06210.7960
Text2Performer9.60124.7817.960.85930.9382
", + "image_path": "9ce8254c3a7eadc5fcdae678c1011e465d711adad0ac2d2efe72a2faab3b66a8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 159, + 282, + 243 + ], + "blocks": [ + { + "bbox": [ + 50, + 159, + 282, + 243 + ], + "lines": [ + { + "bbox": [ + 50, + 159, + 282, + 243 + ], + "spans": [ + { + "bbox": [ + 50, + 159, + 282, + 243 + ], + "type": "image", + "image_path": "fc611437ae5bf137664f8200d8583da06ab4cbe371b7d358e821a21c6a2f07e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 246, + 274, + 258 + ], + "lines": [ + { + "bbox": [ + 58, + 246, + 274, + 258 + ], + "spans": [ + { + "bbox": [ + 58, + 246, + 274, + 258 + ], + "type": "text", + "content": "Figure 7: User Study. We achieve the highest scores." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 278, + 286, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 286, + 326 + ], + "type": "text", + "content": "features are related to identity features. We calculate the average cosine distances between extracted features of the frames and the first frame. The final metrics are obtained by averaging distances among 2,048 generated video clips." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 327, + 286, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 286, + 422 + ], + "type": "text", + "content": "User Study. Users are asked to give three separate scores (the highest score is 4) at three dimensions: 1) The consistency with the text for appearance, 2) The consistency with the text for desired motion, and 3) The overall quality of the generated video in terms of temporal coherence and its realism. A total of 20 users participate in the user study. Each user is presented with 30 videos generated by different methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 430, + 187, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 187, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 187, + 443 + ], + "type": "text", + "content": "5.3. Qualitative Comparisons" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 448, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 556 + ], + "type": "text", + "content": "We show two examples in Fig. 6. In the first example, MMVID and Text2Performer successfully generate desired human motions. However, frames generated by MMVID have drifted identities. StyleGAN-V fails to generate plausible frames as well as consistent motions. As for the second example, only Text2Performer generates corresponding motions. Artifacts appear in the video clips generated by MMVID and StyleGAN-V. The visual comparisons demonstrate the superiority of Text2Performer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 563, + 194, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 194, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 194, + 576 + ], + "type": "text", + "content": "5.4. Quantitative Comparisons" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "The quantitative comparisons are shown in Table 1. In terms of FID, FVD and KVD, our proposed Text2Performer has significant improvements over baselines, which demonstrates the diversity and temporal coherence of the videos generated by our methods. Existing baselines generate videos from entangled human representations and thus suffer from temporal incoherence. As for the identity metrics, our method achieves the best performance on face distance and comparable performance on ReID distance with StyleGAN-V. Since StyleGAN-V is prone to generate videos with small motions, features tend to be similar," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 543, + 235 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 543, + 235 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 543, + 235 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 543, + 235 + ], + "type": "image", + "image_path": "4991f26c8fd6fae6c84895b0df4c4ccb01be5d1e75e35f61b4fdbfc0c75a7253.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 244, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 305, + 244, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 545, + 281 + ], + "type": "text", + "content": "Figure 8: Ablation Studies on (a) Decomposed VQ-Space, (b) Spatial Resolution of Pose Branch, (c) Discrete VQ sampler, and (d) Usage of Codebook." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": "leading to high ReID scores. In our method, thanks to the decomposed VQ-space and continuous VQ-diffuser, temporal and identity consistency can be better achieved. The result of the user study is shown in Fig. 7. Text2Performer achieves the highest scores on all three dimensions. Specifically, the result of the assessment on overall quality is consistent with the other quantitative metrics, which further validates the effectiveness of Text2Performer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 407, + 400, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 400, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 400, + 419 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 426, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 556 + ], + "type": "text", + "content": "Decomposed VQ-Space. The ablation model trains the continuous VQ-diffuser with a unified VQ-space, which fuses the identity and pose information within one feature. As shown in Fig. 8(a), compared to our results, sampling in the unified VQ-Space generates drifted identity and incomplete body structures. This leads to inferior quantitative metrics in Table 2. The unified VQ-Space requires the sampler to handle identity and pose at the same time, which poses burdens for the sampler. With the decomposed VQ-Space, the identity can be easily maintained by fixing the appearance features, which further helps learn motions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "type": "text", + "content": "Spatial Resolution of Pose Branch. In our design, the pose feature has " + }, + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 304, + 558, + 545, + 676 + ], + "type": "text", + "content": " smaller spatial resolution than the appearance feature. In Fig. 8(b), we show the results generated by a decomposed VQVAE where pose and appearance features have the same resolution. Compared to results of the full model, body structures in the side view are incomplete. The quantitative metrics in Table 2 and Fig. 8(b) demonstrate that the smaller pose feature is necessary as the larger feature contains redundant information to harm valid feature disentanglement, thus adding challenges to the sampler." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Discrete VQ sampler in Decomposed VQ-Space. We train a variant of VQ-Diffuser to predict discrete codebook indices with the decomposed VQ-Space. In Fig. 8(c), the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22753" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 67, + 521, + 221 + ], + "blocks": [ + { + "bbox": [ + 72, + 67, + 521, + 221 + ], + "lines": [ + { + "bbox": [ + 72, + 67, + 521, + 221 + ], + "spans": [ + { + "bbox": [ + 72, + 67, + 521, + 221 + ], + "type": "image", + "image_path": "2a0daad2d1921ccb971194ab1f31de5c234de4837d77777e9bbae5af3d9e3f62.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 224, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 47, + 224, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 545, + 237 + ], + "type": "text", + "content": "Figure 9: High-Resolution Results. We show one result generated by a long text sequence. The identity is well maintained." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 73, + 247, + 235, + 354 + ], + "blocks": [ + { + "bbox": [ + 73, + 247, + 235, + 354 + ], + "lines": [ + { + "bbox": [ + 73, + 247, + 235, + 354 + ], + "spans": [ + { + "bbox": [ + 73, + 247, + 235, + 354 + ], + "type": "image", + "image_path": "00b29ec0f0117bcf53c4c297361a9f5801c1bb434f26af13320908fb3f53b79e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 356, + 537, + 369 + ], + "lines": [ + { + "bbox": [ + 54, + 356, + 537, + 369 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 537, + 369 + ], + "type": "text", + "content": "Figure 10: Cross-Dataset Performance. We compare with models trained on large-scale data and domain-specific data." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 236, + 247, + 375, + 354 + ], + "blocks": [ + { + "bbox": [ + 236, + 247, + 375, + 354 + ], + "lines": [ + { + "bbox": [ + 236, + 247, + 375, + 354 + ], + "spans": [ + { + "bbox": [ + 236, + 247, + 375, + 354 + ], + "type": "image", + "image_path": "788fb8a095a8fb5514ee54804ee9f828d6f40acb660717fa7cdd60922b1df37d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 379, + 247, + 519, + 354 + ], + "blocks": [ + { + "bbox": [ + 379, + 247, + 519, + 354 + ], + "lines": [ + { + "bbox": [ + 379, + 247, + 519, + 354 + ], + "spans": [ + { + "bbox": [ + 379, + 247, + 519, + 354 + ], + "type": "image", + "image_path": "3dd2f4f78edbfed8b4ebce517b971cfa2fdd37f206bd81970d3a3200a9e91328.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 403, + 282, + 466 + ], + "blocks": [ + { + "bbox": [ + 59, + 388, + 274, + 399 + ], + "lines": [ + { + "bbox": [ + 59, + 388, + 274, + 399 + ], + "spans": [ + { + "bbox": [ + 59, + 388, + 274, + 399 + ], + "type": "text", + "content": "Table 2: Quantitative Results on Ablation Studies." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 403, + 282, + 466 + ], + "lines": [ + { + "bbox": [ + 48, + 403, + 282, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 282, + 466 + ], + "type": "table", + "html": "
MethodFID ↓FVD ↓KVD ↓Face ↓ReID ↑
w/o decomp.20.52266.9735.520.89360.9213
same res.15.69202.0632.960.86490.9399
disc. sampler10.37308.8873.170.91060.9161
w/o codebook11.31154.9322.800.94370.9109
Full Model10.92134.4320.310.85200.9421
", + "image_path": "a41f40d7a3b33d028b696698e3954c264bb037803f463a67020f3f1f41383c60.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 486, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 617 + ], + "type": "text", + "content": "discrete VQ Sampler can generate complete body structures but fails to generate video clips consistent with the given action control, i.e., turning right from the back to the side. Failure on conditioning on action controls does not harm the quality of each generated frame, and thus this ablation study achieves a comparable FID score as our method. However, our method significantly outperforms this variant on other metrics as shown in Table 2. In VQVAE, some codebook entries have similar contents but with different codebook indices. Therefore, treating the training of the sampler as a classification task makes the model hard to converge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Usage of Codebook in Continuous VQ-Space. In our design, after the prediction of embeddings, the nearest neighbor of embeddings will be retrieved from the codebook. We train an auto-encoder without codebook in the pose branch. The sampler is trained to sample continuous embeddings from the latent space of auto-encoder without the retrieval from codebook. As shown in Fig. 8(d), without the codebook, the sampler results in a disrupted human image at" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 389, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 437 + ], + "type": "text", + "content": "the side view. The quality of generated videos deteriorates as reflected in Table 2. This demonstrates that predicting compact and meaningful continuous embeddings with codebooks enhances the quality of synthesized frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 450, + 421, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 450, + 421, + 463 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 421, + 463 + ], + "type": "text", + "content": "5.6. Additional Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 471, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 519 + ], + "type": "text", + "content": "High-Resolution Results. Text2Performer can generate long videos with high resolutions (512 × 256) as shown in Fig. 1 and Fig. 9. Figure 9 shows an example of high-resolution videos for a long text sequence." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 520, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 640 + ], + "type": "text", + "content": "Cross-Dataset Performance. We further verify the performance of models trained on different datasets. The results are presented in Fig. 10. Compared with CogVideo [24] pretrained on large-scale dataset, our Text2Performer and MMVID succeed to generate plausible specific human content. As for models trained on small-scale domain-specific data, our Text2Performer is comparable to MMVID on the simple iPer dataset [35], and has superior performance on the more challenging Fashion-Text2Video dataset, especially on the completeness of body structures across frames." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Ability to Generate Novel Appearance. To verify that the generated humans are with novel appearances, we retrieve top-2 nearest neighbour images of generated frames using the perceptual distance [28]. As shown in Fig. 11, the generated query frames have different appearances from the top-2 nearest images retrieved from the dataset." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22754" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 72, + 165, + 149 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 149 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 149 + ], + "type": "image", + "image_path": "1da77b9d9e51b4dc40eebfafd041f852e91d73c50992015c987306cd6e814b42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 150, + 282, + 159 + ], + "lines": [ + { + "bbox": [ + 53, + 150, + 282, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 282, + 159 + ], + "type": "text", + "content": "Query Image Top-2 Nearest Images Query Image Top-2 Nearest Images" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 72, + 280, + 149 + ], + "blocks": [ + { + "bbox": [ + 170, + 72, + 280, + 149 + ], + "lines": [ + { + "bbox": [ + 170, + 72, + 280, + 149 + ], + "spans": [ + { + "bbox": [ + 170, + 72, + 280, + 149 + ], + "type": "image", + "image_path": "dd84e38c790c40f4ae63428be22bba459e273829e5dc3fa5a05c2010967ba48c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 170, + 150, + 282, + 159 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 150, + 282, + 159 + ], + "spans": [ + { + "bbox": [ + 170, + 150, + 282, + 159 + ], + "type": "text", + "content": "Query Image Top-2 Nearest Images" + } + ] + } + ], + "index": 3, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 209, + 286, + 236 + ], + "blocks": [ + { + "bbox": [ + 47, + 169, + 287, + 194 + ], + "lines": [ + { + "bbox": [ + 47, + 169, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 169, + 287, + 194 + ], + "type": "text", + "content": "Figure 11: Nearest Neighbour Image Search. Top-2 nearest images of generated images are retrieved." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 209, + 286, + 236 + ], + "lines": [ + { + "bbox": [ + 50, + 209, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 50, + 209, + 286, + 236 + ], + "type": "image", + "image_path": "2f2efce4631626861f257dada972706ce0ef378b11a0fdceead142717aa7fb87.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 246, + 287, + 281 + ], + "lines": [ + { + "bbox": [ + 47, + 246, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 287, + 281 + ], + "type": "text", + "content": "Figure 12: Results on In-the-Wild Dataset. In our current pipeline, We treat the background as a part of appearance information. The results contain some artifacts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 305, + 120, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 305, + 120, + 317 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 120, + 317 + ], + "type": "text", + "content": "6. Discussions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 327, + 287, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 398 + ], + "type": "text", + "content": "We propose a novel Text2Performer for text-driven human video generation. We decompose the VQ-space into appearance and pose representations. Continuous VQ-diffuser is then introduced to sample motions conditioned on texts. These designs empower Text2Performer to synthesize high-quality and high-resolution human videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 399, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 638 + ], + "type": "text", + "content": "Limitations and Future Work. In this paper, we focus on the generation of human appearance and motion with clean backgrounds. The current pipeline is designed for generating human videos without textural backgrounds. It can be extended to generate plausible textual backgrounds by treating the background as a part of appearance information at the training time. The generated results are shown in Fig. 12. There are some artifacts in the generated videos. Generating videos with textual backgrounds is important in real-world applications. In future work, textual backgrounds can be segmented first. The results can be then improved by generating the foreground and textual background separately and fusing them harmoniously using an additional module. In addition, the synthesized human videos are biased toward generating females with dresses. This is because the Fashion Dataset only contains videos of females with dresses. In future applications, more data can be involved in the training to alleviate the bias. Besides, the control modality can also be extended to various types, such as audio and music." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 640, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 640, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 640, + 287, + 688 + ], + "type": "text", + "content": "Potential Negative Societal Impacts. The model may be applied to generate fake videos performing various actions. To alleviate the negative impacts, DeepFake detection methods can be applied to evaluate the realism of videos." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Acknowledgement. This study is supported under the RIE2020 Industry Alignment Fund - Industry Collabora" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "tion Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). This study is also supported by NTU NAP, MOE AcRF Tier 1 (2021-T1-001-088), and the Google PhD Fellowship." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 133, + 365, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 133, + 365, + 145 + ], + "spans": [ + { + "bbox": [ + 307, + 133, + 365, + 145 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 154, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 311, + 154, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 154, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 311, + 154, + 545, + 208 + ], + "type": "text", + "content": "[1] Badour Albahar, Jingwan Lu, Jimei Yang, Zhixin Shu, Eli Shechtman, and Jia-Bin Huang. Pose with style: Detail-preserving pose-guided image synthesis with conditional stylegan. ACM Transactions on Graphics (TOG), 40(6):1-11, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 210, + 547, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 210, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 311, + 210, + 547, + 253 + ], + "type": "text", + "content": "[2] Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces. NeurIPS, 34:17981-17993, 2021. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 256, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 256, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 311, + 256, + 545, + 289 + ], + "type": "text", + "content": "[3] Guha Balakrishnan, Amy Zhao, Adrian V Dalca, Fredo Durand, and John Guttag. Synthesizing images of humans in unseen poses. In CVPR, pages 8340-8348, 2018. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 291, + 545, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 291, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 311, + 291, + 545, + 356 + ], + "type": "text", + "content": "[4] Sam Bond-Taylor, Peter Hessey, Hiroshi Sasaki, Toby P Breckon, and Chris G Willcocks. Unleashing transformers: parallel token prediction with discrete absorbing diffusion for fast high-resolution image generation from vector-quantized codes. In ECCV, pages 170–188. Springer, 2022. 2, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 358, + 545, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 358, + 545, + 391 + ], + "spans": [ + { + "bbox": [ + 311, + 358, + 545, + 391 + ], + "type": "text", + "content": "[5] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ICLR, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 311, + 393, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 393, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 311, + 393, + 545, + 437 + ], + "type": "text", + "content": "[6] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. NeurIPS, 35:31769-31781, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 439, + 545, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 439, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 311, + 439, + 545, + 471 + ], + "type": "text", + "content": "[7] Caroline Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A Efros. Everybody dance now. In ICCV, pages 5933-5942, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 311, + 474, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 474, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 311, + 474, + 545, + 506 + ], + "type": "text", + "content": "[8] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, pages 11315-11325, 2022. 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 311, + 508, + 545, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 508, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 311, + 508, + 545, + 561 + ], + "type": "text", + "content": "[9] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. NeurlPS, 34:19822-19835, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 564, + 545, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 564, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 307, + 564, + 545, + 608 + ], + "type": "text", + "content": "[10] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. NeurIPS, 35:16890-16902, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 610, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 610, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 610, + 545, + 643 + ], + "type": "text", + "content": "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, pages 12873-12883, 2021. 2, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 689 + ], + "type": "text", + "content": "[12] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. In ECCV, pages 1-19. Springer, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[13] Aviv Gabbay and Yedid Hoshen. Demystifying inter-class disentanglement. *ICLR*, 2020. 4" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22755" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[14] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, pages 89-106. Springer, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 173 + ], + "type": "text", + "content": "[15] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long video generation with time-agnostic vqgan and timesensitive transformer. In ECCV, pages 102-118. Springer, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 219 + ], + "type": "text", + "content": "[16] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 221, + 287, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 221, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 48, + 221, + 287, + 265 + ], + "type": "text", + "content": "[17] Lena Gorelick, Moshe Blank, Eli Shechtman, Michal Irani, and Ronen Basri. Actions as space-time shapes. IEEE transactions on pattern analysis and machine intelligence, 29(12):2247-2253, 2007. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 267, + 287, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 267, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 48, + 267, + 287, + 311 + ], + "type": "text", + "content": "[18] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, pages 10696-10706, 2022. 1, 2, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 313, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 313, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 48, + 313, + 287, + 368 + ], + "type": "text", + "content": "[19] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 370, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 370, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 48, + 370, + 287, + 424 + ], + "type": "text", + "content": "[20] Ligong Han, Jian Ren, Hsin-Ying Lee, Francesco Barbieri, Kyle Olszewski, Shervin Minaee, Dimitris Metaxas, and Sergey Tulyakov. Show me what and tell me how: Video synthesis via multimodal conditioning. In CVPR, pages 3615-3625, 2022. 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 470 + ], + "type": "text", + "content": "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 472, + 287, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 287, + 527 + ], + "type": "text", + "content": "[22] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 529, + 287, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 529, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 48, + 529, + 287, + 561 + ], + "type": "text", + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 564, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 564, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 564, + 287, + 597 + ], + "type": "text", + "content": "[24] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. *ICLR*, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "text", + "content": "[25] Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation. NeurIPS, 31, 2018. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 666 + ], + "type": "text", + "content": "[26] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In ICML, pages 9902-9915. PMLR, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[27] Yuming Jiang, Shuai Yang, Haonan Qju, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[28] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In ECCV, pages 694–711. Springer, 2016. 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[29] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4401-4410, 2019. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 143, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 185 + ], + "type": "text", + "content": "[30] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR., pages 8110-8119, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 188, + 545, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 188, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 307, + 188, + 545, + 221 + ], + "type": "text", + "content": "[31] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. NeurIPS, 32, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 224, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 224, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 224, + 545, + 266 + ], + "type": "text", + "content": "[32] Kathleen M Lewis, Srivatsan Varadharajan, and Ira Kemelmacher-Shlizerman. Tryongan: Body-aware try-on via layered interpolation. ACM Transactions on Graphics (TOG), 40(4):1-10, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 269, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 269, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 269, + 545, + 335 + ], + "type": "text", + "content": "[33] Lingjie Liu, Weipeng Xu, Marc Habermann, Michael Zollhöfer, Florian Bernard, Hyeongwoo Kim, Wenping Wang, and Christian Theobalt. Neural human video rendering by learning dynamic textures and rendering-to-video translation. IEEE Transactions on Visualization and Computer Graphics, PP:1-1, 05 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 337, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 545, + 392 + ], + "type": "text", + "content": "[34] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 38(5):1-14, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "type": "text", + "content": "[35] Wen Liu, Zhixin Piao, Jie Min, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis. In ICCV, pages 5904-5913, 2019. 3, 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 440, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 440, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 307, + 440, + 545, + 473 + ], + "type": "text", + "content": "[36] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In ICCV, pages 10985-10995, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 475, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 517 + ], + "type": "text", + "content": "[37] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 563 + ], + "type": "text", + "content": "[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831. PMLR, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 609 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "text", + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[41] Kripasindhu Sarkar, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Humangan: A generative model of human images. In 2021 International Conference on 3D Vision (3DV), pages 258–267. IEEE, 2021. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22756" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 667 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[42] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823, 2015. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "text", + "content": "[43] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. NeurIPS, 32, 2019. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "type": "text", + "content": "[44] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. ICLR, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 229 + ], + "type": "text", + "content": "[45] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In CVPR, pages 11050-11059, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 286, + 274 + ], + "type": "text", + "content": "[46] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In CVPR, pages 3626-3636, 2022. 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 275, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 275, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 275, + 286, + 308 + ], + "type": "text", + "content": "[47] Jianxin Sun, Qiyao Deng, Qi Li, Muyi Sun, Min Ren, and Zhenan Sun. Anyface: Free-style text-to-face synthesis and manipulation. In CVPR, pages 18687-18696, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 309, + 286, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 286, + 352 + ], + "type": "text", + "content": "[48] Yu Tian, Jian Ren, Mengei Chai, Kyle Olszewski, Xi Peng, Dimitris N Metaxas, and Sergey Tulyakov. A good image generator is what you need for high-resolution video synthesis. ICLR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 354, + 286, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 286, + 387 + ], + "type": "text", + "content": "[49] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In CVPR, pages 1526-1535, 2018. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 388, + 286, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 286, + 431 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 286, + 431 + ], + "type": "text", + "content": "[50] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 434, + 286, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 286, + 465 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 286, + 465 + ], + "type": "text", + "content": "[51] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. NeurIPS, 29, 2016. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 467, + 286, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 467, + 286, + 488 + ], + "spans": [ + { + "bbox": [ + 49, + 467, + 286, + 488 + ], + "type": "text", + "content": "[52] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 2, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 490, + 286, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 490, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 490, + 286, + 544 + ], + "type": "text", + "content": "[53] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. ICLR, 2023. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 545, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 545, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 545, + 286, + 578 + ], + "type": "text", + "content": "[54] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 580, + 286, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 580, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 580, + 286, + 622 + ], + "type": "text", + "content": "[55] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. *ICLR*, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 624, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 286, + 667 + ], + "type": "text", + "content": "[56] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose-guided human video generation. arXiv preprint arXiv:1910.09139, 2019. 2, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22757" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/ec345001-5311-431e-90bd-6b91f4b47f2a_content_list.json b/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/ec345001-5311-431e-90bd-6b91f4b47f2a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..81c0429924e3999a787fc39144e8c374adb7e8ac --- /dev/null +++ b/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/ec345001-5311-431e-90bd-6b91f4b47f2a_content_list.json @@ -0,0 +1,2069 @@ +[ + { + "type": "text", + "text": "Text2Room: Extracting Textured 3D Meshes from 2D Text-to-Image Models", + "text_level": 1, + "bbox": [ + 101, + 130, + 869, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lukas Hollein $^{1*}$ Ang Cao $^{2*}$ Andrew Owens $^{2}$ Justin Johnson $^{2}$ Matthias Nießner $^{1}$ $^{1}$ Technical University of Munich \n $^{2}$ University of Michigan", + "bbox": [ + 129, + 179, + 836, + 217 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bdb645acfa17c805cdc9281c49f2227de27f066caee34c8152f2ff7c627d3f83.jpg", + "image_caption": [ + "(a) 3D Mesh Generation from Text" + ], + "image_footnote": [], + "bbox": [ + 81, + 250, + 699, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3fbf50571c12f07c7b65d953b7fda35f01913fb28aef0380854534873099c0f1.jpg", + "image_caption": [ + "(b) Rendered Image + Mesh", + "Figure 1. Textured 3D mesh generation from text prompts. We generate textured 3D meshes from a given text prompt using 2D text-to-image models. (a) The scene is iteratively created from different viewpoints (marked in blue). (b) Our generated mesh contains compelling textures and geometry. We remove the ceiling in the top-down views for better visualization of the scene layout." + ], + "image_footnote": [], + "bbox": [ + 707, + 251, + 893, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 523, + 313, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Text2Room1, a method for generating room-scale textured 3D meshes from a given text prompt as input. To this end, we leverage pre-trained 2D text-to-image models to synthesize a sequence of images from different poses. In order to lift these outputs into a consistent 3D scene representation, we combine monocular depth estimation with a text-conditioned inpainting model. The core idea of our approach is a tailored viewpoint selection such that the content of each image can be fused into a seamless, textured 3D mesh. More specifically, we propose a continuous alignment strategy that iteratively fuses scene frames with the existing geometry to create a seamless mesh. Unlike existing works that focus on generating single objects [56, 41] or zoom-out trajectories [18] from text, our method generates complete 3D scenes with multiple objects and explicit 3D geometry. We evaluate our approach using qualitative and quantitative metrics, demonstrating it as the first method to generate room-scale 3D geometry with compelling textures from only text as input.", + "bbox": [ + 75, + 556, + 473, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 496, + 630, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mesh representations of 3D scenes are a crucial component for many applications, from AR/VR asset creation to computer graphics, yet creating these 3D assets remains a painstaking process that requires considerable expertise. In the 2D domain, recent works have successfully created high-quality images from text using generative models, such as diffusion models [65, 58, 67]. These methods significantly reduce the barriers to creating images that contain a user's desired content, effectively helping towards the democratization of content creation. An emerging line of work has sought to apply similar methods to create 3D models from text [9, 56, 29, 41, 38], yet existing approaches come with a number of significant limitations and lack the generality of 2D text-to-image models.", + "bbox": [ + 496, + 522, + 890, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One of the core challenges of generating 3D models is coping with the lack of available 3D training data, as 3D datasets are vastly smaller than those available in many other applications, such as 2D image synthesis. For example, methods that directly use 3D supervision, such as Chen et al. [9], are often limited to datasets of simple shapes, such as ShapeNet [8]. To address these data limitations, recent methods [56, 29, 41, 38, 88] lift the expressive power of 2D text-to-image models into 3D by formulating 3D generation as an iterative optimization problem in the image domain. This allows them to generate 3D ob", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* joint first authorship", + "bbox": [ + 94, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://lukashoel.github.io/text-to-room", + "bbox": [ + 94, + 887, + 419, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7909", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jects stored in a radiance field representation, demonstrating the ability to generate arbitrary (neural) shapes from text. However, these methods cannot easily be extended to create room-scale 3D structure and texture. The challenge of generating large scenes is ensuring that the generated output is dense and coherent across outward-facing viewpoints, and that these views contain all of the required structures, such as walls, floors, and furniture. Additionally, a mesh remains a desired representation for many end-user tasks, such as rendering on commodity hardware (which requires an additional conversion step as presented in Lin et al. [41]).", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these shortcomings, we propose a method that extracts scene-scale 3D meshes from off-the-shelf 2D text-to-image models. Our method iteratively generates a scene through inpainting and monocular depth estimation. We produce an initial mesh by generating an image from text, and backproject it into 3D using a depth estimation model. Then, we iteratively render the mesh from novel viewpoints. From each one, we fill in holes in the rendered images via inpainting, then fuse the generated content into the mesh (Fig. 1a).", + "bbox": [ + 75, + 257, + 472, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our iterative generation scheme has two important design considerations: how we choose the viewpoints, and how we merge generated scene content with the existing mesh. We first select viewpoints from predefined trajectories that will cover large amounts of scene content, then adaptively select viewpoints that close remaining holes. When merging generated content with the mesh, we align the two depth maps to create smooth transitions, and remove parts of the mesh that contain distorted textures. Together, these decisions lead to large, scene-scale 3D meshes with compelling textures and consistent geometry (Fig. 1b), that can represent a wide range of rooms.", + "bbox": [ + 75, + 407, + 472, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are:", + "bbox": [ + 76, + 590, + 323, + 603 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Generating 3D meshes of room-scale indoor scenes with compelling textures and geometry from any text input.", + "- A method that leverages 2D text-to-image models and monocular depth estimation to lift frames into 3D in an iterative scene generation. Our proposed depth alignment and mesh fusion steps, enable us to create seamless and undistorted geometry and textures.", + "- A two-stage tailored viewpoint selection that samples camera poses from optimal positions to first create the room layout and furniture and then close any remaining holes, creating a watertight mesh." + ], + "bbox": [ + 76, + 604, + 468, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 784, + 218, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-based Generation has seen significant advances due to large-scale image-text datasets [73, 72, 14, 71] and scalable generative model architectures [16, 66, 60, 32], enabling synthesis of novel images from text [20, 3, 54].", + "bbox": [ + 76, + 809, + 468, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, diffusion models [78, 23, 80, 81, 82] achieved impressive results on image synthesis [15, 65, 67, 50, 58]", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "through improvements like latent space denoising [65, 84], faster sampling [23, 79, 52, 34], and better guidance [24].", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In particular, text-to-image methods like Stable Diffusion [65], Imagen [67], GLIDE [50] and DALL-E 2 [58] yield diverse, high-fidelity, and controllable [6, 94] outputs. Text-based generation has been extended to other modalities including audio [35, 17, 28, 70], video [75, 91, 85, 25], and 4D fields [76]. We use text-to-image models by lifting their generated output into complete 3D scene meshes.", + "bbox": [ + 496, + 122, + 893, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-3D. Several methods use 3D data for supervised training of text-to-3D models [9, 51, 5]; however this direction remains challenging due to the lack of large-scale aligned datasets of text and 3D.", + "bbox": [ + 496, + 232, + 890, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Alternative approaches use 2D vision-language models like CLIP [57] to create 3D content by formulating the generation as an optimization problem in the image domain [86, 29, 38, 48, 30] or as object alignment [69]. Related methods refine existing 3D input through text guidance in a similar fashion [46, 10, 87, 62].", + "bbox": [ + 496, + 292, + 890, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent methods [56, 41, 45, 88, 44] combine large text-to-image diffusion models [65, 67] and neural radiance fields [47] to generate 3D objects without training. Other approaches train custom diffusion models on a similar text-to-3D task [39, 49, 12]. In contrast, we use a fixed text-to-image model and extract a 3D mesh representing entire scenes of many objects and structural elements like walls.", + "bbox": [ + 496, + 385, + 890, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D-Consistent View Synthesis from a Single Image. Several methods have been proposed that perform novel-view-synthesis from a single image [64, 90, 74, 61, 19]. Others optimize a neural 3D representation of an object, that can be viewed from arbitrary novel view points [92, 89, 1]. Another line of work performs perpetual view generation [77, 42, 40, 7], synthesizing videos via a render-refine-repeat pattern from a single RGB image that depict a scene along a forward-facing camera trajectory. In very recent concurrent work, Fridman et al. [18] create 3D scenes from text, but focus on this type of 3D-consistent \"zoom-out\" video generation. Instead, we generate complete, textured 3D room geometry from arbitrary trajectories.", + "bbox": [ + 496, + 494, + 890, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 708, + 589, + 723 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method creates a textured 3D mesh of a complete scene from text input. To this end, we continuously fuse generated frames from a 2D text-to-image model at different poses into a joint 3D mesh, creating the scene over time. The core idea of our approach is a two-stage tailored viewpoint selection, that first generates the scene layout and objects and then closes remaining holes in the 3D geometry (Section 3.4). We visualize this workflow in Figure 2. For each pose in both stages, we apply an iterative scene generation scheme to update the mesh (Section 3.1). We first align each frame with the existing geometry with a depth align-", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7910", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d0b92d778b3d5b2fe0a757649ba1e4ed4e9697e0ba97a52b85bd754ccbfbaec0.jpg", + "image_caption": [ + "(a) Scene Generation Stage" + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 656, + 169 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/67d6e7dc793e04bc55296ebbcf382d887b1ed15e3a373e64a5642dbd3a1febba.jpg", + "image_caption": [ + "(b) Scene Completion Stage" + ], + "image_footnote": [], + "bbox": [ + 658, + 88, + 890, + 169 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/76fae7b7534754273b139f805925c072af7ac5b7cf560feef8f74589ffe696b4.jpg", + "image_caption": [ + "Figure 2. Method overview. We iteratively create a textured 3D mesh in two stages. (a) First, we sample predefined poses and text to generate the complete scene layout and furniture. Each new pose (marked in green) adds newly generated geometry to the mesh (depicted by green triangles) in an iterative scene generation scheme (see Figure 3 for details). Blue poses/triangles denote viewpoints that created geometry in a previous iteration. (b) Second, we fill in the remaining unobserved regions by sampling additional poses (marked in red) after the scene layout is defined.", + "Figure 3. Iterative scene generation. For each new pose, we render the current mesh to obtain partial RGB and depth renderings. We complete both, utilizing respective inpainting models and the text prompt. Next, we perform depth alignment (see Section 3.2) and mesh filtering (see Section 3.3) to obtain an optimal next mesh patch, that is finally fused with the existing geometry." + ], + "image_footnote": [], + "bbox": [ + 81, + 290, + 480, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ment strategy (Section 3.2). Next, we triangulate and filter the novel content to merge it into the mesh (Section 3.3).", + "bbox": [ + 75, + 604, + 468, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Iterative 3D Scene Generation", + "text_level": 1, + "bbox": [ + 76, + 643, + 344, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our scene is represented as a mesh $\\mathcal{M} = (\\mathcal{V},\\mathcal{C},\\mathcal{S})$ where the vertices $\\mathcal{V}\\in \\mathbb{R}^{N\\times 3}$ , vertex colors $\\mathcal{C}\\in \\mathbb{R}^{N\\times 3}$ and the face set $S\\in \\mathbb{N}_0^{M\\times 3}$ are generated over time. Input to our method is a set of arbitrary text prompts $\\{P_t\\}_{t = 1}^T$ that corresponds to our selected poses $\\{E_t\\}_{t = 1}^T\\in \\mathbb{R}^{3\\times 4}$ in both stages. Inspired by recent methods [42, 40], we iteratively build up the scene, following a render-refine-repeat pattern. We summarize this iterative scene generation process in Figure 3. Formally, for each step of generation $t$ , we first render the current scene from a novel viewpoint:", + "bbox": [ + 75, + 666, + 468, + 819 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t}, d _ {t}, m _ {t} = r \\left(\\mathcal {M} _ {t}, E _ {t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 828, + 468, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $r$ is a classical rasterization function without shading, $I_{t}$ is the rendered image, $d_{t}$ the rendered depth and $m_{t}$ the image-space mask, that marks pixels without observed", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "content. We then use a fixed text-to-image model $\\mathcal{F}_{t2i}$ to inpaint unobserved pixels according to the text prompt:", + "bbox": [ + 498, + 290, + 890, + 320 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {t} = \\mathcal {F} _ {t 2 i} \\left(I _ {t}, m _ {t}, P _ {t}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 332, + 890, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next, we inpaint unobserved depth by applying a monocular depth estimator $\\mathcal{F}_d$ in our depth alignment (see Section 3.2):", + "bbox": [ + 498, + 364, + 890, + 393 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {d} _ {t} = \\text {p r e d i c t - a n d - a l i g n} \\left(\\mathcal {F} _ {d}, I _ {t}, d _ {t}, m _ {t}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 409, + 890, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we combine the novel content $\\{\\hat{I}_t,\\hat{d}_t,m_t\\}$ with the existing mesh by our fusion scheme (see Section 3.3):", + "bbox": [ + 498, + 436, + 890, + 469 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {t + 1} = f u s e \\left(\\mathcal {M} _ {t}, \\hat {I} _ {t}, \\hat {d} _ {t}, m _ {t}, E _ {t}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 479, + 890, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Depth Alignment Step", + "text_level": 1, + "bbox": [ + 500, + 512, + 705, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To lift a 2D image $I$ into 3D, we predict the per-pixel depth. To correctly combine old and new content, it is necessary that both align with each other. In other words, similar regions in a scene like walls or furniture should be placed at similar depth. However, directly using the predicted depth for backprojection leads to hard cuts and discontinuities in the 3D geometry, since the depth is inconsistent in scale between subsequent viewpoints (see Figure 7a).", + "bbox": [ + 496, + 537, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we perform depth alignment in two-stages. First, we use a state-of-the-art depth inpainting network [4] that takes ground-truth depth $d$ for known parts in the image as input and aligns the prediction to it: $\\hat{d}_p = \\mathcal{F}_d(I,d)$ .", + "bbox": [ + 496, + 659, + 890, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by Liu et al. [42] we then improve the result by optimizing for scale and shift parameters $\\gamma, \\beta \\in \\mathbb{R}$ , aligning predicted and rendered disparity in the least squares sense:", + "bbox": [ + 496, + 720, + 890, + 766 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\gamma , \\beta} \\quad \\left\\| m \\odot \\left(\\frac {\\gamma}{\\hat {d} _ {p}} + \\beta - \\frac {1}{d}\\right) \\right\\| ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 777, + 890, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where we mask out unobserved pixels via $m$ . We can then extract the aligned depth as $\\hat{d} = (\\frac{\\gamma}{\\hat{d}_p} + \\beta)^{-1}$ . Finally, we smooth $\\hat{d}$ by applying a $5 \\times 5$ Gaussian kernel at the mask edges (see supplemental material for more details).", + "bbox": [ + 496, + 833, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7911", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/207a3b4cd3937f72df3864668ca2d82e905f71588dbc74ad1f280b31f55c3aa2.jpg", + "image_caption": [ + "(a) Pixel Triangulation (b) Face Filtering (c) Mesh Fusion" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 215, + 209 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/56b07ccd1dde58e503e1527a1e776c02b6a334341cfbbb86190444be0a8cab52.jpg", + "image_caption": [ + "Figure 4. Visualization of our mesh fusion step. (a) We triangulate an image, such that 4 neighboring pixels (orange dots) create two faces. (b) We filter a face (marked in red), if its surface normal forms a small grazing angle with the viewing direction or if any edge in world space is too long. (c) We fuse the remaining faces (marked in green) with the existing geometry (marked in blue)." + ], + "image_footnote": [], + "bbox": [ + 230, + 88, + 352, + 209 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a78e77db4e1e55dc29b4357bf074357eee9e02e94c4f2e8fb663e01bfad09ec3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 88, + 452, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Mesh Fusion Step", + "text_level": 1, + "bbox": [ + 76, + 342, + 250, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At each step, we insert new content $\\{\\hat{I}_t,\\hat{d}_t,m_t\\}$ into the scene. For that, we first backproject the image-space pixels into a world-space point cloud:", + "bbox": [ + 75, + 366, + 468, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {t} = \\left\\{E _ {t} ^ {- 1} K ^ {- 1} \\cdot \\hat {d} _ {t} [ u, v ] \\cdot (u, v, 1) ^ {T} \\right\\} _ {u = 0, v = 0} ^ {W, H}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 424, + 468, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K \\in \\mathbb{R}^{3 \\times 3}$ are the camera intrinsics and $W, H$ are image width and height, respectively. We then use a simple triangulation scheme (Figure 4a), where each four neighboring pixels $\\{(u, v), (u + 1, v), (u, v + 1), (u + 1, v + 1)\\}$ in the image form two triangles. Since the estimated depth is noisy, this naive triangulation creates stretched out 3D geometry (see Figure 7b). To alleviate this problem, we propose two filters that remove stretched out faces (Figure 4b).", + "bbox": [ + 75, + 458, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we filter faces based on their edge length. We remove a face if the Euclidean distance of any face edge is larger than a threshold $\\delta_{edge}$ . Second, we filter faces based on the angle between surface normal and viewing direction:", + "bbox": [ + 75, + 580, + 468, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = \\left\\{\\left(i _ {0}, i _ {1}, i _ {2}\\right) \\mid n ^ {T} v > \\delta_ {s n} \\right\\} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 659, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S$ is the face set, $(i_0,i_1,i_2)$ are the vertex indices of the triangle, $\\delta_{sn}$ is the threshold, $n\\in \\mathbb{R}^3$ is the normalized face normal, and $v\\in \\mathbb{R}^3$ is the normalized view direction in world space from the camera center towards the average pixel location from which the triangle originated. This avoids creating texture for large regions of the mesh from a comparatively small number of pixels from an image.", + "bbox": [ + 75, + 686, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we fuse together the newly generated mesh patch and the existing geometry (Figure 4c). All faces that are backprojected from pixels falling into the inpainting mask $m_{t}$ are stitched together with their neighboring faces, which are already part of the mesh. Precisely, we continue the triangulation scheme at all edges of $m_{t}$ , but use the existing vertex positions of $\\mathcal{M}_t$ to create the corresponding faces.", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Two-Stage Viewpoint Selection", + "text_level": 1, + "bbox": [ + 500, + 90, + 771, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A key part of our method is the choice of text prompts and camera poses from which the scene is synthesized. Users can in principle choose these inputs arbitrarily to create any desired indoor scene. However, the generated scene can degenerate and contain stretch and hole artifacts, if poses are chosen carelessly (see Figure 7 and supplemental material). To this end, we propose a two-stage viewpoint selection strategy, that samples each next camera pose from optimal positions and refines empty regions subsequently.", + "bbox": [ + 496, + 114, + 890, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generation Stage. In the first stage, we create the main parts of the scene, including the general layout and furniture. We subsequently render predefined trajectories in different directions that eventually cover the whole room. We found generation works best, if each trajectory starts off from a viewpoint with mostly unobserved regions. This generates the outline of the next chunk, while still being connected to the rest of the scene (e.g., see Figure 3). Then, we complete the 3D structure of that chunk by moving and rotating into it subsequently until the end of the trajectory.", + "bbox": [ + 496, + 255, + 890, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, we ensure an optimal observation distance for each pose. We translates camera positions $T_0 \\in \\mathbb{R}^3$ along the look-at direction $L \\in \\mathbb{R}^3$ uniformly: $T_{i+1} = T_i - 0.3L$ . We stop if the mean rendered depth is larger than 0.1 or discard the camera after 10 steps. This avoids views too close to existing geometry. For example, the green pose in Figure 2a is moved back as far as possible into the existing geometry such that it views most of the empty floor region.", + "bbox": [ + 496, + 407, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We create closed room layouts following this principle, by choosing trajectories that generate the next chunks in a circular motion, roughly centered around the origin. We found it helpful to discourage the text-to-image generator from generating furniture in unwanted regions by engineering the text prompts accordingly. For example, for poses looking at the floor or ceiling, we choose text prompts that only contain the words \"floor\" or \"ceiling\", respectively.", + "bbox": [ + 496, + 530, + 890, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Completion Stage. After the first stage, the scene layout and furniture is defined. However, it is impossible to choose sufficient poses $a$ -priori. Since the scene is generated on-the-fly, the mesh contains holes that were not observed by any camera (see Figure 7c). We complete the scene by sampling additional poses $a$ -posteriori, looking at those holes.", + "bbox": [ + 496, + 656, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by trajectory optimization [22, 63], we voxelize the scene into dense uniform cells. We sample random poses in each cell, discarding those being too close to existing geometry. We select one pose per cell that views most unobserved pixels (e.g., see the red poses in Figure 2b).", + "bbox": [ + 496, + 748, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we inpaint the scene from all chosen camera poses following Section 3.1. Similar to Fridman et al. [18], we observe it is important to clean the inpainting masks, because our text-to-image generator can generate better results for large connected regions. Thus, we first inpaint small holes", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7912", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with a classical inpainting algorithm [83] and dilate the remaining holes. We additionally remove all faces that fall into the dilated region and are close to the rendered depth. Please see the supplemental material for more details.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we run Poisson surface reconstruction [33] on the scene mesh. This closes any remaining holes after completion and smoothens out discontinuities. The result is a watertight mesh of the generated scene, that can be rendered with classical rasterization.", + "bbox": [ + 75, + 152, + 468, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 76, + 242, + 165, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We implement mesh rasterization and fusion with Pytorch3D [59]. As our text-to-image model $\\mathcal{F}_{t2i}$ , we utilize a Stable Diffusion [65] model, that is finetuned on the image inpainting task, using additional mask input. We generate a single inpainting proposal and employ a state-of-the-art guided diffusion sampler [43]. As our monocular depth estimator $\\mathcal{F}_d$ , we employ an Iron-Depth [4] model, that is trained on indoor scenes from the ScanNet dataset [13] and augment it for depth inpainting according to Bae et al. [4]. We set $\\delta_{edge} = 0.1$ and $\\delta_{sn} = 0.1$ in all our experiments. During generation, we use 20 different trajectories with 10 frames each sampled between the respective start and end poses. We construct prompts using the guidelines suggested by Pierre [55]. Creating one scene takes approximately 50 minutes on one RTX 3090 GPU.", + "bbox": [ + 75, + 272, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Baselines. To the best of our knowledge, there are no direct baselines that generate textured 3D room geometry from text. We compare against four related methods (please see the supplemental material for more details about baselines).", + "- PureClipNeRF [38]: We compare against text-to-3D methods for generating objects [56, 41, 29, 38, 88] and choose Lee et al. [38] as open-source representative.", + "- Outpainting [58, 53]: We combine outpainting from a Stable Diffusion [65] model with depth estimation and triangulation to create a mesh from an enlarged viewpoint.", + "- Text2Light [11]: We generate RGB panoramas from text using Chen et al. [11]. Estimating 3D mesh structure from a panorama is difficult. Related approaches estimate room layout [93], perform view synthesis [36, 26, 21, 27] or predict $360^{\\circ}$ depth [2, 31]. We perform depth prediction and subsequently apply our mesh fusion step.", + "- Blockade [37]: We apply Blockade [37], which uses a text-to-image diffusion model to produce more expressive RGB panoramas. We then extract the mesh similarly." + ], + "bbox": [ + 75, + 503, + 468, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. The generated 3D geometry is evaluated both quantitatively and qualitatively. We calculate CLIP Score (CS) [57] and Inception Score (IS) [68] on RGB renderings of the respective scenes. Additionally, we conduct a user study and ask $n = 61$ users to score Perceptual Quality ( $PQ$ ) and 3D Structure Completeness ( $3DS$ ) of the whole scene on a scale of 1-5.", + "bbox": [ + 75, + 794, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9a0e2e6b569ce4ea3422066d6f05c3a98075626a5b6f4aba41d890086a0adbb9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method2D MetricsUser Study
CS ↑IS ↑PQ ↑3DS ↑
PureClipNeRF [38]24.061.262.342.38
Outpainting [58, 53]23.101.602.902.58
Text2Light [11]+Ours25.992.212.822.97
Blockade [37]+Ours26.292.133.353.36
Ours w/o alignment26.731.783.122.96
Ours w/o stretch removal27.721.863.283.75
Ours w/o completion27.972.183.723.87
Ours28.022.314.014.19
", + "bbox": [ + 504, + 88, + 890, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison. We report 2D metrics and user study results, including: Clip Score (CS), Inception Score (IS), Perceptual Quality (PQ), and 3D Structure Completeness (3DS). Our method creates scenes with the highest quality.", + "bbox": [ + 498, + 268, + 890, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 352, + 683, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We show top-down views into the scene and RGB renderings from within for our method and baselines in Figure 6. We show additional results of our method in Figure 5. PureClipNeRF [38] creates the key objects of the given text prompt, but does not create a complete 3D structure with floor, walls and ceilings. Outpainting [58, 53] creates high-detail textures, but projection from a single viewpoint creates holes due to occlusion and hinders the creation of complete 3D geometry. Text2Light [11] and Blockade [37] both create a high-detail $360^{\\circ}$ view of a complete scene, but occlusions that cannot be resolved from a single panoramic viewpoint lead to holes in the extracted 3D geometry.", + "bbox": [ + 496, + 377, + 890, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast, our approach creates high-detail textures and geometry, that are fused into a complete 3D scene mesh without holes. The resulting scenes contain flat floors, walls and ceilings, as well as 3D object geometry distributed throughout the scene. When specifying text prompts with a huge variety, the resulting scene contains a diverse set of objects. Please see the supplemental material for more scenes, animated results, intermediate outputs of our baselines (such as the panoramic images) as well as top-down views of meshes, that contain the reconstructed ceilings.", + "bbox": [ + 496, + 560, + 890, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 500, + 724, + 692, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We show quantitative results averaged over multiple scenes in Table 1. We render 60 images from novel viewpoints for each scene to calculate the 2D metrics. We present users with multiple top-down views and renderings for each scene and let them rate each method individually (no side-by-side comparison). Stretched-out geometry and holes in the 3D geometry lead to lower scores for the baselines in all image-based metrics. Our approach achieves the highest scores, because the renderings are complete from arbitrary novel poses, satisfy the given text-prompt and", + "bbox": [ + 496, + 750, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7913", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d2eca9ce87387bb4b59230539c92d088cf169eff0c69a25a872dfbe2afb4f335.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 88, + 436, + 265 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb9852895e792c867f2c9153ce6babf4c15383e7bacf3a169c80284e9f13c862.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 88, + 640, + 176 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/85cb3ff25fbe5385b65da195af10e3c71bb21ae14c1d4ed285cbdc5959d68f08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 89, + 843, + 175 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2e2041852c7c20bccf5e8b9dfe23c6f3d82d4b06e2b7be4fbf1bb70a631444c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 178, + 640, + 266 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/41c2fb0160bb6ce1d137e34ec69bda5b2e94227da28e92b56b0c8a8fac70a26c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 178, + 843, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Editorial Style Photo, Coastal Bathroom, Clawfoot Tub, Seashell, Wicker, Mosaic Tile, Blue and White", + "bbox": [ + 148, + 270, + 820, + 284 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d490125d2ee33df607bb19a86327e4afcda30b628d9ee3f401a42fa10fbe39d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 287, + 398, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/feec8cc2863f2075c549047bbfca4f0c8aa7c0a50e616b0c31d6319261cfdb90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 285, + 640, + 372 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0f2d1a79fa49f25f6f0a034cd2981799a201be4c59af3daa844be026d200ae2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 285, + 843, + 372 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16c4e258a97397ad49226868831fb185695929825561df0fafe92cb295bf5c5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 375, + 640, + 463 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e6fbcfc3552c572956b37bd02099b975c4565c48770e4cbc7ce5d7ff6185d418.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 375, + 843, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A living room with a lit furnace, couch, and cozy curtains, bright lamps that make the room look well-lit", + "bbox": [ + 143, + 465, + 825, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db647f626426a2292c35eff9ef458c3ee87607c2ddf0083b4f5759a873122502.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 482, + 426, + 655 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/15ca036db5bf5940cde79bfa1f66b04ed9f7129e401f712738fa684052ec52c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 481, + 640, + 568 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1fd22bc3048c7b7f28f67b3eb0c36462f2c6889038e2ab2f6d1a984e58dbfbed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 481, + 843, + 569 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afbbc7a61702516a0571f1f09d412731d1f64f18083493f67c0539cb37a33927.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 570, + 640, + 657 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bb525f53e3f383cd059cbe20736408a07be28d5338f624d2a5ea681b1ad177d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 570, + 843, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Editorial Style Photo, Modern Living Room, Large Window, Leather, Glass, Metal, Wood Paneling, Apartment", + "bbox": [ + 124, + 662, + 844, + 676 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/251399c527f029c48a652a0f5a880cbfcc8a0ca5cae9b8fec8486ebe0d696fa6.jpg", + "image_caption": [ + "Figure 5. 3D scene generation results of our method. We show color and shaded geometry renderings from generated scenes with corresponding text prompts. Our method synthesizes realistic meshes satisfying text descriptions. We remove the ceiling in the top-down view for better visualization of the scene layout." + ], + "image_footnote": [], + "bbox": [ + 163, + 679, + 401, + 853 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d1fe957ca3366c375343d3fb08bb1c1335739ec6885d0d06bc6b4aa248254d7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 678, + 640, + 763 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bfbe32a7c7e0ce74676c8b8dbf14c8080842fa5057bcf199bb289fda05771198.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 676, + 843, + 763 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/abd48c60ff69b6a59a731b9277017eccf017e10828ae6d4a62bcb1fffb2565b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 766, + 640, + 854 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d0f77186f66663241a634ce076d5f236a931fb67b7baa52c9aa03205830d021e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 766, + 843, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Editorial Style Photo, Modern Nursery, Table Lamp, Rocking Chair, Tree Wall Decal, Wood, Cotton, Faux Fur", + "bbox": [ + 122, + 858, + 844, + 872 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7914", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cc180d7990d0664bd23cc9d4f05424da14d560fd3eea4a2dd21c1a2b7196be2f.jpg", + "image_caption": [ + "Editorial Style Photo, Industrial Home Office, Steel Shelves, Concrete, Metal, Edison Bulbs, Exposed Ductwork" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 883, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4bead8a57fc0d8fcfc81dbe818a3b5c86589f67a4e4d00571b6ed51932708cc8.jpg", + "image_caption": [ + "a bedroom with a king-size bed and a large wardrobe", + "PureClipNeRF [38]", + "Outpainting [58, 53]", + "Text2Light [11]+Ours", + "Blockade [37]+Ours", + "Ours", + "Figure 6. Qualitative comparison of our method and baselines. PureClipNeRF [38] cannot produce immersive scenes with floors and walls. Outpainting [58, 53] does not produce 3D consistent scenes. Text2Light [11] and Blockade [37] both have holes due to occlusions. In contrast, our method creates complete meshes without holes and high details. We remove the ceiling in the top-down view for better visualization of the scene layout." + ], + "image_footnote": [], + "bbox": [ + 86, + 296, + 883, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "contain high-resolution image features. Users prefer our method, which highlights the quality of our accurate and complete geometry, as well as the RGB texture.", + "bbox": [ + 75, + 588, + 470, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 76, + 643, + 187, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The key ingredients of our method are depth alignment (Section 3.2), mesh fusion (Section 3.3) and the two-stage viewpoint selection (Section 3.4). We demonstrate the importance of each component in Figure 7 and Table 1.", + "bbox": [ + 75, + 666, + 468, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Depth alignment creates seamless scenes. Monocular depth predictions from subsequent frames can be inconsistent in scale. This leads to disconnected components in the mesh that are backprojected from multiple viewpoints (see Figure 7a). Our depth alignment strategy allows fusing multiple frames into a seamless mesh, eventually creating a complete scene with flat floors, walls, ceilings and no holes.", + "bbox": [ + 75, + 729, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stretch removal creates undistorted scene geometry. During mesh fusion, we update the scene geometry with the contents of the next frame. Due to noisy depth prediction, the objects become stretched out, if they are observed", + "bbox": [ + 75, + 840, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from small grazing angles. Thus, we propose two filters (edge length and surface normal thresholds) that alleviate this issue. Instead of baking in stretched-out geometry (see Figure 7b), we disregard the corresponding faces and let the object be completed from a more suitable, later viewpoint.", + "bbox": [ + 496, + 588, + 890, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Two-stage generation creates complete scenes. Our approach chooses camera poses in two stages to create a complete scene without holes. After generating the scene from predefined trajectories, the scene still contains some holes (see Figure 7c). Because the scene is built-up over time, it is impossible to choose camera poses $a$ -priori, that view all unobserved regions. To this end, our completion stage samples poses $a$ -posteriori to refine those regions. The resulting mesh is watertight and contains no holes (see Figure 7d).", + "bbox": [ + 496, + 667, + 892, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Spatially Varying Scene Generation", + "text_level": 1, + "bbox": [ + 500, + 816, + 808, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method can be applied to generate a scene as the combination of multiple text prompts. Specifically, we use separate text prompts for different poses, crafting a set of trajectories that spatially combines scene descriptions. This", + "bbox": [ + 496, + 840, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7915", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e782e9c77329ce1310e5a56e2d0239a7e09936253dc60d2d7ef1599b93f0a242.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 87, + 285, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e791f881e28aef630bc565e2c60edd16ec6e56a89c90166cb463e1b068799fc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 87, + 460, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/333a53c14136c1e5eb79546a8c33a5cab54f49160f7ef6a6dd762f04155e0afa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 90, + 678, + 226 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/86ddbd6b0aac7e63748ff124b8ffc4f800cf145d629885f11fd0d8abe2c81aa4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 88, + 879, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/054d5f3dda57bf7f8d6fdb9471a1591706a6bb1706c3c841c9c863cf0e95a81a.jpg", + "image_caption": [ + "(a) Ours w/o alignment" + ], + "image_footnote": [], + "bbox": [ + 86, + 229, + 285, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9287746b24ed8c0550ea06d2dc2ebdcb740b8194965f4455273c97f30965d986.jpg", + "image_caption": [ + "(b) Ours w/o stretch removal" + ], + "image_footnote": [], + "bbox": [ + 285, + 229, + 482, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4299f0a5e9e71d98de2892563c77d4a937192577c965e69cc711bf77ef3f1aa.jpg", + "image_caption": [ + "(c) Ours w/o completion" + ], + "image_footnote": [], + "bbox": [ + 483, + 229, + 679, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b4b15330369576b7b9e41ef08aba4d3bc87839299c8f86ba1f45e5320c023c1b.jpg", + "image_caption": [ + "(d) Ours (full)" + ], + "image_footnote": [], + "bbox": [ + 681, + 229, + 877, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "living room, couches, curtains, lit furnace, lamps", + "bbox": [ + 320, + 354, + 643, + 368 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c755f295f8051c799a7331e05b23b30ac84c54b05e8dd50fc27058b529e1a396.jpg", + "image_caption": [ + "Figure 7. Ablation study on the key components of our method. Without depth alignment (see Section 3.2), different parts of the scene are disconnected and do not fuse into a seamless mesh. Without edge and surface normal thresholds (see Section 3.3), many faces are stretched out unnaturally. Without completion (see Section 3.4), the mesh has holes in remaining unobserved regions. Our full pipeline creates complete, high-resolution scenes. We remove the ceiling in the top-down view for better visualization of the scene layout." + ], + "image_footnote": [], + "bbox": [ + 96, + 462, + 259, + 550 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c926aef2331ef8854368beeff4730020aa312b0477b3bba85b8762f26dc73461.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 462, + 447, + 550 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/006d84408941ae6fe0888f7fe3b2ba6cfd328ae279fe2e48b69f3f6093298268.jpg", + "image_caption": [ + "kitchen, dinner table, dishwashers, ovens, countertops living room, lit furnace, couch, curtains" + ], + "image_footnote": [], + "bbox": [ + 116, + 584, + 240, + 671 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/806162c801d775f3e05192c3bf5d0667a193e2ee64958ea3f370e7072ac108b4.jpg", + "image_caption": [ + "bathroom, shower, bathtub bedroom, king-size bed, wardrobe", + "Figure 8. Spatially varying scene generation. Our method can create rooms with multiple parts by prompt mixing. We use separate prompts for cameras viewing different parts of the scene. This is a controllable way to create rooms from multiple descriptions." + ], + "image_footnote": [], + "bbox": [ + 264, + 584, + 447, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "can be desired to avoid repeating elements in a complete scene (e.g., multiple couches could be spread out over the whole room when using the same prompt for every camera). Instead, users can specify different object positions through different camera poses and text prompts. It can also be used to design a house comprised of multiple rooms, each with a different type (e.g., a living room that leads to a kitchen).", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e3bce1ef21548390e5a6e0bb425761c89423b7880322207bea90af4e6ecb3a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 463, + 704, + 542 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/547a261cc20522c050bb21615ed515fba8232845c711e2859e1b180560c04f69.jpg", + "image_caption": [ + "Figure 9. Scene generation with layout guidance. Our method can generate scenes from layout guidance. Left: we describe objects with different prompts for cameras facing at different directions. Right: the generated part of the room." + ], + "image_footnote": [], + "bbox": [ + 705, + 462, + 882, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show results that combine multiple text prompts in Figure 8 and Figure 9.", + "bbox": [ + 496, + 633, + 888, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We note that the layout can only be partially controlled by the camera poses, since scene generation can create chunks with larger or smaller extent. We believe this demonstrates an exciting application of our method, that can be further explored in future work.", + "bbox": [ + 496, + 665, + 890, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations", + "text_level": 1, + "bbox": [ + 500, + 753, + 625, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our approach allows to generate 3D room geometry from arbitrary text prompts that are highly detailed and contain consistent geometry. Nevertheless, our method can still fail under certain conditions (see supplemental material). First, our thresholding scheme (see Section 3.3) may not detect all stretched-out regions, which may lead to remaining distortions. Additionally, some holes may still not be completed fully after the second stage (see Section 3.4),", + "bbox": [ + 496, + 779, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7916", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which results in over-smoothed regions after applying poisson reconstruction. Our scene representation does not decompose material from lighting, which bakes in shadows or bright lamps, that are generated from the diffusion model.", + "bbox": [ + 76, + 90, + 472, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 162, + 196, + 178 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have shown a method to generate textured 3D meshes from only text input. We use text-to-image 2D generators to create a sequence of images. The core insight of our method is a tailored viewpoint selection, that allows to create a 3D mesh with seamless geometry and compelling textures. Specifically, we lift the images into a 3D scene, by employing our alignment strategy that iteratively fuses all images into the mesh. Our output meshes represent arbitrary indoor scenes that can be rendered with classical rasterization pipelines. We believe our approach demonstrates an exciting application of large-scale 3D asset creation, that only requires text as input.", + "bbox": [ + 75, + 188, + 472, + 369 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 381, + 243, + 398 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was funded in part by Cisco Systems. It was also supported by the ERC Starting Grant Scan2CAD (804724) as well as the German Research Foundation (DFG) Research Unit \"Learning and Simulation in Visual Computing.\" We also thank Angela Dai for the video voice over.", + "bbox": [ + 76, + 404, + 468, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 500, + 173, + 516 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3D reconstruction, inpainting and generation. arXiv, 2022. 2", + "[2] Manuel Rey Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution $360^{\\circ}$ monocular depth estimation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5", + "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18187-18197, 2021. 2", + "[4] Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Iron-depth: Iterative refinement of single-view depth using surface normal and its uncertainty. In British Machine Vision Conference (BMVC), 2022. 3, 5", + "[5] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh Susskind. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2", + "[6] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2", + "[7] Shengqu Cai, Eric Ryan Chan, Songyou Peng, Mohamad Shahbazi, Anton Obukhov, Luc Van Gool, and Gordon Wetzstein. Diffdreamer: Consistent single-view perpetual view" + ], + "bbox": [ + 86, + 525, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generation with conditional diffusion models. In arXiV, 2022. 2", + "[8] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, L. Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository. ArXiv, abs/1512.03012, 2015. 1", + "[9] Kevin Chen, Christopher Bongsoo Choy, Manolis Savva, Angel X. Chang, Thomas A. Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. ArXiv, abs/1803.08495, 2018. 1, 2", + "[10] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[11] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 5, 7", + "[12] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. arXiv preprint arXiv:2212.04493, 2022. 2", + "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5", + "[14] Karan Desai, Gaurav Kaul, Zubin Trivadi Aysola, and Justin Johnson. Redcaps: Web-curated image-text data created by the people, for the people. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 2", + "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2", + "[16] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868-12878, 2020. 2", + "[17] Seth Forsgren* and Hayk Martiros*.iffusion - Stable diffusion for real-time music generation, https://riffusion.com/about, accessed 2023-03-06, 2022. 2", + "[18] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133, 2023. 1, 2, 4", + "[19] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M. Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. ArXiv, abs/2302.10109, 2023. 2", + "[20] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10686-10696, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7917", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Takayuki Hara and Tatsuya Harada. Enhancement of novel view synthesis using omnidirectional image completion. arXiv preprint arXiv:2203.09957, 2022. 5", + "[22] Benjamin Hepp, Matthias Nießner, and Otmar Hilliges. Plan3d: Viewpoint and trajectory optimization for aerial multi-view stereo reconstruction. ACM Transactions on Graphics (TOG), 38(1):1-17, 2018. 4", + "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2", + "[24] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2", + "[25] Jonathan Ho, Tim Salimans, Alexey A Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems. 2", + "[26] Ching-Yu Hsu, Cheng Sun, and Hwann-Tzong Chen. Moving in a 360 world: Synthesizing panoramic parallaxes from a single panorama. arXiv preprint arXiv:2106.10859, 2021. 5", + "[27] Huajian Huang, Yingshu Chen, Tianjian Zhang, and SaiKit Yeung. 360room: Real-time indoor roaming using geometry-aware $360^{\\circ}$ radiance fields. arXiv preprint arXiv:2208.02705, 2022. 5", + "[28] Qingqing Huang, Daniel S. Park, Tao Wang, Timo I. Denk, Andy Ly, Nanxin Chen, Zhengdong Zhang, Zhishuai Zhang, Jia Yu, C. Frank, Jesse Engel, Quoc V. Le, William Chan, and Weixiang Han. Noise2music: Text-conditioned music generation with diffusion models. ArXiv, abs/2302.03917, 2023. 2", + "[29] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 857-866. IEEE Computer Society, 2022. 1, 2, 5", + "[30] Zutao Jiang, Guangsong Lu, Xiaodan Liang, Jihua Zhu, Wei Zhang, Xiaojun Chang, and Hang Xu. 3d-togo: Towards text-guided cross-category 3d object generation. arXiv preprint arXiv:2212.01103, 2022. 2", + "[31] Lei Jin, Yanyu Xu, Jia Zheng, Junfei Zhang, Rui Tang, Shugong Xu, Jingyi Yu, and Shenghua Gao. Geometric structure based and regularized depth estimation from 360 indoor imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 889-898, 2020. 5", + "[32] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 2", + "[33] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, page 0, 2006. 5", + "[34] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. In ICML Workshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likelihood Models. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations. 2", + "[36] Shreyas Kulkarni, Peng Yin, and Sebastian Scherer. 360fusionnerf: Panoramic neural radiance fields with joint guidance. arXiv preprint arXiv:2209.14265, 2022. 5", + "[37] Blockade Labs. Blockade skybox, https://skybox.blockadelabs.com/, accessed 2023-03-04. 5, 7", + "[38] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022. 1, 2, 5, 7", + "[39] Gang Li, Heliang Zheng, Chaoyue Wang, Chang Li, Changwen Zheng, and Dacheng Tao. 3ddesigner: Towards photorealistic 3d object generation and editing with text-guided diffusion models. arXiv preprint arXiv:2211.14108, 2022. 2", + "[40] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In ECCV, 2022. 2, 3", + "[41] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 1, 2, 5", + "[42] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2021. 2, 3", + "[43] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 5", + "[44] Luke Melas-Kyriazi, C. Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: $360^{\\circ}$ reconstruction of any object from a single image. ArXiv, abs/2302.10663, 2023. 2", + "[45] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. 2", + "[46] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2", + "[47] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, 2020. 2", + "[48] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7918", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Gimin Nam, Mariem Khlifi, Andrew Rodriguez, Alberto Tono, Linqi Zhou, and Paul Guerrero. 3d-ldm: Neural implicit 3d shape generation with latent diffusion models. arXiv preprint arXiv:2212.00842, 2022. 2", + "[50] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 2", + "[51] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2", + "[52] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2", + "[53] OpenAI. Dalle: Introducing outpainting, https://openai.com/blog/dall-e-introducing-outpainting?utm_source=tldrnewsletter, accessed 2023-03-07. 5, 7", + "[54] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2065–2074, 2021. 2", + "[55] Nick St. Pierre. Additive prompting, https://twitter.com/nickfloats/status/1628796348446253057, accessed 2023-03-07. 5", + "[56] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 1, 2, 5", + "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5", + "[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1, 2, 5, 7", + "[59] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 5", + "[60] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2", + "[61] Xuanchi Ren and Xiaolong Wang. Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3553-3563, 2022. 2", + "[62] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2", + "[63] Mike Roberts, Debadeepta Dey, Anh Truong, Sudipta Sinha, Shital Shah, Ashish Kapoor, Pat Hanrahan, and Neel Joshi." + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Submodular trajectory optimization for aerial 3d scanning. In Proceedings of the IEEE International Conference on Computer Vision, pages 5324-5333, 2017. 4", + "[64] C. Rockwell, David F. Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14084-14093, 2021. 2", + "[65] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 1, 2, 5", + "[66] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2", + "[67] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1, 2", + "[68] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 5", + "[69] Aditya Sanghi, Hang Chu, J. Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18582-18592, 2021. 2", + "[70] Flavio Miguel Schneider, Zhijing Jin, and Bernhard Schölkopf. Mousai: Text-to-music generation with long-context latent diffusion. ArXiv, abs/2301.11757, 2023. 2", + "[71] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track. 2", + "[72] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. ArXiv, abs/2111.02114, 2021. 2", + "[73] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2", + "[74] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8025-8035, 2020. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7919", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 2", + "[76] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, and Yaniv Taigman. Text-to-4d dynamic scene generation. ArXiv, abs/2301.11280, 2023. 2", + "[77] Josef Sivic, Biliana K. Kaneva, Antonio Torralba, Shai Avidan, and William T. Freeman. Infinite images: Creating and exploring a large photorealistic virtual space. Proceedings of the IEEE, 98:1391-1407, 2008. 2", + "[78] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2", + "[79] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations. 2", + "[80] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2", + "[81] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 2", + "[82] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations. 2", + "[83] Alexandru Telea. An image inpainting technique based on the fast marching method. Journal of graphics tools, 9(1):23-34, 2004. 5", + "[84] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In Neural Information Processing Systems, 2021. 2", + "[85] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2", + "[86] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3825–3834, 2021. 2", + "[87] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. arXiv preprint arXiv:2212.08070, 2022. 2", + "[88] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022. 1, 2, 5" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[89] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2", + "[90] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7465-7475, 2019. 2", + "[91] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. ArXiv, abs/2212.11565, 2022. 2", + "[92] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with $360^{\\circ}$ views. arXiv e-prints, pages arXiv-2211, 2022. 2", + "[93] Jiale Xu, Jia Zheng, Yanyu Xu, Rui Tang, and Shenghua Gao. Layout-guided novel view synthesis from a single indoor panorama. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16438-16447, 2021. 5", + "[94] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2" + ], + "bbox": [ + 501, + 90, + 892, + 444 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "7920", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/ec345001-5311-431e-90bd-6b91f4b47f2a_model.json b/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/ec345001-5311-431e-90bd-6b91f4b47f2a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5d49402a8d4db3bfc682a1d5bbbe7d5b72b54b63 --- /dev/null +++ b/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/ec345001-5311-431e-90bd-6b91f4b47f2a_model.json @@ -0,0 +1,3260 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.131, + 0.87, + 0.152 + ], + "angle": 0, + "content": "Text2Room: Extracting Textured 3D Meshes from 2D Text-to-Image Models" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.18, + 0.838, + 0.218 + ], + "angle": 0, + "content": "Lukas Hollein\\(^{1*}\\) Ang Cao\\(^{2*}\\) Andrew Owens\\(^{2}\\) Justin Johnson\\(^{2}\\) Matthias Nießner\\(^{1}\\) \n\\(^{1}\\)Technical University of Munich \n\\(^{2}\\)University of Michigan" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.251, + 0.7, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.422, + 0.508, + 0.435 + ], + "angle": 0, + "content": "(a) 3D Mesh Generation from Text" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.252, + 0.895, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.421, + 0.893, + 0.437 + ], + "angle": 0, + "content": "(b) Rendered Image + Mesh" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.445, + 0.893, + 0.488 + ], + "angle": 0, + "content": "Figure 1. Textured 3D mesh generation from text prompts. We generate textured 3D meshes from a given text prompt using 2D text-to-image models. (a) The scene is iteratively created from different viewpoints (marked in blue). (b) Our generated mesh contains compelling textures and geometry. We remove the ceiling in the top-down views for better visualization of the scene layout." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.525, + 0.314, + 0.541 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.474, + 0.846 + ], + "angle": 0, + "content": "We present Text2Room1, a method for generating room-scale textured 3D meshes from a given text prompt as input. To this end, we leverage pre-trained 2D text-to-image models to synthesize a sequence of images from different poses. In order to lift these outputs into a consistent 3D scene representation, we combine monocular depth estimation with a text-conditioned inpainting model. The core idea of our approach is a tailored viewpoint selection such that the content of each image can be fused into a seamless, textured 3D mesh. More specifically, we propose a continuous alignment strategy that iteratively fuses scene frames with the existing geometry to create a seamless mesh. Unlike existing works that focus on generating single objects [56, 41] or zoom-out trajectories [18] from text, our method generates complete 3D scenes with multiple objects and explicit 3D geometry. We evaluate our approach using qualitative and quantitative metrics, demonstrating it as the first method to generate room-scale 3D geometry with compelling textures from only text as input." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.497, + 0.631, + 0.512 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Mesh representations of 3D scenes are a crucial component for many applications, from AR/VR asset creation to computer graphics, yet creating these 3D assets remains a painstaking process that requires considerable expertise. In the 2D domain, recent works have successfully created high-quality images from text using generative models, such as diffusion models [65, 58, 67]. These methods significantly reduce the barriers to creating images that contain a user's desired content, effectively helping towards the democratization of content creation. An emerging line of work has sought to apply similar methods to create 3D models from text [9, 56, 29, 41, 38], yet existing approaches come with a number of significant limitations and lack the generality of 2D text-to-image models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "One of the core challenges of generating 3D models is coping with the lack of available 3D training data, as 3D datasets are vastly smaller than those available in many other applications, such as 2D image synthesis. For example, methods that directly use 3D supervision, such as Chen et al. [9], are often limited to datasets of simple shapes, such as ShapeNet [8]. To address these data limitations, recent methods [56, 29, 41, 38, 88] lift the expressive power of 2D text-to-image models into 3D by formulating 3D generation as an iterative optimization problem in the image domain. This allows them to generate 3D ob" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.223, + 0.888 + ], + "angle": 0, + "content": "* joint first authorship" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.42, + 0.9 + ], + "angle": 0, + "content": "1https://lukashoel.github.io/text-to-room" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.42, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7909" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "jects stored in a radiance field representation, demonstrating the ability to generate arbitrary (neural) shapes from text. However, these methods cannot easily be extended to create room-scale 3D structure and texture. The challenge of generating large scenes is ensuring that the generated output is dense and coherent across outward-facing viewpoints, and that these views contain all of the required structures, such as walls, floors, and furniture. Additionally, a mesh remains a desired representation for many end-user tasks, such as rendering on commodity hardware (which requires an additional conversion step as presented in Lin et al. [41])." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.409 + ], + "angle": 0, + "content": "To address these shortcomings, we propose a method that extracts scene-scale 3D meshes from off-the-shelf 2D text-to-image models. Our method iteratively generates a scene through inpainting and monocular depth estimation. We produce an initial mesh by generating an image from text, and backproject it into 3D using a depth estimation model. Then, we iteratively render the mesh from novel viewpoints. From each one, we fill in holes in the rendered images via inpainting, then fuse the generated content into the mesh (Fig. 1a)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.473, + 0.59 + ], + "angle": 0, + "content": "Our iterative generation scheme has two important design considerations: how we choose the viewpoints, and how we merge generated scene content with the existing mesh. We first select viewpoints from predefined trajectories that will cover large amounts of scene content, then adaptively select viewpoints that close remaining holes. When merging generated content with the mesh, we align the two depth maps to create smooth transitions, and remove parts of the mesh that contain distorted textures. Together, these decisions lead to large, scene-scale 3D meshes with compelling textures and consistent geometry (Fig. 1b), that can represent a wide range of rooms." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.591, + 0.325, + 0.604 + ], + "angle": 0, + "content": "To summarize, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.606, + 0.469, + 0.637 + ], + "angle": 0, + "content": "- Generating 3D meshes of room-scale indoor scenes with compelling textures and geometry from any text input." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.637, + 0.469, + 0.711 + ], + "angle": 0, + "content": "- A method that leverages 2D text-to-image models and monocular depth estimation to lift frames into 3D in an iterative scene generation. Our proposed depth alignment and mesh fusion steps, enable us to create seamless and undistorted geometry and textures." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.712, + 0.469, + 0.772 + ], + "angle": 0, + "content": "- A two-stage tailored viewpoint selection that samples camera poses from optimal positions to first create the room layout and furniture and then close any remaining holes, creating a watertight mesh." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.606, + 0.469, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.785, + 0.22, + 0.801 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.469, + 0.871 + ], + "angle": 0, + "content": "Text-based Generation has seen significant advances due to large-scale image-text datasets [73, 72, 14, 71] and scalable generative model architectures [16, 66, 60, 32], enabling synthesis of novel images from text [20, 3, 54]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Recently, diffusion models [78, 23, 80, 81, 82] achieved impressive results on image synthesis [15, 65, 67, 50, 58]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "through improvements like latent space denoising [65, 84], faster sampling [23, 79, 52, 34], and better guidance [24]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.894, + 0.228 + ], + "angle": 0, + "content": "In particular, text-to-image methods like Stable Diffusion [65], Imagen [67], GLIDE [50] and DALL-E 2 [58] yield diverse, high-fidelity, and controllable [6, 94] outputs. Text-based generation has been extended to other modalities including audio [35, 17, 28, 70], video [75, 91, 85, 25], and 4D fields [76]. We use text-to-image models by lifting their generated output into complete 3D scene meshes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.233, + 0.892, + 0.293 + ], + "angle": 0, + "content": "Text-to-3D. Several methods use 3D data for supervised training of text-to-3D models [9, 51, 5]; however this direction remains challenging due to the lack of large-scale aligned datasets of text and 3D." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Alternative approaches use 2D vision-language models like CLIP [57] to create 3D content by formulating the generation as an optimization problem in the image domain [86, 29, 38, 48, 30] or as object alignment [69]. Related methods refine existing 3D input through text guidance in a similar fashion [46, 10, 87, 62]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.892, + 0.491 + ], + "angle": 0, + "content": "Recent methods [56, 41, 45, 88, 44] combine large text-to-image diffusion models [65, 67] and neural radiance fields [47] to generate 3D objects without training. Other approaches train custom diffusion models on a similar text-to-3D task [39, 49, 12]. In contrast, we use a fixed text-to-image model and extract a 3D mesh representing entire scenes of many objects and structural elements like walls." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.892, + 0.693 + ], + "angle": 0, + "content": "3D-Consistent View Synthesis from a Single Image. Several methods have been proposed that perform novel-view-synthesis from a single image [64, 90, 74, 61, 19]. Others optimize a neural 3D representation of an object, that can be viewed from arbitrary novel view points [92, 89, 1]. Another line of work performs perpetual view generation [77, 42, 40, 7], synthesizing videos via a render-refine-repeat pattern from a single RGB image that depict a scene along a forward-facing camera trajectory. In very recent concurrent work, Fridman et al. [18] create 3D scenes from text, but focus on this type of 3D-consistent \"zoom-out\" video generation. Instead, we generate complete, textured 3D room geometry from arbitrary trajectories." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.709, + 0.591, + 0.724 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our method creates a textured 3D mesh of a complete scene from text input. To this end, we continuously fuse generated frames from a 2D text-to-image model at different poses into a joint 3D mesh, creating the scene over time. The core idea of our approach is a two-stage tailored viewpoint selection, that first generates the scene layout and objects and then closes remaining holes in the 3D geometry (Section 3.4). We visualize this workflow in Figure 2. For each pose in both stages, we apply an iterative scene generation scheme to update the mesh (Section 3.1). We first align each frame with the existing geometry with a depth align-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7910" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.657, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.174, + 0.461, + 0.189 + ], + "angle": 0, + "content": "(a) Scene Generation Stage" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.089, + 0.891, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.174, + 0.866, + 0.189 + ], + "angle": 0, + "content": "(b) Scene Completion Stage" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.196, + 0.894, + 0.266 + ], + "angle": 0, + "content": "Figure 2. Method overview. We iteratively create a textured 3D mesh in two stages. (a) First, we sample predefined poses and text to generate the complete scene layout and furniture. Each new pose (marked in green) adds newly generated geometry to the mesh (depicted by green triangles) in an iterative scene generation scheme (see Figure 3 for details). Blue poses/triangles denote viewpoints that created geometry in a previous iteration. (b) Second, we fill in the remaining unobserved regions by sampling additional poses (marked in red) after the scene layout is defined." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.291, + 0.481, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.499, + 0.47, + 0.584 + ], + "angle": 0, + "content": "Figure 3. Iterative scene generation. For each new pose, we render the current mesh to obtain partial RGB and depth renderings. We complete both, utilizing respective inpainting models and the text prompt. Next, we perform depth alignment (see Section 3.2) and mesh filtering (see Section 3.3) to obtain an optimal next mesh patch, that is finally fused with the existing geometry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.47, + 0.636 + ], + "angle": 0, + "content": "ment strategy (Section 3.2). Next, we triangulate and filter the novel content to merge it into the mesh (Section 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.645, + 0.345, + 0.659 + ], + "angle": 0, + "content": "3.1. Iterative 3D Scene Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.82 + ], + "angle": 0, + "content": "Our scene is represented as a mesh \\(\\mathcal{M} = (\\mathcal{V},\\mathcal{C},\\mathcal{S})\\) where the vertices \\(\\mathcal{V}\\in \\mathbb{R}^{N\\times 3}\\), vertex colors \\(\\mathcal{C}\\in \\mathbb{R}^{N\\times 3}\\) and the face set \\(S\\in \\mathbb{N}_0^{M\\times 3}\\) are generated over time. Input to our method is a set of arbitrary text prompts \\(\\{P_t\\}_{t = 1}^T\\) that corresponds to our selected poses \\(\\{E_t\\}_{t = 1}^T\\in \\mathbb{R}^{3\\times 4}\\) in both stages. Inspired by recent methods [42, 40], we iteratively build up the scene, following a render-refine-repeat pattern. We summarize this iterative scene generation process in Figure 3. Formally, for each step of generation \\(t\\), we first render the current scene from a novel viewpoint:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.829, + 0.469, + 0.845 + ], + "angle": 0, + "content": "\\[\nI _ {t}, d _ {t}, m _ {t} = r \\left(\\mathcal {M} _ {t}, E _ {t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(r\\) is a classical rasterization function without shading, \\(I_{t}\\) is the rendered image, \\(d_{t}\\) the rendered depth and \\(m_{t}\\) the image-space mask, that marks pixels without observed" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.291, + 0.892, + 0.321 + ], + "angle": 0, + "content": "content. We then use a fixed text-to-image model \\(\\mathcal{F}_{t2i}\\) to inpaint unobserved pixels according to the text prompt:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.333, + 0.892, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {t} = \\mathcal {F} _ {t 2 i} \\left(I _ {t}, m _ {t}, P _ {t}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.365, + 0.892, + 0.395 + ], + "angle": 0, + "content": "Next, we inpaint unobserved depth by applying a monocular depth estimator \\(\\mathcal{F}_d\\) in our depth alignment (see Section 3.2):" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.41, + 0.892, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\hat {d} _ {t} = \\text {p r e d i c t - a n d - a l i g n} \\left(\\mathcal {F} _ {d}, I _ {t}, d _ {t}, m _ {t}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.438, + 0.892, + 0.47 + ], + "angle": 0, + "content": "Finally, we combine the novel content \\(\\{\\hat{I}_t,\\hat{d}_t,m_t\\}\\) with the existing mesh by our fusion scheme (see Section 3.3):" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.481, + 0.892, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {t + 1} = f u s e \\left(\\mathcal {M} _ {t}, \\hat {I} _ {t}, \\hat {d} _ {t}, m _ {t}, E _ {t}\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.707, + 0.53 + ], + "angle": 0, + "content": "3.2. Depth Alignment Step" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.659 + ], + "angle": 0, + "content": "To lift a 2D image \\(I\\) into 3D, we predict the per-pixel depth. To correctly combine old and new content, it is necessary that both align with each other. In other words, similar regions in a scene like walls or furniture should be placed at similar depth. However, directly using the predicted depth for backprojection leads to hard cuts and discontinuities in the 3D geometry, since the depth is inconsistent in scale between subsequent viewpoints (see Figure 7a)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.72 + ], + "angle": 0, + "content": "To this end, we perform depth alignment in two-stages. First, we use a state-of-the-art depth inpainting network [4] that takes ground-truth depth \\( d \\) for known parts in the image as input and aligns the prediction to it: \\( \\hat{d}_p = \\mathcal{F}_d(I,d) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.721, + 0.892, + 0.767 + ], + "angle": 0, + "content": "Inspired by Liu et al. [42] we then improve the result by optimizing for scale and shift parameters \\(\\gamma, \\beta \\in \\mathbb{R}\\), aligning predicted and rendered disparity in the least squares sense:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.779, + 0.892, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\gamma , \\beta} \\quad \\left\\| m \\odot \\left(\\frac {\\gamma}{\\hat {d} _ {p}} + \\beta - \\frac {1}{d}\\right) \\right\\| ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.834, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where we mask out unobserved pixels via \\( m \\). We can then extract the aligned depth as \\( \\hat{d} = (\\frac{\\gamma}{\\hat{d}_p} + \\beta)^{-1} \\). Finally, we smooth \\( \\hat{d} \\) by applying a \\( 5 \\times 5 \\) Gaussian kernel at the mask edges (see supplemental material for more details)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7911" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.216, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.089, + 0.353, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.089, + 0.453, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.08, + 0.215, + 0.463, + 0.229 + ], + "angle": 0, + "content": "(a) Pixel Triangulation (b) Face Filtering (c) Mesh Fusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.233, + 0.468, + 0.316 + ], + "angle": 0, + "content": "Figure 4. Visualization of our mesh fusion step. (a) We triangulate an image, such that 4 neighboring pixels (orange dots) create two faces. (b) We filter a face (marked in red), if its surface normal forms a small grazing angle with the viewing direction or if any edge in world space is too long. (c) We fuse the remaining faces (marked in green) with the existing geometry (marked in blue)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.343, + 0.251, + 0.359 + ], + "angle": 0, + "content": "3.3. Mesh Fusion Step" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.469, + 0.413 + ], + "angle": 0, + "content": "At each step, we insert new content \\(\\{\\hat{I}_t,\\hat{d}_t,m_t\\}\\) into the scene. For that, we first backproject the image-space pixels into a world-space point cloud:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.425, + 0.469, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {t} = \\left\\{E _ {t} ^ {- 1} K ^ {- 1} \\cdot \\hat {d} _ {t} [ u, v ] \\cdot (u, v, 1) ^ {T} \\right\\} _ {u = 0, v = 0} ^ {W, H}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.469, + 0.581 + ], + "angle": 0, + "content": "where \\(K \\in \\mathbb{R}^{3 \\times 3}\\) are the camera intrinsics and \\(W, H\\) are image width and height, respectively. We then use a simple triangulation scheme (Figure 4a), where each four neighboring pixels \\(\\{(u, v), (u + 1, v), (u, v + 1), (u + 1, v + 1)\\}\\) in the image form two triangles. Since the estimated depth is noisy, this naive triangulation creates stretched out 3D geometry (see Figure 7b). To alleviate this problem, we propose two filters that remove stretched out faces (Figure 4b)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.469, + 0.643 + ], + "angle": 0, + "content": "First, we filter faces based on their edge length. We remove a face if the Euclidean distance of any face edge is larger than a threshold \\(\\delta_{edge}\\). Second, we filter faces based on the angle between surface normal and viewing direction:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.66, + 0.469, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = \\left\\{\\left(i _ {0}, i _ {1}, i _ {2}\\right) \\mid n ^ {T} v > \\delta_ {s n} \\right\\} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.469, + 0.794 + ], + "angle": 0, + "content": "where \\(S\\) is the face set, \\((i_0,i_1,i_2)\\) are the vertex indices of the triangle, \\(\\delta_{sn}\\) is the threshold, \\(n\\in \\mathbb{R}^3\\) is the normalized face normal, and \\(v\\in \\mathbb{R}^3\\) is the normalized view direction in world space from the camera center towards the average pixel location from which the triangle originated. This avoids creating texture for large regions of the mesh from a comparatively small number of pixels from an image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Finally, we fuse together the newly generated mesh patch and the existing geometry (Figure 4c). All faces that are backprojected from pixels falling into the inpainting mask \\( m_{t} \\) are stitched together with their neighboring faces, which are already part of the mesh. Precisely, we continue the triangulation scheme at all edges of \\( m_{t} \\), but use the existing vertex positions of \\( \\mathcal{M}_t \\) to create the corresponding faces." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.772, + 0.107 + ], + "angle": 0, + "content": "3.4. Two-Stage Viewpoint Selection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.252 + ], + "angle": 0, + "content": "A key part of our method is the choice of text prompts and camera poses from which the scene is synthesized. Users can in principle choose these inputs arbitrarily to create any desired indoor scene. However, the generated scene can degenerate and contain stretch and hole artifacts, if poses are chosen carelessly (see Figure 7 and supplemental material). To this end, we propose a two-stage viewpoint selection strategy, that samples each next camera pose from optimal positions and refines empty regions subsequently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.892, + 0.407 + ], + "angle": 0, + "content": "Generation Stage. In the first stage, we create the main parts of the scene, including the general layout and furniture. We subsequently render predefined trajectories in different directions that eventually cover the whole room. We found generation works best, if each trajectory starts off from a viewpoint with mostly unobserved regions. This generates the outline of the next chunk, while still being connected to the rest of the scene (e.g., see Figure 3). Then, we complete the 3D structure of that chunk by moving and rotating into it subsequently until the end of the trajectory." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.892, + 0.529 + ], + "angle": 0, + "content": "Additionally, we ensure an optimal observation distance for each pose. We translates camera positions \\( T_0 \\in \\mathbb{R}^3 \\) along the look-at direction \\( L \\in \\mathbb{R}^3 \\) uniformly: \\( T_{i+1} = T_i - 0.3L \\). We stop if the mean rendered depth is larger than 0.1 or discard the camera after 10 steps. This avoids views too close to existing geometry. For example, the green pose in Figure 2a is moved back as far as possible into the existing geometry such that it views most of the empty floor region." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.652 + ], + "angle": 0, + "content": "We create closed room layouts following this principle, by choosing trajectories that generate the next chunks in a circular motion, roughly centered around the origin. We found it helpful to discourage the text-to-image generator from generating furniture in unwanted regions by engineering the text prompts accordingly. For example, for poses looking at the floor or ceiling, we choose text prompts that only contain the words \"floor\" or \"ceiling\", respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Completion Stage. After the first stage, the scene layout and furniture is defined. However, it is impossible to choose sufficient poses \\( a \\)-priori. Since the scene is generated on-the-fly, the mesh contains holes that were not observed by any camera (see Figure 7c). We complete the scene by sampling additional poses \\( a \\)-posteriori, looking at those holes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Inspired by trajectory optimization [22, 63], we voxelize the scene into dense uniform cells. We sample random poses in each cell, discarding those being too close to existing geometry. We select one pose per cell that views most unobserved pixels (e.g., see the red poses in Figure 2b)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Next, we inpaint the scene from all chosen camera poses following Section 3.1. Similar to Fridman et al. [18], we observe it is important to clean the inpainting masks, because our text-to-image generator can generate better results for large connected regions. Thus, we first inpaint small holes" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7912" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.152 + ], + "angle": 0, + "content": "with a classical inpainting algorithm [83] and dilate the remaining holes. We additionally remove all faces that fall into the dilated region and are close to the rendered depth. Please see the supplemental material for more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.47, + 0.228 + ], + "angle": 0, + "content": "Finally, we run Poisson surface reconstruction [33] on the scene mesh. This closes any remaining holes after completion and smoothens out discontinuities. The result is a watertight mesh of the generated scene, that can be rendered with classical rasterization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.243, + 0.166, + 0.259 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.47, + 0.501 + ], + "angle": 0, + "content": "Implementation Details. We implement mesh rasterization and fusion with Pytorch3D [59]. As our text-to-image model \\(\\mathcal{F}_{t2i}\\), we utilize a Stable Diffusion [65] model, that is finetuned on the image inpainting task, using additional mask input. We generate a single inpainting proposal and employ a state-of-the-art guided diffusion sampler [43]. As our monocular depth estimator \\(\\mathcal{F}_d\\), we employ an Iron-Depth [4] model, that is trained on indoor scenes from the ScanNet dataset [13] and augment it for depth inpainting according to Bae et al. [4]. We set \\(\\delta_{edge} = 0.1\\) and \\(\\delta_{sn} = 0.1\\) in all our experiments. During generation, we use 20 different trajectories with 10 frames each sampled between the respective start and end poses. We construct prompts using the guidelines suggested by Pierre [55]. Creating one scene takes approximately 50 minutes on one RTX 3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.564 + ], + "angle": 0, + "content": "Baselines. To the best of our knowledge, there are no direct baselines that generate textured 3D room geometry from text. We compare against four related methods (please see the supplemental material for more details about baselines)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.565, + 0.469, + 0.61 + ], + "angle": 0, + "content": "- PureClipNeRF [38]: We compare against text-to-3D methods for generating objects [56, 41, 29, 38, 88] and choose Lee et al. [38] as open-source representative." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.611, + 0.469, + 0.655 + ], + "angle": 0, + "content": "- Outpainting [58, 53]: We combine outpainting from a Stable Diffusion [65] model with depth estimation and triangulation to create a mesh from an enlarged viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.656, + 0.469, + 0.746 + ], + "angle": 0, + "content": "- Text2Light [11]: We generate RGB panoramas from text using Chen et al. [11]. Estimating 3D mesh structure from a panorama is difficult. Related approaches estimate room layout [93], perform view synthesis [36, 26, 21, 27] or predict \\(360^{\\circ}\\) depth [2, 31]. We perform depth prediction and subsequently apply our mesh fusion step." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.747, + 0.469, + 0.791 + ], + "angle": 0, + "content": "- Blockade [37]: We apply Blockade [37], which uses a text-to-image diffusion model to produce more expressive RGB panoramas. We then extract the mesh similarly." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.791 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Evaluation Metrics. The generated 3D geometry is evaluated both quantitatively and qualitatively. We calculate CLIP Score (CS) [57] and Inception Score (IS) [68] on RGB renderings of the respective scenes. Additionally, we conduct a user study and ask \\( n = 61 \\) users to score Perceptual Quality (\\( PQ \\)) and 3D Structure Completeness (\\( 3DS \\)) of the whole scene on a scale of 1-5." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.892, + 0.26 + ], + "angle": 0, + "content": "
Method2D MetricsUser Study
CS ↑IS ↑PQ ↑3DS ↑
PureClipNeRF [38]24.061.262.342.38
Outpainting [58, 53]23.101.602.902.58
Text2Light [11]+Ours25.992.212.822.97
Blockade [37]+Ours26.292.133.353.36
Ours w/o alignment26.731.783.122.96
Ours w/o stretch removal27.721.863.283.75
Ours w/o completion27.972.183.723.87
Ours28.022.314.014.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.27, + 0.892, + 0.327 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison. We report 2D metrics and user study results, including: Clip Score (CS), Inception Score (IS), Perceptual Quality (PQ), and 3D Structure Completeness (3DS). Our method creates scenes with the highest quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.353, + 0.684, + 0.37 + ], + "angle": 0, + "content": "4.1. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.892, + 0.56 + ], + "angle": 0, + "content": "We show top-down views into the scene and RGB renderings from within for our method and baselines in Figure 6. We show additional results of our method in Figure 5. PureClipNeRF [38] creates the key objects of the given text prompt, but does not create a complete 3D structure with floor, walls and ceilings. Outpainting [58, 53] creates high-detail textures, but projection from a single viewpoint creates holes due to occlusion and hinders the creation of complete 3D geometry. Text2Light [11] and Blockade [37] both create a high-detail \\(360^{\\circ}\\) view of a complete scene, but occlusions that cannot be resolved from a single panoramic viewpoint lead to holes in the extracted 3D geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.892, + 0.712 + ], + "angle": 0, + "content": "In contrast, our approach creates high-detail textures and geometry, that are fused into a complete 3D scene mesh without holes. The resulting scenes contain flat floors, walls and ceilings, as well as 3D object geometry distributed throughout the scene. When specifying text prompts with a huge variety, the resulting scene contains a diverse set of objects. Please see the supplemental material for more scenes, animated results, intermediate outputs of our baselines (such as the panoramic images) as well as top-down views of meshes, that contain the reconstructed ceilings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.694, + 0.741 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We show quantitative results averaged over multiple scenes in Table 1. We render 60 images from novel viewpoints for each scene to calculate the 2D metrics. We present users with multiple top-down views and renderings for each scene and let them rate each method individually (no side-by-side comparison). Stretched-out geometry and holes in the 3D geometry lead to lower scores for the baselines in all image-based metrics. Our approach achieves the highest scores, because the renderings are complete from arbitrary novel poses, satisfy the given text-prompt and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7913" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.089, + 0.437, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.089, + 0.641, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.09, + 0.844, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.179, + 0.641, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.179, + 0.844, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.271, + 0.821, + 0.285 + ], + "angle": 0, + "content": "Editorial Style Photo, Coastal Bathroom, Clawfoot Tub, Seashell, Wicker, Mosaic Tile, Blue and White" + }, + { + "type": "image", + "bbox": [ + 0.165, + 0.288, + 0.4, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.286, + 0.641, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.286, + 0.844, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.376, + 0.641, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.376, + 0.844, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.467, + 0.826, + 0.481 + ], + "angle": 0, + "content": "A living room with a lit furnace, couch, and cozy curtains, bright lamps that make the room look well-lit" + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.483, + 0.427, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.482, + 0.641, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.482, + 0.844, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.571, + 0.641, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.571, + 0.844, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.663, + 0.846, + 0.677 + ], + "angle": 0, + "content": "Editorial Style Photo, Modern Living Room, Large Window, Leather, Glass, Metal, Wood Paneling, Apartment" + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.68, + 0.402, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.679, + 0.641, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.678, + 0.844, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.767, + 0.641, + 0.855 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.767, + 0.844, + 0.855 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.859, + 0.846, + 0.873 + ], + "angle": 0, + "content": "Editorial Style Photo, Modern Nursery, Table Lamp, Rocking Chair, Tree Wall Decal, Wood, Cotton, Faux Fur" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.874, + 0.892, + 0.915 + ], + "angle": 0, + "content": "Figure 5. 3D scene generation results of our method. We show color and shaded geometry renderings from generated scenes with corresponding text prompts. Our method synthesizes realistic meshes satisfying text descriptions. We remove the ceiling in the top-down view for better visualization of the scene layout." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7914" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.885, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.281, + 0.85, + 0.296 + ], + "angle": 0, + "content": "Editorial Style Photo, Industrial Home Office, Steel Shelves, Concrete, Metal, Edison Bulbs, Exposed Ductwork" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.297, + 0.885, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.477, + 0.66, + 0.491 + ], + "angle": 0, + "content": "a bedroom with a king-size bed and a large wardrobe" + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.492, + 0.228, + 0.507 + ], + "angle": 0, + "content": "PureClipNeRF [38]" + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.492, + 0.392, + 0.507 + ], + "angle": 0, + "content": "Outpainting [58, 53]" + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.492, + 0.558, + 0.506 + ], + "angle": 0, + "content": "Text2Light [11]+Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.492, + 0.714, + 0.506 + ], + "angle": 0, + "content": "Blockade [37]+Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.492, + 0.822, + 0.505 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.508, + 0.895, + 0.564 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of our method and baselines. PureClipNeRF [38] cannot produce immersive scenes with floors and walls. Outpainting [58, 53] does not produce 3D consistent scenes. Text2Light [11] and Blockade [37] both have holes due to occlusions. In contrast, our method creates complete meshes without holes and high details. We remove the ceiling in the top-down view for better visualization of the scene layout." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.471, + 0.634 + ], + "angle": 0, + "content": "contain high-resolution image features. Users prefer our method, which highlights the quality of our accurate and complete geometry, as well as the RGB texture." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.644, + 0.188, + 0.658 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.728 + ], + "angle": 0, + "content": "The key ingredients of our method are depth alignment (Section 3.2), mesh fusion (Section 3.3) and the two-stage viewpoint selection (Section 3.4). We demonstrate the importance of each component in Figure 7 and Table 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.47, + 0.837 + ], + "angle": 0, + "content": "Depth alignment creates seamless scenes. Monocular depth predictions from subsequent frames can be inconsistent in scale. This leads to disconnected components in the mesh that are backprojected from multiple viewpoints (see Figure 7a). Our depth alignment strategy allows fusing multiple frames into a seamless mesh, eventually creating a complete scene with flat floors, walls, ceilings and no holes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Stretch removal creates undistorted scene geometry. During mesh fusion, we update the scene geometry with the contents of the next frame. Due to noisy depth prediction, the objects become stretched out, if they are observed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.664 + ], + "angle": 0, + "content": "from small grazing angles. Thus, we propose two filters (edge length and surface normal thresholds) that alleviate this issue. Instead of baking in stretched-out geometry (see Figure 7b), we disregard the corresponding faces and let the object be completed from a more suitable, later viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.893, + 0.805 + ], + "angle": 0, + "content": "Two-stage generation creates complete scenes. Our approach chooses camera poses in two stages to create a complete scene without holes. After generating the scene from predefined trajectories, the scene still contains some holes (see Figure 7c). Because the scene is built-up over time, it is impossible to choose camera poses \\(a\\)-priori, that view all unobserved regions. To this end, our completion stage samples poses \\(a\\)-posteriori to refine those regions. The resulting mesh is watertight and contains no holes (see Figure 7d)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.81, + 0.833 + ], + "angle": 0, + "content": "4.4. Spatially Varying Scene Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Our method can be applied to generate a scene as the combination of multiple text prompts. Specifically, we use separate text prompts for different poses, crafting a set of trajectories that spatially combines scene descriptions. This" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7915" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.088, + 0.287, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.088, + 0.461, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.092, + 0.679, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.089, + 0.88, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.231, + 0.286, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.231, + 0.483, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.231, + 0.681, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.231, + 0.879, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.355, + 0.645, + 0.369 + ], + "angle": 0, + "content": "living room, couches, curtains, lit furnace, lamps" + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.37, + 0.265, + 0.384 + ], + "angle": 0, + "content": "(a) Ours w/o alignment" + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.37, + 0.48, + 0.383 + ], + "angle": 0, + "content": "(b) Ours w/o stretch removal" + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.37, + 0.663, + 0.384 + ], + "angle": 0, + "content": "(c) Ours w/o completion" + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.37, + 0.827, + 0.384 + ], + "angle": 0, + "content": "(d) Ours (full)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.385, + 0.894, + 0.442 + ], + "angle": 0, + "content": "Figure 7. Ablation study on the key components of our method. Without depth alignment (see Section 3.2), different parts of the scene are disconnected and do not fuse into a seamless mesh. Without edge and surface normal thresholds (see Section 3.3), many faces are stretched out unnaturally. Without completion (see Section 3.4), the mesh has holes in remaining unobserved regions. Our full pipeline creates complete, high-resolution scenes. We remove the ceiling in the top-down view for better visualization of the scene layout." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.463, + 0.26, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.463, + 0.449, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.555, + 0.452, + 0.584 + ], + "angle": 0, + "content": "kitchen, dinner table, dishwashers, ovens, countertops living room, lit furnace, couch, curtains" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.585, + 0.241, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.585, + 0.448, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.677, + 0.39, + 0.704 + ], + "angle": 0, + "content": "bathroom, shower, bathtub bedroom, king-size bed, wardrobe" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.762 + ], + "angle": 0, + "content": "Figure 8. Spatially varying scene generation. Our method can create rooms with multiple parts by prompt mixing. We use separate prompts for cameras viewing different parts of the scene. This is a controllable way to create rooms from multiple descriptions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "can be desired to avoid repeating elements in a complete scene (e.g., multiple couches could be spread out over the whole room when using the same prompt for every camera). Instead, users can specify different object positions through different camera poses and text prompts. It can also be used to design a house comprised of multiple rooms, each with a different type (e.g., a living room that leads to a kitchen)." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.464, + 0.705, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.463, + 0.883, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.607 + ], + "angle": 0, + "content": "Figure 9. Scene generation with layout guidance. Our method can generate scenes from layout guidance. Left: we describe objects with different prompts for cameras facing at different directions. Right: the generated part of the room." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.89, + 0.665 + ], + "angle": 0, + "content": "We show results that combine multiple text prompts in Figure 8 and Figure 9." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.666, + 0.892, + 0.743 + ], + "angle": 0, + "content": "We note that the layout can only be partially controlled by the camera poses, since scene generation can create chunks with larger or smaller extent. We believe this demonstrates an exciting application of our method, that can be further explored in future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.627, + 0.77 + ], + "angle": 0, + "content": "4.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Our approach allows to generate 3D room geometry from arbitrary text prompts that are highly detailed and contain consistent geometry. Nevertheless, our method can still fail under certain conditions (see supplemental material). First, our thresholding scheme (see Section 3.3) may not detect all stretched-out regions, which may lead to remaining distortions. Additionally, some holes may still not be completed fully after the second stage (see Section 3.4)," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7916" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.152 + ], + "angle": 0, + "content": "which results in over-smoothed regions after applying poisson reconstruction. Our scene representation does not decompose material from lighting, which bakes in shadows or bright lamps, that are generated from the diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.164, + 0.197, + 0.179 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.189, + 0.473, + 0.371 + ], + "angle": 0, + "content": "We have shown a method to generate textured 3D meshes from only text input. We use text-to-image 2D generators to create a sequence of images. The core insight of our method is a tailored viewpoint selection, that allows to create a 3D mesh with seamless geometry and compelling textures. Specifically, we lift the images into a 3D scene, by employing our alignment strategy that iteratively fuses all images into the mesh. Our output meshes represent arbitrary indoor scenes that can be rendered with classical rasterization pipelines. We believe our approach demonstrates an exciting application of large-scale 3D asset creation, that only requires text as input." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.382, + 0.245, + 0.399 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.405, + 0.47, + 0.489 + ], + "angle": 0, + "content": "This work was funded in part by Cisco Systems. It was also supported by the ERC Starting Grant Scan2CAD (804724) as well as the German Research Foundation (DFG) Research Unit \"Learning and Simulation in Visual Computing.\" We also thank Angela Dai for the video voice over." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.501, + 0.174, + 0.517 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.526, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3D reconstruction, inpainting and generation. arXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.582, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[2] Manuel Rey Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution \\(360^{\\circ}\\) monocular depth estimation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.638, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18187-18197, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.693, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[4] Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Iron-depth: Iterative refinement of single-view depth using surface normal and its uncertainty. In British Machine Vision Conference (BMVC), 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.749, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[5] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh Susskind. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.818, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[7] Shengqu Cai, Eric Ryan Chan, Songyou Peng, Mohamad Shahbazi, Anton Obukhov, Luc Van Gool, and Gordon Wetzstein. Diffdreamer: Consistent single-view perpetual view" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.526, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "generation with conditional diffusion models. In arXiV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[8] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, L. Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository. ArXiv, abs/1512.03012, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.193, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[9] Kevin Chen, Christopher Bongsoo Choy, Manolis Savva, Angel X. Chang, Thomas A. Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. ArXiv, abs/1803.08495, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[10] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[11] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[12] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. arXiv preprint arXiv:2212.04493, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[14] Karan Desai, Gaurav Kaul, Zubin Trivadi Aysola, and Justin Johnson. Redcaps: Web-curated image-text data created by the people, for the people. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[16] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868-12878, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[17] Seth Forsgren* and Hayk Martiros*.iffusion - Stable diffusion for real-time music generation, https://riffusion.com/about, accessed 2023-03-06, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[18] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[19] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M. Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. ArXiv, abs/2302.10109, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[20] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10686-10696, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7917" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[21] Takayuki Hara and Tatsuya Harada. Enhancement of novel view synthesis using omnidirectional image completion. arXiv preprint arXiv:2203.09957, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[22] Benjamin Hepp, Matthias Nießner, and Otmar Hilliges. Plan3d: Viewpoint and trajectory optimization for aerial multi-view stereo reconstruction. ACM Transactions on Graphics (TOG), 38(1):1-17, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[24] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Tim Salimans, Alexey A Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.383 + ], + "angle": 0, + "content": "[26] Ching-Yu Hsu, Cheng Sun, and Hwann-Tzong Chen. Moving in a 360 world: Synthesizing panoramic parallaxes from a single panorama. arXiv preprint arXiv:2106.10859, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[27] Huajian Huang, Yingshu Chen, Tianjian Zhang, and SaiKit Yeung. 360room: Real-time indoor roaming using geometry-aware \\(360^{\\circ}\\) radiance fields. arXiv preprint arXiv:2208.02705, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[28] Qingqing Huang, Daniel S. Park, Tao Wang, Timo I. Denk, Andy Ly, Nanxin Chen, Zhengdong Zhang, Zhishuai Zhang, Jia Yu, C. Frank, Jesse Engel, Quoc V. Le, William Chan, and Weixiang Han. Noise2music: Text-conditioned music generation with diffusion models. ArXiv, abs/2302.03917, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[29] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 857-866. IEEE Computer Society, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[30] Zutao Jiang, Guangsong Lu, Xiaodan Liang, Jihua Zhu, Wei Zhang, Xiaojun Chang, and Hang Xu. 3d-togo: Towards text-guided cross-category 3d object generation. arXiv preprint arXiv:2212.01103, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[31] Lei Jin, Yanyu Xu, Jia Zheng, Junfei Zhang, Rui Tang, Shugong Xu, Jingyi Yu, and Shenghua Gao. Geometric structure based and regularized depth estimation from 360 indoor imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 889-898, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[32] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[33] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, page 0, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[34] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. In ICML Workshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likelihood Models. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[35] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[36] Shreyas Kulkarni, Peng Yin, and Sebastian Scherer. 360fusionnerf: Panoramic neural radiance fields with joint guidance. arXiv preprint arXiv:2209.14265, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[37] Blockade Labs. Blockade skybox, https://skybox.blockadelabs.com/, accessed 2023-03-04. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.238, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[38] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.282, + 0.892, + 0.338 + ], + "angle": 0, + "content": "[39] Gang Li, Heliang Zheng, Chaoyue Wang, Chang Li, Changwen Zheng, and Dacheng Tao. 3ddesigner: Towards photorealistic 3d object generation and editing with text-guided diffusion models. arXiv preprint arXiv:2211.14108, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.34, + 0.892, + 0.395 + ], + "angle": 0, + "content": "[40] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.399, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[41] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.54 + ], + "angle": 0, + "content": "[42] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.542, + 0.892, + 0.597 + ], + "angle": 0, + "content": "[43] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.6, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[44] Luke Melas-Kyriazi, C. Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: \\(360^{\\circ}\\) reconstruction of any object from a single image. ArXiv, abs/2302.10663, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.892, + 0.698 + ], + "angle": 0, + "content": "[45] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[46] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.841 + ], + "angle": 0, + "content": "[47] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[48] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7918" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[49] Gimin Nam, Mariem Khlifi, Andrew Rodriguez, Alberto Tono, Linqi Zhou, and Paul Guerrero. 3d-ldm: Neural implicit 3d shape generation with latent diffusion models. arXiv preprint arXiv:2212.00842, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[50] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.218, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[51] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[52] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.33, + 0.47, + 0.383 + ], + "angle": 0, + "content": "[53] OpenAI. Dalle: Introducing outpainting, https://openai.com/blog/dall-e-introducing-outpainting?utm_source=tldrnewsletter, accessed 2023-03-07. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.385, + 0.47, + 0.44 + ], + "angle": 0, + "content": "[54] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2065–2074, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[55] Nick St. Pierre. Additive prompting, https://twitter.com/nickfloats/status/1628796348446253057, accessed 2023-03-07. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.483, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[56] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.525, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.609, + 0.469, + 0.661 + ], + "angle": 0, + "content": "[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.664, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[59] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.72, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[60] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.762, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[61] Xuanchi Ren and Xiaolong Wang. Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3553-3563, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.831, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[62] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[63] Mike Roberts, Debadeepta Dey, Anh Truong, Sudipta Sinha, Shital Shah, Ashish Kapoor, Pat Hanrahan, and Neel Joshi." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Submodular trajectory optimization for aerial 3d scanning. In Proceedings of the IEEE International Conference on Computer Vision, pages 5324-5333, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[64] C. Rockwell, David F. Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14084-14093, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[65] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[66] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[67] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[68] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[69] Aditya Sanghi, Hang Chu, J. Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18582-18592, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[70] Flavio Miguel Schneider, Zhijing Jin, and Bernhard Schölkopf. Mousai: Text-to-music generation with long-context latent diffusion. ArXiv, abs/2301.11757, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[71] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[72] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. ArXiv, abs/2111.02114, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[73] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[74] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8025-8035, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7919" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[75] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[76] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, and Yaniv Taigman. Text-to-4d dynamic scene generation. ArXiv, abs/2301.11280, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[77] Josef Sivic, Biliana K. Kaneva, Antonio Torralba, Shai Avidan, and William T. Freeman. Infinite images: Creating and exploring a large photorealistic virtual space. Proceedings of the IEEE, 98:1391-1407, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.355 + ], + "angle": 0, + "content": "[78] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.357, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[79] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.441 + ], + "angle": 0, + "content": "[80] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.47, + 0.483 + ], + "angle": 0, + "content": "[81] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.47, + 0.551 + ], + "angle": 0, + "content": "[82] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.47, + 0.594 + ], + "angle": 0, + "content": "[83] Alexandru Telea. An image inpainting technique based on the fast marching method. Journal of graphics tools, 9(1):23-34, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.637 + ], + "angle": 0, + "content": "[84] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[85] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[86] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3825–3834, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[87] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. arXiv preprint arXiv:2212.08070, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[88] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.148 + ], + "angle": 0, + "content": "[89] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.205 + ], + "angle": 0, + "content": "[90] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7465-7475, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.893, + 0.274 + ], + "angle": 0, + "content": "[91] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. ArXiv, abs/2212.11565, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.893, + 0.332 + ], + "angle": 0, + "content": "[92] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with \\(360^{\\circ}\\) views. arXiv e-prints, pages arXiv-2211, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.893, + 0.401 + ], + "angle": 0, + "content": "[93] Jiale Xu, Jia Zheng, Yanyu Xu, Rui Tang, and Shenghua Gao. Layout-guided novel view synthesis from a single indoor panorama. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16438-16447, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.893, + 0.445 + ], + "angle": 0, + "content": "[94] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7920" + } + ] +] \ No newline at end of file diff --git a/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/ec345001-5311-431e-90bd-6b91f4b47f2a_origin.pdf b/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/ec345001-5311-431e-90bd-6b91f4b47f2a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3f7c00008128383c720618d305a835d7f5b51571 --- /dev/null +++ b/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/ec345001-5311-431e-90bd-6b91f4b47f2a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0e412b82ab652e5c11e45eb91eaf1a34b7a7e2f897dee35e6cea9eb6fac2931 +size 7277122 diff --git a/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/full.md b/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3cae1171746f22ddd13786f03747ca7ad0555276 --- /dev/null +++ b/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/full.md @@ -0,0 +1,427 @@ +# Text2Room: Extracting Textured 3D Meshes from 2D Text-to-Image Models + +Lukas Hollein $^{1*}$ Ang Cao $^{2*}$ Andrew Owens $^{2}$ Justin Johnson $^{2}$ Matthias Nießner $^{1}$ $^{1}$ Technical University of Munich + $^{2}$ University of Michigan + +![](images/bdb645acfa17c805cdc9281c49f2227de27f066caee34c8152f2ff7c627d3f83.jpg) +(a) 3D Mesh Generation from Text + +![](images/3fbf50571c12f07c7b65d953b7fda35f01913fb28aef0380854534873099c0f1.jpg) +(b) Rendered Image + Mesh +Figure 1. Textured 3D mesh generation from text prompts. We generate textured 3D meshes from a given text prompt using 2D text-to-image models. (a) The scene is iteratively created from different viewpoints (marked in blue). (b) Our generated mesh contains compelling textures and geometry. We remove the ceiling in the top-down views for better visualization of the scene layout. + +# Abstract + +We present Text2Room1, a method for generating room-scale textured 3D meshes from a given text prompt as input. To this end, we leverage pre-trained 2D text-to-image models to synthesize a sequence of images from different poses. In order to lift these outputs into a consistent 3D scene representation, we combine monocular depth estimation with a text-conditioned inpainting model. The core idea of our approach is a tailored viewpoint selection such that the content of each image can be fused into a seamless, textured 3D mesh. More specifically, we propose a continuous alignment strategy that iteratively fuses scene frames with the existing geometry to create a seamless mesh. Unlike existing works that focus on generating single objects [56, 41] or zoom-out trajectories [18] from text, our method generates complete 3D scenes with multiple objects and explicit 3D geometry. We evaluate our approach using qualitative and quantitative metrics, demonstrating it as the first method to generate room-scale 3D geometry with compelling textures from only text as input. + +# 1. Introduction + +Mesh representations of 3D scenes are a crucial component for many applications, from AR/VR asset creation to computer graphics, yet creating these 3D assets remains a painstaking process that requires considerable expertise. In the 2D domain, recent works have successfully created high-quality images from text using generative models, such as diffusion models [65, 58, 67]. These methods significantly reduce the barriers to creating images that contain a user's desired content, effectively helping towards the democratization of content creation. An emerging line of work has sought to apply similar methods to create 3D models from text [9, 56, 29, 41, 38], yet existing approaches come with a number of significant limitations and lack the generality of 2D text-to-image models. + +One of the core challenges of generating 3D models is coping with the lack of available 3D training data, as 3D datasets are vastly smaller than those available in many other applications, such as 2D image synthesis. For example, methods that directly use 3D supervision, such as Chen et al. [9], are often limited to datasets of simple shapes, such as ShapeNet [8]. To address these data limitations, recent methods [56, 29, 41, 38, 88] lift the expressive power of 2D text-to-image models into 3D by formulating 3D generation as an iterative optimization problem in the image domain. This allows them to generate 3D ob + +jects stored in a radiance field representation, demonstrating the ability to generate arbitrary (neural) shapes from text. However, these methods cannot easily be extended to create room-scale 3D structure and texture. The challenge of generating large scenes is ensuring that the generated output is dense and coherent across outward-facing viewpoints, and that these views contain all of the required structures, such as walls, floors, and furniture. Additionally, a mesh remains a desired representation for many end-user tasks, such as rendering on commodity hardware (which requires an additional conversion step as presented in Lin et al. [41]). + +To address these shortcomings, we propose a method that extracts scene-scale 3D meshes from off-the-shelf 2D text-to-image models. Our method iteratively generates a scene through inpainting and monocular depth estimation. We produce an initial mesh by generating an image from text, and backproject it into 3D using a depth estimation model. Then, we iteratively render the mesh from novel viewpoints. From each one, we fill in holes in the rendered images via inpainting, then fuse the generated content into the mesh (Fig. 1a). + +Our iterative generation scheme has two important design considerations: how we choose the viewpoints, and how we merge generated scene content with the existing mesh. We first select viewpoints from predefined trajectories that will cover large amounts of scene content, then adaptively select viewpoints that close remaining holes. When merging generated content with the mesh, we align the two depth maps to create smooth transitions, and remove parts of the mesh that contain distorted textures. Together, these decisions lead to large, scene-scale 3D meshes with compelling textures and consistent geometry (Fig. 1b), that can represent a wide range of rooms. + +To summarize, our contributions are: + +- Generating 3D meshes of room-scale indoor scenes with compelling textures and geometry from any text input. +- A method that leverages 2D text-to-image models and monocular depth estimation to lift frames into 3D in an iterative scene generation. Our proposed depth alignment and mesh fusion steps, enable us to create seamless and undistorted geometry and textures. +- A two-stage tailored viewpoint selection that samples camera poses from optimal positions to first create the room layout and furniture and then close any remaining holes, creating a watertight mesh. + +# 2. Related Work + +Text-based Generation has seen significant advances due to large-scale image-text datasets [73, 72, 14, 71] and scalable generative model architectures [16, 66, 60, 32], enabling synthesis of novel images from text [20, 3, 54]. + +Recently, diffusion models [78, 23, 80, 81, 82] achieved impressive results on image synthesis [15, 65, 67, 50, 58] + +through improvements like latent space denoising [65, 84], faster sampling [23, 79, 52, 34], and better guidance [24]. + +In particular, text-to-image methods like Stable Diffusion [65], Imagen [67], GLIDE [50] and DALL-E 2 [58] yield diverse, high-fidelity, and controllable [6, 94] outputs. Text-based generation has been extended to other modalities including audio [35, 17, 28, 70], video [75, 91, 85, 25], and 4D fields [76]. We use text-to-image models by lifting their generated output into complete 3D scene meshes. + +Text-to-3D. Several methods use 3D data for supervised training of text-to-3D models [9, 51, 5]; however this direction remains challenging due to the lack of large-scale aligned datasets of text and 3D. + +Alternative approaches use 2D vision-language models like CLIP [57] to create 3D content by formulating the generation as an optimization problem in the image domain [86, 29, 38, 48, 30] or as object alignment [69]. Related methods refine existing 3D input through text guidance in a similar fashion [46, 10, 87, 62]. + +Recent methods [56, 41, 45, 88, 44] combine large text-to-image diffusion models [65, 67] and neural radiance fields [47] to generate 3D objects without training. Other approaches train custom diffusion models on a similar text-to-3D task [39, 49, 12]. In contrast, we use a fixed text-to-image model and extract a 3D mesh representing entire scenes of many objects and structural elements like walls. + +3D-Consistent View Synthesis from a Single Image. Several methods have been proposed that perform novel-view-synthesis from a single image [64, 90, 74, 61, 19]. Others optimize a neural 3D representation of an object, that can be viewed from arbitrary novel view points [92, 89, 1]. Another line of work performs perpetual view generation [77, 42, 40, 7], synthesizing videos via a render-refine-repeat pattern from a single RGB image that depict a scene along a forward-facing camera trajectory. In very recent concurrent work, Fridman et al. [18] create 3D scenes from text, but focus on this type of 3D-consistent "zoom-out" video generation. Instead, we generate complete, textured 3D room geometry from arbitrary trajectories. + +# 3. Method + +Our method creates a textured 3D mesh of a complete scene from text input. To this end, we continuously fuse generated frames from a 2D text-to-image model at different poses into a joint 3D mesh, creating the scene over time. The core idea of our approach is a two-stage tailored viewpoint selection, that first generates the scene layout and objects and then closes remaining holes in the 3D geometry (Section 3.4). We visualize this workflow in Figure 2. For each pose in both stages, we apply an iterative scene generation scheme to update the mesh (Section 3.1). We first align each frame with the existing geometry with a depth align- + +![](images/d0b92d778b3d5b2fe0a757649ba1e4ed4e9697e0ba97a52b85bd754ccbfbaec0.jpg) +(a) Scene Generation Stage + +![](images/67d6e7dc793e04bc55296ebbcf382d887b1ed15e3a373e64a5642dbd3a1febba.jpg) +(b) Scene Completion Stage + +![](images/76fae7b7534754273b139f805925c072af7ac5b7cf560feef8f74589ffe696b4.jpg) +Figure 2. Method overview. We iteratively create a textured 3D mesh in two stages. (a) First, we sample predefined poses and text to generate the complete scene layout and furniture. Each new pose (marked in green) adds newly generated geometry to the mesh (depicted by green triangles) in an iterative scene generation scheme (see Figure 3 for details). Blue poses/triangles denote viewpoints that created geometry in a previous iteration. (b) Second, we fill in the remaining unobserved regions by sampling additional poses (marked in red) after the scene layout is defined. +Figure 3. Iterative scene generation. For each new pose, we render the current mesh to obtain partial RGB and depth renderings. We complete both, utilizing respective inpainting models and the text prompt. Next, we perform depth alignment (see Section 3.2) and mesh filtering (see Section 3.3) to obtain an optimal next mesh patch, that is finally fused with the existing geometry. + +ment strategy (Section 3.2). Next, we triangulate and filter the novel content to merge it into the mesh (Section 3.3). + +# 3.1. Iterative 3D Scene Generation + +Our scene is represented as a mesh $\mathcal{M} = (\mathcal{V},\mathcal{C},\mathcal{S})$ where the vertices $\mathcal{V}\in \mathbb{R}^{N\times 3}$ , vertex colors $\mathcal{C}\in \mathbb{R}^{N\times 3}$ and the face set $S\in \mathbb{N}_0^{M\times 3}$ are generated over time. Input to our method is a set of arbitrary text prompts $\{P_t\}_{t = 1}^T$ that corresponds to our selected poses $\{E_t\}_{t = 1}^T\in \mathbb{R}^{3\times 4}$ in both stages. Inspired by recent methods [42, 40], we iteratively build up the scene, following a render-refine-repeat pattern. We summarize this iterative scene generation process in Figure 3. Formally, for each step of generation $t$ , we first render the current scene from a novel viewpoint: + +$$ +I _ {t}, d _ {t}, m _ {t} = r \left(\mathcal {M} _ {t}, E _ {t}\right), \tag {1} +$$ + +where $r$ is a classical rasterization function without shading, $I_{t}$ is the rendered image, $d_{t}$ the rendered depth and $m_{t}$ the image-space mask, that marks pixels without observed + +content. We then use a fixed text-to-image model $\mathcal{F}_{t2i}$ to inpaint unobserved pixels according to the text prompt: + +$$ +\hat {I} _ {t} = \mathcal {F} _ {t 2 i} \left(I _ {t}, m _ {t}, P _ {t}\right). \tag {2} +$$ + +Next, we inpaint unobserved depth by applying a monocular depth estimator $\mathcal{F}_d$ in our depth alignment (see Section 3.2): + +$$ +\hat {d} _ {t} = \text {p r e d i c t - a n d - a l i g n} \left(\mathcal {F} _ {d}, I _ {t}, d _ {t}, m _ {t}\right). \tag {3} +$$ + +Finally, we combine the novel content $\{\hat{I}_t,\hat{d}_t,m_t\}$ with the existing mesh by our fusion scheme (see Section 3.3): + +$$ +\mathcal {M} _ {t + 1} = f u s e \left(\mathcal {M} _ {t}, \hat {I} _ {t}, \hat {d} _ {t}, m _ {t}, E _ {t}\right). \tag {4} +$$ + +# 3.2. Depth Alignment Step + +To lift a 2D image $I$ into 3D, we predict the per-pixel depth. To correctly combine old and new content, it is necessary that both align with each other. In other words, similar regions in a scene like walls or furniture should be placed at similar depth. However, directly using the predicted depth for backprojection leads to hard cuts and discontinuities in the 3D geometry, since the depth is inconsistent in scale between subsequent viewpoints (see Figure 7a). + +To this end, we perform depth alignment in two-stages. First, we use a state-of-the-art depth inpainting network [4] that takes ground-truth depth $d$ for known parts in the image as input and aligns the prediction to it: $\hat{d}_p = \mathcal{F}_d(I,d)$ . + +Inspired by Liu et al. [42] we then improve the result by optimizing for scale and shift parameters $\gamma, \beta \in \mathbb{R}$ , aligning predicted and rendered disparity in the least squares sense: + +$$ +\min _ {\gamma , \beta} \quad \left\| m \odot \left(\frac {\gamma}{\hat {d} _ {p}} + \beta - \frac {1}{d}\right) \right\| ^ {2}, \tag {5} +$$ + +where we mask out unobserved pixels via $m$ . We can then extract the aligned depth as $\hat{d} = (\frac{\gamma}{\hat{d}_p} + \beta)^{-1}$ . Finally, we smooth $\hat{d}$ by applying a $5 \times 5$ Gaussian kernel at the mask edges (see supplemental material for more details). + +![](images/207a3b4cd3937f72df3864668ca2d82e905f71588dbc74ad1f280b31f55c3aa2.jpg) +(a) Pixel Triangulation (b) Face Filtering (c) Mesh Fusion + +![](images/56b07ccd1dde58e503e1527a1e776c02b6a334341cfbbb86190444be0a8cab52.jpg) +Figure 4. Visualization of our mesh fusion step. (a) We triangulate an image, such that 4 neighboring pixels (orange dots) create two faces. (b) We filter a face (marked in red), if its surface normal forms a small grazing angle with the viewing direction or if any edge in world space is too long. (c) We fuse the remaining faces (marked in green) with the existing geometry (marked in blue). + +![](images/a78e77db4e1e55dc29b4357bf074357eee9e02e94c4f2e8fb663e01bfad09ec3.jpg) + +# 3.3. Mesh Fusion Step + +At each step, we insert new content $\{\hat{I}_t,\hat{d}_t,m_t\}$ into the scene. For that, we first backproject the image-space pixels into a world-space point cloud: + +$$ +\mathcal {P} _ {t} = \left\{E _ {t} ^ {- 1} K ^ {- 1} \cdot \hat {d} _ {t} [ u, v ] \cdot (u, v, 1) ^ {T} \right\} _ {u = 0, v = 0} ^ {W, H}, \tag {6} +$$ + +where $K \in \mathbb{R}^{3 \times 3}$ are the camera intrinsics and $W, H$ are image width and height, respectively. We then use a simple triangulation scheme (Figure 4a), where each four neighboring pixels $\{(u, v), (u + 1, v), (u, v + 1), (u + 1, v + 1)\}$ in the image form two triangles. Since the estimated depth is noisy, this naive triangulation creates stretched out 3D geometry (see Figure 7b). To alleviate this problem, we propose two filters that remove stretched out faces (Figure 4b). + +First, we filter faces based on their edge length. We remove a face if the Euclidean distance of any face edge is larger than a threshold $\delta_{edge}$ . Second, we filter faces based on the angle between surface normal and viewing direction: + +$$ +\mathcal {S} = \left\{\left(i _ {0}, i _ {1}, i _ {2}\right) \mid n ^ {T} v > \delta_ {s n} \right\} \tag {7} +$$ + +where $S$ is the face set, $(i_0,i_1,i_2)$ are the vertex indices of the triangle, $\delta_{sn}$ is the threshold, $n\in \mathbb{R}^3$ is the normalized face normal, and $v\in \mathbb{R}^3$ is the normalized view direction in world space from the camera center towards the average pixel location from which the triangle originated. This avoids creating texture for large regions of the mesh from a comparatively small number of pixels from an image. + +Finally, we fuse together the newly generated mesh patch and the existing geometry (Figure 4c). All faces that are backprojected from pixels falling into the inpainting mask $m_{t}$ are stitched together with their neighboring faces, which are already part of the mesh. Precisely, we continue the triangulation scheme at all edges of $m_{t}$ , but use the existing vertex positions of $\mathcal{M}_t$ to create the corresponding faces. + +# 3.4. Two-Stage Viewpoint Selection + +A key part of our method is the choice of text prompts and camera poses from which the scene is synthesized. Users can in principle choose these inputs arbitrarily to create any desired indoor scene. However, the generated scene can degenerate and contain stretch and hole artifacts, if poses are chosen carelessly (see Figure 7 and supplemental material). To this end, we propose a two-stage viewpoint selection strategy, that samples each next camera pose from optimal positions and refines empty regions subsequently. + +Generation Stage. In the first stage, we create the main parts of the scene, including the general layout and furniture. We subsequently render predefined trajectories in different directions that eventually cover the whole room. We found generation works best, if each trajectory starts off from a viewpoint with mostly unobserved regions. This generates the outline of the next chunk, while still being connected to the rest of the scene (e.g., see Figure 3). Then, we complete the 3D structure of that chunk by moving and rotating into it subsequently until the end of the trajectory. + +Additionally, we ensure an optimal observation distance for each pose. We translates camera positions $T_0 \in \mathbb{R}^3$ along the look-at direction $L \in \mathbb{R}^3$ uniformly: $T_{i+1} = T_i - 0.3L$ . We stop if the mean rendered depth is larger than 0.1 or discard the camera after 10 steps. This avoids views too close to existing geometry. For example, the green pose in Figure 2a is moved back as far as possible into the existing geometry such that it views most of the empty floor region. + +We create closed room layouts following this principle, by choosing trajectories that generate the next chunks in a circular motion, roughly centered around the origin. We found it helpful to discourage the text-to-image generator from generating furniture in unwanted regions by engineering the text prompts accordingly. For example, for poses looking at the floor or ceiling, we choose text prompts that only contain the words "floor" or "ceiling", respectively. + +Completion Stage. After the first stage, the scene layout and furniture is defined. However, it is impossible to choose sufficient poses $a$ -priori. Since the scene is generated on-the-fly, the mesh contains holes that were not observed by any camera (see Figure 7c). We complete the scene by sampling additional poses $a$ -posteriori, looking at those holes. + +Inspired by trajectory optimization [22, 63], we voxelize the scene into dense uniform cells. We sample random poses in each cell, discarding those being too close to existing geometry. We select one pose per cell that views most unobserved pixels (e.g., see the red poses in Figure 2b). + +Next, we inpaint the scene from all chosen camera poses following Section 3.1. Similar to Fridman et al. [18], we observe it is important to clean the inpainting masks, because our text-to-image generator can generate better results for large connected regions. Thus, we first inpaint small holes + +with a classical inpainting algorithm [83] and dilate the remaining holes. We additionally remove all faces that fall into the dilated region and are close to the rendered depth. Please see the supplemental material for more details. + +Finally, we run Poisson surface reconstruction [33] on the scene mesh. This closes any remaining holes after completion and smoothens out discontinuities. The result is a watertight mesh of the generated scene, that can be rendered with classical rasterization. + +# 4. Results + +Implementation Details. We implement mesh rasterization and fusion with Pytorch3D [59]. As our text-to-image model $\mathcal{F}_{t2i}$ , we utilize a Stable Diffusion [65] model, that is finetuned on the image inpainting task, using additional mask input. We generate a single inpainting proposal and employ a state-of-the-art guided diffusion sampler [43]. As our monocular depth estimator $\mathcal{F}_d$ , we employ an Iron-Depth [4] model, that is trained on indoor scenes from the ScanNet dataset [13] and augment it for depth inpainting according to Bae et al. [4]. We set $\delta_{edge} = 0.1$ and $\delta_{sn} = 0.1$ in all our experiments. During generation, we use 20 different trajectories with 10 frames each sampled between the respective start and end poses. We construct prompts using the guidelines suggested by Pierre [55]. Creating one scene takes approximately 50 minutes on one RTX 3090 GPU. + +Baselines. To the best of our knowledge, there are no direct baselines that generate textured 3D room geometry from text. We compare against four related methods (please see the supplemental material for more details about baselines). +- PureClipNeRF [38]: We compare against text-to-3D methods for generating objects [56, 41, 29, 38, 88] and choose Lee et al. [38] as open-source representative. +- Outpainting [58, 53]: We combine outpainting from a Stable Diffusion [65] model with depth estimation and triangulation to create a mesh from an enlarged viewpoint. +- Text2Light [11]: We generate RGB panoramas from text using Chen et al. [11]. Estimating 3D mesh structure from a panorama is difficult. Related approaches estimate room layout [93], perform view synthesis [36, 26, 21, 27] or predict $360^{\circ}$ depth [2, 31]. We perform depth prediction and subsequently apply our mesh fusion step. +- Blockade [37]: We apply Blockade [37], which uses a text-to-image diffusion model to produce more expressive RGB panoramas. We then extract the mesh similarly. + +Evaluation Metrics. The generated 3D geometry is evaluated both quantitatively and qualitatively. We calculate CLIP Score (CS) [57] and Inception Score (IS) [68] on RGB renderings of the respective scenes. Additionally, we conduct a user study and ask $n = 61$ users to score Perceptual Quality ( $PQ$ ) and 3D Structure Completeness ( $3DS$ ) of the whole scene on a scale of 1-5. + +
Method2D MetricsUser Study
CS ↑IS ↑PQ ↑3DS ↑
PureClipNeRF [38]24.061.262.342.38
Outpainting [58, 53]23.101.602.902.58
Text2Light [11]+Ours25.992.212.822.97
Blockade [37]+Ours26.292.133.353.36
Ours w/o alignment26.731.783.122.96
Ours w/o stretch removal27.721.863.283.75
Ours w/o completion27.972.183.723.87
Ours28.022.314.014.19
+ +Table 1. Quantitative comparison. We report 2D metrics and user study results, including: Clip Score (CS), Inception Score (IS), Perceptual Quality (PQ), and 3D Structure Completeness (3DS). Our method creates scenes with the highest quality. + +# 4.1. Qualitative Results + +We show top-down views into the scene and RGB renderings from within for our method and baselines in Figure 6. We show additional results of our method in Figure 5. PureClipNeRF [38] creates the key objects of the given text prompt, but does not create a complete 3D structure with floor, walls and ceilings. Outpainting [58, 53] creates high-detail textures, but projection from a single viewpoint creates holes due to occlusion and hinders the creation of complete 3D geometry. Text2Light [11] and Blockade [37] both create a high-detail $360^{\circ}$ view of a complete scene, but occlusions that cannot be resolved from a single panoramic viewpoint lead to holes in the extracted 3D geometry. + +In contrast, our approach creates high-detail textures and geometry, that are fused into a complete 3D scene mesh without holes. The resulting scenes contain flat floors, walls and ceilings, as well as 3D object geometry distributed throughout the scene. When specifying text prompts with a huge variety, the resulting scene contains a diverse set of objects. Please see the supplemental material for more scenes, animated results, intermediate outputs of our baselines (such as the panoramic images) as well as top-down views of meshes, that contain the reconstructed ceilings. + +# 4.2. Quantitative Results + +We show quantitative results averaged over multiple scenes in Table 1. We render 60 images from novel viewpoints for each scene to calculate the 2D metrics. We present users with multiple top-down views and renderings for each scene and let them rate each method individually (no side-by-side comparison). Stretched-out geometry and holes in the 3D geometry lead to lower scores for the baselines in all image-based metrics. Our approach achieves the highest scores, because the renderings are complete from arbitrary novel poses, satisfy the given text-prompt and + +![](images/d2eca9ce87387bb4b59230539c92d088cf169eff0c69a25a872dfbe2afb4f335.jpg) + +![](images/eb9852895e792c867f2c9153ce6babf4c15383e7bacf3a169c80284e9f13c862.jpg) + +![](images/85cb3ff25fbe5385b65da195af10e3c71bb21ae14c1d4ed285cbdc5959d68f08.jpg) + +![](images/2e2041852c7c20bccf5e8b9dfe23c6f3d82d4b06e2b7be4fbf1bb70a631444c0.jpg) + +![](images/41c2fb0160bb6ce1d137e34ec69bda5b2e94227da28e92b56b0c8a8fac70a26c.jpg) + +Editorial Style Photo, Coastal Bathroom, Clawfoot Tub, Seashell, Wicker, Mosaic Tile, Blue and White + +![](images/d490125d2ee33df607bb19a86327e4afcda30b628d9ee3f401a42fa10fbe39d7.jpg) + +![](images/feec8cc2863f2075c549047bbfca4f0c8aa7c0a50e616b0c31d6319261cfdb90.jpg) + +![](images/e0f2d1a79fa49f25f6f0a034cd2981799a201be4c59af3daa844be026d200ae2.jpg) + +![](images/16c4e258a97397ad49226868831fb185695929825561df0fafe92cb295bf5c5f.jpg) + +![](images/e6fbcfc3552c572956b37bd02099b975c4565c48770e4cbc7ce5d7ff6185d418.jpg) + +A living room with a lit furnace, couch, and cozy curtains, bright lamps that make the room look well-lit + +![](images/db647f626426a2292c35eff9ef458c3ee87607c2ddf0083b4f5759a873122502.jpg) + +![](images/15ca036db5bf5940cde79bfa1f66b04ed9f7129e401f712738fa684052ec52c3.jpg) + +![](images/1fd22bc3048c7b7f28f67b3eb0c36462f2c6889038e2ab2f6d1a984e58dbfbed.jpg) + +![](images/afbbc7a61702516a0571f1f09d412731d1f64f18083493f67c0539cb37a33927.jpg) + +![](images/bb525f53e3f383cd059cbe20736408a07be28d5338f624d2a5ea681b1ad177d1.jpg) + +Editorial Style Photo, Modern Living Room, Large Window, Leather, Glass, Metal, Wood Paneling, Apartment + +![](images/251399c527f029c48a652a0f5a880cbfcc8a0ca5cae9b8fec8486ebe0d696fa6.jpg) +Figure 5. 3D scene generation results of our method. We show color and shaded geometry renderings from generated scenes with corresponding text prompts. Our method synthesizes realistic meshes satisfying text descriptions. We remove the ceiling in the top-down view for better visualization of the scene layout. + +![](images/d1fe957ca3366c375343d3fb08bb1c1335739ec6885d0d06bc6b4aa248254d7a.jpg) + +![](images/bfbe32a7c7e0ce74676c8b8dbf14c8080842fa5057bcf199bb289fda05771198.jpg) + +![](images/abd48c60ff69b6a59a731b9277017eccf017e10828ae6d4a62bcb1fffb2565b1.jpg) + +![](images/d0f77186f66663241a634ce076d5f236a931fb67b7baa52c9aa03205830d021e.jpg) + +Editorial Style Photo, Modern Nursery, Table Lamp, Rocking Chair, Tree Wall Decal, Wood, Cotton, Faux Fur + +![](images/cc180d7990d0664bd23cc9d4f05424da14d560fd3eea4a2dd21c1a2b7196be2f.jpg) +Editorial Style Photo, Industrial Home Office, Steel Shelves, Concrete, Metal, Edison Bulbs, Exposed Ductwork + +![](images/4bead8a57fc0d8fcfc81dbe818a3b5c86589f67a4e4d00571b6ed51932708cc8.jpg) +a bedroom with a king-size bed and a large wardrobe +PureClipNeRF [38] +Outpainting [58, 53] +Text2Light [11]+Ours +Blockade [37]+Ours +Ours +Figure 6. Qualitative comparison of our method and baselines. PureClipNeRF [38] cannot produce immersive scenes with floors and walls. Outpainting [58, 53] does not produce 3D consistent scenes. Text2Light [11] and Blockade [37] both have holes due to occlusions. In contrast, our method creates complete meshes without holes and high details. We remove the ceiling in the top-down view for better visualization of the scene layout. + +contain high-resolution image features. Users prefer our method, which highlights the quality of our accurate and complete geometry, as well as the RGB texture. + +# 4.3. Ablations + +The key ingredients of our method are depth alignment (Section 3.2), mesh fusion (Section 3.3) and the two-stage viewpoint selection (Section 3.4). We demonstrate the importance of each component in Figure 7 and Table 1. + +Depth alignment creates seamless scenes. Monocular depth predictions from subsequent frames can be inconsistent in scale. This leads to disconnected components in the mesh that are backprojected from multiple viewpoints (see Figure 7a). Our depth alignment strategy allows fusing multiple frames into a seamless mesh, eventually creating a complete scene with flat floors, walls, ceilings and no holes. + +Stretch removal creates undistorted scene geometry. During mesh fusion, we update the scene geometry with the contents of the next frame. Due to noisy depth prediction, the objects become stretched out, if they are observed + +from small grazing angles. Thus, we propose two filters (edge length and surface normal thresholds) that alleviate this issue. Instead of baking in stretched-out geometry (see Figure 7b), we disregard the corresponding faces and let the object be completed from a more suitable, later viewpoint. + +Two-stage generation creates complete scenes. Our approach chooses camera poses in two stages to create a complete scene without holes. After generating the scene from predefined trajectories, the scene still contains some holes (see Figure 7c). Because the scene is built-up over time, it is impossible to choose camera poses $a$ -priori, that view all unobserved regions. To this end, our completion stage samples poses $a$ -posteriori to refine those regions. The resulting mesh is watertight and contains no holes (see Figure 7d). + +# 4.4. Spatially Varying Scene Generation + +Our method can be applied to generate a scene as the combination of multiple text prompts. Specifically, we use separate text prompts for different poses, crafting a set of trajectories that spatially combines scene descriptions. This + +![](images/e782e9c77329ce1310e5a56e2d0239a7e09936253dc60d2d7ef1599b93f0a242.jpg) + +![](images/e791f881e28aef630bc565e2c60edd16ec6e56a89c90166cb463e1b068799fc5.jpg) + +![](images/333a53c14136c1e5eb79546a8c33a5cab54f49160f7ef6a6dd762f04155e0afa.jpg) + +![](images/86ddbd6b0aac7e63748ff124b8ffc4f800cf145d629885f11fd0d8abe2c81aa4.jpg) + +![](images/054d5f3dda57bf7f8d6fdb9471a1591706a6bb1706c3c841c9c863cf0e95a81a.jpg) +(a) Ours w/o alignment + +![](images/9287746b24ed8c0550ea06d2dc2ebdcb740b8194965f4455273c97f30965d986.jpg) +(b) Ours w/o stretch removal + +![](images/e4299f0a5e9e71d98de2892563c77d4a937192577c965e69cc711bf77ef3f1aa.jpg) +(c) Ours w/o completion + +![](images/b4b15330369576b7b9e41ef08aba4d3bc87839299c8f86ba1f45e5320c023c1b.jpg) +(d) Ours (full) + +living room, couches, curtains, lit furnace, lamps + +![](images/c755f295f8051c799a7331e05b23b30ac84c54b05e8dd50fc27058b529e1a396.jpg) +Figure 7. Ablation study on the key components of our method. Without depth alignment (see Section 3.2), different parts of the scene are disconnected and do not fuse into a seamless mesh. Without edge and surface normal thresholds (see Section 3.3), many faces are stretched out unnaturally. Without completion (see Section 3.4), the mesh has holes in remaining unobserved regions. Our full pipeline creates complete, high-resolution scenes. We remove the ceiling in the top-down view for better visualization of the scene layout. + +![](images/c926aef2331ef8854368beeff4730020aa312b0477b3bba85b8762f26dc73461.jpg) + +![](images/006d84408941ae6fe0888f7fe3b2ba6cfd328ae279fe2e48b69f3f6093298268.jpg) +kitchen, dinner table, dishwashers, ovens, countertops living room, lit furnace, couch, curtains + +![](images/806162c801d775f3e05192c3bf5d0667a193e2ee64958ea3f370e7072ac108b4.jpg) +bathroom, shower, bathtub bedroom, king-size bed, wardrobe +Figure 8. Spatially varying scene generation. Our method can create rooms with multiple parts by prompt mixing. We use separate prompts for cameras viewing different parts of the scene. This is a controllable way to create rooms from multiple descriptions. + +can be desired to avoid repeating elements in a complete scene (e.g., multiple couches could be spread out over the whole room when using the same prompt for every camera). Instead, users can specify different object positions through different camera poses and text prompts. It can also be used to design a house comprised of multiple rooms, each with a different type (e.g., a living room that leads to a kitchen). + +![](images/2e3bce1ef21548390e5a6e0bb425761c89423b7880322207bea90af4e6ecb3a9.jpg) + +![](images/547a261cc20522c050bb21615ed515fba8232845c711e2859e1b180560c04f69.jpg) +Figure 9. Scene generation with layout guidance. Our method can generate scenes from layout guidance. Left: we describe objects with different prompts for cameras facing at different directions. Right: the generated part of the room. + +We show results that combine multiple text prompts in Figure 8 and Figure 9. + +We note that the layout can only be partially controlled by the camera poses, since scene generation can create chunks with larger or smaller extent. We believe this demonstrates an exciting application of our method, that can be further explored in future work. + +# 4.5. Limitations + +Our approach allows to generate 3D room geometry from arbitrary text prompts that are highly detailed and contain consistent geometry. Nevertheless, our method can still fail under certain conditions (see supplemental material). First, our thresholding scheme (see Section 3.3) may not detect all stretched-out regions, which may lead to remaining distortions. Additionally, some holes may still not be completed fully after the second stage (see Section 3.4), + +which results in over-smoothed regions after applying poisson reconstruction. Our scene representation does not decompose material from lighting, which bakes in shadows or bright lamps, that are generated from the diffusion model. + +# 5. Conclusion + +We have shown a method to generate textured 3D meshes from only text input. We use text-to-image 2D generators to create a sequence of images. The core insight of our method is a tailored viewpoint selection, that allows to create a 3D mesh with seamless geometry and compelling textures. Specifically, we lift the images into a 3D scene, by employing our alignment strategy that iteratively fuses all images into the mesh. Our output meshes represent arbitrary indoor scenes that can be rendered with classical rasterization pipelines. We believe our approach demonstrates an exciting application of large-scale 3D asset creation, that only requires text as input. + +# Acknowledgements + +This work was funded in part by Cisco Systems. It was also supported by the ERC Starting Grant Scan2CAD (804724) as well as the German Research Foundation (DFG) Research Unit "Learning and Simulation in Visual Computing." We also thank Angela Dai for the video voice over. + +# References + +[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3D reconstruction, inpainting and generation. arXiv, 2022. 2 +[2] Manuel Rey Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution $360^{\circ}$ monocular depth estimation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5 +[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18187-18197, 2021. 2 +[4] Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Iron-depth: Iterative refinement of single-view depth using surface normal and its uncertainty. In British Machine Vision Conference (BMVC), 2022. 3, 5 +[5] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh Susskind. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2 +[6] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2 +[7] Shengqu Cai, Eric Ryan Chan, Songyou Peng, Mohamad Shahbazi, Anton Obukhov, Luc Van Gool, and Gordon Wetzstein. Diffdreamer: Consistent single-view perpetual view + +generation with conditional diffusion models. In arXiV, 2022. 2 +[8] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, L. Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository. ArXiv, abs/1512.03012, 2015. 1 +[9] Kevin Chen, Christopher Bongsoo Choy, Manolis Savva, Angel X. Chang, Thomas A. Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. ArXiv, abs/1803.08495, 2018. 1, 2 +[10] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[11] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 5, 7 +[12] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. arXiv preprint arXiv:2212.04493, 2022. 2 +[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5 +[14] Karan Desai, Gaurav Kaul, Zubin Trivadi Aysola, and Justin Johnson. Redcaps: Web-curated image-text data created by the people, for the people. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 2 +[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2 +[16] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868-12878, 2020. 2 +[17] Seth Forsgren* and Hayk Martiros*.iffusion - Stable diffusion for real-time music generation, https://riffusion.com/about, accessed 2023-03-06, 2022. 2 +[18] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133, 2023. 1, 2, 4 +[19] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M. Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. ArXiv, abs/2302.10109, 2023. 2 +[20] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10686-10696, 2021. 2 + +[21] Takayuki Hara and Tatsuya Harada. Enhancement of novel view synthesis using omnidirectional image completion. arXiv preprint arXiv:2203.09957, 2022. 5 +[22] Benjamin Hepp, Matthias Nießner, and Otmar Hilliges. Plan3d: Viewpoint and trajectory optimization for aerial multi-view stereo reconstruction. ACM Transactions on Graphics (TOG), 38(1):1-17, 2018. 4 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2 +[24] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2 +[25] Jonathan Ho, Tim Salimans, Alexey A Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems. 2 +[26] Ching-Yu Hsu, Cheng Sun, and Hwann-Tzong Chen. Moving in a 360 world: Synthesizing panoramic parallaxes from a single panorama. arXiv preprint arXiv:2106.10859, 2021. 5 +[27] Huajian Huang, Yingshu Chen, Tianjian Zhang, and SaiKit Yeung. 360room: Real-time indoor roaming using geometry-aware $360^{\circ}$ radiance fields. arXiv preprint arXiv:2208.02705, 2022. 5 +[28] Qingqing Huang, Daniel S. Park, Tao Wang, Timo I. Denk, Andy Ly, Nanxin Chen, Zhengdong Zhang, Zhishuai Zhang, Jia Yu, C. Frank, Jesse Engel, Quoc V. Le, William Chan, and Weixiang Han. Noise2music: Text-conditioned music generation with diffusion models. ArXiv, abs/2302.03917, 2023. 2 +[29] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 857-866. IEEE Computer Society, 2022. 1, 2, 5 +[30] Zutao Jiang, Guangsong Lu, Xiaodan Liang, Jihua Zhu, Wei Zhang, Xiaojun Chang, and Hang Xu. 3d-togo: Towards text-guided cross-category 3d object generation. arXiv preprint arXiv:2212.01103, 2022. 2 +[31] Lei Jin, Yanyu Xu, Jia Zheng, Junfei Zhang, Rui Tang, Shugong Xu, Jingyi Yu, and Shenghua Gao. Geometric structure based and regularized depth estimation from 360 indoor imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 889-898, 2020. 5 +[32] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 2 +[33] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, page 0, 2006. 5 +[34] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. In ICML Workshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likelihood Models. 2 + +[35] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations. 2 +[36] Shreyas Kulkarni, Peng Yin, and Sebastian Scherer. 360fusionnerf: Panoramic neural radiance fields with joint guidance. arXiv preprint arXiv:2209.14265, 2022. 5 +[37] Blockade Labs. Blockade skybox, https://skybox.blockadelabs.com/, accessed 2023-03-04. 5, 7 +[38] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022. 1, 2, 5, 7 +[39] Gang Li, Heliang Zheng, Chaoyue Wang, Chang Li, Changwen Zheng, and Dacheng Tao. 3ddesigner: Towards photorealistic 3d object generation and editing with text-guided diffusion models. arXiv preprint arXiv:2211.14108, 2022. 2 +[40] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In ECCV, 2022. 2, 3 +[41] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 1, 2, 5 +[42] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2021. 2, 3 +[43] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 5 +[44] Luke Melas-Kyriazi, C. Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: $360^{\circ}$ reconstruction of any object from a single image. ArXiv, abs/2302.10663, 2023. 2 +[45] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. 2 +[46] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2 +[47] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, 2020. 2 +[48] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2 + +[49] Gimin Nam, Mariem Khlifi, Andrew Rodriguez, Alberto Tono, Linqi Zhou, and Paul Guerrero. 3d-ldm: Neural implicit 3d shape generation with latent diffusion models. arXiv preprint arXiv:2212.00842, 2022. 2 +[50] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 2 +[51] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2 +[52] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2 +[53] OpenAI. Dalle: Introducing outpainting, https://openai.com/blog/dall-e-introducing-outpainting?utm_source=tldrnewsletter, accessed 2023-03-07. 5, 7 +[54] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2065–2074, 2021. 2 +[55] Nick St. Pierre. Additive prompting, https://twitter.com/nickfloats/status/1628796348446253057, accessed 2023-03-07. 5 +[56] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 1, 2, 5 +[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5 +[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1, 2, 5, 7 +[59] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 5 +[60] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2 +[61] Xuanchi Ren and Xiaolong Wang. Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3553-3563, 2022. 2 +[62] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2 +[63] Mike Roberts, Debadeepta Dey, Anh Truong, Sudipta Sinha, Shital Shah, Ashish Kapoor, Pat Hanrahan, and Neel Joshi. + +Submodular trajectory optimization for aerial 3d scanning. In Proceedings of the IEEE International Conference on Computer Vision, pages 5324-5333, 2017. 4 +[64] C. Rockwell, David F. Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14084-14093, 2021. 2 +[65] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 1, 2, 5 +[66] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2 +[67] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1, 2 +[68] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 5 +[69] Aditya Sanghi, Hang Chu, J. Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18582-18592, 2021. 2 +[70] Flavio Miguel Schneider, Zhijing Jin, and Bernhard Schölkopf. Mousai: Text-to-music generation with long-context latent diffusion. ArXiv, abs/2301.11757, 2023. 2 +[71] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track. 2 +[72] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. ArXiv, abs/2111.02114, 2021. 2 +[73] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2 +[74] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8025-8035, 2020. 2 + +[75] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 2 +[76] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, and Yaniv Taigman. Text-to-4d dynamic scene generation. ArXiv, abs/2301.11280, 2023. 2 +[77] Josef Sivic, Biliana K. Kaneva, Antonio Torralba, Shai Avidan, and William T. Freeman. Infinite images: Creating and exploring a large photorealistic virtual space. Proceedings of the IEEE, 98:1391-1407, 2008. 2 +[78] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2 +[79] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations. 2 +[80] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2 +[81] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 2 +[82] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations. 2 +[83] Alexandru Telea. An image inpainting technique based on the fast marching method. Journal of graphics tools, 9(1):23-34, 2004. 5 +[84] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In Neural Information Processing Systems, 2021. 2 +[85] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2 +[86] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3825–3834, 2021. 2 +[87] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. arXiv preprint arXiv:2212.08070, 2022. 2 +[88] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022. 1, 2, 5 + +[89] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2 +[90] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7465-7475, 2019. 2 +[91] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. ArXiv, abs/2212.11565, 2022. 2 +[92] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with $360^{\circ}$ views. arXiv e-prints, pages arXiv-2211, 2022. 2 +[93] Jiale Xu, Jia Zheng, Yanyu Xu, Rui Tang, and Shenghua Gao. Layout-guided novel view synthesis from a single indoor panorama. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16438-16447, 2021. 5 +[94] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2 \ No newline at end of file diff --git a/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/images.zip b/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..05dd1d7b9c3742df323f0fdbca72ed5b8f4cda4c --- /dev/null +++ b/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ec2551c99a0074753470daee784972e1f6ddc0a2387c8f16ad66a5d1a921d86 +size 1002300 diff --git a/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/layout.json b/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9be9f3fbf477042e02f8428cf293388a4949b0ce --- /dev/null +++ b/2023/Text2Room_ Extracting Textured 3D Meshes from 2D Text-to-Image Models/layout.json @@ -0,0 +1,11395 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 103, + 532, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 103, + 532, + 120 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 532, + 120 + ], + "type": "text", + "content": "Text2Room: Extracting Textured 3D Meshes from 2D Text-to-Image Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "spans": [ + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "text", + "content": "Lukas Hollein" + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "text", + "content": " Ang Cao" + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "text", + "content": " Andrew Owens" + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "text", + "content": " Justin Johnson" + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "text", + "content": " Matthias Nießner" + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "text", + "content": "Technical University of Munich \n" + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 512, + 172 + ], + "type": "text", + "content": "University of Michigan" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 198, + 428, + 331 + ], + "blocks": [ + { + "bbox": [ + 50, + 198, + 428, + 331 + ], + "lines": [ + { + "bbox": [ + 50, + 198, + 428, + 331 + ], + "spans": [ + { + "bbox": [ + 50, + 198, + 428, + 331 + ], + "type": "image", + "image_path": "bdb645acfa17c805cdc9281c49f2227de27f066caee34c8152f2ff7c627d3f83.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 334, + 310, + 344 + ], + "lines": [ + { + "bbox": [ + 169, + 334, + 310, + 344 + ], + "spans": [ + { + "bbox": [ + 169, + 334, + 310, + 344 + ], + "type": "text", + "content": "(a) 3D Mesh Generation from Text" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 433, + 199, + 547, + 329 + ], + "blocks": [ + { + "bbox": [ + 433, + 199, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 433, + 199, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 433, + 199, + 547, + 329 + ], + "type": "image", + "image_path": "3fbf50571c12f07c7b65d953b7fda35f01913fb28aef0380854534873099c0f1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 333, + 546, + 346 + ], + "lines": [ + { + "bbox": [ + 432, + 333, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 432, + 333, + 546, + 346 + ], + "type": "text", + "content": "(b) Rendered Image + Mesh" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 352, + 546, + 386 + ], + "lines": [ + { + "bbox": [ + 45, + 352, + 546, + 386 + ], + "spans": [ + { + "bbox": [ + 45, + 352, + 546, + 386 + ], + "type": "text", + "content": "Figure 1. Textured 3D mesh generation from text prompts. We generate textured 3D meshes from a given text prompt using 2D text-to-image models. (a) The scene is iteratively created from different viewpoints (marked in blue). (b) Our generated mesh contains compelling textures and geometry. We remove the ceiling in the top-down views for better visualization of the scene layout." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 415, + 192, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 415, + 192, + 428 + ], + "spans": [ + { + "bbox": [ + 143, + 415, + 192, + 428 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 441, + 290, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 290, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 290, + 670 + ], + "type": "text", + "content": "We present Text2Room1, a method for generating room-scale textured 3D meshes from a given text prompt as input. To this end, we leverage pre-trained 2D text-to-image models to synthesize a sequence of images from different poses. In order to lift these outputs into a consistent 3D scene representation, we combine monocular depth estimation with a text-conditioned inpainting model. The core idea of our approach is a tailored viewpoint selection such that the content of each image can be fused into a seamless, textured 3D mesh. More specifically, we propose a continuous alignment strategy that iteratively fuses scene frames with the existing geometry to create a seamless mesh. Unlike existing works that focus on generating single objects [56, 41] or zoom-out trajectories [18] from text, our method generates complete 3D scenes with multiple objects and explicit 3D geometry. We evaluate our approach using qualitative and quantitative metrics, demonstrating it as the first method to generate room-scale 3D geometry with compelling textures from only text as input." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 393, + 386, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 386, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 386, + 405 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 414, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 581 + ], + "type": "text", + "content": "Mesh representations of 3D scenes are a crucial component for many applications, from AR/VR asset creation to computer graphics, yet creating these 3D assets remains a painstaking process that requires considerable expertise. In the 2D domain, recent works have successfully created high-quality images from text using generative models, such as diffusion models [65, 58, 67]. These methods significantly reduce the barriers to creating images that contain a user's desired content, effectively helping towards the democratization of content creation. An emerging line of work has sought to apply similar methods to create 3D models from text [9, 56, 29, 41, 38], yet existing approaches come with a number of significant limitations and lack the generality of 2D text-to-image models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "One of the core challenges of generating 3D models is coping with the lack of available 3D training data, as 3D datasets are vastly smaller than those available in many other applications, such as 2D image synthesis. For example, methods that directly use 3D supervision, such as Chen et al. [9], are often limited to datasets of simple shapes, such as ShapeNet [8]. To address these data limitations, recent methods [56, 29, 41, 38, 88] lift the expressive power of 2D text-to-image models into 3D by formulating 3D generation as an iterative optimization problem in the image domain. This allows them to generate 3D ob" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "type": "text", + "content": "* joint first authorship" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 58, + 703, + 257, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 257, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 257, + 712 + ], + "type": "text", + "content": "1https://lukashoel.github.io/text-to-room" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7909" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "jects stored in a radiance field representation, demonstrating the ability to generate arbitrary (neural) shapes from text. However, these methods cannot easily be extended to create room-scale 3D structure and texture. The challenge of generating large scenes is ensuring that the generated output is dense and coherent across outward-facing viewpoints, and that these views contain all of the required structures, such as walls, floors, and furniture. Additionally, a mesh remains a desired representation for many end-user tasks, such as rendering on commodity hardware (which requires an additional conversion step as presented in Lin et al. [41])." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 323 + ], + "type": "text", + "content": "To address these shortcomings, we propose a method that extracts scene-scale 3D meshes from off-the-shelf 2D text-to-image models. Our method iteratively generates a scene through inpainting and monocular depth estimation. We produce an initial mesh by generating an image from text, and backproject it into 3D using a depth estimation model. Then, we iteratively render the mesh from novel viewpoints. From each one, we fill in holes in the rendered images via inpainting, then fuse the generated content into the mesh (Fig. 1a)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 323, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 289, + 467 + ], + "type": "text", + "content": "Our iterative generation scheme has two important design considerations: how we choose the viewpoints, and how we merge generated scene content with the existing mesh. We first select viewpoints from predefined trajectories that will cover large amounts of scene content, then adaptively select viewpoints that close remaining holes. When merging generated content with the mesh, we align the two depth maps to create smooth transitions, and remove parts of the mesh that contain distorted textures. Together, these decisions lead to large, scene-scale 3D meshes with compelling textures and consistent geometry (Fig. 1b), that can represent a wide range of rooms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 468, + 198, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 198, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 198, + 478 + ], + "type": "text", + "content": "To summarize, our contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 287, + 611 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 479, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 287, + 504 + ], + "type": "text", + "content": "- Generating 3D meshes of room-scale indoor scenes with compelling textures and geometry from any text input." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 504, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 287, + 563 + ], + "type": "text", + "content": "- A method that leverages 2D text-to-image models and monocular depth estimation to lift frames into 3D in an iterative scene generation. Our proposed depth alignment and mesh fusion steps, enable us to create seamless and undistorted geometry and textures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 563, + 287, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 287, + 611 + ], + "type": "text", + "content": "- A two-stage tailored viewpoint selection that samples camera poses from optimal positions to first create the room layout and furniture and then close any remaining holes, creating a watertight mesh." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 621, + 134, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 134, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 134, + 634 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 641, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 689 + ], + "type": "text", + "content": "Text-based Generation has seen significant advances due to large-scale image-text datasets [73, 72, 14, 71] and scalable generative model architectures [16, 66, 60, 32], enabling synthesis of novel images from text [20, 3, 54]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Recently, diffusion models [78, 23, 80, 81, 82] achieved impressive results on image synthesis [15, 65, 67, 50, 58]" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "through improvements like latent space denoising [65, 84], faster sampling [23, 79, 52, 34], and better guidance [24]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 97, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 547, + 180 + ], + "type": "text", + "content": "In particular, text-to-image methods like Stable Diffusion [65], Imagen [67], GLIDE [50] and DALL-E 2 [58] yield diverse, high-fidelity, and controllable [6, 94] outputs. Text-based generation has been extended to other modalities including audio [35, 17, 28, 70], video [75, 91, 85, 25], and 4D fields [76]. We use text-to-image models by lifting their generated output into complete 3D scene meshes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 184, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 184, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 545, + 232 + ], + "type": "text", + "content": "Text-to-3D. Several methods use 3D data for supervised training of text-to-3D models [9, 51, 5]; however this direction remains challenging due to the lack of large-scale aligned datasets of text and 3D." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 232, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 304 + ], + "type": "text", + "content": "Alternative approaches use 2D vision-language models like CLIP [57] to create 3D content by formulating the generation as an optimization problem in the image domain [86, 29, 38, 48, 30] or as object alignment [69]. Related methods refine existing 3D input through text guidance in a similar fashion [46, 10, 87, 62]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 305, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 545, + 388 + ], + "type": "text", + "content": "Recent methods [56, 41, 45, 88, 44] combine large text-to-image diffusion models [65, 67] and neural radiance fields [47] to generate 3D objects without training. Other approaches train custom diffusion models on a similar text-to-3D task [39, 49, 12]. In contrast, we use a fixed text-to-image model and extract a 3D mesh representing entire scenes of many objects and structural elements like walls." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 392, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 548 + ], + "type": "text", + "content": "3D-Consistent View Synthesis from a Single Image. Several methods have been proposed that perform novel-view-synthesis from a single image [64, 90, 74, 61, 19]. Others optimize a neural 3D representation of an object, that can be viewed from arbitrary novel view points [92, 89, 1]. Another line of work performs perpetual view generation [77, 42, 40, 7], synthesizing videos via a render-refine-repeat pattern from a single RGB image that depict a scene along a forward-facing camera trajectory. In very recent concurrent work, Fridman et al. [18] create 3D scenes from text, but focus on this type of 3D-consistent \"zoom-out\" video generation. Instead, we generate complete, textured 3D room geometry from arbitrary trajectories." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 561, + 361, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 561, + 361, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 561, + 361, + 573 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Our method creates a textured 3D mesh of a complete scene from text input. To this end, we continuously fuse generated frames from a 2D text-to-image model at different poses into a joint 3D mesh, creating the scene over time. The core idea of our approach is a two-stage tailored viewpoint selection, that first generates the scene layout and objects and then closes remaining holes in the 3D geometry (Section 3.4). We visualize this workflow in Figure 2. For each pose in both stages, we apply an iterative scene generation scheme to update the mesh (Section 3.1). We first align each frame with the existing geometry with a depth align-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7910" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 402, + 134 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 402, + 134 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 402, + 134 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 402, + 134 + ], + "type": "image", + "image_path": "d0b92d778b3d5b2fe0a757649ba1e4ed4e9697e0ba97a52b85bd754ccbfbaec0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 137, + 282, + 149 + ], + "lines": [ + { + "bbox": [ + 170, + 137, + 282, + 149 + ], + "spans": [ + { + "bbox": [ + 170, + 137, + 282, + 149 + ], + "type": "text", + "content": "(a) Scene Generation Stage" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 403, + 70, + 545, + 134 + ], + "blocks": [ + { + "bbox": [ + 403, + 70, + 545, + 134 + ], + "lines": [ + { + "bbox": [ + 403, + 70, + 545, + 134 + ], + "spans": [ + { + "bbox": [ + 403, + 70, + 545, + 134 + ], + "type": "image", + "image_path": "67d6e7dc793e04bc55296ebbcf382d887b1ed15e3a373e64a5642dbd3a1febba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 137, + 529, + 149 + ], + "lines": [ + { + "bbox": [ + 415, + 137, + 529, + 149 + ], + "spans": [ + { + "bbox": [ + 415, + 137, + 529, + 149 + ], + "type": "text", + "content": "(b) Scene Completion Stage" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 230, + 294, + 392 + ], + "blocks": [ + { + "bbox": [ + 46, + 155, + 547, + 210 + ], + "lines": [ + { + "bbox": [ + 46, + 155, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 155, + 547, + 210 + ], + "type": "text", + "content": "Figure 2. Method overview. We iteratively create a textured 3D mesh in two stages. (a) First, we sample predefined poses and text to generate the complete scene layout and furniture. Each new pose (marked in green) adds newly generated geometry to the mesh (depicted by green triangles) in an iterative scene generation scheme (see Figure 3 for details). Blue poses/triangles denote viewpoints that created geometry in a previous iteration. (b) Second, we fill in the remaining unobserved regions by sampling additional poses (marked in red) after the scene layout is defined." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 230, + 294, + 392 + ], + "lines": [ + { + "bbox": [ + 50, + 230, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 50, + 230, + 294, + 392 + ], + "type": "image", + "image_path": "76fae7b7534754273b139f805925c072af7ac5b7cf560feef8f74589ffe696b4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 395, + 287, + 462 + ], + "lines": [ + { + "bbox": [ + 46, + 395, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 287, + 462 + ], + "type": "text", + "content": "Figure 3. Iterative scene generation. For each new pose, we render the current mesh to obtain partial RGB and depth renderings. We complete both, utilizing respective inpainting models and the text prompt. Next, we perform depth alignment (see Section 3.2) and mesh filtering (see Section 3.3) to obtain an optimal next mesh patch, that is finally fused with the existing geometry." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 503 + ], + "type": "text", + "content": "ment strategy (Section 3.2). Next, we triangulate and filter the novel content to merge it into the mesh (Section 3.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 510, + 211, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 211, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 211, + 521 + ], + "type": "text", + "content": "3.1. Iterative 3D Scene Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "text", + "content": "Our scene is represented as a mesh " + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = (\\mathcal{V},\\mathcal{C},\\mathcal{S})" + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "text", + "content": " where the vertices " + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{V}\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "text", + "content": ", vertex colors " + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\mathcal{C}\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "text", + "content": " and the face set " + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "inline_equation", + "content": "S\\in \\mathbb{N}_0^{M\\times 3}" + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "text", + "content": " are generated over time. Input to our method is a set of arbitrary text prompts " + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\{P_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "text", + "content": " that corresponds to our selected poses " + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\{E_t\\}_{t = 1}^T\\in \\mathbb{R}^{3\\times 4}" + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "text", + "content": " in both stages. Inspired by recent methods [42, 40], we iteratively build up the scene, following a render-refine-repeat pattern. We summarize this iterative scene generation process in Figure 3. Formally, for each step of generation " + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 528, + 287, + 649 + ], + "type": "text", + "content": ", we first render the current scene from a novel viewpoint:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 656, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 656, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 117, + 656, + 287, + 669 + ], + "type": "interline_equation", + "content": "I _ {t}, d _ {t}, m _ {t} = r \\left(\\mathcal {M} _ {t}, E _ {t}\\right), \\tag {1}", + "image_path": "a1520d2c0afe1a8cb3f9ce6877d4a11da88c9cfc8cf872302a6bc5810107e52e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " is a classical rasterization function without shading, " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " is the rendered image, " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " the rendered depth and " + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "m_{t}" + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": " the image-space mask, that marks pixels without observed" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 230, + 545, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 230, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 305, + 230, + 545, + 254 + ], + "type": "text", + "content": "content. We then use a fixed text-to-image model " + }, + { + "bbox": [ + 305, + 230, + 545, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{t2i}" + }, + { + "bbox": [ + 305, + 230, + 545, + 254 + ], + "type": "text", + "content": " to inpaint unobserved pixels according to the text prompt:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 380, + 263, + 545, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 263, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 380, + 263, + 545, + 277 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {t} = \\mathcal {F} _ {t 2 i} \\left(I _ {t}, m _ {t}, P _ {t}\\right). \\tag {2}", + "image_path": "8bc908efbd5d236a4e01fc1ea9834d70be69290f7a3664f921b8f761a6f66790.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 289, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 312 + ], + "type": "text", + "content": "Next, we inpaint unobserved depth by applying a monocular depth estimator " + }, + { + "bbox": [ + 305, + 289, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_d" + }, + { + "bbox": [ + 305, + 289, + 545, + 312 + ], + "type": "text", + "content": " in our depth alignment (see Section 3.2):" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 345, + 324, + 545, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 324, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 345, + 324, + 545, + 339 + ], + "type": "interline_equation", + "content": "\\hat {d} _ {t} = \\text {p r e d i c t - a n d - a l i g n} \\left(\\mathcal {F} _ {d}, I _ {t}, d _ {t}, m _ {t}\\right). \\tag {3}", + "image_path": "2f1c82ee44c2134ca035633cba114e19d990e208259bc692a255bcdc5771af4d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 346, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 372 + ], + "type": "text", + "content": "Finally, we combine the novel content " + }, + { + "bbox": [ + 305, + 346, + 545, + 372 + ], + "type": "inline_equation", + "content": "\\{\\hat{I}_t,\\hat{d}_t,m_t\\}" + }, + { + "bbox": [ + 305, + 346, + 545, + 372 + ], + "type": "text", + "content": " with the existing mesh by our fusion scheme (see Section 3.3):" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 354, + 380, + 545, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 380, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 354, + 380, + 545, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {t + 1} = f u s e \\left(\\mathcal {M} _ {t}, \\hat {I} _ {t}, \\hat {d} _ {t}, m _ {t}, E _ {t}\\right). \\tag {4}", + "image_path": "9003ae235344f0960e288c127d4d69793742fd9fd35ee3a256484c9edb3ac4ff.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 406, + 432, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 432, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 432, + 419 + ], + "type": "text", + "content": "3.2. Depth Alignment Step" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "text", + "content": "To lift a 2D image " + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 426, + 545, + 521 + ], + "type": "text", + "content": " into 3D, we predict the per-pixel depth. To correctly combine old and new content, it is necessary that both align with each other. In other words, similar regions in a scene like walls or furniture should be placed at similar depth. However, directly using the predicted depth for backprojection leads to hard cuts and discontinuities in the 3D geometry, since the depth is inconsistent in scale between subsequent viewpoints (see Figure 7a)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "text", + "content": "To this end, we perform depth alignment in two-stages. First, we use a state-of-the-art depth inpainting network [4] that takes ground-truth depth " + }, + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "text", + "content": " for known parts in the image as input and aligns the prediction to it: " + }, + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "inline_equation", + "content": "\\hat{d}_p = \\mathcal{F}_d(I,d)" + }, + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 571, + 545, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 545, + 607 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 545, + 607 + ], + "type": "text", + "content": "Inspired by Liu et al. [42] we then improve the result by optimizing for scale and shift parameters " + }, + { + "bbox": [ + 304, + 571, + 545, + 607 + ], + "type": "inline_equation", + "content": "\\gamma, \\beta \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 571, + 545, + 607 + ], + "type": "text", + "content": ", aligning predicted and rendered disparity in the least squares sense:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 356, + 616, + 545, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 616, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 356, + 616, + 545, + 651 + ], + "type": "interline_equation", + "content": "\\min _ {\\gamma , \\beta} \\quad \\left\\| m \\odot \\left(\\frac {\\gamma}{\\hat {d} _ {p}} + \\beta - \\frac {1}{d}\\right) \\right\\| ^ {2}, \\tag {5}", + "image_path": "c1f6a69d53aae9418bb9b104efce86e83857041ff7111a7e35c3eb78dc2a86ee.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 660, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 660, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 660, + 545, + 713 + ], + "type": "text", + "content": "where we mask out unobserved pixels via " + }, + { + "bbox": [ + 304, + 660, + 545, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 660, + 545, + 713 + ], + "type": "text", + "content": ". We can then extract the aligned depth as " + }, + { + "bbox": [ + 304, + 660, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{d} = (\\frac{\\gamma}{\\hat{d}_p} + \\beta)^{-1}" + }, + { + "bbox": [ + 304, + 660, + 545, + 713 + ], + "type": "text", + "content": ". Finally, we smooth " + }, + { + "bbox": [ + 304, + 660, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{d}" + }, + { + "bbox": [ + 304, + 660, + 545, + 713 + ], + "type": "text", + "content": " by applying a " + }, + { + "bbox": [ + 304, + 660, + 545, + 713 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 304, + 660, + 545, + 713 + ], + "type": "text", + "content": " Gaussian kernel at the mask edges (see supplemental material for more details)." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7911" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 132, + 166 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 132, + 166 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 132, + 166 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 132, + 166 + ], + "type": "image", + "image_path": "207a3b4cd3937f72df3864668ca2d82e905f71588dbc74ad1f280b31f55c3aa2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 48, + 170, + 283, + 181 + ], + "lines": [ + { + "bbox": [ + 48, + 170, + 283, + 181 + ], + "spans": [ + { + "bbox": [ + 48, + 170, + 283, + 181 + ], + "type": "text", + "content": "(a) Pixel Triangulation (b) Face Filtering (c) Mesh Fusion" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 141, + 70, + 216, + 166 + ], + "blocks": [ + { + "bbox": [ + 141, + 70, + 216, + 166 + ], + "lines": [ + { + "bbox": [ + 141, + 70, + 216, + 166 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 216, + 166 + ], + "type": "image", + "image_path": "56b07ccd1dde58e503e1527a1e776c02b6a334341cfbbb86190444be0a8cab52.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 184, + 286, + 250 + ], + "lines": [ + { + "bbox": [ + 46, + 184, + 286, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 184, + 286, + 250 + ], + "type": "text", + "content": "Figure 4. Visualization of our mesh fusion step. (a) We triangulate an image, such that 4 neighboring pixels (orange dots) create two faces. (b) We filter a face (marked in red), if its surface normal forms a small grazing angle with the viewing direction or if any edge in world space is too long. (c) We fuse the remaining faces (marked in green) with the existing geometry (marked in blue)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 221, + 70, + 277, + 166 + ], + "blocks": [ + { + "bbox": [ + 221, + 70, + 277, + 166 + ], + "lines": [ + { + "bbox": [ + 221, + 70, + 277, + 166 + ], + "spans": [ + { + "bbox": [ + 221, + 70, + 277, + 166 + ], + "type": "image", + "image_path": "a78e77db4e1e55dc29b4357bf074357eee9e02e94c4f2e8fb663e01bfad09ec3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 271, + 153, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 271, + 153, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 271, + 153, + 284 + ], + "type": "text", + "content": "3.3. Mesh Fusion Step" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 290, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 287, + 327 + ], + "type": "text", + "content": "At each step, we insert new content " + }, + { + "bbox": [ + 46, + 290, + 287, + 327 + ], + "type": "inline_equation", + "content": "\\{\\hat{I}_t,\\hat{d}_t,m_t\\}" + }, + { + "bbox": [ + 46, + 290, + 287, + 327 + ], + "type": "text", + "content": " into the scene. For that, we first backproject the image-space pixels into a world-space point cloud:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 336, + 287, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 336, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 70, + 336, + 287, + 352 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {t} = \\left\\{E _ {t} ^ {- 1} K ^ {- 1} \\cdot \\hat {d} _ {t} [ u, v ] \\cdot (u, v, 1) ^ {T} \\right\\} _ {u = 0, v = 0} ^ {W, H}, \\tag {6}", + "image_path": "5d44ef362dde5749694874ae08f1a3ef62fe1633a6f8d613eaa3951c9aa7354f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "text", + "content": " are the camera intrinsics and " + }, + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "inline_equation", + "content": "W, H" + }, + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "text", + "content": " are image width and height, respectively. We then use a simple triangulation scheme (Figure 4a), where each four neighboring pixels " + }, + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\{(u, v), (u + 1, v), (u, v + 1), (u + 1, v + 1)\\}" + }, + { + "bbox": [ + 46, + 363, + 287, + 460 + ], + "type": "text", + "content": " in the image form two triangles. Since the estimated depth is noisy, this naive triangulation creates stretched out 3D geometry (see Figure 7b). To alleviate this problem, we propose two filters that remove stretched out faces (Figure 4b)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 460, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 509 + ], + "type": "text", + "content": "First, we filter faces based on their edge length. We remove a face if the Euclidean distance of any face edge is larger than a threshold " + }, + { + "bbox": [ + 46, + 460, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\delta_{edge}" + }, + { + "bbox": [ + 46, + 460, + 287, + 509 + ], + "type": "text", + "content": ". Second, we filter faces based on the angle between surface normal and viewing direction:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 522, + 287, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 522, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 287, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = \\left\\{\\left(i _ {0}, i _ {1}, i _ {2}\\right) \\mid n ^ {T} v > \\delta_ {s n} \\right\\} \\tag {7}", + "image_path": "302bdf455fe7a8d88dbea8ac878c217591f1bf7a8f7c48e9b0ca063a1c636a6d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " is the face set, " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "(i_0,i_1,i_2)" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " are the vertex indices of the triangle, " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\delta_{sn}" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " is the threshold, " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "n\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " is the normalized face normal, and " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "v\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " is the normalized view direction in world space from the camera center towards the average pixel location from which the triangle originated. This avoids creating texture for large regions of the mesh from a comparatively small number of pixels from an image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Finally, we fuse together the newly generated mesh patch and the existing geometry (Figure 4c). All faces that are backprojected from pixels falling into the inpainting mask " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "m_{t}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " are stitched together with their neighboring faces, which are already part of the mesh. Precisely, we continue the triangulation scheme at all edges of " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "m_{t}" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": ", but use the existing vertex positions of " + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": " to create the corresponding faces." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 72, + 472, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 472, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 472, + 84 + ], + "type": "text", + "content": "3.4. Two-Stage Viewpoint Selection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 91, + 545, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 199 + ], + "type": "text", + "content": "A key part of our method is the choice of text prompts and camera poses from which the scene is synthesized. Users can in principle choose these inputs arbitrarily to create any desired indoor scene. However, the generated scene can degenerate and contain stretch and hole artifacts, if poses are chosen carelessly (see Figure 7 and supplemental material). To this end, we propose a two-stage viewpoint selection strategy, that samples each next camera pose from optimal positions and refines empty regions subsequently." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 202, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 545, + 322 + ], + "type": "text", + "content": "Generation Stage. In the first stage, we create the main parts of the scene, including the general layout and furniture. We subsequently render predefined trajectories in different directions that eventually cover the whole room. We found generation works best, if each trajectory starts off from a viewpoint with mostly unobserved regions. This generates the outline of the next chunk, while still being connected to the rest of the scene (e.g., see Figure 3). Then, we complete the 3D structure of that chunk by moving and rotating into it subsequently until the end of the trajectory." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 323, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 418 + ], + "type": "text", + "content": "Additionally, we ensure an optimal observation distance for each pose. We translates camera positions " + }, + { + "bbox": [ + 304, + 323, + 545, + 418 + ], + "type": "inline_equation", + "content": "T_0 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 323, + 545, + 418 + ], + "type": "text", + "content": " along the look-at direction " + }, + { + "bbox": [ + 304, + 323, + 545, + 418 + ], + "type": "inline_equation", + "content": "L \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 323, + 545, + 418 + ], + "type": "text", + "content": " uniformly: " + }, + { + "bbox": [ + 304, + 323, + 545, + 418 + ], + "type": "inline_equation", + "content": "T_{i+1} = T_i - 0.3L" + }, + { + "bbox": [ + 304, + 323, + 545, + 418 + ], + "type": "text", + "content": ". We stop if the mean rendered depth is larger than 0.1 or discard the camera after 10 steps. This avoids views too close to existing geometry. For example, the green pose in Figure 2a is moved back as far as possible into the existing geometry such that it views most of the empty floor region." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 420, + 545, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 516 + ], + "type": "text", + "content": "We create closed room layouts following this principle, by choosing trajectories that generate the next chunks in a circular motion, roughly centered around the origin. We found it helpful to discourage the text-to-image generator from generating furniture in unwanted regions by engineering the text prompts accordingly. For example, for poses looking at the floor or ceiling, we choose text prompts that only contain the words \"floor\" or \"ceiling\", respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 520, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 591 + ], + "type": "text", + "content": "Completion Stage. After the first stage, the scene layout and furniture is defined. However, it is impossible to choose sufficient poses " + }, + { + "bbox": [ + 304, + 520, + 545, + 591 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 520, + 545, + 591 + ], + "type": "text", + "content": "-priori. Since the scene is generated on-the-fly, the mesh contains holes that were not observed by any camera (see Figure 7c). We complete the scene by sampling additional poses " + }, + { + "bbox": [ + 304, + 520, + 545, + 591 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 520, + 545, + 591 + ], + "type": "text", + "content": "-posteriori, looking at those holes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 593, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 652 + ], + "type": "text", + "content": "Inspired by trajectory optimization [22, 63], we voxelize the scene into dense uniform cells. We sample random poses in each cell, discarding those being too close to existing geometry. We select one pose per cell that views most unobserved pixels (e.g., see the red poses in Figure 2b)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "Next, we inpaint the scene from all chosen camera poses following Section 3.1. Similar to Fridman et al. [18], we observe it is important to clean the inpainting masks, because our text-to-image generator can generate better results for large connected regions. Thus, we first inpaint small holes" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7912" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "with a classical inpainting algorithm [83] and dilate the remaining holes. We additionally remove all faces that fall into the dilated region and are close to the rendered depth. Please see the supplemental material for more details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 287, + 180 + ], + "type": "text", + "content": "Finally, we run Poisson surface reconstruction [33] on the scene mesh. This closes any remaining holes after completion and smoothens out discontinuities. The result is a watertight mesh of the generated scene, that can be rendered with classical rasterization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 192, + 101, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 101, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 101, + 205 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 216, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 287, + 396 + ], + "type": "text", + "content": "Implementation Details. We implement mesh rasterization and fusion with Pytorch3D [59]. As our text-to-image model " + }, + { + "bbox": [ + 46, + 216, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{t2i}" + }, + { + "bbox": [ + 46, + 216, + 287, + 396 + ], + "type": "text", + "content": ", we utilize a Stable Diffusion [65] model, that is finetuned on the image inpainting task, using additional mask input. We generate a single inpainting proposal and employ a state-of-the-art guided diffusion sampler [43]. As our monocular depth estimator " + }, + { + "bbox": [ + 46, + 216, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_d" + }, + { + "bbox": [ + 46, + 216, + 287, + 396 + ], + "type": "text", + "content": ", we employ an Iron-Depth [4] model, that is trained on indoor scenes from the ScanNet dataset [13] and augment it for depth inpainting according to Bae et al. [4]. We set " + }, + { + "bbox": [ + 46, + 216, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\delta_{edge} = 0.1" + }, + { + "bbox": [ + 46, + 216, + 287, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\delta_{sn} = 0.1" + }, + { + "bbox": [ + 46, + 216, + 287, + 396 + ], + "type": "text", + "content": " in all our experiments. During generation, we use 20 different trajectories with 10 frames each sampled between the respective start and end poses. We construct prompts using the guidelines suggested by Pierre [55]. Creating one scene takes approximately 50 minutes on one RTX 3090 GPU." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 399, + 287, + 626 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 46, + 399, + 287, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 446 + ], + "type": "text", + "content": "Baselines. To the best of our knowledge, there are no direct baselines that generate textured 3D room geometry from text. We compare against four related methods (please see the supplemental material for more details about baselines)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 447, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 447, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 447, + 287, + 483 + ], + "type": "text", + "content": "- PureClipNeRF [38]: We compare against text-to-3D methods for generating objects [56, 41, 29, 38, 88] and choose Lee et al. [38] as open-source representative." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 483, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 287, + 518 + ], + "type": "text", + "content": "- Outpainting [58, 53]: We combine outpainting from a Stable Diffusion [65] model with depth estimation and triangulation to create a mesh from an enlarged viewpoint." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 519, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 519, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 519, + 287, + 590 + ], + "type": "text", + "content": "- Text2Light [11]: We generate RGB panoramas from text using Chen et al. [11]. Estimating 3D mesh structure from a panorama is difficult. Related approaches estimate room layout [93], perform view synthesis [36, 26, 21, 27] or predict " + }, + { + "bbox": [ + 47, + 519, + 287, + 590 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 47, + 519, + 287, + 590 + ], + "type": "text", + "content": " depth [2, 31]. We perform depth prediction and subsequently apply our mesh fusion step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 591, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 287, + 626 + ], + "type": "text", + "content": "- Blockade [37]: We apply Blockade [37], which uses a text-to-image diffusion model to produce more expressive RGB panoramas. We then extract the mesh similarly." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "content": "Evaluation Metrics. The generated 3D geometry is evaluated both quantitatively and qualitatively. We calculate CLIP Score (CS) [57] and Inception Score (IS) [68] on RGB renderings of the respective scenes. Additionally, we conduct a user study and ask " + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "inline_equation", + "content": "n = 61" + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "content": " users to score Perceptual Quality (" + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "inline_equation", + "content": "PQ" + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "content": ") and 3D Structure Completeness (" + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "inline_equation", + "content": "3DS" + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "content": ") of the whole scene on a scale of 1-5." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 205 + ], + "type": "table", + "html": "
Method2D MetricsUser Study
CS ↑IS ↑PQ ↑3DS ↑
PureClipNeRF [38]24.061.262.342.38
Outpainting [58, 53]23.101.602.902.58
Text2Light [11]+Ours25.992.212.822.97
Blockade [37]+Ours26.292.133.353.36
Ours w/o alignment26.731.783.122.96
Ours w/o stretch removal27.721.863.283.75
Ours w/o completion27.972.183.723.87
Ours28.022.314.014.19
", + "image_path": "9a0e2e6b569ce4ea3422066d6f05c3a98075626a5b6f4aba41d890086a0adbb9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 213, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 258 + ], + "type": "text", + "content": "Table 1. Quantitative comparison. We report 2D metrics and user study results, including: Clip Score (CS), Inception Score (IS), Perceptual Quality (PQ), and 3D Structure Completeness (3DS). Our method creates scenes with the highest quality." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 279, + 418, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 279, + 418, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 279, + 418, + 293 + ], + "type": "text", + "content": "4.1. Qualitative Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 299, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 545, + 443 + ], + "type": "text", + "content": "We show top-down views into the scene and RGB renderings from within for our method and baselines in Figure 6. We show additional results of our method in Figure 5. PureClipNeRF [38] creates the key objects of the given text prompt, but does not create a complete 3D structure with floor, walls and ceilings. Outpainting [58, 53] creates high-detail textures, but projection from a single viewpoint creates holes due to occlusion and hinders the creation of complete 3D geometry. Text2Light [11] and Blockade [37] both create a high-detail " + }, + { + "bbox": [ + 304, + 299, + 545, + 443 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 299, + 545, + 443 + ], + "type": "text", + "content": " view of a complete scene, but occlusions that cannot be resolved from a single panoramic viewpoint lead to holes in the extracted 3D geometry." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 444, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 545, + 563 + ], + "type": "text", + "content": "In contrast, our approach creates high-detail textures and geometry, that are fused into a complete 3D scene mesh without holes. The resulting scenes contain flat floors, walls and ceilings, as well as 3D object geometry distributed throughout the scene. When specifying text prompts with a huge variety, the resulting scene contains a diverse set of objects. Please see the supplemental material for more scenes, animated results, intermediate outputs of our baselines (such as the panoramic images) as well as top-down views of meshes, that contain the reconstructed ceilings." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 574, + 424, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 424, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 424, + 586 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "content": "We show quantitative results averaged over multiple scenes in Table 1. We render 60 images from novel viewpoints for each scene to calculate the 2D metrics. We present users with multiple top-down views and renderings for each scene and let them rate each method individually (no side-by-side comparison). Stretched-out geometry and holes in the 3D geometry lead to lower scores for the baselines in all image-based metrics. Our approach achieves the highest scores, because the renderings are complete from arbitrary novel poses, satisfy the given text-prompt and" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7913" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 70, + 267, + 210 + ], + "blocks": [ + { + "bbox": [ + 76, + 70, + 267, + 210 + ], + "lines": [ + { + "bbox": [ + 76, + 70, + 267, + 210 + ], + "spans": [ + { + "bbox": [ + 76, + 70, + 267, + 210 + ], + "type": "image", + "image_path": "d2eca9ce87387bb4b59230539c92d088cf169eff0c69a25a872dfbe2afb4f335.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 269, + 70, + 392, + 140 + ], + "blocks": [ + { + "bbox": [ + 269, + 70, + 392, + 140 + ], + "lines": [ + { + "bbox": [ + 269, + 70, + 392, + 140 + ], + "spans": [ + { + "bbox": [ + 269, + 70, + 392, + 140 + ], + "type": "image", + "image_path": "eb9852895e792c867f2c9153ce6babf4c15383e7bacf3a169c80284e9f13c862.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 393, + 71, + 516, + 139 + ], + "blocks": [ + { + "bbox": [ + 393, + 71, + 516, + 139 + ], + "lines": [ + { + "bbox": [ + 393, + 71, + 516, + 139 + ], + "spans": [ + { + "bbox": [ + 393, + 71, + 516, + 139 + ], + "type": "image", + "image_path": "85cb3ff25fbe5385b65da195af10e3c71bb21ae14c1d4ed285cbdc5959d68f08.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 269, + 141, + 392, + 211 + ], + "blocks": [ + { + "bbox": [ + 269, + 141, + 392, + 211 + ], + "lines": [ + { + "bbox": [ + 269, + 141, + 392, + 211 + ], + "spans": [ + { + "bbox": [ + 269, + 141, + 392, + 211 + ], + "type": "image", + "image_path": "2e2041852c7c20bccf5e8b9dfe23c6f3d82d4b06e2b7be4fbf1bb70a631444c0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 393, + 141, + 516, + 211 + ], + "blocks": [ + { + "bbox": [ + 393, + 141, + 516, + 211 + ], + "lines": [ + { + "bbox": [ + 393, + 141, + 516, + 211 + ], + "spans": [ + { + "bbox": [ + 393, + 141, + 516, + 211 + ], + "type": "image", + "image_path": "41c2fb0160bb6ce1d137e34ec69bda5b2e94227da28e92b56b0c8a8fac70a26c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 214, + 502, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 214, + 502, + 225 + ], + "spans": [ + { + "bbox": [ + 91, + 214, + 502, + 225 + ], + "type": "text", + "content": "Editorial Style Photo, Coastal Bathroom, Clawfoot Tub, Seashell, Wicker, Mosaic Tile, Blue and White" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 100, + 228, + 244, + 365 + ], + "blocks": [ + { + "bbox": [ + 100, + 228, + 244, + 365 + ], + "lines": [ + { + "bbox": [ + 100, + 228, + 244, + 365 + ], + "spans": [ + { + "bbox": [ + 100, + 228, + 244, + 365 + ], + "type": "image", + "image_path": "d490125d2ee33df607bb19a86327e4afcda30b628d9ee3f401a42fa10fbe39d7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 269, + 226, + 392, + 295 + ], + "blocks": [ + { + "bbox": [ + 269, + 226, + 392, + 295 + ], + "lines": [ + { + "bbox": [ + 269, + 226, + 392, + 295 + ], + "spans": [ + { + "bbox": [ + 269, + 226, + 392, + 295 + ], + "type": "image", + "image_path": "feec8cc2863f2075c549047bbfca4f0c8aa7c0a50e616b0c31d6319261cfdb90.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 393, + 226, + 516, + 295 + ], + "blocks": [ + { + "bbox": [ + 393, + 226, + 516, + 295 + ], + "lines": [ + { + "bbox": [ + 393, + 226, + 516, + 295 + ], + "spans": [ + { + "bbox": [ + 393, + 226, + 516, + 295 + ], + "type": "image", + "image_path": "e0f2d1a79fa49f25f6f0a034cd2981799a201be4c59af3daa844be026d200ae2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 269, + 297, + 392, + 367 + ], + "blocks": [ + { + "bbox": [ + 269, + 297, + 392, + 367 + ], + "lines": [ + { + "bbox": [ + 269, + 297, + 392, + 367 + ], + "spans": [ + { + "bbox": [ + 269, + 297, + 392, + 367 + ], + "type": "image", + "image_path": "16c4e258a97397ad49226868831fb185695929825561df0fafe92cb295bf5c5f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 393, + 297, + 516, + 367 + ], + "blocks": [ + { + "bbox": [ + 393, + 297, + 516, + 367 + ], + "lines": [ + { + "bbox": [ + 393, + 297, + 516, + 367 + ], + "spans": [ + { + "bbox": [ + 393, + 297, + 516, + 367 + ], + "type": "image", + "image_path": "e6fbcfc3552c572956b37bd02099b975c4565c48770e4cbc7ce5d7ff6185d418.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 88, + 369, + 505, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 369, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 88, + 369, + 505, + 380 + ], + "type": "text", + "content": "A living room with a lit furnace, couch, and cozy curtains, bright lamps that make the room look well-lit" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 86, + 382, + 261, + 519 + ], + "blocks": [ + { + "bbox": [ + 86, + 382, + 261, + 519 + ], + "lines": [ + { + "bbox": [ + 86, + 382, + 261, + 519 + ], + "spans": [ + { + "bbox": [ + 86, + 382, + 261, + 519 + ], + "type": "image", + "image_path": "db647f626426a2292c35eff9ef458c3ee87607c2ddf0083b4f5759a873122502.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 269, + 381, + 392, + 450 + ], + "blocks": [ + { + "bbox": [ + 269, + 381, + 392, + 450 + ], + "lines": [ + { + "bbox": [ + 269, + 381, + 392, + 450 + ], + "spans": [ + { + "bbox": [ + 269, + 381, + 392, + 450 + ], + "type": "image", + "image_path": "15ca036db5bf5940cde79bfa1f66b04ed9f7129e401f712738fa684052ec52c3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 393, + 381, + 516, + 451 + ], + "blocks": [ + { + "bbox": [ + 393, + 381, + 516, + 451 + ], + "lines": [ + { + "bbox": [ + 393, + 381, + 516, + 451 + ], + "spans": [ + { + "bbox": [ + 393, + 381, + 516, + 451 + ], + "type": "image", + "image_path": "1fd22bc3048c7b7f28f67b3eb0c36462f2c6889038e2ab2f6d1a984e58dbfbed.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 269, + 452, + 392, + 521 + ], + "blocks": [ + { + "bbox": [ + 269, + 452, + 392, + 521 + ], + "lines": [ + { + "bbox": [ + 269, + 452, + 392, + 521 + ], + "spans": [ + { + "bbox": [ + 269, + 452, + 392, + 521 + ], + "type": "image", + "image_path": "afbbc7a61702516a0571f1f09d412731d1f64f18083493f67c0539cb37a33927.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 393, + 452, + 516, + 521 + ], + "blocks": [ + { + "bbox": [ + 393, + 452, + 516, + 521 + ], + "lines": [ + { + "bbox": [ + 393, + 452, + 516, + 521 + ], + "spans": [ + { + "bbox": [ + 393, + 452, + 516, + 521 + ], + "type": "image", + "image_path": "bb525f53e3f383cd059cbe20736408a07be28d5338f624d2a5ea681b1ad177d1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 76, + 525, + 517, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 525, + 517, + 536 + ], + "spans": [ + { + "bbox": [ + 76, + 525, + 517, + 536 + ], + "type": "text", + "content": "Editorial Style Photo, Modern Living Room, Large Window, Leather, Glass, Metal, Wood Paneling, Apartment" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 100, + 538, + 246, + 676 + ], + "blocks": [ + { + "bbox": [ + 100, + 538, + 246, + 676 + ], + "lines": [ + { + "bbox": [ + 100, + 538, + 246, + 676 + ], + "spans": [ + { + "bbox": [ + 100, + 538, + 246, + 676 + ], + "type": "image", + "image_path": "251399c527f029c48a652a0f5a880cbfcc8a0ca5cae9b8fec8486ebe0d696fa6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 692, + 545, + 724 + ], + "lines": [ + { + "bbox": [ + 46, + 692, + 545, + 724 + ], + "spans": [ + { + "bbox": [ + 46, + 692, + 545, + 724 + ], + "type": "text", + "content": "Figure 5. 3D scene generation results of our method. We show color and shaded geometry renderings from generated scenes with corresponding text prompts. Our method synthesizes realistic meshes satisfying text descriptions. We remove the ceiling in the top-down view for better visualization of the scene layout." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 269, + 537, + 392, + 605 + ], + "blocks": [ + { + "bbox": [ + 269, + 537, + 392, + 605 + ], + "lines": [ + { + "bbox": [ + 269, + 537, + 392, + 605 + ], + "spans": [ + { + "bbox": [ + 269, + 537, + 392, + 605 + ], + "type": "image", + "image_path": "d1fe957ca3366c375343d3fb08bb1c1335739ec6885d0d06bc6b4aa248254d7a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 393, + 536, + 516, + 605 + ], + "blocks": [ + { + "bbox": [ + 393, + 536, + 516, + 605 + ], + "lines": [ + { + "bbox": [ + 393, + 536, + 516, + 605 + ], + "spans": [ + { + "bbox": [ + 393, + 536, + 516, + 605 + ], + "type": "image", + "image_path": "bfbe32a7c7e0ce74676c8b8dbf14c8080842fa5057bcf199bb289fda05771198.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 269, + 607, + 392, + 677 + ], + "blocks": [ + { + "bbox": [ + 269, + 607, + 392, + 677 + ], + "lines": [ + { + "bbox": [ + 269, + 607, + 392, + 677 + ], + "spans": [ + { + "bbox": [ + 269, + 607, + 392, + 677 + ], + "type": "image", + "image_path": "abd48c60ff69b6a59a731b9277017eccf017e10828ae6d4a62bcb1fffb2565b1.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 393, + 607, + 516, + 677 + ], + "blocks": [ + { + "bbox": [ + 393, + 607, + 516, + 677 + ], + "lines": [ + { + "bbox": [ + 393, + 607, + 516, + 677 + ], + "spans": [ + { + "bbox": [ + 393, + 607, + 516, + 677 + ], + "type": "image", + "image_path": "d0f77186f66663241a634ce076d5f236a931fb67b7baa52c9aa03205830d021e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 75, + 680, + 517, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 680, + 517, + 691 + ], + "spans": [ + { + "bbox": [ + 75, + 680, + 517, + 691 + ], + "type": "text", + "content": "Editorial Style Photo, Modern Nursery, Table Lamp, Rocking Chair, Tree Wall Decal, Wood, Cotton, Faux Fur" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7914" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 541, + 220 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 541, + 220 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 541, + 220 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 541, + 220 + ], + "type": "image", + "image_path": "cc180d7990d0664bd23cc9d4f05424da14d560fd3eea4a2dd21c1a2b7196be2f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 222, + 520, + 234 + ], + "lines": [ + { + "bbox": [ + 70, + 222, + 520, + 234 + ], + "spans": [ + { + "bbox": [ + 70, + 222, + 520, + 234 + ], + "type": "text", + "content": "Editorial Style Photo, Industrial Home Office, Steel Shelves, Concrete, Metal, Edison Bulbs, Exposed Ductwork" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 235, + 541, + 375 + ], + "blocks": [ + { + "bbox": [ + 53, + 235, + 541, + 375 + ], + "lines": [ + { + "bbox": [ + 53, + 235, + 541, + 375 + ], + "spans": [ + { + "bbox": [ + 53, + 235, + 541, + 375 + ], + "type": "image", + "image_path": "4bead8a57fc0d8fcfc81dbe818a3b5c86589f67a4e4d00571b6ed51932708cc8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 377, + 403, + 388 + ], + "lines": [ + { + "bbox": [ + 187, + 377, + 403, + 388 + ], + "spans": [ + { + "bbox": [ + 187, + 377, + 403, + 388 + ], + "type": "text", + "content": "a bedroom with a king-size bed and a large wardrobe" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 389, + 139, + 401 + ], + "lines": [ + { + "bbox": [ + 59, + 389, + 139, + 401 + ], + "spans": [ + { + "bbox": [ + 59, + 389, + 139, + 401 + ], + "type": "text", + "content": "PureClipNeRF [38]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 389, + 239, + 401 + ], + "lines": [ + { + "bbox": [ + 156, + 389, + 239, + 401 + ], + "spans": [ + { + "bbox": [ + 156, + 389, + 239, + 401 + ], + "type": "text", + "content": "Outpainting [58, 53]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 389, + 341, + 400 + ], + "lines": [ + { + "bbox": [ + 249, + 389, + 341, + 400 + ], + "spans": [ + { + "bbox": [ + 249, + 389, + 341, + 400 + ], + "type": "text", + "content": "Text2Light [11]+Ours" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 351, + 389, + 436, + 400 + ], + "lines": [ + { + "bbox": [ + 351, + 389, + 436, + 400 + ], + "spans": [ + { + "bbox": [ + 351, + 389, + 436, + 400 + ], + "type": "text", + "content": "Blockade [37]+Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 389, + 503, + 399 + ], + "lines": [ + { + "bbox": [ + 481, + 389, + 503, + 399 + ], + "spans": [ + { + "bbox": [ + 481, + 389, + 503, + 399 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 402, + 547, + 446 + ], + "lines": [ + { + "bbox": [ + 46, + 402, + 547, + 446 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 547, + 446 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of our method and baselines. PureClipNeRF [38] cannot produce immersive scenes with floors and walls. Outpainting [58, 53] does not produce 3D consistent scenes. Text2Light [11] and Blockade [37] both have holes due to occlusions. In contrast, our method creates complete meshes without holes and high details. We remove the ceiling in the top-down view for better visualization of the scene layout." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 466, + 288, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 288, + 502 + ], + "type": "text", + "content": "contain high-resolution image features. Users prefer our method, which highlights the quality of our accurate and complete geometry, as well as the RGB texture." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 510, + 115, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 115, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 115, + 521 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 528, + 287, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 576 + ], + "type": "text", + "content": "The key ingredients of our method are depth alignment (Section 3.2), mesh fusion (Section 3.3) and the two-stage viewpoint selection (Section 3.4). We demonstrate the importance of each component in Figure 7 and Table 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 578, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 662 + ], + "type": "text", + "content": "Depth alignment creates seamless scenes. Monocular depth predictions from subsequent frames can be inconsistent in scale. This leads to disconnected components in the mesh that are backprojected from multiple viewpoints (see Figure 7a). Our depth alignment strategy allows fusing multiple frames into a seamless mesh, eventually creating a complete scene with flat floors, walls, ceilings and no holes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": "Stretch removal creates undistorted scene geometry. During mesh fusion, we update the scene geometry with the contents of the next frame. Due to noisy depth prediction, the objects become stretched out, if they are observed" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 466, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 525 + ], + "type": "text", + "content": "from small grazing angles. Thus, we propose two filters (edge length and surface normal thresholds) that alleviate this issue. Instead of baking in stretched-out geometry (see Figure 7b), we disregard the corresponding faces and let the object be completed from a more suitable, later viewpoint." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 529, + 546, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 546, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 546, + 637 + ], + "type": "text", + "content": "Two-stage generation creates complete scenes. Our approach chooses camera poses in two stages to create a complete scene without holes. After generating the scene from predefined trajectories, the scene still contains some holes (see Figure 7c). Because the scene is built-up over time, it is impossible to choose camera poses " + }, + { + "bbox": [ + 304, + 529, + 546, + 637 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 529, + 546, + 637 + ], + "type": "text", + "content": "-priori, that view all unobserved regions. To this end, our completion stage samples poses " + }, + { + "bbox": [ + 304, + 529, + 546, + 637 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 529, + 546, + 637 + ], + "type": "text", + "content": "-posteriori to refine those regions. The resulting mesh is watertight and contains no holes (see Figure 7d)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 647, + 495, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 495, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 495, + 659 + ], + "type": "text", + "content": "4.4. Spatially Varying Scene Generation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 713 + ], + "type": "text", + "content": "Our method can be applied to generate a scene as the combination of multiple text prompts. Specifically, we use separate text prompts for different poses, crafting a set of trajectories that spatially combines scene descriptions. This" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7915" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 175, + 180 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 175, + 180 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 175, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 175, + 180 + ], + "type": "image", + "image_path": "e782e9c77329ce1310e5a56e2d0239a7e09936253dc60d2d7ef1599b93f0a242.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 187, + 69, + 282, + 180 + ], + "blocks": [ + { + "bbox": [ + 187, + 69, + 282, + 180 + ], + "lines": [ + { + "bbox": [ + 187, + 69, + 282, + 180 + ], + "spans": [ + { + "bbox": [ + 187, + 69, + 282, + 180 + ], + "type": "image", + "image_path": "e791f881e28aef630bc565e2c60edd16ec6e56a89c90166cb463e1b068799fc5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 296, + 72, + 415, + 179 + ], + "blocks": [ + { + "bbox": [ + 296, + 72, + 415, + 179 + ], + "lines": [ + { + "bbox": [ + 296, + 72, + 415, + 179 + ], + "spans": [ + { + "bbox": [ + 296, + 72, + 415, + 179 + ], + "type": "image", + "image_path": "333a53c14136c1e5eb79546a8c33a5cab54f49160f7ef6a6dd762f04155e0afa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 417, + 70, + 538, + 180 + ], + "blocks": [ + { + "bbox": [ + 417, + 70, + 538, + 180 + ], + "lines": [ + { + "bbox": [ + 417, + 70, + 538, + 180 + ], + "spans": [ + { + "bbox": [ + 417, + 70, + 538, + 180 + ], + "type": "image", + "image_path": "86ddbd6b0aac7e63748ff124b8ffc4f800cf145d629885f11fd0d8abe2c81aa4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 53, + 182, + 175, + 277 + ], + "blocks": [ + { + "bbox": [ + 53, + 182, + 175, + 277 + ], + "lines": [ + { + "bbox": [ + 53, + 182, + 175, + 277 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 175, + 277 + ], + "type": "image", + "image_path": "054d5f3dda57bf7f8d6fdb9471a1591706a6bb1706c3c841c9c863cf0e95a81a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 293, + 162, + 304 + ], + "lines": [ + { + "bbox": [ + 67, + 293, + 162, + 304 + ], + "spans": [ + { + "bbox": [ + 67, + 293, + 162, + 304 + ], + "type": "text", + "content": "(a) Ours w/o alignment" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 175, + 182, + 295, + 277 + ], + "blocks": [ + { + "bbox": [ + 175, + 182, + 295, + 277 + ], + "lines": [ + { + "bbox": [ + 175, + 182, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 175, + 182, + 295, + 277 + ], + "type": "image", + "image_path": "9287746b24ed8c0550ea06d2dc2ebdcb740b8194965f4455273c97f30965d986.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 293, + 293, + 303 + ], + "lines": [ + { + "bbox": [ + 177, + 293, + 293, + 303 + ], + "spans": [ + { + "bbox": [ + 177, + 293, + 293, + 303 + ], + "type": "text", + "content": "(b) Ours w/o stretch removal" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 296, + 182, + 416, + 277 + ], + "blocks": [ + { + "bbox": [ + 296, + 182, + 416, + 277 + ], + "lines": [ + { + "bbox": [ + 296, + 182, + 416, + 277 + ], + "spans": [ + { + "bbox": [ + 296, + 182, + 416, + 277 + ], + "type": "image", + "image_path": "e4299f0a5e9e71d98de2892563c77d4a937192577c965e69cc711bf77ef3f1aa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 293, + 405, + 304 + ], + "lines": [ + { + "bbox": [ + 306, + 293, + 405, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 405, + 304 + ], + "type": "text", + "content": "(c) Ours w/o completion" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 417, + 182, + 537, + 277 + ], + "blocks": [ + { + "bbox": [ + 417, + 182, + 537, + 277 + ], + "lines": [ + { + "bbox": [ + 417, + 182, + 537, + 277 + ], + "spans": [ + { + "bbox": [ + 417, + 182, + 537, + 277 + ], + "type": "image", + "image_path": "b4b15330369576b7b9e41ef08aba4d3bc87839299c8f86ba1f45e5320c023c1b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 293, + 506, + 304 + ], + "lines": [ + { + "bbox": [ + 447, + 293, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 447, + 293, + 506, + 304 + ], + "type": "text", + "content": "(d) Ours (full)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 196, + 281, + 394, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 281, + 394, + 292 + ], + "spans": [ + { + "bbox": [ + 196, + 281, + 394, + 292 + ], + "type": "text", + "content": "living room, couches, curtains, lit furnace, lamps" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 366, + 159, + 436 + ], + "blocks": [ + { + "bbox": [ + 46, + 304, + 547, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 547, + 350 + ], + "type": "text", + "content": "Figure 7. Ablation study on the key components of our method. Without depth alignment (see Section 3.2), different parts of the scene are disconnected and do not fuse into a seamless mesh. Without edge and surface normal thresholds (see Section 3.3), many faces are stretched out unnaturally. Without completion (see Section 3.4), the mesh has holes in remaining unobserved regions. Our full pipeline creates complete, high-resolution scenes. We remove the ceiling in the top-down view for better visualization of the scene layout." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 366, + 159, + 436 + ], + "lines": [ + { + "bbox": [ + 59, + 366, + 159, + 436 + ], + "spans": [ + { + "bbox": [ + 59, + 366, + 159, + 436 + ], + "type": "image", + "image_path": "c755f295f8051c799a7331e05b23b30ac84c54b05e8dd50fc27058b529e1a396.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 162, + 366, + 274, + 436 + ], + "blocks": [ + { + "bbox": [ + 162, + 366, + 274, + 436 + ], + "lines": [ + { + "bbox": [ + 162, + 366, + 274, + 436 + ], + "spans": [ + { + "bbox": [ + 162, + 366, + 274, + 436 + ], + "type": "image", + "image_path": "c926aef2331ef8854368beeff4730020aa312b0477b3bba85b8762f26dc73461.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 71, + 463, + 147, + 532 + ], + "blocks": [ + { + "bbox": [ + 58, + 439, + 276, + 462 + ], + "lines": [ + { + "bbox": [ + 58, + 439, + 276, + 462 + ], + "spans": [ + { + "bbox": [ + 58, + 439, + 276, + 462 + ], + "type": "text", + "content": "kitchen, dinner table, dishwashers, ovens, countertops living room, lit furnace, couch, curtains" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 463, + 147, + 532 + ], + "lines": [ + { + "bbox": [ + 71, + 463, + 147, + 532 + ], + "spans": [ + { + "bbox": [ + 71, + 463, + 147, + 532 + ], + "type": "image", + "image_path": "006d84408941ae6fe0888f7fe3b2ba6cfd328ae279fe2e48b69f3f6093298268.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 162, + 463, + 274, + 533 + ], + "blocks": [ + { + "bbox": [ + 162, + 463, + 274, + 533 + ], + "lines": [ + { + "bbox": [ + 162, + 463, + 274, + 533 + ], + "spans": [ + { + "bbox": [ + 162, + 463, + 274, + 533 + ], + "type": "image", + "image_path": "806162c801d775f3e05192c3bf5d0667a193e2ee64958ea3f370e7072ac108b4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 536, + 238, + 557 + ], + "lines": [ + { + "bbox": [ + 96, + 536, + 238, + 557 + ], + "spans": [ + { + "bbox": [ + 96, + 536, + 238, + 557 + ], + "type": "text", + "content": "bathroom, shower, bathtub bedroom, king-size bed, wardrobe" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 558, + 287, + 603 + ], + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 603 + ], + "type": "text", + "content": "Figure 8. Spatially varying scene generation. Our method can create rooms with multiple parts by prompt mixing. We use separate prompts for cameras viewing different parts of the scene. This is a controllable way to create rooms from multiple descriptions." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "can be desired to avoid repeating elements in a complete scene (e.g., multiple couches could be spread out over the whole room when using the same prompt for every camera). Instead, users can specify different object positions through different camera poses and text prompts. It can also be used to design a house comprised of multiple rooms, each with a different type (e.g., a living room that leads to a kitchen)." + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 313, + 367, + 431, + 430 + ], + "blocks": [ + { + "bbox": [ + 313, + 367, + 431, + 430 + ], + "lines": [ + { + "bbox": [ + 313, + 367, + 431, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 431, + 430 + ], + "type": "image", + "image_path": "2e3bce1ef21548390e5a6e0bb425761c89423b7880322207bea90af4e6ecb3a9.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 432, + 366, + 540, + 434 + ], + "blocks": [ + { + "bbox": [ + 432, + 366, + 540, + 434 + ], + "lines": [ + { + "bbox": [ + 432, + 366, + 540, + 434 + ], + "spans": [ + { + "bbox": [ + 432, + 366, + 540, + 434 + ], + "type": "image", + "image_path": "547a261cc20522c050bb21615ed515fba8232845c711e2859e1b180560c04f69.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 436, + 545, + 480 + ], + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 480 + ], + "type": "text", + "content": "Figure 9. Scene generation with layout guidance. Our method can generate scenes from layout guidance. Left: we describe objects with different prompts for cameras facing at different directions. Right: the generated part of the room." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 502, + 544, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 544, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 544, + 526 + ], + "type": "text", + "content": "We show results that combine multiple text prompts in Figure 8 and Figure 9." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 527, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 527, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 527, + 545, + 588 + ], + "type": "text", + "content": "We note that the layout can only be partially controlled by the camera poses, since scene generation can create chunks with larger or smaller extent. We believe this demonstrates an exciting application of our method, that can be further explored in future work." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 597, + 383, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 383, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 383, + 609 + ], + "type": "text", + "content": "4.5. Limitations" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": "Our approach allows to generate 3D room geometry from arbitrary text prompts that are highly detailed and contain consistent geometry. Nevertheless, our method can still fail under certain conditions (see supplemental material). First, our thresholding scheme (see Section 3.3) may not detect all stretched-out regions, which may lead to remaining distortions. Additionally, some holes may still not be completed fully after the second stage (see Section 3.4)," + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7916" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 120 + ], + "type": "text", + "content": "which results in over-smoothed regions after applying poisson reconstruction. Our scene representation does not decompose material from lighting, which bakes in shadows or bright lamps, that are generated from the diffusion model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 129, + 120, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 120, + 141 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 120, + 141 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 149, + 289, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 149, + 289, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 289, + 293 + ], + "type": "text", + "content": "We have shown a method to generate textured 3D meshes from only text input. We use text-to-image 2D generators to create a sequence of images. The core insight of our method is a tailored viewpoint selection, that allows to create a 3D mesh with seamless geometry and compelling textures. Specifically, we lift the images into a 3D scene, by employing our alignment strategy that iteratively fuses all images into the mesh. Our output meshes represent arbitrary indoor scenes that can be rendered with classical rasterization pipelines. We believe our approach demonstrates an exciting application of large-scale 3D asset creation, that only requires text as input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 302, + 149, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 302, + 149, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 302, + 149, + 316 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 320, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 320, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 320, + 287, + 387 + ], + "type": "text", + "content": "This work was funded in part by Cisco Systems. It was also supported by the ERC Starting Grant Scan2CAD (804724) as well as the German Research Foundation (DFG) Research Unit \"Learning and Simulation in Visual Computing.\" We also thank Angela Dai for the video voice over." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 396, + 106, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 396, + 106, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 106, + 409 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 416, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 416, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 416, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 53, + 416, + 287, + 459 + ], + "type": "text", + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J. Mitra, and Paul Guerrero. RenderDiffusion: Image diffusion for 3D reconstruction, inpainting and generation. arXiv, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 460, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 460, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 53, + 460, + 287, + 503 + ], + "type": "text", + "content": "[2] Manuel Rey Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution " + }, + { + "bbox": [ + 53, + 460, + 287, + 503 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 53, + 460, + 287, + 503 + ], + "type": "text", + "content": " monocular depth estimation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 505, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 505, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 53, + 505, + 287, + 548 + ], + "type": "text", + "content": "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18187-18197, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 287, + 591 + ], + "type": "text", + "content": "[4] Gwangbin Bae, Ignas Budvytis, and Roberto Cipolla. Iron-depth: Iterative refinement of single-view depth using surface normal and its uncertainty. In British Machine Vision Conference (BMVC), 2022. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 593, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 593, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 53, + 593, + 287, + 646 + ], + "type": "text", + "content": "[5] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh Susskind. Gaudi: A neural architect for immersive 3d scene generation. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 647, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 287, + 678 + ], + "type": "text", + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 680, + 287, + 712 + ], + "type": "text", + "content": "[7] Shengqu Cai, Eric Ryan Chan, Songyou Peng, Mohamad Shahbazi, Anton Obukhov, Luc Van Gool, and Gordon Wetzstein. Diffdreamer: Consistent single-view perpetual view" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "generation with conditional diffusion models. In arXiV, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 150 + ], + "type": "text", + "content": "[8] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, L. Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository. ArXiv, abs/1512.03012, 2015. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 545, + 205 + ], + "type": "text", + "content": "[9] Kevin Chen, Christopher Bongsoo Choy, Manolis Savva, Angel X. Chang, Thomas A. Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. ArXiv, abs/1803.08495, 2018. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[10] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "text", + "content": "[11] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "type": "text", + "content": "[12] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. arXiv preprint arXiv:2212.04493, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "type": "text", + "content": "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "type": "text", + "content": "[14] Karan Desai, Gaurav Kaul, Zubin Trivadi Aysola, and Justin Johnson. Redcaps: Web-curated image-text data created by the people, for the people. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 445, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 476 + ], + "type": "text", + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 521 + ], + "type": "text", + "content": "[16] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868-12878, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "text", + "content": "[17] Seth Forsgren* and Hayk Martiros*.iffusion - Stable diffusion for real-time music generation, https://riffusion.com/about, accessed 2023-03-06, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "text", + "content": "[18] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. arXiv preprint arXiv:2302.01133, 2023. 1, 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 602, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 655 + ], + "type": "text", + "content": "[19] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M. Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. ArXiv, abs/2302.10109, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[20] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10686-10696, 2021. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7917" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[21] Takayuki Hara and Tatsuya Harada. Enhancement of novel view synthesis using omnidirectional image completion. arXiv preprint arXiv:2203.09957, 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "text", + "content": "[22] Benjamin Hepp, Matthias Nießner, and Otmar Hilliges. Plan3d: Viewpoint and trajectory optimization for aerial multi-view stereo reconstruction. ACM Transactions on Graphics (TOG), 38(1):1-17, 2018. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "text", + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "text", + "content": "[24] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[25] Jonathan Ho, Tim Salimans, Alexey A Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 303 + ], + "type": "text", + "content": "[26] Ching-Yu Hsu, Cheng Sun, and Hwann-Tzong Chen. Moving in a 360 world: Synthesizing panoramic parallaxes from a single panorama. arXiv preprint arXiv:2106.10859, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "text", + "content": "[27] Huajian Huang, Yingshu Chen, Tianjian Zhang, and SaiKit Yeung. 360room: Real-time indoor roaming using geometry-aware " + }, + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "text", + "content": " radiance fields. arXiv preprint arXiv:2208.02705, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 414 + ], + "type": "text", + "content": "[28] Qingqing Huang, Daniel S. Park, Tao Wang, Timo I. Denk, Andy Ly, Nanxin Chen, Zhengdong Zhang, Zhishuai Zhang, Jia Yu, C. Frank, Jesse Engel, Quoc V. Le, William Chan, and Weixiang Han. Noise2music: Text-conditioned music generation with diffusion models. ArXiv, abs/2302.03917, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 470 + ], + "type": "text", + "content": "[29] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 857-866. IEEE Computer Society, 2022. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[30] Zutao Jiang, Guangsong Lu, Xiaodan Liang, Jihua Zhu, Wei Zhang, Xiaojun Chang, and Hang Xu. 3d-togo: Towards text-guided cross-category 3d object generation. arXiv preprint arXiv:2212.01103, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 579 + ], + "type": "text", + "content": "[31] Lei Jin, Yanyu Xu, Jia Zheng, Junfei Zhang, Rui Tang, Shugong Xu, Jingyi Yu, and Shenghua Gao. Geometric structure based and regularized depth estimation from 360 indoor imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 889-898, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "text", + "content": "[32] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "text", + "content": "[33] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, page 0, 2006. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[34] Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models. In ICML Workshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likelihood Models. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[35] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 152 + ], + "type": "text", + "content": "[36] Shreyas Kulkarni, Peng Yin, and Sebastian Scherer. 360fusionnerf: Panoramic neural radiance fields with joint guidance. arXiv preprint arXiv:2209.14265, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 153, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 185 + ], + "type": "text", + "content": "[37] Blockade Labs. Blockade skybox, https://skybox.blockadelabs.com/, accessed 2023-03-04. 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 188, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 188, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 188, + 545, + 220 + ], + "type": "text", + "content": "[38] Han-Hung Lee and Angel X Chang. Understanding pure clip guidance for voxel grid nerf models. arXiv preprint arXiv:2209.15172, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 223, + 545, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 223, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 307, + 223, + 545, + 267 + ], + "type": "text", + "content": "[39] Gang Li, Heliang Zheng, Chaoyue Wang, Chang Li, Changwen Zheng, and Dacheng Tao. 3ddesigner: Towards photorealistic 3d object generation and editing with text-guided diffusion models. arXiv preprint arXiv:2211.14108, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 269, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 269, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 269, + 545, + 312 + ], + "type": "text", + "content": "[40] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 316, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 369 + ], + "type": "text", + "content": "[41] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 1, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 372, + 545, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 427 + ], + "type": "text", + "content": "[42] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2021. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 429, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 429, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 307, + 429, + 545, + 472 + ], + "type": "text", + "content": "[43] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 475, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 508 + ], + "type": "text", + "content": "[44] Luke Melas-Kyriazi, C. Rupprecht, Iro Laina, and Andrea Vedaldi. Realfusion: " + }, + { + "bbox": [ + 307, + 475, + 545, + 508 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 307, + 475, + 545, + 508 + ], + "type": "text", + "content": " reconstruction of any object from a single image. ArXiv, abs/2302.10663, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 510, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 552 + ], + "type": "text", + "content": "[45] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. arXiv preprint arXiv:2211.07600, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "text", + "content": "[46] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 612, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 666 + ], + "type": "text", + "content": "[47] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[48] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7918" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[49] Gimin Nam, Mariem Khlifi, Andrew Rodriguez, Alberto Tono, Linqi Zhou, and Paul Guerrero. 3d-ldm: Neural implicit 3d shape generation with latent diffusion models. arXiv preprint arXiv:2212.00842, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 172 + ], + "type": "text", + "content": "[50] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 172, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 172, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 172, + 288, + 215 + ], + "type": "text", + "content": "[51] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 217, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 288, + 259 + ], + "type": "text", + "content": "[52] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 261, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 261, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 49, + 261, + 287, + 303 + ], + "type": "text", + "content": "[53] OpenAI. Dalle: Introducing outpainting, https://openai.com/blog/dall-e-introducing-outpainting?utm_source=tldrnewsletter, accessed 2023-03-07. 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 304, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 304, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 49, + 304, + 287, + 348 + ], + "type": "text", + "content": "[54] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2065–2074, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 350, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 287, + 380 + ], + "type": "text", + "content": "[55] Nick St. Pierre. Additive prompting, https://twitter.com/nickfloats/status/1628796348446253057, accessed 2023-03-07. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 382, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 382, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 49, + 382, + 287, + 414 + ], + "type": "text", + "content": "[56] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 1, 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 415, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 415, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 49, + 415, + 287, + 480 + ], + "type": "text", + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 482, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 482, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 49, + 482, + 287, + 523 + ], + "type": "text", + "content": "[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 49, + 525, + 287, + 568 + ], + "type": "text", + "content": "[59] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 570, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 49, + 570, + 287, + 602 + ], + "type": "text", + "content": "[60] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 603, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 603, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 49, + 603, + 287, + 656 + ], + "type": "text", + "content": "[61] Xuanchi Ren and Xiaolong Wang. Look outside the room: Synthesizing a consistent long-term 3d scene video from a single image. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3553-3563, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 658, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 287, + 690 + ], + "type": "text", + "content": "[62] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 692, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 287, + 712 + ], + "type": "text", + "content": "[63] Mike Roberts, Debadeepta Dey, Anh Truong, Sudipta Sinha, Shital Shah, Ashish Kapoor, Pat Hanrahan, and Neel Joshi." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Submodular trajectory optimization for aerial 3d scanning. In Proceedings of the IEEE International Conference on Computer Vision, pages 5324-5333, 2017. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 107, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 152 + ], + "type": "text", + "content": "[64] C. Rockwell, David F. Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14084-14093, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 153, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 545, + 206 + ], + "type": "text", + "content": "[65] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 274 + ], + "type": "text", + "content": "[66] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 276, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 352 + ], + "type": "text", + "content": "[67] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, Seyedeh Sara Mahdavi, Raphael Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. ArXiv, abs/2205.11487, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "type": "text", + "content": "[68] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "text", + "content": "[69] Aditya Sanghi, Hang Chu, J. Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18582-18592, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 455, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 487 + ], + "type": "text", + "content": "[70] Flavio Miguel Schneider, Zhijing Jin, and Bernhard Schölkopf. Mousai: Text-to-music generation with long-context latent diffusion. ArXiv, abs/2301.11757, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 490, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 566 + ], + "type": "text", + "content": "[71] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 568, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 621 + ], + "type": "text", + "content": "[72] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. ArXiv, abs/2111.02114, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 624, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 666 + ], + "type": "text", + "content": "[73] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[74] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8025-8035, 2020. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7919" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[75] Uriel Singer, Adam Polyak, Thomas Hayes, Xiaoyue Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. ArXiv, abs/2209.14792, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 183 + ], + "type": "text", + "content": "[76] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, and Yaniv Taigman. Text-to-4d dynamic scene generation. ArXiv, abs/2301.11280, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "text", + "content": "[77] Josef Sivic, Biliana K. Kaneva, Antonio Torralba, Shai Avidan, and William T. Freeman. Infinite images: Creating and exploring a large photorealistic virtual space. Proceedings of the IEEE, 98:1391-1407, 2008. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "type": "text", + "content": "[78] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 282, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 282, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 282, + 287, + 316 + ], + "type": "text", + "content": "[79] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 349 + ], + "type": "text", + "content": "[80] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 382 + ], + "type": "text", + "content": "[81] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438-12448, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 436 + ], + "type": "text", + "content": "[82] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 470 + ], + "type": "text", + "content": "[83] Alexandru Telea. An image inpainting technique based on the fast marching method. Journal of graphics tools, 9(1):23-34, 2004. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 287, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 504 + ], + "type": "text", + "content": "[84] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 504, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 568 + ], + "type": "text", + "content": "[85] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "text", + "content": "[86] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3825–3834, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 667 + ], + "type": "text", + "content": "[87] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. arXiv preprint arXiv:2212.08070, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[88] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. arXiv preprint arXiv:2212.00774, 2022. 1, 2, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 352 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "type": "text", + "content": "[89] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 546, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 162 + ], + "type": "text", + "content": "[90] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7465-7475, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 546, + 217 + ], + "type": "text", + "content": "[91] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. ArXiv, abs/2212.11565, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 218, + 546, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 546, + 262 + ], + "type": "text", + "content": "[92] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with " + }, + { + "bbox": [ + 307, + 218, + 546, + 262 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 307, + 218, + 546, + 262 + ], + "type": "text", + "content": " views. arXiv e-prints, pages arXiv-2211, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 263, + 546, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 546, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 546, + 317 + ], + "type": "text", + "content": "[93] Jiale Xu, Jia Zheng, Yanyu Xu, Rui Tang, and Shenghua Gao. Layout-guided novel view synthesis from a single indoor panorama. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16438-16447, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 319, + 546, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 546, + 352 + ], + "type": "text", + "content": "[94] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7920" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/8023656a-1169-49dd-b8cd-4db39941377d_content_list.json b/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/8023656a-1169-49dd-b8cd-4db39941377d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..93cee5efe7bb7d19c8d93b9f58fabb95249b760e --- /dev/null +++ b/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/8023656a-1169-49dd-b8cd-4db39941377d_content_list.json @@ -0,0 +1,1439 @@ +[ + { + "type": "text", + "text": "Text2Tex: Text-driven Texture Synthesis via Diffusion Models", + "text_level": 1, + "bbox": [ + 173, + 130, + 795, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dave Zhenyu Chen", + "text_level": 1, + "bbox": [ + 86, + 180, + 250, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yawar Siddiqui", + "text_level": 1, + "bbox": [ + 271, + 180, + 401, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hsin-Ying Lee2", + "text_level": 1, + "bbox": [ + 424, + 180, + 550, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sergey Tulyakov2", + "text_level": 1, + "bbox": [ + 571, + 180, + 715, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthias Nießner1", + "text_level": 1, + "bbox": [ + 736, + 180, + 880, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Technical University of Munich", + "text_level": 1, + "bbox": [ + 272, + 198, + 532, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Snap Research", + "text_level": 1, + "bbox": [ + 571, + 199, + 700, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://daveredrum.github.io/Text2Tex/", + "text_level": 1, + "bbox": [ + 261, + 217, + 709, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/325025c2a43659fdd8b91347ff83c17846c00c8bf1bf4ffc6466f79b4a5e8104.jpg", + "image_caption": [ + "Figure 1: We introduce Text2Tex, a text-driven architecture for 3D texture synthesis. Given object geometries and text prompts as input, Text2Tex generates high quality and consistent textures via depth-aware inpainting and refinement." + ], + "image_footnote": [], + "bbox": [ + 86, + 265, + 888, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 525, + 313, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Text2Tex, a novel method for generating high-quality textures for 3D meshes from the given text prompts. Our method incorporates inpainting into a pre-trained depth-aware image diffusion model to progressively synthesize high resolution partial textures from multiple viewpoints. To avoid accumulating inconsistent and stretched artifacts across views, we dynamically segment the rendered view into a generation mask, which represents the generation status of each visible texel. This partitioned view representation guides the depth-aware inpainting model to generate and update partial textures for the corresponding regions. Furthermore, we propose an automatic view sequence generation scheme to determine the next best view for updating the partial texture. Extensive experiments demonstrate that our method significantly outperforms the existing text-driven approaches and GAN-based methods.", + "bbox": [ + 73, + 556, + 473, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 814, + 207, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating high-quality 3D content is an essential component of visual applications in films, games, and upcoming AR/VR scenarios. With an increasing number of 3D content datasets, the computer vision community has wit-", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nessed significant progress in the field of 3D geometry generation [12, 39, 61, 66, 38]. Despite the remarkable success in modeling 3D geometries in recent years, fully automatic 3D content generation is still hindered by the laborious human efforts required to design textures. Therefore, automating the texture design process through alternative guidance, such as text, has become an intriguing but challenging research problem.", + "bbox": [ + 496, + 526, + 892, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, text-to-image generators have shown remarkable progress in the 2D domain leveraging diffusion model architectures, enabling high resolution 2D content generation based on textual descriptions [1, 48]. However, there are notable challenges for producing 3D textures via such 2D vision-language prior knowledge. Specifically, the synthesized textures are expected to be not only with high fidelity to the language cues, but also of high and consistent quality for target meshes. As such, previous attempts to paint 3D geometry from text inputs often fail to deliver well-textured 3D content.", + "bbox": [ + 496, + 652, + 892, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce Text2Tex, a novel texture synthesis method that seamlessly texturizes 3D objects using a pre-trained depth-aware text-to-image diffusion model. The method renders a target mesh from multiple viewpoints and inpaints the missing appearance with a", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18558", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "depth-aware text-to-image diffusion model. Text2Tex follows a generate-then-refine strategy. Our method progressively generates partial textures across viewpoints and backprojects them to texture space. To address stretched and inconsistent artifacts observed from rotated viewpoints, we design a view partitioning technique that computes similarity maps between visible texel's normal vectors and the current view direction. The generation mask created from these similarity maps guides the diffusion process by indicating regions to generate, update, keep, or ignore. This allows us to apply different diffusion strengths to respective regions, inpainting missing appearance and updating stretched artifacts. However, the autoregressive generation process via the diffusion-based image inpainting model presents a new challenge. As the inpainting and updating scheme is conditioned on previously synthesized results, a viewpoint sequence with an ill-defined order or incomplete coverage over the mesh surface may result in unsatisfactory texturization. Therefore, we propose an automatic viewpoint selection technique that progressively selects the next best view. The confidence of each candidate view containing the biggest relative area for generation and updating is estimated, given the partially textured mesh. This approach ensures complete coverage over the mesh surface and a high-quality texture map by consistently updating stretched regions.", + "bbox": [ + 76, + 90, + 472, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of Text2Tex for synthesizing high-quality 3D textures from language cues. The proposed method performs favorably against other language-based texture synthesis methods in terms of FID [23], KID [4], and user study on a subset of the Objaverse dataset [17]. Additionally, our method also outperforms category-specific GAN-based methods on the ShapeNet car dataset [7].", + "bbox": [ + 76, + 484, + 472, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our technical contributions are threefold:", + "bbox": [ + 96, + 604, + 467, + 619 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design a novel method for high-quality texture synthesis by progressively inpainting and updating the 3D textures via depth-aware diffusion models.", + "- We propose an automatic view sequence generation scheme to dynamically determine the order for generating and updating the texture space.", + "- We conduct extensive study on a considerable amount of 3D objects, demonstrating the proposed method is effective for large-scale 3D content generation." + ], + "bbox": [ + 96, + 631, + 467, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 800, + 215, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Generation from 3D and 2D data. To achieve 3D generation, it is natural to train models directly on 3D data. In contrast to 2D images, there are several 3D representations available, each with its unique characteristics, leading to the development of various generative models such as", + "bbox": [ + 76, + 824, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "those based on voxels [30, 55, 59, 62], point clouds [3, 32], meshes [68], signed distance function [12, 13, 14, 16, 36], etc. However, unlike images or videos that are ubiquitous, 3D data is inherently scarce and challenging to collect and annotate. Consequently, the synthesized samples from 3D generative models, trained on 3D data, are of limited quality and diversity, in terms of both structure and texture. Recent works have leveraged differentiable rendering to learn texture generation using only 2D images [20, 56, 65]. However, they are typically trained for specific shape categories and struggle in the quality of textures.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-Guided Generation. Recently, there has been tremendous progress in the vision-language domain [26, 57, 60, 47, 10, 8, 11, 9]. Specifically, the emergence of Contrastive Language-Image Pre-Training (CLIP) [47] has enabled the development of text-guided image generation through its semantically rich representation trained on text-image pairs. Initial efforts [15, 45] incorporated CLIP with different backbones, such as StyleGAN [27] and VQGAN [19]. However, diffusion models [18, 24, 25, 40, 53], which have gained attention due to their superior visual quality and training stability compared to Generative Adversarial Networks (GANs) [21], have recently been trained on large-scale text-image datasets with CLIP encodings [28, 49, 52]. Among these models, Stable Diffusion [1, 52] has garnered significant interest as an open-sourced model with numerous extensions [1, 67] that support different conditional modalities in addition to text prompts, including depth images, poses, sketches, etc. Additionally, CLIP has also been adopted in 3D to perform text-guided shape and texture generation [34, 37]. In this work, we take advantage of the depth-conditioning feature of Stable Diffusion to provide more consistent texturing.", + "bbox": [ + 496, + 261, + 892, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-3D from 2D data. Inspired by the success of Neural Radiance Fields (NeRF) [35], NeRF-based generators [2, 5, 6, 22, 41, 43, 54, 58, 63] have been proposed to learn 3D structures from 2D images using GAN-based frameworks. A new research direction emerged by combining NeRF techniques with flourishing diffusion-based text-to-image models, enabling text-to-3D learning with only 2D supervision. To address the challenge of optimizing a NeRF field, a score-distillation loss is proposed [46] that leverages a pretrained 2D diffusion model as a critic to provide essential gradients. Subsequent efforts have focused on adopting this loss in latent space [13, 33] and in a coarse-to-fine refinement approach [29]. However, optimization-based methods are plagued by long convergence times. A recent concurrent work [51] proposes a non-optimization approach with progressive updates from multiple pre-set viewpoints. In contrast, our method iteratively updates and refines the synthesized textures from automatically selected viewpoint sequences, which minimizes human efforts with designing different viewpoint orders for various geometries.", + "bbox": [ + 496, + 598, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18559", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/af9726b05d572e3372214fe3a3f4d751977d8c7bca91eab3a69ad543c0bf55a9.jpg", + "image_caption": [ + "Figure 2: Overview of Text2Tex. We illustrate the pipeline using a 3D car mesh with a prompt \"golden Porsche\". We progressively generate the texture via a generate-then-refine scheme. In progressive texture generation (Sec. 3.3), we start by rendering the object from an initial preset viewpoint. We generate a new appearance according to the input prompt via a depth-to-image diffusion model, and project the generated image back to the partial texture. Then, we repeat this process until the last preset viewpoint to output the initial textured mesh. In the subsequent texture refinement (Sec. 3.4), we update the initial texture from a sequence of automatically selected viewpoints to refine the stretched and blurry artifacts." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 651, + 296 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fc07aaf42c347ed25813041dfa6a9bf54aab089619c2a563a6a4fe381eccc68e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 89, + 883, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 426, + 166, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The objective of our work is to texture a 3D mesh using a pretrained text-to-image diffusion model. In this section, we begin by laying the foundation of the diffusion model in Sec. 3.1 and depth-aware inpainting model in Sec. 3.2. We then propose a generate-then-refine scheme for progressively synthesizing and updating the 3D textures in a coarse-to-fine fashion. In the progressive texture generation (Sec. 3.3), we paint the visible regions of the input geometry in an incremental fashion, following a sequence of predefined viewpoints. To ensure local and global consistency, we incorporate a view partition to guide the depth-aware inpainting objectives. Subsequently, we introduce an automatic viewpoint selection mechanism (Sec. 3.4) to perform texture refinement and address any issues of texture stretching and inconsistency.", + "bbox": [ + 75, + 452, + 473, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 691, + 205, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use a Denoising Diffusion Probabilistic Model (DDPM) [24] as the generative model. Specifically, to avoid high computational overhead, we adopt the latent diffusion model [52], where an input image $x_0$ is first encoded into latent code $z_0$ before the diffusion process. The forward pass follows a Markov Chain to gradually add noise to the input latent code $z_0$ towards the white Gaussian noise $\\mathcal{N}(0,1)$ . At each step in the forward pass, the noised latent code $z_t$ is obtained by adding a noise variance $\\beta_t$ to the previous latent code $z_{t-1}$ scaled with $\\sqrt{1 - \\beta_t}$ :", + "bbox": [ + 75, + 715, + 468, + 867 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} \\sim \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 883, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The independence property enables direct transformation of the noised latent code $z_{t}$ at an arbitrary time step $t$ from the input latent $z_{0}$ via:", + "bbox": [ + 498, + 428, + 892, + 472 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} \\sim \\mathcal {N} (\\sqrt {\\bar {a}} _ {t} z _ {0}, (1 - \\bar {a} _ {t}) \\mathbf {I}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 484, + 890, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{a}_t$ is the total noise variance, which can be calculated by $\\sum_{t = i}^{T}(1 - \\beta_{t})$ from the noise $\\beta_{t}$ added to the input latent code $z_0$ at each time step.", + "bbox": [ + 498, + 513, + 890, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During inference, the latent estimation $\\hat{z}_{t - 1}$ for the next time step $t - 1$ is obtained by predicting $\\mu_{\\theta}(z_t,t)$ and $\\sigma_{\\theta}(z_t,t)$ of a Gaussian distribution:", + "bbox": [ + 498, + 560, + 892, + 604 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {t - 1} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(z _ {t}, t\\right), \\sigma_ {\\theta} \\left(z _ {t}, t\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 619, + 890, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising strength. To prevent complete randomness during the diffusion process, we introduce a scaling factor $\\gamma, 0 < \\gamma \\leq 1$ , which controls the number of diffusion steps. We assume that a white Gaussian noise $\\mathcal{N}(0,1)$ can be obtained by adding noise to the input latent code $z_0$ through $T$ steps, and the final denoised latent estimation $\\hat{z}_0$ is fully governed by the pure noise. By applying the scaling factor, we can start denoising the latent code at time step $\\gamma T$ to guide the final latent code with the original image information. This technique is applied to refine the previously generated image contents.", + "bbox": [ + 498, + 655, + 890, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Depth-Aware Image Inpainting", + "text_level": 1, + "bbox": [ + 500, + 830, + 774, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core of the texture synthesis lies in painting the missing regions on the mesh surface. The generated texture is expected to be highly faithful to the mesh geometry and the", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18560", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a3f7c5976aa72fad20aaf2fed00b8ed53e16d6d77b6963f2210e064ae2fd7267.jpg", + "image_caption": [ + "Figure 3: We dynamically partition the current view into a generation mask to guide the depth-aware inpainting model. For the \"new\" region, we denoise the new object appearance from white Gaussian noise. For the \"update\" region, we refine the previous texture by denoising the partially noised image segments. We freeze the texture in the \"keep\" region for this view." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 472, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "input text. To achieve this, we build our method on a pretrained depth-to-image model [1, 67] that can produce high-quality images from text while being consistent with depth cues. However, since the Depth2Image model is designed to generate entire images, we need to use an inpainting mask to guide the sampling process. This mask provides explicit hints of which regions to generate or keep fixed, similar to the denoising guidance strategy in RePaint [31].", + "bbox": [ + 75, + 534, + 468, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To condition the denoising on the known regions of the input, we inject a generation mask $\\mathcal{M}$ into the sampling steps. This mask explicitly blends the noised latent code $z_{t}$ and the denoised latent estimation $\\hat{z}_t$ as follows:", + "bbox": [ + 75, + 655, + 470, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {t} = \\hat {z} _ {t} \\odot \\mathcal {M} + z _ {t} \\odot (1 - \\mathcal {M}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 734, + 468, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then decode the final denoised latent estimation $\\hat{z}_0$ to the final output image $\\mathcal{O}$ .", + "bbox": [ + 76, + 758, + 470, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Progressive Texture Generation", + "text_level": 1, + "bbox": [ + 76, + 801, + 357, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the customized Depth2Image model, we are able to paint the object with a high quality image from a single view. To synthesize the appearance of the input geometry, we project the generated 2D views onto the texture space of a normalized 3D object with proper UV parameterization.", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assuming Y-axis as the up-axis in the world coordinate system, we define the viewpoint as $v = (\\theta, \\phi, r)$ , where $\\theta$ is the azimuth angle with respect to the Z-axis, $\\phi$ is the viewpoint elevation angle with respect to the XZ-plane, and $r$ is the distance between the viewpoint and the origin.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, we start by generating the visible but missing texture in an initial viewpoint $v^0$ . We render the object to a depth map $\\mathcal{D}^0$ and a generation mask $\\mathcal{M}^0$ , and then use a customized Depth2Image diffusion model with $\\mathcal{D}^0$ as input and $\\mathcal{M}^0$ as extra guidance to generate a colored image $\\mathcal{O}^0$ . We then back-project the image $\\mathcal{O}^0$ to the visible part of the texture $\\mathcal{T}^0$ . In the subsequent steps, we progressively diffuse the colored images $\\mathcal{O}^k$ and back-project them to the texture $\\mathcal{T}^k$ through a sequence of viewpoints.", + "bbox": [ + 496, + 167, + 892, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We notice that directly inpainting the missing regions on mesh surface often results in inconsistency issues. The issue is mainly caused by the stretched artifacts that occur when the 2D views are projected back onto the curved surface of the mesh. Therefore, we design a dynamic view partitioning strategy to guide the inpainting process with respective generation objectives $\\mathcal{M}$ and denoising strengths $\\gamma$ .", + "bbox": [ + 496, + 303, + 893, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic view partitioning For all viewpoints $\\mathcal{V} = v_{i}, i = 1, \\dots, N$ , we render the similarity mask $S^i$ for each viewpoint $v^k$ and map those values to the texture space. Each pixel in a similarity mask represents the reversed normalized value of the cosine similarity between the normal vectors of the visible faces and the view direction (ranging from 0 to 1). A pixel with a value of 1 indicates that the corresponding face is perpendicular to the view direction. For simplicity, we set the background to 0. In summary, these masks indicate the extent to which a face is rotated away from the viewpoint.", + "bbox": [ + 496, + 431, + 892, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the similarity mask $S^k$ at step $k$ , we segment the rendered view into a generation mask $\\mathcal{M}^k$ , including the following 4 regions, as shown in Fig. 3: 1) New: This region contains pixels that have not yet been textured. We inpaint this region from pure white Gaussian noise, i.e. with denoising strength 1. 2) Update: This region contains pixels that have been textured, but the corresponding similarity score in $S^k$ is greater than all other views. This indicates that those pixels are being observed in a better angle. Therefore, we update this region with a moderate denoising strength $\\gamma_g$ to avoid stretched appearance. 3) Keep: Pixels in this region have been textured, but the corresponding similarity score in $S^k$ is not the highest among all other views. These pixels have already been observed from a better angle, so we keep them fixed. 4) Ignore: This region contains pixels that belong to the background and are irrelevant to the process, so we ignore them throughout the entire process.", + "bbox": [ + 496, + 598, + 893, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the generation mask helps guide the texture inpainting process with accurate generation objectives and ap", + "bbox": [ + 496, + 869, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18561", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/de0fb4cfc55b75d280ee7caa5fb9247df42a3bd65b283bf5809b5cc1503bf40c.jpg", + "image_caption": [ + "Figure 4: In the refinement stage, the sequence of viewpoints are automatically determined by selecting the viewpoint with the maximum normalized area of the \"update\" region at each step. We update 2D views in the \"update\" regions with a moderate diffusion denoising strength. The updated object appearance is then back-projected to the texture space at the end of each refinement step." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "propriate denoising strengths, blurriness and stretches can still exist on the mesh surface. This is because the generation mask is limited to a predefined set of viewpoints, and the seams and stretches on the texture are still visible from a novel viewpoint. To address this issue, we propose a texture refinement technique with an automatic viewpoint selection strategy, which is described in the next section.", + "bbox": [ + 75, + 515, + 468, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Texture Refinement with Automatic Viewpoint Selection", + "text_level": 1, + "bbox": [ + 76, + 633, + 468, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To remove the synthesis artifacts, a straightforward solution is to increase the number of viewpoints. However, the optimal viewpoint sequence can vary for different object geometries, making it difficult to manually pre-set the viewpoint sequence for massive synthesis targets. To address these challenges, we propose an automatic viewpoint selection strategy that effectively prevents stretches and seams, as illustrated in Fig. 4. We densely define a set of refinement viewpoints $\\mathcal{V} = v_{i}, i = 1,\\dots,K$ , where $K$ is larger than $N$ . To distribute the refinement viewpoints evenly, we scatter them on a hemisphere, taking into account that objects are rarely observed from the bottom-up view.", + "bbox": [ + 75, + 672, + 470, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assuming that an initial texture has been applied to the object, the refinement process begins by segmenting the generation masks $\\mathcal{M}$ using the similarity masks $S$ from all", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "available viewpoints. For each of the $K$ refinement viewpoints in $\\mathcal{V}$ , we calculate a view heat $h^i$ from the corresponding generation mask $\\mathcal{M}_i$ , which represents the normalized area of the \"update\" region with respect to the current visible area of the object. The viewpoint $v_i$ that maximizes the view heat is then selected by $\\arg \\max_i h_i = \\frac{1}{N_p} \\sum_{k=1}^{N_p} w_k$ where $N_p$ is the total number of the non-background pixels, and $w_k$ is the scaling factor for the segments in the generation mask. In order to let views with relatively more areas for updating, $w_k$ for the \"update\" region is set to be bigger than that for the \"keep\" region. We dynamically select the next view with the highest view heat for updating. To avoid conflicts with the previously generated textures, we update the \"update\" regions with a mild denoising strength $\\gamma_r$ , which preserves the original appearance cues.", + "bbox": [ + 496, + 90, + 893, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 500, + 349, + 586, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 375, + 717, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply the Depth2Image model from Stable Diffusion v2 [1] as our generation backbone. The denoising strength $\\gamma_{g}$ and $\\gamma_{r}$ are set as 0.5 and 0.3 for the generation and refinement stages, respectively. We define 6 axis-aligned principles viewpoints for generation, and in total 36 viewpoints for refinement, among which we dynamically select only 20 views to reduce time cost. Each synthesis process takes around 15 minutes to complete on an NVIDIA RTX A6000. Our implementation uses the PyTorch [44] framework with PyTorch3D [50] used for rendering and texture projection.", + "bbox": [ + 496, + 396, + 893, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 556, + 676, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data. We evaluate our method on a subset of textured meshes from the Objaverse [17] dataset. We first sample 3 random meshes from each category. To ensure the quality of the input meshes, we manually filter out thin or unrecognizable meshes, such as \"strainer\", \"sweatpants\", and \"legging\", meshes with too simplistic textures, and meshes that do not correspond with their assigned categories. For the purpose of reducing processing time, we also remove overtriangulated and scanned objects. After this curation, there are in total 410 high quality textured meshes across 225 categories for the experiments. Note that the original textures are only used for the evaluation. To compare with GAN-based category-specific approaches, we also report results on the \"car\" objects from the ShapeNet dataset [7]. In particular, we use the 300 meshes from the test set used in [56].", + "bbox": [ + 496, + 580, + 893, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our method against the following state-of-the-art text-driven texture synthesis method: 1) Text2Mesh [34], a neural pipeline that directly optimizes the textures and geometries via a CLIP-based optimization objective. We remove the displacement prediction so that", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18562", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9eb23aece4df7bb0df9d72d4d6ae6a5bd1bac0f9ce28267c2238b96121a089ae.jpg", + "image_caption": [ + "Figure 5: Qualitative comparisons on Objaverse. We compare our textured mesh against CLIPMesh [37], Text2Mesh [34], Latent-Paint [33], and the original textures from Objaverse. In comparison with the baselines, our method produces more consistent and detailed 3D textures with respect to the input geometries. Image best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 890, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "only the surface RGBs are optimized. 2) CLIPMesh [37], a CLIP-based pipeline that deforms a sphere and optimizes the surface RGBs. Similar to Text2Mesh, we remove the shape deformation branch, and the texture colors are directly optimized on the surface of a given geometry. 3) Latent-Paint [33], a texture generation variant of the NeRF-based 3D object generation pipeline Latent-NeRF [33]. It explicitly operates on a given texture map using Stable Diffusion as a prior. In addition to text-guided methods, we also compare with category-specific GAN-based approaches, including Texture Fields [42], SPSG [16], LTG [65], and Texturify [56].", + "bbox": [ + 75, + 547, + 472, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate the generated textures via commonly used image quality and diversity metrics for generative models. Specifically, we report the Frechet Inception Distance (FID) [23] and Kernel Inception Distance $(\\mathrm{KID} \\times 10^{-3})$ [4]. The generated image distribution for these metrics consists of renders of each mesh with the synthesized textures from 20 fixed viewpoints at a resolution of $512 \\times 512$ . For experiments on Objaverse dataset, the real distribution comprises renders of the meshes with the same settings using their artist designed textures. For ex", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "periments on ShapeNet cars, we use the 18991 background segmented images from CompCars dataset [64].", + "bbox": [ + 498, + 547, + 892, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Quantitative results", + "text_level": 1, + "bbox": [ + 500, + 601, + 689, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, we compare our method against the previous SOTA text-driven texture synthesis methods on Objverse objects. As input, we uniformly feed template texts \"a (category)\" to the models. Quantitatively, our method outperforms all baselines by a significant margin (19% improvement in FID and 26% improvement in KID). Such improvements demonstrate that our method is more capable of generating more realistic textures on various object geometries from numerous categories. To demonstrate the effectiveness of our method against the GAN-based approaches on category-specific data, we report experiment results on ShapeNet \"car\" category in Tab. 2. Notably, our method achieves superior performance over the previous GAN-based SOTA texture synthesis method Texturify, improving by 21% in FID and 12% in KID. This indicates our method is more effective with synthesizing realistic textures than GAN based approaches that were trained on specific categories.", + "bbox": [ + 496, + 628, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18563", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ecad2569c5931d313c99782523227924b57d97a57fb94380dce4e7b9f4648dd1.jpg", + "image_caption": [ + "Figure 6: Qualitative comparisons on ShapeNet car. Our method generates sharper and more coherent textures with respect to the geometries compared to the state-of-the-art GAN-based method." + ], + "image_footnote": [], + "bbox": [ + 80, + 85, + 468, + 239 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/534063c2b066a3510ab32f6b7fcc81b2cb01f772daeaa4379bcd950cb389b99a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID ↓KID (x10-3) ↓
Text2Mesh [34]45.38(+9.7)10.40(+2.7)
CLIPMesh [37]43.25(+7.6)12.52(+4.8)
Latent-Paint [33]43.87(+8.1)11.43(+3.7)
Text2Tex (Ours)35.687.74
", + "bbox": [ + 99, + 324, + 446, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User study. We conduct user study to analyze the quality of the synthesized textures and their fidelity to the input text prompts. For each baseline method, we randomly show the users 5 pair of renders from the baseline and our method. The users are requested to choose the one that is more realistic and closer to the text prompts. More details of the questionnaire can be found in the supplemental. In the end, we receive 604 responses across 41 users. The collected preferences are reported in Tab. 3. In comparison to CLIPMesh and Text2Mesh, our method is clearly preferred by the users with preference rate $83.92\\%$ and $76.47\\%$ , respectively. Besides that, more users $(64.18\\%)$ lean towards our method over the competitive baseline Latent-Paint. As can be seen, our method demonstrates the effectiveness in generating high quality textures that are favored by human users.", + "bbox": [ + 75, + 505, + 468, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative analysis", + "text_level": 1, + "bbox": [ + 76, + 756, + 264, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our qualitative results on Objaverse objects against text-driven baselines in Fig. 5. In comparison with CLIP-based methods CLIPMesh and Text2Mesh, our method generates more realistic and consistent textures. In particular, CLIPMesh generates sketchy textures while Text2Mesh produces repetitive patterns. Latent-Paint outputs a consistent texture capturing the semantics of the object well, but the results are often quite blurry. Clearly,", + "bbox": [ + 75, + 780, + 467, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/faa529ab6afd557413d8781bbe996252ed90885d96e4fb2e84e5d9326f1689bf.jpg", + "table_caption": [ + "Table 1: Quantitative comparisons on Objverse subset. Our method performs favorably against state-of-the-art text-driven texture synthesis methods." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓KID (x10-3) ↓
Texture Fields [42]177.15(+130.2)17.14(+12.8)
SPSG [16]110.65(+63.7)9.59(+5.2)
LTG [65]70.76(+23.8)5.72(+1.4)
Texturify [56]59.55(+12.6)4.97(+0.6)
Text2Tex (Ours)46.914.35
", + "bbox": [ + 506, + 88, + 885, + 202 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c0bd159419ae4b1f00ba3939405c89576f294ec6f909be03ff15b9976aee5644.jpg", + "table_caption": [ + "Table 2: Quantitative comparison on the ShapeNet cars. Our method outperform state-of-the-art category-specific GAN-based methods by a significant margin." + ], + "table_footnote": [], + "table_body": "
CLIPMesh ↑Text2Mesh ↑Latent-Paint ↑
Ours83.92%76.47%64.18%
", + "bbox": [ + 506, + 275, + 885, + 323 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b4c9585749651ad51cacee25b38c43e2fe9d6624bf0fff72d56a8aea3e5fc398.jpg", + "table_caption": [ + "Table 3: Percentage of users who prefer our method over the baselines in a user study with 604 responses across 41 participants. Our method is shown to be more favored by human users." + ], + "table_footnote": [], + "table_body": "
w/ Depth2Imgw/ inpaintingw/ updateFID ↓KID (x10-3) ↓
✓XX39.889.78
✓✓X38.199.11
✓✓✓37.098.78
", + "bbox": [ + 503, + 410, + 888, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4: Effect of components in the generation stage. We quantitatively ablate the efficacy of each component. Applying the inpainting and update scheme effectively improves the quality of the synthesized textures.", + "bbox": [ + 498, + 483, + 890, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our method can synthesize more consistent textures with cleaner and richer local details. We also compare our textures with GAN-based generation approach on category-specific objectives. In Fig. 6, we show the textures for ShapeNet cars of our methods and Texturify. Notably, our textures have a much cleaner appearance and provide more details with respect to the input geometries.", + "bbox": [ + 496, + 574, + 890, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation studies", + "text_level": 1, + "bbox": [ + 500, + 694, + 660, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Does depth-aware inpainting and updating help? We show in Fig. 7 that the depth-aware inpainting is essential for producing high quality textures. In particular, the plain Depth2Img model often struggles to produce consistent appearances due to the governance of random noise. When the inpainting scheme is applied, the produced textures are more consistent. However, the textures still appear to be stretched and blurry over the curved mesh surface. These artifacts are amended by the texture updating scheme from the better viewing angles. The effectiveness of the depth-aware inpainting and the updating scheme is reflected in the improved FID and KID scores in Tab. 4.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18564", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/707f8f03a26e274ccb1740e84ce734a092493b02293f3562d38b86f2deb955ba.jpg", + "image_caption": [ + "Figure 7: The proposed inpainting and update technique can effectively generate more consistent textures and eliminate stretched and blurry artifacts. Note that there is no refinement in this ablation." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 472, + 414 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ad3c7772e108a566e034def3cda0ccae91999c0e825cfd74c9b0706449469ea7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
# views05101520
↓FID37.0936.6736.3935.9835.68
↓KID (x10-3)8.788.318.127.987.74
", + "bbox": [ + 81, + 500, + 465, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5: We quantitatively study the effect of selecting different number of viewpoints in the refinement stage. Refining the synthesis results with more viewpoints improves the texture quality.", + "bbox": [ + 75, + 569, + 468, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Does viewpoint selection in refinement stage help? We compare our results with different refinement settings in Tab. 5. When the input geometries are painted with initial textures from the generation stage, the blurry artifacts and projection seams are usually not eliminated due to a limited number of viewpoints. As can be seen in Fig. 8, such flaws can be minimized by refining with more viewpoints. We also showcase the effectiveness of the automatic viewpoint selection technique, as the refinement process does not require any manual efforts with defining and fine-tuning the viewpoint sequence for different shapes.", + "bbox": [ + 75, + 641, + 470, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Limitations.", + "text_level": 1, + "bbox": [ + 76, + 816, + 207, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our method has shown the capability to produce high-quality 3D textures, we have observed that it tends to produce textures with shading effects from the diffusion backbone. Although this issue can be addressed by care", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9d69f53354c4c876204d8c3cf5fd072b538fabc551fb18558616d3ddc27129f6.jpg", + "image_caption": [ + "Figure 8: The proposed automatic viewpoint selection method further improves the texture quality by gradually removing the remaining artifacts from the generation stage." + ], + "image_footnote": [], + "bbox": [ + 506, + 87, + 890, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fully fine-tuning the input prompts, doing so requires additional human engineering effort and may not scale well to massive generation targets. One potential solution is to fine-tune the diffusion model to remove the shading from textures. We acknowledge this challenge and leave it to future work to explore this possibility.", + "bbox": [ + 498, + 491, + 890, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 601, + 617, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a novel method, Text2Tex, for synthesizing high quality textures for 3D meshes from the given text prompts. Our approach leverages a depth-aware image inpainting diffusion model to progressively generate high-resolution partial textures from multiple viewpoints. To avoid accumulating inconsistent and stretching artifacts across viewpoints, we dynamically segment the rendered view into a generation mask, which effectively guides the diffusion model to generate and update the corresponding partial textures. Furthermore, we propose an automatic viewpoint sequence generation scheme that utilizes the generation mask to automatically determine the next best view for refining the generated textures. Extensive experiments demonstrate that our method can effectively synthesize consistent and highly detailed 3D textures for various object geometries without extra manual effort. Overall, we hope our work can inspire more future research in the area of text-to-3D synthesis.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18565", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 78, + 90, + 243, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the ERC Starting Grant Scan2CAD (804724) and the German Research Foundation (DFG) Research Unit \"Learning and Simulation in Visual Computing\". We further thank Guy Gafni and Manuel Dahnert for the helpful discussions, and Angela Dai for the video voiceover.", + "bbox": [ + 76, + 116, + 468, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 224, + 173, + 239 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stable diffusion. https://github.com/Stability-AI/stablediffusion, 2022. 1, 2, 4, 5", + "[2] Rameen Abdal, Hsin-Ying Lee, Peihao Zhu, Menglei Chai, Aliaksandr Siarohin, Peter Wonka, and Sergey Tulyakov. 3davatargan: Bridging domains for personalized editable avatars. In CVPR, 2023. 2", + "[3] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 2", + "[4] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In ICLR, 2018. 2, 6", + "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022: 2", + "[6] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2", + "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 5", + "[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. ScanRefer: 3D object localization in RGB-D scans using natural language. In European Conference on Computer Vision, pages 202-221. Springer, 2020. 2", + "[9] Dave Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. arXiv preprint arXiv:2212.00836, 2022. 2", + "[10] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 487-505. Springer, 2022. 2", + "[11] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3193-3203, 2021. 2", + "[12] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 1, 2" + ], + "bbox": [ + 78, + 250, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, 2023. 2", + "[14] Zezhou Cheng, Mengei Chai, Jian Ren, Hsin-Ying Lee, Kyle Olszewski, Zeng Huang, Subhransu Maji, and Sergey Tulyakov. Cross-modal 3d shape generation and manipulation. In ECCV, 2022. 2", + "[15] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In ECCV, 2022. 2", + "[16] Angela Dai, Yawar Siddiqui, Justus Thies, Julien Valentin, and Matthias Nießner. Spsg: Self-supervised photometric scene generation from rgb-d scans. In CVPR, 2021. 2, 6, 7", + "[17] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 2, 5", + "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2", + "[19] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2", + "[20] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 2", + "[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NIPS, 2014. 2", + "[22] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2", + "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2, 6", + "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2, 3", + "[25] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. arXiv preprint arXiv:2106.15282, 2021. 2", + "[26] Ronghang Hu and Amanpreet Singh. Unit: Multimodal multitask learning with a unified transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1439-1449, 2021. 2", + "[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 2", + "[28] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2", + "[29] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fi" + ], + "bbox": [ + 501, + 92, + 890, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18566", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 2", + "[30] Chieh Hubert Lin, Hsin-Ying Lee, Willi Menapace, Mengei Chai, Aliaksandr Siarohin, Ming-Hsuan Yang, and Sergey Tulyakov. *Infinicity: Infinite-scale city synthesis.* arXiv preprint arXiv:2301.09637, 2023. 2", + "[31] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 4", + "[32] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, 2021. 2", + "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, 2023. 2, 6, 7", + "[34] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In CVPR, 2022. 2, 5, 6, 7", + "[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2", + "[36] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. AutoSDF: Shape priors for 3d completion, reconstruction and generation. In CVPR, 2022. 2", + "[37] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia, 2022. 2, 6, 7", + "[38] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. arXiv preprint arXiv:2212.01206, 2022. 1", + "[39] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International conference on machine learning, pages 7220-7229. PMLR, 2020. 1", + "[40] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2", + "[41] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2", + "[42] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In ICCV, 2019. 6, 7", + "[43] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In CVPR, 2022. 2", + "[44] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 5", + "[45] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 2", + "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden" + ], + "bbox": [ + 78, + 90, + 468, + 887 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "hall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICLR, 2021. 2", + "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2", + "[50] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 5", + "[51] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2", + "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3", + "[53] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. 2022. 2", + "[54] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2", + "[55] Aliaksandr Siarohin, Willi Menapace, Ivan Skorokhodov, Kyle Olszewski, Jian Ren, Hsin-Ying Lee, Mengei Chai, and Sergey Tulyakov. Unsupervised volumetric animation. In CVPR, 2023. 2", + "[56] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In ECCV, 2022. 2, 5, 6, 7", + "[57] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2", + "[58] Ivan Skorokhodov, Aliaksandr Siarohin, Yinghao Xu, Jian Ren, Hsin-Ying Lee, Peter Wonka, and Sergey Tulyakov. 3d generation on imagenet. In ICLR, 2023. 2", + "[59] Edward J Smith and David Meger. Improved adversarial systems for 3d object generation and reconstruction. In CoRL, 2017. 2", + "[60] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 2", + "[61] Jiajun Wu, Chengkai Zhang, Xiuming Zhang, Zhoutong Zhang, William T Freeman, and Joshua B Tenenbaum. Learning shape priors for single-view 3d completion and reconstruction. In Proceedings of the European Conference on" + ], + "bbox": [ + 501, + 92, + 890, + 890 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18567", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision (ECCV), pages 646-662, 2018. 1", + "[62] Jianwen Xie, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Song-Chun Zhu, and Ying Nian Wu. Learning descriptor networks for 3d shape synthesis and analysis. In CVPR, 2018. 2", + "[63] Yinghao Xu, Mengei Chai, Zifan Shi, Sida Peng, Ivan Skorokhodov, Aliaksandr Siarohin, Ceyuan Yang, Yujun Shen, Hsin-Ying Lee, Bolei Zhou, et al. Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In CVPR, 2023. 2", + "[64] Linjie Yang, Ping Luo, Chen Change Loy, and Xiaou Tang. A large-scale car dataset for fine-grained categorization and verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3973-3981, 2015. 6", + "[65] Rui Yu, Yue Dong, Pieter Peers, and Xin Tong. Learning texture generators for 3d shape collections from internet photo sets. In BMVC, 2021. 2, 6, 7", + "[66] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 1", + "[67] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2, 4", + "[68] Song-Hai Zhang, Yuan-Chen Guo, and Qing-Wen Gu. Sketch2model: View-aware 3d modeling from single free-hand sketches. In CVPR, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 470 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18568", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/8023656a-1169-49dd-b8cd-4db39941377d_model.json b/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/8023656a-1169-49dd-b8cd-4db39941377d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..44a90c42ee434bac759afda33a6ec24a05804105 --- /dev/null +++ b/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/8023656a-1169-49dd-b8cd-4db39941377d_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.174, + 0.131, + 0.797, + 0.152 + ], + "angle": 0, + "content": "Text2Tex: Text-driven Texture Synthesis via Diffusion Models" + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.181, + 0.25, + 0.199 + ], + "angle": 0, + "content": "Dave Zhenyu Chen" + }, + { + "type": "title", + "bbox": [ + 0.272, + 0.181, + 0.402, + 0.199 + ], + "angle": 0, + "content": "Yawar Siddiqui" + }, + { + "type": "title", + "bbox": [ + 0.425, + 0.181, + 0.551, + 0.199 + ], + "angle": 0, + "content": "Hsin-Ying Lee2" + }, + { + "type": "title", + "bbox": [ + 0.572, + 0.181, + 0.717, + 0.199 + ], + "angle": 0, + "content": "Sergey Tulyakov2" + }, + { + "type": "title", + "bbox": [ + 0.737, + 0.181, + 0.882, + 0.199 + ], + "angle": 0, + "content": "Matthias Nießner1" + }, + { + "type": "title", + "bbox": [ + 0.274, + 0.199, + 0.533, + 0.217 + ], + "angle": 0, + "content": "1Technical University of Munich" + }, + { + "type": "title", + "bbox": [ + 0.573, + 0.2, + 0.701, + 0.217 + ], + "angle": 0, + "content": "\\(^{2}\\)Snap Research" + }, + { + "type": "title", + "bbox": [ + 0.262, + 0.218, + 0.71, + 0.234 + ], + "angle": 0, + "content": "https://daveredrum.github.io/Text2Tex/" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.266, + 0.889, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.482, + 0.893, + 0.514 + ], + "angle": 0, + "content": "Figure 1: We introduce Text2Tex, a text-driven architecture for 3D texture synthesis. Given object geometries and text prompts as input, Text2Tex generates high quality and consistent textures via depth-aware inpainting and refinement." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.526, + 0.314, + 0.541 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.557, + 0.474, + 0.801 + ], + "angle": 0, + "content": "We present Text2Tex, a novel method for generating high-quality textures for 3D meshes from the given text prompts. Our method incorporates inpainting into a pre-trained depth-aware image diffusion model to progressively synthesize high resolution partial textures from multiple viewpoints. To avoid accumulating inconsistent and stretched artifacts across views, we dynamically segment the rendered view into a generation mask, which represents the generation status of each visible texel. This partitioned view representation guides the depth-aware inpainting model to generate and update partial textures for the corresponding regions. Furthermore, we propose an automatic view sequence generation scheme to determine the next best view for updating the partial texture. Extensive experiments demonstrate that our method significantly outperforms the existing text-driven approaches and GAN-based methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.815, + 0.208, + 0.831 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Generating high-quality 3D content is an essential component of visual applications in films, games, and upcoming AR/VR scenarios. With an increasing number of 3D content datasets, the computer vision community has wit-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.893, + 0.648 + ], + "angle": 0, + "content": "nessed significant progress in the field of 3D geometry generation [12, 39, 61, 66, 38]. Despite the remarkable success in modeling 3D geometries in recent years, fully automatic 3D content generation is still hindered by the laborious human efforts required to design textures. Therefore, automating the texture design process through alternative guidance, such as text, has become an intriguing but challenging research problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.893, + 0.82 + ], + "angle": 0, + "content": "Recently, text-to-image generators have shown remarkable progress in the 2D domain leveraging diffusion model architectures, enabling high resolution 2D content generation based on textual descriptions [1, 48]. However, there are notable challenges for producing 3D textures via such 2D vision-language prior knowledge. Specifically, the synthesized textures are expected to be not only with high fidelity to the language cues, but also of high and consistent quality for target meshes. As such, previous attempts to paint 3D geometry from text inputs often fail to deliver well-textured 3D content." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce Text2Tex, a novel texture synthesis method that seamlessly texturizes 3D objects using a pre-trained depth-aware text-to-image diffusion model. The method renders a target mesh from multiple viewpoints and inpaints the missing appearance with a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18558" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.483 + ], + "angle": 0, + "content": "depth-aware text-to-image diffusion model. Text2Tex follows a generate-then-refine strategy. Our method progressively generates partial textures across viewpoints and backprojects them to texture space. To address stretched and inconsistent artifacts observed from rotated viewpoints, we design a view partitioning technique that computes similarity maps between visible texel's normal vectors and the current view direction. The generation mask created from these similarity maps guides the diffusion process by indicating regions to generate, update, keep, or ignore. This allows us to apply different diffusion strengths to respective regions, inpainting missing appearance and updating stretched artifacts. However, the autoregressive generation process via the diffusion-based image inpainting model presents a new challenge. As the inpainting and updating scheme is conditioned on previously synthesized results, a viewpoint sequence with an ill-defined order or incomplete coverage over the mesh surface may result in unsatisfactory texturization. Therefore, we propose an automatic viewpoint selection technique that progressively selects the next best view. The confidence of each candidate view containing the biggest relative area for generation and updating is estimated, given the partially textured mesh. This approach ensures complete coverage over the mesh surface and a high-quality texture map by consistently updating stretched regions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.485, + 0.473, + 0.606 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of Text2Tex for synthesizing high-quality 3D textures from language cues. The proposed method performs favorably against other language-based texture synthesis methods in terms of FID [23], KID [4], and user study on a subset of the Objaverse dataset [17]. Additionally, our method also outperforms category-specific GAN-based methods on the ShapeNet car dataset [7]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.606, + 0.468, + 0.62 + ], + "angle": 0, + "content": "To summarize, our technical contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.632, + 0.468, + 0.677 + ], + "angle": 0, + "content": "- We design a novel method for high-quality texture synthesis by progressively inpainting and updating the 3D textures via depth-aware diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.687, + 0.468, + 0.732 + ], + "angle": 0, + "content": "- We propose an automatic view sequence generation scheme to dynamically determine the order for generating and updating the texture space." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.742, + 0.468, + 0.787 + ], + "angle": 0, + "content": "- We conduct extensive study on a considerable amount of 3D objects, demonstrating the proposed method is effective for large-scale 3D content generation." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.632, + 0.468, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.216, + 0.817 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "3D Generation from 3D and 2D data. To achieve 3D generation, it is natural to train models directly on 3D data. In contrast to 2D images, there are several 3D representations available, each with its unique characteristics, leading to the development of various generative models such as" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "those based on voxels [30, 55, 59, 62], point clouds [3, 32], meshes [68], signed distance function [12, 13, 14, 16, 36], etc. However, unlike images or videos that are ubiquitous, 3D data is inherently scarce and challenging to collect and annotate. Consequently, the synthesized samples from 3D generative models, trained on 3D data, are of limited quality and diversity, in terms of both structure and texture. Recent works have leveraged differentiable rendering to learn texture generation using only 2D images [20, 56, 65]. However, they are typically trained for specific shape categories and struggle in the quality of textures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.893, + 0.595 + ], + "angle": 0, + "content": "Text-Guided Generation. Recently, there has been tremendous progress in the vision-language domain [26, 57, 60, 47, 10, 8, 11, 9]. Specifically, the emergence of Contrastive Language-Image Pre-Training (CLIP) [47] has enabled the development of text-guided image generation through its semantically rich representation trained on text-image pairs. Initial efforts [15, 45] incorporated CLIP with different backbones, such as StyleGAN [27] and VQGAN [19]. However, diffusion models [18, 24, 25, 40, 53], which have gained attention due to their superior visual quality and training stability compared to Generative Adversarial Networks (GANs) [21], have recently been trained on large-scale text-image datasets with CLIP encodings [28, 49, 52]. Among these models, Stable Diffusion [1, 52] has garnered significant interest as an open-sourced model with numerous extensions [1, 67] that support different conditional modalities in addition to text prompts, including depth images, poses, sketches, etc. Additionally, CLIP has also been adopted in 3D to perform text-guided shape and texture generation [34, 37]. In this work, we take advantage of the depth-conditioning feature of Stable Diffusion to provide more consistent texturing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Text-to-3D from 2D data. Inspired by the success of Neural Radiance Fields (NeRF) [35], NeRF-based generators [2, 5, 6, 22, 41, 43, 54, 58, 63] have been proposed to learn 3D structures from 2D images using GAN-based frameworks. A new research direction emerged by combining NeRF techniques with flourishing diffusion-based text-to-image models, enabling text-to-3D learning with only 2D supervision. To address the challenge of optimizing a NeRF field, a score-distillation loss is proposed [46] that leverages a pretrained 2D diffusion model as a critic to provide essential gradients. Subsequent efforts have focused on adopting this loss in latent space [13, 33] and in a coarse-to-fine refinement approach [29]. However, optimization-based methods are plagued by long convergence times. A recent concurrent work [51] proposes a non-optimization approach with progressive updates from multiple pre-set viewpoints. In contrast, our method iteratively updates and refines the synthesized textures from automatically selected viewpoint sequences, which minimizes human efforts with designing different viewpoint orders for various geometries." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18559" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.652, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.09, + 0.885, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.311, + 0.893, + 0.402 + ], + "angle": 0, + "content": "Figure 2: Overview of Text2Tex. We illustrate the pipeline using a 3D car mesh with a prompt \"golden Porsche\". We progressively generate the texture via a generate-then-refine scheme. In progressive texture generation (Sec. 3.3), we start by rendering the object from an initial preset viewpoint. We generate a new appearance according to the input prompt via a depth-to-image diffusion model, and project the generated image back to the partial texture. Then, we repeat this process until the last preset viewpoint to output the initial textured mesh. In the subsequent texture refinement (Sec. 3.4), we update the initial texture from a sequence of automatically selected viewpoints to refine the stretched and blurry artifacts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.427, + 0.168, + 0.442 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.474, + 0.681 + ], + "angle": 0, + "content": "The objective of our work is to texture a 3D mesh using a pretrained text-to-image diffusion model. In this section, we begin by laying the foundation of the diffusion model in Sec. 3.1 and depth-aware inpainting model in Sec. 3.2. We then propose a generate-then-refine scheme for progressively synthesizing and updating the 3D textures in a coarse-to-fine fashion. In the progressive texture generation (Sec. 3.3), we paint the visible regions of the input geometry in an incremental fashion, following a sequence of predefined viewpoints. To ensure local and global consistency, we incorporate a view partition to guide the depth-aware inpainting objectives. Subsequently, we introduce an automatic viewpoint selection mechanism (Sec. 3.4) to perform texture refinement and address any issues of texture stretching and inconsistency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.692, + 0.206, + 0.708 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.47, + 0.868 + ], + "angle": 0, + "content": "We use a Denoising Diffusion Probabilistic Model (DDPM) [24] as the generative model. Specifically, to avoid high computational overhead, we adopt the latent diffusion model [52], where an input image \\( x_0 \\) is first encoded into latent code \\( z_0 \\) before the diffusion process. The forward pass follows a Markov Chain to gradually add noise to the input latent code \\( z_0 \\) towards the white Gaussian noise \\( \\mathcal{N}(0,1) \\). At each step in the forward pass, the noised latent code \\( z_t \\) is obtained by adding a noise variance \\( \\beta_t \\) to the previous latent code \\( z_{t-1} \\) scaled with \\( \\sqrt{1 - \\beta_t} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.884, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\nz _ {t} \\sim \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.429, + 0.893, + 0.473 + ], + "angle": 0, + "content": "The independence property enables direct transformation of the noised latent code \\( z_{t} \\) at an arbitrary time step \\( t \\) from the input latent \\( z_{0} \\) via:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.485, + 0.892, + 0.503 + ], + "angle": 0, + "content": "\\[\nz _ {t} \\sim \\mathcal {N} (\\sqrt {\\bar {a}} _ {t} z _ {0}, (1 - \\bar {a} _ {t}) \\mathbf {I}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.515, + 0.892, + 0.56 + ], + "angle": 0, + "content": "where \\(\\bar{a}_t\\) is the total noise variance, which can be calculated by \\(\\sum_{t = i}^{T}(1 - \\beta_{t})\\) from the noise \\(\\beta_{t}\\) added to the input latent code \\(z_0\\) at each time step." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.561, + 0.893, + 0.606 + ], + "angle": 0, + "content": "During inference, the latent estimation \\(\\hat{z}_{t - 1}\\) for the next time step \\(t - 1\\) is obtained by predicting \\(\\mu_{\\theta}(z_t,t)\\) and \\(\\sigma_{\\theta}(z_t,t)\\) of a Gaussian distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.621, + 0.892, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {t - 1} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(z _ {t}, t\\right), \\sigma_ {\\theta} \\left(z _ {t}, t\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.656, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Denoising strength. To prevent complete randomness during the diffusion process, we introduce a scaling factor \\(\\gamma, 0 < \\gamma \\leq 1\\), which controls the number of diffusion steps. We assume that a white Gaussian noise \\(\\mathcal{N}(0,1)\\) can be obtained by adding noise to the input latent code \\(z_0\\) through \\(T\\) steps, and the final denoised latent estimation \\(\\hat{z}_0\\) is fully governed by the pure noise. By applying the scaling factor, we can start denoising the latent code at time step \\(\\gamma T\\) to guide the final latent code with the original image information. This technique is applied to refine the previously generated image contents." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.776, + 0.849 + ], + "angle": 0, + "content": "3.2. Depth-Aware Image Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The core of the texture synthesis lies in painting the missing regions on the mesh surface. The generated texture is expected to be highly faithful to the mesh geometry and the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18560" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.473, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.4, + 0.474, + 0.507 + ], + "angle": 0, + "content": "Figure 3: We dynamically partition the current view into a generation mask to guide the depth-aware inpainting model. For the \"new\" region, we denoise the new object appearance from white Gaussian noise. For the \"update\" region, we refine the previous texture by denoising the partially noised image segments. We freeze the texture in the \"keep\" region for this view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.656 + ], + "angle": 0, + "content": "input text. To achieve this, we build our method on a pretrained depth-to-image model [1, 67] that can produce high-quality images from text while being consistent with depth cues. However, since the Depth2Image model is designed to generate entire images, we need to use an inpainting mask to guide the sampling process. This mask provides explicit hints of which regions to generate or keep fixed, similar to the denoising guidance strategy in RePaint [31]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.471, + 0.718 + ], + "angle": 0, + "content": "To condition the denoising on the known regions of the input, we inject a generation mask \\(\\mathcal{M}\\) into the sampling steps. This mask explicitly blends the noised latent code \\(z_{t}\\) and the denoised latent estimation \\(\\hat{z}_t\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.735, + 0.47, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {t} = \\hat {z} _ {t} \\odot \\mathcal {M} + z _ {t} \\odot (1 - \\mathcal {M}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.759, + 0.471, + 0.791 + ], + "angle": 0, + "content": "We then decode the final denoised latent estimation \\(\\hat{z}_0\\) to the final output image \\(\\mathcal{O}\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.358, + 0.817 + ], + "angle": 0, + "content": "3.3. Progressive Texture Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "With the customized Depth2Image model, we are able to paint the object with a high quality image from a single view. To synthesize the appearance of the input geometry, we project the generated 2D views onto the texture space of a normalized 3D object with proper UV parameterization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Assuming Y-axis as the up-axis in the world coordinate system, we define the viewpoint as \\( v = (\\theta, \\phi, r) \\), where \\( \\theta \\) is the azimuth angle with respect to the Z-axis, \\( \\phi \\) is the viewpoint elevation angle with respect to the XZ-plane, and \\( r \\) is the distance between the viewpoint and the origin." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.304 + ], + "angle": 0, + "content": "As shown in Fig. 2, we start by generating the visible but missing texture in an initial viewpoint \\( v^0 \\). We render the object to a depth map \\( \\mathcal{D}^0 \\) and a generation mask \\( \\mathcal{M}^0 \\), and then use a customized Depth2Image diffusion model with \\( \\mathcal{D}^0 \\) as input and \\( \\mathcal{M}^0 \\) as extra guidance to generate a colored image \\( \\mathcal{O}^0 \\). We then back-project the image \\( \\mathcal{O}^0 \\) to the visible part of the texture \\( \\mathcal{T}^0 \\). In the subsequent steps, we progressively diffuse the colored images \\( \\mathcal{O}^k \\) and back-project them to the texture \\( \\mathcal{T}^k \\) through a sequence of viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.894, + 0.412 + ], + "angle": 0, + "content": "We notice that directly inpainting the missing regions on mesh surface often results in inconsistency issues. The issue is mainly caused by the stretched artifacts that occur when the 2D views are projected back onto the curved surface of the mesh. Therefore, we design a dynamic view partitioning strategy to guide the inpainting process with respective generation objectives \\(\\mathcal{M}\\) and denoising strengths \\(\\gamma\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.893, + 0.598 + ], + "angle": 0, + "content": "Dynamic view partitioning For all viewpoints \\(\\mathcal{V} = v_{i}, i = 1, \\dots, N\\), we render the similarity mask \\(S^i\\) for each viewpoint \\(v^k\\) and map those values to the texture space. Each pixel in a similarity mask represents the reversed normalized value of the cosine similarity between the normal vectors of the visible faces and the view direction (ranging from 0 to 1). A pixel with a value of 1 indicates that the corresponding face is perpendicular to the view direction. For simplicity, we set the background to 0. In summary, these masks indicate the extent to which a face is rotated away from the viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Based on the similarity mask \\(S^k\\) at step \\(k\\), we segment the rendered view into a generation mask \\(\\mathcal{M}^k\\), including the following 4 regions, as shown in Fig. 3: 1) New: This region contains pixels that have not yet been textured. We inpaint this region from pure white Gaussian noise, i.e. with denoising strength 1. 2) Update: This region contains pixels that have been textured, but the corresponding similarity score in \\(S^k\\) is greater than all other views. This indicates that those pixels are being observed in a better angle. Therefore, we update this region with a moderate denoising strength \\(\\gamma_g\\) to avoid stretched appearance. 3) Keep: Pixels in this region have been textured, but the corresponding similarity score in \\(S^k\\) is not the highest among all other views. These pixels have already been observed from a better angle, so we keep them fixed. 4) Ignore: This region contains pixels that belong to the background and are irrelevant to the process, so we ignore them throughout the entire process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.894, + 0.903 + ], + "angle": 0, + "content": "While the generation mask helps guide the texture inpainting process with accurate generation objectives and ap" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18561" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.381, + 0.47, + 0.489 + ], + "angle": 0, + "content": "Figure 4: In the refinement stage, the sequence of viewpoints are automatically determined by selecting the viewpoint with the maximum normalized area of the \"update\" region at each step. We update 2D views in the \"update\" regions with a moderate diffusion denoising strength. The updated object appearance is then back-projected to the texture space at the end of each refinement step." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.47, + 0.624 + ], + "angle": 0, + "content": "propriate denoising strengths, blurriness and stretches can still exist on the mesh surface. This is because the generation mask is limited to a predefined set of viewpoints, and the seams and stretches on the texture are still visible from a novel viewpoint. To address this issue, we propose a texture refinement technique with an automatic viewpoint selection strategy, which is described in the next section." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.47, + 0.665 + ], + "angle": 0, + "content": "3.4. Texture Refinement with Automatic Viewpoint Selection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.471, + 0.856 + ], + "angle": 0, + "content": "To remove the synthesis artifacts, a straightforward solution is to increase the number of viewpoints. However, the optimal viewpoint sequence can vary for different object geometries, making it difficult to manually pre-set the viewpoint sequence for massive synthesis targets. To address these challenges, we propose an automatic viewpoint selection strategy that effectively prevents stretches and seams, as illustrated in Fig. 4. We densely define a set of refinement viewpoints \\(\\mathcal{V} = v_{i}, i = 1,\\dots,K\\), where \\(K\\) is larger than \\(N\\). To distribute the refinement viewpoints evenly, we scatter them on a hemisphere, taking into account that objects are rarely observed from the bottom-up view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Assuming that an initial texture has been applied to the object, the refinement process begins by segmenting the generation masks \\(\\mathcal{M}\\) using the similarity masks \\(S\\) from all" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.337 + ], + "angle": 0, + "content": "available viewpoints. For each of the \\(K\\) refinement viewpoints in \\(\\mathcal{V}\\), we calculate a view heat \\(h^i\\) from the corresponding generation mask \\(\\mathcal{M}_i\\), which represents the normalized area of the \"update\" region with respect to the current visible area of the object. The viewpoint \\(v_i\\) that maximizes the view heat is then selected by \\(\\arg \\max_i h_i = \\frac{1}{N_p} \\sum_{k=1}^{N_p} w_k\\) where \\(N_p\\) is the total number of the non-background pixels, and \\(w_k\\) is the scaling factor for the segments in the generation mask. In order to let views with relatively more areas for updating, \\(w_k\\) for the \"update\" region is set to be bigger than that for the \"keep\" region. We dynamically select the next view with the highest view heat for updating. To avoid conflicts with the previously generated textures, we update the \"update\" regions with a mild denoising strength \\(\\gamma_r\\), which preserves the original appearance cues." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.35, + 0.587, + 0.365 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.375, + 0.718, + 0.391 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.895, + 0.55 + ], + "angle": 0, + "content": "We apply the Depth2Image model from Stable Diffusion v2 [1] as our generation backbone. The denoising strength \\(\\gamma_{g}\\) and \\(\\gamma_{r}\\) are set as 0.5 and 0.3 for the generation and refinement stages, respectively. We define 6 axis-aligned principles viewpoints for generation, and in total 36 viewpoints for refinement, among which we dynamically select only 20 views to reduce time cost. Each synthesis process takes around 15 minutes to complete on an NVIDIA RTX A6000. Our implementation uses the PyTorch [44] framework with PyTorch3D [50] used for rendering and texture projection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.558, + 0.677, + 0.574 + ], + "angle": 0, + "content": "4.2. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.894, + 0.809 + ], + "angle": 0, + "content": "Data. We evaluate our method on a subset of textured meshes from the Objaverse [17] dataset. We first sample 3 random meshes from each category. To ensure the quality of the input meshes, we manually filter out thin or unrecognizable meshes, such as \"strainer\", \"sweatpants\", and \"legging\", meshes with too simplistic textures, and meshes that do not correspond with their assigned categories. For the purpose of reducing processing time, we also remove overtriangulated and scanned objects. After this curation, there are in total 410 high quality textured meshes across 225 categories for the experiments. Note that the original textures are only used for the evaluation. To compare with GAN-based category-specific approaches, we also report results on the \"car\" objects from the ShapeNet dataset [7]. In particular, we use the 300 meshes from the test set used in [56]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Baselines. We compare our method against the following state-of-the-art text-driven texture synthesis method: 1) Text2Mesh [34], a neural pipeline that directly optimizes the textures and geometries via a CLIP-based optimization objective. We remove the displacement prediction so that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18562" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.891, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.477, + 0.893, + 0.524 + ], + "angle": 0, + "content": "Figure 5: Qualitative comparisons on Objaverse. We compare our textured mesh against CLIPMesh [37], Text2Mesh [34], Latent-Paint [33], and the original textures from Objaverse. In comparison with the baselines, our method produces more consistent and detailed 3D textures with respect to the input geometries. Image best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.473, + 0.732 + ], + "angle": 0, + "content": "only the surface RGBs are optimized. 2) CLIPMesh [37], a CLIP-based pipeline that deforms a sphere and optimizes the surface RGBs. Similar to Text2Mesh, we remove the shape deformation branch, and the texture colors are directly optimized on the surface of a given geometry. 3) Latent-Paint [33], a texture generation variant of the NeRF-based 3D object generation pipeline Latent-NeRF [33]. It explicitly operates on a given texture map using Stable Diffusion as a prior. In addition to text-guided methods, we also compare with category-specific GAN-based approaches, including Texture Fields [42], SPSG [16], LTG [65], and Texturify [56]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate the generated textures via commonly used image quality and diversity metrics for generative models. Specifically, we report the Frechet Inception Distance (FID) [23] and Kernel Inception Distance \\((\\mathrm{KID} \\times 10^{-3})\\) [4]. The generated image distribution for these metrics consists of renders of each mesh with the synthesized textures from 20 fixed viewpoints at a resolution of \\(512 \\times 512\\). For experiments on Objaverse dataset, the real distribution comprises renders of the meshes with the same settings using their artist designed textures. For ex" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.549, + 0.893, + 0.581 + ], + "angle": 0, + "content": "periments on ShapeNet cars, we use the 18991 background segmented images from CompCars dataset [64]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.69, + 0.618 + ], + "angle": 0, + "content": "4.3. Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In Tab. 1, we compare our method against the previous SOTA text-driven texture synthesis methods on Objverse objects. As input, we uniformly feed template texts \"a (category)\" to the models. Quantitatively, our method outperforms all baselines by a significant margin (19% improvement in FID and 26% improvement in KID). Such improvements demonstrate that our method is more capable of generating more realistic textures on various object geometries from numerous categories. To demonstrate the effectiveness of our method against the GAN-based approaches on category-specific data, we report experiment results on ShapeNet \"car\" category in Tab. 2. Notably, our method achieves superior performance over the previous GAN-based SOTA texture synthesis method Texturify, improving by 21% in FID and 12% in KID. This indicates our method is more effective with synthesizing realistic textures than GAN based approaches that were trained on specific categories." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18563" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.087, + 0.47, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.251, + 0.471, + 0.31 + ], + "angle": 0, + "content": "Figure 6: Qualitative comparisons on ShapeNet car. Our method generates sharper and more coherent textures with respect to the geometries compared to the state-of-the-art GAN-based method." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.325, + 0.447, + 0.424 + ], + "angle": 0, + "content": "
MethodFID ↓KID (x10-3) ↓
Text2Mesh [34]45.38(+9.7)10.40(+2.7)
CLIPMesh [37]43.25(+7.6)12.52(+4.8)
Latent-Paint [33]43.87(+8.1)11.43(+3.7)
Text2Tex (Ours)35.687.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.434, + 0.47, + 0.48 + ], + "angle": 0, + "content": "Table 1: Quantitative comparisons on Objverse subset. Our method performs favorably against state-of-the-art text-driven texture synthesis methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.747 + ], + "angle": 0, + "content": "User study. We conduct user study to analyze the quality of the synthesized textures and their fidelity to the input text prompts. For each baseline method, we randomly show the users 5 pair of renders from the baseline and our method. The users are requested to choose the one that is more realistic and closer to the text prompts. More details of the questionnaire can be found in the supplemental. In the end, we receive 604 responses across 41 users. The collected preferences are reported in Tab. 3. In comparison to CLIPMesh and Text2Mesh, our method is clearly preferred by the users with preference rate \\(83.92\\%\\) and \\(76.47\\%\\), respectively. Besides that, more users \\((64.18\\%)\\) lean towards our method over the competitive baseline Latent-Paint. As can be seen, our method demonstrates the effectiveness in generating high quality textures that are favored by human users." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.265, + 0.773 + ], + "angle": 0, + "content": "4.4. Qualitative analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.468, + 0.903 + ], + "angle": 0, + "content": "We compare our qualitative results on Objaverse objects against text-driven baselines in Fig. 5. In comparison with CLIP-based methods CLIPMesh and Text2Mesh, our method generates more realistic and consistent textures. In particular, CLIPMesh generates sketchy textures while Text2Mesh produces repetitive patterns. Latent-Paint outputs a consistent texture capturing the semantics of the object well, but the results are often quite blurry. Clearly," + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.089, + 0.886, + 0.203 + ], + "angle": 0, + "content": "
MethodFID ↓KID (x10-3) ↓
Texture Fields [42]177.15(+130.2)17.14(+12.8)
SPSG [16]110.65(+63.7)9.59(+5.2)
LTG [65]70.76(+23.8)5.72(+1.4)
Texturify [56]59.55(+12.6)4.97(+0.6)
Text2Tex (Ours)46.914.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.213, + 0.892, + 0.259 + ], + "angle": 0, + "content": "Table 2: Quantitative comparison on the ShapeNet cars. Our method outperform state-of-the-art category-specific GAN-based methods by a significant margin." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.276, + 0.886, + 0.324 + ], + "angle": 0, + "content": "
CLIPMesh ↑Text2Mesh ↑Latent-Paint ↑
Ours83.92%76.47%64.18%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.333, + 0.892, + 0.393 + ], + "angle": 0, + "content": "Table 3: Percentage of users who prefer our method over the baselines in a user study with 604 responses across 41 participants. Our method is shown to be more favored by human users." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.411, + 0.89, + 0.474 + ], + "angle": 0, + "content": "
w/ Depth2Imgw/ inpaintingw/ updateFID ↓KID (x10-3) ↓
✓XX39.889.78
✓✓X38.199.11
✓✓✓37.098.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.484, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Table 4: Effect of components in the generation stage. We quantitatively ablate the efficacy of each component. Applying the inpainting and update scheme effectively improves the quality of the synthesized textures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.681 + ], + "angle": 0, + "content": "our method can synthesize more consistent textures with cleaner and richer local details. We also compare our textures with GAN-based generation approach on category-specific objectives. In Fig. 6, we show the textures for ShapeNet cars of our methods and Texturify. Notably, our textures have a much cleaner appearance and provide more details with respect to the input geometries." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.661, + 0.71 + ], + "angle": 0, + "content": "4.5. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Does depth-aware inpainting and updating help? We show in Fig. 7 that the depth-aware inpainting is essential for producing high quality textures. In particular, the plain Depth2Img model often struggles to produce consistent appearances due to the governance of random noise. When the inpainting scheme is applied, the produced textures are more consistent. However, the textures still appear to be stretched and blurry over the curved mesh surface. These artifacts are amended by the texture updating scheme from the better viewing angles. The effectiveness of the depth-aware inpainting and the updating scheme is reflected in the improved FID and KID scores in Tab. 4." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18564" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.473, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.426, + 0.471, + 0.486 + ], + "angle": 0, + "content": "Figure 7: The proposed inpainting and update technique can effectively generate more consistent textures and eliminate stretched and blurry artifacts. Note that there is no refinement in this ablation." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.5, + 0.467, + 0.56 + ], + "angle": 0, + "content": "
# views05101520
↓FID37.0936.6736.3935.9835.68
↓KID (x10-3)8.788.318.127.987.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.632 + ], + "angle": 0, + "content": "Table 5: We quantitatively study the effect of selecting different number of viewpoints in the refinement stage. Refining the synthesis results with more viewpoints improves the texture quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.471, + 0.809 + ], + "angle": 0, + "content": "Does viewpoint selection in refinement stage help? We compare our results with different refinement settings in Tab. 5. When the input geometries are painted with initial textures from the generation stage, the blurry artifacts and projection seams are usually not eliminated due to a limited number of viewpoints. As can be seen in Fig. 8, such flaws can be minimized by refining with more viewpoints. We also showcase the effectiveness of the automatic viewpoint selection technique, as the refinement process does not require any manual efforts with defining and fine-tuning the viewpoint sequence for different shapes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.208, + 0.833 + ], + "angle": 0, + "content": "4.6. Limitations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "While our method has shown the capability to produce high-quality 3D textures, we have observed that it tends to produce textures with shading effects from the diffusion backbone. Although this issue can be addressed by care" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.088, + 0.891, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.427, + 0.893, + 0.474 + ], + "angle": 0, + "content": "Figure 8: The proposed automatic viewpoint selection method further improves the texture quality by gradually removing the remaining artifacts from the generation stage." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.492, + 0.892, + 0.584 + ], + "angle": 0, + "content": "fully fine-tuning the input prompts, doing so requires additional human engineering effort and may not scale well to massive generation targets. One potential solution is to fine-tune the diffusion model to remove the shading from textures. We acknowledge this challenge and leave it to future work to explore this possibility." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.619, + 0.618 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present a novel method, Text2Tex, for synthesizing high quality textures for 3D meshes from the given text prompts. Our approach leverages a depth-aware image inpainting diffusion model to progressively generate high-resolution partial textures from multiple viewpoints. To avoid accumulating inconsistent and stretching artifacts across viewpoints, we dynamically segment the rendered view into a generation mask, which effectively guides the diffusion model to generate and update the corresponding partial textures. Furthermore, we propose an automatic viewpoint sequence generation scheme that utilizes the generation mask to automatically determine the next best view for refining the generated textures. Extensive experiments demonstrate that our method can effectively synthesize consistent and highly detailed 3D textures for various object geometries without extra manual effort. Overall, we hope our work can inspire more future research in the area of text-to-3D synthesis." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18565" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.091, + 0.245, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.117, + 0.47, + 0.207 + ], + "angle": 0, + "content": "This work was supported by the ERC Starting Grant Scan2CAD (804724) and the German Research Foundation (DFG) Research Unit \"Learning and Simulation in Visual Computing\". We further thank Guy Gafni and Manuel Dahnert for the helpful discussions, and Angela Dai for the video voiceover." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.225, + 0.174, + 0.241 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.251, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[1] Stable diffusion. https://github.com/Stability-AI/stablediffusion, 2022. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.47, + 0.345 + ], + "angle": 0, + "content": "[2] Rameen Abdal, Hsin-Ying Lee, Peihao Zhu, Menglei Chai, Aliaksandr Siarohin, Peter Wonka, and Sergey Tulyakov. 3davatargan: Bridging domains for personalized editable avatars. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[3] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.386, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[4] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In ICLR, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022: 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.469, + 0.547 + ], + "angle": 0, + "content": "[6] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.548, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.469, + 0.67 + ], + "angle": 0, + "content": "[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. ScanRefer: 3D object localization in RGB-D scans using natural language. In European Conference on Computer Vision, pages 202-221. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.671, + 0.469, + 0.723 + ], + "angle": 0, + "content": "[9] Dave Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. arXiv preprint arXiv:2212.00836, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.724, + 0.469, + 0.806 + ], + "angle": 0, + "content": "[10] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 487-505. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.806, + 0.469, + 0.873 + ], + "angle": 0, + "content": "[11] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3193-3203, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.251, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[13] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.148, + 0.892, + 0.201 + ], + "angle": 0, + "content": "[14] Zezhou Cheng, Mengei Chai, Jian Ren, Hsin-Ying Lee, Kyle Olszewski, Zeng Huang, Subhransu Maji, and Sergey Tulyakov. Cross-modal 3d shape generation and manipulation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.202, + 0.892, + 0.256 + ], + "angle": 0, + "content": "[15] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.256, + 0.892, + 0.296 + ], + "angle": 0, + "content": "[16] Angela Dai, Yawar Siddiqui, Justus Thies, Julien Valentin, and Matthias Nießner. Spsg: Self-supervised photometric scene generation from rgb-d scans. In CVPR, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[17] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.364, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[19] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.431, + 0.892, + 0.498 + ], + "angle": 0, + "content": "[20] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.499, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[22] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.892, + 0.632 + ], + "angle": 0, + "content": "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. arXiv preprint arXiv:2106.15282, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.714, + 0.892, + 0.768 + ], + "angle": 0, + "content": "[26] Ronghang Hu and Amanpreet Singh. Unit: Multimodal multitask learning with a unified transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1439-1449, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.768, + 0.892, + 0.808 + ], + "angle": 0, + "content": "[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.808, + 0.892, + 0.862 + ], + "angle": 0, + "content": "[28] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.863, + 0.892, + 0.89 + ], + "angle": 0, + "content": "[29] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fi" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18566" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "dler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[30] Chieh Hubert Lin, Hsin-Ying Lee, Willi Menapace, Mengei Chai, Aliaksandr Siarohin, Ming-Hsuan Yang, and Sergey Tulyakov. *Infinicity: Infinite-scale city synthesis.* arXiv preprint arXiv:2301.09637, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.188, + 0.469, + 0.241 + ], + "angle": 0, + "content": "[31] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.242, + 0.469, + 0.268 + ], + "angle": 0, + "content": "[32] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.268, + 0.469, + 0.308 + ], + "angle": 0, + "content": "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.469, + 0.348 + ], + "angle": 0, + "content": "[34] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In CVPR, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.469, + 0.443 + ], + "angle": 0, + "content": "[36] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. AutoSDF: Shape priors for 3d completion, reconstruction and generation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.443, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[37] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.551 + ], + "angle": 0, + "content": "[38] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. arXiv preprint arXiv:2212.01206, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.469, + 0.606 + ], + "angle": 0, + "content": "[39] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International conference on machine learning, pages 7220-7229. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.469, + 0.632 + ], + "angle": 0, + "content": "[40] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[41] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.469, + 0.724 + ], + "angle": 0, + "content": "[42] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In ICCV, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.726, + 0.469, + 0.78 + ], + "angle": 0, + "content": "[43] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.78, + 0.469, + 0.834 + ], + "angle": 0, + "content": "[44] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.835, + 0.469, + 0.875 + ], + "angle": 0, + "content": "[45] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.875, + 0.469, + 0.888 + ], + "angle": 0, + "content": "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "hall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.12, + 0.892, + 0.187 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.187, + 0.892, + 0.241 + ], + "angle": 0, + "content": "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.242, + 0.892, + 0.282 + ], + "angle": 0, + "content": "[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.282, + 0.892, + 0.336 + ], + "angle": 0, + "content": "[50] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.336, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[51] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.416, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[53] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.892, + 0.497 + ], + "angle": 0, + "content": "[54] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[55] Aliaksandr Siarohin, Willi Menapace, Ivan Skorokhodov, Kyle Olszewski, Jian Ren, Hsin-Ying Lee, Mengei Chai, and Sergey Tulyakov. Unsupervised volumetric animation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.551, + 0.892, + 0.592 + ], + "angle": 0, + "content": "[56] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In ECCV, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.592, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[57] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[58] Ivan Skorokhodov, Aliaksandr Siarohin, Yinghao Xu, Jian Ren, Hsin-Ying Lee, Peter Wonka, and Sergey Tulyakov. 3d generation on imagenet. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.713, + 0.892, + 0.753 + ], + "angle": 0, + "content": "[59] Edward J Smith and David Meger. Improved adversarial systems for 3d object generation and reconstruction. In CoRL, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.753, + 0.892, + 0.835 + ], + "angle": 0, + "content": "[60] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.835, + 0.892, + 0.891 + ], + "angle": 0, + "content": "[61] Jiajun Wu, Chengkai Zhang, Xiuming Zhang, Zhoutong Zhang, William T Freeman, and Joshua B Tenenbaum. Learning shape priors for single-view 3d completion and reconstruction. In Proceedings of the European Conference on" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18567" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.092, + 0.414, + 0.106 + ], + "angle": 0, + "content": "Computer Vision (ECCV), pages 646-662, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.107, + 0.47, + 0.159 + ], + "angle": 0, + "content": "[62] Jianwen Xie, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Song-Chun Zhu, and Ying Nian Wu. Learning descriptor networks for 3d shape synthesis and analysis. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.16, + 0.47, + 0.228 + ], + "angle": 0, + "content": "[63] Yinghao Xu, Mengei Chai, Zifan Shi, Sida Peng, Ivan Skorokhodov, Aliaksandr Siarohin, Ceyuan Yang, Yujun Shen, Hsin-Ying Lee, Bolei Zhou, et al. Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.228, + 0.47, + 0.295 + ], + "angle": 0, + "content": "[64] Linjie Yang, Ping Luo, Chen Change Loy, and Xiaou Tang. A large-scale car dataset for fine-grained categorization and verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3973-3981, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.296, + 0.47, + 0.336 + ], + "angle": 0, + "content": "[65] Rui Yu, Yue Dong, Pieter Peers, and Xin Tong. Learning texture generators for 3d shape collections from internet photo sets. In BMVC, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.337, + 0.47, + 0.39 + ], + "angle": 0, + "content": "[66] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.431 + ], + "angle": 0, + "content": "[67] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.431, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[68] Song-Hai Zhang, Yuan-Chen Guo, and Qing-Wen Gu. Sketch2model: View-aware 3d modeling from single free-hand sketches. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18568" + } + ] +] \ No newline at end of file diff --git a/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/8023656a-1169-49dd-b8cd-4db39941377d_origin.pdf b/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/8023656a-1169-49dd-b8cd-4db39941377d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..079773cc2da0ff4607238da11da5eb9b6a21d846 --- /dev/null +++ b/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/8023656a-1169-49dd-b8cd-4db39941377d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e664713ee6ab15d41b5334cc92b2c7d95c32bbf88fb498433e7152f35e46e53e +size 4213541 diff --git a/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/full.md b/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d681ecbffaba5c9300e32c071d9f672fd66c35bb --- /dev/null +++ b/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/full.md @@ -0,0 +1,293 @@ +# Text2Tex: Text-driven Texture Synthesis via Diffusion Models + +# Dave Zhenyu Chen + +# Yawar Siddiqui + +# Hsin-Ying Lee2 + +# Sergey Tulyakov2 + +# Matthias Nießner1 + +# 1Technical University of Munich + +# $^{2}$ Snap Research + +# https://daveredrum.github.io/Text2Tex/ + +![](images/325025c2a43659fdd8b91347ff83c17846c00c8bf1bf4ffc6466f79b4a5e8104.jpg) +Figure 1: We introduce Text2Tex, a text-driven architecture for 3D texture synthesis. Given object geometries and text prompts as input, Text2Tex generates high quality and consistent textures via depth-aware inpainting and refinement. + +# Abstract + +We present Text2Tex, a novel method for generating high-quality textures for 3D meshes from the given text prompts. Our method incorporates inpainting into a pre-trained depth-aware image diffusion model to progressively synthesize high resolution partial textures from multiple viewpoints. To avoid accumulating inconsistent and stretched artifacts across views, we dynamically segment the rendered view into a generation mask, which represents the generation status of each visible texel. This partitioned view representation guides the depth-aware inpainting model to generate and update partial textures for the corresponding regions. Furthermore, we propose an automatic view sequence generation scheme to determine the next best view for updating the partial texture. Extensive experiments demonstrate that our method significantly outperforms the existing text-driven approaches and GAN-based methods. + +# 1. Introduction + +Generating high-quality 3D content is an essential component of visual applications in films, games, and upcoming AR/VR scenarios. With an increasing number of 3D content datasets, the computer vision community has wit- + +nessed significant progress in the field of 3D geometry generation [12, 39, 61, 66, 38]. Despite the remarkable success in modeling 3D geometries in recent years, fully automatic 3D content generation is still hindered by the laborious human efforts required to design textures. Therefore, automating the texture design process through alternative guidance, such as text, has become an intriguing but challenging research problem. + +Recently, text-to-image generators have shown remarkable progress in the 2D domain leveraging diffusion model architectures, enabling high resolution 2D content generation based on textual descriptions [1, 48]. However, there are notable challenges for producing 3D textures via such 2D vision-language prior knowledge. Specifically, the synthesized textures are expected to be not only with high fidelity to the language cues, but also of high and consistent quality for target meshes. As such, previous attempts to paint 3D geometry from text inputs often fail to deliver well-textured 3D content. + +In this paper, we introduce Text2Tex, a novel texture synthesis method that seamlessly texturizes 3D objects using a pre-trained depth-aware text-to-image diffusion model. The method renders a target mesh from multiple viewpoints and inpaints the missing appearance with a + +depth-aware text-to-image diffusion model. Text2Tex follows a generate-then-refine strategy. Our method progressively generates partial textures across viewpoints and backprojects them to texture space. To address stretched and inconsistent artifacts observed from rotated viewpoints, we design a view partitioning technique that computes similarity maps between visible texel's normal vectors and the current view direction. The generation mask created from these similarity maps guides the diffusion process by indicating regions to generate, update, keep, or ignore. This allows us to apply different diffusion strengths to respective regions, inpainting missing appearance and updating stretched artifacts. However, the autoregressive generation process via the diffusion-based image inpainting model presents a new challenge. As the inpainting and updating scheme is conditioned on previously synthesized results, a viewpoint sequence with an ill-defined order or incomplete coverage over the mesh surface may result in unsatisfactory texturization. Therefore, we propose an automatic viewpoint selection technique that progressively selects the next best view. The confidence of each candidate view containing the biggest relative area for generation and updating is estimated, given the partially textured mesh. This approach ensures complete coverage over the mesh surface and a high-quality texture map by consistently updating stretched regions. + +We demonstrate the effectiveness of Text2Tex for synthesizing high-quality 3D textures from language cues. The proposed method performs favorably against other language-based texture synthesis methods in terms of FID [23], KID [4], and user study on a subset of the Objaverse dataset [17]. Additionally, our method also outperforms category-specific GAN-based methods on the ShapeNet car dataset [7]. + +To summarize, our technical contributions are threefold: + +- We design a novel method for high-quality texture synthesis by progressively inpainting and updating the 3D textures via depth-aware diffusion models. +- We propose an automatic view sequence generation scheme to dynamically determine the order for generating and updating the texture space. +- We conduct extensive study on a considerable amount of 3D objects, demonstrating the proposed method is effective for large-scale 3D content generation. + +# 2. Related work + +3D Generation from 3D and 2D data. To achieve 3D generation, it is natural to train models directly on 3D data. In contrast to 2D images, there are several 3D representations available, each with its unique characteristics, leading to the development of various generative models such as + +those based on voxels [30, 55, 59, 62], point clouds [3, 32], meshes [68], signed distance function [12, 13, 14, 16, 36], etc. However, unlike images or videos that are ubiquitous, 3D data is inherently scarce and challenging to collect and annotate. Consequently, the synthesized samples from 3D generative models, trained on 3D data, are of limited quality and diversity, in terms of both structure and texture. Recent works have leveraged differentiable rendering to learn texture generation using only 2D images [20, 56, 65]. However, they are typically trained for specific shape categories and struggle in the quality of textures. + +Text-Guided Generation. Recently, there has been tremendous progress in the vision-language domain [26, 57, 60, 47, 10, 8, 11, 9]. Specifically, the emergence of Contrastive Language-Image Pre-Training (CLIP) [47] has enabled the development of text-guided image generation through its semantically rich representation trained on text-image pairs. Initial efforts [15, 45] incorporated CLIP with different backbones, such as StyleGAN [27] and VQGAN [19]. However, diffusion models [18, 24, 25, 40, 53], which have gained attention due to their superior visual quality and training stability compared to Generative Adversarial Networks (GANs) [21], have recently been trained on large-scale text-image datasets with CLIP encodings [28, 49, 52]. Among these models, Stable Diffusion [1, 52] has garnered significant interest as an open-sourced model with numerous extensions [1, 67] that support different conditional modalities in addition to text prompts, including depth images, poses, sketches, etc. Additionally, CLIP has also been adopted in 3D to perform text-guided shape and texture generation [34, 37]. In this work, we take advantage of the depth-conditioning feature of Stable Diffusion to provide more consistent texturing. + +Text-to-3D from 2D data. Inspired by the success of Neural Radiance Fields (NeRF) [35], NeRF-based generators [2, 5, 6, 22, 41, 43, 54, 58, 63] have been proposed to learn 3D structures from 2D images using GAN-based frameworks. A new research direction emerged by combining NeRF techniques with flourishing diffusion-based text-to-image models, enabling text-to-3D learning with only 2D supervision. To address the challenge of optimizing a NeRF field, a score-distillation loss is proposed [46] that leverages a pretrained 2D diffusion model as a critic to provide essential gradients. Subsequent efforts have focused on adopting this loss in latent space [13, 33] and in a coarse-to-fine refinement approach [29]. However, optimization-based methods are plagued by long convergence times. A recent concurrent work [51] proposes a non-optimization approach with progressive updates from multiple pre-set viewpoints. In contrast, our method iteratively updates and refines the synthesized textures from automatically selected viewpoint sequences, which minimizes human efforts with designing different viewpoint orders for various geometries. + +![](images/af9726b05d572e3372214fe3a3f4d751977d8c7bca91eab3a69ad543c0bf55a9.jpg) +Figure 2: Overview of Text2Tex. We illustrate the pipeline using a 3D car mesh with a prompt "golden Porsche". We progressively generate the texture via a generate-then-refine scheme. In progressive texture generation (Sec. 3.3), we start by rendering the object from an initial preset viewpoint. We generate a new appearance according to the input prompt via a depth-to-image diffusion model, and project the generated image back to the partial texture. Then, we repeat this process until the last preset viewpoint to output the initial textured mesh. In the subsequent texture refinement (Sec. 3.4), we update the initial texture from a sequence of automatically selected viewpoints to refine the stretched and blurry artifacts. + +![](images/fc07aaf42c347ed25813041dfa6a9bf54aab089619c2a563a6a4fe381eccc68e.jpg) + +# 3. Method + +The objective of our work is to texture a 3D mesh using a pretrained text-to-image diffusion model. In this section, we begin by laying the foundation of the diffusion model in Sec. 3.1 and depth-aware inpainting model in Sec. 3.2. We then propose a generate-then-refine scheme for progressively synthesizing and updating the 3D textures in a coarse-to-fine fashion. In the progressive texture generation (Sec. 3.3), we paint the visible regions of the input geometry in an incremental fashion, following a sequence of predefined viewpoints. To ensure local and global consistency, we incorporate a view partition to guide the depth-aware inpainting objectives. Subsequently, we introduce an automatic viewpoint selection mechanism (Sec. 3.4) to perform texture refinement and address any issues of texture stretching and inconsistency. + +# 3.1. Preliminary + +We use a Denoising Diffusion Probabilistic Model (DDPM) [24] as the generative model. Specifically, to avoid high computational overhead, we adopt the latent diffusion model [52], where an input image $x_0$ is first encoded into latent code $z_0$ before the diffusion process. The forward pass follows a Markov Chain to gradually add noise to the input latent code $z_0$ towards the white Gaussian noise $\mathcal{N}(0,1)$ . At each step in the forward pass, the noised latent code $z_t$ is obtained by adding a noise variance $\beta_t$ to the previous latent code $z_{t-1}$ scaled with $\sqrt{1 - \beta_t}$ : + +$$ +z _ {t} \sim \mathcal {N} \left(\sqrt {1 - \beta_ {t}} z _ {t - 1}, \beta_ {t} \mathbf {I}\right). \tag {1} +$$ + +The independence property enables direct transformation of the noised latent code $z_{t}$ at an arbitrary time step $t$ from the input latent $z_{0}$ via: + +$$ +z _ {t} \sim \mathcal {N} (\sqrt {\bar {a}} _ {t} z _ {0}, (1 - \bar {a} _ {t}) \mathbf {I}), \tag {2} +$$ + +where $\bar{a}_t$ is the total noise variance, which can be calculated by $\sum_{t = i}^{T}(1 - \beta_{t})$ from the noise $\beta_{t}$ added to the input latent code $z_0$ at each time step. + +During inference, the latent estimation $\hat{z}_{t - 1}$ for the next time step $t - 1$ is obtained by predicting $\mu_{\theta}(z_t,t)$ and $\sigma_{\theta}(z_t,t)$ of a Gaussian distribution: + +$$ +\hat {z} _ {t - 1} \sim \mathcal {N} \left(\mu_ {\theta} \left(z _ {t}, t\right), \sigma_ {\theta} \left(z _ {t}, t\right)\right) \tag {3} +$$ + +Denoising strength. To prevent complete randomness during the diffusion process, we introduce a scaling factor $\gamma, 0 < \gamma \leq 1$ , which controls the number of diffusion steps. We assume that a white Gaussian noise $\mathcal{N}(0,1)$ can be obtained by adding noise to the input latent code $z_0$ through $T$ steps, and the final denoised latent estimation $\hat{z}_0$ is fully governed by the pure noise. By applying the scaling factor, we can start denoising the latent code at time step $\gamma T$ to guide the final latent code with the original image information. This technique is applied to refine the previously generated image contents. + +# 3.2. Depth-Aware Image Inpainting + +The core of the texture synthesis lies in painting the missing regions on the mesh surface. The generated texture is expected to be highly faithful to the mesh geometry and the + +![](images/a3f7c5976aa72fad20aaf2fed00b8ed53e16d6d77b6963f2210e064ae2fd7267.jpg) +Figure 3: We dynamically partition the current view into a generation mask to guide the depth-aware inpainting model. For the "new" region, we denoise the new object appearance from white Gaussian noise. For the "update" region, we refine the previous texture by denoising the partially noised image segments. We freeze the texture in the "keep" region for this view. + +input text. To achieve this, we build our method on a pretrained depth-to-image model [1, 67] that can produce high-quality images from text while being consistent with depth cues. However, since the Depth2Image model is designed to generate entire images, we need to use an inpainting mask to guide the sampling process. This mask provides explicit hints of which regions to generate or keep fixed, similar to the denoising guidance strategy in RePaint [31]. + +To condition the denoising on the known regions of the input, we inject a generation mask $\mathcal{M}$ into the sampling steps. This mask explicitly blends the noised latent code $z_{t}$ and the denoised latent estimation $\hat{z}_t$ as follows: + +$$ +\hat {z} _ {t} = \hat {z} _ {t} \odot \mathcal {M} + z _ {t} \odot (1 - \mathcal {M}). \tag {4} +$$ + +We then decode the final denoised latent estimation $\hat{z}_0$ to the final output image $\mathcal{O}$ . + +# 3.3. Progressive Texture Generation + +With the customized Depth2Image model, we are able to paint the object with a high quality image from a single view. To synthesize the appearance of the input geometry, we project the generated 2D views onto the texture space of a normalized 3D object with proper UV parameterization. + +Assuming Y-axis as the up-axis in the world coordinate system, we define the viewpoint as $v = (\theta, \phi, r)$ , where $\theta$ is the azimuth angle with respect to the Z-axis, $\phi$ is the viewpoint elevation angle with respect to the XZ-plane, and $r$ is the distance between the viewpoint and the origin. + +As shown in Fig. 2, we start by generating the visible but missing texture in an initial viewpoint $v^0$ . We render the object to a depth map $\mathcal{D}^0$ and a generation mask $\mathcal{M}^0$ , and then use a customized Depth2Image diffusion model with $\mathcal{D}^0$ as input and $\mathcal{M}^0$ as extra guidance to generate a colored image $\mathcal{O}^0$ . We then back-project the image $\mathcal{O}^0$ to the visible part of the texture $\mathcal{T}^0$ . In the subsequent steps, we progressively diffuse the colored images $\mathcal{O}^k$ and back-project them to the texture $\mathcal{T}^k$ through a sequence of viewpoints. + +We notice that directly inpainting the missing regions on mesh surface often results in inconsistency issues. The issue is mainly caused by the stretched artifacts that occur when the 2D views are projected back onto the curved surface of the mesh. Therefore, we design a dynamic view partitioning strategy to guide the inpainting process with respective generation objectives $\mathcal{M}$ and denoising strengths $\gamma$ . + +Dynamic view partitioning For all viewpoints $\mathcal{V} = v_{i}, i = 1, \dots, N$ , we render the similarity mask $S^i$ for each viewpoint $v^k$ and map those values to the texture space. Each pixel in a similarity mask represents the reversed normalized value of the cosine similarity between the normal vectors of the visible faces and the view direction (ranging from 0 to 1). A pixel with a value of 1 indicates that the corresponding face is perpendicular to the view direction. For simplicity, we set the background to 0. In summary, these masks indicate the extent to which a face is rotated away from the viewpoint. + +Based on the similarity mask $S^k$ at step $k$ , we segment the rendered view into a generation mask $\mathcal{M}^k$ , including the following 4 regions, as shown in Fig. 3: 1) New: This region contains pixels that have not yet been textured. We inpaint this region from pure white Gaussian noise, i.e. with denoising strength 1. 2) Update: This region contains pixels that have been textured, but the corresponding similarity score in $S^k$ is greater than all other views. This indicates that those pixels are being observed in a better angle. Therefore, we update this region with a moderate denoising strength $\gamma_g$ to avoid stretched appearance. 3) Keep: Pixels in this region have been textured, but the corresponding similarity score in $S^k$ is not the highest among all other views. These pixels have already been observed from a better angle, so we keep them fixed. 4) Ignore: This region contains pixels that belong to the background and are irrelevant to the process, so we ignore them throughout the entire process. + +While the generation mask helps guide the texture inpainting process with accurate generation objectives and ap + +![](images/de0fb4cfc55b75d280ee7caa5fb9247df42a3bd65b283bf5809b5cc1503bf40c.jpg) +Figure 4: In the refinement stage, the sequence of viewpoints are automatically determined by selecting the viewpoint with the maximum normalized area of the "update" region at each step. We update 2D views in the "update" regions with a moderate diffusion denoising strength. The updated object appearance is then back-projected to the texture space at the end of each refinement step. + +propriate denoising strengths, blurriness and stretches can still exist on the mesh surface. This is because the generation mask is limited to a predefined set of viewpoints, and the seams and stretches on the texture are still visible from a novel viewpoint. To address this issue, we propose a texture refinement technique with an automatic viewpoint selection strategy, which is described in the next section. + +# 3.4. Texture Refinement with Automatic Viewpoint Selection + +To remove the synthesis artifacts, a straightforward solution is to increase the number of viewpoints. However, the optimal viewpoint sequence can vary for different object geometries, making it difficult to manually pre-set the viewpoint sequence for massive synthesis targets. To address these challenges, we propose an automatic viewpoint selection strategy that effectively prevents stretches and seams, as illustrated in Fig. 4. We densely define a set of refinement viewpoints $\mathcal{V} = v_{i}, i = 1,\dots,K$ , where $K$ is larger than $N$ . To distribute the refinement viewpoints evenly, we scatter them on a hemisphere, taking into account that objects are rarely observed from the bottom-up view. + +Assuming that an initial texture has been applied to the object, the refinement process begins by segmenting the generation masks $\mathcal{M}$ using the similarity masks $S$ from all + +available viewpoints. For each of the $K$ refinement viewpoints in $\mathcal{V}$ , we calculate a view heat $h^i$ from the corresponding generation mask $\mathcal{M}_i$ , which represents the normalized area of the "update" region with respect to the current visible area of the object. The viewpoint $v_i$ that maximizes the view heat is then selected by $\arg \max_i h_i = \frac{1}{N_p} \sum_{k=1}^{N_p} w_k$ where $N_p$ is the total number of the non-background pixels, and $w_k$ is the scaling factor for the segments in the generation mask. In order to let views with relatively more areas for updating, $w_k$ for the "update" region is set to be bigger than that for the "keep" region. We dynamically select the next view with the highest view heat for updating. To avoid conflicts with the previously generated textures, we update the "update" regions with a mild denoising strength $\gamma_r$ , which preserves the original appearance cues. + +# 4. Results + +# 4.1. Implementation Details + +We apply the Depth2Image model from Stable Diffusion v2 [1] as our generation backbone. The denoising strength $\gamma_{g}$ and $\gamma_{r}$ are set as 0.5 and 0.3 for the generation and refinement stages, respectively. We define 6 axis-aligned principles viewpoints for generation, and in total 36 viewpoints for refinement, among which we dynamically select only 20 views to reduce time cost. Each synthesis process takes around 15 minutes to complete on an NVIDIA RTX A6000. Our implementation uses the PyTorch [44] framework with PyTorch3D [50] used for rendering and texture projection. + +# 4.2. Experiment Setup + +Data. We evaluate our method on a subset of textured meshes from the Objaverse [17] dataset. We first sample 3 random meshes from each category. To ensure the quality of the input meshes, we manually filter out thin or unrecognizable meshes, such as "strainer", "sweatpants", and "legging", meshes with too simplistic textures, and meshes that do not correspond with their assigned categories. For the purpose of reducing processing time, we also remove overtriangulated and scanned objects. After this curation, there are in total 410 high quality textured meshes across 225 categories for the experiments. Note that the original textures are only used for the evaluation. To compare with GAN-based category-specific approaches, we also report results on the "car" objects from the ShapeNet dataset [7]. In particular, we use the 300 meshes from the test set used in [56]. + +Baselines. We compare our method against the following state-of-the-art text-driven texture synthesis method: 1) Text2Mesh [34], a neural pipeline that directly optimizes the textures and geometries via a CLIP-based optimization objective. We remove the displacement prediction so that + +![](images/9eb23aece4df7bb0df9d72d4d6ae6a5bd1bac0f9ce28267c2238b96121a089ae.jpg) +Figure 5: Qualitative comparisons on Objaverse. We compare our textured mesh against CLIPMesh [37], Text2Mesh [34], Latent-Paint [33], and the original textures from Objaverse. In comparison with the baselines, our method produces more consistent and detailed 3D textures with respect to the input geometries. Image best viewed in color. + +only the surface RGBs are optimized. 2) CLIPMesh [37], a CLIP-based pipeline that deforms a sphere and optimizes the surface RGBs. Similar to Text2Mesh, we remove the shape deformation branch, and the texture colors are directly optimized on the surface of a given geometry. 3) Latent-Paint [33], a texture generation variant of the NeRF-based 3D object generation pipeline Latent-NeRF [33]. It explicitly operates on a given texture map using Stable Diffusion as a prior. In addition to text-guided methods, we also compare with category-specific GAN-based approaches, including Texture Fields [42], SPSG [16], LTG [65], and Texturify [56]. + +Evaluation metrics. We evaluate the generated textures via commonly used image quality and diversity metrics for generative models. Specifically, we report the Frechet Inception Distance (FID) [23] and Kernel Inception Distance $(\mathrm{KID} \times 10^{-3})$ [4]. The generated image distribution for these metrics consists of renders of each mesh with the synthesized textures from 20 fixed viewpoints at a resolution of $512 \times 512$ . For experiments on Objaverse dataset, the real distribution comprises renders of the meshes with the same settings using their artist designed textures. For ex + +periments on ShapeNet cars, we use the 18991 background segmented images from CompCars dataset [64]. + +# 4.3. Quantitative results + +In Tab. 1, we compare our method against the previous SOTA text-driven texture synthesis methods on Objverse objects. As input, we uniformly feed template texts "a (category)" to the models. Quantitatively, our method outperforms all baselines by a significant margin (19% improvement in FID and 26% improvement in KID). Such improvements demonstrate that our method is more capable of generating more realistic textures on various object geometries from numerous categories. To demonstrate the effectiveness of our method against the GAN-based approaches on category-specific data, we report experiment results on ShapeNet "car" category in Tab. 2. Notably, our method achieves superior performance over the previous GAN-based SOTA texture synthesis method Texturify, improving by 21% in FID and 12% in KID. This indicates our method is more effective with synthesizing realistic textures than GAN based approaches that were trained on specific categories. + +![](images/ecad2569c5931d313c99782523227924b57d97a57fb94380dce4e7b9f4648dd1.jpg) +Figure 6: Qualitative comparisons on ShapeNet car. Our method generates sharper and more coherent textures with respect to the geometries compared to the state-of-the-art GAN-based method. + +
MethodFID ↓KID (x10-3) ↓
Text2Mesh [34]45.38(+9.7)10.40(+2.7)
CLIPMesh [37]43.25(+7.6)12.52(+4.8)
Latent-Paint [33]43.87(+8.1)11.43(+3.7)
Text2Tex (Ours)35.687.74
+ +User study. We conduct user study to analyze the quality of the synthesized textures and their fidelity to the input text prompts. For each baseline method, we randomly show the users 5 pair of renders from the baseline and our method. The users are requested to choose the one that is more realistic and closer to the text prompts. More details of the questionnaire can be found in the supplemental. In the end, we receive 604 responses across 41 users. The collected preferences are reported in Tab. 3. In comparison to CLIPMesh and Text2Mesh, our method is clearly preferred by the users with preference rate $83.92\%$ and $76.47\%$ , respectively. Besides that, more users $(64.18\%)$ lean towards our method over the competitive baseline Latent-Paint. As can be seen, our method demonstrates the effectiveness in generating high quality textures that are favored by human users. + +# 4.4. Qualitative analysis + +We compare our qualitative results on Objaverse objects against text-driven baselines in Fig. 5. In comparison with CLIP-based methods CLIPMesh and Text2Mesh, our method generates more realistic and consistent textures. In particular, CLIPMesh generates sketchy textures while Text2Mesh produces repetitive patterns. Latent-Paint outputs a consistent texture capturing the semantics of the object well, but the results are often quite blurry. Clearly, + +Table 1: Quantitative comparisons on Objverse subset. Our method performs favorably against state-of-the-art text-driven texture synthesis methods. + +
MethodFID ↓KID (x10-3) ↓
Texture Fields [42]177.15(+130.2)17.14(+12.8)
SPSG [16]110.65(+63.7)9.59(+5.2)
LTG [65]70.76(+23.8)5.72(+1.4)
Texturify [56]59.55(+12.6)4.97(+0.6)
Text2Tex (Ours)46.914.35
+ +Table 2: Quantitative comparison on the ShapeNet cars. Our method outperform state-of-the-art category-specific GAN-based methods by a significant margin. + +
CLIPMesh ↑Text2Mesh ↑Latent-Paint ↑
Ours83.92%76.47%64.18%
+ +Table 3: Percentage of users who prefer our method over the baselines in a user study with 604 responses across 41 participants. Our method is shown to be more favored by human users. + +
w/ Depth2Imgw/ inpaintingw/ updateFID ↓KID (x10-3) ↓
✓XX39.889.78
✓✓X38.199.11
✓✓✓37.098.78
+ +Table 4: Effect of components in the generation stage. We quantitatively ablate the efficacy of each component. Applying the inpainting and update scheme effectively improves the quality of the synthesized textures. + +our method can synthesize more consistent textures with cleaner and richer local details. We also compare our textures with GAN-based generation approach on category-specific objectives. In Fig. 6, we show the textures for ShapeNet cars of our methods and Texturify. Notably, our textures have a much cleaner appearance and provide more details with respect to the input geometries. + +# 4.5. Ablation studies + +Does depth-aware inpainting and updating help? We show in Fig. 7 that the depth-aware inpainting is essential for producing high quality textures. In particular, the plain Depth2Img model often struggles to produce consistent appearances due to the governance of random noise. When the inpainting scheme is applied, the produced textures are more consistent. However, the textures still appear to be stretched and blurry over the curved mesh surface. These artifacts are amended by the texture updating scheme from the better viewing angles. The effectiveness of the depth-aware inpainting and the updating scheme is reflected in the improved FID and KID scores in Tab. 4. + +![](images/707f8f03a26e274ccb1740e84ce734a092493b02293f3562d38b86f2deb955ba.jpg) +Figure 7: The proposed inpainting and update technique can effectively generate more consistent textures and eliminate stretched and blurry artifacts. Note that there is no refinement in this ablation. + +
# views05101520
↓FID37.0936.6736.3935.9835.68
↓KID (x10-3)8.788.318.127.987.74
+ +Table 5: We quantitatively study the effect of selecting different number of viewpoints in the refinement stage. Refining the synthesis results with more viewpoints improves the texture quality. + +Does viewpoint selection in refinement stage help? We compare our results with different refinement settings in Tab. 5. When the input geometries are painted with initial textures from the generation stage, the blurry artifacts and projection seams are usually not eliminated due to a limited number of viewpoints. As can be seen in Fig. 8, such flaws can be minimized by refining with more viewpoints. We also showcase the effectiveness of the automatic viewpoint selection technique, as the refinement process does not require any manual efforts with defining and fine-tuning the viewpoint sequence for different shapes. + +# 4.6. Limitations. + +While our method has shown the capability to produce high-quality 3D textures, we have observed that it tends to produce textures with shading effects from the diffusion backbone. Although this issue can be addressed by care + +![](images/9d69f53354c4c876204d8c3cf5fd072b538fabc551fb18558616d3ddc27129f6.jpg) +Figure 8: The proposed automatic viewpoint selection method further improves the texture quality by gradually removing the remaining artifacts from the generation stage. + +fully fine-tuning the input prompts, doing so requires additional human engineering effort and may not scale well to massive generation targets. One potential solution is to fine-tune the diffusion model to remove the shading from textures. We acknowledge this challenge and leave it to future work to explore this possibility. + +# 5. Conclusion + +In this paper, we present a novel method, Text2Tex, for synthesizing high quality textures for 3D meshes from the given text prompts. Our approach leverages a depth-aware image inpainting diffusion model to progressively generate high-resolution partial textures from multiple viewpoints. To avoid accumulating inconsistent and stretching artifacts across viewpoints, we dynamically segment the rendered view into a generation mask, which effectively guides the diffusion model to generate and update the corresponding partial textures. Furthermore, we propose an automatic viewpoint sequence generation scheme that utilizes the generation mask to automatically determine the next best view for refining the generated textures. Extensive experiments demonstrate that our method can effectively synthesize consistent and highly detailed 3D textures for various object geometries without extra manual effort. Overall, we hope our work can inspire more future research in the area of text-to-3D synthesis. + +# Acknowledgements + +This work was supported by the ERC Starting Grant Scan2CAD (804724) and the German Research Foundation (DFG) Research Unit "Learning and Simulation in Visual Computing". We further thank Guy Gafni and Manuel Dahnert for the helpful discussions, and Angela Dai for the video voiceover. + +# References + +[1] Stable diffusion. https://github.com/Stability-AI/stablediffusion, 2022. 1, 2, 4, 5 +[2] Rameen Abdal, Hsin-Ying Lee, Peihao Zhu, Menglei Chai, Aliaksandr Siarohin, Peter Wonka, and Sergey Tulyakov. 3davatargan: Bridging domains for personalized editable avatars. In CVPR, 2023. 2 +[3] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 2 +[4] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In ICLR, 2018. 2, 6 +[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022: 2 +[6] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2 +[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 5 +[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. ScanRefer: 3D object localization in RGB-D scans using natural language. In European Conference on Computer Vision, pages 202-221. Springer, 2020. 2 +[9] Dave Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. arXiv preprint arXiv:2212.00836, 2022. 2 +[10] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 487-505. Springer, 2022. 2 +[11] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3193-3203, 2021. 2 +[12] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 1, 2 + +[13] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, 2023. 2 +[14] Zezhou Cheng, Mengei Chai, Jian Ren, Hsin-Ying Lee, Kyle Olszewski, Zeng Huang, Subhransu Maji, and Sergey Tulyakov. Cross-modal 3d shape generation and manipulation. In ECCV, 2022. 2 +[15] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In ECCV, 2022. 2 +[16] Angela Dai, Yawar Siddiqui, Justus Thies, Julien Valentin, and Matthias Nießner. Spsg: Self-supervised photometric scene generation from rgb-d scans. In CVPR, 2021. 2, 6, 7 +[17] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 2, 5 +[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2 +[19] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2 +[20] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 2 +[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NIPS, 2014. 2 +[22] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2 +[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2, 6 +[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2, 3 +[25] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. arXiv preprint arXiv:2106.15282, 2021. 2 +[26] Ronghang Hu and Amanpreet Singh. Unit: Multimodal multitask learning with a unified transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1439-1449, 2021. 2 +[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 2 +[28] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2 +[29] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fi + +dler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 2 +[30] Chieh Hubert Lin, Hsin-Ying Lee, Willi Menapace, Mengei Chai, Aliaksandr Siarohin, Ming-Hsuan Yang, and Sergey Tulyakov. *Infinicity: Infinite-scale city synthesis.* arXiv preprint arXiv:2301.09637, 2023. 2 +[31] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 4 +[32] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, 2021. 2 +[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, 2023. 2, 6, 7 +[34] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In CVPR, 2022. 2, 5, 6, 7 +[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2 +[36] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. AutoSDF: Shape priors for 3d completion, reconstruction and generation. In CVPR, 2022. 2 +[37] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia, 2022. 2, 6, 7 +[38] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. arXiv preprint arXiv:2212.01206, 2022. 1 +[39] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International conference on machine learning, pages 7220-7229. PMLR, 2020. 1 +[40] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2 +[41] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2 +[42] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In ICCV, 2019. 6, 7 +[43] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In CVPR, 2022. 2 +[44] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 5 +[45] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 2 +[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden + +hall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2 +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICLR, 2021. 2 +[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2 +[50] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 5 +[51] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2 +[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3 +[53] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. 2022. 2 +[54] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2 +[55] Aliaksandr Siarohin, Willi Menapace, Ivan Skorokhodov, Kyle Olszewski, Jian Ren, Hsin-Ying Lee, Mengei Chai, and Sergey Tulyakov. Unsupervised volumetric animation. In CVPR, 2023. 2 +[56] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In ECCV, 2022. 2, 5, 6, 7 +[57] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2 +[58] Ivan Skorokhodov, Aliaksandr Siarohin, Yinghao Xu, Jian Ren, Hsin-Ying Lee, Peter Wonka, and Sergey Tulyakov. 3d generation on imagenet. In ICLR, 2023. 2 +[59] Edward J Smith and David Meger. Improved adversarial systems for 3d object generation and reconstruction. In CoRL, 2017. 2 +[60] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 2 +[61] Jiajun Wu, Chengkai Zhang, Xiuming Zhang, Zhoutong Zhang, William T Freeman, and Joshua B Tenenbaum. Learning shape priors for single-view 3d completion and reconstruction. In Proceedings of the European Conference on + +Computer Vision (ECCV), pages 646-662, 2018. 1 +[62] Jianwen Xie, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Song-Chun Zhu, and Ying Nian Wu. Learning descriptor networks for 3d shape synthesis and analysis. In CVPR, 2018. 2 +[63] Yinghao Xu, Mengei Chai, Zifan Shi, Sida Peng, Ivan Skorokhodov, Aliaksandr Siarohin, Ceyuan Yang, Yujun Shen, Hsin-Ying Lee, Bolei Zhou, et al. Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In CVPR, 2023. 2 +[64] Linjie Yang, Ping Luo, Chen Change Loy, and Xiaou Tang. A large-scale car dataset for fine-grained categorization and verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3973-3981, 2015. 6 +[65] Rui Yu, Yue Dong, Pieter Peers, and Xin Tong. Learning texture generators for 3d shape collections from internet photo sets. In BMVC, 2021. 2, 6, 7 +[66] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 1 +[67] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2, 4 +[68] Song-Hai Zhang, Yuan-Chen Guo, and Qing-Wen Gu. Sketch2model: View-aware 3d modeling from single free-hand sketches. In CVPR, 2021. 2 \ No newline at end of file diff --git a/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/images.zip b/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cd5c994e7b3e05041c1c9dde94967f98102a0aa1 --- /dev/null +++ b/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91366d5b4dc7a9f9726c6bf77b4be1b28d73464a2c9fb620e13cb8177f31de4b +size 623743 diff --git a/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/layout.json b/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f198c3cbb2d711760fb4fef82e67d163196a0c2d --- /dev/null +++ b/2023/Text2Tex_ Text-driven Texture Synthesis via Diffusion Models/layout.json @@ -0,0 +1,8340 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 106, + 103, + 487, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 103, + 487, + 120 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 487, + 120 + ], + "type": "text", + "content": "Text2Tex: Text-driven Texture Synthesis via Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 143, + 153, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 143, + 153, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 143, + 153, + 157 + ], + "type": "text", + "content": "Dave Zhenyu Chen" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 166, + 143, + 246, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 143, + 246, + 157 + ], + "spans": [ + { + "bbox": [ + 166, + 143, + 246, + 157 + ], + "type": "text", + "content": "Yawar Siddiqui" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 260, + 143, + 337, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 143, + 337, + 157 + ], + "spans": [ + { + "bbox": [ + 260, + 143, + 337, + 157 + ], + "type": "text", + "content": "Hsin-Ying Lee2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 350, + 143, + 438, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 143, + 438, + 157 + ], + "spans": [ + { + "bbox": [ + 350, + 143, + 438, + 157 + ], + "type": "text", + "content": "Sergey Tulyakov2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 451, + 143, + 539, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 143, + 539, + 157 + ], + "spans": [ + { + "bbox": [ + 451, + 143, + 539, + 157 + ], + "type": "text", + "content": "Matthias Nießner1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 167, + 157, + 326, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 157, + 326, + 171 + ], + "spans": [ + { + "bbox": [ + 167, + 157, + 326, + 171 + ], + "type": "text", + "content": "1Technical University of Munich" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 350, + 158, + 429, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 158, + 429, + 171 + ], + "spans": [ + { + "bbox": [ + 350, + 158, + 429, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 350, + 158, + 429, + 171 + ], + "type": "text", + "content": "Snap Research" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 160, + 172, + 434, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 172, + 434, + 185 + ], + "spans": [ + { + "bbox": [ + 160, + 172, + 434, + 185 + ], + "type": "text", + "content": "https://daveredrum.github.io/Text2Tex/" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 53, + 210, + 544, + 375 + ], + "blocks": [ + { + "bbox": [ + 53, + 210, + 544, + 375 + ], + "lines": [ + { + "bbox": [ + 53, + 210, + 544, + 375 + ], + "spans": [ + { + "bbox": [ + 53, + 210, + 544, + 375 + ], + "type": "image", + "image_path": "325025c2a43659fdd8b91347ff83c17846c00c8bf1bf4ffc6466f79b4a5e8104.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 381, + 546, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 381, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 546, + 407 + ], + "type": "text", + "content": "Figure 1: We introduce Text2Tex, a text-driven architecture for 3D texture synthesis. Given object geometries and text prompts as input, Text2Tex generates high quality and consistent textures via depth-aware inpainting and refinement." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 416, + 192, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 416, + 192, + 428 + ], + "spans": [ + { + "bbox": [ + 143, + 416, + 192, + 428 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 441, + 290, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 441, + 290, + 634 + ], + "spans": [ + { + "bbox": [ + 45, + 441, + 290, + 634 + ], + "type": "text", + "content": "We present Text2Tex, a novel method for generating high-quality textures for 3D meshes from the given text prompts. Our method incorporates inpainting into a pre-trained depth-aware image diffusion model to progressively synthesize high resolution partial textures from multiple viewpoints. To avoid accumulating inconsistent and stretched artifacts across views, we dynamically segment the rendered view into a generation mask, which represents the generation status of each visible texel. This partitioned view representation guides the depth-aware inpainting model to generate and update partial textures for the corresponding regions. Furthermore, we propose an automatic view sequence generation scheme to determine the next best view for updating the partial texture. Extensive experiments demonstrate that our method significantly outperforms the existing text-driven approaches and GAN-based methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 645, + 127, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 127, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 127, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Generating high-quality 3D content is an essential component of visual applications in films, games, and upcoming AR/VR scenarios. With an increasing number of 3D content datasets, the computer vision community has wit-" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 417, + 546, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 546, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 546, + 513 + ], + "type": "text", + "content": "nessed significant progress in the field of 3D geometry generation [12, 39, 61, 66, 38]. Despite the remarkable success in modeling 3D geometries in recent years, fully automatic 3D content generation is still hindered by the laborious human efforts required to design textures. Therefore, automating the texture design process through alternative guidance, such as text, has become an intriguing but challenging research problem." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 517, + 546, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 546, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 546, + 649 + ], + "type": "text", + "content": "Recently, text-to-image generators have shown remarkable progress in the 2D domain leveraging diffusion model architectures, enabling high resolution 2D content generation based on textual descriptions [1, 48]. However, there are notable challenges for producing 3D textures via such 2D vision-language prior knowledge. Specifically, the synthesized textures are expected to be not only with high fidelity to the language cues, but also of high and consistent quality for target meshes. As such, previous attempts to paint 3D geometry from text inputs often fail to deliver well-textured 3D content." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we introduce Text2Tex, a novel texture synthesis method that seamlessly texturizes 3D objects using a pre-trained depth-aware text-to-image diffusion model. The method renders a target mesh from multiple viewpoints and inpaints the missing appearance with a" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18558" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 382 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 382 + ], + "type": "text", + "content": "depth-aware text-to-image diffusion model. Text2Tex follows a generate-then-refine strategy. Our method progressively generates partial textures across viewpoints and backprojects them to texture space. To address stretched and inconsistent artifacts observed from rotated viewpoints, we design a view partitioning technique that computes similarity maps between visible texel's normal vectors and the current view direction. The generation mask created from these similarity maps guides the diffusion process by indicating regions to generate, update, keep, or ignore. This allows us to apply different diffusion strengths to respective regions, inpainting missing appearance and updating stretched artifacts. However, the autoregressive generation process via the diffusion-based image inpainting model presents a new challenge. As the inpainting and updating scheme is conditioned on previously synthesized results, a viewpoint sequence with an ill-defined order or incomplete coverage over the mesh surface may result in unsatisfactory texturization. Therefore, we propose an automatic viewpoint selection technique that progressively selects the next best view. The confidence of each candidate view containing the biggest relative area for generation and updating is estimated, given the partially textured mesh. This approach ensures complete coverage over the mesh surface and a high-quality texture map by consistently updating stretched regions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 384, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 289, + 479 + ], + "type": "text", + "content": "We demonstrate the effectiveness of Text2Tex for synthesizing high-quality 3D textures from language cues. The proposed method performs favorably against other language-based texture synthesis methods in terms of FID [23], KID [4], and user study on a subset of the Objaverse dataset [17]. Additionally, our method also outperforms category-specific GAN-based methods on the ShapeNet car dataset [7]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 479, + 286, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 479, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 59, + 479, + 286, + 491 + ], + "type": "text", + "content": "To summarize, our technical contributions are threefold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 500, + 286, + 623 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 500, + 286, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 500, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 59, + 500, + 286, + 536 + ], + "type": "text", + "content": "- We design a novel method for high-quality texture synthesis by progressively inpainting and updating the 3D textures via depth-aware diffusion models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 544, + 286, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 544, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 59, + 544, + 286, + 579 + ], + "type": "text", + "content": "- We propose an automatic view sequence generation scheme to dynamically determine the order for generating and updating the texture space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 587, + 286, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 587, + 286, + 623 + ], + "spans": [ + { + "bbox": [ + 59, + 587, + 286, + 623 + ], + "type": "text", + "content": "- We conduct extensive study on a considerable amount of 3D objects, demonstrating the proposed method is effective for large-scale 3D content generation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 634, + 132, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 132, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 132, + 647 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 715 + ], + "type": "text", + "content": "3D Generation from 3D and 2D data. To achieve 3D generation, it is natural to train models directly on 3D data. In contrast to 2D images, there are several 3D representations available, each with its unique characteristics, leading to the development of various generative models such as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "those based on voxels [30, 55, 59, 62], point clouds [3, 32], meshes [68], signed distance function [12, 13, 14, 16, 36], etc. However, unlike images or videos that are ubiquitous, 3D data is inherently scarce and challenging to collect and annotate. Consequently, the synthesized samples from 3D generative models, trained on 3D data, are of limited quality and diversity, in terms of both structure and texture. Recent works have leveraged differentiable rendering to learn texture generation using only 2D images [20, 56, 65]. However, they are typically trained for specific shape categories and struggle in the quality of textures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 207, + 546, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 546, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 546, + 471 + ], + "type": "text", + "content": "Text-Guided Generation. Recently, there has been tremendous progress in the vision-language domain [26, 57, 60, 47, 10, 8, 11, 9]. Specifically, the emergence of Contrastive Language-Image Pre-Training (CLIP) [47] has enabled the development of text-guided image generation through its semantically rich representation trained on text-image pairs. Initial efforts [15, 45] incorporated CLIP with different backbones, such as StyleGAN [27] and VQGAN [19]. However, diffusion models [18, 24, 25, 40, 53], which have gained attention due to their superior visual quality and training stability compared to Generative Adversarial Networks (GANs) [21], have recently been trained on large-scale text-image datasets with CLIP encodings [28, 49, 52]. Among these models, Stable Diffusion [1, 52] has garnered significant interest as an open-sourced model with numerous extensions [1, 67] that support different conditional modalities in addition to text prompts, including depth images, poses, sketches, etc. Additionally, CLIP has also been adopted in 3D to perform text-guided shape and texture generation [34, 37]. In this work, we take advantage of the depth-conditioning feature of Stable Diffusion to provide more consistent texturing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": "Text-to-3D from 2D data. Inspired by the success of Neural Radiance Fields (NeRF) [35], NeRF-based generators [2, 5, 6, 22, 41, 43, 54, 58, 63] have been proposed to learn 3D structures from 2D images using GAN-based frameworks. A new research direction emerged by combining NeRF techniques with flourishing diffusion-based text-to-image models, enabling text-to-3D learning with only 2D supervision. To address the challenge of optimizing a NeRF field, a score-distillation loss is proposed [46] that leverages a pretrained 2D diffusion model as a critic to provide essential gradients. Subsequent efforts have focused on adopting this loss in latent space [13, 33] and in a coarse-to-fine refinement approach [29]. However, optimization-based methods are plagued by long convergence times. A recent concurrent work [51] proposes a non-optimization approach with progressive updates from multiple pre-set viewpoints. In contrast, our method iteratively updates and refines the synthesized textures from automatically selected viewpoint sequences, which minimizes human efforts with designing different viewpoint orders for various geometries." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18559" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 399, + 235 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 399, + 235 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 399, + 235 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 399, + 235 + ], + "type": "image", + "image_path": "af9726b05d572e3372214fe3a3f4d751977d8c7bca91eab3a69ad543c0bf55a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 246, + 546, + 318 + ], + "lines": [ + { + "bbox": [ + 46, + 246, + 546, + 318 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 546, + 318 + ], + "type": "text", + "content": "Figure 2: Overview of Text2Tex. We illustrate the pipeline using a 3D car mesh with a prompt \"golden Porsche\". We progressively generate the texture via a generate-then-refine scheme. In progressive texture generation (Sec. 3.3), we start by rendering the object from an initial preset viewpoint. We generate a new appearance according to the input prompt via a depth-to-image diffusion model, and project the generated image back to the partial texture. Then, we repeat this process until the last preset viewpoint to output the initial textured mesh. In the subsequent texture refinement (Sec. 3.4), we update the initial texture from a sequence of automatically selected viewpoints to refine the stretched and blurry artifacts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 399, + 71, + 541, + 235 + ], + "blocks": [ + { + "bbox": [ + 399, + 71, + 541, + 235 + ], + "lines": [ + { + "bbox": [ + 399, + 71, + 541, + 235 + ], + "spans": [ + { + "bbox": [ + 399, + 71, + 541, + 235 + ], + "type": "image", + "image_path": "fc07aaf42c347ed25813041dfa6a9bf54aab089619c2a563a6a4fe381eccc68e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 338, + 102, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 102, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 102, + 350 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 358, + 290, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 290, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 290, + 539 + ], + "type": "text", + "content": "The objective of our work is to texture a 3D mesh using a pretrained text-to-image diffusion model. In this section, we begin by laying the foundation of the diffusion model in Sec. 3.1 and depth-aware inpainting model in Sec. 3.2. We then propose a generate-then-refine scheme for progressively synthesizing and updating the 3D textures in a coarse-to-fine fashion. In the progressive texture generation (Sec. 3.3), we paint the visible regions of the input geometry in an incremental fashion, following a sequence of predefined viewpoints. To ensure local and global consistency, we incorporate a view partition to guide the depth-aware inpainting objectives. Subsequently, we introduce an automatic viewpoint selection mechanism (Sec. 3.4) to perform texture refinement and address any issues of texture stretching and inconsistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 548, + 126, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 126, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 126, + 560 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "content": "We use a Denoising Diffusion Probabilistic Model (DDPM) [24] as the generative model. Specifically, to avoid high computational overhead, we adopt the latent diffusion model [52], where an input image " + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "content": " is first encoded into latent code " + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "content": " before the diffusion process. The forward pass follows a Markov Chain to gradually add noise to the input latent code " + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "content": " towards the white Gaussian noise " + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "content": ". At each step in the forward pass, the noised latent code " + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "content": " is obtained by adding a noise variance " + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "content": " to the previous latent code " + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "content": " scaled with " + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\sqrt{1 - \\beta_t}" + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 108, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "z _ {t} \\sim \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right). \\tag {1}", + "image_path": "608b59c0ed145f98ad4dc99189ec729abf4c728cfb7ad5f8d2a1a834436d23ef.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 339, + 546, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 339, + 546, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 339, + 546, + 374 + ], + "type": "text", + "content": "The independence property enables direct transformation of the noised latent code " + }, + { + "bbox": [ + 305, + 339, + 546, + 374 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 305, + 339, + 546, + 374 + ], + "type": "text", + "content": " at an arbitrary time step " + }, + { + "bbox": [ + 305, + 339, + 546, + 374 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 339, + 546, + 374 + ], + "type": "text", + "content": " from the input latent " + }, + { + "bbox": [ + 305, + 339, + 546, + 374 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 305, + 339, + 546, + 374 + ], + "type": "text", + "content": " via:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 369, + 384, + 545, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 384, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 369, + 384, + 545, + 398 + ], + "type": "interline_equation", + "content": "z _ {t} \\sim \\mathcal {N} (\\sqrt {\\bar {a}} _ {t} z _ {0}, (1 - \\bar {a} _ {t}) \\mathbf {I}), \\tag {2}", + "image_path": "4dab2e98a7e4605d83de6787cfca109ab8125dbaa7810d869bda24afa90ff277.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\bar{a}_t" + }, + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "text", + "content": " is the total noise variance, which can be calculated by " + }, + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\sum_{t = i}^{T}(1 - \\beta_{t})" + }, + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "text", + "content": " from the noise " + }, + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "text", + "content": " added to the input latent code " + }, + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 305, + 407, + 545, + 443 + ], + "type": "text", + "content": " at each time step." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 444, + 546, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 444, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 444, + 546, + 479 + ], + "type": "text", + "content": "During inference, the latent estimation " + }, + { + "bbox": [ + 305, + 444, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t - 1}" + }, + { + "bbox": [ + 305, + 444, + 546, + 479 + ], + "type": "text", + "content": " for the next time step " + }, + { + "bbox": [ + 305, + 444, + 546, + 479 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 305, + 444, + 546, + 479 + ], + "type": "text", + "content": " is obtained by predicting " + }, + { + "bbox": [ + 305, + 444, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(z_t,t)" + }, + { + "bbox": [ + 305, + 444, + 546, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 444, + 546, + 479 + ], + "type": "inline_equation", + "content": "\\sigma_{\\theta}(z_t,t)" + }, + { + "bbox": [ + 305, + 444, + 546, + 479 + ], + "type": "text", + "content": " of a Gaussian distribution:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 363, + 491, + 545, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 491, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 363, + 491, + 545, + 506 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {t - 1} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(z _ {t}, t\\right), \\sigma_ {\\theta} \\left(z _ {t}, t\\right)\\right) \\tag {3}", + "image_path": "a4d24bcd10fccf5389f87516ed68ec5f1036814cc449fa16c02741529d4354b7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "text", + "content": "Denoising strength. To prevent complete randomness during the diffusion process, we introduce a scaling factor " + }, + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\gamma, 0 < \\gamma \\leq 1" + }, + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "text", + "content": ", which controls the number of diffusion steps. We assume that a white Gaussian noise " + }, + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "text", + "content": " can be obtained by adding noise to the input latent code " + }, + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "text", + "content": " steps, and the final denoised latent estimation " + }, + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "text", + "content": " is fully governed by the pure noise. By applying the scaling factor, we can start denoising the latent code at time step " + }, + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\gamma T" + }, + { + "bbox": [ + 305, + 519, + 545, + 651 + ], + "type": "text", + "content": " to guide the final latent code with the original image information. This technique is applied to refine the previously generated image contents." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 658, + 474, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 474, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 474, + 672 + ], + "type": "text", + "content": "3.2. Depth-Aware Image Inpainting" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "The core of the texture synthesis lies in painting the missing regions on the mesh surface. The generated texture is expected to be highly faithful to the mesh geometry and the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18560" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 289, + 310 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 289, + 310 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 289, + 310 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 289, + 310 + ], + "type": "image", + "image_path": "a3f7c5976aa72fad20aaf2fed00b8ed53e16d6d77b6963f2210e064ae2fd7267.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 316, + 290, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 290, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 290, + 401 + ], + "type": "text", + "content": "Figure 3: We dynamically partition the current view into a generation mask to guide the depth-aware inpainting model. For the \"new\" region, we denoise the new object appearance from white Gaussian noise. For the \"update\" region, we refine the previous texture by denoising the partially noised image segments. We freeze the texture in the \"keep\" region for this view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 423, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 519 + ], + "type": "text", + "content": "input text. To achieve this, we build our method on a pretrained depth-to-image model [1, 67] that can produce high-quality images from text while being consistent with depth cues. However, since the Depth2Image model is designed to generate entire images, we need to use an inpainting mask to guide the sampling process. This mask provides explicit hints of which regions to generate or keep fixed, similar to the denoising guidance strategy in RePaint [31]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 519, + 288, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 288, + 568 + ], + "type": "text", + "content": "To condition the denoising on the known regions of the input, we inject a generation mask " + }, + { + "bbox": [ + 46, + 519, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 519, + 288, + 568 + ], + "type": "text", + "content": " into the sampling steps. This mask explicitly blends the noised latent code " + }, + { + "bbox": [ + 46, + 519, + 288, + 568 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 519, + 288, + 568 + ], + "type": "text", + "content": " and the denoised latent estimation " + }, + { + "bbox": [ + 46, + 519, + 288, + 568 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 46, + 519, + 288, + 568 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 582, + 287, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 582, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 102, + 582, + 287, + 594 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {t} = \\hat {z} _ {t} \\odot \\mathcal {M} + z _ {t} \\odot (1 - \\mathcal {M}). \\tag {4}", + "image_path": "52d0a27d22f074abe3231cad50fd7c42e14e2550f7e35303d1317a73b0dc9acc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 601, + 288, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 288, + 626 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 288, + 626 + ], + "type": "text", + "content": "We then decode the final denoised latent estimation " + }, + { + "bbox": [ + 47, + 601, + 288, + 626 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 47, + 601, + 288, + 626 + ], + "type": "text", + "content": " to the final output image " + }, + { + "bbox": [ + 47, + 601, + 288, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 47, + 601, + 288, + 626 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 635, + 219, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 219, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 219, + 647 + ], + "type": "text", + "content": "3.3. Progressive Texture Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "With the customized Depth2Image model, we are able to paint the object with a high quality image from a single view. To synthesize the appearance of the input geometry, we project the generated 2D views onto the texture space of a normalized 3D object with proper UV parameterization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "Assuming Y-axis as the up-axis in the world coordinate system, we define the viewpoint as " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "v = (\\theta, \\phi, r)" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " is the azimuth angle with respect to the Z-axis, " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " is the viewpoint elevation angle with respect to the XZ-plane, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " is the distance between the viewpoint and the origin." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "text", + "content": "As shown in Fig. 2, we start by generating the visible but missing texture in an initial viewpoint " + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "inline_equation", + "content": "v^0" + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "text", + "content": ". We render the object to a depth map " + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^0" + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "text", + "content": " and a generation mask " + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^0" + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "text", + "content": ", and then use a customized Depth2Image diffusion model with " + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^0" + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "text", + "content": " as input and " + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^0" + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "text", + "content": " as extra guidance to generate a colored image " + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{O}^0" + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "text", + "content": ". We then back-project the image " + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{O}^0" + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "text", + "content": " to the visible part of the texture " + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^0" + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "text", + "content": ". In the subsequent steps, we progressively diffuse the colored images " + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{O}^k" + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "text", + "content": " and back-project them to the texture " + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^k" + }, + { + "bbox": [ + 304, + 133, + 546, + 240 + ], + "type": "text", + "content": " through a sequence of viewpoints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 240, + 547, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 547, + 326 + ], + "type": "text", + "content": "We notice that directly inpainting the missing regions on mesh surface often results in inconsistency issues. The issue is mainly caused by the stretched artifacts that occur when the 2D views are projected back onto the curved surface of the mesh. Therefore, we design a dynamic view partitioning strategy to guide the inpainting process with respective generation objectives " + }, + { + "bbox": [ + 304, + 240, + 547, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 240, + 547, + 326 + ], + "type": "text", + "content": " and denoising strengths " + }, + { + "bbox": [ + 304, + 240, + 547, + 326 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 240, + 547, + 326 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 342, + 546, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 546, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 546, + 473 + ], + "type": "text", + "content": "Dynamic view partitioning For all viewpoints " + }, + { + "bbox": [ + 304, + 342, + 546, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = v_{i}, i = 1, \\dots, N" + }, + { + "bbox": [ + 304, + 342, + 546, + 473 + ], + "type": "text", + "content": ", we render the similarity mask " + }, + { + "bbox": [ + 304, + 342, + 546, + 473 + ], + "type": "inline_equation", + "content": "S^i" + }, + { + "bbox": [ + 304, + 342, + 546, + 473 + ], + "type": "text", + "content": " for each viewpoint " + }, + { + "bbox": [ + 304, + 342, + 546, + 473 + ], + "type": "inline_equation", + "content": "v^k" + }, + { + "bbox": [ + 304, + 342, + 546, + 473 + ], + "type": "text", + "content": " and map those values to the texture space. Each pixel in a similarity mask represents the reversed normalized value of the cosine similarity between the normal vectors of the visible faces and the view direction (ranging from 0 to 1). A pixel with a value of 1 indicates that the corresponding face is perpendicular to the view direction. For simplicity, we set the background to 0. In summary, these masks indicate the extent to which a face is rotated away from the viewpoint." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "text", + "content": "Based on the similarity mask " + }, + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "inline_equation", + "content": "S^k" + }, + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "text", + "content": ", we segment the rendered view into a generation mask " + }, + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^k" + }, + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "text", + "content": ", including the following 4 regions, as shown in Fig. 3: 1) New: This region contains pixels that have not yet been textured. We inpaint this region from pure white Gaussian noise, i.e. with denoising strength 1. 2) Update: This region contains pixels that have been textured, but the corresponding similarity score in " + }, + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "inline_equation", + "content": "S^k" + }, + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "text", + "content": " is greater than all other views. This indicates that those pixels are being observed in a better angle. Therefore, we update this region with a moderate denoising strength " + }, + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "inline_equation", + "content": "\\gamma_g" + }, + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "text", + "content": " to avoid stretched appearance. 3) Keep: Pixels in this region have been textured, but the corresponding similarity score in " + }, + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "inline_equation", + "content": "S^k" + }, + { + "bbox": [ + 304, + 474, + 547, + 689 + ], + "type": "text", + "content": " is not the highest among all other views. These pixels have already been observed from a better angle, so we keep them fixed. 4) Ignore: This region contains pixels that belong to the background and are irrelevant to the process, so we ignore them throughout the entire process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": "While the generation mask helps guide the texture inpainting process with accurate generation objectives and ap" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18561" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 294 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 294 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 294 + ], + "type": "image", + "image_path": "de0fb4cfc55b75d280ee7caa5fb9247df42a3bd65b283bf5809b5cc1503bf40c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 301, + 287, + 387 + ], + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 387 + ], + "type": "text", + "content": "Figure 4: In the refinement stage, the sequence of viewpoints are automatically determined by selecting the viewpoint with the maximum normalized area of the \"update\" region at each step. We update 2D views in the \"update\" regions with a moderate diffusion denoising strength. The updated object appearance is then back-projected to the texture space at the end of each refinement step." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 408, + 287, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 494 + ], + "type": "text", + "content": "propriate denoising strengths, blurriness and stretches can still exist on the mesh surface. This is because the generation mask is limited to a predefined set of viewpoints, and the seams and stretches on the texture are still visible from a novel viewpoint. To address this issue, we propose a texture refinement technique with an automatic viewpoint selection strategy, which is described in the next section." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 502, + 287, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 287, + 526 + ], + "type": "text", + "content": "3.4. Texture Refinement with Automatic Viewpoint Selection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 533, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 677 + ], + "type": "text", + "content": "To remove the synthesis artifacts, a straightforward solution is to increase the number of viewpoints. However, the optimal viewpoint sequence can vary for different object geometries, making it difficult to manually pre-set the viewpoint sequence for massive synthesis targets. To address these challenges, we propose an automatic viewpoint selection strategy that effectively prevents stretches and seams, as illustrated in Fig. 4. We densely define a set of refinement viewpoints " + }, + { + "bbox": [ + 46, + 533, + 288, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = v_{i}, i = 1,\\dots,K" + }, + { + "bbox": [ + 46, + 533, + 288, + 677 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 533, + 288, + 677 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 533, + 288, + 677 + ], + "type": "text", + "content": " is larger than " + }, + { + "bbox": [ + 46, + 533, + 288, + 677 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 533, + 288, + 677 + ], + "type": "text", + "content": ". To distribute the refinement viewpoints evenly, we scatter them on a hemisphere, taking into account that objects are rarely observed from the bottom-up view." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "Assuming that an initial texture has been applied to the object, the refinement process begins by segmenting the generation masks " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " using the similarity masks " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " from all" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "text", + "content": "available viewpoints. For each of the " + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "text", + "content": " refinement viewpoints in " + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "text", + "content": ", we calculate a view heat " + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "inline_equation", + "content": "h^i" + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "text", + "content": " from the corresponding generation mask " + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "text", + "content": ", which represents the normalized area of the \"update\" region with respect to the current visible area of the object. The viewpoint " + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "text", + "content": " that maximizes the view heat is then selected by " + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\arg \\max_i h_i = \\frac{1}{N_p} \\sum_{k=1}^{N_p} w_k" + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "text", + "content": " is the total number of the non-background pixels, and " + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "inline_equation", + "content": "w_k" + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "text", + "content": " is the scaling factor for the segments in the generation mask. In order to let views with relatively more areas for updating, " + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "inline_equation", + "content": "w_k" + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "text", + "content": " for the \"update\" region is set to be bigger than that for the \"keep\" region. We dynamically select the next view with the highest view heat for updating. To avoid conflicts with the previously generated textures, we update the \"update\" regions with a mild denoising strength " + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\gamma_r" + }, + { + "bbox": [ + 304, + 72, + 547, + 266 + ], + "type": "text", + "content": ", which preserves the original appearance cues." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 277, + 359, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 359, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 359, + 289 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 297, + 439, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 439, + 309 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 439, + 309 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 314, + 547, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 547, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 547, + 435 + ], + "type": "text", + "content": "We apply the Depth2Image model from Stable Diffusion v2 [1] as our generation backbone. The denoising strength " + }, + { + "bbox": [ + 304, + 314, + 547, + 435 + ], + "type": "inline_equation", + "content": "\\gamma_{g}" + }, + { + "bbox": [ + 304, + 314, + 547, + 435 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 314, + 547, + 435 + ], + "type": "inline_equation", + "content": "\\gamma_{r}" + }, + { + "bbox": [ + 304, + 314, + 547, + 435 + ], + "type": "text", + "content": " are set as 0.5 and 0.3 for the generation and refinement stages, respectively. We define 6 axis-aligned principles viewpoints for generation, and in total 36 viewpoints for refinement, among which we dynamically select only 20 views to reduce time cost. Each synthesis process takes around 15 minutes to complete on an NVIDIA RTX A6000. Our implementation uses the PyTorch [44] framework with PyTorch3D [50] used for rendering and texture projection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 441, + 414, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 414, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 414, + 454 + ], + "type": "text", + "content": "4.2. Experiment Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 640 + ], + "type": "text", + "content": "Data. We evaluate our method on a subset of textured meshes from the Objaverse [17] dataset. We first sample 3 random meshes from each category. To ensure the quality of the input meshes, we manually filter out thin or unrecognizable meshes, such as \"strainer\", \"sweatpants\", and \"legging\", meshes with too simplistic textures, and meshes that do not correspond with their assigned categories. For the purpose of reducing processing time, we also remove overtriangulated and scanned objects. After this curation, there are in total 410 high quality textured meshes across 225 categories for the experiments. Note that the original textures are only used for the evaluation. To compare with GAN-based category-specific approaches, we also report results on the \"car\" objects from the ShapeNet dataset [7]. In particular, we use the 300 meshes from the test set used in [56]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Baselines. We compare our method against the following state-of-the-art text-driven texture synthesis method: 1) Text2Mesh [34], a neural pipeline that directly optimizes the textures and geometries via a CLIP-based optimization objective. We remove the displacement prediction so that" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18562" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 545, + 371 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 545, + 371 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 545, + 371 + ], + "type": "image", + "image_path": "9eb23aece4df7bb0df9d72d4d6ae6a5bd1bac0f9ce28267c2238b96121a089ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 377, + 546, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 377, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 546, + 415 + ], + "type": "text", + "content": "Figure 5: Qualitative comparisons on Objaverse. We compare our textured mesh against CLIPMesh [37], Text2Mesh [34], Latent-Paint [33], and the original textures from Objaverse. In comparison with the baselines, our method produces more consistent and detailed 3D textures with respect to the input geometries. Image best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 434, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 289, + 579 + ], + "type": "text", + "content": "only the surface RGBs are optimized. 2) CLIPMesh [37], a CLIP-based pipeline that deforms a sphere and optimizes the surface RGBs. Similar to Text2Mesh, we remove the shape deformation branch, and the texture colors are directly optimized on the surface of a given geometry. 3) Latent-Paint [33], a texture generation variant of the NeRF-based 3D object generation pipeline Latent-NeRF [33]. It explicitly operates on a given texture map using Stable Diffusion as a prior. In addition to text-guided methods, we also compare with category-specific GAN-based approaches, including Texture Fields [42], SPSG [16], LTG [65], and Texturify [56]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate the generated textures via commonly used image quality and diversity metrics for generative models. Specifically, we report the Frechet Inception Distance (FID) [23] and Kernel Inception Distance " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\mathrm{KID} \\times 10^{-3})" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": " [4]. The generated image distribution for these metrics consists of renders of each mesh with the synthesized textures from 20 fixed viewpoints at a resolution of " + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": ". For experiments on Objaverse dataset, the real distribution comprises renders of the meshes with the same settings using their artist designed textures. For ex" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 434, + 546, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 546, + 460 + ], + "type": "text", + "content": "periments on ShapeNet cars, we use the 18991 background segmented images from CompCars dataset [64]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 476, + 422, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 422, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 422, + 489 + ], + "type": "text", + "content": "4.3. Quantitative results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": "In Tab. 1, we compare our method against the previous SOTA text-driven texture synthesis methods on Objverse objects. As input, we uniformly feed template texts \"a (category)\" to the models. Quantitatively, our method outperforms all baselines by a significant margin (19% improvement in FID and 26% improvement in KID). Such improvements demonstrate that our method is more capable of generating more realistic textures on various object geometries from numerous categories. To demonstrate the effectiveness of our method against the GAN-based approaches on category-specific data, we report experiment results on ShapeNet \"car\" category in Tab. 2. Notably, our method achieves superior performance over the previous GAN-based SOTA texture synthesis method Texturify, improving by 21% in FID and 12% in KID. This indicates our method is more effective with synthesizing realistic textures than GAN based approaches that were trained on specific categories." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18563" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 68, + 287, + 190 + ], + "blocks": [ + { + "bbox": [ + 49, + 68, + 287, + 190 + ], + "lines": [ + { + "bbox": [ + 49, + 68, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 49, + 68, + 287, + 190 + ], + "type": "image", + "image_path": "ecad2569c5931d313c99782523227924b57d97a57fb94380dce4e7b9f4648dd1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 288, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 288, + 245 + ], + "type": "text", + "content": "Figure 6: Qualitative comparisons on ShapeNet car. Our method generates sharper and more coherent textures with respect to the geometries compared to the state-of-the-art GAN-based method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 257, + 273, + 335 + ], + "blocks": [ + { + "bbox": [ + 61, + 257, + 273, + 335 + ], + "lines": [ + { + "bbox": [ + 61, + 257, + 273, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 257, + 273, + 335 + ], + "type": "table", + "html": "
MethodFID ↓KID (x10-3) ↓
Text2Mesh [34]45.38(+9.7)10.40(+2.7)
CLIPMesh [37]43.25(+7.6)12.52(+4.8)
Latent-Paint [33]43.87(+8.1)11.43(+3.7)
Text2Tex (Ours)35.687.74
", + "image_path": "534063c2b066a3510ab32f6b7fcc81b2cb01f772daeaa4379bcd950cb389b99a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 400, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 591 + ], + "type": "text", + "content": "User study. We conduct user study to analyze the quality of the synthesized textures and their fidelity to the input text prompts. For each baseline method, we randomly show the users 5 pair of renders from the baseline and our method. The users are requested to choose the one that is more realistic and closer to the text prompts. More details of the questionnaire can be found in the supplemental. In the end, we receive 604 responses across 41 users. The collected preferences are reported in Tab. 3. In comparison to CLIPMesh and Text2Mesh, our method is clearly preferred by the users with preference rate " + }, + { + "bbox": [ + 46, + 400, + 287, + 591 + ], + "type": "inline_equation", + "content": "83.92\\%" + }, + { + "bbox": [ + 46, + 400, + 287, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 400, + 287, + 591 + ], + "type": "inline_equation", + "content": "76.47\\%" + }, + { + "bbox": [ + 46, + 400, + 287, + 591 + ], + "type": "text", + "content": ", respectively. Besides that, more users " + }, + { + "bbox": [ + 46, + 400, + 287, + 591 + ], + "type": "inline_equation", + "content": "(64.18\\%)" + }, + { + "bbox": [ + 46, + 400, + 287, + 591 + ], + "type": "text", + "content": " lean towards our method over the competitive baseline Latent-Paint. As can be seen, our method demonstrates the effectiveness in generating high quality textures that are favored by human users." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 599, + 162, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 162, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 162, + 612 + ], + "type": "text", + "content": "4.4. Qualitative analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 286, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 286, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 286, + 715 + ], + "type": "text", + "content": "We compare our qualitative results on Objaverse objects against text-driven baselines in Fig. 5. In comparison with CLIP-based methods CLIPMesh and Text2Mesh, our method generates more realistic and consistent textures. In particular, CLIPMesh generates sketchy textures while Text2Mesh produces repetitive patterns. Latent-Paint outputs a consistent texture capturing the semantics of the object well, but the results are often quite blurry. Clearly," + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 542, + 160 + ], + "blocks": [ + { + "bbox": [ + 46, + 343, + 287, + 380 + ], + "lines": [ + { + "bbox": [ + 46, + 343, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 287, + 380 + ], + "type": "text", + "content": "Table 1: Quantitative comparisons on Objverse subset. Our method performs favorably against state-of-the-art text-driven texture synthesis methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 70, + 542, + 160 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 542, + 160 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 542, + 160 + ], + "type": "table", + "html": "
MethodFID ↓KID (x10-3) ↓
Texture Fields [42]177.15(+130.2)17.14(+12.8)
SPSG [16]110.65(+63.7)9.59(+5.2)
LTG [65]70.76(+23.8)5.72(+1.4)
Texturify [56]59.55(+12.6)4.97(+0.6)
Text2Tex (Ours)46.914.35
", + "image_path": "faa529ab6afd557413d8781bbe996252ed90885d96e4fb2e84e5d9326f1689bf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 218, + 542, + 256 + ], + "blocks": [ + { + "bbox": [ + 305, + 168, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 305, + 168, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 168, + 545, + 205 + ], + "type": "text", + "content": "Table 2: Quantitative comparison on the ShapeNet cars. Our method outperform state-of-the-art category-specific GAN-based methods by a significant margin." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 218, + 542, + 256 + ], + "lines": [ + { + "bbox": [ + 310, + 218, + 542, + 256 + ], + "spans": [ + { + "bbox": [ + 310, + 218, + 542, + 256 + ], + "type": "table", + "html": "
CLIPMesh ↑Text2Mesh ↑Latent-Paint ↑
Ours83.92%76.47%64.18%
", + "image_path": "c0bd159419ae4b1f00ba3939405c89576f294ec6f909be03ff15b9976aee5644.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 325, + 544, + 375 + ], + "blocks": [ + { + "bbox": [ + 305, + 263, + 545, + 311 + ], + "lines": [ + { + "bbox": [ + 305, + 263, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 545, + 311 + ], + "type": "text", + "content": "Table 3: Percentage of users who prefer our method over the baselines in a user study with 604 responses across 41 participants. Our method is shown to be more favored by human users." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 325, + 544, + 375 + ], + "lines": [ + { + "bbox": [ + 308, + 325, + 544, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 325, + 544, + 375 + ], + "type": "table", + "html": "
w/ Depth2Imgw/ inpaintingw/ updateFID ↓KID (x10-3) ↓
✓XX39.889.78
✓✓X38.199.11
✓✓✓37.098.78
", + "image_path": "b4c9585749651ad51cacee25b38c43e2fe9d6624bf0fff72d56a8aea3e5fc398.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 383, + 545, + 431 + ], + "lines": [ + { + "bbox": [ + 305, + 383, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 383, + 545, + 431 + ], + "type": "text", + "content": "Table 4: Effect of components in the generation stage. We quantitatively ablate the efficacy of each component. Applying the inpainting and update scheme effectively improves the quality of the synthesized textures." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 455, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 539 + ], + "type": "text", + "content": "our method can synthesize more consistent textures with cleaner and richer local details. We also compare our textures with GAN-based generation approach on category-specific objectives. In Fig. 6, we show the textures for ShapeNet cars of our methods and Texturify. Notably, our textures have a much cleaner appearance and provide more details with respect to the input geometries." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 550, + 404, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 404, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 404, + 562 + ], + "type": "text", + "content": "4.5. Ablation studies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Does depth-aware inpainting and updating help? We show in Fig. 7 that the depth-aware inpainting is essential for producing high quality textures. In particular, the plain Depth2Img model often struggles to produce consistent appearances due to the governance of random noise. When the inpainting scheme is applied, the produced textures are more consistent. However, the textures still appear to be stretched and blurry over the curved mesh surface. These artifacts are amended by the texture updating scheme from the better viewing angles. The effectiveness of the depth-aware inpainting and the updating scheme is reflected in the improved FID and KID scores in Tab. 4." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18564" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 289, + 328 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 289, + 328 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 289, + 328 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 289, + 328 + ], + "type": "image", + "image_path": "707f8f03a26e274ccb1740e84ce734a092493b02293f3562d38b86f2deb955ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 337, + 288, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 337, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 288, + 384 + ], + "type": "text", + "content": "Figure 7: The proposed inpainting and update technique can effectively generate more consistent textures and eliminate stretched and blurry artifacts. Note that there is no refinement in this ablation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 396, + 285, + 443 + ], + "blocks": [ + { + "bbox": [ + 50, + 396, + 285, + 443 + ], + "lines": [ + { + "bbox": [ + 50, + 396, + 285, + 443 + ], + "spans": [ + { + "bbox": [ + 50, + 396, + 285, + 443 + ], + "type": "table", + "html": "
# views05101520
↓FID37.0936.6736.3935.9835.68
↓KID (x10-3)8.788.318.127.987.74
", + "image_path": "ad3c7772e108a566e034def3cda0ccae91999c0e825cfd74c9b0706449469ea7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 451, + 287, + 500 + ], + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 500 + ], + "type": "text", + "content": "Table 5: We quantitatively study the effect of selecting different number of viewpoints in the refinement stage. Refining the synthesis results with more viewpoints improves the texture quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 508, + 288, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 288, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 288, + 640 + ], + "type": "text", + "content": "Does viewpoint selection in refinement stage help? We compare our results with different refinement settings in Tab. 5. When the input geometries are painted with initial textures from the generation stage, the blurry artifacts and projection seams are usually not eliminated due to a limited number of viewpoints. As can be seen in Fig. 8, such flaws can be minimized by refining with more viewpoints. We also showcase the effectiveness of the automatic viewpoint selection technique, as the refinement process does not require any manual efforts with defining and fine-tuning the viewpoint sequence for different shapes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 647, + 127, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 127, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 127, + 659 + ], + "type": "text", + "content": "4.6. Limitations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "While our method has shown the capability to produce high-quality 3D textures, we have observed that it tends to produce textures with shading effects from the diffusion backbone. Although this issue can be addressed by care" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 545, + 329 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 545, + 329 + ], + "type": "image", + "image_path": "9d69f53354c4c876204d8c3cf5fd072b538fabc551fb18558616d3ddc27129f6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 338, + 546, + 375 + ], + "lines": [ + { + "bbox": [ + 305, + 338, + 546, + 375 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 546, + 375 + ], + "type": "text", + "content": "Figure 8: The proposed automatic viewpoint selection method further improves the texture quality by gradually removing the remaining artifacts from the generation stage." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 389, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 389, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 389, + 545, + 462 + ], + "type": "text", + "content": "fully fine-tuning the input prompts, doing so requires additional human engineering effort and may not scale well to massive generation targets. One potential solution is to fine-tune the diffusion model to remove the shading from textures. We acknowledge this challenge and leave it to future work to explore this possibility." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 476, + 378, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 378, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 378, + 489 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we present a novel method, Text2Tex, for synthesizing high quality textures for 3D meshes from the given text prompts. Our approach leverages a depth-aware image inpainting diffusion model to progressively generate high-resolution partial textures from multiple viewpoints. To avoid accumulating inconsistent and stretching artifacts across viewpoints, we dynamically segment the rendered view into a generation mask, which effectively guides the diffusion model to generate and update the corresponding partial textures. Furthermore, we propose an automatic viewpoint sequence generation scheme that utilizes the generation mask to automatically determine the next best view for refining the generated textures. Extensive experiments demonstrate that our method can effectively synthesize consistent and highly detailed 3D textures for various object geometries without extra manual effort. Overall, we hope our work can inspire more future research in the area of text-to-3D synthesis." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18565" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 149, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 149, + 85 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 149, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 92, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 92, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 47, + 92, + 287, + 163 + ], + "type": "text", + "content": "This work was supported by the ERC Starting Grant Scan2CAD (804724) and the German Research Foundation (DFG) Research Unit \"Learning and Simulation in Visual Computing\". We further thank Guy Gafni and Manuel Dahnert for the helpful discussions, and Angela Dai for the video voiceover." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 178, + 106, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 178, + 106, + 190 + ], + "spans": [ + { + "bbox": [ + 48, + 178, + 106, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 198, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 287, + 229 + ], + "type": "text", + "content": "[1] Stable diffusion. https://github.com/Stability-AI/stablediffusion, 2022. 1, 2, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 230, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 287, + 273 + ], + "type": "text", + "content": "[2] Rameen Abdal, Hsin-Ying Lee, Peihao Zhu, Menglei Chai, Aliaksandr Siarohin, Peter Wonka, and Sergey Tulyakov. 3davatargan: Bridging domains for personalized editable avatars. In CVPR, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 274, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 287, + 304 + ], + "type": "text", + "content": "[3] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 305, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 305, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 305, + 287, + 335 + ], + "type": "text", + "content": "[4] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. In ICLR, 2018. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 336, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 287, + 390 + ], + "type": "text", + "content": "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022: 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 391, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 433 + ], + "type": "text", + "content": "[6] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 434, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 287, + 487 + ], + "type": "text", + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 487, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 530 + ], + "type": "text", + "content": "[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. ScanRefer: 3D object localization in RGB-D scans using natural language. In European Conference on Computer Vision, pages 202-221. Springer, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 531, + 287, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 531, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 53, + 531, + 287, + 572 + ], + "type": "text", + "content": "[9] Dave Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. arXiv preprint arXiv:2212.00836, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 573, + 287, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 573, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 48, + 573, + 287, + 638 + ], + "type": "text", + "content": "[10] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 487-505. Springer, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 638, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 638, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 638, + 287, + 691 + ], + "type": "text", + "content": "[11] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3193-3203, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[12] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 704 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[13] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander Schwing, and Liangyan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 117, + 545, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 117, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 545, + 159 + ], + "type": "text", + "content": "[14] Zezhou Cheng, Mengei Chai, Jian Ren, Hsin-Ying Lee, Kyle Olszewski, Zeng Huang, Subhransu Maji, and Sergey Tulyakov. Cross-modal 3d shape generation and manipulation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 159, + 545, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 159, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 545, + 202 + ], + "type": "text", + "content": "[15] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In ECCV, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 202, + 545, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 202, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 307, + 202, + 545, + 234 + ], + "type": "text", + "content": "[16] Angela Dai, Yawar Siddiqui, Justus Thies, Julien Valentin, and Matthias Nießner. Spsg: Self-supervised photometric scene generation from rgb-d scans. In CVPR, 2021. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 234, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 287 + ], + "type": "text", + "content": "[17] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. arXiv preprint arXiv:2212.08051, 2022. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 288, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 288, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 288, + 545, + 308 + ], + "type": "text", + "content": "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 309, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 340 + ], + "type": "text", + "content": "[19] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 341, + 545, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 341, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 307, + 341, + 545, + 394 + ], + "type": "text", + "content": "[20] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 395, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 395, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 395, + 545, + 426 + ], + "type": "text", + "content": "[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NIPS, 2014. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 426, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 458 + ], + "type": "text", + "content": "[22] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In ICLR, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 459, + 545, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 545, + 500 + ], + "type": "text", + "content": "[23] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 501, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 521 + ], + "type": "text", + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 522, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 564 + ], + "type": "text", + "content": "[25] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. arXiv preprint arXiv:2106.15282, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 565, + 545, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 565, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 307, + 565, + 545, + 608 + ], + "type": "text", + "content": "[26] Ronghang Hu and Amanpreet Singh. Unit: Multimodal multitask learning with a unified transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1439-1449, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 608, + 545, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 608, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 307, + 608, + 545, + 639 + ], + "type": "text", + "content": "[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 639, + 545, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 639, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 307, + 639, + 545, + 682 + ], + "type": "text", + "content": "[28] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 683, + 545, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 683, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 307, + 683, + 545, + 704 + ], + "type": "text", + "content": "[29] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fi" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18566" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 703 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "dler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. arXiv preprint arXiv:2211.10440, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 148 + ], + "type": "text", + "content": "[30] Chieh Hubert Lin, Hsin-Ying Lee, Willi Menapace, Mengei Chai, Aliaksandr Siarohin, Ming-Hsuan Yang, and Sergey Tulyakov. *Infinicity: Infinite-scale city synthesis.* arXiv preprint arXiv:2301.09637, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 148, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 148, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 48, + 148, + 287, + 190 + ], + "type": "text", + "content": "[31] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 191, + 287, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 191, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 48, + 191, + 287, + 212 + ], + "type": "text", + "content": "[32] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 212, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 212, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 212, + 287, + 243 + ], + "type": "text", + "content": "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, 2023. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "text", + "content": "[34] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In CVPR, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 319, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 350 + ], + "type": "text", + "content": "[36] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. AutoSDF: Shape priors for 3d completion, reconstruction and generation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 392 + ], + "type": "text", + "content": "[37] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia, 2022. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 393, + 287, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 436 + ], + "type": "text", + "content": "[38] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. arXiv preprint arXiv:2212.01206, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 287, + 479 + ], + "type": "text", + "content": "[39] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International conference on machine learning, pages 7220-7229. PMLR, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 479, + 287, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 500 + ], + "type": "text", + "content": "[40] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 500, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 531 + ], + "type": "text", + "content": "[41] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 531, + 287, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 287, + 573 + ], + "type": "text", + "content": "[42] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In ICCV, 2019. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 574, + 287, + 617 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 574, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 48, + 574, + 287, + 617 + ], + "type": "text", + "content": "[43] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 617, + 287, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 617, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 48, + 617, + 287, + 660 + ], + "type": "text", + "content": "[44] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 661, + 287, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 661, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 48, + 661, + 287, + 693 + ], + "type": "text", + "content": "[45] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 693, + 287, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 693, + 287, + 703 + ], + "spans": [ + { + "bbox": [ + 48, + 693, + 287, + 703 + ], + "type": "text", + "content": "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 705 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "hall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 148 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICLR, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 148, + 545, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 545, + 190 + ], + "type": "text", + "content": "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 191, + 545, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 191, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 307, + 191, + 545, + 223 + ], + "type": "text", + "content": "[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 223, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 223, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 223, + 545, + 266 + ], + "type": "text", + "content": "[50] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 266, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 298 + ], + "type": "text", + "content": "[51] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. arXiv preprint arXiv:2302.01721, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 298, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 329 + ], + "type": "text", + "content": "[52] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 329, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 329, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 329, + 545, + 361 + ], + "type": "text", + "content": "[53] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 361, + 545, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 393 + ], + "type": "text", + "content": "[54] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 393, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 436 + ], + "type": "text", + "content": "[55] Aliaksandr Siarohin, Willi Menapace, Ivan Skorokhodov, Kyle Olszewski, Jian Ren, Hsin-Ying Lee, Mengei Chai, and Sergey Tulyakov. Unsupervised volumetric animation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 436, + 545, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 436, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 307, + 436, + 545, + 468 + ], + "type": "text", + "content": "[56] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. In ECCV, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 468, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 533 + ], + "type": "text", + "content": "[57] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 533, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 564 + ], + "type": "text", + "content": "[58] Ivan Skorokhodov, Aliaksandr Siarohin, Yinghao Xu, Jian Ren, Hsin-Ying Lee, Peter Wonka, and Sergey Tulyakov. 3d generation on imagenet. In ICLR, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 564, + 545, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 564, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 307, + 564, + 545, + 596 + ], + "type": "text", + "content": "[59] Edward J Smith and David Meger. Improved adversarial systems for 3d object generation and reconstruction. In CoRL, 2017. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 596, + 545, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 596, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 307, + 596, + 545, + 661 + ], + "type": "text", + "content": "[60] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. OFA: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 661, + 545, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 661, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 307, + 661, + 545, + 705 + ], + "type": "text", + "content": "[61] Jiajun Wu, Chengkai Zhang, Xiuming Zhang, Zhoutong Zhang, William T Freeman, and Joshua B Tenenbaum. Learning shape priors for single-view 3d completion and reconstruction. In Proceedings of the European Conference on" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18567" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 373 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 68, + 72, + 253, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 72, + 253, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 72, + 253, + 83 + ], + "type": "text", + "content": "Computer Vision (ECCV), pages 646-662, 2018. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 84, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 84, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 48, + 84, + 287, + 125 + ], + "type": "text", + "content": "[62] Jianwen Xie, Zilong Zheng, Ruiqi Gao, Wenguan Wang, Song-Chun Zhu, and Ying Nian Wu. Learning descriptor networks for 3d shape synthesis and analysis. In CVPR, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 126, + 287, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 126, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 48, + 126, + 287, + 180 + ], + "type": "text", + "content": "[63] Yinghao Xu, Mengei Chai, Zifan Shi, Sida Peng, Ivan Skorokhodov, Aliaksandr Siarohin, Ceyuan Yang, Yujun Shen, Hsin-Ying Lee, Bolei Zhou, et al. Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In CVPR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 180, + 287, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 180, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 180, + 287, + 233 + ], + "type": "text", + "content": "[64] Linjie Yang, Ping Luo, Chen Change Loy, and Xiaou Tang. A large-scale car dataset for fine-grained categorization and verification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3973-3981, 2015. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 234, + 287, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 234, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 234, + 287, + 266 + ], + "type": "text", + "content": "[65] Rui Yu, Yue Dong, Pieter Peers, and Xin Tong. Learning texture generators for 3d shape collections from internet photo sets. In BMVC, 2021. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 266, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 266, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 287, + 308 + ], + "type": "text", + "content": "[66] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 341 + ], + "type": "text", + "content": "[67] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 341, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 341, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 341, + 287, + 373 + ], + "type": "text", + "content": "[68] Song-Hai Zhang, Yuan-Chen Guo, and Qing-Wen Gu. Sketch2model: View-aware 3d modeling from single free-hand sketches. In CVPR, 2021. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18568" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/56c5c84c-dd92-4857-8492-3e83266b25fc_content_list.json b/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/56c5c84c-dd92-4857-8492-3e83266b25fc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9675abdf6748daa571e4f0475386b86391df1e91 --- /dev/null +++ b/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/56c5c84c-dd92-4857-8492-3e83266b25fc_content_list.json @@ -0,0 +1,2412 @@ +[ + { + "type": "text", + "text": "Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators", + "text_level": 1, + "bbox": [ + 158, + 130, + 810, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Levon Khachatryan $^{1*}$ Andranik Movsisyan $^{1*}$ Vahram Tadevosyan $^{1*}$ Roberto Henschel $^{1*}$ Zhangyang Wang $^{1,2}$ Shant Navasardyan $^{1}$ Humphrey Shi $^{1,3}$ $^{1}$ Picsart AI Research (PAIR) $^{2}$ UT Austin $^{3}$ SHI Labs @ Georgia Tech, Oregon & UIUC https://github.com/Picsart-AI-Research/Text2Video-Zero", + "bbox": [ + 114, + 202, + 846, + 276 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5638995bdf120554a11b3d2bc540caa4ab3e8a7c2868cc2d7e7fe856b11f0784.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 286, + 181, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a4d8d0a3ceca72cf40b495bea7ac0f26329a5b6dd01e1531cc72073f7a0e8251.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 287, + 282, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aeeb8c18a627e667284dee6ab977731786cf2633a9eed02ccbf5f00d10389a6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 287, + 383, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e8e563dfa608ee4bb05d24195d74a25e40bca7cd71d46b792d52b0c2e144d75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 287, + 483, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f9323419c2c537317827585f68dc62e75a08942e2c5bf67a185596a625edd231.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 287, + 586, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2cb5b3ba5892262e4f2653806a92caf658daf7fefbe529caef50e48a01a34e66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 287, + 687, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3c8eb8cd473b95a423a37cfe89749e5874467222ce30246c27a17fd6e964d98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 287, + 787, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ef3f447b8f14833e105f220bbfb65c77cab36a0b3a985334e5aab9fab31285f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 287, + 890, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-Video generation: \"a horse galloping on a street\"", + "bbox": [ + 318, + 366, + 648, + 380 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d5f7e351bb85f55105e698774b39c26b3d37d9e15ef182532d35f05c49a2c21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 385, + 179, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0ab0779b112b13190a80c90c982d45723d9d92f348aef7d9c95a2ddd267cdd67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 385, + 279, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39c461fa8f85041f87e27ab52faf1c89a2be2b45fcdf6acef5962b4e82f8c215.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 385, + 380, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b943bb548f0ef3e190658ddb374d22225872c963246fef31a653736ec95e6290.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 385, + 482, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/840293918536d238e5f19ca003a506b66fb15b941f5234520443ae8b6ee22c91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 385, + 583, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0c3cc4cfe40b25a2fd3542f66e350609066bc3f7b46224fb9cd77b59233c84ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 385, + 684, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/635b8291b229f014be55890211a331493e9d02368740bcbfee9f4ef8a6df5412.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 385, + 785, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/819828471aefa4ac20a52e296620fe280800fa82e8aebe8e25ae192f7f83e46c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 385, + 887, + 462 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-Video generation: \"a panda is playing guitar on times square\"", + "bbox": [ + 284, + 463, + 686, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/480ad00bf61e23c3b61d9e02a24e6560632e83738a3d73875f45f49883bf473e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 494, + 225, + 558 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/730b01567f2aae5a32d7af0090815955b6c306e8df495bf04b55b85c0ff9153e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 494, + 379, + 558 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4911268d744d8638ec96d63d1f9037b9b22f564e3fe6b28558cb7095a5fe6b96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 494, + 532, + 558 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a3aeaf6679817680885232c903f0b133184877bb2afe91cf73169deb0cbd9d61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 478, + 630, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec210fa5b5451758b917214eced5d7a9fdd264a0af302f4399c16b7d4aef56e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 478, + 717, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f0804966b7b5bc4001e2b961ba987f85799727ca73df2e167ccb3c1aa7d74b89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 478, + 802, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9242f762bf79b0af603c475a35eb32e70bcf12774fd5f32c533c470d571ff4ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 478, + 887, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d1a694a11d86ed5c7ab1519115b1566b8ebbe2bde282a4fc4e2f5d8b2749b519.jpg", + "image_caption": [ + "Video Instruct-Pix2Pix: \"make it Van Gogh Starry Night style\"" + ], + "image_footnote": [], + "bbox": [ + 80, + 571, + 225, + 635 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d3e9c648aaa0481ddedcfc7d7de88c6c123d85a14ba39031da9e1e6b51a7fff.jpg", + "image_caption": [ + "Figure 1: Our method Text2Video-Zero enables zero-shot video generation using (i) a textual prompt (see rows 1, 2), (ii) a prompt combined with guidance from poses or edges (see lower right), and (iii) Video Instruct-Pix2Pix, i.e., instruction-guided video editing (see lower left). Results are temporally consistent and follow closely the guidance and textual prompts." + ], + "image_footnote": [], + "bbox": [ + 233, + 571, + 379, + 635 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8530c75814d28747a705839bf7314dfbf197d044bd9dea2d0915c7ace3a34787.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 571, + 532, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-Video generation + pose control: \"a bear dancing on the concrete\"", + "bbox": [ + 573, + 542, + 857, + 568 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95aabc830f54679227d25a91795bc51a4a902b50cf7b59265807a052c4655bcc.jpg", + "image_caption": [ + "Text-to-Video generation + edge control: \"white butterfly\"" + ], + "image_footnote": [], + "bbox": [ + 547, + 569, + 630, + 631 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/81454a49861a03f943a212da4832b983a0defbb449084f3f786ee71a4f2f67c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 568, + 715, + 631 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4ec10fefe4c5557de5c9080d4dc0e738df1ec2882fdb1cdafcffde8c06f82e34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 568, + 802, + 631 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3de2cf00b5845d69926f12d24372e2b3d58591b656c7d36840d8c6e9727f1026.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 568, + 887, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 734, + 313, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent text-to-video generation approaches rely on computationally heavy training and require large-scale video datasets. In this paper, we introduce a new task, zero-shot text-to-video generation, and propose a low-cost approach (without any training or optimization) by leveraging the power of existing text-to-image synthesis methods (e.g. Stable Diffusion), making them suitable for the video domain. Our key modifications include (i) enriching the la", + "bbox": [ + 75, + 752, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tent codes of the generated frames with motion dynamics to keep the global scene and the background time consistent; and (ii) reprogramming frame-level self-attention using a new cross-frame attention of each frame on the first frame, to preserve the context, appearance, and identity of the foreground object. Experiments show that this leads to low overhead, yet high-quality and remarkably consistent video generation. Moreover, our approach is not limited to text-to-video synthesis but is also applicable to other tasks such as conditional and content-specialized video generation, and Video Instruct-Pix2Pix, i.e., instruction-guided", + "bbox": [ + 500, + 734, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 96, + 887, + 205, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15954", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "video editing. As experiments show, our method performs comparably or sometimes better than recent approaches, despite not being trained on additional video data. Our code is publicly available at: https://github.com/Picsart-AI-Research/Text2Video-Zero.", + "bbox": [ + 76, + 90, + 470, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 195, + 207, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, generative AI has attracted enormous attention in the computer vision community. With the advent of diffusion models [34, 12, 35, 36], it has become tremendously popular and successful to generate high-quality images from textual prompts, also called text-to-image synthesis [26, 29, 32, 7, 44]. Recent works [14, 33, 11, 42, 5, 21] attempt to extend the success to text-to-video generation and editing tasks, by reusing text-to-image diffusion models in the video domain. While such approaches yield promising outcomes, most of them require substantial training with a massive amount of labeled data which can be costly and unaffordable for many users. With the aim of making video generation cheaper, Tune-A-Video [42] introduces a mechanism that can adopt Stable Diffusion (SD) model [29] for the video domain. The training effort is drastically reduced to tuning one video. While that is much more efficient than previous approaches, it still requires an optimization process. In addition, the generation abilities of Tune-A-Video are limited to text-guided video editing applications; video synthesis from scratch, however, remains out of its reach.", + "bbox": [ + 76, + 220, + 468, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we take one step forward in studying the novel problem of zero-shot, \"training-free\" text-to-video synthesis, which is the task of generating videos from textual prompts without requiring any optimization or finetuning. A key concept of our approach is to modify a pre-trained text-to-image model (e.g., Stable Diffusion), enriching it with temporally consistent generation. By building upon already trained text-to-image models, our method takes advantage of their excellent image generation quality and enhances their applicability to the video domain without performing additional training. To enforce temporal consistency, we present two innovative and lightweight modifications: (1) we first enrich the latent codes of generated frames with motion information to keep the global scene and the background time consistent; (2) we then use cross-frame attention of each frame on the first frame to preserve the context, appearance, and identity of the foreground object throughout the entire sequence. Our experiments show that these simple modifications lead to high-quality and time-consistent video generations (see Fig. 1 and further results in the appendix). Despite the fact that other works train on large-scale video data, our method achieves similar or sometimes even better performance (see Figures 8, 9 and appendix Figures 18, 25, 26). Furthermore, our method is not limited to text-to-video synthesis but is", + "bbox": [ + 76, + 523, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "also applicable to conditional (see Figures 5,6 and appendix Figures 19, 21, 22, 23) and specialized video generation (see Fig. 7), and instruction-guided video editing, which we refer as Video Instruct-Pix2Pix motivated by Instruct-Pix2Pix [2] (see Fig. 9 and appendix Figures 24, 25, 26).", + "bbox": [ + 496, + 90, + 890, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as three-folds:", + "bbox": [ + 517, + 167, + 843, + 180 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A new problem setting of zero-shot text-to-video synthesis, aiming at making text-guided video generation and editing \"freely affordable\". We use only a pretrained text-to-image diffusion model without any further fine-tuning or optimization.", + "- Two novel post-hoc techniques to enforce temporally consistent generation, via encoding motion dynamics in the latent codes, and reprogramming each frame's self-attention using a new cross-frame attention.", + "- A broad variety of applications that demonstrate our method's effectiveness, including conditional and specialized video generation, and Video Instruct-Pix2Pix i.e., video editing by textual instructions." + ], + "bbox": [ + 517, + 189, + 890, + 401 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 414, + 640, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 500, + 439, + 733, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early approaches to text-to-image synthesis relied on methods such as template-based generation [19] and feature matching [28]. However, these methods were limited in their ability to generate realistic and diverse images.", + "bbox": [ + 496, + 462, + 890, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following the success of GANs [8], several other deep learning-based methods were proposed for text-to-image synthesis. These include StackGAN [46], AttnGAN [43], and MirrorGAN [24], which further improve image quality and diversity by introducing novel architectures and attention mechanisms.", + "bbox": [ + 496, + 523, + 890, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Later, with the advancement of transformers [38], new approaches emerged for text-to-image synthesis. Being a 12-billion-parameter transformer model, Dall-E [27] introduces a two-stage training process: First, it generates image tokens, which later are combined with text tokens for joint training of an autoregressive model. Later Parti [45] proposed a method to generate content-rich images with multiple objects. Make-a-Scene [7] enables a control mechanism by segmentation masks for text-to-image generation.", + "bbox": [ + 496, + 613, + 890, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current approaches build upon diffusion models, thereby taking text-to-image synthesis quality to the next level. GLIDE [23] improved Dall-E by adding classifier-free guidance [13]. Later, Dall-E 2 [26] utilizes the contrastive model CLIP [25]. By means of diffusion processes, (i) a mapping from CLIP text encodings to image encodings, and (ii) a CLIP decoder is obtained. LDM / SD [29] applies a diffusion model on lower-resolution encoded signals of VQ-GAN [6], showing competitive quality with a significant gain in speed and efficiency. Imagen [32] shows", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15955", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "incredible performance in text-to-image synthesis by utilizing large language models for text processing. Versatile Diffusion [44] further unifies text-to-image, image-to-text and variations in a single multi-flow diffusion model. Specialisation of text-to-image models to desired styles can be obtained efficiently via few-shot tuning, e.g. using DreamBoth [31] or Specialist Diffusion [18], which employs text-to-image customized data augmentations.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Because of their great image quality, it is desired to exploit text-to-image models for video generation. However, applying diffusion models in the video domain is not straightforward, especially due to their probabilistic generation procedure, making it difficult to ensure temporal consistency. As we show in our ablation experiments in the appendix (see Fig. 14), our modifications are crucial for temporal consistency in terms of both global scene and background motion, and for the preservation of the foreground object identity.", + "bbox": [ + 75, + 215, + 472, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Text-to-Video Generation", + "text_level": 1, + "bbox": [ + 76, + 386, + 310, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-video synthesis is a relatively new research direction. Existing approaches try to leverage autoregressive transformers and diffusion processes for the generation. NUWA [41] introduces a 3D transformer encoder-decoder framework and supports both text-to-image and text-to-video generation. Phenaki [39] introduces a bidirectional masked transformer with a causal attention mechanism that allows the generation of arbitrary-long videos from text prompt sequences. CogVideo [15] extends the text-to-image model CogView 2 [4] by tuning it using a multi-frame-rate hierarchical training strategy to better align text and video clips. Video Diffusion Models (VDM) [14] naturally extend text-to-image diffusion models and train jointly on image and video data. Imagen Video [11] constructs a cascade of video diffusion models and utilizes spatial and temporal super-resolution models to generate high-resolution time-consistent videos. Make-A-Video [33] builds upon a text-to-image synthesis model and leverages video data in an unsupervised manner. Gen-1 [5] extends SD and proposes a structure and content-guided video editing method based on visual or textual descriptions of desired outputs. Tune-A-Video [42] proposes a new task of one-shot video generation by extending and tuning SD on a single reference video.", + "bbox": [ + 75, + 414, + 472, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike the methods mentioned above, our approach is completely training-free, does not require massive computing power or dozens of GPUs, which makes the video generation process affordable for everyone. In this respect, Tunea-Video [42] comes closest to our work, as it reduces the necessary computations to tuning on only one video. However, it still requires an optimization process and its generating ability is heavily restricted by the reference video.", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 89, + 591, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start this section with a brief introduction of diffusion models, particularly Stable Diffusion (SD) [29]. Then we introduce the problem formulation of zero-shot text-to-video synthesis, followed by a subsection presenting our approach. After that, to show the universality of our method, we use it in combination with ControlNet [47] and Dream-Booth [31] diffusion models for generating conditional and specialized videos. Later we demonstrate the power of our approach with the application of instruction-guided video editing, namely, Video Instruct-Pix2Pix.", + "bbox": [ + 496, + 114, + 893, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Stable Diffusion", + "text_level": 1, + "bbox": [ + 500, + 273, + 661, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SD is a diffusion model operating in the latent space of an autoencoder $\\mathcal{D}(\\mathcal{E}(\\cdot))$ , namely VQ-GAN [6] or VQ-VAE [37], where $\\mathcal{E}$ and $\\mathcal{D}$ are the corresponding encoder and decoder, respectively. More precisely, if $x_0 \\in \\mathbb{R}^{h \\times w \\times c}$ is the latent tensor of an input image $\\mathrm{Im}$ given by the autoencoder, i.e. $x_0 = \\mathcal{E}(\\mathrm{Im})$ , diffusion forward process iteratively adds Gaussian noise to the signal $x_0$ :", + "bbox": [ + 496, + 297, + 893, + 404 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I\\right), t = 1,.., T \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 412, + 890, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $q(x_{t}|x_{t - 1})$ is the conditional density of $x_{t}$ given $x_{t - 1}$ , and $\\{\\beta_t\\}_{t = 1}^T$ are hyperparameters. $T$ is chosen to be as large that the forward process completely destroys the initial signal $x_0$ resulting in $x_{T}\\sim \\mathcal{N}(0,I)$ . The goal of SD is then to learn a backward process", + "bbox": [ + 496, + 439, + 893, + 516 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 523, + 890, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for $t = T, \\dots, 1$ , which allows to generate a valid signal $x_0$ from the standard Gaussian noise $x_T$ . To get the final image generated from $x_T$ it remains to pass $x_0$ to the decoder of the initially chosen autoencoder: $\\operatorname{Im} = \\mathcal{D}(x_0)$ .", + "bbox": [ + 496, + 549, + 893, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After learning the abovementioned backward diffusion process (see DDPM [12]) one can apply a deterministic sampling process, called DDIM [35]:", + "bbox": [ + 496, + 609, + 893, + 656 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\left(\\frac {x _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} ^ {t} \\left(x _ {t}\\right)}{\\sqrt {\\alpha_ {t}}}\\right) + \\tag {3} \\\\ \\sqrt {1 - \\alpha_ {t - 1}} e _ {\\theta} ^ {t} (x _ {t}), \\quad t = T, \\dots , 1, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 662, + 890, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t} = \\prod_{i = 1}^{t}(1 - \\beta_{i})$ and", + "bbox": [ + 500, + 728, + 709, + 747 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\theta} ^ {t} (x _ {t}) = \\frac {\\sqrt {1 - \\alpha_ {t}}}{\\beta_ {t}} x _ {t} + \\frac {(1 - \\beta_ {t}) (1 - \\alpha_ {t})}{\\beta_ {t}} \\mu_ {\\theta} (x _ {t}, t). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 753, + 890, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To get a text-to-image synthesis framework, SD guides the diffusion processes with a textual prompt $\\tau$ . Particularly for DDIM sampling, we get:", + "bbox": [ + 496, + 794, + 893, + 842 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\left(\\frac {x _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} ^ {t} \\left(x _ {t} , \\tau\\right)}{\\sqrt {\\alpha_ {t}}}\\right) + \\tag {5} \\\\ \\sqrt {1 - \\alpha_ {t - 1}} \\epsilon_ {\\theta} ^ {t} (x _ {t}, \\tau), \\quad t = T, \\ldots , 1. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 848, + 890, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15956", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is worth noting that in SD, the function $\\epsilon_{\\theta}^{t}(x_{t},\\tau)$ is modeled as a neural network with a UNet-like [30] architecture composed of convolutional and (self- and cross-) attentional blocks. $x_{T}$ is called the latent code of the signal $x_0$ and there is a method [3] to apply a deterministic forward process to reconstruct the latent code $x_{T}$ given a signal $x_0$ . This method is known as DDIM inversion. Sometimes for simplicity, we will call $x_{t},t = 1,\\ldots ,T$ also the latent codes of the initial signal $x_0$ .", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Zero-Shot Text-to-Video Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 250, + 468, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing text-to-video synthesis methods require either costly training on a large-scale (ranging from $1M$ to $15M$ data-points) text-video paired data [41, 15, 39, 11, 14, 5] or tuning on a reference video [42]. To make video generation cheaper and easier, we propose a new problem: zero-shot text-to-video synthesis. Formally, given a text description $\\tau$ and a positive integer $m\\in \\mathbb{N}$ , the goal is to design a function $\\mathcal{F}$ that outputs video frames $\\mathcal{V}\\in \\mathbb{R}^{m\\times H\\times W\\times 3}$ (for predefined resolution $H\\times W$ ) that exhibit temporal consistency. To determine the function $\\mathcal{F}$ , no training or fine-tuning must be performed on a video dataset.", + "bbox": [ + 75, + 277, + 468, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our problem formulation provides a new paradigm for text-to-video. Noticeably, a zero-shot text-to-video method naturally benefits from quality improvements of text-to-image models.", + "bbox": [ + 75, + 448, + 470, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Method", + "text_level": 1, + "bbox": [ + 76, + 531, + 174, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we approach the zero-shot text-to-video task by exploiting the text-to-image synthesis power of Stable Diffusion (SD). As we need to generate videos instead of images, SD should operate on sequences of latent codes. The naive approach is to independently sample $m$ latent codes from standard Gaussian distribution $x_{T}^{1},\\ldots ,x_{T}^{m}\\sim \\mathcal{N}(0,I)$ and apply DDIM sampling to obtain the corresponding tensors $x_0^k$ for $k = 1,\\dots,m$ , followed by decoding to obtain the generated video sequence $\\{\\mathcal{D}(x_0^k)\\}_{k = 1}^m\\in \\mathbb{R}^{m\\times H\\times W\\times 3}$ . However, this leads to completely random generation of images sharing only the semantics described by $\\tau$ but neither object appearance nor motion coherence (see appendix Fig. 14, first row).", + "bbox": [ + 75, + 559, + 468, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue, we propose to (i) introduce motion dynamics between the latent codes $x_{T}^{1}, \\ldots, x_{T}^{m}$ to keep the global scene time consistent and (ii) use cross-frame attention mechanism to preserve the appearance and the identity of the foreground object. Each of the components of our method are described below in detail. The overview of our method can be found in Fig. 2.", + "bbox": [ + 75, + 758, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note, to simplify notation, we will denote the entire sequence of latent codes by $x_{T}^{1:m} = [x_{T}^{1}, \\ldots, x_{T}^{m}]$ .", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Motion dynamics in latent codes" + ], + "code_body": "Require: $\\Delta t\\geq 0,m\\in \\mathbb{N},\\lambda >0,\\delta = (\\delta_x,\\delta_y)\\in$ $\\mathbb{R}^2$ ,Stable Diffusion (SD) \n1: $x_{T}^{1}\\sim \\mathcal{N}(0,I)$ random sample the first latent code \n2: $x_{T^{\\prime}}^{1}\\gets \\mathrm{DDIM\\_Backward}(x_{T}^{1},\\Delta t,SD)\\triangleright$ perform $\\Delta t$ backward steps by SD \n3: for all $k = 2,3,\\ldots ,m$ do \n4: $\\delta^k\\gets \\lambda \\cdot (k - 1)\\delta \\triangleright$ computing global translation vectors \n5: $W_{k}\\leftarrow$ Warping by $\\delta^k\\triangleright$ defining warping functions \n6: $\\tilde{x}_{T^{\\prime}}^{k}\\gets W_{k}(x_{T^{\\prime}}^{1})$ \n7: $x_{T}^{k}\\gets \\mathrm{DDPM\\_Forward}(\\tilde{x}_{T^{\\prime}}^{k},\\Delta t)\\triangleright \\mathrm{DDPM~forward}$ for more motion freedom return $x_{T}^{1:m}$", + "bbox": [ + 501, + 109, + 893, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Motion Dynamics in Latent Codes", + "text_level": 1, + "bbox": [ + 500, + 334, + 792, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of sampling the latent codes $x_{T}^{1:m}$ randomly and independently from the standard Gaussian distribution, we construct them by performing the following steps (see also Alg. 1 and Fig. 2).", + "bbox": [ + 498, + 358, + 892, + 420 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Randomly sample the latent code of the first frame: $x_{T}^{1} \\sim \\mathcal{N}(0, I)$ .", + "2. Perform $\\Delta t \\geq 0$ DDIM backward steps on the latent code $x_{T}^{1}$ by using the SD model and get the corresponding latent $x_{T'}^{1}$ , where $T' = T - \\Delta t$ .", + "3. Define a direction $\\delta = (\\delta_x, \\delta_y) \\in \\mathbb{R}^2$ for the global scene and camera motion. By default $\\delta$ can be the main diagonal direction $\\delta_x = \\delta_y = 1$ .", + "4. For each frame $k = 1,2,\\ldots ,m$ we want to generate, compute the global translation vector $\\delta^k = \\lambda \\cdot (k - 1)\\delta$ , where $\\lambda$ is a hyperparameter controlling the amount of the global motion.", + "5. Apply the constructed motion (translation) flow $\\delta^{1:m}$ to $x_{T'}^1$ , denote the resulting sequence by $\\tilde{x}_{T'}^{1:m}$ :" + ], + "bbox": [ + 511, + 428, + 890, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {x} _ {T ^ {\\prime}} ^ {k} = W _ {k} \\left(x _ {T ^ {\\prime}} ^ {1}\\right) \\text {f o r} k = 1, 2, \\dots , m, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 686, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{k}(x_{T^{\\prime}}^{1})$ is the warping operation for translation by the vector $\\delta^{k}$ .", + "bbox": [ + 529, + 715, + 888, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6. Perform $\\Delta t$ DDPM forward steps on each of the latents $\\tilde{x}_{T'}^{2:m}$ and get the corresponding latent codes $x_{T}^{2:m}$ .", + "bbox": [ + 511, + 755, + 888, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we take the sequence $x_{T}^{1:m}$ as the starting point of the backward (video) diffusion process. As a result, the latent codes generated with our proposed motion dynamics lead to better temporal consistency of the global scene as well as the background, see in the appendix, Sect. 8.1. Yet, the initial latent codes are not constraining enough to describe particular colors, identities or shapes, thus still", + "bbox": [ + 496, + 794, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15957", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/122a24541992a5835ea4cd4ab2532dd7f4bba239b8a46f912763c3fdc5492f96.jpg", + "image_caption": [ + "Figure 2: Method overview: Starting from a randomly sampled latent code $x_{T}^{1}$ , we apply $\\Delta t$ DDIM backward steps to obtain $x_{T'}^{1}$ using a pre-trained Stable Diffusion model (SD). A specified motion field results for each frame $k$ in a warping function $W_{k}$ that turns $x_{T'}^{1}$ to $x_{T'}^{k}$ . By enhancing the latent codes with motion dynamics, we determine the global scene and camera motion and achieve temporal consistency in the background and the global scene. A subsequent DDPM forward application delivers latent codes $x_{T}^{k}$ for $k = 1, \\ldots, m$ . By using the (probabilistic) DDPM method, a greater degree of freedom is achieved with respect to the motion of objects (see appendix Sec. 8.1). Finally, the latent codes are passed to our modified SD model using the proposed cross-frame attention, which uses keys and values from the first frame to generate the image of frame $k = 1, \\ldots, m$ . By using cross-frame attention, the appearance and the identity of the foreground object are preserved throughout the sequence. Optionally, we apply background smoothing. To this end, we employ salient object detection to obtain for each frame $k$ a mask $M^{k}$ indicating the foreground pixels. Finally, for the background (using the mask $M^{k}$ ), a convex combination between the latent code $x_{t}^{1}$ of frame one warped to frame $k$ and the latent code $x_{t}^{k}$ is used to further improve the temporal consistency of the background." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 877, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "leading to temporal inconsistencies, especially for the foreground object.", + "bbox": [ + 75, + 479, + 470, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Reprogramming Cross-Frame Attention", + "text_level": 1, + "bbox": [ + 76, + 527, + 413, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the issue mentioned above, we use a cross-frame attention mechanism to preserve the information about (in particular) the foreground object's appearance, shape, and identity throughout the generated video.", + "bbox": [ + 75, + 551, + 468, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To leverage the power of cross-frame attention and at the same time exploit a pretrained SD without retraining, we replace each of its self-attention layers with a cross-frame attention, with the attention for each frame being on the first frame. More precisely in the original SD UNet architecture $\\epsilon_{\\theta}^{t}(x_{t},\\tau)$ , each self-attention layer takes a feature map $x\\in \\mathbb{R}^{h\\times w\\times c}$ , linearly projects it into query, key, value features $Q,K,V\\in \\mathbb{R}^{h\\times w\\times c}$ , and computes the layer output by the following formula (for simplicity described here for only one attention head) [38]:", + "bbox": [ + 75, + 611, + 470, + 763 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S e l f - A t t n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {c}}\\right) V. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 768, + 468, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our case, each attention layer receives $m$ inputs: $x^{1:m} = [x^1,\\dots,x^m]\\in \\mathbb{R}^{m\\times h\\times w\\times c}$ . Hence, the linear projection layers produce $m$ queries, keys, and values $Q^{1:m},K^{1:m},V^{1:m}$ , respectively.", + "bbox": [ + 75, + 810, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, we replace each self-attention layer with a cross-frame attention of each frame on the first frame as", + "bbox": [ + 75, + 869, + 470, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "follows:", + "bbox": [ + 500, + 479, + 558, + 493 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {C r o s s - F r a m e - A t t n} \\left(Q ^ {k}, K ^ {1: m}, V ^ {1: m}\\right) = \\\\ \\operatorname {S o f t m a x} \\left(\\frac {Q ^ {k} \\left(K ^ {1}\\right) ^ {T}}{\\sqrt {c}}\\right) V ^ {1} \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 503, + 890, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for $k = 1, \\ldots, m$ . By using cross frame attention, the appearance and structure of the objects and background as well as identities are carried over from the first frame to subsequent frames, which significantly increases the temporal consistency of the generated frames (see in the appendix the Figures 14, 16, 22, 23).", + "bbox": [ + 496, + 569, + 892, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3 Background smoothing (Optional)", + "text_level": 1, + "bbox": [ + 500, + 680, + 794, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further improve temporal consistency of the background using a convex combination of background-masked latent codes between the first frame and frame $k$ . This helps especially to generate videos from textual prompts when one or no initial image and no further guidance are provided.", + "bbox": [ + 496, + 703, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In more detail, given the generated sequence of our video generator, $x_0^{1:m}$ , we apply (an in-house solution for) salient object detection [40] to the decoded images to obtain a corresponding foreground mask $M^k$ for each frame $k$ . Then we warp $x_t^1$ according to the employed motion dynamics defined by $W_k$ and denote the result by $\\hat{x}_t^k \\coloneqq W_k(x_t^1)$ .", + "bbox": [ + 496, + 779, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Background smoothing is achieved by a convex combination between the actual latent code $x_{t}^{k}$ and the warped", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15958", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b6f131c5c649c6adec56561c373e34336d84eca3cc6b006750a0ecf743bad24a.jpg", + "image_caption": [ + "Figure 3: Text-to-Video results of our method. Depicted frames show that identities and appearances are temporally consistent and fitting to the textual prompt. For more results, see Appendix Sec. 8." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 854, + 438 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/daf48d03f9add4f33db4bd7ad68c90101df7e7ba8c2d823beb961dd3210654e8.jpg", + "image_caption": [ + "Figure 4: The overview of Text2Video-Zero + ControlNet" + ], + "image_footnote": [], + "bbox": [ + 91, + 500, + 460, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "latent code $\\hat{x}_t^k$ on the background, i.e.,", + "bbox": [ + 76, + 739, + 333, + 756 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {x} _ {t} ^ {k} = M ^ {k} \\odot x _ {t} ^ {k} + (1 - M ^ {k}) \\odot (\\alpha \\hat {x} _ {t} ^ {k} + (1 - \\alpha) x _ {t} ^ {k}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 763, + 468, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for $k = 1, \\ldots, m$ , where $\\alpha$ is a hyperparameter, which we empirically choose $\\alpha = 0.6$ . Finally, DDIM sampling is employed on $\\overline{x}_t^k$ , which delivers video generation with background smoothing. We use background smoothing in our video generation from text when no guidance is provided. For an ablation study on background smoothing, see the appendix, Sec. 8.1.", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Conditional and Specialized Text-to-Video", + "text_level": 1, + "bbox": [ + 500, + 502, + 859, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recently powerful controlling mechanisms [47, 22, 17] emerged to guide the diffusion process for text-to-image generation. Particularly, ControlNet [47] enables to condition the generation process using edges, pose, semantic masks, image depths, etc. However, a direct application of ControlNet in the video domain leads to temporal inconsistencies and to severe changes of object appearance, identity, and the background (see in the appendix Figures 14, 16, 22, 23). It turns out that our modifications on the basic diffusion process for videos result in more consistent videos guided by ControlNet conditions. We would like to point out again that our method does not require any fine-tuning or optimization processes.", + "bbox": [ + 496, + 529, + 890, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "More specifically, ControlNet creates a trainable copy of the encoder (including the middle blocks) of the UNet $\\epsilon_{\\theta}^{t}(x_{t},\\tau)$ while additionally taking the input $x_{t}$ and a condition $c$ , and adds the outputs of each layer to the skip-connections of the original UNet. Here $c$ can be any type of condition, such as edge map, scribbles, pose (body landmarks), depth map, segmentation map, etc. The trainable branch is being trained on a specific domain for each type of the condition $c$ resulting in an effective conditional text-to-image generation mechanism.", + "bbox": [ + 496, + 729, + 892, + 881 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To guide our video generation process with ControlNet", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15959", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d3ad2b3a4b5b049ed5496c6a32387da51b0b14902ac6e8116c10048648b83c44.jpg", + "image_caption": [ + "(a) a panda dancing in Antarctica" + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 436, + 150 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/146925351d27b7ffa1c17ba502561b340a30916edea3f2652b40c89a654fc3dd.jpg", + "image_caption": [ + "(b) a cyborg dancing near a volcano" + ], + "image_footnote": [], + "bbox": [ + 109, + 169, + 436, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7ccb181fc487596aeaab093619c3d96ab8d5083fe3a19785c70d3d74ab20209.jpg", + "image_caption": [ + "(c) an astronaut dancing in the outer space" + ], + "image_footnote": [], + "bbox": [ + 109, + 250, + 436, + 313 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4e34c0bba37c003312b0238e5d92a646b35d900005e18f29110e0d78e93650d.jpg", + "image_caption": [ + "Figure 5: Conditional generation with pose control. For more results see appendix, Sec. 10.", + "(a) oil painting of a girl dancing close-up" + ], + "image_footnote": [], + "bbox": [ + 109, + 391, + 436, + 454 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/517241d129b050d6dca845a4a8b6745fc102cd095e105dd0b0fd23d033d798ac.jpg", + "image_caption": [ + "(b) Cyberpunk boy with a hat dancing close-up", + "Figure 6: Conditional generation with edge control. For more results see appendix, Sec. 9." + ], + "image_footnote": [], + "bbox": [ + 109, + 472, + 436, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we apply our method to the basic diffusion process, i.e. enrich the latent codes $x_{T}^{1:m}$ with motion information and change the self-attention into cross-frame attentions in the main UNet. While adopting the main UNet for video generation task, we apply the ControlNet pretrained copy branch per-frame on each $x_{t}^{k}$ for $k = 1,\\dots,m$ in each diffusion time-step $t = T,\\ldots ,1$ and add the ControlNet branch outputs to the skip-connections of the main UNet.", + "bbox": [ + 75, + 626, + 468, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, for our conditional generation task, we adopted the weights of specialized DreamBooth (DB) [31] models1. This gives us specialized time-consistent video generations (see Fig. 7).", + "bbox": [ + 76, + 768, + 468, + 829 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/825a7a5b591098e7077a2d104041b6b91451f42f4ae2f62bc2616ad6f45e35b0.jpg", + "image_caption": [ + "(a) oil painting of a beautiful girl avatar style" + ], + "image_footnote": [], + "bbox": [ + 532, + 88, + 859, + 150 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1700324d92b6c560ad0c09d2aa1ac3ba41c914f5731b997e0d6df048544f3584.jpg", + "image_caption": [ + "(b) gta-5 style", + "Figure 7: Conditional generation with edge control and DB models." + ], + "image_footnote": [], + "bbox": [ + 532, + 169, + 859, + 232 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5. Video Instruct-Pix2Pix", + "text_level": 1, + "bbox": [ + 500, + 311, + 712, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the rise of text-guided image editing methods such as Prompt2Prompt [9], Instruct-Pix2Pix [2], SDEdit [20], etc., text-guided video editing approaches emerged [1, 16, 42]. While these methods require complex optimization processes, our approach enables the adoption of any SD-based text-guided image editing algorithm to the video domain without any training or fine-tuning. Here we take the text-guided image editing method Instruct-Pix2Pix and combine it with our approach. More precisely, we change the self-attention mechanisms in Instruct-Pix2Pix to cross-frame attentions according to Eq. 8. Our experiments show that this adaptation significantly improves the consistency of the edited videos (see Fig. 9) over the naive per-frame usage of Instruct-Pix2Pix.", + "bbox": [ + 498, + 335, + 890, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 561, + 632, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 587, + 717, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We take the Stable Diffusion [29] code2 with its pretrained weights from version 1.5 as basis and implement our modifications. For each video, we generate $m = 8$ frames with $512 \\times 512$ resolution. However, our framework allows generating any number of frames, either by increasing $m$ , or by employing our method in an auto-regressive fashion where the last generated frame $m$ becomes the first frame in computing the next $m$ frames. For all text-to-video generation experiments, we take $T' = 881$ , $T = 941$ without specific tuning, while for conditional and specialized generation, and for Video Instruct-Pix2Pix, we take $T' = T = 1000$ .", + "bbox": [ + 496, + 611, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For a conditional generation, we use the codebase3 of ControlNet [47]. For specialized models, we take DB [31] models from publicly available sources. For Video Instruct", + "bbox": [ + 498, + 792, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Avatar model: https://civitai.com/models/9968/ avatar-style. GTA-5 model: https://civitai.com/models/1309/gta5-artwork-diffusion.", + "bbox": [ + 76, + 862, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2https://github.com/huggingface/diffusers. We also benefit from the codebase of Tune-A-Video https://github.com/showlab/Tune-A-Video.", + "bbox": [ + 500, + 849, + 890, + 886 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3https://github.com/lllyasviel/ControlNet.", + "bbox": [ + 517, + 887, + 843, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15960", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pix2Pix, we use the codebase of Instruct Pix2Pix [2].", + "bbox": [ + 76, + 90, + 436, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 114, + 261, + 132 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "All applications of Text2Video-Zero show that it successfully generates videos where the global scene and the background are time consistent and the context, appearance, and identity of the foreground object are maintained throughout the entire sequence.", + "bbox": [ + 76, + 138, + 468, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the case of text-to-video, we observe that it generates high-quality videos that are well-aligned to the text prompt (see Fig. 3 and the appendix). For instance, the depicted panda shows a naturally walking on the street. Likewise, using additional guidance from edges or poses (see Fig. 5, Fig. 6 and Fig. 7 and the appendix), high quality videos are generated matching the prompt and the guidance that show great temporal consistency and identity preservation.", + "bbox": [ + 75, + 215, + 468, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Videos generated by Video Instruct-Pix2Pix (see Fig. 1 and the appendix) possess high-fidelity with respect to the input video, while following closely the instruction.", + "bbox": [ + 76, + 335, + 468, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Comparison with Baselines", + "text_level": 1, + "bbox": [ + 76, + 392, + 321, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare our method with two publicly available baselines: CogVideo [15] and Tune-A-Video [42]. Since CogVideo is a text-to-video method we compare with it in pure text-guided video synthesis settings. With Tune-A-Video we compare in our Video Instruct-Pix2Pix setting.", + "bbox": [ + 76, + 415, + 468, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.1 Quantitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 510, + 308, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To show quantitative results, we evaluate the CLIP score [10], which indicates video-text alignment. We randomly take 25 videos generated by CogVideo and synthesize corresponding videos using the same prompts according to our method. The CLIP scores for our method and CogVideo are 31.19 and 29.63, respectively. Our method thus slightly outperforms CogVideo, even though the latter has 9.4 billion parameters and requires large-scale training on videos.", + "bbox": [ + 75, + 534, + 468, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2 Qualitative Comparison", + "text_level": 1, + "bbox": [ + 76, + 675, + 299, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present several results of our method in Fig. 8 and provide a qualitative comparison to CogVideo [15]. Both methods show good temporal consistency throughout the sequence, preserving the identity of the object and background. However, our method shows better text-video alignment. For instance, while our method correctly generates a video of a man riding a bicycle in the sunshine in Fig. 8(b), CogVideo sets the background to moon light. Also in Fig. 8(a), our method correctly shows a man running in the snow, while neither the snow nor a man running are clearly visible in the video generated by CogVideo.", + "bbox": [ + 75, + 699, + 468, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative results of Video Instruct-Pix2Pix and a visual comparison with per-frame Instruct-Pix2Pix and Tune-A-Video are shown in Fig. 9. While Instruct-Pix2Pix shows a good editing performance per frame, it lacks temporal consistency. This becomes evident especially in the video depicting a skiing person, where the snow and the sky are drawn using different styles and colors. Using our Video Instruct-Pix2Pix method, these issues are solved resulting in temporally consistent video edits throughout the sequence.", + "bbox": [ + 500, + 90, + 890, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While Tune-A-Video creates temporally consistent video generations, it is less aligned to the instruction guidance than our method, struggles creating local edits and losses details of the input sequence. This becomes apparent when looking at the edit of the dancer video depicted in Fig. 9 (left side). In contrast to Tune-A-Video, our method draws the entire dress brighter and at the same time better preserves the background, e.g. the wall behind the dancer is almost kept the same. Tune-A-Video draws a severely modified wall. Moreover, our method is more faithful to the input details, e.g., Video Instruct-Pix2Pix draws the dancer using the pose exactly as provided (Fig. 9 left), and shows all skiing persons appearing in the input video (compare last frame of Fig. 9(right)), in constraint to Tune-A-Video. All the above-mentioned weaknesses of Tune-A-Video can also be observed in our additional evaluations that are provided in the appendix, Figures 25, 26.", + "bbox": [ + 496, + 227, + 892, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 494, + 651, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform several ablation studies and provide the results in the appendix, Sect. 8.1. Namely, we analyse background smoothing, $\\Delta t$ , and two main components of our method: making the initial latent codes coherent to a motion, and using cross-frame attention on the first frame instead of self-attention.", + "bbox": [ + 496, + 518, + 890, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 500, + 623, + 772, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The main limitation of this work is the inability to generate longer videos with sequences of actions. Future research may target enriching our method with techniques such as autoregressive scene action generation, while keeping the training-free spirit. Overall, our method is focused on generating video key-frames as in the first stage of Imagen Video [11] and Make-A-Video [33], and can thus be considered as good basis for longer and smoother video generation by integrating temporal upsampling.", + "bbox": [ + 496, + 648, + 890, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 799, + 617, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we addressed the problem of zero-shot text-to-video synthesis and proposed a novel method for time-consistent video generation. Our approach does not require any optimization or fine-tuning, making text-to-video generation and its applications affordable for everyone. We", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "4https://github.com/timothybrooks/instruct-pix2pix.", + "bbox": [ + 76, + 875, + 362, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15961", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6abe8efd60ccf723ec1e9bfcb66f4574cb2f3051b4762cb63fac75b7a0dd275c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 87, + 210, + 155 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/02142f2d2c7c8c9739bc72cbc1006c4817647c9f722476cd81bb514a1632fb45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 88, + 300, + 155 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2f1d2962f5bfec19bf745c3ed67c17bd35952ef1ad7890897081c7ef000ff228.jpg", + "image_caption": [ + "(a) a man is running in the snow" + ], + "image_footnote": [], + "bbox": [ + 300, + 88, + 385, + 155 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4ef23ae7ad417795ac8e9930c23fbe718052fb3aecf46923b05c8940d408116f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 88, + 473, + 155 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8cc03a3fce8e50a676a49e08b32b6c6f2e20cea5cb31f87195034f3043726e62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 88, + 581, + 155 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/14901d1229a7627c775f5d890f9782ec261cd0f088986498f5c1b5d1a91fd28a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 88, + 669, + 155 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/513fb00d780079568f573e4efe0150ebf79212d9c5b007cfd5e2a13cf740a0d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 88, + 756, + 155 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0b0f7fad9e4f521742e716a39bedcd784b1e3a7b6f0498f3a4d07c7eed086b18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 88, + 844, + 155 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/03ad100fe744a479696e12c8e8ed9972586357f6dd7704cc2052c4057e1871aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 172, + 210, + 242 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/78a6506b39f07f779c7501d2e5ebc2373cf7120ecd0843e0aa64ca5104a2c138.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 172, + 299, + 242 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/29a51a72e6f73ce46daffdb36bf09a7f347d6f0f158b85c94c99b84c3b2065cc.jpg", + "image_caption": [ + "(b) a man is riding a bicycle in the sunshine" + ], + "image_footnote": [], + "bbox": [ + 300, + 172, + 385, + 242 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b847fedda237ea961f3095dc645dd670fd215bf04ccd158a6f21eab5143c84e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 172, + 473, + 242 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7d9085b17c5c4083691edca26fac6f0759dabe67ea8c568c41eb9f4232a6f9e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 172, + 581, + 242 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b438330d6528881f97ad4834fa6d6e05228a783afc07052f52f55178df4e180c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 172, + 669, + 242 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6ba3bdb805425ce2ad7bce7cdb062e10ef911cc0169f129d21e69b447148d0f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 172, + 756, + 242 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a8e18d4bf4a8d27fc085b6fe4e1669861fec77c0bce8bd34e84f032b14ca3c14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 172, + 844, + 242 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/42a8be58677d32fc464e3d22d2b1e430442a35ae3112e9d6d776eb18889bfba0.jpg", + "image_caption": [ + "Figure 8: Comparison of our method vs CogVideo on text-to-video generation task (left is ours, right is CogVideo [15]). For more comparisons, see appendix Fig. 18." + ], + "image_footnote": [], + "bbox": [ + 124, + 258, + 210, + 327 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/453efd49d5f4494cf18f24b6db0d44899c0502ce70277990668181abe65016ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 258, + 299, + 327 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2c9fcb8cf777e9e0b26f320d1e131ed8a6dd0b438a884752264add079fa6f1f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 258, + 385, + 327 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2ac10858d4af622e178cc7acaefc2d7c40248dfeb5369f110578b829daac52ab.jpg", + "image_caption": [ + "(c) a man is walking in the rain" + ], + "image_footnote": [], + "bbox": [ + 387, + 258, + 473, + 327 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/65f570a45b25502db7088cd712e9db0c3d4ca04a4ec4f1066e00e85ce661fd30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 258, + 581, + 327 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/43096e243642767407c61fc05c4ca865d4474303adea5f53c21e3d37038fd8e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 258, + 669, + 327 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/99d564c034db10e955e246dbf50133e35caff3a65b5dcee6688c8736f074e14b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 258, + 756, + 327 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a8861c84fe2e35e07e4ac85ad0e4a7b3e35beb99f1a5ae66c3bef9c7578a012f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 258, + 844, + 327 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b5e0f36cb746698940a75050946403f57e4e0d127ca69d23ad45a59261dbd117.jpg", + "image_caption": [ + "Figure 9: Comparison of Video Instruct-Pix2Pix(ours) with Tune-A-Video and per-frame Instruct-Pix2Pix. For more comparisons see our appendix." + ], + "image_footnote": [], + "bbox": [ + 91, + 405, + 890, + 647 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "demonstrated the effectiveness of our method for various applications, including conditional and specialized video generation, and Video Instruct-Pix2Pix, i.e., instruction-guided video editing. Our contributions to the field include presenting a new problem of zero-shot text-to-video synthesis, showing the utilization of text-to-image diffusion models for generating time-consistent videos, and providing evidence of the effectiveness of our method for various video synthesis applications. We believe that our proposed method will open up new possibilities for video generation and editing, making it accessible and affordable for everyone.", + "bbox": [ + 75, + 715, + 472, + 896 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments. This material is partially based upon work supported by the National Science Foundation CAREER Award #2239840, and the National AI Institute for Exceptional Education (Award #2229873) by National Science Foundation and the Institute of Education Sciences, U.S. Department of Education. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation, the Institute of Education Sciences, or the U.S. Department of Education.", + "bbox": [ + 496, + 715, + 893, + 868 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15962", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723. Springer, 2022. 7", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2, 7, 8, 13, 25, 26", + "[3] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 4", + "[4] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022.3", + "[5] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 2, 3, 4", + "[6] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 3", + "[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 89-106. Springer, 2022. 2", + "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2", + "[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 7", + "[10] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 8", + "[11] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 4, 8", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3", + "[13] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 12", + "[14] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 3, 4" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3, 4, 8, 9, 12, 13, 20", + "[16] Yao-Chih Lee, Ji-Ze Genevieve Jang, Yi-Ting Chen, Elizabeth Qiu, and Jia-Bin Huang. Shape-aware text-driven layered video editing. arXiv e-prints, pages arXiv-2301, 2023. 7", + "[17] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 6", + "[18] Haoming Lu, Hazarapet Tunanyan, Kai Wang, Shant Navasardyan, Zhangyang Wang, and Humphrey Shi. Specialist diffusion: Plug-and-play sample-efficient fine-tuning of text-to-image diffusion models to learn any unseen style. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14267-14276, 2023. 3", + "[19] Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Generating images from captions with attention. arXiv preprint arXiv:1511.02793, 2015. 2", + "[20] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 7", + "[21] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 2", + "[22] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 6", + "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[24] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. Mirrorgan: Learning text-to-image generation by redescription. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1514, 2019. 2", + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[27] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever." + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15963", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2", + "[28] Scott E Reed, Zeynep Akata, Santosh Mohan, Samuel Tenka, Bernt Schiele, and Honglak Lee. Learning what and where to draw. Advances in neural information processing systems, 29, 2016. 2", + "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 7", + "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 4", + "[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3, 7, 22", + "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2", + "[33] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 3, 8", + "[34] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2", + "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3", + "[36] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2", + "[37] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3", + "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5", + "[39] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "textual description. arXiv preprint arXiv:2210.02399, 2022. 3,4", + "[40] Wenguan Wang, Qiuxia Lai, Huazhu Fu, Jianbing Shen, Haibin Ling, and Ruigang Yang. Salient object detection in the deep learning era: An in-depth survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3239-3259, 2021. 5", + "[41] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Niwa: Visual synthesis pretraining for neural visual world creation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 720-736. Springer, 2022. 3, 4", + "[42] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2, 3, 4, 7, 8, 12, 13, 26", + "[43] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2", + "[44] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332, 2022. 2, 3", + "[45] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2", + "[46] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2", + "[47] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 3, 6, 7" + ], + "bbox": [ + 503, + 92, + 890, + 681 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15964", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/56c5c84c-dd92-4857-8492-3e83266b25fc_model.json b/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/56c5c84c-dd92-4857-8492-3e83266b25fc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..19a6f51dd623a96334649d2418194c203110e024 --- /dev/null +++ b/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/56c5c84c-dd92-4857-8492-3e83266b25fc_model.json @@ -0,0 +1,3038 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.131, + 0.812, + 0.177 + ], + "angle": 0, + "content": "Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.203, + 0.848, + 0.277 + ], + "angle": 0, + "content": "Levon Khachatryan\\(^{1*}\\) Andranik Movsisyan\\(^{1*}\\) Vahram Tadevosyan\\(^{1*}\\) Roberto Henschel\\(^{1*}\\) Zhangyang Wang\\(^{1,2}\\) Shant Navasardyan\\(^{1}\\) Humphrey Shi\\(^{1,3}\\) \n\\(^{1}\\)Picsart AI Research (PAIR) \\(^{2}\\)UT Austin \\(^{3}\\)SHI Labs @ Georgia Tech, Oregon & UIUC https://github.com/Picsart-AI-Research/Text2Video-Zero" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.287, + 0.182, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.288, + 0.283, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.288, + 0.384, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.288, + 0.485, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.288, + 0.587, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.288, + 0.688, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.288, + 0.789, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.288, + 0.891, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.367, + 0.65, + 0.381 + ], + "angle": 0, + "content": "Text-to-Video generation: \"a horse galloping on a street\"" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.386, + 0.18, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.386, + 0.28, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.386, + 0.381, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.386, + 0.483, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.386, + 0.584, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.386, + 0.685, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.386, + 0.787, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.386, + 0.888, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.464, + 0.687, + 0.478 + ], + "angle": 0, + "content": "Text-to-Video generation: \"a panda is playing guitar on times square\"" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.495, + 0.226, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.495, + 0.38, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.495, + 0.533, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.479, + 0.631, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.479, + 0.718, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.479, + 0.803, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.479, + 0.888, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.573, + 0.226, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.573, + 0.38, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.573, + 0.534, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.575, + 0.544, + 0.858, + 0.569 + ], + "angle": 0, + "content": "Text-to-Video generation + pose control: \"a bear dancing on the concrete\"" + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.57, + 0.631, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.569, + 0.717, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.569, + 0.803, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.569, + 0.888, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.638, + 0.49, + 0.652 + ], + "angle": 0, + "content": "Video Instruct-Pix2Pix: \"make it Van Gogh Starry Night style\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.635, + 0.885, + 0.649 + ], + "angle": 0, + "content": "Text-to-Video generation + edge control: \"white butterfly\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.669, + 0.895, + 0.715 + ], + "angle": 0, + "content": "Figure 1: Our method Text2Video-Zero enables zero-shot video generation using (i) a textual prompt (see rows 1, 2), (ii) a prompt combined with guidance from poses or edges (see lower right), and (iii) Video Instruct-Pix2Pix, i.e., instruction-guided video editing (see lower left). Results are temporally consistent and follow closely the guidance and textual prompts." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.735, + 0.314, + 0.75 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.753, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Recent text-to-video generation approaches rely on computationally heavy training and require large-scale video datasets. In this paper, we introduce a new task, zero-shot text-to-video generation, and propose a low-cost approach (without any training or optimization) by leveraging the power of existing text-to-image synthesis methods (e.g. Stable Diffusion), making them suitable for the video domain. Our key modifications include (i) enriching the la" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.736, + 0.895, + 0.903 + ], + "angle": 0, + "content": "tent codes of the generated frames with motion dynamics to keep the global scene and the background time consistent; and (ii) reprogramming frame-level self-attention using a new cross-frame attention of each frame on the first frame, to preserve the context, appearance, and identity of the foreground object. Experiments show that this leads to low overhead, yet high-quality and remarkably consistent video generation. Moreover, our approach is not limited to text-to-video synthesis but is also applicable to other tasks such as conditional and content-specialized video generation, and Video Instruct-Pix2Pix, i.e., instruction-guided" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.206, + 0.902 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15954" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.166 + ], + "angle": 0, + "content": "video editing. As experiments show, our method performs comparably or sometimes better than recent approaches, despite not being trained on additional video data. Our code is publicly available at: https://github.com/Picsart-AI-Research/Text2Video-Zero." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.196, + 0.208, + 0.212 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.222, + 0.47, + 0.523 + ], + "angle": 0, + "content": "In recent years, generative AI has attracted enormous attention in the computer vision community. With the advent of diffusion models [34, 12, 35, 36], it has become tremendously popular and successful to generate high-quality images from textual prompts, also called text-to-image synthesis [26, 29, 32, 7, 44]. Recent works [14, 33, 11, 42, 5, 21] attempt to extend the success to text-to-video generation and editing tasks, by reusing text-to-image diffusion models in the video domain. While such approaches yield promising outcomes, most of them require substantial training with a massive amount of labeled data which can be costly and unaffordable for many users. With the aim of making video generation cheaper, Tune-A-Video [42] introduces a mechanism that can adopt Stable Diffusion (SD) model [29] for the video domain. The training effort is drastically reduced to tuning one video. While that is much more efficient than previous approaches, it still requires an optimization process. In addition, the generation abilities of Tune-A-Video are limited to text-guided video editing applications; video synthesis from scratch, however, remains out of its reach." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.524, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this paper, we take one step forward in studying the novel problem of zero-shot, \"training-free\" text-to-video synthesis, which is the task of generating videos from textual prompts without requiring any optimization or finetuning. A key concept of our approach is to modify a pre-trained text-to-image model (e.g., Stable Diffusion), enriching it with temporally consistent generation. By building upon already trained text-to-image models, our method takes advantage of their excellent image generation quality and enhances their applicability to the video domain without performing additional training. To enforce temporal consistency, we present two innovative and lightweight modifications: (1) we first enrich the latent codes of generated frames with motion information to keep the global scene and the background time consistent; (2) we then use cross-frame attention of each frame on the first frame to preserve the context, appearance, and identity of the foreground object throughout the entire sequence. Our experiments show that these simple modifications lead to high-quality and time-consistent video generations (see Fig. 1 and further results in the appendix). Despite the fact that other works train on large-scale video data, our method achieves similar or sometimes even better performance (see Figures 8, 9 and appendix Figures 18, 25, 26). Furthermore, our method is not limited to text-to-video synthesis but is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "also applicable to conditional (see Figures 5,6 and appendix Figures 19, 21, 22, 23) and specialized video generation (see Fig. 7), and instruction-guided video editing, which we refer as Video Instruct-Pix2Pix motivated by Instruct-Pix2Pix [2] (see Fig. 9 and appendix Figures 24, 25, 26)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.168, + 0.844, + 0.181 + ], + "angle": 0, + "content": "Our contributions are summarized as three-folds:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.19, + 0.892, + 0.265 + ], + "angle": 0, + "content": "- A new problem setting of zero-shot text-to-video synthesis, aiming at making text-guided video generation and editing \"freely affordable\". We use only a pretrained text-to-image diffusion model without any further fine-tuning or optimization." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.274, + 0.892, + 0.333 + ], + "angle": 0, + "content": "- Two novel post-hoc techniques to enforce temporally consistent generation, via encoding motion dynamics in the latent codes, and reprogramming each frame's self-attention using a new cross-frame attention." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.343, + 0.892, + 0.402 + ], + "angle": 0, + "content": "- A broad variety of applications that demonstrate our method's effectiveness, including conditional and specialized video generation, and Video Instruct-Pix2Pix i.e., video editing by textual instructions." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.19, + 0.892, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.415, + 0.641, + 0.43 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.44, + 0.735, + 0.456 + ], + "angle": 0, + "content": "2.1. Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.523 + ], + "angle": 0, + "content": "Early approaches to text-to-image synthesis relied on methods such as template-based generation [19] and feature matching [28]. However, these methods were limited in their ability to generate realistic and diverse images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.613 + ], + "angle": 0, + "content": "Following the success of GANs [8], several other deep learning-based methods were proposed for text-to-image synthesis. These include StackGAN [46], AttnGAN [43], and MirrorGAN [24], which further improve image quality and diversity by introducing novel architectures and attention mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Later, with the advancement of transformers [38], new approaches emerged for text-to-image synthesis. Being a 12-billion-parameter transformer model, Dall-E [27] introduces a two-stage training process: First, it generates image tokens, which later are combined with text tokens for joint training of an autoregressive model. Later Parti [45] proposed a method to generate content-rich images with multiple objects. Make-a-Scene [7] enables a control mechanism by segmentation masks for text-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Current approaches build upon diffusion models, thereby taking text-to-image synthesis quality to the next level. GLIDE [23] improved Dall-E by adding classifier-free guidance [13]. Later, Dall-E 2 [26] utilizes the contrastive model CLIP [25]. By means of diffusion processes, (i) a mapping from CLIP text encodings to image encodings, and (ii) a CLIP decoder is obtained. LDM / SD [29] applies a diffusion model on lower-resolution encoded signals of VQ-GAN [6], showing competitive quality with a significant gain in speed and efficiency. Imagen [32] shows" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15955" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "incredible performance in text-to-image synthesis by utilizing large language models for text processing. Versatile Diffusion [44] further unifies text-to-image, image-to-text and variations in a single multi-flow diffusion model. Specialisation of text-to-image models to desired styles can be obtained efficiently via few-shot tuning, e.g. using DreamBoth [31] or Specialist Diffusion [18], which employs text-to-image customized data augmentations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.216, + 0.473, + 0.368 + ], + "angle": 0, + "content": "Because of their great image quality, it is desired to exploit text-to-image models for video generation. However, applying diffusion models in the video domain is not straightforward, especially due to their probabilistic generation procedure, making it difficult to ensure temporal consistency. As we show in our ablation experiments in the appendix (see Fig. 14), our modifications are crucial for temporal consistency in terms of both global scene and background motion, and for the preservation of the foreground object identity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.387, + 0.312, + 0.403 + ], + "angle": 0, + "content": "2.2. Text-to-Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.473, + 0.777 + ], + "angle": 0, + "content": "Text-to-video synthesis is a relatively new research direction. Existing approaches try to leverage autoregressive transformers and diffusion processes for the generation. NUWA [41] introduces a 3D transformer encoder-decoder framework and supports both text-to-image and text-to-video generation. Phenaki [39] introduces a bidirectional masked transformer with a causal attention mechanism that allows the generation of arbitrary-long videos from text prompt sequences. CogVideo [15] extends the text-to-image model CogView 2 [4] by tuning it using a multi-frame-rate hierarchical training strategy to better align text and video clips. Video Diffusion Models (VDM) [14] naturally extend text-to-image diffusion models and train jointly on image and video data. Imagen Video [11] constructs a cascade of video diffusion models and utilizes spatial and temporal super-resolution models to generate high-resolution time-consistent videos. Make-A-Video [33] builds upon a text-to-image synthesis model and leverages video data in an unsupervised manner. Gen-1 [5] extends SD and proposes a structure and content-guided video editing method based on visual or textual descriptions of desired outputs. Tune-A-Video [42] proposes a new task of one-shot video generation by extending and tuning SD on a single reference video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Unlike the methods mentioned above, our approach is completely training-free, does not require massive computing power or dozens of GPUs, which makes the video generation process affordable for everyone. In this respect, Tunea-Video [42] comes closest to our work, as it reduces the necessary computations to tuning on only one video. However, it still requires an optimization process and its generating ability is heavily restricted by the reference video." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.593, + 0.107 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.268 + ], + "angle": 0, + "content": "We start this section with a brief introduction of diffusion models, particularly Stable Diffusion (SD) [29]. Then we introduce the problem formulation of zero-shot text-to-video synthesis, followed by a subsection presenting our approach. After that, to show the universality of our method, we use it in combination with ControlNet [47] and Dream-Booth [31] diffusion models for generating conditional and specialized videos. Later we demonstrate the power of our approach with the application of instruction-guided video editing, namely, Video Instruct-Pix2Pix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.275, + 0.662, + 0.29 + ], + "angle": 0, + "content": "3.1. Stable Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.895, + 0.405 + ], + "angle": 0, + "content": "SD is a diffusion model operating in the latent space of an autoencoder \\(\\mathcal{D}(\\mathcal{E}(\\cdot))\\), namely VQ-GAN [6] or VQ-VAE [37], where \\(\\mathcal{E}\\) and \\(\\mathcal{D}\\) are the corresponding encoder and decoder, respectively. More precisely, if \\(x_0 \\in \\mathbb{R}^{h \\times w \\times c}\\) is the latent tensor of an input image \\(\\mathrm{Im}\\) given by the autoencoder, i.e. \\(x_0 = \\mathcal{E}(\\mathrm{Im})\\), diffusion forward process iteratively adds Gaussian noise to the signal \\(x_0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.413, + 0.892, + 0.432 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I\\right), t = 1,.., T \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.895, + 0.517 + ], + "angle": 0, + "content": "where \\( q(x_{t}|x_{t - 1}) \\) is the conditional density of \\( x_{t} \\) given \\( x_{t - 1} \\), and \\( \\{\\beta_t\\}_{t = 1}^T \\) are hyperparameters. \\( T \\) is chosen to be as large that the forward process completely destroys the initial signal \\( x_0 \\) resulting in \\( x_{T}\\sim \\mathcal{N}(0,I) \\). The goal of SD is then to learn a backward process" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.525, + 0.892, + 0.543 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.895, + 0.61 + ], + "angle": 0, + "content": "for \\( t = T, \\dots, 1 \\), which allows to generate a valid signal \\( x_0 \\) from the standard Gaussian noise \\( x_T \\). To get the final image generated from \\( x_T \\) it remains to pass \\( x_0 \\) to the decoder of the initially chosen autoencoder: \\( \\operatorname{Im} = \\mathcal{D}(x_0) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.895, + 0.657 + ], + "angle": 0, + "content": "After learning the abovementioned backward diffusion process (see DDPM [12]) one can apply a deterministic sampling process, called DDIM [35]:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.664, + 0.892, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\left(\\frac {x _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} ^ {t} \\left(x _ {t}\\right)}{\\sqrt {\\alpha_ {t}}}\\right) + \\tag {3} \\\\ \\sqrt {1 - \\alpha_ {t - 1}} e _ {\\theta} ^ {t} (x _ {t}), \\quad t = T, \\dots , 1, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.729, + 0.71, + 0.748 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} = \\prod_{i = 1}^{t}(1 - \\beta_{i})\\) and" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.754, + 0.892, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\theta} ^ {t} (x _ {t}) = \\frac {\\sqrt {1 - \\alpha_ {t}}}{\\beta_ {t}} x _ {t} + \\frac {(1 - \\beta_ {t}) (1 - \\alpha_ {t})}{\\beta_ {t}} \\mu_ {\\theta} (x _ {t}, t). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.843 + ], + "angle": 0, + "content": "To get a text-to-image synthesis framework, SD guides the diffusion processes with a textual prompt \\(\\tau\\). Particularly for DDIM sampling, we get:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.849, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\left(\\frac {x _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} ^ {t} \\left(x _ {t} , \\tau\\right)}{\\sqrt {\\alpha_ {t}}}\\right) + \\tag {5} \\\\ \\sqrt {1 - \\alpha_ {t - 1}} \\epsilon_ {\\theta} ^ {t} (x _ {t}, \\tau), \\quad t = T, \\ldots , 1. \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15956" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.228 + ], + "angle": 0, + "content": "It is worth noting that in SD, the function \\(\\epsilon_{\\theta}^{t}(x_{t},\\tau)\\) is modeled as a neural network with a UNet-like [30] architecture composed of convolutional and (self- and cross-) attentional blocks. \\(x_{T}\\) is called the latent code of the signal \\(x_0\\) and there is a method [3] to apply a deterministic forward process to reconstruct the latent code \\(x_{T}\\) given a signal \\(x_0\\). This method is known as DDIM inversion. Sometimes for simplicity, we will call \\(x_{t},t = 1,\\ldots ,T\\) also the latent codes of the initial signal \\(x_0\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.25, + 0.47, + 0.266 + ], + "angle": 0, + "content": "3.2. Zero-Shot Text-to-Video Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.47, + 0.445 + ], + "angle": 0, + "content": "Existing text-to-video synthesis methods require either costly training on a large-scale (ranging from \\(1M\\) to \\(15M\\) data-points) text-video paired data [41, 15, 39, 11, 14, 5] or tuning on a reference video [42]. To make video generation cheaper and easier, we propose a new problem: zero-shot text-to-video synthesis. Formally, given a text description \\(\\tau\\) and a positive integer \\(m\\in \\mathbb{N}\\), the goal is to design a function \\(\\mathcal{F}\\) that outputs video frames \\(\\mathcal{V}\\in \\mathbb{R}^{m\\times H\\times W\\times 3}\\) (for predefined resolution \\(H\\times W\\)) that exhibit temporal consistency. To determine the function \\(\\mathcal{F}\\), no training or fine-tuning must be performed on a video dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.471, + 0.51 + ], + "angle": 0, + "content": "Our problem formulation provides a new paradigm for text-to-video. Noticeably, a zero-shot text-to-video method naturally benefits from quality improvements of text-to-image models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.532, + 0.175, + 0.546 + ], + "angle": 0, + "content": "3.3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.756 + ], + "angle": 0, + "content": "In this paper, we approach the zero-shot text-to-video task by exploiting the text-to-image synthesis power of Stable Diffusion (SD). As we need to generate videos instead of images, SD should operate on sequences of latent codes. The naive approach is to independently sample \\( m \\) latent codes from standard Gaussian distribution \\( x_{T}^{1},\\ldots ,x_{T}^{m}\\sim \\mathcal{N}(0,I) \\) and apply DDIM sampling to obtain the corresponding tensors \\( x_0^k \\) for \\( k = 1,\\dots,m \\), followed by decoding to obtain the generated video sequence \\( \\{\\mathcal{D}(x_0^k)\\}_{k = 1}^m\\in \\mathbb{R}^{m\\times H\\times W\\times 3} \\). However, this leads to completely random generation of images sharing only the semantics described by \\( \\tau \\) but neither object appearance nor motion coherence (see appendix Fig. 14, first row)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.76, + 0.469, + 0.867 + ], + "angle": 0, + "content": "To address this issue, we propose to (i) introduce motion dynamics between the latent codes \\( x_{T}^{1}, \\ldots, x_{T}^{m} \\) to keep the global scene time consistent and (ii) use cross-frame attention mechanism to preserve the appearance and the identity of the foreground object. Each of the components of our method are described below in detail. The overview of our method can be found in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Note, to simplify notation, we will denote the entire sequence of latent codes by \\( x_{T}^{1:m} = [x_{T}^{1}, \\ldots, x_{T}^{m}] \\)." + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.808, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Motion dynamics in latent codes" + }, + { + "type": "algorithm", + "bbox": [ + 0.502, + 0.11, + 0.894, + 0.31 + ], + "angle": 0, + "content": "Require: \\(\\Delta t\\geq 0,m\\in \\mathbb{N},\\lambda >0,\\delta = (\\delta_x,\\delta_y)\\in\\) \\(\\mathbb{R}^2\\) ,Stable Diffusion (SD) \n1: \\(x_{T}^{1}\\sim \\mathcal{N}(0,I)\\) random sample the first latent code \n2: \\(x_{T^{\\prime}}^{1}\\gets \\mathrm{DDIM\\_Backward}(x_{T}^{1},\\Delta t,SD)\\triangleright\\) perform \\(\\Delta t\\) backward steps by SD \n3: for all \\(k = 2,3,\\ldots ,m\\) do \n4: \\(\\delta^k\\gets \\lambda \\cdot (k - 1)\\delta \\triangleright\\) computing global translation vectors \n5: \\(W_{k}\\leftarrow\\) Warping by \\(\\delta^k\\triangleright\\) defining warping functions \n6: \\(\\tilde{x}_{T^{\\prime}}^{k}\\gets W_{k}(x_{T^{\\prime}}^{1})\\) \n7: \\(x_{T}^{k}\\gets \\mathrm{DDPM\\_Forward}(\\tilde{x}_{T^{\\prime}}^{k},\\Delta t)\\triangleright \\mathrm{DDPM~forward}\\) for more motion freedom return \\(x_{T}^{1:m}\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.335, + 0.794, + 0.35 + ], + "angle": 0, + "content": "3.3.1 Motion Dynamics in Latent Codes" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.359, + 0.893, + 0.421 + ], + "angle": 0, + "content": "Instead of sampling the latent codes \\( x_{T}^{1:m} \\) randomly and independently from the standard Gaussian distribution, we construct them by performing the following steps (see also Alg. 1 and Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.429, + 0.89, + 0.461 + ], + "angle": 0, + "content": "1. Randomly sample the latent code of the first frame: \\( x_{T}^{1} \\sim \\mathcal{N}(0, I) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.892, + 0.515 + ], + "angle": 0, + "content": "2. Perform \\(\\Delta t \\geq 0\\) DDIM backward steps on the latent code \\(x_{T}^{1}\\) by using the SD model and get the corresponding latent \\(x_{T'}^{1}\\), where \\(T' = T - \\Delta t\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.892, + 0.57 + ], + "angle": 0, + "content": "3. Define a direction \\(\\delta = (\\delta_x, \\delta_y) \\in \\mathbb{R}^2\\) for the global scene and camera motion. By default \\(\\delta\\) can be the main diagonal direction \\(\\delta_x = \\delta_y = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.892, + 0.639 + ], + "angle": 0, + "content": "4. For each frame \\( k = 1,2,\\ldots ,m \\) we want to generate, compute the global translation vector \\( \\delta^k = \\lambda \\cdot (k - 1)\\delta \\), where \\( \\lambda \\) is a hyperparameter controlling the amount of the global motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.89, + 0.68 + ], + "angle": 0, + "content": "5. Apply the constructed motion (translation) flow \\(\\delta^{1:m}\\) to \\(x_{T'}^1\\), denote the resulting sequence by \\(\\tilde{x}_{T'}^{1:m}\\):" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.429, + 0.892, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.688, + 0.891, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\tilde {x} _ {T ^ {\\prime}} ^ {k} = W _ {k} \\left(x _ {T ^ {\\prime}} ^ {1}\\right) \\text {f o r} k = 1, 2, \\dots , m, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.716, + 0.89, + 0.747 + ], + "angle": 0, + "content": "where \\( W_{k}(x_{T^{\\prime}}^{1}) \\) is the warping operation for translation by the vector \\( \\delta^{k} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.89, + 0.788 + ], + "angle": 0, + "content": "6. Perform \\(\\Delta t\\) DDPM forward steps on each of the latents \\(\\tilde{x}_{T'}^{2:m}\\) and get the corresponding latent codes \\(x_{T}^{2:m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Then we take the sequence \\( x_{T}^{1:m} \\) as the starting point of the backward (video) diffusion process. As a result, the latent codes generated with our proposed motion dynamics lead to better temporal consistency of the global scene as well as the background, see in the appendix, Sect. 8.1. Yet, the initial latent codes are not constraining enough to describe particular colors, identities or shapes, thus still" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15957" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.878, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.271, + 0.896, + 0.455 + ], + "angle": 0, + "content": "Figure 2: Method overview: Starting from a randomly sampled latent code \\( x_{T}^{1} \\), we apply \\( \\Delta t \\) DDIM backward steps to obtain \\( x_{T'}^{1} \\) using a pre-trained Stable Diffusion model (SD). A specified motion field results for each frame \\( k \\) in a warping function \\( W_{k} \\) that turns \\( x_{T'}^{1} \\) to \\( x_{T'}^{k} \\). By enhancing the latent codes with motion dynamics, we determine the global scene and camera motion and achieve temporal consistency in the background and the global scene. A subsequent DDPM forward application delivers latent codes \\( x_{T}^{k} \\) for \\( k = 1, \\ldots, m \\). By using the (probabilistic) DDPM method, a greater degree of freedom is achieved with respect to the motion of objects (see appendix Sec. 8.1). Finally, the latent codes are passed to our modified SD model using the proposed cross-frame attention, which uses keys and values from the first frame to generate the image of frame \\( k = 1, \\ldots, m \\). By using cross-frame attention, the appearance and the identity of the foreground object are preserved throughout the sequence. Optionally, we apply background smoothing. To this end, we employ salient object detection to obtain for each frame \\( k \\) a mask \\( M^{k} \\) indicating the foreground pixels. Finally, for the background (using the mask \\( M^{k} \\)), a convex combination between the latent code \\( x_{t}^{1} \\) of frame one warped to frame \\( k \\) and the latent code \\( x_{t}^{k} \\) is used to further improve the temporal consistency of the background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.471, + 0.511 + ], + "angle": 0, + "content": "leading to temporal inconsistencies, especially for the foreground object." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.414, + 0.543 + ], + "angle": 0, + "content": "3.3.2 Reprogramming Cross-Frame Attention" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.612 + ], + "angle": 0, + "content": "To address the issue mentioned above, we use a cross-frame attention mechanism to preserve the information about (in particular) the foreground object's appearance, shape, and identity throughout the generated video." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.471, + 0.764 + ], + "angle": 0, + "content": "To leverage the power of cross-frame attention and at the same time exploit a pretrained SD without retraining, we replace each of its self-attention layers with a cross-frame attention, with the attention for each frame being on the first frame. More precisely in the original SD UNet architecture \\(\\epsilon_{\\theta}^{t}(x_{t},\\tau)\\), each self-attention layer takes a feature map \\(x\\in \\mathbb{R}^{h\\times w\\times c}\\), linearly projects it into query, key, value features \\(Q,K,V\\in \\mathbb{R}^{h\\times w\\times c}\\), and computes the layer output by the following formula (for simplicity described here for only one attention head) [38]:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.77, + 0.47, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S e l f - A t t n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {c}}\\right) V. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.87 + ], + "angle": 0, + "content": "In our case, each attention layer receives \\(m\\) inputs: \\(x^{1:m} = [x^1,\\dots,x^m]\\in \\mathbb{R}^{m\\times h\\times w\\times c}\\). Hence, the linear projection layers produce \\(m\\) queries, keys, and values \\(Q^{1:m},K^{1:m},V^{1:m}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Therefore, we replace each self-attention layer with a cross-frame attention of each frame on the first frame as" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.481, + 0.559, + 0.494 + ], + "angle": 0, + "content": "follows:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.504, + 0.892, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {C r o s s - F r a m e - A t t n} \\left(Q ^ {k}, K ^ {1: m}, V ^ {1: m}\\right) = \\\\ \\operatorname {S o f t m a x} \\left(\\frac {Q ^ {k} \\left(K ^ {1}\\right) ^ {T}}{\\sqrt {c}}\\right) V ^ {1} \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.893, + 0.661 + ], + "angle": 0, + "content": "for \\( k = 1, \\ldots, m \\). By using cross frame attention, the appearance and structure of the objects and background as well as identities are carried over from the first frame to subsequent frames, which significantly increases the temporal consistency of the generated frames (see in the appendix the Figures 14, 16, 22, 23)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.795, + 0.696 + ], + "angle": 0, + "content": "3.3.3 Background smoothing (Optional)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.78 + ], + "angle": 0, + "content": "We further improve temporal consistency of the background using a convex combination of background-masked latent codes between the first frame and frame \\( k \\). This helps especially to generate videos from textual prompts when one or no initial image and no further guidance are provided." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.871 + ], + "angle": 0, + "content": "In more detail, given the generated sequence of our video generator, \\( x_0^{1:m} \\), we apply (an in-house solution for) salient object detection [40] to the decoded images to obtain a corresponding foreground mask \\( M^k \\) for each frame \\( k \\). Then we warp \\( x_t^1 \\) according to the employed motion dynamics defined by \\( W_k \\) and denote the result by \\( \\hat{x}_t^k \\coloneqq W_k(x_t^1) \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Background smoothing is achieved by a convex combination between the actual latent code \\( x_{t}^{k} \\) and the warped" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15958" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.089, + 0.855, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.446, + 0.893, + 0.478 + ], + "angle": 0, + "content": "Figure 3: Text-to-Video results of our method. Depicted frames show that identities and appearances are temporally consistent and fitting to the textual prompt. For more results, see Appendix Sec. 8." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.5, + 0.462, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.698, + 0.466, + 0.714 + ], + "angle": 0, + "content": "Figure 4: The overview of Text2Video-Zero + ControlNet" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.74, + 0.334, + 0.757 + ], + "angle": 0, + "content": "latent code \\(\\hat{x}_t^k\\) on the background, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.765, + 0.469, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\bar {x} _ {t} ^ {k} = M ^ {k} \\odot x _ {t} ^ {k} + (1 - M ^ {k}) \\odot (\\alpha \\hat {x} _ {t} ^ {k} + (1 - \\alpha) x _ {t} ^ {k}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "for \\( k = 1, \\ldots, m \\), where \\( \\alpha \\) is a hyperparameter, which we empirically choose \\( \\alpha = 0.6 \\). Finally, DDIM sampling is employed on \\( \\overline{x}_t^k \\), which delivers video generation with background smoothing. We use background smoothing in our video generation from text when no guidance is provided. For an ablation study on background smoothing, see the appendix, Sec. 8.1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.503, + 0.86, + 0.52 + ], + "angle": 0, + "content": "3.4. Conditional and Specialized Text-to-Video" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Recently powerful controlling mechanisms [47, 22, 17] emerged to guide the diffusion process for text-to-image generation. Particularly, ControlNet [47] enables to condition the generation process using edges, pose, semantic masks, image depths, etc. However, a direct application of ControlNet in the video domain leads to temporal inconsistencies and to severe changes of object appearance, identity, and the background (see in the appendix Figures 14, 16, 22, 23). It turns out that our modifications on the basic diffusion process for videos result in more consistent videos guided by ControlNet conditions. We would like to point out again that our method does not require any fine-tuning or optimization processes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.893, + 0.882 + ], + "angle": 0, + "content": "More specifically, ControlNet creates a trainable copy of the encoder (including the middle blocks) of the UNet \\(\\epsilon_{\\theta}^{t}(x_{t},\\tau)\\) while additionally taking the input \\(x_{t}\\) and a condition \\(c\\), and adds the outputs of each layer to the skip-connections of the original UNet. Here \\(c\\) can be any type of condition, such as edge map, scribbles, pose (body landmarks), depth map, segmentation map, etc. The trainable branch is being trained on a specific domain for each type of the condition \\(c\\) resulting in an effective conditional text-to-image generation mechanism." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To guide our video generation process with ControlNet" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15959" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.437, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.153, + 0.373, + 0.165 + ], + "angle": 0, + "content": "(a) a panda dancing in Antarctica" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.17, + 0.437, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.234, + 0.381, + 0.246 + ], + "angle": 0, + "content": "(b) a cyborg dancing near a volcano" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.251, + 0.437, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.315, + 0.399, + 0.328 + ], + "angle": 0, + "content": "(c) an astronaut dancing in the outer space" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.335, + 0.47, + 0.366 + ], + "angle": 0, + "content": "Figure 5: Conditional generation with pose control. For more results see appendix, Sec. 10." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.392, + 0.437, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.456, + 0.396, + 0.469 + ], + "angle": 0, + "content": "(a) oil painting of a girl dancing close-up" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.473, + 0.437, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.537, + 0.414, + 0.55 + ], + "angle": 0, + "content": "(b) Cyberpunk boy with a hat dancing close-up" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.557, + 0.47, + 0.588 + ], + "angle": 0, + "content": "Figure 6: Conditional generation with edge control. For more results see appendix, Sec. 9." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.749 + ], + "angle": 0, + "content": "we apply our method to the basic diffusion process, i.e. enrich the latent codes \\( x_{T}^{1:m} \\) with motion information and change the self-attention into cross-frame attentions in the main UNet. While adopting the main UNet for video generation task, we apply the ControlNet pretrained copy branch per-frame on each \\( x_{t}^{k} \\) for \\( k = 1,\\dots,m \\) in each diffusion time-step \\( t = T,\\ldots ,1 \\) and add the ControlNet branch outputs to the skip-connections of the main UNet." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.769, + 0.469, + 0.83 + ], + "angle": 0, + "content": "Furthermore, for our conditional generation task, we adopted the weights of specialized DreamBooth (DB) [31] models1. This gives us specialized time-consistent video generations (see Fig. 7)." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.089, + 0.86, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.152, + 0.83, + 0.166 + ], + "angle": 0, + "content": "(a) oil painting of a beautiful girl avatar style" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.17, + 0.86, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.234, + 0.74, + 0.247 + ], + "angle": 0, + "content": "(b) gta-5 style" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.254, + 0.892, + 0.283 + ], + "angle": 0, + "content": "Figure 7: Conditional generation with edge control and DB models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.312, + 0.713, + 0.327 + ], + "angle": 0, + "content": "3.5. Video Instruct-Pix2Pix" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.336, + 0.892, + 0.547 + ], + "angle": 0, + "content": "With the rise of text-guided image editing methods such as Prompt2Prompt [9], Instruct-Pix2Pix [2], SDEdit [20], etc., text-guided video editing approaches emerged [1, 16, 42]. While these methods require complex optimization processes, our approach enables the adoption of any SD-based text-guided image editing algorithm to the video domain without any training or fine-tuning. Here we take the text-guided image editing method Instruct-Pix2Pix and combine it with our approach. More precisely, we change the self-attention mechanisms in Instruct-Pix2Pix to cross-frame attentions according to Eq. 8. Our experiments show that this adaptation significantly improves the consistency of the edited videos (see Fig. 9) over the naive per-frame usage of Instruct-Pix2Pix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.563, + 0.633, + 0.58 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.588, + 0.718, + 0.604 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.793 + ], + "angle": 0, + "content": "We take the Stable Diffusion [29] code2 with its pretrained weights from version 1.5 as basis and implement our modifications. For each video, we generate \\( m = 8 \\) frames with \\( 512 \\times 512 \\) resolution. However, our framework allows generating any number of frames, either by increasing \\( m \\), or by employing our method in an auto-regressive fashion where the last generated frame \\( m \\) becomes the first frame in computing the next \\( m \\) frames. For all text-to-video generation experiments, we take \\( T' = 881 \\), \\( T = 941 \\) without specific tuning, while for conditional and specialized generation, and for Video Instruct-Pix2Pix, we take \\( T' = T = 1000 \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.794, + 0.892, + 0.84 + ], + "angle": 0, + "content": "For a conditional generation, we use the codebase3 of ControlNet [47]. For specialized models, we take DB [31] models from publicly available sources. For Video Instruct" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.863, + 0.468, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Avatar model: https://civitai.com/models/9968/ avatar-style. GTA-5 model: https://civitai.com/models/1309/gta5-artwork-diffusion." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.851, + 0.892, + 0.887 + ], + "angle": 0, + "content": "2https://github.com/huggingface/diffusers. We also benefit from the codebase of Tune-A-Video https://github.com/showlab/Tune-A-Video." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.888, + 0.844, + 0.9 + ], + "angle": 0, + "content": "3https://github.com/lllyasviel/ControlNet." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.851, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15960" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.091, + 0.437, + 0.108 + ], + "angle": 0, + "content": "Pix2Pix, we use the codebase of Instruct Pix2Pix [2]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.116, + 0.262, + 0.133 + ], + "angle": 0, + "content": "4.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.14, + 0.47, + 0.216 + ], + "angle": 0, + "content": "All applications of Text2Video-Zero show that it successfully generates videos where the global scene and the background are time consistent and the context, appearance, and identity of the foreground object are maintained throughout the entire sequence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.469, + 0.337 + ], + "angle": 0, + "content": "In the case of text-to-video, we observe that it generates high-quality videos that are well-aligned to the text prompt (see Fig. 3 and the appendix). For instance, the depicted panda shows a naturally walking on the street. Likewise, using additional guidance from edges or poses (see Fig. 5, Fig. 6 and Fig. 7 and the appendix), high quality videos are generated matching the prompt and the guidance that show great temporal consistency and identity preservation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.469, + 0.382 + ], + "angle": 0, + "content": "Videos generated by Video Instruct-Pix2Pix (see Fig. 1 and the appendix) possess high-fidelity with respect to the input video, while following closely the instruction." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.393, + 0.323, + 0.409 + ], + "angle": 0, + "content": "4.3. Comparison with Baselines" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.416, + 0.469, + 0.492 + ], + "angle": 0, + "content": "We compare our method with two publicly available baselines: CogVideo [15] and Tune-A-Video [42]. Since CogVideo is a text-to-video method we compare with it in pure text-guided video synthesis settings. With Tune-A-Video we compare in our Video Instruct-Pix2Pix setting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.511, + 0.31, + 0.526 + ], + "angle": 0, + "content": "4.3.1 Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.469, + 0.656 + ], + "angle": 0, + "content": "To show quantitative results, we evaluate the CLIP score [10], which indicates video-text alignment. We randomly take 25 videos generated by CogVideo and synthesize corresponding videos using the same prompts according to our method. The CLIP scores for our method and CogVideo are 31.19 and 29.63, respectively. Our method thus slightly outperforms CogVideo, even though the latter has 9.4 billion parameters and requires large-scale training on videos." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.676, + 0.3, + 0.691 + ], + "angle": 0, + "content": "4.3.2 Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.469, + 0.865 + ], + "angle": 0, + "content": "We present several results of our method in Fig. 8 and provide a qualitative comparison to CogVideo [15]. Both methods show good temporal consistency throughout the sequence, preserving the identity of the object and background. However, our method shows better text-video alignment. For instance, while our method correctly generates a video of a man riding a bicycle in the sunshine in Fig. 8(b), CogVideo sets the background to moon light. Also in Fig. 8(a), our method correctly shows a man running in the snow, while neither the snow nor a man running are clearly visible in the video generated by CogVideo." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "Qualitative results of Video Instruct-Pix2Pix and a visual comparison with per-frame Instruct-Pix2Pix and Tune-A-Video are shown in Fig. 9. While Instruct-Pix2Pix shows a good editing performance per frame, it lacks temporal consistency. This becomes evident especially in the video depicting a skiing person, where the snow and the sky are drawn using different styles and colors. Using our Video Instruct-Pix2Pix method, these issues are solved resulting in temporally consistent video edits throughout the sequence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.893, + 0.485 + ], + "angle": 0, + "content": "While Tune-A-Video creates temporally consistent video generations, it is less aligned to the instruction guidance than our method, struggles creating local edits and losses details of the input sequence. This becomes apparent when looking at the edit of the dancer video depicted in Fig. 9 (left side). In contrast to Tune-A-Video, our method draws the entire dress brighter and at the same time better preserves the background, e.g. the wall behind the dancer is almost kept the same. Tune-A-Video draws a severely modified wall. Moreover, our method is more faithful to the input details, e.g., Video Instruct-Pix2Pix draws the dancer using the pose exactly as provided (Fig. 9 left), and shows all skiing persons appearing in the input video (compare last frame of Fig. 9(right)), in constraint to Tune-A-Video. All the above-mentioned weaknesses of Tune-A-Video can also be observed in our additional evaluations that are provided in the appendix, Figures 25, 26." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.653, + 0.511 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.609 + ], + "angle": 0, + "content": "We perform several ablation studies and provide the results in the appendix, Sect. 8.1. Namely, we analyse background smoothing, \\(\\Delta t\\), and two main components of our method: making the initial latent codes coherent to a motion, and using cross-frame attention on the first frame instead of self-attention." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.624, + 0.774, + 0.64 + ], + "angle": 0, + "content": "5. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.786 + ], + "angle": 0, + "content": "The main limitation of this work is the inability to generate longer videos with sequences of actions. Future research may target enriching our method with techniques such as autoregressive scene action generation, while keeping the training-free spirit. Overall, our method is focused on generating video key-frames as in the first stage of Imagen Video [11] and Make-A-Video [33], and can thus be considered as good basis for longer and smoother video generation by integrating temporal upsampling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.619, + 0.815 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we addressed the problem of zero-shot text-to-video synthesis and proposed a novel method for time-consistent video generation. Our approach does not require any optimization or fine-tuning, making text-to-video generation and its applications affordable for everyone. We" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.363, + 0.901 + ], + "angle": 0, + "content": "4https://github.com/timothybrooks/instruct-pix2pix." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15961" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.088, + 0.212, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.089, + 0.301, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.089, + 0.387, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.089, + 0.475, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.089, + 0.583, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.089, + 0.671, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.089, + 0.758, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.089, + 0.845, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.157, + 0.581, + 0.17 + ], + "angle": 0, + "content": "(a) a man is running in the snow" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.174, + 0.212, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.213, + 0.174, + 0.3, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.174, + 0.387, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.174, + 0.475, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.174, + 0.582, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.174, + 0.67, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.174, + 0.758, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.174, + 0.846, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.243, + 0.615, + 0.256 + ], + "angle": 0, + "content": "(b) a man is riding a bicycle in the sunshine" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.26, + 0.212, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.213, + 0.26, + 0.3, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.26, + 0.387, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.26, + 0.475, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.26, + 0.583, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.26, + 0.67, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.26, + 0.758, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.26, + 0.846, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.329, + 0.578, + 0.342 + ], + "angle": 0, + "content": "(c) a man is walking in the rain" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.349, + 0.892, + 0.38 + ], + "angle": 0, + "content": "Figure 8: Comparison of our method vs CogVideo on text-to-video generation task (left is ours, right is CogVideo [15]). For more comparisons, see appendix Fig. 18." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.406, + 0.891, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.659, + 0.89, + 0.69 + ], + "angle": 0, + "content": "Figure 9: Comparison of Video Instruct-Pix2Pix(ours) with Tune-A-Video and per-frame Instruct-Pix2Pix. For more comparisons see our appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.473, + 0.897 + ], + "angle": 0, + "content": "demonstrated the effectiveness of our method for various applications, including conditional and specialized video generation, and Video Instruct-Pix2Pix, i.e., instruction-guided video editing. Our contributions to the field include presenting a new problem of zero-shot text-to-video synthesis, showing the utilization of text-to-image diffusion models for generating time-consistent videos, and providing evidence of the effectiveness of our method for various video synthesis applications. We believe that our proposed method will open up new possibilities for video generation and editing, making it accessible and affordable for everyone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.895, + 0.869 + ], + "angle": 0, + "content": "Acknowledgments. This material is partially based upon work supported by the National Science Foundation CAREER Award #2239840, and the National AI Institute for Exceptional Education (Award #2229873) by National Science Foundation and the Institute of Education Sciences, U.S. Department of Education. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation, the Institute of Education Sciences, or the U.S. Department of Education." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15962" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.47, + 0.228 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2, 7, 8, 13, 25, 26" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[3] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.47, + 0.327 + ], + "angle": 0, + "content": "[4] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.47, + 0.386 + ], + "angle": 0, + "content": "[5] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.47, + 0.444 + ], + "angle": 0, + "content": "[6] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.47, + 0.528 + ], + "angle": 0, + "content": "[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 89-106. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.47, + 0.585 + ], + "angle": 0, + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.587, + 0.47, + 0.642 + ], + "angle": 0, + "content": "[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.644, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[10] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.701, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[11] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.772, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.816, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[13] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.846, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[15] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3, 4, 8, 9, 12, 13, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.203 + ], + "angle": 0, + "content": "[16] Yao-Chih Lee, Ji-Ze Genevieve Jang, Yi-Ting Chen, Elizabeth Qiu, and Jia-Bin Huang. Shape-aware text-driven layered video editing. arXiv e-prints, pages arXiv-2301, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.894, + 0.287 + ], + "angle": 0, + "content": "[17] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.893, + 0.383 + ], + "angle": 0, + "content": "[18] Haoming Lu, Hazarapet Tunanyan, Kai Wang, Shant Navasardyan, Zhangyang Wang, and Humphrey Shi. Specialist diffusion: Plug-and-play sample-efficient fine-tuning of text-to-image diffusion models to learn any unseen style. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14267-14276, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.893, + 0.426 + ], + "angle": 0, + "content": "[19] Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Generating images from captions with attention. arXiv preprint arXiv:1511.02793, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.893, + 0.481 + ], + "angle": 0, + "content": "[20] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.893, + 0.538 + ], + "angle": 0, + "content": "[21] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.893, + 0.594 + ], + "angle": 0, + "content": "[22] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.893, + 0.662 + ], + "angle": 0, + "content": "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.893, + 0.732 + ], + "angle": 0, + "content": "[24] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. Mirrorgan: Learning text-to-image generation by redescription. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1514, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.893, + 0.814 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.893, + 0.871 + ], + "angle": 0, + "content": "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[27] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15963" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.133 + ], + "angle": 0, + "content": "Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[28] Scott E Reed, Zeynep Akata, Santosh Mohan, Samuel Tenka, Bernt Schiele, and Honglak Lee. Learning what and where to draw. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.193, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.47, + 0.347 + ], + "angle": 0, + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3, 7, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[33] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[34] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[36] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[37] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "textual description. arXiv preprint arXiv:2210.02399, 2022. 3,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[40] Wenguan Wang, Qiuxia Lai, Huazhu Fu, Jianbing Shen, Haibin Ling, and Ruigang Yang. Salient object detection in the deep learning era: An in-depth survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3239-3259, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[41] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Niwa: Visual synthesis pretraining for neural visual world creation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 720-736. Springer, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[42] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2, 3, 4, 7, 8, 12, 13, 26" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.892, + 0.429 + ], + "angle": 0, + "content": "[43] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.431, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[44] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.488, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[45] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.559, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[46] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.642, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[47] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15964" + } + ] +] \ No newline at end of file diff --git a/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/56c5c84c-dd92-4857-8492-3e83266b25fc_origin.pdf b/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/56c5c84c-dd92-4857-8492-3e83266b25fc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0ea6850a3017c852cec716d92344f70eea9b27af --- /dev/null +++ b/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/56c5c84c-dd92-4857-8492-3e83266b25fc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e40aeef748adf4bd4f2d2afec1c0096d0120924b3b4a719b294cf4ed48480dd0 +size 3285961 diff --git a/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/full.md b/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/full.md new file mode 100644 index 0000000000000000000000000000000000000000..28c3239f8ea9ac7cb8a6d42d9a467d846656d29a --- /dev/null +++ b/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/full.md @@ -0,0 +1,449 @@ +# Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators + +Levon Khachatryan $^{1*}$ Andranik Movsisyan $^{1*}$ Vahram Tadevosyan $^{1*}$ Roberto Henschel $^{1*}$ Zhangyang Wang $^{1,2}$ Shant Navasardyan $^{1}$ Humphrey Shi $^{1,3}$ $^{1}$ Picsart AI Research (PAIR) $^{2}$ UT Austin $^{3}$ SHI Labs @ Georgia Tech, Oregon & UIUC https://github.com/Picsart-AI-Research/Text2Video-Zero + +![](images/5638995bdf120554a11b3d2bc540caa4ab3e8a7c2868cc2d7e7fe856b11f0784.jpg) + +![](images/a4d8d0a3ceca72cf40b495bea7ac0f26329a5b6dd01e1531cc72073f7a0e8251.jpg) + +![](images/aeeb8c18a627e667284dee6ab977731786cf2633a9eed02ccbf5f00d10389a6c.jpg) + +![](images/6e8e563dfa608ee4bb05d24195d74a25e40bca7cd71d46b792d52b0c2e144d75.jpg) + +![](images/f9323419c2c537317827585f68dc62e75a08942e2c5bf67a185596a625edd231.jpg) + +![](images/2cb5b3ba5892262e4f2653806a92caf658daf7fefbe529caef50e48a01a34e66.jpg) + +![](images/f3c8eb8cd473b95a423a37cfe89749e5874467222ce30246c27a17fd6e964d98.jpg) + +![](images/1ef3f447b8f14833e105f220bbfb65c77cab36a0b3a985334e5aab9fab31285f.jpg) + +Text-to-Video generation: "a horse galloping on a street" + +![](images/1d5f7e351bb85f55105e698774b39c26b3d37d9e15ef182532d35f05c49a2c21.jpg) + +![](images/0ab0779b112b13190a80c90c982d45723d9d92f348aef7d9c95a2ddd267cdd67.jpg) + +![](images/39c461fa8f85041f87e27ab52faf1c89a2be2b45fcdf6acef5962b4e82f8c215.jpg) + +![](images/b943bb548f0ef3e190658ddb374d22225872c963246fef31a653736ec95e6290.jpg) + +![](images/840293918536d238e5f19ca003a506b66fb15b941f5234520443ae8b6ee22c91.jpg) + +![](images/0c3cc4cfe40b25a2fd3542f66e350609066bc3f7b46224fb9cd77b59233c84ac.jpg) + +![](images/635b8291b229f014be55890211a331493e9d02368740bcbfee9f4ef8a6df5412.jpg) + +![](images/819828471aefa4ac20a52e296620fe280800fa82e8aebe8e25ae192f7f83e46c.jpg) + +Text-to-Video generation: "a panda is playing guitar on times square" + +![](images/480ad00bf61e23c3b61d9e02a24e6560632e83738a3d73875f45f49883bf473e.jpg) + +![](images/730b01567f2aae5a32d7af0090815955b6c306e8df495bf04b55b85c0ff9153e.jpg) + +![](images/4911268d744d8638ec96d63d1f9037b9b22f564e3fe6b28558cb7095a5fe6b96.jpg) + +![](images/a3aeaf6679817680885232c903f0b133184877bb2afe91cf73169deb0cbd9d61.jpg) + +![](images/ec210fa5b5451758b917214eced5d7a9fdd264a0af302f4399c16b7d4aef56e0.jpg) + +![](images/f0804966b7b5bc4001e2b961ba987f85799727ca73df2e167ccb3c1aa7d74b89.jpg) + +![](images/9242f762bf79b0af603c475a35eb32e70bcf12774fd5f32c533c470d571ff4ce.jpg) + +![](images/d1a694a11d86ed5c7ab1519115b1566b8ebbe2bde282a4fc4e2f5d8b2749b519.jpg) +Video Instruct-Pix2Pix: "make it Van Gogh Starry Night style" + +![](images/6d3e9c648aaa0481ddedcfc7d7de88c6c123d85a14ba39031da9e1e6b51a7fff.jpg) +Figure 1: Our method Text2Video-Zero enables zero-shot video generation using (i) a textual prompt (see rows 1, 2), (ii) a prompt combined with guidance from poses or edges (see lower right), and (iii) Video Instruct-Pix2Pix, i.e., instruction-guided video editing (see lower left). Results are temporally consistent and follow closely the guidance and textual prompts. + +![](images/8530c75814d28747a705839bf7314dfbf197d044bd9dea2d0915c7ace3a34787.jpg) + +Text-to-Video generation + pose control: "a bear dancing on the concrete" + +![](images/95aabc830f54679227d25a91795bc51a4a902b50cf7b59265807a052c4655bcc.jpg) +Text-to-Video generation + edge control: "white butterfly" + +![](images/81454a49861a03f943a212da4832b983a0defbb449084f3f786ee71a4f2f67c5.jpg) + +![](images/4ec10fefe4c5557de5c9080d4dc0e738df1ec2882fdb1cdafcffde8c06f82e34.jpg) + +![](images/3de2cf00b5845d69926f12d24372e2b3d58591b656c7d36840d8c6e9727f1026.jpg) + +# Abstract + +Recent text-to-video generation approaches rely on computationally heavy training and require large-scale video datasets. In this paper, we introduce a new task, zero-shot text-to-video generation, and propose a low-cost approach (without any training or optimization) by leveraging the power of existing text-to-image synthesis methods (e.g. Stable Diffusion), making them suitable for the video domain. Our key modifications include (i) enriching the la + +tent codes of the generated frames with motion dynamics to keep the global scene and the background time consistent; and (ii) reprogramming frame-level self-attention using a new cross-frame attention of each frame on the first frame, to preserve the context, appearance, and identity of the foreground object. Experiments show that this leads to low overhead, yet high-quality and remarkably consistent video generation. Moreover, our approach is not limited to text-to-video synthesis but is also applicable to other tasks such as conditional and content-specialized video generation, and Video Instruct-Pix2Pix, i.e., instruction-guided + +video editing. As experiments show, our method performs comparably or sometimes better than recent approaches, despite not being trained on additional video data. Our code is publicly available at: https://github.com/Picsart-AI-Research/Text2Video-Zero. + +# 1. Introduction + +In recent years, generative AI has attracted enormous attention in the computer vision community. With the advent of diffusion models [34, 12, 35, 36], it has become tremendously popular and successful to generate high-quality images from textual prompts, also called text-to-image synthesis [26, 29, 32, 7, 44]. Recent works [14, 33, 11, 42, 5, 21] attempt to extend the success to text-to-video generation and editing tasks, by reusing text-to-image diffusion models in the video domain. While such approaches yield promising outcomes, most of them require substantial training with a massive amount of labeled data which can be costly and unaffordable for many users. With the aim of making video generation cheaper, Tune-A-Video [42] introduces a mechanism that can adopt Stable Diffusion (SD) model [29] for the video domain. The training effort is drastically reduced to tuning one video. While that is much more efficient than previous approaches, it still requires an optimization process. In addition, the generation abilities of Tune-A-Video are limited to text-guided video editing applications; video synthesis from scratch, however, remains out of its reach. + +In this paper, we take one step forward in studying the novel problem of zero-shot, "training-free" text-to-video synthesis, which is the task of generating videos from textual prompts without requiring any optimization or finetuning. A key concept of our approach is to modify a pre-trained text-to-image model (e.g., Stable Diffusion), enriching it with temporally consistent generation. By building upon already trained text-to-image models, our method takes advantage of their excellent image generation quality and enhances their applicability to the video domain without performing additional training. To enforce temporal consistency, we present two innovative and lightweight modifications: (1) we first enrich the latent codes of generated frames with motion information to keep the global scene and the background time consistent; (2) we then use cross-frame attention of each frame on the first frame to preserve the context, appearance, and identity of the foreground object throughout the entire sequence. Our experiments show that these simple modifications lead to high-quality and time-consistent video generations (see Fig. 1 and further results in the appendix). Despite the fact that other works train on large-scale video data, our method achieves similar or sometimes even better performance (see Figures 8, 9 and appendix Figures 18, 25, 26). Furthermore, our method is not limited to text-to-video synthesis but is + +also applicable to conditional (see Figures 5,6 and appendix Figures 19, 21, 22, 23) and specialized video generation (see Fig. 7), and instruction-guided video editing, which we refer as Video Instruct-Pix2Pix motivated by Instruct-Pix2Pix [2] (see Fig. 9 and appendix Figures 24, 25, 26). + +Our contributions are summarized as three-folds: + +- A new problem setting of zero-shot text-to-video synthesis, aiming at making text-guided video generation and editing "freely affordable". We use only a pretrained text-to-image diffusion model without any further fine-tuning or optimization. +- Two novel post-hoc techniques to enforce temporally consistent generation, via encoding motion dynamics in the latent codes, and reprogramming each frame's self-attention using a new cross-frame attention. +- A broad variety of applications that demonstrate our method's effectiveness, including conditional and specialized video generation, and Video Instruct-Pix2Pix i.e., video editing by textual instructions. + +# 2. Related Work + +# 2.1. Text-to-Image Generation + +Early approaches to text-to-image synthesis relied on methods such as template-based generation [19] and feature matching [28]. However, these methods were limited in their ability to generate realistic and diverse images. + +Following the success of GANs [8], several other deep learning-based methods were proposed for text-to-image synthesis. These include StackGAN [46], AttnGAN [43], and MirrorGAN [24], which further improve image quality and diversity by introducing novel architectures and attention mechanisms. + +Later, with the advancement of transformers [38], new approaches emerged for text-to-image synthesis. Being a 12-billion-parameter transformer model, Dall-E [27] introduces a two-stage training process: First, it generates image tokens, which later are combined with text tokens for joint training of an autoregressive model. Later Parti [45] proposed a method to generate content-rich images with multiple objects. Make-a-Scene [7] enables a control mechanism by segmentation masks for text-to-image generation. + +Current approaches build upon diffusion models, thereby taking text-to-image synthesis quality to the next level. GLIDE [23] improved Dall-E by adding classifier-free guidance [13]. Later, Dall-E 2 [26] utilizes the contrastive model CLIP [25]. By means of diffusion processes, (i) a mapping from CLIP text encodings to image encodings, and (ii) a CLIP decoder is obtained. LDM / SD [29] applies a diffusion model on lower-resolution encoded signals of VQ-GAN [6], showing competitive quality with a significant gain in speed and efficiency. Imagen [32] shows + +incredible performance in text-to-image synthesis by utilizing large language models for text processing. Versatile Diffusion [44] further unifies text-to-image, image-to-text and variations in a single multi-flow diffusion model. Specialisation of text-to-image models to desired styles can be obtained efficiently via few-shot tuning, e.g. using DreamBoth [31] or Specialist Diffusion [18], which employs text-to-image customized data augmentations. + +Because of their great image quality, it is desired to exploit text-to-image models for video generation. However, applying diffusion models in the video domain is not straightforward, especially due to their probabilistic generation procedure, making it difficult to ensure temporal consistency. As we show in our ablation experiments in the appendix (see Fig. 14), our modifications are crucial for temporal consistency in terms of both global scene and background motion, and for the preservation of the foreground object identity. + +# 2.2. Text-to-Video Generation + +Text-to-video synthesis is a relatively new research direction. Existing approaches try to leverage autoregressive transformers and diffusion processes for the generation. NUWA [41] introduces a 3D transformer encoder-decoder framework and supports both text-to-image and text-to-video generation. Phenaki [39] introduces a bidirectional masked transformer with a causal attention mechanism that allows the generation of arbitrary-long videos from text prompt sequences. CogVideo [15] extends the text-to-image model CogView 2 [4] by tuning it using a multi-frame-rate hierarchical training strategy to better align text and video clips. Video Diffusion Models (VDM) [14] naturally extend text-to-image diffusion models and train jointly on image and video data. Imagen Video [11] constructs a cascade of video diffusion models and utilizes spatial and temporal super-resolution models to generate high-resolution time-consistent videos. Make-A-Video [33] builds upon a text-to-image synthesis model and leverages video data in an unsupervised manner. Gen-1 [5] extends SD and proposes a structure and content-guided video editing method based on visual or textual descriptions of desired outputs. Tune-A-Video [42] proposes a new task of one-shot video generation by extending and tuning SD on a single reference video. + +Unlike the methods mentioned above, our approach is completely training-free, does not require massive computing power or dozens of GPUs, which makes the video generation process affordable for everyone. In this respect, Tunea-Video [42] comes closest to our work, as it reduces the necessary computations to tuning on only one video. However, it still requires an optimization process and its generating ability is heavily restricted by the reference video. + +# 3. Method + +We start this section with a brief introduction of diffusion models, particularly Stable Diffusion (SD) [29]. Then we introduce the problem formulation of zero-shot text-to-video synthesis, followed by a subsection presenting our approach. After that, to show the universality of our method, we use it in combination with ControlNet [47] and Dream-Booth [31] diffusion models for generating conditional and specialized videos. Later we demonstrate the power of our approach with the application of instruction-guided video editing, namely, Video Instruct-Pix2Pix. + +# 3.1. Stable Diffusion + +SD is a diffusion model operating in the latent space of an autoencoder $\mathcal{D}(\mathcal{E}(\cdot))$ , namely VQ-GAN [6] or VQ-VAE [37], where $\mathcal{E}$ and $\mathcal{D}$ are the corresponding encoder and decoder, respectively. More precisely, if $x_0 \in \mathbb{R}^{h \times w \times c}$ is the latent tensor of an input image $\mathrm{Im}$ given by the autoencoder, i.e. $x_0 = \mathcal{E}(\mathrm{Im})$ , diffusion forward process iteratively adds Gaussian noise to the signal $x_0$ : + +$$ +q \left(x _ {t} \mid x _ {t - 1}\right) = \mathcal {N} \left(x _ {t}; \sqrt {1 - \beta_ {t}} x _ {t - 1}, \beta_ {t} I\right), t = 1,.., T \tag {1} +$$ + +where $q(x_{t}|x_{t - 1})$ is the conditional density of $x_{t}$ given $x_{t - 1}$ , and $\{\beta_t\}_{t = 1}^T$ are hyperparameters. $T$ is chosen to be as large that the forward process completely destroys the initial signal $x_0$ resulting in $x_{T}\sim \mathcal{N}(0,I)$ . The goal of SD is then to learn a backward process + +$$ +p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right) = \mathcal {N} \left(x _ {t - 1}; \mu_ {\theta} \left(x _ {t}, t\right), \Sigma_ {\theta} \left(x _ {t}, t\right)\right) \tag {2} +$$ + +for $t = T, \dots, 1$ , which allows to generate a valid signal $x_0$ from the standard Gaussian noise $x_T$ . To get the final image generated from $x_T$ it remains to pass $x_0$ to the decoder of the initially chosen autoencoder: $\operatorname{Im} = \mathcal{D}(x_0)$ . + +After learning the abovementioned backward diffusion process (see DDPM [12]) one can apply a deterministic sampling process, called DDIM [35]: + +$$ +\begin{array}{l} x _ {t - 1} = \sqrt {\alpha_ {t - 1}} \left(\frac {x _ {t} - \sqrt {1 - \alpha_ {t}} \epsilon_ {\theta} ^ {t} \left(x _ {t}\right)}{\sqrt {\alpha_ {t}}}\right) + \tag {3} \\ \sqrt {1 - \alpha_ {t - 1}} e _ {\theta} ^ {t} (x _ {t}), \quad t = T, \dots , 1, \\ \end{array} +$$ + +where $\alpha_{t} = \prod_{i = 1}^{t}(1 - \beta_{i})$ and + +$$ +\epsilon_ {\theta} ^ {t} (x _ {t}) = \frac {\sqrt {1 - \alpha_ {t}}}{\beta_ {t}} x _ {t} + \frac {(1 - \beta_ {t}) (1 - \alpha_ {t})}{\beta_ {t}} \mu_ {\theta} (x _ {t}, t). \tag {4} +$$ + +To get a text-to-image synthesis framework, SD guides the diffusion processes with a textual prompt $\tau$ . Particularly for DDIM sampling, we get: + +$$ +\begin{array}{l} x _ {t - 1} = \sqrt {\alpha_ {t - 1}} \left(\frac {x _ {t} - \sqrt {1 - \alpha_ {t}} \epsilon_ {\theta} ^ {t} \left(x _ {t} , \tau\right)}{\sqrt {\alpha_ {t}}}\right) + \tag {5} \\ \sqrt {1 - \alpha_ {t - 1}} \epsilon_ {\theta} ^ {t} (x _ {t}, \tau), \quad t = T, \ldots , 1. \\ \end{array} +$$ + +It is worth noting that in SD, the function $\epsilon_{\theta}^{t}(x_{t},\tau)$ is modeled as a neural network with a UNet-like [30] architecture composed of convolutional and (self- and cross-) attentional blocks. $x_{T}$ is called the latent code of the signal $x_0$ and there is a method [3] to apply a deterministic forward process to reconstruct the latent code $x_{T}$ given a signal $x_0$ . This method is known as DDIM inversion. Sometimes for simplicity, we will call $x_{t},t = 1,\ldots ,T$ also the latent codes of the initial signal $x_0$ . + +# 3.2. Zero-Shot Text-to-Video Problem Formulation + +Existing text-to-video synthesis methods require either costly training on a large-scale (ranging from $1M$ to $15M$ data-points) text-video paired data [41, 15, 39, 11, 14, 5] or tuning on a reference video [42]. To make video generation cheaper and easier, we propose a new problem: zero-shot text-to-video synthesis. Formally, given a text description $\tau$ and a positive integer $m\in \mathbb{N}$ , the goal is to design a function $\mathcal{F}$ that outputs video frames $\mathcal{V}\in \mathbb{R}^{m\times H\times W\times 3}$ (for predefined resolution $H\times W$ ) that exhibit temporal consistency. To determine the function $\mathcal{F}$ , no training or fine-tuning must be performed on a video dataset. + +Our problem formulation provides a new paradigm for text-to-video. Noticeably, a zero-shot text-to-video method naturally benefits from quality improvements of text-to-image models. + +# 3.3. Method + +In this paper, we approach the zero-shot text-to-video task by exploiting the text-to-image synthesis power of Stable Diffusion (SD). As we need to generate videos instead of images, SD should operate on sequences of latent codes. The naive approach is to independently sample $m$ latent codes from standard Gaussian distribution $x_{T}^{1},\ldots ,x_{T}^{m}\sim \mathcal{N}(0,I)$ and apply DDIM sampling to obtain the corresponding tensors $x_0^k$ for $k = 1,\dots,m$ , followed by decoding to obtain the generated video sequence $\{\mathcal{D}(x_0^k)\}_{k = 1}^m\in \mathbb{R}^{m\times H\times W\times 3}$ . However, this leads to completely random generation of images sharing only the semantics described by $\tau$ but neither object appearance nor motion coherence (see appendix Fig. 14, first row). + +To address this issue, we propose to (i) introduce motion dynamics between the latent codes $x_{T}^{1}, \ldots, x_{T}^{m}$ to keep the global scene time consistent and (ii) use cross-frame attention mechanism to preserve the appearance and the identity of the foreground object. Each of the components of our method are described below in detail. The overview of our method can be found in Fig. 2. + +Note, to simplify notation, we will denote the entire sequence of latent codes by $x_{T}^{1:m} = [x_{T}^{1}, \ldots, x_{T}^{m}]$ . + +Algorithm 1 Motion dynamics in latent codes +Require: $\Delta t\geq 0,m\in \mathbb{N},\lambda >0,\delta = (\delta_x,\delta_y)\in$ $\mathbb{R}^2$ ,Stable Diffusion (SD) +1: $x_{T}^{1}\sim \mathcal{N}(0,I)$ random sample the first latent code +2: $x_{T^{\prime}}^{1}\gets \mathrm{DDIM\_Backward}(x_{T}^{1},\Delta t,SD)\triangleright$ perform $\Delta t$ backward steps by SD +3: for all $k = 2,3,\ldots ,m$ do +4: $\delta^k\gets \lambda \cdot (k - 1)\delta \triangleright$ computing global translation vectors +5: $W_{k}\leftarrow$ Warping by $\delta^k\triangleright$ defining warping functions +6: $\tilde{x}_{T^{\prime}}^{k}\gets W_{k}(x_{T^{\prime}}^{1})$ +7: $x_{T}^{k}\gets \mathrm{DDPM\_Forward}(\tilde{x}_{T^{\prime}}^{k},\Delta t)\triangleright \mathrm{DDPM~forward}$ for more motion freedom return $x_{T}^{1:m}$ + +# 3.3.1 Motion Dynamics in Latent Codes + +Instead of sampling the latent codes $x_{T}^{1:m}$ randomly and independently from the standard Gaussian distribution, we construct them by performing the following steps (see also Alg. 1 and Fig. 2). + +1. Randomly sample the latent code of the first frame: $x_{T}^{1} \sim \mathcal{N}(0, I)$ . +2. Perform $\Delta t \geq 0$ DDIM backward steps on the latent code $x_{T}^{1}$ by using the SD model and get the corresponding latent $x_{T'}^{1}$ , where $T' = T - \Delta t$ . +3. Define a direction $\delta = (\delta_x, \delta_y) \in \mathbb{R}^2$ for the global scene and camera motion. By default $\delta$ can be the main diagonal direction $\delta_x = \delta_y = 1$ . +4. For each frame $k = 1,2,\ldots ,m$ we want to generate, compute the global translation vector $\delta^k = \lambda \cdot (k - 1)\delta$ , where $\lambda$ is a hyperparameter controlling the amount of the global motion. +5. Apply the constructed motion (translation) flow $\delta^{1:m}$ to $x_{T'}^1$ , denote the resulting sequence by $\tilde{x}_{T'}^{1:m}$ : + +$$ +\tilde {x} _ {T ^ {\prime}} ^ {k} = W _ {k} \left(x _ {T ^ {\prime}} ^ {1}\right) \text {f o r} k = 1, 2, \dots , m, \tag {6} +$$ + +where $W_{k}(x_{T^{\prime}}^{1})$ is the warping operation for translation by the vector $\delta^{k}$ . + +6. Perform $\Delta t$ DDPM forward steps on each of the latents $\tilde{x}_{T'}^{2:m}$ and get the corresponding latent codes $x_{T}^{2:m}$ . + +Then we take the sequence $x_{T}^{1:m}$ as the starting point of the backward (video) diffusion process. As a result, the latent codes generated with our proposed motion dynamics lead to better temporal consistency of the global scene as well as the background, see in the appendix, Sect. 8.1. Yet, the initial latent codes are not constraining enough to describe particular colors, identities or shapes, thus still + +![](images/122a24541992a5835ea4cd4ab2532dd7f4bba239b8a46f912763c3fdc5492f96.jpg) +Figure 2: Method overview: Starting from a randomly sampled latent code $x_{T}^{1}$ , we apply $\Delta t$ DDIM backward steps to obtain $x_{T'}^{1}$ using a pre-trained Stable Diffusion model (SD). A specified motion field results for each frame $k$ in a warping function $W_{k}$ that turns $x_{T'}^{1}$ to $x_{T'}^{k}$ . By enhancing the latent codes with motion dynamics, we determine the global scene and camera motion and achieve temporal consistency in the background and the global scene. A subsequent DDPM forward application delivers latent codes $x_{T}^{k}$ for $k = 1, \ldots, m$ . By using the (probabilistic) DDPM method, a greater degree of freedom is achieved with respect to the motion of objects (see appendix Sec. 8.1). Finally, the latent codes are passed to our modified SD model using the proposed cross-frame attention, which uses keys and values from the first frame to generate the image of frame $k = 1, \ldots, m$ . By using cross-frame attention, the appearance and the identity of the foreground object are preserved throughout the sequence. Optionally, we apply background smoothing. To this end, we employ salient object detection to obtain for each frame $k$ a mask $M^{k}$ indicating the foreground pixels. Finally, for the background (using the mask $M^{k}$ ), a convex combination between the latent code $x_{t}^{1}$ of frame one warped to frame $k$ and the latent code $x_{t}^{k}$ is used to further improve the temporal consistency of the background. + +leading to temporal inconsistencies, especially for the foreground object. + +# 3.3.2 Reprogramming Cross-Frame Attention + +To address the issue mentioned above, we use a cross-frame attention mechanism to preserve the information about (in particular) the foreground object's appearance, shape, and identity throughout the generated video. + +To leverage the power of cross-frame attention and at the same time exploit a pretrained SD without retraining, we replace each of its self-attention layers with a cross-frame attention, with the attention for each frame being on the first frame. More precisely in the original SD UNet architecture $\epsilon_{\theta}^{t}(x_{t},\tau)$ , each self-attention layer takes a feature map $x\in \mathbb{R}^{h\times w\times c}$ , linearly projects it into query, key, value features $Q,K,V\in \mathbb{R}^{h\times w\times c}$ , and computes the layer output by the following formula (for simplicity described here for only one attention head) [38]: + +$$ +\operatorname {S e l f - A t t n} (Q, K, V) = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {c}}\right) V. \tag {7} +$$ + +In our case, each attention layer receives $m$ inputs: $x^{1:m} = [x^1,\dots,x^m]\in \mathbb{R}^{m\times h\times w\times c}$ . Hence, the linear projection layers produce $m$ queries, keys, and values $Q^{1:m},K^{1:m},V^{1:m}$ , respectively. + +Therefore, we replace each self-attention layer with a cross-frame attention of each frame on the first frame as + +follows: + +$$ +\begin{array}{l} \operatorname {C r o s s - F r a m e - A t t n} \left(Q ^ {k}, K ^ {1: m}, V ^ {1: m}\right) = \\ \operatorname {S o f t m a x} \left(\frac {Q ^ {k} \left(K ^ {1}\right) ^ {T}}{\sqrt {c}}\right) V ^ {1} \tag {8} \\ \end{array} +$$ + +for $k = 1, \ldots, m$ . By using cross frame attention, the appearance and structure of the objects and background as well as identities are carried over from the first frame to subsequent frames, which significantly increases the temporal consistency of the generated frames (see in the appendix the Figures 14, 16, 22, 23). + +# 3.3.3 Background smoothing (Optional) + +We further improve temporal consistency of the background using a convex combination of background-masked latent codes between the first frame and frame $k$ . This helps especially to generate videos from textual prompts when one or no initial image and no further guidance are provided. + +In more detail, given the generated sequence of our video generator, $x_0^{1:m}$ , we apply (an in-house solution for) salient object detection [40] to the decoded images to obtain a corresponding foreground mask $M^k$ for each frame $k$ . Then we warp $x_t^1$ according to the employed motion dynamics defined by $W_k$ and denote the result by $\hat{x}_t^k \coloneqq W_k(x_t^1)$ . + +Background smoothing is achieved by a convex combination between the actual latent code $x_{t}^{k}$ and the warped + +![](images/b6f131c5c649c6adec56561c373e34336d84eca3cc6b006750a0ecf743bad24a.jpg) +Figure 3: Text-to-Video results of our method. Depicted frames show that identities and appearances are temporally consistent and fitting to the textual prompt. For more results, see Appendix Sec. 8. + +![](images/daf48d03f9add4f33db4bd7ad68c90101df7e7ba8c2d823beb961dd3210654e8.jpg) +Figure 4: The overview of Text2Video-Zero + ControlNet + +latent code $\hat{x}_t^k$ on the background, i.e., + +$$ +\bar {x} _ {t} ^ {k} = M ^ {k} \odot x _ {t} ^ {k} + (1 - M ^ {k}) \odot (\alpha \hat {x} _ {t} ^ {k} + (1 - \alpha) x _ {t} ^ {k}), \tag {9} +$$ + +for $k = 1, \ldots, m$ , where $\alpha$ is a hyperparameter, which we empirically choose $\alpha = 0.6$ . Finally, DDIM sampling is employed on $\overline{x}_t^k$ , which delivers video generation with background smoothing. We use background smoothing in our video generation from text when no guidance is provided. For an ablation study on background smoothing, see the appendix, Sec. 8.1. + +# 3.4. Conditional and Specialized Text-to-Video + +Recently powerful controlling mechanisms [47, 22, 17] emerged to guide the diffusion process for text-to-image generation. Particularly, ControlNet [47] enables to condition the generation process using edges, pose, semantic masks, image depths, etc. However, a direct application of ControlNet in the video domain leads to temporal inconsistencies and to severe changes of object appearance, identity, and the background (see in the appendix Figures 14, 16, 22, 23). It turns out that our modifications on the basic diffusion process for videos result in more consistent videos guided by ControlNet conditions. We would like to point out again that our method does not require any fine-tuning or optimization processes. + +More specifically, ControlNet creates a trainable copy of the encoder (including the middle blocks) of the UNet $\epsilon_{\theta}^{t}(x_{t},\tau)$ while additionally taking the input $x_{t}$ and a condition $c$ , and adds the outputs of each layer to the skip-connections of the original UNet. Here $c$ can be any type of condition, such as edge map, scribbles, pose (body landmarks), depth map, segmentation map, etc. The trainable branch is being trained on a specific domain for each type of the condition $c$ resulting in an effective conditional text-to-image generation mechanism. + +To guide our video generation process with ControlNet + +![](images/d3ad2b3a4b5b049ed5496c6a32387da51b0b14902ac6e8116c10048648b83c44.jpg) +(a) a panda dancing in Antarctica + +![](images/146925351d27b7ffa1c17ba502561b340a30916edea3f2652b40c89a654fc3dd.jpg) +(b) a cyborg dancing near a volcano + +![](images/a7ccb181fc487596aeaab093619c3d96ab8d5083fe3a19785c70d3d74ab20209.jpg) +(c) an astronaut dancing in the outer space + +![](images/a4e34c0bba37c003312b0238e5d92a646b35d900005e18f29110e0d78e93650d.jpg) +Figure 5: Conditional generation with pose control. For more results see appendix, Sec. 10. +(a) oil painting of a girl dancing close-up + +![](images/517241d129b050d6dca845a4a8b6745fc102cd095e105dd0b0fd23d033d798ac.jpg) +(b) Cyberpunk boy with a hat dancing close-up +Figure 6: Conditional generation with edge control. For more results see appendix, Sec. 9. + +we apply our method to the basic diffusion process, i.e. enrich the latent codes $x_{T}^{1:m}$ with motion information and change the self-attention into cross-frame attentions in the main UNet. While adopting the main UNet for video generation task, we apply the ControlNet pretrained copy branch per-frame on each $x_{t}^{k}$ for $k = 1,\dots,m$ in each diffusion time-step $t = T,\ldots ,1$ and add the ControlNet branch outputs to the skip-connections of the main UNet. + +Furthermore, for our conditional generation task, we adopted the weights of specialized DreamBooth (DB) [31] models1. This gives us specialized time-consistent video generations (see Fig. 7). + +![](images/825a7a5b591098e7077a2d104041b6b91451f42f4ae2f62bc2616ad6f45e35b0.jpg) +(a) oil painting of a beautiful girl avatar style + +![](images/1700324d92b6c560ad0c09d2aa1ac3ba41c914f5731b997e0d6df048544f3584.jpg) +(b) gta-5 style +Figure 7: Conditional generation with edge control and DB models. + +# 3.5. Video Instruct-Pix2Pix + +With the rise of text-guided image editing methods such as Prompt2Prompt [9], Instruct-Pix2Pix [2], SDEdit [20], etc., text-guided video editing approaches emerged [1, 16, 42]. While these methods require complex optimization processes, our approach enables the adoption of any SD-based text-guided image editing algorithm to the video domain without any training or fine-tuning. Here we take the text-guided image editing method Instruct-Pix2Pix and combine it with our approach. More precisely, we change the self-attention mechanisms in Instruct-Pix2Pix to cross-frame attentions according to Eq. 8. Our experiments show that this adaptation significantly improves the consistency of the edited videos (see Fig. 9) over the naive per-frame usage of Instruct-Pix2Pix. + +# 4. Experiments + +# 4.1. Implementation Details + +We take the Stable Diffusion [29] code2 with its pretrained weights from version 1.5 as basis and implement our modifications. For each video, we generate $m = 8$ frames with $512 \times 512$ resolution. However, our framework allows generating any number of frames, either by increasing $m$ , or by employing our method in an auto-regressive fashion where the last generated frame $m$ becomes the first frame in computing the next $m$ frames. For all text-to-video generation experiments, we take $T' = 881$ , $T = 941$ without specific tuning, while for conditional and specialized generation, and for Video Instruct-Pix2Pix, we take $T' = T = 1000$ . + +For a conditional generation, we use the codebase3 of ControlNet [47]. For specialized models, we take DB [31] models from publicly available sources. For Video Instruct + +Pix2Pix, we use the codebase of Instruct Pix2Pix [2]. + +# 4.2. Qualitative Results + +All applications of Text2Video-Zero show that it successfully generates videos where the global scene and the background are time consistent and the context, appearance, and identity of the foreground object are maintained throughout the entire sequence. + +In the case of text-to-video, we observe that it generates high-quality videos that are well-aligned to the text prompt (see Fig. 3 and the appendix). For instance, the depicted panda shows a naturally walking on the street. Likewise, using additional guidance from edges or poses (see Fig. 5, Fig. 6 and Fig. 7 and the appendix), high quality videos are generated matching the prompt and the guidance that show great temporal consistency and identity preservation. + +Videos generated by Video Instruct-Pix2Pix (see Fig. 1 and the appendix) possess high-fidelity with respect to the input video, while following closely the instruction. + +# 4.3. Comparison with Baselines + +We compare our method with two publicly available baselines: CogVideo [15] and Tune-A-Video [42]. Since CogVideo is a text-to-video method we compare with it in pure text-guided video synthesis settings. With Tune-A-Video we compare in our Video Instruct-Pix2Pix setting. + +# 4.3.1 Quantitative Comparison + +To show quantitative results, we evaluate the CLIP score [10], which indicates video-text alignment. We randomly take 25 videos generated by CogVideo and synthesize corresponding videos using the same prompts according to our method. The CLIP scores for our method and CogVideo are 31.19 and 29.63, respectively. Our method thus slightly outperforms CogVideo, even though the latter has 9.4 billion parameters and requires large-scale training on videos. + +# 4.3.2 Qualitative Comparison + +We present several results of our method in Fig. 8 and provide a qualitative comparison to CogVideo [15]. Both methods show good temporal consistency throughout the sequence, preserving the identity of the object and background. However, our method shows better text-video alignment. For instance, while our method correctly generates a video of a man riding a bicycle in the sunshine in Fig. 8(b), CogVideo sets the background to moon light. Also in Fig. 8(a), our method correctly shows a man running in the snow, while neither the snow nor a man running are clearly visible in the video generated by CogVideo. + +Qualitative results of Video Instruct-Pix2Pix and a visual comparison with per-frame Instruct-Pix2Pix and Tune-A-Video are shown in Fig. 9. While Instruct-Pix2Pix shows a good editing performance per frame, it lacks temporal consistency. This becomes evident especially in the video depicting a skiing person, where the snow and the sky are drawn using different styles and colors. Using our Video Instruct-Pix2Pix method, these issues are solved resulting in temporally consistent video edits throughout the sequence. + +While Tune-A-Video creates temporally consistent video generations, it is less aligned to the instruction guidance than our method, struggles creating local edits and losses details of the input sequence. This becomes apparent when looking at the edit of the dancer video depicted in Fig. 9 (left side). In contrast to Tune-A-Video, our method draws the entire dress brighter and at the same time better preserves the background, e.g. the wall behind the dancer is almost kept the same. Tune-A-Video draws a severely modified wall. Moreover, our method is more faithful to the input details, e.g., Video Instruct-Pix2Pix draws the dancer using the pose exactly as provided (Fig. 9 left), and shows all skiing persons appearing in the input video (compare last frame of Fig. 9(right)), in constraint to Tune-A-Video. All the above-mentioned weaknesses of Tune-A-Video can also be observed in our additional evaluations that are provided in the appendix, Figures 25, 26. + +# 4.4. Ablation Study + +We perform several ablation studies and provide the results in the appendix, Sect. 8.1. Namely, we analyse background smoothing, $\Delta t$ , and two main components of our method: making the initial latent codes coherent to a motion, and using cross-frame attention on the first frame instead of self-attention. + +# 5. Limitations and Future Work + +The main limitation of this work is the inability to generate longer videos with sequences of actions. Future research may target enriching our method with techniques such as autoregressive scene action generation, while keeping the training-free spirit. Overall, our method is focused on generating video key-frames as in the first stage of Imagen Video [11] and Make-A-Video [33], and can thus be considered as good basis for longer and smoother video generation by integrating temporal upsampling. + +# 6. Conclusion + +In this paper, we addressed the problem of zero-shot text-to-video synthesis and proposed a novel method for time-consistent video generation. Our approach does not require any optimization or fine-tuning, making text-to-video generation and its applications affordable for everyone. We + +![](images/6abe8efd60ccf723ec1e9bfcb66f4574cb2f3051b4762cb63fac75b7a0dd275c.jpg) + +![](images/02142f2d2c7c8c9739bc72cbc1006c4817647c9f722476cd81bb514a1632fb45.jpg) + +![](images/2f1d2962f5bfec19bf745c3ed67c17bd35952ef1ad7890897081c7ef000ff228.jpg) +(a) a man is running in the snow + +![](images/4ef23ae7ad417795ac8e9930c23fbe718052fb3aecf46923b05c8940d408116f.jpg) + +![](images/8cc03a3fce8e50a676a49e08b32b6c6f2e20cea5cb31f87195034f3043726e62.jpg) + +![](images/14901d1229a7627c775f5d890f9782ec261cd0f088986498f5c1b5d1a91fd28a.jpg) + +![](images/513fb00d780079568f573e4efe0150ebf79212d9c5b007cfd5e2a13cf740a0d3.jpg) + +![](images/0b0f7fad9e4f521742e716a39bedcd784b1e3a7b6f0498f3a4d07c7eed086b18.jpg) + +![](images/03ad100fe744a479696e12c8e8ed9972586357f6dd7704cc2052c4057e1871aa.jpg) + +![](images/78a6506b39f07f779c7501d2e5ebc2373cf7120ecd0843e0aa64ca5104a2c138.jpg) + +![](images/29a51a72e6f73ce46daffdb36bf09a7f347d6f0f158b85c94c99b84c3b2065cc.jpg) +(b) a man is riding a bicycle in the sunshine + +![](images/b847fedda237ea961f3095dc645dd670fd215bf04ccd158a6f21eab5143c84e5.jpg) + +![](images/7d9085b17c5c4083691edca26fac6f0759dabe67ea8c568c41eb9f4232a6f9e0.jpg) + +![](images/b438330d6528881f97ad4834fa6d6e05228a783afc07052f52f55178df4e180c.jpg) + +![](images/6ba3bdb805425ce2ad7bce7cdb062e10ef911cc0169f129d21e69b447148d0f5.jpg) + +![](images/a8e18d4bf4a8d27fc085b6fe4e1669861fec77c0bce8bd34e84f032b14ca3c14.jpg) + +![](images/42a8be58677d32fc464e3d22d2b1e430442a35ae3112e9d6d776eb18889bfba0.jpg) +Figure 8: Comparison of our method vs CogVideo on text-to-video generation task (left is ours, right is CogVideo [15]). For more comparisons, see appendix Fig. 18. + +![](images/453efd49d5f4494cf18f24b6db0d44899c0502ce70277990668181abe65016ae.jpg) + +![](images/2c9fcb8cf777e9e0b26f320d1e131ed8a6dd0b438a884752264add079fa6f1f8.jpg) + +![](images/2ac10858d4af622e178cc7acaefc2d7c40248dfeb5369f110578b829daac52ab.jpg) +(c) a man is walking in the rain + +![](images/65f570a45b25502db7088cd712e9db0c3d4ca04a4ec4f1066e00e85ce661fd30.jpg) + +![](images/43096e243642767407c61fc05c4ca865d4474303adea5f53c21e3d37038fd8e5.jpg) + +![](images/99d564c034db10e955e246dbf50133e35caff3a65b5dcee6688c8736f074e14b.jpg) + +![](images/a8861c84fe2e35e07e4ac85ad0e4a7b3e35beb99f1a5ae66c3bef9c7578a012f.jpg) + +![](images/b5e0f36cb746698940a75050946403f57e4e0d127ca69d23ad45a59261dbd117.jpg) +Figure 9: Comparison of Video Instruct-Pix2Pix(ours) with Tune-A-Video and per-frame Instruct-Pix2Pix. For more comparisons see our appendix. + +demonstrated the effectiveness of our method for various applications, including conditional and specialized video generation, and Video Instruct-Pix2Pix, i.e., instruction-guided video editing. Our contributions to the field include presenting a new problem of zero-shot text-to-video synthesis, showing the utilization of text-to-image diffusion models for generating time-consistent videos, and providing evidence of the effectiveness of our method for various video synthesis applications. We believe that our proposed method will open up new possibilities for video generation and editing, making it accessible and affordable for everyone. + +Acknowledgments. This material is partially based upon work supported by the National Science Foundation CAREER Award #2239840, and the National AI Institute for Exceptional Education (Award #2229873) by National Science Foundation and the Institute of Education Sciences, U.S. Department of Education. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation, the Institute of Education Sciences, or the U.S. Department of Education. + +# References + +[1] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723. Springer, 2022. 7 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2, 7, 8, 13, 25, 26 +[3] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 4 +[4] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022.3 +[5] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 2, 3, 4 +[6] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 3 +[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 89-106. Springer, 2022. 2 +[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2 +[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 7 +[10] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 8 +[11] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 4, 8 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3 +[13] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 12 +[14] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 3, 4 + +[15] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3, 4, 8, 9, 12, 13, 20 +[16] Yao-Chih Lee, Ji-Ze Genevieve Jang, Yi-Ting Chen, Elizabeth Qiu, and Jia-Bin Huang. Shape-aware text-driven layered video editing. arXiv e-prints, pages arXiv-2301, 2023. 7 +[17] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 6 +[18] Haoming Lu, Hazarapet Tunanyan, Kai Wang, Shant Navasardyan, Zhangyang Wang, and Humphrey Shi. Specialist diffusion: Plug-and-play sample-efficient fine-tuning of text-to-image diffusion models to learn any unseen style. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14267-14276, 2023. 3 +[19] Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Generating images from captions with attention. arXiv preprint arXiv:1511.02793, 2015. 2 +[20] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 7 +[21] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 2 +[22] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 6 +[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[24] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. Mirrorgan: Learning text-to-image generation by redescription. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1514, 2019. 2 +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[27] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. + +Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2 +[28] Scott E Reed, Zeynep Akata, Santosh Mohan, Samuel Tenka, Bernt Schiele, and Honglak Lee. Learning what and where to draw. Advances in neural information processing systems, 29, 2016. 2 +[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 7 +[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 4 +[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3, 7, 22 +[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2 +[33] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 3, 8 +[34] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2 +[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3 +[36] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2 +[37] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3 +[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5 +[39] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain + +textual description. arXiv preprint arXiv:2210.02399, 2022. 3,4 +[40] Wenguan Wang, Qiuxia Lai, Huazhu Fu, Jianbing Shen, Haibin Ling, and Ruigang Yang. Salient object detection in the deep learning era: An in-depth survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3239-3259, 2021. 5 +[41] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Niwa: Visual synthesis pretraining for neural visual world creation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 720-736. Springer, 2022. 3, 4 +[42] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2, 3, 4, 7, 8, 12, 13, 26 +[43] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2 +[44] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332, 2022. 2, 3 +[45] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2 +[46] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2 +[47] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 3, 6, 7 \ No newline at end of file diff --git a/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/images.zip b/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..18f64a51dcd7ba9f6debfec7c8df970ce0617866 --- /dev/null +++ b/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebd6874a4e6d36480d48e19a5eb311fc2ab80500c7bbaac238c0f1a3cd672621 +size 824200 diff --git a/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/layout.json b/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ffbdced1bfcb9945187f5bb0011c5c2801e57595 --- /dev/null +++ b/2023/Text2Video-Zero_ Text-to-Image Diffusion Models are Zero-Shot Video Generators/layout.json @@ -0,0 +1,12664 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 496, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 496, + 140 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 496, + 140 + ], + "type": "text", + "content": "Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "spans": [ + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "text", + "content": "Levon Khachatryan" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "text", + "content": " Andranik Movsisyan" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "text", + "content": " Vahram Tadevosyan" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "text", + "content": " Roberto Henschel" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "text", + "content": " Zhangyang Wang" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "text", + "content": " Shant Navasardyan" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "text", + "content": " Humphrey Shi" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "text", + "content": "Picsart AI Research (PAIR) " + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "text", + "content": "UT Austin " + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 160, + 518, + 219 + ], + "type": "text", + "content": "SHI Labs @ Georgia Tech, Oregon & UIUC https://github.com/Picsart-AI-Research/Text2Video-Zero" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 227, + 111, + 288 + ], + "blocks": [ + { + "bbox": [ + 50, + 227, + 111, + 288 + ], + "lines": [ + { + "bbox": [ + 50, + 227, + 111, + 288 + ], + "spans": [ + { + "bbox": [ + 50, + 227, + 111, + 288 + ], + "type": "image", + "image_path": "5638995bdf120554a11b3d2bc540caa4ab3e8a7c2868cc2d7e7fe856b11f0784.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 112, + 228, + 173, + 288 + ], + "blocks": [ + { + "bbox": [ + 112, + 228, + 173, + 288 + ], + "lines": [ + { + "bbox": [ + 112, + 228, + 173, + 288 + ], + "spans": [ + { + "bbox": [ + 112, + 228, + 173, + 288 + ], + "type": "image", + "image_path": "a4d8d0a3ceca72cf40b495bea7ac0f26329a5b6dd01e1531cc72073f7a0e8251.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 174, + 228, + 235, + 288 + ], + "blocks": [ + { + "bbox": [ + 174, + 228, + 235, + 288 + ], + "lines": [ + { + "bbox": [ + 174, + 228, + 235, + 288 + ], + "spans": [ + { + "bbox": [ + 174, + 228, + 235, + 288 + ], + "type": "image", + "image_path": "aeeb8c18a627e667284dee6ab977731786cf2633a9eed02ccbf5f00d10389a6c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 235, + 228, + 296, + 288 + ], + "blocks": [ + { + "bbox": [ + 235, + 228, + 296, + 288 + ], + "lines": [ + { + "bbox": [ + 235, + 228, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 235, + 228, + 296, + 288 + ], + "type": "image", + "image_path": "6e8e563dfa608ee4bb05d24195d74a25e40bca7cd71d46b792d52b0c2e144d75.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 298, + 228, + 359, + 288 + ], + "blocks": [ + { + "bbox": [ + 298, + 228, + 359, + 288 + ], + "lines": [ + { + "bbox": [ + 298, + 228, + 359, + 288 + ], + "spans": [ + { + "bbox": [ + 298, + 228, + 359, + 288 + ], + "type": "image", + "image_path": "f9323419c2c537317827585f68dc62e75a08942e2c5bf67a185596a625edd231.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 360, + 228, + 421, + 288 + ], + "blocks": [ + { + "bbox": [ + 360, + 228, + 421, + 288 + ], + "lines": [ + { + "bbox": [ + 360, + 228, + 421, + 288 + ], + "spans": [ + { + "bbox": [ + 360, + 228, + 421, + 288 + ], + "type": "image", + "image_path": "2cb5b3ba5892262e4f2653806a92caf658daf7fefbe529caef50e48a01a34e66.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 422, + 228, + 482, + 288 + ], + "blocks": [ + { + "bbox": [ + 422, + 228, + 482, + 288 + ], + "lines": [ + { + "bbox": [ + 422, + 228, + 482, + 288 + ], + "spans": [ + { + "bbox": [ + 422, + 228, + 482, + 288 + ], + "type": "image", + "image_path": "f3c8eb8cd473b95a423a37cfe89749e5874467222ce30246c27a17fd6e964d98.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 484, + 228, + 545, + 288 + ], + "blocks": [ + { + "bbox": [ + 484, + 228, + 545, + 288 + ], + "lines": [ + { + "bbox": [ + 484, + 228, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 484, + 228, + 545, + 288 + ], + "type": "image", + "image_path": "1ef3f447b8f14833e105f220bbfb65c77cab36a0b3a985334e5aab9fab31285f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 195, + 290, + 397, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 290, + 397, + 301 + ], + "spans": [ + { + "bbox": [ + 195, + 290, + 397, + 301 + ], + "type": "text", + "content": "Text-to-Video generation: \"a horse galloping on a street\"" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 49, + 305, + 110, + 366 + ], + "blocks": [ + { + "bbox": [ + 49, + 305, + 110, + 366 + ], + "lines": [ + { + "bbox": [ + 49, + 305, + 110, + 366 + ], + "spans": [ + { + "bbox": [ + 49, + 305, + 110, + 366 + ], + "type": "image", + "image_path": "1d5f7e351bb85f55105e698774b39c26b3d37d9e15ef182532d35f05c49a2c21.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 111, + 305, + 171, + 366 + ], + "blocks": [ + { + "bbox": [ + 111, + 305, + 171, + 366 + ], + "lines": [ + { + "bbox": [ + 111, + 305, + 171, + 366 + ], + "spans": [ + { + "bbox": [ + 111, + 305, + 171, + 366 + ], + "type": "image", + "image_path": "0ab0779b112b13190a80c90c982d45723d9d92f348aef7d9c95a2ddd267cdd67.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 172, + 305, + 233, + 366 + ], + "blocks": [ + { + "bbox": [ + 172, + 305, + 233, + 366 + ], + "lines": [ + { + "bbox": [ + 172, + 305, + 233, + 366 + ], + "spans": [ + { + "bbox": [ + 172, + 305, + 233, + 366 + ], + "type": "image", + "image_path": "39c461fa8f85041f87e27ab52faf1c89a2be2b45fcdf6acef5962b4e82f8c215.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 235, + 305, + 295, + 366 + ], + "blocks": [ + { + "bbox": [ + 235, + 305, + 295, + 366 + ], + "lines": [ + { + "bbox": [ + 235, + 305, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 235, + 305, + 295, + 366 + ], + "type": "image", + "image_path": "b943bb548f0ef3e190658ddb374d22225872c963246fef31a653736ec95e6290.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 296, + 305, + 357, + 366 + ], + "blocks": [ + { + "bbox": [ + 296, + 305, + 357, + 366 + ], + "lines": [ + { + "bbox": [ + 296, + 305, + 357, + 366 + ], + "spans": [ + { + "bbox": [ + 296, + 305, + 357, + 366 + ], + "type": "image", + "image_path": "840293918536d238e5f19ca003a506b66fb15b941f5234520443ae8b6ee22c91.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 359, + 305, + 419, + 366 + ], + "blocks": [ + { + "bbox": [ + 359, + 305, + 419, + 366 + ], + "lines": [ + { + "bbox": [ + 359, + 305, + 419, + 366 + ], + "spans": [ + { + "bbox": [ + 359, + 305, + 419, + 366 + ], + "type": "image", + "image_path": "0c3cc4cfe40b25a2fd3542f66e350609066bc3f7b46224fb9cd77b59233c84ac.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 420, + 305, + 481, + 366 + ], + "blocks": [ + { + "bbox": [ + 420, + 305, + 481, + 366 + ], + "lines": [ + { + "bbox": [ + 420, + 305, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 420, + 305, + 481, + 366 + ], + "type": "image", + "image_path": "635b8291b229f014be55890211a331493e9d02368740bcbfee9f4ef8a6df5412.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 482, + 305, + 543, + 366 + ], + "blocks": [ + { + "bbox": [ + 482, + 305, + 543, + 366 + ], + "lines": [ + { + "bbox": [ + 482, + 305, + 543, + 366 + ], + "spans": [ + { + "bbox": [ + 482, + 305, + 543, + 366 + ], + "type": "image", + "image_path": "819828471aefa4ac20a52e296620fe280800fa82e8aebe8e25ae192f7f83e46c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 174, + 367, + 420, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 367, + 420, + 378 + ], + "spans": [ + { + "bbox": [ + 174, + 367, + 420, + 378 + ], + "type": "text", + "content": "Text-to-Video generation: \"a panda is playing guitar on times square\"" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 49, + 392, + 138, + 442 + ], + "blocks": [ + { + "bbox": [ + 49, + 392, + 138, + 442 + ], + "lines": [ + { + "bbox": [ + 49, + 392, + 138, + 442 + ], + "spans": [ + { + "bbox": [ + 49, + 392, + 138, + 442 + ], + "type": "image", + "image_path": "480ad00bf61e23c3b61d9e02a24e6560632e83738a3d73875f45f49883bf473e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 143, + 392, + 232, + 442 + ], + "blocks": [ + { + "bbox": [ + 143, + 392, + 232, + 442 + ], + "lines": [ + { + "bbox": [ + 143, + 392, + 232, + 442 + ], + "spans": [ + { + "bbox": [ + 143, + 392, + 232, + 442 + ], + "type": "image", + "image_path": "730b01567f2aae5a32d7af0090815955b6c306e8df495bf04b55b85c0ff9153e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 237, + 392, + 326, + 442 + ], + "blocks": [ + { + "bbox": [ + 237, + 392, + 326, + 442 + ], + "lines": [ + { + "bbox": [ + 237, + 392, + 326, + 442 + ], + "spans": [ + { + "bbox": [ + 237, + 392, + 326, + 442 + ], + "type": "image", + "image_path": "4911268d744d8638ec96d63d1f9037b9b22f564e3fe6b28558cb7095a5fe6b96.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 335, + 379, + 386, + 430 + ], + "blocks": [ + { + "bbox": [ + 335, + 379, + 386, + 430 + ], + "lines": [ + { + "bbox": [ + 335, + 379, + 386, + 430 + ], + "spans": [ + { + "bbox": [ + 335, + 379, + 386, + 430 + ], + "type": "image", + "image_path": "a3aeaf6679817680885232c903f0b133184877bb2afe91cf73169deb0cbd9d61.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 388, + 379, + 439, + 430 + ], + "blocks": [ + { + "bbox": [ + 388, + 379, + 439, + 430 + ], + "lines": [ + { + "bbox": [ + 388, + 379, + 439, + 430 + ], + "spans": [ + { + "bbox": [ + 388, + 379, + 439, + 430 + ], + "type": "image", + "image_path": "ec210fa5b5451758b917214eced5d7a9fdd264a0af302f4399c16b7d4aef56e0.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 441, + 379, + 491, + 430 + ], + "blocks": [ + { + "bbox": [ + 441, + 379, + 491, + 430 + ], + "lines": [ + { + "bbox": [ + 441, + 379, + 491, + 430 + ], + "spans": [ + { + "bbox": [ + 441, + 379, + 491, + 430 + ], + "type": "image", + "image_path": "f0804966b7b5bc4001e2b961ba987f85799727ca73df2e167ccb3c1aa7d74b89.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 493, + 379, + 543, + 430 + ], + "blocks": [ + { + "bbox": [ + 493, + 379, + 543, + 430 + ], + "lines": [ + { + "bbox": [ + 493, + 379, + 543, + 430 + ], + "spans": [ + { + "bbox": [ + 493, + 379, + 543, + 430 + ], + "type": "image", + "image_path": "9242f762bf79b0af603c475a35eb32e70bcf12774fd5f32c533c470d571ff4ce.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 49, + 453, + 138, + 503 + ], + "blocks": [ + { + "bbox": [ + 49, + 453, + 138, + 503 + ], + "lines": [ + { + "bbox": [ + 49, + 453, + 138, + 503 + ], + "spans": [ + { + "bbox": [ + 49, + 453, + 138, + 503 + ], + "type": "image", + "image_path": "d1a694a11d86ed5c7ab1519115b1566b8ebbe2bde282a4fc4e2f5d8b2749b519.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 505, + 299, + 516 + ], + "lines": [ + { + "bbox": [ + 75, + 505, + 299, + 516 + ], + "spans": [ + { + "bbox": [ + 75, + 505, + 299, + 516 + ], + "type": "text", + "content": "Video Instruct-Pix2Pix: \"make it Van Gogh Starry Night style\"" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 143, + 453, + 232, + 503 + ], + "blocks": [ + { + "bbox": [ + 143, + 453, + 232, + 503 + ], + "lines": [ + { + "bbox": [ + 143, + 453, + 232, + 503 + ], + "spans": [ + { + "bbox": [ + 143, + 453, + 232, + 503 + ], + "type": "image", + "image_path": "6d3e9c648aaa0481ddedcfc7d7de88c6c123d85a14ba39031da9e1e6b51a7fff.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 529, + 547, + 566 + ], + "lines": [ + { + "bbox": [ + 46, + 529, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 547, + 566 + ], + "type": "text", + "content": "Figure 1: Our method Text2Video-Zero enables zero-shot video generation using (i) a textual prompt (see rows 1, 2), (ii) a prompt combined with guidance from poses or edges (see lower right), and (iii) Video Instruct-Pix2Pix, i.e., instruction-guided video editing (see lower left). Results are temporally consistent and follow closely the guidance and textual prompts." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 238, + 453, + 326, + 503 + ], + "blocks": [ + { + "bbox": [ + 238, + 453, + 326, + 503 + ], + "lines": [ + { + "bbox": [ + 238, + 453, + 326, + 503 + ], + "spans": [ + { + "bbox": [ + 238, + 453, + 326, + 503 + ], + "type": "image", + "image_path": "8530c75814d28747a705839bf7314dfbf197d044bd9dea2d0915c7ace3a34787.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 351, + 430, + 525, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 430, + 525, + 450 + ], + "spans": [ + { + "bbox": [ + 351, + 430, + 525, + 450 + ], + "type": "text", + "content": "Text-to-Video generation + pose control: \"a bear dancing on the concrete\"" + } + ] + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 335, + 451, + 386, + 500 + ], + "blocks": [ + { + "bbox": [ + 335, + 451, + 386, + 500 + ], + "lines": [ + { + "bbox": [ + 335, + 451, + 386, + 500 + ], + "spans": [ + { + "bbox": [ + 335, + 451, + 386, + 500 + ], + "type": "image", + "image_path": "95aabc830f54679227d25a91795bc51a4a902b50cf7b59265807a052c4655bcc.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 502, + 541, + 514 + ], + "lines": [ + { + "bbox": [ + 334, + 502, + 541, + 514 + ], + "spans": [ + { + "bbox": [ + 334, + 502, + 541, + 514 + ], + "type": "text", + "content": "Text-to-Video generation + edge control: \"white butterfly\"" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 388, + 450, + 438, + 500 + ], + "blocks": [ + { + "bbox": [ + 388, + 450, + 438, + 500 + ], + "lines": [ + { + "bbox": [ + 388, + 450, + 438, + 500 + ], + "spans": [ + { + "bbox": [ + 388, + 450, + 438, + 500 + ], + "type": "image", + "image_path": "81454a49861a03f943a212da4832b983a0defbb449084f3f786ee71a4f2f67c5.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 440, + 450, + 491, + 500 + ], + "blocks": [ + { + "bbox": [ + 440, + 450, + 491, + 500 + ], + "lines": [ + { + "bbox": [ + 440, + 450, + 491, + 500 + ], + "spans": [ + { + "bbox": [ + 440, + 450, + 491, + 500 + ], + "type": "image", + "image_path": "4ec10fefe4c5557de5c9080d4dc0e738df1ec2882fdb1cdafcffde8c06f82e34.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 493, + 450, + 543, + 500 + ], + "blocks": [ + { + "bbox": [ + 493, + 450, + 543, + 500 + ], + "lines": [ + { + "bbox": [ + 493, + 450, + 543, + 500 + ], + "spans": [ + { + "bbox": [ + 493, + 450, + 543, + 500 + ], + "type": "image", + "image_path": "3de2cf00b5845d69926f12d24372e2b3d58591b656c7d36840d8c6e9727f1026.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 143, + 582, + 192, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 582, + 192, + 594 + ], + "spans": [ + { + "bbox": [ + 143, + 582, + 192, + 594 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 46, + 596, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 596, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 596, + 288, + 693 + ], + "type": "text", + "content": "Recent text-to-video generation approaches rely on computationally heavy training and require large-scale video datasets. In this paper, we introduce a new task, zero-shot text-to-video generation, and propose a low-cost approach (without any training or optimization) by leveraging the power of existing text-to-image synthesis methods (e.g. Stable Diffusion), making them suitable for the video domain. Our key modifications include (i) enriching the la" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 306, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 547, + 715 + ], + "type": "text", + "content": "tent codes of the generated frames with motion dynamics to keep the global scene and the background time consistent; and (ii) reprogramming frame-level self-attention using a new cross-frame attention of each frame on the first frame, to preserve the context, appearance, and identity of the foreground object. Experiments show that this leads to low overhead, yet high-quality and remarkably consistent video generation. Moreover, our approach is not limited to text-to-video synthesis but is also applicable to other tasks such as conditional and content-specialized video generation, and Video Instruct-Pix2Pix, i.e., instruction-guided" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 703, + 126, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 126, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 126, + 714 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15954" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 131 + ], + "type": "text", + "content": "video editing. As experiments show, our method performs comparably or sometimes better than recent approaches, despite not being trained on additional video data. Our code is publicly available at: https://github.com/Picsart-AI-Research/Text2Video-Zero." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 155, + 127, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 155, + 127, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 155, + 127, + 167 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 175, + 287, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 287, + 414 + ], + "type": "text", + "content": "In recent years, generative AI has attracted enormous attention in the computer vision community. With the advent of diffusion models [34, 12, 35, 36], it has become tremendously popular and successful to generate high-quality images from textual prompts, also called text-to-image synthesis [26, 29, 32, 7, 44]. Recent works [14, 33, 11, 42, 5, 21] attempt to extend the success to text-to-video generation and editing tasks, by reusing text-to-image diffusion models in the video domain. While such approaches yield promising outcomes, most of them require substantial training with a massive amount of labeled data which can be costly and unaffordable for many users. With the aim of making video generation cheaper, Tune-A-Video [42] introduces a mechanism that can adopt Stable Diffusion (SD) model [29] for the video domain. The training effort is drastically reduced to tuning one video. While that is much more efficient than previous approaches, it still requires an optimization process. In addition, the generation abilities of Tune-A-Video are limited to text-guided video editing applications; video synthesis from scratch, however, remains out of its reach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 415, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 288, + 714 + ], + "type": "text", + "content": "In this paper, we take one step forward in studying the novel problem of zero-shot, \"training-free\" text-to-video synthesis, which is the task of generating videos from textual prompts without requiring any optimization or finetuning. A key concept of our approach is to modify a pre-trained text-to-image model (e.g., Stable Diffusion), enriching it with temporally consistent generation. By building upon already trained text-to-image models, our method takes advantage of their excellent image generation quality and enhances their applicability to the video domain without performing additional training. To enforce temporal consistency, we present two innovative and lightweight modifications: (1) we first enrich the latent codes of generated frames with motion information to keep the global scene and the background time consistent; (2) we then use cross-frame attention of each frame on the first frame to preserve the context, appearance, and identity of the foreground object throughout the entire sequence. Our experiments show that these simple modifications lead to high-quality and time-consistent video generations (see Fig. 1 and further results in the appendix). Despite the fact that other works train on large-scale video data, our method achieves similar or sometimes even better performance (see Figures 8, 9 and appendix Figures 18, 25, 26). Furthermore, our method is not limited to text-to-video synthesis but is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 132 + ], + "type": "text", + "content": "also applicable to conditional (see Figures 5,6 and appendix Figures 19, 21, 22, 23) and specialized video generation (see Fig. 7), and instruction-guided video editing, which we refer as Video Instruct-Pix2Pix motivated by Instruct-Pix2Pix [2] (see Fig. 9 and appendix Figures 24, 25, 26)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 133, + 516, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 133, + 516, + 143 + ], + "spans": [ + { + "bbox": [ + 317, + 133, + 516, + 143 + ], + "type": "text", + "content": "Our contributions are summarized as three-folds:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 150, + 545, + 318 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 150, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 150, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 150, + 545, + 209 + ], + "type": "text", + "content": "- A new problem setting of zero-shot text-to-video synthesis, aiming at making text-guided video generation and editing \"freely affordable\". We use only a pretrained text-to-image diffusion model without any further fine-tuning or optimization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 217, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 545, + 263 + ], + "type": "text", + "content": "- Two novel post-hoc techniques to enforce temporally consistent generation, via encoding motion dynamics in the latent codes, and reprogramming each frame's self-attention using a new cross-frame attention." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 271, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 271, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 317, + 271, + 545, + 318 + ], + "type": "text", + "content": "- A broad variety of applications that demonstrate our method's effectiveness, including conditional and specialized video generation, and Video Instruct-Pix2Pix i.e., video editing by textual instructions." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 328, + 392, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 328, + 392, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 328, + 392, + 340 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 348, + 449, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 449, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 449, + 361 + ], + "type": "text", + "content": "2.1. Text-to-Image Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 366, + 545, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 414 + ], + "type": "text", + "content": "Early approaches to text-to-image synthesis relied on methods such as template-based generation [19] and feature matching [28]. However, these methods were limited in their ability to generate realistic and diverse images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 415, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 485 + ], + "type": "text", + "content": "Following the success of GANs [8], several other deep learning-based methods were proposed for text-to-image synthesis. These include StackGAN [46], AttnGAN [43], and MirrorGAN [24], which further improve image quality and diversity by introducing novel architectures and attention mechanisms." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "content": "Later, with the advancement of transformers [38], new approaches emerged for text-to-image synthesis. Being a 12-billion-parameter transformer model, Dall-E [27] introduces a two-stage training process: First, it generates image tokens, which later are combined with text tokens for joint training of an autoregressive model. Later Parti [45] proposed a method to generate content-rich images with multiple objects. Make-a-Scene [7] enables a control mechanism by segmentation masks for text-to-image generation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Current approaches build upon diffusion models, thereby taking text-to-image synthesis quality to the next level. GLIDE [23] improved Dall-E by adding classifier-free guidance [13]. Later, Dall-E 2 [26] utilizes the contrastive model CLIP [25]. By means of diffusion processes, (i) a mapping from CLIP text encodings to image encodings, and (ii) a CLIP decoder is obtained. LDM / SD [29] applies a diffusion model on lower-resolution encoded signals of VQ-GAN [6], showing competitive quality with a significant gain in speed and efficiency. Imagen [32] shows" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15955" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "incredible performance in text-to-image synthesis by utilizing large language models for text processing. Versatile Diffusion [44] further unifies text-to-image, image-to-text and variations in a single multi-flow diffusion model. Specialisation of text-to-image models to desired styles can be obtained efficiently via few-shot tuning, e.g. using DreamBoth [31] or Specialist Diffusion [18], which employs text-to-image customized data augmentations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 171, + 289, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 289, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 289, + 291 + ], + "type": "text", + "content": "Because of their great image quality, it is desired to exploit text-to-image models for video generation. However, applying diffusion models in the video domain is not straightforward, especially due to their probabilistic generation procedure, making it difficult to ensure temporal consistency. As we show in our ablation experiments in the appendix (see Fig. 14), our modifications are crucial for temporal consistency in terms of both global scene and background motion, and for the preservation of the foreground object identity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 306, + 190, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 306, + 190, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 190, + 319 + ], + "type": "text", + "content": "2.2. Text-to-Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 328, + 289, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 289, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 289, + 615 + ], + "type": "text", + "content": "Text-to-video synthesis is a relatively new research direction. Existing approaches try to leverage autoregressive transformers and diffusion processes for the generation. NUWA [41] introduces a 3D transformer encoder-decoder framework and supports both text-to-image and text-to-video generation. Phenaki [39] introduces a bidirectional masked transformer with a causal attention mechanism that allows the generation of arbitrary-long videos from text prompt sequences. CogVideo [15] extends the text-to-image model CogView 2 [4] by tuning it using a multi-frame-rate hierarchical training strategy to better align text and video clips. Video Diffusion Models (VDM) [14] naturally extend text-to-image diffusion models and train jointly on image and video data. Imagen Video [11] constructs a cascade of video diffusion models and utilizes spatial and temporal super-resolution models to generate high-resolution time-consistent videos. Make-A-Video [33] builds upon a text-to-image synthesis model and leverages video data in an unsupervised manner. Gen-1 [5] extends SD and proposes a structure and content-guided video editing method based on visual or textual descriptions of desired outputs. Tune-A-Video [42] proposes a new task of one-shot video generation by extending and tuning SD on a single reference video." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Unlike the methods mentioned above, our approach is completely training-free, does not require massive computing power or dozens of GPUs, which makes the video generation process affordable for everyone. In this respect, Tunea-Video [42] comes closest to our work, as it reduces the necessary computations to tuning on only one video. However, it still requires an optimization process and its generating ability is heavily restricted by the reference video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 71, + 362, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 362, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 362, + 84 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 212 + ], + "type": "text", + "content": "We start this section with a brief introduction of diffusion models, particularly Stable Diffusion (SD) [29]. Then we introduce the problem formulation of zero-shot text-to-video synthesis, followed by a subsection presenting our approach. After that, to show the universality of our method, we use it in combination with ControlNet [47] and Dream-Booth [31] diffusion models for generating conditional and specialized videos. Later we demonstrate the power of our approach with the application of instruction-guided video editing, namely, Video Instruct-Pix2Pix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 217, + 405, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 217, + 405, + 229 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 405, + 229 + ], + "type": "text", + "content": "3.1. Stable Diffusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "text", + "content": "SD is a diffusion model operating in the latent space of an autoencoder " + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\mathcal{E}(\\cdot))" + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "text", + "content": ", namely VQ-GAN [6] or VQ-VAE [37], where " + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "text", + "content": " are the corresponding encoder and decoder, respectively. More precisely, if " + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "inline_equation", + "content": "x_0 \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "text", + "content": " is the latent tensor of an input image " + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "inline_equation", + "content": "\\mathrm{Im}" + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "text", + "content": " given by the autoencoder, i.e. " + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "inline_equation", + "content": "x_0 = \\mathcal{E}(\\mathrm{Im})" + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "text", + "content": ", diffusion forward process iteratively adds Gaussian noise to the signal " + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 327, + 545, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 327, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 311, + 327, + 545, + 342 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I\\right), t = 1,.., T \\tag {1}", + "image_path": "3721b93d32e6acaa278b9a1e5f96c5715c879cdefffed8c57764ce17a15eb605.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "inline_equation", + "content": "q(x_{t}|x_{t - 1})" + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "text", + "content": " is the conditional density of " + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "inline_equation", + "content": "\\{\\beta_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "text", + "content": " are hyperparameters. " + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "text", + "content": " is chosen to be as large that the forward process completely destroys the initial signal " + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "text", + "content": " resulting in " + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "inline_equation", + "content": "x_{T}\\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 304, + 348, + 547, + 409 + ], + "type": "text", + "content": ". The goal of SD is then to learn a backward process" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 333, + 415, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 415, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 333, + 415, + 545, + 430 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right) \\tag {2}", + "image_path": "ac76fd0a760f3ef3ab27de560972f55c66408eb1d9a5834d8f7d5cb164bd60fb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "inline_equation", + "content": "t = T, \\dots, 1" + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "text", + "content": ", which allows to generate a valid signal " + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "text", + "content": " from the standard Gaussian noise " + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "inline_equation", + "content": "x_T" + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "text", + "content": ". To get the final image generated from " + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "inline_equation", + "content": "x_T" + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "text", + "content": " it remains to pass " + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "text", + "content": " to the decoder of the initially chosen autoencoder: " + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "inline_equation", + "content": "\\operatorname{Im} = \\mathcal{D}(x_0)" + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 483, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 520 + ], + "type": "text", + "content": "After learning the abovementioned backward diffusion process (see DDPM [12]) one can apply a deterministic sampling process, called DDIM [35]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 337, + 525, + 545, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 525, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 337, + 525, + 545, + 570 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\left(\\frac {x _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} ^ {t} \\left(x _ {t}\\right)}{\\sqrt {\\alpha_ {t}}}\\right) + \\tag {3} \\\\ \\sqrt {1 - \\alpha_ {t - 1}} e _ {\\theta} ^ {t} (x _ {t}), \\quad t = T, \\dots , 1, \\\\ \\end{array}", + "image_path": "1df77b684fd4dfe860f87cd8590050f7840d8115c09de0df9a3ad96ab880712f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 577, + 434, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 577, + 434, + 592 + ], + "spans": [ + { + "bbox": [ + 306, + 577, + 434, + 592 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 577, + 434, + 592 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = \\prod_{i = 1}^{t}(1 - \\beta_{i})" + }, + { + "bbox": [ + 306, + 577, + 434, + 592 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 597, + 545, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 545, + 624 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\theta} ^ {t} (x _ {t}) = \\frac {\\sqrt {1 - \\alpha_ {t}}}{\\beta_ {t}} x _ {t} + \\frac {(1 - \\beta_ {t}) (1 - \\alpha_ {t})}{\\beta_ {t}} \\mu_ {\\theta} (x _ {t}, t). \\tag {4}", + "image_path": "865071fb80e15468472329d220b1c48f621582327750f63b9dd30f43d6dfd9d6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 629, + 547, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 667 + ], + "type": "text", + "content": "To get a text-to-image synthesis framework, SD guides the diffusion processes with a textual prompt " + }, + { + "bbox": [ + 304, + 629, + 547, + 667 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 629, + 547, + 667 + ], + "type": "text", + "content": ". Particularly for DDIM sampling, we get:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 332, + 672, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 672, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 332, + 672, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\left(\\frac {x _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} ^ {t} \\left(x _ {t} , \\tau\\right)}{\\sqrt {\\alpha_ {t}}}\\right) + \\tag {5} \\\\ \\sqrt {1 - \\alpha_ {t - 1}} \\epsilon_ {\\theta} ^ {t} (x _ {t}, \\tau), \\quad t = T, \\ldots , 1. \\\\ \\end{array}", + "image_path": "7107c36d0b130a1f19a4234bf44121998009115375839699f1924d5a4015cb91.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15956" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "It is worth noting that in SD, the function " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{t}(x_{t},\\tau)" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " is modeled as a neural network with a UNet-like [30] architecture composed of convolutional and (self- and cross-) attentional blocks. " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " is called the latent code of the signal " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " and there is a method [3] to apply a deterministic forward process to reconstruct the latent code " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " given a signal " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ". This method is known as DDIM inversion. Sometimes for simplicity, we will call " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "x_{t},t = 1,\\ldots ,T" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " also the latent codes of the initial signal " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 198, + 287, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 198, + 287, + 210 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 287, + 210 + ], + "type": "text", + "content": "3.2. Zero-Shot Text-to-Video Problem Formulation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": "Existing text-to-video synthesis methods require either costly training on a large-scale (ranging from " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "1M" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "15M" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " data-points) text-video paired data [41, 15, 39, 11, 14, 5] or tuning on a reference video [42]. To make video generation cheaper and easier, we propose a new problem: zero-shot text-to-video synthesis. Formally, given a text description " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " and a positive integer " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "m\\in \\mathbb{N}" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": ", the goal is to design a function " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " that outputs video frames " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{V}\\in \\mathbb{R}^{m\\times H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " (for predefined resolution " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": ") that exhibit temporal consistency. To determine the function " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": ", no training or fine-tuning must be performed on a video dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 355, + 288, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 288, + 403 + ], + "type": "text", + "content": "Our problem formulation provides a new paradigm for text-to-video. Noticeably, a zero-shot text-to-video method naturally benefits from quality improvements of text-to-image models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 421, + 107, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 421, + 107, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 421, + 107, + 432 + ], + "type": "text", + "content": "3.3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "content": "In this paper, we approach the zero-shot text-to-video task by exploiting the text-to-image synthesis power of Stable Diffusion (SD). As we need to generate videos instead of images, SD should operate on sequences of latent codes. The naive approach is to independently sample " + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "content": " latent codes from standard Gaussian distribution " + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "inline_equation", + "content": "x_{T}^{1},\\ldots ,x_{T}^{m}\\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "content": " and apply DDIM sampling to obtain the corresponding tensors " + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "inline_equation", + "content": "x_0^k" + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "inline_equation", + "content": "k = 1,\\dots,m" + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "content": ", followed by decoding to obtain the generated video sequence " + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}(x_0^k)\\}_{k = 1}^m\\in \\mathbb{R}^{m\\times H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "content": ". However, this leads to completely random generation of images sharing only the semantics described by " + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 443, + 287, + 598 + ], + "type": "text", + "content": " but neither object appearance nor motion coherence (see appendix Fig. 14, first row)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 601, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 686 + ], + "type": "text", + "content": "To address this issue, we propose to (i) introduce motion dynamics between the latent codes " + }, + { + "bbox": [ + 46, + 601, + 287, + 686 + ], + "type": "inline_equation", + "content": "x_{T}^{1}, \\ldots, x_{T}^{m}" + }, + { + "bbox": [ + 46, + 601, + 287, + 686 + ], + "type": "text", + "content": " to keep the global scene time consistent and (ii) use cross-frame attention mechanism to preserve the appearance and the identity of the foreground object. Each of the components of our method are described below in detail. The overview of our method can be found in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Note, to simplify notation, we will denote the entire sequence of latent codes by " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "x_{T}^{1:m} = [x_{T}^{1}, \\ldots, x_{T}^{m}]" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "type": "code", + "bbox": [ + 307, + 87, + 547, + 245 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 494, + 84 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 494, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 494, + 84 + ], + "type": "text", + "content": "Algorithm 1 Motion dynamics in latent codes" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\Delta t\\geq 0,m\\in \\mathbb{N},\\lambda >0,\\delta = (\\delta_x,\\delta_y)\\in" + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "text", + "content": " ,Stable Diffusion (SD) \n1: " + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "inline_equation", + "content": "x_{T}^{1}\\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "text", + "content": " random sample the first latent code \n2: " + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "inline_equation", + "content": "x_{T^{\\prime}}^{1}\\gets \\mathrm{DDIM\\_Backward}(x_{T}^{1},\\Delta t,SD)\\triangleright" + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "text", + "content": " perform " + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "text", + "content": " backward steps by SD \n3: for all " + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "inline_equation", + "content": "k = 2,3,\\ldots ,m" + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "text", + "content": " do \n4: " + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\delta^k\\gets \\lambda \\cdot (k - 1)\\delta \\triangleright" + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "text", + "content": " computing global translation vectors \n5: " + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "inline_equation", + "content": "W_{k}\\leftarrow" + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "text", + "content": " Warping by " + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\delta^k\\triangleright" + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "text", + "content": " defining warping functions \n6: " + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{T^{\\prime}}^{k}\\gets W_{k}(x_{T^{\\prime}}^{1})" + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "inline_equation", + "content": "x_{T}^{k}\\gets \\mathrm{DDPM\\_Forward}(\\tilde{x}_{T^{\\prime}}^{k},\\Delta t)\\triangleright \\mathrm{DDPM~forward}" + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "text", + "content": " for more motion freedom return " + }, + { + "bbox": [ + 307, + 87, + 547, + 245 + ], + "type": "inline_equation", + "content": "x_{T}^{1:m}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 306, + 265, + 485, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 265, + 485, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 265, + 485, + 277 + ], + "type": "text", + "content": "3.3.1 Motion Dynamics in Latent Codes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 284, + 546, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 284, + 546, + 333 + ], + "spans": [ + { + "bbox": [ + 305, + 284, + 546, + 333 + ], + "type": "text", + "content": "Instead of sampling the latent codes " + }, + { + "bbox": [ + 305, + 284, + 546, + 333 + ], + "type": "inline_equation", + "content": "x_{T}^{1:m}" + }, + { + "bbox": [ + 305, + 284, + 546, + 333 + ], + "type": "text", + "content": " randomly and independently from the standard Gaussian distribution, we construct them by performing the following steps (see also Alg. 1 and Fig. 2)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 339, + 545, + 538 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 313, + 339, + 544, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 544, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 544, + 365 + ], + "type": "text", + "content": "1. Randomly sample the latent code of the first frame: " + }, + { + "bbox": [ + 313, + 339, + 544, + 365 + ], + "type": "inline_equation", + "content": "x_{T}^{1} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 339, + 544, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 370, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 545, + 407 + ], + "type": "text", + "content": "2. Perform " + }, + { + "bbox": [ + 313, + 370, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\Delta t \\geq 0" + }, + { + "bbox": [ + 313, + 370, + 545, + 407 + ], + "type": "text", + "content": " DDIM backward steps on the latent code " + }, + { + "bbox": [ + 313, + 370, + 545, + 407 + ], + "type": "inline_equation", + "content": "x_{T}^{1}" + }, + { + "bbox": [ + 313, + 370, + 545, + 407 + ], + "type": "text", + "content": " by using the SD model and get the corresponding latent " + }, + { + "bbox": [ + 313, + 370, + 545, + 407 + ], + "type": "inline_equation", + "content": "x_{T'}^{1}" + }, + { + "bbox": [ + 313, + 370, + 545, + 407 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 370, + 545, + 407 + ], + "type": "inline_equation", + "content": "T' = T - \\Delta t" + }, + { + "bbox": [ + 313, + 370, + 545, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 414, + 545, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 545, + 451 + ], + "type": "text", + "content": "3. Define a direction " + }, + { + "bbox": [ + 313, + 414, + 545, + 451 + ], + "type": "inline_equation", + "content": "\\delta = (\\delta_x, \\delta_y) \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 414, + 545, + 451 + ], + "type": "text", + "content": " for the global scene and camera motion. By default " + }, + { + "bbox": [ + 313, + 414, + 545, + 451 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 414, + 545, + 451 + ], + "type": "text", + "content": " can be the main diagonal direction " + }, + { + "bbox": [ + 313, + 414, + 545, + 451 + ], + "type": "inline_equation", + "content": "\\delta_x = \\delta_y = 1" + }, + { + "bbox": [ + 313, + 414, + 545, + 451 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 458, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 545, + 506 + ], + "type": "text", + "content": "4. For each frame " + }, + { + "bbox": [ + 313, + 458, + 545, + 506 + ], + "type": "inline_equation", + "content": "k = 1,2,\\ldots ,m" + }, + { + "bbox": [ + 313, + 458, + 545, + 506 + ], + "type": "text", + "content": " we want to generate, compute the global translation vector " + }, + { + "bbox": [ + 313, + 458, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\delta^k = \\lambda \\cdot (k - 1)\\delta" + }, + { + "bbox": [ + 313, + 458, + 545, + 506 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 458, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 458, + 545, + 506 + ], + "type": "text", + "content": " is a hyperparameter controlling the amount of the global motion." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 513, + 544, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 544, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 544, + 538 + ], + "type": "text", + "content": "5. Apply the constructed motion (translation) flow " + }, + { + "bbox": [ + 313, + 513, + 544, + 538 + ], + "type": "inline_equation", + "content": "\\delta^{1:m}" + }, + { + "bbox": [ + 313, + 513, + 544, + 538 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 513, + 544, + 538 + ], + "type": "inline_equation", + "content": "x_{T'}^1" + }, + { + "bbox": [ + 313, + 513, + 544, + 538 + ], + "type": "text", + "content": ", denote the resulting sequence by " + }, + { + "bbox": [ + 313, + 513, + 544, + 538 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{T'}^{1:m}" + }, + { + "bbox": [ + 313, + 513, + 544, + 538 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 359, + 544, + 545, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 544, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 359, + 544, + 545, + 559 + ], + "type": "interline_equation", + "content": "\\tilde {x} _ {T ^ {\\prime}} ^ {k} = W _ {k} \\left(x _ {T ^ {\\prime}} ^ {1}\\right) \\text {f o r} k = 1, 2, \\dots , m, \\tag {6}", + "image_path": "4a584d81238f79e2fe2545a1e721d597423daecca2f484c30000b2d9043fba92.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 324, + 567, + 544, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 567, + 544, + 591 + ], + "spans": [ + { + "bbox": [ + 324, + 567, + 544, + 591 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 324, + 567, + 544, + 591 + ], + "type": "inline_equation", + "content": "W_{k}(x_{T^{\\prime}}^{1})" + }, + { + "bbox": [ + 324, + 567, + 544, + 591 + ], + "type": "text", + "content": " is the warping operation for translation by the vector " + }, + { + "bbox": [ + 324, + 567, + 544, + 591 + ], + "type": "inline_equation", + "content": "\\delta^{k}" + }, + { + "bbox": [ + 324, + 567, + 544, + 591 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 598, + 544, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 544, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 544, + 624 + ], + "type": "text", + "content": "6. Perform " + }, + { + "bbox": [ + 313, + 598, + 544, + 624 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 313, + 598, + 544, + 624 + ], + "type": "text", + "content": " DDPM forward steps on each of the latents " + }, + { + "bbox": [ + 313, + 598, + 544, + 624 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{T'}^{2:m}" + }, + { + "bbox": [ + 313, + 598, + 544, + 624 + ], + "type": "text", + "content": " and get the corresponding latent codes " + }, + { + "bbox": [ + 313, + 598, + 544, + 624 + ], + "type": "inline_equation", + "content": "x_{T}^{2:m}" + }, + { + "bbox": [ + 313, + 598, + 544, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "content": "Then we take the sequence " + }, + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "inline_equation", + "content": "x_{T}^{1:m}" + }, + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "content": " as the starting point of the backward (video) diffusion process. As a result, the latent codes generated with our proposed motion dynamics lead to better temporal consistency of the global scene as well as the background, see in the appendix, Sect. 8.1. Yet, the initial latent codes are not constraining enough to describe particular colors, identities or shapes, thus still" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15957" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 537, + 208 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 537, + 208 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 537, + 208 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 537, + 208 + ], + "type": "image", + "image_path": "122a24541992a5835ea4cd4ab2532dd7f4bba239b8a46f912763c3fdc5492f96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "lines": [ + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": "Figure 2: Method overview: Starting from a randomly sampled latent code " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "x_{T}^{1}" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": ", we apply " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": " DDIM backward steps to obtain " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "x_{T'}^{1}" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": " using a pre-trained Stable Diffusion model (SD). A specified motion field results for each frame " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": " in a warping function " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": " that turns " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "x_{T'}^{1}" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "x_{T'}^{k}" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": ". By enhancing the latent codes with motion dynamics, we determine the global scene and camera motion and achieve temporal consistency in the background and the global scene. A subsequent DDPM forward application delivers latent codes " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "x_{T}^{k}" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "k = 1, \\ldots, m" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": ". By using the (probabilistic) DDPM method, a greater degree of freedom is achieved with respect to the motion of objects (see appendix Sec. 8.1). Finally, the latent codes are passed to our modified SD model using the proposed cross-frame attention, which uses keys and values from the first frame to generate the image of frame " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "k = 1, \\ldots, m" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": ". By using cross-frame attention, the appearance and the identity of the foreground object are preserved throughout the sequence. Optionally, we apply background smoothing. To this end, we employ salient object detection to obtain for each frame " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": " a mask " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "M^{k}" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": " indicating the foreground pixels. Finally, for the background (using the mask " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "M^{k}" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": "), a convex combination between the latent code " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "x_{t}^{1}" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": " of frame one warped to frame " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": " and the latent code " + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "inline_equation", + "content": "x_{t}^{k}" + }, + { + "bbox": [ + 46, + 214, + 548, + 360 + ], + "type": "text", + "content": " is used to further improve the temporal consistency of the background." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 380, + 288, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 288, + 404 + ], + "type": "text", + "content": "leading to temporal inconsistencies, especially for the foreground object." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 418, + 253, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 253, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 253, + 430 + ], + "type": "text", + "content": "3.3.2 Reprogramming Cross-Frame Attention" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 484 + ], + "type": "text", + "content": "To address the issue mentioned above, we use a cross-frame attention mechanism to preserve the information about (in particular) the foreground object's appearance, shape, and identity throughout the generated video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 484, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 288, + 605 + ], + "type": "text", + "content": "To leverage the power of cross-frame attention and at the same time exploit a pretrained SD without retraining, we replace each of its self-attention layers with a cross-frame attention, with the attention for each frame being on the first frame. More precisely in the original SD UNet architecture " + }, + { + "bbox": [ + 46, + 484, + 288, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{t}(x_{t},\\tau)" + }, + { + "bbox": [ + 46, + 484, + 288, + 605 + ], + "type": "text", + "content": ", each self-attention layer takes a feature map " + }, + { + "bbox": [ + 46, + 484, + 288, + 605 + ], + "type": "inline_equation", + "content": "x\\in \\mathbb{R}^{h\\times w\\times c}" + }, + { + "bbox": [ + 46, + 484, + 288, + 605 + ], + "type": "text", + "content": ", linearly projects it into query, key, value features " + }, + { + "bbox": [ + 46, + 484, + 288, + 605 + ], + "type": "inline_equation", + "content": "Q,K,V\\in \\mathbb{R}^{h\\times w\\times c}" + }, + { + "bbox": [ + 46, + 484, + 288, + 605 + ], + "type": "text", + "content": ", and computes the layer output by the following formula (for simplicity described here for only one attention head) [38]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 609, + 287, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 609, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 76, + 609, + 287, + 637 + ], + "type": "interline_equation", + "content": "\\operatorname {S e l f - A t t n} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {c}}\\right) V. \\tag {7}", + "image_path": "9b0bf6469245a22e4984e155a0128174d7785b133953ead54e732df3cb6a5736.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 689 + ], + "type": "text", + "content": "In our case, each attention layer receives " + }, + { + "bbox": [ + 46, + 642, + 287, + 689 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 642, + 287, + 689 + ], + "type": "text", + "content": " inputs: " + }, + { + "bbox": [ + 46, + 642, + 287, + 689 + ], + "type": "inline_equation", + "content": "x^{1:m} = [x^1,\\dots,x^m]\\in \\mathbb{R}^{m\\times h\\times w\\times c}" + }, + { + "bbox": [ + 46, + 642, + 287, + 689 + ], + "type": "text", + "content": ". Hence, the linear projection layers produce " + }, + { + "bbox": [ + 46, + 642, + 287, + 689 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 642, + 287, + 689 + ], + "type": "text", + "content": " queries, keys, and values " + }, + { + "bbox": [ + 46, + 642, + 287, + 689 + ], + "type": "inline_equation", + "content": "Q^{1:m},K^{1:m},V^{1:m}" + }, + { + "bbox": [ + 46, + 642, + 287, + 689 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 712 + ], + "type": "text", + "content": "Therefore, we replace each self-attention layer with a cross-frame attention of each frame on the first frame as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 380, + 342, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 342, + 391 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 342, + 391 + ], + "type": "text", + "content": "follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 345, + 399, + 545, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 345, + 399, + 545, + 443 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {C r o s s - F r a m e - A t t n} \\left(Q ^ {k}, K ^ {1: m}, V ^ {1: m}\\right) = \\\\ \\operatorname {S o f t m a x} \\left(\\frac {Q ^ {k} \\left(K ^ {1}\\right) ^ {T}}{\\sqrt {c}}\\right) V ^ {1} \\tag {8} \\\\ \\end{array}", + "image_path": "34274776f0b26da26e16875a5096c8f3e3f6efdfb418e7c3b6fdd683794ef76b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 451, + 546, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 546, + 523 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 546, + 523 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 304, + 451, + 546, + 523 + ], + "type": "inline_equation", + "content": "k = 1, \\ldots, m" + }, + { + "bbox": [ + 304, + 451, + 546, + 523 + ], + "type": "text", + "content": ". By using cross frame attention, the appearance and structure of the objects and background as well as identities are carried over from the first frame to subsequent frames, which significantly increases the temporal consistency of the generated frames (see in the appendix the Figures 14, 16, 22, 23)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 539, + 486, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 486, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 486, + 551 + ], + "type": "text", + "content": "3.3.3 Background smoothing (Optional)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "text", + "content": "We further improve temporal consistency of the background using a convex combination of background-masked latent codes between the first frame and frame " + }, + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "text", + "content": ". This helps especially to generate videos from textual prompts when one or no initial image and no further guidance are provided." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "In more detail, given the generated sequence of our video generator, " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "x_0^{1:m}" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": ", we apply (an in-house solution for) salient object detection [40] to the decoded images to obtain a corresponding foreground mask " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "M^k" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": " for each frame " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": ". Then we warp " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "x_t^1" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": " according to the employed motion dynamics defined by " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "W_k" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": " and denote the result by " + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\hat{x}_t^k \\coloneqq W_k(x_t^1)" + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Background smoothing is achieved by a convex combination between the actual latent code " + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "x_{t}^{k}" + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": " and the warped" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15958" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 523, + 347 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 523, + 347 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 523, + 347 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 523, + 347 + ], + "type": "image", + "image_path": "b6f131c5c649c6adec56561c373e34336d84eca3cc6b006750a0ecf743bad24a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 353, + 546, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 353, + 546, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 546, + 378 + ], + "type": "text", + "content": "Figure 3: Text-to-Video results of our method. Depicted frames show that identities and appearances are temporally consistent and fitting to the textual prompt. For more results, see Appendix Sec. 8." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 396, + 282, + 544 + ], + "blocks": [ + { + "bbox": [ + 56, + 396, + 282, + 544 + ], + "lines": [ + { + "bbox": [ + 56, + 396, + 282, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 282, + 544 + ], + "type": "image", + "image_path": "daf48d03f9add4f33db4bd7ad68c90101df7e7ba8c2d823beb961dd3210654e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 552, + 285, + 565 + ], + "lines": [ + { + "bbox": [ + 50, + 552, + 285, + 565 + ], + "spans": [ + { + "bbox": [ + 50, + 552, + 285, + 565 + ], + "type": "text", + "content": "Figure 4: The overview of Text2Video-Zero + ControlNet" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 586, + 204, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 204, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 204, + 599 + ], + "type": "text", + "content": "latent code " + }, + { + "bbox": [ + 47, + 586, + 204, + 599 + ], + "type": "inline_equation", + "content": "\\hat{x}_t^k" + }, + { + "bbox": [ + 47, + 586, + 204, + 599 + ], + "type": "text", + "content": " on the background, i.e.," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 605, + 287, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 287, + 622 + ], + "type": "interline_equation", + "content": "\\bar {x} _ {t} ^ {k} = M ^ {k} \\odot x _ {t} ^ {k} + (1 - M ^ {k}) \\odot (\\alpha \\hat {x} _ {t} ^ {k} + (1 - \\alpha) x _ {t} ^ {k}), \\tag {9}", + "image_path": "cf6c9e558cceb54734ecc6e66583639696a4916d3bd858ecff3234b4faadbc49.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "k = 1, \\ldots, m" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": " is a hyperparameter, which we empirically choose " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\alpha = 0.6" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": ". Finally, DDIM sampling is employed on " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\overline{x}_t^k" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": ", which delivers video generation with background smoothing. We use background smoothing in our video generation from text when no guidance is provided. For an ablation study on background smoothing, see the appendix, Sec. 8.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 398, + 526, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 398, + 526, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 526, + 411 + ], + "type": "text", + "content": "3.4. Conditional and Specialized Text-to-Video" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 419, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 545, + 576 + ], + "type": "text", + "content": "Recently powerful controlling mechanisms [47, 22, 17] emerged to guide the diffusion process for text-to-image generation. Particularly, ControlNet [47] enables to condition the generation process using edges, pose, semantic masks, image depths, etc. However, a direct application of ControlNet in the video domain leads to temporal inconsistencies and to severe changes of object appearance, identity, and the background (see in the appendix Figures 14, 16, 22, 23). It turns out that our modifications on the basic diffusion process for videos result in more consistent videos guided by ControlNet conditions. We would like to point out again that our method does not require any fine-tuning or optimization processes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 578, + 546, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 546, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 546, + 698 + ], + "type": "text", + "content": "More specifically, ControlNet creates a trainable copy of the encoder (including the middle blocks) of the UNet " + }, + { + "bbox": [ + 304, + 578, + 546, + 698 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{t}(x_{t},\\tau)" + }, + { + "bbox": [ + 304, + 578, + 546, + 698 + ], + "type": "text", + "content": " while additionally taking the input " + }, + { + "bbox": [ + 304, + 578, + 546, + 698 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 578, + 546, + 698 + ], + "type": "text", + "content": " and a condition " + }, + { + "bbox": [ + 304, + 578, + 546, + 698 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 578, + 546, + 698 + ], + "type": "text", + "content": ", and adds the outputs of each layer to the skip-connections of the original UNet. Here " + }, + { + "bbox": [ + 304, + 578, + 546, + 698 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 578, + 546, + 698 + ], + "type": "text", + "content": " can be any type of condition, such as edge map, scribbles, pose (body landmarks), depth map, segmentation map, etc. The trainable branch is being trained on a specific domain for each type of the condition " + }, + { + "bbox": [ + 304, + 578, + 546, + 698 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 578, + 546, + 698 + ], + "type": "text", + "content": " resulting in an effective conditional text-to-image generation mechanism." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "To guide our video generation process with ControlNet" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15959" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 267, + 119 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 267, + 119 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 267, + 119 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 267, + 119 + ], + "type": "image", + "image_path": "d3ad2b3a4b5b049ed5496c6a32387da51b0b14902ac6e8116c10048648b83c44.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 121, + 228, + 130 + ], + "lines": [ + { + "bbox": [ + 107, + 121, + 228, + 130 + ], + "spans": [ + { + "bbox": [ + 107, + 121, + 228, + 130 + ], + "type": "text", + "content": "(a) a panda dancing in Antarctica" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 134, + 267, + 183 + ], + "blocks": [ + { + "bbox": [ + 67, + 134, + 267, + 183 + ], + "lines": [ + { + "bbox": [ + 67, + 134, + 267, + 183 + ], + "spans": [ + { + "bbox": [ + 67, + 134, + 267, + 183 + ], + "type": "image", + "image_path": "146925351d27b7ffa1c17ba502561b340a30916edea3f2652b40c89a654fc3dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 185, + 233, + 194 + ], + "lines": [ + { + "bbox": [ + 102, + 185, + 233, + 194 + ], + "spans": [ + { + "bbox": [ + 102, + 185, + 233, + 194 + ], + "type": "text", + "content": "(b) a cyborg dancing near a volcano" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 67, + 198, + 267, + 248 + ], + "blocks": [ + { + "bbox": [ + 67, + 198, + 267, + 248 + ], + "lines": [ + { + "bbox": [ + 67, + 198, + 267, + 248 + ], + "spans": [ + { + "bbox": [ + 67, + 198, + 267, + 248 + ], + "type": "image", + "image_path": "a7ccb181fc487596aeaab093619c3d96ab8d5083fe3a19785c70d3d74ab20209.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 249, + 244, + 259 + ], + "lines": [ + { + "bbox": [ + 91, + 249, + 244, + 259 + ], + "spans": [ + { + "bbox": [ + 91, + 249, + 244, + 259 + ], + "type": "text", + "content": "(c) an astronaut dancing in the outer space" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 67, + 310, + 267, + 360 + ], + "blocks": [ + { + "bbox": [ + 47, + 265, + 287, + 289 + ], + "lines": [ + { + "bbox": [ + 47, + 265, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 287, + 289 + ], + "type": "text", + "content": "Figure 5: Conditional generation with pose control. For more results see appendix, Sec. 10." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 310, + 267, + 360 + ], + "lines": [ + { + "bbox": [ + 67, + 310, + 267, + 360 + ], + "spans": [ + { + "bbox": [ + 67, + 310, + 267, + 360 + ], + "type": "image", + "image_path": "a4e34c0bba37c003312b0238e5d92a646b35d900005e18f29110e0d78e93650d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 361, + 242, + 371 + ], + "lines": [ + { + "bbox": [ + 93, + 361, + 242, + 371 + ], + "spans": [ + { + "bbox": [ + 93, + 361, + 242, + 371 + ], + "type": "text", + "content": "(a) oil painting of a girl dancing close-up" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 67, + 374, + 267, + 424 + ], + "blocks": [ + { + "bbox": [ + 67, + 374, + 267, + 424 + ], + "lines": [ + { + "bbox": [ + 67, + 374, + 267, + 424 + ], + "spans": [ + { + "bbox": [ + 67, + 374, + 267, + 424 + ], + "type": "image", + "image_path": "517241d129b050d6dca845a4a8b6745fc102cd095e105dd0b0fd23d033d798ac.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 425, + 253, + 435 + ], + "lines": [ + { + "bbox": [ + 82, + 425, + 253, + 435 + ], + "spans": [ + { + "bbox": [ + 82, + 425, + 253, + 435 + ], + "type": "text", + "content": "(b) Cyberpunk boy with a hat dancing close-up" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 441, + 287, + 465 + ], + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 465 + ], + "type": "text", + "content": "Figure 6: Conditional generation with edge control. For more results see appendix, Sec. 9." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "content": "we apply our method to the basic diffusion process, i.e. enrich the latent codes " + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "x_{T}^{1:m}" + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "content": " with motion information and change the self-attention into cross-frame attentions in the main UNet. While adopting the main UNet for video generation task, we apply the ControlNet pretrained copy branch per-frame on each " + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "x_{t}^{k}" + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "k = 1,\\dots,m" + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "content": " in each diffusion time-step " + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "inline_equation", + "content": "t = T,\\ldots ,1" + }, + { + "bbox": [ + 46, + 496, + 287, + 593 + ], + "type": "text", + "content": " and add the ControlNet branch outputs to the skip-connections of the main UNet." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 609, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 287, + 657 + ], + "type": "text", + "content": "Furthermore, for our conditional generation task, we adopted the weights of specialized DreamBooth (DB) [31] models1. This gives us specialized time-consistent video generations (see Fig. 7)." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 526, + 119 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 526, + 119 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 526, + 119 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 526, + 119 + ], + "type": "image", + "image_path": "825a7a5b591098e7077a2d104041b6b91451f42f4ae2f62bc2616ad6f45e35b0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 120, + 507, + 131 + ], + "lines": [ + { + "bbox": [ + 345, + 120, + 507, + 131 + ], + "spans": [ + { + "bbox": [ + 345, + 120, + 507, + 131 + ], + "type": "text", + "content": "(a) oil painting of a beautiful girl avatar style" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 326, + 134, + 526, + 184 + ], + "blocks": [ + { + "bbox": [ + 326, + 134, + 526, + 184 + ], + "lines": [ + { + "bbox": [ + 326, + 134, + 526, + 184 + ], + "spans": [ + { + "bbox": [ + 326, + 134, + 526, + 184 + ], + "type": "image", + "image_path": "1700324d92b6c560ad0c09d2aa1ac3ba41c914f5731b997e0d6df048544f3584.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 185, + 452, + 195 + ], + "lines": [ + { + "bbox": [ + 400, + 185, + 452, + 195 + ], + "spans": [ + { + "bbox": [ + 400, + 185, + 452, + 195 + ], + "type": "text", + "content": "(b) gta-5 style" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 201, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 306, + 201, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 545, + 224 + ], + "type": "text", + "content": "Figure 7: Conditional generation with edge control and DB models." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 247, + 436, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 247, + 436, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 436, + 258 + ], + "type": "text", + "content": "3.5. Video Instruct-Pix2Pix" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 266, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 545, + 433 + ], + "type": "text", + "content": "With the rise of text-guided image editing methods such as Prompt2Prompt [9], Instruct-Pix2Pix [2], SDEdit [20], etc., text-guided video editing approaches emerged [1, 16, 42]. While these methods require complex optimization processes, our approach enables the adoption of any SD-based text-guided image editing algorithm to the video domain without any training or fine-tuning. Here we take the text-guided image editing method Instruct-Pix2Pix and combine it with our approach. More precisely, we change the self-attention mechanisms in Instruct-Pix2Pix to cross-frame attentions according to Eq. 8. Our experiments show that this adaptation significantly improves the consistency of the edited videos (see Fig. 9) over the naive per-frame usage of Instruct-Pix2Pix." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 445, + 387, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 445, + 387, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 445, + 387, + 459 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 465, + 439, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 439, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 439, + 478 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": "We take the Stable Diffusion [29] code2 with its pretrained weights from version 1.5 as basis and implement our modifications. For each video, we generate " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "m = 8" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": " frames with " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": " resolution. However, our framework allows generating any number of frames, either by increasing " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": ", or by employing our method in an auto-regressive fashion where the last generated frame " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": " becomes the first frame in computing the next " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": " frames. For all text-to-video generation experiments, we take " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "T' = 881" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "T = 941" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": " without specific tuning, while for conditional and specialized generation, and for Video Instruct-Pix2Pix, we take " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "T' = T = 1000" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 628, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 628, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 628, + 545, + 665 + ], + "type": "text", + "content": "For a conditional generation, we use the codebase3 of ControlNet [47]. For specialized models, we take DB [31] models from publicly available sources. For Video Instruct" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 683, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 286, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 47, + 683, + 286, + 712 + ], + "type": "text", + "content": "Avatar model: https://civitai.com/models/9968/ avatar-style. GTA-5 model: https://civitai.com/models/1309/gta5-artwork-diffusion." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 673, + 545, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 545, + 702 + ], + "type": "text", + "content": "2https://github.com/huggingface/diffusers. We also benefit from the codebase of Tune-A-Video https://github.com/showlab/Tune-A-Video." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 703, + 516, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 516, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 516, + 712 + ], + "type": "text", + "content": "3https://github.com/lllyasviel/ControlNet." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15960" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 267, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 267, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 267, + 85 + ], + "type": "text", + "content": "Pix2Pix, we use the codebase of Instruct Pix2Pix [2]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 160, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 160, + 105 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 160, + 105 + ], + "type": "text", + "content": "4.2. Qualitative Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 110, + 287, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 110, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 287, + 171 + ], + "type": "text", + "content": "All applications of Text2Video-Zero show that it successfully generates videos where the global scene and the background are time consistent and the context, appearance, and identity of the foreground object are maintained throughout the entire sequence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 171, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 266 + ], + "type": "text", + "content": "In the case of text-to-video, we observe that it generates high-quality videos that are well-aligned to the text prompt (see Fig. 3 and the appendix). For instance, the depicted panda shows a naturally walking on the street. Likewise, using additional guidance from edges or poses (see Fig. 5, Fig. 6 and Fig. 7 and the appendix), high quality videos are generated matching the prompt and the guidance that show great temporal consistency and identity preservation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 266, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 287, + 302 + ], + "type": "text", + "content": "Videos generated by Video Instruct-Pix2Pix (see Fig. 1 and the appendix) possess high-fidelity with respect to the input video, while following closely the instruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 311, + 197, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 197, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 197, + 323 + ], + "type": "text", + "content": "4.3. Comparison with Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 329, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 329, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 287, + 389 + ], + "type": "text", + "content": "We compare our method with two publicly available baselines: CogVideo [15] and Tune-A-Video [42]. Since CogVideo is a text-to-video method we compare with it in pure text-guided video synthesis settings. With Tune-A-Video we compare in our Video Instruct-Pix2Pix setting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 404, + 189, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 404, + 189, + 416 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 189, + 416 + ], + "type": "text", + "content": "4.3.1 Quantitative Comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 423, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 519 + ], + "type": "text", + "content": "To show quantitative results, we evaluate the CLIP score [10], which indicates video-text alignment. We randomly take 25 videos generated by CogVideo and synthesize corresponding videos using the same prompts according to our method. The CLIP scores for our method and CogVideo are 31.19 and 29.63, respectively. Our method thus slightly outperforms CogVideo, even though the latter has 9.4 billion parameters and requires large-scale training on videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 535, + 183, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 183, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 183, + 547 + ], + "type": "text", + "content": "4.3.2 Qualitative Comparison" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 554, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 685 + ], + "type": "text", + "content": "We present several results of our method in Fig. 8 and provide a qualitative comparison to CogVideo [15]. Both methods show good temporal consistency throughout the sequence, preserving the identity of the object and background. However, our method shows better text-video alignment. For instance, while our method correctly generates a video of a man riding a bicycle in the sunshine in Fig. 8(b), CogVideo sets the background to moon light. Also in Fig. 8(a), our method correctly shows a man running in the snow, while neither the snow nor a man running are clearly visible in the video generated by CogVideo." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 179 + ], + "type": "text", + "content": "Qualitative results of Video Instruct-Pix2Pix and a visual comparison with per-frame Instruct-Pix2Pix and Tune-A-Video are shown in Fig. 9. While Instruct-Pix2Pix shows a good editing performance per frame, it lacks temporal consistency. This becomes evident especially in the video depicting a skiing person, where the snow and the sky are drawn using different styles and colors. Using our Video Instruct-Pix2Pix method, these issues are solved resulting in temporally consistent video edits throughout the sequence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 180, + 546, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 546, + 384 + ], + "type": "text", + "content": "While Tune-A-Video creates temporally consistent video generations, it is less aligned to the instruction guidance than our method, struggles creating local edits and losses details of the input sequence. This becomes apparent when looking at the edit of the dancer video depicted in Fig. 9 (left side). In contrast to Tune-A-Video, our method draws the entire dress brighter and at the same time better preserves the background, e.g. the wall behind the dancer is almost kept the same. Tune-A-Video draws a severely modified wall. Moreover, our method is more faithful to the input details, e.g., Video Instruct-Pix2Pix draws the dancer using the pose exactly as provided (Fig. 9 left), and shows all skiing persons appearing in the input video (compare last frame of Fig. 9(right)), in constraint to Tune-A-Video. All the above-mentioned weaknesses of Tune-A-Video can also be observed in our additional evaluations that are provided in the appendix, Figures 25, 26." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 392, + 399, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 399, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 399, + 404 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 411, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 482 + ], + "type": "text", + "content": "We perform several ablation studies and provide the results in the appendix, Sect. 8.1. Namely, we analyse background smoothing, " + }, + { + "bbox": [ + 304, + 411, + 545, + 482 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 411, + 545, + 482 + ], + "type": "text", + "content": ", and two main components of our method: making the initial latent codes coherent to a motion, and using cross-frame attention on the first frame instead of self-attention." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 494, + 473, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 494, + 473, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 494, + 473, + 506 + ], + "type": "text", + "content": "5. Limitations and Future Work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 514, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 622 + ], + "type": "text", + "content": "The main limitation of this work is the inability to generate longer videos with sequences of actions. Future research may target enriching our method with techniques such as autoregressive scene action generation, while keeping the training-free spirit. Overall, our method is focused on generating video key-frames as in the first stage of Imagen Video [11] and Make-A-Video [33], and can thus be considered as good basis for longer and smoother video generation by integrating temporal upsampling." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 633, + 378, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 378, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 378, + 645 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we addressed the problem of zero-shot text-to-video synthesis and proposed a novel method for time-consistent video generation. Our approach does not require any optimization or fine-tuning, making text-to-video generation and its applications affordable for everyone. We" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 222, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 222, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 222, + 713 + ], + "type": "text", + "content": "4https://github.com/timothybrooks/instruct-pix2pix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15961" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 69, + 129, + 123 + ], + "blocks": [ + { + "bbox": [ + 76, + 69, + 129, + 123 + ], + "lines": [ + { + "bbox": [ + 76, + 69, + 129, + 123 + ], + "spans": [ + { + "bbox": [ + 76, + 69, + 129, + 123 + ], + "type": "image", + "image_path": "6abe8efd60ccf723ec1e9bfcb66f4574cb2f3051b4762cb63fac75b7a0dd275c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 130, + 70, + 184, + 123 + ], + "blocks": [ + { + "bbox": [ + 130, + 70, + 184, + 123 + ], + "lines": [ + { + "bbox": [ + 130, + 70, + 184, + 123 + ], + "spans": [ + { + "bbox": [ + 130, + 70, + 184, + 123 + ], + "type": "image", + "image_path": "02142f2d2c7c8c9739bc72cbc1006c4817647c9f722476cd81bb514a1632fb45.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 184, + 70, + 236, + 123 + ], + "blocks": [ + { + "bbox": [ + 184, + 70, + 236, + 123 + ], + "lines": [ + { + "bbox": [ + 184, + 70, + 236, + 123 + ], + "spans": [ + { + "bbox": [ + 184, + 70, + 236, + 123 + ], + "type": "image", + "image_path": "2f1d2962f5bfec19bf745c3ed67c17bd35952ef1ad7890897081c7ef000ff228.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 124, + 355, + 134 + ], + "lines": [ + { + "bbox": [ + 238, + 124, + 355, + 134 + ], + "spans": [ + { + "bbox": [ + 238, + 124, + 355, + 134 + ], + "type": "text", + "content": "(a) a man is running in the snow" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 237, + 70, + 290, + 123 + ], + "blocks": [ + { + "bbox": [ + 237, + 70, + 290, + 123 + ], + "lines": [ + { + "bbox": [ + 237, + 70, + 290, + 123 + ], + "spans": [ + { + "bbox": [ + 237, + 70, + 290, + 123 + ], + "type": "image", + "image_path": "4ef23ae7ad417795ac8e9930c23fbe718052fb3aecf46923b05c8940d408116f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 302, + 70, + 356, + 123 + ], + "blocks": [ + { + "bbox": [ + 302, + 70, + 356, + 123 + ], + "lines": [ + { + "bbox": [ + 302, + 70, + 356, + 123 + ], + "spans": [ + { + "bbox": [ + 302, + 70, + 356, + 123 + ], + "type": "image", + "image_path": "8cc03a3fce8e50a676a49e08b32b6c6f2e20cea5cb31f87195034f3043726e62.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 358, + 70, + 410, + 123 + ], + "blocks": [ + { + "bbox": [ + 358, + 70, + 410, + 123 + ], + "lines": [ + { + "bbox": [ + 358, + 70, + 410, + 123 + ], + "spans": [ + { + "bbox": [ + 358, + 70, + 410, + 123 + ], + "type": "image", + "image_path": "14901d1229a7627c775f5d890f9782ec261cd0f088986498f5c1b5d1a91fd28a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 410, + 70, + 463, + 123 + ], + "blocks": [ + { + "bbox": [ + 410, + 70, + 463, + 123 + ], + "lines": [ + { + "bbox": [ + 410, + 70, + 463, + 123 + ], + "spans": [ + { + "bbox": [ + 410, + 70, + 463, + 123 + ], + "type": "image", + "image_path": "513fb00d780079568f573e4efe0150ebf79212d9c5b007cfd5e2a13cf740a0d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 465, + 70, + 517, + 123 + ], + "blocks": [ + { + "bbox": [ + 465, + 70, + 517, + 123 + ], + "lines": [ + { + "bbox": [ + 465, + 70, + 517, + 123 + ], + "spans": [ + { + "bbox": [ + 465, + 70, + 517, + 123 + ], + "type": "image", + "image_path": "0b0f7fad9e4f521742e716a39bedcd784b1e3a7b6f0498f3a4d07c7eed086b18.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 76, + 137, + 129, + 192 + ], + "blocks": [ + { + "bbox": [ + 76, + 137, + 129, + 192 + ], + "lines": [ + { + "bbox": [ + 76, + 137, + 129, + 192 + ], + "spans": [ + { + "bbox": [ + 76, + 137, + 129, + 192 + ], + "type": "image", + "image_path": "03ad100fe744a479696e12c8e8ed9972586357f6dd7704cc2052c4057e1871aa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 130, + 137, + 183, + 192 + ], + "blocks": [ + { + "bbox": [ + 130, + 137, + 183, + 192 + ], + "lines": [ + { + "bbox": [ + 130, + 137, + 183, + 192 + ], + "spans": [ + { + "bbox": [ + 130, + 137, + 183, + 192 + ], + "type": "image", + "image_path": "78a6506b39f07f779c7501d2e5ebc2373cf7120ecd0843e0aa64ca5104a2c138.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 184, + 137, + 236, + 192 + ], + "blocks": [ + { + "bbox": [ + 184, + 137, + 236, + 192 + ], + "lines": [ + { + "bbox": [ + 184, + 137, + 236, + 192 + ], + "spans": [ + { + "bbox": [ + 184, + 137, + 236, + 192 + ], + "type": "image", + "image_path": "29a51a72e6f73ce46daffdb36bf09a7f347d6f0f158b85c94c99b84c3b2065cc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 192, + 376, + 202 + ], + "lines": [ + { + "bbox": [ + 217, + 192, + 376, + 202 + ], + "spans": [ + { + "bbox": [ + 217, + 192, + 376, + 202 + ], + "type": "text", + "content": "(b) a man is riding a bicycle in the sunshine" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 237, + 137, + 290, + 192 + ], + "blocks": [ + { + "bbox": [ + 237, + 137, + 290, + 192 + ], + "lines": [ + { + "bbox": [ + 237, + 137, + 290, + 192 + ], + "spans": [ + { + "bbox": [ + 237, + 137, + 290, + 192 + ], + "type": "image", + "image_path": "b847fedda237ea961f3095dc645dd670fd215bf04ccd158a6f21eab5143c84e5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 303, + 137, + 356, + 192 + ], + "blocks": [ + { + "bbox": [ + 303, + 137, + 356, + 192 + ], + "lines": [ + { + "bbox": [ + 303, + 137, + 356, + 192 + ], + "spans": [ + { + "bbox": [ + 303, + 137, + 356, + 192 + ], + "type": "image", + "image_path": "7d9085b17c5c4083691edca26fac6f0759dabe67ea8c568c41eb9f4232a6f9e0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 357, + 137, + 410, + 192 + ], + "blocks": [ + { + "bbox": [ + 357, + 137, + 410, + 192 + ], + "lines": [ + { + "bbox": [ + 357, + 137, + 410, + 192 + ], + "spans": [ + { + "bbox": [ + 357, + 137, + 410, + 192 + ], + "type": "image", + "image_path": "b438330d6528881f97ad4834fa6d6e05228a783afc07052f52f55178df4e180c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 410, + 137, + 463, + 192 + ], + "blocks": [ + { + "bbox": [ + 410, + 137, + 463, + 192 + ], + "lines": [ + { + "bbox": [ + 410, + 137, + 463, + 192 + ], + "spans": [ + { + "bbox": [ + 410, + 137, + 463, + 192 + ], + "type": "image", + "image_path": "6ba3bdb805425ce2ad7bce7cdb062e10ef911cc0169f129d21e69b447148d0f5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 465, + 137, + 517, + 192 + ], + "blocks": [ + { + "bbox": [ + 465, + 137, + 517, + 192 + ], + "lines": [ + { + "bbox": [ + 465, + 137, + 517, + 192 + ], + "spans": [ + { + "bbox": [ + 465, + 137, + 517, + 192 + ], + "type": "image", + "image_path": "a8e18d4bf4a8d27fc085b6fe4e1669861fec77c0bce8bd34e84f032b14ca3c14.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 76, + 205, + 129, + 259 + ], + "blocks": [ + { + "bbox": [ + 76, + 205, + 129, + 259 + ], + "lines": [ + { + "bbox": [ + 76, + 205, + 129, + 259 + ], + "spans": [ + { + "bbox": [ + 76, + 205, + 129, + 259 + ], + "type": "image", + "image_path": "42a8be58677d32fc464e3d22d2b1e430442a35ae3112e9d6d776eb18889bfba0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 276, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 545, + 300 + ], + "type": "text", + "content": "Figure 8: Comparison of our method vs CogVideo on text-to-video generation task (left is ours, right is CogVideo [15]). For more comparisons, see appendix Fig. 18." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 130, + 205, + 183, + 259 + ], + "blocks": [ + { + "bbox": [ + 130, + 205, + 183, + 259 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 183, + 259 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 183, + 259 + ], + "type": "image", + "image_path": "453efd49d5f4494cf18f24b6db0d44899c0502ce70277990668181abe65016ae.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 184, + 205, + 236, + 259 + ], + "blocks": [ + { + "bbox": [ + 184, + 205, + 236, + 259 + ], + "lines": [ + { + "bbox": [ + 184, + 205, + 236, + 259 + ], + "spans": [ + { + "bbox": [ + 184, + 205, + 236, + 259 + ], + "type": "image", + "image_path": "2c9fcb8cf777e9e0b26f320d1e131ed8a6dd0b438a884752264add079fa6f1f8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 237, + 205, + 290, + 259 + ], + "blocks": [ + { + "bbox": [ + 237, + 205, + 290, + 259 + ], + "lines": [ + { + "bbox": [ + 237, + 205, + 290, + 259 + ], + "spans": [ + { + "bbox": [ + 237, + 205, + 290, + 259 + ], + "type": "image", + "image_path": "2ac10858d4af622e178cc7acaefc2d7c40248dfeb5369f110578b829daac52ab.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 260, + 353, + 270 + ], + "lines": [ + { + "bbox": [ + 240, + 260, + 353, + 270 + ], + "spans": [ + { + "bbox": [ + 240, + 260, + 353, + 270 + ], + "type": "text", + "content": "(c) a man is walking in the rain" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 303, + 205, + 356, + 259 + ], + "blocks": [ + { + "bbox": [ + 303, + 205, + 356, + 259 + ], + "lines": [ + { + "bbox": [ + 303, + 205, + 356, + 259 + ], + "spans": [ + { + "bbox": [ + 303, + 205, + 356, + 259 + ], + "type": "image", + "image_path": "65f570a45b25502db7088cd712e9db0c3d4ca04a4ec4f1066e00e85ce661fd30.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 357, + 205, + 410, + 259 + ], + "blocks": [ + { + "bbox": [ + 357, + 205, + 410, + 259 + ], + "lines": [ + { + "bbox": [ + 357, + 205, + 410, + 259 + ], + "spans": [ + { + "bbox": [ + 357, + 205, + 410, + 259 + ], + "type": "image", + "image_path": "43096e243642767407c61fc05c4ca865d4474303adea5f53c21e3d37038fd8e5.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 410, + 205, + 463, + 259 + ], + "blocks": [ + { + "bbox": [ + 410, + 205, + 463, + 259 + ], + "lines": [ + { + "bbox": [ + 410, + 205, + 463, + 259 + ], + "spans": [ + { + "bbox": [ + 410, + 205, + 463, + 259 + ], + "type": "image", + "image_path": "99d564c034db10e955e246dbf50133e35caff3a65b5dcee6688c8736f074e14b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 465, + 205, + 517, + 259 + ], + "blocks": [ + { + "bbox": [ + 465, + 205, + 517, + 259 + ], + "lines": [ + { + "bbox": [ + 465, + 205, + 517, + 259 + ], + "spans": [ + { + "bbox": [ + 465, + 205, + 517, + 259 + ], + "type": "image", + "image_path": "a8861c84fe2e35e07e4ac85ad0e4a7b3e35beb99f1a5ae66c3bef9c7578a012f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 56, + 321, + 545, + 513 + ], + "blocks": [ + { + "bbox": [ + 56, + 321, + 545, + 513 + ], + "lines": [ + { + "bbox": [ + 56, + 321, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 545, + 513 + ], + "type": "image", + "image_path": "b5e0f36cb746698940a75050946403f57e4e0d127ca69d23ad45a59261dbd117.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 521, + 544, + 546 + ], + "lines": [ + { + "bbox": [ + 46, + 521, + 544, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 544, + 546 + ], + "type": "text", + "content": "Figure 9: Comparison of Video Instruct-Pix2Pix(ours) with Tune-A-Video and per-frame Instruct-Pix2Pix. For more comparisons see our appendix." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 567, + 289, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 289, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 289, + 710 + ], + "type": "text", + "content": "demonstrated the effectiveness of our method for various applications, including conditional and specialized video generation, and Video Instruct-Pix2Pix, i.e., instruction-guided video editing. Our contributions to the field include presenting a new problem of zero-shot text-to-video synthesis, showing the utilization of text-to-image diffusion models for generating time-consistent videos, and providing evidence of the effectiveness of our method for various video synthesis applications. We believe that our proposed method will open up new possibilities for video generation and editing, making it accessible and affordable for everyone." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 547, + 688 + ], + "type": "text", + "content": "Acknowledgments. This material is partially based upon work supported by the National Science Foundation CAREER Award #2239840, and the National AI Institute for Exceptional Education (Award #2229873) by National Science Foundation and the Institute of Education Sciences, U.S. Department of Education. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation, the Institute of Education Sciences, or the U.S. Department of Education." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15962" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "text", + "content": "[1] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 707-723. Springer, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 287, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 287, + 180 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 2, 7, 8, 13, 25, 26" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "text", + "content": "[3] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 258 + ], + "type": "text", + "content": "[4] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers. arXiv preprint arXiv:2204.14217, 2022.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 287, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 287, + 305 + ], + "type": "text", + "content": "[5] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. arXiv preprint arXiv:2302.03011, 2023. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 307, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 287, + 351 + ], + "type": "text", + "content": "[6] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 352, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 418 + ], + "type": "text", + "content": "[7] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 89-106. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 419, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 287, + 463 + ], + "type": "text", + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 464, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 287, + 508 + ], + "type": "text", + "content": "[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "text", + "content": "[10] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "text", + "content": "[11] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3, 4, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 611, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 644 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 668 + ], + "type": "text", + "content": "[13] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[14] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 3, 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[15] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3, 4, 8, 9, 12, 13, 20" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 546, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 160 + ], + "type": "text", + "content": "[16] Yao-Chih Lee, Ji-Ze Genevieve Jang, Yi-Ting Chen, Elizabeth Qiu, and Jia-Bin Huang. Shape-aware text-driven layered video editing. arXiv e-prints, pages arXiv-2301, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 162, + 547, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 547, + 227 + ], + "type": "text", + "content": "[17] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 228, + 546, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 546, + 303 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 546, + 303 + ], + "type": "text", + "content": "[18] Haoming Lu, Hazarapet Tunanyan, Kai Wang, Shant Navasardyan, Zhangyang Wang, and Humphrey Shi. Specialist diffusion: Plug-and-play sample-efficient fine-tuning of text-to-image diffusion models to learn any unseen style. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14267-14276, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 304, + 546, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 546, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 546, + 337 + ], + "type": "text", + "content": "[19] Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Generating images from captions with attention. arXiv preprint arXiv:1511.02793, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 338, + 546, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 546, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 546, + 380 + ], + "type": "text", + "content": "[20] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 382, + 546, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 546, + 426 + ], + "type": "text", + "content": "[21] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 426, + 546, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 546, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 546, + 470 + ], + "type": "text", + "content": "[22] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 471, + 546, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 546, + 524 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 546, + 524 + ], + "type": "text", + "content": "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 525, + 546, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 546, + 579 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 546, + 579 + ], + "type": "text", + "content": "[24] Tingting Qiao, Jing Zhang, Duanqing Xu, and Dacheng Tao. Mirrorgan: Learning text-to-image generation by redescription. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1505-1514, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 581, + 546, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 546, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 546, + 644 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 546, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 546, + 689 + ], + "type": "text", + "content": "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "type": "text", + "content": "[27] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15963" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 105 + ], + "type": "text", + "content": "Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[28] Scott E Reed, Zeynep Akata, Santosh Mohan, Samuel Tenka, Bernt Schiele, and Honglak Lee. Learning what and where to draw. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 287, + 207 + ], + "type": "text", + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 287, + 274 + ], + "type": "text", + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 276, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 287, + 329 + ], + "type": "text", + "content": "[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 3, 7, 22" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 397 + ], + "type": "text", + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "text", + "content": "[33] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 455, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 509 + ], + "type": "text", + "content": "[34] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 511, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 543 + ], + "type": "text", + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "text", + "content": "[36] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "text", + "content": "[37] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 668 + ], + "type": "text", + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[39] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 540 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "textual description. arXiv preprint arXiv:2210.02399, 2022. 3,4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[40] Wenguan Wang, Qiuxia Lai, Huazhu Fu, Jianbing Shen, Haibin Ling, and Ruigang Yang. Salient object detection in the deep learning era: An in-depth survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3239-3259, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 152, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 217 + ], + "type": "text", + "content": "[41] Chenfei Wu, Jian Liang, Lei Ji, Fan Yang, Yuejian Fang, Daxin Jiang, and Nan Duan. Niwa: Visual synthesis pretraining for neural visual world creation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 720-736. Springer, 2022. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 219, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 273 + ], + "type": "text", + "content": "[42] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 2, 3, 4, 7, 8, 12, 13, 26" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 274, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 339 + ], + "type": "text", + "content": "[43] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Finegrained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 341, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 341, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 341, + 545, + 384 + ], + "type": "text", + "content": "[44] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. arXiv preprint arXiv:2211.08332, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 386, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 386, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 386, + 545, + 441 + ], + "type": "text", + "content": "[45] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 442, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 442, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 308, + 442, + 545, + 506 + ], + "type": "text", + "content": "[46] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 508, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 508, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 308, + 508, + 545, + 540 + ], + "type": "text", + "content": "[47] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 3, 6, 7" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15964" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/21a8728d-cd5e-44be-95fc-d79147ae763b_content_list.json b/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/21a8728d-cd5e-44be-95fc-d79147ae763b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ab6f8aa5c9018e8fa3459a7d3b0dcf6137f36990 --- /dev/null +++ b/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/21a8728d-cd5e-44be-95fc-d79147ae763b_content_list.json @@ -0,0 +1,1825 @@ +[ + { + "type": "text", + "text": "TextMania: Enriching Visual Feature by Text-driven Manifold Augmentation", + "text_level": 1, + "bbox": [ + 94, + 130, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Moon Ye-Bin", + "bbox": [ + 112, + 179, + 233, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jisoo Kim", + "bbox": [ + 272, + 180, + 366, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongyeob Kim3", + "bbox": [ + 405, + 180, + 539, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kilho Son", + "bbox": [ + 578, + 180, + 669, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tae-Hyun $\\mathrm{Oh}^{1,5,6}$", + "bbox": [ + 710, + 180, + 849, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dept. of Electrical Engineering and $^{5}$ Grad. School of Artificial Intelligence, POSTECH", + "bbox": [ + 135, + 208, + 833, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Columbia University", + "bbox": [ + 176, + 227, + 352, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Sungkyunkwan University", + "bbox": [ + 392, + 227, + 611, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Microsoft Azure", + "bbox": [ + 651, + 227, + 792, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ Institute for Convergence Research and Education in Advanced Technology, Yonsei University", + "bbox": [ + 107, + 244, + 861, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://textmania.github.io/", + "bbox": [ + 341, + 267, + 620, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 316, + 313, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose TextManiA, a text-driven manifold augmentation method that semantically enriches visual feature spaces, regardless of class distribution. TextManiA augments visual data with intra-class semantic perturbation by exploiting easy-to-understand visually mimetic words, i.e., attributes. This work is built on an interesting hypothesis that general language models, e.g., BERT and GPT, encompass visual information to some extent, even without training on visual training data. Given the hypothesis, TextManiA transfers pre-trained text representation obtained from a well-established large language encoder to a target visual feature space being learned. Our extensive analysis hints that the language encoder indeed encompasses visual information at least useful to augment visual representation. Our experiments demonstrate that TextManiA is particularly powerful in scarce samples with class imbalance as well as even distribution. We also show compatibility with the label mix-based approaches in evenly distributed scarce data.", + "bbox": [ + 75, + 349, + 473, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 648, + 209, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning models, e.g., neural networks, are known to perform well on visual recognition tasks when training and testing datasets present similar distributions [4]. However, their performance often degrades considerably when evaluated in subtly different distributions [69]. One effective way to enhance the generalization ability of a model against such data distribution shifts would be data augmentation [18, 86, 85, 44, 41, 73]. Augmenting data enlarges the support of the training distribution formed by given samples and yields the effect of increasing the amount of data even without additional laborious data collection. By training on augmented data, decision boundaries are smoothed, and the generalization ability of the model is improved [73].", + "bbox": [ + 75, + 674, + 470, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There has been a distinctive and successful line of research for label mix-based data augmentation, such as", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77558d45f938bf35ae6f8307b24cc555b6a5a3f130c0e3d77f83544e7a98931e.jpg", + "image_caption": [ + "Figure 1. Illustration of TextMania. Our method augments the target visual feature by leveraging text embedding of the visually mimetic words, which are comprehensible and semantically rich. For example, when the text of the existing class \"bull\" is manipulated as \"red bull\" by adding the attribute \"red,\" we can get augmented visual features by reflecting the difference of text embeddings. In this way, TextMania densifies sparse visual feature space using various attributes text." + ], + "image_footnote": [], + "bbox": [ + 509, + 323, + 893, + 435 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mixup [86], CutMix [85], and manifold Mixup [73], which are effective for model generalization and calibration [25]. The effectiveness of those label mix-based approaches is attributed to semantic perturbation by label mixing [86, 73, 85]. This is a distinctive property from other lines of data augmentation methods, e.g., [74, 44, 41, 64, 13], where they synthesize diverse virtual data that appear differently but retain class semantics of original contents. However, we found that the performance of mix-based augmentation methods is noticeably degraded when training with skewed class distribution having scarce samples for non-major classes, i.e., long-tailed distribution. In real-world, data often exhibit long-tailed class distribution (e.g., Pareto distribution), which cannot be dealt with the prevalent mix-based approaches. This motivates us to seek a semantically rich data augmentation effective for limited data regimes, including long-tailed distribution, scarce data, and few-shot cases.", + "bbox": [ + 496, + 593, + 895, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose TextMania, a text-driven manifold augmentation for visual features, which is effective for long-tail classes and scarce data. Our TextMania is based", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2526", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on an interesting hypothesis that general language models, e.g., BERT [17] and GPT [55], have learned visual information to some extent that can be transferred to visual feature spaces even with no visual training data. With this hypothesis, we semantically enrich the target visual feature space to be trained by leveraging visually mimetic texts, encoded with general language models and transferred to the target space. Specifically, TextMania encodes meaningful attributes such as \"red\" and \"large\" to vectors by computing the difference between text embeddings with and without attributes. We add the attribute embeddings to target visual features to mimic those attributes on the target visual feature space. Figure 1 illustrates the augmentation process of TextMania. The input feature (e.g., the visual feature of \"bull\") is manipulated by adding the attribute vector induced by the attribute text (e.g., \"red\"), which yields the augmented visual feature (e.g., \"red bull\"). Thanks to the text modality properties, the augmentations generated by TextMania are symbolic, human-interpretable, and easily controllable.", + "bbox": [ + 75, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach applies semantic perturbation on a different level to that of the label mix-based methods [86, 73, 85]. The mix-based methods augment a sample from a combination of two different class samples, i.e., applying semantic perturbation in an inter-class way. This further aggravates the class imbalance problem in the long-tailed (skewed) class distribution cases. Our TextManiA, whereas, perturbs data in an intra-class way. A sample per each class is selected, and we enrich the semantic granularity of the class using the sample, thus enabling us to better maintain the amount of augmentation balances in the long-tailed class distribution cases. Moreover, TextManiA can densify around the training samples by extrapolating the class semantics along augmented semantic attribute axes. With this, our method can be combined with the label mix-based methods to further improve performance in evenly distributed sparse data cases because they are complementary.", + "bbox": [ + 75, + 378, + 472, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To empirically support that our attribute vectors transformed from text embeddings are reasonably designed, we devise two visualization-based analyses: with t-SNE [72] and a latent inversion technique. These demonstrate that attribute vectors lead to visually interpretable manifold augmentation of input. We also evaluate our method with two different tasks in scarce data regimes: few-shot object detection and image classification with deficient datasets and long-tail datasets. Our experiments demonstrate that TextManiA is an effective and model-agnostic data augmentation method, especially in scarce data cases, by exploiting the favors of zero-shot attributes. Also, additional studies show the versatility and compatibility of the design of TextManiA. Our key contributions are summarized as:", + "bbox": [ + 75, + 637, + 472, + 849 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose TextMania, which enriches the visual features by conveying attribute information from the text embedding to the target visual feature space.", + "- We validate our hypothesis of the existence of embedded visual knowledge in pre-trained language encoders despite no training on visual data.", + "- We demonstrate that TextManiA is especially helpful in augmenting sparse samples in long-tail class cases.", + "- We show that our TextMania is complementary to other augmentation methods, and in particular, the combination of our TextMania and manifold Mixup [73] noticeably improves the performance in deficient data cases." + ], + "bbox": [ + 500, + 90, + 893, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 297, + 640, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We brief the related work in the following three perspectives: image data augmentation, foundation models, and target application tasks. In this work, our TextMania augments data by leveraging the text encoder of CLIP [54], BERT [17], or GPT-2² [55]. For main target applications, we focus on long-tail and small data classification and few-shot object detection tasks in the data-scarce regimes.", + "bbox": [ + 496, + 321, + 893, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Data Augmentation. Image data augmentation can be largely divided into whether semantic perturbation exists. Semantic perturbation, in specific, can be further split into methods with or without label mixing. Methods [18, 66, 57, 2, 24, 5, 30, 74, 44, 1, 41, 64, 13, 12] without semantic perturbation, which have no label change, contain primitive image processing and transformation operations. This includes photometric (e.g., color jitter, contrast, blur, noise, etc.) and geometric (e.g., horizontal reflection, rotation, etc.) operations, and advanced augmentations, including Cutout [18] and adaptive combinations [13, 12].", + "bbox": [ + 496, + 429, + 895, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, Mixup [86], CutMix [85], and manifold Mixup [73] execute semantic perturbation along with label mixing. Mixup interpolates two whole input images pixelwisely, CutMix interpolates a partial region of an image with another, and manifold Mixup mixes features from the images. These mix-based methods also augment labels of samples by an inter-class semantic perturbation, where labels of two different class samples are mixed. While the mixed label is known to be effective for generalization and model calibration effects [25], we found that the mix-based methods are heavily affected by class distribution due to sampling from two sources; thus, their effect is restricted to evenly distributed datasets. For datasets with skewed class distributions with tails, the sampling probabilities between major and minor classes would significantly differ, which can exaggerate biased sampling to major classes and makes minor classes more minor.", + "bbox": [ + 496, + 595, + 893, + 851 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "For example, if data size of major classes is 10 times larger than minor classes, the probability of choosing a pair of source samples from the major classes is approximately 100 times more than that of minor classes.", + "bbox": [ + 76, + 862, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$^{2}$ GTP-2 is the decoder-only architecture, but we use it as a text embedding extractor, so we call it text-encoder.", + "bbox": [ + 500, + 875, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2527", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our TextManiA, on the other hand, is applied to all of the given samples uniformly regardless of class distribution. TextManiA densifies around the sample features by perturbing and enriching the semantic meaning of them at an intra-class level, which does not change the label. Moreover, because of the different semantic granularity of perturbation between TextManiA (intra-class) and mix-based methods (inter-class), two methods can be used complementarily when class imbalance does not exist.", + "bbox": [ + 75, + 90, + 472, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Foundation Models. Recent foundation models [84, 54, 43, 33, 56, 17, 55] have shown a successful case of reflecting human nuances with visually imitated word composition. Particularly, language models, e.g., BERT [17] and GPT [55], show their ability not only in language tasks [78] but also in vision-language multi-modal tasks [62, 22]. Contrastive Language-Image Pretraining (CLIP) [54] also achieves huge success in various tasks even in zero-shot recognition. Follow-up studies show that CLIP representation is effective in conducting other visual tasks by bridging vision and language, e.g., 2D image generation [23, 39, 35, 49], image manipulation [51, 35] and synthesis [21], and even 3D domain tasks [83, 32, 47].", + "bbox": [ + 75, + 229, + 473, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In TextMania, we focus on estimating attribute features by exploiting BERT, GPT-2, or CLIP text encoder alone. Distinctively, we only transfer the estimated attribute feature to augment visual features in a different space, which makes our work different from knowledge distillation [29] of foundation models [15, 76, 63]. Rather, our design is an instance of the module neural network structure [27, 3], where recent module-based designs procedurally train the whole model module-by-module with the guidance of the well pretrained module, e.g., [38, 60, 50, 68, 26]. Also, our work is applicable agnostically to architectures; thus, more flexibly applicable than fine-tuning of foundation models [77].", + "bbox": [ + 75, + 428, + 473, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Long-tail Classification. In real world, visual data follow a long-tailed distribution which induces class imbalance and leads to performance degrading [81]. A representative line of the methods for long-tail classification is rebalancing [7, 14, 58], which resamples data or reweights the loss for tail classes. However, improvement in performance of the tail classes comes with the sacrifice of head class performance. Note that TextMania densifies all the given samples regardless of the class imbalance, and whereby the model is trained with reasonable variations of training samples for every class at least, which improves the performance while minimizing sacrifice of the head class.", + "bbox": [ + 75, + 611, + 473, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-Shot Object Detection (FSOD). We tackle FSOD, one of the sparse sample problems, to demonstrate the effectiveness of TextManIA and its model architecture agnostic property. FSOD handles novel object classes after the base training for object detection tasks. The model rapidly adapts to novel classes using few data by matching-based [42, 9, 79] or fine-tuning based [31, 67, 75, 53, 82]", + "bbox": [ + 75, + 794, + 473, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/14fc0c522639b723a5dcb4df77a05f91e74c4c4f7b25f1bc665e81cb2b58631b.jpg", + "image_caption": [ + "Figure 2. Overview of TextMania. Given flower image $\\mathbf{I}_0$ and class \"flower\" $T_0$ , we construct the variant text $T_{1}$ by adding the attribute \"red\" on $T_{0}$ . Embeddings of $T_{0}$ and $T_{1}$ are computed with text encoders, e.g. CLIP [54], BERT [17], or GPT-2 [55] and their difference vector, $\\Delta_{0\\rightarrow 1} = \\mathbf{e}_{T_1} - \\mathbf{e}_{T_0}$ is added to the image feature $\\mathbf{f}_{I_0}$ after projection $\\mathrm{proj}(\\cdot)$ and weight $\\alpha$ . We make the target feature space semantically rich and plausible by adding the difference vector, which embeds interpretable information." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 877, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "methods. TextManiA is evaluated with the fine-tuning-based FSOD approach [75], which facilitates to use general model architectures.", + "bbox": [ + 496, + 542, + 893, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. TextMania", + "text_level": 1, + "bbox": [ + 500, + 601, + 620, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In image classification, the class label is typically utilized only as a supervision for measuring the loss. We, instead, propose to treat the class label as additional information, the text describing the class, and derive semantic information from it. However, class label as a text description itself is too coarse to represent rich semantics within a class. For example, a class label \"dog\" does not represent all the details of the description such as \"small size of the brown colored dog.\" To enrich the detailed semantics over the given coarse class texts, we leverage the attribute words, such as \"small size\" and \"brown colored,\" that can visually modify objects in images at the semantic level.", + "bbox": [ + 496, + 626, + 893, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Main Idea", + "text_level": 1, + "bbox": [ + 500, + 816, + 617, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main idea of TextManiA is to densify distribution around sparse training samples on the target feature space, making it semantically rich through the difference vectors having plausible attribute information, as depicted in Fig. 1.", + "bbox": [ + 496, + 840, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2528", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 illustrates how TextMania augment data. Suppose we have an image $\\mathbf{I}_0$ and corresponding class label $T_0$ . The model generally learns the target task using the image $\\mathbf{I}_0$ as an input and the class label $T_0$ as supervision. In this work, we also consider the class label $T_0$ as text information and extract the embedding vector $\\mathbf{e}_{T_0} \\in \\mathbb{R}^{d_c}$ using text encoder, e.g., CLIP [54], BERT [17], or GPT-2 [55], where $d_c$ is the text embedding dimension. For obtaining an embedding vector $\\mathbf{e}_{T_0}$ , we use the text embedding of the encoder output directly when using CLIP text encoder, or use the average vector of all the embeddings of the sentence when using other language models such as BERT or GPT-2.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, text input $T_0$ is formed with class name and pre-defined prompts, such as \"a photo of,\" \"a picture of,\" and \"a sketch of.\" We also synthesize another text input variant $T_1$ by adding color or size attribute words, such as \"red\" and \"big,\" and compute the embedding vector $\\mathbf{e}_{T_1} \\in \\mathbb{R}^{d_c}$ . Numerous variants can be created with various attribute words and their combinations, but we explain the case of one variant for convenience. Based on the word vector analogy [48], we hypothesize that the relationship between $T_0$ and $T_1$ is maintained in the text embedding space, i.e., the difference vector $\\Delta_{0 \\to 1} = \\mathbf{e}_{T_1} - \\mathbf{e}_{T_0}$ would contain the information of added attributes (this hypothesis is validated in Sec. 3.2). To exploit the difference vector from text embeddings, we design our method on the manifold.", + "bbox": [ + 75, + 272, + 472, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can obtain such diverse attribute vectors from various attribute text templates; however, their representation space is not directly related to the visual feature space of the target model we are interested in. To bridge the gap, we project the difference vector to the target feature space with a learnable linear projection layer $\\operatorname{proj}(\\cdot)$ . Then, we add the projected difference vector to the target image feature $\\mathbf{f}_{I_0} \\in \\mathbb{R}^{d_t}$ obtained from the target task encoder with the input image $\\mathbf{I}_0$ , where $d_t$ is the target feature space dimension. A linear layer for $\\operatorname{proj}(\\cdot)$ would be sufficient to transfer cross-modal information, referring to the cross-modal transferability under the contrastive learning case [87] and our experiments.", + "bbox": [ + 75, + 484, + 472, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To inject the stochasticity, a mixing weight $\\alpha \\in \\mathbb{R}$ is introduced and randomly sampled from the clamped Normal distribution in the range over 0.1. Then, we have the augmented feature vector $\\hat{\\mathbf{f}}_{I_0}$ as,", + "bbox": [ + 75, + 665, + 472, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {f}} _ {I _ {0}} = \\mathbf {f} _ {I _ {0}} + \\alpha \\cdot \\operatorname {p r o j} \\left(\\boldsymbol {\\Delta} _ {0 \\rightarrow 1}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 734, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the cases having $d_{t} = d_{c}$ , we can set $\\operatorname{proj}(\\cdot)$ operation to be an identity mapping without any learnable parameter.", + "bbox": [ + 75, + 761, + 468, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train the target task model with this augmented feature vector, whose class label is still $T_{0}$ . We note that computing difference attribute vectors with text encoder is computationally expensive. For efficient training, we pre-compute all possible combinations of difference vectors $\\{\\Delta\\}$ and store", + "bbox": [ + 75, + 791, + 470, + 867 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bc60c8e6a2bf4e2e451789ef5f7ba5c648ecc92f97d504671678e4824e1ed920.jpg", + "image_caption": [ + "Label Mix" + ], + "image_footnote": [], + "bbox": [ + 501, + 104, + 584, + 160 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1c5a82e344d30d839bf097d51a75a558511fce77c4ed6f1cfe5b59b212c9d4bf.jpg", + "image_caption": [ + "TextMania" + ], + "image_footnote": [], + "bbox": [ + 612, + 90, + 733, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2e3e6bad334560fe657c439c8e3d91bcc6695576f1651cb24d5777fbcc9aa415.jpg", + "image_caption": [ + "TextManiA + Label Mix", + "Figure 3. Comparison of a typical label mix-based augmentation, TextManiA, and their combination, in a sparse data case. Given two samples for two classes, the augmented samples by a label mix-based method are at points lying on six lines resulting from the combinations of samples, i.e., inter-class perturbation. TextManiA densifies the sample distribution along semantic attribute axes in an intra-class. The combination of them yields synergy of their respective advantages." + ], + "image_footnote": [], + "bbox": [ + 754, + 89, + 887, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "them in a look-up table because class names and attributes can be pre-determined and unchanged during training.", + "bbox": [ + 496, + 354, + 890, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different from knowledge distillation [15, 76, 63, 29, 36], TextManiA does not transfer-learn the text embeddings directly. Instead, the difference vector projected onto the target domain is injected into the target model, allowing our method to be applied to arbitrary target models. Since the visual feature augmentation is solely controlled by text, TextManiA is human-interpretable and easily controllable.", + "bbox": [ + 496, + 383, + 892, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to label mix-based augmentations [86, 73, 85], TextManiA has advantages in imbalanced data distribution. We suppose a scenario where few samples are in one class and many samples are in another class. The augmented points by a mix-based method would be located only on the interpolation lines between the given samples, which limits the augmentation effects, as depicted in Fig. 3. If we apply a mix-based method in the long-tailed class distribution cases, i.e., notably skewed distribution, the class imbalance problem is further aggravated, and augmentation is more biased toward major classes. In contrast, TextManiA can equally densify all the given samples since it augments each sample independently. Thus, TextManiA can be used in general regardless of the imbalance factor of class distribution.", + "bbox": [ + 496, + 489, + 893, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the other hand, in another scenario with small training data but with uniform class distribution, both TextMania and mix-based methods would increase diverse combinations of samples by augmentation in respective aspects, which leads to complementary performance improvement. This will be empirically demonstrated in Sec. 4.", + "bbox": [ + 496, + 702, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Characteristics of Attribute Embedding", + "text_level": 1, + "bbox": [ + 500, + 801, + 841, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To scrutinize the relationship between the text $T_{0}$ , text variant $T_{1}$ and the attribute embedding $\\Delta_{0\\rightarrow 1}$ , we visualize their distribution and discuss the characteristics. We also visualize the difference vector to verify the hypothesis that the difference vector embeds its corresponding attribute.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3It was shown that simple algebraic operations can be performed on the word vectors, e.g., king - man + woman ≈ queen on the embedding space.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2529", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d7d4dbd599ac5c5f805e51e6bb53a90cd7c45c828819bbef8f9ea9ded9b15a26.jpg", + "image_caption": [ + "Figure 4. The t-SNE plot of difference vectors (e.g., \"brown dog\" - \"dog\") projected to visual feature space. The colors of the points represent color attributes used for computing the difference vector, and we use all the classes in CIFAR-100 for this plot. As a comparison, the colored points in the red circle show direct color-text embedding (e.g., \"brown\") projected to the visual feature space." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 468, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Embedding Difference vs. Direct Text Embedding. When guessing the difference between two texts, e.g., \"brown X\" - \"X,\" it would be \"brown.\" Someone may think of using the text embedding directly obtained from \"brown\" instead of our attribute embedding from \"brown X\" - \"X.\" To understand the difference between the two representations, we visualize the difference vectors and text embeddings with BERT and CLIP text encoder in Fig. 4. While the direct text embeddings in the red circle of Fig. 4 are clustered no matter with different color-texts, the difference vectors are well clustered dependent on the color. This observation indicates that the difference vector is more effective in augmenting the visual feature space than text embedding. In addition, the difference vectors obtained from the same attribute word are similarly clustered regardless of the class \"X\" but slightly different. It may imply our attribute embedding has subtle difference awareness on granularity according to class.", + "bbox": [ + 75, + 398, + 470, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that Fig. 4 presents difference vectors in the visual feature space, and we also observe similar distributions of difference vectors in the original text embedding space, i.e., t-SNE is not employed for clustering the direct text embeddings. This observation supports our hypothesis that general language models, e.g., BERT or GPT, have learned visual information to some extent. It, also, demonstrates the visual information is properly transferred to the target visual feature space.", + "bbox": [ + 75, + 657, + 470, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Do We Need to Rule out Unrealistic Attributes? One can be curious about how TextMania handles the unrealistic attribute, such as \"blue cow.\" We intentionally include such unrealistic attributes, motivated by other contexts in self- and semi-supervised learning [10, 65, 8], where they showed the strong benefit of unnatural strong augmentations to train neural networks. This observation regarding strong", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0f96afdd265fdda275adad38210782c4f070163ff9970d0b39bfea273e97afc8.jpg", + "image_caption": [ + "Figure 5. The attribute embedding visualization through image manipulation examples. We analyze how an image is manipulated when a difference vector, containing specific attribute information, is injected to the original image feature. (Top) We visualize an example of generated image given a specific class and its manipulated pair by size and color attribute, respectively. (Bottom) From left to right, we visualize the real image of \"Butterfly\" and its manipulated image pair with gigantic, green, and dot pattern, respectively." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 696, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0df78f704b5fea91b93d95831053246926a44e560afdf522bb7d6e932f8d1bc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 89, + 890, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "augmentation is consistent with the design of TextMania containing unrealistic attributes.", + "bbox": [ + 496, + 383, + 890, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Does Difference Vector Embed Attribute? To visually understand whether attribute editing is reflected while maintaining class information, we attempt to manipulate images by changing their features with the difference vectors $\\Delta_{0\\rightarrow 1} = \\mathbf{e}_{T_1} - \\mathbf{e}_{T_0}$ , i.e., we want to visualize the change effect between $\\mathbf{f}_{\\mathbf{I}_0}$ and $\\mathbf{e}_{a_1} = \\mathbf{f}_{\\mathbf{I}_0} + \\alpha \\Delta_{0\\rightarrow 1}$ in image domain. To see the effect in image domain, we need to invert the change from $\\mathbf{e}_{\\mathbf{I}_0}$ to $\\mathbf{e}_{a_1}$ in image domain, which can be formulated as the following optimization problem,", + "bbox": [ + 496, + 412, + 893, + 550 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\mathbf {I}} \\| E _ {i} (\\mathbf {I}) - \\mathbf {e} _ {a _ {1}} \\| _ {1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 561, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $E_{i}(\\cdot):\\mathbf{I}\\rightarrow \\mathbf{f}$ denotes the image encoder in Fig. 2. Direct optimization in Eq. (2) is known to be difficult [88]; thus, we parameterize a given image with an image generator $G_{\\theta}$ with a latent code $\\mathbf{z}$ , i.e., $\\mathbf{I}(\\theta) = G_{\\theta}(\\mathbf{z})$ , which is known to ease the optimization [71]. Then, we can obtain the visualization by the following optimization over $\\theta$", + "bbox": [ + 496, + 588, + 893, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\theta} \\| E _ {i} \\left(G _ {\\theta} (\\mathbf {z})\\right) - \\mathbf {e} _ {a _ {1}} \\| _ {1}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 691, + 890, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $E_{i}(\\cdot)$ , $\\mathbf{z}$ , and $\\mathbf{e}_{a_1}$ are frozen during the optimization. Since our goal is to see the move from $\\mathbf{f}_{\\mathbf{I}_0} = E_i(\\mathbf{I}_0)$ to $\\mathbf{e}_{a_1}$ for the query $\\mathbf{I}_0$ , we initialize $\\theta$ and $\\mathbf{z}$ such that $G_{\\theta}(\\mathbf{z}) \\simeq \\mathbf{I}_0$ by the GAN inversion [88]. We use IC-GAN [52] for the image generator and the text embeddings are obtained from the CLIP text encoders. Note that $E_{i}(\\cdot)$ is trained with random perturbation of transformed attributes; thus, the visualization through $E_{i}(\\cdot)$ is different from that of CLIP encoders. Details can be found in the supplementary material.", + "bbox": [ + 496, + 719, + 893, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 5 shows that the manipulated image reflects the added attribute, i.e., the size of the dog is reduced by the size attribute \"small,\" and the bird becomes yellow by injecting", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2530", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the color attribute \"yellow.\" The manipulated results imply that 1) the difference vector indeed embeds the attributes while preserving its semantics, and 2) our augmentation on the feature space may have analogous effects to an image-level augmentation but without implementing complicated image perturbation operations. Note that these visualizations are for analysis purposes but not for competing with any existing image manipulation methods.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 224, + 209, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate TextMania in various cases presenting sparse data with different tasks: long-tail classification in Sec. 4.1, evenly distributed scarce data classification in Sec. 4.2, and few-shot object detection in Sec. 4.3. We also conduct additional studies demonstrating the effectiveness of the design of our method and the versatility of TextMania in Sec. 4.4. Additional experimental results and details can be found in the supplementary material.", + "bbox": [ + 75, + 250, + 470, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Long-tail Classification", + "text_level": 1, + "bbox": [ + 76, + 378, + 290, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Setting. We compare TextMania with the mix-based augmentations on the CIFAR-100-LT [14] and ImageNet-LT [46] datasets, where LT stands for long-tailed distribution. They are artificially truncated to have a long-tail from each original dataset, CIFAR-100 [37] and ImageNet2012 [16]. Long-tail datasets usually have three sets of classes: Many-shot (more than 100 images), Medium-shot (20-100 images), and Few-shot (less than 20 images).", + "bbox": [ + 75, + 402, + 468, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For CIFAR-100-LT, we control the imbalance factor (IF) [11] computed as the ratio of samples in the head to tail class, $N_{1} / N_{K}$ , where $N_{k} = |\\mathcal{D}_{k}|$ , and $\\mathcal{D}_k$ is the set of samples belonging to the class $k\\in \\{1,\\dots ,K\\}$ . A larger value of the IF represents a more severe imbalance in data, which is more challenging. We evaluate the performance according to different IFs of 100, 50, and 10.", + "bbox": [ + 75, + 523, + 470, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilize ResNet18 as the baseline on CIFAR-100-LT and ResNext50 on ImageNet-LT. We use the validation set of the original datasets to measure the Top-1 accuracy. Note that we apply each augmentation on all the samples without carefully selecting a set of classes in Table 1.", + "bbox": [ + 75, + 628, + 468, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. Table 1 presents the long-tail classification results on CIFAR-100-LT, which show consistent improvement with TextManiA. Also, TextManiA with various text encoders achieves analogous improvement trend regardless of the imbalance factor but marginal degradation on Many class of IF=100 when using general language model, BERT and GPT-2. While the performance gain is from leaking pretrained language information, it is surprising and a virtue that the language models never exposed to any image can improve the visual recognition performance. In comparison to single usage of mix-based augmentations, our method shows higher accuracy because of uniform effects of TextManiA on samples regardless of class imbalance. The mix-based", + "bbox": [ + 75, + 704, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ea13a965ad005c108d25030166264b4c9a33351a2234946a63df2f874f2855c9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(a) AugmentationImbalance Factor (IF)
1005010
Baseline38.3943.3359.29
TextMania (CLIP)40.65 (+2.26)46.48 (+3.15)60.17 (+0.88)
TextMania (BERT)41.10 (+2.71)47.17 (+3.84)60.67 (+1.38)
TextMania (GPT-2)41.20 (+2.81)46.93 (+3.60)60.94 (+1.65)
Cutout [18]37.5142.2859.26
+ TextMania40.35 (+2.84)45.48 (+3.20)61.31 (+2.05)
Cutmix [85]37.9343.3459.30
+ TextMania40.22 (+2.29)45.36 (+2.02)61.30 (+2.00)
Mixup [86]36.7540.7757.50
+ TextMania38.40 (+1.65)43.33 (+2.56)59.80 (+2.30)
ManiMixup [73]35.7240.5155.26
+ TextMania38.60 (+2.88)43.22 (+2.71)59.35 (+4.09)
", + "bbox": [ + 504, + 89, + 888, + 292 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6523db0eef49408cc14f792da49abb4376f83d8848cfd01c380cd6973d2e0692.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(b) AugmentationSet of Classes (IF=100)
ManyMediumFew
Baseline71.1138.423.00
TextMania (CLIP)71.14 (+0.03)40.28 (+1.86)7.53 (+4.53)
TextMania (BERT)70.22 (-0.89)40.73 (+2.31)9.41 (+6.41)
TextMania (GPT-2)70.60 (-0.51)40.61 (+2.19)9.93 (+6.93)
Cutout71.5435.941.06
+ TextMania71.94 (+0.83)40.97 (+2.55)4.03 (+3.03)
Cutmix72.0237.170.90
+ TextMania72.37 (+0.35)40.80 (+3.63)3.90 (+3.00)
Mixup71.9733.620.36
+ TextMania71.97 (+0.00)36.77 (+3.15)1.83 (+1.47)
ManiMixup72.9729.510.70
+ TextMania73.20 (+0.23)36.80 (+7.29)0.76 (+0.06)
", + "bbox": [ + 503, + 295, + 888, + 496 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1dc19a112f3e178e1de1d029236b1ec79bbc31c02089cbcb973cf40038049d92.jpg", + "table_caption": [ + "Table 1. Long-tail classification results (\\%) on CIFAR-100-LT with ResNet18. (a) The accuracy with respect to the different imbalance factors, i.e., IF=\\{100, 50, 10\\}. (b) The accuracy of each class set with IF=100. Baseline contains random horizontal flip, random crop and rotation, and normalization, applied in all experiments. TextMania without parenthesis uses CLIP for the text encoder." + ], + "table_footnote": [], + "table_body": "
Aug.CBSAllManyMediumFew
Baseline38.3971.1138.423.00
Cutmix✓38.2371.7737.791.90
Mixup✓38.7371.6037.643.16
ManiMixup✓38.5671.2537.882.80
TextMania40.6571.1440.287.53
", + "bbox": [ + 500, + 599, + 888, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison to label mix-based augmentations with classbalanced sampling (CBS) on CIFAR-100-LT with IF=100. CBS samples two classes first and then samples data in each classes.", + "bbox": [ + 496, + 694, + 893, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods, on the other hand, sample two data points from the total dataset, where the probability that a tail class sample contributes to a resulting augmented sample is very low. Even with class-balanced sampling on mixed-based augmentation in Table 2, TextMania performs better, further demonstrating our effectiveness.", + "bbox": [ + 496, + 763, + 893, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Particularly in Table 1(b), the mix-based methods have degraded performance in the Medium and Few-shot classes, while our TextManiA improves performance. Combining", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2531", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6f4e5f996a15556d1e8002b8e5818c815b07aec4dfa054de184bd865de5b7ff9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AugmentationSet of ClassesTotal
ManyMediumFew
Baseline85.3470.4742.8072.24
TextMania85.4071.7548.4973.68
Cutout [18]85.0270.3242.9172.07
+ TextMania85.3371.7048.5473.65
Cutmix [85]84.8569.9035.8270.77
+ TextMania85.3071.9347.0473.52
Mixup [86]84.9670.2740.2071.63
+ TextMania84.5569.9535.7270.66
ManiMixup [73]84.8469.9738.8371.24
+ TextMania85.4271.9846.6573.54
", + "bbox": [ + 81, + 88, + 467, + 282 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a08592c8349aacf8bc7fc21c6bc5512feaad0ace90c9bba57762dba6676e944a.jpg", + "table_caption": [ + "Table 3. Long-tail classification results (\\%) on ImageNet-LT with ViT, and color the value as best and second best. Baseline contains random horizontal flip, random resize crop, color jitter, and normalization, applied in all experiments." + ], + "table_footnote": [], + "table_body": "
MethodManyMediumFewAll
LWS [34]63.3448.0827.1951.14
cRT [34]61.8046.2027.4049.60
cRT+TextManiA62.7448.6029.6751.47
", + "bbox": [ + 78, + 369, + 467, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the mix-based methods with TextMania improves overall performance, but the tendency to sacrifice the Medium and Few-shot classes is the same as before combining. Additionally, while Cutout has performance degradation due to the information loss [85], it is not affected by skewness due to no mix between inter-classes; thus, the performance is higher than the mix-based one in the long-tailed distribution.", + "bbox": [ + 75, + 527, + 468, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also evaluate our TextMania on the large-scale dataset ImageNet-LT. In Table 3, the best and second best results are with TextMania, which demonstrate that our augmentation method is also effective in the large-scale long-tailed data distribution, consistent with the CIFAR-100-LT results in Table 1. The improvement with TextMania implies the importance of intra-class perturbation, which can uniformly affect the samples regardless of the skewness of the class distribution.", + "bbox": [ + 75, + 638, + 470, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 4, we compare with LWS [34], cRT [34], and TextMania on cRT. LWS and cRT are one of effective methods in recent long-tailed recognition. The result shows that TextMania on cRT achieves the best results compared to the counterparts in all classes except for the Many class, wherefrom ours achieves second best. Overall, TextMania improves well-established works, e.g., LWS, and cRT, and it demonstrates the compatibility of our method.", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/09df1ff5ebbf78baf0c14a610326dc7f560aaedb2df8b76ac7aef207e907ff9f.jpg", + "table_caption": [ + "Table 4. Long-tail classification results $(\\%)$ on ImageNet-LT with ResNext50. We compare with LWS, cRT, and TextManiA on cRT, and color the value as best and second best." + ], + "table_footnote": [], + "table_body": "
AugmentationAcc.
Top-1Top-5
Baseline31.1059.14
Cutout32.0360.53
Cutmix32.4361.04
Mixup32.7262.47
ManiMixup33.7463.29
TextMania34.52 (+3.42)65.74 (+6.60)
Cutout + TextMania33.91 (+2.81)61.58 (+2.44)
Cutmix + TextMania35.61 (+4.51)63.82 (+4.68)
Mixup + TextMania37.97 (+6.87)66.75 (+7.61)
ManiMixup + TextMania38.02 (+6.92)67.28 (+8.14)
", + "bbox": [ + 504, + 88, + 890, + 281 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/02ca0ab3382dabb3c5bde3e5331c10a9e9de55fdda4a7708b827ed171db266b1.jpg", + "table_caption": [ + "Table 5. Classification results (\\%) on CIFAR-100-10% with ResNet18. Baseline represents random horizontal flip, random crop, and normalization, basically applied in all experiments. The parentheses stand for the improvement compared to the Baseline." + ], + "table_footnote": [], + "table_body": "
AugmentationAcc.
Top-1Top-5
Baseline65.3789.82
Cutout69.1791.12
Cutmix69.8291.76
Mixup67.5490.23
TextManiA70.81 (+5.44)92.37 (+2.55)
Cutout + TextManiA69.71 (+4.34)91.32 (+1.50)
Cutmix + TextManiA71.05 (+5.68)92.22 (+2.40)
Mixup + TextManiA70.56 (+5.19)91.58 (+1.76)
", + "bbox": [ + 514, + 352, + 879, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Classification results (\\%) on CIFAR-100-10% with VIT-Tiny. The configuration follows Table 5. The parentheses stands for the improvement compared to the Baseline.", + "bbox": [ + 496, + 525, + 893, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evenly Distributed Scarce Data Classification", + "text_level": 1, + "bbox": [ + 498, + 598, + 883, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental Setting. For evaluating the effectiveness of TextMania on the scarce dataset, we use $10\\%$ data of the CIFAR-100 [37] and Tiny/ImageNet [40] datasets, named CIFAR-100- $10\\%$ and Tiny/ImageNet- $10\\%$ , respectively. CIFAR-100 has 100 classes with 500 training images per class, but we only use randomly sampled 50 images per class. Tiny/ImageNet is a subset of ImageNet-1k [61] with 100k images and 200 classes, but we use 10k images (50 images per class) for simulating a small dataset. Note that the evaluation set is same with those of the original datasets.", + "bbox": [ + 496, + 625, + 893, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The baseline models of scarce data classification are ResNet18 [28] and ViT-Tiny [19]. Due to the space limit, details of training can be found in the supplementary material. Results. We demonstrate the effectiveness of TextMania compared to mix-based augmentations on evenly distributed scarce datasets. As in Table 5 for CIFAR-100-10%, TextMania outperforms other methods when a single augmentation is used. Furthermore, the effect is amplified when", + "bbox": [ + 496, + 777, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2532", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3e916db14f4bd6c462c0a202e802b337a5240544bb2a4a8e4db92c192c06011a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AugmentationAcc.
Top-1Top-5
Baseline25.9450.53
Cutout26.4150.28
Cutmix25.9449.67
Mixup29.3454.10
ManiMixup28.4353.25
TextMania29.37 (+3.43)52.37 (+1.84)
Cutout + TextMania29.14 (+3.20)52.60 (+2.07)
Cutmix + TextMania29.86 (+3.92)54.31 (+3.78)
Mixup + TextMania31.15 (+5.21)56.71 (+6.18)
ManiMixup + TextMania32.39 (+6.35)58.25 (+7.72)
", + "bbox": [ + 78, + 88, + 467, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our method and mix-based methods are combined, with particularly good compatibility with Manifold Mixup. The results demonstrate the importance of intra-class semantic perturbation along with inter-class in scarce data settings. This tendency is also observed with another baseline architecture in Table 6, and datasets in Table 7, implying that TextMania is model-agnostic to be applied. The overall results demonstrate the potential of TextMania to enrich the visual feature space using text modalities and develop more accurate and robust models in scarce data regimes.", + "bbox": [ + 75, + 354, + 472, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Few-shot Object Detection", + "text_level": 1, + "bbox": [ + 76, + 513, + 316, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experimental Setting. We evaluate TextMania on the PASCAL VOC [20] and MS-COCO [45] datasets with a few-shot division following Wang et al. [75]. For VOC, we have three random splits, which have different divisions into 15 base classes and 5 novel classes among the 20 total classes, and $K = 1, 2, 3, 5, 10$ objects are sampled from the novel classes. We utilize the VOC2007 test set for evaluation with AP50 metrics and train with the combination of the VOC2007 and VOC2012 train/val set. For COCO, the base classes are disjoint with VOC classes while the remaining classes are used as novel classes, and $K = 1, 3, 5, 10, 30$ objects are sampled from the novel classes for few-shot finetuning. We use 5k images from the validation set in COCO for evaluation with mAP metrics and the rest for training.", + "bbox": [ + 75, + 537, + 470, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The baseline [80] is the Faster R-CNN [59] trained with the base classes first and then fine-tuned with the novel classes. TextMania is applied to the novel class samples during the fine-tuning stage. Following the prior studies, all the reported results are averaged over 10 repeated runs.", + "bbox": [ + 75, + 750, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. Note that we apply TextManiA only on the classification head; thus, the quality of the regressed bounding boxes will remain similar as before applying TextManiA. As shown in Table 8 for VOC and Table 9 for COCO, TextManiA improves the AP by improving only the clas", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/edab6cafb3aaaf5a930a20d8172bff3ffff57c299dfbe58d2b2d63504afed035.jpg", + "table_caption": [ + "Table 7. Classification results on Tiny/ImageNet-10% with ResNet18. The configuration follows Table 5. The parentheses represent the improvement compared to the Baseline." + ], + "table_footnote": [], + "table_body": "
SplitAug.K-shot
123510
AllBaseline12.8216.6520.0420.6423.19
TextManiA17.74 (+4.92)22.40 (+5.75)23.37 (+3.33)25.09 (+4.45)24.22 (+1.03)
1Baseline15.1118.8222.6121.9723.74
TextManiA21.94 (+6.83)26.44 (+7.62)23.66 (+1.05)25.88 (+3.91)25.14 (+1.40)
2Baseline10.8614.2218.6719.3422.49
TextManiA14.64 (+3.78)18.49 (+4.27)23.28 (+4.61)23.06 (+3.72)24.44 (+1.95)
3Baseline12.4916.9018.8420.6123.35
TextManiA16.65 (+4.16)22.26 (+5.36)23.16 (+4.32)26.33 (+5.72)25.08 (+1.73)
", + "bbox": [ + 504, + 88, + 890, + 282 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0f1d2a46dc5edb00155f1168085d10c25ba4fa59920b57fb5ef7d2a7fd545807.jpg", + "table_caption": [ + "Table 8. Few-shot object detection results (AP50) on VOC. The value in the parentheses indicates the improvement compared to the Baseline of each split set." + ], + "table_footnote": [], + "table_body": "
Aug.K- shot
1351030
Baseline3.434.666.109.1112.78
TextMania5.39(+1.96)6.47(+1.81)7.80(+1.70)10.03(+0.92)13.60(+0.82)
", + "bbox": [ + 503, + 344, + 890, + 422 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1b9dbe97fb62e77f284fd5ebf733613db92332104d366d12232fc51ff1af9e5f.jpg", + "table_caption": [ + "Table 9. Few-shot object detection results (mAP) on COCO. The configuration follows Table 8." + ], + "table_footnote": [], + "table_body": "
Aug.ManyMediumFewIF=100IF=50IF=10
(a)Baseline71.1138.423.0038.3943.3359.29
Random71.3738.552.9038.4343.2860.39
TextManiA70.2240.739.4141.1047.1760.67
(b)Direct.71.3438.644.3238.6643.4459.82
Concat.68.0235.825.3536.9842.6859.44
", + "bbox": [ + 500, + 468, + 890, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10. Comparison to (a) random perturbation, and (b) direct text and concatenated embeddings on CIFAR-100-LT.", + "bbox": [ + 498, + 571, + 890, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sification accuracy, where the result is in a similar line to the analysis [6] that classification error weighs more than localization error. The improvement is clearer when $K$ is low. The results demonstrate the applicability of TextManiA to enhance the classification accuracy of detection models.", + "bbox": [ + 496, + 626, + 893, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Further Analyses", + "text_level": 1, + "bbox": [ + 500, + 710, + 669, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Random Baseline. In Table 10-(a), we compare our method with the Random baseline. We randomly sample a vector from a Normal distribution $\\mathcal{N}(0,1)$ and use it instead of the difference vector, i.e., augmenting visual features with random perturbations on the same manifold of visual features.", + "bbox": [ + 496, + 734, + 893, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The result shows that the Random baseline improves performance by serving as intra-perturb, but marginal compared to our method considering semantics additionally, which implies that semantic information embedded in the difference vector guides the augmentation more effective direction rather than random.", + "bbox": [ + 496, + 810, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2533", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5a10c70d03d264e0277408f8c88401b4b3e44ddc7edd4b39b3ae3238973047c6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CLIP Arch.Aug.ZSLP-10%LP-Full
ResNet50Baseline39.4750.1863.64
TextManiA-52.8364.17
ResNet101Baseline45.1757.3768.60
TextManiA-59.4969.12
ViT-BBaseline58.2173.3079.99
TextManiA-73.3579.58
", + "bbox": [ + 98, + 88, + 446, + 209 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Effectiveness of Difference Vectors. While we use the difference vectors by subtracting the embeddings with and without attribute words for TextMania, there could be another way to extract the attribute information. In Table 10(b), we compare with counterparts, direct text embedding (Direct.) and concatenated embeddings (Concat.). For the Direct method, we use the text embedding computed from the attribute word directly instead of the difference vector. For the Concat method, we concatenate the text embeddings from with and without attribute words, e.g., [“bull”||“red bull”], and use it instead of the difference vector.", + "bbox": [ + 75, + 296, + 470, + 462 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The results show that using difference vector (TextMania) outperforms using direct text embedding or concatenated embeddings, and imply that remaining contextual information after subtraction plays an important role in doing intra-perturbation in a semantic way. Although the word \"blue\" can function as both an adjective and a noun, its exact role in a sentence cannot be determined solely based on the word itself. Our intention of subtraction is for attribute words to act as a modifier in the sentence motivated by word analogy. When we computed the cosine similarity, embeddings derived directly from \"red\" and those obtained from the difference exhibited low similarity because they contain different contextual information despite the same origin of a word.", + "bbox": [ + 75, + 468, + 468, + 680 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Linear Probing with Advanced Models. Further demonstrating the compatibility of TextMania, we apply our method during linear probing of the model. In Table 11, we test VL-LTR [70], the state-of-the-art model in long-tail", + "bbox": [ + 76, + 686, + 254, + 821 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/83cbf6b36ea2d14683fae3c85b1d9df665eca467f280b46ee1212a1036c8ddd9.jpg", + "table_caption": [ + "Table 12. Classification results (\\%) of CLIP with zero-shot (ZS) and linear-probe (LP) on Full and $10\\%$ CIFAR-100. We apply our TextManiA to the linear-probed CLIP." + ], + "table_footnote": [], + "table_body": "
ModelLP-Full
VL-LTR61.04
+TextMania61.82
", + "bbox": [ + 282, + 691, + 452, + 752 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "classification, on CIFAR-100. In Table 12, we use a CLIP image encoder [54] with various architectures as the baseline model and linear-probe the model on both $10\\%$ and full data of CIFAR-100. The results demonstrate that TextMania is compatible with linear-probed CLIP and VL-LTR models.", + "bbox": [ + 75, + 821, + 468, + 897 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ablation Study on Attributes.", + "text_level": 1, + "bbox": [ + 500, + 90, + 709, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In TextManiA, we have considered color and size attributes. To confirm the effect of each attribute, we conduct an ablation study on attributes in Table 13. The result shows that while each attribute brings non-trivial gain, using both brings more gain. We", + "bbox": [ + 500, + 106, + 712, + 224 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/0881c7011910077db4156c338dd864183cc830784a7098e85f0517a4dd8b6580.jpg", + "table_caption": [ + "Table 11. Comparison between the SOTA model with and without TextMania during linear probing on CIFAR-100." + ], + "table_footnote": [], + "table_body": "
ColorSizeAcc.
31.10
✓33.48
✓33.89
✓✓34.52
", + "bbox": [ + 759, + 92, + 870, + 171 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 13. Ablation study on the attributes with CIFAR-100-10%.", + "bbox": [ + 741, + 176, + 890, + 217 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "believe that there are additional attributes we could use and a more effective method for selecting appropriate attributes, but leave it for future work.", + "bbox": [ + 498, + 227, + 890, + 271 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 309, + 617, + 324 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To mitigate the scarce data problem in long-tailed data distribution, small dataset, and few-shot cases, we propose a text-driven visual feature manifold augmentation method, TextManiA. Our method densifies around all the given individual visual features by adding a difference vector stem from the text embedding. While the mix-based augmentations inflict semantic perturbation in an inter-class way by label mixing, TextManiA perturbs the semantic meaning of the visual features at an intra-class level, i.e., having semantic perturbation while maintaining its class. The intraclass semantic perturbation is achieved by transferring the attribute-embedded vectors to visual feature space.", + "bbox": [ + 496, + 342, + 893, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To scrutinize the design of our estimated attribute embedding, we conduct visualization-based analyses: t-SNE plot and simple manipulation tests. The results empirically demonstrate that TextManiA readily enriches the sparse samples with comprehensible manipulation, since the general language models also reflect some extent of visual information. The experiment on the long-tail classification validates the effectiveness of our method, especially on the highly skewed class distribution. We additionally show the compatibility of TextManiA with other augmentation methods or other models in scarce data cases and during linear probing. In this work, note that we only use color and size as attributes; thus, there would be room for further investigation of other effective attributes.", + "bbox": [ + 496, + 531, + 893, + 743 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgment. This work was partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub; No.2022-0-00124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities; No. 2020-0-00004, Development of Provisional Intelligence based on Long-term Visual Memory Network).", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2534", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. *ICLR*, 2017. 2", + "[2] Guozhong An. The effects of adding noise during backpropagation training on a generalization performance. Neural computation, 8(3):643-674, 1996. 2", + "[3] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In CVPR, 2016. 3", + "[4] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151-175, 2010. 1", + "[5] Chris M Bishop. Training with noise is equivalent to tikhonov regularization. Neural computation, 7(1):108-116, 1995. 2", + "[6] Ali Borji and Seyed Mehdi Iranmanesh. Empirical upper bound in object detection and more. arXiv preprint arXiv:1911.12451, 2019. 8", + "[7] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural networks, 2018. 3", + "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 5", + "[9] Tung-I Chen, Yueh-Cheng Liu, Hung-Ting Su, Yu-Cheng Chang, Yu-Hsiang Lin, Jia-Fong Yeh, Wen-Chin Chen, and Winston Hsu. Dual-awareness attention for few-shot object detection. IEEE TMM, 2021. 3", + "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 5", + "[11] Peng Chu, Xiao Bian, Shaopeng Liu, and Haibin Ling. Feature space augmentation for long-tailed data. In ECCV, 2020. 6", + "[12] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. In CVPR, 2019. 2", + "[13] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPRW, 2020. 1, 2", + "[14] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 3, 6", + "[15] Wenliang Dai, Lu Hou, Lifeng Shang, Xin Jiang, Qun Liu, and Pascale Fung. Enabling multimodal generation on clip via vision-language knowledge distillation. ACL, 2022. 3, 4", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6", + "[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. ACL, 2018. 2, 3, 4", + "[18] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 1, 2, 6, 7", + "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 7", + "[20] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 8", + "[21] Kevin Frans, Lisa B Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. arXiv preprint arXiv:2106.14843, 2021. 3", + "[22] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. *ICLR*, 2023. 3", + "[23] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM TOG, 2022. 3", + "[24] Caglar Gulcehre, Marcin Moczulski, Misha Denil, and Yoshua Bengio. Noisy activation functions. In ICML, pages 3059-3068. PMLR, 2016. 2", + "[25] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, 2017. 1, 2", + "[26] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In CVPR, 2023. 3", + "[27] Bart LM Happel and Jacob MJ Murre. Design and evolution of modular neural network architectures. Neural networks, 7(6-7):985–1004, 1994. 3", + "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7", + "[29] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3, 4", + "[30] Lasse Holmstrom and Petri Koistinen. Using additive noise in back-propagation training. IEEE transactions on neural networks, 3(1):24-38, 1992. 2", + "[31] Hanzhe Hu, Shuai Bai, Aoxue Li, Jinshi Cui, and Liwei Wang. Dense relation distillation with context-aware aggregation for few-shot object detection. In CVPR, 2021. 3", + "[32] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, 2022. 3", + "[33] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3", + "[34] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In ICLR, 2020. 7", + "[35] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 3", + "[36] Youmin Kim, Jinbae Park, YounHo Jang, Muhammad Ali, Tae-Hyun Oh, and Sung-Ho Bae. Distilling global and local logits with densely connected relations. In ICCV, 2021. 4" + ], + "bbox": [ + 503, + 93, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2535", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, Citeseer, 2009. 6, 7", + "[38] Ashish Kumar, Zipeng Fu, Deepak Pathak, and Jitendra Malik. Rma: Rapid motor adaptation for legged robots. In Robotics: Science and Systems, 2021. 3", + "[39] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In CVPR, 2022. 3", + "[40] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 2015. 7", + "[41] Boyi Li, Felix Wu, Ser-Nam Lim, Serge Belongie, and Kilian Q Weinberger. On feature normalization and data augmentation. In CVPR, 2021. 1, 2", + "[42] Bohao Li, Boyu Yang, Chang Liu, Feng Liu, Rongrong Ji, and Qixiang Ye. Beyond max-margin: Class margin equilibrium for few-shot object detection. In CVPR, 2021. 3", + "[43] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 3", + "[44] Pan Li, Da Li, Wei Li, Shaogang Gong, Yanwei Fu, and Timothy M Hospedales. A simple feature augmentation for domain generalization. In ICCV, 2021. 1, 2", + "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 8", + "[46] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In CVPR, 2019. 6", + "[47] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In CVPR, 2022. 3", + "[48] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In ICLR, 2013. 4", + "[49] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 3", + "[50] Tae-Hyun Oh, Tali Dekel, Changil Kim, Inbar Mosseri, William T Freeman, Michael Rubinstein, and Wojciech Matusik. Speech2face: Learning the face behind a voice. In CVPR, 2019. 3", + "[51] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 3", + "[52] Guim Perarnau, Joost van de Weijer, Bogdan Raducanu, and Jose M. Álvarez. Invertible conditional gans for image editing. In CVPR, 2016. 5", + "[53] Limeng Qiao, Yuxuan Zhao, Zhiyuan Li, Xi Qiu, Jianan Wu, and Chi Zhang. Defrcn: Decoupled faster r-cnn for few-shot object detection. In ICCV, 2021. 3", + "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4, 9", + "[55] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2, 3, 4", + "[56] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3", + "[57] Russell Reed and Robert J MarksII. Neural smithing: supervised learning in feedforward artificial neural networks. Mit Press, 1999. 2", + "[58] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In ICML, 2018. 3", + "[59] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 8", + "[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3", + "[61] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 7", + "[62] Fawaz Sammani, Tanmoy Mukherjee, and Nikos Deligiannis. Nlx-gpt: A model for natural language explanations in vision and vision-language tasks. In CVPR, 2022. 3", + "[63] Gyungin Shin, Weidi Xie, and Samuel Albanie. Namedmask: Distilling segmenters from complementary foundation models. arXiv preprint arXiv:2209.11228, 2022. 3, 4", + "[64] Patrice Y Simard, Yann A LeCun, John S Denker, and Bernard Victorri. Transformation invariance in pattern recognition—tangent distance and tangent propagation. In Neural networks: tricks of the trade, pages 239–274. Springer, 1998. 1, 2", + "[65] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 5", + "[66] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 2014. 2", + "[67] Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. Fsce: Few-shot object detection via contrastive proposal encoding. In CVPR, 2021. 3", + "[68] Kim Sung-Bin, Arda Senocak, Hyunwoo Ha, Andrew Owens, and Tae-Hyun Oh. Sound to visual scene generation by audio-to-visual latent alignment. In CVPR, 2023. 3", + "[69] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. 1", + "[70] Changyao Tian, Wenhai Wang, Xizhou Zhu, Jifeng Dai, and Yu Qiao. Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition. In ECCV, 2022. 9" + ], + "bbox": [ + 503, + 92, + 893, + 897 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2536", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In CVPR, 2018. 5", + "[72] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 2", + "[73] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In ICML, 2019. 1, 2, 4, 6, 7", + "[74] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and León Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 1, 2", + "[75] Xin Wang, Thomas E Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. In ICML, 2020. 3, 8", + "[76] Zhecan Wang, Noel Codella, Yen-Chun Chen, Luowei Zhou, Xiyang Dai, Bin Xiao, Jianwei Yang, Haoxuan You, Kai-Wei Chang, Shih-fu Chang, et al. Multimodal adaptive distillation for leveraging unimodal encoders for vision-language tasks. arXiv preprint arXiv:2204.10496, 2022. 3, 4", + "[77] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In CVPR, 2022. 3", + "[78] Chien-Sheng Wu, Steven Hoi, Richard Socher, and Caiming Xiong. Tod-bert: Pre-trained natural language understanding for task-oriented dialogue. EMNLP, 2020. 3", + "[79] Yang Xiao and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. In ECCV, 2020. 3", + "[80] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In ICCV, 2019. 8", + "[81] Lu Yang, He Jiang, Qing Song, and Jun Guo. A survey on long-tailed visual recognition. IJCV, 2022. 3", + "[82] Moon Ye-Bin, Dongmin Choi, Yongjin Kwon, Junsik Kim, and Tae-Hyun Oh. Eninst: Enhancing weakly-supervised low-shot instance segmentation. arXiv preprint arXiv:2302.09765, 2023. 3", + "[83] Kim Youwang, Kim Ji-Yeon, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. ECCV, 2022. 3", + "[84] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 3", + "[85] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. 1, 2, 4, 6, 7", + "[86] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 1, 2, 4, 6, 7" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[87] Yuhui Zhang, Jeff Z HaoChen, Shih-Cheng Huang, Kuan-Chieh Wang, James Zou, and Serena Yeung. Diagnosing and rectifying vision models using language. In ICLR, 2023. 4", + "[88] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In ECCV, 2020. 5" + ], + "bbox": [ + 501, + 92, + 893, + 174 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2537", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/21a8728d-cd5e-44be-95fc-d79147ae763b_model.json b/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/21a8728d-cd5e-44be-95fc-d79147ae763b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..77ff569cc2cb587a272f8c4dc2d7f014dc53eac4 --- /dev/null +++ b/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/21a8728d-cd5e-44be-95fc-d79147ae763b_model.json @@ -0,0 +1,2853 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.131, + 0.877, + 0.153 + ], + "angle": 0, + "content": "TextMania: Enriching Visual Feature by Text-driven Manifold Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.18, + 0.234, + 0.198 + ], + "angle": 0, + "content": "Moon Ye-Bin" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.181, + 0.367, + 0.198 + ], + "angle": 0, + "content": "Jisoo Kim" + }, + { + "type": "text", + "bbox": [ + 0.406, + 0.181, + 0.54, + 0.2 + ], + "angle": 0, + "content": "Hongyeob Kim3" + }, + { + "type": "text", + "bbox": [ + 0.579, + 0.181, + 0.671, + 0.198 + ], + "angle": 0, + "content": "Kilho Son" + }, + { + "type": "text", + "bbox": [ + 0.711, + 0.181, + 0.85, + 0.199 + ], + "angle": 0, + "content": "Tae-Hyun \\(\\mathrm{Oh}^{1,5,6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.209, + 0.834, + 0.227 + ], + "angle": 0, + "content": "\\(^{1}\\)Dept. of Electrical Engineering and \\(^{5}\\)Grad. School of Artificial Intelligence, POSTECH" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.228, + 0.353, + 0.245 + ], + "angle": 0, + "content": "2Columbia University" + }, + { + "type": "text", + "bbox": [ + 0.393, + 0.228, + 0.612, + 0.246 + ], + "angle": 0, + "content": "3Sungkyunkwan University" + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.228, + 0.794, + 0.244 + ], + "angle": 0, + "content": "4 Microsoft Azure" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.245, + 0.862, + 0.264 + ], + "angle": 0, + "content": "\\(^{6}\\)Institute for Convergence Research and Education in Advanced Technology, Yonsei University" + }, + { + "type": "text", + "bbox": [ + 0.343, + 0.268, + 0.621, + 0.283 + ], + "angle": 0, + "content": "https://textmania.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.318, + 0.314, + 0.334 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.474, + 0.622 + ], + "angle": 0, + "content": "We propose TextManiA, a text-driven manifold augmentation method that semantically enriches visual feature spaces, regardless of class distribution. TextManiA augments visual data with intra-class semantic perturbation by exploiting easy-to-understand visually mimetic words, i.e., attributes. This work is built on an interesting hypothesis that general language models, e.g., BERT and GPT, encompass visual information to some extent, even without training on visual training data. Given the hypothesis, TextManiA transfers pre-trained text representation obtained from a well-established large language encoder to a target visual feature space being learned. Our extensive analysis hints that the language encoder indeed encompasses visual information at least useful to augment visual representation. Our experiments demonstrate that TextManiA is particularly powerful in scarce samples with class imbalance as well as even distribution. We also show compatibility with the label mix-based approaches in evenly distributed scarce data." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.649, + 0.21, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.87 + ], + "angle": 0, + "content": "Learning models, e.g., neural networks, are known to perform well on visual recognition tasks when training and testing datasets present similar distributions [4]. However, their performance often degrades considerably when evaluated in subtly different distributions [69]. One effective way to enhance the generalization ability of a model against such data distribution shifts would be data augmentation [18, 86, 85, 44, 41, 73]. Augmenting data enlarges the support of the training distribution formed by given samples and yields the effect of increasing the amount of data even without additional laborious data collection. By training on augmented data, decision boundaries are smoothed, and the generalization ability of the model is improved [73]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.901 + ], + "angle": 0, + "content": "There has been a distinctive and successful line of research for label mix-based data augmentation, such as" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.324, + 0.895, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.447, + 0.896, + 0.558 + ], + "angle": 0, + "content": "Figure 1. Illustration of TextMania. Our method augments the target visual feature by leveraging text embedding of the visually mimetic words, which are comprehensible and semantically rich. For example, when the text of the existing class \"bull\" is manipulated as \"red bull\" by adding the attribute \"red,\" we can get augmented visual features by reflecting the difference of text embeddings. In this way, TextMania densifies sparse visual feature space using various attributes text." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.897, + 0.852 + ], + "angle": 0, + "content": "Mixup [86], CutMix [85], and manifold Mixup [73], which are effective for model generalization and calibration [25]. The effectiveness of those label mix-based approaches is attributed to semantic perturbation by label mixing [86, 73, 85]. This is a distinctive property from other lines of data augmentation methods, e.g., [74, 44, 41, 64, 13], where they synthesize diverse virtual data that appear differently but retain class semantics of original contents. However, we found that the performance of mix-based augmentation methods is noticeably degraded when training with skewed class distribution having scarce samples for non-major classes, i.e., long-tailed distribution. In real-world, data often exhibit long-tailed class distribution (e.g., Pareto distribution), which cannot be dealt with the prevalent mix-based approaches. This motivates us to seek a semantically rich data augmentation effective for limited data regimes, including long-tailed distribution, scarce data, and few-shot cases." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose TextMania, a text-driven manifold augmentation for visual features, which is effective for long-tail classes and scarce data. Our TextMania is based" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2526" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.379 + ], + "angle": 0, + "content": "on an interesting hypothesis that general language models, e.g., BERT [17] and GPT [55], have learned visual information to some extent that can be transferred to visual feature spaces even with no visual training data. With this hypothesis, we semantically enrich the target visual feature space to be trained by leveraging visually mimetic texts, encoded with general language models and transferred to the target space. Specifically, TextMania encodes meaningful attributes such as \"red\" and \"large\" to vectors by computing the difference between text embeddings with and without attributes. We add the attribute embeddings to target visual features to mimic those attributes on the target visual feature space. Figure 1 illustrates the augmentation process of TextMania. The input feature (e.g., the visual feature of \"bull\") is manipulated by adding the attribute vector induced by the attribute text (e.g., \"red\"), which yields the augmented visual feature (e.g., \"red bull\"). Thanks to the text modality properties, the augmentations generated by TextMania are symbolic, human-interpretable, and easily controllable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.473, + 0.637 + ], + "angle": 0, + "content": "Our approach applies semantic perturbation on a different level to that of the label mix-based methods [86, 73, 85]. The mix-based methods augment a sample from a combination of two different class samples, i.e., applying semantic perturbation in an inter-class way. This further aggravates the class imbalance problem in the long-tailed (skewed) class distribution cases. Our TextManiA, whereas, perturbs data in an intra-class way. A sample per each class is selected, and we enrich the semantic granularity of the class using the sample, thus enabling us to better maintain the amount of augmentation balances in the long-tailed class distribution cases. Moreover, TextManiA can densify around the training samples by extrapolating the class semantics along augmented semantic attribute axes. With this, our method can be combined with the label mix-based methods to further improve performance in evenly distributed sparse data cases because they are complementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.473, + 0.851 + ], + "angle": 0, + "content": "To empirically support that our attribute vectors transformed from text embeddings are reasonably designed, we devise two visualization-based analyses: with t-SNE [72] and a latent inversion technique. These demonstrate that attribute vectors lead to visually interpretable manifold augmentation of input. We also evaluate our method with two different tasks in scarce data regimes: few-shot object detection and image classification with deficient datasets and long-tail datasets. Our experiments demonstrate that TextManiA is an effective and model-agnostic data augmentation method, especially in scarce data cases, by exploiting the favors of zero-shot attributes. Also, additional studies show the versatility and compatibility of the design of TextManiA. Our key contributions are summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.137 + ], + "angle": 0, + "content": "- We propose TextMania, which enriches the visual features by conveying attribute information from the text embedding to the target visual feature space." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.141, + 0.894, + 0.186 + ], + "angle": 0, + "content": "- We validate our hypothesis of the existence of embedded visual knowledge in pre-trained language encoders despite no training on visual data." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.189, + 0.892, + 0.22 + ], + "angle": 0, + "content": "- We demonstrate that TextManiA is especially helpful in augmenting sparse samples in long-tail class cases." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.224, + 0.895, + 0.284 + ], + "angle": 0, + "content": "- We show that our TextMania is complementary to other augmentation methods, and in particular, the combination of our TextMania and manifold Mixup [73] noticeably improves the performance in deficient data cases." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.298, + 0.642, + 0.314 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.895, + 0.429 + ], + "angle": 0, + "content": "We brief the related work in the following three perspectives: image data augmentation, foundation models, and target application tasks. In this work, our TextMania augments data by leveraging the text encoder of CLIP [54], BERT [17], or GPT-2² [55]. For main target applications, we focus on long-tail and small data classification and few-shot object detection tasks in the data-scarce regimes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.897, + 0.596 + ], + "angle": 0, + "content": "Image Data Augmentation. Image data augmentation can be largely divided into whether semantic perturbation exists. Semantic perturbation, in specific, can be further split into methods with or without label mixing. Methods [18, 66, 57, 2, 24, 5, 30, 74, 44, 1, 41, 64, 13, 12] without semantic perturbation, which have no label change, contain primitive image processing and transformation operations. This includes photometric (e.g., color jitter, contrast, blur, noise, etc.) and geometric (e.g., horizontal reflection, rotation, etc.) operations, and advanced augmentations, including Cutout [18] and adaptive combinations [13, 12]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.895, + 0.852 + ], + "angle": 0, + "content": "In contrast, Mixup [86], CutMix [85], and manifold Mixup [73] execute semantic perturbation along with label mixing. Mixup interpolates two whole input images pixelwisely, CutMix interpolates a partial region of an image with another, and manifold Mixup mixes features from the images. These mix-based methods also augment labels of samples by an inter-class semantic perturbation, where labels of two different class samples are mixed. While the mixed label is known to be effective for generalization and model calibration effects [25], we found that the mix-based methods are heavily affected by class distribution due to sampling from two sources; thus, their effect is restricted to evenly distributed datasets. For datasets with skewed class distributions with tails, the sampling probabilities between major and minor classes would significantly differ, which can exaggerate biased sampling to major classes and makes minor classes more minor." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.863, + 0.472, + 0.903 + ], + "angle": 0, + "content": "For example, if data size of major classes is 10 times larger than minor classes, the probability of choosing a pair of source samples from the major classes is approximately 100 times more than that of minor classes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.9 + ], + "angle": 0, + "content": "\\(^{2}\\)GTP-2 is the decoder-only architecture, but we use it as a text embedding extractor, so we call it text-encoder." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2527" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.226 + ], + "angle": 0, + "content": "Our TextManiA, on the other hand, is applied to all of the given samples uniformly regardless of class distribution. TextManiA densifies around the sample features by perturbing and enriching the semantic meaning of them at an intra-class level, which does not change the label. Moreover, because of the different semantic granularity of perturbation between TextManiA (intra-class) and mix-based methods (inter-class), two methods can be used complementarily when class imbalance does not exist." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.474, + 0.426 + ], + "angle": 0, + "content": "Foundation Models. Recent foundation models [84, 54, 43, 33, 56, 17, 55] have shown a successful case of reflecting human nuances with visually imitated word composition. Particularly, language models, e.g., BERT [17] and GPT [55], show their ability not only in language tasks [78] but also in vision-language multi-modal tasks [62, 22]. Contrastive Language-Image Pretraining (CLIP) [54] also achieves huge success in various tasks even in zero-shot recognition. Follow-up studies show that CLIP representation is effective in conducting other visual tasks by bridging vision and language, e.g., 2D image generation [23, 39, 35, 49], image manipulation [51, 35] and synthesis [21], and even 3D domain tasks [83, 32, 47]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.474, + 0.61 + ], + "angle": 0, + "content": "In TextMania, we focus on estimating attribute features by exploiting BERT, GPT-2, or CLIP text encoder alone. Distinctively, we only transfer the estimated attribute feature to augment visual features in a different space, which makes our work different from knowledge distillation [29] of foundation models [15, 76, 63]. Rather, our design is an instance of the module neural network structure [27, 3], where recent module-based designs procedurally train the whole model module-by-module with the guidance of the well pretrained module, e.g., [38, 60, 50, 68, 26]. Also, our work is applicable agnostically to architectures; thus, more flexibly applicable than fine-tuning of foundation models [77]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.474, + 0.793 + ], + "angle": 0, + "content": "Long-tail Classification. In real world, visual data follow a long-tailed distribution which induces class imbalance and leads to performance degrading [81]. A representative line of the methods for long-tail classification is rebalancing [7, 14, 58], which resamples data or reweights the loss for tail classes. However, improvement in performance of the tail classes comes with the sacrifice of head class performance. Note that TextMania densifies all the given samples regardless of the class imbalance, and whereby the model is trained with reasonable variations of training samples for every class at least, which improves the performance while minimizing sacrifice of the head class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.474, + 0.903 + ], + "angle": 0, + "content": "Few-Shot Object Detection (FSOD). We tackle FSOD, one of the sparse sample problems, to demonstrate the effectiveness of TextManIA and its model architecture agnostic property. FSOD handles novel object classes after the base training for object detection tasks. The model rapidly adapts to novel classes using few data by matching-based [42, 9, 79] or fine-tuning based [31, 67, 75, 53, 82]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.878, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.406, + 0.895, + 0.518 + ], + "angle": 0, + "content": "Figure 2. Overview of TextMania. Given flower image \\(\\mathbf{I}_0\\) and class \"flower\" \\(T_0\\), we construct the variant text \\(T_{1}\\) by adding the attribute \"red\" on \\(T_{0}\\). Embeddings of \\(T_{0}\\) and \\(T_{1}\\) are computed with text encoders, e.g. CLIP [54], BERT [17], or GPT-2 [55] and their difference vector, \\(\\Delta_{0\\rightarrow 1} = \\mathbf{e}_{T_1} - \\mathbf{e}_{T_0}\\) is added to the image feature \\(\\mathbf{f}_{I_0}\\) after projection \\(\\mathrm{proj}(\\cdot)\\) and weight \\(\\alpha\\). We make the target feature space semantically rich and plausible by adding the difference vector, which embeds interpretable information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.895, + 0.588 + ], + "angle": 0, + "content": "methods. TextManiA is evaluated with the fine-tuning-based FSOD approach [75], which facilitates to use general model architectures." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.622, + 0.617 + ], + "angle": 0, + "content": "3. TextMania" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.895, + 0.809 + ], + "angle": 0, + "content": "In image classification, the class label is typically utilized only as a supervision for measuring the loss. We, instead, propose to treat the class label as additional information, the text describing the class, and derive semantic information from it. However, class label as a text description itself is too coarse to represent rich semantics within a class. For example, a class label \"dog\" does not represent all the details of the description such as \"small size of the brown colored dog.\" To enrich the detailed semantics over the given coarse class texts, we leverage the attribute words, such as \"small size\" and \"brown colored,\" that can visually modify objects in images at the semantic level." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.618, + 0.832 + ], + "angle": 0, + "content": "3.1. Main Idea" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.897, + 0.903 + ], + "angle": 0, + "content": "The main idea of TextManiA is to densify distribution around sparse training samples on the target feature space, making it semantically rich through the difference vectors having plausible attribute information, as depicted in Fig. 1." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2528" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "Figure 2 illustrates how TextMania augment data. Suppose we have an image \\(\\mathbf{I}_0\\) and corresponding class label \\(T_0\\). The model generally learns the target task using the image \\(\\mathbf{I}_0\\) as an input and the class label \\(T_0\\) as supervision. In this work, we also consider the class label \\(T_0\\) as text information and extract the embedding vector \\(\\mathbf{e}_{T_0} \\in \\mathbb{R}^{d_c}\\) using text encoder, e.g., CLIP [54], BERT [17], or GPT-2 [55], where \\(d_c\\) is the text embedding dimension. For obtaining an embedding vector \\(\\mathbf{e}_{T_0}\\), we use the text embedding of the encoder output directly when using CLIP text encoder, or use the average vector of all the embeddings of the sentence when using other language models such as BERT or GPT-2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.485 + ], + "angle": 0, + "content": "Specifically, text input \\( T_0 \\) is formed with class name and pre-defined prompts, such as \"a photo of,\" \"a picture of,\" and \"a sketch of.\" We also synthesize another text input variant \\( T_1 \\) by adding color or size attribute words, such as \"red\" and \"big,\" and compute the embedding vector \\( \\mathbf{e}_{T_1} \\in \\mathbb{R}^{d_c} \\). Numerous variants can be created with various attribute words and their combinations, but we explain the case of one variant for convenience. Based on the word vector analogy [48], we hypothesize that the relationship between \\( T_0 \\) and \\( T_1 \\) is maintained in the text embedding space, i.e., the difference vector \\( \\Delta_{0 \\to 1} = \\mathbf{e}_{T_1} - \\mathbf{e}_{T_0} \\) would contain the information of added attributes (this hypothesis is validated in Sec. 3.2). To exploit the difference vector from text embeddings, we design our method on the manifold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.473, + 0.666 + ], + "angle": 0, + "content": "We can obtain such diverse attribute vectors from various attribute text templates; however, their representation space is not directly related to the visual feature space of the target model we are interested in. To bridge the gap, we project the difference vector to the target feature space with a learnable linear projection layer \\(\\operatorname{proj}(\\cdot)\\). Then, we add the projected difference vector to the target image feature \\(\\mathbf{f}_{I_0} \\in \\mathbb{R}^{d_t}\\) obtained from the target task encoder with the input image \\(\\mathbf{I}_0\\), where \\(d_t\\) is the target feature space dimension. A linear layer for \\(\\operatorname{proj}(\\cdot)\\) would be sufficient to transfer cross-modal information, referring to the cross-modal transferability under the contrastive learning case [87] and our experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.473, + 0.726 + ], + "angle": 0, + "content": "To inject the stochasticity, a mixing weight \\(\\alpha \\in \\mathbb{R}\\) is introduced and randomly sampled from the clamped Normal distribution in the range over 0.1. Then, we have the augmented feature vector \\(\\hat{\\mathbf{f}}_{I_0}\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.735, + 0.47, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {f}} _ {I _ {0}} = \\mathbf {f} _ {I _ {0}} + \\alpha \\cdot \\operatorname {p r o j} \\left(\\boldsymbol {\\Delta} _ {0 \\rightarrow 1}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.47, + 0.792 + ], + "angle": 0, + "content": "For the cases having \\( d_{t} = d_{c} \\), we can set \\( \\operatorname{proj}(\\cdot) \\) operation to be an identity mapping without any learnable parameter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.471, + 0.868 + ], + "angle": 0, + "content": "We train the target task model with this augmented feature vector, whose class label is still \\( T_{0} \\). We note that computing difference attribute vectors with text encoder is computationally expensive. For efficient training, we pre-compute all possible combinations of difference vectors \\( \\{\\Delta\\} \\) and store" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.105, + 0.585, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.179, + 0.572, + 0.19 + ], + "angle": 0, + "content": "Label Mix" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.092, + 0.735, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.179, + 0.703, + 0.19 + ], + "angle": 0, + "content": "TextMania" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.09, + 0.888, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.179, + 0.89, + 0.19 + ], + "angle": 0, + "content": "TextManiA + Label Mix" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.218, + 0.895, + 0.33 + ], + "angle": 0, + "content": "Figure 3. Comparison of a typical label mix-based augmentation, TextManiA, and their combination, in a sparse data case. Given two samples for two classes, the augmented samples by a label mix-based method are at points lying on six lines resulting from the combinations of samples, i.e., inter-class perturbation. TextManiA densifies the sample distribution along semantic attribute axes in an intra-class. The combination of them yields synergy of their respective advantages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.385 + ], + "angle": 0, + "content": "them in a look-up table because class names and attributes can be pre-determined and unchanged during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.893, + 0.491 + ], + "angle": 0, + "content": "Different from knowledge distillation [15, 76, 63, 29, 36], TextManiA does not transfer-learn the text embeddings directly. Instead, the difference vector projected onto the target domain is injected into the target model, allowing our method to be applied to arbitrary target models. Since the visual feature augmentation is solely controlled by text, TextManiA is human-interpretable and easily controllable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.895, + 0.702 + ], + "angle": 0, + "content": "Compared to label mix-based augmentations [86, 73, 85], TextManiA has advantages in imbalanced data distribution. We suppose a scenario where few samples are in one class and many samples are in another class. The augmented points by a mix-based method would be located only on the interpolation lines between the given samples, which limits the augmentation effects, as depicted in Fig. 3. If we apply a mix-based method in the long-tailed class distribution cases, i.e., notably skewed distribution, the class imbalance problem is further aggravated, and augmentation is more biased toward major classes. In contrast, TextManiA can equally densify all the given samples since it augments each sample independently. Thus, TextManiA can be used in general regardless of the imbalance factor of class distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.793 + ], + "angle": 0, + "content": "On the other hand, in another scenario with small training data but with uniform class distribution, both TextMania and mix-based methods would increase diverse combinations of samples by augmentation in respective aspects, which leads to complementary performance improvement. This will be empirically demonstrated in Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.843, + 0.818 + ], + "angle": 0, + "content": "3.2. Characteristics of Attribute Embedding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To scrutinize the relationship between the text \\( T_{0} \\), text variant \\( T_{1} \\) and the attribute embedding \\( \\Delta_{0\\rightarrow 1} \\), we visualize their distribution and discuss the characteristics. We also visualize the difference vector to verify the hypothesis that the difference vector embeds its corresponding attribute." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "3It was shown that simple algebraic operations can be performed on the word vectors, e.g., king - man + woman ≈ queen on the embedding space." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2529" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.469, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.286, + 0.472, + 0.371 + ], + "angle": 0, + "content": "Figure 4. The t-SNE plot of difference vectors (e.g., \"brown dog\" - \"dog\") projected to visual feature space. The colors of the points represent color attributes used for computing the difference vector, and we use all the classes in CIFAR-100 for this plot. As a comparison, the colored points in the red circle show direct color-text embedding (e.g., \"brown\") projected to the visual feature space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.472, + 0.656 + ], + "angle": 0, + "content": "Embedding Difference vs. Direct Text Embedding. When guessing the difference between two texts, e.g., \"brown X\" - \"X,\" it would be \"brown.\" Someone may think of using the text embedding directly obtained from \"brown\" instead of our attribute embedding from \"brown X\" - \"X.\" To understand the difference between the two representations, we visualize the difference vectors and text embeddings with BERT and CLIP text encoder in Fig. 4. While the direct text embeddings in the red circle of Fig. 4 are clustered no matter with different color-texts, the difference vectors are well clustered dependent on the color. This observation indicates that the difference vector is more effective in augmenting the visual feature space than text embedding. In addition, the difference vectors obtained from the same attribute word are similarly clustered regardless of the class \"X\" but slightly different. It may imply our attribute embedding has subtle difference awareness on granularity according to class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.472, + 0.793 + ], + "angle": 0, + "content": "Note that Fig. 4 presents difference vectors in the visual feature space, and we also observe similar distributions of difference vectors in the original text embedding space, i.e., t-SNE is not employed for clustering the direct text embeddings. This observation supports our hypothesis that general language models, e.g., BERT or GPT, have learned visual information to some extent. It, also, demonstrates the visual information is properly transferred to the target visual feature space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Do We Need to Rule out Unrealistic Attributes? One can be curious about how TextMania handles the unrealistic attribute, such as \"blue cow.\" We intentionally include such unrealistic attributes, motivated by other contexts in self- and semi-supervised learning [10, 65, 8], where they showed the strong benefit of unnatural strong augmentations to train neural networks. This observation regarding strong" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.697, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.09, + 0.892, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.246, + 0.895, + 0.358 + ], + "angle": 0, + "content": "Figure 5. The attribute embedding visualization through image manipulation examples. We analyze how an image is manipulated when a difference vector, containing specific attribute information, is injected to the original image feature. (Top) We visualize an example of generated image given a specific class and its manipulated pair by size and color attribute, respectively. (Bottom) From left to right, we visualize the real image of \"Butterfly\" and its manipulated image pair with gigantic, green, and dot pattern, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.413 + ], + "angle": 0, + "content": "augmentation is consistent with the design of TextMania containing unrealistic attributes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.895, + 0.551 + ], + "angle": 0, + "content": "Does Difference Vector Embed Attribute? To visually understand whether attribute editing is reflected while maintaining class information, we attempt to manipulate images by changing their features with the difference vectors \\(\\Delta_{0\\rightarrow 1} = \\mathbf{e}_{T_1} - \\mathbf{e}_{T_0}\\), i.e., we want to visualize the change effect between \\(\\mathbf{f}_{\\mathbf{I}_0}\\) and \\(\\mathbf{e}_{a_1} = \\mathbf{f}_{\\mathbf{I}_0} + \\alpha \\Delta_{0\\rightarrow 1}\\) in image domain. To see the effect in image domain, we need to invert the change from \\(\\mathbf{e}_{\\mathbf{I}_0}\\) to \\(\\mathbf{e}_{a_1}\\) in image domain, which can be formulated as the following optimization problem," + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.562, + 0.892, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\mathbf {I}} \\| E _ {i} (\\mathbf {I}) - \\mathbf {e} _ {a _ {1}} \\| _ {1}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.895, + 0.68 + ], + "angle": 0, + "content": "where \\(E_{i}(\\cdot):\\mathbf{I}\\rightarrow \\mathbf{f}\\) denotes the image encoder in Fig. 2. Direct optimization in Eq. (2) is known to be difficult [88]; thus, we parameterize a given image with an image generator \\(G_{\\theta}\\) with a latent code \\(\\mathbf{z}\\), i.e., \\(\\mathbf{I}(\\theta) = G_{\\theta}(\\mathbf{z})\\), which is known to ease the optimization [71]. Then, we can obtain the visualization by the following optimization over \\(\\theta\\)" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.692, + 0.892, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\theta} \\| E _ {i} \\left(G _ {\\theta} (\\mathbf {z})\\right) - \\mathbf {e} _ {a _ {1}} \\| _ {1}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.855 + ], + "angle": 0, + "content": "where \\( E_{i}(\\cdot) \\), \\( \\mathbf{z} \\), and \\( \\mathbf{e}_{a_1} \\) are frozen during the optimization. Since our goal is to see the move from \\( \\mathbf{f}_{\\mathbf{I}_0} = E_i(\\mathbf{I}_0) \\) to \\( \\mathbf{e}_{a_1} \\) for the query \\( \\mathbf{I}_0 \\), we initialize \\( \\theta \\) and \\( \\mathbf{z} \\) such that \\( G_{\\theta}(\\mathbf{z}) \\simeq \\mathbf{I}_0 \\) by the GAN inversion [88]. We use IC-GAN [52] for the image generator and the text embeddings are obtained from the CLIP text encoders. Note that \\( E_{i}(\\cdot) \\) is trained with random perturbation of transformed attributes; thus, the visualization through \\( E_{i}(\\cdot) \\) is different from that of CLIP encoders. Details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Figure 5 shows that the manipulated image reflects the added attribute, i.e., the size of the dog is reduced by the size attribute \"small,\" and the bird becomes yellow by injecting" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2530" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "the color attribute \"yellow.\" The manipulated results imply that 1) the difference vector indeed embeds the attributes while preserving its semantics, and 2) our augmentation on the feature space may have analogous effects to an image-level augmentation but without implementing complicated image perturbation operations. Note that these visualizations are for analysis purposes but not for competing with any existing image manipulation methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.225, + 0.21, + 0.243 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.251, + 0.471, + 0.372 + ], + "angle": 0, + "content": "We evaluate TextMania in various cases presenting sparse data with different tasks: long-tail classification in Sec. 4.1, evenly distributed scarce data classification in Sec. 4.2, and few-shot object detection in Sec. 4.3. We also conduct additional studies demonstrating the effectiveness of the design of our method and the versatility of TextMania in Sec. 4.4. Additional experimental results and details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.38, + 0.292, + 0.396 + ], + "angle": 0, + "content": "4.1. Long-tail Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.524 + ], + "angle": 0, + "content": "Experimental Setting. We compare TextMania with the mix-based augmentations on the CIFAR-100-LT [14] and ImageNet-LT [46] datasets, where LT stands for long-tailed distribution. They are artificially truncated to have a long-tail from each original dataset, CIFAR-100 [37] and ImageNet2012 [16]. Long-tail datasets usually have three sets of classes: Many-shot (more than 100 images), Medium-shot (20-100 images), and Few-shot (less than 20 images)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.471, + 0.629 + ], + "angle": 0, + "content": "For CIFAR-100-LT, we control the imbalance factor (IF) [11] computed as the ratio of samples in the head to tail class, \\( N_{1} / N_{K} \\), where \\( N_{k} = |\\mathcal{D}_{k}| \\), and \\( \\mathcal{D}_k \\) is the set of samples belonging to the class \\( k\\in \\{1,\\dots ,K\\} \\). A larger value of the IF represents a more severe imbalance in data, which is more challenging. We evaluate the performance according to different IFs of 100, 50, and 10." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.704 + ], + "angle": 0, + "content": "We utilize ResNet18 as the baseline on CIFAR-100-LT and ResNext50 on ImageNet-LT. We use the validation set of the original datasets to measure the Top-1 accuracy. Note that we apply each augmentation on all the samples without carefully selecting a set of classes in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Results. Table 1 presents the long-tail classification results on CIFAR-100-LT, which show consistent improvement with TextManiA. Also, TextManiA with various text encoders achieves analogous improvement trend regardless of the imbalance factor but marginal degradation on Many class of IF=100 when using general language model, BERT and GPT-2. While the performance gain is from leaking pretrained language information, it is surprising and a virtue that the language models never exposed to any image can improve the visual recognition performance. In comparison to single usage of mix-based augmentations, our method shows higher accuracy because of uniform effects of TextManiA on samples regardless of class imbalance. The mix-based" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.89, + 0.294 + ], + "angle": 0, + "content": "
(a) AugmentationImbalance Factor (IF)
1005010
Baseline38.3943.3359.29
TextMania (CLIP)40.65 (+2.26)46.48 (+3.15)60.17 (+0.88)
TextMania (BERT)41.10 (+2.71)47.17 (+3.84)60.67 (+1.38)
TextMania (GPT-2)41.20 (+2.81)46.93 (+3.60)60.94 (+1.65)
Cutout [18]37.5142.2859.26
+ TextMania40.35 (+2.84)45.48 (+3.20)61.31 (+2.05)
Cutmix [85]37.9343.3459.30
+ TextMania40.22 (+2.29)45.36 (+2.02)61.30 (+2.00)
Mixup [86]36.7540.7757.50
+ TextMania38.40 (+1.65)43.33 (+2.56)59.80 (+2.30)
ManiMixup [73]35.7240.5155.26
+ TextMania38.60 (+2.88)43.22 (+2.71)59.35 (+4.09)
" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.296, + 0.89, + 0.497 + ], + "angle": 0, + "content": "
(b) AugmentationSet of Classes (IF=100)
ManyMediumFew
Baseline71.1138.423.00
TextMania (CLIP)71.14 (+0.03)40.28 (+1.86)7.53 (+4.53)
TextMania (BERT)70.22 (-0.89)40.73 (+2.31)9.41 (+6.41)
TextMania (GPT-2)70.60 (-0.51)40.61 (+2.19)9.93 (+6.93)
Cutout71.5435.941.06
+ TextMania71.94 (+0.83)40.97 (+2.55)4.03 (+3.03)
Cutmix72.0237.170.90
+ TextMania72.37 (+0.35)40.80 (+3.63)3.90 (+3.00)
Mixup71.9733.620.36
+ TextMania71.97 (+0.00)36.77 (+3.15)1.83 (+1.47)
ManiMixup72.9729.510.70
+ TextMania73.20 (+0.23)36.80 (+7.29)0.76 (+0.06)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.502, + 0.894, + 0.587 + ], + "angle": 0, + "content": "Table 1. Long-tail classification results (\\%) on CIFAR-100-LT with ResNet18. (a) The accuracy with respect to the different imbalance factors, i.e., IF=\\{100, 50, 10\\}. (b) The accuracy of each class set with IF=100. Baseline contains random horizontal flip, random crop and rotation, and normalization, applied in all experiments. TextMania without parenthesis uses CLIP for the text encoder." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.6, + 0.89, + 0.691 + ], + "angle": 0, + "content": "
Aug.CBSAllManyMediumFew
Baseline38.3971.1138.423.00
Cutmix✓38.2371.7737.791.90
Mixup✓38.7371.6037.643.16
ManiMixup✓38.5671.2537.882.80
TextMania40.6571.1440.287.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.695, + 0.894, + 0.738 + ], + "angle": 0, + "content": "Table 2. Comparison to label mix-based augmentations with classbalanced sampling (CBS) on CIFAR-100-LT with IF=100. CBS samples two classes first and then samples data in each classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.854 + ], + "angle": 0, + "content": "methods, on the other hand, sample two data points from the total dataset, where the probability that a tail class sample contributes to a resulting augmented sample is very low. Even with class-balanced sampling on mixed-based augmentation in Table 2, TextMania performs better, further demonstrating our effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Particularly in Table 1(b), the mix-based methods have degraded performance in the Medium and Few-shot classes, while our TextManiA improves performance. Combining" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2531" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.284 + ], + "angle": 0, + "content": "
AugmentationSet of ClassesTotal
ManyMediumFew
Baseline85.3470.4742.8072.24
TextMania85.4071.7548.4973.68
Cutout [18]85.0270.3242.9172.07
+ TextMania85.3371.7048.5473.65
Cutmix [85]84.8569.9035.8270.77
+ TextMania85.3071.9347.0473.52
Mixup [86]84.9670.2740.2071.63
+ TextMania84.5569.9535.7270.66
ManiMixup [73]84.8469.9738.8371.24
+ TextMania85.4271.9846.6573.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.289, + 0.472, + 0.348 + ], + "angle": 0, + "content": "Table 3. Long-tail classification results (\\%) on ImageNet-LT with ViT, and color the value as best and second best. Baseline contains random horizontal flip, random resize crop, color jitter, and normalization, applied in all experiments." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.37, + 0.468, + 0.434 + ], + "angle": 0, + "content": "
MethodManyMediumFewAll
LWS [34]63.3448.0827.1951.14
cRT [34]61.8046.2027.4049.60
cRT+TextManiA62.7448.6029.6751.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.438, + 0.47, + 0.482 + ], + "angle": 0, + "content": "Table 4. Long-tail classification results \\((\\%)\\) on ImageNet-LT with ResNext50. We compare with LWS, cRT, and TextManiA on cRT, and color the value as best and second best." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.47, + 0.635 + ], + "angle": 0, + "content": "the mix-based methods with TextMania improves overall performance, but the tendency to sacrifice the Medium and Few-shot classes is the same as before combining. Additionally, while Cutout has performance degradation due to the information loss [85], it is not affected by skewness due to no mix between inter-classes; thus, the performance is higher than the mix-based one in the long-tailed distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.471, + 0.775 + ], + "angle": 0, + "content": "We also evaluate our TextMania on the large-scale dataset ImageNet-LT. In Table 3, the best and second best results are with TextMania, which demonstrate that our augmentation method is also effective in the large-scale long-tailed data distribution, consistent with the CIFAR-100-LT results in Table 1. The improvement with TextMania implies the importance of intra-class perturbation, which can uniformly affect the samples regardless of the skewness of the class distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In Table 4, we compare with LWS [34], cRT [34], and TextMania on cRT. LWS and cRT are one of effective methods in recent long-tailed recognition. The result shows that TextMania on cRT achieves the best results compared to the counterparts in all classes except for the Many class, wherefrom ours achieves second best. Overall, TextMania improves well-established works, e.g., LWS, and cRT, and it demonstrates the compatibility of our method." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.891, + 0.282 + ], + "angle": 0, + "content": "
AugmentationAcc.
Top-1Top-5
Baseline31.1059.14
Cutout32.0360.53
Cutmix32.4361.04
Mixup32.7262.47
ManiMixup33.7463.29
TextMania34.52 (+3.42)65.74 (+6.60)
Cutout + TextMania33.91 (+2.81)61.58 (+2.44)
Cutmix + TextMania35.61 (+4.51)63.82 (+4.68)
Mixup + TextMania37.97 (+6.87)66.75 (+7.61)
ManiMixup + TextMania38.02 (+6.92)67.28 (+8.14)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.286, + 0.893, + 0.344 + ], + "angle": 0, + "content": "Table 5. Classification results (\\%) on CIFAR-100-10% with ResNet18. Baseline represents random horizontal flip, random crop, and normalization, basically applied in all experiments. The parentheses stand for the improvement compared to the Baseline." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.353, + 0.88, + 0.522 + ], + "angle": 0, + "content": "
AugmentationAcc.
Top-1Top-5
Baseline65.3789.82
Cutout69.1791.12
Cutmix69.8291.76
Mixup67.5490.23
TextManiA70.81 (+5.44)92.37 (+2.55)
Cutout + TextManiA69.71 (+4.34)91.32 (+1.50)
Cutmix + TextManiA71.05 (+5.68)92.22 (+2.40)
Mixup + TextManiA70.56 (+5.19)91.58 (+1.76)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.526, + 0.894, + 0.57 + ], + "angle": 0, + "content": "Table 6. Classification results (\\%) on CIFAR-100-10% with VIT-Tiny. The configuration follows Table 5. The parentheses stands for the improvement compared to the Baseline." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.599, + 0.885, + 0.617 + ], + "angle": 0, + "content": "4.2. Evenly Distributed Scarce Data Classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.895, + 0.776 + ], + "angle": 0, + "content": "Experimental Setting. For evaluating the effectiveness of TextMania on the scarce dataset, we use \\(10\\%\\) data of the CIFAR-100 [37] and Tiny/ImageNet [40] datasets, named CIFAR-100- \\(10\\%\\) and Tiny/ImageNet- \\(10\\%\\), respectively. CIFAR-100 has 100 classes with 500 training images per class, but we only use randomly sampled 50 images per class. Tiny/ImageNet is a subset of ImageNet-1k [61] with 100k images and 200 classes, but we use 10k images (50 images per class) for simulating a small dataset. Note that the evaluation set is same with those of the original datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.778, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The baseline models of scarce data classification are ResNet18 [28] and ViT-Tiny [19]. Due to the space limit, details of training can be found in the supplementary material. Results. We demonstrate the effectiveness of TextMania compared to mix-based augmentations on evenly distributed scarce datasets. As in Table 5 for CIFAR-100-10%, TextMania outperforms other methods when a single augmentation is used. Furthermore, the effect is amplified when" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2532" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.468, + 0.284 + ], + "angle": 0, + "content": "
AugmentationAcc.
Top-1Top-5
Baseline25.9450.53
Cutout26.4150.28
Cutmix25.9449.67
Mixup29.3454.10
ManiMixup28.4353.25
TextMania29.37 (+3.43)52.37 (+1.84)
Cutout + TextMania29.14 (+3.20)52.60 (+2.07)
Cutmix + TextMania29.86 (+3.92)54.31 (+3.78)
Mixup + TextMania31.15 (+5.21)56.71 (+6.18)
ManiMixup + TextMania32.39 (+6.35)58.25 (+7.72)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.289, + 0.471, + 0.332 + ], + "angle": 0, + "content": "Table 7. Classification results on Tiny/ImageNet-10% with ResNet18. The configuration follows Table 5. The parentheses represent the improvement compared to the Baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.473, + 0.508 + ], + "angle": 0, + "content": "our method and mix-based methods are combined, with particularly good compatibility with Manifold Mixup. The results demonstrate the importance of intra-class semantic perturbation along with inter-class in scarce data settings. This tendency is also observed with another baseline architecture in Table 6, and datasets in Table 7, implying that TextMania is model-agnostic to be applied. The overall results demonstrate the potential of TextMania to enrich the visual feature space using text modalities and develop more accurate and robust models in scarce data regimes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.318, + 0.531 + ], + "angle": 0, + "content": "4.3. Few-shot Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.472, + 0.75 + ], + "angle": 0, + "content": "Experimental Setting. We evaluate TextMania on the PASCAL VOC [20] and MS-COCO [45] datasets with a few-shot division following Wang et al. [75]. For VOC, we have three random splits, which have different divisions into 15 base classes and 5 novel classes among the 20 total classes, and \\( K = 1, 2, 3, 5, 10 \\) objects are sampled from the novel classes. We utilize the VOC2007 test set for evaluation with AP50 metrics and train with the combination of the VOC2007 and VOC2012 train/val set. For COCO, the base classes are disjoint with VOC classes while the remaining classes are used as novel classes, and \\( K = 1, 3, 5, 10, 30 \\) objects are sampled from the novel classes for few-shot finetuning. We use 5k images from the validation set in COCO for evaluation with mAP metrics and the rest for training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.825 + ], + "angle": 0, + "content": "The baseline [80] is the Faster R-CNN [59] trained with the base classes first and then fine-tuned with the novel classes. TextMania is applied to the novel class samples during the fine-tuning stage. Following the prior studies, all the reported results are averaged over 10 repeated runs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Results. Note that we apply TextManiA only on the classification head; thus, the quality of the regressed bounding boxes will remain similar as before applying TextManiA. As shown in Table 8 for VOC and Table 9 for COCO, TextManiA improves the AP by improving only the clas" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.892, + 0.284 + ], + "angle": 0, + "content": "
SplitAug.K-shot
123510
AllBaseline12.8216.6520.0420.6423.19
TextManiA17.74 (+4.92)22.40 (+5.75)23.37 (+3.33)25.09 (+4.45)24.22 (+1.03)
1Baseline15.1118.8222.6121.9723.74
TextManiA21.94 (+6.83)26.44 (+7.62)23.66 (+1.05)25.88 (+3.91)25.14 (+1.40)
2Baseline10.8614.2218.6719.3422.49
TextManiA14.64 (+3.78)18.49 (+4.27)23.28 (+4.61)23.06 (+3.72)24.44 (+1.95)
3Baseline12.4916.9018.8420.6123.35
TextManiA16.65 (+4.16)22.26 (+5.36)23.16 (+4.32)26.33 (+5.72)25.08 (+1.73)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.289, + 0.893, + 0.332 + ], + "angle": 0, + "content": "Table 8. Few-shot object detection results (AP50) on VOC. The value in the parentheses indicates the improvement compared to the Baseline of each split set." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.345, + 0.891, + 0.423 + ], + "angle": 0, + "content": "
Aug.K- shot
1351030
Baseline3.434.666.109.1112.78
TextMania5.39(+1.96)6.47(+1.81)7.80(+1.70)10.03(+0.92)13.60(+0.82)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.428, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Table 9. Few-shot object detection results (mAP) on COCO. The configuration follows Table 8." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.469, + 0.891, + 0.568 + ], + "angle": 0, + "content": "
Aug.ManyMediumFewIF=100IF=50IF=10
(a)Baseline71.1138.423.0038.3943.3359.29
Random71.3738.552.9038.4343.2860.39
TextManiA70.2240.739.4141.1047.1760.67
(b)Direct.71.3438.644.3238.6643.4459.82
Concat.68.0235.825.3536.9842.6859.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.572, + 0.892, + 0.601 + ], + "angle": 0, + "content": "Table 10. Comparison to (a) random perturbation, and (b) direct text and concatenated embeddings on CIFAR-100-LT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.894, + 0.703 + ], + "angle": 0, + "content": "sification accuracy, where the result is in a similar line to the analysis [6] that classification error weighs more than localization error. The improvement is clearer when \\( K \\) is low. The results demonstrate the applicability of TextManiA to enhance the classification accuracy of detection models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.671, + 0.728 + ], + "angle": 0, + "content": "4.4. Further Analyses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.81 + ], + "angle": 0, + "content": "Random Baseline. In Table 10-(a), we compare our method with the Random baseline. We randomly sample a vector from a Normal distribution \\(\\mathcal{N}(0,1)\\) and use it instead of the difference vector, i.e., augmenting visual features with random perturbations on the same manifold of visual features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.9 + ], + "angle": 0, + "content": "The result shows that the Random baseline improves performance by serving as intra-perturb, but marginal compared to our method considering semantics additionally, which implies that semantic information embedded in the difference vector guides the augmentation more effective direction rather than random." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2533" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.447, + 0.21 + ], + "angle": 0, + "content": "
CLIP Arch.Aug.ZSLP-10%LP-Full
ResNet50Baseline39.4750.1863.64
TextManiA-52.8364.17
ResNet101Baseline45.1757.3768.60
TextManiA-59.4969.12
ViT-BBaseline58.2173.3079.99
TextManiA-73.3579.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.215, + 0.47, + 0.258 + ], + "angle": 0, + "content": "Table 12. Classification results (\\%) of CLIP with zero-shot (ZS) and linear-probe (LP) on Full and \\(10\\%\\) CIFAR-100. We apply our TextManiA to the linear-probed CLIP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.471, + 0.463 + ], + "angle": 0, + "content": "Effectiveness of Difference Vectors. While we use the difference vectors by subtracting the embeddings with and without attribute words for TextMania, there could be another way to extract the attribute information. In Table 10(b), we compare with counterparts, direct text embedding (Direct.) and concatenated embeddings (Concat.). For the Direct method, we use the text embedding computed from the attribute word directly instead of the difference vector. For the Concat method, we concatenate the text embeddings from with and without attribute words, e.g., [“bull”||“red bull”], and use it instead of the difference vector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.681 + ], + "angle": 0, + "content": "The results show that using difference vector (TextMania) outperforms using direct text embedding or concatenated embeddings, and imply that remaining contextual information after subtraction plays an important role in doing intra-perturbation in a semantic way. Although the word \"blue\" can function as both an adjective and a noun, its exact role in a sentence cannot be determined solely based on the word itself. Our intention of subtraction is for attribute words to act as a modifier in the sentence motivated by word analogy. When we computed the cosine similarity, embeddings derived directly from \"red\" and those obtained from the difference exhibited low similarity because they contain different contextual information despite the same origin of a word." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.687, + 0.256, + 0.822 + ], + "angle": 0, + "content": "Linear Probing with Advanced Models. Further demonstrating the compatibility of TextMania, we apply our method during linear probing of the model. In Table 11, we test VL-LTR [70], the state-of-the-art model in long-tail" + }, + { + "type": "table", + "bbox": [ + 0.283, + 0.693, + 0.454, + 0.753 + ], + "angle": 0, + "content": "
ModelLP-Full
VL-LTR61.04
+TextMania61.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.285, + 0.758, + 0.47, + 0.813 + ], + "angle": 0, + "content": "Table 11. Comparison between the SOTA model with and without TextMania during linear probing on CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.822, + 0.47, + 0.898 + ], + "angle": 0, + "content": "classification, on CIFAR-100. In Table 12, we use a CLIP image encoder [54] with various architectures as the baseline model and linear-probe the model on both \\(10\\%\\) and full data of CIFAR-100. The results demonstrate that TextMania is compatible with linear-probed CLIP and VL-LTR models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.71, + 0.106 + ], + "angle": 0, + "content": "Ablation Study on Attributes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.107, + 0.714, + 0.226 + ], + "angle": 0, + "content": "In TextManiA, we have considered color and size attributes. To confirm the effect of each attribute, we conduct an ablation study on attributes in Table 13. The result shows that while each attribute brings non-trivial gain, using both brings more gain. We" + }, + { + "type": "table", + "bbox": [ + 0.761, + 0.093, + 0.871, + 0.172 + ], + "angle": 0, + "content": "
ColorSizeAcc.
31.10
✓33.48
✓33.89
✓✓34.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.743, + 0.177, + 0.892, + 0.218 + ], + "angle": 0, + "content": "Table 13. Ablation study on the attributes with CIFAR-100-10%." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.228, + 0.892, + 0.272 + ], + "angle": 0, + "content": "believe that there are additional attributes we could use and a more effective method for selecting appropriate attributes, but leave it for future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.31, + 0.619, + 0.325 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.895, + 0.525 + ], + "angle": 0, + "content": "To mitigate the scarce data problem in long-tailed data distribution, small dataset, and few-shot cases, we propose a text-driven visual feature manifold augmentation method, TextManiA. Our method densifies around all the given individual visual features by adding a difference vector stem from the text embedding. While the mix-based augmentations inflict semantic perturbation in an inter-class way by label mixing, TextManiA perturbs the semantic meaning of the visual features at an intra-class level, i.e., having semantic perturbation while maintaining its class. The intraclass semantic perturbation is achieved by transferring the attribute-embedded vectors to visual feature space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.895, + 0.744 + ], + "angle": 0, + "content": "To scrutinize the design of our estimated attribute embedding, we conduct visualization-based analyses: t-SNE plot and simple manipulation tests. The results empirically demonstrate that TextManiA readily enriches the sparse samples with comprehensible manipulation, since the general language models also reflect some extent of visual information. The experiment on the long-tail classification validates the effectiveness of our method, especially on the highly skewed class distribution. We additionally show the compatibility of TextManiA with other augmentation methods or other models in scarce data cases and during linear probing. In this work, note that we only use color and size as attributes; thus, there would be room for further investigation of other effective attributes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This work was partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub; No.2022-0-00124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities; No. 2020-0-00004, Development of Provisional Intelligence based on Long-term Visual Memory Network)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2534" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. *ICLR*, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.2 + ], + "angle": 0, + "content": "[2] Guozhong An. The effects of adding noise during backpropagation training on a generalization performance. Neural computation, 8(3):643-674, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.202, + 0.469, + 0.228 + ], + "angle": 0, + "content": "[3] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.231, + 0.471, + 0.285 + ], + "angle": 0, + "content": "[4] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151-175, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.288, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[5] Chris M Bishop. Training with noise is equivalent to tikhonov regularization. Neural computation, 7(1):108-116, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.317, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[6] Ali Borji and Seyed Mehdi Iranmanesh. Empirical upper bound in object detection and more. arXiv preprint arXiv:1911.12451, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.36, + 0.471, + 0.4 + ], + "angle": 0, + "content": "[7] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural networks, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.402, + 0.471, + 0.442 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.446, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[9] Tung-I Chen, Yueh-Cheng Liu, Hung-Ting Su, Yu-Cheng Chang, Yu-Hsiang Lin, Jia-Fong Yeh, Wen-Chin Chen, and Winston Hsu. Dual-awareness attention for few-shot object detection. IEEE TMM, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.502, + 0.469, + 0.529 + ], + "angle": 0, + "content": "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.471, + 0.571 + ], + "angle": 0, + "content": "[11] Peng Chu, Xiao Bian, Shaopeng Liu, and Haibin Ling. Feature space augmentation for long-tailed data. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.471, + 0.614 + ], + "angle": 0, + "content": "[12] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[13] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPRW, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.471, + 0.7 + ], + "angle": 0, + "content": "[14] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.471, + 0.743 + ], + "angle": 0, + "content": "[15] Wenliang Dai, Lu Hou, Lifeng Shang, Xin Jiang, Qun Liu, and Pascale Fung. Enabling multimodal generation on clip via vision-language knowledge distillation. ACL, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.471, + 0.786 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.829 + ], + "angle": 0, + "content": "[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. ACL, 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[18] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.894, + 0.134 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.893, + 0.176 + ], + "angle": 0, + "content": "[20] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.893, + 0.218 + ], + "angle": 0, + "content": "[21] Kevin Frans, Lisa B Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. arXiv preprint arXiv:2106.14843, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.22, + 0.894, + 0.273 + ], + "angle": 0, + "content": "[22] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. *ICLR*, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.894, + 0.327 + ], + "angle": 0, + "content": "[23] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM TOG, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.893, + 0.37 + ], + "angle": 0, + "content": "[24] Caglar Gulcehre, Marcin Moczulski, Misha Denil, and Yoshua Bengio. Noisy activation functions. In ICML, pages 3059-3068. PMLR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.373, + 0.894, + 0.411 + ], + "angle": 0, + "content": "[25] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.893, + 0.453 + ], + "angle": 0, + "content": "[26] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.893, + 0.495 + ], + "angle": 0, + "content": "[27] Bart LM Happel and Jacob MJ Murre. Design and evolution of modular neural network architectures. Neural networks, 7(6-7):985–1004, 1994. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.893, + 0.536 + ], + "angle": 0, + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.893, + 0.579 + ], + "angle": 0, + "content": "[29] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.893, + 0.621 + ], + "angle": 0, + "content": "[30] Lasse Holmstrom and Petri Koistinen. Using additive noise in back-propagation training. IEEE transactions on neural networks, 3(1):24-38, 1992. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.624, + 0.893, + 0.663 + ], + "angle": 0, + "content": "[31] Hanzhe Hu, Shuai Bai, Aoxue Li, Jinshi Cui, and Liwei Wang. Dense relation distillation with context-aware aggregation for few-shot object detection. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.893, + 0.704 + ], + "angle": 0, + "content": "[32] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.893, + 0.761 + ], + "angle": 0, + "content": "[33] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.893, + 0.815 + ], + "angle": 0, + "content": "[34] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In ICLR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[35] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[36] Youmin Kim, Jinbae Park, YounHo Jang, Muhammad Ali, Tae-Hyun Oh, and Sung-Ho Bae. Distilling global and local logits with densely connected relations. In ICCV, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2535" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[37] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, Citeseer, 2009. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.471, + 0.176 + ], + "angle": 0, + "content": "[38] Ashish Kumar, Zipeng Fu, Deepak Pathak, and Jitendra Malik. Rma: Rapid motor adaptation for legged robots. In Robotics: Science and Systems, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.178, + 0.469, + 0.206 + ], + "angle": 0, + "content": "[39] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.208, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[40] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[41] Boyi Li, Felix Wu, Ser-Nam Lim, Serge Belongie, and Kilian Q Weinberger. On feature normalization and data augmentation. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.279, + 0.469, + 0.32 + ], + "angle": 0, + "content": "[42] Bohao Li, Boyu Yang, Chang Liu, Feng Liu, Rongrong Ji, and Qixiang Ye. Beyond max-margin: Class margin equilibrium for few-shot object detection. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.322, + 0.471, + 0.376 + ], + "angle": 0, + "content": "[43] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.378, + 0.469, + 0.419 + ], + "angle": 0, + "content": "[44] Pan Li, Da Li, Wei Li, Shaogang Gong, Yanwei Fu, and Timothy M Hospedales. A simple feature augmentation for domain generalization. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.421, + 0.469, + 0.474 + ], + "angle": 0, + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.477, + 0.469, + 0.518 + ], + "angle": 0, + "content": "[46] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.52, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[47] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.563, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[48] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In ICLR, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.605, + 0.469, + 0.671 + ], + "angle": 0, + "content": "[49] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.675, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[50] Tae-Hyun Oh, Tali Dekel, Changil Kim, Inbar Mosseri, William T Freeman, Michael Rubinstein, and Wojciech Matusik. Speech2face: Learning the face behind a voice. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.732, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[51] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.775, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[52] Guim Perarnau, Joost van de Weijer, Bogdan Raducanu, and Jose M. Álvarez. Invertible conditional gans for image editing. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.817, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[53] Limeng Qiao, Yuxuan Zhao, Zhiyuan Li, Xi Qiu, Jianan Wu, and Chi Zhang. Defrcn: Decoupled faster r-cnn for few-shot object detection. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.893, + 0.119 + ], + "angle": 0, + "content": "transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.894, + 0.174 + ], + "angle": 0, + "content": "[55] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.893, + 0.218 + ], + "angle": 0, + "content": "[56] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.22, + 0.893, + 0.259 + ], + "angle": 0, + "content": "[57] Russell Reed and Robert J MarksII. Neural smithing: supervised learning in feedforward artificial neural networks. Mit Press, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.893, + 0.3 + ], + "angle": 0, + "content": "[58] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In ICML, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.893, + 0.343 + ], + "angle": 0, + "content": "[59] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.893, + 0.385 + ], + "angle": 0, + "content": "[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.387, + 0.893, + 0.441 + ], + "angle": 0, + "content": "[61] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.443, + 0.893, + 0.483 + ], + "angle": 0, + "content": "[62] Fawaz Sammani, Tanmoy Mukherjee, and Nikos Deligiannis. Nlx-gpt: A model for natural language explanations in vision and vision-language tasks. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.485, + 0.893, + 0.524 + ], + "angle": 0, + "content": "[63] Gyungin Shin, Weidi Xie, and Samuel Albanie. Namedmask: Distilling segmenters from complementary foundation models. arXiv preprint arXiv:2209.11228, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.893, + 0.593 + ], + "angle": 0, + "content": "[64] Patrice Y Simard, Yann A LeCun, John S Denker, and Bernard Victorri. Transformation invariance in pattern recognition—tangent distance and tangent propagation. In Neural networks: tricks of the trade, pages 239–274. Springer, 1998. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.893, + 0.663 + ], + "angle": 0, + "content": "[65] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.893, + 0.719 + ], + "angle": 0, + "content": "[66] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.893, + 0.761 + ], + "angle": 0, + "content": "[67] Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. Fsce: Few-shot object detection via contrastive proposal encoding. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.893, + 0.803 + ], + "angle": 0, + "content": "[68] Kim Sung-Bin, Arda Senocak, Hyunwoo Ha, Andrew Owens, and Tae-Hyun Oh. Sound to visual scene generation by audio-to-visual latent alignment. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.805, + 0.893, + 0.845 + ], + "angle": 0, + "content": "[69] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.893, + 0.898 + ], + "angle": 0, + "content": "[70] Changyao Tian, Wenhai Wang, Xizhou Zhu, Jifeng Dai, and Yu Qiao. Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition. In ECCV, 2022. 9" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2536" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "[71] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[72] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.204 + ], + "angle": 0, + "content": "[73] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In ICML, 2019. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.472, + 0.273 + ], + "angle": 0, + "content": "[74] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and León Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.472, + 0.316 + ], + "angle": 0, + "content": "[75] Xin Wang, Thomas E Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. In ICML, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.317, + 0.472, + 0.386 + ], + "angle": 0, + "content": "[76] Zhecan Wang, Noel Codella, Yen-Chun Chen, Luowei Zhou, Xiyang Dai, Bin Xiao, Jianwei Yang, Haoxuan You, Kai-Wei Chang, Shih-fu Chang, et al. Multimodal adaptive distillation for leveraging unimodal encoders for vision-language tasks. arXiv preprint arXiv:2204.10496, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.472, + 0.455 + ], + "angle": 0, + "content": "[77] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.472, + 0.497 + ], + "angle": 0, + "content": "[78] Chien-Sheng Wu, Steven Hoi, Richard Socher, and Caiming Xiong. Tod-bert: Pre-trained natural language understanding for task-oriented dialogue. EMNLP, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.472, + 0.538 + ], + "angle": 0, + "content": "[79] Yang Xiao and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.472, + 0.581 + ], + "angle": 0, + "content": "[80] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In ICCV, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.472, + 0.609 + ], + "angle": 0, + "content": "[81] Lu Yang, He Jiang, Qing Song, and Jun Guo. A survey on long-tailed visual recognition. IJCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.61, + 0.472, + 0.663 + ], + "angle": 0, + "content": "[82] Moon Ye-Bin, Dongmin Choi, Yongjin Kwon, Junsik Kim, and Tae-Hyun Oh. Eninst: Enhancing weakly-supervised low-shot instance segmentation. arXiv preprint arXiv:2302.09765, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.472, + 0.705 + ], + "angle": 0, + "content": "[83] Kim Youwang, Kim Ji-Yeon, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.472, + 0.803 + ], + "angle": 0, + "content": "[84] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.472, + 0.859 + ], + "angle": 0, + "content": "[85] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[86] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 1, 2, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[87] Yuhui Zhang, Jeff Z HaoChen, Shih-Cheng Huang, Kuan-Chieh Wang, James Zou, and Serena Yeung. Diagnosing and rectifying vision models using language. In ICLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[88] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In ECCV, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2537" + } + ] +] \ No newline at end of file diff --git a/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/21a8728d-cd5e-44be-95fc-d79147ae763b_origin.pdf b/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/21a8728d-cd5e-44be-95fc-d79147ae763b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0a9cb8ffc64faf8562435a6de26aadc10a38bb15 --- /dev/null +++ b/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/21a8728d-cd5e-44be-95fc-d79147ae763b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9c06b7789548ec1c59550cfb37f5a19a0779560a7d92fcc65b2fb00bc0a8960 +size 1783480 diff --git a/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/full.md b/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..074588a26c126f0bab0344fff94946757f2408c3 --- /dev/null +++ b/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/full.md @@ -0,0 +1,376 @@ +# TextMania: Enriching Visual Feature by Text-driven Manifold Augmentation + +Moon Ye-Bin + +Jisoo Kim + +Hongyeob Kim3 + +Kilho Son + +Tae-Hyun $\mathrm{Oh}^{1,5,6}$ + +$^{1}$ Dept. of Electrical Engineering and $^{5}$ Grad. School of Artificial Intelligence, POSTECH + +2Columbia University + +3Sungkyunkwan University + +4 Microsoft Azure + +$^{6}$ Institute for Convergence Research and Education in Advanced Technology, Yonsei University + +https://textmania.github.io/ + +# Abstract + +We propose TextManiA, a text-driven manifold augmentation method that semantically enriches visual feature spaces, regardless of class distribution. TextManiA augments visual data with intra-class semantic perturbation by exploiting easy-to-understand visually mimetic words, i.e., attributes. This work is built on an interesting hypothesis that general language models, e.g., BERT and GPT, encompass visual information to some extent, even without training on visual training data. Given the hypothesis, TextManiA transfers pre-trained text representation obtained from a well-established large language encoder to a target visual feature space being learned. Our extensive analysis hints that the language encoder indeed encompasses visual information at least useful to augment visual representation. Our experiments demonstrate that TextManiA is particularly powerful in scarce samples with class imbalance as well as even distribution. We also show compatibility with the label mix-based approaches in evenly distributed scarce data. + +# 1. Introduction + +Learning models, e.g., neural networks, are known to perform well on visual recognition tasks when training and testing datasets present similar distributions [4]. However, their performance often degrades considerably when evaluated in subtly different distributions [69]. One effective way to enhance the generalization ability of a model against such data distribution shifts would be data augmentation [18, 86, 85, 44, 41, 73]. Augmenting data enlarges the support of the training distribution formed by given samples and yields the effect of increasing the amount of data even without additional laborious data collection. By training on augmented data, decision boundaries are smoothed, and the generalization ability of the model is improved [73]. + +There has been a distinctive and successful line of research for label mix-based data augmentation, such as + +![](images/77558d45f938bf35ae6f8307b24cc555b6a5a3f130c0e3d77f83544e7a98931e.jpg) +Figure 1. Illustration of TextMania. Our method augments the target visual feature by leveraging text embedding of the visually mimetic words, which are comprehensible and semantically rich. For example, when the text of the existing class "bull" is manipulated as "red bull" by adding the attribute "red," we can get augmented visual features by reflecting the difference of text embeddings. In this way, TextMania densifies sparse visual feature space using various attributes text. + +Mixup [86], CutMix [85], and manifold Mixup [73], which are effective for model generalization and calibration [25]. The effectiveness of those label mix-based approaches is attributed to semantic perturbation by label mixing [86, 73, 85]. This is a distinctive property from other lines of data augmentation methods, e.g., [74, 44, 41, 64, 13], where they synthesize diverse virtual data that appear differently but retain class semantics of original contents. However, we found that the performance of mix-based augmentation methods is noticeably degraded when training with skewed class distribution having scarce samples for non-major classes, i.e., long-tailed distribution. In real-world, data often exhibit long-tailed class distribution (e.g., Pareto distribution), which cannot be dealt with the prevalent mix-based approaches. This motivates us to seek a semantically rich data augmentation effective for limited data regimes, including long-tailed distribution, scarce data, and few-shot cases. + +In this work, we propose TextMania, a text-driven manifold augmentation for visual features, which is effective for long-tail classes and scarce data. Our TextMania is based + +on an interesting hypothesis that general language models, e.g., BERT [17] and GPT [55], have learned visual information to some extent that can be transferred to visual feature spaces even with no visual training data. With this hypothesis, we semantically enrich the target visual feature space to be trained by leveraging visually mimetic texts, encoded with general language models and transferred to the target space. Specifically, TextMania encodes meaningful attributes such as "red" and "large" to vectors by computing the difference between text embeddings with and without attributes. We add the attribute embeddings to target visual features to mimic those attributes on the target visual feature space. Figure 1 illustrates the augmentation process of TextMania. The input feature (e.g., the visual feature of "bull") is manipulated by adding the attribute vector induced by the attribute text (e.g., "red"), which yields the augmented visual feature (e.g., "red bull"). Thanks to the text modality properties, the augmentations generated by TextMania are symbolic, human-interpretable, and easily controllable. + +Our approach applies semantic perturbation on a different level to that of the label mix-based methods [86, 73, 85]. The mix-based methods augment a sample from a combination of two different class samples, i.e., applying semantic perturbation in an inter-class way. This further aggravates the class imbalance problem in the long-tailed (skewed) class distribution cases. Our TextManiA, whereas, perturbs data in an intra-class way. A sample per each class is selected, and we enrich the semantic granularity of the class using the sample, thus enabling us to better maintain the amount of augmentation balances in the long-tailed class distribution cases. Moreover, TextManiA can densify around the training samples by extrapolating the class semantics along augmented semantic attribute axes. With this, our method can be combined with the label mix-based methods to further improve performance in evenly distributed sparse data cases because they are complementary. + +To empirically support that our attribute vectors transformed from text embeddings are reasonably designed, we devise two visualization-based analyses: with t-SNE [72] and a latent inversion technique. These demonstrate that attribute vectors lead to visually interpretable manifold augmentation of input. We also evaluate our method with two different tasks in scarce data regimes: few-shot object detection and image classification with deficient datasets and long-tail datasets. Our experiments demonstrate that TextManiA is an effective and model-agnostic data augmentation method, especially in scarce data cases, by exploiting the favors of zero-shot attributes. Also, additional studies show the versatility and compatibility of the design of TextManiA. Our key contributions are summarized as: + +- We propose TextMania, which enriches the visual features by conveying attribute information from the text embedding to the target visual feature space. +- We validate our hypothesis of the existence of embedded visual knowledge in pre-trained language encoders despite no training on visual data. +- We demonstrate that TextManiA is especially helpful in augmenting sparse samples in long-tail class cases. +- We show that our TextMania is complementary to other augmentation methods, and in particular, the combination of our TextMania and manifold Mixup [73] noticeably improves the performance in deficient data cases. + +# 2. Related Work + +We brief the related work in the following three perspectives: image data augmentation, foundation models, and target application tasks. In this work, our TextMania augments data by leveraging the text encoder of CLIP [54], BERT [17], or GPT-2² [55]. For main target applications, we focus on long-tail and small data classification and few-shot object detection tasks in the data-scarce regimes. + +Image Data Augmentation. Image data augmentation can be largely divided into whether semantic perturbation exists. Semantic perturbation, in specific, can be further split into methods with or without label mixing. Methods [18, 66, 57, 2, 24, 5, 30, 74, 44, 1, 41, 64, 13, 12] without semantic perturbation, which have no label change, contain primitive image processing and transformation operations. This includes photometric (e.g., color jitter, contrast, blur, noise, etc.) and geometric (e.g., horizontal reflection, rotation, etc.) operations, and advanced augmentations, including Cutout [18] and adaptive combinations [13, 12]. + +In contrast, Mixup [86], CutMix [85], and manifold Mixup [73] execute semantic perturbation along with label mixing. Mixup interpolates two whole input images pixelwisely, CutMix interpolates a partial region of an image with another, and manifold Mixup mixes features from the images. These mix-based methods also augment labels of samples by an inter-class semantic perturbation, where labels of two different class samples are mixed. While the mixed label is known to be effective for generalization and model calibration effects [25], we found that the mix-based methods are heavily affected by class distribution due to sampling from two sources; thus, their effect is restricted to evenly distributed datasets. For datasets with skewed class distributions with tails, the sampling probabilities between major and minor classes would significantly differ, which can exaggerate biased sampling to major classes and makes minor classes more minor. + +Our TextManiA, on the other hand, is applied to all of the given samples uniformly regardless of class distribution. TextManiA densifies around the sample features by perturbing and enriching the semantic meaning of them at an intra-class level, which does not change the label. Moreover, because of the different semantic granularity of perturbation between TextManiA (intra-class) and mix-based methods (inter-class), two methods can be used complementarily when class imbalance does not exist. + +Foundation Models. Recent foundation models [84, 54, 43, 33, 56, 17, 55] have shown a successful case of reflecting human nuances with visually imitated word composition. Particularly, language models, e.g., BERT [17] and GPT [55], show their ability not only in language tasks [78] but also in vision-language multi-modal tasks [62, 22]. Contrastive Language-Image Pretraining (CLIP) [54] also achieves huge success in various tasks even in zero-shot recognition. Follow-up studies show that CLIP representation is effective in conducting other visual tasks by bridging vision and language, e.g., 2D image generation [23, 39, 35, 49], image manipulation [51, 35] and synthesis [21], and even 3D domain tasks [83, 32, 47]. + +In TextMania, we focus on estimating attribute features by exploiting BERT, GPT-2, or CLIP text encoder alone. Distinctively, we only transfer the estimated attribute feature to augment visual features in a different space, which makes our work different from knowledge distillation [29] of foundation models [15, 76, 63]. Rather, our design is an instance of the module neural network structure [27, 3], where recent module-based designs procedurally train the whole model module-by-module with the guidance of the well pretrained module, e.g., [38, 60, 50, 68, 26]. Also, our work is applicable agnostically to architectures; thus, more flexibly applicable than fine-tuning of foundation models [77]. + +Long-tail Classification. In real world, visual data follow a long-tailed distribution which induces class imbalance and leads to performance degrading [81]. A representative line of the methods for long-tail classification is rebalancing [7, 14, 58], which resamples data or reweights the loss for tail classes. However, improvement in performance of the tail classes comes with the sacrifice of head class performance. Note that TextMania densifies all the given samples regardless of the class imbalance, and whereby the model is trained with reasonable variations of training samples for every class at least, which improves the performance while minimizing sacrifice of the head class. + +Few-Shot Object Detection (FSOD). We tackle FSOD, one of the sparse sample problems, to demonstrate the effectiveness of TextManIA and its model architecture agnostic property. FSOD handles novel object classes after the base training for object detection tasks. The model rapidly adapts to novel classes using few data by matching-based [42, 9, 79] or fine-tuning based [31, 67, 75, 53, 82] + +![](images/14fc0c522639b723a5dcb4df77a05f91e74c4c4f7b25f1bc665e81cb2b58631b.jpg) +Figure 2. Overview of TextMania. Given flower image $\mathbf{I}_0$ and class "flower" $T_0$ , we construct the variant text $T_{1}$ by adding the attribute "red" on $T_{0}$ . Embeddings of $T_{0}$ and $T_{1}$ are computed with text encoders, e.g. CLIP [54], BERT [17], or GPT-2 [55] and their difference vector, $\Delta_{0\rightarrow 1} = \mathbf{e}_{T_1} - \mathbf{e}_{T_0}$ is added to the image feature $\mathbf{f}_{I_0}$ after projection $\mathrm{proj}(\cdot)$ and weight $\alpha$ . We make the target feature space semantically rich and plausible by adding the difference vector, which embeds interpretable information. + +methods. TextManiA is evaluated with the fine-tuning-based FSOD approach [75], which facilitates to use general model architectures. + +# 3. TextMania + +In image classification, the class label is typically utilized only as a supervision for measuring the loss. We, instead, propose to treat the class label as additional information, the text describing the class, and derive semantic information from it. However, class label as a text description itself is too coarse to represent rich semantics within a class. For example, a class label "dog" does not represent all the details of the description such as "small size of the brown colored dog." To enrich the detailed semantics over the given coarse class texts, we leverage the attribute words, such as "small size" and "brown colored," that can visually modify objects in images at the semantic level. + +# 3.1. Main Idea + +The main idea of TextManiA is to densify distribution around sparse training samples on the target feature space, making it semantically rich through the difference vectors having plausible attribute information, as depicted in Fig. 1. + +Figure 2 illustrates how TextMania augment data. Suppose we have an image $\mathbf{I}_0$ and corresponding class label $T_0$ . The model generally learns the target task using the image $\mathbf{I}_0$ as an input and the class label $T_0$ as supervision. In this work, we also consider the class label $T_0$ as text information and extract the embedding vector $\mathbf{e}_{T_0} \in \mathbb{R}^{d_c}$ using text encoder, e.g., CLIP [54], BERT [17], or GPT-2 [55], where $d_c$ is the text embedding dimension. For obtaining an embedding vector $\mathbf{e}_{T_0}$ , we use the text embedding of the encoder output directly when using CLIP text encoder, or use the average vector of all the embeddings of the sentence when using other language models such as BERT or GPT-2. + +Specifically, text input $T_0$ is formed with class name and pre-defined prompts, such as "a photo of," "a picture of," and "a sketch of." We also synthesize another text input variant $T_1$ by adding color or size attribute words, such as "red" and "big," and compute the embedding vector $\mathbf{e}_{T_1} \in \mathbb{R}^{d_c}$ . Numerous variants can be created with various attribute words and their combinations, but we explain the case of one variant for convenience. Based on the word vector analogy [48], we hypothesize that the relationship between $T_0$ and $T_1$ is maintained in the text embedding space, i.e., the difference vector $\Delta_{0 \to 1} = \mathbf{e}_{T_1} - \mathbf{e}_{T_0}$ would contain the information of added attributes (this hypothesis is validated in Sec. 3.2). To exploit the difference vector from text embeddings, we design our method on the manifold. + +We can obtain such diverse attribute vectors from various attribute text templates; however, their representation space is not directly related to the visual feature space of the target model we are interested in. To bridge the gap, we project the difference vector to the target feature space with a learnable linear projection layer $\operatorname{proj}(\cdot)$ . Then, we add the projected difference vector to the target image feature $\mathbf{f}_{I_0} \in \mathbb{R}^{d_t}$ obtained from the target task encoder with the input image $\mathbf{I}_0$ , where $d_t$ is the target feature space dimension. A linear layer for $\operatorname{proj}(\cdot)$ would be sufficient to transfer cross-modal information, referring to the cross-modal transferability under the contrastive learning case [87] and our experiments. + +To inject the stochasticity, a mixing weight $\alpha \in \mathbb{R}$ is introduced and randomly sampled from the clamped Normal distribution in the range over 0.1. Then, we have the augmented feature vector $\hat{\mathbf{f}}_{I_0}$ as, + +$$ +\hat {\mathbf {f}} _ {I _ {0}} = \mathbf {f} _ {I _ {0}} + \alpha \cdot \operatorname {p r o j} \left(\boldsymbol {\Delta} _ {0 \rightarrow 1}\right). \tag {1} +$$ + +For the cases having $d_{t} = d_{c}$ , we can set $\operatorname{proj}(\cdot)$ operation to be an identity mapping without any learnable parameter. + +We train the target task model with this augmented feature vector, whose class label is still $T_{0}$ . We note that computing difference attribute vectors with text encoder is computationally expensive. For efficient training, we pre-compute all possible combinations of difference vectors $\{\Delta\}$ and store + +![](images/bc60c8e6a2bf4e2e451789ef5f7ba5c648ecc92f97d504671678e4824e1ed920.jpg) +Label Mix + +![](images/1c5a82e344d30d839bf097d51a75a558511fce77c4ed6f1cfe5b59b212c9d4bf.jpg) +TextMania + +![](images/2e3e6bad334560fe657c439c8e3d91bcc6695576f1651cb24d5777fbcc9aa415.jpg) +TextManiA + Label Mix +Figure 3. Comparison of a typical label mix-based augmentation, TextManiA, and their combination, in a sparse data case. Given two samples for two classes, the augmented samples by a label mix-based method are at points lying on six lines resulting from the combinations of samples, i.e., inter-class perturbation. TextManiA densifies the sample distribution along semantic attribute axes in an intra-class. The combination of them yields synergy of their respective advantages. + +them in a look-up table because class names and attributes can be pre-determined and unchanged during training. + +Different from knowledge distillation [15, 76, 63, 29, 36], TextManiA does not transfer-learn the text embeddings directly. Instead, the difference vector projected onto the target domain is injected into the target model, allowing our method to be applied to arbitrary target models. Since the visual feature augmentation is solely controlled by text, TextManiA is human-interpretable and easily controllable. + +Compared to label mix-based augmentations [86, 73, 85], TextManiA has advantages in imbalanced data distribution. We suppose a scenario where few samples are in one class and many samples are in another class. The augmented points by a mix-based method would be located only on the interpolation lines between the given samples, which limits the augmentation effects, as depicted in Fig. 3. If we apply a mix-based method in the long-tailed class distribution cases, i.e., notably skewed distribution, the class imbalance problem is further aggravated, and augmentation is more biased toward major classes. In contrast, TextManiA can equally densify all the given samples since it augments each sample independently. Thus, TextManiA can be used in general regardless of the imbalance factor of class distribution. + +On the other hand, in another scenario with small training data but with uniform class distribution, both TextMania and mix-based methods would increase diverse combinations of samples by augmentation in respective aspects, which leads to complementary performance improvement. This will be empirically demonstrated in Sec. 4. + +# 3.2. Characteristics of Attribute Embedding + +To scrutinize the relationship between the text $T_{0}$ , text variant $T_{1}$ and the attribute embedding $\Delta_{0\rightarrow 1}$ , we visualize their distribution and discuss the characteristics. We also visualize the difference vector to verify the hypothesis that the difference vector embeds its corresponding attribute. + +![](images/d7d4dbd599ac5c5f805e51e6bb53a90cd7c45c828819bbef8f9ea9ded9b15a26.jpg) +Figure 4. The t-SNE plot of difference vectors (e.g., "brown dog" - "dog") projected to visual feature space. The colors of the points represent color attributes used for computing the difference vector, and we use all the classes in CIFAR-100 for this plot. As a comparison, the colored points in the red circle show direct color-text embedding (e.g., "brown") projected to the visual feature space. + +Embedding Difference vs. Direct Text Embedding. When guessing the difference between two texts, e.g., "brown X" - "X," it would be "brown." Someone may think of using the text embedding directly obtained from "brown" instead of our attribute embedding from "brown X" - "X." To understand the difference between the two representations, we visualize the difference vectors and text embeddings with BERT and CLIP text encoder in Fig. 4. While the direct text embeddings in the red circle of Fig. 4 are clustered no matter with different color-texts, the difference vectors are well clustered dependent on the color. This observation indicates that the difference vector is more effective in augmenting the visual feature space than text embedding. In addition, the difference vectors obtained from the same attribute word are similarly clustered regardless of the class "X" but slightly different. It may imply our attribute embedding has subtle difference awareness on granularity according to class. + +Note that Fig. 4 presents difference vectors in the visual feature space, and we also observe similar distributions of difference vectors in the original text embedding space, i.e., t-SNE is not employed for clustering the direct text embeddings. This observation supports our hypothesis that general language models, e.g., BERT or GPT, have learned visual information to some extent. It, also, demonstrates the visual information is properly transferred to the target visual feature space. + +Do We Need to Rule out Unrealistic Attributes? One can be curious about how TextMania handles the unrealistic attribute, such as "blue cow." We intentionally include such unrealistic attributes, motivated by other contexts in self- and semi-supervised learning [10, 65, 8], where they showed the strong benefit of unnatural strong augmentations to train neural networks. This observation regarding strong + +![](images/0f96afdd265fdda275adad38210782c4f070163ff9970d0b39bfea273e97afc8.jpg) +Figure 5. The attribute embedding visualization through image manipulation examples. We analyze how an image is manipulated when a difference vector, containing specific attribute information, is injected to the original image feature. (Top) We visualize an example of generated image given a specific class and its manipulated pair by size and color attribute, respectively. (Bottom) From left to right, we visualize the real image of "Butterfly" and its manipulated image pair with gigantic, green, and dot pattern, respectively. + +![](images/0df78f704b5fea91b93d95831053246926a44e560afdf522bb7d6e932f8d1bc4.jpg) + +augmentation is consistent with the design of TextMania containing unrealistic attributes. + +Does Difference Vector Embed Attribute? To visually understand whether attribute editing is reflected while maintaining class information, we attempt to manipulate images by changing their features with the difference vectors $\Delta_{0\rightarrow 1} = \mathbf{e}_{T_1} - \mathbf{e}_{T_0}$ , i.e., we want to visualize the change effect between $\mathbf{f}_{\mathbf{I}_0}$ and $\mathbf{e}_{a_1} = \mathbf{f}_{\mathbf{I}_0} + \alpha \Delta_{0\rightarrow 1}$ in image domain. To see the effect in image domain, we need to invert the change from $\mathbf{e}_{\mathbf{I}_0}$ to $\mathbf{e}_{a_1}$ in image domain, which can be formulated as the following optimization problem, + +$$ +\arg \min _ {\mathbf {I}} \| E _ {i} (\mathbf {I}) - \mathbf {e} _ {a _ {1}} \| _ {1}, \tag {2} +$$ + +where $E_{i}(\cdot):\mathbf{I}\rightarrow \mathbf{f}$ denotes the image encoder in Fig. 2. Direct optimization in Eq. (2) is known to be difficult [88]; thus, we parameterize a given image with an image generator $G_{\theta}$ with a latent code $\mathbf{z}$ , i.e., $\mathbf{I}(\theta) = G_{\theta}(\mathbf{z})$ , which is known to ease the optimization [71]. Then, we can obtain the visualization by the following optimization over $\theta$ + +$$ +\arg \min _ {\theta} \| E _ {i} \left(G _ {\theta} (\mathbf {z})\right) - \mathbf {e} _ {a _ {1}} \| _ {1}, \tag {3} +$$ + +where $E_{i}(\cdot)$ , $\mathbf{z}$ , and $\mathbf{e}_{a_1}$ are frozen during the optimization. Since our goal is to see the move from $\mathbf{f}_{\mathbf{I}_0} = E_i(\mathbf{I}_0)$ to $\mathbf{e}_{a_1}$ for the query $\mathbf{I}_0$ , we initialize $\theta$ and $\mathbf{z}$ such that $G_{\theta}(\mathbf{z}) \simeq \mathbf{I}_0$ by the GAN inversion [88]. We use IC-GAN [52] for the image generator and the text embeddings are obtained from the CLIP text encoders. Note that $E_{i}(\cdot)$ is trained with random perturbation of transformed attributes; thus, the visualization through $E_{i}(\cdot)$ is different from that of CLIP encoders. Details can be found in the supplementary material. + +Figure 5 shows that the manipulated image reflects the added attribute, i.e., the size of the dog is reduced by the size attribute "small," and the bird becomes yellow by injecting + +the color attribute "yellow." The manipulated results imply that 1) the difference vector indeed embeds the attributes while preserving its semantics, and 2) our augmentation on the feature space may have analogous effects to an image-level augmentation but without implementing complicated image perturbation operations. Note that these visualizations are for analysis purposes but not for competing with any existing image manipulation methods. + +# 4. Experiments + +We evaluate TextMania in various cases presenting sparse data with different tasks: long-tail classification in Sec. 4.1, evenly distributed scarce data classification in Sec. 4.2, and few-shot object detection in Sec. 4.3. We also conduct additional studies demonstrating the effectiveness of the design of our method and the versatility of TextMania in Sec. 4.4. Additional experimental results and details can be found in the supplementary material. + +# 4.1. Long-tail Classification + +Experimental Setting. We compare TextMania with the mix-based augmentations on the CIFAR-100-LT [14] and ImageNet-LT [46] datasets, where LT stands for long-tailed distribution. They are artificially truncated to have a long-tail from each original dataset, CIFAR-100 [37] and ImageNet2012 [16]. Long-tail datasets usually have three sets of classes: Many-shot (more than 100 images), Medium-shot (20-100 images), and Few-shot (less than 20 images). + +For CIFAR-100-LT, we control the imbalance factor (IF) [11] computed as the ratio of samples in the head to tail class, $N_{1} / N_{K}$ , where $N_{k} = |\mathcal{D}_{k}|$ , and $\mathcal{D}_k$ is the set of samples belonging to the class $k\in \{1,\dots ,K\}$ . A larger value of the IF represents a more severe imbalance in data, which is more challenging. We evaluate the performance according to different IFs of 100, 50, and 10. + +We utilize ResNet18 as the baseline on CIFAR-100-LT and ResNext50 on ImageNet-LT. We use the validation set of the original datasets to measure the Top-1 accuracy. Note that we apply each augmentation on all the samples without carefully selecting a set of classes in Table 1. + +Results. Table 1 presents the long-tail classification results on CIFAR-100-LT, which show consistent improvement with TextManiA. Also, TextManiA with various text encoders achieves analogous improvement trend regardless of the imbalance factor but marginal degradation on Many class of IF=100 when using general language model, BERT and GPT-2. While the performance gain is from leaking pretrained language information, it is surprising and a virtue that the language models never exposed to any image can improve the visual recognition performance. In comparison to single usage of mix-based augmentations, our method shows higher accuracy because of uniform effects of TextManiA on samples regardless of class imbalance. The mix-based + +
(a) AugmentationImbalance Factor (IF)
1005010
Baseline38.3943.3359.29
TextMania (CLIP)40.65 (+2.26)46.48 (+3.15)60.17 (+0.88)
TextMania (BERT)41.10 (+2.71)47.17 (+3.84)60.67 (+1.38)
TextMania (GPT-2)41.20 (+2.81)46.93 (+3.60)60.94 (+1.65)
Cutout [18]37.5142.2859.26
+ TextMania40.35 (+2.84)45.48 (+3.20)61.31 (+2.05)
Cutmix [85]37.9343.3459.30
+ TextMania40.22 (+2.29)45.36 (+2.02)61.30 (+2.00)
Mixup [86]36.7540.7757.50
+ TextMania38.40 (+1.65)43.33 (+2.56)59.80 (+2.30)
ManiMixup [73]35.7240.5155.26
+ TextMania38.60 (+2.88)43.22 (+2.71)59.35 (+4.09)
+ +
(b) AugmentationSet of Classes (IF=100)
ManyMediumFew
Baseline71.1138.423.00
TextMania (CLIP)71.14 (+0.03)40.28 (+1.86)7.53 (+4.53)
TextMania (BERT)70.22 (-0.89)40.73 (+2.31)9.41 (+6.41)
TextMania (GPT-2)70.60 (-0.51)40.61 (+2.19)9.93 (+6.93)
Cutout71.5435.941.06
+ TextMania71.94 (+0.83)40.97 (+2.55)4.03 (+3.03)
Cutmix72.0237.170.90
+ TextMania72.37 (+0.35)40.80 (+3.63)3.90 (+3.00)
Mixup71.9733.620.36
+ TextMania71.97 (+0.00)36.77 (+3.15)1.83 (+1.47)
ManiMixup72.9729.510.70
+ TextMania73.20 (+0.23)36.80 (+7.29)0.76 (+0.06)
+ +Table 1. Long-tail classification results (\%) on CIFAR-100-LT with ResNet18. (a) The accuracy with respect to the different imbalance factors, i.e., IF=\{100, 50, 10\}. (b) The accuracy of each class set with IF=100. Baseline contains random horizontal flip, random crop and rotation, and normalization, applied in all experiments. TextMania without parenthesis uses CLIP for the text encoder. + +
Aug.CBSAllManyMediumFew
Baseline38.3971.1138.423.00
Cutmix✓38.2371.7737.791.90
Mixup✓38.7371.6037.643.16
ManiMixup✓38.5671.2537.882.80
TextMania40.6571.1440.287.53
+ +Table 2. Comparison to label mix-based augmentations with classbalanced sampling (CBS) on CIFAR-100-LT with IF=100. CBS samples two classes first and then samples data in each classes. + +methods, on the other hand, sample two data points from the total dataset, where the probability that a tail class sample contributes to a resulting augmented sample is very low. Even with class-balanced sampling on mixed-based augmentation in Table 2, TextMania performs better, further demonstrating our effectiveness. + +Particularly in Table 1(b), the mix-based methods have degraded performance in the Medium and Few-shot classes, while our TextManiA improves performance. Combining + +
AugmentationSet of ClassesTotal
ManyMediumFew
Baseline85.3470.4742.8072.24
TextMania85.4071.7548.4973.68
Cutout [18]85.0270.3242.9172.07
+ TextMania85.3371.7048.5473.65
Cutmix [85]84.8569.9035.8270.77
+ TextMania85.3071.9347.0473.52
Mixup [86]84.9670.2740.2071.63
+ TextMania84.5569.9535.7270.66
ManiMixup [73]84.8469.9738.8371.24
+ TextMania85.4271.9846.6573.54
+ +Table 3. Long-tail classification results (\%) on ImageNet-LT with ViT, and color the value as best and second best. Baseline contains random horizontal flip, random resize crop, color jitter, and normalization, applied in all experiments. + +
MethodManyMediumFewAll
LWS [34]63.3448.0827.1951.14
cRT [34]61.8046.2027.4049.60
cRT+TextManiA62.7448.6029.6751.47
+ +the mix-based methods with TextMania improves overall performance, but the tendency to sacrifice the Medium and Few-shot classes is the same as before combining. Additionally, while Cutout has performance degradation due to the information loss [85], it is not affected by skewness due to no mix between inter-classes; thus, the performance is higher than the mix-based one in the long-tailed distribution. + +We also evaluate our TextMania on the large-scale dataset ImageNet-LT. In Table 3, the best and second best results are with TextMania, which demonstrate that our augmentation method is also effective in the large-scale long-tailed data distribution, consistent with the CIFAR-100-LT results in Table 1. The improvement with TextMania implies the importance of intra-class perturbation, which can uniformly affect the samples regardless of the skewness of the class distribution. + +In Table 4, we compare with LWS [34], cRT [34], and TextMania on cRT. LWS and cRT are one of effective methods in recent long-tailed recognition. The result shows that TextMania on cRT achieves the best results compared to the counterparts in all classes except for the Many class, wherefrom ours achieves second best. Overall, TextMania improves well-established works, e.g., LWS, and cRT, and it demonstrates the compatibility of our method. + +Table 4. Long-tail classification results $(\%)$ on ImageNet-LT with ResNext50. We compare with LWS, cRT, and TextManiA on cRT, and color the value as best and second best. + +
AugmentationAcc.
Top-1Top-5
Baseline31.1059.14
Cutout32.0360.53
Cutmix32.4361.04
Mixup32.7262.47
ManiMixup33.7463.29
TextMania34.52 (+3.42)65.74 (+6.60)
Cutout + TextMania33.91 (+2.81)61.58 (+2.44)
Cutmix + TextMania35.61 (+4.51)63.82 (+4.68)
Mixup + TextMania37.97 (+6.87)66.75 (+7.61)
ManiMixup + TextMania38.02 (+6.92)67.28 (+8.14)
+ +Table 5. Classification results (\%) on CIFAR-100-10% with ResNet18. Baseline represents random horizontal flip, random crop, and normalization, basically applied in all experiments. The parentheses stand for the improvement compared to the Baseline. + +
AugmentationAcc.
Top-1Top-5
Baseline65.3789.82
Cutout69.1791.12
Cutmix69.8291.76
Mixup67.5490.23
TextManiA70.81 (+5.44)92.37 (+2.55)
Cutout + TextManiA69.71 (+4.34)91.32 (+1.50)
Cutmix + TextManiA71.05 (+5.68)92.22 (+2.40)
Mixup + TextManiA70.56 (+5.19)91.58 (+1.76)
+ +Table 6. Classification results (\%) on CIFAR-100-10% with VIT-Tiny. The configuration follows Table 5. The parentheses stands for the improvement compared to the Baseline. + +# 4.2. Evenly Distributed Scarce Data Classification + +Experimental Setting. For evaluating the effectiveness of TextMania on the scarce dataset, we use $10\%$ data of the CIFAR-100 [37] and Tiny/ImageNet [40] datasets, named CIFAR-100- $10\%$ and Tiny/ImageNet- $10\%$ , respectively. CIFAR-100 has 100 classes with 500 training images per class, but we only use randomly sampled 50 images per class. Tiny/ImageNet is a subset of ImageNet-1k [61] with 100k images and 200 classes, but we use 10k images (50 images per class) for simulating a small dataset. Note that the evaluation set is same with those of the original datasets. + +The baseline models of scarce data classification are ResNet18 [28] and ViT-Tiny [19]. Due to the space limit, details of training can be found in the supplementary material. Results. We demonstrate the effectiveness of TextMania compared to mix-based augmentations on evenly distributed scarce datasets. As in Table 5 for CIFAR-100-10%, TextMania outperforms other methods when a single augmentation is used. Furthermore, the effect is amplified when + +
AugmentationAcc.
Top-1Top-5
Baseline25.9450.53
Cutout26.4150.28
Cutmix25.9449.67
Mixup29.3454.10
ManiMixup28.4353.25
TextMania29.37 (+3.43)52.37 (+1.84)
Cutout + TextMania29.14 (+3.20)52.60 (+2.07)
Cutmix + TextMania29.86 (+3.92)54.31 (+3.78)
Mixup + TextMania31.15 (+5.21)56.71 (+6.18)
ManiMixup + TextMania32.39 (+6.35)58.25 (+7.72)
+ +our method and mix-based methods are combined, with particularly good compatibility with Manifold Mixup. The results demonstrate the importance of intra-class semantic perturbation along with inter-class in scarce data settings. This tendency is also observed with another baseline architecture in Table 6, and datasets in Table 7, implying that TextMania is model-agnostic to be applied. The overall results demonstrate the potential of TextMania to enrich the visual feature space using text modalities and develop more accurate and robust models in scarce data regimes. + +# 4.3. Few-shot Object Detection + +Experimental Setting. We evaluate TextMania on the PASCAL VOC [20] and MS-COCO [45] datasets with a few-shot division following Wang et al. [75]. For VOC, we have three random splits, which have different divisions into 15 base classes and 5 novel classes among the 20 total classes, and $K = 1, 2, 3, 5, 10$ objects are sampled from the novel classes. We utilize the VOC2007 test set for evaluation with AP50 metrics and train with the combination of the VOC2007 and VOC2012 train/val set. For COCO, the base classes are disjoint with VOC classes while the remaining classes are used as novel classes, and $K = 1, 3, 5, 10, 30$ objects are sampled from the novel classes for few-shot finetuning. We use 5k images from the validation set in COCO for evaluation with mAP metrics and the rest for training. + +The baseline [80] is the Faster R-CNN [59] trained with the base classes first and then fine-tuned with the novel classes. TextMania is applied to the novel class samples during the fine-tuning stage. Following the prior studies, all the reported results are averaged over 10 repeated runs. + +Results. Note that we apply TextManiA only on the classification head; thus, the quality of the regressed bounding boxes will remain similar as before applying TextManiA. As shown in Table 8 for VOC and Table 9 for COCO, TextManiA improves the AP by improving only the clas + +Table 7. Classification results on Tiny/ImageNet-10% with ResNet18. The configuration follows Table 5. The parentheses represent the improvement compared to the Baseline. + +
SplitAug.K-shot
123510
AllBaseline12.8216.6520.0420.6423.19
TextManiA17.74 (+4.92)22.40 (+5.75)23.37 (+3.33)25.09 (+4.45)24.22 (+1.03)
1Baseline15.1118.8222.6121.9723.74
TextManiA21.94 (+6.83)26.44 (+7.62)23.66 (+1.05)25.88 (+3.91)25.14 (+1.40)
2Baseline10.8614.2218.6719.3422.49
TextManiA14.64 (+3.78)18.49 (+4.27)23.28 (+4.61)23.06 (+3.72)24.44 (+1.95)
3Baseline12.4916.9018.8420.6123.35
TextManiA16.65 (+4.16)22.26 (+5.36)23.16 (+4.32)26.33 (+5.72)25.08 (+1.73)
+ +Table 8. Few-shot object detection results (AP50) on VOC. The value in the parentheses indicates the improvement compared to the Baseline of each split set. + +
Aug.K- shot
1351030
Baseline3.434.666.109.1112.78
TextMania5.39(+1.96)6.47(+1.81)7.80(+1.70)10.03(+0.92)13.60(+0.82)
+ +Table 9. Few-shot object detection results (mAP) on COCO. The configuration follows Table 8. + +
Aug.ManyMediumFewIF=100IF=50IF=10
(a)Baseline71.1138.423.0038.3943.3359.29
Random71.3738.552.9038.4343.2860.39
TextManiA70.2240.739.4141.1047.1760.67
(b)Direct.71.3438.644.3238.6643.4459.82
Concat.68.0235.825.3536.9842.6859.44
+ +Table 10. Comparison to (a) random perturbation, and (b) direct text and concatenated embeddings on CIFAR-100-LT. + +sification accuracy, where the result is in a similar line to the analysis [6] that classification error weighs more than localization error. The improvement is clearer when $K$ is low. The results demonstrate the applicability of TextManiA to enhance the classification accuracy of detection models. + +# 4.4. Further Analyses + +Random Baseline. In Table 10-(a), we compare our method with the Random baseline. We randomly sample a vector from a Normal distribution $\mathcal{N}(0,1)$ and use it instead of the difference vector, i.e., augmenting visual features with random perturbations on the same manifold of visual features. + +The result shows that the Random baseline improves performance by serving as intra-perturb, but marginal compared to our method considering semantics additionally, which implies that semantic information embedded in the difference vector guides the augmentation more effective direction rather than random. + +
CLIP Arch.Aug.ZSLP-10%LP-Full
ResNet50Baseline39.4750.1863.64
TextManiA-52.8364.17
ResNet101Baseline45.1757.3768.60
TextManiA-59.4969.12
ViT-BBaseline58.2173.3079.99
TextManiA-73.3579.58
+ +Effectiveness of Difference Vectors. While we use the difference vectors by subtracting the embeddings with and without attribute words for TextMania, there could be another way to extract the attribute information. In Table 10(b), we compare with counterparts, direct text embedding (Direct.) and concatenated embeddings (Concat.). For the Direct method, we use the text embedding computed from the attribute word directly instead of the difference vector. For the Concat method, we concatenate the text embeddings from with and without attribute words, e.g., [“bull”||“red bull”], and use it instead of the difference vector. + +The results show that using difference vector (TextMania) outperforms using direct text embedding or concatenated embeddings, and imply that remaining contextual information after subtraction plays an important role in doing intra-perturbation in a semantic way. Although the word "blue" can function as both an adjective and a noun, its exact role in a sentence cannot be determined solely based on the word itself. Our intention of subtraction is for attribute words to act as a modifier in the sentence motivated by word analogy. When we computed the cosine similarity, embeddings derived directly from "red" and those obtained from the difference exhibited low similarity because they contain different contextual information despite the same origin of a word. + +Linear Probing with Advanced Models. Further demonstrating the compatibility of TextMania, we apply our method during linear probing of the model. In Table 11, we test VL-LTR [70], the state-of-the-art model in long-tail + +Table 12. Classification results (\%) of CLIP with zero-shot (ZS) and linear-probe (LP) on Full and $10\%$ CIFAR-100. We apply our TextManiA to the linear-probed CLIP. + +
ModelLP-Full
VL-LTR61.04
+TextMania61.82
+ +classification, on CIFAR-100. In Table 12, we use a CLIP image encoder [54] with various architectures as the baseline model and linear-probe the model on both $10\%$ and full data of CIFAR-100. The results demonstrate that TextMania is compatible with linear-probed CLIP and VL-LTR models. + +# Ablation Study on Attributes. + +In TextManiA, we have considered color and size attributes. To confirm the effect of each attribute, we conduct an ablation study on attributes in Table 13. The result shows that while each attribute brings non-trivial gain, using both brings more gain. We + +Table 11. Comparison between the SOTA model with and without TextMania during linear probing on CIFAR-100. + +
ColorSizeAcc.
31.10
✓33.48
✓33.89
✓✓34.52
+ +Table 13. Ablation study on the attributes with CIFAR-100-10%. + +believe that there are additional attributes we could use and a more effective method for selecting appropriate attributes, but leave it for future work. + +# 5. Conclusion + +To mitigate the scarce data problem in long-tailed data distribution, small dataset, and few-shot cases, we propose a text-driven visual feature manifold augmentation method, TextManiA. Our method densifies around all the given individual visual features by adding a difference vector stem from the text embedding. While the mix-based augmentations inflict semantic perturbation in an inter-class way by label mixing, TextManiA perturbs the semantic meaning of the visual features at an intra-class level, i.e., having semantic perturbation while maintaining its class. The intraclass semantic perturbation is achieved by transferring the attribute-embedded vectors to visual feature space. + +To scrutinize the design of our estimated attribute embedding, we conduct visualization-based analyses: t-SNE plot and simple manipulation tests. The results empirically demonstrate that TextManiA readily enriches the sparse samples with comprehensible manipulation, since the general language models also reflect some extent of visual information. The experiment on the long-tail classification validates the effectiveness of our method, especially on the highly skewed class distribution. We additionally show the compatibility of TextManiA with other augmentation methods or other models in scarce data cases and during linear probing. In this work, note that we only use color and size as attributes; thus, there would be room for further investigation of other effective attributes. + +Acknowledgment. This work was partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub; No.2022-0-00124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities; No. 2020-0-00004, Development of Provisional Intelligence based on Long-term Visual Memory Network). + +# References + +[1] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. *ICLR*, 2017. 2 +[2] Guozhong An. The effects of adding noise during backpropagation training on a generalization performance. Neural computation, 8(3):643-674, 1996. 2 +[3] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In CVPR, 2016. 3 +[4] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151-175, 2010. 1 +[5] Chris M Bishop. Training with noise is equivalent to tikhonov regularization. Neural computation, 7(1):108-116, 1995. 2 +[6] Ali Borji and Seyed Mehdi Iranmanesh. Empirical upper bound in object detection and more. arXiv preprint arXiv:1911.12451, 2019. 8 +[7] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural networks, 2018. 3 +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 5 +[9] Tung-I Chen, Yueh-Cheng Liu, Hung-Ting Su, Yu-Cheng Chang, Yu-Hsiang Lin, Jia-Fong Yeh, Wen-Chin Chen, and Winston Hsu. Dual-awareness attention for few-shot object detection. IEEE TMM, 2021. 3 +[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 5 +[11] Peng Chu, Xiao Bian, Shaopeng Liu, and Haibin Ling. Feature space augmentation for long-tailed data. In ECCV, 2020. 6 +[12] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. In CVPR, 2019. 2 +[13] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPRW, 2020. 1, 2 +[14] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 3, 6 +[15] Wenliang Dai, Lu Hou, Lifeng Shang, Xin Jiang, Qun Liu, and Pascale Fung. Enabling multimodal generation on clip via vision-language knowledge distillation. ACL, 2022. 3, 4 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6 +[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. ACL, 2018. 2, 3, 4 +[18] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 1, 2, 6, 7 +[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 7 +[20] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 8 +[21] Kevin Frans, Lisa B Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. arXiv preprint arXiv:2106.14843, 2021. 3 +[22] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. *ICLR*, 2023. 3 +[23] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM TOG, 2022. 3 +[24] Caglar Gulcehre, Marcin Moczulski, Misha Denil, and Yoshua Bengio. Noisy activation functions. In ICML, pages 3059-3068. PMLR, 2016. 2 +[25] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, 2017. 1, 2 +[26] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In CVPR, 2023. 3 +[27] Bart LM Happel and Jacob MJ Murre. Design and evolution of modular neural network architectures. Neural networks, 7(6-7):985–1004, 1994. 3 +[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7 +[29] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3, 4 +[30] Lasse Holmstrom and Petri Koistinen. Using additive noise in back-propagation training. IEEE transactions on neural networks, 3(1):24-38, 1992. 2 +[31] Hanzhe Hu, Shuai Bai, Aoxue Li, Jinshi Cui, and Liwei Wang. Dense relation distillation with context-aware aggregation for few-shot object detection. In CVPR, 2021. 3 +[32] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, 2022. 3 +[33] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3 +[34] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In ICLR, 2020. 7 +[35] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 3 +[36] Youmin Kim, Jinbae Park, YounHo Jang, Muhammad Ali, Tae-Hyun Oh, and Sung-Ho Bae. Distilling global and local logits with densely connected relations. In ICCV, 2021. 4 + +[37] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, Citeseer, 2009. 6, 7 +[38] Ashish Kumar, Zipeng Fu, Deepak Pathak, and Jitendra Malik. Rma: Rapid motor adaptation for legged robots. In Robotics: Science and Systems, 2021. 3 +[39] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In CVPR, 2022. 3 +[40] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 2015. 7 +[41] Boyi Li, Felix Wu, Ser-Nam Lim, Serge Belongie, and Kilian Q Weinberger. On feature normalization and data augmentation. In CVPR, 2021. 1, 2 +[42] Bohao Li, Boyu Yang, Chang Liu, Feng Liu, Rongrong Ji, and Qixiang Ye. Beyond max-margin: Class margin equilibrium for few-shot object detection. In CVPR, 2021. 3 +[43] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 3 +[44] Pan Li, Da Li, Wei Li, Shaogang Gong, Yanwei Fu, and Timothy M Hospedales. A simple feature augmentation for domain generalization. In ICCV, 2021. 1, 2 +[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 8 +[46] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In CVPR, 2019. 6 +[47] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In CVPR, 2022. 3 +[48] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In ICLR, 2013. 4 +[49] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 3 +[50] Tae-Hyun Oh, Tali Dekel, Changil Kim, Inbar Mosseri, William T Freeman, Michael Rubinstein, and Wojciech Matusik. Speech2face: Learning the face behind a voice. In CVPR, 2019. 3 +[51] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 3 +[52] Guim Perarnau, Joost van de Weijer, Bogdan Raducanu, and Jose M. Álvarez. Invertible conditional gans for image editing. In CVPR, 2016. 5 +[53] Limeng Qiao, Yuxuan Zhao, Zhiyuan Li, Xi Qiu, Jianan Wu, and Chi Zhang. Defrcn: Decoupled faster r-cnn for few-shot object detection. In ICCV, 2021. 3 +[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning + +transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4, 9 +[55] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2, 3, 4 +[56] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3 +[57] Russell Reed and Robert J MarksII. Neural smithing: supervised learning in feedforward artificial neural networks. Mit Press, 1999. 2 +[58] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In ICML, 2018. 3 +[59] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 8 +[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3 +[61] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 7 +[62] Fawaz Sammani, Tanmoy Mukherjee, and Nikos Deligiannis. Nlx-gpt: A model for natural language explanations in vision and vision-language tasks. In CVPR, 2022. 3 +[63] Gyungin Shin, Weidi Xie, and Samuel Albanie. Namedmask: Distilling segmenters from complementary foundation models. arXiv preprint arXiv:2209.11228, 2022. 3, 4 +[64] Patrice Y Simard, Yann A LeCun, John S Denker, and Bernard Victorri. Transformation invariance in pattern recognition—tangent distance and tangent propagation. In Neural networks: tricks of the trade, pages 239–274. Springer, 1998. 1, 2 +[65] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 5 +[66] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 2014. 2 +[67] Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. Fsce: Few-shot object detection via contrastive proposal encoding. In CVPR, 2021. 3 +[68] Kim Sung-Bin, Arda Senocak, Hyunwoo Ha, Andrew Owens, and Tae-Hyun Oh. Sound to visual scene generation by audio-to-visual latent alignment. In CVPR, 2023. 3 +[69] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. 1 +[70] Changyao Tian, Wenhai Wang, Xizhou Zhu, Jifeng Dai, and Yu Qiao. Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition. In ECCV, 2022. 9 + +[71] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In CVPR, 2018. 5 +[72] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 2 +[73] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In ICML, 2019. 1, 2, 4, 6, 7 +[74] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and León Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 1, 2 +[75] Xin Wang, Thomas E Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. In ICML, 2020. 3, 8 +[76] Zhecan Wang, Noel Codella, Yen-Chun Chen, Luowei Zhou, Xiyang Dai, Bin Xiao, Jianwei Yang, Haoxuan You, Kai-Wei Chang, Shih-fu Chang, et al. Multimodal adaptive distillation for leveraging unimodal encoders for vision-language tasks. arXiv preprint arXiv:2204.10496, 2022. 3, 4 +[77] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In CVPR, 2022. 3 +[78] Chien-Sheng Wu, Steven Hoi, Richard Socher, and Caiming Xiong. Tod-bert: Pre-trained natural language understanding for task-oriented dialogue. EMNLP, 2020. 3 +[79] Yang Xiao and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. In ECCV, 2020. 3 +[80] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In ICCV, 2019. 8 +[81] Lu Yang, He Jiang, Qing Song, and Jun Guo. A survey on long-tailed visual recognition. IJCV, 2022. 3 +[82] Moon Ye-Bin, Dongmin Choi, Yongjin Kwon, Junsik Kim, and Tae-Hyun Oh. Eninst: Enhancing weakly-supervised low-shot instance segmentation. arXiv preprint arXiv:2302.09765, 2023. 3 +[83] Kim Youwang, Kim Ji-Yeon, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. ECCV, 2022. 3 +[84] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 3 +[85] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. 1, 2, 4, 6, 7 +[86] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 1, 2, 4, 6, 7 + +[87] Yuhui Zhang, Jeff Z HaoChen, Shih-Cheng Huang, Kuan-Chieh Wang, James Zou, and Serena Yeung. Diagnosing and rectifying vision models using language. In ICLR, 2023. 4 +[88] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In ECCV, 2020. 5 \ No newline at end of file diff --git a/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/images.zip b/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d9b1a3d8469b06ee82110b77c553ff2b5f465b14 --- /dev/null +++ b/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42e97658543667394e4e4f83f79e0239ce06ca58e27bce3f3c16f96dac34e850 +size 603187 diff --git a/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/layout.json b/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3350fb80f56536048cd05b75fc428fe18838b97a --- /dev/null +++ b/2023/TextManiA_ Enriching Visual Feature by Text-driven Manifold Augmentation/layout.json @@ -0,0 +1,10277 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 536, + 121 + ], + "type": "text", + "content": "TextMania: Enriching Visual Feature by Text-driven Manifold Augmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 142, + 143, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 142, + 143, + 156 + ], + "spans": [ + { + "bbox": [ + 69, + 142, + 143, + 156 + ], + "type": "text", + "content": "Moon Ye-Bin" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 143, + 224, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 143, + 224, + 156 + ], + "spans": [ + { + "bbox": [ + 167, + 143, + 224, + 156 + ], + "type": "text", + "content": "Jisoo Kim" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 248, + 143, + 330, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 143, + 330, + 158 + ], + "spans": [ + { + "bbox": [ + 248, + 143, + 330, + 158 + ], + "type": "text", + "content": "Hongyeob Kim3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 354, + 143, + 410, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 143, + 410, + 156 + ], + "spans": [ + { + "bbox": [ + 354, + 143, + 410, + 156 + ], + "type": "text", + "content": "Kilho Son" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 435, + 143, + 520, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 435, + 143, + 520, + 157 + ], + "spans": [ + { + "bbox": [ + 435, + 143, + 520, + 157 + ], + "type": "text", + "content": "Tae-Hyun " + }, + { + "bbox": [ + 435, + 143, + 520, + 157 + ], + "type": "inline_equation", + "content": "\\mathrm{Oh}^{1,5,6}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 83, + 165, + 510, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 165, + 510, + 179 + ], + "spans": [ + { + "bbox": [ + 83, + 165, + 510, + 179 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 165, + 510, + 179 + ], + "type": "text", + "content": "Dept. of Electrical Engineering and " + }, + { + "bbox": [ + 83, + 165, + 510, + 179 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 83, + 165, + 510, + 179 + ], + "type": "text", + "content": "Grad. School of Artificial Intelligence, POSTECH" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 108, + 180, + 216, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 180, + 216, + 194 + ], + "spans": [ + { + "bbox": [ + 108, + 180, + 216, + 194 + ], + "type": "text", + "content": "2Columbia University" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 240, + 180, + 374, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 240, + 180, + 374, + 194 + ], + "spans": [ + { + "bbox": [ + 240, + 180, + 374, + 194 + ], + "type": "text", + "content": "3Sungkyunkwan University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 399, + 180, + 485, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 180, + 485, + 193 + ], + "spans": [ + { + "bbox": [ + 399, + 180, + 485, + 193 + ], + "type": "text", + "content": "4 Microsoft Azure" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 66, + 194, + 527, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 194, + 527, + 209 + ], + "spans": [ + { + "bbox": [ + 66, + 194, + 527, + 209 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 66, + 194, + 527, + 209 + ], + "type": "text", + "content": "Institute for Convergence Research and Education in Advanced Technology, Yonsei University" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 209, + 212, + 380, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 212, + 380, + 224 + ], + "spans": [ + { + "bbox": [ + 209, + 212, + 380, + 224 + ], + "type": "text", + "content": "https://textmania.github.io/" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "spans": [ + { + "bbox": [ + 143, + 251, + 192, + 264 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 290, + 492 + ], + "type": "text", + "content": "We propose TextManiA, a text-driven manifold augmentation method that semantically enriches visual feature spaces, regardless of class distribution. TextManiA augments visual data with intra-class semantic perturbation by exploiting easy-to-understand visually mimetic words, i.e., attributes. This work is built on an interesting hypothesis that general language models, e.g., BERT and GPT, encompass visual information to some extent, even without training on visual training data. Given the hypothesis, TextManiA transfers pre-trained text representation obtained from a well-established large language encoder to a target visual feature space being learned. Our extensive analysis hints that the language encoder indeed encompasses visual information at least useful to augment visual representation. Our experiments demonstrate that TextManiA is particularly powerful in scarce samples with class imbalance as well as even distribution. We also show compatibility with the label mix-based approaches in evenly distributed scarce data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 534, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 689 + ], + "type": "text", + "content": "Learning models, e.g., neural networks, are known to perform well on visual recognition tasks when training and testing datasets present similar distributions [4]. However, their performance often degrades considerably when evaluated in subtly different distributions [69]. One effective way to enhance the generalization ability of a model against such data distribution shifts would be data augmentation [18, 86, 85, 44, 41, 73]. Augmenting data enlarges the support of the training distribution formed by given samples and yields the effect of increasing the amount of data even without additional laborious data collection. By training on augmented data, decision boundaries are smoothed, and the generalization ability of the model is improved [73]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "There has been a distinctive and successful line of research for label mix-based data augmentation, such as" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 312, + 256, + 547, + 345 + ], + "blocks": [ + { + "bbox": [ + 312, + 256, + 547, + 345 + ], + "lines": [ + { + "bbox": [ + 312, + 256, + 547, + 345 + ], + "spans": [ + { + "bbox": [ + 312, + 256, + 547, + 345 + ], + "type": "image", + "image_path": "77558d45f938bf35ae6f8307b24cc555b6a5a3f130c0e3d77f83544e7a98931e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 354, + 548, + 441 + ], + "lines": [ + { + "bbox": [ + 304, + 354, + 548, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 548, + 441 + ], + "type": "text", + "content": "Figure 1. Illustration of TextMania. Our method augments the target visual feature by leveraging text embedding of the visually mimetic words, which are comprehensible and semantically rich. For example, when the text of the existing class \"bull\" is manipulated as \"red bull\" by adding the attribute \"red,\" we can get augmented visual features by reflecting the difference of text embeddings. In this way, TextMania densifies sparse visual feature space using various attributes text." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 470, + 548, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 548, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 548, + 674 + ], + "type": "text", + "content": "Mixup [86], CutMix [85], and manifold Mixup [73], which are effective for model generalization and calibration [25]. The effectiveness of those label mix-based approaches is attributed to semantic perturbation by label mixing [86, 73, 85]. This is a distinctive property from other lines of data augmentation methods, e.g., [74, 44, 41, 64, 13], where they synthesize diverse virtual data that appear differently but retain class semantics of original contents. However, we found that the performance of mix-based augmentation methods is noticeably degraded when training with skewed class distribution having scarce samples for non-major classes, i.e., long-tailed distribution. In real-world, data often exhibit long-tailed class distribution (e.g., Pareto distribution), which cannot be dealt with the prevalent mix-based approaches. This motivates us to seek a semantically rich data augmentation effective for limited data regimes, including long-tailed distribution, scarce data, and few-shot cases." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "In this work, we propose TextMania, a text-driven manifold augmentation for visual features, which is effective for long-tail classes and scarce data. Our TextMania is based" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2526" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "content": "on an interesting hypothesis that general language models, e.g., BERT [17] and GPT [55], have learned visual information to some extent that can be transferred to visual feature spaces even with no visual training data. With this hypothesis, we semantically enrich the target visual feature space to be trained by leveraging visually mimetic texts, encoded with general language models and transferred to the target space. Specifically, TextMania encodes meaningful attributes such as \"red\" and \"large\" to vectors by computing the difference between text embeddings with and without attributes. We add the attribute embeddings to target visual features to mimic those attributes on the target visual feature space. Figure 1 illustrates the augmentation process of TextMania. The input feature (e.g., the visual feature of \"bull\") is manipulated by adding the attribute vector induced by the attribute text (e.g., \"red\"), which yields the augmented visual feature (e.g., \"red bull\"). Thanks to the text modality properties, the augmentations generated by TextMania are symbolic, human-interpretable, and easily controllable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 504 + ], + "type": "text", + "content": "Our approach applies semantic perturbation on a different level to that of the label mix-based methods [86, 73, 85]. The mix-based methods augment a sample from a combination of two different class samples, i.e., applying semantic perturbation in an inter-class way. This further aggravates the class imbalance problem in the long-tailed (skewed) class distribution cases. Our TextManiA, whereas, perturbs data in an intra-class way. A sample per each class is selected, and we enrich the semantic granularity of the class using the sample, thus enabling us to better maintain the amount of augmentation balances in the long-tailed class distribution cases. Moreover, TextManiA can densify around the training samples by extrapolating the class semantics along augmented semantic attribute axes. With this, our method can be combined with the label mix-based methods to further improve performance in evenly distributed sparse data cases because they are complementary." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 505, + 289, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 289, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 289, + 673 + ], + "type": "text", + "content": "To empirically support that our attribute vectors transformed from text embeddings are reasonably designed, we devise two visualization-based analyses: with t-SNE [72] and a latent inversion technique. These demonstrate that attribute vectors lead to visually interpretable manifold augmentation of input. We also evaluate our method with two different tasks in scarce data regimes: few-shot object detection and image classification with deficient datasets and long-tail datasets. Our experiments demonstrate that TextManiA is an effective and model-agnostic data augmentation method, especially in scarce data cases, by exploiting the favors of zero-shot attributes. Also, additional studies show the versatility and compatibility of the design of TextManiA. Our key contributions are summarized as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 72, + 547, + 224 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 72, + 547, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 108 + ], + "type": "text", + "content": "- We propose TextMania, which enriches the visual features by conveying attribute information from the text embedding to the target visual feature space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 111, + 547, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 111, + 547, + 147 + ], + "spans": [ + { + "bbox": [ + 306, + 111, + 547, + 147 + ], + "type": "text", + "content": "- We validate our hypothesis of the existence of embedded visual knowledge in pre-trained language encoders despite no training on visual data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 149, + 545, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 149, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 306, + 149, + 545, + 174 + ], + "type": "text", + "content": "- We demonstrate that TextManiA is especially helpful in augmenting sparse samples in long-tail class cases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 177, + 547, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 177, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 177, + 547, + 224 + ], + "type": "text", + "content": "- We show that our TextMania is complementary to other augmentation methods, and in particular, the combination of our TextMania and manifold Mixup [73] noticeably improves the performance in deficient data cases." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 236, + 392, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 392, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 392, + 248 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 255, + 547, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 339 + ], + "type": "text", + "content": "We brief the related work in the following three perspectives: image data augmentation, foundation models, and target application tasks. In this work, our TextMania augments data by leveraging the text encoder of CLIP [54], BERT [17], or GPT-2² [55]. For main target applications, we focus on long-tail and small data classification and few-shot object detection tasks in the data-scarce regimes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 340, + 548, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 548, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 548, + 472 + ], + "type": "text", + "content": "Image Data Augmentation. Image data augmentation can be largely divided into whether semantic perturbation exists. Semantic perturbation, in specific, can be further split into methods with or without label mixing. Methods [18, 66, 57, 2, 24, 5, 30, 74, 44, 1, 41, 64, 13, 12] without semantic perturbation, which have no label change, contain primitive image processing and transformation operations. This includes photometric (e.g., color jitter, contrast, blur, noise, etc.) and geometric (e.g., horizontal reflection, rotation, etc.) operations, and advanced augmentations, including Cutout [18] and adaptive combinations [13, 12]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 472, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 674 + ], + "type": "text", + "content": "In contrast, Mixup [86], CutMix [85], and manifold Mixup [73] execute semantic perturbation along with label mixing. Mixup interpolates two whole input images pixelwisely, CutMix interpolates a partial region of an image with another, and manifold Mixup mixes features from the images. These mix-based methods also augment labels of samples by an inter-class semantic perturbation, where labels of two different class samples are mixed. While the mixed label is known to be effective for generalization and model calibration effects [25], we found that the mix-based methods are heavily affected by class distribution due to sampling from two sources; thus, their effect is restricted to evenly distributed datasets. For datasets with skewed class distributions with tails, the sampling probabilities between major and minor classes would significantly differ, which can exaggerate biased sampling to major classes and makes minor classes more minor." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 683, + 288, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 288, + 715 + ], + "type": "text", + "content": "For example, if data size of major classes is 10 times larger than minor classes, the probability of choosing a pair of source samples from the major classes is approximately 100 times more than that of minor classes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "text", + "content": "GTP-2 is the decoder-only architecture, but we use it as a text embedding extractor, so we call it text-encoder." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2527" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": "Our TextManiA, on the other hand, is applied to all of the given samples uniformly regardless of class distribution. TextManiA densifies around the sample features by perturbing and enriching the semantic meaning of them at an intra-class level, which does not change the label. Moreover, because of the different semantic granularity of perturbation between TextManiA (intra-class) and mix-based methods (inter-class), two methods can be used complementarily when class imbalance does not exist." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 290, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 290, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 290, + 337 + ], + "type": "text", + "content": "Foundation Models. Recent foundation models [84, 54, 43, 33, 56, 17, 55] have shown a successful case of reflecting human nuances with visually imitated word composition. Particularly, language models, e.g., BERT [17] and GPT [55], show their ability not only in language tasks [78] but also in vision-language multi-modal tasks [62, 22]. Contrastive Language-Image Pretraining (CLIP) [54] also achieves huge success in various tasks even in zero-shot recognition. Follow-up studies show that CLIP representation is effective in conducting other visual tasks by bridging vision and language, e.g., 2D image generation [23, 39, 35, 49], image manipulation [51, 35] and synthesis [21], and even 3D domain tasks [83, 32, 47]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 339, + 290, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 290, + 483 + ], + "type": "text", + "content": "In TextMania, we focus on estimating attribute features by exploiting BERT, GPT-2, or CLIP text encoder alone. Distinctively, we only transfer the estimated attribute feature to augment visual features in a different space, which makes our work different from knowledge distillation [29] of foundation models [15, 76, 63]. Rather, our design is an instance of the module neural network structure [27, 3], where recent module-based designs procedurally train the whole model module-by-module with the guidance of the well pretrained module, e.g., [38, 60, 50, 68, 26]. Also, our work is applicable agnostically to architectures; thus, more flexibly applicable than fine-tuning of foundation models [77]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 484, + 290, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 290, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 290, + 628 + ], + "type": "text", + "content": "Long-tail Classification. In real world, visual data follow a long-tailed distribution which induces class imbalance and leads to performance degrading [81]. A representative line of the methods for long-tail classification is rebalancing [7, 14, 58], which resamples data or reweights the loss for tail classes. However, improvement in performance of the tail classes comes with the sacrifice of head class performance. Note that TextMania densifies all the given samples regardless of the class imbalance, and whereby the model is trained with reasonable variations of training samples for every class at least, which improves the performance while minimizing sacrifice of the head class." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 290, + 715 + ], + "type": "text", + "content": "Few-Shot Object Detection (FSOD). We tackle FSOD, one of the sparse sample problems, to demonstrate the effectiveness of TextManIA and its model architecture agnostic property. FSOD handles novel object classes after the base training for object detection tasks. The model rapidly adapts to novel classes using few data by matching-based [42, 9, 79] or fine-tuning based [31, 67, 75, 53, 82]" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 537, + 316 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 537, + 316 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 537, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 537, + 316 + ], + "type": "image", + "image_path": "14fc0c522639b723a5dcb4df77a05f91e74c4c4f7b25f1bc665e81cb2b58631b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "lines": [ + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "text", + "content": "Figure 2. Overview of TextMania. Given flower image " + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_0" + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "text", + "content": " and class \"flower\" " + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "text", + "content": ", we construct the variant text " + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "text", + "content": " by adding the attribute \"red\" on " + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "text", + "content": ". Embeddings of " + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "text", + "content": " are computed with text encoders, e.g. CLIP [54], BERT [17], or GPT-2 [55] and their difference vector, " + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "inline_equation", + "content": "\\Delta_{0\\rightarrow 1} = \\mathbf{e}_{T_1} - \\mathbf{e}_{T_0}" + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "text", + "content": " is added to the image feature " + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{I_0}" + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "text", + "content": " after projection " + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "inline_equation", + "content": "\\mathrm{proj}(\\cdot)" + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "text", + "content": " and weight " + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 321, + 547, + 410 + ], + "type": "text", + "content": ". We make the target feature space semantically rich and plausible by adding the difference vector, which embeds interpretable information." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 430, + 547, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 465 + ], + "type": "text", + "content": "methods. TextManiA is evaluated with the fine-tuning-based FSOD approach [75], which facilitates to use general model architectures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 476, + 380, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 380, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 380, + 488 + ], + "type": "text", + "content": "3. TextMania" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 640 + ], + "type": "text", + "content": "In image classification, the class label is typically utilized only as a supervision for measuring the loss. We, instead, propose to treat the class label as additional information, the text describing the class, and derive semantic information from it. However, class label as a text description itself is too coarse to represent rich semantics within a class. For example, a class label \"dog\" does not represent all the details of the description such as \"small size of the brown colored dog.\" To enrich the detailed semantics over the given coarse class texts, we leverage the attribute words, such as \"small size\" and \"brown colored,\" that can visually modify objects in images at the semantic level." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 647, + 378, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 378, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 378, + 658 + ], + "type": "text", + "content": "3.1. Main Idea" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 666, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 548, + 715 + ], + "type": "text", + "content": "The main idea of TextManiA is to densify distribution around sparse training samples on the target feature space, making it semantically rich through the difference vectors having plausible attribute information, as depicted in Fig. 1." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2528" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "Figure 2 illustrates how TextMania augment data. Suppose we have an image " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_0" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " and corresponding class label " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": ". The model generally learns the target task using the image " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_0" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " as an input and the class label " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " as supervision. In this work, we also consider the class label " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " as text information and extract the embedding vector " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{T_0} \\in \\mathbb{R}^{d_c}" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " using text encoder, e.g., CLIP [54], BERT [17], or GPT-2 [55], where " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "d_c" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " is the text embedding dimension. For obtaining an embedding vector " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{T_0}" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": ", we use the text embedding of the encoder output directly when using CLIP text encoder, or use the average vector of all the embeddings of the sentence when using other language models such as BERT or GPT-2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "text", + "content": "Specifically, text input " + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "text", + "content": " is formed with class name and pre-defined prompts, such as \"a photo of,\" \"a picture of,\" and \"a sketch of.\" We also synthesize another text input variant " + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "inline_equation", + "content": "T_1" + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "text", + "content": " by adding color or size attribute words, such as \"red\" and \"big,\" and compute the embedding vector " + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{T_1} \\in \\mathbb{R}^{d_c}" + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "text", + "content": ". Numerous variants can be created with various attribute words and their combinations, but we explain the case of one variant for convenience. Based on the word vector analogy [48], we hypothesize that the relationship between " + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "inline_equation", + "content": "T_1" + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "text", + "content": " is maintained in the text embedding space, i.e., the difference vector " + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\Delta_{0 \\to 1} = \\mathbf{e}_{T_1} - \\mathbf{e}_{T_0}" + }, + { + "bbox": [ + 46, + 216, + 289, + 384 + ], + "type": "text", + "content": " would contain the information of added attributes (this hypothesis is validated in Sec. 3.2). To exploit the difference vector from text embeddings, we design our method on the manifold." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 384, + 289, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 289, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 289, + 527 + ], + "type": "text", + "content": "We can obtain such diverse attribute vectors from various attribute text templates; however, their representation space is not directly related to the visual feature space of the target model we are interested in. To bridge the gap, we project the difference vector to the target feature space with a learnable linear projection layer " + }, + { + "bbox": [ + 46, + 384, + 289, + 527 + ], + "type": "inline_equation", + "content": "\\operatorname{proj}(\\cdot)" + }, + { + "bbox": [ + 46, + 384, + 289, + 527 + ], + "type": "text", + "content": ". Then, we add the projected difference vector to the target image feature " + }, + { + "bbox": [ + 46, + 384, + 289, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{I_0} \\in \\mathbb{R}^{d_t}" + }, + { + "bbox": [ + 46, + 384, + 289, + 527 + ], + "type": "text", + "content": " obtained from the target task encoder with the input image " + }, + { + "bbox": [ + 46, + 384, + 289, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_0" + }, + { + "bbox": [ + 46, + 384, + 289, + 527 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 384, + 289, + 527 + ], + "type": "inline_equation", + "content": "d_t" + }, + { + "bbox": [ + 46, + 384, + 289, + 527 + ], + "type": "text", + "content": " is the target feature space dimension. A linear layer for " + }, + { + "bbox": [ + 46, + 384, + 289, + 527 + ], + "type": "inline_equation", + "content": "\\operatorname{proj}(\\cdot)" + }, + { + "bbox": [ + 46, + 384, + 289, + 527 + ], + "type": "text", + "content": " would be sufficient to transfer cross-modal information, referring to the cross-modal transferability under the contrastive learning case [87] and our experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 527, + 289, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 289, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 289, + 574 + ], + "type": "text", + "content": "To inject the stochasticity, a mixing weight " + }, + { + "bbox": [ + 46, + 527, + 289, + 574 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 527, + 289, + 574 + ], + "type": "text", + "content": " is introduced and randomly sampled from the clamped Normal distribution in the range over 0.1. Then, we have the augmented feature vector " + }, + { + "bbox": [ + 46, + 527, + 289, + 574 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_{I_0}" + }, + { + "bbox": [ + 46, + 527, + 289, + 574 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 582, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 582, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 287, + 597 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {f}} _ {I _ {0}} = \\mathbf {f} _ {I _ {0}} + \\alpha \\cdot \\operatorname {p r o j} \\left(\\boldsymbol {\\Delta} _ {0 \\rightarrow 1}\\right). \\tag {1}", + "image_path": "f72882bc047c861611f8bb714eced60338aa871b8db16aca2b55c550980e93e4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "text", + "content": "For the cases having " + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "inline_equation", + "content": "d_{t} = d_{c}" + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "text", + "content": ", we can set " + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "inline_equation", + "content": "\\operatorname{proj}(\\cdot)" + }, + { + "bbox": [ + 46, + 603, + 287, + 627 + ], + "type": "text", + "content": " operation to be an identity mapping without any learnable parameter." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 627, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 288, + 687 + ], + "type": "text", + "content": "We train the target task model with this augmented feature vector, whose class label is still " + }, + { + "bbox": [ + 46, + 627, + 288, + 687 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 46, + 627, + 288, + 687 + ], + "type": "text", + "content": ". We note that computing difference attribute vectors with text encoder is computationally expensive. For efficient training, we pre-compute all possible combinations of difference vectors " + }, + { + "bbox": [ + 46, + 627, + 288, + 687 + ], + "type": "inline_equation", + "content": "\\{\\Delta\\}" + }, + { + "bbox": [ + 46, + 627, + 288, + 687 + ], + "type": "text", + "content": " and store" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 83, + 358, + 127 + ], + "blocks": [ + { + "bbox": [ + 307, + 83, + 358, + 127 + ], + "lines": [ + { + "bbox": [ + 307, + 83, + 358, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 83, + 358, + 127 + ], + "type": "image", + "image_path": "bc60c8e6a2bf4e2e451789ef5f7ba5c648ecc92f97d504671678e4824e1ed920.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 141, + 350, + 150 + ], + "lines": [ + { + "bbox": [ + 313, + 141, + 350, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 141, + 350, + 150 + ], + "type": "text", + "content": "Label Mix" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 375, + 72, + 449, + 135 + ], + "blocks": [ + { + "bbox": [ + 375, + 72, + 449, + 135 + ], + "lines": [ + { + "bbox": [ + 375, + 72, + 449, + 135 + ], + "spans": [ + { + "bbox": [ + 375, + 72, + 449, + 135 + ], + "type": "image", + "image_path": "1c5a82e344d30d839bf097d51a75a558511fce77c4ed6f1cfe5b59b212c9d4bf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 141, + 430, + 150 + ], + "lines": [ + { + "bbox": [ + 391, + 141, + 430, + 150 + ], + "spans": [ + { + "bbox": [ + 391, + 141, + 430, + 150 + ], + "type": "text", + "content": "TextMania" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 462, + 71, + 543, + 139 + ], + "blocks": [ + { + "bbox": [ + 462, + 71, + 543, + 139 + ], + "lines": [ + { + "bbox": [ + 462, + 71, + 543, + 139 + ], + "spans": [ + { + "bbox": [ + 462, + 71, + 543, + 139 + ], + "type": "image", + "image_path": "2e3e6bad334560fe657c439c8e3d91bcc6695576f1651cb24d5777fbcc9aa415.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 141, + 544, + 150 + ], + "lines": [ + { + "bbox": [ + 463, + 141, + 544, + 150 + ], + "spans": [ + { + "bbox": [ + 463, + 141, + 544, + 150 + ], + "type": "text", + "content": "TextManiA + Label Mix" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 172, + 547, + 261 + ], + "lines": [ + { + "bbox": [ + 304, + 172, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 547, + 261 + ], + "type": "text", + "content": "Figure 3. Comparison of a typical label mix-based augmentation, TextManiA, and their combination, in a sparse data case. Given two samples for two classes, the augmented samples by a label mix-based method are at points lying on six lines resulting from the combinations of samples, i.e., inter-class perturbation. TextManiA densifies the sample distribution along semantic attribute axes in an intra-class. The combination of them yields synergy of their respective advantages." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 281, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 304 + ], + "type": "text", + "content": "them in a look-up table because class names and attributes can be pre-determined and unchanged during training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 304, + 546, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 546, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 546, + 388 + ], + "type": "text", + "content": "Different from knowledge distillation [15, 76, 63, 29, 36], TextManiA does not transfer-learn the text embeddings directly. Instead, the difference vector projected onto the target domain is injected into the target model, allowing our method to be applied to arbitrary target models. Since the visual feature augmentation is solely controlled by text, TextManiA is human-interpretable and easily controllable." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 388, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 555 + ], + "type": "text", + "content": "Compared to label mix-based augmentations [86, 73, 85], TextManiA has advantages in imbalanced data distribution. We suppose a scenario where few samples are in one class and many samples are in another class. The augmented points by a mix-based method would be located only on the interpolation lines between the given samples, which limits the augmentation effects, as depicted in Fig. 3. If we apply a mix-based method in the long-tailed class distribution cases, i.e., notably skewed distribution, the class imbalance problem is further aggravated, and augmentation is more biased toward major classes. In contrast, TextManiA can equally densify all the given samples since it augments each sample independently. Thus, TextManiA can be used in general regardless of the imbalance factor of class distribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 556, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 628 + ], + "type": "text", + "content": "On the other hand, in another scenario with small training data but with uniform class distribution, both TextMania and mix-based methods would increase diverse combinations of samples by augmentation in respective aspects, which leads to complementary performance improvement. This will be empirically demonstrated in Sec. 4." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 635, + 515, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 515, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 515, + 647 + ], + "type": "text", + "content": "3.2. Characteristics of Attribute Embedding" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "To scrutinize the relationship between the text " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", text variant " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " and the attribute embedding " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Delta_{0\\rightarrow 1}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", we visualize their distribution and discuss the characteristics. We also visualize the difference vector to verify the hypothesis that the difference vector embeds its corresponding attribute." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "3It was shown that simple algebraic operations can be performed on the word vectors, e.g., king - man + woman ≈ queen on the embedding space." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2529" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 287, + 220 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 287, + 220 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 287, + 220 + ], + "type": "image", + "image_path": "d7d4dbd599ac5c5f805e51e6bb53a90cd7c45c828819bbef8f9ea9ded9b15a26.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 226, + 288, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 226, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 288, + 293 + ], + "type": "text", + "content": "Figure 4. The t-SNE plot of difference vectors (e.g., \"brown dog\" - \"dog\") projected to visual feature space. The colors of the points represent color attributes used for computing the difference vector, and we use all the classes in CIFAR-100 for this plot. As a comparison, the colored points in the red circle show direct color-text embedding (e.g., \"brown\") projected to the visual feature space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 288, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 519 + ], + "type": "text", + "content": "Embedding Difference vs. Direct Text Embedding. When guessing the difference between two texts, e.g., \"brown X\" - \"X,\" it would be \"brown.\" Someone may think of using the text embedding directly obtained from \"brown\" instead of our attribute embedding from \"brown X\" - \"X.\" To understand the difference between the two representations, we visualize the difference vectors and text embeddings with BERT and CLIP text encoder in Fig. 4. While the direct text embeddings in the red circle of Fig. 4 are clustered no matter with different color-texts, the difference vectors are well clustered dependent on the color. This observation indicates that the difference vector is more effective in augmenting the visual feature space than text embedding. In addition, the difference vectors obtained from the same attribute word are similarly clustered regardless of the class \"X\" but slightly different. It may imply our attribute embedding has subtle difference awareness on granularity according to class." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 628 + ], + "type": "text", + "content": "Note that Fig. 4 presents difference vectors in the visual feature space, and we also observe similar distributions of difference vectors in the original text embedding space, i.e., t-SNE is not employed for clustering the direct text embeddings. This observation supports our hypothesis that general language models, e.g., BERT or GPT, have learned visual information to some extent. It, also, demonstrates the visual information is properly transferred to the target visual feature space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Do We Need to Rule out Unrealistic Attributes? One can be curious about how TextMania handles the unrealistic attribute, such as \"blue cow.\" We intentionally include such unrealistic attributes, motivated by other contexts in self- and semi-supervised learning [10, 65, 8], where they showed the strong benefit of unnatural strong augmentations to train neural networks. This observation regarding strong" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 426, + 190 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 426, + 190 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 426, + 190 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 426, + 190 + ], + "type": "image", + "image_path": "0f96afdd265fdda275adad38210782c4f070163ff9970d0b39bfea273e97afc8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 194, + 547, + 283 + ], + "lines": [ + { + "bbox": [ + 304, + 194, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 547, + 283 + ], + "type": "text", + "content": "Figure 5. The attribute embedding visualization through image manipulation examples. We analyze how an image is manipulated when a difference vector, containing specific attribute information, is injected to the original image feature. (Top) We visualize an example of generated image given a specific class and its manipulated pair by size and color attribute, respectively. (Bottom) From left to right, we visualize the real image of \"Butterfly\" and its manipulated image pair with gigantic, green, and dot pattern, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 426, + 71, + 545, + 190 + ], + "blocks": [ + { + "bbox": [ + 426, + 71, + 545, + 190 + ], + "lines": [ + { + "bbox": [ + 426, + 71, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 426, + 71, + 545, + 190 + ], + "type": "image", + "image_path": "0df78f704b5fea91b93d95831053246926a44e560afdf522bb7d6e932f8d1bc4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 304, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 327 + ], + "type": "text", + "content": "augmentation is consistent with the design of TextMania containing unrealistic attributes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "text", + "content": "Does Difference Vector Embed Attribute? To visually understand whether attribute editing is reflected while maintaining class information, we attempt to manipulate images by changing their features with the difference vectors " + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "inline_equation", + "content": "\\Delta_{0\\rightarrow 1} = \\mathbf{e}_{T_1} - \\mathbf{e}_{T_0}" + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "text", + "content": ", i.e., we want to visualize the change effect between " + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{I}_0}" + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{a_1} = \\mathbf{f}_{\\mathbf{I}_0} + \\alpha \\Delta_{0\\rightarrow 1}" + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "text", + "content": " in image domain. To see the effect in image domain, we need to invert the change from " + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathbf{I}_0}" + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{a_1}" + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "text", + "content": " in image domain, which can be formulated as the following optimization problem," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 370, + 445, + 545, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 445, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 370, + 445, + 545, + 458 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\mathbf {I}} \\| E _ {i} (\\mathbf {I}) - \\mathbf {e} _ {a _ {1}} \\| _ {1}, \\tag {2}", + "image_path": "8de6345df10998b42ee8a0f0c161b20cec17afc94879b230f05cf260e8cf0e75.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 466, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 547, + 538 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 466, + 547, + 538 + ], + "type": "inline_equation", + "content": "E_{i}(\\cdot):\\mathbf{I}\\rightarrow \\mathbf{f}" + }, + { + "bbox": [ + 304, + 466, + 547, + 538 + ], + "type": "text", + "content": " denotes the image encoder in Fig. 2. Direct optimization in Eq. (2) is known to be difficult [88]; thus, we parameterize a given image with an image generator " + }, + { + "bbox": [ + 304, + 466, + 547, + 538 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 304, + 466, + 547, + 538 + ], + "type": "text", + "content": " with a latent code " + }, + { + "bbox": [ + 304, + 466, + 547, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 466, + 547, + 538 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 466, + 547, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(\\theta) = G_{\\theta}(\\mathbf{z})" + }, + { + "bbox": [ + 304, + 466, + 547, + 538 + ], + "type": "text", + "content": ", which is known to ease the optimization [71]. Then, we can obtain the visualization by the following optimization over " + }, + { + "bbox": [ + 304, + 466, + 547, + 538 + ], + "type": "inline_equation", + "content": "\\theta" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 548, + 545, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 548, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 359, + 548, + 545, + 561 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\theta} \\| E _ {i} \\left(G _ {\\theta} (\\mathbf {z})\\right) - \\mathbf {e} _ {a _ {1}} \\| _ {1}, \\tag {3}", + "image_path": "b02abc4239068ec8df003ad91d30bc378aaeeb82e79ccb31adda292df7d234af.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "E_{i}(\\cdot)" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{a_1}" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " are frozen during the optimization. Since our goal is to see the move from " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{I}_0} = E_i(\\mathbf{I}_0)" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{a_1}" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " for the query " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_0" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": ", we initialize " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "G_{\\theta}(\\mathbf{z}) \\simeq \\mathbf{I}_0" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " by the GAN inversion [88]. We use IC-GAN [52] for the image generator and the text embeddings are obtained from the CLIP text encoders. Note that " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "E_{i}(\\cdot)" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " is trained with random perturbation of transformed attributes; thus, the visualization through " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "E_{i}(\\cdot)" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " is different from that of CLIP encoders. Details can be found in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Figure 5 shows that the manipulated image reflects the added attribute, i.e., the size of the dog is reduced by the size attribute \"small,\" and the bird becomes yellow by injecting" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2530" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "the color attribute \"yellow.\" The manipulated results imply that 1) the difference vector indeed embeds the attributes while preserving its semantics, and 2) our augmentation on the feature space may have analogous effects to an image-level augmentation but without implementing complicated image perturbation operations. Note that these visualizations are for analysis purposes but not for competing with any existing image manipulation methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 178, + 128, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 128, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 128, + 192 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 288, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 288, + 294 + ], + "type": "text", + "content": "We evaluate TextMania in various cases presenting sparse data with different tasks: long-tail classification in Sec. 4.1, evenly distributed scarce data classification in Sec. 4.2, and few-shot object detection in Sec. 4.3. We also conduct additional studies demonstrating the effectiveness of the design of our method and the versatility of TextMania in Sec. 4.4. Additional experimental results and details can be found in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 300, + 178, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 178, + 313 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 178, + 313 + ], + "type": "text", + "content": "4.1. Long-tail Classification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 415 + ], + "type": "text", + "content": "Experimental Setting. We compare TextMania with the mix-based augmentations on the CIFAR-100-LT [14] and ImageNet-LT [46] datasets, where LT stands for long-tailed distribution. They are artificially truncated to have a long-tail from each original dataset, CIFAR-100 [37] and ImageNet2012 [16]. Long-tail datasets usually have three sets of classes: Many-shot (more than 100 images), Medium-shot (20-100 images), and Few-shot (less than 20 images)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 415, + 288, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 288, + 498 + ], + "type": "text", + "content": "For CIFAR-100-LT, we control the imbalance factor (IF) [11] computed as the ratio of samples in the head to tail class, " + }, + { + "bbox": [ + 46, + 415, + 288, + 498 + ], + "type": "inline_equation", + "content": "N_{1} / N_{K}" + }, + { + "bbox": [ + 46, + 415, + 288, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 415, + 288, + 498 + ], + "type": "inline_equation", + "content": "N_{k} = |\\mathcal{D}_{k}|" + }, + { + "bbox": [ + 46, + 415, + 288, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 415, + 288, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k" + }, + { + "bbox": [ + 46, + 415, + 288, + 498 + ], + "type": "text", + "content": " is the set of samples belonging to the class " + }, + { + "bbox": [ + 46, + 415, + 288, + 498 + ], + "type": "inline_equation", + "content": "k\\in \\{1,\\dots ,K\\}" + }, + { + "bbox": [ + 46, + 415, + 288, + 498 + ], + "type": "text", + "content": ". A larger value of the IF represents a more severe imbalance in data, which is more challenging. We evaluate the performance according to different IFs of 100, 50, and 10." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 557 + ], + "type": "text", + "content": "We utilize ResNet18 as the baseline on CIFAR-100-LT and ResNext50 on ImageNet-LT. We use the validation set of the original datasets to measure the Top-1 accuracy. Note that we apply each augmentation on all the samples without carefully selecting a set of classes in Table 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": "Results. Table 1 presents the long-tail classification results on CIFAR-100-LT, which show consistent improvement with TextManiA. Also, TextManiA with various text encoders achieves analogous improvement trend regardless of the imbalance factor but marginal degradation on Many class of IF=100 when using general language model, BERT and GPT-2. While the performance gain is from leaking pretrained language information, it is surprising and a virtue that the language models never exposed to any image can improve the visual recognition performance. In comparison to single usage of mix-based augmentations, our method shows higher accuracy because of uniform effects of TextManiA on samples regardless of class imbalance. The mix-based" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 544, + 232 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 544, + 232 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 544, + 232 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 544, + 232 + ], + "type": "table", + "html": "
(a) AugmentationImbalance Factor (IF)
1005010
Baseline38.3943.3359.29
TextMania (CLIP)40.65 (+2.26)46.48 (+3.15)60.17 (+0.88)
TextMania (BERT)41.10 (+2.71)47.17 (+3.84)60.67 (+1.38)
TextMania (GPT-2)41.20 (+2.81)46.93 (+3.60)60.94 (+1.65)
Cutout [18]37.5142.2859.26
+ TextMania40.35 (+2.84)45.48 (+3.20)61.31 (+2.05)
Cutmix [85]37.9343.3459.30
+ TextMania40.22 (+2.29)45.36 (+2.02)61.30 (+2.00)
Mixup [86]36.7540.7757.50
+ TextMania38.40 (+1.65)43.33 (+2.56)59.80 (+2.30)
ManiMixup [73]35.7240.5155.26
+ TextMania38.60 (+2.88)43.22 (+2.71)59.35 (+4.09)
", + "image_path": "ea13a965ad005c108d25030166264b4c9a33351a2234946a63df2f874f2855c9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 234, + 544, + 393 + ], + "blocks": [ + { + "bbox": [ + 308, + 234, + 544, + 393 + ], + "lines": [ + { + "bbox": [ + 308, + 234, + 544, + 393 + ], + "spans": [ + { + "bbox": [ + 308, + 234, + 544, + 393 + ], + "type": "table", + "html": "
(b) AugmentationSet of Classes (IF=100)
ManyMediumFew
Baseline71.1138.423.00
TextMania (CLIP)71.14 (+0.03)40.28 (+1.86)7.53 (+4.53)
TextMania (BERT)70.22 (-0.89)40.73 (+2.31)9.41 (+6.41)
TextMania (GPT-2)70.60 (-0.51)40.61 (+2.19)9.93 (+6.93)
Cutout71.5435.941.06
+ TextMania71.94 (+0.83)40.97 (+2.55)4.03 (+3.03)
Cutmix72.0237.170.90
+ TextMania72.37 (+0.35)40.80 (+3.63)3.90 (+3.00)
Mixup71.9733.620.36
+ TextMania71.97 (+0.00)36.77 (+3.15)1.83 (+1.47)
ManiMixup72.9729.510.70
+ TextMania73.20 (+0.23)36.80 (+7.29)0.76 (+0.06)
", + "image_path": "6523db0eef49408cc14f792da49abb4376f83d8848cfd01c380cd6973d2e0692.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 475, + 544, + 547 + ], + "blocks": [ + { + "bbox": [ + 304, + 397, + 547, + 464 + ], + "lines": [ + { + "bbox": [ + 304, + 397, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 547, + 464 + ], + "type": "text", + "content": "Table 1. Long-tail classification results (\\%) on CIFAR-100-LT with ResNet18. (a) The accuracy with respect to the different imbalance factors, i.e., IF=\\{100, 50, 10\\}. (b) The accuracy of each class set with IF=100. Baseline contains random horizontal flip, random crop and rotation, and normalization, applied in all experiments. TextMania without parenthesis uses CLIP for the text encoder." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 475, + 544, + 547 + ], + "lines": [ + { + "bbox": [ + 306, + 475, + 544, + 547 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 544, + 547 + ], + "type": "table", + "html": "
Aug.CBSAllManyMediumFew
Baseline38.3971.1138.423.00
Cutmix✓38.2371.7737.791.90
Mixup✓38.7371.6037.643.16
ManiMixup✓38.5671.2537.882.80
TextMania40.6571.1440.287.53
", + "image_path": "1dc19a112f3e178e1de1d029236b1ec79bbc31c02089cbcb973cf40038049d92.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 550, + 547, + 584 + ], + "lines": [ + { + "bbox": [ + 304, + 550, + 547, + 584 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 547, + 584 + ], + "type": "text", + "content": "Table 2. Comparison to label mix-based augmentations with classbalanced sampling (CBS) on CIFAR-100-LT with IF=100. CBS samples two classes first and then samples data in each classes." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 676 + ], + "type": "text", + "content": "methods, on the other hand, sample two data points from the total dataset, where the probability that a tail class sample contributes to a resulting augmented sample is very low. Even with class-balanced sampling on mixed-based augmentation in Table 2, TextMania performs better, further demonstrating our effectiveness." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Particularly in Table 1(b), the mix-based methods have degraded performance in the Medium and Few-shot classes, while our TextManiA improves performance. Combining" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2531" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 224 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 224 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 224 + ], + "type": "table", + "html": "
AugmentationSet of ClassesTotal
ManyMediumFew
Baseline85.3470.4742.8072.24
TextMania85.4071.7548.4973.68
Cutout [18]85.0270.3242.9172.07
+ TextMania85.3371.7048.5473.65
Cutmix [85]84.8569.9035.8270.77
+ TextMania85.3071.9347.0473.52
Mixup [86]84.9670.2740.2071.63
+ TextMania84.5569.9535.7270.66
ManiMixup [73]84.8469.9738.8371.24
+ TextMania85.4271.9846.6573.54
", + "image_path": "6f4e5f996a15556d1e8002b8e5818c815b07aec4dfa054de184bd865de5b7ff9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 293, + 286, + 343 + ], + "blocks": [ + { + "bbox": [ + 46, + 228, + 288, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 288, + 275 + ], + "type": "text", + "content": "Table 3. Long-tail classification results (\\%) on ImageNet-LT with ViT, and color the value as best and second best. Baseline contains random horizontal flip, random resize crop, color jitter, and normalization, applied in all experiments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 293, + 286, + 343 + ], + "lines": [ + { + "bbox": [ + 48, + 293, + 286, + 343 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 286, + 343 + ], + "type": "table", + "html": "
MethodManyMediumFewAll
LWS [34]63.3448.0827.1951.14
cRT [34]61.8046.2027.4049.60
cRT+TextManiA62.7448.6029.6751.47
", + "image_path": "a08592c8349aacf8bc7fc21c6bc5512feaad0ace90c9bba57762dba6676e944a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "type": "text", + "content": "the mix-based methods with TextMania improves overall performance, but the tendency to sacrifice the Medium and Few-shot classes is the same as before combining. Additionally, while Cutout has performance degradation due to the information loss [85], it is not affected by skewness due to no mix between inter-classes; thus, the performance is higher than the mix-based one in the long-tailed distribution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 506, + 288, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 288, + 613 + ], + "type": "text", + "content": "We also evaluate our TextMania on the large-scale dataset ImageNet-LT. In Table 3, the best and second best results are with TextMania, which demonstrate that our augmentation method is also effective in the large-scale long-tailed data distribution, consistent with the CIFAR-100-LT results in Table 1. The improvement with TextMania implies the importance of intra-class perturbation, which can uniformly affect the samples regardless of the skewness of the class distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "In Table 4, we compare with LWS [34], cRT [34], and TextMania on cRT. LWS and cRT are one of effective methods in recent long-tailed recognition. The result shows that TextMania on cRT achieves the best results compared to the counterparts in all classes except for the Many class, wherefrom ours achieves second best. Overall, TextMania improves well-established works, e.g., LWS, and cRT, and it demonstrates the compatibility of our method." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 223 + ], + "blocks": [ + { + "bbox": [ + 46, + 346, + 287, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 381 + ], + "type": "text", + "content": "Table 4. Long-tail classification results " + }, + { + "bbox": [ + 46, + 346, + 287, + 381 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 346, + 287, + 381 + ], + "type": "text", + "content": " on ImageNet-LT with ResNext50. We compare with LWS, cRT, and TextManiA on cRT, and color the value as best and second best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 223 + ], + "type": "table", + "html": "
AugmentationAcc.
Top-1Top-5
Baseline31.1059.14
Cutout32.0360.53
Cutmix32.4361.04
Mixup32.7262.47
ManiMixup33.7463.29
TextMania34.52 (+3.42)65.74 (+6.60)
Cutout + TextMania33.91 (+2.81)61.58 (+2.44)
Cutmix + TextMania35.61 (+4.51)63.82 (+4.68)
Mixup + TextMania37.97 (+6.87)66.75 (+7.61)
ManiMixup + TextMania38.02 (+6.92)67.28 (+8.14)
", + "image_path": "09df1ff5ebbf78baf0c14a610326dc7f560aaedb2df8b76ac7aef207e907ff9f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 279, + 538, + 413 + ], + "blocks": [ + { + "bbox": [ + 304, + 226, + 546, + 272 + ], + "lines": [ + { + "bbox": [ + 304, + 226, + 546, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 226, + 546, + 272 + ], + "type": "text", + "content": "Table 5. Classification results (\\%) on CIFAR-100-10% with ResNet18. Baseline represents random horizontal flip, random crop, and normalization, basically applied in all experiments. The parentheses stand for the improvement compared to the Baseline." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 279, + 538, + 413 + ], + "lines": [ + { + "bbox": [ + 315, + 279, + 538, + 413 + ], + "spans": [ + { + "bbox": [ + 315, + 279, + 538, + 413 + ], + "type": "table", + "html": "
AugmentationAcc.
Top-1Top-5
Baseline65.3789.82
Cutout69.1791.12
Cutmix69.8291.76
Mixup67.5490.23
TextManiA70.81 (+5.44)92.37 (+2.55)
Cutout + TextManiA69.71 (+4.34)91.32 (+1.50)
Cutmix + TextManiA71.05 (+5.68)92.22 (+2.40)
Mixup + TextManiA70.56 (+5.19)91.58 (+1.76)
", + "image_path": "02ca0ab3382dabb3c5bde3e5331c10a9e9de55fdda4a7708b827ed171db266b1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 416, + 547, + 451 + ], + "lines": [ + { + "bbox": [ + 304, + 416, + 547, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 547, + 451 + ], + "type": "text", + "content": "Table 6. Classification results (\\%) on CIFAR-100-10% with VIT-Tiny. The configuration follows Table 5. The parentheses stands for the improvement compared to the Baseline." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 474, + 541, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 474, + 541, + 488 + ], + "spans": [ + { + "bbox": [ + 305, + 474, + 541, + 488 + ], + "type": "text", + "content": "4.2. Evenly Distributed Scarce Data Classification" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 495, + 547, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 614 + ], + "type": "text", + "content": "Experimental Setting. For evaluating the effectiveness of TextMania on the scarce dataset, we use " + }, + { + "bbox": [ + 304, + 495, + 547, + 614 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 495, + 547, + 614 + ], + "type": "text", + "content": " data of the CIFAR-100 [37] and Tiny/ImageNet [40] datasets, named CIFAR-100- " + }, + { + "bbox": [ + 304, + 495, + 547, + 614 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 495, + 547, + 614 + ], + "type": "text", + "content": " and Tiny/ImageNet- " + }, + { + "bbox": [ + 304, + 495, + 547, + 614 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 495, + 547, + 614 + ], + "type": "text", + "content": ", respectively. CIFAR-100 has 100 classes with 500 training images per class, but we only use randomly sampled 50 images per class. Tiny/ImageNet is a subset of ImageNet-1k [61] with 100k images and 200 classes, but we use 10k images (50 images per class) for simulating a small dataset. Note that the evaluation set is same with those of the original datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 616, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 547, + 714 + ], + "type": "text", + "content": "The baseline models of scarce data classification are ResNet18 [28] and ViT-Tiny [19]. Due to the space limit, details of training can be found in the supplementary material. Results. We demonstrate the effectiveness of TextMania compared to mix-based augmentations on evenly distributed scarce datasets. As in Table 5 for CIFAR-100-10%, TextMania outperforms other methods when a single augmentation is used. Furthermore, the effect is amplified when" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2532" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 286, + 224 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 224 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 224 + ], + "type": "table", + "html": "
AugmentationAcc.
Top-1Top-5
Baseline25.9450.53
Cutout26.4150.28
Cutmix25.9449.67
Mixup29.3454.10
ManiMixup28.4353.25
TextMania29.37 (+3.43)52.37 (+1.84)
Cutout + TextMania29.14 (+3.20)52.60 (+2.07)
Cutmix + TextMania29.86 (+3.92)54.31 (+3.78)
Mixup + TextMania31.15 (+5.21)56.71 (+6.18)
ManiMixup + TextMania32.39 (+6.35)58.25 (+7.72)
", + "image_path": "3e916db14f4bd6c462c0a202e802b337a5240544bb2a4a8e4db92c192c06011a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 281, + 289, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 289, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 289, + 402 + ], + "type": "text", + "content": "our method and mix-based methods are combined, with particularly good compatibility with Manifold Mixup. The results demonstrate the importance of intra-class semantic perturbation along with inter-class in scarce data settings. This tendency is also observed with another baseline architecture in Table 6, and datasets in Table 7, implying that TextMania is model-agnostic to be applied. The overall results demonstrate the potential of TextMania to enrich the visual feature space using text modalities and develop more accurate and robust models in scarce data regimes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 407, + 194, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 194, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 194, + 420 + ], + "type": "text", + "content": "4.3. Few-shot Object Detection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 288, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 594 + ], + "type": "text", + "content": "Experimental Setting. We evaluate TextMania on the PASCAL VOC [20] and MS-COCO [45] datasets with a few-shot division following Wang et al. [75]. For VOC, we have three random splits, which have different divisions into 15 base classes and 5 novel classes among the 20 total classes, and " + }, + { + "bbox": [ + 46, + 426, + 288, + 594 + ], + "type": "inline_equation", + "content": "K = 1, 2, 3, 5, 10" + }, + { + "bbox": [ + 46, + 426, + 288, + 594 + ], + "type": "text", + "content": " objects are sampled from the novel classes. We utilize the VOC2007 test set for evaluation with AP50 metrics and train with the combination of the VOC2007 and VOC2012 train/val set. For COCO, the base classes are disjoint with VOC classes while the remaining classes are used as novel classes, and " + }, + { + "bbox": [ + 46, + 426, + 288, + 594 + ], + "type": "inline_equation", + "content": "K = 1, 3, 5, 10, 30" + }, + { + "bbox": [ + 46, + 426, + 288, + 594 + ], + "type": "text", + "content": " objects are sampled from the novel classes for few-shot finetuning. We use 5k images from the validation set in COCO for evaluation with mAP metrics and the rest for training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 653 + ], + "type": "text", + "content": "The baseline [80] is the Faster R-CNN [59] trained with the base classes first and then fine-tuned with the novel classes. TextMania is applied to the novel class samples during the fine-tuning stage. Following the prior studies, all the reported results are averaged over 10 repeated runs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "Results. Note that we apply TextManiA only on the classification head; thus, the quality of the regressed bounding boxes will remain similar as before applying TextManiA. As shown in Table 8 for VOC and Table 9 for COCO, TextManiA improves the AP by improving only the clas" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 545, + 224 + ], + "blocks": [ + { + "bbox": [ + 46, + 228, + 288, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 288, + 262 + ], + "type": "text", + "content": "Table 7. Classification results on Tiny/ImageNet-10% with ResNet18. The configuration follows Table 5. The parentheses represent the improvement compared to the Baseline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 224 + ], + "type": "table", + "html": "
SplitAug.K-shot
123510
AllBaseline12.8216.6520.0420.6423.19
TextManiA17.74 (+4.92)22.40 (+5.75)23.37 (+3.33)25.09 (+4.45)24.22 (+1.03)
1Baseline15.1118.8222.6121.9723.74
TextManiA21.94 (+6.83)26.44 (+7.62)23.66 (+1.05)25.88 (+3.91)25.14 (+1.40)
2Baseline10.8614.2218.6719.3422.49
TextManiA14.64 (+3.78)18.49 (+4.27)23.28 (+4.61)23.06 (+3.72)24.44 (+1.95)
3Baseline12.4916.9018.8420.6123.35
TextManiA16.65 (+4.16)22.26 (+5.36)23.16 (+4.32)26.33 (+5.72)25.08 (+1.73)
", + "image_path": "edab6cafb3aaaf5a930a20d8172bff3ffff57c299dfbe58d2b2d63504afed035.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 273, + 545, + 335 + ], + "blocks": [ + { + "bbox": [ + 305, + 228, + 546, + 262 + ], + "lines": [ + { + "bbox": [ + 305, + 228, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 546, + 262 + ], + "type": "text", + "content": "Table 8. Few-shot object detection results (AP50) on VOC. The value in the parentheses indicates the improvement compared to the Baseline of each split set." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 273, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 335 + ], + "type": "table", + "html": "
Aug.K- shot
1351030
Baseline3.434.666.109.1112.78
TextMania5.39(+1.96)6.47(+1.81)7.80(+1.70)10.03(+0.92)13.60(+0.82)
", + "image_path": "0f1d2a46dc5edb00155f1168085d10c25ba4fa59920b57fb5ef7d2a7fd545807.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 306, + 371, + 545, + 449 + ], + "blocks": [ + { + "bbox": [ + 305, + 338, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 338, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 545, + 361 + ], + "type": "text", + "content": "Table 9. Few-shot object detection results (mAP) on COCO. The configuration follows Table 8." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 371, + 545, + 449 + ], + "lines": [ + { + "bbox": [ + 306, + 371, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 545, + 449 + ], + "type": "table", + "html": "
Aug.ManyMediumFewIF=100IF=50IF=10
(a)Baseline71.1138.423.0038.3943.3359.29
Random71.3738.552.9038.4343.2860.39
TextManiA70.2240.739.4141.1047.1760.67
(b)Direct.71.3438.644.3238.6643.4459.82
Concat.68.0235.825.3536.9842.6859.44
", + "image_path": "1b9dbe97fb62e77f284fd5ebf733613db92332104d366d12232fc51ff1af9e5f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 453, + 545, + 475 + ], + "lines": [ + { + "bbox": [ + 305, + 453, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 545, + 475 + ], + "type": "text", + "content": "Table 10. Comparison to (a) random perturbation, and (b) direct text and concatenated embeddings on CIFAR-100-LT." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 496, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 556 + ], + "type": "text", + "content": "sification accuracy, where the result is in a similar line to the analysis [6] that classification error weighs more than localization error. The improvement is clearer when " + }, + { + "bbox": [ + 304, + 496, + 547, + 556 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 496, + 547, + 556 + ], + "type": "text", + "content": " is low. The results demonstrate the applicability of TextManiA to enhance the classification accuracy of detection models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 563, + 410, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 410, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 410, + 576 + ], + "type": "text", + "content": "4.4. Further Analyses" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 641 + ], + "type": "text", + "content": "Random Baseline. In Table 10-(a), we compare our method with the Random baseline. We randomly sample a vector from a Normal distribution " + }, + { + "bbox": [ + 304, + 582, + 547, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 304, + 582, + 547, + 641 + ], + "type": "text", + "content": " and use it instead of the difference vector, i.e., augmenting visual features with random perturbations on the same manifold of visual features." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 712 + ], + "type": "text", + "content": "The result shows that the Random baseline improves performance by serving as intra-perturb, but marginal compared to our method considering semantics additionally, which implies that semantic information embedded in the difference vector guides the augmentation more effective direction rather than random." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2533" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 273, + 166 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 273, + 166 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 273, + 166 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 273, + 166 + ], + "type": "table", + "html": "
CLIP Arch.Aug.ZSLP-10%LP-Full
ResNet50Baseline39.4750.1863.64
TextManiA-52.8364.17
ResNet101Baseline45.1757.3768.60
TextManiA-59.4969.12
ViT-BBaseline58.2173.3079.99
TextManiA-73.3579.58
", + "image_path": "5a10c70d03d264e0277408f8c88401b4b3e44ddc7edd4b39b3ae3238973047c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 235, + 288, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 288, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 288, + 366 + ], + "type": "text", + "content": "Effectiveness of Difference Vectors. While we use the difference vectors by subtracting the embeddings with and without attribute words for TextMania, there could be another way to extract the attribute information. In Table 10(b), we compare with counterparts, direct text embedding (Direct.) and concatenated embeddings (Concat.). For the Direct method, we use the text embedding computed from the attribute word directly instead of the difference vector. For the Concat method, we concatenate the text embeddings from with and without attribute words, e.g., [“bull”||“red bull”], and use it instead of the difference vector." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 371, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 539 + ], + "type": "text", + "content": "The results show that using difference vector (TextMania) outperforms using direct text embedding or concatenated embeddings, and imply that remaining contextual information after subtraction plays an important role in doing intra-perturbation in a semantic way. Although the word \"blue\" can function as both an adjective and a noun, its exact role in a sentence cannot be determined solely based on the word itself. Our intention of subtraction is for attribute words to act as a modifier in the sentence motivated by word analogy. When we computed the cosine similarity, embeddings derived directly from \"red\" and those obtained from the difference exhibited low similarity because they contain different contextual information despite the same origin of a word." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 544, + 156, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 156, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 156, + 651 + ], + "type": "text", + "content": "Linear Probing with Advanced Models. Further demonstrating the compatibility of TextMania, we apply our method during linear probing of the model. In Table 11, we test VL-LTR [70], the state-of-the-art model in long-tail" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 173, + 548, + 277, + 596 + ], + "blocks": [ + { + "bbox": [ + 47, + 170, + 287, + 204 + ], + "lines": [ + { + "bbox": [ + 47, + 170, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 170, + 287, + 204 + ], + "type": "text", + "content": "Table 12. Classification results (\\%) of CLIP with zero-shot (ZS) and linear-probe (LP) on Full and " + }, + { + "bbox": [ + 47, + 170, + 287, + 204 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 170, + 287, + 204 + ], + "type": "text", + "content": " CIFAR-100. We apply our TextManiA to the linear-probed CLIP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 173, + 548, + 277, + 596 + ], + "lines": [ + { + "bbox": [ + 173, + 548, + 277, + 596 + ], + "spans": [ + { + "bbox": [ + 173, + 548, + 277, + 596 + ], + "type": "table", + "html": "
ModelLP-Full
VL-LTR61.04
+TextMania61.82
", + "image_path": "83cbf6b36ea2d14683fae3c85b1d9df665eca467f280b46ee1212a1036c8ddd9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 651, + 287, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 651, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 46, + 651, + 287, + 711 + ], + "type": "text", + "content": "classification, on CIFAR-100. In Table 12, we use a CLIP image encoder [54] with various architectures as the baseline model and linear-probe the model on both " + }, + { + "bbox": [ + 46, + 651, + 287, + 711 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 651, + 287, + 711 + ], + "type": "text", + "content": " and full data of CIFAR-100. The results demonstrate that TextMania is compatible with linear-probed CLIP and VL-LTR models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 434, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 434, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 434, + 83 + ], + "type": "text", + "content": "Ablation Study on Attributes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 84, + 436, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 84, + 436, + 178 + ], + "spans": [ + { + "bbox": [ + 306, + 84, + 436, + 178 + ], + "type": "text", + "content": "In TextManiA, we have considered color and size attributes. To confirm the effect of each attribute, we conduct an ablation study on attributes in Table 13. The result shows that while each attribute brings non-trivial gain, using both brings more gain. We" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 465, + 73, + 533, + 136 + ], + "blocks": [ + { + "bbox": [ + 174, + 600, + 287, + 643 + ], + "lines": [ + { + "bbox": [ + 174, + 600, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 174, + 600, + 287, + 643 + ], + "type": "text", + "content": "Table 11. Comparison between the SOTA model with and without TextMania during linear probing on CIFAR-100." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 465, + 73, + 533, + 136 + ], + "lines": [ + { + "bbox": [ + 465, + 73, + 533, + 136 + ], + "spans": [ + { + "bbox": [ + 465, + 73, + 533, + 136 + ], + "type": "table", + "html": "
ColorSizeAcc.
31.10
✓33.48
✓33.89
✓✓34.52
", + "image_path": "0881c7011910077db4156c338dd864183cc830784a7098e85f0517a4dd8b6580.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 454, + 140, + 545, + 172 + ], + "lines": [ + { + "bbox": [ + 454, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 454, + 140, + 545, + 172 + ], + "type": "text", + "content": "Table 13. Ablation study on the attributes with CIFAR-100-10%." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 180, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 545, + 215 + ], + "type": "text", + "content": "believe that there are additional attributes we could use and a more effective method for selecting appropriate attributes, but leave it for future work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 245, + 378, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 245, + 378, + 257 + ], + "spans": [ + { + "bbox": [ + 306, + 245, + 378, + 257 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 271, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 547, + 415 + ], + "type": "text", + "content": "To mitigate the scarce data problem in long-tailed data distribution, small dataset, and few-shot cases, we propose a text-driven visual feature manifold augmentation method, TextManiA. Our method densifies around all the given individual visual features by adding a difference vector stem from the text embedding. While the mix-based augmentations inflict semantic perturbation in an inter-class way by label mixing, TextManiA perturbs the semantic meaning of the visual features at an intra-class level, i.e., having semantic perturbation while maintaining its class. The intraclass semantic perturbation is achieved by transferring the attribute-embedded vectors to visual feature space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 421, + 547, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 547, + 589 + ], + "type": "text", + "content": "To scrutinize the design of our estimated attribute embedding, we conduct visualization-based analyses: t-SNE plot and simple manipulation tests. The results empirically demonstrate that TextManiA readily enriches the sparse samples with comprehensible manipulation, since the general language models also reflect some extent of visual information. The experiment on the long-tail classification validates the effectiveness of our method, especially on the highly skewed class distribution. We additionally show the compatibility of TextManiA with other augmentation methods or other models in scarce data cases and during linear probing. In this work, note that we only use color and size as attributes; thus, there would be room for further investigation of other effective attributes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgment. This work was partially supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2021-0-02068, Artificial Intelligence Innovation Hub; No.2022-0-00124, Development of Artificial Intelligence Technology for Self-Improving Competency-Aware Learning Capabilities; No. 2020-0-00004, Development of Provisional Intelligence based on Long-term Visual Memory Network)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2534" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. *ICLR*, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "text", + "content": "[2] Guozhong An. The effects of adding noise during backpropagation training on a generalization performance. Neural computation, 8(3):643-674, 1996. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 287, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 287, + 180 + ], + "type": "text", + "content": "[3] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module networks. In CVPR, 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 182, + 288, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 225 + ], + "type": "text", + "content": "[4] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151-175, 2010. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 228, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 228, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 287, + 249 + ], + "type": "text", + "content": "[5] Chris M Bishop. Training with noise is equivalent to tikhonov regularization. Neural computation, 7(1):108-116, 1995. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 251, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 282 + ], + "type": "text", + "content": "[6] Ali Borji and Seyed Mehdi Iranmanesh. Empirical upper bound in object detection and more. arXiv preprint arXiv:1911.12451, 2019. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 285, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 288, + 316 + ], + "type": "text", + "content": "[7] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural networks, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 318, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 350 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 353, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 287, + 395 + ], + "type": "text", + "content": "[9] Tung-I Chen, Yueh-Cheng Liu, Hung-Ting Su, Yu-Cheng Chang, Yu-Hsiang Lin, Jia-Fong Yeh, Wen-Chin Chen, and Winston Hsu. Dual-awareness attention for few-shot object detection. IEEE TMM, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 397, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 397, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 397, + 287, + 418 + ], + "type": "text", + "content": "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 420, + 288, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 288, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 288, + 452 + ], + "type": "text", + "content": "[11] Peng Chu, Xiao Bian, Shaopeng Liu, and Haibin Ling. Feature space augmentation for long-tailed data. In ECCV, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 454, + 288, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 288, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 288, + 486 + ], + "type": "text", + "content": "[12] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. In CVPR, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 488, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 288, + 520 + ], + "type": "text", + "content": "[13] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPRW, 2020. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 522, + 288, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 554 + ], + "type": "text", + "content": "[14] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In CVPR, 2019. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 555, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 588 + ], + "type": "text", + "content": "[15] Wenliang Dai, Lu Hou, Lifeng Shang, Xin Jiang, Qun Liu, and Pascale Fung. Enabling multimodal generation on clip via vision-language knowledge distillation. ACL, 2022. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 590, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 622 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 624, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 656 + ], + "type": "text", + "content": "[17] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. ACL, 2018. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "text", + "content": "[18] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 48, + 692, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 288, + 712 + ], + "type": "text", + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 39, + "blocks": [ + { + "bbox": [ + 326, + 74, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 547, + 106 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 107, + 546, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 546, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 546, + 139 + ], + "type": "text", + "content": "[20] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 140, + 546, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 546, + 172 + ], + "type": "text", + "content": "[21] Kevin Frans, Lisa B Soros, and Olaf Witkowski. Clipdraw: Exploring text-to-drawing synthesis through language-image encoders. arXiv preprint arXiv:2106.14843, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 174, + 547, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 174, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 174, + 547, + 216 + ], + "type": "text", + "content": "[22] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. *ICLR*, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 217, + 547, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 547, + 258 + ], + "type": "text", + "content": "[23] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM TOG, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 261, + 546, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 546, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 546, + 293 + ], + "type": "text", + "content": "[24] Caglar Gulcehre, Marcin Moczulski, Misha Denil, and Yoshua Bengio. Noisy activation functions. In ICML, pages 3059-3068. PMLR, 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 295, + 547, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 295, + 547, + 325 + ], + "spans": [ + { + "bbox": [ + 308, + 295, + 547, + 325 + ], + "type": "text", + "content": "[25] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, 2017. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 327, + 546, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 546, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 546, + 358 + ], + "type": "text", + "content": "[26] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In CVPR, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 361, + 546, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 546, + 392 + ], + "type": "text", + "content": "[27] Bart LM Happel and Jacob MJ Murre. Design and evolution of modular neural network architectures. Neural networks, 7(6-7):985–1004, 1994. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 394, + 546, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 546, + 424 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 546, + 424 + ], + "type": "text", + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 426, + 546, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 546, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 546, + 458 + ], + "type": "text", + "content": "[29] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3, 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 460, + 546, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 546, + 491 + ], + "type": "text", + "content": "[30] Lasse Holmstrom and Petri Koistinen. Using additive noise in back-propagation training. IEEE transactions on neural networks, 3(1):24-38, 1992. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 494, + 546, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 494, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 494, + 546, + 525 + ], + "type": "text", + "content": "[31] Hanzhe Hu, Shuai Bai, Aoxue Li, Jinshi Cui, and Liwei Wang. Dense relation distillation with context-aware aggregation for few-shot object detection. In CVPR, 2021. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 526, + 546, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 546, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 546, + 557 + ], + "type": "text", + "content": "[32] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, 2022. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 559, + 546, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 546, + 602 + ], + "type": "text", + "content": "[33] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 308, + 604, + 546, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 546, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 546, + 645 + ], + "type": "text", + "content": "[34] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In ICLR, 2020. 7" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 308, + 647, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 546, + 679 + ], + "type": "text", + "content": "[35] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In CVPR, 2022. 3" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 308, + 681, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 546, + 712 + ], + "type": "text", + "content": "[36] Youmin Kim, Jinbae Park, YounHo Jang, Muhammad Ali, Tae-Hyun Oh, and Sung-Ho Bae. Distilling global and local logits with densely connected relations. In ICCV, 2021. 4" + } + ] + } + ], + "index": 38 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2535" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[37] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, Citeseer, 2009. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 288, + 139 + ], + "type": "text", + "content": "[38] Ashish Kumar, Zipeng Fu, Deepak Pathak, and Jitendra Malik. Rma: Rapid motor adaptation for legged robots. In Robotics: Science and Systems, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 140, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 287, + 163 + ], + "type": "text", + "content": "[39] Gihyun Kwon and Jong Chul Ye. Clipstyler: Image style transfer with a single text condition. In CVPR, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 164, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 164, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 164, + 287, + 185 + ], + "type": "text", + "content": "[40] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 2015. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 186, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 288, + 218 + ], + "type": "text", + "content": "[41] Boyi Li, Felix Wu, Ser-Nam Lim, Serge Belongie, and Kilian Q Weinberger. On feature normalization and data augmentation. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 220, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 253 + ], + "type": "text", + "content": "[42] Bohao Li, Boyu Yang, Chang Liu, Feng Liu, Rongrong Ji, and Qixiang Ye. Beyond max-margin: Class margin equilibrium for few-shot object detection. In CVPR, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 255, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 255, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 255, + 288, + 297 + ], + "type": "text", + "content": "[43] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 299, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 299, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 49, + 299, + 287, + 331 + ], + "type": "text", + "content": "[44] Pan Li, Da Li, Wei Li, Shaogang Gong, Yanwei Fu, and Timothy M Hospedales. A simple feature augmentation for domain generalization. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 333, + 287, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 333, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 49, + 333, + 287, + 375 + ], + "type": "text", + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 377, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 377, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 49, + 377, + 287, + 410 + ], + "type": "text", + "content": "[46] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In CVPR, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 411, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 287, + 443 + ], + "type": "text", + "content": "[47] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In CVPR, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 445, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 287, + 476 + ], + "type": "text", + "content": "[48] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In ICLR, 2013. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 479, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 479, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 49, + 479, + 287, + 531 + ], + "type": "text", + "content": "[49] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 534, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 534, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 49, + 534, + 287, + 577 + ], + "type": "text", + "content": "[50] Tae-Hyun Oh, Tali Dekel, Changil Kim, Inbar Mosseri, William T Freeman, Michael Rubinstein, and Wojciech Matusik. Speech2face: Learning the face behind a voice. In CVPR, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 579, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 579, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 49, + 579, + 287, + 612 + ], + "type": "text", + "content": "[51] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 613, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 287, + 644 + ], + "type": "text", + "content": "[52] Guim Perarnau, Joost van de Weijer, Bogdan Raducanu, and Jose M. Álvarez. Invertible conditional gans for image editing. In CVPR, 2016. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 679 + ], + "type": "text", + "content": "[53] Limeng Qiao, Yuxuan Zhao, Zhiyuan Li, Xi Qiu, Jianan Wu, and Chi Zhang. Defrcn: Decoupled faster r-cnn for few-shot object detection. In ICCV, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "text", + "content": "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 711 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 327, + 73, + 546, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 546, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 546, + 94 + ], + "type": "text", + "content": "transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4, 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 95, + 547, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 547, + 137 + ], + "type": "text", + "content": "[55] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 140, + 546, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 546, + 172 + ], + "type": "text", + "content": "[56] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 174, + 546, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 174, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 174, + 546, + 205 + ], + "type": "text", + "content": "[57] Russell Reed and Robert J MarksII. Neural smithing: supervised learning in feedforward artificial neural networks. Mit Press, 1999. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 206, + 546, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 546, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 546, + 237 + ], + "type": "text", + "content": "[58] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In ICML, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 239, + 546, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 546, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 546, + 271 + ], + "type": "text", + "content": "[59] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 273, + 546, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 546, + 304 + ], + "type": "text", + "content": "[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 306, + 546, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 306, + 546, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 306, + 546, + 349 + ], + "type": "text", + "content": "[61] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 350, + 546, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 546, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 546, + 382 + ], + "type": "text", + "content": "[62] Fawaz Sammani, Tanmoy Mukherjee, and Nikos Deligiannis. Nlx-gpt: A model for natural language explanations in vision and vision-language tasks. In CVPR, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 384, + 546, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 546, + 415 + ], + "type": "text", + "content": "[63] Gyungin Shin, Weidi Xie, and Samuel Albanie. Namedmask: Distilling segmenters from complementary foundation models. arXiv preprint arXiv:2209.11228, 2022. 3, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 416, + 546, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 546, + 469 + ], + "type": "text", + "content": "[64] Patrice Y Simard, Yann A LeCun, John S Denker, and Bernard Victorri. Transformation invariance in pattern recognition—tangent distance and tangent propagation. In Neural networks: tricks of the trade, pages 239–274. Springer, 1998. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 472, + 546, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 546, + 525 + ], + "type": "text", + "content": "[65] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. In NeurIPS, 2020. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 526, + 546, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 546, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 546, + 569 + ], + "type": "text", + "content": "[66] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 2014. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 571, + 546, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 546, + 602 + ], + "type": "text", + "content": "[67] Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. Fsce: Few-shot object detection via contrastive proposal encoding. In CVPR, 2021. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 604, + 546, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 546, + 635 + ], + "type": "text", + "content": "[68] Kim Sung-Bin, Arda Senocak, Hyunwoo Ha, Andrew Owens, and Tae-Hyun Oh. Sound to visual scene generation by audio-to-visual latent alignment. In CVPR, 2023. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 637, + 546, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 637, + 546, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 637, + 546, + 669 + ], + "type": "text", + "content": "[69] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 670, + 546, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 546, + 711 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 546, + 711 + ], + "type": "text", + "content": "[70] Changyao Tian, Wenhai Wang, Xizhou Zhu, Jifeng Dai, and Yu Qiao. Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition. In ECCV, 2022. 9" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2536" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 287, + 95 + ], + "type": "text", + "content": "[71] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In CVPR, 2018. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 117 + ], + "type": "text", + "content": "[72] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 118, + 288, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 161 + ], + "type": "text", + "content": "[73] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In ICML, 2019. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 162, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 216 + ], + "type": "text", + "content": "[74] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and León Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 250 + ], + "type": "text", + "content": "[75] Xin Wang, Thomas E Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. In ICML, 2020. 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 251, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 288, + 305 + ], + "type": "text", + "content": "[76] Zhecan Wang, Noel Codella, Yen-Chun Chen, Luowei Zhou, Xiyang Dai, Bin Xiao, Jianwei Yang, Haoxuan You, Kai-Wei Chang, Shih-fu Chang, et al. Multimodal adaptive distillation for leveraging unimodal encoders for vision-language tasks. arXiv preprint arXiv:2204.10496, 2022. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "type": "text", + "content": "[77] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 393 + ], + "type": "text", + "content": "[78] Chien-Sheng Wu, Steven Hoi, Richard Socher, and Caiming Xiong. Tod-bert: Pre-trained natural language understanding for task-oriented dialogue. EMNLP, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 288, + 426 + ], + "type": "text", + "content": "[79] Yang Xiao and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. In ECCV, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 426, + 288, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 460 + ], + "type": "text", + "content": "[80] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In ICCV, 2019. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 460, + 288, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 482 + ], + "type": "text", + "content": "[81] Lu Yang, He Jiang, Qing Song, and Jun Guo. A survey on long-tailed visual recognition. IJCV, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 483, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 483, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 288, + 525 + ], + "type": "text", + "content": "[82] Moon Ye-Bin, Dongmin Choi, Yongjin Kwon, Junsik Kim, and Tae-Hyun Oh. Eninst: Enhancing weakly-supervised low-shot instance segmentation. arXiv preprint arXiv:2302.09765, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 526, + 288, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 288, + 558 + ], + "type": "text", + "content": "[83] Kim Youwang, Kim Ji-Yeon, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. ECCV, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 559, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 635 + ], + "type": "text", + "content": "[84] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, Ce Liu, Mengchen Liu, Zicheng Liu, Yumao Lu, Yu Shi, Lijuan Wang, Jianfeng Wang, Bin Xiao, Zhen Xiao, Jianwei Yang, Michael Zeng, Luowei Zhou, and Pengchuan Zhang. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "type": "text", + "content": "[85] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[86] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[87] Yuhui Zhang, Jeff Z HaoChen, Shih-Cheng Huang, Kuan-Chieh Wang, James Zou, and Serena Yeung. Diagnosing and rectifying vision models using language. In ICLR, 2023. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 138 + ], + "type": "text", + "content": "[88] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In ECCV, 2020. 5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2537" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/a412d306-b042-443c-958a-505db74fd5e7_content_list.json b/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/a412d306-b042-443c-958a-505db74fd5e7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5879b51f71ed5183fc790c9c354676d339032799 --- /dev/null +++ b/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/a412d306-b042-443c-958a-505db74fd5e7_content_list.json @@ -0,0 +1,2015 @@ +[ + { + "type": "text", + "text": "TextPSG: Panoptic Scene Graph Generation from Textual Descriptions", + "text_level": 1, + "bbox": [ + 125, + 130, + 844, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengyang Zhao1 Yikang Shen2 Zhenfang Chen2 Mingyu Ding3 Chuang Gan2,4 \n1Peking University 2MIT-IBM Watson AI Lab \n3UC Berkley 4UMass Amherst", + "bbox": [ + 263, + 179, + 704, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Panoptic Scene Graph has recently been proposed for comprehensive scene understanding. However, previous works adopt a fully-supervised learning manner, requiring large amounts of pixel-wise densely-annotated data, which is always tedious and expensive to obtain. To address this limitation, we study a new problem of Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG). The key idea is to leverage the large collection of free image-caption data on the Web alone to generate panoptic scene graphs. The problem is very challenging for three constraints: 1) no location priors; 2) no explicit links between visual regions and textual entities; and 3) no predefined concept sets. To tackle this problem, we propose a new framework TextPSG consisting of four modules, i.e., a region grouper, an entity grounder, a segment merger, and a label generator, with several novel techniques. The region grouper first groups image pixels into different segments and the entity grounder then aligns visual segments with language entities based on the textual description of the segment being referred to. The grounding results can thus serve as pseudo labels enabling the segment merger to learn the segment similarity as well as guiding the label generator to learn object semantics and relation predicates, resulting in a fine-grained structured scene understanding. Our framework is effective, significantly outperforming the baselines and achieving strong out-of-distribution robustness. We perform comprehensive ablation studies to corroborate the effectiveness of our design choices and provide an in-depth analysis to highlight future directions. Our code, data, and results are available on our project page: https://vis-ww.c.s.umass.edu/TextPSG.", + "bbox": [ + 76, + 316, + 473, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 814, + 209, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A scene graph is a directed-graph-based abstract representation of the objects and their relations within a scene. It has been widely utilized to develop a structured scene understanding of object semantics, locations, and relations,", + "bbox": [ + 76, + 839, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/065d0ed580478e3af013e3fcf045d049fb6c45002c8ef8797ce8355a2355b286.jpg", + "image_caption": [ + "Figure 1. Problem Overview. Different from the traditional bbox-based form of the scene graph as shown in (a), Caption-to-PSG aims to generate the mask-based panoptic scene graph. In Caption-to-PSG, the model has no access to any location priors, explicit region-entity links, or pre-defined concept sets. Consequently, the model is required to learn partitioning and grounding as illustrated in (b), as well as object semantics and relation predicates as illustrated in (c), all purely from textual descriptions." + ], + "image_footnote": [], + "bbox": [ + 503, + 284, + 624, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0254e5361ce1c24aa5539ff5c6c3b04b2ce638ce0bbed7ee183c246ba9132828.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 285, + 888, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which facilitates a variety of downstream applications, such as image generation [14, 8], visual reasoning [42, 1, 39], and robotics [2, 10].", + "bbox": [ + 496, + 582, + 890, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Typically, in a scene graph, each node denotes an object in the scene located by a bounding box (bbox) with a semantic label, and each directed edge denotes the relation between a pair of objects with a predicate label. Nonetheless, a recent work [49] points out that such a bbox-based form of scene graph is not ideal enough. Firstly, compared with pixel-wise segmentation masks, bboxes are less fine-grained and may contain some noisy pixels belonging to other objects, limiting the applications for some downstream tasks. For example, as shown in Fig. 1 (a), about half of the pixels in the yellow_bbox for girl belong to wall. Secondly, it is challenging for bboxes to cover the entire scene without ambiguities caused by overlaps, which prevents a scene graph from including every object in the scene for a complete description. To this end, the work [49] proposes the concept of Panoptic Scene Graph (PSG), in which each object is grounded by a panoptic segmentation mask, to reach a comprehensive structured scene representation.", + "bbox": [ + 496, + 628, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2839", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, all existing works [49, 45] approach PSG generation through a fully-supervised manner, i.e., learning to perform panoptic segmentation and relation prediction from manually-annotated datasets with explicit supervision for both segmentation and relation prediction. Unfortunately, it is extremely labor-intensive to build such datasets, making it difficult to scale up to cover more complex scenes, object semantics, and relation predicates, thus significantly limiting the generalizability and the application of these methods to the real world. For instance, the current PSG dataset [49] only covers 133 object semantics and 56 relation predicates.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To relieve the reliance on densely-annotated data, weakly-supervised methods [52, 58, 22] for scene graph generation are promising. These methods could induce scene graphs from image-caption pairs, which can be easily harvested from the Web for free. Even so, they still rely heavily on two strong preconditions, i.e., a powerful region proposal network (e.g., [35]) and a pre-defined set of object semantics and relation predicates. Although these preconditions facilitate the learning process of the methods, they also limit the generalizability for locating novel objects (unforeseen objects for the region proposal network) and constrain the understanding into the limited concept set.", + "bbox": [ + 75, + 260, + 472, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by previous weakly-supervised methods, we introduce a new problem, Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG), to explore a holistic structured scene understanding without labor-intensive data annotation. Considering the limitation of the preconditions mentioned, we set three constraints to Caption-to-PSG to reach a more comprehensive and generalizable understanding, which results in a very challenging problem: a) only image-caption pairs are provided during training, without any location priors in either region proposals or location supervision; b) the explicit links between regions in images and entities in captions are missing; c) no concept sets are pre-defined, i.e., neither object semantics nor relation predicates are known beforehand.", + "bbox": [ + 75, + 444, + 472, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given these three constraints, we argue that there are two key challenges for the model to solve the problem. Firstly, the model should learn to ground entities in language onto the visual scene without explicit location supervision, i.e., the ability to perform partitioning and grounding, as shown in Fig. 1 (b), should be developed purely from textual descriptions. Secondly, during training, the model should also learn the object semantics and relation predicates from textual descriptions, as shown in Fig. 1 (c), without pre-defined fixed object and relation vocabularies. By solving these challenges, the model could associate visual scene patterns with textual descriptions, gradually acquire common sense among them, and finally reach a more comprehensive and generalizable understanding, including novel object location, extensive semantics recognition, and complex relation analysis, which is more suitable to the real world.", + "bbox": [ + 75, + 659, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With these considerations, we propose a novel framework, TextPSG, as the first step towards this challenging problem. TextPSG consists of a series of modules to cooperate with each other, i.e., a region grouper, an entity grounder, a segment merger, and a label generator. The region grouper learns to merge image regions into several segments in a hierarchical way based on object semantics, similar to [48]. The entity grouder employs a fine-grained contrastive learning strategy [51] to bridge the textual description and the visual content, grounding entities in the caption onto the image segments. With the entity-grounding results as pseudo labels, the segment merger learns similarity matrices to merge small image segments during inference, while the label generator learns the prediction of object semantics and relation predicates. Specifically, in the segment merger, we propose to leverage the grounding as explicit supervision for merging, compared with [48] which learns merging in a fully implicit manner, to improve the ability of location. In the label generator, different from all previous pipelines for scene graph generation, we reformulate the label prediction as an auto-regressive generation problem rather than a classification problem, and employ a pre-trained language model [21] as the decoder to leverage the pre-learned common sense. We further design a novel prompt-embedding-based technique (PET) to better incorporate common sense from the language model. Our experiments show that TextPSG significantly outperforms the baselines and achieves strong out-of-distribution (OOD) robustness. Comprehensive ablation studies corroborate the effectiveness of our design choices. As a side product, the proposed grouder and merger modules also have been observed to enhance text-supervised semantic segmentation.", + "bbox": [ + 496, + 90, + 893, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In spite of the promising performance of TextPSG, certain challenges persist. We delve into an in-depth analysis of the failure cases, provide a model diagnosis, and discuss potential future directions for enhancing our framework.", + "bbox": [ + 496, + 574, + 893, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, our contributions are as follows:", + "bbox": [ + 519, + 635, + 813, + 648 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new problem, Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG), to alleviate the burden of human annotation for PSG by learning purely from the weak supervision of captions.", + "- We propose a new modularized framework, TextPSG, with several novel techniques, which significantly outperforms the baselines and achieves strong OOD robustness. We demonstrate that the proposed modules in TextPSG can also facilitate text-supervised semantic segmentation.", + "- We perform an in-depth failure case analysis with a model diagnosis, and further highlight future directions." + ], + "bbox": [ + 500, + 648, + 890, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 829, + 640, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Bbox-based Scene Graph Generation. Bbox-based scene graph generation aims to create a structured representation of object semantics, locations, and relations in the", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2840", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scene, where each object is identified by a bbox. Most of existing works [50, 47, 41, 12, 24] follow a fully-supervised approach to learn the generation from densely-annotated datasets [18, 13], which requires significant human labors. To reduce the labeling effort, some weakly-supervised methods have been proposed [32, 55, 57, 40]. Recent works [52, 58, 22] further explore learning scene graph generation from image-caption pairs. However, they all rely on off-the-shelf region proposal networks for the location of objects in the scene, which are typically pretrained on pre-defined fixed sets of object semantics, limiting their generalizability to locating unforeseen objects. To reach a more granular and accurate grounding, [17] proposes to ground each object by segmentation. A recent work [49] further introduces the concept of PSG, where each object is identified by a panoptic segmentation mask, as a more comprehensive scene representation.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-supervised Semantic Segmentation (TSSS). TSSS [48, 20, 26, 9, 27, 59] aims to learn image pixel semantic labeling from image-caption pairs without fine-grained pixel-wise annotations. Similar to TSSS, our proposed Caption-to-PSG aims to learn to connect visual regions and textual entities from only image-caption pairs and has the potential to leverage the large collection of free data on the Web. However, different from TSSS, Caption-to-PSG further requires the model to learn the relations among different visual regions, resulting in a higher-order structured understanding of visual scenes. In addition to unknown object semantics, Caption-to-PSG does not assume any pre-defined relation predicate concepts.", + "bbox": [ + 75, + 347, + 472, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Grounding. Our work is also related to visual grounding [16, 56, 29, 11, 6, 33], which grounds entities in language onto objects in images. Early works [54, 53, 7] on visual grounding typically detect object proposals [35, 43] from images first and then match them with language descriptions by putting features of both modalities into the same feature space, which are in a fully-supervised learning manner. There are also some weakly-supervised grounding methods [15, 36, 5] which relieve the need for dense regional annotations by multiple instance learning [15] or learning to reconstruct [36]. Different from them, Caption-to-PSG is more challenging since it requires grounding fine-grained object relations between entities without region proposal networks for a pre-defined object vocabulary.", + "bbox": [ + 75, + 544, + 472, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 768, + 284, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG). A PSG $\\mathcal{G} = (\\mathcal{V},\\mathcal{E})$ is a directed graph representation of the objects and the relations among them in a scene image $I\\in \\mathbb{R}^{H\\times W\\times 3}$ . Each node $v_{i}\\in \\mathcal{V}$ denotes an object in $I$ located by a panoptic segmentation mask $m_i\\in \\{0,1\\}^{H\\times W}$ with an object semantic label $o_i\\in \\mathcal{C}_o$ , and each directed edge $e_{ij}\\in \\mathcal{E}$ denotes a", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pair of subject $o_i$ and object $o_j$ with a relation predicate label $r_{ij} \\in \\mathcal{C}_r$ , where $\\mathcal{C}_o$ and $\\mathcal{C}_r$ are the defined concept sets of object semantics and relation predicates. Note that for a PSG, it is constrained that all segmentation masks could not overlap, i.e., $\\sum_{i=1}^{|\\mathcal{V}|} m_i \\leq 1^{H \\times W}$ .", + "bbox": [ + 496, + 90, + 890, + 169 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a large collection of paired scene images and textual descriptions $S = \\{(I_i, T_i)\\}_i$ , Caption-to-PSG aims to learn PSG generation from purely text descriptions for a holistic structured scene understanding, i.e., during training, only $S$ is provided as supervision, while during inference, with a scene observation $I'$ as input, the model is required to generate a corresponding PSG $\\mathcal{G}'$ .", + "bbox": [ + 496, + 169, + 893, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Three Constraints. Note that in Caption-to-PSG, three important constraints are set to reach a more comprehensive and generalizable scene understanding: a) no location priors: different from all previous scene graph generation methods, neither pre-trained region proposal networks nor location supervision are allowed; b) no explicit region-entity links: the links between regions in the image $I$ and entities in the textual description $T$ are not provided; c) no pre-defined concept sets: the target concept sets defined for inference, i.e., object semantics $\\mathcal{C}_o$ and relation predicates $\\mathcal{C}_r$ , are unknown during training.", + "bbox": [ + 496, + 276, + 893, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 498, + 458, + 589, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. As illustrated in Fig. 2, our proposed framework for Caption-to-PSG, TextPSG, contains four modules to cooperate with each other: a region grouper, an entity grounder, a segment merger, and a label generator.", + "bbox": [ + 496, + 484, + 890, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, TextPSG takes batched image-caption pairs as input. For each pair, the image is passed through the region grouper to be partitioned into several image segments, while the caption is first pre-processed to extract its linguistic structure as a text graph and then taken by the entity grounder to ground textual entities in the graph onto the image segments. With the grounding results as pseudo labels, the segment merger learns similarity matrices between small image segments for further merging during inference, while the label generator learns the prediction of object semantics and relation predicates.", + "bbox": [ + 496, + 546, + 893, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During inference, for each input image, the image segments output from the region grouper are directly passed to the segment merger to be further merged according to the learned similarity matrices, and then fed to the label generator to predict the object semantic labels and the relation predicate labels.", + "bbox": [ + 496, + 713, + 890, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Text Graph Preprocessing", + "text_level": 1, + "bbox": [ + 498, + 815, + 736, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following previous methods [52, 58, 22] that leverage a rule-based language parser [46] based on [37] to preprocess textual descriptions, in TextPSG, we employ the same parser to extract linguistic structures from captions. Addi", + "bbox": [ + 496, + 839, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2841", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1088337b43e50200b645149a7c81c28956c4478e71e9ee822b3f5e51652acf2b.jpg", + "image_caption": [ + "Figure 2. Framework Overview of TextPSG. The framework consists of four modules cooperating with each other: a region grouper to merge regions in the input image into several segments, an entity grounder to ground entities in the caption onto the image segments, a segment merger to learn similarity matrices to merge small image segments during inference, and a label generator to learn the prediction of object semantics and relation predicates. The solid arrows indicate the training flow, while the dash arrows indicate the inference flow. The arrows from the region grouper to the label generator indicating the segment feature and mask query are omitted." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 893, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tionally, inspired by the success of open information extraction (OpenIE) [3] in natural language processing, we also employ an OpenIE system from Stanford CoreNLP [28] for extraction as a supplement. After merging, for each caption, we obtain its linguistic structure represented in a text graph, where each node denotes an entity, and each directed edge denotes the relation between an entity pair.", + "bbox": [ + 75, + 454, + 472, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Region Grouper", + "text_level": 1, + "bbox": [ + 76, + 568, + 238, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With a scene image as input, the region grouper aims to merge the regions with similar object semantics into several segments and extract corresponding high-level features.", + "bbox": [ + 75, + 592, + 468, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our region grouper follows the hierarchical design of GroupViT [48]. Given an input image, the grouper first splits the image into $N$ non-overlapping patches as the initial image segments $\\{\\mathbf{s}_i^0\\}_{i = 1}^N$ . These segments are then passed through $K$ grouping layers, where they are merged into larger, arbitrary-shaped segments progressively. Specifically, within each grouping layer $\\mathbf{Grp}_k(k = 1,2,\\dots ,K)$ , $H_{k}$ grouping centers $\\{\\mathbf{c}_i^k\\}_{i = 1}^{H_k}$ could be learned. The grouping operation is performed through an attention mechanism between the centers and the segments, merging $H_{k - 1}$ input segments into $H_{k}$ larger ones, i.e.,", + "bbox": [ + 75, + 638, + 470, + 805 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {s} _ {i} ^ {k} \\right\\} _ {i = 1} ^ {H _ {k}} = \\mathbf {G r p} _ {k} \\big (\\left\\{\\mathbf {c} _ {i} ^ {k} \\right\\} _ {i = 1} ^ {H _ {k}}, \\left\\{\\mathbf {s} _ {i} ^ {k - 1} \\right\\} _ {i = 1} ^ {H _ {k - 1}} \\big).\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 811, + 415, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that $H_0 = N$ . After the hierarchical grouping, multiple groups of segments $\\{\\mathbf{s}_i^k\\}_{i=1}^{H_k}$ at different grouping stages are obtained. More details about the design of $\\{\\mathbf{Grp}_k\\}_{k=1}^K$ can be found in the supplementary material.", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Entity Grounder", + "text_level": 1, + "bbox": [ + 500, + 454, + 666, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the explicit region-entity links are not provided, bridging the textual description and the visual content automatically plays an important role in solving Caption-to-PSG. Inspired by FILIP [51], in TextPSG, we employ a similar fine-grained contrastive learning strategy to perform region-entity alignment.", + "bbox": [ + 496, + 479, + 890, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each grouping stage $k$ , on image side, the grounder projects the segment group $\\{\\mathbf{s}_i^k\\}_{i=1}^{H_k}$ into a new feature space $\\mathcal{F}$ by a multi-layer perceptron (MLP) $\\mathbf{Proj}_k^I$ to obtain segment embeddings $\\{\\mathbf{x}_i^k\\}_{i=1}^{H_k}$ . On text side, the input caption is first tokenized into $M$ tokens $\\{\\mathbf{t}_i\\}_{i=1}^M$ , which are then processed by a Transformer [44] $\\mathbf{Tfm}^T$ to propagate information between each other. A recurrent neural network (RNN) $\\mathbf{Rnn}$ further merges the tokens corresponding to the same entity, encoding tokens into their associated weights one by one and utilizing weighted sum to merge the token features into a singular entity feature. Finally, these entity features are projected to the same feature space $\\mathcal{F}$ by a MLP $\\mathbf{Proj}^T$ to obtain entity embeddings $\\{\\mathbf{y}_i\\}_{i=1}^E$ , where $E$ denotes the number of entities in the caption.", + "bbox": [ + 496, + 571, + 892, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the segment embeddings and the entity embeddings in the shared feature space $\\mathcal{F}$ , we compute their token-wise similarities. Specifically, for the $i$ -th segment, we compute its cosine similarities with all entities to obtain the token-wise similarity from the $i$ -th segment to the caption $p_i^k$ via", + "bbox": [ + 496, + 787, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} ^ {k} = \\max _ {1 \\leq j \\leq E} \\cos [ \\mathbf {x} _ {i} ^ {k}, \\mathbf {y} _ {j} ],\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 878, + 779, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2842", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\cos[\\cdot, \\cdot]$ denotes the cosine similarity operation. Note that different from the original FILIP [51], in the scenario of region-entity alignment, some regions in the scene may not be described in the caption, while some entities in the caption may not exist in the scene. To tackle this problem, we propose to set a filtering threshold $\\theta$ , where pairs with similarity lower than $\\theta$ will be considered in different semantics and filtered out. The fine-grained similarity from the image to the caption $p^k$ can thus be computed via", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np ^ {k} = \\frac {1}{\\sum_ {i = 1} ^ {H _ {k}} {\\mathbf {1} _ {p _ {i} ^ {k} > \\theta}}} \\sum_ {i = 1} ^ {H _ {k}} (p _ {i} ^ {k} \\cdot \\mathbf {1} _ {p _ {i} ^ {k} > \\theta}).\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 234, + 393, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, we can also compute the token-wise similarity from the $j$ -th entity to the image $q_j^k$ via", + "bbox": [ + 76, + 284, + 468, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq _ {j} ^ {k} = \\max _ {1 \\leq i \\leq H _ {k}} \\cos [ \\mathbf {x} _ {i} ^ {k}, \\mathbf {y} _ {j} ],\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 323, + 359, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the fine-grained similarity from the caption to the image $q^{k}$ via", + "bbox": [ + 75, + 356, + 468, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq ^ {k} = \\frac {1}{\\sum_ {j = 1} ^ {E} \\mathbf {1} _ {q _ {j} ^ {k} > \\theta}} \\sum_ {j = 1} ^ {E} (q _ {j} ^ {k} \\cdot \\mathbf {1} _ {q _ {j} ^ {k} > \\theta}).\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 387, + 393, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Denoting the training batch with batch size $B$ as $\\{(I_i,T_i)\\}_{i = 1}^B$ , the fine-grained similarity from the image $I_{i}$ to the caption $T_{j}$ as $p^{k,i\\rightarrow j}$ and from the caption $T_{j}$ to the image $I_{i}$ as $q^{k,j\\rightarrow i}$ , the image-to-text fine-grained contrastive loss $\\mathcal{L}_{fine}^{k,I\\to T}$ and the text-to-image fine-grained contrastive loss $\\mathcal{L}_{fine}^{k,T\\to I}$ can then be formulated as", + "bbox": [ + 76, + 431, + 468, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f i n e} ^ {k, I \\rightarrow T} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\frac {\\exp \\left(p ^ {k , i \\rightarrow i} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(p ^ {k , i \\rightarrow j} / \\tau\\right)},\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 537, + 408, + 594 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f i n e} ^ {k, T \\rightarrow I} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\frac {\\exp \\left(q ^ {k , i \\rightarrow i} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(q ^ {k , i \\rightarrow j} / \\tau\\right)},\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 602, + 408, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is a learnable temperature. The total fine-grained contrastive loss is", + "bbox": [ + 76, + 662, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f i n e} ^ {k} = \\frac {1}{2} (\\mathcal {L} _ {f i n e} ^ {k, I \\rightarrow T} + \\mathcal {L} _ {f i n e} ^ {k, T \\rightarrow I}).\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 696, + 380, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By minimizing $\\mathcal{L}_{fine}^{k}$ at all grouping stages during training, our framework could reach a meaningful fine-grained alignment automatically, i.e., for the $i$ -th segment $\\mathbf{s}_i^k$ , the $l_i^k$ -th entity satisfying", + "bbox": [ + 76, + 734, + 468, + 796 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl_{i}^{k} = \\operatorname *{arg max}_{1\\leq j\\leq E}\\cos [\\mathbf{x}_{i}^{k},\\mathbf{y}_{j}]\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 801, + 359, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tends to have a similar semantics with $\\mathbf{s}_i^k$ . We thus obtain $\\{l_i^k\\}_{i=1}^{H_k}$ as the grounding results for the image segments $\\{\\mathbf{s}_i^k\\}_{i=1}^{H_k}$ . A further explanation of the automatic meaningful alignment can be found in the supplementary material.", + "bbox": [ + 76, + 838, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Segment Merger", + "text_level": 1, + "bbox": [ + 500, + 90, + 665, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To improve the ability of location, we propose to leverage the entity-grounding results as explicit supervision to learn a group of similarity matrices between image segments for small segments merging during inference, compared with [48] that learns the merging fully implicitly.", + "bbox": [ + 496, + 113, + 890, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each grouping stage $k$ , we compute the cosine similarity between each pair of image segments, which is then linearly re-scaled into $[0,1]$ to formulate a similarity matrix $\\mathbf{Sim}_k \\in [0,1]^{H_k \\times H_k}$ , where", + "bbox": [ + 496, + 189, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S i m} _ {k} [ i, j ] = \\frac {1}{2} \\left(\\cos \\left[ \\mathbf {x} _ {i} ^ {k}, \\mathbf {x} _ {j} ^ {k} \\right] + 1\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 257, + 808, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further leverage $\\{l_i^k\\}_{i = 1}^{H_k}$ as pseudo labels to build a pseudo target matrix $\\mathbf{Sim}_k^{target} \\in \\{0,1\\}^{H_k \\times H_k}$ , where", + "bbox": [ + 498, + 295, + 890, + 328 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S i m} _ {k} ^ {t a r g e t} [ i, j ] = \\left\\{ \\begin{array}{l} 1, \\text {i f} l _ {i} ^ {k} = l _ {j} ^ {k} \\wedge \\cos [ \\mathbf {x} _ {i} ^ {k}, \\mathbf {y} _ {l _ {i} ^ {k}} ] > \\theta \\\\ \\qquad \\wedge \\cos [ \\mathbf {x} _ {j} ^ {k}, \\mathbf {y} _ {l _ {j} ^ {k}} ] > \\theta , \\\\ 0, \\text {o t h e r w i s e}. \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 335, + 867, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The similarity loss for the stage $k$ is then formulated as", + "bbox": [ + 519, + 407, + 883, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s i m} ^ {k} = \\frac {1}{H _ {k} ^ {2}} \\| \\mathbf {S i m} _ {k} - \\mathbf {S i m} _ {k} ^ {\\text {t a r g e t}} \\| _ {F} ^ {2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 430, + 818, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Label Generator", + "text_level": 1, + "bbox": [ + 500, + 469, + 665, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addressing the challenge of no pre-defined concept sets, the previous work [58] proposes to build a large vocabulary for learning during training and use WordNet [31] to correlate predictions within this vocabulary to the target concepts during inference. However, there are two limitations to the previous method. Firstly, compared with the extensive object semantics and relation predicates contained in textual descriptions, despite the large vocabulary established, it is inevitable that some classes will be overlooked. Secondly, leveraging WordNet to match vocabulary predictions to targets is not accurate and robust enough, for WordNet may only reach a coarse matching with multiple target concepts. This imprecision is particularly pronounced for relation predicates relative to object semantics.", + "bbox": [ + 496, + 493, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given these limitations, we introduce a novel approach in TextPSG. Instead of approaching label prediction of objects and relations as a traditional classification problem, we reformulate it as an auto-regressive generation problem, which eliminates the necessity for pre-defined concept sets.", + "bbox": [ + 496, + 704, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared with a vanilla RNN, we employ a pre-trained vision language model BLIP [21] to leverage the pre-learned common sense. BLIP can take an image as input and output a caption to describe the image. In TextPSG, we borrowed the pre-trained decoder module from BLIP to perform the generation of object and relation labels.", + "bbox": [ + 496, + 779, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, the label generator takes the caption-parsed text graph, the segment features from the region", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2843", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "grouper, and the grounding results $\\{l_i^k\\}_{i = 1}^{H_k}$ from the entity grouper as input. It filters out the segments with tokenwise similarity lower than the threshold $\\theta$ , merges the segments mapped to the same entity, and queries the corresponding image masks from the region grouper. Then, $E_{k}$ image masks $\\{\\mathbf{m}_i^k\\}_{i = 1}^{E_k}$ with their pseudo entity labels $\\{b_i^k\\}_{i = 1}^{E_k}$ can be obtained, where each $b_{i}^{k}$ is one entity in the text graph. $E_{k}\\leq E$ because some textual entities may not exist in the image.", + "bbox": [ + 76, + 89, + 472, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt-embedding-based technique (PET). To better incorporate common sense from the vision language model, we further design a novel PET for label generation. For object prediction, the decoder takes the segment features and the image mask $\\mathbf{m}_i^k$ , using a prompt", + "bbox": [ + 75, + 228, + 468, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a photo of [ENT]", + "text_level": 1, + "bbox": [ + 202, + 314, + 341, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to guide the object generation, where the [ENT] token is expected to be the pseudo label $b_{i}^{k}$ . For relation prediction, the decoder takes the segment features and an image mask pair $(\\mathbf{m}_i^k,\\mathbf{m}_j^k)$ as input, using a prompt", + "bbox": [ + 75, + 333, + 468, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a photo of [SUB] and [OBJ] what is their relation [REL]", + "text_level": 1, + "bbox": [ + 161, + 404, + 383, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to guide the relation generation, where the [SUB] and [OBJ] tokens are embedded by the pseudo labels $b_{i}^{k}$ and $b_{j}^{k}$ , and the [REL] token is expected to be the relation predicate between $(b_{i}^{k}, b_{j}^{k})$ with $b_{i}^{k}$ as subject and $b_{j}^{k}$ as object in the text graph. To enhance relation generation, we further design three learnable positional embeddings $\\mathbf{f}_{sub}$ , $\\mathbf{f}_{obj}$ , $\\mathbf{f}_{region}$ for indicating the different regions in the segment features. Two cross-entropy losses $\\mathcal{L}_{ent}^{k}$ , $\\mathcal{L}_{rel}^{k}$ are used to supervise the generation of the [ENT] and [REL] tokens, maximizing the likelihood of the target label strings, respectively. More details are in the supplementary material.", + "bbox": [ + 75, + 439, + 472, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.6. Inference", + "text_level": 1, + "bbox": [ + 76, + 616, + 187, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During inference, the target concepts of object semantics $\\mathcal{C}_o$ and relation predicates $\\mathcal{C}_r$ are known. With an image $I$ as input and an inference stage index $k_{inf}$ specified, the region grouper first partitions $I$ into several candidate segments $\\{\\mathbf{s}_i^{k_{inf}}\\}_{i=1}^{H_{k_{inf}}}$ , which are then passed through the segment merger to obtain the similarity matrix $\\mathbf{Sim}_{k_{inf}}$ . We formulate the segment merging as a spectral clustering problem and perform the graph cut [38] on $\\mathbf{Sim}_{l_{inf}}$ for clustering. To improve the accuracy, we employ a matrix recovery method [25] to reduce the noise in $\\mathbf{Sim}_{l_{inf}}$ . In this step, the segments with similar semantics tend to be merged into the same cluster. For each cluster and each pair of clusters, the label generator uses a similar PET to generate the object semantics and the relation predicates. For every label within sets $\\mathcal{C}_o$ and $\\mathcal{C}_r$ , the label generator computes its generation probability. Subsequently, these probabilities are used to rank the concepts, selecting the most probable as the", + "bbox": [ + 75, + 638, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "final prediction. Note that between object and relation prediction, to convert semantic segmentation into instance segmentation, we identify each connected component in the semantic segmentation to be an instance, for simplicity. More details about inference are in the supplementary material.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments and Results", + "text_level": 1, + "bbox": [ + 498, + 179, + 733, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We train our model with the COCO Caption dataset [4], which involves 123,287 images with each labeled by 5 independent human-generated captions. Following the 2017 split, we use 118,287 images with their captions for training. We evaluate models with the Panoptic Scene Graph dataset [49] for its pixel-wise labeling as well as its high-quality object and relation annotation. We further merge the object semantics with ambiguities. After merging, 127 object semantics and 56 relation predicates are finally obtained for evaluation. More details about the datasets can be found in the supplementary material.", + "bbox": [ + 496, + 204, + 890, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocol and Metrics. Following all previous works in scene graph generation, we evaluate the quality of a generated scene graph by viewing it as a set of subject-predicate-object triplets. We evaluate models on two tasks: Visual Phrase Detection (PhrDet) and Scene Graph Detection (SGDet). PhrDet aims to detect the whole phrase of subject-Predicate-object with a union location of subject and object. It is considered to be correct if the phrase labels are all correct and the union location matches the ground truth with the intersection over union (IoU) greater than 0.5. SGDet further requires a more accurate location, i.e., the location of subject and object should match the ground truth with IoU greater than 0.5 respectively.", + "bbox": [ + 496, + 369, + 892, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use No-Graph-Constraint-X Recall@K (NXR@K, %) to measure the ability of generation. Recall@K computes the recall between the top-k generated triplets with the ground truth. No-Graph-Constraint-X indicates that at most X predicate labels could be predicted for each subject-object pair. Since some predicates defined in [49] are not exclusive, such as on and sitting on, NXR@K could be a more reasonable metric compared with Recall@K.", + "bbox": [ + 496, + 566, + 890, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We consider several baselines for Caption-to-PSG in the following experiments. Firstly, we design four baselines that strictly follow the constraints of Caption-to-PSG, where objects are located by bbox proposals generated by selective search [43]:", + "bbox": [ + 496, + 686, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Random is the most naive baseline where all object semantics and relation predicates are randomly predicted.", + "- Prior augments Random by performing label prediction based on the statistical priors in the training set.", + "- MIL performs the alignment between proposals and textual entities by multiple instance learning [30]. Similar to [58], it formulates the object label prediction as a classification problem in a large pre-built vocabulary, with" + ], + "bbox": [ + 500, + 771, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2844", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9203dd379636268710f0c1a5714ae2d11246c664bb209ace8265a5eba0959924.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative Comparison of Different Methods on Caption-to-PSG. 'Proposal' indicates how the method obtains bbox proposals. 'Target' indicates whether the concept sets for inference are known during training. 'Mode' indicates the mode used for evaluation." + ], + "table_body": "
MethodModePhrDetSGDet
ModelProposalTargetN3R50N3R100N5R50N5R100N3R50N3R100N5R50N5R100
SGGNLS-cDetector✓bbox9.6911.4510.2412.226.767.817.28.65
Randomxbbox0.020.030.020.030.010.020.020.03
PriorSelective Searchxbbox0.040.070.050.070.030.060.050.07
MILxbbox1.972.182.042.611.21.351.561.97
SGCLIPxbbox3.023.453.383.712.132.32.392.7
SGGNLS-oDetectorxbbox6.26.796.927.933.964.214.535.02
Ours-xmask8.289.169.0610.513.323.633.714.18
Ours-xbbox11.3712.7412.2414.374.294.774.825.48
", + "bbox": [ + 84, + 88, + 893, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "WordNet [31] employed during inference. The relation labels are predicted with statistical priors, similar to Prior.", + "bbox": [ + 83, + 311, + 467, + 343 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SGCLIP employs the pre-trained CLIP [34] to predict both object semantic labels and relation predicate labels.", + "Secondly, to further benchmark the performance of our framework, we set two additional baselines based on [58] by gradually removing the constraints of Caption-to-PSG:", + "- SGGNLS-o [58] extracts proposals with a detector [35] pre-trained on OpenImage [19]. It formulates the object and relation label prediction as a classification problem within a large pre-built vocabulary, with WordNet [31] employed during inference.", + "- SGGNLS-c [58] uses the same proposals as SGGNLS-o. In SGGNLS-c, the target concept sets for inference are known during training. It formulates the label prediction as a classification problem in these target concept sets." + ], + "bbox": [ + 76, + 348, + 468, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More design details are in the supplementary material.", + "bbox": [ + 96, + 580, + 455, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. Following GroupViT [48], we set $K = 2$ , $H_{1} = 64$ , and $H_{2} = 8$ for our region grouper. We leverage general pre-trained models for weight initialization. We employ the pre-trained GroupViT for the region grouper as well as $\\mathbf{Tfm}^T$ in the entity grounder, and the pre-trained BLIP [21] decoder for the label generator. During training, $\\mathbf{Tfm}^T$ and the label generator are frozen. During inference, we set $k_{inf} = 1$ . More implementation details can be found in the supplementary material.", + "bbox": [ + 75, + 597, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Main Results on Caption-to-PSG", + "text_level": 1, + "bbox": [ + 76, + 741, + 366, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. Our quantitative results on Caption-to-PSG are shown in Tab. 1. To make a fair comparison with bbox-based scene graphs generated by baselines, we evaluate our generated PSGs in both mask and bbox mode. For the latter, all masks in both prediction and ground truth are converted into bboxes (i.e., the mask area's enclosing rectangle) for evaluation, resulting in an easier setting than the former. The results show that our framework (Ours) significantly outperforms all the baselines under the same", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "constraints on both PhrDet (14.37 vs. 3.71 N5R100) and SGDet (5.48 vs. 2.7 N5R100). Our method also shows better results compared with SGGNLS-o on all metrics and all tasks (on PhrDet, 14.37 vs. 7.93 N5R100; on SGDet, 5.48 vs. 5.02 N5R100) although SGGNLS-o utilizes location priors by leveraging a pre-trained detector. The results demonstrate that our framework is more effective for learning a good panoptic structured scene understanding.", + "bbox": [ + 496, + 311, + 890, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. We provide typical qualitative results in Fig. 3 to further show our framework's effectiveness. Compared with SGGNLS-o, our framework has the following advantages. First, our framework is able to provide fine-grained semantic labels to each pixel in the image to reach a panoptic understanding, while SGGNLS-o can only provide sparse bboxes produced by the pre-trained detector. Note that categories with irregular shapes (e.g., trees in Fig. 3) are hard to be labeled precisely by bboxes. Second, compared with SGGNLS-o, our framework can generate more comprehensive object semantics and relation predicates, such as \"dry grass field\" and \"land at\" in Fig. 3, showing the open-vocabulary potential of our framework. More qualitative results are in the supplementary material.", + "bbox": [ + 496, + 435, + 892, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. OOD Robustness Analysis", + "text_level": 1, + "bbox": [ + 498, + 656, + 736, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further analyze another key advantage of our framework, i.e., the robustness in OOD cases. Since SGGNLS-c and SGGNLS-o both rely on a pre-trained detector to locate objects, their performance highly depends on whether object semantics in the scene are covered by the detector.", + "bbox": [ + 496, + 674, + 890, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the object semantics [19] covered by the detector, we split the ground truth triplets into an in-distribution (ID) set and an OOD set. For triplets within the ID set, both the subject and object semantics are covered, while for triplets in the OOD set, at least one of the semantics is not covered. As shown in Tab. 2, both SGGNLS-c and SGGNLS-o suffer a significant performance drop from the ID set to the OOD set. On the OOD set, the triplets can hardly be retrieved. However, our framework, with the ability of location learned from purely text descriptions, can", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2845", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5fd646a0764fe88146fa3c41513539928ff5b9acebc30c769b9610d834715363.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 93, + 233, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e89251f4d08ca6a339cd4f66c3dfd9afd8363fa94e6a3b3217ec9e15a4ed343.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 107, + 357, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c2c9e3f598c9d86058f817e83a877ac96010667f702a6bfb1f858c2ce578417b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 95, + 555, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba4b6199c1d58515ae730fc2e95c3dbd2a1f7a5d4bbec65696d9bc2150d4c5ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 95, + 700, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f7188de0a861be9c3a7b18d3c91db4c0e157d7d39c6f0db9974027b2b742017f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 94, + 872, + 171 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8f5e498031f6f214f84c89768d1d5d55761c6c8ecd8b5e37c5723d4a86f5dd27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 174, + 233, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/757c43a99d609870c03d35ef36d4c3aa5b76dfeb106ae0996af7516593eb7887.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 238, + 189, + 387, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9946d972a85ce5a277190f051c4a2ac8d255ca12edbee4611f3eaa379867cde2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 174, + 555, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd097e5a5a4cff4e736bf4ca5458115ca08768dc20062efc4ac13bc4fb6fd4db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 181, + 717, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3fb94a60134ede9c03df6ecf4d475950e28983c6dce3822db506f42c701ae58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 183, + 883, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0db44181bf9f88f3acf4aae6678debcd4f1a7e4aee7b9396ead8dc7579bf256c.jpg", + "image_caption": [ + "(a) Results of SGGNLS-o" + ], + "image_footnote": [], + "bbox": [ + 83, + 255, + 233, + 334 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3fcd7be67b3d6b4fefa7c213cb4887cc98865565b0092549333f3582d046ec37.jpg", + "image_caption": [ + "Figure 3. Qualitative Comparison between SGGNLS-o (a) and Ours (b). For each method, the results of object location are shown on the left, while the results of scene graph generation are shown on the right. For Ours, scene graphs predicted within the given concept sets are provided in the middle column, and scene graphs directly predicted through the auto-regressive generation (i.e., an open-vocabulary manner) in the label generator are additionally provided in the right column." + ], + "image_footnote": [], + "bbox": [ + 254, + 263, + 357, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/081cd78f56524175490a5712006eec3d1bb73aca015a857908e90d5bc81638b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 255, + 555, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/76979ea156d605c127316bcd9c315ff7e2a2bab5f7a79d86baaf2b89d4bc0158.jpg", + "image_caption": [ + "(b) Results of Ours" + ], + "image_footnote": [], + "bbox": [ + 573, + 250, + 717, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6358bff9e31c151ed9f639abb8b6f7d260924235a5251b491291fd88401901d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 248, + 887, + 334 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f064641e11259d5b630acd13a1f96da324906685465a0194940b8ef9e7960080.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SetModelTargetModePhrDetSGDet
N3R100N5R100N3R100N5R100
IDSGGNLS-c✓bbox16.7618.4810.4511.86
SGGNLS-oXbbox11.5513.647.138.47
OursXmask9.2710.453.283.76
OursXbbox13.3514.824.635.36
OODSGGNLS-c✓bbox0000
SGGNLS-oXbbox0.050.0600
OursXmask8.479.764.074.51
OursXbbox10.1811.695.235.72
", + "bbox": [ + 86, + 422, + 467, + 549 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/348ce7a6ea8ed39b82c146262caa4072f4848abb77315441b0c974afd6084865.jpg", + "table_caption": [ + "Table 2. Analysis on OOD Robustness. 'Set' indicates the triplet set used for evaluation." + ], + "table_footnote": [], + "table_body": "
Stage#SegCutPhrDetSGDet
N3R100N5R100N3R100N5R100
164×10.7311.393.183.51
164✓12.7414.374.775.48
28×9.2411.033.534.35
28✓6.788.452.463.21
", + "bbox": [ + 84, + 590, + 467, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "reach similar performance on both sets, which demonstrates the OOD robustness of our framework for PSG generation.", + "bbox": [ + 76, + 726, + 467, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 770, + 240, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct additional ablation studies to evaluate the effectiveness of our design choices. For all following experiments, we report N3R100 and N5R100 evaluated in bbox mode for simplicity. We answer the following questions. Q1: Does the explicit learning of merging in the segment merger helps provide better image segments? Q2: Is the generation-based label prediction better than the", + "bbox": [ + 75, + 794, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a4b7e0dc9db5d965a6d0fb1d610f4b73e68b2dbcf2bfa4a54eb884ecb24a4aaf.jpg", + "table_caption": [ + "Table 3. Ablation Study on the Segment Merger. 'Stage' indicates the grouping stage where image segments used for merging are from. $\\# \\mathrm{Seg}$ indicates the number of image segments. Cut indicates whether the graph-cut-based segment merging is applied." + ], + "table_footnote": [], + "table_body": "
Label PredictionModelPhrDetSGDet
N3R100N5R100N3R100N5R100
Cls + WordNet-8.829.362.362.72
GenRNN9.1210.442.653.07
Gen w/o PETBLIP [21]2.332.580.450.6
Gen w/ PETBLIP [21]12.6414.284.775.49
", + "bbox": [ + 503, + 422, + 890, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation Study on the Label Generator. 'Cls' indicates classification. 'Gen' indicates generation.", + "bbox": [ + 500, + 494, + 890, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "classification-based prediction? Q3: Does the pre-learned common sense from the pre-trained BLIP [21] helps with the label prediction? Q4: Does the PET helps incorporate the pre-learned common sense for label prediction?", + "bbox": [ + 496, + 532, + 890, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 3, we compare different strategies of image segment merging during inference. Row 1&2 denote that the $H_{1} = 64$ segments from the first grouping stage are used for further merging, while row 3&4 denote that the $H_{2} = 8$ segments from the second stage are used. The results show that applying the graph cut to merge the segments from the first stage could reach the best performance, corroborating that compared with the fully implicit learning of merging, the explicit learning of merging can provide better segments (row 2 vs 3, answering Q1).", + "bbox": [ + 496, + 595, + 890, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 4, we compare different designs of the label generator. Keeping the other modules the same, we change the label generator (row 4) into three different designs, i.e., classification within a large pre-built vocabulary followed by WordNet [31] for target matching (row 1), generation with a vanilla RNN (row 2), generation with the BLIP decoder but without the PET (row 3). The results show that with the constraint of no pre-defined concept sets, compared with formulating the label prediction into a classification problem, formulating it into a generation problem is a bet-", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2846", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ter choice (row 1 vs 2&4, answering Q2). By employing the pre-trained BLIP for leveraging the pre-learned common sense, the prediction could be further boosted (row 2 vs 4, answering Q3). And the PET is very important for incorporating the common sense from the pre-trained model (row 3 vs 4, answering Q4).", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "More ablation studies for the design evaluation can be found in the supplementary material.", + "bbox": [ + 76, + 181, + 468, + 213 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.4. Application on TSSS", + "text_level": 1, + "bbox": [ + 76, + 222, + 272, + 239 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As a side product, we observe that our entity grounder and segment merger can also enhance TSSS. Based on the original GroupViT [48], we replace the multi", + "bbox": [ + 75, + 247, + 243, + 351 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/ee0a269e928c9212d8d1a757fc9dac593efa6df6604ea30ed35bb3e51722ec61.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmIoU
GroupViT [48]24.28
GroupViT† [48]24.72
Ours26.87
", + "bbox": [ + 282, + 258, + 460, + 319 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 5. Results on TSSS. † indicates finetuned.", + "bbox": [ + 272, + 320, + 467, + 345 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "label contrastive loss with our entity grounder and segment merger. Then we finetune the model on the COCO Caption dataset [4]. As shown in Tab. 5, compared with GroupViT directly finetuned on [4], the explicit learning of merging in our modules can boost the model with an absolute $2.15\\%$ improvement of mean Intersection over Union (mIoU, $\\%$ ) on COCO [23], which demonstrates the effectiveness of our proposed modules on better object location.", + "bbox": [ + 75, + 352, + 468, + 473 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 483, + 194, + 498 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Failure Case Analysis. Despite the impressive performance of TextPSG, there are still challenges to address. Upon analyzing the failure cases for PSG generation, we identify three specific limitations of TextPSG that contribute to these failures. a) The strategy we use to convert semantic segmentation into instance segmentation is not entirely effective. For simplicity, in TextPSG, we identify each connected component in the semantic segmentation to be an individual object instance. However, this strategy may fail when instances overlap or are occluded, resulting in either an underestimation or an overestimation of instances. b) Our framework faces difficulty in locating small objects in the scene due to limitations in resolution and the grouping strategy for location. c) The relation prediction of our framework requires enhancement, as it is not adequately conditioned on the image. While the label generator uses both image features and predicted object semantics to determine the relation, it sometimes seems to lean heavily on the object semantics, potentially neglecting the actual image content. Examples of failure cases for each of these limitations can be found in the supplementary material.", + "bbox": [ + 75, + 507, + 468, + 824 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Model Diagnosis. For a clearer understanding of the efficacy of our framework, we conduct a model diagnosis to answer the following question: why does our framework only achieve semantic segmentation through learning, rather than panoptic segmentation (and thus requires further", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "segmentation conversion to obtain instance segmentation)?", + "bbox": [ + 500, + 90, + 888, + 106 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c626cb00f1fd940999acb0183a3d1210f60281923e2db9d49a498916ba1926eb.jpg", + "image_caption": [ + "(a) A white sheep and a black sheep standing on the grass under the sky.", + "Figure 4. Region-Entity Alignment Results of Captions in Different Granularity. Two captions in different granularity are used to execute region-entity alignment with the same image, with (a) one describing the two sheep individually while (b) the other merges them in plural form." + ], + "image_footnote": [], + "bbox": [ + 516, + 119, + 629, + 215 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ddc720bb97a834e199c54994b057015092868968038267bc7df10d73c4c98e9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 119, + 746, + 215 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8feeeffbb255dd913d4f772a2159f22072d1a7bb396920dba6d87e54cb7da40c.jpg", + "image_caption": [ + "(b) Pair of sheep standing on open grassy field on clear day." + ], + "image_footnote": [], + "bbox": [ + 753, + 119, + 870, + 215 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Fig. 4, we use two captions in different granularity to execute region-entity alignment. It shows that our framework has the capability to assign distinct masks to individual instances. However, the nature of caption data, where captions often merge objects of the same semantics in plural form, limits our framework from differentiating instances. It is the weak supervision provided by the caption data that constrains our framework.", + "bbox": [ + 496, + 321, + 890, + 441 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "More diagnoses are in the supplementary material.", + "bbox": [ + 517, + 443, + 852, + 458 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Future Directions. In response to the limitations discussed, we outline several potential directions for enhancing our framework: a) a refined and sophisticated strategy for segmentation conversion; b) increasing the input resolution, though this may introduce greater computational demands; c) a more suitable image-conditioned reasoning mechanism for relation prediction; d) a superior image-caption-pair dataset with more detailed granularity in captions to achieve panoptic segmentation through learning.", + "bbox": [ + 496, + 458, + 890, + 594 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 603, + 617, + 618 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We take the first step towards the novel problem Caption-to-PSG, aiming to learn PSG generation purely from language. To tackle this challenging problem, we propose a new modularized framework TextPSG with several novel techniques, which significantly outperforms the baselines and achieves strong OOD robustness. This paves the path to a more comprehensive and generalizable panoptic structured scene understanding. There are still bottle-necks in TextPSG to be explored in future work, including a) a more sophisticated strategy for segmentation conversion; b) a more suitable image-conditioned reasoning mechanism for relation prediction; c) a superior image-caption-pair dataset for panoptic segmentation through learning.", + "bbox": [ + 496, + 622, + 890, + 819 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the DSO grant DSOCO21072, IBM, and gift funding from MERL, Cisco, and Amazon. We would also like to thank the computation support from AiMOS, a server cluster for the IBM Research AI Hardware Center.", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2847", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Somak Aditya, Yezhou Yang, Chitta Baral, Yiannis Aloimonos, and Cornelia Fermüller. Image understanding using vision and reasoning through scene description graph. Computer Vision and Image Understanding, 2017. 1", + "[2] S. Amiri, Kishan Chandan, and Shiqi Zhang. Reasoning with scene graphs for robot planning under partial observability. IEEE Robotics and Automation Letters, 7:5560-5567, 2022. 1", + "[3] Gabor Angeli, Melvin Johnson, and Christopher D. Manning. Leveraging linguistic structure for open domain information extraction. In Annual Meeting of the Association for Computational Linguistics, 2015. 4", + "[4] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6, 9", + "[5] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee K Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. In Proc. 57th Annual Meeting of the Association for Computational Linguistics, 2019. 3", + "[6] Zhenfang Chen, Jiayuan Mao, Jiajun Wu, Kwan-Yee K Wong, Joshua B. Tenenbaum, and Chuang Gan. Grounding physical concepts of objects and events through dynamic visual reasoning. In International Conference on Learning Representations, 2021. 3", + "[7] Zhenfang Chen, Peng Wang, Lin Ma, Kwan-Yee K Wong, and Qi Wu. Cops-ref: A new dataset and task on compositional referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10086-10095, 2020. 3", + "[8] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D. Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In CVPR, 2020. 1", + "[9] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3", + "[10] Samir Yitzhak Gadre, Kiana Ehsani, Shuran Song, and Roozbeh Mottaghi. Continuous scene representations for embodied ai. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14829-14839, 2022. 1", + "[11] Chuang Gan, Yandong Li, Haoxiang Li, Chen Sun, and Boqing Gong. Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1811-1820, 2017. 3", + "[12] Jiumiang Gu, Handong Zhao, Zhe Lin, Sheng Li, Jianfei Cai, and Mingyang Ling. Scene graph generation with external knowledge and image reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 3", + "[14] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In CVPR, 2018. 1", + "[15] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3128-3137, 2015. 3", + "[16] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 3", + "[17] Siddhesh Khandelwal, Mohammed Suhail, and Leonid Sigal. Segmentation-grounded scene graph generation. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15859-15869, 2021. 3", + "[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Fei-Fei Li. Visual Genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32-73, 2016. 3", + "[19] Alina Kuznetsova, Hassan Rom, Neil Gordon Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4. International Journal of Computer Vision, 128:1956-1981, 2018. 7", + "[20] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022. 3", + "[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 5, 7, 8", + "[22] Xingchen Li, Long Chen, Wenbo Ma, Yi Yang, and Jun Xiao. Integrating object-aware and interaction-aware knowledge for weakly supervised scene graph generation. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2, 3", + "[23] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 9", + "[24] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3", + "[25] Guangcan Liu, Zhouchen Lin, Shuicheng Yan, Ju Sun, Yong Yu, and Yi Ma. Robust recovery of subspace structures by low-rank representation. IEEE transactions on pattern analysis and machine intelligence, 35(1):171-184, 2012. 6" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2848", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7086-7096, 2022. 3", + "[27] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. arXiv preprint arXiv:2211.14813, 2022. 3", + "[28] Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David McClosky. The Stanford CoreNLP natural language processing toolkit. In Association for Computational Linguistics (ACL) System Demonstrations, pages 55–60, 2014. 4", + "[29] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 3", + "[30] Oded Maron and Tomás Lozano-Pérez. A framework for multiple-instance learning. Advances in neural information processing systems, 10, 1997. 6", + "[31] George A. Miller. WordNet: A lexical database for English. In Speech and Natural Language: Proceedings of a Workshop Held at Harriman, New York, February 23-26, 1992, 1992. 5, 7, 8", + "[32] Julia Peyre, Ivan Laptev, Cordelia Schmid, and Josef Sivic. Weakly-supervised learning of visual relations. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5189-5198, 2017. 3", + "[33] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649, 2015. 3", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763, 2021. 7", + "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2, 3, 7", + "[36] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 817-834. Springer, 2016. 3", + "[37] Sebastian Schuster, Ranjay Krishna, Angel X. Chang, Li Fei-Fei, and Christopher D. Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In VL@EMNLP, 2015. 3", + "[38] Jianbo Shi and J. Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000. 6" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8376-8384, 2019. 1", + "[40] Jing Shi, Yiwu Zhong, Ning Xu, Yin Li, and Chenliang Xu. A simple baseline for weakly-supervised scene graph generation. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16373-16382, 2021. 3", + "[41] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3", + "[42] Damien Teney, Lingqiao Liu, and Anton van den Hengel. Graph-structured representations for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1", + "[43] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104:154-171, 2013. 3, 6", + "[44] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 4", + "[45] Qixun Wang, Xiaofeng Guo, and Haofan Wang. 1st place solution for psg competition with eccv'22 sensehuman workshop. ArXiv, abs/2302.02651, 2023. 2", + "[46] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6602–6611, 2019. 3", + "[47] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5419, 2017. 3", + "[48] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. GroupViT: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18134-18144, June 2022. 2, 3, 4, 5, 7, 9", + "[49] Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. In ECCV, 2022. 1, 2, 3, 6", + "[50] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 3", + "[51] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. FILIP: Fine-grained interactive language-image pre-training. In International Conference on Learning Representations, 2022. 2, 4, 5", + "[52] Keren Ye and Adriana Kovashka. Linguistic structures as weak supervision for visual scene graph generation. In Pro-" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2849", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 2, 3", + "[53] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1307-1315, 2018. 3", + "[54] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3", + "[55] Alireza Zareian, Svebor Karaman, and Shih-Fu Chang. Weakly supervised visual semantic parsing. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3733-3742, 2020. 3", + "[56] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 3", + "[57] Hanwang Zhang, Zawlin Kyaw, Jinyang Yu, and Shih-Fu Chang. Ppr-fcn: Weakly supervised visual relation detection via parallel pairwise r-fcn. In Proceedings of the IEEE international conference on computer vision, pages 4233-4241, 2017. 3", + "[58] Yiwu Zhong, Jing Shi, Jianwei Yang, Chenliang Xu, and Yin Li. Learning to generate scene graph from natural language supervision. In ICCV, 2021. 2, 3, 5, 6, 7", + "[59] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVIII, pages 696-712. Springer, 2022. 3" + ], + "bbox": [ + 78, + 92, + 468, + 583 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2850", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/a412d306-b042-443c-958a-505db74fd5e7_model.json b/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/a412d306-b042-443c-958a-505db74fd5e7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ca0b336998e2aefc6bd6db529e0d0e3b0be653c5 --- /dev/null +++ b/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/a412d306-b042-443c-958a-505db74fd5e7_model.json @@ -0,0 +1,2732 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.131, + 0.846, + 0.155 + ], + "angle": 0, + "content": "TextPSG: Panoptic Scene Graph Generation from Textual Descriptions" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.18, + 0.705, + 0.253 + ], + "angle": 0, + "content": "Chengyang Zhao1 Yikang Shen2 Zhenfang Chen2 Mingyu Ding3 Chuang Gan2,4 \n1Peking University 2MIT-IBM Watson AI Lab \n3UC Berkley 4UMass Amherst" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.474, + 0.786 + ], + "angle": 0, + "content": "Panoptic Scene Graph has recently been proposed for comprehensive scene understanding. However, previous works adopt a fully-supervised learning manner, requiring large amounts of pixel-wise densely-annotated data, which is always tedious and expensive to obtain. To address this limitation, we study a new problem of Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG). The key idea is to leverage the large collection of free image-caption data on the Web alone to generate panoptic scene graphs. The problem is very challenging for three constraints: 1) no location priors; 2) no explicit links between visual regions and textual entities; and 3) no predefined concept sets. To tackle this problem, we propose a new framework TextPSG consisting of four modules, i.e., a region grouper, an entity grounder, a segment merger, and a label generator, with several novel techniques. The region grouper first groups image pixels into different segments and the entity grounder then aligns visual segments with language entities based on the textual description of the segment being referred to. The grounding results can thus serve as pseudo labels enabling the segment merger to learn the segment similarity as well as guiding the label generator to learn object semantics and relation predicates, resulting in a fine-grained structured scene understanding. Our framework is effective, significantly outperforming the baselines and achieving strong out-of-distribution robustness. We perform comprehensive ablation studies to corroborate the effectiveness of our design choices and provide an in-depth analysis to highlight future directions. Our code, data, and results are available on our project page: https://vis-ww.c.s.umass.edu/TextPSG." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.815, + 0.21, + 0.831 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "A scene graph is a directed-graph-based abstract representation of the objects and their relations within a scene. It has been widely utilized to develop a structured scene understanding of object semantics, locations, and relations," + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.285, + 0.625, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.286, + 0.89, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.461, + 0.893, + 0.572 + ], + "angle": 0, + "content": "Figure 1. Problem Overview. Different from the traditional bbox-based form of the scene graph as shown in (a), Caption-to-PSG aims to generate the mask-based panoptic scene graph. In Caption-to-PSG, the model has no access to any location priors, explicit region-entity links, or pre-defined concept sets. Consequently, the model is required to learn partitioning and grounding as illustrated in (b), as well as object semantics and relation predicates as illustrated in (c), all purely from textual descriptions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.629 + ], + "angle": 0, + "content": "which facilitates a variety of downstream applications, such as image generation [14, 8], visual reasoning [42, 1, 39], and robotics [2, 10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Typically, in a scene graph, each node denotes an object in the scene located by a bounding box (bbox) with a semantic label, and each directed edge denotes the relation between a pair of objects with a predicate label. Nonetheless, a recent work [49] points out that such a bbox-based form of scene graph is not ideal enough. Firstly, compared with pixel-wise segmentation masks, bboxes are less fine-grained and may contain some noisy pixels belonging to other objects, limiting the applications for some downstream tasks. For example, as shown in Fig. 1 (a), about half of the pixels in the yellow_bbox for girl belong to wall. Secondly, it is challenging for bboxes to cover the entire scene without ambiguities caused by overlaps, which prevents a scene graph from including every object in the scene for a complete description. To this end, the work [49] proposes the concept of Panoptic Scene Graph (PSG), in which each object is grounded by a panoptic segmentation mask, to reach a comprehensive structured scene representation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2839" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "However, all existing works [49, 45] approach PSG generation through a fully-supervised manner, i.e., learning to perform panoptic segmentation and relation prediction from manually-annotated datasets with explicit supervision for both segmentation and relation prediction. Unfortunately, it is extremely labor-intensive to build such datasets, making it difficult to scale up to cover more complex scenes, object semantics, and relation predicates, thus significantly limiting the generalizability and the application of these methods to the real world. For instance, the current PSG dataset [49] only covers 133 object semantics and 56 relation predicates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.442 + ], + "angle": 0, + "content": "To relieve the reliance on densely-annotated data, weakly-supervised methods [52, 58, 22] for scene graph generation are promising. These methods could induce scene graphs from image-caption pairs, which can be easily harvested from the Web for free. Even so, they still rely heavily on two strong preconditions, i.e., a powerful region proposal network (e.g., [35]) and a pre-defined set of object semantics and relation predicates. Although these preconditions facilitate the learning process of the methods, they also limit the generalizability for locating novel objects (unforeseen objects for the region proposal network) and constrain the understanding into the limited concept set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.473, + 0.656 + ], + "angle": 0, + "content": "Inspired by previous weakly-supervised methods, we introduce a new problem, Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG), to explore a holistic structured scene understanding without labor-intensive data annotation. Considering the limitation of the preconditions mentioned, we set three constraints to Caption-to-PSG to reach a more comprehensive and generalizable understanding, which results in a very challenging problem: a) only image-caption pairs are provided during training, without any location priors in either region proposals or location supervision; b) the explicit links between regions in images and entities in captions are missing; c) no concept sets are pre-defined, i.e., neither object semantics nor relation predicates are known beforehand." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Given these three constraints, we argue that there are two key challenges for the model to solve the problem. Firstly, the model should learn to ground entities in language onto the visual scene without explicit location supervision, i.e., the ability to perform partitioning and grounding, as shown in Fig. 1 (b), should be developed purely from textual descriptions. Secondly, during training, the model should also learn the object semantics and relation predicates from textual descriptions, as shown in Fig. 1 (c), without pre-defined fixed object and relation vocabularies. By solving these challenges, the model could associate visual scene patterns with textual descriptions, gradually acquire common sense among them, and finally reach a more comprehensive and generalizable understanding, including novel object location, extensive semantics recognition, and complex relation analysis, which is more suitable to the real world." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.574 + ], + "angle": 0, + "content": "With these considerations, we propose a novel framework, TextPSG, as the first step towards this challenging problem. TextPSG consists of a series of modules to cooperate with each other, i.e., a region grouper, an entity grounder, a segment merger, and a label generator. The region grouper learns to merge image regions into several segments in a hierarchical way based on object semantics, similar to [48]. The entity grouder employs a fine-grained contrastive learning strategy [51] to bridge the textual description and the visual content, grounding entities in the caption onto the image segments. With the entity-grounding results as pseudo labels, the segment merger learns similarity matrices to merge small image segments during inference, while the label generator learns the prediction of object semantics and relation predicates. Specifically, in the segment merger, we propose to leverage the grounding as explicit supervision for merging, compared with [48] which learns merging in a fully implicit manner, to improve the ability of location. In the label generator, different from all previous pipelines for scene graph generation, we reformulate the label prediction as an auto-regressive generation problem rather than a classification problem, and employ a pre-trained language model [21] as the decoder to leverage the pre-learned common sense. We further design a novel prompt-embedding-based technique (PET) to better incorporate common sense from the language model. Our experiments show that TextPSG significantly outperforms the baselines and achieves strong out-of-distribution (OOD) robustness. Comprehensive ablation studies corroborate the effectiveness of our design choices. As a side product, the proposed grouder and merger modules also have been observed to enhance text-supervised semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.895, + 0.635 + ], + "angle": 0, + "content": "In spite of the promising performance of TextPSG, certain challenges persist. We delve into an in-depth analysis of the failure cases, provide a model diagnosis, and discuss potential future directions for enhancing our framework." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.636, + 0.814, + 0.65 + ], + "angle": 0, + "content": "To sum up, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.65, + 0.892, + 0.711 + ], + "angle": 0, + "content": "- We introduce a new problem, Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG), to alleviate the burden of human annotation for PSG by learning purely from the weak supervision of captions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.714, + 0.892, + 0.789 + ], + "angle": 0, + "content": "- We propose a new modularized framework, TextPSG, with several novel techniques, which significantly outperforms the baselines and achieves strong OOD robustness. We demonstrate that the proposed modules in TextPSG can also facilitate text-supervised semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.793, + 0.892, + 0.823 + ], + "angle": 0, + "content": "- We perform an in-depth failure case analysis with a model diagnosis, and further highlight future directions." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.65, + 0.892, + 0.823 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.642, + 0.847 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Bbox-based Scene Graph Generation. Bbox-based scene graph generation aims to create a structured representation of object semantics, locations, and relations in the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2840" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "scene, where each object is identified by a bbox. Most of existing works [50, 47, 41, 12, 24] follow a fully-supervised approach to learn the generation from densely-annotated datasets [18, 13], which requires significant human labors. To reduce the labeling effort, some weakly-supervised methods have been proposed [32, 55, 57, 40]. Recent works [52, 58, 22] further explore learning scene graph generation from image-caption pairs. However, they all rely on off-the-shelf region proposal networks for the location of objects in the scene, which are typically pretrained on pre-defined fixed sets of object semantics, limiting their generalizability to locating unforeseen objects. To reach a more granular and accurate grounding, [17] proposes to ground each object by segmentation. A recent work [49] further introduces the concept of PSG, where each object is identified by a panoptic segmentation mask, as a more comprehensive scene representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.544 + ], + "angle": 0, + "content": "Text-supervised Semantic Segmentation (TSSS). TSSS [48, 20, 26, 9, 27, 59] aims to learn image pixel semantic labeling from image-caption pairs without fine-grained pixel-wise annotations. Similar to TSSS, our proposed Caption-to-PSG aims to learn to connect visual regions and textual entities from only image-caption pairs and has the potential to leverage the large collection of free data on the Web. However, different from TSSS, Caption-to-PSG further requires the model to learn the relations among different visual regions, resulting in a higher-order structured understanding of visual scenes. In addition to unknown object semantics, Caption-to-PSG does not assume any pre-defined relation predicate concepts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.473, + 0.758 + ], + "angle": 0, + "content": "Visual Grounding. Our work is also related to visual grounding [16, 56, 29, 11, 6, 33], which grounds entities in language onto objects in images. Early works [54, 53, 7] on visual grounding typically detect object proposals [35, 43] from images first and then match them with language descriptions by putting features of both modalities into the same feature space, which are in a fully-supervised learning manner. There are also some weakly-supervised grounding methods [15, 36, 5] which relieve the need for dense regional annotations by multiple instance learning [15] or learning to reconstruct [36]. Different from them, Caption-to-PSG is more challenging since it requires grounding fine-grained object relations between entities without region proposal networks for a pre-defined object vocabulary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.285, + 0.787 + ], + "angle": 0, + "content": "3. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.904 + ], + "angle": 0, + "content": "Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG). A PSG \\(\\mathcal{G} = (\\mathcal{V},\\mathcal{E})\\) is a directed graph representation of the objects and the relations among them in a scene image \\(I\\in \\mathbb{R}^{H\\times W\\times 3}\\). Each node \\(v_{i}\\in \\mathcal{V}\\) denotes an object in \\(I\\) located by a panoptic segmentation mask \\(m_i\\in \\{0,1\\}^{H\\times W}\\) with an object semantic label \\(o_i\\in \\mathcal{C}_o\\), and each directed edge \\(e_{ij}\\in \\mathcal{E}\\) denotes a" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.892, + 0.17 + ], + "angle": 0, + "content": "pair of subject \\( o_i \\) and object \\( o_j \\) with a relation predicate label \\( r_{ij} \\in \\mathcal{C}_r \\), where \\( \\mathcal{C}_o \\) and \\( \\mathcal{C}_r \\) are the defined concept sets of object semantics and relation predicates. Note that for a PSG, it is constrained that all segmentation masks could not overlap, i.e., \\( \\sum_{i=1}^{|\\mathcal{V}|} m_i \\leq 1^{H \\times W} \\)." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.17, + 0.894, + 0.276 + ], + "angle": 0, + "content": "Given a large collection of paired scene images and textual descriptions \\( S = \\{(I_i, T_i)\\}_i \\), Caption-to-PSG aims to learn PSG generation from purely text descriptions for a holistic structured scene understanding, i.e., during training, only \\( S \\) is provided as supervision, while during inference, with a scene observation \\( I' \\) as input, the model is required to generate a corresponding PSG \\( \\mathcal{G}' \\)." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.277, + 0.894, + 0.444 + ], + "angle": 0, + "content": "Three Constraints. Note that in Caption-to-PSG, three important constraints are set to reach a more comprehensive and generalizable scene understanding: a) no location priors: different from all previous scene graph generation methods, neither pre-trained region proposal networks nor location supervision are allowed; b) no explicit region-entity links: the links between regions in the image \\(I\\) and entities in the textual description \\(T\\) are not provided; c) no pre-defined concept sets: the target concept sets defined for inference, i.e., object semantics \\(\\mathcal{C}_o\\) and relation predicates \\(\\mathcal{C}_r\\), are unknown during training." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.459, + 0.591, + 0.475 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.485, + 0.892, + 0.546 + ], + "angle": 0, + "content": "Overview. As illustrated in Fig. 2, our proposed framework for Caption-to-PSG, TextPSG, contains four modules to cooperate with each other: a region grouper, an entity grounder, a segment merger, and a label generator." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.547, + 0.894, + 0.713 + ], + "angle": 0, + "content": "During training, TextPSG takes batched image-caption pairs as input. For each pair, the image is passed through the region grouper to be partitioned into several image segments, while the caption is first pre-processed to extract its linguistic structure as a text graph and then taken by the entity grounder to ground textual entities in the graph onto the image segments. With the grounding results as pseudo labels, the segment merger learns similarity matrices between small image segments for further merging during inference, while the label generator learns the prediction of object semantics and relation predicates." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.714, + 0.892, + 0.805 + ], + "angle": 0, + "content": "During inference, for each input image, the image segments output from the region grouper are directly passed to the segment merger to be further merged according to the learned similarity matrices, and then fed to the label generator to predict the object semantic labels and the relation predicate labels." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.816, + 0.738, + 0.833 + ], + "angle": 0, + "content": "4.1. Text Graph Preprocessing" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.84, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Following previous methods [52, 58, 22] that leverage a rule-based language parser [46] based on [37] to preprocess textual descriptions, in TextPSG, we employ the same parser to extract linguistic structures from captions. Addi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2841" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.369, + 0.895, + 0.44 + ], + "angle": 0, + "content": "Figure 2. Framework Overview of TextPSG. The framework consists of four modules cooperating with each other: a region grouper to merge regions in the input image into several segments, an entity grounder to ground entities in the caption onto the image segments, a segment merger to learn similarity matrices to merge small image segments during inference, and a label generator to learn the prediction of object semantics and relation predicates. The solid arrows indicate the training flow, while the dash arrows indicate the inference flow. The arrows from the region grouper to the label generator indicating the segment feature and mask query are omitted." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.473, + 0.562 + ], + "angle": 0, + "content": "tionally, inspired by the success of open information extraction (OpenIE) [3] in natural language processing, we also employ an OpenIE system from Stanford CoreNLP [28] for extraction as a supplement. After merging, for each caption, we obtain its linguistic structure represented in a text graph, where each node denotes an entity, and each directed edge denotes the relation between an entity pair." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.569, + 0.24, + 0.586 + ], + "angle": 0, + "content": "4.2. Region Grouper" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.47, + 0.638 + ], + "angle": 0, + "content": "With a scene image as input, the region grouper aims to merge the regions with similar object semantics into several segments and extract corresponding high-level features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.471, + 0.806 + ], + "angle": 0, + "content": "Our region grouper follows the hierarchical design of GroupViT [48]. Given an input image, the grouper first splits the image into \\(N\\) non-overlapping patches as the initial image segments \\(\\{\\mathbf{s}_i^0\\}_{i = 1}^N\\). These segments are then passed through \\(K\\) grouping layers, where they are merged into larger, arbitrary-shaped segments progressively. Specifically, within each grouping layer \\(\\mathbf{Grp}_k(k = 1,2,\\dots ,K)\\), \\(H_{k}\\) grouping centers \\(\\{\\mathbf{c}_i^k\\}_{i = 1}^{H_k}\\) could be learned. The grouping operation is performed through an attention mechanism between the centers and the segments, merging \\(H_{k - 1}\\) input segments into \\(H_{k}\\) larger ones, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.813, + 0.416, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {s} _ {i} ^ {k} \\right\\} _ {i = 1} ^ {H _ {k}} = \\mathbf {G r p} _ {k} \\big (\\left\\{\\mathbf {c} _ {i} ^ {k} \\right\\} _ {i = 1} ^ {H _ {k}}, \\left\\{\\mathbf {s} _ {i} ^ {k - 1} \\right\\} _ {i = 1} ^ {H _ {k - 1}} \\big).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Note that \\( H_0 = N \\). After the hierarchical grouping, multiple groups of segments \\( \\{\\mathbf{s}_i^k\\}_{i=1}^{H_k} \\) at different grouping stages are obtained. More details about the design of \\( \\{\\mathbf{Grp}_k\\}_{k=1}^K \\) can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.667, + 0.47 + ], + "angle": 0, + "content": "4.3. Entity Grounder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.892, + 0.571 + ], + "angle": 0, + "content": "Since the explicit region-entity links are not provided, bridging the textual description and the visual content automatically plays an important role in solving Caption-to-PSG. Inspired by FILIP [51], in TextPSG, we employ a similar fine-grained contrastive learning strategy to perform region-entity alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.573, + 0.893, + 0.787 + ], + "angle": 0, + "content": "For each grouping stage \\(k\\), on image side, the grounder projects the segment group \\(\\{\\mathbf{s}_i^k\\}_{i=1}^{H_k}\\) into a new feature space \\(\\mathcal{F}\\) by a multi-layer perceptron (MLP) \\(\\mathbf{Proj}_k^I\\) to obtain segment embeddings \\(\\{\\mathbf{x}_i^k\\}_{i=1}^{H_k}\\). On text side, the input caption is first tokenized into \\(M\\) tokens \\(\\{\\mathbf{t}_i\\}_{i=1}^M\\), which are then processed by a Transformer [44] \\(\\mathbf{Tfm}^T\\) to propagate information between each other. A recurrent neural network (RNN) \\(\\mathbf{Rnn}\\) further merges the tokens corresponding to the same entity, encoding tokens into their associated weights one by one and utilizing weighted sum to merge the token features into a singular entity feature. Finally, these entity features are projected to the same feature space \\(\\mathcal{F}\\) by a MLP \\(\\mathbf{Proj}^T\\) to obtain entity embeddings \\(\\{\\mathbf{y}_i\\}_{i=1}^E\\), where \\(E\\) denotes the number of entities in the caption." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.788, + 0.892, + 0.865 + ], + "angle": 0, + "content": "With the segment embeddings and the entity embeddings in the shared feature space \\(\\mathcal{F}\\), we compute their token-wise similarities. Specifically, for the \\(i\\)-th segment, we compute its cosine similarities with all entities to obtain the token-wise similarity from the \\(i\\)-th segment to the caption \\(p_i^k\\) via" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.879, + 0.78, + 0.904 + ], + "angle": 0, + "content": "\\[\np _ {i} ^ {k} = \\max _ {1 \\leq j \\leq E} \\cos [ \\mathbf {x} _ {i} ^ {k}, \\mathbf {y} _ {j} ],\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2842" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "where \\(\\cos[\\cdot, \\cdot]\\) denotes the cosine similarity operation. Note that different from the original FILIP [51], in the scenario of region-entity alignment, some regions in the scene may not be described in the caption, while some entities in the caption may not exist in the scene. To tackle this problem, we propose to set a filtering threshold \\(\\theta\\), where pairs with similarity lower than \\(\\theta\\) will be considered in different semantics and filtered out. The fine-grained similarity from the image to the caption \\(p^k\\) can thus be computed via" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.236, + 0.395, + 0.279 + ], + "angle": 0, + "content": "\\[\np ^ {k} = \\frac {1}{\\sum_ {i = 1} ^ {H _ {k}} {\\mathbf {1} _ {p _ {i} ^ {k} > \\theta}}} \\sum_ {i = 1} ^ {H _ {k}} (p _ {i} ^ {k} \\cdot \\mathbf {1} _ {p _ {i} ^ {k} > \\theta}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.285, + 0.47, + 0.317 + ], + "angle": 0, + "content": "Similarly, we can also compute the token-wise similarity from the \\(j\\)-th entity to the image \\(q_j^k\\) via" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.324, + 0.361, + 0.35 + ], + "angle": 0, + "content": "\\[\nq _ {j} ^ {k} = \\max _ {1 \\leq i \\leq H _ {k}} \\cos [ \\mathbf {x} _ {i} ^ {k}, \\mathbf {y} _ {j} ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.469, + 0.388 + ], + "angle": 0, + "content": "and the fine-grained similarity from the caption to the image \\( q^{k} \\) via" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.388, + 0.395, + 0.429 + ], + "angle": 0, + "content": "\\[\nq ^ {k} = \\frac {1}{\\sum_ {j = 1} ^ {E} \\mathbf {1} _ {q _ {j} ^ {k} > \\theta}} \\sum_ {j = 1} ^ {E} (q _ {j} ^ {k} \\cdot \\mathbf {1} _ {q _ {j} ^ {k} > \\theta}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.432, + 0.47, + 0.53 + ], + "angle": 0, + "content": "Denoting the training batch with batch size \\(B\\) as \\(\\{(I_i,T_i)\\}_{i = 1}^B\\), the fine-grained similarity from the image \\(I_{i}\\) to the caption \\(T_{j}\\) as \\(p^{k,i\\rightarrow j}\\) and from the caption \\(T_{j}\\) to the image \\(I_{i}\\) as \\(q^{k,j\\rightarrow i}\\), the image-to-text fine-grained contrastive loss \\(\\mathcal{L}_{fine}^{k,I\\to T}\\) and the text-to-image fine-grained contrastive loss \\(\\mathcal{L}_{fine}^{k,T\\to I}\\) can then be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.538, + 0.41, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f i n e} ^ {k, I \\rightarrow T} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\frac {\\exp \\left(p ^ {k , i \\rightarrow i} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(p ^ {k , i \\rightarrow j} / \\tau\\right)},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.603, + 0.41, + 0.659 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f i n e} ^ {k, T \\rightarrow I} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\frac {\\exp \\left(q ^ {k , i \\rightarrow i} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(q ^ {k , i \\rightarrow j} / \\tau\\right)},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.663, + 0.47, + 0.692 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a learnable temperature. The total fine-grained contrastive loss is" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.698, + 0.382, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f i n e} ^ {k} = \\frac {1}{2} (\\mathcal {L} _ {f i n e} ^ {k, I \\rightarrow T} + \\mathcal {L} _ {f i n e} ^ {k, T \\rightarrow I}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.735, + 0.47, + 0.797 + ], + "angle": 0, + "content": "By minimizing \\(\\mathcal{L}_{fine}^{k}\\) at all grouping stages during training, our framework could reach a meaningful fine-grained alignment automatically, i.e., for the \\(i\\)-th segment \\(\\mathbf{s}_i^k\\), the \\(l_i^k\\)-th entity satisfying" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.803, + 0.36, + 0.83 + ], + "angle": 0, + "content": "\\[\nl_{i}^{k} = \\operatorname *{arg max}_{1\\leq j\\leq E}\\cos [\\mathbf{x}_{i}^{k},\\mathbf{y}_{j}]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.839, + 0.47, + 0.902 + ], + "angle": 0, + "content": "tends to have a similar semantics with \\(\\mathbf{s}_i^k\\). We thus obtain \\(\\{l_i^k\\}_{i=1}^{H_k}\\) as the grounding results for the image segments \\(\\{\\mathbf{s}_i^k\\}_{i=1}^{H_k}\\). A further explanation of the automatic meaningful alignment can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.666, + 0.108 + ], + "angle": 0, + "content": "4.4. Segment Merger" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.189 + ], + "angle": 0, + "content": "To improve the ability of location, we propose to leverage the entity-grounding results as explicit supervision to learn a group of similarity matrices between image segments for small segments merging during inference, compared with [48] that learns the merging fully implicitly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.19, + 0.892, + 0.251 + ], + "angle": 0, + "content": "For each grouping stage \\( k \\), we compute the cosine similarity between each pair of image segments, which is then linearly re-scaled into \\( [0,1] \\) to formulate a similarity matrix \\( \\mathbf{Sim}_k \\in [0,1]^{H_k \\times H_k} \\), where" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.258, + 0.81, + 0.287 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S i m} _ {k} [ i, j ] = \\frac {1}{2} \\left(\\cos \\left[ \\mathbf {x} _ {i} ^ {k}, \\mathbf {x} _ {j} ^ {k} \\right] + 1\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.296, + 0.892, + 0.329 + ], + "angle": 0, + "content": "We further leverage \\(\\{l_i^k\\}_{i = 1}^{H_k}\\) as pseudo labels to build a pseudo target matrix \\(\\mathbf{Sim}_k^{target} \\in \\{0,1\\}^{H_k \\times H_k}\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.336, + 0.868, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S i m} _ {k} ^ {t a r g e t} [ i, j ] = \\left\\{ \\begin{array}{l} 1, \\text {i f} l _ {i} ^ {k} = l _ {j} ^ {k} \\wedge \\cos [ \\mathbf {x} _ {i} ^ {k}, \\mathbf {y} _ {l _ {i} ^ {k}} ] > \\theta \\\\ \\qquad \\wedge \\cos [ \\mathbf {x} _ {j} ^ {k}, \\mathbf {y} _ {l _ {j} ^ {k}} ] > \\theta , \\\\ 0, \\text {o t h e r w i s e}. \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.408, + 0.884, + 0.423 + ], + "angle": 0, + "content": "The similarity loss for the stage \\( k \\) is then formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.431, + 0.819, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s i m} ^ {k} = \\frac {1}{H _ {k} ^ {2}} \\| \\mathbf {S i m} _ {k} - \\mathbf {S i m} _ {k} ^ {\\text {t a r g e t}} \\| _ {F} ^ {2}.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.666, + 0.485 + ], + "angle": 0, + "content": "4.5. Label Generator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.704 + ], + "angle": 0, + "content": "In addressing the challenge of no pre-defined concept sets, the previous work [58] proposes to build a large vocabulary for learning during training and use WordNet [31] to correlate predictions within this vocabulary to the target concepts during inference. However, there are two limitations to the previous method. Firstly, compared with the extensive object semantics and relation predicates contained in textual descriptions, despite the large vocabulary established, it is inevitable that some classes will be overlooked. Secondly, leveraging WordNet to match vocabulary predictions to targets is not accurate and robust enough, for WordNet may only reach a coarse matching with multiple target concepts. This imprecision is particularly pronounced for relation predicates relative to object semantics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Given these limitations, we introduce a novel approach in TextPSG. Instead of approaching label prediction of objects and relations as a traditional classification problem, we reformulate it as an auto-regressive generation problem, which eliminates the necessity for pre-defined concept sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Compared with a vanilla RNN, we employ a pre-trained vision language model BLIP [21] to leverage the pre-learned common sense. BLIP can take an image as input and output a caption to describe the image. In TextPSG, we borrowed the pre-trained decoder module from BLIP to perform the generation of object and relation labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "During training, the label generator takes the caption-parsed text graph, the segment features from the region" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2843" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.09, + 0.473, + 0.228 + ], + "angle": 0, + "content": "grouper, and the grounding results \\(\\{l_i^k\\}_{i = 1}^{H_k}\\) from the entity grouper as input. It filters out the segments with tokenwise similarity lower than the threshold \\(\\theta\\), merges the segments mapped to the same entity, and queries the corresponding image masks from the region grouper. Then, \\(E_{k}\\) image masks \\(\\{\\mathbf{m}_i^k\\}_{i = 1}^{E_k}\\) with their pseudo entity labels \\(\\{b_i^k\\}_{i = 1}^{E_k}\\) can be obtained, where each \\(b_{i}^{k}\\) is one entity in the text graph. \\(E_{k}\\leq E\\) because some textual entities may not exist in the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.47, + 0.306 + ], + "angle": 0, + "content": "Prompt-embedding-based technique (PET). To better incorporate common sense from the vision language model, we further design a novel PET for label generation. For object prediction, the decoder takes the segment features and the image mask \\(\\mathbf{m}_i^k\\), using a prompt" + }, + { + "type": "title", + "bbox": [ + 0.203, + 0.315, + 0.343, + 0.332 + ], + "angle": 0, + "content": "a photo of [ENT]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.47, + 0.398 + ], + "angle": 0, + "content": "to guide the object generation, where the [ENT] token is expected to be the pseudo label \\( b_{i}^{k} \\). For relation prediction, the decoder takes the segment features and an image mask pair \\( (\\mathbf{m}_i^k,\\mathbf{m}_j^k) \\) as input, using a prompt" + }, + { + "type": "title", + "bbox": [ + 0.162, + 0.405, + 0.385, + 0.438 + ], + "angle": 0, + "content": "a photo of [SUB] and [OBJ] what is their relation [REL]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.473, + 0.61 + ], + "angle": 0, + "content": "to guide the relation generation, where the [SUB] and [OBJ] tokens are embedded by the pseudo labels \\( b_{i}^{k} \\) and \\( b_{j}^{k} \\), and the [REL] token is expected to be the relation predicate between \\( (b_{i}^{k}, b_{j}^{k}) \\) with \\( b_{i}^{k} \\) as subject and \\( b_{j}^{k} \\) as object in the text graph. To enhance relation generation, we further design three learnable positional embeddings \\( \\mathbf{f}_{sub} \\), \\( \\mathbf{f}_{obj} \\), \\( \\mathbf{f}_{region} \\) for indicating the different regions in the segment features. Two cross-entropy losses \\( \\mathcal{L}_{ent}^{k} \\), \\( \\mathcal{L}_{rel}^{k} \\) are used to supervise the generation of the [ENT] and [REL] tokens, maximizing the likelihood of the target label strings, respectively. More details are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.617, + 0.188, + 0.632 + ], + "angle": 0, + "content": "4.6. Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.473, + 0.903 + ], + "angle": 0, + "content": "During inference, the target concepts of object semantics \\(\\mathcal{C}_o\\) and relation predicates \\(\\mathcal{C}_r\\) are known. With an image \\(I\\) as input and an inference stage index \\(k_{inf}\\) specified, the region grouper first partitions \\(I\\) into several candidate segments \\(\\{\\mathbf{s}_i^{k_{inf}}\\}_{i=1}^{H_{k_{inf}}}\\), which are then passed through the segment merger to obtain the similarity matrix \\(\\mathbf{Sim}_{k_{inf}}\\). We formulate the segment merging as a spectral clustering problem and perform the graph cut [38] on \\(\\mathbf{Sim}_{l_{inf}}\\) for clustering. To improve the accuracy, we employ a matrix recovery method [25] to reduce the noise in \\(\\mathbf{Sim}_{l_{inf}}\\). In this step, the segments with similar semantics tend to be merged into the same cluster. For each cluster and each pair of clusters, the label generator uses a similar PET to generate the object semantics and the relation predicates. For every label within sets \\(\\mathcal{C}_o\\) and \\(\\mathcal{C}_r\\), the label generator computes its generation probability. Subsequently, these probabilities are used to rank the concepts, selecting the most probable as the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "final prediction. Note that between object and relation prediction, to convert semantic segmentation into instance segmentation, we identify each connected component in the semantic segmentation to be an instance, for simplicity. More details about inference are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.18, + 0.734, + 0.196 + ], + "angle": 0, + "content": "5. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.205, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Datasets. We train our model with the COCO Caption dataset [4], which involves 123,287 images with each labeled by 5 independent human-generated captions. Following the 2017 split, we use 118,287 images with their captions for training. We evaluate models with the Panoptic Scene Graph dataset [49] for its pixel-wise labeling as well as its high-quality object and relation annotation. We further merge the object semantics with ambiguities. After merging, 127 object semantics and 56 relation predicates are finally obtained for evaluation. More details about the datasets can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.893, + 0.567 + ], + "angle": 0, + "content": "Evaluation Protocol and Metrics. Following all previous works in scene graph generation, we evaluate the quality of a generated scene graph by viewing it as a set of subject-predicate-object triplets. We evaluate models on two tasks: Visual Phrase Detection (PhrDet) and Scene Graph Detection (SGDet). PhrDet aims to detect the whole phrase of subject-Predicate-object with a union location of subject and object. It is considered to be correct if the phrase labels are all correct and the union location matches the ground truth with the intersection over union (IoU) greater than 0.5. SGDet further requires a more accurate location, i.e., the location of subject and object should match the ground truth with IoU greater than 0.5 respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.687 + ], + "angle": 0, + "content": "We use No-Graph-Constraint-X Recall@K (NXR@K, %) to measure the ability of generation. Recall@K computes the recall between the top-k generated triplets with the ground truth. No-Graph-Constraint-X indicates that at most X predicate labels could be predicted for each subject-object pair. Since some predicates defined in [49] are not exclusive, such as on and sitting on, NXR@K could be a more reasonable metric compared with Recall@K." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Baselines. We consider several baselines for Caption-to-PSG in the following experiments. Firstly, we design four baselines that strictly follow the constraints of Caption-to-PSG, where objects are located by bbox proposals generated by selective search [43]:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.772, + 0.891, + 0.801 + ], + "angle": 0, + "content": "- Random is the most naive baseline where all object semantics and relation predicates are randomly predicted." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.806, + 0.892, + 0.837 + ], + "angle": 0, + "content": "- Prior augments Random by performing label prediction based on the statistical priors in the training set." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "- MIL performs the alignment between proposals and textual entities by multiple instance learning [30]. Similar to [58], it formulates the object label prediction as a classification problem in a large pre-built vocabulary, with" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.772, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2844" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.894, + 0.261 + ], + "angle": 0, + "content": "
MethodModePhrDetSGDet
ModelProposalTargetN3R50N3R100N5R50N5R100N3R50N3R100N5R50N5R100
SGGNLS-cDetector✓bbox9.6911.4510.2412.226.767.817.28.65
Randomxbbox0.020.030.020.030.010.020.020.03
PriorSelective Searchxbbox0.040.070.050.070.030.060.050.07
MILxbbox1.972.182.042.611.21.351.561.97
SGCLIPxbbox3.023.453.383.712.132.32.392.7
SGGNLS-oDetectorxbbox6.26.796.927.933.964.214.535.02
Ours-xmask8.289.169.0610.513.323.633.714.18
Ours-xbbox11.3712.7412.2414.374.294.774.825.48
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.262, + 0.893, + 0.29 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparison of Different Methods on Caption-to-PSG. 'Proposal' indicates how the method obtains bbox proposals. 'Target' indicates whether the concept sets for inference are known during training. 'Mode' indicates the mode used for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.313, + 0.468, + 0.344 + ], + "angle": 0, + "content": "WordNet [31] employed during inference. The relation labels are predicted with statistical priors, similar to Prior." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.349, + 0.469, + 0.379 + ], + "angle": 0, + "content": "- SGCLIP employs the pre-trained CLIP [34] to predict both object semantic labels and relation predicate labels." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.383, + 0.469, + 0.429 + ], + "angle": 0, + "content": "Secondly, to further benchmark the performance of our framework, we set two additional baselines based on [58] by gradually removing the constraints of Caption-to-PSG:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.433, + 0.469, + 0.509 + ], + "angle": 0, + "content": "- SGGNLS-o [58] extracts proposals with a detector [35] pre-trained on OpenImage [19]. It formulates the object and relation label prediction as a classification problem within a large pre-built vocabulary, with WordNet [31] employed during inference." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.513, + 0.468, + 0.573 + ], + "angle": 0, + "content": "- SGGNLS-c [58] uses the same proposals as SGGNLS-o. In SGGNLS-c, the target concept sets for inference are known during training. It formulates the label prediction as a classification problem in these target concept sets." + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.349, + 0.469, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.582, + 0.456, + 0.598 + ], + "angle": 0, + "content": "More design details are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.469, + 0.734 + ], + "angle": 0, + "content": "Implementation Details. Following GroupViT [48], we set \\( K = 2 \\), \\( H_{1} = 64 \\), and \\( H_{2} = 8 \\) for our region grouper. We leverage general pre-trained models for weight initialization. We employ the pre-trained GroupViT for the region grouper as well as \\( \\mathbf{Tfm}^T \\) in the entity grounder, and the pre-trained BLIP [21] decoder for the label generator. During training, \\( \\mathbf{Tfm}^T \\) and the label generator are frozen. During inference, we set \\( k_{inf} = 1 \\). More implementation details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.367, + 0.759 + ], + "angle": 0, + "content": "5.1. Main Results on Caption-to-PSG" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. Our quantitative results on Caption-to-PSG are shown in Tab. 1. To make a fair comparison with bbox-based scene graphs generated by baselines, we evaluate our generated PSGs in both mask and bbox mode. For the latter, all masks in both prediction and ground truth are converted into bboxes (i.e., the mask area's enclosing rectangle) for evaluation, resulting in an easier setting than the former. The results show that our framework (Ours) significantly outperforms all the baselines under the same" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.435 + ], + "angle": 0, + "content": "constraints on both PhrDet (14.37 vs. 3.71 N5R100) and SGDet (5.48 vs. 2.7 N5R100). Our method also shows better results compared with SGGNLS-o on all metrics and all tasks (on PhrDet, 14.37 vs. 7.93 N5R100; on SGDet, 5.48 vs. 5.02 N5R100) although SGGNLS-o utilizes location priors by leveraging a pre-trained detector. The results demonstrate that our framework is more effective for learning a good panoptic structured scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.893, + 0.646 + ], + "angle": 0, + "content": "Qualitative Results. We provide typical qualitative results in Fig. 3 to further show our framework's effectiveness. Compared with SGGNLS-o, our framework has the following advantages. First, our framework is able to provide fine-grained semantic labels to each pixel in the image to reach a panoptic understanding, while SGGNLS-o can only provide sparse bboxes produced by the pre-trained detector. Note that categories with irregular shapes (e.g., trees in Fig. 3) are hard to be labeled precisely by bboxes. Second, compared with SGGNLS-o, our framework can generate more comprehensive object semantics and relation predicates, such as \"dry grass field\" and \"land at\" in Fig. 3, showing the open-vocabulary potential of our framework. More qualitative results are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.657, + 0.738, + 0.673 + ], + "angle": 0, + "content": "5.2. OOD Robustness Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.749 + ], + "angle": 0, + "content": "We further analyze another key advantage of our framework, i.e., the robustness in OOD cases. Since SGGNLS-c and SGGNLS-o both rely on a pre-trained detector to locate objects, their performance highly depends on whether object semantics in the scene are covered by the detector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Based on the object semantics [19] covered by the detector, we split the ground truth triplets into an in-distribution (ID) set and an OOD set. For triplets within the ID set, both the subject and object semantics are covered, while for triplets in the OOD set, at least one of the semantics is not covered. As shown in Tab. 2, both SGGNLS-c and SGGNLS-o suffer a significant performance drop from the ID set to the OOD set. On the OOD set, the triplets can hardly be retrieved. However, our framework, with the ability of location learned from purely text descriptions, can" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2845" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.094, + 0.235, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.108, + 0.359, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.096, + 0.557, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.096, + 0.701, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.095, + 0.873, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.175, + 0.235, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.19, + 0.388, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.175, + 0.556, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.183, + 0.718, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.184, + 0.884, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.256, + 0.234, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.265, + 0.358, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.256, + 0.556, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.251, + 0.718, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.249, + 0.888, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.345, + 0.296, + 0.356 + ], + "angle": 0, + "content": "(a) Results of SGGNLS-o" + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.346, + 0.69, + 0.356 + ], + "angle": 0, + "content": "(b) Results of Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.357, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Figure 3. Qualitative Comparison between SGGNLS-o (a) and Ours (b). For each method, the results of object location are shown on the left, while the results of scene graph generation are shown on the right. For Ours, scene graphs predicted within the given concept sets are provided in the middle column, and scene graphs directly predicted through the auto-regressive generation (i.e., an open-vocabulary manner) in the label generator are additionally provided in the right column." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.424, + 0.468, + 0.55 + ], + "angle": 0, + "content": "
SetModelTargetModePhrDetSGDet
N3R100N5R100N3R100N5R100
IDSGGNLS-c✓bbox16.7618.4810.4511.86
SGGNLS-oXbbox11.5513.647.138.47
OursXmask9.2710.453.283.76
OursXbbox13.3514.824.635.36
OODSGGNLS-c✓bbox0000
SGGNLS-oXbbox0.050.0600
OursXmask8.479.764.074.51
OursXbbox10.1811.695.235.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.551, + 0.469, + 0.577 + ], + "angle": 0, + "content": "Table 2. Analysis on OOD Robustness. 'Set' indicates the triplet set used for evaluation." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.591, + 0.468, + 0.663 + ], + "angle": 0, + "content": "
Stage#SegCutPhrDetSGDet
N3R100N5R100N3R100N5R100
164×10.7311.393.183.51
164✓12.7414.374.775.48
28×9.2411.033.534.35
28✓6.788.452.463.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.663, + 0.468, + 0.718 + ], + "angle": 0, + "content": "Table 3. Ablation Study on the Segment Merger. 'Stage' indicates the grouping stage where image segments used for merging are from. \\(\\# \\mathrm{Seg}\\) indicates the number of image segments. Cut indicates whether the graph-cut-based segment merging is applied." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.727, + 0.468, + 0.757 + ], + "angle": 0, + "content": "reach similar performance on both sets, which demonstrates the OOD robustness of our framework for PSG generation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.771, + 0.241, + 0.785 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.9 + ], + "angle": 0, + "content": "We conduct additional ablation studies to evaluate the effectiveness of our design choices. For all following experiments, we report N3R100 and N5R100 evaluated in bbox mode for simplicity. We answer the following questions. Q1: Does the explicit learning of merging in the segment merger helps provide better image segments? Q2: Is the generation-based label prediction better than the" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.424, + 0.892, + 0.495 + ], + "angle": 0, + "content": "
Label PredictionModelPhrDetSGDet
N3R100N5R100N3R100N5R100
Cls + WordNet-8.829.362.362.72
GenRNN9.1210.442.653.07
Gen w/o PETBLIP [21]2.332.580.450.6
Gen w/ PETBLIP [21]12.6414.284.775.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.495, + 0.891, + 0.522 + ], + "angle": 0, + "content": "Table 4. Ablation Study on the Label Generator. 'Cls' indicates classification. 'Gen' indicates generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.594 + ], + "angle": 0, + "content": "classification-based prediction? Q3: Does the pre-learned common sense from the pre-trained BLIP [21] helps with the label prediction? Q4: Does the PET helps incorporate the pre-learned common sense for label prediction?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.747 + ], + "angle": 0, + "content": "In Tab. 3, we compare different strategies of image segment merging during inference. Row 1&2 denote that the \\( H_{1} = 64 \\) segments from the first grouping stage are used for further merging, while row 3&4 denote that the \\( H_{2} = 8 \\) segments from the second stage are used. The results show that applying the graph cut to merge the segments from the first stage could reach the best performance, corroborating that compared with the fully implicit learning of merging, the explicit learning of merging can provide better segments (row 2 vs 3, answering Q1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Tab. 4, we compare different designs of the label generator. Keeping the other modules the same, we change the label generator (row 4) into three different designs, i.e., classification within a large pre-built vocabulary followed by WordNet [31] for target matching (row 1), generation with a vanilla RNN (row 2), generation with the BLIP decoder but without the PET (row 3). The results show that with the constraint of no pre-defined concept sets, compared with formulating the label prediction into a classification problem, formulating it into a generation problem is a bet-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2846" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "ter choice (row 1 vs 2&4, answering Q2). By employing the pre-trained BLIP for leveraging the pre-learned common sense, the prediction could be further boosted (row 2 vs 4, answering Q3). And the PET is very important for incorporating the common sense from the pre-trained model (row 3 vs 4, answering Q4)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.183, + 0.47, + 0.214 + ], + "angle": 0, + "content": "More ablation studies for the design evaluation can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.223, + 0.274, + 0.24 + ], + "angle": 0, + "content": "5.4. Application on TSSS" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.248, + 0.244, + 0.352 + ], + "angle": 0, + "content": "As a side product, we observe that our entity grounder and segment merger can also enhance TSSS. Based on the original GroupViT [48], we replace the multi" + }, + { + "type": "table", + "bbox": [ + 0.283, + 0.259, + 0.462, + 0.32 + ], + "angle": 0, + "content": "
MethodmIoU
GroupViT [48]24.28
GroupViT† [48]24.72
Ours26.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.274, + 0.321, + 0.468, + 0.347 + ], + "angle": 0, + "content": "Table 5. Results on TSSS. † indicates finetuned." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.474 + ], + "angle": 0, + "content": "label contrastive loss with our entity grounder and segment merger. Then we finetune the model on the COCO Caption dataset [4]. As shown in Tab. 5, compared with GroupViT directly finetuned on [4], the explicit learning of merging in our modules can boost the model with an absolute \\(2.15\\%\\) improvement of mean Intersection over Union (mIoU, \\(\\%\\)) on COCO [23], which demonstrates the effectiveness of our proposed modules on better object location." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.484, + 0.195, + 0.499 + ], + "angle": 0, + "content": "5.5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Failure Case Analysis. Despite the impressive performance of TextPSG, there are still challenges to address. Upon analyzing the failure cases for PSG generation, we identify three specific limitations of TextPSG that contribute to these failures. a) The strategy we use to convert semantic segmentation into instance segmentation is not entirely effective. For simplicity, in TextPSG, we identify each connected component in the semantic segmentation to be an individual object instance. However, this strategy may fail when instances overlap or are occluded, resulting in either an underestimation or an overestimation of instances. b) Our framework faces difficulty in locating small objects in the scene due to limitations in resolution and the grouping strategy for location. c) The relation prediction of our framework requires enhancement, as it is not adequately conditioned on the image. While the label generator uses both image features and predicted object semantics to determine the relation, it sometimes seems to lean heavily on the object semantics, potentially neglecting the actual image content. Examples of failure cases for each of these limitations can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Model Diagnosis. For a clearer understanding of the efficacy of our framework, we conduct a model diagnosis to answer the following question: why does our framework only achieve semantic segmentation through learning, rather than panoptic segmentation (and thus requires further" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.89, + 0.107 + ], + "angle": 0, + "content": "segmentation conversion to obtain instance segmentation)?" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.121, + 0.63, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.216, + 0.7, + 0.234 + ], + "angle": 0, + "content": "(a) A white sheep and a black sheep standing on the grass under the sky." + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.121, + 0.747, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.121, + 0.871, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.216, + 0.875, + 0.234 + ], + "angle": 0, + "content": "(b) Pair of sheep standing on open grassy field on clear day." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.237, + 0.892, + 0.307 + ], + "angle": 0, + "content": "Figure 4. Region-Entity Alignment Results of Captions in Different Granularity. Two captions in different granularity are used to execute region-entity alignment with the same image, with (a) one describing the two sheep individually while (b) the other merges them in plural form." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.442 + ], + "angle": 0, + "content": "In Fig. 4, we use two captions in different granularity to execute region-entity alignment. It shows that our framework has the capability to assign distinct masks to individual instances. However, the nature of caption data, where captions often merge objects of the same semantics in plural form, limits our framework from differentiating instances. It is the weak supervision provided by the caption data that constrains our framework." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.444, + 0.854, + 0.459 + ], + "angle": 0, + "content": "More diagnoses are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Future Directions. In response to the limitations discussed, we outline several potential directions for enhancing our framework: a) a refined and sophisticated strategy for segmentation conversion; b) increasing the input resolution, though this may introduce greater computational demands; c) a more suitable image-conditioned reasoning mechanism for relation prediction; d) a superior image-caption-pair dataset with more detailed granularity in captions to achieve panoptic segmentation through learning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.619, + 0.619 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.82 + ], + "angle": 0, + "content": "We take the first step towards the novel problem Caption-to-PSG, aiming to learn PSG generation purely from language. To tackle this challenging problem, we propose a new modularized framework TextPSG with several novel techniques, which significantly outperforms the baselines and achieves strong OOD robustness. This paves the path to a more comprehensive and generalizable panoptic structured scene understanding. There are still bottle-necks in TextPSG to be explored in future work, including a) a more sophisticated strategy for segmentation conversion; b) a more suitable image-conditioned reasoning mechanism for relation prediction; c) a superior image-caption-pair dataset for panoptic segmentation through learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the DSO grant DSOCO21072, IBM, and gift funding from MERL, Cisco, and Amazon. We would also like to thank the computation support from AiMOS, a server cluster for the IBM Research AI Hardware Center." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2847" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.469, + 0.171 + ], + "angle": 0, + "content": "[1] Somak Aditya, Yezhou Yang, Chitta Baral, Yiannis Aloimonos, and Cornelia Fermüller. Image understanding using vision and reasoning through scene description graph. Computer Vision and Image Understanding, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.469, + 0.228 + ], + "angle": 0, + "content": "[2] S. Amiri, Kishan Chandan, and Shiqi Zhang. Reasoning with scene graphs for robot planning under partial observability. IEEE Robotics and Automation Letters, 7:5560-5567, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[3] Gabor Angeli, Melvin Johnson, and Christopher D. Manning. Leveraging linguistic structure for open domain information extraction. In Annual Meeting of the Association for Computational Linguistics, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.289, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[4] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[5] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee K Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. In Proc. 57th Annual Meeting of the Association for Computational Linguistics, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.404, + 0.469, + 0.471 + ], + "angle": 0, + "content": "[6] Zhenfang Chen, Jiayuan Mao, Jiajun Wu, Kwan-Yee K Wong, Joshua B. Tenenbaum, and Chuang Gan. Grounding physical concepts of objects and events through dynamic visual reasoning. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.475, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[7] Zhenfang Chen, Peng Wang, Lin Ma, Kwan-Yee K Wong, and Qi Wu. Cops-ref: A new dataset and task on compositional referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10086-10095, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.547, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[8] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D. Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.604, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[9] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[10] Samir Yitzhak Gadre, Kiana Ehsani, Shuran Song, and Roozbeh Mottaghi. Continuous scene representations for embodied ai. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14829-14839, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[11] Chuang Gan, Yandong Li, Haoxiang Li, Chen Sun, and Boqing Gong. Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1811-1820, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Jiumiang Gu, Handong Zhao, Zhe Lin, Sheng Li, Jianfei Cai, and Mingyang Ling. Scene graph generation with external knowledge and image reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[14] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[15] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3128-3137, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[16] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.317, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[17] Siddhesh Khandelwal, Mohammed Suhail, and Leonid Sigal. Segmentation-grounded scene graph generation. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15859-15869, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Fei-Fei Li. Visual Genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32-73, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[19] Alina Kuznetsova, Hassan Rom, Neil Gordon Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4. International Journal of Computer Vision, 128:1956-1981, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[20] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[22] Xingchen Li, Long Chen, Wenbo Ma, Yi Yang, and Jun Xiao. Integrating object-aware and interaction-aware knowledge for weakly supervised scene graph generation. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[24] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Guangcan Liu, Zhouchen Lin, Shuicheng Yan, Ju Sun, Yong Yu, and Yi Ma. Robust recovery of subspace structures by low-rank representation. IEEE transactions on pattern analysis and machine intelligence, 35(1):171-184, 2012. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2848" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7086-7096, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.204 + ], + "angle": 0, + "content": "[27] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. arXiv preprint arXiv:2211.14813, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.205, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[28] Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David McClosky. The Stanford CoreNLP natural language processing toolkit. In Association for Computational Linguistics (ACL) System Demonstrations, pages 55–60, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[29] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.344, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[30] Oded Maron and Tomás Lozano-Pérez. A framework for multiple-instance learning. Advances in neural information processing systems, 10, 1997. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.386, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[31] George A. Miller. WordNet: A lexical database for English. In Speech and Natural Language: Proceedings of a Workshop Held at Harriman, New York, February 23-26, 1992, 1992. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[32] Julia Peyre, Ivan Laptev, Cordelia Schmid, and Josef Sivic. Weakly-supervised learning of visual relations. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5189-5198, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.497, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[33] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.581, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.665, + 0.469, + 0.719 + ], + "angle": 0, + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.721, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[36] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 817-834. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.804, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[37] Sebastian Schuster, Ranjay Krishna, Angel X. Chang, Li Fei-Fei, and Christopher D. Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In VL@EMNLP, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[38] Jianbo Shi and J. Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[39] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8376-8384, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[40] Jing Shi, Yiwu Zhong, Ning Xu, Yin Li, and Chenliang Xu. A simple baseline for weakly-supervised scene graph generation. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16373-16382, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[41] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[42] Damien Teney, Lingqiao Liu, and Anton van den Hengel. Graph-structured representations for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[43] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104:154-171, 2013. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[44] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[45] Qixun Wang, Xiaofeng Guo, and Haofan Wang. 1st place solution for psg competition with eccv'22 sensehuman workshop. ArXiv, abs/2302.02651, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[46] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6602–6611, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[47] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5419, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[48] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. GroupViT: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18134-18144, June 2022. 2, 3, 4, 5, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[49] Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. In ECCV, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[50] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[51] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. FILIP: Fine-grained interactive language-image pre-training. In International Conference on Learning Representations, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[52] Keren Ye and Adriana Kovashka. Linguistic structures as weak supervision for visual scene graph generation. In Pro-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2849" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[53] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1307-1315, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[54] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[55] Alireza Zareian, Svebor Karaman, and Shih-Fu Chang. Weakly supervised visual semantic parsing. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3733-3742, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[56] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.469, + 0.471 + ], + "angle": 0, + "content": "[57] Hanwang Zhang, Zawlin Kyaw, Jinyang Yu, and Shih-Fu Chang. Ppr-fcn: Weakly supervised visual relation detection via parallel pairwise r-fcn. In Proceedings of the IEEE international conference on computer vision, pages 4233-4241, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.474, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[58] Yiwu Zhong, Jing Shi, Jianwei Yang, Chenliang Xu, and Yin Li. Learning to generate scene graph from natural language supervision. In ICCV, 2021. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.469, + 0.584 + ], + "angle": 0, + "content": "[59] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVIII, pages 696-712. Springer, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2850" + } + ] +] \ No newline at end of file diff --git a/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/a412d306-b042-443c-958a-505db74fd5e7_origin.pdf b/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/a412d306-b042-443c-958a-505db74fd5e7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6a40203d376f3bb7d1dd8286888bb8eed1558796 --- /dev/null +++ b/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/a412d306-b042-443c-958a-505db74fd5e7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d20b855b8a99db5a3b7319416a15f7b1477df4cc0f6d2b468c05914aa250e301 +size 3544375 diff --git a/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/full.md b/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9c6b35d159d3b64553101ebb8dc6e727d9512dd8 --- /dev/null +++ b/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/full.md @@ -0,0 +1,403 @@ +# TextPSG: Panoptic Scene Graph Generation from Textual Descriptions + +Chengyang Zhao1 Yikang Shen2 Zhenfang Chen2 Mingyu Ding3 Chuang Gan2,4 +1Peking University 2MIT-IBM Watson AI Lab +3UC Berkley 4UMass Amherst + +# Abstract + +Panoptic Scene Graph has recently been proposed for comprehensive scene understanding. However, previous works adopt a fully-supervised learning manner, requiring large amounts of pixel-wise densely-annotated data, which is always tedious and expensive to obtain. To address this limitation, we study a new problem of Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG). The key idea is to leverage the large collection of free image-caption data on the Web alone to generate panoptic scene graphs. The problem is very challenging for three constraints: 1) no location priors; 2) no explicit links between visual regions and textual entities; and 3) no predefined concept sets. To tackle this problem, we propose a new framework TextPSG consisting of four modules, i.e., a region grouper, an entity grounder, a segment merger, and a label generator, with several novel techniques. The region grouper first groups image pixels into different segments and the entity grounder then aligns visual segments with language entities based on the textual description of the segment being referred to. The grounding results can thus serve as pseudo labels enabling the segment merger to learn the segment similarity as well as guiding the label generator to learn object semantics and relation predicates, resulting in a fine-grained structured scene understanding. Our framework is effective, significantly outperforming the baselines and achieving strong out-of-distribution robustness. We perform comprehensive ablation studies to corroborate the effectiveness of our design choices and provide an in-depth analysis to highlight future directions. Our code, data, and results are available on our project page: https://vis-ww.c.s.umass.edu/TextPSG. + +# 1. Introduction + +A scene graph is a directed-graph-based abstract representation of the objects and their relations within a scene. It has been widely utilized to develop a structured scene understanding of object semantics, locations, and relations, + +![](images/065d0ed580478e3af013e3fcf045d049fb6c45002c8ef8797ce8355a2355b286.jpg) +Figure 1. Problem Overview. Different from the traditional bbox-based form of the scene graph as shown in (a), Caption-to-PSG aims to generate the mask-based panoptic scene graph. In Caption-to-PSG, the model has no access to any location priors, explicit region-entity links, or pre-defined concept sets. Consequently, the model is required to learn partitioning and grounding as illustrated in (b), as well as object semantics and relation predicates as illustrated in (c), all purely from textual descriptions. + +![](images/0254e5361ce1c24aa5539ff5c6c3b04b2ce638ce0bbed7ee183c246ba9132828.jpg) + +which facilitates a variety of downstream applications, such as image generation [14, 8], visual reasoning [42, 1, 39], and robotics [2, 10]. + +Typically, in a scene graph, each node denotes an object in the scene located by a bounding box (bbox) with a semantic label, and each directed edge denotes the relation between a pair of objects with a predicate label. Nonetheless, a recent work [49] points out that such a bbox-based form of scene graph is not ideal enough. Firstly, compared with pixel-wise segmentation masks, bboxes are less fine-grained and may contain some noisy pixels belonging to other objects, limiting the applications for some downstream tasks. For example, as shown in Fig. 1 (a), about half of the pixels in the yellow_bbox for girl belong to wall. Secondly, it is challenging for bboxes to cover the entire scene without ambiguities caused by overlaps, which prevents a scene graph from including every object in the scene for a complete description. To this end, the work [49] proposes the concept of Panoptic Scene Graph (PSG), in which each object is grounded by a panoptic segmentation mask, to reach a comprehensive structured scene representation. + +However, all existing works [49, 45] approach PSG generation through a fully-supervised manner, i.e., learning to perform panoptic segmentation and relation prediction from manually-annotated datasets with explicit supervision for both segmentation and relation prediction. Unfortunately, it is extremely labor-intensive to build such datasets, making it difficult to scale up to cover more complex scenes, object semantics, and relation predicates, thus significantly limiting the generalizability and the application of these methods to the real world. For instance, the current PSG dataset [49] only covers 133 object semantics and 56 relation predicates. + +To relieve the reliance on densely-annotated data, weakly-supervised methods [52, 58, 22] for scene graph generation are promising. These methods could induce scene graphs from image-caption pairs, which can be easily harvested from the Web for free. Even so, they still rely heavily on two strong preconditions, i.e., a powerful region proposal network (e.g., [35]) and a pre-defined set of object semantics and relation predicates. Although these preconditions facilitate the learning process of the methods, they also limit the generalizability for locating novel objects (unforeseen objects for the region proposal network) and constrain the understanding into the limited concept set. + +Inspired by previous weakly-supervised methods, we introduce a new problem, Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG), to explore a holistic structured scene understanding without labor-intensive data annotation. Considering the limitation of the preconditions mentioned, we set three constraints to Caption-to-PSG to reach a more comprehensive and generalizable understanding, which results in a very challenging problem: a) only image-caption pairs are provided during training, without any location priors in either region proposals or location supervision; b) the explicit links between regions in images and entities in captions are missing; c) no concept sets are pre-defined, i.e., neither object semantics nor relation predicates are known beforehand. + +Given these three constraints, we argue that there are two key challenges for the model to solve the problem. Firstly, the model should learn to ground entities in language onto the visual scene without explicit location supervision, i.e., the ability to perform partitioning and grounding, as shown in Fig. 1 (b), should be developed purely from textual descriptions. Secondly, during training, the model should also learn the object semantics and relation predicates from textual descriptions, as shown in Fig. 1 (c), without pre-defined fixed object and relation vocabularies. By solving these challenges, the model could associate visual scene patterns with textual descriptions, gradually acquire common sense among them, and finally reach a more comprehensive and generalizable understanding, including novel object location, extensive semantics recognition, and complex relation analysis, which is more suitable to the real world. + +With these considerations, we propose a novel framework, TextPSG, as the first step towards this challenging problem. TextPSG consists of a series of modules to cooperate with each other, i.e., a region grouper, an entity grounder, a segment merger, and a label generator. The region grouper learns to merge image regions into several segments in a hierarchical way based on object semantics, similar to [48]. The entity grouder employs a fine-grained contrastive learning strategy [51] to bridge the textual description and the visual content, grounding entities in the caption onto the image segments. With the entity-grounding results as pseudo labels, the segment merger learns similarity matrices to merge small image segments during inference, while the label generator learns the prediction of object semantics and relation predicates. Specifically, in the segment merger, we propose to leverage the grounding as explicit supervision for merging, compared with [48] which learns merging in a fully implicit manner, to improve the ability of location. In the label generator, different from all previous pipelines for scene graph generation, we reformulate the label prediction as an auto-regressive generation problem rather than a classification problem, and employ a pre-trained language model [21] as the decoder to leverage the pre-learned common sense. We further design a novel prompt-embedding-based technique (PET) to better incorporate common sense from the language model. Our experiments show that TextPSG significantly outperforms the baselines and achieves strong out-of-distribution (OOD) robustness. Comprehensive ablation studies corroborate the effectiveness of our design choices. As a side product, the proposed grouder and merger modules also have been observed to enhance text-supervised semantic segmentation. + +In spite of the promising performance of TextPSG, certain challenges persist. We delve into an in-depth analysis of the failure cases, provide a model diagnosis, and discuss potential future directions for enhancing our framework. + +To sum up, our contributions are as follows: + +- We introduce a new problem, Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG), to alleviate the burden of human annotation for PSG by learning purely from the weak supervision of captions. +- We propose a new modularized framework, TextPSG, with several novel techniques, which significantly outperforms the baselines and achieves strong OOD robustness. We demonstrate that the proposed modules in TextPSG can also facilitate text-supervised semantic segmentation. +- We perform an in-depth failure case analysis with a model diagnosis, and further highlight future directions. + +# 2. Related Work + +Bbox-based Scene Graph Generation. Bbox-based scene graph generation aims to create a structured representation of object semantics, locations, and relations in the + +scene, where each object is identified by a bbox. Most of existing works [50, 47, 41, 12, 24] follow a fully-supervised approach to learn the generation from densely-annotated datasets [18, 13], which requires significant human labors. To reduce the labeling effort, some weakly-supervised methods have been proposed [32, 55, 57, 40]. Recent works [52, 58, 22] further explore learning scene graph generation from image-caption pairs. However, they all rely on off-the-shelf region proposal networks for the location of objects in the scene, which are typically pretrained on pre-defined fixed sets of object semantics, limiting their generalizability to locating unforeseen objects. To reach a more granular and accurate grounding, [17] proposes to ground each object by segmentation. A recent work [49] further introduces the concept of PSG, where each object is identified by a panoptic segmentation mask, as a more comprehensive scene representation. + +Text-supervised Semantic Segmentation (TSSS). TSSS [48, 20, 26, 9, 27, 59] aims to learn image pixel semantic labeling from image-caption pairs without fine-grained pixel-wise annotations. Similar to TSSS, our proposed Caption-to-PSG aims to learn to connect visual regions and textual entities from only image-caption pairs and has the potential to leverage the large collection of free data on the Web. However, different from TSSS, Caption-to-PSG further requires the model to learn the relations among different visual regions, resulting in a higher-order structured understanding of visual scenes. In addition to unknown object semantics, Caption-to-PSG does not assume any pre-defined relation predicate concepts. + +Visual Grounding. Our work is also related to visual grounding [16, 56, 29, 11, 6, 33], which grounds entities in language onto objects in images. Early works [54, 53, 7] on visual grounding typically detect object proposals [35, 43] from images first and then match them with language descriptions by putting features of both modalities into the same feature space, which are in a fully-supervised learning manner. There are also some weakly-supervised grounding methods [15, 36, 5] which relieve the need for dense regional annotations by multiple instance learning [15] or learning to reconstruct [36]. Different from them, Caption-to-PSG is more challenging since it requires grounding fine-grained object relations between entities without region proposal networks for a pre-defined object vocabulary. + +# 3. Problem Formulation + +Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG). A PSG $\mathcal{G} = (\mathcal{V},\mathcal{E})$ is a directed graph representation of the objects and the relations among them in a scene image $I\in \mathbb{R}^{H\times W\times 3}$ . Each node $v_{i}\in \mathcal{V}$ denotes an object in $I$ located by a panoptic segmentation mask $m_i\in \{0,1\}^{H\times W}$ with an object semantic label $o_i\in \mathcal{C}_o$ , and each directed edge $e_{ij}\in \mathcal{E}$ denotes a + +pair of subject $o_i$ and object $o_j$ with a relation predicate label $r_{ij} \in \mathcal{C}_r$ , where $\mathcal{C}_o$ and $\mathcal{C}_r$ are the defined concept sets of object semantics and relation predicates. Note that for a PSG, it is constrained that all segmentation masks could not overlap, i.e., $\sum_{i=1}^{|\mathcal{V}|} m_i \leq 1^{H \times W}$ . + +Given a large collection of paired scene images and textual descriptions $S = \{(I_i, T_i)\}_i$ , Caption-to-PSG aims to learn PSG generation from purely text descriptions for a holistic structured scene understanding, i.e., during training, only $S$ is provided as supervision, while during inference, with a scene observation $I'$ as input, the model is required to generate a corresponding PSG $\mathcal{G}'$ . + +Three Constraints. Note that in Caption-to-PSG, three important constraints are set to reach a more comprehensive and generalizable scene understanding: a) no location priors: different from all previous scene graph generation methods, neither pre-trained region proposal networks nor location supervision are allowed; b) no explicit region-entity links: the links between regions in the image $I$ and entities in the textual description $T$ are not provided; c) no pre-defined concept sets: the target concept sets defined for inference, i.e., object semantics $\mathcal{C}_o$ and relation predicates $\mathcal{C}_r$ , are unknown during training. + +# 4. Method + +Overview. As illustrated in Fig. 2, our proposed framework for Caption-to-PSG, TextPSG, contains four modules to cooperate with each other: a region grouper, an entity grounder, a segment merger, and a label generator. + +During training, TextPSG takes batched image-caption pairs as input. For each pair, the image is passed through the region grouper to be partitioned into several image segments, while the caption is first pre-processed to extract its linguistic structure as a text graph and then taken by the entity grounder to ground textual entities in the graph onto the image segments. With the grounding results as pseudo labels, the segment merger learns similarity matrices between small image segments for further merging during inference, while the label generator learns the prediction of object semantics and relation predicates. + +During inference, for each input image, the image segments output from the region grouper are directly passed to the segment merger to be further merged according to the learned similarity matrices, and then fed to the label generator to predict the object semantic labels and the relation predicate labels. + +# 4.1. Text Graph Preprocessing + +Following previous methods [52, 58, 22] that leverage a rule-based language parser [46] based on [37] to preprocess textual descriptions, in TextPSG, we employ the same parser to extract linguistic structures from captions. Addi + +![](images/1088337b43e50200b645149a7c81c28956c4478e71e9ee822b3f5e51652acf2b.jpg) +Figure 2. Framework Overview of TextPSG. The framework consists of four modules cooperating with each other: a region grouper to merge regions in the input image into several segments, an entity grounder to ground entities in the caption onto the image segments, a segment merger to learn similarity matrices to merge small image segments during inference, and a label generator to learn the prediction of object semantics and relation predicates. The solid arrows indicate the training flow, while the dash arrows indicate the inference flow. The arrows from the region grouper to the label generator indicating the segment feature and mask query are omitted. + +tionally, inspired by the success of open information extraction (OpenIE) [3] in natural language processing, we also employ an OpenIE system from Stanford CoreNLP [28] for extraction as a supplement. After merging, for each caption, we obtain its linguistic structure represented in a text graph, where each node denotes an entity, and each directed edge denotes the relation between an entity pair. + +# 4.2. Region Grouper + +With a scene image as input, the region grouper aims to merge the regions with similar object semantics into several segments and extract corresponding high-level features. + +Our region grouper follows the hierarchical design of GroupViT [48]. Given an input image, the grouper first splits the image into $N$ non-overlapping patches as the initial image segments $\{\mathbf{s}_i^0\}_{i = 1}^N$ . These segments are then passed through $K$ grouping layers, where they are merged into larger, arbitrary-shaped segments progressively. Specifically, within each grouping layer $\mathbf{Grp}_k(k = 1,2,\dots ,K)$ , $H_{k}$ grouping centers $\{\mathbf{c}_i^k\}_{i = 1}^{H_k}$ could be learned. The grouping operation is performed through an attention mechanism between the centers and the segments, merging $H_{k - 1}$ input segments into $H_{k}$ larger ones, i.e., + +$$ +\left\{\mathbf {s} _ {i} ^ {k} \right\} _ {i = 1} ^ {H _ {k}} = \mathbf {G r p} _ {k} \big (\left\{\mathbf {c} _ {i} ^ {k} \right\} _ {i = 1} ^ {H _ {k}}, \left\{\mathbf {s} _ {i} ^ {k - 1} \right\} _ {i = 1} ^ {H _ {k - 1}} \big). +$$ + +Note that $H_0 = N$ . After the hierarchical grouping, multiple groups of segments $\{\mathbf{s}_i^k\}_{i=1}^{H_k}$ at different grouping stages are obtained. More details about the design of $\{\mathbf{Grp}_k\}_{k=1}^K$ can be found in the supplementary material. + +# 4.3. Entity Grounder + +Since the explicit region-entity links are not provided, bridging the textual description and the visual content automatically plays an important role in solving Caption-to-PSG. Inspired by FILIP [51], in TextPSG, we employ a similar fine-grained contrastive learning strategy to perform region-entity alignment. + +For each grouping stage $k$ , on image side, the grounder projects the segment group $\{\mathbf{s}_i^k\}_{i=1}^{H_k}$ into a new feature space $\mathcal{F}$ by a multi-layer perceptron (MLP) $\mathbf{Proj}_k^I$ to obtain segment embeddings $\{\mathbf{x}_i^k\}_{i=1}^{H_k}$ . On text side, the input caption is first tokenized into $M$ tokens $\{\mathbf{t}_i\}_{i=1}^M$ , which are then processed by a Transformer [44] $\mathbf{Tfm}^T$ to propagate information between each other. A recurrent neural network (RNN) $\mathbf{Rnn}$ further merges the tokens corresponding to the same entity, encoding tokens into their associated weights one by one and utilizing weighted sum to merge the token features into a singular entity feature. Finally, these entity features are projected to the same feature space $\mathcal{F}$ by a MLP $\mathbf{Proj}^T$ to obtain entity embeddings $\{\mathbf{y}_i\}_{i=1}^E$ , where $E$ denotes the number of entities in the caption. + +With the segment embeddings and the entity embeddings in the shared feature space $\mathcal{F}$ , we compute their token-wise similarities. Specifically, for the $i$ -th segment, we compute its cosine similarities with all entities to obtain the token-wise similarity from the $i$ -th segment to the caption $p_i^k$ via + +$$ +p _ {i} ^ {k} = \max _ {1 \leq j \leq E} \cos [ \mathbf {x} _ {i} ^ {k}, \mathbf {y} _ {j} ], +$$ + +where $\cos[\cdot, \cdot]$ denotes the cosine similarity operation. Note that different from the original FILIP [51], in the scenario of region-entity alignment, some regions in the scene may not be described in the caption, while some entities in the caption may not exist in the scene. To tackle this problem, we propose to set a filtering threshold $\theta$ , where pairs with similarity lower than $\theta$ will be considered in different semantics and filtered out. The fine-grained similarity from the image to the caption $p^k$ can thus be computed via + +$$ +p ^ {k} = \frac {1}{\sum_ {i = 1} ^ {H _ {k}} {\mathbf {1} _ {p _ {i} ^ {k} > \theta}}} \sum_ {i = 1} ^ {H _ {k}} (p _ {i} ^ {k} \cdot \mathbf {1} _ {p _ {i} ^ {k} > \theta}). +$$ + +Similarly, we can also compute the token-wise similarity from the $j$ -th entity to the image $q_j^k$ via + +$$ +q _ {j} ^ {k} = \max _ {1 \leq i \leq H _ {k}} \cos [ \mathbf {x} _ {i} ^ {k}, \mathbf {y} _ {j} ], +$$ + +and the fine-grained similarity from the caption to the image $q^{k}$ via + +$$ +q ^ {k} = \frac {1}{\sum_ {j = 1} ^ {E} \mathbf {1} _ {q _ {j} ^ {k} > \theta}} \sum_ {j = 1} ^ {E} (q _ {j} ^ {k} \cdot \mathbf {1} _ {q _ {j} ^ {k} > \theta}). +$$ + +Denoting the training batch with batch size $B$ as $\{(I_i,T_i)\}_{i = 1}^B$ , the fine-grained similarity from the image $I_{i}$ to the caption $T_{j}$ as $p^{k,i\rightarrow j}$ and from the caption $T_{j}$ to the image $I_{i}$ as $q^{k,j\rightarrow i}$ , the image-to-text fine-grained contrastive loss $\mathcal{L}_{fine}^{k,I\to T}$ and the text-to-image fine-grained contrastive loss $\mathcal{L}_{fine}^{k,T\to I}$ can then be formulated as + +$$ +\mathcal {L} _ {f i n e} ^ {k, I \rightarrow T} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \frac {\exp \left(p ^ {k , i \rightarrow i} / \tau\right)}{\sum_ {j = 1} ^ {B} \exp \left(p ^ {k , i \rightarrow j} / \tau\right)}, +$$ + +$$ +\mathcal {L} _ {f i n e} ^ {k, T \rightarrow I} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \frac {\exp \left(q ^ {k , i \rightarrow i} / \tau\right)}{\sum_ {j = 1} ^ {B} \exp \left(q ^ {k , i \rightarrow j} / \tau\right)}, +$$ + +where $\tau$ is a learnable temperature. The total fine-grained contrastive loss is + +$$ +\mathcal {L} _ {f i n e} ^ {k} = \frac {1}{2} (\mathcal {L} _ {f i n e} ^ {k, I \rightarrow T} + \mathcal {L} _ {f i n e} ^ {k, T \rightarrow I}). +$$ + +By minimizing $\mathcal{L}_{fine}^{k}$ at all grouping stages during training, our framework could reach a meaningful fine-grained alignment automatically, i.e., for the $i$ -th segment $\mathbf{s}_i^k$ , the $l_i^k$ -th entity satisfying + +$$ +l_{i}^{k} = \operatorname *{arg max}_{1\leq j\leq E}\cos [\mathbf{x}_{i}^{k},\mathbf{y}_{j}] +$$ + +tends to have a similar semantics with $\mathbf{s}_i^k$ . We thus obtain $\{l_i^k\}_{i=1}^{H_k}$ as the grounding results for the image segments $\{\mathbf{s}_i^k\}_{i=1}^{H_k}$ . A further explanation of the automatic meaningful alignment can be found in the supplementary material. + +# 4.4. Segment Merger + +To improve the ability of location, we propose to leverage the entity-grounding results as explicit supervision to learn a group of similarity matrices between image segments for small segments merging during inference, compared with [48] that learns the merging fully implicitly. + +For each grouping stage $k$ , we compute the cosine similarity between each pair of image segments, which is then linearly re-scaled into $[0,1]$ to formulate a similarity matrix $\mathbf{Sim}_k \in [0,1]^{H_k \times H_k}$ , where + +$$ +\mathbf {S i m} _ {k} [ i, j ] = \frac {1}{2} \left(\cos \left[ \mathbf {x} _ {i} ^ {k}, \mathbf {x} _ {j} ^ {k} \right] + 1\right). +$$ + +We further leverage $\{l_i^k\}_{i = 1}^{H_k}$ as pseudo labels to build a pseudo target matrix $\mathbf{Sim}_k^{target} \in \{0,1\}^{H_k \times H_k}$ , where + +$$ +\mathbf {S i m} _ {k} ^ {t a r g e t} [ i, j ] = \left\{ \begin{array}{l} 1, \text {i f} l _ {i} ^ {k} = l _ {j} ^ {k} \wedge \cos [ \mathbf {x} _ {i} ^ {k}, \mathbf {y} _ {l _ {i} ^ {k}} ] > \theta \\ \qquad \wedge \cos [ \mathbf {x} _ {j} ^ {k}, \mathbf {y} _ {l _ {j} ^ {k}} ] > \theta , \\ 0, \text {o t h e r w i s e}. \end{array} \right. +$$ + +The similarity loss for the stage $k$ is then formulated as + +$$ +\mathcal {L} _ {s i m} ^ {k} = \frac {1}{H _ {k} ^ {2}} \| \mathbf {S i m} _ {k} - \mathbf {S i m} _ {k} ^ {\text {t a r g e t}} \| _ {F} ^ {2}. +$$ + +# 4.5. Label Generator + +In addressing the challenge of no pre-defined concept sets, the previous work [58] proposes to build a large vocabulary for learning during training and use WordNet [31] to correlate predictions within this vocabulary to the target concepts during inference. However, there are two limitations to the previous method. Firstly, compared with the extensive object semantics and relation predicates contained in textual descriptions, despite the large vocabulary established, it is inevitable that some classes will be overlooked. Secondly, leveraging WordNet to match vocabulary predictions to targets is not accurate and robust enough, for WordNet may only reach a coarse matching with multiple target concepts. This imprecision is particularly pronounced for relation predicates relative to object semantics. + +Given these limitations, we introduce a novel approach in TextPSG. Instead of approaching label prediction of objects and relations as a traditional classification problem, we reformulate it as an auto-regressive generation problem, which eliminates the necessity for pre-defined concept sets. + +Compared with a vanilla RNN, we employ a pre-trained vision language model BLIP [21] to leverage the pre-learned common sense. BLIP can take an image as input and output a caption to describe the image. In TextPSG, we borrowed the pre-trained decoder module from BLIP to perform the generation of object and relation labels. + +During training, the label generator takes the caption-parsed text graph, the segment features from the region + +grouper, and the grounding results $\{l_i^k\}_{i = 1}^{H_k}$ from the entity grouper as input. It filters out the segments with tokenwise similarity lower than the threshold $\theta$ , merges the segments mapped to the same entity, and queries the corresponding image masks from the region grouper. Then, $E_{k}$ image masks $\{\mathbf{m}_i^k\}_{i = 1}^{E_k}$ with their pseudo entity labels $\{b_i^k\}_{i = 1}^{E_k}$ can be obtained, where each $b_{i}^{k}$ is one entity in the text graph. $E_{k}\leq E$ because some textual entities may not exist in the image. + +Prompt-embedding-based technique (PET). To better incorporate common sense from the vision language model, we further design a novel PET for label generation. For object prediction, the decoder takes the segment features and the image mask $\mathbf{m}_i^k$ , using a prompt + +# a photo of [ENT] + +to guide the object generation, where the [ENT] token is expected to be the pseudo label $b_{i}^{k}$ . For relation prediction, the decoder takes the segment features and an image mask pair $(\mathbf{m}_i^k,\mathbf{m}_j^k)$ as input, using a prompt + +# a photo of [SUB] and [OBJ] what is their relation [REL] + +to guide the relation generation, where the [SUB] and [OBJ] tokens are embedded by the pseudo labels $b_{i}^{k}$ and $b_{j}^{k}$ , and the [REL] token is expected to be the relation predicate between $(b_{i}^{k}, b_{j}^{k})$ with $b_{i}^{k}$ as subject and $b_{j}^{k}$ as object in the text graph. To enhance relation generation, we further design three learnable positional embeddings $\mathbf{f}_{sub}$ , $\mathbf{f}_{obj}$ , $\mathbf{f}_{region}$ for indicating the different regions in the segment features. Two cross-entropy losses $\mathcal{L}_{ent}^{k}$ , $\mathcal{L}_{rel}^{k}$ are used to supervise the generation of the [ENT] and [REL] tokens, maximizing the likelihood of the target label strings, respectively. More details are in the supplementary material. + +# 4.6. Inference + +During inference, the target concepts of object semantics $\mathcal{C}_o$ and relation predicates $\mathcal{C}_r$ are known. With an image $I$ as input and an inference stage index $k_{inf}$ specified, the region grouper first partitions $I$ into several candidate segments $\{\mathbf{s}_i^{k_{inf}}\}_{i=1}^{H_{k_{inf}}}$ , which are then passed through the segment merger to obtain the similarity matrix $\mathbf{Sim}_{k_{inf}}$ . We formulate the segment merging as a spectral clustering problem and perform the graph cut [38] on $\mathbf{Sim}_{l_{inf}}$ for clustering. To improve the accuracy, we employ a matrix recovery method [25] to reduce the noise in $\mathbf{Sim}_{l_{inf}}$ . In this step, the segments with similar semantics tend to be merged into the same cluster. For each cluster and each pair of clusters, the label generator uses a similar PET to generate the object semantics and the relation predicates. For every label within sets $\mathcal{C}_o$ and $\mathcal{C}_r$ , the label generator computes its generation probability. Subsequently, these probabilities are used to rank the concepts, selecting the most probable as the + +final prediction. Note that between object and relation prediction, to convert semantic segmentation into instance segmentation, we identify each connected component in the semantic segmentation to be an instance, for simplicity. More details about inference are in the supplementary material. + +# 5. Experiments and Results + +Datasets. We train our model with the COCO Caption dataset [4], which involves 123,287 images with each labeled by 5 independent human-generated captions. Following the 2017 split, we use 118,287 images with their captions for training. We evaluate models with the Panoptic Scene Graph dataset [49] for its pixel-wise labeling as well as its high-quality object and relation annotation. We further merge the object semantics with ambiguities. After merging, 127 object semantics and 56 relation predicates are finally obtained for evaluation. More details about the datasets can be found in the supplementary material. + +Evaluation Protocol and Metrics. Following all previous works in scene graph generation, we evaluate the quality of a generated scene graph by viewing it as a set of subject-predicate-object triplets. We evaluate models on two tasks: Visual Phrase Detection (PhrDet) and Scene Graph Detection (SGDet). PhrDet aims to detect the whole phrase of subject-Predicate-object with a union location of subject and object. It is considered to be correct if the phrase labels are all correct and the union location matches the ground truth with the intersection over union (IoU) greater than 0.5. SGDet further requires a more accurate location, i.e., the location of subject and object should match the ground truth with IoU greater than 0.5 respectively. + +We use No-Graph-Constraint-X Recall@K (NXR@K, %) to measure the ability of generation. Recall@K computes the recall between the top-k generated triplets with the ground truth. No-Graph-Constraint-X indicates that at most X predicate labels could be predicted for each subject-object pair. Since some predicates defined in [49] are not exclusive, such as on and sitting on, NXR@K could be a more reasonable metric compared with Recall@K. + +Baselines. We consider several baselines for Caption-to-PSG in the following experiments. Firstly, we design four baselines that strictly follow the constraints of Caption-to-PSG, where objects are located by bbox proposals generated by selective search [43]: + +- Random is the most naive baseline where all object semantics and relation predicates are randomly predicted. +- Prior augments Random by performing label prediction based on the statistical priors in the training set. +- MIL performs the alignment between proposals and textual entities by multiple instance learning [30]. Similar to [58], it formulates the object label prediction as a classification problem in a large pre-built vocabulary, with + +
MethodModePhrDetSGDet
ModelProposalTargetN3R50N3R100N5R50N5R100N3R50N3R100N5R50N5R100
SGGNLS-cDetector✓bbox9.6911.4510.2412.226.767.817.28.65
Randomxbbox0.020.030.020.030.010.020.020.03
PriorSelective Searchxbbox0.040.070.050.070.030.060.050.07
MILxbbox1.972.182.042.611.21.351.561.97
SGCLIPxbbox3.023.453.383.712.132.32.392.7
SGGNLS-oDetectorxbbox6.26.796.927.933.964.214.535.02
Ours-xmask8.289.169.0610.513.323.633.714.18
Ours-xbbox11.3712.7412.2414.374.294.774.825.48
+ +Table 1. Quantitative Comparison of Different Methods on Caption-to-PSG. 'Proposal' indicates how the method obtains bbox proposals. 'Target' indicates whether the concept sets for inference are known during training. 'Mode' indicates the mode used for evaluation. + +WordNet [31] employed during inference. The relation labels are predicted with statistical priors, similar to Prior. + +- SGCLIP employs the pre-trained CLIP [34] to predict both object semantic labels and relation predicate labels. +Secondly, to further benchmark the performance of our framework, we set two additional baselines based on [58] by gradually removing the constraints of Caption-to-PSG: +- SGGNLS-o [58] extracts proposals with a detector [35] pre-trained on OpenImage [19]. It formulates the object and relation label prediction as a classification problem within a large pre-built vocabulary, with WordNet [31] employed during inference. +- SGGNLS-c [58] uses the same proposals as SGGNLS-o. In SGGNLS-c, the target concept sets for inference are known during training. It formulates the label prediction as a classification problem in these target concept sets. + +More design details are in the supplementary material. + +Implementation Details. Following GroupViT [48], we set $K = 2$ , $H_{1} = 64$ , and $H_{2} = 8$ for our region grouper. We leverage general pre-trained models for weight initialization. We employ the pre-trained GroupViT for the region grouper as well as $\mathbf{Tfm}^T$ in the entity grounder, and the pre-trained BLIP [21] decoder for the label generator. During training, $\mathbf{Tfm}^T$ and the label generator are frozen. During inference, we set $k_{inf} = 1$ . More implementation details can be found in the supplementary material. + +# 5.1. Main Results on Caption-to-PSG + +Quantitative Results. Our quantitative results on Caption-to-PSG are shown in Tab. 1. To make a fair comparison with bbox-based scene graphs generated by baselines, we evaluate our generated PSGs in both mask and bbox mode. For the latter, all masks in both prediction and ground truth are converted into bboxes (i.e., the mask area's enclosing rectangle) for evaluation, resulting in an easier setting than the former. The results show that our framework (Ours) significantly outperforms all the baselines under the same + +constraints on both PhrDet (14.37 vs. 3.71 N5R100) and SGDet (5.48 vs. 2.7 N5R100). Our method also shows better results compared with SGGNLS-o on all metrics and all tasks (on PhrDet, 14.37 vs. 7.93 N5R100; on SGDet, 5.48 vs. 5.02 N5R100) although SGGNLS-o utilizes location priors by leveraging a pre-trained detector. The results demonstrate that our framework is more effective for learning a good panoptic structured scene understanding. + +Qualitative Results. We provide typical qualitative results in Fig. 3 to further show our framework's effectiveness. Compared with SGGNLS-o, our framework has the following advantages. First, our framework is able to provide fine-grained semantic labels to each pixel in the image to reach a panoptic understanding, while SGGNLS-o can only provide sparse bboxes produced by the pre-trained detector. Note that categories with irregular shapes (e.g., trees in Fig. 3) are hard to be labeled precisely by bboxes. Second, compared with SGGNLS-o, our framework can generate more comprehensive object semantics and relation predicates, such as "dry grass field" and "land at" in Fig. 3, showing the open-vocabulary potential of our framework. More qualitative results are in the supplementary material. + +# 5.2. OOD Robustness Analysis + +We further analyze another key advantage of our framework, i.e., the robustness in OOD cases. Since SGGNLS-c and SGGNLS-o both rely on a pre-trained detector to locate objects, their performance highly depends on whether object semantics in the scene are covered by the detector. + +Based on the object semantics [19] covered by the detector, we split the ground truth triplets into an in-distribution (ID) set and an OOD set. For triplets within the ID set, both the subject and object semantics are covered, while for triplets in the OOD set, at least one of the semantics is not covered. As shown in Tab. 2, both SGGNLS-c and SGGNLS-o suffer a significant performance drop from the ID set to the OOD set. On the OOD set, the triplets can hardly be retrieved. However, our framework, with the ability of location learned from purely text descriptions, can + +![](images/5fd646a0764fe88146fa3c41513539928ff5b9acebc30c769b9610d834715363.jpg) + +![](images/4e89251f4d08ca6a339cd4f66c3dfd9afd8363fa94e6a3b3217ec9e15a4ed343.jpg) + +![](images/c2c9e3f598c9d86058f817e83a877ac96010667f702a6bfb1f858c2ce578417b.jpg) + +![](images/ba4b6199c1d58515ae730fc2e95c3dbd2a1f7a5d4bbec65696d9bc2150d4c5ca.jpg) + +![](images/f7188de0a861be9c3a7b18d3c91db4c0e157d7d39c6f0db9974027b2b742017f.jpg) + +![](images/8f5e498031f6f214f84c89768d1d5d55761c6c8ecd8b5e37c5723d4a86f5dd27.jpg) + +![](images/757c43a99d609870c03d35ef36d4c3aa5b76dfeb106ae0996af7516593eb7887.jpg) + +![](images/9946d972a85ce5a277190f051c4a2ac8d255ca12edbee4611f3eaa379867cde2.jpg) + +![](images/cd097e5a5a4cff4e736bf4ca5458115ca08768dc20062efc4ac13bc4fb6fd4db.jpg) + +![](images/c3fb94a60134ede9c03df6ecf4d475950e28983c6dce3822db506f42c701ae58.jpg) + +![](images/0db44181bf9f88f3acf4aae6678debcd4f1a7e4aee7b9396ead8dc7579bf256c.jpg) +(a) Results of SGGNLS-o + +![](images/3fcd7be67b3d6b4fefa7c213cb4887cc98865565b0092549333f3582d046ec37.jpg) +Figure 3. Qualitative Comparison between SGGNLS-o (a) and Ours (b). For each method, the results of object location are shown on the left, while the results of scene graph generation are shown on the right. For Ours, scene graphs predicted within the given concept sets are provided in the middle column, and scene graphs directly predicted through the auto-regressive generation (i.e., an open-vocabulary manner) in the label generator are additionally provided in the right column. + +![](images/081cd78f56524175490a5712006eec3d1bb73aca015a857908e90d5bc81638b6.jpg) + +![](images/76979ea156d605c127316bcd9c315ff7e2a2bab5f7a79d86baaf2b89d4bc0158.jpg) +(b) Results of Ours + +![](images/6358bff9e31c151ed9f639abb8b6f7d260924235a5251b491291fd88401901d9.jpg) + +
SetModelTargetModePhrDetSGDet
N3R100N5R100N3R100N5R100
IDSGGNLS-c✓bbox16.7618.4810.4511.86
SGGNLS-oXbbox11.5513.647.138.47
OursXmask9.2710.453.283.76
OursXbbox13.3514.824.635.36
OODSGGNLS-c✓bbox0000
SGGNLS-oXbbox0.050.0600
OursXmask8.479.764.074.51
OursXbbox10.1811.695.235.72
+ +Table 2. Analysis on OOD Robustness. 'Set' indicates the triplet set used for evaluation. + +
Stage#SegCutPhrDetSGDet
N3R100N5R100N3R100N5R100
164×10.7311.393.183.51
164✓12.7414.374.775.48
28×9.2411.033.534.35
28✓6.788.452.463.21
+ +reach similar performance on both sets, which demonstrates the OOD robustness of our framework for PSG generation. + +# 5.3. Ablation Studies + +We conduct additional ablation studies to evaluate the effectiveness of our design choices. For all following experiments, we report N3R100 and N5R100 evaluated in bbox mode for simplicity. We answer the following questions. Q1: Does the explicit learning of merging in the segment merger helps provide better image segments? Q2: Is the generation-based label prediction better than the + +Table 3. Ablation Study on the Segment Merger. 'Stage' indicates the grouping stage where image segments used for merging are from. $\# \mathrm{Seg}$ indicates the number of image segments. Cut indicates whether the graph-cut-based segment merging is applied. + +
Label PredictionModelPhrDetSGDet
N3R100N5R100N3R100N5R100
Cls + WordNet-8.829.362.362.72
GenRNN9.1210.442.653.07
Gen w/o PETBLIP [21]2.332.580.450.6
Gen w/ PETBLIP [21]12.6414.284.775.49
+ +Table 4. Ablation Study on the Label Generator. 'Cls' indicates classification. 'Gen' indicates generation. + +classification-based prediction? Q3: Does the pre-learned common sense from the pre-trained BLIP [21] helps with the label prediction? Q4: Does the PET helps incorporate the pre-learned common sense for label prediction? + +In Tab. 3, we compare different strategies of image segment merging during inference. Row 1&2 denote that the $H_{1} = 64$ segments from the first grouping stage are used for further merging, while row 3&4 denote that the $H_{2} = 8$ segments from the second stage are used. The results show that applying the graph cut to merge the segments from the first stage could reach the best performance, corroborating that compared with the fully implicit learning of merging, the explicit learning of merging can provide better segments (row 2 vs 3, answering Q1). + +In Tab. 4, we compare different designs of the label generator. Keeping the other modules the same, we change the label generator (row 4) into three different designs, i.e., classification within a large pre-built vocabulary followed by WordNet [31] for target matching (row 1), generation with a vanilla RNN (row 2), generation with the BLIP decoder but without the PET (row 3). The results show that with the constraint of no pre-defined concept sets, compared with formulating the label prediction into a classification problem, formulating it into a generation problem is a bet- + +ter choice (row 1 vs 2&4, answering Q2). By employing the pre-trained BLIP for leveraging the pre-learned common sense, the prediction could be further boosted (row 2 vs 4, answering Q3). And the PET is very important for incorporating the common sense from the pre-trained model (row 3 vs 4, answering Q4). + +More ablation studies for the design evaluation can be found in the supplementary material. + +# 5.4. Application on TSSS + +As a side product, we observe that our entity grounder and segment merger can also enhance TSSS. Based on the original GroupViT [48], we replace the multi + +
MethodmIoU
GroupViT [48]24.28
GroupViT† [48]24.72
Ours26.87
+ +Table 5. Results on TSSS. † indicates finetuned. + +label contrastive loss with our entity grounder and segment merger. Then we finetune the model on the COCO Caption dataset [4]. As shown in Tab. 5, compared with GroupViT directly finetuned on [4], the explicit learning of merging in our modules can boost the model with an absolute $2.15\%$ improvement of mean Intersection over Union (mIoU, $\%$ ) on COCO [23], which demonstrates the effectiveness of our proposed modules on better object location. + +# 5.5. Discussion + +Failure Case Analysis. Despite the impressive performance of TextPSG, there are still challenges to address. Upon analyzing the failure cases for PSG generation, we identify three specific limitations of TextPSG that contribute to these failures. a) The strategy we use to convert semantic segmentation into instance segmentation is not entirely effective. For simplicity, in TextPSG, we identify each connected component in the semantic segmentation to be an individual object instance. However, this strategy may fail when instances overlap or are occluded, resulting in either an underestimation or an overestimation of instances. b) Our framework faces difficulty in locating small objects in the scene due to limitations in resolution and the grouping strategy for location. c) The relation prediction of our framework requires enhancement, as it is not adequately conditioned on the image. While the label generator uses both image features and predicted object semantics to determine the relation, it sometimes seems to lean heavily on the object semantics, potentially neglecting the actual image content. Examples of failure cases for each of these limitations can be found in the supplementary material. + +Model Diagnosis. For a clearer understanding of the efficacy of our framework, we conduct a model diagnosis to answer the following question: why does our framework only achieve semantic segmentation through learning, rather than panoptic segmentation (and thus requires further + +segmentation conversion to obtain instance segmentation)? + +![](images/c626cb00f1fd940999acb0183a3d1210f60281923e2db9d49a498916ba1926eb.jpg) +(a) A white sheep and a black sheep standing on the grass under the sky. +Figure 4. Region-Entity Alignment Results of Captions in Different Granularity. Two captions in different granularity are used to execute region-entity alignment with the same image, with (a) one describing the two sheep individually while (b) the other merges them in plural form. + +![](images/ddc720bb97a834e199c54994b057015092868968038267bc7df10d73c4c98e9b.jpg) + +![](images/8feeeffbb255dd913d4f772a2159f22072d1a7bb396920dba6d87e54cb7da40c.jpg) +(b) Pair of sheep standing on open grassy field on clear day. + +In Fig. 4, we use two captions in different granularity to execute region-entity alignment. It shows that our framework has the capability to assign distinct masks to individual instances. However, the nature of caption data, where captions often merge objects of the same semantics in plural form, limits our framework from differentiating instances. It is the weak supervision provided by the caption data that constrains our framework. + +More diagnoses are in the supplementary material. + +Future Directions. In response to the limitations discussed, we outline several potential directions for enhancing our framework: a) a refined and sophisticated strategy for segmentation conversion; b) increasing the input resolution, though this may introduce greater computational demands; c) a more suitable image-conditioned reasoning mechanism for relation prediction; d) a superior image-caption-pair dataset with more detailed granularity in captions to achieve panoptic segmentation through learning. + +# 6. Conclusion + +We take the first step towards the novel problem Caption-to-PSG, aiming to learn PSG generation purely from language. To tackle this challenging problem, we propose a new modularized framework TextPSG with several novel techniques, which significantly outperforms the baselines and achieves strong OOD robustness. This paves the path to a more comprehensive and generalizable panoptic structured scene understanding. There are still bottle-necks in TextPSG to be explored in future work, including a) a more sophisticated strategy for segmentation conversion; b) a more suitable image-conditioned reasoning mechanism for relation prediction; c) a superior image-caption-pair dataset for panoptic segmentation through learning. + +Acknowledgements. This work was supported by the DSO grant DSOCO21072, IBM, and gift funding from MERL, Cisco, and Amazon. We would also like to thank the computation support from AiMOS, a server cluster for the IBM Research AI Hardware Center. + +# References + +[1] Somak Aditya, Yezhou Yang, Chitta Baral, Yiannis Aloimonos, and Cornelia Fermüller. Image understanding using vision and reasoning through scene description graph. Computer Vision and Image Understanding, 2017. 1 +[2] S. Amiri, Kishan Chandan, and Shiqi Zhang. Reasoning with scene graphs for robot planning under partial observability. IEEE Robotics and Automation Letters, 7:5560-5567, 2022. 1 +[3] Gabor Angeli, Melvin Johnson, and Christopher D. Manning. Leveraging linguistic structure for open domain information extraction. In Annual Meeting of the Association for Computational Linguistics, 2015. 4 +[4] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6, 9 +[5] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee K Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. In Proc. 57th Annual Meeting of the Association for Computational Linguistics, 2019. 3 +[6] Zhenfang Chen, Jiayuan Mao, Jiajun Wu, Kwan-Yee K Wong, Joshua B. Tenenbaum, and Chuang Gan. Grounding physical concepts of objects and events through dynamic visual reasoning. In International Conference on Learning Representations, 2021. 3 +[7] Zhenfang Chen, Peng Wang, Lin Ma, Kwan-Yee K Wong, and Qi Wu. Cops-ref: A new dataset and task on compositional referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10086-10095, 2020. 3 +[8] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D. Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In CVPR, 2020. 1 +[9] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3 +[10] Samir Yitzhak Gadre, Kiana Ehsani, Shuran Song, and Roozbeh Mottaghi. Continuous scene representations for embodied ai. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14829-14839, 2022. 1 +[11] Chuang Gan, Yandong Li, Haoxiang Li, Chen Sun, and Boqing Gong. Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1811-1820, 2017. 3 +[12] Jiumiang Gu, Handong Zhao, Zhe Lin, Sheng Li, Jianfei Cai, and Mingyang Ling. Scene graph generation with external knowledge and image reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3 + +[13] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 3 +[14] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In CVPR, 2018. 1 +[15] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3128-3137, 2015. 3 +[16] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 3 +[17] Siddhesh Khandelwal, Mohammed Suhail, and Leonid Sigal. Segmentation-grounded scene graph generation. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15859-15869, 2021. 3 +[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Fei-Fei Li. Visual Genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32-73, 2016. 3 +[19] Alina Kuznetsova, Hassan Rom, Neil Gordon Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4. International Journal of Computer Vision, 128:1956-1981, 2018. 7 +[20] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022. 3 +[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 5, 7, 8 +[22] Xingchen Li, Long Chen, Wenbo Ma, Yi Yang, and Jun Xiao. Integrating object-aware and interaction-aware knowledge for weakly supervised scene graph generation. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2, 3 +[23] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 9 +[24] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3 +[25] Guangcan Liu, Zhouchen Lin, Shuicheng Yan, Ju Sun, Yong Yu, and Yi Ma. Robust recovery of subspace structures by low-rank representation. IEEE transactions on pattern analysis and machine intelligence, 35(1):171-184, 2012. 6 + +[26] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7086-7096, 2022. 3 +[27] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. arXiv preprint arXiv:2211.14813, 2022. 3 +[28] Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David McClosky. The Stanford CoreNLP natural language processing toolkit. In Association for Computational Linguistics (ACL) System Demonstrations, pages 55–60, 2014. 4 +[29] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 3 +[30] Oded Maron and Tomás Lozano-Pérez. A framework for multiple-instance learning. Advances in neural information processing systems, 10, 1997. 6 +[31] George A. Miller. WordNet: A lexical database for English. In Speech and Natural Language: Proceedings of a Workshop Held at Harriman, New York, February 23-26, 1992, 1992. 5, 7, 8 +[32] Julia Peyre, Ivan Laptev, Cordelia Schmid, and Josef Sivic. Weakly-supervised learning of visual relations. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5189-5198, 2017. 3 +[33] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649, 2015. 3 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763, 2021. 7 +[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2, 3, 7 +[36] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 817-834. Springer, 2016. 3 +[37] Sebastian Schuster, Ranjay Krishna, Angel X. Chang, Li Fei-Fei, and Christopher D. Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In VL@EMNLP, 2015. 3 +[38] Jianbo Shi and J. Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000. 6 + +[39] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8376-8384, 2019. 1 +[40] Jing Shi, Yiwu Zhong, Ning Xu, Yin Li, and Chenliang Xu. A simple baseline for weakly-supervised scene graph generation. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16373-16382, 2021. 3 +[41] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3 +[42] Damien Teney, Lingqiao Liu, and Anton van den Hengel. Graph-structured representations for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1 +[43] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104:154-171, 2013. 3, 6 +[44] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 4 +[45] Qixun Wang, Xiaofeng Guo, and Haofan Wang. 1st place solution for psg competition with eccv'22 sensehuman workshop. ArXiv, abs/2302.02651, 2023. 2 +[46] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6602–6611, 2019. 3 +[47] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5419, 2017. 3 +[48] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. GroupViT: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18134-18144, June 2022. 2, 3, 4, 5, 7, 9 +[49] Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. In ECCV, 2022. 1, 2, 3, 6 +[50] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 3 +[51] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. FILIP: Fine-grained interactive language-image pre-training. In International Conference on Learning Representations, 2022. 2, 4, 5 +[52] Keren Ye and Adriana Kovashka. Linguistic structures as weak supervision for visual scene graph generation. In Pro- + +ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 2, 3 +[53] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1307-1315, 2018. 3 +[54] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3 +[55] Alireza Zareian, Svebor Karaman, and Shih-Fu Chang. Weakly supervised visual semantic parsing. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3733-3742, 2020. 3 +[56] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 3 +[57] Hanwang Zhang, Zawlin Kyaw, Jinyang Yu, and Shih-Fu Chang. Ppr-fcn: Weakly supervised visual relation detection via parallel pairwise r-fcn. In Proceedings of the IEEE international conference on computer vision, pages 4233-4241, 2017. 3 +[58] Yiwu Zhong, Jing Shi, Jianwei Yang, Chenliang Xu, and Yin Li. Learning to generate scene graph from natural language supervision. In ICCV, 2021. 2, 3, 5, 6, 7 +[59] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVIII, pages 696-712. Springer, 2022. 3 \ No newline at end of file diff --git a/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/images.zip b/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4c16e8447075d94edb6ee6a0590f72dd1ce7e162 --- /dev/null +++ b/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca3c6df5c47126a161a0ec3634e1ee161b272bd36e927004c4269501ae71f5e9 +size 468620 diff --git a/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/layout.json b/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e4c39a0d4b4c2204e5d80f2f5daefe4c367470a6 --- /dev/null +++ b/2023/TextPSG_ Panoptic Scene Graph Generation from Textual Descriptions/layout.json @@ -0,0 +1,10873 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 517, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 517, + 122 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 517, + 122 + ], + "type": "text", + "content": "TextPSG: Panoptic Scene Graph Generation from Textual Descriptions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 142, + 431, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 142, + 431, + 200 + ], + "spans": [ + { + "bbox": [ + 161, + 142, + 431, + 200 + ], + "type": "text", + "content": "Chengyang Zhao1 Yikang Shen2 Zhenfang Chen2 Mingyu Ding3 Chuang Gan2,4 \n1Peking University 2MIT-IBM Watson AI Lab \n3UC Berkley 4UMass Amherst" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 251, + 290, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 290, + 622 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 290, + 622 + ], + "type": "text", + "content": "Panoptic Scene Graph has recently been proposed for comprehensive scene understanding. However, previous works adopt a fully-supervised learning manner, requiring large amounts of pixel-wise densely-annotated data, which is always tedious and expensive to obtain. To address this limitation, we study a new problem of Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG). The key idea is to leverage the large collection of free image-caption data on the Web alone to generate panoptic scene graphs. The problem is very challenging for three constraints: 1) no location priors; 2) no explicit links between visual regions and textual entities; and 3) no predefined concept sets. To tackle this problem, we propose a new framework TextPSG consisting of four modules, i.e., a region grouper, an entity grounder, a segment merger, and a label generator, with several novel techniques. The region grouper first groups image pixels into different segments and the entity grounder then aligns visual segments with language entities based on the textual description of the segment being referred to. The grounding results can thus serve as pseudo labels enabling the segment merger to learn the segment similarity as well as guiding the label generator to learn object semantics and relation predicates, resulting in a fine-grained structured scene understanding. Our framework is effective, significantly outperforming the baselines and achieving strong out-of-distribution robustness. We perform comprehensive ablation studies to corroborate the effectiveness of our design choices and provide an in-depth analysis to highlight future directions. Our code, data, and results are available on our project page: https://vis-ww.c.s.umass.edu/TextPSG." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 645, + 128, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 128, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 128, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 715 + ], + "type": "text", + "content": "A scene graph is a directed-graph-based abstract representation of the objects and their relations within a scene. It has been widely utilized to develop a structured scene understanding of object semantics, locations, and relations," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 225, + 382, + 361 + ], + "blocks": [ + { + "bbox": [ + 308, + 225, + 382, + 361 + ], + "lines": [ + { + "bbox": [ + 308, + 225, + 382, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 225, + 382, + 361 + ], + "type": "image", + "image_path": "065d0ed580478e3af013e3fcf045d049fb6c45002c8ef8797ce8355a2355b286.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 365, + 546, + 453 + ], + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 453 + ], + "type": "text", + "content": "Figure 1. Problem Overview. Different from the traditional bbox-based form of the scene graph as shown in (a), Caption-to-PSG aims to generate the mask-based panoptic scene graph. In Caption-to-PSG, the model has no access to any location priors, explicit region-entity links, or pre-defined concept sets. Consequently, the model is required to learn partitioning and grounding as illustrated in (b), as well as object semantics and relation predicates as illustrated in (c), all purely from textual descriptions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 389, + 226, + 544, + 361 + ], + "blocks": [ + { + "bbox": [ + 389, + 226, + 544, + 361 + ], + "lines": [ + { + "bbox": [ + 389, + 226, + 544, + 361 + ], + "spans": [ + { + "bbox": [ + 389, + 226, + 544, + 361 + ], + "type": "image", + "image_path": "0254e5361ce1c24aa5539ff5c6c3b04b2ce638ce0bbed7ee183c246ba9132828.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 461, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 498 + ], + "type": "text", + "content": "which facilitates a variety of downstream applications, such as image generation [14, 8], visual reasoning [42, 1, 39], and robotics [2, 10]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": "Typically, in a scene graph, each node denotes an object in the scene located by a bounding box (bbox) with a semantic label, and each directed edge denotes the relation between a pair of objects with a predicate label. Nonetheless, a recent work [49] points out that such a bbox-based form of scene graph is not ideal enough. Firstly, compared with pixel-wise segmentation masks, bboxes are less fine-grained and may contain some noisy pixels belonging to other objects, limiting the applications for some downstream tasks. For example, as shown in Fig. 1 (a), about half of the pixels in the yellow_bbox for girl belong to wall. Secondly, it is challenging for bboxes to cover the entire scene without ambiguities caused by overlaps, which prevents a scene graph from including every object in the scene for a complete description. To this end, the work [49] proposes the concept of Panoptic Scene Graph (PSG), in which each object is grounded by a panoptic segmentation mask, to reach a comprehensive structured scene representation." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2839" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "However, all existing works [49, 45] approach PSG generation through a fully-supervised manner, i.e., learning to perform panoptic segmentation and relation prediction from manually-annotated datasets with explicit supervision for both segmentation and relation prediction. Unfortunately, it is extremely labor-intensive to build such datasets, making it difficult to scale up to cover more complex scenes, object semantics, and relation predicates, thus significantly limiting the generalizability and the application of these methods to the real world. For instance, the current PSG dataset [49] only covers 133 object semantics and 56 relation predicates." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 206, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 350 + ], + "type": "text", + "content": "To relieve the reliance on densely-annotated data, weakly-supervised methods [52, 58, 22] for scene graph generation are promising. These methods could induce scene graphs from image-caption pairs, which can be easily harvested from the Web for free. Even so, they still rely heavily on two strong preconditions, i.e., a powerful region proposal network (e.g., [35]) and a pre-defined set of object semantics and relation predicates. Although these preconditions facilitate the learning process of the methods, they also limit the generalizability for locating novel objects (unforeseen objects for the region proposal network) and constrain the understanding into the limited concept set." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 352, + 289, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 289, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 289, + 519 + ], + "type": "text", + "content": "Inspired by previous weakly-supervised methods, we introduce a new problem, Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG), to explore a holistic structured scene understanding without labor-intensive data annotation. Considering the limitation of the preconditions mentioned, we set three constraints to Caption-to-PSG to reach a more comprehensive and generalizable understanding, which results in a very challenging problem: a) only image-caption pairs are provided during training, without any location priors in either region proposals or location supervision; b) the explicit links between regions in images and entities in captions are missing; c) no concept sets are pre-defined, i.e., neither object semantics nor relation predicates are known beforehand." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": "Given these three constraints, we argue that there are two key challenges for the model to solve the problem. Firstly, the model should learn to ground entities in language onto the visual scene without explicit location supervision, i.e., the ability to perform partitioning and grounding, as shown in Fig. 1 (b), should be developed purely from textual descriptions. Secondly, during training, the model should also learn the object semantics and relation predicates from textual descriptions, as shown in Fig. 1 (c), without pre-defined fixed object and relation vocabularies. By solving these challenges, the model could associate visual scene patterns with textual descriptions, gradually acquire common sense among them, and finally reach a more comprehensive and generalizable understanding, including novel object location, extensive semantics recognition, and complex relation analysis, which is more suitable to the real world." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 454 + ], + "type": "text", + "content": "With these considerations, we propose a novel framework, TextPSG, as the first step towards this challenging problem. TextPSG consists of a series of modules to cooperate with each other, i.e., a region grouper, an entity grounder, a segment merger, and a label generator. The region grouper learns to merge image regions into several segments in a hierarchical way based on object semantics, similar to [48]. The entity grouder employs a fine-grained contrastive learning strategy [51] to bridge the textual description and the visual content, grounding entities in the caption onto the image segments. With the entity-grounding results as pseudo labels, the segment merger learns similarity matrices to merge small image segments during inference, while the label generator learns the prediction of object semantics and relation predicates. Specifically, in the segment merger, we propose to leverage the grounding as explicit supervision for merging, compared with [48] which learns merging in a fully implicit manner, to improve the ability of location. In the label generator, different from all previous pipelines for scene graph generation, we reformulate the label prediction as an auto-regressive generation problem rather than a classification problem, and employ a pre-trained language model [21] as the decoder to leverage the pre-learned common sense. We further design a novel prompt-embedding-based technique (PET) to better incorporate common sense from the language model. Our experiments show that TextPSG significantly outperforms the baselines and achieves strong out-of-distribution (OOD) robustness. Comprehensive ablation studies corroborate the effectiveness of our design choices. As a side product, the proposed grouder and merger modules also have been observed to enhance text-supervised semantic segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 455, + 547, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 547, + 502 + ], + "type": "text", + "content": "In spite of the promising performance of TextPSG, certain challenges persist. We delve into an in-depth analysis of the failure cases, provide a model diagnosis, and discuss potential future directions for enhancing our framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 503, + 498, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 503, + 498, + 514 + ], + "spans": [ + { + "bbox": [ + 318, + 503, + 498, + 514 + ], + "type": "text", + "content": "To sum up, our contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 514, + 545, + 651 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 514, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 545, + 563 + ], + "type": "text", + "content": "- We introduce a new problem, Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG), to alleviate the burden of human annotation for PSG by learning purely from the weak supervision of captions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 565, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 565, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 565, + 545, + 624 + ], + "type": "text", + "content": "- We propose a new modularized framework, TextPSG, with several novel techniques, which significantly outperforms the baselines and achieves strong OOD robustness. We demonstrate that the proposed modules in TextPSG can also facilitate text-supervised semantic segmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 628, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 628, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 628, + 545, + 651 + ], + "type": "text", + "content": "- We perform an in-depth failure case analysis with a model diagnosis, and further highlight future directions." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 657, + 392, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 392, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 392, + 670 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Bbox-based Scene Graph Generation. Bbox-based scene graph generation aims to create a structured representation of object semantics, locations, and relations in the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2840" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "scene, where each object is identified by a bbox. Most of existing works [50, 47, 41, 12, 24] follow a fully-supervised approach to learn the generation from densely-annotated datasets [18, 13], which requires significant human labors. To reduce the labeling effort, some weakly-supervised methods have been proposed [32, 55, 57, 40]. Recent works [52, 58, 22] further explore learning scene graph generation from image-caption pairs. However, they all rely on off-the-shelf region proposal networks for the location of objects in the scene, which are typically pretrained on pre-defined fixed sets of object semantics, limiting their generalizability to locating unforeseen objects. To reach a more granular and accurate grounding, [17] proposes to ground each object by segmentation. A recent work [49] further introduces the concept of PSG, where each object is identified by a panoptic segmentation mask, as a more comprehensive scene representation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 430 + ], + "type": "text", + "content": "Text-supervised Semantic Segmentation (TSSS). TSSS [48, 20, 26, 9, 27, 59] aims to learn image pixel semantic labeling from image-caption pairs without fine-grained pixel-wise annotations. Similar to TSSS, our proposed Caption-to-PSG aims to learn to connect visual regions and textual entities from only image-caption pairs and has the potential to leverage the large collection of free data on the Web. However, different from TSSS, Caption-to-PSG further requires the model to learn the relations among different visual regions, resulting in a higher-order structured understanding of visual scenes. In addition to unknown object semantics, Caption-to-PSG does not assume any pre-defined relation predicate concepts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 431, + 289, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 289, + 600 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 289, + 600 + ], + "type": "text", + "content": "Visual Grounding. Our work is also related to visual grounding [16, 56, 29, 11, 6, 33], which grounds entities in language onto objects in images. Early works [54, 53, 7] on visual grounding typically detect object proposals [35, 43] from images first and then match them with language descriptions by putting features of both modalities into the same feature space, which are in a fully-supervised learning manner. There are also some weakly-supervised grounding methods [15, 36, 5] which relieve the need for dense regional annotations by multiple instance learning [15] or learning to reconstruct [36]. Different from them, Caption-to-PSG is more challenging since it requires grounding fine-grained object relations between entities without region proposal networks for a pre-defined object vocabulary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 609, + 174, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 174, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 174, + 623 + ], + "type": "text", + "content": "3. Problem Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Panoptic Scene Graph Generation from Purely Textual Descriptions (Caption-to-PSG). A PSG " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V},\\mathcal{E})" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " is a directed graph representation of the objects and the relations among them in a scene image " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "I\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ". Each node " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "v_{i}\\in \\mathcal{V}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " denotes an object in " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " located by a panoptic segmentation mask " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "m_i\\in \\{0,1\\}^{H\\times W}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " with an object semantic label " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "o_i\\in \\mathcal{C}_o" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ", and each directed edge " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "e_{ij}\\in \\mathcal{E}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " denotes a" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "text", + "content": "pair of subject " + }, + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "text", + "content": " and object " + }, + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "text", + "content": " with a relation predicate label " + }, + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "r_{ij} \\in \\mathcal{C}_r" + }, + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_o" + }, + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_r" + }, + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "text", + "content": " are the defined concept sets of object semantics and relation predicates. Note that for a PSG, it is constrained that all segmentation masks could not overlap, i.e., " + }, + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{|\\mathcal{V}|} m_i \\leq 1^{H \\times W}" + }, + { + "bbox": [ + 304, + 72, + 545, + 134 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "text", + "content": "Given a large collection of paired scene images and textual descriptions " + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "inline_equation", + "content": "S = \\{(I_i, T_i)\\}_i" + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "text", + "content": ", Caption-to-PSG aims to learn PSG generation from purely text descriptions for a holistic structured scene understanding, i.e., during training, only " + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "text", + "content": " is provided as supervision, while during inference, with a scene observation " + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "text", + "content": " as input, the model is required to generate a corresponding PSG " + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{G}'" + }, + { + "bbox": [ + 304, + 134, + 547, + 218 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 219, + 547, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 547, + 351 + ], + "type": "text", + "content": "Three Constraints. Note that in Caption-to-PSG, three important constraints are set to reach a more comprehensive and generalizable scene understanding: a) no location priors: different from all previous scene graph generation methods, neither pre-trained region proposal networks nor location supervision are allowed; b) no explicit region-entity links: the links between regions in the image " + }, + { + "bbox": [ + 304, + 219, + 547, + 351 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 219, + 547, + 351 + ], + "type": "text", + "content": " and entities in the textual description " + }, + { + "bbox": [ + 304, + 219, + 547, + 351 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 219, + 547, + 351 + ], + "type": "text", + "content": " are not provided; c) no pre-defined concept sets: the target concept sets defined for inference, i.e., object semantics " + }, + { + "bbox": [ + 304, + 219, + 547, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_o" + }, + { + "bbox": [ + 304, + 219, + 547, + 351 + ], + "type": "text", + "content": " and relation predicates " + }, + { + "bbox": [ + 304, + 219, + 547, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_r" + }, + { + "bbox": [ + 304, + 219, + 547, + 351 + ], + "type": "text", + "content": ", are unknown during training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 363, + 361, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 361, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 361, + 376 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 384, + 545, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 432 + ], + "type": "text", + "content": "Overview. As illustrated in Fig. 2, our proposed framework for Caption-to-PSG, TextPSG, contains four modules to cooperate with each other: a region grouper, an entity grounder, a segment merger, and a label generator." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 547, + 564 + ], + "type": "text", + "content": "During training, TextPSG takes batched image-caption pairs as input. For each pair, the image is passed through the region grouper to be partitioned into several image segments, while the caption is first pre-processed to extract its linguistic structure as a text graph and then taken by the entity grounder to ground textual entities in the graph onto the image segments. With the grounding results as pseudo labels, the segment merger learns similarity matrices between small image segments for further merging during inference, while the label generator learns the prediction of object semantics and relation predicates." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 565, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 637 + ], + "type": "text", + "content": "During inference, for each input image, the image segments output from the region grouper are directly passed to the segment merger to be further merged according to the learned similarity matrices, and then fed to the label generator to predict the object semantic labels and the relation predicate labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 646, + 451, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 646, + 451, + 659 + ], + "spans": [ + { + "bbox": [ + 305, + 646, + 451, + 659 + ], + "type": "text", + "content": "4.1. Text Graph Preprocessing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": "Following previous methods [52, 58, 22] that leverage a rule-based language parser [46] based on [37] to preprocess textual descriptions, in TextPSG, we employ the same parser to extract linguistic structures from captions. Addi" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2841" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 547, + 290 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 290 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 290 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 290 + ], + "type": "image", + "image_path": "1088337b43e50200b645149a7c81c28956c4478e71e9ee822b3f5e51652acf2b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 292, + 547, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 292, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 547, + 348 + ], + "type": "text", + "content": "Figure 2. Framework Overview of TextPSG. The framework consists of four modules cooperating with each other: a region grouper to merge regions in the input image into several segments, an entity grounder to ground entities in the caption onto the image segments, a segment merger to learn similarity matrices to merge small image segments during inference, and a label generator to learn the prediction of object semantics and relation predicates. The solid arrows indicate the training flow, while the dash arrows indicate the inference flow. The arrows from the region grouper to the label generator indicating the segment feature and mask query are omitted." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 360, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 289, + 445 + ], + "type": "text", + "content": "tionally, inspired by the success of open information extraction (OpenIE) [3] in natural language processing, we also employ an OpenIE system from Stanford CoreNLP [28] for extraction as a supplement. After merging, for each caption, we obtain its linguistic structure represented in a text graph, where each node denotes an entity, and each directed edge denotes the relation between an entity pair." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 450, + 146, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 146, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 146, + 464 + ], + "type": "text", + "content": "4.2. Region Grouper" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 469, + 287, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 505 + ], + "type": "text", + "content": "With a scene image as input, the region grouper aims to merge the regions with similar object semantics into several segments and extract corresponding high-level features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "text", + "content": "Our region grouper follows the hierarchical design of GroupViT [48]. Given an input image, the grouper first splits the image into " + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "text", + "content": " non-overlapping patches as the initial image segments " + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_i^0\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "text", + "content": ". These segments are then passed through " + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "text", + "content": " grouping layers, where they are merged into larger, arbitrary-shaped segments progressively. Specifically, within each grouping layer " + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{Grp}_k(k = 1,2,\\dots ,K)" + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "inline_equation", + "content": "H_{k}" + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "text", + "content": " grouping centers " + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_i^k\\}_{i = 1}^{H_k}" + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "text", + "content": " could be learned. The grouping operation is performed through an attention mechanism between the centers and the segments, merging " + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "inline_equation", + "content": "H_{k - 1}" + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "text", + "content": " input segments into " + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "inline_equation", + "content": "H_{k}" + }, + { + "bbox": [ + 46, + 506, + 288, + 638 + ], + "type": "text", + "content": " larger ones, i.e.," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 81, + 643, + 254, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 643, + 254, + 659 + ], + "spans": [ + { + "bbox": [ + 81, + 643, + 254, + 659 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {s} _ {i} ^ {k} \\right\\} _ {i = 1} ^ {H _ {k}} = \\mathbf {G r p} _ {k} \\big (\\left\\{\\mathbf {c} _ {i} ^ {k} \\right\\} _ {i = 1} ^ {H _ {k}}, \\left\\{\\mathbf {s} _ {i} ^ {k - 1} \\right\\} _ {i = 1} ^ {H _ {k - 1}} \\big).", + "image_path": "0f8bbb95d5f0ddc136a3a7f03429b340a32c2bf9188fce8322e2b6a49663d2bc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "H_0 = N" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ". After the hierarchical grouping, multiple groups of segments " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_i^k\\}_{i=1}^{H_k}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " at different grouping stages are obtained. More details about the design of " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Grp}_k\\}_{k=1}^K" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " can be found in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 360, + 408, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 408, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 408, + 372 + ], + "type": "text", + "content": "4.3. Entity Grounder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 380, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 452 + ], + "type": "text", + "content": "Since the explicit region-entity links are not provided, bridging the textual description and the visual content automatically plays an important role in solving Caption-to-PSG. Inspired by FILIP [51], in TextPSG, we employ a similar fine-grained contrastive learning strategy to perform region-entity alignment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": "For each grouping stage " + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": ", on image side, the grounder projects the segment group " + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_i^k\\}_{i=1}^{H_k}" + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": " into a new feature space " + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": " by a multi-layer perceptron (MLP) " + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{Proj}_k^I" + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": " to obtain segment embeddings " + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^k\\}_{i=1}^{H_k}" + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": ". On text side, the input caption is first tokenized into " + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": " tokens " + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{t}_i\\}_{i=1}^M" + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": ", which are then processed by a Transformer [44] " + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{Tfm}^T" + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": " to propagate information between each other. A recurrent neural network (RNN) " + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{Rnn}" + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": " further merges the tokens corresponding to the same entity, encoding tokens into their associated weights one by one and utilizing weighted sum to merge the token features into a singular entity feature. Finally, these entity features are projected to the same feature space " + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": " by a MLP " + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{Proj}^T" + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": " to obtain entity embeddings " + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{y}_i\\}_{i=1}^E" + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 453, + 546, + 623 + ], + "type": "text", + "content": " denotes the number of entities in the caption." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "type": "text", + "content": "With the segment embeddings and the entity embeddings in the shared feature space " + }, + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "type": "text", + "content": ", we compute their token-wise similarities. Specifically, for the " + }, + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "type": "text", + "content": "-th segment, we compute its cosine similarities with all entities to obtain the token-wise similarity from the " + }, + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "type": "text", + "content": "-th segment to the caption " + }, + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "type": "inline_equation", + "content": "p_i^k" + }, + { + "bbox": [ + 304, + 624, + 545, + 685 + ], + "type": "text", + "content": " via" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 374, + 696, + 477, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 696, + 477, + 715 + ], + "spans": [ + { + "bbox": [ + 374, + 696, + 477, + 715 + ], + "type": "interline_equation", + "content": "p _ {i} ^ {k} = \\max _ {1 \\leq j \\leq E} \\cos [ \\mathbf {x} _ {i} ^ {k}, \\mathbf {y} _ {j} ],", + "image_path": "a0de75b0b1a2c36ee88b61386a89f9ef42ecc77e24cae16ba95f225721a5fb21.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2842" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\cos[\\cdot, \\cdot]" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " denotes the cosine similarity operation. Note that different from the original FILIP [51], in the scenario of region-entity alignment, some regions in the scene may not be described in the caption, while some entities in the caption may not exist in the scene. To tackle this problem, we propose to set a filtering threshold " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ", where pairs with similarity lower than " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " will be considered in different semantics and filtered out. The fine-grained similarity from the image to the caption " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "p^k" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " can thus be computed via" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 186, + 241, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 186, + 241, + 220 + ], + "spans": [ + { + "bbox": [ + 91, + 186, + 241, + 220 + ], + "type": "interline_equation", + "content": "p ^ {k} = \\frac {1}{\\sum_ {i = 1} ^ {H _ {k}} {\\mathbf {1} _ {p _ {i} ^ {k} > \\theta}}} \\sum_ {i = 1} ^ {H _ {k}} (p _ {i} ^ {k} \\cdot \\mathbf {1} _ {p _ {i} ^ {k} > \\theta}).", + "image_path": "79409ea6f6d2edf1b425debb9cf9ffcb5278e52b77ba4381b44cd629af038382.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 225, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 225, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 225, + 287, + 251 + ], + "type": "text", + "content": "Similarly, we can also compute the token-wise similarity from the " + }, + { + "bbox": [ + 47, + 225, + 287, + 251 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 225, + 287, + 251 + ], + "type": "text", + "content": "-th entity to the image " + }, + { + "bbox": [ + 47, + 225, + 287, + 251 + ], + "type": "inline_equation", + "content": "q_j^k" + }, + { + "bbox": [ + 47, + 225, + 287, + 251 + ], + "type": "text", + "content": " via" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 256, + 220, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 256, + 220, + 277 + ], + "spans": [ + { + "bbox": [ + 113, + 256, + 220, + 277 + ], + "type": "interline_equation", + "content": "q _ {j} ^ {k} = \\max _ {1 \\leq i \\leq H _ {k}} \\cos [ \\mathbf {x} _ {i} ^ {k}, \\mathbf {y} _ {j} ],", + "image_path": "60e169c0a6b5ffd0f47e51de9f9b08374233e9d79281564aac09632ce27f6666.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 282, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 307 + ], + "type": "text", + "content": "and the fine-grained similarity from the caption to the image " + }, + { + "bbox": [ + 46, + 282, + 287, + 307 + ], + "type": "inline_equation", + "content": "q^{k}" + }, + { + "bbox": [ + 46, + 282, + 287, + 307 + ], + "type": "text", + "content": " via" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 307, + 241, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 307, + 241, + 339 + ], + "spans": [ + { + "bbox": [ + 91, + 307, + 241, + 339 + ], + "type": "interline_equation", + "content": "q ^ {k} = \\frac {1}{\\sum_ {j = 1} ^ {E} \\mathbf {1} _ {q _ {j} ^ {k} > \\theta}} \\sum_ {j = 1} ^ {E} (q _ {j} ^ {k} \\cdot \\mathbf {1} _ {q _ {j} ^ {k} > \\theta}).", + "image_path": "a190ebcdbec86455a224765ba2573b407ad4c081d330487eebbb2da396581e10.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "text", + "content": "Denoting the training batch with batch size " + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\{(I_i,T_i)\\}_{i = 1}^B" + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "text", + "content": ", the fine-grained similarity from the image " + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "text", + "content": " to the caption " + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "inline_equation", + "content": "p^{k,i\\rightarrow j}" + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "text", + "content": " and from the caption " + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "text", + "content": " to the image " + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "inline_equation", + "content": "q^{k,j\\rightarrow i}" + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "text", + "content": ", the image-to-text fine-grained contrastive loss " + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fine}^{k,I\\to T}" + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "text", + "content": " and the text-to-image fine-grained contrastive loss " + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fine}^{k,T\\to I}" + }, + { + "bbox": [ + 47, + 342, + 287, + 419 + ], + "type": "text", + "content": " can then be formulated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 83, + 426, + 250, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 426, + 250, + 471 + ], + "spans": [ + { + "bbox": [ + 83, + 426, + 250, + 471 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f i n e} ^ {k, I \\rightarrow T} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\frac {\\exp \\left(p ^ {k , i \\rightarrow i} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(p ^ {k , i \\rightarrow j} / \\tau\\right)},", + "image_path": "0bef895c7e6539188b7a5d12a51121f953d9dde199312f331d419cfd59d7e90b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 83, + 477, + 250, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 477, + 250, + 521 + ], + "spans": [ + { + "bbox": [ + 83, + 477, + 250, + 521 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f i n e} ^ {k, T \\rightarrow I} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\frac {\\exp \\left(q ^ {k , i \\rightarrow i} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(q ^ {k , i \\rightarrow j} / \\tau\\right)},", + "image_path": "9ec63c681c400952d420d9779ed397f8ead860efd13372326e77ebd052182e29.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 525, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 287, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 525, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 525, + 287, + 548 + ], + "type": "text", + "content": " is a learnable temperature. The total fine-grained contrastive loss is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 100, + 552, + 233, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 552, + 233, + 575 + ], + "spans": [ + { + "bbox": [ + 100, + 552, + 233, + 575 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f i n e} ^ {k} = \\frac {1}{2} (\\mathcal {L} _ {f i n e} ^ {k, I \\rightarrow T} + \\mathcal {L} _ {f i n e} ^ {k, T \\rightarrow I}).", + "image_path": "5f2f363f38d9d5cd24d1ba33c4a9f6e47894a825541278c41d7a6bb0fa01d47d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 582, + 287, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 631 + ], + "type": "text", + "content": "By minimizing " + }, + { + "bbox": [ + 47, + 582, + 287, + 631 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fine}^{k}" + }, + { + "bbox": [ + 47, + 582, + 287, + 631 + ], + "type": "text", + "content": " at all grouping stages during training, our framework could reach a meaningful fine-grained alignment automatically, i.e., for the " + }, + { + "bbox": [ + 47, + 582, + 287, + 631 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 582, + 287, + 631 + ], + "type": "text", + "content": "-th segment " + }, + { + "bbox": [ + 47, + 582, + 287, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i^k" + }, + { + "bbox": [ + 47, + 582, + 287, + 631 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 47, + 582, + 287, + 631 + ], + "type": "inline_equation", + "content": "l_i^k" + }, + { + "bbox": [ + 47, + 582, + 287, + 631 + ], + "type": "text", + "content": "-th entity satisfying" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 113, + 635, + 220, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 635, + 220, + 657 + ], + "spans": [ + { + "bbox": [ + 113, + 635, + 220, + 657 + ], + "type": "interline_equation", + "content": "l_{i}^{k} = \\operatorname *{arg max}_{1\\leq j\\leq E}\\cos [\\mathbf{x}_{i}^{k},\\mathbf{y}_{j}]", + "image_path": "bd8c97839e3957772ae803323abd4c1d8f779edc2803dc4756f5ea12838e92e5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "text", + "content": "tends to have a similar semantics with " + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i^k" + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "text", + "content": ". We thus obtain " + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{l_i^k\\}_{i=1}^{H_k}" + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "text", + "content": " as the grounding results for the image segments " + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_i^k\\}_{i=1}^{H_k}" + }, + { + "bbox": [ + 47, + 664, + 287, + 714 + ], + "type": "text", + "content": ". A further explanation of the automatic meaningful alignment can be found in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 407, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 407, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 407, + 85 + ], + "type": "text", + "content": "4.4. Segment Merger" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 90, + 545, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 149 + ], + "type": "text", + "content": "To improve the ability of location, we propose to leverage the entity-grounding results as explicit supervision to learn a group of similarity matrices between image segments for small segments merging during inference, compared with [48] that learns the merging fully implicitly." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 150, + 545, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 150, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 150, + 545, + 198 + ], + "type": "text", + "content": "For each grouping stage " + }, + { + "bbox": [ + 304, + 150, + 545, + 198 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 150, + 545, + 198 + ], + "type": "text", + "content": ", we compute the cosine similarity between each pair of image segments, which is then linearly re-scaled into " + }, + { + "bbox": [ + 304, + 150, + 545, + 198 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 304, + 150, + 545, + 198 + ], + "type": "text", + "content": " to formulate a similarity matrix " + }, + { + "bbox": [ + 304, + 150, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{Sim}_k \\in [0,1]^{H_k \\times H_k}" + }, + { + "bbox": [ + 304, + 150, + 545, + 198 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 355, + 204, + 495, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 204, + 495, + 227 + ], + "spans": [ + { + "bbox": [ + 355, + 204, + 495, + 227 + ], + "type": "interline_equation", + "content": "\\mathbf {S i m} _ {k} [ i, j ] = \\frac {1}{2} \\left(\\cos \\left[ \\mathbf {x} _ {i} ^ {k}, \\mathbf {x} _ {j} ^ {k} \\right] + 1\\right).", + "image_path": "ba11145f03e45903e5d898f4facd5be2d177e89d5f44a82fe37dba9cb8c32754.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 234, + 545, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 234, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 305, + 234, + 545, + 260 + ], + "type": "text", + "content": "We further leverage " + }, + { + "bbox": [ + 305, + 234, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\{l_i^k\\}_{i = 1}^{H_k}" + }, + { + "bbox": [ + 305, + 234, + 545, + 260 + ], + "type": "text", + "content": " as pseudo labels to build a pseudo target matrix " + }, + { + "bbox": [ + 305, + 234, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{Sim}_k^{target} \\in \\{0,1\\}^{H_k \\times H_k}" + }, + { + "bbox": [ + 305, + 234, + 545, + 260 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 266, + 531, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 266, + 531, + 316 + ], + "spans": [ + { + "bbox": [ + 318, + 266, + 531, + 316 + ], + "type": "interline_equation", + "content": "\\mathbf {S i m} _ {k} ^ {t a r g e t} [ i, j ] = \\left\\{ \\begin{array}{l} 1, \\text {i f} l _ {i} ^ {k} = l _ {j} ^ {k} \\wedge \\cos [ \\mathbf {x} _ {i} ^ {k}, \\mathbf {y} _ {l _ {i} ^ {k}} ] > \\theta \\\\ \\qquad \\wedge \\cos [ \\mathbf {x} _ {j} ^ {k}, \\mathbf {y} _ {l _ {j} ^ {k}} ] > \\theta , \\\\ 0, \\text {o t h e r w i s e}. \\end{array} \\right.", + "image_path": "6bb131bfc993f6d6ece696acdc9b51ecf426554da86649ce7191c46c5aeffbeb.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 323, + 541, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 323, + 541, + 335 + ], + "spans": [ + { + "bbox": [ + 318, + 323, + 541, + 335 + ], + "type": "text", + "content": "The similarity loss for the stage " + }, + { + "bbox": [ + 318, + 323, + 541, + 335 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 318, + 323, + 541, + 335 + ], + "type": "text", + "content": " is then formulated as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 350, + 341, + 501, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 341, + 501, + 366 + ], + "spans": [ + { + "bbox": [ + 350, + 341, + 501, + 366 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s i m} ^ {k} = \\frac {1}{H _ {k} ^ {2}} \\| \\mathbf {S i m} _ {k} - \\mathbf {S i m} _ {k} ^ {\\text {t a r g e t}} \\| _ {F} ^ {2}.", + "image_path": "f6974ec2db98693afc20edb3c969c2e6e2dab0e4f28b291c77fca7454832a270.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 372, + 407, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 407, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 407, + 384 + ], + "type": "text", + "content": "4.5. Label Generator" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 391, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 557 + ], + "type": "text", + "content": "In addressing the challenge of no pre-defined concept sets, the previous work [58] proposes to build a large vocabulary for learning during training and use WordNet [31] to correlate predictions within this vocabulary to the target concepts during inference. However, there are two limitations to the previous method. Firstly, compared with the extensive object semantics and relation predicates contained in textual descriptions, despite the large vocabulary established, it is inevitable that some classes will be overlooked. Secondly, leveraging WordNet to match vocabulary predictions to targets is not accurate and robust enough, for WordNet may only reach a coarse matching with multiple target concepts. This imprecision is particularly pronounced for relation predicates relative to object semantics." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "type": "text", + "content": "Given these limitations, we introduce a novel approach in TextPSG. Instead of approaching label prediction of objects and relations as a traditional classification problem, we reformulate it as an auto-regressive generation problem, which eliminates the necessity for pre-defined concept sets." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "Compared with a vanilla RNN, we employ a pre-trained vision language model BLIP [21] to leverage the pre-learned common sense. BLIP can take an image as input and output a caption to describe the image. In TextPSG, we borrowed the pre-trained decoder module from BLIP to perform the generation of object and relation labels." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "During training, the label generator takes the caption-parsed text graph, the segment features from the region" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2843" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "text", + "content": "grouper, and the grounding results " + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\{l_i^k\\}_{i = 1}^{H_k}" + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "text", + "content": " from the entity grouper as input. It filters out the segments with tokenwise similarity lower than the threshold " + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "text", + "content": ", merges the segments mapped to the same entity, and queries the corresponding image masks from the region grouper. Then, " + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "inline_equation", + "content": "E_{k}" + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "text", + "content": " image masks " + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{m}_i^k\\}_{i = 1}^{E_k}" + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "text", + "content": " with their pseudo entity labels " + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "inline_equation", + "content": "\\{b_i^k\\}_{i = 1}^{E_k}" + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "text", + "content": " can be obtained, where each " + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "inline_equation", + "content": "b_{i}^{k}" + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "text", + "content": " is one entity in the text graph. " + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "inline_equation", + "content": "E_{k}\\leq E" + }, + { + "bbox": [ + 47, + 71, + 289, + 180 + ], + "type": "text", + "content": " because some textual entities may not exist in the image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 181, + 287, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 287, + 242 + ], + "type": "text", + "content": "Prompt-embedding-based technique (PET). To better incorporate common sense from the vision language model, we further design a novel PET for label generation. For object prediction, the decoder takes the segment features and the image mask " + }, + { + "bbox": [ + 46, + 181, + 287, + 242 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i^k" + }, + { + "bbox": [ + 46, + 181, + 287, + 242 + ], + "type": "text", + "content": ", using a prompt" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 124, + 249, + 209, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 249, + 209, + 262 + ], + "spans": [ + { + "bbox": [ + 124, + 249, + 209, + 262 + ], + "type": "text", + "content": "a photo of [ENT]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 264, + 287, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 287, + 315 + ], + "type": "text", + "content": "to guide the object generation, where the [ENT] token is expected to be the pseudo label " + }, + { + "bbox": [ + 46, + 264, + 287, + 315 + ], + "type": "inline_equation", + "content": "b_{i}^{k}" + }, + { + "bbox": [ + 46, + 264, + 287, + 315 + ], + "type": "text", + "content": ". For relation prediction, the decoder takes the segment features and an image mask pair " + }, + { + "bbox": [ + 46, + 264, + 287, + 315 + ], + "type": "inline_equation", + "content": "(\\mathbf{m}_i^k,\\mathbf{m}_j^k)" + }, + { + "bbox": [ + 46, + 264, + 287, + 315 + ], + "type": "text", + "content": " as input, using a prompt" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 320, + 235, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 320, + 235, + 346 + ], + "spans": [ + { + "bbox": [ + 99, + 320, + 235, + 346 + ], + "type": "text", + "content": "a photo of [SUB] and [OBJ] what is their relation [REL]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "text", + "content": "to guide the relation generation, where the [SUB] and [OBJ] tokens are embedded by the pseudo labels " + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "inline_equation", + "content": "b_{i}^{k}" + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "inline_equation", + "content": "b_{j}^{k}" + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "text", + "content": ", and the [REL] token is expected to be the relation predicate between " + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "inline_equation", + "content": "(b_{i}^{k}, b_{j}^{k})" + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "inline_equation", + "content": "b_{i}^{k}" + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "text", + "content": " as subject and " + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "inline_equation", + "content": "b_{j}^{k}" + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "text", + "content": " as object in the text graph. To enhance relation generation, we further design three learnable positional embeddings " + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{sub}" + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{obj}" + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{region}" + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "text", + "content": " for indicating the different regions in the segment features. Two cross-entropy losses " + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ent}^{k}" + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rel}^{k}" + }, + { + "bbox": [ + 46, + 348, + 289, + 483 + ], + "type": "text", + "content": " are used to supervise the generation of the [ENT] and [REL] tokens, maximizing the likelihood of the target label strings, respectively. More details are in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 488, + 115, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 488, + 115, + 500 + ], + "spans": [ + { + "bbox": [ + 47, + 488, + 115, + 500 + ], + "type": "text", + "content": "4.6. Inference" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": "During inference, the target concepts of object semantics " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_o" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " and relation predicates " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_r" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " are known. With an image " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " as input and an inference stage index " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "k_{inf}" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " specified, the region grouper first partitions " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " into several candidate segments " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_i^{k_{inf}}\\}_{i=1}^{H_{k_{inf}}}" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": ", which are then passed through the segment merger to obtain the similarity matrix " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Sim}_{k_{inf}}" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": ". We formulate the segment merging as a spectral clustering problem and perform the graph cut [38] on " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Sim}_{l_{inf}}" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " for clustering. To improve the accuracy, we employ a matrix recovery method [25] to reduce the noise in " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Sim}_{l_{inf}}" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": ". In this step, the segments with similar semantics tend to be merged into the same cluster. For each cluster and each pair of clusters, the label generator uses a similar PET to generate the object semantics and the relation predicates. For every label within sets " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_o" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_r" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": ", the label generator computes its generation probability. Subsequently, these probabilities are used to rank the concepts, selecting the most probable as the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "final prediction. Note that between object and relation prediction, to convert semantic segmentation into instance segmentation, we identify each connected component in the semantic segmentation to be an instance, for simplicity. More details about inference are in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 142, + 449, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 142, + 449, + 155 + ], + "spans": [ + { + "bbox": [ + 305, + 142, + 449, + 155 + ], + "type": "text", + "content": "5. Experiments and Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 162, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 162, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 162, + 545, + 293 + ], + "type": "text", + "content": "Datasets. We train our model with the COCO Caption dataset [4], which involves 123,287 images with each labeled by 5 independent human-generated captions. Following the 2017 split, we use 118,287 images with their captions for training. We evaluate models with the Panoptic Scene Graph dataset [49] for its pixel-wise labeling as well as its high-quality object and relation annotation. We further merge the object semantics with ambiguities. After merging, 127 object semantics and 56 relation predicates are finally obtained for evaluation. More details about the datasets can be found in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 293, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 546, + 449 + ], + "type": "text", + "content": "Evaluation Protocol and Metrics. Following all previous works in scene graph generation, we evaluate the quality of a generated scene graph by viewing it as a set of subject-predicate-object triplets. We evaluate models on two tasks: Visual Phrase Detection (PhrDet) and Scene Graph Detection (SGDet). PhrDet aims to detect the whole phrase of subject-Predicate-object with a union location of subject and object. It is considered to be correct if the phrase labels are all correct and the union location matches the ground truth with the intersection over union (IoU) greater than 0.5. SGDet further requires a more accurate location, i.e., the location of subject and object should match the ground truth with IoU greater than 0.5 respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "content": "We use No-Graph-Constraint-X Recall@K (NXR@K, %) to measure the ability of generation. Recall@K computes the recall between the top-k generated triplets with the ground truth. No-Graph-Constraint-X indicates that at most X predicate labels could be predicted for each subject-object pair. Since some predicates defined in [49] are not exclusive, such as on and sitting on, NXR@K could be a more reasonable metric compared with Recall@K." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 605 + ], + "type": "text", + "content": "Baselines. We consider several baselines for Caption-to-PSG in the following experiments. Firstly, we design four baselines that strictly follow the constraints of Caption-to-PSG, where objects are located by bbox proposals generated by selective search [43]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 611, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 306, + 611, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 545, + 634 + ], + "type": "text", + "content": "- Random is the most naive baseline where all object semantics and relation predicates are randomly predicted." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 638, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 638, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 638, + 545, + 662 + ], + "type": "text", + "content": "- Prior augments Random by performing label prediction based on the statistical priors in the training set." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 665, + 547, + 714 + ], + "type": "text", + "content": "- MIL performs the alignment between proposals and textual entities by multiple instance learning [30]. Similar to [58], it formulates the object label prediction as a classification problem in a large pre-built vocabulary, with" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2844" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 547, + 206 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 547, + 206 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 547, + 206 + ], + "type": "table", + "html": "
MethodModePhrDetSGDet
ModelProposalTargetN3R50N3R100N5R50N5R100N3R50N3R100N5R50N5R100
SGGNLS-cDetector✓bbox9.6911.4510.2412.226.767.817.28.65
Randomxbbox0.020.030.020.030.010.020.020.03
PriorSelective Searchxbbox0.040.070.050.070.030.060.050.07
MILxbbox1.972.182.042.611.21.351.561.97
SGCLIPxbbox3.023.453.383.712.132.32.392.7
SGGNLS-oDetectorxbbox6.26.796.927.933.964.214.535.02
Ours-xmask8.289.169.0610.513.323.633.714.18
Ours-xbbox11.3712.7412.2414.374.294.774.825.48
", + "image_path": "9203dd379636268710f0c1a5714ae2d11246c664bb209ace8265a5eba0959924.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 207, + 546, + 229 + ], + "lines": [ + { + "bbox": [ + 47, + 207, + 546, + 229 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 546, + 229 + ], + "type": "text", + "content": "Table 1. Quantitative Comparison of Different Methods on Caption-to-PSG. 'Proposal' indicates how the method obtains bbox proposals. 'Target' indicates whether the concept sets for inference are known during training. 'Mode' indicates the mode used for evaluation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 51, + 247, + 286, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 247, + 286, + 272 + ], + "spans": [ + { + "bbox": [ + 51, + 247, + 286, + 272 + ], + "type": "text", + "content": "WordNet [31] employed during inference. The relation labels are predicted with statistical priors, similar to Prior." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 276, + 287, + 453 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 276, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 276, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 276, + 287, + 300 + ], + "type": "text", + "content": "- SGCLIP employs the pre-trained CLIP [34] to predict both object semantic labels and relation predicate labels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 303, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 303, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 51, + 303, + 287, + 339 + ], + "type": "text", + "content": "Secondly, to further benchmark the performance of our framework, we set two additional baselines based on [58] by gradually removing the constraints of Caption-to-PSG:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 287, + 403 + ], + "type": "text", + "content": "- SGGNLS-o [58] extracts proposals with a detector [35] pre-trained on OpenImage [19]. It formulates the object and relation label prediction as a classification problem within a large pre-built vocabulary, with WordNet [31] employed during inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 406, + 286, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 286, + 453 + ], + "type": "text", + "content": "- SGGNLS-c [58] uses the same proposals as SGGNLS-o. In SGGNLS-c, the target concept sets for inference are known during training. It formulates the label prediction as a classification problem in these target concept sets." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 59, + 460, + 279, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 460, + 279, + 473 + ], + "spans": [ + { + "bbox": [ + 59, + 460, + 279, + 473 + ], + "type": "text", + "content": "More design details are in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": "Implementation Details. Following GroupViT [48], we set " + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "inline_equation", + "content": "H_{1} = 64" + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "inline_equation", + "content": "H_{2} = 8" + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": " for our region grouper. We leverage general pre-trained models for weight initialization. We employ the pre-trained GroupViT for the region grouper as well as " + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{Tfm}^T" + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": " in the entity grounder, and the pre-trained BLIP [21] decoder for the label generator. During training, " + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{Tfm}^T" + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": " and the label generator are frozen. During inference, we set " + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "inline_equation", + "content": "k_{inf} = 1" + }, + { + "bbox": [ + 46, + 473, + 287, + 581 + ], + "type": "text", + "content": ". More implementation details can be found in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 587, + 224, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 224, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 224, + 601 + ], + "type": "text", + "content": "5.1. Main Results on Caption-to-PSG" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Quantitative Results. Our quantitative results on Caption-to-PSG are shown in Tab. 1. To make a fair comparison with bbox-based scene graphs generated by baselines, we evaluate our generated PSGs in both mask and bbox mode. For the latter, all masks in both prediction and ground truth are converted into bboxes (i.e., the mask area's enclosing rectangle) for evaluation, resulting in an easier setting than the former. The results show that our framework (Ours) significantly outperforms all the baselines under the same" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 247, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 344 + ], + "type": "text", + "content": "constraints on both PhrDet (14.37 vs. 3.71 N5R100) and SGDet (5.48 vs. 2.7 N5R100). Our method also shows better results compared with SGGNLS-o on all metrics and all tasks (on PhrDet, 14.37 vs. 7.93 N5R100; on SGDet, 5.48 vs. 5.02 N5R100) although SGGNLS-o utilizes location priors by leveraging a pre-trained detector. The results demonstrate that our framework is more effective for learning a good panoptic structured scene understanding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 345, + 546, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 546, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 546, + 511 + ], + "type": "text", + "content": "Qualitative Results. We provide typical qualitative results in Fig. 3 to further show our framework's effectiveness. Compared with SGGNLS-o, our framework has the following advantages. First, our framework is able to provide fine-grained semantic labels to each pixel in the image to reach a panoptic understanding, while SGGNLS-o can only provide sparse bboxes produced by the pre-trained detector. Note that categories with irregular shapes (e.g., trees in Fig. 3) are hard to be labeled precisely by bboxes. Second, compared with SGGNLS-o, our framework can generate more comprehensive object semantics and relation predicates, such as \"dry grass field\" and \"land at\" in Fig. 3, showing the open-vocabulary potential of our framework. More qualitative results are in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 520, + 451, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 520, + 451, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 520, + 451, + 533 + ], + "type": "text", + "content": "5.2. OOD Robustness Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 593 + ], + "type": "text", + "content": "We further analyze another key advantage of our framework, i.e., the robustness in OOD cases. Since SGGNLS-c and SGGNLS-o both rely on a pre-trained detector to locate objects, their performance highly depends on whether object semantics in the scene are covered by the detector." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Based on the object semantics [19] covered by the detector, we split the ground truth triplets into an in-distribution (ID) set and an OOD set. For triplets within the ID set, both the subject and object semantics are covered, while for triplets in the OOD set, at least one of the semantics is not covered. As shown in Tab. 2, both SGGNLS-c and SGGNLS-o suffer a significant performance drop from the ID set to the OOD set. On the OOD set, the triplets can hardly be retrieved. However, our framework, with the ability of location learned from purely text descriptions, can" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2845" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 143, + 137 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 143, + 137 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 143, + 137 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 143, + 137 + ], + "type": "image", + "image_path": "5fd646a0764fe88146fa3c41513539928ff5b9acebc30c769b9610d834715363.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 85, + 219, + 121 + ], + "blocks": [ + { + "bbox": [ + 156, + 85, + 219, + 121 + ], + "lines": [ + { + "bbox": [ + 156, + 85, + 219, + 121 + ], + "spans": [ + { + "bbox": [ + 156, + 85, + 219, + 121 + ], + "type": "image", + "image_path": "4e89251f4d08ca6a339cd4f66c3dfd9afd8363fa94e6a3b3217ec9e15a4ed343.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 249, + 76, + 340, + 137 + ], + "blocks": [ + { + "bbox": [ + 249, + 76, + 340, + 137 + ], + "lines": [ + { + "bbox": [ + 249, + 76, + 340, + 137 + ], + "spans": [ + { + "bbox": [ + 249, + 76, + 340, + 137 + ], + "type": "image", + "image_path": "c2c9e3f598c9d86058f817e83a877ac96010667f702a6bfb1f858c2ce578417b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 342, + 76, + 429, + 137 + ], + "blocks": [ + { + "bbox": [ + 342, + 76, + 429, + 137 + ], + "lines": [ + { + "bbox": [ + 342, + 76, + 429, + 137 + ], + "spans": [ + { + "bbox": [ + 342, + 76, + 429, + 137 + ], + "type": "image", + "image_path": "ba4b6199c1d58515ae730fc2e95c3dbd2a1f7a5d4bbec65696d9bc2150d4c5ca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 446, + 75, + 534, + 136 + ], + "blocks": [ + { + "bbox": [ + 446, + 75, + 534, + 136 + ], + "lines": [ + { + "bbox": [ + 446, + 75, + 534, + 136 + ], + "spans": [ + { + "bbox": [ + 446, + 75, + 534, + 136 + ], + "type": "image", + "image_path": "f7188de0a861be9c3a7b18d3c91db4c0e157d7d39c6f0db9974027b2b742017f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 138, + 143, + 201 + ], + "blocks": [ + { + "bbox": [ + 51, + 138, + 143, + 201 + ], + "lines": [ + { + "bbox": [ + 51, + 138, + 143, + 201 + ], + "spans": [ + { + "bbox": [ + 51, + 138, + 143, + 201 + ], + "type": "image", + "image_path": "8f5e498031f6f214f84c89768d1d5d55761c6c8ecd8b5e37c5723d4a86f5dd27.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 146, + 150, + 237, + 186 + ], + "blocks": [ + { + "bbox": [ + 146, + 150, + 237, + 186 + ], + "lines": [ + { + "bbox": [ + 146, + 150, + 237, + 186 + ], + "spans": [ + { + "bbox": [ + 146, + 150, + 237, + 186 + ], + "type": "image", + "image_path": "757c43a99d609870c03d35ef36d4c3aa5b76dfeb106ae0996af7516593eb7887.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 249, + 138, + 340, + 200 + ], + "blocks": [ + { + "bbox": [ + 249, + 138, + 340, + 200 + ], + "lines": [ + { + "bbox": [ + 249, + 138, + 340, + 200 + ], + "spans": [ + { + "bbox": [ + 249, + 138, + 340, + 200 + ], + "type": "image", + "image_path": "9946d972a85ce5a277190f051c4a2ac8d255ca12edbee4611f3eaa379867cde2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 343, + 144, + 439, + 186 + ], + "blocks": [ + { + "bbox": [ + 343, + 144, + 439, + 186 + ], + "lines": [ + { + "bbox": [ + 343, + 144, + 439, + 186 + ], + "spans": [ + { + "bbox": [ + 343, + 144, + 439, + 186 + ], + "type": "image", + "image_path": "cd097e5a5a4cff4e736bf4ca5458115ca08768dc20062efc4ac13bc4fb6fd4db.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 444, + 145, + 541, + 186 + ], + "blocks": [ + { + "bbox": [ + 444, + 145, + 541, + 186 + ], + "lines": [ + { + "bbox": [ + 444, + 145, + 541, + 186 + ], + "spans": [ + { + "bbox": [ + 444, + 145, + 541, + 186 + ], + "type": "image", + "image_path": "c3fb94a60134ede9c03df6ecf4d475950e28983c6dce3822db506f42c701ae58.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 51, + 202, + 143, + 265 + ], + "blocks": [ + { + "bbox": [ + 51, + 202, + 143, + 265 + ], + "lines": [ + { + "bbox": [ + 51, + 202, + 143, + 265 + ], + "spans": [ + { + "bbox": [ + 51, + 202, + 143, + 265 + ], + "type": "image", + "image_path": "0db44181bf9f88f3acf4aae6678debcd4f1a7e4aee7b9396ead8dc7579bf256c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 273, + 181, + 281 + ], + "lines": [ + { + "bbox": [ + 107, + 273, + 181, + 281 + ], + "spans": [ + { + "bbox": [ + 107, + 273, + 181, + 281 + ], + "type": "text", + "content": "(a) Results of SGGNLS-o" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 156, + 209, + 219, + 259 + ], + "blocks": [ + { + "bbox": [ + 156, + 209, + 219, + 259 + ], + "lines": [ + { + "bbox": [ + 156, + 209, + 219, + 259 + ], + "spans": [ + { + "bbox": [ + 156, + 209, + 219, + 259 + ], + "type": "image", + "image_path": "3fcd7be67b3d6b4fefa7c213cb4887cc98865565b0092549333f3582d046ec37.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 282, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 282, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 545, + 326 + ], + "type": "text", + "content": "Figure 3. Qualitative Comparison between SGGNLS-o (a) and Ours (b). For each method, the results of object location are shown on the left, while the results of scene graph generation are shown on the right. For Ours, scene graphs predicted within the given concept sets are provided in the middle column, and scene graphs directly predicted through the auto-regressive generation (i.e., an open-vocabulary manner) in the label generator are additionally provided in the right column." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 249, + 202, + 340, + 264 + ], + "blocks": [ + { + "bbox": [ + 249, + 202, + 340, + 264 + ], + "lines": [ + { + "bbox": [ + 249, + 202, + 340, + 264 + ], + "spans": [ + { + "bbox": [ + 249, + 202, + 340, + 264 + ], + "type": "image", + "image_path": "081cd78f56524175490a5712006eec3d1bb73aca015a857908e90d5bc81638b6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 351, + 198, + 439, + 266 + ], + "blocks": [ + { + "bbox": [ + 351, + 198, + 439, + 266 + ], + "lines": [ + { + "bbox": [ + 351, + 198, + 439, + 266 + ], + "spans": [ + { + "bbox": [ + 351, + 198, + 439, + 266 + ], + "type": "image", + "image_path": "76979ea156d605c127316bcd9c315ff7e2a2bab5f7a79d86baaf2b89d4bc0158.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 274, + 422, + 281 + ], + "lines": [ + { + "bbox": [ + 365, + 274, + 422, + 281 + ], + "spans": [ + { + "bbox": [ + 365, + 274, + 422, + 281 + ], + "type": "text", + "content": "(b) Results of Ours" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 455, + 197, + 543, + 265 + ], + "blocks": [ + { + "bbox": [ + 455, + 197, + 543, + 265 + ], + "lines": [ + { + "bbox": [ + 455, + 197, + 543, + 265 + ], + "spans": [ + { + "bbox": [ + 455, + 197, + 543, + 265 + ], + "type": "image", + "image_path": "6358bff9e31c151ed9f639abb8b6f7d260924235a5251b491291fd88401901d9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 53, + 335, + 286, + 435 + ], + "blocks": [ + { + "bbox": [ + 53, + 335, + 286, + 435 + ], + "lines": [ + { + "bbox": [ + 53, + 335, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 286, + 435 + ], + "type": "table", + "html": "
SetModelTargetModePhrDetSGDet
N3R100N5R100N3R100N5R100
IDSGGNLS-c✓bbox16.7618.4810.4511.86
SGGNLS-oXbbox11.5513.647.138.47
OursXmask9.2710.453.283.76
OursXbbox13.3514.824.635.36
OODSGGNLS-c✓bbox0000
SGGNLS-oXbbox0.050.0600
OursXmask8.479.764.074.51
OursXbbox10.1811.695.235.72
", + "image_path": "f064641e11259d5b630acd13a1f96da324906685465a0194940b8ef9e7960080.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 52, + 468, + 286, + 525 + ], + "blocks": [ + { + "bbox": [ + 47, + 436, + 287, + 456 + ], + "lines": [ + { + "bbox": [ + 47, + 436, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 287, + 456 + ], + "type": "text", + "content": "Table 2. Analysis on OOD Robustness. 'Set' indicates the triplet set used for evaluation." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 468, + 286, + 525 + ], + "lines": [ + { + "bbox": [ + 52, + 468, + 286, + 525 + ], + "spans": [ + { + "bbox": [ + 52, + 468, + 286, + 525 + ], + "type": "table", + "html": "
Stage#SegCutPhrDetSGDet
N3R100N5R100N3R100N5R100
164×10.7311.393.183.51
164✓12.7414.374.775.48
28×9.2411.033.534.35
28✓6.788.452.463.21
", + "image_path": "348ce7a6ea8ed39b82c146262caa4072f4848abb77315441b0c974afd6084865.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 575, + 286, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 286, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 286, + 599 + ], + "type": "text", + "content": "reach similar performance on both sets, which demonstrates the OOD robustness of our framework for PSG generation." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 610, + 147, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 147, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 147, + 621 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 712 + ], + "type": "text", + "content": "We conduct additional ablation studies to evaluate the effectiveness of our design choices. For all following experiments, we report N3R100 and N5R100 evaluated in bbox mode for simplicity. We answer the following questions. Q1: Does the explicit learning of merging in the segment merger helps provide better image segments? Q2: Is the generation-based label prediction better than the" + } + ] + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 308, + 335, + 545, + 392 + ], + "blocks": [ + { + "bbox": [ + 47, + 525, + 286, + 568 + ], + "lines": [ + { + "bbox": [ + 47, + 525, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 286, + 568 + ], + "type": "text", + "content": "Table 3. Ablation Study on the Segment Merger. 'Stage' indicates the grouping stage where image segments used for merging are from. " + }, + { + "bbox": [ + 47, + 525, + 286, + 568 + ], + "type": "inline_equation", + "content": "\\# \\mathrm{Seg}" + }, + { + "bbox": [ + 47, + 525, + 286, + 568 + ], + "type": "text", + "content": " indicates the number of image segments. Cut indicates whether the graph-cut-based segment merging is applied." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 335, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 308, + 335, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 335, + 545, + 392 + ], + "type": "table", + "html": "
Label PredictionModelPhrDetSGDet
N3R100N5R100N3R100N5R100
Cls + WordNet-8.829.362.362.72
GenRNN9.1210.442.653.07
Gen w/o PETBLIP [21]2.332.580.450.6
Gen w/ PETBLIP [21]12.6414.284.775.49
", + "image_path": "a4b7e0dc9db5d965a6d0fb1d610f4b73e68b2dbcf2bfa4a54eb884ecb24a4aaf.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 392, + 545, + 413 + ], + "lines": [ + { + "bbox": [ + 306, + 392, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 545, + 413 + ], + "type": "text", + "content": "Table 4. Ablation Study on the Label Generator. 'Cls' indicates classification. 'Gen' indicates generation." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 422, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 470 + ], + "type": "text", + "content": "classification-based prediction? Q3: Does the pre-learned common sense from the pre-trained BLIP [21] helps with the label prediction? Q4: Does the PET helps incorporate the pre-learned common sense for label prediction?" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 472, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 591 + ], + "type": "text", + "content": "In Tab. 3, we compare different strategies of image segment merging during inference. Row 1&2 denote that the " + }, + { + "bbox": [ + 304, + 472, + 545, + 591 + ], + "type": "inline_equation", + "content": "H_{1} = 64" + }, + { + "bbox": [ + 304, + 472, + 545, + 591 + ], + "type": "text", + "content": " segments from the first grouping stage are used for further merging, while row 3&4 denote that the " + }, + { + "bbox": [ + 304, + 472, + 545, + 591 + ], + "type": "inline_equation", + "content": "H_{2} = 8" + }, + { + "bbox": [ + 304, + 472, + 545, + 591 + ], + "type": "text", + "content": " segments from the second stage are used. The results show that applying the graph cut to merge the segments from the first stage could reach the best performance, corroborating that compared with the fully implicit learning of merging, the explicit learning of merging can provide better segments (row 2 vs 3, answering Q1)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "In Tab. 4, we compare different designs of the label generator. Keeping the other modules the same, we change the label generator (row 4) into three different designs, i.e., classification within a large pre-built vocabulary followed by WordNet [31] for target matching (row 1), generation with a vanilla RNN (row 2), generation with the BLIP decoder but without the PET (row 3). The results show that with the constraint of no pre-defined concept sets, compared with formulating the label prediction into a classification problem, formulating it into a generation problem is a bet-" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2846" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "ter choice (row 1 vs 2&4, answering Q2). By employing the pre-trained BLIP for leveraging the pre-learned common sense, the prediction could be further boosted (row 2 vs 4, answering Q3). And the PET is very important for incorporating the common sense from the pre-trained model (row 3 vs 4, answering Q4)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 144, + 287, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 144, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 287, + 169 + ], + "type": "text", + "content": "More ablation studies for the design evaluation can be found in the supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 176, + 167, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 176, + 167, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 167, + 190 + ], + "type": "text", + "content": "5.4. Application on TSSS" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 196, + 149, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 196, + 149, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 149, + 278 + ], + "type": "text", + "content": "As a side product, we observe that our entity grounder and segment merger can also enhance TSSS. Based on the original GroupViT [48], we replace the multi" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 173, + 205, + 282, + 253 + ], + "blocks": [ + { + "bbox": [ + 173, + 205, + 282, + 253 + ], + "lines": [ + { + "bbox": [ + 173, + 205, + 282, + 253 + ], + "spans": [ + { + "bbox": [ + 173, + 205, + 282, + 253 + ], + "type": "table", + "html": "
MethodmIoU
GroupViT [48]24.28
GroupViT† [48]24.72
Ours26.87
", + "image_path": "ee0a269e928c9212d8d1a757fc9dac593efa6df6604ea30ed35bb3e51722ec61.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 167, + 254, + 286, + 274 + ], + "lines": [ + { + "bbox": [ + 167, + 254, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 167, + 254, + 286, + 274 + ], + "type": "text", + "content": "Table 5. Results on TSSS. † indicates finetuned." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 279, + 287, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 375 + ], + "type": "text", + "content": "label contrastive loss with our entity grounder and segment merger. Then we finetune the model on the COCO Caption dataset [4]. As shown in Tab. 5, compared with GroupViT directly finetuned on [4], the explicit learning of merging in our modules can boost the model with an absolute " + }, + { + "bbox": [ + 46, + 279, + 287, + 375 + ], + "type": "inline_equation", + "content": "2.15\\%" + }, + { + "bbox": [ + 46, + 279, + 287, + 375 + ], + "type": "text", + "content": " improvement of mean Intersection over Union (mIoU, " + }, + { + "bbox": [ + 46, + 279, + 287, + 375 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 46, + 279, + 287, + 375 + ], + "type": "text", + "content": ") on COCO [23], which demonstrates the effectiveness of our proposed modules on better object location." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 383, + 119, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 119, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 119, + 395 + ], + "type": "text", + "content": "5.5. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 402, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 653 + ], + "type": "text", + "content": "Failure Case Analysis. Despite the impressive performance of TextPSG, there are still challenges to address. Upon analyzing the failure cases for PSG generation, we identify three specific limitations of TextPSG that contribute to these failures. a) The strategy we use to convert semantic segmentation into instance segmentation is not entirely effective. For simplicity, in TextPSG, we identify each connected component in the semantic segmentation to be an individual object instance. However, this strategy may fail when instances overlap or are occluded, resulting in either an underestimation or an overestimation of instances. b) Our framework faces difficulty in locating small objects in the scene due to limitations in resolution and the grouping strategy for location. c) The relation prediction of our framework requires enhancement, as it is not adequately conditioned on the image. While the label generator uses both image features and predicted object semantics to determine the relation, it sometimes seems to lean heavily on the object semantics, potentially neglecting the actual image content. Examples of failure cases for each of these limitations can be found in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "Model Diagnosis. For a clearer understanding of the efficacy of our framework, we conduct a model diagnosis to answer the following question: why does our framework only achieve semantic segmentation through learning, rather than panoptic segmentation (and thus requires further" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 544, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 544, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 544, + 84 + ], + "type": "text", + "content": "segmentation conversion to obtain instance segmentation)?" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 95, + 385, + 171 + ], + "blocks": [ + { + "bbox": [ + 316, + 95, + 385, + 171 + ], + "lines": [ + { + "bbox": [ + 316, + 95, + 385, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 385, + 171 + ], + "type": "image", + "image_path": "c626cb00f1fd940999acb0183a3d1210f60281923e2db9d49a498916ba1926eb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 171, + 428, + 185 + ], + "lines": [ + { + "bbox": [ + 339, + 171, + 428, + 185 + ], + "spans": [ + { + "bbox": [ + 339, + 171, + 428, + 185 + ], + "type": "text", + "content": "(a) A white sheep and a black sheep standing on the grass under the sky." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 187, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 305, + 187, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 187, + 545, + 243 + ], + "type": "text", + "content": "Figure 4. Region-Entity Alignment Results of Captions in Different Granularity. Two captions in different granularity are used to execute region-entity alignment with the same image, with (a) one describing the two sheep individually while (b) the other merges them in plural form." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 387, + 95, + 457, + 171 + ], + "blocks": [ + { + "bbox": [ + 387, + 95, + 457, + 171 + ], + "lines": [ + { + "bbox": [ + 387, + 95, + 457, + 171 + ], + "spans": [ + { + "bbox": [ + 387, + 95, + 457, + 171 + ], + "type": "image", + "image_path": "ddc720bb97a834e199c54994b057015092868968038267bc7df10d73c4c98e9b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 461, + 95, + 533, + 171 + ], + "blocks": [ + { + "bbox": [ + 461, + 95, + 533, + 171 + ], + "lines": [ + { + "bbox": [ + 461, + 95, + 533, + 171 + ], + "spans": [ + { + "bbox": [ + 461, + 95, + 533, + 171 + ], + "type": "image", + "image_path": "8feeeffbb255dd913d4f772a2159f22072d1a7bb396920dba6d87e54cb7da40c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 171, + 535, + 185 + ], + "lines": [ + { + "bbox": [ + 461, + 171, + 535, + 185 + ], + "spans": [ + { + "bbox": [ + 461, + 171, + 535, + 185 + ], + "type": "text", + "content": "(b) Pair of sheep standing on open grassy field on clear day." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "text", + "content": "In Fig. 4, we use two captions in different granularity to execute region-entity alignment. It shows that our framework has the capability to assign distinct masks to individual instances. However, the nature of caption data, where captions often merge objects of the same semantics in plural form, limits our framework from differentiating instances. It is the weak supervision provided by the caption data that constrains our framework." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 351, + 522, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 351, + 522, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 351, + 522, + 363 + ], + "type": "text", + "content": "More diagnoses are in the supplementary material." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "text", + "content": "Future Directions. In response to the limitations discussed, we outline several potential directions for enhancing our framework: a) a refined and sophisticated strategy for segmentation conversion; b) increasing the input resolution, though this may introduce greater computational demands; c) a more suitable image-conditioned reasoning mechanism for relation prediction; d) a superior image-caption-pair dataset with more detailed granularity in captions to achieve panoptic segmentation through learning." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 493, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 649 + ], + "type": "text", + "content": "We take the first step towards the novel problem Caption-to-PSG, aiming to learn PSG generation purely from language. To tackle this challenging problem, we propose a new modularized framework TextPSG with several novel techniques, which significantly outperforms the baselines and achieves strong OOD robustness. This paves the path to a more comprehensive and generalizable panoptic structured scene understanding. There are still bottle-necks in TextPSG to be explored in future work, including a) a more sophisticated strategy for segmentation conversion; b) a more suitable image-conditioned reasoning mechanism for relation prediction; c) a superior image-caption-pair dataset for panoptic segmentation through learning." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the DSO grant DSOCO21072, IBM, and gift funding from MERL, Cisco, and Amazon. We would also like to thank the computation support from AiMOS, a server cluster for the IBM Research AI Hardware Center." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2847" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Somak Aditya, Yezhou Yang, Chitta Baral, Yiannis Aloimonos, and Cornelia Fermüller. Image understanding using vision and reasoning through scene description graph. Computer Vision and Image Understanding, 2017. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 180 + ], + "type": "text", + "content": "[2] S. Amiri, Kishan Chandan, and Shiqi Zhang. Reasoning with scene graphs for robot planning under partial observability. IEEE Robotics and Automation Letters, 7:5560-5567, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 227 + ], + "type": "text", + "content": "[3] Gabor Angeli, Melvin Johnson, and Christopher D. Manning. Leveraging linguistic structure for open domain information extraction. In Annual Meeting of the Association for Computational Linguistics, 2015. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 271 + ], + "type": "text", + "content": "[4] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 274, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 287, + 317 + ], + "type": "text", + "content": "[5] Zhenfang Chen, Lin Ma, Wenhan Luo, and Kwan-Yee K Wong. Weakly-supervised spatio-temporally grounding natural sentence in video. In Proc. 57th Annual Meeting of the Association for Computational Linguistics, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 287, + 373 + ], + "type": "text", + "content": "[6] Zhenfang Chen, Jiayuan Mao, Jiajun Wu, Kwan-Yee K Wong, Joshua B. Tenenbaum, and Chuang Gan. Grounding physical concepts of objects and events through dynamic visual reasoning. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 376, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 287, + 430 + ], + "type": "text", + "content": "[7] Zhenfang Chen, Peng Wang, Lin Ma, Kwan-Yee K Wong, and Qi Wu. Cops-ref: A new dataset and task on compositional referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10086-10095, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 433, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 433, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 433, + 287, + 475 + ], + "type": "text", + "content": "[8] Helisa Dhamo, Azade Farshad, Iro Laina, Nassir Navab, Gregory D. Hager, Federico Tombari, and Christian Rupprecht. Semantic image manipulation using scene graphs. In CVPR, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 478, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 287, + 532 + ], + "type": "text", + "content": "[9] Xiaoyi Dong, Yinglin Zheng, Jianmin Bao, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. arXiv preprint arXiv:2208.12262, 2022.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 534, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 588 + ], + "type": "text", + "content": "[10] Samir Yitzhak Gadre, Kiana Ehsani, Shuran Song, and Roozbeh Mottaghi. Continuous scene representations for embodied ai. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14829-14839, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 655 + ], + "type": "text", + "content": "[11] Chuang Gan, Yandong Li, Haoxiang Li, Chen Sun, and Boqing Gong. Vqs: Linking segmentations to questions and answers for supervised attention in vqa and question-focused semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1811-1820, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Jiumiang Gu, Handong Zhao, Zhe Lin, Sheng Li, Jianfei Cai, and Mingyang Ling. Scene graph generation with external knowledge and image reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 149 + ], + "type": "text", + "content": "[14] Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image generation from scene graphs. In CVPR, 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 151, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 547, + 194 + ], + "type": "text", + "content": "[15] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3128-3137, 2015. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 248 + ], + "type": "text", + "content": "[16] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 545, + 293 + ], + "type": "text", + "content": "[17] Siddhesh Khandelwal, Mohammed Suhail, and Leonid Sigal. Segmentation-grounded scene graph generation. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15859-15869, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 294, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 360 + ], + "type": "text", + "content": "[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Fei-Fei Li. Visual Genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32-73, 2016. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 361, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 425 + ], + "type": "text", + "content": "[19] Alina Kuznetsova, Hassan Rom, Neil Gordon Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4. International Journal of Computer Vision, 128:1956-1981, 2018. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 459 + ], + "type": "text", + "content": "[20] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. arXiv preprint arXiv:2201.03546, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "text", + "content": "[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 5, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 557 + ], + "type": "text", + "content": "[22] Xingchen Li, Long Chen, Wenbo Ma, Yi Yang, and Jun Xiao. Integrating object-aware and interaction-aware knowledge for weakly supervised scene graph generation. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 559, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 612 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 9" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 614, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 667 + ], + "type": "text", + "content": "[24] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[25] Guangcan Liu, Zhouchen Lin, Shuicheng Yan, Ju Sun, Yong Yu, and Yi Ma. Robust recovery of subspace structures by low-rank representation. IEEE transactions on pattern analysis and machine intelligence, 35(1):171-184, 2012. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2848" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7086-7096, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "text", + "content": "[27] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. arXiv preprint arXiv:2211.14813, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 162, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 162, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 287, + 216 + ], + "type": "text", + "content": "[28] Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David McClosky. The Stanford CoreNLP natural language processing toolkit. In Association for Computational Linguistics (ACL) System Demonstrations, pages 55–60, 2014. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 217, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 287, + 271 + ], + "type": "text", + "content": "[29] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 272, + 287, + 304 + ], + "type": "text", + "content": "[30] Oded Maron and Tomás Lozano-Pérez. A framework for multiple-instance learning. Advances in neural information processing systems, 10, 1997. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 305, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 305, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 49, + 305, + 287, + 348 + ], + "type": "text", + "content": "[31] George A. Miller. WordNet: A lexical database for English. In Speech and Natural Language: Proceedings of a Workshop Held at Harriman, New York, February 23-26, 1992, 1992. 5, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 287, + 392 + ], + "type": "text", + "content": "[32] Julia Peyre, Ivan Laptev, Cordelia Schmid, and Josef Sivic. Weakly-supervised learning of visual relations. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5189-5198, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 393, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 393, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 49, + 393, + 287, + 458 + ], + "type": "text", + "content": "[33] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 460, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 460, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 49, + 460, + 287, + 524 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763, 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 526, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 526, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 49, + 526, + 287, + 569 + ], + "type": "text", + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2, 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 571, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 571, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 49, + 571, + 287, + 635 + ], + "type": "text", + "content": "[36] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 817-834. Springer, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 49, + 636, + 287, + 679 + ], + "type": "text", + "content": "[37] Sebastian Schuster, Ranjay Krishna, Angel X. Chang, Li Fei-Fei, and Christopher D. Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In VL@EMNLP, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "text", + "content": "[38] Jianbo Shi and J. Malik. Normalized cuts and image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[39] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8376-8384, 2019. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[40] Jing Shi, Yiwu Zhong, Ning Xu, Yin Li, and Chenliang Xu. A simple baseline for weakly-supervised scene graph generation. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16373-16382, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[41] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "text", + "content": "[42] Damien Teney, Lingqiao Liu, and Anton van den Hengel. Graph-structured representations for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 264, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 307 + ], + "type": "text", + "content": "[43] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104:154-171, 2013. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "type": "text", + "content": "[44] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 343, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 375 + ], + "type": "text", + "content": "[45] Qixun Wang, Xiaofeng Guo, and Haofan Wang. 1st place solution for psg competition with eccv'22 sensehuman workshop. ArXiv, abs/2302.02651, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 377, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 442 + ], + "type": "text", + "content": "[46] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6602–6611, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "type": "text", + "content": "[47] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5419, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 490, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 554 + ], + "type": "text", + "content": "[48] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. GroupViT: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18134-18144, June 2022. 2, 3, 4, 5, 7, 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 556, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 588 + ], + "type": "text", + "content": "[49] Jingkang Yang, Yi Zhe Ang, Zujin Guo, Kaiyang Zhou, Wayne Zhang, and Ziwei Liu. Panoptic scene graph generation. In ECCV, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "type": "text", + "content": "[50] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In Proceedings of the European Conference on Computer Vision (ECCV), September 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[51] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. FILIP: Fine-grained interactive language-image pre-training. In International Conference on Learning Representations, 2022. 2, 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "text", + "content": "[52] Keren Ye and Adriana Kovashka. Linguistic structures as weak supervision for visual scene graph generation. In Pro-" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2849" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 462 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[53] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1307-1315, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 217 + ], + "type": "text", + "content": "[54] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "text", + "content": "[55] Alireza Zareian, Svebor Karaman, and Shih-Fu Chang. Weakly supervised visual semantic parsing. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3733-3742, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 317 + ], + "type": "text", + "content": "[56] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "type": "text", + "content": "[57] Hanwang Zhang, Zawlin Kyaw, Jinyang Yu, and Shih-Fu Chang. Ppr-fcn: Weakly supervised visual relation detection via parallel pairwise r-fcn. In Proceedings of the IEEE international conference on computer vision, pages 4233-4241, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 375, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 375, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 287, + 407 + ], + "type": "text", + "content": "[58] Yiwu Zhong, Jing Shi, Jianwei Yang, Chenliang Xu, and Yin Li. Learning to generate scene graph from natural language supervision. In ICCV, 2021. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 462 + ], + "type": "text", + "content": "[59] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVIII, pages 696-712. Springer, 2022. 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2850" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/295801ea-f88d-446b-aeea-24167c36963d_content_list.json b/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/295801ea-f88d-446b-aeea-24167c36963d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0ed756e4ebea865024684288bd2d97f02e40ae75 --- /dev/null +++ b/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/295801ea-f88d-446b-aeea-24167c36963d_content_list.json @@ -0,0 +1,1800 @@ +[ + { + "type": "text", + "text": "Texture Generation on 3D Meshes with Point-UV Diffusion", + "text_level": 1, + "bbox": [ + 187, + 130, + 782, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Yu", + "bbox": [ + 109, + 179, + 176, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peng Dai", + "bbox": [ + 217, + 180, + 300, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenbo Li2", + "bbox": [ + 339, + 180, + 428, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lan Ma3", + "bbox": [ + 467, + 180, + 540, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhengzhe Liu $^{2\\dagger}$", + "bbox": [ + 578, + 180, + 705, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaojuan Qi $^{1\\dagger}$", + "bbox": [ + 745, + 180, + 859, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Hong Kong, 2The Chinese University of Hong Kong, 3TCL Corporate Research", + "bbox": [ + 86, + 215, + 883, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yuxin, daipeng, xjqi}@eee.hku.hk", + "bbox": [ + 114, + 236, + 388, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wenboli,zzliu}@cse.cuhk.edu.hk", + "bbox": [ + 423, + 236, + 696, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rubyma@tcl.com", + "bbox": [ + 730, + 237, + 857, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b22925ff733e95162010d693f891239383078d8f8361fa87400e6d9beb215271.jpg", + "image_caption": [ + "Figure 1. A gallery of generated results by our Point-UV diffusion. Our method is capable of processing meshes of any genus, generating diversified, geometry-compatible, and high-fidelity textures." + ], + "image_footnote": [], + "bbox": [ + 89, + 263, + 885, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 547, + 313, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we focus on synthesizing high-quality textures on 3D meshes. We present Point-UV diffusion, a coarse-to-fine pipeline that marries the denoising diffusion model with UV mapping to generate 3D consistent and high-quality texture images in UV space. We start with introducing a point diffusion model to synthesize low-frequency texture components with our tailored style guidance to tackle the biased color distribution. The derived coarse texture offers global consistency and serves as a condition for the subsequent UV diffusion stage, aiding in regularizing the model to generate a 3D consistent UV texture image. Then, a UV diffusion model with hybrid conditions is developed to enhance the texture fidelity in the 2D UV space. Our method can process meshes of any genus, generating diversified, geometry-compatible, and high-fidelity textures. Code is available at https://cvmilab.github.io/Point-UV-Diffusion.", + "bbox": [ + 73, + 573, + 473, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 547, + 630, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Texturing 3D meshes is a fundamental task in computer vision and graphics. It enhances the visual richness of 3D objects, thereby facilitating their application in various fields such as video games, 3D movies, and AR/VR technologies. However, generating high-quality textures can be daunting and time-consuming, often requiring specialized knowledge and resources. As such, there is a pressing need for an efficient approach to automatically create high-quality textures on 3D meshes.", + "bbox": [ + 496, + 578, + 893, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the substantial progress the community has made in 2D image synthesis and 3D shape generation using GANs [13, 20, 44, 11] or diffusion models [34, 36, 35, 19], crafting realistic textures on mesh surfaces remains challenging. One major difficulty stems from the need for suitable 3D representations for texture synthesis. Early approaches investigate the use of voxels [5, 45, 6] or point clouds [10] and synthesize point/voxel colors. However, they can only afford to synthesize low-resolution results with low-fidelity textures due to memory and model complexity constraints. In response, Texture Fields [29] adopts an implicit representation with the potential to synthesize", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†: Corresponding authors", + "bbox": [ + 101, + 886, + 240, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4206", + "bbox": [ + 482, + 924, + 516, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "high-resolution textures, but still could not yield satisfactory results as shown in Figure 2 (a): over-smoothed results. Most recently, Siddiqui et al. [38] propose to parameterize the shape as tetrahedral meshes and introduce tetrahedral mesh convolution to enhance local details. Albeit improving results, tetrahedral parameterization inevitably destroys geometric details of the input mesh and thus cannot faithfully preserve the original structure. As shown in Figure 2 (b), the delicate structures of the chair's back are absent. Moreover, the generative models utilized in these methods are limited to GANs [29, 38] and VAEs [29, 12]. The more advanced diffusion model, which could potentially open up new avenues for high-quality texture generation, remains insufficiently explored.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we delve into a novel texture representation based on UV maps and investigate the advanced diffusion model for texture generation. The 2D nature of the UV map enables it to circumvent the cost of high-resolution point/voxel representations. Besides, the UV map is compatible with arbitrary mesh topologies, thereby preserving the original geometric structures. However, while promising, direct integration of the UV map representation with a 2D diffusion model presents challenges in synthesizing seamless textures, leading to severe artifacts, as shown in Figure 2 (c). This occurs because the UV mapping process fragments the continuous texture on the 3D surface into isolated patches on the 2D UV plane (see Figure 3).", + "bbox": [ + 75, + 305, + 468, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we introduce Point-UV diffusion, a two-stage coarse-to-fine framework consisting of point diffusion and UV diffusion. Specifically, we initially design a point diffusion model to generate color for sampled points that act as low-frequency texture components. This model is equipped with a style guidance mechanism that alleviates the impact of biased color distributions in the dataset and facilitates diversity during inference. Next, we project these colorized points onto the 2D UV space with 3D coordinate interpolation, thereby generating a coarse texture image that maintains 3D consistency and continuity. Given the coarse textured image, we develop a UV diffusion model with elaborately designed hybrid conditions to improve the quality of the textures (see Figure 2 (ours) and Figure 1).", + "bbox": [ + 75, + 505, + 468, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In short, our contributions are as follows: 1) We propose a new framework for texture generation for given meshes. Our representation can handle meshes with arbitrary topology and is able to faithfully preserve geometric structures. 2) To the best of our knowledge, we are the first to train a diffusion model specifically for mesh texture generation. Our coarse-to-fine framework allows us to enjoy the efficiency of 2D representation while enhancing 3D consistency. 3) We compare our approach with multiple methods in unconditional generation and achieve state-of-the-art results. Furthermore, we demonstrate that our method can be easily extended to scenarios with text-conditioning and", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9892278130303e08163bccc166ec305808fbec62493a36a1e330353c20108282.jpg", + "image_caption": [ + "(a) Texture Fields vs Ours" + ], + "image_footnote": [], + "bbox": [ + 506, + 93, + 653, + 261 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c61bc7531d7ce27c511cb3be73337055d95bf097cd97b0bf8611eed49c31584c.jpg", + "image_caption": [ + "(b) Texturify vs Ours" + ], + "image_footnote": [], + "bbox": [ + 656, + 93, + 774, + 261 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/55a88ce9b2911b3d4ea4a7da214f77dd9b1cc1e98ed13e135ca289646d74c39d.jpg", + "image_caption": [ + "(c) 2D diffusion vs Ours" + ], + "image_footnote": [], + "bbox": [ + 759, + 93, + 890, + 261 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30221e74c987260583731d2865db7cc1d63613dff5bb7263268b952ee83092c3.jpg", + "image_caption": [ + "Figure 2. Comparisons with different methods. Our generative results (a) possess high-quality details, (b) faithfully preserve the mesh structure, and (c) are better consistent with the given shape, compared with Texture Fields [29], Texturify [38] and 2D diffusion [16], respectively.", + "Textured Mesh", + "Figure 3. Illustration of UV mapping process. It establishes connections between the 2D texture map and the surface appearance of 3D shape." + ], + "image_footnote": [], + "bbox": [ + 524, + 369, + 730, + 494 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8da5aa7d15164e4c4214b2694020aeb680d8c4b36c6a7df4785e57607da9f6ee.jpg", + "image_caption": [ + "Texture Image" + ], + "image_footnote": [], + "bbox": [ + 735, + 369, + 892, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image-conditioning.", + "bbox": [ + 500, + 585, + 635, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 617, + 640, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Texture generation. Early works [41, 43, 6] propose using voxels to represent colors. However, due to a cubic increase in memory usage and computational cost with rising resolution, these methods produce only coarse, low-resolution textures. For meshes, recent work in texture generation [38] predicts the color for each face, but the representation capacity remains largely constrained by the mesh's resolution. Another set of works [30, 42, 26, 12] leverage UV mapping, but they require either spherical parameterization or part segmentation and, therefore, are confined to handling low-genus shapes. Other works [32, 14] rely on image exemplars, while our approach focuses on unconditional generation. Recently, implicit functions attract increasing attention for generative tasks [11, 31, 39, 3, 9, 37, 29, 22]. Most of this research centers on generating 3D-aware images [39, 3, 9, 37], rather than synthesizing textures for given 3D meshes. Texture Fields [29] represents", + "bbox": [ + 496, + 643, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4207", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the most similar work to our task, which, however, tends to produce over-smoothed results.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some methods focus on test-time optimization for generation tasks. For instance, Text2Mesh [24] employs CLIP [33] to design a loss function. This function stylizes a 3D mesh to align with a target text prompt, predicting both the color and displacement for each mesh vertex. More recently, DreamFusion [31] and Magic3D [21] leverage pretrained stable diffusion [35] for score distillation, creating either NeRF [25] or a 3D mesh. However, these approaches often demand extensive optimization time or necessitate alterations to the existing mesh geometry [24]. In this paper, our objective is to generate textures for specified full-3D meshes with arbitrary topology. This goal entails two fundamental requirements: 1) preserving the original mesh's geometric structure; 2) producing a 3D texture representation exportable as either vertex color or a uv-texture-image, instead of merely generating-rendering multi-view or 3D-aware images.", + "bbox": [ + 75, + 121, + 472, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models for 3D generation. In addition to 2D image generation, diffusion models recently gain significant attention in 3D generation, leading to many related works [49, 18, 19, 28, 31, 47, 1]. For instance, Zhou et al. [49] introduce point-voxel diffusion for point cloud generation and completion. Hui et al. [19] and Hu et al. [18] suggest a compact wavelet domain representation for shapes, enabling higher-quality shape creation via diffusion models. Nichol et al. [28] present Point-E, a system that uses a text-conditional strategy to produce colored 3D point clouds. This approach enables an efficient synthesis of intricate 3D shapes from textual prompts. Yet, this system yields lower-resolution point clouds, often missing detailed shapes and textures. In this work, we develop a brand new 3D diffusion model for texture image synthesis, which allows high-fidelity and 3D-consistent texture generation.", + "bbox": [ + 75, + 397, + 472, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 652, + 217, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Denoising diffusion model.", + "text_level": 1, + "bbox": [ + 76, + 676, + 316, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion model [40, 16] is a kind of likelihood-based generative model that has gained significant attention recently. It learns the data distribution $q(x_0)$ by progressive denoising from a prior Gaussian distribution. Given a sample from the data distribution $x_0 \\sim q(x_0)$ , a fixed forward process $q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0) = \\prod_{t=1}^{T} q(\\mathbf{x}_t \\mid \\mathbf{x}_{t-1})$ is used to perturb the data with Gaussian kernels $q(\\mathbf{x}_t \\mid \\mathbf{x}_{t-1}) \\coloneqq \\mathcal{N}\\left(\\sqrt{1 - \\beta_t} \\mathbf{x}_{t-1}, \\beta_t \\mathbf{I}\\right)$ , producing increasingly noisy latent variables $\\{x_1, x_2, \\ldots, x_T\\}$ . Then, a parameterized Markov process $p_\\theta(\\mathbf{x}_{0:T}) = p(\\mathbf{x}_T) \\prod_{t=1}^{T} p_\\theta(\\mathbf{x}_{t-1} \\mid \\mathbf{x}_t)$ with transition kernel $p_\\theta(\\mathbf{x}_{t-1} \\mid \\mathbf{x}_t) \\coloneqq \\mathcal{N}\\left(\\mu_\\theta(\\mathbf{x}_t, t), \\sigma_t^2 \\mathbf{I}\\right)$ is optimized through maximizing a variational lower bound of log data likelihood, which essentially targets to match the joint", + "bbox": [ + 75, + 700, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "distribution $q(\\pmb{x}_{0:T})$ :", + "bbox": [ + 500, + 90, + 645, + 107 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E} _ {q \\left(\\mathbf {x} _ {0}\\right)} \\left[ \\log p _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\right] \\geq \\operatorname {E} _ {q \\left(\\mathbf {x} _ {0: T}\\right)} \\left[ \\log \\frac {p _ {\\theta} \\left(\\mathbf {x} _ {0 : T}\\right)}{q \\left(\\mathbf {x} _ {1 : T} \\mid \\mathbf {x} _ {0}\\right)} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 116, + 893, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After training, novel samples can then be generated via iterative sampling from $p_{\\theta}(\\boldsymbol{x}_{t-1} \\mid \\boldsymbol{x}_t)$ following:", + "bbox": [ + 498, + 160, + 892, + 191 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t - 1} = \\mu_ {\\theta} (\\mathbf {x} _ {t}, t) + \\sqrt {\\beta_ {t}} \\mathbf {z}, \\mathbf {z} \\sim \\mathcal {N} (0, \\mathbf {I}).\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 200, + 834, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models have shown impressive capabilities in generating high-quality and diverse content. This paper proposes a texture generation framework based on diffusion model.", + "bbox": [ + 498, + 231, + 893, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. UV mapping and challenges.", + "text_level": 1, + "bbox": [ + 500, + 300, + 756, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UV mapping is a method of surface parameterization that translates a 3D surface into a 2D image, effectively creating a 2D coordinate system known as a UV map for a polygonal mesh $S$ . This is achieved by explicitly assigning UV coordinates to each vertex of the mesh. Further, an arbitrary surface coordinate on the mesh can be mapped to its 2D coordinate through barycentric interpolation:", + "bbox": [ + 498, + 323, + 892, + 430 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n(u, v) = f (p) \\quad f: S \\rightarrow \\Omega , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 441, + 892, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f$ represents the UV mapping process.", + "bbox": [ + 498, + 469, + 799, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 3 \"UV warping\", we can apply textures to the mesh surface by generating a high-resolution texture image in the UV space. Besides texture, we can also create this kind of 2D map for surface normals and point coordinates. Our objective is to generate a high-quality 2D UV texture image of the given mesh. However, the mapping process requires cutting the continuous texture on the 3D shape into a series of individual patches in the 2D UV plane, as depicted in Figure 3. This fragmentation makes it challenging for the generative model to directly learn the 3D adjacency relationships of the patches within the 2D texture UV map. Consequently, this can lead to discontinuity and inconsistency issues when the generated texture map is applied back to the 3D mesh surface. As shown in Figure 2 \"2D Diffusion\", the diffusion model generates inconsistent textures and suffers from discontinuity issues.", + "bbox": [ + 496, + 484, + 893, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 739, + 591, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The challenge mentioned in Section 3.2 motivates us to propose a coarse-to-fine framework for texture image synthesis, namely Point-UV diffusion, illustrated in Figure 4. To start, we design a 3D point diffusion model to colorize a set of sampled points on the mesh surface, as shown in Figure 4 (Top). This stage leverages the 3D topology for predicting low-frequency colors on the mesh, without being affected by the discontinuity of the UV map. Based on the color components generated by the coarse stage, we then", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4208", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/03bdf3a5f17d771afd0ee6fb532bea7a0253e87315c91a4981dc2a4aed1ff5a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 93, + 883, + 256 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1df90fc3e88730a3a8b4a0cc36e39f51d1602f2414a9d3442ad1d525fe477bcc.jpg", + "image_caption": [ + "Figure 4. The overview of our Point-UV-diffusion framework. (Top) The coarse stage first samples a point cloud on the mesh surface, and then predicts the color for each point using a 3D diffusion model conditioned on shape features including surface normal, mask, and coordinates. Then the points are mapped to the 2D UV map and the remaining uncolored points are filled up by tri-linear interpolation based on the 3D coordinates. (Bottom) The fine stage predicts high-quality textures with a 2D diffusion model conditioned on the shape attributes and the coarse texture image." + ], + "image_footnote": [], + "bbox": [ + 86, + 260, + 883, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "establish a 2D diffusion model in the UV space. This enhances the fidelity of the generated texture, as depicted in Figure 4 (Bottom).", + "bbox": [ + 75, + 497, + 470, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Coarse Stage: Point Diffusion", + "text_level": 1, + "bbox": [ + 76, + 551, + 344, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the coarse stage, we begin by executing farthest point sampling (FPS) on the mesh surface, deriving a point set consisting of $K$ points. These points are defined by their coordinates $\\mathbf{z}_{\\mathrm{coord}}$ and colors $\\mathbf{z}_0$ . During training, a forward diffusion process degrades the clean colors $\\mathbf{z}_0$ , transforming them into a noisy state $\\mathbf{z}_t$ . The noise level is dictated by the time step $t$ , where $t \\in \\{0,1,\\dots,T\\}$ .", + "bbox": [ + 75, + 575, + 468, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our network is trained to reverse this diffusion process, aiming to denoise $\\mathbf{z}_t$ back to its original clean colors. This denoising network is informed by three pre-computed maps: a coordinate map $\\mathbf{x}_{\\mathrm{coord}}$ , a normal map $\\mathbf{x}_{\\mathrm{normal}}$ , and a mask map $\\mathbf{x}_{\\mathrm{mask}}$ (details in Section 3.2). Unless stated otherwise, these conditions concatenate along the channel dimension, culminating in what we term the shape map $\\mathbf{x}_{\\mathrm{shape}}$ :", + "bbox": [ + 75, + 681, + 470, + 789 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\text {s h a p e}} = \\left(\\left[ \\mathbf {x} _ {\\text {n o r m a l}}, \\mathbf {x} _ {\\text {m a s k}}, \\mathbf {x} _ {\\text {c o o r d}} \\right]\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 801, + 468, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our network architecture is constructed upon PVCNN [23], drawing similarities to point-voxel diffusion [49]. However, we introduce slight modifications to amplify the integration of global shape information, contrasting with the approach in [49] which primarily relies on point coordinates. Ini", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tially, we employ a lightweight shape encoder $E_{\\phi}$ to extract a global shape embedding $f_{g}$ from $\\mathbf{x}_{\\mathrm{shape}}$ . This embedding is subsequently fed into the 3D network $G_{\\theta_1}^1$ along with $\\mathbf{z}_t$ , $\\mathbf{z}_{\\mathrm{coord}}$ , and $t$ to predict the color $\\hat{\\mathbf{z}}_0$ :", + "bbox": [ + 498, + 497, + 892, + 559 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {g} = E _ {\\phi} \\left(\\mathbf {x} _ {\\text {s h a p e}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 563, + 890, + 587 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {0} = G _ {\\theta_ {1}} ^ {1} \\left(\\left[ \\mathbf {z} _ {\\mathrm {c o o r d}}, \\mathbf {z} _ {t}, f _ {g}, t \\right]\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 583, + 790, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Style guidance. We observe that the synthesized colors are largely influenced by the predominant colors within the dataset (for instance, textures in the ShapeNet \"chair\" category are typically white, pure magenta, or wood-colored), leading to a lack of diversity in the outputs. To address this bias and promote color diversity, we introduce a style guidance mechanism. This is achieved by flattening each $\\mathbf{z}_0$ into a unidimensional vector, followed by employing PCA to extract the principal component coefficients, thus reducing the dimensionality of $\\mathbf{z}_0$ . Subsequently, K-means clustering is utilized to assign a style label to each $\\mathbf{z}_0$ . As depicted in Figure 5, shapes within a particular cluster exhibit similar color styles. During training, we provide the network with an additional style label $z_{\\mathrm{style}}$ as a condition, which is referred to as style guidance:", + "bbox": [ + 496, + 616, + 892, + 843 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {0} = G _ {\\theta_ {1}} ^ {1} \\left(\\left[ \\mathbf {z} _ {\\text {c o o r d}}, \\mathbf {z} _ {t}, f _ {g}, t, z _ {\\text {s t y l e}} \\right]\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 847, + 890, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, we can guide the network to predict the desired color during inference by providing a certain style condi", + "bbox": [ + 496, + 869, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4209", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion, thereby alleviating the biased color issue.", + "bbox": [ + 76, + 90, + 383, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse texture image. With the colored point clouds, we then project them onto the 2D UV space based on the pre-calculated UV mapping. Following this, we perform KNN interpolation based on the 3D coordinates to assign colors to the remaining uncolored pixels, thereby generating a coarse texture image $\\mathbf{x}_{\\mathrm{coarse}}$ for the given mesh. This coarse texture image offers a coherent base color initialization and functions as a conditional element to guide the high-fidelity texture generation in the second stage, helping to avoid discontinuity caused by UV mapping.", + "bbox": [ + 75, + 135, + 472, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Fine Stage: UV Diffusion", + "text_level": 1, + "bbox": [ + 76, + 301, + 307, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further refine the coarse texture image, we design a fine stage using 2D diffusion in the UV space, as depicted in Figure 4 (Bottom). In addition to the conditions used in the coarse stage, we incorporate the coarse texture image $\\mathbf{x}_{\\mathrm{coarse}}$ as an additional condition. We apply a 2D U-Net $G_{\\theta_2}^2$ combined with self-attention modules to learn the high-quality texture image $\\hat{\\mathbf{x}}_0$ , from a noisy texture image $\\mathbf{x}_t$ :", + "bbox": [ + 76, + 325, + 468, + 434 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {0} = G _ {\\theta_ {2}} ^ {2} \\left(\\left[ \\mathbf {x} _ {\\text {s h a p e}}, \\mathbf {x} _ {t}, \\mathbf {x} _ {\\text {c o a r s e}}, t \\right]\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 446, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hybrid condition. During training, we employ FPS on the ground-truth texture image, $\\mathbf{x}_0$ , followed by interpolation to simulate the coarse texture image, $\\mathbf{x}_{\\mathrm{coarse}}$ . However, in joint cascaded testing involving both stages, we note that the output quality of $\\hat{\\mathbf{x}}_{\\mathrm{coarse}}$ from the first stage doesn't always align flawlessly with the quality of $\\mathbf{x}_{\\mathrm{coarse}}$ . Such mismatches can influence the performance of the subsequent stage. To address this, we introduce a hybrid conditioning method aimed at narrowing this discrepancy. Firstly, we create a smooth texture image $\\mathbf{x}_{\\mathrm{smooth}}$ (see Figure 6), inspired by the blur augmentation described in [17] for cascaded diffusion models. In particular, we segment the mask map into multiple discrete regions using four-connectivity detection, then perform average color pooling within each region based on its connectivity. After this procedure, $\\mathbf{x}_{\\mathrm{smooth}}$ maintains merely the regional color, ensuring more consistent alignment across both training and testing phases. Then, we combine $\\mathbf{x}_{\\mathrm{coarse}}$ with $\\mathbf{x}_{\\mathrm{smooth}}$ , using a certain probability $p_{\\mathrm{hybrid}}$ during training. Thus, the network is forced to be capable of generating textures even when adopting a weaker condition $\\mathbf{x}_{\\mathrm{smooth}}$ in the fine stage during inference. Considering that the diffusion model first generates low-frequency information and then higher one [8], we also explore a condition-truncated sampling, as detailed discussed in Section 5.4, where we condition both maps for generation during the initial time steps and then exclusively utilize the smooth map for the remainder of the generation.", + "bbox": [ + 75, + 493, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/78e68ad0434138d143723c3880402adbd98827bbb432950edb198cb32e5506dc.jpg", + "image_caption": [ + "(a) Samples from cluster 1" + ], + "image_footnote": [], + "bbox": [ + 509, + 92, + 669, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/355323337c1383ea88af947ed41b22f456a8e2ee1f29a46adba57c4c190866bd.jpg", + "image_caption": [ + "(d) Samples from cluster 4" + ], + "image_footnote": [], + "bbox": [ + 678, + 92, + 883, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b80f5fa28dbd811834d417bd7fee31866b25e4db8cdd6595ad89ac370118ef38.jpg", + "image_caption": [ + "Figure 5. Illustration of samples across various clusters. Shapes within a particular cluster exhibit analogous color styles. However, there exists an imbalance in the quantity of shapes among different clusters, leading to challenges in unbiased synthesis.", + "(a) Texture image", + "Figure 6. Varieties of texture images. (a) Original texture image enriched with high-frequency details, (b) Coarse texture image, and (c) Smooth texture image." + ], + "image_footnote": [], + "bbox": [ + 509, + 303, + 627, + 396 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d2b50a6733fa74d83efc1a693d73306df119eb23950b24af615cd5fd3178b772.jpg", + "image_caption": [ + "(b) Coarse texture image" + ], + "image_footnote": [], + "bbox": [ + 635, + 303, + 751, + 396 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8eaabc9f4446cc5dfd47c1d7b23a300f7a4de380b347cc2d3005cd19a031a678.jpg", + "image_caption": [ + "(c) Smooth texture image" + ], + "image_footnote": [], + "bbox": [ + 764, + 303, + 883, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 481, + 632, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 506, + 818, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments across four categories of the ShapeNet dataset [4], namely, chair, table, car, and bench. Before training, we use an open-source UV-Atlas tool [46] to generate the UV map and pre-process the dataset to obtain the shape maps and ground-truth texture images. We sample $K = 4096$ points in our coarse stage and synthesize a texture image in the UV space with a resolution of $512 \\times 512$ for the fine stage. For training the diffusion models, akin to [1], we predict the clean signals. This approach provides more stable training than predicting the noise component as recommended by [16]. We employ the cosine noise scheduling [27] ranging from 0.0001 to 0.02 over 1,024 time steps for both stages. In the fine stage, we leverage the noise scaling strategy from [7] and also incorporate a rendering loss. Specifically, we randomly select four views and render the mesh using the predicted UV map and the ground-truth UV map to produce $1024 \\times 1024$ images. Subsequently, we crop $224 \\times 224$ patches from these images and compute the corresponding $L_{1}$ loss. For the hybrid condition in the fine stage, we use $p_{\\mathrm{hybrid}} = 0.3$ and sweep over condition-truncated time $t_{c}$ (see Figure 12).", + "bbox": [ + 496, + 530, + 890, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared methods. We compare our method with state-of-the-art approaches, including Texturify [38], Texture", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4210", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/86b797e91ed0ca00ebed9a3cc70e0b93e0e6164bedfeaae3d8a5080965b9cba0.jpg", + "image_caption": [ + "Figure 7. Our texture synthesis results in different categories. The results on chairs, tables, cars, and benches demonstrate that our approach can generate natural, vivid, and diverse textures, even though the given 3D shapes are challenging, such as cutouts on the bench." + ], + "image_footnote": [], + "bbox": [ + 106, + 95, + 867, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fields [29], and PVD [49], in the context of unconditional texture generation. PVD was originally designed for point cloud generation instead of texture generation, so it cannot be directly compared with our approach. To facilitate comparison, we modify this framework to learn RGB values from the sampled point cloud. This extension, referred to as \"PVD-Tex\", serves as a baseline for directly learning texture in the point space using the diffusion model.", + "bbox": [ + 75, + 414, + 472, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Unconditional Texture Generation", + "text_level": 1, + "bbox": [ + 76, + 546, + 377, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Gallary. First, we showcase our texture generation results for each category without additional conditions. The results in Figure 7 and Figure 1 demonstrate the remarkable performance of our method, which is able to generate appealing textures with fine details and preserve the geometric structures. Note that our method is compatible with meshes of diverse topology and intricate geometric details.", + "bbox": [ + 75, + 571, + 470, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparisons. As depicted in Figure 8, our approach excels at generating high-quality textures while maintaining the geometric intricacies of the input mesh. Firstly, the results of Texture Fields [29] appear deficient in high-frequency details. Additionally, while PVD-Tex [49] is able to produce spatially varying colors to some extent (e.g., the car in Figure 8), it falls short in synthesizing intricate high-frequency details. Lastly, even though Texturify [38] demonstrates an ability to generate finer textures, it compromises on preserving slender structures (i.e. chair).", + "bbox": [ + 75, + 696, + 468, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparisons. To quantitatively compare with existing works, we follow [38] and assess the genera", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion quality using Frechet Inception Distance (FID)[15] and Kernel Inception Distance (KID)[2], metrics that are widely used for evaluating image generation models. To this end, we render $512 \\times 512$ images from each generated textured mesh and ground-truth textured mesh using four distinct camera views. Table 1 presents the quantitative comparisons with current approaches, revealing that our method surpasses existing works.", + "bbox": [ + 496, + 414, + 893, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Conditional Texture Generation", + "text_level": 1, + "bbox": [ + 498, + 545, + 782, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we demonstrate the capability of our framework to synthesize textures conditioned on either text prompts or a single-view image. We conduct our experiments on the chair and table categories. For the text condition, we utilize text-shape pairs as provided in [6] (with additional corrections for text accuracy). For the image condition, we randomly render a view from the ground-truth mesh. To infuse the network with condition-specific information, we use the pre-trained vision-language model CLIP [33] to extract the corresponding embedding from either the image or the text. This embedding is then fed into a simple MLP to incorporate the information into the diffusion model as a condition for both training and inference. As depicted in Figure 9, our method succeeds in generating textures that align well with the given text descriptions or images. We also compare our approach with Text2Mesh [24], a test-time optimization method. Text2Mesh takes around 10 minutes per instance, while ours only requires 30 seconds. Importantly, to adapt to our task which requires preserving the mesh geometry, we freeze the geometry deformation branch of Text2Mesh to generate only colors. As shown in Figure 10, Text2Mesh", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4211", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc1e5f50344813ec6fab5477aca5d6115d6fa928b654080272d06ad0cfea7468.jpg", + "image_caption": [ + "Figure 8. Qualitative comparisons with existing works. Given pure 3D shapes, Texturify [38] generates textures on the surface of 3D shape but tends to damage the topology; Texture fields [29] and PVD-Tex [49] produce textures with limited details. On the contrary, our approach faithfully preserves topology and produces realistic appearances." + ], + "image_footnote": [], + "bbox": [ + 89, + 97, + 460, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is bound by vertex resolution, while our method can deliver detailed visuals even on low-resolution meshes, offering a distinct advantage in graphics.", + "bbox": [ + 75, + 560, + 468, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 614, + 241, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Coarse-to-fine diffusion. To manifest the effectiveness of our coarse-to-fine diffusion strategy, we conduct experiments on two baselines \"w/o coarse stage\" and \"w/o fine stage\". The \"w/o coarse stage\" configuration indicates that we directly generate the UV texture image using the fine stage, bypassing the initialization from the coarse stage. The \"w/o fine stage\" configuration signifies the result of the coarse stage, after assigning colors to the uncolored pixels using KNN interpolation. In both scenarios, the model produces inferior outcomes relative to our full model, as shown in Table 2 and Figure 11. Absent the coarse stage, the generated result suffers from noticeably inconsistent colors. Without the fine stage, the results are over-smoothed.", + "bbox": [ + 75, + 638, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hybrid condition. As shown in Table 2 and Figure 11 \"coarse condition\", the fine stage cannot generate high-quality textures when it is exclusively conditioned on the", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d65f2fe611a112f9326e484b0374fcb845e223f7ce4fafd1259ec5be1d3b39c7.jpg", + "image_caption": [ + "Figure 9. Results of conditional texture generation. Our method is adaptable to craft textures guided by text descriptions (rows 1-3) or single-view images (rows 4-6)." + ], + "image_footnote": [], + "bbox": [ + 511, + 90, + 888, + 526 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9697bacf708f3c7ee5d0feeb1e4398bbe83b2a70456554c5d10c05e5e5f8cbcb.jpg", + "image_caption": [ + "Figure 10. Comparison with Text2Mesh. We freeze the geometry deformation branch of Text2Mesh to adapt to our task. Ours can generate high-frequency details for low-resolution mesh." + ], + "image_footnote": [], + "bbox": [ + 532, + 585, + 867, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "coarse map. Further, the generated quality remains unsatisfactory if we apply a hybrid condition entirely during training (i.e., $p_{\\mathrm{hybrid}} = 1.0$ ), as shown in \" $p_{\\mathrm{hybrid}} = 1.0$ \". We attribute this to the network's propensity to depend solely on the coarse map during training and generation. In contrast, our full model (i.e., hybrid condition with $p_{\\mathrm{hybrid}} = 0.3$ ) achieves significant improvement both", + "bbox": [ + 496, + 795, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4212", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/87145427ccf2b3f6d2b01820679a7938dc1e30769356d617c5a32e0980749cdd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ChairCarTableBench
MethodsFID ↓KID ↓FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓
Texture Fields [29]24.241.07156.3813.6468.964.2062.712.96
Texturify [38]27.801.3273.164.71----
PVD-Tex [49]15.520.6259.473.7416.120.5528.940.39
Ours9.880.2226.890.689.630.1523.090.15
", + "bbox": [ + 169, + 88, + 802, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84a3def2507621a83f179e2207ec41d7871a84169839c3e98f7ea23eb33bf6cf.jpg", + "image_caption": [ + "(a) untextured mesh" + ], + "image_footnote": [], + "bbox": [ + 127, + 224, + 207, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de8ba1fb2bbc89c540f09921c6de151892921d8d4601c509a949c7e569b92b51.jpg", + "image_caption": [ + "(b) w/o coarse stage" + ], + "image_footnote": [], + "bbox": [ + 236, + 224, + 316, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/982320679431f2013186f29101ba9df8eea6815de214fb528438f7143ddba134.jpg", + "image_caption": [ + "(c) w/o fine stage" + ], + "image_footnote": [], + "bbox": [ + 344, + 224, + 424, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fa9d3f85ce3244cde44c54ba778651e23f9615245bc9f5138883c8ace80c0938.jpg", + "image_caption": [ + "(d) coarse condition" + ], + "image_footnote": [], + "bbox": [ + 130, + 351, + 207, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b14609cb0ea646c3e293935110cac2f5b0df18528fa22ebf0fcb096eaa07d55d.jpg", + "image_caption": [ + "(e) $p_{hybrid} = 1.0$" + ], + "image_footnote": [], + "bbox": [ + 241, + 352, + 320, + 458 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca8050151c919536b4463c759fad95ce7b7bb56b3ff226b18a716ccd92f8bf8b.jpg", + "image_caption": [ + "(f) ours", + "Figure 11. Ablation studies. Without our two-stage generation pipeline or hybrid condition designs, the generated results are inconsistent with the 3D shape or lack of high-frequency details." + ], + "image_footnote": [], + "bbox": [ + 348, + 352, + 426, + 459 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8b00a2e0803c4dc49d128bfa2856df111f68a6da9652301dabdecdef83f95cf3.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons with existing works. Ours outperforms other approaches on both FID [15] and KID $(\\times 10^{2})$ [2]." + ], + "table_footnote": [], + "table_body": "
FID ↓KID ↓
w/o fine stage17.880.76
w/o coarse stage15.110.49
coarse condition14.930.56
\\(p_{hybrid}=1.0\\)15.250.59
Ours9.880.22
", + "bbox": [ + 145, + 535, + 397, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "qualitatively and quantitatively, as presented in Table 2 and Figure 11 \"ours\". Furthermore, we also sweep over the effect of condition-truncated time $t_c$ for inference. As Figure 12 shows, $t_c = 0.4$ strikes a sweet point. The model utilizes information from both conditions maps to better generate low-frequency components during the first $40\\%$ of the sampling timesteps. After that, the coarse map $\\mathbf{x}_{\\mathrm{coarse}}$ is dropped, and the model further focuses on fine-grained detail generation without relying on the coarse map.", + "bbox": [ + 75, + 684, + 470, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Style guidance. The style guidance is aimed at addressing the issue of insufficient diversity due to color distribution bias in the dataset. As a result, we do not utilize FID or KID for evaluation since they evaluate the distribution sim", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ff88d7cbada0b4dfd13deae9aa6e0f98f5afe296871f23f4d2367ad3d83af73d.jpg", + "table_caption": [ + "Table 2. Ablation studies. This table shows the effectiveness of each component in our proposed method." + ], + "table_footnote": [], + "table_body": "
MethodsPreference↑LPIPS↑
Ours49.8%0.083
Texturify [38]15.9%0.086
PVD-Tex [49]29.2%0.029
Texture Fields [29]5.1%0.005
", + "bbox": [ + 535, + 219, + 852, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. LPIPS and user study. This table shows the average LPIPS and preference via user study in the chair category.", + "bbox": [ + 500, + 299, + 890, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The diversity of the LPIPS is a key feature that is crucial to the quality of the generated results and ground truth. However, the ground-truth distribution shows a strong bias towards particular colors, making these metrics inappropriate for assessing generative texture diversity. In contrast, we adopt LPIPS [48] to measure the pairwise similarity among five textures generated by our method given the same input mesh, where a larger diversity in textures will lead to a higher LPIPS value. In Figure 13, we report the quantitative results for 500 shapes. Our approach achieves a higher LPIPS in most of the evaluated cases, indicating better diversity. Besides, as shown in Figure 14, we present the results of generating three textures randomly for the shapes without and with style guidance. In the latter case, we uniformly sample three style labels as style inputs. It is evident that without style guidance, the generated textures are nearly identical, and the color styles of different shapes tend to be similar. With the introduction of style guidance, however, the diversity of colors has significantly increased. We also conduct a comparison with other methods in terms of diversity, as well as a quality assessment through user studies, as shown in Table 3.", + "bbox": [ + 496, + 354, + 890, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion, Limitations, and Conclusions", + "text_level": 1, + "bbox": [ + 500, + 690, + 867, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents Point-UV diffusion, a brand-new framework that employs a coarse-to-fine pipeline to generate textures for 3D meshes. We begin with a 3D diffusion model to synthesize low-frequency texture components from point clouds, which maintain 3D consistency. We then refine the textures using a 2D UV-space diffusion model. Our method is compatible with meshes with arbitrary topology and can faithfully preserve the geometry structure. We further demonstrate the flexibility of our framework by extending it to conditional generative models.", + "bbox": [ + 496, + 717, + 890, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite its merits, our method has inherent limitations. Similar to other methodologies relying on 3D data for train-", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4213", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5fd0d599729542c9b7c974173dd1ac7d5e71b924af103adc5dced05045c12332.jpg", + "image_caption": [ + "Figure 12. Examination of condition-truncated time. The horizontal axis denotes the percentage of time points when the coarse texture image is conditioned during inference. The left vertical axis indicates the FID value, while the right vertical axis shows the KID value $(\\times 10^{2})$ ." + ], + "image_footnote": [], + "bbox": [ + 102, + 94, + 442, + 239 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/06390beb6f5eec44e05400f76334549a09b8030b3ba6da06b328c2c9b0ab768a.jpg", + "image_caption": [ + "Figure 13. Diversity measurement. Utilizing style guidance leads to larger LPIPS scores, indicating enhanced generation diversity." + ], + "image_footnote": [], + "bbox": [ + 102, + 329, + 444, + 529 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ing, our technique is upper-bounded by the scope and diversity of current 3D datasets. This restriction poses challenges in generating textures that parallel the depth of effects seen in 2D image synthesis. Moreover, our method's efficacy relies on the quality of UV mapping. Our approach faces difficulties in rendering high-quality results for meshes where the UV mapping produces excessive fragmented cuts, resulting in fragmented artifacts. This phenomenon is commonly observed in the car category, as shown in the Figure 15. We believe the emergence of larger and more diverse 3D datasets would be helpful for generating superior-quality textures. Further advancements in UV parameterization would also be beneficial in augmenting our method's consistency.", + "bbox": [ + 75, + 589, + 468, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 815, + 264, + 832 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work has been supported by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant No. 17202422), and RGC matching fund scheme (RMGS). Part", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/7dd3488f2ed95f997cfcd5c595ebd414538e0b8d0a8a065ad88e85c438c55789.jpg", + "image_caption": [ + "(a) w/o style guidance", + "Result 1" + ], + "image_footnote": [], + "bbox": [ + 550, + 111, + 637, + 310 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1c4718cca7ba0e778997d2ff96a834b0db2dde9ea762c238ecb6c23b4d182451.jpg", + "image_caption": [ + "Result 2" + ], + "image_footnote": [], + "bbox": [ + 661, + 111, + 740, + 309 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4890e18c992d21c254ca1d2ce295014cf1c69273442006ab745df2fe643f5c7e.jpg", + "image_caption": [ + "Result 3" + ], + "image_footnote": [], + "bbox": [ + 767, + 111, + 846, + 309 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ce8b49180e1a128a3c18a36a4ed30a43e3aa5778fcc29b6351b3bac82173037d.jpg", + "image_caption": [ + "(b) w/ style guidance", + "Result 1", + "Figure 14. Efficacy of style guidance. The model without style guidance produces consistent colors across different random seeds (rows 1-2). In contrast, with style guidance, diverse outcomes emerge (rows 3-4)." + ], + "image_footnote": [], + "bbox": [ + 547, + 354, + 629, + 551 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c7e8219b8cc398c024d35bc258208c08fe92b14bb737544b5a608a3377b07c9b.jpg", + "image_caption": [ + "Result 2" + ], + "image_footnote": [], + "bbox": [ + 655, + 354, + 735, + 551 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/74411651e70bb3a0ad2617dfa06e510a30e900647cec2688889e10165142ce83.jpg", + "image_caption": [ + "Result 3" + ], + "image_footnote": [], + "bbox": [ + 761, + 354, + 841, + 551 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/37ea41a649f5635d3496f1b5e4805874ba17d11c31569c9b66cd288641236060.jpg", + "image_caption": [ + "(a) Generated Texture Image", + "Figure 15. Failure case. Our approach still struggles to generate seamless results when there are too many fragmented cuts of the UV map." + ], + "image_footnote": [], + "bbox": [ + 540, + 665, + 669, + 765 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c79cf71ae6423c18fb1df3c0a9a3ae93b90dd28b4932b2f24394bb28a8f1c2a8.jpg", + "image_caption": [ + "(b) Textured Mesh" + ], + "image_footnote": [], + "bbox": [ + 696, + 685, + 861, + 758 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4214", + "bbox": [ + 480, + 924, + 514, + 936 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. arXiv preprint arXiv:2211.09869, 2022. 3, 5", + "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6, 8", + "[3] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 2", + "[4] Angel X. Chang, Thomas Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], 2015. 5", + "[5] Bindita Chaudhuri, Nikolaos Sarafianos, Linda Shapiro, and Tony Tung. Semi-supervised synthesis of high-resolution editable textures for 3d humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7991-8000, 2021. 1", + "[6] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III 14, pages 100-116. Springer, 2019. 1, 2, 6", + "[7] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 5", + "[8] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5", + "[9] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10673-10683, 2022. 2", + "[10] Jun Gao, Wenzheng Chen, Tommy Xiang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. Advances In Neural Information Processing Systems, 33:9936-9947, 2020. 1", + "[11] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 1, 2", + "[12] Lin Gao, Tong Wu, Yu-Jie Yuan, Ming-Xian Lin, Yu-Kun Lai, and Hao Zhang. Tm-net: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG), 40(6):1-15, 2021. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. In Advances in neural information processing systems, pages 5767-5777, 2017. 1", + "[14] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Learning a neural 3d texture space from 2d exemplars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8356-8364, 2020. 2", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. NIPS, 2017. 6, 8", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models, 2020. 2, 3, 5", + "[17] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23(47):1-33, 2022. 5", + "[18] Jingyu Hu, Ka-Hei Hui, Zhengzhe Liu, Ruihui Li, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation, inversion, and manipulation. arXiv preprint arXiv:2302.00190, 2023. 3", + "[19] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation. In SIGGRAPH Asia 2022 Conference Papers, pages 1–9, 2022. 1, 3", + "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1", + "[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3", + "[22] Zhengzhe Liu, Peng Dai, Ruihui Li, Xiaojuan Qi, and Chi-Wing Fu. Iss: Image as stetting stone for text-guided 3d shape generation. International Conference on Learning Representations, 2023. 2", + "[23] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Pointvoxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems, 32, 2019. 4", + "[24] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3, 6", + "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[26] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4215", + "bbox": [ + 482, + 924, + 514, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Alex Nichol and Prafulla Dhariwal. Improved Denoising Diffusion Probabilistic Models, 2021. 5", + "[28] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3", + "[29] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 1, 2, 6, 7, 8", + "[30] Dario Pavllo, Graham Spinks, Thomas Hofmann, Marie-Francine Moens, and Aurelien Lucchi. Convolutional generation of textured 3d meshes. Advances in Neural Information Processing Systems, 33:870-882, 2020. 2", + "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations, 2023. 2, 3", + "[32] Tiziano Portenier, Siavash Arjomand Bigdeli, and Orcun Goksel. Gramgan: Deep 3d texture synthesis from 2d exemplars. Advances in Neural Information Processing Systems, 33:6994-7004, 2020. 2", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6", + "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 3", + "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 1", + "[37] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 2", + "[38] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. arXiv preprint arXiv:2204.02411, 2022. 2, 5, 6, 7, 8", + "[39] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. Advances in Neural Information Processing Systems, 35:24487-24501, 2022. 2", + "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Confer" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ence on Machine Learning, pages 2256-2265. PMLR, 2015. 3", + "[41] Yongbin Sun, Ziwei Liu, Yue Wang, and Sanjay E Sarma. Im2avatar: Colorful 3d reconstruction from a single image. arXiv preprint arXiv:1804.06375, 2018. 2", + "[42] Shubham Tulsiani, Nilesh Kulkarni, and Abhinav Gupta. Implicit mesh reconstruction from unannotated image collections. arXiv preprint arXiv:2007.08504, 2020. 2", + "[43] Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2626-2634, 2017. 2", + "[44] Hao Wang, Guosheng Lin, Steven Hoi, and Chunyan Miao. Cycle-consistent inverse GAN for text-to-image synthesis. ACM MM, 2021. 1", + "[45] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12456-12465, 2021. 1", + "[46] Jonathan Young. xatlas, 2018. https://github.com/jpcy/xatlas.5", + "[47] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 3", + "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 8", + "[49] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 3, 4, 6, 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 602 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4216", + "bbox": [ + 482, + 925, + 514, + 936 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/295801ea-f88d-446b-aeea-24167c36963d_model.json b/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/295801ea-f88d-446b-aeea-24167c36963d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..78c98a5a6c5821f576e7eb1799fb323e55bf0300 --- /dev/null +++ b/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/295801ea-f88d-446b-aeea-24167c36963d_model.json @@ -0,0 +1,2609 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.188, + 0.131, + 0.784, + 0.151 + ], + "angle": 0, + "content": "Texture Generation on 3D Meshes with Point-UV Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.18, + 0.178, + 0.198 + ], + "angle": 0, + "content": "Xin Yu" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.181, + 0.301, + 0.2 + ], + "angle": 0, + "content": "Peng Dai" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.181, + 0.429, + 0.198 + ], + "angle": 0, + "content": "Wenbo Li2" + }, + { + "type": "text", + "bbox": [ + 0.468, + 0.181, + 0.542, + 0.198 + ], + "angle": 0, + "content": "Lan Ma3" + }, + { + "type": "text", + "bbox": [ + 0.579, + 0.181, + 0.706, + 0.199 + ], + "angle": 0, + "content": "Zhengzhe Liu\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.746, + 0.181, + 0.86, + 0.2 + ], + "angle": 0, + "content": "Xiaojuan Qi\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.216, + 0.884, + 0.235 + ], + "angle": 0, + "content": "1The University of Hong Kong, 2The Chinese University of Hong Kong, 3TCL Corporate Research" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.237, + 0.39, + 0.252 + ], + "angle": 0, + "content": "{yuxin, daipeng, xjqi}@eee.hku.hk" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.237, + 0.697, + 0.252 + ], + "angle": 0, + "content": "{wenboli,zzliu}@cse.cuhk.edu.hk" + }, + { + "type": "text", + "bbox": [ + 0.732, + 0.238, + 0.859, + 0.251 + ], + "angle": 0, + "content": "rubyma@tcl.com" + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.264, + 0.887, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.509, + 0.895, + 0.539 + ], + "angle": 0, + "content": "Figure 1. A gallery of generated results by our Point-UV diffusion. Our method is capable of processing meshes of any genus, generating diversified, geometry-compatible, and high-fidelity textures." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.549, + 0.314, + 0.564 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.574, + 0.474, + 0.831 + ], + "angle": 0, + "content": "In this work, we focus on synthesizing high-quality textures on 3D meshes. We present Point-UV diffusion, a coarse-to-fine pipeline that marries the denoising diffusion model with UV mapping to generate 3D consistent and high-quality texture images in UV space. We start with introducing a point diffusion model to synthesize low-frequency texture components with our tailored style guidance to tackle the biased color distribution. The derived coarse texture offers global consistency and serves as a condition for the subsequent UV diffusion stage, aiding in regularizing the model to generate a 3D consistent UV texture image. Then, a UV diffusion model with hybrid conditions is developed to enhance the texture fidelity in the 2D UV space. Our method can process meshes of any genus, generating diversified, geometry-compatible, and high-fidelity textures. Code is available at https://cvmilab.github.io/Point-UV-Diffusion." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.549, + 0.631, + 0.564 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.895, + 0.715 + ], + "angle": 0, + "content": "Texturing 3D meshes is a fundamental task in computer vision and graphics. It enhances the visual richness of 3D objects, thereby facilitating their application in various fields such as video games, 3D movies, and AR/VR technologies. However, generating high-quality textures can be daunting and time-consuming, often requiring specialized knowledge and resources. As such, there is a pressing need for an efficient approach to automatically create high-quality textures on 3D meshes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Despite the substantial progress the community has made in 2D image synthesis and 3D shape generation using GANs [13, 20, 44, 11] or diffusion models [34, 36, 35, 19], crafting realistic textures on mesh surfaces remains challenging. One major difficulty stems from the need for suitable 3D representations for texture synthesis. Early approaches investigate the use of voxels [5, 45, 6] or point clouds [10] and synthesize point/voxel colors. However, they can only afford to synthesize low-resolution results with low-fidelity textures due to memory and model complexity constraints. In response, Texture Fields [29] adopts an implicit representation with the potential to synthesize" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.887, + 0.241, + 0.901 + ], + "angle": 0, + "content": "†: Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.517, + 0.937 + ], + "angle": 0, + "content": "4206" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "high-resolution textures, but still could not yield satisfactory results as shown in Figure 2 (a): over-smoothed results. Most recently, Siddiqui et al. [38] propose to parameterize the shape as tetrahedral meshes and introduce tetrahedral mesh convolution to enhance local details. Albeit improving results, tetrahedral parameterization inevitably destroys geometric details of the input mesh and thus cannot faithfully preserve the original structure. As shown in Figure 2 (b), the delicate structures of the chair's back are absent. Moreover, the generative models utilized in these methods are limited to GANs [29, 38] and VAEs [29, 12]. The more advanced diffusion model, which could potentially open up new avenues for high-quality texture generation, remains insufficiently explored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.47, + 0.502 + ], + "angle": 0, + "content": "In this paper, we delve into a novel texture representation based on UV maps and investigate the advanced diffusion model for texture generation. The 2D nature of the UV map enables it to circumvent the cost of high-resolution point/voxel representations. Besides, the UV map is compatible with arbitrary mesh topologies, thereby preserving the original geometric structures. However, while promising, direct integration of the UV map representation with a 2D diffusion model presents challenges in synthesizing seamless textures, leading to severe artifacts, as shown in Figure 2 (c). This occurs because the UV mapping process fragments the continuous texture on the 3D surface into isolated patches on the 2D UV plane (see Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.718 + ], + "angle": 0, + "content": "To this end, we introduce Point-UV diffusion, a two-stage coarse-to-fine framework consisting of point diffusion and UV diffusion. Specifically, we initially design a point diffusion model to generate color for sampled points that act as low-frequency texture components. This model is equipped with a style guidance mechanism that alleviates the impact of biased color distributions in the dataset and facilitates diversity during inference. Next, we project these colorized points onto the 2D UV space with 3D coordinate interpolation, thereby generating a coarse texture image that maintains 3D consistency and continuity. Given the coarse textured image, we develop a UV diffusion model with elaborately designed hybrid conditions to improve the quality of the textures (see Figure 2 (ours) and Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In short, our contributions are as follows: 1) We propose a new framework for texture generation for given meshes. Our representation can handle meshes with arbitrary topology and is able to faithfully preserve geometric structures. 2) To the best of our knowledge, we are the first to train a diffusion model specifically for mesh texture generation. Our coarse-to-fine framework allows us to enjoy the efficiency of 2D representation while enhancing 3D consistency. 3) We compare our approach with multiple methods in unconditional generation and achieve state-of-the-art results. Furthermore, we demonstrate that our method can be easily extended to scenarios with text-conditioning and" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.094, + 0.655, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.267, + 0.628, + 0.277 + ], + "angle": 0, + "content": "(a) Texture Fields vs Ours" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.094, + 0.775, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.267, + 0.753, + 0.277 + ], + "angle": 0, + "content": "(b) Texturify vs Ours" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.094, + 0.891, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.267, + 0.884, + 0.277 + ], + "angle": 0, + "content": "(c) 2D diffusion vs Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.287, + 0.893, + 0.357 + ], + "angle": 0, + "content": "Figure 2. Comparisons with different methods. Our generative results (a) possess high-quality details, (b) faithfully preserve the mesh structure, and (c) are better consistent with the given shape, compared with Texture Fields [29], Texturify [38] and 2D diffusion [16], respectively." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.371, + 0.732, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.497, + 0.641, + 0.507 + ], + "angle": 0, + "content": "Textured Mesh" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.37, + 0.893, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.498, + 0.849, + 0.509 + ], + "angle": 0, + "content": "Texture Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.516, + 0.893, + 0.558 + ], + "angle": 0, + "content": "Figure 3. Illustration of UV mapping process. It establishes connections between the 2D texture map and the surface appearance of 3D shape." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.586, + 0.637, + 0.601 + ], + "angle": 0, + "content": "image-conditioning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.618, + 0.642, + 0.633 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Texture generation. Early works [41, 43, 6] propose using voxels to represent colors. However, due to a cubic increase in memory usage and computational cost with rising resolution, these methods produce only coarse, low-resolution textures. For meshes, recent work in texture generation [38] predicts the color for each face, but the representation capacity remains largely constrained by the mesh's resolution. Another set of works [30, 42, 26, 12] leverage UV mapping, but they require either spherical parameterization or part segmentation and, therefore, are confined to handling low-genus shapes. Other works [32, 14] rely on image exemplars, while our approach focuses on unconditional generation. Recently, implicit functions attract increasing attention for generative tasks [11, 31, 39, 3, 9, 37, 29, 22]. Most of this research centers on generating 3D-aware images [39, 3, 9, 37], rather than synthesizing textures for given 3D meshes. Texture Fields [29] represents" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4207" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "the most similar work to our task, which, however, tends to produce over-smoothed results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.473, + 0.381 + ], + "angle": 0, + "content": "Some methods focus on test-time optimization for generation tasks. For instance, Text2Mesh [24] employs CLIP [33] to design a loss function. This function stylizes a 3D mesh to align with a target text prompt, predicting both the color and displacement for each mesh vertex. More recently, DreamFusion [31] and Magic3D [21] leverage pretrained stable diffusion [35] for score distillation, creating either NeRF [25] or a 3D mesh. However, these approaches often demand extensive optimization time or necessitate alterations to the existing mesh geometry [24]. In this paper, our objective is to generate textures for specified full-3D meshes with arbitrary topology. This goal entails two fundamental requirements: 1) preserving the original mesh's geometric structure; 2) producing a 3D texture representation exportable as either vertex color or a uv-texture-image, instead of merely generating-rendering multi-view or 3D-aware images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.473, + 0.641 + ], + "angle": 0, + "content": "Diffusion models for 3D generation. In addition to 2D image generation, diffusion models recently gain significant attention in 3D generation, leading to many related works [49, 18, 19, 28, 31, 47, 1]. For instance, Zhou et al. [49] introduce point-voxel diffusion for point cloud generation and completion. Hui et al. [19] and Hu et al. [18] suggest a compact wavelet domain representation for shapes, enabling higher-quality shape creation via diffusion models. Nichol et al. [28] present Point-E, a system that uses a text-conditional strategy to produce colored 3D point clouds. This approach enables an efficient synthesis of intricate 3D shapes from textual prompts. Yet, this system yields lower-resolution point clouds, often missing detailed shapes and textures. In this work, we develop a brand new 3D diffusion model for texture image synthesis, which allows high-fidelity and 3D-consistent texture generation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.653, + 0.218, + 0.669 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.318, + 0.695 + ], + "angle": 0, + "content": "3.1. Denoising diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Diffusion model [40, 16] is a kind of likelihood-based generative model that has gained significant attention recently. It learns the data distribution \\(q(x_0)\\) by progressive denoising from a prior Gaussian distribution. Given a sample from the data distribution \\(x_0 \\sim q(x_0)\\), a fixed forward process \\(q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0) = \\prod_{t=1}^{T} q(\\mathbf{x}_t \\mid \\mathbf{x}_{t-1})\\) is used to perturb the data with Gaussian kernels \\(q(\\mathbf{x}_t \\mid \\mathbf{x}_{t-1}) \\coloneqq \\mathcal{N}\\left(\\sqrt{1 - \\beta_t} \\mathbf{x}_{t-1}, \\beta_t \\mathbf{I}\\right)\\), producing increasingly noisy latent variables \\(\\{x_1, x_2, \\ldots, x_T\\}\\). Then, a parameterized Markov process \\(p_\\theta(\\mathbf{x}_{0:T}) = p(\\mathbf{x}_T) \\prod_{t=1}^{T} p_\\theta(\\mathbf{x}_{t-1} \\mid \\mathbf{x}_t)\\) with transition kernel \\(p_\\theta(\\mathbf{x}_{t-1} \\mid \\mathbf{x}_t) \\coloneqq \\mathcal{N}\\left(\\mu_\\theta(\\mathbf{x}_t, t), \\sigma_t^2 \\mathbf{I}\\right)\\) is optimized through maximizing a variational lower bound of log data likelihood, which essentially targets to match the joint" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.646, + 0.108 + ], + "angle": 0, + "content": "distribution \\(q(\\pmb{x}_{0:T})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.117, + 0.894, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E} _ {q \\left(\\mathbf {x} _ {0}\\right)} \\left[ \\log p _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\right] \\geq \\operatorname {E} _ {q \\left(\\mathbf {x} _ {0: T}\\right)} \\left[ \\log \\frac {p _ {\\theta} \\left(\\mathbf {x} _ {0 : T}\\right)}{q \\left(\\mathbf {x} _ {1 : T} \\mid \\mathbf {x} _ {0}\\right)} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.161, + 0.893, + 0.192 + ], + "angle": 0, + "content": "After training, novel samples can then be generated via iterative sampling from \\(p_{\\theta}(\\boldsymbol{x}_{t-1} \\mid \\boldsymbol{x}_t)\\) following:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.202, + 0.836, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t - 1} = \\mu_ {\\theta} (\\mathbf {x} _ {t}, t) + \\sqrt {\\beta_ {t}} \\mathbf {z}, \\mathbf {z} \\sim \\mathcal {N} (0, \\mathbf {I}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.232, + 0.894, + 0.292 + ], + "angle": 0, + "content": "Diffusion models have shown impressive capabilities in generating high-quality and diverse content. This paper proposes a texture generation framework based on diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.301, + 0.757, + 0.318 + ], + "angle": 0, + "content": "3.2. UV mapping and challenges." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.324, + 0.893, + 0.431 + ], + "angle": 0, + "content": "UV mapping is a method of surface parameterization that translates a 3D surface into a 2D image, effectively creating a 2D coordinate system known as a UV map for a polygonal mesh \\( S \\). This is achieved by explicitly assigning UV coordinates to each vertex of the mesh. Further, an arbitrary surface coordinate on the mesh can be mapped to its 2D coordinate through barycentric interpolation:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.442, + 0.893, + 0.459 + ], + "angle": 0, + "content": "\\[\n(u, v) = f (p) \\quad f: S \\rightarrow \\Omega , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.47, + 0.8, + 0.485 + ], + "angle": 0, + "content": "where \\( f \\) represents the UV mapping process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.895, + 0.728 + ], + "angle": 0, + "content": "As illustrated in Figure 3 \"UV warping\", we can apply textures to the mesh surface by generating a high-resolution texture image in the UV space. Besides texture, we can also create this kind of 2D map for surface normals and point coordinates. Our objective is to generate a high-quality 2D UV texture image of the given mesh. However, the mapping process requires cutting the continuous texture on the 3D shape into a series of individual patches in the 2D UV plane, as depicted in Figure 3. This fragmentation makes it challenging for the generative model to directly learn the 3D adjacency relationships of the patches within the 2D texture UV map. Consequently, this can lead to discontinuity and inconsistency issues when the generated texture map is applied back to the 3D mesh surface. As shown in Figure 2 \"2D Diffusion\", the diffusion model generates inconsistent textures and suffers from discontinuity issues." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.592, + 0.755 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The challenge mentioned in Section 3.2 motivates us to propose a coarse-to-fine framework for texture image synthesis, namely Point-UV diffusion, illustrated in Figure 4. To start, we design a 3D point diffusion model to colorize a set of sampled points on the mesh surface, as shown in Figure 4 (Top). This stage leverages the 3D topology for predicting low-frequency colors on the mesh, without being affected by the discontinuity of the UV map. Based on the color components generated by the coarse stage, we then" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4208" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.094, + 0.885, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.261, + 0.885, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.403, + 0.896, + 0.475 + ], + "angle": 0, + "content": "Figure 4. The overview of our Point-UV-diffusion framework. (Top) The coarse stage first samples a point cloud on the mesh surface, and then predicts the color for each point using a 3D diffusion model conditioned on shape features including surface normal, mask, and coordinates. Then the points are mapped to the 2D UV map and the remaining uncolored points are filled up by tri-linear interpolation based on the 3D coordinates. (Bottom) The fine stage predicts high-quality textures with a 2D diffusion model conditioned on the shape attributes and the coarse texture image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.471, + 0.545 + ], + "angle": 0, + "content": "establish a 2D diffusion model in the UV space. This enhances the fidelity of the generated texture, as depicted in Figure 4 (Bottom)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.553, + 0.345, + 0.57 + ], + "angle": 0, + "content": "4.1. Coarse Stage: Point Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.682 + ], + "angle": 0, + "content": "In the coarse stage, we begin by executing farthest point sampling (FPS) on the mesh surface, deriving a point set consisting of \\(K\\) points. These points are defined by their coordinates \\(\\mathbf{z}_{\\mathrm{coord}}\\) and colors \\(\\mathbf{z}_0\\). During training, a forward diffusion process degrades the clean colors \\(\\mathbf{z}_0\\), transforming them into a noisy state \\(\\mathbf{z}_t\\). The noise level is dictated by the time step \\(t\\), where \\(t \\in \\{0,1,\\dots,T\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.471, + 0.79 + ], + "angle": 0, + "content": "Our network is trained to reverse this diffusion process, aiming to denoise \\(\\mathbf{z}_t\\) back to its original clean colors. This denoising network is informed by three pre-computed maps: a coordinate map \\(\\mathbf{x}_{\\mathrm{coord}}\\), a normal map \\(\\mathbf{x}_{\\mathrm{normal}}\\), and a mask map \\(\\mathbf{x}_{\\mathrm{mask}}\\) (details in Section 3.2). Unless stated otherwise, these conditions concatenate along the channel dimension, culminating in what we term the shape map \\(\\mathbf{x}_{\\mathrm{shape}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.803, + 0.47, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\text {s h a p e}} = \\left(\\left[ \\mathbf {x} _ {\\text {n o r m a l}}, \\mathbf {x} _ {\\text {m a s k}}, \\mathbf {x} _ {\\text {c o o r d}} \\right]\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our network architecture is constructed upon PVCNN [23], drawing similarities to point-voxel diffusion [49]. However, we introduce slight modifications to amplify the integration of global shape information, contrasting with the approach in [49] which primarily relies on point coordinates. Ini" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.498, + 0.893, + 0.56 + ], + "angle": 0, + "content": "tially, we employ a lightweight shape encoder \\( E_{\\phi} \\) to extract a global shape embedding \\( f_{g} \\) from \\( \\mathbf{x}_{\\mathrm{shape}} \\). This embedding is subsequently fed into the 3D network \\( G_{\\theta_1}^1 \\) along with \\( \\mathbf{z}_t \\), \\( \\mathbf{z}_{\\mathrm{coord}} \\), and \\( t \\) to predict the color \\( \\hat{\\mathbf{z}}_0 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.564, + 0.892, + 0.588 + ], + "angle": 0, + "content": "\\[\nf _ {g} = E _ {\\phi} \\left(\\mathbf {x} _ {\\text {s h a p e}}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.584, + 0.792, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {0} = G _ {\\theta_ {1}} ^ {1} \\left(\\left[ \\mathbf {z} _ {\\mathrm {c o o r d}}, \\mathbf {z} _ {t}, f _ {g}, t \\right]\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.893, + 0.844 + ], + "angle": 0, + "content": "Style guidance. We observe that the synthesized colors are largely influenced by the predominant colors within the dataset (for instance, textures in the ShapeNet \"chair\" category are typically white, pure magenta, or wood-colored), leading to a lack of diversity in the outputs. To address this bias and promote color diversity, we introduce a style guidance mechanism. This is achieved by flattening each \\(\\mathbf{z}_0\\) into a unidimensional vector, followed by employing PCA to extract the principal component coefficients, thus reducing the dimensionality of \\(\\mathbf{z}_0\\). Subsequently, K-means clustering is utilized to assign a style label to each \\(\\mathbf{z}_0\\). As depicted in Figure 5, shapes within a particular cluster exhibit similar color styles. During training, we provide the network with an additional style label \\(z_{\\mathrm{style}}\\) as a condition, which is referred to as style guidance:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.848, + 0.892, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {0} = G _ {\\theta_ {1}} ^ {1} \\left(\\left[ \\mathbf {z} _ {\\text {c o o r d}}, \\mathbf {z} _ {t}, f _ {g}, t, z _ {\\text {s t y l e}} \\right]\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "In this way, we can guide the network to predict the desired color during inference by providing a certain style condi" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4209" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.385, + 0.106 + ], + "angle": 0, + "content": "tion, thereby alleviating the biased color issue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.136, + 0.473, + 0.288 + ], + "angle": 0, + "content": "Coarse texture image. With the colored point clouds, we then project them onto the 2D UV space based on the pre-calculated UV mapping. Following this, we perform KNN interpolation based on the 3D coordinates to assign colors to the remaining uncolored pixels, thereby generating a coarse texture image \\(\\mathbf{x}_{\\mathrm{coarse}}\\) for the given mesh. This coarse texture image offers a coherent base color initialization and functions as a conditional element to guide the high-fidelity texture generation in the second stage, helping to avoid discontinuity caused by UV mapping." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.302, + 0.308, + 0.318 + ], + "angle": 0, + "content": "4.2. Fine Stage: UV Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.327, + 0.47, + 0.435 + ], + "angle": 0, + "content": "To further refine the coarse texture image, we design a fine stage using 2D diffusion in the UV space, as depicted in Figure 4 (Bottom). In addition to the conditions used in the coarse stage, we incorporate the coarse texture image \\(\\mathbf{x}_{\\mathrm{coarse}}\\) as an additional condition. We apply a 2D U-Net \\(G_{\\theta_2}^2\\) combined with self-attention modules to learn the high-quality texture image \\(\\hat{\\mathbf{x}}_0\\), from a noisy texture image \\(\\mathbf{x}_t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.448, + 0.469, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {0} = G _ {\\theta_ {2}} ^ {2} \\left(\\left[ \\mathbf {x} _ {\\text {s h a p e}}, \\mathbf {x} _ {t}, \\mathbf {x} _ {\\text {c o a r s e}}, t \\right]\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Hybrid condition. During training, we employ FPS on the ground-truth texture image, \\(\\mathbf{x}_0\\), followed by interpolation to simulate the coarse texture image, \\(\\mathbf{x}_{\\mathrm{coarse}}\\). However, in joint cascaded testing involving both stages, we note that the output quality of \\(\\hat{\\mathbf{x}}_{\\mathrm{coarse}}\\) from the first stage doesn't always align flawlessly with the quality of \\(\\mathbf{x}_{\\mathrm{coarse}}\\). Such mismatches can influence the performance of the subsequent stage. To address this, we introduce a hybrid conditioning method aimed at narrowing this discrepancy. Firstly, we create a smooth texture image \\(\\mathbf{x}_{\\mathrm{smooth}}\\) (see Figure 6), inspired by the blur augmentation described in [17] for cascaded diffusion models. In particular, we segment the mask map into multiple discrete regions using four-connectivity detection, then perform average color pooling within each region based on its connectivity. After this procedure, \\(\\mathbf{x}_{\\mathrm{smooth}}\\) maintains merely the regional color, ensuring more consistent alignment across both training and testing phases. Then, we combine \\(\\mathbf{x}_{\\mathrm{coarse}}\\) with \\(\\mathbf{x}_{\\mathrm{smooth}}\\), using a certain probability \\(p_{\\mathrm{hybrid}}\\) during training. Thus, the network is forced to be capable of generating textures even when adopting a weaker condition \\(\\mathbf{x}_{\\mathrm{smooth}}\\) in the fine stage during inference. Considering that the diffusion model first generates low-frequency information and then higher one [8], we also explore a condition-truncated sampling, as detailed discussed in Section 5.4, where we condition both maps for generation during the initial time steps and then exclusively utilize the smooth map for the remainder of the generation." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.093, + 0.67, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.219, + 0.628, + 0.229 + ], + "angle": 0, + "content": "(a) Samples from cluster 1" + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.093, + 0.885, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.219, + 0.822, + 0.229 + ], + "angle": 0, + "content": "(d) Samples from cluster 4" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.233, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Figure 5. Illustration of samples across various clusters. Shapes within a particular cluster exhibit analogous color styles. However, there exists an imbalance in the quantity of shapes among different clusters, leading to challenges in unbiased synthesis." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.304, + 0.628, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.399, + 0.609, + 0.41 + ], + "angle": 0, + "content": "(a) Texture image" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.304, + 0.753, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.399, + 0.751, + 0.41 + ], + "angle": 0, + "content": "(b) Coarse texture image" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.304, + 0.884, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.399, + 0.882, + 0.41 + ], + "angle": 0, + "content": "(c) Smooth texture image" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.415, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Figure 6. Varieties of texture images. (a) Original texture image enriched with high-frequency details, (b) Coarse texture image, and (c) Smooth texture image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.633, + 0.499 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.507, + 0.819, + 0.524 + ], + "angle": 0, + "content": "5.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.85 + ], + "angle": 0, + "content": "We conduct experiments across four categories of the ShapeNet dataset [4], namely, chair, table, car, and bench. Before training, we use an open-source UV-Atlas tool [46] to generate the UV map and pre-process the dataset to obtain the shape maps and ground-truth texture images. We sample \\( K = 4096 \\) points in our coarse stage and synthesize a texture image in the UV space with a resolution of \\( 512 \\times 512 \\) for the fine stage. For training the diffusion models, akin to [1], we predict the clean signals. This approach provides more stable training than predicting the noise component as recommended by [16]. We employ the cosine noise scheduling [27] ranging from 0.0001 to 0.02 over 1,024 time steps for both stages. In the fine stage, we leverage the noise scaling strategy from [7] and also incorporate a rendering loss. Specifically, we randomly select four views and render the mesh using the predicted UV map and the ground-truth UV map to produce \\( 1024 \\times 1024 \\) images. Subsequently, we crop \\( 224 \\times 224 \\) patches from these images and compute the corresponding \\( L_{1} \\) loss. For the hybrid condition in the fine stage, we use \\( p_{\\mathrm{hybrid}} = 0.3 \\) and sweep over condition-truncated time \\( t_{c} \\) (see Figure 12)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Compared methods. We compare our method with state-of-the-art approaches, including Texturify [38], Texture" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4210" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.096, + 0.868, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.362, + 0.895, + 0.392 + ], + "angle": 0, + "content": "Figure 7. Our texture synthesis results in different categories. The results on chairs, tables, cars, and benches demonstrate that our approach can generate natural, vivid, and diverse textures, even though the given 3D shapes are challenging, such as cutouts on the bench." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.473, + 0.538 + ], + "angle": 0, + "content": "Fields [29], and PVD [49], in the context of unconditional texture generation. PVD was originally designed for point cloud generation instead of texture generation, so it cannot be directly compared with our approach. To facilitate comparison, we modify this framework to learn RGB values from the sampled point cloud. This extension, referred to as \"PVD-Tex\", serves as a baseline for directly learning texture in the point space using the diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.547, + 0.379, + 0.563 + ], + "angle": 0, + "content": "5.2. Unconditional Texture Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.471, + 0.678 + ], + "angle": 0, + "content": "Gallary. First, we showcase our texture generation results for each category without additional conditions. The results in Figure 7 and Figure 1 demonstrate the remarkable performance of our method, which is able to generate appealing textures with fine details and preserve the geometric structures. Note that our method is compatible with meshes of diverse topology and intricate geometric details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.47, + 0.851 + ], + "angle": 0, + "content": "Qualitative comparisons. As depicted in Figure 8, our approach excels at generating high-quality textures while maintaining the geometric intricacies of the input mesh. Firstly, the results of Texture Fields [29] appear deficient in high-frequency details. Additionally, while PVD-Tex [49] is able to produce spatially varying colors to some extent (e.g., the car in Figure 8), it falls short in synthesizing intricate high-frequency details. Lastly, even though Texturify [38] demonstrates an ability to generate finer textures, it compromises on preserving slender structures (i.e. chair)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Quantitative comparisons. To quantitatively compare with existing works, we follow [38] and assess the genera" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.895, + 0.538 + ], + "angle": 0, + "content": "tion quality using Frechet Inception Distance (FID)[15] and Kernel Inception Distance (KID)[2], metrics that are widely used for evaluating image generation models. To this end, we render \\( 512 \\times 512 \\) images from each generated textured mesh and ground-truth textured mesh using four distinct camera views. Table 1 presents the quantitative comparisons with current approaches, revealing that our method surpasses existing works." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.546, + 0.783, + 0.561 + ], + "angle": 0, + "content": "5.3. Conditional Texture Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In addition, we demonstrate the capability of our framework to synthesize textures conditioned on either text prompts or a single-view image. We conduct our experiments on the chair and table categories. For the text condition, we utilize text-shape pairs as provided in [6] (with additional corrections for text accuracy). For the image condition, we randomly render a view from the ground-truth mesh. To infuse the network with condition-specific information, we use the pre-trained vision-language model CLIP [33] to extract the corresponding embedding from either the image or the text. This embedding is then fed into a simple MLP to incorporate the information into the diffusion model as a condition for both training and inference. As depicted in Figure 9, our method succeeds in generating textures that align well with the given text descriptions or images. We also compare our approach with Text2Mesh [24], a test-time optimization method. Text2Mesh takes around 10 minutes per instance, while ours only requires 30 seconds. Importantly, to adapt to our task which requires preserving the mesh geometry, we freeze the geometry deformation branch of Text2Mesh to generate only colors. As shown in Figure 10, Text2Mesh" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.515, + 0.938 + ], + "angle": 0, + "content": "4211" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.098, + 0.462, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.452, + 0.472, + 0.537 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparisons with existing works. Given pure 3D shapes, Texturify [38] generates textures on the surface of 3D shape but tends to damage the topology; Texture fields [29] and PVD-Tex [49] produce textures with limited details. On the contrary, our approach faithfully preserves topology and produces realistic appearances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.47, + 0.607 + ], + "angle": 0, + "content": "is bound by vertex resolution, while our method can deliver detailed visuals even on low-resolution meshes, offering a distinct advantage in graphics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.616, + 0.242, + 0.631 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.837 + ], + "angle": 0, + "content": "Coarse-to-fine diffusion. To manifest the effectiveness of our coarse-to-fine diffusion strategy, we conduct experiments on two baselines \"w/o coarse stage\" and \"w/o fine stage\". The \"w/o coarse stage\" configuration indicates that we directly generate the UV texture image using the fine stage, bypassing the initialization from the coarse stage. The \"w/o fine stage\" configuration signifies the result of the coarse stage, after assigning colors to the uncolored pixels using KNN interpolation. In both scenarios, the model produces inferior outcomes relative to our full model, as shown in Table 2 and Figure 11. Absent the coarse stage, the generated result suffers from noticeably inconsistent colors. Without the fine stage, the results are over-smoothed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Hybrid condition. As shown in Table 2 and Figure 11 \"coarse condition\", the fine stage cannot generate high-quality textures when it is exclusively conditioned on the" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.091, + 0.889, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.532, + 0.892, + 0.575 + ], + "angle": 0, + "content": "Figure 9. Results of conditional texture generation. Our method is adaptable to craft textures guided by text descriptions (rows 1-3) or single-view images (rows 4-6)." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.587, + 0.868, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.728, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Figure 10. Comparison with Text2Mesh. We freeze the geometry deformation branch of Text2Mesh to adapt to our task. Ours can generate high-frequency details for low-resolution mesh." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.903 + ], + "angle": 0, + "content": "coarse map. Further, the generated quality remains unsatisfactory if we apply a hybrid condition entirely during training (i.e., \\( p_{\\mathrm{hybrid}} = 1.0 \\)), as shown in \" \\( p_{\\mathrm{hybrid}} = 1.0 \\)\". We attribute this to the network's propensity to depend solely on the coarse map during training and generation. In contrast, our full model (i.e., hybrid condition with \\( p_{\\mathrm{hybrid}} = 0.3 \\)) achieves significant improvement both" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4212" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.171, + 0.089, + 0.803, + 0.184 + ], + "angle": 0, + "content": "
ChairCarTableBench
MethodsFID ↓KID ↓FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓
Texture Fields [29]24.241.07156.3813.6468.964.2062.712.96
Texturify [38]27.801.3273.164.71----
PVD-Tex [49]15.520.6259.473.7416.120.5528.940.39
Ours9.880.2226.890.689.630.1523.090.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.184, + 0.867, + 0.199 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with existing works. Ours outperforms other approaches on both FID [15] and KID \\((\\times 10^{2})\\) [2]." + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.226, + 0.208, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.338, + 0.225, + 0.349 + ], + "angle": 0, + "content": "(a) untextured mesh" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.226, + 0.317, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.337, + 0.34, + 0.349 + ], + "angle": 0, + "content": "(b) w/o coarse stage" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.226, + 0.426, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.337, + 0.437, + 0.349 + ], + "angle": 0, + "content": "(c) w/o fine stage" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.352, + 0.209, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.465, + 0.225, + 0.477 + ], + "angle": 0, + "content": "(d) coarse condition" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.353, + 0.321, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.465, + 0.333, + 0.477 + ], + "angle": 0, + "content": "(e) \\(p_{hybrid} = 1.0\\)" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.353, + 0.427, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.466, + 0.411, + 0.477 + ], + "angle": 0, + "content": "(f) ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.525 + ], + "angle": 0, + "content": "Figure 11. Ablation studies. Without our two-stage generation pipeline or hybrid condition designs, the generated results are inconsistent with the 3D shape or lack of high-frequency details." + }, + { + "type": "table", + "bbox": [ + 0.147, + 0.536, + 0.398, + 0.631 + ], + "angle": 0, + "content": "
FID ↓KID ↓
w/o fine stage17.880.76
w/o coarse stage15.110.49
coarse condition14.930.56
\\(p_{hybrid}=1.0\\)15.250.59
Ours9.880.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.632, + 0.47, + 0.66 + ], + "angle": 0, + "content": "Table 2. Ablation studies. This table shows the effectiveness of each component in our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.471, + 0.822 + ], + "angle": 0, + "content": "qualitatively and quantitatively, as presented in Table 2 and Figure 11 \"ours\". Furthermore, we also sweep over the effect of condition-truncated time \\( t_c \\) for inference. As Figure 12 shows, \\( t_c = 0.4 \\) strikes a sweet point. The model utilizes information from both conditions maps to better generate low-frequency components during the first \\( 40\\% \\) of the sampling timesteps. After that, the coarse map \\( \\mathbf{x}_{\\mathrm{coarse}} \\) is dropped, and the model further focuses on fine-grained detail generation without relying on the coarse map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Style guidance. The style guidance is aimed at addressing the issue of insufficient diversity due to color distribution bias in the dataset. As a result, we do not utilize FID or KID for evaluation since they evaluate the distribution sim" + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.22, + 0.853, + 0.299 + ], + "angle": 0, + "content": "
MethodsPreference↑LPIPS↑
Ours49.8%0.083
Texturify [38]15.9%0.086
PVD-Tex [49]29.2%0.029
Texture Fields [29]5.1%0.005
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.3, + 0.892, + 0.328 + ], + "angle": 0, + "content": "Table 3. LPIPS and user study. This table shows the average LPIPS and preference via user study in the chair category." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.892, + 0.673 + ], + "angle": 0, + "content": "The diversity of the LPIPS is a key feature that is crucial to the quality of the generated results and ground truth. However, the ground-truth distribution shows a strong bias towards particular colors, making these metrics inappropriate for assessing generative texture diversity. In contrast, we adopt LPIPS [48] to measure the pairwise similarity among five textures generated by our method given the same input mesh, where a larger diversity in textures will lead to a higher LPIPS value. In Figure 13, we report the quantitative results for 500 shapes. Our approach achieves a higher LPIPS in most of the evaluated cases, indicating better diversity. Besides, as shown in Figure 14, we present the results of generating three textures randomly for the shapes without and with style guidance. In the latter case, we uniformly sample three style labels as style inputs. It is evident that without style guidance, the generated textures are nearly identical, and the color styles of different shapes tend to be similar. With the introduction of style guidance, however, the diversity of colors has significantly increased. We also conduct a comparison with other methods in terms of diversity, as well as a quality assessment through user studies, as shown in Table 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.691, + 0.868, + 0.708 + ], + "angle": 0, + "content": "6. Discussion, Limitations, and Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.869 + ], + "angle": 0, + "content": "This paper presents Point-UV diffusion, a brand-new framework that employs a coarse-to-fine pipeline to generate textures for 3D meshes. We begin with a 3D diffusion model to synthesize low-frequency texture components from point clouds, which maintain 3D consistency. We then refine the textures using a 2D UV-space diffusion model. Our method is compatible with meshes with arbitrary topology and can faithfully preserve the geometry structure. We further demonstrate the flexibility of our framework by extending it to conditional generative models." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Despite its merits, our method has inherent limitations. Similar to other methodologies relying on 3D data for train-" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4213" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.095, + 0.443, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.471, + 0.318 + ], + "angle": 0, + "content": "Figure 12. Examination of condition-truncated time. The horizontal axis denotes the percentage of time points when the coarse texture image is conditioned during inference. The left vertical axis indicates the FID value, while the right vertical axis shows the KID value \\((\\times 10^{2})\\)." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.33, + 0.446, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.566 + ], + "angle": 0, + "content": "Figure 13. Diversity measurement. Utilizing style guidance leads to larger LPIPS scores, indicating enhanced generation diversity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.47, + 0.803 + ], + "angle": 0, + "content": "ing, our technique is upper-bounded by the scope and diversity of current 3D datasets. This restriction poses challenges in generating textures that parallel the depth of effects seen in 2D image synthesis. Moreover, our method's efficacy relies on the quality of UV mapping. Our approach faces difficulties in rendering high-quality results for meshes where the UV mapping produces excessive fragmented cuts, resulting in fragmented artifacts. This phenomenon is commonly observed in the car category, as shown in the Figure 15. We believe the emergence of larger and more diverse 3D datasets would be helpful for generating superior-quality textures. Further advancements in UV parameterization would also be beneficial in augmenting our method's consistency." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.816, + 0.265, + 0.833 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "This work has been supported by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant No. 17202422), and RGC matching fund scheme (RMGS). Part" + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.09, + 0.689, + 0.105 + ], + "angle": 0, + "content": "(a) w/o style guidance" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.112, + 0.638, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.314, + 0.614, + 0.327 + ], + "angle": 0, + "content": "Result 1" + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.112, + 0.741, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.314, + 0.718, + 0.327 + ], + "angle": 0, + "content": "Result 2" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.112, + 0.847, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.314, + 0.827, + 0.327 + ], + "angle": 0, + "content": "Result 3" + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.332, + 0.681, + 0.348 + ], + "angle": 0, + "content": "(b) w/ style guidance" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.356, + 0.63, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.557, + 0.614, + 0.57 + ], + "angle": 0, + "content": "Result 1" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.356, + 0.736, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.557, + 0.718, + 0.57 + ], + "angle": 0, + "content": "Result 2" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.356, + 0.842, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.557, + 0.827, + 0.57 + ], + "angle": 0, + "content": "Result 3" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.578, + 0.892, + 0.634 + ], + "angle": 0, + "content": "Figure 14. Efficacy of style guidance. The model without style guidance produces consistent colors across different random seeds (rows 1-2). In contrast, with style guidance, diverse outcomes emerge (rows 3-4)." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.666, + 0.671, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.769, + 0.664, + 0.78 + ], + "angle": 0, + "content": "(a) Generated Texture Image" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.686, + 0.862, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.769, + 0.822, + 0.779 + ], + "angle": 0, + "content": "(b) Textured Mesh" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.787, + 0.892, + 0.828 + ], + "angle": 0, + "content": "Figure 15. Failure case. Our approach still struggles to generate seamless results when there are too many fragmented cuts of the UV map." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4214" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. arXiv preprint arXiv:2211.09869, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[3] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.38 + ], + "angle": 0, + "content": "[4] Angel X. Chang, Thomas Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.452 + ], + "angle": 0, + "content": "[5] Bindita Chaudhuri, Nikolaos Sarafianos, Linda Shapiro, and Tony Tung. Semi-supervised synthesis of high-resolution editable textures for 3d humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7991-8000, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.549 + ], + "angle": 0, + "content": "[6] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III 14, pages 100-116. Springer, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.47, + 0.578 + ], + "angle": 0, + "content": "[7] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.472, + 0.648 + ], + "angle": 0, + "content": "[8] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[9] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10673-10683, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[10] Jun Gao, Wenzheng Chen, Tommy Xiang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. Advances In Neural Information Processing Systems, 33:9936-9947, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[11] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Lin Gao, Tong Wu, Yu-Jie Yuan, Ming-Xian Lin, Yu-Kun Lai, and Hao Zhang. Tm-net: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG), 40(6):1-15, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.148 + ], + "angle": 0, + "content": "[13] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. In Advances in neural information processing systems, pages 5767-5777, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[14] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Learning a neural 3d texture space from 2d exemplars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8356-8364, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.893, + 0.261 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. NIPS, 2017. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.262, + 0.893, + 0.29 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.893, + 0.347 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23(47):1-33, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.893, + 0.404 + ], + "angle": 0, + "content": "[18] Jingyu Hu, Ka-Hei Hui, Zhengzhe Liu, Ruihui Li, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation, inversion, and manipulation. arXiv preprint arXiv:2302.00190, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.893, + 0.46 + ], + "angle": 0, + "content": "[19] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation. In SIGGRAPH Asia 2022 Conference Papers, pages 1–9, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.893, + 0.518 + ], + "angle": 0, + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.893, + 0.603 + ], + "angle": 0, + "content": "[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.893, + 0.659 + ], + "angle": 0, + "content": "[22] Zhengzhe Liu, Peng Dai, Ruihui Li, Xiaojuan Qi, and Chi-Wing Fu. Iss: Image as stetting stone for text-guided 3d shape generation. International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.893, + 0.703 + ], + "angle": 0, + "content": "[23] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Pointvoxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems, 32, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.893, + 0.772 + ], + "angle": 0, + "content": "[24] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.893, + 0.843 + ], + "angle": 0, + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[26] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.925, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4215" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "[27] Alex Nichol and Prafulla Dhariwal. Improved Denoising Diffusion Probabilistic Models, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[28] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[29] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[30] Dario Pavllo, Graham Spinks, Thomas Hofmann, Marie-Francine Moens, and Aurelien Lucchi. Convolutional generation of textured 3d meshes. Advances in Neural Information Processing Systems, 33:870-882, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.348 + ], + "angle": 0, + "content": "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[32] Tiziano Portenier, Siavash Arjomand Bigdeli, and Orcun Goksel. Gramgan: Deep 3d texture synthesis from 2d exemplars. Advances in Neural Information Processing Systems, 33:6994-7004, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.47, + 0.603 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[37] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[38] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. arXiv preprint arXiv:2204.02411, 2022. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[39] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. Advances in Neural Information Processing Systems, 35:24487-24501, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Confer" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "ence on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[41] Yongbin Sun, Ziwei Liu, Yue Wang, and Sanjay E Sarma. Im2avatar: Colorful 3d reconstruction from a single image. arXiv preprint arXiv:1804.06375, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[42] Shubham Tulsiani, Nilesh Kulkarni, and Abhinav Gupta. Implicit mesh reconstruction from unannotated image collections. arXiv preprint arXiv:2007.08504, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[43] Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2626-2634, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[44] Hao Wang, Guosheng Lin, Steven Hoi, and Chunyan Miao. Cycle-consistent inverse GAN for text-to-image synthesis. ACM MM, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[45] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12456-12465, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[46] Jonathan Young. xatlas, 2018. https://github.com/jpcy/xatlas.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[47] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[49] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 3, 4, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.926, + 0.516, + 0.937 + ], + "angle": 0, + "content": "4216" + } + ] +] \ No newline at end of file diff --git a/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/295801ea-f88d-446b-aeea-24167c36963d_origin.pdf b/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/295801ea-f88d-446b-aeea-24167c36963d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5a24e5d4d062b8037d8e03c2d08e1ed52a09a4a7 --- /dev/null +++ b/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/295801ea-f88d-446b-aeea-24167c36963d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b2eb95b927cd59090744dc33a3e2023759f60f33d9a3ac6b5c25963edac6dbf +size 5154671 diff --git a/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/full.md b/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e8bce9fc893c730d8f3fa0ef698519a0b5767d1c --- /dev/null +++ b/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/full.md @@ -0,0 +1,362 @@ +# Texture Generation on 3D Meshes with Point-UV Diffusion + +Xin Yu + +Peng Dai + +Wenbo Li2 + +Lan Ma3 + +Zhengzhe Liu $^{2\dagger}$ + +Xiaojuan Qi $^{1\dagger}$ + +1The University of Hong Kong, 2The Chinese University of Hong Kong, 3TCL Corporate Research + +{yuxin, daipeng, xjqi}@eee.hku.hk + +{wenboli,zzliu}@cse.cuhk.edu.hk + +rubyma@tcl.com + +![](images/b22925ff733e95162010d693f891239383078d8f8361fa87400e6d9beb215271.jpg) +Figure 1. A gallery of generated results by our Point-UV diffusion. Our method is capable of processing meshes of any genus, generating diversified, geometry-compatible, and high-fidelity textures. + +# Abstract + +In this work, we focus on synthesizing high-quality textures on 3D meshes. We present Point-UV diffusion, a coarse-to-fine pipeline that marries the denoising diffusion model with UV mapping to generate 3D consistent and high-quality texture images in UV space. We start with introducing a point diffusion model to synthesize low-frequency texture components with our tailored style guidance to tackle the biased color distribution. The derived coarse texture offers global consistency and serves as a condition for the subsequent UV diffusion stage, aiding in regularizing the model to generate a 3D consistent UV texture image. Then, a UV diffusion model with hybrid conditions is developed to enhance the texture fidelity in the 2D UV space. Our method can process meshes of any genus, generating diversified, geometry-compatible, and high-fidelity textures. Code is available at https://cvmilab.github.io/Point-UV-Diffusion. + +# 1. Introduction + +Texturing 3D meshes is a fundamental task in computer vision and graphics. It enhances the visual richness of 3D objects, thereby facilitating their application in various fields such as video games, 3D movies, and AR/VR technologies. However, generating high-quality textures can be daunting and time-consuming, often requiring specialized knowledge and resources. As such, there is a pressing need for an efficient approach to automatically create high-quality textures on 3D meshes. + +Despite the substantial progress the community has made in 2D image synthesis and 3D shape generation using GANs [13, 20, 44, 11] or diffusion models [34, 36, 35, 19], crafting realistic textures on mesh surfaces remains challenging. One major difficulty stems from the need for suitable 3D representations for texture synthesis. Early approaches investigate the use of voxels [5, 45, 6] or point clouds [10] and synthesize point/voxel colors. However, they can only afford to synthesize low-resolution results with low-fidelity textures due to memory and model complexity constraints. In response, Texture Fields [29] adopts an implicit representation with the potential to synthesize + +high-resolution textures, but still could not yield satisfactory results as shown in Figure 2 (a): over-smoothed results. Most recently, Siddiqui et al. [38] propose to parameterize the shape as tetrahedral meshes and introduce tetrahedral mesh convolution to enhance local details. Albeit improving results, tetrahedral parameterization inevitably destroys geometric details of the input mesh and thus cannot faithfully preserve the original structure. As shown in Figure 2 (b), the delicate structures of the chair's back are absent. Moreover, the generative models utilized in these methods are limited to GANs [29, 38] and VAEs [29, 12]. The more advanced diffusion model, which could potentially open up new avenues for high-quality texture generation, remains insufficiently explored. + +In this paper, we delve into a novel texture representation based on UV maps and investigate the advanced diffusion model for texture generation. The 2D nature of the UV map enables it to circumvent the cost of high-resolution point/voxel representations. Besides, the UV map is compatible with arbitrary mesh topologies, thereby preserving the original geometric structures. However, while promising, direct integration of the UV map representation with a 2D diffusion model presents challenges in synthesizing seamless textures, leading to severe artifacts, as shown in Figure 2 (c). This occurs because the UV mapping process fragments the continuous texture on the 3D surface into isolated patches on the 2D UV plane (see Figure 3). + +To this end, we introduce Point-UV diffusion, a two-stage coarse-to-fine framework consisting of point diffusion and UV diffusion. Specifically, we initially design a point diffusion model to generate color for sampled points that act as low-frequency texture components. This model is equipped with a style guidance mechanism that alleviates the impact of biased color distributions in the dataset and facilitates diversity during inference. Next, we project these colorized points onto the 2D UV space with 3D coordinate interpolation, thereby generating a coarse texture image that maintains 3D consistency and continuity. Given the coarse textured image, we develop a UV diffusion model with elaborately designed hybrid conditions to improve the quality of the textures (see Figure 2 (ours) and Figure 1). + +In short, our contributions are as follows: 1) We propose a new framework for texture generation for given meshes. Our representation can handle meshes with arbitrary topology and is able to faithfully preserve geometric structures. 2) To the best of our knowledge, we are the first to train a diffusion model specifically for mesh texture generation. Our coarse-to-fine framework allows us to enjoy the efficiency of 2D representation while enhancing 3D consistency. 3) We compare our approach with multiple methods in unconditional generation and achieve state-of-the-art results. Furthermore, we demonstrate that our method can be easily extended to scenarios with text-conditioning and + +![](images/9892278130303e08163bccc166ec305808fbec62493a36a1e330353c20108282.jpg) +(a) Texture Fields vs Ours + +![](images/c61bc7531d7ce27c511cb3be73337055d95bf097cd97b0bf8611eed49c31584c.jpg) +(b) Texturify vs Ours + +![](images/55a88ce9b2911b3d4ea4a7da214f77dd9b1cc1e98ed13e135ca289646d74c39d.jpg) +(c) 2D diffusion vs Ours + +![](images/30221e74c987260583731d2865db7cc1d63613dff5bb7263268b952ee83092c3.jpg) +Figure 2. Comparisons with different methods. Our generative results (a) possess high-quality details, (b) faithfully preserve the mesh structure, and (c) are better consistent with the given shape, compared with Texture Fields [29], Texturify [38] and 2D diffusion [16], respectively. +Textured Mesh +Figure 3. Illustration of UV mapping process. It establishes connections between the 2D texture map and the surface appearance of 3D shape. + +![](images/8da5aa7d15164e4c4214b2694020aeb680d8c4b36c6a7df4785e57607da9f6ee.jpg) +Texture Image + +image-conditioning. + +# 2. Related Work + +Texture generation. Early works [41, 43, 6] propose using voxels to represent colors. However, due to a cubic increase in memory usage and computational cost with rising resolution, these methods produce only coarse, low-resolution textures. For meshes, recent work in texture generation [38] predicts the color for each face, but the representation capacity remains largely constrained by the mesh's resolution. Another set of works [30, 42, 26, 12] leverage UV mapping, but they require either spherical parameterization or part segmentation and, therefore, are confined to handling low-genus shapes. Other works [32, 14] rely on image exemplars, while our approach focuses on unconditional generation. Recently, implicit functions attract increasing attention for generative tasks [11, 31, 39, 3, 9, 37, 29, 22]. Most of this research centers on generating 3D-aware images [39, 3, 9, 37], rather than synthesizing textures for given 3D meshes. Texture Fields [29] represents + +the most similar work to our task, which, however, tends to produce over-smoothed results. + +Some methods focus on test-time optimization for generation tasks. For instance, Text2Mesh [24] employs CLIP [33] to design a loss function. This function stylizes a 3D mesh to align with a target text prompt, predicting both the color and displacement for each mesh vertex. More recently, DreamFusion [31] and Magic3D [21] leverage pretrained stable diffusion [35] for score distillation, creating either NeRF [25] or a 3D mesh. However, these approaches often demand extensive optimization time or necessitate alterations to the existing mesh geometry [24]. In this paper, our objective is to generate textures for specified full-3D meshes with arbitrary topology. This goal entails two fundamental requirements: 1) preserving the original mesh's geometric structure; 2) producing a 3D texture representation exportable as either vertex color or a uv-texture-image, instead of merely generating-rendering multi-view or 3D-aware images. + +Diffusion models for 3D generation. In addition to 2D image generation, diffusion models recently gain significant attention in 3D generation, leading to many related works [49, 18, 19, 28, 31, 47, 1]. For instance, Zhou et al. [49] introduce point-voxel diffusion for point cloud generation and completion. Hui et al. [19] and Hu et al. [18] suggest a compact wavelet domain representation for shapes, enabling higher-quality shape creation via diffusion models. Nichol et al. [28] present Point-E, a system that uses a text-conditional strategy to produce colored 3D point clouds. This approach enables an efficient synthesis of intricate 3D shapes from textual prompts. Yet, this system yields lower-resolution point clouds, often missing detailed shapes and textures. In this work, we develop a brand new 3D diffusion model for texture image synthesis, which allows high-fidelity and 3D-consistent texture generation. + +# 3. Preliminaries + +# 3.1. Denoising diffusion model. + +Diffusion model [40, 16] is a kind of likelihood-based generative model that has gained significant attention recently. It learns the data distribution $q(x_0)$ by progressive denoising from a prior Gaussian distribution. Given a sample from the data distribution $x_0 \sim q(x_0)$ , a fixed forward process $q(\mathbf{x}_{1:T} \mid \mathbf{x}_0) = \prod_{t=1}^{T} q(\mathbf{x}_t \mid \mathbf{x}_{t-1})$ is used to perturb the data with Gaussian kernels $q(\mathbf{x}_t \mid \mathbf{x}_{t-1}) \coloneqq \mathcal{N}\left(\sqrt{1 - \beta_t} \mathbf{x}_{t-1}, \beta_t \mathbf{I}\right)$ , producing increasingly noisy latent variables $\{x_1, x_2, \ldots, x_T\}$ . Then, a parameterized Markov process $p_\theta(\mathbf{x}_{0:T}) = p(\mathbf{x}_T) \prod_{t=1}^{T} p_\theta(\mathbf{x}_{t-1} \mid \mathbf{x}_t)$ with transition kernel $p_\theta(\mathbf{x}_{t-1} \mid \mathbf{x}_t) \coloneqq \mathcal{N}\left(\mu_\theta(\mathbf{x}_t, t), \sigma_t^2 \mathbf{I}\right)$ is optimized through maximizing a variational lower bound of log data likelihood, which essentially targets to match the joint + +distribution $q(\pmb{x}_{0:T})$ : + +$$ +\operatorname {E} _ {q \left(\mathbf {x} _ {0}\right)} \left[ \log p _ {\theta} \left(\mathbf {x} _ {0}\right) \right] \geq \operatorname {E} _ {q \left(\mathbf {x} _ {0: T}\right)} \left[ \log \frac {p _ {\theta} \left(\mathbf {x} _ {0 : T}\right)}{q \left(\mathbf {x} _ {1 : T} \mid \mathbf {x} _ {0}\right)} \right]. \tag {1} +$$ + +After training, novel samples can then be generated via iterative sampling from $p_{\theta}(\boldsymbol{x}_{t-1} \mid \boldsymbol{x}_t)$ following: + +$$ +\mathbf {x} _ {t - 1} = \mu_ {\theta} (\mathbf {x} _ {t}, t) + \sqrt {\beta_ {t}} \mathbf {z}, \mathbf {z} \sim \mathcal {N} (0, \mathbf {I}). +$$ + +Diffusion models have shown impressive capabilities in generating high-quality and diverse content. This paper proposes a texture generation framework based on diffusion model. + +# 3.2. UV mapping and challenges. + +UV mapping is a method of surface parameterization that translates a 3D surface into a 2D image, effectively creating a 2D coordinate system known as a UV map for a polygonal mesh $S$ . This is achieved by explicitly assigning UV coordinates to each vertex of the mesh. Further, an arbitrary surface coordinate on the mesh can be mapped to its 2D coordinate through barycentric interpolation: + +$$ +(u, v) = f (p) \quad f: S \rightarrow \Omega , \tag {2} +$$ + +where $f$ represents the UV mapping process. + +As illustrated in Figure 3 "UV warping", we can apply textures to the mesh surface by generating a high-resolution texture image in the UV space. Besides texture, we can also create this kind of 2D map for surface normals and point coordinates. Our objective is to generate a high-quality 2D UV texture image of the given mesh. However, the mapping process requires cutting the continuous texture on the 3D shape into a series of individual patches in the 2D UV plane, as depicted in Figure 3. This fragmentation makes it challenging for the generative model to directly learn the 3D adjacency relationships of the patches within the 2D texture UV map. Consequently, this can lead to discontinuity and inconsistency issues when the generated texture map is applied back to the 3D mesh surface. As shown in Figure 2 "2D Diffusion", the diffusion model generates inconsistent textures and suffers from discontinuity issues. + +# 4. Method + +The challenge mentioned in Section 3.2 motivates us to propose a coarse-to-fine framework for texture image synthesis, namely Point-UV diffusion, illustrated in Figure 4. To start, we design a 3D point diffusion model to colorize a set of sampled points on the mesh surface, as shown in Figure 4 (Top). This stage leverages the 3D topology for predicting low-frequency colors on the mesh, without being affected by the discontinuity of the UV map. Based on the color components generated by the coarse stage, we then + +![](images/03bdf3a5f17d771afd0ee6fb532bea7a0253e87315c91a4981dc2a4aed1ff5a1.jpg) + +![](images/1df90fc3e88730a3a8b4a0cc36e39f51d1602f2414a9d3442ad1d525fe477bcc.jpg) +Figure 4. The overview of our Point-UV-diffusion framework. (Top) The coarse stage first samples a point cloud on the mesh surface, and then predicts the color for each point using a 3D diffusion model conditioned on shape features including surface normal, mask, and coordinates. Then the points are mapped to the 2D UV map and the remaining uncolored points are filled up by tri-linear interpolation based on the 3D coordinates. (Bottom) The fine stage predicts high-quality textures with a 2D diffusion model conditioned on the shape attributes and the coarse texture image. + +establish a 2D diffusion model in the UV space. This enhances the fidelity of the generated texture, as depicted in Figure 4 (Bottom). + +# 4.1. Coarse Stage: Point Diffusion + +In the coarse stage, we begin by executing farthest point sampling (FPS) on the mesh surface, deriving a point set consisting of $K$ points. These points are defined by their coordinates $\mathbf{z}_{\mathrm{coord}}$ and colors $\mathbf{z}_0$ . During training, a forward diffusion process degrades the clean colors $\mathbf{z}_0$ , transforming them into a noisy state $\mathbf{z}_t$ . The noise level is dictated by the time step $t$ , where $t \in \{0,1,\dots,T\}$ . + +Our network is trained to reverse this diffusion process, aiming to denoise $\mathbf{z}_t$ back to its original clean colors. This denoising network is informed by three pre-computed maps: a coordinate map $\mathbf{x}_{\mathrm{coord}}$ , a normal map $\mathbf{x}_{\mathrm{normal}}$ , and a mask map $\mathbf{x}_{\mathrm{mask}}$ (details in Section 3.2). Unless stated otherwise, these conditions concatenate along the channel dimension, culminating in what we term the shape map $\mathbf{x}_{\mathrm{shape}}$ : + +$$ +\mathbf {x} _ {\text {s h a p e}} = \left(\left[ \mathbf {x} _ {\text {n o r m a l}}, \mathbf {x} _ {\text {m a s k}}, \mathbf {x} _ {\text {c o o r d}} \right]\right). \tag {3} +$$ + +Our network architecture is constructed upon PVCNN [23], drawing similarities to point-voxel diffusion [49]. However, we introduce slight modifications to amplify the integration of global shape information, contrasting with the approach in [49] which primarily relies on point coordinates. Ini + +tially, we employ a lightweight shape encoder $E_{\phi}$ to extract a global shape embedding $f_{g}$ from $\mathbf{x}_{\mathrm{shape}}$ . This embedding is subsequently fed into the 3D network $G_{\theta_1}^1$ along with $\mathbf{z}_t$ , $\mathbf{z}_{\mathrm{coord}}$ , and $t$ to predict the color $\hat{\mathbf{z}}_0$ : + +$$ +f _ {g} = E _ {\phi} \left(\mathbf {x} _ {\text {s h a p e}}\right), \tag {4} +$$ + +$$ +\hat {\mathbf {z}} _ {0} = G _ {\theta_ {1}} ^ {1} \left(\left[ \mathbf {z} _ {\mathrm {c o o r d}}, \mathbf {z} _ {t}, f _ {g}, t \right]\right). +$$ + +Style guidance. We observe that the synthesized colors are largely influenced by the predominant colors within the dataset (for instance, textures in the ShapeNet "chair" category are typically white, pure magenta, or wood-colored), leading to a lack of diversity in the outputs. To address this bias and promote color diversity, we introduce a style guidance mechanism. This is achieved by flattening each $\mathbf{z}_0$ into a unidimensional vector, followed by employing PCA to extract the principal component coefficients, thus reducing the dimensionality of $\mathbf{z}_0$ . Subsequently, K-means clustering is utilized to assign a style label to each $\mathbf{z}_0$ . As depicted in Figure 5, shapes within a particular cluster exhibit similar color styles. During training, we provide the network with an additional style label $z_{\mathrm{style}}$ as a condition, which is referred to as style guidance: + +$$ +\hat {\mathbf {z}} _ {0} = G _ {\theta_ {1}} ^ {1} \left(\left[ \mathbf {z} _ {\text {c o o r d}}, \mathbf {z} _ {t}, f _ {g}, t, z _ {\text {s t y l e}} \right]\right). \tag {5} +$$ + +In this way, we can guide the network to predict the desired color during inference by providing a certain style condi + +tion, thereby alleviating the biased color issue. + +Coarse texture image. With the colored point clouds, we then project them onto the 2D UV space based on the pre-calculated UV mapping. Following this, we perform KNN interpolation based on the 3D coordinates to assign colors to the remaining uncolored pixels, thereby generating a coarse texture image $\mathbf{x}_{\mathrm{coarse}}$ for the given mesh. This coarse texture image offers a coherent base color initialization and functions as a conditional element to guide the high-fidelity texture generation in the second stage, helping to avoid discontinuity caused by UV mapping. + +# 4.2. Fine Stage: UV Diffusion + +To further refine the coarse texture image, we design a fine stage using 2D diffusion in the UV space, as depicted in Figure 4 (Bottom). In addition to the conditions used in the coarse stage, we incorporate the coarse texture image $\mathbf{x}_{\mathrm{coarse}}$ as an additional condition. We apply a 2D U-Net $G_{\theta_2}^2$ combined with self-attention modules to learn the high-quality texture image $\hat{\mathbf{x}}_0$ , from a noisy texture image $\mathbf{x}_t$ : + +$$ +\hat {\mathbf {x}} _ {0} = G _ {\theta_ {2}} ^ {2} \left(\left[ \mathbf {x} _ {\text {s h a p e}}, \mathbf {x} _ {t}, \mathbf {x} _ {\text {c o a r s e}}, t \right]\right). \tag {6} +$$ + +Hybrid condition. During training, we employ FPS on the ground-truth texture image, $\mathbf{x}_0$ , followed by interpolation to simulate the coarse texture image, $\mathbf{x}_{\mathrm{coarse}}$ . However, in joint cascaded testing involving both stages, we note that the output quality of $\hat{\mathbf{x}}_{\mathrm{coarse}}$ from the first stage doesn't always align flawlessly with the quality of $\mathbf{x}_{\mathrm{coarse}}$ . Such mismatches can influence the performance of the subsequent stage. To address this, we introduce a hybrid conditioning method aimed at narrowing this discrepancy. Firstly, we create a smooth texture image $\mathbf{x}_{\mathrm{smooth}}$ (see Figure 6), inspired by the blur augmentation described in [17] for cascaded diffusion models. In particular, we segment the mask map into multiple discrete regions using four-connectivity detection, then perform average color pooling within each region based on its connectivity. After this procedure, $\mathbf{x}_{\mathrm{smooth}}$ maintains merely the regional color, ensuring more consistent alignment across both training and testing phases. Then, we combine $\mathbf{x}_{\mathrm{coarse}}$ with $\mathbf{x}_{\mathrm{smooth}}$ , using a certain probability $p_{\mathrm{hybrid}}$ during training. Thus, the network is forced to be capable of generating textures even when adopting a weaker condition $\mathbf{x}_{\mathrm{smooth}}$ in the fine stage during inference. Considering that the diffusion model first generates low-frequency information and then higher one [8], we also explore a condition-truncated sampling, as detailed discussed in Section 5.4, where we condition both maps for generation during the initial time steps and then exclusively utilize the smooth map for the remainder of the generation. + +![](images/78e68ad0434138d143723c3880402adbd98827bbb432950edb198cb32e5506dc.jpg) +(a) Samples from cluster 1 + +![](images/355323337c1383ea88af947ed41b22f456a8e2ee1f29a46adba57c4c190866bd.jpg) +(d) Samples from cluster 4 + +![](images/b80f5fa28dbd811834d417bd7fee31866b25e4db8cdd6595ad89ac370118ef38.jpg) +Figure 5. Illustration of samples across various clusters. Shapes within a particular cluster exhibit analogous color styles. However, there exists an imbalance in the quantity of shapes among different clusters, leading to challenges in unbiased synthesis. +(a) Texture image +Figure 6. Varieties of texture images. (a) Original texture image enriched with high-frequency details, (b) Coarse texture image, and (c) Smooth texture image. + +![](images/d2b50a6733fa74d83efc1a693d73306df119eb23950b24af615cd5fd3178b772.jpg) +(b) Coarse texture image + +![](images/8eaabc9f4446cc5dfd47c1d7b23a300f7a4de380b347cc2d3005cd19a031a678.jpg) +(c) Smooth texture image + +# 5. Experiments + +# 5.1. Datasets and Implementation Details + +We conduct experiments across four categories of the ShapeNet dataset [4], namely, chair, table, car, and bench. Before training, we use an open-source UV-Atlas tool [46] to generate the UV map and pre-process the dataset to obtain the shape maps and ground-truth texture images. We sample $K = 4096$ points in our coarse stage and synthesize a texture image in the UV space with a resolution of $512 \times 512$ for the fine stage. For training the diffusion models, akin to [1], we predict the clean signals. This approach provides more stable training than predicting the noise component as recommended by [16]. We employ the cosine noise scheduling [27] ranging from 0.0001 to 0.02 over 1,024 time steps for both stages. In the fine stage, we leverage the noise scaling strategy from [7] and also incorporate a rendering loss. Specifically, we randomly select four views and render the mesh using the predicted UV map and the ground-truth UV map to produce $1024 \times 1024$ images. Subsequently, we crop $224 \times 224$ patches from these images and compute the corresponding $L_{1}$ loss. For the hybrid condition in the fine stage, we use $p_{\mathrm{hybrid}} = 0.3$ and sweep over condition-truncated time $t_{c}$ (see Figure 12). + +Compared methods. We compare our method with state-of-the-art approaches, including Texturify [38], Texture + +![](images/86b797e91ed0ca00ebed9a3cc70e0b93e0e6164bedfeaae3d8a5080965b9cba0.jpg) +Figure 7. Our texture synthesis results in different categories. The results on chairs, tables, cars, and benches demonstrate that our approach can generate natural, vivid, and diverse textures, even though the given 3D shapes are challenging, such as cutouts on the bench. + +Fields [29], and PVD [49], in the context of unconditional texture generation. PVD was originally designed for point cloud generation instead of texture generation, so it cannot be directly compared with our approach. To facilitate comparison, we modify this framework to learn RGB values from the sampled point cloud. This extension, referred to as "PVD-Tex", serves as a baseline for directly learning texture in the point space using the diffusion model. + +# 5.2. Unconditional Texture Generation + +Gallary. First, we showcase our texture generation results for each category without additional conditions. The results in Figure 7 and Figure 1 demonstrate the remarkable performance of our method, which is able to generate appealing textures with fine details and preserve the geometric structures. Note that our method is compatible with meshes of diverse topology and intricate geometric details. + +Qualitative comparisons. As depicted in Figure 8, our approach excels at generating high-quality textures while maintaining the geometric intricacies of the input mesh. Firstly, the results of Texture Fields [29] appear deficient in high-frequency details. Additionally, while PVD-Tex [49] is able to produce spatially varying colors to some extent (e.g., the car in Figure 8), it falls short in synthesizing intricate high-frequency details. Lastly, even though Texturify [38] demonstrates an ability to generate finer textures, it compromises on preserving slender structures (i.e. chair). + +Quantitative comparisons. To quantitatively compare with existing works, we follow [38] and assess the genera + +tion quality using Frechet Inception Distance (FID)[15] and Kernel Inception Distance (KID)[2], metrics that are widely used for evaluating image generation models. To this end, we render $512 \times 512$ images from each generated textured mesh and ground-truth textured mesh using four distinct camera views. Table 1 presents the quantitative comparisons with current approaches, revealing that our method surpasses existing works. + +# 5.3. Conditional Texture Generation + +In addition, we demonstrate the capability of our framework to synthesize textures conditioned on either text prompts or a single-view image. We conduct our experiments on the chair and table categories. For the text condition, we utilize text-shape pairs as provided in [6] (with additional corrections for text accuracy). For the image condition, we randomly render a view from the ground-truth mesh. To infuse the network with condition-specific information, we use the pre-trained vision-language model CLIP [33] to extract the corresponding embedding from either the image or the text. This embedding is then fed into a simple MLP to incorporate the information into the diffusion model as a condition for both training and inference. As depicted in Figure 9, our method succeeds in generating textures that align well with the given text descriptions or images. We also compare our approach with Text2Mesh [24], a test-time optimization method. Text2Mesh takes around 10 minutes per instance, while ours only requires 30 seconds. Importantly, to adapt to our task which requires preserving the mesh geometry, we freeze the geometry deformation branch of Text2Mesh to generate only colors. As shown in Figure 10, Text2Mesh + +![](images/fc1e5f50344813ec6fab5477aca5d6115d6fa928b654080272d06ad0cfea7468.jpg) +Figure 8. Qualitative comparisons with existing works. Given pure 3D shapes, Texturify [38] generates textures on the surface of 3D shape but tends to damage the topology; Texture fields [29] and PVD-Tex [49] produce textures with limited details. On the contrary, our approach faithfully preserves topology and produces realistic appearances. + +is bound by vertex resolution, while our method can deliver detailed visuals even on low-resolution meshes, offering a distinct advantage in graphics. + +# 5.4. Ablation Studies + +Coarse-to-fine diffusion. To manifest the effectiveness of our coarse-to-fine diffusion strategy, we conduct experiments on two baselines "w/o coarse stage" and "w/o fine stage". The "w/o coarse stage" configuration indicates that we directly generate the UV texture image using the fine stage, bypassing the initialization from the coarse stage. The "w/o fine stage" configuration signifies the result of the coarse stage, after assigning colors to the uncolored pixels using KNN interpolation. In both scenarios, the model produces inferior outcomes relative to our full model, as shown in Table 2 and Figure 11. Absent the coarse stage, the generated result suffers from noticeably inconsistent colors. Without the fine stage, the results are over-smoothed. + +Hybrid condition. As shown in Table 2 and Figure 11 "coarse condition", the fine stage cannot generate high-quality textures when it is exclusively conditioned on the + +![](images/d65f2fe611a112f9326e484b0374fcb845e223f7ce4fafd1259ec5be1d3b39c7.jpg) +Figure 9. Results of conditional texture generation. Our method is adaptable to craft textures guided by text descriptions (rows 1-3) or single-view images (rows 4-6). + +![](images/9697bacf708f3c7ee5d0feeb1e4398bbe83b2a70456554c5d10c05e5e5f8cbcb.jpg) +Figure 10. Comparison with Text2Mesh. We freeze the geometry deformation branch of Text2Mesh to adapt to our task. Ours can generate high-frequency details for low-resolution mesh. + +coarse map. Further, the generated quality remains unsatisfactory if we apply a hybrid condition entirely during training (i.e., $p_{\mathrm{hybrid}} = 1.0$ ), as shown in " $p_{\mathrm{hybrid}} = 1.0$ ". We attribute this to the network's propensity to depend solely on the coarse map during training and generation. In contrast, our full model (i.e., hybrid condition with $p_{\mathrm{hybrid}} = 0.3$ ) achieves significant improvement both + +
ChairCarTableBench
MethodsFID ↓KID ↓FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓
Texture Fields [29]24.241.07156.3813.6468.964.2062.712.96
Texturify [38]27.801.3273.164.71----
PVD-Tex [49]15.520.6259.473.7416.120.5528.940.39
Ours9.880.2226.890.689.630.1523.090.15
+ +![](images/84a3def2507621a83f179e2207ec41d7871a84169839c3e98f7ea23eb33bf6cf.jpg) +(a) untextured mesh + +![](images/de8ba1fb2bbc89c540f09921c6de151892921d8d4601c509a949c7e569b92b51.jpg) +(b) w/o coarse stage + +![](images/982320679431f2013186f29101ba9df8eea6815de214fb528438f7143ddba134.jpg) +(c) w/o fine stage + +![](images/fa9d3f85ce3244cde44c54ba778651e23f9615245bc9f5138883c8ace80c0938.jpg) +(d) coarse condition + +![](images/b14609cb0ea646c3e293935110cac2f5b0df18528fa22ebf0fcb096eaa07d55d.jpg) +(e) $p_{hybrid} = 1.0$ + +![](images/ca8050151c919536b4463c759fad95ce7b7bb56b3ff226b18a716ccd92f8bf8b.jpg) +(f) ours +Figure 11. Ablation studies. Without our two-stage generation pipeline or hybrid condition designs, the generated results are inconsistent with the 3D shape or lack of high-frequency details. + +Table 1. Quantitative comparisons with existing works. Ours outperforms other approaches on both FID [15] and KID $(\times 10^{2})$ [2]. + +
FID ↓KID ↓
w/o fine stage17.880.76
w/o coarse stage15.110.49
coarse condition14.930.56
\(p_{hybrid}=1.0\)15.250.59
Ours9.880.22
+ +qualitatively and quantitatively, as presented in Table 2 and Figure 11 "ours". Furthermore, we also sweep over the effect of condition-truncated time $t_c$ for inference. As Figure 12 shows, $t_c = 0.4$ strikes a sweet point. The model utilizes information from both conditions maps to better generate low-frequency components during the first $40\%$ of the sampling timesteps. After that, the coarse map $\mathbf{x}_{\mathrm{coarse}}$ is dropped, and the model further focuses on fine-grained detail generation without relying on the coarse map. + +Style guidance. The style guidance is aimed at addressing the issue of insufficient diversity due to color distribution bias in the dataset. As a result, we do not utilize FID or KID for evaluation since they evaluate the distribution sim + +Table 2. Ablation studies. This table shows the effectiveness of each component in our proposed method. + +
MethodsPreference↑LPIPS↑
Ours49.8%0.083
Texturify [38]15.9%0.086
PVD-Tex [49]29.2%0.029
Texture Fields [29]5.1%0.005
+ +Table 3. LPIPS and user study. This table shows the average LPIPS and preference via user study in the chair category. + +The diversity of the LPIPS is a key feature that is crucial to the quality of the generated results and ground truth. However, the ground-truth distribution shows a strong bias towards particular colors, making these metrics inappropriate for assessing generative texture diversity. In contrast, we adopt LPIPS [48] to measure the pairwise similarity among five textures generated by our method given the same input mesh, where a larger diversity in textures will lead to a higher LPIPS value. In Figure 13, we report the quantitative results for 500 shapes. Our approach achieves a higher LPIPS in most of the evaluated cases, indicating better diversity. Besides, as shown in Figure 14, we present the results of generating three textures randomly for the shapes without and with style guidance. In the latter case, we uniformly sample three style labels as style inputs. It is evident that without style guidance, the generated textures are nearly identical, and the color styles of different shapes tend to be similar. With the introduction of style guidance, however, the diversity of colors has significantly increased. We also conduct a comparison with other methods in terms of diversity, as well as a quality assessment through user studies, as shown in Table 3. + +# 6. Discussion, Limitations, and Conclusions + +This paper presents Point-UV diffusion, a brand-new framework that employs a coarse-to-fine pipeline to generate textures for 3D meshes. We begin with a 3D diffusion model to synthesize low-frequency texture components from point clouds, which maintain 3D consistency. We then refine the textures using a 2D UV-space diffusion model. Our method is compatible with meshes with arbitrary topology and can faithfully preserve the geometry structure. We further demonstrate the flexibility of our framework by extending it to conditional generative models. + +Despite its merits, our method has inherent limitations. Similar to other methodologies relying on 3D data for train- + +![](images/5fd0d599729542c9b7c974173dd1ac7d5e71b924af103adc5dced05045c12332.jpg) +Figure 12. Examination of condition-truncated time. The horizontal axis denotes the percentage of time points when the coarse texture image is conditioned during inference. The left vertical axis indicates the FID value, while the right vertical axis shows the KID value $(\times 10^{2})$ . + +![](images/06390beb6f5eec44e05400f76334549a09b8030b3ba6da06b328c2c9b0ab768a.jpg) +Figure 13. Diversity measurement. Utilizing style guidance leads to larger LPIPS scores, indicating enhanced generation diversity. + +ing, our technique is upper-bounded by the scope and diversity of current 3D datasets. This restriction poses challenges in generating textures that parallel the depth of effects seen in 2D image synthesis. Moreover, our method's efficacy relies on the quality of UV mapping. Our approach faces difficulties in rendering high-quality results for meshes where the UV mapping produces excessive fragmented cuts, resulting in fragmented artifacts. This phenomenon is commonly observed in the car category, as shown in the Figure 15. We believe the emergence of larger and more diverse 3D datasets would be helpful for generating superior-quality textures. Further advancements in UV parameterization would also be beneficial in augmenting our method's consistency. + +# 7. Acknowledgements + +This work has been supported by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant No. 17202422), and RGC matching fund scheme (RMGS). Part + +![](images/7dd3488f2ed95f997cfcd5c595ebd414538e0b8d0a8a065ad88e85c438c55789.jpg) +(a) w/o style guidance +Result 1 + +![](images/1c4718cca7ba0e778997d2ff96a834b0db2dde9ea762c238ecb6c23b4d182451.jpg) +Result 2 + +![](images/4890e18c992d21c254ca1d2ce295014cf1c69273442006ab745df2fe643f5c7e.jpg) +Result 3 + +![](images/ce8b49180e1a128a3c18a36a4ed30a43e3aa5778fcc29b6351b3bac82173037d.jpg) +(b) w/ style guidance +Result 1 +Figure 14. Efficacy of style guidance. The model without style guidance produces consistent colors across different random seeds (rows 1-2). In contrast, with style guidance, diverse outcomes emerge (rows 3-4). + +![](images/c7e8219b8cc398c024d35bc258208c08fe92b14bb737544b5a608a3377b07c9b.jpg) +Result 2 + +![](images/74411651e70bb3a0ad2617dfa06e510a30e900647cec2688889e10165142ce83.jpg) +Result 3 + +![](images/37ea41a649f5635d3496f1b5e4805874ba17d11c31569c9b66cd288641236060.jpg) +(a) Generated Texture Image +Figure 15. Failure case. Our approach still struggles to generate seamless results when there are too many fragmented cuts of the UV map. + +![](images/c79cf71ae6423c18fb1df3c0a9a3ae93b90dd28b4932b2f24394bb28a8f1c2a8.jpg) +(b) Textured Mesh + +of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust. + +# References + +[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. arXiv preprint arXiv:2211.09869, 2022. 3, 5 +[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6, 8 +[3] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 2 +[4] Angel X. Chang, Thomas Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], 2015. 5 +[5] Bindita Chaudhuri, Nikolaos Sarafianos, Linda Shapiro, and Tony Tung. Semi-supervised synthesis of high-resolution editable textures for 3d humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7991-8000, 2021. 1 +[6] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III 14, pages 100-116. Springer, 2019. 1, 2, 6 +[7] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 5 +[8] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5 +[9] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10673-10683, 2022. 2 +[10] Jun Gao, Wenzheng Chen, Tommy Xiang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. Advances In Neural Information Processing Systems, 33:9936-9947, 2020. 1 +[11] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 1, 2 +[12] Lin Gao, Tong Wu, Yu-Jie Yuan, Ming-Xian Lin, Yu-Kun Lai, and Hao Zhang. Tm-net: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG), 40(6):1-15, 2021. 2 + +[13] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. In Advances in neural information processing systems, pages 5767-5777, 2017. 1 +[14] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Learning a neural 3d texture space from 2d exemplars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8356-8364, 2020. 2 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. NIPS, 2017. 6, 8 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models, 2020. 2, 3, 5 +[17] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23(47):1-33, 2022. 5 +[18] Jingyu Hu, Ka-Hei Hui, Zhengzhe Liu, Ruihui Li, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation, inversion, and manipulation. arXiv preprint arXiv:2302.00190, 2023. 3 +[19] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation. In SIGGRAPH Asia 2022 Conference Papers, pages 1–9, 2022. 1, 3 +[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1 +[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3 +[22] Zhengzhe Liu, Peng Dai, Ruihui Li, Xiaojuan Qi, and Chi-Wing Fu. Iss: Image as stetting stone for text-guided 3d shape generation. International Conference on Learning Representations, 2023. 2 +[23] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Pointvoxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems, 32, 2019. 4 +[24] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3, 6 +[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[26] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2 + +[27] Alex Nichol and Prafulla Dhariwal. Improved Denoising Diffusion Probabilistic Models, 2021. 5 +[28] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3 +[29] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 1, 2, 6, 7, 8 +[30] Dario Pavllo, Graham Spinks, Thomas Hofmann, Marie-Francine Moens, and Aurelien Lucchi. Convolutional generation of textured 3d meshes. Advances in Neural Information Processing Systems, 33:870-882, 2020. 2 +[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations, 2023. 2, 3 +[32] Tiziano Portenier, Siavash Arjomand Bigdeli, and Orcun Goksel. Gramgan: Deep 3d texture synthesis from 2d exemplars. Advances in Neural Information Processing Systems, 33:6994-7004, 2020. 2 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6 +[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 3 +[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 1 +[37] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 2 +[38] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. arXiv preprint arXiv:2204.02411, 2022. 2, 5, 6, 7, 8 +[39] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. Advances in Neural Information Processing Systems, 35:24487-24501, 2022. 2 +[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Confer + +ence on Machine Learning, pages 2256-2265. PMLR, 2015. 3 +[41] Yongbin Sun, Ziwei Liu, Yue Wang, and Sanjay E Sarma. Im2avatar: Colorful 3d reconstruction from a single image. arXiv preprint arXiv:1804.06375, 2018. 2 +[42] Shubham Tulsiani, Nilesh Kulkarni, and Abhinav Gupta. Implicit mesh reconstruction from unannotated image collections. arXiv preprint arXiv:2007.08504, 2020. 2 +[43] Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2626-2634, 2017. 2 +[44] Hao Wang, Guosheng Lin, Steven Hoi, and Chunyan Miao. Cycle-consistent inverse GAN for text-to-image synthesis. ACM MM, 2021. 1 +[45] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12456-12465, 2021. 1 +[46] Jonathan Young. xatlas, 2018. https://github.com/jpcy/xatlas.5 +[47] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 3 +[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 8 +[49] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 3, 4, 6, 7, 8 \ No newline at end of file diff --git a/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/images.zip b/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..708e0a1cb54e68359618281a67283f3adde0b2ef --- /dev/null +++ b/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3695b60a458fad71ff15efd673d58e0cdbbed8343019e7617a542d1a93062cec +size 779771 diff --git a/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/layout.json b/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..57908a024de06b6ad26b7b1e91e5bc13007e2fb3 --- /dev/null +++ b/2023/Texture Generation on 3D Meshes with Point-UV Diffusion/layout.json @@ -0,0 +1,9619 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 479, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 479, + 119 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 479, + 119 + ], + "type": "text", + "content": "Texture Generation on 3D Meshes with Point-UV Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 142, + 108, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 142, + 108, + 156 + ], + "spans": [ + { + "bbox": [ + 67, + 142, + 108, + 156 + ], + "type": "text", + "content": "Xin Yu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 143, + 184, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 143, + 184, + 158 + ], + "spans": [ + { + "bbox": [ + 133, + 143, + 184, + 158 + ], + "type": "text", + "content": "Peng Dai" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 143, + 262, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 143, + 262, + 156 + ], + "spans": [ + { + "bbox": [ + 208, + 143, + 262, + 156 + ], + "type": "text", + "content": "Wenbo Li2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 286, + 143, + 331, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 143, + 331, + 156 + ], + "spans": [ + { + "bbox": [ + 286, + 143, + 331, + 156 + ], + "type": "text", + "content": "Lan Ma3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 354, + 143, + 432, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 143, + 432, + 157 + ], + "spans": [ + { + "bbox": [ + 354, + 143, + 432, + 157 + ], + "type": "text", + "content": "Zhengzhe Liu" + }, + { + "bbox": [ + 354, + 143, + 432, + 157 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 456, + 143, + 526, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 456, + 143, + 526, + 158 + ], + "spans": [ + { + "bbox": [ + 456, + 143, + 526, + 158 + ], + "type": "text", + "content": "Xiaojuan Qi" + }, + { + "bbox": [ + 456, + 143, + 526, + 158 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 171, + 541, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 541, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 541, + 186 + ], + "type": "text", + "content": "1The University of Hong Kong, 2The Chinese University of Hong Kong, 3TCL Corporate Research" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 187, + 238, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 187, + 238, + 199 + ], + "spans": [ + { + "bbox": [ + 70, + 187, + 238, + 199 + ], + "type": "text", + "content": "{yuxin, daipeng, xjqi}@eee.hku.hk" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 259, + 187, + 426, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 187, + 426, + 199 + ], + "spans": [ + { + "bbox": [ + 259, + 187, + 426, + 199 + ], + "type": "text", + "content": "{wenboli,zzliu}@cse.cuhk.edu.hk" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 447, + 188, + 525, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 447, + 188, + 525, + 198 + ], + "spans": [ + { + "bbox": [ + 447, + 188, + 525, + 198 + ], + "type": "text", + "content": "rubyma@tcl.com" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 55, + 209, + 542, + 399 + ], + "blocks": [ + { + "bbox": [ + 55, + 209, + 542, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 542, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 542, + 399 + ], + "type": "image", + "image_path": "b22925ff733e95162010d693f891239383078d8f8361fa87400e6d9beb215271.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 403, + 547, + 426 + ], + "lines": [ + { + "bbox": [ + 46, + 403, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 547, + 426 + ], + "type": "text", + "content": "Figure 1. A gallery of generated results by our Point-UV diffusion. Our method is capable of processing meshes of any genus, generating diversified, geometry-compatible, and high-fidelity textures." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 434, + 192, + 446 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 434, + 192, + 446 + ], + "spans": [ + { + "bbox": [ + 143, + 434, + 192, + 446 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 45, + 454, + 290, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 454, + 290, + 658 + ], + "spans": [ + { + "bbox": [ + 45, + 454, + 290, + 658 + ], + "type": "text", + "content": "In this work, we focus on synthesizing high-quality textures on 3D meshes. We present Point-UV diffusion, a coarse-to-fine pipeline that marries the denoising diffusion model with UV mapping to generate 3D consistent and high-quality texture images in UV space. We start with introducing a point diffusion model to synthesize low-frequency texture components with our tailored style guidance to tackle the biased color distribution. The derived coarse texture offers global consistency and serves as a condition for the subsequent UV diffusion stage, aiding in regularizing the model to generate a 3D consistent UV texture image. Then, a UV diffusion model with hybrid conditions is developed to enhance the texture fidelity in the 2D UV space. Our method can process meshes of any genus, generating diversified, geometry-compatible, and high-fidelity textures. Code is available at https://cvmilab.github.io/Point-UV-Diffusion." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 434, + 386, + 446 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 386, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 386, + 446 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 458, + 547, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 547, + 566 + ], + "type": "text", + "content": "Texturing 3D meshes is a fundamental task in computer vision and graphics. It enhances the visual richness of 3D objects, thereby facilitating their application in various fields such as video games, 3D movies, and AR/VR technologies. However, generating high-quality textures can be daunting and time-consuming, often requiring specialized knowledge and resources. As such, there is a pressing need for an efficient approach to automatically create high-quality textures on 3D meshes." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "Despite the substantial progress the community has made in 2D image synthesis and 3D shape generation using GANs [13, 20, 44, 11] or diffusion models [34, 36, 35, 19], crafting realistic textures on mesh surfaces remains challenging. One major difficulty stems from the need for suitable 3D representations for texture synthesis. Early approaches investigate the use of voxels [5, 45, 6] or point clouds [10] and synthesize point/voxel colors. However, they can only afford to synthesize low-resolution results with low-fidelity textures due to memory and model complexity constraints. In response, Texture Fields [29] adopts an implicit representation with the potential to synthesize" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 702, + 147, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 702, + 147, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 702, + 147, + 713 + ], + "type": "text", + "content": "†: Corresponding authors" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 732, + 316, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 316, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 316, + 742 + ], + "type": "text", + "content": "4206" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "high-resolution textures, but still could not yield satisfactory results as shown in Figure 2 (a): over-smoothed results. Most recently, Siddiqui et al. [38] propose to parameterize the shape as tetrahedral meshes and introduce tetrahedral mesh convolution to enhance local details. Albeit improving results, tetrahedral parameterization inevitably destroys geometric details of the input mesh and thus cannot faithfully preserve the original structure. As shown in Figure 2 (b), the delicate structures of the chair's back are absent. Moreover, the generative models utilized in these methods are limited to GANs [29, 38] and VAEs [29, 12]. The more advanced diffusion model, which could potentially open up new avenues for high-quality texture generation, remains insufficiently explored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 287, + 397 + ], + "type": "text", + "content": "In this paper, we delve into a novel texture representation based on UV maps and investigate the advanced diffusion model for texture generation. The 2D nature of the UV map enables it to circumvent the cost of high-resolution point/voxel representations. Besides, the UV map is compatible with arbitrary mesh topologies, thereby preserving the original geometric structures. However, while promising, direct integration of the UV map representation with a 2D diffusion model presents challenges in synthesizing seamless textures, leading to severe artifacts, as shown in Figure 2 (c). This occurs because the UV mapping process fragments the continuous texture on the 3D surface into isolated patches on the 2D UV plane (see Figure 3)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 400, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 568 + ], + "type": "text", + "content": "To this end, we introduce Point-UV diffusion, a two-stage coarse-to-fine framework consisting of point diffusion and UV diffusion. Specifically, we initially design a point diffusion model to generate color for sampled points that act as low-frequency texture components. This model is equipped with a style guidance mechanism that alleviates the impact of biased color distributions in the dataset and facilitates diversity during inference. Next, we project these colorized points onto the 2D UV space with 3D coordinate interpolation, thereby generating a coarse texture image that maintains 3D consistency and continuity. Given the coarse textured image, we develop a UV diffusion model with elaborately designed hybrid conditions to improve the quality of the textures (see Figure 2 (ours) and Figure 1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "In short, our contributions are as follows: 1) We propose a new framework for texture generation for given meshes. Our representation can handle meshes with arbitrary topology and is able to faithfully preserve geometric structures. 2) To the best of our knowledge, we are the first to train a diffusion model specifically for mesh texture generation. Our coarse-to-fine framework allows us to enjoy the efficiency of 2D representation while enhancing 3D consistency. 3) We compare our approach with multiple methods in unconditional generation and achieve state-of-the-art results. Furthermore, we demonstrate that our method can be easily extended to scenarios with text-conditioning and" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 74, + 400, + 207 + ], + "blocks": [ + { + "bbox": [ + 310, + 74, + 400, + 207 + ], + "lines": [ + { + "bbox": [ + 310, + 74, + 400, + 207 + ], + "spans": [ + { + "bbox": [ + 310, + 74, + 400, + 207 + ], + "type": "image", + "image_path": "9892278130303e08163bccc166ec305808fbec62493a36a1e330353c20108282.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 211, + 384, + 219 + ], + "lines": [ + { + "bbox": [ + 313, + 211, + 384, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 384, + 219 + ], + "type": "text", + "content": "(a) Texture Fields vs Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 402, + 74, + 474, + 207 + ], + "blocks": [ + { + "bbox": [ + 402, + 74, + 474, + 207 + ], + "lines": [ + { + "bbox": [ + 402, + 74, + 474, + 207 + ], + "spans": [ + { + "bbox": [ + 402, + 74, + 474, + 207 + ], + "type": "image", + "image_path": "c61bc7531d7ce27c511cb3be73337055d95bf097cd97b0bf8611eed49c31584c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 211, + 460, + 219 + ], + "lines": [ + { + "bbox": [ + 403, + 211, + 460, + 219 + ], + "spans": [ + { + "bbox": [ + 403, + 211, + 460, + 219 + ], + "type": "text", + "content": "(b) Texturify vs Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 465, + 74, + 545, + 207 + ], + "blocks": [ + { + "bbox": [ + 465, + 74, + 545, + 207 + ], + "lines": [ + { + "bbox": [ + 465, + 74, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 465, + 74, + 545, + 207 + ], + "type": "image", + "image_path": "55a88ce9b2911b3d4ea4a7da214f77dd9b1cc1e98ed13e135ca289646d74c39d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 211, + 541, + 219 + ], + "lines": [ + { + "bbox": [ + 475, + 211, + 541, + 219 + ], + "spans": [ + { + "bbox": [ + 475, + 211, + 541, + 219 + ], + "type": "text", + "content": "(c) 2D diffusion vs Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 293, + 447, + 392 + ], + "blocks": [ + { + "bbox": [ + 304, + 227, + 546, + 282 + ], + "lines": [ + { + "bbox": [ + 304, + 227, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 546, + 282 + ], + "type": "text", + "content": "Figure 2. Comparisons with different methods. Our generative results (a) possess high-quality details, (b) faithfully preserve the mesh structure, and (c) are better consistent with the given shape, compared with Texture Fields [29], Texturify [38] and 2D diffusion [16], respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 293, + 447, + 392 + ], + "lines": [ + { + "bbox": [ + 321, + 293, + 447, + 392 + ], + "spans": [ + { + "bbox": [ + 321, + 293, + 447, + 392 + ], + "type": "image", + "image_path": "30221e74c987260583731d2865db7cc1d63613dff5bb7263268b952ee83092c3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 393, + 392, + 401 + ], + "lines": [ + { + "bbox": [ + 348, + 393, + 392, + 401 + ], + "spans": [ + { + "bbox": [ + 348, + 393, + 392, + 401 + ], + "type": "text", + "content": "Textured Mesh" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 408, + 546, + 441 + ], + "lines": [ + { + "bbox": [ + 304, + 408, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 546, + 441 + ], + "type": "text", + "content": "Figure 3. Illustration of UV mapping process. It establishes connections between the 2D texture map and the surface appearance of 3D shape." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 450, + 293, + 546, + 390 + ], + "blocks": [ + { + "bbox": [ + 450, + 293, + 546, + 390 + ], + "lines": [ + { + "bbox": [ + 450, + 293, + 546, + 390 + ], + "spans": [ + { + "bbox": [ + 450, + 293, + 546, + 390 + ], + "type": "image", + "image_path": "8da5aa7d15164e4c4214b2694020aeb680d8c4b36c6a7df4785e57607da9f6ee.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 394, + 519, + 403 + ], + "lines": [ + { + "bbox": [ + 477, + 394, + 519, + 403 + ], + "spans": [ + { + "bbox": [ + 477, + 394, + 519, + 403 + ], + "type": "text", + "content": "Texture Image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 464, + 389, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 389, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 389, + 475 + ], + "type": "text", + "content": "image-conditioning." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 489, + 392, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 392, + 501 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 392, + 501 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "Texture generation. Early works [41, 43, 6] propose using voxels to represent colors. However, due to a cubic increase in memory usage and computational cost with rising resolution, these methods produce only coarse, low-resolution textures. For meshes, recent work in texture generation [38] predicts the color for each face, but the representation capacity remains largely constrained by the mesh's resolution. Another set of works [30, 42, 26, 12] leverage UV mapping, but they require either spherical parameterization or part segmentation and, therefore, are confined to handling low-genus shapes. Other works [32, 14] rely on image exemplars, while our approach focuses on unconditional generation. Recently, implicit functions attract increasing attention for generative tasks [11, 31, 39, 3, 9, 37, 29, 22]. Most of this research centers on generating 3D-aware images [39, 3, 9, 37], rather than synthesizing textures for given 3D meshes. Texture Fields [29] represents" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4207" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "the most similar work to our task, which, however, tends to produce over-smoothed results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 289, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 289, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 289, + 301 + ], + "type": "text", + "content": "Some methods focus on test-time optimization for generation tasks. For instance, Text2Mesh [24] employs CLIP [33] to design a loss function. This function stylizes a 3D mesh to align with a target text prompt, predicting both the color and displacement for each mesh vertex. More recently, DreamFusion [31] and Magic3D [21] leverage pretrained stable diffusion [35] for score distillation, creating either NeRF [25] or a 3D mesh. However, these approaches often demand extensive optimization time or necessitate alterations to the existing mesh geometry [24]. In this paper, our objective is to generate textures for specified full-3D meshes with arbitrary topology. This goal entails two fundamental requirements: 1) preserving the original mesh's geometric structure; 2) producing a 3D texture representation exportable as either vertex color or a uv-texture-image, instead of merely generating-rendering multi-view or 3D-aware images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 315, + 289, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 289, + 507 + ], + "type": "text", + "content": "Diffusion models for 3D generation. In addition to 2D image generation, diffusion models recently gain significant attention in 3D generation, leading to many related works [49, 18, 19, 28, 31, 47, 1]. For instance, Zhou et al. [49] introduce point-voxel diffusion for point cloud generation and completion. Hui et al. [19] and Hu et al. [18] suggest a compact wavelet domain representation for shapes, enabling higher-quality shape creation via diffusion models. Nichol et al. [28] present Point-E, a system that uses a text-conditional strategy to produce colored 3D point clouds. This approach enables an efficient synthesis of intricate 3D shapes from textual prompts. Yet, this system yields lower-resolution point clouds, often missing detailed shapes and textures. In this work, we develop a brand new 3D diffusion model for texture image synthesis, which allows high-fidelity and 3D-consistent texture generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 517, + 133, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 133, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 133, + 529 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 536, + 194, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 194, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 194, + 550 + ], + "type": "text", + "content": "3.1. Denoising diffusion model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "text", + "content": "Diffusion model [40, 16] is a kind of likelihood-based generative model that has gained significant attention recently. It learns the data distribution " + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "text", + "content": " by progressive denoising from a prior Gaussian distribution. Given a sample from the data distribution " + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "inline_equation", + "content": "x_0 \\sim q(x_0)" + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "text", + "content": ", a fixed forward process " + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{1:T} \\mid \\mathbf{x}_0) = \\prod_{t=1}^{T} q(\\mathbf{x}_t \\mid \\mathbf{x}_{t-1})" + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "text", + "content": " is used to perturb the data with Gaussian kernels " + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t \\mid \\mathbf{x}_{t-1}) \\coloneqq \\mathcal{N}\\left(\\sqrt{1 - \\beta_t} \\mathbf{x}_{t-1}, \\beta_t \\mathbf{I}\\right)" + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "text", + "content": ", producing increasingly noisy latent variables " + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{x_1, x_2, \\ldots, x_T\\}" + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "text", + "content": ". Then, a parameterized Markov process " + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "inline_equation", + "content": "p_\\theta(\\mathbf{x}_{0:T}) = p(\\mathbf{x}_T) \\prod_{t=1}^{T} p_\\theta(\\mathbf{x}_{t-1} \\mid \\mathbf{x}_t)" + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "text", + "content": " with transition kernel " + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "inline_equation", + "content": "p_\\theta(\\mathbf{x}_{t-1} \\mid \\mathbf{x}_t) \\coloneqq \\mathcal{N}\\left(\\mu_\\theta(\\mathbf{x}_t, t), \\sigma_t^2 \\mathbf{I}\\right)" + }, + { + "bbox": [ + 46, + 555, + 289, + 715 + ], + "type": "text", + "content": " is optimized through maximizing a variational lower bound of log data likelihood, which essentially targets to match the joint" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 395, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 395, + 85 + ], + "type": "text", + "content": "distribution " + }, + { + "bbox": [ + 306, + 72, + 395, + 85 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_{0:T})" + }, + { + "bbox": [ + 306, + 72, + 395, + 85 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 92, + 547, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 92, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 92, + 547, + 121 + ], + "type": "interline_equation", + "content": "\\operatorname {E} _ {q \\left(\\mathbf {x} _ {0}\\right)} \\left[ \\log p _ {\\theta} \\left(\\mathbf {x} _ {0}\\right) \\right] \\geq \\operatorname {E} _ {q \\left(\\mathbf {x} _ {0: T}\\right)} \\left[ \\log \\frac {p _ {\\theta} \\left(\\mathbf {x} _ {0 : T}\\right)}{q \\left(\\mathbf {x} _ {1 : T} \\mid \\mathbf {x} _ {0}\\right)} \\right]. \\tag {1}", + "image_path": "875258fba8c2f0ea2a9c60c2fdec1d17fa26c69482f722406bd56d4737180ad9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 127, + 546, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 127, + 546, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 127, + 546, + 152 + ], + "type": "text", + "content": "After training, novel samples can then be generated via iterative sampling from " + }, + { + "bbox": [ + 305, + 127, + 546, + 152 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x}_{t-1} \\mid \\boldsymbol{x}_t)" + }, + { + "bbox": [ + 305, + 127, + 546, + 152 + ], + "type": "text", + "content": " following:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 341, + 159, + 511, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 159, + 511, + 175 + ], + "spans": [ + { + "bbox": [ + 341, + 159, + 511, + 175 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t - 1} = \\mu_ {\\theta} (\\mathbf {x} _ {t}, t) + \\sqrt {\\beta_ {t}} \\mathbf {z}, \\mathbf {z} \\sim \\mathcal {N} (0, \\mathbf {I}).", + "image_path": "5f22a8d2d9c60f4e055cc8751e55ca3ce0f023839c43b203498b86abc672697f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 183, + 547, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 183, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 305, + 183, + 547, + 231 + ], + "type": "text", + "content": "Diffusion models have shown impressive capabilities in generating high-quality and diverse content. This paper proposes a texture generation framework based on diffusion model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 238, + 463, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 238, + 463, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 238, + 463, + 251 + ], + "type": "text", + "content": "3.2. UV mapping and challenges." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 256, + 546, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 256, + 546, + 341 + ], + "spans": [ + { + "bbox": [ + 305, + 256, + 546, + 341 + ], + "type": "text", + "content": "UV mapping is a method of surface parameterization that translates a 3D surface into a 2D image, effectively creating a 2D coordinate system known as a UV map for a polygonal mesh " + }, + { + "bbox": [ + 305, + 256, + 546, + 341 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 256, + 546, + 341 + ], + "type": "text", + "content": ". This is achieved by explicitly assigning UV coordinates to each vertex of the mesh. Further, an arbitrary surface coordinate on the mesh can be mapped to its 2D coordinate through barycentric interpolation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 368, + 350, + 546, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 350, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 368, + 350, + 546, + 363 + ], + "type": "interline_equation", + "content": "(u, v) = f (p) \\quad f: S \\rightarrow \\Omega , \\tag {2}", + "image_path": "5f7de6e1d41550b6179e66cfa8a258d539ccdd44798a5951f29919405b211df3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 372, + 489, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 489, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 489, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 372, + 489, + 384 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 372, + 489, + 384 + ], + "type": "text", + "content": " represents the UV mapping process." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 384, + 547, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 576 + ], + "type": "text", + "content": "As illustrated in Figure 3 \"UV warping\", we can apply textures to the mesh surface by generating a high-resolution texture image in the UV space. Besides texture, we can also create this kind of 2D map for surface normals and point coordinates. Our objective is to generate a high-quality 2D UV texture image of the given mesh. However, the mapping process requires cutting the continuous texture on the 3D shape into a series of individual patches in the 2D UV plane, as depicted in Figure 3. This fragmentation makes it challenging for the generative model to directly learn the 3D adjacency relationships of the patches within the 2D texture UV map. Consequently, this can lead to discontinuity and inconsistency issues when the generated texture map is applied back to the 3D mesh surface. As shown in Figure 2 \"2D Diffusion\", the diffusion model generates inconsistent textures and suffers from discontinuity issues." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 586, + 362, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 362, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 362, + 597 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "The challenge mentioned in Section 3.2 motivates us to propose a coarse-to-fine framework for texture image synthesis, namely Point-UV diffusion, illustrated in Figure 4. To start, we design a 3D point diffusion model to colorize a set of sampled points on the mesh surface, as shown in Figure 4 (Top). This stage leverages the 3D topology for predicting low-frequency colors on the mesh, without being affected by the discontinuity of the UV map. Based on the color components generated by the coarse stage, we then" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4208" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 74, + 541, + 203 + ], + "blocks": [ + { + "bbox": [ + 53, + 74, + 541, + 203 + ], + "lines": [ + { + "bbox": [ + 53, + 74, + 541, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 74, + 541, + 203 + ], + "type": "image", + "image_path": "03bdf3a5f17d771afd0ee6fb532bea7a0253e87315c91a4981dc2a4aed1ff5a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 206, + 541, + 313 + ], + "blocks": [ + { + "bbox": [ + 53, + 206, + 541, + 313 + ], + "lines": [ + { + "bbox": [ + 53, + 206, + 541, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 541, + 313 + ], + "type": "image", + "image_path": "1df90fc3e88730a3a8b4a0cc36e39f51d1602f2414a9d3442ad1d525fe477bcc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 319, + 548, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 548, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 548, + 376 + ], + "type": "text", + "content": "Figure 4. The overview of our Point-UV-diffusion framework. (Top) The coarse stage first samples a point cloud on the mesh surface, and then predicts the color for each point using a 3D diffusion model conditioned on shape features including surface normal, mask, and coordinates. Then the points are mapped to the 2D UV map and the remaining uncolored points are filled up by tri-linear interpolation based on the 3D coordinates. (Bottom) The fine stage predicts high-quality textures with a 2D diffusion model conditioned on the shape attributes and the coarse texture image." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 394, + 288, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 288, + 431 + ], + "type": "text", + "content": "establish a 2D diffusion model in the UV space. This enhances the fidelity of the generated texture, as depicted in Figure 4 (Bottom)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 437, + 211, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 211, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 211, + 451 + ], + "type": "text", + "content": "4.1. Coarse Stage: Point Diffusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "text", + "content": "In the coarse stage, we begin by executing farthest point sampling (FPS) on the mesh surface, deriving a point set consisting of " + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "text", + "content": " points. These points are defined by their coordinates " + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{coord}}" + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "text", + "content": " and colors " + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "text", + "content": ". During training, a forward diffusion process degrades the clean colors " + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "text", + "content": ", transforming them into a noisy state " + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "text", + "content": ". The noise level is dictated by the time step " + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "inline_equation", + "content": "t \\in \\{0,1,\\dots,T\\}" + }, + { + "bbox": [ + 46, + 456, + 287, + 540 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 540, + 288, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 288, + 625 + ], + "type": "text", + "content": "Our network is trained to reverse this diffusion process, aiming to denoise " + }, + { + "bbox": [ + 46, + 540, + 288, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 46, + 540, + 288, + 625 + ], + "type": "text", + "content": " back to its original clean colors. This denoising network is informed by three pre-computed maps: a coordinate map " + }, + { + "bbox": [ + 46, + 540, + 288, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{coord}}" + }, + { + "bbox": [ + 46, + 540, + 288, + 625 + ], + "type": "text", + "content": ", a normal map " + }, + { + "bbox": [ + 46, + 540, + 288, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 46, + 540, + 288, + 625 + ], + "type": "text", + "content": ", and a mask map " + }, + { + "bbox": [ + 46, + 540, + 288, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 46, + 540, + 288, + 625 + ], + "type": "text", + "content": " (details in Section 3.2). Unless stated otherwise, these conditions concatenate along the channel dimension, culminating in what we term the shape map " + }, + { + "bbox": [ + 46, + 540, + 288, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{shape}}" + }, + { + "bbox": [ + 46, + 540, + 288, + 625 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 98, + 635, + 287, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 635, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 98, + 635, + 287, + 649 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\text {s h a p e}} = \\left(\\left[ \\mathbf {x} _ {\\text {n o r m a l}}, \\mathbf {x} _ {\\text {m a s k}}, \\mathbf {x} _ {\\text {c o o r d}} \\right]\\right). \\tag {3}", + "image_path": "6f1683025e42bb1762ef125947f21337756aae3710f0a587ff6ee9047a157868.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Our network architecture is constructed upon PVCNN [23], drawing similarities to point-voxel diffusion [49]. However, we introduce slight modifications to amplify the integration of global shape information, contrasting with the approach in [49] which primarily relies on point coordinates. Ini" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "text", + "content": "tially, we employ a lightweight shape encoder " + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "inline_equation", + "content": "E_{\\phi}" + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "text", + "content": " to extract a global shape embedding " + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "inline_equation", + "content": "f_{g}" + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{shape}}" + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "text", + "content": ". This embedding is subsequently fed into the 3D network " + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "inline_equation", + "content": "G_{\\theta_1}^1" + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "text", + "content": " along with " + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{coord}}" + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "text", + "content": " to predict the color " + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_0" + }, + { + "bbox": [ + 305, + 394, + 546, + 443 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 366, + 446, + 545, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 446, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 366, + 446, + 545, + 465 + ], + "type": "interline_equation", + "content": "f _ {g} = E _ {\\phi} \\left(\\mathbf {x} _ {\\text {s h a p e}}\\right), \\tag {4}", + "image_path": "ee55e989a33557f5973702a17c24905dd4b6a7cad44b2cc05a0dc0de9bb354c0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 367, + 462, + 484, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 462, + 484, + 476 + ], + "spans": [ + { + "bbox": [ + 367, + 462, + 484, + 476 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {0} = G _ {\\theta_ {1}} ^ {1} \\left(\\left[ \\mathbf {z} _ {\\mathrm {c o o r d}}, \\mathbf {z} _ {t}, f _ {g}, t \\right]\\right).", + "image_path": "69e644e24ac0174a4d3939ec71b99557d73e594a572f3c1c915448c3d92196b9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 488, + 546, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 546, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 546, + 668 + ], + "type": "text", + "content": "Style guidance. We observe that the synthesized colors are largely influenced by the predominant colors within the dataset (for instance, textures in the ShapeNet \"chair\" category are typically white, pure magenta, or wood-colored), leading to a lack of diversity in the outputs. To address this bias and promote color diversity, we introduce a style guidance mechanism. This is achieved by flattening each " + }, + { + "bbox": [ + 304, + 488, + 546, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 304, + 488, + 546, + 668 + ], + "type": "text", + "content": " into a unidimensional vector, followed by employing PCA to extract the principal component coefficients, thus reducing the dimensionality of " + }, + { + "bbox": [ + 304, + 488, + 546, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 304, + 488, + 546, + 668 + ], + "type": "text", + "content": ". Subsequently, K-means clustering is utilized to assign a style label to each " + }, + { + "bbox": [ + 304, + 488, + 546, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 304, + 488, + 546, + 668 + ], + "type": "text", + "content": ". As depicted in Figure 5, shapes within a particular cluster exhibit similar color styles. During training, we provide the network with an additional style label " + }, + { + "bbox": [ + 304, + 488, + 546, + 668 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{style}}" + }, + { + "bbox": [ + 304, + 488, + 546, + 668 + ], + "type": "text", + "content": " as a condition, which is referred to as style guidance:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 354, + 671, + 545, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 671, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 354, + 671, + 545, + 687 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {0} = G _ {\\theta_ {1}} ^ {1} \\left(\\left[ \\mathbf {z} _ {\\text {c o o r d}}, \\mathbf {z} _ {t}, f _ {g}, t, z _ {\\text {s t y l e}} \\right]\\right). \\tag {5}", + "image_path": "4a152750c86443505ab61eaa11b07b6f8c9ead4eb9f7a5e24082fd7cef041ad6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 546, + 714 + ], + "type": "text", + "content": "In this way, we can guide the network to predict the desired color during inference by providing a certain style condi" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4209" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 235, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 235, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 235, + 83 + ], + "type": "text", + "content": "tion, thereby alleviating the biased color issue." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 107, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 107, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 107, + 289, + 228 + ], + "type": "text", + "content": "Coarse texture image. With the colored point clouds, we then project them onto the 2D UV space based on the pre-calculated UV mapping. Following this, we perform KNN interpolation based on the 3D coordinates to assign colors to the remaining uncolored pixels, thereby generating a coarse texture image " + }, + { + "bbox": [ + 46, + 107, + 289, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{coarse}}" + }, + { + "bbox": [ + 46, + 107, + 289, + 228 + ], + "type": "text", + "content": " for the given mesh. This coarse texture image offers a coherent base color initialization and functions as a conditional element to guide the high-fidelity texture generation in the second stage, helping to avoid discontinuity caused by UV mapping." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 239, + 188, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 188, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 188, + 251 + ], + "type": "text", + "content": "4.2. Fine Stage: UV Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 258, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 287, + 344 + ], + "type": "text", + "content": "To further refine the coarse texture image, we design a fine stage using 2D diffusion in the UV space, as depicted in Figure 4 (Bottom). In addition to the conditions used in the coarse stage, we incorporate the coarse texture image " + }, + { + "bbox": [ + 47, + 258, + 287, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{coarse}}" + }, + { + "bbox": [ + 47, + 258, + 287, + 344 + ], + "type": "text", + "content": " as an additional condition. We apply a 2D U-Net " + }, + { + "bbox": [ + 47, + 258, + 287, + 344 + ], + "type": "inline_equation", + "content": "G_{\\theta_2}^2" + }, + { + "bbox": [ + 47, + 258, + 287, + 344 + ], + "type": "text", + "content": " combined with self-attention modules to learn the high-quality texture image " + }, + { + "bbox": [ + 47, + 258, + 287, + 344 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 47, + 258, + 287, + 344 + ], + "type": "text", + "content": ", from a noisy texture image " + }, + { + "bbox": [ + 47, + 258, + 287, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 47, + 258, + 287, + 344 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 354, + 287, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 354, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 99, + 354, + 287, + 369 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {0} = G _ {\\theta_ {2}} ^ {2} \\left(\\left[ \\mathbf {x} _ {\\text {s h a p e}}, \\mathbf {x} _ {t}, \\mathbf {x} _ {\\text {c o a r s e}}, t \\right]\\right). \\tag {6}", + "image_path": "abd3ddb97de2ce5a06338f16528cafead138491cac0038b20c4f6bc4e24b5fb0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "content": "Hybrid condition. During training, we employ FPS on the ground-truth texture image, " + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "content": ", followed by interpolation to simulate the coarse texture image, " + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{coarse}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "content": ". However, in joint cascaded testing involving both stages, we note that the output quality of " + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\mathrm{coarse}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "content": " from the first stage doesn't always align flawlessly with the quality of " + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{coarse}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "content": ". Such mismatches can influence the performance of the subsequent stage. To address this, we introduce a hybrid conditioning method aimed at narrowing this discrepancy. Firstly, we create a smooth texture image " + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "content": " (see Figure 6), inspired by the blur augmentation described in [17] for cascaded diffusion models. In particular, we segment the mask map into multiple discrete regions using four-connectivity detection, then perform average color pooling within each region based on its connectivity. After this procedure, " + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "content": " maintains merely the regional color, ensuring more consistent alignment across both training and testing phases. Then, we combine " + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{coarse}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "content": ", using a certain probability " + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{hybrid}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "content": " during training. Thus, the network is forced to be capable of generating textures even when adopting a weaker condition " + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 714 + ], + "type": "text", + "content": " in the fine stage during inference. Considering that the diffusion model first generates low-frequency information and then higher one [8], we also explore a condition-truncated sampling, as detailed discussed in Section 5.4, where we condition both maps for generation during the initial time steps and then exclusively utilize the smooth map for the remainder of the generation." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 73, + 410, + 171 + ], + "blocks": [ + { + "bbox": [ + 312, + 73, + 410, + 171 + ], + "lines": [ + { + "bbox": [ + 312, + 73, + 410, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 410, + 171 + ], + "type": "image", + "image_path": "78e68ad0434138d143723c3880402adbd98827bbb432950edb198cb32e5506dc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 173, + 384, + 181 + ], + "lines": [ + { + "bbox": [ + 321, + 173, + 384, + 181 + ], + "spans": [ + { + "bbox": [ + 321, + 173, + 384, + 181 + ], + "type": "text", + "content": "(a) Samples from cluster 1" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 415, + 73, + 541, + 171 + ], + "blocks": [ + { + "bbox": [ + 415, + 73, + 541, + 171 + ], + "lines": [ + { + "bbox": [ + 415, + 73, + 541, + 171 + ], + "spans": [ + { + "bbox": [ + 415, + 73, + 541, + 171 + ], + "type": "image", + "image_path": "355323337c1383ea88af947ed41b22f456a8e2ee1f29a46adba57c4c190866bd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 173, + 503, + 181 + ], + "lines": [ + { + "bbox": [ + 438, + 173, + 503, + 181 + ], + "spans": [ + { + "bbox": [ + 438, + 173, + 503, + 181 + ], + "type": "text", + "content": "(d) Samples from cluster 4" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 240, + 384, + 314 + ], + "blocks": [ + { + "bbox": [ + 305, + 184, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 305, + 184, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 184, + 545, + 228 + ], + "type": "text", + "content": "Figure 5. Illustration of samples across various clusters. Shapes within a particular cluster exhibit analogous color styles. However, there exists an imbalance in the quantity of shapes among different clusters, leading to challenges in unbiased synthesis." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 240, + 384, + 314 + ], + "lines": [ + { + "bbox": [ + 312, + 240, + 384, + 314 + ], + "spans": [ + { + "bbox": [ + 312, + 240, + 384, + 314 + ], + "type": "image", + "image_path": "b80f5fa28dbd811834d417bd7fee31866b25e4db8cdd6595ad89ac370118ef38.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 316, + 372, + 324 + ], + "lines": [ + { + "bbox": [ + 321, + 316, + 372, + 324 + ], + "spans": [ + { + "bbox": [ + 321, + 316, + 372, + 324 + ], + "type": "text", + "content": "(a) Texture image" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 328, + 545, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 328, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 328, + 545, + 361 + ], + "type": "text", + "content": "Figure 6. Varieties of texture images. (a) Original texture image enriched with high-frequency details, (b) Coarse texture image, and (c) Smooth texture image." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 389, + 240, + 460, + 314 + ], + "blocks": [ + { + "bbox": [ + 389, + 240, + 460, + 314 + ], + "lines": [ + { + "bbox": [ + 389, + 240, + 460, + 314 + ], + "spans": [ + { + "bbox": [ + 389, + 240, + 460, + 314 + ], + "type": "image", + "image_path": "d2b50a6733fa74d83efc1a693d73306df119eb23950b24af615cd5fd3178b772.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 316, + 459, + 324 + ], + "lines": [ + { + "bbox": [ + 389, + 316, + 459, + 324 + ], + "spans": [ + { + "bbox": [ + 389, + 316, + 459, + 324 + ], + "type": "text", + "content": "(b) Coarse texture image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 468, + 240, + 541, + 314 + ], + "blocks": [ + { + "bbox": [ + 468, + 240, + 541, + 314 + ], + "lines": [ + { + "bbox": [ + 468, + 240, + 541, + 314 + ], + "spans": [ + { + "bbox": [ + 468, + 240, + 541, + 314 + ], + "type": "image", + "image_path": "8eaabc9f4446cc5dfd47c1d7b23a300f7a4de380b347cc2d3005cd19a031a678.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 316, + 539, + 324 + ], + "lines": [ + { + "bbox": [ + 468, + 316, + 539, + 324 + ], + "spans": [ + { + "bbox": [ + 468, + 316, + 539, + 324 + ], + "type": "text", + "content": "(c) Smooth texture image" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 381, + 387, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 387, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 387, + 395 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 401, + 501, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 401, + 501, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 501, + 415 + ], + "type": "text", + "content": "5.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "text", + "content": "We conduct experiments across four categories of the ShapeNet dataset [4], namely, chair, table, car, and bench. Before training, we use an open-source UV-Atlas tool [46] to generate the UV map and pre-process the dataset to obtain the shape maps and ground-truth texture images. We sample " + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "inline_equation", + "content": "K = 4096" + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "text", + "content": " points in our coarse stage and synthesize a texture image in the UV space with a resolution of " + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "text", + "content": " for the fine stage. For training the diffusion models, akin to [1], we predict the clean signals. This approach provides more stable training than predicting the noise component as recommended by [16]. We employ the cosine noise scheduling [27] ranging from 0.0001 to 0.02 over 1,024 time steps for both stages. In the fine stage, we leverage the noise scaling strategy from [7] and also incorporate a rendering loss. Specifically, we randomly select four views and render the mesh using the predicted UV map and the ground-truth UV map to produce " + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "text", + "content": " images. Subsequently, we crop " + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "text", + "content": " patches from these images and compute the corresponding " + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "text", + "content": " loss. For the hybrid condition in the fine stage, we use " + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{hybrid}} = 0.3" + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "text", + "content": " and sweep over condition-truncated time " + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "inline_equation", + "content": "t_{c}" + }, + { + "bbox": [ + 304, + 420, + 545, + 673 + ], + "type": "text", + "content": " (see Figure 12)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Compared methods. We compare our method with state-of-the-art approaches, including Texturify [38], Texture" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4210" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 76, + 531, + 277 + ], + "blocks": [ + { + "bbox": [ + 65, + 76, + 531, + 277 + ], + "lines": [ + { + "bbox": [ + 65, + 76, + 531, + 277 + ], + "spans": [ + { + "bbox": [ + 65, + 76, + 531, + 277 + ], + "type": "image", + "image_path": "86b797e91ed0ca00ebed9a3cc70e0b93e0e6164bedfeaae3d8a5080965b9cba0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 286, + 547, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 286, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 547, + 310 + ], + "type": "text", + "content": "Figure 7. Our texture synthesis results in different categories. The results on chairs, tables, cars, and benches demonstrate that our approach can generate natural, vivid, and diverse textures, even though the given 3D shapes are challenging, such as cutouts on the bench." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 328, + 289, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 289, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 289, + 426 + ], + "type": "text", + "content": "Fields [29], and PVD [49], in the context of unconditional texture generation. PVD was originally designed for point cloud generation instead of texture generation, so it cannot be directly compared with our approach. To facilitate comparison, we modify this framework to learn RGB values from the sampled point cloud. This extension, referred to as \"PVD-Tex\", serves as a baseline for directly learning texture in the point space using the diffusion model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 433, + 231, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 433, + 231, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 433, + 231, + 445 + ], + "type": "text", + "content": "5.2. Unconditional Texture Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 453, + 288, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 288, + 536 + ], + "type": "text", + "content": "Gallary. First, we showcase our texture generation results for each category without additional conditions. The results in Figure 7 and Figure 1 demonstrate the remarkable performance of our method, which is able to generate appealing textures with fine details and preserve the geometric structures. Note that our method is compatible with meshes of diverse topology and intricate geometric details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 552, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 673 + ], + "type": "text", + "content": "Qualitative comparisons. As depicted in Figure 8, our approach excels at generating high-quality textures while maintaining the geometric intricacies of the input mesh. Firstly, the results of Texture Fields [29] appear deficient in high-frequency details. Additionally, while PVD-Tex [49] is able to produce spatially varying colors to some extent (e.g., the car in Figure 8), it falls short in synthesizing intricate high-frequency details. Lastly, even though Texturify [38] demonstrates an ability to generate finer textures, it compromises on preserving slender structures (i.e. chair)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Quantitative comparisons. To quantitatively compare with existing works, we follow [38] and assess the genera" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 328, + 547, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 547, + 426 + ], + "type": "text", + "content": "tion quality using Frechet Inception Distance (FID)[15] and Kernel Inception Distance (KID)[2], metrics that are widely used for evaluating image generation models. To this end, we render " + }, + { + "bbox": [ + 304, + 328, + 547, + 426 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 328, + 547, + 426 + ], + "type": "text", + "content": " images from each generated textured mesh and ground-truth textured mesh using four distinct camera views. Table 1 presents the quantitative comparisons with current approaches, revealing that our method surpasses existing works." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 432, + 479, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 432, + 479, + 444 + ], + "spans": [ + { + "bbox": [ + 305, + 432, + 479, + 444 + ], + "type": "text", + "content": "5.3. Conditional Texture Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "In addition, we demonstrate the capability of our framework to synthesize textures conditioned on either text prompts or a single-view image. We conduct our experiments on the chair and table categories. For the text condition, we utilize text-shape pairs as provided in [6] (with additional corrections for text accuracy). For the image condition, we randomly render a view from the ground-truth mesh. To infuse the network with condition-specific information, we use the pre-trained vision-language model CLIP [33] to extract the corresponding embedding from either the image or the text. This embedding is then fed into a simple MLP to incorporate the information into the diffusion model as a condition for both training and inference. As depicted in Figure 9, our method succeeds in generating textures that align well with the given text descriptions or images. We also compare our approach with Text2Mesh [24], a test-time optimization method. Text2Mesh takes around 10 minutes per instance, while ours only requires 30 seconds. Importantly, to adapt to our task which requires preserving the mesh geometry, we freeze the geometry deformation branch of Text2Mesh to generate only colors. As shown in Figure 10, Text2Mesh" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4211" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 77, + 282, + 351 + ], + "blocks": [ + { + "bbox": [ + 55, + 77, + 282, + 351 + ], + "lines": [ + { + "bbox": [ + 55, + 77, + 282, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 77, + 282, + 351 + ], + "type": "image", + "image_path": "fc1e5f50344813ec6fab5477aca5d6115d6fa928b654080272d06ad0cfea7468.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 357, + 288, + 425 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 288, + 425 + ], + "type": "text", + "content": "Figure 8. Qualitative comparisons with existing works. Given pure 3D shapes, Texturify [38] generates textures on the surface of 3D shape but tends to damage the topology; Texture fields [29] and PVD-Tex [49] produce textures with limited details. On the contrary, our approach faithfully preserves topology and produces realistic appearances." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 444, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 480 + ], + "type": "text", + "content": "is bound by vertex resolution, while our method can deliver detailed visuals even on low-resolution meshes, offering a distinct advantage in graphics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 487, + 148, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 148, + 499 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 148, + 499 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 506, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 662 + ], + "type": "text", + "content": "Coarse-to-fine diffusion. To manifest the effectiveness of our coarse-to-fine diffusion strategy, we conduct experiments on two baselines \"w/o coarse stage\" and \"w/o fine stage\". The \"w/o coarse stage\" configuration indicates that we directly generate the UV texture image using the fine stage, bypassing the initialization from the coarse stage. The \"w/o fine stage\" configuration signifies the result of the coarse stage, after assigning colors to the uncolored pixels using KNN interpolation. In both scenarios, the model produces inferior outcomes relative to our full model, as shown in Table 2 and Figure 11. Absent the coarse stage, the generated result suffers from noticeably inconsistent colors. Without the fine stage, the results are over-smoothed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Hybrid condition. As shown in Table 2 and Figure 11 \"coarse condition\", the fine stage cannot generate high-quality textures when it is exclusively conditioned on the" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 313, + 72, + 544, + 417 + ], + "blocks": [ + { + "bbox": [ + 313, + 72, + 544, + 417 + ], + "lines": [ + { + "bbox": [ + 313, + 72, + 544, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 544, + 417 + ], + "type": "image", + "image_path": "d65f2fe611a112f9326e484b0374fcb845e223f7ce4fafd1259ec5be1d3b39c7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 421, + 545, + 455 + ], + "lines": [ + { + "bbox": [ + 305, + 421, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 305, + 421, + 545, + 455 + ], + "type": "text", + "content": "Figure 9. Results of conditional texture generation. Our method is adaptable to craft textures guided by text descriptions (rows 1-3) or single-view images (rows 4-6)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 326, + 464, + 531, + 574 + ], + "blocks": [ + { + "bbox": [ + 326, + 464, + 531, + 574 + ], + "lines": [ + { + "bbox": [ + 326, + 464, + 531, + 574 + ], + "spans": [ + { + "bbox": [ + 326, + 464, + 531, + 574 + ], + "type": "image", + "image_path": "9697bacf708f3c7ee5d0feeb1e4398bbe83b2a70456554c5d10c05e5e5f8cbcb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 576, + 545, + 609 + ], + "lines": [ + { + "bbox": [ + 305, + 576, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 305, + 576, + 545, + 609 + ], + "type": "text", + "content": "Figure 10. Comparison with Text2Mesh. We freeze the geometry deformation branch of Text2Mesh to adapt to our task. Ours can generate high-frequency details for low-resolution mesh." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": "coarse map. Further, the generated quality remains unsatisfactory if we apply a hybrid condition entirely during training (i.e., " + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{hybrid}} = 1.0" + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": "), as shown in \" " + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{hybrid}} = 1.0" + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": "\". We attribute this to the network's propensity to depend solely on the coarse map during training and generation. In contrast, our full model (i.e., hybrid condition with " + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{hybrid}} = 0.3" + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": ") achieves significant improvement both" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4212" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 104, + 70, + 491, + 145 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 491, + 145 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 491, + 145 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 491, + 145 + ], + "type": "table", + "html": "
ChairCarTableBench
MethodsFID ↓KID ↓FID ↓KID ↓FID ↓KID ↓FID ↓KID ↓
Texture Fields [29]24.241.07156.3813.6468.964.2062.712.96
Texturify [38]27.801.3273.164.71----
PVD-Tex [49]15.520.6259.473.7416.120.5528.940.39
Ours9.880.2226.890.689.630.1523.090.15
", + "image_path": "87145427ccf2b3f6d2b01820679a7938dc1e30769356d617c5a32e0980749cdd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 78, + 178, + 127, + 264 + ], + "blocks": [ + { + "bbox": [ + 78, + 178, + 127, + 264 + ], + "lines": [ + { + "bbox": [ + 78, + 178, + 127, + 264 + ], + "spans": [ + { + "bbox": [ + 78, + 178, + 127, + 264 + ], + "type": "image", + "image_path": "84a3def2507621a83f179e2207ec41d7871a84169839c3e98f7ea23eb33bf6cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 267, + 137, + 276 + ], + "lines": [ + { + "bbox": [ + 78, + 267, + 137, + 276 + ], + "spans": [ + { + "bbox": [ + 78, + 267, + 137, + 276 + ], + "type": "text", + "content": "(a) untextured mesh" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 145, + 178, + 194, + 264 + ], + "blocks": [ + { + "bbox": [ + 145, + 178, + 194, + 264 + ], + "lines": [ + { + "bbox": [ + 145, + 178, + 194, + 264 + ], + "spans": [ + { + "bbox": [ + 145, + 178, + 194, + 264 + ], + "type": "image", + "image_path": "de8ba1fb2bbc89c540f09921c6de151892921d8d4601c509a949c7e569b92b51.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 266, + 208, + 276 + ], + "lines": [ + { + "bbox": [ + 148, + 266, + 208, + 276 + ], + "spans": [ + { + "bbox": [ + 148, + 266, + 208, + 276 + ], + "type": "text", + "content": "(b) w/o coarse stage" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 211, + 178, + 260, + 264 + ], + "blocks": [ + { + "bbox": [ + 211, + 178, + 260, + 264 + ], + "lines": [ + { + "bbox": [ + 211, + 178, + 260, + 264 + ], + "spans": [ + { + "bbox": [ + 211, + 178, + 260, + 264 + ], + "type": "image", + "image_path": "982320679431f2013186f29101ba9df8eea6815de214fb528438f7143ddba134.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 266, + 267, + 276 + ], + "lines": [ + { + "bbox": [ + 215, + 266, + 267, + 276 + ], + "spans": [ + { + "bbox": [ + 215, + 266, + 267, + 276 + ], + "type": "text", + "content": "(c) w/o fine stage" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 80, + 278, + 127, + 361 + ], + "blocks": [ + { + "bbox": [ + 80, + 278, + 127, + 361 + ], + "lines": [ + { + "bbox": [ + 80, + 278, + 127, + 361 + ], + "spans": [ + { + "bbox": [ + 80, + 278, + 127, + 361 + ], + "type": "image", + "image_path": "fa9d3f85ce3244cde44c54ba778651e23f9615245bc9f5138883c8ace80c0938.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 368, + 137, + 377 + ], + "lines": [ + { + "bbox": [ + 78, + 368, + 137, + 377 + ], + "spans": [ + { + "bbox": [ + 78, + 368, + 137, + 377 + ], + "type": "text", + "content": "(d) coarse condition" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 148, + 279, + 196, + 363 + ], + "blocks": [ + { + "bbox": [ + 148, + 279, + 196, + 363 + ], + "lines": [ + { + "bbox": [ + 148, + 279, + 196, + 363 + ], + "spans": [ + { + "bbox": [ + 148, + 279, + 196, + 363 + ], + "type": "image", + "image_path": "b14609cb0ea646c3e293935110cac2f5b0df18528fa22ebf0fcb096eaa07d55d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 368, + 203, + 377 + ], + "lines": [ + { + "bbox": [ + 151, + 368, + 203, + 377 + ], + "spans": [ + { + "bbox": [ + 151, + 368, + 203, + 377 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 151, + 368, + 203, + 377 + ], + "type": "inline_equation", + "content": "p_{hybrid} = 1.0" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 213, + 279, + 261, + 364 + ], + "blocks": [ + { + "bbox": [ + 213, + 279, + 261, + 364 + ], + "lines": [ + { + "bbox": [ + 213, + 279, + 261, + 364 + ], + "spans": [ + { + "bbox": [ + 213, + 279, + 261, + 364 + ], + "type": "image", + "image_path": "ca8050151c919536b4463c759fad95ce7b7bb56b3ff226b18a716ccd92f8bf8b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 369, + 251, + 377 + ], + "lines": [ + { + "bbox": [ + 228, + 369, + 251, + 377 + ], + "spans": [ + { + "bbox": [ + 228, + 369, + 251, + 377 + ], + "type": "text", + "content": "(f) ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 381, + 287, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 415 + ], + "type": "text", + "content": "Figure 11. Ablation studies. Without our two-stage generation pipeline or hybrid condition designs, the generated results are inconsistent with the 3D shape or lack of high-frequency details." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 89, + 424, + 243, + 499 + ], + "blocks": [ + { + "bbox": [ + 47, + 145, + 530, + 157 + ], + "lines": [ + { + "bbox": [ + 47, + 145, + 530, + 157 + ], + "spans": [ + { + "bbox": [ + 47, + 145, + 530, + 157 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with existing works. Ours outperforms other approaches on both FID [15] and KID " + }, + { + "bbox": [ + 47, + 145, + 530, + 157 + ], + "type": "inline_equation", + "content": "(\\times 10^{2})" + }, + { + "bbox": [ + 47, + 145, + 530, + 157 + ], + "type": "text", + "content": " [2]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 424, + 243, + 499 + ], + "lines": [ + { + "bbox": [ + 89, + 424, + 243, + 499 + ], + "spans": [ + { + "bbox": [ + 89, + 424, + 243, + 499 + ], + "type": "table", + "html": "
FID ↓KID ↓
w/o fine stage17.880.76
w/o coarse stage15.110.49
coarse condition14.930.56
\\(p_{hybrid}=1.0\\)15.250.59
Ours9.880.22
", + "image_path": "8b00a2e0803c4dc49d128bfa2856df111f68a6da9652301dabdecdef83f95cf3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 542, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 288, + 651 + ], + "type": "text", + "content": "qualitatively and quantitatively, as presented in Table 2 and Figure 11 \"ours\". Furthermore, we also sweep over the effect of condition-truncated time " + }, + { + "bbox": [ + 46, + 542, + 288, + 651 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 46, + 542, + 288, + 651 + ], + "type": "text", + "content": " for inference. As Figure 12 shows, " + }, + { + "bbox": [ + 46, + 542, + 288, + 651 + ], + "type": "inline_equation", + "content": "t_c = 0.4" + }, + { + "bbox": [ + 46, + 542, + 288, + 651 + ], + "type": "text", + "content": " strikes a sweet point. The model utilizes information from both conditions maps to better generate low-frequency components during the first " + }, + { + "bbox": [ + 46, + 542, + 288, + 651 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 542, + 288, + 651 + ], + "type": "text", + "content": " of the sampling timesteps. After that, the coarse map " + }, + { + "bbox": [ + 46, + 542, + 288, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{coarse}}" + }, + { + "bbox": [ + 46, + 542, + 288, + 651 + ], + "type": "text", + "content": " is dropped, and the model further focuses on fine-grained detail generation without relying on the coarse map." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Style guidance. The style guidance is aimed at addressing the issue of insufficient diversity due to color distribution bias in the dataset. As a result, we do not utilize FID or KID for evaluation since they evaluate the distribution sim" + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 328, + 174, + 522, + 236 + ], + "blocks": [ + { + "bbox": [ + 46, + 500, + 287, + 522 + ], + "lines": [ + { + "bbox": [ + 46, + 500, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 287, + 522 + ], + "type": "text", + "content": "Table 2. Ablation studies. This table shows the effectiveness of each component in our proposed method." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 174, + 522, + 236 + ], + "lines": [ + { + "bbox": [ + 328, + 174, + 522, + 236 + ], + "spans": [ + { + "bbox": [ + 328, + 174, + 522, + 236 + ], + "type": "table", + "html": "
MethodsPreference↑LPIPS↑
Ours49.8%0.083
Texturify [38]15.9%0.086
PVD-Tex [49]29.2%0.029
Texture Fields [29]5.1%0.005
", + "image_path": "ff88d7cbada0b4dfd13deae9aa6e0f98f5afe296871f23f4d2367ad3d83af73d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 237, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 306, + 237, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 306, + 237, + 545, + 259 + ], + "type": "text", + "content": "Table 3. LPIPS and user study. This table shows the average LPIPS and preference via user study in the chair category." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 281, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 533 + ], + "type": "text", + "content": "The diversity of the LPIPS is a key feature that is crucial to the quality of the generated results and ground truth. However, the ground-truth distribution shows a strong bias towards particular colors, making these metrics inappropriate for assessing generative texture diversity. In contrast, we adopt LPIPS [48] to measure the pairwise similarity among five textures generated by our method given the same input mesh, where a larger diversity in textures will lead to a higher LPIPS value. In Figure 13, we report the quantitative results for 500 shapes. Our approach achieves a higher LPIPS in most of the evaluated cases, indicating better diversity. Besides, as shown in Figure 14, we present the results of generating three textures randomly for the shapes without and with style guidance. In the latter case, we uniformly sample three style labels as style inputs. It is evident that without style guidance, the generated textures are nearly identical, and the color styles of different shapes tend to be similar. With the introduction of style guidance, however, the diversity of colors has significantly increased. We also conduct a comparison with other methods in terms of diversity, as well as a quality assessment through user studies, as shown in Table 3." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 547, + 531, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 547, + 531, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 547, + 531, + 560 + ], + "type": "text", + "content": "6. Discussion, Limitations, and Conclusions" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 568, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 688 + ], + "type": "text", + "content": "This paper presents Point-UV diffusion, a brand-new framework that employs a coarse-to-fine pipeline to generate textures for 3D meshes. We begin with a 3D diffusion model to synthesize low-frequency texture components from point clouds, which maintain 3D consistency. We then refine the textures using a 2D UV-space diffusion model. Our method is compatible with meshes with arbitrary topology and can faithfully preserve the geometry structure. We further demonstrate the flexibility of our framework by extending it to conditional generative models." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Despite its merits, our method has inherent limitations. Similar to other methodologies relying on 3D data for train-" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4213" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 75, + 271, + 190 + ], + "blocks": [ + { + "bbox": [ + 63, + 75, + 271, + 190 + ], + "lines": [ + { + "bbox": [ + 63, + 75, + 271, + 190 + ], + "spans": [ + { + "bbox": [ + 63, + 75, + 271, + 190 + ], + "type": "image", + "image_path": "5fd0d599729542c9b7c974173dd1ac7d5e71b924af103adc5dced05045c12332.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 288, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 288, + 251 + ], + "type": "text", + "content": "Figure 12. Examination of condition-truncated time. The horizontal axis denotes the percentage of time points when the coarse texture image is conditioned during inference. The left vertical axis indicates the FID value, while the right vertical axis shows the KID value " + }, + { + "bbox": [ + 46, + 198, + 288, + 251 + ], + "type": "inline_equation", + "content": "(\\times 10^{2})" + }, + { + "bbox": [ + 46, + 198, + 288, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 261, + 272, + 419 + ], + "blocks": [ + { + "bbox": [ + 63, + 261, + 272, + 419 + ], + "lines": [ + { + "bbox": [ + 63, + 261, + 272, + 419 + ], + "spans": [ + { + "bbox": [ + 63, + 261, + 272, + 419 + ], + "type": "image", + "image_path": "06390beb6f5eec44e05400f76334549a09b8030b3ba6da06b328c2c9b0ab768a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 425, + 287, + 448 + ], + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 448 + ], + "type": "text", + "content": "Figure 13. Diversity measurement. Utilizing style guidance leads to larger LPIPS scores, indicating enhanced generation diversity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 467, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 287, + 635 + ], + "type": "text", + "content": "ing, our technique is upper-bounded by the scope and diversity of current 3D datasets. This restriction poses challenges in generating textures that parallel the depth of effects seen in 2D image synthesis. Moreover, our method's efficacy relies on the quality of UV mapping. Our approach faces difficulties in rendering high-quality results for meshes where the UV mapping produces excessive fragmented cuts, resulting in fragmented artifacts. This phenomenon is commonly observed in the car category, as shown in the Figure 15. We believe the emergence of larger and more diverse 3D datasets would be helpful for generating superior-quality textures. Further advancements in UV parameterization would also be beneficial in augmenting our method's consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 646, + 162, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 162, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 162, + 659 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "This work has been supported by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant No. 17202422), and RGC matching fund scheme (RMGS). Part" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 337, + 88, + 390, + 246 + ], + "blocks": [ + { + "bbox": [ + 326, + 71, + 421, + 83 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 421, + 83 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 421, + 83 + ], + "type": "text", + "content": "(a) w/o style guidance" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 88, + 390, + 246 + ], + "lines": [ + { + "bbox": [ + 337, + 88, + 390, + 246 + ], + "spans": [ + { + "bbox": [ + 337, + 88, + 390, + 246 + ], + "type": "image", + "image_path": "7dd3488f2ed95f997cfcd5c595ebd414538e0b8d0a8a065ad88e85c438c55789.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 248, + 375, + 258 + ], + "lines": [ + { + "bbox": [ + 341, + 248, + 375, + 258 + ], + "spans": [ + { + "bbox": [ + 341, + 248, + 375, + 258 + ], + "type": "text", + "content": "Result 1" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 405, + 88, + 453, + 245 + ], + "blocks": [ + { + "bbox": [ + 405, + 88, + 453, + 245 + ], + "lines": [ + { + "bbox": [ + 405, + 88, + 453, + 245 + ], + "spans": [ + { + "bbox": [ + 405, + 88, + 453, + 245 + ], + "type": "image", + "image_path": "1c4718cca7ba0e778997d2ff96a834b0db2dde9ea762c238ecb6c23b4d182451.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 248, + 439, + 258 + ], + "lines": [ + { + "bbox": [ + 403, + 248, + 439, + 258 + ], + "spans": [ + { + "bbox": [ + 403, + 248, + 439, + 258 + ], + "type": "text", + "content": "Result 2" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 470, + 88, + 518, + 245 + ], + "blocks": [ + { + "bbox": [ + 470, + 88, + 518, + 245 + ], + "lines": [ + { + "bbox": [ + 470, + 88, + 518, + 245 + ], + "spans": [ + { + "bbox": [ + 470, + 88, + 518, + 245 + ], + "type": "image", + "image_path": "4890e18c992d21c254ca1d2ce295014cf1c69273442006ab745df2fe643f5c7e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 248, + 506, + 258 + ], + "lines": [ + { + "bbox": [ + 470, + 248, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 470, + 248, + 506, + 258 + ], + "type": "text", + "content": "Result 3" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 335, + 281, + 385, + 437 + ], + "blocks": [ + { + "bbox": [ + 326, + 262, + 416, + 275 + ], + "lines": [ + { + "bbox": [ + 326, + 262, + 416, + 275 + ], + "spans": [ + { + "bbox": [ + 326, + 262, + 416, + 275 + ], + "type": "text", + "content": "(b) w/ style guidance" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 281, + 385, + 437 + ], + "lines": [ + { + "bbox": [ + 335, + 281, + 385, + 437 + ], + "spans": [ + { + "bbox": [ + 335, + 281, + 385, + 437 + ], + "type": "image", + "image_path": "ce8b49180e1a128a3c18a36a4ed30a43e3aa5778fcc29b6351b3bac82173037d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 441, + 375, + 451 + ], + "lines": [ + { + "bbox": [ + 341, + 441, + 375, + 451 + ], + "spans": [ + { + "bbox": [ + 341, + 441, + 375, + 451 + ], + "type": "text", + "content": "Result 1" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 457, + 545, + 502 + ], + "lines": [ + { + "bbox": [ + 305, + 457, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 457, + 545, + 502 + ], + "type": "text", + "content": "Figure 14. Efficacy of style guidance. The model without style guidance produces consistent colors across different random seeds (rows 1-2). In contrast, with style guidance, diverse outcomes emerge (rows 3-4)." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 401, + 281, + 450, + 437 + ], + "blocks": [ + { + "bbox": [ + 401, + 281, + 450, + 437 + ], + "lines": [ + { + "bbox": [ + 401, + 281, + 450, + 437 + ], + "spans": [ + { + "bbox": [ + 401, + 281, + 450, + 437 + ], + "type": "image", + "image_path": "c7e8219b8cc398c024d35bc258208c08fe92b14bb737544b5a608a3377b07c9b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 441, + 439, + 451 + ], + "lines": [ + { + "bbox": [ + 403, + 441, + 439, + 451 + ], + "spans": [ + { + "bbox": [ + 403, + 441, + 439, + 451 + ], + "type": "text", + "content": "Result 2" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 466, + 281, + 515, + 437 + ], + "blocks": [ + { + "bbox": [ + 466, + 281, + 515, + 437 + ], + "lines": [ + { + "bbox": [ + 466, + 281, + 515, + 437 + ], + "spans": [ + { + "bbox": [ + 466, + 281, + 515, + 437 + ], + "type": "image", + "image_path": "74411651e70bb3a0ad2617dfa06e510a30e900647cec2688889e10165142ce83.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 441, + 506, + 451 + ], + "lines": [ + { + "bbox": [ + 470, + 441, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 470, + 441, + 506, + 451 + ], + "type": "text", + "content": "Result 3" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 331, + 527, + 410, + 606 + ], + "blocks": [ + { + "bbox": [ + 331, + 527, + 410, + 606 + ], + "lines": [ + { + "bbox": [ + 331, + 527, + 410, + 606 + ], + "spans": [ + { + "bbox": [ + 331, + 527, + 410, + 606 + ], + "type": "image", + "image_path": "37ea41a649f5635d3496f1b5e4805874ba17d11c31569c9b66cd288641236060.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 609, + 406, + 617 + ], + "lines": [ + { + "bbox": [ + 328, + 609, + 406, + 617 + ], + "spans": [ + { + "bbox": [ + 328, + 609, + 406, + 617 + ], + "type": "text", + "content": "(a) Generated Texture Image" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 623, + 545, + 655 + ], + "lines": [ + { + "bbox": [ + 306, + 623, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 545, + 655 + ], + "type": "text", + "content": "Figure 15. Failure case. Our approach still struggles to generate seamless results when there are too many fragmented cuts of the UV map." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 426, + 543, + 527, + 601 + ], + "blocks": [ + { + "bbox": [ + 426, + 543, + 527, + 601 + ], + "lines": [ + { + "bbox": [ + 426, + 543, + 527, + 601 + ], + "spans": [ + { + "bbox": [ + 426, + 543, + 527, + 601 + ], + "type": "image", + "image_path": "c79cf71ae6423c18fb1df3c0a9a3ae93b90dd28b4932b2f24394bb28a8f1c2a8.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 609, + 503, + 616 + ], + "lines": [ + { + "bbox": [ + 452, + 609, + 503, + 616 + ], + "spans": [ + { + "bbox": [ + 452, + 609, + 503, + 616 + ], + "type": "text", + "content": "(b) Textured Mesh" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 315, + 742 + ], + "type": "text", + "content": "4214" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. arXiv preprint arXiv:2211.09869, 2022. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 179 + ], + "type": "text", + "content": "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "type": "text", + "content": "[3] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 300 + ], + "type": "text", + "content": "[4] Angel X. Chang, Thomas Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], 2015. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "type": "text", + "content": "[5] Bindita Chaudhuri, Nikolaos Sarafianos, Linda Shapiro, and Tony Tung. Semi-supervised synthesis of high-resolution editable textures for 3d humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7991-8000, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 434 + ], + "type": "text", + "content": "[6] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III 14, pages 100-116. Springer, 2019. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 435, + 287, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 287, + 457 + ], + "type": "text", + "content": "[7] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "type": "text", + "content": "[8] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 568 + ], + "type": "text", + "content": "[9] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10673-10683, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "text", + "content": "[10] Jun Gao, Wenzheng Chen, Tommy Xiang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. Advances In Neural Information Processing Systems, 33:9936-9947, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[11] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[12] Lin Gao, Tong Wu, Yu-Jie Yuan, Ming-Xian Lin, Yu-Kun Lai, and Hao Zhang. Tm-net: Deep generative networks for textured meshes. ACM Transactions on Graphics (TOG), 40(6):1-15, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 117 + ], + "type": "text", + "content": "[13] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. In Advances in neural information processing systems, pages 5767-5777, 2017. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "text", + "content": "[14] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Learning a neural 3d texture space from 2d exemplars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8356-8364, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 546, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 546, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 546, + 206 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. NIPS, 2017. 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 207, + 546, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 207, + 546, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 207, + 546, + 229 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models, 2020. 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 231, + 546, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 546, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 546, + 274 + ], + "type": "text", + "content": "[17] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23(47):1-33, 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 276, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 546, + 319 + ], + "type": "text", + "content": "[18] Jingyu Hu, Ka-Hei Hui, Zhengzhe Liu, Ruihui Li, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation, inversion, and manipulation. arXiv preprint arXiv:2302.00190, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 321, + 546, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 546, + 364 + ], + "type": "text", + "content": "[19] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation. In SIGGRAPH Asia 2022 Conference Papers, pages 1–9, 2022. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 546, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 546, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 546, + 410 + ], + "type": "text", + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 411, + 546, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 546, + 477 + ], + "type": "text", + "content": "[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 478, + 546, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 546, + 521 + ], + "type": "text", + "content": "[22] Zhengzhe Liu, Peng Dai, Ruihui Li, Xiaojuan Qi, and Chi-Wing Fu. Iss: Image as stetting stone for text-guided 3d shape generation. International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 522, + 546, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 546, + 556 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 546, + 556 + ], + "type": "text", + "content": "[23] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Pointvoxel cnn for efficient 3d deep learning. Advances in Neural Information Processing Systems, 32, 2019. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 557, + 546, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 546, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 546, + 611 + ], + "type": "text", + "content": "[24] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 546, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 546, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 546, + 667 + ], + "type": "text", + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 669, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 546, + 713 + ], + "type": "text", + "content": "[26] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 732, + 315, + 742 + ], + "type": "text", + "content": "4215" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "type": "text", + "content": "[27] Alex Nichol and Prafulla Dhariwal. Improved Denoising Diffusion Probabilistic Models, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[28] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 196 + ], + "type": "text", + "content": "[29] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "text", + "content": "[30] Dario Pavllo, Graham Spinks, Thomas Hofmann, Marie-Francine Moens, and Aurelien Lucchi. Convolutional generation of textured 3d meshes. Advances in Neural Information Processing Systems, 33:870-882, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 275 + ], + "type": "text", + "content": "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. International Conference on Learning Representations, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "text", + "content": "[32] Tiziano Portenier, Siavash Arjomand Bigdeli, and Orcun Goksel. Gramgan: Deep 3d texture synthesis from 2d exemplars. Advances in Neural Information Processing Systems, 33:6994-7004, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 377, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 420 + ], + "type": "text", + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 477 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "text", + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "text", + "content": "[37] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[38] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces. arXiv preprint arXiv:2204.02411, 2022. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[39] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. Advances in Neural Information Processing Systems, 35:24487-24501, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[40] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Confer" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 477 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "ence on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[41] Yongbin Sun, Ziwei Liu, Yue Wang, and Sanjay E Sarma. Im2avatar: Colorful 3d reconstruction from a single image. arXiv preprint arXiv:1804.06375, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "text", + "content": "[42] Shubham Tulsiani, Nilesh Kulkarni, and Abhinav Gupta. Implicit mesh reconstruction from unannotated image collections. arXiv preprint arXiv:2007.08504, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 164, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 218 + ], + "type": "text", + "content": "[43] Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2626-2634, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "text", + "content": "[44] Hao Wang, Guosheng Lin, Steven Hoi, and Chunyan Miao. Cycle-consistent inverse GAN for text-to-image synthesis. ACM MM, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 254, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 307 + ], + "type": "text", + "content": "[45] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12456-12465, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 309, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 331 + ], + "type": "text", + "content": "[46] Jonathan Young. xatlas, 2018. https://github.com/jpcy/xatlas.5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 333, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 545, + 376 + ], + "type": "text", + "content": "[47] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. arXiv preprint arXiv:2210.06978, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 377, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 431 + ], + "type": "text", + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 477 + ], + "type": "text", + "content": "[49] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5826–5835, 2021. 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "spans": [ + { + "bbox": [ + 295, + 733, + 315, + 742 + ], + "type": "text", + "content": "4216" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/5378f03d-031c-45ab-a522-6e62326e01af_content_list.json b/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/5378f03d-031c-45ab-a522-6e62326e01af_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cdd6d5342e43dbb86d13486ae5a394979e7cb24e --- /dev/null +++ b/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/5378f03d-031c-45ab-a522-6e62326e01af_content_list.json @@ -0,0 +1,1781 @@ +[ + { + "type": "text", + "text": "Texture Learning Domain Randomization for Domain Generalized Segmentation", + "text_level": 1, + "bbox": [ + 78, + 161, + 890, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sunghwan Kim", + "bbox": [ + 262, + 214, + 389, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dae-hwan Kim", + "bbox": [ + 428, + 216, + 552, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hoseong Kim*", + "bbox": [ + 593, + 216, + 710, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Agency for Defense Development (ADD)", + "bbox": [ + 319, + 235, + 648, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sssshwan，dhkim7，hoseongkim}@add.re.kr", + "bbox": [ + 312, + 256, + 655, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 302, + 312, + 317 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Neural Networks (DNNs)-based semantic segmentation models trained on a source domain often struggle to generalize to unseen target domains, i.e., a domain gap problem. Texture often contributes to the domain gap, making DNNs vulnerable to domain shift because they are prone to be texture-biased. Existing Domain Generalized Semantic Segmentation (DGSS) methods have alleviated the domain gap problem by guiding models to prioritize shape over texture. On the other hand, shape and texture are two prominent and complementary cues in semantic segmentation. This paper argues that leveraging texture is crucial for improving performance in DGSS. Specifically, we propose a novel framework, coined Texture Learning Domain Randomization (TLDR). TLDR includes two novel losses to effectively enhance texture learning in DGSS: (1) a texture regularization loss to prevent overfitting to source domain textures by using texture features from an ImageNet pretrained model and (2) a texture generalization loss that utilizes random style images to learn diverse texture representations in a self-supervised manner. Extensive experimental results demonstrate the superiority of the proposed TLDR; e.g., TLDR achieves 46.5 mIoU on GTA $\\rightarrow$ Cityscapes using ResNet-50, which improves the prior state-of-the-art method by 1.9 mIoU. The source code is available at https://github.com/ssssshwan/TLDR.", + "bbox": [ + 75, + 333, + 473, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 77, + 707, + 208, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation is an essential task in computer vision with many real-world applications, such as autonomous vehicles, augmented reality, and medical imaging. Deep Neural Networks (DNNs)-based semantic segmentation models work well when the data distributions are consistent between a source domain and target domains [1,5-7,22,34,55]. However, the performance of the models tends to degrade significantly in practical settings that in", + "bbox": [ + 75, + 731, + 468, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae0071673eeea8563b05c01168db66a957eddeeb2b50ce19075e12e7dc3ad200.jpg", + "image_caption": [ + "Unseen Target Image" + ], + "image_footnote": [], + "bbox": [ + 507, + 302, + 695, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f8a852b6b9d1742bca238f8ef771078e2b33d29d99d4d88cb3957a50d26220fd.jpg", + "image_caption": [ + "(a) Normalization & Whitening" + ], + "image_footnote": [], + "bbox": [ + 697, + 302, + 887, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/81d2ee6b752b10cd2c5a91a9f446266da65d16f984f80ca6c14149233d2d2d84.jpg", + "image_caption": [ + "(b) Domain Randomization" + ], + "image_footnote": [], + "bbox": [ + 507, + 396, + 695, + 475 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9f789ebab88827300ca44af2cc259dd87268ab796d6d1269d1c1f8606b7fa095.jpg", + "image_caption": [ + "(c) TLDR (ours)", + "Figure 1. Segmentation results for an image from an unseen domain (i.e., Cityscapes [12]), using models trained on GTA [47] with DGSS methods. (a-b) Existing methods [10, 59] have difficulty in distinguishing between road, sidewalk, and terrain which have similar shapes and contexts, as the texture is not sufficiently considered during the training process. (c) Our Texture Learning Domain Randomization (TLDR) can distinguish the classes effectively as we utilize the texture as prediction cues." + ], + "image_footnote": [], + "bbox": [ + 697, + 396, + 885, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "volve unseen out-of-distribution scenarios, also known as a domain gap problem. Many domain adaptation and generalization methods have been proposed to solve the domain gap problem [10, 24, 28, 42-44, 54, 59, 60]. Domain adaptation assumes accessibility of target domain images, differing from domain generalization. This paper addresses Domain Generalized Semantic Segmentation (DGSS), which aims to train models that can generalize to diverse unseen domains by training on a single source domain.", + "bbox": [ + 497, + 629, + 890, + 754 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing DGSS methods have attempted to address the domain gap problem by guiding models to focus on shape rather than texture. Given that texture often varies across different domains (e.g., synthetic/real and sunny/rainy/foggy), DNNs are susceptible to domain shift because they tend to be texture-biased [19, 39]. Accordingly, there are two main approaches for the DGSS methods. The first approach is Normalization and Whitening", + "bbox": [ + 497, + 755, + 892, + 866 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 129, + 49, + 200, + 84 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 253, + 48, + 789, + 87 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 95, + 853, + 223, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "677", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e7da4b5f40d6d796a760fb8112d63d57bb22a188473c1008c2dd804920c2ba03.jpg", + "image_caption": [ + "Original Source", + "Figure 2. Reconstructed source images from the feature maps of (a) normalization and whitening and (b) domain randomization. Texture features are often omitted in existing DGSS methods." + ], + "image_footnote": [], + "bbox": [ + 90, + 134, + 208, + 182 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3263f88a5d78f936c88be60963d0c6ba7807a1d643a4629ef9bf0c3b0bbc73f2.jpg", + "image_caption": [ + "(a) Norm. & Whiten." + ], + "image_footnote": [], + "bbox": [ + 213, + 134, + 332, + 182 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8d525b63b1d1b0b786906e033988aee3a74fc12f8eca834a9a08ef14bdbdc140.jpg", + "image_caption": [ + "(b) Domain Random." + ], + "image_footnote": [], + "bbox": [ + 334, + 134, + 453, + 182 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/574fefa65ff72ecd66d6e237388c56e580afdf452b7bfd87cd366d6b02abaf56.jpg", + "image_caption": [ + "(a) Shape Features", + "Figure 3. The t-SNE [52] plots for the road, sidewalk, and terrain classes from Cityscapes [12] that have similar shapes. While the shape features (Canny edge [3]) are entangled in (a), the texture features (Gram-matrix) of these classes are clearly separated in (b). The plots are based on an ImageNet pre-trained model." + ], + "image_footnote": [], + "bbox": [ + 112, + 256, + 260, + 315 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/777bc6c73a6047f506102ae422d0deacf5ab2b81255b3aabbaa0349536a68580.jpg", + "image_caption": [ + "(b) Texture Features" + ], + "image_footnote": [], + "bbox": [ + 287, + 256, + 435, + 315 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(NW), which involves normalizing and whitening the features [10, 42-44]. It is possible to remove domain-specific texture features and learn domain-invariant shape features with NW (see Figure 2a). The second approach is Domain Randomization (DR), which trains by transforming source images into randomly stylized images [24,28,45,54,59-61]. The model learns domain-invariant shape features because texture cues are mostly replaced by random styles (see Figure 2b) [38, 51].", + "bbox": [ + 75, + 407, + 468, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the existing methods are effective at making the models focus on shape features, they need to give more consideration to texture features. In addition to utilizing shape features like edges and structures, DNNs also use texture features such as patterns, color variations, and histograms as important cues for prediction [18]. Particularly in semantic segmentation, texture plays a crucial role in accurately maintaining the boundaries between objects [26, 63].", + "bbox": [ + 75, + 533, + 468, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Figure 1 demonstrates the results of predicting an unseen image from Cityscapes [12] using DGSS methods trained on GTA [47]. One can see that the models trained with NW and DR have difficulty distinguishing between the road, sidewalk, and terrain classes, which have similar shapes and contexts. In order to determine these classes accurately, it is necessary to use texture cues. This assertion is further emphasized through t-SNE [52] plots of the classes: the shape features are entangled in Figure 3a, whereas the texture features are clearly separated in Figure 3b. Meanwhile, some textures remain relatively unchanged across domains in DGSS as shown in Figure 4. Based on these observations, we suggest utilizing texture as valuable cues for prediction.", + "bbox": [ + 75, + 644, + 468, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose Texture Learning Domain Randomization (TLDR), which enables DGSS models to learn both shape and texture. Since only source domain images are available", + "bbox": [ + 75, + 824, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f130c7df74f6a7f53667abe8be917e6f04a64181f981afe366380cf53dcda358.jpg", + "image_caption": [ + "Figure 4. Visualization of texture for the road, sidewalk, and terrain classes from GTA [47], Cityscapes [12], BDD [58], and Mapillary [40] datasets. For each class, there are commonalities in texture across the datasets." + ], + "image_footnote": [], + "bbox": [ + 514, + 124, + 877, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in DGSS, texture must be learned from them. To accurately capture the texture, we leverage the source domain images without any modification, which we refer to as the original source images. The stylized source images from DR are more focused on learning shape features, and the original source images are more focused on learning texture features (Section 4.1). To further improve texture learning in DGSS, we propose a texture regularization loss and a texture generalization loss. While there are commonalities in texture across different domains, there are also clear texture differences. Thus, if the model overfits source domain textures, it will result in a performance drop. To mitigate this problem, we propose the texture regularization loss to prevent overfitting to source domain textures by using texture features from an ImageNet pre-trained model (Section 4.2). Since only source domain textures alone may not be sufficient for learning general texture representations, we propose the texture generalization loss that utilizes random style images to learn diverse texture representations in a self-supervised manner (Section 4.3).", + "bbox": [ + 497, + 404, + 890, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contribution can be summarized into three aspects. First, to the best of our knowledge, we are approaching DGSS for the first time from both the shape and texture perspectives. We argue that leveraging texture is essential in distinguishing between classes with similar shapes despite the domain gap. Second, to enhance texture learning in DGSS, we introduce two novel losses: the texture regularization loss and the texture generalization loss. Third, extensive experiments over multiple DGSS tasks show that our proposed TLDR achieves state-of-the-art performance. Our method attains $46.5\\mathrm{mIoU}$ on $\\mathrm{GTA} \\rightarrow$ Cityscapes using ResNet-50, surpassing the prior state-of-the-art method by $1.9\\mathrm{mIoU}$ .", + "bbox": [ + 497, + 686, + 890, + 865 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "678", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/37e9b48855740e1681c51867b98a94f7d1b42bad75a75d268823c0e5bc0be709.jpg", + "image_caption": [ + "(a) Style Transfer Module (STM)" + ], + "image_footnote": [], + "bbox": [ + 80, + 129, + 539, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c2332a36b4f7268c22715b33c201c0891d41b72ce84065419b1077d1f2ed42cf.jpg", + "image_caption": [ + "(b) Texture Extraction Operator (TEO)", + "Figure 5. (a) Visualization of using the Style Transfer Module (STM) to transform an original source image into stylized source images with random style images. (b) Illustration of the process of using Texture Extraction Operator (TEO) to extract only texture features, i.e., a Gram-matrix, from a feature map." + ], + "image_footnote": [], + "bbox": [ + 571, + 125, + 884, + 203 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 75, + 312, + 218, + 326 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain Generalization (DG). DG aims to learn DNNs that perform well on multiple unseen domains [36]. DG has primarily been studied in image classification. A number of studies have been proposed to address DG, including adversarial learning [27, 31, 50], data augmentation [46, 53, 56, 62], meta-learning [2, 14, 33], ensemble learning [57, 62], and self-supervised learning [4, 29].", + "bbox": [ + 75, + 336, + 468, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent studies have attempted to train domain generalized models by preserving ImageNet pre-trained feature representations as much as possible. Chen et al. defined DG as a life-long learning problem [33] and tried to utilize ImageNet pre-trained weights to prevent catastrophic forgetting [9]. Contrastive learning [8] and attentional pooling [8, 37] were introduced to enhance the capturing of semantic knowledge in ImageNet features. In this paper, we regularize texture representations of DNNs with the ImageNet features. To the best of our knowledge, this is the first attempt to extract a specific semantic concept (i.e. texture) from the ImageNet features for regularization in DG.", + "bbox": [ + 75, + 435, + 468, + 600 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain Generalized Semantic Segmentation (DGSS). DGSS is in its early stages and has yet to receive much research attention. Existing DGSS methods have tried to alleviate a domain gap problem through two main approaches: Normalization and Whitening (NW) and Domain Randomization (DR). NW trains by normalizing the mean and standard deviation of source features and whitening the covariance of source features. This process eliminates domain-specific features, allowing the model to learn domain-invariant features. Pan et al. introduced instance normalization to remove domain-specific features [42]. Pan et al. proposed switchable whitening to decorrelate features [43], and Choi et al. proposed Instance Selective Whitening (ISW) to enhance the ability to whiten domain-specific features [10]. Peng et al. tried to normalize and whiten features in a category-wise manner [44].", + "bbox": [ + 75, + 602, + 468, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DR trains by transforming source images into randomly stylized images. The model is guided to capture domain-invariant shape features since texture cues are substituted", + "bbox": [ + 75, + 824, + 468, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with random styles. Yue et al. presented a method for considering consistency among multiple stylized images [59]. Peng et al. distinguished between global and local texture in the randomization process [45]. Huang et al. proposed DR in frequency space [24]. Adversarial learning [61] and self-supervised learning [54] have been used as attempts to make style learnable rather than using random style images. It has been shown that utilizing content and style of ImageNet [28] and ImageNet pre-trained features [60] can aid in learning generalized features in DR.", + "bbox": [ + 497, + 313, + 890, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing DR methods have yet to comprehensively address DGSS from both shape and texture perspectives. Several studies [28, 59] have stated that the success of DR is attributed to its ability to learn various styles, resulting in improved performance in generalized domains. However, from a different perspective, we consider that the effectiveness of DR is due to the model becoming more focused on shape, as discussed in recent DG methods [38, 51]. Therefore, we assume that the model needs to learn texture for further performance improvement in DR. We propose a novel approach for learning texture in DGSS without overfitting to source domain textures.", + "bbox": [ + 497, + 459, + 890, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 499, + 657, + 636, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Domain Randomization with Style Transfer", + "text_level": 1, + "bbox": [ + 499, + 686, + 870, + 701 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adopt DR as a baseline DGSS method. We use a neural style transfer method at each epoch to transform each original source image into a different random style. If edge information is lost during the style transfer process, it may cause a mismatch with the semantic label, leading to a decrease in performance. We utilize photoWCT [30] known as an edge-preserving style transfer method. Random style images are sampled from ImageNet [13] validation set. Figure 5a is a visualization of using the Style Transfer Module (STM) to transform an original source image into stylized source images with random style images.", + "bbox": [ + 497, + 713, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "679", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d7777b91dc5dc3f15613d57574226280baccf338af342435c6095641e0828cb3.jpg", + "image_caption": [ + "Figure 6. An overview of our proposed Texture Learning Domain Randomization (TLDR) framework. The stylized source image $x^{sr}$ (green) is obtained by stylizing the original source image $x^s$ (blue) with the random style image $x^r$ (purple). The stylized task loss $\\mathcal{L}_{\\mathrm{styl}}$ focuses on learning shape from $x^{sr}$ , and the original task loss $\\mathcal{L}_{\\mathrm{orig}}$ focuses on learning texture from $x^s$ . The texture regularization loss $\\mathcal{L}_{\\mathrm{TR}}$ enforces the consistency between the Gram-matrices of the ImageNet model $f_I$ and the task model $f_T$ for $x^s$ . The texture generalization loss $\\mathcal{L}_{\\mathrm{TG}}$ enforces the consistency between the Gram-matrices of the task model $f_T$ for $x^{sr}$ and $x^r$ . Random Style Masking (RSM) selects only the random style features when applying the texture generalization loss." + ], + "image_footnote": [], + "bbox": [ + 78, + 123, + 890, + 350 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Texture Extraction with Gram-matrix", + "text_level": 1, + "bbox": [ + 75, + 459, + 406, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Texture is a regional descriptor that can offer measurements for both local structural (e.g., pattern) and global statistical (e.g., overall distribution of colors) properties of an image [21]. It has been shown that texture can be represented by pairwise correlations between features extracted by DNNs, also known as a Gram-matrix [16, 17, 32].", + "bbox": [ + 75, + 480, + 468, + 562 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the Gram-matrix to extract only texture features from a feature map. The Gram-matrix $G \\in \\mathbb{R}^{C \\times C}$ for the vectorized matrix $F \\in \\mathbb{R}^{C \\times HW}$ of the feature map is defined as Equation 1. $C, H$ , and $W$ denote the channel, height, and width of the feature map, respectively.", + "bbox": [ + 75, + 562, + 468, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {i, j} = F _ {i} \\cdot F _ {j}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 641, + 468, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\cdot$ represents a dot product, $F_{i}$ and $F_{j}$ are the $i^{th}$ and $j^{th}$ row vectors of $F$ , respectively. $G_{i,j}$ is the entry at the $i^{th}$ row and $j^{th}$ column of $G$ . Each entry of the Grammatrix indicates a pairwise correlation between the features corresponding to a texture feature. In this paper, the operator extracting texture features is called Texture Extraction Operator (TEO). Figure 5b illustrates the process of TEO.", + "bbox": [ + 75, + 663, + 472, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Approach", + "text_level": 1, + "bbox": [ + 75, + 774, + 186, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section describes our proposed Texture Learning Domain Randomization (TLDR) framework. TLDR learns texture features in addition to learning shape features through domain randomization. TLDR consists of four losses: a stylized task loss, an original task loss, a tex", + "bbox": [ + 75, + 796, + 470, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ture regularization loss, and a texture generalization loss. The stylized task loss focuses on learning shape, and the original task loss focuses on learning texture. The texture regularization loss and the texture generalization loss prevent overfitting to source domain textures. Figure 6 is an overview of TLDR. The frozen ImageNet pre-trained encoder is denoted as $f_{I}$ , i.e., the ImageNet model. The training encoder is denoted as $f_{T}$ , i.e., the task model. The training semantic segmentation decoder is denoted as $h_{T}$ .", + "bbox": [ + 497, + 459, + 890, + 586 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Task Losses", + "text_level": 1, + "bbox": [ + 497, + 593, + 628, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stylized task loss. We denote an original source image $x^{s}$ and its semantic label $y$ . $x^{sr}$ is a stylized source image obtained by stylizing an original source image $x^{s}$ with a random style image $x^{r}$ . The prediction result for $x^{sr}$ of the model is $p^{sr} = h_{T}(f_{T}(x^{sr}))$ . Then the stylized task loss $\\mathcal{L}_{\\mathrm{styl}}$ is given by Equation 2.", + "bbox": [ + 497, + 619, + 892, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s t y l}} = \\mathrm {C E} \\left(p ^ {s r}, y\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 631, + 712, + 892, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{CE}(\\cdot)$ represents the categorical cross-entropy loss. The stylized task loss encourages the model to focus on shape features during training since the texture cues are mostly replaced by random styles [38, 51].", + "bbox": [ + 497, + 738, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Original task loss. The model struggles to learn texture from the stylized source images as the texture cues are mostly substituted with random styles. To accurately capture the source domain textures, the model is trained on the original source images. The prediction result for $x^{s}$ of the", + "bbox": [ + 497, + 796, + 892, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "680", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model is $p^s = h_T(f_T(x^s))$ . The original task loss $\\mathcal{L}_{\\mathrm{orig}}$ is given by Equation 3.", + "bbox": [ + 75, + 125, + 468, + 153 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o r i g}} = \\mathrm {C E} \\left(p ^ {s}, y\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 211, + 160, + 468, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DNNs tend to prioritize texture cues without restrictions [19, 39]. The original task loss guides the model to concentrate on texture features during training.", + "bbox": [ + 75, + 181, + 468, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Texture Regularization Loss", + "text_level": 1, + "bbox": [ + 77, + 230, + 329, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If the model is trained on the original source images without regularization, it will overfit the source domain textures [19, 39]. We regularize texture representations of the task model using the ImageNet model, which encodes diverse feature representations [8]. However, it is important to note that ImageNet features include not only texture features, but also other semantic features such as shape features. We therefore assume that regularizing the entire feature may interfere with the task model learning texture. To address this issue, we propose to apply TEO to extract only texture features from ImageNet features and regularize the task model with the extracted texture features (see Table 5 for ablation).", + "bbox": [ + 75, + 251, + 468, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $F_{I}^{l,s}$ and $F_{T}^{l,s}$ denote the vectorized feature maps in layer $l$ of the ImageNet model $f_{I}$ and the task model $f_{T}$ for $x^{s}$ , respectively. The Gram-matrices $G_{I}^{l,s}$ and $G_{T}^{l,s}$ from $F_{I}^{l,s}$ and $F_{T}^{l,s}$ are the texture features of the original source image as seen by $f_{I}$ and $f_{T}$ , respectively. The contribution of the $l^{th}$ layer to the texture regularization loss is $\\| G_{I}^{l,s} - G_{T}^{l,s}\\|_{2}$ . The total texture regularization loss is given by Equation 4.", + "bbox": [ + 75, + 429, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T R}} = \\sum_ {l = 1} ^ {L} \\frac {u _ {l}}{C _ {l} H _ {l} W _ {l}} \\| G _ {I} ^ {l, s} - G _ {T} ^ {l, s} \\| _ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 149, + 547, + 468, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L$ is the number of feature map layers, and $u_{l}$ is a weighting factor for the contribution of the $l^{th}$ layer to $\\mathcal{L}_{\\mathrm{TR}}$ . $C_l, H_l,$ and $W_{l}$ denote the channel, height and width of the $l^{th}$ layer feature map, respectively. We reduce the value of $u_{l}$ as $l$ increases, considering that fewer texture features are encoded in feature maps as layers become deeper [25].", + "bbox": [ + 75, + 587, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Texture Generalization Loss", + "text_level": 1, + "bbox": [ + 75, + 679, + 329, + 692 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We supplement texture learning from the random style images for more diverse texture representations. Since the random style images are unlabeled, the texture representations should be learned self-supervised. Note that the random style image $x^{r}$ and the stylized source image $x^{sr}$ share some texture features. To encourage learning of diverse texture representations, we induce the texture features of $x^{sr}$ to become more similar to those of $x^{r}$ as much as possible while preserving source texture features.", + "bbox": [ + 75, + 700, + 468, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $F_T^{l,r}$ and $F_T^{l,sr}$ denote the vectorized feature maps in layer $l$ of $f_T$ for the random style image $x^r$ and the stylized source image $x^{sr}$ , respectively. $G_T^{l,r}$ and $G_T^{l,sr}$ are the", + "bbox": [ + 75, + 824, + 468, + 866 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/33b977f9373ea0340b092317e794bc0edf6abe173e5f0166c1809898363a9ebe.jpg", + "image_caption": [ + "Figure 7. Random Style Masking (RSM) masks only the entries corresponding to the random style features (purple) in the Grammatrices. The difference between $G_T^{l,s\\tau}$ and $G_T^{l,s}$ is calculated as $D^l$ . The entries with values greater than a certain threshold $\\tau$ in $D^l$ are masked as $M^l$ ." + ], + "image_footnote": [], + "bbox": [ + 502, + 124, + 889, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "corresponding Gram-matrices. Our goal is basically to set an objective that makes $G_T^{l,r}$ and $G_T^{l,sr}$ consistent. However, $G_T^{l, sr}$ includes both random style features and remaining source texture features. Applying the constraints to the entire $G_T^{l, sr}$ also imposes the objective on the source texture features in the stylized source image. To select only the random style features when enforcing the consistency between $G_T^{l,r}$ and $G_T^{l, sr}$ , we propose Random Style Masking (RSM), inspired by ISW [10]. Figure 7 is an illustration of RSM. We assume that the entries corresponding to the random style features are activated in $G_T^{l, sr}$ but deactivated in $G_T^{l,s}$ . Considering $D^l = G_T^{l, sr} - G_T^{l,s}$ , the entries corresponding to the random style features are expected to be larger than a certain threshold $\\tau$ . We denote the mask for the entries representing the random style features as $M^l$ (see Equation 5).", + "bbox": [ + 497, + 312, + 890, + 524 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i, j} ^ {l} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} D _ {i, j} ^ {l} > \\tau \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 533, + 889, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i$ and $j$ are the row and column indices in each matrix, respectively. The threshold $\\tau$ is determined empirically (see Table 6 for ablation). We only apply the objective to selected random style features by RSM. The contribution of the $l^{th}$ layer to the texture generalization loss is $\\|(G_T^{l,r} - G_T^{l,sr})\\odot M^l\\|_2$ , where $\\odot$ represents an element-wise product. The total texture generalization loss is given by Equation 6.", + "bbox": [ + 497, + 572, + 890, + 682 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T G}} = \\sum_ {l = 1} ^ {L} \\frac {v _ {l}}{C _ {l} H _ {l} W _ {l}} \\| \\left(G _ {T} ^ {l, r} - G _ {T} ^ {l, s r}\\right) \\odot M ^ {l} \\| _ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 686, + 890, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $v_{l}$ is a weighting factor for the contribution of the $l^{th}$ layer to $\\mathcal{L}_{\\mathrm{TG}}$ . We also reduce the value of $v_{l}$ as $l$ increases for the same reason as $u_{l}$ in Equation 4.", + "bbox": [ + 497, + 726, + 889, + 769 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Full Objective", + "text_level": 1, + "bbox": [ + 499, + 775, + 643, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the training progresses, the task losses decrease while the texture regularization loss and the texture generalization loss remain relatively constant. We add a Linear Decay Factor (LDF) to the texture regularization and generalization losses to balance their scale with the task losses", + "bbox": [ + 497, + 796, + 890, + 865 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "681", + "bbox": [ + 472, + 887, + 495, + 897 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/91e175ab136db1080102171b13881ca60d47673ee1b5cb9afc021c57431ca827.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodEncoderCBMS
DRPC [59]37.4232.1434.12-
RobustNet [10]36.5835.2040.3328.30
SAN-SAW [44]39.7537.3441.8630.79
SiamDoGe [54]ResNet-5042.9637.5440.6428.34
WildNet [28]44.6238.4246.0931.34
SHADE [60]44.6539.2843.34-
TLDR (ours)46.5142.5846.1836.30
DRPC [59]42.5338.7238.0529.67
GTR [45]43.7039.6039.1029.32
FSDR [24]44.8041.2043.40-
SAN-SAW [44]ResNet-10145.3341.1840.7731.84
WildNet [28]45.7941.7347.0832.51
SHADE [60]46.6643.6645.50-
TLDR (ours)47.5844.8848.8039.35
", + "bbox": [ + 85, + 123, + 460, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(see Table 6 for ablation). The LDF at iteration $t$ is set $w(t) = (1 - t / t_{\\mathrm{total}})$ , where $t_{\\mathrm{total}}$ denotes the total number of iterations. Our full objective is given by Equation 7.", + "bbox": [ + 75, + 414, + 468, + 456 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\alpha_ {\\text {o r i g}} \\mathcal {L} _ {\\text {o r i g}} + \\alpha_ {\\text {s t y l}} \\mathcal {L} _ {\\text {s t y l}} + w (t) \\mathcal {L} _ {\\text {T R}} + w (t) \\mathcal {L} _ {\\text {T G}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 87, + 473, + 468, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha_{\\mathrm{orig}}$ and $\\alpha_{\\mathrm{styl}}$ are the weights for the original task loss and the stylized task loss, respectively.", + "bbox": [ + 75, + 501, + 468, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 75, + 540, + 208, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 75, + 562, + 294, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. As synthetic datasets, GTA [47] consists of 24,966 images with a resolution of $1914 \\times 1052$ . It has 12,403, 6,382, and 6,181 images for training, validation, and test sets. SYNTHIA [48] contains 9,400 images with a resolution of $1280 \\times 760$ . It has 6,580 and 2,820 images for training and validation sets, respectively.", + "bbox": [ + 75, + 584, + 468, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As real-world datasets, Cityscapes [12] consists of 2,975 training images and 500 validation images with a resolution of $2048 \\times 1024$ . BDD [58] contains 7,000 training images and 1,000 validation images with a resolution of $1280 \\times 720$ . Mapillary [40] involves 18,000 training images and 2,000 validation images with diverse resolutions. For brevity, we denote GTA, SYNTHIA, Cityscapes, BDD, and Mapillary as G, S, C, B, and M, respectively.", + "bbox": [ + 75, + 667, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Network architecture. We conduct experiments using ResNet [23] as an encoder architecture and DeepLabV3+ [7] as a semantic segmentation decoder architecture. In all experiments, encoders are initialized with an ImageNet [13] pre-trained model.", + "bbox": [ + 75, + 780, + 468, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. We adopt an AdamW [35] optimizer. An initial", + "bbox": [ + 75, + 851, + 468, + 866 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/03756d1aacf16ae3cf6f87b84f1e7880c4006d46fe6636650c25b277d0fc9a5c.jpg", + "table_caption": [ + "Table 1. Comparison of mIoU (%; higher is better) between DGSS methods trained on GTA and evaluated on C, B, M, and S. The best and second best results are highlighted and underlined, respectively. Our method is marked in gray." + ], + "table_footnote": [], + "table_body": "
MethodEncoderCBMG
DRPC [59]35.6531.5332.7428.75
SAN-SAW [44]ResNet-5038.9235.2434.5229.16
TLDR (ours)41.8834.3536.7935.90
DRPC [59]37.5834.3434.1229.24
GTR [45]39.7035.3036.4028.71
FSDR [24]ResNet-10140.8039.6037.40-
SAN-SAW [44]40.8735.9837.2630.79
TLDR (ours)42.6035.4637.4637.77
", + "bbox": [ + 507, + 123, + 884, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of mIoU (%; higher is better) between DGSS methods trained on SYNTHIA and evaluated on C, B, M, and G. The best and second best results are highlighted and underlined, respectively. Our method is marked in gray.", + "bbox": [ + 497, + 261, + 890, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "learning rate is set to $3 \\times 10^{-5}$ for the encoder and $3 \\times 10^{-4}$ for the decoder, $40k$ training iterations, a batch size of 4. A weight decay is set to 0.01, with a linear warmup [20] over $t_{\\mathrm{warm}} = 1k$ iterations, followed by a linear decay. We use random scaling in the range [0.5, 2.0] and random cropping with a size of $768 \\times 768$ . We apply additional data augmentation techniques, including random flipping and color jittering. We set the texture regularization parameters as $u_l = 5 \\times 10^{-l-2}$ , and the texture generalization parameters as $v_l = 5 \\times 10^{-l-2}$ . The original task loss and the stylized task loss weights are set to $\\alpha_{\\mathrm{orig}} = 0.5$ and $\\alpha_{\\mathrm{styl}} = 0.5$ , respectively. We set the RSM threshold $\\tau = 0.1$ .", + "bbox": [ + 497, + 332, + 890, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison with DGSS methods", + "text_level": 1, + "bbox": [ + 497, + 508, + 791, + 524 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To measure generalization capacity in unseen domains, we train on a single source domain and evaluate multiple unseen domains. We conduct experiments on two settings, (1) $\\mathrm{G} \\rightarrow \\{\\mathrm{C}, \\mathrm{B}, \\mathrm{M}, \\mathrm{S}\\}$ and (2) $\\mathrm{S} \\rightarrow \\{\\mathrm{C}, \\mathrm{B}, \\mathrm{M}, \\mathrm{G}\\}$ . We repeat each benchmark three times, each time with a different random seed, and report the average results. We evaluate our method using ResNet-50 and ResNet-101 encoders. We use mean Intersection over Union (mIoU) [15] as the evaluation metric. The best and second best results are highlighted and underlined in tables, respectively.", + "bbox": [ + 497, + 530, + 890, + 668 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tables 1 and 2 show the generalization performance of models trained on GTA [47] and SYNTHIA [48], respectively. Our TLDR generally outperforms other DGSS methods in most benchmarks. In particular, we improve the $\\mathrm{G}\\rightarrow \\mathrm{C}$ benchmark by $+1.9\\mathrm{mIoU}$ and $+0.9\\mathrm{mIoU}$ on ResNet-50 and ResNet-101 encoders, respectively.", + "bbox": [ + 497, + 668, + 890, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Texture Awareness of Model", + "text_level": 1, + "bbox": [ + 497, + 761, + 754, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We design experiments to verify whether the performance improvement of TLDR is actually due to being aware of texture. In the experiments, we compare TLDR to plain Domain Randomization (DR), which is only trained with the stylized task loss. We train on GTA using a ResNet-101 encoder. For convenience, we refer to the models trained", + "bbox": [ + 497, + 783, + 890, + 866 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "682", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/652eb5ac72abe96eea1bc816e7c474a770983d1c4a4bfadd87a61db0e7ce2366.jpg", + "image_caption": [ + "Figure 8. Visualization of class activation maps for different classes using DR and our TLDR. The classes are road, sidewalk, and terrain. Only the classes are displayed as opaque for better visualization in the unseen images. The TLDR model tends to have activation throughout broader areas than the DR model, suggesting it can rely on more texture cues when making predictions." + ], + "image_footnote": [], + "bbox": [ + 80, + 124, + 463, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f291afe86909639704d8d49ef21f6cc14997bcd912729e375c7f7312d3a1bf74.jpg", + "image_caption": [ + "Figure 9. Comparison of the confusion matrices between DR and our TLDR for classes with similar shapes. The TLDR model tends to have less confusion and higher accuracy than the DR model." + ], + "image_footnote": [], + "bbox": [ + 85, + 368, + 465, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with DR and TLDR as the DR model and the TLDR model, respectively. The models are evaluated on Cityscapes in the experiments if no specific dataset is mentioned.", + "bbox": [ + 75, + 584, + 468, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Class activation map. To validate whether the models utilize texture cues for prediction, we generate class activation maps and analyze the contribution of texture features to the predictions. We apply Grad-CAM [49] to road, sidewalk, and terrain on the DR and TLDR models. Figure 8 displays the attention heat maps generated for each class, highlighting the regions the models focus on during prediction. Our analysis reveals that the DR model tends to have high activation in edge regions, while the TLDR model tends to have activation throughout broader areas. One can infer that these broader areas include texture cues, which the TLDR model uses as valuable prediction cues.", + "bbox": [ + 75, + 628, + 468, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Confusion matrix. We compare confusion matrices between the two models. Figure 9 shows the confusion matrices of the DR model and the TLDR model for shape-similar classes. The TLDR model has lower false positive rates in classifying sidewalk and terrain as road, with rates of", + "bbox": [ + 75, + 796, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79702bec39ca23c74a93bf52fd95e09694b242311f446047eb305b0edea6ce36.jpg", + "image_caption": [ + "Figure 10. Qualitative results of DR and our TLDR on unseen images. The unseen images are from Cityscapes, BDD, Mapillary, and SYNTHIA in row order. The TLDR model provides better prediction results for shape-similar classes than the DR model, as seen in the white boxes." + ], + "image_footnote": [], + "bbox": [ + 505, + 124, + 885, + 290 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a760cbb8db630cc9a6981ec688dbcb8a3f2692ee473a780fb6f86eddbde8b663.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LayerDRTLDR
TextureShapeTextureShape
l = 166.5%29.4%66.7%29.2%
l = 256.6%39.8%57.3%39.1%
l = 332.0%63.5%33.8%61.9%
l = 419.1%54.3%20.1%52.5%
", + "bbox": [ + 547, + 375, + 840, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Layer-wise dimensionality of texture and shape in the latent representations of DR and our TLDR, evaluated in percentages $(\\%)$ . The TLDR model stores more texture information than the DR model across all the layers.", + "bbox": [ + 497, + 473, + 890, + 524 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$19.6\\%$ and $6.9\\%$ , respectively, compared to the DR model, which has rates of $27.6\\%$ and $11.6\\%$ (see Figure 9a). Also, there is an increase in the accuracy of each class. There are also clear reductions in confusion between building, wall & fence and car, truck & bus (see Figures 9b and 9c).", + "bbox": [ + 497, + 543, + 890, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. Figure 10 shows the qualitative results of the DR and TLDR models in various domains. The TLDR model provides better prediction results for shape-similar classes than the DR model (see white boxes).", + "bbox": [ + 497, + 615, + 890, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dimensionality analysis. We investigate whether the TLDR model encodes additional texture information compared to the DR model. We use the method proposed by Islam et al. to quantify the dimensions corresponding to texture and shape in latent representations [25]. The method estimates the dimensionality of a semantic concept by computing mutual information between feature representations of input pairs that exhibit the same semantic concept. A detailed explanation of the experiment is in Appendix A. Table 3 shows the estimated dimensions (\\%) for texture and shape in each layer of the DR and TLDR models. At every layer, it is apparent that the TLDR model encodes more texture information than the DR model.", + "bbox": [ + 497, + 672, + 890, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The experimental results demonstrate that the TLDR", + "bbox": [ + 517, + 852, + 890, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "683", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e4b5fbedbcc69f9cacf9ea831e8de411a1faf330fe4a297d3ee1139cca540473.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LorigLstylLTRLTGCBMS
1✓---36.1231.3830.8937.43
2-✓--39.4839.2143.4531.96
3✓✓--43.6243.9444.0638.75
4✓✓✓-46.8546.0746.5238.64
5✓✓-✓46.3644.0147.4938.10
6✓✓✓✓47.5844.8848.8039.35
", + "bbox": [ + 84, + 122, + 458, + 225 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8523074a3f641654e5e7d6045388634fe1b6648d888e0dfcecf2a6ab961fd457.jpg", + "table_caption": [ + "Table 4. Ablation experiment on each loss in TLDR. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray." + ], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{\\text{TR}} \\)\\( \\mathcal{L}_{\\text{TG}} \\)TEOCBMS
1✓--44.9544.0944.2439.07
2✓-✓46.8546.0746.5238.64
3-✓-44.2543.5346.9038.96
4-✓✓46.3644.0147.4938.10
5✓✓-45.7044.1346.2038.57
6✓✓✓47.5844.8848.8039.35
", + "bbox": [ + 104, + 286, + 438, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model can effectively distinguish between classes with similar shapes by encoding and using texture as additional discriminative cues.", + "bbox": [ + 75, + 456, + 468, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 77, + 507, + 228, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In ablation experiments, we use ResNet-101 as the encoder, train on GTA, and evaluate on C, B, M, and S. The best results are highlighted in tables.", + "bbox": [ + 75, + 529, + 468, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss components. We investigate how each loss component contributes to overall performance. Table 4 shows the mIoU performance change with respect to the ablation of loss components. There is a large performance improvement when the stylized task loss and the original task loss are used simultaneously (cf. row 3). It seems to be the effect of using shape and texture cues as complementary. There is an additional performance improvement when the texture regularization loss and the texture generalization loss are added separately (cf. rows 4 and 5). The overall performance is the best when both losses are used (cf. row 6).", + "bbox": [ + 75, + 573, + 468, + 725 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Texture Extraction Operator (TEO). We use TEO to extract only texture features from feature maps. We conduct experiments to verify whether TEO leads to performance improvement in the texture regularization loss and the texture generalization loss. In the absence of TEO, $\\mathcal{L}_{\\mathrm{TR}}$ and $\\mathcal{L}_{\\mathrm{TG}}$ refer to the calculation of direct consistency in the feature maps without any alterations. Table 5 shows the results of TEO ablation experiments. The results show that the performance is improved in the presence of TEO (cf. rows 2, 4, and 6) in both $\\mathcal{L}_{\\mathrm{TR}}$ and $\\mathcal{L}_{\\mathrm{TG}}$ compared to the absence of", + "bbox": [ + 75, + 726, + 468, + 866 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/889925936ac137f9dca060fedb1fdf1cb9ef94ffad64c2bfc88977fdec289e89.jpg", + "table_caption": [ + "Table 5. Ablation experiments on TEO. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray." + ], + "table_footnote": [], + "table_body": "
CaseCBMS
1w/o RSM46.5744.8947.1738.87
2w/ RSM (τ=0.01)46.9244.0147.7538.61
3w/ RSM (τ=0.1)47.5844.8848.8039.35
4w/o LDF46.7744.8446.8739.72
5w/ LDF47.5844.8848.8039.35
", + "bbox": [ + 534, + 122, + 857, + 207 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/14a3867c4093fd3aba8ef0b29b540f264d6991903dbf9b1441629d18792271fa.jpg", + "table_caption": [ + "Table 6. Ablation experiments on the design choices of TLDR. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray." + ], + "table_footnote": [], + "table_body": "
Random StyleCBMS
1WikiArt [41]45.0944.5845.1639.52
2DTD [11]45.6545.1045.2939.78
3ImageNet val [13]47.5844.8848.8039.35
", + "bbox": [ + 527, + 271, + 860, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation experiment on different random style datasets. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray.", + "bbox": [ + 497, + 340, + 890, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TEO (cf. rows 1, 3, and 5).", + "bbox": [ + 499, + 400, + 680, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Random Style Masking (RSM). We conduct an experiment to validate the effectiveness of RSM. As shown in Table 6, the model with RSM (cf. row 3) performs better than the model without RSM (cf. row 1). Also, when the threshold of RSM is decreased to $\\tau = 0.01$ (cf. row 2), the performance is higher than without RSM (cf. row 1) but lower than the default setting $\\tau = 0.1$ (cf. row 3). The experiment suggests that maintaining an appropriate RSM threshold can aid in learning random styles.", + "bbox": [ + 497, + 416, + 890, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Linear Decay Factor (LDF). When analyzing the texture regularization and generalization losses, we observe that the losses do not naturally decrease as much as the task losses (see Appendix B). The texture regularization loss remains relatively constant likely due to the frozen ImageNet model. The texture generalization loss is also relatively stable likely due to the different combinations of source images and random style images used in each epoch. We design LDF to match the scale of the losses. As shown in Table 6, the model with LDF (cf. row 5) performs better than the model without LDF (cf. row 4).", + "bbox": [ + 497, + 545, + 890, + 697 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Random style dataset. We analyze how performance changes when training on various random style datasets. Table 7 shows the performance when using WikiArt [41] dataset, Describable Texture Dataset (DTD) [11] and ImageNet [13] validation set as the random style dataset. The overall performance is the best when using the ImageNet validation set as the random style dataset (cf. row 3). The performances remain consistently demonstrated even using the other two datasets (cf. rows 1 and 2). The performance improvement when using ImageNet is probably due to the superiority of the ImageNet pre-trained model to extract the texture present in ImageNet.", + "bbox": [ + 497, + 700, + 890, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "684", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 77, + 124, + 194, + 138 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Texture often contributes to the domain gap in DGSS. Existing DGSS methods have attempted to eliminate or randomize texture features. However, this paper argues that texture remains supplementary prediction cues for shape despite the domain gap. Accordingly, we proposed TLDR to learn texture features without overfitting to source domain textures in DGSS. TLDR includes novel texture regularization and generalization losses, using a Gram-matrix as a key component. We conducted a diverse set of experiments to demonstrate that TLDR effectively differentiates shape-similar classes by leveraging texture as a prediction cue. The experiments on multiple DGSS tasks show that our TLDR achieves state-of-the-art performance.", + "bbox": [ + 75, + 148, + 472, + 328 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitation. During the experiments, we found that texture differences between domains vary class-wise. Learning more source domain textures for classes with small texture differences may help with generalization, but it may be advantageous for classes with high texture differences to learn less of them. In our method, there is no class-wise prescription for different texture differences. We leave it as an interesting future work.", + "bbox": [ + 75, + 331, + 468, + 442 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 77, + 457, + 243, + 472 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We sincerely thank Chanyong Lee and Eunjin Koh for their constructive discussions and support. We also appreciate Chaehyeon Lim, Sihyun Yu, Seokhyun Moon, and Youngju Yoo for providing insightful feedback. This work was supported by the Agency for Defense Development (ADD) grant funded by the Korea government (912855101).", + "bbox": [ + 75, + 480, + 468, + 578 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 77, + 592, + 173, + 606 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vijay Badrinarayan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 39(12):2481-2495, 2017. 1", + "[2] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 3", + "[3] John Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), (6):679-698, 1986. 2", + "[4] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2229-2238, 2019. 3", + "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image" + ], + "bbox": [ + 84, + 616, + 468, + 866 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 40(4):834-848, 2017. 1", + "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1", + "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 801-818, 2018. 1, 6", + "[8] Wuyang Chen, Zhiding Yu, Shalini De Mello, Sifei Liu, Jose M Alvarez, Zhangyang Wang, and Anima Anandkumar. Contrastive syn-to-real generalization. arXiv preprint arXiv:2104.02290, 2021. 3, 5", + "[9] Wuyang Chen, Zhiding Yu, Zhangyang Wang, and Animashree Anandkumar. Automated synthetic-to-real generalization. In Proceedings of the International Conference on Machine Learning (ICML), pages 1746-1756, 2020. 3", + "[10] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11580-11590, 2021. 1, 2, 3, 5, 6", + "[11] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606-3613, 2014. 8", + "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3213-3223, 2016. 1, 2, 6", + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248-255, 2009. 3, 6, 8", + "[14] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 3", + "[15] Mark Everingham, SM Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision (IJCV), 111(1):98-136, 2015. 6", + "[16] Leon Gatys, Alexander S Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. Advances in Neural Information Processing Systems (NeurIPS), 28, 2015. 4" + ], + "bbox": [ + 502, + 125, + 890, + 864 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "685", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 4", + "[18] Yunhao Ge, Yao Xiao, Zhi Xu, Xingrui Wang, and Laurent Itti. Contributions of Shape, Texture, and Color in Visual Recognition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 369-386, 2022. 2", + "[19] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 1, 5", + "[20] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 6", + "[21] Robert M Haralick. Statistical and structural approaches to texture. Proceedings of the IEEE, 67(5):786-804, 1979. 4", + "[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2961-2969, 2017. 1", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6", + "[24] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6891-6902, 2021. 1, 2, 3, 6", + "[25] Md Amirul Islam, Matthew Kowal, Patrick Esser, Sen Jia, Bjorn Ommer, Konstantinos G Derpanis, and Neil Bruce. Shape or texture: Understanding discriminative features in cnns. arXiv preprint arXiv:2101.11604, 2021. 5, 7", + "[26] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and Statistical Texture Knowledge Distillation for Semantic Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16876-16885, 2022. 2", + "[27] Yunpei Jia, Jie Zhang, Shiguang Shan, and Xilin Chen. Single-side domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8484-8493, 2020. 3", + "[28] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. WildNet: Learning Domain Generalized Semantic Segmentation from the Wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9936-9946, 2022. 1, 2, 3, 6", + "[29] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 78, + 125, + 468, + 865 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition (CVPR), pages 224-233, 2021. 3", + "[30] Yijun Li, Ming-Yu Liu, Xueting Li, Ming-Hsuan Yang, and Jan Kautz. A closed-form solution to photorealistic image stylization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 453-468, 2018. 3", + "[31] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 624-639, 2018. 3", + "[32] Yanghao Li, Naiyan Wang, Jiaying Liu, and Xiaodi Hou. Demystifying neural style transfer. arXiv preprint arXiv:1701.01036, 2017. 4", + "[33] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 40(12):2935-2947, 2017. 3", + "[34] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3431-3440, 2015. 1", + "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[36] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain generalization via invariant feature representation. In Proceedings of the International Conference on Machine Learning (ICML), pages 10-18, 2013. 3", + "[37] Gilhyun Nam, Gyeongjae Choi, and Kyungmin Lee. GCISG: Guided Causal Invariant Learning for Improved Syn-to-Real Generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 656–672, 2022. 3", + "[38] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8690-8699, 2021. 2, 3, 4", + "[39] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. Advances in Neural Information Processing Systems (NeurIPS), 34:23296-23308, 2021. 1, 5", + "[40] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4990-4999, 2017. 2, 6", + "[41] Kiri Nichol. Painter by numbers, wikiart. Kiri Nichol, 2016. 8", + "[42] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 1, 2, 3", + "[43] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1863-1871, 2019. 1, 2, 3" + ], + "bbox": [ + 502, + 127, + 890, + 864 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "686", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2594-2605, 2022. 1, 2, 3, 6", + "[45] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2, 3, 6", + "[46] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12556-12565, 2020. 3", + "[47] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Proceedings of the European conference on computer vision (ECCV), pages 102-118, 2016. 1, 2, 6", + "[48] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3234-3243, 2016. 6", + "[49] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 618-626, 2017. 7", + "[50] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10023-10031, 2019. 3", + "[51] Nathan Somavarapu, Chih-Yao Ma, and Zsolt Kira. Frustratingly simple domain generalization via image stylization. arXiv preprint arXiv:2006.11207, 2020. 2, 3, 4", + "[52] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research (JMLR), 9(11), 2008. 2", + "[53] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 3", + "[54] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. SiamDoGe: Domain Generalizable Semantic Segmentation Using Siamese Network. In Proceedings of the European Conference on Computer Vision (ECCV), pages 603-620, 2022. 1, 2, 3, 6", + "[55] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems (NeurIPS), 34:12077-12090, 2021. 1", + "[56] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generaliza" + ], + "bbox": [ + 78, + 125, + 468, + 866 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14383-14392, 2021. 3", + "[57] Zheng Xu, Wen Li, Li Niu, and Dong Xu. Exploiting low-rank structure from latent domains for domain generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 628-643, 2014. 3", + "[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2636-2645, 2020. 2, 6", + "[59] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2100-2110, 2019. 1, 2, 3, 6", + "[60] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 535-552, 2022. 1, 2, 3, 6", + "[61] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. arXiv preprint arXiv:2207.04892, 2022. 2, 3", + "[62] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. arXiv preprint arXiv:2104.02008, 2021. 3", + "[63] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12537-12546, 2021. 2" + ], + "bbox": [ + 502, + 125, + 890, + 602 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "687", + "bbox": [ + 472, + 887, + 497, + 897 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/5378f03d-031c-45ab-a522-6e62326e01af_model.json b/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/5378f03d-031c-45ab-a522-6e62326e01af_model.json new file mode 100644 index 0000000000000000000000000000000000000000..320386c4998f4a011e856a194296632e2e040c99 --- /dev/null +++ b/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/5378f03d-031c-45ab-a522-6e62326e01af_model.json @@ -0,0 +1,2598 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.13, + 0.051, + 0.201, + 0.085 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.254, + 0.049, + 0.79, + 0.088 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.162, + 0.892, + 0.182 + ], + "angle": 0, + "content": "Texture Learning Domain Randomization for Domain Generalized Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.216, + 0.391, + 0.232 + ], + "angle": 0, + "content": "Sunghwan Kim" + }, + { + "type": "text", + "bbox": [ + 0.43, + 0.217, + 0.553, + 0.232 + ], + "angle": 0, + "content": "Dae-hwan Kim" + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.217, + 0.712, + 0.233 + ], + "angle": 0, + "content": "Hoseong Kim*" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.236, + 0.65, + 0.252 + ], + "angle": 0, + "content": "Agency for Defense Development (ADD)" + }, + { + "type": "text", + "bbox": [ + 0.313, + 0.257, + 0.657, + 0.271 + ], + "angle": 0, + "content": "{sssshwan，dhkim7，hoseongkim}@add.re.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.304, + 0.313, + 0.318 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.474, + 0.681 + ], + "angle": 0, + "content": "Deep Neural Networks (DNNs)-based semantic segmentation models trained on a source domain often struggle to generalize to unseen target domains, i.e., a domain gap problem. Texture often contributes to the domain gap, making DNNs vulnerable to domain shift because they are prone to be texture-biased. Existing Domain Generalized Semantic Segmentation (DGSS) methods have alleviated the domain gap problem by guiding models to prioritize shape over texture. On the other hand, shape and texture are two prominent and complementary cues in semantic segmentation. This paper argues that leveraging texture is crucial for improving performance in DGSS. Specifically, we propose a novel framework, coined Texture Learning Domain Randomization (TLDR). TLDR includes two novel losses to effectively enhance texture learning in DGSS: (1) a texture regularization loss to prevent overfitting to source domain textures by using texture features from an ImageNet pretrained model and (2) a texture generalization loss that utilizes random style images to learn diverse texture representations in a self-supervised manner. Extensive experimental results demonstrate the superiority of the proposed TLDR; e.g., TLDR achieves 46.5 mIoU on GTA \\(\\rightarrow\\) Cityscapes using ResNet-50, which improves the prior state-of-the-art method by 1.9 mIoU. The source code is available at https://github.com/ssssshwan/TLDR." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.708, + 0.21, + 0.722 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.844 + ], + "angle": 0, + "content": "Semantic segmentation is an essential task in computer vision with many real-world applications, such as autonomous vehicles, augmented reality, and medical imaging. Deep Neural Networks (DNNs)-based semantic segmentation models work well when the data distributions are consistent between a source domain and target domains [1,5-7,22,34,55]. However, the performance of the models tends to degrade significantly in practical settings that in" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.303, + 0.697, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.384, + 0.662, + 0.395 + ], + "angle": 0, + "content": "Unseen Target Image" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.303, + 0.888, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.384, + 0.877, + 0.395 + ], + "angle": 0, + "content": "(a) Normalization & Whitening" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.397, + 0.697, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.477, + 0.676, + 0.488 + ], + "angle": 0, + "content": "(b) Domain Randomization" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.397, + 0.887, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.477, + 0.837, + 0.489 + ], + "angle": 0, + "content": "(c) TLDR (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.603 + ], + "angle": 0, + "content": "Figure 1. Segmentation results for an image from an unseen domain (i.e., Cityscapes [12]), using models trained on GTA [47] with DGSS methods. (a-b) Existing methods [10, 59] have difficulty in distinguishing between road, sidewalk, and terrain which have similar shapes and contexts, as the texture is not sufficiently considered during the training process. (c) Our Texture Learning Domain Randomization (TLDR) can distinguish the classes effectively as we utilize the texture as prediction cues." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.755 + ], + "angle": 0, + "content": "volve unseen out-of-distribution scenarios, also known as a domain gap problem. Many domain adaptation and generalization methods have been proposed to solve the domain gap problem [10, 24, 28, 42-44, 54, 59, 60]. Domain adaptation assumes accessibility of target domain images, differing from domain generalization. This paper addresses Domain Generalized Semantic Segmentation (DGSS), which aims to train models that can generalize to diverse unseen domains by training on a single source domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.893, + 0.868 + ], + "angle": 0, + "content": "Existing DGSS methods have attempted to address the domain gap problem by guiding models to focus on shape rather than texture. Given that texture often varies across different domains (e.g., synthetic/real and sunny/rainy/foggy), DNNs are susceptible to domain shift because they tend to be texture-biased [19, 39]. Accordingly, there are two main approaches for the DGSS methods. The first approach is Normalization and Whitening" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.854, + 0.224, + 0.867 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "677" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.125, + 0.191, + 0.135 + ], + "angle": 0, + "content": "Original Source" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.135, + 0.21, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.125, + 0.321, + 0.135 + ], + "angle": 0, + "content": "(a) Norm. & Whiten." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.135, + 0.333, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.125, + 0.445, + 0.135 + ], + "angle": 0, + "content": "(b) Domain Random." + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.135, + 0.455, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.194, + 0.471, + 0.233 + ], + "angle": 0, + "content": "Figure 2. Reconstructed source images from the feature maps of (a) normalization and whitening and (b) domain randomization. Texture features are often omitted in existing DGSS methods." + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.247, + 0.232, + 0.257 + ], + "angle": 0, + "content": "(a) Shape Features" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.257, + 0.261, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.247, + 0.41, + 0.257 + ], + "angle": 0, + "content": "(b) Texture Features" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.257, + 0.436, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.327, + 0.47, + 0.392 + ], + "angle": 0, + "content": "Figure 3. The t-SNE [52] plots for the road, sidewalk, and terrain classes from Cityscapes [12] that have similar shapes. While the shape features (Canny edge [3]) are entangled in (a), the texture features (Gram-matrix) of these classes are clearly separated in (b). The plots are based on an ImageNet pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.469, + 0.532 + ], + "angle": 0, + "content": "(NW), which involves normalizing and whitening the features [10, 42-44]. It is possible to remove domain-specific texture features and learn domain-invariant shape features with NW (see Figure 2a). The second approach is Domain Randomization (DR), which trains by transforming source images into randomly stylized images [24,28,45,54,59-61]. The model learns domain-invariant shape features because texture cues are mostly replaced by random styles (see Figure 2b) [38, 51]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.469, + 0.644 + ], + "angle": 0, + "content": "While the existing methods are effective at making the models focus on shape features, they need to give more consideration to texture features. In addition to utilizing shape features like edges and structures, DNNs also use texture features such as patterns, color variations, and histograms as important cues for prediction [18]. Particularly in semantic segmentation, texture plays a crucial role in accurately maintaining the boundaries between objects [26, 63]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Figure 1 demonstrates the results of predicting an unseen image from Cityscapes [12] using DGSS methods trained on GTA [47]. One can see that the models trained with NW and DR have difficulty distinguishing between the road, sidewalk, and terrain classes, which have similar shapes and contexts. In order to determine these classes accurately, it is necessary to use texture cues. This assertion is further emphasized through t-SNE [52] plots of the classes: the shape features are entangled in Figure 3a, whereas the texture features are clearly separated in Figure 3b. Meanwhile, some textures remain relatively unchanged across domains in DGSS as shown in Figure 4. Based on these observations, we suggest utilizing texture as valuable cues for prediction." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.867 + ], + "angle": 0, + "content": "We propose Texture Learning Domain Randomization (TLDR), which enables DGSS models to learn both shape and texture. Since only source domain images are available" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.125, + 0.878, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.332, + 0.892, + 0.382 + ], + "angle": 0, + "content": "Figure 4. Visualization of texture for the road, sidewalk, and terrain classes from GTA [47], Cityscapes [12], BDD [58], and Mapillary [40] datasets. For each class, there are commonalities in texture across the datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.682 + ], + "angle": 0, + "content": "in DGSS, texture must be learned from them. To accurately capture the texture, we leverage the source domain images without any modification, which we refer to as the original source images. The stylized source images from DR are more focused on learning shape features, and the original source images are more focused on learning texture features (Section 4.1). To further improve texture learning in DGSS, we propose a texture regularization loss and a texture generalization loss. While there are commonalities in texture across different domains, there are also clear texture differences. Thus, if the model overfits source domain textures, it will result in a performance drop. To mitigate this problem, we propose the texture regularization loss to prevent overfitting to source domain textures by using texture features from an ImageNet pre-trained model (Section 4.2). Since only source domain textures alone may not be sufficient for learning general texture representations, we propose the texture generalization loss that utilizes random style images to learn diverse texture representations in a self-supervised manner (Section 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.866 + ], + "angle": 0, + "content": "Our contribution can be summarized into three aspects. First, to the best of our knowledge, we are approaching DGSS for the first time from both the shape and texture perspectives. We argue that leveraging texture is essential in distinguishing between classes with similar shapes despite the domain gap. Second, to enhance texture learning in DGSS, we introduce two novel losses: the texture regularization loss and the texture generalization loss. Third, extensive experiments over multiple DGSS tasks show that our proposed TLDR achieves state-of-the-art performance. Our method attains \\(46.5\\mathrm{mIoU}\\) on \\(\\mathrm{GTA} \\rightarrow\\) Cityscapes using ResNet-50, surpassing the prior state-of-the-art method by \\(1.9\\mathrm{mIoU}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "678" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.13, + 0.541, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.226, + 0.408, + 0.238 + ], + "angle": 0, + "content": "(a) Style Transfer Module (STM)" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.127, + 0.885, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.224, + 0.836, + 0.237 + ], + "angle": 0, + "content": "(b) Texture Extraction Operator (TEO)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.252, + 0.893, + 0.292 + ], + "angle": 0, + "content": "Figure 5. (a) Visualization of using the Style Transfer Module (STM) to transform an original source image into stylized source images with random style images. (b) Illustration of the process of using Texture Extraction Operator (TEO) to extract only texture features, i.e., a Gram-matrix, from a feature map." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.313, + 0.219, + 0.327 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.469, + 0.435 + ], + "angle": 0, + "content": "Domain Generalization (DG). DG aims to learn DNNs that perform well on multiple unseen domains [36]. DG has primarily been studied in image classification. A number of studies have been proposed to address DG, including adversarial learning [27, 31, 50], data augmentation [46, 53, 56, 62], meta-learning [2, 14, 33], ensemble learning [57, 62], and self-supervised learning [4, 29]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.469, + 0.601 + ], + "angle": 0, + "content": "Recent studies have attempted to train domain generalized models by preserving ImageNet pre-trained feature representations as much as possible. Chen et al. defined DG as a life-long learning problem [33] and tried to utilize ImageNet pre-trained weights to prevent catastrophic forgetting [9]. Contrastive learning [8] and attentional pooling [8, 37] were introduced to enhance the capturing of semantic knowledge in ImageNet features. In this paper, we regularize texture representations of DNNs with the ImageNet features. To the best of our knowledge, this is the first attempt to extract a specific semantic concept (i.e. texture) from the ImageNet features for regularization in DG." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.603, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Domain Generalized Semantic Segmentation (DGSS). DGSS is in its early stages and has yet to receive much research attention. Existing DGSS methods have tried to alleviate a domain gap problem through two main approaches: Normalization and Whitening (NW) and Domain Randomization (DR). NW trains by normalizing the mean and standard deviation of source features and whitening the covariance of source features. This process eliminates domain-specific features, allowing the model to learn domain-invariant features. Pan et al. introduced instance normalization to remove domain-specific features [42]. Pan et al. proposed switchable whitening to decorrelate features [43], and Choi et al. proposed Instance Selective Whitening (ISW) to enhance the ability to whiten domain-specific features [10]. Peng et al. tried to normalize and whiten features in a category-wise manner [44]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.867 + ], + "angle": 0, + "content": "DR trains by transforming source images into randomly stylized images. The model is guided to capture domain-invariant shape features since texture cues are substituted" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.892, + 0.453 + ], + "angle": 0, + "content": "with random styles. Yue et al. presented a method for considering consistency among multiple stylized images [59]. Peng et al. distinguished between global and local texture in the randomization process [45]. Huang et al. proposed DR in frequency space [24]. Adversarial learning [61] and self-supervised learning [54] have been used as attempts to make style learnable rather than using random style images. It has been shown that utilizing content and style of ImageNet [28] and ImageNet pre-trained features [60] can aid in learning generalized features in DR." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.626 + ], + "angle": 0, + "content": "Existing DR methods have yet to comprehensively address DGSS from both shape and texture perspectives. Several studies [28, 59] have stated that the success of DR is attributed to its ability to learn various styles, resulting in improved performance in generalized domains. However, from a different perspective, we consider that the effectiveness of DR is due to the model becoming more focused on shape, as discussed in recent DG methods [38, 51]. Therefore, we assume that the model needs to learn texture for further performance improvement in DR. We propose a novel approach for learning texture in DGSS without overfitting to source domain textures." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.658, + 0.637, + 0.672 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.687, + 0.871, + 0.702 + ], + "angle": 0, + "content": "3.1. Domain Randomization with Style Transfer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.867 + ], + "angle": 0, + "content": "We adopt DR as a baseline DGSS method. We use a neural style transfer method at each epoch to transform each original source image into a different random style. If edge information is lost during the style transfer process, it may cause a mismatch with the semantic label, leading to a decrease in performance. We utilize photoWCT [30] known as an edge-preserving style transfer method. Random style images are sampled from ImageNet [13] validation set. Figure 5a is a visualization of using the Style Transfer Module (STM) to transform an original source image into stylized source images with random style images." + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "679" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.124, + 0.891, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.359, + 0.896, + 0.437 + ], + "angle": 0, + "content": "Figure 6. An overview of our proposed Texture Learning Domain Randomization (TLDR) framework. The stylized source image \\( x^{sr} \\) (green) is obtained by stylizing the original source image \\( x^s \\) (blue) with the random style image \\( x^r \\) (purple). The stylized task loss \\( \\mathcal{L}_{\\mathrm{styl}} \\) focuses on learning shape from \\( x^{sr} \\), and the original task loss \\( \\mathcal{L}_{\\mathrm{orig}} \\) focuses on learning texture from \\( x^s \\). The texture regularization loss \\( \\mathcal{L}_{\\mathrm{TR}} \\) enforces the consistency between the Gram-matrices of the ImageNet model \\( f_I \\) and the task model \\( f_T \\) for \\( x^s \\). The texture generalization loss \\( \\mathcal{L}_{\\mathrm{TG}} \\) enforces the consistency between the Gram-matrices of the task model \\( f_T \\) for \\( x^{sr} \\) and \\( x^r \\). Random Style Masking (RSM) selects only the random style features when applying the texture generalization loss." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.46, + 0.407, + 0.473 + ], + "angle": 0, + "content": "3.2. Texture Extraction with Gram-matrix" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.47, + 0.564 + ], + "angle": 0, + "content": "Texture is a regional descriptor that can offer measurements for both local structural (e.g., pattern) and global statistical (e.g., overall distribution of colors) properties of an image [21]. It has been shown that texture can be represented by pairwise correlations between features extracted by DNNs, also known as a Gram-matrix [16, 17, 32]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.634 + ], + "angle": 0, + "content": "We use the Gram-matrix to extract only texture features from a feature map. The Gram-matrix \\( G \\in \\mathbb{R}^{C \\times C} \\) for the vectorized matrix \\( F \\in \\mathbb{R}^{C \\times HW} \\) of the feature map is defined as Equation 1. \\( C, H \\), and \\( W \\) denote the channel, height, and width of the feature map, respectively." + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.642, + 0.47, + 0.657 + ], + "angle": 0, + "content": "\\[\nG _ {i, j} = F _ {i} \\cdot F _ {j}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.473, + 0.764 + ], + "angle": 0, + "content": "where \\(\\cdot\\) represents a dot product, \\(F_{i}\\) and \\(F_{j}\\) are the \\(i^{th}\\) and \\(j^{th}\\) row vectors of \\(F\\), respectively. \\(G_{i,j}\\) is the entry at the \\(i^{th}\\) row and \\(j^{th}\\) column of \\(G\\). Each entry of the Grammatrix indicates a pairwise correlation between the features corresponding to a texture feature. In this paper, the operator extracting texture features is called Texture Extraction Operator (TEO). Figure 5b illustrates the process of TEO." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.775, + 0.188, + 0.79 + ], + "angle": 0, + "content": "4. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.797, + 0.471, + 0.868 + ], + "angle": 0, + "content": "This section describes our proposed Texture Learning Domain Randomization (TLDR) framework. TLDR learns texture features in addition to learning shape features through domain randomization. TLDR consists of four losses: a stylized task loss, an original task loss, a tex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.587 + ], + "angle": 0, + "content": "ture regularization loss, and a texture generalization loss. The stylized task loss focuses on learning shape, and the original task loss focuses on learning texture. The texture regularization loss and the texture generalization loss prevent overfitting to source domain textures. Figure 6 is an overview of TLDR. The frozen ImageNet pre-trained encoder is denoted as \\( f_{I} \\), i.e., the ImageNet model. The training encoder is denoted as \\( f_{T} \\), i.e., the task model. The training semantic segmentation decoder is denoted as \\( h_{T} \\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.595, + 0.629, + 0.609 + ], + "angle": 0, + "content": "4.1. Task Losses" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.893, + 0.705 + ], + "angle": 0, + "content": "Stylized task loss. We denote an original source image \\( x^{s} \\) and its semantic label \\( y \\). \\( x^{sr} \\) is a stylized source image obtained by stylizing an original source image \\( x^{s} \\) with a random style image \\( x^{r} \\). The prediction result for \\( x^{sr} \\) of the model is \\( p^{sr} = h_{T}(f_{T}(x^{sr})) \\). Then the stylized task loss \\( \\mathcal{L}_{\\mathrm{styl}} \\) is given by Equation 2." + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.713, + 0.893, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s t y l}} = \\mathrm {C E} \\left(p ^ {s r}, y\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.739, + 0.892, + 0.795 + ], + "angle": 0, + "content": "where \\(\\mathrm{CE}(\\cdot)\\) represents the categorical cross-entropy loss. The stylized task loss encourages the model to focus on shape features during training since the texture cues are mostly replaced by random styles [38, 51]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.893, + 0.868 + ], + "angle": 0, + "content": "Original task loss. The model struggles to learn texture from the stylized source images as the texture cues are mostly substituted with random styles. To accurately capture the source domain textures, the model is trained on the original source images. The prediction result for \\( x^{s} \\) of the" + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "680" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.47, + 0.154 + ], + "angle": 0, + "content": "model is \\( p^s = h_T(f_T(x^s)) \\). The original task loss \\( \\mathcal{L}_{\\mathrm{orig}} \\) is given by Equation 3." + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.161, + 0.469, + 0.176 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o r i g}} = \\mathrm {C E} \\left(p ^ {s}, y\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.182, + 0.47, + 0.224 + ], + "angle": 0, + "content": "DNNs tend to prioritize texture cues without restrictions [19, 39]. The original task loss guides the model to concentrate on texture features during training." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.231, + 0.331, + 0.244 + ], + "angle": 0, + "content": "4.2. Texture Regularization Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.252, + 0.469, + 0.431 + ], + "angle": 0, + "content": "If the model is trained on the original source images without regularization, it will overfit the source domain textures [19, 39]. We regularize texture representations of the task model using the ImageNet model, which encodes diverse feature representations [8]. However, it is important to note that ImageNet features include not only texture features, but also other semantic features such as shape features. We therefore assume that regularizing the entire feature may interfere with the task model learning texture. To address this issue, we propose to apply TEO to extract only texture features from ImageNet features and regularize the task model with the extracted texture features (see Table 5 for ablation)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.469, + 0.543 + ], + "angle": 0, + "content": "Let \\( F_{I}^{l,s} \\) and \\( F_{T}^{l,s} \\) denote the vectorized feature maps in layer \\( l \\) of the ImageNet model \\( f_{I} \\) and the task model \\( f_{T} \\) for \\( x^{s} \\), respectively. The Gram-matrices \\( G_{I}^{l,s} \\) and \\( G_{T}^{l,s} \\) from \\( F_{I}^{l,s} \\) and \\( F_{T}^{l,s} \\) are the texture features of the original source image as seen by \\( f_{I} \\) and \\( f_{T} \\), respectively. The contribution of the \\( l^{th} \\) layer to the texture regularization loss is \\( \\| G_{I}^{l,s} - G_{T}^{l,s}\\|_{2} \\). The total texture regularization loss is given by Equation 4." + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.548, + 0.469, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T R}} = \\sum_ {l = 1} ^ {L} \\frac {u _ {l}}{C _ {l} H _ {l} W _ {l}} \\| G _ {I} ^ {l, s} - G _ {T} ^ {l, s} \\| _ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.589, + 0.469, + 0.673 + ], + "angle": 0, + "content": "where \\(L\\) is the number of feature map layers, and \\(u_{l}\\) is a weighting factor for the contribution of the \\(l^{th}\\) layer to \\(\\mathcal{L}_{\\mathrm{TR}}\\). \\(C_l, H_l,\\) and \\(W_{l}\\) denote the channel, height and width of the \\(l^{th}\\) layer feature map, respectively. We reduce the value of \\(u_{l}\\) as \\(l\\) increases, considering that fewer texture features are encoded in feature maps as layers become deeper [25]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.331, + 0.693 + ], + "angle": 0, + "content": "4.3. Texture Generalization Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.469, + 0.824 + ], + "angle": 0, + "content": "We supplement texture learning from the random style images for more diverse texture representations. Since the random style images are unlabeled, the texture representations should be learned self-supervised. Note that the random style image \\( x^{r} \\) and the stylized source image \\( x^{sr} \\) share some texture features. To encourage learning of diverse texture representations, we induce the texture features of \\( x^{sr} \\) to become more similar to those of \\( x^{r} \\) as much as possible while preserving source texture features." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.868 + ], + "angle": 0, + "content": "Let \\( F_T^{l,r} \\) and \\( F_T^{l,sr} \\) denote the vectorized feature maps in layer \\( l \\) of \\( f_T \\) for the random style image \\( x^r \\) and the stylized source image \\( x^{sr} \\), respectively. \\( G_T^{l,r} \\) and \\( G_T^{l,sr} \\) are the" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.125, + 0.89, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.238, + 0.892, + 0.302 + ], + "angle": 0, + "content": "Figure 7. Random Style Masking (RSM) masks only the entries corresponding to the random style features (purple) in the Grammatrices. The difference between \\( G_T^{l,s\\tau} \\) and \\( G_T^{l,s} \\) is calculated as \\( D^l \\). The entries with values greater than a certain threshold \\( \\tau \\) in \\( D^l \\) are masked as \\( M^l \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.313, + 0.892, + 0.525 + ], + "angle": 0, + "content": "corresponding Gram-matrices. Our goal is basically to set an objective that makes \\( G_T^{l,r} \\) and \\( G_T^{l,sr} \\) consistent. However, \\( G_T^{l, sr} \\) includes both random style features and remaining source texture features. Applying the constraints to the entire \\( G_T^{l, sr} \\) also imposes the objective on the source texture features in the stylized source image. To select only the random style features when enforcing the consistency between \\( G_T^{l,r} \\) and \\( G_T^{l, sr} \\), we propose Random Style Masking (RSM), inspired by ISW [10]. Figure 7 is an illustration of RSM. We assume that the entries corresponding to the random style features are activated in \\( G_T^{l, sr} \\) but deactivated in \\( G_T^{l,s} \\). Considering \\( D^l = G_T^{l, sr} - G_T^{l,s} \\), the entries corresponding to the random style features are expected to be larger than a certain threshold \\( \\tau \\). We denote the mask for the entries representing the random style features as \\( M^l \\) (see Equation 5)." + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.534, + 0.89, + 0.571 + ], + "angle": 0, + "content": "\\[\nM _ {i, j} ^ {l} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} D _ {i, j} ^ {l} > \\tau \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.573, + 0.891, + 0.683 + ], + "angle": 0, + "content": "where \\(i\\) and \\(j\\) are the row and column indices in each matrix, respectively. The threshold \\(\\tau\\) is determined empirically (see Table 6 for ablation). We only apply the objective to selected random style features by RSM. The contribution of the \\(l^{th}\\) layer to the texture generalization loss is \\(\\|(G_T^{l,r} - G_T^{l,sr})\\odot M^l\\|_2\\), where \\(\\odot\\) represents an element-wise product. The total texture generalization loss is given by Equation 6." + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.687, + 0.891, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T G}} = \\sum_ {l = 1} ^ {L} \\frac {v _ {l}}{C _ {l} H _ {l} W _ {l}} \\| \\left(G _ {T} ^ {l, r} - G _ {T} ^ {l, s r}\\right) \\odot M ^ {l} \\| _ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.727, + 0.89, + 0.77 + ], + "angle": 0, + "content": "where \\( v_{l} \\) is a weighting factor for the contribution of the \\( l^{th} \\) layer to \\( \\mathcal{L}_{\\mathrm{TG}} \\). We also reduce the value of \\( v_{l} \\) as \\( l \\) increases for the same reason as \\( u_{l} \\) in Equation 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.776, + 0.645, + 0.791 + ], + "angle": 0, + "content": "4.4. Full Objective" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.798, + 0.891, + 0.866 + ], + "angle": 0, + "content": "As the training progresses, the task losses decrease while the texture regularization loss and the texture generalization loss remain relatively constant. We add a Linear Decay Factor (LDF) to the texture regularization and generalization losses to balance their scale with the task losses" + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.496, + 0.898 + ], + "angle": 0, + "content": "681" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.124, + 0.462, + 0.334 + ], + "angle": 0, + "content": "
MethodEncoderCBMS
DRPC [59]37.4232.1434.12-
RobustNet [10]36.5835.2040.3328.30
SAN-SAW [44]39.7537.3441.8630.79
SiamDoGe [54]ResNet-5042.9637.5440.6428.34
WildNet [28]44.6238.4246.0931.34
SHADE [60]44.6539.2843.34-
TLDR (ours)46.5142.5846.1836.30
DRPC [59]42.5338.7238.0529.67
GTR [45]43.7039.6039.1029.32
FSDR [24]44.8041.2043.40-
SAN-SAW [44]ResNet-10145.3341.1840.7731.84
WildNet [28]45.7941.7347.0832.51
SHADE [60]46.6643.6645.50-
TLDR (ours)47.5844.8848.8039.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.345, + 0.47, + 0.397 + ], + "angle": 0, + "content": "Table 1. Comparison of mIoU (%; higher is better) between DGSS methods trained on GTA and evaluated on C, B, M, and S. The best and second best results are highlighted and underlined, respectively. Our method is marked in gray." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.415, + 0.47, + 0.457 + ], + "angle": 0, + "content": "(see Table 6 for ablation). The LDF at iteration \\( t \\) is set \\( w(t) = (1 - t / t_{\\mathrm{total}}) \\), where \\( t_{\\mathrm{total}} \\) denotes the total number of iterations. Our full objective is given by Equation 7." + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.474, + 0.469, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\alpha_ {\\text {o r i g}} \\mathcal {L} _ {\\text {o r i g}} + \\alpha_ {\\text {s t y l}} \\mathcal {L} _ {\\text {s t y l}} + w (t) \\mathcal {L} _ {\\text {T R}} + w (t) \\mathcal {L} _ {\\text {T G}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.469, + 0.53 + ], + "angle": 0, + "content": "where \\(\\alpha_{\\mathrm{orig}}\\) and \\(\\alpha_{\\mathrm{styl}}\\) are the weights for the original task loss and the stylized task loss, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.541, + 0.21, + 0.557 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.564, + 0.295, + 0.579 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.469, + 0.668 + ], + "angle": 0, + "content": "Datasets. As synthetic datasets, GTA [47] consists of 24,966 images with a resolution of \\(1914 \\times 1052\\). It has 12,403, 6,382, and 6,181 images for training, validation, and test sets. SYNTHIA [48] contains 9,400 images with a resolution of \\(1280 \\times 760\\). It has 6,580 and 2,820 images for training and validation sets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.469, + 0.78 + ], + "angle": 0, + "content": "As real-world datasets, Cityscapes [12] consists of 2,975 training images and 500 validation images with a resolution of \\(2048 \\times 1024\\). BDD [58] contains 7,000 training images and 1,000 validation images with a resolution of \\(1280 \\times 720\\). Mapillary [40] involves 18,000 training images and 2,000 validation images with diverse resolutions. For brevity, we denote GTA, SYNTHIA, Cityscapes, BDD, and Mapillary as G, S, C, B, and M, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.85 + ], + "angle": 0, + "content": "Network architecture. We conduct experiments using ResNet [23] as an encoder architecture and DeepLabV3+ [7] as a semantic segmentation decoder architecture. In all experiments, encoders are initialized with an ImageNet [13] pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.852, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Training. We adopt an AdamW [35] optimizer. An initial" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.124, + 0.885, + 0.252 + ], + "angle": 0, + "content": "
MethodEncoderCBMG
DRPC [59]35.6531.5332.7428.75
SAN-SAW [44]ResNet-5038.9235.2434.5229.16
TLDR (ours)41.8834.3536.7935.90
DRPC [59]37.5834.3434.1229.24
GTR [45]39.7035.3036.4028.71
FSDR [24]ResNet-10140.8039.6037.40-
SAN-SAW [44]40.8735.9837.2630.79
TLDR (ours)42.6035.4637.4637.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.262, + 0.892, + 0.314 + ], + "angle": 0, + "content": "Table 2. Comparison of mIoU (%; higher is better) between DGSS methods trained on SYNTHIA and evaluated on C, B, M, and G. The best and second best results are highlighted and underlined, respectively. Our method is marked in gray." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.5 + ], + "angle": 0, + "content": "learning rate is set to \\(3 \\times 10^{-5}\\) for the encoder and \\(3 \\times 10^{-4}\\) for the decoder, \\(40k\\) training iterations, a batch size of 4. A weight decay is set to 0.01, with a linear warmup [20] over \\(t_{\\mathrm{warm}} = 1k\\) iterations, followed by a linear decay. We use random scaling in the range [0.5, 2.0] and random cropping with a size of \\(768 \\times 768\\). We apply additional data augmentation techniques, including random flipping and color jittering. We set the texture regularization parameters as \\(u_l = 5 \\times 10^{-l-2}\\), and the texture generalization parameters as \\(v_l = 5 \\times 10^{-l-2}\\). The original task loss and the stylized task loss weights are set to \\(\\alpha_{\\mathrm{orig}} = 0.5\\) and \\(\\alpha_{\\mathrm{styl}} = 0.5\\), respectively. We set the RSM threshold \\(\\tau = 0.1\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.509, + 0.792, + 0.525 + ], + "angle": 0, + "content": "5.2. Comparison with DGSS methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.892, + 0.669 + ], + "angle": 0, + "content": "To measure generalization capacity in unseen domains, we train on a single source domain and evaluate multiple unseen domains. We conduct experiments on two settings, (1) \\( \\mathrm{G} \\rightarrow \\{\\mathrm{C}, \\mathrm{B}, \\mathrm{M}, \\mathrm{S}\\} \\) and (2) \\( \\mathrm{S} \\rightarrow \\{\\mathrm{C}, \\mathrm{B}, \\mathrm{M}, \\mathrm{G}\\} \\). We repeat each benchmark three times, each time with a different random seed, and report the average results. We evaluate our method using ResNet-50 and ResNet-101 encoders. We use mean Intersection over Union (mIoU) [15] as the evaluation metric. The best and second best results are highlighted and underlined in tables, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.753 + ], + "angle": 0, + "content": "Tables 1 and 2 show the generalization performance of models trained on GTA [47] and SYNTHIA [48], respectively. Our TLDR generally outperforms other DGSS methods in most benchmarks. In particular, we improve the \\(\\mathrm{G}\\rightarrow \\mathrm{C}\\) benchmark by \\(+1.9\\mathrm{mIoU}\\) and \\(+0.9\\mathrm{mIoU}\\) on ResNet-50 and ResNet-101 encoders, respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.762, + 0.755, + 0.776 + ], + "angle": 0, + "content": "5.3. Texture Awareness of Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.892, + 0.867 + ], + "angle": 0, + "content": "We design experiments to verify whether the performance improvement of TLDR is actually due to being aware of texture. In the experiments, we compare TLDR to plain Domain Randomization (DR), which is only trained with the stylized task loss. We train on GTA using a ResNet-101 encoder. For convenience, we refer to the models trained" + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "682" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.125, + 0.465, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.28, + 0.47, + 0.357 + ], + "angle": 0, + "content": "Figure 8. Visualization of class activation maps for different classes using DR and our TLDR. The classes are road, sidewalk, and terrain. Only the classes are displayed as opaque for better visualization in the unseen images. The TLDR model tends to have activation throughout broader areas than the DR model, suggesting it can rely on more texture cues when making predictions." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.369, + 0.466, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.529, + 0.47, + 0.568 + ], + "angle": 0, + "content": "Figure 9. Comparison of the confusion matrices between DR and our TLDR for classes with similar shapes. The TLDR model tends to have less confusion and higher accuracy than the DR model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.469, + 0.627 + ], + "angle": 0, + "content": "with DR and TLDR as the DR model and the TLDR model, respectively. The models are evaluated on Cityscapes in the experiments if no specific dataset is mentioned." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.795 + ], + "angle": 0, + "content": "Class activation map. To validate whether the models utilize texture cues for prediction, we generate class activation maps and analyze the contribution of texture features to the predictions. We apply Grad-CAM [49] to road, sidewalk, and terrain on the DR and TLDR models. Figure 8 displays the attention heat maps generated for each class, highlighting the regions the models focus on during prediction. Our analysis reveals that the DR model tends to have high activation in edge regions, while the TLDR model tends to have activation throughout broader areas. One can infer that these broader areas include texture cues, which the TLDR model uses as valuable prediction cues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.798, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Confusion matrix. We compare confusion matrices between the two models. Figure 9 shows the confusion matrices of the DR model and the TLDR model for shape-similar classes. The TLDR model has lower false positive rates in classifying sidewalk and terrain as road, with rates of" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.125, + 0.886, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.3, + 0.892, + 0.363 + ], + "angle": 0, + "content": "Figure 10. Qualitative results of DR and our TLDR on unseen images. The unseen images are from Cityscapes, BDD, Mapillary, and SYNTHIA in row order. The TLDR model provides better prediction results for shape-similar classes than the DR model, as seen in the white boxes." + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.376, + 0.842, + 0.466 + ], + "angle": 0, + "content": "
LayerDRTLDR
TextureShapeTextureShape
l = 166.5%29.4%66.7%29.2%
l = 256.6%39.8%57.3%39.1%
l = 332.0%63.5%33.8%61.9%
l = 419.1%54.3%20.1%52.5%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.475, + 0.892, + 0.526 + ], + "angle": 0, + "content": "Table 3. Layer-wise dimensionality of texture and shape in the latent representations of DR and our TLDR, evaluated in percentages \\((\\%)\\). The TLDR model stores more texture information than the DR model across all the layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.892, + 0.613 + ], + "angle": 0, + "content": "\\(19.6\\%\\) and \\(6.9\\%\\), respectively, compared to the DR model, which has rates of \\(27.6\\%\\) and \\(11.6\\%\\) (see Figure 9a). Also, there is an increase in the accuracy of each class. There are also clear reductions in confusion between building, wall & fence and car, truck & bus (see Figures 9b and 9c)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.892, + 0.671 + ], + "angle": 0, + "content": "Qualitative results. Figure 10 shows the qualitative results of the DR and TLDR models in various domains. The TLDR model provides better prediction results for shape-similar classes than the DR model (see white boxes)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Dimensionality analysis. We investigate whether the TLDR model encodes additional texture information compared to the DR model. We use the method proposed by Islam et al. to quantify the dimensions corresponding to texture and shape in latent representations [25]. The method estimates the dimensionality of a semantic concept by computing mutual information between feature representations of input pairs that exhibit the same semantic concept. A detailed explanation of the experiment is in Appendix A. Table 3 shows the estimated dimensions (\\%) for texture and shape in each layer of the DR and TLDR models. At every layer, it is apparent that the TLDR model encodes more texture information than the DR model." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.853, + 0.892, + 0.867 + ], + "angle": 0, + "content": "The experimental results demonstrate that the TLDR" + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "683" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.085, + 0.123, + 0.46, + 0.226 + ], + "angle": 0, + "content": "
LorigLstylLTRLTGCBMS
1✓---36.1231.3830.8937.43
2-✓--39.4839.2143.4531.96
3✓✓--43.6243.9444.0638.75
4✓✓✓-46.8546.0746.5238.64
5✓✓-✓46.3644.0147.4938.10
6✓✓✓✓47.5844.8848.8039.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.234, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Table 4. Ablation experiment on each loss in TLDR. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.287, + 0.44, + 0.389 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{\\text{TR}} \\)\\( \\mathcal{L}_{\\text{TG}} \\)TEOCBMS
1✓--44.9544.0944.2439.07
2✓-✓46.8546.0746.5238.64
3-✓-44.2543.5346.9038.96
4-✓✓46.3644.0147.4938.10
5✓✓-45.7044.1346.2038.57
6✓✓✓47.5844.8848.8039.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.398, + 0.47, + 0.438 + ], + "angle": 0, + "content": "Table 5. Ablation experiments on TEO. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.457, + 0.469, + 0.498 + ], + "angle": 0, + "content": "model can effectively distinguish between classes with similar shapes by encoding and using texture as additional discriminative cues." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.508, + 0.23, + 0.523 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.53, + 0.469, + 0.571 + ], + "angle": 0, + "content": "In ablation experiments, we use ResNet-101 as the encoder, train on GTA, and evaluate on C, B, M, and S. The best results are highlighted in tables." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.469, + 0.726 + ], + "angle": 0, + "content": "Loss components. We investigate how each loss component contributes to overall performance. Table 4 shows the mIoU performance change with respect to the ablation of loss components. There is a large performance improvement when the stylized task loss and the original task loss are used simultaneously (cf. row 3). It seems to be the effect of using shape and texture cues as complementary. There is an additional performance improvement when the texture regularization loss and the texture generalization loss are added separately (cf. rows 4 and 5). The overall performance is the best when both losses are used (cf. row 6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Texture Extraction Operator (TEO). We use TEO to extract only texture features from feature maps. We conduct experiments to verify whether TEO leads to performance improvement in the texture regularization loss and the texture generalization loss. In the absence of TEO, \\(\\mathcal{L}_{\\mathrm{TR}}\\) and \\(\\mathcal{L}_{\\mathrm{TG}}\\) refer to the calculation of direct consistency in the feature maps without any alterations. Table 5 shows the results of TEO ablation experiments. The results show that the performance is improved in the presence of TEO (cf. rows 2, 4, and 6) in both \\(\\mathcal{L}_{\\mathrm{TR}}\\) and \\(\\mathcal{L}_{\\mathrm{TG}}\\) compared to the absence of" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.123, + 0.858, + 0.209 + ], + "angle": 0, + "content": "
CaseCBMS
1w/o RSM46.5744.8947.1738.87
2w/ RSM (τ=0.01)46.9244.0147.7538.61
3w/ RSM (τ=0.1)47.5844.8848.8039.35
4w/o LDF46.7744.8446.8739.72
5w/ LDF47.5844.8848.8039.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.218, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Table 6. Ablation experiments on the design choices of TLDR. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.272, + 0.862, + 0.333 + ], + "angle": 0, + "content": "
Random StyleCBMS
1WikiArt [41]45.0944.5845.1639.52
2DTD [11]45.6545.1045.2939.78
3ImageNet val [13]47.5844.8848.8039.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.342, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Table 7. Ablation experiment on different random style datasets. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.401, + 0.681, + 0.415 + ], + "angle": 0, + "content": "TEO (cf. rows 1, 3, and 5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.543 + ], + "angle": 0, + "content": "Random Style Masking (RSM). We conduct an experiment to validate the effectiveness of RSM. As shown in Table 6, the model with RSM (cf. row 3) performs better than the model without RSM (cf. row 1). Also, when the threshold of RSM is decreased to \\(\\tau = 0.01\\) (cf. row 2), the performance is higher than without RSM (cf. row 1) but lower than the default setting \\(\\tau = 0.1\\) (cf. row 3). The experiment suggests that maintaining an appropriate RSM threshold can aid in learning random styles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.698 + ], + "angle": 0, + "content": "Linear Decay Factor (LDF). When analyzing the texture regularization and generalization losses, we observe that the losses do not naturally decrease as much as the task losses (see Appendix B). The texture regularization loss remains relatively constant likely due to the frozen ImageNet model. The texture generalization loss is also relatively stable likely due to the different combinations of source images and random style images used in each epoch. We design LDF to match the scale of the losses. As shown in Table 6, the model with LDF (cf. row 5) performs better than the model without LDF (cf. row 4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Random style dataset. We analyze how performance changes when training on various random style datasets. Table 7 shows the performance when using WikiArt [41] dataset, Describable Texture Dataset (DTD) [11] and ImageNet [13] validation set as the random style dataset. The overall performance is the best when using the ImageNet validation set as the random style dataset (cf. row 3). The performances remain consistently demonstrated even using the other two datasets (cf. rows 1 and 2). The performance improvement when using ImageNet is probably due to the superiority of the ImageNet pre-trained model to extract the texture present in ImageNet." + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "684" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.125, + 0.196, + 0.14 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.149, + 0.473, + 0.329 + ], + "angle": 0, + "content": "Texture often contributes to the domain gap in DGSS. Existing DGSS methods have attempted to eliminate or randomize texture features. However, this paper argues that texture remains supplementary prediction cues for shape despite the domain gap. Accordingly, we proposed TLDR to learn texture features without overfitting to source domain textures in DGSS. TLDR includes novel texture regularization and generalization losses, using a Gram-matrix as a key component. We conducted a diverse set of experiments to demonstrate that TLDR effectively differentiates shape-similar classes by leveraging texture as a prediction cue. The experiments on multiple DGSS tasks show that our TLDR achieves state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.444 + ], + "angle": 0, + "content": "Limitation. During the experiments, we found that texture differences between domains vary class-wise. Learning more source domain textures for classes with small texture differences may help with generalization, but it may be advantageous for classes with high texture differences to learn less of them. In our method, there is no class-wise prescription for different texture differences. We leave it as an interesting future work." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.458, + 0.245, + 0.473 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.579 + ], + "angle": 0, + "content": "We sincerely thank Chanyong Lee and Eunjin Koh for their constructive discussions and support. We also appreciate Chaehyeon Lim, Sihyun Yu, Seokhyun Moon, and Youngju Yoo for providing insightful feedback. This work was supported by the Agency for Defense Development (ADD) grant funded by the Korea government (912855101)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.593, + 0.174, + 0.608 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.617, + 0.47, + 0.68 + ], + "angle": 0, + "content": "[1] Vijay Badrinarayan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 39(12):2481-2495, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.683, + 0.47, + 0.734 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.736, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[3] John Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), (6):679-698, 1986. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.776, + 0.469, + 0.839 + ], + "angle": 0, + "content": "[4] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2229-2238, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.841, + 0.469, + 0.867 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image" + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.617, + 0.47, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.127, + 0.892, + 0.178 + ], + "angle": 0, + "content": "segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 40(4):834-848, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.18, + 0.892, + 0.229 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.232, + 0.892, + 0.295 + ], + "angle": 0, + "content": "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 801-818, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.297, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[8] Wuyang Chen, Zhiding Yu, Shalini De Mello, Sifei Liu, Jose M Alvarez, Zhangyang Wang, and Anima Anandkumar. Contrastive syn-to-real generalization. arXiv preprint arXiv:2104.02290, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.349, + 0.892, + 0.4 + ], + "angle": 0, + "content": "[9] Wuyang Chen, Zhiding Yu, Zhangyang Wang, and Animashree Anandkumar. Automated synthetic-to-real generalization. In Proceedings of the International Conference on Machine Learning (ICML), pages 1746-1756, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.402, + 0.892, + 0.478 + ], + "angle": 0, + "content": "[10] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11580-11590, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.479, + 0.892, + 0.541 + ], + "angle": 0, + "content": "[11] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606-3613, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.544, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3213-3223, 2016. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.634, + 0.892, + 0.696 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248-255, 2009. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.699, + 0.892, + 0.749 + ], + "angle": 0, + "content": "[14] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.751, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[15] Mark Everingham, SM Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision (IJCV), 111(1):98-136, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.816, + 0.892, + 0.865 + ], + "angle": 0, + "content": "[16] Leon Gatys, Alexander S Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. Advances in Neural Information Processing Systems (NeurIPS), 28, 2015. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "685" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[17] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.47, + 0.241 + ], + "angle": 0, + "content": "[18] Yunhao Ge, Yao Xiao, Zhi Xu, Xingrui Wang, and Laurent Itti. Contributions of Shape, Texture, and Color in Visual Recognition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 369-386, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.243, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[19] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.469, + 0.368 + ], + "angle": 0, + "content": "[20] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.369, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[21] Robert M Haralick. Statistical and structural approaches to texture. Proceedings of the IEEE, 67(5):786-804, 1979. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.396, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2961-2969, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[24] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6891-6902, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.611 + ], + "angle": 0, + "content": "[25] Md Amirul Islam, Matthew Kowal, Patrick Esser, Sen Jia, Bjorn Ommer, Konstantinos G Derpanis, and Neil Bruce. Shape or texture: Understanding discriminative features in cnns. arXiv preprint arXiv:2101.11604, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.612, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[26] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and Statistical Texture Knowledge Distillation for Semantic Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16876-16885, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.75 + ], + "angle": 0, + "content": "[27] Yunpei Jia, Jie Zhang, Shiguang Shan, and Xilin Chen. Single-side domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8484-8493, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.752, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[28] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. WildNet: Learning Domain Generalized Semantic Segmentation from the Wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9936-9946, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.866 + ], + "angle": 0, + "content": "[29] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.128, + 0.891, + 0.152 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 224-233, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.153, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[30] Yijun Li, Ming-Yu Liu, Xueting Li, Ming-Hsuan Yang, and Jan Kautz. A closed-form solution to photorealistic image stylization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 453-468, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.267 + ], + "angle": 0, + "content": "[31] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 624-639, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.268, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[32] Yanghao Li, Naiyan Wang, Jiaying Liu, and Xiaodi Hou. Demystifying neural style transfer. arXiv preprint arXiv:1701.01036, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[33] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 40(12):2935-2947, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.394 + ], + "angle": 0, + "content": "[34] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3431-3440, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.395, + 0.892, + 0.42 + ], + "angle": 0, + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[36] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain generalization via invariant feature representation. In Proceedings of the International Conference on Machine Learning (ICML), pages 10-18, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.485, + 0.892, + 0.535 + ], + "angle": 0, + "content": "[37] Gilhyun Nam, Gyeongjae Choi, and Kyungmin Lee. GCISG: Guided Causal Invariant Learning for Improved Syn-to-Real Generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 656–672, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.536, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[38] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8690-8699, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.599, + 0.892, + 0.661 + ], + "angle": 0, + "content": "[39] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. Advances in Neural Information Processing Systems (NeurIPS), 34:23296-23308, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.662, + 0.892, + 0.726 + ], + "angle": 0, + "content": "[40] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4990-4999, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.727, + 0.892, + 0.75 + ], + "angle": 0, + "content": "[41] Kiri Nichol. Painter by numbers, wikiart. Kiri Nichol, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.752, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[42] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.865 + ], + "angle": 0, + "content": "[43] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1863-1871, 2019. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "686" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[44] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2594-2605, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.243 + ], + "angle": 0, + "content": "[45] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.245, + 0.469, + 0.295 + ], + "angle": 0, + "content": "[46] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12556-12565, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.296, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[47] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Proceedings of the European conference on computer vision (ECCV), pages 102-118, 2016. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[48] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3234-3243, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[49] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 618-626, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.47, + 0.567 + ], + "angle": 0, + "content": "[50] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10023-10031, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.469, + 0.606 + ], + "angle": 0, + "content": "[51] Nathan Somavarapu, Chih-Yao Ma, and Zsolt Kira. Frustratingly simple domain generalization via image stylization. arXiv preprint arXiv:2006.11207, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[52] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research (JMLR), 9(11), 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.709 + ], + "angle": 0, + "content": "[53] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.712, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[54] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. SiamDoGe: Domain Generalizable Semantic Segmentation Using Siamese Network. In Proceedings of the European Conference on Computer Vision (ECCV), pages 603-620, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.777, + 0.469, + 0.839 + ], + "angle": 0, + "content": "[55] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems (NeurIPS), 34:12077-12090, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.841, + 0.469, + 0.867 + ], + "angle": 0, + "content": "[56] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generaliza" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.127, + 0.892, + 0.165 + ], + "angle": 0, + "content": "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14383-14392, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[57] Zheng Xu, Wen Li, Li Niu, and Dong Xu. Exploiting low-rank structure from latent domains for domain generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 628-643, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.294 + ], + "angle": 0, + "content": "[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2636-2645, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[59] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2100-2110, 2019. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.892, + 0.449 + ], + "angle": 0, + "content": "[60] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 535-552, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.45, + 0.892, + 0.499 + ], + "angle": 0, + "content": "[61] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. arXiv preprint arXiv:2207.04892, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.501, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[62] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. arXiv preprint arXiv:2104.02008, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.541, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[63] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12537-12546, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.473, + 0.888, + 0.498, + 0.898 + ], + "angle": 0, + "content": "687" + } + ] +] \ No newline at end of file diff --git a/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/5378f03d-031c-45ab-a522-6e62326e01af_origin.pdf b/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/5378f03d-031c-45ab-a522-6e62326e01af_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b2c4aca4cbb37738c57e156017489136a0cd0893 --- /dev/null +++ b/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/5378f03d-031c-45ab-a522-6e62326e01af_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0aba43248b71d54adccab7a6aeeee7ea51fad48b4acd51452ca8c44592b9eee5 +size 1719489 diff --git a/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/full.md b/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1fe6638ebccb1f10688d29febee2671bd8b1880a --- /dev/null +++ b/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/full.md @@ -0,0 +1,360 @@ +# Texture Learning Domain Randomization for Domain Generalized Segmentation + +Sunghwan Kim + +Dae-hwan Kim + +Hoseong Kim* + +Agency for Defense Development (ADD) + +{sssshwan，dhkim7，hoseongkim}@add.re.kr + +# Abstract + +Deep Neural Networks (DNNs)-based semantic segmentation models trained on a source domain often struggle to generalize to unseen target domains, i.e., a domain gap problem. Texture often contributes to the domain gap, making DNNs vulnerable to domain shift because they are prone to be texture-biased. Existing Domain Generalized Semantic Segmentation (DGSS) methods have alleviated the domain gap problem by guiding models to prioritize shape over texture. On the other hand, shape and texture are two prominent and complementary cues in semantic segmentation. This paper argues that leveraging texture is crucial for improving performance in DGSS. Specifically, we propose a novel framework, coined Texture Learning Domain Randomization (TLDR). TLDR includes two novel losses to effectively enhance texture learning in DGSS: (1) a texture regularization loss to prevent overfitting to source domain textures by using texture features from an ImageNet pretrained model and (2) a texture generalization loss that utilizes random style images to learn diverse texture representations in a self-supervised manner. Extensive experimental results demonstrate the superiority of the proposed TLDR; e.g., TLDR achieves 46.5 mIoU on GTA $\rightarrow$ Cityscapes using ResNet-50, which improves the prior state-of-the-art method by 1.9 mIoU. The source code is available at https://github.com/ssssshwan/TLDR. + +# 1. Introduction + +Semantic segmentation is an essential task in computer vision with many real-world applications, such as autonomous vehicles, augmented reality, and medical imaging. Deep Neural Networks (DNNs)-based semantic segmentation models work well when the data distributions are consistent between a source domain and target domains [1,5-7,22,34,55]. However, the performance of the models tends to degrade significantly in practical settings that in + +![](images/ae0071673eeea8563b05c01168db66a957eddeeb2b50ce19075e12e7dc3ad200.jpg) +Unseen Target Image + +![](images/f8a852b6b9d1742bca238f8ef771078e2b33d29d99d4d88cb3957a50d26220fd.jpg) +(a) Normalization & Whitening + +![](images/81d2ee6b752b10cd2c5a91a9f446266da65d16f984f80ca6c14149233d2d2d84.jpg) +(b) Domain Randomization + +![](images/9f789ebab88827300ca44af2cc259dd87268ab796d6d1269d1c1f8606b7fa095.jpg) +(c) TLDR (ours) +Figure 1. Segmentation results for an image from an unseen domain (i.e., Cityscapes [12]), using models trained on GTA [47] with DGSS methods. (a-b) Existing methods [10, 59] have difficulty in distinguishing between road, sidewalk, and terrain which have similar shapes and contexts, as the texture is not sufficiently considered during the training process. (c) Our Texture Learning Domain Randomization (TLDR) can distinguish the classes effectively as we utilize the texture as prediction cues. + +volve unseen out-of-distribution scenarios, also known as a domain gap problem. Many domain adaptation and generalization methods have been proposed to solve the domain gap problem [10, 24, 28, 42-44, 54, 59, 60]. Domain adaptation assumes accessibility of target domain images, differing from domain generalization. This paper addresses Domain Generalized Semantic Segmentation (DGSS), which aims to train models that can generalize to diverse unseen domains by training on a single source domain. + +Existing DGSS methods have attempted to address the domain gap problem by guiding models to focus on shape rather than texture. Given that texture often varies across different domains (e.g., synthetic/real and sunny/rainy/foggy), DNNs are susceptible to domain shift because they tend to be texture-biased [19, 39]. Accordingly, there are two main approaches for the DGSS methods. The first approach is Normalization and Whitening + +![](images/e7da4b5f40d6d796a760fb8112d63d57bb22a188473c1008c2dd804920c2ba03.jpg) +Original Source +Figure 2. Reconstructed source images from the feature maps of (a) normalization and whitening and (b) domain randomization. Texture features are often omitted in existing DGSS methods. + +![](images/3263f88a5d78f936c88be60963d0c6ba7807a1d643a4629ef9bf0c3b0bbc73f2.jpg) +(a) Norm. & Whiten. + +![](images/8d525b63b1d1b0b786906e033988aee3a74fc12f8eca834a9a08ef14bdbdc140.jpg) +(b) Domain Random. + +![](images/574fefa65ff72ecd66d6e237388c56e580afdf452b7bfd87cd366d6b02abaf56.jpg) +(a) Shape Features +Figure 3. The t-SNE [52] plots for the road, sidewalk, and terrain classes from Cityscapes [12] that have similar shapes. While the shape features (Canny edge [3]) are entangled in (a), the texture features (Gram-matrix) of these classes are clearly separated in (b). The plots are based on an ImageNet pre-trained model. + +![](images/777bc6c73a6047f506102ae422d0deacf5ab2b81255b3aabbaa0349536a68580.jpg) +(b) Texture Features + +(NW), which involves normalizing and whitening the features [10, 42-44]. It is possible to remove domain-specific texture features and learn domain-invariant shape features with NW (see Figure 2a). The second approach is Domain Randomization (DR), which trains by transforming source images into randomly stylized images [24,28,45,54,59-61]. The model learns domain-invariant shape features because texture cues are mostly replaced by random styles (see Figure 2b) [38, 51]. + +While the existing methods are effective at making the models focus on shape features, they need to give more consideration to texture features. In addition to utilizing shape features like edges and structures, DNNs also use texture features such as patterns, color variations, and histograms as important cues for prediction [18]. Particularly in semantic segmentation, texture plays a crucial role in accurately maintaining the boundaries between objects [26, 63]. + +Figure 1 demonstrates the results of predicting an unseen image from Cityscapes [12] using DGSS methods trained on GTA [47]. One can see that the models trained with NW and DR have difficulty distinguishing between the road, sidewalk, and terrain classes, which have similar shapes and contexts. In order to determine these classes accurately, it is necessary to use texture cues. This assertion is further emphasized through t-SNE [52] plots of the classes: the shape features are entangled in Figure 3a, whereas the texture features are clearly separated in Figure 3b. Meanwhile, some textures remain relatively unchanged across domains in DGSS as shown in Figure 4. Based on these observations, we suggest utilizing texture as valuable cues for prediction. + +We propose Texture Learning Domain Randomization (TLDR), which enables DGSS models to learn both shape and texture. Since only source domain images are available + +![](images/f130c7df74f6a7f53667abe8be917e6f04a64181f981afe366380cf53dcda358.jpg) +Figure 4. Visualization of texture for the road, sidewalk, and terrain classes from GTA [47], Cityscapes [12], BDD [58], and Mapillary [40] datasets. For each class, there are commonalities in texture across the datasets. + +in DGSS, texture must be learned from them. To accurately capture the texture, we leverage the source domain images without any modification, which we refer to as the original source images. The stylized source images from DR are more focused on learning shape features, and the original source images are more focused on learning texture features (Section 4.1). To further improve texture learning in DGSS, we propose a texture regularization loss and a texture generalization loss. While there are commonalities in texture across different domains, there are also clear texture differences. Thus, if the model overfits source domain textures, it will result in a performance drop. To mitigate this problem, we propose the texture regularization loss to prevent overfitting to source domain textures by using texture features from an ImageNet pre-trained model (Section 4.2). Since only source domain textures alone may not be sufficient for learning general texture representations, we propose the texture generalization loss that utilizes random style images to learn diverse texture representations in a self-supervised manner (Section 4.3). + +Our contribution can be summarized into three aspects. First, to the best of our knowledge, we are approaching DGSS for the first time from both the shape and texture perspectives. We argue that leveraging texture is essential in distinguishing between classes with similar shapes despite the domain gap. Second, to enhance texture learning in DGSS, we introduce two novel losses: the texture regularization loss and the texture generalization loss. Third, extensive experiments over multiple DGSS tasks show that our proposed TLDR achieves state-of-the-art performance. Our method attains $46.5\mathrm{mIoU}$ on $\mathrm{GTA} \rightarrow$ Cityscapes using ResNet-50, surpassing the prior state-of-the-art method by $1.9\mathrm{mIoU}$ . + +![](images/37e9b48855740e1681c51867b98a94f7d1b42bad75a75d268823c0e5bc0be709.jpg) +(a) Style Transfer Module (STM) + +![](images/c2332a36b4f7268c22715b33c201c0891d41b72ce84065419b1077d1f2ed42cf.jpg) +(b) Texture Extraction Operator (TEO) +Figure 5. (a) Visualization of using the Style Transfer Module (STM) to transform an original source image into stylized source images with random style images. (b) Illustration of the process of using Texture Extraction Operator (TEO) to extract only texture features, i.e., a Gram-matrix, from a feature map. + +# 2. Related Work + +Domain Generalization (DG). DG aims to learn DNNs that perform well on multiple unseen domains [36]. DG has primarily been studied in image classification. A number of studies have been proposed to address DG, including adversarial learning [27, 31, 50], data augmentation [46, 53, 56, 62], meta-learning [2, 14, 33], ensemble learning [57, 62], and self-supervised learning [4, 29]. + +Recent studies have attempted to train domain generalized models by preserving ImageNet pre-trained feature representations as much as possible. Chen et al. defined DG as a life-long learning problem [33] and tried to utilize ImageNet pre-trained weights to prevent catastrophic forgetting [9]. Contrastive learning [8] and attentional pooling [8, 37] were introduced to enhance the capturing of semantic knowledge in ImageNet features. In this paper, we regularize texture representations of DNNs with the ImageNet features. To the best of our knowledge, this is the first attempt to extract a specific semantic concept (i.e. texture) from the ImageNet features for regularization in DG. + +Domain Generalized Semantic Segmentation (DGSS). DGSS is in its early stages and has yet to receive much research attention. Existing DGSS methods have tried to alleviate a domain gap problem through two main approaches: Normalization and Whitening (NW) and Domain Randomization (DR). NW trains by normalizing the mean and standard deviation of source features and whitening the covariance of source features. This process eliminates domain-specific features, allowing the model to learn domain-invariant features. Pan et al. introduced instance normalization to remove domain-specific features [42]. Pan et al. proposed switchable whitening to decorrelate features [43], and Choi et al. proposed Instance Selective Whitening (ISW) to enhance the ability to whiten domain-specific features [10]. Peng et al. tried to normalize and whiten features in a category-wise manner [44]. + +DR trains by transforming source images into randomly stylized images. The model is guided to capture domain-invariant shape features since texture cues are substituted + +with random styles. Yue et al. presented a method for considering consistency among multiple stylized images [59]. Peng et al. distinguished between global and local texture in the randomization process [45]. Huang et al. proposed DR in frequency space [24]. Adversarial learning [61] and self-supervised learning [54] have been used as attempts to make style learnable rather than using random style images. It has been shown that utilizing content and style of ImageNet [28] and ImageNet pre-trained features [60] can aid in learning generalized features in DR. + +Existing DR methods have yet to comprehensively address DGSS from both shape and texture perspectives. Several studies [28, 59] have stated that the success of DR is attributed to its ability to learn various styles, resulting in improved performance in generalized domains. However, from a different perspective, we consider that the effectiveness of DR is due to the model becoming more focused on shape, as discussed in recent DG methods [38, 51]. Therefore, we assume that the model needs to learn texture for further performance improvement in DR. We propose a novel approach for learning texture in DGSS without overfitting to source domain textures. + +# 3. Preliminaries + +# 3.1. Domain Randomization with Style Transfer + +We adopt DR as a baseline DGSS method. We use a neural style transfer method at each epoch to transform each original source image into a different random style. If edge information is lost during the style transfer process, it may cause a mismatch with the semantic label, leading to a decrease in performance. We utilize photoWCT [30] known as an edge-preserving style transfer method. Random style images are sampled from ImageNet [13] validation set. Figure 5a is a visualization of using the Style Transfer Module (STM) to transform an original source image into stylized source images with random style images. + +![](images/d7777b91dc5dc3f15613d57574226280baccf338af342435c6095641e0828cb3.jpg) +Figure 6. An overview of our proposed Texture Learning Domain Randomization (TLDR) framework. The stylized source image $x^{sr}$ (green) is obtained by stylizing the original source image $x^s$ (blue) with the random style image $x^r$ (purple). The stylized task loss $\mathcal{L}_{\mathrm{styl}}$ focuses on learning shape from $x^{sr}$ , and the original task loss $\mathcal{L}_{\mathrm{orig}}$ focuses on learning texture from $x^s$ . The texture regularization loss $\mathcal{L}_{\mathrm{TR}}$ enforces the consistency between the Gram-matrices of the ImageNet model $f_I$ and the task model $f_T$ for $x^s$ . The texture generalization loss $\mathcal{L}_{\mathrm{TG}}$ enforces the consistency between the Gram-matrices of the task model $f_T$ for $x^{sr}$ and $x^r$ . Random Style Masking (RSM) selects only the random style features when applying the texture generalization loss. + +# 3.2. Texture Extraction with Gram-matrix + +Texture is a regional descriptor that can offer measurements for both local structural (e.g., pattern) and global statistical (e.g., overall distribution of colors) properties of an image [21]. It has been shown that texture can be represented by pairwise correlations between features extracted by DNNs, also known as a Gram-matrix [16, 17, 32]. + +We use the Gram-matrix to extract only texture features from a feature map. The Gram-matrix $G \in \mathbb{R}^{C \times C}$ for the vectorized matrix $F \in \mathbb{R}^{C \times HW}$ of the feature map is defined as Equation 1. $C, H$ , and $W$ denote the channel, height, and width of the feature map, respectively. + +$$ +G _ {i, j} = F _ {i} \cdot F _ {j}, \tag {1} +$$ + +where $\cdot$ represents a dot product, $F_{i}$ and $F_{j}$ are the $i^{th}$ and $j^{th}$ row vectors of $F$ , respectively. $G_{i,j}$ is the entry at the $i^{th}$ row and $j^{th}$ column of $G$ . Each entry of the Grammatrix indicates a pairwise correlation between the features corresponding to a texture feature. In this paper, the operator extracting texture features is called Texture Extraction Operator (TEO). Figure 5b illustrates the process of TEO. + +# 4. Approach + +This section describes our proposed Texture Learning Domain Randomization (TLDR) framework. TLDR learns texture features in addition to learning shape features through domain randomization. TLDR consists of four losses: a stylized task loss, an original task loss, a tex + +ture regularization loss, and a texture generalization loss. The stylized task loss focuses on learning shape, and the original task loss focuses on learning texture. The texture regularization loss and the texture generalization loss prevent overfitting to source domain textures. Figure 6 is an overview of TLDR. The frozen ImageNet pre-trained encoder is denoted as $f_{I}$ , i.e., the ImageNet model. The training encoder is denoted as $f_{T}$ , i.e., the task model. The training semantic segmentation decoder is denoted as $h_{T}$ . + +# 4.1. Task Losses + +Stylized task loss. We denote an original source image $x^{s}$ and its semantic label $y$ . $x^{sr}$ is a stylized source image obtained by stylizing an original source image $x^{s}$ with a random style image $x^{r}$ . The prediction result for $x^{sr}$ of the model is $p^{sr} = h_{T}(f_{T}(x^{sr}))$ . Then the stylized task loss $\mathcal{L}_{\mathrm{styl}}$ is given by Equation 2. + +$$ +\mathcal {L} _ {\text {s t y l}} = \mathrm {C E} \left(p ^ {s r}, y\right), \tag {2} +$$ + +where $\mathrm{CE}(\cdot)$ represents the categorical cross-entropy loss. The stylized task loss encourages the model to focus on shape features during training since the texture cues are mostly replaced by random styles [38, 51]. + +Original task loss. The model struggles to learn texture from the stylized source images as the texture cues are mostly substituted with random styles. To accurately capture the source domain textures, the model is trained on the original source images. The prediction result for $x^{s}$ of the + +model is $p^s = h_T(f_T(x^s))$ . The original task loss $\mathcal{L}_{\mathrm{orig}}$ is given by Equation 3. + +$$ +\mathcal {L} _ {\text {o r i g}} = \mathrm {C E} \left(p ^ {s}, y\right). \tag {3} +$$ + +DNNs tend to prioritize texture cues without restrictions [19, 39]. The original task loss guides the model to concentrate on texture features during training. + +# 4.2. Texture Regularization Loss + +If the model is trained on the original source images without regularization, it will overfit the source domain textures [19, 39]. We regularize texture representations of the task model using the ImageNet model, which encodes diverse feature representations [8]. However, it is important to note that ImageNet features include not only texture features, but also other semantic features such as shape features. We therefore assume that regularizing the entire feature may interfere with the task model learning texture. To address this issue, we propose to apply TEO to extract only texture features from ImageNet features and regularize the task model with the extracted texture features (see Table 5 for ablation). + +Let $F_{I}^{l,s}$ and $F_{T}^{l,s}$ denote the vectorized feature maps in layer $l$ of the ImageNet model $f_{I}$ and the task model $f_{T}$ for $x^{s}$ , respectively. The Gram-matrices $G_{I}^{l,s}$ and $G_{T}^{l,s}$ from $F_{I}^{l,s}$ and $F_{T}^{l,s}$ are the texture features of the original source image as seen by $f_{I}$ and $f_{T}$ , respectively. The contribution of the $l^{th}$ layer to the texture regularization loss is $\| G_{I}^{l,s} - G_{T}^{l,s}\|_{2}$ . The total texture regularization loss is given by Equation 4. + +$$ +\mathcal {L} _ {\mathrm {T R}} = \sum_ {l = 1} ^ {L} \frac {u _ {l}}{C _ {l} H _ {l} W _ {l}} \| G _ {I} ^ {l, s} - G _ {T} ^ {l, s} \| _ {2}, \tag {4} +$$ + +where $L$ is the number of feature map layers, and $u_{l}$ is a weighting factor for the contribution of the $l^{th}$ layer to $\mathcal{L}_{\mathrm{TR}}$ . $C_l, H_l,$ and $W_{l}$ denote the channel, height and width of the $l^{th}$ layer feature map, respectively. We reduce the value of $u_{l}$ as $l$ increases, considering that fewer texture features are encoded in feature maps as layers become deeper [25]. + +# 4.3. Texture Generalization Loss + +We supplement texture learning from the random style images for more diverse texture representations. Since the random style images are unlabeled, the texture representations should be learned self-supervised. Note that the random style image $x^{r}$ and the stylized source image $x^{sr}$ share some texture features. To encourage learning of diverse texture representations, we induce the texture features of $x^{sr}$ to become more similar to those of $x^{r}$ as much as possible while preserving source texture features. + +Let $F_T^{l,r}$ and $F_T^{l,sr}$ denote the vectorized feature maps in layer $l$ of $f_T$ for the random style image $x^r$ and the stylized source image $x^{sr}$ , respectively. $G_T^{l,r}$ and $G_T^{l,sr}$ are the + +![](images/33b977f9373ea0340b092317e794bc0edf6abe173e5f0166c1809898363a9ebe.jpg) +Figure 7. Random Style Masking (RSM) masks only the entries corresponding to the random style features (purple) in the Grammatrices. The difference between $G_T^{l,s\tau}$ and $G_T^{l,s}$ is calculated as $D^l$ . The entries with values greater than a certain threshold $\tau$ in $D^l$ are masked as $M^l$ . + +corresponding Gram-matrices. Our goal is basically to set an objective that makes $G_T^{l,r}$ and $G_T^{l,sr}$ consistent. However, $G_T^{l, sr}$ includes both random style features and remaining source texture features. Applying the constraints to the entire $G_T^{l, sr}$ also imposes the objective on the source texture features in the stylized source image. To select only the random style features when enforcing the consistency between $G_T^{l,r}$ and $G_T^{l, sr}$ , we propose Random Style Masking (RSM), inspired by ISW [10]. Figure 7 is an illustration of RSM. We assume that the entries corresponding to the random style features are activated in $G_T^{l, sr}$ but deactivated in $G_T^{l,s}$ . Considering $D^l = G_T^{l, sr} - G_T^{l,s}$ , the entries corresponding to the random style features are expected to be larger than a certain threshold $\tau$ . We denote the mask for the entries representing the random style features as $M^l$ (see Equation 5). + +$$ +M _ {i, j} ^ {l} = \left\{ \begin{array}{l l} 1, & \text {i f} D _ {i, j} ^ {l} > \tau \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {5} +$$ + +where $i$ and $j$ are the row and column indices in each matrix, respectively. The threshold $\tau$ is determined empirically (see Table 6 for ablation). We only apply the objective to selected random style features by RSM. The contribution of the $l^{th}$ layer to the texture generalization loss is $\|(G_T^{l,r} - G_T^{l,sr})\odot M^l\|_2$ , where $\odot$ represents an element-wise product. The total texture generalization loss is given by Equation 6. + +$$ +\mathcal {L} _ {\mathrm {T G}} = \sum_ {l = 1} ^ {L} \frac {v _ {l}}{C _ {l} H _ {l} W _ {l}} \| \left(G _ {T} ^ {l, r} - G _ {T} ^ {l, s r}\right) \odot M ^ {l} \| _ {2}, \tag {6} +$$ + +where $v_{l}$ is a weighting factor for the contribution of the $l^{th}$ layer to $\mathcal{L}_{\mathrm{TG}}$ . We also reduce the value of $v_{l}$ as $l$ increases for the same reason as $u_{l}$ in Equation 4. + +# 4.4. Full Objective + +As the training progresses, the task losses decrease while the texture regularization loss and the texture generalization loss remain relatively constant. We add a Linear Decay Factor (LDF) to the texture regularization and generalization losses to balance their scale with the task losses + +
MethodEncoderCBMS
DRPC [59]37.4232.1434.12-
RobustNet [10]36.5835.2040.3328.30
SAN-SAW [44]39.7537.3441.8630.79
SiamDoGe [54]ResNet-5042.9637.5440.6428.34
WildNet [28]44.6238.4246.0931.34
SHADE [60]44.6539.2843.34-
TLDR (ours)46.5142.5846.1836.30
DRPC [59]42.5338.7238.0529.67
GTR [45]43.7039.6039.1029.32
FSDR [24]44.8041.2043.40-
SAN-SAW [44]ResNet-10145.3341.1840.7731.84
WildNet [28]45.7941.7347.0832.51
SHADE [60]46.6643.6645.50-
TLDR (ours)47.5844.8848.8039.35
+ +(see Table 6 for ablation). The LDF at iteration $t$ is set $w(t) = (1 - t / t_{\mathrm{total}})$ , where $t_{\mathrm{total}}$ denotes the total number of iterations. Our full objective is given by Equation 7. + +$$ +\mathcal {L} _ {\text {t o t a l}} = \alpha_ {\text {o r i g}} \mathcal {L} _ {\text {o r i g}} + \alpha_ {\text {s t y l}} \mathcal {L} _ {\text {s t y l}} + w (t) \mathcal {L} _ {\text {T R}} + w (t) \mathcal {L} _ {\text {T G}}, \tag {7} +$$ + +where $\alpha_{\mathrm{orig}}$ and $\alpha_{\mathrm{styl}}$ are the weights for the original task loss and the stylized task loss, respectively. + +# 5. Experiments + +# 5.1. Implementation Details + +Datasets. As synthetic datasets, GTA [47] consists of 24,966 images with a resolution of $1914 \times 1052$ . It has 12,403, 6,382, and 6,181 images for training, validation, and test sets. SYNTHIA [48] contains 9,400 images with a resolution of $1280 \times 760$ . It has 6,580 and 2,820 images for training and validation sets, respectively. + +As real-world datasets, Cityscapes [12] consists of 2,975 training images and 500 validation images with a resolution of $2048 \times 1024$ . BDD [58] contains 7,000 training images and 1,000 validation images with a resolution of $1280 \times 720$ . Mapillary [40] involves 18,000 training images and 2,000 validation images with diverse resolutions. For brevity, we denote GTA, SYNTHIA, Cityscapes, BDD, and Mapillary as G, S, C, B, and M, respectively. + +Network architecture. We conduct experiments using ResNet [23] as an encoder architecture and DeepLabV3+ [7] as a semantic segmentation decoder architecture. In all experiments, encoders are initialized with an ImageNet [13] pre-trained model. + +Training. We adopt an AdamW [35] optimizer. An initial + +Table 1. Comparison of mIoU (%; higher is better) between DGSS methods trained on GTA and evaluated on C, B, M, and S. The best and second best results are highlighted and underlined, respectively. Our method is marked in gray. + +
MethodEncoderCBMG
DRPC [59]35.6531.5332.7428.75
SAN-SAW [44]ResNet-5038.9235.2434.5229.16
TLDR (ours)41.8834.3536.7935.90
DRPC [59]37.5834.3434.1229.24
GTR [45]39.7035.3036.4028.71
FSDR [24]ResNet-10140.8039.6037.40-
SAN-SAW [44]40.8735.9837.2630.79
TLDR (ours)42.6035.4637.4637.77
+ +Table 2. Comparison of mIoU (%; higher is better) between DGSS methods trained on SYNTHIA and evaluated on C, B, M, and G. The best and second best results are highlighted and underlined, respectively. Our method is marked in gray. + +learning rate is set to $3 \times 10^{-5}$ for the encoder and $3 \times 10^{-4}$ for the decoder, $40k$ training iterations, a batch size of 4. A weight decay is set to 0.01, with a linear warmup [20] over $t_{\mathrm{warm}} = 1k$ iterations, followed by a linear decay. We use random scaling in the range [0.5, 2.0] and random cropping with a size of $768 \times 768$ . We apply additional data augmentation techniques, including random flipping and color jittering. We set the texture regularization parameters as $u_l = 5 \times 10^{-l-2}$ , and the texture generalization parameters as $v_l = 5 \times 10^{-l-2}$ . The original task loss and the stylized task loss weights are set to $\alpha_{\mathrm{orig}} = 0.5$ and $\alpha_{\mathrm{styl}} = 0.5$ , respectively. We set the RSM threshold $\tau = 0.1$ . + +# 5.2. Comparison with DGSS methods + +To measure generalization capacity in unseen domains, we train on a single source domain and evaluate multiple unseen domains. We conduct experiments on two settings, (1) $\mathrm{G} \rightarrow \{\mathrm{C}, \mathrm{B}, \mathrm{M}, \mathrm{S}\}$ and (2) $\mathrm{S} \rightarrow \{\mathrm{C}, \mathrm{B}, \mathrm{M}, \mathrm{G}\}$ . We repeat each benchmark three times, each time with a different random seed, and report the average results. We evaluate our method using ResNet-50 and ResNet-101 encoders. We use mean Intersection over Union (mIoU) [15] as the evaluation metric. The best and second best results are highlighted and underlined in tables, respectively. + +Tables 1 and 2 show the generalization performance of models trained on GTA [47] and SYNTHIA [48], respectively. Our TLDR generally outperforms other DGSS methods in most benchmarks. In particular, we improve the $\mathrm{G}\rightarrow \mathrm{C}$ benchmark by $+1.9\mathrm{mIoU}$ and $+0.9\mathrm{mIoU}$ on ResNet-50 and ResNet-101 encoders, respectively. + +# 5.3. Texture Awareness of Model + +We design experiments to verify whether the performance improvement of TLDR is actually due to being aware of texture. In the experiments, we compare TLDR to plain Domain Randomization (DR), which is only trained with the stylized task loss. We train on GTA using a ResNet-101 encoder. For convenience, we refer to the models trained + +![](images/652eb5ac72abe96eea1bc816e7c474a770983d1c4a4bfadd87a61db0e7ce2366.jpg) +Figure 8. Visualization of class activation maps for different classes using DR and our TLDR. The classes are road, sidewalk, and terrain. Only the classes are displayed as opaque for better visualization in the unseen images. The TLDR model tends to have activation throughout broader areas than the DR model, suggesting it can rely on more texture cues when making predictions. + +![](images/f291afe86909639704d8d49ef21f6cc14997bcd912729e375c7f7312d3a1bf74.jpg) +Figure 9. Comparison of the confusion matrices between DR and our TLDR for classes with similar shapes. The TLDR model tends to have less confusion and higher accuracy than the DR model. + +with DR and TLDR as the DR model and the TLDR model, respectively. The models are evaluated on Cityscapes in the experiments if no specific dataset is mentioned. + +Class activation map. To validate whether the models utilize texture cues for prediction, we generate class activation maps and analyze the contribution of texture features to the predictions. We apply Grad-CAM [49] to road, sidewalk, and terrain on the DR and TLDR models. Figure 8 displays the attention heat maps generated for each class, highlighting the regions the models focus on during prediction. Our analysis reveals that the DR model tends to have high activation in edge regions, while the TLDR model tends to have activation throughout broader areas. One can infer that these broader areas include texture cues, which the TLDR model uses as valuable prediction cues. + +Confusion matrix. We compare confusion matrices between the two models. Figure 9 shows the confusion matrices of the DR model and the TLDR model for shape-similar classes. The TLDR model has lower false positive rates in classifying sidewalk and terrain as road, with rates of + +![](images/79702bec39ca23c74a93bf52fd95e09694b242311f446047eb305b0edea6ce36.jpg) +Figure 10. Qualitative results of DR and our TLDR on unseen images. The unseen images are from Cityscapes, BDD, Mapillary, and SYNTHIA in row order. The TLDR model provides better prediction results for shape-similar classes than the DR model, as seen in the white boxes. + +
LayerDRTLDR
TextureShapeTextureShape
l = 166.5%29.4%66.7%29.2%
l = 256.6%39.8%57.3%39.1%
l = 332.0%63.5%33.8%61.9%
l = 419.1%54.3%20.1%52.5%
+ +Table 3. Layer-wise dimensionality of texture and shape in the latent representations of DR and our TLDR, evaluated in percentages $(\%)$ . The TLDR model stores more texture information than the DR model across all the layers. + +$19.6\%$ and $6.9\%$ , respectively, compared to the DR model, which has rates of $27.6\%$ and $11.6\%$ (see Figure 9a). Also, there is an increase in the accuracy of each class. There are also clear reductions in confusion between building, wall & fence and car, truck & bus (see Figures 9b and 9c). + +Qualitative results. Figure 10 shows the qualitative results of the DR and TLDR models in various domains. The TLDR model provides better prediction results for shape-similar classes than the DR model (see white boxes). + +Dimensionality analysis. We investigate whether the TLDR model encodes additional texture information compared to the DR model. We use the method proposed by Islam et al. to quantify the dimensions corresponding to texture and shape in latent representations [25]. The method estimates the dimensionality of a semantic concept by computing mutual information between feature representations of input pairs that exhibit the same semantic concept. A detailed explanation of the experiment is in Appendix A. Table 3 shows the estimated dimensions (\%) for texture and shape in each layer of the DR and TLDR models. At every layer, it is apparent that the TLDR model encodes more texture information than the DR model. + +The experimental results demonstrate that the TLDR + +
LorigLstylLTRLTGCBMS
1✓---36.1231.3830.8937.43
2-✓--39.4839.2143.4531.96
3✓✓--43.6243.9444.0638.75
4✓✓✓-46.8546.0746.5238.64
5✓✓-✓46.3644.0147.4938.10
6✓✓✓✓47.5844.8848.8039.35
+ +Table 4. Ablation experiment on each loss in TLDR. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray. + +
\( \mathcal{L}_{\text{TR}} \)\( \mathcal{L}_{\text{TG}} \)TEOCBMS
1✓--44.9544.0944.2439.07
2✓-✓46.8546.0746.5238.64
3-✓-44.2543.5346.9038.96
4-✓✓46.3644.0147.4938.10
5✓✓-45.7044.1346.2038.57
6✓✓✓47.5844.8848.8039.35
+ +model can effectively distinguish between classes with similar shapes by encoding and using texture as additional discriminative cues. + +# 5.4. Ablation Study + +In ablation experiments, we use ResNet-101 as the encoder, train on GTA, and evaluate on C, B, M, and S. The best results are highlighted in tables. + +Loss components. We investigate how each loss component contributes to overall performance. Table 4 shows the mIoU performance change with respect to the ablation of loss components. There is a large performance improvement when the stylized task loss and the original task loss are used simultaneously (cf. row 3). It seems to be the effect of using shape and texture cues as complementary. There is an additional performance improvement when the texture regularization loss and the texture generalization loss are added separately (cf. rows 4 and 5). The overall performance is the best when both losses are used (cf. row 6). + +Texture Extraction Operator (TEO). We use TEO to extract only texture features from feature maps. We conduct experiments to verify whether TEO leads to performance improvement in the texture regularization loss and the texture generalization loss. In the absence of TEO, $\mathcal{L}_{\mathrm{TR}}$ and $\mathcal{L}_{\mathrm{TG}}$ refer to the calculation of direct consistency in the feature maps without any alterations. Table 5 shows the results of TEO ablation experiments. The results show that the performance is improved in the presence of TEO (cf. rows 2, 4, and 6) in both $\mathcal{L}_{\mathrm{TR}}$ and $\mathcal{L}_{\mathrm{TG}}$ compared to the absence of + +Table 5. Ablation experiments on TEO. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray. + +
CaseCBMS
1w/o RSM46.5744.8947.1738.87
2w/ RSM (τ=0.01)46.9244.0147.7538.61
3w/ RSM (τ=0.1)47.5844.8848.8039.35
4w/o LDF46.7744.8446.8739.72
5w/ LDF47.5844.8848.8039.35
+ +Table 6. Ablation experiments on the design choices of TLDR. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray. + +
Random StyleCBMS
1WikiArt [41]45.0944.5845.1639.52
2DTD [11]45.6545.1045.2939.78
3ImageNet val [13]47.5844.8848.8039.35
+ +Table 7. Ablation experiment on different random style datasets. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray. + +TEO (cf. rows 1, 3, and 5). + +Random Style Masking (RSM). We conduct an experiment to validate the effectiveness of RSM. As shown in Table 6, the model with RSM (cf. row 3) performs better than the model without RSM (cf. row 1). Also, when the threshold of RSM is decreased to $\tau = 0.01$ (cf. row 2), the performance is higher than without RSM (cf. row 1) but lower than the default setting $\tau = 0.1$ (cf. row 3). The experiment suggests that maintaining an appropriate RSM threshold can aid in learning random styles. + +Linear Decay Factor (LDF). When analyzing the texture regularization and generalization losses, we observe that the losses do not naturally decrease as much as the task losses (see Appendix B). The texture regularization loss remains relatively constant likely due to the frozen ImageNet model. The texture generalization loss is also relatively stable likely due to the different combinations of source images and random style images used in each epoch. We design LDF to match the scale of the losses. As shown in Table 6, the model with LDF (cf. row 5) performs better than the model without LDF (cf. row 4). + +Random style dataset. We analyze how performance changes when training on various random style datasets. Table 7 shows the performance when using WikiArt [41] dataset, Describable Texture Dataset (DTD) [11] and ImageNet [13] validation set as the random style dataset. The overall performance is the best when using the ImageNet validation set as the random style dataset (cf. row 3). The performances remain consistently demonstrated even using the other two datasets (cf. rows 1 and 2). The performance improvement when using ImageNet is probably due to the superiority of the ImageNet pre-trained model to extract the texture present in ImageNet. + +# 6. Conclusion + +Texture often contributes to the domain gap in DGSS. Existing DGSS methods have attempted to eliminate or randomize texture features. However, this paper argues that texture remains supplementary prediction cues for shape despite the domain gap. Accordingly, we proposed TLDR to learn texture features without overfitting to source domain textures in DGSS. TLDR includes novel texture regularization and generalization losses, using a Gram-matrix as a key component. We conducted a diverse set of experiments to demonstrate that TLDR effectively differentiates shape-similar classes by leveraging texture as a prediction cue. The experiments on multiple DGSS tasks show that our TLDR achieves state-of-the-art performance. + +Limitation. During the experiments, we found that texture differences between domains vary class-wise. Learning more source domain textures for classes with small texture differences may help with generalization, but it may be advantageous for classes with high texture differences to learn less of them. In our method, there is no class-wise prescription for different texture differences. We leave it as an interesting future work. + +# Acknowledgements + +We sincerely thank Chanyong Lee and Eunjin Koh for their constructive discussions and support. We also appreciate Chaehyeon Lim, Sihyun Yu, Seokhyun Moon, and Youngju Yoo for providing insightful feedback. This work was supported by the Agency for Defense Development (ADD) grant funded by the Korea government (912855101). + +# References + +[1] Vijay Badrinarayan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 39(12):2481-2495, 2017. 1 +[2] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 3 +[3] John Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), (6):679-698, 1986. 2 +[4] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2229-2238, 2019. 3 +[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image + +segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 40(4):834-848, 2017. 1 +[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1 +[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 801-818, 2018. 1, 6 +[8] Wuyang Chen, Zhiding Yu, Shalini De Mello, Sifei Liu, Jose M Alvarez, Zhangyang Wang, and Anima Anandkumar. Contrastive syn-to-real generalization. arXiv preprint arXiv:2104.02290, 2021. 3, 5 +[9] Wuyang Chen, Zhiding Yu, Zhangyang Wang, and Animashree Anandkumar. Automated synthetic-to-real generalization. In Proceedings of the International Conference on Machine Learning (ICML), pages 1746-1756, 2020. 3 +[10] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11580-11590, 2021. 1, 2, 3, 5, 6 +[11] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606-3613, 2014. 8 +[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3213-3223, 2016. 1, 2, 6 +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248-255, 2009. 3, 6, 8 +[14] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 3 +[15] Mark Everingham, SM Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision (IJCV), 111(1):98-136, 2015. 6 +[16] Leon Gatys, Alexander S Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. Advances in Neural Information Processing Systems (NeurIPS), 28, 2015. 4 + +[17] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 4 +[18] Yunhao Ge, Yao Xiao, Zhi Xu, Xingrui Wang, and Laurent Itti. Contributions of Shape, Texture, and Color in Visual Recognition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 369-386, 2022. 2 +[19] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 1, 5 +[20] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 6 +[21] Robert M Haralick. Statistical and structural approaches to texture. Proceedings of the IEEE, 67(5):786-804, 1979. 4 +[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2961-2969, 2017. 1 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6 +[24] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6891-6902, 2021. 1, 2, 3, 6 +[25] Md Amirul Islam, Matthew Kowal, Patrick Esser, Sen Jia, Bjorn Ommer, Konstantinos G Derpanis, and Neil Bruce. Shape or texture: Understanding discriminative features in cnns. arXiv preprint arXiv:2101.11604, 2021. 5, 7 +[26] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and Statistical Texture Knowledge Distillation for Semantic Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16876-16885, 2022. 2 +[27] Yunpei Jia, Jie Zhang, Shiguang Shan, and Xilin Chen. Single-side domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8484-8493, 2020. 3 +[28] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. WildNet: Learning Domain Generalized Semantic Segmentation from the Wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9936-9946, 2022. 1, 2, 3, 6 +[29] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition (CVPR), pages 224-233, 2021. 3 +[30] Yijun Li, Ming-Yu Liu, Xueting Li, Ming-Hsuan Yang, and Jan Kautz. A closed-form solution to photorealistic image stylization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 453-468, 2018. 3 +[31] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 624-639, 2018. 3 +[32] Yanghao Li, Naiyan Wang, Jiaying Liu, and Xiaodi Hou. Demystifying neural style transfer. arXiv preprint arXiv:1701.01036, 2017. 4 +[33] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 40(12):2935-2947, 2017. 3 +[34] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3431-3440, 2015. 1 +[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[36] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain generalization via invariant feature representation. In Proceedings of the International Conference on Machine Learning (ICML), pages 10-18, 2013. 3 +[37] Gilhyun Nam, Gyeongjae Choi, and Kyungmin Lee. GCISG: Guided Causal Invariant Learning for Improved Syn-to-Real Generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 656–672, 2022. 3 +[38] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8690-8699, 2021. 2, 3, 4 +[39] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. Advances in Neural Information Processing Systems (NeurIPS), 34:23296-23308, 2021. 1, 5 +[40] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4990-4999, 2017. 2, 6 +[41] Kiri Nichol. Painter by numbers, wikiart. Kiri Nichol, 2016. 8 +[42] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 1, 2, 3 +[43] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1863-1871, 2019. 1, 2, 3 + +[44] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2594-2605, 2022. 1, 2, 3, 6 +[45] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2, 3, 6 +[46] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12556-12565, 2020. 3 +[47] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Proceedings of the European conference on computer vision (ECCV), pages 102-118, 2016. 1, 2, 6 +[48] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3234-3243, 2016. 6 +[49] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 618-626, 2017. 7 +[50] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10023-10031, 2019. 3 +[51] Nathan Somavarapu, Chih-Yao Ma, and Zsolt Kira. Frustratingly simple domain generalization via image stylization. arXiv preprint arXiv:2006.11207, 2020. 2, 3, 4 +[52] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research (JMLR), 9(11), 2008. 2 +[53] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 3 +[54] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. SiamDoGe: Domain Generalizable Semantic Segmentation Using Siamese Network. In Proceedings of the European Conference on Computer Vision (ECCV), pages 603-620, 2022. 1, 2, 3, 6 +[55] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems (NeurIPS), 34:12077-12090, 2021. 1 +[56] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generaliza + +tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14383-14392, 2021. 3 +[57] Zheng Xu, Wen Li, Li Niu, and Dong Xu. Exploiting low-rank structure from latent domains for domain generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 628-643, 2014. 3 +[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2636-2645, 2020. 2, 6 +[59] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2100-2110, 2019. 1, 2, 3, 6 +[60] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 535-552, 2022. 1, 2, 3, 6 +[61] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. arXiv preprint arXiv:2207.04892, 2022. 2, 3 +[62] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. arXiv preprint arXiv:2104.02008, 2021. 3 +[63] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12537-12546, 2021. 2 \ No newline at end of file diff --git a/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/images.zip b/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3dbbd64ad16eaa0958d07b358322034fa0dbfbe3 --- /dev/null +++ b/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07f65f0ab4afed1e055ecc7aba394c391d5e9e45d727d49f1cc7f7739ea93d8f +size 556621 diff --git a/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/layout.json b/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0cd28ec6240d84ab278de2c6855b5f37062f2a7b --- /dev/null +++ b/2023/Texture Learning Domain Randomization for Domain Generalized Segmentation/layout.json @@ -0,0 +1,10967 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 47, + 136, + 530, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 136, + 530, + 153 + ], + "spans": [ + { + "bbox": [ + 47, + 136, + 530, + 153 + ], + "type": "text", + "content": "Texture Learning Domain Randomization for Domain Generalized Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 156, + 181, + 232, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 181, + 232, + 195 + ], + "spans": [ + { + "bbox": [ + 156, + 181, + 232, + 195 + ], + "type": "text", + "content": "Sunghwan Kim" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 255, + 182, + 329, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 182, + 329, + 195 + ], + "spans": [ + { + "bbox": [ + 255, + 182, + 329, + 195 + ], + "type": "text", + "content": "Dae-hwan Kim" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 353, + 182, + 423, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 182, + 423, + 196 + ], + "spans": [ + { + "bbox": [ + 353, + 182, + 423, + 196 + ], + "type": "text", + "content": "Hoseong Kim*" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 190, + 198, + 386, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 198, + 386, + 212 + ], + "spans": [ + { + "bbox": [ + 190, + 198, + 386, + 212 + ], + "type": "text", + "content": "Agency for Defense Development (ADD)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 186, + 216, + 390, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 216, + 390, + 228 + ], + "spans": [ + { + "bbox": [ + 186, + 216, + 390, + 228 + ], + "type": "text", + "content": "{sssshwan，dhkim7，hoseongkim}@add.re.kr" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 139, + 255, + 186, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 255, + 186, + 267 + ], + "spans": [ + { + "bbox": [ + 139, + 255, + 186, + 267 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 281, + 282, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 281, + 282, + 573 + ], + "spans": [ + { + "bbox": [ + 45, + 281, + 282, + 573 + ], + "type": "text", + "content": "Deep Neural Networks (DNNs)-based semantic segmentation models trained on a source domain often struggle to generalize to unseen target domains, i.e., a domain gap problem. Texture often contributes to the domain gap, making DNNs vulnerable to domain shift because they are prone to be texture-biased. Existing Domain Generalized Semantic Segmentation (DGSS) methods have alleviated the domain gap problem by guiding models to prioritize shape over texture. On the other hand, shape and texture are two prominent and complementary cues in semantic segmentation. This paper argues that leveraging texture is crucial for improving performance in DGSS. Specifically, we propose a novel framework, coined Texture Learning Domain Randomization (TLDR). TLDR includes two novel losses to effectively enhance texture learning in DGSS: (1) a texture regularization loss to prevent overfitting to source domain textures by using texture features from an ImageNet pretrained model and (2) a texture generalization loss that utilizes random style images to learn diverse texture representations in a self-supervised manner. Extensive experimental results demonstrate the superiority of the proposed TLDR; e.g., TLDR achieves 46.5 mIoU on GTA " + }, + { + "bbox": [ + 45, + 281, + 282, + 573 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 45, + 281, + 282, + 573 + ], + "type": "text", + "content": " Cityscapes using ResNet-50, which improves the prior state-of-the-art method by 1.9 mIoU. The source code is available at https://github.com/ssssshwan/TLDR." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 596, + 124, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 596, + 124, + 607 + ], + "spans": [ + { + "bbox": [ + 46, + 596, + 124, + 607 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 616, + 279, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 616, + 279, + 710 + ], + "spans": [ + { + "bbox": [ + 45, + 616, + 279, + 710 + ], + "type": "text", + "content": "Semantic segmentation is an essential task in computer vision with many real-world applications, such as autonomous vehicles, augmented reality, and medical imaging. Deep Neural Networks (DNNs)-based semantic segmentation models work well when the data distributions are consistent between a source domain and target domains [1,5-7,22,34,55]. However, the performance of the models tends to degrade significantly in practical settings that in" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 302, + 255, + 414, + 321 + ], + "blocks": [ + { + "bbox": [ + 302, + 255, + 414, + 321 + ], + "lines": [ + { + "bbox": [ + 302, + 255, + 414, + 321 + ], + "spans": [ + { + "bbox": [ + 302, + 255, + 414, + 321 + ], + "type": "image", + "image_path": "ae0071673eeea8563b05c01168db66a957eddeeb2b50ce19075e12e7dc3ad200.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 323, + 393, + 332 + ], + "lines": [ + { + "bbox": [ + 326, + 323, + 393, + 332 + ], + "spans": [ + { + "bbox": [ + 326, + 323, + 393, + 332 + ], + "type": "text", + "content": "Unseen Target Image" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 415, + 255, + 528, + 321 + ], + "blocks": [ + { + "bbox": [ + 415, + 255, + 528, + 321 + ], + "lines": [ + { + "bbox": [ + 415, + 255, + 528, + 321 + ], + "spans": [ + { + "bbox": [ + 415, + 255, + 528, + 321 + ], + "type": "image", + "image_path": "f8a852b6b9d1742bca238f8ef771078e2b33d29d99d4d88cb3957a50d26220fd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 323, + 521, + 332 + ], + "lines": [ + { + "bbox": [ + 421, + 323, + 521, + 332 + ], + "spans": [ + { + "bbox": [ + 421, + 323, + 521, + 332 + ], + "type": "text", + "content": "(a) Normalization & Whitening" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 302, + 334, + 414, + 400 + ], + "blocks": [ + { + "bbox": [ + 302, + 334, + 414, + 400 + ], + "lines": [ + { + "bbox": [ + 302, + 334, + 414, + 400 + ], + "spans": [ + { + "bbox": [ + 302, + 334, + 414, + 400 + ], + "type": "image", + "image_path": "81d2ee6b752b10cd2c5a91a9f446266da65d16f984f80ca6c14149233d2d2d84.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 401, + 402, + 410 + ], + "lines": [ + { + "bbox": [ + 314, + 401, + 402, + 410 + ], + "spans": [ + { + "bbox": [ + 314, + 401, + 402, + 410 + ], + "type": "text", + "content": "(b) Domain Randomization" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 415, + 334, + 527, + 399 + ], + "blocks": [ + { + "bbox": [ + 415, + 334, + 527, + 399 + ], + "lines": [ + { + "bbox": [ + 415, + 334, + 527, + 399 + ], + "spans": [ + { + "bbox": [ + 415, + 334, + 527, + 399 + ], + "type": "image", + "image_path": "9f789ebab88827300ca44af2cc259dd87268ab796d6d1269d1c1f8606b7fa095.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 401, + 498, + 411 + ], + "lines": [ + { + "bbox": [ + 445, + 401, + 498, + 411 + ], + "spans": [ + { + "bbox": [ + 445, + 401, + 498, + 411 + ], + "type": "text", + "content": "(c) TLDR (ours)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 421, + 530, + 507 + ], + "lines": [ + { + "bbox": [ + 296, + 421, + 530, + 507 + ], + "spans": [ + { + "bbox": [ + 296, + 421, + 530, + 507 + ], + "type": "text", + "content": "Figure 1. Segmentation results for an image from an unseen domain (i.e., Cityscapes [12]), using models trained on GTA [47] with DGSS methods. (a-b) Existing methods [10, 59] have difficulty in distinguishing between road, sidewalk, and terrain which have similar shapes and contexts, as the texture is not sufficiently considered during the training process. (c) Our Texture Learning Domain Randomization (TLDR) can distinguish the classes effectively as we utilize the texture as prediction cues." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 530, + 530, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 530, + 530, + 635 + ], + "spans": [ + { + "bbox": [ + 296, + 530, + 530, + 635 + ], + "type": "text", + "content": "volve unseen out-of-distribution scenarios, also known as a domain gap problem. Many domain adaptation and generalization methods have been proposed to solve the domain gap problem [10, 24, 28, 42-44, 54, 59, 60]. Domain adaptation assumes accessibility of target domain images, differing from domain generalization. This paper addresses Domain Generalized Semantic Segmentation (DGSS), which aims to train models that can generalize to diverse unseen domains by training on a single source domain." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 636, + 531, + 730 + ], + "type": "text", + "content": "Existing DGSS methods have attempted to address the domain gap problem by guiding models to focus on shape rather than texture. Given that texture often varies across different domains (e.g., synthetic/real and sunny/rainy/foggy), DNNs are susceptible to domain shift because they tend to be texture-biased [19, 39]. Accordingly, there are two main approaches for the DGSS methods. The first approach is Normalization and Whitening" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 42, + 119, + 71 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 42, + 119, + 71 + ], + "spans": [ + { + "bbox": [ + 77, + 42, + 119, + 71 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 151, + 41, + 470, + 74 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 41, + 470, + 74 + ], + "spans": [ + { + "bbox": [ + 151, + 41, + 470, + 74 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 719, + 133, + 730 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 719, + 133, + 730 + ], + "spans": [ + { + "bbox": [ + 57, + 719, + 133, + 730 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "677" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 113, + 124, + 154 + ], + "blocks": [ + { + "bbox": [ + 66, + 105, + 113, + 113 + ], + "lines": [ + { + "bbox": [ + 66, + 105, + 113, + 113 + ], + "spans": [ + { + "bbox": [ + 66, + 105, + 113, + 113 + ], + "type": "text", + "content": "Original Source" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 113, + 124, + 154 + ], + "lines": [ + { + "bbox": [ + 54, + 113, + 124, + 154 + ], + "spans": [ + { + "bbox": [ + 54, + 113, + 124, + 154 + ], + "type": "image", + "image_path": "e7da4b5f40d6d796a760fb8112d63d57bb22a188473c1008c2dd804920c2ba03.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 163, + 280, + 196 + ], + "lines": [ + { + "bbox": [ + 45, + 163, + 280, + 196 + ], + "spans": [ + { + "bbox": [ + 45, + 163, + 280, + 196 + ], + "type": "text", + "content": "Figure 2. Reconstructed source images from the feature maps of (a) normalization and whitening and (b) domain randomization. Texture features are often omitted in existing DGSS methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 127, + 113, + 198, + 154 + ], + "blocks": [ + { + "bbox": [ + 131, + 105, + 190, + 113 + ], + "lines": [ + { + "bbox": [ + 131, + 105, + 190, + 113 + ], + "spans": [ + { + "bbox": [ + 131, + 105, + 190, + 113 + ], + "type": "text", + "content": "(a) Norm. & Whiten." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 127, + 113, + 198, + 154 + ], + "lines": [ + { + "bbox": [ + 127, + 113, + 198, + 154 + ], + "spans": [ + { + "bbox": [ + 127, + 113, + 198, + 154 + ], + "type": "image", + "image_path": "3263f88a5d78f936c88be60963d0c6ba7807a1d643a4629ef9bf0c3b0bbc73f2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 199, + 113, + 270, + 154 + ], + "blocks": [ + { + "bbox": [ + 202, + 105, + 264, + 113 + ], + "lines": [ + { + "bbox": [ + 202, + 105, + 264, + 113 + ], + "spans": [ + { + "bbox": [ + 202, + 105, + 264, + 113 + ], + "type": "text", + "content": "(b) Domain Random." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 199, + 113, + 270, + 154 + ], + "lines": [ + { + "bbox": [ + 199, + 113, + 270, + 154 + ], + "spans": [ + { + "bbox": [ + 199, + 113, + 270, + 154 + ], + "type": "image", + "image_path": "8d525b63b1d1b0b786906e033988aee3a74fc12f8eca834a9a08ef14bdbdc140.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 67, + 216, + 155, + 266 + ], + "blocks": [ + { + "bbox": [ + 81, + 207, + 138, + 216 + ], + "lines": [ + { + "bbox": [ + 81, + 207, + 138, + 216 + ], + "spans": [ + { + "bbox": [ + 81, + 207, + 138, + 216 + ], + "type": "text", + "content": "(a) Shape Features" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 216, + 155, + 266 + ], + "lines": [ + { + "bbox": [ + 67, + 216, + 155, + 266 + ], + "spans": [ + { + "bbox": [ + 67, + 216, + 155, + 266 + ], + "type": "image", + "image_path": "574fefa65ff72ecd66d6e237388c56e580afdf452b7bfd87cd366d6b02abaf56.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 275, + 279, + 330 + ], + "lines": [ + { + "bbox": [ + 45, + 275, + 279, + 330 + ], + "spans": [ + { + "bbox": [ + 45, + 275, + 279, + 330 + ], + "type": "text", + "content": "Figure 3. The t-SNE [52] plots for the road, sidewalk, and terrain classes from Cityscapes [12] that have similar shapes. While the shape features (Canny edge [3]) are entangled in (a), the texture features (Gram-matrix) of these classes are clearly separated in (b). The plots are based on an ImageNet pre-trained model." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 171, + 216, + 259, + 266 + ], + "blocks": [ + { + "bbox": [ + 184, + 207, + 243, + 216 + ], + "lines": [ + { + "bbox": [ + 184, + 207, + 243, + 216 + ], + "spans": [ + { + "bbox": [ + 184, + 207, + 243, + 216 + ], + "type": "text", + "content": "(b) Texture Features" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 216, + 259, + 266 + ], + "lines": [ + { + "bbox": [ + 171, + 216, + 259, + 266 + ], + "spans": [ + { + "bbox": [ + 171, + 216, + 259, + 266 + ], + "type": "image", + "image_path": "777bc6c73a6047f506102ae422d0deacf5ab2b81255b3aabbaa0349536a68580.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 343, + 279, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 343, + 279, + 447 + ], + "spans": [ + { + "bbox": [ + 45, + 343, + 279, + 447 + ], + "type": "text", + "content": "(NW), which involves normalizing and whitening the features [10, 42-44]. It is possible to remove domain-specific texture features and learn domain-invariant shape features with NW (see Figure 2a). The second approach is Domain Randomization (DR), which trains by transforming source images into randomly stylized images [24,28,45,54,59-61]. The model learns domain-invariant shape features because texture cues are mostly replaced by random styles (see Figure 2b) [38, 51]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 45, + 449, + 279, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 449, + 279, + 542 + ], + "spans": [ + { + "bbox": [ + 45, + 449, + 279, + 542 + ], + "type": "text", + "content": "While the existing methods are effective at making the models focus on shape features, they need to give more consideration to texture features. In addition to utilizing shape features like edges and structures, DNNs also use texture features such as patterns, color variations, and histograms as important cues for prediction [18]. Particularly in semantic segmentation, texture plays a crucial role in accurately maintaining the boundaries between objects [26, 63]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 45, + 543, + 279, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 543, + 279, + 693 + ], + "spans": [ + { + "bbox": [ + 45, + 543, + 279, + 693 + ], + "type": "text", + "content": "Figure 1 demonstrates the results of predicting an unseen image from Cityscapes [12] using DGSS methods trained on GTA [47]. One can see that the models trained with NW and DR have difficulty distinguishing between the road, sidewalk, and terrain classes, which have similar shapes and contexts. In order to determine these classes accurately, it is necessary to use texture cues. This assertion is further emphasized through t-SNE [52] plots of the classes: the shape features are entangled in Figure 3a, whereas the texture features are clearly separated in Figure 3b. Meanwhile, some textures remain relatively unchanged across domains in DGSS as shown in Figure 4. Based on these observations, we suggest utilizing texture as valuable cues for prediction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": "We propose Texture Learning Domain Randomization (TLDR), which enables DGSS models to learn both shape and texture. Since only source domain images are available" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 306, + 105, + 522, + 271 + ], + "blocks": [ + { + "bbox": [ + 306, + 105, + 522, + 271 + ], + "lines": [ + { + "bbox": [ + 306, + 105, + 522, + 271 + ], + "spans": [ + { + "bbox": [ + 306, + 105, + 522, + 271 + ], + "type": "image", + "image_path": "f130c7df74f6a7f53667abe8be917e6f04a64181f981afe366380cf53dcda358.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 279, + 530, + 321 + ], + "lines": [ + { + "bbox": [ + 296, + 279, + 530, + 321 + ], + "spans": [ + { + "bbox": [ + 296, + 279, + 530, + 321 + ], + "type": "text", + "content": "Figure 4. Visualization of texture for the road, sidewalk, and terrain classes from GTA [47], Cityscapes [12], BDD [58], and Mapillary [40] datasets. For each class, there are commonalities in texture across the datasets." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 341, + 530, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 341, + 530, + 574 + ], + "spans": [ + { + "bbox": [ + 296, + 341, + 530, + 574 + ], + "type": "text", + "content": "in DGSS, texture must be learned from them. To accurately capture the texture, we leverage the source domain images without any modification, which we refer to as the original source images. The stylized source images from DR are more focused on learning shape features, and the original source images are more focused on learning texture features (Section 4.1). To further improve texture learning in DGSS, we propose a texture regularization loss and a texture generalization loss. While there are commonalities in texture across different domains, there are also clear texture differences. Thus, if the model overfits source domain textures, it will result in a performance drop. To mitigate this problem, we propose the texture regularization loss to prevent overfitting to source domain textures by using texture features from an ImageNet pre-trained model (Section 4.2). Since only source domain textures alone may not be sufficient for learning general texture representations, we propose the texture generalization loss that utilizes random style images to learn diverse texture representations in a self-supervised manner (Section 4.3)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 578, + 530, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 578, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 296, + 578, + 530, + 729 + ], + "type": "text", + "content": "Our contribution can be summarized into three aspects. First, to the best of our knowledge, we are approaching DGSS for the first time from both the shape and texture perspectives. We argue that leveraging texture is essential in distinguishing between classes with similar shapes despite the domain gap. Second, to enhance texture learning in DGSS, we introduce two novel losses: the texture regularization loss and the texture generalization loss. Third, extensive experiments over multiple DGSS tasks show that our proposed TLDR achieves state-of-the-art performance. Our method attains " + }, + { + "bbox": [ + 296, + 578, + 530, + 729 + ], + "type": "inline_equation", + "content": "46.5\\mathrm{mIoU}" + }, + { + "bbox": [ + 296, + 578, + 530, + 729 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 296, + 578, + 530, + 729 + ], + "type": "inline_equation", + "content": "\\mathrm{GTA} \\rightarrow" + }, + { + "bbox": [ + 296, + 578, + 530, + 729 + ], + "type": "text", + "content": " Cityscapes using ResNet-50, surpassing the prior state-of-the-art method by " + }, + { + "bbox": [ + 296, + 578, + 530, + 729 + ], + "type": "inline_equation", + "content": "1.9\\mathrm{mIoU}" + }, + { + "bbox": [ + 296, + 578, + 530, + 729 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "678" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 109, + 321, + 186 + ], + "blocks": [ + { + "bbox": [ + 48, + 109, + 321, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 109, + 321, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 109, + 321, + 186 + ], + "type": "image", + "image_path": "37e9b48855740e1681c51867b98a94f7d1b42bad75a75d268823c0e5bc0be709.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 190, + 242, + 200 + ], + "lines": [ + { + "bbox": [ + 130, + 190, + 242, + 200 + ], + "spans": [ + { + "bbox": [ + 130, + 190, + 242, + 200 + ], + "type": "text", + "content": "(a) Style Transfer Module (STM)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 340, + 106, + 526, + 171 + ], + "blocks": [ + { + "bbox": [ + 340, + 106, + 526, + 171 + ], + "lines": [ + { + "bbox": [ + 340, + 106, + 526, + 171 + ], + "spans": [ + { + "bbox": [ + 340, + 106, + 526, + 171 + ], + "type": "image", + "image_path": "c2332a36b4f7268c22715b33c201c0891d41b72ce84065419b1077d1f2ed42cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 188, + 497, + 199 + ], + "lines": [ + { + "bbox": [ + 366, + 188, + 497, + 199 + ], + "spans": [ + { + "bbox": [ + 366, + 188, + 497, + 199 + ], + "type": "text", + "content": "(b) Texture Extraction Operator (TEO)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 212, + 531, + 245 + ], + "lines": [ + { + "bbox": [ + 45, + 212, + 531, + 245 + ], + "spans": [ + { + "bbox": [ + 45, + 212, + 531, + 245 + ], + "type": "text", + "content": "Figure 5. (a) Visualization of using the Style Transfer Module (STM) to transform an original source image into stylized source images with random style images. (b) Illustration of the process of using Texture Extraction Operator (TEO) to extract only texture features, i.e., a Gram-matrix, from a feature map." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 263, + 130, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 263, + 130, + 275 + ], + "spans": [ + { + "bbox": [ + 45, + 263, + 130, + 275 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 283, + 279, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 283, + 279, + 366 + ], + "spans": [ + { + "bbox": [ + 45, + 283, + 279, + 366 + ], + "type": "text", + "content": "Domain Generalization (DG). DG aims to learn DNNs that perform well on multiple unseen domains [36]. DG has primarily been studied in image classification. A number of studies have been proposed to address DG, including adversarial learning [27, 31, 50], data augmentation [46, 53, 56, 62], meta-learning [2, 14, 33], ensemble learning [57, 62], and self-supervised learning [4, 29]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 367, + 279, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 367, + 279, + 506 + ], + "spans": [ + { + "bbox": [ + 45, + 367, + 279, + 506 + ], + "type": "text", + "content": "Recent studies have attempted to train domain generalized models by preserving ImageNet pre-trained feature representations as much as possible. Chen et al. defined DG as a life-long learning problem [33] and tried to utilize ImageNet pre-trained weights to prevent catastrophic forgetting [9]. Contrastive learning [8] and attentional pooling [8, 37] were introduced to enhance the capturing of semantic knowledge in ImageNet features. In this paper, we regularize texture representations of DNNs with the ImageNet features. To the best of our knowledge, this is the first attempt to extract a specific semantic concept (i.e. texture) from the ImageNet features for regularization in DG." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 507, + 279, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 507, + 279, + 693 + ], + "spans": [ + { + "bbox": [ + 45, + 507, + 279, + 693 + ], + "type": "text", + "content": "Domain Generalized Semantic Segmentation (DGSS). DGSS is in its early stages and has yet to receive much research attention. Existing DGSS methods have tried to alleviate a domain gap problem through two main approaches: Normalization and Whitening (NW) and Domain Randomization (DR). NW trains by normalizing the mean and standard deviation of source features and whitening the covariance of source features. This process eliminates domain-specific features, allowing the model to learn domain-invariant features. Pan et al. introduced instance normalization to remove domain-specific features [42]. Pan et al. proposed switchable whitening to decorrelate features [43], and Choi et al. proposed Instance Selective Whitening (ISW) to enhance the ability to whiten domain-specific features [10]. Peng et al. tried to normalize and whiten features in a category-wise manner [44]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": "DR trains by transforming source images into randomly stylized images. The model is guided to capture domain-invariant shape features since texture cues are substituted" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 264, + 530, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 264, + 530, + 381 + ], + "spans": [ + { + "bbox": [ + 296, + 264, + 530, + 381 + ], + "type": "text", + "content": "with random styles. Yue et al. presented a method for considering consistency among multiple stylized images [59]. Peng et al. distinguished between global and local texture in the randomization process [45]. Huang et al. proposed DR in frequency space [24]. Adversarial learning [61] and self-supervised learning [54] have been used as attempts to make style learnable rather than using random style images. It has been shown that utilizing content and style of ImageNet [28] and ImageNet pre-trained features [60] can aid in learning generalized features in DR." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 387, + 530, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 387, + 530, + 527 + ], + "spans": [ + { + "bbox": [ + 296, + 387, + 530, + 527 + ], + "type": "text", + "content": "Existing DR methods have yet to comprehensively address DGSS from both shape and texture perspectives. Several studies [28, 59] have stated that the success of DR is attributed to its ability to learn various styles, resulting in improved performance in generalized domains. However, from a different perspective, we consider that the effectiveness of DR is due to the model becoming more focused on shape, as discussed in recent DG methods [38, 51]. Therefore, we assume that the model needs to learn texture for further performance improvement in DR. We propose a novel approach for learning texture in DGSS without overfitting to source domain textures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 297, + 554, + 379, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 554, + 379, + 565 + ], + "spans": [ + { + "bbox": [ + 297, + 554, + 379, + 565 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 578, + 518, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 578, + 518, + 591 + ], + "spans": [ + { + "bbox": [ + 297, + 578, + 518, + 591 + ], + "type": "text", + "content": "3.1. Domain Randomization with Style Transfer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 601, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 601, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 601, + 530, + 730 + ], + "type": "text", + "content": "We adopt DR as a baseline DGSS method. We use a neural style transfer method at each epoch to transform each original source image into a different random style. If edge information is lost during the style transfer process, it may cause a mismatch with the semantic label, leading to a decrease in performance. We utilize photoWCT [30] known as an edge-preserving style transfer method. Random style images are sampled from ImageNet [13] validation set. Figure 5a is a visualization of using the Style Transfer Module (STM) to transform an original source image into stylized source images with random style images." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "679" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 104, + 530, + 295 + ], + "blocks": [ + { + "bbox": [ + 47, + 104, + 530, + 295 + ], + "lines": [ + { + "bbox": [ + 47, + 104, + 530, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 104, + 530, + 295 + ], + "type": "image", + "image_path": "d7777b91dc5dc3f15613d57574226280baccf338af342435c6095641e0828cb3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "lines": [ + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "spans": [ + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": "Figure 6. An overview of our proposed Texture Learning Domain Randomization (TLDR) framework. The stylized source image " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "x^{sr}" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": " (green) is obtained by stylizing the original source image " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "x^s" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": " (blue) with the random style image " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "x^r" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": " (purple). The stylized task loss " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{styl}}" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": " focuses on learning shape from " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "x^{sr}" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": ", and the original task loss " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{orig}}" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": " focuses on learning texture from " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "x^s" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": ". The texture regularization loss " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TR}}" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": " enforces the consistency between the Gram-matrices of the ImageNet model " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "f_I" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": " and the task model " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "f_T" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "x^s" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": ". The texture generalization loss " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TG}}" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": " enforces the consistency between the Gram-matrices of the task model " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "f_T" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "x^{sr}" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "inline_equation", + "content": "x^r" + }, + { + "bbox": [ + 45, + 302, + 533, + 367 + ], + "type": "text", + "content": ". Random Style Masking (RSM) selects only the random style features when applying the texture generalization loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 387, + 242, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 387, + 242, + 398 + ], + "spans": [ + { + "bbox": [ + 45, + 387, + 242, + 398 + ], + "type": "text", + "content": "3.2. Texture Extraction with Gram-matrix" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 405, + 279, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 405, + 279, + 474 + ], + "spans": [ + { + "bbox": [ + 45, + 405, + 279, + 474 + ], + "type": "text", + "content": "Texture is a regional descriptor that can offer measurements for both local structural (e.g., pattern) and global statistical (e.g., overall distribution of colors) properties of an image [21]. It has been shown that texture can be represented by pairwise correlations between features extracted by DNNs, also known as a Gram-matrix [16, 17, 32]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 474, + 279, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 474, + 279, + 533 + ], + "spans": [ + { + "bbox": [ + 45, + 474, + 279, + 533 + ], + "type": "text", + "content": "We use the Gram-matrix to extract only texture features from a feature map. The Gram-matrix " + }, + { + "bbox": [ + 45, + 474, + 279, + 533 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{C \\times C}" + }, + { + "bbox": [ + 45, + 474, + 279, + 533 + ], + "type": "text", + "content": " for the vectorized matrix " + }, + { + "bbox": [ + 45, + 474, + 279, + 533 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{C \\times HW}" + }, + { + "bbox": [ + 45, + 474, + 279, + 533 + ], + "type": "text", + "content": " of the feature map is defined as Equation 1. " + }, + { + "bbox": [ + 45, + 474, + 279, + 533 + ], + "type": "inline_equation", + "content": "C, H" + }, + { + "bbox": [ + 45, + 474, + 279, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 474, + 279, + 533 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 45, + 474, + 279, + 533 + ], + "type": "text", + "content": " denote the channel, height, and width of the feature map, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 540, + 279, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 540, + 279, + 553 + ], + "spans": [ + { + "bbox": [ + 130, + 540, + 279, + 553 + ], + "type": "interline_equation", + "content": "G _ {i, j} = F _ {i} \\cdot F _ {j}, \\tag {1}", + "image_path": "b40202ab66901e9fb9f61555b2b4abb3bd9c264df7de43b9302819d3bc107b70.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "spans": [ + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "text", + "content": " represents a dot product, " + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "inline_equation", + "content": "F_{j}" + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "text", + "content": " row vectors of " + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "inline_equation", + "content": "G_{i,j}" + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "text", + "content": " is the entry at the " + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "text", + "content": " row and " + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "text", + "content": " column of " + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 45, + 559, + 281, + 643 + ], + "type": "text", + "content": ". Each entry of the Grammatrix indicates a pairwise correlation between the features corresponding to a texture feature. In this paper, the operator extracting texture features is called Texture Extraction Operator (TEO). Figure 5b illustrates the process of TEO." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 652, + 111, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 652, + 111, + 665 + ], + "spans": [ + { + "bbox": [ + 45, + 652, + 111, + 665 + ], + "type": "text", + "content": "4. Approach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 280, + 730 + ], + "type": "text", + "content": "This section describes our proposed Texture Learning Domain Randomization (TLDR) framework. TLDR learns texture features in addition to learning shape features through domain randomization. TLDR consists of four losses: a stylized task loss, an original task loss, a tex" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 296, + 387, + 530, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 387, + 530, + 494 + ], + "spans": [ + { + "bbox": [ + 296, + 387, + 530, + 494 + ], + "type": "text", + "content": "ture regularization loss, and a texture generalization loss. The stylized task loss focuses on learning shape, and the original task loss focuses on learning texture. The texture regularization loss and the texture generalization loss prevent overfitting to source domain textures. Figure 6 is an overview of TLDR. The frozen ImageNet pre-trained encoder is denoted as " + }, + { + "bbox": [ + 296, + 387, + 530, + 494 + ], + "type": "inline_equation", + "content": "f_{I}" + }, + { + "bbox": [ + 296, + 387, + 530, + 494 + ], + "type": "text", + "content": ", i.e., the ImageNet model. The training encoder is denoted as " + }, + { + "bbox": [ + 296, + 387, + 530, + 494 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 296, + 387, + 530, + 494 + ], + "type": "text", + "content": ", i.e., the task model. The training semantic segmentation decoder is denoted as " + }, + { + "bbox": [ + 296, + 387, + 530, + 494 + ], + "type": "inline_equation", + "content": "h_{T}" + }, + { + "bbox": [ + 296, + 387, + 530, + 494 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 500, + 374, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 500, + 374, + 512 + ], + "spans": [ + { + "bbox": [ + 296, + 500, + 374, + 512 + ], + "type": "text", + "content": "4.1. Task Losses" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "spans": [ + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "text", + "content": "Stylized task loss. We denote an original source image " + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "inline_equation", + "content": "x^{s}" + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "text", + "content": " and its semantic label " + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "inline_equation", + "content": "x^{sr}" + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "text", + "content": " is a stylized source image obtained by stylizing an original source image " + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "inline_equation", + "content": "x^{s}" + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "text", + "content": " with a random style image " + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "inline_equation", + "content": "x^{r}" + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "text", + "content": ". The prediction result for " + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "inline_equation", + "content": "x^{sr}" + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "text", + "content": " of the model is " + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "inline_equation", + "content": "p^{sr} = h_{T}(f_{T}(x^{sr}))" + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "text", + "content": ". Then the stylized task loss " + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{styl}}" + }, + { + "bbox": [ + 296, + 522, + 531, + 593 + ], + "type": "text", + "content": " is given by Equation 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 376, + 600, + 531, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 600, + 531, + 614 + ], + "spans": [ + { + "bbox": [ + 376, + 600, + 531, + 614 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s t y l}} = \\mathrm {C E} \\left(p ^ {s r}, y\\right), \\tag {2}", + "image_path": "b4084d499911922d484bcdfc867cc11a8b8c12a8c9ee38bc611bef7392f9729c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 622, + 530, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 622, + 530, + 669 + ], + "spans": [ + { + "bbox": [ + 296, + 622, + 530, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 622, + 530, + 669 + ], + "type": "inline_equation", + "content": "\\mathrm{CE}(\\cdot)" + }, + { + "bbox": [ + 296, + 622, + 530, + 669 + ], + "type": "text", + "content": " represents the categorical cross-entropy loss. The stylized task loss encourages the model to focus on shape features during training since the texture cues are mostly replaced by random styles [38, 51]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 671, + 531, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 671, + 531, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 671, + 531, + 730 + ], + "type": "text", + "content": "Original task loss. The model struggles to learn texture from the stylized source images as the texture cues are mostly substituted with random styles. To accurately capture the source domain textures, the model is trained on the original source images. The prediction result for " + }, + { + "bbox": [ + 296, + 671, + 531, + 730 + ], + "type": "inline_equation", + "content": "x^{s}" + }, + { + "bbox": [ + 296, + 671, + 531, + 730 + ], + "type": "text", + "content": " of the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "680" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "spans": [ + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "type": "text", + "content": "model is " + }, + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "type": "inline_equation", + "content": "p^s = h_T(f_T(x^s))" + }, + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "type": "text", + "content": ". The original task loss " + }, + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{orig}}" + }, + { + "bbox": [ + 45, + 106, + 279, + 129 + ], + "type": "text", + "content": " is given by Equation 3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 135, + 279, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 135, + 279, + 148 + ], + "spans": [ + { + "bbox": [ + 126, + 135, + 279, + 148 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o r i g}} = \\mathrm {C E} \\left(p ^ {s}, y\\right). \\tag {3}", + "image_path": "65d66987a481786f0a51b57b0d01f9724a7c1aa77a36b64a203cbdfe1dd963e2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 153, + 279, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 153, + 279, + 188 + ], + "spans": [ + { + "bbox": [ + 45, + 153, + 279, + 188 + ], + "type": "text", + "content": "DNNs tend to prioritize texture cues without restrictions [19, 39]. The original task loss guides the model to concentrate on texture features during training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 194, + 196, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 196, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 196, + 205 + ], + "type": "text", + "content": "4.2. Texture Regularization Loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 212, + 279, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 212, + 279, + 362 + ], + "spans": [ + { + "bbox": [ + 45, + 212, + 279, + 362 + ], + "type": "text", + "content": "If the model is trained on the original source images without regularization, it will overfit the source domain textures [19, 39]. We regularize texture representations of the task model using the ImageNet model, which encodes diverse feature representations [8]. However, it is important to note that ImageNet features include not only texture features, but also other semantic features such as shape features. We therefore assume that regularizing the entire feature may interfere with the task model learning texture. To address this issue, we propose to apply TEO to extract only texture features from ImageNet features and regularize the task model with the extracted texture features (see Table 5 for ablation)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "spans": [ + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "F_{I}^{l,s}" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "F_{T}^{l,s}" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": " denote the vectorized feature maps in layer " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": " of the ImageNet model " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "f_{I}" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": " and the task model " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "x^{s}" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": ", respectively. The Gram-matrices " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "G_{I}^{l,s}" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "G_{T}^{l,s}" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "F_{I}^{l,s}" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "F_{T}^{l,s}" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": " are the texture features of the original source image as seen by " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "f_{I}" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "f_{T}" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": ", respectively. The contribution of the " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": " layer to the texture regularization loss is " + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "inline_equation", + "content": "\\| G_{I}^{l,s} - G_{T}^{l,s}\\|_{2}" + }, + { + "bbox": [ + 45, + 362, + 279, + 457 + ], + "type": "text", + "content": ". The total texture regularization loss is given by Equation 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 461, + 279, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 461, + 279, + 493 + ], + "spans": [ + { + "bbox": [ + 89, + 461, + 279, + 493 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T R}} = \\sum_ {l = 1} ^ {L} \\frac {u _ {l}}{C _ {l} H _ {l} W _ {l}} \\| G _ {I} ^ {l, s} - G _ {T} ^ {l, s} \\| _ {2}, \\tag {4}", + "image_path": "3d5513d8836f18d0923d8adaabe87c038f58a5aefbc8e3f557dbf9587d1111d1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "spans": [ + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "text", + "content": " is the number of feature map layers, and " + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "inline_equation", + "content": "u_{l}" + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "text", + "content": " is a weighting factor for the contribution of the " + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "text", + "content": " layer to " + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TR}}" + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "inline_equation", + "content": "C_l, H_l," + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "inline_equation", + "content": "W_{l}" + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "text", + "content": " denote the channel, height and width of the " + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "text", + "content": " layer feature map, respectively. We reduce the value of " + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "inline_equation", + "content": "u_{l}" + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 45, + 495, + 279, + 566 + ], + "type": "text", + "content": " increases, considering that fewer texture features are encoded in feature maps as layers become deeper [25]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 572, + 196, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 572, + 196, + 583 + ], + "spans": [ + { + "bbox": [ + 45, + 572, + 196, + 583 + ], + "type": "text", + "content": "4.3. Texture Generalization Loss" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 590, + 279, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 590, + 279, + 693 + ], + "spans": [ + { + "bbox": [ + 45, + 590, + 279, + 693 + ], + "type": "text", + "content": "We supplement texture learning from the random style images for more diverse texture representations. Since the random style images are unlabeled, the texture representations should be learned self-supervised. Note that the random style image " + }, + { + "bbox": [ + 45, + 590, + 279, + 693 + ], + "type": "inline_equation", + "content": "x^{r}" + }, + { + "bbox": [ + 45, + 590, + 279, + 693 + ], + "type": "text", + "content": " and the stylized source image " + }, + { + "bbox": [ + 45, + 590, + 279, + 693 + ], + "type": "inline_equation", + "content": "x^{sr}" + }, + { + "bbox": [ + 45, + 590, + 279, + 693 + ], + "type": "text", + "content": " share some texture features. To encourage learning of diverse texture representations, we induce the texture features of " + }, + { + "bbox": [ + 45, + 590, + 279, + 693 + ], + "type": "inline_equation", + "content": "x^{sr}" + }, + { + "bbox": [ + 45, + 590, + 279, + 693 + ], + "type": "text", + "content": " to become more similar to those of " + }, + { + "bbox": [ + 45, + 590, + 279, + 693 + ], + "type": "inline_equation", + "content": "x^{r}" + }, + { + "bbox": [ + 45, + 590, + 279, + 693 + ], + "type": "text", + "content": " as much as possible while preserving source texture features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "inline_equation", + "content": "F_T^{l,r}" + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "inline_equation", + "content": "F_T^{l,sr}" + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": " denote the vectorized feature maps in layer " + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "inline_equation", + "content": "f_T" + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": " for the random style image " + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "inline_equation", + "content": "x^r" + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": " and the stylized source image " + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "inline_equation", + "content": "x^{sr}" + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "inline_equation", + "content": "G_T^{l,r}" + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "inline_equation", + "content": "G_T^{l,sr}" + }, + { + "bbox": [ + 45, + 694, + 279, + 730 + ], + "type": "text", + "content": " are the" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 299, + 105, + 529, + 191 + ], + "blocks": [ + { + "bbox": [ + 299, + 105, + 529, + 191 + ], + "lines": [ + { + "bbox": [ + 299, + 105, + 529, + 191 + ], + "spans": [ + { + "bbox": [ + 299, + 105, + 529, + 191 + ], + "type": "image", + "image_path": "33b977f9373ea0340b092317e794bc0edf6abe173e5f0166c1809898363a9ebe.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "lines": [ + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "spans": [ + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "type": "text", + "content": "Figure 7. Random Style Masking (RSM) masks only the entries corresponding to the random style features (purple) in the Grammatrices. The difference between " + }, + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "type": "inline_equation", + "content": "G_T^{l,s\\tau}" + }, + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "type": "inline_equation", + "content": "G_T^{l,s}" + }, + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "type": "text", + "content": " is calculated as " + }, + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "type": "inline_equation", + "content": "D^l" + }, + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "type": "text", + "content": ". The entries with values greater than a certain threshold " + }, + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "type": "inline_equation", + "content": "D^l" + }, + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "type": "text", + "content": " are masked as " + }, + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "type": "inline_equation", + "content": "M^l" + }, + { + "bbox": [ + 296, + 200, + 530, + 254 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "spans": [ + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "text", + "content": "corresponding Gram-matrices. Our goal is basically to set an objective that makes " + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "inline_equation", + "content": "G_T^{l,r}" + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "inline_equation", + "content": "G_T^{l,sr}" + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "text", + "content": " consistent. However, " + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "inline_equation", + "content": "G_T^{l, sr}" + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "text", + "content": " includes both random style features and remaining source texture features. Applying the constraints to the entire " + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "inline_equation", + "content": "G_T^{l, sr}" + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "text", + "content": " also imposes the objective on the source texture features in the stylized source image. To select only the random style features when enforcing the consistency between " + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "inline_equation", + "content": "G_T^{l,r}" + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "inline_equation", + "content": "G_T^{l, sr}" + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "text", + "content": ", we propose Random Style Masking (RSM), inspired by ISW [10]. Figure 7 is an illustration of RSM. We assume that the entries corresponding to the random style features are activated in " + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "inline_equation", + "content": "G_T^{l, sr}" + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "text", + "content": " but deactivated in " + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "inline_equation", + "content": "G_T^{l,s}" + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "text", + "content": ". Considering " + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "inline_equation", + "content": "D^l = G_T^{l, sr} - G_T^{l,s}" + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "text", + "content": ", the entries corresponding to the random style features are expected to be larger than a certain threshold " + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "text", + "content": ". We denote the mask for the entries representing the random style features as " + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "inline_equation", + "content": "M^l" + }, + { + "bbox": [ + 296, + 263, + 530, + 442 + ], + "type": "text", + "content": " (see Equation 5)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 449, + 529, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 449, + 529, + 480 + ], + "spans": [ + { + "bbox": [ + 361, + 449, + 529, + 480 + ], + "type": "interline_equation", + "content": "M _ {i, j} ^ {l} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} D _ {i, j} ^ {l} > \\tau \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}", + "image_path": "bc0a628e63718a00e09ac5f91447fbe3aa5a08592868fd1b96b8877eaf12e83f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "spans": [ + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "text", + "content": " are the row and column indices in each matrix, respectively. The threshold " + }, + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "text", + "content": " is determined empirically (see Table 6 for ablation). We only apply the objective to selected random style features by RSM. The contribution of the " + }, + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "text", + "content": " layer to the texture generalization loss is " + }, + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "inline_equation", + "content": "\\|(G_T^{l,r} - G_T^{l,sr})\\odot M^l\\|_2" + }, + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 296, + 482, + 530, + 575 + ], + "type": "text", + "content": " represents an element-wise product. The total texture generalization loss is given by Equation 6." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 323, + 578, + 530, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 578, + 530, + 609 + ], + "spans": [ + { + "bbox": [ + 323, + 578, + 530, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T G}} = \\sum_ {l = 1} ^ {L} \\frac {v _ {l}}{C _ {l} H _ {l} W _ {l}} \\| \\left(G _ {T} ^ {l, r} - G _ {T} ^ {l, s r}\\right) \\odot M ^ {l} \\| _ {2}, \\tag {6}", + "image_path": "94728b40a5bad25e586724ef3aa4e1e92474c67559a194d189ba76f25452f605.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "spans": [ + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "text", + "content": " is a weighting factor for the contribution of the " + }, + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "text", + "content": " layer to " + }, + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TG}}" + }, + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "text", + "content": ". We also reduce the value of " + }, + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "text", + "content": " increases for the same reason as " + }, + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "inline_equation", + "content": "u_{l}" + }, + { + "bbox": [ + 296, + 612, + 529, + 648 + ], + "type": "text", + "content": " in Equation 4." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 297, + 653, + 383, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 653, + 383, + 666 + ], + "spans": [ + { + "bbox": [ + 297, + 653, + 383, + 666 + ], + "type": "text", + "content": "4.4. Full Objective" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 671, + 530, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 671, + 530, + 729 + ], + "spans": [ + { + "bbox": [ + 296, + 671, + 530, + 729 + ], + "type": "text", + "content": "As the training progresses, the task losses decrease while the texture regularization loss and the texture generalization loss remain relatively constant. We add a Linear Decay Factor (LDF) to the texture regularization and generalization losses to balance their scale with the task losses" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 295, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 295, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 295, + 756 + ], + "type": "text", + "content": "681" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 104, + 274, + 281 + ], + "blocks": [ + { + "bbox": [ + 51, + 104, + 274, + 281 + ], + "lines": [ + { + "bbox": [ + 51, + 104, + 274, + 281 + ], + "spans": [ + { + "bbox": [ + 51, + 104, + 274, + 281 + ], + "type": "table", + "html": "
MethodEncoderCBMS
DRPC [59]37.4232.1434.12-
RobustNet [10]36.5835.2040.3328.30
SAN-SAW [44]39.7537.3441.8630.79
SiamDoGe [54]ResNet-5042.9637.5440.6428.34
WildNet [28]44.6238.4246.0931.34
SHADE [60]44.6539.2843.34-
TLDR (ours)46.5142.5846.1836.30
DRPC [59]42.5338.7238.0529.67
GTR [45]43.7039.6039.1029.32
FSDR [24]44.8041.2043.40-
SAN-SAW [44]ResNet-10145.3341.1840.7731.84
WildNet [28]45.7941.7347.0832.51
SHADE [60]46.6643.6645.50-
TLDR (ours)47.5844.8848.8039.35
", + "image_path": "91e175ab136db1080102171b13881ca60d47673ee1b5cb9afc021c57431ca827.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 349, + 279, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 349, + 279, + 384 + ], + "spans": [ + { + "bbox": [ + 45, + 349, + 279, + 384 + ], + "type": "text", + "content": "(see Table 6 for ablation). The LDF at iteration " + }, + { + "bbox": [ + 45, + 349, + 279, + 384 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 45, + 349, + 279, + 384 + ], + "type": "text", + "content": " is set " + }, + { + "bbox": [ + 45, + 349, + 279, + 384 + ], + "type": "inline_equation", + "content": "w(t) = (1 - t / t_{\\mathrm{total}})" + }, + { + "bbox": [ + 45, + 349, + 279, + 384 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 349, + 279, + 384 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{total}}" + }, + { + "bbox": [ + 45, + 349, + 279, + 384 + ], + "type": "text", + "content": " denotes the total number of iterations. Our full objective is given by Equation 7." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 399, + 279, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 399, + 279, + 422 + ], + "spans": [ + { + "bbox": [ + 52, + 399, + 279, + 422 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\alpha_ {\\text {o r i g}} \\mathcal {L} _ {\\text {o r i g}} + \\alpha_ {\\text {s t y l}} \\mathcal {L} _ {\\text {s t y l}} + w (t) \\mathcal {L} _ {\\text {T R}} + w (t) \\mathcal {L} _ {\\text {T G}}, \\tag {7}", + "image_path": "e02291aa610ffd29551d10fd9413cec3a75e2cc9a3b9d715371bcdbb60c30f66.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 422, + 279, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 422, + 279, + 446 + ], + "spans": [ + { + "bbox": [ + 45, + 422, + 279, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 45, + 422, + 279, + 446 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{orig}}" + }, + { + "bbox": [ + 45, + 422, + 279, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 422, + 279, + 446 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{styl}}" + }, + { + "bbox": [ + 45, + 422, + 279, + 446 + ], + "type": "text", + "content": " are the weights for the original task loss and the stylized task loss, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 455, + 124, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 455, + 124, + 468 + ], + "spans": [ + { + "bbox": [ + 45, + 455, + 124, + 468 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 474, + 175, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 474, + 175, + 487 + ], + "spans": [ + { + "bbox": [ + 45, + 474, + 175, + 487 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 492, + 279, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 492, + 279, + 562 + ], + "spans": [ + { + "bbox": [ + 45, + 492, + 279, + 562 + ], + "type": "text", + "content": "Datasets. As synthetic datasets, GTA [47] consists of 24,966 images with a resolution of " + }, + { + "bbox": [ + 45, + 492, + 279, + 562 + ], + "type": "inline_equation", + "content": "1914 \\times 1052" + }, + { + "bbox": [ + 45, + 492, + 279, + 562 + ], + "type": "text", + "content": ". It has 12,403, 6,382, and 6,181 images for training, validation, and test sets. SYNTHIA [48] contains 9,400 images with a resolution of " + }, + { + "bbox": [ + 45, + 492, + 279, + 562 + ], + "type": "inline_equation", + "content": "1280 \\times 760" + }, + { + "bbox": [ + 45, + 492, + 279, + 562 + ], + "type": "text", + "content": ". It has 6,580 and 2,820 images for training and validation sets, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 562, + 279, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 562, + 279, + 656 + ], + "spans": [ + { + "bbox": [ + 45, + 562, + 279, + 656 + ], + "type": "text", + "content": "As real-world datasets, Cityscapes [12] consists of 2,975 training images and 500 validation images with a resolution of " + }, + { + "bbox": [ + 45, + 562, + 279, + 656 + ], + "type": "inline_equation", + "content": "2048 \\times 1024" + }, + { + "bbox": [ + 45, + 562, + 279, + 656 + ], + "type": "text", + "content": ". BDD [58] contains 7,000 training images and 1,000 validation images with a resolution of " + }, + { + "bbox": [ + 45, + 562, + 279, + 656 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 45, + 562, + 279, + 656 + ], + "type": "text", + "content": ". Mapillary [40] involves 18,000 training images and 2,000 validation images with diverse resolutions. For brevity, we denote GTA, SYNTHIA, Cityscapes, BDD, and Mapillary as G, S, C, B, and M, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 657, + 279, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 657, + 279, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 657, + 279, + 715 + ], + "type": "text", + "content": "Network architecture. We conduct experiments using ResNet [23] as an encoder architecture and DeepLabV3+ [7] as a semantic segmentation decoder architecture. In all experiments, encoders are initialized with an ImageNet [13] pre-trained model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 717, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 717, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 717, + 279, + 730 + ], + "type": "text", + "content": "Training. We adopt an AdamW [35] optimizer. An initial" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 302, + 104, + 526, + 212 + ], + "blocks": [ + { + "bbox": [ + 45, + 290, + 279, + 334 + ], + "lines": [ + { + "bbox": [ + 45, + 290, + 279, + 334 + ], + "spans": [ + { + "bbox": [ + 45, + 290, + 279, + 334 + ], + "type": "text", + "content": "Table 1. Comparison of mIoU (%; higher is better) between DGSS methods trained on GTA and evaluated on C, B, M, and S. The best and second best results are highlighted and underlined, respectively. Our method is marked in gray." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 302, + 104, + 526, + 212 + ], + "lines": [ + { + "bbox": [ + 302, + 104, + 526, + 212 + ], + "spans": [ + { + "bbox": [ + 302, + 104, + 526, + 212 + ], + "type": "table", + "html": "
MethodEncoderCBMG
DRPC [59]35.6531.5332.7428.75
SAN-SAW [44]ResNet-5038.9235.2434.5229.16
TLDR (ours)41.8834.3536.7935.90
DRPC [59]37.5834.3434.1229.24
GTR [45]39.7035.3036.4028.71
FSDR [24]ResNet-10140.8039.6037.40-
SAN-SAW [44]40.8735.9837.2630.79
TLDR (ours)42.6035.4637.4637.77
", + "image_path": "03756d1aacf16ae3cf6f87b84f1e7880c4006d46fe6636650c25b277d0fc9a5c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 220, + 530, + 264 + ], + "lines": [ + { + "bbox": [ + 296, + 220, + 530, + 264 + ], + "spans": [ + { + "bbox": [ + 296, + 220, + 530, + 264 + ], + "type": "text", + "content": "Table 2. Comparison of mIoU (%; higher is better) between DGSS methods trained on SYNTHIA and evaluated on C, B, M, and G. The best and second best results are highlighted and underlined, respectively. Our method is marked in gray." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "spans": [ + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "text", + "content": "learning rate is set to " + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-5}" + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "text", + "content": " for the encoder and " + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-4}" + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "text", + "content": " for the decoder, " + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "inline_equation", + "content": "40k" + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "text", + "content": " training iterations, a batch size of 4. A weight decay is set to 0.01, with a linear warmup [20] over " + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{warm}} = 1k" + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "text", + "content": " iterations, followed by a linear decay. We use random scaling in the range [0.5, 2.0] and random cropping with a size of " + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "inline_equation", + "content": "768 \\times 768" + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "text", + "content": ". We apply additional data augmentation techniques, including random flipping and color jittering. We set the texture regularization parameters as " + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "inline_equation", + "content": "u_l = 5 \\times 10^{-l-2}" + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "text", + "content": ", and the texture generalization parameters as " + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "inline_equation", + "content": "v_l = 5 \\times 10^{-l-2}" + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "text", + "content": ". The original task loss and the stylized task loss weights are set to " + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{orig}} = 0.5" + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{styl}} = 0.5" + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "text", + "content": ", respectively. We set the RSM threshold " + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 296, + 280, + 530, + 421 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 428, + 471, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 428, + 471, + 442 + ], + "spans": [ + { + "bbox": [ + 296, + 428, + 471, + 442 + ], + "type": "text", + "content": "5.2. Comparison with DGSS methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 447, + 530, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 447, + 530, + 563 + ], + "spans": [ + { + "bbox": [ + 296, + 447, + 530, + 563 + ], + "type": "text", + "content": "To measure generalization capacity in unseen domains, we train on a single source domain and evaluate multiple unseen domains. We conduct experiments on two settings, (1) " + }, + { + "bbox": [ + 296, + 447, + 530, + 563 + ], + "type": "inline_equation", + "content": "\\mathrm{G} \\rightarrow \\{\\mathrm{C}, \\mathrm{B}, \\mathrm{M}, \\mathrm{S}\\}" + }, + { + "bbox": [ + 296, + 447, + 530, + 563 + ], + "type": "text", + "content": " and (2) " + }, + { + "bbox": [ + 296, + 447, + 530, + 563 + ], + "type": "inline_equation", + "content": "\\mathrm{S} \\rightarrow \\{\\mathrm{C}, \\mathrm{B}, \\mathrm{M}, \\mathrm{G}\\}" + }, + { + "bbox": [ + 296, + 447, + 530, + 563 + ], + "type": "text", + "content": ". We repeat each benchmark three times, each time with a different random seed, and report the average results. We evaluate our method using ResNet-50 and ResNet-101 encoders. We use mean Intersection over Union (mIoU) [15] as the evaluation metric. The best and second best results are highlighted and underlined in tables, respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 563, + 530, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 563, + 530, + 634 + ], + "spans": [ + { + "bbox": [ + 296, + 563, + 530, + 634 + ], + "type": "text", + "content": "Tables 1 and 2 show the generalization performance of models trained on GTA [47] and SYNTHIA [48], respectively. Our TLDR generally outperforms other DGSS methods in most benchmarks. In particular, we improve the " + }, + { + "bbox": [ + 296, + 563, + 530, + 634 + ], + "type": "inline_equation", + "content": "\\mathrm{G}\\rightarrow \\mathrm{C}" + }, + { + "bbox": [ + 296, + 563, + 530, + 634 + ], + "type": "text", + "content": " benchmark by " + }, + { + "bbox": [ + 296, + 563, + 530, + 634 + ], + "type": "inline_equation", + "content": "+1.9\\mathrm{mIoU}" + }, + { + "bbox": [ + 296, + 563, + 530, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 563, + 530, + 634 + ], + "type": "inline_equation", + "content": "+0.9\\mathrm{mIoU}" + }, + { + "bbox": [ + 296, + 563, + 530, + 634 + ], + "type": "text", + "content": " on ResNet-50 and ResNet-101 encoders, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 641, + 449, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 641, + 449, + 653 + ], + "spans": [ + { + "bbox": [ + 296, + 641, + 449, + 653 + ], + "type": "text", + "content": "5.3. Texture Awareness of Model" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 660, + 530, + 730 + ], + "type": "text", + "content": "We design experiments to verify whether the performance improvement of TLDR is actually due to being aware of texture. In the experiments, we compare TLDR to plain Domain Randomization (DR), which is only trained with the stylized task loss. We train on GTA using a ResNet-101 encoder. For convenience, we refer to the models trained" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "682" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 105, + 276, + 228 + ], + "blocks": [ + { + "bbox": [ + 48, + 105, + 276, + 228 + ], + "lines": [ + { + "bbox": [ + 48, + 105, + 276, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 105, + 276, + 228 + ], + "type": "image", + "image_path": "652eb5ac72abe96eea1bc816e7c474a770983d1c4a4bfadd87a61db0e7ce2366.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 235, + 279, + 300 + ], + "lines": [ + { + "bbox": [ + 45, + 235, + 279, + 300 + ], + "spans": [ + { + "bbox": [ + 45, + 235, + 279, + 300 + ], + "type": "text", + "content": "Figure 8. Visualization of class activation maps for different classes using DR and our TLDR. The classes are road, sidewalk, and terrain. Only the classes are displayed as opaque for better visualization in the unseen images. The TLDR model tends to have activation throughout broader areas than the DR model, suggesting it can rely on more texture cues when making predictions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 310, + 277, + 436 + ], + "blocks": [ + { + "bbox": [ + 51, + 310, + 277, + 436 + ], + "lines": [ + { + "bbox": [ + 51, + 310, + 277, + 436 + ], + "spans": [ + { + "bbox": [ + 51, + 310, + 277, + 436 + ], + "type": "image", + "image_path": "f291afe86909639704d8d49ef21f6cc14997bcd912729e375c7f7312d3a1bf74.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 445, + 279, + 478 + ], + "lines": [ + { + "bbox": [ + 45, + 445, + 279, + 478 + ], + "spans": [ + { + "bbox": [ + 45, + 445, + 279, + 478 + ], + "type": "text", + "content": "Figure 9. Comparison of the confusion matrices between DR and our TLDR for classes with similar shapes. The TLDR model tends to have less confusion and higher accuracy than the DR model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 492, + 279, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 492, + 279, + 527 + ], + "spans": [ + { + "bbox": [ + 45, + 492, + 279, + 527 + ], + "type": "text", + "content": "with DR and TLDR as the DR model and the TLDR model, respectively. The models are evaluated on Cityscapes in the experiments if no specific dataset is mentioned." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 529, + 279, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 529, + 279, + 669 + ], + "spans": [ + { + "bbox": [ + 45, + 529, + 279, + 669 + ], + "type": "text", + "content": "Class activation map. To validate whether the models utilize texture cues for prediction, we generate class activation maps and analyze the contribution of texture features to the predictions. We apply Grad-CAM [49] to road, sidewalk, and terrain on the DR and TLDR models. Figure 8 displays the attention heat maps generated for each class, highlighting the regions the models focus on during prediction. Our analysis reveals that the DR model tends to have high activation in edge regions, while the TLDR model tends to have activation throughout broader areas. One can infer that these broader areas include texture cues, which the TLDR model uses as valuable prediction cues." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 671, + 279, + 730 + ], + "type": "text", + "content": "Confusion matrix. We compare confusion matrices between the two models. Figure 9 shows the confusion matrices of the DR model and the TLDR model for shape-similar classes. The TLDR model has lower false positive rates in classifying sidewalk and terrain as road, with rates of" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 301, + 105, + 527, + 245 + ], + "blocks": [ + { + "bbox": [ + 301, + 105, + 527, + 245 + ], + "lines": [ + { + "bbox": [ + 301, + 105, + 527, + 245 + ], + "spans": [ + { + "bbox": [ + 301, + 105, + 527, + 245 + ], + "type": "image", + "image_path": "79702bec39ca23c74a93bf52fd95e09694b242311f446047eb305b0edea6ce36.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 252, + 530, + 305 + ], + "lines": [ + { + "bbox": [ + 296, + 252, + 530, + 305 + ], + "spans": [ + { + "bbox": [ + 296, + 252, + 530, + 305 + ], + "type": "text", + "content": "Figure 10. Qualitative results of DR and our TLDR on unseen images. The unseen images are from Cityscapes, BDD, Mapillary, and SYNTHIA in row order. The TLDR model provides better prediction results for shape-similar classes than the DR model, as seen in the white boxes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 326, + 316, + 500, + 392 + ], + "blocks": [ + { + "bbox": [ + 326, + 316, + 500, + 392 + ], + "lines": [ + { + "bbox": [ + 326, + 316, + 500, + 392 + ], + "spans": [ + { + "bbox": [ + 326, + 316, + 500, + 392 + ], + "type": "table", + "html": "
LayerDRTLDR
TextureShapeTextureShape
l = 166.5%29.4%66.7%29.2%
l = 256.6%39.8%57.3%39.1%
l = 332.0%63.5%33.8%61.9%
l = 419.1%54.3%20.1%52.5%
", + "image_path": "a760cbb8db630cc9a6981ec688dbcb8a3f2692ee473a780fb6f86eddbde8b663.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 399, + 530, + 442 + ], + "lines": [ + { + "bbox": [ + 296, + 399, + 530, + 442 + ], + "spans": [ + { + "bbox": [ + 296, + 399, + 530, + 442 + ], + "type": "text", + "content": "Table 3. Layer-wise dimensionality of texture and shape in the latent representations of DR and our TLDR, evaluated in percentages " + }, + { + "bbox": [ + 296, + 399, + 530, + 442 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 296, + 399, + 530, + 442 + ], + "type": "text", + "content": ". The TLDR model stores more texture information than the DR model across all the layers." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 458, + 530, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 458, + 530, + 516 + ], + "spans": [ + { + "bbox": [ + 296, + 458, + 530, + 516 + ], + "type": "inline_equation", + "content": "19.6\\%" + }, + { + "bbox": [ + 296, + 458, + 530, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 458, + 530, + 516 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 296, + 458, + 530, + 516 + ], + "type": "text", + "content": ", respectively, compared to the DR model, which has rates of " + }, + { + "bbox": [ + 296, + 458, + 530, + 516 + ], + "type": "inline_equation", + "content": "27.6\\%" + }, + { + "bbox": [ + 296, + 458, + 530, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 458, + 530, + 516 + ], + "type": "inline_equation", + "content": "11.6\\%" + }, + { + "bbox": [ + 296, + 458, + 530, + 516 + ], + "type": "text", + "content": " (see Figure 9a). Also, there is an increase in the accuracy of each class. There are also clear reductions in confusion between building, wall & fence and car, truck & bus (see Figures 9b and 9c)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 518, + 530, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 518, + 530, + 564 + ], + "spans": [ + { + "bbox": [ + 296, + 518, + 530, + 564 + ], + "type": "text", + "content": "Qualitative results. Figure 10 shows the qualitative results of the DR and TLDR models in various domains. The TLDR model provides better prediction results for shape-similar classes than the DR model (see white boxes)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 566, + 530, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 566, + 530, + 717 + ], + "spans": [ + { + "bbox": [ + 296, + 566, + 530, + 717 + ], + "type": "text", + "content": "Dimensionality analysis. We investigate whether the TLDR model encodes additional texture information compared to the DR model. We use the method proposed by Islam et al. to quantify the dimensions corresponding to texture and shape in latent representations [25]. The method estimates the dimensionality of a semantic concept by computing mutual information between feature representations of input pairs that exhibit the same semantic concept. A detailed explanation of the experiment is in Appendix A. Table 3 shows the estimated dimensions (\\%) for texture and shape in each layer of the DR and TLDR models. At every layer, it is apparent that the TLDR model encodes more texture information than the DR model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 718, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 718, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 308, + 718, + 530, + 730 + ], + "type": "text", + "content": "The experimental results demonstrate that the TLDR" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "683" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 103, + 273, + 190 + ], + "blocks": [ + { + "bbox": [ + 50, + 103, + 273, + 190 + ], + "lines": [ + { + "bbox": [ + 50, + 103, + 273, + 190 + ], + "spans": [ + { + "bbox": [ + 50, + 103, + 273, + 190 + ], + "type": "table", + "html": "
LorigLstylLTRLTGCBMS
1✓---36.1231.3830.8937.43
2-✓--39.4839.2143.4531.96
3✓✓--43.6243.9444.0638.75
4✓✓✓-46.8546.0746.5238.64
5✓✓-✓46.3644.0147.4938.10
6✓✓✓✓47.5844.8848.8039.35
", + "image_path": "e4b5fbedbcc69f9cacf9ea831e8de411a1faf330fe4a297d3ee1139cca540473.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 241, + 261, + 327 + ], + "blocks": [ + { + "bbox": [ + 45, + 197, + 279, + 229 + ], + "lines": [ + { + "bbox": [ + 45, + 197, + 279, + 229 + ], + "spans": [ + { + "bbox": [ + 45, + 197, + 279, + 229 + ], + "type": "text", + "content": "Table 4. Ablation experiment on each loss in TLDR. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 241, + 261, + 327 + ], + "lines": [ + { + "bbox": [ + 62, + 241, + 261, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 261, + 327 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{\\text{TR}} \\)\\( \\mathcal{L}_{\\text{TG}} \\)TEOCBMS
1✓--44.9544.0944.2439.07
2✓-✓46.8546.0746.5238.64
3-✓-44.2543.5346.9038.96
4-✓✓46.3644.0147.4938.10
5✓✓-45.7044.1346.2038.57
6✓✓✓47.5844.8848.8039.35
", + "image_path": "8523074a3f641654e5e7d6045388634fe1b6648d888e0dfcecf2a6ab961fd457.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 384, + 279, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 384, + 279, + 419 + ], + "spans": [ + { + "bbox": [ + 45, + 384, + 279, + 419 + ], + "type": "text", + "content": "model can effectively distinguish between classes with similar shapes by encoding and using texture as additional discriminative cues." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 427, + 136, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 136, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 136, + 440 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 446, + 279, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 446, + 279, + 480 + ], + "spans": [ + { + "bbox": [ + 45, + 446, + 279, + 480 + ], + "type": "text", + "content": "In ablation experiments, we use ResNet-101 as the encoder, train on GTA, and evaluate on C, B, M, and S. The best results are highlighted in tables." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 483, + 279, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 483, + 279, + 611 + ], + "spans": [ + { + "bbox": [ + 45, + 483, + 279, + 611 + ], + "type": "text", + "content": "Loss components. We investigate how each loss component contributes to overall performance. Table 4 shows the mIoU performance change with respect to the ablation of loss components. There is a large performance improvement when the stylized task loss and the original task loss are used simultaneously (cf. row 3). It seems to be the effect of using shape and texture cues as complementary. There is an additional performance improvement when the texture regularization loss and the texture generalization loss are added separately (cf. rows 4 and 5). The overall performance is the best when both losses are used (cf. row 6)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 612, + 279, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 612, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 45, + 612, + 279, + 730 + ], + "type": "text", + "content": "Texture Extraction Operator (TEO). We use TEO to extract only texture features from feature maps. We conduct experiments to verify whether TEO leads to performance improvement in the texture regularization loss and the texture generalization loss. In the absence of TEO, " + }, + { + "bbox": [ + 45, + 612, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TR}}" + }, + { + "bbox": [ + 45, + 612, + 279, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 612, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TG}}" + }, + { + "bbox": [ + 45, + 612, + 279, + 730 + ], + "type": "text", + "content": " refer to the calculation of direct consistency in the feature maps without any alterations. Table 5 shows the results of TEO ablation experiments. The results show that the performance is improved in the presence of TEO (cf. rows 2, 4, and 6) in both " + }, + { + "bbox": [ + 45, + 612, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TR}}" + }, + { + "bbox": [ + 45, + 612, + 279, + 730 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 612, + 279, + 730 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TG}}" + }, + { + "bbox": [ + 45, + 612, + 279, + 730 + ], + "type": "text", + "content": " compared to the absence of" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 103, + 510, + 175 + ], + "blocks": [ + { + "bbox": [ + 45, + 335, + 279, + 368 + ], + "lines": [ + { + "bbox": [ + 45, + 335, + 279, + 368 + ], + "spans": [ + { + "bbox": [ + 45, + 335, + 279, + 368 + ], + "type": "text", + "content": "Table 5. Ablation experiments on TEO. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 103, + 510, + 175 + ], + "lines": [ + { + "bbox": [ + 318, + 103, + 510, + 175 + ], + "spans": [ + { + "bbox": [ + 318, + 103, + 510, + 175 + ], + "type": "table", + "html": "
CaseCBMS
1w/o RSM46.5744.8947.1738.87
2w/ RSM (τ=0.01)46.9244.0147.7538.61
3w/ RSM (τ=0.1)47.5844.8848.8039.35
4w/o LDF46.7744.8446.8739.72
5w/ LDF47.5844.8848.8039.35
", + "image_path": "889925936ac137f9dca060fedb1fdf1cb9ef94ffad64c2bfc88977fdec289e89.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 314, + 229, + 512, + 280 + ], + "blocks": [ + { + "bbox": [ + 296, + 183, + 530, + 217 + ], + "lines": [ + { + "bbox": [ + 296, + 183, + 530, + 217 + ], + "spans": [ + { + "bbox": [ + 296, + 183, + 530, + 217 + ], + "type": "text", + "content": "Table 6. Ablation experiments on the design choices of TLDR. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 229, + 512, + 280 + ], + "lines": [ + { + "bbox": [ + 314, + 229, + 512, + 280 + ], + "spans": [ + { + "bbox": [ + 314, + 229, + 512, + 280 + ], + "type": "table", + "html": "
Random StyleCBMS
1WikiArt [41]45.0944.5845.1639.52
2DTD [11]45.6545.1045.2939.78
3ImageNet val [13]47.5844.8848.8039.35
", + "image_path": "14a3867c4093fd3aba8ef0b29b540f264d6991903dbf9b1441629d18792271fa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 287, + 530, + 320 + ], + "lines": [ + { + "bbox": [ + 296, + 287, + 530, + 320 + ], + "spans": [ + { + "bbox": [ + 296, + 287, + 530, + 320 + ], + "type": "text", + "content": "Table 7. Ablation experiment on different random style datasets. The model is trained on GTA using ResNet-101. The best results are highlighted. The default setting is marked in gray." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 297, + 337, + 405, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 337, + 405, + 349 + ], + "spans": [ + { + "bbox": [ + 297, + 337, + 405, + 349 + ], + "type": "text", + "content": "TEO (cf. rows 1, 3, and 5)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 351, + 530, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 351, + 530, + 457 + ], + "spans": [ + { + "bbox": [ + 296, + 351, + 530, + 457 + ], + "type": "text", + "content": "Random Style Masking (RSM). We conduct an experiment to validate the effectiveness of RSM. As shown in Table 6, the model with RSM (cf. row 3) performs better than the model without RSM (cf. row 1). Also, when the threshold of RSM is decreased to " + }, + { + "bbox": [ + 296, + 351, + 530, + 457 + ], + "type": "inline_equation", + "content": "\\tau = 0.01" + }, + { + "bbox": [ + 296, + 351, + 530, + 457 + ], + "type": "text", + "content": " (cf. row 2), the performance is higher than without RSM (cf. row 1) but lower than the default setting " + }, + { + "bbox": [ + 296, + 351, + 530, + 457 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 296, + 351, + 530, + 457 + ], + "type": "text", + "content": " (cf. row 3). The experiment suggests that maintaining an appropriate RSM threshold can aid in learning random styles." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 459, + 530, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 459, + 530, + 587 + ], + "spans": [ + { + "bbox": [ + 296, + 459, + 530, + 587 + ], + "type": "text", + "content": "Linear Decay Factor (LDF). When analyzing the texture regularization and generalization losses, we observe that the losses do not naturally decrease as much as the task losses (see Appendix B). The texture regularization loss remains relatively constant likely due to the frozen ImageNet model. The texture generalization loss is also relatively stable likely due to the different combinations of source images and random style images used in each epoch. We design LDF to match the scale of the losses. As shown in Table 6, the model with LDF (cf. row 5) performs better than the model without LDF (cf. row 4)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "spans": [ + { + "bbox": [ + 296, + 590, + 530, + 730 + ], + "type": "text", + "content": "Random style dataset. We analyze how performance changes when training on various random style datasets. Table 7 shows the performance when using WikiArt [41] dataset, Describable Texture Dataset (DTD) [11] and ImageNet [13] validation set as the random style dataset. The overall performance is the best when using the ImageNet validation set as the random style dataset (cf. row 3). The performances remain consistently demonstrated even using the other two datasets (cf. rows 1 and 2). The performance improvement when using ImageNet is probably due to the superiority of the ImageNet pre-trained model to extract the texture present in ImageNet." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "684" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 105, + 116, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 105, + 116, + 117 + ], + "spans": [ + { + "bbox": [ + 46, + 105, + 116, + 117 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 125, + 281, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 125, + 281, + 277 + ], + "spans": [ + { + "bbox": [ + 45, + 125, + 281, + 277 + ], + "type": "text", + "content": "Texture often contributes to the domain gap in DGSS. Existing DGSS methods have attempted to eliminate or randomize texture features. However, this paper argues that texture remains supplementary prediction cues for shape despite the domain gap. Accordingly, we proposed TLDR to learn texture features without overfitting to source domain textures in DGSS. TLDR includes novel texture regularization and generalization losses, using a Gram-matrix as a key component. We conducted a diverse set of experiments to demonstrate that TLDR effectively differentiates shape-similar classes by leveraging texture as a prediction cue. The experiments on multiple DGSS tasks show that our TLDR achieves state-of-the-art performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 279, + 279, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 279, + 279, + 373 + ], + "spans": [ + { + "bbox": [ + 45, + 279, + 279, + 373 + ], + "type": "text", + "content": "Limitation. During the experiments, we found that texture differences between domains vary class-wise. Learning more source domain textures for classes with small texture differences may help with generalization, but it may be advantageous for classes with high texture differences to learn less of them. In our method, there is no class-wise prescription for different texture differences. We leave it as an interesting future work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 385, + 145, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 145, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 145, + 398 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 405, + 279, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 405, + 279, + 487 + ], + "spans": [ + { + "bbox": [ + 45, + 405, + 279, + 487 + ], + "type": "text", + "content": "We sincerely thank Chanyong Lee and Eunjin Koh for their constructive discussions and support. We also appreciate Chaehyeon Lim, Sihyun Yu, Seokhyun Moon, and Youngju Yoo for providing insightful feedback. This work was supported by the Agency for Defense Development (ADD) grant funded by the Korea government (912855101)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 499, + 103, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 103, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 103, + 511 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 50, + 519, + 279, + 730 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 50, + 519, + 279, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 519, + 279, + 572 + ], + "spans": [ + { + "bbox": [ + 50, + 519, + 279, + 572 + ], + "type": "text", + "content": "[1] Vijay Badrinarayan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 39(12):2481-2495, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 51, + 575, + 279, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 575, + 279, + 618 + ], + "spans": [ + { + "bbox": [ + 51, + 575, + 279, + 618 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 619, + 279, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 619, + 279, + 650 + ], + "spans": [ + { + "bbox": [ + 51, + 619, + 279, + 650 + ], + "type": "text", + "content": "[3] John Canny. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), (6):679-698, 1986. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 653, + 279, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 653, + 279, + 706 + ], + "spans": [ + { + "bbox": [ + 51, + 653, + 279, + 706 + ], + "type": "text", + "content": "[4] Fabio M Carlucci, Antonio D'Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. Domain generalization by solving jigsaw puzzles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2229-2238, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 51, + 708, + 279, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 708, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 51, + 708, + 279, + 730 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 530, + 728 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 317, + 106, + 530, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 530, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 530, + 149 + ], + "type": "text", + "content": "segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 40(4):834-848, 2017. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 303, + 151, + 530, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 151, + 530, + 192 + ], + "spans": [ + { + "bbox": [ + 303, + 151, + 530, + 192 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 303, + 195, + 530, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 195, + 530, + 248 + ], + "spans": [ + { + "bbox": [ + 303, + 195, + 530, + 248 + ], + "type": "text", + "content": "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 801-818, 2018. 1, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 250, + 530, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 250, + 530, + 292 + ], + "spans": [ + { + "bbox": [ + 303, + 250, + 530, + 292 + ], + "type": "text", + "content": "[8] Wuyang Chen, Zhiding Yu, Shalini De Mello, Sifei Liu, Jose M Alvarez, Zhangyang Wang, and Anima Anandkumar. Contrastive syn-to-real generalization. arXiv preprint arXiv:2104.02290, 2021. 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 303, + 293, + 530, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 293, + 530, + 336 + ], + "spans": [ + { + "bbox": [ + 303, + 293, + 530, + 336 + ], + "type": "text", + "content": "[9] Wuyang Chen, Zhiding Yu, Zhangyang Wang, and Animashree Anandkumar. Automated synthetic-to-real generalization. In Proceedings of the International Conference on Machine Learning (ICML), pages 1746-1756, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 338, + 530, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 338, + 530, + 402 + ], + "spans": [ + { + "bbox": [ + 299, + 338, + 530, + 402 + ], + "type": "text", + "content": "[10] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11580-11590, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 403, + 530, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 403, + 530, + 455 + ], + "spans": [ + { + "bbox": [ + 299, + 403, + 530, + 455 + ], + "type": "text", + "content": "[11] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606-3613, 2014. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 458, + 530, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 458, + 530, + 531 + ], + "spans": [ + { + "bbox": [ + 299, + 458, + 530, + 531 + ], + "type": "text", + "content": "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3213-3223, 2016. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 533, + 530, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 533, + 530, + 586 + ], + "spans": [ + { + "bbox": [ + 299, + 533, + 530, + 586 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248-255, 2009. 3, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 588, + 530, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 588, + 530, + 630 + ], + "spans": [ + { + "bbox": [ + 299, + 588, + 530, + 630 + ], + "type": "text", + "content": "[14] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. Advances in Neural Information Processing Systems (NeurIPS), 32, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 632, + 530, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 632, + 530, + 684 + ], + "spans": [ + { + "bbox": [ + 299, + 632, + 530, + 684 + ], + "type": "text", + "content": "[15] Mark Everingham, SM Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision (IJCV), 111(1):98-136, 2015. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 687, + 530, + 728 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 687, + 530, + 728 + ], + "spans": [ + { + "bbox": [ + 299, + 687, + 530, + 728 + ], + "type": "text", + "content": "[16] Leon Gatys, Alexander S Ecker, and Matthias Bethge. Texture synthesis using convolutional neural networks. Advances in Neural Information Processing Systems (NeurIPS), 28, 2015. 4" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "685" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 729 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 159 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 159 + ], + "type": "text", + "content": "[17] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2414-2423, 2016. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 160, + 279, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 160, + 279, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 160, + 279, + 202 + ], + "type": "text", + "content": "[18] Yunhao Ge, Yao Xiao, Zhi Xu, Xingrui Wang, and Laurent Itti. Contributions of Shape, Texture, and Color in Visual Recognition. In Proceedings of the European Conference on Computer Vision (ECCV), pages 369-386, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 204, + 279, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 204, + 279, + 256 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 279, + 256 + ], + "type": "text", + "content": "[19] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv preprint arXiv:1811.12231, 2018. 1, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 257, + 279, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 279, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 279, + 309 + ], + "type": "text", + "content": "[20] Priya Goyal, Piotr Dólar, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 310, + 279, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 279, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 279, + 331 + ], + "type": "text", + "content": "[21] Robert M Haralick. Statistical and structural approaches to texture. Proceedings of the IEEE, 67(5):786-804, 1979. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 333, + 279, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 333, + 279, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 333, + 279, + 374 + ], + "type": "text", + "content": "[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2961-2969, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 375, + 279, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 375, + 279, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 375, + 279, + 417 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 418, + 279, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 279, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 279, + 471 + ], + "type": "text", + "content": "[24] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6891-6902, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 472, + 279, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 279, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 279, + 514 + ], + "type": "text", + "content": "[25] Md Amirul Islam, Matthew Kowal, Patrick Esser, Sen Jia, Bjorn Ommer, Konstantinos G Derpanis, and Neil Bruce. Shape or texture: Understanding discriminative features in cnns. arXiv preprint arXiv:2101.11604, 2021. 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 515, + 279, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 279, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 279, + 578 + ], + "type": "text", + "content": "[26] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and Statistical Texture Knowledge Distillation for Semantic Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16876-16885, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 579, + 279, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 279, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 279, + 631 + ], + "type": "text", + "content": "[27] Yunpei Jia, Jie Zhang, Shiguang Shan, and Xilin Chen. Single-side domain generalization for face anti-spoofing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8484-8493, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 633, + 279, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 279, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 279, + 685 + ], + "type": "text", + "content": "[28] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. WildNet: Learning Domain Generalized Semantic Segmentation from the Wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9936-9946, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 279, + 729 + ], + "type": "text", + "content": "[29] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 107, + 530, + 728 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 318, + 107, + 530, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 107, + 530, + 127 + ], + "spans": [ + { + "bbox": [ + 318, + 107, + 530, + 127 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition (CVPR), pages 224-233, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 128, + 530, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 128, + 530, + 171 + ], + "spans": [ + { + "bbox": [ + 299, + 128, + 530, + 171 + ], + "type": "text", + "content": "[30] Yijun Li, Ming-Yu Liu, Xueting Li, Ming-Hsuan Yang, and Jan Kautz. A closed-form solution to photorealistic image stylization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 453-468, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 172, + 530, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 172, + 530, + 224 + ], + "spans": [ + { + "bbox": [ + 299, + 172, + 530, + 224 + ], + "type": "text", + "content": "[31] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European Conference on Computer Vision (ECCV), pages 624-639, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 225, + 530, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 225, + 530, + 255 + ], + "spans": [ + { + "bbox": [ + 299, + 225, + 530, + 255 + ], + "type": "text", + "content": "[32] Yanghao Li, Naiyan Wang, Jiaying Liu, and Xiaodi Hou. Demystifying neural style transfer. arXiv preprint arXiv:1701.01036, 2017. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 256, + 530, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 256, + 530, + 288 + ], + "spans": [ + { + "bbox": [ + 299, + 256, + 530, + 288 + ], + "type": "text", + "content": "[33] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 40(12):2935-2947, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 289, + 530, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 289, + 530, + 331 + ], + "spans": [ + { + "bbox": [ + 299, + 289, + 530, + 331 + ], + "type": "text", + "content": "[34] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3431-3440, 2015. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 332, + 530, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 332, + 530, + 353 + ], + "spans": [ + { + "bbox": [ + 299, + 332, + 530, + 353 + ], + "type": "text", + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 354, + 530, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 354, + 530, + 406 + ], + "spans": [ + { + "bbox": [ + 299, + 354, + 530, + 406 + ], + "type": "text", + "content": "[36] Krikamol Muandet, David Balduzzi, and Bernhard Scholkopf. Domain generalization via invariant feature representation. In Proceedings of the International Conference on Machine Learning (ICML), pages 10-18, 2013. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 408, + 530, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 408, + 530, + 450 + ], + "spans": [ + { + "bbox": [ + 299, + 408, + 530, + 450 + ], + "type": "text", + "content": "[37] Gilhyun Nam, Gyeongjae Choi, and Kyungmin Lee. GCISG: Guided Causal Invariant Learning for Improved Syn-to-Real Generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 656–672, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 451, + 530, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 451, + 530, + 503 + ], + "spans": [ + { + "bbox": [ + 299, + 451, + 530, + 503 + ], + "type": "text", + "content": "[38] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8690-8699, 2021. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 504, + 530, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 504, + 530, + 556 + ], + "spans": [ + { + "bbox": [ + 299, + 504, + 530, + 556 + ], + "type": "text", + "content": "[39] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. Advances in Neural Information Processing Systems (NeurIPS), 34:23296-23308, 2021. 1, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 557, + 530, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 557, + 530, + 611 + ], + "spans": [ + { + "bbox": [ + 299, + 557, + 530, + 611 + ], + "type": "text", + "content": "[40] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4990-4999, 2017. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 612, + 530, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 612, + 530, + 631 + ], + "spans": [ + { + "bbox": [ + 299, + 612, + 530, + 631 + ], + "type": "text", + "content": "[41] Kiri Nichol. Painter by numbers, wikiart. Kiri Nichol, 2016. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 633, + 530, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 633, + 530, + 676 + ], + "spans": [ + { + "bbox": [ + 299, + 633, + 530, + 676 + ], + "type": "text", + "content": "[42] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 1, 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 676, + 530, + 728 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 676, + 530, + 728 + ], + "spans": [ + { + "bbox": [ + 299, + 676, + 530, + 728 + ], + "type": "text", + "content": "[43] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1863-1871, 2019. 1, 2, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "686" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 730 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 47, + 106, + 279, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 106, + 279, + 159 + ], + "spans": [ + { + "bbox": [ + 47, + 106, + 279, + 159 + ], + "type": "text", + "content": "[44] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2594-2605, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 161, + 279, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 161, + 279, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 161, + 279, + 204 + ], + "type": "text", + "content": "[45] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE Transactions on Image Processing, 30:6594-6608, 2021. 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 206, + 279, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 279, + 248 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 279, + 248 + ], + "type": "text", + "content": "[46] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12556-12565, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 249, + 279, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 249, + 279, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 249, + 279, + 292 + ], + "type": "text", + "content": "[47] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Proceedings of the European conference on computer vision (ECCV), pages 102-118, 2016. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 293, + 279, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 279, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 279, + 357 + ], + "type": "text", + "content": "[48] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3234-3243, 2016. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 358, + 279, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 279, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 279, + 421 + ], + "type": "text", + "content": "[49] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 618-626, 2017. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 423, + 279, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 279, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 279, + 477 + ], + "type": "text", + "content": "[50] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10023-10031, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 478, + 279, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 279, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 279, + 510 + ], + "type": "text", + "content": "[51] Nathan Somavarapu, Chih-Yao Ma, and Zsolt Kira. Frustratingly simple domain generalization via image stylization. arXiv preprint arXiv:2006.11207, 2020. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 511, + 279, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 279, + 543 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 279, + 543 + ], + "type": "text", + "content": "[52] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research (JMLR), 9(11), 2008. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 544, + 279, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 279, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 279, + 596 + ], + "type": "text", + "content": "[53] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. Advances in Neural Information Processing Systems (NeurIPS), 31, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 599, + 279, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 279, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 279, + 652 + ], + "type": "text", + "content": "[54] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. SiamDoGe: Domain Generalizable Semantic Segmentation Using Siamese Network. In Proceedings of the European Conference on Computer Vision (ECCV), pages 603-620, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 654, + 279, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 279, + 706 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 279, + 706 + ], + "type": "text", + "content": "[55] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. SegFormer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems (NeurIPS), 34:12077-12090, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 708, + 279, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 708, + 279, + 730 + ], + "spans": [ + { + "bbox": [ + 47, + 708, + 279, + 730 + ], + "type": "text", + "content": "[56] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generaliza" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 106, + 530, + 507 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 317, + 106, + 530, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 530, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 530, + 138 + ], + "type": "text", + "content": "tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14383-14392, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 140, + 530, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 140, + 530, + 182 + ], + "spans": [ + { + "bbox": [ + 299, + 140, + 530, + 182 + ], + "type": "text", + "content": "[57] Zheng Xu, Wen Li, Li Niu, and Dong Xu. Exploiting low-rank structure from latent domains for domain generalization. In Proceedings of the European Conference on Computer Vision (ECCV), pages 628-643, 2014. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 184, + 530, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 184, + 530, + 247 + ], + "spans": [ + { + "bbox": [ + 299, + 184, + 530, + 247 + ], + "type": "text", + "content": "[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2636-2645, 2020. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 249, + 530, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 249, + 530, + 322 + ], + "spans": [ + { + "bbox": [ + 299, + 249, + 530, + 322 + ], + "type": "text", + "content": "[59] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2100-2110, 2019. 1, 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 324, + 530, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 324, + 530, + 378 + ], + "spans": [ + { + "bbox": [ + 299, + 324, + 530, + 378 + ], + "type": "text", + "content": "[60] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 535-552, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 378, + 530, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 378, + 530, + 420 + ], + "spans": [ + { + "bbox": [ + 299, + 378, + 530, + 420 + ], + "type": "text", + "content": "[61] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. arXiv preprint arXiv:2207.04892, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 421, + 530, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 421, + 530, + 453 + ], + "spans": [ + { + "bbox": [ + 299, + 421, + 530, + 453 + ], + "type": "text", + "content": "[62] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain generalization with mixstyle. arXiv preprint arXiv:2104.02008, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 455, + 530, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 455, + 530, + 507 + ], + "spans": [ + { + "bbox": [ + 299, + 455, + 530, + 507 + ], + "type": "text", + "content": "[63] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12537-12546, 2021. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "spans": [ + { + "bbox": [ + 281, + 747, + 296, + 756 + ], + "type": "text", + "content": "687" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/b4b142a9-63c6-4b31-9bbd-a4e1e31c2d72_content_list.json b/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/b4b142a9-63c6-4b31-9bbd-a4e1e31c2d72_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..012d5bdc67395f30f2837b027768b66cc5b84e80 --- /dev/null +++ b/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/b4b142a9-63c6-4b31-9bbd-a4e1e31c2d72_content_list.json @@ -0,0 +1,1551 @@ +[ + { + "type": "text", + "text": "The Devil is in the Crack Orientation: A New Perspective for Crack Detection", + "text_level": 1, + "bbox": [ + 94, + 130, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Guanming Zhu, Zun Liu, Jie Chen, Jianqiang Li Shenzhen University, Shenzhen, China", + "bbox": [ + 104, + 180, + 864, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chenzhuangzhuang2016，2018101068，zhuguanming2022}@email.szu.edu.cn", + "bbox": [ + 184, + 218, + 784, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jin.zhang, chenjie, zunliu, lijq}@szu.edu.cn", + "bbox": [ + 285, + 236, + 683, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cracks are usually curve-like structures that are the focus of many computer-vision applications (e.g., road safety inspection and surface inspection of the industrial facilities). The existing pixel-based crack segmentation methods rely on time-consuming and costly pixel-level annotations. And the object-based crack detection methods exploit the horizontal box to detect the crack without considering crack orientation, resulting in scale variation and intraclass variation. Considering this, we provide a new perspective for crack detection that models the cracks as a series of sub-cracks with the corresponding orientation. However, the vanilla adaptation of the existing oriented object detection methods to the crack detection tasks will result in limited performance, due to the boundary discontinuity issue and the ambiguities in sub-crack orientation. In this paper, we propose a first-of-its-kind oriented sub-crack detector, dubbed as CrackDet, which is derived from a novel piecewise angle definition, to ease the boundary discontinuity problem. And then, we propose a multi-branch angle regression loss for learning sub-crack orientation and variance together. Since there are no related benchmarks, we construct three fully annotated datasets, namely, ORC, ONPP, and OCCSD, which involve various cracks in road pavement and industrial facilities. Experiments show that our approach outperforms state-of-the-art crack detectors.", + "bbox": [ + 76, + 319, + 473, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 742, + 209, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Crack is a type of defect that appears on the surfaces of many physical structures, e.g., the road pavement [41], the industrial facilities [8]. Inspecting and repairing cracks are crucial tasks for avoiding expansion of the harm to these structures [77]. Besides, a recent study reveals that recognizing the surface cracks caused by earthquakes in time can avoid secondary damage to people's properties", + "bbox": [ + 75, + 768, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/353ce6bab1dc650c20cb1584d0fcd775c70a88ecf20b729933de78f7d1ae718b.jpg", + "image_caption": [ + "Images", + "Segmentation", + "Detection", + "Ours", + "(a) Visualization of different perspectives for crack detection" + ], + "image_footnote": [], + "bbox": [ + 496, + 301, + 893, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ebdacb2e49e1b25ae2264cb8df0e79f084d4aa4945f46ef7e5c2a3f4fbf3371d.jpg", + "image_caption": [ + "(b) Boundary discontinuity", + "Figure 1. Cracks tend to be curve-like structures [77] in real-world applications. (a) We compare our orientation-based perspective with others for crack detection. (b) Boundary discontinuity issue in the existing 180-degree definition. (c) The sub-crack orientation is ambiguous in some cases due to the irregular sub-crack shape." + ], + "image_footnote": [], + "bbox": [ + 509, + 547, + 679, + 623 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/38ce88cbf2af0d07252c02137be66dbeaa620a0a9172513e275408cd4ffd606b.jpg", + "image_caption": [ + "(c) Ambiguities in orientation" + ], + "image_footnote": [], + "bbox": [ + 707, + 542, + 879, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and lives [67]. However, manual inspection is costly, time-consuming, and prone to human error [34]. Hence, automatic image-based crack detection plays an essential role in a real-world application of crack identification [22, 67].", + "bbox": [ + 496, + 744, + 890, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, image-based crack detection methods can be roughly divided into three categories [36]: patch-level classification [4, 8], object-level detection [54, 31], and pixel-level crack segmentation [76, 49]. Although the first two methods take less labeling cost than the last one, they can only provide a coarse crack localization. Moreover, we are", + "bbox": [ + 496, + 809, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author Email: chenjie@szu.edu.cn", + "bbox": [ + 94, + 886, + 370, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6653", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "motivated by the fact that a crack is a line object that has a certain width from a local perspective [77]. Then, we are encouraged to start from an orientation-based perspective that models the cracks as a series of oriented sub-cracks (see Fig. 1 (a)). The advantage of this perspective is three-fold: (1) eliminating the scale variation, (2) reducing the intraclass variation, and (3) taking less labeling cost than the segmentation methods. More specifically, in the case of our practice, the price and time cost of the pixel-level annotations are $15 \\times$ and $30 \\times$ higher than our oriented sub-crack annotations, respectively. Besides, considering real-world applications, existing works [33, 35] show great power by taking a detection-then-segmentation strategy, i.e., it first identifies and localizes cracks by a detector and then obtains crack details by segmentation-based methods. Hence, the detector based on our perspective can serve as a complement to segmentation-based methods, as it can provide a more fine-grained localization (see Fig. 1 a).", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, it is difficult to directly adopt the state-of-art oriented detectors [16, 39, 23, 37] to crack detection. The reason is two-fold: (1) several existing works (e.g., ReDet [16], DRN [39]) simply ignore the boundary discontinuity problem, and (2) although the remaining works make great efforts on the above problem, these works still suffer from ambiguity issue in the sub-crack orientation. Among existing oriented detection methods, those detectors based on five parameters $(x,y,h,w,\\theta)$ dominate [58], where $(x,y)$ indicates the center point and $(h,w,\\theta)$ denotes height, width, and angle, respectively. As shown in Fig. 1 (b), the boundary discontinuity problem indicates that small rotations for oriented objects around the angle period's boundary result in a sudden sharp increase of loss [61]. Moreover, Fig. 1 (c) provides some cases to illustrate the inherent ambiguities in sub-crack orientation. When the growth direction of a crack changes frequently, the orientation of the sub-crack is unclear to judge, making it hard to learn regression functions for estimating orientation.", + "bbox": [ + 75, + 367, + 472, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper starts from a new perspective for crack detection that models the cracks as a series of oriented subcracks. After that, we adopt CenterNet [73] as our baseline due to its simplicity. Next, we propose a new piecewise angle definition to ease the boundary discontinuity issue. Based on our piecewise definition, we propose the CrackDet that contains a parallel multi-branch architecture, each of which is responsible for predicting a fixed range of angles. Then, to address the ambiguities in sub-crack orientation, we propose a novel multi-branch angle regression loss, namely MAR Loss, which is based on Wasserstein distances [1] for both angle regression and angle variance estimation. The advantage of coupling CrackDet with MAR Loss is three-fold: (1) the ambiguities in sub-crack orientation can be captured to a certain extent. (2) The learned variance from each branch is beneficial during the post-process. We", + "bbox": [ + 75, + 659, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "acquire the estimated variances from each branch and obtain the detection results by variance voting. (3) Since crack orientation is significant in a real-world application [38], the learned variances reveal the level of confidence in the angle prediction, it can potentially help to judge the overall crack orientation. Our main contributions are summarized as follows:", + "bbox": [ + 496, + 90, + 890, + 195 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) To the best of our knowledge, we are the first ones to start from an oriented sub-crack detection perspective for crack detection by modeling cracks as a series of sub-cracks with their inherent orientations.", + "(2) Propose a first-of-its-kind oriented sub-crack detector, called as CrackDet, which stems from the proposed piecewise angle definition (see Sec. 3.1), to ease the boundary discontinuity issue.", + "(3) Propose a multi-branch angle regression loss that enables CrackDet to capture ambiguities in sub-crack orientation via jointly learning sub-crack orientation and the corresponding orientation variance (see Sec. 3.2 and 3.3).", + "(4) A tale of three datasets has been constructed, namely, ORC, ONPP, and OCCSD. These datasets contain orientation annotations of the sub-cracks in real-world industrial facilities and road pavements, to facilitate the research about oriented sub-crack detection." + ], + "bbox": [ + 496, + 196, + 890, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 467, + 637, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By following the previous study [36], the vision-based crack detection can be roughly clustered into three categories: patch-level classification [4, 8], object-level detection [54, 31], and pixel-level segmentation [77, 48].", + "bbox": [ + 496, + 492, + 890, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Patch-level crack classification. Recently, deep learning-based methods are more powerful than the traditional methods (e.g., HOG [42], LBP[42], Gaussian process, SVM [42]) in the patch-level crack classification tasks [36]. Among deep learning-based methods, variants of convolutional neural network (CNN) followed by a fully connected (FC) layer are used to recognize patch-level crack samples. Typically, ResNet [17] and VGG [47] are popular backbones among existing backbones. Moreover, focal loss [26] focuses on the hard samples to learn discriminative features. However, patch-level methods fail in providing orientation and the fine-grained localization of crack objects [36].", + "bbox": [ + 496, + 553, + 890, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object-level crack detection. For object-level crack detection, a bounding box is exploited to describe the location and the coarse size of a crack object. Most existing crack detectors mainly adopt the general object detectors, such as Faster R-CNN [44], RetinaNet [26], Yolo [43], CenterNet [12], and SSD [28]. Specifically, Fen et al. [13] first design a hybrid approach that uses Faster R-CNN to detect the crack patch. Then, as the crack orientation is important in the post-process, a deep CNN is exploited to obtain the orientation of each detected patch. Ma et al. [31] design an automatic crack detection system by using YOLO v3 [75]. To", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6654", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "detect various cracks, Wang et al. [51] propose a crack detection model by using CenterNet to extract crack features. Moreover, a RetinaNet-based crack detector is proposed to detect the crack in asphalt pavement [50]. Unfortunately, as shown in Fig. 1, the object-level crack detection methods suffer from scale variation and intra-class variation. Moreover, it also can not directly provide the orientation and the fine-grained localization of cracks.", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pixel-level crack segmentation. Since those horizontal bounding boxes can not describe the exact shape of the crack, segmentation-based methods are proposed to detect cracks at the pixel level. Yang et al. [59] design an automatic crack segmentation pipeline by using fully convolutional networks [30] and achieve a noticeable performance gain over traditional approaches [76, 46]. By coupling with Deeplab v3+ [5], CrackSeg [48] is proposed to learn rich features for different scales. Moreover, aiming to learn high-level crack representations, DeepCrack [77] is proposed to capture the line structures by feature fusion. Considering different receptive fields, FPHBN [57] makes great efforts for crack detection by aggregating pyramid features. Besides, by addressing the imbalance problem, Li et al. [22] propose an adaptive weighted cross-entropy loss function, which serves as a complement to the existing methods. However, these segmentation-based methods heavily rely on high-quality labeled datasets [6, 72], which involves a high labeling cost.", + "bbox": [ + 75, + 214, + 470, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to the above methods, we start from a new perspective that turns crack detection into oriented subcrack detection. Based on this perspective, we are allowed to provide a fine-grained localization of crack objects (See details in Fig. 1), instead of a coarse localization by the patch-level classification or object-level detection methods. Hence, considering the existing detectionthen-segmentation strategy [33, 35], the detector based on our perspective can serve as a better candidate for the segmentation-based methods. Besides, when considering a dynamic field without historical human supervision, our perspective can greatly ease the labeling cost than the pixellevel segmentation methods.", + "bbox": [ + 75, + 505, + 468, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Oriented object detection. Since our oriented sub-crack detection perspective belongs to the oriented object detection category, we give a brief review of related research on this scope. Recently, many powerful rotation detectors are mainly derived from horizontal object detectors by adding an orientation regression branch [23]. For example, ICN [2], ROI-Transformer [11], SCRDet [64], CAD-Net [70], CSL[60], $\\mathbf{R}^3\\mathrm{Det}$ [62], ReDet [16], Oriented R-CNN [55], and DCL [58] bring satisfied detection results on DOTA dataset [53] and HRSC2016 dataset [29]. Gliding Vertex [56] and RSDet [40] share the same purpose of considering quadrilateral regression prediction. Note that, the boundary discontinuity problem is caused by the periodicity of the", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "angle when dealing with the angel-based orientation estimation. Then, Yang et al. [60] take a new perspective that turns the angular regression tasks into the angular classification tasks [58]. Moreover, Oriented RepPoints [23] utilize the flexible adaptive points as the box representation to achieve oriented object detection. Unlike these methods, our CrackDet addresses the boundary discontinuity problem via the new piecewise angle definition, which is motivated by the divide-and-conquer strategy. Besides, although KLD [65] and GWD [63] also start from a distribution perspective, it converts the center point, height, width, and orientation of a rotated bounding box as a 2-D Gaussian distribution, and does not focus on learning angle variance. Instead, the proposed CrackDet assumes that the ground-truth orientation obeys a single-variate gaussian distribution, and aims to learn the sub-crack orientation and its variance together.", + "bbox": [ + 496, + 90, + 890, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our method and datasets", + "text_level": 1, + "bbox": [ + 498, + 345, + 736, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework is shown in Fig. 2. We first introduce the piecewise angle definition in Sec. 3.1, and the network architecture of CrackDet in Sec. 3.2. Then, we give the details of the proposed MAR Loss for estimating ambiguities in sub-crack orientation in Sec. 3.3. Finally, we present the inference step of CrackDet with estimated angle variances in Sec. 3.4, and the proposed datasets in Sec. 3.5.", + "bbox": [ + 496, + 369, + 890, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Piecewise angle definition", + "text_level": 1, + "bbox": [ + 498, + 484, + 732, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we recap the boundary discontinuity problem and inconsistency regression problem in the existing 180-degree five-parameter definition. Then, we discuss the details of the proposed piecewise angle definition.", + "bbox": [ + 496, + 508, + 890, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rethinking the five-parameter definition. Fig. 2 (a) shows that the existing five-parameter definition [53] with the 180-degree angular range includes $(x, y, h, w, \\theta)$ , where $(x, y)$ indicates the center point and $[h, w, \\theta]$ denotes height, width, and angle, respectively. Then, the boundary discontinuity problem refers to a sharp increase of loss at the boundary caused by the periodicity of the angle. As shown in Fig. 3 (a), it can be further explained that when the angle of a rotated box is close to $180^{\\circ}$ , the angle under the five-parameter definition will suddenly change from $180^{\\circ}$ to $0^{\\circ}$ , resulting a sharp increase of angle regression loss. Moreover, the inconsistency regression problem can be attributed to the rotation symmetry in object detection, which includes the equivariant orientation and invariant shape estimation for a rotated bounding box [68]. Specifically speaking, rotation-equivariant features are desired for orientation estimation. In contrast, the shape of the bounding box does not change after a rotation. Thus, rotation-invariant features are needed for predicting its invariant height and width [16]. Discussion on our piecewise angle definition. We are inspired by the divide-and-conquer strategy [7] and propose the following piecewise angle definition, shown in", + "bbox": [ + 496, + 569, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6655", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f78265a71a5b0e24a85cc2a82660ea83321cd092bfb6563b85b68983066bcad5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 85, + 261, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f1c1fc28a8475903ae0cf41fc6f4a5ff7337a4b1b39025cc7c7dab6598b69616.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 266, + 88, + 875, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7819f9a0c288ed74cbd4b303221544d3c03bdd41a301475d5f26b358f4b397e3.jpg", + "image_caption": [ + "(a) The five-parameter definition", + "(b) The proposed piecewise angle definition", + "(c) Overall architecture of the proposed CrackDet", + "Figure 2. (a) The existing five-parameter definition with long-side definition. (b) Our piecewise angle definition is inspired by the divide-and-conquer strategy and built upon the above five-parameter definition. It breaks down the existing 180-degree regression tasks into four sub-tasks for addressing the discontinuity problem. (c) Based on our piecewise angle definition, we present an overview of CrackDet that can estimate standard deviations (std) along with angle regression for capturing ambiguities in the sub-crack orientation. Note that, the heatmap branch and offset branch are responsible for the center (i.e., $x$ and $y$ ) of the predicted bounding box as in the original CenterNet." + ], + "image_footnote": [], + "bbox": [ + 93, + 258, + 861, + 411 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/47a5c9a2891c667b5976e9ce2a1dc021e9b6f8418d068c3c3caa1686a2b11413.jpg", + "image_caption": [ + "(a) The 180-degree definition", + "(b) Our piecewise definition", + "Figure 3. The horizontal axis denotes the angle of a rotated box in anti-clockwise order. And the vertical axis denotes the angle definition in different methods. Due to the periodicity of the angle, Figure (a) shows that there is a sudden change when the angle of a rotated box closes to $180^{\\circ}$ with the 180-degree definition.", + "Fig. 2 (b). On the one hand, Fig. 3 (b) shows that there is no sudden change of the angle in our piecewise angle definition. Thus, ours is free of the boundary discontinuity problem. On the other hand, rotation equivariance can be easily achieved by group convolutions [52], while rotation invariance relies on a larger capacity network or a larger number of training samples. Hence, we focus on rotation-equivariant features. And then, we are allowed to transform the invariant size estimation into equivariant size estimation by adding a factor $\\sin (\\theta)$ or $\\cos (\\theta)$ to the height and width (see Fig. 2 (b)). That way, the box size is enforced to change equivalently when the box orientation changes. Ac" + ], + "image_footnote": [], + "bbox": [ + 88, + 535, + 454, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cordingly, both box size and angle regressions can benefit from rotation-equivariant features.", + "bbox": [ + 498, + 527, + 890, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Network architecture", + "text_level": 1, + "bbox": [ + 500, + 571, + 700, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As discussed before, we aim to extract rotation-equivariant features for box size regression and angle regression. Therefore, we adopt rotation-equivariant networks as the backbone to extract rotation-equivariant features. Then, we design a parallel multi-branch architecture based on the proposed piecewise angle definition. The overall architecture of CrackDet is shown in Fig. 2 (c).", + "bbox": [ + 496, + 597, + 890, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rotation-equivariant backbone. We first adopt CenterNet [73] as our baseline, which takes an object as a single point (i.e., the center point of the bounding box) and regresses the object size and offset. Then, we re-implement all layers of the fully-convolutional encoder-decoder networks (i.e., up-convolutional residual networks [20]) in CenterNet based on e2cnn [52], named as ReEDNet. Importantly, e2cnn includes rotation-equivariant convolution, rotation-equivariant up-convolution, and pooling, etc. Thanks to the rotation weight sharing and group representations in e2cnn, our ReEDNet takes smaller parameters than the original encoder-decoder networks in CenterNet and enjoys the capability of equivariance.", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6656", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/43798c108a9d6fc78b12f450717250e477abbc44e0ec7400827ef807a167d362.jpg", + "image_caption": [ + "(a) Branch 1", + "Figure 4. A visualization of our angle-based variance voting. Firstly, the red, blue, orange, and teal textboxes are the corresponding standard deviation obtained from branch 1, 2, 3, and 4, respectively. Then, for each oriented bounding box, we select brach $i$ that has the smallest standard deviation to obtain its final detection result. Note that, the red background box is associated with the selected branch." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 240, + 205 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6baa85e4c3754dac30c8e401d3cb750f626808d80587c6cbf7ba8985ac0c6e5a.jpg", + "image_caption": [ + "(b) Branch 2" + ], + "image_footnote": [], + "bbox": [ + 240, + 88, + 545, + 205 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a98286860ac998d5d81dffa88fbd579c4b86c8b3eff7a8e3e38e74220b875f71.jpg", + "image_caption": [ + "(d) Branch 4", + "(e) After variance voting" + ], + "image_footnote": [], + "bbox": [ + 545, + 88, + 883, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-branch architecture. According to our piecewise angle definition, we define a valid range $[l_i, r_i)$ for each branch $i$ . During the training, we only select the ground truth boxes whose angles fall in the corresponding valid range of one branch. Specifically, referring to the existing 180-degree definition, for an oriented bounding box with angle $\\theta$ on the input image, it is valid for branch $i$ when:", + "bbox": [ + 75, + 287, + 470, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {i} \\leq \\theta < r _ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 402, + 468, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2 (c), the valid ranges of four branches are set to $[0^{\\circ}, 45^{\\circ})$ , $[45^{\\circ}, 90^{\\circ})$ , $[90^{\\circ}, 135^{\\circ})$ and $[135^{\\circ}, 180^{\\circ})$ , respectively. For branch $i$ , we redefine the ground-truth $\\theta$ , height $h$ , and width $w$ in the 180-degree definition into $\\theta_{i}$ , $h_{i}$ , and $w_{i}$ , according to our piecewise definition in Fig. 2 (b). Moreover, we aim to estimate the angle and angle confidence together. To this end, each branch of our parallel multi-branch architecture contains three detection heads on the top of the rotation-equivariant backbone for predicting size (i.e., height and width), angle, and angle confidence. Note that, we simply implement each head with a fully-connected layer. That way, our network enables us to predict multiple probability distributions instead of only one angle. For simplicity, we assume each predicted angle from each branch is independent and obeys a single-variate gaussian distribution. The corresponding equation is shown as follows:", + "bbox": [ + 75, + 426, + 468, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\Theta_ {i}} (\\theta) = \\frac {1}{\\sqrt {2 \\pi \\sigma_ {i} ^ {2}}} e ^ {- \\frac {\\left(\\theta - \\theta_ {i} ^ {e}\\right) ^ {2}}{2 \\sigma_ {i} ^ {2}}}, i \\in \\{1, 2, 3, 4 \\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 686, + 468, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Theta_{i}$ is the parameters of the angle detection head from branch $i$ , $\\theta_{i}^{e}$ denotes the estimated angle from branch $i$ , and standard deviation $\\sigma_{i}$ measures the angle confidence of the estimation on branch $i$ . When $\\sigma_{i}$ is very close to 0, it means the branch $i$ is extremely confident about the estimated angle. Accordingly, we suppose to have the redefined ground-truth angle $\\theta_{i}$ that is in the valid range of branch $i$ . Similarly, $\\theta_{i}$ can also be formulated as a gaussian distribution $\\mathcal{N}(\\theta_i,\\sigma_{gt}^2)$ with its standard deviation $\\sigma_{gt}\\to 0$ . Then, this gaussian distribution can be viewed as: $P_{i}^{gt}(\\theta) = \\delta (\\theta -\\theta_{i})$ , where $\\delta (\\cdot)$ indicates Dirac delta function.", + "bbox": [ + 75, + 731, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. The proposed MAR Loss", + "text_level": 1, + "bbox": [ + 500, + 287, + 728, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Suppose the angle of a ground-truth bounding box lies in the valid range of branch $i$ , our multi-branch angle regression loss (MAR Loss) aims to minimize the distribution discrepancy between the predicted angle distribution from branch $i$ and the redefined ground-truth angle distribution. Meanwhile, we expect the other three branches to be able to predict larger variance so that we can filter those inaccurate predictions by variance voting in the inference step. More specifically, we first exploit the Wasserstein distance as the distance metric and minimize the distance between $P_{\\Theta_i}(\\theta)$ and $P_i^{gt}(\\theta)$ . Secondly, we maximize the estimated variance from the other three branches. The equation is shown as:", + "bbox": [ + 496, + 311, + 890, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {M A R} = \\frac {D _ {W} \\left(P _ {\\Theta_ {i}} (\\theta) \\| P _ {i} ^ {g t} (\\theta)\\right)}{\\frac {1}{2} + \\sigma_ {i} ^ {2}} - \\sum_ {j = 1, j \\neq i} ^ {4} \\sigma_ {j} ^ {2} \\tag {3} \\\\ = \\frac {\\| \\theta_ {i} ^ {e} - \\theta_ {i} \\| _ {2} ^ {2} + \\sigma_ {i} ^ {2}}{\\frac {1}{2} + \\sigma_ {i} ^ {2}} - \\sum_ {j = 1, j \\neq i} ^ {4} \\sigma_ {j} ^ {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 497, + 890, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D_W(\\cdot)$ is unfolded based on Wasserstein distance [3]:", + "bbox": [ + 500, + 592, + 890, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {W} \\left(P _ {\\Theta_ {i}} (\\theta) \\| P _ {i} ^ {g t} (\\theta)\\right) = \\| \\theta_ {i} ^ {e} - \\theta_ {i} \\| _ {2} ^ {2} + \\sigma_ {i} ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 636, + 890, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Importantly, we exploit Wasserstein distance instead of KL-Divergence [18] for angle regression, because KL-Divergence heavily relies on a non-negligible intersection between two distributions [1, 18]. Note that, when $\\theta_i^e$ is estimated accurately, i.e., $\\| \\theta_i^e -\\theta_i\\| _2^2\\to 0$ , the branch $i$ is expected to predict smaller variance. In addition, a term $\\frac{1}{2} +\\sigma_i^2$ is added for the following reason: due to the ambiguities in sub-crack orientation, when the angle $\\theta_{i}^{e}$ is estimated inaccurately, i.e., $\\| \\theta_i^e -\\theta_i\\| _2^2 >\\frac{1}{2}$ , we expect the branch $i$ to be able to predict larger variance $\\sigma_i^2$ so that $L_{MAR}$ will be lower. Moreover, according to our piecewise angle definition, the size regression loss $L_{\\mathrm{size}}$ for the redefined ground truth height $h_i$ and width $w_{i}$ is defined as $L_{\\mathrm{size}} = |h_i^e -h_i| + |w_i^e -w_i|$ where $h_i^e$ and $w_{i}^{e}$ indicate the estimated height and width from the valid branch $i$ . Then,", + "bbox": [ + 496, + 671, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6657", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the overall training objective is shown below:", + "bbox": [ + 76, + 90, + 379, + 106 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {C r a c k D e t}} = L _ {k} + \\lambda_ {o f f} L _ {o f f} + \\lambda_ {s i z e} L _ {s i z e} + \\lambda_ {M A R} L _ {M A R} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 116, + 468, + 145 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{k}$ and $L_{off}$ are the losses of center point recognition and offset regression by following CenterNet [73]. The hyper-parameter $\\lambda_{off}$ , $\\lambda_{size}$ , and $\\lambda_{MAR}$ are constant factors to control the balance of the above losses.", + "bbox": [ + 76, + 147, + 468, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Inference", + "text_level": 1, + "bbox": [ + 76, + 215, + 186, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the inference step, there are four branches, each of which will predict the height, width, angle, and standard deviation. Thus, we can not directly decide which one should be adopted. Fortunately, with the help of MAR Loss, we design an angle-based variance voting for obtaining the final bounding boxes. Firstly, we extract the peaks in the heatmap by following CenterNet [73]. Then, we get standard deviations $\\sigma_{1},\\sigma_{2},\\sigma_{3},\\sigma_{4}$ from four branches at each peak. As shown in Fig. 4, for each peak, we choose branch $i$ that has the minimum estimated variance to obtain the raw output: $i^{*} = \\mathrm{argmin}_{i = \\{1,2,3,4\\}}\\sigma_{i}^{2}$ . Next, according to CenterNet [73], we obtain the center (i.e., $x$ and $y$ ) of the oriented bounding box according to the above peaks and the offset branch. Finally, we acquire the final height $h_o = h_{i^*}^e\\backslash \\Delta (\\theta_{i^*}^e)$ , width $w_{o} = w_{i^{*}}^{e}\\backslash \\Delta (\\theta_{i^{*}}^{e})$ , and angle $\\theta_o = \\Gamma (\\theta_{i^*}^e)$ of the bounding box by the following equation:", + "bbox": [ + 76, + 239, + 468, + 492 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma \\left(\\theta_ {i} ^ {e}\\right) \\triangleq \\left\\{ \\begin{array}{l l} \\theta_ {i} ^ {e}, & i = 1, \\\\ 9 0 ^ {\\circ} - \\theta_ {i} ^ {e}, & i = 2, \\\\ 9 0 ^ {\\circ} + \\theta_ {i} ^ {e}, & i = 3, \\\\ 1 8 0 ^ {\\circ} - \\theta_ {i} ^ {e}, & i = 4. \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 492, + 468, + 555 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\left(\\theta_ {i} ^ {e}\\right) \\triangleq \\left\\{ \\begin{array}{l l} \\cos \\left(\\theta_ {i} ^ {e}\\right), & i = 1, \\\\ \\sin \\left(9 0 ^ {\\circ} - \\theta_ {i} ^ {e}\\right), & i = 2, \\\\ \\sin \\left(9 0 ^ {\\circ} + \\theta_ {i} ^ {e}\\right), & i = 3, \\\\ - \\cos \\left(1 8 0 ^ {\\circ} - \\theta_ {i} ^ {e}\\right), & i = 4. \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 560, + 468, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $i$ denotes the index of the selected branch. $\\theta_{i}^{e}$ indicate the estimated angle from branch $i$ . It is worth noting that the remaining inference step is the same as CenterNet [73].", + "bbox": [ + 76, + 627, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. The proposed datasets", + "text_level": 1, + "bbox": [ + 76, + 681, + 284, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To promote the research on oriented sub-crack detection, we propose ONPP, ORC, and OCCSD datasets, which are collected from industrial facilities, road pavement, and various buildings in real-world applications.", + "bbox": [ + 76, + 704, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ONPP: This dataset is collected from the industrial facilities in real-world applications by a high-resolution camera, comprising 200 images with the size of $7360 \\times 4912$ . The width of the crack in the collected images varies from 0.05 mm to $10\\mathrm{mm}$ . Moreover, the collected images contain different kinds of noise, such as various concrete types and light intensity. To extend this dataset without compromising the resolution, we directly slice these images into $512 \\times 512$ pixels, constructing a final dataset with 3,104 samples.", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ORC: This dataset is collected from road pavement by a mobile vehicle equipped with a high-resolution camera. It contains 300 images with a size of $3489 \\times 3489$ . The images of this dataset include various sizes of cracks in the road. It also contains a lot of noise, such as well covers and various appearances of the crosswalk. Similar to ONPP, we slice these images into $512 \\times 512$ pixels, constructing a final road pavement dataset with 1,303 samples.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OCCSD: This dataset is extended from the concrete crack segmentation dataset [71], which contains 458 high-resolution images ( $4032 \\times 3024$ pixels). We construct an oriented sub-crack detection dataset with 1,875 samples by slicing these images into $512 \\times 512$ pixels and relabeling them from an oriented sub-crack detection perspective.", + "bbox": [ + 496, + 213, + 890, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 320, + 630, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment details", + "text_level": 1, + "bbox": [ + 500, + 345, + 681, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation testbed. We conduct experiments for oriented sub-crack detection on three datasets, i.e., ONPP, ORC, and OCCSD. By following the existing crack detection work [8, 31], we divide each dataset into the training, validation, and test set, with the proportion of $8:1:1$ . Moreover, to further verify the effectiveness of CrackDet, we also provide the experiment results on the HRSC2016 dataset [29]. HRSC2016 dataset contains 1,061 samples. Its resolution ranges from $300 \\times 300$ to $1500 \\times 900$ . For fairness, according to previous works [55, 16, 23], we exploit the training set (436 images) and validation set (181 images) for training, and the testing set (444 images) for evaluation.", + "bbox": [ + 496, + 369, + 890, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We adopt Adam as the optimizer for training. The total epoch for training the CrackDet is set as 60, 60, 60, and 140 on ONPP, ORC, OCCSD, and HRSC2016 datasets, respectively. For the first three datasets, the initial learning rate is set to $4e - 4$ and reduced by a factor of 10 after 20, and 40 epochs. For the last one, the initial learning rate is $4e - 4$ and reduced by a factor of 10 after 90, and 120 epochs by following the previous work [39]. The batch size for ONPP, ORC, OCCSD, and HRSC2016 is set to 32, 32, 32, and 8, respectively. More importantly, the input resolutions of four datasets are $512 \\times 512$ , $512 \\times 512$ , $512 \\times 512$ , and $768 \\times 768$ , respectively. By following the previous work [16], we only use random flipping for data augmentation. We implement our ReEDNet based on ResNet50 and e2cnn [52], and denote it as ReED-R-50. We train all models on four RTX 3090 GPUs and test on a single RTX 3090 GPU. According to our hyper-parameter sensitivity study, we set $\\lambda_{off} = 0.1$ , $\\lambda_{size} = 0.2$ , and $\\lambda_{MAR} = 0.1$ in our next experiments.", + "bbox": [ + 496, + 551, + 890, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. Following previous works [75, 54], we evaluate the performance of CrackDet on oriented crack detection, in terms of Precision and Recall. Moreover, we adopt the mean orientation error (MOE) to verify the effec", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6658", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/53174c2de169aff004830959668837e160c4b535a157c8be2ad1f297187b53f7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodONPPORC
Precision ↑Recall ↑MOE ↓Precision ↑Recall ↑MOE ↓
FCrack-O2020 [13]0.70370.82630.27850.66310.87140.3397
ReCrack-O2021 [50]0.75060.67190.39870.74320.82540.3282
CABF-FCOS-O2021 [69]0.78930.84130.32630.76300.88510.3744
Y3Crack-O2022[31]0.71340.73790.33500.71030.85140.3632
CenterCrack-O2022 [51]0.72260.61320.35740.77310.81020.3085
Yolo-ViT-O2022 [54]0.77660.83790.35030.79820.84610.2978
CrackDet0.82040.91060.16800.85780.89260.1166
", + "bbox": [ + 151, + 87, + 816, + 233 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Comparison with the state-of-the-art crack detectors on ONPP and ORC. MOE denotes the mean orientation error.", + "bbox": [ + 120, + 233, + 844, + 246 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tiveness of the proposed method for estimating crack orientation. For the detection accuracy on the HRSC2016, we adopt the mean average precision (mAP) as an evaluation criterion, which is consistent with VOC2007 metrics.", + "bbox": [ + 75, + 271, + 468, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with crack detection approaches", + "text_level": 1, + "bbox": [ + 76, + 380, + 460, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, our method is compared to several baselines including state-of-the-art robust crack detection methods: FCrack uses Faster R-CNN to detect the crack patch [13], $\\mathrm{Y}^{3}\\mathrm{Crack}$ designs a intelligent system based on YOLO v3 for detecting crack [31], CenterCrack proposes a crack detection model by using CenterNet [51], ReCrack utilizes the RetinaNet to find the crack in asphalt pavement [50], CABF-FCOS proposes a one-stage network that can detect various defects including crack [69]. Yolo-ViT applies the Transformer to YOLO v5 for crack detection [54]. Note that, all of these methods can not detect the orientation of cracks. For a fair comparison, we add a branch to these methods for the regression of crack orientation and denote them as FCrack-O, $\\mathrm{Y}^{3}\\mathrm{Crack - O}$ , CenterCrack-O, ReCrack-O, CABF-FCOS-O, and Yolo-ViT-O respectively. All experiments are independently carried out on ONPP and ORC datasets. Table 1 provides the performance of various approaches from our oriented sub-crack detection perspective for crack detection. Bold denotes the best results. It can be found that the proposed method compares favorably to other competitive crack detection approaches. For example, compared with the Yolo-ViT-O, the precision and recall of CrackDet are 0.8204 and 0.9106 while the Yolo-ViT-O only achieves 0.7766 and 0.8379 on ONPP dataset. For the ORC dataset, CrackDet brings performance gains over Yolo-ViT-O around $6.0\\%$ and $4.7\\%$ under the precision and recall metrics. Moreover, in terms of MOE metric, the proposed CrackDet achieves consistent improvements on all datasets. The results above together show the necessity that the boundary discontinuity problem and the ambiguities in orientation should be treated properly under the oriented sub-crack detection tasks.", + "bbox": [ + 75, + 417, + 472, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Discussion on the segmentation-based methods", + "text_level": 1, + "bbox": [ + 498, + 271, + 890, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, considering the detection-then-segmentation strategy in the real-world application [33, 35], we further verify that the proposed method can serve as a complement to segmentation-based methods. More specifically, for the pixel-level segmentation task, we replace the oriented box annotation with pixel-level annotation for the OCCSD dataset. Table 3 shows that the proposed method can improve the segmentation results, especially for the segmentation model with limited performance. The reason attributes that the proposed method can provide a more fine-grained localization, which can effectively promote the segmentation model being influenced by the noise background.", + "bbox": [ + 496, + 294, + 890, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison with oriented detection methods", + "text_level": 1, + "bbox": [ + 498, + 500, + 880, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further validate our method, we conduct a series of experiments on both crack datasets and HRSC2016, and report quantitative results to verify the effectiveness of CrackDet. As CrackDet shows the purpose of detecting the oriented object, we compare it with existing state-of-the-art oriented detectors with the help of [74], including Oriented RepPoints [23], Oriented R-CNN [55], SASM [19], etc.", + "bbox": [ + 496, + 523, + 890, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on oriented sub-crack detection. Table 2 provides the comparison results over the existing powerful oriented object detectors. We can observe that, first, our CrackDet outperforms the existing detectors by a large margin in terms of Precision. The reason is that the proposed piecewise angle definition can promote the proposed detector to take advantage of rotation-equivariant features and be free of the boundary discontinuity problem. Secondly, CrackDet yields significant and consistent improvement on all three datasets in terms of MOE metric. The reason can be explained as CrackDet can estimate the angle confidence with the help of the multi-branch architecture and MAR Loss. Thus, according to the corresponding angle confidence from each branch, the angle-based variance voting improves the angle estimation results.", + "bbox": [ + 496, + 643, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on HRSC2016. For fairness, we follow the same settings in Oriented RepPoints [23], and report the results", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6659", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/916d7f96728ac977e3ec6636996257302ec1e15c82ffeb9741371cb3f7e140a9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodONPPORC
Precision ↑Recall ↑MOE ↓Precision ↑Recall ↑MOE ↓
\\(CSL_{ECCV'20}\\) [60]0.73090.81460.26330.79430.74370.3285
\\(R^3\\text{Det}_{AAAI'21}\\) [62]0.78350.74330.29770.76970.73120.4280
\\(S^2ANet_{TGRS'21}\\) [15]0.77730.77730.41330.70680.78680.4640
\\(ReDet_{CVPR'21}\\) [16]0.78440.80590.27430.71910.86840.3335
Beyond Bounding-Box \\(CVPR'21\\) [14]0.77900.84680.30030.60760.90390.3441
Oriented R-CNN\\(ICCV'21\\) [55]0.80320.73550.21720.70870.74160.3470
\\(GWD_{ICML'21}\\) [63]0.71690.78840.44010.75760.86320.3014
\\(KLD_{NeurIPS'21}\\) [65]0.74660.64070.34460.74240.90760.2818
\\(SASM_{AAAI'22}\\) [19]0.74160.83030.30620.76530.90380.3489
Oriented RepPoints \\(CVPR'22\\) [23]0.78800.87820.30080.78510.91010.3497
CrackDet0.82040.91060.16800.85780.89260.1166
", + "bbox": [ + 116, + 88, + 849, + 295 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c38e7c8391ee00ba1e8a5408a6b33a2db948630f52c7cf3ac76f91ea3e66cf5d.jpg", + "table_caption": [ + "Table 2. We provide a detailed comparison over general oriented detection approaches on ONPP and ORC." + ], + "table_footnote": [], + "table_body": "
Precision ↑Crackformer ICCV'21 [27]JTFN ICCV'21 [10]LIOT TIP'22 [45]
w/o CrackDet0.7810.7540.803
w/ CrackDet0.8140.7980.819
", + "bbox": [ + 81, + 329, + 470, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "under the VOC2007 metric. Table 4 shows the comparison results. The proposed CrackDet achieves the best performance, even compared with strong FPN-based baselines. Such results further validate that the proposed CrackDet has the potential in more oriented object detection tasks.", + "bbox": [ + 75, + 412, + 467, + 488 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3540a8799e5df1ec3397b5989c60fa1f9bc1a5c47540eade57c0f86f4b9eb76a.jpg", + "table_caption": [ + "Table 3. Segmentation performance with CrackDet on OCCSD." + ], + "table_footnote": [], + "table_body": "
MethodsBackbonemAP50(07)↑
R3Det AAAI'21 [62]R-101-FPN89.26
R3Det-DCL CVPR'21 [58]R-101-FPN89.46
FPN-CSLECCV'20 [60]R-101-FPN89.62
ReDet CVPR'21 [16]Re-R-5090.46
DAL AAAI'21 [32]R-101-FPN89.77
S2A-NetTGRS'21 [15]R-101-FPN90.17
S2A-Net-DHRec PAMI'22 [37]R-101-FPN90.22
Oriented R-CNN ICCV'21 [55]R-50-FPN90.40
Oriented RepPoints CVPR'22 [23]R-50-FPN90.38
CrackDetReED-R-5090.60
", + "bbox": [ + 81, + 501, + 467, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 709, + 225, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To examine the contribution of each component in our proposed detector: rotation-equivariant network, the number of valid ranges in the proposed piecewise angle definition, and MAR Loss, a series of ablation experiments are performed on ONPP and ORC datasets.", + "bbox": [ + 75, + 733, + 467, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on the rotation-equivariant network. As expected, CrackDet achieves better results by combining the rotation-equivariant network. Table 5 shows that it improves the precision, recall, and MOE by $0.91\\%$ , $1.5\\%$ , and $0.023$ with the help of the rotation-equivariant network, respectively. The reason behind this effect is that rotation-", + "bbox": [ + 75, + 810, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "equivariant features extracted by the rotation-equivariant network are beneficial to size and angle regression within the proposed piecewise angle definition.", + "bbox": [ + 498, + 333, + 890, + 378 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a951cd280356b5878ddc49e34ba3ba5cd91b61e01c1e50479d80ddb13c8c9b0b.jpg", + "table_caption": [ + "Table 4. We report results on the HRSC2016 test set. mAP(07) denotes detection results under VOC2007 mAP metrics." + ], + "table_footnote": [], + "table_body": "
MethodsONPP
Precision ↑Recall ↑MOE ↓
w/o Rotation-equivariant network [17]0.81130.89520.1914
w/ Rotation-equivariant network0.82040.91060.1680
", + "bbox": [ + 503, + 390, + 890, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on the number of valid angle ranges. We study the effect of the number of valid angle ranges in our piecewise angle definition. For fairness, we do not add the factor $\\sin(\\theta)$ or $\\cos(\\theta)$ for all settings. Table 6 shows the results using one to five ranges. The corresponding results demonstrate that our four ranges definition improves over the single range (baseline) with a 0.14 decrease. As can be noticed, the five ranges do not bring further improvement over the four ranges. By considering the complexity and performance, we choose four ranges as the default setting.", + "bbox": [ + 498, + 487, + 890, + 638 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/824c8efc3b8abfb9cf485ffa98635fbd88d282d4fb362d6e80b6df5b8e8705b5.jpg", + "table_caption": [ + "Table 5. Performance comparisons of with- and without rotation-equivariant backbone on ONPP dataset." + ], + "table_footnote": [], + "table_body": "
NumValid Angle RangesMOE ↓
1[0°, 180°)0.33
2[0°, 90°), [90°, 180°)0.30
3[0°, 60°), [60°, 120°), [120°, 180°)0.21
4[0°, 45°), [45°, 90°), [90°, 135°), [135°, 180°)0.19
5[0°, 36°), [36°, 72°), [72°, 108°), [108°, 144°), [144°, 180°)0.22
", + "bbox": [ + 503, + 651, + 883, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Results from the different number of ranges on ONPP.", + "bbox": [ + 500, + 720, + 872, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on MAR loss. It is to be expected that Crack-Det combined with MAR Loss achieves consistent improvements on ONPP and ORC datasets, shown in Table 7. The reason is three-fold: (1) KL Loss may fail when two distributions do not have a non-negligible intersection. (2) By learning to estimate high variances for the ambiguous crack orientation during training, our model is able to learn confidence in predicting sub-crack orientation. (3) MAR Loss incorporates learning angle confidence which can potentially help the network to learn discriminative features. The", + "bbox": [ + 498, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6660", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "learned variances through our MAR Loss are interpretable (see Fig. 4). Our network will output higher variances for ambiguous crack orientation, which can be useful in judging the overall orientation in real-world applications [38].", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/7a7a73e6bf306c8a7188de8368db1ef9c9e2170e68c6d767edb650c119aa752f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsONPPORC
Precision ↑Recall ↑MOE ↓Precision ↑Recall ↑MOE ↓
KL Loss [18]0.78650.87180.24250.81490.85370.1986
MAR Loss0.82040.91060.16800.85780.89260.1166
", + "bbox": [ + 81, + 164, + 467, + 208 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 7. Performance comparisons of MAR Loss and KL Loss.", + "bbox": [ + 76, + 210, + 450, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 256, + 194, + 272 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, by starting from a new perspective for crack detection, we propose CrackDet, a first-of-its-kind oriented sub-crack detector that overcomes the boundary discontinuity problem. Meanwhile, we propose the MAR loss to capture the ambiguities in crack orientation by combining CrackDet. The experiments demonstrate that our method outperforms state-of-the-art crack detectors and oriented detectors. On one hand, as we also construct three datasets for oriented sub-crack detection, this work should also give a direction to researchers for designing the crack detector from the oriented sub-crack detection perspective. On the other hand, it highlights the key drawback of existing perspectives of crack detection (i.e., a high labeling cost of pixel-level segmentation, and a coarse crack localization of patch-level classification and object-level detection). We believe this paper has the potential to inspire the following works: (1) multi-model learning [24, 25] for boosting the detection performance in a novel field with limited target data [21], (2) semi-supervised learning and active learning [9] to further alleviate its labeling cost in real-world applications, and (3) oriented sub-crack detection via horizontal annotations [66].", + "bbox": [ + 75, + 282, + 470, + 613 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported by the National Nature Science Foundation of China under Grants 62073225,62203134, 61972263, 62072315, the National Key RD Program of China under Grants 2020YFA0908700, the Natural Science Foundation of Guangdong Province Outstanding Youth Program under Grants 2019B151502018, Shenzhen Science and Technology Innovation Commission (20220809141216003, JCYJ20210324093808021, JCYJ20220531102817040), the Guangdong \"Pearl River Talent Recruitment Program\" under Grant 2019ZT08X603, the Guangdong \"Pearl River Talent Plan\" under Grant 2019JC01X235.", + "bbox": [ + 75, + 636, + 468, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 832, + 173, + 848 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Proc. ICML, pages 214-223, 2017. 2, 5", + "bbox": [ + 84, + 858, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[2] Seyed Majid Azimi, Eleonora Vig, Reza Bahmanyar, Marco Körner, and Peter Reinartz. Towards multi-class object detection in unconstrained remote sensing imagery. In Proc. ACCV, pages 150-165, 2018. 3", + "[3] Raktim Bhattacharya. Data assimilation in optimal transport framework. arXiv preprint arXiv:2005.08670, 2020. 5", + "[4] Fu-Chen Chen and Mohammad R Jahanshahi. Nb-cnn: Deep learning-based crack detection using convolutional neural network and naive bayes data fusion. IEEE Transactions on Industrial Electronics, 65(5):4392-4400, 2017. 1, 2", + "[5] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proc. ECCV, pages 801-818, 2018. 3", + "[6] Minghao Chen, Hongyang Xue, and Deng Cai. Domain adaptation for semantic segmentation with maximum squares loss. In Proc. ICCV, pages 2090-2099, 2019. 3", + "[7] Qiang Chen, Yingming Wang, Tong Yang, Xiangyu Zhang, Jian Cheng, and Jian Sun. You only look one-level feature. In Proc. CVPR, pages 13039-13048, 2021. 3", + "[8] Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Jie Chen, Zun Liu, and Jianqiang Li. Geometry-aware guided loss for deep crack recognition. In Proc. CVPR, pages 4703-4712, 2022. 1, 2, 6", + "[9] Zhuangzhuang Chen, Jin Zhang, Pan Wang, Jie Chen, and Jianqiang Li. When active learning meets implicit semantic data augmentation. In Proc. ECCV, pages 56-72, 2022. 9", + "[10] Mingfei Cheng, Kaili Zhao, Xuhong Guo, Yajing Xu, and Jun Guo. Joint topology-preserving and feature-refinement network for curvilinear structure segmentation. In Proc. ICCV, pages 7147-7156, 2021. 8", + "[11] Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning roi transformer for oriented object detection in aerial images. In Proc. CVPR, pages 2849-2858, 2019. 3", + "[12] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proc. ICCV, pages 6569-6578, 2019. 2", + "[13] Fen Fang, Liyuan Li, Ying Gu, Hongyuan Zhu, and Joo-Hwee Lim. A novel hybrid approach for crack detection. PR, 107:107474, 2020. 2, 7", + "[14] Zonghao Guo, Chang Liu, Xiaosong Zhang, Jianbin Jiao, Xi-angyang Ji, and Qixiang Ye. Beyond bounding-box: Convex-hull feature adaptation for oriented and densely packed object detection. In Proc. CVPR, 2021. 8", + "[15] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2021. 8", + "[16] Jiaming Han, Jian Ding, Nan Xue, and Gui-Song Xia. Redet: A rotation-equivariant detector for aerial object detection. In Proc. CVPR, pages 2786-2795, 2021. 2, 3, 6, 8", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In Proc. ECCV, pages 630-645, 2016. 2, 8", + "[18] Yihui He, Chenchen Zhu, Jianren Wang, Marios Savvides, and Xiangyu Zhang. Bounding box regression with uncertainty for accurate object detection. In Proc. CVPR, pages 2888-2897, 2019. 5, 9" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6661", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Liping Hou, Ke Lu, Jian Xue, and Yuqiu Li. Shape-adaptive selection and measurement for oriented object detection. In Proc. AAAI, 2022. 7, 8", + "[20] Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, et al. Speed/accuracy trade-offs for modern convolutional object detectors. In Proc. CVPR, pages 7310-7311, 2017. 4", + "[21] Jianqiang Li, Zhuangzhuang Chen, Jie Chen, and Qiuzhen Lin. Diversity-sensitive generative adversarial network for terrain mapping under limited human intervention. IEEE Transactions on Cybernetics, 51(12):6029-6040, 2020. 9", + "[22] Kai Li, Bo Wang, Yingjie Tian, and Zhiquan Qi. Fast and accurate road crack detection based on adaptive cost-sensitive loss function. IEEE Transactions on Cybernetics, 2021. 1, 3", + "[23] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented leptopoints for aerial object detection. In Proc. CVPR, pages 1829-1838, 2022. 2, 3, 6, 7, 8", + "[24] Wenrui Li and Xiaopeng Fan. Image-text alignment and retrieval using light-weight transformer. In Proc. ICASSP, pages 4758-4762, 2022. 9", + "[25] Wenrui Li, Zhengyu Ma, Liang-Jian Deng, Xiaopeng Fan, and Yonghong Tian. Neuron-based spiking transmission and reasoning network for robust image-text retrieval. IEEE TCSVT, 2022. 9", + "[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dálár. Focal loss for dense object detection. In Proc. ICCV, pages 2980-2988, 2017. 2", + "[27] Huajun Liu, Xiangyu Miao, Christoph Mertz, Chengzhong Xu, and Hui Kong. Crackformer: Transformer network for fine-grained crack detection. In Proc. ICCV, pages 3783-3792, 2021. 8", + "[28] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In Proc. ECCV, pages 21-37, 2016. 2", + "[29] Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang. A high resolution optical satellite image dataset for ship recognition and some new baselines. In International conference on pattern recognition applications and methods, volume 2, pages 324-331, 2017. 3, 6", + "[30] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proc. CVPR, pages 3431-3440, 2015. 3", + "[31] Duo Ma, Hongyuan Fang, Niannian Wang, Chao Zhang, Jiaxiu Dong, and Haobang Hu. Automatic detection and counting system for pavement cracks based on pcgan and yolo-mf. IEEE Transactions on Intelligent Transportation Systems, 2022. 1, 2, 6, 7", + "[32] Qi Ming, Zhiqiang Zhou, Lingjuan Miao, Hongwei Zhang, and Linhao Li. Dynamic anchor learning for arbitrary-oriented object detection. In Proc. AAAI, pages 2355-2363, 2021. 8", + "[33] Mayank Mishra, Vipul Jain, Saurabh Kumar Singh, and Damodar Maity. Two-stage method based on the you only look once framework and image segmentation for crack detection in concrete structures. Architecture, Structures and Construction, pages 1-18, 2022. 2, 3, 7" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Hafiz Suliman Munawar, Ahmed WA Hammad, Assed Haddad, Carlos Alberto Pereira Soares, and S Travis Waller. Image-based crack detection methods: A review. Infrastructures, 6(8):115, 2021. 1", + "[35] Nhung Hong Thi Nguyen, Stuart Perry, Don Bone, Ha Thanh Le, and Thuy Thi Nguyen. Two-stage convolutional neural network for road crack detection and segmentation. Expert Systems with Applications, 186:115718, 2021. 2, 3, 7", + "[36] Son Dong Nguyen, Thai Son Tran, Van Phuc Tran, Hyun Jong Lee, Md Piran, Van Phuc Le, et al. Deep learning-based crack detection: A survey. International Journal of Pavement Research and Technology, 2022. 1, 2", + "[37] Guangtao Nie and Hua Huang. Multi-oriented object detection in aerial images with double horizontal rectangles. IEEE TPAMI, 2022. 2, 8", + "[38] Sijun Niu and Vikas Srivastava. Simulation trained cnn for accurate embedded crack length, location, and orientation prediction from ultrasound measurements. International Journal of Solids and Structures, 242:111521, 2022. 2, 9", + "[39] Xingjia Pan, Yuqiang Ren, Kekai Sheng, Weiming Dong, Haolei Yuan, Xiaowei Guo, Chongyang Ma, and Changsheng Xu. Dynamic refinement network for oriented and densely packed object detection. In Proc. CVPR, pages 11207-11216, 2020. 2, 6", + "[40] Wen Qian, Xue Yang, Silong Peng, Junchi Yan, and Yue Guo. Learning modulated loss for rotated object detection. In Proc. AAAI, pages 2458-2466, 2021. 3", + "[41] Zhong Qu, Wen Chen, Shi-Yan Wang, Tu-Ming Yi, and Ling Liu. A crack detection algorithm for concrete pavement based on attention mechanism and multi-features fusion. IEEE Transactions on Intelligent Transportation Systems, 2021. 1", + "[42] Marcos Quintana, Juan Torres, and José Manuel Menéndez. A simplified computer vision system for road surface inspection and maintenance. IEEE Transactions on Intelligent Transportation Systems, 17(3):608-619, 2015. 2", + "[43] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proc. CVPR, pages 779-788, 2016. 2", + "[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Proc. NeurIPS, 2015. 2", + "[45] Tianyi Shi, Nicolas Boutry, Yongchao Xu, and Thierry Géraud. Local intensity order transformation for robust curvilinear object segmentation. IEEE Transactions on Image Processing, 31:2557-2569, 2022. 8", + "[46] Yong Shi, Limeng Cui, Zhiquan Qi, Fan Meng, and Zhensong Chen. Automatic road crack detection using random structured forests. IEEE Transactions on Intelligent Transportation Systems, 17(12):3434-3445, 2016. 3", + "[47] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2", + "[48] Weidong Song, Guohui Jia, Hong Zhu, Di Jia, and Lin Gao. Automated pavement crack damage detection using deep multiscale convolutional features. Journal of Advanced Transportation, 2020, 2020. 2, 3" + ], + "bbox": [ + 501, + 90, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6662", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Xinzi Sun, Yuanchang Xie, Liming Jiang, Yu Cao, and Benyuan Liu. Dma-net: Deeplab with multi-scale attention for pavement crack segmentation. IEEE Transactions on Intelligent Transportation Systems, 23(10):18392-18403, 2022. 1", + "[50] Van Phuc Tran, Thai Son Tran, and Hyun Jong Lee. One stage detector (retinanet)-based crack detection for asphalt pavements considering pavement distresses and surface objects. Journal of Civil Structural Health Monitoring, pages 205-222, 2021. 3, 7", + "[51] Ruoxin Wang and Chi Fai Cheung. Centernet-based defect detection for additive manufacturing. Expert Systems with Applications, 188:116000, 2022. 3, 7", + "[52] Maurice Weiler and Gabriele Cesa. General e (2)-equivariant steerable cnns. In Proc. NeurIPS, 2019. 4, 6", + "[53] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In Proc. CVPR, pages 3974-3983, 2018. 3", + "[54] Xuezhi Xiang, Zhiyuan Wang, and Yulong Qiao. An improved yolov5 crack detection method combined with transformer. IEEE Sensors Journal, 2022. 1, 2, 6, 7", + "[55] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented r-cnn for object detection. In Proc. ICCV, pages 3520-3529, 2021. 3, 6, 7, 8", + "[56] Yongchao Xu, Mingtao Fu, Qimeng Wang, Yukang Wang, Kai Chen, Gui-Song Xia, and Xiang Bai. Gliding vertex on the horizontal bounding box for multi-oriented object detection. IEEE TPAMI, 43(4):1452-1459, 2020. 3", + "[57] Fan Yang, Lei Zhang, Sijia Yu, Danil Prokhorov, Xue Mei, and Haibin Ling. Feature pyramid and hierarchical boosting network for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 21(4):1525-1535, 2019. 3", + "[58] Xue Yang, Liping Hou, Yue Zhou, Wentao Wang, and Junchi Yan. Dense label encoding for boundary discontinuity free rotation detection. In Proc. CVPR, 2021. 2, 3, 8", + "[59] Xincong Yang, Heng Li, Yantao Yu, Xiaochun Luo, Ting Huang, and Xu Yang. Automatic pixel-level crack detection and measurement using fully convolutional network. Computer-Aided Civil and Infrastructure Engineering, 33(12):1090-1109, 2018. 3", + "[60] Xue Yang and Junchi Yan. Arbitrary-oriented object detection with circular smooth label. In Proc. ECCV, pages 677-694, 2020. 3, 8", + "[61] Xue Yang and Junchi Yan. On the arbitrary-oriented object detection: Classification based approaches revisited. IJCV, 2022. 2", + "[62] Xue Yang, Junchi Yan, Ziming Feng, and Tao He. R3det: Refined single-stage detector with feature refinement for rotating object. In Proc. AAAI, 2021. 3, 8", + "[63] Xue Yang, Junchi Yan, Qi Ming, Wentao Wang, Xiaopeng Zhang, and Qi Tian. Rethinking rotated object detection with gaussian Wasserstein distance loss. In Proc. ICML, 2021. 3, 8", + "[64] X Yang, J Yang, J Yan, Y Zhang, T Zhang, Z Guo, X Sun, and K SCRDet Fu. Towards more robust detection for small," + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cluttered and rotated objects. In Proc. ICCV, volume 27, pages 8232-8241, 2019. 3", + "[65] Xue Yang, Xiaojiang Yang, Jirui Yang, Qi Ming, Wentao Wang, Qi Tian, and Junchi Yan. Learning high-precision bounding box for rotated object detection via kullback-leibler divergence. In Proc. NeurIPS, 2021. 3, 8", + "[66] Xue Yang, Gefan Zhang, Wentong Li, Yue Zhou, Xuehui Wang, and Junchi Yan. H2rbox: Horizontal box annotation is all you need for oriented object detection. In Proc. ICLR, 2022. 9", + "[67] Dawen Yu, Shunping Ji, Xue Li, Zhaode Yuan, and Chaoyong Shen. Earthquake crack detection from aerial images using a deformable convolutional neural network. IEEE Transactions on Geoscience and Remote Sensing, 2022. 1", + "[68] Hong-Xing Yu, Jiajun Wu, and Li Yi. Rotationally equivariant 3d object detection. In Proc. CVPR, pages 1456-1464, 2022. 3", + "[69] Jianbo Yu, Xun Cheng, and Qingfeng Li. Surface defect detection of steel strips based on anchor-free network with channel attention and bidirectional feature fusion. IEEE Transactions on Instrumentation and Measurement, 2021. 7", + "[70] Gongjie Zhang, Shijian Lu, and Wei Zhang. Cad-net: A context-aware detection network for objects in remote sensing imagery. IEEE Transactions on Geoscience and Remote Sensing, 57(12):10015-10024, 2019. 3", + "[71] Lei Zhang, Fan Yang, Yimin Daniel Zhang, and Ying Julie Zhu. Road crack detection using deep convolutional neural network. In Proc. ICIP, pages 3708-3712, 2016. 6", + "[72] Xiaoyu Zhao, Wenlian Huang, Jie Chen, Zhuangzhuang Chen, and Jianqiang Li. Automatic thin crack segmentation with deep context aggregation network. In Proc. ICARM, pages 206-212, 2022. 3", + "[73] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 2, 4, 6", + "[74] Yue Zhou, Xue Yang, Gefan Zhang, Jiabao Wang, Yanyi Liu, Liping Hou, Xue Jiang, Xingzhao Liu, Junchi Yan, Chengqi Lyu, et al. Mmrotate: A rotated object detection benchmark using pytorch. In Proc. ACMM, pages 7331-7334, 2022. 7", + "[75] Junqing Zhu, Jingtao Zhong, Tao Ma, Xiaoming Huang, Weiguang Zhang, and Yang Zhou. Pavement distress detection using convolutional neural networks with images captured via uav. Automation in Construction, 2022. 2, 6", + "[76] Qin Zou, Yu Cao, Qingquan Li, Qingzhou Mao, and Song Wang. Cracktree: Automatic crack detection from pavement images. Pattern Recognition Letters, 33(3):227-238, 2012. 1, 3", + "[77] Qin Zou, Zheng Zhang, Qingquan Li, Xianbiao Qi, Qian Wang, and Song Wang. Deepcrack: Learning hierarchical convolutional features for crack detection. IEEE TIP, 28(3):1498-1512, 2018. 1, 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 827 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6663", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/b4b142a9-63c6-4b31-9bbd-a4e1e31c2d72_model.json b/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/b4b142a9-63c6-4b31-9bbd-a4e1e31c2d72_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a0e1a6b0ae1fd8f82bc0ac87c3a142027813da0a --- /dev/null +++ b/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/b4b142a9-63c6-4b31-9bbd-a4e1e31c2d72_model.json @@ -0,0 +1,2543 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.131, + 0.877, + 0.153 + ], + "angle": 0, + "content": "The Devil is in the Crack Orientation: A New Perspective for Crack Detection" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.181, + 0.865, + 0.217 + ], + "angle": 0, + "content": "Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Guanming Zhu, Zun Liu, Jie Chen, Jianqiang Li Shenzhen University, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.219, + 0.785, + 0.234 + ], + "angle": 0, + "content": "{chenzhuangzhuang2016，2018101068，zhuguanming2022}@email.szu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.237, + 0.684, + 0.252 + ], + "angle": 0, + "content": "{jin.zhang, chenjie, zunliu, lijq}@szu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.32, + 0.474, + 0.699 + ], + "angle": 0, + "content": "Cracks are usually curve-like structures that are the focus of many computer-vision applications (e.g., road safety inspection and surface inspection of the industrial facilities). The existing pixel-based crack segmentation methods rely on time-consuming and costly pixel-level annotations. And the object-based crack detection methods exploit the horizontal box to detect the crack without considering crack orientation, resulting in scale variation and intraclass variation. Considering this, we provide a new perspective for crack detection that models the cracks as a series of sub-cracks with the corresponding orientation. However, the vanilla adaptation of the existing oriented object detection methods to the crack detection tasks will result in limited performance, due to the boundary discontinuity issue and the ambiguities in sub-crack orientation. In this paper, we propose a first-of-its-kind oriented sub-crack detector, dubbed as CrackDet, which is derived from a novel piecewise angle definition, to ease the boundary discontinuity problem. And then, we propose a multi-branch angle regression loss for learning sub-crack orientation and variance together. Since there are no related benchmarks, we construct three fully annotated datasets, namely, ORC, ONPP, and OCCSD, which involve various cracks in road pavement and industrial facilities. Experiments show that our approach outperforms state-of-the-art crack detectors." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.743, + 0.21, + 0.759 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Crack is a type of defect that appears on the surfaces of many physical structures, e.g., the road pavement [41], the industrial facilities [8]. Inspecting and repairing cracks are crucial tasks for avoiding expansion of the harm to these structures [77]. Besides, a recent study reveals that recognizing the surface cracks caused by earthquakes in time can avoid secondary damage to people's properties" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.287, + 0.572, + 0.3 + ], + "angle": 0, + "content": "Images" + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.287, + 0.692, + 0.3 + ], + "angle": 0, + "content": "Segmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.287, + 0.779, + 0.3 + ], + "angle": 0, + "content": "Detection" + }, + { + "type": "image_caption", + "bbox": [ + 0.824, + 0.287, + 0.863, + 0.3 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.302, + 0.895, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.529, + 0.884, + 0.542 + ], + "angle": 0, + "content": "(a) Visualization of different perspectives for crack detection" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.549, + 0.681, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.628, + 0.686, + 0.641 + ], + "angle": 0, + "content": "(b) Boundary discontinuity" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.543, + 0.88, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.628, + 0.892, + 0.641 + ], + "angle": 0, + "content": "(c) Ambiguities in orientation" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.642, + 0.892, + 0.711 + ], + "angle": 0, + "content": "Figure 1. Cracks tend to be curve-like structures [77] in real-world applications. (a) We compare our orientation-based perspective with others for crack detection. (b) Boundary discontinuity issue in the existing 180-degree definition. (c) The sub-crack orientation is ambiguous in some cases due to the irregular sub-crack shape." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.745, + 0.892, + 0.806 + ], + "angle": 0, + "content": "and lives [67]. However, manual inspection is costly, time-consuming, and prone to human error [34]. Hence, automatic image-based crack detection plays an essential role in a real-world application of crack identification [22, 67]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Currently, image-based crack detection methods can be roughly divided into three categories [36]: patch-level classification [4, 8], object-level detection [54, 31], and pixel-level crack segmentation [76, 49]. Although the first two methods take less labeling cost than the last one, they can only provide a coarse crack localization. Moreover, we are" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.371, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author Email: chenjie@szu.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6653" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "motivated by the fact that a crack is a line object that has a certain width from a local perspective [77]. Then, we are encouraged to start from an orientation-based perspective that models the cracks as a series of oriented sub-cracks (see Fig. 1 (a)). The advantage of this perspective is three-fold: (1) eliminating the scale variation, (2) reducing the intraclass variation, and (3) taking less labeling cost than the segmentation methods. More specifically, in the case of our practice, the price and time cost of the pixel-level annotations are \\(15 \\times\\) and \\(30 \\times\\) higher than our oriented sub-crack annotations, respectively. Besides, considering real-world applications, existing works [33, 35] show great power by taking a detection-then-segmentation strategy, i.e., it first identifies and localizes cracks by a detector and then obtains crack details by segmentation-based methods. Hence, the detector based on our perspective can serve as a complement to segmentation-based methods, as it can provide a more fine-grained localization (see Fig. 1 a)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.655 + ], + "angle": 0, + "content": "However, it is difficult to directly adopt the state-of-art oriented detectors [16, 39, 23, 37] to crack detection. The reason is two-fold: (1) several existing works (e.g., ReDet [16], DRN [39]) simply ignore the boundary discontinuity problem, and (2) although the remaining works make great efforts on the above problem, these works still suffer from ambiguity issue in the sub-crack orientation. Among existing oriented detection methods, those detectors based on five parameters \\((x,y,h,w,\\theta)\\) dominate [58], where \\((x,y)\\) indicates the center point and \\((h,w,\\theta)\\) denotes height, width, and angle, respectively. As shown in Fig. 1 (b), the boundary discontinuity problem indicates that small rotations for oriented objects around the angle period's boundary result in a sudden sharp increase of loss [61]. Moreover, Fig. 1 (c) provides some cases to illustrate the inherent ambiguities in sub-crack orientation. When the growth direction of a crack changes frequently, the orientation of the sub-crack is unclear to judge, making it hard to learn regression functions for estimating orientation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.902 + ], + "angle": 0, + "content": "This paper starts from a new perspective for crack detection that models the cracks as a series of oriented subcracks. After that, we adopt CenterNet [73] as our baseline due to its simplicity. Next, we propose a new piecewise angle definition to ease the boundary discontinuity issue. Based on our piecewise definition, we propose the CrackDet that contains a parallel multi-branch architecture, each of which is responsible for predicting a fixed range of angles. Then, to address the ambiguities in sub-crack orientation, we propose a novel multi-branch angle regression loss, namely MAR Loss, which is based on Wasserstein distances [1] for both angle regression and angle variance estimation. The advantage of coupling CrackDet with MAR Loss is three-fold: (1) the ambiguities in sub-crack orientation can be captured to a certain extent. (2) The learned variance from each branch is beneficial during the post-process. We" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.196 + ], + "angle": 0, + "content": "acquire the estimated variances from each branch and obtain the detection results by variance voting. (3) Since crack orientation is significant in a real-world application [38], the learned variances reveal the level of confidence in the angle prediction, it can potentially help to judge the overall crack orientation. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.257 + ], + "angle": 0, + "content": "(1) To the best of our knowledge, we are the first ones to start from an oriented sub-crack detection perspective for crack detection by modeling cracks as a series of sub-cracks with their inherent orientations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.318 + ], + "angle": 0, + "content": "(2) Propose a first-of-its-kind oriented sub-crack detector, called as CrackDet, which stems from the proposed piecewise angle definition (see Sec. 3.1), to ease the boundary discontinuity issue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.378 + ], + "angle": 0, + "content": "(3) Propose a multi-branch angle regression loss that enables CrackDet to capture ambiguities in sub-crack orientation via jointly learning sub-crack orientation and the corresponding orientation variance (see Sec. 3.2 and 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.453 + ], + "angle": 0, + "content": "(4) A tale of three datasets has been constructed, namely, ORC, ONPP, and OCCSD. These datasets contain orientation annotations of the sub-cracks in real-world industrial facilities and road pavements, to facilitate the research about oriented sub-crack detection." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.638, + 0.484 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.553 + ], + "angle": 0, + "content": "By following the previous study [36], the vision-based crack detection can be roughly clustered into three categories: patch-level classification [4, 8], object-level detection [54, 31], and pixel-level segmentation [77, 48]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.735 + ], + "angle": 0, + "content": "Patch-level crack classification. Recently, deep learning-based methods are more powerful than the traditional methods (e.g., HOG [42], LBP[42], Gaussian process, SVM [42]) in the patch-level crack classification tasks [36]. Among deep learning-based methods, variants of convolutional neural network (CNN) followed by a fully connected (FC) layer are used to recognize patch-level crack samples. Typically, ResNet [17] and VGG [47] are popular backbones among existing backbones. Moreover, focal loss [26] focuses on the hard samples to learn discriminative features. However, patch-level methods fail in providing orientation and the fine-grained localization of crack objects [36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Object-level crack detection. For object-level crack detection, a bounding box is exploited to describe the location and the coarse size of a crack object. Most existing crack detectors mainly adopt the general object detectors, such as Faster R-CNN [44], RetinaNet [26], Yolo [43], CenterNet [12], and SSD [28]. Specifically, Fen et al. [13] first design a hybrid approach that uses Faster R-CNN to detect the crack patch. Then, as the crack orientation is important in the post-process, a deep CNN is exploited to obtain the orientation of each detected patch. Ma et al. [31] design an automatic crack detection system by using YOLO v3 [75]. To" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6654" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.212 + ], + "angle": 0, + "content": "detect various cracks, Wang et al. [51] propose a crack detection model by using CenterNet to extract crack features. Moreover, a RetinaNet-based crack detector is proposed to detect the crack in asphalt pavement [50]. Unfortunately, as shown in Fig. 1, the object-level crack detection methods suffer from scale variation and intra-class variation. Moreover, it also can not directly provide the orientation and the fine-grained localization of cracks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.471, + 0.502 + ], + "angle": 0, + "content": "Pixel-level crack segmentation. Since those horizontal bounding boxes can not describe the exact shape of the crack, segmentation-based methods are proposed to detect cracks at the pixel level. Yang et al. [59] design an automatic crack segmentation pipeline by using fully convolutional networks [30] and achieve a noticeable performance gain over traditional approaches [76, 46]. By coupling with Deeplab v3+ [5], CrackSeg [48] is proposed to learn rich features for different scales. Moreover, aiming to learn high-level crack representations, DeepCrack [77] is proposed to capture the line structures by feature fusion. Considering different receptive fields, FPHBN [57] makes great efforts for crack detection by aggregating pyramid features. Besides, by addressing the imbalance problem, Li et al. [22] propose an adaptive weighted cross-entropy loss function, which serves as a complement to the existing methods. However, these segmentation-based methods heavily rely on high-quality labeled datasets [6, 72], which involves a high labeling cost." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.702 + ], + "angle": 0, + "content": "Compared to the above methods, we start from a new perspective that turns crack detection into oriented subcrack detection. Based on this perspective, we are allowed to provide a fine-grained localization of crack objects (See details in Fig. 1), instead of a coarse localization by the patch-level classification or object-level detection methods. Hence, considering the existing detectionthen-segmentation strategy [33, 35], the detector based on our perspective can serve as a better candidate for the segmentation-based methods. Besides, when considering a dynamic field without historical human supervision, our perspective can greatly ease the labeling cost than the pixellevel segmentation methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Oriented object detection. Since our oriented sub-crack detection perspective belongs to the oriented object detection category, we give a brief review of related research on this scope. Recently, many powerful rotation detectors are mainly derived from horizontal object detectors by adding an orientation regression branch [23]. For example, ICN [2], ROI-Transformer [11], SCRDet [64], CAD-Net [70], CSL[60], \\(\\mathbf{R}^3\\mathrm{Det}\\) [62], ReDet [16], Oriented R-CNN [55], and DCL [58] bring satisfied detection results on DOTA dataset [53] and HRSC2016 dataset [29]. Gliding Vertex [56] and RSDet [40] share the same purpose of considering quadrilateral regression prediction. Note that, the boundary discontinuity problem is caused by the periodicity of the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.334 + ], + "angle": 0, + "content": "angle when dealing with the angel-based orientation estimation. Then, Yang et al. [60] take a new perspective that turns the angular regression tasks into the angular classification tasks [58]. Moreover, Oriented RepPoints [23] utilize the flexible adaptive points as the box representation to achieve oriented object detection. Unlike these methods, our CrackDet addresses the boundary discontinuity problem via the new piecewise angle definition, which is motivated by the divide-and-conquer strategy. Besides, although KLD [65] and GWD [63] also start from a distribution perspective, it converts the center point, height, width, and orientation of a rotated bounding box as a 2-D Gaussian distribution, and does not focus on learning angle variance. Instead, the proposed CrackDet assumes that the ground-truth orientation obeys a single-variate gaussian distribution, and aims to learn the sub-crack orientation and its variance together." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.346, + 0.738, + 0.362 + ], + "angle": 0, + "content": "3. Our method and datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.477 + ], + "angle": 0, + "content": "The overall framework is shown in Fig. 2. We first introduce the piecewise angle definition in Sec. 3.1, and the network architecture of CrackDet in Sec. 3.2. Then, we give the details of the proposed MAR Loss for estimating ambiguities in sub-crack orientation in Sec. 3.3. Finally, we present the inference step of CrackDet with estimated angle variances in Sec. 3.4, and the proposed datasets in Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.485, + 0.733, + 0.501 + ], + "angle": 0, + "content": "3.1. Piecewise angle definition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.569 + ], + "angle": 0, + "content": "In this section, we recap the boundary discontinuity problem and inconsistency regression problem in the existing 180-degree five-parameter definition. Then, we discuss the details of the proposed piecewise angle definition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Rethinking the five-parameter definition. Fig. 2 (a) shows that the existing five-parameter definition [53] with the 180-degree angular range includes \\((x, y, h, w, \\theta)\\), where \\((x, y)\\) indicates the center point and \\([h, w, \\theta]\\) denotes height, width, and angle, respectively. Then, the boundary discontinuity problem refers to a sharp increase of loss at the boundary caused by the periodicity of the angle. As shown in Fig. 3 (a), it can be further explained that when the angle of a rotated box is close to \\(180^{\\circ}\\), the angle under the five-parameter definition will suddenly change from \\(180^{\\circ}\\) to \\(0^{\\circ}\\), resulting a sharp increase of angle regression loss. Moreover, the inconsistency regression problem can be attributed to the rotation symmetry in object detection, which includes the equivariant orientation and invariant shape estimation for a rotated bounding box [68]. Specifically speaking, rotation-equivariant features are desired for orientation estimation. In contrast, the shape of the bounding box does not change after a rotation. Thus, rotation-invariant features are needed for predicting its invariant height and width [16]. Discussion on our piecewise angle definition. We are inspired by the divide-and-conquer strategy [7] and propose the following piecewise angle definition, shown in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6655" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.262, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.089, + 0.877, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.24, + 0.362, + 0.259 + ], + "angle": 0, + "content": "(a) The five-parameter definition" + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.24, + 0.779, + 0.259 + ], + "angle": 0, + "content": "(b) The proposed piecewise angle definition" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.26, + 0.862, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.414, + 0.687, + 0.431 + ], + "angle": 0, + "content": "(c) Overall architecture of the proposed CrackDet" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.434, + 0.892, + 0.504 + ], + "angle": 0, + "content": "Figure 2. (a) The existing five-parameter definition with long-side definition. (b) Our piecewise angle definition is inspired by the divide-and-conquer strategy and built upon the above five-parameter definition. It breaks down the existing 180-degree regression tasks into four sub-tasks for addressing the discontinuity problem. (c) Based on our piecewise angle definition, we present an overview of CrackDet that can estimate standard deviations (std) along with angle regression for capturing ambiguities in the sub-crack orientation. Note that, the heatmap branch and offset branch are responsible for the center (i.e., \\( x \\) and \\( y \\)) of the predicted bounding box as in the original CenterNet." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.536, + 0.455, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.599, + 0.269, + 0.613 + ], + "angle": 0, + "content": "(a) The 180-degree definition" + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.599, + 0.456, + 0.613 + ], + "angle": 0, + "content": "(b) Our piecewise definition" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.682 + ], + "angle": 0, + "content": "Figure 3. The horizontal axis denotes the angle of a rotated box in anti-clockwise order. And the vertical axis denotes the angle definition in different methods. Due to the periodicity of the angle, Figure (a) shows that there is a sudden change when the angle of a rotated box closes to \\(180^{\\circ}\\) with the 180-degree definition." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Fig. 2 (b). On the one hand, Fig. 3 (b) shows that there is no sudden change of the angle in our piecewise angle definition. Thus, ours is free of the boundary discontinuity problem. On the other hand, rotation equivariance can be easily achieved by group convolutions [52], while rotation invariance relies on a larger capacity network or a larger number of training samples. Hence, we focus on rotation-equivariant features. And then, we are allowed to transform the invariant size estimation into equivariant size estimation by adding a factor \\(\\sin (\\theta)\\) or \\(\\cos (\\theta)\\) to the height and width (see Fig. 2 (b)). That way, the box size is enforced to change equivalently when the box orientation changes. Ac" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.529, + 0.892, + 0.56 + ], + "angle": 0, + "content": "cordingly, both box size and angle regressions can benefit from rotation-equivariant features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.573, + 0.702, + 0.588 + ], + "angle": 0, + "content": "3.2. Network architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.703 + ], + "angle": 0, + "content": "As discussed before, we aim to extract rotation-equivariant features for box size regression and angle regression. Therefore, we adopt rotation-equivariant networks as the backbone to extract rotation-equivariant features. Then, we design a parallel multi-branch architecture based on the proposed piecewise angle definition. The overall architecture of CrackDet is shown in Fig. 2 (c)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Rotation-equivariant backbone. We first adopt CenterNet [73] as our baseline, which takes an object as a single point (i.e., the center point of the bounding box) and regresses the object size and offset. Then, we re-implement all layers of the fully-convolutional encoder-decoder networks (i.e., up-convolutional residual networks [20]) in CenterNet based on e2cnn [52], named as ReEDNet. Importantly, e2cnn includes rotation-equivariant convolution, rotation-equivariant up-convolution, and pooling, etc. Thanks to the rotation weight sharing and group representations in e2cnn, our ReEDNet takes smaller parameters than the original encoder-decoder networks in CenterNet and enjoys the capability of equivariance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6656" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.241, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.207, + 0.208, + 0.218 + ], + "angle": 0, + "content": "(a) Branch 1" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.089, + 0.546, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.207, + 0.358, + 0.218 + ], + "angle": 0, + "content": "(b) Branch 2" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.089, + 0.885, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.207, + 0.666, + 0.219 + ], + "angle": 0, + "content": "(d) Branch 4" + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.207, + 0.88, + 0.22 + ], + "angle": 0, + "content": "(e) After variance voting" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.222, + 0.894, + 0.264 + ], + "angle": 0, + "content": "Figure 4. A visualization of our angle-based variance voting. Firstly, the red, blue, orange, and teal textboxes are the corresponding standard deviation obtained from branch 1, 2, 3, and 4, respectively. Then, for each oriented bounding box, we select brach \\(i\\) that has the smallest standard deviation to obtain its final detection result. Note that, the red background box is associated with the selected branch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.471, + 0.395 + ], + "angle": 0, + "content": "Multi-branch architecture. According to our piecewise angle definition, we define a valid range \\([l_i, r_i)\\) for each branch \\(i\\). During the training, we only select the ground truth boxes whose angles fall in the corresponding valid range of one branch. Specifically, referring to the existing 180-degree definition, for an oriented bounding box with angle \\(\\theta\\) on the input image, it is valid for branch \\(i\\) when:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.403, + 0.469, + 0.42 + ], + "angle": 0, + "content": "\\[\nl _ {i} \\leq \\theta < r _ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.47, + 0.683 + ], + "angle": 0, + "content": "As shown in Fig. 2 (c), the valid ranges of four branches are set to \\([0^{\\circ}, 45^{\\circ})\\), \\([45^{\\circ}, 90^{\\circ})\\), \\([90^{\\circ}, 135^{\\circ})\\) and \\([135^{\\circ}, 180^{\\circ})\\), respectively. For branch \\(i\\), we redefine the ground-truth \\(\\theta\\), height \\(h\\), and width \\(w\\) in the 180-degree definition into \\(\\theta_{i}\\), \\(h_{i}\\), and \\(w_{i}\\), according to our piecewise definition in Fig. 2 (b). Moreover, we aim to estimate the angle and angle confidence together. To this end, each branch of our parallel multi-branch architecture contains three detection heads on the top of the rotation-equivariant backbone for predicting size (i.e., height and width), angle, and angle confidence. Note that, we simply implement each head with a fully-connected layer. That way, our network enables us to predict multiple probability distributions instead of only one angle. For simplicity, we assume each predicted angle from each branch is independent and obeys a single-variate gaussian distribution. The corresponding equation is shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.687, + 0.47, + 0.728 + ], + "angle": 0, + "content": "\\[\nP _ {\\Theta_ {i}} (\\theta) = \\frac {1}{\\sqrt {2 \\pi \\sigma_ {i} ^ {2}}} e ^ {- \\frac {\\left(\\theta - \\theta_ {i} ^ {e}\\right) ^ {2}}{2 \\sigma_ {i} ^ {2}}}, i \\in \\{1, 2, 3, 4 \\} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where \\(\\Theta_{i}\\) is the parameters of the angle detection head from branch \\(i\\), \\(\\theta_{i}^{e}\\) denotes the estimated angle from branch \\(i\\), and standard deviation \\(\\sigma_{i}\\) measures the angle confidence of the estimation on branch \\(i\\). When \\(\\sigma_{i}\\) is very close to 0, it means the branch \\(i\\) is extremely confident about the estimated angle. Accordingly, we suppose to have the redefined ground-truth angle \\(\\theta_{i}\\) that is in the valid range of branch \\(i\\). Similarly, \\(\\theta_{i}\\) can also be formulated as a gaussian distribution \\(\\mathcal{N}(\\theta_i,\\sigma_{gt}^2)\\) with its standard deviation \\(\\sigma_{gt}\\to 0\\). Then, this gaussian distribution can be viewed as: \\(P_{i}^{gt}(\\theta) = \\delta (\\theta -\\theta_{i})\\), where \\(\\delta (\\cdot)\\) indicates Dirac delta function." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.288, + 0.729, + 0.305 + ], + "angle": 0, + "content": "3.3. The proposed MAR Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.482 + ], + "angle": 0, + "content": "Suppose the angle of a ground-truth bounding box lies in the valid range of branch \\(i\\), our multi-branch angle regression loss (MAR Loss) aims to minimize the distribution discrepancy between the predicted angle distribution from branch \\(i\\) and the redefined ground-truth angle distribution. Meanwhile, we expect the other three branches to be able to predict larger variance so that we can filter those inaccurate predictions by variance voting in the inference step. More specifically, we first exploit the Wasserstein distance as the distance metric and minimize the distance between \\(P_{\\Theta_i}(\\theta)\\) and \\(P_i^{gt}(\\theta)\\). Secondly, we maximize the estimated variance from the other three branches. The equation is shown as:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.498, + 0.892, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {M A R} = \\frac {D _ {W} \\left(P _ {\\Theta_ {i}} (\\theta) \\| P _ {i} ^ {g t} (\\theta)\\right)}{\\frac {1}{2} + \\sigma_ {i} ^ {2}} - \\sum_ {j = 1, j \\neq i} ^ {4} \\sigma_ {j} ^ {2} \\tag {3} \\\\ = \\frac {\\| \\theta_ {i} ^ {e} - \\theta_ {i} \\| _ {2} ^ {2} + \\sigma_ {i} ^ {2}}{\\frac {1}{2} + \\sigma_ {i} ^ {2}} - \\sum_ {j = 1, j \\neq i} ^ {4} \\sigma_ {j} ^ {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.593, + 0.891, + 0.623 + ], + "angle": 0, + "content": "where \\(D_W(\\cdot)\\) is unfolded based on Wasserstein distance [3]:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.637, + 0.892, + 0.657 + ], + "angle": 0, + "content": "\\[\nD _ {W} \\left(P _ {\\Theta_ {i}} (\\theta) \\| P _ {i} ^ {g t} (\\theta)\\right) = \\| \\theta_ {i} ^ {e} - \\theta_ {i} \\| _ {2} ^ {2} + \\sigma_ {i} ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Importantly, we exploit Wasserstein distance instead of KL-Divergence [18] for angle regression, because KL-Divergence heavily relies on a non-negligible intersection between two distributions [1, 18]. Note that, when \\(\\theta_i^e\\) is estimated accurately, i.e., \\(\\| \\theta_i^e -\\theta_i\\| _2^2\\to 0\\), the branch \\(i\\) is expected to predict smaller variance. In addition, a term \\(\\frac{1}{2} +\\sigma_i^2\\) is added for the following reason: due to the ambiguities in sub-crack orientation, when the angle \\(\\theta_{i}^{e}\\) is estimated inaccurately, i.e., \\(\\| \\theta_i^e -\\theta_i\\| _2^2 >\\frac{1}{2}\\), we expect the branch \\(i\\) to be able to predict larger variance \\(\\sigma_i^2\\) so that \\(L_{MAR}\\) will be lower. Moreover, according to our piecewise angle definition, the size regression loss \\(L_{\\mathrm{size}}\\) for the redefined ground truth height \\(h_i\\) and width \\(w_{i}\\) is defined as \\(L_{\\mathrm{size}} = |h_i^e -h_i| + |w_i^e -w_i|\\) where \\(h_i^e\\) and \\(w_{i}^{e}\\) indicate the estimated height and width from the valid branch \\(i\\). Then," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6657" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.38, + 0.107 + ], + "angle": 0, + "content": "the overall training objective is shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.117, + 0.469, + 0.146 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {C r a c k D e t}} = L _ {k} + \\lambda_ {o f f} L _ {o f f} + \\lambda_ {s i z e} L _ {s i z e} + \\lambda_ {M A R} L _ {M A R} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.148, + 0.469, + 0.207 + ], + "angle": 0, + "content": "where \\( L_{k} \\) and \\( L_{off} \\) are the losses of center point recognition and offset regression by following CenterNet [73]. The hyper-parameter \\( \\lambda_{off} \\), \\( \\lambda_{size} \\), and \\( \\lambda_{MAR} \\) are constant factors to control the balance of the above losses." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.217, + 0.187, + 0.231 + ], + "angle": 0, + "content": "3.4. Inference" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.24, + 0.47, + 0.493 + ], + "angle": 0, + "content": "At the inference step, there are four branches, each of which will predict the height, width, angle, and standard deviation. Thus, we can not directly decide which one should be adopted. Fortunately, with the help of MAR Loss, we design an angle-based variance voting for obtaining the final bounding boxes. Firstly, we extract the peaks in the heatmap by following CenterNet [73]. Then, we get standard deviations \\(\\sigma_{1},\\sigma_{2},\\sigma_{3},\\sigma_{4}\\) from four branches at each peak. As shown in Fig. 4, for each peak, we choose branch \\(i\\) that has the minimum estimated variance to obtain the raw output: \\(i^{*} = \\mathrm{argmin}_{i = \\{1,2,3,4\\}}\\sigma_{i}^{2}\\). Next, according to CenterNet [73], we obtain the center (i.e., \\(x\\) and \\(y\\)) of the oriented bounding box according to the above peaks and the offset branch. Finally, we acquire the final height \\(h_o = h_{i^*}^e\\backslash \\Delta (\\theta_{i^*}^e)\\), width \\(w_{o} = w_{i^{*}}^{e}\\backslash \\Delta (\\theta_{i^{*}}^{e})\\), and angle \\(\\theta_o = \\Gamma (\\theta_{i^*}^e)\\) of the bounding box by the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.493, + 0.469, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\Gamma \\left(\\theta_ {i} ^ {e}\\right) \\triangleq \\left\\{ \\begin{array}{l l} \\theta_ {i} ^ {e}, & i = 1, \\\\ 9 0 ^ {\\circ} - \\theta_ {i} ^ {e}, & i = 2, \\\\ 9 0 ^ {\\circ} + \\theta_ {i} ^ {e}, & i = 3, \\\\ 1 8 0 ^ {\\circ} - \\theta_ {i} ^ {e}, & i = 4. \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.561, + 0.469, + 0.624 + ], + "angle": 0, + "content": "\\[\n\\Delta \\left(\\theta_ {i} ^ {e}\\right) \\triangleq \\left\\{ \\begin{array}{l l} \\cos \\left(\\theta_ {i} ^ {e}\\right), & i = 1, \\\\ \\sin \\left(9 0 ^ {\\circ} - \\theta_ {i} ^ {e}\\right), & i = 2, \\\\ \\sin \\left(9 0 ^ {\\circ} + \\theta_ {i} ^ {e}\\right), & i = 3, \\\\ - \\cos \\left(1 8 0 ^ {\\circ} - \\theta_ {i} ^ {e}\\right), & i = 4. \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.628, + 0.469, + 0.673 + ], + "angle": 0, + "content": "where \\(i\\) denotes the index of the selected branch. \\(\\theta_{i}^{e}\\) indicate the estimated angle from branch \\(i\\). It is worth noting that the remaining inference step is the same as CenterNet [73]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.682, + 0.285, + 0.698 + ], + "angle": 0, + "content": "3.5. The proposed datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.469, + 0.764 + ], + "angle": 0, + "content": "To promote the research on oriented sub-crack detection, we propose ONPP, ORC, and OCCSD datasets, which are collected from industrial facilities, road pavement, and various buildings in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "ONPP: This dataset is collected from the industrial facilities in real-world applications by a high-resolution camera, comprising 200 images with the size of \\(7360 \\times 4912\\). The width of the crack in the collected images varies from 0.05 mm to \\(10\\mathrm{mm}\\). Moreover, the collected images contain different kinds of noise, such as various concrete types and light intensity. To extend this dataset without compromising the resolution, we directly slice these images into \\(512 \\times 512\\) pixels, constructing a final dataset with 3,104 samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "ORC: This dataset is collected from road pavement by a mobile vehicle equipped with a high-resolution camera. It contains 300 images with a size of \\(3489 \\times 3489\\). The images of this dataset include various sizes of cracks in the road. It also contains a lot of noise, such as well covers and various appearances of the crosswalk. Similar to ONPP, we slice these images into \\(512 \\times 512\\) pixels, constructing a final road pavement dataset with 1,303 samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.214, + 0.892, + 0.304 + ], + "angle": 0, + "content": "OCCSD: This dataset is extended from the concrete crack segmentation dataset [71], which contains 458 high-resolution images (\\(4032 \\times 3024\\) pixels). We construct an oriented sub-crack detection dataset with 1,875 samples by slicing these images into \\(512 \\times 512\\) pixels and relabeling them from an oriented sub-crack detection perspective." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.321, + 0.632, + 0.338 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.346, + 0.682, + 0.362 + ], + "angle": 0, + "content": "4.1. Experiment details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Evaluation testbed. We conduct experiments for oriented sub-crack detection on three datasets, i.e., ONPP, ORC, and OCCSD. By following the existing crack detection work [8, 31], we divide each dataset into the training, validation, and test set, with the proportion of \\(8:1:1\\). Moreover, to further verify the effectiveness of CrackDet, we also provide the experiment results on the HRSC2016 dataset [29]. HRSC2016 dataset contains 1,061 samples. Its resolution ranges from \\(300 \\times 300\\) to \\(1500 \\times 900\\). For fairness, according to previous works [55, 16, 23], we exploit the training set (436 images) and validation set (181 images) for training, and the testing set (444 images) for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Implementation details. We adopt Adam as the optimizer for training. The total epoch for training the CrackDet is set as 60, 60, 60, and 140 on ONPP, ORC, OCCSD, and HRSC2016 datasets, respectively. For the first three datasets, the initial learning rate is set to \\(4e - 4\\) and reduced by a factor of 10 after 20, and 40 epochs. For the last one, the initial learning rate is \\(4e - 4\\) and reduced by a factor of 10 after 90, and 120 epochs by following the previous work [39]. The batch size for ONPP, ORC, OCCSD, and HRSC2016 is set to 32, 32, 32, and 8, respectively. More importantly, the input resolutions of four datasets are \\(512 \\times 512\\), \\(512 \\times 512\\), \\(512 \\times 512\\), and \\(768 \\times 768\\), respectively. By following the previous work [16], we only use random flipping for data augmentation. We implement our ReEDNet based on ResNet50 and e2cnn [52], and denote it as ReED-R-50. We train all models on four RTX 3090 GPUs and test on a single RTX 3090 GPU. According to our hyper-parameter sensitivity study, we set \\(\\lambda_{off} = 0.1\\), \\(\\lambda_{size} = 0.2\\), and \\(\\lambda_{MAR} = 0.1\\) in our next experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation metrics. Following previous works [75, 54], we evaluate the performance of CrackDet on oriented crack detection, in terms of Precision and Recall. Moreover, we adopt the mean orientation error (MOE) to verify the effec" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6658" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.153, + 0.088, + 0.817, + 0.234 + ], + "angle": 0, + "content": "
MethodONPPORC
Precision ↑Recall ↑MOE ↓Precision ↑Recall ↑MOE ↓
FCrack-O2020 [13]0.70370.82630.27850.66310.87140.3397
ReCrack-O2021 [50]0.75060.67190.39870.74320.82540.3282
CABF-FCOS-O2021 [69]0.78930.84130.32630.76300.88510.3744
Y3Crack-O2022[31]0.71340.73790.33500.71030.85140.3632
CenterCrack-O2022 [51]0.72260.61320.35740.77310.81020.3085
Yolo-ViT-O2022 [54]0.77660.83790.35030.79820.84610.2978
CrackDet0.82040.91060.16800.85780.89260.1166
" + }, + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.234, + 0.846, + 0.247 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-art crack detectors on ONPP and ORC. MOE denotes the mean orientation error." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.47, + 0.332 + ], + "angle": 0, + "content": "tiveness of the proposed method for estimating crack orientation. For the detection accuracy on the HRSC2016, we adopt the mean average precision (mAP) as an evaluation criterion, which is consistent with VOC2007 metrics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.381, + 0.462, + 0.399 + ], + "angle": 0, + "content": "4.2. Comparison with crack detection approaches" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.473, + 0.901 + ], + "angle": 0, + "content": "In this section, our method is compared to several baselines including state-of-the-art robust crack detection methods: FCrack uses Faster R-CNN to detect the crack patch [13], \\(\\mathrm{Y}^{3}\\mathrm{Crack}\\) designs a intelligent system based on YOLO v3 for detecting crack [31], CenterCrack proposes a crack detection model by using CenterNet [51], ReCrack utilizes the RetinaNet to find the crack in asphalt pavement [50], CABF-FCOS proposes a one-stage network that can detect various defects including crack [69]. Yolo-ViT applies the Transformer to YOLO v5 for crack detection [54]. Note that, all of these methods can not detect the orientation of cracks. For a fair comparison, we add a branch to these methods for the regression of crack orientation and denote them as FCrack-O, \\(\\mathrm{Y}^{3}\\mathrm{Crack - O}\\), CenterCrack-O, ReCrack-O, CABF-FCOS-O, and Yolo-ViT-O respectively. All experiments are independently carried out on ONPP and ORC datasets. Table 1 provides the performance of various approaches from our oriented sub-crack detection perspective for crack detection. Bold denotes the best results. It can be found that the proposed method compares favorably to other competitive crack detection approaches. For example, compared with the Yolo-ViT-O, the precision and recall of CrackDet are 0.8204 and 0.9106 while the Yolo-ViT-O only achieves 0.7766 and 0.8379 on ONPP dataset. For the ORC dataset, CrackDet brings performance gains over Yolo-ViT-O around \\(6.0\\%\\) and \\(4.7\\%\\) under the precision and recall metrics. Moreover, in terms of MOE metric, the proposed CrackDet achieves consistent improvements on all datasets. The results above together show the necessity that the boundary discontinuity problem and the ambiguities in orientation should be treated properly under the oriented sub-crack detection tasks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.272, + 0.892, + 0.288 + ], + "angle": 0, + "content": "4.3. Discussion on the segmentation-based methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.892, + 0.492 + ], + "angle": 0, + "content": "In this section, considering the detection-then-segmentation strategy in the real-world application [33, 35], we further verify that the proposed method can serve as a complement to segmentation-based methods. More specifically, for the pixel-level segmentation task, we replace the oriented box annotation with pixel-level annotation for the OCCSD dataset. Table 3 shows that the proposed method can improve the segmentation results, especially for the segmentation model with limited performance. The reason attributes that the proposed method can provide a more fine-grained localization, which can effectively promote the segmentation model being influenced by the noise background." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.5, + 0.882, + 0.516 + ], + "angle": 0, + "content": "4.4. Comparison with oriented detection methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.644 + ], + "angle": 0, + "content": "To further validate our method, we conduct a series of experiments on both crack datasets and HRSC2016, and report quantitative results to verify the effectiveness of CrackDet. As CrackDet shows the purpose of detecting the oriented object, we compare it with existing state-of-the-art oriented detectors with the help of [74], including Oriented RepPoints [23], Oriented R-CNN [55], SASM [19], etc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Results on oriented sub-crack detection. Table 2 provides the comparison results over the existing powerful oriented object detectors. We can observe that, first, our CrackDet outperforms the existing detectors by a large margin in terms of Precision. The reason is that the proposed piecewise angle definition can promote the proposed detector to take advantage of rotation-equivariant features and be free of the boundary discontinuity problem. Secondly, CrackDet yields significant and consistent improvement on all three datasets in terms of MOE metric. The reason can be explained as CrackDet can estimate the angle confidence with the help of the multi-branch architecture and MAR Loss. Thus, according to the corresponding angle confidence from each branch, the angle-based variance voting improves the angle estimation results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results on HRSC2016. For fairness, we follow the same settings in Oriented RepPoints [23], and report the results" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6659" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.089, + 0.85, + 0.296 + ], + "angle": 0, + "content": "
MethodONPPORC
Precision ↑Recall ↑MOE ↓Precision ↑Recall ↑MOE ↓
\\(CSL_{ECCV'20}\\) [60]0.73090.81460.26330.79430.74370.3285
\\(R^3\\text{Det}_{AAAI'21}\\) [62]0.78350.74330.29770.76970.73120.4280
\\(S^2ANet_{TGRS'21}\\) [15]0.77730.77730.41330.70680.78680.4640
\\(ReDet_{CVPR'21}\\) [16]0.78440.80590.27430.71910.86840.3335
Beyond Bounding-Box \\(CVPR'21\\) [14]0.77900.84680.30030.60760.90390.3441
Oriented R-CNN\\(ICCV'21\\) [55]0.80320.73550.21720.70870.74160.3470
\\(GWD_{ICML'21}\\) [63]0.71690.78840.44010.75760.86320.3014
\\(KLD_{NeurIPS'21}\\) [65]0.74660.64070.34460.74240.90760.2818
\\(SASM_{AAAI'22}\\) [19]0.74160.83030.30620.76530.90380.3489
Oriented RepPoints \\(CVPR'22\\) [23]0.78800.87820.30080.78510.91010.3497
CrackDet0.82040.91060.16800.85780.89260.1166
" + }, + { + "type": "table_caption", + "bbox": [ + 0.168, + 0.296, + 0.799, + 0.308 + ], + "angle": 0, + "content": "Table 2. We provide a detailed comparison over general oriented detection approaches on ONPP and ORC." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.33, + 0.472, + 0.37 + ], + "angle": 0, + "content": "
Precision ↑Crackformer ICCV'21 [27]JTFN ICCV'21 [10]LIOT TIP'22 [45]
w/o CrackDet0.7810.7540.803
w/ CrackDet0.8140.7980.819
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.373, + 0.455, + 0.387 + ], + "angle": 0, + "content": "Table 3. Segmentation performance with CrackDet on OCCSD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.468, + 0.489 + ], + "angle": 0, + "content": "under the VOC2007 metric. Table 4 shows the comparison results. The proposed CrackDet achieves the best performance, even compared with strong FPN-based baselines. Such results further validate that the proposed CrackDet has the potential in more oriented object detection tasks." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.502, + 0.468, + 0.656 + ], + "angle": 0, + "content": "
MethodsBackbonemAP50(07)↑
R3Det AAAI'21 [62]R-101-FPN89.26
R3Det-DCL CVPR'21 [58]R-101-FPN89.46
FPN-CSLECCV'20 [60]R-101-FPN89.62
ReDet CVPR'21 [16]Re-R-5090.46
DAL AAAI'21 [32]R-101-FPN89.77
S2A-NetTGRS'21 [15]R-101-FPN90.17
S2A-Net-DHRec PAMI'22 [37]R-101-FPN90.22
Oriented R-CNN ICCV'21 [55]R-50-FPN90.40
Oriented RepPoints CVPR'22 [23]R-50-FPN90.38
CrackDetReED-R-5090.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.659, + 0.468, + 0.686 + ], + "angle": 0, + "content": "Table 4. We report results on the HRSC2016 test set. mAP(07) denotes detection results under VOC2007 mAP metrics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.227, + 0.727 + ], + "angle": 0, + "content": "4.5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.468, + 0.809 + ], + "angle": 0, + "content": "To examine the contribution of each component in our proposed detector: rotation-equivariant network, the number of valid ranges in the proposed piecewise angle definition, and MAR Loss, a series of ablation experiments are performed on ONPP and ORC datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Evaluation on the rotation-equivariant network. As expected, CrackDet achieves better results by combining the rotation-equivariant network. Table 5 shows that it improves the precision, recall, and MOE by \\(0.91\\%\\), \\(1.5\\%\\), and \\(0.023\\) with the help of the rotation-equivariant network, respectively. The reason behind this effect is that rotation-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.334, + 0.892, + 0.379 + ], + "angle": 0, + "content": "equivariant features extracted by the rotation-equivariant network are beneficial to size and angle regression within the proposed piecewise angle definition." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.442 + ], + "angle": 0, + "content": "
MethodsONPP
Precision ↑Recall ↑MOE ↓
w/o Rotation-equivariant network [17]0.81130.89520.1914
w/ Rotation-equivariant network0.82040.91060.1680
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.445, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Table 5. Performance comparisons of with- and without rotation-equivariant backbone on ONPP dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.488, + 0.892, + 0.639 + ], + "angle": 0, + "content": "Evaluation on the number of valid angle ranges. We study the effect of the number of valid angle ranges in our piecewise angle definition. For fairness, we do not add the factor \\( \\sin(\\theta) \\) or \\( \\cos(\\theta) \\) for all settings. Table 6 shows the results using one to five ranges. The corresponding results demonstrate that our four ranges definition improves over the single range (baseline) with a 0.14 decrease. As can be noticed, the five ranges do not bring further improvement over the four ranges. By considering the complexity and performance, we choose four ranges as the default setting." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.652, + 0.885, + 0.722 + ], + "angle": 0, + "content": "
NumValid Angle RangesMOE ↓
1[0°, 180°)0.33
2[0°, 90°), [90°, 180°)0.30
3[0°, 60°), [60°, 120°), [120°, 180°)0.21
4[0°, 45°), [45°, 90°), [90°, 135°), [135°, 180°)0.19
5[0°, 36°), [36°, 72°), [72°, 108°), [108°, 144°), [144°, 180°)0.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.722, + 0.874, + 0.735 + ], + "angle": 0, + "content": "Table 6. Results from the different number of ranges on ONPP." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation on MAR loss. It is to be expected that Crack-Det combined with MAR Loss achieves consistent improvements on ONPP and ORC datasets, shown in Table 7. The reason is three-fold: (1) KL Loss may fail when two distributions do not have a non-negligible intersection. (2) By learning to estimate high variances for the ambiguous crack orientation during training, our model is able to learn confidence in predicting sub-crack orientation. (3) MAR Loss incorporates learning angle confidence which can potentially help the network to learn discriminative features. The" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6660" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "learned variances through our MAR Loss are interpretable (see Fig. 4). Our network will output higher variances for ambiguous crack orientation, which can be useful in judging the overall orientation in real-world applications [38]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.165, + 0.468, + 0.209 + ], + "angle": 0, + "content": "
MethodsONPPORC
Precision ↑Recall ↑MOE ↓Precision ↑Recall ↑MOE ↓
KL Loss [18]0.78650.87180.24250.81490.85370.1986
MAR Loss0.82040.91060.16800.85780.89260.1166
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.212, + 0.452, + 0.226 + ], + "angle": 0, + "content": "Table 7. Performance comparisons of MAR Loss and KL Loss." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.257, + 0.195, + 0.273 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.471, + 0.614 + ], + "angle": 0, + "content": "In this work, by starting from a new perspective for crack detection, we propose CrackDet, a first-of-its-kind oriented sub-crack detector that overcomes the boundary discontinuity problem. Meanwhile, we propose the MAR loss to capture the ambiguities in crack orientation by combining CrackDet. The experiments demonstrate that our method outperforms state-of-the-art crack detectors and oriented detectors. On one hand, as we also construct three datasets for oriented sub-crack detection, this work should also give a direction to researchers for designing the crack detector from the oriented sub-crack detection perspective. On the other hand, it highlights the key drawback of existing perspectives of crack detection (i.e., a high labeling cost of pixel-level segmentation, and a coarse crack localization of patch-level classification and object-level detection). We believe this paper has the potential to inspire the following works: (1) multi-model learning [24, 25] for boosting the detection performance in a novel field with limited target data [21], (2) semi-supervised learning and active learning [9] to further alleviate its labeling cost in real-world applications, and (3) oriented sub-crack detection via horizontal annotations [66]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.47, + 0.803 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported by the National Nature Science Foundation of China under Grants 62073225,62203134, 61972263, 62072315, the National Key RD Program of China under Grants 2020YFA0908700, the Natural Science Foundation of Guangdong Province Outstanding Youth Program under Grants 2019B151502018, Shenzhen Science and Technology Innovation Commission (20220809141216003, JCYJ20210324093808021, JCYJ20220531102817040), the Guangdong \"Pearl River Talent Recruitment Program\" under Grant 2019ZT08X603, the Guangdong \"Pearl River Talent Plan\" under Grant 2019JC01X235." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.833, + 0.174, + 0.849 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Proc. ICML, pages 214-223, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[2] Seyed Majid Azimi, Eleonora Vig, Reza Bahmanyar, Marco Körner, and Peter Reinartz. Towards multi-class object detection in unconstrained remote sensing imagery. In Proc. ACCV, pages 150-165, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.149, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[3] Raktim Bhattacharya. Data assimilation in optimal transport framework. arXiv preprint arXiv:2005.08670, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[4] Fu-Chen Chen and Mohammad R Jahanshahi. Nb-cnn: Deep learning-based crack detection using convolutional neural network and naive bayes data fusion. IEEE Transactions on Industrial Electronics, 65(5):4392-4400, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proc. ECCV, pages 801-818, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.292, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[6] Minghao Chen, Hongyang Xue, and Deng Cai. Domain adaptation for semantic segmentation with maximum squares loss. In Proc. ICCV, pages 2090-2099, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[7] Qiang Chen, Yingming Wang, Tong Yang, Xiangyu Zhang, Jian Cheng, and Jian Sun. You only look one-level feature. In Proc. CVPR, pages 13039-13048, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.377, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[8] Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Jie Chen, Zun Liu, and Jianqiang Li. Geometry-aware guided loss for deep crack recognition. In Proc. CVPR, pages 4703-4712, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.434, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[9] Zhuangzhuang Chen, Jin Zhang, Pan Wang, Jie Chen, and Jianqiang Li. When active learning meets implicit semantic data augmentation. In Proc. ECCV, pages 56-72, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[10] Mingfei Cheng, Kaili Zhao, Xuhong Guo, Yajing Xu, and Jun Guo. Joint topology-preserving and feature-refinement network for curvilinear structure segmentation. In Proc. ICCV, pages 7147-7156, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.533, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[11] Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning roi transformer for oriented object detection in aerial images. In Proc. CVPR, pages 2849-2858, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.575, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[12] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proc. ICCV, pages 6569-6578, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.618, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[13] Fen Fang, Liyuan Li, Ying Gu, Hongyuan Zhu, and Joo-Hwee Lim. A novel hybrid approach for crack detection. PR, 107:107474, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.661, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[14] Zonghao Guo, Chang Liu, Xiaosong Zhang, Jianbin Jiao, Xi-angyang Ji, and Qixiang Ye. Beyond bounding-box: Convex-hull feature adaptation for oriented and densely packed object detection. In Proc. CVPR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.718, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[15] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[16] Jiaming Han, Jian Ding, Nan Xue, and Gui-Song Xia. Redet: A rotation-equivariant detector for aerial object detection. In Proc. CVPR, pages 2786-2795, 2021. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.803, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In Proc. ECCV, pages 630-645, 2016. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[18] Yihui He, Chenchen Zhu, Jianren Wang, Marios Savvides, and Xiangyu Zhang. Bounding box regression with uncertainty for accurate object detection. In Proc. CVPR, pages 2888-2897, 2019. 5, 9" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6661" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[19] Liping Hou, Ke Lu, Jian Xue, and Yuqiu Li. Shape-adaptive selection and measurement for oriented object detection. In Proc. AAAI, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[20] Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, et al. Speed/accuracy trade-offs for modern convolutional object detectors. In Proc. CVPR, pages 7310-7311, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.26 + ], + "angle": 0, + "content": "[21] Jianqiang Li, Zhuangzhuang Chen, Jie Chen, and Qiuzhen Lin. Diversity-sensitive generative adversarial network for terrain mapping under limited human intervention. IEEE Transactions on Cybernetics, 51(12):6029-6040, 2020. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[22] Kai Li, Bo Wang, Yingjie Tian, and Zhiquan Qi. Fast and accurate road crack detection based on adaptive cost-sensitive loss function. IEEE Transactions on Cybernetics, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.344 + ], + "angle": 0, + "content": "[23] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented leptopoints for aerial object detection. In Proc. CVPR, pages 1829-1838, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[24] Wenrui Li and Xiaopeng Fan. Image-text alignment and retrieval using light-weight transformer. In Proc. ICASSP, pages 4758-4762, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[25] Wenrui Li, Zhengyu Ma, Liang-Jian Deng, Xiaopeng Fan, and Yonghong Tian. Neuron-based spiking transmission and reasoning network for robust image-text retrieval. IEEE TCSVT, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.471, + 0.483 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dálár. Focal loss for dense object detection. In Proc. ICCV, pages 2980-2988, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[27] Huajun Liu, Xiangyu Miao, Christoph Mertz, Chengzhong Xu, and Hui Kong. Crackformer: Transformer network for fine-grained crack detection. In Proc. ICCV, pages 3783-3792, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.595 + ], + "angle": 0, + "content": "[28] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In Proc. ECCV, pages 21-37, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[29] Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang. A high resolution optical satellite image dataset for ship recognition and some new baselines. In International conference on pattern recognition applications and methods, volume 2, pages 324-331, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.706 + ], + "angle": 0, + "content": "[30] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proc. CVPR, pages 3431-3440, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.775 + ], + "angle": 0, + "content": "[31] Duo Ma, Hongyuan Fang, Niannian Wang, Chao Zhang, Jiaxiu Dong, and Haobang Hu. Automatic detection and counting system for pavement cracks based on pcgan and yolo-mf. IEEE Transactions on Intelligent Transportation Systems, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[32] Qi Ming, Zhiqiang Zhou, Lingjuan Miao, Hongwei Zhang, and Linhao Li. Dynamic anchor learning for arbitrary-oriented object detection. In Proc. AAAI, pages 2355-2363, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[33] Mayank Mishra, Vipul Jain, Saurabh Kumar Singh, and Damodar Maity. Two-stage method based on the you only look once framework and image segmentation for crack detection in concrete structures. Architecture, Structures and Construction, pages 1-18, 2022. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[34] Hafiz Suliman Munawar, Ahmed WA Hammad, Assed Haddad, Carlos Alberto Pereira Soares, and S Travis Waller. Image-based crack detection methods: A review. Infrastructures, 6(8):115, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[35] Nhung Hong Thi Nguyen, Stuart Perry, Don Bone, Ha Thanh Le, and Thuy Thi Nguyen. Two-stage convolutional neural network for road crack detection and segmentation. Expert Systems with Applications, 186:115718, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.894, + 0.262 + ], + "angle": 0, + "content": "[36] Son Dong Nguyen, Thai Son Tran, Van Phuc Tran, Hyun Jong Lee, Md Piran, Van Phuc Le, et al. Deep learning-based crack detection: A survey. International Journal of Pavement Research and Technology, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.893, + 0.304 + ], + "angle": 0, + "content": "[37] Guangtao Nie and Hua Huang. Multi-oriented object detection in aerial images with double horizontal rectangles. IEEE TPAMI, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.893, + 0.361 + ], + "angle": 0, + "content": "[38] Sijun Niu and Vikas Srivastava. Simulation trained cnn for accurate embedded crack length, location, and orientation prediction from ultrasound measurements. International Journal of Solids and Structures, 242:111521, 2022. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.361, + 0.893, + 0.432 + ], + "angle": 0, + "content": "[39] Xingjia Pan, Yuqiang Ren, Kekai Sheng, Weiming Dong, Haolei Yuan, Xiaowei Guo, Chongyang Ma, and Changsheng Xu. Dynamic refinement network for oriented and densely packed object detection. In Proc. CVPR, pages 11207-11216, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.432, + 0.893, + 0.475 + ], + "angle": 0, + "content": "[40] Wen Qian, Xue Yang, Silong Peng, Junchi Yan, and Yue Guo. Learning modulated loss for rotated object detection. In Proc. AAAI, pages 2458-2466, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.893, + 0.545 + ], + "angle": 0, + "content": "[41] Zhong Qu, Wen Chen, Shi-Yan Wang, Tu-Ming Yi, and Ling Liu. A crack detection algorithm for concrete pavement based on attention mechanism and multi-features fusion. IEEE Transactions on Intelligent Transportation Systems, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.893, + 0.602 + ], + "angle": 0, + "content": "[42] Marcos Quintana, Juan Torres, and José Manuel Menéndez. A simplified computer vision system for road surface inspection and maintenance. IEEE Transactions on Intelligent Transportation Systems, 17(3):608-619, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.602, + 0.893, + 0.645 + ], + "angle": 0, + "content": "[43] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proc. CVPR, pages 779-788, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.893, + 0.688 + ], + "angle": 0, + "content": "[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Proc. NeurIPS, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.893, + 0.745 + ], + "angle": 0, + "content": "[45] Tianyi Shi, Nicolas Boutry, Yongchao Xu, and Thierry Géraud. Local intensity order transformation for robust curvilinear object segmentation. IEEE Transactions on Image Processing, 31:2557-2569, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.893, + 0.801 + ], + "angle": 0, + "content": "[46] Yong Shi, Limeng Cui, Zhiquan Qi, Fan Meng, and Zhensong Chen. Automatic road crack detection using random structured forests. IEEE Transactions on Intelligent Transportation Systems, 17(12):3434-3445, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.893, + 0.845 + ], + "angle": 0, + "content": "[47] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[48] Weidong Song, Guohui Jia, Hong Zhu, Di Jia, and Lin Gao. Automated pavement crack damage detection using deep multiscale convolutional features. Journal of Advanced Transportation, 2020, 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6662" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[49] Xinzi Sun, Yuanchang Xie, Liming Jiang, Yu Cao, and Benyuan Liu. Dma-net: Deeplab with multi-scale attention for pavement crack segmentation. IEEE Transactions on Intelligent Transportation Systems, 23(10):18392-18403, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[50] Van Phuc Tran, Thai Son Tran, and Hyun Jong Lee. One stage detector (retinanet)-based crack detection for asphalt pavements considering pavement distresses and surface objects. Journal of Civil Structural Health Monitoring, pages 205-222, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[51] Ruoxin Wang and Chi Fai Cheung. Centernet-based defect detection for additive manufacturing. Expert Systems with Applications, 188:116000, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.303 + ], + "angle": 0, + "content": "[52] Maurice Weiler and Gabriele Cesa. General e (2)-equivariant steerable cnns. In Proc. NeurIPS, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[53] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In Proc. CVPR, pages 3974-3983, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[54] Xuezhi Xiang, Zhiyuan Wang, and Yulong Qiao. An improved yolov5 crack detection method combined with transformer. IEEE Sensors Journal, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[55] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented r-cnn for object detection. In Proc. ICCV, pages 3520-3529, 2021. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[56] Yongchao Xu, Mingtao Fu, Qimeng Wang, Yukang Wang, Kai Chen, Gui-Song Xia, and Xiang Bai. Gliding vertex on the horizontal bounding box for multi-oriented object detection. IEEE TPAMI, 43(4):1452-1459, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[57] Fan Yang, Lei Zhang, Sijia Yu, Danil Prokhorov, Xue Mei, and Haibin Ling. Feature pyramid and hierarchical boosting network for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 21(4):1525-1535, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.47, + 0.616 + ], + "angle": 0, + "content": "[58] Xue Yang, Liping Hou, Yue Zhou, Wentao Wang, and Junchi Yan. Dense label encoding for boundary discontinuity free rotation detection. In Proc. CVPR, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[59] Xincong Yang, Heng Li, Yantao Yu, Xiaochun Luo, Ting Huang, and Xu Yang. Automatic pixel-level crack detection and measurement using fully convolutional network. Computer-Aided Civil and Infrastructure Engineering, 33(12):1090-1109, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[60] Xue Yang and Junchi Yan. Arbitrary-oriented object detection with circular smooth label. In Proc. ECCV, pages 677-694, 2020. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[61] Xue Yang and Junchi Yan. On the arbitrary-oriented object detection: Classification based approaches revisited. IJCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[62] Xue Yang, Junchi Yan, Ziming Feng, and Tao He. R3det: Refined single-stage detector with feature refinement for rotating object. In Proc. AAAI, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[63] Xue Yang, Junchi Yan, Qi Ming, Wentao Wang, Xiaopeng Zhang, and Qi Tian. Rethinking rotated object detection with gaussian Wasserstein distance loss. In Proc. ICML, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[64] X Yang, J Yang, J Yan, Y Zhang, T Zhang, Z Guo, X Sun, and K SCRDet Fu. Towards more robust detection for small," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "cluttered and rotated objects. In Proc. ICCV, volume 27, pages 8232-8241, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[65] Xue Yang, Xiaojiang Yang, Jirui Yang, Qi Ming, Wentao Wang, Qi Tian, and Junchi Yan. Learning high-precision bounding box for rotated object detection via kullback-leibler divergence. In Proc. NeurIPS, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[66] Xue Yang, Gefan Zhang, Wentong Li, Yue Zhou, Xuehui Wang, and Junchi Yan. H2rbox: Horizontal box annotation is all you need for oriented object detection. In Proc. ICLR, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[67] Dawen Yu, Shunping Ji, Xue Li, Zhaode Yuan, and Chaoyong Shen. Earthquake crack detection from aerial images using a deformable convolutional neural network. IEEE Transactions on Geoscience and Remote Sensing, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[68] Hong-Xing Yu, Jiajun Wu, and Li Yi. Rotationally equivariant 3d object detection. In Proc. CVPR, pages 1456-1464, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[69] Jianbo Yu, Xun Cheng, and Qingfeng Li. Surface defect detection of steel strips based on anchor-free network with channel attention and bidirectional feature fusion. IEEE Transactions on Instrumentation and Measurement, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[70] Gongjie Zhang, Shijian Lu, and Wei Zhang. Cad-net: A context-aware detection network for objects in remote sensing imagery. IEEE Transactions on Geoscience and Remote Sensing, 57(12):10015-10024, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[71] Lei Zhang, Fan Yang, Yimin Daniel Zhang, and Ying Julie Zhu. Road crack detection using deep convolutional neural network. In Proc. ICIP, pages 3708-3712, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[72] Xiaoyu Zhao, Wenlian Huang, Jie Chen, Zhuangzhuang Chen, and Jianqiang Li. Automatic thin crack segmentation with deep context aggregation network. In Proc. ICARM, pages 206-212, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[73] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[74] Yue Zhou, Xue Yang, Gefan Zhang, Jiabao Wang, Yanyi Liu, Liping Hou, Xue Jiang, Xingzhao Liu, Junchi Yan, Chengqi Lyu, et al. Mmrotate: A rotated object detection benchmark using pytorch. In Proc. ACMM, pages 7331-7334, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[75] Junqing Zhu, Jingtao Zhong, Tao Ma, Xiaoming Huang, Weiguang Zhang, and Yang Zhou. Pavement distress detection using convolutional neural networks with images captured via uav. Automation in Construction, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[76] Qin Zou, Yu Cao, Qingquan Li, Qingzhou Mao, and Song Wang. Cracktree: Automatic crack detection from pavement images. Pattern Recognition Letters, 33(3):227-238, 2012. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[77] Qin Zou, Zheng Zhang, Qingquan Li, Xianbiao Qi, Qian Wang, and Song Wang. Deepcrack: Learning hierarchical convolutional features for crack detection. IEEE TIP, 28(3):1498-1512, 2018. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.828 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6663" + } + ] +] \ No newline at end of file diff --git a/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/b4b142a9-63c6-4b31-9bbd-a4e1e31c2d72_origin.pdf b/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/b4b142a9-63c6-4b31-9bbd-a4e1e31c2d72_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..512ad1f31d5d969da996e2e5f6544427992c04e9 --- /dev/null +++ b/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/b4b142a9-63c6-4b31-9bbd-a4e1e31c2d72_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1a7482ca09fb376e74e425afc15a057668832dbca8bde8ddca89318b422eb7a +size 4979013 diff --git a/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/full.md b/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fb83a680ff21c5ad685c80c2bd6c71514954478a --- /dev/null +++ b/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/full.md @@ -0,0 +1,335 @@ +# The Devil is in the Crack Orientation: A New Perspective for Crack Detection + +Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Guanming Zhu, Zun Liu, Jie Chen, Jianqiang Li Shenzhen University, Shenzhen, China + +{chenzhuangzhuang2016，2018101068，zhuguanming2022}@email.szu.edu.cn + +{jin.zhang, chenjie, zunliu, lijq}@szu.edu.cn + +# Abstract + +Cracks are usually curve-like structures that are the focus of many computer-vision applications (e.g., road safety inspection and surface inspection of the industrial facilities). The existing pixel-based crack segmentation methods rely on time-consuming and costly pixel-level annotations. And the object-based crack detection methods exploit the horizontal box to detect the crack without considering crack orientation, resulting in scale variation and intraclass variation. Considering this, we provide a new perspective for crack detection that models the cracks as a series of sub-cracks with the corresponding orientation. However, the vanilla adaptation of the existing oriented object detection methods to the crack detection tasks will result in limited performance, due to the boundary discontinuity issue and the ambiguities in sub-crack orientation. In this paper, we propose a first-of-its-kind oriented sub-crack detector, dubbed as CrackDet, which is derived from a novel piecewise angle definition, to ease the boundary discontinuity problem. And then, we propose a multi-branch angle regression loss for learning sub-crack orientation and variance together. Since there are no related benchmarks, we construct three fully annotated datasets, namely, ORC, ONPP, and OCCSD, which involve various cracks in road pavement and industrial facilities. Experiments show that our approach outperforms state-of-the-art crack detectors. + +# 1. Introduction + +Crack is a type of defect that appears on the surfaces of many physical structures, e.g., the road pavement [41], the industrial facilities [8]. Inspecting and repairing cracks are crucial tasks for avoiding expansion of the harm to these structures [77]. Besides, a recent study reveals that recognizing the surface cracks caused by earthquakes in time can avoid secondary damage to people's properties + +![](images/353ce6bab1dc650c20cb1584d0fcd775c70a88ecf20b729933de78f7d1ae718b.jpg) +Images +Segmentation +Detection +Ours +(a) Visualization of different perspectives for crack detection + +![](images/ebdacb2e49e1b25ae2264cb8df0e79f084d4aa4945f46ef7e5c2a3f4fbf3371d.jpg) +(b) Boundary discontinuity +Figure 1. Cracks tend to be curve-like structures [77] in real-world applications. (a) We compare our orientation-based perspective with others for crack detection. (b) Boundary discontinuity issue in the existing 180-degree definition. (c) The sub-crack orientation is ambiguous in some cases due to the irregular sub-crack shape. + +![](images/38ce88cbf2af0d07252c02137be66dbeaa620a0a9172513e275408cd4ffd606b.jpg) +(c) Ambiguities in orientation + +and lives [67]. However, manual inspection is costly, time-consuming, and prone to human error [34]. Hence, automatic image-based crack detection plays an essential role in a real-world application of crack identification [22, 67]. + +Currently, image-based crack detection methods can be roughly divided into three categories [36]: patch-level classification [4, 8], object-level detection [54, 31], and pixel-level crack segmentation [76, 49]. Although the first two methods take less labeling cost than the last one, they can only provide a coarse crack localization. Moreover, we are + +motivated by the fact that a crack is a line object that has a certain width from a local perspective [77]. Then, we are encouraged to start from an orientation-based perspective that models the cracks as a series of oriented sub-cracks (see Fig. 1 (a)). The advantage of this perspective is three-fold: (1) eliminating the scale variation, (2) reducing the intraclass variation, and (3) taking less labeling cost than the segmentation methods. More specifically, in the case of our practice, the price and time cost of the pixel-level annotations are $15 \times$ and $30 \times$ higher than our oriented sub-crack annotations, respectively. Besides, considering real-world applications, existing works [33, 35] show great power by taking a detection-then-segmentation strategy, i.e., it first identifies and localizes cracks by a detector and then obtains crack details by segmentation-based methods. Hence, the detector based on our perspective can serve as a complement to segmentation-based methods, as it can provide a more fine-grained localization (see Fig. 1 a). + +However, it is difficult to directly adopt the state-of-art oriented detectors [16, 39, 23, 37] to crack detection. The reason is two-fold: (1) several existing works (e.g., ReDet [16], DRN [39]) simply ignore the boundary discontinuity problem, and (2) although the remaining works make great efforts on the above problem, these works still suffer from ambiguity issue in the sub-crack orientation. Among existing oriented detection methods, those detectors based on five parameters $(x,y,h,w,\theta)$ dominate [58], where $(x,y)$ indicates the center point and $(h,w,\theta)$ denotes height, width, and angle, respectively. As shown in Fig. 1 (b), the boundary discontinuity problem indicates that small rotations for oriented objects around the angle period's boundary result in a sudden sharp increase of loss [61]. Moreover, Fig. 1 (c) provides some cases to illustrate the inherent ambiguities in sub-crack orientation. When the growth direction of a crack changes frequently, the orientation of the sub-crack is unclear to judge, making it hard to learn regression functions for estimating orientation. + +This paper starts from a new perspective for crack detection that models the cracks as a series of oriented subcracks. After that, we adopt CenterNet [73] as our baseline due to its simplicity. Next, we propose a new piecewise angle definition to ease the boundary discontinuity issue. Based on our piecewise definition, we propose the CrackDet that contains a parallel multi-branch architecture, each of which is responsible for predicting a fixed range of angles. Then, to address the ambiguities in sub-crack orientation, we propose a novel multi-branch angle regression loss, namely MAR Loss, which is based on Wasserstein distances [1] for both angle regression and angle variance estimation. The advantage of coupling CrackDet with MAR Loss is three-fold: (1) the ambiguities in sub-crack orientation can be captured to a certain extent. (2) The learned variance from each branch is beneficial during the post-process. We + +acquire the estimated variances from each branch and obtain the detection results by variance voting. (3) Since crack orientation is significant in a real-world application [38], the learned variances reveal the level of confidence in the angle prediction, it can potentially help to judge the overall crack orientation. Our main contributions are summarized as follows: + +(1) To the best of our knowledge, we are the first ones to start from an oriented sub-crack detection perspective for crack detection by modeling cracks as a series of sub-cracks with their inherent orientations. +(2) Propose a first-of-its-kind oriented sub-crack detector, called as CrackDet, which stems from the proposed piecewise angle definition (see Sec. 3.1), to ease the boundary discontinuity issue. +(3) Propose a multi-branch angle regression loss that enables CrackDet to capture ambiguities in sub-crack orientation via jointly learning sub-crack orientation and the corresponding orientation variance (see Sec. 3.2 and 3.3). +(4) A tale of three datasets has been constructed, namely, ORC, ONPP, and OCCSD. These datasets contain orientation annotations of the sub-cracks in real-world industrial facilities and road pavements, to facilitate the research about oriented sub-crack detection. + +# 2. Related work + +By following the previous study [36], the vision-based crack detection can be roughly clustered into three categories: patch-level classification [4, 8], object-level detection [54, 31], and pixel-level segmentation [77, 48]. + +Patch-level crack classification. Recently, deep learning-based methods are more powerful than the traditional methods (e.g., HOG [42], LBP[42], Gaussian process, SVM [42]) in the patch-level crack classification tasks [36]. Among deep learning-based methods, variants of convolutional neural network (CNN) followed by a fully connected (FC) layer are used to recognize patch-level crack samples. Typically, ResNet [17] and VGG [47] are popular backbones among existing backbones. Moreover, focal loss [26] focuses on the hard samples to learn discriminative features. However, patch-level methods fail in providing orientation and the fine-grained localization of crack objects [36]. + +Object-level crack detection. For object-level crack detection, a bounding box is exploited to describe the location and the coarse size of a crack object. Most existing crack detectors mainly adopt the general object detectors, such as Faster R-CNN [44], RetinaNet [26], Yolo [43], CenterNet [12], and SSD [28]. Specifically, Fen et al. [13] first design a hybrid approach that uses Faster R-CNN to detect the crack patch. Then, as the crack orientation is important in the post-process, a deep CNN is exploited to obtain the orientation of each detected patch. Ma et al. [31] design an automatic crack detection system by using YOLO v3 [75]. To + +detect various cracks, Wang et al. [51] propose a crack detection model by using CenterNet to extract crack features. Moreover, a RetinaNet-based crack detector is proposed to detect the crack in asphalt pavement [50]. Unfortunately, as shown in Fig. 1, the object-level crack detection methods suffer from scale variation and intra-class variation. Moreover, it also can not directly provide the orientation and the fine-grained localization of cracks. + +Pixel-level crack segmentation. Since those horizontal bounding boxes can not describe the exact shape of the crack, segmentation-based methods are proposed to detect cracks at the pixel level. Yang et al. [59] design an automatic crack segmentation pipeline by using fully convolutional networks [30] and achieve a noticeable performance gain over traditional approaches [76, 46]. By coupling with Deeplab v3+ [5], CrackSeg [48] is proposed to learn rich features for different scales. Moreover, aiming to learn high-level crack representations, DeepCrack [77] is proposed to capture the line structures by feature fusion. Considering different receptive fields, FPHBN [57] makes great efforts for crack detection by aggregating pyramid features. Besides, by addressing the imbalance problem, Li et al. [22] propose an adaptive weighted cross-entropy loss function, which serves as a complement to the existing methods. However, these segmentation-based methods heavily rely on high-quality labeled datasets [6, 72], which involves a high labeling cost. + +Compared to the above methods, we start from a new perspective that turns crack detection into oriented subcrack detection. Based on this perspective, we are allowed to provide a fine-grained localization of crack objects (See details in Fig. 1), instead of a coarse localization by the patch-level classification or object-level detection methods. Hence, considering the existing detectionthen-segmentation strategy [33, 35], the detector based on our perspective can serve as a better candidate for the segmentation-based methods. Besides, when considering a dynamic field without historical human supervision, our perspective can greatly ease the labeling cost than the pixellevel segmentation methods. + +Oriented object detection. Since our oriented sub-crack detection perspective belongs to the oriented object detection category, we give a brief review of related research on this scope. Recently, many powerful rotation detectors are mainly derived from horizontal object detectors by adding an orientation regression branch [23]. For example, ICN [2], ROI-Transformer [11], SCRDet [64], CAD-Net [70], CSL[60], $\mathbf{R}^3\mathrm{Det}$ [62], ReDet [16], Oriented R-CNN [55], and DCL [58] bring satisfied detection results on DOTA dataset [53] and HRSC2016 dataset [29]. Gliding Vertex [56] and RSDet [40] share the same purpose of considering quadrilateral regression prediction. Note that, the boundary discontinuity problem is caused by the periodicity of the + +angle when dealing with the angel-based orientation estimation. Then, Yang et al. [60] take a new perspective that turns the angular regression tasks into the angular classification tasks [58]. Moreover, Oriented RepPoints [23] utilize the flexible adaptive points as the box representation to achieve oriented object detection. Unlike these methods, our CrackDet addresses the boundary discontinuity problem via the new piecewise angle definition, which is motivated by the divide-and-conquer strategy. Besides, although KLD [65] and GWD [63] also start from a distribution perspective, it converts the center point, height, width, and orientation of a rotated bounding box as a 2-D Gaussian distribution, and does not focus on learning angle variance. Instead, the proposed CrackDet assumes that the ground-truth orientation obeys a single-variate gaussian distribution, and aims to learn the sub-crack orientation and its variance together. + +# 3. Our method and datasets + +The overall framework is shown in Fig. 2. We first introduce the piecewise angle definition in Sec. 3.1, and the network architecture of CrackDet in Sec. 3.2. Then, we give the details of the proposed MAR Loss for estimating ambiguities in sub-crack orientation in Sec. 3.3. Finally, we present the inference step of CrackDet with estimated angle variances in Sec. 3.4, and the proposed datasets in Sec. 3.5. + +# 3.1. Piecewise angle definition + +In this section, we recap the boundary discontinuity problem and inconsistency regression problem in the existing 180-degree five-parameter definition. Then, we discuss the details of the proposed piecewise angle definition. + +Rethinking the five-parameter definition. Fig. 2 (a) shows that the existing five-parameter definition [53] with the 180-degree angular range includes $(x, y, h, w, \theta)$ , where $(x, y)$ indicates the center point and $[h, w, \theta]$ denotes height, width, and angle, respectively. Then, the boundary discontinuity problem refers to a sharp increase of loss at the boundary caused by the periodicity of the angle. As shown in Fig. 3 (a), it can be further explained that when the angle of a rotated box is close to $180^{\circ}$ , the angle under the five-parameter definition will suddenly change from $180^{\circ}$ to $0^{\circ}$ , resulting a sharp increase of angle regression loss. Moreover, the inconsistency regression problem can be attributed to the rotation symmetry in object detection, which includes the equivariant orientation and invariant shape estimation for a rotated bounding box [68]. Specifically speaking, rotation-equivariant features are desired for orientation estimation. In contrast, the shape of the bounding box does not change after a rotation. Thus, rotation-invariant features are needed for predicting its invariant height and width [16]. Discussion on our piecewise angle definition. We are inspired by the divide-and-conquer strategy [7] and propose the following piecewise angle definition, shown in + +![](images/f78265a71a5b0e24a85cc2a82660ea83321cd092bfb6563b85b68983066bcad5.jpg) + +![](images/f1c1fc28a8475903ae0cf41fc6f4a5ff7337a4b1b39025cc7c7dab6598b69616.jpg) + +![](images/7819f9a0c288ed74cbd4b303221544d3c03bdd41a301475d5f26b358f4b397e3.jpg) +(a) The five-parameter definition +(b) The proposed piecewise angle definition +(c) Overall architecture of the proposed CrackDet +Figure 2. (a) The existing five-parameter definition with long-side definition. (b) Our piecewise angle definition is inspired by the divide-and-conquer strategy and built upon the above five-parameter definition. It breaks down the existing 180-degree regression tasks into four sub-tasks for addressing the discontinuity problem. (c) Based on our piecewise angle definition, we present an overview of CrackDet that can estimate standard deviations (std) along with angle regression for capturing ambiguities in the sub-crack orientation. Note that, the heatmap branch and offset branch are responsible for the center (i.e., $x$ and $y$ ) of the predicted bounding box as in the original CenterNet. + +![](images/47a5c9a2891c667b5976e9ce2a1dc021e9b6f8418d068c3c3caa1686a2b11413.jpg) +(a) The 180-degree definition +(b) Our piecewise definition +Figure 3. The horizontal axis denotes the angle of a rotated box in anti-clockwise order. And the vertical axis denotes the angle definition in different methods. Due to the periodicity of the angle, Figure (a) shows that there is a sudden change when the angle of a rotated box closes to $180^{\circ}$ with the 180-degree definition. +Fig. 2 (b). On the one hand, Fig. 3 (b) shows that there is no sudden change of the angle in our piecewise angle definition. Thus, ours is free of the boundary discontinuity problem. On the other hand, rotation equivariance can be easily achieved by group convolutions [52], while rotation invariance relies on a larger capacity network or a larger number of training samples. Hence, we focus on rotation-equivariant features. And then, we are allowed to transform the invariant size estimation into equivariant size estimation by adding a factor $\sin (\theta)$ or $\cos (\theta)$ to the height and width (see Fig. 2 (b)). That way, the box size is enforced to change equivalently when the box orientation changes. Ac + +cordingly, both box size and angle regressions can benefit from rotation-equivariant features. + +# 3.2. Network architecture + +As discussed before, we aim to extract rotation-equivariant features for box size regression and angle regression. Therefore, we adopt rotation-equivariant networks as the backbone to extract rotation-equivariant features. Then, we design a parallel multi-branch architecture based on the proposed piecewise angle definition. The overall architecture of CrackDet is shown in Fig. 2 (c). + +Rotation-equivariant backbone. We first adopt CenterNet [73] as our baseline, which takes an object as a single point (i.e., the center point of the bounding box) and regresses the object size and offset. Then, we re-implement all layers of the fully-convolutional encoder-decoder networks (i.e., up-convolutional residual networks [20]) in CenterNet based on e2cnn [52], named as ReEDNet. Importantly, e2cnn includes rotation-equivariant convolution, rotation-equivariant up-convolution, and pooling, etc. Thanks to the rotation weight sharing and group representations in e2cnn, our ReEDNet takes smaller parameters than the original encoder-decoder networks in CenterNet and enjoys the capability of equivariance. + +![](images/43798c108a9d6fc78b12f450717250e477abbc44e0ec7400827ef807a167d362.jpg) +(a) Branch 1 +Figure 4. A visualization of our angle-based variance voting. Firstly, the red, blue, orange, and teal textboxes are the corresponding standard deviation obtained from branch 1, 2, 3, and 4, respectively. Then, for each oriented bounding box, we select brach $i$ that has the smallest standard deviation to obtain its final detection result. Note that, the red background box is associated with the selected branch. + +![](images/6baa85e4c3754dac30c8e401d3cb750f626808d80587c6cbf7ba8985ac0c6e5a.jpg) +(b) Branch 2 + +![](images/a98286860ac998d5d81dffa88fbd579c4b86c8b3eff7a8e3e38e74220b875f71.jpg) +(d) Branch 4 +(e) After variance voting + +Multi-branch architecture. According to our piecewise angle definition, we define a valid range $[l_i, r_i)$ for each branch $i$ . During the training, we only select the ground truth boxes whose angles fall in the corresponding valid range of one branch. Specifically, referring to the existing 180-degree definition, for an oriented bounding box with angle $\theta$ on the input image, it is valid for branch $i$ when: + +$$ +l _ {i} \leq \theta < r _ {i}. \tag {1} +$$ + +As shown in Fig. 2 (c), the valid ranges of four branches are set to $[0^{\circ}, 45^{\circ})$ , $[45^{\circ}, 90^{\circ})$ , $[90^{\circ}, 135^{\circ})$ and $[135^{\circ}, 180^{\circ})$ , respectively. For branch $i$ , we redefine the ground-truth $\theta$ , height $h$ , and width $w$ in the 180-degree definition into $\theta_{i}$ , $h_{i}$ , and $w_{i}$ , according to our piecewise definition in Fig. 2 (b). Moreover, we aim to estimate the angle and angle confidence together. To this end, each branch of our parallel multi-branch architecture contains three detection heads on the top of the rotation-equivariant backbone for predicting size (i.e., height and width), angle, and angle confidence. Note that, we simply implement each head with a fully-connected layer. That way, our network enables us to predict multiple probability distributions instead of only one angle. For simplicity, we assume each predicted angle from each branch is independent and obeys a single-variate gaussian distribution. The corresponding equation is shown as follows: + +$$ +P _ {\Theta_ {i}} (\theta) = \frac {1}{\sqrt {2 \pi \sigma_ {i} ^ {2}}} e ^ {- \frac {\left(\theta - \theta_ {i} ^ {e}\right) ^ {2}}{2 \sigma_ {i} ^ {2}}}, i \in \{1, 2, 3, 4 \} \tag {2} +$$ + +where $\Theta_{i}$ is the parameters of the angle detection head from branch $i$ , $\theta_{i}^{e}$ denotes the estimated angle from branch $i$ , and standard deviation $\sigma_{i}$ measures the angle confidence of the estimation on branch $i$ . When $\sigma_{i}$ is very close to 0, it means the branch $i$ is extremely confident about the estimated angle. Accordingly, we suppose to have the redefined ground-truth angle $\theta_{i}$ that is in the valid range of branch $i$ . Similarly, $\theta_{i}$ can also be formulated as a gaussian distribution $\mathcal{N}(\theta_i,\sigma_{gt}^2)$ with its standard deviation $\sigma_{gt}\to 0$ . Then, this gaussian distribution can be viewed as: $P_{i}^{gt}(\theta) = \delta (\theta -\theta_{i})$ , where $\delta (\cdot)$ indicates Dirac delta function. + +# 3.3. The proposed MAR Loss + +Suppose the angle of a ground-truth bounding box lies in the valid range of branch $i$ , our multi-branch angle regression loss (MAR Loss) aims to minimize the distribution discrepancy between the predicted angle distribution from branch $i$ and the redefined ground-truth angle distribution. Meanwhile, we expect the other three branches to be able to predict larger variance so that we can filter those inaccurate predictions by variance voting in the inference step. More specifically, we first exploit the Wasserstein distance as the distance metric and minimize the distance between $P_{\Theta_i}(\theta)$ and $P_i^{gt}(\theta)$ . Secondly, we maximize the estimated variance from the other three branches. The equation is shown as: + +$$ +\begin{array}{l} L _ {M A R} = \frac {D _ {W} \left(P _ {\Theta_ {i}} (\theta) \| P _ {i} ^ {g t} (\theta)\right)}{\frac {1}{2} + \sigma_ {i} ^ {2}} - \sum_ {j = 1, j \neq i} ^ {4} \sigma_ {j} ^ {2} \tag {3} \\ = \frac {\| \theta_ {i} ^ {e} - \theta_ {i} \| _ {2} ^ {2} + \sigma_ {i} ^ {2}}{\frac {1}{2} + \sigma_ {i} ^ {2}} - \sum_ {j = 1, j \neq i} ^ {4} \sigma_ {j} ^ {2} \\ \end{array} +$$ + +where $D_W(\cdot)$ is unfolded based on Wasserstein distance [3]: + +$$ +D _ {W} \left(P _ {\Theta_ {i}} (\theta) \| P _ {i} ^ {g t} (\theta)\right) = \| \theta_ {i} ^ {e} - \theta_ {i} \| _ {2} ^ {2} + \sigma_ {i} ^ {2}. \tag {4} +$$ + +Importantly, we exploit Wasserstein distance instead of KL-Divergence [18] for angle regression, because KL-Divergence heavily relies on a non-negligible intersection between two distributions [1, 18]. Note that, when $\theta_i^e$ is estimated accurately, i.e., $\| \theta_i^e -\theta_i\| _2^2\to 0$ , the branch $i$ is expected to predict smaller variance. In addition, a term $\frac{1}{2} +\sigma_i^2$ is added for the following reason: due to the ambiguities in sub-crack orientation, when the angle $\theta_{i}^{e}$ is estimated inaccurately, i.e., $\| \theta_i^e -\theta_i\| _2^2 >\frac{1}{2}$ , we expect the branch $i$ to be able to predict larger variance $\sigma_i^2$ so that $L_{MAR}$ will be lower. Moreover, according to our piecewise angle definition, the size regression loss $L_{\mathrm{size}}$ for the redefined ground truth height $h_i$ and width $w_{i}$ is defined as $L_{\mathrm{size}} = |h_i^e -h_i| + |w_i^e -w_i|$ where $h_i^e$ and $w_{i}^{e}$ indicate the estimated height and width from the valid branch $i$ . Then, + +the overall training objective is shown below: + +$$ +L _ {\text {C r a c k D e t}} = L _ {k} + \lambda_ {o f f} L _ {o f f} + \lambda_ {s i z e} L _ {s i z e} + \lambda_ {M A R} L _ {M A R} \tag {5} +$$ + +where $L_{k}$ and $L_{off}$ are the losses of center point recognition and offset regression by following CenterNet [73]. The hyper-parameter $\lambda_{off}$ , $\lambda_{size}$ , and $\lambda_{MAR}$ are constant factors to control the balance of the above losses. + +# 3.4. Inference + +At the inference step, there are four branches, each of which will predict the height, width, angle, and standard deviation. Thus, we can not directly decide which one should be adopted. Fortunately, with the help of MAR Loss, we design an angle-based variance voting for obtaining the final bounding boxes. Firstly, we extract the peaks in the heatmap by following CenterNet [73]. Then, we get standard deviations $\sigma_{1},\sigma_{2},\sigma_{3},\sigma_{4}$ from four branches at each peak. As shown in Fig. 4, for each peak, we choose branch $i$ that has the minimum estimated variance to obtain the raw output: $i^{*} = \mathrm{argmin}_{i = \{1,2,3,4\}}\sigma_{i}^{2}$ . Next, according to CenterNet [73], we obtain the center (i.e., $x$ and $y$ ) of the oriented bounding box according to the above peaks and the offset branch. Finally, we acquire the final height $h_o = h_{i^*}^e\backslash \Delta (\theta_{i^*}^e)$ , width $w_{o} = w_{i^{*}}^{e}\backslash \Delta (\theta_{i^{*}}^{e})$ , and angle $\theta_o = \Gamma (\theta_{i^*}^e)$ of the bounding box by the following equation: + +$$ +\Gamma \left(\theta_ {i} ^ {e}\right) \triangleq \left\{ \begin{array}{l l} \theta_ {i} ^ {e}, & i = 1, \\ 9 0 ^ {\circ} - \theta_ {i} ^ {e}, & i = 2, \\ 9 0 ^ {\circ} + \theta_ {i} ^ {e}, & i = 3, \\ 1 8 0 ^ {\circ} - \theta_ {i} ^ {e}, & i = 4. \end{array} \right. \tag {6} +$$ + +$$ +\Delta \left(\theta_ {i} ^ {e}\right) \triangleq \left\{ \begin{array}{l l} \cos \left(\theta_ {i} ^ {e}\right), & i = 1, \\ \sin \left(9 0 ^ {\circ} - \theta_ {i} ^ {e}\right), & i = 2, \\ \sin \left(9 0 ^ {\circ} + \theta_ {i} ^ {e}\right), & i = 3, \\ - \cos \left(1 8 0 ^ {\circ} - \theta_ {i} ^ {e}\right), & i = 4. \end{array} \right. \tag {7} +$$ + +where $i$ denotes the index of the selected branch. $\theta_{i}^{e}$ indicate the estimated angle from branch $i$ . It is worth noting that the remaining inference step is the same as CenterNet [73]. + +# 3.5. The proposed datasets + +To promote the research on oriented sub-crack detection, we propose ONPP, ORC, and OCCSD datasets, which are collected from industrial facilities, road pavement, and various buildings in real-world applications. + +ONPP: This dataset is collected from the industrial facilities in real-world applications by a high-resolution camera, comprising 200 images with the size of $7360 \times 4912$ . The width of the crack in the collected images varies from 0.05 mm to $10\mathrm{mm}$ . Moreover, the collected images contain different kinds of noise, such as various concrete types and light intensity. To extend this dataset without compromising the resolution, we directly slice these images into $512 \times 512$ pixels, constructing a final dataset with 3,104 samples. + +ORC: This dataset is collected from road pavement by a mobile vehicle equipped with a high-resolution camera. It contains 300 images with a size of $3489 \times 3489$ . The images of this dataset include various sizes of cracks in the road. It also contains a lot of noise, such as well covers and various appearances of the crosswalk. Similar to ONPP, we slice these images into $512 \times 512$ pixels, constructing a final road pavement dataset with 1,303 samples. + +OCCSD: This dataset is extended from the concrete crack segmentation dataset [71], which contains 458 high-resolution images ( $4032 \times 3024$ pixels). We construct an oriented sub-crack detection dataset with 1,875 samples by slicing these images into $512 \times 512$ pixels and relabeling them from an oriented sub-crack detection perspective. + +# 4. Experiments + +# 4.1. Experiment details + +Evaluation testbed. We conduct experiments for oriented sub-crack detection on three datasets, i.e., ONPP, ORC, and OCCSD. By following the existing crack detection work [8, 31], we divide each dataset into the training, validation, and test set, with the proportion of $8:1:1$ . Moreover, to further verify the effectiveness of CrackDet, we also provide the experiment results on the HRSC2016 dataset [29]. HRSC2016 dataset contains 1,061 samples. Its resolution ranges from $300 \times 300$ to $1500 \times 900$ . For fairness, according to previous works [55, 16, 23], we exploit the training set (436 images) and validation set (181 images) for training, and the testing set (444 images) for evaluation. + +Implementation details. We adopt Adam as the optimizer for training. The total epoch for training the CrackDet is set as 60, 60, 60, and 140 on ONPP, ORC, OCCSD, and HRSC2016 datasets, respectively. For the first three datasets, the initial learning rate is set to $4e - 4$ and reduced by a factor of 10 after 20, and 40 epochs. For the last one, the initial learning rate is $4e - 4$ and reduced by a factor of 10 after 90, and 120 epochs by following the previous work [39]. The batch size for ONPP, ORC, OCCSD, and HRSC2016 is set to 32, 32, 32, and 8, respectively. More importantly, the input resolutions of four datasets are $512 \times 512$ , $512 \times 512$ , $512 \times 512$ , and $768 \times 768$ , respectively. By following the previous work [16], we only use random flipping for data augmentation. We implement our ReEDNet based on ResNet50 and e2cnn [52], and denote it as ReED-R-50. We train all models on four RTX 3090 GPUs and test on a single RTX 3090 GPU. According to our hyper-parameter sensitivity study, we set $\lambda_{off} = 0.1$ , $\lambda_{size} = 0.2$ , and $\lambda_{MAR} = 0.1$ in our next experiments. + +Evaluation metrics. Following previous works [75, 54], we evaluate the performance of CrackDet on oriented crack detection, in terms of Precision and Recall. Moreover, we adopt the mean orientation error (MOE) to verify the effec + +
MethodONPPORC
Precision ↑Recall ↑MOE ↓Precision ↑Recall ↑MOE ↓
FCrack-O2020 [13]0.70370.82630.27850.66310.87140.3397
ReCrack-O2021 [50]0.75060.67190.39870.74320.82540.3282
CABF-FCOS-O2021 [69]0.78930.84130.32630.76300.88510.3744
Y3Crack-O2022[31]0.71340.73790.33500.71030.85140.3632
CenterCrack-O2022 [51]0.72260.61320.35740.77310.81020.3085
Yolo-ViT-O2022 [54]0.77660.83790.35030.79820.84610.2978
CrackDet0.82040.91060.16800.85780.89260.1166
+ +Table 1. Comparison with the state-of-the-art crack detectors on ONPP and ORC. MOE denotes the mean orientation error. + +tiveness of the proposed method for estimating crack orientation. For the detection accuracy on the HRSC2016, we adopt the mean average precision (mAP) as an evaluation criterion, which is consistent with VOC2007 metrics. + +# 4.2. Comparison with crack detection approaches + +In this section, our method is compared to several baselines including state-of-the-art robust crack detection methods: FCrack uses Faster R-CNN to detect the crack patch [13], $\mathrm{Y}^{3}\mathrm{Crack}$ designs a intelligent system based on YOLO v3 for detecting crack [31], CenterCrack proposes a crack detection model by using CenterNet [51], ReCrack utilizes the RetinaNet to find the crack in asphalt pavement [50], CABF-FCOS proposes a one-stage network that can detect various defects including crack [69]. Yolo-ViT applies the Transformer to YOLO v5 for crack detection [54]. Note that, all of these methods can not detect the orientation of cracks. For a fair comparison, we add a branch to these methods for the regression of crack orientation and denote them as FCrack-O, $\mathrm{Y}^{3}\mathrm{Crack - O}$ , CenterCrack-O, ReCrack-O, CABF-FCOS-O, and Yolo-ViT-O respectively. All experiments are independently carried out on ONPP and ORC datasets. Table 1 provides the performance of various approaches from our oriented sub-crack detection perspective for crack detection. Bold denotes the best results. It can be found that the proposed method compares favorably to other competitive crack detection approaches. For example, compared with the Yolo-ViT-O, the precision and recall of CrackDet are 0.8204 and 0.9106 while the Yolo-ViT-O only achieves 0.7766 and 0.8379 on ONPP dataset. For the ORC dataset, CrackDet brings performance gains over Yolo-ViT-O around $6.0\%$ and $4.7\%$ under the precision and recall metrics. Moreover, in terms of MOE metric, the proposed CrackDet achieves consistent improvements on all datasets. The results above together show the necessity that the boundary discontinuity problem and the ambiguities in orientation should be treated properly under the oriented sub-crack detection tasks. + +# 4.3. Discussion on the segmentation-based methods + +In this section, considering the detection-then-segmentation strategy in the real-world application [33, 35], we further verify that the proposed method can serve as a complement to segmentation-based methods. More specifically, for the pixel-level segmentation task, we replace the oriented box annotation with pixel-level annotation for the OCCSD dataset. Table 3 shows that the proposed method can improve the segmentation results, especially for the segmentation model with limited performance. The reason attributes that the proposed method can provide a more fine-grained localization, which can effectively promote the segmentation model being influenced by the noise background. + +# 4.4. Comparison with oriented detection methods + +To further validate our method, we conduct a series of experiments on both crack datasets and HRSC2016, and report quantitative results to verify the effectiveness of CrackDet. As CrackDet shows the purpose of detecting the oriented object, we compare it with existing state-of-the-art oriented detectors with the help of [74], including Oriented RepPoints [23], Oriented R-CNN [55], SASM [19], etc. + +Results on oriented sub-crack detection. Table 2 provides the comparison results over the existing powerful oriented object detectors. We can observe that, first, our CrackDet outperforms the existing detectors by a large margin in terms of Precision. The reason is that the proposed piecewise angle definition can promote the proposed detector to take advantage of rotation-equivariant features and be free of the boundary discontinuity problem. Secondly, CrackDet yields significant and consistent improvement on all three datasets in terms of MOE metric. The reason can be explained as CrackDet can estimate the angle confidence with the help of the multi-branch architecture and MAR Loss. Thus, according to the corresponding angle confidence from each branch, the angle-based variance voting improves the angle estimation results. + +Results on HRSC2016. For fairness, we follow the same settings in Oriented RepPoints [23], and report the results + +
MethodONPPORC
Precision ↑Recall ↑MOE ↓Precision ↑Recall ↑MOE ↓
\(CSL_{ECCV'20}\) [60]0.73090.81460.26330.79430.74370.3285
\(R^3\text{Det}_{AAAI'21}\) [62]0.78350.74330.29770.76970.73120.4280
\(S^2ANet_{TGRS'21}\) [15]0.77730.77730.41330.70680.78680.4640
\(ReDet_{CVPR'21}\) [16]0.78440.80590.27430.71910.86840.3335
Beyond Bounding-Box \(CVPR'21\) [14]0.77900.84680.30030.60760.90390.3441
Oriented R-CNN\(ICCV'21\) [55]0.80320.73550.21720.70870.74160.3470
\(GWD_{ICML'21}\) [63]0.71690.78840.44010.75760.86320.3014
\(KLD_{NeurIPS'21}\) [65]0.74660.64070.34460.74240.90760.2818
\(SASM_{AAAI'22}\) [19]0.74160.83030.30620.76530.90380.3489
Oriented RepPoints \(CVPR'22\) [23]0.78800.87820.30080.78510.91010.3497
CrackDet0.82040.91060.16800.85780.89260.1166
+ +Table 2. We provide a detailed comparison over general oriented detection approaches on ONPP and ORC. + +
Precision ↑Crackformer ICCV'21 [27]JTFN ICCV'21 [10]LIOT TIP'22 [45]
w/o CrackDet0.7810.7540.803
w/ CrackDet0.8140.7980.819
+ +under the VOC2007 metric. Table 4 shows the comparison results. The proposed CrackDet achieves the best performance, even compared with strong FPN-based baselines. Such results further validate that the proposed CrackDet has the potential in more oriented object detection tasks. + +Table 3. Segmentation performance with CrackDet on OCCSD. + +
MethodsBackbonemAP50(07)↑
R3Det AAAI'21 [62]R-101-FPN89.26
R3Det-DCL CVPR'21 [58]R-101-FPN89.46
FPN-CSLECCV'20 [60]R-101-FPN89.62
ReDet CVPR'21 [16]Re-R-5090.46
DAL AAAI'21 [32]R-101-FPN89.77
S2A-NetTGRS'21 [15]R-101-FPN90.17
S2A-Net-DHRec PAMI'22 [37]R-101-FPN90.22
Oriented R-CNN ICCV'21 [55]R-50-FPN90.40
Oriented RepPoints CVPR'22 [23]R-50-FPN90.38
CrackDetReED-R-5090.60
+ +# 4.5. Ablation study + +To examine the contribution of each component in our proposed detector: rotation-equivariant network, the number of valid ranges in the proposed piecewise angle definition, and MAR Loss, a series of ablation experiments are performed on ONPP and ORC datasets. + +Evaluation on the rotation-equivariant network. As expected, CrackDet achieves better results by combining the rotation-equivariant network. Table 5 shows that it improves the precision, recall, and MOE by $0.91\%$ , $1.5\%$ , and $0.023$ with the help of the rotation-equivariant network, respectively. The reason behind this effect is that rotation- + +equivariant features extracted by the rotation-equivariant network are beneficial to size and angle regression within the proposed piecewise angle definition. + +Table 4. We report results on the HRSC2016 test set. mAP(07) denotes detection results under VOC2007 mAP metrics. + +
MethodsONPP
Precision ↑Recall ↑MOE ↓
w/o Rotation-equivariant network [17]0.81130.89520.1914
w/ Rotation-equivariant network0.82040.91060.1680
+ +Evaluation on the number of valid angle ranges. We study the effect of the number of valid angle ranges in our piecewise angle definition. For fairness, we do not add the factor $\sin(\theta)$ or $\cos(\theta)$ for all settings. Table 6 shows the results using one to five ranges. The corresponding results demonstrate that our four ranges definition improves over the single range (baseline) with a 0.14 decrease. As can be noticed, the five ranges do not bring further improvement over the four ranges. By considering the complexity and performance, we choose four ranges as the default setting. + +Table 5. Performance comparisons of with- and without rotation-equivariant backbone on ONPP dataset. + +
NumValid Angle RangesMOE ↓
1[0°, 180°)0.33
2[0°, 90°), [90°, 180°)0.30
3[0°, 60°), [60°, 120°), [120°, 180°)0.21
4[0°, 45°), [45°, 90°), [90°, 135°), [135°, 180°)0.19
5[0°, 36°), [36°, 72°), [72°, 108°), [108°, 144°), [144°, 180°)0.22
+ +Table 6. Results from the different number of ranges on ONPP. + +Evaluation on MAR loss. It is to be expected that Crack-Det combined with MAR Loss achieves consistent improvements on ONPP and ORC datasets, shown in Table 7. The reason is three-fold: (1) KL Loss may fail when two distributions do not have a non-negligible intersection. (2) By learning to estimate high variances for the ambiguous crack orientation during training, our model is able to learn confidence in predicting sub-crack orientation. (3) MAR Loss incorporates learning angle confidence which can potentially help the network to learn discriminative features. The + +learned variances through our MAR Loss are interpretable (see Fig. 4). Our network will output higher variances for ambiguous crack orientation, which can be useful in judging the overall orientation in real-world applications [38]. + +
MethodsONPPORC
Precision ↑Recall ↑MOE ↓Precision ↑Recall ↑MOE ↓
KL Loss [18]0.78650.87180.24250.81490.85370.1986
MAR Loss0.82040.91060.16800.85780.89260.1166
+ +Table 7. Performance comparisons of MAR Loss and KL Loss. + +# 5. Conclusion + +In this work, by starting from a new perspective for crack detection, we propose CrackDet, a first-of-its-kind oriented sub-crack detector that overcomes the boundary discontinuity problem. Meanwhile, we propose the MAR loss to capture the ambiguities in crack orientation by combining CrackDet. The experiments demonstrate that our method outperforms state-of-the-art crack detectors and oriented detectors. On one hand, as we also construct three datasets for oriented sub-crack detection, this work should also give a direction to researchers for designing the crack detector from the oriented sub-crack detection perspective. On the other hand, it highlights the key drawback of existing perspectives of crack detection (i.e., a high labeling cost of pixel-level segmentation, and a coarse crack localization of patch-level classification and object-level detection). We believe this paper has the potential to inspire the following works: (1) multi-model learning [24, 25] for boosting the detection performance in a novel field with limited target data [21], (2) semi-supervised learning and active learning [9] to further alleviate its labeling cost in real-world applications, and (3) oriented sub-crack detection via horizontal annotations [66]. + +Acknowledgements. This work is supported by the National Nature Science Foundation of China under Grants 62073225,62203134, 61972263, 62072315, the National Key RD Program of China under Grants 2020YFA0908700, the Natural Science Foundation of Guangdong Province Outstanding Youth Program under Grants 2019B151502018, Shenzhen Science and Technology Innovation Commission (20220809141216003, JCYJ20210324093808021, JCYJ20220531102817040), the Guangdong "Pearl River Talent Recruitment Program" under Grant 2019ZT08X603, the Guangdong "Pearl River Talent Plan" under Grant 2019JC01X235. + +# References + +[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Proc. ICML, pages 214-223, 2017. 2, 5 + +[2] Seyed Majid Azimi, Eleonora Vig, Reza Bahmanyar, Marco Körner, and Peter Reinartz. Towards multi-class object detection in unconstrained remote sensing imagery. In Proc. ACCV, pages 150-165, 2018. 3 +[3] Raktim Bhattacharya. Data assimilation in optimal transport framework. arXiv preprint arXiv:2005.08670, 2020. 5 +[4] Fu-Chen Chen and Mohammad R Jahanshahi. Nb-cnn: Deep learning-based crack detection using convolutional neural network and naive bayes data fusion. IEEE Transactions on Industrial Electronics, 65(5):4392-4400, 2017. 1, 2 +[5] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proc. ECCV, pages 801-818, 2018. 3 +[6] Minghao Chen, Hongyang Xue, and Deng Cai. Domain adaptation for semantic segmentation with maximum squares loss. In Proc. ICCV, pages 2090-2099, 2019. 3 +[7] Qiang Chen, Yingming Wang, Tong Yang, Xiangyu Zhang, Jian Cheng, and Jian Sun. You only look one-level feature. In Proc. CVPR, pages 13039-13048, 2021. 3 +[8] Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Jie Chen, Zun Liu, and Jianqiang Li. Geometry-aware guided loss for deep crack recognition. In Proc. CVPR, pages 4703-4712, 2022. 1, 2, 6 +[9] Zhuangzhuang Chen, Jin Zhang, Pan Wang, Jie Chen, and Jianqiang Li. When active learning meets implicit semantic data augmentation. In Proc. ECCV, pages 56-72, 2022. 9 +[10] Mingfei Cheng, Kaili Zhao, Xuhong Guo, Yajing Xu, and Jun Guo. Joint topology-preserving and feature-refinement network for curvilinear structure segmentation. In Proc. ICCV, pages 7147-7156, 2021. 8 +[11] Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning roi transformer for oriented object detection in aerial images. In Proc. CVPR, pages 2849-2858, 2019. 3 +[12] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proc. ICCV, pages 6569-6578, 2019. 2 +[13] Fen Fang, Liyuan Li, Ying Gu, Hongyuan Zhu, and Joo-Hwee Lim. A novel hybrid approach for crack detection. PR, 107:107474, 2020. 2, 7 +[14] Zonghao Guo, Chang Liu, Xiaosong Zhang, Jianbin Jiao, Xi-angyang Ji, and Qixiang Ye. Beyond bounding-box: Convex-hull feature adaptation for oriented and densely packed object detection. In Proc. CVPR, 2021. 8 +[15] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2021. 8 +[16] Jiaming Han, Jian Ding, Nan Xue, and Gui-Song Xia. Redet: A rotation-equivariant detector for aerial object detection. In Proc. CVPR, pages 2786-2795, 2021. 2, 3, 6, 8 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In Proc. ECCV, pages 630-645, 2016. 2, 8 +[18] Yihui He, Chenchen Zhu, Jianren Wang, Marios Savvides, and Xiangyu Zhang. Bounding box regression with uncertainty for accurate object detection. In Proc. CVPR, pages 2888-2897, 2019. 5, 9 + +[19] Liping Hou, Ke Lu, Jian Xue, and Yuqiu Li. Shape-adaptive selection and measurement for oriented object detection. In Proc. AAAI, 2022. 7, 8 +[20] Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, et al. Speed/accuracy trade-offs for modern convolutional object detectors. In Proc. CVPR, pages 7310-7311, 2017. 4 +[21] Jianqiang Li, Zhuangzhuang Chen, Jie Chen, and Qiuzhen Lin. Diversity-sensitive generative adversarial network for terrain mapping under limited human intervention. IEEE Transactions on Cybernetics, 51(12):6029-6040, 2020. 9 +[22] Kai Li, Bo Wang, Yingjie Tian, and Zhiquan Qi. Fast and accurate road crack detection based on adaptive cost-sensitive loss function. IEEE Transactions on Cybernetics, 2021. 1, 3 +[23] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented leptopoints for aerial object detection. In Proc. CVPR, pages 1829-1838, 2022. 2, 3, 6, 7, 8 +[24] Wenrui Li and Xiaopeng Fan. Image-text alignment and retrieval using light-weight transformer. In Proc. ICASSP, pages 4758-4762, 2022. 9 +[25] Wenrui Li, Zhengyu Ma, Liang-Jian Deng, Xiaopeng Fan, and Yonghong Tian. Neuron-based spiking transmission and reasoning network for robust image-text retrieval. IEEE TCSVT, 2022. 9 +[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dálár. Focal loss for dense object detection. In Proc. ICCV, pages 2980-2988, 2017. 2 +[27] Huajun Liu, Xiangyu Miao, Christoph Mertz, Chengzhong Xu, and Hui Kong. Crackformer: Transformer network for fine-grained crack detection. In Proc. ICCV, pages 3783-3792, 2021. 8 +[28] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In Proc. ECCV, pages 21-37, 2016. 2 +[29] Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang. A high resolution optical satellite image dataset for ship recognition and some new baselines. In International conference on pattern recognition applications and methods, volume 2, pages 324-331, 2017. 3, 6 +[30] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proc. CVPR, pages 3431-3440, 2015. 3 +[31] Duo Ma, Hongyuan Fang, Niannian Wang, Chao Zhang, Jiaxiu Dong, and Haobang Hu. Automatic detection and counting system for pavement cracks based on pcgan and yolo-mf. IEEE Transactions on Intelligent Transportation Systems, 2022. 1, 2, 6, 7 +[32] Qi Ming, Zhiqiang Zhou, Lingjuan Miao, Hongwei Zhang, and Linhao Li. Dynamic anchor learning for arbitrary-oriented object detection. In Proc. AAAI, pages 2355-2363, 2021. 8 +[33] Mayank Mishra, Vipul Jain, Saurabh Kumar Singh, and Damodar Maity. Two-stage method based on the you only look once framework and image segmentation for crack detection in concrete structures. Architecture, Structures and Construction, pages 1-18, 2022. 2, 3, 7 + +[34] Hafiz Suliman Munawar, Ahmed WA Hammad, Assed Haddad, Carlos Alberto Pereira Soares, and S Travis Waller. Image-based crack detection methods: A review. Infrastructures, 6(8):115, 2021. 1 +[35] Nhung Hong Thi Nguyen, Stuart Perry, Don Bone, Ha Thanh Le, and Thuy Thi Nguyen. Two-stage convolutional neural network for road crack detection and segmentation. Expert Systems with Applications, 186:115718, 2021. 2, 3, 7 +[36] Son Dong Nguyen, Thai Son Tran, Van Phuc Tran, Hyun Jong Lee, Md Piran, Van Phuc Le, et al. Deep learning-based crack detection: A survey. International Journal of Pavement Research and Technology, 2022. 1, 2 +[37] Guangtao Nie and Hua Huang. Multi-oriented object detection in aerial images with double horizontal rectangles. IEEE TPAMI, 2022. 2, 8 +[38] Sijun Niu and Vikas Srivastava. Simulation trained cnn for accurate embedded crack length, location, and orientation prediction from ultrasound measurements. International Journal of Solids and Structures, 242:111521, 2022. 2, 9 +[39] Xingjia Pan, Yuqiang Ren, Kekai Sheng, Weiming Dong, Haolei Yuan, Xiaowei Guo, Chongyang Ma, and Changsheng Xu. Dynamic refinement network for oriented and densely packed object detection. In Proc. CVPR, pages 11207-11216, 2020. 2, 6 +[40] Wen Qian, Xue Yang, Silong Peng, Junchi Yan, and Yue Guo. Learning modulated loss for rotated object detection. In Proc. AAAI, pages 2458-2466, 2021. 3 +[41] Zhong Qu, Wen Chen, Shi-Yan Wang, Tu-Ming Yi, and Ling Liu. A crack detection algorithm for concrete pavement based on attention mechanism and multi-features fusion. IEEE Transactions on Intelligent Transportation Systems, 2021. 1 +[42] Marcos Quintana, Juan Torres, and José Manuel Menéndez. A simplified computer vision system for road surface inspection and maintenance. IEEE Transactions on Intelligent Transportation Systems, 17(3):608-619, 2015. 2 +[43] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proc. CVPR, pages 779-788, 2016. 2 +[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Proc. NeurIPS, 2015. 2 +[45] Tianyi Shi, Nicolas Boutry, Yongchao Xu, and Thierry Géraud. Local intensity order transformation for robust curvilinear object segmentation. IEEE Transactions on Image Processing, 31:2557-2569, 2022. 8 +[46] Yong Shi, Limeng Cui, Zhiquan Qi, Fan Meng, and Zhensong Chen. Automatic road crack detection using random structured forests. IEEE Transactions on Intelligent Transportation Systems, 17(12):3434-3445, 2016. 3 +[47] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2 +[48] Weidong Song, Guohui Jia, Hong Zhu, Di Jia, and Lin Gao. Automated pavement crack damage detection using deep multiscale convolutional features. Journal of Advanced Transportation, 2020, 2020. 2, 3 + +[49] Xinzi Sun, Yuanchang Xie, Liming Jiang, Yu Cao, and Benyuan Liu. Dma-net: Deeplab with multi-scale attention for pavement crack segmentation. IEEE Transactions on Intelligent Transportation Systems, 23(10):18392-18403, 2022. 1 +[50] Van Phuc Tran, Thai Son Tran, and Hyun Jong Lee. One stage detector (retinanet)-based crack detection for asphalt pavements considering pavement distresses and surface objects. Journal of Civil Structural Health Monitoring, pages 205-222, 2021. 3, 7 +[51] Ruoxin Wang and Chi Fai Cheung. Centernet-based defect detection for additive manufacturing. Expert Systems with Applications, 188:116000, 2022. 3, 7 +[52] Maurice Weiler and Gabriele Cesa. General e (2)-equivariant steerable cnns. In Proc. NeurIPS, 2019. 4, 6 +[53] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In Proc. CVPR, pages 3974-3983, 2018. 3 +[54] Xuezhi Xiang, Zhiyuan Wang, and Yulong Qiao. An improved yolov5 crack detection method combined with transformer. IEEE Sensors Journal, 2022. 1, 2, 6, 7 +[55] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented r-cnn for object detection. In Proc. ICCV, pages 3520-3529, 2021. 3, 6, 7, 8 +[56] Yongchao Xu, Mingtao Fu, Qimeng Wang, Yukang Wang, Kai Chen, Gui-Song Xia, and Xiang Bai. Gliding vertex on the horizontal bounding box for multi-oriented object detection. IEEE TPAMI, 43(4):1452-1459, 2020. 3 +[57] Fan Yang, Lei Zhang, Sijia Yu, Danil Prokhorov, Xue Mei, and Haibin Ling. Feature pyramid and hierarchical boosting network for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 21(4):1525-1535, 2019. 3 +[58] Xue Yang, Liping Hou, Yue Zhou, Wentao Wang, and Junchi Yan. Dense label encoding for boundary discontinuity free rotation detection. In Proc. CVPR, 2021. 2, 3, 8 +[59] Xincong Yang, Heng Li, Yantao Yu, Xiaochun Luo, Ting Huang, and Xu Yang. Automatic pixel-level crack detection and measurement using fully convolutional network. Computer-Aided Civil and Infrastructure Engineering, 33(12):1090-1109, 2018. 3 +[60] Xue Yang and Junchi Yan. Arbitrary-oriented object detection with circular smooth label. In Proc. ECCV, pages 677-694, 2020. 3, 8 +[61] Xue Yang and Junchi Yan. On the arbitrary-oriented object detection: Classification based approaches revisited. IJCV, 2022. 2 +[62] Xue Yang, Junchi Yan, Ziming Feng, and Tao He. R3det: Refined single-stage detector with feature refinement for rotating object. In Proc. AAAI, 2021. 3, 8 +[63] Xue Yang, Junchi Yan, Qi Ming, Wentao Wang, Xiaopeng Zhang, and Qi Tian. Rethinking rotated object detection with gaussian Wasserstein distance loss. In Proc. ICML, 2021. 3, 8 +[64] X Yang, J Yang, J Yan, Y Zhang, T Zhang, Z Guo, X Sun, and K SCRDet Fu. Towards more robust detection for small, + +cluttered and rotated objects. In Proc. ICCV, volume 27, pages 8232-8241, 2019. 3 +[65] Xue Yang, Xiaojiang Yang, Jirui Yang, Qi Ming, Wentao Wang, Qi Tian, and Junchi Yan. Learning high-precision bounding box for rotated object detection via kullback-leibler divergence. In Proc. NeurIPS, 2021. 3, 8 +[66] Xue Yang, Gefan Zhang, Wentong Li, Yue Zhou, Xuehui Wang, and Junchi Yan. H2rbox: Horizontal box annotation is all you need for oriented object detection. In Proc. ICLR, 2022. 9 +[67] Dawen Yu, Shunping Ji, Xue Li, Zhaode Yuan, and Chaoyong Shen. Earthquake crack detection from aerial images using a deformable convolutional neural network. IEEE Transactions on Geoscience and Remote Sensing, 2022. 1 +[68] Hong-Xing Yu, Jiajun Wu, and Li Yi. Rotationally equivariant 3d object detection. In Proc. CVPR, pages 1456-1464, 2022. 3 +[69] Jianbo Yu, Xun Cheng, and Qingfeng Li. Surface defect detection of steel strips based on anchor-free network with channel attention and bidirectional feature fusion. IEEE Transactions on Instrumentation and Measurement, 2021. 7 +[70] Gongjie Zhang, Shijian Lu, and Wei Zhang. Cad-net: A context-aware detection network for objects in remote sensing imagery. IEEE Transactions on Geoscience and Remote Sensing, 57(12):10015-10024, 2019. 3 +[71] Lei Zhang, Fan Yang, Yimin Daniel Zhang, and Ying Julie Zhu. Road crack detection using deep convolutional neural network. In Proc. ICIP, pages 3708-3712, 2016. 6 +[72] Xiaoyu Zhao, Wenlian Huang, Jie Chen, Zhuangzhuang Chen, and Jianqiang Li. Automatic thin crack segmentation with deep context aggregation network. In Proc. ICARM, pages 206-212, 2022. 3 +[73] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 2, 4, 6 +[74] Yue Zhou, Xue Yang, Gefan Zhang, Jiabao Wang, Yanyi Liu, Liping Hou, Xue Jiang, Xingzhao Liu, Junchi Yan, Chengqi Lyu, et al. Mmrotate: A rotated object detection benchmark using pytorch. In Proc. ACMM, pages 7331-7334, 2022. 7 +[75] Junqing Zhu, Jingtao Zhong, Tao Ma, Xiaoming Huang, Weiguang Zhang, and Yang Zhou. Pavement distress detection using convolutional neural networks with images captured via uav. Automation in Construction, 2022. 2, 6 +[76] Qin Zou, Yu Cao, Qingquan Li, Qingzhou Mao, and Song Wang. Cracktree: Automatic crack detection from pavement images. Pattern Recognition Letters, 33(3):227-238, 2012. 1, 3 +[77] Qin Zou, Zheng Zhang, Qingquan Li, Xianbiao Qi, Qian Wang, and Song Wang. Deepcrack: Learning hierarchical convolutional features for crack detection. IEEE TIP, 28(3):1498-1512, 2018. 1, 2, 3 \ No newline at end of file diff --git a/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/images.zip b/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..52a13866e717a0099e660166034f4307c00710a5 --- /dev/null +++ b/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ed9fa5c418dce502237dd9ce40d832ea770b148ed08ecd295a35848c53289e2 +size 567727 diff --git a/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/layout.json b/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eb0fff9a7cc02af5ae819768ee13659c34b018d8 --- /dev/null +++ b/2023/The Devil is in the Crack Orientation_ A New Perspective for Crack Detection/layout.json @@ -0,0 +1,10048 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 536, + 121 + ], + "type": "text", + "content": "The Devil is in the Crack Orientation: A New Perspective for Crack Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 143, + 529, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 143, + 529, + 171 + ], + "spans": [ + { + "bbox": [ + 64, + 143, + 529, + 171 + ], + "type": "text", + "content": "Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Guanming Zhu, Zun Liu, Jie Chen, Jianqiang Li Shenzhen University, Shenzhen, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 173, + 480, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 173, + 480, + 185 + ], + "spans": [ + { + "bbox": [ + 113, + 173, + 480, + 185 + ], + "type": "text", + "content": "{chenzhuangzhuang2016，2018101068，zhuguanming2022}@email.szu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 175, + 187, + 418, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 187, + 418, + 199 + ], + "spans": [ + { + "bbox": [ + 175, + 187, + 418, + 199 + ], + "type": "text", + "content": "{jin.zhang, chenjie, zunliu, lijq}@szu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 253, + 290, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 290, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 290, + 553 + ], + "type": "text", + "content": "Cracks are usually curve-like structures that are the focus of many computer-vision applications (e.g., road safety inspection and surface inspection of the industrial facilities). The existing pixel-based crack segmentation methods rely on time-consuming and costly pixel-level annotations. And the object-based crack detection methods exploit the horizontal box to detect the crack without considering crack orientation, resulting in scale variation and intraclass variation. Considering this, we provide a new perspective for crack detection that models the cracks as a series of sub-cracks with the corresponding orientation. However, the vanilla adaptation of the existing oriented object detection methods to the crack detection tasks will result in limited performance, due to the boundary discontinuity issue and the ambiguities in sub-crack orientation. In this paper, we propose a first-of-its-kind oriented sub-crack detector, dubbed as CrackDet, which is derived from a novel piecewise angle definition, to ease the boundary discontinuity problem. And then, we propose a multi-branch angle regression loss for learning sub-crack orientation and variance together. Since there are no related benchmarks, we construct three fully annotated datasets, namely, ORC, ONPP, and OCCSD, which involve various cracks in road pavement and industrial facilities. Experiments show that our approach outperforms state-of-the-art crack detectors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 588, + 128, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 588, + 128, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 588, + 128, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 694 + ], + "type": "text", + "content": "Crack is a type of defect that appears on the surfaces of many physical structures, e.g., the road pavement [41], the industrial facilities [8]. Inspecting and repairing cracks are crucial tasks for avoiding expansion of the harm to these structures [77]. Besides, a recent study reveals that recognizing the surface cracks caused by earthquakes in time can avoid secondary damage to people's properties" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 304, + 239, + 547, + 418 + ], + "blocks": [ + { + "bbox": [ + 317, + 227, + 350, + 237 + ], + "lines": [ + { + "bbox": [ + 317, + 227, + 350, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 227, + 350, + 237 + ], + "type": "text", + "content": "Images" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 362, + 227, + 423, + 237 + ], + "lines": [ + { + "bbox": [ + 362, + 227, + 423, + 237 + ], + "spans": [ + { + "bbox": [ + 362, + 227, + 423, + 237 + ], + "type": "text", + "content": "Segmentation" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 434, + 227, + 476, + 237 + ], + "lines": [ + { + "bbox": [ + 434, + 227, + 476, + 237 + ], + "spans": [ + { + "bbox": [ + 434, + 227, + 476, + 237 + ], + "type": "text", + "content": "Detection" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 504, + 227, + 528, + 237 + ], + "lines": [ + { + "bbox": [ + 504, + 227, + 528, + 237 + ], + "spans": [ + { + "bbox": [ + 504, + 227, + 528, + 237 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 239, + 547, + 418 + ], + "lines": [ + { + "bbox": [ + 304, + 239, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 547, + 418 + ], + "type": "image", + "image_path": "353ce6bab1dc650c20cb1584d0fcd775c70a88ecf20b729933de78f7d1ae718b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 418, + 541, + 429 + ], + "lines": [ + { + "bbox": [ + 318, + 418, + 541, + 429 + ], + "spans": [ + { + "bbox": [ + 318, + 418, + 541, + 429 + ], + "type": "text", + "content": "(a) Visualization of different perspectives for crack detection" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 312, + 434, + 416, + 494 + ], + "blocks": [ + { + "bbox": [ + 312, + 434, + 416, + 494 + ], + "lines": [ + { + "bbox": [ + 312, + 434, + 416, + 494 + ], + "spans": [ + { + "bbox": [ + 312, + 434, + 416, + 494 + ], + "type": "image", + "image_path": "ebdacb2e49e1b25ae2264cb8df0e79f084d4aa4945f46ef7e5c2a3f4fbf3371d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 497, + 419, + 507 + ], + "lines": [ + { + "bbox": [ + 318, + 497, + 419, + 507 + ], + "spans": [ + { + "bbox": [ + 318, + 497, + 419, + 507 + ], + "type": "text", + "content": "(b) Boundary discontinuity" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 508, + 545, + 563 + ], + "lines": [ + { + "bbox": [ + 306, + 508, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 508, + 545, + 563 + ], + "type": "text", + "content": "Figure 1. Cracks tend to be curve-like structures [77] in real-world applications. (a) We compare our orientation-based perspective with others for crack detection. (b) Boundary discontinuity issue in the existing 180-degree definition. (c) The sub-crack orientation is ambiguous in some cases due to the irregular sub-crack shape." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 433, + 430, + 538, + 494 + ], + "blocks": [ + { + "bbox": [ + 433, + 430, + 538, + 494 + ], + "lines": [ + { + "bbox": [ + 433, + 430, + 538, + 494 + ], + "spans": [ + { + "bbox": [ + 433, + 430, + 538, + 494 + ], + "type": "image", + "image_path": "38ce88cbf2af0d07252c02137be66dbeaa620a0a9172513e275408cd4ffd606b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 497, + 545, + 507 + ], + "lines": [ + { + "bbox": [ + 436, + 497, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 436, + 497, + 545, + 507 + ], + "type": "text", + "content": "(c) Ambiguities in orientation" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 590, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 638 + ], + "type": "text", + "content": "and lives [67]. However, manual inspection is costly, time-consuming, and prone to human error [34]. Hence, automatic image-based crack detection plays an essential role in a real-world application of crack identification [22, 67]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 713 + ], + "type": "text", + "content": "Currently, image-based crack detection methods can be roughly divided into three categories [36]: patch-level classification [4, 8], object-level detection [54, 31], and pixel-level crack segmentation [76, 49]. Although the first two methods take less labeling cost than the last one, they can only provide a coarse crack localization. Moreover, we are" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 227, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 227, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 227, + 712 + ], + "type": "text", + "content": "*Corresponding Author Email: chenjie@szu.edu.cn" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6653" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "motivated by the fact that a crack is a line object that has a certain width from a local perspective [77]. Then, we are encouraged to start from an orientation-based perspective that models the cracks as a series of oriented sub-cracks (see Fig. 1 (a)). The advantage of this perspective is three-fold: (1) eliminating the scale variation, (2) reducing the intraclass variation, and (3) taking less labeling cost than the segmentation methods. More specifically, in the case of our practice, the price and time cost of the pixel-level annotations are " + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "inline_equation", + "content": "15 \\times" + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "inline_equation", + "content": "30 \\times" + }, + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": " higher than our oriented sub-crack annotations, respectively. Besides, considering real-world applications, existing works [33, 35] show great power by taking a detection-then-segmentation strategy, i.e., it first identifies and localizes cracks by a detector and then obtains crack details by segmentation-based methods. Hence, the detector based on our perspective can serve as a complement to segmentation-based methods, as it can provide a more fine-grained localization (see Fig. 1 a)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 291, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 518 + ], + "type": "text", + "content": "However, it is difficult to directly adopt the state-of-art oriented detectors [16, 39, 23, 37] to crack detection. The reason is two-fold: (1) several existing works (e.g., ReDet [16], DRN [39]) simply ignore the boundary discontinuity problem, and (2) although the remaining works make great efforts on the above problem, these works still suffer from ambiguity issue in the sub-crack orientation. Among existing oriented detection methods, those detectors based on five parameters " + }, + { + "bbox": [ + 46, + 291, + 289, + 518 + ], + "type": "inline_equation", + "content": "(x,y,h,w,\\theta)" + }, + { + "bbox": [ + 46, + 291, + 289, + 518 + ], + "type": "text", + "content": " dominate [58], where " + }, + { + "bbox": [ + 46, + 291, + 289, + 518 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 291, + 289, + 518 + ], + "type": "text", + "content": " indicates the center point and " + }, + { + "bbox": [ + 46, + 291, + 289, + 518 + ], + "type": "inline_equation", + "content": "(h,w,\\theta)" + }, + { + "bbox": [ + 46, + 291, + 289, + 518 + ], + "type": "text", + "content": " denotes height, width, and angle, respectively. As shown in Fig. 1 (b), the boundary discontinuity problem indicates that small rotations for oriented objects around the angle period's boundary result in a sudden sharp increase of loss [61]. Moreover, Fig. 1 (c) provides some cases to illustrate the inherent ambiguities in sub-crack orientation. When the growth direction of a crack changes frequently, the orientation of the sub-crack is unclear to judge, making it hard to learn regression functions for estimating orientation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": "This paper starts from a new perspective for crack detection that models the cracks as a series of oriented subcracks. After that, we adopt CenterNet [73] as our baseline due to its simplicity. Next, we propose a new piecewise angle definition to ease the boundary discontinuity issue. Based on our piecewise definition, we propose the CrackDet that contains a parallel multi-branch architecture, each of which is responsible for predicting a fixed range of angles. Then, to address the ambiguities in sub-crack orientation, we propose a novel multi-branch angle regression loss, namely MAR Loss, which is based on Wasserstein distances [1] for both angle regression and angle variance estimation. The advantage of coupling CrackDet with MAR Loss is three-fold: (1) the ambiguities in sub-crack orientation can be captured to a certain extent. (2) The learned variance from each branch is beneficial during the post-process. We" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "type": "text", + "content": "acquire the estimated variances from each branch and obtain the detection results by variance voting. (3) Since crack orientation is significant in a real-world application [38], the learned variances reveal the level of confidence in the angle prediction, it can potentially help to judge the overall crack orientation. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 156, + 545, + 358 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 304, + 156, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 203 + ], + "type": "text", + "content": "(1) To the best of our knowledge, we are the first ones to start from an oriented sub-crack detection perspective for crack detection by modeling cracks as a series of sub-cracks with their inherent orientations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 204, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 251 + ], + "type": "text", + "content": "(2) Propose a first-of-its-kind oriented sub-crack detector, called as CrackDet, which stems from the proposed piecewise angle definition (see Sec. 3.1), to ease the boundary discontinuity issue." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 251, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 299 + ], + "type": "text", + "content": "(3) Propose a multi-branch angle regression loss that enables CrackDet to capture ambiguities in sub-crack orientation via jointly learning sub-crack orientation and the corresponding orientation variance (see Sec. 3.2 and 3.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 300, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 358 + ], + "type": "text", + "content": "(4) A tale of three datasets has been constructed, namely, ORC, ONPP, and OCCSD. These datasets contain orientation annotations of the sub-cracks in real-world industrial facilities and road pavements, to facilitate the research about oriented sub-crack detection." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 370, + 390, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 390, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 390, + 383 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 390, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 437 + ], + "type": "text", + "content": "By following the previous study [36], the vision-based crack detection can be roughly clustered into three categories: patch-level classification [4, 8], object-level detection [54, 31], and pixel-level segmentation [77, 48]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 438, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 582 + ], + "type": "text", + "content": "Patch-level crack classification. Recently, deep learning-based methods are more powerful than the traditional methods (e.g., HOG [42], LBP[42], Gaussian process, SVM [42]) in the patch-level crack classification tasks [36]. Among deep learning-based methods, variants of convolutional neural network (CNN) followed by a fully connected (FC) layer are used to recognize patch-level crack samples. Typically, ResNet [17] and VGG [47] are popular backbones among existing backbones. Moreover, focal loss [26] focuses on the hard samples to learn discriminative features. However, patch-level methods fail in providing orientation and the fine-grained localization of crack objects [36]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Object-level crack detection. For object-level crack detection, a bounding box is exploited to describe the location and the coarse size of a crack object. Most existing crack detectors mainly adopt the general object detectors, such as Faster R-CNN [44], RetinaNet [26], Yolo [43], CenterNet [12], and SSD [28]. Specifically, Fen et al. [13] first design a hybrid approach that uses Faster R-CNN to detect the crack patch. Then, as the crack orientation is important in the post-process, a deep CNN is exploited to obtain the orientation of each detected patch. Ma et al. [31] design an automatic crack detection system by using YOLO v3 [75]. To" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6654" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "detect various cracks, Wang et al. [51] propose a crack detection model by using CenterNet to extract crack features. Moreover, a RetinaNet-based crack detector is proposed to detect the crack in asphalt pavement [50]. Unfortunately, as shown in Fig. 1, the object-level crack detection methods suffer from scale variation and intra-class variation. Moreover, it also can not directly provide the orientation and the fine-grained localization of cracks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 288, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 288, + 397 + ], + "type": "text", + "content": "Pixel-level crack segmentation. Since those horizontal bounding boxes can not describe the exact shape of the crack, segmentation-based methods are proposed to detect cracks at the pixel level. Yang et al. [59] design an automatic crack segmentation pipeline by using fully convolutional networks [30] and achieve a noticeable performance gain over traditional approaches [76, 46]. By coupling with Deeplab v3+ [5], CrackSeg [48] is proposed to learn rich features for different scales. Moreover, aiming to learn high-level crack representations, DeepCrack [77] is proposed to capture the line structures by feature fusion. Considering different receptive fields, FPHBN [57] makes great efforts for crack detection by aggregating pyramid features. Besides, by addressing the imbalance problem, Li et al. [22] propose an adaptive weighted cross-entropy loss function, which serves as a complement to the existing methods. However, these segmentation-based methods heavily rely on high-quality labeled datasets [6, 72], which involves a high labeling cost." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 400, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 555 + ], + "type": "text", + "content": "Compared to the above methods, we start from a new perspective that turns crack detection into oriented subcrack detection. Based on this perspective, we are allowed to provide a fine-grained localization of crack objects (See details in Fig. 1), instead of a coarse localization by the patch-level classification or object-level detection methods. Hence, considering the existing detectionthen-segmentation strategy [33, 35], the detector based on our perspective can serve as a better candidate for the segmentation-based methods. Besides, when considering a dynamic field without historical human supervision, our perspective can greatly ease the labeling cost than the pixellevel segmentation methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Oriented object detection. Since our oriented sub-crack detection perspective belongs to the oriented object detection category, we give a brief review of related research on this scope. Recently, many powerful rotation detectors are mainly derived from horizontal object detectors by adding an orientation regression branch [23]. For example, ICN [2], ROI-Transformer [11], SCRDet [64], CAD-Net [70], CSL[60], " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^3\\mathrm{Det}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " [62], ReDet [16], Oriented R-CNN [55], and DCL [58] bring satisfied detection results on DOTA dataset [53] and HRSC2016 dataset [29]. Gliding Vertex [56] and RSDet [40] share the same purpose of considering quadrilateral regression prediction. Note that, the boundary discontinuity problem is caused by the periodicity of the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 264 + ], + "type": "text", + "content": "angle when dealing with the angel-based orientation estimation. Then, Yang et al. [60] take a new perspective that turns the angular regression tasks into the angular classification tasks [58]. Moreover, Oriented RepPoints [23] utilize the flexible adaptive points as the box representation to achieve oriented object detection. Unlike these methods, our CrackDet addresses the boundary discontinuity problem via the new piecewise angle definition, which is motivated by the divide-and-conquer strategy. Besides, although KLD [65] and GWD [63] also start from a distribution perspective, it converts the center point, height, width, and orientation of a rotated bounding box as a 2-D Gaussian distribution, and does not focus on learning angle variance. Instead, the proposed CrackDet assumes that the ground-truth orientation obeys a single-variate gaussian distribution, and aims to learn the sub-crack orientation and its variance together." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 274, + 451, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 274, + 451, + 286 + ], + "spans": [ + { + "bbox": [ + 305, + 274, + 451, + 286 + ], + "type": "text", + "content": "3. Our method and datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 293, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 377 + ], + "type": "text", + "content": "The overall framework is shown in Fig. 2. We first introduce the piecewise angle definition in Sec. 3.1, and the network architecture of CrackDet in Sec. 3.2. Then, we give the details of the proposed MAR Loss for estimating ambiguities in sub-crack orientation in Sec. 3.3. Finally, we present the inference step of CrackDet with estimated angle variances in Sec. 3.4, and the proposed datasets in Sec. 3.5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 384, + 448, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 448, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 448, + 396 + ], + "type": "text", + "content": "3.1. Piecewise angle definition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 403, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 450 + ], + "type": "text", + "content": "In this section, we recap the boundary discontinuity problem and inconsistency regression problem in the existing 180-degree five-parameter definition. Then, we discuss the details of the proposed piecewise angle definition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "text", + "content": "Rethinking the five-parameter definition. Fig. 2 (a) shows that the existing five-parameter definition [53] with the 180-degree angular range includes " + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "inline_equation", + "content": "(x, y, h, w, \\theta)" + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "text", + "content": " indicates the center point and " + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "inline_equation", + "content": "[h, w, \\theta]" + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "text", + "content": " denotes height, width, and angle, respectively. Then, the boundary discontinuity problem refers to a sharp increase of loss at the boundary caused by the periodicity of the angle. As shown in Fig. 3 (a), it can be further explained that when the angle of a rotated box is close to " + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "text", + "content": ", the angle under the five-parameter definition will suddenly change from " + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 304, + 451, + 545, + 714 + ], + "type": "text", + "content": ", resulting a sharp increase of angle regression loss. Moreover, the inconsistency regression problem can be attributed to the rotation symmetry in object detection, which includes the equivariant orientation and invariant shape estimation for a rotated bounding box [68]. Specifically speaking, rotation-equivariant features are desired for orientation estimation. In contrast, the shape of the bounding box does not change after a rotation. Thus, rotation-invariant features are needed for predicting its invariant height and width [16]. Discussion on our piecewise angle definition. We are inspired by the divide-and-conquer strategy [7] and propose the following piecewise angle definition, shown in" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6655" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 160, + 187 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 160, + 187 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 160, + 187 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 160, + 187 + ], + "type": "image", + "image_path": "f78265a71a5b0e24a85cc2a82660ea83321cd092bfb6563b85b68983066bcad5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 163, + 70, + 536, + 188 + ], + "blocks": [ + { + "bbox": [ + 163, + 70, + 536, + 188 + ], + "lines": [ + { + "bbox": [ + 163, + 70, + 536, + 188 + ], + "spans": [ + { + "bbox": [ + 163, + 70, + 536, + 188 + ], + "type": "image", + "image_path": "f1c1fc28a8475903ae0cf41fc6f4a5ff7337a4b1b39025cc7c7dab6598b69616.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 205, + 527, + 326 + ], + "blocks": [ + { + "bbox": [ + 57, + 190, + 221, + 205 + ], + "lines": [ + { + "bbox": [ + 57, + 190, + 221, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 190, + 221, + 205 + ], + "type": "text", + "content": "(a) The five-parameter definition" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 261, + 190, + 476, + 205 + ], + "lines": [ + { + "bbox": [ + 261, + 190, + 476, + 205 + ], + "spans": [ + { + "bbox": [ + 261, + 190, + 476, + 205 + ], + "type": "text", + "content": "(b) The proposed piecewise angle definition" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 205, + 527, + 326 + ], + "lines": [ + { + "bbox": [ + 57, + 205, + 527, + 326 + ], + "spans": [ + { + "bbox": [ + 57, + 205, + 527, + 326 + ], + "type": "image", + "image_path": "7819f9a0c288ed74cbd4b303221544d3c03bdd41a301475d5f26b358f4b397e3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 327, + 420, + 341 + ], + "lines": [ + { + "bbox": [ + 171, + 327, + 420, + 341 + ], + "spans": [ + { + "bbox": [ + 171, + 327, + 420, + 341 + ], + "type": "text", + "content": "(c) Overall architecture of the proposed CrackDet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 343, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 343, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 545, + 399 + ], + "type": "text", + "content": "Figure 2. (a) The existing five-parameter definition with long-side definition. (b) Our piecewise angle definition is inspired by the divide-and-conquer strategy and built upon the above five-parameter definition. It breaks down the existing 180-degree regression tasks into four sub-tasks for addressing the discontinuity problem. (c) Based on our piecewise angle definition, we present an overview of CrackDet that can estimate standard deviations (std) along with angle regression for capturing ambiguities in the sub-crack orientation. Note that, the heatmap branch and offset branch are responsible for the center (i.e., " + }, + { + "bbox": [ + 46, + 343, + 545, + 399 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 343, + 545, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 343, + 545, + 399 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 343, + 545, + 399 + ], + "type": "text", + "content": ") of the predicted bounding box as in the original CenterNet." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 54, + 424, + 278, + 467 + ], + "blocks": [ + { + "bbox": [ + 54, + 424, + 278, + 467 + ], + "lines": [ + { + "bbox": [ + 54, + 424, + 278, + 467 + ], + "spans": [ + { + "bbox": [ + 54, + 424, + 278, + 467 + ], + "type": "image", + "image_path": "47a5c9a2891c667b5976e9ce2a1dc021e9b6f8418d068c3c3caa1686a2b11413.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 474, + 164, + 485 + ], + "lines": [ + { + "bbox": [ + 55, + 474, + 164, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 164, + 485 + ], + "type": "text", + "content": "(a) The 180-degree definition" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 174, + 474, + 279, + 485 + ], + "lines": [ + { + "bbox": [ + 174, + 474, + 279, + 485 + ], + "spans": [ + { + "bbox": [ + 174, + 474, + 279, + 485 + ], + "type": "text", + "content": "(b) Our piecewise definition" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 486, + 287, + 540 + ], + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 540 + ], + "type": "text", + "content": "Figure 3. The horizontal axis denotes the angle of a rotated box in anti-clockwise order. And the vertical axis denotes the angle definition in different methods. Due to the periodicity of the angle, Figure (a) shows that there is a sudden change when the angle of a rotated box closes to " + }, + { + "bbox": [ + 46, + 486, + 287, + 540 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 46, + 486, + 287, + 540 + ], + "type": "text", + "content": " with the 180-degree definition." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "Fig. 2 (b). On the one hand, Fig. 3 (b) shows that there is no sudden change of the angle in our piecewise angle definition. Thus, ours is free of the boundary discontinuity problem. On the other hand, rotation equivariance can be easily achieved by group convolutions [52], while rotation invariance relies on a larger capacity network or a larger number of training samples. Hence, we focus on rotation-equivariant features. And then, we are allowed to transform the invariant size estimation into equivariant size estimation by adding a factor " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\sin (\\theta)" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\cos (\\theta)" + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": " to the height and width (see Fig. 2 (b)). That way, the box size is enforced to change equivalently when the box orientation changes. Ac" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 418, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 443 + ], + "type": "text", + "content": "cordingly, both box size and angle regressions can benefit from rotation-equivariant features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 453, + 429, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 429, + 465 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 429, + 465 + ], + "type": "text", + "content": "3.2. Network architecture" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 473, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 556 + ], + "type": "text", + "content": "As discussed before, we aim to extract rotation-equivariant features for box size regression and angle regression. Therefore, we adopt rotation-equivariant networks as the backbone to extract rotation-equivariant features. Then, we design a parallel multi-branch architecture based on the proposed piecewise angle definition. The overall architecture of CrackDet is shown in Fig. 2 (c)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "Rotation-equivariant backbone. We first adopt CenterNet [73] as our baseline, which takes an object as a single point (i.e., the center point of the bounding box) and regresses the object size and offset. Then, we re-implement all layers of the fully-convolutional encoder-decoder networks (i.e., up-convolutional residual networks [20]) in CenterNet based on e2cnn [52], named as ReEDNet. Importantly, e2cnn includes rotation-equivariant convolution, rotation-equivariant up-convolution, and pooling, etc. Thanks to the rotation weight sharing and group representations in e2cnn, our ReEDNet takes smaller parameters than the original encoder-decoder networks in CenterNet and enjoys the capability of equivariance." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6656" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 147, + 163 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 147, + 163 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 147, + 163 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 147, + 163 + ], + "type": "image", + "image_path": "43798c108a9d6fc78b12f450717250e477abbc44e0ec7400827ef807a167d362.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 163, + 127, + 172 + ], + "lines": [ + { + "bbox": [ + 81, + 163, + 127, + 172 + ], + "spans": [ + { + "bbox": [ + 81, + 163, + 127, + 172 + ], + "type": "text", + "content": "(a) Branch 1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 175, + 547, + 209 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 547, + 209 + ], + "type": "text", + "content": "Figure 4. A visualization of our angle-based variance voting. Firstly, the red, blue, orange, and teal textboxes are the corresponding standard deviation obtained from branch 1, 2, 3, and 4, respectively. Then, for each oriented bounding box, we select brach " + }, + { + "bbox": [ + 46, + 175, + 547, + 209 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 175, + 547, + 209 + ], + "type": "text", + "content": " that has the smallest standard deviation to obtain its final detection result. Note that, the red background box is associated with the selected branch." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 147, + 70, + 334, + 163 + ], + "blocks": [ + { + "bbox": [ + 147, + 70, + 334, + 163 + ], + "lines": [ + { + "bbox": [ + 147, + 70, + 334, + 163 + ], + "spans": [ + { + "bbox": [ + 147, + 70, + 334, + 163 + ], + "type": "image", + "image_path": "6baa85e4c3754dac30c8e401d3cb750f626808d80587c6cbf7ba8985ac0c6e5a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 163, + 219, + 172 + ], + "lines": [ + { + "bbox": [ + 171, + 163, + 219, + 172 + ], + "spans": [ + { + "bbox": [ + 171, + 163, + 219, + 172 + ], + "type": "text", + "content": "(b) Branch 2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 334, + 70, + 541, + 163 + ], + "blocks": [ + { + "bbox": [ + 334, + 70, + 541, + 163 + ], + "lines": [ + { + "bbox": [ + 334, + 70, + 541, + 163 + ], + "spans": [ + { + "bbox": [ + 334, + 70, + 541, + 163 + ], + "type": "image", + "image_path": "a98286860ac998d5d81dffa88fbd579c4b86c8b3eff7a8e3e38e74220b875f71.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 163, + 407, + 173 + ], + "lines": [ + { + "bbox": [ + 361, + 163, + 407, + 173 + ], + "spans": [ + { + "bbox": [ + 361, + 163, + 407, + 173 + ], + "type": "text", + "content": "(d) Branch 4" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 163, + 538, + 174 + ], + "lines": [ + { + "bbox": [ + 447, + 163, + 538, + 174 + ], + "spans": [ + { + "bbox": [ + 447, + 163, + 538, + 174 + ], + "type": "text", + "content": "(e) After variance voting" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 228, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 288, + 312 + ], + "type": "text", + "content": "Multi-branch architecture. According to our piecewise angle definition, we define a valid range " + }, + { + "bbox": [ + 46, + 228, + 288, + 312 + ], + "type": "inline_equation", + "content": "[l_i, r_i)" + }, + { + "bbox": [ + 46, + 228, + 288, + 312 + ], + "type": "text", + "content": " for each branch " + }, + { + "bbox": [ + 46, + 228, + 288, + 312 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 228, + 288, + 312 + ], + "type": "text", + "content": ". During the training, we only select the ground truth boxes whose angles fall in the corresponding valid range of one branch. Specifically, referring to the existing 180-degree definition, for an oriented bounding box with angle " + }, + { + "bbox": [ + 46, + 228, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 228, + 288, + 312 + ], + "type": "text", + "content": " on the input image, it is valid for branch " + }, + { + "bbox": [ + 46, + 228, + 288, + 312 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 228, + 288, + 312 + ], + "type": "text", + "content": " when:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 140, + 319, + 287, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 319, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 140, + 319, + 287, + 332 + ], + "type": "interline_equation", + "content": "l _ {i} \\leq \\theta < r _ {i}. \\tag {1}", + "image_path": "cd6096c1bb5bd42c9e5c77433d6d81271d8e3fa8533693a8c55c2ab6b23435a3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "text", + "content": "As shown in Fig. 2 (c), the valid ranges of four branches are set to " + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "inline_equation", + "content": "[0^{\\circ}, 45^{\\circ})" + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "inline_equation", + "content": "[45^{\\circ}, 90^{\\circ})" + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "inline_equation", + "content": "[90^{\\circ}, 135^{\\circ})" + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "inline_equation", + "content": "[135^{\\circ}, 180^{\\circ})" + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "text", + "content": ", respectively. For branch " + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "text", + "content": ", we redefine the ground-truth " + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "text", + "content": ", height " + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "text", + "content": ", and width " + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "text", + "content": " in the 180-degree definition into " + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "inline_equation", + "content": "h_{i}" + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 46, + 338, + 287, + 540 + ], + "type": "text", + "content": ", according to our piecewise definition in Fig. 2 (b). Moreover, we aim to estimate the angle and angle confidence together. To this end, each branch of our parallel multi-branch architecture contains three detection heads on the top of the rotation-equivariant backbone for predicting size (i.e., height and width), angle, and angle confidence. Note that, we simply implement each head with a fully-connected layer. That way, our network enables us to predict multiple probability distributions instead of only one angle. For simplicity, we assume each predicted angle from each branch is independent and obeys a single-variate gaussian distribution. The corresponding equation is shown as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 80, + 544, + 287, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 544, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 80, + 544, + 287, + 576 + ], + "type": "interline_equation", + "content": "P _ {\\Theta_ {i}} (\\theta) = \\frac {1}{\\sqrt {2 \\pi \\sigma_ {i} ^ {2}}} e ^ {- \\frac {\\left(\\theta - \\theta_ {i} ^ {e}\\right) ^ {2}}{2 \\sigma_ {i} ^ {2}}}, i \\in \\{1, 2, 3, 4 \\} \\tag {2}", + "image_path": "1ddae73d8623e2c79aa7646112d2398f6cc5966d1e5d97f72d87f141a2ed83ff.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Theta_{i}" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": " is the parameters of the angle detection head from branch " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{e}" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": " denotes the estimated angle from branch " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": ", and standard deviation " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": " measures the angle confidence of the estimation on branch " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": " is very close to 0, it means the branch " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": " is extremely confident about the estimated angle. Accordingly, we suppose to have the redefined ground-truth angle " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": " that is in the valid range of branch " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": ". Similarly, " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": " can also be formulated as a gaussian distribution " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\theta_i,\\sigma_{gt}^2)" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": " with its standard deviation " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{gt}\\to 0" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": ". Then, this gaussian distribution can be viewed as: " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "P_{i}^{gt}(\\theta) = \\delta (\\theta -\\theta_{i})" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\delta (\\cdot)" + }, + { + "bbox": [ + 46, + 579, + 288, + 714 + ], + "type": "text", + "content": " indicates Dirac delta function." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 228, + 446, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 446, + 241 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 446, + 241 + ], + "type": "text", + "content": "3.3. The proposed MAR Loss" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 247, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 381 + ], + "type": "text", + "content": "Suppose the angle of a ground-truth bounding box lies in the valid range of branch " + }, + { + "bbox": [ + 304, + 247, + 545, + 381 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 247, + 545, + 381 + ], + "type": "text", + "content": ", our multi-branch angle regression loss (MAR Loss) aims to minimize the distribution discrepancy between the predicted angle distribution from branch " + }, + { + "bbox": [ + 304, + 247, + 545, + 381 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 247, + 545, + 381 + ], + "type": "text", + "content": " and the redefined ground-truth angle distribution. Meanwhile, we expect the other three branches to be able to predict larger variance so that we can filter those inaccurate predictions by variance voting in the inference step. More specifically, we first exploit the Wasserstein distance as the distance metric and minimize the distance between " + }, + { + "bbox": [ + 304, + 247, + 545, + 381 + ], + "type": "inline_equation", + "content": "P_{\\Theta_i}(\\theta)" + }, + { + "bbox": [ + 304, + 247, + 545, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 247, + 545, + 381 + ], + "type": "inline_equation", + "content": "P_i^{gt}(\\theta)" + }, + { + "bbox": [ + 304, + 247, + 545, + 381 + ], + "type": "text", + "content": ". Secondly, we maximize the estimated variance from the other three branches. The equation is shown as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 337, + 394, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 394, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 337, + 394, + 545, + 456 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {M A R} = \\frac {D _ {W} \\left(P _ {\\Theta_ {i}} (\\theta) \\| P _ {i} ^ {g t} (\\theta)\\right)}{\\frac {1}{2} + \\sigma_ {i} ^ {2}} - \\sum_ {j = 1, j \\neq i} ^ {4} \\sigma_ {j} ^ {2} \\tag {3} \\\\ = \\frac {\\| \\theta_ {i} ^ {e} - \\theta_ {i} \\| _ {2} ^ {2} + \\sigma_ {i} ^ {2}}{\\frac {1}{2} + \\sigma_ {i} ^ {2}} - \\sum_ {j = 1, j \\neq i} ^ {4} \\sigma_ {j} ^ {2} \\\\ \\end{array}", + "image_path": "79bb0589915183a30891e22a7d2fa96855b47cfbc77e9bba55f8d6b13891c14b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 469, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 469, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 469, + 545, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 469, + 545, + 493 + ], + "type": "inline_equation", + "content": "D_W(\\cdot)" + }, + { + "bbox": [ + 306, + 469, + 545, + 493 + ], + "type": "text", + "content": " is unfolded based on Wasserstein distance [3]:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 337, + 504, + 545, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 504, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 337, + 504, + 545, + 520 + ], + "type": "interline_equation", + "content": "D _ {W} \\left(P _ {\\Theta_ {i}} (\\theta) \\| P _ {i} ^ {g t} (\\theta)\\right) = \\| \\theta_ {i} ^ {e} - \\theta_ {i} \\| _ {2} ^ {2} + \\sigma_ {i} ^ {2}. \\tag {4}", + "image_path": "76989b05fd93cd5e33e548d4ab0a1de981bc76258ba20fa026cdb7dedbd0d0de.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": "Importantly, we exploit Wasserstein distance instead of KL-Divergence [18] for angle regression, because KL-Divergence heavily relies on a non-negligible intersection between two distributions [1, 18]. Note that, when " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_i^e" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": " is estimated accurately, i.e., " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\| \\theta_i^e -\\theta_i\\| _2^2\\to 0" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": ", the branch " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": " is expected to predict smaller variance. In addition, a term " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\frac{1}{2} +\\sigma_i^2" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": " is added for the following reason: due to the ambiguities in sub-crack orientation, when the angle " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{e}" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": " is estimated inaccurately, i.e., " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\| \\theta_i^e -\\theta_i\\| _2^2 >\\frac{1}{2}" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": ", we expect the branch " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": " to be able to predict larger variance " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_i^2" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": " so that " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{MAR}" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": " will be lower. Moreover, according to our piecewise angle definition, the size regression loss " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{size}}" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": " for the redefined ground truth height " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{size}} = |h_i^e -h_i| + |w_i^e -w_i|" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_i^e" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "w_{i}^{e}" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": " indicate the estimated height and width from the valid branch " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": ". Then," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6657" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 232, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 232, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 232, + 84 + ], + "type": "text", + "content": "the overall training objective is shown below:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 287, + 115 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 92, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 92, + 287, + 115 + ], + "type": "interline_equation", + "content": "L _ {\\text {C r a c k D e t}} = L _ {k} + \\lambda_ {o f f} L _ {o f f} + \\lambda_ {s i z e} L _ {s i z e} + \\lambda_ {M A R} L _ {M A R} \\tag {5}", + "image_path": "ffcbcb79dd2b918ab619be42feb99061aa12b9c29d68ed3922cd677274c6e22d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 117, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 117, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 47, + 117, + 287, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 117, + 287, + 163 + ], + "type": "inline_equation", + "content": "L_{k}" + }, + { + "bbox": [ + 47, + 117, + 287, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 117, + 287, + 163 + ], + "type": "inline_equation", + "content": "L_{off}" + }, + { + "bbox": [ + 47, + 117, + 287, + 163 + ], + "type": "text", + "content": " are the losses of center point recognition and offset regression by following CenterNet [73]. The hyper-parameter " + }, + { + "bbox": [ + 47, + 117, + 287, + 163 + ], + "type": "inline_equation", + "content": "\\lambda_{off}" + }, + { + "bbox": [ + 47, + 117, + 287, + 163 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 117, + 287, + 163 + ], + "type": "inline_equation", + "content": "\\lambda_{size}" + }, + { + "bbox": [ + 47, + 117, + 287, + 163 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 117, + 287, + 163 + ], + "type": "inline_equation", + "content": "\\lambda_{MAR}" + }, + { + "bbox": [ + 47, + 117, + 287, + 163 + ], + "type": "text", + "content": " are constant factors to control the balance of the above losses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 171, + 114, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 171, + 114, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 171, + 114, + 182 + ], + "type": "text", + "content": "3.4. Inference" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "text", + "content": "At the inference step, there are four branches, each of which will predict the height, width, angle, and standard deviation. Thus, we can not directly decide which one should be adopted. Fortunately, with the help of MAR Loss, we design an angle-based variance voting for obtaining the final bounding boxes. Firstly, we extract the peaks in the heatmap by following CenterNet [73]. Then, we get standard deviations " + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\sigma_{1},\\sigma_{2},\\sigma_{3},\\sigma_{4}" + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "text", + "content": " from four branches at each peak. As shown in Fig. 4, for each peak, we choose branch " + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "text", + "content": " that has the minimum estimated variance to obtain the raw output: " + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "inline_equation", + "content": "i^{*} = \\mathrm{argmin}_{i = \\{1,2,3,4\\}}\\sigma_{i}^{2}" + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "text", + "content": ". Next, according to CenterNet [73], we obtain the center (i.e., " + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "text", + "content": ") of the oriented bounding box according to the above peaks and the offset branch. Finally, we acquire the final height " + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "inline_equation", + "content": "h_o = h_{i^*}^e\\backslash \\Delta (\\theta_{i^*}^e)" + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "text", + "content": ", width " + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "inline_equation", + "content": "w_{o} = w_{i^{*}}^{e}\\backslash \\Delta (\\theta_{i^{*}}^{e})" + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "text", + "content": ", and angle " + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\theta_o = \\Gamma (\\theta_{i^*}^e)" + }, + { + "bbox": [ + 47, + 190, + 287, + 390 + ], + "type": "text", + "content": " of the bounding box by the following equation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 390, + 287, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 390, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 96, + 390, + 287, + 440 + ], + "type": "interline_equation", + "content": "\\Gamma \\left(\\theta_ {i} ^ {e}\\right) \\triangleq \\left\\{ \\begin{array}{l l} \\theta_ {i} ^ {e}, & i = 1, \\\\ 9 0 ^ {\\circ} - \\theta_ {i} ^ {e}, & i = 2, \\\\ 9 0 ^ {\\circ} + \\theta_ {i} ^ {e}, & i = 3, \\\\ 1 8 0 ^ {\\circ} - \\theta_ {i} ^ {e}, & i = 4. \\end{array} \\right. \\tag {6}", + "image_path": "897fb68ba56fb58799e4c34cd8cd88738335b88abf70a7b456c208a7cd326174.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 444, + 287, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 444, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 79, + 444, + 287, + 494 + ], + "type": "interline_equation", + "content": "\\Delta \\left(\\theta_ {i} ^ {e}\\right) \\triangleq \\left\\{ \\begin{array}{l l} \\cos \\left(\\theta_ {i} ^ {e}\\right), & i = 1, \\\\ \\sin \\left(9 0 ^ {\\circ} - \\theta_ {i} ^ {e}\\right), & i = 2, \\\\ \\sin \\left(9 0 ^ {\\circ} + \\theta_ {i} ^ {e}\\right), & i = 3, \\\\ - \\cos \\left(1 8 0 ^ {\\circ} - \\theta_ {i} ^ {e}\\right), & i = 4. \\end{array} \\right. \\tag {7}", + "image_path": "cddcba0b6098c38b810cc8cd1fabf18f85be4caaea07758739209151acd2891c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "type": "text", + "content": " denotes the index of the selected branch. " + }, + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\theta_{i}^{e}" + }, + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "type": "text", + "content": " indicate the estimated angle from branch " + }, + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 497, + 287, + 533 + ], + "type": "text", + "content": ". It is worth noting that the remaining inference step is the same as CenterNet [73]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 540, + 174, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 174, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 174, + 552 + ], + "type": "text", + "content": "3.5. The proposed datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 558, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 287, + 605 + ], + "type": "text", + "content": "To promote the research on oriented sub-crack detection, we propose ONPP, ORC, and OCCSD datasets, which are collected from industrial facilities, road pavement, and various buildings in real-world applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "ONPP: This dataset is collected from the industrial facilities in real-world applications by a high-resolution camera, comprising 200 images with the size of " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "7360 \\times 4912" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ". The width of the crack in the collected images varies from 0.05 mm to " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "10\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ". Moreover, the collected images contain different kinds of noise, such as various concrete types and light intensity. To extend this dataset without compromising the resolution, we directly slice these images into " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " pixels, constructing a final dataset with 3,104 samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "ORC: This dataset is collected from road pavement by a mobile vehicle equipped with a high-resolution camera. It contains 300 images with a size of " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "3489 \\times 3489" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": ". The images of this dataset include various sizes of cracks in the road. It also contains a lot of noise, such as well covers and various appearances of the crosswalk. Similar to ONPP, we slice these images into " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " pixels, constructing a final road pavement dataset with 1,303 samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 169, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 169, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 545, + 240 + ], + "type": "text", + "content": "OCCSD: This dataset is extended from the concrete crack segmentation dataset [71], which contains 458 high-resolution images (" + }, + { + "bbox": [ + 304, + 169, + 545, + 240 + ], + "type": "inline_equation", + "content": "4032 \\times 3024" + }, + { + "bbox": [ + 304, + 169, + 545, + 240 + ], + "type": "text", + "content": " pixels). We construct an oriented sub-crack detection dataset with 1,875 samples by slicing these images into " + }, + { + "bbox": [ + 304, + 169, + 545, + 240 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 169, + 545, + 240 + ], + "type": "text", + "content": " pixels and relabeling them from an oriented sub-crack detection perspective." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 254, + 386, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 254, + 386, + 267 + ], + "spans": [ + { + "bbox": [ + 306, + 254, + 386, + 267 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 274, + 417, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 417, + 286 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 417, + 286 + ], + "type": "text", + "content": "4.1. Experiment details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 293, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 437 + ], + "type": "text", + "content": "Evaluation testbed. We conduct experiments for oriented sub-crack detection on three datasets, i.e., ONPP, ORC, and OCCSD. By following the existing crack detection work [8, 31], we divide each dataset into the training, validation, and test set, with the proportion of " + }, + { + "bbox": [ + 304, + 293, + 545, + 437 + ], + "type": "inline_equation", + "content": "8:1:1" + }, + { + "bbox": [ + 304, + 293, + 545, + 437 + ], + "type": "text", + "content": ". Moreover, to further verify the effectiveness of CrackDet, we also provide the experiment results on the HRSC2016 dataset [29]. HRSC2016 dataset contains 1,061 samples. Its resolution ranges from " + }, + { + "bbox": [ + 304, + 293, + 545, + 437 + ], + "type": "inline_equation", + "content": "300 \\times 300" + }, + { + "bbox": [ + 304, + 293, + 545, + 437 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 293, + 545, + 437 + ], + "type": "inline_equation", + "content": "1500 \\times 900" + }, + { + "bbox": [ + 304, + 293, + 545, + 437 + ], + "type": "text", + "content": ". For fairness, according to previous works [55, 16, 23], we exploit the training set (436 images) and validation set (181 images) for training, and the testing set (444 images) for evaluation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": "Implementation details. We adopt Adam as the optimizer for training. The total epoch for training the CrackDet is set as 60, 60, 60, and 140 on ONPP, ORC, OCCSD, and HRSC2016 datasets, respectively. For the first three datasets, the initial learning rate is set to " + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "inline_equation", + "content": "4e - 4" + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": " and reduced by a factor of 10 after 20, and 40 epochs. For the last one, the initial learning rate is " + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "inline_equation", + "content": "4e - 4" + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": " and reduced by a factor of 10 after 90, and 120 epochs by following the previous work [39]. The batch size for ONPP, ORC, OCCSD, and HRSC2016 is set to 32, 32, 32, and 8, respectively. More importantly, the input resolutions of four datasets are " + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "inline_equation", + "content": "768 \\times 768" + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": ", respectively. By following the previous work [16], we only use random flipping for data augmentation. We implement our ReEDNet based on ResNet50 and e2cnn [52], and denote it as ReED-R-50. We train all models on four RTX 3090 GPUs and test on a single RTX 3090 GPU. According to our hyper-parameter sensitivity study, we set " + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{off} = 0.1" + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{size} = 0.2" + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{MAR} = 0.1" + }, + { + "bbox": [ + 304, + 437, + 545, + 665 + ], + "type": "text", + "content": " in our next experiments." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Evaluation metrics. Following previous works [75, 54], we evaluate the performance of CrackDet on oriented crack detection, in terms of Precision and Recall. Moreover, we adopt the mean orientation error (MOE) to verify the effec" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6658" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 69, + 500, + 185 + ], + "blocks": [ + { + "bbox": [ + 93, + 69, + 500, + 185 + ], + "lines": [ + { + "bbox": [ + 93, + 69, + 500, + 185 + ], + "spans": [ + { + "bbox": [ + 93, + 69, + 500, + 185 + ], + "type": "table", + "html": "
MethodONPPORC
Precision ↑Recall ↑MOE ↓Precision ↑Recall ↑MOE ↓
FCrack-O2020 [13]0.70370.82630.27850.66310.87140.3397
ReCrack-O2021 [50]0.75060.67190.39870.74320.82540.3282
CABF-FCOS-O2021 [69]0.78930.84130.32630.76300.88510.3744
Y3Crack-O2022[31]0.71340.73790.33500.71030.85140.3632
CenterCrack-O2022 [51]0.72260.61320.35740.77310.81020.3085
Yolo-ViT-O2022 [54]0.77660.83790.35030.79820.84610.2978
CrackDet0.82040.91060.16800.85780.89260.1166
", + "image_path": "53174c2de169aff004830959668837e160c4b535a157c8be2ad1f297187b53f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 185, + 517, + 195 + ], + "lines": [ + { + "bbox": [ + 74, + 185, + 517, + 195 + ], + "spans": [ + { + "bbox": [ + 74, + 185, + 517, + 195 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-art crack detectors on ONPP and ORC. MOE denotes the mean orientation error." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 215, + 287, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 287, + 262 + ], + "type": "text", + "content": "tiveness of the proposed method for estimating crack orientation. For the detection accuracy on the HRSC2016, we adopt the mean average precision (mAP) as an evaluation criterion, which is consistent with VOC2007 metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 301, + 282, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 282, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 282, + 316 + ], + "type": "text", + "content": "4.2. Comparison with crack detection approaches" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 331, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 289, + 713 + ], + "type": "text", + "content": "In this section, our method is compared to several baselines including state-of-the-art robust crack detection methods: FCrack uses Faster R-CNN to detect the crack patch [13], " + }, + { + "bbox": [ + 46, + 331, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Y}^{3}\\mathrm{Crack}" + }, + { + "bbox": [ + 46, + 331, + 289, + 713 + ], + "type": "text", + "content": " designs a intelligent system based on YOLO v3 for detecting crack [31], CenterCrack proposes a crack detection model by using CenterNet [51], ReCrack utilizes the RetinaNet to find the crack in asphalt pavement [50], CABF-FCOS proposes a one-stage network that can detect various defects including crack [69]. Yolo-ViT applies the Transformer to YOLO v5 for crack detection [54]. Note that, all of these methods can not detect the orientation of cracks. For a fair comparison, we add a branch to these methods for the regression of crack orientation and denote them as FCrack-O, " + }, + { + "bbox": [ + 46, + 331, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Y}^{3}\\mathrm{Crack - O}" + }, + { + "bbox": [ + 46, + 331, + 289, + 713 + ], + "type": "text", + "content": ", CenterCrack-O, ReCrack-O, CABF-FCOS-O, and Yolo-ViT-O respectively. All experiments are independently carried out on ONPP and ORC datasets. Table 1 provides the performance of various approaches from our oriented sub-crack detection perspective for crack detection. Bold denotes the best results. It can be found that the proposed method compares favorably to other competitive crack detection approaches. For example, compared with the Yolo-ViT-O, the precision and recall of CrackDet are 0.8204 and 0.9106 while the Yolo-ViT-O only achieves 0.7766 and 0.8379 on ONPP dataset. For the ORC dataset, CrackDet brings performance gains over Yolo-ViT-O around " + }, + { + "bbox": [ + 46, + 331, + 289, + 713 + ], + "type": "inline_equation", + "content": "6.0\\%" + }, + { + "bbox": [ + 46, + 331, + 289, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 331, + 289, + 713 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 46, + 331, + 289, + 713 + ], + "type": "text", + "content": " under the precision and recall metrics. Moreover, in terms of MOE metric, the proposed CrackDet achieves consistent improvements on all datasets. The results above together show the necessity that the boundary discontinuity problem and the ambiguities in orientation should be treated properly under the oriented sub-crack detection tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 215, + 545, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 215, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 215, + 545, + 228 + ], + "type": "text", + "content": "4.3. Discussion on the segmentation-based methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 233, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 545, + 389 + ], + "type": "text", + "content": "In this section, considering the detection-then-segmentation strategy in the real-world application [33, 35], we further verify that the proposed method can serve as a complement to segmentation-based methods. More specifically, for the pixel-level segmentation task, we replace the oriented box annotation with pixel-level annotation for the OCCSD dataset. Table 3 shows that the proposed method can improve the segmentation results, especially for the segmentation model with limited performance. The reason attributes that the proposed method can provide a more fine-grained localization, which can effectively promote the segmentation model being influenced by the noise background." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 396, + 539, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 396, + 539, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 539, + 408 + ], + "type": "text", + "content": "4.4. Comparison with oriented detection methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 510 + ], + "type": "text", + "content": "To further validate our method, we conduct a series of experiments on both crack datasets and HRSC2016, and report quantitative results to verify the effectiveness of CrackDet. As CrackDet shows the purpose of detecting the oriented object, we compare it with existing state-of-the-art oriented detectors with the help of [74], including Oriented RepPoints [23], Oriented R-CNN [55], SASM [19], etc." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 510, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 689 + ], + "type": "text", + "content": "Results on oriented sub-crack detection. Table 2 provides the comparison results over the existing powerful oriented object detectors. We can observe that, first, our CrackDet outperforms the existing detectors by a large margin in terms of Precision. The reason is that the proposed piecewise angle definition can promote the proposed detector to take advantage of rotation-equivariant features and be free of the boundary discontinuity problem. Secondly, CrackDet yields significant and consistent improvement on all three datasets in terms of MOE metric. The reason can be explained as CrackDet can estimate the angle confidence with the help of the multi-branch architecture and MAR Loss. Thus, according to the corresponding angle confidence from each branch, the angle-based variance voting improves the angle estimation results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Results on HRSC2016. For fairness, we follow the same settings in Oriented RepPoints [23], and report the results" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6659" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 70, + 520, + 234 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 520, + 234 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 520, + 234 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 520, + 234 + ], + "type": "table", + "html": "
MethodONPPORC
Precision ↑Recall ↑MOE ↓Precision ↑Recall ↑MOE ↓
\\(CSL_{ECCV'20}\\) [60]0.73090.81460.26330.79430.74370.3285
\\(R^3\\text{Det}_{AAAI'21}\\) [62]0.78350.74330.29770.76970.73120.4280
\\(S^2ANet_{TGRS'21}\\) [15]0.77730.77730.41330.70680.78680.4640
\\(ReDet_{CVPR'21}\\) [16]0.78440.80590.27430.71910.86840.3335
Beyond Bounding-Box \\(CVPR'21\\) [14]0.77900.84680.30030.60760.90390.3441
Oriented R-CNN\\(ICCV'21\\) [55]0.80320.73550.21720.70870.74160.3470
\\(GWD_{ICML'21}\\) [63]0.71690.78840.44010.75760.86320.3014
\\(KLD_{NeurIPS'21}\\) [65]0.74660.64070.34460.74240.90760.2818
\\(SASM_{AAAI'22}\\) [19]0.74160.83030.30620.76530.90380.3489
Oriented RepPoints \\(CVPR'22\\) [23]0.78800.87820.30080.78510.91010.3497
CrackDet0.82040.91060.16800.85780.89260.1166
", + "image_path": "916d7f96728ac977e3ec6636996257302ec1e15c82ffeb9741371cb3f7e140a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 261, + 288, + 293 + ], + "blocks": [ + { + "bbox": [ + 102, + 234, + 488, + 243 + ], + "lines": [ + { + "bbox": [ + 102, + 234, + 488, + 243 + ], + "spans": [ + { + "bbox": [ + 102, + 234, + 488, + 243 + ], + "type": "text", + "content": "Table 2. We provide a detailed comparison over general oriented detection approaches on ONPP and ORC." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 261, + 288, + 293 + ], + "lines": [ + { + "bbox": [ + 50, + 261, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 50, + 261, + 288, + 293 + ], + "type": "table", + "html": "
Precision ↑Crackformer ICCV'21 [27]JTFN ICCV'21 [10]LIOT TIP'22 [45]
w/o CrackDet0.7810.7540.803
w/ CrackDet0.8140.7980.819
", + "image_path": "c38e7c8391ee00ba1e8a5408a6b33a2db948630f52c7cf3ac76f91ea3e66cf5d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 286, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 286, + 387 + ], + "type": "text", + "content": "under the VOC2007 metric. Table 4 shows the comparison results. The proposed CrackDet achieves the best performance, even compared with strong FPN-based baselines. Such results further validate that the proposed CrackDet has the potential in more oriented object detection tasks." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 50, + 397, + 286, + 519 + ], + "blocks": [ + { + "bbox": [ + 47, + 295, + 278, + 306 + ], + "lines": [ + { + "bbox": [ + 47, + 295, + 278, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 278, + 306 + ], + "type": "text", + "content": "Table 3. Segmentation performance with CrackDet on OCCSD." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 397, + 286, + 519 + ], + "lines": [ + { + "bbox": [ + 50, + 397, + 286, + 519 + ], + "spans": [ + { + "bbox": [ + 50, + 397, + 286, + 519 + ], + "type": "table", + "html": "
MethodsBackbonemAP50(07)↑
R3Det AAAI'21 [62]R-101-FPN89.26
R3Det-DCL CVPR'21 [58]R-101-FPN89.46
FPN-CSLECCV'20 [60]R-101-FPN89.62
ReDet CVPR'21 [16]Re-R-5090.46
DAL AAAI'21 [32]R-101-FPN89.77
S2A-NetTGRS'21 [15]R-101-FPN90.17
S2A-Net-DHRec PAMI'22 [37]R-101-FPN90.22
Oriented R-CNN ICCV'21 [55]R-50-FPN90.40
Oriented RepPoints CVPR'22 [23]R-50-FPN90.38
CrackDetReED-R-5090.60
", + "image_path": "3540a8799e5df1ec3397b5989c60fa1f9bc1a5c47540eade57c0f86f4b9eb76a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 562, + 138, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 138, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 138, + 575 + ], + "type": "text", + "content": "4.5. Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 581, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 286, + 640 + ], + "type": "text", + "content": "To examine the contribution of each component in our proposed detector: rotation-equivariant network, the number of valid ranges in the proposed piecewise angle definition, and MAR Loss, a series of ablation experiments are performed on ONPP and ORC datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "content": "Evaluation on the rotation-equivariant network. As expected, CrackDet achieves better results by combining the rotation-equivariant network. Table 5 shows that it improves the precision, recall, and MOE by " + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "inline_equation", + "content": "0.91\\%" + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "inline_equation", + "content": "0.023" + }, + { + "bbox": [ + 46, + 642, + 286, + 713 + ], + "type": "text", + "content": " with the help of the rotation-equivariant network, respectively. The reason behind this effect is that rotation-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 264, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 264, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 264, + 545, + 300 + ], + "type": "text", + "content": "equivariant features extracted by the rotation-equivariant network are beneficial to size and angle regression within the proposed piecewise angle definition." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 309, + 545, + 350 + ], + "blocks": [ + { + "bbox": [ + 46, + 521, + 286, + 543 + ], + "lines": [ + { + "bbox": [ + 46, + 521, + 286, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 286, + 543 + ], + "type": "text", + "content": "Table 4. We report results on the HRSC2016 test set. mAP(07) denotes detection results under VOC2007 mAP metrics." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 309, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 350 + ], + "type": "table", + "html": "
MethodsONPP
Precision ↑Recall ↑MOE ↓
w/o Rotation-equivariant network [17]0.81130.89520.1914
w/ Rotation-equivariant network0.82040.91060.1680
", + "image_path": "a951cd280356b5878ddc49e34ba3ba5cd91b61e01c1e50479d80ddb13c8c9b0b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 386, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 386, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 545, + 506 + ], + "type": "text", + "content": "Evaluation on the number of valid angle ranges. We study the effect of the number of valid angle ranges in our piecewise angle definition. For fairness, we do not add the factor " + }, + { + "bbox": [ + 305, + 386, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\sin(\\theta)" + }, + { + "bbox": [ + 305, + 386, + 545, + 506 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 386, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\cos(\\theta)" + }, + { + "bbox": [ + 305, + 386, + 545, + 506 + ], + "type": "text", + "content": " for all settings. Table 6 shows the results using one to five ranges. The corresponding results demonstrate that our four ranges definition improves over the single range (baseline) with a 0.14 decrease. As can be noticed, the five ranges do not bring further improvement over the four ranges. By considering the complexity and performance, we choose four ranges as the default setting." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 308, + 516, + 541, + 571 + ], + "blocks": [ + { + "bbox": [ + 305, + 352, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 305, + 352, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 352, + 545, + 373 + ], + "type": "text", + "content": "Table 5. Performance comparisons of with- and without rotation-equivariant backbone on ONPP dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 516, + 541, + 571 + ], + "lines": [ + { + "bbox": [ + 308, + 516, + 541, + 571 + ], + "spans": [ + { + "bbox": [ + 308, + 516, + 541, + 571 + ], + "type": "table", + "html": "
NumValid Angle RangesMOE ↓
1[0°, 180°)0.33
2[0°, 90°), [90°, 180°)0.30
3[0°, 60°), [60°, 120°), [120°, 180°)0.21
4[0°, 45°), [45°, 90°), [90°, 135°), [135°, 180°)0.19
5[0°, 36°), [36°, 72°), [72°, 108°), [108°, 144°), [144°, 180°)0.22
", + "image_path": "824c8efc3b8abfb9cf485ffa98635fbd88d282d4fb362d6e80b6df5b8e8705b5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 571, + 534, + 582 + ], + "lines": [ + { + "bbox": [ + 306, + 571, + 534, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 534, + 582 + ], + "type": "text", + "content": "Table 6. Results from the different number of ranges on ONPP." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 713 + ], + "type": "text", + "content": "Evaluation on MAR loss. It is to be expected that Crack-Det combined with MAR Loss achieves consistent improvements on ONPP and ORC datasets, shown in Table 7. The reason is three-fold: (1) KL Loss may fail when two distributions do not have a non-negligible intersection. (2) By learning to estimate high variances for the ambiguous crack orientation during training, our model is able to learn confidence in predicting sub-crack orientation. (3) MAR Loss incorporates learning angle confidence which can potentially help the network to learn discriminative features. The" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6660" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "learned variances through our MAR Loss are interpretable (see Fig. 4). Our network will output higher variances for ambiguous crack orientation, which can be useful in judging the overall orientation in real-world applications [38]." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 130, + 286, + 165 + ], + "blocks": [ + { + "bbox": [ + 50, + 130, + 286, + 165 + ], + "lines": [ + { + "bbox": [ + 50, + 130, + 286, + 165 + ], + "spans": [ + { + "bbox": [ + 50, + 130, + 286, + 165 + ], + "type": "table", + "html": "
MethodsONPPORC
Precision ↑Recall ↑MOE ↓Precision ↑Recall ↑MOE ↓
KL Loss [18]0.78650.87180.24250.81490.85370.1986
MAR Loss0.82040.91060.16800.85780.89260.1166
", + "image_path": "7a7a73e6bf306c8a7188de8368db1ef9c9e2170e68c6d767edb650c119aa752f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 167, + 276, + 178 + ], + "lines": [ + { + "bbox": [ + 47, + 167, + 276, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 167, + 276, + 178 + ], + "type": "text", + "content": "Table 7. Performance comparisons of MAR Loss and KL Loss." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 203, + 119, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 203, + 119, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 203, + 119, + 216 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 224, + 288, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 288, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 288, + 486 + ], + "type": "text", + "content": "In this work, by starting from a new perspective for crack detection, we propose CrackDet, a first-of-its-kind oriented sub-crack detector that overcomes the boundary discontinuity problem. Meanwhile, we propose the MAR loss to capture the ambiguities in crack orientation by combining CrackDet. The experiments demonstrate that our method outperforms state-of-the-art crack detectors and oriented detectors. On one hand, as we also construct three datasets for oriented sub-crack detection, this work should also give a direction to researchers for designing the crack detector from the oriented sub-crack detection perspective. On the other hand, it highlights the key drawback of existing perspectives of crack detection (i.e., a high labeling cost of pixel-level segmentation, and a coarse crack localization of patch-level classification and object-level detection). We believe this paper has the potential to inspire the following works: (1) multi-model learning [24, 25] for boosting the detection performance in a novel field with limited target data [21], (2) semi-supervised learning and active learning [9] to further alleviate its labeling cost in real-world applications, and (3) oriented sub-crack detection via horizontal annotations [66]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 504, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 287, + 635 + ], + "type": "text", + "content": "Acknowledgements. This work is supported by the National Nature Science Foundation of China under Grants 62073225,62203134, 61972263, 62072315, the National Key RD Program of China under Grants 2020YFA0908700, the Natural Science Foundation of Guangdong Province Outstanding Youth Program under Grants 2019B151502018, Shenzhen Science and Technology Innovation Commission (20220809141216003, JCYJ20210324093808021, JCYJ20220531102817040), the Guangdong \"Pearl River Talent Recruitment Program\" under Grant 2019ZT08X603, the Guangdong \"Pearl River Talent Plan\" under Grant 2019JC01X235." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 659, + 106, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 659, + 106, + 672 + ], + "spans": [ + { + "bbox": [ + 48, + 659, + 106, + 672 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 52, + 680, + 286, + 713 + ], + "type": "text", + "content": "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In Proc. ICML, pages 214-223, 2017. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 311, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 311, + 73, + 545, + 116 + ], + "type": "text", + "content": "[2] Seyed Majid Azimi, Eleonora Vig, Reza Bahmanyar, Marco Körner, and Peter Reinartz. Towards multi-class object detection in unconstrained remote sensing imagery. In Proc. ACCV, pages 150-165, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 118, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 118, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 311, + 118, + 545, + 138 + ], + "type": "text", + "content": "[3] Raktim Bhattacharya. Data assimilation in optimal transport framework. arXiv preprint arXiv:2005.08670, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 311, + 141, + 545, + 184 + ], + "type": "text", + "content": "[4] Fu-Chen Chen and Mohammad R Jahanshahi. Nb-cnn: Deep learning-based crack detection using convolutional neural network and naive bayes data fusion. IEEE Transactions on Industrial Electronics, 65(5):4392-4400, 2017. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 311, + 186, + 545, + 228 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proc. ECCV, pages 801-818, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 311, + 231, + 545, + 262 + ], + "type": "text", + "content": "[6] Minghao Chen, Hongyang Xue, and Deng Cai. Domain adaptation for semantic segmentation with maximum squares loss. In Proc. ICCV, pages 2090-2099, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 311, + 264, + 545, + 297 + ], + "type": "text", + "content": "[7] Qiang Chen, Yingming Wang, Tong Yang, Xiangyu Zhang, Jian Cheng, and Jian Sun. You only look one-level feature. In Proc. CVPR, pages 13039-13048, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 298, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 298, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 311, + 298, + 545, + 341 + ], + "type": "text", + "content": "[8] Zhuangzhuang Chen, Jin Zhang, Zhuonan Lai, Jie Chen, Zun Liu, and Jianqiang Li. Geometry-aware guided loss for deep crack recognition. In Proc. CVPR, pages 4703-4712, 2022. 1, 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 343, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 343, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 311, + 343, + 545, + 375 + ], + "type": "text", + "content": "[9] Zhuangzhuang Chen, Jin Zhang, Pan Wang, Jie Chen, and Jianqiang Li. When active learning meets implicit semantic data augmentation. In Proc. ECCV, pages 56-72, 2022. 9" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "text", + "content": "[10] Mingfei Cheng, Kaili Zhao, Xuhong Guo, Yajing Xu, and Jun Guo. Joint topology-preserving and feature-refinement network for curvilinear structure segmentation. In Proc. ICCV, pages 7147-7156, 2021. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "type": "text", + "content": "[11] Jian Ding, Nan Xue, Yang Long, Gui-Song Xia, and Qikai Lu. Learning roi transformer for oriented object detection in aerial images. In Proc. CVPR, pages 2849-2858, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 487 + ], + "type": "text", + "content": "[12] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet: Keypoint triplets for object detection. In Proc. ICCV, pages 6569-6578, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "text", + "content": "[13] Fen Fang, Liyuan Li, Ying Gu, Hongyuan Zhu, and Joo-Hwee Lim. A novel hybrid approach for crack detection. PR, 107:107474, 2020. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "text", + "content": "[14] Zonghao Guo, Chang Liu, Xiaosong Zhang, Jianbin Jiao, Xi-angyang Ji, and Qixiang Ye. Beyond bounding-box: Convex-hull feature adaptation for oriented and densely packed object detection. In Proc. CVPR, 2021. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 568, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 600 + ], + "type": "text", + "content": "[15] Jiaming Han, Jian Ding, Jie Li, and Gui-Song Xia. Align deep features for oriented object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2021. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "text", + "content": "[16] Jiaming Han, Jian Ding, Nan Xue, and Gui-Song Xia. Redet: A rotation-equivariant detector for aerial object detection. In Proc. CVPR, pages 2786-2795, 2021. 2, 3, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 667 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In Proc. ECCV, pages 630-645, 2016. 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[18] Yihui He, Chenchen Zhu, Jianren Wang, Marios Savvides, and Xiangyu Zhang. Bounding box regression with uncertainty for accurate object detection. In Proc. CVPR, pages 2888-2897, 2019. 5, 9" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6661" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[19] Liping Hou, Ke Lu, Jian Xue, and Yuqiu Li. Shape-adaptive selection and measurement for oriented object detection. In Proc. AAAI, 2022. 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 162 + ], + "type": "text", + "content": "[20] Jonathan Huang, Vivek Rathod, Chen Sun, Menglong Zhu, Anoop Korattikara, Alireza Fathi, Ian Fischer, Zbigniew Wojna, Yang Song, Sergio Guadarrama, et al. Speed/accuracy trade-offs for modern convolutional object detectors. In Proc. CVPR, pages 7310-7311, 2017. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "text", + "content": "[21] Jianqiang Li, Zhuangzhuang Chen, Jie Chen, and Qiuzhen Lin. Diversity-sensitive generative adversarial network for terrain mapping under limited human intervention. IEEE Transactions on Cybernetics, 51(12):6029-6040, 2020. 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "type": "text", + "content": "[22] Kai Li, Bo Wang, Yingjie Tian, and Zhiquan Qi. Fast and accurate road crack detection based on adaptive cost-sensitive loss function. IEEE Transactions on Cybernetics, 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 288, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 272 + ], + "type": "text", + "content": "[23] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented leptopoints for aerial object detection. In Proc. CVPR, pages 1829-1838, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 304 + ], + "type": "text", + "content": "[24] Wenrui Li and Xiaopeng Fan. Image-text alignment and retrieval using light-weight transformer. In Proc. ICASSP, pages 4758-4762, 2022. 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 304, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 348 + ], + "type": "text", + "content": "[25] Wenrui Li, Zhengyu Ma, Liang-Jian Deng, Xiaopeng Fan, and Yonghong Tian. Neuron-based spiking transmission and reasoning network for robust image-text retrieval. IEEE TCSVT, 2022. 9" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 349, + 288, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 288, + 382 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dálár. Focal loss for dense object detection. In Proc. ICCV, pages 2980-2988, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 383, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 426 + ], + "type": "text", + "content": "[27] Huajun Liu, Xiangyu Miao, Christoph Mertz, Chengzhong Xu, and Hui Kong. Crackformer: Transformer network for fine-grained crack detection. In Proc. ICCV, pages 3783-3792, 2021. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 426, + 288, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 471 + ], + "type": "text", + "content": "[28] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In Proc. ECCV, pages 21-37, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "text", + "content": "[29] Zikun Liu, Liu Yuan, Lubin Weng, and Yiping Yang. A high resolution optical satellite image dataset for ship recognition and some new baselines. In International conference on pattern recognition applications and methods, volume 2, pages 324-331, 2017. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 288, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 559 + ], + "type": "text", + "content": "[30] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proc. CVPR, pages 3431-3440, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "text", + "content": "[31] Duo Ma, Hongyuan Fang, Niannian Wang, Chao Zhang, Jiaxiu Dong, and Haobang Hu. Automatic detection and counting system for pavement cracks based on pcgan and yolo-mf. IEEE Transactions on Intelligent Transportation Systems, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 657 + ], + "type": "text", + "content": "[32] Qi Ming, Zhiqiang Zhou, Lingjuan Miao, Hongwei Zhang, and Linhao Li. Dynamic anchor learning for arbitrary-oriented object detection. In Proc. AAAI, pages 2355-2363, 2021. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[33] Mayank Mishra, Vipul Jain, Saurabh Kumar Singh, and Damodar Maity. Two-stage method based on the you only look once framework and image segmentation for crack detection in concrete structures. Architecture, Structures and Construction, pages 1-18, 2022. 2, 3, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "text", + "content": "[34] Hafiz Suliman Munawar, Ahmed WA Hammad, Assed Haddad, Carlos Alberto Pereira Soares, and S Travis Waller. Image-based crack detection methods: A review. Infrastructures, 6(8):115, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "text", + "content": "[35] Nhung Hong Thi Nguyen, Stuart Perry, Don Bone, Ha Thanh Le, and Thuy Thi Nguyen. Two-stage convolutional neural network for road crack detection and segmentation. Expert Systems with Applications, 186:115718, 2021. 2, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 162, + 547, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 547, + 207 + ], + "type": "text", + "content": "[36] Son Dong Nguyen, Thai Son Tran, Van Phuc Tran, Hyun Jong Lee, Md Piran, Van Phuc Le, et al. Deep learning-based crack detection: A survey. International Journal of Pavement Research and Technology, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 207, + 546, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 546, + 240 + ], + "type": "text", + "content": "[37] Guangtao Nie and Hua Huang. Multi-oriented object detection in aerial images with double horizontal rectangles. IEEE TPAMI, 2022. 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 241, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 546, + 285 + ], + "type": "text", + "content": "[38] Sijun Niu and Vikas Srivastava. Simulation trained cnn for accurate embedded crack length, location, and orientation prediction from ultrasound measurements. International Journal of Solids and Structures, 242:111521, 2022. 2, 9" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 285, + 546, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 285, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 546, + 342 + ], + "type": "text", + "content": "[39] Xingjia Pan, Yuqiang Ren, Kekai Sheng, Weiming Dong, Haolei Yuan, Xiaowei Guo, Chongyang Ma, and Changsheng Xu. Dynamic refinement network for oriented and densely packed object detection. In Proc. CVPR, pages 11207-11216, 2020. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 342, + 546, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 546, + 376 + ], + "type": "text", + "content": "[40] Wen Qian, Xue Yang, Silong Peng, Junchi Yan, and Yue Guo. Learning modulated loss for rotated object detection. In Proc. AAAI, pages 2458-2466, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 376, + 546, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 546, + 431 + ], + "type": "text", + "content": "[41] Zhong Qu, Wen Chen, Shi-Yan Wang, Tu-Ming Yi, and Ling Liu. A crack detection algorithm for concrete pavement based on attention mechanism and multi-features fusion. IEEE Transactions on Intelligent Transportation Systems, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 432, + 546, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 546, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 546, + 476 + ], + "type": "text", + "content": "[42] Marcos Quintana, Juan Torres, and José Manuel Menéndez. A simplified computer vision system for road surface inspection and maintenance. IEEE Transactions on Intelligent Transportation Systems, 17(3):608-619, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 476, + 546, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 476, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 476, + 546, + 510 + ], + "type": "text", + "content": "[43] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proc. CVPR, pages 779-788, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 510, + 546, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 546, + 544 + ], + "type": "text", + "content": "[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Proc. NeurIPS, 2015. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 545, + 546, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 546, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 546, + 590 + ], + "type": "text", + "content": "[45] Tianyi Shi, Nicolas Boutry, Yongchao Xu, and Thierry Géraud. Local intensity order transformation for robust curvilinear object segmentation. IEEE Transactions on Image Processing, 31:2557-2569, 2022. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 590, + 546, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 546, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 546, + 634 + ], + "type": "text", + "content": "[46] Yong Shi, Limeng Cui, Zhiquan Qi, Fan Meng, and Zhensong Chen. Automatic road crack detection using random structured forests. IEEE Transactions on Intelligent Transportation Systems, 17(12):3434-3445, 2016. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 635, + 546, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 546, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 546, + 669 + ], + "type": "text", + "content": "[47] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 669, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 546, + 713 + ], + "type": "text", + "content": "[48] Weidong Song, Guohui Jia, Hong Zhu, Di Jia, and Lin Gao. Automated pavement crack damage detection using deep multiscale convolutional features. Journal of Advanced Transportation, 2020, 2020. 2, 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6662" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[49] Xinzi Sun, Yuanchang Xie, Liming Jiang, Yu Cao, and Benyuan Liu. Dma-net: Deeplab with multi-scale attention for pavement crack segmentation. IEEE Transactions on Intelligent Transportation Systems, 23(10):18392-18403, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "text", + "content": "[50] Van Phuc Tran, Thai Son Tran, and Hyun Jong Lee. One stage detector (retinanet)-based crack detection for asphalt pavements considering pavement distresses and surface objects. Journal of Civil Structural Health Monitoring, pages 205-222, 2021. 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "text", + "content": "[51] Ruoxin Wang and Chi Fai Cheung. Centernet-based defect detection for additive manufacturing. Expert Systems with Applications, 188:116000, 2022. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 239 + ], + "type": "text", + "content": "[52] Maurice Weiler and Gabriele Cesa. General e (2)-equivariant steerable cnns. In Proc. NeurIPS, 2019. 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "text", + "content": "[53] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In Proc. CVPR, pages 3974-3983, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "type": "text", + "content": "[54] Xuezhi Xiang, Zhiyuan Wang, and Yulong Qiao. An improved yolov5 crack detection method combined with transformer. IEEE Sensors Journal, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 320, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 353 + ], + "type": "text", + "content": "[55] Xingxing Xie, Gong Cheng, Jiabao Wang, Xiwen Yao, and Junwei Han. Oriented r-cnn for object detection. In Proc. ICCV, pages 3520-3529, 2021. 3, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "text", + "content": "[56] Yongchao Xu, Mingtao Fu, Qimeng Wang, Yukang Wang, Kai Chen, Gui-Song Xia, and Xiang Bai. Gliding vertex on the horizontal bounding box for multi-oriented object detection. IEEE TPAMI, 43(4):1452-1459, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 453 + ], + "type": "text", + "content": "[57] Fan Yang, Lei Zhang, Sijia Yu, Danil Prokhorov, Xue Mei, and Haibin Ling. Feature pyramid and hierarchical boosting network for pavement crack detection. IEEE Transactions on Intelligent Transportation Systems, 21(4):1525-1535, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "text", + "content": "[58] Xue Yang, Liping Hou, Yue Zhou, Wentao Wang, and Junchi Yan. Dense label encoding for boundary discontinuity free rotation detection. In Proc. CVPR, 2021. 2, 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 488, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 544 + ], + "type": "text", + "content": "[59] Xincong Yang, Heng Li, Yantao Yu, Xiaochun Luo, Ting Huang, and Xu Yang. Automatic pixel-level crack detection and measurement using fully convolutional network. Computer-Aided Civil and Infrastructure Engineering, 33(12):1090-1109, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "text", + "content": "[60] Xue Yang and Junchi Yan. Arbitrary-oriented object detection with circular smooth label. In Proc. ECCV, pages 677-694, 2020. 3, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "text", + "content": "[61] Xue Yang and Junchi Yan. On the arbitrary-oriented object detection: Classification based approaches revisited. IJCV, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "text", + "content": "[62] Xue Yang, Junchi Yan, Ziming Feng, and Tao He. R3det: Refined single-stage detector with feature refinement for rotating object. In Proc. AAAI, 2021. 3, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[63] Xue Yang, Junchi Yan, Qi Ming, Wentao Wang, Xiaopeng Zhang, and Qi Tian. Rethinking rotated object detection with gaussian Wasserstein distance loss. In Proc. ICML, 2021. 3, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[64] X Yang, J Yang, J Yan, Y Zhang, T Zhang, Z Guo, X Sun, and K SCRDet Fu. Towards more robust detection for small," + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 655 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "cluttered and rotated objects. In Proc. ICCV, volume 27, pages 8232-8241, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[65] Xue Yang, Xiaojiang Yang, Jirui Yang, Qi Ming, Wentao Wang, Qi Tian, and Junchi Yan. Learning high-precision bounding box for rotated object detection via kullback-leibler divergence. In Proc. NeurIPS, 2021. 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 183 + ], + "type": "text", + "content": "[66] Xue Yang, Gefan Zhang, Wentong Li, Yue Zhou, Xuehui Wang, and Junchi Yan. H2rbox: Horizontal box annotation is all you need for oriented object detection. In Proc. ICLR, 2022. 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 185, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 229 + ], + "type": "text", + "content": "[67] Dawen Yu, Shunping Ji, Xue Li, Zhaode Yuan, and Chaoyong Shen. Earthquake crack detection from aerial images using a deformable convolutional neural network. IEEE Transactions on Geoscience and Remote Sensing, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "text", + "content": "[68] Hong-Xing Yu, Jiajun Wu, and Li Yi. Rotationally equivariant 3d object detection. In Proc. CVPR, pages 1456-1464, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[69] Jianbo Yu, Xun Cheng, and Qingfeng Li. Surface defect detection of steel strips based on anchor-free network with channel attention and bidirectional feature fusion. IEEE Transactions on Instrumentation and Measurement, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 353 + ], + "type": "text", + "content": "[70] Gongjie Zhang, Shijian Lu, and Wei Zhang. Cad-net: A context-aware detection network for objects in remote sensing imagery. IEEE Transactions on Geoscience and Remote Sensing, 57(12):10015-10024, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "text", + "content": "[71] Lei Zhang, Fan Yang, Yimin Daniel Zhang, and Ying Julie Zhu. Road crack detection using deep convolutional neural network. In Proc. ICIP, pages 3708-3712, 2016. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[72] Xiaoyu Zhao, Wenlian Huang, Jie Chen, Zhuangzhuang Chen, and Jianqiang Li. Automatic thin crack segmentation with deep context aggregation network. In Proc. ICARM, pages 206-212, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 464 + ], + "type": "text", + "content": "[73] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 2, 4, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 466, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 520 + ], + "type": "text", + "content": "[74] Yue Zhou, Xue Yang, Gefan Zhang, Jiabao Wang, Yanyi Liu, Liping Hou, Xue Jiang, Xingzhao Liu, Junchi Yan, Chengqi Lyu, et al. Mmrotate: A rotated object detection benchmark using pytorch. In Proc. ACMM, pages 7331-7334, 2022. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 566 + ], + "type": "text", + "content": "[75] Junqing Zhu, Jingtao Zhong, Tao Ma, Xiaoming Huang, Weiguang Zhang, and Yang Zhou. Pavement distress detection using convolutional neural networks with images captured via uav. Automation in Construction, 2022. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 567, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 610 + ], + "type": "text", + "content": "[76] Qin Zou, Yu Cao, Qingquan Li, Qingzhou Mao, and Song Wang. Cracktree: Automatic crack detection from pavement images. Pattern Recognition Letters, 33(3):227-238, 2012. 1, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 612, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 655 + ], + "type": "text", + "content": "[77] Qin Zou, Zheng Zhang, Qingquan Li, Xianbiao Qi, Qian Wang, and Song Wang. Deepcrack: Learning hierarchical convolutional features for crack detection. IEEE TIP, 28(3):1498-1512, 2018. 1, 2, 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6663" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/6b883d6c-8e38-4d25-996f-79941abad307_content_list.json b/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/6b883d6c-8e38-4d25-996f-79941abad307_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e67a47c50c6b15c6aa3bfbb11e8d5ba3f48c5694 --- /dev/null +++ b/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/6b883d6c-8e38-4d25-996f-79941abad307_content_list.json @@ -0,0 +1,1775 @@ +[ + { + "type": "text", + "text": "The Devil is in the Upsampling: Architectural Decisions Made Simpler for Denoising with Deep Image Prior", + "text_level": 1, + "bbox": [ + 112, + 128, + 857, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yilin Liu $^{1*}$ , Jiang Li $^{1*}$ , Yunkui Pang $^{1*}$ , Dong Nie $^{1}$ , and Pew-Thian Yap $^{1}$", + "bbox": [ + 196, + 186, + 772, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of North Carolina at Chapel Hill", + "bbox": [ + 308, + 209, + 661, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yilinliu, yunkuipa, dongnie} @cs.unc.edu, jiang.li@unc.edu, ptyap@med.unc.edu", + "bbox": [ + 241, + 227, + 728, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Image Prior (DIP) shows that some network architectures inherently tend towards generating smooth images while resisting noise, a phenomenon known as spectral bias. Image denoising is a natural application of this property. Although denoising with DIP mitigates the need for large training sets, two often intertwined practical challenges need to be overcome: architectural design and noise fitting. Existing methods either handcraft or search for suitable architectures from a vast design space, due to the limited understanding of how architectural choices affect the denoising outcome. In this study, we demonstrate from a frequency perspective that unlearnt upsampling is the main driving force behind the denoising phenomenon with DIP. This finding leads to straightforward strategies for identifying a suitable architecture for every image without laborious search. Extensive experiments show that the estimated architectures achieve superior denoising results than existing methods with up to $95\\%$ fewer parameters. Thanks to this under-parameterization, the resulting architectures are less prone to noise-fitting1.", + "bbox": [ + 73, + 318, + 473, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 652, + 209, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image denoising is useful on its own and can be a plug-in module for many other image restoration tasks [37, 38, 4]. Deep neural networks have become the tool of choice for image denoising owing to their ability to learn natural image priors from large-scale datasets. Yet, Deep Image Prior (DIP) [33] requires only a single degraded image for image resotration. Remarkably, DIP shows that a randomly initialized convolutional neural network (CNN) can regularize image restoration through its architecture and early-stopping optimization. This is inspired by the phenomenon that some network architectures act inherently as image priors, favoring generating smooth, natural images and re", + "bbox": [ + 75, + 679, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd7b06dc5cbb5a34ad44df7ce728b9d897eaf43e119c39ed18a5c170b4952f40.jpg", + "image_caption": [ + "Noisy 20.34/0.648" + ], + "image_footnote": [], + "bbox": [ + 501, + 285, + 691, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca844309f68f2a77e921b730de11049ec9e94f261727dbd29fa2054654496d2e.jpg", + "image_caption": [ + "DIP NAS-DIP 8/0.821 21.92/0.680", + "ISNAS-DIP Ours Ground Truth 21.04/0.600 25.22/0.841 PSNR $\\uparrow$ /SSIM", + "Deep Decoder Ours Ground Truth 27.33/0.794 32.01/0.924 PSNR ↑ / SSIM ↑" + ], + "image_footnote": [], + "bbox": [ + 699, + 285, + 890, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e322c5a3024523e2a6b8746a745f90b99b6313a2b1429156322313ef3046831.jpg", + "image_caption": [ + "DIP 28.39/0.824", + "ConvDecoder 28.91/0.853", + "Figure 1: Top: Performance (SSIM↑) under different levels of Gaussian noises. Bottom: Denoising of a fine-grained $(1^{st}\\mathrm{row})$ and a coarse-grained image $(2^{nd}\\mathrm{row})$ . Most existing methods, including the recent image-specific ISNAS-DIP[1], struggle to perform well simultaneously in both cases. Our simple strategies are flexible in image-specific architectural adaptation without requiring a search. Moreover, the results of the lightweight ConvDecoder[10] and Deep Decoder[15] suggest that without proper model setups, under-parameterization itself can neither ensure good denoising performance nor remove the need for early-stopping." + ], + "image_footnote": [], + "bbox": [ + 514, + 467, + 578, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sisting noises or degradations. However, image denoising with DIP is greatly influenced by architectural design [3, 1, 33, 15, 17], and the associated tendency to fit the original noisy image, i.e., over-fitting [15, 16].", + "bbox": [ + 496, + 744, + 890, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Architectural design for DIP remains an open problem. One prevailing view is that model under-parameterization limits noise over-fitting and thus mitigates the need for early stopping [15]. However, our experiments reveal that multiple model architectures can exist under a similar parameter budget, where inappropriate model setups can still lead to", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* These authors contribute equally.", + "bbox": [ + 96, + 875, + 287, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/YilinLiu97/FasterDIP-devil-in-upsampling.git", + "bbox": [ + 96, + 887, + 446, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12408", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "noise fitting and over-smoothing (Fig.1, Fig.3). Another line of work automates the architecture identification using Neural Architecture Search (NAS)[6, 17, 1]. Without prior knowledge on suitable architectures, extensive search incurs substantial computational costs, prohibiting image-wise NAS for optimal restoration [1]. Arican et al. [1] narrow the search space using training-free metrics, but the need for candidate comparison dramatically prolongs the restoration time ( $\\sim$ 7 hours/image). Moreover, NAS-based models, along with many DIP models, are typically heavily parameterized and prone to over-fitting, as corroborated in our experiments. Thus, their performance largely depends on the timing of early stopping, which is typically image-specific and hard to pinpoint without access to ground truth.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Directly identifying an effective under-parameterized architecture for each image is challenging in light of the vast number of different architectures and the absence of ground truth for explicit supervision. To simplify the architectural decisions for DIP, we rethink the architectural influences on its performance in the context of image denoising.", + "bbox": [ + 75, + 304, + 468, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We start by noting that denoising performance is attributable to only a few components, primarily the unlearnt upsampling operations. Our frequency analysis reveals that the fixed upsampling operations tend to bias the architecture towards low-frequency contents more strongly than linear or convolutional layers, critically influencing both the peak PSNR and the point of early stopping to avoid noise-fitting.", + "bbox": [ + 75, + 397, + 468, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Importantly, this finding leads to empirical discovery on the roles of typical architectural components in DIP: assuming a standard hourglass network, i) simply scaling the depth and width can balance smoothing and preservation of details, due to the low-pass filtering effects of the upsampling operations inserted in-between the layers. As we observed, a wider and shallower network is better at preserving details and therefore suitable for fine-grained images. This suggests that the \"optimal\" architecture can vary across images, and image texture should be considered for more effective denoising. ii) Skip connections make a deep network perform similarly as a shallower one likely by reducing the \"effective upsampling rate\". This implies the possibility of discarding skip connections to simplify DIP architectural design.", + "bbox": [ + 75, + 505, + 470, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on these insights, we find it sufficient to restrict the design choices to only the network depth and width, reducing the problem to searching through only a handful of sub-networks and adapting them according to the level of image details. This can be done as a pre-processing step without costly searching or evaluation. We show that this simple strategy works with both the hourglass and decoder structures, and with proper setups, the estimated networks can denoise while preserving the details better than the larger networks with only $5\\% \\sim 40\\%$ number of parameters. The resulting under-parameterization alleviates the", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "need for early stopping. Our contributions are as follows:", + "bbox": [ + 498, + 90, + 879, + 107 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We pinpoint that unlearnt upsampling is the main driving force behind the spectral bias of DIP.", + "- Leveraging this insight, we empirically identify the influences of depth, width and skip connections, along with their correlations with image texture, allowing for quick, effective and more interpretable architectural design for every image without the laborious search.", + "- We are the first to associate DIP architectural design with image texture. To promote future research, we build a Texture-DIP Dataset consisting of images from three popular datasets, reclassified into several predefined width choices – validated through experiments, according to textural complexity.", + "- We show that with proper setups, a highly underparameterized subnetwork could match and even outperform larger counterparts, especially at a higher noise level. We conducted extensive synthetic and real-world noise removal experiments to validate our findings and approach." + ], + "bbox": [ + 517, + 112, + 893, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 436, + 640, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DIP Variants. Deep Decoder [15] is an underparameterized network proposed to avoid early stopping. However, our investigations show that underparameterization alone is not sufficient for denoising. For instance, ConvDecoder [10], a convolutional variant of Deep Decoder, contains more parameters but demonstrates less tendency to over-fit (Fig 1). In contrast to NAS-based methods[6, 17, 1], our strategy leverages the observed relationship between the architecture and the image to prevent the exhaustive search. Other variants such as DIP-RED [24] and DIP-TV [20] augment DIP with additional priors.", + "bbox": [ + 496, + 462, + 890, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early-stopping criterion. Cheng et al. [7] perform posterior inference in DIP so as to prevent the need for early stopping by adding Gaussian noise to the gradients. Similarly, Shi et al. [29] regularize the matrix norm of the network weights to alleviate performance decay. However, tuning the regularization granularity can be challenging (Suppl. B). Jo et al. [18] combine DIP with Stein's unbiased risk estimator (SURE) [32] for training without clean images, but SURE is limited to only a few known noise types [31, 23]. Wang et al. [35] propose to track the running variance of the outputs, but this also introduces new hyper-parameters that require non-trivial tuning.", + "bbox": [ + 496, + 628, + 890, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Investigation and Method", + "text_level": 1, + "bbox": [ + 500, + 821, + 740, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 847, + 640, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Image Prior. A noisy image $\\mathbf{y} \\in \\mathbb{R}^N$ can be modeled as: $\\mathbf{y} = \\mathbf{x} + \\mathbf{n}$ , where $\\mathbf{x} \\in \\mathbb{R}^N$ is the clean counterpart to", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12409", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/336a54dcb45b4de8b5f091f3c4d7770b40ff40f20124a31f58824faa30a4195e.jpg", + "image_caption": [ + "Figure 2: Spectral density of images of different levels of textural complexity, obtained by azimuthally integrating over the power spectrum of the image." + ], + "image_footnote": [], + "bbox": [ + 119, + 90, + 405, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be recovered and $\\mathbf{n}$ is assumed to be $i.i.d$ . Gaussian Noise drawn from $\\mathcal{N}(0, \\sigma^2\\mathbf{I})$ with $\\mathbf{I}$ being the identity matrix. DIP parameterizes the clean image $\\mathbf{x}$ via a network $\\mathbf{G}_{\\theta}$ and is optimized to fit the noisy image $\\mathbf{y}$ , formulated as:", + "bbox": [ + 75, + 300, + 470, + 361 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} = \\underset {\\theta} {\\arg \\min } \\mathcal {L} (\\mathbf {y}; \\mathbf {G} _ {\\theta} (\\mathbf {z})), \\quad \\mathbf {x} ^ {*} = \\mathbf {G} _ {\\theta^ {*}} (\\mathbf {z}), \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 375, + 468, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{z}$ denotes the fixed white noise input tensor. Such parameterization allows lower-frequency contents to be fitted before the higher-frequency ones [3, 29], exhibiting high impedance to image noises or degradations. Early-stopping is often required to prevent excessive high frequencies from being fitted, such as noise.", + "bbox": [ + 75, + 411, + 468, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image complexity. Ideally, DIP should denoise an image while preserving textural details. We define image complexity by its texture, which can be characterized by its power spectral density (PSD). The spectral power of a natural image typically follows an exponential decay from low frequencies to high frequencies [30]. High-frequency components correspond to fine features such as details, while low-frequency components correspond to coarse structures. Hence, fine-grained images contain more high frequencies than coarse-grained images, giving a flatter PSD curve in (Fig.2). Each image is scored based on its texture features, as described in Sec.3.5.", + "bbox": [ + 75, + 502, + 470, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. The importance of upsampling", + "text_level": 1, + "bbox": [ + 76, + 695, + 349, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first identified the core architecture components that affect the denoising performance of DIP. To this end, we analyzed a decoder-only architecture by removing the encoder from the typical encoder-decoder architecture [33], since a decoder is the minimum requirement for reconstructing the final image. Our base model for analysis is a 6-layered convolutional decoder (Conv-Decoder[10]), with $128 \\times 3$ filters per layer except for the last regression layer, followed by batch normalization, ReLU activation function and upsampling. We further simplify the setup by replacing the spatial filters with pixel-wise $1 \\times 1$ filters, constructing a non-convolutional variant dubbed MLP-Decoder.", + "bbox": [ + 75, + 719, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1509a510253d9b6a0f16ce25070f2617c3e0561aa8977fefe6c4a92d046c0c5c.jpg", + "image_caption": [ + "Figure 3: Influences of architecture components on image denoising. (a) Without convolutional layers, upsampling still enables denoising. Transposed convolutions lead to faster performance drop due to earlier noise-fitting than bilinear upsampling. Removing the upsampling decreases denoising capability, which cannot be compensated by simply reducing the number of layers, i.e., under-parameterization. (b) Convolutional layers alone exhibit certain denoising effects but necessitate early stopping. Increasing the number of convolutional layers achieves higher peak PSNR at the expense of earlier noise fitting (see red arrows). Better results achieved when combined with upsampling." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 220 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Results for MLP-Decoder, shown in Fig.3 (a), suggest that upsampling plays a vital role. Removal of upsampling results in significant performance degradation of denoising, which cannot be compensated by simply reducing the size of the network for under-parameterization. This holds true for other tasks such as super-resolution (Suppl.A). Fig.3 (b) shows that learnable spatial filters alone also enable denoising, in contrast to the pixel-wise filters, re-affirming the frequency bias of the convolutional layers [3]. However, the effects vanish as the network size increases and noise-fitting occurs sonner. Convolutional layers together with upsampling achieve better results as manifested in the higher peak PSNR and longer denoising effects.", + "bbox": [ + 496, + 410, + 890, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discussion. These results suggest that an appropriate upsampling operation is crucial for effective network image priors, and that under-parameterization alone is insufficient. Different upsampling operations induce varying extents of denoising effects: transposed convolutions [25] tend to fit noise faster than bilinear upsampling, necessitating early stopping. In the next section, we investigate the behaviors of these upsampling operations from a signal-processing perspective to gain insights into their influences on denoising performance.", + "bbox": [ + 496, + 607, + 890, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Spectral effects of upsampling", + "text_level": 1, + "bbox": [ + 500, + 770, + 767, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The design choice for upsampling is typically standard: bilinear/nearest neighbor (NN) interpolation or transposed convolutions. These upsampling operations can all be decomposed into two steps: (i) zero-insertion and (ii) filtering. Given a target upsampling factor R, the low-resolution feature map is first interleaved with (R-1) rows/columns of zeros to increase its sampling rate, and then convolved with", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12410", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ddcf37cccf623cbe135d083ec5ef6a96a7eec4aa25faef68756d24bba29ef6b9.jpg", + "image_caption": [ + "Figure 4: (Left) Frequency responses of the tested LPFs. Different LPFs result in upsamplers with different extents of smoothing. NN interpolation preserves most signals in the passband but also the high-frequency replica in the stopband; $\\mathcal{L}_{-100}$ attenuates the signals most significantly ( $\\sim 100dB$ ) and suppresses the high-frequency replica most. (Right) Denoising results on coarse- and fine-textured images from Set9. Top rows: Peak PSNR values. Bottom rows: PSNR values at the last training iteration." + ], + "image_footnote": [], + "bbox": [ + 80, + 99, + 468, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a low-pass filter (LPF) to remove the alias high frequencies introduced by zero-insertion. The key difference between the upsampling operations lies in the nature of the filters. The filters for transposed convolutions are learnable, but are fixed for bilinear and nearest neighbor upsampling.", + "bbox": [ + 75, + 393, + 468, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To better illustrate this, we consider the case of a 1D signal $x(n), n = 0, \\dots, N - 1$ , and its discrete Fourier representation $X(k) = \\sum_{n=0}^{N-1} x(n) e^{\\frac{-i2\\pi}{N} k n}, k = 0, \\dots, N - 1$ .", + "bbox": [ + 76, + 469, + 468, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For an upsampling factor of 2, we have", + "bbox": [ + 96, + 517, + 357, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} X ^ {\\mathrm {u p}} (\\hat {k}) = \\sum_ {n = 0} ^ {2 N - 1} x ^ {\\mathrm {u p}} (n) e ^ {- \\frac {i 2 \\pi}{2 N} \\hat {k} n} (2) \\\\ = \\sum_ {n = 0} ^ {N - 1} x ^ {\\mathrm {u p}} (2 n) e ^ {\\frac {- i 2 \\pi}{2 N} (2 n) \\hat {k}} + \\sum_ {n = 0} ^ {N - 1} x ^ {\\mathrm {u p}} (2 n + 1) e ^ {\\frac {- i 2 \\pi}{2 N} (2 n + 1) \\hat {k}}, (3) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 542, + 478, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{k} = 0,\\dots ,2N - 1,x^{\\mathrm{up}}(2n) = x(n),x^{\\mathrm{up}}(2n + 1) = 0$ due to interleaved zero insertion. Hence, for $0\\leq \\hat{k} < N$ $X^{\\mathrm{up}}(\\hat{k}) = X(k)$", + "bbox": [ + 76, + 655, + 468, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For $\\hat{k} \\geq N$ , let $k' = \\hat{k} - N$ , $k' = 0, \\dots, N - 1$ , we have", + "bbox": [ + 94, + 704, + 465, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} X ^ {\\mathrm {u p}} (\\hat {k}) = \\sum_ {n = 0} ^ {N - 1} x (n) e ^ {- \\frac {i 2 \\pi}{2 N} \\left(k ^ {\\prime} + N\\right) 2 n} (4) \\\\ = \\sum_ {n = 0} ^ {N - 1} x (n) e ^ {\\frac {- i 2 \\pi}{N} n k ^ {\\prime} - i 2 n \\pi} = X \\left(k ^ {\\prime}\\right), (5) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 729, + 468, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Eq.5 exploits the periodicity of the complex exponential function. Thus, zero-insertion will preserve the original spectrum at $[0,N]$ (passband) and additionally create a mirrored copy at $[N,2N - 1]$ (stopband). The high-frequency replica beyond $N$ should be suppressed by the", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/52c7e87bfed4a05d4da40f74bd39c56d70bbb0d36286edf81546ac6ee8b64a41.jpg", + "table_caption": [ + "Table 1: The iteration number [iter./5000] at the peak PSNR for different upsamplers. The upsamplers are designed to mainly differ in the stopband. The peak PSNR is reached more slowly when the attenuation is stronger (from left to right)." + ], + "table_footnote": [], + "table_body": "
NNL14L15L16L17
Coarse-grained17831589168122052214
Fine-grained24242942389843614957
", + "bbox": [ + 537, + 155, + 849, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "subsequent LPF to avoid image artifacts. According to duality and convolution theorem, convolving with NN or bilinear interpolation filter is equivalent to multiplication of $X^{up}(\\hat{k})$ with a Sinc or Sinc² function corresponding to low-pass filtering, while transposed convolutional filter may not necessarily be low-pass as it depends on the optimization objective.", + "bbox": [ + 498, + 234, + 890, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We experimentally demonstrate that different upsampling operations bias the architecture towards different spectral properties. Specifically, we constructed four upsamplers by first interleaving the input with zeros and then convolving it with handcrafted LPFs (shortened as $\\mathcal{L}$ ): $\\mathcal{L}14$ , $\\mathcal{L}15$ , $\\mathcal{L}_{-60}$ and $\\mathcal{L}_{-100}$ , with the subscript denoting the decayed dB. By construction, $\\mathcal{L}14$ and $\\mathcal{L}15$ are very close to NN in the passband ( $< 0.03dB$ ) and only differ in the stopband. The frequency responses of the compared LPFs are detailed in Fig. 4. We applied the customized upsamplers on ConvDecoder and tested them on the fine- and coarse-textured images from Set9 respectively. Similar findings hold for the encoder-decoder architecture (Suppl. C).", + "bbox": [ + 496, + 340, + 892, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From Fig. 4, upsampling critically influences both the peak PSNR value and the timing for early stopping with respect to images of different textural complexities. Upsamplers (NN, $\\mathcal{L}14$ , $\\mathcal{L}15$ ) with less attenuation on the high-frequency replica are beneficial for generating fine-grained images, but they tend to cause noise-fitting especially for coarser-grained images. Also, they are generally the fastest to reach the peak PSNR (Table 1). This explains why the transposed convolution requires early-stopping, as the filters may not learn to attenuate the introduced high frequencies effectively. Similar issues with the learned upsampling are also prevalent in generative models [28, 5, 11, 12].", + "bbox": [ + 496, + 537, + 890, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the other hand, bilinear and $\\mathcal{L}_{-60}$ exhibit a greater amount of attenuation, more strongly biasing the network against high frequencies and leading to longer-lasting denoising effects. They turn out to work sufficiently well for both kinds of images, especially on the coarser-grained ones which are typically the majority in the dataset. $\\mathrm{LPF}_{-100}$ tends to over-smooth the output and performs the worst as it attenuates both the passband and stopband signals substantially, though not requiring early stopping.", + "bbox": [ + 496, + 719, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion. These results lead us to conclude that the upsamplers with fixed LPFs are key to the denoising effects of DIP for their tendency towards smooth images (i.e., re", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12411", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b3773ab6ef4badce1d35f251185add8ecc5c54dd7dce50b1077dc48aaf7e6882.jpg", + "image_caption": [ + "Figure 5: Influences of width, depth, and skip connections, assuming upsampling inserted in-between the layers. (a) Tendency to over-fit coarse-grained images increases with width. SSIM scores averaged across the depths. (b) Tendency to over-smooth fine-grained images increases with depth, but alleviated by skip connections. SSIM scores are averaged across the widths.", + "Figure 6: Position of upsampling. Placing the upsampling close to the input (or encoder) can easily cause over-fitting, regardless of the scaling factor. When close to the output (end of the decoder), upsampling with large scaling factors (e.g., $32 \\times$ ) causes over-smoothing." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 472, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "duced high-frequency contents), which aligns well with the spectral statistics of natural images (Fig.2). Probably due to a good balance between the denoising performance and persistency, bilinear upsampling has been widely adopted in DIP models for various applications [33, 15, 13, 16].", + "bbox": [ + 75, + 458, + 468, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Interactions with other architecture elements", + "text_level": 1, + "bbox": [ + 76, + 544, + 459, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After establishing the significance of upsampling, we will now consider how it may interact with other common architectural elements in affecting the ultimate output.", + "bbox": [ + 75, + 568, + 468, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Convolutions + non-linearity. Ideal upsampling does not modify the signal representations but only expands the spectrum for the subsequent layers to add new frequency contents. Convolution followed by nonlinearity such as ReLU, is the only operation capable of introducing arbitrarily high frequencies [19]. Increasing the number of layers (depth) or channels (width) enhances the capability of generating new high frequencies, as theoretically and empirically proved in [26].", + "bbox": [ + 75, + 614, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, using an excessive number of layers or channels can accelerate the learning of both details and noise. However, the effects can be attenuated by the fixed upsampling operations between the layers. As shown in Fig.5, when using fewer upsampling operations (i.e., a shallower network), increasing the width of the network increases the tendency to cause overfitting on simpler images while benefiting the more complex images. Increasing the number of upsampling operations (i.e., a deeper network) can alleviate over-fitting with stronger attenuation but results in", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/136d16f28ea08ca0d7873758b72c1e9c5c13c4577fed7c728787dc63ae32fb88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 90, + 854, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "blurry outputs for fine-grained images. Increasing only the upsampling factors without adding more layers can make the output even more blurry (Fig.6). In other words, the final output is determined by the balance between the generation of high frequencies by the layers and signal attenuation caused by the upsampling operations.", + "bbox": [ + 496, + 361, + 890, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Skip connections between the encoder and the decoder often complicate the design space [6]. While they are not directly responsible for denoising, they may lower the effective upsampling rate, making deep networks perform similarly to shallower ones, i.e., resulting in lesser suppression of the high-frequency replica. This finding is validated on a large-scale experiment comprising 7424 architectures (Suppl. D). As exemplified in Fig. 5 (b), skip connections notably ameliorate the over-smoothing issue induced by the same deep network. Overall, skip connections exert a more pronounced influence on the deeper networks compared to the shallower ones, as evident in the smaller deviation observed when depth $\\leq 3$ in Fig.5 (b).", + "bbox": [ + 496, + 453, + 892, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Practical application to architectural design", + "text_level": 1, + "bbox": [ + 500, + 662, + 870, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the above findings and analysis, we argue that it is possible to estimate an effective architecture for each image without extensive search. Assuming every decoder layer is followed by a $2 \\times$ bilinear upsampling layer, our strategies are as follows:", + "bbox": [ + 496, + 686, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Depth estimation. We have shown that increased depth tends to over-smooth the output, affecting fine-grained images much more than on coarse-grained images. For fine-grained images, we have three options: a) add more skip connections to a deep network; b) simply use a shallower one; c) keep all the layers but reduce down-/up-sampling layers. We recommend c for decoder-only architectures since they are already under-parameterized; for hourglass network this can easily lead to over-fitting (Fig.6). We find", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12412", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f141af135de27177544c1a98154a26eb4ab025584958a34ac6301c8291e7222a.jpg", + "table_caption": [ + "Table 2: Quantitative results on Gaussian noise. $\\sigma$ denotes the noise level. All methods were trained with a fixed iteration number (3000) throughout the experiments. The highest score is in bold, and the second highest is underlined." + ], + "table_footnote": [], + "table_body": "
DatasetsDIP [33]Deep Decoder [15]ConvDecoder [10]NAS-DIP [6]ISNAS-DIP [1]Ours
σ = 25
Set9 [9]PSNR30.1028.4528.5126.3729.1130.26
SSIM0.8930.8480.8540.7530.8620.900
Set12 [37]PSNR26.9725.9825.7820.8624.1028.14
SSIM0.8120.7890.7860.5340.7450.884
CBSD68 [27]PSNR28.9325.5025.1923.8024.5128.57
SSIM0.8920.8090.7930.6930.7450.888
σ = 50
Set9 [9]PSNR25.0425.2225.0121.0723.9126.13
SSIM0.7610.7640.7690.5930.6980.833
Set12 [37]PSNR22.1520.4422.7218.9219.2024.59
SSIM0.6230.6870.7060.4760.5370.805
CBSD68 [27]PSNR23.7423.5224.0617.9219.9324.17
SSIM0.7460.7250.7670.3230.5730.774
# Params (Millions)2.3M0.1M0.89M4.4MVaried0.05M~0.92M
", + "bbox": [ + 138, + 128, + 831, + 344 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/45628dba47b89aaffc631b486f58596d2808b2818e39c40189232e9ec56d72de.jpg", + "table_caption": [ + "Table 3: Comparisons of desired properties. Restoration time is computed on an image of size $512 \\times 512$ with 3000 iterations." + ], + "table_footnote": [], + "table_body": "
NAS [6]ISNAS [1]Ours
Image-SpecificX✓✓
Architecture Search3 days5 mins-
Per-image Restoration~23 mins~7 hrs~6 mins
Early Stopping Required?YesYesNo
", + "bbox": [ + 104, + 406, + 434, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbf{b}$ is generally better than $\\mathbf{a}$ in trading off between good performance and the need for early stopping, especially under higher-level noise, as shown in our results section. This also holds for coarse-grained images as they are not sensitive to depth. More specifically, we find a 2-level hourglass network sufficient for both types of images (Suppl. D).", + "bbox": [ + 75, + 502, + 468, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Width estimation. Width is crucial for learning sufficient details while avoiding over-fitting especially to a shallow network (less signal attenuation). We find that the width needs to be set according to textural complexity of the image: a finer-grained image requires more channels per layer and vice versa. To further validate this, we treated width estimation as a classification problem, and trained three SVMs [8] with texture features as the inputs to classify the images into three width choices \\{32, 64, 128\\}. Note that these widths were empirically chosen for the datasets used in this work and are by no means optimal for all cases, but tuning for other images should be straightforward. The classification results and analysis are in Sec.4.4.", + "bbox": [ + 75, + 595, + 470, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image scoring. To more robustly classify image texture, we extracted both spatial and frequency features and trained a Decision Tree [22] for feature selection. We used the classic Gray Level Co-occurrence Matrix (GLCM) [14] for deriving the spatial texture features from each image. The most useful features turn out to be the following: correlations measured at $0^{\\circ}$ , homogeneity at $45^{\\circ}$ , and contrast at", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f1dbb0f52bd4533939d9e89658830faad14aee39f3c9a3141a22d78cc32ab6d5.jpg", + "table_caption": [ + "Table 4: Quantitative evaluation on Poisson noise. Deep Decoder shortened as DD and ConvDecoder as CD." + ], + "table_footnote": [], + "table_body": "
Noise scaleDIPDDCDNASISNASOurs
ζ = 0.01PSNR31.5829.8129.5129.7030.1230.95
SSIM0.9150.8800.8750.8640.8750.910
ζ = 0.1PSNR22.6623.9124.9415.7217.6824.99
SSIM0.6400.7180.7760.4170.4960.789
ζ = 0.2PSNR20.3621.1321.8712.9914.6422.55
SSIM0.5630.6090.6720.3110.3830.774
", + "bbox": [ + 500, + 406, + 887, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$0^{\\circ}$ . Frequency features are captured using a 1D PSD vector obtained by first converting the 2D PSD from Cartesian coordinates $S(k,l)$ to polar coordinates, and then azimuthally averaging over $\\theta$ , defined as $\\hat{S}(r) = \\frac{1}{2\\pi} \\int_{0}^{2\\pi} S(r,\\theta)d\\theta$ with $r = \\sqrt{k^2 + l^2}$ and $\\theta = \\mathrm{atan2}(k,l)$ , which represents the mean magnitude of the frequencies with respect to the radial distance $r$ . Refer to Suppl. E for more details.", + "bbox": [ + 498, + 550, + 890, + 661 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 690, + 633, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 720, + 718, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted the experiments on three popular datasets and a real-world noisy dataset: Set9 [9] consisting of 9 colored images, Set12 [37] consisting of 12 grey-scaled images, CBSD68 [27] consisting of 68 colored images, and PolyU [36] consisting of 100 real noisy and clean image pairs. We first report our results with 2-level hourglass architecture with the same components as in DIP [33] when comparing with the existing methods, and then extend our strategy to ConvDecoder [10], a decoder-only architecture. All models were trained for 3000 iterations.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12413", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ed947076a7849f316d2efa54dcc6c5c499c2b5ee6f2882d3d6704c78a66a2fe.jpg", + "image_caption": [ + "Noisy (20.27/0.583)" + ], + "image_footnote": [], + "bbox": [ + 178, + 103, + 326, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1469b980b9ea844c4e685756dc96e1d7670331c9b29ec2b88987bad293e04a45.jpg", + "image_caption": [ + "ConvDecoder (25.19/0.781)" + ], + "image_footnote": [], + "bbox": [ + 330, + 103, + 480, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab3d3054eaba591aeb2c8982926ef77599bd75500f3947f44f00289bdf614231.jpg", + "image_caption": [ + "Deep Decoder (25.36/0.789)" + ], + "image_footnote": [], + "bbox": [ + 483, + 103, + 633, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3f4e4196aeadf937c5961b77a6513ebb78d759fcb37d6f7824be0c25a2498491.jpg", + "image_caption": [ + "ISNAS-DIP (25.43/0.779)" + ], + "image_footnote": [], + "bbox": [ + 638, + 103, + 787, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06ef85d3d530455ece4bd00addffec3d56f2580e88d1a35ab6464b4b338be4cc.jpg", + "image_caption": [ + "NAS (26.72/0.836)" + ], + "image_footnote": [], + "bbox": [ + 178, + 229, + 328, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1323e7f67f513cdab527ec41de7554210b98e8c690f893eef3613f6f1b562af4.jpg", + "image_caption": [ + "DIP (27.86/0.874)" + ], + "image_footnote": [], + "bbox": [ + 331, + 229, + 480, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67582ec77a84183f7de8db533caf1d60fa81111476ec83e3603175f6bd4ea176.jpg", + "image_caption": [ + "Ours (28.18/0.882)", + "Figure 7: Denoising results on a fine-grained image (\"kodim01\" from Set9) with Gaussian noise ( $\\sigma = 25$ ). Our estimated architecture for this image is a two-level hourglass network with one skip connection and 128 channels, which is much smaller than others." + ], + "image_footnote": [], + "bbox": [ + 483, + 229, + 633, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b77b57c8e2de72f074de6fca9db92b036d2082d7207c8aa720cf50e6834b84f.jpg", + "image_caption": [ + "Ground Truth (PSNR/SSIM)" + ], + "image_footnote": [], + "bbox": [ + 638, + 229, + 787, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0058c9b2fdfa291bdb1c390a12ffb53bb03366b450d5619c45958d57927e0adb.jpg", + "image_caption": [ + "Noisy (36.42/0.967)" + ], + "image_footnote": [], + "bbox": [ + 96, + 419, + 210, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75df0402a1c28fcbad7ba1d2e3ab25b19e25436b164dd8545a6790faeff6e2ae.jpg", + "image_caption": [ + "ConvDecoder (35.02/0.951)" + ], + "image_footnote": [], + "bbox": [ + 212, + 419, + 326, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad64a683a5131ea5d44efeb63e963c282884fd89bfd817921d4149cf8b5f3d6d.jpg", + "image_caption": [ + "Deep Decoder (35.37/0.955)" + ], + "image_footnote": [], + "bbox": [ + 328, + 419, + 442, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1655adbd05a377d4846d7d2310229dad9abf46080f24bc0915b94a84e9264ce.jpg", + "image_caption": [ + "DIP w/o. skip (35.14/0.951)" + ], + "image_footnote": [], + "bbox": [ + 96, + 518, + 210, + 607 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc48ba180b0a394d4ed183f2e2c5fb91ded25c4e4a34caba6d541873bfff2a2f.jpg", + "image_caption": [ + "DIP w. skip (36.38/0.964)" + ], + "image_footnote": [], + "bbox": [ + 212, + 518, + 326, + 607 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/715e737117a44b78e9e39dca6ef387335141163f0b6139a788e680989667ff40.jpg", + "image_caption": [ + "ISNAS-DIP (36.86/0.968)" + ], + "image_footnote": [], + "bbox": [ + 328, + 518, + 442, + 607 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69abfed3152aacbe7dd8d90197074fd1bd81ef8bea3b961383991eb84d056d08.jpg", + "image_caption": [ + "NAS-DIP (37.18/0.972)", + "Figure 8: Denoising results on a real-world noisy image." + ], + "image_footnote": [], + "bbox": [ + 96, + 619, + 210, + 708 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/accfa68a40c9f4ce01eebbf3d9770701c5d0fd0c999bdb1d2d510fd468fc2ab4.jpg", + "image_caption": [ + "Ours (37.99/0.979)" + ], + "image_footnote": [], + "bbox": [ + 212, + 619, + 326, + 708 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/367a38b6107b30610a8d7141364f87029972d41d242bbfc99d2deb110ebe0e5a.jpg", + "image_caption": [ + "Ground Truth (PSNR/SSIM)" + ], + "image_footnote": [], + "bbox": [ + 328, + 619, + 442, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. DIP variants", + "text_level": 1, + "bbox": [ + 76, + 767, + 210, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Gaussian Noise. Compared with the base DIP network, the properly-designed under-parameterized networks estimated with our strategy perform on par at a mild noise level while excel at a higher noise level (Table 2). This cannot be solely explained by under-parameterization since Deep Decoder and ConvDecoder contain similar or even fewer parameters while are unable to achieve similar results. Fig.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7 shows that a shallow and wide network can preserve the details better than many deeper ones. Note that DIP is a 5-level hourglass network with full skip connections, which by our standard can also well preserve the details. In fact, it is a very strong baseline under mild noise. NAS-DIP and ISNAS-DIP suffer from various degrees of over-fitting on different datasets. This also suggests that the point of optimal stopping varies across images. Besides, they are time-intensive in either searching or evaluation (Table 3). Similar conclusions hold for Poisson noise, which was tested on Set9 [9] as shown in Table 4.", + "bbox": [ + 498, + 407, + 890, + 573 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/656015fdbe9d687cd9928dd5ed3701a7a1d186f835135b01f3674148a159ee4e.jpg", + "table_caption": [ + "Table 5: Quantitative evaluation on PolyU, a real-world noisy dataset. Deep Decoder and ConvDecoder shortened as DD, CD." + ], + "table_footnote": [], + "table_body": "
DIPDDCDNASISNASOurs
PSNR38.1537.2237.0037.8337.7838.05
SSIM0.9820.9780.9760.9820.9770.984
", + "bbox": [ + 509, + 627, + 875, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real-World Noise. We additionally evaluated all methods on the PolyU dataset [36]. We applied again the two-level hourglass architectures with variable width estimated for the images. Table 5 summarizes the numerical results. Fig.9 shows that our method tends to preserve more details, though this may not be reflected by the metrics. Simply removing the skip connections from DIP causing blurring, similar to the decoder-only architectures.", + "bbox": [ + 498, + 714, + 890, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Decoder-only Architectures", + "text_level": 1, + "bbox": [ + 500, + 845, + 746, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We applied our strategy to ConvDecoder and tested it on all three datasets. Here we keep all 5 layers but remove", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12414", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/28aeaa1c0c8c8d047aab2abceef81fe065a4151db3b3301da5219c2f6da5866b.jpg", + "table_caption": [ + "Table 6: Application to ConvDecoder." + ], + "table_footnote": [], + "table_body": "
Datasetsσ = 25σ = 50
BeforeAfterBeforeAfter
Set9PSNR28.5128.7425.0125.11
SSIM0.8540.8730.7690.784
Set12PSNR25.7926.9822.7223.01
SSIM0.7860.8540.7060.742
CBSD68PSNR25.1928.2924.3624.12
SSIM0.7930.8770.7670.768
# Params (Millions)0.89M0.06M~0.89M0.89M0.06M~0.89M
", + "bbox": [ + 78, + 114, + 468, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f00b73fd16c073224f1444236f30f04e25bef6a6c304715618d2358f5c9ae981.jpg", + "image_caption": [ + "Figure 10: (a) Width critically influences a shallower network, while (b) it rarely has any impact on a sufficiently deep network." + ], + "image_footnote": [], + "bbox": [ + 78, + 250, + 468, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a certain number of upsampling layers and scale the width accordingly for the images. These simple changes effectively alleviate the over-smoothing issue that often brought by ConvDecoder as shown in Fig.8 and improve the quantitative results as shown in Table 6.", + "bbox": [ + 75, + 455, + 468, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. More Analysis on Depth and Width", + "text_level": 1, + "bbox": [ + 76, + 542, + 385, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In practice, we find it more efficient to first determine the depth, and then the width. This is because when the network is deep enough (strong attenuation due to upsampling), width becomes less influential, as evident in Fig.10 (b) and the small deviation in Fig.5 (b). However, to relax the need for early stopping at a higher noise level, one may prefer a shallower one, where width matters more (Fig.10 (a)). In this regard, we use the texture features of the images to predict the desired width. This makes intuitive sense as noise-fitting is associated with the amount of high frequency contents in the image, which is manifested in its texture. This is corroborated by the 0.86 Micro-average AUC score shown in Fig.11 (a). The \"optimal\" width labels we used to train the classifiers were obtained by experimenting with all three width choices on a 2-level hourglass architecture. These width labels are also applicable to decoder architectures as demonstrated by our experiments.", + "bbox": [ + 75, + 566, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although the width choices seem limited, we found the images robust to the choice of width to some extent, and simple averaging also works well for ambiguous cases (Appendix D). In fact, some images have multiple width labels. We included these cases in the released Texture-DIP dataset.", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5754ede4b38c3a18f2c7260c34f2cc2734a08603c2a0cb3c43a85a57863d98e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 92, + 694, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0aae8a7306d6ea4f82205c810e0384c9b0a54ebacacdf046dc32d0e52274ccdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 92, + 893, + 217 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7eeacf01310fd3804c86f5d5045c665fc81138200bfdda49fe2f9013737f218d.jpg", + "image_caption": [ + "Figure 9: Qualitative improvement on ConvDecoder simply by removing three upsampling layers in this case while preserving all the convolutional layers." + ], + "image_footnote": [], + "bbox": [ + 526, + 287, + 754, + 439 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/023522865d6b37451d03b9139f8be488c6b04d5796bf1fd4903996ed75ae284a.jpg", + "image_caption": [ + "Figure 11: (a) ROC curves with AUC scores for width classification on images from Set9, Set12 and CBSD68. 5-fold cross-validation is performed and repeated 10 times. (b) Overview of our Texture-DIP Dataset." + ], + "image_footnote": [], + "bbox": [ + 767, + 286, + 864, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. What about Transformers?", + "text_level": 1, + "bbox": [ + 500, + 532, + 745, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transformers [34] have become integral parts of modern deep neural network architectures. We experimented with Swin-UNet [2], an encoder-decoder pure transformer consisting of Swin Transformer blocks [21] and skip connections. Note that Swin-UNet relies on patch merging/expanding layers for $2 \\times$ down-/upsampling without involving unlearnt upsampling. We replaced all the patch expanding layers with $2 \\times$ bilinear upsampling operations and compared the performance of the modified network with the original. We did not replace the patch merging layers with the corresponding downsampling layers or strided convolutions because we did not observe any significant difference. The results presented in Fig. 12 are consistent with our findings on CNNs. For more implementation details, please refer to our code.", + "bbox": [ + 496, + 556, + 890, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 500, + 799, + 769, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented simple but effective solutions to the challenging DIP architectural design in the context of image denoising. Leveraging the spectral effects of upsampling and its relationships with other architectural components, we show that simple architectural changes yield highly-", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12415", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a20380dbac06f8ee4b129d45b992dd3d329d48d091e1a0ff19dcf072f4dd2fde.jpg", + "image_caption": [ + "Noisy (20.38/0.465)" + ], + "image_footnote": [], + "bbox": [ + 156, + 116, + 310, + 275 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0bf8fbc3e4aa8754ff75e318b7c07895151dff8208364b062f46e979f7198bc6.jpg", + "image_caption": [ + "Swin-UNet (21.42/0.497)" + ], + "image_footnote": [], + "bbox": [ + 325, + 114, + 478, + 273 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/765cb6455ab9c03b8d43dd72c56a6cfe4f45c5826ec026b10dfe48c4dfd89b56.jpg", + "image_caption": [ + "Swin-UNet w. Bilinear Up (30.61/0.928)", + "Figure 12: Visual comparisons on the transformer-based Swin-UNet [2] with and without bilinear upsampling." + ], + "image_footnote": [], + "bbox": [ + 491, + 114, + 643, + 273 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/4958944630f6e3e4c6bfda76d1a568a1d2dbeb501d1f9164b9ce574cf566ecea.jpg", + "image_caption": [ + "Ground Truth (PSNR/SSIM)" + ], + "image_footnote": [], + "bbox": [ + 655, + 114, + 807, + 273 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "effective under-parameterized networks that could surpass the larger counterparts and does not critically rely on early-stopping. Although this work focuses on denoising, insights learned about upsampling can be employed in the future to understand the relationship between architectural characteristics and other image restoration tasks. We hope our study could encourage efficient architectural design for DIP and image synthesis in general.", + "bbox": [ + 75, + 325, + 468, + 446 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 459, + 243, + 474 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the United States National Institutes of Health (NIH) under Grant CA266702 and Grant EB008374.", + "bbox": [ + 75, + 483, + 468, + 527 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 541, + 173, + 556 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Metin Ersin Arican, Ozgur Kara, Gustav Bredell, and Ender Konukoglu. Isnas-dip: Image-specific neural architecture search for deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1960-1968, 2022. 1, 2, 6", + "[2] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 8, 9", + "[3] Prithvijit Chakrabarty and Subhransu Maji. The spectral bias of the deep image prior. arXiv preprint arXiv:1912.08905, 2019. 1, 3", + "[4] Stanley H Chan, Xiran Wang, and Omar A Elgendy. Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging, 3(1):84-98, 2016. 1", + "[5] Keshigeyan Chandrasegaran, Ngoc-Trung Tran, and Ngai-Man Cheung. A closer look at fourier spectrum discrepancies for cnn-generated images detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7200-7209, 2021. 4", + "[6] Yun-Chun Chen, Chen Gao, Esther Robb, and Jia-Bin Huang. Nas-dip: Learning deep image prior with neural" + ], + "bbox": [ + 84, + 566, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "architecture search. In European Conference on Computer Vision, pages 442-459. Springer, 2020. 2, 5, 6", + "[7] Zezhou Cheng, Matheus Gadelha, Subhransu Maji, and Daniel Sheldon. A bayesian perspective on the deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5451, 2019. 2", + "[8] Corinna Cortes and Vladimir Vapnik. Support-vector networks. Machine learning, 20:273-297, 1995. 6", + "[9] Kostadin Dabov, Alessandro Foi, and Karen Egiazarian. Video denoising by sparse 3d transform-domain collaborative filtering. In 2007 15th European Signal Processing Conference, pages 145-149. IEEE, 2007. 6, 7", + "[10] Mohammad Zalbagi Darestani and Reinhard Heckel. Accelerated mri with un-trained neural networks. IEEE Transactions on Computational Imaging, 7:724-733, 2021. 1, 2, 3, 6", + "[11] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7890–7899, 2020. 4", + "[12] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 4", + "[13] Yosef Gandelsman, Assaf Shocher, and Michal Irani. \"double-dip\": unsupervised image decomposition via coupled deep-image-priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11026-11035, 2019. 5", + "[14] Robert M Haralick, Karthikeyan Shanmugam, and Its' Hak Dinstein. Textural features for image classification. IEEE Transactions on systems, man, and cybernetics, (6):610-621, 1973. 6", + "[15] Reinhard Heckel and Paul Hand. Deep decoder: Concise image representations from untrained non-convolutional networks. arXiv preprint arXiv:1810.03982, 2018. 1, 2, 5, 6", + "[16] Reinhard Heckel and Mahdi Soltanolkotabi. Denoising and regularization via exploiting the structural bias of convolu" + ], + "bbox": [ + 503, + 327, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12416", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional generators. arXiv preprint arXiv:1910.14634, 2019. 1,5", + "[17] Kary Ho, Andrew Gilbert, Hailin Jin, and John Collomosse. Neural architecture search for deep image prior. Computers & graphics, 98:188-196, 2021. 1, 2", + "[18] Yeonsik Jo, Se Young Chun, and Jonghyun Choi. Rethinking deep image prior for denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5087-5096, 2021. 2", + "[19] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 5", + "[20] Jiaming Liu, Yu Sun, Xiaojian Xu, and Ulugbek S Kamilov. Image restoration using total variation regularized deep image prior. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7715-7719. IEEE, 2019. 2", + "[21] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 8", + "[22] Wei-Yin Loh. Classification and regression trees. Wiley interdisciplinary reviews: data mining and knowledge discovery, 1(1):14-23, 2011. 6", + "[23] Florian Luisier, Thierry Blu, and Michael Unser. Image denoising in mixed poisson-gaussian noise. IEEE Transactions on image processing, 20(3):696-708, 2010. 2", + "[24] Gary Mataev, Peyman Milanfar, and Michael Elad. Deepred: Deep image prior powered by red. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 2", + "[25] Augustus Odena, Vincent Dumoulin, and Chris Olah. Deconvolution and checkerboard artifacts. Distill, 1(10):e3, 2016. 3", + "[26] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 5", + "[27] Stefan Roth and Michael J Black. Fields of experts. International Journal of Computer Vision, 82(2):205-229, 2009. 6", + "[28] Katja Schwarz, Yiyi Liao, and Andreas Geiger. On the frequency bias of generative models. Advances in Neural Information Processing Systems, 34:18126-18136, 2021. 4", + "[29] Zenglin Shi, Pascal Mettes, Subhransu Maji, and Cees GM Snoek. On measuring and controlling the spectral bias of the deep image prior. International Journal of Computer Vision, 130(4):885-908, 2022. 2, 3", + "[30] Eero P Simoncelli and Bruno A Olshausen. Natural image statistics and neural representation. Annual review of neuroscience, 24(1):1193-1216, 2001. 3", + "[31] Shakarim Soltanayev and Se Young Chun. Training deep learning based denoisers without ground truth data. Advances in neural information processing systems, 31, 2018. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Charles M Stein. Estimation of the mean of a multivariate normal distribution. The annals of Statistics, pages 1135-1151, 1981. 2", + "[33] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446-9454, 2018. 1, 3, 5, 6", + "[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 8", + "[35] Hengkang Wang, Taihui Li, Zhong Zhuang, Tiancong Chen, Hengyue Liang, and Ju Sun. Early stopping for deep image prior. arXiv preprint arXiv:2112.06074, 2021. 2", + "[36] Jun Xu, Hui Li, Zhetong Liang, David Zhang, and Lei Zhang. Real-world noisy image denoising: A new benchmark. arXiv preprint arXiv:1804.02603, 2018. 6, 7", + "[37] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 6", + "[38] Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3929-3938, 2017. 1" + ], + "bbox": [ + 501, + 92, + 890, + 446 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12417", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/6b883d6c-8e38-4d25-996f-79941abad307_model.json b/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/6b883d6c-8e38-4d25-996f-79941abad307_model.json new file mode 100644 index 0000000000000000000000000000000000000000..72d128c62406fd4813d92a0920979a66c273fe54 --- /dev/null +++ b/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/6b883d6c-8e38-4d25-996f-79941abad307_model.json @@ -0,0 +1,2541 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.13, + 0.859, + 0.177 + ], + "angle": 0, + "content": "The Devil is in the Upsampling: Architectural Decisions Made Simpler for Denoising with Deep Image Prior" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.188, + 0.773, + 0.208 + ], + "angle": 0, + "content": "Yilin Liu\\(^{1*}\\), Jiang Li\\(^{1*}\\), Yunkui Pang\\(^{1*}\\), Dong Nie\\(^{1}\\), and Pew-Thian Yap\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.21, + 0.662, + 0.228 + ], + "angle": 0, + "content": "1University of North Carolina at Chapel Hill" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.228, + 0.73, + 0.242 + ], + "angle": 0, + "content": "{yilinliu, yunkuipa, dongnie} @cs.unc.edu, jiang.li@unc.edu, ptyap@med.unc.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.319, + 0.474, + 0.624 + ], + "angle": 0, + "content": "Deep Image Prior (DIP) shows that some network architectures inherently tend towards generating smooth images while resisting noise, a phenomenon known as spectral bias. Image denoising is a natural application of this property. Although denoising with DIP mitigates the need for large training sets, two often intertwined practical challenges need to be overcome: architectural design and noise fitting. Existing methods either handcraft or search for suitable architectures from a vast design space, due to the limited understanding of how architectural choices affect the denoising outcome. In this study, we demonstrate from a frequency perspective that unlearnt upsampling is the main driving force behind the denoising phenomenon with DIP. This finding leads to straightforward strategies for identifying a suitable architecture for every image without laborious search. Extensive experiments show that the estimated architectures achieve superior denoising results than existing methods with up to \\(95\\%\\) fewer parameters. Thanks to this under-parameterization, the resulting architectures are less prone to noise-fitting1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.654, + 0.21, + 0.67 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Image denoising is useful on its own and can be a plug-in module for many other image restoration tasks [37, 38, 4]. Deep neural networks have become the tool of choice for image denoising owing to their ability to learn natural image priors from large-scale datasets. Yet, Deep Image Prior (DIP) [33] requires only a single degraded image for image resotration. Remarkably, DIP shows that a randomly initialized convolutional neural network (CNN) can regularize image restoration through its architecture and early-stopping optimization. This is inspired by the phenomenon that some network architectures act inherently as image priors, favoring generating smooth, natural images and re" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.286, + 0.692, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.415, + 0.578, + 0.461 + ], + "angle": 0, + "content": "Noisy 20.34/0.648" + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.415, + 0.693, + 0.461 + ], + "angle": 0, + "content": "DIP NAS-DIP 8/0.821 21.92/0.680" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.286, + 0.891, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.415, + 0.891, + 0.461 + ], + "angle": 0, + "content": "ISNAS-DIP Ours Ground Truth 21.04/0.600 25.22/0.841 PSNR \\(\\uparrow\\) /SSIM" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.468, + 0.579, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.468, + 0.644, + 0.532 + ], + "angle": 0, + "content": "DIP 28.39/0.824" + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.468, + 0.705, + 0.532 + ], + "angle": 0, + "content": "ConvDecoder 28.91/0.853" + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.468, + 0.891, + 0.532 + ], + "angle": 0, + "content": "Deep Decoder Ours Ground Truth 27.33/0.794 32.01/0.924 PSNR ↑ / SSIM ↑" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.559, + 0.895, + 0.713 + ], + "angle": 0, + "content": "Figure 1: Top: Performance (SSIM↑) under different levels of Gaussian noises. Bottom: Denoising of a fine-grained \\((1^{st}\\mathrm{row})\\) and a coarse-grained image \\((2^{nd}\\mathrm{row})\\). Most existing methods, including the recent image-specific ISNAS-DIP[1], struggle to perform well simultaneously in both cases. Our simple strategies are flexible in image-specific architectural adaptation without requiring a search. Moreover, the results of the lightweight ConvDecoder[10] and Deep Decoder[15] suggest that without proper model setups, under-parameterization itself can neither ensure good denoising performance nor remove the need for early-stopping." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.892, + 0.808 + ], + "angle": 0, + "content": "sisting noises or degradations. However, image denoising with DIP is greatly influenced by architectural design [3, 1, 33, 15, 17], and the associated tendency to fit the original noisy image, i.e., over-fitting [15, 16]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Architectural design for DIP remains an open problem. One prevailing view is that model under-parameterization limits noise over-fitting and thus mitigates the need for early stopping [15]. However, our experiments reveal that multiple model architectures can exist under a similar parameter budget, where inappropriate model setups can still lead to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.875, + 0.289, + 0.888 + ], + "angle": 0, + "content": "* These authors contribute equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.447, + 0.901 + ], + "angle": 0, + "content": "1https://github.com/YilinLiu97/FasterDIP-devil-in-upsampling.git" + }, + { + "type": "list", + "bbox": [ + 0.098, + 0.875, + 0.447, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12408" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "noise fitting and over-smoothing (Fig.1, Fig.3). Another line of work automates the architecture identification using Neural Architecture Search (NAS)[6, 17, 1]. Without prior knowledge on suitable architectures, extensive search incurs substantial computational costs, prohibiting image-wise NAS for optimal restoration [1]. Arican et al. [1] narrow the search space using training-free metrics, but the need for candidate comparison dramatically prolongs the restoration time (\\(\\sim\\) 7 hours/image). Moreover, NAS-based models, along with many DIP models, are typically heavily parameterized and prone to over-fitting, as corroborated in our experiments. Thus, their performance largely depends on the timing of early stopping, which is typically image-specific and hard to pinpoint without access to ground truth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.47, + 0.397 + ], + "angle": 0, + "content": "Directly identifying an effective under-parameterized architecture for each image is challenging in light of the vast number of different architectures and the absence of ground truth for explicit supervision. To simplify the architectural decisions for DIP, we rethink the architectural influences on its performance in the context of image denoising." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.47, + 0.504 + ], + "angle": 0, + "content": "We start by noting that denoising performance is attributable to only a few components, primarily the unlearnt upsampling operations. Our frequency analysis reveals that the fixed upsampling operations tend to bias the architecture towards low-frequency contents more strongly than linear or convolutional layers, critically influencing both the peak PSNR and the point of early stopping to avoid noise-fitting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.471, + 0.734 + ], + "angle": 0, + "content": "Importantly, this finding leads to empirical discovery on the roles of typical architectural components in DIP: assuming a standard hourglass network, i) simply scaling the depth and width can balance smoothing and preservation of details, due to the low-pass filtering effects of the upsampling operations inserted in-between the layers. As we observed, a wider and shallower network is better at preserving details and therefore suitable for fine-grained images. This suggests that the \"optimal\" architecture can vary across images, and image texture should be considered for more effective denoising. ii) Skip connections make a deep network perform similarly as a shallower one likely by reducing the \"effective upsampling rate\". This implies the possibility of discarding skip connections to simplify DIP architectural design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Based on these insights, we find it sufficient to restrict the design choices to only the network depth and width, reducing the problem to searching through only a handful of sub-networks and adapting them according to the level of image details. This can be done as a pre-processing step without costly searching or evaluation. We show that this simple strategy works with both the hourglass and decoder structures, and with proper setups, the estimated networks can denoise while preserving the details better than the larger networks with only \\(5\\% \\sim 40\\%\\) number of parameters. The resulting under-parameterization alleviates the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.88, + 0.108 + ], + "angle": 0, + "content": "need for early stopping. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.113, + 0.892, + 0.144 + ], + "angle": 0, + "content": "- We pinpoint that unlearnt upsampling is the main driving force behind the spectral bias of DIP." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.152, + 0.894, + 0.229 + ], + "angle": 0, + "content": "- Leveraging this insight, we empirically identify the influences of depth, width and skip connections, along with their correlations with image texture, allowing for quick, effective and more interpretable architectural design for every image without the laborious search." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.236, + 0.892, + 0.328 + ], + "angle": 0, + "content": "- We are the first to associate DIP architectural design with image texture. To promote future research, we build a Texture-DIP Dataset consisting of images from three popular datasets, reclassified into several predefined width choices – validated through experiments, according to textural complexity." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.335, + 0.892, + 0.427 + ], + "angle": 0, + "content": "- We show that with proper setups, a highly underparameterized subnetwork could match and even outperform larger counterparts, especially at a higher noise level. We conducted extensive synthetic and real-world noise removal experiments to validate our findings and approach." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.113, + 0.894, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.438, + 0.642, + 0.455 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.63 + ], + "angle": 0, + "content": "DIP Variants. Deep Decoder [15] is an underparameterized network proposed to avoid early stopping. However, our investigations show that underparameterization alone is not sufficient for denoising. For instance, ConvDecoder [10], a convolutional variant of Deep Decoder, contains more parameters but demonstrates less tendency to over-fit (Fig 1). In contrast to NAS-based methods[6, 17, 1], our strategy leverages the observed relationship between the architecture and the image to prevent the exhaustive search. Other variants such as DIP-RED [24] and DIP-TV [20] augment DIP with additional priors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.812 + ], + "angle": 0, + "content": "Early-stopping criterion. Cheng et al. [7] perform posterior inference in DIP so as to prevent the need for early stopping by adding Gaussian noise to the gradients. Similarly, Shi et al. [29] regularize the matrix norm of the network weights to alleviate performance decay. However, tuning the regularization granularity can be challenging (Suppl. B). Jo et al. [18] combine DIP with Stein's unbiased risk estimator (SURE) [32] for training without clean images, but SURE is limited to only a few known noise types [31, 23]. Wang et al. [35] propose to track the running variance of the outputs, but this also introduces new hyper-parameters that require non-trivial tuning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.822, + 0.741, + 0.839 + ], + "angle": 0, + "content": "3. Investigation and Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.642, + 0.863 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.87, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Deep Image Prior. A noisy image \\(\\mathbf{y} \\in \\mathbb{R}^N\\) can be modeled as: \\(\\mathbf{y} = \\mathbf{x} + \\mathbf{n}\\), where \\(\\mathbf{x} \\in \\mathbb{R}^N\\) is the clean counterpart to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12409" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.091, + 0.406, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.23, + 0.471, + 0.273 + ], + "angle": 0, + "content": "Figure 2: Spectral density of images of different levels of textural complexity, obtained by azimuthally integrating over the power spectrum of the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.471, + 0.362 + ], + "angle": 0, + "content": "be recovered and \\(\\mathbf{n}\\) is assumed to be \\(i.i.d\\). Gaussian Noise drawn from \\(\\mathcal{N}(0, \\sigma^2\\mathbf{I})\\) with \\(\\mathbf{I}\\) being the identity matrix. DIP parameterizes the clean image \\(\\mathbf{x}\\) via a network \\(\\mathbf{G}_{\\theta}\\) and is optimized to fit the noisy image \\(\\mathbf{y}\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.375, + 0.47, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} = \\underset {\\theta} {\\arg \\min } \\mathcal {L} (\\mathbf {y}; \\mathbf {G} _ {\\theta} (\\mathbf {z})), \\quad \\mathbf {x} ^ {*} = \\mathbf {G} _ {\\theta^ {*}} (\\mathbf {z}), \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.47, + 0.501 + ], + "angle": 0, + "content": "where \\( \\mathbf{z} \\) denotes the fixed white noise input tensor. Such parameterization allows lower-frequency contents to be fitted before the higher-frequency ones [3, 29], exhibiting high impedance to image noises or degradations. Early-stopping is often required to prevent excessive high frequencies from being fitted, such as noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.471, + 0.684 + ], + "angle": 0, + "content": "Image complexity. Ideally, DIP should denoise an image while preserving textural details. We define image complexity by its texture, which can be characterized by its power spectral density (PSD). The spectral power of a natural image typically follows an exponential decay from low frequencies to high frequencies [30]. High-frequency components correspond to fine features such as details, while low-frequency components correspond to coarse structures. Hence, fine-grained images contain more high frequencies than coarse-grained images, giving a flatter PSD curve in (Fig.2). Each image is scored based on its texture features, as described in Sec.3.5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.351, + 0.714 + ], + "angle": 0, + "content": "3.2. The importance of upsampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We first identified the core architecture components that affect the denoising performance of DIP. To this end, we analyzed a decoder-only architecture by removing the encoder from the typical encoder-decoder architecture [33], since a decoder is the minimum requirement for reconstructing the final image. Our base model for analysis is a 6-layered convolutional decoder (Conv-Decoder[10]), with \\(128 \\times 3\\) filters per layer except for the last regression layer, followed by batch normalization, ReLU activation function and upsampling. We further simplify the setup by replacing the spatial filters with pixel-wise \\(1 \\times 1\\) filters, constructing a non-convolutional variant dubbed MLP-Decoder." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.23, + 0.894, + 0.382 + ], + "angle": 0, + "content": "Figure 3: Influences of architecture components on image denoising. (a) Without convolutional layers, upsampling still enables denoising. Transposed convolutions lead to faster performance drop due to earlier noise-fitting than bilinear upsampling. Removing the upsampling decreases denoising capability, which cannot be compensated by simply reducing the number of layers, i.e., under-parameterization. (b) Convolutional layers alone exhibit certain denoising effects but necessitate early stopping. Increasing the number of convolutional layers achieves higher peak PSNR at the expense of earlier noise fitting (see red arrows). Better results achieved when combined with upsampling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.607 + ], + "angle": 0, + "content": "Results for MLP-Decoder, shown in Fig.3 (a), suggest that upsampling plays a vital role. Removal of upsampling results in significant performance degradation of denoising, which cannot be compensated by simply reducing the size of the network for under-parameterization. This holds true for other tasks such as super-resolution (Suppl.A). Fig.3 (b) shows that learnable spatial filters alone also enable denoising, in contrast to the pixel-wise filters, re-affirming the frequency bias of the convolutional layers [3]. However, the effects vanish as the network size increases and noise-fitting occurs sonner. Convolutional layers together with upsampling achieve better results as manifested in the higher peak PSNR and longer denoising effects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.759 + ], + "angle": 0, + "content": "Discussion. These results suggest that an appropriate upsampling operation is crucial for effective network image priors, and that under-parameterization alone is insufficient. Different upsampling operations induce varying extents of denoising effects: transposed convolutions [25] tend to fit noise faster than bilinear upsampling, necessitating early stopping. In the next section, we investigate the behaviors of these upsampling operations from a signal-processing perspective to gain insights into their influences on denoising performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.768, + 0.789 + ], + "angle": 0, + "content": "3.3. Spectral effects of upsampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The design choice for upsampling is typically standard: bilinear/nearest neighbor (NN) interpolation or transposed convolutions. These upsampling operations can all be decomposed into two steps: (i) zero-insertion and (ii) filtering. Given a target upsampling factor R, the low-resolution feature map is first interleaved with (R-1) rows/columns of zeros to increase its sampling rate, and then convolved with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12410" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.101, + 0.47, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.274, + 0.47, + 0.386 + ], + "angle": 0, + "content": "Figure 4: (Left) Frequency responses of the tested LPFs. Different LPFs result in upsamplers with different extents of smoothing. NN interpolation preserves most signals in the passband but also the high-frequency replica in the stopband; \\(\\mathcal{L}_{-100}\\) attenuates the signals most significantly (\\(\\sim 100dB\\)) and suppresses the high-frequency replica most. (Right) Denoising results on coarse- and fine-textured images from Set9. Top rows: Peak PSNR values. Bottom rows: PSNR values at the last training iteration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.469, + 0.47 + ], + "angle": 0, + "content": "a low-pass filter (LPF) to remove the alias high frequencies introduced by zero-insertion. The key difference between the upsampling operations lies in the nature of the filters. The filters for transposed convolutions are learnable, but are fixed for bilinear and nearest neighbor upsampling." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.47, + 0.469, + 0.518 + ], + "angle": 0, + "content": "To better illustrate this, we consider the case of a 1D signal \\(x(n), n = 0, \\dots, N - 1\\), and its discrete Fourier representation \\(X(k) = \\sum_{n=0}^{N-1} x(n) e^{\\frac{-i2\\pi}{N} k n}, k = 0, \\dots, N - 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.518, + 0.358, + 0.533 + ], + "angle": 0, + "content": "For an upsampling factor of 2, we have" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.543, + 0.479, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} X ^ {\\mathrm {u p}} (\\hat {k}) = \\sum_ {n = 0} ^ {2 N - 1} x ^ {\\mathrm {u p}} (n) e ^ {- \\frac {i 2 \\pi}{2 N} \\hat {k} n} (2) \\\\ = \\sum_ {n = 0} ^ {N - 1} x ^ {\\mathrm {u p}} (2 n) e ^ {\\frac {- i 2 \\pi}{2 N} (2 n) \\hat {k}} + \\sum_ {n = 0} ^ {N - 1} x ^ {\\mathrm {u p}} (2 n + 1) e ^ {\\frac {- i 2 \\pi}{2 N} (2 n + 1) \\hat {k}}, (3) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.656, + 0.469, + 0.704 + ], + "angle": 0, + "content": "where \\(\\hat{k} = 0,\\dots ,2N - 1,x^{\\mathrm{up}}(2n) = x(n),x^{\\mathrm{up}}(2n + 1) = 0\\) due to interleaved zero insertion. Hence, for \\(0\\leq \\hat{k} < N\\) \\(X^{\\mathrm{up}}(\\hat{k}) = X(k)\\)" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.705, + 0.466, + 0.72 + ], + "angle": 0, + "content": "For \\(\\hat{k} \\geq N\\), let \\(k' = \\hat{k} - N\\), \\(k' = 0, \\dots, N - 1\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.731, + 0.469, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} X ^ {\\mathrm {u p}} (\\hat {k}) = \\sum_ {n = 0} ^ {N - 1} x (n) e ^ {- \\frac {i 2 \\pi}{2 N} \\left(k ^ {\\prime} + N\\right) 2 n} (4) \\\\ = \\sum_ {n = 0} ^ {N - 1} x (n) e ^ {\\frac {- i 2 \\pi}{N} n k ^ {\\prime} - i 2 n \\pi} = X \\left(k ^ {\\prime}\\right), (5) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where Eq.5 exploits the periodicity of the complex exponential function. Thus, zero-insertion will preserve the original spectrum at \\([0,N]\\) (passband) and additionally create a mirrored copy at \\([N,2N - 1]\\) (stopband). The high-frequency replica beyond \\(N\\) should be suppressed by the" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Table 1: The iteration number [iter./5000] at the peak PSNR for different upsamplers. The upsamplers are designed to mainly differ in the stopband. The peak PSNR is reached more slowly when the attenuation is stronger (from left to right)." + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.156, + 0.851, + 0.211 + ], + "angle": 0, + "content": "
NNL14L15L16L17
Coarse-grained17831589168122052214
Fine-grained24242942389843614957
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.236, + 0.892, + 0.34 + ], + "angle": 0, + "content": "subsequent LPF to avoid image artifacts. According to duality and convolution theorem, convolving with NN or bilinear interpolation filter is equivalent to multiplication of \\( X^{up}(\\hat{k}) \\) with a Sinc or Sinc² function corresponding to low-pass filtering, while transposed convolutional filter may not necessarily be low-pass as it depends on the optimization objective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.893, + 0.538 + ], + "angle": 0, + "content": "We experimentally demonstrate that different upsampling operations bias the architecture towards different spectral properties. Specifically, we constructed four upsamplers by first interleaving the input with zeros and then convolving it with handcrafted LPFs (shortened as \\(\\mathcal{L}\\)): \\(\\mathcal{L}14\\), \\(\\mathcal{L}15\\), \\(\\mathcal{L}_{-60}\\) and \\(\\mathcal{L}_{-100}\\), with the subscript denoting the decayed dB. By construction, \\(\\mathcal{L}14\\) and \\(\\mathcal{L}15\\) are very close to NN in the passband (\\(< 0.03dB\\)) and only differ in the stopband. The frequency responses of the compared LPFs are detailed in Fig. 4. We applied the customized upsamplers on ConvDecoder and tested them on the fine- and coarse-textured images from Set9 respectively. Similar findings hold for the encoder-decoder architecture (Suppl. C)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.719 + ], + "angle": 0, + "content": "From Fig. 4, upsampling critically influences both the peak PSNR value and the timing for early stopping with respect to images of different textural complexities. Upsamplers (NN, \\(\\mathcal{L}14\\), \\(\\mathcal{L}15\\)) with less attenuation on the high-frequency replica are beneficial for generating fine-grained images, but they tend to cause noise-fitting especially for coarser-grained images. Also, they are generally the fastest to reach the peak PSNR (Table 1). This explains why the transposed convolution requires early-stopping, as the filters may not learn to attenuate the introduced high frequencies effectively. Similar issues with the learned upsampling are also prevalent in generative models [28, 5, 11, 12]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.856 + ], + "angle": 0, + "content": "On the other hand, bilinear and \\(\\mathcal{L}_{-60}\\) exhibit a greater amount of attenuation, more strongly biasing the network against high frequencies and leading to longer-lasting denoising effects. They turn out to work sufficiently well for both kinds of images, especially on the coarser-grained ones which are typically the majority in the dataset. \\(\\mathrm{LPF}_{-100}\\) tends to over-smooth the output and performs the worst as it attenuates both the passband and stopband signals substantially, though not requiring early stopping." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Discussion. These results lead us to conclude that the upsamplers with fixed LPFs are key to the denoising effects of DIP for their tendency towards smooth images (i.e., re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12411" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.473, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.348, + 0.47, + 0.432 + ], + "angle": 0, + "content": "Figure 5: Influences of width, depth, and skip connections, assuming upsampling inserted in-between the layers. (a) Tendency to over-fit coarse-grained images increases with width. SSIM scores averaged across the depths. (b) Tendency to over-smooth fine-grained images increases with depth, but alleviated by skip connections. SSIM scores are averaged across the widths." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.536 + ], + "angle": 0, + "content": "duced high-frequency contents), which aligns well with the spectral statistics of natural images (Fig.2). Probably due to a good balance between the denoising performance and persistency, bilinear upsampling has been widely adopted in DIP models for various applications [33, 15, 13, 16]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.46, + 0.56 + ], + "angle": 0, + "content": "3.4. Interactions with other architecture elements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.469, + 0.614 + ], + "angle": 0, + "content": "After establishing the significance of upsampling, we will now consider how it may interact with other common architectural elements in affecting the ultimate output." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.469, + 0.75 + ], + "angle": 0, + "content": "Convolutions + non-linearity. Ideal upsampling does not modify the signal representations but only expands the spectrum for the subsequent layers to add new frequency contents. Convolution followed by nonlinearity such as ReLU, is the only operation capable of introducing arbitrarily high frequencies [19]. Increasing the number of layers (depth) or channels (width) enhances the capability of generating new high frequencies, as theoretically and empirically proved in [26]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Intuitively, using an excessive number of layers or channels can accelerate the learning of both details and noise. However, the effects can be attenuated by the fixed upsampling operations between the layers. As shown in Fig.5, when using fewer upsampling operations (i.e., a shallower network), increasing the width of the network increases the tendency to cause overfitting on simpler images while benefiting the more complex images. Increasing the number of upsampling operations (i.e., a deeper network) can alleviate over-fitting with stronger attenuation but results in" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.091, + 0.855, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.263, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Figure 6: Position of upsampling. Placing the upsampling close to the input (or encoder) can easily cause over-fitting, regardless of the scaling factor. When close to the output (end of the decoder), upsampling with large scaling factors (e.g., \\(32 \\times\\)) causes over-smoothing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.453 + ], + "angle": 0, + "content": "blurry outputs for fine-grained images. Increasing only the upsampling factors without adding more layers can make the output even more blurry (Fig.6). In other words, the final output is determined by the balance between the generation of high frequencies by the layers and signal attenuation caused by the upsampling operations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.454, + 0.893, + 0.651 + ], + "angle": 0, + "content": "Skip connections between the encoder and the decoder often complicate the design space [6]. While they are not directly responsible for denoising, they may lower the effective upsampling rate, making deep networks perform similarly to shallower ones, i.e., resulting in lesser suppression of the high-frequency replica. This finding is validated on a large-scale experiment comprising 7424 architectures (Suppl. D). As exemplified in Fig. 5 (b), skip connections notably ameliorate the over-smoothing issue induced by the same deep network. Overall, skip connections exert a more pronounced influence on the deeper networks compared to the shallower ones, as evident in the smaller deviation observed when depth \\(\\leq 3\\) in Fig.5 (b)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.664, + 0.871, + 0.681 + ], + "angle": 0, + "content": "3.5. Practical application to architectural design" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Based on the above findings and analysis, we argue that it is possible to estimate an effective architecture for each image without extensive search. Assuming every decoder layer is followed by a \\(2 \\times\\) bilinear upsampling layer, our strategies are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Depth estimation. We have shown that increased depth tends to over-smooth the output, affecting fine-grained images much more than on coarse-grained images. For fine-grained images, we have three options: a) add more skip connections to a deep network; b) simply use a shallower one; c) keep all the layers but reduce down-/up-sampling layers. We recommend c for decoder-only architectures since they are already under-parameterized; for hourglass network this can easily lead to over-fitting (Fig.6). We find" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12412" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 2: Quantitative results on Gaussian noise. \\( \\sigma \\) denotes the noise level. All methods were trained with a fixed iteration number (3000) throughout the experiments. The highest score is in bold, and the second highest is underlined." + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.129, + 0.833, + 0.345 + ], + "angle": 0, + "content": "
DatasetsDIP [33]Deep Decoder [15]ConvDecoder [10]NAS-DIP [6]ISNAS-DIP [1]Ours
σ = 25
Set9 [9]PSNR30.1028.4528.5126.3729.1130.26
SSIM0.8930.8480.8540.7530.8620.900
Set12 [37]PSNR26.9725.9825.7820.8624.1028.14
SSIM0.8120.7890.7860.5340.7450.884
CBSD68 [27]PSNR28.9325.5025.1923.8024.5128.57
SSIM0.8920.8090.7930.6930.7450.888
σ = 50
Set9 [9]PSNR25.0425.2225.0121.0723.9126.13
SSIM0.7610.7640.7690.5930.6980.833
Set12 [37]PSNR22.1520.4422.7218.9219.2024.59
SSIM0.6230.6870.7060.4760.5370.805
CBSD68 [27]PSNR23.7423.5224.0617.9219.9324.17
SSIM0.7460.7250.7670.3230.5730.774
# Params (Millions)2.3M0.1M0.89M4.4MVaried0.05M~0.92M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.367, + 0.47, + 0.396 + ], + "angle": 0, + "content": "Table 3: Comparisons of desired properties. Restoration time is computed on an image of size \\(512 \\times 512\\) with 3000 iterations." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.407, + 0.436, + 0.473 + ], + "angle": 0, + "content": "
NAS [6]ISNAS [1]Ours
Image-SpecificX✓✓
Architecture Search3 days5 mins-
Per-image Restoration~23 mins~7 hrs~6 mins
Early Stopping Required?YesYesNo
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.594 + ], + "angle": 0, + "content": "\\(\\mathbf{b}\\) is generally better than \\(\\mathbf{a}\\) in trading off between good performance and the need for early stopping, especially under higher-level noise, as shown in our results section. This also holds for coarse-grained images as they are not sensitive to depth. More specifically, we find a 2-level hourglass network sufficient for both types of images (Suppl. D)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.471, + 0.793 + ], + "angle": 0, + "content": "Width estimation. Width is crucial for learning sufficient details while avoiding over-fitting especially to a shallow network (less signal attenuation). We find that the width needs to be set according to textural complexity of the image: a finer-grained image requires more channels per layer and vice versa. To further validate this, we treated width estimation as a classification problem, and trained three SVMs [8] with texture features as the inputs to classify the images into three width choices \\{32, 64, 128\\}. Note that these widths were empirically chosen for the datasets used in this work and are by no means optimal for all cases, but tuning for other images should be straightforward. The classification results and analysis are in Sec.4.4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Image scoring. To more robustly classify image texture, we extracted both spatial and frequency features and trained a Decision Tree [22] for feature selection. We used the classic Gray Level Co-occurrence Matrix (GLCM) [14] for deriving the spatial texture features from each image. The most useful features turn out to be the following: correlations measured at \\(0^{\\circ}\\), homogeneity at \\(45^{\\circ}\\), and contrast at" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.367, + 0.892, + 0.396 + ], + "angle": 0, + "content": "Table 4: Quantitative evaluation on Poisson noise. Deep Decoder shortened as DD and ConvDecoder as CD." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.407, + 0.888, + 0.503 + ], + "angle": 0, + "content": "
Noise scaleDIPDDCDNASISNASOurs
ζ = 0.01PSNR31.5829.8129.5129.7030.1230.95
SSIM0.9150.8800.8750.8640.8750.910
ζ = 0.1PSNR22.6623.9124.9415.7217.6824.99
SSIM0.6400.7180.7760.4170.4960.789
ζ = 0.2PSNR20.3621.1321.8712.9914.6422.55
SSIM0.5630.6090.6720.3110.3830.774
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.551, + 0.892, + 0.662 + ], + "angle": 0, + "content": "\\(0^{\\circ}\\). Frequency features are captured using a 1D PSD vector obtained by first converting the 2D PSD from Cartesian coordinates \\(S(k,l)\\) to polar coordinates, and then azimuthally averaging over \\(\\theta\\), defined as \\(\\hat{S}(r) = \\frac{1}{2\\pi} \\int_{0}^{2\\pi} S(r,\\theta)d\\theta\\) with \\(r = \\sqrt{k^2 + l^2}\\) and \\(\\theta = \\mathrm{atan2}(k,l)\\), which represents the mean magnitude of the frequencies with respect to the radial distance \\(r\\). Refer to Suppl. E for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.691, + 0.634, + 0.709 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.721, + 0.719, + 0.739 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We conducted the experiments on three popular datasets and a real-world noisy dataset: Set9 [9] consisting of 9 colored images, Set12 [37] consisting of 12 grey-scaled images, CBSD68 [27] consisting of 68 colored images, and PolyU [36] consisting of 100 real noisy and clean image pairs. We first report our results with 2-level hourglass architecture with the same components as in DIP [33] when comparing with the existing methods, and then extend our strategy to ConvDecoder [10], a decoder-only architecture. All models were trained for 3000 iterations." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12413" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.092, + 0.293, + 0.102 + ], + "angle": 0, + "content": "Noisy (20.27/0.583)" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.104, + 0.328, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.22, + 0.293, + 0.229 + ], + "angle": 0, + "content": "NAS (26.72/0.836)" + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.093, + 0.465, + 0.102 + ], + "angle": 0, + "content": "ConvDecoder (25.19/0.781)" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.104, + 0.482, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.22, + 0.445, + 0.229 + ], + "angle": 0, + "content": "DIP (27.86/0.874)" + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.093, + 0.619, + 0.102 + ], + "angle": 0, + "content": "Deep Decoder (25.36/0.789)" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.104, + 0.635, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.22, + 0.598, + 0.229 + ], + "angle": 0, + "content": "Ours (28.18/0.882)" + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.093, + 0.765, + 0.102 + ], + "angle": 0, + "content": "ISNAS-DIP (25.43/0.779)" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.104, + 0.788, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.22, + 0.773, + 0.229 + ], + "angle": 0, + "content": "Ground Truth (PSNR/SSIM)" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.231, + 0.329, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.231, + 0.481, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.231, + 0.635, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.231, + 0.788, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.353, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Figure 7: Denoising results on a fine-grained image (\"kodim01\" from Set9) with Gaussian noise (\\(\\sigma = 25\\)). Our estimated architecture for this image is a two-level hourglass network with one skip connection and 128 channels, which is much smaller than others." + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.409, + 0.192, + 0.418 + ], + "angle": 0, + "content": "Noisy (36.42/0.967)" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.42, + 0.211, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.409, + 0.323, + 0.418 + ], + "angle": 0, + "content": "ConvDecoder (35.02/0.951)" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.42, + 0.328, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.409, + 0.441, + 0.418 + ], + "angle": 0, + "content": "Deep Decoder (35.37/0.955)" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.42, + 0.444, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.51, + 0.207, + 0.519 + ], + "angle": 0, + "content": "DIP w/o. skip (35.14/0.951)" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.52, + 0.211, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.51, + 0.317, + 0.519 + ], + "angle": 0, + "content": "DIP w. skip (36.38/0.964)" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.52, + 0.328, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.511, + 0.434, + 0.519 + ], + "angle": 0, + "content": "ISNAS-DIP (36.86/0.968)" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.52, + 0.444, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.611, + 0.198, + 0.62 + ], + "angle": 0, + "content": "NAS-DIP (37.18/0.972)" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.621, + 0.211, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.612, + 0.306, + 0.62 + ], + "angle": 0, + "content": "Ours (37.99/0.979)" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.621, + 0.328, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.612, + 0.442, + 0.62 + ], + "angle": 0, + "content": "Ground Truth (PSNR/SSIM)" + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.621, + 0.444, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.72, + 0.445, + 0.733 + ], + "angle": 0, + "content": "Figure 8: Denoising results on a real-world noisy image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.768, + 0.212, + 0.782 + ], + "angle": 0, + "content": "4.2. DIP variants" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Gaussian Noise. Compared with the base DIP network, the properly-designed under-parameterized networks estimated with our strategy perform on par at a mild noise level while excel at a higher noise level (Table 2). This cannot be solely explained by under-parameterization since Deep Decoder and ConvDecoder contain similar or even fewer parameters while are unable to achieve similar results. Fig." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.409, + 0.892, + 0.574 + ], + "angle": 0, + "content": "7 shows that a shallow and wide network can preserve the details better than many deeper ones. Note that DIP is a 5-level hourglass network with full skip connections, which by our standard can also well preserve the details. In fact, it is a very strong baseline under mild noise. NAS-DIP and ISNAS-DIP suffer from various degrees of over-fitting on different datasets. This also suggests that the point of optimal stopping varies across images. Besides, they are time-intensive in either searching or evaluation (Table 3). Similar conclusions hold for Poisson noise, which was tested on Set9 [9] as shown in Table 4." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.588, + 0.891, + 0.616 + ], + "angle": 0, + "content": "Table 5: Quantitative evaluation on PolyU, a real-world noisy dataset. Deep Decoder and ConvDecoder shortened as DD, CD." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.628, + 0.876, + 0.687 + ], + "angle": 0, + "content": "
DIPDDCDNASISNASOurs
PSNR38.1537.2237.0037.8337.7838.05
SSIM0.9820.9780.9760.9820.9770.984
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.715, + 0.892, + 0.836 + ], + "angle": 0, + "content": "Real-World Noise. We additionally evaluated all methods on the PolyU dataset [36]. We applied again the two-level hourglass architectures with variable width estimated for the images. Table 5 summarizes the numerical results. Fig.9 shows that our method tends to preserve more details, though this may not be reflected by the metrics. Simply removing the skip connections from DIP causing blurring, similar to the decoder-only architectures." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.747, + 0.863 + ], + "angle": 0, + "content": "4.3. Decoder-only Architectures" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We applied our strategy to ConvDecoder and tested it on all three datasets. Here we keep all 5 layers but remove" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12414" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.09, + 0.393, + 0.104 + ], + "angle": 0, + "content": "Table 6: Application to ConvDecoder." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.235 + ], + "angle": 0, + "content": "
Datasetsσ = 25σ = 50
BeforeAfterBeforeAfter
Set9PSNR28.5128.7425.0125.11
SSIM0.8540.8730.7690.784
Set12PSNR25.7926.9822.7223.01
SSIM0.7860.8540.7060.742
CBSD68PSNR25.1928.2924.3624.12
SSIM0.7930.8770.7670.768
# Params (Millions)0.89M0.06M~0.89M0.89M0.06M~0.89M
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.251, + 0.47, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.386, + 0.47, + 0.428 + ], + "angle": 0, + "content": "Figure 9: Qualitative improvement on ConvDecoder simply by removing three upsampling layers in this case while preserving all the convolutional layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.47, + 0.531 + ], + "angle": 0, + "content": "a certain number of upsampling layers and scale the width accordingly for the images. These simple changes effectively alleviate the over-smoothing issue that often brought by ConvDecoder as shown in Fig.8 and improve the quantitative results as shown in Table 6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.544, + 0.387, + 0.56 + ], + "angle": 0, + "content": "4.4. More Analysis on Depth and Width" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.825 + ], + "angle": 0, + "content": "In practice, we find it more efficient to first determine the depth, and then the width. This is because when the network is deep enough (strong attenuation due to upsampling), width becomes less influential, as evident in Fig.10 (b) and the small deviation in Fig.5 (b). However, to relax the need for early stopping at a higher noise level, one may prefer a shallower one, where width matters more (Fig.10 (a)). In this regard, we use the texture features of the images to predict the desired width. This makes intuitive sense as noise-fitting is associated with the amount of high frequency contents in the image, which is manifested in its texture. This is corroborated by the 0.86 Micro-average AUC score shown in Fig.11 (a). The \"optimal\" width labels we used to train the classifiers were obtained by experimenting with all three width choices on a 2-level hourglass architecture. These width labels are also applicable to decoder architectures as demonstrated by our experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Although the width choices seem limited, we found the images robust to the choice of width to some extent, and simple averaging also works well for ambiguous cases (Appendix D). In fact, some images have multiple width labels. We included these cases in the released Texture-DIP dataset." + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.093, + 0.695, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.093, + 0.895, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.243, + 0.892, + 0.271 + ], + "angle": 0, + "content": "Figure 10: (a) Width critically influences a shallower network, while (b) it rarely has any impact on a sufficiently deep network." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.288, + 0.755, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.287, + 0.866, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.45, + 0.894, + 0.505 + ], + "angle": 0, + "content": "Figure 11: (a) ROC curves with AUC scores for width classification on images from Set9, Set12 and CBSD68. 5-fold cross-validation is performed and repeated 10 times. (b) Overview of our Texture-DIP Dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.534, + 0.746, + 0.549 + ], + "angle": 0, + "content": "4.5. What about Transformers?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.785 + ], + "angle": 0, + "content": "Transformers [34] have become integral parts of modern deep neural network architectures. We experimented with Swin-UNet [2], an encoder-decoder pure transformer consisting of Swin Transformer blocks [21] and skip connections. Note that Swin-UNet relies on patch merging/expanding layers for \\(2 \\times\\) down-/upsampling without involving unlearnt upsampling. We replaced all the patch expanding layers with \\(2 \\times\\) bilinear upsampling operations and compared the performance of the modified network with the original. We did not replace the patch merging layers with the corresponding downsampling layers or strided convolutions because we did not observe any significant difference. The results presented in Fig. 12 are consistent with our findings on CNNs. For more implementation details, please refer to our code." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.77, + 0.815 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We presented simple but effective solutions to the challenging DIP architectural design in the context of image denoising. Leveraging the spectral effects of upsampling and its relationships with other architectural components, we show that simple architectural changes yield highly-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12415" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.092, + 0.273, + 0.115 + ], + "angle": 0, + "content": "Noisy (20.38/0.465)" + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.117, + 0.311, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.091, + 0.438, + 0.114 + ], + "angle": 0, + "content": "Swin-UNet (21.42/0.497)" + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.116, + 0.48, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.091, + 0.64, + 0.114 + ], + "angle": 0, + "content": "Swin-UNet w. Bilinear Up (30.61/0.928)" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.116, + 0.644, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.092, + 0.773, + 0.115 + ], + "angle": 0, + "content": "Ground Truth (PSNR/SSIM)" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.116, + 0.808, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.285, + 0.813, + 0.3 + ], + "angle": 0, + "content": "Figure 12: Visual comparisons on the transformer-based Swin-UNet [2] with and without bilinear upsampling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.327, + 0.47, + 0.448 + ], + "angle": 0, + "content": "effective under-parameterized networks that could surpass the larger counterparts and does not critically rely on early-stopping. Although this work focuses on denoising, insights learned about upsampling can be employed in the future to understand the relationship between architectural characteristics and other image restoration tasks. We hope our study could encourage efficient architectural design for DIP and image synthesis in general." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.46, + 0.245, + 0.476 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.47, + 0.529 + ], + "angle": 0, + "content": "This work was supported in part by the United States National Institutes of Health (NIH) under Grant CA266702 and Grant EB008374." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.542, + 0.174, + 0.557 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.567, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[1] Metin Ersin Arican, Ozgur Kara, Gustav Bredell, and Ender Konukoglu. Isnas-dip: Image-specific neural architecture search for deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1960-1968, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.638, + 0.47, + 0.706 + ], + "angle": 0, + "content": "[2] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.707, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[3] Prithvijit Chakrabarty and Subhransu Maji. The spectral bias of the deep image prior. arXiv preprint arXiv:1912.08905, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.749, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[4] Stanley H Chan, Xiran Wang, and Omar A Elgendy. Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging, 3(1):84-98, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.805, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[5] Keshigeyan Chandrasegaran, Ngoc-Trung Tran, and Ngai-Man Cheung. A closer look at fourier spectrum discrepancies for cnn-generated images detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7200-7209, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[6] Yun-Chun Chen, Chen Gao, Esther Robb, and Jia-Bin Huang. Nas-dip: Learning deep image prior with neural" + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.567, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.328, + 0.892, + 0.355 + ], + "angle": 0, + "content": "architecture search. In European Conference on Computer Vision, pages 442-459. Springer, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.357, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[7] Zezhou Cheng, Matheus Gadelha, Subhransu Maji, and Daniel Sheldon. A bayesian perspective on the deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5451, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.426, + 0.891, + 0.453 + ], + "angle": 0, + "content": "[8] Corinna Cortes and Vladimir Vapnik. Support-vector networks. Machine learning, 20:273-297, 1995. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.455, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[9] Kostadin Dabov, Alessandro Foi, and Karen Egiazarian. Video denoising by sparse 3d transform-domain collaborative filtering. In 2007 15th European Signal Processing Conference, pages 145-149. IEEE, 2007. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[10] Mohammad Zalbagi Darestani and Reinhard Heckel. Accelerated mri with un-trained neural networks. IEEE Transactions on Computational Imaging, 7:724-733, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[11] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7890–7899, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[12] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[13] Yosef Gandelsman, Assaf Shocher, and Michal Irani. \"double-dip\": unsupervised image decomposition via coupled deep-image-priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11026-11035, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[14] Robert M Haralick, Karthikeyan Shanmugam, and Its' Hak Dinstein. Textural features for image classification. IEEE Transactions on systems, man, and cybernetics, (6):610-621, 1973. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[15] Reinhard Heckel and Paul Hand. Deep decoder: Concise image representations from untrained non-convolutional networks. arXiv preprint arXiv:1810.03982, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[16] Reinhard Heckel and Mahdi Soltanolkotabi. Denoising and regularization via exploiting the structural bias of convolu" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.328, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12416" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "tional generators. arXiv preprint arXiv:1910.14634, 2019. 1,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[17] Kary Ho, Andrew Gilbert, Hailin Jin, and John Collomosse. Neural architecture search for deep image prior. Computers & graphics, 98:188-196, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[18] Yeonsik Jo, Se Young Chun, and Jonghyun Choi. Rethinking deep image prior for denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5087-5096, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[19] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[20] Jiaming Liu, Yu Sun, Xiaojian Xu, and Ulugbek S Kamilov. Image restoration using total variation regularized deep image prior. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7715-7719. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[21] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[22] Wei-Yin Loh. Classification and regression trees. Wiley interdisciplinary reviews: data mining and knowledge discovery, 1(1):14-23, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.469, + 0.497 + ], + "angle": 0, + "content": "[23] Florian Luisier, Thierry Blu, and Michael Unser. Image denoising in mixed poisson-gaussian noise. IEEE Transactions on image processing, 20(3):696-708, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[24] Gary Mataev, Peyman Milanfar, and Michael Elad. Deepred: Deep image prior powered by red. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.593 + ], + "angle": 0, + "content": "[25] Augustus Odena, Vincent Dumoulin, and Chris Olah. Deconvolution and checkerboard artifacts. Distill, 1(10):e3, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[26] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[27] Stefan Roth and Michael J Black. Fields of experts. International Journal of Computer Vision, 82(2):205-229, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.748 + ], + "angle": 0, + "content": "[28] Katja Schwarz, Yiyi Liao, and Andreas Geiger. On the frequency bias of generative models. Advances in Neural Information Processing Systems, 34:18126-18136, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[29] Zenglin Shi, Pascal Mettes, Subhransu Maji, and Cees GM Snoek. On measuring and controlling the spectral bias of the deep image prior. International Journal of Computer Vision, 130(4):885-908, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[30] Eero P Simoncelli and Bruno A Olshausen. Natural image statistics and neural representation. Annual review of neuroscience, 24(1):1193-1216, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[31] Shakarim Soltanayev and Se Young Chun. Training deep learning based denoisers without ground truth data. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[32] Charles M Stein. Estimation of the mean of a multivariate normal distribution. The annals of Statistics, pages 1135-1151, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[33] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446-9454, 2018. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[35] Hengkang Wang, Taihui Li, Zhong Zhuang, Tiancong Chen, Hengyue Liang, and Ju Sun. Early stopping for deep image prior. arXiv preprint arXiv:2112.06074, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[36] Jun Xu, Hui Li, Zhetong Liang, David Zhang, and Lei Zhang. Real-world noisy image denoising: A new benchmark. arXiv preprint arXiv:1804.02603, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[37] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[38] Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3929-3938, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12417" + } + ] +] \ No newline at end of file diff --git a/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/6b883d6c-8e38-4d25-996f-79941abad307_origin.pdf b/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/6b883d6c-8e38-4d25-996f-79941abad307_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..576150e2b27af8c079c17f8d546781fa2acb1c03 --- /dev/null +++ b/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/6b883d6c-8e38-4d25-996f-79941abad307_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e219638ff815acb43a4ee965c3cc53a33d37498197e6434f693b21eb5aec51d2 +size 10294451 diff --git a/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/full.md b/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..79561688774dbb2a81aa90d9f5b8419d23316481 --- /dev/null +++ b/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/full.md @@ -0,0 +1,345 @@ +# The Devil is in the Upsampling: Architectural Decisions Made Simpler for Denoising with Deep Image Prior + +Yilin Liu $^{1*}$ , Jiang Li $^{1*}$ , Yunkui Pang $^{1*}$ , Dong Nie $^{1}$ , and Pew-Thian Yap $^{1}$ + +1University of North Carolina at Chapel Hill + +{yilinliu, yunkuipa, dongnie} @cs.unc.edu, jiang.li@unc.edu, ptyap@med.unc.edu + +# Abstract + +Deep Image Prior (DIP) shows that some network architectures inherently tend towards generating smooth images while resisting noise, a phenomenon known as spectral bias. Image denoising is a natural application of this property. Although denoising with DIP mitigates the need for large training sets, two often intertwined practical challenges need to be overcome: architectural design and noise fitting. Existing methods either handcraft or search for suitable architectures from a vast design space, due to the limited understanding of how architectural choices affect the denoising outcome. In this study, we demonstrate from a frequency perspective that unlearnt upsampling is the main driving force behind the denoising phenomenon with DIP. This finding leads to straightforward strategies for identifying a suitable architecture for every image without laborious search. Extensive experiments show that the estimated architectures achieve superior denoising results than existing methods with up to $95\%$ fewer parameters. Thanks to this under-parameterization, the resulting architectures are less prone to noise-fitting1. + +# 1. Introduction + +Image denoising is useful on its own and can be a plug-in module for many other image restoration tasks [37, 38, 4]. Deep neural networks have become the tool of choice for image denoising owing to their ability to learn natural image priors from large-scale datasets. Yet, Deep Image Prior (DIP) [33] requires only a single degraded image for image resotration. Remarkably, DIP shows that a randomly initialized convolutional neural network (CNN) can regularize image restoration through its architecture and early-stopping optimization. This is inspired by the phenomenon that some network architectures act inherently as image priors, favoring generating smooth, natural images and re + +![](images/bd7b06dc5cbb5a34ad44df7ce728b9d897eaf43e119c39ed18a5c170b4952f40.jpg) +Noisy 20.34/0.648 + +![](images/ca844309f68f2a77e921b730de11049ec9e94f261727dbd29fa2054654496d2e.jpg) +DIP NAS-DIP 8/0.821 21.92/0.680 +ISNAS-DIP Ours Ground Truth 21.04/0.600 25.22/0.841 PSNR $\uparrow$ /SSIM +Deep Decoder Ours Ground Truth 27.33/0.794 32.01/0.924 PSNR ↑ / SSIM ↑ + +![](images/4e322c5a3024523e2a6b8746a745f90b99b6313a2b1429156322313ef3046831.jpg) +DIP 28.39/0.824 +ConvDecoder 28.91/0.853 +Figure 1: Top: Performance (SSIM↑) under different levels of Gaussian noises. Bottom: Denoising of a fine-grained $(1^{st}\mathrm{row})$ and a coarse-grained image $(2^{nd}\mathrm{row})$ . Most existing methods, including the recent image-specific ISNAS-DIP[1], struggle to perform well simultaneously in both cases. Our simple strategies are flexible in image-specific architectural adaptation without requiring a search. Moreover, the results of the lightweight ConvDecoder[10] and Deep Decoder[15] suggest that without proper model setups, under-parameterization itself can neither ensure good denoising performance nor remove the need for early-stopping. + +sisting noises or degradations. However, image denoising with DIP is greatly influenced by architectural design [3, 1, 33, 15, 17], and the associated tendency to fit the original noisy image, i.e., over-fitting [15, 16]. + +Architectural design for DIP remains an open problem. One prevailing view is that model under-parameterization limits noise over-fitting and thus mitigates the need for early stopping [15]. However, our experiments reveal that multiple model architectures can exist under a similar parameter budget, where inappropriate model setups can still lead to + +noise fitting and over-smoothing (Fig.1, Fig.3). Another line of work automates the architecture identification using Neural Architecture Search (NAS)[6, 17, 1]. Without prior knowledge on suitable architectures, extensive search incurs substantial computational costs, prohibiting image-wise NAS for optimal restoration [1]. Arican et al. [1] narrow the search space using training-free metrics, but the need for candidate comparison dramatically prolongs the restoration time ( $\sim$ 7 hours/image). Moreover, NAS-based models, along with many DIP models, are typically heavily parameterized and prone to over-fitting, as corroborated in our experiments. Thus, their performance largely depends on the timing of early stopping, which is typically image-specific and hard to pinpoint without access to ground truth. + +Directly identifying an effective under-parameterized architecture for each image is challenging in light of the vast number of different architectures and the absence of ground truth for explicit supervision. To simplify the architectural decisions for DIP, we rethink the architectural influences on its performance in the context of image denoising. + +We start by noting that denoising performance is attributable to only a few components, primarily the unlearnt upsampling operations. Our frequency analysis reveals that the fixed upsampling operations tend to bias the architecture towards low-frequency contents more strongly than linear or convolutional layers, critically influencing both the peak PSNR and the point of early stopping to avoid noise-fitting. + +Importantly, this finding leads to empirical discovery on the roles of typical architectural components in DIP: assuming a standard hourglass network, i) simply scaling the depth and width can balance smoothing and preservation of details, due to the low-pass filtering effects of the upsampling operations inserted in-between the layers. As we observed, a wider and shallower network is better at preserving details and therefore suitable for fine-grained images. This suggests that the "optimal" architecture can vary across images, and image texture should be considered for more effective denoising. ii) Skip connections make a deep network perform similarly as a shallower one likely by reducing the "effective upsampling rate". This implies the possibility of discarding skip connections to simplify DIP architectural design. + +Based on these insights, we find it sufficient to restrict the design choices to only the network depth and width, reducing the problem to searching through only a handful of sub-networks and adapting them according to the level of image details. This can be done as a pre-processing step without costly searching or evaluation. We show that this simple strategy works with both the hourglass and decoder structures, and with proper setups, the estimated networks can denoise while preserving the details better than the larger networks with only $5\% \sim 40\%$ number of parameters. The resulting under-parameterization alleviates the + +need for early stopping. Our contributions are as follows: + +- We pinpoint that unlearnt upsampling is the main driving force behind the spectral bias of DIP. +- Leveraging this insight, we empirically identify the influences of depth, width and skip connections, along with their correlations with image texture, allowing for quick, effective and more interpretable architectural design for every image without the laborious search. +- We are the first to associate DIP architectural design with image texture. To promote future research, we build a Texture-DIP Dataset consisting of images from three popular datasets, reclassified into several predefined width choices – validated through experiments, according to textural complexity. +- We show that with proper setups, a highly underparameterized subnetwork could match and even outperform larger counterparts, especially at a higher noise level. We conducted extensive synthetic and real-world noise removal experiments to validate our findings and approach. + +# 2. Related Work + +DIP Variants. Deep Decoder [15] is an underparameterized network proposed to avoid early stopping. However, our investigations show that underparameterization alone is not sufficient for denoising. For instance, ConvDecoder [10], a convolutional variant of Deep Decoder, contains more parameters but demonstrates less tendency to over-fit (Fig 1). In contrast to NAS-based methods[6, 17, 1], our strategy leverages the observed relationship between the architecture and the image to prevent the exhaustive search. Other variants such as DIP-RED [24] and DIP-TV [20] augment DIP with additional priors. + +Early-stopping criterion. Cheng et al. [7] perform posterior inference in DIP so as to prevent the need for early stopping by adding Gaussian noise to the gradients. Similarly, Shi et al. [29] regularize the matrix norm of the network weights to alleviate performance decay. However, tuning the regularization granularity can be challenging (Suppl. B). Jo et al. [18] combine DIP with Stein's unbiased risk estimator (SURE) [32] for training without clean images, but SURE is limited to only a few known noise types [31, 23]. Wang et al. [35] propose to track the running variance of the outputs, but this also introduces new hyper-parameters that require non-trivial tuning. + +# 3. Investigation and Method + +# 3.1. Preliminaries + +Deep Image Prior. A noisy image $\mathbf{y} \in \mathbb{R}^N$ can be modeled as: $\mathbf{y} = \mathbf{x} + \mathbf{n}$ , where $\mathbf{x} \in \mathbb{R}^N$ is the clean counterpart to + +![](images/336a54dcb45b4de8b5f091f3c4d7770b40ff40f20124a31f58824faa30a4195e.jpg) +Figure 2: Spectral density of images of different levels of textural complexity, obtained by azimuthally integrating over the power spectrum of the image. + +be recovered and $\mathbf{n}$ is assumed to be $i.i.d$ . Gaussian Noise drawn from $\mathcal{N}(0, \sigma^2\mathbf{I})$ with $\mathbf{I}$ being the identity matrix. DIP parameterizes the clean image $\mathbf{x}$ via a network $\mathbf{G}_{\theta}$ and is optimized to fit the noisy image $\mathbf{y}$ , formulated as: + +$$ +\theta^ {*} = \underset {\theta} {\arg \min } \mathcal {L} (\mathbf {y}; \mathbf {G} _ {\theta} (\mathbf {z})), \quad \mathbf {x} ^ {*} = \mathbf {G} _ {\theta^ {*}} (\mathbf {z}), \quad (1) +$$ + +where $\mathbf{z}$ denotes the fixed white noise input tensor. Such parameterization allows lower-frequency contents to be fitted before the higher-frequency ones [3, 29], exhibiting high impedance to image noises or degradations. Early-stopping is often required to prevent excessive high frequencies from being fitted, such as noise. + +Image complexity. Ideally, DIP should denoise an image while preserving textural details. We define image complexity by its texture, which can be characterized by its power spectral density (PSD). The spectral power of a natural image typically follows an exponential decay from low frequencies to high frequencies [30]. High-frequency components correspond to fine features such as details, while low-frequency components correspond to coarse structures. Hence, fine-grained images contain more high frequencies than coarse-grained images, giving a flatter PSD curve in (Fig.2). Each image is scored based on its texture features, as described in Sec.3.5. + +# 3.2. The importance of upsampling + +We first identified the core architecture components that affect the denoising performance of DIP. To this end, we analyzed a decoder-only architecture by removing the encoder from the typical encoder-decoder architecture [33], since a decoder is the minimum requirement for reconstructing the final image. Our base model for analysis is a 6-layered convolutional decoder (Conv-Decoder[10]), with $128 \times 3$ filters per layer except for the last regression layer, followed by batch normalization, ReLU activation function and upsampling. We further simplify the setup by replacing the spatial filters with pixel-wise $1 \times 1$ filters, constructing a non-convolutional variant dubbed MLP-Decoder. + +![](images/1509a510253d9b6a0f16ce25070f2617c3e0561aa8977fefe6c4a92d046c0c5c.jpg) +Figure 3: Influences of architecture components on image denoising. (a) Without convolutional layers, upsampling still enables denoising. Transposed convolutions lead to faster performance drop due to earlier noise-fitting than bilinear upsampling. Removing the upsampling decreases denoising capability, which cannot be compensated by simply reducing the number of layers, i.e., under-parameterization. (b) Convolutional layers alone exhibit certain denoising effects but necessitate early stopping. Increasing the number of convolutional layers achieves higher peak PSNR at the expense of earlier noise fitting (see red arrows). Better results achieved when combined with upsampling. + +Results for MLP-Decoder, shown in Fig.3 (a), suggest that upsampling plays a vital role. Removal of upsampling results in significant performance degradation of denoising, which cannot be compensated by simply reducing the size of the network for under-parameterization. This holds true for other tasks such as super-resolution (Suppl.A). Fig.3 (b) shows that learnable spatial filters alone also enable denoising, in contrast to the pixel-wise filters, re-affirming the frequency bias of the convolutional layers [3]. However, the effects vanish as the network size increases and noise-fitting occurs sonner. Convolutional layers together with upsampling achieve better results as manifested in the higher peak PSNR and longer denoising effects. + +Discussion. These results suggest that an appropriate upsampling operation is crucial for effective network image priors, and that under-parameterization alone is insufficient. Different upsampling operations induce varying extents of denoising effects: transposed convolutions [25] tend to fit noise faster than bilinear upsampling, necessitating early stopping. In the next section, we investigate the behaviors of these upsampling operations from a signal-processing perspective to gain insights into their influences on denoising performance. + +# 3.3. Spectral effects of upsampling + +The design choice for upsampling is typically standard: bilinear/nearest neighbor (NN) interpolation or transposed convolutions. These upsampling operations can all be decomposed into two steps: (i) zero-insertion and (ii) filtering. Given a target upsampling factor R, the low-resolution feature map is first interleaved with (R-1) rows/columns of zeros to increase its sampling rate, and then convolved with + +![](images/ddcf37cccf623cbe135d083ec5ef6a96a7eec4aa25faef68756d24bba29ef6b9.jpg) +Figure 4: (Left) Frequency responses of the tested LPFs. Different LPFs result in upsamplers with different extents of smoothing. NN interpolation preserves most signals in the passband but also the high-frequency replica in the stopband; $\mathcal{L}_{-100}$ attenuates the signals most significantly ( $\sim 100dB$ ) and suppresses the high-frequency replica most. (Right) Denoising results on coarse- and fine-textured images from Set9. Top rows: Peak PSNR values. Bottom rows: PSNR values at the last training iteration. + +a low-pass filter (LPF) to remove the alias high frequencies introduced by zero-insertion. The key difference between the upsampling operations lies in the nature of the filters. The filters for transposed convolutions are learnable, but are fixed for bilinear and nearest neighbor upsampling. + +To better illustrate this, we consider the case of a 1D signal $x(n), n = 0, \dots, N - 1$ , and its discrete Fourier representation $X(k) = \sum_{n=0}^{N-1} x(n) e^{\frac{-i2\pi}{N} k n}, k = 0, \dots, N - 1$ . + +For an upsampling factor of 2, we have + +$$ +\begin{array}{l} X ^ {\mathrm {u p}} (\hat {k}) = \sum_ {n = 0} ^ {2 N - 1} x ^ {\mathrm {u p}} (n) e ^ {- \frac {i 2 \pi}{2 N} \hat {k} n} (2) \\ = \sum_ {n = 0} ^ {N - 1} x ^ {\mathrm {u p}} (2 n) e ^ {\frac {- i 2 \pi}{2 N} (2 n) \hat {k}} + \sum_ {n = 0} ^ {N - 1} x ^ {\mathrm {u p}} (2 n + 1) e ^ {\frac {- i 2 \pi}{2 N} (2 n + 1) \hat {k}}, (3) \\ \end{array} +$$ + +where $\hat{k} = 0,\dots ,2N - 1,x^{\mathrm{up}}(2n) = x(n),x^{\mathrm{up}}(2n + 1) = 0$ due to interleaved zero insertion. Hence, for $0\leq \hat{k} < N$ $X^{\mathrm{up}}(\hat{k}) = X(k)$ + +For $\hat{k} \geq N$ , let $k' = \hat{k} - N$ , $k' = 0, \dots, N - 1$ , we have + +$$ +\begin{array}{l} X ^ {\mathrm {u p}} (\hat {k}) = \sum_ {n = 0} ^ {N - 1} x (n) e ^ {- \frac {i 2 \pi}{2 N} \left(k ^ {\prime} + N\right) 2 n} (4) \\ = \sum_ {n = 0} ^ {N - 1} x (n) e ^ {\frac {- i 2 \pi}{N} n k ^ {\prime} - i 2 n \pi} = X \left(k ^ {\prime}\right), (5) \\ \end{array} +$$ + +where Eq.5 exploits the periodicity of the complex exponential function. Thus, zero-insertion will preserve the original spectrum at $[0,N]$ (passband) and additionally create a mirrored copy at $[N,2N - 1]$ (stopband). The high-frequency replica beyond $N$ should be suppressed by the + +Table 1: The iteration number [iter./5000] at the peak PSNR for different upsamplers. The upsamplers are designed to mainly differ in the stopband. The peak PSNR is reached more slowly when the attenuation is stronger (from left to right). + +
NNL14L15L16L17
Coarse-grained17831589168122052214
Fine-grained24242942389843614957
+ +subsequent LPF to avoid image artifacts. According to duality and convolution theorem, convolving with NN or bilinear interpolation filter is equivalent to multiplication of $X^{up}(\hat{k})$ with a Sinc or Sinc² function corresponding to low-pass filtering, while transposed convolutional filter may not necessarily be low-pass as it depends on the optimization objective. + +We experimentally demonstrate that different upsampling operations bias the architecture towards different spectral properties. Specifically, we constructed four upsamplers by first interleaving the input with zeros and then convolving it with handcrafted LPFs (shortened as $\mathcal{L}$ ): $\mathcal{L}14$ , $\mathcal{L}15$ , $\mathcal{L}_{-60}$ and $\mathcal{L}_{-100}$ , with the subscript denoting the decayed dB. By construction, $\mathcal{L}14$ and $\mathcal{L}15$ are very close to NN in the passband ( $< 0.03dB$ ) and only differ in the stopband. The frequency responses of the compared LPFs are detailed in Fig. 4. We applied the customized upsamplers on ConvDecoder and tested them on the fine- and coarse-textured images from Set9 respectively. Similar findings hold for the encoder-decoder architecture (Suppl. C). + +From Fig. 4, upsampling critically influences both the peak PSNR value and the timing for early stopping with respect to images of different textural complexities. Upsamplers (NN, $\mathcal{L}14$ , $\mathcal{L}15$ ) with less attenuation on the high-frequency replica are beneficial for generating fine-grained images, but they tend to cause noise-fitting especially for coarser-grained images. Also, they are generally the fastest to reach the peak PSNR (Table 1). This explains why the transposed convolution requires early-stopping, as the filters may not learn to attenuate the introduced high frequencies effectively. Similar issues with the learned upsampling are also prevalent in generative models [28, 5, 11, 12]. + +On the other hand, bilinear and $\mathcal{L}_{-60}$ exhibit a greater amount of attenuation, more strongly biasing the network against high frequencies and leading to longer-lasting denoising effects. They turn out to work sufficiently well for both kinds of images, especially on the coarser-grained ones which are typically the majority in the dataset. $\mathrm{LPF}_{-100}$ tends to over-smooth the output and performs the worst as it attenuates both the passband and stopband signals substantially, though not requiring early stopping. + +Discussion. These results lead us to conclude that the upsamplers with fixed LPFs are key to the denoising effects of DIP for their tendency towards smooth images (i.e., re + +![](images/b3773ab6ef4badce1d35f251185add8ecc5c54dd7dce50b1077dc48aaf7e6882.jpg) +Figure 5: Influences of width, depth, and skip connections, assuming upsampling inserted in-between the layers. (a) Tendency to over-fit coarse-grained images increases with width. SSIM scores averaged across the depths. (b) Tendency to over-smooth fine-grained images increases with depth, but alleviated by skip connections. SSIM scores are averaged across the widths. +Figure 6: Position of upsampling. Placing the upsampling close to the input (or encoder) can easily cause over-fitting, regardless of the scaling factor. When close to the output (end of the decoder), upsampling with large scaling factors (e.g., $32 \times$ ) causes over-smoothing. + +duced high-frequency contents), which aligns well with the spectral statistics of natural images (Fig.2). Probably due to a good balance between the denoising performance and persistency, bilinear upsampling has been widely adopted in DIP models for various applications [33, 15, 13, 16]. + +# 3.4. Interactions with other architecture elements + +After establishing the significance of upsampling, we will now consider how it may interact with other common architectural elements in affecting the ultimate output. + +Convolutions + non-linearity. Ideal upsampling does not modify the signal representations but only expands the spectrum for the subsequent layers to add new frequency contents. Convolution followed by nonlinearity such as ReLU, is the only operation capable of introducing arbitrarily high frequencies [19]. Increasing the number of layers (depth) or channels (width) enhances the capability of generating new high frequencies, as theoretically and empirically proved in [26]. + +Intuitively, using an excessive number of layers or channels can accelerate the learning of both details and noise. However, the effects can be attenuated by the fixed upsampling operations between the layers. As shown in Fig.5, when using fewer upsampling operations (i.e., a shallower network), increasing the width of the network increases the tendency to cause overfitting on simpler images while benefiting the more complex images. Increasing the number of upsampling operations (i.e., a deeper network) can alleviate over-fitting with stronger attenuation but results in + +![](images/136d16f28ea08ca0d7873758b72c1e9c5c13c4577fed7c728787dc63ae32fb88.jpg) + +blurry outputs for fine-grained images. Increasing only the upsampling factors without adding more layers can make the output even more blurry (Fig.6). In other words, the final output is determined by the balance between the generation of high frequencies by the layers and signal attenuation caused by the upsampling operations. + +Skip connections between the encoder and the decoder often complicate the design space [6]. While they are not directly responsible for denoising, they may lower the effective upsampling rate, making deep networks perform similarly to shallower ones, i.e., resulting in lesser suppression of the high-frequency replica. This finding is validated on a large-scale experiment comprising 7424 architectures (Suppl. D). As exemplified in Fig. 5 (b), skip connections notably ameliorate the over-smoothing issue induced by the same deep network. Overall, skip connections exert a more pronounced influence on the deeper networks compared to the shallower ones, as evident in the smaller deviation observed when depth $\leq 3$ in Fig.5 (b). + +# 3.5. Practical application to architectural design + +Based on the above findings and analysis, we argue that it is possible to estimate an effective architecture for each image without extensive search. Assuming every decoder layer is followed by a $2 \times$ bilinear upsampling layer, our strategies are as follows: + +Depth estimation. We have shown that increased depth tends to over-smooth the output, affecting fine-grained images much more than on coarse-grained images. For fine-grained images, we have three options: a) add more skip connections to a deep network; b) simply use a shallower one; c) keep all the layers but reduce down-/up-sampling layers. We recommend c for decoder-only architectures since they are already under-parameterized; for hourglass network this can easily lead to over-fitting (Fig.6). We find + +Table 2: Quantitative results on Gaussian noise. $\sigma$ denotes the noise level. All methods were trained with a fixed iteration number (3000) throughout the experiments. The highest score is in bold, and the second highest is underlined. + +
DatasetsDIP [33]Deep Decoder [15]ConvDecoder [10]NAS-DIP [6]ISNAS-DIP [1]Ours
σ = 25
Set9 [9]PSNR30.1028.4528.5126.3729.1130.26
SSIM0.8930.8480.8540.7530.8620.900
Set12 [37]PSNR26.9725.9825.7820.8624.1028.14
SSIM0.8120.7890.7860.5340.7450.884
CBSD68 [27]PSNR28.9325.5025.1923.8024.5128.57
SSIM0.8920.8090.7930.6930.7450.888
σ = 50
Set9 [9]PSNR25.0425.2225.0121.0723.9126.13
SSIM0.7610.7640.7690.5930.6980.833
Set12 [37]PSNR22.1520.4422.7218.9219.2024.59
SSIM0.6230.6870.7060.4760.5370.805
CBSD68 [27]PSNR23.7423.5224.0617.9219.9324.17
SSIM0.7460.7250.7670.3230.5730.774
# Params (Millions)2.3M0.1M0.89M4.4MVaried0.05M~0.92M
+ +Table 3: Comparisons of desired properties. Restoration time is computed on an image of size $512 \times 512$ with 3000 iterations. + +
NAS [6]ISNAS [1]Ours
Image-SpecificX✓✓
Architecture Search3 days5 mins-
Per-image Restoration~23 mins~7 hrs~6 mins
Early Stopping Required?YesYesNo
+ +$\mathbf{b}$ is generally better than $\mathbf{a}$ in trading off between good performance and the need for early stopping, especially under higher-level noise, as shown in our results section. This also holds for coarse-grained images as they are not sensitive to depth. More specifically, we find a 2-level hourglass network sufficient for both types of images (Suppl. D). + +Width estimation. Width is crucial for learning sufficient details while avoiding over-fitting especially to a shallow network (less signal attenuation). We find that the width needs to be set according to textural complexity of the image: a finer-grained image requires more channels per layer and vice versa. To further validate this, we treated width estimation as a classification problem, and trained three SVMs [8] with texture features as the inputs to classify the images into three width choices \{32, 64, 128\}. Note that these widths were empirically chosen for the datasets used in this work and are by no means optimal for all cases, but tuning for other images should be straightforward. The classification results and analysis are in Sec.4.4. + +Image scoring. To more robustly classify image texture, we extracted both spatial and frequency features and trained a Decision Tree [22] for feature selection. We used the classic Gray Level Co-occurrence Matrix (GLCM) [14] for deriving the spatial texture features from each image. The most useful features turn out to be the following: correlations measured at $0^{\circ}$ , homogeneity at $45^{\circ}$ , and contrast at + +Table 4: Quantitative evaluation on Poisson noise. Deep Decoder shortened as DD and ConvDecoder as CD. + +
Noise scaleDIPDDCDNASISNASOurs
ζ = 0.01PSNR31.5829.8129.5129.7030.1230.95
SSIM0.9150.8800.8750.8640.8750.910
ζ = 0.1PSNR22.6623.9124.9415.7217.6824.99
SSIM0.6400.7180.7760.4170.4960.789
ζ = 0.2PSNR20.3621.1321.8712.9914.6422.55
SSIM0.5630.6090.6720.3110.3830.774
+ +$0^{\circ}$ . Frequency features are captured using a 1D PSD vector obtained by first converting the 2D PSD from Cartesian coordinates $S(k,l)$ to polar coordinates, and then azimuthally averaging over $\theta$ , defined as $\hat{S}(r) = \frac{1}{2\pi} \int_{0}^{2\pi} S(r,\theta)d\theta$ with $r = \sqrt{k^2 + l^2}$ and $\theta = \mathrm{atan2}(k,l)$ , which represents the mean magnitude of the frequencies with respect to the radial distance $r$ . Refer to Suppl. E for more details. + +# 4. Experiments + +# 4.1. Implementation Details + +We conducted the experiments on three popular datasets and a real-world noisy dataset: Set9 [9] consisting of 9 colored images, Set12 [37] consisting of 12 grey-scaled images, CBSD68 [27] consisting of 68 colored images, and PolyU [36] consisting of 100 real noisy and clean image pairs. We first report our results with 2-level hourglass architecture with the same components as in DIP [33] when comparing with the existing methods, and then extend our strategy to ConvDecoder [10], a decoder-only architecture. All models were trained for 3000 iterations. + +![](images/9ed947076a7849f316d2efa54dcc6c5c499c2b5ee6f2882d3d6704c78a66a2fe.jpg) +Noisy (20.27/0.583) + +![](images/1469b980b9ea844c4e685756dc96e1d7670331c9b29ec2b88987bad293e04a45.jpg) +ConvDecoder (25.19/0.781) + +![](images/ab3d3054eaba591aeb2c8982926ef77599bd75500f3947f44f00289bdf614231.jpg) +Deep Decoder (25.36/0.789) + +![](images/3f4e4196aeadf937c5961b77a6513ebb78d759fcb37d6f7824be0c25a2498491.jpg) +ISNAS-DIP (25.43/0.779) + +![](images/06ef85d3d530455ece4bd00addffec3d56f2580e88d1a35ab6464b4b338be4cc.jpg) +NAS (26.72/0.836) + +![](images/1323e7f67f513cdab527ec41de7554210b98e8c690f893eef3613f6f1b562af4.jpg) +DIP (27.86/0.874) + +![](images/67582ec77a84183f7de8db533caf1d60fa81111476ec83e3603175f6bd4ea176.jpg) +Ours (28.18/0.882) +Figure 7: Denoising results on a fine-grained image ("kodim01" from Set9) with Gaussian noise ( $\sigma = 25$ ). Our estimated architecture for this image is a two-level hourglass network with one skip connection and 128 channels, which is much smaller than others. + +![](images/3b77b57c8e2de72f074de6fca9db92b036d2082d7207c8aa720cf50e6834b84f.jpg) +Ground Truth (PSNR/SSIM) + +![](images/0058c9b2fdfa291bdb1c390a12ffb53bb03366b450d5619c45958d57927e0adb.jpg) +Noisy (36.42/0.967) + +![](images/75df0402a1c28fcbad7ba1d2e3ab25b19e25436b164dd8545a6790faeff6e2ae.jpg) +ConvDecoder (35.02/0.951) + +![](images/ad64a683a5131ea5d44efeb63e963c282884fd89bfd817921d4149cf8b5f3d6d.jpg) +Deep Decoder (35.37/0.955) + +![](images/b1655adbd05a377d4846d7d2310229dad9abf46080f24bc0915b94a84e9264ce.jpg) +DIP w/o. skip (35.14/0.951) + +![](images/bc48ba180b0a394d4ed183f2e2c5fb91ded25c4e4a34caba6d541873bfff2a2f.jpg) +DIP w. skip (36.38/0.964) + +![](images/715e737117a44b78e9e39dca6ef387335141163f0b6139a788e680989667ff40.jpg) +ISNAS-DIP (36.86/0.968) + +![](images/69abfed3152aacbe7dd8d90197074fd1bd81ef8bea3b961383991eb84d056d08.jpg) +NAS-DIP (37.18/0.972) +Figure 8: Denoising results on a real-world noisy image. + +![](images/accfa68a40c9f4ce01eebbf3d9770701c5d0fd0c999bdb1d2d510fd468fc2ab4.jpg) +Ours (37.99/0.979) + +![](images/367a38b6107b30610a8d7141364f87029972d41d242bbfc99d2deb110ebe0e5a.jpg) +Ground Truth (PSNR/SSIM) + +# 4.2. DIP variants + +Gaussian Noise. Compared with the base DIP network, the properly-designed under-parameterized networks estimated with our strategy perform on par at a mild noise level while excel at a higher noise level (Table 2). This cannot be solely explained by under-parameterization since Deep Decoder and ConvDecoder contain similar or even fewer parameters while are unable to achieve similar results. Fig. + +7 shows that a shallow and wide network can preserve the details better than many deeper ones. Note that DIP is a 5-level hourglass network with full skip connections, which by our standard can also well preserve the details. In fact, it is a very strong baseline under mild noise. NAS-DIP and ISNAS-DIP suffer from various degrees of over-fitting on different datasets. This also suggests that the point of optimal stopping varies across images. Besides, they are time-intensive in either searching or evaluation (Table 3). Similar conclusions hold for Poisson noise, which was tested on Set9 [9] as shown in Table 4. + +Table 5: Quantitative evaluation on PolyU, a real-world noisy dataset. Deep Decoder and ConvDecoder shortened as DD, CD. + +
DIPDDCDNASISNASOurs
PSNR38.1537.2237.0037.8337.7838.05
SSIM0.9820.9780.9760.9820.9770.984
+ +Real-World Noise. We additionally evaluated all methods on the PolyU dataset [36]. We applied again the two-level hourglass architectures with variable width estimated for the images. Table 5 summarizes the numerical results. Fig.9 shows that our method tends to preserve more details, though this may not be reflected by the metrics. Simply removing the skip connections from DIP causing blurring, similar to the decoder-only architectures. + +# 4.3. Decoder-only Architectures + +We applied our strategy to ConvDecoder and tested it on all three datasets. Here we keep all 5 layers but remove + +Table 6: Application to ConvDecoder. + +
Datasetsσ = 25σ = 50
BeforeAfterBeforeAfter
Set9PSNR28.5128.7425.0125.11
SSIM0.8540.8730.7690.784
Set12PSNR25.7926.9822.7223.01
SSIM0.7860.8540.7060.742
CBSD68PSNR25.1928.2924.3624.12
SSIM0.7930.8770.7670.768
# Params (Millions)0.89M0.06M~0.89M0.89M0.06M~0.89M
+ +![](images/f00b73fd16c073224f1444236f30f04e25bef6a6c304715618d2358f5c9ae981.jpg) +Figure 10: (a) Width critically influences a shallower network, while (b) it rarely has any impact on a sufficiently deep network. + +a certain number of upsampling layers and scale the width accordingly for the images. These simple changes effectively alleviate the over-smoothing issue that often brought by ConvDecoder as shown in Fig.8 and improve the quantitative results as shown in Table 6. + +# 4.4. More Analysis on Depth and Width + +In practice, we find it more efficient to first determine the depth, and then the width. This is because when the network is deep enough (strong attenuation due to upsampling), width becomes less influential, as evident in Fig.10 (b) and the small deviation in Fig.5 (b). However, to relax the need for early stopping at a higher noise level, one may prefer a shallower one, where width matters more (Fig.10 (a)). In this regard, we use the texture features of the images to predict the desired width. This makes intuitive sense as noise-fitting is associated with the amount of high frequency contents in the image, which is manifested in its texture. This is corroborated by the 0.86 Micro-average AUC score shown in Fig.11 (a). The "optimal" width labels we used to train the classifiers were obtained by experimenting with all three width choices on a 2-level hourglass architecture. These width labels are also applicable to decoder architectures as demonstrated by our experiments. + +Although the width choices seem limited, we found the images robust to the choice of width to some extent, and simple averaging also works well for ambiguous cases (Appendix D). In fact, some images have multiple width labels. We included these cases in the released Texture-DIP dataset. + +![](images/5754ede4b38c3a18f2c7260c34f2cc2734a08603c2a0cb3c43a85a57863d98e7.jpg) + +![](images/0aae8a7306d6ea4f82205c810e0384c9b0a54ebacacdf046dc32d0e52274ccdd.jpg) + +![](images/7eeacf01310fd3804c86f5d5045c665fc81138200bfdda49fe2f9013737f218d.jpg) +Figure 9: Qualitative improvement on ConvDecoder simply by removing three upsampling layers in this case while preserving all the convolutional layers. + +![](images/023522865d6b37451d03b9139f8be488c6b04d5796bf1fd4903996ed75ae284a.jpg) +Figure 11: (a) ROC curves with AUC scores for width classification on images from Set9, Set12 and CBSD68. 5-fold cross-validation is performed and repeated 10 times. (b) Overview of our Texture-DIP Dataset. + +# 4.5. What about Transformers? + +Transformers [34] have become integral parts of modern deep neural network architectures. We experimented with Swin-UNet [2], an encoder-decoder pure transformer consisting of Swin Transformer blocks [21] and skip connections. Note that Swin-UNet relies on patch merging/expanding layers for $2 \times$ down-/upsampling without involving unlearnt upsampling. We replaced all the patch expanding layers with $2 \times$ bilinear upsampling operations and compared the performance of the modified network with the original. We did not replace the patch merging layers with the corresponding downsampling layers or strided convolutions because we did not observe any significant difference. The results presented in Fig. 12 are consistent with our findings on CNNs. For more implementation details, please refer to our code. + +# 5. Conclusion and Future Work + +We presented simple but effective solutions to the challenging DIP architectural design in the context of image denoising. Leveraging the spectral effects of upsampling and its relationships with other architectural components, we show that simple architectural changes yield highly- + +![](images/a20380dbac06f8ee4b129d45b992dd3d329d48d091e1a0ff19dcf072f4dd2fde.jpg) +Noisy (20.38/0.465) + +![](images/0bf8fbc3e4aa8754ff75e318b7c07895151dff8208364b062f46e979f7198bc6.jpg) +Swin-UNet (21.42/0.497) + +![](images/765cb6455ab9c03b8d43dd72c56a6cfe4f45c5826ec026b10dfe48c4dfd89b56.jpg) +Swin-UNet w. Bilinear Up (30.61/0.928) +Figure 12: Visual comparisons on the transformer-based Swin-UNet [2] with and without bilinear upsampling. + +![](images/4958944630f6e3e4c6bfda76d1a568a1d2dbeb501d1f9164b9ce574cf566ecea.jpg) +Ground Truth (PSNR/SSIM) + +effective under-parameterized networks that could surpass the larger counterparts and does not critically rely on early-stopping. Although this work focuses on denoising, insights learned about upsampling can be employed in the future to understand the relationship between architectural characteristics and other image restoration tasks. We hope our study could encourage efficient architectural design for DIP and image synthesis in general. + +# Acknowledgements + +This work was supported in part by the United States National Institutes of Health (NIH) under Grant CA266702 and Grant EB008374. + +# References + +[1] Metin Ersin Arican, Ozgur Kara, Gustav Bredell, and Ender Konukoglu. Isnas-dip: Image-specific neural architecture search for deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1960-1968, 2022. 1, 2, 6 +[2] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 8, 9 +[3] Prithvijit Chakrabarty and Subhransu Maji. The spectral bias of the deep image prior. arXiv preprint arXiv:1912.08905, 2019. 1, 3 +[4] Stanley H Chan, Xiran Wang, and Omar A Elgendy. Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging, 3(1):84-98, 2016. 1 +[5] Keshigeyan Chandrasegaran, Ngoc-Trung Tran, and Ngai-Man Cheung. A closer look at fourier spectrum discrepancies for cnn-generated images detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7200-7209, 2021. 4 +[6] Yun-Chun Chen, Chen Gao, Esther Robb, and Jia-Bin Huang. Nas-dip: Learning deep image prior with neural + +architecture search. In European Conference on Computer Vision, pages 442-459. Springer, 2020. 2, 5, 6 +[7] Zezhou Cheng, Matheus Gadelha, Subhransu Maji, and Daniel Sheldon. A bayesian perspective on the deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5451, 2019. 2 +[8] Corinna Cortes and Vladimir Vapnik. Support-vector networks. Machine learning, 20:273-297, 1995. 6 +[9] Kostadin Dabov, Alessandro Foi, and Karen Egiazarian. Video denoising by sparse 3d transform-domain collaborative filtering. In 2007 15th European Signal Processing Conference, pages 145-149. IEEE, 2007. 6, 7 +[10] Mohammad Zalbagi Darestani and Reinhard Heckel. Accelerated mri with un-trained neural networks. IEEE Transactions on Computational Imaging, 7:724-733, 2021. 1, 2, 3, 6 +[11] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7890–7899, 2020. 4 +[12] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 4 +[13] Yosef Gandelsman, Assaf Shocher, and Michal Irani. "double-dip": unsupervised image decomposition via coupled deep-image-priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11026-11035, 2019. 5 +[14] Robert M Haralick, Karthikeyan Shanmugam, and Its' Hak Dinstein. Textural features for image classification. IEEE Transactions on systems, man, and cybernetics, (6):610-621, 1973. 6 +[15] Reinhard Heckel and Paul Hand. Deep decoder: Concise image representations from untrained non-convolutional networks. arXiv preprint arXiv:1810.03982, 2018. 1, 2, 5, 6 +[16] Reinhard Heckel and Mahdi Soltanolkotabi. Denoising and regularization via exploiting the structural bias of convolu + +tional generators. arXiv preprint arXiv:1910.14634, 2019. 1,5 +[17] Kary Ho, Andrew Gilbert, Hailin Jin, and John Collomosse. Neural architecture search for deep image prior. Computers & graphics, 98:188-196, 2021. 1, 2 +[18] Yeonsik Jo, Se Young Chun, and Jonghyun Choi. Rethinking deep image prior for denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5087-5096, 2021. 2 +[19] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 5 +[20] Jiaming Liu, Yu Sun, Xiaojian Xu, and Ulugbek S Kamilov. Image restoration using total variation regularized deep image prior. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7715-7719. IEEE, 2019. 2 +[21] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 8 +[22] Wei-Yin Loh. Classification and regression trees. Wiley interdisciplinary reviews: data mining and knowledge discovery, 1(1):14-23, 2011. 6 +[23] Florian Luisier, Thierry Blu, and Michael Unser. Image denoising in mixed poisson-gaussian noise. IEEE Transactions on image processing, 20(3):696-708, 2010. 2 +[24] Gary Mataev, Peyman Milanfar, and Michael Elad. Deepred: Deep image prior powered by red. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 2 +[25] Augustus Odena, Vincent Dumoulin, and Chris Olah. Deconvolution and checkerboard artifacts. Distill, 1(10):e3, 2016. 3 +[26] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 5 +[27] Stefan Roth and Michael J Black. Fields of experts. International Journal of Computer Vision, 82(2):205-229, 2009. 6 +[28] Katja Schwarz, Yiyi Liao, and Andreas Geiger. On the frequency bias of generative models. Advances in Neural Information Processing Systems, 34:18126-18136, 2021. 4 +[29] Zenglin Shi, Pascal Mettes, Subhransu Maji, and Cees GM Snoek. On measuring and controlling the spectral bias of the deep image prior. International Journal of Computer Vision, 130(4):885-908, 2022. 2, 3 +[30] Eero P Simoncelli and Bruno A Olshausen. Natural image statistics and neural representation. Annual review of neuroscience, 24(1):1193-1216, 2001. 3 +[31] Shakarim Soltanayev and Se Young Chun. Training deep learning based denoisers without ground truth data. Advances in neural information processing systems, 31, 2018. 2 + +[32] Charles M Stein. Estimation of the mean of a multivariate normal distribution. The annals of Statistics, pages 1135-1151, 1981. 2 +[33] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446-9454, 2018. 1, 3, 5, 6 +[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 8 +[35] Hengkang Wang, Taihui Li, Zhong Zhuang, Tiancong Chen, Hengyue Liang, and Ju Sun. Early stopping for deep image prior. arXiv preprint arXiv:2112.06074, 2021. 2 +[36] Jun Xu, Hui Li, Zhetong Liang, David Zhang, and Lei Zhang. Real-world noisy image denoising: A new benchmark. arXiv preprint arXiv:1804.02603, 2018. 6, 7 +[37] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 6 +[38] Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3929-3938, 2017. 1 \ No newline at end of file diff --git a/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/images.zip b/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d1bb6ed32cf41b9bbc9e9ab580336991c840377f --- /dev/null +++ b/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19d18a26a86caa81a9885a1c24f7dfe4161429a96d318734fd9d8f4473c0482d +size 716041 diff --git a/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/layout.json b/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..38f0e9885a0320f0db93e9a5368cc57ba285236d --- /dev/null +++ b/2023/The Devil is in the Upsampling_ Architectural Decisions Made Simpler for Denoising with Deep Image Prior/layout.json @@ -0,0 +1,9364 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 102, + 525, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 102, + 525, + 140 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 525, + 140 + ], + "type": "text", + "content": "The Devil is in the Upsampling: Architectural Decisions Made Simpler for Denoising with Deep Image Prior" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 148, + 473, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 148, + 473, + 164 + ], + "spans": [ + { + "bbox": [ + 120, + 148, + 473, + 164 + ], + "type": "text", + "content": "Yilin Liu" + }, + { + "bbox": [ + 120, + 148, + 473, + 164 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 120, + 148, + 473, + 164 + ], + "type": "text", + "content": ", Jiang Li" + }, + { + "bbox": [ + 120, + 148, + 473, + 164 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 120, + 148, + 473, + 164 + ], + "type": "text", + "content": ", Yunkui Pang" + }, + { + "bbox": [ + 120, + 148, + 473, + 164 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 120, + 148, + 473, + 164 + ], + "type": "text", + "content": ", Dong Nie" + }, + { + "bbox": [ + 120, + 148, + 473, + 164 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 148, + 473, + 164 + ], + "type": "text", + "content": ", and Pew-Thian Yap" + }, + { + "bbox": [ + 120, + 148, + 473, + 164 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 189, + 166, + 405, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 166, + 405, + 180 + ], + "spans": [ + { + "bbox": [ + 189, + 166, + 405, + 180 + ], + "type": "text", + "content": "1University of North Carolina at Chapel Hill" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 148, + 180, + 446, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 180, + 446, + 191 + ], + "spans": [ + { + "bbox": [ + 148, + 180, + 446, + 191 + ], + "type": "text", + "content": "{yilinliu, yunkuipa, dongnie} @cs.unc.edu, jiang.li@unc.edu, ptyap@med.unc.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 252, + 290, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 252, + 290, + 494 + ], + "spans": [ + { + "bbox": [ + 45, + 252, + 290, + 494 + ], + "type": "text", + "content": "Deep Image Prior (DIP) shows that some network architectures inherently tend towards generating smooth images while resisting noise, a phenomenon known as spectral bias. Image denoising is a natural application of this property. Although denoising with DIP mitigates the need for large training sets, two often intertwined practical challenges need to be overcome: architectural design and noise fitting. Existing methods either handcraft or search for suitable architectures from a vast design space, due to the limited understanding of how architectural choices affect the denoising outcome. In this study, we demonstrate from a frequency perspective that unlearnt upsampling is the main driving force behind the denoising phenomenon with DIP. This finding leads to straightforward strategies for identifying a suitable architecture for every image without laborious search. Extensive experiments show that the estimated architectures achieve superior denoising results than existing methods with up to " + }, + { + "bbox": [ + 45, + 252, + 290, + 494 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 45, + 252, + 290, + 494 + ], + "type": "text", + "content": " fewer parameters. Thanks to this under-parameterization, the resulting architectures are less prone to noise-fitting1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 517, + 128, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 128, + 530 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 128, + 530 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 538, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 684 + ], + "type": "text", + "content": "Image denoising is useful on its own and can be a plug-in module for many other image restoration tasks [37, 38, 4]. Deep neural networks have become the tool of choice for image denoising owing to their ability to learn natural image priors from large-scale datasets. Yet, Deep Image Prior (DIP) [33] requires only a single degraded image for image resotration. Remarkably, DIP shows that a randomly initialized convolutional neural network (CNN) can regularize image restoration through its architecture and early-stopping optimization. This is inspired by the phenomenon that some network architectures act inherently as image priors, favoring generating smooth, natural images and re" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 226, + 423, + 323 + ], + "blocks": [ + { + "bbox": [ + 307, + 226, + 423, + 323 + ], + "lines": [ + { + "bbox": [ + 307, + 226, + 423, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 226, + 423, + 323 + ], + "type": "image", + "image_path": "bd7b06dc5cbb5a34ad44df7ce728b9d897eaf43e119c39ed18a5c170b4952f40.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 328, + 353, + 365 + ], + "lines": [ + { + "bbox": [ + 312, + 328, + 353, + 365 + ], + "spans": [ + { + "bbox": [ + 312, + 328, + 353, + 365 + ], + "type": "text", + "content": "Noisy 20.34/0.648" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 428, + 226, + 545, + 323 + ], + "blocks": [ + { + "bbox": [ + 372, + 328, + 424, + 365 + ], + "lines": [ + { + "bbox": [ + 372, + 328, + 424, + 365 + ], + "spans": [ + { + "bbox": [ + 372, + 328, + 424, + 365 + ], + "type": "text", + "content": "DIP NAS-DIP 8/0.821 21.92/0.680" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 428, + 226, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 428, + 226, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 428, + 226, + 545, + 323 + ], + "type": "image", + "image_path": "ca844309f68f2a77e921b730de11049ec9e94f261727dbd29fa2054654496d2e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 328, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 428, + 328, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 428, + 328, + 545, + 365 + ], + "type": "text", + "content": "ISNAS-DIP Ours Ground Truth 21.04/0.600 25.22/0.841 PSNR " + }, + { + "bbox": [ + 428, + 328, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 428, + 328, + 545, + 365 + ], + "type": "text", + "content": " /SSIM" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 431, + 370, + 545, + 421 + ], + "lines": [ + { + "bbox": [ + 431, + 370, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 431, + 370, + 545, + 421 + ], + "type": "text", + "content": "Deep Decoder Ours Ground Truth 27.33/0.794 32.01/0.924 PSNR ↑ / SSIM ↑" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 315, + 370, + 354, + 422 + ], + "blocks": [ + { + "bbox": [ + 315, + 370, + 354, + 422 + ], + "lines": [ + { + "bbox": [ + 315, + 370, + 354, + 422 + ], + "spans": [ + { + "bbox": [ + 315, + 370, + 354, + 422 + ], + "type": "image", + "image_path": "4e322c5a3024523e2a6b8746a745f90b99b6313a2b1429156322313ef3046831.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 370, + 394, + 421 + ], + "lines": [ + { + "bbox": [ + 354, + 370, + 394, + 421 + ], + "spans": [ + { + "bbox": [ + 354, + 370, + 394, + 421 + ], + "type": "text", + "content": "DIP 28.39/0.824" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 394, + 370, + 431, + 421 + ], + "lines": [ + { + "bbox": [ + 394, + 370, + 431, + 421 + ], + "spans": [ + { + "bbox": [ + 394, + 370, + 431, + 421 + ], + "type": "text", + "content": "ConvDecoder 28.91/0.853" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 442, + 547, + 564 + ], + "lines": [ + { + "bbox": [ + 304, + 442, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 442, + 547, + 564 + ], + "type": "text", + "content": "Figure 1: Top: Performance (SSIM↑) under different levels of Gaussian noises. Bottom: Denoising of a fine-grained " + }, + { + "bbox": [ + 304, + 442, + 547, + 564 + ], + "type": "inline_equation", + "content": "(1^{st}\\mathrm{row})" + }, + { + "bbox": [ + 304, + 442, + 547, + 564 + ], + "type": "text", + "content": " and a coarse-grained image " + }, + { + "bbox": [ + 304, + 442, + 547, + 564 + ], + "type": "inline_equation", + "content": "(2^{nd}\\mathrm{row})" + }, + { + "bbox": [ + 304, + 442, + 547, + 564 + ], + "type": "text", + "content": ". Most existing methods, including the recent image-specific ISNAS-DIP[1], struggle to perform well simultaneously in both cases. Our simple strategies are flexible in image-specific architectural adaptation without requiring a search. Moreover, the results of the lightweight ConvDecoder[10] and Deep Decoder[15] suggest that without proper model setups, under-parameterization itself can neither ensure good denoising performance nor remove the need for early-stopping." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 590, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 639 + ], + "type": "text", + "content": "sisting noises or degradations. However, image denoising with DIP is greatly influenced by architectural design [3, 1, 33, 15, 17], and the associated tendency to fit the original noisy image, i.e., over-fitting [15, 16]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "Architectural design for DIP remains an open problem. One prevailing view is that model under-parameterization limits noise over-fitting and thus mitigates the need for early stopping [15]. However, our experiments reveal that multiple model architectures can exist under a similar parameter budget, where inappropriate model setups can still lead to" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 693, + 176, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 693, + 176, + 703 + ], + "spans": [ + { + "bbox": [ + 59, + 693, + 176, + 703 + ], + "type": "text", + "content": "* These authors contribute equally." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 59, + 703, + 273, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 273, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 273, + 713 + ], + "type": "text", + "content": "1https://github.com/YilinLiu97/FasterDIP-devil-in-upsampling.git" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12408" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "noise fitting and over-smoothing (Fig.1, Fig.3). Another line of work automates the architecture identification using Neural Architecture Search (NAS)[6, 17, 1]. Without prior knowledge on suitable architectures, extensive search incurs substantial computational costs, prohibiting image-wise NAS for optimal restoration [1]. Arican et al. [1] narrow the search space using training-free metrics, but the need for candidate comparison dramatically prolongs the restoration time (" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " 7 hours/image). Moreover, NAS-based models, along with many DIP models, are typically heavily parameterized and prone to over-fitting, as corroborated in our experiments. Thus, their performance largely depends on the timing of early stopping, which is typically image-specific and hard to pinpoint without access to ground truth." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 241, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 314 + ], + "type": "text", + "content": "Directly identifying an effective under-parameterized architecture for each image is challenging in light of the vast number of different architectures and the absence of ground truth for explicit supervision. To simplify the architectural decisions for DIP, we rethink the architectural influences on its performance in the context of image denoising." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 399 + ], + "type": "text", + "content": "We start by noting that denoising performance is attributable to only a few components, primarily the unlearnt upsampling operations. Our frequency analysis reveals that the fixed upsampling operations tend to bias the architecture towards low-frequency contents more strongly than linear or convolutional layers, critically influencing both the peak PSNR and the point of early stopping to avoid noise-fitting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 400, + 288, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 288, + 581 + ], + "type": "text", + "content": "Importantly, this finding leads to empirical discovery on the roles of typical architectural components in DIP: assuming a standard hourglass network, i) simply scaling the depth and width can balance smoothing and preservation of details, due to the low-pass filtering effects of the upsampling operations inserted in-between the layers. As we observed, a wider and shallower network is better at preserving details and therefore suitable for fine-grained images. This suggests that the \"optimal\" architecture can vary across images, and image texture should be considered for more effective denoising. ii) Skip connections make a deep network perform similarly as a shallower one likely by reducing the \"effective upsampling rate\". This implies the possibility of discarding skip connections to simplify DIP architectural design." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Based on these insights, we find it sufficient to restrict the design choices to only the network depth and width, reducing the problem to searching through only a handful of sub-networks and adapting them according to the level of image details. This can be done as a pre-processing step without costly searching or evaluation. We show that this simple strategy works with both the hourglass and decoder structures, and with proper setups, the estimated networks can denoise while preserving the details better than the larger networks with only " + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "inline_equation", + "content": "5\\% \\sim 40\\%" + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": " number of parameters. The resulting under-parameterization alleviates the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 538, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 538, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 538, + 85 + ], + "type": "text", + "content": "need for early stopping. Our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 89, + 547, + 338 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 89, + 545, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 89, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 317, + 89, + 545, + 114 + ], + "type": "text", + "content": "- We pinpoint that unlearnt upsampling is the main driving force behind the spectral bias of DIP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 120, + 547, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 120, + 547, + 181 + ], + "spans": [ + { + "bbox": [ + 317, + 120, + 547, + 181 + ], + "type": "text", + "content": "- Leveraging this insight, we empirically identify the influences of depth, width and skip connections, along with their correlations with image texture, allowing for quick, effective and more interpretable architectural design for every image without the laborious search." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 186, + 545, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 545, + 259 + ], + "type": "text", + "content": "- We are the first to associate DIP architectural design with image texture. To promote future research, we build a Texture-DIP Dataset consisting of images from three popular datasets, reclassified into several predefined width choices – validated through experiments, according to textural complexity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 265, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 265, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 265, + 545, + 338 + ], + "type": "text", + "content": "- We show that with proper setups, a highly underparameterized subnetwork could match and even outperform larger counterparts, especially at a higher noise level. We conducted extensive synthetic and real-world noise removal experiments to validate our findings and approach." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 346, + 392, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 392, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 392, + 360 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 498 + ], + "type": "text", + "content": "DIP Variants. Deep Decoder [15] is an underparameterized network proposed to avoid early stopping. However, our investigations show that underparameterization alone is not sufficient for denoising. For instance, ConvDecoder [10], a convolutional variant of Deep Decoder, contains more parameters but demonstrates less tendency to over-fit (Fig 1). In contrast to NAS-based methods[6, 17, 1], our strategy leverages the observed relationship between the architecture and the image to prevent the exhaustive search. Other variants such as DIP-RED [24] and DIP-TV [20] augment DIP with additional priors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 643 + ], + "type": "text", + "content": "Early-stopping criterion. Cheng et al. [7] perform posterior inference in DIP so as to prevent the need for early stopping by adding Gaussian noise to the gradients. Similarly, Shi et al. [29] regularize the matrix norm of the network weights to alleviate performance decay. However, tuning the regularization granularity can be challenging (Suppl. B). Jo et al. [18] combine DIP with Stein's unbiased risk estimator (SURE) [32] for training without clean images, but SURE is limited to only a few known noise types [31, 23]. Wang et al. [35] propose to track the running variance of the outputs, but this also introduces new hyper-parameters that require non-trivial tuning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 651, + 453, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 453, + 664 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 453, + 664 + ], + "type": "text", + "content": "3. Investigation and Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 671, + 392, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 392, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 392, + 683 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Deep Image Prior. A noisy image " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " can be modeled as: " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = \\mathbf{x} + \\mathbf{n}" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " is the clean counterpart to" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12409" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 72, + 248, + 175 + ], + "blocks": [ + { + "bbox": [ + 73, + 72, + 248, + 175 + ], + "lines": [ + { + "bbox": [ + 73, + 72, + 248, + 175 + ], + "spans": [ + { + "bbox": [ + 73, + 72, + 248, + 175 + ], + "type": "image", + "image_path": "336a54dcb45b4de8b5f091f3c4d7770b40ff40f20124a31f58824faa30a4195e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 288, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 288, + 216 + ], + "type": "text", + "content": "Figure 2: Spectral density of images of different levels of textural complexity, obtained by azimuthally integrating over the power spectrum of the image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "text", + "content": "be recovered and " + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "text", + "content": " is assumed to be " + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "inline_equation", + "content": "i.i.d" + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "text", + "content": ". Gaussian Noise drawn from " + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, \\sigma^2\\mathbf{I})" + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "text", + "content": " being the identity matrix. DIP parameterizes the clean image " + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "text", + "content": " via a network " + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\theta}" + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "text", + "content": " and is optimized to fit the noisy image " + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 46, + 238, + 288, + 286 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 297, + 287, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 297, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 75, + 297, + 287, + 316 + ], + "type": "interline_equation", + "content": "\\theta^ {*} = \\underset {\\theta} {\\arg \\min } \\mathcal {L} (\\mathbf {y}; \\mathbf {G} _ {\\theta} (\\mathbf {z})), \\quad \\mathbf {x} ^ {*} = \\mathbf {G} _ {\\theta^ {*}} (\\mathbf {z}), \\quad (1)", + "image_path": "a39175d7f63257c995905fadb5d380704730b8a8b79855478f48e9326000cd94.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 326, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 326, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 326, + 287, + 396 + ], + "type": "text", + "content": " denotes the fixed white noise input tensor. Such parameterization allows lower-frequency contents to be fitted before the higher-frequency ones [3, 29], exhibiting high impedance to image noises or degradations. Early-stopping is often required to prevent excessive high frequencies from being fitted, such as noise." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 398, + 288, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 288, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 288, + 541 + ], + "type": "text", + "content": "Image complexity. Ideally, DIP should denoise an image while preserving textural details. We define image complexity by its texture, which can be characterized by its power spectral density (PSD). The spectral power of a natural image typically follows an exponential decay from low frequencies to high frequencies [30]. High-frequency components correspond to fine features such as details, while low-frequency components correspond to coarse structures. Hence, fine-grained images contain more high frequencies than coarse-grained images, giving a flatter PSD curve in (Fig.2). Each image is scored based on its texture features, as described in Sec.3.5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 551, + 214, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 214, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 214, + 565 + ], + "type": "text", + "content": "3.2. The importance of upsampling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": "We first identified the core architecture components that affect the denoising performance of DIP. To this end, we analyzed a decoder-only architecture by removing the encoder from the typical encoder-decoder architecture [33], since a decoder is the minimum requirement for reconstructing the final image. Our base model for analysis is a 6-layered convolutional decoder (Conv-Decoder[10]), with " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "128 \\times 3" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " filters per layer except for the last regression layer, followed by batch normalization, ReLU activation function and upsampling. We further simplify the setup by replacing the spatial filters with pixel-wise " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " filters, constructing a non-convolutional variant dubbed MLP-Decoder." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 175 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 175 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 175 + ], + "type": "image", + "image_path": "1509a510253d9b6a0f16ce25070f2617c3e0561aa8977fefe6c4a92d046c0c5c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 182, + 547, + 302 + ], + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 302 + ], + "type": "text", + "content": "Figure 3: Influences of architecture components on image denoising. (a) Without convolutional layers, upsampling still enables denoising. Transposed convolutions lead to faster performance drop due to earlier noise-fitting than bilinear upsampling. Removing the upsampling decreases denoising capability, which cannot be compensated by simply reducing the number of layers, i.e., under-parameterization. (b) Convolutional layers alone exhibit certain denoising effects but necessitate early stopping. Increasing the number of convolutional layers achieves higher peak PSNR at the expense of earlier noise fitting (see red arrows). Better results achieved when combined with upsampling." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 325, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 480 + ], + "type": "text", + "content": "Results for MLP-Decoder, shown in Fig.3 (a), suggest that upsampling plays a vital role. Removal of upsampling results in significant performance degradation of denoising, which cannot be compensated by simply reducing the size of the network for under-parameterization. This holds true for other tasks such as super-resolution (Suppl.A). Fig.3 (b) shows that learnable spatial filters alone also enable denoising, in contrast to the pixel-wise filters, re-affirming the frequency bias of the convolutional layers [3]. However, the effects vanish as the network size increases and noise-fitting occurs sonner. Convolutional layers together with upsampling achieve better results as manifested in the higher peak PSNR and longer denoising effects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 481, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 601 + ], + "type": "text", + "content": "Discussion. These results suggest that an appropriate upsampling operation is crucial for effective network image priors, and that under-parameterization alone is insufficient. Different upsampling operations induce varying extents of denoising effects: transposed convolutions [25] tend to fit noise faster than bilinear upsampling, necessitating early stopping. In the next section, we investigate the behaviors of these upsampling operations from a signal-processing perspective to gain insights into their influences on denoising performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 610, + 470, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 470, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 470, + 624 + ], + "type": "text", + "content": "3.3. Spectral effects of upsampling" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "The design choice for upsampling is typically standard: bilinear/nearest neighbor (NN) interpolation or transposed convolutions. These upsampling operations can all be decomposed into two steps: (i) zero-insertion and (ii) filtering. Given a target upsampling factor R, the low-resolution feature map is first interleaved with (R-1) rows/columns of zeros to increase its sampling rate, and then convolved with" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12410" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 79, + 287, + 207 + ], + "blocks": [ + { + "bbox": [ + 49, + 79, + 287, + 207 + ], + "lines": [ + { + "bbox": [ + 49, + 79, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 79, + 287, + 207 + ], + "type": "image", + "image_path": "ddcf37cccf623cbe135d083ec5ef6a96a7eec4aa25faef68756d24bba29ef6b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 287, + 305 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 305 + ], + "type": "text", + "content": "Figure 4: (Left) Frequency responses of the tested LPFs. Different LPFs result in upsamplers with different extents of smoothing. NN interpolation preserves most signals in the passband but also the high-frequency replica in the stopband; " + }, + { + "bbox": [ + 46, + 217, + 287, + 305 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{-100}" + }, + { + "bbox": [ + 46, + 217, + 287, + 305 + ], + "type": "text", + "content": " attenuates the signals most significantly (" + }, + { + "bbox": [ + 46, + 217, + 287, + 305 + ], + "type": "inline_equation", + "content": "\\sim 100dB" + }, + { + "bbox": [ + 46, + 217, + 287, + 305 + ], + "type": "text", + "content": ") and suppresses the high-frequency replica most. (Right) Denoising results on coarse- and fine-textured images from Set9. Top rows: Peak PSNR values. Bottom rows: PSNR values at the last training iteration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 287, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 372 + ], + "type": "text", + "content": "a low-pass filter (LPF) to remove the alias high frequencies introduced by zero-insertion. The key difference between the upsampling operations lies in the nature of the filters. The filters for transposed convolutions are learnable, but are fixed for bilinear and nearest neighbor upsampling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 372, + 287, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 287, + 410 + ], + "type": "text", + "content": "To better illustrate this, we consider the case of a 1D signal " + }, + { + "bbox": [ + 47, + 372, + 287, + 410 + ], + "type": "inline_equation", + "content": "x(n), n = 0, \\dots, N - 1" + }, + { + "bbox": [ + 47, + 372, + 287, + 410 + ], + "type": "text", + "content": ", and its discrete Fourier representation " + }, + { + "bbox": [ + 47, + 372, + 287, + 410 + ], + "type": "inline_equation", + "content": "X(k) = \\sum_{n=0}^{N-1} x(n) e^{\\frac{-i2\\pi}{N} k n}, k = 0, \\dots, N - 1" + }, + { + "bbox": [ + 47, + 372, + 287, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 410, + 219, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 410, + 219, + 422 + ], + "spans": [ + { + "bbox": [ + 59, + 410, + 219, + 422 + ], + "type": "text", + "content": "For an upsampling factor of 2, we have" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 430, + 293, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 293, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 293, + 509 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} X ^ {\\mathrm {u p}} (\\hat {k}) = \\sum_ {n = 0} ^ {2 N - 1} x ^ {\\mathrm {u p}} (n) e ^ {- \\frac {i 2 \\pi}{2 N} \\hat {k} n} (2) \\\\ = \\sum_ {n = 0} ^ {N - 1} x ^ {\\mathrm {u p}} (2 n) e ^ {\\frac {- i 2 \\pi}{2 N} (2 n) \\hat {k}} + \\sum_ {n = 0} ^ {N - 1} x ^ {\\mathrm {u p}} (2 n + 1) e ^ {\\frac {- i 2 \\pi}{2 N} (2 n + 1) \\hat {k}}, (3) \\\\ \\end{array}", + "image_path": "68296be2471df748950988f2a70222da24cba734e356096463a52688cc565070.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 519, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 519, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 519, + 287, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 519, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\hat{k} = 0,\\dots ,2N - 1,x^{\\mathrm{up}}(2n) = x(n),x^{\\mathrm{up}}(2n + 1) = 0" + }, + { + "bbox": [ + 47, + 519, + 287, + 557 + ], + "type": "text", + "content": " due to interleaved zero insertion. Hence, for " + }, + { + "bbox": [ + 47, + 519, + 287, + 557 + ], + "type": "inline_equation", + "content": "0\\leq \\hat{k} < N" + }, + { + "bbox": [ + 47, + 519, + 287, + 557 + ], + "type": "inline_equation", + "content": "X^{\\mathrm{up}}(\\hat{k}) = X(k)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 558, + 285, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 558, + 285, + 570 + ], + "spans": [ + { + "bbox": [ + 58, + 558, + 285, + 570 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 58, + 558, + 285, + 570 + ], + "type": "inline_equation", + "content": "\\hat{k} \\geq N" + }, + { + "bbox": [ + 58, + 558, + 285, + 570 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 58, + 558, + 285, + 570 + ], + "type": "inline_equation", + "content": "k' = \\hat{k} - N" + }, + { + "bbox": [ + 58, + 558, + 285, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 58, + 558, + 285, + 570 + ], + "type": "inline_equation", + "content": "k' = 0, \\dots, N - 1" + }, + { + "bbox": [ + 58, + 558, + 285, + 570 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 578, + 287, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 578, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 75, + 578, + 287, + 646 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} X ^ {\\mathrm {u p}} (\\hat {k}) = \\sum_ {n = 0} ^ {N - 1} x (n) e ^ {- \\frac {i 2 \\pi}{2 N} \\left(k ^ {\\prime} + N\\right) 2 n} (4) \\\\ = \\sum_ {n = 0} ^ {N - 1} x (n) e ^ {\\frac {- i 2 \\pi}{N} n k ^ {\\prime} - i 2 n \\pi} = X \\left(k ^ {\\prime}\\right), (5) \\\\ \\end{array}", + "image_path": "8ca68ac21ee19c555a1f7fb00bbc0cb4e3cb8535d7b09814947e07a0a6ffc8dc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "where Eq.5 exploits the periodicity of the complex exponential function. Thus, zero-insertion will preserve the original spectrum at " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "[0,N]" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " (passband) and additionally create a mirrored copy at " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "[N,2N - 1]" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " (stopband). The high-frequency replica beyond " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " should be suppressed by the" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 329, + 123, + 520, + 167 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "type": "text", + "content": "Table 1: The iteration number [iter./5000] at the peak PSNR for different upsamplers. The upsamplers are designed to mainly differ in the stopband. The peak PSNR is reached more slowly when the attenuation is stronger (from left to right)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 123, + 520, + 167 + ], + "lines": [ + { + "bbox": [ + 329, + 123, + 520, + 167 + ], + "spans": [ + { + "bbox": [ + 329, + 123, + 520, + 167 + ], + "type": "table", + "html": "
NNL14L15L16L17
Coarse-grained17831589168122052214
Fine-grained24242942389843614957
", + "image_path": "52c7e87bfed4a05d4da40f74bd39c56d70bbb0d36286edf81546ac6ee8b64a41.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "text", + "content": "subsequent LPF to avoid image artifacts. According to duality and convolution theorem, convolving with NN or bilinear interpolation filter is equivalent to multiplication of " + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "inline_equation", + "content": "X^{up}(\\hat{k})" + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "text", + "content": " with a Sinc or Sinc² function corresponding to low-pass filtering, while transposed convolutional filter may not necessarily be low-pass as it depends on the optimization objective." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "text", + "content": "We experimentally demonstrate that different upsampling operations bias the architecture towards different spectral properties. Specifically, we constructed four upsamplers by first interleaving the input with zeros and then convolving it with handcrafted LPFs (shortened as " + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "text", + "content": "): " + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}14" + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}15" + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{-60}" + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{-100}" + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "text", + "content": ", with the subscript denoting the decayed dB. By construction, " + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}14" + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{L}15" + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "text", + "content": " are very close to NN in the passband (" + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "inline_equation", + "content": "< 0.03dB" + }, + { + "bbox": [ + 304, + 270, + 546, + 426 + ], + "type": "text", + "content": ") and only differ in the stopband. The frequency responses of the compared LPFs are detailed in Fig. 4. We applied the customized upsamplers on ConvDecoder and tested them on the fine- and coarse-textured images from Set9 respectively. Similar findings hold for the encoder-decoder architecture (Suppl. C)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 426, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 569 + ], + "type": "text", + "content": "From Fig. 4, upsampling critically influences both the peak PSNR value and the timing for early stopping with respect to images of different textural complexities. Upsamplers (NN, " + }, + { + "bbox": [ + 304, + 426, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}14" + }, + { + "bbox": [ + 304, + 426, + 545, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 426, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{L}15" + }, + { + "bbox": [ + 304, + 426, + 545, + 569 + ], + "type": "text", + "content": ") with less attenuation on the high-frequency replica are beneficial for generating fine-grained images, but they tend to cause noise-fitting especially for coarser-grained images. Also, they are generally the fastest to reach the peak PSNR (Table 1). This explains why the transposed convolution requires early-stopping, as the filters may not learn to attenuate the introduced high frequencies effectively. Similar issues with the learned upsampling are also prevalent in generative models [28, 5, 11, 12]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "type": "text", + "content": "On the other hand, bilinear and " + }, + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{-60}" + }, + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "type": "text", + "content": " exhibit a greater amount of attenuation, more strongly biasing the network against high frequencies and leading to longer-lasting denoising effects. They turn out to work sufficiently well for both kinds of images, especially on the coarser-grained ones which are typically the majority in the dataset. " + }, + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{LPF}_{-100}" + }, + { + "bbox": [ + 304, + 570, + 545, + 677 + ], + "type": "text", + "content": " tends to over-smooth the output and performs the worst as it attenuates both the passband and stopband signals substantially, though not requiring early stopping." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Discussion. These results lead us to conclude that the upsamplers with fixed LPFs are key to the denoising effects of DIP for their tendency towards smooth images (i.e., re" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12411" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 289, + 257 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 289, + 257 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 289, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 289, + 257 + ], + "type": "image", + "image_path": "b3773ab6ef4badce1d35f251185add8ecc5c54dd7dce50b1077dc48aaf7e6882.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 275, + 287, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 287, + 342 + ], + "type": "text", + "content": "Figure 5: Influences of width, depth, and skip connections, assuming upsampling inserted in-between the layers. (a) Tendency to over-fit coarse-grained images increases with width. SSIM scores averaged across the depths. (b) Tendency to over-smooth fine-grained images increases with depth, but alleviated by skip connections. SSIM scores are averaged across the widths." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 208, + 545, + 263 + ], + "lines": [ + { + "bbox": [ + 305, + 208, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 305, + 208, + 545, + 263 + ], + "type": "text", + "content": "Figure 6: Position of upsampling. Placing the upsampling close to the input (or encoder) can easily cause over-fitting, regardless of the scaling factor. When close to the output (end of the decoder), upsampling with large scaling factors (e.g., " + }, + { + "bbox": [ + 305, + 208, + 545, + 263 + ], + "type": "inline_equation", + "content": "32 \\times" + }, + { + "bbox": [ + 305, + 208, + 545, + 263 + ], + "type": "text", + "content": ") causes over-smoothing." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 363, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 424 + ], + "type": "text", + "content": "duced high-frequency contents), which aligns well with the spectral statistics of natural images (Fig.2). Probably due to a good balance between the denoising performance and persistency, bilinear upsampling has been widely adopted in DIP models for various applications [33, 15, 13, 16]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 431, + 281, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 281, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 281, + 443 + ], + "type": "text", + "content": "3.4. Interactions with other architecture elements" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 450, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 486 + ], + "type": "text", + "content": "After establishing the significance of upsampling, we will now consider how it may interact with other common architectural elements in affecting the ultimate output." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 487, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 594 + ], + "type": "text", + "content": "Convolutions + non-linearity. Ideal upsampling does not modify the signal representations but only expands the spectrum for the subsequent layers to add new frequency contents. Convolution followed by nonlinearity such as ReLU, is the only operation capable of introducing arbitrarily high frequencies [19]. Increasing the number of layers (depth) or channels (width) enhances the capability of generating new high frequencies, as theoretically and empirically proved in [26]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Intuitively, using an excessive number of layers or channels can accelerate the learning of both details and noise. However, the effects can be attenuated by the fixed upsampling operations between the layers. As shown in Fig.5, when using fewer upsampling operations (i.e., a shallower network), increasing the width of the network increases the tendency to cause overfitting on simpler images while benefiting the more complex images. Increasing the number of upsampling operations (i.e., a deeper network) can alleviate over-fitting with stronger attenuation but results in" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 331, + 72, + 523, + 201 + ], + "blocks": [ + { + "bbox": [ + 331, + 72, + 523, + 201 + ], + "lines": [ + { + "bbox": [ + 331, + 72, + 523, + 201 + ], + "spans": [ + { + "bbox": [ + 331, + 72, + 523, + 201 + ], + "type": "image", + "image_path": "136d16f28ea08ca0d7873758b72c1e9c5c13c4577fed7c728787dc63ae32fb88.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 286, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 358 + ], + "type": "text", + "content": "blurry outputs for fine-grained images. Increasing only the upsampling factors without adding more layers can make the output even more blurry (Fig.6). In other words, the final output is determined by the balance between the generation of high frequencies by the layers and signal attenuation caused by the upsampling operations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 359, + 546, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 359, + 546, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 546, + 515 + ], + "type": "text", + "content": "Skip connections between the encoder and the decoder often complicate the design space [6]. While they are not directly responsible for denoising, they may lower the effective upsampling rate, making deep networks perform similarly to shallower ones, i.e., resulting in lesser suppression of the high-frequency replica. This finding is validated on a large-scale experiment comprising 7424 architectures (Suppl. D). As exemplified in Fig. 5 (b), skip connections notably ameliorate the over-smoothing issue induced by the same deep network. Overall, skip connections exert a more pronounced influence on the deeper networks compared to the shallower ones, as evident in the smaller deviation observed when depth " + }, + { + "bbox": [ + 304, + 359, + 546, + 515 + ], + "type": "inline_equation", + "content": "\\leq 3" + }, + { + "bbox": [ + 304, + 359, + 546, + 515 + ], + "type": "text", + "content": " in Fig.5 (b)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 525, + 533, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 533, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 533, + 539 + ], + "type": "text", + "content": "3.5. Practical application to architectural design" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": "Based on the above findings and analysis, we argue that it is possible to estimate an effective architecture for each image without extensive search. Assuming every decoder layer is followed by a " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " bilinear upsampling layer, our strategies are as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Depth estimation. We have shown that increased depth tends to over-smooth the output, affecting fine-grained images much more than on coarse-grained images. For fine-grained images, we have three options: a) add more skip connections to a deep network; b) simply use a shallower one; c) keep all the layers but reduce down-/up-sampling layers. We recommend c for decoder-only architectures since they are already under-parameterized; for hourglass network this can easily lead to over-fitting (Fig.6). We find" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12412" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 102, + 509, + 273 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 2: Quantitative results on Gaussian noise. " + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": " denotes the noise level. All methods were trained with a fixed iteration number (3000) throughout the experiments. The highest score is in bold, and the second highest is underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 102, + 509, + 273 + ], + "lines": [ + { + "bbox": [ + 85, + 102, + 509, + 273 + ], + "spans": [ + { + "bbox": [ + 85, + 102, + 509, + 273 + ], + "type": "table", + "html": "
DatasetsDIP [33]Deep Decoder [15]ConvDecoder [10]NAS-DIP [6]ISNAS-DIP [1]Ours
σ = 25
Set9 [9]PSNR30.1028.4528.5126.3729.1130.26
SSIM0.8930.8480.8540.7530.8620.900
Set12 [37]PSNR26.9725.9825.7820.8624.1028.14
SSIM0.8120.7890.7860.5340.7450.884
CBSD68 [27]PSNR28.9325.5025.1923.8024.5128.57
SSIM0.8920.8090.7930.6930.7450.888
σ = 50
Set9 [9]PSNR25.0425.2225.0121.0723.9126.13
SSIM0.7610.7640.7690.5930.6980.833
Set12 [37]PSNR22.1520.4422.7218.9219.2024.59
SSIM0.6230.6870.7060.4760.5370.805
CBSD68 [27]PSNR23.7423.5224.0617.9219.9324.17
SSIM0.7460.7250.7670.3230.5730.774
# Params (Millions)2.3M0.1M0.89M4.4MVaried0.05M~0.92M
", + "image_path": "f141af135de27177544c1a98154a26eb4ab025584958a34ac6301c8291e7222a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 64, + 322, + 266, + 374 + ], + "blocks": [ + { + "bbox": [ + 47, + 290, + 287, + 313 + ], + "lines": [ + { + "bbox": [ + 47, + 290, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 47, + 290, + 287, + 313 + ], + "type": "text", + "content": "Table 3: Comparisons of desired properties. Restoration time is computed on an image of size " + }, + { + "bbox": [ + 47, + 290, + 287, + 313 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 47, + 290, + 287, + 313 + ], + "type": "text", + "content": " with 3000 iterations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 322, + 266, + 374 + ], + "lines": [ + { + "bbox": [ + 64, + 322, + 266, + 374 + ], + "spans": [ + { + "bbox": [ + 64, + 322, + 266, + 374 + ], + "type": "table", + "html": "
NAS [6]ISNAS [1]Ours
Image-SpecificX✓✓
Architecture Search3 days5 mins-
Per-image Restoration~23 mins~7 hrs~6 mins
Early Stopping Required?YesYesNo
", + "image_path": "45628dba47b89aaffc631b486f58596d2808b2818e39c40189232e9ec56d72de.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "type": "text", + "content": " is generally better than " + }, + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 46, + 398, + 287, + 470 + ], + "type": "text", + "content": " in trading off between good performance and the need for early stopping, especially under higher-level noise, as shown in our results section. This also holds for coarse-grained images as they are not sensitive to depth. More specifically, we find a 2-level hourglass network sufficient for both types of images (Suppl. D)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 472, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 628 + ], + "type": "text", + "content": "Width estimation. Width is crucial for learning sufficient details while avoiding over-fitting especially to a shallow network (less signal attenuation). We find that the width needs to be set according to textural complexity of the image: a finer-grained image requires more channels per layer and vice versa. To further validate this, we treated width estimation as a classification problem, and trained three SVMs [8] with texture features as the inputs to classify the images into three width choices \\{32, 64, 128\\}. Note that these widths were empirically chosen for the datasets used in this work and are by no means optimal for all cases, but tuning for other images should be straightforward. The classification results and analysis are in Sec.4.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Image scoring. To more robustly classify image texture, we extracted both spatial and frequency features and trained a Decision Tree [22] for feature selection. We used the classic Gray Level Co-occurrence Matrix (GLCM) [14] for deriving the spatial texture features from each image. The most useful features turn out to be the following: correlations measured at " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ", homogeneity at " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ", and contrast at" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 322, + 543, + 398 + ], + "blocks": [ + { + "bbox": [ + 305, + 290, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 305, + 290, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 545, + 313 + ], + "type": "text", + "content": "Table 4: Quantitative evaluation on Poisson noise. Deep Decoder shortened as DD and ConvDecoder as CD." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 322, + 543, + 398 + ], + "lines": [ + { + "bbox": [ + 306, + 322, + 543, + 398 + ], + "spans": [ + { + "bbox": [ + 306, + 322, + 543, + 398 + ], + "type": "table", + "html": "
Noise scaleDIPDDCDNASISNASOurs
ζ = 0.01PSNR31.5829.8129.5129.7030.1230.95
SSIM0.9150.8800.8750.8640.8750.910
ζ = 0.1PSNR22.6623.9124.9415.7217.6824.99
SSIM0.6400.7180.7760.4170.4960.789
ζ = 0.2PSNR20.3621.1321.8712.9914.6422.55
SSIM0.5630.6090.6720.3110.3830.774
", + "image_path": "f1dbb0f52bd4533939d9e89658830faad14aee39f3c9a3141a22d78cc32ab6d5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "text", + "content": ". Frequency features are captured using a 1D PSD vector obtained by first converting the 2D PSD from Cartesian coordinates " + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "inline_equation", + "content": "S(k,l)" + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "text", + "content": " to polar coordinates, and then azimuthally averaging over " + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "text", + "content": ", defined as " + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "inline_equation", + "content": "\\hat{S}(r) = \\frac{1}{2\\pi} \\int_{0}^{2\\pi} S(r,\\theta)d\\theta" + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "inline_equation", + "content": "r = \\sqrt{k^2 + l^2}" + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "inline_equation", + "content": "\\theta = \\mathrm{atan2}(k,l)" + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "text", + "content": ", which represents the mean magnitude of the frequencies with respect to the radial distance " + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 305, + 436, + 545, + 524 + ], + "type": "text", + "content": ". Refer to Suppl. E for more details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 547, + 388, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 547, + 388, + 561 + ], + "spans": [ + { + "bbox": [ + 306, + 547, + 388, + 561 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 571, + 440, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 571, + 440, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 571, + 440, + 585 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "We conducted the experiments on three popular datasets and a real-world noisy dataset: Set9 [9] consisting of 9 colored images, Set12 [37] consisting of 12 grey-scaled images, CBSD68 [27] consisting of 68 colored images, and PolyU [36] consisting of 100 real noisy and clean image pairs. We first report our results with 2-level hourglass architecture with the same components as in DIP [33] when comparing with the existing methods, and then extend our strategy to ConvDecoder [10], a decoder-only architecture. All models were trained for 3000 iterations." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12413" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 82, + 200, + 171 + ], + "blocks": [ + { + "bbox": [ + 127, + 72, + 179, + 80 + ], + "lines": [ + { + "bbox": [ + 127, + 72, + 179, + 80 + ], + "spans": [ + { + "bbox": [ + 127, + 72, + 179, + 80 + ], + "type": "text", + "content": "Noisy (20.27/0.583)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 109, + 82, + 200, + 171 + ], + "lines": [ + { + "bbox": [ + 109, + 82, + 200, + 171 + ], + "spans": [ + { + "bbox": [ + 109, + 82, + 200, + 171 + ], + "type": "image", + "image_path": "9ed947076a7849f316d2efa54dcc6c5c499c2b5ee6f2882d3d6704c78a66a2fe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 202, + 82, + 294, + 171 + ], + "blocks": [ + { + "bbox": [ + 213, + 73, + 284, + 80 + ], + "lines": [ + { + "bbox": [ + 213, + 73, + 284, + 80 + ], + "spans": [ + { + "bbox": [ + 213, + 73, + 284, + 80 + ], + "type": "text", + "content": "ConvDecoder (25.19/0.781)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 202, + 82, + 294, + 171 + ], + "lines": [ + { + "bbox": [ + 202, + 82, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 202, + 82, + 294, + 171 + ], + "type": "image", + "image_path": "1469b980b9ea844c4e685756dc96e1d7670331c9b29ec2b88987bad293e04a45.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 296, + 82, + 388, + 171 + ], + "blocks": [ + { + "bbox": [ + 306, + 73, + 378, + 80 + ], + "lines": [ + { + "bbox": [ + 306, + 73, + 378, + 80 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 378, + 80 + ], + "type": "text", + "content": "Deep Decoder (25.36/0.789)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 82, + 388, + 171 + ], + "lines": [ + { + "bbox": [ + 296, + 82, + 388, + 171 + ], + "spans": [ + { + "bbox": [ + 296, + 82, + 388, + 171 + ], + "type": "image", + "image_path": "ab3d3054eaba591aeb2c8982926ef77599bd75500f3947f44f00289bdf614231.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 391, + 82, + 482, + 171 + ], + "blocks": [ + { + "bbox": [ + 402, + 73, + 468, + 80 + ], + "lines": [ + { + "bbox": [ + 402, + 73, + 468, + 80 + ], + "spans": [ + { + "bbox": [ + 402, + 73, + 468, + 80 + ], + "type": "text", + "content": "ISNAS-DIP (25.43/0.779)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 82, + 482, + 171 + ], + "lines": [ + { + "bbox": [ + 391, + 82, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 391, + 82, + 482, + 171 + ], + "type": "image", + "image_path": "3f4e4196aeadf937c5961b77a6513ebb78d759fcb37d6f7824be0c25a2498491.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 109, + 182, + 201, + 271 + ], + "blocks": [ + { + "bbox": [ + 130, + 174, + 179, + 181 + ], + "lines": [ + { + "bbox": [ + 130, + 174, + 179, + 181 + ], + "spans": [ + { + "bbox": [ + 130, + 174, + 179, + 181 + ], + "type": "text", + "content": "NAS (26.72/0.836)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 109, + 182, + 201, + 271 + ], + "lines": [ + { + "bbox": [ + 109, + 182, + 201, + 271 + ], + "spans": [ + { + "bbox": [ + 109, + 182, + 201, + 271 + ], + "type": "image", + "image_path": "06ef85d3d530455ece4bd00addffec3d56f2580e88d1a35ab6464b4b338be4cc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 203, + 182, + 294, + 271 + ], + "blocks": [ + { + "bbox": [ + 225, + 174, + 272, + 181 + ], + "lines": [ + { + "bbox": [ + 225, + 174, + 272, + 181 + ], + "spans": [ + { + "bbox": [ + 225, + 174, + 272, + 181 + ], + "type": "text", + "content": "DIP (27.86/0.874)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 203, + 182, + 294, + 271 + ], + "lines": [ + { + "bbox": [ + 203, + 182, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 203, + 182, + 294, + 271 + ], + "type": "image", + "image_path": "1323e7f67f513cdab527ec41de7554210b98e8c690f893eef3613f6f1b562af4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 296, + 182, + 388, + 271 + ], + "blocks": [ + { + "bbox": [ + 316, + 174, + 365, + 181 + ], + "lines": [ + { + "bbox": [ + 316, + 174, + 365, + 181 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 365, + 181 + ], + "type": "text", + "content": "Ours (28.18/0.882)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 182, + 388, + 271 + ], + "lines": [ + { + "bbox": [ + 296, + 182, + 388, + 271 + ], + "spans": [ + { + "bbox": [ + 296, + 182, + 388, + 271 + ], + "type": "image", + "image_path": "67582ec77a84183f7de8db533caf1d60fa81111476ec83e3603175f6bd4ea176.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 279, + 545, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 279, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 545, + 301 + ], + "type": "text", + "content": "Figure 7: Denoising results on a fine-grained image (\"kodim01\" from Set9) with Gaussian noise (" + }, + { + "bbox": [ + 46, + 279, + 545, + 301 + ], + "type": "inline_equation", + "content": "\\sigma = 25" + }, + { + "bbox": [ + 46, + 279, + 545, + 301 + ], + "type": "text", + "content": "). Our estimated architecture for this image is a two-level hourglass network with one skip connection and 128 channels, which is much smaller than others." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 391, + 182, + 482, + 271 + ], + "blocks": [ + { + "bbox": [ + 396, + 174, + 473, + 181 + ], + "lines": [ + { + "bbox": [ + 396, + 174, + 473, + 181 + ], + "spans": [ + { + "bbox": [ + 396, + 174, + 473, + 181 + ], + "type": "text", + "content": "Ground Truth (PSNR/SSIM)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 182, + 482, + 271 + ], + "lines": [ + { + "bbox": [ + 391, + 182, + 482, + 271 + ], + "spans": [ + { + "bbox": [ + 391, + 182, + 482, + 271 + ], + "type": "image", + "image_path": "3b77b57c8e2de72f074de6fca9db92b036d2082d7207c8aa720cf50e6834b84f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 59, + 332, + 129, + 401 + ], + "blocks": [ + { + "bbox": [ + 72, + 323, + 117, + 331 + ], + "lines": [ + { + "bbox": [ + 72, + 323, + 117, + 331 + ], + "spans": [ + { + "bbox": [ + 72, + 323, + 117, + 331 + ], + "type": "text", + "content": "Noisy (36.42/0.967)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 332, + 129, + 401 + ], + "lines": [ + { + "bbox": [ + 59, + 332, + 129, + 401 + ], + "spans": [ + { + "bbox": [ + 59, + 332, + 129, + 401 + ], + "type": "image", + "image_path": "0058c9b2fdfa291bdb1c390a12ffb53bb03366b450d5619c45958d57927e0adb.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 130, + 332, + 200, + 401 + ], + "blocks": [ + { + "bbox": [ + 134, + 323, + 197, + 331 + ], + "lines": [ + { + "bbox": [ + 134, + 323, + 197, + 331 + ], + "spans": [ + { + "bbox": [ + 134, + 323, + 197, + 331 + ], + "type": "text", + "content": "ConvDecoder (35.02/0.951)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 332, + 200, + 401 + ], + "lines": [ + { + "bbox": [ + 130, + 332, + 200, + 401 + ], + "spans": [ + { + "bbox": [ + 130, + 332, + 200, + 401 + ], + "type": "image", + "image_path": "75df0402a1c28fcbad7ba1d2e3ab25b19e25436b164dd8545a6790faeff6e2ae.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 201, + 332, + 271, + 401 + ], + "blocks": [ + { + "bbox": [ + 205, + 323, + 269, + 331 + ], + "lines": [ + { + "bbox": [ + 205, + 323, + 269, + 331 + ], + "spans": [ + { + "bbox": [ + 205, + 323, + 269, + 331 + ], + "type": "text", + "content": "Deep Decoder (35.37/0.955)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 201, + 332, + 271, + 401 + ], + "lines": [ + { + "bbox": [ + 201, + 332, + 271, + 401 + ], + "spans": [ + { + "bbox": [ + 201, + 332, + 271, + 401 + ], + "type": "image", + "image_path": "ad64a683a5131ea5d44efeb63e963c282884fd89bfd817921d4149cf8b5f3d6d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 59, + 411, + 129, + 481 + ], + "blocks": [ + { + "bbox": [ + 64, + 403, + 126, + 411 + ], + "lines": [ + { + "bbox": [ + 64, + 403, + 126, + 411 + ], + "spans": [ + { + "bbox": [ + 64, + 403, + 126, + 411 + ], + "type": "text", + "content": "DIP w/o. skip (35.14/0.951)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 411, + 129, + 481 + ], + "lines": [ + { + "bbox": [ + 59, + 411, + 129, + 481 + ], + "spans": [ + { + "bbox": [ + 59, + 411, + 129, + 481 + ], + "type": "image", + "image_path": "b1655adbd05a377d4846d7d2310229dad9abf46080f24bc0915b94a84e9264ce.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 130, + 411, + 200, + 481 + ], + "blocks": [ + { + "bbox": [ + 137, + 403, + 194, + 411 + ], + "lines": [ + { + "bbox": [ + 137, + 403, + 194, + 411 + ], + "spans": [ + { + "bbox": [ + 137, + 403, + 194, + 411 + ], + "type": "text", + "content": "DIP w. skip (36.38/0.964)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 411, + 200, + 481 + ], + "lines": [ + { + "bbox": [ + 130, + 411, + 200, + 481 + ], + "spans": [ + { + "bbox": [ + 130, + 411, + 200, + 481 + ], + "type": "image", + "image_path": "bc48ba180b0a394d4ed183f2e2c5fb91ded25c4e4a34caba6d541873bfff2a2f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 201, + 411, + 271, + 481 + ], + "blocks": [ + { + "bbox": [ + 208, + 404, + 265, + 411 + ], + "lines": [ + { + "bbox": [ + 208, + 404, + 265, + 411 + ], + "spans": [ + { + "bbox": [ + 208, + 404, + 265, + 411 + ], + "type": "text", + "content": "ISNAS-DIP (36.86/0.968)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 201, + 411, + 271, + 481 + ], + "lines": [ + { + "bbox": [ + 201, + 411, + 271, + 481 + ], + "spans": [ + { + "bbox": [ + 201, + 411, + 271, + 481 + ], + "type": "image", + "image_path": "715e737117a44b78e9e39dca6ef387335141163f0b6139a788e680989667ff40.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 59, + 491, + 129, + 561 + ], + "blocks": [ + { + "bbox": [ + 67, + 483, + 121, + 491 + ], + "lines": [ + { + "bbox": [ + 67, + 483, + 121, + 491 + ], + "spans": [ + { + "bbox": [ + 67, + 483, + 121, + 491 + ], + "type": "text", + "content": "NAS-DIP (37.18/0.972)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 491, + 129, + 561 + ], + "lines": [ + { + "bbox": [ + 59, + 491, + 129, + 561 + ], + "spans": [ + { + "bbox": [ + 59, + 491, + 129, + 561 + ], + "type": "image", + "image_path": "69abfed3152aacbe7dd8d90197074fd1bd81ef8bea3b961383991eb84d056d08.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 570, + 272, + 580 + ], + "lines": [ + { + "bbox": [ + 61, + 570, + 272, + 580 + ], + "spans": [ + { + "bbox": [ + 61, + 570, + 272, + 580 + ], + "type": "text", + "content": "Figure 8: Denoising results on a real-world noisy image." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 130, + 491, + 200, + 561 + ], + "blocks": [ + { + "bbox": [ + 143, + 484, + 187, + 491 + ], + "lines": [ + { + "bbox": [ + 143, + 484, + 187, + 491 + ], + "spans": [ + { + "bbox": [ + 143, + 484, + 187, + 491 + ], + "type": "text", + "content": "Ours (37.99/0.979)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 491, + 200, + 561 + ], + "lines": [ + { + "bbox": [ + 130, + 491, + 200, + 561 + ], + "spans": [ + { + "bbox": [ + 130, + 491, + 200, + 561 + ], + "type": "image", + "image_path": "accfa68a40c9f4ce01eebbf3d9770701c5d0fd0c999bdb1d2d510fd468fc2ab4.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 201, + 491, + 271, + 561 + ], + "blocks": [ + { + "bbox": [ + 203, + 484, + 270, + 491 + ], + "lines": [ + { + "bbox": [ + 203, + 484, + 270, + 491 + ], + "spans": [ + { + "bbox": [ + 203, + 484, + 270, + 491 + ], + "type": "text", + "content": "Ground Truth (PSNR/SSIM)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 201, + 491, + 271, + 561 + ], + "lines": [ + { + "bbox": [ + 201, + 491, + 271, + 561 + ], + "spans": [ + { + "bbox": [ + 201, + 491, + 271, + 561 + ], + "type": "image", + "image_path": "367a38b6107b30610a8d7141364f87029972d41d242bbfc99d2deb110ebe0e5a.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 47, + 608, + 129, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 129, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 129, + 619 + ], + "type": "text", + "content": "4.2. DIP variants" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Gaussian Noise. Compared with the base DIP network, the properly-designed under-parameterized networks estimated with our strategy perform on par at a mild noise level while excel at a higher noise level (Table 2). This cannot be solely explained by under-parameterization since Deep Decoder and ConvDecoder contain similar or even fewer parameters while are unable to achieve similar results. Fig." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 305, + 323, + 545, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 454 + ], + "type": "text", + "content": "7 shows that a shallow and wide network can preserve the details better than many deeper ones. Note that DIP is a 5-level hourglass network with full skip connections, which by our standard can also well preserve the details. In fact, it is a very strong baseline under mild noise. NAS-DIP and ISNAS-DIP suffer from various degrees of over-fitting on different datasets. This also suggests that the point of optimal stopping varies across images. Besides, they are time-intensive in either searching or evaluation (Table 3). Similar conclusions hold for Poisson noise, which was tested on Set9 [9] as shown in Table 4." + } + ] + } + ], + "index": 38 + }, + { + "type": "table", + "bbox": [ + 312, + 497, + 536, + 544 + ], + "blocks": [ + { + "bbox": [ + 306, + 465, + 545, + 487 + ], + "lines": [ + { + "bbox": [ + 306, + 465, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 545, + 487 + ], + "type": "text", + "content": "Table 5: Quantitative evaluation on PolyU, a real-world noisy dataset. Deep Decoder and ConvDecoder shortened as DD, CD." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 497, + 536, + 544 + ], + "lines": [ + { + "bbox": [ + 312, + 497, + 536, + 544 + ], + "spans": [ + { + "bbox": [ + 312, + 497, + 536, + 544 + ], + "type": "table", + "html": "
DIPDDCDNASISNASOurs
PSNR38.1537.2237.0037.8337.7838.05
SSIM0.9820.9780.9760.9820.9770.984
", + "image_path": "656015fdbe9d687cd9928dd5ed3701a7a1d186f835135b01f3674148a159ee4e.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "table_body" + } + ], + "index": 40 + }, + { + "bbox": [ + 305, + 566, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 566, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 305, + 566, + 545, + 662 + ], + "type": "text", + "content": "Real-World Noise. We additionally evaluated all methods on the PolyU dataset [36]. We applied again the two-level hourglass architectures with variable width estimated for the images. Table 5 summarizes the numerical results. Fig.9 shows that our method tends to preserve more details, though this may not be reflected by the metrics. Simply removing the skip connections from DIP causing blurring, similar to the decoder-only architectures." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 306, + 670, + 457, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 457, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 457, + 683 + ], + "type": "text", + "content": "4.3. Decoder-only Architectures" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We applied our strategy to ConvDecoder and tested it on all three datasets. Here we keep all 5 layers but remove" + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12414" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 91, + 287, + 186 + ], + "blocks": [ + { + "bbox": [ + 94, + 71, + 240, + 82 + ], + "lines": [ + { + "bbox": [ + 94, + 71, + 240, + 82 + ], + "spans": [ + { + "bbox": [ + 94, + 71, + 240, + 82 + ], + "type": "text", + "content": "Table 6: Application to ConvDecoder." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 91, + 287, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 91, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 91, + 287, + 186 + ], + "type": "table", + "html": "
Datasetsσ = 25σ = 50
BeforeAfterBeforeAfter
Set9PSNR28.5128.7425.0125.11
SSIM0.8540.8730.7690.784
Set12PSNR25.7926.9822.7223.01
SSIM0.7860.8540.7060.742
CBSD68PSNR25.1928.2924.3624.12
SSIM0.7930.8770.7670.768
# Params (Millions)0.89M0.06M~0.89M0.89M0.06M~0.89M
", + "image_path": "28aeaa1c0c8c8d047aab2abceef81fe065a4151db3b3301da5219c2f6da5866b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 198, + 287, + 285 + ], + "blocks": [ + { + "bbox": [ + 48, + 198, + 287, + 285 + ], + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 285 + ], + "type": "image", + "image_path": "f00b73fd16c073224f1444236f30f04e25bef6a6c304715618d2358f5c9ae981.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 192, + 545, + 214 + ], + "lines": [ + { + "bbox": [ + 305, + 192, + 545, + 214 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 545, + 214 + ], + "type": "text", + "content": "Figure 10: (a) Width critically influences a shallower network, while (b) it rarely has any impact on a sufficiently deep network." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 420 + ], + "type": "text", + "content": "a certain number of upsampling layers and scale the width accordingly for the images. These simple changes effectively alleviate the over-smoothing issue that often brought by ConvDecoder as shown in Fig.8 and improve the quantitative results as shown in Table 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 430, + 236, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 236, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 236, + 443 + ], + "type": "text", + "content": "4.4. More Analysis on Depth and Width" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 653 + ], + "type": "text", + "content": "In practice, we find it more efficient to first determine the depth, and then the width. This is because when the network is deep enough (strong attenuation due to upsampling), width becomes less influential, as evident in Fig.10 (b) and the small deviation in Fig.5 (b). However, to relax the need for early stopping at a higher noise level, one may prefer a shallower one, where width matters more (Fig.10 (a)). In this regard, we use the texture features of the images to predict the desired width. This makes intuitive sense as noise-fitting is associated with the amount of high frequency contents in the image, which is manifested in its texture. This is corroborated by the 0.86 Micro-average AUC score shown in Fig.11 (a). The \"optimal\" width labels we used to train the classifiers were obtained by experimenting with all three width choices on a 2-level hourglass architecture. These width labels are also applicable to decoder architectures as demonstrated by our experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "Although the width choices seem limited, we found the images robust to the choice of width to some extent, and simple averaging also works well for ambiguous cases (Appendix D). In fact, some images have multiple width labels. We included these cases in the released Texture-DIP dataset." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 73, + 425, + 173 + ], + "blocks": [ + { + "bbox": [ + 306, + 73, + 425, + 173 + ], + "lines": [ + { + "bbox": [ + 306, + 73, + 425, + 173 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 425, + 173 + ], + "type": "image", + "image_path": "5754ede4b38c3a18f2c7260c34f2cc2734a08603c2a0cb3c43a85a57863d98e7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 428, + 73, + 547, + 172 + ], + "blocks": [ + { + "bbox": [ + 428, + 73, + 547, + 172 + ], + "lines": [ + { + "bbox": [ + 428, + 73, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 428, + 73, + 547, + 172 + ], + "type": "image", + "image_path": "0aae8a7306d6ea4f82205c810e0384c9b0a54ebacacdf046dc32d0e52274ccdd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 322, + 228, + 462, + 348 + ], + "blocks": [ + { + "bbox": [ + 47, + 305, + 287, + 338 + ], + "lines": [ + { + "bbox": [ + 47, + 305, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 287, + 338 + ], + "type": "text", + "content": "Figure 9: Qualitative improvement on ConvDecoder simply by removing three upsampling layers in this case while preserving all the convolutional layers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 322, + 228, + 462, + 348 + ], + "lines": [ + { + "bbox": [ + 322, + 228, + 462, + 348 + ], + "spans": [ + { + "bbox": [ + 322, + 228, + 462, + 348 + ], + "type": "image", + "image_path": "7eeacf01310fd3804c86f5d5045c665fc81138200bfdda49fe2f9013737f218d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 470, + 227, + 529, + 348 + ], + "blocks": [ + { + "bbox": [ + 470, + 227, + 529, + 348 + ], + "lines": [ + { + "bbox": [ + 470, + 227, + 529, + 348 + ], + "spans": [ + { + "bbox": [ + 470, + 227, + 529, + 348 + ], + "type": "image", + "image_path": "023522865d6b37451d03b9139f8be488c6b04d5796bf1fd4903996ed75ae284a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 356, + 547, + 399 + ], + "lines": [ + { + "bbox": [ + 305, + 356, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 547, + 399 + ], + "type": "text", + "content": "Figure 11: (a) ROC curves with AUC scores for width classification on images from Set9, Set12 and CBSD68. 5-fold cross-validation is performed and repeated 10 times. (b) Overview of our Texture-DIP Dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 422, + 456, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 456, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 456, + 434 + ], + "type": "text", + "content": "4.5. What about Transformers?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 441, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 621 + ], + "type": "text", + "content": "Transformers [34] have become integral parts of modern deep neural network architectures. We experimented with Swin-UNet [2], an encoder-decoder pure transformer consisting of Swin Transformer blocks [21] and skip connections. Note that Swin-UNet relies on patch merging/expanding layers for " + }, + { + "bbox": [ + 304, + 441, + 545, + 621 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 304, + 441, + 545, + 621 + ], + "type": "text", + "content": " down-/upsampling without involving unlearnt upsampling. We replaced all the patch expanding layers with " + }, + { + "bbox": [ + 304, + 441, + 545, + 621 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 304, + 441, + 545, + 621 + ], + "type": "text", + "content": " bilinear upsampling operations and compared the performance of the modified network with the original. We did not replace the patch merging layers with the corresponding downsampling layers or strided convolutions because we did not observe any significant difference. The results presented in Fig. 12 are consistent with our findings on CNNs. For more implementation details, please refer to our code." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 633, + 471, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 471, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 471, + 645 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "We presented simple but effective solutions to the challenging DIP architectural design in the context of image denoising. Leveraging the spectral effects of upsampling and its relationships with other architectural components, we show that simple architectural changes yield highly-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12415" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 92, + 190, + 218 + ], + "blocks": [ + { + "bbox": [ + 120, + 72, + 167, + 91 + ], + "lines": [ + { + "bbox": [ + 120, + 72, + 167, + 91 + ], + "spans": [ + { + "bbox": [ + 120, + 72, + 167, + 91 + ], + "type": "text", + "content": "Noisy (20.38/0.465)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 96, + 92, + 190, + 218 + ], + "lines": [ + { + "bbox": [ + 96, + 92, + 190, + 218 + ], + "spans": [ + { + "bbox": [ + 96, + 92, + 190, + 218 + ], + "type": "image", + "image_path": "a20380dbac06f8ee4b129d45b992dd3d329d48d091e1a0ff19dcf072f4dd2fde.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 199, + 91, + 293, + 217 + ], + "blocks": [ + { + "bbox": [ + 223, + 72, + 268, + 90 + ], + "lines": [ + { + "bbox": [ + 223, + 72, + 268, + 90 + ], + "spans": [ + { + "bbox": [ + 223, + 72, + 268, + 90 + ], + "type": "text", + "content": "Swin-UNet (21.42/0.497)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 199, + 91, + 293, + 217 + ], + "lines": [ + { + "bbox": [ + 199, + 91, + 293, + 217 + ], + "spans": [ + { + "bbox": [ + 199, + 91, + 293, + 217 + ], + "type": "image", + "image_path": "0bf8fbc3e4aa8754ff75e318b7c07895151dff8208364b062f46e979f7198bc6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 301, + 91, + 394, + 217 + ], + "blocks": [ + { + "bbox": [ + 302, + 72, + 391, + 90 + ], + "lines": [ + { + "bbox": [ + 302, + 72, + 391, + 90 + ], + "spans": [ + { + "bbox": [ + 302, + 72, + 391, + 90 + ], + "type": "text", + "content": "Swin-UNet w. Bilinear Up (30.61/0.928)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 91, + 394, + 217 + ], + "lines": [ + { + "bbox": [ + 301, + 91, + 394, + 217 + ], + "spans": [ + { + "bbox": [ + 301, + 91, + 394, + 217 + ], + "type": "image", + "image_path": "765cb6455ab9c03b8d43dd72c56a6cfe4f45c5826ec026b10dfe48c4dfd89b56.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 225, + 497, + 237 + ], + "lines": [ + { + "bbox": [ + 94, + 225, + 497, + 237 + ], + "spans": [ + { + "bbox": [ + 94, + 225, + 497, + 237 + ], + "type": "text", + "content": "Figure 12: Visual comparisons on the transformer-based Swin-UNet [2] with and without bilinear upsampling." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 401, + 91, + 494, + 217 + ], + "blocks": [ + { + "bbox": [ + 424, + 72, + 473, + 91 + ], + "lines": [ + { + "bbox": [ + 424, + 72, + 473, + 91 + ], + "spans": [ + { + "bbox": [ + 424, + 72, + 473, + 91 + ], + "type": "text", + "content": "Ground Truth (PSNR/SSIM)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 401, + 91, + 494, + 217 + ], + "lines": [ + { + "bbox": [ + 401, + 91, + 494, + 217 + ], + "spans": [ + { + "bbox": [ + 401, + 91, + 494, + 217 + ], + "type": "image", + "image_path": "4958944630f6e3e4c6bfda76d1a568a1d2dbeb501d1f9164b9ce574cf566ecea.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 258, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 354 + ], + "type": "text", + "content": "effective under-parameterized networks that could surpass the larger counterparts and does not critically rely on early-stopping. Although this work focuses on denoising, insights learned about upsampling can be employed in the future to understand the relationship between architectural characteristics and other image restoration tasks. We hope our study could encourage efficient architectural design for DIP and image synthesis in general." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 364, + 149, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 364, + 149, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 149, + 376 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 383, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 287, + 418 + ], + "type": "text", + "content": "This work was supported in part by the United States National Institutes of Health (NIH) under Grant CA266702 and Grant EB008374." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 429, + 106, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 429, + 106, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 429, + 106, + 441 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 52, + 449, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 52, + 449, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 449, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 52, + 449, + 287, + 503 + ], + "type": "text", + "content": "[1] Metin Ersin Arican, Ozgur Kara, Gustav Bredell, and Ender Konukoglu. Isnas-dip: Image-specific neural architecture search for deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1960-1968, 2022. 1, 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 505, + 287, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 505, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 53, + 505, + 287, + 559 + ], + "type": "text", + "content": "[2] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 8, 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 559, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 287, + 591 + ], + "type": "text", + "content": "[3] Prithvijit Chakrabarty and Subhransu Maji. The spectral bias of the deep image prior. arXiv preprint arXiv:1912.08905, 2019. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 593, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 593, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 593, + 287, + 635 + ], + "type": "text", + "content": "[4] Stanley H Chan, Xiran Wang, and Omar A Elgendy. Plug-and-play admm for image restoration: Fixed-point convergence and applications. IEEE Transactions on Computational Imaging, 3(1):84-98, 2016. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 53, + 637, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 637, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 53, + 637, + 288, + 691 + ], + "type": "text", + "content": "[5] Keshigeyan Chandrasegaran, Ngoc-Trung Tran, and Ngai-Man Cheung. A closer look at fourier spectrum discrepancies for cnn-generated images detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7200-7209, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 53, + 692, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 287, + 714 + ], + "type": "text", + "content": "[6] Yun-Chun Chen, Chen Gao, Esther Robb, and Jia-Bin Huang. Nas-dip: Learning deep image prior with neural" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 259, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 259, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 259, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 326, + 259, + 545, + 281 + ], + "type": "text", + "content": "architecture search. In European Conference on Computer Vision, pages 442-459. Springer, 2020. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 282, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 282, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 312, + 282, + 545, + 335 + ], + "type": "text", + "content": "[7] Zezhou Cheng, Matheus Gadelha, Subhransu Maji, and Daniel Sheldon. A bayesian perspective on the deep image prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5451, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 337, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 337, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 312, + 337, + 545, + 358 + ], + "type": "text", + "content": "[8] Corinna Cortes and Vladimir Vapnik. Support-vector networks. Machine learning, 20:273-297, 1995. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 360, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 360, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 360, + 545, + 403 + ], + "type": "text", + "content": "[9] Kostadin Dabov, Alessandro Foi, and Karen Egiazarian. Video denoising by sparse 3d transform-domain collaborative filtering. In 2007 15th European Signal Processing Conference, pages 145-149. IEEE, 2007. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 404, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 446 + ], + "type": "text", + "content": "[10] Mohammad Zalbagi Darestani and Reinhard Heckel. Accelerated mri with un-trained neural networks. IEEE Transactions on Computational Imaging, 7:724-733, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[11] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7890–7899, 2020. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "text", + "content": "[12] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 559, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 612 + ], + "type": "text", + "content": "[13] Yosef Gandelsman, Assaf Shocher, and Michal Irani. \"double-dip\": unsupervised image decomposition via coupled deep-image-priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11026-11035, 2019. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "type": "text", + "content": "[14] Robert M Haralick, Karthikeyan Shanmugam, and Its' Hak Dinstein. Textural features for image classification. IEEE Transactions on systems, man, and cybernetics, (6):610-621, 1973. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 691 + ], + "type": "text", + "content": "[15] Reinhard Heckel and Paul Hand. Deep decoder: Concise image representations from untrained non-convolutional networks. arXiv preprint arXiv:1810.03982, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[16] Reinhard Heckel and Mahdi Soltanolkotabi. Denoising and regularization via exploiting the structural bias of convolu" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12416" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "text", + "content": "tional generators. arXiv preprint arXiv:1910.14634, 2019. 1,5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "text", + "content": "[17] Kary Ho, Andrew Gilbert, Hailin Jin, and John Collomosse. Neural architecture search for deep image prior. Computers & graphics, 98:188-196, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "text", + "content": "[18] Yeonsik Jo, Se Young Chun, and Jonghyun Choi. Rethinking deep image prior for denoising. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5087-5096, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 216 + ], + "type": "text", + "content": "[19] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "text", + "content": "[20] Jiaming Liu, Yu Sun, Xiaojian Xu, and Ulugbek S Kamilov. Image restoration using total variation regularized deep image prior. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7715-7719. IEEE, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 327 + ], + "type": "text", + "content": "[21] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 360 + ], + "type": "text", + "content": "[22] Wei-Yin Loh. Classification and regression trees. Wiley interdisciplinary reviews: data mining and knowledge discovery, 1(1):14-23, 2011. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 393 + ], + "type": "text", + "content": "[23] Florian Luisier, Thierry Blu, and Michael Unser. Image denoising in mixed poisson-gaussian noise. IEEE Transactions on image processing, 20(3):696-708, 2010. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 437 + ], + "type": "text", + "content": "[24] Gary Mataev, Peyman Milanfar, and Michael Elad. Deepred: Deep image prior powered by red. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 437, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 469 + ], + "type": "text", + "content": "[25] Augustus Odena, Vincent Dumoulin, and Chris Olah. Deconvolution and checkerboard artifacts. Distill, 1(10):e3, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "text", + "content": "[26] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 557 + ], + "type": "text", + "content": "[27] Stefan Roth and Michael J Black. Fields of experts. International Journal of Computer Vision, 82(2):205-229, 2009. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 287, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 592 + ], + "type": "text", + "content": "[28] Katja Schwarz, Yiyi Liao, and Andreas Geiger. On the frequency bias of generative models. Advances in Neural Information Processing Systems, 34:18126-18136, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[29] Zenglin Shi, Pascal Mettes, Subhransu Maji, and Cees GM Snoek. On measuring and controlling the spectral bias of the deep image prior. International Journal of Computer Vision, 130(4):885-908, 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "text", + "content": "[30] Eero P Simoncelli and Bruno A Olshausen. Natural image statistics and neural representation. Annual review of neuroscience, 24(1):1193-1216, 2001. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[31] Shakarim Soltanayev and Se Young Chun. Training deep learning based denoisers without ground truth data. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 354 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[32] Charles M Stein. Estimation of the mean of a multivariate normal distribution. The annals of Statistics, pages 1135-1151, 1981. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[33] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9446-9454, 2018. 1, 3, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "type": "text", + "content": "[35] Hengkang Wang, Taihui Li, Zhong Zhuang, Tiancong Chen, Hengyue Liang, and Ju Sun. Early stopping for deep image prior. arXiv preprint arXiv:2112.06074, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 263 + ], + "type": "text", + "content": "[36] Jun Xu, Hui Li, Zhetong Liang, David Zhang, and Lei Zhang. Real-world noisy image denoising: A new benchmark. arXiv preprint arXiv:1804.02603, 2018. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[37] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 26(7):3142-3155, 2017. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 354 + ], + "type": "text", + "content": "[38] Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3929-3938, 2017. 1" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12417" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/5b230860-aa61-4f33-9fe7-2407ba62f713_content_list.json b/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/5b230860-aa61-4f33-9fe7-2407ba62f713_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..06228a02b15c7245717d66456fac0242a5e2aab4 --- /dev/null +++ b/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/5b230860-aa61-4f33-9fe7-2407ba62f713_content_list.json @@ -0,0 +1,1452 @@ +[ + { + "type": "text", + "text": "The effectiveness of MAE pre-pretraining for billion-scale pretraining", + "text_level": 1, + "bbox": [ + 135, + 130, + 834, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mannat Singh*,† Quentin Duval* Kalyan Vasudev Alwala* Haoqi Fan Vaibhav Aggarwal Aaron Adcock Armand Joulin Piotr Dólár Christoph Feichtenhofer Ross Girshick Rohit Girdhar Ishan Misra Meta AI", + "bbox": [ + 228, + 181, + 738, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper revisits the standard pretrain-then-finetune paradigm used in computer vision for visual recognition tasks. Typically, state-of-the-art foundation models are pretrained using large scale (weakly) supervised datasets with billions of images. We introduce an additional pretraining stage that is simple and uses the self-supervised MAE technique to initialize the model. While MAE has only been shown to scale with the size of models, we find that it scales with the size of the training dataset as well. Thus, our MAE-based pre-training scales with both model and data size making it applicable for training foundation models. Pre-training consistently improves both the model convergence and the downstream transfer performance across a range of model scales (millions to billions of parameters), and dataset sizes (millions to billions of images). We measure the effectiveness of pre-training on 10 different visual recognition tasks spanning image classification, video recognition, object detection, low-shot classification and zero-shot recognition. Our largest model achieves new state-of-the-art results on iNaturalist-18 $(91.3\\%)$ , 1-shot ImageNet-1k $(62.1\\%)$ , and zero-shot transfer on Food-101 $(96.2\\%)$ . Our study reveals that model initialization plays a significant role, even for web-scale pretraining with billions of images.", + "bbox": [ + 75, + 318, + 473, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 209, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The pretrain-then-finetune paradigm in visual recognition has enabled high performance visual recognition models across a range of tasks such as image classification [52, 59, 70], video action recognition [25, 27, 28], object detection [11, 89], 3D etc. Typically, pretraining consists of training a model using a pretraining task on large scale data. The resulting pretrained models learn general purpose visual representations that can be used for a range of target tasks, often with limited labeled data, by transfer learning.", + "bbox": [ + 75, + 722, + 470, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we show that an initial stage of pre", + "bbox": [ + 94, + 859, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76441380b84add5d1f18b755d1ff430c2cacc528d03599debfa5d6b714e0409f.jpg", + "image_caption": [ + "Figure 1: MAE pre-pretraining improves performance. Transfer performance of a ViT-L architecture trained with self-supervised pretraining (MAE), weakly supervised pretraining on billions of images (WSP), and our pre-pretraining $(\\mathrm{MAE} \\rightarrow \\mathrm{WSP})$ that initializes the model with MAE and then pretrains with WSP. Pre-pretraining consistently improves performance." + ], + "image_footnote": [], + "bbox": [ + 504, + 285, + 888, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pretraining before the standard pretraining task can improve vision models across a variety of different tasks. Our method combines two common pretraining tasks in vision: (1) weakly supervised pretraining that uses weak, often noisy, signals such as text or image hashtags as supervision, and (2) self-supervised pretraining that only uses the data without additional supervision. Both forms of pretraining start training with a randomly initialized model and have proven effective at learning general purpose vision models. While there have been attempts to combine both these forms of pretraining [54, 69], they are typically used independently in the pretrain-then-finetune two stage paradigm [23, 70, 84].", + "bbox": [ + 496, + 642, + 893, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work we explore the combination of self- and weakly-supervised learning in a simple pre-pretraining framework, as follows. We first begin with the Masked Autoencoder (MAE) [33] self-supervised learning technique to pre-pretrain vision models without using any labels. After", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal technical contribution. †Project Lead.", + "bbox": [ + 99, + 886, + 344, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5484", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "initializing from the pre-pretrained model, we use standard weakly supervised pretraining on billions of images with noisy labels. We perform a large-scale empirical study to measure the effectiveness of pre-pretraining on 10 different visual recognition tasks spanning image classification, video recognition, object detection, low-shot classification and zero-shot recognition. Our study reveals that pre-pretrain initialization improves the performance for the weakly supervised models, and this improvement holds even at billion scale weakly labeled data, and across vision tasks (Figure 1). It also improves the model convergence during pretraining, leading to an efficient way of training large scale vision models. Pre-pretrain further enjoys the computational efficiency of the MAE approach, making it simple and scalable. Finally, we show that by using pre-pretraining, both self-supervised learning and weakly supervised learning can be combined for improved model performance for billion-scale data.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pre-pretrain is related to 'intermediate finetuning' [5, 50] which introduces a stage after pretraining to better align the pretrained features with the downstream task using labeled data. In contrast, pre-pretrain serves as a better way to initialize a model before pretraining. Since we leverage MAE for pre-pretraining, we do not need additional information or labels for this stage and can re-use the pretraining data. This makes pre-pretrain convenient and simple to use with existing pretraining datasets.", + "bbox": [ + 75, + 347, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our study on large-scale pre-pretraining reveals that model initialization plays a significant role, even for web-scale pretraining, and pre-pretraining is a simple and promising technique in that direction. In particular, we show that (i) MAE not only scales with model size as shown in [33], but also with the size of the training data (Figure 2). (ii) Pre-pretraining improves both the model convergence and the final downstream performance for different sized models (millions to billions of parameters) trained on different sized datasets (millions to billions of labels). (iii) Using pre-pretraining combines the benefits of both self-supervised learning and large scale weakly-supervised learning, and our models achieve excellent performance on a variety of different visual recognition tasks (Figure 1). Most prominently, our model sets new state-of-the-art results on iNaturalist-18 image classification $(91.3\\%)$ , 1-shot ImageNet-1k image classification $(62.1\\%)$ , and zero-shot transfer on Food-101 $(96.2\\%)$ .", + "bbox": [ + 75, + 484, + 472, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 768, + 218, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised pretraining of transferrable representations on large labeled datasets [21, 43, 63] and employing them for downstream recognition tasks, has emerged as a powerful approach in computer vision. It has spurred rapid progress on various tasks including image classification [22, 57, 60], object detection/segmentation [29, 62], image captioning [45, 81] and video action recognition [13, 24, 68]. While useful,", + "bbox": [ + 75, + 794, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "such representations are often limited by the scale and diversity of the supervision in the pretraining datasets. Hence, recent work has probed the effectiveness, robustness, and fairness of these representations [1, 20, 40, 44, 61, 66, 74]. Self-supervised pretraining is a promising alternative to learn these representation without relying on large well-labeled datasets. Initial works focused on reconstructions methods [77] before moving to other pretraining tasks such as solving jigsaw puzzles [55], constractive learning [15, 34] or joint embedding approaches [3, 4, 11, 12, 31, 89]. With the advent of Vision Transformers [23], approaches based on reconstructions such as [5, 33, 79] got renewed interest for their simplicity and state of the art performance. Of particular interest to us is MAE [33] for its state of the art performance on many transfer tasks [25, 28, 33, 46, 75] and its computational efficiency. Given the lack of supervision during pretraining, these representations often require significant finetuning to align to downstream tasks.", + "bbox": [ + 496, + 90, + 893, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weakly supervised pretraining (WSP) is a middle-ground between supervised and self-supervised pretraining. Instead of ignoring annotations completely as in self-supervised pretraining, or requiring exhaustive labels as in supervised pretraining, WSP relies on the large quantity of \"free\" annotation available on the internet. These annotations occur as image-text pairs [59, 65], where the text can additionally be processed to produce pseudo labels. Of particular interest to us is the latter, i.e. approaches which leverage multi-label classification on noisy labels [27, 52, 70, 84] which have shown state of the art fine-tuning performance, and at the same time can be adapted using image-text data to gain zero-shot capabilities [85]. In this work, we explore WSP in conjunction with self-supervised pre-pretraining, and show faster convergence and stronger performance.", + "bbox": [ + 496, + 363, + 895, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Setup", + "text_level": 1, + "bbox": [ + 500, + 603, + 573, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our goal is to empirically study the effectiveness of self-supervised pre-pretraining as a precursor to billion scale weakly supervised pretraining for representation learning. Given the simplicity and efficiency of Masked AutoEncoding (MAE) [33], we leverage it as the self-supervised prepretraining approach. Our study shows that MAE scales with the size of the pretraining dataset and model size, and combining it with weak supervision improves large scale vision models. Additionally, such a combination leads to faster convergence and is a simple, scalable way to learn visual representations at scale. We describe our setup and the approaches in detail next.", + "bbox": [ + 496, + 628, + 893, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Architecture. We use the Vision Transformer (ViT) [23] architecture as the visual encoder for all our experiments. ViTs employ minimal vision-specific inductive biases combined with the standard transformer architecture [76], and yet have emerged as an architecture of choice for a wide variety of visual and multimodal recognition tasks [2, 28, 84]. We train", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5485", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ViT models at various scales in terms of number of parameters, including ViT-B (86M), ViT-L (307M), and ViT-H (632M). We also train on larger 1.9B and 6.5B parameter ViT models, which we call ViT-2B and ViT-6.5B, respectively (Appendix Table 8). As is common practice [23, 84], we train models of sizes ViT-B, ViT-L with a patch size of 16 and larger models with a patch size of 14. We pretrain with a $224 \\times 224$ resolution for all models.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-pretraining (MAE) [33] learns visual representations from image datasets without using any labels. We choose this approach as it is simple to implement and scales very effectively with large ViT model sizes due to patch dropping as described next. MAE randomly masks $75\\%$ of an image and trains the model to reconstruct the masked input image by minimizing the pixel reconstruction error. The target pixel values for a given patch are normalized by the mean and standard deviation of all pixels in it. Coupled with the ViT architecture, MAE can be trained by only processing the $25\\%$ unmasked image patches. A separate, smaller, decoder is then used to reconstruct the missing part of the input. This asymmetrical design makes training the encoder extremely efficient, allowing for scaling visual encoder sizes.", + "bbox": [ + 75, + 210, + 472, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weakly-supervised pretraining (WSP) leverages images with associated 'weak' supervision for training models. In particular, we focus on internet images and use their associated text information as supervision. We convert the text into a discrete set of labels, specifically leveraging hash-tag information [27, 52, 70]. We then use a multi-label classification loss to train models. We refer to this method as WSP.", + "bbox": [ + 75, + 422, + 472, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbf{MAE} \\rightarrow \\mathbf{WSP}$ first trains the encoder using the MAE self-supervised method using only the images. This pretraining stage initializes the model while simultaneously being computationally efficient because of the masking used in MAE. In the second stage, we pretrain the encoder using both the image and associated weak supervision. This combination outperforms using either strategy in isolation, i.e., an MAE model or a weakly supervised model trained from scratch.", + "bbox": [ + 75, + 527, + 472, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 676, + 209, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We empirically evaluate and analyze large scale MAE pre-pretraining using Instagram data on a variety of different visual recognition tasks. We describe the datasets used for pretraining and evaluation, followed by analysis of the pretraining design decisions, and finally the downstream transfer evaluation of our learned representation.", + "bbox": [ + 75, + 702, + 470, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Datasets and training details", + "text_level": 1, + "bbox": [ + 76, + 801, + 333, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pretraining dataset. We use Instagram-3B (IG-3B) a billion-scale multi-label dataset sourced from Instagram (IG). This multi-label dataset contains 28K classes and 3B unique images, resampled to 5B total images, and was produced by running the dataset generation pipeline from SWAG [70]", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/247dc7231748062648a64500245223397462614fb96974a2dc7ee8d8af713715.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetTask#cls#train#val
ImageNet-1k (IN1k) [64]Image cls.10001M50K
iNaturalist-18 (iNat18) [36]Fine-grained cls.8142437K24K
ImageNetv2 (INv2) [61]Image cls.1000-10K
ImageNet-ReaL (IN-ReaL) [7]Image cls.1000-50K
ObjectNet (ON) [6]Image cls.113-19K
Food-101 (F-101) [9]Image cls.101N/A25K
COCO [49]Obj. det.80118K5K
LVIS [32]Obj. det.1K100K20K
Kinetics-400 (K400) [43]Action cls.400220K20K
Something Something-v2 (SSv2) [30]Action cls.174169K25K
", + "bbox": [ + 506, + 88, + 890, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: Evaluation datasets used to evaluate MAE $\\rightarrow$ WSP on image classification, object detection, and video action recognition tasks. The table reports the task, number of classes (#cls), number of training samples (#train), and number of validation samples (#val) for each dataset.", + "bbox": [ + 496, + 233, + 893, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "without modification. Compared to [70], our version of the dataset has $16\\%$ fewer images (3.0B vs. 3.6B), but we were able to reproduce the results from [70] with our version. We obtain labels using an automated process wherein we first obtain hashtags from the associated image captions, and then map the hashtags to WordNet synsets following [70]. After this processing, we get the weakly labeled IG dataset that contains images and their associated labels.", + "bbox": [ + 496, + 334, + 893, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation datasets. We evaluate MAE $\\rightarrow$ WSP on a variety of different downstream visual recognition tasks. To evaluate our model on image classification, we use the standard ImageNet-1k [64] (IN1k) dataset, and also the long-tailed and fine-grained iNaturalist-18 [36] (iNat18) dataset. For object detection and segmentation, we use the popular COCO [49] dataset, and also LVIS [32], a large vocabulary dataset for long tailed object recognition. We evaluate video classification performance using two popular action recognition datasets, Kinetics-400 [43] (K400) and Something Something-v2 [30] (SSv2). For zero-shot transfer, we evaluate on IN1k and Food-101 [9] (F-101). We also evaluate the robustness of our models on test sets which overlap with IN1k classes, specifically ImageNetv2 [61] (INv2), ImageNet-ReaL [7] (IN-ReaL), and ObjectNet [6] (ON). Please see Table 1 for more details.", + "bbox": [ + 496, + 455, + 895, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MAE pretraining details. We follow [33] to train MAE models on IG-3B without using any labels. We mask $75\\%$ of the image for this training and train the model for 1 epoch over the dataset. We follow the same hyperparameters used in [33] for pretraining on IN1k.", + "bbox": [ + 496, + 699, + 893, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Supervised pretraining details. We train with a supervised cross-entropy loss on IG-3B using the hashtags as labels. This model is trained by default with random weight initialization and we use the training hyperparameters from [70].", + "bbox": [ + 496, + 777, + 893, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using pre-pretraining. When using pre-pretraining, we first train a model from scratch using MAE on the IG dataset. We then use the weights of the MAE encoder and perform supervised pretraining using the cross-entropy loss as described", + "bbox": [ + 496, + 840, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5486", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1da6f92a9a9f6a1b08fe1b5ae2983335a664f0cd1ecb53082b91d3f341f15060.jpg", + "image_caption": [ + "Figure 2: Scaling MAE with model and dataset size. We plot MAE's performance when pretrained on ImageNet-1k or Instagram-3B and finetuned on downstream tasks. MAE scales to billion parameters sized models using just IN1k pretraining. Larger models show improved scaling behavior when pretrained with the much larger IG-3B dataset. MAE pretrained on IN1k data point is missing for the 2 billion model as training at that scale was unstable on both COCO and LVIS datasets. Tabulated results in Table 19." + ], + "image_footnote": [], + "bbox": [ + 81, + 92, + 277, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0ad97bb96c50fc9fd5f9ba5ed3b5cc6cce2192c7030235da98e8835b342be4e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 90, + 475, + 279 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c14ccc52e24c3ea920dcf17c4e45039b75658860e5ab962f7efccf21eb1ded12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 90, + 681, + 280 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bd3d13e7ae81b85391bd5023aad7f6e4c0e97f068a2ff5a12a19051fa123684e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 90, + 880, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "above. We reuse the same hyperparameters and training details as [70], i.e. there is no hyperparameter search needed for MAE $\\rightarrow$ WSP, and we train for 1 epoch on IG-3B.", + "bbox": [ + 75, + 382, + 470, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Zero-shot training and evaluation details. To impart zero shot understanding capabilities to our models, we use the LiT approach from [85]. For LiT, we use the original (image, caption) pairs from the IG-3B dataset. We freeze the image encoder, and train a text encoder to encode the image captions and match the text embeddings to the associated image embedding using a CLIP loss [59]. We train the text encoder for 1 epoch. For evaluation, we follow [59] - we use the text encoder to compute embeddings from the templated text descriptions of classes and use the cosine similarity of the image and text embeddings as the classification score.", + "bbox": [ + 75, + 428, + 468, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For full training details and hyperparameters, please refer to Appendix A.", + "bbox": [ + 76, + 594, + 470, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Scaling MAE pretraining to large data", + "text_level": 1, + "bbox": [ + 76, + 633, + 410, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since our pre-pretraining uses MAE in the very first stage, we first study how MAE behaves on the large scale IG-3B dataset. We compare the performance of MAE pretraining on the large scale IG-3B with the original MAE [33] models trained on IN1k for 1600 epochs. We train models of varying sizes, from ViT-B to ViT-H as in [33]. To test the scaling behavior further, we also train MAE on the ViT-2B model with 2B parameters. We measure the performance of the resulting models in Figure 2 on four different vision tasks.", + "bbox": [ + 75, + 657, + 468, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that using the IG-3B data provides consistent gains over IN1k for all vision tasks, and the gain increases for larger models. These experiments show that MAE scales with the size of the pretraining dataset, and benefits from using billions of images from IG-3B. He et al. [33]'s findings were limited to the fact that MAE scales with the size of the model, and thus our findings on MAE scaling with the size", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the pretraining data are complementary to theirs.", + "bbox": [ + 498, + 382, + 839, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our ViT-2B model pretrained on IG-3B improves upon the best results from [33] on image classification, attaining $88.4\\%$ on IN1k $(+0.6\\%)$ and $89.3\\%$ on iNat18 $(+2.5\\%)$ at $518\\times 518$ resolution. The gains on detection are equally encouraging, with our ViT-2B reaching $53.6\\mathrm{AP}^{\\mathrm{box}}$ on LVIS $(+2.1$ over [33]) and $59.8\\mathrm{AP}^{\\mathrm{box}}$ on COCO $(+1.2$ over [33]). Lastly, we highlight the simplicity of our setup, since we use the same hyperparameters as [33] to train MAE on IG-3B, and also for the largest ViT-2B model, without encountering any training instabilities or needing any extra tweaks.", + "bbox": [ + 496, + 397, + 893, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. MAE pre-pretraining", + "text_level": 1, + "bbox": [ + 500, + 559, + 702, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the promising aspects of MAE as a pretraining approach from § 4.2, specifically that MAE (i) trains with larger models and datasets without needing any tuning (ii) shows gains when scaling model and / or dataset size (iii) is efficient to train, we investigate it as a pre-pretraining approach for supervised pretraining (WSP).", + "bbox": [ + 496, + 582, + 893, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 1 shows the performance of a ViT-L with MAE pretraining, supervised pretraining (WSP), or MAE pre-pretraining followed by supervised pretraining $(\\mathrm{MAE}\\rightarrow \\mathrm{WSP})$ . We see that MAE and WSP have different strengths. MAE has strong performance for object detection, and full finetuned image classification. However, MAE underperforms on tasks where the model is not finetuned, such as linear classifiers, zero-shot, or low-shot classification - situations where WSP performs better. For these evaluations MAE lags behind WSP by more than 10 points, which is why the results for MAE are not visible in Figure 1. For video classification, MAE performs significantly better than WSP on SSv2, but lags behind it on K400.", + "bbox": [ + 496, + 674, + 893, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathrm{MAE} \\rightarrow \\mathrm{WSP}$ outperforms either of MAE or WSP pretraining on most evaluations, across image classification,", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5487", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3c7ffcdf7c176b6b90ec756754dbef4b37f2f3a7afb4c1f0561ad6de8cc6c36c.jpg", + "image_caption": [ + "IN1k linear probe (accuracy)", + "Figure 3: MAE pre-pretraining scales with model size. Across model sizes, MAE $\\rightarrow$ WSP outperforms a WSP only model, and shows strong scaling behavior. Most notably, a 2B MAE $\\rightarrow$ WSP model outperforms a 6.5B WSP model." + ], + "image_footnote": [], + "bbox": [ + 81, + 113, + 321, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b4f5842507f9fd3e2f729967dd03b3b95d626d7eb48155e2003ce9d9b0f851a4.jpg", + "image_caption": [ + "IN1k linear probe (accuracy)", + "Figure 4: Varying the number of pre-pretraining epochs used to initialize the model for WSP pretraining. Pre-pretraining leads to improved convergence, providing higher performance using fewer number of WSP pretraining epochs." + ], + "image_footnote": [], + "bbox": [ + 357, + 111, + 619, + 294 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7738b885c32d02d3ea14e7e4b1fcdc50f6d39747622305acaa76a6eff8d54eb2.jpg", + "image_caption": [ + "IN1k linear probe (accuracy)", + "Figure 5: MAE $\\rightarrow$ WSP is more FLOPs efficient than WSP. Across a wide range of training FLOP profiles for a ViT-B computed by varying WSP (and MAE) epochs, MAE $\\rightarrow$ WSP outperforms a baseline WSP only model." + ], + "image_footnote": [], + "bbox": [ + 656, + 113, + 893, + 295 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7f30545ce98bfc625cdb27d47fba29ab20bbae10cddd6d0a9b1b39089cfddb68.jpg", + "image_caption": [ + "Figure 6: MAE initialization for medium scale data. Results for a ViT-L trained on IN21k with MAE, WSP, and MAE $\\rightarrow$ WSP. MAE pre-pretraining improves WSP results by a wide margin." + ], + "image_footnote": [], + "bbox": [ + 81, + 411, + 465, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "video recognition, zero-shot evaluation, object detection, etc. Given that all baselines are trained at billion-scale, these results show that MAE $\\rightarrow$ WSP is a simple yet promising strategy to improve performance while requiring no extra data or tuning. Next, we ablate the key aspects of MAE $\\rightarrow$ WSP.", + "bbox": [ + 75, + 729, + 470, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Effect of model size. In Figure 3 we study the effect of prepretraining compared to random initialization for different model sizes. After initialization, all models are pretrained using WSP on the IG-3B dataset, and we measure the transfer performance on IN1k using linear probing. We observe that MAE pre-pretraining gives consistent gains over the", + "bbox": [ + 75, + 809, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "WSP baseline across all model sizes, ranging from 86M to 2B parameters. The gains over the WSP baseline increase for larger model sizes showing that pre-pretraining shows promising scaling behavior with model sizes. Notably, a 2B MAE $\\rightarrow$ WSP model outperforms a larger 6.5B WSP model. Number of pre-pretraining epochs. We vary the number of MAE pre-pretraining and the number of WSP pretraining epochs to understand their effect on the final recognition performance. We study this in Figure 4.", + "bbox": [ + 496, + 412, + 893, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-pretraining improves results over the standard pretraining (random initialization, w/o pre-pretraining), and provides large gains with fewer WSP pretraining epochs. Pre-pretraining also leads to faster convergence since even a small amount of pre-pretraining for 0.1 epochs provides improvements. Increasing the epochs of pre-pretraining provide a larger improvement, and the gains saturate at 1 epoch of pre-pretraining. Finally, pre-pretraining's gains do not diminish even after 4 epochs of WSP (20 billion samples) showing the value of pre-pretraining at scale. We also note that these gains are independent of the evaluation protocol, and we observed them with full finetuning on IN1k as well.", + "bbox": [ + 496, + 551, + 895, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training efficiency. Figure 5 shows a comparison between WSP and MAE $\\rightarrow$ WSP when comparing training FLOPs. For the same training FLOPs, MAE $\\rightarrow$ WSP achieves better transfer performance compared to WSP, and is up to $2\\times$ more efficient. Pre-pretraining's training efficiency holds over a large $10\\times$ compute window.", + "bbox": [ + 496, + 733, + 893, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different datasets for pre-pretraining. We also evaluate the performance of MAE $\\rightarrow$ WSP when pre-pretraining MAE on the much smaller ImageNet-1k dataset (Appendix Table 21) and find that pre-pretraining remains just as effective. This allows reusing pretrained MAE models in practice.", + "bbox": [ + 496, + 824, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5488", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/04cdf05af39b8c9623dcfffeb60e3018056e5de2b31e75d30829fdd6f5d65fed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDatasetArch.Res.ImageNet-1kiNat18
IN1kINv2ReaLON
MAE [33]IN1kViT-H44887.8---86.8
SWAG [70]IG-3.6BViT-H51888.681.190.569.586.0
DINOv2 [56]LVD-142MViT-g44888.9----
Florence [82]FLD-900MCoSwin-H51290.1----
ViT [23]JFT-300MViT-H51888.6-90.7--
Scale-ViT [84]JFT-3BViT-L38488.580.490.4--
Scale-ViT [84]JFT-3BViT-G51890.583.390.870.5-
SwinV2 [50]IN-ext-70MSwinV2-G64090.284.0---
CoCa [81]JFT-3B + ALIGNN-1.8BCoCa-2B57691.0----
MAE→WSPIG-3BViT-H51889.382.390.872.690.5
MAE→WSPIG-3BViT-2B51889.783.090.975.891.3
", + "bbox": [ + 81, + 103, + 540, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different datasets for pre-pretraining and pretraining. We investigate the effect of the dataset used in pre-pretraining and pretraining by using IN21k [21] for all methods, including MAE and MAE $\\rightarrow$ WSP. Compared to IG-3B, IN21k is more curated, smaller (14M images), and has cleaner labels (21K classes), where each image is labeled with one class from the WordNet synsets [53]. For evaluating zero shot performance, we use the PMD [69] dataset for LiT training. For full details about the hyperparameters, refer to Appendix A.", + "bbox": [ + 75, + 395, + 470, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 6 compares the performance of MAE, WSP and $\\mathrm{MAE} \\rightarrow \\mathrm{WSP}$ when pretrained on this IN21k dataset. We notice a similar trend as when pretraining on IG-3B where $\\mathrm{MAE} \\rightarrow \\mathrm{WSP}$ outperforms both MAE and WSP. This shows that MAE pre-pretraining works with datasets of different scales and distributions.", + "bbox": [ + 75, + 532, + 468, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Transfer Evaluation", + "text_level": 1, + "bbox": [ + 76, + 633, + 267, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare with state-of-the-art research on image and video classification, detection and segmentation, low-shot image classification, zero shot transfer, robustness analysis. ImageNet-1k image classification. Table 2 shows the performance of different methods on IN1k. MAE $\\rightarrow$ WSP gets the best performance for a ViT-H sized model (89.3%). As we report in the appendix, MAE $\\rightarrow$ WSP also achieves the best performance for ViT-L (88.8%). Recent methods such as Scale-ViT [84] are better on IN1k and we hypothesize that this gap stems mainly from the differences in the pretraining datasets (IG-3B vs. JFT-3B). We also compute linear performance using frozen features on IN1k at 224px resolution. Our models produce strong features which outperform other methods with WSP objectives. They also surpass the performance of the self-supervised DINOv2 [56] model optimized to produce strong frozen representations.", + "bbox": [ + 75, + 657, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6f7cecc212925ff8e28d6b3acd37d80c67f3ec8b526a99b6e5649ed148b1592d.jpg", + "table_caption": [ + "Table 2: Image classification results. We report finetuning results on IN1k and iNat18 and note the pretraining dataset, architecture and finetuning resolution. We also evaluate the performance of our IN1k finetuned models on multiple test sets to measure robustness. Our models are robust and also push the state-of-the-art on the challenging fine-grained and long-tailed iNat18 dataset." + ], + "table_footnote": [], + "table_body": "
MethodDatasetArch.Res.K400SSv2
Florence [82]FLD-900MCoSwin-H38486.5-
SwinV2 [50]IN-ext-70MSwinV2-G38486.8-
CoCa [81]JFT-3B + ALIGNN-1.8BCoCa-2B57688.9-
Results with models pretrained on videos
MaskFeat [78]K400MViT-L22484.3-
MAE [25, 75]K400ViT-L22485.274.0
OmniMAE [28]IN1k + SSv2ViT-L22484.074.2
MAE [25, 75]K400ViT-H22486.6-
MAE→WSPIG-3BViT-L22486.074.4
", + "bbox": [ + 552, + 126, + 898, + 279 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/99149ae5d25b8867f8b25e12fb2b07b93427bf268fe228dff477e470c55f747e.jpg", + "table_caption": [ + "Table 3: Video classification results on Kinetics-400 and Something Something-v2. Our models generalize well to video action recognition tasks, despite not seeing any videos during pretraining." + ], + "table_footnote": [], + "table_body": "
MethodDatasetArch.IN1k Linear
SWAG [70]IG-3.6BViT-H85.8
OpenCLIP [41]LAION-2BViT-G86.2
DINOv2 [56]LVD-142MViT-g86.5
MAE→WSPIG-3BViT-H87.0
MAE→WSPIG-3BViT-2B88.1
", + "bbox": [ + 542, + 391, + 849, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Robustness for image classification. We evaluate the robustness of our models finetuned on IN1k on additional test sets in Table 2 and notice that, despite MAE→WSP being $\\sim 1\\%$ behind Scale-ViT on IN1k, it is significantly more robust and generalizes better on these additional test sets – MAE→WSP gets the highest reported performance on ImageNet-ReaL and ObjectNet for IN1k finetuned models.", + "bbox": [ + 496, + 477, + 893, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalization in image classification. We evaluate the generalization of our model on additional fine-grained image classification using iNaturalist-18. iNat18 is a challenging long-tailed and fine-grained dataset with images of multiple species of visually similar plants and animals. Our ViT-2B sets a new state-of-the-art result on iNat18 (+4.5% over [33]) Video classification. In Table 3 we investigate how MAE→WSP's pretraining transfers to video action classification on K400 and SSv2. MAE→WSP is competitive with state-of-the-art methods, including ones that pretrain on videos, whereas our models are only pretrained on images. Specifically, our ViT-L gets the highest reported performance on both video datasets. For all video finetuning, we use relative position embeddings [24], which improves our performance by 0.6% on K400 for a ViT-L. Overall, the results indicate the promise of MAE pre-pretraining for building strong video understanding models.", + "bbox": [ + 496, + 583, + 893, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Low-shot image classification. We evaluate the label efficiency of our models using a few examples per class for finetuning. We use two datasets, IN1k and iNat18, with $K$ shots (labeled examples per class), $K \\in \\{1,5,10\\}$ . For", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5489", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/30912965ae66cac80befd575b4e04b34f1a9ca7744361083254d82aee62fa36e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDatasetArch.IN1kiNat18
1-shot5-shot10-shot1-shot5-shot10-shot
Results with different pretraining datasets
CLIP [59]WIT-400MViT-L41.366.271.321.949.058.5
OpenCLIP [41]LAION-2BViT-H44.370.074.926.054.663.7
OpenCLIP [41]LAION-2BViT-G46.372.977.226.355.765.1
Scale-ViT†[84]JFT-3BViT-G-83.084.9---
DINO [12]IN1kViT-B/845.864.669.019.845.955.9
MSN [3]IN1kViT-L/757.172.174.417.038.048.1
MAE [33]IN1kViT-H-57.970.8-48.568.2
SWAG [70]IG-3.6BViT-H59.478.781.030.162.872.3
MAE→WSPIG-3BViT-H57.179.882.531.767.876.1
MAE→WSPIG-3BViT-2B62.181.583.735.572.880.3
", + "bbox": [ + 80, + 109, + 553, + 286 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9ae6dafa25c7f370c77040d3cb808efe53f5fe4762e47843a2449b4c7e7899a8.jpg", + "table_caption": [ + "Table 4: Low shot image classification. We compare the performance of our models using just a few examples per class for ImageNet-1k and iNaturalist-18. MAE $\\rightarrow$ WSP excels at classification even with just 1 example per class. Pretraining on large scale data can outperform techniques designed to work well in the low data regime, like MSN. We evaluate and report results for each technique with the best protocol, except for when the checkpoints are not available $^{\\dagger}$ ." + ], + "table_footnote": [], + "table_body": "
MethodDatasetArch.Res.IN1kF-101
Results with different pretraining datasets
CLIP [59]WIT-400MViT-L/1433676.293.8
OpenCLIP [41]LAION-2BViT-H22478.092.5
OpenCLIP [41]LAION-2BViT-G22480.192.9
Florence [82]FLD-900MCoSwin-H38483.795.1
Scale-ViT [84]JFT-3BViT-L22480.8-
+ LiT [85]+ ALIGNN-3.6B
Scale-ViT [84]JFT-3BViT-g28885.2-
+ LiT [85]+ ALIGNN-3.6B
CoCa [81]JFT-3B + ALIGNN-1.8BCoCa-2B57686.3-
MAE→WSPIG-3BViT-H22480.895.8
MAE→WSPIG-3BViT-2B22481.996.2
", + "bbox": [ + 558, + 102, + 903, + 280 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/421db912e0c03b654b08fe01c36db1738bd2a251a63c449097ca39cb0d719c18.jpg", + "table_caption": [ + "Table 5: Zero shot image classification results. We evaluate zero-shot transfer on IN1k and Food-101. Our models push the state-of-the-art on F-101, while being competitive on IN1k. The best performing models on IN1k train on JFT3B and ALIGN, and the performance on the two datasets is not well correlated, exemplifying the impact of pretraining dataset choice on zero-shot transfer performance." + ], + "table_footnote": [], + "table_body": "
MethodDatasetArch.FrameworkAPboxAPmask
Results with a different detection framework
Winner 2021 [26]IN21kCBNetv2HTC-49.2
MAE [46]IN1kViTDet-HCascade51.546.6
SWAG [70]IG-3BViTDet-HCascade47.142.1
MAE→WSPIG-3BViTDet-HCascade50.845.5
MAE→WSPIG-3BViTDet-2BCascade51.846.1
", + "bbox": [ + 91, + 407, + 460, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "iNat18, as some classes have less than $K$ images, we adapt our setting to consider at most $K$ shots. For each value of $K$ , we generate 5 splits of the original dataset using 5 different random seeds and report the mean top-1 accuracy.", + "bbox": [ + 75, + 636, + 468, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate two protocols for low-shot finetuning – linear classifiers and Adapters [37], both of which keep the entire model parameters frozen and introduce a few trainable parameters. We evaluated multiple Adapters proposed for ViTs – LoRA [38], AdaptFormer [14], and VPT [42]. We found that VPT performed the best while being robust to the choice of hyperparameters, and outperforms linear classifiers for our models. For other works, we report with the best protocol. Full details in Appendix B.", + "bbox": [ + 75, + 700, + 470, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 shows a comparison with state-of-the-art methods on low-shot IN1k and iNat18, including foundational and self-supervised models. Our models show impressive low-shot performance on both IN1k and iNat18, reaching $83.7\\%$", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7cfdd741a986fa2f790085a09520e7b4cd27fbd4f174c529b32c86b8e4ac92b3.jpg", + "table_caption": [ + "Table 6: Detection and Segmentation results on LVIS (v1 val). We compare MAE $\\rightarrow$ WSP with prior work and report the detection $(\\mathrm{AP}^{\\mathrm{box}})$ and instance segmentation performance $(\\mathrm{AP}^{\\mathrm{mask}})$ . Our ViT-H significantly outperforms SWAG which uses the same pretraining and model size, but without pre-pretraining. Our detection performance also improves with the larger 2B parameter model." + ], + "table_footnote": [], + "table_body": "
MethodDatasetArch.FrameworkAPboxAPmask
Results with different detection framework
CBNetv2 [48]IN21k2 ×Swin-LHTC59.151.0
SwinV2-L [50]IN21kSwinV2-LHTC++58.951.2
Co-DETR [90]IN21kSwin-LCo-DETR58.5-
Methods that pretrain on a detection dataset
Florence [82]FLD-900M + FLOD-9M†CoSwin-HDyHead [19]62.0-
DINO [87]IN21k + O365†Swin-L-63.2-
FocalNet [80]IN21k + O365†FocalNet-HDINO [87]64.2-
Co-DETR [90]IN21k + O365†MixMIM-gCo-DETR64.4-
MAE [46]IN1kViTDet-HCascade58.750.9
SWAG [70]IG-3BViTDet-HCascade55.747.9
MAE→WSPIG-3BViTDet-HCascade57.749.6
MAE→WSPIG-3BViTDet-2BCascade58.050.1
", + "bbox": [ + 504, + 407, + 893, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7: Detection and Segmentation on COCO (val2017). We report the detection $(\\mathrm{AP}^{\\mathrm{box}})$ and instance segmentation performance $(\\mathrm{AP}^{\\mathrm{mask}})$ . Our models outperform SWAG which uses the same pretraining, but without pre-pretraining. $^\\dagger$ Large scale detection datasets - using additional detection data like Objects365 has been shown to boost performance by $5.6~\\mathrm{mAP}$ [67].", + "bbox": [ + 496, + 609, + 893, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and $80.3\\%$ top-1 accuracy with only 10 labeled examples per class, and reaching the highest reported performance with just one labeled example per class on IN1k of $62.1\\%$ .", + "bbox": [ + 496, + 729, + 893, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot transfer. Strong foundational models are expected to also have a good open world understanding of visual concepts. To equip our pretrained vision encoders with such capabilities, we utilize LiT [85]. We initialize the text encoder from an XLM-R Large [17] model. Table 5 shows the zero-shot transfer performance of our models on IN1k, and F-101. Our ViT-2B attains $81.9\\%$ accuracy on IN1k, outperforming a similarly sized OpenCLIP model.", + "bbox": [ + 496, + 779, + 895, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5490", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7998ea378450e1a5619c31d6d7cec1fcc0fae2a261e499bfe783e062ac012b13.jpg", + "image_caption": [ + "Figure 7: Pre-pretraining with MAE significantly boosts performance on detection. Scaling the model size with WSP only does not improve performance on detection. Adding MAE as prepretraining helps MAE $\\rightarrow$ WSP scaling on detection." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 269, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/55cbfa016d97cc918ca93cc6a3b8d78e9394aab699c8894eb83a2fce446d2b11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 90, + 460, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our results lag behind other works which pretrain on datasets such as JFT-3B and ALIGN, highlighting the importance of the pretraining dataset for performance. This data-advantage is also observed in the finetuning results discussed in Table 2, where JFT-3B provides best IN1k accuracy. On Food-101 we attain the highest zero-shot transfer accuracy of $96.2\\%$ . The performance on the two datasets is not well correlated, further demonstrating the impact the pretraining dataset can have on a particular zero-shot task.", + "bbox": [ + 75, + 352, + 468, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Detection and segmentation. Next, we evaluate models on detection and instance segmentation, on the LVIS [32] (Table 6) and COCO [49] datasets (Table 7). We use the Cascade Mask R-CNN framework [35], with the ViTDet [47] architecture, and initialize the backbone with our pretrained models. For finetuning our models, we start with the hyperparameters from [47] and adapt them for our models, full details are in Appendix B. We also perform a system level comparison with other state-of-the-art works, but note that drawing meaningful conclusions out of this is difficult, on account of the multiple differences in the detection frameworks, model architectures, and datasets.", + "bbox": [ + 75, + 489, + 470, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On both benchmarks, MAE $\\rightarrow$ WSP considerably outperforms the weakly supervised SWAG [70] model, demonstrating the benefit of our additional pre-pretraining stage. On the long-tailed LVIS dataset, MAE $\\rightarrow$ WSP outperforms MAE IN1k pretraining on detection AP [47], but lags slightly behind on COCO. MAE's strong performance on detection using both IN1k and IG-3B (Figure 2) can potentially be explained by the fact that it is trained to reproduce images with $75\\%$ masking, whereas WSP is only trained to predict one or a few salient objects in an image. Lastly, methods which use additional detection data, such as Objects365 [67] or FLOD-9M [82], have strong detection performance.", + "bbox": [ + 75, + 672, + 470, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analyzing detection performance. We further inspect the benefits of pre-pretraining for detection in Figure 7. Unlike other tasks like image / video classification, scaling model", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10731e96835c7150bcae427746d72e196863719bfb9ed7090769ec48cb690214.jpg", + "image_caption": [ + "Figure 8: Dissecting detection performance. LVIS performance based on the size of objects (left), or the class occurrence frequency (right). MAE is stronger than WSP at detecting small objects and frequent classes, and is worse on rare classes. MAE $\\rightarrow$ WSP outperforms WSP on all axes. It is much closer to MAE on smaller objects and rare classes, and also outperforms MAE elsewhere." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 691, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da1ddd1635dcf4bd9d9a4a0cd4298368e0b67b371544c267405f860094bb7fe2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 92, + 895, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "size using WSP pretraining does not improve detection performance. However, adding MAE pre-pretraining provides consistent gains and allows WSP to scale with model size.", + "bbox": [ + 496, + 372, + 893, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We dissect the performance of ViT-L models trained with MAE, WSP, and $\\mathrm{MAE} \\rightarrow \\mathrm{WSP}$ in Figure 8 based on the size of the object, or by the frequency of the object's class. We observe that $\\mathrm{MAE} \\rightarrow \\mathrm{WSP}$ performs better than WSP on all tasks - detecting rare to frequent classes across small to large object sizes. $\\mathrm{MAE} \\rightarrow \\mathrm{WSP}$ improves over MAE at detecting rare objects, presumably because of the diversity in the IG-3B labels, and improves at detecting frequent classes over WSP due to MAE pre-pretraining.", + "bbox": [ + 496, + 417, + 893, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 566, + 617, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced pre-pretraining which is an initial stage in the standard pretrain-then-finetune paradigm. Prepretraining uses MAE and thus, does not need additional supervision and can be conveniently added to web-scale training pipelines. We show that pre-pretraining improves downstream performance on multiple different recognition tasks, improves model convergence, and is overall more efficient than standard weakly-supervised pretraining. Our self-supervised pre-pretraining improves results for models trained with billions of labels, showing that it is a scalable technique that matters even at web-scale. The benefits of using pre-pretraining hold across varying model sizes, and different pretraining data distributions showing that it is a robust technique. Finally, pre-pretraining naturally and successfully combines the two most common pretraining strategies - self-supervised and weakly-supervised learning. Our results suggest that model initialization plays a significant role in the final performance, training dynamics etc. even for web-scale training with billions of parameter updates and labels, and should be further investigated.", + "bbox": [ + 496, + 592, + 893, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5491", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 90, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samira Abnar, Mostafa Dehghani, Behnam Neyshabur, and Hanie Sedghi. Exploring the limits of large scale pre-training. arXiv preprint arXiv:2110.02095, 2021.", + "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A video vision transformer. In CVPR, 2021.", + "[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022.", + "[4] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. arXiv preprint arXiv:2301.08243, 2023.", + "[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.", + "[6] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, 2019.", + "[7] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020.", + "[8] Daniel Bolya, Sean Foley, James Hays, and Judy Hoffman. Tide: A general toolbox for identifying object detection errors. In ECCV, 2020.", + "[9] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV. Springer, 2014.", + "[10] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020.", + "[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020.", + "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021.", + "[13] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017.", + "[14] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. arXiv preprint arXiv:2205.13535, 2022.", + "[15] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020.", + "[16] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. *ICLR*, 2020.", + "[17] Alexis Conneau, Kartikay Khandelwal, Naman Goyal," + ], + "bbox": [ + 78, + 114, + 470, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In ACL, 2020.", + "[18] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR, 2020.", + "[19] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In CVPR, 2021.", + "[20] Terrance De Vries, Ishan Misra, Changhan Wang, and Laurens Van der Maaten. Does object recognition work for everyone? In CVPR Workshop, 2019.", + "[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009.", + "[22] Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng, and Trevor Darrell. Decaf: A deep convolutional activation feature for generic visual recognition. In ICML, 2014.", + "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021.", + "[24] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021.", + "[25] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022.", + "[26] WeiFu Fu, CongChong Nie, Ting Sun, Jun Liu, TianLiang Zhang, and Yong Liu. Lvis challenge track technical report 1st place solution: distribution balanced and boundary refinement for large vocabulary instance segmentation. arXiv preprint arXiv:2111.02668, 2021.", + "[27] Deepti Ghadiyaram, Du Tran, and Dhruv Mahajan. Large-scale weakly-supervised pre-training for video action recognition. In CVPR, 2019.", + "[28] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimae: Single model masked pretraining on images and videos. In CVPR, 2023.", + "[29] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. CVPR, 2013.", + "[30] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The \"something something\" video database for learning and evaluating visual common sense. In ICCV, 2017.", + "[31] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020." + ], + "bbox": [ + 501, + 92, + 892, + 891 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5492", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019.", + "[33] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022.", + "[34] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019.", + "[35] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017.", + "[36] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018.", + "[37] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019.", + "[38] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022.", + "[39] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016.", + "[40] Badr Youbi Idrissi, Diane Bouchacourt, Randall Balestriero, Ivan Evtimov, Caner Hazirbas, Nicolas Ballas, Pascal Vincent, Michal Drozdzal, David Lopez-Paz, and Mark Ibrahim. Imagenet-x: Understanding model mistakes with factor of variation annotations. arXiv preprint arXiv:2211.01866, 2022.", + "[41] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021.", + "[42] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022.", + "[43] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, AMustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017.", + "[44] S Kornblith, J Shlens, and QV Le. Do better imagenet models transfer better? arxiv 2018. In CVPR, 2019.", + "[45] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020.", + "[46] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022.", + "[47] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022.", + "[48] Tingting Liang, Xiaojie Chu, Yudong Liu, Yongtao Wang, Zhi Tang, Wei Chu, Jingdong Chen, and Haibin Ling. Cbnet: A composite backbone network architecture for object detection." + ], + "bbox": [ + 78, + 92, + 470, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Transactions on Image Processing, 2022.", + "[49] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014.", + "[50] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022.", + "[51] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017.", + "[52] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018.", + "[53] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 1995.", + "[54] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022.", + "[55] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016.", + "[56] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023.", + "[57] Karol J Piczak. Esc: Dataset for environmental sound classification. In ACM MM, 2015.", + "[58] Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization, 1992.", + "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021.", + "[60] Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. Cnn features off-the-shelf: an astounding baseline for recognition. In CVPR Workshops, 2014.", + "[61] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imageNet? In ICML, 2019.", + "[62] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015.", + "[63] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. In NeurIPS, 2021.", + "[64] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015." + ], + "bbox": [ + 503, + 93, + 892, + 888 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5493", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022.", + "[66] Vaishaal Shankar, Achal Dave, Rebecca Roelofs, Deva Ramanan, Benjamin Recht, and Ludwig Schmidt. Do image classifiers generalize across time? In ICCV, 2021.", + "[67] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In ICCV, 2019.", + "[68] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS, 2014.", + "[69] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022.", + "[70] Mannat Singh, Laura Gustafson, Aaron Adcock, Vinicius de Freitas Reis, Bugra Gedik, Raj Prateek Kosaraju, Dhruv Mahajan, Ross Girshick, Piotr Dólár, and Laurens van der Maaten. Revisiting weakly supervised pre-training of visual perception models. In CVPR, 2022.", + "[71] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014.", + "[72] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021.", + "[73] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016.", + "[74] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. In NeurIPS, 2020.", + "[75] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022.", + "[76] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017.", + "[77] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and León Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. JMLR, 2010.", + "[78] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022.", + "[79] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022.", + "[80] Jianwei Yang, Chunyuan Li, and Jianfeng Gao. Focal modulation networks. arXiv preprint arXiv:2203.11926, 2022.", + "[81] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive" + ], + "bbox": [ + 78, + 92, + 470, + 888 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "captioners are image-text foundation models. TMLR, 2022.", + "[82] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021.", + "[83] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019.", + "[84] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, 2022.", + "[85] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In CVPR, 2022.", + "[86] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018.", + "[87] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2022.", + "[88] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, 2020.", + "[89] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. ICLR, 2022.", + "[90] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. arXiv preprint arXiv:2211.12860, 2022." + ], + "bbox": [ + 503, + 92, + 893, + 508 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5494", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/5b230860-aa61-4f33-9fe7-2407ba62f713_model.json b/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/5b230860-aa61-4f33-9fe7-2407ba62f713_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5f6ab74bc192b067addfdb3f49a119e9cd7f2a85 --- /dev/null +++ b/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/5b230860-aa61-4f33-9fe7-2407ba62f713_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.836, + 0.154 + ], + "angle": 0, + "content": "The effectiveness of MAE pre-pretraining for billion-scale pretraining" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.183, + 0.74, + 0.252 + ], + "angle": 0, + "content": "Mannat Singh*,† Quentin Duval* Kalyan Vasudev Alwala* Haoqi Fan Vaibhav Aggarwal Aaron Adcock Armand Joulin Piotr Dólár Christoph Feichtenhofer Ross Girshick Rohit Girdhar Ishan Misra Meta AI" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.668 + ], + "angle": 0, + "content": "This paper revisits the standard pretrain-then-finetune paradigm used in computer vision for visual recognition tasks. Typically, state-of-the-art foundation models are pretrained using large scale (weakly) supervised datasets with billions of images. We introduce an additional pretraining stage that is simple and uses the self-supervised MAE technique to initialize the model. While MAE has only been shown to scale with the size of models, we find that it scales with the size of the training dataset as well. Thus, our MAE-based pre-training scales with both model and data size making it applicable for training foundation models. Pre-training consistently improves both the model convergence and the downstream transfer performance across a range of model scales (millions to billions of parameters), and dataset sizes (millions to billions of images). We measure the effectiveness of pre-training on 10 different visual recognition tasks spanning image classification, video recognition, object detection, low-shot classification and zero-shot recognition. Our largest model achieves new state-of-the-art results on iNaturalist-18 \\((91.3\\%)\\), 1-shot ImageNet-1k \\((62.1\\%)\\), and zero-shot transfer on Food-101 \\((96.2\\%)\\). Our study reveals that model initialization plays a significant role, even for web-scale pretraining with billions of images." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.697, + 0.21, + 0.713 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.471, + 0.859 + ], + "angle": 0, + "content": "The pretrain-then-finetune paradigm in visual recognition has enabled high performance visual recognition models across a range of tasks such as image classification [52, 59, 70], video action recognition [25, 27, 28], object detection [11, 89], 3D etc. Typically, pretraining consists of training a model using a pretraining task on large scale data. The resulting pretrained models learn general purpose visual representations that can be used for a range of target tasks, often with limited labeled data, by transfer learning." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.86, + 0.472, + 0.876 + ], + "angle": 0, + "content": "In this paper, we show that an initial stage of pre" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.286, + 0.889, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.53, + 0.896, + 0.614 + ], + "angle": 0, + "content": "Figure 1: MAE pre-pretraining improves performance. Transfer performance of a ViT-L architecture trained with self-supervised pretraining (MAE), weakly supervised pretraining on billions of images (WSP), and our pre-pretraining \\((\\mathrm{MAE} \\rightarrow \\mathrm{WSP})\\) that initializes the model with MAE and then pretrains with WSP. Pre-pretraining consistently improves performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.895, + 0.824 + ], + "angle": 0, + "content": "pretraining before the standard pretraining task can improve vision models across a variety of different tasks. Our method combines two common pretraining tasks in vision: (1) weakly supervised pretraining that uses weak, often noisy, signals such as text or image hashtags as supervision, and (2) self-supervised pretraining that only uses the data without additional supervision. Both forms of pretraining start training with a randomly initialized model and have proven effective at learning general purpose vision models. While there have been attempts to combine both these forms of pretraining [54, 69], they are typically used independently in the pretrain-then-finetune two stage paradigm [23, 70, 84]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this work we explore the combination of self- and weakly-supervised learning in a simple pre-pretraining framework, as follows. We first begin with the Masked Autoencoder (MAE) [33] self-supervised learning technique to pre-pretrain vision models without using any labels. After" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.345, + 0.901 + ], + "angle": 0, + "content": "*Equal technical contribution. †Project Lead." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5484" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "initializing from the pre-pretrained model, we use standard weakly supervised pretraining on billions of images with noisy labels. We perform a large-scale empirical study to measure the effectiveness of pre-pretraining on 10 different visual recognition tasks spanning image classification, video recognition, object detection, low-shot classification and zero-shot recognition. Our study reveals that pre-pretrain initialization improves the performance for the weakly supervised models, and this improvement holds even at billion scale weakly labeled data, and across vision tasks (Figure 1). It also improves the model convergence during pretraining, leading to an efficient way of training large scale vision models. Pre-pretrain further enjoys the computational efficiency of the MAE approach, making it simple and scalable. Finally, we show that by using pre-pretraining, both self-supervised learning and weakly supervised learning can be combined for improved model performance for billion-scale data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.485 + ], + "angle": 0, + "content": "Pre-pretrain is related to 'intermediate finetuning' [5, 50] which introduces a stage after pretraining to better align the pretrained features with the downstream task using labeled data. In contrast, pre-pretrain serves as a better way to initialize a model before pretraining. Since we leverage MAE for pre-pretraining, we do not need additional information or labels for this stage and can re-use the pretraining data. This makes pre-pretrain convenient and simple to use with existing pretraining datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.473, + 0.756 + ], + "angle": 0, + "content": "Our study on large-scale pre-pretraining reveals that model initialization plays a significant role, even for web-scale pretraining, and pre-pretraining is a simple and promising technique in that direction. In particular, we show that (i) MAE not only scales with model size as shown in [33], but also with the size of the training data (Figure 2). (ii) Pre-pretraining improves both the model convergence and the final downstream performance for different sized models (millions to billions of parameters) trained on different sized datasets (millions to billions of labels). (iii) Using pre-pretraining combines the benefits of both self-supervised learning and large scale weakly-supervised learning, and our models achieve excellent performance on a variety of different visual recognition tasks (Figure 1). Most prominently, our model sets new state-of-the-art results on iNaturalist-18 image classification \\((91.3\\%)\\), 1-shot ImageNet-1k image classification \\((62.1\\%)\\), and zero-shot transfer on Food-101 \\((96.2\\%)\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.77, + 0.22, + 0.787 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Supervised pretraining of transferrable representations on large labeled datasets [21, 43, 63] and employing them for downstream recognition tasks, has emerged as a powerful approach in computer vision. It has spurred rapid progress on various tasks including image classification [22, 57, 60], object detection/segmentation [29, 62], image captioning [45, 81] and video action recognition [13, 24, 68]. While useful," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.363 + ], + "angle": 0, + "content": "such representations are often limited by the scale and diversity of the supervision in the pretraining datasets. Hence, recent work has probed the effectiveness, robustness, and fairness of these representations [1, 20, 40, 44, 61, 66, 74]. Self-supervised pretraining is a promising alternative to learn these representation without relying on large well-labeled datasets. Initial works focused on reconstructions methods [77] before moving to other pretraining tasks such as solving jigsaw puzzles [55], constractive learning [15, 34] or joint embedding approaches [3, 4, 11, 12, 31, 89]. With the advent of Vision Transformers [23], approaches based on reconstructions such as [5, 33, 79] got renewed interest for their simplicity and state of the art performance. Of particular interest to us is MAE [33] for its state of the art performance on many transfer tasks [25, 28, 33, 46, 75] and its computational efficiency. Given the lack of supervision during pretraining, these representations often require significant finetuning to align to downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.897, + 0.591 + ], + "angle": 0, + "content": "Weakly supervised pretraining (WSP) is a middle-ground between supervised and self-supervised pretraining. Instead of ignoring annotations completely as in self-supervised pretraining, or requiring exhaustive labels as in supervised pretraining, WSP relies on the large quantity of \"free\" annotation available on the internet. These annotations occur as image-text pairs [59, 65], where the text can additionally be processed to produce pseudo labels. Of particular interest to us is the latter, i.e. approaches which leverage multi-label classification on noisy labels [27, 52, 70, 84] which have shown state of the art fine-tuning performance, and at the same time can be adapted using image-text data to gain zero-shot capabilities [85]. In this work, we explore WSP in conjunction with self-supervised pre-pretraining, and show faster convergence and stronger performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.574, + 0.621 + ], + "angle": 0, + "content": "3. Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.81 + ], + "angle": 0, + "content": "Our goal is to empirically study the effectiveness of self-supervised pre-pretraining as a precursor to billion scale weakly supervised pretraining for representation learning. Given the simplicity and efficiency of Masked AutoEncoding (MAE) [33], we leverage it as the self-supervised prepretraining approach. Our study shows that MAE scales with the size of the pretraining dataset and model size, and combining it with weak supervision improves large scale vision models. Additionally, such a combination leads to faster convergence and is a simple, scalable way to learn visual representations at scale. We describe our setup and the approaches in detail next." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Architecture. We use the Vision Transformer (ViT) [23] architecture as the visual encoder for all our experiments. ViTs employ minimal vision-specific inductive biases combined with the standard transformer architecture [76], and yet have emerged as an architecture of choice for a wide variety of visual and multimodal recognition tasks [2, 28, 84]. We train" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5485" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.211 + ], + "angle": 0, + "content": "ViT models at various scales in terms of number of parameters, including ViT-B (86M), ViT-L (307M), and ViT-H (632M). We also train on larger 1.9B and 6.5B parameter ViT models, which we call ViT-2B and ViT-6.5B, respectively (Appendix Table 8). As is common practice [23, 84], we train models of sizes ViT-B, ViT-L with a patch size of 16 and larger models with a patch size of 14. We pretrain with a \\(224 \\times 224\\) resolution for all models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.473, + 0.424 + ], + "angle": 0, + "content": "Pre-pretraining (MAE) [33] learns visual representations from image datasets without using any labels. We choose this approach as it is simple to implement and scales very effectively with large ViT model sizes due to patch dropping as described next. MAE randomly masks \\(75\\%\\) of an image and trains the model to reconstruct the masked input image by minimizing the pixel reconstruction error. The target pixel values for a given patch are normalized by the mean and standard deviation of all pixels in it. Coupled with the ViT architecture, MAE can be trained by only processing the \\(25\\%\\) unmasked image patches. A separate, smaller, decoder is then used to reconstruct the missing part of the input. This asymmetrical design makes training the encoder extremely efficient, allowing for scaling visual encoder sizes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.473, + 0.528 + ], + "angle": 0, + "content": "Weakly-supervised pretraining (WSP) leverages images with associated 'weak' supervision for training models. In particular, we focus on internet images and use their associated text information as supervision. We convert the text into a discrete set of labels, specifically leveraging hash-tag information [27, 52, 70]. We then use a multi-label classification loss to train models. We refer to this method as WSP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.473, + 0.664 + ], + "angle": 0, + "content": "\\(\\mathbf{MAE} \\rightarrow \\mathbf{WSP}\\) first trains the encoder using the MAE self-supervised method using only the images. This pretraining stage initializes the model while simultaneously being computationally efficient because of the masking used in MAE. In the second stage, we pretrain the encoder using both the image and associated weak supervision. This combination outperforms using either strategy in isolation, i.e., an MAE model or a weakly supervised model trained from scratch." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.21, + 0.695 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.471, + 0.794 + ], + "angle": 0, + "content": "We empirically evaluate and analyze large scale MAE pre-pretraining using Instagram data on a variety of different visual recognition tasks. We describe the datasets used for pretraining and evaluation, followed by analysis of the pretraining design decisions, and finally the downstream transfer evaluation of our learned representation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.334, + 0.818 + ], + "angle": 0, + "content": "4.1. Datasets and training details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Pretraining dataset. We use Instagram-3B (IG-3B) a billion-scale multi-label dataset sourced from Instagram (IG). This multi-label dataset contains 28K classes and 3B unique images, resampled to 5B total images, and was produced by running the dataset generation pipeline from SWAG [70]" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.089, + 0.892, + 0.226 + ], + "angle": 0, + "content": "
DatasetTask#cls#train#val
ImageNet-1k (IN1k) [64]Image cls.10001M50K
iNaturalist-18 (iNat18) [36]Fine-grained cls.8142437K24K
ImageNetv2 (INv2) [61]Image cls.1000-10K
ImageNet-ReaL (IN-ReaL) [7]Image cls.1000-50K
ObjectNet (ON) [6]Image cls.113-19K
Food-101 (F-101) [9]Image cls.101N/A25K
COCO [49]Obj. det.80118K5K
LVIS [32]Obj. det.1K100K20K
Kinetics-400 (K400) [43]Action cls.400220K20K
Something Something-v2 (SSv2) [30]Action cls.174169K25K
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.234, + 0.894, + 0.304 + ], + "angle": 0, + "content": "Table 1: Evaluation datasets used to evaluate MAE \\(\\rightarrow\\) WSP on image classification, object detection, and video action recognition tasks. The table reports the task, number of classes (#cls), number of training samples (#train), and number of validation samples (#val) for each dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.895, + 0.456 + ], + "angle": 0, + "content": "without modification. Compared to [70], our version of the dataset has \\(16\\%\\) fewer images (3.0B vs. 3.6B), but we were able to reproduce the results from [70] with our version. We obtain labels using an automated process wherein we first obtain hashtags from the associated image captions, and then map the hashtags to WordNet synsets following [70]. After this processing, we get the weakly labeled IG dataset that contains images and their associated labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.897, + 0.699 + ], + "angle": 0, + "content": "Evaluation datasets. We evaluate MAE \\(\\rightarrow\\) WSP on a variety of different downstream visual recognition tasks. To evaluate our model on image classification, we use the standard ImageNet-1k [64] (IN1k) dataset, and also the long-tailed and fine-grained iNaturalist-18 [36] (iNat18) dataset. For object detection and segmentation, we use the popular COCO [49] dataset, and also LVIS [32], a large vocabulary dataset for long tailed object recognition. We evaluate video classification performance using two popular action recognition datasets, Kinetics-400 [43] (K400) and Something Something-v2 [30] (SSv2). For zero-shot transfer, we evaluate on IN1k and Food-101 [9] (F-101). We also evaluate the robustness of our models on test sets which overlap with IN1k classes, specifically ImageNetv2 [61] (INv2), ImageNet-ReaL [7] (IN-ReaL), and ObjectNet [6] (ON). Please see Table 1 for more details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.895, + 0.776 + ], + "angle": 0, + "content": "MAE pretraining details. We follow [33] to train MAE models on IG-3B without using any labels. We mask \\(75\\%\\) of the image for this training and train the model for 1 epoch over the dataset. We follow the same hyperparameters used in [33] for pretraining on IN1k." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.778, + 0.895, + 0.84 + ], + "angle": 0, + "content": "Supervised pretraining details. We train with a supervised cross-entropy loss on IG-3B using the hashtags as labels. This model is trained by default with random weight initialization and we use the training hyperparameters from [70]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Using pre-pretraining. When using pre-pretraining, we first train a model from scratch using MAE on the IG dataset. We then use the weights of the MAE encoder and perform supervised pretraining using the cross-entropy loss as described" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5486" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.093, + 0.278, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.092, + 0.477, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.092, + 0.682, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.092, + 0.882, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.895, + 0.343 + ], + "angle": 0, + "content": "Figure 2: Scaling MAE with model and dataset size. We plot MAE's performance when pretrained on ImageNet-1k or Instagram-3B and finetuned on downstream tasks. MAE scales to billion parameters sized models using just IN1k pretraining. Larger models show improved scaling behavior when pretrained with the much larger IG-3B dataset. MAE pretrained on IN1k data point is missing for the 2 billion model as training at that scale was unstable on both COCO and LVIS datasets. Tabulated results in Table 19." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.471, + 0.427 + ], + "angle": 0, + "content": "above. We reuse the same hyperparameters and training details as [70], i.e. there is no hyperparameter search needed for MAE \\(\\rightarrow\\) WSP, and we train for 1 epoch on IG-3B." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.594 + ], + "angle": 0, + "content": "Zero-shot training and evaluation details. To impart zero shot understanding capabilities to our models, we use the LiT approach from [85]. For LiT, we use the original (image, caption) pairs from the IG-3B dataset. We freeze the image encoder, and train a text encoder to encode the image captions and match the text embeddings to the associated image embedding using a CLIP loss [59]. We train the text encoder for 1 epoch. For evaluation, we follow [59] - we use the text encoder to compute embeddings from the templated text descriptions of classes and use the cosine similarity of the image and text embeddings as the classification score." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.595, + 0.471, + 0.625 + ], + "angle": 0, + "content": "For full training details and hyperparameters, please refer to Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.411, + 0.652 + ], + "angle": 0, + "content": "4.2. Scaling MAE pretraining to large data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Since our pre-pretraining uses MAE in the very first stage, we first study how MAE behaves on the large scale IG-3B dataset. We compare the performance of MAE pretraining on the large scale IG-3B with the original MAE [33] models trained on IN1k for 1600 epochs. We train models of varying sizes, from ViT-B to ViT-H as in [33]. To test the scaling behavior further, we also train MAE on the ViT-2B model with 2B parameters. We measure the performance of the resulting models in Figure 2 on four different vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We observe that using the IG-3B data provides consistent gains over IN1k for all vision tasks, and the gain increases for larger models. These experiments show that MAE scales with the size of the pretraining dataset, and benefits from using billions of images from IG-3B. He et al. [33]'s findings were limited to the fact that MAE scales with the size of the model, and thus our findings on MAE scaling with the size" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.383, + 0.84, + 0.397 + ], + "angle": 0, + "content": "of the pretraining data are complementary to theirs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.895, + 0.55 + ], + "angle": 0, + "content": "Our ViT-2B model pretrained on IG-3B improves upon the best results from [33] on image classification, attaining \\(88.4\\%\\) on IN1k \\((+0.6\\%)\\) and \\(89.3\\%\\) on iNat18 \\((+2.5\\%)\\) at \\(518\\times 518\\) resolution. The gains on detection are equally encouraging, with our ViT-2B reaching \\(53.6\\mathrm{AP}^{\\mathrm{box}}\\) on LVIS \\((+2.1\\) over [33]) and \\(59.8\\mathrm{AP}^{\\mathrm{box}}\\) on COCO \\((+1.2\\) over [33]). Lastly, we highlight the simplicity of our setup, since we use the same hyperparameters as [33] to train MAE on IG-3B, and also for the largest ViT-2B model, without encountering any training instabilities or needing any extra tweaks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.704, + 0.575 + ], + "angle": 0, + "content": "4.3. MAE pre-pretraining" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.894, + 0.674 + ], + "angle": 0, + "content": "Given the promising aspects of MAE as a pretraining approach from § 4.2, specifically that MAE (i) trains with larger models and datasets without needing any tuning (ii) shows gains when scaling model and / or dataset size (iii) is efficient to train, we investigate it as a pre-pretraining approach for supervised pretraining (WSP)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Figure 1 shows the performance of a ViT-L with MAE pretraining, supervised pretraining (WSP), or MAE pre-pretraining followed by supervised pretraining \\((\\mathrm{MAE}\\rightarrow \\mathrm{WSP})\\). We see that MAE and WSP have different strengths. MAE has strong performance for object detection, and full finetuned image classification. However, MAE underperforms on tasks where the model is not finetuned, such as linear classifiers, zero-shot, or low-shot classification - situations where WSP performs better. For these evaluations MAE lags behind WSP by more than 10 points, which is why the results for MAE are not visible in Figure 1. For video classification, MAE performs significantly better than WSP on SSv2, but lags behind it on K400." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "\\(\\mathrm{MAE} \\rightarrow \\mathrm{WSP}\\) outperforms either of MAE or WSP pretraining on most evaluations, across image classification," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5487" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.093, + 0.299, + 0.108 + ], + "angle": 0, + "content": "IN1k linear probe (accuracy)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.114, + 0.322, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.093, + 0.586, + 0.108 + ], + "angle": 0, + "content": "IN1k linear probe (accuracy)" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.112, + 0.62, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.093, + 0.874, + 0.108 + ], + "angle": 0, + "content": "IN1k linear probe (accuracy)" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.114, + 0.895, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.301, + 0.319, + 0.386 + ], + "angle": 0, + "content": "Figure 3: MAE pre-pretraining scales with model size. Across model sizes, MAE \\(\\rightarrow\\) WSP outperforms a WSP only model, and shows strong scaling behavior. Most notably, a 2B MAE \\(\\rightarrow\\) WSP model outperforms a 6.5B WSP model." + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.302, + 0.647, + 0.372 + ], + "angle": 0, + "content": "Figure 4: Varying the number of pre-pretraining epochs used to initialize the model for WSP pretraining. Pre-pretraining leads to improved convergence, providing higher performance using fewer number of WSP pretraining epochs." + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.303, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Figure 5: MAE \\(\\rightarrow\\) WSP is more FLOPs efficient than WSP. Across a wide range of training FLOP profiles for a ViT-B computed by varying WSP (and MAE) epochs, MAE \\(\\rightarrow\\) WSP outperforms a baseline WSP only model." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.412, + 0.466, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.655, + 0.472, + 0.699 + ], + "angle": 0, + "content": "Figure 6: MAE initialization for medium scale data. Results for a ViT-L trained on IN21k with MAE, WSP, and MAE \\(\\rightarrow\\) WSP. MAE pre-pretraining improves WSP results by a wide margin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.472, + 0.808 + ], + "angle": 0, + "content": "video recognition, zero-shot evaluation, object detection, etc. Given that all baselines are trained at billion-scale, these results show that MAE \\(\\rightarrow\\) WSP is a simple yet promising strategy to improve performance while requiring no extra data or tuning. Next, we ablate the key aspects of MAE \\(\\rightarrow\\) WSP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Effect of model size. In Figure 3 we study the effect of prepretraining compared to random initialization for different model sizes. After initialization, all models are pretrained using WSP on the IG-3B dataset, and we measure the transfer performance on IN1k using linear probing. We observe that MAE pre-pretraining gives consistent gains over the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.895, + 0.551 + ], + "angle": 0, + "content": "WSP baseline across all model sizes, ranging from 86M to 2B parameters. The gains over the WSP baseline increase for larger model sizes showing that pre-pretraining shows promising scaling behavior with model sizes. Notably, a 2B MAE \\(\\rightarrow\\) WSP model outperforms a larger 6.5B WSP model. Number of pre-pretraining epochs. We vary the number of MAE pre-pretraining and the number of WSP pretraining epochs to understand their effect on the final recognition performance. We study this in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.896, + 0.734 + ], + "angle": 0, + "content": "Pre-pretraining improves results over the standard pretraining (random initialization, w/o pre-pretraining), and provides large gains with fewer WSP pretraining epochs. Pre-pretraining also leads to faster convergence since even a small amount of pre-pretraining for 0.1 epochs provides improvements. Increasing the epochs of pre-pretraining provide a larger improvement, and the gains saturate at 1 epoch of pre-pretraining. Finally, pre-pretraining's gains do not diminish even after 4 epochs of WSP (20 billion samples) showing the value of pre-pretraining at scale. We also note that these gains are independent of the evaluation protocol, and we observed them with full finetuning on IN1k as well." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.895, + 0.825 + ], + "angle": 0, + "content": "Training efficiency. Figure 5 shows a comparison between WSP and MAE \\(\\rightarrow\\) WSP when comparing training FLOPs. For the same training FLOPs, MAE \\(\\rightarrow\\) WSP achieves better transfer performance compared to WSP, and is up to \\(2\\times\\) more efficient. Pre-pretraining's training efficiency holds over a large \\(10\\times\\) compute window." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Different datasets for pre-pretraining. We also evaluate the performance of MAE \\(\\rightarrow\\) WSP when pre-pretraining MAE on the much smaller ImageNet-1k dataset (Appendix Table 21) and find that pre-pretraining remains just as effective. This allows reusing pretrained MAE models in practice." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5488" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.104, + 0.541, + 0.289 + ], + "angle": 0, + "content": "
MethodDatasetArch.Res.ImageNet-1kiNat18
IN1kINv2ReaLON
MAE [33]IN1kViT-H44887.8---86.8
SWAG [70]IG-3.6BViT-H51888.681.190.569.586.0
DINOv2 [56]LVD-142MViT-g44888.9----
Florence [82]FLD-900MCoSwin-H51290.1----
ViT [23]JFT-300MViT-H51888.6-90.7--
Scale-ViT [84]JFT-3BViT-L38488.580.490.4--
Scale-ViT [84]JFT-3BViT-G51890.583.390.870.5-
SwinV2 [50]IN-ext-70MSwinV2-G64090.284.0---
CoCa [81]JFT-3B + ALIGNN-1.8BCoCa-2B57691.0----
MAE→WSPIG-3BViT-H51889.382.390.872.690.5
MAE→WSPIG-3BViT-2B51889.783.090.975.891.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.3, + 0.546, + 0.369 + ], + "angle": 0, + "content": "Table 2: Image classification results. We report finetuning results on IN1k and iNat18 and note the pretraining dataset, architecture and finetuning resolution. We also evaluate the performance of our IN1k finetuned models on multiple test sets to measure robustness. Our models are robust and also push the state-of-the-art on the challenging fine-grained and long-tailed iNat18 dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.471, + 0.532 + ], + "angle": 0, + "content": "Different datasets for pre-pretraining and pretraining. We investigate the effect of the dataset used in pre-pretraining and pretraining by using IN21k [21] for all methods, including MAE and MAE \\(\\rightarrow\\) WSP. Compared to IG-3B, IN21k is more curated, smaller (14M images), and has cleaner labels (21K classes), where each image is labeled with one class from the WordNet synsets [53]. For evaluating zero shot performance, we use the PMD [69] dataset for LiT training. For full details about the hyperparameters, refer to Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.623 + ], + "angle": 0, + "content": "Figure 6 compares the performance of MAE, WSP and \\(\\mathrm{MAE} \\rightarrow \\mathrm{WSP}\\) when pretrained on this IN21k dataset. We notice a similar trend as when pretraining on IG-3B where \\(\\mathrm{MAE} \\rightarrow \\mathrm{WSP}\\) outperforms both MAE and WSP. This shows that MAE pre-pretraining works with datasets of different scales and distributions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.269, + 0.649 + ], + "angle": 0, + "content": "4.4. Transfer Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We compare with state-of-the-art research on image and video classification, detection and segmentation, low-shot image classification, zero shot transfer, robustness analysis. ImageNet-1k image classification. Table 2 shows the performance of different methods on IN1k. MAE \\(\\rightarrow\\) WSP gets the best performance for a ViT-H sized model (89.3%). As we report in the appendix, MAE \\(\\rightarrow\\) WSP also achieves the best performance for ViT-L (88.8%). Recent methods such as Scale-ViT [84] are better on IN1k and we hypothesize that this gap stems mainly from the differences in the pretraining datasets (IG-3B vs. JFT-3B). We also compute linear performance using frozen features on IN1k at 224px resolution. Our models produce strong features which outperform other methods with WSP objectives. They also surpass the performance of the self-supervised DINOv2 [56] model optimized to produce strong frozen representations." + }, + { + "type": "table", + "bbox": [ + 0.553, + 0.127, + 0.899, + 0.28 + ], + "angle": 0, + "content": "
MethodDatasetArch.Res.K400SSv2
Florence [82]FLD-900MCoSwin-H38486.5-
SwinV2 [50]IN-ext-70MSwinV2-G38486.8-
CoCa [81]JFT-3B + ALIGNN-1.8BCoCa-2B57688.9-
Results with models pretrained on videos
MaskFeat [78]K400MViT-L22484.3-
MAE [25, 75]K400ViT-L22485.274.0
OmniMAE [28]IN1k + SSv2ViT-L22484.074.2
MAE [25, 75]K400ViT-H22486.6-
MAE→WSPIG-3BViT-L22486.074.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.29, + 0.902, + 0.346 + ], + "angle": 0, + "content": "Table 3: Video classification results on Kinetics-400 and Something Something-v2. Our models generalize well to video action recognition tasks, despite not seeing any videos during pretraining." + }, + { + "type": "table", + "bbox": [ + 0.544, + 0.392, + 0.851, + 0.474 + ], + "angle": 0, + "content": "
MethodDatasetArch.IN1k Linear
SWAG [70]IG-3.6BViT-H85.8
OpenCLIP [41]LAION-2BViT-G86.2
DINOv2 [56]LVD-142MViT-g86.5
MAE→WSPIG-3BViT-H87.0
MAE→WSPIG-3BViT-2B88.1
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.584 + ], + "angle": 0, + "content": "Robustness for image classification. We evaluate the robustness of our models finetuned on IN1k on additional test sets in Table 2 and notice that, despite MAE→WSP being \\(\\sim 1\\%\\) behind Scale-ViT on IN1k, it is significantly more robust and generalizes better on these additional test sets – MAE→WSP gets the highest reported performance on ImageNet-ReaL and ObjectNet for IN1k finetuned models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.841 + ], + "angle": 0, + "content": "Generalization in image classification. We evaluate the generalization of our model on additional fine-grained image classification using iNaturalist-18. iNat18 is a challenging long-tailed and fine-grained dataset with images of multiple species of visually similar plants and animals. Our ViT-2B sets a new state-of-the-art result on iNat18 (+4.5% over [33]) Video classification. In Table 3 we investigate how MAE→WSP's pretraining transfers to video action classification on K400 and SSv2. MAE→WSP is competitive with state-of-the-art methods, including ones that pretrain on videos, whereas our models are only pretrained on images. Specifically, our ViT-L gets the highest reported performance on both video datasets. For all video finetuning, we use relative position embeddings [24], which improves our performance by 0.6% on K400 for a ViT-L. Overall, the results indicate the promise of MAE pre-pretraining for building strong video understanding models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Low-shot image classification. We evaluate the label efficiency of our models using a few examples per class for finetuning. We use two datasets, IN1k and iNat18, with \\( K \\) shots (labeled examples per class), \\( K \\in \\{1,5,10\\} \\). For" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5489" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.111, + 0.554, + 0.287 + ], + "angle": 0, + "content": "
MethodDatasetArch.IN1kiNat18
1-shot5-shot10-shot1-shot5-shot10-shot
Results with different pretraining datasets
CLIP [59]WIT-400MViT-L41.366.271.321.949.058.5
OpenCLIP [41]LAION-2BViT-H44.370.074.926.054.663.7
OpenCLIP [41]LAION-2BViT-G46.372.977.226.355.765.1
Scale-ViT†[84]JFT-3BViT-G-83.084.9---
DINO [12]IN1kViT-B/845.864.669.019.845.955.9
MSN [3]IN1kViT-L/757.172.174.417.038.048.1
MAE [33]IN1kViT-H-57.970.8-48.568.2
SWAG [70]IG-3.6BViT-H59.478.781.030.162.872.3
MAE→WSPIG-3BViT-H57.179.882.531.767.876.1
MAE→WSPIG-3BViT-2B62.181.583.735.572.880.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.294, + 0.554, + 0.38 + ], + "angle": 0, + "content": "Table 4: Low shot image classification. We compare the performance of our models using just a few examples per class for ImageNet-1k and iNaturalist-18. MAE \\(\\rightarrow\\) WSP excels at classification even with just 1 example per class. Pretraining on large scale data can outperform techniques designed to work well in the low data regime, like MSN. We evaluate and report results for each technique with the best protocol, except for when the checkpoints are not available\\(^{\\dagger}\\)." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.103, + 0.905, + 0.281 + ], + "angle": 0, + "content": "
MethodDatasetArch.Res.IN1kF-101
Results with different pretraining datasets
CLIP [59]WIT-400MViT-L/1433676.293.8
OpenCLIP [41]LAION-2BViT-H22478.092.5
OpenCLIP [41]LAION-2BViT-G22480.192.9
Florence [82]FLD-900MCoSwin-H38483.795.1
Scale-ViT [84]JFT-3BViT-L22480.8-
+ LiT [85]+ ALIGNN-3.6B
Scale-ViT [84]JFT-3BViT-g28885.2-
+ LiT [85]+ ALIGNN-3.6B
CoCa [81]JFT-3B + ALIGNN-1.8BCoCa-2B57686.3-
MAE→WSPIG-3BViT-H22480.895.8
MAE→WSPIG-3BViT-2B22481.996.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.289, + 0.905, + 0.387 + ], + "angle": 0, + "content": "Table 5: Zero shot image classification results. We evaluate zero-shot transfer on IN1k and Food-101. Our models push the state-of-the-art on F-101, while being competitive on IN1k. The best performing models on IN1k train on JFT3B and ALIGN, and the performance on the two datasets is not well correlated, exemplifying the impact of pretraining dataset choice on zero-shot transfer performance." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.409, + 0.462, + 0.511 + ], + "angle": 0, + "content": "
MethodDatasetArch.FrameworkAPboxAPmask
Results with a different detection framework
Winner 2021 [26]IN21kCBNetv2HTC-49.2
MAE [46]IN1kViTDet-HCascade51.546.6
SWAG [70]IG-3BViTDet-HCascade47.142.1
MAE→WSPIG-3BViTDet-HCascade50.845.5
MAE→WSPIG-3BViTDet-2BCascade51.846.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.52, + 0.472, + 0.604 + ], + "angle": 0, + "content": "Table 6: Detection and Segmentation results on LVIS (v1 val). We compare MAE \\(\\rightarrow\\) WSP with prior work and report the detection \\((\\mathrm{AP}^{\\mathrm{box}})\\) and instance segmentation performance \\((\\mathrm{AP}^{\\mathrm{mask}})\\). Our ViT-H significantly outperforms SWAG which uses the same pretraining and model size, but without pre-pretraining. Our detection performance also improves with the larger 2B parameter model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.47, + 0.699 + ], + "angle": 0, + "content": "iNat18, as some classes have less than \\( K \\) images, we adapt our setting to consider at most \\( K \\) shots. For each value of \\( K \\), we generate 5 splits of the original dataset using 5 different random seeds and report the mean top-1 accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.701, + 0.471, + 0.838 + ], + "angle": 0, + "content": "We evaluate two protocols for low-shot finetuning – linear classifiers and Adapters [37], both of which keep the entire model parameters frozen and introduce a few trainable parameters. We evaluated multiple Adapters proposed for ViTs – LoRA [38], AdaptFormer [14], and VPT [42]. We found that VPT performed the best while being robust to the choice of hyperparameters, and outperforms linear classifiers for our models. For other works, we report with the best protocol. Full details in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Table 4 shows a comparison with state-of-the-art methods on low-shot IN1k and iNat18, including foundational and self-supervised models. Our models show impressive low-shot performance on both IN1k and iNat18, reaching \\(83.7\\%\\)" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.409, + 0.895, + 0.602 + ], + "angle": 0, + "content": "
MethodDatasetArch.FrameworkAPboxAPmask
Results with different detection framework
CBNetv2 [48]IN21k2 ×Swin-LHTC59.151.0
SwinV2-L [50]IN21kSwinV2-LHTC++58.951.2
Co-DETR [90]IN21kSwin-LCo-DETR58.5-
Methods that pretrain on a detection dataset
Florence [82]FLD-900M + FLOD-9M†CoSwin-HDyHead [19]62.0-
DINO [87]IN21k + O365†Swin-L-63.2-
FocalNet [80]IN21k + O365†FocalNet-HDINO [87]64.2-
Co-DETR [90]IN21k + O365†MixMIM-gCo-DETR64.4-
MAE [46]IN1kViTDet-HCascade58.750.9
SWAG [70]IG-3BViTDet-HCascade55.747.9
MAE→WSPIG-3BViTDet-HCascade57.749.6
MAE→WSPIG-3BViTDet-2BCascade58.050.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.611, + 0.895, + 0.695 + ], + "angle": 0, + "content": "Table 7: Detection and Segmentation on COCO (val2017). We report the detection \\((\\mathrm{AP}^{\\mathrm{box}})\\) and instance segmentation performance \\((\\mathrm{AP}^{\\mathrm{mask}})\\). Our models outperform SWAG which uses the same pretraining, but without pre-pretraining. \\(^\\dagger\\)Large scale detection datasets - using additional detection data like Objects365 has been shown to boost performance by \\(5.6~\\mathrm{mAP}\\) [67]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.895, + 0.777 + ], + "angle": 0, + "content": "and \\(80.3\\%\\) top-1 accuracy with only 10 labeled examples per class, and reaching the highest reported performance with just one labeled example per class on IN1k of \\(62.1\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Zero-shot transfer. Strong foundational models are expected to also have a good open world understanding of visual concepts. To equip our pretrained vision encoders with such capabilities, we utilize LiT [85]. We initialize the text encoder from an XLM-R Large [17] model. Table 5 shows the zero-shot transfer performance of our models on IN1k, and F-101. Our ViT-2B attains \\(81.9\\%\\) accuracy on IN1k, outperforming a similarly sized OpenCLIP model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5490" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.27, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.092, + 0.462, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.473, + 0.311 + ], + "angle": 0, + "content": "Figure 7: Pre-pretraining with MAE significantly boosts performance on detection. Scaling the model size with WSP only does not improve performance on detection. Adding MAE as prepretraining helps MAE \\(\\rightarrow\\) WSP scaling on detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.488 + ], + "angle": 0, + "content": "Our results lag behind other works which pretrain on datasets such as JFT-3B and ALIGN, highlighting the importance of the pretraining dataset for performance. This data-advantage is also observed in the finetuning results discussed in Table 2, where JFT-3B provides best IN1k accuracy. On Food-101 we attain the highest zero-shot transfer accuracy of \\(96.2\\%\\). The performance on the two datasets is not well correlated, further demonstrating the impact the pretraining dataset can have on a particular zero-shot task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.471, + 0.671 + ], + "angle": 0, + "content": "Detection and segmentation. Next, we evaluate models on detection and instance segmentation, on the LVIS [32] (Table 6) and COCO [49] datasets (Table 7). We use the Cascade Mask R-CNN framework [35], with the ViTDet [47] architecture, and initialize the backbone with our pretrained models. For finetuning our models, we start with the hyperparameters from [47] and adapt them for our models, full details are in Appendix B. We also perform a system level comparison with other state-of-the-art works, but note that drawing meaningful conclusions out of this is difficult, on account of the multiple differences in the detection frameworks, model architectures, and datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.471, + 0.854 + ], + "angle": 0, + "content": "On both benchmarks, MAE \\(\\rightarrow\\) WSP considerably outperforms the weakly supervised SWAG [70] model, demonstrating the benefit of our additional pre-pretraining stage. On the long-tailed LVIS dataset, MAE \\(\\rightarrow\\) WSP outperforms MAE IN1k pretraining on detection AP [47], but lags slightly behind on COCO. MAE's strong performance on detection using both IN1k and IG-3B (Figure 2) can potentially be explained by the fact that it is trained to reproduce images with \\(75\\%\\) masking, whereas WSP is only trained to predict one or a few salient objects in an image. Lastly, methods which use additional detection data, such as Objects365 [67] or FLOD-9M [82], have strong detection performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Analyzing detection performance. We further inspect the benefits of pre-pretraining for detection in Figure 7. Unlike other tasks like image / video classification, scaling model" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.093, + 0.692, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.093, + 0.897, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.25, + 0.895, + 0.334 + ], + "angle": 0, + "content": "Figure 8: Dissecting detection performance. LVIS performance based on the size of objects (left), or the class occurrence frequency (right). MAE is stronger than WSP at detecting small objects and frequent classes, and is worse on rare classes. MAE \\(\\rightarrow\\) WSP outperforms WSP on all axes. It is much closer to MAE on smaller objects and rare classes, and also outperforms MAE elsewhere." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.895, + 0.418 + ], + "angle": 0, + "content": "size using WSP pretraining does not improve detection performance. However, adding MAE pre-pretraining provides consistent gains and allows WSP to scale with model size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.895, + 0.554 + ], + "angle": 0, + "content": "We dissect the performance of ViT-L models trained with MAE, WSP, and \\(\\mathrm{MAE} \\rightarrow \\mathrm{WSP}\\) in Figure 8 based on the size of the object, or by the frequency of the object's class. We observe that \\(\\mathrm{MAE} \\rightarrow \\mathrm{WSP}\\) performs better than WSP on all tasks - detecting rare to frequent classes across small to large object sizes. \\(\\mathrm{MAE} \\rightarrow \\mathrm{WSP}\\) improves over MAE at detecting rare objects, presumably because of the diversity in the IG-3B labels, and improves at detecting frequent classes over WSP due to MAE pre-pretraining." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.568, + 0.619, + 0.584 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.895, + 0.895 + ], + "angle": 0, + "content": "We introduced pre-pretraining which is an initial stage in the standard pretrain-then-finetune paradigm. Prepretraining uses MAE and thus, does not need additional supervision and can be conveniently added to web-scale training pipelines. We show that pre-pretraining improves downstream performance on multiple different recognition tasks, improves model convergence, and is overall more efficient than standard weakly-supervised pretraining. Our self-supervised pre-pretraining improves results for models trained with billions of labels, showing that it is a scalable technique that matters even at web-scale. The benefits of using pre-pretraining hold across varying model sizes, and different pretraining data distributions showing that it is a robust technique. Finally, pre-pretraining naturally and successfully combines the two most common pretraining strategies - self-supervised and weakly-supervised learning. Our results suggest that model initialization plays a significant role in the final performance, training dynamics etc. even for web-scale training with billions of parameter updates and labels, and should be further investigated." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5491" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.091, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.156 + ], + "angle": 0, + "content": "[1] Samira Abnar, Mostafa Dehghani, Behnam Neyshabur, and Hanie Sedghi. Exploring the limits of large scale pre-training. arXiv preprint arXiv:2110.02095, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.196 + ], + "angle": 0, + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A video vision transformer. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.197, + 0.472, + 0.251 + ], + "angle": 0, + "content": "[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.251, + 0.472, + 0.318 + ], + "angle": 0, + "content": "[4] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. arXiv preprint arXiv:2301.08243, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.472, + 0.359 + ], + "angle": 0, + "content": "[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.472, + 0.426 + ], + "angle": 0, + "content": "[6] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.472, + 0.468 + ], + "angle": 0, + "content": "[7] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.468, + 0.472, + 0.507 + ], + "angle": 0, + "content": "[8] Daniel Bolya, Sean Foley, James Hays, and Judy Hoffman. Tide: A general toolbox for identifying object detection errors. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.507, + 0.472, + 0.548 + ], + "angle": 0, + "content": "[9] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.472, + 0.602 + ], + "angle": 0, + "content": "[10] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.472, + 0.655 + ], + "angle": 0, + "content": "[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.472, + 0.709 + ], + "angle": 0, + "content": "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.71, + 0.472, + 0.749 + ], + "angle": 0, + "content": "[13] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.472, + 0.805 + ], + "angle": 0, + "content": "[14] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. arXiv preprint arXiv:2205.13535, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.805, + 0.472, + 0.845 + ], + "angle": 0, + "content": "[15] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.472, + 0.885 + ], + "angle": 0, + "content": "[16] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. *ICLR*, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.885, + 0.472, + 0.898 + ], + "angle": 0, + "content": "[17] Alexis Conneau, Kartikay Khandelwal, Naman Goyal," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In ACL, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.148, + 0.893, + 0.188 + ], + "angle": 0, + "content": "[18] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.188, + 0.893, + 0.242 + ], + "angle": 0, + "content": "[19] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.242, + 0.893, + 0.283 + ], + "angle": 0, + "content": "[20] Terrance De Vries, Ishan Misra, Changhan Wang, and Laurens Van der Maaten. Does object recognition work for everyone? In CVPR Workshop, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.283, + 0.893, + 0.322 + ], + "angle": 0, + "content": "[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.322, + 0.893, + 0.376 + ], + "angle": 0, + "content": "[22] Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng, and Trevor Darrell. Decaf: A deep convolutional activation feature for generic visual recognition. In ICML, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.893, + 0.445 + ], + "angle": 0, + "content": "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.445, + 0.893, + 0.485 + ], + "angle": 0, + "content": "[24] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.485, + 0.893, + 0.525 + ], + "angle": 0, + "content": "[25] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.893, + 0.593 + ], + "angle": 0, + "content": "[26] WeiFu Fu, CongChong Nie, Ting Sun, Jun Liu, TianLiang Zhang, and Yong Liu. Lvis challenge track technical report 1st place solution: distribution balanced and boundary refinement for large vocabulary instance segmentation. arXiv preprint arXiv:2111.02668, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.593, + 0.893, + 0.633 + ], + "angle": 0, + "content": "[27] Deepti Ghadiyaram, Du Tran, and Dhruv Mahajan. Large-scale weakly-supervised pre-training for video action recognition. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.893, + 0.687 + ], + "angle": 0, + "content": "[28] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimae: Single model masked pretraining on images and videos. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.893, + 0.728 + ], + "angle": 0, + "content": "[29] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. CVPR, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.728, + 0.893, + 0.822 + ], + "angle": 0, + "content": "[30] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The \"something something\" video database for learning and evaluating visual common sense. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.822, + 0.893, + 0.892 + ], + "angle": 0, + "content": "[31] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5492" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[32] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.173 + ], + "angle": 0, + "content": "[33] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.175, + 0.471, + 0.214 + ], + "angle": 0, + "content": "[34] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.215, + 0.471, + 0.24 + ], + "angle": 0, + "content": "[35] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.241, + 0.471, + 0.294 + ], + "angle": 0, + "content": "[36] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.471, + 0.349 + ], + "angle": 0, + "content": "[37] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.471, + 0.402 + ], + "angle": 0, + "content": "[38] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.471, + 0.442 + ], + "angle": 0, + "content": "[39] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.443, + 0.471, + 0.524 + ], + "angle": 0, + "content": "[40] Badr Youbi Idrissi, Diane Bouchacourt, Randall Balestriero, Ivan Evtimov, Caner Hazirbas, Nicolas Ballas, Pascal Vincent, Michal Drozdzal, David Lopez-Paz, and Mark Ibrahim. Imagenet-x: Understanding model mistakes with factor of variation annotations. arXiv preprint arXiv:2211.01866, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[41] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[42] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.62, + 0.471, + 0.687 + ], + "angle": 0, + "content": "[43] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, AMustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.471, + 0.713 + ], + "angle": 0, + "content": "[44] S Kornblith, J Shlens, and QV Le. Do better imagenet models transfer better? arxiv 2018. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.471, + 0.768 + ], + "angle": 0, + "content": "[45] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.769, + 0.471, + 0.807 + ], + "angle": 0, + "content": "[46] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.808, + 0.471, + 0.848 + ], + "angle": 0, + "content": "[47] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.849, + 0.471, + 0.889 + ], + "angle": 0, + "content": "[48] Tingting Liang, Xiaojie Chu, Yudong Liu, Yongtao Wang, Zhi Tang, Wei Chu, Jingdong Chen, and Haibin Ling. Cbnet: A composite backbone network architecture for object detection." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.815, + 0.106 + ], + "angle": 0, + "content": "IEEE Transactions on Image Processing, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.106, + 0.893, + 0.16 + ], + "angle": 0, + "content": "[49] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.161, + 0.893, + 0.214 + ], + "angle": 0, + "content": "[50] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.215, + 0.893, + 0.241 + ], + "angle": 0, + "content": "[51] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.242, + 0.893, + 0.295 + ], + "angle": 0, + "content": "[52] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.295, + 0.893, + 0.321 + ], + "angle": 0, + "content": "[53] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 1995." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.322, + 0.893, + 0.361 + ], + "angle": 0, + "content": "[54] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.893, + 0.4 + ], + "angle": 0, + "content": "[55] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.402, + 0.893, + 0.525 + ], + "angle": 0, + "content": "[56] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.893, + 0.551 + ], + "angle": 0, + "content": "[57] Karol J Piczak. Esc: Dataset for environmental sound classification. In ACM MM, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.552, + 0.893, + 0.592 + ], + "angle": 0, + "content": "[58] Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization, 1992." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.592, + 0.893, + 0.659 + ], + "angle": 0, + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.893, + 0.7 + ], + "angle": 0, + "content": "[60] Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. Cnn features off-the-shelf: an astounding baseline for recognition. In CVPR Workshops, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.893, + 0.74 + ], + "angle": 0, + "content": "[61] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imageNet? In ICML, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.741, + 0.893, + 0.78 + ], + "angle": 0, + "content": "[62] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.781, + 0.893, + 0.819 + ], + "angle": 0, + "content": "[63] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.82, + 0.893, + 0.889 + ], + "angle": 0, + "content": "[64] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.893, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5493" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[65] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.161, + 0.472, + 0.202 + ], + "angle": 0, + "content": "[66] Vaishaal Shankar, Achal Dave, Rebecca Roelofs, Deva Ramanan, Benjamin Recht, and Ludwig Schmidt. Do image classifiers generalize across time? In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.202, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[67] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.256, + 0.472, + 0.295 + ], + "angle": 0, + "content": "[68] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.296, + 0.472, + 0.35 + ], + "angle": 0, + "content": "[69] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.472, + 0.418 + ], + "angle": 0, + "content": "[70] Mannat Singh, Laura Gustafson, Aaron Adcock, Vinicius de Freitas Reis, Bugra Gedik, Raj Prateek Kosaraju, Dhruv Mahajan, Ross Girshick, Piotr Dólár, and Laurens van der Maaten. Revisiting weakly supervised pre-training of visual perception models. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.472, + 0.459 + ], + "angle": 0, + "content": "[71] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.459, + 0.472, + 0.513 + ], + "angle": 0, + "content": "[72] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.472, + 0.552 + ], + "angle": 0, + "content": "[73] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.472, + 0.607 + ], + "angle": 0, + "content": "[74] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.472, + 0.66 + ], + "angle": 0, + "content": "[75] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.472, + 0.702 + ], + "angle": 0, + "content": "[76] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.472, + 0.756 + ], + "angle": 0, + "content": "[77] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and León Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. JMLR, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.756, + 0.472, + 0.796 + ], + "angle": 0, + "content": "[78] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.796, + 0.472, + 0.836 + ], + "angle": 0, + "content": "[79] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.836, + 0.472, + 0.863 + ], + "angle": 0, + "content": "[80] Jianwei Yang, Chunyuan Li, and Jianfeng Gao. Focal modulation networks. arXiv preprint arXiv:2203.11926, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.863, + 0.472, + 0.89 + ], + "angle": 0, + "content": "[81] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.885, + 0.107 + ], + "angle": 0, + "content": "captioners are image-text foundation models. TMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.107, + 0.894, + 0.173 + ], + "angle": 0, + "content": "[82] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.173, + 0.894, + 0.228 + ], + "angle": 0, + "content": "[83] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.228, + 0.894, + 0.255 + ], + "angle": 0, + "content": "[84] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.255, + 0.894, + 0.307 + ], + "angle": 0, + "content": "[85] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.308, + 0.893, + 0.349 + ], + "angle": 0, + "content": "[86] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.893, + 0.403 + ], + "angle": 0, + "content": "[87] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.403, + 0.893, + 0.43 + ], + "angle": 0, + "content": "[88] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.43, + 0.893, + 0.469 + ], + "angle": 0, + "content": "[89] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.893, + 0.51 + ], + "angle": 0, + "content": "[90] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. arXiv preprint arXiv:2211.12860, 2022." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5494" + } + ] +] \ No newline at end of file diff --git a/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/5b230860-aa61-4f33-9fe7-2407ba62f713_origin.pdf b/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/5b230860-aa61-4f33-9fe7-2407ba62f713_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a3706db239c49046aa11d98b274acb6b4b4ee46 --- /dev/null +++ b/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/5b230860-aa61-4f33-9fe7-2407ba62f713_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14ad6c70e26f805fe35dd2110e4e36431e996fa9a6dec6bd8f520d9379b62500 +size 168773 diff --git a/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/full.md b/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f555cac15b109627a8e4b64bc1ed85d25f989324 --- /dev/null +++ b/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/full.md @@ -0,0 +1,307 @@ +# The effectiveness of MAE pre-pretraining for billion-scale pretraining + +Mannat Singh*,† Quentin Duval* Kalyan Vasudev Alwala* Haoqi Fan Vaibhav Aggarwal Aaron Adcock Armand Joulin Piotr Dólár Christoph Feichtenhofer Ross Girshick Rohit Girdhar Ishan Misra Meta AI + +# Abstract + +This paper revisits the standard pretrain-then-finetune paradigm used in computer vision for visual recognition tasks. Typically, state-of-the-art foundation models are pretrained using large scale (weakly) supervised datasets with billions of images. We introduce an additional pretraining stage that is simple and uses the self-supervised MAE technique to initialize the model. While MAE has only been shown to scale with the size of models, we find that it scales with the size of the training dataset as well. Thus, our MAE-based pre-training scales with both model and data size making it applicable for training foundation models. Pre-training consistently improves both the model convergence and the downstream transfer performance across a range of model scales (millions to billions of parameters), and dataset sizes (millions to billions of images). We measure the effectiveness of pre-training on 10 different visual recognition tasks spanning image classification, video recognition, object detection, low-shot classification and zero-shot recognition. Our largest model achieves new state-of-the-art results on iNaturalist-18 $(91.3\%)$ , 1-shot ImageNet-1k $(62.1\%)$ , and zero-shot transfer on Food-101 $(96.2\%)$ . Our study reveals that model initialization plays a significant role, even for web-scale pretraining with billions of images. + +# 1. Introduction + +The pretrain-then-finetune paradigm in visual recognition has enabled high performance visual recognition models across a range of tasks such as image classification [52, 59, 70], video action recognition [25, 27, 28], object detection [11, 89], 3D etc. Typically, pretraining consists of training a model using a pretraining task on large scale data. The resulting pretrained models learn general purpose visual representations that can be used for a range of target tasks, often with limited labeled data, by transfer learning. + +In this paper, we show that an initial stage of pre + +![](images/76441380b84add5d1f18b755d1ff430c2cacc528d03599debfa5d6b714e0409f.jpg) +Figure 1: MAE pre-pretraining improves performance. Transfer performance of a ViT-L architecture trained with self-supervised pretraining (MAE), weakly supervised pretraining on billions of images (WSP), and our pre-pretraining $(\mathrm{MAE} \rightarrow \mathrm{WSP})$ that initializes the model with MAE and then pretrains with WSP. Pre-pretraining consistently improves performance. + +pretraining before the standard pretraining task can improve vision models across a variety of different tasks. Our method combines two common pretraining tasks in vision: (1) weakly supervised pretraining that uses weak, often noisy, signals such as text or image hashtags as supervision, and (2) self-supervised pretraining that only uses the data without additional supervision. Both forms of pretraining start training with a randomly initialized model and have proven effective at learning general purpose vision models. While there have been attempts to combine both these forms of pretraining [54, 69], they are typically used independently in the pretrain-then-finetune two stage paradigm [23, 70, 84]. + +In this work we explore the combination of self- and weakly-supervised learning in a simple pre-pretraining framework, as follows. We first begin with the Masked Autoencoder (MAE) [33] self-supervised learning technique to pre-pretrain vision models without using any labels. After + +initializing from the pre-pretrained model, we use standard weakly supervised pretraining on billions of images with noisy labels. We perform a large-scale empirical study to measure the effectiveness of pre-pretraining on 10 different visual recognition tasks spanning image classification, video recognition, object detection, low-shot classification and zero-shot recognition. Our study reveals that pre-pretrain initialization improves the performance for the weakly supervised models, and this improvement holds even at billion scale weakly labeled data, and across vision tasks (Figure 1). It also improves the model convergence during pretraining, leading to an efficient way of training large scale vision models. Pre-pretrain further enjoys the computational efficiency of the MAE approach, making it simple and scalable. Finally, we show that by using pre-pretraining, both self-supervised learning and weakly supervised learning can be combined for improved model performance for billion-scale data. + +Pre-pretrain is related to 'intermediate finetuning' [5, 50] which introduces a stage after pretraining to better align the pretrained features with the downstream task using labeled data. In contrast, pre-pretrain serves as a better way to initialize a model before pretraining. Since we leverage MAE for pre-pretraining, we do not need additional information or labels for this stage and can re-use the pretraining data. This makes pre-pretrain convenient and simple to use with existing pretraining datasets. + +Our study on large-scale pre-pretraining reveals that model initialization plays a significant role, even for web-scale pretraining, and pre-pretraining is a simple and promising technique in that direction. In particular, we show that (i) MAE not only scales with model size as shown in [33], but also with the size of the training data (Figure 2). (ii) Pre-pretraining improves both the model convergence and the final downstream performance for different sized models (millions to billions of parameters) trained on different sized datasets (millions to billions of labels). (iii) Using pre-pretraining combines the benefits of both self-supervised learning and large scale weakly-supervised learning, and our models achieve excellent performance on a variety of different visual recognition tasks (Figure 1). Most prominently, our model sets new state-of-the-art results on iNaturalist-18 image classification $(91.3\%)$ , 1-shot ImageNet-1k image classification $(62.1\%)$ , and zero-shot transfer on Food-101 $(96.2\%)$ . + +# 2. Related Work + +Supervised pretraining of transferrable representations on large labeled datasets [21, 43, 63] and employing them for downstream recognition tasks, has emerged as a powerful approach in computer vision. It has spurred rapid progress on various tasks including image classification [22, 57, 60], object detection/segmentation [29, 62], image captioning [45, 81] and video action recognition [13, 24, 68]. While useful, + +such representations are often limited by the scale and diversity of the supervision in the pretraining datasets. Hence, recent work has probed the effectiveness, robustness, and fairness of these representations [1, 20, 40, 44, 61, 66, 74]. Self-supervised pretraining is a promising alternative to learn these representation without relying on large well-labeled datasets. Initial works focused on reconstructions methods [77] before moving to other pretraining tasks such as solving jigsaw puzzles [55], constractive learning [15, 34] or joint embedding approaches [3, 4, 11, 12, 31, 89]. With the advent of Vision Transformers [23], approaches based on reconstructions such as [5, 33, 79] got renewed interest for their simplicity and state of the art performance. Of particular interest to us is MAE [33] for its state of the art performance on many transfer tasks [25, 28, 33, 46, 75] and its computational efficiency. Given the lack of supervision during pretraining, these representations often require significant finetuning to align to downstream tasks. + +Weakly supervised pretraining (WSP) is a middle-ground between supervised and self-supervised pretraining. Instead of ignoring annotations completely as in self-supervised pretraining, or requiring exhaustive labels as in supervised pretraining, WSP relies on the large quantity of "free" annotation available on the internet. These annotations occur as image-text pairs [59, 65], where the text can additionally be processed to produce pseudo labels. Of particular interest to us is the latter, i.e. approaches which leverage multi-label classification on noisy labels [27, 52, 70, 84] which have shown state of the art fine-tuning performance, and at the same time can be adapted using image-text data to gain zero-shot capabilities [85]. In this work, we explore WSP in conjunction with self-supervised pre-pretraining, and show faster convergence and stronger performance. + +# 3. Setup + +Our goal is to empirically study the effectiveness of self-supervised pre-pretraining as a precursor to billion scale weakly supervised pretraining for representation learning. Given the simplicity and efficiency of Masked AutoEncoding (MAE) [33], we leverage it as the self-supervised prepretraining approach. Our study shows that MAE scales with the size of the pretraining dataset and model size, and combining it with weak supervision improves large scale vision models. Additionally, such a combination leads to faster convergence and is a simple, scalable way to learn visual representations at scale. We describe our setup and the approaches in detail next. + +Architecture. We use the Vision Transformer (ViT) [23] architecture as the visual encoder for all our experiments. ViTs employ minimal vision-specific inductive biases combined with the standard transformer architecture [76], and yet have emerged as an architecture of choice for a wide variety of visual and multimodal recognition tasks [2, 28, 84]. We train + +ViT models at various scales in terms of number of parameters, including ViT-B (86M), ViT-L (307M), and ViT-H (632M). We also train on larger 1.9B and 6.5B parameter ViT models, which we call ViT-2B and ViT-6.5B, respectively (Appendix Table 8). As is common practice [23, 84], we train models of sizes ViT-B, ViT-L with a patch size of 16 and larger models with a patch size of 14. We pretrain with a $224 \times 224$ resolution for all models. + +Pre-pretraining (MAE) [33] learns visual representations from image datasets without using any labels. We choose this approach as it is simple to implement and scales very effectively with large ViT model sizes due to patch dropping as described next. MAE randomly masks $75\%$ of an image and trains the model to reconstruct the masked input image by minimizing the pixel reconstruction error. The target pixel values for a given patch are normalized by the mean and standard deviation of all pixels in it. Coupled with the ViT architecture, MAE can be trained by only processing the $25\%$ unmasked image patches. A separate, smaller, decoder is then used to reconstruct the missing part of the input. This asymmetrical design makes training the encoder extremely efficient, allowing for scaling visual encoder sizes. + +Weakly-supervised pretraining (WSP) leverages images with associated 'weak' supervision for training models. In particular, we focus on internet images and use their associated text information as supervision. We convert the text into a discrete set of labels, specifically leveraging hash-tag information [27, 52, 70]. We then use a multi-label classification loss to train models. We refer to this method as WSP. + +$\mathbf{MAE} \rightarrow \mathbf{WSP}$ first trains the encoder using the MAE self-supervised method using only the images. This pretraining stage initializes the model while simultaneously being computationally efficient because of the masking used in MAE. In the second stage, we pretrain the encoder using both the image and associated weak supervision. This combination outperforms using either strategy in isolation, i.e., an MAE model or a weakly supervised model trained from scratch. + +# 4. Experiments + +We empirically evaluate and analyze large scale MAE pre-pretraining using Instagram data on a variety of different visual recognition tasks. We describe the datasets used for pretraining and evaluation, followed by analysis of the pretraining design decisions, and finally the downstream transfer evaluation of our learned representation. + +# 4.1. Datasets and training details + +Pretraining dataset. We use Instagram-3B (IG-3B) a billion-scale multi-label dataset sourced from Instagram (IG). This multi-label dataset contains 28K classes and 3B unique images, resampled to 5B total images, and was produced by running the dataset generation pipeline from SWAG [70] + +
DatasetTask#cls#train#val
ImageNet-1k (IN1k) [64]Image cls.10001M50K
iNaturalist-18 (iNat18) [36]Fine-grained cls.8142437K24K
ImageNetv2 (INv2) [61]Image cls.1000-10K
ImageNet-ReaL (IN-ReaL) [7]Image cls.1000-50K
ObjectNet (ON) [6]Image cls.113-19K
Food-101 (F-101) [9]Image cls.101N/A25K
COCO [49]Obj. det.80118K5K
LVIS [32]Obj. det.1K100K20K
Kinetics-400 (K400) [43]Action cls.400220K20K
Something Something-v2 (SSv2) [30]Action cls.174169K25K
+ +Table 1: Evaluation datasets used to evaluate MAE $\rightarrow$ WSP on image classification, object detection, and video action recognition tasks. The table reports the task, number of classes (#cls), number of training samples (#train), and number of validation samples (#val) for each dataset. + +without modification. Compared to [70], our version of the dataset has $16\%$ fewer images (3.0B vs. 3.6B), but we were able to reproduce the results from [70] with our version. We obtain labels using an automated process wherein we first obtain hashtags from the associated image captions, and then map the hashtags to WordNet synsets following [70]. After this processing, we get the weakly labeled IG dataset that contains images and their associated labels. + +Evaluation datasets. We evaluate MAE $\rightarrow$ WSP on a variety of different downstream visual recognition tasks. To evaluate our model on image classification, we use the standard ImageNet-1k [64] (IN1k) dataset, and also the long-tailed and fine-grained iNaturalist-18 [36] (iNat18) dataset. For object detection and segmentation, we use the popular COCO [49] dataset, and also LVIS [32], a large vocabulary dataset for long tailed object recognition. We evaluate video classification performance using two popular action recognition datasets, Kinetics-400 [43] (K400) and Something Something-v2 [30] (SSv2). For zero-shot transfer, we evaluate on IN1k and Food-101 [9] (F-101). We also evaluate the robustness of our models on test sets which overlap with IN1k classes, specifically ImageNetv2 [61] (INv2), ImageNet-ReaL [7] (IN-ReaL), and ObjectNet [6] (ON). Please see Table 1 for more details. + +MAE pretraining details. We follow [33] to train MAE models on IG-3B without using any labels. We mask $75\%$ of the image for this training and train the model for 1 epoch over the dataset. We follow the same hyperparameters used in [33] for pretraining on IN1k. + +Supervised pretraining details. We train with a supervised cross-entropy loss on IG-3B using the hashtags as labels. This model is trained by default with random weight initialization and we use the training hyperparameters from [70]. + +Using pre-pretraining. When using pre-pretraining, we first train a model from scratch using MAE on the IG dataset. We then use the weights of the MAE encoder and perform supervised pretraining using the cross-entropy loss as described + +![](images/1da6f92a9a9f6a1b08fe1b5ae2983335a664f0cd1ecb53082b91d3f341f15060.jpg) +Figure 2: Scaling MAE with model and dataset size. We plot MAE's performance when pretrained on ImageNet-1k or Instagram-3B and finetuned on downstream tasks. MAE scales to billion parameters sized models using just IN1k pretraining. Larger models show improved scaling behavior when pretrained with the much larger IG-3B dataset. MAE pretrained on IN1k data point is missing for the 2 billion model as training at that scale was unstable on both COCO and LVIS datasets. Tabulated results in Table 19. + +![](images/0ad97bb96c50fc9fd5f9ba5ed3b5cc6cce2192c7030235da98e8835b342be4e2.jpg) + +![](images/c14ccc52e24c3ea920dcf17c4e45039b75658860e5ab962f7efccf21eb1ded12.jpg) + +![](images/bd3d13e7ae81b85391bd5023aad7f6e4c0e97f068a2ff5a12a19051fa123684e.jpg) + +above. We reuse the same hyperparameters and training details as [70], i.e. there is no hyperparameter search needed for MAE $\rightarrow$ WSP, and we train for 1 epoch on IG-3B. + +Zero-shot training and evaluation details. To impart zero shot understanding capabilities to our models, we use the LiT approach from [85]. For LiT, we use the original (image, caption) pairs from the IG-3B dataset. We freeze the image encoder, and train a text encoder to encode the image captions and match the text embeddings to the associated image embedding using a CLIP loss [59]. We train the text encoder for 1 epoch. For evaluation, we follow [59] - we use the text encoder to compute embeddings from the templated text descriptions of classes and use the cosine similarity of the image and text embeddings as the classification score. + +For full training details and hyperparameters, please refer to Appendix A. + +# 4.2. Scaling MAE pretraining to large data + +Since our pre-pretraining uses MAE in the very first stage, we first study how MAE behaves on the large scale IG-3B dataset. We compare the performance of MAE pretraining on the large scale IG-3B with the original MAE [33] models trained on IN1k for 1600 epochs. We train models of varying sizes, from ViT-B to ViT-H as in [33]. To test the scaling behavior further, we also train MAE on the ViT-2B model with 2B parameters. We measure the performance of the resulting models in Figure 2 on four different vision tasks. + +We observe that using the IG-3B data provides consistent gains over IN1k for all vision tasks, and the gain increases for larger models. These experiments show that MAE scales with the size of the pretraining dataset, and benefits from using billions of images from IG-3B. He et al. [33]'s findings were limited to the fact that MAE scales with the size of the model, and thus our findings on MAE scaling with the size + +of the pretraining data are complementary to theirs. + +Our ViT-2B model pretrained on IG-3B improves upon the best results from [33] on image classification, attaining $88.4\%$ on IN1k $(+0.6\%)$ and $89.3\%$ on iNat18 $(+2.5\%)$ at $518\times 518$ resolution. The gains on detection are equally encouraging, with our ViT-2B reaching $53.6\mathrm{AP}^{\mathrm{box}}$ on LVIS $(+2.1$ over [33]) and $59.8\mathrm{AP}^{\mathrm{box}}$ on COCO $(+1.2$ over [33]). Lastly, we highlight the simplicity of our setup, since we use the same hyperparameters as [33] to train MAE on IG-3B, and also for the largest ViT-2B model, without encountering any training instabilities or needing any extra tweaks. + +# 4.3. MAE pre-pretraining + +Given the promising aspects of MAE as a pretraining approach from § 4.2, specifically that MAE (i) trains with larger models and datasets without needing any tuning (ii) shows gains when scaling model and / or dataset size (iii) is efficient to train, we investigate it as a pre-pretraining approach for supervised pretraining (WSP). + +Figure 1 shows the performance of a ViT-L with MAE pretraining, supervised pretraining (WSP), or MAE pre-pretraining followed by supervised pretraining $(\mathrm{MAE}\rightarrow \mathrm{WSP})$ . We see that MAE and WSP have different strengths. MAE has strong performance for object detection, and full finetuned image classification. However, MAE underperforms on tasks where the model is not finetuned, such as linear classifiers, zero-shot, or low-shot classification - situations where WSP performs better. For these evaluations MAE lags behind WSP by more than 10 points, which is why the results for MAE are not visible in Figure 1. For video classification, MAE performs significantly better than WSP on SSv2, but lags behind it on K400. + +$\mathrm{MAE} \rightarrow \mathrm{WSP}$ outperforms either of MAE or WSP pretraining on most evaluations, across image classification, + +![](images/3c7ffcdf7c176b6b90ec756754dbef4b37f2f3a7afb4c1f0561ad6de8cc6c36c.jpg) +IN1k linear probe (accuracy) +Figure 3: MAE pre-pretraining scales with model size. Across model sizes, MAE $\rightarrow$ WSP outperforms a WSP only model, and shows strong scaling behavior. Most notably, a 2B MAE $\rightarrow$ WSP model outperforms a 6.5B WSP model. + +![](images/b4f5842507f9fd3e2f729967dd03b3b95d626d7eb48155e2003ce9d9b0f851a4.jpg) +IN1k linear probe (accuracy) +Figure 4: Varying the number of pre-pretraining epochs used to initialize the model for WSP pretraining. Pre-pretraining leads to improved convergence, providing higher performance using fewer number of WSP pretraining epochs. + +![](images/7738b885c32d02d3ea14e7e4b1fcdc50f6d39747622305acaa76a6eff8d54eb2.jpg) +IN1k linear probe (accuracy) +Figure 5: MAE $\rightarrow$ WSP is more FLOPs efficient than WSP. Across a wide range of training FLOP profiles for a ViT-B computed by varying WSP (and MAE) epochs, MAE $\rightarrow$ WSP outperforms a baseline WSP only model. + +![](images/7f30545ce98bfc625cdb27d47fba29ab20bbae10cddd6d0a9b1b39089cfddb68.jpg) +Figure 6: MAE initialization for medium scale data. Results for a ViT-L trained on IN21k with MAE, WSP, and MAE $\rightarrow$ WSP. MAE pre-pretraining improves WSP results by a wide margin. + +video recognition, zero-shot evaluation, object detection, etc. Given that all baselines are trained at billion-scale, these results show that MAE $\rightarrow$ WSP is a simple yet promising strategy to improve performance while requiring no extra data or tuning. Next, we ablate the key aspects of MAE $\rightarrow$ WSP. + +Effect of model size. In Figure 3 we study the effect of prepretraining compared to random initialization for different model sizes. After initialization, all models are pretrained using WSP on the IG-3B dataset, and we measure the transfer performance on IN1k using linear probing. We observe that MAE pre-pretraining gives consistent gains over the + +WSP baseline across all model sizes, ranging from 86M to 2B parameters. The gains over the WSP baseline increase for larger model sizes showing that pre-pretraining shows promising scaling behavior with model sizes. Notably, a 2B MAE $\rightarrow$ WSP model outperforms a larger 6.5B WSP model. Number of pre-pretraining epochs. We vary the number of MAE pre-pretraining and the number of WSP pretraining epochs to understand their effect on the final recognition performance. We study this in Figure 4. + +Pre-pretraining improves results over the standard pretraining (random initialization, w/o pre-pretraining), and provides large gains with fewer WSP pretraining epochs. Pre-pretraining also leads to faster convergence since even a small amount of pre-pretraining for 0.1 epochs provides improvements. Increasing the epochs of pre-pretraining provide a larger improvement, and the gains saturate at 1 epoch of pre-pretraining. Finally, pre-pretraining's gains do not diminish even after 4 epochs of WSP (20 billion samples) showing the value of pre-pretraining at scale. We also note that these gains are independent of the evaluation protocol, and we observed them with full finetuning on IN1k as well. + +Training efficiency. Figure 5 shows a comparison between WSP and MAE $\rightarrow$ WSP when comparing training FLOPs. For the same training FLOPs, MAE $\rightarrow$ WSP achieves better transfer performance compared to WSP, and is up to $2\times$ more efficient. Pre-pretraining's training efficiency holds over a large $10\times$ compute window. + +Different datasets for pre-pretraining. We also evaluate the performance of MAE $\rightarrow$ WSP when pre-pretraining MAE on the much smaller ImageNet-1k dataset (Appendix Table 21) and find that pre-pretraining remains just as effective. This allows reusing pretrained MAE models in practice. + +
MethodDatasetArch.Res.ImageNet-1kiNat18
IN1kINv2ReaLON
MAE [33]IN1kViT-H44887.8---86.8
SWAG [70]IG-3.6BViT-H51888.681.190.569.586.0
DINOv2 [56]LVD-142MViT-g44888.9----
Florence [82]FLD-900MCoSwin-H51290.1----
ViT [23]JFT-300MViT-H51888.6-90.7--
Scale-ViT [84]JFT-3BViT-L38488.580.490.4--
Scale-ViT [84]JFT-3BViT-G51890.583.390.870.5-
SwinV2 [50]IN-ext-70MSwinV2-G64090.284.0---
CoCa [81]JFT-3B + ALIGNN-1.8BCoCa-2B57691.0----
MAE→WSPIG-3BViT-H51889.382.390.872.690.5
MAE→WSPIG-3BViT-2B51889.783.090.975.891.3
+ +Different datasets for pre-pretraining and pretraining. We investigate the effect of the dataset used in pre-pretraining and pretraining by using IN21k [21] for all methods, including MAE and MAE $\rightarrow$ WSP. Compared to IG-3B, IN21k is more curated, smaller (14M images), and has cleaner labels (21K classes), where each image is labeled with one class from the WordNet synsets [53]. For evaluating zero shot performance, we use the PMD [69] dataset for LiT training. For full details about the hyperparameters, refer to Appendix A. + +Figure 6 compares the performance of MAE, WSP and $\mathrm{MAE} \rightarrow \mathrm{WSP}$ when pretrained on this IN21k dataset. We notice a similar trend as when pretraining on IG-3B where $\mathrm{MAE} \rightarrow \mathrm{WSP}$ outperforms both MAE and WSP. This shows that MAE pre-pretraining works with datasets of different scales and distributions. + +# 4.4. Transfer Evaluation + +We compare with state-of-the-art research on image and video classification, detection and segmentation, low-shot image classification, zero shot transfer, robustness analysis. ImageNet-1k image classification. Table 2 shows the performance of different methods on IN1k. MAE $\rightarrow$ WSP gets the best performance for a ViT-H sized model (89.3%). As we report in the appendix, MAE $\rightarrow$ WSP also achieves the best performance for ViT-L (88.8%). Recent methods such as Scale-ViT [84] are better on IN1k and we hypothesize that this gap stems mainly from the differences in the pretraining datasets (IG-3B vs. JFT-3B). We also compute linear performance using frozen features on IN1k at 224px resolution. Our models produce strong features which outperform other methods with WSP objectives. They also surpass the performance of the self-supervised DINOv2 [56] model optimized to produce strong frozen representations. + +Table 2: Image classification results. We report finetuning results on IN1k and iNat18 and note the pretraining dataset, architecture and finetuning resolution. We also evaluate the performance of our IN1k finetuned models on multiple test sets to measure robustness. Our models are robust and also push the state-of-the-art on the challenging fine-grained and long-tailed iNat18 dataset. + +
MethodDatasetArch.Res.K400SSv2
Florence [82]FLD-900MCoSwin-H38486.5-
SwinV2 [50]IN-ext-70MSwinV2-G38486.8-
CoCa [81]JFT-3B + ALIGNN-1.8BCoCa-2B57688.9-
Results with models pretrained on videos
MaskFeat [78]K400MViT-L22484.3-
MAE [25, 75]K400ViT-L22485.274.0
OmniMAE [28]IN1k + SSv2ViT-L22484.074.2
MAE [25, 75]K400ViT-H22486.6-
MAE→WSPIG-3BViT-L22486.074.4
+ +Table 3: Video classification results on Kinetics-400 and Something Something-v2. Our models generalize well to video action recognition tasks, despite not seeing any videos during pretraining. + +
MethodDatasetArch.IN1k Linear
SWAG [70]IG-3.6BViT-H85.8
OpenCLIP [41]LAION-2BViT-G86.2
DINOv2 [56]LVD-142MViT-g86.5
MAE→WSPIG-3BViT-H87.0
MAE→WSPIG-3BViT-2B88.1
+ +Robustness for image classification. We evaluate the robustness of our models finetuned on IN1k on additional test sets in Table 2 and notice that, despite MAE→WSP being $\sim 1\%$ behind Scale-ViT on IN1k, it is significantly more robust and generalizes better on these additional test sets – MAE→WSP gets the highest reported performance on ImageNet-ReaL and ObjectNet for IN1k finetuned models. + +Generalization in image classification. We evaluate the generalization of our model on additional fine-grained image classification using iNaturalist-18. iNat18 is a challenging long-tailed and fine-grained dataset with images of multiple species of visually similar plants and animals. Our ViT-2B sets a new state-of-the-art result on iNat18 (+4.5% over [33]) Video classification. In Table 3 we investigate how MAE→WSP's pretraining transfers to video action classification on K400 and SSv2. MAE→WSP is competitive with state-of-the-art methods, including ones that pretrain on videos, whereas our models are only pretrained on images. Specifically, our ViT-L gets the highest reported performance on both video datasets. For all video finetuning, we use relative position embeddings [24], which improves our performance by 0.6% on K400 for a ViT-L. Overall, the results indicate the promise of MAE pre-pretraining for building strong video understanding models. + +Low-shot image classification. We evaluate the label efficiency of our models using a few examples per class for finetuning. We use two datasets, IN1k and iNat18, with $K$ shots (labeled examples per class), $K \in \{1,5,10\}$ . For + +
MethodDatasetArch.IN1kiNat18
1-shot5-shot10-shot1-shot5-shot10-shot
Results with different pretraining datasets
CLIP [59]WIT-400MViT-L41.366.271.321.949.058.5
OpenCLIP [41]LAION-2BViT-H44.370.074.926.054.663.7
OpenCLIP [41]LAION-2BViT-G46.372.977.226.355.765.1
Scale-ViT†[84]JFT-3BViT-G-83.084.9---
DINO [12]IN1kViT-B/845.864.669.019.845.955.9
MSN [3]IN1kViT-L/757.172.174.417.038.048.1
MAE [33]IN1kViT-H-57.970.8-48.568.2
SWAG [70]IG-3.6BViT-H59.478.781.030.162.872.3
MAE→WSPIG-3BViT-H57.179.882.531.767.876.1
MAE→WSPIG-3BViT-2B62.181.583.735.572.880.3
+ +Table 4: Low shot image classification. We compare the performance of our models using just a few examples per class for ImageNet-1k and iNaturalist-18. MAE $\rightarrow$ WSP excels at classification even with just 1 example per class. Pretraining on large scale data can outperform techniques designed to work well in the low data regime, like MSN. We evaluate and report results for each technique with the best protocol, except for when the checkpoints are not available $^{\dagger}$ . + +
MethodDatasetArch.Res.IN1kF-101
Results with different pretraining datasets
CLIP [59]WIT-400MViT-L/1433676.293.8
OpenCLIP [41]LAION-2BViT-H22478.092.5
OpenCLIP [41]LAION-2BViT-G22480.192.9
Florence [82]FLD-900MCoSwin-H38483.795.1
Scale-ViT [84]JFT-3BViT-L22480.8-
+ LiT [85]+ ALIGNN-3.6B
Scale-ViT [84]JFT-3BViT-g28885.2-
+ LiT [85]+ ALIGNN-3.6B
CoCa [81]JFT-3B + ALIGNN-1.8BCoCa-2B57686.3-
MAE→WSPIG-3BViT-H22480.895.8
MAE→WSPIG-3BViT-2B22481.996.2
+ +Table 5: Zero shot image classification results. We evaluate zero-shot transfer on IN1k and Food-101. Our models push the state-of-the-art on F-101, while being competitive on IN1k. The best performing models on IN1k train on JFT3B and ALIGN, and the performance on the two datasets is not well correlated, exemplifying the impact of pretraining dataset choice on zero-shot transfer performance. + +
MethodDatasetArch.FrameworkAPboxAPmask
Results with a different detection framework
Winner 2021 [26]IN21kCBNetv2HTC-49.2
MAE [46]IN1kViTDet-HCascade51.546.6
SWAG [70]IG-3BViTDet-HCascade47.142.1
MAE→WSPIG-3BViTDet-HCascade50.845.5
MAE→WSPIG-3BViTDet-2BCascade51.846.1
+ +iNat18, as some classes have less than $K$ images, we adapt our setting to consider at most $K$ shots. For each value of $K$ , we generate 5 splits of the original dataset using 5 different random seeds and report the mean top-1 accuracy. + +We evaluate two protocols for low-shot finetuning – linear classifiers and Adapters [37], both of which keep the entire model parameters frozen and introduce a few trainable parameters. We evaluated multiple Adapters proposed for ViTs – LoRA [38], AdaptFormer [14], and VPT [42]. We found that VPT performed the best while being robust to the choice of hyperparameters, and outperforms linear classifiers for our models. For other works, we report with the best protocol. Full details in Appendix B. + +Table 4 shows a comparison with state-of-the-art methods on low-shot IN1k and iNat18, including foundational and self-supervised models. Our models show impressive low-shot performance on both IN1k and iNat18, reaching $83.7\%$ + +Table 6: Detection and Segmentation results on LVIS (v1 val). We compare MAE $\rightarrow$ WSP with prior work and report the detection $(\mathrm{AP}^{\mathrm{box}})$ and instance segmentation performance $(\mathrm{AP}^{\mathrm{mask}})$ . Our ViT-H significantly outperforms SWAG which uses the same pretraining and model size, but without pre-pretraining. Our detection performance also improves with the larger 2B parameter model. + +
MethodDatasetArch.FrameworkAPboxAPmask
Results with different detection framework
CBNetv2 [48]IN21k2 ×Swin-LHTC59.151.0
SwinV2-L [50]IN21kSwinV2-LHTC++58.951.2
Co-DETR [90]IN21kSwin-LCo-DETR58.5-
Methods that pretrain on a detection dataset
Florence [82]FLD-900M + FLOD-9M†CoSwin-HDyHead [19]62.0-
DINO [87]IN21k + O365†Swin-L-63.2-
FocalNet [80]IN21k + O365†FocalNet-HDINO [87]64.2-
Co-DETR [90]IN21k + O365†MixMIM-gCo-DETR64.4-
MAE [46]IN1kViTDet-HCascade58.750.9
SWAG [70]IG-3BViTDet-HCascade55.747.9
MAE→WSPIG-3BViTDet-HCascade57.749.6
MAE→WSPIG-3BViTDet-2BCascade58.050.1
+ +Table 7: Detection and Segmentation on COCO (val2017). We report the detection $(\mathrm{AP}^{\mathrm{box}})$ and instance segmentation performance $(\mathrm{AP}^{\mathrm{mask}})$ . Our models outperform SWAG which uses the same pretraining, but without pre-pretraining. $^\dagger$ Large scale detection datasets - using additional detection data like Objects365 has been shown to boost performance by $5.6~\mathrm{mAP}$ [67]. + +and $80.3\%$ top-1 accuracy with only 10 labeled examples per class, and reaching the highest reported performance with just one labeled example per class on IN1k of $62.1\%$ . + +Zero-shot transfer. Strong foundational models are expected to also have a good open world understanding of visual concepts. To equip our pretrained vision encoders with such capabilities, we utilize LiT [85]. We initialize the text encoder from an XLM-R Large [17] model. Table 5 shows the zero-shot transfer performance of our models on IN1k, and F-101. Our ViT-2B attains $81.9\%$ accuracy on IN1k, outperforming a similarly sized OpenCLIP model. + +![](images/7998ea378450e1a5619c31d6d7cec1fcc0fae2a261e499bfe783e062ac012b13.jpg) +Figure 7: Pre-pretraining with MAE significantly boosts performance on detection. Scaling the model size with WSP only does not improve performance on detection. Adding MAE as prepretraining helps MAE $\rightarrow$ WSP scaling on detection. + +![](images/55cbfa016d97cc918ca93cc6a3b8d78e9394aab699c8894eb83a2fce446d2b11.jpg) + +Our results lag behind other works which pretrain on datasets such as JFT-3B and ALIGN, highlighting the importance of the pretraining dataset for performance. This data-advantage is also observed in the finetuning results discussed in Table 2, where JFT-3B provides best IN1k accuracy. On Food-101 we attain the highest zero-shot transfer accuracy of $96.2\%$ . The performance on the two datasets is not well correlated, further demonstrating the impact the pretraining dataset can have on a particular zero-shot task. + +Detection and segmentation. Next, we evaluate models on detection and instance segmentation, on the LVIS [32] (Table 6) and COCO [49] datasets (Table 7). We use the Cascade Mask R-CNN framework [35], with the ViTDet [47] architecture, and initialize the backbone with our pretrained models. For finetuning our models, we start with the hyperparameters from [47] and adapt them for our models, full details are in Appendix B. We also perform a system level comparison with other state-of-the-art works, but note that drawing meaningful conclusions out of this is difficult, on account of the multiple differences in the detection frameworks, model architectures, and datasets. + +On both benchmarks, MAE $\rightarrow$ WSP considerably outperforms the weakly supervised SWAG [70] model, demonstrating the benefit of our additional pre-pretraining stage. On the long-tailed LVIS dataset, MAE $\rightarrow$ WSP outperforms MAE IN1k pretraining on detection AP [47], but lags slightly behind on COCO. MAE's strong performance on detection using both IN1k and IG-3B (Figure 2) can potentially be explained by the fact that it is trained to reproduce images with $75\%$ masking, whereas WSP is only trained to predict one or a few salient objects in an image. Lastly, methods which use additional detection data, such as Objects365 [67] or FLOD-9M [82], have strong detection performance. + +Analyzing detection performance. We further inspect the benefits of pre-pretraining for detection in Figure 7. Unlike other tasks like image / video classification, scaling model + +![](images/10731e96835c7150bcae427746d72e196863719bfb9ed7090769ec48cb690214.jpg) +Figure 8: Dissecting detection performance. LVIS performance based on the size of objects (left), or the class occurrence frequency (right). MAE is stronger than WSP at detecting small objects and frequent classes, and is worse on rare classes. MAE $\rightarrow$ WSP outperforms WSP on all axes. It is much closer to MAE on smaller objects and rare classes, and also outperforms MAE elsewhere. + +![](images/da1ddd1635dcf4bd9d9a4a0cd4298368e0b67b371544c267405f860094bb7fe2.jpg) + +size using WSP pretraining does not improve detection performance. However, adding MAE pre-pretraining provides consistent gains and allows WSP to scale with model size. + +We dissect the performance of ViT-L models trained with MAE, WSP, and $\mathrm{MAE} \rightarrow \mathrm{WSP}$ in Figure 8 based on the size of the object, or by the frequency of the object's class. We observe that $\mathrm{MAE} \rightarrow \mathrm{WSP}$ performs better than WSP on all tasks - detecting rare to frequent classes across small to large object sizes. $\mathrm{MAE} \rightarrow \mathrm{WSP}$ improves over MAE at detecting rare objects, presumably because of the diversity in the IG-3B labels, and improves at detecting frequent classes over WSP due to MAE pre-pretraining. + +# 5. Conclusion + +We introduced pre-pretraining which is an initial stage in the standard pretrain-then-finetune paradigm. Prepretraining uses MAE and thus, does not need additional supervision and can be conveniently added to web-scale training pipelines. We show that pre-pretraining improves downstream performance on multiple different recognition tasks, improves model convergence, and is overall more efficient than standard weakly-supervised pretraining. Our self-supervised pre-pretraining improves results for models trained with billions of labels, showing that it is a scalable technique that matters even at web-scale. The benefits of using pre-pretraining hold across varying model sizes, and different pretraining data distributions showing that it is a robust technique. Finally, pre-pretraining naturally and successfully combines the two most common pretraining strategies - self-supervised and weakly-supervised learning. Our results suggest that model initialization plays a significant role in the final performance, training dynamics etc. even for web-scale training with billions of parameter updates and labels, and should be further investigated. + +# References + +[1] Samira Abnar, Mostafa Dehghani, Behnam Neyshabur, and Hanie Sedghi. Exploring the limits of large scale pre-training. arXiv preprint arXiv:2110.02095, 2021. +[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A video vision transformer. In CVPR, 2021. +[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022. +[4] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. arXiv preprint arXiv:2301.08243, 2023. +[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. +[6] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, 2019. +[7] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020. +[8] Daniel Bolya, Sean Foley, James Hays, and Judy Hoffman. Tide: A general toolbox for identifying object detection errors. In ECCV, 2020. +[9] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV. Springer, 2014. +[10] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. +[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. +[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. +[13] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017. +[14] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. arXiv preprint arXiv:2205.13535, 2022. +[15] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. +[16] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. *ICLR*, 2020. +[17] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, + +Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In ACL, 2020. +[18] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR, 2020. +[19] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In CVPR, 2021. +[20] Terrance De Vries, Ishan Misra, Changhan Wang, and Laurens Van der Maaten. Does object recognition work for everyone? In CVPR Workshop, 2019. +[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. +[22] Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng, and Trevor Darrell. Decaf: A deep convolutional activation feature for generic visual recognition. In ICML, 2014. +[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. +[24] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. +[25] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022. +[26] WeiFu Fu, CongChong Nie, Ting Sun, Jun Liu, TianLiang Zhang, and Yong Liu. Lvis challenge track technical report 1st place solution: distribution balanced and boundary refinement for large vocabulary instance segmentation. arXiv preprint arXiv:2111.02668, 2021. +[27] Deepti Ghadiyaram, Du Tran, and Dhruv Mahajan. Large-scale weakly-supervised pre-training for video action recognition. In CVPR, 2019. +[28] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimae: Single model masked pretraining on images and videos. In CVPR, 2023. +[29] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. CVPR, 2013. +[30] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The "something something" video database for learning and evaluating visual common sense. In ICCV, 2017. +[31] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020. + +[32] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. +[33] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. +[34] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. +[35] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. +[36] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. +[37] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019. +[38] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. +[39] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016. +[40] Badr Youbi Idrissi, Diane Bouchacourt, Randall Balestriero, Ivan Evtimov, Caner Hazirbas, Nicolas Ballas, Pascal Vincent, Michal Drozdzal, David Lopez-Paz, and Mark Ibrahim. Imagenet-x: Understanding model mistakes with factor of variation annotations. arXiv preprint arXiv:2211.01866, 2022. +[41] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. +[42] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. +[43] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, AMustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. +[44] S Kornblith, J Shlens, and QV Le. Do better imagenet models transfer better? arxiv 2018. In CVPR, 2019. +[45] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020. +[46] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022. +[47] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022. +[48] Tingting Liang, Xiaojie Chu, Yudong Liu, Yongtao Wang, Zhi Tang, Wei Chu, Jingdong Chen, and Haibin Ling. Cbnet: A composite backbone network architecture for object detection. + +IEEE Transactions on Image Processing, 2022. +[49] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. +[50] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. +[51] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. +[52] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018. +[53] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 1995. +[54] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. +[55] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016. +[56] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. +[57] Karol J Piczak. Esc: Dataset for environmental sound classification. In ACM MM, 2015. +[58] Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization, 1992. +[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. +[60] Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. Cnn features off-the-shelf: an astounding baseline for recognition. In CVPR Workshops, 2014. +[61] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imageNet? In ICML, 2019. +[62] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. +[63] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. In NeurIPS, 2021. +[64] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. + +[65] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. +[66] Vaishaal Shankar, Achal Dave, Rebecca Roelofs, Deva Ramanan, Benjamin Recht, and Ludwig Schmidt. Do image classifiers generalize across time? In ICCV, 2021. +[67] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In ICCV, 2019. +[68] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS, 2014. +[69] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. +[70] Mannat Singh, Laura Gustafson, Aaron Adcock, Vinicius de Freitas Reis, Bugra Gedik, Raj Prateek Kosaraju, Dhruv Mahajan, Ross Girshick, Piotr Dólár, and Laurens van der Maaten. Revisiting weakly supervised pre-training of visual perception models. In CVPR, 2022. +[71] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014. +[72] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. +[73] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. +[74] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. In NeurIPS, 2020. +[75] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022. +[76] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. +[77] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and León Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. JMLR, 2010. +[78] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022. +[79] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. +[80] Jianwei Yang, Chunyuan Li, and Jianfeng Gao. Focal modulation networks. arXiv preprint arXiv:2203.11926, 2022. +[81] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive + +captioners are image-text foundation models. TMLR, 2022. +[82] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. +[83] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. +[84] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, 2022. +[85] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In CVPR, 2022. +[86] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. +[87] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2022. +[88] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, 2020. +[89] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. ICLR, 2022. +[90] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. arXiv preprint arXiv:2211.12860, 2022. \ No newline at end of file diff --git a/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/images.zip b/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7a0581c1b86dc914cc8081b7020c3ecbe9dfe3b0 --- /dev/null +++ b/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85c56d06854cb1f26ac0107e2758a8ac09d9aa4aee7358c2504e3ce928a59c37 +size 552570 diff --git a/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/layout.json b/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..98019b4462a3badc8ac24704311fe8d42ec566ac --- /dev/null +++ b/2023/The Effectiveness of MAE Pre-Pretraining for Billion-Scale Pretraining/layout.json @@ -0,0 +1,9062 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 511, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 511, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 511, + 121 + ], + "type": "text", + "content": "The effectiveness of MAE pre-pretraining for billion-scale pretraining" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 144, + 452, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 144, + 452, + 199 + ], + "spans": [ + { + "bbox": [ + 140, + 144, + 452, + 199 + ], + "type": "text", + "content": "Mannat Singh*,† Quentin Duval* Kalyan Vasudev Alwala* Haoqi Fan Vaibhav Aggarwal Aaron Adcock Armand Joulin Piotr Dólár Christoph Feichtenhofer Ross Girshick Rohit Girdhar Ishan Misra Meta AI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "type": "text", + "content": "This paper revisits the standard pretrain-then-finetune paradigm used in computer vision for visual recognition tasks. Typically, state-of-the-art foundation models are pretrained using large scale (weakly) supervised datasets with billions of images. We introduce an additional pretraining stage that is simple and uses the self-supervised MAE technique to initialize the model. While MAE has only been shown to scale with the size of models, we find that it scales with the size of the training dataset as well. Thus, our MAE-based pre-training scales with both model and data size making it applicable for training foundation models. Pre-training consistently improves both the model convergence and the downstream transfer performance across a range of model scales (millions to billions of parameters), and dataset sizes (millions to billions of images). We measure the effectiveness of pre-training on 10 different visual recognition tasks spanning image classification, video recognition, object detection, low-shot classification and zero-shot recognition. Our largest model achieves new state-of-the-art results on iNaturalist-18 " + }, + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "type": "inline_equation", + "content": "(91.3\\%)" + }, + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "type": "text", + "content": ", 1-shot ImageNet-1k " + }, + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "type": "inline_equation", + "content": "(62.1\\%)" + }, + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "type": "text", + "content": ", and zero-shot transfer on Food-101 " + }, + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "type": "inline_equation", + "content": "(96.2\\%)" + }, + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "type": "text", + "content": ". Our study reveals that model initialization plays a significant role, even for web-scale pretraining with billions of images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 128, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 128, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 128, + 564 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 572, + 288, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 288, + 680 + ], + "type": "text", + "content": "The pretrain-then-finetune paradigm in visual recognition has enabled high performance visual recognition models across a range of tasks such as image classification [52, 59, 70], video action recognition [25, 27, 28], object detection [11, 89], 3D etc. Typically, pretraining consists of training a model using a pretraining task on large scale data. The resulting pretrained models learn general purpose visual representations that can be used for a range of target tasks, often with limited labeled data, by transfer learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 681, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 681, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 681, + 288, + 693 + ], + "type": "text", + "content": "In this paper, we show that an initial stage of pre" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 226, + 544, + 411 + ], + "blocks": [ + { + "bbox": [ + 309, + 226, + 544, + 411 + ], + "lines": [ + { + "bbox": [ + 309, + 226, + 544, + 411 + ], + "spans": [ + { + "bbox": [ + 309, + 226, + 544, + 411 + ], + "type": "image", + "image_path": "76441380b84add5d1f18b755d1ff430c2cacc528d03599debfa5d6b714e0409f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 419, + 548, + 486 + ], + "lines": [ + { + "bbox": [ + 305, + 419, + 548, + 486 + ], + "spans": [ + { + "bbox": [ + 305, + 419, + 548, + 486 + ], + "type": "text", + "content": "Figure 1: MAE pre-pretraining improves performance. Transfer performance of a ViT-L architecture trained with self-supervised pretraining (MAE), weakly supervised pretraining on billions of images (WSP), and our pre-pretraining " + }, + { + "bbox": [ + 305, + 419, + 548, + 486 + ], + "type": "inline_equation", + "content": "(\\mathrm{MAE} \\rightarrow \\mathrm{WSP})" + }, + { + "bbox": [ + 305, + 419, + 548, + 486 + ], + "type": "text", + "content": " that initializes the model with MAE and then pretrains with WSP. Pre-pretraining consistently improves performance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 509, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 652 + ], + "type": "text", + "content": "pretraining before the standard pretraining task can improve vision models across a variety of different tasks. Our method combines two common pretraining tasks in vision: (1) weakly supervised pretraining that uses weak, often noisy, signals such as text or image hashtags as supervision, and (2) self-supervised pretraining that only uses the data without additional supervision. Both forms of pretraining start training with a randomly initialized model and have proven effective at learning general purpose vision models. While there have been attempts to combine both these forms of pretraining [54, 69], they are typically used independently in the pretrain-then-finetune two stage paradigm [23, 70, 84]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "In this work we explore the combination of self- and weakly-supervised learning in a simple pre-pretraining framework, as follows. We first begin with the Masked Autoencoder (MAE) [33] self-supervised learning technique to pre-pretrain vision models without using any labels. After" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 211, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 211, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 211, + 713 + ], + "type": "text", + "content": "*Equal technical contribution. †Project Lead." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5484" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "initializing from the pre-pretrained model, we use standard weakly supervised pretraining on billions of images with noisy labels. We perform a large-scale empirical study to measure the effectiveness of pre-pretraining on 10 different visual recognition tasks spanning image classification, video recognition, object detection, low-shot classification and zero-shot recognition. Our study reveals that pre-pretrain initialization improves the performance for the weakly supervised models, and this improvement holds even at billion scale weakly labeled data, and across vision tasks (Figure 1). It also improves the model convergence during pretraining, leading to an efficient way of training large scale vision models. Pre-pretrain further enjoys the computational efficiency of the MAE approach, making it simple and scalable. Finally, we show that by using pre-pretraining, both self-supervised learning and weakly supervised learning can be combined for improved model performance for billion-scale data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 384 + ], + "type": "text", + "content": "Pre-pretrain is related to 'intermediate finetuning' [5, 50] which introduces a stage after pretraining to better align the pretrained features with the downstream task using labeled data. In contrast, pre-pretrain serves as a better way to initialize a model before pretraining. Since we leverage MAE for pre-pretraining, we do not need additional information or labels for this stage and can re-use the pretraining data. This makes pre-pretrain convenient and simple to use with existing pretraining datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 384, + 289, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 289, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 289, + 598 + ], + "type": "text", + "content": "Our study on large-scale pre-pretraining reveals that model initialization plays a significant role, even for web-scale pretraining, and pre-pretraining is a simple and promising technique in that direction. In particular, we show that (i) MAE not only scales with model size as shown in [33], but also with the size of the training data (Figure 2). (ii) Pre-pretraining improves both the model convergence and the final downstream performance for different sized models (millions to billions of parameters) trained on different sized datasets (millions to billions of labels). (iii) Using pre-pretraining combines the benefits of both self-supervised learning and large scale weakly-supervised learning, and our models achieve excellent performance on a variety of different visual recognition tasks (Figure 1). Most prominently, our model sets new state-of-the-art results on iNaturalist-18 image classification " + }, + { + "bbox": [ + 46, + 384, + 289, + 598 + ], + "type": "inline_equation", + "content": "(91.3\\%)" + }, + { + "bbox": [ + 46, + 384, + 289, + 598 + ], + "type": "text", + "content": ", 1-shot ImageNet-1k image classification " + }, + { + "bbox": [ + 46, + 384, + 289, + 598 + ], + "type": "inline_equation", + "content": "(62.1\\%)" + }, + { + "bbox": [ + 46, + 384, + 289, + 598 + ], + "type": "text", + "content": ", and zero-shot transfer on Food-101 " + }, + { + "bbox": [ + 46, + 384, + 289, + 598 + ], + "type": "inline_equation", + "content": "(96.2\\%)" + }, + { + "bbox": [ + 46, + 384, + 289, + 598 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 609, + 134, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 134, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 134, + 623 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": "Supervised pretraining of transferrable representations on large labeled datasets [21, 43, 63] and employing them for downstream recognition tasks, has emerged as a powerful approach in computer vision. It has spurred rapid progress on various tasks including image classification [22, 57, 60], object detection/segmentation [29, 62], image captioning [45, 81] and video action recognition [13, 24, 68]. While useful," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 287 + ], + "type": "text", + "content": "such representations are often limited by the scale and diversity of the supervision in the pretraining datasets. Hence, recent work has probed the effectiveness, robustness, and fairness of these representations [1, 20, 40, 44, 61, 66, 74]. Self-supervised pretraining is a promising alternative to learn these representation without relying on large well-labeled datasets. Initial works focused on reconstructions methods [77] before moving to other pretraining tasks such as solving jigsaw puzzles [55], constractive learning [15, 34] or joint embedding approaches [3, 4, 11, 12, 31, 89]. With the advent of Vision Transformers [23], approaches based on reconstructions such as [5, 33, 79] got renewed interest for their simplicity and state of the art performance. Of particular interest to us is MAE [33] for its state of the art performance on many transfer tasks [25, 28, 33, 46, 75] and its computational efficiency. Given the lack of supervision during pretraining, these representations often require significant finetuning to align to downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 288, + 548, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 548, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 548, + 468 + ], + "type": "text", + "content": "Weakly supervised pretraining (WSP) is a middle-ground between supervised and self-supervised pretraining. Instead of ignoring annotations completely as in self-supervised pretraining, or requiring exhaustive labels as in supervised pretraining, WSP relies on the large quantity of \"free\" annotation available on the internet. These annotations occur as image-text pairs [59, 65], where the text can additionally be processed to produce pseudo labels. Of particular interest to us is the latter, i.e. approaches which leverage multi-label classification on noisy labels [27, 52, 70, 84] which have shown state of the art fine-tuning performance, and at the same time can be adapted using image-text data to gain zero-shot capabilities [85]. In this work, we explore WSP in conjunction with self-supervised pre-pretraining, and show faster convergence and stronger performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 478, + 351, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 351, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 351, + 491 + ], + "type": "text", + "content": "3. Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 641 + ], + "type": "text", + "content": "Our goal is to empirically study the effectiveness of self-supervised pre-pretraining as a precursor to billion scale weakly supervised pretraining for representation learning. Given the simplicity and efficiency of Masked AutoEncoding (MAE) [33], we leverage it as the self-supervised prepretraining approach. Our study shows that MAE scales with the size of the pretraining dataset and model size, and combining it with weak supervision improves large scale vision models. Additionally, such a combination leads to faster convergence and is a simple, scalable way to learn visual representations at scale. We describe our setup and the approaches in detail next." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Architecture. We use the Vision Transformer (ViT) [23] architecture as the visual encoder for all our experiments. ViTs employ minimal vision-specific inductive biases combined with the standard transformer architecture [76], and yet have emerged as an architecture of choice for a wide variety of visual and multimodal recognition tasks [2, 28, 84]. We train" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5485" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "ViT models at various scales in terms of number of parameters, including ViT-B (86M), ViT-L (307M), and ViT-H (632M). We also train on larger 1.9B and 6.5B parameter ViT models, which we call ViT-2B and ViT-6.5B, respectively (Appendix Table 8). As is common practice [23, 84], we train models of sizes ViT-B, ViT-L with a patch size of 16 and larger models with a patch size of 14. We pretrain with a " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": " resolution for all models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 167, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 289, + 335 + ], + "type": "text", + "content": "Pre-pretraining (MAE) [33] learns visual representations from image datasets without using any labels. We choose this approach as it is simple to implement and scales very effectively with large ViT model sizes due to patch dropping as described next. MAE randomly masks " + }, + { + "bbox": [ + 46, + 167, + 289, + 335 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 167, + 289, + 335 + ], + "type": "text", + "content": " of an image and trains the model to reconstruct the masked input image by minimizing the pixel reconstruction error. The target pixel values for a given patch are normalized by the mean and standard deviation of all pixels in it. Coupled with the ViT architecture, MAE can be trained by only processing the " + }, + { + "bbox": [ + 46, + 167, + 289, + 335 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 167, + 289, + 335 + ], + "type": "text", + "content": " unmasked image patches. A separate, smaller, decoder is then used to reconstruct the missing part of the input. This asymmetrical design makes training the encoder extremely efficient, allowing for scaling visual encoder sizes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 335, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 418 + ], + "type": "text", + "content": "Weakly-supervised pretraining (WSP) leverages images with associated 'weak' supervision for training models. In particular, we focus on internet images and use their associated text information as supervision. We convert the text into a discrete set of labels, specifically leveraging hash-tag information [27, 52, 70]. We then use a multi-label classification loss to train models. We refer to this method as WSP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 418, + 289, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 289, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 289, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{MAE} \\rightarrow \\mathbf{WSP}" + }, + { + "bbox": [ + 46, + 418, + 289, + 525 + ], + "type": "text", + "content": " first trains the encoder using the MAE self-supervised method using only the images. This pretraining stage initializes the model while simultaneously being computationally efficient because of the masking used in MAE. In the second stage, we pretrain the encoder using both the image and associated weak supervision. This combination outperforms using either strategy in isolation, i.e., an MAE model or a weakly supervised model trained from scratch." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 556, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 288, + 628 + ], + "type": "text", + "content": "We empirically evaluate and analyze large scale MAE pre-pretraining using Instagram data on a variety of different visual recognition tasks. We describe the datasets used for pretraining and evaluation, followed by analysis of the pretraining design decisions, and finally the downstream transfer evaluation of our learned representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 635, + 204, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 204, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 204, + 647 + ], + "type": "text", + "content": "4.1. Datasets and training details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Pretraining dataset. We use Instagram-3B (IG-3B) a billion-scale multi-label dataset sourced from Instagram (IG). This multi-label dataset contains 28K classes and 3B unique images, resampled to 5B total images, and was produced by running the dataset generation pipeline from SWAG [70]" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 545, + 178 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 545, + 178 + ], + "type": "table", + "html": "
DatasetTask#cls#train#val
ImageNet-1k (IN1k) [64]Image cls.10001M50K
iNaturalist-18 (iNat18) [36]Fine-grained cls.8142437K24K
ImageNetv2 (INv2) [61]Image cls.1000-10K
ImageNet-ReaL (IN-ReaL) [7]Image cls.1000-50K
ObjectNet (ON) [6]Image cls.113-19K
Food-101 (F-101) [9]Image cls.101N/A25K
COCO [49]Obj. det.80118K5K
LVIS [32]Obj. det.1K100K20K
Kinetics-400 (K400) [43]Action cls.400220K20K
Something Something-v2 (SSv2) [30]Action cls.174169K25K
", + "image_path": "247dc7231748062648a64500245223397462614fb96974a2dc7ee8d8af713715.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 185, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 304, + 185, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 547, + 240 + ], + "type": "text", + "content": "Table 1: Evaluation datasets used to evaluate MAE " + }, + { + "bbox": [ + 304, + 185, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 185, + 547, + 240 + ], + "type": "text", + "content": " WSP on image classification, object detection, and video action recognition tasks. The table reports the task, number of classes (#cls), number of training samples (#train), and number of validation samples (#val) for each dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 265, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 547, + 361 + ], + "type": "text", + "content": "without modification. Compared to [70], our version of the dataset has " + }, + { + "bbox": [ + 304, + 265, + 547, + 361 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 304, + 265, + 547, + 361 + ], + "type": "text", + "content": " fewer images (3.0B vs. 3.6B), but we were able to reproduce the results from [70] with our version. We obtain labels using an automated process wherein we first obtain hashtags from the associated image captions, and then map the hashtags to WordNet synsets following [70]. After this processing, we get the weakly labeled IG dataset that contains images and their associated labels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 361, + 548, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 548, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 548, + 553 + ], + "type": "text", + "content": "Evaluation datasets. We evaluate MAE " + }, + { + "bbox": [ + 304, + 361, + 548, + 553 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 361, + 548, + 553 + ], + "type": "text", + "content": " WSP on a variety of different downstream visual recognition tasks. To evaluate our model on image classification, we use the standard ImageNet-1k [64] (IN1k) dataset, and also the long-tailed and fine-grained iNaturalist-18 [36] (iNat18) dataset. For object detection and segmentation, we use the popular COCO [49] dataset, and also LVIS [32], a large vocabulary dataset for long tailed object recognition. We evaluate video classification performance using two popular action recognition datasets, Kinetics-400 [43] (K400) and Something Something-v2 [30] (SSv2). For zero-shot transfer, we evaluate on IN1k and Food-101 [9] (F-101). We also evaluate the robustness of our models on test sets which overlap with IN1k classes, specifically ImageNetv2 [61] (INv2), ImageNet-ReaL [7] (IN-ReaL), and ObjectNet [6] (ON). Please see Table 1 for more details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "type": "text", + "content": "MAE pretraining details. We follow [33] to train MAE models on IG-3B without using any labels. We mask " + }, + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 554, + 547, + 614 + ], + "type": "text", + "content": " of the image for this training and train the model for 1 epoch over the dataset. We follow the same hyperparameters used in [33] for pretraining on IN1k." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 547, + 665 + ], + "type": "text", + "content": "Supervised pretraining details. We train with a supervised cross-entropy loss on IG-3B using the hashtags as labels. This model is trained by default with random weight initialization and we use the training hyperparameters from [70]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "content": "Using pre-pretraining. When using pre-pretraining, we first train a model from scratch using MAE on the IG dataset. We then use the weights of the MAE encoder and perform supervised pretraining using the cross-entropy loss as described" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5486" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 73, + 170, + 223 + ], + "blocks": [ + { + "bbox": [ + 50, + 73, + 170, + 223 + ], + "lines": [ + { + "bbox": [ + 50, + 73, + 170, + 223 + ], + "spans": [ + { + "bbox": [ + 50, + 73, + 170, + 223 + ], + "type": "image", + "image_path": "1da6f92a9a9f6a1b08fe1b5ae2983335a664f0cd1ecb53082b91d3f341f15060.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 227, + 547, + 271 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 547, + 271 + ], + "type": "text", + "content": "Figure 2: Scaling MAE with model and dataset size. We plot MAE's performance when pretrained on ImageNet-1k or Instagram-3B and finetuned on downstream tasks. MAE scales to billion parameters sized models using just IN1k pretraining. Larger models show improved scaling behavior when pretrained with the much larger IG-3B dataset. MAE pretrained on IN1k data point is missing for the 2 billion model as training at that scale was unstable on both COCO and LVIS datasets. Tabulated results in Table 19." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 72, + 291, + 221 + ], + "blocks": [ + { + "bbox": [ + 173, + 72, + 291, + 221 + ], + "lines": [ + { + "bbox": [ + 173, + 72, + 291, + 221 + ], + "spans": [ + { + "bbox": [ + 173, + 72, + 291, + 221 + ], + "type": "image", + "image_path": "0ad97bb96c50fc9fd5f9ba5ed3b5cc6cce2192c7030235da98e8835b342be4e2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 296, + 72, + 417, + 222 + ], + "blocks": [ + { + "bbox": [ + 296, + 72, + 417, + 222 + ], + "lines": [ + { + "bbox": [ + 296, + 72, + 417, + 222 + ], + "spans": [ + { + "bbox": [ + 296, + 72, + 417, + 222 + ], + "type": "image", + "image_path": "c14ccc52e24c3ea920dcf17c4e45039b75658860e5ab962f7efccf21eb1ded12.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 421, + 72, + 539, + 222 + ], + "blocks": [ + { + "bbox": [ + 421, + 72, + 539, + 222 + ], + "lines": [ + { + "bbox": [ + 421, + 72, + 539, + 222 + ], + "spans": [ + { + "bbox": [ + 421, + 72, + 539, + 222 + ], + "type": "image", + "image_path": "bd3d13e7ae81b85391bd5023aad7f6e4c0e97f068a2ff5a12a19051fa123684e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 303, + 288, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 288, + 338 + ], + "type": "text", + "content": "above. We reuse the same hyperparameters and training details as [70], i.e. there is no hyperparameter search needed for MAE " + }, + { + "bbox": [ + 46, + 303, + 288, + 338 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 303, + 288, + 338 + ], + "type": "text", + "content": " WSP, and we train for 1 epoch on IG-3B." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 339, + 287, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 470 + ], + "type": "text", + "content": "Zero-shot training and evaluation details. To impart zero shot understanding capabilities to our models, we use the LiT approach from [85]. For LiT, we use the original (image, caption) pairs from the IG-3B dataset. We freeze the image encoder, and train a text encoder to encode the image captions and match the text embeddings to the associated image embedding using a CLIP loss [59]. We train the text encoder for 1 epoch. For evaluation, we follow [59] - we use the text encoder to compute embeddings from the templated text descriptions of classes and use the cosine similarity of the image and text embeddings as the classification score." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 471, + 288, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 471, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 288, + 495 + ], + "type": "text", + "content": "For full training details and hyperparameters, please refer to Appendix A." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 502, + 251, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 251, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 251, + 516 + ], + "type": "text", + "content": "4.2. Scaling MAE pretraining to large data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 521, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 629 + ], + "type": "text", + "content": "Since our pre-pretraining uses MAE in the very first stage, we first study how MAE behaves on the large scale IG-3B dataset. We compare the performance of MAE pretraining on the large scale IG-3B with the original MAE [33] models trained on IN1k for 1600 epochs. We train models of varying sizes, from ViT-B to ViT-H as in [33]. To test the scaling behavior further, we also train MAE on the ViT-2B model with 2B parameters. We measure the performance of the resulting models in Figure 2 on four different vision tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "We observe that using the IG-3B data provides consistent gains over IN1k for all vision tasks, and the gain increases for larger models. These experiments show that MAE scales with the size of the pretraining dataset, and benefits from using billions of images from IG-3B. He et al. [33]'s findings were limited to the fact that MAE scales with the size of the model, and thus our findings on MAE scaling with the size" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 303, + 514, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 303, + 514, + 314 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 514, + 314 + ], + "type": "text", + "content": "of the pretraining data are complementary to theirs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "text", + "content": "Our ViT-2B model pretrained on IG-3B improves upon the best results from [33] on image classification, attaining " + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "inline_equation", + "content": "88.4\\%" + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "text", + "content": " on IN1k " + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "inline_equation", + "content": "(+0.6\\%)" + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "inline_equation", + "content": "89.3\\%" + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "text", + "content": " on iNat18 " + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "inline_equation", + "content": "(+2.5\\%)" + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "inline_equation", + "content": "518\\times 518" + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "text", + "content": " resolution. The gains on detection are equally encouraging, with our ViT-2B reaching " + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "inline_equation", + "content": "53.6\\mathrm{AP}^{\\mathrm{box}}" + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "text", + "content": " on LVIS " + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "inline_equation", + "content": "(+2.1" + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "text", + "content": " over [33]) and " + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "inline_equation", + "content": "59.8\\mathrm{AP}^{\\mathrm{box}}" + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "text", + "content": " on COCO " + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "inline_equation", + "content": "(+1.2" + }, + { + "bbox": [ + 304, + 315, + 547, + 435 + ], + "type": "text", + "content": " over [33]). Lastly, we highlight the simplicity of our setup, since we use the same hyperparameters as [33] to train MAE on IG-3B, and also for the largest ViT-2B model, without encountering any training instabilities or needing any extra tweaks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 443, + 430, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 430, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 430, + 455 + ], + "type": "text", + "content": "4.3. MAE pre-pretraining" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 461, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 533 + ], + "type": "text", + "content": "Given the promising aspects of MAE as a pretraining approach from § 4.2, specifically that MAE (i) trains with larger models and datasets without needing any tuning (ii) shows gains when scaling model and / or dataset size (iii) is efficient to train, we investigate it as a pre-pretraining approach for supervised pretraining (WSP)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 689 + ], + "type": "text", + "content": "Figure 1 shows the performance of a ViT-L with MAE pretraining, supervised pretraining (WSP), or MAE pre-pretraining followed by supervised pretraining " + }, + { + "bbox": [ + 304, + 534, + 547, + 689 + ], + "type": "inline_equation", + "content": "(\\mathrm{MAE}\\rightarrow \\mathrm{WSP})" + }, + { + "bbox": [ + 304, + 534, + 547, + 689 + ], + "type": "text", + "content": ". We see that MAE and WSP have different strengths. MAE has strong performance for object detection, and full finetuned image classification. However, MAE underperforms on tasks where the model is not finetuned, such as linear classifiers, zero-shot, or low-shot classification - situations where WSP performs better. For these evaluations MAE lags behind WSP by more than 10 points, which is why the results for MAE are not visible in Figure 1. For video classification, MAE performs significantly better than WSP on SSv2, but lags behind it on K400." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{MAE} \\rightarrow \\mathrm{WSP}" + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": " outperforms either of MAE or WSP pretraining on most evaluations, across image classification," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5487" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 90, + 197, + 232 + ], + "blocks": [ + { + "bbox": [ + 76, + 73, + 182, + 85 + ], + "lines": [ + { + "bbox": [ + 76, + 73, + 182, + 85 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 182, + 85 + ], + "type": "text", + "content": "IN1k linear probe (accuracy)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 90, + 197, + 232 + ], + "lines": [ + { + "bbox": [ + 50, + 90, + 197, + 232 + ], + "spans": [ + { + "bbox": [ + 50, + 90, + 197, + 232 + ], + "type": "image", + "image_path": "3c7ffcdf7c176b6b90ec756754dbef4b37f2f3a7afb4c1f0561ad6de8cc6c36c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 238, + 195, + 305 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 195, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 195, + 305 + ], + "type": "text", + "content": "Figure 3: MAE pre-pretraining scales with model size. Across model sizes, MAE " + }, + { + "bbox": [ + 46, + 238, + 195, + 305 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 238, + 195, + 305 + ], + "type": "text", + "content": " WSP outperforms a WSP only model, and shows strong scaling behavior. Most notably, a 2B MAE " + }, + { + "bbox": [ + 46, + 238, + 195, + 305 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 238, + 195, + 305 + ], + "type": "text", + "content": " WSP model outperforms a 6.5B WSP model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 219, + 88, + 379, + 233 + ], + "blocks": [ + { + "bbox": [ + 251, + 73, + 358, + 85 + ], + "lines": [ + { + "bbox": [ + 251, + 73, + 358, + 85 + ], + "spans": [ + { + "bbox": [ + 251, + 73, + 358, + 85 + ], + "type": "text", + "content": "IN1k linear probe (accuracy)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 219, + 88, + 379, + 233 + ], + "lines": [ + { + "bbox": [ + 219, + 88, + 379, + 233 + ], + "spans": [ + { + "bbox": [ + 219, + 88, + 379, + 233 + ], + "type": "image", + "image_path": "b4f5842507f9fd3e2f729967dd03b3b95d626d7eb48155e2003ce9d9b0f851a4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 239, + 395, + 294 + ], + "lines": [ + { + "bbox": [ + 198, + 239, + 395, + 294 + ], + "spans": [ + { + "bbox": [ + 198, + 239, + 395, + 294 + ], + "type": "text", + "content": "Figure 4: Varying the number of pre-pretraining epochs used to initialize the model for WSP pretraining. Pre-pretraining leads to improved convergence, providing higher performance using fewer number of WSP pretraining epochs." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 402, + 90, + 547, + 234 + ], + "blocks": [ + { + "bbox": [ + 427, + 73, + 534, + 85 + ], + "lines": [ + { + "bbox": [ + 427, + 73, + 534, + 85 + ], + "spans": [ + { + "bbox": [ + 427, + 73, + 534, + 85 + ], + "type": "text", + "content": "IN1k linear probe (accuracy)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 402, + 90, + 547, + 234 + ], + "lines": [ + { + "bbox": [ + 402, + 90, + 547, + 234 + ], + "spans": [ + { + "bbox": [ + 402, + 90, + 547, + 234 + ], + "type": "image", + "image_path": "7738b885c32d02d3ea14e7e4b1fcdc50f6d39747622305acaa76a6eff8d54eb2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 239, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 398, + 239, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 398, + 239, + 547, + 307 + ], + "type": "text", + "content": "Figure 5: MAE " + }, + { + "bbox": [ + 398, + 239, + 547, + 307 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 398, + 239, + 547, + 307 + ], + "type": "text", + "content": " WSP is more FLOPs efficient than WSP. Across a wide range of training FLOP profiles for a ViT-B computed by varying WSP (and MAE) epochs, MAE " + }, + { + "bbox": [ + 398, + 239, + 547, + 307 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 398, + 239, + 547, + 307 + ], + "type": "text", + "content": " WSP outperforms a baseline WSP only model." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 326, + 285, + 511 + ], + "blocks": [ + { + "bbox": [ + 50, + 326, + 285, + 511 + ], + "lines": [ + { + "bbox": [ + 50, + 326, + 285, + 511 + ], + "spans": [ + { + "bbox": [ + 50, + 326, + 285, + 511 + ], + "type": "image", + "image_path": "7f30545ce98bfc625cdb27d47fba29ab20bbae10cddd6d0a9b1b39089cfddb68.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 518, + 288, + 553 + ], + "lines": [ + { + "bbox": [ + 46, + 518, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 288, + 553 + ], + "type": "text", + "content": "Figure 6: MAE initialization for medium scale data. Results for a ViT-L trained on IN21k with MAE, WSP, and MAE " + }, + { + "bbox": [ + 46, + 518, + 288, + 553 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 518, + 288, + 553 + ], + "type": "text", + "content": " WSP. MAE pre-pretraining improves WSP results by a wide margin." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 578, + 288, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 639 + ], + "type": "text", + "content": "video recognition, zero-shot evaluation, object detection, etc. Given that all baselines are trained at billion-scale, these results show that MAE " + }, + { + "bbox": [ + 46, + 578, + 288, + 639 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 578, + 288, + 639 + ], + "type": "text", + "content": " WSP is a simple yet promising strategy to improve performance while requiring no extra data or tuning. Next, we ablate the key aspects of MAE " + }, + { + "bbox": [ + 46, + 578, + 288, + 639 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 578, + 288, + 639 + ], + "type": "text", + "content": " WSP." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 641, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 289, + 714 + ], + "type": "text", + "content": "Effect of model size. In Figure 3 we study the effect of prepretraining compared to random initialization for different model sizes. After initialization, all models are pretrained using WSP on the IG-3B dataset, and we measure the transfer performance on IN1k using linear probing. We observe that MAE pre-pretraining gives consistent gains over the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "text", + "content": "WSP baseline across all model sizes, ranging from 86M to 2B parameters. The gains over the WSP baseline increase for larger model sizes showing that pre-pretraining shows promising scaling behavior with model sizes. Notably, a 2B MAE " + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 327, + 547, + 436 + ], + "type": "text", + "content": " WSP model outperforms a larger 6.5B WSP model. Number of pre-pretraining epochs. We vary the number of MAE pre-pretraining and the number of WSP pretraining epochs to understand their effect on the final recognition performance. We study this in Figure 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 437, + 548, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 548, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 548, + 581 + ], + "type": "text", + "content": "Pre-pretraining improves results over the standard pretraining (random initialization, w/o pre-pretraining), and provides large gains with fewer WSP pretraining epochs. Pre-pretraining also leads to faster convergence since even a small amount of pre-pretraining for 0.1 epochs provides improvements. Increasing the epochs of pre-pretraining provide a larger improvement, and the gains saturate at 1 epoch of pre-pretraining. Finally, pre-pretraining's gains do not diminish even after 4 epochs of WSP (20 billion samples) showing the value of pre-pretraining at scale. We also note that these gains are independent of the evaluation protocol, and we observed them with full finetuning on IN1k as well." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 581, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 547, + 653 + ], + "type": "text", + "content": "Training efficiency. Figure 5 shows a comparison between WSP and MAE " + }, + { + "bbox": [ + 304, + 581, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 581, + 547, + 653 + ], + "type": "text", + "content": " WSP when comparing training FLOPs. For the same training FLOPs, MAE " + }, + { + "bbox": [ + 304, + 581, + 547, + 653 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 581, + 547, + 653 + ], + "type": "text", + "content": " WSP achieves better transfer performance compared to WSP, and is up to " + }, + { + "bbox": [ + 304, + 581, + 547, + 653 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 304, + 581, + 547, + 653 + ], + "type": "text", + "content": " more efficient. Pre-pretraining's training efficiency holds over a large " + }, + { + "bbox": [ + 304, + 581, + 547, + 653 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 304, + 581, + 547, + 653 + ], + "type": "text", + "content": " compute window." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": "Different datasets for pre-pretraining. We also evaluate the performance of MAE " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": " WSP when pre-pretraining MAE on the much smaller ImageNet-1k dataset (Appendix Table 21) and find that pre-pretraining remains just as effective. This allows reusing pretrained MAE models in practice." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5488" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 82, + 331, + 228 + ], + "blocks": [ + { + "bbox": [ + 50, + 82, + 331, + 228 + ], + "lines": [ + { + "bbox": [ + 50, + 82, + 331, + 228 + ], + "spans": [ + { + "bbox": [ + 50, + 82, + 331, + 228 + ], + "type": "table", + "html": "
MethodDatasetArch.Res.ImageNet-1kiNat18
IN1kINv2ReaLON
MAE [33]IN1kViT-H44887.8---86.8
SWAG [70]IG-3.6BViT-H51888.681.190.569.586.0
DINOv2 [56]LVD-142MViT-g44888.9----
Florence [82]FLD-900MCoSwin-H51290.1----
ViT [23]JFT-300MViT-H51888.6-90.7--
Scale-ViT [84]JFT-3BViT-L38488.580.490.4--
Scale-ViT [84]JFT-3BViT-G51890.583.390.870.5-
SwinV2 [50]IN-ext-70MSwinV2-G64090.284.0---
CoCa [81]JFT-3B + ALIGNN-1.8BCoCa-2B57691.0----
MAE→WSPIG-3BViT-H51889.382.390.872.690.5
MAE→WSPIG-3BViT-2B51889.783.090.975.891.3
", + "image_path": "04cdf05af39b8c9623dcfffeb60e3018056e5de2b31e75d30829fdd6f5d65fed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 313, + 288, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 288, + 421 + ], + "type": "text", + "content": "Different datasets for pre-pretraining and pretraining. We investigate the effect of the dataset used in pre-pretraining and pretraining by using IN21k [21] for all methods, including MAE and MAE " + }, + { + "bbox": [ + 46, + 313, + 288, + 421 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 313, + 288, + 421 + ], + "type": "text", + "content": " WSP. Compared to IG-3B, IN21k is more curated, smaller (14M images), and has cleaner labels (21K classes), where each image is labeled with one class from the WordNet synsets [53]. For evaluating zero shot performance, we use the PMD [69] dataset for LiT training. For full details about the hyperparameters, refer to Appendix A." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 493 + ], + "type": "text", + "content": "Figure 6 compares the performance of MAE, WSP and " + }, + { + "bbox": [ + 46, + 422, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\mathrm{MAE} \\rightarrow \\mathrm{WSP}" + }, + { + "bbox": [ + 46, + 422, + 287, + 493 + ], + "type": "text", + "content": " when pretrained on this IN21k dataset. We notice a similar trend as when pretraining on IG-3B where " + }, + { + "bbox": [ + 46, + 422, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\mathrm{MAE} \\rightarrow \\mathrm{WSP}" + }, + { + "bbox": [ + 46, + 422, + 287, + 493 + ], + "type": "text", + "content": " outperforms both MAE and WSP. This shows that MAE pre-pretraining works with datasets of different scales and distributions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 502, + 164, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 164, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 164, + 514 + ], + "type": "text", + "content": "4.4. Transfer Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "text", + "content": "We compare with state-of-the-art research on image and video classification, detection and segmentation, low-shot image classification, zero shot transfer, robustness analysis. ImageNet-1k image classification. Table 2 shows the performance of different methods on IN1k. MAE " + }, + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "text", + "content": " WSP gets the best performance for a ViT-H sized model (89.3%). As we report in the appendix, MAE " + }, + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 521, + 288, + 713 + ], + "type": "text", + "content": " WSP also achieves the best performance for ViT-L (88.8%). Recent methods such as Scale-ViT [84] are better on IN1k and we hypothesize that this gap stems mainly from the differences in the pretraining datasets (IG-3B vs. JFT-3B). We also compute linear performance using frozen features on IN1k at 224px resolution. Our models produce strong features which outperform other methods with WSP objectives. They also surpass the performance of the self-supervised DINOv2 [56] model optimized to produce strong frozen representations." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 338, + 100, + 550, + 221 + ], + "blocks": [ + { + "bbox": [ + 46, + 237, + 334, + 292 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 334, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 334, + 292 + ], + "type": "text", + "content": "Table 2: Image classification results. We report finetuning results on IN1k and iNat18 and note the pretraining dataset, architecture and finetuning resolution. We also evaluate the performance of our IN1k finetuned models on multiple test sets to measure robustness. Our models are robust and also push the state-of-the-art on the challenging fine-grained and long-tailed iNat18 dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 338, + 100, + 550, + 221 + ], + "lines": [ + { + "bbox": [ + 338, + 100, + 550, + 221 + ], + "spans": [ + { + "bbox": [ + 338, + 100, + 550, + 221 + ], + "type": "table", + "html": "
MethodDatasetArch.Res.K400SSv2
Florence [82]FLD-900MCoSwin-H38486.5-
SwinV2 [50]IN-ext-70MSwinV2-G38486.8-
CoCa [81]JFT-3B + ALIGNN-1.8BCoCa-2B57688.9-
Results with models pretrained on videos
MaskFeat [78]K400MViT-L22484.3-
MAE [25, 75]K400ViT-L22485.274.0
OmniMAE [28]IN1k + SSv2ViT-L22484.074.2
MAE [25, 75]K400ViT-H22486.6-
MAE→WSPIG-3BViT-L22486.074.4
", + "image_path": "6f7cecc212925ff8e28d6b3acd37d80c67f3ec8b526a99b6e5649ed148b1592d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 332, + 310, + 520, + 375 + ], + "blocks": [ + { + "bbox": [ + 334, + 229, + 552, + 274 + ], + "lines": [ + { + "bbox": [ + 334, + 229, + 552, + 274 + ], + "spans": [ + { + "bbox": [ + 334, + 229, + 552, + 274 + ], + "type": "text", + "content": "Table 3: Video classification results on Kinetics-400 and Something Something-v2. Our models generalize well to video action recognition tasks, despite not seeing any videos during pretraining." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 310, + 520, + 375 + ], + "lines": [ + { + "bbox": [ + 332, + 310, + 520, + 375 + ], + "spans": [ + { + "bbox": [ + 332, + 310, + 520, + 375 + ], + "type": "table", + "html": "
MethodDatasetArch.IN1k Linear
SWAG [70]IG-3.6BViT-H85.8
OpenCLIP [41]LAION-2BViT-G86.2
DINOv2 [56]LVD-142MViT-g86.5
MAE→WSPIG-3BViT-H87.0
MAE→WSPIG-3BViT-2B88.1
", + "image_path": "99149ae5d25b8867f8b25e12fb2b07b93427bf268fe228dff477e470c55f747e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 378, + 547, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 462 + ], + "type": "text", + "content": "Robustness for image classification. We evaluate the robustness of our models finetuned on IN1k on additional test sets in Table 2 and notice that, despite MAE→WSP being " + }, + { + "bbox": [ + 304, + 378, + 547, + 462 + ], + "type": "inline_equation", + "content": "\\sim 1\\%" + }, + { + "bbox": [ + 304, + 378, + 547, + 462 + ], + "type": "text", + "content": " behind Scale-ViT on IN1k, it is significantly more robust and generalizes better on these additional test sets – MAE→WSP gets the highest reported performance on ImageNet-ReaL and ObjectNet for IN1k finetuned models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 462, + 547, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 666 + ], + "type": "text", + "content": "Generalization in image classification. We evaluate the generalization of our model on additional fine-grained image classification using iNaturalist-18. iNat18 is a challenging long-tailed and fine-grained dataset with images of multiple species of visually similar plants and animals. Our ViT-2B sets a new state-of-the-art result on iNat18 (+4.5% over [33]) Video classification. In Table 3 we investigate how MAE→WSP's pretraining transfers to video action classification on K400 and SSv2. MAE→WSP is competitive with state-of-the-art methods, including ones that pretrain on videos, whereas our models are only pretrained on images. Specifically, our ViT-L gets the highest reported performance on both video datasets. For all video finetuning, we use relative position embeddings [24], which improves our performance by 0.6% on K400 for a ViT-L. Overall, the results indicate the promise of MAE pre-pretraining for building strong video understanding models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Low-shot image classification. We evaluate the label efficiency of our models using a few examples per class for finetuning. We use two datasets, IN1k and iNat18, with " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " shots (labeled examples per class), " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "K \\in \\{1,5,10\\}" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": ". For" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5489" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 87, + 339, + 227 + ], + "blocks": [ + { + "bbox": [ + 49, + 87, + 339, + 227 + ], + "lines": [ + { + "bbox": [ + 49, + 87, + 339, + 227 + ], + "spans": [ + { + "bbox": [ + 49, + 87, + 339, + 227 + ], + "type": "table", + "html": "
MethodDatasetArch.IN1kiNat18
1-shot5-shot10-shot1-shot5-shot10-shot
Results with different pretraining datasets
CLIP [59]WIT-400MViT-L41.366.271.321.949.058.5
OpenCLIP [41]LAION-2BViT-H44.370.074.926.054.663.7
OpenCLIP [41]LAION-2BViT-G46.372.977.226.355.765.1
Scale-ViT†[84]JFT-3BViT-G-83.084.9---
DINO [12]IN1kViT-B/845.864.669.019.845.955.9
MSN [3]IN1kViT-L/757.172.174.417.038.048.1
MAE [33]IN1kViT-H-57.970.8-48.568.2
SWAG [70]IG-3.6BViT-H59.478.781.030.162.872.3
MAE→WSPIG-3BViT-H57.179.882.531.767.876.1
MAE→WSPIG-3BViT-2B62.181.583.735.572.880.3
", + "image_path": "30912965ae66cac80befd575b4e04b34f1a9ca7744361083254d82aee62fa36e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 342, + 81, + 553, + 222 + ], + "blocks": [ + { + "bbox": [ + 46, + 232, + 339, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 339, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 339, + 300 + ], + "type": "text", + "content": "Table 4: Low shot image classification. We compare the performance of our models using just a few examples per class for ImageNet-1k and iNaturalist-18. MAE " + }, + { + "bbox": [ + 46, + 232, + 339, + 300 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 232, + 339, + 300 + ], + "type": "text", + "content": " WSP excels at classification even with just 1 example per class. Pretraining on large scale data can outperform techniques designed to work well in the low data regime, like MSN. We evaluate and report results for each technique with the best protocol, except for when the checkpoints are not available" + }, + { + "bbox": [ + 46, + 232, + 339, + 300 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 46, + 232, + 339, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 81, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 342, + 81, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 342, + 81, + 553, + 222 + ], + "type": "table", + "html": "
MethodDatasetArch.Res.IN1kF-101
Results with different pretraining datasets
CLIP [59]WIT-400MViT-L/1433676.293.8
OpenCLIP [41]LAION-2BViT-H22478.092.5
OpenCLIP [41]LAION-2BViT-G22480.192.9
Florence [82]FLD-900MCoSwin-H38483.795.1
Scale-ViT [84]JFT-3BViT-L22480.8-
+ LiT [85]+ ALIGNN-3.6B
Scale-ViT [84]JFT-3BViT-g28885.2-
+ LiT [85]+ ALIGNN-3.6B
CoCa [81]JFT-3B + ALIGNN-1.8BCoCa-2B57686.3-
MAE→WSPIG-3BViT-H22480.895.8
MAE→WSPIG-3BViT-2B22481.996.2
", + "image_path": "9ae6dafa25c7f370c77040d3cb808efe53f5fe4762e47843a2449b4c7e7899a8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 323, + 282, + 404 + ], + "blocks": [ + { + "bbox": [ + 339, + 228, + 553, + 306 + ], + "lines": [ + { + "bbox": [ + 339, + 228, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 339, + 228, + 553, + 306 + ], + "type": "text", + "content": "Table 5: Zero shot image classification results. We evaluate zero-shot transfer on IN1k and Food-101. Our models push the state-of-the-art on F-101, while being competitive on IN1k. The best performing models on IN1k train on JFT3B and ALIGN, and the performance on the two datasets is not well correlated, exemplifying the impact of pretraining dataset choice on zero-shot transfer performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 323, + 282, + 404 + ], + "lines": [ + { + "bbox": [ + 56, + 323, + 282, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 323, + 282, + 404 + ], + "type": "table", + "html": "
MethodDatasetArch.FrameworkAPboxAPmask
Results with a different detection framework
Winner 2021 [26]IN21kCBNetv2HTC-49.2
MAE [46]IN1kViTDet-HCascade51.546.6
SWAG [70]IG-3BViTDet-HCascade47.142.1
MAE→WSPIG-3BViTDet-HCascade50.845.5
MAE→WSPIG-3BViTDet-2BCascade51.846.1
", + "image_path": "421db912e0c03b654b08fe01c36db1738bd2a251a63c449097ca39cb0d719c18.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 504, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 287, + 553 + ], + "type": "text", + "content": "iNat18, as some classes have less than " + }, + { + "bbox": [ + 46, + 504, + 287, + 553 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 504, + 287, + 553 + ], + "type": "text", + "content": " images, we adapt our setting to consider at most " + }, + { + "bbox": [ + 46, + 504, + 287, + 553 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 504, + 287, + 553 + ], + "type": "text", + "content": " shots. For each value of " + }, + { + "bbox": [ + 46, + 504, + 287, + 553 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 504, + 287, + 553 + ], + "type": "text", + "content": ", we generate 5 splits of the original dataset using 5 different random seeds and report the mean top-1 accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 555, + 288, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 288, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 288, + 663 + ], + "type": "text", + "content": "We evaluate two protocols for low-shot finetuning – linear classifiers and Adapters [37], both of which keep the entire model parameters frozen and introduce a few trainable parameters. We evaluated multiple Adapters proposed for ViTs – LoRA [38], AdaptFormer [14], and VPT [42]. We found that VPT performed the best while being robust to the choice of hyperparameters, and outperforms linear classifiers for our models. For other works, we report with the best protocol. Full details in Appendix B." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Table 4 shows a comparison with state-of-the-art methods on low-shot IN1k and iNat18, including foundational and self-supervised models. Our models show impressive low-shot performance on both IN1k and iNat18, reaching " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "83.7\\%" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 323, + 547, + 476 + ], + "blocks": [ + { + "bbox": [ + 46, + 411, + 288, + 478 + ], + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 478 + ], + "type": "text", + "content": "Table 6: Detection and Segmentation results on LVIS (v1 val). We compare MAE " + }, + { + "bbox": [ + 46, + 411, + 288, + 478 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 411, + 288, + 478 + ], + "type": "text", + "content": " WSP with prior work and report the detection " + }, + { + "bbox": [ + 46, + 411, + 288, + 478 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}^{\\mathrm{box}})" + }, + { + "bbox": [ + 46, + 411, + 288, + 478 + ], + "type": "text", + "content": " and instance segmentation performance " + }, + { + "bbox": [ + 46, + 411, + 288, + 478 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}^{\\mathrm{mask}})" + }, + { + "bbox": [ + 46, + 411, + 288, + 478 + ], + "type": "text", + "content": ". Our ViT-H significantly outperforms SWAG which uses the same pretraining and model size, but without pre-pretraining. Our detection performance also improves with the larger 2B parameter model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 323, + 547, + 476 + ], + "lines": [ + { + "bbox": [ + 309, + 323, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 309, + 323, + 547, + 476 + ], + "type": "table", + "html": "
MethodDatasetArch.FrameworkAPboxAPmask
Results with different detection framework
CBNetv2 [48]IN21k2 ×Swin-LHTC59.151.0
SwinV2-L [50]IN21kSwinV2-LHTC++58.951.2
Co-DETR [90]IN21kSwin-LCo-DETR58.5-
Methods that pretrain on a detection dataset
Florence [82]FLD-900M + FLOD-9M†CoSwin-HDyHead [19]62.0-
DINO [87]IN21k + O365†Swin-L-63.2-
FocalNet [80]IN21k + O365†FocalNet-HDINO [87]64.2-
Co-DETR [90]IN21k + O365†MixMIM-gCo-DETR64.4-
MAE [46]IN1kViTDet-HCascade58.750.9
SWAG [70]IG-3BViTDet-HCascade55.747.9
MAE→WSPIG-3BViTDet-HCascade57.749.6
MAE→WSPIG-3BViTDet-2BCascade58.050.1
", + "image_path": "7cfdd741a986fa2f790085a09520e7b4cd27fbd4f174c529b32c86b8e4ac92b3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 483, + 547, + 550 + ], + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 550 + ], + "type": "text", + "content": "Table 7: Detection and Segmentation on COCO (val2017). We report the detection " + }, + { + "bbox": [ + 304, + 483, + 547, + 550 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}^{\\mathrm{box}})" + }, + { + "bbox": [ + 304, + 483, + 547, + 550 + ], + "type": "text", + "content": " and instance segmentation performance " + }, + { + "bbox": [ + 304, + 483, + 547, + 550 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}^{\\mathrm{mask}})" + }, + { + "bbox": [ + 304, + 483, + 547, + 550 + ], + "type": "text", + "content": ". Our models outperform SWAG which uses the same pretraining, but without pre-pretraining. " + }, + { + "bbox": [ + 304, + 483, + 547, + 550 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 304, + 483, + 547, + 550 + ], + "type": "text", + "content": "Large scale detection datasets - using additional detection data like Objects365 has been shown to boost performance by " + }, + { + "bbox": [ + 304, + 483, + 547, + 550 + ], + "type": "inline_equation", + "content": "5.6~\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 483, + 547, + 550 + ], + "type": "text", + "content": " [67]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 578, + 547, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 547, + 615 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 304, + 578, + 547, + 615 + ], + "type": "inline_equation", + "content": "80.3\\%" + }, + { + "bbox": [ + 304, + 578, + 547, + 615 + ], + "type": "text", + "content": " top-1 accuracy with only 10 labeled examples per class, and reaching the highest reported performance with just one labeled example per class on IN1k of " + }, + { + "bbox": [ + 304, + 578, + 547, + 615 + ], + "type": "inline_equation", + "content": "62.1\\%" + }, + { + "bbox": [ + 304, + 578, + 547, + 615 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "content": "Zero-shot transfer. Strong foundational models are expected to also have a good open world understanding of visual concepts. To equip our pretrained vision encoders with such capabilities, we utilize LiT [85]. We initialize the text encoder from an XLM-R Large [17] model. Table 5 shows the zero-shot transfer performance of our models on IN1k, and F-101. Our ViT-2B attains " + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "inline_equation", + "content": "81.9\\%" + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "content": " accuracy on IN1k, outperforming a similarly sized OpenCLIP model." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5490" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 165, + 194 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 165, + 194 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 165, + 194 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 165, + 194 + ], + "type": "image", + "image_path": "7998ea378450e1a5619c31d6d7cec1fcc0fae2a261e499bfe783e062ac012b13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 289, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 289, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 289, + 246 + ], + "type": "text", + "content": "Figure 7: Pre-pretraining with MAE significantly boosts performance on detection. Scaling the model size with WSP only does not improve performance on detection. Adding MAE as prepretraining helps MAE " + }, + { + "bbox": [ + 46, + 201, + 289, + 246 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 201, + 289, + 246 + ], + "type": "text", + "content": " WSP scaling on detection." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 72, + 282, + 195 + ], + "blocks": [ + { + "bbox": [ + 168, + 72, + 282, + 195 + ], + "lines": [ + { + "bbox": [ + 168, + 72, + 282, + 195 + ], + "spans": [ + { + "bbox": [ + 168, + 72, + 282, + 195 + ], + "type": "image", + "image_path": "55cbfa016d97cc918ca93cc6a3b8d78e9394aab699c8894eb83a2fce446d2b11.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 279, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 386 + ], + "type": "text", + "content": "Our results lag behind other works which pretrain on datasets such as JFT-3B and ALIGN, highlighting the importance of the pretraining dataset for performance. This data-advantage is also observed in the finetuning results discussed in Table 2, where JFT-3B provides best IN1k accuracy. On Food-101 we attain the highest zero-shot transfer accuracy of " + }, + { + "bbox": [ + 46, + 279, + 287, + 386 + ], + "type": "inline_equation", + "content": "96.2\\%" + }, + { + "bbox": [ + 46, + 279, + 287, + 386 + ], + "type": "text", + "content": ". The performance on the two datasets is not well correlated, further demonstrating the impact the pretraining dataset can have on a particular zero-shot task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 388, + 288, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 288, + 531 + ], + "type": "text", + "content": "Detection and segmentation. Next, we evaluate models on detection and instance segmentation, on the LVIS [32] (Table 6) and COCO [49] datasets (Table 7). We use the Cascade Mask R-CNN framework [35], with the ViTDet [47] architecture, and initialize the backbone with our pretrained models. For finetuning our models, we start with the hyperparameters from [47] and adapt them for our models, full details are in Appendix B. We also perform a system level comparison with other state-of-the-art works, but note that drawing meaningful conclusions out of this is difficult, on account of the multiple differences in the detection frameworks, model architectures, and datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 676 + ], + "type": "text", + "content": "On both benchmarks, MAE " + }, + { + "bbox": [ + 46, + 533, + 288, + 676 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 533, + 288, + 676 + ], + "type": "text", + "content": " WSP considerably outperforms the weakly supervised SWAG [70] model, demonstrating the benefit of our additional pre-pretraining stage. On the long-tailed LVIS dataset, MAE " + }, + { + "bbox": [ + 46, + 533, + 288, + 676 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 533, + 288, + 676 + ], + "type": "text", + "content": " WSP outperforms MAE IN1k pretraining on detection AP [47], but lags slightly behind on COCO. MAE's strong performance on detection using both IN1k and IG-3B (Figure 2) can potentially be explained by the fact that it is trained to reproduce images with " + }, + { + "bbox": [ + 46, + 533, + 288, + 676 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 533, + 288, + 676 + ], + "type": "text", + "content": " masking, whereas WSP is only trained to predict one or a few salient objects in an image. Lastly, methods which use additional detection data, such as Objects365 [67] or FLOD-9M [82], have strong detection performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Analyzing detection performance. We further inspect the benefits of pre-pretraining for detection in Figure 7. Unlike other tasks like image / video classification, scaling model" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 423, + 191 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 423, + 191 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 423, + 191 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 423, + 191 + ], + "type": "image", + "image_path": "10731e96835c7150bcae427746d72e196863719bfb9ed7090769ec48cb690214.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 198, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 264 + ], + "type": "text", + "content": "Figure 8: Dissecting detection performance. LVIS performance based on the size of objects (left), or the class occurrence frequency (right). MAE is stronger than WSP at detecting small objects and frequent classes, and is worse on rare classes. MAE " + }, + { + "bbox": [ + 304, + 198, + 547, + 264 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 198, + 547, + 264 + ], + "type": "text", + "content": " WSP outperforms WSP on all axes. It is much closer to MAE on smaller objects and rare classes, and also outperforms MAE elsewhere." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 433, + 73, + 548, + 191 + ], + "blocks": [ + { + "bbox": [ + 433, + 73, + 548, + 191 + ], + "lines": [ + { + "bbox": [ + 433, + 73, + 548, + 191 + ], + "spans": [ + { + "bbox": [ + 433, + 73, + 548, + 191 + ], + "type": "image", + "image_path": "da1ddd1635dcf4bd9d9a4a0cd4298368e0b67b371544c267405f860094bb7fe2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 295, + 547, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 547, + 331 + ], + "type": "text", + "content": "size using WSP pretraining does not improve detection performance. However, adding MAE pre-pretraining provides consistent gains and allows WSP to scale with model size." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 331, + 547, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 547, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 547, + 438 + ], + "type": "text", + "content": "We dissect the performance of ViT-L models trained with MAE, WSP, and " + }, + { + "bbox": [ + 304, + 331, + 547, + 438 + ], + "type": "inline_equation", + "content": "\\mathrm{MAE} \\rightarrow \\mathrm{WSP}" + }, + { + "bbox": [ + 304, + 331, + 547, + 438 + ], + "type": "text", + "content": " in Figure 8 based on the size of the object, or by the frequency of the object's class. We observe that " + }, + { + "bbox": [ + 304, + 331, + 547, + 438 + ], + "type": "inline_equation", + "content": "\\mathrm{MAE} \\rightarrow \\mathrm{WSP}" + }, + { + "bbox": [ + 304, + 331, + 547, + 438 + ], + "type": "text", + "content": " performs better than WSP on all tasks - detecting rare to frequent classes across small to large object sizes. " + }, + { + "bbox": [ + 304, + 331, + 547, + 438 + ], + "type": "inline_equation", + "content": "\\mathrm{MAE} \\rightarrow \\mathrm{WSP}" + }, + { + "bbox": [ + 304, + 331, + 547, + 438 + ], + "type": "text", + "content": " improves over MAE at detecting rare objects, presumably because of the diversity in the IG-3B labels, and improves at detecting frequent classes over WSP due to MAE pre-pretraining." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 449, + 378, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 378, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 378, + 462 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 469, + 547, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 547, + 708 + ], + "type": "text", + "content": "We introduced pre-pretraining which is an initial stage in the standard pretrain-then-finetune paradigm. Prepretraining uses MAE and thus, does not need additional supervision and can be conveniently added to web-scale training pipelines. We show that pre-pretraining improves downstream performance on multiple different recognition tasks, improves model convergence, and is overall more efficient than standard weakly-supervised pretraining. Our self-supervised pre-pretraining improves results for models trained with billions of labels, showing that it is a scalable technique that matters even at web-scale. The benefits of using pre-pretraining hold across varying model sizes, and different pretraining data distributions showing that it is a robust technique. Finally, pre-pretraining naturally and successfully combines the two most common pretraining strategies - self-supervised and weakly-supervised learning. Our results suggest that model initialization plays a significant role in the final performance, training dynamics etc. even for web-scale training with billions of parameter updates and labels, and should be further investigated." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5491" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 711 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "text", + "content": "[1] Samira Abnar, Mostafa Dehghani, Behnam Neyshabur, and Hanie Sedghi. Exploring the limits of large scale pre-training. arXiv preprint arXiv:2110.02095, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 155 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 155 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 155 + ], + "type": "text", + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A video vision transformer. In CVPR, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 156, + 288, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 156, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 53, + 156, + 288, + 198 + ], + "type": "text", + "content": "[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 198, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 288, + 251 + ], + "type": "text", + "content": "[4] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. arXiv preprint arXiv:2301.08243, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 284 + ], + "type": "text", + "content": "[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 284, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 288, + 337 + ], + "type": "text", + "content": "[6] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 337, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 288, + 370 + ], + "type": "text", + "content": "[7] Lucas Beyer, Olivier J Henaff, Alexander Kolesnikov, Xiaohua Zhai, and Aaron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 370, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 370, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 370, + 288, + 401 + ], + "type": "text", + "content": "[8] Daniel Bolya, Sean Foley, James Hays, and Judy Hoffman. Tide: A general toolbox for identifying object detection errors. In ECCV, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 401, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 401, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 401, + 288, + 434 + ], + "type": "text", + "content": "[9] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV. Springer, 2014." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 434, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 476 + ], + "type": "text", + "content": "[10] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 476, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 288, + 518 + ], + "type": "text", + "content": "[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 519, + 288, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 288, + 561 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 288, + 561 + ], + "type": "text", + "content": "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 562, + 288, + 593 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 562, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 48, + 562, + 288, + 593 + ], + "type": "text", + "content": "[13] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 593, + 288, + 637 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 288, + 637 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 288, + 637 + ], + "type": "text", + "content": "[14] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. arXiv preprint arXiv:2205.13535, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 637, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 637, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 637, + 288, + 669 + ], + "type": "text", + "content": "[15] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 669, + 288, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 700 + ], + "type": "text", + "content": "[16] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. *ICLR*, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 700, + 288, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 700, + 288, + 711 + ], + "spans": [ + { + "bbox": [ + 48, + 700, + 288, + 711 + ], + "type": "text", + "content": "[17] Alexis Conneau, Kartikay Khandelwal, Naman Goyal," + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 706 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 116 + ], + "type": "text", + "content": "Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In ACL, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 117, + 546, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 117, + 546, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 546, + 148 + ], + "type": "text", + "content": "[18] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 148, + 546, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 148, + 546, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 546, + 191 + ], + "type": "text", + "content": "[19] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In CVPR, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 191, + 546, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 191, + 546, + 224 + ], + "spans": [ + { + "bbox": [ + 307, + 191, + 546, + 224 + ], + "type": "text", + "content": "[20] Terrance De Vries, Ishan Misra, Changhan Wang, and Laurens Van der Maaten. Does object recognition work for everyone? In CVPR Workshop, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 224, + 546, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 224, + 546, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 224, + 546, + 255 + ], + "type": "text", + "content": "[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 255, + 546, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 546, + 297 + ], + "type": "text", + "content": "[22] Jeff Donahue, Yangqing Jia, Oriol Vinyals, Judy Hoffman, Ning Zhang, Eric Tzeng, and Trevor Darrell. Decaf: A deep convolutional activation feature for generic visual recognition. In ICML, 2014." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 297, + 546, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 546, + 352 + ], + "type": "text", + "content": "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 352, + 546, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 352, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 352, + 546, + 384 + ], + "type": "text", + "content": "[24] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 384, + 546, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 546, + 415 + ], + "type": "text", + "content": "[25] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 415, + 546, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 546, + 469 + ], + "type": "text", + "content": "[26] WeiFu Fu, CongChong Nie, Ting Sun, Jun Liu, TianLiang Zhang, and Yong Liu. Lvis challenge track technical report 1st place solution: distribution balanced and boundary refinement for large vocabulary instance segmentation. arXiv preprint arXiv:2111.02668, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 469, + 546, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 469, + 546, + 501 + ], + "spans": [ + { + "bbox": [ + 307, + 469, + 546, + 501 + ], + "type": "text", + "content": "[27] Deepti Ghadiyaram, Du Tran, and Dhruv Mahajan. Large-scale weakly-supervised pre-training for video action recognition. In CVPR, 2019." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 501, + 546, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 546, + 544 + ], + "type": "text", + "content": "[28] Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Omnimae: Single model masked pretraining on images and videos. In CVPR, 2023." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 544, + 546, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 546, + 576 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 546, + 576 + ], + "type": "text", + "content": "[29] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. CVPR, 2013." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 576, + 546, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 576, + 546, + 651 + ], + "spans": [ + { + "bbox": [ + 307, + 576, + 546, + 651 + ], + "type": "text", + "content": "[30] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The \"something something\" video database for learning and evaluating visual common sense. In ICCV, 2017." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 651, + 546, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 651, + 546, + 706 + ], + "spans": [ + { + "bbox": [ + 307, + 651, + 546, + 706 + ], + "type": "text", + "content": "[31] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5492" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 704 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 287, + 105 + ], + "type": "text", + "content": "[32] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 137 + ], + "type": "text", + "content": "[33] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 138, + 288, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 138, + 288, + 169 + ], + "spans": [ + { + "bbox": [ + 48, + 138, + 288, + 169 + ], + "type": "text", + "content": "[34] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 170, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 170, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 48, + 170, + 288, + 190 + ], + "type": "text", + "content": "[35] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 190, + 288, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 190, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 190, + 288, + 232 + ], + "type": "text", + "content": "[36] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 233, + 288, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 288, + 276 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 288, + 276 + ], + "type": "text", + "content": "[37] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In ICML, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 276, + 288, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 318 + ], + "type": "text", + "content": "[38] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 319, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 288, + 350 + ], + "type": "text", + "content": "[39] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 350, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 415 + ], + "type": "text", + "content": "[40] Badr Youbi Idrissi, Diane Bouchacourt, Randall Balestriero, Ivan Evtimov, Caner Hazirbas, Nicolas Ballas, Pascal Vincent, Michal Drozdzal, David Lopez-Paz, and Mark Ibrahim. Imagenet-x: Understanding model mistakes with factor of variation annotations. arXiv preprint arXiv:2211.01866, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "text", + "content": "[41] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 459, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 288, + 491 + ], + "type": "text", + "content": "[42] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 544 + ], + "type": "text", + "content": "[43] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, AMustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 288, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 564 + ], + "type": "text", + "content": "[44] S Kornblith, J Shlens, and QV Le. Do better imagenet models transfer better? arxiv 2018. In CVPR, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 565, + 288, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 288, + 608 + ], + "type": "text", + "content": "[45] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In ECCV, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 609, + 288, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 609, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 48, + 609, + 288, + 639 + ], + "type": "text", + "content": "[46] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 639, + 288, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 639, + 288, + 671 + ], + "spans": [ + { + "bbox": [ + 48, + 639, + 288, + 671 + ], + "type": "text", + "content": "[47] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In ECCV, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 672, + 288, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 672, + 288, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 672, + 288, + 704 + ], + "type": "text", + "content": "[48] Tingting Liang, Xiaojie Chu, Yudong Liu, Yongtao Wang, Zhi Tang, Wei Chu, Jingdong Chen, and Haibin Ling. Cbnet: A composite backbone network architecture for object detection." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 546, + 704 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 327, + 74, + 498, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 498, + 83 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 498, + 83 + ], + "type": "text", + "content": "IEEE Transactions on Image Processing, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 83, + 546, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 83, + 546, + 126 + ], + "spans": [ + { + "bbox": [ + 308, + 83, + 546, + 126 + ], + "type": "text", + "content": "[49] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 127, + 546, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 127, + 546, + 169 + ], + "spans": [ + { + "bbox": [ + 308, + 127, + 546, + 169 + ], + "type": "text", + "content": "[50] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 170, + 546, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 170, + 546, + 190 + ], + "spans": [ + { + "bbox": [ + 308, + 170, + 546, + 190 + ], + "type": "text", + "content": "[51] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 191, + 546, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 191, + 546, + 233 + ], + "spans": [ + { + "bbox": [ + 308, + 191, + 546, + 233 + ], + "type": "text", + "content": "[52] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In ECCV, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 233, + 546, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 233, + 546, + 254 + ], + "spans": [ + { + "bbox": [ + 308, + 233, + 546, + 254 + ], + "type": "text", + "content": "[53] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 1995." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 255, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 546, + 285 + ], + "type": "text", + "content": "[54] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 286, + 546, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 546, + 316 + ], + "type": "text", + "content": "[55] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 318, + 546, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 318, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 318, + 546, + 415 + ], + "type": "text", + "content": "[56] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 416, + 546, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 546, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 546, + 436 + ], + "type": "text", + "content": "[57] Karol J Piczak. Esc: Dataset for environmental sound classification. In ACM MM, 2015." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 437, + 546, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 546, + 468 + ], + "type": "text", + "content": "[58] Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization, 1992." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 468, + 546, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 546, + 521 + ], + "type": "text", + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 522, + 546, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 546, + 554 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 546, + 554 + ], + "type": "text", + "content": "[60] Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. Cnn features off-the-shelf: an astounding baseline for recognition. In CVPR Workshops, 2014." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 554, + 546, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 546, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 546, + 586 + ], + "type": "text", + "content": "[61] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize to imageNet? In ICML, 2019." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 586, + 546, + 617 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 586, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 308, + 586, + 546, + 617 + ], + "type": "text", + "content": "[62] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 618, + 546, + 648 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 618, + 546, + 648 + ], + "spans": [ + { + "bbox": [ + 308, + 618, + 546, + 648 + ], + "type": "text", + "content": "[63] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik-Manor. Imagenet-21k pretraining for the masses. In NeurIPS, 2021." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 649, + 546, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 649, + 546, + 704 + ], + "spans": [ + { + "bbox": [ + 308, + 649, + 546, + 704 + ], + "type": "text", + "content": "[64] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015." + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5493" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 704 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 287, + 127 + ], + "type": "text", + "content": "[65] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 127, + 288, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 127, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 127, + 288, + 159 + ], + "type": "text", + "content": "[66] Vaishaal Shankar, Achal Dave, Rebecca Roelofs, Deva Ramanan, Benjamin Recht, and Ludwig Schmidt. Do image classifiers generalize across time? In ICCV, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 159, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 159, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 159, + 288, + 201 + ], + "type": "text", + "content": "[67] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In ICCV, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 202, + 288, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 202, + 288, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 202, + 288, + 233 + ], + "type": "text", + "content": "[68] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS, 2014." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 234, + 288, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 234, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 234, + 288, + 277 + ], + "type": "text", + "content": "[69] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 288, + 331 + ], + "type": "text", + "content": "[70] Mannat Singh, Laura Gustafson, Aaron Adcock, Vinicius de Freitas Reis, Bugra Gedik, Raj Prateek Kosaraju, Dhruv Mahajan, Ross Girshick, Piotr Dólár, and Laurens van der Maaten. Revisiting weakly supervised pre-training of visual perception models. In CVPR, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 363 + ], + "type": "text", + "content": "[71] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 363, + 288, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 363, + 288, + 406 + ], + "spans": [ + { + "bbox": [ + 48, + 363, + 288, + 406 + ], + "type": "text", + "content": "[72] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 406, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 288, + 437 + ], + "type": "text", + "content": "[73] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "type": "text", + "content": "[74] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. In NeurIPS, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 480, + 288, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 288, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 288, + 522 + ], + "type": "text", + "content": "[75] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 523, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 555 + ], + "type": "text", + "content": "[76] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 288, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 598 + ], + "type": "text", + "content": "[77] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and León Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. JMLR, 2010." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 598, + 288, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 598, + 288, + 630 + ], + "spans": [ + { + "bbox": [ + 48, + 598, + 288, + 630 + ], + "type": "text", + "content": "[78] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 630, + 288, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 630, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 48, + 630, + 288, + 662 + ], + "type": "text", + "content": "[79] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 662, + 288, + 683 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 662, + 288, + 683 + ], + "spans": [ + { + "bbox": [ + 48, + 662, + 288, + 683 + ], + "type": "text", + "content": "[80] Jianwei Yang, Chunyuan Li, and Jianfeng Gao. Focal modulation networks. arXiv preprint arXiv:2203.11926, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 683, + 288, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 683, + 288, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 683, + 288, + 704 + ], + "type": "text", + "content": "[81] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 403 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 541, + 84 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 541, + 84 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 541, + 84 + ], + "type": "text", + "content": "captioners are image-text foundation models. TMLR, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 84, + 547, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 84, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 308, + 84, + 547, + 137 + ], + "type": "text", + "content": "[82] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 137, + 547, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 137, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 308, + 137, + 547, + 180 + ], + "type": "text", + "content": "[83] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 180, + 547, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 180, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 308, + 180, + 547, + 201 + ], + "type": "text", + "content": "[84] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In CVPR, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 201, + 547, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 201, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 547, + 243 + ], + "type": "text", + "content": "[85] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In CVPR, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 243, + 546, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 546, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 546, + 276 + ], + "type": "text", + "content": "[86] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 276, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 546, + 319 + ], + "type": "text", + "content": "[87] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 319, + 546, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 546, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 546, + 340 + ], + "type": "text", + "content": "[88] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 340, + 546, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 340, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 340, + 546, + 371 + ], + "type": "text", + "content": "[89] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. ICLR, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 371, + 546, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 546, + 403 + ], + "type": "text", + "content": "[90] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. arXiv preprint arXiv:2211.12860, 2022." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5494" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/e705b187-21ba-4812-b0e4-1feb752f1cdd_content_list.json b/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/e705b187-21ba-4812-b0e4-1feb752f1cdd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..471980cf0bed6d67ba0acb392bc7c25ac184adab --- /dev/null +++ b/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/e705b187-21ba-4812-b0e4-1feb752f1cdd_content_list.json @@ -0,0 +1,1826 @@ +[ + { + "type": "text", + "text": "The Euclidean Space is Evil: Hyperbolic Attribute Editing for Few-shot Image Generation", + "text_level": 1, + "bbox": [ + 192, + 130, + 779, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lingxiao Li1 Yi Zhang2 Shuhui Wang3*", + "bbox": [ + 308, + 202, + 665, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Columbia University 2 University of Oxford", + "bbox": [ + 292, + 237, + 676, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Institute of Computing Technology, Chinese Academy of Sciences", + "bbox": [ + 215, + 256, + 751, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "113504@columbia.edu, wolf5965@ox.ac.uk, wangshuhui@ict.ac.cn", + "bbox": [ + 218, + 277, + 750, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot image generation is a challenging task since it aims to generate diverse new images for an unseen category with only a few images. Existing methods suffer from the trade-off between the quality and diversity of generated images. To tackle this problem, we propose Hyperbolic Attribute Editing (HAE), a simple yet effective method. Unlike other methods that work in Euclidean space, HAE captures the hierarchy among images using data from seen categories in hyperbolic space. Given a well-trained HAE, images of unseen categories can be generated by moving the latent code of a given image toward any meaningful directions in the Poincaré disk with a fixing radius. Most importantly, the hyperbolic space allows us to control the semantic diversity of the generated images by setting different radii in the disk. Extensive experiments and visualizations demonstrate that HAE is capable of not only generating images with promising quality and diversity using limited data but achieving a highly controllable and interpretable editing process. Code is available at https://github.com/lingxiao-li/HAE.", + "bbox": [ + 75, + 358, + 473, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 686, + 209, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to the persistent development of deep learning, the task of image generation has received significant research attention in recent years. Specifically, the Generative Adversarial Networks (GANs) [21] and its variants (e.g., StyleGANv2 [34]) have succeeded in generating high-fidelity and realistic images, requiring a large number of high-quality data for model training. However, considering the long-tail distribution and data imbalance widely exists among different image categories [30], it is difficult for GANs to be trained on categories with sufficient training images to generate new realistic and diverse images for a", + "bbox": [ + 75, + 712, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/798e260bd080ccee3be0c563ba81492de9d13b09e3997c3ed9d29f1b17138eb7.jpg", + "image_caption": [ + "Figure 1: Illustration of hierarchical attribute editing in hyperbolic space. Hyperbolic space can naturally and compactly encode semantic hierarchical structures within a large image data corpus. Changing the high-level, i.e., category-relevant attribute $\\Delta \\mathbf{w}^r$ changes the category of an image. While changing low-level or category-irrelevant attribute $\\Delta \\mathbf{w}^{ir}$ varies images within categories." + ], + "image_footnote": [], + "bbox": [ + 504, + 325, + 893, + 463 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "category with only a few images. This task is referred to as few-shot image generation [10, 26, 30, 28, 29, 27, 15]. A variety of tasks can benefit from improvements in few-shot image generation, for instance, low-data detection [17] and few-shot classification [54, 57].", + "bbox": [ + 496, + 594, + 893, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In general, existing GAN-based few-shot image generation mechanisms can be classified into three categories. Transfer-based methods [10, 39] introduce meta-learning or domain adaptation on GANs to generate new images by enforcing knowledge transfer among categories. Fusion-based methods [2, 23, 30, 28] perform feature fusion of multiple input images in a feature space and generate images via decoding the fused features back to image space. However, the output is still highly similar to the source images. Transformation-based methods [1, 29, 27, 15] find intra-category transformations or inject random perturbations to conditional unseen category samples to generate images without tedious fine-tuning. By representing the images in the Euclidean feature space, the above learning mechanisms tend to be over-complicated, and the generated", + "bbox": [ + 496, + 674, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22714", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "images are often collapsed due to limited diversity.", + "bbox": [ + 76, + 90, + 413, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Similar to the ubiquity of hierarchies in language [44, 56, 13], the semantic hierarchy is also common in images [35, 11]. As Fig. 1 shows, the semantic hierarchies constructed in the language domain can be instantiated with visual images. From the visual perspective, an image can be regarded as a collection of attributes of multiple levels. High-level attributes, a.k.a. category-relevant attributes, define the category of an image, such as the shape and color of an animal [15]. For instance, in the middle row of Fig. 1, changing the high-level attributes of the given image of a Shih-Tzu dog, the category can be changed to a Rhodesian Ridgeback Dog. While the low-level or fine-grained attributes, including expressions, postures, etc., that vary within the category as shown at the bottom of Fig. 1, are called category-irrelevant attributes. Therefore, an image can also be viewed as a descendant of another image with the same category-relevant attributes by adding fine-grained category-irrelevant attributes to its parent image. To edit the visual attributes for high-quality image generation, it is crucial to capture the attribute hierarchy within the large image data corpus and find a good representation space. Ideally, we aim to construct a hierarchical visual representation in a latent space that allows us to change the category of an image by moving the latent code in a category-relevant direction, and perform few-shot image generation by moving the code in a category-irrelevant direction.", + "bbox": [ + 75, + 109, + 472, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unfortunately, the Euclidean space and its corresponding distance metrics used by existing GAN-based methods can not facilitate the hierarchical attribute representation, thus the design of complicated attribute disentangling and editing mechanisms seems to be crucial for the generation quality. Inspired by the application of hyperbolic space in images [35] and videos [55], we found that the metrics introduced in hyperbolic geometry can naturally and compactly encode hierarchical structures. Unlike the general affine spaces, e.g., the Euclidean space, hyperbolic spaces can be viewed as the continuous analog of a tree since tree-like graphs can be embedded in finite-dimension with minimal distortion [44]. This property of hyperbolic space provides continuous and up to infinite semantic levels for attribute editing, allowing us to robustly generate diverse images with only a few images from unseen categories with simple operations.", + "bbox": [ + 75, + 505, + 472, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above findings, we propose a simple but effective Hyperbolic Attribute Editing (HAE) method for few-shot image generation. Our method is based on the observation that hierarchical latent code manipulation can be easily implemented in Hyperbolic space. The core of HAE is mapping the latent vectors from the Euclidean space $\\mathbb{R}^n$ to a hyperbolic space $\\mathbb{D}^n$ . We minimize a supervised classification loss function to ensure the images are hierarchically embedded in hyperbolic space. Once we capture the", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "attribute hierarchy among images, we can generate new images of unseen categories by moving the latent code from one leaf to another with the same parents by fixing the radius. Most importantly, the hyperbolic space allows us to control the semantic diversity of generated images by setting different radii in the Poincaré disk. Those operations can well facilitate continually hierarchical attribute editing in hyperbolic space for flexible few-shot image generation with both quality and diversity.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 517, + 227, + 844, + 242 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a simple yet effective method for few-shot image generation, i.e., hyperbolic attribute editing. In order to capture the hierarchy among images, we use hyperbolic space as the latent space. To the best of our knowledge, HAE is the first attempt to use hyperbolic latent spaces for few-shot image generation.", + "- We show that in our designed hyperbolic latent space, the semantic hierarchical attribute relations among images can be reflected by their distances to the center of the Poincaré disk.", + "- Extensive experiments and visualization suggest that HAE achieves stable few-shot image generation with state-of-the-art quality and diversity. Unlike other few-shot image generation methods, HAE allows us to generate images with better control of diversity by changing the semantic levels of attributes we want to edit." + ], + "bbox": [ + 517, + 253, + 890, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 542, + 640, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot image generation. Recently, diverse methods have been proposed for few-shot image generation. The transfer-based methods [10, 39] which introduce meta-learning or domain adaptation on GANs can hardly generate realistic images. While fusion-based methods that fuse the features by matching the random vector with the conditional images [28] or formulating the problem as a conditional generating task [23, 30] suffer from the limited diversity of generated images. Furthermore, transformation-based methods [1, 29, 27, 15] can generate images with only one conditional image by focusing on either capturing the cross-category or intra-category transformations by injecting random perturbations [1]. Nevertheless, the transformation captured by those methods is not very consistent. Ding et al. [15, 14] propose the \"editing-based\" perspective, the intra-category transformation can be modeled as category-irrelevant image editing based on one sample instead of pairs of samples. Most recently, Zhu et al. [60] fine-tune powerful diffusion models (DMs) [25] pre-trained on large source domains on limited target data to generate diverse and high quality images. DMs outperform GANs [21] on sample quality with a more controllable training process", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22715", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9372c400b973e3f1a5586f6beafe464dec6f37c4b21f87584f07c83e34f39aef.jpg", + "image_caption": [ + "Figure 2: The overview of HAE. The Hyper layer is a hyperbolic feedforward layer called Möbius linear layer which is used to project the latent code from Euclidean space $\\mathbb{R}^n$ to hyperbolic space $\\mathbb{D}^n$ [7]. $\\bar{z}_{\\mathbb{D}i}$ can be viewed as the \"parent\" or average code of $z_{\\mathbb{D}i}$ and $z_{\\mathbb{D}i}'$ . One can generate diverse images without changing the category by moving the latent code from one child to another of the same parent in the hyperbolic space." + ], + "image_footnote": [], + "bbox": [ + 86, + 97, + 883, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "at the cost less flexibility and editability, since they denoise images in the image space rather than operate in the latent space. Furthermore, the inference process of DMs is much slower than GANs [52].", + "bbox": [ + 75, + 393, + 470, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hyperbolic Embedding. The use of hyperbolic space in deep learning [44, 45, 56, 55, 35] is a pioneering work in recent years. It was first used in natural language processing for hierarchical language representation [44, 45, 56]. The Riemannian optimization algorithms are used to optimize models in hyperbolic space [5, 3]. As hyperbolic space is successfully applied to represent hierarchical data, Ganea et al. [18] derives hyperbolic versions of tools in neural networks including multinomial logistic regression, feed-forward, and recurrent neural networks. Following this, hyperbolic geometry is used in image [35], video [55], and graph data [7, 47]. Most recently, Lazcano et al [36] shows that hyperbolic space outperforms traditional Euclidean space in image generation using HGAN. However, the hierarchy and controllability of hyperbolic space remain uninvestigated in HGAN, as the generator is still governed by Gaussian samples in Euclidean space.", + "bbox": [ + 75, + 458, + 468, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent Code Manipulation. It has been shown that the latent spaces of GANs are able to encode rich semantic information [20, 32, 50]. One of the popular approaches is finding linear directions corresponding to changes in a given binary labeled attributes, which might be difficult to obtain for new datasets and could require manual labeling effort [50, 20, 12]. Others [8, 58, 41, 31, 9] try to find semantic directions in an unsupervised manner. For instance, PCA is applied in the latent space to create interpretable controls for synthesizing images [31, 9]. Most recent works [53, 51] directly compute in the close form to find the meaningful semantic direction without training and optimization. In com", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "parison, our work HAE focuses on attributes in different semantic levels in the latent space rather than trying hard to find disentangled interpretable directions as previous works.", + "bbox": [ + 498, + 393, + 893, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 454, + 591, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of HAE is shown in Fig. 2, we first give a detailed explanation of getting the hierarchical representations in the hyperbolic space, and then we introduce the framework of HAE and explain the loss functions.", + "bbox": [ + 498, + 479, + 893, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Hierarchical Representation", + "text_level": 1, + "bbox": [ + 500, + 550, + 756, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The major issue of our study is how to obtain the hierarchical representation from real images to facilitate editing in different semantic levels, as illustrated in Fig. 1. Therefore, hyperbolic space is introduced as the latent space to achieve this goal.", + "bbox": [ + 498, + 574, + 890, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike Euclidean spaces with their zero curvature and spherical spaces with their positive curvature, hyperbolic spaces with negative curvature have been shown that it is more appropriate for learning hierarchical representation [44, 45]. Informally, hyperbolic space can be viewed as a continuous analogy of trees [44]. One important feature of hyperbolic space is that the length grows exponentially with its radius while linearly in Euclidean space. This property allows hyperbolic space to be naturally compatible with hierarchical data [22] including text, images, videos, etc.", + "bbox": [ + 496, + 650, + 892, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The $n$ -dimensional hyperbolic space can be formally defined as a homogeneous, simply connected $n$ -dimensional Riemannian manifold, denoted as $\\mathbb{H}^n$ with constant negative sectional curvature1. We choose to work in the Poincaré disk from five isometric models of hyperbolic", + "bbox": [ + 496, + 801, + 893, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1The curvature of the hyperbolic space $c$ is set as $-1$ in this work.", + "bbox": [ + 517, + 886, + 867, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22716", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5f0c8e82bef45e739ea6849503339b029492966b9a2749af717d4385f97fde25.jpg", + "image_caption": [ + "Figure 3: Illustration of the property of hyperbolic space on the Poincaré disk. Given two latent codes of Spaniel $z_{\\mathbb{D}}^{c_1}$ and $z_{\\mathbb{D}}^{c_1'}$ (red dots) on the edge of Poincaré disk, the geodesic between these two points is the brown curve rather than a straight line in Euclidean space. Therefore, their average latent code is calculated as $\\bar{z}_{\\mathbb{D}}^{c_1}$ (pink dot) which is closer to the center $O$ (still a Spaniel, but less fine-grained). While the latent code of a tiger $z_{\\mathbb{D}}^{c_2}$ (blue dot) locates far from the latent code of a Spaniel. Thus, the hyperbolic average code of tiger and Spaniel $\\bar{z}_{\\mathbb{D}}^{c_{12}}$ (purple dot) is closer to the center $O$ than $\\bar{z}_{\\mathbb{D}}^{c_1}$ which is more abstract (a feline contains features from both tiger and Spaniel)." + ], + "image_footnote": [], + "bbox": [ + 99, + 92, + 444, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "space defined in [6] since it is commonly used in gradient-based learning [44, 18, 45, 56, 55, 35]. The Poincaré disk model $\\left(\\mathbb{D}^n,g^{\\mathbb{D}}\\right)$ is defined by the manifold $\\mathbb{D}^n = \\{x\\in \\mathbb{R}^n:\\| x\\| < 1\\}$ equipped with the following Riemannian metric:", + "bbox": [ + 76, + 530, + 468, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {x} ^ {\\mathbb {D}} = \\lambda_ {x} ^ {2} g ^ {E}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 604, + 468, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{x} = \\frac{2}{1 - \\|x\\|^{2}}$ and $g^{E}$ is the Euclidean metric tensor $g^{E} = \\mathbf{I}^{n}$ . The induced distance between two points $\\mathbf{x},\\mathbf{y}\\in \\mathbb{D}^n$ can be defined by:", + "bbox": [ + 76, + 628, + 468, + 678 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathbb {D}} (\\mathbf {x}, \\mathbf {y}) = \\operatorname {a r c c o s h} \\left(1 + 2 \\frac {\\| \\mathbf {x} - \\mathbf {y} \\| ^ {2}}{\\left(1 - \\| \\mathbf {x} \\| ^ {2}\\right) \\left(1 - \\| \\mathbf {y} \\| ^ {2}\\right)}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 686, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recall that a geodesic is a locally minimized-length curve between two points. In the hyperboloid model, the geodesic can be defined as the curve created by intersecting the plane defined by two points and the origin with the hyperboloid [38]. Thus, the mean of two latent codes in hyperbolic space locates at the mid-point of the geodesic that is closer to the origin. This is the key desired feature of hyperbolic space, i.e., the mean between two leaf embeddings is not another leaf embedding, but the hierarchical parent of them [55]. This feature allows us to generate new images by moving the latent code from one leaf to another with the", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "same parents. We can also change the semantic levels of attributes by determining how abstract their parent is.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This unique property is visualized in Fig. 3 on a 2-D Poincaré disk. The image embedding near the edge of the ball (with a large radius) represents a more fine-grained image while the embedding near the center (which has a smaller radius) represents an image with abstract features (an average face).", + "bbox": [ + 498, + 122, + 893, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the hyperbolic space shares similar features with trees, it is continuous. In other words, there is no fixed number of hierarchy levels. Instead, there is a continuum from very fine-grained (near the edge of Poincaré disk) to very abstract (near the origin).", + "bbox": [ + 498, + 214, + 893, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Network Architecture", + "text_level": 1, + "bbox": [ + 500, + 301, + 705, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although we aim to embed and edit real images in hyperbolic space, the whole network does not need to be implemented in a hyperbolic manner. Instead, we can take advantage of the number of existing GAN inversion models and optimization algorithms that have been fine-tuned for Euclidean space.", + "bbox": [ + 498, + 325, + 890, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve image editing, we need to embed the image back into the latent space. In particular, we select pSp [49] as the backbone of HAE to encode images to the $\\mathcal{W}^+$ -space of StyleGAN2 [34]:", + "bbox": [ + 498, + 417, + 890, + 477 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {i} = \\mathrm {p S p} \\left(x _ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 492, + 890, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{w}_i\\in \\mathbb{R}^{18\\times 512}$ is the corresponding latent vector of $x_{i}$ in the $\\mathcal{W}^+$ -space.", + "bbox": [ + 498, + 520, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To manipulate latent code in hyperbolic space, we need to define a bijective map from $\\mathbb{R}^n$ to $\\mathbb{D}_c^n$ to map Euclidean vectors to the hyperbolic space and vice versa. A manifold is a differentiable topological space that locally resembles the Euclidean space $\\mathbb{R}^n$ [37, 38]. For $x\\in \\mathbb{D}^n$ , one can define the tangent space $T_{x}\\mathbb{D}_{c}^{n}$ of $\\mathbb{D}_c^n$ at $x$ as the first order linear approximation of $\\mathbb{D}_c^n$ around $x$ . Therefore, this bijective map can be performed by exponential and logarithmic maps. Specifically, the exponential map $\\exp_{\\mathbf{x}}^{c}:T_{\\mathbf{x}}\\mathbb{D}_{c}^{n}\\cong \\mathbb{R}^{n}\\to \\mathbb{D}_{c}^{n}$ , maps from the tangent spaces into the manifold. While the logarithmic map $\\log_{\\mathbf{x}}^{c}:\\mathbb{D}_{c}^{n}\\to T_{\\mathbf{x}}\\mathbb{D}_{c}^{n}\\cong \\mathbb{R}^{n}$ is the reverse map of the exponential map.", + "bbox": [ + 498, + 553, + 892, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use exponential and logarithmic maps at origin $\\mathbf{0}$ for the transformation between the Euclidean and hyperbolic representations. After getting $\\mathbf{w}_i$ in the $\\mathcal{W}^+$ -space, we first use a Multi-layer Perceptron (MLP) encoder to reduce the dimension of latent vectors in Euclidean space. Then we apply an exponential map to project the Euclidean latent code $z_{\\mathbb{R}i}$ to hyperbolic space. After that, we use the hyperbolic feed-forward layer as [18] to obtain the final hierarchical representation $z_{\\mathbb{D}}$ as shown in Fig. 2:", + "bbox": [ + 498, + 734, + 890, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\mathbb {D} i} = f ^ {\\otimes_ {c}} \\left(\\exp_ {\\mathbf {0}} ^ {c} \\left(\\mathrm {M L P} _ {E} (\\mathbf {w} _ {i})\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 883, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22717", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8dac1d1444f682bdbff58e0c7b8fae390dfb8bc5f46ce3b9eef0509faecf5432.jpg", + "image_caption": [ + "Figure 4: One-shot image generation from HAE on Animal Faces, Flowers, and VGGFaces." + ], + "image_footnote": [], + "bbox": [ + 163, + 90, + 472, + 229 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7255d8b40b12993f122298bd7ea3eeb3ebea7af22453f64765212fb3e83278c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 90, + 805, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f^{\\otimes_c}$ is the Möbius translation of feed-forward layer $f$ as the map from $\\mathbb{D}_c^n$ to $\\mathbb{D}_c^m$ , denoted as Möbius linear layer.", + "bbox": [ + 75, + 271, + 468, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the hyperbolic representation $z_{\\mathbb{D}}$ needs to be projected back to the $\\mathcal{W}^+$ -space of StyleGAN2. In practice, this is achieved by applying a logarithmic map followed by an MLP decoder:", + "bbox": [ + 75, + 301, + 468, + 361 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {i} ^ {\\prime} = \\operatorname {M L P} _ {D} \\left(\\log_ {\\mathbf {0}} ^ {c} \\left(z _ {\\mathbb {D} i}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 373, + 468, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $\\mathbf{w}_i^{\\prime}$ will be fed into a pre-trained StyleGAN2's generator $G$ to reconstruct the image $x_{i}^{\\prime}$ .", + "bbox": [ + 76, + 401, + 468, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Loss Function", + "text_level": 1, + "bbox": [ + 76, + 441, + 220, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss function of HAE consists of two parts: the Hyperbolic loss ensures to get the hierarchical representation in the hyperbolic space and the reconstruction loss guarantees the quality of reconstruction images.", + "bbox": [ + 75, + 465, + 468, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hyperbolic Loss. To learn the semantic hierarchical representation of real images in hyperbolic space, we minimize the distance between latent codes of images with similar categories and attributes while pushing away the latent codes from different categories. We choose the supervised approach to achieve this. In order to perform multi-class classification on the Poincaré disk defined in Sec. 3.1, one needs to generalize multinomial logistic regression (MLR) to the Poincaré disk defined in [18]. An extra linear layer needs to be trained for the classification and the softmax probability can be computed as: Given $K$ classes and $k \\in \\{1, \\ldots, K\\}$ , $p_k \\in \\mathbb{D}_c^n$ , $a_k \\in T_{p_k} \\mathbb{D}_c^n \\setminus \\{\\mathbf{0}\\}$ :", + "bbox": [ + 75, + 525, + 468, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p (y = k \\mid x) \\propto \\exp \\left(\\frac {\\lambda_ {p _ {k}} ^ {c} \\| a _ {k} \\|}{\\sqrt {c}} \\sinh^ {- 1} \\right. \\\\ \\left(\\frac {2 \\sqrt {c} \\langle - p _ {k} \\oplus_ {c} x , a _ {k} \\rangle}{\\left(1 - c \\| - p _ {k} \\oplus_ {c} x \\| ^ {2}\\right) \\| a _ {k} \\|\\right)\\left. \\right), \\\\ \\forall x \\in \\mathbb {D} _ {c} ^ {n}. \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 718, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\oplus_{c}$ denotes the Möbius addition defined in [35] with fixed sectional curvature of the space, denoted by $c$ .", + "bbox": [ + 75, + 840, + 467, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After getting the softmax result for each class, one can use negative log-likelihood loss (NLL Loss) to calculate the", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "hyperbolic loss:", + "bbox": [ + 500, + 272, + 607, + 286 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {h y p e r}} = - \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\log \\left(p _ {n}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 297, + 890, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the batch size and $p_n$ is the probability predicted by the model for the correct class.", + "bbox": [ + 498, + 349, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned in Sec. 3.1, the distance between points grows exponentially with their radius in the Poincaré disk. In order to minimize Eq. (7), the latent codes of fine-grained images will be pushed to the edge of the ball to maximize the distances between different categories while the embedding of abstract images (images have common features from many categories) will be located near the center of the ball. Since hyperbolic space is continuous and differentiable, we are able to optimize Eq. (7) with stochastic gradient descent, which learns the hierarchy of the images.", + "bbox": [ + 498, + 378, + 890, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reconstruction Loss. In order to guarantee the quality of the generated images, we first use the $\\mathcal{L}_2$ loss and LPIPS loss used in pSp [49], given image $x_{i}$ :", + "bbox": [ + 498, + 530, + 890, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2} \\left(x _ {i}\\right) = \\left\\| x _ {i} - \\mathrm {H A E} \\left(x _ {i}\\right) \\right\\| _ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 588, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L P I P S}} \\left(x _ {i}\\right) = \\| F \\left(x _ {i}\\right) - F (\\mathrm {H A E} \\left(x _ {i}\\right)) \\| _ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 631, + 890, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F(\\cdot)$ denotes the perceptual feature extractor.", + "bbox": [ + 498, + 654, + 841, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the pSp encoder and StyleGAN2 generator are pretrained, we only train the neural layers between the encoder and generator of HAE. To further guarantee the network to better project back to the $\\mathcal{W}^+$ -space, the reconstructed $\\mathbf{w}_i'$ should be the same as the original $\\mathbf{w}_i$ :", + "bbox": [ + 496, + 669, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c}} \\left(w _ {i}\\right) = \\left\\| \\mathbf {w} _ {\\mathrm {i}} - \\mathbf {w} _ {i} ^ {\\prime} \\right\\| _ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 756, + 890, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{w}_i^{\\prime}$ can be calculated by Eq. (4) and Eq. (5).", + "bbox": [ + 500, + 785, + 833, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss function is:", + "bbox": [ + 519, + 800, + 712, + 813 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {2} \\left(x _ {i}\\right) + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {L P I P S}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {3} \\mathcal {L} _ {\\text {h y p e r}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 827, + 890, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1},\\lambda_{2}$ and $\\lambda_{3}$ are trade-off adaptive parameters. This curated set of loss functions ensures the model learns the hierarchical representation and reconstructs images.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22718", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5fa26aab41441e0d1d0c821143c058787a2d03e1ad5b6b95a160abf7c17914ae.jpg", + "image_caption": [ + "Figure 5: Images generated by HAE by adding the same perturbation on the latent code of a given image with different hyperbolic radii on Animal Faces and Flowers." + ], + "image_footnote": [], + "bbox": [ + 161, + 88, + 810, + 265 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0c6302289e3089c5e2b0504fa904a12bc7da1c482eef0dd11c4b89f9ff215842.jpg", + "image_caption": [ + "Figure 6: Interpolations in hyperbolic space along the edge of the Poincaré disk (with $r_{\\mathbb{D}} = 6.2126$ ) on three datasets." + ], + "image_footnote": [], + "bbox": [ + 78, + 318, + 468, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Image Generation", + "text_level": 1, + "bbox": [ + 76, + 491, + 253, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To study the generating quality of the model, a straightforward way is to generate new images via interpolation between two designated images, or random perturbation.", + "bbox": [ + 76, + 513, + 468, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In hyperbolic space, the shortest path with the induced distance between two points is given by the geodesic defined in Eq. (2). The geodesic equation between two embeddings $z_{\\mathbb{D}\\mathbf{i}}$ and $z_{\\mathbb{D}\\mathbf{j}}$ , denoted by $\\gamma z_{\\mathbb{D}i}\\to z_{\\mathbb{D}j}(t)$ , is given by", + "bbox": [ + 76, + 560, + 468, + 622 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {z _ {\\mathbb {D} i} \\rightarrow z _ {\\mathbb {D} j}} (t) = z _ {\\mathbb {D} i} \\oplus_ {c} t \\otimes_ {c} \\left(\\left(- z _ {\\mathbb {D} i}\\right) \\oplus_ {c} z _ {\\mathbb {D} j}\\right), t \\in [ 0, 1 ], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 643, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\oplus_{c}$ denotes the Möbius addition with aforementioned sectional curvature $c$ , with details in supplementary material.", + "bbox": [ + 76, + 674, + 468, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt the following method to achieve generating via perturbation: For a given image $x_{i}$ , we first rescale its embedding $z_{\\mathbb{D}i}$ to the desired radius $r_{\\mathbb{D}}$ . Then we sample a random vector from seen categories in $z_{\\mathbb{D}j}$ with radius $r_{\\mathbb{D}}$ fixed and take the geodesic as the direction of perturbation to generate images.", + "bbox": [ + 76, + 719, + 468, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 821, + 200, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 847, + 294, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the training stage, we first train a StyleGAN2 [33] and pSp [49] with seen categories. Given a trained pSp, the", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ec7a49ed547804cdbf2da6ee8e40cbb514291eee3b8ed893497fff31ea49295.jpg", + "image_caption": [ + "Figure 7: Interpolations by moving the latent codes from the edge to the center of the Poincaré disk (from fine-grained to abstract) on three datasets." + ], + "image_footnote": [], + "bbox": [ + 504, + 311, + 888, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MLP encoder $\\mathrm{MLP}_E$ is an 8-layer MLP with a Leaky-ReLU activation function. The dimension of the latent code in hyperbolic space is chosen to be 512. More details can be found in the supplementary.", + "bbox": [ + 498, + 506, + 890, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Datasets", + "text_level": 1, + "bbox": [ + 500, + 577, + 601, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on Animal Faces [40], Flowers [46], and VGGFaces [48] following the settings described in [15].", + "bbox": [ + 498, + 599, + 890, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Animal Faces. We randomly select 119 categories as seen for training and leave 30 as unseen categories for testing.", + "bbox": [ + 498, + 646, + 890, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Flowers. The Flowers [46] dataset is split into 85 seen categories for training and 17 unseen categories for testing.", + "bbox": [ + 498, + 676, + 890, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VGGFaces. For VGGFaces [48], we randomly select 1802 categories for training and 572 for testing.", + "bbox": [ + 498, + 707, + 890, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Analysis of Hierarchical Feature Editing", + "text_level": 1, + "bbox": [ + 500, + 747, + 846, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We analyze the properties of the learned hierarchical representations and how the levels of attributes relate to their locations of latent codes in hyperbolic space.", + "bbox": [ + 498, + 771, + 890, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As we mentioned in Sec. 3.1, there is a continuum from fine-grained attributes to abstract attributes, corresponding to the points from the peripheral to the center of the ball. We define the hyperbolic radius $r_{\\mathbb{D}}^2$ as the hyperbolic dis", + "bbox": [ + 498, + 816, + 890, + 877 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2The radius of the Poincaré disk in our experiment is about 6.2126", + "bbox": [ + 514, + 886, + 872, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22719", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0c9858734b7907de3c669b7523ce7b7fafa9770080660a70faed799bec9d9052.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSettingsFlowersAnimal FacesVGG Faces*
FID(↓)LPIPS(↑)FID(↓)LPIPS(↑)FID(↓)LPIPS(↑)
DAWSON [39]3-shot188.960.0583208.680.0642137.820.0769
MatchingGAN [28]3-shot143.350.1627148.520.1514118.620.1695
F2GAN [30]3-shot120.480.2172117.740.1831109.160.2125
LoFGAN [23]3-shot79.330.3862112.810.496420.310.2869
DeltaGAN [27]1-shot109.780.391289.810.441880.120.3146
Disco-FUNIT [29]1-shot90.120.443671.440.4511--
AGE [15]1-shot45.960.430528.040.557534.860.3294
SAGE [14]1-shot43.520.439227.430.544834.970.3232
HAE (Ours)1-shot50.100.473926.330.563635.930.5919
", + "bbox": [ + 83, + 88, + 885, + 282 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/942c39ebdac414ba3d3e6cae4b84fd4bf91f8b62056ed8a3d19fa21782be31d4.jpg", + "table_caption": [ + "Table 1: FID(↓) and LPIPS(↑) of images generated by different methods for unseen categories on three datasets. Bold indicates the best results and underline indicates the second best results. VGGFaces is marked with * because different methods report different numbers of unseen categories on this dataset(e.g. 552 in LoFGAN, 96 in DeltaGAN, 497 in L2GAN, and 572 in AGE and SAGE). Note that: disco-FUNIT [29] does not provide pre-trained models on VGG Faces [48] dataset." + ], + "table_footnote": [], + "table_body": "
InputMatchingGANLoFGANAGEHAE(ours)
", + "bbox": [ + 84, + 347, + 888, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 8: Comparison between images generated by MatchingGAN, LoFGAN, AGE, and HAE on Flowers, Animal Faces, and VGGFaces. Zoom in to see the details. Note that: SAGE [14] has not released code and pre-trained models.", + "bbox": [ + 75, + 585, + 892, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tance of the given latent code to the center of the Poincaré disk. To study the influence of the radius of embeddings in hyperbolic space, we run several experiments with different choices of $r_{\\mathbb{D}}$ .", + "bbox": [ + 75, + 633, + 470, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperbolic Perturbation and Interpolation. As mentioned in Sec. 3.4, we demonstrate the results of perturbation and interpolation. In addition to the choice of perturbation, we can set the intensity of the perturbation by controlling both the step distance and radius as shown in Fig. 5. The results show that level of semantic attributes is highly related to $r_{\\mathbb{D}}$ . With the radius becoming smaller, the attributes become more abstract. We further visualize this property of hyperbolic space by moving the latent codes of the given image from the edge of the Poincaré disk to the center. As Fig. 7 shows, the images change from very fine-grained to very abstract (the average of all images). The results in Fig. 6 show that we can achieve smooth interpola", + "bbox": [ + 75, + 704, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion in hyperbolic space without any distortion. The results demonstrate that with HAE, we can freely control the editing geodesically and hierarchically.", + "bbox": [ + 496, + 633, + 890, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Few-shot Image Generation", + "text_level": 1, + "bbox": [ + 498, + 694, + 750, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As Fig. 5 shows, the image categories will be changed when $r_{\\mathbb{D}}$ is smaller than about 4, and the category-irrelevant attributes of images will be changed when $r_{\\mathbb{D}}$ is larger than about 5. The embeddings of Animal Faces are visualized in 2-D Poincaré disk using UMAP [43] shown in Fig. 9. As Fig. 6 shows, the posture and the angle of the images will be changed at the early stage of interpolation without changing the category. Thus, the images can be generated by moving the latent code of a given image to some randomly selected semantic direction within the cluster of the category. In practice, we select $r_{\\mathbb{D}} = 6.21$ and step size of perturbation as 8 to achieve few-shot image generation as", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22720", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c88386d26e9c5591abf100e3ba414141752462191b577ad1bc9ad47b6ea573ce.jpg", + "image_caption": [ + "Figure 9: UMAP visualization of hyperbolic 2-D embeddings of Animal Faces dataset. We observe that similar categories are clustered and positioned near the boundary, while ambiguous samples are located near the center. Zoom in to see the details." + ], + "image_footnote": [], + "bbox": [ + 116, + 97, + 433, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 4 shows the diverse images generated by adding random perturbations from seen categories. We conduct three experiments to show that HAE can achieve promising few-shot image generation. More examples of generated images are available in the supplementary.", + "bbox": [ + 75, + 416, + 468, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Comparison with State-of-the-art. We calculate the FID [24] and LPIPS [59] to evaluate the fidelity and diversity of the generated images following one-shot settings in [15, 14]. The comparison results are shown in Tab. 1. Our method achieves the best scores on most of the FID and LPIPS metrics compared with state-of-the-art few-shot image generation methods, which indicates that our method not only improves the model from the semantic aspect but also achieves state-of-the-art performance on the traditional evaluation metrics. Specifically, the LPIPS score of HAE beats SOTA model SAGE on all three datasets since HAE can generate more diverse images.", + "bbox": [ + 75, + 493, + 468, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Evaluation. We qualitatively compare our method with MatchingGAN [28], LoFGAN [23], DeltaGAN [27] and AGE [15]. As shown in Fig. 8, HAE can generate images with diversity and fine-grained details. More importantly, the newly generated images have more semantic diversity than others. For instance, the shadow and skin color of the generated faces change with the light condition, and this effect looks more natural. We further conduct a user study by randomly selecting 60 (20 from each dataset) images with generated variants using AGE and HAE. 50 users from different backgrounds are asked to rate the results only based on diversity and quality external information. This is achieved by randomly shuffling the order of images pairwisely and inside any pair. HAE won by a ratio of $58.1\\%$ (1743/3000) over AGE (more details in supplementary).", + "bbox": [ + 75, + 674, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dff337300422ef151eb250686ca8c16928f26d0703f014aab3eecbb8e295ba58.jpg", + "image_caption": [ + "Figure 10: Manipulate images from different categories with the same perturbation (Target 1&2)." + ], + "image_footnote": [], + "bbox": [ + 532, + 88, + 870, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transferability. If we move latent codes at category-irrelevant levels, the target perturbation is transferable across all categories. We edit the images from three categories with the same editing direction, the output images are shown in Fig. 10. It demonstrates that HAE achieves a highly controllable and interpretable editing process.", + "bbox": [ + 498, + 295, + 890, + 386 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5f9f560479b0d9136517568e4b67b3c443063013777ce7ca054ae3a3676eeca1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFlowersAnimal FacesVGG Faces
FIDLPIPSFIDLPIPSFIDLPIPS
SAGE [14]43.520.439227.430.544834.970.3232
HAE(Euc)54.620.429325.270.512938.460.5908
HAE(Hyp)50.100.473926.330.563635.930.5919
", + "bbox": [ + 504, + 398, + 888, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2: $\\mathrm{{FID}}\\left( \\downarrow \\right)$ and LPIPS(↑) of images generated by HAE in different geometries for unseen categories on three datasets. Bold indicates the best results and underline indicates the second best results.", + "bbox": [ + 498, + 486, + 890, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 560, + 653, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "HAE in Euclidean. We re-trained HAE models in Euclidean space with the NLL loss to validate the performance gain in Tab. 1 is due to the hierarchical hyperbolic representation rather than the disentanglement caused by Eq. (7). The quantitative comparison is shown in Tab. 2. It shows that the hyperbolic space boosts the performance, especially for the LPIPS score, since the latent code is more disentangled in hyperbolic space [19]. This finding is also supported by the UMAP visualization in Fig. 9. More details can be found in the supplementary material.", + "bbox": [ + 496, + 583, + 890, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyperbolic Radius versus Truncation. StyleGAN [33] uses truncation trick [42, 4, 33, 34] in $\\mathcal{W}$ -space to achieve the balance between the image quality and diversity. The experiments in [33, 34] also show that the truncation level in $\\mathcal{W}^+$ -space control the level of abstraction of the generated images. We conduct the experiments in Sec. 3.4 using truncation to validate the gains of hyperbolic space. The results are illustrated in Fig. 11 and Fig. 12. As Fig. 11 shows, the category of the image changes along with the posture of the dog as the truncation gets smaller, while the category-relevant attributes do not change when the hyperbolic radius", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22721", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3aed1d95f93aeb57be8519dddde61107148227012a529009129067feb1f24650.jpg", + "image_caption": [ + "Figure 11: Top: Interpolations by moving the latent codes from the edge to the center in hyperbolic space. Bottom: Interpolation with different truncation in Euclidean space. Zoom in to see the details." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 467, + 188 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/9b86bda6924f822f791aa9d247f422d569275c81d1d371940aaf6f6fc7a33c96.jpg", + "image_caption": [ + "Figure 12: Images generated by HAE(Hyp) and HAE(Euc) by adding the same perturbation on the latent code of a given image with different settings of hyperbolic radius and truncation. Zoom in to see the details." + ], + "image_footnote": [], + "bbox": [ + 78, + 253, + 467, + 383 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "$(r_{\\mathbb{D}})$ is large. This can also be proved in Fig. 12. The category remains the same after adding perturbation when $r_{\\mathbb{D}}$ is large, while the truncation can not control semantic-level editing. This shows that Euclidean space can only capture scale-based hierarchy rather than the semantic hierarchy.", + "bbox": [ + 75, + 460, + 467, + 535 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Downstream Task. We conduct data augmentation via HAE for image classification on Animal Faces [48]. We randomly select 30, 35, and 35 images for each category as train, val, and test, respectively. Following [23], a ResNet-18 backbone is initialized from the seen categories. Then the model is fine-tuned on the unseen categories referred to as the baseline. 60 images are generated for each unseen category as data augmentation. The result is presented in Tab. 3. The diversity and quality of generated images are primarily controlled by the hyperbolic radii $r_{\\mathbb{D}}$ . As the radius becomes smaller, HAE generates images of higher diversity, but categories (referring to high-level attributes) also gradually change to others. $r_{\\mathbb{D}} = 6$ achieves the best performance on the classification experiment. However, the performance drops when the radius is smaller than 4.5. This is because the semantic attributes change too much and thus mislead the classifier.", + "bbox": [ + 75, + 536, + 467, + 791 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.6. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 76, + 801, + 339, + 816 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Although HAE achieves reliable hierarchical attribute editing in hyperbolic space for few-shot image generation, there are several limitations.", + "bbox": [ + 75, + 824, + 467, + 868 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "First, the boundary of category changing is hard to be quantified since the hierarchical levels are continuous in", + "bbox": [ + 75, + 869, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1ba5faaa2f5eb0bfd2f9d8c6da29214feb58a9db9ba01ac1828d8084cf15e10f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Hyperbolic RadiusAccuracyFID(↓)LPIPS(↑)
baseline58.67--
6.060.1046.890.4520
5.559.5248.680.4651
5.059.0552.080.4823
4.559.1460.870.5174
4.058.5765.830.5386
3.556.8668.440.6034
3.054.1469.400.6316
", + "bbox": [ + 501, + 88, + 890, + 262 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 3: Ablation of same perturbation on different radii on Animal Faces.", + "bbox": [ + 500, + 267, + 888, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "the hyperbolic space. Users need to find a \"safe\" boundary by trying different radii and step sizes of perturbation before generating new images. However, from another perspective, this continuity of hierarchy provides flexibility for users to set different boundaries for different downstream tasks as they need.", + "bbox": [ + 496, + 323, + 890, + 412 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Second, the performance of HAE is limited by the pretrained styleGAN and the inversion method. If the input image can not be well embedded, the editing will also fail. This problem can be solved by changing more powerful backbones, e.g., ViT [16], in future work.", + "bbox": [ + 496, + 417, + 890, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, we use supervised learning to get the hierarchical embedding in hyperbolic space. However, the number of images in existing datasets with labels for generation tasks is relatively small, which makes the embeddings in the hyperbolic space not evenly distributed. The solution to this problem is simple, use unsupervised learning with large-scale high-quality datasets.", + "bbox": [ + 496, + 497, + 890, + 603 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 628, + 617, + 643 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we propose a simple yet effective method HAE to edit hierarchical attributes in hyperbolic space. After learning the semantic hierarchy from images, our model is able to edit continuous semantic hierarchical features of images for flexible few-shot image generation in the hyperbolic space. Experiments demonstrate that HAE is capable of achieving not only stable few-shot image generation with state-of-the-art quality and diversity but a controllable and interpretable editing process. Future work includes the combination of HAE and large pretrained models and applications to more downstream tasks.", + "bbox": [ + 496, + 659, + 890, + 824 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by the National Key R&D Program of China under Grant 2018AAA0102000, in part by National Natural Science Foundation of China: 62022083 and 62236008.", + "bbox": [ + 496, + 840, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22722", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Antreas Antoniou, Amos Storkey, and Harrison Edwards. Data augmentation generative adversarial networks. arXiv preprint arXiv:1711.04340, 2017. 1, 2", + "[2] Sergey Bartunov and Dmitry P. Vetrov. Few-shot generative modelling with generative matching networks. In AISTATS, 2018. 1", + "[3] Silvère Bonnabel. Stochastic gradient descent on riemannian manifolds. IEEE Transactions on Automatic Control, 58(9):2217-2229, 2013. 3", + "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. In ICLR, 2019. 8", + "[5] Gary Bécigneul and Octavian-Eugen Ganea. Riemannian adaptive optimization methods. In ICLR, 2019. 3", + "[6] James W Cannon, William J Floyd, Richard Kenyon, and Walter R Parry. Hyperbolic geometry. Flavors of geometry, 31:59-115, 1997. 4", + "[7] Ines Chami, Rex Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. In NeurIPS, page 4868-4879, 2019. 3", + "[8] Anton Cherepkov, Andrey Voynov, and Artem Babenko. Navigating the gan parameter space for semantic image editing. In CVPR, pages 3670-3679, 2021. 3", + "[9] Jaewoong Choi, Junho Lee, Changyeon Yoon, Jung Ho Park, Geonho Hwang, and Myungjoo Kang. Do not escape from the manifold: Discovering the local coordinates on the latent space of gans. In ICLR, 2022. 3", + "[10] Louis Clouatre and Marc Demers. Figr: Few-shot image generation with reptile. arXiv:1901.02199, 2019. 1, 2", + "[11] Jiali Cui, Ying Nian Wu, and Tian Han. Learning joint latent space ebm prior model for multi-layer generator. In CVPR, pages 3603-3612, 2023. 2", + "[12] Emily Denton, Ben Hutchinson, Margaret Mitchell, and Timnit Gebru. Detecting bias with generative counterfactual face attribute augmentation. CoRR, abs/1906.06439, 2019. 3", + "[13] Bhuwan Dhingra, Chris Shallue, Mohammad Norouzi, Andrew Dai, and George Dahl. Embedding text in hyperbolic spaces. arXiv preprint arXiv:1806.04313, 2018. 2", + "[14] Guanqi Ding, Xinzhe Han, Shuhui Wang, Xin Jin, Dandan Tu, and Qingming Huang. Stable attribute group editing for reliable few-shot image generation. arXiv preprint arXiv:2302.00179, 2023. 2, 7, 8", + "[15] Guanqi Ding, Xinzhe Han, Shuhui Wang, Shuzhe Wu, Xin Jin, Dandan Tu, and Qingming Huang. Attribute group editing for reliable few-shot image generation. In CVPR, pages 11184-11193, 2022. 1, 2, 6, 7, 8", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 9", + "[17] Kun Fu, Tengfei Zhang, Yue Zhang, Menglong Yan, Zhonghan Chang, Zhengyuan Zhang, and Xian Sun. Meta-ssd: Towards fast adaptation for few-shot object detection with meta-learning. IEEE Access, 7:77597-77606, 2019. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. In NeurIPS, pages 5345-5355, 2018. 3, 4, 5", + "[19] Songwei Ge, Shlok Mishra, Simon Kornblith, Chun-Liang Li, and David Jacobs. Hyperbolic contrastive learning for visual representations beyond objects. In CVPR, 2023. 8", + "[20] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Phillip Isola. Ganalyze: Toward visual definitions of cognitive image properties. In CVPR, page 5744-5753, 2019. 3", + "[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2", + "[22] Michael Gromov. Hyperbolic groups. In Essays in group theory, 1987. 3", + "[23] Zheng Gu, Wenbin Li, Jing Huo, Lei Wang, and Yang Gao. Lofgan: Fusing local representations for fewshot image generation. In ICCV, 2021. 1, 2, 7, 8, 9", + "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 8", + "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In ICLR, 2020. 2", + "[26] Yan Hong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Deltagan: Towards diverse few-shot image generation with sample-specific delta. In CVPR, 2020. 1", + "[27] Yan Hong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Deltagan: Towards diverse few-shot image generation with sample-specific delta. In ECCV, 2022. 1, 2, 7, 8", + "[28] Yan Hong, Li Niu, Jianfu Zhang, and Liqing Zhang. Matchinggan: Matching-based few-shot image generation. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2020. 1, 2, 7, 8", + "[29] Yan Hong, Li Niu, Jianfu Zhang, and Liqing Zhang. Few-shot image generation using discrete content representation. In Proceedings of the 30th ACM International Conference on Multimedia, MM '22, page 2796-2804, New York, NY, USA, 2022. Association for Computing Machinery. 1, 2, 7", + "[30] Yan Hong, Li Niu, Jianfu Zhang, Weijie Zhao, Chen Fu, and Liqing Zhang. F2gan: Fusing-and-filling gan for few-shot image generation. In Proceedings of the 28th ACM International Conference on Multimedia, MM '20, page 2535-2543. Association for Computing Machinery, 2020. 1, 2, 7", + "[31] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In NeurIPS, 2020. 3", + "[32] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the\" steerability\" of generative adversarial networks. In ICLR, 2020. 3", + "[33] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4217-4228, 2019. 6, 8", + "[34] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, pages 8107-8116, 2020. 1, 4, 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22723", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In CVPR, pages 6417-6427, 2020. 2, 3, 4, 5", + "[36] Diego Lazcano, Nicolas Fredes Franco, and Werner Creixell. Hgan: Hyperbolic generative adversarial network. IEEE Access, 9:96309-96320, 2021. 3", + "[37] John M Lee. Riemannian manifolds: an introduction to curvature. Springer Science & Business Media, 176, 2006. 4", + "[38] John M Lee. Introduction to Smooth Manifolds. Springer, 2013. 4", + "[39] Weixin Liang, Zixuan Liu, and Can Liu. Dawson: A domain adaptive few shot generation framework. arXiv preprint arXiv:2001.00576, 2020. 1, 2, 7", + "[40] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsueprivised image-to-image translation. In ICCV, 2019. 6", + "[41] Yu-Ding Lu, Hsin-Ying Lee, Hung-Yu Tseng, and Ming-Hsuan Yang. Unsupervised discovery of disentangled manifolds in gans. arXiv preprint arXiv:2011.11842, 2020. 3", + "[42] Marco Marchesi. Megapixel size image creation using generative adversarial networks. CoRR, abs/1706.00082, 2017. 8", + "[43] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Grossberger. Umap: Uniform manifold approximation and projection. The Journal of Open Source Software, 3(29):861, 2018. 7", + "[44] Maximillian Nickel and Douwe Kiela. Generative visual manipulation on the natural image manifold. In ECCV, 2017. 2, 3, 4", + "[45] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In ICML, 2018. 3, 4", + "[46] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 6", + "[47] Jiwoong Park, Junho Cho, Hyung Jin Chang, and Jin Young Choi. Unsupervised hyperbolic representation learning via message passing auto-encoders. In CVPR, pages 5512–5522, 2021. 3", + "[48] Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In British Machine Vision Conference, 2015. 6, 7, 9", + "[49] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In CVPR, pages 2287–2296, 2021. 4, 5, 6", + "[50] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, pages 9240-9249, 2020. 3", + "[51] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, pages 1532-1540, 2021. 3", + "[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Nurit Spingarn-Eliezer, Ron Banner, and Tomer Michaeli. Gan \"steerability\" without optimization. In ICLR, 2021. 3", + "[54] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H.S. Torr, and Timothy M. Hospedales. Learning to compare: Relation network for few-shot learning. In CVPR, pages 1199-1208, 2018. 1", + "[55] Dídac Surís, Ruoshi Liu, and Carl Vondrick. Learning the predictability of the future. In CVPR, pages 12602-12612, 2021. 2, 3, 4", + "[56] Alexandru Tifrea, Gary Bécigneul, and OctavianEugen Ganea. Poincaré glove: Hyperbolic word embeddings. In ICLR, 2019. 2, 3, 4", + "[57] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NeurIPS, 2016. 1", + "[58] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In ICML, pages 9786-9796, 2020. 3", + "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 8", + "[60] Jingyuan Zhu, Huimin Ma, Jiansheng Chen, and Jian Yuan. Few-shot image generation with diffusion models. arXiv preprint arXiv:2211.03264, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 446 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22724", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/e705b187-21ba-4812-b0e4-1feb752f1cdd_model.json b/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/e705b187-21ba-4812-b0e4-1feb752f1cdd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..47f8e0a531a9555ad2d1cc960df66448a1fb05e1 --- /dev/null +++ b/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/e705b187-21ba-4812-b0e4-1feb752f1cdd_model.json @@ -0,0 +1,2510 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.193, + 0.131, + 0.78, + 0.177 + ], + "angle": 0, + "content": "The Euclidean Space is Evil: Hyperbolic Attribute Editing for Few-shot Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.203, + 0.666, + 0.223 + ], + "angle": 0, + "content": "Lingxiao Li1 Yi Zhang2 Shuhui Wang3*" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.238, + 0.677, + 0.257 + ], + "angle": 0, + "content": "1 Columbia University 2 University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.257, + 0.753, + 0.275 + ], + "angle": 0, + "content": "3 Institute of Computing Technology, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.278, + 0.751, + 0.291 + ], + "angle": 0, + "content": "113504@columbia.edu, wolf5965@ox.ac.uk, wangshuhui@ict.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.474, + 0.661 + ], + "angle": 0, + "content": "Few-shot image generation is a challenging task since it aims to generate diverse new images for an unseen category with only a few images. Existing methods suffer from the trade-off between the quality and diversity of generated images. To tackle this problem, we propose Hyperbolic Attribute Editing (HAE), a simple yet effective method. Unlike other methods that work in Euclidean space, HAE captures the hierarchy among images using data from seen categories in hyperbolic space. Given a well-trained HAE, images of unseen categories can be generated by moving the latent code of a given image toward any meaningful directions in the Poincaré disk with a fixing radius. Most importantly, the hyperbolic space allows us to control the semantic diversity of the generated images by setting different radii in the disk. Extensive experiments and visualizations demonstrate that HAE is capable of not only generating images with promising quality and diversity using limited data but achieving a highly controllable and interpretable editing process. Code is available at https://github.com/lingxiao-li/HAE." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.688, + 0.21, + 0.704 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Due to the persistent development of deep learning, the task of image generation has received significant research attention in recent years. Specifically, the Generative Adversarial Networks (GANs) [21] and its variants (e.g., StyleGANv2 [34]) have succeeded in generating high-fidelity and realistic images, requiring a large number of high-quality data for model training. However, considering the long-tail distribution and data imbalance widely exists among different image categories [30], it is difficult for GANs to be trained on categories with sufficient training images to generate new realistic and diverse images for a" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.327, + 0.895, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.47, + 0.895, + 0.568 + ], + "angle": 0, + "content": "Figure 1: Illustration of hierarchical attribute editing in hyperbolic space. Hyperbolic space can naturally and compactly encode semantic hierarchical structures within a large image data corpus. Changing the high-level, i.e., category-relevant attribute \\(\\Delta \\mathbf{w}^r\\) changes the category of an image. While changing low-level or category-irrelevant attribute \\(\\Delta \\mathbf{w}^{ir}\\) varies images within categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.895, + 0.671 + ], + "angle": 0, + "content": "category with only a few images. This task is referred to as few-shot image generation [10, 26, 30, 28, 29, 27, 15]. A variety of tasks can benefit from improvements in few-shot image generation, for instance, low-data detection [17] and few-shot classification [54, 57]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.896, + 0.903 + ], + "angle": 0, + "content": "In general, existing GAN-based few-shot image generation mechanisms can be classified into three categories. Transfer-based methods [10, 39] introduce meta-learning or domain adaptation on GANs to generate new images by enforcing knowledge transfer among categories. Fusion-based methods [2, 23, 30, 28] perform feature fusion of multiple input images in a feature space and generate images via decoding the fused features back to image space. However, the output is still highly similar to the source images. Transformation-based methods [1, 29, 27, 15] find intra-category transformations or inject random perturbations to conditional unseen category samples to generate images without tedious fine-tuning. By representing the images in the Euclidean feature space, the above learning mechanisms tend to be over-complicated, and the generated" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22714" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.414, + 0.107 + ], + "angle": 0, + "content": "images are often collapsed due to limited diversity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.11, + 0.473, + 0.504 + ], + "angle": 0, + "content": "Similar to the ubiquity of hierarchies in language [44, 56, 13], the semantic hierarchy is also common in images [35, 11]. As Fig. 1 shows, the semantic hierarchies constructed in the language domain can be instantiated with visual images. From the visual perspective, an image can be regarded as a collection of attributes of multiple levels. High-level attributes, a.k.a. category-relevant attributes, define the category of an image, such as the shape and color of an animal [15]. For instance, in the middle row of Fig. 1, changing the high-level attributes of the given image of a Shih-Tzu dog, the category can be changed to a Rhodesian Ridgeback Dog. While the low-level or fine-grained attributes, including expressions, postures, etc., that vary within the category as shown at the bottom of Fig. 1, are called category-irrelevant attributes. Therefore, an image can also be viewed as a descendant of another image with the same category-relevant attributes by adding fine-grained category-irrelevant attributes to its parent image. To edit the visual attributes for high-quality image generation, it is crucial to capture the attribute hierarchy within the large image data corpus and find a good representation space. Ideally, we aim to construct a hierarchical visual representation in a latent space that allows us to change the category of an image by moving the latent code in a category-relevant direction, and perform few-shot image generation by moving the code in a category-irrelevant direction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.473, + 0.763 + ], + "angle": 0, + "content": "Unfortunately, the Euclidean space and its corresponding distance metrics used by existing GAN-based methods can not facilitate the hierarchical attribute representation, thus the design of complicated attribute disentangling and editing mechanisms seems to be crucial for the generation quality. Inspired by the application of hyperbolic space in images [35] and videos [55], we found that the metrics introduced in hyperbolic geometry can naturally and compactly encode hierarchical structures. Unlike the general affine spaces, e.g., the Euclidean space, hyperbolic spaces can be viewed as the continuous analog of a tree since tree-like graphs can be embedded in finite-dimension with minimal distortion [44]. This property of hyperbolic space provides continuous and up to infinite semantic levels for attribute editing, allowing us to robustly generate diverse images with only a few images from unseen categories with simple operations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Based on the above findings, we propose a simple but effective Hyperbolic Attribute Editing (HAE) method for few-shot image generation. Our method is based on the observation that hierarchical latent code manipulation can be easily implemented in Hyperbolic space. The core of HAE is mapping the latent vectors from the Euclidean space \\(\\mathbb{R}^n\\) to a hyperbolic space \\(\\mathbb{D}^n\\). We minimize a supervised classification loss function to ensure the images are hierarchically embedded in hyperbolic space. Once we capture the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "attribute hierarchy among images, we can generate new images of unseen categories by moving the latent code from one leaf to another with the same parents by fixing the radius. Most importantly, the hyperbolic space allows us to control the semantic diversity of generated images by setting different radii in the Poincaré disk. Those operations can well facilitate continually hierarchical attribute editing in hyperbolic space for flexible few-shot image generation with both quality and diversity." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.228, + 0.846, + 0.243 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.254, + 0.892, + 0.345 + ], + "angle": 0, + "content": "- We propose a simple yet effective method for few-shot image generation, i.e., hyperbolic attribute editing. In order to capture the hierarchy among images, we use hyperbolic space as the latent space. To the best of our knowledge, HAE is the first attempt to use hyperbolic latent spaces for few-shot image generation." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.356, + 0.892, + 0.416 + ], + "angle": 0, + "content": "- We show that in our designed hyperbolic latent space, the semantic hierarchical attribute relations among images can be reflected by their distances to the center of the Poincaré disk." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.428, + 0.892, + 0.518 + ], + "angle": 0, + "content": "- Extensive experiments and visualization suggest that HAE achieves stable few-shot image generation with state-of-the-art quality and diversity. Unlike other few-shot image generation methods, HAE allows us to generate images with better control of diversity by changing the semantic levels of attributes we want to edit." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.254, + 0.892, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.543, + 0.642, + 0.559 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Few-shot image generation. Recently, diverse methods have been proposed for few-shot image generation. The transfer-based methods [10, 39] which introduce meta-learning or domain adaptation on GANs can hardly generate realistic images. While fusion-based methods that fuse the features by matching the random vector with the conditional images [28] or formulating the problem as a conditional generating task [23, 30] suffer from the limited diversity of generated images. Furthermore, transformation-based methods [1, 29, 27, 15] can generate images with only one conditional image by focusing on either capturing the cross-category or intra-category transformations by injecting random perturbations [1]. Nevertheless, the transformation captured by those methods is not very consistent. Ding et al. [15, 14] propose the \"editing-based\" perspective, the intra-category transformation can be modeled as category-irrelevant image editing based on one sample instead of pairs of samples. Most recently, Zhu et al. [60] fine-tune powerful diffusion models (DMs) [25] pre-trained on large source domains on limited target data to generate diverse and high quality images. DMs outperform GANs [21] on sample quality with a more controllable training process" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22715" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.098, + 0.885, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.895, + 0.378 + ], + "angle": 0, + "content": "Figure 2: The overview of HAE. The Hyper layer is a hyperbolic feedforward layer called Möbius linear layer which is used to project the latent code from Euclidean space \\(\\mathbb{R}^n\\) to hyperbolic space \\(\\mathbb{D}^n\\) [7]. \\(\\bar{z}_{\\mathbb{D}i}\\) can be viewed as the \"parent\" or average code of \\(z_{\\mathbb{D}i}\\) and \\(z_{\\mathbb{D}i}'\\). One can generate diverse images without changing the category by moving the latent code from one child to another of the same parent in the hyperbolic space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.471, + 0.456 + ], + "angle": 0, + "content": "at the cost less flexibility and editability, since they denoise images in the image space rather than operate in the latent space. Furthermore, the inference process of DMs is much slower than GANs [52]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.717 + ], + "angle": 0, + "content": "Hyperbolic Embedding. The use of hyperbolic space in deep learning [44, 45, 56, 55, 35] is a pioneering work in recent years. It was first used in natural language processing for hierarchical language representation [44, 45, 56]. The Riemannian optimization algorithms are used to optimize models in hyperbolic space [5, 3]. As hyperbolic space is successfully applied to represent hierarchical data, Ganea et al. [18] derives hyperbolic versions of tools in neural networks including multinomial logistic regression, feed-forward, and recurrent neural networks. Following this, hyperbolic geometry is used in image [35], video [55], and graph data [7, 47]. Most recently, Lazcano et al [36] shows that hyperbolic space outperforms traditional Euclidean space in image generation using HGAN. However, the hierarchy and controllability of hyperbolic space remain uninvestigated in HGAN, as the generator is still governed by Gaussian samples in Euclidean space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Latent Code Manipulation. It has been shown that the latent spaces of GANs are able to encode rich semantic information [20, 32, 50]. One of the popular approaches is finding linear directions corresponding to changes in a given binary labeled attributes, which might be difficult to obtain for new datasets and could require manual labeling effort [50, 20, 12]. Others [8, 58, 41, 31, 9] try to find semantic directions in an unsupervised manner. For instance, PCA is applied in the latent space to create interpretable controls for synthesizing images [31, 9]. Most recent works [53, 51] directly compute in the close form to find the meaningful semantic direction without training and optimization. In com" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.395, + 0.895, + 0.442 + ], + "angle": 0, + "content": "parison, our work HAE focuses on attributes in different semantic levels in the latent space rather than trying hard to find disentangled interpretable directions as previous works." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.593, + 0.47 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.48, + 0.894, + 0.542 + ], + "angle": 0, + "content": "The overall framework of HAE is shown in Fig. 2, we first give a detailed explanation of getting the hierarchical representations in the hyperbolic space, and then we introduce the framework of HAE and explain the loss functions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.551, + 0.757, + 0.568 + ], + "angle": 0, + "content": "3.1. Hierarchical Representation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.575, + 0.892, + 0.65 + ], + "angle": 0, + "content": "The major issue of our study is how to obtain the hierarchical representation from real images to facilitate editing in different semantic levels, as illustrated in Fig. 1. Therefore, hyperbolic space is introduced as the latent space to achieve this goal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.893, + 0.801 + ], + "angle": 0, + "content": "Unlike Euclidean spaces with their zero curvature and spherical spaces with their positive curvature, hyperbolic spaces with negative curvature have been shown that it is more appropriate for learning hierarchical representation [44, 45]. Informally, hyperbolic space can be viewed as a continuous analogy of trees [44]. One important feature of hyperbolic space is that the length grows exponentially with its radius while linearly in Euclidean space. This property allows hyperbolic space to be naturally compatible with hierarchical data [22] including text, images, videos, etc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.802, + 0.894, + 0.879 + ], + "angle": 0, + "content": "The \\(n\\)-dimensional hyperbolic space can be formally defined as a homogeneous, simply connected \\(n\\)-dimensional Riemannian manifold, denoted as \\(\\mathbb{H}^n\\) with constant negative sectional curvature1. We choose to work in the Poincaré disk from five isometric models of hyperbolic" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.868, + 0.902 + ], + "angle": 0, + "content": "1The curvature of the hyperbolic space \\(c\\) is set as \\(-1\\) in this work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22716" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.093, + 0.445, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.359, + 0.473, + 0.513 + ], + "angle": 0, + "content": "Figure 3: Illustration of the property of hyperbolic space on the Poincaré disk. Given two latent codes of Spaniel \\( z_{\\mathbb{D}}^{c_1} \\) and \\( z_{\\mathbb{D}}^{c_1'} \\) (red dots) on the edge of Poincaré disk, the geodesic between these two points is the brown curve rather than a straight line in Euclidean space. Therefore, their average latent code is calculated as \\( \\bar{z}_{\\mathbb{D}}^{c_1} \\) (pink dot) which is closer to the center \\( O \\) (still a Spaniel, but less fine-grained). While the latent code of a tiger \\( z_{\\mathbb{D}}^{c_2} \\) (blue dot) locates far from the latent code of a Spaniel. Thus, the hyperbolic average code of tiger and Spaniel \\( \\bar{z}_{\\mathbb{D}}^{c_{12}} \\) (purple dot) is closer to the center \\( O \\) than \\( \\bar{z}_{\\mathbb{D}}^{c_1} \\) which is more abstract (a feline contains features from both tiger and Spaniel)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.531, + 0.47, + 0.605 + ], + "angle": 0, + "content": "space defined in [6] since it is commonly used in gradient-based learning [44, 18, 45, 56, 55, 35]. The Poincaré disk model \\(\\left(\\mathbb{D}^n,g^{\\mathbb{D}}\\right)\\) is defined by the manifold \\(\\mathbb{D}^n = \\{x\\in \\mathbb{R}^n:\\| x\\| < 1\\}\\) equipped with the following Riemannian metric:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.605, + 0.469, + 0.623 + ], + "angle": 0, + "content": "\\[\ng _ {x} ^ {\\mathbb {D}} = \\lambda_ {x} ^ {2} g ^ {E}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.629, + 0.47, + 0.679 + ], + "angle": 0, + "content": "where \\(\\lambda_{x} = \\frac{2}{1 - \\|x\\|^{2}}\\) and \\(g^{E}\\) is the Euclidean metric tensor \\(g^{E} = \\mathbf{I}^{n}\\). The induced distance between two points \\(\\mathbf{x},\\mathbf{y}\\in \\mathbb{D}^n\\) can be defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.688, + 0.469, + 0.734 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathbb {D}} (\\mathbf {x}, \\mathbf {y}) = \\operatorname {a r c c o s h} \\left(1 + 2 \\frac {\\| \\mathbf {x} - \\mathbf {y} \\| ^ {2}}{\\left(1 - \\| \\mathbf {x} \\| ^ {2}\\right) \\left(1 - \\| \\mathbf {y} \\| ^ {2}\\right)}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Recall that a geodesic is a locally minimized-length curve between two points. In the hyperboloid model, the geodesic can be defined as the curve created by intersecting the plane defined by two points and the origin with the hyperboloid [38]. Thus, the mean of two latent codes in hyperbolic space locates at the mid-point of the geodesic that is closer to the origin. This is the key desired feature of hyperbolic space, i.e., the mean between two leaf embeddings is not another leaf embedding, but the hierarchical parent of them [55]. This feature allows us to generate new images by moving the latent code from one leaf to another with the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "same parents. We can also change the semantic levels of attributes by determining how abstract their parent is." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.123, + 0.894, + 0.213 + ], + "angle": 0, + "content": "This unique property is visualized in Fig. 3 on a 2-D Poincaré disk. The image embedding near the edge of the ball (with a large radius) represents a more fine-grained image while the embedding near the center (which has a smaller radius) represents an image with abstract features (an average face)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.215, + 0.894, + 0.29 + ], + "angle": 0, + "content": "Although the hyperbolic space shares similar features with trees, it is continuous. In other words, there is no fixed number of hierarchy levels. Instead, there is a continuum from very fine-grained (near the edge of Poincaré disk) to very abstract (near the origin)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.302, + 0.706, + 0.316 + ], + "angle": 0, + "content": "3.2. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.326, + 0.892, + 0.417 + ], + "angle": 0, + "content": "Although we aim to embed and edit real images in hyperbolic space, the whole network does not need to be implemented in a hyperbolic manner. Instead, we can take advantage of the number of existing GAN inversion models and optimization algorithms that have been fine-tuned for Euclidean space." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.418, + 0.892, + 0.478 + ], + "angle": 0, + "content": "To achieve image editing, we need to embed the image back into the latent space. In particular, we select pSp [49] as the backbone of HAE to encode images to the \\(\\mathcal{W}^+\\)-space of StyleGAN2 [34]:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.493, + 0.891, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {i} = \\mathrm {p S p} \\left(x _ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.521, + 0.892, + 0.553 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}_i\\in \\mathbb{R}^{18\\times 512}\\) is the corresponding latent vector of \\(x_{i}\\) in the \\(\\mathcal{W}^+\\) -space." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.554, + 0.893, + 0.735 + ], + "angle": 0, + "content": "To manipulate latent code in hyperbolic space, we need to define a bijective map from \\(\\mathbb{R}^n\\) to \\(\\mathbb{D}_c^n\\) to map Euclidean vectors to the hyperbolic space and vice versa. A manifold is a differentiable topological space that locally resembles the Euclidean space \\(\\mathbb{R}^n\\) [37, 38]. For \\(x\\in \\mathbb{D}^n\\), one can define the tangent space \\(T_{x}\\mathbb{D}_{c}^{n}\\) of \\(\\mathbb{D}_c^n\\) at \\(x\\) as the first order linear approximation of \\(\\mathbb{D}_c^n\\) around \\(x\\). Therefore, this bijective map can be performed by exponential and logarithmic maps. Specifically, the exponential map \\(\\exp_{\\mathbf{x}}^{c}:T_{\\mathbf{x}}\\mathbb{D}_{c}^{n}\\cong \\mathbb{R}^{n}\\to \\mathbb{D}_{c}^{n}\\), maps from the tangent spaces into the manifold. While the logarithmic map \\(\\log_{\\mathbf{x}}^{c}:\\mathbb{D}_{c}^{n}\\to T_{\\mathbf{x}}\\mathbb{D}_{c}^{n}\\cong \\mathbb{R}^{n}\\) is the reverse map of the exponential map." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.892, + 0.872 + ], + "angle": 0, + "content": "We use exponential and logarithmic maps at origin \\(\\mathbf{0}\\) for the transformation between the Euclidean and hyperbolic representations. After getting \\(\\mathbf{w}_i\\) in the \\(\\mathcal{W}^+\\)-space, we first use a Multi-layer Perceptron (MLP) encoder to reduce the dimension of latent vectors in Euclidean space. Then we apply an exponential map to project the Euclidean latent code \\(z_{\\mathbb{R}i}\\) to hyperbolic space. After that, we use the hyperbolic feed-forward layer as [18] to obtain the final hierarchical representation \\(z_{\\mathbb{D}}\\) as shown in Fig. 2:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.885, + 0.891, + 0.902 + ], + "angle": 0, + "content": "\\[\nz _ {\\mathbb {D} i} = f ^ {\\otimes_ {c}} \\left(\\exp_ {\\mathbf {0}} ^ {c} \\left(\\mathrm {M L P} _ {E} (\\mathbf {w} _ {i})\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22717" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.165, + 0.092, + 0.473, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.092, + 0.807, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.238, + 0.757, + 0.253 + ], + "angle": 0, + "content": "Figure 4: One-shot image generation from HAE on Animal Faces, Flowers, and VGGFaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.47, + 0.302 + ], + "angle": 0, + "content": "where \\( f^{\\otimes_c} \\) is the Möbius translation of feed-forward layer \\( f \\) as the map from \\( \\mathbb{D}_c^n \\) to \\( \\mathbb{D}_c^m \\), denoted as Möbius linear layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.469, + 0.362 + ], + "angle": 0, + "content": "Finally, the hyperbolic representation \\( z_{\\mathbb{D}} \\) needs to be projected back to the \\( \\mathcal{W}^+ \\)-space of StyleGAN2. In practice, this is achieved by applying a logarithmic map followed by an MLP decoder:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.374, + 0.469, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {i} ^ {\\prime} = \\operatorname {M L P} _ {D} \\left(\\log_ {\\mathbf {0}} ^ {c} \\left(z _ {\\mathbb {D} i}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.402, + 0.47, + 0.434 + ], + "angle": 0, + "content": "and \\(\\mathbf{w}_i^{\\prime}\\) will be fed into a pre-trained StyleGAN2's generator \\(G\\) to reconstruct the image \\(x_{i}^{\\prime}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.443, + 0.222, + 0.457 + ], + "angle": 0, + "content": "3.3. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.469, + 0.526 + ], + "angle": 0, + "content": "The loss function of HAE consists of two parts: the Hyperbolic loss ensures to get the hierarchical representation in the hyperbolic space and the reconstruction loss guarantees the quality of reconstruction images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.469, + 0.709 + ], + "angle": 0, + "content": "Hyperbolic Loss. To learn the semantic hierarchical representation of real images in hyperbolic space, we minimize the distance between latent codes of images with similar categories and attributes while pushing away the latent codes from different categories. We choose the supervised approach to achieve this. In order to perform multi-class classification on the Poincaré disk defined in Sec. 3.1, one needs to generalize multinomial logistic regression (MLR) to the Poincaré disk defined in [18]. An extra linear layer needs to be trained for the classification and the softmax probability can be computed as: Given \\(K\\) classes and \\(k \\in \\{1, \\ldots, K\\}\\), \\(p_k \\in \\mathbb{D}_c^n\\), \\(a_k \\in T_{p_k} \\mathbb{D}_c^n \\setminus \\{\\mathbf{0}\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.719, + 0.469, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p (y = k \\mid x) \\propto \\exp \\left(\\frac {\\lambda_ {p _ {k}} ^ {c} \\| a _ {k} \\|}{\\sqrt {c}} \\sinh^ {- 1} \\right. \\\\ \\left(\\frac {2 \\sqrt {c} \\langle - p _ {k} \\oplus_ {c} x , a _ {k} \\rangle}{\\left(1 - c \\| - p _ {k} \\oplus_ {c} x \\| ^ {2}\\right) \\| a _ {k} \\|\\right)\\left. \\right), \\\\ \\forall x \\in \\mathbb {D} _ {c} ^ {n}. \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.468, + 0.871 + ], + "angle": 0, + "content": "where \\(\\oplus_{c}\\) denotes the Möbius addition defined in [35] with fixed sectional curvature of the space, denoted by \\(c\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "After getting the softmax result for each class, one can use negative log-likelihood loss (NLL Loss) to calculate the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.273, + 0.609, + 0.287 + ], + "angle": 0, + "content": "hyperbolic loss:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.298, + 0.892, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {h y p e r}} = - \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\log \\left(p _ {n}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.35, + 0.892, + 0.379 + ], + "angle": 0, + "content": "where \\(N\\) is the batch size and \\(p_n\\) is the probability predicted by the model for the correct class." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.38, + 0.892, + 0.531 + ], + "angle": 0, + "content": "As mentioned in Sec. 3.1, the distance between points grows exponentially with their radius in the Poincaré disk. In order to minimize Eq. (7), the latent codes of fine-grained images will be pushed to the edge of the ball to maximize the distances between different categories while the embedding of abstract images (images have common features from many categories) will be located near the center of the ball. Since hyperbolic space is continuous and differentiable, we are able to optimize Eq. (7) with stochastic gradient descent, which learns the hierarchy of the images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.531, + 0.892, + 0.577 + ], + "angle": 0, + "content": "Reconstruction Loss. In order to guarantee the quality of the generated images, we first use the \\(\\mathcal{L}_2\\) loss and LPIPS loss used in pSp [49], given image \\(x_{i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.589, + 0.891, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2} \\left(x _ {i}\\right) = \\left\\| x _ {i} - \\mathrm {H A E} \\left(x _ {i}\\right) \\right\\| _ {2}. \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.632, + 0.891, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L P I P S}} \\left(x _ {i}\\right) = \\| F \\left(x _ {i}\\right) - F (\\mathrm {H A E} \\left(x _ {i}\\right)) \\| _ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.655, + 0.842, + 0.669 + ], + "angle": 0, + "content": "where \\(F(\\cdot)\\) denotes the perceptual feature extractor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.745 + ], + "angle": 0, + "content": "Since the pSp encoder and StyleGAN2 generator are pretrained, we only train the neural layers between the encoder and generator of HAE. To further guarantee the network to better project back to the \\(\\mathcal{W}^+\\)-space, the reconstructed \\(\\mathbf{w}_i'\\) should be the same as the original \\(\\mathbf{w}_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.757, + 0.891, + 0.774 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c}} \\left(w _ {i}\\right) = \\left\\| \\mathbf {w} _ {\\mathrm {i}} - \\mathbf {w} _ {i} ^ {\\prime} \\right\\| _ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.786, + 0.834, + 0.801 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}_i^{\\prime}\\) can be calculated by Eq. (4) and Eq. (5)." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.801, + 0.713, + 0.814 + ], + "angle": 0, + "content": "The overall loss function is:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.828, + 0.891, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {2} \\left(x _ {i}\\right) + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {L P I P S}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {3} \\mathcal {L} _ {\\text {h y p e r}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\lambda_{1},\\lambda_{2}\\) and \\(\\lambda_{3}\\) are trade-off adaptive parameters. This curated set of loss functions ensures the model learns the hierarchical representation and reconstructs images." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22718" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.162, + 0.089, + 0.812, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.272, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Figure 5: Images generated by HAE by adding the same perturbation on the latent code of a given image with different hyperbolic radii on Animal Faces and Flowers." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.319, + 0.47, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.44, + 0.47, + 0.469 + ], + "angle": 0, + "content": "Figure 6: Interpolations in hyperbolic space along the edge of the Poincaré disk (with \\( r_{\\mathbb{D}} = 6.2126 \\)) on three datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.492, + 0.254, + 0.508 + ], + "angle": 0, + "content": "3.4. Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.515, + 0.47, + 0.56 + ], + "angle": 0, + "content": "To study the generating quality of the model, a straightforward way is to generate new images via interpolation between two designated images, or random perturbation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.561, + 0.47, + 0.623 + ], + "angle": 0, + "content": "In hyperbolic space, the shortest path with the induced distance between two points is given by the geodesic defined in Eq. (2). The geodesic equation between two embeddings \\(z_{\\mathbb{D}\\mathbf{i}}\\) and \\(z_{\\mathbb{D}\\mathbf{j}}\\), denoted by \\(\\gamma z_{\\mathbb{D}i}\\to z_{\\mathbb{D}j}(t)\\), is given by" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.644, + 0.469, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {z _ {\\mathbb {D} i} \\rightarrow z _ {\\mathbb {D} j}} (t) = z _ {\\mathbb {D} i} \\oplus_ {c} t \\otimes_ {c} \\left(\\left(- z _ {\\mathbb {D} i}\\right) \\oplus_ {c} z _ {\\mathbb {D} j}\\right), t \\in [ 0, 1 ], \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.47, + 0.718 + ], + "angle": 0, + "content": "where \\(\\oplus_{c}\\) denotes the Möbius addition with aforementioned sectional curvature \\(c\\), with details in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.469, + 0.811 + ], + "angle": 0, + "content": "We adopt the following method to achieve generating via perturbation: For a given image \\( x_{i} \\), we first rescale its embedding \\( z_{\\mathbb{D}i} \\) to the desired radius \\( r_{\\mathbb{D}} \\). Then we sample a random vector from seen categories in \\( z_{\\mathbb{D}j} \\) with radius \\( r_{\\mathbb{D}} \\) fixed and take the geodesic as the direction of perturbation to generate images." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.822, + 0.202, + 0.839 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.295, + 0.864 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In the training stage, we first train a StyleGAN2 [33] and pSp [49] with seen categories. Given a trained pSp, the" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.312, + 0.89, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.445, + 0.893, + 0.486 + ], + "angle": 0, + "content": "Figure 7: Interpolations by moving the latent codes from the edge to the center of the Poincaré disk (from fine-grained to abstract) on three datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.507, + 0.892, + 0.568 + ], + "angle": 0, + "content": "MLP encoder \\(\\mathrm{MLP}_E\\) is an 8-layer MLP with a Leaky-ReLU activation function. The dimension of the latent code in hyperbolic space is chosen to be 512. More details can be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.578, + 0.602, + 0.592 + ], + "angle": 0, + "content": "4.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.601, + 0.892, + 0.646 + ], + "angle": 0, + "content": "We evaluate our method on Animal Faces [40], Flowers [46], and VGGFaces [48] following the settings described in [15]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.647, + 0.892, + 0.677 + ], + "angle": 0, + "content": "Animal Faces. We randomly select 119 categories as seen for training and leave 30 as unseen categories for testing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.678, + 0.892, + 0.707 + ], + "angle": 0, + "content": "Flowers. The Flowers [46] dataset is split into 85 seen categories for training and 17 unseen categories for testing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.708, + 0.892, + 0.739 + ], + "angle": 0, + "content": "VGGFaces. For VGGFaces [48], we randomly select 1802 categories for training and 572 for testing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.748, + 0.848, + 0.765 + ], + "angle": 0, + "content": "4.3. Analysis of Hierarchical Feature Editing" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.772, + 0.892, + 0.817 + ], + "angle": 0, + "content": "We analyze the properties of the learned hierarchical representations and how the levels of attributes relate to their locations of latent codes in hyperbolic space." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.818, + 0.892, + 0.878 + ], + "angle": 0, + "content": "As we mentioned in Sec. 3.1, there is a continuum from fine-grained attributes to abstract attributes, corresponding to the points from the peripheral to the center of the ball. We define the hyperbolic radius \\(r_{\\mathbb{D}}^2\\) as the hyperbolic dis" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.874, + 0.9 + ], + "angle": 0, + "content": "2The radius of the Poincaré disk in our experiment is about 6.2126" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22719" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.887, + 0.284 + ], + "angle": 0, + "content": "
MethodSettingsFlowersAnimal FacesVGG Faces*
FID(↓)LPIPS(↑)FID(↓)LPIPS(↑)FID(↓)LPIPS(↑)
DAWSON [39]3-shot188.960.0583208.680.0642137.820.0769
MatchingGAN [28]3-shot143.350.1627148.520.1514118.620.1695
F2GAN [30]3-shot120.480.2172117.740.1831109.160.2125
LoFGAN [23]3-shot79.330.3862112.810.496420.310.2869
DeltaGAN [27]1-shot109.780.391289.810.441880.120.3146
Disco-FUNIT [29]1-shot90.120.443671.440.4511--
AGE [15]1-shot45.960.430528.040.557534.860.3294
SAGE [14]1-shot43.520.439227.430.544834.970.3232
HAE (Ours)1-shot50.100.473926.330.563635.930.5919
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.289, + 0.895, + 0.346 + ], + "angle": 0, + "content": "Table 1: FID(↓) and LPIPS(↑) of images generated by different methods for unseen categories on three datasets. Bold indicates the best results and underline indicates the second best results. VGGFaces is marked with * because different methods report different numbers of unseen categories on this dataset(e.g. 552 in LoFGAN, 96 in DeltaGAN, 497 in L2GAN, and 572 in AGE and SAGE). Note that: disco-FUNIT [29] does not provide pre-trained models on VGG Faces [48] dataset." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.348, + 0.889, + 0.58 + ], + "angle": 0, + "content": "
InputMatchingGANLoFGANAGEHAE(ours)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.587, + 0.893, + 0.617 + ], + "angle": 0, + "content": "Figure 8: Comparison between images generated by MatchingGAN, LoFGAN, AGE, and HAE on Flowers, Animal Faces, and VGGFaces. Zoom in to see the details. Note that: SAGE [14] has not released code and pre-trained models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.635, + 0.471, + 0.695 + ], + "angle": 0, + "content": "tance of the given latent code to the center of the Poincaré disk. To study the influence of the radius of embeddings in hyperbolic space, we run several experiments with different choices of \\( r_{\\mathbb{D}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Hyperbolic Perturbation and Interpolation. As mentioned in Sec. 3.4, we demonstrate the results of perturbation and interpolation. In addition to the choice of perturbation, we can set the intensity of the perturbation by controlling both the step distance and radius as shown in Fig. 5. The results show that level of semantic attributes is highly related to \\( r_{\\mathbb{D}} \\). With the radius becoming smaller, the attributes become more abstract. We further visualize this property of hyperbolic space by moving the latent codes of the given image from the edge of the Poincaré disk to the center. As Fig. 7 shows, the images change from very fine-grained to very abstract (the average of all images). The results in Fig. 6 show that we can achieve smooth interpola" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.892, + 0.681 + ], + "angle": 0, + "content": "tion in hyperbolic space without any distortion. The results demonstrate that with HAE, we can freely control the editing geodesically and hierarchically." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.695, + 0.75, + 0.712 + ], + "angle": 0, + "content": "4.4. Few-shot Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As Fig. 5 shows, the image categories will be changed when \\( r_{\\mathbb{D}} \\) is smaller than about 4, and the category-irrelevant attributes of images will be changed when \\( r_{\\mathbb{D}} \\) is larger than about 5. The embeddings of Animal Faces are visualized in 2-D Poincaré disk using UMAP [43] shown in Fig. 9. As Fig. 6 shows, the posture and the angle of the images will be changed at the early stage of interpolation without changing the category. Thus, the images can be generated by moving the latent code of a given image to some randomly selected semantic direction within the cluster of the category. In practice, we select \\( r_{\\mathbb{D}} = 6.21 \\) and step size of perturbation as 8 to achieve few-shot image generation as" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22720" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.098, + 0.434, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.353, + 0.473, + 0.409 + ], + "angle": 0, + "content": "Figure 9: UMAP visualization of hyperbolic 2-D embeddings of Animal Faces dataset. We observe that similar categories are clustered and positioned near the boundary, while ambiguous samples are located near the center. Zoom in to see the details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.47, + 0.493 + ], + "angle": 0, + "content": "Fig. 4 shows the diverse images generated by adding random perturbations from seen categories. We conduct three experiments to show that HAE can achieve promising few-shot image generation. More examples of generated images are available in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.47, + 0.674 + ], + "angle": 0, + "content": "Quantitative Comparison with State-of-the-art. We calculate the FID [24] and LPIPS [59] to evaluate the fidelity and diversity of the generated images following one-shot settings in [15, 14]. The comparison results are shown in Tab. 1. Our method achieves the best scores on most of the FID and LPIPS metrics compared with state-of-the-art few-shot image generation methods, which indicates that our method not only improves the model from the semantic aspect but also achieves state-of-the-art performance on the traditional evaluation metrics. Specifically, the LPIPS score of HAE beats SOTA model SAGE on all three datasets since HAE can generate more diverse images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Qualitative Evaluation. We qualitatively compare our method with MatchingGAN [28], LoFGAN [23], DeltaGAN [27] and AGE [15]. As shown in Fig. 8, HAE can generate images with diversity and fine-grained details. More importantly, the newly generated images have more semantic diversity than others. For instance, the shadow and skin color of the generated faces change with the light condition, and this effect looks more natural. We further conduct a user study by randomly selecting 60 (20 from each dataset) images with generated variants using AGE and HAE. 50 users from different backgrounds are asked to rate the results only based on diversity and quality external information. This is achieved by randomly shuffling the order of images pairwisely and inside any pair. HAE won by a ratio of \\(58.1\\%\\) (1743/3000) over AGE (more details in supplementary)." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.089, + 0.872, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.249, + 0.892, + 0.277 + ], + "angle": 0, + "content": "Figure 10: Manipulate images from different categories with the same perturbation (Target 1&2)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.296, + 0.892, + 0.387 + ], + "angle": 0, + "content": "Transferability. If we move latent codes at category-irrelevant levels, the target perturbation is transferable across all categories. We edit the images from three categories with the same editing direction, the output images are shown in Fig. 10. It demonstrates that HAE achieves a highly controllable and interpretable editing process." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.399, + 0.889, + 0.482 + ], + "angle": 0, + "content": "
MethodFlowersAnimal FacesVGG Faces
FIDLPIPSFIDLPIPSFIDLPIPS
SAGE [14]43.520.439227.430.544834.970.3232
HAE(Euc)54.620.429325.270.512938.460.5908
HAE(Hyp)50.100.473926.330.563635.930.5919
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.487, + 0.892, + 0.542 + ], + "angle": 0, + "content": "Table 2: \\( \\mathrm{{FID}}\\left( \\downarrow \\right) \\) and LPIPS(↑) of images generated by HAE in different geometries for unseen categories on three datasets. Bold indicates the best results and underline indicates the second best results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.655, + 0.577 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.734 + ], + "angle": 0, + "content": "HAE in Euclidean. We re-trained HAE models in Euclidean space with the NLL loss to validate the performance gain in Tab. 1 is due to the hierarchical hyperbolic representation rather than the disentanglement caused by Eq. (7). The quantitative comparison is shown in Tab. 2. It shows that the hyperbolic space boosts the performance, especially for the LPIPS score, since the latent code is more disentangled in hyperbolic space [19]. This finding is also supported by the UMAP visualization in Fig. 9. More details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Hyperbolic Radius versus Truncation. StyleGAN [33] uses truncation trick [42, 4, 33, 34] in \\(\\mathcal{W}\\)-space to achieve the balance between the image quality and diversity. The experiments in [33, 34] also show that the truncation level in \\(\\mathcal{W}^+\\)-space control the level of abstraction of the generated images. We conduct the experiments in Sec. 3.4 using truncation to validate the gains of hyperbolic space. The results are illustrated in Fig. 11 and Fig. 12. As Fig. 11 shows, the category of the image changes along with the posture of the dog as the truncation gets smaller, while the category-relevant attributes do not change when the hyperbolic radius" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22721" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.192, + 0.47, + 0.247 + ], + "angle": 0, + "content": "Figure 11: Top: Interpolations by moving the latent codes from the edge to the center in hyperbolic space. Bottom: Interpolation with different truncation in Euclidean space. Zoom in to see the details." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.255, + 0.468, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.386, + 0.468, + 0.441 + ], + "angle": 0, + "content": "Figure 12: Images generated by HAE(Hyp) and HAE(Euc) by adding the same perturbation on the latent code of a given image with different settings of hyperbolic radius and truncation. Zoom in to see the details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.468, + 0.536 + ], + "angle": 0, + "content": "\\((r_{\\mathbb{D}})\\) is large. This can also be proved in Fig. 12. The category remains the same after adding perturbation when \\(r_{\\mathbb{D}}\\) is large, while the truncation can not control semantic-level editing. This shows that Euclidean space can only capture scale-based hierarchy rather than the semantic hierarchy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.468, + 0.792 + ], + "angle": 0, + "content": "Downstream Task. We conduct data augmentation via HAE for image classification on Animal Faces [48]. We randomly select 30, 35, and 35 images for each category as train, val, and test, respectively. Following [23], a ResNet-18 backbone is initialized from the seen categories. Then the model is fine-tuned on the unseen categories referred to as the baseline. 60 images are generated for each unseen category as data augmentation. The result is presented in Tab. 3. The diversity and quality of generated images are primarily controlled by the hyperbolic radii \\( r_{\\mathbb{D}} \\). As the radius becomes smaller, HAE generates images of higher diversity, but categories (referring to high-level attributes) also gradually change to others. \\( r_{\\mathbb{D}} = 6 \\) achieves the best performance on the classification experiment. However, the performance drops when the radius is smaller than 4.5. This is because the semantic attributes change too much and thus mislead the classifier." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.341, + 0.817 + ], + "angle": 0, + "content": "4.6. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.869 + ], + "angle": 0, + "content": "Although HAE achieves reliable hierarchical attribute editing in hyperbolic space for few-shot image generation, there are several limitations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.9 + ], + "angle": 0, + "content": "First, the boundary of category changing is hard to be quantified since the hierarchical levels are continuous in" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.263 + ], + "angle": 0, + "content": "
Hyperbolic RadiusAccuracyFID(↓)LPIPS(↑)
baseline58.67--
6.060.1046.890.4520
5.559.5248.680.4651
5.059.0552.080.4823
4.559.1460.870.5174
4.058.5765.830.5386
3.556.8668.440.6034
3.054.1469.400.6316
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.268, + 0.89, + 0.294 + ], + "angle": 0, + "content": "Table 3: Ablation of same perturbation on different radii on Animal Faces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.414 + ], + "angle": 0, + "content": "the hyperbolic space. Users need to find a \"safe\" boundary by trying different radii and step sizes of perturbation before generating new images. However, from another perspective, this continuity of hierarchy provides flexibility for users to set different boundaries for different downstream tasks as they need." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.892, + 0.495 + ], + "angle": 0, + "content": "Second, the performance of HAE is limited by the pretrained styleGAN and the inversion method. If the input image can not be well embedded, the editing will also fail. This problem can be solved by changing more powerful backbones, e.g., ViT [16], in future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.892, + 0.604 + ], + "angle": 0, + "content": "Finally, we use supervised learning to get the hierarchical embedding in hyperbolic space. However, the number of images in existing datasets with labels for generation tasks is relatively small, which makes the embeddings in the hyperbolic space not evenly distributed. The solution to this problem is simple, use unsupervised learning with large-scale high-quality datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.63, + 0.619, + 0.645 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.825 + ], + "angle": 0, + "content": "In this work, we propose a simple yet effective method HAE to edit hierarchical attributes in hyperbolic space. After learning the semantic hierarchy from images, our model is able to edit continuous semantic hierarchical features of images for flexible few-shot image generation in the hyperbolic space. Experiments demonstrate that HAE is capable of achieving not only stable few-shot image generation with state-of-the-art quality and diversity but a controllable and interpretable editing process. Future work includes the combination of HAE and large pretrained models and applications to more downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by the National Key R&D Program of China under Grant 2018AAA0102000, in part by National Natural Science Foundation of China: 62022083 and 62236008." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22722" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Antreas Antoniou, Amos Storkey, and Harrison Edwards. Data augmentation generative adversarial networks. arXiv preprint arXiv:1711.04340, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[2] Sergey Bartunov and Dmitry P. Vetrov. Few-shot generative modelling with generative matching networks. In AISTATS, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.471, + 0.24 + ], + "angle": 0, + "content": "[3] Silvère Bonnabel. Stochastic gradient descent on riemannian manifolds. IEEE Transactions on Automatic Control, 58(9):2217-2229, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. In ICLR, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.469, + 0.311 + ], + "angle": 0, + "content": "[5] Gary Bécigneul and Octavian-Eugen Ganea. Riemannian adaptive optimization methods. In ICLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[6] James W Cannon, William J Floyd, Richard Kenyon, and Walter R Parry. Hyperbolic geometry. Flavors of geometry, 31:59-115, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.469, + 0.396 + ], + "angle": 0, + "content": "[7] Ines Chami, Rex Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. In NeurIPS, page 4868-4879, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.398, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[8] Anton Cherepkov, Andrey Voynov, and Artem Babenko. Navigating the gan parameter space for semantic image editing. In CVPR, pages 3670-3679, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[9] Jaewoong Choi, Junho Lee, Changyeon Yoon, Jung Ho Park, Geonho Hwang, and Myungjoo Kang. Do not escape from the manifold: Discovering the local coordinates on the latent space of gans. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.495, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[10] Louis Clouatre and Marc Demers. Figr: Few-shot image generation with reptile. arXiv:1901.02199, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.524, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[11] Jiali Cui, Ying Nian Wu, and Tian Han. Learning joint latent space ebm prior model for multi-layer generator. In CVPR, pages 3603-3612, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.566, + 0.469, + 0.606 + ], + "angle": 0, + "content": "[12] Emily Denton, Ben Hutchinson, Margaret Mitchell, and Timnit Gebru. Detecting bias with generative counterfactual face attribute augmentation. CoRR, abs/1906.06439, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.648 + ], + "angle": 0, + "content": "[13] Bhuwan Dhingra, Chris Shallue, Mohammad Norouzi, Andrew Dai, and George Dahl. Embedding text in hyperbolic spaces. arXiv preprint arXiv:1806.04313, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[14] Guanqi Ding, Xinzhe Han, Shuhui Wang, Xin Jin, Dandan Tu, and Qingming Huang. Stable attribute group editing for reliable few-shot image generation. arXiv preprint arXiv:2302.00179, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[15] Guanqi Ding, Xinzhe Han, Shuhui Wang, Shuzhe Wu, Xin Jin, Dandan Tu, and Qingming Huang. Attribute group editing for reliable few-shot image generation. In CVPR, pages 11184-11193, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[17] Kun Fu, Tengfei Zhang, Yue Zhang, Menglong Yan, Zhonghan Chang, Zhengyuan Zhang, and Xian Sun. Meta-ssd: Towards fast adaptation for few-shot object detection with meta-learning. IEEE Access, 7:77597-77606, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[18] Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. In NeurIPS, pages 5345-5355, 2018. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[19] Songwei Ge, Shlok Mishra, Simon Kornblith, Chun-Liang Li, and David Jacobs. Hyperbolic contrastive learning for visual representations beyond objects. In CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[20] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Phillip Isola. Ganalyze: Toward visual definitions of cognitive image properties. In CVPR, page 5744-5753, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[22] Michael Gromov. Hyperbolic groups. In Essays in group theory, 1987. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[23] Zheng Gu, Wenbin Li, Jing Huo, Lei Wang, and Yang Gao. Lofgan: Fusing local representations for fewshot image generation. In ICCV, 2021. 1, 2, 7, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[26] Yan Hong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Deltagan: Towards diverse few-shot image generation with sample-specific delta. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.522 + ], + "angle": 0, + "content": "[27] Yan Hong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Deltagan: Towards diverse few-shot image generation with sample-specific delta. In ECCV, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[28] Yan Hong, Li Niu, Jianfu Zhang, and Liqing Zhang. Matchinggan: Matching-based few-shot image generation. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2020. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[29] Yan Hong, Li Niu, Jianfu Zhang, and Liqing Zhang. Few-shot image generation using discrete content representation. In Proceedings of the 30th ACM International Conference on Multimedia, MM '22, page 2796-2804, New York, NY, USA, 2022. Association for Computing Machinery. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[30] Yan Hong, Li Niu, Jianfu Zhang, Weijie Zhao, Chen Fu, and Liqing Zhang. F2gan: Fusing-and-filling gan for few-shot image generation. In Proceedings of the 28th ACM International Conference on Multimedia, MM '20, page 2535-2543. Association for Computing Machinery, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[31] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[32] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the\" steerability\" of generative adversarial networks. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[33] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4217-4228, 2019. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[34] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, pages 8107-8116, 2020. 1, 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22723" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[35] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In CVPR, pages 6417-6427, 2020. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.471, + 0.19 + ], + "angle": 0, + "content": "[36] Diego Lazcano, Nicolas Fredes Franco, and Werner Creixell. Hgan: Hyperbolic generative adversarial network. IEEE Access, 9:96309-96320, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.471, + 0.221 + ], + "angle": 0, + "content": "[37] John M Lee. Riemannian manifolds: an introduction to curvature. Springer Science & Business Media, 176, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.224, + 0.47, + 0.249 + ], + "angle": 0, + "content": "[38] John M Lee. Introduction to Smooth Manifolds. Springer, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.253, + 0.47, + 0.294 + ], + "angle": 0, + "content": "[39] Weixin Liang, Zixuan Liu, and Can Liu. Dawson: A domain adaptive few shot generation framework. arXiv preprint arXiv:2001.00576, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.297, + 0.469, + 0.338 + ], + "angle": 0, + "content": "[40] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsueprivised image-to-image translation. In ICCV, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.34, + 0.469, + 0.381 + ], + "angle": 0, + "content": "[41] Yu-Ding Lu, Hsin-Ying Lee, Hung-Yu Tseng, and Ming-Hsuan Yang. Unsupervised discovery of disentangled manifolds in gans. arXiv preprint arXiv:2011.11842, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.384, + 0.469, + 0.423 + ], + "angle": 0, + "content": "[42] Marco Marchesi. Megapixel size image creation using generative adversarial networks. CoRR, abs/1706.00082, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.469, + 0.48 + ], + "angle": 0, + "content": "[43] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Grossberger. Umap: Uniform manifold approximation and projection. The Journal of Open Source Software, 3(29):861, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[44] Maximillian Nickel and Douwe Kiela. Generative visual manipulation on the natural image manifold. In ECCV, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.527, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[45] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In ICML, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[46] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.469, + 0.682 + ], + "angle": 0, + "content": "[47] Jiwoong Park, Junho Cho, Hyung Jin Chang, and Jin Young Choi. Unsupervised hyperbolic representation learning via message passing auto-encoders. In CVPR, pages 5512–5522, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.726 + ], + "angle": 0, + "content": "[48] Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In British Machine Vision Conference, 2015. 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[49] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In CVPR, pages 2287–2296, 2021. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.469, + 0.827 + ], + "angle": 0, + "content": "[50] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, pages 9240-9249, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.469, + 0.869 + ], + "angle": 0, + "content": "[51] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, pages 1532-1540, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.891, + 0.12 + ], + "angle": 0, + "content": "[53] Nurit Spingarn-Eliezer, Ron Banner, and Tomer Michaeli. Gan \"steerability\" without optimization. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[54] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H.S. Torr, and Timothy M. Hospedales. Learning to compare: Relation network for few-shot learning. In CVPR, pages 1199-1208, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[55] Dídac Surís, Ruoshi Liu, and Carl Vondrick. Learning the predictability of the future. In CVPR, pages 12602-12612, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[56] Alexandru Tifrea, Gary Bécigneul, and OctavianEugen Ganea. Poincaré glove: Hyperbolic word embeddings. In ICLR, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[57] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NeurIPS, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[58] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In ICML, pages 9786-9796, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.35, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[60] Jingyuan Zhu, Huimin Ma, Jiansheng Chen, and Jian Yuan. Few-shot image generation with diffusion models. arXiv preprint arXiv:2211.03264, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22724" + } + ] +] \ No newline at end of file diff --git a/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/e705b187-21ba-4812-b0e4-1feb752f1cdd_origin.pdf b/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/e705b187-21ba-4812-b0e4-1feb752f1cdd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bbf8a8e1350d799e60339e3097211df500fe5902 --- /dev/null +++ b/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/e705b187-21ba-4812-b0e4-1feb752f1cdd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:623de25dbe1f2d160b1c95b98f915a98d8b02cc516b1687fb580fa20bfe81619 +size 7969604 diff --git a/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/full.md b/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7701337f8c7273a55ee84fb65aa48215e22e781d --- /dev/null +++ b/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/full.md @@ -0,0 +1,364 @@ +# The Euclidean Space is Evil: Hyperbolic Attribute Editing for Few-shot Image Generation + +Lingxiao Li1 Yi Zhang2 Shuhui Wang3* + +1 Columbia University 2 University of Oxford + +3 Institute of Computing Technology, Chinese Academy of Sciences + +113504@columbia.edu, wolf5965@ox.ac.uk, wangshuhui@ict.ac.cn + +# Abstract + +Few-shot image generation is a challenging task since it aims to generate diverse new images for an unseen category with only a few images. Existing methods suffer from the trade-off between the quality and diversity of generated images. To tackle this problem, we propose Hyperbolic Attribute Editing (HAE), a simple yet effective method. Unlike other methods that work in Euclidean space, HAE captures the hierarchy among images using data from seen categories in hyperbolic space. Given a well-trained HAE, images of unseen categories can be generated by moving the latent code of a given image toward any meaningful directions in the Poincaré disk with a fixing radius. Most importantly, the hyperbolic space allows us to control the semantic diversity of the generated images by setting different radii in the disk. Extensive experiments and visualizations demonstrate that HAE is capable of not only generating images with promising quality and diversity using limited data but achieving a highly controllable and interpretable editing process. Code is available at https://github.com/lingxiao-li/HAE. + +# 1. Introduction + +Due to the persistent development of deep learning, the task of image generation has received significant research attention in recent years. Specifically, the Generative Adversarial Networks (GANs) [21] and its variants (e.g., StyleGANv2 [34]) have succeeded in generating high-fidelity and realistic images, requiring a large number of high-quality data for model training. However, considering the long-tail distribution and data imbalance widely exists among different image categories [30], it is difficult for GANs to be trained on categories with sufficient training images to generate new realistic and diverse images for a + +![](images/798e260bd080ccee3be0c563ba81492de9d13b09e3997c3ed9d29f1b17138eb7.jpg) +Figure 1: Illustration of hierarchical attribute editing in hyperbolic space. Hyperbolic space can naturally and compactly encode semantic hierarchical structures within a large image data corpus. Changing the high-level, i.e., category-relevant attribute $\Delta \mathbf{w}^r$ changes the category of an image. While changing low-level or category-irrelevant attribute $\Delta \mathbf{w}^{ir}$ varies images within categories. + +category with only a few images. This task is referred to as few-shot image generation [10, 26, 30, 28, 29, 27, 15]. A variety of tasks can benefit from improvements in few-shot image generation, for instance, low-data detection [17] and few-shot classification [54, 57]. + +In general, existing GAN-based few-shot image generation mechanisms can be classified into three categories. Transfer-based methods [10, 39] introduce meta-learning or domain adaptation on GANs to generate new images by enforcing knowledge transfer among categories. Fusion-based methods [2, 23, 30, 28] perform feature fusion of multiple input images in a feature space and generate images via decoding the fused features back to image space. However, the output is still highly similar to the source images. Transformation-based methods [1, 29, 27, 15] find intra-category transformations or inject random perturbations to conditional unseen category samples to generate images without tedious fine-tuning. By representing the images in the Euclidean feature space, the above learning mechanisms tend to be over-complicated, and the generated + +images are often collapsed due to limited diversity. + +Similar to the ubiquity of hierarchies in language [44, 56, 13], the semantic hierarchy is also common in images [35, 11]. As Fig. 1 shows, the semantic hierarchies constructed in the language domain can be instantiated with visual images. From the visual perspective, an image can be regarded as a collection of attributes of multiple levels. High-level attributes, a.k.a. category-relevant attributes, define the category of an image, such as the shape and color of an animal [15]. For instance, in the middle row of Fig. 1, changing the high-level attributes of the given image of a Shih-Tzu dog, the category can be changed to a Rhodesian Ridgeback Dog. While the low-level or fine-grained attributes, including expressions, postures, etc., that vary within the category as shown at the bottom of Fig. 1, are called category-irrelevant attributes. Therefore, an image can also be viewed as a descendant of another image with the same category-relevant attributes by adding fine-grained category-irrelevant attributes to its parent image. To edit the visual attributes for high-quality image generation, it is crucial to capture the attribute hierarchy within the large image data corpus and find a good representation space. Ideally, we aim to construct a hierarchical visual representation in a latent space that allows us to change the category of an image by moving the latent code in a category-relevant direction, and perform few-shot image generation by moving the code in a category-irrelevant direction. + +Unfortunately, the Euclidean space and its corresponding distance metrics used by existing GAN-based methods can not facilitate the hierarchical attribute representation, thus the design of complicated attribute disentangling and editing mechanisms seems to be crucial for the generation quality. Inspired by the application of hyperbolic space in images [35] and videos [55], we found that the metrics introduced in hyperbolic geometry can naturally and compactly encode hierarchical structures. Unlike the general affine spaces, e.g., the Euclidean space, hyperbolic spaces can be viewed as the continuous analog of a tree since tree-like graphs can be embedded in finite-dimension with minimal distortion [44]. This property of hyperbolic space provides continuous and up to infinite semantic levels for attribute editing, allowing us to robustly generate diverse images with only a few images from unseen categories with simple operations. + +Based on the above findings, we propose a simple but effective Hyperbolic Attribute Editing (HAE) method for few-shot image generation. Our method is based on the observation that hierarchical latent code manipulation can be easily implemented in Hyperbolic space. The core of HAE is mapping the latent vectors from the Euclidean space $\mathbb{R}^n$ to a hyperbolic space $\mathbb{D}^n$ . We minimize a supervised classification loss function to ensure the images are hierarchically embedded in hyperbolic space. Once we capture the + +attribute hierarchy among images, we can generate new images of unseen categories by moving the latent code from one leaf to another with the same parents by fixing the radius. Most importantly, the hyperbolic space allows us to control the semantic diversity of generated images by setting different radii in the Poincaré disk. Those operations can well facilitate continually hierarchical attribute editing in hyperbolic space for flexible few-shot image generation with both quality and diversity. + +Our contributions can be summarized as follows: + +- We propose a simple yet effective method for few-shot image generation, i.e., hyperbolic attribute editing. In order to capture the hierarchy among images, we use hyperbolic space as the latent space. To the best of our knowledge, HAE is the first attempt to use hyperbolic latent spaces for few-shot image generation. +- We show that in our designed hyperbolic latent space, the semantic hierarchical attribute relations among images can be reflected by their distances to the center of the Poincaré disk. +- Extensive experiments and visualization suggest that HAE achieves stable few-shot image generation with state-of-the-art quality and diversity. Unlike other few-shot image generation methods, HAE allows us to generate images with better control of diversity by changing the semantic levels of attributes we want to edit. + +# 2. Related Work + +Few-shot image generation. Recently, diverse methods have been proposed for few-shot image generation. The transfer-based methods [10, 39] which introduce meta-learning or domain adaptation on GANs can hardly generate realistic images. While fusion-based methods that fuse the features by matching the random vector with the conditional images [28] or formulating the problem as a conditional generating task [23, 30] suffer from the limited diversity of generated images. Furthermore, transformation-based methods [1, 29, 27, 15] can generate images with only one conditional image by focusing on either capturing the cross-category or intra-category transformations by injecting random perturbations [1]. Nevertheless, the transformation captured by those methods is not very consistent. Ding et al. [15, 14] propose the "editing-based" perspective, the intra-category transformation can be modeled as category-irrelevant image editing based on one sample instead of pairs of samples. Most recently, Zhu et al. [60] fine-tune powerful diffusion models (DMs) [25] pre-trained on large source domains on limited target data to generate diverse and high quality images. DMs outperform GANs [21] on sample quality with a more controllable training process + +![](images/9372c400b973e3f1a5586f6beafe464dec6f37c4b21f87584f07c83e34f39aef.jpg) +Figure 2: The overview of HAE. The Hyper layer is a hyperbolic feedforward layer called Möbius linear layer which is used to project the latent code from Euclidean space $\mathbb{R}^n$ to hyperbolic space $\mathbb{D}^n$ [7]. $\bar{z}_{\mathbb{D}i}$ can be viewed as the "parent" or average code of $z_{\mathbb{D}i}$ and $z_{\mathbb{D}i}'$ . One can generate diverse images without changing the category by moving the latent code from one child to another of the same parent in the hyperbolic space. + +at the cost less flexibility and editability, since they denoise images in the image space rather than operate in the latent space. Furthermore, the inference process of DMs is much slower than GANs [52]. + +Hyperbolic Embedding. The use of hyperbolic space in deep learning [44, 45, 56, 55, 35] is a pioneering work in recent years. It was first used in natural language processing for hierarchical language representation [44, 45, 56]. The Riemannian optimization algorithms are used to optimize models in hyperbolic space [5, 3]. As hyperbolic space is successfully applied to represent hierarchical data, Ganea et al. [18] derives hyperbolic versions of tools in neural networks including multinomial logistic regression, feed-forward, and recurrent neural networks. Following this, hyperbolic geometry is used in image [35], video [55], and graph data [7, 47]. Most recently, Lazcano et al [36] shows that hyperbolic space outperforms traditional Euclidean space in image generation using HGAN. However, the hierarchy and controllability of hyperbolic space remain uninvestigated in HGAN, as the generator is still governed by Gaussian samples in Euclidean space. + +Latent Code Manipulation. It has been shown that the latent spaces of GANs are able to encode rich semantic information [20, 32, 50]. One of the popular approaches is finding linear directions corresponding to changes in a given binary labeled attributes, which might be difficult to obtain for new datasets and could require manual labeling effort [50, 20, 12]. Others [8, 58, 41, 31, 9] try to find semantic directions in an unsupervised manner. For instance, PCA is applied in the latent space to create interpretable controls for synthesizing images [31, 9]. Most recent works [53, 51] directly compute in the close form to find the meaningful semantic direction without training and optimization. In com + +parison, our work HAE focuses on attributes in different semantic levels in the latent space rather than trying hard to find disentangled interpretable directions as previous works. + +# 3. Method + +The overall framework of HAE is shown in Fig. 2, we first give a detailed explanation of getting the hierarchical representations in the hyperbolic space, and then we introduce the framework of HAE and explain the loss functions. + +# 3.1. Hierarchical Representation + +The major issue of our study is how to obtain the hierarchical representation from real images to facilitate editing in different semantic levels, as illustrated in Fig. 1. Therefore, hyperbolic space is introduced as the latent space to achieve this goal. + +Unlike Euclidean spaces with their zero curvature and spherical spaces with their positive curvature, hyperbolic spaces with negative curvature have been shown that it is more appropriate for learning hierarchical representation [44, 45]. Informally, hyperbolic space can be viewed as a continuous analogy of trees [44]. One important feature of hyperbolic space is that the length grows exponentially with its radius while linearly in Euclidean space. This property allows hyperbolic space to be naturally compatible with hierarchical data [22] including text, images, videos, etc. + +The $n$ -dimensional hyperbolic space can be formally defined as a homogeneous, simply connected $n$ -dimensional Riemannian manifold, denoted as $\mathbb{H}^n$ with constant negative sectional curvature1. We choose to work in the Poincaré disk from five isometric models of hyperbolic + +![](images/5f0c8e82bef45e739ea6849503339b029492966b9a2749af717d4385f97fde25.jpg) +Figure 3: Illustration of the property of hyperbolic space on the Poincaré disk. Given two latent codes of Spaniel $z_{\mathbb{D}}^{c_1}$ and $z_{\mathbb{D}}^{c_1'}$ (red dots) on the edge of Poincaré disk, the geodesic between these two points is the brown curve rather than a straight line in Euclidean space. Therefore, their average latent code is calculated as $\bar{z}_{\mathbb{D}}^{c_1}$ (pink dot) which is closer to the center $O$ (still a Spaniel, but less fine-grained). While the latent code of a tiger $z_{\mathbb{D}}^{c_2}$ (blue dot) locates far from the latent code of a Spaniel. Thus, the hyperbolic average code of tiger and Spaniel $\bar{z}_{\mathbb{D}}^{c_{12}}$ (purple dot) is closer to the center $O$ than $\bar{z}_{\mathbb{D}}^{c_1}$ which is more abstract (a feline contains features from both tiger and Spaniel). + +space defined in [6] since it is commonly used in gradient-based learning [44, 18, 45, 56, 55, 35]. The Poincaré disk model $\left(\mathbb{D}^n,g^{\mathbb{D}}\right)$ is defined by the manifold $\mathbb{D}^n = \{x\in \mathbb{R}^n:\| x\| < 1\}$ equipped with the following Riemannian metric: + +$$ +g _ {x} ^ {\mathbb {D}} = \lambda_ {x} ^ {2} g ^ {E}, \tag {1} +$$ + +where $\lambda_{x} = \frac{2}{1 - \|x\|^{2}}$ and $g^{E}$ is the Euclidean metric tensor $g^{E} = \mathbf{I}^{n}$ . The induced distance between two points $\mathbf{x},\mathbf{y}\in \mathbb{D}^n$ can be defined by: + +$$ +d _ {\mathbb {D}} (\mathbf {x}, \mathbf {y}) = \operatorname {a r c c o s h} \left(1 + 2 \frac {\| \mathbf {x} - \mathbf {y} \| ^ {2}}{\left(1 - \| \mathbf {x} \| ^ {2}\right) \left(1 - \| \mathbf {y} \| ^ {2}\right)}\right). \tag {2} +$$ + +Recall that a geodesic is a locally minimized-length curve between two points. In the hyperboloid model, the geodesic can be defined as the curve created by intersecting the plane defined by two points and the origin with the hyperboloid [38]. Thus, the mean of two latent codes in hyperbolic space locates at the mid-point of the geodesic that is closer to the origin. This is the key desired feature of hyperbolic space, i.e., the mean between two leaf embeddings is not another leaf embedding, but the hierarchical parent of them [55]. This feature allows us to generate new images by moving the latent code from one leaf to another with the + +same parents. We can also change the semantic levels of attributes by determining how abstract their parent is. + +This unique property is visualized in Fig. 3 on a 2-D Poincaré disk. The image embedding near the edge of the ball (with a large radius) represents a more fine-grained image while the embedding near the center (which has a smaller radius) represents an image with abstract features (an average face). + +Although the hyperbolic space shares similar features with trees, it is continuous. In other words, there is no fixed number of hierarchy levels. Instead, there is a continuum from very fine-grained (near the edge of Poincaré disk) to very abstract (near the origin). + +# 3.2. Network Architecture + +Although we aim to embed and edit real images in hyperbolic space, the whole network does not need to be implemented in a hyperbolic manner. Instead, we can take advantage of the number of existing GAN inversion models and optimization algorithms that have been fine-tuned for Euclidean space. + +To achieve image editing, we need to embed the image back into the latent space. In particular, we select pSp [49] as the backbone of HAE to encode images to the $\mathcal{W}^+$ -space of StyleGAN2 [34]: + +$$ +\mathbf {w} _ {i} = \mathrm {p S p} \left(x _ {i}\right), \tag {3} +$$ + +where $\mathbf{w}_i\in \mathbb{R}^{18\times 512}$ is the corresponding latent vector of $x_{i}$ in the $\mathcal{W}^+$ -space. + +To manipulate latent code in hyperbolic space, we need to define a bijective map from $\mathbb{R}^n$ to $\mathbb{D}_c^n$ to map Euclidean vectors to the hyperbolic space and vice versa. A manifold is a differentiable topological space that locally resembles the Euclidean space $\mathbb{R}^n$ [37, 38]. For $x\in \mathbb{D}^n$ , one can define the tangent space $T_{x}\mathbb{D}_{c}^{n}$ of $\mathbb{D}_c^n$ at $x$ as the first order linear approximation of $\mathbb{D}_c^n$ around $x$ . Therefore, this bijective map can be performed by exponential and logarithmic maps. Specifically, the exponential map $\exp_{\mathbf{x}}^{c}:T_{\mathbf{x}}\mathbb{D}_{c}^{n}\cong \mathbb{R}^{n}\to \mathbb{D}_{c}^{n}$ , maps from the tangent spaces into the manifold. While the logarithmic map $\log_{\mathbf{x}}^{c}:\mathbb{D}_{c}^{n}\to T_{\mathbf{x}}\mathbb{D}_{c}^{n}\cong \mathbb{R}^{n}$ is the reverse map of the exponential map. + +We use exponential and logarithmic maps at origin $\mathbf{0}$ for the transformation between the Euclidean and hyperbolic representations. After getting $\mathbf{w}_i$ in the $\mathcal{W}^+$ -space, we first use a Multi-layer Perceptron (MLP) encoder to reduce the dimension of latent vectors in Euclidean space. Then we apply an exponential map to project the Euclidean latent code $z_{\mathbb{R}i}$ to hyperbolic space. After that, we use the hyperbolic feed-forward layer as [18] to obtain the final hierarchical representation $z_{\mathbb{D}}$ as shown in Fig. 2: + +$$ +z _ {\mathbb {D} i} = f ^ {\otimes_ {c}} \left(\exp_ {\mathbf {0}} ^ {c} \left(\mathrm {M L P} _ {E} (\mathbf {w} _ {i})\right)\right), \tag {4} +$$ + +![](images/8dac1d1444f682bdbff58e0c7b8fae390dfb8bc5f46ce3b9eef0509faecf5432.jpg) +Figure 4: One-shot image generation from HAE on Animal Faces, Flowers, and VGGFaces. + +![](images/7255d8b40b12993f122298bd7ea3eeb3ebea7af22453f64765212fb3e83278c5.jpg) + +where $f^{\otimes_c}$ is the Möbius translation of feed-forward layer $f$ as the map from $\mathbb{D}_c^n$ to $\mathbb{D}_c^m$ , denoted as Möbius linear layer. + +Finally, the hyperbolic representation $z_{\mathbb{D}}$ needs to be projected back to the $\mathcal{W}^+$ -space of StyleGAN2. In practice, this is achieved by applying a logarithmic map followed by an MLP decoder: + +$$ +\mathbf {w} _ {i} ^ {\prime} = \operatorname {M L P} _ {D} \left(\log_ {\mathbf {0}} ^ {c} \left(z _ {\mathbb {D} i}\right)\right), \tag {5} +$$ + +and $\mathbf{w}_i^{\prime}$ will be fed into a pre-trained StyleGAN2's generator $G$ to reconstruct the image $x_{i}^{\prime}$ . + +# 3.3. Loss Function + +The loss function of HAE consists of two parts: the Hyperbolic loss ensures to get the hierarchical representation in the hyperbolic space and the reconstruction loss guarantees the quality of reconstruction images. + +Hyperbolic Loss. To learn the semantic hierarchical representation of real images in hyperbolic space, we minimize the distance between latent codes of images with similar categories and attributes while pushing away the latent codes from different categories. We choose the supervised approach to achieve this. In order to perform multi-class classification on the Poincaré disk defined in Sec. 3.1, one needs to generalize multinomial logistic regression (MLR) to the Poincaré disk defined in [18]. An extra linear layer needs to be trained for the classification and the softmax probability can be computed as: Given $K$ classes and $k \in \{1, \ldots, K\}$ , $p_k \in \mathbb{D}_c^n$ , $a_k \in T_{p_k} \mathbb{D}_c^n \setminus \{\mathbf{0}\}$ : + +$$ +\begin{array}{l} p (y = k \mid x) \propto \exp \left(\frac {\lambda_ {p _ {k}} ^ {c} \| a _ {k} \|}{\sqrt {c}} \sinh^ {- 1} \right. \\ \left(\frac {2 \sqrt {c} \langle - p _ {k} \oplus_ {c} x , a _ {k} \rangle}{\left(1 - c \| - p _ {k} \oplus_ {c} x \| ^ {2}\right) \| a _ {k} \|\right)\left. \right), \\ \forall x \in \mathbb {D} _ {c} ^ {n}. \tag {6} \\ \end{array} +$$ + +where $\oplus_{c}$ denotes the Möbius addition defined in [35] with fixed sectional curvature of the space, denoted by $c$ . + +After getting the softmax result for each class, one can use negative log-likelihood loss (NLL Loss) to calculate the + +hyperbolic loss: + +$$ +\mathcal {L} _ {\text {h y p e r}} = - \frac {1}{N} \sum_ {n = 1} ^ {N} \log \left(p _ {n}\right), \tag {7} +$$ + +where $N$ is the batch size and $p_n$ is the probability predicted by the model for the correct class. + +As mentioned in Sec. 3.1, the distance between points grows exponentially with their radius in the Poincaré disk. In order to minimize Eq. (7), the latent codes of fine-grained images will be pushed to the edge of the ball to maximize the distances between different categories while the embedding of abstract images (images have common features from many categories) will be located near the center of the ball. Since hyperbolic space is continuous and differentiable, we are able to optimize Eq. (7) with stochastic gradient descent, which learns the hierarchy of the images. + +Reconstruction Loss. In order to guarantee the quality of the generated images, we first use the $\mathcal{L}_2$ loss and LPIPS loss used in pSp [49], given image $x_{i}$ : + +$$ +\mathcal {L} _ {2} \left(x _ {i}\right) = \left\| x _ {i} - \mathrm {H A E} \left(x _ {i}\right) \right\| _ {2}. \tag {8} +$$ + +$$ +\mathcal {L} _ {\mathrm {L P I P S}} \left(x _ {i}\right) = \| F \left(x _ {i}\right) - F (\mathrm {H A E} \left(x _ {i}\right)) \| _ {2}, \tag {9} +$$ + +where $F(\cdot)$ denotes the perceptual feature extractor. + +Since the pSp encoder and StyleGAN2 generator are pretrained, we only train the neural layers between the encoder and generator of HAE. To further guarantee the network to better project back to the $\mathcal{W}^+$ -space, the reconstructed $\mathbf{w}_i'$ should be the same as the original $\mathbf{w}_i$ : + +$$ +\mathcal {L} _ {\text {r e c}} \left(w _ {i}\right) = \left\| \mathbf {w} _ {\mathrm {i}} - \mathbf {w} _ {i} ^ {\prime} \right\| _ {2}, \tag {10} +$$ + +where $\mathbf{w}_i^{\prime}$ can be calculated by Eq. (4) and Eq. (5). + +The overall loss function is: + +$$ +\mathcal {L} = \mathcal {L} _ {2} \left(x _ {i}\right) + \lambda_ {1} \mathcal {L} _ {\mathrm {L P I P S}} + \lambda_ {2} \mathcal {L} _ {\text {r e c}} + \lambda_ {3} \mathcal {L} _ {\text {h y p e r}}, \tag {11} +$$ + +where $\lambda_{1},\lambda_{2}$ and $\lambda_{3}$ are trade-off adaptive parameters. This curated set of loss functions ensures the model learns the hierarchical representation and reconstructs images. + +![](images/5fa26aab41441e0d1d0c821143c058787a2d03e1ad5b6b95a160abf7c17914ae.jpg) +Figure 5: Images generated by HAE by adding the same perturbation on the latent code of a given image with different hyperbolic radii on Animal Faces and Flowers. + +![](images/0c6302289e3089c5e2b0504fa904a12bc7da1c482eef0dd11c4b89f9ff215842.jpg) +Figure 6: Interpolations in hyperbolic space along the edge of the Poincaré disk (with $r_{\mathbb{D}} = 6.2126$ ) on three datasets. + +# 3.4. Image Generation + +To study the generating quality of the model, a straightforward way is to generate new images via interpolation between two designated images, or random perturbation. + +In hyperbolic space, the shortest path with the induced distance between two points is given by the geodesic defined in Eq. (2). The geodesic equation between two embeddings $z_{\mathbb{D}\mathbf{i}}$ and $z_{\mathbb{D}\mathbf{j}}$ , denoted by $\gamma z_{\mathbb{D}i}\to z_{\mathbb{D}j}(t)$ , is given by + +$$ +\gamma_ {z _ {\mathbb {D} i} \rightarrow z _ {\mathbb {D} j}} (t) = z _ {\mathbb {D} i} \oplus_ {c} t \otimes_ {c} \left(\left(- z _ {\mathbb {D} i}\right) \oplus_ {c} z _ {\mathbb {D} j}\right), t \in [ 0, 1 ], \tag {12} +$$ + +where $\oplus_{c}$ denotes the Möbius addition with aforementioned sectional curvature $c$ , with details in supplementary material. + +We adopt the following method to achieve generating via perturbation: For a given image $x_{i}$ , we first rescale its embedding $z_{\mathbb{D}i}$ to the desired radius $r_{\mathbb{D}}$ . Then we sample a random vector from seen categories in $z_{\mathbb{D}j}$ with radius $r_{\mathbb{D}}$ fixed and take the geodesic as the direction of perturbation to generate images. + +# 4. Experiment + +# 4.1. Implementation Details + +In the training stage, we first train a StyleGAN2 [33] and pSp [49] with seen categories. Given a trained pSp, the + +![](images/9ec7a49ed547804cdbf2da6ee8e40cbb514291eee3b8ed893497fff31ea49295.jpg) +Figure 7: Interpolations by moving the latent codes from the edge to the center of the Poincaré disk (from fine-grained to abstract) on three datasets. + +MLP encoder $\mathrm{MLP}_E$ is an 8-layer MLP with a Leaky-ReLU activation function. The dimension of the latent code in hyperbolic space is chosen to be 512. More details can be found in the supplementary. + +# 4.2. Datasets + +We evaluate our method on Animal Faces [40], Flowers [46], and VGGFaces [48] following the settings described in [15]. + +Animal Faces. We randomly select 119 categories as seen for training and leave 30 as unseen categories for testing. + +Flowers. The Flowers [46] dataset is split into 85 seen categories for training and 17 unseen categories for testing. + +VGGFaces. For VGGFaces [48], we randomly select 1802 categories for training and 572 for testing. + +# 4.3. Analysis of Hierarchical Feature Editing + +We analyze the properties of the learned hierarchical representations and how the levels of attributes relate to their locations of latent codes in hyperbolic space. + +As we mentioned in Sec. 3.1, there is a continuum from fine-grained attributes to abstract attributes, corresponding to the points from the peripheral to the center of the ball. We define the hyperbolic radius $r_{\mathbb{D}}^2$ as the hyperbolic dis + +
MethodSettingsFlowersAnimal FacesVGG Faces*
FID(↓)LPIPS(↑)FID(↓)LPIPS(↑)FID(↓)LPIPS(↑)
DAWSON [39]3-shot188.960.0583208.680.0642137.820.0769
MatchingGAN [28]3-shot143.350.1627148.520.1514118.620.1695
F2GAN [30]3-shot120.480.2172117.740.1831109.160.2125
LoFGAN [23]3-shot79.330.3862112.810.496420.310.2869
DeltaGAN [27]1-shot109.780.391289.810.441880.120.3146
Disco-FUNIT [29]1-shot90.120.443671.440.4511--
AGE [15]1-shot45.960.430528.040.557534.860.3294
SAGE [14]1-shot43.520.439227.430.544834.970.3232
HAE (Ours)1-shot50.100.473926.330.563635.930.5919
+ +Table 1: FID(↓) and LPIPS(↑) of images generated by different methods for unseen categories on three datasets. Bold indicates the best results and underline indicates the second best results. VGGFaces is marked with * because different methods report different numbers of unseen categories on this dataset(e.g. 552 in LoFGAN, 96 in DeltaGAN, 497 in L2GAN, and 572 in AGE and SAGE). Note that: disco-FUNIT [29] does not provide pre-trained models on VGG Faces [48] dataset. + +
InputMatchingGANLoFGANAGEHAE(ours)
+ +Figure 8: Comparison between images generated by MatchingGAN, LoFGAN, AGE, and HAE on Flowers, Animal Faces, and VGGFaces. Zoom in to see the details. Note that: SAGE [14] has not released code and pre-trained models. + +tance of the given latent code to the center of the Poincaré disk. To study the influence of the radius of embeddings in hyperbolic space, we run several experiments with different choices of $r_{\mathbb{D}}$ . + +Hyperbolic Perturbation and Interpolation. As mentioned in Sec. 3.4, we demonstrate the results of perturbation and interpolation. In addition to the choice of perturbation, we can set the intensity of the perturbation by controlling both the step distance and radius as shown in Fig. 5. The results show that level of semantic attributes is highly related to $r_{\mathbb{D}}$ . With the radius becoming smaller, the attributes become more abstract. We further visualize this property of hyperbolic space by moving the latent codes of the given image from the edge of the Poincaré disk to the center. As Fig. 7 shows, the images change from very fine-grained to very abstract (the average of all images). The results in Fig. 6 show that we can achieve smooth interpola + +tion in hyperbolic space without any distortion. The results demonstrate that with HAE, we can freely control the editing geodesically and hierarchically. + +# 4.4. Few-shot Image Generation + +As Fig. 5 shows, the image categories will be changed when $r_{\mathbb{D}}$ is smaller than about 4, and the category-irrelevant attributes of images will be changed when $r_{\mathbb{D}}$ is larger than about 5. The embeddings of Animal Faces are visualized in 2-D Poincaré disk using UMAP [43] shown in Fig. 9. As Fig. 6 shows, the posture and the angle of the images will be changed at the early stage of interpolation without changing the category. Thus, the images can be generated by moving the latent code of a given image to some randomly selected semantic direction within the cluster of the category. In practice, we select $r_{\mathbb{D}} = 6.21$ and step size of perturbation as 8 to achieve few-shot image generation as + +![](images/c88386d26e9c5591abf100e3ba414141752462191b577ad1bc9ad47b6ea573ce.jpg) +Figure 9: UMAP visualization of hyperbolic 2-D embeddings of Animal Faces dataset. We observe that similar categories are clustered and positioned near the boundary, while ambiguous samples are located near the center. Zoom in to see the details. + +Fig. 4 shows the diverse images generated by adding random perturbations from seen categories. We conduct three experiments to show that HAE can achieve promising few-shot image generation. More examples of generated images are available in the supplementary. + +Quantitative Comparison with State-of-the-art. We calculate the FID [24] and LPIPS [59] to evaluate the fidelity and diversity of the generated images following one-shot settings in [15, 14]. The comparison results are shown in Tab. 1. Our method achieves the best scores on most of the FID and LPIPS metrics compared with state-of-the-art few-shot image generation methods, which indicates that our method not only improves the model from the semantic aspect but also achieves state-of-the-art performance on the traditional evaluation metrics. Specifically, the LPIPS score of HAE beats SOTA model SAGE on all three datasets since HAE can generate more diverse images. + +Qualitative Evaluation. We qualitatively compare our method with MatchingGAN [28], LoFGAN [23], DeltaGAN [27] and AGE [15]. As shown in Fig. 8, HAE can generate images with diversity and fine-grained details. More importantly, the newly generated images have more semantic diversity than others. For instance, the shadow and skin color of the generated faces change with the light condition, and this effect looks more natural. We further conduct a user study by randomly selecting 60 (20 from each dataset) images with generated variants using AGE and HAE. 50 users from different backgrounds are asked to rate the results only based on diversity and quality external information. This is achieved by randomly shuffling the order of images pairwisely and inside any pair. HAE won by a ratio of $58.1\%$ (1743/3000) over AGE (more details in supplementary). + +![](images/dff337300422ef151eb250686ca8c16928f26d0703f014aab3eecbb8e295ba58.jpg) +Figure 10: Manipulate images from different categories with the same perturbation (Target 1&2). + +Transferability. If we move latent codes at category-irrelevant levels, the target perturbation is transferable across all categories. We edit the images from three categories with the same editing direction, the output images are shown in Fig. 10. It demonstrates that HAE achieves a highly controllable and interpretable editing process. + +
MethodFlowersAnimal FacesVGG Faces
FIDLPIPSFIDLPIPSFIDLPIPS
SAGE [14]43.520.439227.430.544834.970.3232
HAE(Euc)54.620.429325.270.512938.460.5908
HAE(Hyp)50.100.473926.330.563635.930.5919
+ +Table 2: $\mathrm{{FID}}\left( \downarrow \right)$ and LPIPS(↑) of images generated by HAE in different geometries for unseen categories on three datasets. Bold indicates the best results and underline indicates the second best results. + +# 4.5. Ablation Study + +HAE in Euclidean. We re-trained HAE models in Euclidean space with the NLL loss to validate the performance gain in Tab. 1 is due to the hierarchical hyperbolic representation rather than the disentanglement caused by Eq. (7). The quantitative comparison is shown in Tab. 2. It shows that the hyperbolic space boosts the performance, especially for the LPIPS score, since the latent code is more disentangled in hyperbolic space [19]. This finding is also supported by the UMAP visualization in Fig. 9. More details can be found in the supplementary material. + +Hyperbolic Radius versus Truncation. StyleGAN [33] uses truncation trick [42, 4, 33, 34] in $\mathcal{W}$ -space to achieve the balance between the image quality and diversity. The experiments in [33, 34] also show that the truncation level in $\mathcal{W}^+$ -space control the level of abstraction of the generated images. We conduct the experiments in Sec. 3.4 using truncation to validate the gains of hyperbolic space. The results are illustrated in Fig. 11 and Fig. 12. As Fig. 11 shows, the category of the image changes along with the posture of the dog as the truncation gets smaller, while the category-relevant attributes do not change when the hyperbolic radius + +![](images/3aed1d95f93aeb57be8519dddde61107148227012a529009129067feb1f24650.jpg) +Figure 11: Top: Interpolations by moving the latent codes from the edge to the center in hyperbolic space. Bottom: Interpolation with different truncation in Euclidean space. Zoom in to see the details. + +![](images/9b86bda6924f822f791aa9d247f422d569275c81d1d371940aaf6f6fc7a33c96.jpg) +Figure 12: Images generated by HAE(Hyp) and HAE(Euc) by adding the same perturbation on the latent code of a given image with different settings of hyperbolic radius and truncation. Zoom in to see the details. + +$(r_{\mathbb{D}})$ is large. This can also be proved in Fig. 12. The category remains the same after adding perturbation when $r_{\mathbb{D}}$ is large, while the truncation can not control semantic-level editing. This shows that Euclidean space can only capture scale-based hierarchy rather than the semantic hierarchy. + +Downstream Task. We conduct data augmentation via HAE for image classification on Animal Faces [48]. We randomly select 30, 35, and 35 images for each category as train, val, and test, respectively. Following [23], a ResNet-18 backbone is initialized from the seen categories. Then the model is fine-tuned on the unseen categories referred to as the baseline. 60 images are generated for each unseen category as data augmentation. The result is presented in Tab. 3. The diversity and quality of generated images are primarily controlled by the hyperbolic radii $r_{\mathbb{D}}$ . As the radius becomes smaller, HAE generates images of higher diversity, but categories (referring to high-level attributes) also gradually change to others. $r_{\mathbb{D}} = 6$ achieves the best performance on the classification experiment. However, the performance drops when the radius is smaller than 4.5. This is because the semantic attributes change too much and thus mislead the classifier. + +# 4.6. Limitations and Future Work + +Although HAE achieves reliable hierarchical attribute editing in hyperbolic space for few-shot image generation, there are several limitations. + +First, the boundary of category changing is hard to be quantified since the hierarchical levels are continuous in + +
Hyperbolic RadiusAccuracyFID(↓)LPIPS(↑)
baseline58.67--
6.060.1046.890.4520
5.559.5248.680.4651
5.059.0552.080.4823
4.559.1460.870.5174
4.058.5765.830.5386
3.556.8668.440.6034
3.054.1469.400.6316
+ +Table 3: Ablation of same perturbation on different radii on Animal Faces. + +the hyperbolic space. Users need to find a "safe" boundary by trying different radii and step sizes of perturbation before generating new images. However, from another perspective, this continuity of hierarchy provides flexibility for users to set different boundaries for different downstream tasks as they need. + +Second, the performance of HAE is limited by the pretrained styleGAN and the inversion method. If the input image can not be well embedded, the editing will also fail. This problem can be solved by changing more powerful backbones, e.g., ViT [16], in future work. + +Finally, we use supervised learning to get the hierarchical embedding in hyperbolic space. However, the number of images in existing datasets with labels for generation tasks is relatively small, which makes the embeddings in the hyperbolic space not evenly distributed. The solution to this problem is simple, use unsupervised learning with large-scale high-quality datasets. + +# 5. Conclusion + +In this work, we propose a simple yet effective method HAE to edit hierarchical attributes in hyperbolic space. After learning the semantic hierarchy from images, our model is able to edit continuous semantic hierarchical features of images for flexible few-shot image generation in the hyperbolic space. Experiments demonstrate that HAE is capable of achieving not only stable few-shot image generation with state-of-the-art quality and diversity but a controllable and interpretable editing process. Future work includes the combination of HAE and large pretrained models and applications to more downstream tasks. + +Acknowledgement. This work was supported in part by the National Key R&D Program of China under Grant 2018AAA0102000, in part by National Natural Science Foundation of China: 62022083 and 62236008. + +# References + +[1] Antreas Antoniou, Amos Storkey, and Harrison Edwards. Data augmentation generative adversarial networks. arXiv preprint arXiv:1711.04340, 2017. 1, 2 +[2] Sergey Bartunov and Dmitry P. Vetrov. Few-shot generative modelling with generative matching networks. In AISTATS, 2018. 1 +[3] Silvère Bonnabel. Stochastic gradient descent on riemannian manifolds. IEEE Transactions on Automatic Control, 58(9):2217-2229, 2013. 3 +[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. In ICLR, 2019. 8 +[5] Gary Bécigneul and Octavian-Eugen Ganea. Riemannian adaptive optimization methods. In ICLR, 2019. 3 +[6] James W Cannon, William J Floyd, Richard Kenyon, and Walter R Parry. Hyperbolic geometry. Flavors of geometry, 31:59-115, 1997. 4 +[7] Ines Chami, Rex Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. In NeurIPS, page 4868-4879, 2019. 3 +[8] Anton Cherepkov, Andrey Voynov, and Artem Babenko. Navigating the gan parameter space for semantic image editing. In CVPR, pages 3670-3679, 2021. 3 +[9] Jaewoong Choi, Junho Lee, Changyeon Yoon, Jung Ho Park, Geonho Hwang, and Myungjoo Kang. Do not escape from the manifold: Discovering the local coordinates on the latent space of gans. In ICLR, 2022. 3 +[10] Louis Clouatre and Marc Demers. Figr: Few-shot image generation with reptile. arXiv:1901.02199, 2019. 1, 2 +[11] Jiali Cui, Ying Nian Wu, and Tian Han. Learning joint latent space ebm prior model for multi-layer generator. In CVPR, pages 3603-3612, 2023. 2 +[12] Emily Denton, Ben Hutchinson, Margaret Mitchell, and Timnit Gebru. Detecting bias with generative counterfactual face attribute augmentation. CoRR, abs/1906.06439, 2019. 3 +[13] Bhuwan Dhingra, Chris Shallue, Mohammad Norouzi, Andrew Dai, and George Dahl. Embedding text in hyperbolic spaces. arXiv preprint arXiv:1806.04313, 2018. 2 +[14] Guanqi Ding, Xinzhe Han, Shuhui Wang, Xin Jin, Dandan Tu, and Qingming Huang. Stable attribute group editing for reliable few-shot image generation. arXiv preprint arXiv:2302.00179, 2023. 2, 7, 8 +[15] Guanqi Ding, Xinzhe Han, Shuhui Wang, Shuzhe Wu, Xin Jin, Dandan Tu, and Qingming Huang. Attribute group editing for reliable few-shot image generation. In CVPR, pages 11184-11193, 2022. 1, 2, 6, 7, 8 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 9 +[17] Kun Fu, Tengfei Zhang, Yue Zhang, Menglong Yan, Zhonghan Chang, Zhengyuan Zhang, and Xian Sun. Meta-ssd: Towards fast adaptation for few-shot object detection with meta-learning. IEEE Access, 7:77597-77606, 2019. 1 + +[18] Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. In NeurIPS, pages 5345-5355, 2018. 3, 4, 5 +[19] Songwei Ge, Shlok Mishra, Simon Kornblith, Chun-Liang Li, and David Jacobs. Hyperbolic contrastive learning for visual representations beyond objects. In CVPR, 2023. 8 +[20] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Phillip Isola. Ganalyze: Toward visual definitions of cognitive image properties. In CVPR, page 5744-5753, 2019. 3 +[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2 +[22] Michael Gromov. Hyperbolic groups. In Essays in group theory, 1987. 3 +[23] Zheng Gu, Wenbin Li, Jing Huo, Lei Wang, and Yang Gao. Lofgan: Fusing local representations for fewshot image generation. In ICCV, 2021. 1, 2, 7, 8, 9 +[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 8 +[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In ICLR, 2020. 2 +[26] Yan Hong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Deltagan: Towards diverse few-shot image generation with sample-specific delta. In CVPR, 2020. 1 +[27] Yan Hong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Deltagan: Towards diverse few-shot image generation with sample-specific delta. In ECCV, 2022. 1, 2, 7, 8 +[28] Yan Hong, Li Niu, Jianfu Zhang, and Liqing Zhang. Matchinggan: Matching-based few-shot image generation. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2020. 1, 2, 7, 8 +[29] Yan Hong, Li Niu, Jianfu Zhang, and Liqing Zhang. Few-shot image generation using discrete content representation. In Proceedings of the 30th ACM International Conference on Multimedia, MM '22, page 2796-2804, New York, NY, USA, 2022. Association for Computing Machinery. 1, 2, 7 +[30] Yan Hong, Li Niu, Jianfu Zhang, Weijie Zhao, Chen Fu, and Liqing Zhang. F2gan: Fusing-and-filling gan for few-shot image generation. In Proceedings of the 28th ACM International Conference on Multimedia, MM '20, page 2535-2543. Association for Computing Machinery, 2020. 1, 2, 7 +[31] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In NeurIPS, 2020. 3 +[32] Ali Jahanian, Lucy Chai, and Phillip Isola. "On the" steerability" of generative adversarial networks. In ICLR, 2020. 3 +[33] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4217-4228, 2019. 6, 8 +[34] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, pages 8107-8116, 2020. 1, 4, 8 + +[35] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In CVPR, pages 6417-6427, 2020. 2, 3, 4, 5 +[36] Diego Lazcano, Nicolas Fredes Franco, and Werner Creixell. Hgan: Hyperbolic generative adversarial network. IEEE Access, 9:96309-96320, 2021. 3 +[37] John M Lee. Riemannian manifolds: an introduction to curvature. Springer Science & Business Media, 176, 2006. 4 +[38] John M Lee. Introduction to Smooth Manifolds. Springer, 2013. 4 +[39] Weixin Liang, Zixuan Liu, and Can Liu. Dawson: A domain adaptive few shot generation framework. arXiv preprint arXiv:2001.00576, 2020. 1, 2, 7 +[40] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsueprivised image-to-image translation. In ICCV, 2019. 6 +[41] Yu-Ding Lu, Hsin-Ying Lee, Hung-Yu Tseng, and Ming-Hsuan Yang. Unsupervised discovery of disentangled manifolds in gans. arXiv preprint arXiv:2011.11842, 2020. 3 +[42] Marco Marchesi. Megapixel size image creation using generative adversarial networks. CoRR, abs/1706.00082, 2017. 8 +[43] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Grossberger. Umap: Uniform manifold approximation and projection. The Journal of Open Source Software, 3(29):861, 2018. 7 +[44] Maximillian Nickel and Douwe Kiela. Generative visual manipulation on the natural image manifold. In ECCV, 2017. 2, 3, 4 +[45] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In ICML, 2018. 3, 4 +[46] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 6 +[47] Jiwoong Park, Junho Cho, Hyung Jin Chang, and Jin Young Choi. Unsupervised hyperbolic representation learning via message passing auto-encoders. In CVPR, pages 5512–5522, 2021. 3 +[48] Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In British Machine Vision Conference, 2015. 6, 7, 9 +[49] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In CVPR, pages 2287–2296, 2021. 4, 5, 6 +[50] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, pages 9240-9249, 2020. 3 +[51] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, pages 1532-1540, 2021. 3 +[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3 + +[53] Nurit Spingarn-Eliezer, Ron Banner, and Tomer Michaeli. Gan "steerability" without optimization. In ICLR, 2021. 3 +[54] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H.S. Torr, and Timothy M. Hospedales. Learning to compare: Relation network for few-shot learning. In CVPR, pages 1199-1208, 2018. 1 +[55] Dídac Surís, Ruoshi Liu, and Carl Vondrick. Learning the predictability of the future. In CVPR, pages 12602-12612, 2021. 2, 3, 4 +[56] Alexandru Tifrea, Gary Bécigneul, and OctavianEugen Ganea. Poincaré glove: Hyperbolic word embeddings. In ICLR, 2019. 2, 3, 4 +[57] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NeurIPS, 2016. 1 +[58] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In ICML, pages 9786-9796, 2020. 3 +[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 8 +[60] Jingyuan Zhu, Huimin Ma, Jiansheng Chen, and Jian Yuan. Few-shot image generation with diffusion models. arXiv preprint arXiv:2211.03264, 2022. 2 \ No newline at end of file diff --git a/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/images.zip b/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..499395dbc883e24be32cfa0055122375459eee9a --- /dev/null +++ b/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86fd8d7698fecf04e3e7e76242a8aac7ee1f060bc2e8aca6bf0b288e3abcaa87 +size 796228 diff --git a/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/layout.json b/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e92b72c4b7da6fa55a462098a2912630b79a1de5 --- /dev/null +++ b/2023/The Euclidean Space is Evil_ Hyperbolic Attribute Editing for Few-shot Image Generation/layout.json @@ -0,0 +1,9688 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 118, + 103, + 477, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 103, + 477, + 140 + ], + "spans": [ + { + "bbox": [ + 118, + 103, + 477, + 140 + ], + "type": "text", + "content": "The Euclidean Space is Evil: Hyperbolic Attribute Editing for Few-shot Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 189, + 160, + 407, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 160, + 407, + 176 + ], + "spans": [ + { + "bbox": [ + 189, + 160, + 407, + 176 + ], + "type": "text", + "content": "Lingxiao Li1 Yi Zhang2 Shuhui Wang3*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 179, + 188, + 414, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 188, + 414, + 203 + ], + "spans": [ + { + "bbox": [ + 179, + 188, + 414, + 203 + ], + "type": "text", + "content": "1 Columbia University 2 University of Oxford" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 203, + 460, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 203, + 460, + 217 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 460, + 217 + ], + "type": "text", + "content": "3 Institute of Computing Technology, Chinese Academy of Sciences" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 220, + 459, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 220, + 459, + 230 + ], + "spans": [ + { + "bbox": [ + 134, + 220, + 459, + 230 + ], + "type": "text", + "content": "113504@columbia.edu, wolf5965@ox.ac.uk, wangshuhui@ict.ac.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 284, + 290, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 290, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 290, + 523 + ], + "type": "text", + "content": "Few-shot image generation is a challenging task since it aims to generate diverse new images for an unseen category with only a few images. Existing methods suffer from the trade-off between the quality and diversity of generated images. To tackle this problem, we propose Hyperbolic Attribute Editing (HAE), a simple yet effective method. Unlike other methods that work in Euclidean space, HAE captures the hierarchy among images using data from seen categories in hyperbolic space. Given a well-trained HAE, images of unseen categories can be generated by moving the latent code of a given image toward any meaningful directions in the Poincaré disk with a fixing radius. Most importantly, the hyperbolic space allows us to control the semantic diversity of the generated images by setting different radii in the disk. Extensive experiments and visualizations demonstrate that HAE is capable of not only generating images with promising quality and diversity using limited data but achieving a highly controllable and interpretable editing process. Code is available at https://github.com/lingxiao-li/HAE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 544, + 128, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 128, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 128, + 557 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 564, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 696 + ], + "type": "text", + "content": "Due to the persistent development of deep learning, the task of image generation has received significant research attention in recent years. Specifically, the Generative Adversarial Networks (GANs) [21] and its variants (e.g., StyleGANv2 [34]) have succeeded in generating high-fidelity and realistic images, requiring a large number of high-quality data for model training. However, considering the long-tail distribution and data imbalance widely exists among different image categories [30], it is difficult for GANs to be trained on categories with sufficient training images to generate new realistic and diverse images for a" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 258, + 547, + 367 + ], + "blocks": [ + { + "bbox": [ + 309, + 258, + 547, + 367 + ], + "lines": [ + { + "bbox": [ + 309, + 258, + 547, + 367 + ], + "spans": [ + { + "bbox": [ + 309, + 258, + 547, + 367 + ], + "type": "image", + "image_path": "798e260bd080ccee3be0c563ba81492de9d13b09e3997c3ed9d29f1b17138eb7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 372, + 547, + 449 + ], + "lines": [ + { + "bbox": [ + 304, + 372, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 547, + 449 + ], + "type": "text", + "content": "Figure 1: Illustration of hierarchical attribute editing in hyperbolic space. Hyperbolic space can naturally and compactly encode semantic hierarchical structures within a large image data corpus. Changing the high-level, i.e., category-relevant attribute " + }, + { + "bbox": [ + 304, + 372, + 547, + 449 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{w}^r" + }, + { + "bbox": [ + 304, + 372, + 547, + 449 + ], + "type": "text", + "content": " changes the category of an image. While changing low-level or category-irrelevant attribute " + }, + { + "bbox": [ + 304, + 372, + 547, + 449 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{w}^{ir}" + }, + { + "bbox": [ + 304, + 372, + 547, + 449 + ], + "type": "text", + "content": " varies images within categories." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 471, + 547, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 547, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 547, + 531 + ], + "type": "text", + "content": "category with only a few images. This task is referred to as few-shot image generation [10, 26, 30, 28, 29, 27, 15]. A variety of tasks can benefit from improvements in few-shot image generation, for instance, low-data detection [17] and few-shot classification [54, 57]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 548, + 715 + ], + "type": "text", + "content": "In general, existing GAN-based few-shot image generation mechanisms can be classified into three categories. Transfer-based methods [10, 39] introduce meta-learning or domain adaptation on GANs to generate new images by enforcing knowledge transfer among categories. Fusion-based methods [2, 23, 30, 28] perform feature fusion of multiple input images in a feature space and generate images via decoding the fused features back to image space. However, the output is still highly similar to the source images. Transformation-based methods [1, 29, 27, 15] find intra-category transformations or inject random perturbations to conditional unseen category samples to generate images without tedious fine-tuning. By representing the images in the Euclidean feature space, the above learning mechanisms tend to be over-complicated, and the generated" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22714" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 253, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 253, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 253, + 84 + ], + "type": "text", + "content": "images are often collapsed due to limited diversity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 87, + 289, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 87, + 289, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 87, + 289, + 399 + ], + "type": "text", + "content": "Similar to the ubiquity of hierarchies in language [44, 56, 13], the semantic hierarchy is also common in images [35, 11]. As Fig. 1 shows, the semantic hierarchies constructed in the language domain can be instantiated with visual images. From the visual perspective, an image can be regarded as a collection of attributes of multiple levels. High-level attributes, a.k.a. category-relevant attributes, define the category of an image, such as the shape and color of an animal [15]. For instance, in the middle row of Fig. 1, changing the high-level attributes of the given image of a Shih-Tzu dog, the category can be changed to a Rhodesian Ridgeback Dog. While the low-level or fine-grained attributes, including expressions, postures, etc., that vary within the category as shown at the bottom of Fig. 1, are called category-irrelevant attributes. Therefore, an image can also be viewed as a descendant of another image with the same category-relevant attributes by adding fine-grained category-irrelevant attributes to its parent image. To edit the visual attributes for high-quality image generation, it is crucial to capture the attribute hierarchy within the large image data corpus and find a good representation space. Ideally, we aim to construct a hierarchical visual representation in a latent space that allows us to change the category of an image by moving the latent code in a category-relevant direction, and perform few-shot image generation by moving the code in a category-irrelevant direction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 400, + 289, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 289, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 289, + 604 + ], + "type": "text", + "content": "Unfortunately, the Euclidean space and its corresponding distance metrics used by existing GAN-based methods can not facilitate the hierarchical attribute representation, thus the design of complicated attribute disentangling and editing mechanisms seems to be crucial for the generation quality. Inspired by the application of hyperbolic space in images [35] and videos [55], we found that the metrics introduced in hyperbolic geometry can naturally and compactly encode hierarchical structures. Unlike the general affine spaces, e.g., the Euclidean space, hyperbolic spaces can be viewed as the continuous analog of a tree since tree-like graphs can be embedded in finite-dimension with minimal distortion [44]. This property of hyperbolic space provides continuous and up to infinite semantic levels for attribute editing, allowing us to robustly generate diverse images with only a few images from unseen categories with simple operations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "Based on the above findings, we propose a simple but effective Hyperbolic Attribute Editing (HAE) method for few-shot image generation. Our method is based on the observation that hierarchical latent code manipulation can be easily implemented in Hyperbolic space. The core of HAE is mapping the latent vectors from the Euclidean space " + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": " to a hyperbolic space " + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{D}^n" + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": ". We minimize a supervised classification loss function to ensure the images are hierarchically embedded in hyperbolic space. Once we capture the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "attribute hierarchy among images, we can generate new images of unseen categories by moving the latent code from one leaf to another with the same parents by fixing the radius. Most importantly, the hyperbolic space allows us to control the semantic diversity of generated images by setting different radii in the Poincaré disk. Those operations can well facilitate continually hierarchical attribute editing in hyperbolic space for flexible few-shot image generation with both quality and diversity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 180, + 517, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 180, + 517, + 192 + ], + "spans": [ + { + "bbox": [ + 317, + 180, + 517, + 192 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 201, + 545, + 410 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 201, + 545, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 201, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 201, + 545, + 273 + ], + "type": "text", + "content": "- We propose a simple yet effective method for few-shot image generation, i.e., hyperbolic attribute editing. In order to capture the hierarchy among images, we use hyperbolic space as the latent space. To the best of our knowledge, HAE is the first attempt to use hyperbolic latent spaces for few-shot image generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 281, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 281, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 317, + 281, + 545, + 329 + ], + "type": "text", + "content": "- We show that in our designed hyperbolic latent space, the semantic hierarchical attribute relations among images can be reflected by their distances to the center of the Poincaré disk." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 338, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 545, + 410 + ], + "type": "text", + "content": "- Extensive experiments and visualization suggest that HAE achieves stable few-shot image generation with state-of-the-art quality and diversity. Unlike other few-shot image generation methods, HAE allows us to generate images with better control of diversity by changing the semantic levels of attributes we want to edit." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 430, + 392, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 392, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 392, + 442 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "Few-shot image generation. Recently, diverse methods have been proposed for few-shot image generation. The transfer-based methods [10, 39] which introduce meta-learning or domain adaptation on GANs can hardly generate realistic images. While fusion-based methods that fuse the features by matching the random vector with the conditional images [28] or formulating the problem as a conditional generating task [23, 30] suffer from the limited diversity of generated images. Furthermore, transformation-based methods [1, 29, 27, 15] can generate images with only one conditional image by focusing on either capturing the cross-category or intra-category transformations by injecting random perturbations [1]. Nevertheless, the transformation captured by those methods is not very consistent. Ding et al. [15, 14] propose the \"editing-based\" perspective, the intra-category transformation can be modeled as category-irrelevant image editing based on one sample instead of pairs of samples. Most recently, Zhu et al. [60] fine-tune powerful diffusion models (DMs) [25] pre-trained on large source domains on limited target data to generate diverse and high quality images. DMs outperform GANs [21] on sample quality with a more controllable training process" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22715" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 77, + 541, + 246 + ], + "blocks": [ + { + "bbox": [ + 53, + 77, + 541, + 246 + ], + "lines": [ + { + "bbox": [ + 53, + 77, + 541, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 77, + 541, + 246 + ], + "type": "image", + "image_path": "9372c400b973e3f1a5586f6beafe464dec6f37c4b21f87584f07c83e34f39aef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 252, + 547, + 299 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 547, + 299 + ], + "type": "text", + "content": "Figure 2: The overview of HAE. The Hyper layer is a hyperbolic feedforward layer called Möbius linear layer which is used to project the latent code from Euclidean space " + }, + { + "bbox": [ + 46, + 252, + 547, + 299 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 46, + 252, + 547, + 299 + ], + "type": "text", + "content": " to hyperbolic space " + }, + { + "bbox": [ + 46, + 252, + 547, + 299 + ], + "type": "inline_equation", + "content": "\\mathbb{D}^n" + }, + { + "bbox": [ + 46, + 252, + 547, + 299 + ], + "type": "text", + "content": " [7]. " + }, + { + "bbox": [ + 46, + 252, + 547, + 299 + ], + "type": "inline_equation", + "content": "\\bar{z}_{\\mathbb{D}i}" + }, + { + "bbox": [ + 46, + 252, + 547, + 299 + ], + "type": "text", + "content": " can be viewed as the \"parent\" or average code of " + }, + { + "bbox": [ + 46, + 252, + 547, + 299 + ], + "type": "inline_equation", + "content": "z_{\\mathbb{D}i}" + }, + { + "bbox": [ + 46, + 252, + 547, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 252, + 547, + 299 + ], + "type": "inline_equation", + "content": "z_{\\mathbb{D}i}'" + }, + { + "bbox": [ + 46, + 252, + 547, + 299 + ], + "type": "text", + "content": ". One can generate diverse images without changing the category by moving the latent code from one child to another of the same parent in the hyperbolic space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 288, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 361 + ], + "type": "text", + "content": "at the cost less flexibility and editability, since they denoise images in the image space rather than operate in the latent space. Furthermore, the inference process of DMs is much slower than GANs [52]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 363, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 567 + ], + "type": "text", + "content": "Hyperbolic Embedding. The use of hyperbolic space in deep learning [44, 45, 56, 55, 35] is a pioneering work in recent years. It was first used in natural language processing for hierarchical language representation [44, 45, 56]. The Riemannian optimization algorithms are used to optimize models in hyperbolic space [5, 3]. As hyperbolic space is successfully applied to represent hierarchical data, Ganea et al. [18] derives hyperbolic versions of tools in neural networks including multinomial logistic regression, feed-forward, and recurrent neural networks. Following this, hyperbolic geometry is used in image [35], video [55], and graph data [7, 47]. Most recently, Lazcano et al [36] shows that hyperbolic space outperforms traditional Euclidean space in image generation using HGAN. However, the hierarchy and controllability of hyperbolic space remain uninvestigated in HGAN, as the generator is still governed by Gaussian samples in Euclidean space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Latent Code Manipulation. It has been shown that the latent spaces of GANs are able to encode rich semantic information [20, 32, 50]. One of the popular approaches is finding linear directions corresponding to changes in a given binary labeled attributes, which might be difficult to obtain for new datasets and could require manual labeling effort [50, 20, 12]. Others [8, 58, 41, 31, 9] try to find semantic directions in an unsupervised manner. For instance, PCA is applied in the latent space to create interpretable controls for synthesizing images [31, 9]. Most recent works [53, 51] directly compute in the close form to find the meaningful semantic direction without training and optimization. In com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 312, + 547, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 547, + 350 + ], + "type": "text", + "content": "parison, our work HAE focuses on attributes in different semantic levels in the latent space rather than trying hard to find disentangled interpretable directions as previous works." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 360, + 362, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 362, + 372 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 362, + 372 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 380, + 547, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 547, + 429 + ], + "type": "text", + "content": "The overall framework of HAE is shown in Fig. 2, we first give a detailed explanation of getting the hierarchical representations in the hyperbolic space, and then we introduce the framework of HAE and explain the loss functions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 436, + 463, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 436, + 463, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 436, + 463, + 449 + ], + "type": "text", + "content": "3.1. Hierarchical Representation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 455, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 455, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 455, + 545, + 514 + ], + "type": "text", + "content": "The major issue of our study is how to obtain the hierarchical representation from real images to facilitate editing in different semantic levels, as illustrated in Fig. 1. Therefore, hyperbolic space is introduced as the latent space to achieve this goal." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 515, + 546, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 546, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 546, + 634 + ], + "type": "text", + "content": "Unlike Euclidean spaces with their zero curvature and spherical spaces with their positive curvature, hyperbolic spaces with negative curvature have been shown that it is more appropriate for learning hierarchical representation [44, 45]. Informally, hyperbolic space can be viewed as a continuous analogy of trees [44]. One important feature of hyperbolic space is that the length grows exponentially with its radius while linearly in Euclidean space. This property allows hyperbolic space to be naturally compatible with hierarchical data [22] including text, images, videos, etc." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "text", + "content": "-dimensional hyperbolic space can be formally defined as a homogeneous, simply connected " + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "text", + "content": "-dimensional Riemannian manifold, denoted as " + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\mathbb{H}^n" + }, + { + "bbox": [ + 304, + 635, + 547, + 696 + ], + "type": "text", + "content": " with constant negative sectional curvature1. We choose to work in the Poincaré disk from five isometric models of hyperbolic" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 531, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 531, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 531, + 714 + ], + "type": "text", + "content": "1The curvature of the hyperbolic space " + }, + { + "bbox": [ + 317, + 702, + 531, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 317, + 702, + 531, + 714 + ], + "type": "text", + "content": " is set as " + }, + { + "bbox": [ + 317, + 702, + 531, + 714 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 317, + 702, + 531, + 714 + ], + "type": "text", + "content": " in this work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22716" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 272, + 277 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 272, + 277 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 272, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 272, + 277 + ], + "type": "image", + "image_path": "5f0c8e82bef45e739ea6849503339b029492966b9a2749af717d4385f97fde25.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "text", + "content": "Figure 3: Illustration of the property of hyperbolic space on the Poincaré disk. Given two latent codes of Spaniel " + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "inline_equation", + "content": "z_{\\mathbb{D}}^{c_1}" + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "inline_equation", + "content": "z_{\\mathbb{D}}^{c_1'}" + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "text", + "content": " (red dots) on the edge of Poincaré disk, the geodesic between these two points is the brown curve rather than a straight line in Euclidean space. Therefore, their average latent code is calculated as " + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "inline_equation", + "content": "\\bar{z}_{\\mathbb{D}}^{c_1}" + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "text", + "content": " (pink dot) which is closer to the center " + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "text", + "content": " (still a Spaniel, but less fine-grained). While the latent code of a tiger " + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "inline_equation", + "content": "z_{\\mathbb{D}}^{c_2}" + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "text", + "content": " (blue dot) locates far from the latent code of a Spaniel. Thus, the hyperbolic average code of tiger and Spaniel " + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "inline_equation", + "content": "\\bar{z}_{\\mathbb{D}}^{c_{12}}" + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "text", + "content": " (purple dot) is closer to the center " + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "text", + "content": " than " + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "inline_equation", + "content": "\\bar{z}_{\\mathbb{D}}^{c_1}" + }, + { + "bbox": [ + 46, + 284, + 289, + 406 + ], + "type": "text", + "content": " which is more abstract (a feline contains features from both tiger and Spaniel)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 420, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 287, + 479 + ], + "type": "text", + "content": "space defined in [6] since it is commonly used in gradient-based learning [44, 18, 45, 56, 55, 35]. The Poincaré disk model " + }, + { + "bbox": [ + 47, + 420, + 287, + 479 + ], + "type": "inline_equation", + "content": "\\left(\\mathbb{D}^n,g^{\\mathbb{D}}\\right)" + }, + { + "bbox": [ + 47, + 420, + 287, + 479 + ], + "type": "text", + "content": " is defined by the manifold " + }, + { + "bbox": [ + 47, + 420, + 287, + 479 + ], + "type": "inline_equation", + "content": "\\mathbb{D}^n = \\{x\\in \\mathbb{R}^n:\\| x\\| < 1\\}" + }, + { + "bbox": [ + 47, + 420, + 287, + 479 + ], + "type": "text", + "content": " equipped with the following Riemannian metric:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 479, + 287, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 479, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 140, + 479, + 287, + 493 + ], + "type": "interline_equation", + "content": "g _ {x} ^ {\\mathbb {D}} = \\lambda_ {x} ^ {2} g ^ {E}, \\tag {1}", + "image_path": "fc85940c51d9c3b3d848574f4319da6d5bc2f60f01c9a3571814b3587b6f7f47.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 498, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 287, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 498, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\lambda_{x} = \\frac{2}{1 - \\|x\\|^{2}}" + }, + { + "bbox": [ + 47, + 498, + 287, + 537 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 498, + 287, + 537 + ], + "type": "inline_equation", + "content": "g^{E}" + }, + { + "bbox": [ + 47, + 498, + 287, + 537 + ], + "type": "text", + "content": " is the Euclidean metric tensor " + }, + { + "bbox": [ + 47, + 498, + 287, + 537 + ], + "type": "inline_equation", + "content": "g^{E} = \\mathbf{I}^{n}" + }, + { + "bbox": [ + 47, + 498, + 287, + 537 + ], + "type": "text", + "content": ". The induced distance between two points " + }, + { + "bbox": [ + 47, + 498, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{x},\\mathbf{y}\\in \\mathbb{D}^n" + }, + { + "bbox": [ + 47, + 498, + 287, + 537 + ], + "type": "text", + "content": " can be defined by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 287, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 287, + 581 + ], + "type": "interline_equation", + "content": "d _ {\\mathbb {D}} (\\mathbf {x}, \\mathbf {y}) = \\operatorname {a r c c o s h} \\left(1 + 2 \\frac {\\| \\mathbf {x} - \\mathbf {y} \\| ^ {2}}{\\left(1 - \\| \\mathbf {x} \\| ^ {2}\\right) \\left(1 - \\| \\mathbf {y} \\| ^ {2}\\right)}\\right). \\tag {2}", + "image_path": "809beac9fa5681c76698ef9ff14d845b4bb78b726b4346e55e47e85d337ae681.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Recall that a geodesic is a locally minimized-length curve between two points. In the hyperboloid model, the geodesic can be defined as the curve created by intersecting the plane defined by two points and the origin with the hyperboloid [38]. Thus, the mean of two latent codes in hyperbolic space locates at the mid-point of the geodesic that is closer to the origin. This is the key desired feature of hyperbolic space, i.e., the mean between two leaf embeddings is not another leaf embedding, but the hierarchical parent of them [55]. This feature allows us to generate new images by moving the latent code from one leaf to another with the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "same parents. We can also change the semantic levels of attributes by determining how abstract their parent is." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 97, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 97, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 97, + 547, + 168 + ], + "type": "text", + "content": "This unique property is visualized in Fig. 3 on a 2-D Poincaré disk. The image embedding near the edge of the ball (with a large radius) represents a more fine-grained image while the embedding near the center (which has a smaller radius) represents an image with abstract features (an average face)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 170, + 547, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 170, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 305, + 170, + 547, + 229 + ], + "type": "text", + "content": "Although the hyperbolic space shares similar features with trees, it is continuous. In other words, there is no fixed number of hierarchy levels. Instead, there is a continuum from very fine-grained (near the edge of Poincaré disk) to very abstract (near the origin)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 239, + 432, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 432, + 250 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 432, + 250 + ], + "type": "text", + "content": "3.2. Network Architecture" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 258, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 258, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 545, + 330 + ], + "type": "text", + "content": "Although we aim to embed and edit real images in hyperbolic space, the whole network does not need to be implemented in a hyperbolic manner. Instead, we can take advantage of the number of existing GAN inversion models and optimization algorithms that have been fine-tuned for Euclidean space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 331, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 378 + ], + "type": "text", + "content": "To achieve image editing, we need to embed the image back into the latent space. In particular, we select pSp [49] as the backbone of HAE to encode images to the " + }, + { + "bbox": [ + 305, + 331, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 305, + 331, + 545, + 378 + ], + "type": "text", + "content": "-space of StyleGAN2 [34]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 392, + 390, + 545, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 390, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 392, + 390, + 545, + 403 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {i} = \\mathrm {p S p} \\left(x _ {i}\\right), \\tag {3}", + "image_path": "90dfde39b24deeb7537908226a9cdeb441610d68de412dc926aaee8cccbb3929.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 412, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 412, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 545, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 412, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i\\in \\mathbb{R}^{18\\times 512}" + }, + { + "bbox": [ + 305, + 412, + 545, + 437 + ], + "type": "text", + "content": " is the corresponding latent vector of " + }, + { + "bbox": [ + 305, + 412, + 545, + 437 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 412, + 545, + 437 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 305, + 412, + 545, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 305, + 412, + 545, + 437 + ], + "type": "text", + "content": " -space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "text", + "content": "To manipulate latent code in hyperbolic space, we need to define a bijective map from " + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c^n" + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "text", + "content": " to map Euclidean vectors to the hyperbolic space and vice versa. A manifold is a differentiable topological space that locally resembles the Euclidean space " + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "text", + "content": " [37, 38]. For " + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "inline_equation", + "content": "x\\in \\mathbb{D}^n" + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "text", + "content": ", one can define the tangent space " + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "inline_equation", + "content": "T_{x}\\mathbb{D}_{c}^{n}" + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c^n" + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "text", + "content": " as the first order linear approximation of " + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c^n" + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "text", + "content": " around " + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "text", + "content": ". Therefore, this bijective map can be performed by exponential and logarithmic maps. Specifically, the exponential map " + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "inline_equation", + "content": "\\exp_{\\mathbf{x}}^{c}:T_{\\mathbf{x}}\\mathbb{D}_{c}^{n}\\cong \\mathbb{R}^{n}\\to \\mathbb{D}_{c}^{n}" + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "text", + "content": ", maps from the tangent spaces into the manifold. While the logarithmic map " + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "inline_equation", + "content": "\\log_{\\mathbf{x}}^{c}:\\mathbb{D}_{c}^{n}\\to T_{\\mathbf{x}}\\mathbb{D}_{c}^{n}\\cong \\mathbb{R}^{n}" + }, + { + "bbox": [ + 305, + 438, + 546, + 582 + ], + "type": "text", + "content": " is the reverse map of the exponential map." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 582, + 545, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 690 + ], + "type": "text", + "content": "We use exponential and logarithmic maps at origin " + }, + { + "bbox": [ + 305, + 582, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 305, + 582, + 545, + 690 + ], + "type": "text", + "content": " for the transformation between the Euclidean and hyperbolic representations. After getting " + }, + { + "bbox": [ + 305, + 582, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 305, + 582, + 545, + 690 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 305, + 582, + 545, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 305, + 582, + 545, + 690 + ], + "type": "text", + "content": "-space, we first use a Multi-layer Perceptron (MLP) encoder to reduce the dimension of latent vectors in Euclidean space. Then we apply an exponential map to project the Euclidean latent code " + }, + { + "bbox": [ + 305, + 582, + 545, + 690 + ], + "type": "inline_equation", + "content": "z_{\\mathbb{R}i}" + }, + { + "bbox": [ + 305, + 582, + 545, + 690 + ], + "type": "text", + "content": " to hyperbolic space. After that, we use the hyperbolic feed-forward layer as [18] to obtain the final hierarchical representation " + }, + { + "bbox": [ + 305, + 582, + 545, + 690 + ], + "type": "inline_equation", + "content": "z_{\\mathbb{D}}" + }, + { + "bbox": [ + 305, + 582, + 545, + 690 + ], + "type": "text", + "content": " as shown in Fig. 2:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 362, + 700, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 362, + 700, + 545, + 714 + ], + "type": "interline_equation", + "content": "z _ {\\mathbb {D} i} = f ^ {\\otimes_ {c}} \\left(\\exp_ {\\mathbf {0}} ^ {c} \\left(\\mathrm {M L P} _ {E} (\\mathbf {w} _ {i})\\right)\\right), \\tag {4}", + "image_path": "a176a215b426a4d45d845e64acf27ba30e6cf33721a27a0f48c60f07df32604f.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22717" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 72, + 289, + 182 + ], + "blocks": [ + { + "bbox": [ + 100, + 72, + 289, + 182 + ], + "lines": [ + { + "bbox": [ + 100, + 72, + 289, + 182 + ], + "spans": [ + { + "bbox": [ + 100, + 72, + 289, + 182 + ], + "type": "image", + "image_path": "8dac1d1444f682bdbff58e0c7b8fae390dfb8bc5f46ce3b9eef0509faecf5432.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 188, + 463, + 200 + ], + "lines": [ + { + "bbox": [ + 129, + 188, + 463, + 200 + ], + "spans": [ + { + "bbox": [ + 129, + 188, + 463, + 200 + ], + "type": "text", + "content": "Figure 4: One-shot image generation from HAE on Animal Faces, Flowers, and VGGFaces." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 493, + 182 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 493, + 182 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 493, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 493, + 182 + ], + "type": "image", + "image_path": "7255d8b40b12993f122298bd7ea3eeb3ebea7af22453f64765212fb3e83278c5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 215, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 287, + 239 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 215, + 287, + 239 + ], + "type": "inline_equation", + "content": "f^{\\otimes_c}" + }, + { + "bbox": [ + 46, + 215, + 287, + 239 + ], + "type": "text", + "content": " is the Möbius translation of feed-forward layer " + }, + { + "bbox": [ + 46, + 215, + 287, + 239 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 215, + 287, + 239 + ], + "type": "text", + "content": " as the map from " + }, + { + "bbox": [ + 46, + 215, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c^n" + }, + { + "bbox": [ + 46, + 215, + 287, + 239 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 215, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c^m" + }, + { + "bbox": [ + 46, + 215, + 287, + 239 + ], + "type": "text", + "content": ", denoted as Möbius linear layer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 239, + 287, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 287, + 286 + ], + "type": "text", + "content": "Finally, the hyperbolic representation " + }, + { + "bbox": [ + 46, + 239, + 287, + 286 + ], + "type": "inline_equation", + "content": "z_{\\mathbb{D}}" + }, + { + "bbox": [ + 46, + 239, + 287, + 286 + ], + "type": "text", + "content": " needs to be projected back to the " + }, + { + "bbox": [ + 46, + 239, + 287, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 46, + 239, + 287, + 286 + ], + "type": "text", + "content": "-space of StyleGAN2. In practice, this is achieved by applying a logarithmic map followed by an MLP decoder:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 296, + 287, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 296, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 116, + 296, + 287, + 310 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {i} ^ {\\prime} = \\operatorname {M L P} _ {D} \\left(\\log_ {\\mathbf {0}} ^ {c} \\left(z _ {\\mathbb {D} i}\\right)\\right), \\tag {5}", + "image_path": "c9614012fee235f8a59c8f58253273c8c3701cef17901403158a56ed3a718880.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 318, + 287, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 287, + 343 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 47, + 318, + 287, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i^{\\prime}" + }, + { + "bbox": [ + 47, + 318, + 287, + 343 + ], + "type": "text", + "content": " will be fed into a pre-trained StyleGAN2's generator " + }, + { + "bbox": [ + 47, + 318, + 287, + 343 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 318, + 287, + 343 + ], + "type": "text", + "content": " to reconstruct the image " + }, + { + "bbox": [ + 47, + 318, + 287, + 343 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime}" + }, + { + "bbox": [ + 47, + 318, + 287, + 343 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 350, + 135, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 135, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 135, + 361 + ], + "type": "text", + "content": "3.3. Loss Function" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 369, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 416 + ], + "type": "text", + "content": "The loss function of HAE consists of two parts: the Hyperbolic loss ensures to get the hierarchical representation in the hyperbolic space and the reconstruction loss guarantees the quality of reconstruction images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 416, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 561 + ], + "type": "text", + "content": "Hyperbolic Loss. To learn the semantic hierarchical representation of real images in hyperbolic space, we minimize the distance between latent codes of images with similar categories and attributes while pushing away the latent codes from different categories. We choose the supervised approach to achieve this. In order to perform multi-class classification on the Poincaré disk defined in Sec. 3.1, one needs to generalize multinomial logistic regression (MLR) to the Poincaré disk defined in [18]. An extra linear layer needs to be trained for the classification and the softmax probability can be computed as: Given " + }, + { + "bbox": [ + 46, + 416, + 287, + 561 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 416, + 287, + 561 + ], + "type": "text", + "content": " classes and " + }, + { + "bbox": [ + 46, + 416, + 287, + 561 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\ldots, K\\}" + }, + { + "bbox": [ + 46, + 416, + 287, + 561 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 416, + 287, + 561 + ], + "type": "inline_equation", + "content": "p_k \\in \\mathbb{D}_c^n" + }, + { + "bbox": [ + 46, + 416, + 287, + 561 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 416, + 287, + 561 + ], + "type": "inline_equation", + "content": "a_k \\in T_{p_k} \\mathbb{D}_c^n \\setminus \\{\\mathbf{0}\\}" + }, + { + "bbox": [ + 46, + 416, + 287, + 561 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 569, + 287, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 287, + 665 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p (y = k \\mid x) \\propto \\exp \\left(\\frac {\\lambda_ {p _ {k}} ^ {c} \\| a _ {k} \\|}{\\sqrt {c}} \\sinh^ {- 1} \\right. \\\\ \\left(\\frac {2 \\sqrt {c} \\langle - p _ {k} \\oplus_ {c} x , a _ {k} \\rangle}{\\left(1 - c \\| - p _ {k} \\oplus_ {c} x \\| ^ {2}\\right) \\| a _ {k} \\|\\right)\\left. \\right), \\\\ \\forall x \\in \\mathbb {D} _ {c} ^ {n}. \\tag {6} \\\\ \\end{array}", + "image_path": "bb790b87a5a9799c101955b491e29dfe2e84c6974746f8ba0c8c0abce4c61790.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 666, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 286, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 666, + 286, + 689 + ], + "type": "inline_equation", + "content": "\\oplus_{c}" + }, + { + "bbox": [ + 46, + 666, + 286, + 689 + ], + "type": "text", + "content": " denotes the Möbius addition defined in [35] with fixed sectional curvature of the space, denoted by " + }, + { + "bbox": [ + 46, + 666, + 286, + 689 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 666, + 286, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "After getting the softmax result for each class, one can use negative log-likelihood loss (NLL Loss) to calculate the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 216, + 372, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 216, + 372, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 216, + 372, + 227 + ], + "type": "text", + "content": "hyperbolic loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 370, + 236, + 545, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 236, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 370, + 236, + 545, + 269 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {h y p e r}} = - \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\log \\left(p _ {n}\\right), \\tag {7}", + "image_path": "750a66538d0b57cec7c70452c9a5eb2c3962015da45648e89235378bdcc9a457.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 277, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 545, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 277, + 545, + 300 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 277, + 545, + 300 + ], + "type": "text", + "content": " is the batch size and " + }, + { + "bbox": [ + 305, + 277, + 545, + 300 + ], + "type": "inline_equation", + "content": "p_n" + }, + { + "bbox": [ + 305, + 277, + 545, + 300 + ], + "type": "text", + "content": " is the probability predicted by the model for the correct class." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 300, + 545, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 420 + ], + "type": "text", + "content": "As mentioned in Sec. 3.1, the distance between points grows exponentially with their radius in the Poincaré disk. In order to minimize Eq. (7), the latent codes of fine-grained images will be pushed to the edge of the ball to maximize the distances between different categories while the embedding of abstract images (images have common features from many categories) will be located near the center of the ball. Since hyperbolic space is continuous and differentiable, we are able to optimize Eq. (7) with stochastic gradient descent, which learns the hierarchy of the images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 420, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 420, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 545, + 456 + ], + "type": "text", + "content": "Reconstruction Loss. In order to guarantee the quality of the generated images, we first use the " + }, + { + "bbox": [ + 305, + 420, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 305, + 420, + 545, + 456 + ], + "type": "text", + "content": " loss and LPIPS loss used in pSp [49], given image " + }, + { + "bbox": [ + 305, + 420, + 545, + 456 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 305, + 420, + 545, + 456 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 367, + 466, + 545, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 466, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 367, + 466, + 545, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2} \\left(x _ {i}\\right) = \\left\\| x _ {i} - \\mathrm {H A E} \\left(x _ {i}\\right) \\right\\| _ {2}. \\tag {8}", + "image_path": "b309108a28764c884f9849b99e795295025ad9c5b72554164e17dcce3cec63d3.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 344, + 500, + 545, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 500, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 344, + 500, + 545, + 513 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L P I P S}} \\left(x _ {i}\\right) = \\| F \\left(x _ {i}\\right) - F (\\mathrm {H A E} \\left(x _ {i}\\right)) \\| _ {2}, \\tag {9}", + "image_path": "dc87579187d955cc4764c4c0ca65a0233dba3830c78bba0ab8d7832a87c5bcb9.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 518, + 515, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 518, + 515, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 518, + 515, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 518, + 515, + 529 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 305, + 518, + 515, + 529 + ], + "type": "text", + "content": " denotes the perceptual feature extractor." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "type": "text", + "content": "Since the pSp encoder and StyleGAN2 generator are pretrained, we only train the neural layers between the encoder and generator of HAE. To further guarantee the network to better project back to the " + }, + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "type": "text", + "content": "-space, the reconstructed " + }, + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i'" + }, + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "type": "text", + "content": " should be the same as the original " + }, + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i" + }, + { + "bbox": [ + 304, + 530, + 545, + 590 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 374, + 599, + 545, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 599, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 374, + 599, + 545, + 613 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c}} \\left(w _ {i}\\right) = \\left\\| \\mathbf {w} _ {\\mathrm {i}} - \\mathbf {w} _ {i} ^ {\\prime} \\right\\| _ {2}, \\tag {10}", + "image_path": "0b67ea74f081bfada6e075b57d273899597db21255c625f8cf6452960710aca6.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 622, + 510, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 510, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 510, + 634 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 622, + 510, + 634 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i^{\\prime}" + }, + { + "bbox": [ + 306, + 622, + 510, + 634 + ], + "type": "text", + "content": " can be calculated by Eq. (4) and Eq. (5)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 634, + 436, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 634, + 436, + 644 + ], + "spans": [ + { + "bbox": [ + 318, + 634, + 436, + 644 + ], + "type": "text", + "content": "The overall loss function is:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 324, + 655, + 545, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 655, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 324, + 655, + 545, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {2} \\left(x _ {i}\\right) + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {L P I P S}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {3} \\mathcal {L} _ {\\text {h y p e r}}, \\tag {11}", + "image_path": "f2f84026417fe773e7aff617ee842295ef29391c3538e31c5c7e7a4372ba2b57.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{1},\\lambda_{2}" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{3}" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " are trade-off adaptive parameters. This curated set of loss functions ensures the model learns the hierarchical representation and reconstructs images." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22718" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 70, + 496, + 210 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 496, + 210 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 496, + 210 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 496, + 210 + ], + "type": "image", + "image_path": "5fa26aab41441e0d1d0c821143c058787a2d03e1ad5b6b95a160abf7c17914ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 215, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 215, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 547, + 237 + ], + "type": "text", + "content": "Figure 5: Images generated by HAE by adding the same perturbation on the latent code of a given image with different hyperbolic radii on Animal Faces and Flowers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 252, + 287, + 344 + ], + "blocks": [ + { + "bbox": [ + 48, + 252, + 287, + 344 + ], + "lines": [ + { + "bbox": [ + 48, + 252, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 48, + 252, + 287, + 344 + ], + "type": "image", + "image_path": "0c6302289e3089c5e2b0504fa904a12bc7da1c482eef0dd11c4b89f9ff215842.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 348, + 287, + 371 + ], + "lines": [ + { + "bbox": [ + 47, + 348, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 287, + 371 + ], + "type": "text", + "content": "Figure 6: Interpolations in hyperbolic space along the edge of the Poincaré disk (with " + }, + { + "bbox": [ + 47, + 348, + 287, + 371 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{D}} = 6.2126" + }, + { + "bbox": [ + 47, + 348, + 287, + 371 + ], + "type": "text", + "content": ") on three datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 389, + 155, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 155, + 402 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 155, + 402 + ], + "type": "text", + "content": "3.4. Image Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 407, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 287, + 443 + ], + "type": "text", + "content": "To study the generating quality of the model, a straightforward way is to generate new images via interpolation between two designated images, or random perturbation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 444, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 287, + 493 + ], + "type": "text", + "content": "In hyperbolic space, the shortest path with the induced distance between two points is given by the geodesic defined in Eq. (2). The geodesic equation between two embeddings " + }, + { + "bbox": [ + 47, + 444, + 287, + 493 + ], + "type": "inline_equation", + "content": "z_{\\mathbb{D}\\mathbf{i}}" + }, + { + "bbox": [ + 47, + 444, + 287, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 444, + 287, + 493 + ], + "type": "inline_equation", + "content": "z_{\\mathbb{D}\\mathbf{j}}" + }, + { + "bbox": [ + 47, + 444, + 287, + 493 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 47, + 444, + 287, + 493 + ], + "type": "inline_equation", + "content": "\\gamma z_{\\mathbb{D}i}\\to z_{\\mathbb{D}j}(t)" + }, + { + "bbox": [ + 47, + 444, + 287, + 493 + ], + "type": "text", + "content": ", is given by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 510, + 287, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 287, + 533 + ], + "type": "interline_equation", + "content": "\\gamma_ {z _ {\\mathbb {D} i} \\rightarrow z _ {\\mathbb {D} j}} (t) = z _ {\\mathbb {D} i} \\oplus_ {c} t \\otimes_ {c} \\left(\\left(- z _ {\\mathbb {D} i}\\right) \\oplus_ {c} z _ {\\mathbb {D} j}\\right), t \\in [ 0, 1 ], \\tag {12}", + "image_path": "3575f67d2a8ec4395a783f68eac30a8d41e8f2610e0b7814099925493f378912.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 534, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 287, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 534, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\oplus_{c}" + }, + { + "bbox": [ + 47, + 534, + 287, + 568 + ], + "type": "text", + "content": " denotes the Möbius addition with aforementioned sectional curvature " + }, + { + "bbox": [ + 47, + 534, + 287, + 568 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 534, + 287, + 568 + ], + "type": "text", + "content": ", with details in supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 570, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 287, + 642 + ], + "type": "text", + "content": "We adopt the following method to achieve generating via perturbation: For a given image " + }, + { + "bbox": [ + 47, + 570, + 287, + 642 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 47, + 570, + 287, + 642 + ], + "type": "text", + "content": ", we first rescale its embedding " + }, + { + "bbox": [ + 47, + 570, + 287, + 642 + ], + "type": "inline_equation", + "content": "z_{\\mathbb{D}i}" + }, + { + "bbox": [ + 47, + 570, + 287, + 642 + ], + "type": "text", + "content": " to the desired radius " + }, + { + "bbox": [ + 47, + 570, + 287, + 642 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{D}}" + }, + { + "bbox": [ + 47, + 570, + 287, + 642 + ], + "type": "text", + "content": ". Then we sample a random vector from seen categories in " + }, + { + "bbox": [ + 47, + 570, + 287, + 642 + ], + "type": "inline_equation", + "content": "z_{\\mathbb{D}j}" + }, + { + "bbox": [ + 47, + 570, + 287, + 642 + ], + "type": "text", + "content": " with radius " + }, + { + "bbox": [ + 47, + 570, + 287, + 642 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{D}}" + }, + { + "bbox": [ + 47, + 570, + 287, + 642 + ], + "type": "text", + "content": " fixed and take the geodesic as the direction of perturbation to generate images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 651, + 123, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 123, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 123, + 664 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 671, + 180, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 180, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 180, + 684 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "In the training stage, we first train a StyleGAN2 [33] and pSp [49] with seen categories. Given a trained pSp, the" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 247, + 544, + 348 + ], + "blocks": [ + { + "bbox": [ + 309, + 247, + 544, + 348 + ], + "lines": [ + { + "bbox": [ + 309, + 247, + 544, + 348 + ], + "spans": [ + { + "bbox": [ + 309, + 247, + 544, + 348 + ], + "type": "image", + "image_path": "9ec7a49ed547804cdbf2da6ee8e40cbb514291eee3b8ed893497fff31ea49295.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 352, + 546, + 384 + ], + "lines": [ + { + "bbox": [ + 305, + 352, + 546, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 352, + 546, + 384 + ], + "type": "text", + "content": "Figure 7: Interpolations by moving the latent codes from the edge to the center of the Poincaré disk (from fine-grained to abstract) on three datasets." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 401, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 401, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 401, + 545, + 449 + ], + "type": "text", + "content": "MLP encoder " + }, + { + "bbox": [ + 305, + 401, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_E" + }, + { + "bbox": [ + 305, + 401, + 545, + 449 + ], + "type": "text", + "content": " is an 8-layer MLP with a Leaky-ReLU activation function. The dimension of the latent code in hyperbolic space is chosen to be 512. More details can be found in the supplementary." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 457, + 368, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 457, + 368, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 368, + 468 + ], + "type": "text", + "content": "4.2. Datasets" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 475, + 545, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 475, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 305, + 475, + 545, + 511 + ], + "type": "text", + "content": "We evaluate our method on Animal Faces [40], Flowers [46], and VGGFaces [48] following the settings described in [15]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 512, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 512, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 512, + 545, + 536 + ], + "type": "text", + "content": "Animal Faces. We randomly select 119 categories as seen for training and leave 30 as unseen categories for testing." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 536, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 536, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 305, + 536, + 545, + 559 + ], + "type": "text", + "content": "Flowers. The Flowers [46] dataset is split into 85 seen categories for training and 17 unseen categories for testing." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 560, + 545, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 560, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 305, + 560, + 545, + 585 + ], + "type": "text", + "content": "VGGFaces. For VGGFaces [48], we randomly select 1802 categories for training and 572 for testing." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 592, + 518, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 592, + 518, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 592, + 518, + 605 + ], + "type": "text", + "content": "4.3. Analysis of Hierarchical Feature Editing" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 611, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 611, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 611, + 545, + 647 + ], + "type": "text", + "content": "We analyze the properties of the learned hierarchical representations and how the levels of attributes relate to their locations of latent codes in hyperbolic space." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 647, + 545, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 647, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 305, + 647, + 545, + 695 + ], + "type": "text", + "content": "As we mentioned in Sec. 3.1, there is a continuum from fine-grained attributes to abstract attributes, corresponding to the points from the peripheral to the center of the ball. We define the hyperbolic radius " + }, + { + "bbox": [ + 305, + 647, + 545, + 695 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{D}}^2" + }, + { + "bbox": [ + 305, + 647, + 545, + 695 + ], + "type": "text", + "content": " as the hyperbolic dis" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 534, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 534, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 534, + 712 + ], + "type": "text", + "content": "2The radius of the Poincaré disk in our experiment is about 6.2126" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22719" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 542, + 224 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 542, + 224 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 542, + 224 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 542, + 224 + ], + "type": "table", + "html": "
MethodSettingsFlowersAnimal FacesVGG Faces*
FID(↓)LPIPS(↑)FID(↓)LPIPS(↑)FID(↓)LPIPS(↑)
DAWSON [39]3-shot188.960.0583208.680.0642137.820.0769
MatchingGAN [28]3-shot143.350.1627148.520.1514118.620.1695
F2GAN [30]3-shot120.480.2172117.740.1831109.160.2125
LoFGAN [23]3-shot79.330.3862112.810.496420.310.2869
DeltaGAN [27]1-shot109.780.391289.810.441880.120.3146
Disco-FUNIT [29]1-shot90.120.443671.440.4511--
AGE [15]1-shot45.960.430528.040.557534.860.3294
SAGE [14]1-shot43.520.439227.430.544834.970.3232
HAE (Ours)1-shot50.100.473926.330.563635.930.5919
", + "image_path": "0c9858734b7907de3c669b7523ce7b7fafa9770080660a70faed799bec9d9052.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 275, + 544, + 459 + ], + "blocks": [ + { + "bbox": [ + 46, + 228, + 547, + 274 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 274 + ], + "type": "text", + "content": "Table 1: FID(↓) and LPIPS(↑) of images generated by different methods for unseen categories on three datasets. Bold indicates the best results and underline indicates the second best results. VGGFaces is marked with * because different methods report different numbers of unseen categories on this dataset(e.g. 552 in LoFGAN, 96 in DeltaGAN, 497 in L2GAN, and 572 in AGE and SAGE). Note that: disco-FUNIT [29] does not provide pre-trained models on VGG Faces [48] dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 275, + 544, + 459 + ], + "lines": [ + { + "bbox": [ + 52, + 275, + 544, + 459 + ], + "spans": [ + { + "bbox": [ + 52, + 275, + 544, + 459 + ], + "type": "table", + "html": "
InputMatchingGANLoFGANAGEHAE(ours)
", + "image_path": "942c39ebdac414ba3d3e6cae4b84fd4bf91f8b62056ed8a3d19fa21782be31d4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 464, + 546, + 488 + ], + "lines": [ + { + "bbox": [ + 46, + 464, + 546, + 488 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 546, + 488 + ], + "type": "text", + "content": "Figure 8: Comparison between images generated by MatchingGAN, LoFGAN, AGE, and HAE on Flowers, Animal Faces, and VGGFaces. Zoom in to see the details. Note that: SAGE [14] has not released code and pre-trained models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 502, + 288, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 288, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 288, + 550 + ], + "type": "text", + "content": "tance of the given latent code to the center of the Poincaré disk. To study the influence of the radius of embeddings in hyperbolic space, we run several experiments with different choices of " + }, + { + "bbox": [ + 46, + 502, + 288, + 550 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{D}}" + }, + { + "bbox": [ + 46, + 502, + 288, + 550 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": "Hyperbolic Perturbation and Interpolation. As mentioned in Sec. 3.4, we demonstrate the results of perturbation and interpolation. In addition to the choice of perturbation, we can set the intensity of the perturbation by controlling both the step distance and radius as shown in Fig. 5. The results show that level of semantic attributes is highly related to " + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{D}}" + }, + { + "bbox": [ + 46, + 558, + 289, + 715 + ], + "type": "text", + "content": ". With the radius becoming smaller, the attributes become more abstract. We further visualize this property of hyperbolic space by moving the latent codes of the given image from the edge of the Poincaré disk to the center. As Fig. 7 shows, the images change from very fine-grained to very abstract (the average of all images). The results in Fig. 6 show that we can achieve smooth interpola" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 539 + ], + "type": "text", + "content": "tion in hyperbolic space without any distortion. The results demonstrate that with HAE, we can freely control the editing geodesically and hierarchically." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 550, + 459, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 550, + 459, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 550, + 459, + 563 + ], + "type": "text", + "content": "4.4. Few-shot Image Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "As Fig. 5 shows, the image categories will be changed when " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{D}}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " is smaller than about 4, and the category-irrelevant attributes of images will be changed when " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{D}}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " is larger than about 5. The embeddings of Animal Faces are visualized in 2-D Poincaré disk using UMAP [43] shown in Fig. 9. As Fig. 6 shows, the posture and the angle of the images will be changed at the early stage of interpolation without changing the category. Thus, the images can be generated by moving the latent code of a given image to some randomly selected semantic direction within the cluster of the category. In practice, we select " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{D}} = 6.21" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " and step size of perturbation as 8 to achieve few-shot image generation as" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22720" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 77, + 265, + 274 + ], + "blocks": [ + { + "bbox": [ + 71, + 77, + 265, + 274 + ], + "lines": [ + { + "bbox": [ + 71, + 77, + 265, + 274 + ], + "spans": [ + { + "bbox": [ + 71, + 77, + 265, + 274 + ], + "type": "image", + "image_path": "c88386d26e9c5591abf100e3ba414141752462191b577ad1bc9ad47b6ea573ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 279, + 289, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 279, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 289, + 323 + ], + "type": "text", + "content": "Figure 9: UMAP visualization of hyperbolic 2-D embeddings of Animal Faces dataset. We observe that similar categories are clustered and positioned near the boundary, while ambiguous samples are located near the center. Zoom in to see the details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 330, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 287, + 390 + ], + "type": "text", + "content": "Fig. 4 shows the diverse images generated by adding random perturbations from seen categories. We conduct three experiments to show that HAE can achieve promising few-shot image generation. More examples of generated images are available in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 391, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 287, + 533 + ], + "type": "text", + "content": "Quantitative Comparison with State-of-the-art. We calculate the FID [24] and LPIPS [59] to evaluate the fidelity and diversity of the generated images following one-shot settings in [15, 14]. The comparison results are shown in Tab. 1. Our method achieves the best scores on most of the FID and LPIPS metrics compared with state-of-the-art few-shot image generation methods, which indicates that our method not only improves the model from the semantic aspect but also achieves state-of-the-art performance on the traditional evaluation metrics. Specifically, the LPIPS score of HAE beats SOTA model SAGE on all three datasets since HAE can generate more diverse images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": "Qualitative Evaluation. We qualitatively compare our method with MatchingGAN [28], LoFGAN [23], DeltaGAN [27] and AGE [15]. As shown in Fig. 8, HAE can generate images with diversity and fine-grained details. More importantly, the newly generated images have more semantic diversity than others. For instance, the shadow and skin color of the generated faces change with the light condition, and this effect looks more natural. We further conduct a user study by randomly selecting 60 (20 from each dataset) images with generated variants using AGE and HAE. 50 users from different backgrounds are asked to rate the results only based on diversity and quality external information. This is achieved by randomly shuffling the order of images pairwisely and inside any pair. HAE won by a ratio of " + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "58.1\\%" + }, + { + "bbox": [ + 46, + 534, + 288, + 714 + ], + "type": "text", + "content": " (1743/3000) over AGE (more details in supplementary)." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 533, + 192 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 533, + 192 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 533, + 192 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 533, + 192 + ], + "type": "image", + "image_path": "dff337300422ef151eb250686ca8c16928f26d0703f014aab3eecbb8e295ba58.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 197, + 545, + 219 + ], + "lines": [ + { + "bbox": [ + 305, + 197, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 305, + 197, + 545, + 219 + ], + "type": "text", + "content": "Figure 10: Manipulate images from different categories with the same perturbation (Target 1&2)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 234, + 545, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 234, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 234, + 545, + 306 + ], + "type": "text", + "content": "Transferability. If we move latent codes at category-irrelevant levels, the target perturbation is transferable across all categories. We edit the images from three categories with the same editing direction, the output images are shown in Fig. 10. It demonstrates that HAE achieves a highly controllable and interpretable editing process." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 316, + 544, + 381 + ], + "blocks": [ + { + "bbox": [ + 309, + 316, + 544, + 381 + ], + "lines": [ + { + "bbox": [ + 309, + 316, + 544, + 381 + ], + "spans": [ + { + "bbox": [ + 309, + 316, + 544, + 381 + ], + "type": "table", + "html": "
MethodFlowersAnimal FacesVGG Faces
FIDLPIPSFIDLPIPSFIDLPIPS
SAGE [14]43.520.439227.430.544834.970.3232
HAE(Euc)54.620.429325.270.512938.460.5908
HAE(Hyp)50.100.473926.330.563635.930.5919
", + "image_path": "5f9f560479b0d9136517568e4b67b3c443063013777ce7ca054ae3a3676eeca1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 385, + 545, + 429 + ], + "lines": [ + { + "bbox": [ + 305, + 385, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 545, + 429 + ], + "type": "text", + "content": "Table 2: " + }, + { + "bbox": [ + 305, + 385, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\mathrm{{FID}}\\left( \\downarrow \\right)" + }, + { + "bbox": [ + 305, + 385, + 545, + 429 + ], + "type": "text", + "content": " and LPIPS(↑) of images generated by HAE in different geometries for unseen categories on three datasets. Bold indicates the best results and underline indicates the second best results." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 444, + 400, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 400, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 400, + 456 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 581 + ], + "type": "text", + "content": "HAE in Euclidean. We re-trained HAE models in Euclidean space with the NLL loss to validate the performance gain in Tab. 1 is due to the hierarchical hyperbolic representation rather than the disentanglement caused by Eq. (7). The quantitative comparison is shown in Tab. 2. It shows that the hyperbolic space boosts the performance, especially for the LPIPS score, since the latent code is more disentangled in hyperbolic space [19]. This finding is also supported by the UMAP visualization in Fig. 9. More details can be found in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Hyperbolic Radius versus Truncation. StyleGAN [33] uses truncation trick [42, 4, 33, 34] in " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "-space to achieve the balance between the image quality and diversity. The experiments in [33, 34] also show that the truncation level in " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^+" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "-space control the level of abstraction of the generated images. We conduct the experiments in Sec. 3.4 using truncation to validate the gains of hyperbolic space. The results are illustrated in Fig. 11 and Fig. 12. As Fig. 11 shows, the category of the image changes along with the posture of the dog as the truncation gets smaller, while the category-relevant attributes do not change when the hyperbolic radius" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22721" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 286, + 149 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 149 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 149 + ], + "type": "image", + "image_path": "3aed1d95f93aeb57be8519dddde61107148227012a529009129067feb1f24650.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 152, + 287, + 195 + ], + "lines": [ + { + "bbox": [ + 46, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 287, + 195 + ], + "type": "text", + "content": "Figure 11: Top: Interpolations by moving the latent codes from the edge to the center in hyperbolic space. Bottom: Interpolation with different truncation in Euclidean space. Zoom in to see the details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 201, + 286, + 304 + ], + "blocks": [ + { + "bbox": [ + 48, + 201, + 286, + 304 + ], + "lines": [ + { + "bbox": [ + 48, + 201, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 286, + 304 + ], + "type": "image", + "image_path": "9b86bda6924f822f791aa9d247f422d569275c81d1d371940aaf6f6fc7a33c96.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 305, + 286, + 349 + ], + "lines": [ + { + "bbox": [ + 46, + 305, + 286, + 349 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 286, + 349 + ], + "type": "text", + "content": "Figure 12: Images generated by HAE(Hyp) and HAE(Euc) by adding the same perturbation on the latent code of a given image with different settings of hyperbolic radius and truncation. Zoom in to see the details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 365, + 286, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 286, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 286, + 424 + ], + "type": "inline_equation", + "content": "(r_{\\mathbb{D}})" + }, + { + "bbox": [ + 46, + 365, + 286, + 424 + ], + "type": "text", + "content": " is large. This can also be proved in Fig. 12. The category remains the same after adding perturbation when " + }, + { + "bbox": [ + 46, + 365, + 286, + 424 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{D}}" + }, + { + "bbox": [ + 46, + 365, + 286, + 424 + ], + "type": "text", + "content": " is large, while the truncation can not control semantic-level editing. This shows that Euclidean space can only capture scale-based hierarchy rather than the semantic hierarchy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 425, + 286, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 286, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 286, + 627 + ], + "type": "text", + "content": "Downstream Task. We conduct data augmentation via HAE for image classification on Animal Faces [48]. We randomly select 30, 35, and 35 images for each category as train, val, and test, respectively. Following [23], a ResNet-18 backbone is initialized from the seen categories. Then the model is fine-tuned on the unseen categories referred to as the baseline. 60 images are generated for each unseen category as data augmentation. The result is presented in Tab. 3. The diversity and quality of generated images are primarily controlled by the hyperbolic radii " + }, + { + "bbox": [ + 46, + 425, + 286, + 627 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{D}}" + }, + { + "bbox": [ + 46, + 425, + 286, + 627 + ], + "type": "text", + "content": ". As the radius becomes smaller, HAE generates images of higher diversity, but categories (referring to high-level attributes) also gradually change to others. " + }, + { + "bbox": [ + 46, + 425, + 286, + 627 + ], + "type": "inline_equation", + "content": "r_{\\mathbb{D}} = 6" + }, + { + "bbox": [ + 46, + 425, + 286, + 627 + ], + "type": "text", + "content": " achieves the best performance on the classification experiment. However, the performance drops when the radius is smaller than 4.5. This is because the semantic attributes change too much and thus mislead the classifier." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 635, + 208, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 208, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 208, + 647 + ], + "type": "text", + "content": "4.6. Limitations and Future Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 688 + ], + "type": "text", + "content": "Although HAE achieves reliable hierarchical attribute editing in hyperbolic space for few-shot image generation, there are several limitations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 712 + ], + "type": "text", + "content": "First, the boundary of category changing is hard to be quantified since the hierarchical levels are continuous in" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 545, + 208 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 208 + ], + "type": "table", + "html": "
Hyperbolic RadiusAccuracyFID(↓)LPIPS(↑)
baseline58.67--
6.060.1046.890.4520
5.559.5248.680.4651
5.059.0552.080.4823
4.559.1460.870.5174
4.058.5765.830.5386
3.556.8668.440.6034
3.054.1469.400.6316
", + "image_path": "1ba5faaa2f5eb0bfd2f9d8c6da29214feb58a9db9ba01ac1828d8084cf15e10f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 212, + 544, + 232 + ], + "lines": [ + { + "bbox": [ + 306, + 212, + 544, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 544, + 232 + ], + "type": "text", + "content": "Table 3: Ablation of same perturbation on different radii on Animal Faces." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 256, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 327 + ], + "type": "text", + "content": "the hyperbolic space. Users need to find a \"safe\" boundary by trying different radii and step sizes of perturbation before generating new images. However, from another perspective, this continuity of hierarchy provides flexibility for users to set different boundaries for different downstream tasks as they need." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "text", + "content": "Second, the performance of HAE is limited by the pretrained styleGAN and the inversion method. If the input image can not be well embedded, the editing will also fail. This problem can be solved by changing more powerful backbones, e.g., ViT [16], in future work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 394, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 545, + 478 + ], + "type": "text", + "content": "Finally, we use supervised learning to get the hierarchical embedding in hyperbolic space. However, the number of images in existing datasets with labels for generation tasks is relatively small, which makes the embeddings in the hyperbolic space not evenly distributed. The solution to this problem is simple, use unsupervised learning with large-scale high-quality datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 498, + 378, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 378, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 378, + 510 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 522, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 653 + ], + "type": "text", + "content": "In this work, we propose a simple yet effective method HAE to edit hierarchical attributes in hyperbolic space. After learning the semantic hierarchy from images, our model is able to edit continuous semantic hierarchical features of images for flexible few-shot image generation in the hyperbolic space. Experiments demonstrate that HAE is capable of achieving not only stable few-shot image generation with state-of-the-art quality and diversity but a controllable and interpretable editing process. Future work includes the combination of HAE and large pretrained models and applications to more downstream tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by the National Key R&D Program of China under Grant 2018AAA0102000, in part by National Natural Science Foundation of China: 62022083 and 62236008." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22722" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Antreas Antoniou, Amos Storkey, and Harrison Edwards. Data augmentation generative adversarial networks. arXiv preprint arXiv:1711.04340, 2017. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "text", + "content": "[2] Sergey Bartunov and Dmitry P. Vetrov. Few-shot generative modelling with generative matching networks. In AISTATS, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "type": "text", + "content": "[3] Silvère Bonnabel. Stochastic gradient descent on riemannian manifolds. IEEE Transactions on Automatic Control, 58(9):2217-2229, 2013. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 223 + ], + "type": "text", + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. In ICLR, 2019. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 224, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 246 + ], + "type": "text", + "content": "[5] Gary Bécigneul and Octavian-Eugen Ganea. Riemannian adaptive optimization methods. In ICLR, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 247, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 278 + ], + "type": "text", + "content": "[6] James W Cannon, William J Floyd, Richard Kenyon, and Walter R Parry. Hyperbolic geometry. Flavors of geometry, 31:59-115, 1997. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 280, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 313 + ], + "type": "text", + "content": "[7] Ines Chami, Rex Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. In NeurIPS, page 4868-4879, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 315, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 287, + 346 + ], + "type": "text", + "content": "[8] Anton Cherepkov, Andrey Voynov, and Artem Babenko. Navigating the gan parameter space for semantic image editing. In CVPR, pages 3670-3679, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 347, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 390 + ], + "type": "text", + "content": "[9] Jaewoong Choi, Junho Lee, Changyeon Yoon, Jung Ho Park, Geonho Hwang, and Myungjoo Kang. Do not escape from the manifold: Discovering the local coordinates on the latent space of gans. In ICLR, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 392, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 392, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 287, + 413 + ], + "type": "text", + "content": "[10] Louis Clouatre and Marc Demers. Figr: Few-shot image generation with reptile. arXiv:1901.02199, 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 415, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 446 + ], + "type": "text", + "content": "[11] Jiali Cui, Ying Nian Wu, and Tian Han. Learning joint latent space ebm prior model for multi-layer generator. In CVPR, pages 3603-3612, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 448, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 448, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 287, + 479 + ], + "type": "text", + "content": "[12] Emily Denton, Ben Hutchinson, Margaret Mitchell, and Timnit Gebru. Detecting bias with generative counterfactual face attribute augmentation. CoRR, abs/1906.06439, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 481, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 513 + ], + "type": "text", + "content": "[13] Bhuwan Dhingra, Chris Shallue, Mohammad Norouzi, Andrew Dai, and George Dahl. Embedding text in hyperbolic spaces. arXiv preprint arXiv:1806.04313, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "type": "text", + "content": "[14] Guanqi Ding, Xinzhe Han, Shuhui Wang, Xin Jin, Dandan Tu, and Qingming Huang. Stable attribute group editing for reliable few-shot image generation. arXiv preprint arXiv:2302.00179, 2023. 2, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 559, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 601 + ], + "type": "text", + "content": "[15] Guanqi Ding, Xinzhe Han, Shuhui Wang, Shuzhe Wu, Xin Jin, Dandan Tu, and Qingming Huang. Attribute group editing for reliable few-shot image generation. In CVPR, pages 11184-11193, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 603, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 667 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 9" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[17] Kun Fu, Tengfei Zhang, Yue Zhang, Menglong Yan, Zhonghan Chang, Zhengyuan Zhang, and Xian Sun. Meta-ssd: Towards fast adaptation for few-shot object detection with meta-learning. IEEE Access, 7:77597-77606, 2019. 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[18] Octavian-Eugen Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. In NeurIPS, pages 5345-5355, 2018. 3, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "type": "text", + "content": "[19] Songwei Ge, Shlok Mishra, Simon Kornblith, Chun-Liang Li, and David Jacobs. Hyperbolic contrastive learning for visual representations beyond objects. In CVPR, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "text", + "content": "[20] Lore Goetschalckx, Alex Andonian, Aude Oliva, and Phillip Isola. Ganalyze: Toward visual definitions of cognitive image properties. In CVPR, page 5744-5753, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "type": "text", + "content": "[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 217, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 237 + ], + "type": "text", + "content": "[22] Michael Gromov. Hyperbolic groups. In Essays in group theory, 1987. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 239, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 270 + ], + "type": "text", + "content": "[23] Zheng Gu, Wenbin Li, Jing Huo, Lei Wang, and Yang Gao. Lofgan: Fusing local representations for fewshot image generation. In ICCV, 2021. 1, 2, 7, 8, 9" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 272, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 315 + ], + "type": "text", + "content": "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 316, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 337 + ], + "type": "text", + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In ICLR, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "text", + "content": "[26] Yan Hong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Deltagan: Towards diverse few-shot image generation with sample-specific delta. In CVPR, 2020. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 372, + 545, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 413 + ], + "type": "text", + "content": "[27] Yan Hong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Deltagan: Towards diverse few-shot image generation with sample-specific delta. In ECCV, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 458 + ], + "type": "text", + "content": "[28] Yan Hong, Li Niu, Jianfu Zhang, and Liqing Zhang. Matchinggan: Matching-based few-shot image generation. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2020. 1, 2, 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 514 + ], + "type": "text", + "content": "[29] Yan Hong, Li Niu, Jianfu Zhang, and Liqing Zhang. Few-shot image generation using discrete content representation. In Proceedings of the 30th ACM International Conference on Multimedia, MM '22, page 2796-2804, New York, NY, USA, 2022. Association for Computing Machinery. 1, 2, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 515, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 545, + 568 + ], + "type": "text", + "content": "[30] Yan Hong, Li Niu, Jianfu Zhang, Weijie Zhao, Chen Fu, and Liqing Zhang. F2gan: Fusing-and-filling gan for few-shot image generation. In Proceedings of the 28th ACM International Conference on Multimedia, MM '20, page 2535-2543. Association for Computing Machinery, 2020. 1, 2, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 570, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 601 + ], + "type": "text", + "content": "[31] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 603, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 633 + ], + "type": "text", + "content": "[32] Ali Jahanian, Lucy Chai, and Phillip Isola. \"On the\" steerability\" of generative adversarial networks. In ICLR, 2020. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "type": "text", + "content": "[33] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, pages 4217-4228, 2019. 6, 8" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[34] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In CVPR, pages 8107-8116, 2020. 1, 4, 8" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22723" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[35] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In CVPR, pages 6417-6427, 2020. 2, 3, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 150 + ], + "type": "text", + "content": "[36] Diego Lazcano, Nicolas Fredes Franco, and Werner Creixell. Hgan: Hyperbolic generative adversarial network. IEEE Access, 9:96309-96320, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 175 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 175 + ], + "type": "text", + "content": "[37] John M Lee. Riemannian manifolds: an introduction to curvature. Springer Science & Business Media, 176, 2006. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 177, + 287, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 177, + 287, + 197 + ], + "spans": [ + { + "bbox": [ + 48, + 177, + 287, + 197 + ], + "type": "text", + "content": "[38] John M Lee. Introduction to Smooth Manifolds. Springer, 2013. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 200, + 287, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 200, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 200, + 287, + 232 + ], + "type": "text", + "content": "[39] Weixin Liang, Zixuan Liu, and Can Liu. Dawson: A domain adaptive few shot generation framework. arXiv preprint arXiv:2001.00576, 2020. 1, 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 235, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 235, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 287, + 267 + ], + "type": "text", + "content": "[40] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsueprivised image-to-image translation. In ICCV, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 269, + 287, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 269, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 287, + 301 + ], + "type": "text", + "content": "[41] Yu-Ding Lu, Hsin-Ying Lee, Hung-Yu Tseng, and Ming-Hsuan Yang. Unsupervised discovery of disentangled manifolds in gans. arXiv preprint arXiv:2011.11842, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 304, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 335 + ], + "type": "text", + "content": "[42] Marco Marchesi. Megapixel size image creation using generative adversarial networks. CoRR, abs/1706.00082, 2017. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 337, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 380 + ], + "type": "text", + "content": "[43] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Grossberger. Umap: Uniform manifold approximation and projection. The Journal of Open Source Software, 3(29):861, 2018. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 383, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 414 + ], + "type": "text", + "content": "[44] Maximillian Nickel and Douwe Kiela. Generative visual manipulation on the natural image manifold. In ECCV, 2017. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 417, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 417, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 417, + 287, + 449 + ], + "type": "text", + "content": "[45] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In ICML, 2018. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 453, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 495 + ], + "type": "text", + "content": "[46] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 498, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 540 + ], + "type": "text", + "content": "[47] Jiwoong Park, Junho Cho, Hyung Jin Chang, and Jin Young Choi. Unsupervised hyperbolic representation learning via message passing auto-encoders. In CVPR, pages 5512–5522, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 543, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 574 + ], + "type": "text", + "content": "[48] Omkar M. Parkhi, Andrea Vedaldi, and Andrew Zisserman. Deep face recognition. In British Machine Vision Conference, 2015. 6, 7, 9" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "text", + "content": "[49] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In CVPR, pages 2287–2296, 2021. 4, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 623, + 287, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 654 + ], + "type": "text", + "content": "[50] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, pages 9240-9249, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 657, + 287, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 688 + ], + "type": "text", + "content": "[51] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, pages 1532-1540, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 712 + ], + "type": "text", + "content": "[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 354 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[53] Nurit Spingarn-Eliezer, Ron Banner, and Tomer Michaeli. Gan \"steerability\" without optimization. In ICLR, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[54] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H.S. Torr, and Timothy M. Hospedales. Learning to compare: Relation network for few-shot learning. In CVPR, pages 1199-1208, 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[55] Dídac Surís, Ruoshi Liu, and Carl Vondrick. Learning the predictability of the future. In CVPR, pages 12602-12612, 2021. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 175, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 206 + ], + "type": "text", + "content": "[56] Alexandru Tifrea, Gary Bécigneul, and OctavianEugen Ganea. Poincaré glove: Hyperbolic word embeddings. In ICLR, 2019. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "text", + "content": "[57] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NeurIPS, 2016. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "text", + "content": "[58] Andrey Voynov and Artem Babenko. Unsupervised discovery of interpretable directions in the gan latent space. In ICML, pages 9786-9796, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 319 + ], + "type": "text", + "content": "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 322, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 354 + ], + "type": "text", + "content": "[60] Jingyuan Zhu, Huimin Ma, Jiansheng Chen, and Jian Yuan. Few-shot image generation with diffusion models. arXiv preprint arXiv:2211.03264, 2022. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22724" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Making and Breaking of Camouflage/b706bafb-dae1-4647-ad2f-702e05fcaa82_content_list.json b/2023/The Making and Breaking of Camouflage/b706bafb-dae1-4647-ad2f-702e05fcaa82_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..585d1b627deed93e066ea12bf60ac17012544162 --- /dev/null +++ b/2023/The Making and Breaking of Camouflage/b706bafb-dae1-4647-ad2f-702e05fcaa82_content_list.json @@ -0,0 +1,2131 @@ +[ + { + "type": "text", + "text": "The Making and Breaking of Camouflage", + "text_level": 1, + "bbox": [ + 274, + 130, + 696, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hala Lamdouar1", + "bbox": [ + 251, + 179, + 383, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weidi Xie $^{1,2}$", + "bbox": [ + 421, + 179, + 522, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrew Zisserman", + "bbox": [ + 558, + 180, + 715, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Geometry Group, University of Oxford", + "bbox": [ + 125, + 198, + 496, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ CMIC, Shanghai Jiao Tong University", + "bbox": [ + 532, + 198, + 844, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lamdouar,weidi,az}@robots.ox.ac.uk", + "bbox": [ + 331, + 219, + 640, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Not all camouflages are equally effective, as even a partially visible contour or a slight color difference can make the animal stand out and break its camouflage. In this paper, we address the question of what makes a camouflage successful, by proposing three scores for automatically assessing its effectiveness. In particular, we show that camouflage can be measured by the similarity between background and foreground features and boundary visibility. We use these camouflage scores to assess and compare all available camouflage datasets. We also incorporate the proposed camouflage score into a generative model as an auxiliary loss and show that effective camouflage images or videos can be synthesised in a scalable manner. The generated synthetic dataset is used to train a transformer-based model for segmenting camouflaged animals in videos. Experimentally, we demonstrate state-of-the-art camouflage breaking performance on the public MoCA-Mask benchmark.", + "bbox": [ + 75, + 301, + 473, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 587, + 209, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camouflage has long been a subject of interest and fascination for the scientific community, especially evolutionary biologists who consider it as an excellent example of species adaptation. In order to confuse predators or to hide from prey, and increase their chances of survival in their natural habitat, animals have developed numerous camouflage mechanisms, e.g., disruptive coloration and background matching. Some species have even evolved to develop an adaptive camouflage, e.g., an arctic fox loses its white fur to better match the brown grey of the new season's landscape. Perhaps the most dramatic camouflage adaptation is the cuttlefish; it changes its patterns dynamically and rapidly as it moves from one spot to the other, constantly adapting and improving its camouflage.", + "bbox": [ + 75, + 612, + 468, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This search for optimal camouflage has inspired numerous works in the computer vision community, such as [12, 27], that tackled camouflage as a problem of optimal texture synthesis, making 3D objects non-detectable in a given scene. Others have addressed camouflage as a highly", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/acdb0323b3a207303092250111f36da1588f1f7b145b21d66b78260c3a8276ef.jpg", + "image_caption": [ + "a" + ], + "image_footnote": [], + "bbox": [ + 501, + 267, + 629, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/758a7bb71018f45b0f7b16043e3b56ce11b372d43202f6f0e7276681a2c98c14.jpg", + "image_caption": [ + "b" + ], + "image_footnote": [], + "bbox": [ + 632, + 267, + 759, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6dfef7ca4e28c02701f98c863384965930bab9386775378b22ed1ec4136de47.jpg", + "image_caption": [ + "C" + ], + "image_footnote": [], + "bbox": [ + 763, + 267, + 890, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "score function", + "bbox": [ + 633, + 362, + 751, + 380 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "camouflageranking: $c > a > b$", + "bbox": [ + 578, + 391, + 792, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1: The three images depict the same animal, an Arctic fox, as it adapts its appearance to better blend with the changing landscape of the new season. While images $a$ and $c$ exhibit better background similarity than image $b$ , the fox boundary is more visible in image $a$ than image $c$ . We assess the effectiveness of camouflage by measuring the degree of ambiguity it creates with respect to its background.", + "bbox": [ + 496, + 415, + 893, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "challenging object segmentation task in [9, 15, 20, 23]. Efforts have been made in collecting large scale camouflage datasets [9, 20] with costly annotation. In fact, camouflaged animals often exhibit complex shapes and thin structures that add to the boundary ambiguity and make the annotation highly time-consuming. Fan et al. report up to $60\\mathrm{min}$ per image to provide accurate pixel-level annotations for their dataset COD10K [9]. While another line of research turned to camouflage breaking in sequences by taking advantage of motion cues [1, 5, 18, 19, 40], the camouflage data scarcity is even more extreme for videos. Recently, Sim2Real [18, 39] training has shown to be very effective for motion segmentation. By training on the optical flows of synthetic videos, these models can generalise to real videos without suffering from the domain gaps.", + "bbox": [ + 496, + 535, + 893, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we start by asking the question: \"what are the properties of a successful camouflage?\" To answer this question, we investigate three scoring functions for quantifying the effectiveness of camouflage, namely, reconstruction fidelity score $(S_{R_f})$ , boundary score $(S_b)$ and intra-image Fréchet score $(d_{\\mathcal{F}})$ . These scores are later adopted for two critical roles, (i) they are used to assess the relevance of existing camouflage datasets and act as quality-indicator in collecting new camouflage data; (ii)", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "832", + "bbox": [ + 485, + 924, + 511, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "they can be used as a proxy loss for image and video inpainting/generation, where we establish a synthetic camouflage data generation pipeline with a Generative Adversarial Network (GAN), that can simultaneously generate high-quality camouflage examples and masks of the camouflaged animals. We further train a Transformer-based architecture on the generated camouflage video sequences, and demonstrate a performance boost over training on only the (small scale) real data.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we make the following contributions: First, we introduce three scoring functions to measure the effectiveness of a given camouflage in still images and videos. We use these camouflage scores to rank the images in existing datasets in terms of camouflage success, and also show that the rankings are largely in accordance with human-produced rankings. Second, we incorporate the camouflage score into a generative model, establishing a scalable pipeline for generating high-quality camouflage images or videos, along with the pixelwise segmentation mask of the camouflaged animals; Third, we show that a Transformer-based model trained on the synthetically generated data can achieve state-of-the-art performance on the MoCA-Mask video camouflaged animal segmentation benchmark.", + "bbox": [ + 75, + 229, + 472, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 460, + 215, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camouflage Evaluation. Although there are no previous computational works that directly assess camouflage, as far as we know, there have been a number of human studies. Previous works proposed methods to evaluate camouflage by analysing the human viewers' eye movements [22, 37] or conducting subjective perceptual experiments [12, 27, 32]. In [12, 27], participants were asked to indicate the camouflaged object and their answers were analysed in terms of accuracy and time needed per image. Similarly, Skurowski et al. [32] asked human volunteers to rate CHAMELEON [32] images from 1 to 5 in terms of camouflage effectiveness. They produced a score after compensating for personal bias. We validate our proposed camouflage scoring functions by comparing our rankings to these human-based rankings.", + "bbox": [ + 75, + 487, + 472, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motion Segmentation. The goal of this task is to partition the frames of a video sequence into background and independently moving objects. Early approaches focused on clustering motion patterns by grouping pixels with similar trajectories [3, 17, 21, 24, 25, 26, 31]. Another line of research tackled the problem by compensating for the background motion [1, 2, 19], via registering consecutive frames, or explicitly estimating camera motion. More recently, in [40], a Transformer-like architecture is used to reconstruct the input flows, and the segmentation masks are generated as a side product, while [6] decomposes the flow filed into regions by fitting affine transformations. The most", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "closely related to ours is [18, 39], where the authors adopt a Sim2Real strategy by training the model on optical flow computed from synthetic videos. These models can generalise towards real videos without fine-tuning. In this paper, we train a model on synthetic camouflage RGB sequences.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Frechet Inception Distance and its variants. Also known as Wasserstein-2 distance [38], Fréchet distance (FD) is a metric quantifying the difference between two probability distributions [10]. Lately, [13] introduced the Fréchet Inception Distance (FID) in the context of generative models. Under the assumption that real images and generated images follow Gaussian distributions, they compute the FD of the two distributions. Note that here, each image is represented by a vector obtained from the last pooling layer of InceptionV3 [35]. Following their steps, [30] adapted FID to the image re-targeting task and introduced the Single Image Fréchet Inception Distance (SIFID). Instead of comparing entire image datasets, they compare two images and consider the distributions of their feature vectors obtained from earlier layer of InceptionV3 [35]. Recently, [7] included FID in the training objective and introduced Fréchet-GAN. Our work takes inspiration from [7, 30] and investigates a FID of regions within the same image as an auxiliary loss.", + "bbox": [ + 496, + 170, + 892, + 443 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Measuring the effectiveness of camouflage", + "text_level": 1, + "bbox": [ + 498, + 455, + 875, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Assuming there exists a dataset, $\\mathcal{D} = \\{(I_1,m_1),\\ldots ,$ $(I_N,m_N)\\}$ , where $I_{i}\\in \\mathbb{R}^{H\\times W\\times 3}$ refers to the images, and $m_i\\in \\{0,1\\}^{H\\times W\\times 1}$ denotes a binary segmentation mask of the camouflaged animal. Here, we aim to design a scoring function that takes the image and segmentation mask as input, and outputs a scalar value $S$ that can quantify the effectiveness of camouflage, i.e., how successfully an animal blends into its background, termed as the camouflage score,", + "bbox": [ + 496, + 482, + 890, + 603 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nS: (I _ {i}, m _ {i}) \\longmapsto s _ {i} \\in [ 0, 1 ]\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 630, + 787, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "specifically, for $i \\neq j$ and $s_i < s_j$ , the animal in $I_j$ , indicated by the mask $m_j$ , is more concealed than the animal in $I_i$ , indicated by the mask $m_i$ . Having such a scoring function enables various applications, first, it enables to rank the images from one dataset, in terms of the effectiveness of camouflage, and generate camouflage-relevant statistics for the dataset; second, it can serve as the objective for low-level image generation tasks, for example, image inpainting, editing, etc.", + "bbox": [ + 496, + 657, + 890, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the following sections, we investigate three such scoring functions, specifically, we exploit the animal's mask and quantify the key aspects that determine the camouflage success both perceptually (Sec. 3.1) and probabilistically (Sec. 3.2). Note that, we take into account the local aspect of camouflage, i.e., instead of processing the entire background region, we only focus on the immediate", + "bbox": [ + 496, + 795, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "833", + "bbox": [ + 485, + 924, + 509, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c6ff09a76b9d508514d200550ab067f87aa64e16710d61143837536fab88a82.jpg", + "image_caption": [ + "1" + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 176, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1759ca146b405d8dabfef343fbec1b5c33972393ce87ede6d026804762e706a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 179, + 90, + 276, + 141 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6085a984a7ccbde1dcba2ab54db69eed384d4fcc384e1f928c38adbc40d1f684.jpg", + "image_caption": [ + "$I_{bg}$" + ], + "image_footnote": [], + "bbox": [ + 179, + 142, + 276, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f0ddbae7d3aaaf399ac03767840eb5b6ae4b66acf19036927fb7ef5d19cb4ea4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 90, + 375, + 140 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/49bb0f76d2411510dcabfc805a9c12eac5e5abacf324da3fc03f2e361baa5415.jpg", + "image_caption": [ + "Figure 2: The reconstruction fidelity score $S_{R_f}$ evaluates the similarities between the original foreground $I_{fg}$ and its reconstruction from background features $\\Psi_{bg}(I_{fg})$ . The top example shows a case where the animal exhibits different visual patterns (color) from its background $(S_{R_f} = 0.11)$ , while the bottom example shows a better background matching $(S_{R_f} = 0.82)$ ." + ], + "image_footnote": [], + "bbox": [ + 279, + 142, + 375, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/81dfcb9dc60e8f7c80ec6ba0f4ae8c4c422310ffc4a6bb754bd6a2835efa3078.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 90, + 475, + 141 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/71fa186d169a144de8fb90476ef32242a700c72953a1d9f26fc7551798689c07.jpg", + "image_caption": [ + "$\\Psi_{bg}(I_{fg})$" + ], + "image_footnote": [], + "bbox": [ + 377, + 142, + 475, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "surround of the animal. Hence, when referring to background, we mean the local background and only consider the cropped images centred around the object.", + "bbox": [ + 75, + 339, + 468, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Perceptual camouflage score", + "text_level": 1, + "bbox": [ + 76, + 393, + 331, + 410 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we define the camouflage score through a perceptual lens, specifically, we measure the foreground and background similarity (Sect. 3.1.1), along with the boundary visibility (Sect. 3.1.2).", + "bbox": [ + 75, + 416, + 468, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Reconstruction fidelity score", + "text_level": 1, + "bbox": [ + 76, + 489, + 328, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We attempt to reconstruct the foreground region with patches from the background, and propose a score to quantitatively measure the discrepancy between the original image and its reconstruction. Intuitively, for successful camouflage, we should be able to reconstruct the foreground perfectly by copying patches from its close neighborhood.", + "bbox": [ + 75, + 513, + 468, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, for a given image $(I)$ and segmentation mask $(m)$ , subscript has been ignored for simplicity, we consider a trimap separation and define the foreground and background regions using morphological erosion and dilation of the mask,", + "bbox": [ + 75, + 604, + 468, + 678 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathrm {f g}}, I _ {\\mathrm {b g}} = m _ {\\mathrm {f g}} \\odot I, m _ {\\mathrm {b g}} \\odot I \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 691, + 468, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $m_{\\mathrm{fg}} = \\operatorname{erode}(m)$ and $m_{\\mathrm{bg}} = 1 - \\operatorname{dilate}(m)$ .", + "bbox": [ + 76, + 719, + 410, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To reconstruct the foreground region, we replace the foreground patch with the closest one in the background. Here a patch is a $n \\times n$ pixel region ( $n = 7$ in our case with an overlap of 3) and the patchwise similarity is computed by exploiting the low-level perceptual similarity, i.e., comparing corresponding RGB values. This reconstruction method is inspired by the texture generation and inpainting approaches, such as [8]. In practice, we implement it with fast approximate nearest neighbor search for efficiency.", + "bbox": [ + 75, + 734, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The reconstruction fidelity score is computed by assessing the difference value between the foreground region and", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1943e4b80c021397a9deb3c8d7e5c1c2515ae3ac3362302d2719334a9a40aa4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 90, + 591, + 141 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d7d53a5c03c2d669c20cc43dffa73bc5353dccf66a7254f19be6f04eac09352e.jpg", + "image_caption": [ + "I" + ], + "image_footnote": [], + "bbox": [ + 496, + 142, + 591, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/66fb6378b291b130de9d698e6f20bfc45618b095b2cb48a688fbf333b138f7c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 90, + 687, + 141 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2a267669f36fb0477be3d48781fcada221b81ea83127b12a460e6fe96381523a.jpg", + "image_caption": [ + "Trimap" + ], + "image_footnote": [], + "bbox": [ + 596, + 142, + 687, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/13e2f1c90b7851132a4ee4256a4cf3d668266358acef757e31f5d6424b048dfc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 90, + 784, + 141 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5f03892f6c37fa35cee3086f08952931687805dcf804a6cd95ce6f3b8464cea7.jpg", + "image_caption": [ + "C", + "Figure 3: Trimap regions and animal contours. From left to right: Input image; Trimap partition (foreground region in yellow, background in purple, and boundary region in green); Contours $C$ computed on the original image; and ground truth Contours $C_{gt}$ . The top example shows less contour agreement along the boundary region ( $S_b = 0.72$ ) than the bottom example ( $S_b = 0.40$ )." + ], + "image_footnote": [], + "bbox": [ + 691, + 142, + 785, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/267e36e0db0341525401d83dbe0e77f5207a9175695eb9822a168d31da6e4d0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 98, + 870, + 138 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/24deecae6fa693f42c16a9961cc19aec933802e5cf6ce7705d4cab5b575995fe.jpg", + "image_caption": [ + "$C_{gt}$" + ], + "image_footnote": [], + "bbox": [ + 787, + 143, + 870, + 185 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "its reconstruction, specifically, we count the number of foreground pixels that have been successfully reconstructed from the background:", + "bbox": [ + 496, + 339, + 892, + 383 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {R _ {f}} (I, m) = \\frac {1}{N _ {\\mathrm {f g}}} \\sum_ {(i, j) \\in I _ {\\mathrm {f g}}} R _ {f} (i, j) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 395, + 890, + 434 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {f} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} | | I _ {\\mathrm {f g}} - \\Psi_ {I _ {\\mathrm {b g}}} \\left(I _ {\\mathrm {f g}}\\right) | | _ {2} < \\lambda | | I _ {\\mathrm {f g}} | | _ {2} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 436, + 890, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\Psi_{I_{bg}}(.)$ denotes the reconstruction operation, $N_{\\mathrm{fg}} = |m_{\\mathrm{fg}}|$ is the total number of pixels in the foreground region, and $\\lambda$ is a threshold ( $\\lambda = 0.2$ in our case). A higher $S_{R_f}$ score, means that the animal's visual attributes are well represented in the background. Conversely, a low $S_{R_f}$ indicates that the animal's appearance is composed of unique features that makes it standout from its surrounding. In Fig. 2, we present examples of camouflage evaluation by reconstruction fidelity.", + "bbox": [ + 496, + 488, + 890, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 Boundary visibility score", + "text_level": 1, + "bbox": [ + 500, + 622, + 725, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the trimap representation, we can easily extract the boundary region, $I_{\\mathrm{b}} = m_{\\mathrm{b}} \\odot I$ , where $m_{\\mathrm{b}} = (1 - m_{\\mathrm{bg}}) - m_{\\mathrm{fg}}$ . Here, we aim to measure the animal's boundary properties, e.g., contour visibility, as they can also provide visually evident cues for camouflage breaking. In particular, we adopt an off-the-shelf contour extraction method [29], and run it on the original image and foreground mask, to generate the images' contour ( $C$ ) and ground truth animal's contour ( $C_{\\mathrm{gt}}$ ), as shown in Fig. 3.", + "bbox": [ + 496, + 647, + 890, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We express the agreement between predicted contours $C$ and ground truth contours $C_{\\mathrm{gt}}$ over the boundary region with the F1 metric:", + "bbox": [ + 496, + 782, + 890, + 827 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {b} (I, m) = 1 - \\mathrm {F} 1 \\left(m _ {b} \\odot C _ {\\mathrm {g t}}, m _ {b} \\odot C\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 840, + 890, + 858 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This score penalises the boundary pixels that are predicted as contour in both $C$ and $C_{\\mathrm{gt}}$ . While we do not expect a", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "834", + "bbox": [ + 485, + 924, + 509, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "perfect contour agreement for visible edges, we consider $S_{b}$ as a reasonable approximation, given the shape and size of the boundary region as a thin margin centred around the animal. If a boundary pixel is predicted as a contour in $C_{\\mathrm{gt}}$ but not in $C$ , this means that the animal's edge is not visible in the original image. If a pixel is predicted as a contour in $C$ but not $C_{\\mathrm{gt}}$ , this indicates the presence of distracting elements, such as complex texture patterns on the foreground animal, which also affects the visibility of the animal's contour and therefore improves its camouflage. A perfectly camouflaged animal would have little contour agreement and $S_{b} = 1$ .", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To get the final perceptual camouflage score, we linearly combine both reconstruction fidelity score and boundary visibility score:", + "bbox": [ + 75, + 273, + 470, + 319 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\alpha} = (1 - \\alpha) S _ {R _ {f}} + \\alpha S _ {b} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 333, + 468, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We will describe our method for setting the weighting parameter $\\alpha$ in the Experiment section (Sec. 6).", + "bbox": [ + 76, + 364, + 468, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Probabilistic scoring function", + "text_level": 1, + "bbox": [ + 76, + 407, + 339, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to using the low-level RGB representation, in this section, we consider the pixelwise representation in the feature space, and propose a differentiable metric that compares the probabilistic distribution between the foreground and background regions, which acts as a proxy for the score $S_{\\alpha}$ described in previous section, and can be used directly as a differentiable loss function in image generation.", + "bbox": [ + 75, + 431, + 468, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider the Intra-Image Fréchet Inception Distance, specifically, we compute the feature maps for the foreground and background image regions:", + "bbox": [ + 75, + 539, + 468, + 584 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {f g}, f _ {b g} = \\Phi_ {\\mathrm {v}} \\left(I _ {f g}\\right), \\Phi_ {\\mathrm {v}} \\left(I _ {b g}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 598, + 468, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\Phi_{\\mathrm{v}}(\\cdot)$ denotes a pre-trained Inception network [34, 35], truncated at early layers. We take inspiration from [30], and adapt the single image Fréchet distance to measure the deviation between feature distributions of different regions within the same image. We adopt the Fréchet hypothesis with respect to our regions, i.e., the features of the foreground and background follow multivariate Gaussian distributions: $f_{\\mathrm{fg}} \\sim \\mathcal{N}(\\mu_{\\mathrm{fg}}, \\Sigma_{\\mathrm{fg}})$ and $f_{\\mathrm{bg}} \\sim \\mathcal{N}(\\mu_{\\mathrm{bg}}, \\Sigma_{\\mathrm{bg}})$ . The intra-image Fréchet distance can be formulated as follows:", + "bbox": [ + 75, + 628, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathcal {F}} ^ {2} (I, m) = \\left\\| \\mu_ {\\mathrm {f g}} - \\mu_ {\\mathrm {b g}} \\right\\| _ {2} ^ {2} + T r \\left(\\Sigma_ {\\mathrm {f g}} + \\Sigma_ {\\mathrm {b g}} - 2 \\left(\\Sigma_ {\\mathrm {f g}} \\Sigma_ {\\mathrm {b g}}\\right) ^ {1 / 2}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 792, + 478, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that, when $(I,m)$ is the output from a generative model, $d_{\\mathcal{F}}^2$ is differentiable with respect to its parameters, and can therefore be used as an auxiliary loss for optimising the image generation procedure, i.e., generate effective camouflage examples.", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Generating camouflage video sequences", + "text_level": 1, + "bbox": [ + 498, + 90, + 856, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we propose a scalable pipeline for generating images with concealed animals. In particular, we incorporate the differentiable intra-image Fréchet distance in image generation, explicitly optimising the camouflage score of the image. We first detail the static image generation framework, then describe how we use the images it generates to create camouflage video sequences.", + "bbox": [ + 496, + 114, + 893, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Camouflage image generation", + "text_level": 1, + "bbox": [ + 498, + 229, + 764, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to generate the image with a camouflaged animal, and its corresponding segmentation mask, we adopt a Generative Adversarial Network (GAN), where a generator is fed with a latent vector $z \\sim \\mathcal{N}(0,1)$ and learns to predict the pair of realistic image and segmentation mask, such that a discriminator cannot differentiate it from real samples.", + "bbox": [ + 496, + 253, + 890, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $(x_{i},m_{i})$ denote an image and segmentation pair sampled from the training dataset, a subset of COD10K [9] is used here. We adopt the conventional label notation, i.e. $y = 1$ for (real) examples sampled from the training dataset and $y = 0$ for (fake) examples from the generator, and consider the generative adversarial loss functions [11]:", + "bbox": [ + 496, + 344, + 893, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} = E _ {\\left(x _ {i}, m _ {i}\\right)} \\left[ - \\log \\left(D \\left(x _ {i}, m _ {i}\\right)\\right) \\right] + E _ {z} \\left[ - \\log \\left(1 - D (G (z))\\right) \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 444, + 911, + 462 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G} = E _ {z} \\left[ - \\log (D (G (z))) \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 464, + 687, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enforce coherence between images and their segmentation masks, we feed the discriminator with additional fake pairs, consisting of real images coupled with unpaired real masks. This can be formulated as an additional coherence loss term that minimises the probability of assigning incorrect labels, i.e. $y = 1$ , to $(x_{i},m_{j})$ sampled from the training set such that $i\\neq j$ :", + "bbox": [ + 496, + 489, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C o h}} = E _ {\\left(x _ {i}, m _ {j}\\right) i \\neq j} [ - \\log (1 - D \\left(x _ {i}, m _ {j}\\right)) ] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 606, + 890, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To increase the camouflage effectiveness in the generated examples, we adopt the intra-image Fréchet distance in the generator loss, as an auxiliary loss term, and train our camouflage image generator with the following loss:", + "bbox": [ + 496, + 633, + 890, + 695 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\tilde {D}} = \\mathcal {L} _ {D} + \\mathcal {L} _ {\\mathrm {C o h}} \\quad \\mathcal {L} _ {\\tilde {G}} = \\mathcal {L} _ {G} + \\beta d _ {\\mathcal {F}} ^ {2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 704, + 890, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present an overview of our data generation framework in Fig. 4.", + "bbox": [ + 496, + 732, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Camouflage video generation", + "text_level": 1, + "bbox": [ + 500, + 771, + 759, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a camouflage image and corresponding segmentation mask, generated using the method above, we can create video sequences by applying different motions to the foreground and background, in a similar manner to [18]. Specifically, we first inpaint the backgrounds with an off-the-shelf model proposed by Suvorov et al. [33] and overlay the synthetic animal (extracted from the original generated image", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "835", + "bbox": [ + 485, + 924, + 509, + 936 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/43d5b3a4c21a12801039fe84cedce83142917067f3d79df01c89be4f126f27c2.jpg", + "image_caption": [ + "Figure 4: Our framework consists of a synthetic camouflage data generation pipeline (left), where we train a generator $G$ in a GAN setting, to create camouflage images and masks, while encouraging camouflage effectiveness via minimising $\\mathcal{L}_{\\mathcal{F}}$ . The generated samples are then transformed into synthetic video sequences (middle), following our method presented in Sec. 4.2. The transformer-based motion segmentation architecture (right), for camouflage breaking in videos, is trained on the synthetic video sequences. The architecture is described in Sec. 5." + ], + "image_footnote": [], + "bbox": [ + 116, + 89, + 254, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/87982e9137d9d52a0682cd4d20f167dcac222049a339fbd9973432f2a959dd3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 88, + 379, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/273e9d1cc6704b62007721a334ae0c53e98ce159e2e471b72d76bee842bdacb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 88, + 852, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by using the mask) at a random location within the image in the first frame, then following a random translational trajectory in the following frames. We incorporate a different translational motion for the background and include random sub-sequences where the foreground and background undergo the same motion to simulate momentarily static objects. As we train our generator with intra-image Fréchet loss $(\\mathcal{L}_{\\mathcal{F}})$ , the generated images exhibit strong similarities between the foreground and the background, and even if the object is placed at a different location from the original, we expect it to remain concealed within its surrounding.", + "bbox": [ + 75, + 351, + 473, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Learning to break camouflage", + "text_level": 1, + "bbox": [ + 76, + 545, + 354, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we train a transformer-based architecture on the synthetic dataset, and demonstrate its effectiveness for breaking the camouflage in real videos, for example, MoCA [19]. We build on two previous architectures, namely, the motion segmentation from [18] and the Search Identification Network from [9]. While the first architecture processes sequences of optical flow, the second takes single RGB images as inputs and treats them separately. Our proposed model, shown in Fig. 4, takes both the optical flow and RGB frame sequences as input. The flow is computed with RAFT [36] from the RGB sequence, and processed with a motion encoder, followed by a transformer encoder. The RGB sequence is processed with the appearance encoder from [9], pre-trained for framewise camouflage segmentation. Then both streams are aggregated as inputs to a Mask2Former-like [4] transformer decoder and pixel decoder to produce the high resolution segmentation mask from the motion stream.", + "bbox": [ + 75, + 575, + 468, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion encoder. To encode the motion cues, we use a light-weight convNet that takes as input a sequence of optical flows, $\\mathbf{I}_{\\mathbf{m}} = \\{I_{m_1}, I_{m_2},.., I_{m_t}\\} \\in \\mathcal{R}^{t \\times c_0 \\times h \\times w}$ and out", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "puts motion features:", + "bbox": [ + 500, + 351, + 643, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{f _ {m _ {1}}, f _ {m _ {2}}, \\dots , f _ {m _ {t}} \\right\\} = \\Phi_ {\\text {m o t i o n}} \\left(\\mathbf {I} _ {\\mathbf {m}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 382, + 810, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where each flow frame is separately embedded.", + "bbox": [ + 500, + 416, + 812, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion transformer encoder. A transformer encoder takes as input the motion features, concatenated along the sequence, together with learned spatial and temporal positional encodings $(\\mathrm{Pos}_s,\\mathrm{Pos}_t)$ , as indicated in Fig 4. $\\mathrm{Pos}_t$ specifies the frame, and $\\mathrm{Pos}_s$ specifies the position in the spatial feature map output of the motion encoder. The output of the transformer is a set of enriched motion features.", + "bbox": [ + 496, + 436, + 890, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Appearance encoder. Here, we adopt a SINet-v2 [9] architecture, that encodes the RGB sequence, $\\mathbf{I}_{\\mathbf{a}} = \\{I_{a_1}, I_{a_2},.., I_{a_t}\\} \\in \\mathcal{R}^{t \\times c_1 \\times h \\times w}$ into appearance features:", + "bbox": [ + 500, + 549, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{f _ {a _ {1}}, f _ {a _ {2}}, \\dots , f _ {a _ {t}} \\right\\} = \\Phi_ {\\text {a p p}} \\left(\\mathbf {I} _ {\\mathbf {a}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 608, + 792, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Again, each RGB frame is processed separately by SINet.", + "bbox": [ + 500, + 642, + 880, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformer decoder. A transformer-based decoder takes the output of the motion transformer encoder together with the appearance features and a learnable query for the mask embedding. In a similar manner to Mask2Former [4], the query attends to multiple resolutions of the motion features concatenated with the appearance features and produces a mask embedding for the moving object.", + "bbox": [ + 496, + 662, + 890, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pixel Decoder. Similarly to the pixel decoder in Mask2Former, a light-weight convNet decoder is used with skip-connections to recover high-resolution segmentation masks $\\{\\hat{m}_1,\\hat{m}_2,\\dots ,\\hat{m}_t\\}$ from the motion features and the mask embedding. This is shown as the blue box in Fig. 4.", + "bbox": [ + 496, + 773, + 890, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training objective. We train the motion segmentation model on our synthetic video sequences using the binary cross-entropy loss $\\mathcal{L}_{BCE}$ .", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "836", + "bbox": [ + 485, + 925, + 511, + 936 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 76, + 90, + 209, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we start by introducing the datasets involved in this paper, followed by the implementation details. Our experiments present a thorough analysis of the proposed camouflage scores and demonstrate their effectiveness in our training framework.", + "bbox": [ + 76, + 116, + 468, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 207, + 179, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here we describe the publicly available camouflage datasets that we included in our experiments, as shown in Fig. 5, as well as the synthetic camouflage datasets that we generated.", + "bbox": [ + 76, + 232, + 468, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CHAMELEON [32] is one of the first camouflage image datasets. It contains 76 images collected using Google image search with the 'camouflaged animal' keyword and include ground truth manual annotations.", + "bbox": [ + 76, + 296, + 468, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CAMO [20]. The Camouflaged Object dataset consists of 2500 images collected from the internet, of which 1250 images (1000 training sub-set and 250 testing sub-set) contain at least one camouflage instance from 8 categories with manual pixelwise annotations provided.", + "bbox": [ + 76, + 361, + 468, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "COD10K [9]. COD10K contains 10,000 images collected from photography websites of which 5066 depict camouflaged animals (3040 training sub-set and 2026 testing subset), organised into 10 classes, and 78 sub-classes (69 camouflaged). Note that, in our camouflage evaluation experiments we only used the camouflage subset of this dataset.", + "bbox": [ + 76, + 441, + 468, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Camouflaged Animals [1]. This dataset consists of 9 video sequences of camouflage animals from 6 categories.", + "bbox": [ + 76, + 536, + 468, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MoCA [19] is the first large-scale camouflage video dataset. It contains 141 video sequences (37K frames) totalling 67 animal categories. Recently, other works have curated this dataset by selecting only the videos with more prominent motion (locomotion) [40], and others have provided dense pixel annotation in MoCA-Mask [5]. We use the latter version in our experiments.", + "bbox": [ + 76, + 570, + 468, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Camouflaged cuboids [12, 27]. This dataset was created for texture synthesis and consists of multiple-view scenes, where cuboids were placed at a predefined location then synthetically camouflaged. In our evaluation, we consider the 4-views generated textures [12] from 36 scenes as well as the cuboids masks.", + "bbox": [ + 76, + 680, + 468, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic Camouflage Images. Using the method described in Sec. 4.1, we generate a synthetic camouflage dataset, of $5K$ frames, by discriminating against real camouflage image and annotation masks from COD10K.", + "bbox": [ + 76, + 775, + 468, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic Camouflage Videos. We generate $1K$ sequences of 30 frames each, incorporating static sequences using the framework described in Sec. 4.2. We split these into 800 sequences for training and 200 for testing.", + "bbox": [ + 76, + 840, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6f1b01b60fa09cb7a1475102ac7312ac02003bd5e91c5053782ebfe5639ed8d2.jpg", + "image_caption": [ + "Figure 5: Randomly sampled examples from the camouflage datasets included in our work. For the video datasets, Camouflaged Animals and MoCA-Mask, we show an example sequence. For the multiple-view dataset Camouflaged cuboids (bottom), we show example views from a scene with the 4-view texture synthesis method from [12]." + ], + "image_footnote": [], + "bbox": [ + 498, + 95, + 888, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 830, + 714, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiments, for a given camouflage example, the kernels for the morphological operations are selected from a range of values [1,10], so that the resulting annotation mask", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "837", + "bbox": [ + 485, + 924, + 509, + 935 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is reduced by $20\\%$ for erosion and extended by $20\\%$ for dilation. This allows the erosion/dilation to be adapted to the size of the camouflaged animal, while always keeping a reasonably large region for addressing the boundary effects.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For our synthetic camouflage image generation, we adopt a Style-GAN architecture [41] and train it on the camouflage images from COD10K to generate $256 \\times 256$ images and masks pairs. We use $\\beta = 0.1$ to weight our intra-image Fréchet auxiliary loss. For numerical stability, we adopt the Newton Schultz iterations when calculating the matrix square root term in Intra-Image Fréchet Inception Distance, with $T = 100$ iterations.", + "bbox": [ + 75, + 162, + 468, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When generating the image sequences, we compute optical flows with RAFT [36], and train our moving camouflaged animal segmentation model. We use batches of size 2 and sequences of size 5. We first train on our synthetically generated dataset with a learning rate of $5 \\times 10^{-4}$ for 500 iterations then fine-tune on the training subset of MoCA-Mask.", + "bbox": [ + 75, + 286, + 468, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Results", + "text_level": 1, + "bbox": [ + 76, + 387, + 171, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This section presents qualitative and quantitative results, to demonstrate the effectiveness of our score functions and the benefit of including $d_{\\mathcal{F}}^2$ in the training loop.", + "bbox": [ + 75, + 411, + 468, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3.1 Evaluation on camouflage effectiveness", + "text_level": 1, + "bbox": [ + 76, + 465, + 401, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ranking camouflage images in COD10K. We can rank the images based on the effectiveness of the animal's camouflage, with our proposed scoring functions. In the left part of Fig. 6, we show the four best scored images from the large scale COD10K dataset. We can make the following observations: (i) the best examples with respect to $S_{R_f}$ (top), which favours the background matching of the animals without taking into account the boundary region, all exhibit visible boundaries, with the exception of image (a). This is especially noticeable in the rabbit example (d) along the ears and shadow regions; (ii) the best four examples with respect to the boundary score $S_b$ are all from the caterpillar subclass, mostly the baron type. These insects have thin, and transparent boundaries, that makes them the perfect candidates for this score. However, with the exception of (a) and (c), the animals still stands out as they exhibit colors and patterns that are not present within their background; (iii) the $S_\\alpha$ combining both scores shows highly effective camouflage examples. In contrast, the right part of Fig. 6 shows the lowest scored images for each approach. For $S_{R_f}$ , we find examples with low background matching in the first row, with two ant examples, i.e., (e), (f). While the boundary score penalises examples with high contrasts between the animal and its background, that results in more visible contours. This is also the case for the lowest $S_\\alpha$ score examples.", + "bbox": [ + 75, + 489, + 470, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset level camouflage comparison. We compute the", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "camouflage score for all the camouflage image and video datasets and report the results in Tab. 1. For the image datasets, the dataset level score is computed as the mean of the image scores; for the video dataset the score is first computed per frame, and per-video average computed, then the dataset level score is the mean of the averages. Our experiments show that MoCA-Mask [5, 19] contains the most successful camouflages according to our scoring. While COD10K [9] subsets are balanced in terms of camouflage effectiveness, we find that, for CAMO [20], the test subset yields higher scores than the training subset and therefore contains more challenging examples.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We note that Camouflaged cuboids dataset yields higher $S_{R_f}$ and $S_{\\alpha}$ scores compared to our synthetic datasets. This is due to the fact that the model used in [12, 27] is only trained to produce optimal texture for a predefined region (cuboid) in a particular location of a given scene. However, our model learns to output a new image with a randomly located camouflaged object of random shape, thus offering more scalability and diversity compared to the cuboids dataset which cannot be used to train a model for breaking real camouflage.", + "bbox": [ + 496, + 273, + 890, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison to human-produced rankings. We compare our rankings to the human scoring system based on the rating of CHAMELEON [32] and the time-to-find for Camouflaged cuboids [12]. To search for the optimal $\\alpha$ parameter in Equation (5), we select 15 images from CHAMELEON and compare their $S_{\\alpha}$ ranking with ground truth using kendl- $\\tau$ metric [16]. For $\\alpha = 0.35$ , we obtain kendl- $\\tau = 0.51$ on this validation set that we excluded from our test reported in Tab. 2. We can draw the following observations: (i) we find that the boundary score $S_{b}$ produces more agreement with the ground truth ranking than $S_{R_{f}}$ , suggesting that human observers tend to pay more attention to contour visibility than background matching; (ii) while comparing to the ranking from Camouflaged cuboids, we found negative correlations for $S_{R_{f}}$ and $d_{\\mathcal{F}}$ , we conjecture that this may be due to the nature of the dataset in [12], i.e., all the textured cuboids are synthetically generated with very high background matching; (iii) for both experiments, we obtain the strongest agreement with the $S_{\\alpha}$ score combining both background similarity and boundary blending.", + "bbox": [ + 496, + 428, + 890, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further analysis of the synthetic datasets. We adopt the FID [13] and IS [30] metrics in Table 4. These metrics assess: (i) the similarity between the set of synthetic images output by the generator and the set of real images used to train it (FID); and (ii) the clarity of the object and the diversity of images in terms of classes (IS). While they are not intended for measuring camouflage success, which is the focus of this work, IS could somewhat be (inversely) linked to $S_{b}$ as the object clarity and boundary visibility are close. Note that the object clarity (IS) is decreased with $\\mathcal{L}_{\\mathcal{F}}$ which is the intended purpose of improving camouflage, however,", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "838", + "bbox": [ + 485, + 924, + 509, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/baf0bf0fcb2fab0b8bf954cb197bfe3f0b254b42dd3957dbe58611c4affd46e8.jpg", + "image_caption": [ + "Figure 6: Left: Top-4 scored examples from COD10K for $S_{Rf}(top)$ , $S_b(midle)$ and $S_{\\alpha}$ (bottom). Right: Lowest scored examples from COD10K. For each example, we show the obtained score and the corresponding ground truth mask." + ], + "image_footnote": [], + "bbox": [ + 125, + 90, + 849, + 330 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/29f34b477c02d008cf88a5717e005baeb535c9fada4a858cf862256c470ca993.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Results of the proposed scores on natural camouflage datasets (top) and synthetically generated camouflage datasets (bottom). We report the mean scores for the single image datasets. For the video and multiple view datasets, we compute the mean per sequence and scene respectively. Syn. Camouflage w. $\\mathcal{L}_{\\mathcal{F}}$ refers to the synthetic image dataset that we generated while minimising $\\mathcal{L}_{\\mathcal{F}}$ and Syn. Camouflage Video its sequence version." + ], + "table_body": "
DatasetsData typeSRf↑Sb↑Sα↑d2F↓
CHAMELEON [32]Image0.6940.4450.6070.70
CAMO Train [20]Image0.6720.4510.5951.01
CAMO Test [20]Image0.6830.4700.6081.00
COD10K Train [9]Image0.6550.4330.5770.90
COD10K Test [9]Image0.6570.4310.5780.90
Camouflaged Animals [1]Video0.6740.5360.6261.60
MoCA-Mask Train [5, 19]Video0.8500.4430.7071.14
MoCA-Mask Test [5, 19]Video0.7330.4640.6392.51
Camouflaged cuboids[12, 27]Multi-view0.8940.4330.7336.2
Syn. Camouflage w.o. L_FImage0.6080.4320.5461.36
Syn. Camouflage w. L_FImage0.6790.4470.5981.13
Syn. Camouflage VideoVideo0.6580.4300.5781.18
", + "bbox": [ + 78, + 382, + 522, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "this effect is not maintained in the sequence generation as the animal changes location within its background.", + "bbox": [ + 75, + 683, + 468, + 712 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bb44cf13ae0b476826ded89f74e7fe3658fe9898dae140f6d28404e688b20345.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
datasetFID↓IS↑
Syn. Camouflage w.o. L_F75.013.49
Syn. Camouflage w. L_F76.343.13
Syn. Camouflage Video86.423.85
", + "bbox": [ + 81, + 713, + 475, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3.2 Model ablations", + "text_level": 1, + "bbox": [ + 76, + 791, + 243, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the effectiveness of $\\mathcal{L}_{\\mathcal{F}}$ . Fig. 7 shows the generated samples from the GAN pipeline, trained with and without the intra-image Fréchet Distance. Our generator is able to produce realistic object masks of complex shapes and thin structures similar to those usually encountered in the camouflage datasets. The produced images exhibit rich back", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cbf4a818e2727880036dce2fb812edd28d8183909a428bb447da4c3205ab33a0.jpg", + "table_caption": [ + "Table 4: Overall quality of our synthetic datasets." + ], + "table_footnote": [], + "table_body": "
ScoresCHAMELEON [32]Camouflaged cuboids[12, 27]
SRf0.01-0.07
Sb0.030.42
Sα0.420.41
d2F0.10-0.17
", + "bbox": [ + 532, + 382, + 893, + 459 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fa68bab1bd615959477e3eaec9cb17f52959692beed33a6e93953d0818f5c6f3.jpg", + "table_caption": [ + "Table 2: Kendall- $\\tau$ metric between rankings produced via our scores and human scoring ground truth." + ], + "table_footnote": [ + "Table 3: Ablation study on the different components of the motion segmentation architecture and training datasets. We evaluate on MoCA-Mask for all experiments. (S: our synthetic dataset, MM: MoCA-Mask)." + ], + "table_body": "
ModelTraining datasetAppearanceEncoderTransformerEncoderTransformerAggregationmIoU↑F↑E↑MAE↓
AS14.520.457.39.4
BMM15.320.657.35.1
CS+MM16.021.859.84.3
DS+MM✓20.522.659.83.8
ES+MM✓✓23.023.557.01.6
FS+MM✓✓22.723.258.83.6
GMM✓✓✓23.424.762.02.4
HS+MM✓✓✓30.834.377.01.8
", + "bbox": [ + 535, + 497, + 885, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "grounds with realistic nature-like elements mimicking rocks and coral structures present in the real camouflage data. Adding the $\\mathcal{L}_{\\mathcal{F}}$ loss results in generating images with better blending in the background, both qualitatively and quantitatively, as shown by the $S_{\\alpha}$ computed for both datasets generated with and without the intra-image Fréchet Distance in Tab. 1. We refer the reader to the supplementary material for more examples of generated samples.", + "bbox": [ + 496, + 683, + 890, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the architecture design and gain from the synthetic dataset. We present in Tab. 3 an ablation study detailing the gain from the main components of our architecture, as well as the benefit of first training on our synthetic dataset then fine-tuning on MoCA-Mask (S+MM), for model H, as opposed to only training on MoCA-Mask (MM), for G.", + "bbox": [ + 496, + 809, + 892, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "839", + "bbox": [ + 485, + 924, + 511, + 936 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e56138070bfb2aa0549ae5bad0396b6546b425f700cba994dd51f61044fae60.jpg", + "image_caption": [ + "Figure 7: Generated camouflage images and masks: For the examples on the right, the generator was trained with $\\mathcal{L}_{\\mathcal{F}}$" + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 849, + 203 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/6c24a53ec9d6905bc1fa53804f0499382f71c07cfb142f3d4693ba064e87aee0.jpg", + "image_caption": [ + "Figure 8: Qualitative results on MoCA-Mask. From top to bottom: Appearance sequence $I_{a}$ , flow sequence $I_{m}$ , predicted segmentations and ground truth segmentations." + ], + "image_footnote": [], + "bbox": [ + 63, + 234, + 519, + 393 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/c27bf7424fdff0cb049118ebc43f0c8fff27bf0e3295f3e8086e840c5bb1832d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelRGBMotionmIoU↑F↑E↑MAE↓
SINet [9]✓20.223.169.92.8
SINet-V2 [9]✓18.020.464.23.1
SegMaR [14]✓12.222.580.34.5
ZoomNet [28]✓18.828.770.82.5
SLTNet [5]✓✓27.231.175.92.7
MG [40]✓12.716.856.16.7
Ours-flow✓17.821.560.73.7
Ours✓✓30.834.377.01.8
", + "bbox": [ + 526, + 238, + 888, + 392 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 5: Results on MoCA-Mask. Ours-flow refers to the optical flow only version of our architecture, excluding the appearance stream.", + "bbox": [ + 522, + 398, + 892, + 444 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6.3.3 Results on MoCA", + "text_level": 1, + "bbox": [ + 76, + 458, + 253, + 472 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We train our model on the synthetically generated camouflage images and fine-tune on the training set of MoCA-Mask. Tab. 5 presents our camouflage object segmentation results on the test set of MoCA-Mask. Ours refers to our proposed method with SINet-V2 as the appearance encoder and Ours-flow refers to the optical flow only version of our architecture, excluding the appearance stream. Our main model outperforms RGB and Motion-based methods on MoCA-Mask. Fig. 8 presents qualitative results of our segmentation model. Note that our model is robust to degraded optical flow and static animals.", + "bbox": [ + 75, + 484, + 468, + 650 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6.4. Limitations", + "text_level": 1, + "bbox": [ + 76, + 662, + 202, + 678 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While our image generation method encourages background matching through the $\\mathcal{L}_{\\mathcal{F}}$ loss term minimization, the sequence generating is not guaranteed to maintain object concealment. A possible solution could be to use the proposed $S_{\\alpha}$ camouflage score to curate the generated sequence dataset and filter out the most visible objects.", + "bbox": [ + 75, + 686, + 468, + 777 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our proposed camouflage scores use the ground truth annotation and analyse the different regions it defines. We found that our scoring system penalises specific cases of camouflage by occlusion, where elements from the background are partially occluding the animal and helping improve its camouflage. For instance, the grass in example (g) from the last row of Fig. 6 is not treated as part of the animal and therefore not considered in our background similarity", + "bbox": [ + 75, + 780, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "assessment. One can argue that this is due to the ambiguity in the provided annotations and for such cases, extra amodal annotation should be also considered.", + "bbox": [ + 498, + 458, + 890, + 502 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 518, + 617, + 535 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present three score functions for computationally assessing the effectiveness of camouflage in images and videos. By evaluating the similarity with background and the boundary visibility, our combined perceptual score is strongly correlated with human perceptual ranking systems on two different datasets. We demonstrate that training a generative model with our differentiable camouflage function improves the effectiveness of generated camouflage examples and can be used to generate challenging synthetic camouflage datasets to train models to break camouflage.", + "bbox": [ + 496, + 546, + 890, + 699 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 710, + 645, + 726 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We are grateful to Przemysław Skurowski for providing the human perception study data on the CHAMELEON dataset and Andrew Owens, Rui Guo and Oscar de Lima for providing the data for the Camouflaged cuboids dataset. We thank Tengda Han for fruitful discussions. This research is supported by the UK EPSRC funded CDT in Autonomous Intelligent Machines and Systems (AIMS), the EPSRC Programme Grant VisualAI EP/T028572/1, a Schlumberger Studentship, and a Royal Society Research Professorship. WX is supported by the National Key R&D Program of China (No. 2022ZD0161400).", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "840", + "bbox": [ + 485, + 925, + 511, + 936 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pia Bideau and Erik Learned-Miller. It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In ECCV, pages 433-449, 2016.", + "[2] Pia Bideau, Rakesh R Menon, and Erik Learned-Miller. Moa-net: self-supervised motion segmentation. In Proc. ECCV Workshop, 2018.", + "[3] Thomas Brox and Jitendra Malik. Object segmentation by long term analysis of point trajectories. In European conference on computer vision, pages 282-295, 2010.", + "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022.", + "[5] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022.", + "[6] Subhabrata Choudhury, Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Guess What Moves: Unsupervised Video and Image Segmentation by Anticipating Motion. In Proc. BMVC, 2022.", + "[7] Khoa D Doan, Saurav Manchanda, Fengjiao Wang, Sathiya Keerthi, Avradeep Bhowmik, and Chandan K Reddy. Image generation via minimizing frechet distance in discriminator feature space. arXiv preprint arXiv:2003.11774, 2020.", + "[8] A. Efros and T. Leung. Texture synthesis by non-parametric sampling. In Proc. ICCV, pages 1039-1046, Sep 1999.", + "[9] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021.", + "[10] Maurice Fréchet. Sur la distance de deux lois de probabilité. Comptes Rendus Hebdomadaires des Seances de L'Académie des Sciences, pages 689-692, 1957.", + "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, pages 2672-2680, 2014.", + "[12] Rui Guo, Jasmine Collins, Oscar de Lima, and Andrew Owens. Ganmouflage: 3d object nondetection with texture fields. In arXiv preprint arXiv:2201.07202, 2022.", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017.", + "[14] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In CVPR, pages 4713-4722, 2022.", + "[15] Nobukatsu Kajiura, Hong Liu, and Shin'ichi Satoh. Improving camouflaged object detection with the uncertainty of pseudo-edge labels. In ACM Multimedia Asia, pages 1-7. 2021.", + "[16] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1/2):81-93, 1938.", + "[17] Margret Keuper, Bjoern Andres, and Thomas Brox. Motion trajectory segmentation via minimum cost multicuts. In ICCV, pages 3271-3279, 2015.", + "[18] Hala Lamdouar, Weidi Xie, and Andrew Zisserman. Seg-" + ], + "bbox": [ + 78, + 114, + 467, + 896 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "menting invisible moving objects. In Proc. BMVC, 2021.", + "[19] Hala Lamdouar, Charig Yang, Weidi Xie, and Andrew Zisserman. Betrayed by motion: Camouflaged object discovery via motion segmentation. In Proc. ACCV, 2020.", + "[20] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding, pages 45-56, 2019.", + "[21] José Lezama, Karteek Alahari, Josef Sivic, and Ivan Laptev. Track to the future: Spatio-temporal video segmentation with long-range motion cues. In CVPR, pages 3369-3376, 2011.", + "[22] Chiuhsiang Joe Lin, Chi-Chan Chang, and Yung-Hui Lee. Evaluating camouflage design using eye movement data. Applied ergonomics, pages 714-723, 2014.", + "[23] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In CVPR, pages 8772-8781, 2021.", + "[24] Peter Ochs and Thomas Brox. Object segmentation in video: a hierarchical variational approach for turning point trajectories into dense regions. In ICCV, 2011.", + "[25] Peter Ochs and Thomas Brox. Higher order motion models and spectral clustering. In CVPR, pages 614-621, 2012.", + "[26] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. volume 36, pages 1187-1200, 2013.", + "[27] Andrew Owens, Connelly Barnes, Alex Flint, Hanumant Singh, and William Freeman. Camouflaging an object from many viewpoints. In CVPR, pages 2782-2789, 2014.", + "[28] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages 2160-2170, 2022.", + "[29] Xavier Soria Poma, Edgar Riba, and Angel Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In Proc. WACV, pages 1923-1932, 2020.", + "[30] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In ICCV, pages 4570-4580, 2019.", + "[31] Josef Sivic, Frederik Schaffalitzky, and Andrew Zisserman. Object level grouping for video shots. In Proc. ECCV. Springer-Verlag, 2004.", + "[32] Przemysław Skurowski, Hassan Abdulameer, J Blaszczyk, Tomasz Depta, Adam Kornacki, and P Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2018. URL: https://www.polsl.pl/rau6/chameleon-database-animal-camouflage-analysis/.", + "[33] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021.", + "[34] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proc. CVPR, 2015.", + "[35] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception ar-" + ], + "bbox": [ + 501, + 92, + 890, + 887 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "841", + "bbox": [ + 486, + 925, + 509, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "chitecture for computer vision. In CVPR, pages 2818-2826, 2016.", + "[36] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020.", + "[37] Jolyon Troscianko, John Skelhorn, and Martin Stevens. Quantifying camouflage: how to predict detectability from appearance. BMC evolutionary biology, 17(1):1-13, 2017.", + "[38] Cédric Villani. Optimal transport: old and new, volume 338. Springer, 2009.", + "[39] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. In NeurIPS, 2022.", + "[40] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proc. ICCV, 2021.", + "[41] Bowen Zhang, Shuyang Gu, Bo Zhang, Jianmin Bao, Dong Chen, Fang Wen, Yong Wang, and Baining Guo. Styleswin: Transformer-based gan for high-resolution image generation. In CVPR, pages 11304-11314, 2022." + ], + "bbox": [ + 78, + 90, + 467, + 347 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "842", + "bbox": [ + 486, + 925, + 509, + 935 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/The Making and Breaking of Camouflage/b706bafb-dae1-4647-ad2f-702e05fcaa82_model.json b/2023/The Making and Breaking of Camouflage/b706bafb-dae1-4647-ad2f-702e05fcaa82_model.json new file mode 100644 index 0000000000000000000000000000000000000000..56e63bcdcb559a5da779916ac67183a39f84c1bf --- /dev/null +++ b/2023/The Making and Breaking of Camouflage/b706bafb-dae1-4647-ad2f-702e05fcaa82_model.json @@ -0,0 +1,2664 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.275, + 0.131, + 0.697, + 0.154 + ], + "angle": 0, + "content": "The Making and Breaking of Camouflage" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.18, + 0.385, + 0.197 + ], + "angle": 0, + "content": "Hala Lamdouar1" + }, + { + "type": "text", + "bbox": [ + 0.422, + 0.18, + 0.523, + 0.197 + ], + "angle": 0, + "content": "Weidi Xie\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.559, + 0.181, + 0.717, + 0.197 + ], + "angle": 0, + "content": "Andrew Zisserman" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.199, + 0.498, + 0.217 + ], + "angle": 0, + "content": "Visual Geometry Group, University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.199, + 0.845, + 0.218 + ], + "angle": 0, + "content": "\\(^{2}\\)CMIC, Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.22, + 0.641, + 0.234 + ], + "angle": 0, + "content": "{lamdouar,weidi,az}@robots.ox.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.558 + ], + "angle": 0, + "content": "Not all camouflages are equally effective, as even a partially visible contour or a slight color difference can make the animal stand out and break its camouflage. In this paper, we address the question of what makes a camouflage successful, by proposing three scores for automatically assessing its effectiveness. In particular, we show that camouflage can be measured by the similarity between background and foreground features and boundary visibility. We use these camouflage scores to assess and compare all available camouflage datasets. We also incorporate the proposed camouflage score into a generative model as an auxiliary loss and show that effective camouflage images or videos can be synthesised in a scalable manner. The generated synthetic dataset is used to train a transformer-based model for segmenting camouflaged animals in videos. Experimentally, we demonstrate state-of-the-art camouflage breaking performance on the public MoCA-Mask benchmark." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.588, + 0.21, + 0.604 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Camouflage has long been a subject of interest and fascination for the scientific community, especially evolutionary biologists who consider it as an excellent example of species adaptation. In order to confuse predators or to hide from prey, and increase their chances of survival in their natural habitat, animals have developed numerous camouflage mechanisms, e.g., disruptive coloration and background matching. Some species have even evolved to develop an adaptive camouflage, e.g., an arctic fox loses its white fur to better match the brown grey of the new season's landscape. Perhaps the most dramatic camouflage adaptation is the cuttlefish; it changes its patterns dynamically and rapidly as it moves from one spot to the other, constantly adapting and improving its camouflage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "This search for optimal camouflage has inspired numerous works in the computer vision community, such as [12, 27], that tackled camouflage as a problem of optimal texture synthesis, making 3D objects non-detectable in a given scene. Others have addressed camouflage as a highly" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.268, + 0.63, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.345, + 0.572, + 0.355 + ], + "angle": 0, + "content": "a" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.268, + 0.76, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.344, + 0.7, + 0.356 + ], + "angle": 0, + "content": "b" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.268, + 0.892, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.345, + 0.832, + 0.355 + ], + "angle": 0, + "content": "C" + }, + { + "type": "text", + "bbox": [ + 0.634, + 0.363, + 0.753, + 0.381 + ], + "angle": 0, + "content": "score function" + }, + { + "type": "text", + "bbox": [ + 0.58, + 0.392, + 0.794, + 0.406 + ], + "angle": 0, + "content": "camouflageranking: \\(c > a > b\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.895, + 0.523 + ], + "angle": 0, + "content": "Figure 1: The three images depict the same animal, an Arctic fox, as it adapts its appearance to better blend with the changing landscape of the new season. While images \\( a \\) and \\( c \\) exhibit better background similarity than image \\( b \\), the fox boundary is more visible in image \\( a \\) than image \\( c \\). We assess the effectiveness of camouflage by measuring the degree of ambiguity it creates with respect to its background." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.895, + 0.764 + ], + "angle": 0, + "content": "challenging object segmentation task in [9, 15, 20, 23]. Efforts have been made in collecting large scale camouflage datasets [9, 20] with costly annotation. In fact, camouflaged animals often exhibit complex shapes and thin structures that add to the boundary ambiguity and make the annotation highly time-consuming. Fan et al. report up to \\(60\\mathrm{min}\\) per image to provide accurate pixel-level annotations for their dataset COD10K [9]. While another line of research turned to camouflage breaking in sequences by taking advantage of motion cues [1, 5, 18, 19, 40], the camouflage data scarcity is even more extreme for videos. Recently, Sim2Real [18, 39] training has shown to be very effective for motion segmentation. By training on the optical flows of synthetic videos, these models can generalise to real videos without suffering from the domain gaps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we start by asking the question: \"what are the properties of a successful camouflage?\" To answer this question, we investigate three scoring functions for quantifying the effectiveness of camouflage, namely, reconstruction fidelity score \\((S_{R_f})\\), boundary score \\((S_b)\\) and intra-image Fréchet score \\((d_{\\mathcal{F}})\\). These scores are later adopted for two critical roles, (i) they are used to assess the relevance of existing camouflage datasets and act as quality-indicator in collecting new camouflage data; (ii)" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.925, + 0.512, + 0.937 + ], + "angle": 0, + "content": "832" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "they can be used as a proxy loss for image and video inpainting/generation, where we establish a synthetic camouflage data generation pipeline with a Generative Adversarial Network (GAN), that can simultaneously generate high-quality camouflage examples and masks of the camouflaged animals. We further train a Transformer-based architecture on the generated camouflage video sequences, and demonstrate a performance boost over training on only the (small scale) real data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.473, + 0.442 + ], + "angle": 0, + "content": "In summary, we make the following contributions: First, we introduce three scoring functions to measure the effectiveness of a given camouflage in still images and videos. We use these camouflage scores to rank the images in existing datasets in terms of camouflage success, and also show that the rankings are largely in accordance with human-produced rankings. Second, we incorporate the camouflage score into a generative model, establishing a scalable pipeline for generating high-quality camouflage images or videos, along with the pixelwise segmentation mask of the camouflaged animals; Third, we show that a Transformer-based model trained on the synthetically generated data can achieve state-of-the-art performance on the MoCA-Mask video camouflaged animal segmentation benchmark." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.461, + 0.216, + 0.476 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.473, + 0.716 + ], + "angle": 0, + "content": "Camouflage Evaluation. Although there are no previous computational works that directly assess camouflage, as far as we know, there have been a number of human studies. Previous works proposed methods to evaluate camouflage by analysing the human viewers' eye movements [22, 37] or conducting subjective perceptual experiments [12, 27, 32]. In [12, 27], participants were asked to indicate the camouflaged object and their answers were analysed in terms of accuracy and time needed per image. Similarly, Skurowski et al. [32] asked human volunteers to rate CHAMELEON [32] images from 1 to 5 in terms of camouflage effectiveness. They produced a score after compensating for personal bias. We validate our proposed camouflage scoring functions by comparing our rankings to these human-based rankings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Motion Segmentation. The goal of this task is to partition the frames of a video sequence into background and independently moving objects. Early approaches focused on clustering motion patterns by grouping pixels with similar trajectories [3, 17, 21, 24, 25, 26, 31]. Another line of research tackled the problem by compensating for the background motion [1, 2, 19], via registering consecutive frames, or explicitly estimating camera motion. More recently, in [40], a Transformer-like architecture is used to reconstruct the input flows, and the segmentation masks are generated as a side product, while [6] decomposes the flow filed into regions by fitting affine transformations. The most" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "closely related to ours is [18, 39], where the authors adopt a Sim2Real strategy by training the model on optical flow computed from synthetic videos. These models can generalise towards real videos without fine-tuning. In this paper, we train a model on synthetic camouflage RGB sequences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.171, + 0.893, + 0.444 + ], + "angle": 0, + "content": "Frechet Inception Distance and its variants. Also known as Wasserstein-2 distance [38], Fréchet distance (FD) is a metric quantifying the difference between two probability distributions [10]. Lately, [13] introduced the Fréchet Inception Distance (FID) in the context of generative models. Under the assumption that real images and generated images follow Gaussian distributions, they compute the FD of the two distributions. Note that here, each image is represented by a vector obtained from the last pooling layer of InceptionV3 [35]. Following their steps, [30] adapted FID to the image re-targeting task and introduced the Single Image Fréchet Inception Distance (SIFID). Instead of comparing entire image datasets, they compare two images and consider the distributions of their feature vectors obtained from earlier layer of InceptionV3 [35]. Recently, [7] included FID in the training objective and introduced Fréchet-GAN. Our work takes inspiration from [7, 30] and investigates a FID of regions within the same image as an auxiliary loss." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.457, + 0.877, + 0.475 + ], + "angle": 0, + "content": "3. Measuring the effectiveness of camouflage" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.892, + 0.604 + ], + "angle": 0, + "content": "Assuming there exists a dataset, \\(\\mathcal{D} = \\{(I_1,m_1),\\ldots ,\\) \\((I_N,m_N)\\}\\) , where \\(I_{i}\\in \\mathbb{R}^{H\\times W\\times 3}\\) refers to the images, and \\(m_i\\in \\{0,1\\}^{H\\times W\\times 1}\\) denotes a binary segmentation mask of the camouflaged animal. Here, we aim to design a scoring function that takes the image and segmentation mask as input, and outputs a scalar value \\(S\\) that can quantify the effectiveness of camouflage, i.e., how successfully an animal blends into its background, termed as the camouflage score," + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.631, + 0.788, + 0.649 + ], + "angle": 0, + "content": "\\[\nS: (I _ {i}, m _ {i}) \\longmapsto s _ {i} \\in [ 0, 1 ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.795 + ], + "angle": 0, + "content": "specifically, for \\( i \\neq j \\) and \\( s_i < s_j \\), the animal in \\( I_j \\), indicated by the mask \\( m_j \\), is more concealed than the animal in \\( I_i \\), indicated by the mask \\( m_i \\). Having such a scoring function enables various applications, first, it enables to rank the images from one dataset, in terms of the effectiveness of camouflage, and generate camouflage-relevant statistics for the dataset; second, it can serve as the objective for low-level image generation tasks, for example, image inpainting, editing, etc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In the following sections, we investigate three such scoring functions, specifically, we exploit the animal's mask and quantify the key aspects that determine the camouflage success both perceptually (Sec. 3.1) and probabilistically (Sec. 3.2). Note that, we take into account the local aspect of camouflage, i.e., instead of processing the entire background region, we only focus on the immediate" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.925, + 0.511, + 0.937 + ], + "angle": 0, + "content": "833" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.092, + 0.177, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.195, + 0.134, + 0.204 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.092, + 0.277, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.143, + 0.277, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.195, + 0.238, + 0.208 + ], + "angle": 0, + "content": "\\(I_{bg}\\)" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.092, + 0.376, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.143, + 0.376, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.092, + 0.476, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.143, + 0.476, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.195, + 0.452, + 0.208 + ], + "angle": 0, + "content": "\\(\\Psi_{bg}(I_{fg})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.216, + 0.48, + 0.323 + ], + "angle": 0, + "content": "Figure 2: The reconstruction fidelity score \\( S_{R_f} \\) evaluates the similarities between the original foreground \\( I_{fg} \\) and its reconstruction from background features \\( \\Psi_{bg}(I_{fg}) \\). The top example shows a case where the animal exhibits different visual patterns (color) from its background \\( (S_{R_f} = 0.11) \\), while the bottom example shows a better background matching \\( (S_{R_f} = 0.82) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.386 + ], + "angle": 0, + "content": "surround of the animal. Hence, when referring to background, we mean the local background and only consider the cropped images centred around the object." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.395, + 0.332, + 0.411 + ], + "angle": 0, + "content": "3.1. Perceptual camouflage score" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.47, + 0.478 + ], + "angle": 0, + "content": "In this section, we define the camouflage score through a perceptual lens, specifically, we measure the foreground and background similarity (Sect. 3.1.1), along with the boundary visibility (Sect. 3.1.2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.49, + 0.33, + 0.505 + ], + "angle": 0, + "content": "3.1.1 Reconstruction fidelity score" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.469, + 0.604 + ], + "angle": 0, + "content": "We attempt to reconstruct the foreground region with patches from the background, and propose a score to quantitatively measure the discrepancy between the original image and its reconstruction. Intuitively, for successful camouflage, we should be able to reconstruct the foreground perfectly by copying patches from its close neighborhood." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.469, + 0.679 + ], + "angle": 0, + "content": "Formally, for a given image \\((I)\\) and segmentation mask \\((m)\\), subscript has been ignored for simplicity, we consider a trimap separation and define the foreground and background regions using morphological erosion and dilation of the mask," + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.692, + 0.469, + 0.709 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathrm {f g}}, I _ {\\mathrm {b g}} = m _ {\\mathrm {f g}} \\odot I, m _ {\\mathrm {b g}} \\odot I \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.411, + 0.735 + ], + "angle": 0, + "content": "where \\(m_{\\mathrm{fg}} = \\operatorname{erode}(m)\\) and \\(m_{\\mathrm{bg}} = 1 - \\operatorname{dilate}(m)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.469, + 0.87 + ], + "angle": 0, + "content": "To reconstruct the foreground region, we replace the foreground patch with the closest one in the background. Here a patch is a \\( n \\times n \\) pixel region (\\( n = 7 \\) in our case with an overlap of 3) and the patchwise similarity is computed by exploiting the low-level perceptual similarity, i.e., comparing corresponding RGB values. This reconstruction method is inspired by the texture generation and inpainting approaches, such as [8]. In practice, we implement it with fast approximate nearest neighbor search for efficiency." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The reconstruction fidelity score is computed by assessing the difference value between the foreground region and" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.092, + 0.592, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.143, + 0.592, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.195, + 0.551, + 0.205 + ], + "angle": 0, + "content": "I" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.092, + 0.688, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.143, + 0.688, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.195, + 0.668, + 0.206 + ], + "angle": 0, + "content": "Trimap" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.092, + 0.785, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.143, + 0.786, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.195, + 0.741, + 0.204 + ], + "angle": 0, + "content": "C" + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.099, + 0.871, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.145, + 0.871, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.833, + 0.195, + 0.852, + 0.207 + ], + "angle": 0, + "content": "\\(C_{gt}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.217, + 0.89, + 0.324 + ], + "angle": 0, + "content": "Figure 3: Trimap regions and animal contours. From left to right: Input image; Trimap partition (foreground region in yellow, background in purple, and boundary region in green); Contours \\( C \\) computed on the original image; and ground truth Contours \\( C_{gt} \\). The top example shows less contour agreement along the boundary region (\\( S_b = 0.72 \\)) than the bottom example (\\( S_b = 0.40 \\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.893, + 0.385 + ], + "angle": 0, + "content": "its reconstruction, specifically, we count the number of foreground pixels that have been successfully reconstructed from the background:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.396, + 0.892, + 0.435 + ], + "angle": 0, + "content": "\\[\nS _ {R _ {f}} (I, m) = \\frac {1}{N _ {\\mathrm {f g}}} \\sum_ {(i, j) \\in I _ {\\mathrm {f g}}} R _ {f} (i, j) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.438, + 0.891, + 0.479 + ], + "angle": 0, + "content": "\\[\nR _ {f} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} | | I _ {\\mathrm {f g}} - \\Psi_ {I _ {\\mathrm {b g}}} \\left(I _ {\\mathrm {f g}}\\right) | | _ {2} < \\lambda | | I _ {\\mathrm {f g}} | | _ {2} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.611 + ], + "angle": 0, + "content": "\\(\\Psi_{I_{bg}}(.)\\) denotes the reconstruction operation, \\(N_{\\mathrm{fg}} = |m_{\\mathrm{fg}}|\\) is the total number of pixels in the foreground region, and \\(\\lambda\\) is a threshold (\\(\\lambda = 0.2\\) in our case). A higher \\(S_{R_f}\\) score, means that the animal's visual attributes are well represented in the background. Conversely, a low \\(S_{R_f}\\) indicates that the animal's appearance is composed of unique features that makes it standout from its surrounding. In Fig. 2, we present examples of camouflage evaluation by reconstruction fidelity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.623, + 0.727, + 0.638 + ], + "angle": 0, + "content": "3.1.2 Boundary visibility score" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.892, + 0.783 + ], + "angle": 0, + "content": "With the trimap representation, we can easily extract the boundary region, \\( I_{\\mathrm{b}} = m_{\\mathrm{b}} \\odot I \\), where \\( m_{\\mathrm{b}} = (1 - m_{\\mathrm{bg}}) - m_{\\mathrm{fg}} \\). Here, we aim to measure the animal's boundary properties, e.g., contour visibility, as they can also provide visually evident cues for camouflage breaking. In particular, we adopt an off-the-shelf contour extraction method [29], and run it on the original image and foreground mask, to generate the images' contour (\\( C \\)) and ground truth animal's contour (\\( C_{\\mathrm{gt}} \\)), as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.892, + 0.828 + ], + "angle": 0, + "content": "We express the agreement between predicted contours \\(C\\) and ground truth contours \\(C_{\\mathrm{gt}}\\) over the boundary region with the F1 metric:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.842, + 0.891, + 0.859 + ], + "angle": 0, + "content": "\\[\nS _ {b} (I, m) = 1 - \\mathrm {F} 1 \\left(m _ {b} \\odot C _ {\\mathrm {g t}}, m _ {b} \\odot C\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "This score penalises the boundary pixels that are predicted as contour in both \\(C\\) and \\(C_{\\mathrm{gt}}\\). While we do not expect a" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.925, + 0.511, + 0.937 + ], + "angle": 0, + "content": "834" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "perfect contour agreement for visible edges, we consider \\( S_{b} \\) as a reasonable approximation, given the shape and size of the boundary region as a thin margin centred around the animal. If a boundary pixel is predicted as a contour in \\( C_{\\mathrm{gt}} \\) but not in \\( C \\), this means that the animal's edge is not visible in the original image. If a pixel is predicted as a contour in \\( C \\) but not \\( C_{\\mathrm{gt}} \\), this indicates the presence of distracting elements, such as complex texture patterns on the foreground animal, which also affects the visibility of the animal's contour and therefore improves its camouflage. A perfectly camouflaged animal would have little contour agreement and \\( S_{b} = 1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.472, + 0.32 + ], + "angle": 0, + "content": "To get the final perceptual camouflage score, we linearly combine both reconstruction fidelity score and boundary visibility score:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.334, + 0.47, + 0.352 + ], + "angle": 0, + "content": "\\[\nS _ {\\alpha} = (1 - \\alpha) S _ {R _ {f}} + \\alpha S _ {b} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.365, + 0.47, + 0.396 + ], + "angle": 0, + "content": "We will describe our method for setting the weighting parameter \\(\\alpha\\) in the Experiment section (Sec. 6)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.408, + 0.341, + 0.423 + ], + "angle": 0, + "content": "3.2. Probabilistic scoring function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.47, + 0.538 + ], + "angle": 0, + "content": "In addition to using the low-level RGB representation, in this section, we consider the pixelwise representation in the feature space, and propose a differentiable metric that compares the probabilistic distribution between the foreground and background regions, which acts as a proxy for the score \\( S_{\\alpha} \\) described in previous section, and can be used directly as a differentiable loss function in image generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.47, + 0.585 + ], + "angle": 0, + "content": "We consider the Intra-Image Fréchet Inception Distance, specifically, we compute the feature maps for the foreground and background image regions:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.599, + 0.47, + 0.617 + ], + "angle": 0, + "content": "\\[\nf _ {f g}, f _ {b g} = \\Phi_ {\\mathrm {v}} \\left(I _ {f g}\\right), \\Phi_ {\\mathrm {v}} \\left(I _ {b g}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Here \\(\\Phi_{\\mathrm{v}}(\\cdot)\\) denotes a pre-trained Inception network [34, 35], truncated at early layers. We take inspiration from [30], and adapt the single image Fréchet distance to measure the deviation between feature distributions of different regions within the same image. We adopt the Fréchet hypothesis with respect to our regions, i.e., the features of the foreground and background follow multivariate Gaussian distributions: \\(f_{\\mathrm{fg}} \\sim \\mathcal{N}(\\mu_{\\mathrm{fg}}, \\Sigma_{\\mathrm{fg}})\\) and \\(f_{\\mathrm{bg}} \\sim \\mathcal{N}(\\mu_{\\mathrm{bg}}, \\Sigma_{\\mathrm{bg}})\\). The intra-image Fréchet distance can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.793, + 0.479, + 0.813 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathcal {F}} ^ {2} (I, m) = \\left\\| \\mu_ {\\mathrm {f g}} - \\mu_ {\\mathrm {b g}} \\right\\| _ {2} ^ {2} + T r \\left(\\Sigma_ {\\mathrm {f g}} + \\Sigma_ {\\mathrm {b g}} - 2 \\left(\\Sigma_ {\\mathrm {f g}} \\Sigma_ {\\mathrm {b g}}\\right) ^ {1 / 2}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Note that, when \\((I,m)\\) is the output from a generative model, \\(d_{\\mathcal{F}}^2\\) is differentiable with respect to its parameters, and can therefore be used as an auxiliary loss for optimising the image generation procedure, i.e., generate effective camouflage examples." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.857, + 0.108 + ], + "angle": 0, + "content": "4. Generating camouflage video sequences" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.894, + 0.222 + ], + "angle": 0, + "content": "In this section, we propose a scalable pipeline for generating images with concealed animals. In particular, we incorporate the differentiable intra-image Fréchet distance in image generation, explicitly optimising the camouflage score of the image. We first detail the static image generation framework, then describe how we use the images it generates to create camouflage video sequences." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.231, + 0.766, + 0.246 + ], + "angle": 0, + "content": "4.1. Camouflage image generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.254, + 0.892, + 0.344 + ], + "angle": 0, + "content": "In order to generate the image with a camouflaged animal, and its corresponding segmentation mask, we adopt a Generative Adversarial Network (GAN), where a generator is fed with a latent vector \\( z \\sim \\mathcal{N}(0,1) \\) and learns to predict the pair of realistic image and segmentation mask, such that a discriminator cannot differentiate it from real samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.894, + 0.435 + ], + "angle": 0, + "content": "Let \\((x_{i},m_{i})\\) denote an image and segmentation pair sampled from the training dataset, a subset of COD10K [9] is used here. We adopt the conventional label notation, i.e. \\(y = 1\\) for (real) examples sampled from the training dataset and \\(y = 0\\) for (fake) examples from the generator, and consider the generative adversarial loss functions [11]:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.445, + 0.912, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} = E _ {\\left(x _ {i}, m _ {i}\\right)} \\left[ - \\log \\left(D \\left(x _ {i}, m _ {i}\\right)\\right) \\right] + E _ {z} \\left[ - \\log \\left(1 - D (G (z))\\right) \\right]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.465, + 0.688, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G} = E _ {z} \\left[ - \\log (D (G (z))) \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.597 + ], + "angle": 0, + "content": "To enforce coherence between images and their segmentation masks, we feed the discriminator with additional fake pairs, consisting of real images coupled with unpaired real masks. This can be formulated as an additional coherence loss term that minimises the probability of assigning incorrect labels, i.e. \\( y = 1 \\), to \\( (x_{i},m_{j}) \\) sampled from the training set such that \\( i\\neq j \\):" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.607, + 0.892, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C o h}} = E _ {\\left(x _ {i}, m _ {j}\\right) i \\neq j} [ - \\log (1 - D \\left(x _ {i}, m _ {j}\\right)) ] \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.892, + 0.696 + ], + "angle": 0, + "content": "To increase the camouflage effectiveness in the generated examples, we adopt the intra-image Fréchet distance in the generator loss, as an auxiliary loss term, and train our camouflage image generator with the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.705, + 0.892, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\tilde {D}} = \\mathcal {L} _ {D} + \\mathcal {L} _ {\\mathrm {C o h}} \\quad \\mathcal {L} _ {\\tilde {G}} = \\mathcal {L} _ {G} + \\beta d _ {\\mathcal {F}} ^ {2} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.763 + ], + "angle": 0, + "content": "We present an overview of our data generation framework in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.76, + 0.788 + ], + "angle": 0, + "content": "4.2. Camouflage video generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given a camouflage image and corresponding segmentation mask, generated using the method above, we can create video sequences by applying different motions to the foreground and background, in a similar manner to [18]. Specifically, we first inpaint the backgrounds with an off-the-shelf model proposed by Suvorov et al. [33] and overlay the synthetic animal (extracted from the original generated image" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.925, + 0.511, + 0.937 + ], + "angle": 0, + "content": "835" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.09, + 0.256, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.089, + 0.38, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.089, + 0.853, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.267, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Figure 4: Our framework consists of a synthetic camouflage data generation pipeline (left), where we train a generator \\( G \\) in a GAN setting, to create camouflage images and masks, while encouraging camouflage effectiveness via minimising \\( \\mathcal{L}_{\\mathcal{F}} \\). The generated samples are then transformed into synthetic video sequences (middle), following our method presented in Sec. 4.2. The transformer-based motion segmentation architecture (right), for camouflage breaking in videos, is trained on the synthetic video sequences. The architecture is described in Sec. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.474, + 0.518 + ], + "angle": 0, + "content": "by using the mask) at a random location within the image in the first frame, then following a random translational trajectory in the following frames. We incorporate a different translational motion for the background and include random sub-sequences where the foreground and background undergo the same motion to simulate momentarily static objects. As we train our generator with intra-image Fréchet loss \\((\\mathcal{L}_{\\mathcal{F}})\\), the generated images exhibit strong similarities between the foreground and the background, and even if the object is placed at a different location from the original, we expect it to remain concealed within its surrounding." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.546, + 0.355, + 0.563 + ], + "angle": 0, + "content": "5. Learning to break camouflage" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.47, + 0.846 + ], + "angle": 0, + "content": "In this section, we train a transformer-based architecture on the synthetic dataset, and demonstrate its effectiveness for breaking the camouflage in real videos, for example, MoCA [19]. We build on two previous architectures, namely, the motion segmentation from [18] and the Search Identification Network from [9]. While the first architecture processes sequences of optical flow, the second takes single RGB images as inputs and treats them separately. Our proposed model, shown in Fig. 4, takes both the optical flow and RGB frame sequences as input. The flow is computed with RAFT [36] from the RGB sequence, and processed with a motion encoder, followed by a transformer encoder. The RGB sequence is processed with the appearance encoder from [9], pre-trained for framewise camouflage segmentation. Then both streams are aggregated as inputs to a Mask2Former-like [4] transformer decoder and pixel decoder to produce the high resolution segmentation mask from the motion stream." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Motion encoder. To encode the motion cues, we use a light-weight convNet that takes as input a sequence of optical flows, \\(\\mathbf{I}_{\\mathbf{m}} = \\{I_{m_1}, I_{m_2},.., I_{m_t}\\} \\in \\mathcal{R}^{t \\times c_0 \\times h \\times w}\\) and out" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.352, + 0.644, + 0.366 + ], + "angle": 0, + "content": "puts motion features:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.383, + 0.811, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\left\\{f _ {m _ {1}}, f _ {m _ {2}}, \\dots , f _ {m _ {t}} \\right\\} = \\Phi_ {\\text {m o t i o n}} \\left(\\mathbf {I} _ {\\mathbf {m}}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.417, + 0.813, + 0.432 + ], + "angle": 0, + "content": "where each flow frame is separately embedded." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.892, + 0.543 + ], + "angle": 0, + "content": "Motion transformer encoder. A transformer encoder takes as input the motion features, concatenated along the sequence, together with learned spatial and temporal positional encodings \\((\\mathrm{Pos}_s,\\mathrm{Pos}_t)\\), as indicated in Fig 4. \\(\\mathrm{Pos}_t\\) specifies the frame, and \\(\\mathrm{Pos}_s\\) specifies the position in the spatial feature map output of the motion encoder. The output of the transformer is a set of enriched motion features." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.55, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Appearance encoder. Here, we adopt a SINet-v2 [9] architecture, that encodes the RGB sequence, \\(\\mathbf{I}_{\\mathbf{a}} = \\{I_{a_1}, I_{a_2},.., I_{a_t}\\} \\in \\mathcal{R}^{t \\times c_1 \\times h \\times w}\\) into appearance features:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.609, + 0.794, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\left\\{f _ {a _ {1}}, f _ {a _ {2}}, \\dots , f _ {a _ {t}} \\right\\} = \\Phi_ {\\text {a p p}} \\left(\\mathbf {I} _ {\\mathbf {a}}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.643, + 0.882, + 0.658 + ], + "angle": 0, + "content": "Again, each RGB frame is processed separately by SINet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Transformer decoder. A transformer-based decoder takes the output of the motion transformer encoder together with the appearance features and a learnable query for the mask embedding. In a similar manner to Mask2Former [4], the query attends to multiple resolutions of the motion features concatenated with the appearance features and produces a mask embedding for the moving object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.775, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Pixel Decoder. Similarly to the pixel decoder in Mask2Former, a light-weight convNet decoder is used with skip-connections to recover high-resolution segmentation masks \\(\\{\\hat{m}_1,\\hat{m}_2,\\dots ,\\hat{m}_t\\}\\) from the motion features and the mask embedding. This is shown as the blue box in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Training objective. We train the motion segmentation model on our synthetic video sequences using the binary cross-entropy loss \\(\\mathcal{L}_{BCE}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.926, + 0.512, + 0.937 + ], + "angle": 0, + "content": "836" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.21, + 0.108 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.117, + 0.47, + 0.193 + ], + "angle": 0, + "content": "In this section, we start by introducing the datasets involved in this paper, followed by the implementation details. Our experiments present a thorough analysis of the proposed camouflage scores and demonstrate their effectiveness in our training framework." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.208, + 0.18, + 0.223 + ], + "angle": 0, + "content": "6.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.233, + 0.469, + 0.293 + ], + "angle": 0, + "content": "Here we describe the publicly available camouflage datasets that we included in our experiments, as shown in Fig. 5, as well as the synthetic camouflage datasets that we generated." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.297, + 0.469, + 0.358 + ], + "angle": 0, + "content": "CHAMELEON [32] is one of the first camouflage image datasets. It contains 76 images collected using Google image search with the 'camouflaged animal' keyword and include ground truth manual annotations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.362, + 0.469, + 0.437 + ], + "angle": 0, + "content": "CAMO [20]. The Camouflaged Object dataset consists of 2500 images collected from the internet, of which 1250 images (1000 training sub-set and 250 testing sub-set) contain at least one camouflage instance from 8 categories with manual pixelwise annotations provided." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.442, + 0.469, + 0.532 + ], + "angle": 0, + "content": "COD10K [9]. COD10K contains 10,000 images collected from photography websites of which 5066 depict camouflaged animals (3040 training sub-set and 2026 testing subset), organised into 10 classes, and 78 sub-classes (69 camouflaged). Note that, in our camouflage evaluation experiments we only used the camouflage subset of this dataset." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.537, + 0.469, + 0.567 + ], + "angle": 0, + "content": "Camouflaged Animals [1]. This dataset consists of 9 video sequences of camouflage animals from 6 categories." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.571, + 0.469, + 0.676 + ], + "angle": 0, + "content": "MoCA [19] is the first large-scale camouflage video dataset. It contains 141 video sequences (37K frames) totalling 67 animal categories. Recently, other works have curated this dataset by selecting only the videos with more prominent motion (locomotion) [40], and others have provided dense pixel annotation in MoCA-Mask [5]. We use the latter version in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.681, + 0.469, + 0.77 + ], + "angle": 0, + "content": "Camouflaged cuboids [12, 27]. This dataset was created for texture synthesis and consists of multiple-view scenes, where cuboids were placed at a predefined location then synthetically camouflaged. In our evaluation, we consider the 4-views generated textures [12] from 36 scenes as well as the cuboids masks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.776, + 0.469, + 0.837 + ], + "angle": 0, + "content": "Synthetic Camouflage Images. Using the method described in Sec. 4.1, we generate a synthetic camouflage dataset, of \\(5K\\) frames, by discriminating against real camouflage image and annotation masks from COD10K." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Synthetic Camouflage Videos. We generate \\(1K\\) sequences of 30 frames each, incorporating static sequences using the framework described in Sec. 4.2. We split these into 800 sequences for training and 200 for testing." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.096, + 0.889, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.731, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Figure 5: Randomly sampled examples from the camouflage datasets included in our work. For the video datasets, Camouflaged Animals and MoCA-Mask, we show an example sequence. For the multiple-view dataset Camouflaged cuboids (bottom), we show example views from a scene with the 4-view texture synthesis method from [12]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.715, + 0.848 + ], + "angle": 0, + "content": "6.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In our experiments, for a given camouflage example, the kernels for the morphological operations are selected from a range of values [1,10], so that the resulting annotation mask" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.925, + 0.511, + 0.936 + ], + "angle": 0, + "content": "837" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "is reduced by \\(20\\%\\) for erosion and extended by \\(20\\%\\) for dilation. This allows the erosion/dilation to be adapted to the size of the camouflaged animal, while always keeping a reasonably large region for addressing the boundary effects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.163, + 0.47, + 0.283 + ], + "angle": 0, + "content": "For our synthetic camouflage image generation, we adopt a Style-GAN architecture [41] and train it on the camouflage images from COD10K to generate \\(256 \\times 256\\) images and masks pairs. We use \\(\\beta = 0.1\\) to weight our intra-image Fréchet auxiliary loss. For numerical stability, we adopt the Newton Schultz iterations when calculating the matrix square root term in Intra-Image Fréchet Inception Distance, with \\(T = 100\\) iterations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.47, + 0.379 + ], + "angle": 0, + "content": "When generating the image sequences, we compute optical flows with RAFT [36], and train our moving camouflaged animal segmentation model. We use batches of size 2 and sequences of size 5. We first train on our synthetically generated dataset with a learning rate of \\(5 \\times 10^{-4}\\) for 500 iterations then fine-tune on the training subset of MoCA-Mask." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.388, + 0.172, + 0.403 + ], + "angle": 0, + "content": "6.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.47, + 0.459 + ], + "angle": 0, + "content": "This section presents qualitative and quantitative results, to demonstrate the effectiveness of our score functions and the benefit of including \\( d_{\\mathcal{F}}^2 \\) in the training loop." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.467, + 0.403, + 0.482 + ], + "angle": 0, + "content": "6.3.1 Evaluation on camouflage effectiveness" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.471, + 0.884 + ], + "angle": 0, + "content": "Ranking camouflage images in COD10K. We can rank the images based on the effectiveness of the animal's camouflage, with our proposed scoring functions. In the left part of Fig. 6, we show the four best scored images from the large scale COD10K dataset. We can make the following observations: (i) the best examples with respect to \\( S_{R_f} \\) (top), which favours the background matching of the animals without taking into account the boundary region, all exhibit visible boundaries, with the exception of image (a). This is especially noticeable in the rabbit example (d) along the ears and shadow regions; (ii) the best four examples with respect to the boundary score \\( S_b \\) are all from the caterpillar subclass, mostly the baron type. These insects have thin, and transparent boundaries, that makes them the perfect candidates for this score. However, with the exception of (a) and (c), the animals still stands out as they exhibit colors and patterns that are not present within their background; (iii) the \\( S_\\alpha \\) combining both scores shows highly effective camouflage examples. In contrast, the right part of Fig. 6 shows the lowest scored images for each approach. For \\( S_{R_f} \\), we find examples with low background matching in the first row, with two ant examples, i.e., (e), (f). While the boundary score penalises examples with high contrasts between the animal and its background, that results in more visible contours. This is also the case for the lowest \\( S_\\alpha \\) score examples." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Dataset level camouflage comparison. We compute the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "camouflage score for all the camouflage image and video datasets and report the results in Tab. 1. For the image datasets, the dataset level score is computed as the mean of the image scores; for the video dataset the score is first computed per frame, and per-video average computed, then the dataset level score is the mean of the averages. Our experiments show that MoCA-Mask [5, 19] contains the most successful camouflages according to our scoring. While COD10K [9] subsets are balanced in terms of camouflage effectiveness, we find that, for CAMO [20], the test subset yields higher scores than the training subset and therefore contains more challenging examples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.892, + 0.426 + ], + "angle": 0, + "content": "We note that Camouflaged cuboids dataset yields higher \\( S_{R_f} \\) and \\( S_{\\alpha} \\) scores compared to our synthetic datasets. This is due to the fact that the model used in [12, 27] is only trained to produce optimal texture for a predefined region (cuboid) in a particular location of a given scene. However, our model learns to output a new image with a randomly located camouflaged object of random shape, thus offering more scalability and diversity compared to the cuboids dataset which cannot be used to train a model for breaking real camouflage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Comparison to human-produced rankings. We compare our rankings to the human scoring system based on the rating of CHAMELEON [32] and the time-to-find for Camouflaged cuboids [12]. To search for the optimal \\(\\alpha\\) parameter in Equation (5), we select 15 images from CHAMELEON and compare their \\(S_{\\alpha}\\) ranking with ground truth using kendl-\\(\\tau\\) metric [16]. For \\(\\alpha = 0.35\\), we obtain kendl-\\(\\tau = 0.51\\) on this validation set that we excluded from our test reported in Tab. 2. We can draw the following observations: (i) we find that the boundary score \\(S_{b}\\) produces more agreement with the ground truth ranking than \\(S_{R_{f}}\\), suggesting that human observers tend to pay more attention to contour visibility than background matching; (ii) while comparing to the ranking from Camouflaged cuboids, we found negative correlations for \\(S_{R_{f}}\\) and \\(d_{\\mathcal{F}}\\), we conjecture that this may be due to the nature of the dataset in [12], i.e., all the textured cuboids are synthetically generated with very high background matching; (iii) for both experiments, we obtain the strongest agreement with the \\(S_{\\alpha}\\) score combining both background similarity and boundary blending." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Further analysis of the synthetic datasets. We adopt the FID [13] and IS [30] metrics in Table 4. These metrics assess: (i) the similarity between the set of synthetic images output by the generator and the set of real images used to train it (FID); and (ii) the clarity of the object and the diversity of images in terms of classes (IS). While they are not intended for measuring camouflage success, which is the focus of this work, IS could somewhat be (inversely) linked to \\( S_{b} \\) as the object clarity and boundary visibility are close. Note that the object clarity (IS) is decreased with \\( \\mathcal{L}_{\\mathcal{F}} \\) which is the intended purpose of improving camouflage, however," + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.925, + 0.511, + 0.937 + ], + "angle": 0, + "content": "838" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.126, + 0.092, + 0.85, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.895, + 0.374 + ], + "angle": 0, + "content": "Figure 6: Left: Top-4 scored examples from COD10K for \\( S_{Rf}(top) \\), \\( S_b(midle) \\) and \\( S_{\\alpha} \\) (bottom). Right: Lowest scored examples from COD10K. For each example, we show the obtained score and the corresponding ground truth mask." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.383, + 0.523, + 0.562 + ], + "angle": 0, + "content": "
DatasetsData typeSRf↑Sb↑Sα↑d2F↓
CHAMELEON [32]Image0.6940.4450.6070.70
CAMO Train [20]Image0.6720.4510.5951.01
CAMO Test [20]Image0.6830.4700.6081.00
COD10K Train [9]Image0.6550.4330.5770.90
COD10K Test [9]Image0.6570.4310.5780.90
Camouflaged Animals [1]Video0.6740.5360.6261.60
MoCA-Mask Train [5, 19]Video0.8500.4430.7071.14
MoCA-Mask Test [5, 19]Video0.7330.4640.6392.51
Camouflaged cuboids[12, 27]Multi-view0.8940.4330.7336.2
Syn. Camouflage w.o. L_FImage0.6080.4320.5461.36
Syn. Camouflage w. L_FImage0.6790.4470.5981.13
Syn. Camouflage VideoVideo0.6580.4300.5781.18
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.565, + 0.528, + 0.673 + ], + "angle": 0, + "content": "Table 1: Results of the proposed scores on natural camouflage datasets (top) and synthetically generated camouflage datasets (bottom). We report the mean scores for the single image datasets. For the video and multiple view datasets, we compute the mean per sequence and scene respectively. Syn. Camouflage w. \\(\\mathcal{L}_{\\mathcal{F}}\\) refers to the synthetic image dataset that we generated while minimising \\(\\mathcal{L}_{\\mathcal{F}}\\) and Syn. Camouflage Video its sequence version." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.47, + 0.713 + ], + "angle": 0, + "content": "this effect is not maintained in the sequence generation as the animal changes location within its background." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.714, + 0.477, + 0.772 + ], + "angle": 0, + "content": "
datasetFID↓IS↑
Syn. Camouflage w.o. L_F75.013.49
Syn. Camouflage w. L_F76.343.13
Syn. Camouflage Video86.423.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.774, + 0.446, + 0.789 + ], + "angle": 0, + "content": "Table 4: Overall quality of our synthetic datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.792, + 0.245, + 0.806 + ], + "angle": 0, + "content": "6.3.2 Model ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "On the effectiveness of \\(\\mathcal{L}_{\\mathcal{F}}\\). Fig. 7 shows the generated samples from the GAN pipeline, trained with and without the intra-image Fréchet Distance. Our generator is able to produce realistic object masks of complex shapes and thin structures similar to those usually encountered in the camouflage datasets. The produced images exhibit rich back" + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.383, + 0.895, + 0.46 + ], + "angle": 0, + "content": "
ScoresCHAMELEON [32]Camouflaged cuboids[12, 27]
SRf0.01-0.07
Sb0.030.42
Sα0.420.41
d2F0.10-0.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.534, + 0.464, + 0.889, + 0.494 + ], + "angle": 0, + "content": "Table 2: Kendall- \\(\\tau\\) metric between rankings produced via our scores and human scoring ground truth." + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.498, + 0.887, + 0.609 + ], + "angle": 0, + "content": "
ModelTraining datasetAppearanceEncoderTransformerEncoderTransformerAggregationmIoU↑F↑E↑MAE↓
AS14.520.457.39.4
BMM15.320.657.35.1
CS+MM16.021.859.84.3
DS+MM✓20.522.659.83.8
ES+MM✓✓23.023.557.01.6
FS+MM✓✓22.723.258.83.6
GMM✓✓✓23.424.762.02.4
HS+MM✓✓✓30.834.377.01.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.533, + 0.61, + 0.889, + 0.672 + ], + "angle": 0, + "content": "Table 3: Ablation study on the different components of the motion segmentation architecture and training datasets. We evaluate on MoCA-Mask for all experiments. (S: our synthetic dataset, MM: MoCA-Mask)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.892, + 0.806 + ], + "angle": 0, + "content": "grounds with realistic nature-like elements mimicking rocks and coral structures present in the real camouflage data. Adding the \\(\\mathcal{L}_{\\mathcal{F}}\\) loss results in generating images with better blending in the background, both qualitatively and quantitatively, as shown by the \\(S_{\\alpha}\\) computed for both datasets generated with and without the intra-image Fréchet Distance in Tab. 1. We refer the reader to the supplementary material for more examples of generated samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.902 + ], + "angle": 0, + "content": "On the architecture design and gain from the synthetic dataset. We present in Tab. 3 an ablation study detailing the gain from the main components of our architecture, as well as the benefit of first training on our synthetic dataset then fine-tuning on MoCA-Mask (S+MM), for model H, as opposed to only training on MoCA-Mask (MM), for G." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.925, + 0.512, + 0.937 + ], + "angle": 0, + "content": "839" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.85, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.212, + 0.88, + 0.227 + ], + "angle": 0, + "content": "Figure 7: Generated camouflage images and masks: For the examples on the right, the generator was trained with \\(\\mathcal{L}_{\\mathcal{F}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.065, + 0.235, + 0.52, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.4, + 0.496, + 0.446 + ], + "angle": 0, + "content": "Figure 8: Qualitative results on MoCA-Mask. From top to bottom: Appearance sequence \\( I_{a} \\), flow sequence \\( I_{m} \\), predicted segmentations and ground truth segmentations." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.239, + 0.89, + 0.393 + ], + "angle": 0, + "content": "
ModelRGBMotionmIoU↑F↑E↑MAE↓
SINet [9]✓20.223.169.92.8
SINet-V2 [9]✓18.020.464.23.1
SegMaR [14]✓12.222.580.34.5
ZoomNet [28]✓18.828.770.82.5
SLTNet [5]✓✓27.231.175.92.7
MG [40]✓12.716.856.16.7
Ours-flow✓17.821.560.73.7
Ours✓✓30.834.377.01.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.4, + 0.893, + 0.445 + ], + "angle": 0, + "content": "Table 5: Results on MoCA-Mask. Ours-flow refers to the optical flow only version of our architecture, excluding the appearance stream." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.459, + 0.254, + 0.473 + ], + "angle": 0, + "content": "6.3.3 Results on MoCA" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.651 + ], + "angle": 0, + "content": "We train our model on the synthetically generated camouflage images and fine-tune on the training set of MoCA-Mask. Tab. 5 presents our camouflage object segmentation results on the test set of MoCA-Mask. Ours refers to our proposed method with SINet-V2 as the appearance encoder and Ours-flow refers to the optical flow only version of our architecture, excluding the appearance stream. Our main model outperforms RGB and Motion-based methods on MoCA-Mask. Fig. 8 presents qualitative results of our segmentation model. Note that our model is robust to degraded optical flow and static animals." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.664, + 0.204, + 0.679 + ], + "angle": 0, + "content": "6.4. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.469, + 0.779 + ], + "angle": 0, + "content": "While our image generation method encourages background matching through the \\(\\mathcal{L}_{\\mathcal{F}}\\) loss term minimization, the sequence generating is not guaranteed to maintain object concealment. A possible solution could be to use the proposed \\(S_{\\alpha}\\) camouflage score to curate the generated sequence dataset and filter out the most visible objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Our proposed camouflage scores use the ground truth annotation and analyse the different regions it defines. We found that our scoring system penalises specific cases of camouflage by occlusion, where elements from the background are partially occluding the animal and helping improve its camouflage. For instance, the grass in example (g) from the last row of Fig. 6 is not treated as part of the animal and therefore not considered in our background similarity" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.459, + 0.892, + 0.503 + ], + "angle": 0, + "content": "assessment. One can argue that this is due to the ambiguity in the provided annotations and for such cases, extra amodal annotation should be also considered." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.52, + 0.619, + 0.536 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.7 + ], + "angle": 0, + "content": "We present three score functions for computationally assessing the effectiveness of camouflage in images and videos. By evaluating the similarity with background and the boundary visibility, our combined perceptual score is strongly correlated with human perceptual ranking systems on two different datasets. We demonstrate that training a generative model with our differentiable camouflage function improves the effectiveness of generated camouflage examples and can be used to generate challenging synthetic camouflage datasets to train models to break camouflage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.647, + 0.727 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We are grateful to Przemysław Skurowski for providing the human perception study data on the CHAMELEON dataset and Andrew Owens, Rui Guo and Oscar de Lima for providing the data for the Camouflaged cuboids dataset. We thank Tengda Han for fruitful discussions. This research is supported by the UK EPSRC funded CDT in Autonomous Intelligent Machines and Systems (AIMS), the EPSRC Programme Grant VisualAI EP/T028572/1, a Schlumberger Studentship, and a Royal Society Research Professorship. WX is supported by the National Key R&D Program of China (No. 2022ZD0161400)." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.926, + 0.512, + 0.937 + ], + "angle": 0, + "content": "840" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Pia Bideau and Erik Learned-Miller. It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In ECCV, pages 433-449, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.468, + 0.197 + ], + "angle": 0, + "content": "[2] Pia Bideau, Rakesh R Menon, and Erik Learned-Miller. Moa-net: self-supervised motion segmentation. In Proc. ECCV Workshop, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.197, + 0.468, + 0.237 + ], + "angle": 0, + "content": "[3] Thomas Brox and Jitendra Malik. Object segmentation by long term analysis of point trajectories. In European conference on computer vision, pages 282-295, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.237, + 0.468, + 0.291 + ], + "angle": 0, + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.468, + 0.345 + ], + "angle": 0, + "content": "[5] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.468, + 0.399 + ], + "angle": 0, + "content": "[6] Subhabrata Choudhury, Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Guess What Moves: Unsupervised Video and Image Segmentation by Anticipating Motion. In Proc. BMVC, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.399, + 0.468, + 0.453 + ], + "angle": 0, + "content": "[7] Khoa D Doan, Saurav Manchanda, Fengjiao Wang, Sathiya Keerthi, Avradeep Bhowmik, and Chandan K Reddy. Image generation via minimizing frechet distance in discriminator feature space. arXiv preprint arXiv:2003.11774, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[8] A. Efros and T. Leung. Texture synthesis by non-parametric sampling. In Proc. ICCV, pages 1039-1046, Sep 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.468, + 0.52 + ], + "angle": 0, + "content": "[9] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.52, + 0.468, + 0.56 + ], + "angle": 0, + "content": "[10] Maurice Fréchet. Sur la distance de deux lois de probabilité. Comptes Rendus Hebdomadaires des Seances de L'Académie des Sciences, pages 689-692, 1957." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.56, + 0.468, + 0.613 + ], + "angle": 0, + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, pages 2672-2680, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.613, + 0.468, + 0.654 + ], + "angle": 0, + "content": "[12] Rui Guo, Jasmine Collins, Oscar de Lima, and Andrew Owens. Ganmouflage: 3d object nondetection with texture fields. In arXiv preprint arXiv:2201.07202, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.654, + 0.468, + 0.707 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.708, + 0.468, + 0.761 + ], + "angle": 0, + "content": "[14] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In CVPR, pages 4713-4722, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.761, + 0.468, + 0.815 + ], + "angle": 0, + "content": "[15] Nobukatsu Kajiura, Hong Liu, and Shin'ichi Satoh. Improving camouflaged object detection with the uncertainty of pseudo-edge labels. In ACM Multimedia Asia, pages 1-7. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.816, + 0.468, + 0.842 + ], + "angle": 0, + "content": "[16] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1/2):81-93, 1938." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.843, + 0.468, + 0.883 + ], + "angle": 0, + "content": "[17] Margret Keuper, Bjoern Andres, and Thomas Brox. Motion trajectory segmentation via minimum cost multicuts. In ICCV, pages 3271-3279, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.883, + 0.468, + 0.897 + ], + "angle": 0, + "content": "[18] Hala Lamdouar, Weidi Xie, and Andrew Zisserman. Seg-" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.468, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.874, + 0.107 + ], + "angle": 0, + "content": "menting invisible moving objects. In Proc. BMVC, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.107, + 0.891, + 0.146 + ], + "angle": 0, + "content": "[19] Hala Lamdouar, Charig Yang, Weidi Xie, and Andrew Zisserman. Betrayed by motion: Camouflaged object discovery via motion segmentation. In Proc. ACCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.147, + 0.892, + 0.2 + ], + "angle": 0, + "content": "[20] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding, pages 45-56, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.201, + 0.892, + 0.253 + ], + "angle": 0, + "content": "[21] José Lezama, Karteek Alahari, Josef Sivic, and Ivan Laptev. Track to the future: Spatio-temporal video segmentation with long-range motion cues. In CVPR, pages 3369-3376, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.254, + 0.892, + 0.295 + ], + "angle": 0, + "content": "[22] Chiuhsiang Joe Lin, Chi-Chan Chang, and Yung-Hui Lee. Evaluating camouflage design using eye movement data. Applied ergonomics, pages 714-723, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.295, + 0.892, + 0.335 + ], + "angle": 0, + "content": "[23] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In CVPR, pages 8772-8781, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[24] Peter Ochs and Thomas Brox. Object segmentation in video: a hierarchical variational approach for turning point trajectories into dense regions. In ICCV, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[25] Peter Ochs and Thomas Brox. Higher order motion models and spectral clustering. In CVPR, pages 614-621, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.402, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[26] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. volume 36, pages 1187-1200, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[27] Andrew Owens, Connelly Barnes, Alex Flint, Hanumant Singh, and William Freeman. Camouflaging an object from many viewpoints. In CVPR, pages 2782-2789, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.482, + 0.892, + 0.535 + ], + "angle": 0, + "content": "[28] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages 2160-2170, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.535, + 0.892, + 0.577 + ], + "angle": 0, + "content": "[29] Xavier Soria Poma, Edgar Riba, and Angel Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In Proc. WACV, pages 1923-1932, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[30] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In ICCV, pages 4570-4580, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[31] Josef Sivic, Frederik Schaffalitzky, and Andrew Zisserman. Object level grouping for video shots. In Proc. ECCV. Springer-Verlag, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.725 + ], + "angle": 0, + "content": "[32] Przemysław Skurowski, Hassan Abdulameer, J Blaszczyk, Tomasz Depta, Adam Kornacki, and P Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2018. URL: https://www.polsl.pl/rau6/chameleon-database-animal-camouflage-analysis/." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.725, + 0.892, + 0.806 + ], + "angle": 0, + "content": "[33] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.806, + 0.892, + 0.86 + ], + "angle": 0, + "content": "[34] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proc. CVPR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.888 + ], + "angle": 0, + "content": "[35] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception ar-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.926, + 0.51, + 0.937 + ], + "angle": 0, + "content": "841" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "chitecture for computer vision. In CVPR, pages 2818-2826, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.12, + 0.468, + 0.146 + ], + "angle": 0, + "content": "[36] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.147, + 0.468, + 0.187 + ], + "angle": 0, + "content": "[37] Jolyon Troscianko, John Skelhorn, and Martin Stevens. Quantifying camouflage: how to predict detectability from appearance. BMC evolutionary biology, 17(1):1-13, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.187, + 0.468, + 0.213 + ], + "angle": 0, + "content": "[38] Cédric Villani. Optimal transport: old and new, volume 338. Springer, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.213, + 0.468, + 0.252 + ], + "angle": 0, + "content": "[39] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. In NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.253, + 0.468, + 0.293 + ], + "angle": 0, + "content": "[40] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proc. ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.468, + 0.348 + ], + "angle": 0, + "content": "[41] Bowen Zhang, Shuyang Gu, Bo Zhang, Jianmin Bao, Dong Chen, Fang Wen, Yong Wang, and Baining Guo. Styleswin: Transformer-based gan for high-resolution image generation. In CVPR, pages 11304-11314, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.926, + 0.511, + 0.936 + ], + "angle": 0, + "content": "842" + } + ] +] \ No newline at end of file diff --git a/2023/The Making and Breaking of Camouflage/b706bafb-dae1-4647-ad2f-702e05fcaa82_origin.pdf b/2023/The Making and Breaking of Camouflage/b706bafb-dae1-4647-ad2f-702e05fcaa82_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8234f85b4faa36f5dfa095e6efee77c37550fe57 --- /dev/null +++ b/2023/The Making and Breaking of Camouflage/b706bafb-dae1-4647-ad2f-702e05fcaa82_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc85521234ed6f9674b8fac81a4fef51b7c01897dcc36e707d9ac26f6599290d +size 4897684 diff --git a/2023/The Making and Breaking of Camouflage/full.md b/2023/The Making and Breaking of Camouflage/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eb7bef26e6a15843c25d519e4b9d014fa00f52e7 --- /dev/null +++ b/2023/The Making and Breaking of Camouflage/full.md @@ -0,0 +1,418 @@ +# The Making and Breaking of Camouflage + +Hala Lamdouar1 + +Weidi Xie $^{1,2}$ + +Andrew Zisserman + +Visual Geometry Group, University of Oxford + +$^{2}$ CMIC, Shanghai Jiao Tong University + +{lamdouar,weidi,az}@robots.ox.ac.uk + +# Abstract + +Not all camouflages are equally effective, as even a partially visible contour or a slight color difference can make the animal stand out and break its camouflage. In this paper, we address the question of what makes a camouflage successful, by proposing three scores for automatically assessing its effectiveness. In particular, we show that camouflage can be measured by the similarity between background and foreground features and boundary visibility. We use these camouflage scores to assess and compare all available camouflage datasets. We also incorporate the proposed camouflage score into a generative model as an auxiliary loss and show that effective camouflage images or videos can be synthesised in a scalable manner. The generated synthetic dataset is used to train a transformer-based model for segmenting camouflaged animals in videos. Experimentally, we demonstrate state-of-the-art camouflage breaking performance on the public MoCA-Mask benchmark. + +# 1. Introduction + +Camouflage has long been a subject of interest and fascination for the scientific community, especially evolutionary biologists who consider it as an excellent example of species adaptation. In order to confuse predators or to hide from prey, and increase their chances of survival in their natural habitat, animals have developed numerous camouflage mechanisms, e.g., disruptive coloration and background matching. Some species have even evolved to develop an adaptive camouflage, e.g., an arctic fox loses its white fur to better match the brown grey of the new season's landscape. Perhaps the most dramatic camouflage adaptation is the cuttlefish; it changes its patterns dynamically and rapidly as it moves from one spot to the other, constantly adapting and improving its camouflage. + +This search for optimal camouflage has inspired numerous works in the computer vision community, such as [12, 27], that tackled camouflage as a problem of optimal texture synthesis, making 3D objects non-detectable in a given scene. Others have addressed camouflage as a highly + +![](images/acdb0323b3a207303092250111f36da1588f1f7b145b21d66b78260c3a8276ef.jpg) +a + +![](images/758a7bb71018f45b0f7b16043e3b56ce11b372d43202f6f0e7276681a2c98c14.jpg) +b + +![](images/d6dfef7ca4e28c02701f98c863384965930bab9386775378b22ed1ec4136de47.jpg) +C + +score function + +camouflageranking: $c > a > b$ + +Figure 1: The three images depict the same animal, an Arctic fox, as it adapts its appearance to better blend with the changing landscape of the new season. While images $a$ and $c$ exhibit better background similarity than image $b$ , the fox boundary is more visible in image $a$ than image $c$ . We assess the effectiveness of camouflage by measuring the degree of ambiguity it creates with respect to its background. + +challenging object segmentation task in [9, 15, 20, 23]. Efforts have been made in collecting large scale camouflage datasets [9, 20] with costly annotation. In fact, camouflaged animals often exhibit complex shapes and thin structures that add to the boundary ambiguity and make the annotation highly time-consuming. Fan et al. report up to $60\mathrm{min}$ per image to provide accurate pixel-level annotations for their dataset COD10K [9]. While another line of research turned to camouflage breaking in sequences by taking advantage of motion cues [1, 5, 18, 19, 40], the camouflage data scarcity is even more extreme for videos. Recently, Sim2Real [18, 39] training has shown to be very effective for motion segmentation. By training on the optical flows of synthetic videos, these models can generalise to real videos without suffering from the domain gaps. + +In this paper, we start by asking the question: "what are the properties of a successful camouflage?" To answer this question, we investigate three scoring functions for quantifying the effectiveness of camouflage, namely, reconstruction fidelity score $(S_{R_f})$ , boundary score $(S_b)$ and intra-image Fréchet score $(d_{\mathcal{F}})$ . These scores are later adopted for two critical roles, (i) they are used to assess the relevance of existing camouflage datasets and act as quality-indicator in collecting new camouflage data; (ii) + +they can be used as a proxy loss for image and video inpainting/generation, where we establish a synthetic camouflage data generation pipeline with a Generative Adversarial Network (GAN), that can simultaneously generate high-quality camouflage examples and masks of the camouflaged animals. We further train a Transformer-based architecture on the generated camouflage video sequences, and demonstrate a performance boost over training on only the (small scale) real data. + +In summary, we make the following contributions: First, we introduce three scoring functions to measure the effectiveness of a given camouflage in still images and videos. We use these camouflage scores to rank the images in existing datasets in terms of camouflage success, and also show that the rankings are largely in accordance with human-produced rankings. Second, we incorporate the camouflage score into a generative model, establishing a scalable pipeline for generating high-quality camouflage images or videos, along with the pixelwise segmentation mask of the camouflaged animals; Third, we show that a Transformer-based model trained on the synthetically generated data can achieve state-of-the-art performance on the MoCA-Mask video camouflaged animal segmentation benchmark. + +# 2. Related work + +Camouflage Evaluation. Although there are no previous computational works that directly assess camouflage, as far as we know, there have been a number of human studies. Previous works proposed methods to evaluate camouflage by analysing the human viewers' eye movements [22, 37] or conducting subjective perceptual experiments [12, 27, 32]. In [12, 27], participants were asked to indicate the camouflaged object and their answers were analysed in terms of accuracy and time needed per image. Similarly, Skurowski et al. [32] asked human volunteers to rate CHAMELEON [32] images from 1 to 5 in terms of camouflage effectiveness. They produced a score after compensating for personal bias. We validate our proposed camouflage scoring functions by comparing our rankings to these human-based rankings. + +Motion Segmentation. The goal of this task is to partition the frames of a video sequence into background and independently moving objects. Early approaches focused on clustering motion patterns by grouping pixels with similar trajectories [3, 17, 21, 24, 25, 26, 31]. Another line of research tackled the problem by compensating for the background motion [1, 2, 19], via registering consecutive frames, or explicitly estimating camera motion. More recently, in [40], a Transformer-like architecture is used to reconstruct the input flows, and the segmentation masks are generated as a side product, while [6] decomposes the flow filed into regions by fitting affine transformations. The most + +closely related to ours is [18, 39], where the authors adopt a Sim2Real strategy by training the model on optical flow computed from synthetic videos. These models can generalise towards real videos without fine-tuning. In this paper, we train a model on synthetic camouflage RGB sequences. + +Frechet Inception Distance and its variants. Also known as Wasserstein-2 distance [38], Fréchet distance (FD) is a metric quantifying the difference between two probability distributions [10]. Lately, [13] introduced the Fréchet Inception Distance (FID) in the context of generative models. Under the assumption that real images and generated images follow Gaussian distributions, they compute the FD of the two distributions. Note that here, each image is represented by a vector obtained from the last pooling layer of InceptionV3 [35]. Following their steps, [30] adapted FID to the image re-targeting task and introduced the Single Image Fréchet Inception Distance (SIFID). Instead of comparing entire image datasets, they compare two images and consider the distributions of their feature vectors obtained from earlier layer of InceptionV3 [35]. Recently, [7] included FID in the training objective and introduced Fréchet-GAN. Our work takes inspiration from [7, 30] and investigates a FID of regions within the same image as an auxiliary loss. + +# 3. Measuring the effectiveness of camouflage + +Assuming there exists a dataset, $\mathcal{D} = \{(I_1,m_1),\ldots ,$ $(I_N,m_N)\}$ , where $I_{i}\in \mathbb{R}^{H\times W\times 3}$ refers to the images, and $m_i\in \{0,1\}^{H\times W\times 1}$ denotes a binary segmentation mask of the camouflaged animal. Here, we aim to design a scoring function that takes the image and segmentation mask as input, and outputs a scalar value $S$ that can quantify the effectiveness of camouflage, i.e., how successfully an animal blends into its background, termed as the camouflage score, + +$$ +S: (I _ {i}, m _ {i}) \longmapsto s _ {i} \in [ 0, 1 ] +$$ + +specifically, for $i \neq j$ and $s_i < s_j$ , the animal in $I_j$ , indicated by the mask $m_j$ , is more concealed than the animal in $I_i$ , indicated by the mask $m_i$ . Having such a scoring function enables various applications, first, it enables to rank the images from one dataset, in terms of the effectiveness of camouflage, and generate camouflage-relevant statistics for the dataset; second, it can serve as the objective for low-level image generation tasks, for example, image inpainting, editing, etc. + +In the following sections, we investigate three such scoring functions, specifically, we exploit the animal's mask and quantify the key aspects that determine the camouflage success both perceptually (Sec. 3.1) and probabilistically (Sec. 3.2). Note that, we take into account the local aspect of camouflage, i.e., instead of processing the entire background region, we only focus on the immediate + +![](images/0c6ff09a76b9d508514d200550ab067f87aa64e16710d61143837536fab88a82.jpg) +1 + +![](images/1759ca146b405d8dabfef343fbec1b5c33972393ce87ede6d026804762e706a1.jpg) + +![](images/6085a984a7ccbde1dcba2ab54db69eed384d4fcc384e1f928c38adbc40d1f684.jpg) +$I_{bg}$ + +![](images/f0ddbae7d3aaaf399ac03767840eb5b6ae4b66acf19036927fb7ef5d19cb4ea4.jpg) + +![](images/49bb0f76d2411510dcabfc805a9c12eac5e5abacf324da3fc03f2e361baa5415.jpg) +Figure 2: The reconstruction fidelity score $S_{R_f}$ evaluates the similarities between the original foreground $I_{fg}$ and its reconstruction from background features $\Psi_{bg}(I_{fg})$ . The top example shows a case where the animal exhibits different visual patterns (color) from its background $(S_{R_f} = 0.11)$ , while the bottom example shows a better background matching $(S_{R_f} = 0.82)$ . + +![](images/81dfcb9dc60e8f7c80ec6ba0f4ae8c4c422310ffc4a6bb754bd6a2835efa3078.jpg) + +![](images/71fa186d169a144de8fb90476ef32242a700c72953a1d9f26fc7551798689c07.jpg) +$\Psi_{bg}(I_{fg})$ + +surround of the animal. Hence, when referring to background, we mean the local background and only consider the cropped images centred around the object. + +# 3.1. Perceptual camouflage score + +In this section, we define the camouflage score through a perceptual lens, specifically, we measure the foreground and background similarity (Sect. 3.1.1), along with the boundary visibility (Sect. 3.1.2). + +# 3.1.1 Reconstruction fidelity score + +We attempt to reconstruct the foreground region with patches from the background, and propose a score to quantitatively measure the discrepancy between the original image and its reconstruction. Intuitively, for successful camouflage, we should be able to reconstruct the foreground perfectly by copying patches from its close neighborhood. + +Formally, for a given image $(I)$ and segmentation mask $(m)$ , subscript has been ignored for simplicity, we consider a trimap separation and define the foreground and background regions using morphological erosion and dilation of the mask, + +$$ +I _ {\mathrm {f g}}, I _ {\mathrm {b g}} = m _ {\mathrm {f g}} \odot I, m _ {\mathrm {b g}} \odot I \tag {1} +$$ + +where $m_{\mathrm{fg}} = \operatorname{erode}(m)$ and $m_{\mathrm{bg}} = 1 - \operatorname{dilate}(m)$ . + +To reconstruct the foreground region, we replace the foreground patch with the closest one in the background. Here a patch is a $n \times n$ pixel region ( $n = 7$ in our case with an overlap of 3) and the patchwise similarity is computed by exploiting the low-level perceptual similarity, i.e., comparing corresponding RGB values. This reconstruction method is inspired by the texture generation and inpainting approaches, such as [8]. In practice, we implement it with fast approximate nearest neighbor search for efficiency. + +The reconstruction fidelity score is computed by assessing the difference value between the foreground region and + +![](images/1943e4b80c021397a9deb3c8d7e5c1c2515ae3ac3362302d2719334a9a40aa4e.jpg) + +![](images/d7d53a5c03c2d669c20cc43dffa73bc5353dccf66a7254f19be6f04eac09352e.jpg) +I + +![](images/66fb6378b291b130de9d698e6f20bfc45618b095b2cb48a688fbf333b138f7c1.jpg) + +![](images/2a267669f36fb0477be3d48781fcada221b81ea83127b12a460e6fe96381523a.jpg) +Trimap + +![](images/13e2f1c90b7851132a4ee4256a4cf3d668266358acef757e31f5d6424b048dfc.jpg) + +![](images/5f03892f6c37fa35cee3086f08952931687805dcf804a6cd95ce6f3b8464cea7.jpg) +C +Figure 3: Trimap regions and animal contours. From left to right: Input image; Trimap partition (foreground region in yellow, background in purple, and boundary region in green); Contours $C$ computed on the original image; and ground truth Contours $C_{gt}$ . The top example shows less contour agreement along the boundary region ( $S_b = 0.72$ ) than the bottom example ( $S_b = 0.40$ ). + +![](images/267e36e0db0341525401d83dbe0e77f5207a9175695eb9822a168d31da6e4d0e.jpg) + +![](images/24deecae6fa693f42c16a9961cc19aec933802e5cf6ce7705d4cab5b575995fe.jpg) +$C_{gt}$ + +its reconstruction, specifically, we count the number of foreground pixels that have been successfully reconstructed from the background: + +$$ +S _ {R _ {f}} (I, m) = \frac {1}{N _ {\mathrm {f g}}} \sum_ {(i, j) \in I _ {\mathrm {f g}}} R _ {f} (i, j) \tag {2} +$$ + +$$ +R _ {f} (i, j) = \left\{ \begin{array}{l l} 1 & \text {i f} | | I _ {\mathrm {f g}} - \Psi_ {I _ {\mathrm {b g}}} \left(I _ {\mathrm {f g}}\right) | | _ {2} < \lambda | | I _ {\mathrm {f g}} | | _ {2} \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {3} +$$ + +$\Psi_{I_{bg}}(.)$ denotes the reconstruction operation, $N_{\mathrm{fg}} = |m_{\mathrm{fg}}|$ is the total number of pixels in the foreground region, and $\lambda$ is a threshold ( $\lambda = 0.2$ in our case). A higher $S_{R_f}$ score, means that the animal's visual attributes are well represented in the background. Conversely, a low $S_{R_f}$ indicates that the animal's appearance is composed of unique features that makes it standout from its surrounding. In Fig. 2, we present examples of camouflage evaluation by reconstruction fidelity. + +# 3.1.2 Boundary visibility score + +With the trimap representation, we can easily extract the boundary region, $I_{\mathrm{b}} = m_{\mathrm{b}} \odot I$ , where $m_{\mathrm{b}} = (1 - m_{\mathrm{bg}}) - m_{\mathrm{fg}}$ . Here, we aim to measure the animal's boundary properties, e.g., contour visibility, as they can also provide visually evident cues for camouflage breaking. In particular, we adopt an off-the-shelf contour extraction method [29], and run it on the original image and foreground mask, to generate the images' contour ( $C$ ) and ground truth animal's contour ( $C_{\mathrm{gt}}$ ), as shown in Fig. 3. + +We express the agreement between predicted contours $C$ and ground truth contours $C_{\mathrm{gt}}$ over the boundary region with the F1 metric: + +$$ +S _ {b} (I, m) = 1 - \mathrm {F} 1 \left(m _ {b} \odot C _ {\mathrm {g t}}, m _ {b} \odot C\right) \tag {4} +$$ + +This score penalises the boundary pixels that are predicted as contour in both $C$ and $C_{\mathrm{gt}}$ . While we do not expect a + +perfect contour agreement for visible edges, we consider $S_{b}$ as a reasonable approximation, given the shape and size of the boundary region as a thin margin centred around the animal. If a boundary pixel is predicted as a contour in $C_{\mathrm{gt}}$ but not in $C$ , this means that the animal's edge is not visible in the original image. If a pixel is predicted as a contour in $C$ but not $C_{\mathrm{gt}}$ , this indicates the presence of distracting elements, such as complex texture patterns on the foreground animal, which also affects the visibility of the animal's contour and therefore improves its camouflage. A perfectly camouflaged animal would have little contour agreement and $S_{b} = 1$ . + +To get the final perceptual camouflage score, we linearly combine both reconstruction fidelity score and boundary visibility score: + +$$ +S _ {\alpha} = (1 - \alpha) S _ {R _ {f}} + \alpha S _ {b} \tag {5} +$$ + +We will describe our method for setting the weighting parameter $\alpha$ in the Experiment section (Sec. 6). + +# 3.2. Probabilistic scoring function + +In addition to using the low-level RGB representation, in this section, we consider the pixelwise representation in the feature space, and propose a differentiable metric that compares the probabilistic distribution between the foreground and background regions, which acts as a proxy for the score $S_{\alpha}$ described in previous section, and can be used directly as a differentiable loss function in image generation. + +We consider the Intra-Image Fréchet Inception Distance, specifically, we compute the feature maps for the foreground and background image regions: + +$$ +f _ {f g}, f _ {b g} = \Phi_ {\mathrm {v}} \left(I _ {f g}\right), \Phi_ {\mathrm {v}} \left(I _ {b g}\right) \tag {6} +$$ + +Here $\Phi_{\mathrm{v}}(\cdot)$ denotes a pre-trained Inception network [34, 35], truncated at early layers. We take inspiration from [30], and adapt the single image Fréchet distance to measure the deviation between feature distributions of different regions within the same image. We adopt the Fréchet hypothesis with respect to our regions, i.e., the features of the foreground and background follow multivariate Gaussian distributions: $f_{\mathrm{fg}} \sim \mathcal{N}(\mu_{\mathrm{fg}}, \Sigma_{\mathrm{fg}})$ and $f_{\mathrm{bg}} \sim \mathcal{N}(\mu_{\mathrm{bg}}, \Sigma_{\mathrm{bg}})$ . The intra-image Fréchet distance can be formulated as follows: + +$$ +d _ {\mathcal {F}} ^ {2} (I, m) = \left\| \mu_ {\mathrm {f g}} - \mu_ {\mathrm {b g}} \right\| _ {2} ^ {2} + T r \left(\Sigma_ {\mathrm {f g}} + \Sigma_ {\mathrm {b g}} - 2 \left(\Sigma_ {\mathrm {f g}} \Sigma_ {\mathrm {b g}}\right) ^ {1 / 2}\right) +$$ + +Note that, when $(I,m)$ is the output from a generative model, $d_{\mathcal{F}}^2$ is differentiable with respect to its parameters, and can therefore be used as an auxiliary loss for optimising the image generation procedure, i.e., generate effective camouflage examples. + +# 4. Generating camouflage video sequences + +In this section, we propose a scalable pipeline for generating images with concealed animals. In particular, we incorporate the differentiable intra-image Fréchet distance in image generation, explicitly optimising the camouflage score of the image. We first detail the static image generation framework, then describe how we use the images it generates to create camouflage video sequences. + +# 4.1. Camouflage image generation + +In order to generate the image with a camouflaged animal, and its corresponding segmentation mask, we adopt a Generative Adversarial Network (GAN), where a generator is fed with a latent vector $z \sim \mathcal{N}(0,1)$ and learns to predict the pair of realistic image and segmentation mask, such that a discriminator cannot differentiate it from real samples. + +Let $(x_{i},m_{i})$ denote an image and segmentation pair sampled from the training dataset, a subset of COD10K [9] is used here. We adopt the conventional label notation, i.e. $y = 1$ for (real) examples sampled from the training dataset and $y = 0$ for (fake) examples from the generator, and consider the generative adversarial loss functions [11]: + +$$ +\mathcal {L} _ {D} = E _ {\left(x _ {i}, m _ {i}\right)} \left[ - \log \left(D \left(x _ {i}, m _ {i}\right)\right) \right] + E _ {z} \left[ - \log \left(1 - D (G (z))\right) \right] +$$ + +$$ +\mathcal {L} _ {G} = E _ {z} \left[ - \log (D (G (z))) \right] +$$ + +To enforce coherence between images and their segmentation masks, we feed the discriminator with additional fake pairs, consisting of real images coupled with unpaired real masks. This can be formulated as an additional coherence loss term that minimises the probability of assigning incorrect labels, i.e. $y = 1$ , to $(x_{i},m_{j})$ sampled from the training set such that $i\neq j$ : + +$$ +\mathcal {L} _ {\mathrm {C o h}} = E _ {\left(x _ {i}, m _ {j}\right) i \neq j} [ - \log (1 - D \left(x _ {i}, m _ {j}\right)) ] \tag {7} +$$ + +To increase the camouflage effectiveness in the generated examples, we adopt the intra-image Fréchet distance in the generator loss, as an auxiliary loss term, and train our camouflage image generator with the following loss: + +$$ +\mathcal {L} _ {\tilde {D}} = \mathcal {L} _ {D} + \mathcal {L} _ {\mathrm {C o h}} \quad \mathcal {L} _ {\tilde {G}} = \mathcal {L} _ {G} + \beta d _ {\mathcal {F}} ^ {2} \tag {8} +$$ + +We present an overview of our data generation framework in Fig. 4. + +# 4.2. Camouflage video generation + +Given a camouflage image and corresponding segmentation mask, generated using the method above, we can create video sequences by applying different motions to the foreground and background, in a similar manner to [18]. Specifically, we first inpaint the backgrounds with an off-the-shelf model proposed by Suvorov et al. [33] and overlay the synthetic animal (extracted from the original generated image + +![](images/43d5b3a4c21a12801039fe84cedce83142917067f3d79df01c89be4f126f27c2.jpg) +Figure 4: Our framework consists of a synthetic camouflage data generation pipeline (left), where we train a generator $G$ in a GAN setting, to create camouflage images and masks, while encouraging camouflage effectiveness via minimising $\mathcal{L}_{\mathcal{F}}$ . The generated samples are then transformed into synthetic video sequences (middle), following our method presented in Sec. 4.2. The transformer-based motion segmentation architecture (right), for camouflage breaking in videos, is trained on the synthetic video sequences. The architecture is described in Sec. 5. + +![](images/87982e9137d9d52a0682cd4d20f167dcac222049a339fbd9973432f2a959dd3f.jpg) + +![](images/273e9d1cc6704b62007721a334ae0c53e98ce159e2e471b72d76bee842bdacb1.jpg) + +by using the mask) at a random location within the image in the first frame, then following a random translational trajectory in the following frames. We incorporate a different translational motion for the background and include random sub-sequences where the foreground and background undergo the same motion to simulate momentarily static objects. As we train our generator with intra-image Fréchet loss $(\mathcal{L}_{\mathcal{F}})$ , the generated images exhibit strong similarities between the foreground and the background, and even if the object is placed at a different location from the original, we expect it to remain concealed within its surrounding. + +# 5. Learning to break camouflage + +In this section, we train a transformer-based architecture on the synthetic dataset, and demonstrate its effectiveness for breaking the camouflage in real videos, for example, MoCA [19]. We build on two previous architectures, namely, the motion segmentation from [18] and the Search Identification Network from [9]. While the first architecture processes sequences of optical flow, the second takes single RGB images as inputs and treats them separately. Our proposed model, shown in Fig. 4, takes both the optical flow and RGB frame sequences as input. The flow is computed with RAFT [36] from the RGB sequence, and processed with a motion encoder, followed by a transformer encoder. The RGB sequence is processed with the appearance encoder from [9], pre-trained for framewise camouflage segmentation. Then both streams are aggregated as inputs to a Mask2Former-like [4] transformer decoder and pixel decoder to produce the high resolution segmentation mask from the motion stream. + +Motion encoder. To encode the motion cues, we use a light-weight convNet that takes as input a sequence of optical flows, $\mathbf{I}_{\mathbf{m}} = \{I_{m_1}, I_{m_2},.., I_{m_t}\} \in \mathcal{R}^{t \times c_0 \times h \times w}$ and out + +puts motion features: + +$$ +\left\{f _ {m _ {1}}, f _ {m _ {2}}, \dots , f _ {m _ {t}} \right\} = \Phi_ {\text {m o t i o n}} \left(\mathbf {I} _ {\mathbf {m}}\right) +$$ + +where each flow frame is separately embedded. + +Motion transformer encoder. A transformer encoder takes as input the motion features, concatenated along the sequence, together with learned spatial and temporal positional encodings $(\mathrm{Pos}_s,\mathrm{Pos}_t)$ , as indicated in Fig 4. $\mathrm{Pos}_t$ specifies the frame, and $\mathrm{Pos}_s$ specifies the position in the spatial feature map output of the motion encoder. The output of the transformer is a set of enriched motion features. + +Appearance encoder. Here, we adopt a SINet-v2 [9] architecture, that encodes the RGB sequence, $\mathbf{I}_{\mathbf{a}} = \{I_{a_1}, I_{a_2},.., I_{a_t}\} \in \mathcal{R}^{t \times c_1 \times h \times w}$ into appearance features: + +$$ +\left\{f _ {a _ {1}}, f _ {a _ {2}}, \dots , f _ {a _ {t}} \right\} = \Phi_ {\text {a p p}} \left(\mathbf {I} _ {\mathbf {a}}\right) +$$ + +Again, each RGB frame is processed separately by SINet. + +Transformer decoder. A transformer-based decoder takes the output of the motion transformer encoder together with the appearance features and a learnable query for the mask embedding. In a similar manner to Mask2Former [4], the query attends to multiple resolutions of the motion features concatenated with the appearance features and produces a mask embedding for the moving object. + +Pixel Decoder. Similarly to the pixel decoder in Mask2Former, a light-weight convNet decoder is used with skip-connections to recover high-resolution segmentation masks $\{\hat{m}_1,\hat{m}_2,\dots ,\hat{m}_t\}$ from the motion features and the mask embedding. This is shown as the blue box in Fig. 4. + +Training objective. We train the motion segmentation model on our synthetic video sequences using the binary cross-entropy loss $\mathcal{L}_{BCE}$ . + +# 6. Experiments + +In this section, we start by introducing the datasets involved in this paper, followed by the implementation details. Our experiments present a thorough analysis of the proposed camouflage scores and demonstrate their effectiveness in our training framework. + +# 6.1. Datasets + +Here we describe the publicly available camouflage datasets that we included in our experiments, as shown in Fig. 5, as well as the synthetic camouflage datasets that we generated. + +CHAMELEON [32] is one of the first camouflage image datasets. It contains 76 images collected using Google image search with the 'camouflaged animal' keyword and include ground truth manual annotations. + +CAMO [20]. The Camouflaged Object dataset consists of 2500 images collected from the internet, of which 1250 images (1000 training sub-set and 250 testing sub-set) contain at least one camouflage instance from 8 categories with manual pixelwise annotations provided. + +COD10K [9]. COD10K contains 10,000 images collected from photography websites of which 5066 depict camouflaged animals (3040 training sub-set and 2026 testing subset), organised into 10 classes, and 78 sub-classes (69 camouflaged). Note that, in our camouflage evaluation experiments we only used the camouflage subset of this dataset. + +Camouflaged Animals [1]. This dataset consists of 9 video sequences of camouflage animals from 6 categories. + +MoCA [19] is the first large-scale camouflage video dataset. It contains 141 video sequences (37K frames) totalling 67 animal categories. Recently, other works have curated this dataset by selecting only the videos with more prominent motion (locomotion) [40], and others have provided dense pixel annotation in MoCA-Mask [5]. We use the latter version in our experiments. + +Camouflaged cuboids [12, 27]. This dataset was created for texture synthesis and consists of multiple-view scenes, where cuboids were placed at a predefined location then synthetically camouflaged. In our evaluation, we consider the 4-views generated textures [12] from 36 scenes as well as the cuboids masks. + +Synthetic Camouflage Images. Using the method described in Sec. 4.1, we generate a synthetic camouflage dataset, of $5K$ frames, by discriminating against real camouflage image and annotation masks from COD10K. + +Synthetic Camouflage Videos. We generate $1K$ sequences of 30 frames each, incorporating static sequences using the framework described in Sec. 4.2. We split these into 800 sequences for training and 200 for testing. + +![](images/6f1b01b60fa09cb7a1475102ac7312ac02003bd5e91c5053782ebfe5639ed8d2.jpg) +Figure 5: Randomly sampled examples from the camouflage datasets included in our work. For the video datasets, Camouflaged Animals and MoCA-Mask, we show an example sequence. For the multiple-view dataset Camouflaged cuboids (bottom), we show example views from a scene with the 4-view texture synthesis method from [12]. + +# 6.2. Implementation details + +In our experiments, for a given camouflage example, the kernels for the morphological operations are selected from a range of values [1,10], so that the resulting annotation mask + +is reduced by $20\%$ for erosion and extended by $20\%$ for dilation. This allows the erosion/dilation to be adapted to the size of the camouflaged animal, while always keeping a reasonably large region for addressing the boundary effects. + +For our synthetic camouflage image generation, we adopt a Style-GAN architecture [41] and train it on the camouflage images from COD10K to generate $256 \times 256$ images and masks pairs. We use $\beta = 0.1$ to weight our intra-image Fréchet auxiliary loss. For numerical stability, we adopt the Newton Schultz iterations when calculating the matrix square root term in Intra-Image Fréchet Inception Distance, with $T = 100$ iterations. + +When generating the image sequences, we compute optical flows with RAFT [36], and train our moving camouflaged animal segmentation model. We use batches of size 2 and sequences of size 5. We first train on our synthetically generated dataset with a learning rate of $5 \times 10^{-4}$ for 500 iterations then fine-tune on the training subset of MoCA-Mask. + +# 6.3. Results + +This section presents qualitative and quantitative results, to demonstrate the effectiveness of our score functions and the benefit of including $d_{\mathcal{F}}^2$ in the training loop. + +# 6.3.1 Evaluation on camouflage effectiveness + +Ranking camouflage images in COD10K. We can rank the images based on the effectiveness of the animal's camouflage, with our proposed scoring functions. In the left part of Fig. 6, we show the four best scored images from the large scale COD10K dataset. We can make the following observations: (i) the best examples with respect to $S_{R_f}$ (top), which favours the background matching of the animals without taking into account the boundary region, all exhibit visible boundaries, with the exception of image (a). This is especially noticeable in the rabbit example (d) along the ears and shadow regions; (ii) the best four examples with respect to the boundary score $S_b$ are all from the caterpillar subclass, mostly the baron type. These insects have thin, and transparent boundaries, that makes them the perfect candidates for this score. However, with the exception of (a) and (c), the animals still stands out as they exhibit colors and patterns that are not present within their background; (iii) the $S_\alpha$ combining both scores shows highly effective camouflage examples. In contrast, the right part of Fig. 6 shows the lowest scored images for each approach. For $S_{R_f}$ , we find examples with low background matching in the first row, with two ant examples, i.e., (e), (f). While the boundary score penalises examples with high contrasts between the animal and its background, that results in more visible contours. This is also the case for the lowest $S_\alpha$ score examples. + +Dataset level camouflage comparison. We compute the + +camouflage score for all the camouflage image and video datasets and report the results in Tab. 1. For the image datasets, the dataset level score is computed as the mean of the image scores; for the video dataset the score is first computed per frame, and per-video average computed, then the dataset level score is the mean of the averages. Our experiments show that MoCA-Mask [5, 19] contains the most successful camouflages according to our scoring. While COD10K [9] subsets are balanced in terms of camouflage effectiveness, we find that, for CAMO [20], the test subset yields higher scores than the training subset and therefore contains more challenging examples. + +We note that Camouflaged cuboids dataset yields higher $S_{R_f}$ and $S_{\alpha}$ scores compared to our synthetic datasets. This is due to the fact that the model used in [12, 27] is only trained to produce optimal texture for a predefined region (cuboid) in a particular location of a given scene. However, our model learns to output a new image with a randomly located camouflaged object of random shape, thus offering more scalability and diversity compared to the cuboids dataset which cannot be used to train a model for breaking real camouflage. + +Comparison to human-produced rankings. We compare our rankings to the human scoring system based on the rating of CHAMELEON [32] and the time-to-find for Camouflaged cuboids [12]. To search for the optimal $\alpha$ parameter in Equation (5), we select 15 images from CHAMELEON and compare their $S_{\alpha}$ ranking with ground truth using kendl- $\tau$ metric [16]. For $\alpha = 0.35$ , we obtain kendl- $\tau = 0.51$ on this validation set that we excluded from our test reported in Tab. 2. We can draw the following observations: (i) we find that the boundary score $S_{b}$ produces more agreement with the ground truth ranking than $S_{R_{f}}$ , suggesting that human observers tend to pay more attention to contour visibility than background matching; (ii) while comparing to the ranking from Camouflaged cuboids, we found negative correlations for $S_{R_{f}}$ and $d_{\mathcal{F}}$ , we conjecture that this may be due to the nature of the dataset in [12], i.e., all the textured cuboids are synthetically generated with very high background matching; (iii) for both experiments, we obtain the strongest agreement with the $S_{\alpha}$ score combining both background similarity and boundary blending. + +Further analysis of the synthetic datasets. We adopt the FID [13] and IS [30] metrics in Table 4. These metrics assess: (i) the similarity between the set of synthetic images output by the generator and the set of real images used to train it (FID); and (ii) the clarity of the object and the diversity of images in terms of classes (IS). While they are not intended for measuring camouflage success, which is the focus of this work, IS could somewhat be (inversely) linked to $S_{b}$ as the object clarity and boundary visibility are close. Note that the object clarity (IS) is decreased with $\mathcal{L}_{\mathcal{F}}$ which is the intended purpose of improving camouflage, however, + +![](images/baf0bf0fcb2fab0b8bf954cb197bfe3f0b254b42dd3957dbe58611c4affd46e8.jpg) +Figure 6: Left: Top-4 scored examples from COD10K for $S_{Rf}(top)$ , $S_b(midle)$ and $S_{\alpha}$ (bottom). Right: Lowest scored examples from COD10K. For each example, we show the obtained score and the corresponding ground truth mask. + +
DatasetsData typeSRf↑Sb↑Sα↑d2F↓
CHAMELEON [32]Image0.6940.4450.6070.70
CAMO Train [20]Image0.6720.4510.5951.01
CAMO Test [20]Image0.6830.4700.6081.00
COD10K Train [9]Image0.6550.4330.5770.90
COD10K Test [9]Image0.6570.4310.5780.90
Camouflaged Animals [1]Video0.6740.5360.6261.60
MoCA-Mask Train [5, 19]Video0.8500.4430.7071.14
MoCA-Mask Test [5, 19]Video0.7330.4640.6392.51
Camouflaged cuboids[12, 27]Multi-view0.8940.4330.7336.2
Syn. Camouflage w.o. L_FImage0.6080.4320.5461.36
Syn. Camouflage w. L_FImage0.6790.4470.5981.13
Syn. Camouflage VideoVideo0.6580.4300.5781.18
+ +Table 1: Results of the proposed scores on natural camouflage datasets (top) and synthetically generated camouflage datasets (bottom). We report the mean scores for the single image datasets. For the video and multiple view datasets, we compute the mean per sequence and scene respectively. Syn. Camouflage w. $\mathcal{L}_{\mathcal{F}}$ refers to the synthetic image dataset that we generated while minimising $\mathcal{L}_{\mathcal{F}}$ and Syn. Camouflage Video its sequence version. + +this effect is not maintained in the sequence generation as the animal changes location within its background. + +
datasetFID↓IS↑
Syn. Camouflage w.o. L_F75.013.49
Syn. Camouflage w. L_F76.343.13
Syn. Camouflage Video86.423.85
+ +# 6.3.2 Model ablations + +On the effectiveness of $\mathcal{L}_{\mathcal{F}}$ . Fig. 7 shows the generated samples from the GAN pipeline, trained with and without the intra-image Fréchet Distance. Our generator is able to produce realistic object masks of complex shapes and thin structures similar to those usually encountered in the camouflage datasets. The produced images exhibit rich back + +Table 4: Overall quality of our synthetic datasets. + +
ScoresCHAMELEON [32]Camouflaged cuboids[12, 27]
SRf0.01-0.07
Sb0.030.42
Sα0.420.41
d2F0.10-0.17
+ +Table 2: Kendall- $\tau$ metric between rankings produced via our scores and human scoring ground truth. + +
ModelTraining datasetAppearanceEncoderTransformerEncoderTransformerAggregationmIoU↑F↑E↑MAE↓
AS14.520.457.39.4
BMM15.320.657.35.1
CS+MM16.021.859.84.3
DS+MM✓20.522.659.83.8
ES+MM✓✓23.023.557.01.6
FS+MM✓✓22.723.258.83.6
GMM✓✓✓23.424.762.02.4
HS+MM✓✓✓30.834.377.01.8
+ +Table 3: Ablation study on the different components of the motion segmentation architecture and training datasets. We evaluate on MoCA-Mask for all experiments. (S: our synthetic dataset, MM: MoCA-Mask). + +grounds with realistic nature-like elements mimicking rocks and coral structures present in the real camouflage data. Adding the $\mathcal{L}_{\mathcal{F}}$ loss results in generating images with better blending in the background, both qualitatively and quantitatively, as shown by the $S_{\alpha}$ computed for both datasets generated with and without the intra-image Fréchet Distance in Tab. 1. We refer the reader to the supplementary material for more examples of generated samples. + +On the architecture design and gain from the synthetic dataset. We present in Tab. 3 an ablation study detailing the gain from the main components of our architecture, as well as the benefit of first training on our synthetic dataset then fine-tuning on MoCA-Mask (S+MM), for model H, as opposed to only training on MoCA-Mask (MM), for G. + +![](images/7e56138070bfb2aa0549ae5bad0396b6546b425f700cba994dd51f61044fae60.jpg) +Figure 7: Generated camouflage images and masks: For the examples on the right, the generator was trained with $\mathcal{L}_{\mathcal{F}}$ + +![](images/6c24a53ec9d6905bc1fa53804f0499382f71c07cfb142f3d4693ba064e87aee0.jpg) +Figure 8: Qualitative results on MoCA-Mask. From top to bottom: Appearance sequence $I_{a}$ , flow sequence $I_{m}$ , predicted segmentations and ground truth segmentations. + +
ModelRGBMotionmIoU↑F↑E↑MAE↓
SINet [9]✓20.223.169.92.8
SINet-V2 [9]✓18.020.464.23.1
SegMaR [14]✓12.222.580.34.5
ZoomNet [28]✓18.828.770.82.5
SLTNet [5]✓✓27.231.175.92.7
MG [40]✓12.716.856.16.7
Ours-flow✓17.821.560.73.7
Ours✓✓30.834.377.01.8
+ +Table 5: Results on MoCA-Mask. Ours-flow refers to the optical flow only version of our architecture, excluding the appearance stream. + +# 6.3.3 Results on MoCA + +We train our model on the synthetically generated camouflage images and fine-tune on the training set of MoCA-Mask. Tab. 5 presents our camouflage object segmentation results on the test set of MoCA-Mask. Ours refers to our proposed method with SINet-V2 as the appearance encoder and Ours-flow refers to the optical flow only version of our architecture, excluding the appearance stream. Our main model outperforms RGB and Motion-based methods on MoCA-Mask. Fig. 8 presents qualitative results of our segmentation model. Note that our model is robust to degraded optical flow and static animals. + +# 6.4. Limitations + +While our image generation method encourages background matching through the $\mathcal{L}_{\mathcal{F}}$ loss term minimization, the sequence generating is not guaranteed to maintain object concealment. A possible solution could be to use the proposed $S_{\alpha}$ camouflage score to curate the generated sequence dataset and filter out the most visible objects. + +Our proposed camouflage scores use the ground truth annotation and analyse the different regions it defines. We found that our scoring system penalises specific cases of camouflage by occlusion, where elements from the background are partially occluding the animal and helping improve its camouflage. For instance, the grass in example (g) from the last row of Fig. 6 is not treated as part of the animal and therefore not considered in our background similarity + +assessment. One can argue that this is due to the ambiguity in the provided annotations and for such cases, extra amodal annotation should be also considered. + +# 7. Conclusion + +We present three score functions for computationally assessing the effectiveness of camouflage in images and videos. By evaluating the similarity with background and the boundary visibility, our combined perceptual score is strongly correlated with human perceptual ranking systems on two different datasets. We demonstrate that training a generative model with our differentiable camouflage function improves the effectiveness of generated camouflage examples and can be used to generate challenging synthetic camouflage datasets to train models to break camouflage. + +# Acknowledgement + +We are grateful to Przemysław Skurowski for providing the human perception study data on the CHAMELEON dataset and Andrew Owens, Rui Guo and Oscar de Lima for providing the data for the Camouflaged cuboids dataset. We thank Tengda Han for fruitful discussions. This research is supported by the UK EPSRC funded CDT in Autonomous Intelligent Machines and Systems (AIMS), the EPSRC Programme Grant VisualAI EP/T028572/1, a Schlumberger Studentship, and a Royal Society Research Professorship. WX is supported by the National Key R&D Program of China (No. 2022ZD0161400). + +# References + +[1] Pia Bideau and Erik Learned-Miller. It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In ECCV, pages 433-449, 2016. +[2] Pia Bideau, Rakesh R Menon, and Erik Learned-Miller. Moa-net: self-supervised motion segmentation. In Proc. ECCV Workshop, 2018. +[3] Thomas Brox and Jitendra Malik. Object segmentation by long term analysis of point trajectories. In European conference on computer vision, pages 282-295, 2010. +[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. +[5] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022. +[6] Subhabrata Choudhury, Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Guess What Moves: Unsupervised Video and Image Segmentation by Anticipating Motion. In Proc. BMVC, 2022. +[7] Khoa D Doan, Saurav Manchanda, Fengjiao Wang, Sathiya Keerthi, Avradeep Bhowmik, and Chandan K Reddy. Image generation via minimizing frechet distance in discriminator feature space. arXiv preprint arXiv:2003.11774, 2020. +[8] A. Efros and T. Leung. Texture synthesis by non-parametric sampling. In Proc. ICCV, pages 1039-1046, Sep 1999. +[9] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. +[10] Maurice Fréchet. Sur la distance de deux lois de probabilité. Comptes Rendus Hebdomadaires des Seances de L'Académie des Sciences, pages 689-692, 1957. +[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, pages 2672-2680, 2014. +[12] Rui Guo, Jasmine Collins, Oscar de Lima, and Andrew Owens. Ganmouflage: 3d object nondetection with texture fields. In arXiv preprint arXiv:2201.07202, 2022. +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. +[14] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In CVPR, pages 4713-4722, 2022. +[15] Nobukatsu Kajiura, Hong Liu, and Shin'ichi Satoh. Improving camouflaged object detection with the uncertainty of pseudo-edge labels. In ACM Multimedia Asia, pages 1-7. 2021. +[16] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1/2):81-93, 1938. +[17] Margret Keuper, Bjoern Andres, and Thomas Brox. Motion trajectory segmentation via minimum cost multicuts. In ICCV, pages 3271-3279, 2015. +[18] Hala Lamdouar, Weidi Xie, and Andrew Zisserman. Seg- + +menting invisible moving objects. In Proc. BMVC, 2021. +[19] Hala Lamdouar, Charig Yang, Weidi Xie, and Andrew Zisserman. Betrayed by motion: Camouflaged object discovery via motion segmentation. In Proc. ACCV, 2020. +[20] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding, pages 45-56, 2019. +[21] José Lezama, Karteek Alahari, Josef Sivic, and Ivan Laptev. Track to the future: Spatio-temporal video segmentation with long-range motion cues. In CVPR, pages 3369-3376, 2011. +[22] Chiuhsiang Joe Lin, Chi-Chan Chang, and Yung-Hui Lee. Evaluating camouflage design using eye movement data. Applied ergonomics, pages 714-723, 2014. +[23] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In CVPR, pages 8772-8781, 2021. +[24] Peter Ochs and Thomas Brox. Object segmentation in video: a hierarchical variational approach for turning point trajectories into dense regions. In ICCV, 2011. +[25] Peter Ochs and Thomas Brox. Higher order motion models and spectral clustering. In CVPR, pages 614-621, 2012. +[26] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. volume 36, pages 1187-1200, 2013. +[27] Andrew Owens, Connelly Barnes, Alex Flint, Hanumant Singh, and William Freeman. Camouflaging an object from many viewpoints. In CVPR, pages 2782-2789, 2014. +[28] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages 2160-2170, 2022. +[29] Xavier Soria Poma, Edgar Riba, and Angel Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In Proc. WACV, pages 1923-1932, 2020. +[30] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In ICCV, pages 4570-4580, 2019. +[31] Josef Sivic, Frederik Schaffalitzky, and Andrew Zisserman. Object level grouping for video shots. In Proc. ECCV. Springer-Verlag, 2004. +[32] Przemysław Skurowski, Hassan Abdulameer, J Blaszczyk, Tomasz Depta, Adam Kornacki, and P Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2018. URL: https://www.polsl.pl/rau6/chameleon-database-animal-camouflage-analysis/. +[33] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. +[34] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proc. CVPR, 2015. +[35] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception ar- + +chitecture for computer vision. In CVPR, pages 2818-2826, 2016. +[36] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. +[37] Jolyon Troscianko, John Skelhorn, and Martin Stevens. Quantifying camouflage: how to predict detectability from appearance. BMC evolutionary biology, 17(1):1-13, 2017. +[38] Cédric Villani. Optimal transport: old and new, volume 338. Springer, 2009. +[39] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. In NeurIPS, 2022. +[40] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proc. ICCV, 2021. +[41] Bowen Zhang, Shuyang Gu, Bo Zhang, Jianmin Bao, Dong Chen, Fang Wen, Yong Wang, and Baining Guo. Styleswin: Transformer-based gan for high-resolution image generation. In CVPR, pages 11304-11314, 2022. \ No newline at end of file diff --git a/2023/The Making and Breaking of Camouflage/images.zip b/2023/The Making and Breaking of Camouflage/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..72d5d5d815b70b8cb623a7791f9e65d784ff9606 --- /dev/null +++ b/2023/The Making and Breaking of Camouflage/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b85cfac558e5d52633f318ee38ff1503ec40f0a829cedeb911ad74b4e372259b +size 880868 diff --git a/2023/The Making and Breaking of Camouflage/layout.json b/2023/The Making and Breaking of Camouflage/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ea39e25d16787b81a5972824ce3daea3b935d733 --- /dev/null +++ b/2023/The Making and Breaking of Camouflage/layout.json @@ -0,0 +1,10789 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 168, + 103, + 426, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 103, + 426, + 121 + ], + "spans": [ + { + "bbox": [ + 168, + 103, + 426, + 121 + ], + "type": "text", + "content": "The Making and Breaking of Camouflage" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 142, + 235, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 142, + 235, + 156 + ], + "spans": [ + { + "bbox": [ + 154, + 142, + 235, + 156 + ], + "type": "text", + "content": "Hala Lamdouar1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 258, + 142, + 320, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 142, + 320, + 156 + ], + "spans": [ + { + "bbox": [ + 258, + 142, + 320, + 156 + ], + "type": "text", + "content": "Weidi Xie" + }, + { + "bbox": [ + 258, + 142, + 320, + 156 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 342, + 143, + 438, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 143, + 438, + 156 + ], + "spans": [ + { + "bbox": [ + 342, + 143, + 438, + 156 + ], + "type": "text", + "content": "Andrew Zisserman" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 157, + 304, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 157, + 304, + 171 + ], + "spans": [ + { + "bbox": [ + 77, + 157, + 304, + 171 + ], + "type": "text", + "content": "Visual Geometry Group, University of Oxford" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 157, + 517, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 157, + 517, + 172 + ], + "spans": [ + { + "bbox": [ + 326, + 157, + 517, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 326, + 157, + 517, + 172 + ], + "type": "text", + "content": "CMIC, Shanghai Jiao Tong University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 203, + 174, + 392, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 174, + 392, + 185 + ], + "spans": [ + { + "bbox": [ + 203, + 174, + 392, + 185 + ], + "type": "text", + "content": "{lamdouar,weidi,az}@robots.ox.ac.uk" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "text", + "content": "Not all camouflages are equally effective, as even a partially visible contour or a slight color difference can make the animal stand out and break its camouflage. In this paper, we address the question of what makes a camouflage successful, by proposing three scores for automatically assessing its effectiveness. In particular, we show that camouflage can be measured by the similarity between background and foreground features and boundary visibility. We use these camouflage scores to assess and compare all available camouflage datasets. We also incorporate the proposed camouflage score into a generative model as an auxiliary loss and show that effective camouflage images or videos can be synthesised in a scalable manner. The generated synthetic dataset is used to train a transformer-based model for segmenting camouflaged animals in videos. Experimentally, we demonstrate state-of-the-art camouflage breaking performance on the public MoCA-Mask benchmark." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 465, + 128, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 128, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 128, + 478 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 485, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 653 + ], + "type": "text", + "content": "Camouflage has long been a subject of interest and fascination for the scientific community, especially evolutionary biologists who consider it as an excellent example of species adaptation. In order to confuse predators or to hide from prey, and increase their chances of survival in their natural habitat, animals have developed numerous camouflage mechanisms, e.g., disruptive coloration and background matching. Some species have even evolved to develop an adaptive camouflage, e.g., an arctic fox loses its white fur to better match the brown grey of the new season's landscape. Perhaps the most dramatic camouflage adaptation is the cuttlefish; it changes its patterns dynamically and rapidly as it moves from one spot to the other, constantly adapting and improving its camouflage." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "This search for optimal camouflage has inspired numerous works in the computer vision community, such as [12, 27], that tackled camouflage as a problem of optimal texture synthesis, making 3D objects non-detectable in a given scene. Others have addressed camouflage as a highly" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 212, + 385, + 271 + ], + "blocks": [ + { + "bbox": [ + 307, + 212, + 385, + 271 + ], + "lines": [ + { + "bbox": [ + 307, + 212, + 385, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 385, + 271 + ], + "type": "image", + "image_path": "acdb0323b3a207303092250111f36da1588f1f7b145b21d66b78260c3a8276ef.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 273, + 350, + 281 + ], + "lines": [ + { + "bbox": [ + 343, + 273, + 350, + 281 + ], + "spans": [ + { + "bbox": [ + 343, + 273, + 350, + 281 + ], + "type": "text", + "content": "a" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 387, + 212, + 465, + 271 + ], + "blocks": [ + { + "bbox": [ + 387, + 212, + 465, + 271 + ], + "lines": [ + { + "bbox": [ + 387, + 212, + 465, + 271 + ], + "spans": [ + { + "bbox": [ + 387, + 212, + 465, + 271 + ], + "type": "image", + "image_path": "758a7bb71018f45b0f7b16043e3b56ce11b372d43202f6f0e7276681a2c98c14.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 272, + 428, + 281 + ], + "lines": [ + { + "bbox": [ + 421, + 272, + 428, + 281 + ], + "spans": [ + { + "bbox": [ + 421, + 272, + 428, + 281 + ], + "type": "text", + "content": "b" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 467, + 212, + 545, + 271 + ], + "blocks": [ + { + "bbox": [ + 467, + 212, + 545, + 271 + ], + "lines": [ + { + "bbox": [ + 467, + 212, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 467, + 212, + 545, + 271 + ], + "type": "image", + "image_path": "d6dfef7ca4e28c02701f98c863384965930bab9386775378b22ed1ec4136de47.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 273, + 509, + 281 + ], + "lines": [ + { + "bbox": [ + 503, + 273, + 509, + 281 + ], + "spans": [ + { + "bbox": [ + 503, + 273, + 509, + 281 + ], + "type": "text", + "content": "C" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 388, + 287, + 460, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 287, + 460, + 301 + ], + "spans": [ + { + "bbox": [ + 388, + 287, + 460, + 301 + ], + "type": "text", + "content": "score function" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 354, + 310, + 485, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 310, + 485, + 321 + ], + "spans": [ + { + "bbox": [ + 354, + 310, + 485, + 321 + ], + "type": "text", + "content": "camouflageranking: " + }, + { + "bbox": [ + 354, + 310, + 485, + 321 + ], + "type": "inline_equation", + "content": "c > a > b" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 329, + 547, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 547, + 414 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 547, + 414 + ], + "type": "text", + "content": "Figure 1: The three images depict the same animal, an Arctic fox, as it adapts its appearance to better blend with the changing landscape of the new season. While images " + }, + { + "bbox": [ + 304, + 329, + 547, + 414 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 329, + 547, + 414 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 329, + 547, + 414 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 329, + 547, + 414 + ], + "type": "text", + "content": " exhibit better background similarity than image " + }, + { + "bbox": [ + 304, + 329, + 547, + 414 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 329, + 547, + 414 + ], + "type": "text", + "content": ", the fox boundary is more visible in image " + }, + { + "bbox": [ + 304, + 329, + 547, + 414 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 329, + 547, + 414 + ], + "type": "text", + "content": " than image " + }, + { + "bbox": [ + 304, + 329, + 547, + 414 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 329, + 547, + 414 + ], + "type": "text", + "content": ". We assess the effectiveness of camouflage by measuring the degree of ambiguity it creates with respect to its background." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 424, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 547, + 605 + ], + "type": "text", + "content": "challenging object segmentation task in [9, 15, 20, 23]. Efforts have been made in collecting large scale camouflage datasets [9, 20] with costly annotation. In fact, camouflaged animals often exhibit complex shapes and thin structures that add to the boundary ambiguity and make the annotation highly time-consuming. Fan et al. report up to " + }, + { + "bbox": [ + 304, + 424, + 547, + 605 + ], + "type": "inline_equation", + "content": "60\\mathrm{min}" + }, + { + "bbox": [ + 304, + 424, + 547, + 605 + ], + "type": "text", + "content": " per image to provide accurate pixel-level annotations for their dataset COD10K [9]. While another line of research turned to camouflage breaking in sequences by taking advantage of motion cues [1, 5, 18, 19, 40], the camouflage data scarcity is even more extreme for videos. Recently, Sim2Real [18, 39] training has shown to be very effective for motion segmentation. By training on the optical flows of synthetic videos, these models can generalise to real videos without suffering from the domain gaps." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we start by asking the question: \"what are the properties of a successful camouflage?\" To answer this question, we investigate three scoring functions for quantifying the effectiveness of camouflage, namely, reconstruction fidelity score " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "(S_{R_f})" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", boundary score " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "(S_b)" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " and intra-image Fréchet score " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "(d_{\\mathcal{F}})" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ". These scores are later adopted for two critical roles, (i) they are used to assess the relevance of existing camouflage datasets and act as quality-indicator in collecting new camouflage data; (ii)" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 297, + 732, + 313, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 732, + 313, + 742 + ], + "spans": [ + { + "bbox": [ + 297, + 732, + 313, + 742 + ], + "type": "text", + "content": "832" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "they can be used as a proxy loss for image and video inpainting/generation, where we establish a synthetic camouflage data generation pipeline with a Generative Adversarial Network (GAN), that can simultaneously generate high-quality camouflage examples and masks of the camouflaged animals. We further train a Transformer-based architecture on the generated camouflage video sequences, and demonstrate a performance boost over training on only the (small scale) real data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "text", + "content": "In summary, we make the following contributions: First, we introduce three scoring functions to measure the effectiveness of a given camouflage in still images and videos. We use these camouflage scores to rank the images in existing datasets in terms of camouflage success, and also show that the rankings are largely in accordance with human-produced rankings. Second, we incorporate the camouflage score into a generative model, establishing a scalable pipeline for generating high-quality camouflage images or videos, along with the pixelwise segmentation mask of the camouflaged animals; Third, we show that a Transformer-based model trained on the synthetically generated data can achieve state-of-the-art performance on the MoCA-Mask video camouflaged animal segmentation benchmark." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 365, + 132, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 132, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 132, + 376 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 386, + 289, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 289, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 289, + 567 + ], + "type": "text", + "content": "Camouflage Evaluation. Although there are no previous computational works that directly assess camouflage, as far as we know, there have been a number of human studies. Previous works proposed methods to evaluate camouflage by analysing the human viewers' eye movements [22, 37] or conducting subjective perceptual experiments [12, 27, 32]. In [12, 27], participants were asked to indicate the camouflaged object and their answers were analysed in terms of accuracy and time needed per image. Similarly, Skurowski et al. [32] asked human volunteers to rate CHAMELEON [32] images from 1 to 5 in terms of camouflage effectiveness. They produced a score after compensating for personal bias. We validate our proposed camouflage scoring functions by comparing our rankings to these human-based rankings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "Motion Segmentation. The goal of this task is to partition the frames of a video sequence into background and independently moving objects. Early approaches focused on clustering motion patterns by grouping pixels with similar trajectories [3, 17, 21, 24, 25, 26, 31]. Another line of research tackled the problem by compensating for the background motion [1, 2, 19], via registering consecutive frames, or explicitly estimating camera motion. More recently, in [40], a Transformer-like architecture is used to reconstruct the input flows, and the segmentation masks are generated as a side product, while [6] decomposes the flow filed into regions by fitting affine transformations. The most" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "closely related to ours is [18, 39], where the authors adopt a Sim2Real strategy by training the model on optical flow computed from synthetic videos. These models can generalise towards real videos without fine-tuning. In this paper, we train a model on synthetic camouflage RGB sequences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 135, + 546, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 135, + 546, + 351 + ], + "spans": [ + { + "bbox": [ + 304, + 135, + 546, + 351 + ], + "type": "text", + "content": "Frechet Inception Distance and its variants. Also known as Wasserstein-2 distance [38], Fréchet distance (FD) is a metric quantifying the difference between two probability distributions [10]. Lately, [13] introduced the Fréchet Inception Distance (FID) in the context of generative models. Under the assumption that real images and generated images follow Gaussian distributions, they compute the FD of the two distributions. Note that here, each image is represented by a vector obtained from the last pooling layer of InceptionV3 [35]. Following their steps, [30] adapted FID to the image re-targeting task and introduced the Single Image Fréchet Inception Distance (SIFID). Instead of comparing entire image datasets, they compare two images and consider the distributions of their feature vectors obtained from earlier layer of InceptionV3 [35]. Recently, [7] included FID in the training objective and introduced Fréchet-GAN. Our work takes inspiration from [7, 30] and investigates a FID of regions within the same image as an auxiliary loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 361, + 536, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 536, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 536, + 376 + ], + "type": "text", + "content": "3. Measuring the effectiveness of camouflage" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 382, + 545, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 545, + 478 + ], + "type": "text", + "content": "Assuming there exists a dataset, " + }, + { + "bbox": [ + 304, + 382, + 545, + 478 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(I_1,m_1),\\ldots ," + }, + { + "bbox": [ + 304, + 382, + 545, + 478 + ], + "type": "inline_equation", + "content": "(I_N,m_N)\\}" + }, + { + "bbox": [ + 304, + 382, + 545, + 478 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 304, + 382, + 545, + 478 + ], + "type": "inline_equation", + "content": "I_{i}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 304, + 382, + 545, + 478 + ], + "type": "text", + "content": " refers to the images, and " + }, + { + "bbox": [ + 304, + 382, + 545, + 478 + ], + "type": "inline_equation", + "content": "m_i\\in \\{0,1\\}^{H\\times W\\times 1}" + }, + { + "bbox": [ + 304, + 382, + 545, + 478 + ], + "type": "text", + "content": " denotes a binary segmentation mask of the camouflaged animal. Here, we aim to design a scoring function that takes the image and segmentation mask as input, and outputs a scalar value " + }, + { + "bbox": [ + 304, + 382, + 545, + 478 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 382, + 545, + 478 + ], + "type": "text", + "content": " that can quantify the effectiveness of camouflage, i.e., how successfully an animal blends into its background, termed as the camouflage score," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 369, + 499, + 482, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 499, + 482, + 514 + ], + "spans": [ + { + "bbox": [ + 369, + 499, + 482, + 514 + ], + "type": "interline_equation", + "content": "S: (I _ {i}, m _ {i}) \\longmapsto s _ {i} \\in [ 0, 1 ]", + "image_path": "44ebb5904f36de3d756636697ad5e2977477f9439bd9124b00bcc1ca1785fab6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": "specifically, for " + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "inline_equation", + "content": "s_i < s_j" + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": ", the animal in " + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": ", indicated by the mask " + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": ", is more concealed than the animal in " + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": ", indicated by the mask " + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": ". Having such a scoring function enables various applications, first, it enables to rank the images from one dataset, in terms of the effectiveness of camouflage, and generate camouflage-relevant statistics for the dataset; second, it can serve as the objective for low-level image generation tasks, for example, image inpainting, editing, etc." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 715 + ], + "type": "text", + "content": "In the following sections, we investigate three such scoring functions, specifically, we exploit the animal's mask and quantify the key aspects that determine the camouflage success both perceptually (Sec. 3.1) and probabilistically (Sec. 3.2). Note that, we take into account the local aspect of camouflage, i.e., instead of processing the entire background region, we only focus on the immediate" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 732, + 312, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 732, + 312, + 742 + ], + "spans": [ + { + "bbox": [ + 297, + 732, + 312, + 742 + ], + "type": "text", + "content": "833" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 108, + 152 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 108, + 152 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 108, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 108, + 152 + ], + "type": "image", + "image_path": "0c6ff09a76b9d508514d200550ab067f87aa64e16710d61143837536fab88a82.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 154, + 82, + 161 + ], + "lines": [ + { + "bbox": [ + 76, + 154, + 82, + 161 + ], + "spans": [ + { + "bbox": [ + 76, + 154, + 82, + 161 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 110, + 72, + 169, + 112 + ], + "blocks": [ + { + "bbox": [ + 110, + 72, + 169, + 112 + ], + "lines": [ + { + "bbox": [ + 110, + 72, + 169, + 112 + ], + "spans": [ + { + "bbox": [ + 110, + 72, + 169, + 112 + ], + "type": "image", + "image_path": "1759ca146b405d8dabfef343fbec1b5c33972393ce87ede6d026804762e706a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 110, + 113, + 169, + 152 + ], + "blocks": [ + { + "bbox": [ + 110, + 113, + 169, + 152 + ], + "lines": [ + { + "bbox": [ + 110, + 113, + 169, + 152 + ], + "spans": [ + { + "bbox": [ + 110, + 113, + 169, + 152 + ], + "type": "image", + "image_path": "6085a984a7ccbde1dcba2ab54db69eed384d4fcc384e1f928c38adbc40d1f684.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 154, + 145, + 164 + ], + "lines": [ + { + "bbox": [ + 134, + 154, + 145, + 164 + ], + "spans": [ + { + "bbox": [ + 134, + 154, + 145, + 164 + ], + "type": "inline_equation", + "content": "I_{bg}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 171, + 72, + 230, + 111 + ], + "blocks": [ + { + "bbox": [ + 171, + 72, + 230, + 111 + ], + "lines": [ + { + "bbox": [ + 171, + 72, + 230, + 111 + ], + "spans": [ + { + "bbox": [ + 171, + 72, + 230, + 111 + ], + "type": "image", + "image_path": "f0ddbae7d3aaaf399ac03767840eb5b6ae4b66acf19036927fb7ef5d19cb4ea4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 171, + 113, + 230, + 152 + ], + "blocks": [ + { + "bbox": [ + 171, + 113, + 230, + 152 + ], + "lines": [ + { + "bbox": [ + 171, + 113, + 230, + 152 + ], + "spans": [ + { + "bbox": [ + 171, + 113, + 230, + 152 + ], + "type": "image", + "image_path": "49bb0f76d2411510dcabfc805a9c12eac5e5abacf324da3fc03f2e361baa5415.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 293, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 293, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 293, + 255 + ], + "type": "text", + "content": "Figure 2: The reconstruction fidelity score " + }, + { + "bbox": [ + 46, + 171, + 293, + 255 + ], + "type": "inline_equation", + "content": "S_{R_f}" + }, + { + "bbox": [ + 46, + 171, + 293, + 255 + ], + "type": "text", + "content": " evaluates the similarities between the original foreground " + }, + { + "bbox": [ + 46, + 171, + 293, + 255 + ], + "type": "inline_equation", + "content": "I_{fg}" + }, + { + "bbox": [ + 46, + 171, + 293, + 255 + ], + "type": "text", + "content": " and its reconstruction from background features " + }, + { + "bbox": [ + 46, + 171, + 293, + 255 + ], + "type": "inline_equation", + "content": "\\Psi_{bg}(I_{fg})" + }, + { + "bbox": [ + 46, + 171, + 293, + 255 + ], + "type": "text", + "content": ". The top example shows a case where the animal exhibits different visual patterns (color) from its background " + }, + { + "bbox": [ + 46, + 171, + 293, + 255 + ], + "type": "inline_equation", + "content": "(S_{R_f} = 0.11)" + }, + { + "bbox": [ + 46, + 171, + 293, + 255 + ], + "type": "text", + "content": ", while the bottom example shows a better background matching " + }, + { + "bbox": [ + 46, + 171, + 293, + 255 + ], + "type": "inline_equation", + "content": "(S_{R_f} = 0.82)" + }, + { + "bbox": [ + 46, + 171, + 293, + 255 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 231, + 72, + 291, + 112 + ], + "blocks": [ + { + "bbox": [ + 231, + 72, + 291, + 112 + ], + "lines": [ + { + "bbox": [ + 231, + 72, + 291, + 112 + ], + "spans": [ + { + "bbox": [ + 231, + 72, + 291, + 112 + ], + "type": "image", + "image_path": "81dfcb9dc60e8f7c80ec6ba0f4ae8c4c422310ffc4a6bb754bd6a2835efa3078.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 231, + 113, + 291, + 152 + ], + "blocks": [ + { + "bbox": [ + 231, + 113, + 291, + 152 + ], + "lines": [ + { + "bbox": [ + 231, + 113, + 291, + 152 + ], + "spans": [ + { + "bbox": [ + 231, + 113, + 291, + 152 + ], + "type": "image", + "image_path": "71fa186d169a144de8fb90476ef32242a700c72953a1d9f26fc7551798689c07.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 154, + 276, + 164 + ], + "lines": [ + { + "bbox": [ + 246, + 154, + 276, + 164 + ], + "spans": [ + { + "bbox": [ + 246, + 154, + 276, + 164 + ], + "type": "inline_equation", + "content": "\\Psi_{bg}(I_{fg})" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 269, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 305 + ], + "type": "text", + "content": "surround of the animal. Hence, when referring to background, we mean the local background and only consider the cropped images centred around the object." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 312, + 203, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 203, + 325 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 203, + 325 + ], + "type": "text", + "content": "3.1. Perceptual camouflage score" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 330, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 287, + 378 + ], + "type": "text", + "content": "In this section, we define the camouflage score through a perceptual lens, specifically, we measure the foreground and background similarity (Sect. 3.1.1), along with the boundary visibility (Sect. 3.1.2)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 388, + 201, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 201, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 201, + 399 + ], + "type": "text", + "content": "3.1.1 Reconstruction fidelity score" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 407, + 287, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 478 + ], + "type": "text", + "content": "We attempt to reconstruct the foreground region with patches from the background, and propose a score to quantitatively measure the discrepancy between the original image and its reconstruction. Intuitively, for successful camouflage, we should be able to reconstruct the foreground perfectly by copying patches from its close neighborhood." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 479, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 537 + ], + "type": "text", + "content": "Formally, for a given image " + }, + { + "bbox": [ + 46, + 479, + 287, + 537 + ], + "type": "inline_equation", + "content": "(I)" + }, + { + "bbox": [ + 46, + 479, + 287, + 537 + ], + "type": "text", + "content": " and segmentation mask " + }, + { + "bbox": [ + 46, + 479, + 287, + 537 + ], + "type": "inline_equation", + "content": "(m)" + }, + { + "bbox": [ + 46, + 479, + 287, + 537 + ], + "type": "text", + "content": ", subscript has been ignored for simplicity, we consider a trimap separation and define the foreground and background regions using morphological erosion and dilation of the mask," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 548, + 287, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 548, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 107, + 548, + 287, + 561 + ], + "type": "interline_equation", + "content": "I _ {\\mathrm {f g}}, I _ {\\mathrm {b g}} = m _ {\\mathrm {f g}} \\odot I, m _ {\\mathrm {b g}} \\odot I \\tag {1}", + "image_path": "431993cbb2be77408567b9dbb9d32769ff9b31a93d0b5588b66eb8b5d7daf69e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 570, + 251, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 251, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 251, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 570, + 251, + 582 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{fg}} = \\operatorname{erode}(m)" + }, + { + "bbox": [ + 47, + 570, + 251, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 570, + 251, + 582 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{bg}} = 1 - \\operatorname{dilate}(m)" + }, + { + "bbox": [ + 47, + 570, + 251, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": "To reconstruct the foreground region, we replace the foreground patch with the closest one in the background. Here a patch is a " + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": " pixel region (" + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "n = 7" + }, + { + "bbox": [ + 46, + 582, + 287, + 689 + ], + "type": "text", + "content": " in our case with an overlap of 3) and the patchwise similarity is computed by exploiting the low-level perceptual similarity, i.e., comparing corresponding RGB values. This reconstruction method is inspired by the texture generation and inpainting approaches, such as [8]. In practice, we implement it with fast approximate nearest neighbor search for efficiency." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "The reconstruction fidelity score is computed by assessing the difference value between the foreground region and" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 362, + 112 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 362, + 112 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 362, + 112 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 362, + 112 + ], + "type": "image", + "image_path": "1943e4b80c021397a9deb3c8d7e5c1c2515ae3ac3362302d2719334a9a40aa4e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 304, + 113, + 362, + 152 + ], + "blocks": [ + { + "bbox": [ + 304, + 113, + 362, + 152 + ], + "lines": [ + { + "bbox": [ + 304, + 113, + 362, + 152 + ], + "spans": [ + { + "bbox": [ + 304, + 113, + 362, + 152 + ], + "type": "image", + "image_path": "d7d53a5c03c2d669c20cc43dffa73bc5353dccf66a7254f19be6f04eac09352e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 154, + 337, + 162 + ], + "lines": [ + { + "bbox": [ + 331, + 154, + 337, + 162 + ], + "spans": [ + { + "bbox": [ + 331, + 154, + 337, + 162 + ], + "type": "text", + "content": "I" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 365, + 72, + 421, + 112 + ], + "blocks": [ + { + "bbox": [ + 365, + 72, + 421, + 112 + ], + "lines": [ + { + "bbox": [ + 365, + 72, + 421, + 112 + ], + "spans": [ + { + "bbox": [ + 365, + 72, + 421, + 112 + ], + "type": "image", + "image_path": "66fb6378b291b130de9d698e6f20bfc45618b095b2cb48a688fbf333b138f7c1.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 365, + 113, + 421, + 152 + ], + "blocks": [ + { + "bbox": [ + 365, + 113, + 421, + 152 + ], + "lines": [ + { + "bbox": [ + 365, + 113, + 421, + 152 + ], + "spans": [ + { + "bbox": [ + 365, + 113, + 421, + 152 + ], + "type": "image", + "image_path": "2a267669f36fb0477be3d48781fcada221b81ea83127b12a460e6fe96381523a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 154, + 408, + 163 + ], + "lines": [ + { + "bbox": [ + 380, + 154, + 408, + 163 + ], + "spans": [ + { + "bbox": [ + 380, + 154, + 408, + 163 + ], + "type": "text", + "content": "Trimap" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 423, + 72, + 480, + 112 + ], + "blocks": [ + { + "bbox": [ + 423, + 72, + 480, + 112 + ], + "lines": [ + { + "bbox": [ + 423, + 72, + 480, + 112 + ], + "spans": [ + { + "bbox": [ + 423, + 72, + 480, + 112 + ], + "type": "image", + "image_path": "13e2f1c90b7851132a4ee4256a4cf3d668266358acef757e31f5d6424b048dfc.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 423, + 113, + 481, + 152 + ], + "blocks": [ + { + "bbox": [ + 423, + 113, + 481, + 152 + ], + "lines": [ + { + "bbox": [ + 423, + 113, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 423, + 113, + 481, + 152 + ], + "type": "image", + "image_path": "5f03892f6c37fa35cee3086f08952931687805dcf804a6cd95ce6f3b8464cea7.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 154, + 453, + 161 + ], + "lines": [ + { + "bbox": [ + 446, + 154, + 453, + 161 + ], + "spans": [ + { + "bbox": [ + 446, + 154, + 453, + 161 + ], + "type": "text", + "content": "C" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 171, + 544, + 256 + ], + "lines": [ + { + "bbox": [ + 301, + 171, + 544, + 256 + ], + "spans": [ + { + "bbox": [ + 301, + 171, + 544, + 256 + ], + "type": "text", + "content": "Figure 3: Trimap regions and animal contours. From left to right: Input image; Trimap partition (foreground region in yellow, background in purple, and boundary region in green); Contours " + }, + { + "bbox": [ + 301, + 171, + 544, + 256 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 301, + 171, + 544, + 256 + ], + "type": "text", + "content": " computed on the original image; and ground truth Contours " + }, + { + "bbox": [ + 301, + 171, + 544, + 256 + ], + "type": "inline_equation", + "content": "C_{gt}" + }, + { + "bbox": [ + 301, + 171, + 544, + 256 + ], + "type": "text", + "content": ". The top example shows less contour agreement along the boundary region (" + }, + { + "bbox": [ + 301, + 171, + 544, + 256 + ], + "type": "inline_equation", + "content": "S_b = 0.72" + }, + { + "bbox": [ + 301, + 171, + 544, + 256 + ], + "type": "text", + "content": ") than the bottom example (" + }, + { + "bbox": [ + 301, + 171, + 544, + 256 + ], + "type": "inline_equation", + "content": "S_b = 0.40" + }, + { + "bbox": [ + 301, + 171, + 544, + 256 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 491, + 78, + 533, + 110 + ], + "blocks": [ + { + "bbox": [ + 491, + 78, + 533, + 110 + ], + "lines": [ + { + "bbox": [ + 491, + 78, + 533, + 110 + ], + "spans": [ + { + "bbox": [ + 491, + 78, + 533, + 110 + ], + "type": "image", + "image_path": "267e36e0db0341525401d83dbe0e77f5207a9175695eb9822a168d31da6e4d0e.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 482, + 114, + 533, + 147 + ], + "blocks": [ + { + "bbox": [ + 482, + 114, + 533, + 147 + ], + "lines": [ + { + "bbox": [ + 482, + 114, + 533, + 147 + ], + "spans": [ + { + "bbox": [ + 482, + 114, + 533, + 147 + ], + "type": "image", + "image_path": "24deecae6fa693f42c16a9961cc19aec933802e5cf6ce7705d4cab5b575995fe.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 154, + 521, + 163 + ], + "lines": [ + { + "bbox": [ + 509, + 154, + 521, + 163 + ], + "spans": [ + { + "bbox": [ + 509, + 154, + 521, + 163 + ], + "type": "inline_equation", + "content": "C_{gt}" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 269, + 546, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 546, + 304 + ], + "type": "text", + "content": "its reconstruction, specifically, we count the number of foreground pixels that have been successfully reconstructed from the background:" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 318, + 313, + 545, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 313, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 318, + 313, + 545, + 344 + ], + "type": "interline_equation", + "content": "S _ {R _ {f}} (I, m) = \\frac {1}{N _ {\\mathrm {f g}}} \\sum_ {(i, j) \\in I _ {\\mathrm {f g}}} R _ {f} (i, j) \\tag {2}", + "image_path": "a40c4eff825ff4453f10353b0996362af85a4aa927e7a7c3b48737fdf4e5f460.jpg" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 319, + 346, + 545, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 346, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 319, + 346, + 545, + 379 + ], + "type": "interline_equation", + "content": "R _ {f} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} | | I _ {\\mathrm {f g}} - \\Psi_ {I _ {\\mathrm {b g}}} \\left(I _ {\\mathrm {f g}}\\right) | | _ {2} < \\lambda | | I _ {\\mathrm {f g}} | | _ {2} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {3}", + "image_path": "48b1e459fb6004ce703737605a904bd8a95c71e6b8f005cc085a51efee184a32.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\Psi_{I_{bg}}(.)" + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": " denotes the reconstruction operation, " + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{fg}} = |m_{\\mathrm{fg}}|" + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": " is the total number of pixels in the foreground region, and " + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": " is a threshold (" + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\lambda = 0.2" + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": " in our case). A higher " + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "inline_equation", + "content": "S_{R_f}" + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": " score, means that the animal's visual attributes are well represented in the background. Conversely, a low " + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "inline_equation", + "content": "S_{R_f}" + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": " indicates that the animal's appearance is composed of unique features that makes it standout from its surrounding. In Fig. 2, we present examples of camouflage evaluation by reconstruction fidelity." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 493, + 444, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 493, + 444, + 505 + ], + "spans": [ + { + "bbox": [ + 306, + 493, + 444, + 505 + ], + "type": "text", + "content": "3.1.2 Boundary visibility score" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 513, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 545, + 620 + ], + "type": "text", + "content": "With the trimap representation, we can easily extract the boundary region, " + }, + { + "bbox": [ + 304, + 513, + 545, + 620 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{b}} = m_{\\mathrm{b}} \\odot I" + }, + { + "bbox": [ + 304, + 513, + 545, + 620 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 513, + 545, + 620 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{b}} = (1 - m_{\\mathrm{bg}}) - m_{\\mathrm{fg}}" + }, + { + "bbox": [ + 304, + 513, + 545, + 620 + ], + "type": "text", + "content": ". Here, we aim to measure the animal's boundary properties, e.g., contour visibility, as they can also provide visually evident cues for camouflage breaking. In particular, we adopt an off-the-shelf contour extraction method [29], and run it on the original image and foreground mask, to generate the images' contour (" + }, + { + "bbox": [ + 304, + 513, + 545, + 620 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 513, + 545, + 620 + ], + "type": "text", + "content": ") and ground truth animal's contour (" + }, + { + "bbox": [ + 304, + 513, + 545, + 620 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{gt}}" + }, + { + "bbox": [ + 304, + 513, + 545, + 620 + ], + "type": "text", + "content": "), as shown in Fig. 3." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 304, + 620, + 545, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 545, + 655 + ], + "type": "text", + "content": "We express the agreement between predicted contours " + }, + { + "bbox": [ + 304, + 620, + 545, + 655 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 620, + 545, + 655 + ], + "type": "text", + "content": " and ground truth contours " + }, + { + "bbox": [ + 304, + 620, + 545, + 655 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{gt}}" + }, + { + "bbox": [ + 304, + 620, + 545, + 655 + ], + "type": "text", + "content": " over the boundary region with the F1 metric:" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 342, + 666, + 545, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 666, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 342, + 666, + 545, + 680 + ], + "type": "interline_equation", + "content": "S _ {b} (I, m) = 1 - \\mathrm {F} 1 \\left(m _ {b} \\odot C _ {\\mathrm {g t}}, m _ {b} \\odot C\\right) \\tag {4}", + "image_path": "fa07a37e80a6776d082796050ae6dd3c23652083383f5bdf422458c7977aabbf.jpg" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "This score penalises the boundary pixels that are predicted as contour in both " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{gt}}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": ". While we do not expect a" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 732, + 312, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 732, + 312, + 742 + ], + "spans": [ + { + "bbox": [ + 297, + 732, + 312, + 742 + ], + "type": "text", + "content": "834" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "perfect contour agreement for visible edges, we consider " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "S_{b}" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " as a reasonable approximation, given the shape and size of the boundary region as a thin margin centred around the animal. If a boundary pixel is predicted as a contour in " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{gt}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " but not in " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": ", this means that the animal's edge is not visible in the original image. If a pixel is predicted as a contour in " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": " but not " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{gt}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": ", this indicates the presence of distracting elements, such as complex texture patterns on the foreground animal, which also affects the visibility of the animal's contour and therefore improves its camouflage. A perfectly camouflaged animal would have little contour agreement and " + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "inline_equation", + "content": "S_{b} = 1" + }, + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 288, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 288, + 253 + ], + "type": "text", + "content": "To get the final perceptual camouflage score, we linearly combine both reconstruction fidelity score and boundary visibility score:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 264, + 287, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 264, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 114, + 264, + 287, + 278 + ], + "type": "interline_equation", + "content": "S _ {\\alpha} = (1 - \\alpha) S _ {R _ {f}} + \\alpha S _ {b} \\tag {5}", + "image_path": "8c9005d5012445d2d4d51373fa5651271a20302cad5f4f4c48202257eca652b2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 289, + 287, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 287, + 313 + ], + "type": "text", + "content": "We will describe our method for setting the weighting parameter " + }, + { + "bbox": [ + 47, + 289, + 287, + 313 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 289, + 287, + 313 + ], + "type": "text", + "content": " in the Experiment section (Sec. 6)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 323, + 208, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 208, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 208, + 335 + ], + "type": "text", + "content": "3.2. Probabilistic scoring function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": "In addition to using the low-level RGB representation, in this section, we consider the pixelwise representation in the feature space, and propose a differentiable metric that compares the probabilistic distribution between the foreground and background regions, which acts as a proxy for the score " + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": " described in previous section, and can be used directly as a differentiable loss function in image generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 427, + 287, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 287, + 463 + ], + "type": "text", + "content": "We consider the Intra-Image Fréchet Inception Distance, specifically, we compute the feature maps for the foreground and background image regions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 474, + 287, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 474, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 287, + 488 + ], + "type": "interline_equation", + "content": "f _ {f g}, f _ {b g} = \\Phi_ {\\mathrm {v}} \\left(I _ {f g}\\right), \\Phi_ {\\mathrm {v}} \\left(I _ {b g}\\right) \\tag {6}", + "image_path": "9bbdc7c3f5db8d22e1f4d8881bff88373f9ff516cace6191fb7b473335601816.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{v}}(\\cdot)" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " denotes a pre-trained Inception network [34, 35], truncated at early layers. We take inspiration from [30], and adapt the single image Fréchet distance to measure the deviation between feature distributions of different regions within the same image. We adopt the Fréchet hypothesis with respect to our regions, i.e., the features of the foreground and background follow multivariate Gaussian distributions: " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{fg}} \\sim \\mathcal{N}(\\mu_{\\mathrm{fg}}, \\Sigma_{\\mathrm{fg}})" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{bg}} \\sim \\mathcal{N}(\\mu_{\\mathrm{bg}}, \\Sigma_{\\mathrm{bg}})" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": ". The intra-image Fréchet distance can be formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 628, + 293, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 293, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 293, + 643 + ], + "type": "interline_equation", + "content": "d _ {\\mathcal {F}} ^ {2} (I, m) = \\left\\| \\mu_ {\\mathrm {f g}} - \\mu_ {\\mathrm {b g}} \\right\\| _ {2} ^ {2} + T r \\left(\\Sigma_ {\\mathrm {f g}} + \\Sigma_ {\\mathrm {b g}} - 2 \\left(\\Sigma_ {\\mathrm {f g}} \\Sigma_ {\\mathrm {b g}}\\right) ^ {1 / 2}\\right)", + "image_path": "ebf2008f8c9f2bb467bbf7b1fa98d26e4113351dc98109bf8c0b4b582cc581d0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "Note that, when " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "(I,m)" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " is the output from a generative model, " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "d_{\\mathcal{F}}^2" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " is differentiable with respect to its parameters, and can therefore be used as an auxiliary loss for optimising the image generation procedure, i.e., generate effective camouflage examples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 524, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 524, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 524, + 85 + ], + "type": "text", + "content": "4. Generating camouflage video sequences" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 175 + ], + "type": "text", + "content": "In this section, we propose a scalable pipeline for generating images with concealed animals. In particular, we incorporate the differentiable intra-image Fréchet distance in image generation, explicitly optimising the camouflage score of the image. We first detail the static image generation framework, then describe how we use the images it generates to create camouflage video sequences." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 182, + 468, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 182, + 468, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 468, + 194 + ], + "type": "text", + "content": "4.1. Camouflage image generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 201, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 272 + ], + "type": "text", + "content": "In order to generate the image with a camouflaged animal, and its corresponding segmentation mask, we adopt a Generative Adversarial Network (GAN), where a generator is fed with a latent vector " + }, + { + "bbox": [ + 304, + 201, + 545, + 272 + ], + "type": "inline_equation", + "content": "z \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 304, + 201, + 545, + 272 + ], + "type": "text", + "content": " and learns to predict the pair of realistic image and segmentation mask, such that a discriminator cannot differentiate it from real samples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 273, + 547, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 547, + 344 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 273, + 547, + 344 + ], + "type": "inline_equation", + "content": "(x_{i},m_{i})" + }, + { + "bbox": [ + 304, + 273, + 547, + 344 + ], + "type": "text", + "content": " denote an image and segmentation pair sampled from the training dataset, a subset of COD10K [9] is used here. We adopt the conventional label notation, i.e. " + }, + { + "bbox": [ + 304, + 273, + 547, + 344 + ], + "type": "inline_equation", + "content": "y = 1" + }, + { + "bbox": [ + 304, + 273, + 547, + 344 + ], + "type": "text", + "content": " for (real) examples sampled from the training dataset and " + }, + { + "bbox": [ + 304, + 273, + 547, + 344 + ], + "type": "inline_equation", + "content": "y = 0" + }, + { + "bbox": [ + 304, + 273, + 547, + 344 + ], + "type": "text", + "content": " for (fake) examples from the generator, and consider the generative adversarial loss functions [11]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 352, + 558, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 352, + 558, + 366 + ], + "spans": [ + { + "bbox": [ + 306, + 352, + 558, + 366 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} = E _ {\\left(x _ {i}, m _ {i}\\right)} \\left[ - \\log \\left(D \\left(x _ {i}, m _ {i}\\right)\\right) \\right] + E _ {z} \\left[ - \\log \\left(1 - D (G (z))\\right) \\right]", + "image_path": "26a4092d43f83e22a740c6eabea60cb07f75b8a68ceef8d520f629c7ac66ac3d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 368, + 421, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 368, + 421, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 368, + 421, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G} = E _ {z} \\left[ - \\log (D (G (z))) \\right]", + "image_path": "8606bd1cae258f34489cb8d44d28935764fa26c6faa9e035a8f392f34dedeebc.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 388, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 472 + ], + "type": "text", + "content": "To enforce coherence between images and their segmentation masks, we feed the discriminator with additional fake pairs, consisting of real images coupled with unpaired real masks. This can be formulated as an additional coherence loss term that minimises the probability of assigning incorrect labels, i.e. " + }, + { + "bbox": [ + 304, + 388, + 545, + 472 + ], + "type": "inline_equation", + "content": "y = 1" + }, + { + "bbox": [ + 304, + 388, + 545, + 472 + ], + "type": "text", + "content": ", to " + }, + { + "bbox": [ + 304, + 388, + 545, + 472 + ], + "type": "inline_equation", + "content": "(x_{i},m_{j})" + }, + { + "bbox": [ + 304, + 388, + 545, + 472 + ], + "type": "text", + "content": " sampled from the training set such that " + }, + { + "bbox": [ + 304, + 388, + 545, + 472 + ], + "type": "inline_equation", + "content": "i\\neq j" + }, + { + "bbox": [ + 304, + 388, + 545, + 472 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 339, + 480, + 545, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 480, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 339, + 480, + 545, + 495 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C o h}} = E _ {\\left(x _ {i}, m _ {j}\\right) i \\neq j} [ - \\log (1 - D \\left(x _ {i}, m _ {j}\\right)) ] \\tag {7}", + "image_path": "7105bf90627d13d52235e70bae869315b68edec94547542f2b249547734c8c68.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 502, + 545, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 551 + ], + "type": "text", + "content": "To increase the camouflage effectiveness in the generated examples, we adopt the intra-image Fréchet distance in the generator loss, as an auxiliary loss term, and train our camouflage image generator with the following loss:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 343, + 558, + 545, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 558, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 343, + 558, + 545, + 573 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\tilde {D}} = \\mathcal {L} _ {D} + \\mathcal {L} _ {\\mathrm {C o h}} \\quad \\mathcal {L} _ {\\tilde {G}} = \\mathcal {L} _ {G} + \\beta d _ {\\mathcal {F}} ^ {2} \\tag {8}", + "image_path": "f223a7096c68dd373b0ea1c8626f3cb27a78c9b528c4dfefa17eccf904c36968.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 580, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 604 + ], + "type": "text", + "content": "We present an overview of our data generation framework in Fig. 4." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 611, + 465, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 465, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 465, + 624 + ], + "type": "text", + "content": "4.2. Camouflage video generation" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Given a camouflage image and corresponding segmentation mask, generated using the method above, we can create video sequences by applying different motions to the foreground and background, in a similar manner to [18]. Specifically, we first inpaint the backgrounds with an off-the-shelf model proposed by Suvorov et al. [33] and overlay the synthetic animal (extracted from the original generated image" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 732, + 312, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 732, + 312, + 742 + ], + "spans": [ + { + "bbox": [ + 297, + 732, + 312, + 742 + ], + "type": "text", + "content": "835" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 156, + 204 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 156, + 204 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 156, + 204 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 156, + 204 + ], + "type": "image", + "image_path": "43d5b3a4c21a12801039fe84cedce83142917067f3d79df01c89be4f126f27c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 211, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 45, + 211, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 45, + 211, + 545, + 270 + ], + "type": "text", + "content": "Figure 4: Our framework consists of a synthetic camouflage data generation pipeline (left), where we train a generator " + }, + { + "bbox": [ + 45, + 211, + 545, + 270 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 45, + 211, + 545, + 270 + ], + "type": "text", + "content": " in a GAN setting, to create camouflage images and masks, while encouraging camouflage effectiveness via minimising " + }, + { + "bbox": [ + 45, + 211, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{F}}" + }, + { + "bbox": [ + 45, + 211, + 545, + 270 + ], + "type": "text", + "content": ". The generated samples are then transformed into synthetic video sequences (middle), following our method presented in Sec. 4.2. The transformer-based motion segmentation architecture (right), for camouflage breaking in videos, is trained on the synthetic video sequences. The architecture is described in Sec. 5." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 158, + 70, + 232, + 204 + ], + "blocks": [ + { + "bbox": [ + 158, + 70, + 232, + 204 + ], + "lines": [ + { + "bbox": [ + 158, + 70, + 232, + 204 + ], + "spans": [ + { + "bbox": [ + 158, + 70, + 232, + 204 + ], + "type": "image", + "image_path": "87982e9137d9d52a0682cd4d20f167dcac222049a339fbd9973432f2a959dd3f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 232, + 70, + 522, + 204 + ], + "blocks": [ + { + "bbox": [ + 232, + 70, + 522, + 204 + ], + "lines": [ + { + "bbox": [ + 232, + 70, + 522, + 204 + ], + "spans": [ + { + "bbox": [ + 232, + 70, + 522, + 204 + ], + "type": "image", + "image_path": "273e9d1cc6704b62007721a334ae0c53e98ce159e2e471b72d76bee842bdacb1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 278, + 290, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 290, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 290, + 410 + ], + "type": "text", + "content": "by using the mask) at a random location within the image in the first frame, then following a random translational trajectory in the following frames. We incorporate a different translational motion for the background and include random sub-sequences where the foreground and background undergo the same motion to simulate momentarily static objects. As we train our generator with intra-image Fréchet loss " + }, + { + "bbox": [ + 46, + 278, + 290, + 410 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathcal{F}})" + }, + { + "bbox": [ + 46, + 278, + 290, + 410 + ], + "type": "text", + "content": ", the generated images exhibit strong similarities between the foreground and the background, and even if the object is placed at a different location from the original, we expect it to remain concealed within its surrounding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 432, + 217, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 217, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 217, + 445 + ], + "type": "text", + "content": "5. Learning to break camouflage" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 456, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 670 + ], + "type": "text", + "content": "In this section, we train a transformer-based architecture on the synthetic dataset, and demonstrate its effectiveness for breaking the camouflage in real videos, for example, MoCA [19]. We build on two previous architectures, namely, the motion segmentation from [18] and the Search Identification Network from [9]. While the first architecture processes sequences of optical flow, the second takes single RGB images as inputs and treats them separately. Our proposed model, shown in Fig. 4, takes both the optical flow and RGB frame sequences as input. The flow is computed with RAFT [36] from the RGB sequence, and processed with a motion encoder, followed by a transformer encoder. The RGB sequence is processed with the appearance encoder from [9], pre-trained for framewise camouflage segmentation. Then both streams are aggregated as inputs to a Mask2Former-like [4] transformer decoder and pixel decoder to produce the high resolution segmentation mask from the motion stream." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Motion encoder. To encode the motion cues, we use a light-weight convNet that takes as input a sequence of optical flows, " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathbf{m}} = \\{I_{m_1}, I_{m_2},.., I_{m_t}\\} \\in \\mathcal{R}^{t \\times c_0 \\times h \\times w}" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " and out" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 278, + 394, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 278, + 394, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 278, + 394, + 289 + ], + "type": "text", + "content": "puts motion features:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 355, + 303, + 496, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 303, + 496, + 317 + ], + "spans": [ + { + "bbox": [ + 355, + 303, + 496, + 317 + ], + "type": "interline_equation", + "content": "\\left\\{f _ {m _ {1}}, f _ {m _ {2}}, \\dots , f _ {m _ {t}} \\right\\} = \\Phi_ {\\text {m o t i o n}} \\left(\\mathbf {I} _ {\\mathbf {m}}\\right)", + "image_path": "75cfb4e8821694344e13b422555cb5a1d74426c55d6d655c77f825c23de2a910.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 330, + 497, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 497, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 497, + 342 + ], + "type": "text", + "content": "where each flow frame is separately embedded." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": "Motion transformer encoder. A transformer encoder takes as input the motion features, concatenated along the sequence, together with learned spatial and temporal positional encodings " + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "(\\mathrm{Pos}_s,\\mathrm{Pos}_t)" + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": ", as indicated in Fig 4. " + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\mathrm{Pos}_t" + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": " specifies the frame, and " + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "inline_equation", + "content": "\\mathrm{Pos}_s" + }, + { + "bbox": [ + 304, + 346, + 545, + 430 + ], + "type": "text", + "content": " specifies the position in the spatial feature map output of the motion encoder. The output of the transformer is a set of enriched motion features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 435, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 545, + 472 + ], + "type": "text", + "content": "Appearance encoder. Here, we adopt a SINet-v2 [9] architecture, that encodes the RGB sequence, " + }, + { + "bbox": [ + 306, + 435, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathbf{a}} = \\{I_{a_1}, I_{a_2},.., I_{a_t}\\} \\in \\mathcal{R}^{t \\times c_1 \\times h \\times w}" + }, + { + "bbox": [ + 306, + 435, + 545, + 472 + ], + "type": "text", + "content": " into appearance features:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 482, + 485, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 482, + 485, + 496 + ], + "spans": [ + { + "bbox": [ + 365, + 482, + 485, + 496 + ], + "type": "interline_equation", + "content": "\\left\\{f _ {a _ {1}}, f _ {a _ {2}}, \\dots , f _ {a _ {t}} \\right\\} = \\Phi_ {\\text {a p p}} \\left(\\mathbf {I} _ {\\mathbf {a}}\\right)", + "image_path": "063f16e40f16c597545c1b55e689dda3a51e0ce9475c2e8ba75d642d89b61669.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 509, + 539, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 539, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 539, + 521 + ], + "type": "text", + "content": "Again, each RGB frame is processed separately by SINet." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 525, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 609 + ], + "type": "text", + "content": "Transformer decoder. A transformer-based decoder takes the output of the motion transformer encoder together with the appearance features and a learnable query for the mask embedding. In a similar manner to Mask2Former [4], the query attends to multiple resolutions of the motion features concatenated with the appearance features and produces a mask embedding for the moving object." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 613, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 545, + 674 + ], + "type": "text", + "content": "Pixel Decoder. Similarly to the pixel decoder in Mask2Former, a light-weight convNet decoder is used with skip-connections to recover high-resolution segmentation masks " + }, + { + "bbox": [ + 304, + 613, + 545, + 674 + ], + "type": "inline_equation", + "content": "\\{\\hat{m}_1,\\hat{m}_2,\\dots ,\\hat{m}_t\\}" + }, + { + "bbox": [ + 304, + 613, + 545, + 674 + ], + "type": "text", + "content": " from the motion features and the mask embedding. This is shown as the blue box in Fig. 4." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Training objective. We train the motion segmentation model on our synthetic video sequences using the binary cross-entropy loss " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{BCE}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 733, + 313, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 733, + 313, + 742 + ], + "spans": [ + { + "bbox": [ + 297, + 733, + 313, + 742 + ], + "type": "text", + "content": "836" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 128, + 85 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 92, + 287, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 92, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 92, + 287, + 152 + ], + "type": "text", + "content": "In this section, we start by introducing the datasets involved in this paper, followed by the implementation details. Our experiments present a thorough analysis of the proposed camouflage scores and demonstrate their effectiveness in our training framework." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 164, + 110, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 164, + 110, + 176 + ], + "spans": [ + { + "bbox": [ + 47, + 164, + 110, + 176 + ], + "type": "text", + "content": "6.1. Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 184, + 287, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 184, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 184, + 287, + 232 + ], + "type": "text", + "content": "Here we describe the publicly available camouflage datasets that we included in our experiments, as shown in Fig. 5, as well as the synthetic camouflage datasets that we generated." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 235, + 287, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 235, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 235, + 287, + 283 + ], + "type": "text", + "content": "CHAMELEON [32] is one of the first camouflage image datasets. It contains 76 images collected using Google image search with the 'camouflaged animal' keyword and include ground truth manual annotations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 286, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 286, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 287, + 346 + ], + "type": "text", + "content": "CAMO [20]. The Camouflaged Object dataset consists of 2500 images collected from the internet, of which 1250 images (1000 training sub-set and 250 testing sub-set) contain at least one camouflage instance from 8 categories with manual pixelwise annotations provided." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 350, + 287, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 287, + 421 + ], + "type": "text", + "content": "COD10K [9]. COD10K contains 10,000 images collected from photography websites of which 5066 depict camouflaged animals (3040 training sub-set and 2026 testing subset), organised into 10 classes, and 78 sub-classes (69 camouflaged). Note that, in our camouflage evaluation experiments we only used the camouflage subset of this dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 425, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 425, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 425, + 287, + 449 + ], + "type": "text", + "content": "Camouflaged Animals [1]. This dataset consists of 9 video sequences of camouflage animals from 6 categories." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 452, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 452, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 47, + 452, + 287, + 535 + ], + "type": "text", + "content": "MoCA [19] is the first large-scale camouflage video dataset. It contains 141 video sequences (37K frames) totalling 67 animal categories. Recently, other works have curated this dataset by selecting only the videos with more prominent motion (locomotion) [40], and others have provided dense pixel annotation in MoCA-Mask [5]. We use the latter version in our experiments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 539, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 287, + 609 + ], + "type": "text", + "content": "Camouflaged cuboids [12, 27]. This dataset was created for texture synthesis and consists of multiple-view scenes, where cuboids were placed at a predefined location then synthetically camouflaged. In our evaluation, we consider the 4-views generated textures [12] from 36 scenes as well as the cuboids masks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 614, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 287, + 662 + ], + "type": "text", + "content": "Synthetic Camouflage Images. Using the method described in Sec. 4.1, we generate a synthetic camouflage dataset, of " + }, + { + "bbox": [ + 47, + 614, + 287, + 662 + ], + "type": "inline_equation", + "content": "5K" + }, + { + "bbox": [ + 47, + 614, + 287, + 662 + ], + "type": "text", + "content": " frames, by discriminating against real camouflage image and annotation masks from COD10K." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "content": "Synthetic Camouflage Videos. We generate " + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "inline_equation", + "content": "1K" + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "content": " sequences of 30 frames each, incorporating static sequences using the framework described in Sec. 4.2. We split these into 800 sequences for training and 200 for testing." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 305, + 76, + 544, + 571 + ], + "blocks": [ + { + "bbox": [ + 305, + 76, + 544, + 571 + ], + "lines": [ + { + "bbox": [ + 305, + 76, + 544, + 571 + ], + "spans": [ + { + "bbox": [ + 305, + 76, + 544, + 571 + ], + "type": "image", + "image_path": "6f1b01b60fa09cb7a1475102ac7312ac02003bd5e91c5053782ebfe5639ed8d2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 578, + 545, + 651 + ], + "lines": [ + { + "bbox": [ + 305, + 578, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 305, + 578, + 545, + 651 + ], + "type": "text", + "content": "Figure 5: Randomly sampled examples from the camouflage datasets included in our work. For the video datasets, Camouflaged Animals and MoCA-Mask, we show an example sequence. For the multiple-view dataset Camouflaged cuboids (bottom), we show example views from a scene with the 4-view texture synthesis method from [12]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 658, + 437, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 437, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 437, + 671 + ], + "type": "text", + "content": "6.2. Implementation details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "In our experiments, for a given camouflage example, the kernels for the morphological operations are selected from a range of values [1,10], so that the resulting annotation mask" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 732, + 312, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 732, + 312, + 741 + ], + "spans": [ + { + "bbox": [ + 297, + 732, + 312, + 741 + ], + "type": "text", + "content": "837" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "is reduced by " + }, + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": " for erosion and extended by " + }, + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": " for dilation. This allows the erosion/dilation to be adapted to the size of the camouflaged animal, while always keeping a reasonably large region for addressing the boundary effects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 129, + 287, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 129, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 129, + 287, + 224 + ], + "type": "text", + "content": "For our synthetic camouflage image generation, we adopt a Style-GAN architecture [41] and train it on the camouflage images from COD10K to generate " + }, + { + "bbox": [ + 46, + 129, + 287, + 224 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 129, + 287, + 224 + ], + "type": "text", + "content": " images and masks pairs. We use " + }, + { + "bbox": [ + 46, + 129, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\beta = 0.1" + }, + { + "bbox": [ + 46, + 129, + 287, + 224 + ], + "type": "text", + "content": " to weight our intra-image Fréchet auxiliary loss. For numerical stability, we adopt the Newton Schultz iterations when calculating the matrix square root term in Intra-Image Fréchet Inception Distance, with " + }, + { + "bbox": [ + 46, + 129, + 287, + 224 + ], + "type": "inline_equation", + "content": "T = 100" + }, + { + "bbox": [ + 46, + 129, + 287, + 224 + ], + "type": "text", + "content": " iterations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 227, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 287, + 300 + ], + "type": "text", + "content": "When generating the image sequences, we compute optical flows with RAFT [36], and train our moving camouflaged animal segmentation model. We use batches of size 2 and sequences of size 5. We first train on our synthetically generated dataset with a learning rate of " + }, + { + "bbox": [ + 46, + 227, + 287, + 300 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 227, + 287, + 300 + ], + "type": "text", + "content": " for 500 iterations then fine-tune on the training subset of MoCA-Mask." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 307, + 105, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 105, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 105, + 319 + ], + "type": "text", + "content": "6.3. Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 326, + 287, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 363 + ], + "type": "text", + "content": "This section presents qualitative and quantitative results, to demonstrate the effectiveness of our score functions and the benefit of including " + }, + { + "bbox": [ + 46, + 326, + 287, + 363 + ], + "type": "inline_equation", + "content": "d_{\\mathcal{F}}^2" + }, + { + "bbox": [ + 46, + 326, + 287, + 363 + ], + "type": "text", + "content": " in the training loop." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 369, + 246, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 246, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 246, + 381 + ], + "type": "text", + "content": "6.3.1 Evaluation on camouflage effectiveness" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 388, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 288, + 700 + ], + "type": "text", + "content": "Ranking camouflage images in COD10K. We can rank the images based on the effectiveness of the animal's camouflage, with our proposed scoring functions. In the left part of Fig. 6, we show the four best scored images from the large scale COD10K dataset. We can make the following observations: (i) the best examples with respect to " + }, + { + "bbox": [ + 46, + 388, + 288, + 700 + ], + "type": "inline_equation", + "content": "S_{R_f}" + }, + { + "bbox": [ + 46, + 388, + 288, + 700 + ], + "type": "text", + "content": " (top), which favours the background matching of the animals without taking into account the boundary region, all exhibit visible boundaries, with the exception of image (a). This is especially noticeable in the rabbit example (d) along the ears and shadow regions; (ii) the best four examples with respect to the boundary score " + }, + { + "bbox": [ + 46, + 388, + 288, + 700 + ], + "type": "inline_equation", + "content": "S_b" + }, + { + "bbox": [ + 46, + 388, + 288, + 700 + ], + "type": "text", + "content": " are all from the caterpillar subclass, mostly the baron type. These insects have thin, and transparent boundaries, that makes them the perfect candidates for this score. However, with the exception of (a) and (c), the animals still stands out as they exhibit colors and patterns that are not present within their background; (iii) the " + }, + { + "bbox": [ + 46, + 388, + 288, + 700 + ], + "type": "inline_equation", + "content": "S_\\alpha" + }, + { + "bbox": [ + 46, + 388, + 288, + 700 + ], + "type": "text", + "content": " combining both scores shows highly effective camouflage examples. In contrast, the right part of Fig. 6 shows the lowest scored images for each approach. For " + }, + { + "bbox": [ + 46, + 388, + 288, + 700 + ], + "type": "inline_equation", + "content": "S_{R_f}" + }, + { + "bbox": [ + 46, + 388, + 288, + 700 + ], + "type": "text", + "content": ", we find examples with low background matching in the first row, with two ant examples, i.e., (e), (f). While the boundary score penalises examples with high contrasts between the animal and its background, that results in more visible contours. This is also the case for the lowest " + }, + { + "bbox": [ + 46, + 388, + 288, + 700 + ], + "type": "inline_equation", + "content": "S_\\alpha" + }, + { + "bbox": [ + 46, + 388, + 288, + 700 + ], + "type": "text", + "content": " score examples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Dataset level camouflage comparison. We compute the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "camouflage score for all the camouflage image and video datasets and report the results in Tab. 1. For the image datasets, the dataset level score is computed as the mean of the image scores; for the video dataset the score is first computed per frame, and per-video average computed, then the dataset level score is the mean of the averages. Our experiments show that MoCA-Mask [5, 19] contains the most successful camouflages according to our scoring. While COD10K [9] subsets are balanced in terms of camouflage effectiveness, we find that, for CAMO [20], the test subset yields higher scores than the training subset and therefore contains more challenging examples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "content": "We note that Camouflaged cuboids dataset yields higher " + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "inline_equation", + "content": "S_{R_f}" + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 304, + 217, + 545, + 337 + ], + "type": "text", + "content": " scores compared to our synthetic datasets. This is due to the fact that the model used in [12, 27] is only trained to produce optimal texture for a predefined region (cuboid) in a particular location of a given scene. However, our model learns to output a new image with a randomly located camouflaged object of random shape, thus offering more scalability and diversity compared to the cuboids dataset which cannot be used to train a model for breaking real camouflage." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "text", + "content": "Comparison to human-produced rankings. We compare our rankings to the human scoring system based on the rating of CHAMELEON [32] and the time-to-find for Camouflaged cuboids [12]. To search for the optimal " + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "text", + "content": " parameter in Equation (5), we select 15 images from CHAMELEON and compare their " + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "text", + "content": " ranking with ground truth using kendl-" + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "text", + "content": " metric [16]. For " + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\alpha = 0.35" + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "text", + "content": ", we obtain kendl-" + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "inline_equation", + "content": "\\tau = 0.51" + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "text", + "content": " on this validation set that we excluded from our test reported in Tab. 2. We can draw the following observations: (i) we find that the boundary score " + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "inline_equation", + "content": "S_{b}" + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "text", + "content": " produces more agreement with the ground truth ranking than " + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "inline_equation", + "content": "S_{R_{f}}" + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "text", + "content": ", suggesting that human observers tend to pay more attention to contour visibility than background matching; (ii) while comparing to the ranking from Camouflaged cuboids, we found negative correlations for " + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "inline_equation", + "content": "S_{R_{f}}" + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "inline_equation", + "content": "d_{\\mathcal{F}}" + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "text", + "content": ", we conjecture that this may be due to the nature of the dataset in [12], i.e., all the textured cuboids are synthetically generated with very high background matching; (iii) for both experiments, we obtain the strongest agreement with the " + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 304, + 339, + 545, + 579 + ], + "type": "text", + "content": " score combining both background similarity and boundary blending." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Further analysis of the synthetic datasets. We adopt the FID [13] and IS [30] metrics in Table 4. These metrics assess: (i) the similarity between the set of synthetic images output by the generator and the set of real images used to train it (FID); and (ii) the clarity of the object and the diversity of images in terms of classes (IS). While they are not intended for measuring camouflage success, which is the focus of this work, IS could somewhat be (inversely) linked to " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{b}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " as the object clarity and boundary visibility are close. Note that the object clarity (IS) is decreased with " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{F}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " which is the intended purpose of improving camouflage, however," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 732, + 312, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 732, + 312, + 742 + ], + "spans": [ + { + "bbox": [ + 297, + 732, + 312, + 742 + ], + "type": "text", + "content": "838" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 72, + 520, + 262 + ], + "blocks": [ + { + "bbox": [ + 77, + 72, + 520, + 262 + ], + "lines": [ + { + "bbox": [ + 77, + 72, + 520, + 262 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 520, + 262 + ], + "type": "image", + "image_path": "baf0bf0fcb2fab0b8bf954cb197bfe3f0b254b42dd3957dbe58611c4affd46e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 270, + 547, + 296 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 547, + 296 + ], + "type": "text", + "content": "Figure 6: Left: Top-4 scored examples from COD10K for " + }, + { + "bbox": [ + 46, + 270, + 547, + 296 + ], + "type": "inline_equation", + "content": "S_{Rf}(top)" + }, + { + "bbox": [ + 46, + 270, + 547, + 296 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 270, + 547, + 296 + ], + "type": "inline_equation", + "content": "S_b(midle)" + }, + { + "bbox": [ + 46, + 270, + 547, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 270, + 547, + 296 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 46, + 270, + 547, + 296 + ], + "type": "text", + "content": " (bottom). Right: Lowest scored examples from COD10K. For each example, we show the obtained score and the corresponding ground truth mask." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 303, + 320, + 445 + ], + "blocks": [ + { + "bbox": [ + 48, + 303, + 320, + 445 + ], + "lines": [ + { + "bbox": [ + 48, + 303, + 320, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 303, + 320, + 445 + ], + "type": "table", + "html": "
DatasetsData typeSRf↑Sb↑Sα↑d2F↓
CHAMELEON [32]Image0.6940.4450.6070.70
CAMO Train [20]Image0.6720.4510.5951.01
CAMO Test [20]Image0.6830.4700.6081.00
COD10K Train [9]Image0.6550.4330.5770.90
COD10K Test [9]Image0.6570.4310.5780.90
Camouflaged Animals [1]Video0.6740.5360.6261.60
MoCA-Mask Train [5, 19]Video0.8500.4430.7071.14
MoCA-Mask Test [5, 19]Video0.7330.4640.6392.51
Camouflaged cuboids[12, 27]Multi-view0.8940.4330.7336.2
Syn. Camouflage w.o. L_FImage0.6080.4320.5461.36
Syn. Camouflage w. L_FImage0.6790.4470.5981.13
Syn. Camouflage VideoVideo0.6580.4300.5781.18
", + "image_path": "29f34b477c02d008cf88a5717e005baeb535c9fada4a858cf862256c470ca993.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 447, + 323, + 533 + ], + "lines": [ + { + "bbox": [ + 46, + 447, + 323, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 323, + 533 + ], + "type": "text", + "content": "Table 1: Results of the proposed scores on natural camouflage datasets (top) and synthetically generated camouflage datasets (bottom). We report the mean scores for the single image datasets. For the video and multiple view datasets, we compute the mean per sequence and scene respectively. Syn. Camouflage w. " + }, + { + "bbox": [ + 46, + 447, + 323, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{F}}" + }, + { + "bbox": [ + 46, + 447, + 323, + 533 + ], + "type": "text", + "content": " refers to the synthetic image dataset that we generated while minimising " + }, + { + "bbox": [ + 46, + 447, + 323, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{F}}" + }, + { + "bbox": [ + 46, + 447, + 323, + 533 + ], + "type": "text", + "content": " and Syn. Camouflage Video its sequence version." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 541, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 564 + ], + "type": "text", + "content": "this effect is not maintained in the sequence generation as the animal changes location within its background." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 50, + 565, + 291, + 611 + ], + "blocks": [ + { + "bbox": [ + 50, + 565, + 291, + 611 + ], + "lines": [ + { + "bbox": [ + 50, + 565, + 291, + 611 + ], + "spans": [ + { + "bbox": [ + 50, + 565, + 291, + 611 + ], + "type": "table", + "html": "
datasetFID↓IS↑
Syn. Camouflage w.o. L_F75.013.49
Syn. Camouflage w. L_F76.343.13
Syn. Camouflage Video86.423.85
", + "image_path": "bb44cf13ae0b476826ded89f74e7fe3658fe9898dae140f6d28404e688b20345.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 627, + 149, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 149, + 638 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 149, + 638 + ], + "type": "text", + "content": "6.3.2 Model ablations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "On the effectiveness of " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{F}}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": ". Fig. 7 shows the generated samples from the GAN pipeline, trained with and without the intra-image Fréchet Distance. Our generator is able to produce realistic object masks of complex shapes and thin structures similar to those usually encountered in the camouflage datasets. The produced images exhibit rich back" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 326, + 303, + 547, + 364 + ], + "blocks": [ + { + "bbox": [ + 61, + 613, + 272, + 624 + ], + "lines": [ + { + "bbox": [ + 61, + 613, + 272, + 624 + ], + "spans": [ + { + "bbox": [ + 61, + 613, + 272, + 624 + ], + "type": "text", + "content": "Table 4: Overall quality of our synthetic datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 303, + 547, + 364 + ], + "lines": [ + { + "bbox": [ + 326, + 303, + 547, + 364 + ], + "spans": [ + { + "bbox": [ + 326, + 303, + 547, + 364 + ], + "type": "table", + "html": "
ScoresCHAMELEON [32]Camouflaged cuboids[12, 27]
SRf0.01-0.07
Sb0.030.42
Sα0.420.41
d2F0.10-0.17
", + "image_path": "cbf4a818e2727880036dce2fb812edd28d8183909a428bb447da4c3205ab33a0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 328, + 394, + 542, + 482 + ], + "blocks": [ + { + "bbox": [ + 326, + 367, + 544, + 391 + ], + "lines": [ + { + "bbox": [ + 326, + 367, + 544, + 391 + ], + "spans": [ + { + "bbox": [ + 326, + 367, + 544, + 391 + ], + "type": "text", + "content": "Table 2: Kendall- " + }, + { + "bbox": [ + 326, + 367, + 544, + 391 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 326, + 367, + 544, + 391 + ], + "type": "text", + "content": " metric between rankings produced via our scores and human scoring ground truth." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 394, + 542, + 482 + ], + "lines": [ + { + "bbox": [ + 328, + 394, + 542, + 482 + ], + "spans": [ + { + "bbox": [ + 328, + 394, + 542, + 482 + ], + "type": "table", + "html": "
ModelTraining datasetAppearanceEncoderTransformerEncoderTransformerAggregationmIoU↑F↑E↑MAE↓
AS14.520.457.39.4
BMM15.320.657.35.1
CS+MM16.021.859.84.3
DS+MM✓20.522.659.83.8
ES+MM✓✓23.023.557.01.6
FS+MM✓✓22.723.258.83.6
GMM✓✓✓23.424.762.02.4
HS+MM✓✓✓30.834.377.01.8
", + "image_path": "fa68bab1bd615959477e3eaec9cb17f52959692beed33a6e93953d0818f5c6f3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 326, + 483, + 544, + 532 + ], + "lines": [ + { + "bbox": [ + 326, + 483, + 544, + 532 + ], + "spans": [ + { + "bbox": [ + 326, + 483, + 544, + 532 + ], + "type": "text", + "content": "Table 3: Ablation study on the different components of the motion segmentation architecture and training datasets. We evaluate on MoCA-Mask for all experiments. (S: our synthetic dataset, MM: MoCA-Mask)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 541, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 545, + 638 + ], + "type": "text", + "content": "grounds with realistic nature-like elements mimicking rocks and coral structures present in the real camouflage data. Adding the " + }, + { + "bbox": [ + 304, + 541, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{F}}" + }, + { + "bbox": [ + 304, + 541, + 545, + 638 + ], + "type": "text", + "content": " loss results in generating images with better blending in the background, both qualitatively and quantitatively, as shown by the " + }, + { + "bbox": [ + 304, + 541, + 545, + 638 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 304, + 541, + 545, + 638 + ], + "type": "text", + "content": " computed for both datasets generated with and without the intra-image Fréchet Distance in Tab. 1. We refer the reader to the supplementary material for more examples of generated samples." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 714 + ], + "type": "text", + "content": "On the architecture design and gain from the synthetic dataset. We present in Tab. 3 an ablation study detailing the gain from the main components of our architecture, as well as the benefit of first training on our synthetic dataset then fine-tuning on MoCA-Mask (S+MM), for model H, as opposed to only training on MoCA-Mask (MM), for G." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 732, + 313, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 732, + 313, + 742 + ], + "spans": [ + { + "bbox": [ + 297, + 732, + 313, + 742 + ], + "type": "text", + "content": "839" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 520, + 161 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 520, + 161 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 520, + 161 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 520, + 161 + ], + "type": "image", + "image_path": "7e56138070bfb2aa0549ae5bad0396b6546b425f700cba994dd51f61044fae60.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 167, + 538, + 179 + ], + "lines": [ + { + "bbox": [ + 53, + 167, + 538, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 167, + 538, + 179 + ], + "type": "text", + "content": "Figure 7: Generated camouflage images and masks: For the examples on the right, the generator was trained with " + }, + { + "bbox": [ + 53, + 167, + 538, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{F}}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 39, + 186, + 318, + 312 + ], + "blocks": [ + { + "bbox": [ + 39, + 186, + 318, + 312 + ], + "lines": [ + { + "bbox": [ + 39, + 186, + 318, + 312 + ], + "spans": [ + { + "bbox": [ + 39, + 186, + 318, + 312 + ], + "type": "image", + "image_path": "6c24a53ec9d6905bc1fa53804f0499382f71c07cfb142f3d4693ba064e87aee0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 316, + 303, + 353 + ], + "lines": [ + { + "bbox": [ + 53, + 316, + 303, + 353 + ], + "spans": [ + { + "bbox": [ + 53, + 316, + 303, + 353 + ], + "type": "text", + "content": "Figure 8: Qualitative results on MoCA-Mask. From top to bottom: Appearance sequence " + }, + { + "bbox": [ + 53, + 316, + 303, + 353 + ], + "type": "inline_equation", + "content": "I_{a}" + }, + { + "bbox": [ + 53, + 316, + 303, + 353 + ], + "type": "text", + "content": ", flow sequence " + }, + { + "bbox": [ + 53, + 316, + 303, + 353 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 53, + 316, + 303, + 353 + ], + "type": "text", + "content": ", predicted segmentations and ground truth segmentations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 322, + 189, + 544, + 311 + ], + "blocks": [ + { + "bbox": [ + 322, + 189, + 544, + 311 + ], + "lines": [ + { + "bbox": [ + 322, + 189, + 544, + 311 + ], + "spans": [ + { + "bbox": [ + 322, + 189, + 544, + 311 + ], + "type": "table", + "html": "
ModelRGBMotionmIoU↑F↑E↑MAE↓
SINet [9]✓20.223.169.92.8
SINet-V2 [9]✓18.020.464.23.1
SegMaR [14]✓12.222.580.34.5
ZoomNet [28]✓18.828.770.82.5
SLTNet [5]✓✓27.231.175.92.7
MG [40]✓12.716.856.16.7
Ours-flow✓17.821.560.73.7
Ours✓✓30.834.377.01.8
", + "image_path": "c27bf7424fdff0cb049118ebc43f0c8fff27bf0e3295f3e8086e840c5bb1832d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 320, + 316, + 546, + 352 + ], + "lines": [ + { + "bbox": [ + 320, + 316, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 320, + 316, + 546, + 352 + ], + "type": "text", + "content": "Table 5: Results on MoCA-Mask. Ours-flow refers to the optical flow only version of our architecture, excluding the appearance stream." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 363, + 155, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 155, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 155, + 374 + ], + "type": "text", + "content": "6.3.3 Results on MoCA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 384, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 515 + ], + "type": "text", + "content": "We train our model on the synthetically generated camouflage images and fine-tune on the training set of MoCA-Mask. Tab. 5 presents our camouflage object segmentation results on the test set of MoCA-Mask. Ours refers to our proposed method with SINet-V2 as the appearance encoder and Ours-flow refers to the optical flow only version of our architecture, excluding the appearance stream. Our main model outperforms RGB and Motion-based methods on MoCA-Mask. Fig. 8 presents qualitative results of our segmentation model. Note that our model is robust to degraded optical flow and static animals." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 525, + 124, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 124, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 124, + 537 + ], + "type": "text", + "content": "6.4. Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": "While our image generation method encourages background matching through the " + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{F}}" + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": " loss term minimization, the sequence generating is not guaranteed to maintain object concealment. A possible solution could be to use the proposed " + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "inline_equation", + "content": "S_{\\alpha}" + }, + { + "bbox": [ + 46, + 544, + 287, + 616 + ], + "type": "text", + "content": " camouflage score to curate the generated sequence dataset and filter out the most visible objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "Our proposed camouflage scores use the ground truth annotation and analyse the different regions it defines. We found that our scoring system penalises specific cases of camouflage by occlusion, where elements from the background are partially occluding the animal and helping improve its camouflage. For instance, the grass in example (g) from the last row of Fig. 6 is not treated as part of the animal and therefore not considered in our background similarity" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 363, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 398 + ], + "type": "text", + "content": "assessment. One can argue that this is due to the ambiguity in the provided annotations and for such cases, extra amodal annotation should be also considered." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 411, + 378, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 378, + 424 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 378, + 424 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 433, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 554 + ], + "type": "text", + "content": "We present three score functions for computationally assessing the effectiveness of camouflage in images and videos. By evaluating the similarity with background and the boundary visibility, our combined perceptual score is strongly correlated with human perceptual ranking systems on two different datasets. We demonstrate that training a generative model with our differentiable camouflage function improves the effectiveness of generated camouflage examples and can be used to generate challenging synthetic camouflage datasets to train models to break camouflage." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 563, + 395, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 395, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 395, + 575 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "We are grateful to Przemysław Skurowski for providing the human perception study data on the CHAMELEON dataset and Andrew Owens, Rui Guo and Oscar de Lima for providing the data for the Camouflaged cuboids dataset. We thank Tengda Han for fruitful discussions. This research is supported by the UK EPSRC funded CDT in Autonomous Intelligent Machines and Systems (AIMS), the EPSRC Programme Grant VisualAI EP/T028572/1, a Schlumberger Studentship, and a Royal Society Research Professorship. WX is supported by the National Key R&D Program of China (No. 2022ZD0161400)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 733, + 313, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 733, + 313, + 742 + ], + "spans": [ + { + "bbox": [ + 297, + 733, + 313, + 742 + ], + "type": "text", + "content": "840" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 710 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Pia Bideau and Erik Learned-Miller. It's moving! a probabilistic model for causal motion segmentation in moving camera videos. In ECCV, pages 433-449, 2016." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 286, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 286, + 156 + ], + "type": "text", + "content": "[2] Pia Bideau, Rakesh R Menon, and Erik Learned-Miller. Moa-net: self-supervised motion segmentation. In Proc. ECCV Workshop, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 156, + 286, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 156, + 286, + 187 + ], + "spans": [ + { + "bbox": [ + 53, + 156, + 286, + 187 + ], + "type": "text", + "content": "[3] Thomas Brox and Jitendra Malik. Object segmentation by long term analysis of point trajectories. In European conference on computer vision, pages 282-295, 2010." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 187, + 286, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 187, + 286, + 230 + ], + "spans": [ + { + "bbox": [ + 53, + 187, + 286, + 230 + ], + "type": "text", + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 230, + 286, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 286, + 273 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 286, + 273 + ], + "type": "text", + "content": "[5] Xuelian Cheng, Huan Xiong, Deng-Ping Fan, Yiran Zhong, Mehrtash Harandi, Tom Drummond, and Zongyuan Ge. Implicit motion handling for video camouflaged object detection. In CVPR, pages 13864-13873, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 274, + 286, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 286, + 316 + ], + "type": "text", + "content": "[6] Subhabrata Choudhury, Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Guess What Moves: Unsupervised Video and Image Segmentation by Anticipating Motion. In Proc. BMVC, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 316, + 286, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 316, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 316, + 286, + 358 + ], + "type": "text", + "content": "[7] Khoa D Doan, Saurav Manchanda, Fengjiao Wang, Sathiya Keerthi, Avradeep Bhowmik, and Chandan K Reddy. Image generation via minimizing frechet distance in discriminator feature space. arXiv preprint arXiv:2003.11774, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 358, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 286, + 379 + ], + "type": "text", + "content": "[8] A. Efros and T. Leung. Texture synthesis by non-parametric sampling. In Proc. ICCV, pages 1039-1046, Sep 1999." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 286, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 411 + ], + "type": "text", + "content": "[9] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 411, + 286, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 286, + 443 + ], + "type": "text", + "content": "[10] Maurice Fréchet. Sur la distance de deux lois de probabilité. Comptes Rendus Hebdomadaires des Seances de L'Académie des Sciences, pages 689-692, 1957." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 443, + 286, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 443, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 443, + 286, + 485 + ], + "type": "text", + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, pages 2672-2680, 2014." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 485, + 286, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 286, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 286, + 517 + ], + "type": "text", + "content": "[12] Rui Guo, Jasmine Collins, Oscar de Lima, and Andrew Owens. Ganmouflage: 3d object nondetection with texture fields. In arXiv preprint arXiv:2201.07202, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 517, + 286, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 517, + 286, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 517, + 286, + 559 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 560, + 286, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 560, + 286, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 560, + 286, + 602 + ], + "type": "text", + "content": "[14] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In CVPR, pages 4713-4722, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 602, + 286, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 286, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 286, + 645 + ], + "type": "text", + "content": "[15] Nobukatsu Kajiura, Hong Liu, and Shin'ichi Satoh. Improving camouflaged object detection with the uncertainty of pseudo-edge labels. In ACM Multimedia Asia, pages 1-7. 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 646, + 286, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 286, + 666 + ], + "type": "text", + "content": "[16] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1/2):81-93, 1938." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 667, + 286, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 667, + 286, + 699 + ], + "spans": [ + { + "bbox": [ + 48, + 667, + 286, + 699 + ], + "type": "text", + "content": "[17] Margret Keuper, Bjoern Andres, and Thomas Brox. Motion trajectory segmentation via minimum cost multicuts. In ICCV, pages 3271-3279, 2015." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 699, + 286, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 699, + 286, + 710 + ], + "spans": [ + { + "bbox": [ + 48, + 699, + 286, + 710 + ], + "type": "text", + "content": "[18] Hala Lamdouar, Weidi Xie, and Andrew Zisserman. Seg-" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 703 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 326, + 73, + 534, + 84 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 534, + 84 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 534, + 84 + ], + "type": "text", + "content": "menting invisible moving objects. In Proc. BMVC, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 84, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 84, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 308, + 84, + 545, + 115 + ], + "type": "text", + "content": "[19] Hala Lamdouar, Charig Yang, Weidi Xie, and Andrew Zisserman. Betrayed by motion: Camouflaged object discovery via motion segmentation. In Proc. ACCV, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 116, + 545, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 116, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 545, + 158 + ], + "type": "text", + "content": "[20] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, MinhTriet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer Vision and Image Understanding, pages 45-56, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 159, + 545, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 159, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 545, + 200 + ], + "type": "text", + "content": "[21] José Lezama, Karteek Alahari, Josef Sivic, and Ivan Laptev. Track to the future: Spatio-temporal video segmentation with long-range motion cues. In CVPR, pages 3369-3376, 2011." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 201, + 545, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 201, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 307, + 201, + 545, + 233 + ], + "type": "text", + "content": "[22] Chiuhsiang Joe Lin, Chi-Chan Chang, and Yung-Hui Lee. Evaluating camouflage design using eye movement data. Applied ergonomics, pages 714-723, 2014." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 233, + 545, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 233, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 307, + 233, + 545, + 265 + ], + "type": "text", + "content": "[23] Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan. Camouflaged object segmentation with distraction mining. In CVPR, pages 8772-8781, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 265, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 297 + ], + "type": "text", + "content": "[24] Peter Ochs and Thomas Brox. Object segmentation in video: a hierarchical variational approach for turning point trajectories into dense regions. In ICCV, 2011." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 297, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 318 + ], + "type": "text", + "content": "[25] Peter Ochs and Thomas Brox. Higher order motion models and spectral clustering. In CVPR, pages 614-621, 2012." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 318, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 318, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 318, + 545, + 350 + ], + "type": "text", + "content": "[26] Peter Ochs, Jitendra Malik, and Thomas Brox. Segmentation of moving objects by long term video analysis. volume 36, pages 1187-1200, 2013." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "type": "text", + "content": "[27] Andrew Owens, Connelly Barnes, Alex Flint, Hanumant Singh, and William Freeman. Camouflaging an object from many viewpoints. In CVPR, pages 2782-2789, 2014." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 381, + 545, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 381, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 307, + 381, + 545, + 423 + ], + "type": "text", + "content": "[28] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In CVPR, pages 2160-2170, 2022." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 423, + 545, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 423, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 307, + 423, + 545, + 456 + ], + "type": "text", + "content": "[29] Xavier Soria Poma, Edgar Riba, and Angel Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In Proc. WACV, pages 1923-1932, 2020." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "text", + "content": "[30] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image. In ICCV, pages 4570-4580, 2019." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 488, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 520 + ], + "type": "text", + "content": "[31] Josef Sivic, Frederik Schaffalitzky, and Andrew Zisserman. Object level grouping for video shots. In Proc. ECCV. Springer-Verlag, 2004." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 520, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 574 + ], + "type": "text", + "content": "[32] Przemysław Skurowski, Hassan Abdulameer, J Blaszczyk, Tomasz Depta, Adam Kornacki, and P Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2018. URL: https://www.polsl.pl/rau6/chameleon-database-animal-camouflage-analysis/." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 574, + 545, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 574, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 307, + 574, + 545, + 638 + ], + "type": "text", + "content": "[33] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 638, + 545, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 638, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 307, + 638, + 545, + 681 + ], + "type": "text", + "content": "[34] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proc. CVPR, 2015." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 681, + 545, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 703 + ], + "type": "text", + "content": "[35] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception ar-" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 733, + 312, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 733, + 312, + 742 + ], + "spans": [ + { + "bbox": [ + 298, + 733, + 312, + 742 + ], + "type": "text", + "content": "841" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 275 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "chitecture for computer vision. In CVPR, pages 2818-2826, 2016." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 286, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 286, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 286, + 115 + ], + "type": "text", + "content": "[36] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 116, + 286, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 116, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 116, + 286, + 148 + ], + "type": "text", + "content": "[37] Jolyon Troscianko, John Skelhorn, and Martin Stevens. Quantifying camouflage: how to predict detectability from appearance. BMC evolutionary biology, 17(1):1-13, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 148, + 286, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 148, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 48, + 148, + 286, + 168 + ], + "type": "text", + "content": "[38] Cédric Villani. Optimal transport: old and new, volume 338. Springer, 2009." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 168, + 286, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 168, + 286, + 199 + ], + "spans": [ + { + "bbox": [ + 48, + 168, + 286, + 199 + ], + "type": "text", + "content": "[39] Junyu Xie, Weidi Xie, and Andrew Zisserman. Segmenting moving objects via an object-centric layered representation. In NeurIPS, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 200, + 286, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 200, + 286, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 200, + 286, + 232 + ], + "type": "text", + "content": "[40] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proc. ICCV, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 232, + 286, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 286, + 275 + ], + "type": "text", + "content": "[41] Bowen Zhang, Shuyang Gu, Bo Zhang, Jianmin Bao, Dong Chen, Fang Wen, Yong Wang, and Baining Guo. Styleswin: Transformer-based gan for high-resolution image generation. In CVPR, pages 11304-11314, 2022." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 733, + 312, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 733, + 312, + 741 + ], + "spans": [ + { + "bbox": [ + 298, + 733, + 312, + 741 + ], + "type": "text", + "content": "842" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/b27d12fb-927b-406e-bce3-95c8be702c01_content_list.json b/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/b27d12fb-927b-406e-bce3-95c8be702c01_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..676e9ba70c3ef7a396cf114a3c4f4365ff07f000 --- /dev/null +++ b/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/b27d12fb-927b-406e-bce3-95c8be702c01_content_list.json @@ -0,0 +1,1866 @@ +[ + { + "type": "text", + "text": "The Perils of Learning From Unlabeled Data: Backdoor Attacks on Semi-supervised Learning", + "text_level": 1, + "bbox": [ + 243, + 130, + 727, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Virat Shejwalkar \nUMass Amherst", + "bbox": [ + 202, + 203, + 339, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vshejwalkar@cs.umass.edu", + "bbox": [ + 161, + 242, + 377, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lingjuan Lyu \nSony AI", + "bbox": [ + 459, + 203, + 571, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lingjuan.lv@sony.com", + "bbox": [ + 426, + 242, + 604, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amir Houmansadr \nUMass Amherst", + "bbox": [ + 653, + 203, + 805, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "amir@cs.umass.edu", + "bbox": [ + 653, + 242, + 805, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-supervised learning (SSL) is gaining popularity as it reduces cost of machine learning (ML) by training high performance models using unlabeled data. In this paper, we reveal that the key feature of SSL, i.e., learning from (non-inspected) unlabeled data, exposes SSL to strong poisoning attacks that can significantly damage its security. Poisoning is a long-standing problem in conventional supervised ML, but we argue that, as SSL relies on non-inspected unlabeled data, poisoning poses a more significant threat to SSL.", + "bbox": [ + 75, + 324, + 468, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We demonstrate this by designing a backdoor poisoning attack on SSL that can be conducted by a weak adversary with no knowledge of the target SSL pipeline. This is unlike prior poisoning attacks on supervised ML that assume strong adversaries with impractical capabilities. We show that by poisoning only $0.2\\%$ of the unlabeled training data, our (weak) adversary can successfully cause misclassification on more than $80\\%$ of test inputs (when they contain the backdoor trigger). Our attack remains effective across different benchmark datasets and SSL algorithms, and even circumvents state-of-the-art defenses against backdoor attacks. Our work raises significant concerns about the security of SSL in real-world security critical applications.", + "bbox": [ + 75, + 460, + 468, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 209, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Machine learning (ML) models perform better with increased amounts of training data [13, 12]. However, conventional supervised ML requires labeling large amounts of training data, an expensive [11] and error prone [31, 26] process that makes it prohibitively expensive, especially with today's exploding training data sizes.", + "bbox": [ + 75, + 717, + 468, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-supervised learning (SSL) addresses this major challenge by significantly reducing the need for labeled training data: SSL uses a combination of a small, high-quality labeled data (expensive data) with a large, low-quality unlabeled data (cheap data) to train models. For instance, the FixMatch [40] SSL algorithm combines only", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "40 labeled with $50k$ unlabeled data to achieve a $90\\%$ accuracy on CIFAR10. Training SSL involves two loss functions: a supervised loss (e.g., cross-entropy [30] over true labels) on labeled training data and an unsupervised loss (e.g., cross-entropy over pseudo-labels [22]) on unlabeled training data. Different SSL algorithms primarily differ in terms of how they compute their unsupervised losses.", + "bbox": [ + 496, + 292, + 890, + 398 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SSL has gained popularity in both academia [52, 46, 47] and industry [40, 41, 3, 2], as recent SSL algorithms offer state-of-the-art performances comparable or even superior to supervised techniques—but with no need of large well-inspected labeled data. For instance, due to their effective use of unlabeled data, with less than $10\\%$ of training data labeled, FixMatch [40] outperforms supervised ML.", + "bbox": [ + 496, + 398, + 892, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlabeled data enables poisoning by weak adversaries: Multiple researches have demonstrated the data poisoning threat to supervised ML [18, 28, 34, 36, 50, 44, 38]. However, as the training data in supervised ML undergo an extensive and careful inspection, these attacks assume strong adversaries with the knowledge of model parameters [28], training data [44, 4, 29], its distribution [50], or the ML algorithm. Such strong adversaries are important to evaluate worse-case security of a system, but are irrelevant in practice [38]. On the other hand, the key feature of SSL that makes it attractive to real-world applications is its ability to leverage large amounts of—raw, non-inspected—unlabeled data, e.g., the data scraped off the Internet. We argue that the use of non-inspected data by SSL presents a unique threat to its security, as it allows even the most naive adversaries (with no knowledge of training algorithm, data, etc.) to poison SSL models by simply fabricating malicious unlabeled data. Unfortunately, this ostensible threat is largely unexplored in the SSL literature.", + "bbox": [ + 496, + 506, + 892, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this gap, in this paper, we take the first step towards understanding this threat by studying the possibility of backdoor attacks against SSL in real-world settings. Backdoor attacks aim to install a backdoor function in the target model, such that the backdoored target model will misclassify any test input to the adversary chosen target class when patched with a specific backdoor trigger, but", + "bbox": [ + 496, + 795, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4730", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "will correctly classify test inputs without the trigger.", + "bbox": [ + 76, + 90, + 421, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing backdoor attacks fail on SSL: There exist numerous backdoor attacks in the literature, however, except one attack—DeHiB [48], all of the prior attacks consider supervised ML. Our preliminary evaluations show that all of the existing state-of-the-art (SOTA) attacks, including DeHiB, completely fail against SSL under our realistic threat model (Section 3.1). Hence, to learn from these failures, we first systematically evaluate five SOTA backdoor attacks from three categories against five SOTA SSL algorithms, under our practical, unlabeled data poisoning threat model.", + "bbox": [ + 75, + 106, + 467, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our systematic evaluation leads to the following three major lessons that not only guide our attack design, but can be useful building blocks for (future) backdoor attacks against SSL: (1) Backdoor attacks on SSL should be clean-label style attacks, i.e., poisoning data should be selected from the distribution of target class $y^{t}$ ; (2) Backdoor triggers should be of the same size as the poisoning sample, to circumvent strong augmentations, e.g., cutout [15], that all modern SSL algorithms use; (3) Backdoor triggers should be resistant to noise and with repetitive patterns1 to withstand large amounts of random noises due to strong augmentations, e.g., RandAugment [10], in SSL.", + "bbox": [ + 75, + 258, + 467, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our SSL-tailored backdoor method: The high-level intuition behind our backdoor attack is as follows. All modern SSL algorithms learn via a self-feedback mechanism, called pseudo-labeling, i.e., if current state of target model $f_{\\theta}$ has high confidence prediction $\\tilde{y}$ for an unlabeled sample $\\mathbf{x}$ , then they use $(\\mathbf{x}, \\tilde{y})$ as a labeled sample for further training. We exploit pseudo-labeling and design a clean-label attack that poisons unlabeled data only from the distribution of $y^{t}$ . Our attack patiently waits for $f_{\\theta}$ to correctly label a poisoning sample $(\\mathbf{x} + T)$ as $y^{t}$ , where $T$ is our pre-determined backdoor trigger. As $f_{\\theta}$ trains further on $((\\mathbf{x} + T), y^{t})$ , our attack forces $f_{\\theta}$ to associate features of our simple trigger $T$ , instead of the complex features of $\\mathbf{x}$ , with $y^{t}$ , thereby installing the backdoor in the target model.", + "bbox": [ + 75, + 439, + 467, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note that, we consider the most challenging setting for designing attacks with the least capable and knowledgeable data poisoning adversary. Generally, trigger generation for data poisoning backdoor attacks is formalized as a bi-level optimization problem [29], however such attacks are well-known to be very expensive, and yet ineffective [29, 38]. Instead, our lessons lead us to a simple yet effective static, repetitive grid pattern backdoor trigger (Figure 2).", + "bbox": [ + 75, + 651, + 467, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "**Evaluations:** We demonstrate the strength of our attack via an extensive evaluation against five SOTA SSL and one supervised ML algorithm, using four benchmark image classification tasks commonly used in the SSL literature. We note that our attack significantly outperforms prior attacks from both SSL and supervised ML literature.", + "bbox": [ + 75, + 773, + 467, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We measure success of our attacks using ASR metric: ASR measures the $\\%$ of test inputs from non-target classes that the backdoored model classifies to the target class when patched with backdoor trigger. For the most combinations of algorithms and datasets, our attacks achieve high attack success rates (ASRs) $(>80\\%)$ , while poisoning just $0.2\\%$ of entire training data. For instance, our attacks have more than $90\\%$ ASR against CIFAR100 and more than $80\\%$ ASR against CIFAR10. For SVHN and STL10, our attack has more than $80\\%$ ASR with two exceptions each. While, under our practical threat model, DeHib attack achieves $0\\%$ ASR even with $20\\times$ more poisoning data. Through a systematic experiment design in Section 5.1.4, we show that our intuition aligns with the dynamics of our attacks and justify their strength. Our attack is highly stealthy, as (1) according to $L_{\\infty}$ -norm metric commonly used [50] for stealth measurement, it minimally perturbs the poisoning data and (2) it produces backdoored models which have high accuracy (close to non-backdoored models) on non-backdoored test inputs. We perform comprehensive ablation study (Section 5.2) to demonstrate the high efficacy of our attacks as we vary (1) size of labeled data, (2) backdoor target class, and (3) size of poisoning data.", + "bbox": [ + 500, + 90, + 890, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we show that our attack remains highly effective even when SSL is paired individually with five SOTA defenses against backdoor attacks that are agnostic to learning algorithms. To defend against such unlabeled data poisoning, we argue for SSL to depart from its philosophy of not inspecting unlabeled training data, and instead, pre-process/inspect the unlabeled data and/or design SSL algorithms that are robust-by-design to such poisoning.", + "bbox": [ + 496, + 439, + 890, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries and Related Work", + "text_level": 1, + "bbox": [ + 500, + 574, + 794, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Semi-supervised Learning (SSL)", + "text_level": 1, + "bbox": [ + 500, + 599, + 784, + 614 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised ML requires completely labeled data, $D^{l}$ , which can be prohibitively expensive due to expensive manual labelling involved. SSL reduces this cost by using very few labeled $D^{l}$ and plenty of unlabeled data $D^{u}$ . SSL uses a convex combination of a supervised loss $\\mathcal{L}_{l}$ on $D^{l}$ and an unsupervised loss $\\mathcal{L}_{u}$ on $D^{u}$ . Modern state-of-the-art SSL algorithms rely on two key building blocks: pseudolabeling [22] and consistency regularization [14, 35, 21].", + "bbox": [ + 496, + 622, + 890, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pseudo-labeling uses the current model, $f_{\\theta}$ , to obtain artificial pseudo-labels for $D^{u}$ and only retains the data on which $f_{\\theta}$ has high confidence. Assume $q_{b} = f_{\\theta}(y|u_{b})$ are the predictions of $f_{\\theta}$ on the batch $u_{b}$ of unlabeled data. Then pseudo-labeling loss can be formalized as: $\\frac{1}{|u_b|}\\sum_{b = 1}^{|u_b|}\\mathbb{1}(\\max (q_b)\\geq \\tau)H(\\hat{q}_b,q_b); \\hat{q}_b = \\operatorname {argmax}(q_b), H(. )$ is cross-entropy and $\\tau$ is confidence threshold.", + "bbox": [ + 496, + 744, + 890, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consistency regularization trains $f_{\\theta}$ to output similar predictions for perturbed versions of the same input. It uses stochastic augmentations $a(\\mathbf{x}_u)$ to perturb an unlabeled", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Repetitive pixel patterns are the patterns on which if we zoom in on any part, we get similar pattern. For examples, check Figures 1 and 10.", + "bbox": [ + 76, + 875, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4731", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sample $\\mathbf{x}_u$ and forces $f_{\\theta}$ to have similar outputs on multiple $a(\\mathbf{x}_u)$ 's using the following loss: $\\sum_{b=1}^{|u_b|} \\| f_{\\theta}(y | a(u_b)) - f_{\\theta}(y | a(u_b))\\|_2^2$ , where $a(.)$ produces different output every time it is applied to a batch $u_b$ of unlabeled data. Below we describe the five SOTA SSL algorithms we consider in this work.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) MixMatch [3] combines various prior semi-supervised learning techniques. For an unlabeled sample, MixMatch generates $K$ weakly augmented versions of the unlabeled sample, computes outputs of the current model $f_{\\theta}$ for the $K$ versions, averages them, and sharpens the average prediction by raising all its probabilities by a power of 1/temperature and re-normalizing; it uses the sharpened prediction as the label of the unlabeled sample. Finally, it uses mixup regularization [53] on the combination of labeled and unlabeled data and trains the model using cross-entropy loss.", + "(2) Unsupervised data augmentation (UDA) [46] shows significant improvements in semi-supervised performances by just replacing the simple weak augmentations of MixMatch with a strong augmentation called Randaugment [10]. In an iteration, Randaugment randomly selects a few augmentations from a large set of augmentations and applies them to images.", + "(3) ReMixMatch [3] builds on MixMatch by making multiple modifications, including 1) it replaces the simple weak augmentation in MixMatch with Autoagument [9], 2) it uses augmentation anchoring to improve consistency regularization, i.e., it uses the prediction on a weakly augmented version of unlabeled sample as the target prediction for a strongly augmented version of the unlabeled sample, and 3) it uses distribution alignment, i.e., it normalizes the new model predictions on unlabeled data using the running average of model predictions on unlabeled data. This significantly boosts the performance of resulting model.", + "(4) FixMatch [40] simplifies the complex ReMixMatch algorithm by proposing to use a combination of Pseudolabeling and consistency regularization based on augmentation anchoring (discussed above). FixMatch significantly improves semi-supervised algorithms, especially in the low labeled data regimes.", + "(5) FlexMatch [52] proposes curriculum pseudo labelling (CPL) approach to leverage unlabeled data according to model's learning status. The main idea behind CPL is to flexibly adjust the thresholds used for pseudo-labeling for different classes at each training iteration in order to select more information unlabeled data and their pseudo-labels. CPL can be combined with other algorithms, e.g., UDA." + ], + "bbox": [ + 75, + 184, + 468, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Backdoor Attacks", + "text_level": 1, + "bbox": [ + 76, + 816, + 253, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A backdoor adversary aims to implant a backdoor functionality into a target model. That is, given an input $(\\mathbf{x},y^{*})$ with true label $y^{*}$ , the backdoored target model $f_{\\theta}^{b}$ should output an adversary-desired backdoor target label $y^{t}$ for", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the input patched with a pre-specified backdoor trigger $T$ , but it should output the correct label for the benign input, i.e., $f_{\\theta}^{b}(\\mathbf{x} + T) \\mapsto y^{t}$ and $f_{\\theta}^{b}(\\mathbf{x}) \\mapsto y^{*}$ . There are two major types of backdoor attacks: (1) dirty-label attacks [19, 7, 36, 51, 33, 23] that poison both the features $\\mathbf{x}$ and labels $y^{*}$ of benign, labeled data to obtain poisoning data $D^{p}$ . (2) On the other hand, clean-label attacks obtain $D^{p}$ by poisoning only the features $\\mathbf{x}$ of benign data.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Backdoor attacks on SSL, unfortunately, have not received significant attention from the scientific community. [48, 49, 16] study backdoor attacks against SSL. However, all of these attacks assume access to $D^{l}$ , the labeled training data of SSL. This assumption renders their applicability questionable in realistic settings. For clarity of presentation, we discuss the details of these attacks in Appendix A.1, where we demonstrate (Table 2) and justify why these attacks fail to backdoor SSL.", + "bbox": [ + 496, + 213, + 890, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Backdoor Attack Methodology", + "text_level": 1, + "bbox": [ + 498, + 366, + 821, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first discuss threat model of our attack, followed by our intuition behind backdoor attacks on SSL. Next, we note that, effectively backdoorsing semi-supervised learning (SSL) does not need a new backdoor attack, but requires careful adoption of existing backdoor attacks. We discuss three major lessons learned from systematically evaluating SOTA attacks under our threat model (Section 3.1). Finally, we detail our SOTA backdoor attack based on our intuition and the lessons learned.", + "bbox": [ + 496, + 392, + 890, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Threat Model", + "text_level": 1, + "bbox": [ + 500, + 541, + 643, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider a victim model trainer who collects data from multiple, potentially untrusted sources to train a ML model using SSL for a classification task with $C$ classes.", + "bbox": [ + 496, + 566, + 890, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversary's goal: A backdoor adversary aims to install a backdoor function in the victim's target model. We denote the models without (benign) and with backdoor by $f_{\\theta}^{b}$ and $f_{\\theta}$ , respectively. Adversary's backdoor goal is to force $f_{\\theta}^{b}$ to incorrectly classify all the test inputs from non-target classes to the adversary-desired target class $y^{t}$ , when they are patched with a pre-specified backdoor trigger, $T$ . Adversary's stealth goal aims that $f_{\\theta}^{b}$ should correctly classify all the benign test inputs, i.e., any input without $T$ .", + "bbox": [ + 496, + 612, + 890, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversary's knowledge: As discussed in Section 1, we consider the most naive, real-world adversary with minimum knowledge of the SSL pipeline. More specifically, we assume that the adversary has no knowledge of the labeled or unlabeled training data and does not possess any data from true distribution; they just know the classification task, i.e., CIFAR10 or SVHN. Our adversary knows the details of the target SSL algorithm, but does not know model architecture, e.g., ResNet or VGG, i.e., our attacks are model architecture agnostic.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4732", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversary's capabilities: Due to our emphasis on practicality of our threat model, we consider a data poisoning adversary [38]. Specifically, our adversary can poison only the unlabeled data of SSL pipeline, and cannot poison or even access the model, code or the labeled data of SSL.", + "bbox": [ + 75, + 90, + 472, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Intuition behind our backdoor attack", + "text_level": 1, + "bbox": [ + 76, + 175, + 401, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For brevity, we discuss our intuition for FixMatch [40], but it applies to any SSL algorithms that use pseudolabeling and consistency regularization (Section 2.1).", + "bbox": [ + 75, + 199, + 468, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As explained in Section 2.1, FixMatch trains parameters $\\theta$ to learn a function $f_{\\theta}$ from the labeled data $D^{l}$ and assigns a pseudo-label $\\hat{y}$ to an unlabeled sample $\\mathbf{x} \\in D^{u}$ . Then it further trains $\\theta$ using $(\\mathbf{x}, \\hat{y})$ to improve $f_{\\theta}$ . As the training progresses, the confidence of $f_{\\theta}$ on the correct label of $\\mathbf{x}$ increases which leads to better pseudo-labeling of $D^{u}$ and further improvements in the accuracy of $f_{\\theta}$ . In other words, FixMatch learns via a self-feedback mechanism.", + "bbox": [ + 75, + 244, + 470, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recall that, our realistic data poisoning adversary cannot alter either the SSL training pipeline or the well-inspected labeled training data $D^{l}$ . Now, the first part of our intuition is that our attacks should be of clean-label type, i.e., we select unlabeled data $X^{y^t}$ from the target class, $y^t$ , and patch it with backdoor trigger $T$ to obtain $X^p$ , i.e., $X^p = X^{y^t} + T$ ; next section demonstrates the necessity of this condition. The second part of our intuition is that in initial part of training, FixMatch will assign the desired pseudo-labels $y^t$ to $X^p$ due to the original features $X^{y^t}$ of $X^p$ . However, due to the presence of backdoor trigger, $T$ , on all $X^p s$ , the model will be forced to eventually learn a much simpler task of associating $T$ to $y^t$ .", + "bbox": [ + 75, + 366, + 470, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further understand this, consider three benign samples $\\mathbf{x}_{i\\in \\{1,2,3\\}}$ with target class $y^{t}$ as their true label. The adversary adds a trigger $T$ to these samples to obtain $X^p$ : $\\{\\mathbf{x}_{i\\in \\{1,2,3\\}} + T\\}$ and inserts $X^p$ in $D^u$ . Note that, initially during training, FixMatch learns the association $f_{\\theta}:\\mathcal{X}\\mapsto \\mathcal{V}$ between feature and label spaces only through $D^l$ . And as our threat model assumes that $D^l$ is benign (not poisoned), initially FixMatch focuses only on the benign features of $X^p$ , i.e., on $\\mathbf{x}_{i\\in \\{1,2,3\\}}$ and assigns the correct label $y^t$ to all $X^p$ samples. This in turn forces FixMatch to learn from $(\\mathbf{x}_{i\\in \\{1,2,3\\}} + T,y^t)$ . As $T$ is present in all $X^p$ samples, FixMatch incorrectly learns the simpler task of associating the static trigger $T$ with $y^t$ , instead of the difficult task of associating the complex and dynamic benign features of $\\mathbf{x}_{i\\in \\{1,2,3\\}}$ with $y^t$ ; we very our intuition in Section 5.1.4.", + "bbox": [ + 75, + 564, + 470, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Lessons from systematic evaluation of existing backdoor attacks against SSL", + "text_level": 1, + "bbox": [ + 76, + 801, + 468, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Table 1, we categorize existing SOTA attacks on supervised ML and SSL in three types: dirty label, clean label small trigger and clean label adversarial samples. We evaluate representative attacks from each category and pro", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/26b29a50c1ca59ee37af82cd221b8acd0fab762a56ab82a6eb5131b4be9b42a8.jpg", + "image_caption": [ + "Figure 1: Our backdoor trigger and a poisoning sample." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 880, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ef99f0d41cd3df2ea4ff3f12cb2e622154211804d1aaa4b4312fce5133f4c690.jpg", + "image_caption": [ + "Figure 2: Our backdoor trigger has three parameters: pixel intensity $\\alpha$ , pixel gap $g$ , and pixel width $w$ . For presentation clarity, we use high pixel intensity here, but in experiments we use low intensities to ensure attack stealth." + ], + "image_footnote": [], + "bbox": [ + 575, + 224, + 815, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vide justification for their failure against SSL, and the corresponding lesson that will guide future attack designs on SSL. Due to space limit, we present the key lessons here and defer detailed discussions to Appendix A.", + "bbox": [ + 496, + 401, + 890, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lesson-1: Backdoor attacks against semi-supervised learning should be clean-label type, i.e., the poisoning samples should be from the backdoor target class. Without this condition, model will be forced to learn to associate $T$ with different classes (i.e., original classes of poisoning data $X^p$ ), and effectively, model will simply ignore $T$ .", + "bbox": [ + 496, + 462, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lesson-2: The trigger should have same size as the entire sample (images in our case), to ensure that all the augmented instances of a poisoning sample contain majority of the trigger. This is necessary due to augmentations in SSL that occlude part of an image, e.g., cutout and cutmix, which all the modern SOTA SSL algorithms use.", + "bbox": [ + 496, + 553, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lesson-3: The trigger should be noise-resistant and with a repetitive pattern. This is necessary, again, to circumvent the occluding augmentations described above.", + "bbox": [ + 496, + 645, + 890, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We believe that the above lessons give the minimum constraints to design backdoor attacks on SSL. But, they need not be exhaustive and may need modifications, e.g., based on different threat models and SSL algorithms.", + "bbox": [ + 496, + 691, + 890, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Our State-of-the-art backdoor attack", + "text_level": 1, + "bbox": [ + 500, + 762, + 821, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on our intuition and the three lessons detailed above, we develop a clean-label style backdoor attack using a specific static trigger pattern. Figure 1 depicts our static backdoor trigger and a corresponding poisoning image; we present more images for CIFAR, SVHN, and STL10", + "bbox": [ + 496, + 787, + 890, + 863 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2Repetitive pixel patterns are the patterns on which if we zoom in on any part, we get similar pattern. For examples, check Figures 1 and 10.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4733", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a5ed83480c69aa2c0f019e8a2bf11e34fe65ab3488122419f4c7e61f68c51bb8.jpg", + "table_caption": [ + "Table 1: The left column shows types of backdoor attacks based on specific characteristics, middle column lists existing attacks of each type. Right-most column presents lessons we learn from evaluating one/two representative attacks (in bold) of each type." + ], + "table_footnote": [], + "table_body": "
Attack characteristic/typeExisting attacks of given typeLesson from evaluations
Dirty labelDeHiB [48], DL-Badnets [18], DL-Blend [7], Facehack [36]Attack should be a clean-label attack, i.e., poisoning samples should be from backdoor target class.
Clean-label small triggerCL-Badnets [50], CL-Blend [7]Trigger should span the entire sample/image to avoid cropping/covering by strong augmentations.
Clean-label adversarial samplesNarcissus [50], Label consistent [44], non-repeating trigger patterns, HTBA [34], SAA [42], Embedding [54]Trigger should be noise-resistant and its pattern should be repetitive so that even a part of trigger can install a backdoor.
", + "bbox": [ + 107, + 138, + 848, + 268 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9f02b90f6d6e381435a4d27f1c35ae7eaaf81aa579e1b9b89372cf83f46e33b5.jpg", + "table_caption": [ + "Table 2: Impacts of existing backdoor attacks (Section 2.2) on various semi-supervised algorithms for CIFAR10 data. We poison $0.2\\%$ (100 samples) of all the training data. DeHib* is the original attack with the knowledge of labeled training data $D^l$ while DeHib is the attack without the knowledge of $D^l$ ." + ], + "table_footnote": [], + "table_body": "
AlgorithmDeHiB* ASR (%)DeHiB ASR (%)CL-Badnets ASR (%)LC ASR (%)Narcissus ASR (%)
Mixmatch [3]22.01.09.11.12.2
Remixmatch [2]10.90.90.00.00.0
UDA [46]21.21.25.10.00.0
Fixmatch [40]35.80.910.20.11.3
Flexmatch [52]16.91.29.10.11.1
", + "bbox": [ + 222, + 321, + 733, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "datasets in Figures 14, 15, and 16 in Appendix C. Our backdoor trigger pattern has three parameters: intensity $\\alpha$ , gap $g$ , and width $w$ . $\\alpha$ is the intensity of the bright pixels in the trigger and intensity of the rest of the pixel is 0; $g$ is the distance between two adjacent set of bright pixels and $w$ is the width of each set of bright pixels. Note that the size of our trigger is the same as that of the sample and has a fairly repetitive pattern, hence it satisfies both Lessons-2 and 3.", + "bbox": [ + 75, + 419, + 470, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To summarize our attack: we select a set of samples from the target class (to satisfy Lesson-1, poison them by adding the trigger to them, and inject these poisoned samples into the unlabeled data. As we will show in Section 5.1.1 (Table 3), with poisoning just $0.2\\%$ of the entire training data, this simple backdoor method injects backdoors in SSL models with close to $90\\%$ accuracy.", + "bbox": [ + 75, + 541, + 468, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, it is worth mentioning that, there are many possible triggers that follow aforementioned lessons, but choice of our specific trigger is based on various triggers patterns we investigated in our initial explorations (Figure 10). Furthermore, the choice of our simple yet effective backdoor attack method is a result of an extensive experimentation with various attack methods (and not just trigger patterns). In Appendix C.1, we discuss alternate attack methods we explored but found them unsuccessful at backdoorsing SSL.", + "bbox": [ + 75, + 647, + 470, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental setup", + "text_level": 1, + "bbox": [ + 76, + 799, + 266, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our backdoor attacks using four datasets (CIFAR10, SVHN, CIFAR100 and STL10) commonly used to benchmark semi-supervised algorithms. Due to space limits, we defer the details of datasets, model architectures and hyperparameters to Appendix B.", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Performance metrics", + "text_level": 1, + "bbox": [ + 498, + 419, + 697, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Clean accuracy (CA) [18] measures the accuracy of a model on clean test data without any backdoor trigger $T$ . Backdoors models should have high CA to ensure that the backdoor attack does not impact their benign functionality to ensure the attack stealth.", + "bbox": [ + 496, + 441, + 890, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backdoor attack success rate (ASR) [18] measures the accuracy of a model on the backdoored test data from the nontarget classes patched with $T$ . For a successful backdoor attack, backdoored model should have high ASR.", + "bbox": [ + 496, + 518, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Target class accuracy (TA) [50] measures the accuracy of the clean test data from the target backdoor class which does not contain any $T$ . For a backdoored model, TA should be high to ensure the attack stealth.", + "bbox": [ + 496, + 579, + 890, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Empirical Results", + "text_level": 1, + "bbox": [ + 498, + 648, + 674, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Our attacks effectively backdoor SSL", + "text_level": 1, + "bbox": [ + 498, + 666, + 823, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we demonstrate the superiority of our backdoor attacks over various baseline attacks in terms of three metrics from Section 4.1. Note that, we poison at most $0.2\\%$ of the entire unlabeled data, which is significantly lower than what prior attacks use, e.g., $10\\%$ in DeHiB [48, 49]. Backdoor injection at such low poisoning percentages is extremely challenging as we aim to backdoor the entire test data and not just a single sample as in [5].", + "bbox": [ + 496, + 690, + 890, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.1 High attack success rate (ASR)", + "text_level": 1, + "bbox": [ + 498, + 830, + 769, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Table 3, \"p%\" column shows poisoning percentage and ASR columns show the results. Our backdoor attacks outperform all the baseline backdoor attacks by very large", + "bbox": [ + 496, + 854, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4734", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b67be20d300ab0d3a80b7dcd9131783c9be416b037fa72ca1ff37fc66625bfd2.jpg", + "table_caption": [ + "Table 3: Impacts of backdoor attacks on various semi-supervised (SSL) algorithms (Section 2.1) under the unlabeled data poisoning threat model (Section 3.1). For all datasets, our attack (Section 3.4) significantly outperforms the baseline backdoor attack (DeHiB) against SSL and various clean-label attacks against supervised learning (Section 2.2). Best results are in bold." + ], + "table_footnote": [], + "table_body": "
CIFAR10AlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]92.20.093.50.292.115.394.291.10.094.991.11.492.13092.296.894.6
Remixmatch [2]91.30.094.90.291.01.195.091.30.095.990.82.194.83090.684.394.5
UDA [46]89.50.097.40.188.18.296.989.11.098.689.11.197.22089.681.596.7
Fixmatch [40]91.10.097.50.291.910.197.891.20.098.090.91.195.82093.588.197.6
Flexmatch [52]94.30.097.10.293.96.497.094.10.098.594.22.397.02093.887.996.9
SVHNAlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]94.40.095.40.294.55.493.894.50.096.194.43.295.03093.283.795.8
Remixmatch [2]87.60.095.50.288.01.295.487.10.095.988.11.795.93087.651.195.4
UDA [46]95.00.096.30.294.91.196.094.20.096.094.81.196.62094.995.595.8
Fixmatch [40]94.50.096.30.294.93.197.194.20.097.094.83.296.42094.597.193.9
Flexmatch [52]85.40.096.30.288.91.296.986.10.096.786.82.296.42083.950.196.6
STL10AlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]86.70.086.30.286.39.286.787.11.187.086.11.186.14086.486.287.9
Remixmatch [2]91.70.090.60.291.24.190.691.90.991.191.31.191.04091.282.291.4
UDA [46]88.10.077.50.288.15.577.189.00.177.988.51.777.43088.657.180.4
Fixmatch [40]92.10.086.10.292.213.186.692.10.086.992.02.286.23091.892.487.3
Flexmatch [52]88.10.088.80.288.16.588.188.40.988.087.81.787.93087.849.885.8
CIFAR100AlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]71.60.067.20.271.930.167.572.01.568.372.31.168.13071.692.869.0
Remixmatch [2]73.30.059.10.273.318.959.373.21.160.273.20.559.93073.197.158.2
Fixmatch [40]71.30.049.30.270.622.049.871.41.150.171.42.349.81071.191.848.9
", + "bbox": [ + 83, + 135, + 885, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "margins for all the combinations of datasets and algorithms. More specifically, for various settings, ASRs of our attacks are at least $80\\%$ more than ASRs of Narcissus and DeHiB attacks, while they are at least $60\\%$ more than clean label (CL)-Badnets attacks. Due to space limits, we discuss all the baseline attacks in Appendix A. For UDA + CIFAR10, ASR is $81.5\\%$ with poisoning just a $0.1\\%$ of training data.", + "bbox": [ + 75, + 415, + 468, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Narcissus and $\\mathrm{DeHiB}^3$ attacks achieve close to $0\\%$ ASR for most combinations of datasets and SSL algorithms. As discussed in Appendix A.3, this is expected because all SSL algorithms use strong augmentations which easily obfuscate the dynamic backdoor triggers of these attacks. CL-Badnets attack exhibits relatively higher ASR performances, which is due to the static pattern of its triggers. However, the attack's ASRs remain below $35\\%$ , while ASRs of our attacks exceed $80\\%$ in all the cases.", + "bbox": [ + 75, + 521, + 470, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.2 Negligible impact on clean accuracy (CA)", + "text_level": 1, + "bbox": [ + 76, + 679, + 419, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"CA\" columns in Table 3 show the results. First note that, as Table 6 shows, we use significantly more labeled data for MixMatch than for the other semi-supervised algorithms, and therefore, for some datasets, MixMatch achieves higher accuracy than ReMixMatch or FixMatch. Note from Table 3 that our attacks are highly stealthy as they reduce CA by less than $1.5\\%$ . Baseline attacks also reduce CA negligibly, but their ASRs are very low. Interestingly, for some combinations of dataset and algorithms, we observe an in", + "bbox": [ + 75, + 703, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "crease in CA when we mount our attacks, e.g., for CIFAR10 + FixMatch, CA increases from $91.1\\%$ in the benign setting to $93.5\\%$ , i.e., $2.4\\%$ absolute increase. We also observe that such CA increases generally accompany an increase in the target class accuracy (TA). This is because our attacks add a specific trigger to a subset of target class data and give the model an extra signal to better learn the target class. This improves the TA, and hence, also increases CA.", + "bbox": [ + 500, + 415, + 890, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.3 Negligible impact on target class accuracy (TA)", + "text_level": 1, + "bbox": [ + 500, + 555, + 880, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"TA\" columns in Table 3 show the results. Our attacks remain stealthy with respect to TA as well, as they incur negligible $(< 3\\%)$ reduction in TA. The baseline attacks also do not reduce TAs, but their ASRs are very low. For STL10 + FlexMatch, we observe the maximum, $3\\%$ , reduction in TA. This is because the number of samples for a class that FlexMatch uses during training is inversely proportional to the confidence of the model on that class; the addition of backdoor trigger to the target class data increases the models' confidence on the target class and reduces the target class data that FlexMatch uses for training. On the other hand, in some cases, TA increases as discussed above.", + "bbox": [ + 496, + 578, + 890, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.4 Why and how our attacks work against SSL?", + "text_level": 1, + "bbox": [ + 500, + 770, + 867, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Below, we explain why and how our attack backdoors SSL. For brevity, we limit our discussions to FixMatch and ReMixMatch on CIFAR10 data with target class, $y^{t} = 7$ , i.e., \"horse\", but the insights apply to other SSL algorithms.", + "bbox": [ + 496, + 794, + 890, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FixMatch: FixMatch (as discussed in Section 3.2) uses the current model, $f_{\\theta}$ and assigns hard pseudo-labels to unlabeled data, $D^{u}$ , on which $f_{\\theta}$ has high confidence. Hence,", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3Note that, the original DeHib attack makes an unrealistic assumption, i.e., access to the labeled portion, $D^{l}$ , of the training data. Hence, for a fair comparison, instead of exact $D^{l}$ , we assume that the attacker has some labeled data with same distribution as $D^{l}$ .", + "bbox": [ + 75, + 849, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4735", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/851a6c262d43590641532fb4ab09a22e2e72da6a9cb639a2a9d81d4ad2348168.jpg", + "image_caption": [ + "Figure 3: Dynamics of our backdoor attacks: Initially, SSL assigns the backdoor target class $y^{t}$ as pseudo-labels to poisoning data. Then, our attack forces the model to learn simpler task of associating the backdoor trigger to $y^{t}$ ." + ], + "image_footnote": [], + "bbox": [ + 68, + 90, + 472, + 236 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/67c1d28eb40587aaf17762b758f006f706979a4ce8f11fe906ae63b0848ee29d.jpg", + "table_caption": [ + "Table 4: Backdoor attacks' invisibility as $L_{\\infty}$ -norm of their trigger for CIFAR10. Stealthy attacks have small norms." + ], + "table_footnote": [], + "table_body": "
CL-BadnetsNarcissusDeHiBOurs
At train time255/25532/25532/25530/255
At test time255/25532/25532/25530/255
", + "bbox": [ + 107, + 344, + 434, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to understand why and how our attacks work against FixMatch, in Figure 3-(left), we plot averages of hard pseudolabels of FixMatch on backdoored (poisoning) unlabeled data, $X^p$ and our attack's ASRs as SSL progresses. As training progresses, FixMatch assigns the $y^t$ to more and more of $X^p$ . This forces $f_{\\theta}$ to shift its objective from learning the difficult salient features of the target class to learning much simpler backdoor trigger. Hence, backdoor ASR increases as the average pseudo-label shifts to $y^t = 7$ .", + "bbox": [ + 75, + 392, + 468, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ReMixMatch: As detailed in Section 2.1, ReMixMatch averages predictions on a few augmented versions of an $\\mathbf{x} \\in X^p$ and then uses distribution alignment to compute a prediction vector that it uses as a soft label to train $f_{\\theta}$ . Hence, to understand our backdoor attack on ReMixMatch, in Figure 3-(right), as the training progresses, we plot the average of $f_{\\theta}$ 's confidence on $y^t$ for $X^p$ , and backdoor ASRs. Initially ReMixMatch assigns low confidence to $y^t$ that is due to the distribution alignment, which ensures that ReMixMatch does not assign very high confidence to any single class. However, note from Figure 3 that, once $f_{\\theta}$ learns the salient features of $y^t$ from $X^p$ (with $y^t$ as true label), $f_{\\theta}$ assigns very high confidence to $y^t$ . Next, similar to FixMatch, $f_{\\theta}$ is forced to learn to associate trigger with $y^t$ .", + "bbox": [ + 75, + 529, + 468, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Summary: Our backdoor attacks exploit the high performance of modern SSL algorithms: As our intuition hypothesized in Section 3.2, once they achieve high confidences on $y^{t}$ , our attack forces $f_{\\theta}$ to associate the simple trigger pattern of our attack with the target class, thereby installing the backdoor.", + "bbox": [ + 75, + 741, + 468, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1.5 Additional effectiveness metrics", + "text_level": 1, + "bbox": [ + 76, + 845, + 352, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visibility of backdoor trigger: The unlabeled data of SSL pipeline is never inspected, hence we believe that the visi", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bility of our backdoor triggers is not a significant concern. Nevertheless, following [50], we measure the visibility of backdoor attack as the $L_{\\infty}$ -norm of their backdoor trigger $T$ . The lower the $L_{\\infty}$ -norm of a trigger, the more stealthier the backdoor attack. Table 4 shows the $L_{\\infty}$ -norms of $T$ used for CIFAR10. We note that $L_{\\infty}$ -norm, i.e., visibility, of our $T$ is lower than that of all the baseline attacks. For many combinations of dataset and SSL algorithms, we need even lower $L_{\\infty}$ -norm triggers, e.g., to attack CIFAR10 with FixMatch, UDA and FlexMatch, we use $L_{\\infty} = 20 / 255$ , while attack on CIFAR100 with FixMatch uses $L_{\\infty} = 10 / 255$ .", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficacy against strong augmentations: In this section, we show that our attacks not only work against SSL, but generally perform well against strong augmentations (SA). To this end, we evaluate CL-Badnets, Narcissus and our attack against supervised ML (SML) with and without SAs (we use RandAugment [10]) and provide results in Table 5 for CIFAR10 and CIFAR100. We poison $0.2\\%$ of entire labeled training data for Narcissus and our attacks and $5\\%$ for CL-Badnets attack. We note that although CL-Badnets works well against SML without SAs, it completely fails when we use SAs. On the other hand, our attack works well against SML with and without SAs. Interestingly, Narcissus also works against SML with SAs, but completely fails against SSL (Section 5.1.1 and Appendix A.3). This is because, unlike in SSL, in SML, Narcissus already has the target label for its poisoning data. To summarize, our static pattern backdoor attack is a general attack against strong augmentations and can serve as a building block of backdoor attacks on numerous learning paradigms that use strong augmentations, e.g., self-supervised learning [6].", + "bbox": [ + 496, + 258, + 892, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 573, + 648, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(1) Impact of sizes of labeled training data $(D^{l})$ : Figure 4 plots ASR, CA and TA for our backdoor attacks when we vary $|D^{l}|$ . Due to resource constraints, we experiment with a subset of combinations from Table 3 and use trigger intensities as reported in Table 3 for the combinations.", + "bbox": [ + 496, + 597, + 890, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We note that ASRs remain above $70\\%$ in almost all cases, however we observe a dataset dependent pattern: increasing $|D^l|$ , ASRs first reduce and then increase for CIFAR10, but vice-versa for SVHN (Figure 11). We leave analyses of this phenomena to future work. For FixMatch, we observe that ASRs are almost always above $90\\%$ . This is because FixMatch has high TA and uses hard pseudo-labels, and hence, all poisoning data, $X^p$ , is correctly pseudolabeled as the backdoor target class, $y^t$ . Consequently, the model learns to associate the trigger pattern with $y^t$ . For CIFAR10 + ReMixMatch, we see that TAs are comparable to FixMatch but ASRs are lower. This is because ReMixMatch uses multiple regularizations, including mixup [53] that uses a convex combination of two randomly selected samples and their labels from training data to train the model,", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4736", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9b4ce001142d5688b71459d0d1581cfadcc5612362b21b567426c60a12114674.jpg", + "table_caption": [ + "Table 5: Efficacy of various backdoor attacks against supervised ML (SML) with and without strong augmentations (SA)." + ], + "table_footnote": [], + "table_body": "
AlgorithmCIFAR10CIFAR100
CL-BadnetsNarcissusOur attackCL-BadnetsNarcissusOur attack
CAASRTACAASRTACAASRTACAASRTACAASRTACAASRTA
SML94.783.495.794.6100.096.594.599.895.380.275.379.080.198.186.280.296.890.0
SML + SA94.40.096.794.499.596.894.488.994.980.40.076.080.092.184.380.280.289.0
", + "bbox": [ + 78, + 104, + 893, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c556492bb9f47abc1b1526f5673785b2e4942197ffc3f0b3bd3e6cab4771857b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 174, + 281, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b891b8133becfc0db657fb79d096063a757abe5732abb498f5d3b48bb78dd56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 174, + 460, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/686fffba4f4b3010fd3b3b0ecab9c96222b38bd613ea2ecb5b46b531e4e5a20c.jpg", + "image_caption": [ + "Figure 4: Impacts of varying labeled training data size, $|D^l|$ , for CIFAR10 and {FixMatch, ReMixMatch} algorithms. Upper row shows ASRs and lower row shows CAs and TAs." + ], + "image_footnote": [], + "bbox": [ + 84, + 316, + 279, + 463 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/593a8e1d0b7fb14531477160b91c7024b9e811755a6550b896d59cdd4bc7bfa7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 316, + 460, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which reduces the effective trigger intensity and hence reduces the ASR. Due to space limits, we defer SVHN results and their discussion to Appendix C.", + "bbox": [ + 75, + 532, + 467, + 578 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(2) Impact of backdoor target class $(y^{t})$ : Figure 5 plots ASR, CA and TA of our backdoor attacks for different backdoor target classes, $y^{t}$ ; poisoning data $X^{p}$ is $0.2\\%$ of the total training data.", + "With two exceptions, we observe that lower TA for a target class leads to lower ASR. For instance, in CIFAR10 with FixMatch, when $y^{t}$ is 2 and 3, TAs are $72\\%$ and $65\\%$ , respectively. Due to low TAs, FixMatch assigns $y^{t}$ to smaller proportions of $X^{p}$ , which reduces ASRs. Note that, Carlini [5] also observed that targeted attacks are more effective against better performing SSL algorithms. We observe similar phenomena for CIFAR10 with ReMixMatch and $y^{t} \\in \\{3,5\\}$ , and SVHN with FixMatch and $y^{t} \\in \\{3,5\\}$ . However, we observe that for some classes, e.g., CIFAR10 with FixMatch and $y^{t} \\in \\{6,8\\}$ , TAs are high but ASRs are close to $65\\%$ . We suspect that this is because features of these $y^{t}$ s are too simple to learn, and hence, model correctly ignores the backdoor pattern. Finally, we note that for majority of classes, our attack's ASRs remain above $60\\%$ .", + "(3) Varying the size of unlabeled poisoning data $(X^p)$ : Figure 6 plots ASR, CA and TA for our backdoor attacks" + ], + "bbox": [ + 75, + 580, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0be57d5ab5db1fb3d56d79d5879f659c274e66e8a157ed3d695133d2bdb03968.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 167, + 689, + 309 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99e92580986a1066b3d4b3f63ce7ece2227ca82f2e2834a06713f3a7fe84508d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 167, + 888, + 310 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a597b79ecd7fdb30d373f5fa2f2a396ae49561182a4d6c2a6c37894666d7dae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 315, + 687, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ebd75d7ad0c54c4515146a326c64e733d0d490cecc7002287aa23ae8f8bb3d54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 315, + 888, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/210e2c335db9154670e0684ec6028bcabf9404cf64d5f2b2e62644380ff85af0.jpg", + "image_caption": [ + "Figure 5: ASR, CA and TA of our backdoor attacks for different backdoor target classes, $y^{t}$ ." + ], + "image_footnote": [], + "bbox": [ + 488, + 460, + 687, + 612 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dbdb784c24485f6a60a050406d66859353979a355c08445886a04480c03a2d8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 460, + 888, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with varying $|X^p|$ . More specifically, we vary $|X^p| \\in \\{0.1, 0.15, 0.2, 0.3, 0.4, 0.5\\}$ of the entire training data size. Here, we use labeled data sizes as in Table 6. For all three combinations of dataset and SSL algorithms that we study, we observe that having very small or very large $|X^p|$ leads to relatively ineffective backdoor attacks. This is because at low $|X^p|$ , although almost all of the $X^p$ get the target label, $y^t$ , they are not sufficient to install a backdoor in the target model. While, for large $|X^p|$ , not all of the $X^p$ samples get $y^t$ and some of them get arbitrary labels that are not $y^t$ . This forces model to associate a single trigger pattern with multiple labels and effectively model completely ignores the trigger, which reduces backdoor ASR.", + "bbox": [ + 496, + 656, + 890, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Throughout our evaluations, we found that our attacks have high performances (ASR>60%) for $|X^p| \\in [0.2, 0.4]\\%$ of the entire training data size. Furthermore,", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4737", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a147a2085a663be00db382217a38fe313450dc9a57e267783a26c8b1683263d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 68, + 85, + 256, + 223 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1dbf1a152b9e8e668e0cad82fcb7a3ae3b71cd7b90e7b9e0d539d8b2907769a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 85, + 454, + 222 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/2f7147a00c0c52ad864071dc5de1b018db5a2bef255af5f6f081892f6f957aba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 68, + 224, + 256, + 364 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1361255578e11827fba38627728dd9ca9b653be8134564cfb91b8d3a3ec97700.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 263, + 224, + 454, + 363 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/cc3d8c174c3a8da6075d2c35b4ada8e2f949ea21084fc4364f91430969af9eb0.jpg", + "image_caption": [ + "Figure 6: Impact of varying the sizes of poisoning data on ASR, CA and TA of our backdoor attacks." + ], + "image_footnote": [], + "bbox": [ + 68, + 366, + 256, + 513 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/37a68b8b80046fcdfd08140096b63333dc353231f68b7c0fd52d790017f0f47b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 366, + 454, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "within these ranges, our attacks remain stealthy and do not significantly impact CA or TA of the backdoored models.", + "bbox": [ + 76, + 556, + 468, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.3. Defenses against backdoor attacks on SSL", + "text_level": 1, + "bbox": [ + 76, + 599, + 434, + 616 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Due to space limits, here we give highlights of our evaluations of efficacy of five existing backdoor defenses against our attack and defer detailed discussion to Appendix C.2; Table 7 and Figure 12 in Appendix C.2 show the results.", + "bbox": [ + 76, + 625, + 468, + 685 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Fine-tuning (FT) and fine-pruning (FP) both can reduce ASR of our backdoor attack, however it comes at a significant reduction in CA of the resulting models.", + "(2) Neural attention distillation (NAD) [24] performed best among the defenses we evaluate: for CIFAR10, NAD reduces ASR by $22.1\\%$ for FixMatch and by $23\\%$ for ReMixMatch, but it does not perform as well for SVHN. Nonetheless, even with NAD, our attack still raises significant concerns as its ASR against NAD is always $>60\\%$ .", + "(3) Strip [17] works well when backdoor ASR is very high but for moderate ASRs $\\in [80,90]\\%$ it fails to detect backdoor. For instance, Strip successfully identifies over $90\\%$ of the backdoored test inputs, but it completely fails against CIFAR10 + FixMatch/ReMixMatch and SVHN +" + ], + "bbox": [ + 76, + 686, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "MixMatch.", + "text_level": 1, + "bbox": [ + 500, + 90, + 576, + 104 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "(4) Anti-backdoor learning [25] (ABL) completely fails against SSL, because, SSL training extensively uses strong augmentations, and hence, the unsupervised loss on poisoning unlabeled data remains almost the same as that on benign unlabeled data (Figure 13 in Appendix C.2).", + "bbox": [ + 496, + 106, + 890, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "How to defend SSL from unlabeled data poisoning? We find that some of the SOTA post-processing (FT, FP, NAD, Strip) or in-processing (ABL) defenses cannot defend SSL from our attacks. In other words, current SSL practice of using non-inspected unlabeled training data makes it highly vulnerable to poisoning. Hence, we argue for SSL to depart from its philosophy of not at all inspecting its unlabeled training data and pre-process the unlabeled data to thwart poisoning attacks. Such pre-processing can be tailored to our attack, e.g., check for existence of patterns that follow our three lessons, or check for any abnormal frequency artifacts [51]. We leave pursuing this direction to future work.", + "bbox": [ + 496, + 181, + 892, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 378, + 625, + 395 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Semi-supervised learning (SSL) allows training on large unlabeled data without any inspection, thereby significantly reducing the cost of ML training. Unfortunately, as we show, this key feature can facilitate strong data poisoning attacks on SSL: a naive adversary, without any knowledge of training data or model architecture, can poison just $0.2\\%$ of the entire training data to install a strong backdoor functionality in SSL models. Our attack remains effective against various SSL algorithms and benchmark datasets, and even circumvents state-of-the-art defenses against backdoor attacks.", + "bbox": [ + 496, + 405, + 890, + 569 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that, in contrast to numerous prior works [37, 48, 50, 44, 7, 4, 5, 29], in this work we considered a much weaker, hence more realistic, adversary. Due to our weak adversarial assumptions and simple attack methodology, all of the existing and future SSL applications can use our attack to measure and enhance their robustness against backdoor poisoning.", + "bbox": [ + 496, + 571, + 890, + 676 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Backdoor attacks can have severe consequences in practice, e.g., gaining unauthorized access to a system [7] or denying services to minorities [38]. Hence, a major implication of our study is that real-world SSL applications cannot rely on non-inspected unlabeled data and must preprocess/inspect unlabeled training data and/or design SSL algorithms that are robust to unlabeled data poisoning.", + "bbox": [ + 496, + 678, + 890, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 799, + 666, + 815 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in-part by the NSF grants 2131910 and 1953786 and by DARPA under Agreement No. HR00112190125. Approved for public release; distribution is unlimited. This work was also supported in-part by Sony AI.", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4738", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In International conference on machine learning, pages 284-293. PMLR, 2018. 12", + "[2] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching and augmentation anchoring. In International Conference on Learning Representations, 2019. 1, 5, 6", + "[3] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in Neural Information Processing Systems, 32, 2019. 1, 3, 5, 6, 14", + "[4] B. Biggio, B. Nelson, and P. Laskov. Poisoning attacks against support vector machines. In Proceedings of 29th International Conference on Machine Learning, 2012. 1, 9", + "[5] Nicholas Carlini. Poisoning the unlabeled dataset of {Semi-Supervised} learning. In 30th USENIX Security Symposium (USENIX Security 21), pages 1577-1592, 2021. 5, 8, 9, 14, 15", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 7", + "[7] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 3, 5, 9", + "[8] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 14", + "[9] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 3", + "[10] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 2, 3, 7", + "[11] Aron Culotta and Andrew McCallum. Reducing labeling effort for structured prediction tasks. In AAAI, volume 5, pages 746-751, 2005. 1", + "[12] Jia Deng. A large-scale hierarchical image database. Proc. of IEEE Computer Vision and Pattern Recognition, 2009, 2009. 1", + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1", + "[14] Emily Denton, Sam Gross, and Rob Fergus. Semi-supervised learning with context-conditional generative adversarial networks. arXiv preprint arXiv:1611.06430, 2016. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 2, 12", + "[16] Le Feng, Sheng Li, Zhenxing Qian, and Xinpeng Zhang. Unlabeled backdoor poisoning in semi-supervised learning. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 3", + "[17] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 113–125, 2019. 9, 16", + "[18] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 1, 5", + "[19] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 7:47230-47244, 2019. 3, 13", + "[20] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 14", + "[21] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2", + "[22] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1, 2", + "[23] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021. 3", + "[24] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. In International Conference on Learning Representations, 2020. 9, 16", + "[25] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems, 34:14900-14912, 2021. 9, 16", + "[26] Yuncheng Li, Jianchao Yang, Yale Song, Liangliang Cao, Jiebo Luo, and Li-Jia Li. Learning from noisy labels with distillation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1910-1918, 2017. 1", + "[27] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273-294. Springer, 2018. 16", + "[28] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In 25th Annual Network And Distributed System Security Symposium (NDSS 2018). Internet Soc, 2018. 1", + "[29] Luis Muñoz-González, Battista Biggio, Ambra Demontis, Andrea Paudice, Vasin Wongrassamee, Emil C Lupu, and" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4739", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Fabio Roli. Towards poisoning of deep learning algorithms with back-gradient optimization. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pages 27-38. ACM, 2017. 1, 2, 9", + "[30] Kevin P Murphy. Machine learning: a probabilistic perspective. MIT press, 2012. 1", + "[31] Nagarajan Natarajan, Inderjit S Dhillon, Pradeep K Ravikumar, and Ambuj Tewari. Learning with noisy labels. Advances in neural information processing systems, 26, 2013. 1", + "[32] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 14", + "[33] Tuan Anh Nguyen and Anh Tran. Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems, 33:3454-3464, 2020. 3", + "[34] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11957-11965, 2020. 1, 5", + "[35] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 2", + "[36] Esha Sarkar, Hadjer Benkraouda, and Michail Maniatakos. Facehack: Triggering backdoored facial recognition systems using facial characteristics. arXiv preprint arXiv:2006.11623, 2020. 1, 3, 5", + "[37] Virat Shejwalkar and Amir Houmansadr. Manipulating the byzantine: Optimizing model poisoning attacks and defenses for federated learning. In The Network and Distributed System Security Symposium (NDSS), 2021. 9", + "[38] V. Shejwalkar, A. Houmansadr, P. Kairouz, and D. Ramage. Back to the drawing board: A critical evaluation of poisoning attacks on production federated learning. In 2022 2022 IEEE Symposium on Security and Privacy (SP) (SP), pages 1117-1134, Los Alamitos, CA, USA, may 2022. IEEE Computer Society. 1, 2, 4, 9", + "[39] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019. 12", + "[40] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems, 33:596-608, 2020. 1, 3, 4, 5, 6, 14", + "[41] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 1", + "[42] Hossein Souri, Micah Goldblum, Liam Fowl, Rama Chellappa, and Tom Goldstein. Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. arXiv preprint arXiv:2106.08970, 2021. 5" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Luke Taylor and Geoff Nitschke. Improving deep learning with generic data augmentation. In 2018 IEEE Symposium Series on Computational Intelligence (SSCI), pages 1542-1547. IEEE, 2018. 12", + "[44] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 1, 5, 9, 13", + "[45] Yidong Wang, Hao Chen, Yue Fan, Hao Wu, Bowen Zhang, Wenxin Hou, Yuhao Chen, and Jindong Wang. Torchssl: A pytorch-based toolbox for semi-supervised learning. https://github.com/TorchSSL/TorchSSL, 2021. [Online; accessed 03-July-2022]. 14", + "[46] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems, 33:6256-6268, 2020. 1, 3, 5, 6", + "[47] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 1", + "[48] Zhicong Yan, Gaolei Li, Yuan Tlan, Jun Wu, Shenghong Li, Mingzhe Chen, and H Vincent Poor. Dehib: Deep hidden backdoor attack on semi-supervised learning via adversarial perturbation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 10585-10593, 2021. 2, 3, 5, 9, 12", + "[49] Zhicong Yan, Jun Wu, Gaolei Li, Shenghong Li, and Mohsen Guizani. Deep neural backdoor in semi-supervised learning: threats and countermeasures. IEEE Transactions on Information Forensics and Security, 16:4827-4842, 2021. 3, 5", + "[50] Yi Zeng, Minzhou Pan, Hoang Anh Just, Lingjuan Lyu, Meikang Qiu, and Ruoxi Jia. Narcissus: A practical clean-label backdoor attack with limited information. arXiv preprint arXiv:2204.05255, 2022. 1, 2, 5, 7, 9, 12, 13", + "[51] Yi Zeng, Won Park, Z Morley Mao, and Ruoxi Jia. Rethinking the backdoor attacks' triggers: A frequency perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16473-16481, 2021. 3, 9", + "[52] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 5, 6", + "[53] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 3, 7", + "[54] Haoti Zhong, Cong Liao, Anna Cinzia Squicciarini, Sencun Zhu, and David Miller. Backdoor embedding in convolutional neural network models via invisible perturbation. In Proceedings of the Tenth ACM Conference on Data and Application Security and Privacy, pages 97-108, 2020. 5" + ], + "bbox": [ + 501, + 92, + 890, + 839 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4740", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/b27d12fb-927b-406e-bce3-95c8be702c01_model.json b/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/b27d12fb-927b-406e-bce3-95c8be702c01_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8f6108d3f6971aa45764be52f453be92034e74e6 --- /dev/null +++ b/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/b27d12fb-927b-406e-bce3-95c8be702c01_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.244, + 0.131, + 0.728, + 0.177 + ], + "angle": 0, + "content": "The Perils of Learning From Unlabeled Data: Backdoor Attacks on Semi-supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.203, + 0.204, + 0.341, + 0.238 + ], + "angle": 0, + "content": "Virat Shejwalkar \nUMass Amherst" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.243, + 0.378, + 0.256 + ], + "angle": 0, + "content": "vshejwalkar@cs.umass.edu" + }, + { + "type": "text", + "bbox": [ + 0.46, + 0.204, + 0.572, + 0.239 + ], + "angle": 0, + "content": "Lingjuan Lyu \nSony AI" + }, + { + "type": "text", + "bbox": [ + 0.427, + 0.243, + 0.606, + 0.257 + ], + "angle": 0, + "content": "lingjuan.lv@sony.com" + }, + { + "type": "text", + "bbox": [ + 0.655, + 0.204, + 0.807, + 0.238 + ], + "angle": 0, + "content": "Amir Houmansadr \nUMass Amherst" + }, + { + "type": "text", + "bbox": [ + 0.655, + 0.243, + 0.807, + 0.256 + ], + "angle": 0, + "content": "amir@cs.umass.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.47, + 0.461 + ], + "angle": 0, + "content": "Semi-supervised learning (SSL) is gaining popularity as it reduces cost of machine learning (ML) by training high performance models using unlabeled data. In this paper, we reveal that the key feature of SSL, i.e., learning from (non-inspected) unlabeled data, exposes SSL to strong poisoning attacks that can significantly damage its security. Poisoning is a long-standing problem in conventional supervised ML, but we argue that, as SSL relies on non-inspected unlabeled data, poisoning poses a more significant threat to SSL." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.659 + ], + "angle": 0, + "content": "We demonstrate this by designing a backdoor poisoning attack on SSL that can be conducted by a weak adversary with no knowledge of the target SSL pipeline. This is unlike prior poisoning attacks on supervised ML that assume strong adversaries with impractical capabilities. We show that by poisoning only \\(0.2\\%\\) of the unlabeled training data, our (weak) adversary can successfully cause misclassification on more than \\(80\\%\\) of test inputs (when they contain the backdoor trigger). Our attack remains effective across different benchmark datasets and SSL algorithms, and even circumvents state-of-the-art defenses against backdoor attacks. Our work raises significant concerns about the security of SSL in real-world security critical applications." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.692, + 0.21, + 0.708 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Machine learning (ML) models perform better with increased amounts of training data [13, 12]. However, conventional supervised ML requires labeling large amounts of training data, an expensive [11] and error prone [31, 26] process that makes it prohibitively expensive, especially with today's exploding training data sizes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Semi-supervised learning (SSL) addresses this major challenge by significantly reducing the need for labeled training data: SSL uses a combination of a small, high-quality labeled data (expensive data) with a large, low-quality unlabeled data (cheap data) to train models. For instance, the FixMatch [40] SSL algorithm combines only" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.892, + 0.399 + ], + "angle": 0, + "content": "40 labeled with \\(50k\\) unlabeled data to achieve a \\(90\\%\\) accuracy on CIFAR10. Training SSL involves two loss functions: a supervised loss (e.g., cross-entropy [30] over true labels) on labeled training data and an unsupervised loss (e.g., cross-entropy over pseudo-labels [22]) on unlabeled training data. Different SSL algorithms primarily differ in terms of how they compute their unsupervised losses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.893, + 0.506 + ], + "angle": 0, + "content": "SSL has gained popularity in both academia [52, 46, 47] and industry [40, 41, 3, 2], as recent SSL algorithms offer state-of-the-art performances comparable or even superior to supervised techniques—but with no need of large well-inspected labeled data. For instance, due to their effective use of unlabeled data, with less than \\(10\\%\\) of training data labeled, FixMatch [40] outperforms supervised ML." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.794 + ], + "angle": 0, + "content": "Unlabeled data enables poisoning by weak adversaries: Multiple researches have demonstrated the data poisoning threat to supervised ML [18, 28, 34, 36, 50, 44, 38]. However, as the training data in supervised ML undergo an extensive and careful inspection, these attacks assume strong adversaries with the knowledge of model parameters [28], training data [44, 4, 29], its distribution [50], or the ML algorithm. Such strong adversaries are important to evaluate worse-case security of a system, but are irrelevant in practice [38]. On the other hand, the key feature of SSL that makes it attractive to real-world applications is its ability to leverage large amounts of—raw, non-inspected—unlabeled data, e.g., the data scraped off the Internet. We argue that the use of non-inspected data by SSL presents a unique threat to its security, as it allows even the most naive adversaries (with no knowledge of training algorithm, data, etc.) to poison SSL models by simply fabricating malicious unlabeled data. Unfortunately, this ostensible threat is largely unexplored in the SSL literature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To address this gap, in this paper, we take the first step towards understanding this threat by studying the possibility of backdoor attacks against SSL in real-world settings. Backdoor attacks aim to install a backdoor function in the target model, such that the backdoored target model will misclassify any test input to the adversary chosen target class when patched with a specific backdoor trigger, but" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4730" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.423, + 0.106 + ], + "angle": 0, + "content": "will correctly classify test inputs without the trigger." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.258 + ], + "angle": 0, + "content": "Existing backdoor attacks fail on SSL: There exist numerous backdoor attacks in the literature, however, except one attack—DeHiB [48], all of the prior attacks consider supervised ML. Our preliminary evaluations show that all of the existing state-of-the-art (SOTA) attacks, including DeHiB, completely fail against SSL under our realistic threat model (Section 3.1). Hence, to learn from these failures, we first systematically evaluate five SOTA backdoor attacks from three categories against five SOTA SSL algorithms, under our practical, unlabeled data poisoning threat model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.468, + 0.44 + ], + "angle": 0, + "content": "Our systematic evaluation leads to the following three major lessons that not only guide our attack design, but can be useful building blocks for (future) backdoor attacks against SSL: (1) Backdoor attacks on SSL should be clean-label style attacks, i.e., poisoning data should be selected from the distribution of target class \\( y^{t} \\); (2) Backdoor triggers should be of the same size as the poisoning sample, to circumvent strong augmentations, e.g., cutout [15], that all modern SSL algorithms use; (3) Backdoor triggers should be resistant to noise and with repetitive patterns1 to withstand large amounts of random noises due to strong augmentations, e.g., RandAugment [10], in SSL." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.468, + 0.651 + ], + "angle": 0, + "content": "Our SSL-tailored backdoor method: The high-level intuition behind our backdoor attack is as follows. All modern SSL algorithms learn via a self-feedback mechanism, called pseudo-labeling, i.e., if current state of target model \\( f_{\\theta} \\) has high confidence prediction \\( \\tilde{y} \\) for an unlabeled sample \\( \\mathbf{x} \\), then they use \\( (\\mathbf{x}, \\tilde{y}) \\) as a labeled sample for further training. We exploit pseudo-labeling and design a clean-label attack that poisons unlabeled data only from the distribution of \\( y^{t} \\). Our attack patiently waits for \\( f_{\\theta} \\) to correctly label a poisoning sample \\( (\\mathbf{x} + T) \\) as \\( y^{t} \\), where \\( T \\) is our pre-determined backdoor trigger. As \\( f_{\\theta} \\) trains further on \\( ((\\mathbf{x} + T), y^{t}) \\), our attack forces \\( f_{\\theta} \\) to associate features of our simple trigger \\( T \\), instead of the complex features of \\( \\mathbf{x} \\), with \\( y^{t} \\), thereby installing the backdoor in the target model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.468, + 0.773 + ], + "angle": 0, + "content": "Note that, we consider the most challenging setting for designing attacks with the least capable and knowledgeable data poisoning adversary. Generally, trigger generation for data poisoning backdoor attacks is formalized as a bi-level optimization problem [29], however such attacks are well-known to be very expensive, and yet ineffective [29, 38]. Instead, our lessons lead us to a simple yet effective static, repetitive grid pattern backdoor trigger (Figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.774, + 0.468, + 0.864 + ], + "angle": 0, + "content": "**Evaluations:** We demonstrate the strength of our attack via an extensive evaluation against five SOTA SSL and one supervised ML algorithm, using four benchmark image classification tasks commonly used in the SSL literature. We note that our attack significantly outperforms prior attacks from both SSL and supervised ML literature." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.438 + ], + "angle": 0, + "content": "We measure success of our attacks using ASR metric: ASR measures the \\(\\%\\) of test inputs from non-target classes that the backdoored model classifies to the target class when patched with backdoor trigger. For the most combinations of algorithms and datasets, our attacks achieve high attack success rates (ASRs) \\((>80\\%)\\), while poisoning just \\(0.2\\%\\) of entire training data. For instance, our attacks have more than \\(90\\%\\) ASR against CIFAR100 and more than \\(80\\%\\) ASR against CIFAR10. For SVHN and STL10, our attack has more than \\(80\\%\\) ASR with two exceptions each. While, under our practical threat model, DeHib attack achieves \\(0\\%\\) ASR even with \\(20\\times\\) more poisoning data. Through a systematic experiment design in Section 5.1.4, we show that our intuition aligns with the dynamics of our attacks and justify their strength. Our attack is highly stealthy, as (1) according to \\(L_{\\infty}\\)-norm metric commonly used [50] for stealth measurement, it minimally perturbs the poisoning data and (2) it produces backdoored models which have high accuracy (close to non-backdoored models) on non-backdoored test inputs. We perform comprehensive ablation study (Section 5.2) to demonstrate the high efficacy of our attacks as we vary (1) size of labeled data, (2) backdoor target class, and (3) size of poisoning data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Finally, we show that our attack remains highly effective even when SSL is paired individually with five SOTA defenses against backdoor attacks that are agnostic to learning algorithms. To defend against such unlabeled data poisoning, we argue for SSL to depart from its philosophy of not inspecting unlabeled training data, and instead, pre-process/inspect the unlabeled data and/or design SSL algorithms that are robust-by-design to such poisoning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.795, + 0.59 + ], + "angle": 0, + "content": "2. Preliminaries and Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.6, + 0.785, + 0.616 + ], + "angle": 0, + "content": "2.1. Semi-supervised Learning (SSL)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.745 + ], + "angle": 0, + "content": "Supervised ML requires completely labeled data, \\( D^{l} \\), which can be prohibitively expensive due to expensive manual labelling involved. SSL reduces this cost by using very few labeled \\( D^{l} \\) and plenty of unlabeled data \\( D^{u} \\). SSL uses a convex combination of a supervised loss \\( \\mathcal{L}_{l} \\) on \\( D^{l} \\) and an unsupervised loss \\( \\mathcal{L}_{u} \\) on \\( D^{u} \\). Modern state-of-the-art SSL algorithms rely on two key building blocks: pseudolabeling [22] and consistency regularization [14, 35, 21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Pseudo-labeling uses the current model, \\( f_{\\theta} \\), to obtain artificial pseudo-labels for \\( D^{u} \\) and only retains the data on which \\( f_{\\theta} \\) has high confidence. Assume \\( q_{b} = f_{\\theta}(y|u_{b}) \\) are the predictions of \\( f_{\\theta} \\) on the batch \\( u_{b} \\) of unlabeled data. Then pseudo-labeling loss can be formalized as: \\( \\frac{1}{|u_b|}\\sum_{b = 1}^{|u_b|}\\mathbb{1}(\\max (q_b)\\geq \\tau)H(\\hat{q}_b,q_b); \\hat{q}_b = \\operatorname {argmax}(q_b), H(. ) \\) is cross-entropy and \\( \\tau \\) is confidence threshold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Consistency regularization trains \\( f_{\\theta} \\) to output similar predictions for perturbed versions of the same input. It uses stochastic augmentations \\( a(\\mathbf{x}_u) \\) to perturb an unlabeled" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.468, + 0.901 + ], + "angle": 0, + "content": "1Repetitive pixel patterns are the patterns on which if we zoom in on any part, we get similar pattern. For examples, check Figures 1 and 10." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4731" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.183 + ], + "angle": 0, + "content": "sample \\(\\mathbf{x}_u\\) and forces \\(f_{\\theta}\\) to have similar outputs on multiple \\(a(\\mathbf{x}_u)\\)'s using the following loss: \\(\\sum_{b=1}^{|u_b|} \\| f_{\\theta}(y | a(u_b)) - f_{\\theta}(y | a(u_b))\\|_2^2\\), where \\(a(.)\\) produces different output every time it is applied to a batch \\(u_b\\) of unlabeled data. Below we describe the five SOTA SSL algorithms we consider in this work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.47, + 0.336 + ], + "angle": 0, + "content": "(1) MixMatch [3] combines various prior semi-supervised learning techniques. For an unlabeled sample, MixMatch generates \\( K \\) weakly augmented versions of the unlabeled sample, computes outputs of the current model \\( f_{\\theta} \\) for the \\( K \\) versions, averages them, and sharpens the average prediction by raising all its probabilities by a power of 1/temperature and re-normalizing; it uses the sharpened prediction as the label of the unlabeled sample. Finally, it uses mixup regularization [53] on the combination of labeled and unlabeled data and trains the model using cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.47, + 0.442 + ], + "angle": 0, + "content": "(2) Unsupervised data augmentation (UDA) [46] shows significant improvements in semi-supervised performances by just replacing the simple weak augmentations of MixMatch with a strong augmentation called Randaugment [10]. In an iteration, Randaugment randomly selects a few augmentations from a large set of augmentations and applies them to images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.47, + 0.609 + ], + "angle": 0, + "content": "(3) ReMixMatch [3] builds on MixMatch by making multiple modifications, including 1) it replaces the simple weak augmentation in MixMatch with Autoagument [9], 2) it uses augmentation anchoring to improve consistency regularization, i.e., it uses the prediction on a weakly augmented version of unlabeled sample as the target prediction for a strongly augmented version of the unlabeled sample, and 3) it uses distribution alignment, i.e., it normalizes the new model predictions on unlabeled data using the running average of model predictions on unlabeled data. This significantly boosts the performance of resulting model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.699 + ], + "angle": 0, + "content": "(4) FixMatch [40] simplifies the complex ReMixMatch algorithm by proposing to use a combination of Pseudolabeling and consistency regularization based on augmentation anchoring (discussed above). FixMatch significantly improves semi-supervised algorithms, especially in the low labeled data regimes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.807 + ], + "angle": 0, + "content": "(5) FlexMatch [52] proposes curriculum pseudo labelling (CPL) approach to leverage unlabeled data according to model's learning status. The main idea behind CPL is to flexibly adjust the thresholds used for pseudo-labeling for different classes at each training iteration in order to select more information unlabeled data and their pseudo-labels. CPL can be combined with other algorithms, e.g., UDA." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.185, + 0.47, + 0.807 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.817, + 0.254, + 0.832 + ], + "angle": 0, + "content": "2.2. Backdoor Attacks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "A backdoor adversary aims to implant a backdoor functionality into a target model. That is, given an input \\((\\mathbf{x},y^{*})\\) with true label \\(y^{*}\\), the backdoored target model \\(f_{\\theta}^{b}\\) should output an adversary-desired backdoor target label \\(y^{t}\\) for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "the input patched with a pre-specified backdoor trigger \\( T \\), but it should output the correct label for the benign input, i.e., \\( f_{\\theta}^{b}(\\mathbf{x} + T) \\mapsto y^{t} \\) and \\( f_{\\theta}^{b}(\\mathbf{x}) \\mapsto y^{*} \\). There are two major types of backdoor attacks: (1) dirty-label attacks [19, 7, 36, 51, 33, 23] that poison both the features \\( \\mathbf{x} \\) and labels \\( y^{*} \\) of benign, labeled data to obtain poisoning data \\( D^{p} \\). (2) On the other hand, clean-label attacks obtain \\( D^{p} \\) by poisoning only the features \\( \\mathbf{x} \\) of benign data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.214, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Backdoor attacks on SSL, unfortunately, have not received significant attention from the scientific community. [48, 49, 16] study backdoor attacks against SSL. However, all of these attacks assume access to \\( D^{l} \\), the labeled training data of SSL. This assumption renders their applicability questionable in realistic settings. For clarity of presentation, we discuss the details of these attacks in Appendix A.1, where we demonstrate (Table 2) and justify why these attacks fail to backdoor SSL." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.367, + 0.822, + 0.385 + ], + "angle": 0, + "content": "3. Our Backdoor Attack Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.528 + ], + "angle": 0, + "content": "We first discuss threat model of our attack, followed by our intuition behind backdoor attacks on SSL. Next, we note that, effectively backdoorsing semi-supervised learning (SSL) does not need a new backdoor attack, but requires careful adoption of existing backdoor attacks. We discuss three major lessons learned from systematically evaluating SOTA attacks under our threat model (Section 3.1). Finally, we detail our SOTA backdoor attack based on our intuition and the lessons learned." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.542, + 0.644, + 0.557 + ], + "angle": 0, + "content": "3.1. Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.612 + ], + "angle": 0, + "content": "We consider a victim model trainer who collects data from multiple, potentially untrusted sources to train a ML model using SSL for a classification task with \\( C \\) classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Adversary's goal: A backdoor adversary aims to install a backdoor function in the victim's target model. We denote the models without (benign) and with backdoor by \\( f_{\\theta}^{b} \\) and \\( f_{\\theta} \\), respectively. Adversary's backdoor goal is to force \\( f_{\\theta}^{b} \\) to incorrectly classify all the test inputs from non-target classes to the adversary-desired target class \\( y^{t} \\), when they are patched with a pre-specified backdoor trigger, \\( T \\). Adversary's stealth goal aims that \\( f_{\\theta}^{b} \\) should correctly classify all the benign test inputs, i.e., any input without \\( T \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Adversary's knowledge: As discussed in Section 1, we consider the most naive, real-world adversary with minimum knowledge of the SSL pipeline. More specifically, we assume that the adversary has no knowledge of the labeled or unlabeled training data and does not possess any data from true distribution; they just know the classification task, i.e., CIFAR10 or SVHN. Our adversary knows the details of the target SSL algorithm, but does not know model architecture, e.g., ResNet or VGG, i.e., our attacks are model architecture agnostic." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4732" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.168 + ], + "angle": 0, + "content": "Adversary's capabilities: Due to our emphasis on practicality of our threat model, we consider a data poisoning adversary [38]. Specifically, our adversary can poison only the unlabeled data of SSL pipeline, and cannot poison or even access the model, code or the labeled data of SSL." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.176, + 0.403, + 0.19 + ], + "angle": 0, + "content": "3.2. Intuition behind our backdoor attack" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.47, + 0.246 + ], + "angle": 0, + "content": "For brevity, we discuss our intuition for FixMatch [40], but it applies to any SSL algorithms that use pseudolabeling and consistency regularization (Section 2.1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.471, + 0.366 + ], + "angle": 0, + "content": "As explained in Section 2.1, FixMatch trains parameters \\(\\theta\\) to learn a function \\(f_{\\theta}\\) from the labeled data \\(D^{l}\\) and assigns a pseudo-label \\(\\hat{y}\\) to an unlabeled sample \\(\\mathbf{x} \\in D^{u}\\). Then it further trains \\(\\theta\\) using \\((\\mathbf{x}, \\hat{y})\\) to improve \\(f_{\\theta}\\). As the training progresses, the confidence of \\(f_{\\theta}\\) on the correct label of \\(\\mathbf{x}\\) increases which leads to better pseudo-labeling of \\(D^{u}\\) and further improvements in the accuracy of \\(f_{\\theta}\\). In other words, FixMatch learns via a self-feedback mechanism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.471, + 0.564 + ], + "angle": 0, + "content": "Recall that, our realistic data poisoning adversary cannot alter either the SSL training pipeline or the well-inspected labeled training data \\( D^{l} \\). Now, the first part of our intuition is that our attacks should be of clean-label type, i.e., we select unlabeled data \\( X^{y^t} \\) from the target class, \\( y^t \\), and patch it with backdoor trigger \\( T \\) to obtain \\( X^p \\), i.e., \\( X^p = X^{y^t} + T \\); next section demonstrates the necessity of this condition. The second part of our intuition is that in initial part of training, FixMatch will assign the desired pseudo-labels \\( y^t \\) to \\( X^p \\) due to the original features \\( X^{y^t} \\) of \\( X^p \\). However, due to the presence of backdoor trigger, \\( T \\), on all \\( X^p s \\), the model will be forced to eventually learn a much simpler task of associating \\( T \\) to \\( y^t \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.471, + 0.793 + ], + "angle": 0, + "content": "To further understand this, consider three benign samples \\(\\mathbf{x}_{i\\in \\{1,2,3\\}}\\) with target class \\(y^{t}\\) as their true label. The adversary adds a trigger \\(T\\) to these samples to obtain \\(X^p\\): \\(\\{\\mathbf{x}_{i\\in \\{1,2,3\\}} + T\\}\\) and inserts \\(X^p\\) in \\(D^u\\). Note that, initially during training, FixMatch learns the association \\(f_{\\theta}:\\mathcal{X}\\mapsto \\mathcal{V}\\) between feature and label spaces only through \\(D^l\\). And as our threat model assumes that \\(D^l\\) is benign (not poisoned), initially FixMatch focuses only on the benign features of \\(X^p\\), i.e., on \\(\\mathbf{x}_{i\\in \\{1,2,3\\}}\\) and assigns the correct label \\(y^t\\) to all \\(X^p\\) samples. This in turn forces FixMatch to learn from \\((\\mathbf{x}_{i\\in \\{1,2,3\\}} + T,y^t)\\). As \\(T\\) is present in all \\(X^p\\) samples, FixMatch incorrectly learns the simpler task of associating the static trigger \\(T\\) with \\(y^t\\), instead of the difficult task of associating the complex and dynamic benign features of \\(\\mathbf{x}_{i\\in \\{1,2,3\\}}\\) with \\(y^t\\); we very our intuition in Section 5.1.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.47, + 0.833 + ], + "angle": 0, + "content": "3.3. Lessons from systematic evaluation of existing backdoor attacks against SSL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In Table 1, we categorize existing SOTA attacks on supervised ML and SSL in three types: dirty label, clean label small trigger and clean label adversarial samples. We evaluate representative attacks from each category and pro" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.092, + 0.882, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.208, + 0.866, + 0.223 + ], + "angle": 0, + "content": "Figure 1: Our backdoor trigger and a poisoning sample." + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.226, + 0.816, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.33, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Figure 2: Our backdoor trigger has three parameters: pixel intensity \\(\\alpha\\), pixel gap \\(g\\), and pixel width \\(w\\). For presentation clarity, we use high pixel intensity here, but in experiments we use low intensities to ensure attack stealth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.463 + ], + "angle": 0, + "content": "vide justification for their failure against SSL, and the corresponding lesson that will guide future attack designs on SSL. Due to space limit, we present the key lessons here and defer detailed discussions to Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Lesson-1: Backdoor attacks against semi-supervised learning should be clean-label type, i.e., the poisoning samples should be from the backdoor target class. Without this condition, model will be forced to learn to associate \\( T \\) with different classes (i.e., original classes of poisoning data \\( X^p \\)), and effectively, model will simply ignore \\( T \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.645 + ], + "angle": 0, + "content": "Lesson-2: The trigger should have same size as the entire sample (images in our case), to ensure that all the augmented instances of a poisoning sample contain majority of the trigger. This is necessary due to augmentations in SSL that occlude part of an image, e.g., cutout and cutmix, which all the modern SOTA SSL algorithms use." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.646, + 0.892, + 0.691 + ], + "angle": 0, + "content": "Lesson-3: The trigger should be noise-resistant and with a repetitive pattern. This is necessary, again, to circumvent the occluding augmentations described above." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.892, + 0.753 + ], + "angle": 0, + "content": "We believe that the above lessons give the minimum constraints to design backdoor attacks on SSL. But, they need not be exhaustive and may need modifications, e.g., based on different threat models and SSL algorithms." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.822, + 0.779 + ], + "angle": 0, + "content": "3.4. Our State-of-the-art backdoor attack" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.788, + 0.892, + 0.864 + ], + "angle": 0, + "content": "Based on our intuition and the three lessons detailed above, we develop a clean-label style backdoor attack using a specific static trigger pattern. Figure 1 depicts our static backdoor trigger and a corresponding poisoning image; we present more images for CIFAR, SVHN, and STL10" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2Repetitive pixel patterns are the patterns on which if we zoom in on any part, we get similar pattern. For examples, check Figures 1 and 10." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4733" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.135 + ], + "angle": 0, + "content": "Table 1: The left column shows types of backdoor attacks based on specific characteristics, middle column lists existing attacks of each type. Right-most column presents lessons we learn from evaluating one/two representative attacks (in bold) of each type." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.14, + 0.849, + 0.27 + ], + "angle": 0, + "content": "
Attack characteristic/typeExisting attacks of given typeLesson from evaluations
Dirty labelDeHiB [48], DL-Badnets [18], DL-Blend [7], Facehack [36]Attack should be a clean-label attack, i.e., poisoning samples should be from backdoor target class.
Clean-label small triggerCL-Badnets [50], CL-Blend [7]Trigger should span the entire sample/image to avoid cropping/covering by strong augmentations.
Clean-label adversarial samplesNarcissus [50], Label consistent [44], non-repeating trigger patterns, HTBA [34], SAA [42], Embedding [54]Trigger should be noise-resistant and its pattern should be repetitive so that even a part of trigger can install a backdoor.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.272, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Table 2: Impacts of existing backdoor attacks (Section 2.2) on various semi-supervised algorithms for CIFAR10 data. We poison \\(0.2\\%\\) (100 samples) of all the training data. DeHib* is the original attack with the knowledge of labeled training data \\(D^l\\) while DeHib is the attack without the knowledge of \\(D^l\\)." + }, + { + "type": "table", + "bbox": [ + 0.223, + 0.323, + 0.735, + 0.412 + ], + "angle": 0, + "content": "
AlgorithmDeHiB* ASR (%)DeHiB ASR (%)CL-Badnets ASR (%)LC ASR (%)Narcissus ASR (%)
Mixmatch [3]22.01.09.11.12.2
Remixmatch [2]10.90.90.00.00.0
UDA [46]21.21.25.10.00.0
Fixmatch [40]35.80.910.20.11.3
Flexmatch [52]16.91.29.10.11.1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.471, + 0.541 + ], + "angle": 0, + "content": "datasets in Figures 14, 15, and 16 in Appendix C. Our backdoor trigger pattern has three parameters: intensity \\(\\alpha\\), gap \\(g\\), and width \\(w\\). \\(\\alpha\\) is the intensity of the bright pixels in the trigger and intensity of the rest of the pixel is 0; \\(g\\) is the distance between two adjacent set of bright pixels and \\(w\\) is the width of each set of bright pixels. Note that the size of our trigger is the same as that of the sample and has a fairly repetitive pattern, hence it satisfies both Lessons-2 and 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.648 + ], + "angle": 0, + "content": "To summarize our attack: we select a set of samples from the target class (to satisfy Lesson-1, poison them by adding the trigger to them, and inject these poisoned samples into the unlabeled data. As we will show in Section 5.1.1 (Table 3), with poisoning just \\(0.2\\%\\) of the entire training data, this simple backdoor method injects backdoors in SSL models with close to \\(90\\%\\) accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.471, + 0.785 + ], + "angle": 0, + "content": "Finally, it is worth mentioning that, there are many possible triggers that follow aforementioned lessons, but choice of our specific trigger is based on various triggers patterns we investigated in our initial explorations (Figure 10). Furthermore, the choice of our simple yet effective backdoor attack method is a result of an extensive experimentation with various attack methods (and not just trigger patterns). In Appendix C.1, we discuss alternate attack methods we explored but found them unsuccessful at backdoorsing SSL." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.267, + 0.817 + ], + "angle": 0, + "content": "4. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We evaluate our backdoor attacks using four datasets (CIFAR10, SVHN, CIFAR100 and STL10) commonly used to benchmark semi-supervised algorithms. Due to space limits, we defer the details of datasets, model architectures and hyperparameters to Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.42, + 0.698, + 0.435 + ], + "angle": 0, + "content": "4.1. Performance metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.517 + ], + "angle": 0, + "content": "Clean accuracy (CA) [18] measures the accuracy of a model on clean test data without any backdoor trigger \\( T \\). Backdoors models should have high CA to ensure that the backdoor attack does not impact their benign functionality to ensure the attack stealth." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Backdoor attack success rate (ASR) [18] measures the accuracy of a model on the backdoored test data from the nontarget classes patched with \\( T \\). For a successful backdoor attack, backdoored model should have high ASR." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Target class accuracy (TA) [50] measures the accuracy of the clean test data from the target backdoor class which does not contain any \\( T \\). For a backdoored model, TA should be high to ensure the attack stealth." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.649, + 0.676, + 0.665 + ], + "angle": 0, + "content": "5. Empirical Results" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.667, + 0.824, + 0.683 + ], + "angle": 0, + "content": "5.1. Our attacks effectively backdoor SSL" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.691, + 0.892, + 0.812 + ], + "angle": 0, + "content": "In this section, we demonstrate the superiority of our backdoor attacks over various baseline attacks in terms of three metrics from Section 4.1. Note that, we poison at most \\(0.2\\%\\) of the entire unlabeled data, which is significantly lower than what prior attacks use, e.g., \\(10\\%\\) in DeHiB [48, 49]. Backdoor injection at such low poisoning percentages is extremely challenging as we aim to backdoor the entire test data and not just a single sample as in [5]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.832, + 0.771, + 0.846 + ], + "angle": 0, + "content": "5.1.1 High attack success rate (ASR)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In Table 3, \"p%\" column shows poisoning percentage and ASR columns show the results. Our backdoor attacks outperform all the baseline backdoor attacks by very large" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4734" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.135 + ], + "angle": 0, + "content": "Table 3: Impacts of backdoor attacks on various semi-supervised (SSL) algorithms (Section 2.1) under the unlabeled data poisoning threat model (Section 3.1). For all datasets, our attack (Section 3.4) significantly outperforms the baseline backdoor attack (DeHiB) against SSL and various clean-label attacks against supervised learning (Section 2.2). Best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.136, + 0.887, + 0.409 + ], + "angle": 0, + "content": "
CIFAR10AlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]92.20.093.50.292.115.394.291.10.094.991.11.492.13092.296.894.6
Remixmatch [2]91.30.094.90.291.01.195.091.30.095.990.82.194.83090.684.394.5
UDA [46]89.50.097.40.188.18.296.989.11.098.689.11.197.22089.681.596.7
Fixmatch [40]91.10.097.50.291.910.197.891.20.098.090.91.195.82093.588.197.6
Flexmatch [52]94.30.097.10.293.96.497.094.10.098.594.22.397.02093.887.996.9
SVHNAlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]94.40.095.40.294.55.493.894.50.096.194.43.295.03093.283.795.8
Remixmatch [2]87.60.095.50.288.01.295.487.10.095.988.11.795.93087.651.195.4
UDA [46]95.00.096.30.294.91.196.094.20.096.094.81.196.62094.995.595.8
Fixmatch [40]94.50.096.30.294.93.197.194.20.097.094.83.296.42094.597.193.9
Flexmatch [52]85.40.096.30.288.91.296.986.10.096.786.82.296.42083.950.196.6
STL10AlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]86.70.086.30.286.39.286.787.11.187.086.11.186.14086.486.287.9
Remixmatch [2]91.70.090.60.291.24.190.691.90.991.191.31.191.04091.282.291.4
UDA [46]88.10.077.50.288.15.577.189.00.177.988.51.777.43088.657.180.4
Fixmatch [40]92.10.086.10.292.213.186.692.10.086.992.02.286.23091.892.487.3
Flexmatch [52]88.10.088.80.288.16.588.188.40.988.087.81.787.93087.849.885.8
CIFAR100AlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]71.60.067.20.271.930.167.572.01.568.372.31.168.13071.692.869.0
Remixmatch [2]73.30.059.10.273.318.959.373.21.160.273.20.559.93073.197.158.2
Fixmatch [40]71.30.049.30.270.622.049.871.41.150.171.42.349.81071.191.848.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.522 + ], + "angle": 0, + "content": "margins for all the combinations of datasets and algorithms. More specifically, for various settings, ASRs of our attacks are at least \\(80\\%\\) more than ASRs of Narcissus and DeHiB attacks, while they are at least \\(60\\%\\) more than clean label (CL)-Badnets attacks. Due to space limits, we discuss all the baseline attacks in Appendix A. For UDA + CIFAR10, ASR is \\(81.5\\%\\) with poisoning just a \\(0.1\\%\\) of training data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.471, + 0.658 + ], + "angle": 0, + "content": "Narcissus and \\(\\mathrm{DeHiB}^3\\) attacks achieve close to \\(0\\%\\) ASR for most combinations of datasets and SSL algorithms. As discussed in Appendix A.3, this is expected because all SSL algorithms use strong augmentations which easily obfuscate the dynamic backdoor triggers of these attacks. CL-Badnets attack exhibits relatively higher ASR performances, which is due to the static pattern of its triggers. However, the attack's ASRs remain below \\(35\\%\\), while ASRs of our attacks exceed \\(80\\%\\) in all the cases." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.421, + 0.696 + ], + "angle": 0, + "content": "5.1.2 Negligible impact on clean accuracy (CA)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.84 + ], + "angle": 0, + "content": "\"CA\" columns in Table 3 show the results. First note that, as Table 6 shows, we use significantly more labeled data for MixMatch than for the other semi-supervised algorithms, and therefore, for some datasets, MixMatch achieves higher accuracy than ReMixMatch or FixMatch. Note from Table 3 that our attacks are highly stealthy as they reduce CA by less than \\(1.5\\%\\). Baseline attacks also reduce CA negligibly, but their ASRs are very low. Interestingly, for some combinations of dataset and algorithms, we observe an in" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.416, + 0.892, + 0.537 + ], + "angle": 0, + "content": "crease in CA when we mount our attacks, e.g., for CIFAR10 + FixMatch, CA increases from \\(91.1\\%\\) in the benign setting to \\(93.5\\%\\), i.e., \\(2.4\\%\\) absolute increase. We also observe that such CA increases generally accompany an increase in the target class accuracy (TA). This is because our attacks add a specific trigger to a subset of target class data and give the model an extra signal to better learn the target class. This improves the TA, and hence, also increases CA." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.556, + 0.882, + 0.571 + ], + "angle": 0, + "content": "5.1.3 Negligible impact on target class accuracy (TA)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.761 + ], + "angle": 0, + "content": "\"TA\" columns in Table 3 show the results. Our attacks remain stealthy with respect to TA as well, as they incur negligible \\((< 3\\%)\\) reduction in TA. The baseline attacks also do not reduce TAs, but their ASRs are very low. For STL10 + FlexMatch, we observe the maximum, \\(3\\%\\), reduction in TA. This is because the number of samples for a class that FlexMatch uses during training is inversely proportional to the confidence of the model on that class; the addition of backdoor trigger to the target class data increases the models' confidence on the target class and reduces the target class data that FlexMatch uses for training. On the other hand, in some cases, TA increases as discussed above." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.869, + 0.787 + ], + "angle": 0, + "content": "5.1.4 Why and how our attacks work against SSL?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Below, we explain why and how our attack backdoors SSL. For brevity, we limit our discussions to FixMatch and ReMixMatch on CIFAR10 data with target class, \\( y^{t} = 7 \\), i.e., \"horse\", but the insights apply to other SSL algorithms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "FixMatch: FixMatch (as discussed in Section 3.2) uses the current model, \\( f_{\\theta} \\) and assigns hard pseudo-labels to unlabeled data, \\( D^{u} \\), on which \\( f_{\\theta} \\) has high confidence. Hence," + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.85, + 0.47, + 0.9 + ], + "angle": 0, + "content": "3Note that, the original DeHib attack makes an unrealistic assumption, i.e., access to the labeled portion, \\( D^{l} \\), of the training data. Hence, for a fair comparison, instead of exact \\( D^{l} \\), we assume that the attacker has some labeled data with same distribution as \\( D^{l} \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4735" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.07, + 0.092, + 0.473, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.47, + 0.306 + ], + "angle": 0, + "content": "Figure 3: Dynamics of our backdoor attacks: Initially, SSL assigns the backdoor target class \\( y^{t} \\) as pseudo-labels to poisoning data. Then, our attack forces the model to learn simpler task of associating the backdoor trigger to \\( y^{t} \\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.314, + 0.47, + 0.344 + ], + "angle": 0, + "content": "Table 4: Backdoor attacks' invisibility as \\( L_{\\infty} \\)-norm of their trigger for CIFAR10. Stealthy attacks have small norms." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.345, + 0.436, + 0.38 + ], + "angle": 0, + "content": "
CL-BadnetsNarcissusDeHiBOurs
At train time255/25532/25532/25530/255
At test time255/25532/25532/25530/255
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.47, + 0.529 + ], + "angle": 0, + "content": "to understand why and how our attacks work against FixMatch, in Figure 3-(left), we plot averages of hard pseudolabels of FixMatch on backdoored (poisoning) unlabeled data, \\( X^p \\) and our attack's ASRs as SSL progresses. As training progresses, FixMatch assigns the \\( y^t \\) to more and more of \\( X^p \\). This forces \\( f_{\\theta} \\) to shift its objective from learning the difficult salient features of the target class to learning much simpler backdoor trigger. Hence, backdoor ASR increases as the average pseudo-label shifts to \\( y^t = 7 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.741 + ], + "angle": 0, + "content": "ReMixMatch: As detailed in Section 2.1, ReMixMatch averages predictions on a few augmented versions of an \\(\\mathbf{x} \\in X^p\\) and then uses distribution alignment to compute a prediction vector that it uses as a soft label to train \\(f_{\\theta}\\). Hence, to understand our backdoor attack on ReMixMatch, in Figure 3-(right), as the training progresses, we plot the average of \\(f_{\\theta}\\)'s confidence on \\(y^t\\) for \\(X^p\\), and backdoor ASRs. Initially ReMixMatch assigns low confidence to \\(y^t\\) that is due to the distribution alignment, which ensures that ReMixMatch does not assign very high confidence to any single class. However, note from Figure 3 that, once \\(f_{\\theta}\\) learns the salient features of \\(y^t\\) from \\(X^p\\) (with \\(y^t\\) as true label), \\(f_{\\theta}\\) assigns very high confidence to \\(y^t\\). Next, similar to FixMatch, \\(f_{\\theta}\\) is forced to learn to associate trigger with \\(y^t\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.469, + 0.833 + ], + "angle": 0, + "content": "Summary: Our backdoor attacks exploit the high performance of modern SSL algorithms: As our intuition hypothesized in Section 3.2, once they achieve high confidences on \\( y^{t} \\), our attack forces \\( f_{\\theta} \\) to associate the simple trigger pattern of our attack with the target class, thereby installing the backdoor." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.353, + 0.86 + ], + "angle": 0, + "content": "5.1.5 Additional effectiveness metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Visibility of backdoor trigger: The unlabeled data of SSL pipeline is never inspected, hence we believe that the visi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "bility of our backdoor triggers is not a significant concern. Nevertheless, following [50], we measure the visibility of backdoor attack as the \\( L_{\\infty} \\)-norm of their backdoor trigger \\( T \\). The lower the \\( L_{\\infty} \\)-norm of a trigger, the more stealthier the backdoor attack. Table 4 shows the \\( L_{\\infty} \\)-norms of \\( T \\) used for CIFAR10. We note that \\( L_{\\infty} \\)-norm, i.e., visibility, of our \\( T \\) is lower than that of all the baseline attacks. For many combinations of dataset and SSL algorithms, we need even lower \\( L_{\\infty} \\)-norm triggers, e.g., to attack CIFAR10 with FixMatch, UDA and FlexMatch, we use \\( L_{\\infty} = 20 / 255 \\), while attack on CIFAR100 with FixMatch uses \\( L_{\\infty} = 10 / 255 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.893, + 0.562 + ], + "angle": 0, + "content": "Efficacy against strong augmentations: In this section, we show that our attacks not only work against SSL, but generally perform well against strong augmentations (SA). To this end, we evaluate CL-Badnets, Narcissus and our attack against supervised ML (SML) with and without SAs (we use RandAugment [10]) and provide results in Table 5 for CIFAR10 and CIFAR100. We poison \\(0.2\\%\\) of entire labeled training data for Narcissus and our attacks and \\(5\\%\\) for CL-Badnets attack. We note that although CL-Badnets works well against SML without SAs, it completely fails when we use SAs. On the other hand, our attack works well against SML with and without SAs. Interestingly, Narcissus also works against SML with SAs, but completely fails against SSL (Section 5.1.1 and Appendix A.3). This is because, unlike in SSL, in SML, Narcissus already has the target label for its poisoning data. To summarize, our static pattern backdoor attack is a general attack against strong augmentations and can serve as a building block of backdoor attacks on numerous learning paradigms that use strong augmentations, e.g., self-supervised learning [6]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.65, + 0.59 + ], + "angle": 0, + "content": "5.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.673 + ], + "angle": 0, + "content": "(1) Impact of sizes of labeled training data \\((D^{l})\\): Figure 4 plots ASR, CA and TA for our backdoor attacks when we vary \\(|D^{l}|\\). Due to resource constraints, we experiment with a subset of combinations from Table 3 and use trigger intensities as reported in Table 3 for the combinations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "We note that ASRs remain above \\(70\\%\\) in almost all cases, however we observe a dataset dependent pattern: increasing \\(|D^l|\\), ASRs first reduce and then increase for CIFAR10, but vice-versa for SVHN (Figure 11). We leave analyses of this phenomena to future work. For FixMatch, we observe that ASRs are almost always above \\(90\\%\\). This is because FixMatch has high TA and uses hard pseudo-labels, and hence, all poisoning data, \\(X^p\\), is correctly pseudolabeled as the backdoor target class, \\(y^t\\). Consequently, the model learns to associate the trigger pattern with \\(y^t\\). For CIFAR10 + ReMixMatch, we see that TAs are comparable to FixMatch but ASRs are lower. This is because ReMixMatch uses multiple regularizations, including mixup [53] that uses a convex combination of two randomly selected samples and their labels from training data to train the model," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4736" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.085, + 0.09, + 0.882, + 0.104 + ], + "angle": 0, + "content": "Table 5: Efficacy of various backdoor attacks against supervised ML (SML) with and without strong augmentations (SA)." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.105, + 0.894, + 0.16 + ], + "angle": 0, + "content": "
AlgorithmCIFAR10CIFAR100
CL-BadnetsNarcissusOur attackCL-BadnetsNarcissusOur attack
CAASRTACAASRTACAASRTACAASRTACAASRTACAASRTA
SML94.783.495.794.6100.096.594.599.895.380.275.379.080.198.186.280.296.890.0
SML + SA94.40.096.794.499.596.894.488.994.980.40.076.080.092.184.380.280.289.0
" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.175, + 0.282, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.175, + 0.462, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.317, + 0.28, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.317, + 0.462, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.471, + 0.468, + 0.517 + ], + "angle": 0, + "content": "Figure 4: Impacts of varying labeled training data size, \\( |D^l| \\), for CIFAR10 and {FixMatch, ReMixMatch} algorithms. Upper row shows ASRs and lower row shows CAs and TAs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.468, + 0.579 + ], + "angle": 0, + "content": "which reduces the effective trigger intensity and hence reduces the ASR. Due to space limits, we defer SVHN results and their discussion to Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.468, + 0.642 + ], + "angle": 0, + "content": "(2) Impact of backdoor target class \\((y^{t})\\): Figure 5 plots ASR, CA and TA of our backdoor attacks for different backdoor target classes, \\(y^{t}\\); poisoning data \\(X^{p}\\) is \\(0.2\\%\\) of the total training data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.468, + 0.868 + ], + "angle": 0, + "content": "With two exceptions, we observe that lower TA for a target class leads to lower ASR. For instance, in CIFAR10 with FixMatch, when \\( y^{t} \\) is 2 and 3, TAs are \\( 72\\% \\) and \\( 65\\% \\), respectively. Due to low TAs, FixMatch assigns \\( y^{t} \\) to smaller proportions of \\( X^{p} \\), which reduces ASRs. Note that, Carlini [5] also observed that targeted attacks are more effective against better performing SSL algorithms. We observe similar phenomena for CIFAR10 with ReMixMatch and \\( y^{t} \\in \\{3,5\\} \\), and SVHN with FixMatch and \\( y^{t} \\in \\{3,5\\} \\). However, we observe that for some classes, e.g., CIFAR10 with FixMatch and \\( y^{t} \\in \\{6,8\\} \\), TAs are high but ASRs are close to \\( 65\\% \\). We suspect that this is because features of these \\( y^{t} \\)s are too simple to learn, and hence, model correctly ignores the backdoor pattern. Finally, we note that for majority of classes, our attack's ASRs remain above \\( 60\\% \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "(3) Varying the size of unlabeled poisoning data \\((X^p)\\): Figure 6 plots ASR, CA and TA for our backdoor attacks" + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.581, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.169, + 0.691, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.169, + 0.89, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.316, + 0.689, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.316, + 0.889, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.461, + 0.688, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.461, + 0.889, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.621, + 0.892, + 0.652 + ], + "angle": 0, + "content": "Figure 5: ASR, CA and TA of our backdoor attacks for different backdoor target classes, \\( y^{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.853 + ], + "angle": 0, + "content": "with varying \\( |X^p| \\). More specifically, we vary \\( |X^p| \\in \\{0.1, 0.15, 0.2, 0.3, 0.4, 0.5\\} \\) of the entire training data size. Here, we use labeled data sizes as in Table 6. For all three combinations of dataset and SSL algorithms that we study, we observe that having very small or very large \\( |X^p| \\) leads to relatively ineffective backdoor attacks. This is because at low \\( |X^p| \\), although almost all of the \\( X^p \\) get the target label, \\( y^t \\), they are not sufficient to install a backdoor in the target model. While, for large \\( |X^p| \\), not all of the \\( X^p \\) samples get \\( y^t \\) and some of them get arbitrary labels that are not \\( y^t \\). This forces model to associate a single trigger pattern with multiple labels and effectively model completely ignores the trigger, which reduces backdoor ASR." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Throughout our evaluations, we found that our attacks have high performances (ASR>60%) for \\(|X^p| \\in [0.2, 0.4]\\%\\) of the entire training data size. Furthermore," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4737" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.07, + 0.087, + 0.258, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.087, + 0.455, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.069, + 0.226, + 0.258, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.226, + 0.455, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.07, + 0.367, + 0.258, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.367, + 0.455, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.52, + 0.47, + 0.55 + ], + "angle": 0, + "content": "Figure 6: Impact of varying the sizes of poisoning data on ASR, CA and TA of our backdoor attacks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.558, + 0.47, + 0.589 + ], + "angle": 0, + "content": "within these ranges, our attacks remain stealthy and do not significantly impact CA or TA of the backdoored models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.436, + 0.617 + ], + "angle": 0, + "content": "5.3. Defenses against backdoor attacks on SSL" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.626, + 0.47, + 0.686 + ], + "angle": 0, + "content": "Due to space limits, here we give highlights of our evaluations of efficacy of five existing backdoor defenses against our attack and defer detailed discussion to Appendix C.2; Table 7 and Figure 12 in Appendix C.2 show the results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.687, + 0.469, + 0.732 + ], + "angle": 0, + "content": "(1) Fine-tuning (FT) and fine-pruning (FP) both can reduce ASR of our backdoor attack, however it comes at a significant reduction in CA of the resulting models." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.733, + 0.469, + 0.824 + ], + "angle": 0, + "content": "(2) Neural attention distillation (NAD) [24] performed best among the defenses we evaluate: for CIFAR10, NAD reduces ASR by \\(22.1\\%\\) for FixMatch and by \\(23\\%\\) for ReMixMatch, but it does not perform as well for SVHN. Nonetheless, even with NAD, our attack still raises significant concerns as its ASR against NAD is always \\(>60\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "(3) Strip [17] works well when backdoor ASR is very high but for moderate ASRs \\(\\in [80,90]\\%\\) it fails to detect backdoor. For instance, Strip successfully identifies over \\(90\\%\\) of the backdoored test inputs, but it completely fails against CIFAR10 + FixMatch/ReMixMatch and SVHN +" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.687, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.578, + 0.105 + ], + "angle": 0, + "content": "MixMatch." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.182 + ], + "angle": 0, + "content": "(4) Anti-backdoor learning [25] (ABL) completely fails against SSL, because, SSL training extensively uses strong augmentations, and hence, the unsupervised loss on poisoning unlabeled data remains almost the same as that on benign unlabeled data (Figure 13 in Appendix C.2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.893, + 0.365 + ], + "angle": 0, + "content": "How to defend SSL from unlabeled data poisoning? We find that some of the SOTA post-processing (FT, FP, NAD, Strip) or in-processing (ABL) defenses cannot defend SSL from our attacks. In other words, current SSL practice of using non-inspected unlabeled training data makes it highly vulnerable to poisoning. Hence, we argue for SSL to depart from its philosophy of not at all inspecting its unlabeled training data and pre-process the unlabeled data to thwart poisoning attacks. Such pre-processing can be tailored to our attack, e.g., check for existence of patterns that follow our three lessons, or check for any abnormal frequency artifacts [51]. We leave pursuing this direction to future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.38, + 0.627, + 0.396 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.57 + ], + "angle": 0, + "content": "Semi-supervised learning (SSL) allows training on large unlabeled data without any inspection, thereby significantly reducing the cost of ML training. Unfortunately, as we show, this key feature can facilitate strong data poisoning attacks on SSL: a naive adversary, without any knowledge of training data or model architecture, can poison just \\(0.2\\%\\) of the entire training data to install a strong backdoor functionality in SSL models. Our attack remains effective against various SSL algorithms and benchmark datasets, and even circumvents state-of-the-art defenses against backdoor attacks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Note that, in contrast to numerous prior works [37, 48, 50, 44, 7, 4, 5, 29], in this work we considered a much weaker, hence more realistic, adversary. Due to our weak adversarial assumptions and simple attack methodology, all of the existing and future SSL applications can use our attack to measure and enhance their robustness against backdoor poisoning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.892, + 0.785 + ], + "angle": 0, + "content": "Backdoor attacks can have severe consequences in practice, e.g., gaining unauthorized access to a system [7] or denying services to minorities [38]. Hence, a major implication of our study is that real-world SSL applications cannot rely on non-inspected unlabeled data and must preprocess/inspect unlabeled training data and/or design SSL algorithms that are robust to unlabeled data poisoning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.667, + 0.816 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in-part by the NSF grants 2131910 and 1953786 and by DARPA under Agreement No. HR00112190125. Approved for public release; distribution is unlimited. This work was also supported in-part by Sony AI." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4738" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In International conference on machine learning, pages 284-293. PMLR, 2018. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching and augmentation anchoring. In International Conference on Learning Representations, 2019. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.309 + ], + "angle": 0, + "content": "[3] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in Neural Information Processing Systems, 32, 2019. 1, 3, 5, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[4] B. Biggio, B. Nelson, and P. Laskov. Poisoning attacks against support vector machines. In Proceedings of 29th International Conference on Machine Learning, 2012. 1, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.47, + 0.407 + ], + "angle": 0, + "content": "[5] Nicholas Carlini. Poisoning the unlabeled dataset of {Semi-Supervised} learning. In 30th USENIX Security Symposium (USENIX Security 21), pages 1577-1592, 2021. 5, 8, 9, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.465 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.521 + ], + "angle": 0, + "content": "[7] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 3, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[8] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.47, + 0.646 + ], + "angle": 0, + "content": "[9] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.717 + ], + "angle": 0, + "content": "[10] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[11] Aron Culotta and Andrew McCallum. Reducing labeling effort for structured prediction tasks. In AAAI, volume 5, pages 746-751, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.802 + ], + "angle": 0, + "content": "[12] Jia Deng. A large-scale hierarchical image database. Proc. of IEEE Computer Vision and Pattern Recognition, 2009, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.859 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Emily Denton, Sam Gross, and Rob Fergus. Semi-supervised learning with context-conditional generative adversarial networks. arXiv preprint arXiv:1611.06430, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[15] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 2, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[16] Le Feng, Sheng Li, Zhenxing Qian, and Xinpeng Zhang. Unlabeled backdoor poisoning in semi-supervised learning. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[17] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 113–125, 2019. 9, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[18] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[19] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 7:47230-47244, 2019. 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[20] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[21] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.497 + ], + "angle": 0, + "content": "[22] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[23] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[24] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. In International Conference on Learning Representations, 2020. 9, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.624, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[25] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems, 34:14900-14912, 2021. 9, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[26] Yuncheng Li, Jianchao Yang, Yale Song, Liangliang Cao, Jiebo Luo, and Li-Jia Li. Learning from noisy labels with distillation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1910-1918, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[27] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273-294. Springer, 2018. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[28] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In 25th Annual Network And Distributed System Security Symposium (NDSS 2018). Internet Soc, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Luis Muñoz-González, Battista Biggio, Ambra Demontis, Andrea Paudice, Vasin Wongrassamee, Emil C Lupu, and" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4739" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Fabio Roli. Towards poisoning of deep learning algorithms with back-gradient optimization. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pages 27-38. ACM, 2017. 1, 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[30] Kevin P Murphy. Machine learning: a probabilistic perspective. MIT press, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[31] Nagarajan Natarajan, Inderjit S Dhillon, Pradeep K Ravikumar, and Ambuj Tewari. Learning with noisy labels. Advances in neural information processing systems, 26, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[32] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[33] Tuan Anh Nguyen and Anh Tran. Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems, 33:3454-3464, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[34] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11957-11965, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.407, + 0.469, + 0.461 + ], + "angle": 0, + "content": "[35] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[36] Esha Sarkar, Hadjer Benkraouda, and Michail Maniatakos. Facehack: Triggering backdoored facial recognition systems using facial characteristics. arXiv preprint arXiv:2006.11623, 2020. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.521, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[37] Virat Shejwalkar and Amir Houmansadr. Manipulating the byzantine: Optimizing model poisoning attacks and defenses for federated learning. In The Network and Distributed System Security Symposium (NDSS), 2021. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[38] V. Shejwalkar, A. Houmansadr, P. Kairouz, and D. Ramage. Back to the drawing board: A critical evaluation of poisoning attacks on production federated learning. In 2022 2022 IEEE Symposium on Security and Privacy (SP) (SP), pages 1117-1134, Los Alamitos, CA, USA, may 2022. IEEE Computer Society. 1, 2, 4, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.662, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[39] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[40] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems, 33:596-608, 2020. 1, 3, 4, 5, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[41] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[42] Hossein Souri, Micah Goldblum, Liam Fowl, Rama Chellappa, and Tom Goldstein. Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. arXiv preprint arXiv:2106.08970, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[43] Luke Taylor and Geoff Nitschke. Improving deep learning with generic data augmentation. In 2018 IEEE Symposium Series on Computational Intelligence (SSCI), pages 1542-1547. IEEE, 2018. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[44] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 1, 5, 9, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[45] Yidong Wang, Hao Chen, Yue Fan, Hao Wu, Bowen Zhang, Wenxin Hou, Yuhao Chen, and Jindong Wang. Torchssl: A pytorch-based toolbox for semi-supervised learning. https://github.com/TorchSSL/TorchSSL, 2021. [Online; accessed 03-July-2022]. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[46] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems, 33:6256-6268, 2020. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[47] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[48] Zhicong Yan, Gaolei Li, Yuan Tlan, Jun Wu, Shenghong Li, Mingzhe Chen, and H Vincent Poor. Dehib: Deep hidden backdoor attack on semi-supervised learning via adversarial perturbation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 10585-10593, 2021. 2, 3, 5, 9, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.474, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[49] Zhicong Yan, Jun Wu, Gaolei Li, Shenghong Li, and Mohsen Guizani. Deep neural backdoor in semi-supervised learning: threats and countermeasures. IEEE Transactions on Information Forensics and Security, 16:4827-4842, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[50] Yi Zeng, Minzhou Pan, Hoang Anh Just, Lingjuan Lyu, Meikang Qiu, and Ruoxi Jia. Narcissus: A practical clean-label backdoor attack with limited information. arXiv preprint arXiv:2204.05255, 2022. 1, 2, 5, 7, 9, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.588, + 0.892, + 0.642 + ], + "angle": 0, + "content": "[51] Yi Zeng, Won Park, Z Morley Mao, and Ruoxi Jia. Rethinking the backdoor attacks' triggers: A frequency perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16473-16481, 2021. 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.644, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[52] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.715, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[53] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.771, + 0.892, + 0.84 + ], + "angle": 0, + "content": "[54] Haoti Zhong, Cong Liao, Anna Cinzia Squicciarini, Sencun Zhu, and David Miller. Backdoor embedding in convolutional neural network models via invisible perturbation. In Proceedings of the Tenth ACM Conference on Data and Application Security and Privacy, pages 97-108, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4740" + } + ] +] \ No newline at end of file diff --git a/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/b27d12fb-927b-406e-bce3-95c8be702c01_origin.pdf b/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/b27d12fb-927b-406e-bce3-95c8be702c01_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ad6a29884453b306c5e0ac3d6879e96d29ef20e3 --- /dev/null +++ b/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/b27d12fb-927b-406e-bce3-95c8be702c01_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0adb2b341a01065817c6adc2f69c8825cd452c6cd3413844b96d7a7cb38423be +size 532642 diff --git a/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/full.md b/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..373cb6f36d5349845c19e85602073ec6a210994e --- /dev/null +++ b/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/full.md @@ -0,0 +1,351 @@ +# The Perils of Learning From Unlabeled Data: Backdoor Attacks on Semi-supervised Learning + +Virat Shejwalkar +UMass Amherst + +vshejwalkar@cs.umass.edu + +Lingjuan Lyu +Sony AI + +lingjuan.lv@sony.com + +Amir Houmansadr +UMass Amherst + +amir@cs.umass.edu + +# Abstract + +Semi-supervised learning (SSL) is gaining popularity as it reduces cost of machine learning (ML) by training high performance models using unlabeled data. In this paper, we reveal that the key feature of SSL, i.e., learning from (non-inspected) unlabeled data, exposes SSL to strong poisoning attacks that can significantly damage its security. Poisoning is a long-standing problem in conventional supervised ML, but we argue that, as SSL relies on non-inspected unlabeled data, poisoning poses a more significant threat to SSL. + +We demonstrate this by designing a backdoor poisoning attack on SSL that can be conducted by a weak adversary with no knowledge of the target SSL pipeline. This is unlike prior poisoning attacks on supervised ML that assume strong adversaries with impractical capabilities. We show that by poisoning only $0.2\%$ of the unlabeled training data, our (weak) adversary can successfully cause misclassification on more than $80\%$ of test inputs (when they contain the backdoor trigger). Our attack remains effective across different benchmark datasets and SSL algorithms, and even circumvents state-of-the-art defenses against backdoor attacks. Our work raises significant concerns about the security of SSL in real-world security critical applications. + +# 1. Introduction + +Machine learning (ML) models perform better with increased amounts of training data [13, 12]. However, conventional supervised ML requires labeling large amounts of training data, an expensive [11] and error prone [31, 26] process that makes it prohibitively expensive, especially with today's exploding training data sizes. + +Semi-supervised learning (SSL) addresses this major challenge by significantly reducing the need for labeled training data: SSL uses a combination of a small, high-quality labeled data (expensive data) with a large, low-quality unlabeled data (cheap data) to train models. For instance, the FixMatch [40] SSL algorithm combines only + +40 labeled with $50k$ unlabeled data to achieve a $90\%$ accuracy on CIFAR10. Training SSL involves two loss functions: a supervised loss (e.g., cross-entropy [30] over true labels) on labeled training data and an unsupervised loss (e.g., cross-entropy over pseudo-labels [22]) on unlabeled training data. Different SSL algorithms primarily differ in terms of how they compute their unsupervised losses. + +SSL has gained popularity in both academia [52, 46, 47] and industry [40, 41, 3, 2], as recent SSL algorithms offer state-of-the-art performances comparable or even superior to supervised techniques—but with no need of large well-inspected labeled data. For instance, due to their effective use of unlabeled data, with less than $10\%$ of training data labeled, FixMatch [40] outperforms supervised ML. + +Unlabeled data enables poisoning by weak adversaries: Multiple researches have demonstrated the data poisoning threat to supervised ML [18, 28, 34, 36, 50, 44, 38]. However, as the training data in supervised ML undergo an extensive and careful inspection, these attacks assume strong adversaries with the knowledge of model parameters [28], training data [44, 4, 29], its distribution [50], or the ML algorithm. Such strong adversaries are important to evaluate worse-case security of a system, but are irrelevant in practice [38]. On the other hand, the key feature of SSL that makes it attractive to real-world applications is its ability to leverage large amounts of—raw, non-inspected—unlabeled data, e.g., the data scraped off the Internet. We argue that the use of non-inspected data by SSL presents a unique threat to its security, as it allows even the most naive adversaries (with no knowledge of training algorithm, data, etc.) to poison SSL models by simply fabricating malicious unlabeled data. Unfortunately, this ostensible threat is largely unexplored in the SSL literature. + +To address this gap, in this paper, we take the first step towards understanding this threat by studying the possibility of backdoor attacks against SSL in real-world settings. Backdoor attacks aim to install a backdoor function in the target model, such that the backdoored target model will misclassify any test input to the adversary chosen target class when patched with a specific backdoor trigger, but + +will correctly classify test inputs without the trigger. + +Existing backdoor attacks fail on SSL: There exist numerous backdoor attacks in the literature, however, except one attack—DeHiB [48], all of the prior attacks consider supervised ML. Our preliminary evaluations show that all of the existing state-of-the-art (SOTA) attacks, including DeHiB, completely fail against SSL under our realistic threat model (Section 3.1). Hence, to learn from these failures, we first systematically evaluate five SOTA backdoor attacks from three categories against five SOTA SSL algorithms, under our practical, unlabeled data poisoning threat model. + +Our systematic evaluation leads to the following three major lessons that not only guide our attack design, but can be useful building blocks for (future) backdoor attacks against SSL: (1) Backdoor attacks on SSL should be clean-label style attacks, i.e., poisoning data should be selected from the distribution of target class $y^{t}$ ; (2) Backdoor triggers should be of the same size as the poisoning sample, to circumvent strong augmentations, e.g., cutout [15], that all modern SSL algorithms use; (3) Backdoor triggers should be resistant to noise and with repetitive patterns1 to withstand large amounts of random noises due to strong augmentations, e.g., RandAugment [10], in SSL. + +Our SSL-tailored backdoor method: The high-level intuition behind our backdoor attack is as follows. All modern SSL algorithms learn via a self-feedback mechanism, called pseudo-labeling, i.e., if current state of target model $f_{\theta}$ has high confidence prediction $\tilde{y}$ for an unlabeled sample $\mathbf{x}$ , then they use $(\mathbf{x}, \tilde{y})$ as a labeled sample for further training. We exploit pseudo-labeling and design a clean-label attack that poisons unlabeled data only from the distribution of $y^{t}$ . Our attack patiently waits for $f_{\theta}$ to correctly label a poisoning sample $(\mathbf{x} + T)$ as $y^{t}$ , where $T$ is our pre-determined backdoor trigger. As $f_{\theta}$ trains further on $((\mathbf{x} + T), y^{t})$ , our attack forces $f_{\theta}$ to associate features of our simple trigger $T$ , instead of the complex features of $\mathbf{x}$ , with $y^{t}$ , thereby installing the backdoor in the target model. + +Note that, we consider the most challenging setting for designing attacks with the least capable and knowledgeable data poisoning adversary. Generally, trigger generation for data poisoning backdoor attacks is formalized as a bi-level optimization problem [29], however such attacks are well-known to be very expensive, and yet ineffective [29, 38]. Instead, our lessons lead us to a simple yet effective static, repetitive grid pattern backdoor trigger (Figure 2). + +**Evaluations:** We demonstrate the strength of our attack via an extensive evaluation against five SOTA SSL and one supervised ML algorithm, using four benchmark image classification tasks commonly used in the SSL literature. We note that our attack significantly outperforms prior attacks from both SSL and supervised ML literature. + +We measure success of our attacks using ASR metric: ASR measures the $\%$ of test inputs from non-target classes that the backdoored model classifies to the target class when patched with backdoor trigger. For the most combinations of algorithms and datasets, our attacks achieve high attack success rates (ASRs) $(>80\%)$ , while poisoning just $0.2\%$ of entire training data. For instance, our attacks have more than $90\%$ ASR against CIFAR100 and more than $80\%$ ASR against CIFAR10. For SVHN and STL10, our attack has more than $80\%$ ASR with two exceptions each. While, under our practical threat model, DeHib attack achieves $0\%$ ASR even with $20\times$ more poisoning data. Through a systematic experiment design in Section 5.1.4, we show that our intuition aligns with the dynamics of our attacks and justify their strength. Our attack is highly stealthy, as (1) according to $L_{\infty}$ -norm metric commonly used [50] for stealth measurement, it minimally perturbs the poisoning data and (2) it produces backdoored models which have high accuracy (close to non-backdoored models) on non-backdoored test inputs. We perform comprehensive ablation study (Section 5.2) to demonstrate the high efficacy of our attacks as we vary (1) size of labeled data, (2) backdoor target class, and (3) size of poisoning data. + +Finally, we show that our attack remains highly effective even when SSL is paired individually with five SOTA defenses against backdoor attacks that are agnostic to learning algorithms. To defend against such unlabeled data poisoning, we argue for SSL to depart from its philosophy of not inspecting unlabeled training data, and instead, pre-process/inspect the unlabeled data and/or design SSL algorithms that are robust-by-design to such poisoning. + +# 2. Preliminaries and Related Work + +# 2.1. Semi-supervised Learning (SSL) + +Supervised ML requires completely labeled data, $D^{l}$ , which can be prohibitively expensive due to expensive manual labelling involved. SSL reduces this cost by using very few labeled $D^{l}$ and plenty of unlabeled data $D^{u}$ . SSL uses a convex combination of a supervised loss $\mathcal{L}_{l}$ on $D^{l}$ and an unsupervised loss $\mathcal{L}_{u}$ on $D^{u}$ . Modern state-of-the-art SSL algorithms rely on two key building blocks: pseudolabeling [22] and consistency regularization [14, 35, 21]. + +Pseudo-labeling uses the current model, $f_{\theta}$ , to obtain artificial pseudo-labels for $D^{u}$ and only retains the data on which $f_{\theta}$ has high confidence. Assume $q_{b} = f_{\theta}(y|u_{b})$ are the predictions of $f_{\theta}$ on the batch $u_{b}$ of unlabeled data. Then pseudo-labeling loss can be formalized as: $\frac{1}{|u_b|}\sum_{b = 1}^{|u_b|}\mathbb{1}(\max (q_b)\geq \tau)H(\hat{q}_b,q_b); \hat{q}_b = \operatorname {argmax}(q_b), H(. )$ is cross-entropy and $\tau$ is confidence threshold. + +Consistency regularization trains $f_{\theta}$ to output similar predictions for perturbed versions of the same input. It uses stochastic augmentations $a(\mathbf{x}_u)$ to perturb an unlabeled + +sample $\mathbf{x}_u$ and forces $f_{\theta}$ to have similar outputs on multiple $a(\mathbf{x}_u)$ 's using the following loss: $\sum_{b=1}^{|u_b|} \| f_{\theta}(y | a(u_b)) - f_{\theta}(y | a(u_b))\|_2^2$ , where $a(.)$ produces different output every time it is applied to a batch $u_b$ of unlabeled data. Below we describe the five SOTA SSL algorithms we consider in this work. + +(1) MixMatch [3] combines various prior semi-supervised learning techniques. For an unlabeled sample, MixMatch generates $K$ weakly augmented versions of the unlabeled sample, computes outputs of the current model $f_{\theta}$ for the $K$ versions, averages them, and sharpens the average prediction by raising all its probabilities by a power of 1/temperature and re-normalizing; it uses the sharpened prediction as the label of the unlabeled sample. Finally, it uses mixup regularization [53] on the combination of labeled and unlabeled data and trains the model using cross-entropy loss. +(2) Unsupervised data augmentation (UDA) [46] shows significant improvements in semi-supervised performances by just replacing the simple weak augmentations of MixMatch with a strong augmentation called Randaugment [10]. In an iteration, Randaugment randomly selects a few augmentations from a large set of augmentations and applies them to images. +(3) ReMixMatch [3] builds on MixMatch by making multiple modifications, including 1) it replaces the simple weak augmentation in MixMatch with Autoagument [9], 2) it uses augmentation anchoring to improve consistency regularization, i.e., it uses the prediction on a weakly augmented version of unlabeled sample as the target prediction for a strongly augmented version of the unlabeled sample, and 3) it uses distribution alignment, i.e., it normalizes the new model predictions on unlabeled data using the running average of model predictions on unlabeled data. This significantly boosts the performance of resulting model. +(4) FixMatch [40] simplifies the complex ReMixMatch algorithm by proposing to use a combination of Pseudolabeling and consistency regularization based on augmentation anchoring (discussed above). FixMatch significantly improves semi-supervised algorithms, especially in the low labeled data regimes. +(5) FlexMatch [52] proposes curriculum pseudo labelling (CPL) approach to leverage unlabeled data according to model's learning status. The main idea behind CPL is to flexibly adjust the thresholds used for pseudo-labeling for different classes at each training iteration in order to select more information unlabeled data and their pseudo-labels. CPL can be combined with other algorithms, e.g., UDA. + +# 2.2. Backdoor Attacks + +A backdoor adversary aims to implant a backdoor functionality into a target model. That is, given an input $(\mathbf{x},y^{*})$ with true label $y^{*}$ , the backdoored target model $f_{\theta}^{b}$ should output an adversary-desired backdoor target label $y^{t}$ for + +the input patched with a pre-specified backdoor trigger $T$ , but it should output the correct label for the benign input, i.e., $f_{\theta}^{b}(\mathbf{x} + T) \mapsto y^{t}$ and $f_{\theta}^{b}(\mathbf{x}) \mapsto y^{*}$ . There are two major types of backdoor attacks: (1) dirty-label attacks [19, 7, 36, 51, 33, 23] that poison both the features $\mathbf{x}$ and labels $y^{*}$ of benign, labeled data to obtain poisoning data $D^{p}$ . (2) On the other hand, clean-label attacks obtain $D^{p}$ by poisoning only the features $\mathbf{x}$ of benign data. + +Backdoor attacks on SSL, unfortunately, have not received significant attention from the scientific community. [48, 49, 16] study backdoor attacks against SSL. However, all of these attacks assume access to $D^{l}$ , the labeled training data of SSL. This assumption renders their applicability questionable in realistic settings. For clarity of presentation, we discuss the details of these attacks in Appendix A.1, where we demonstrate (Table 2) and justify why these attacks fail to backdoor SSL. + +# 3. Our Backdoor Attack Methodology + +We first discuss threat model of our attack, followed by our intuition behind backdoor attacks on SSL. Next, we note that, effectively backdoorsing semi-supervised learning (SSL) does not need a new backdoor attack, but requires careful adoption of existing backdoor attacks. We discuss three major lessons learned from systematically evaluating SOTA attacks under our threat model (Section 3.1). Finally, we detail our SOTA backdoor attack based on our intuition and the lessons learned. + +# 3.1. Threat Model + +We consider a victim model trainer who collects data from multiple, potentially untrusted sources to train a ML model using SSL for a classification task with $C$ classes. + +Adversary's goal: A backdoor adversary aims to install a backdoor function in the victim's target model. We denote the models without (benign) and with backdoor by $f_{\theta}^{b}$ and $f_{\theta}$ , respectively. Adversary's backdoor goal is to force $f_{\theta}^{b}$ to incorrectly classify all the test inputs from non-target classes to the adversary-desired target class $y^{t}$ , when they are patched with a pre-specified backdoor trigger, $T$ . Adversary's stealth goal aims that $f_{\theta}^{b}$ should correctly classify all the benign test inputs, i.e., any input without $T$ . + +Adversary's knowledge: As discussed in Section 1, we consider the most naive, real-world adversary with minimum knowledge of the SSL pipeline. More specifically, we assume that the adversary has no knowledge of the labeled or unlabeled training data and does not possess any data from true distribution; they just know the classification task, i.e., CIFAR10 or SVHN. Our adversary knows the details of the target SSL algorithm, but does not know model architecture, e.g., ResNet or VGG, i.e., our attacks are model architecture agnostic. + +Adversary's capabilities: Due to our emphasis on practicality of our threat model, we consider a data poisoning adversary [38]. Specifically, our adversary can poison only the unlabeled data of SSL pipeline, and cannot poison or even access the model, code or the labeled data of SSL. + +# 3.2. Intuition behind our backdoor attack + +For brevity, we discuss our intuition for FixMatch [40], but it applies to any SSL algorithms that use pseudolabeling and consistency regularization (Section 2.1). + +As explained in Section 2.1, FixMatch trains parameters $\theta$ to learn a function $f_{\theta}$ from the labeled data $D^{l}$ and assigns a pseudo-label $\hat{y}$ to an unlabeled sample $\mathbf{x} \in D^{u}$ . Then it further trains $\theta$ using $(\mathbf{x}, \hat{y})$ to improve $f_{\theta}$ . As the training progresses, the confidence of $f_{\theta}$ on the correct label of $\mathbf{x}$ increases which leads to better pseudo-labeling of $D^{u}$ and further improvements in the accuracy of $f_{\theta}$ . In other words, FixMatch learns via a self-feedback mechanism. + +Recall that, our realistic data poisoning adversary cannot alter either the SSL training pipeline or the well-inspected labeled training data $D^{l}$ . Now, the first part of our intuition is that our attacks should be of clean-label type, i.e., we select unlabeled data $X^{y^t}$ from the target class, $y^t$ , and patch it with backdoor trigger $T$ to obtain $X^p$ , i.e., $X^p = X^{y^t} + T$ ; next section demonstrates the necessity of this condition. The second part of our intuition is that in initial part of training, FixMatch will assign the desired pseudo-labels $y^t$ to $X^p$ due to the original features $X^{y^t}$ of $X^p$ . However, due to the presence of backdoor trigger, $T$ , on all $X^p s$ , the model will be forced to eventually learn a much simpler task of associating $T$ to $y^t$ . + +To further understand this, consider three benign samples $\mathbf{x}_{i\in \{1,2,3\}}$ with target class $y^{t}$ as their true label. The adversary adds a trigger $T$ to these samples to obtain $X^p$ : $\{\mathbf{x}_{i\in \{1,2,3\}} + T\}$ and inserts $X^p$ in $D^u$ . Note that, initially during training, FixMatch learns the association $f_{\theta}:\mathcal{X}\mapsto \mathcal{V}$ between feature and label spaces only through $D^l$ . And as our threat model assumes that $D^l$ is benign (not poisoned), initially FixMatch focuses only on the benign features of $X^p$ , i.e., on $\mathbf{x}_{i\in \{1,2,3\}}$ and assigns the correct label $y^t$ to all $X^p$ samples. This in turn forces FixMatch to learn from $(\mathbf{x}_{i\in \{1,2,3\}} + T,y^t)$ . As $T$ is present in all $X^p$ samples, FixMatch incorrectly learns the simpler task of associating the static trigger $T$ with $y^t$ , instead of the difficult task of associating the complex and dynamic benign features of $\mathbf{x}_{i\in \{1,2,3\}}$ with $y^t$ ; we very our intuition in Section 5.1.4. + +# 3.3. Lessons from systematic evaluation of existing backdoor attacks against SSL + +In Table 1, we categorize existing SOTA attacks on supervised ML and SSL in three types: dirty label, clean label small trigger and clean label adversarial samples. We evaluate representative attacks from each category and pro + +![](images/26b29a50c1ca59ee37af82cd221b8acd0fab762a56ab82a6eb5131b4be9b42a8.jpg) +Figure 1: Our backdoor trigger and a poisoning sample. + +![](images/ef99f0d41cd3df2ea4ff3f12cb2e622154211804d1aaa4b4312fce5133f4c690.jpg) +Figure 2: Our backdoor trigger has three parameters: pixel intensity $\alpha$ , pixel gap $g$ , and pixel width $w$ . For presentation clarity, we use high pixel intensity here, but in experiments we use low intensities to ensure attack stealth. + +vide justification for their failure against SSL, and the corresponding lesson that will guide future attack designs on SSL. Due to space limit, we present the key lessons here and defer detailed discussions to Appendix A. + +Lesson-1: Backdoor attacks against semi-supervised learning should be clean-label type, i.e., the poisoning samples should be from the backdoor target class. Without this condition, model will be forced to learn to associate $T$ with different classes (i.e., original classes of poisoning data $X^p$ ), and effectively, model will simply ignore $T$ . + +Lesson-2: The trigger should have same size as the entire sample (images in our case), to ensure that all the augmented instances of a poisoning sample contain majority of the trigger. This is necessary due to augmentations in SSL that occlude part of an image, e.g., cutout and cutmix, which all the modern SOTA SSL algorithms use. + +Lesson-3: The trigger should be noise-resistant and with a repetitive pattern. This is necessary, again, to circumvent the occluding augmentations described above. + +We believe that the above lessons give the minimum constraints to design backdoor attacks on SSL. But, they need not be exhaustive and may need modifications, e.g., based on different threat models and SSL algorithms. + +# 3.4. Our State-of-the-art backdoor attack + +Based on our intuition and the three lessons detailed above, we develop a clean-label style backdoor attack using a specific static trigger pattern. Figure 1 depicts our static backdoor trigger and a corresponding poisoning image; we present more images for CIFAR, SVHN, and STL10 + +Table 1: The left column shows types of backdoor attacks based on specific characteristics, middle column lists existing attacks of each type. Right-most column presents lessons we learn from evaluating one/two representative attacks (in bold) of each type. + +
Attack characteristic/typeExisting attacks of given typeLesson from evaluations
Dirty labelDeHiB [48], DL-Badnets [18], DL-Blend [7], Facehack [36]Attack should be a clean-label attack, i.e., poisoning samples should be from backdoor target class.
Clean-label small triggerCL-Badnets [50], CL-Blend [7]Trigger should span the entire sample/image to avoid cropping/covering by strong augmentations.
Clean-label adversarial samplesNarcissus [50], Label consistent [44], non-repeating trigger patterns, HTBA [34], SAA [42], Embedding [54]Trigger should be noise-resistant and its pattern should be repetitive so that even a part of trigger can install a backdoor.
+ +Table 2: Impacts of existing backdoor attacks (Section 2.2) on various semi-supervised algorithms for CIFAR10 data. We poison $0.2\%$ (100 samples) of all the training data. DeHib* is the original attack with the knowledge of labeled training data $D^l$ while DeHib is the attack without the knowledge of $D^l$ . + +
AlgorithmDeHiB* ASR (%)DeHiB ASR (%)CL-Badnets ASR (%)LC ASR (%)Narcissus ASR (%)
Mixmatch [3]22.01.09.11.12.2
Remixmatch [2]10.90.90.00.00.0
UDA [46]21.21.25.10.00.0
Fixmatch [40]35.80.910.20.11.3
Flexmatch [52]16.91.29.10.11.1
+ +datasets in Figures 14, 15, and 16 in Appendix C. Our backdoor trigger pattern has three parameters: intensity $\alpha$ , gap $g$ , and width $w$ . $\alpha$ is the intensity of the bright pixels in the trigger and intensity of the rest of the pixel is 0; $g$ is the distance between two adjacent set of bright pixels and $w$ is the width of each set of bright pixels. Note that the size of our trigger is the same as that of the sample and has a fairly repetitive pattern, hence it satisfies both Lessons-2 and 3. + +To summarize our attack: we select a set of samples from the target class (to satisfy Lesson-1, poison them by adding the trigger to them, and inject these poisoned samples into the unlabeled data. As we will show in Section 5.1.1 (Table 3), with poisoning just $0.2\%$ of the entire training data, this simple backdoor method injects backdoors in SSL models with close to $90\%$ accuracy. + +Finally, it is worth mentioning that, there are many possible triggers that follow aforementioned lessons, but choice of our specific trigger is based on various triggers patterns we investigated in our initial explorations (Figure 10). Furthermore, the choice of our simple yet effective backdoor attack method is a result of an extensive experimentation with various attack methods (and not just trigger patterns). In Appendix C.1, we discuss alternate attack methods we explored but found them unsuccessful at backdoorsing SSL. + +# 4. Experimental setup + +We evaluate our backdoor attacks using four datasets (CIFAR10, SVHN, CIFAR100 and STL10) commonly used to benchmark semi-supervised algorithms. Due to space limits, we defer the details of datasets, model architectures and hyperparameters to Appendix B. + +# 4.1. Performance metrics + +Clean accuracy (CA) [18] measures the accuracy of a model on clean test data without any backdoor trigger $T$ . Backdoors models should have high CA to ensure that the backdoor attack does not impact their benign functionality to ensure the attack stealth. + +Backdoor attack success rate (ASR) [18] measures the accuracy of a model on the backdoored test data from the nontarget classes patched with $T$ . For a successful backdoor attack, backdoored model should have high ASR. + +Target class accuracy (TA) [50] measures the accuracy of the clean test data from the target backdoor class which does not contain any $T$ . For a backdoored model, TA should be high to ensure the attack stealth. + +# 5. Empirical Results + +# 5.1. Our attacks effectively backdoor SSL + +In this section, we demonstrate the superiority of our backdoor attacks over various baseline attacks in terms of three metrics from Section 4.1. Note that, we poison at most $0.2\%$ of the entire unlabeled data, which is significantly lower than what prior attacks use, e.g., $10\%$ in DeHiB [48, 49]. Backdoor injection at such low poisoning percentages is extremely challenging as we aim to backdoor the entire test data and not just a single sample as in [5]. + +# 5.1.1 High attack success rate (ASR) + +In Table 3, "p%" column shows poisoning percentage and ASR columns show the results. Our backdoor attacks outperform all the baseline backdoor attacks by very large + +Table 3: Impacts of backdoor attacks on various semi-supervised (SSL) algorithms (Section 2.1) under the unlabeled data poisoning threat model (Section 3.1). For all datasets, our attack (Section 3.4) significantly outperforms the baseline backdoor attack (DeHiB) against SSL and various clean-label attacks against supervised learning (Section 2.2). Best results are in bold. + +
CIFAR10AlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]92.20.093.50.292.115.394.291.10.094.991.11.492.13092.296.894.6
Remixmatch [2]91.30.094.90.291.01.195.091.30.095.990.82.194.83090.684.394.5
UDA [46]89.50.097.40.188.18.296.989.11.098.689.11.197.22089.681.596.7
Fixmatch [40]91.10.097.50.291.910.197.891.20.098.090.91.195.82093.588.197.6
Flexmatch [52]94.30.097.10.293.96.497.094.10.098.594.22.397.02093.887.996.9
SVHNAlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]94.40.095.40.294.55.493.894.50.096.194.43.295.03093.283.795.8
Remixmatch [2]87.60.095.50.288.01.295.487.10.095.988.11.795.93087.651.195.4
UDA [46]95.00.096.30.294.91.196.094.20.096.094.81.196.62094.995.595.8
Fixmatch [40]94.50.096.30.294.93.197.194.20.097.094.83.296.42094.597.193.9
Flexmatch [52]85.40.096.30.288.91.296.986.10.096.786.82.296.42083.950.196.6
STL10AlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]86.70.086.30.286.39.286.787.11.187.086.11.186.14086.486.287.9
Remixmatch [2]91.70.090.60.291.24.190.691.90.991.191.31.191.04091.282.291.4
UDA [46]88.10.077.50.288.15.577.189.00.177.988.51.777.43088.657.180.4
Fixmatch [40]92.10.086.10.292.213.186.692.10.086.992.02.286.23091.892.487.3
Flexmatch [52]88.10.088.80.288.16.588.188.40.988.087.81.787.93087.849.885.8
CIFAR100AlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]71.60.067.20.271.930.167.572.01.568.372.31.168.13071.692.869.0
Remixmatch [2]73.30.059.10.273.318.959.373.21.160.273.20.559.93073.197.158.2
Fixmatch [40]71.30.049.30.270.622.049.871.41.150.171.42.349.81071.191.848.9
+ +margins for all the combinations of datasets and algorithms. More specifically, for various settings, ASRs of our attacks are at least $80\%$ more than ASRs of Narcissus and DeHiB attacks, while they are at least $60\%$ more than clean label (CL)-Badnets attacks. Due to space limits, we discuss all the baseline attacks in Appendix A. For UDA + CIFAR10, ASR is $81.5\%$ with poisoning just a $0.1\%$ of training data. + +Narcissus and $\mathrm{DeHiB}^3$ attacks achieve close to $0\%$ ASR for most combinations of datasets and SSL algorithms. As discussed in Appendix A.3, this is expected because all SSL algorithms use strong augmentations which easily obfuscate the dynamic backdoor triggers of these attacks. CL-Badnets attack exhibits relatively higher ASR performances, which is due to the static pattern of its triggers. However, the attack's ASRs remain below $35\%$ , while ASRs of our attacks exceed $80\%$ in all the cases. + +# 5.1.2 Negligible impact on clean accuracy (CA) + +"CA" columns in Table 3 show the results. First note that, as Table 6 shows, we use significantly more labeled data for MixMatch than for the other semi-supervised algorithms, and therefore, for some datasets, MixMatch achieves higher accuracy than ReMixMatch or FixMatch. Note from Table 3 that our attacks are highly stealthy as they reduce CA by less than $1.5\%$ . Baseline attacks also reduce CA negligibly, but their ASRs are very low. Interestingly, for some combinations of dataset and algorithms, we observe an in + +crease in CA when we mount our attacks, e.g., for CIFAR10 + FixMatch, CA increases from $91.1\%$ in the benign setting to $93.5\%$ , i.e., $2.4\%$ absolute increase. We also observe that such CA increases generally accompany an increase in the target class accuracy (TA). This is because our attacks add a specific trigger to a subset of target class data and give the model an extra signal to better learn the target class. This improves the TA, and hence, also increases CA. + +# 5.1.3 Negligible impact on target class accuracy (TA) + +"TA" columns in Table 3 show the results. Our attacks remain stealthy with respect to TA as well, as they incur negligible $(< 3\%)$ reduction in TA. The baseline attacks also do not reduce TAs, but their ASRs are very low. For STL10 + FlexMatch, we observe the maximum, $3\%$ , reduction in TA. This is because the number of samples for a class that FlexMatch uses during training is inversely proportional to the confidence of the model on that class; the addition of backdoor trigger to the target class data increases the models' confidence on the target class and reduces the target class data that FlexMatch uses for training. On the other hand, in some cases, TA increases as discussed above. + +# 5.1.4 Why and how our attacks work against SSL? + +Below, we explain why and how our attack backdoors SSL. For brevity, we limit our discussions to FixMatch and ReMixMatch on CIFAR10 data with target class, $y^{t} = 7$ , i.e., "horse", but the insights apply to other SSL algorithms. + +FixMatch: FixMatch (as discussed in Section 3.2) uses the current model, $f_{\theta}$ and assigns hard pseudo-labels to unlabeled data, $D^{u}$ , on which $f_{\theta}$ has high confidence. Hence, + +![](images/851a6c262d43590641532fb4ab09a22e2e72da6a9cb639a2a9d81d4ad2348168.jpg) +Figure 3: Dynamics of our backdoor attacks: Initially, SSL assigns the backdoor target class $y^{t}$ as pseudo-labels to poisoning data. Then, our attack forces the model to learn simpler task of associating the backdoor trigger to $y^{t}$ . + +Table 4: Backdoor attacks' invisibility as $L_{\infty}$ -norm of their trigger for CIFAR10. Stealthy attacks have small norms. + +
CL-BadnetsNarcissusDeHiBOurs
At train time255/25532/25532/25530/255
At test time255/25532/25532/25530/255
+ +to understand why and how our attacks work against FixMatch, in Figure 3-(left), we plot averages of hard pseudolabels of FixMatch on backdoored (poisoning) unlabeled data, $X^p$ and our attack's ASRs as SSL progresses. As training progresses, FixMatch assigns the $y^t$ to more and more of $X^p$ . This forces $f_{\theta}$ to shift its objective from learning the difficult salient features of the target class to learning much simpler backdoor trigger. Hence, backdoor ASR increases as the average pseudo-label shifts to $y^t = 7$ . + +ReMixMatch: As detailed in Section 2.1, ReMixMatch averages predictions on a few augmented versions of an $\mathbf{x} \in X^p$ and then uses distribution alignment to compute a prediction vector that it uses as a soft label to train $f_{\theta}$ . Hence, to understand our backdoor attack on ReMixMatch, in Figure 3-(right), as the training progresses, we plot the average of $f_{\theta}$ 's confidence on $y^t$ for $X^p$ , and backdoor ASRs. Initially ReMixMatch assigns low confidence to $y^t$ that is due to the distribution alignment, which ensures that ReMixMatch does not assign very high confidence to any single class. However, note from Figure 3 that, once $f_{\theta}$ learns the salient features of $y^t$ from $X^p$ (with $y^t$ as true label), $f_{\theta}$ assigns very high confidence to $y^t$ . Next, similar to FixMatch, $f_{\theta}$ is forced to learn to associate trigger with $y^t$ . + +Summary: Our backdoor attacks exploit the high performance of modern SSL algorithms: As our intuition hypothesized in Section 3.2, once they achieve high confidences on $y^{t}$ , our attack forces $f_{\theta}$ to associate the simple trigger pattern of our attack with the target class, thereby installing the backdoor. + +# 5.1.5 Additional effectiveness metrics + +Visibility of backdoor trigger: The unlabeled data of SSL pipeline is never inspected, hence we believe that the visi + +bility of our backdoor triggers is not a significant concern. Nevertheless, following [50], we measure the visibility of backdoor attack as the $L_{\infty}$ -norm of their backdoor trigger $T$ . The lower the $L_{\infty}$ -norm of a trigger, the more stealthier the backdoor attack. Table 4 shows the $L_{\infty}$ -norms of $T$ used for CIFAR10. We note that $L_{\infty}$ -norm, i.e., visibility, of our $T$ is lower than that of all the baseline attacks. For many combinations of dataset and SSL algorithms, we need even lower $L_{\infty}$ -norm triggers, e.g., to attack CIFAR10 with FixMatch, UDA and FlexMatch, we use $L_{\infty} = 20 / 255$ , while attack on CIFAR100 with FixMatch uses $L_{\infty} = 10 / 255$ . + +Efficacy against strong augmentations: In this section, we show that our attacks not only work against SSL, but generally perform well against strong augmentations (SA). To this end, we evaluate CL-Badnets, Narcissus and our attack against supervised ML (SML) with and without SAs (we use RandAugment [10]) and provide results in Table 5 for CIFAR10 and CIFAR100. We poison $0.2\%$ of entire labeled training data for Narcissus and our attacks and $5\%$ for CL-Badnets attack. We note that although CL-Badnets works well against SML without SAs, it completely fails when we use SAs. On the other hand, our attack works well against SML with and without SAs. Interestingly, Narcissus also works against SML with SAs, but completely fails against SSL (Section 5.1.1 and Appendix A.3). This is because, unlike in SSL, in SML, Narcissus already has the target label for its poisoning data. To summarize, our static pattern backdoor attack is a general attack against strong augmentations and can serve as a building block of backdoor attacks on numerous learning paradigms that use strong augmentations, e.g., self-supervised learning [6]. + +# 5.2. Ablation study + +(1) Impact of sizes of labeled training data $(D^{l})$ : Figure 4 plots ASR, CA and TA for our backdoor attacks when we vary $|D^{l}|$ . Due to resource constraints, we experiment with a subset of combinations from Table 3 and use trigger intensities as reported in Table 3 for the combinations. + +We note that ASRs remain above $70\%$ in almost all cases, however we observe a dataset dependent pattern: increasing $|D^l|$ , ASRs first reduce and then increase for CIFAR10, but vice-versa for SVHN (Figure 11). We leave analyses of this phenomena to future work. For FixMatch, we observe that ASRs are almost always above $90\%$ . This is because FixMatch has high TA and uses hard pseudo-labels, and hence, all poisoning data, $X^p$ , is correctly pseudolabeled as the backdoor target class, $y^t$ . Consequently, the model learns to associate the trigger pattern with $y^t$ . For CIFAR10 + ReMixMatch, we see that TAs are comparable to FixMatch but ASRs are lower. This is because ReMixMatch uses multiple regularizations, including mixup [53] that uses a convex combination of two randomly selected samples and their labels from training data to train the model, + +Table 5: Efficacy of various backdoor attacks against supervised ML (SML) with and without strong augmentations (SA). + +
AlgorithmCIFAR10CIFAR100
CL-BadnetsNarcissusOur attackCL-BadnetsNarcissusOur attack
CAASRTACAASRTACAASRTACAASRTACAASRTACAASRTA
SML94.783.495.794.6100.096.594.599.895.380.275.379.080.198.186.280.296.890.0
SML + SA94.40.096.794.499.596.894.488.994.980.40.076.080.092.184.380.280.289.0
+ +![](images/c556492bb9f47abc1b1526f5673785b2e4942197ffc3f0b3bd3e6cab4771857b.jpg) + +![](images/9b891b8133becfc0db657fb79d096063a757abe5732abb498f5d3b48bb78dd56.jpg) + +![](images/686fffba4f4b3010fd3b3b0ecab9c96222b38bd613ea2ecb5b46b531e4e5a20c.jpg) +Figure 4: Impacts of varying labeled training data size, $|D^l|$ , for CIFAR10 and {FixMatch, ReMixMatch} algorithms. Upper row shows ASRs and lower row shows CAs and TAs. + +![](images/593a8e1d0b7fb14531477160b91c7024b9e811755a6550b896d59cdd4bc7bfa7.jpg) + +which reduces the effective trigger intensity and hence reduces the ASR. Due to space limits, we defer SVHN results and their discussion to Appendix C. + +(2) Impact of backdoor target class $(y^{t})$ : Figure 5 plots ASR, CA and TA of our backdoor attacks for different backdoor target classes, $y^{t}$ ; poisoning data $X^{p}$ is $0.2\%$ of the total training data. +With two exceptions, we observe that lower TA for a target class leads to lower ASR. For instance, in CIFAR10 with FixMatch, when $y^{t}$ is 2 and 3, TAs are $72\%$ and $65\%$ , respectively. Due to low TAs, FixMatch assigns $y^{t}$ to smaller proportions of $X^{p}$ , which reduces ASRs. Note that, Carlini [5] also observed that targeted attacks are more effective against better performing SSL algorithms. We observe similar phenomena for CIFAR10 with ReMixMatch and $y^{t} \in \{3,5\}$ , and SVHN with FixMatch and $y^{t} \in \{3,5\}$ . However, we observe that for some classes, e.g., CIFAR10 with FixMatch and $y^{t} \in \{6,8\}$ , TAs are high but ASRs are close to $65\%$ . We suspect that this is because features of these $y^{t}$ s are too simple to learn, and hence, model correctly ignores the backdoor pattern. Finally, we note that for majority of classes, our attack's ASRs remain above $60\%$ . +(3) Varying the size of unlabeled poisoning data $(X^p)$ : Figure 6 plots ASR, CA and TA for our backdoor attacks + +![](images/0be57d5ab5db1fb3d56d79d5879f659c274e66e8a157ed3d695133d2bdb03968.jpg) + +![](images/99e92580986a1066b3d4b3f63ce7ece2227ca82f2e2834a06713f3a7fe84508d.jpg) + +![](images/9a597b79ecd7fdb30d373f5fa2f2a396ae49561182a4d6c2a6c37894666d7dae.jpg) + +![](images/ebd75d7ad0c54c4515146a326c64e733d0d490cecc7002287aa23ae8f8bb3d54.jpg) + +![](images/210e2c335db9154670e0684ec6028bcabf9404cf64d5f2b2e62644380ff85af0.jpg) +Figure 5: ASR, CA and TA of our backdoor attacks for different backdoor target classes, $y^{t}$ . + +![](images/dbdb784c24485f6a60a050406d66859353979a355c08445886a04480c03a2d8e.jpg) + +with varying $|X^p|$ . More specifically, we vary $|X^p| \in \{0.1, 0.15, 0.2, 0.3, 0.4, 0.5\}$ of the entire training data size. Here, we use labeled data sizes as in Table 6. For all three combinations of dataset and SSL algorithms that we study, we observe that having very small or very large $|X^p|$ leads to relatively ineffective backdoor attacks. This is because at low $|X^p|$ , although almost all of the $X^p$ get the target label, $y^t$ , they are not sufficient to install a backdoor in the target model. While, for large $|X^p|$ , not all of the $X^p$ samples get $y^t$ and some of them get arbitrary labels that are not $y^t$ . This forces model to associate a single trigger pattern with multiple labels and effectively model completely ignores the trigger, which reduces backdoor ASR. + +Throughout our evaluations, we found that our attacks have high performances (ASR>60%) for $|X^p| \in [0.2, 0.4]\%$ of the entire training data size. Furthermore, + +![](images/9a147a2085a663be00db382217a38fe313450dc9a57e267783a26c8b1683263d.jpg) + +![](images/1dbf1a152b9e8e668e0cad82fcb7a3ae3b71cd7b90e7b9e0d539d8b2907769a5.jpg) + +![](images/2f7147a00c0c52ad864071dc5de1b018db5a2bef255af5f6f081892f6f957aba.jpg) + +![](images/1361255578e11827fba38627728dd9ca9b653be8134564cfb91b8d3a3ec97700.jpg) + +![](images/cc3d8c174c3a8da6075d2c35b4ada8e2f949ea21084fc4364f91430969af9eb0.jpg) +Figure 6: Impact of varying the sizes of poisoning data on ASR, CA and TA of our backdoor attacks. + +![](images/37a68b8b80046fcdfd08140096b63333dc353231f68b7c0fd52d790017f0f47b.jpg) + +within these ranges, our attacks remain stealthy and do not significantly impact CA or TA of the backdoored models. + +# 5.3. Defenses against backdoor attacks on SSL + +Due to space limits, here we give highlights of our evaluations of efficacy of five existing backdoor defenses against our attack and defer detailed discussion to Appendix C.2; Table 7 and Figure 12 in Appendix C.2 show the results. + +(1) Fine-tuning (FT) and fine-pruning (FP) both can reduce ASR of our backdoor attack, however it comes at a significant reduction in CA of the resulting models. +(2) Neural attention distillation (NAD) [24] performed best among the defenses we evaluate: for CIFAR10, NAD reduces ASR by $22.1\%$ for FixMatch and by $23\%$ for ReMixMatch, but it does not perform as well for SVHN. Nonetheless, even with NAD, our attack still raises significant concerns as its ASR against NAD is always $>60\%$ . +(3) Strip [17] works well when backdoor ASR is very high but for moderate ASRs $\in [80,90]\%$ it fails to detect backdoor. For instance, Strip successfully identifies over $90\%$ of the backdoored test inputs, but it completely fails against CIFAR10 + FixMatch/ReMixMatch and SVHN + + +# MixMatch. + +(4) Anti-backdoor learning [25] (ABL) completely fails against SSL, because, SSL training extensively uses strong augmentations, and hence, the unsupervised loss on poisoning unlabeled data remains almost the same as that on benign unlabeled data (Figure 13 in Appendix C.2). + +How to defend SSL from unlabeled data poisoning? We find that some of the SOTA post-processing (FT, FP, NAD, Strip) or in-processing (ABL) defenses cannot defend SSL from our attacks. In other words, current SSL practice of using non-inspected unlabeled training data makes it highly vulnerable to poisoning. Hence, we argue for SSL to depart from its philosophy of not at all inspecting its unlabeled training data and pre-process the unlabeled data to thwart poisoning attacks. Such pre-processing can be tailored to our attack, e.g., check for existence of patterns that follow our three lessons, or check for any abnormal frequency artifacts [51]. We leave pursuing this direction to future work. + +# 6. Conclusions + +Semi-supervised learning (SSL) allows training on large unlabeled data without any inspection, thereby significantly reducing the cost of ML training. Unfortunately, as we show, this key feature can facilitate strong data poisoning attacks on SSL: a naive adversary, without any knowledge of training data or model architecture, can poison just $0.2\%$ of the entire training data to install a strong backdoor functionality in SSL models. Our attack remains effective against various SSL algorithms and benchmark datasets, and even circumvents state-of-the-art defenses against backdoor attacks. + +Note that, in contrast to numerous prior works [37, 48, 50, 44, 7, 4, 5, 29], in this work we considered a much weaker, hence more realistic, adversary. Due to our weak adversarial assumptions and simple attack methodology, all of the existing and future SSL applications can use our attack to measure and enhance their robustness against backdoor poisoning. + +Backdoor attacks can have severe consequences in practice, e.g., gaining unauthorized access to a system [7] or denying services to minorities [38]. Hence, a major implication of our study is that real-world SSL applications cannot rely on non-inspected unlabeled data and must preprocess/inspect unlabeled training data and/or design SSL algorithms that are robust to unlabeled data poisoning. + +# Acknowledgements + +This work was supported in-part by the NSF grants 2131910 and 1953786 and by DARPA under Agreement No. HR00112190125. Approved for public release; distribution is unlimited. This work was also supported in-part by Sony AI. + +# References + +[1] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In International conference on machine learning, pages 284-293. PMLR, 2018. 12 +[2] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching and augmentation anchoring. In International Conference on Learning Representations, 2019. 1, 5, 6 +[3] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in Neural Information Processing Systems, 32, 2019. 1, 3, 5, 6, 14 +[4] B. Biggio, B. Nelson, and P. Laskov. Poisoning attacks against support vector machines. In Proceedings of 29th International Conference on Machine Learning, 2012. 1, 9 +[5] Nicholas Carlini. Poisoning the unlabeled dataset of {Semi-Supervised} learning. In 30th USENIX Security Symposium (USENIX Security 21), pages 1577-1592, 2021. 5, 8, 9, 14, 15 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 7 +[7] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 3, 5, 9 +[8] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 14 +[9] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 3 +[10] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 2, 3, 7 +[11] Aron Culotta and Andrew McCallum. Reducing labeling effort for structured prediction tasks. In AAAI, volume 5, pages 746-751, 2005. 1 +[12] Jia Deng. A large-scale hierarchical image database. Proc. of IEEE Computer Vision and Pattern Recognition, 2009, 2009. 1 +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1 +[14] Emily Denton, Sam Gross, and Rob Fergus. Semi-supervised learning with context-conditional generative adversarial networks. arXiv preprint arXiv:1611.06430, 2016. 2 + +[15] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 2, 12 +[16] Le Feng, Sheng Li, Zhenxing Qian, and Xinpeng Zhang. Unlabeled backdoor poisoning in semi-supervised learning. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 3 +[17] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 113–125, 2019. 9, 16 +[18] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 1, 5 +[19] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 7:47230-47244, 2019. 3, 13 +[20] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 14 +[21] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2 +[22] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1, 2 +[23] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021. 3 +[24] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. In International Conference on Learning Representations, 2020. 9, 16 +[25] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems, 34:14900-14912, 2021. 9, 16 +[26] Yuncheng Li, Jianchao Yang, Yale Song, Liangliang Cao, Jiebo Luo, and Li-Jia Li. Learning from noisy labels with distillation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1910-1918, 2017. 1 +[27] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273-294. Springer, 2018. 16 +[28] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In 25th Annual Network And Distributed System Security Symposium (NDSS 2018). Internet Soc, 2018. 1 +[29] Luis Muñoz-González, Battista Biggio, Ambra Demontis, Andrea Paudice, Vasin Wongrassamee, Emil C Lupu, and + +Fabio Roli. Towards poisoning of deep learning algorithms with back-gradient optimization. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pages 27-38. ACM, 2017. 1, 2, 9 +[30] Kevin P Murphy. Machine learning: a probabilistic perspective. MIT press, 2012. 1 +[31] Nagarajan Natarajan, Inderjit S Dhillon, Pradeep K Ravikumar, and Ambuj Tewari. Learning with noisy labels. Advances in neural information processing systems, 26, 2013. 1 +[32] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 14 +[33] Tuan Anh Nguyen and Anh Tran. Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems, 33:3454-3464, 2020. 3 +[34] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11957-11965, 2020. 1, 5 +[35] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 2 +[36] Esha Sarkar, Hadjer Benkraouda, and Michail Maniatakos. Facehack: Triggering backdoored facial recognition systems using facial characteristics. arXiv preprint arXiv:2006.11623, 2020. 1, 3, 5 +[37] Virat Shejwalkar and Amir Houmansadr. Manipulating the byzantine: Optimizing model poisoning attacks and defenses for federated learning. In The Network and Distributed System Security Symposium (NDSS), 2021. 9 +[38] V. Shejwalkar, A. Houmansadr, P. Kairouz, and D. Ramage. Back to the drawing board: A critical evaluation of poisoning attacks on production federated learning. In 2022 2022 IEEE Symposium on Security and Privacy (SP) (SP), pages 1117-1134, Los Alamitos, CA, USA, may 2022. IEEE Computer Society. 1, 2, 4, 9 +[39] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019. 12 +[40] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems, 33:596-608, 2020. 1, 3, 4, 5, 6, 14 +[41] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 1 +[42] Hossein Souri, Micah Goldblum, Liam Fowl, Rama Chellappa, and Tom Goldstein. Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. arXiv preprint arXiv:2106.08970, 2021. 5 + +[43] Luke Taylor and Geoff Nitschke. Improving deep learning with generic data augmentation. In 2018 IEEE Symposium Series on Computational Intelligence (SSCI), pages 1542-1547. IEEE, 2018. 12 +[44] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 1, 5, 9, 13 +[45] Yidong Wang, Hao Chen, Yue Fan, Hao Wu, Bowen Zhang, Wenxin Hou, Yuhao Chen, and Jindong Wang. Torchssl: A pytorch-based toolbox for semi-supervised learning. https://github.com/TorchSSL/TorchSSL, 2021. [Online; accessed 03-July-2022]. 14 +[46] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems, 33:6256-6268, 2020. 1, 3, 5, 6 +[47] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 1 +[48] Zhicong Yan, Gaolei Li, Yuan Tlan, Jun Wu, Shenghong Li, Mingzhe Chen, and H Vincent Poor. Dehib: Deep hidden backdoor attack on semi-supervised learning via adversarial perturbation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 10585-10593, 2021. 2, 3, 5, 9, 12 +[49] Zhicong Yan, Jun Wu, Gaolei Li, Shenghong Li, and Mohsen Guizani. Deep neural backdoor in semi-supervised learning: threats and countermeasures. IEEE Transactions on Information Forensics and Security, 16:4827-4842, 2021. 3, 5 +[50] Yi Zeng, Minzhou Pan, Hoang Anh Just, Lingjuan Lyu, Meikang Qiu, and Ruoxi Jia. Narcissus: A practical clean-label backdoor attack with limited information. arXiv preprint arXiv:2204.05255, 2022. 1, 2, 5, 7, 9, 12, 13 +[51] Yi Zeng, Won Park, Z Morley Mao, and Ruoxi Jia. Rethinking the backdoor attacks' triggers: A frequency perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16473-16481, 2021. 3, 9 +[52] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 5, 6 +[53] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 3, 7 +[54] Haoti Zhong, Cong Liao, Anna Cinzia Squicciarini, Sencun Zhu, and David Miller. Backdoor embedding in convolutional neural network models via invisible perturbation. In Proceedings of the Tenth ACM Conference on Data and Application Security and Privacy, pages 97-108, 2020. 5 \ No newline at end of file diff --git a/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/images.zip b/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b3638ae6dc7b48cfaefb8fd6de16cedc78b6fd4d --- /dev/null +++ b/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6288c1db2bb5d6d31f9405b30d05f6b14ec1e56a90a43c3b9ba053f3e82fd808 +size 637164 diff --git a/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/layout.json b/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..21f2a404130bacf491527b0712276d1377812fe1 --- /dev/null +++ b/2023/The Perils of Learning From Unlabeled Data_ Backdoor Attacks on Semi-supervised Learning/layout.json @@ -0,0 +1,12797 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 103, + 445, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 103, + 445, + 140 + ], + "spans": [ + { + "bbox": [ + 149, + 103, + 445, + 140 + ], + "type": "text", + "content": "The Perils of Learning From Unlabeled Data: Backdoor Attacks on Semi-supervised Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 161, + 208, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 161, + 208, + 188 + ], + "spans": [ + { + "bbox": [ + 124, + 161, + 208, + 188 + ], + "type": "text", + "content": "Virat Shejwalkar \nUMass Amherst" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 192, + 231, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 192, + 231, + 202 + ], + "spans": [ + { + "bbox": [ + 99, + 192, + 231, + 202 + ], + "type": "text", + "content": "vshejwalkar@cs.umass.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 281, + 161, + 350, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 161, + 350, + 189 + ], + "spans": [ + { + "bbox": [ + 281, + 161, + 350, + 189 + ], + "type": "text", + "content": "Lingjuan Lyu \nSony AI" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 261, + 192, + 370, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 192, + 370, + 203 + ], + "spans": [ + { + "bbox": [ + 261, + 192, + 370, + 203 + ], + "type": "text", + "content": "lingjuan.lv@sony.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 400, + 161, + 493, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 161, + 493, + 188 + ], + "spans": [ + { + "bbox": [ + 400, + 161, + 493, + 188 + ], + "type": "text", + "content": "Amir Houmansadr \nUMass Amherst" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 400, + 192, + 493, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 192, + 493, + 202 + ], + "spans": [ + { + "bbox": [ + 400, + 192, + 493, + 202 + ], + "type": "text", + "content": "amir@cs.umass.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 365 + ], + "type": "text", + "content": "Semi-supervised learning (SSL) is gaining popularity as it reduces cost of machine learning (ML) by training high performance models using unlabeled data. In this paper, we reveal that the key feature of SSL, i.e., learning from (non-inspected) unlabeled data, exposes SSL to strong poisoning attacks that can significantly damage its security. Poisoning is a long-standing problem in conventional supervised ML, but we argue that, as SSL relies on non-inspected unlabeled data, poisoning poses a more significant threat to SSL." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 365, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 521 + ], + "type": "text", + "content": "We demonstrate this by designing a backdoor poisoning attack on SSL that can be conducted by a weak adversary with no knowledge of the target SSL pipeline. This is unlike prior poisoning attacks on supervised ML that assume strong adversaries with impractical capabilities. We show that by poisoning only " + }, + { + "bbox": [ + 46, + 365, + 287, + 521 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 521 + ], + "type": "text", + "content": " of the unlabeled training data, our (weak) adversary can successfully cause misclassification on more than " + }, + { + "bbox": [ + 46, + 365, + 287, + 521 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 365, + 287, + 521 + ], + "type": "text", + "content": " of test inputs (when they contain the backdoor trigger). Our attack remains effective across different benchmark datasets and SSL algorithms, and even circumvents state-of-the-art defenses against backdoor attacks. Our work raises significant concerns about the security of SSL in real-world security critical applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 128, + 560 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 640 + ], + "type": "text", + "content": "Machine learning (ML) models perform better with increased amounts of training data [13, 12]. However, conventional supervised ML requires labeling large amounts of training data, an expensive [11] and error prone [31, 26] process that makes it prohibitively expensive, especially with today's exploding training data sizes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Semi-supervised learning (SSL) addresses this major challenge by significantly reducing the need for labeled training data: SSL uses a combination of a small, high-quality labeled data (expensive data) with a large, low-quality unlabeled data (cheap data) to train models. For instance, the FixMatch [40] SSL algorithm combines only" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 232, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 316 + ], + "type": "text", + "content": "40 labeled with " + }, + { + "bbox": [ + 304, + 232, + 545, + 316 + ], + "type": "inline_equation", + "content": "50k" + }, + { + "bbox": [ + 304, + 232, + 545, + 316 + ], + "type": "text", + "content": " unlabeled data to achieve a " + }, + { + "bbox": [ + 304, + 232, + 545, + 316 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 232, + 545, + 316 + ], + "type": "text", + "content": " accuracy on CIFAR10. Training SSL involves two loss functions: a supervised loss (e.g., cross-entropy [30] over true labels) on labeled training data and an unsupervised loss (e.g., cross-entropy over pseudo-labels [22]) on unlabeled training data. Different SSL algorithms primarily differ in terms of how they compute their unsupervised losses." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 316, + 546, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 546, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 546, + 400 + ], + "type": "text", + "content": "SSL has gained popularity in both academia [52, 46, 47] and industry [40, 41, 3, 2], as recent SSL algorithms offer state-of-the-art performances comparable or even superior to supervised techniques—but with no need of large well-inspected labeled data. For instance, due to their effective use of unlabeled data, with less than " + }, + { + "bbox": [ + 304, + 316, + 546, + 400 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 316, + 546, + 400 + ], + "type": "text", + "content": " of training data labeled, FixMatch [40] outperforms supervised ML." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 401, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 628 + ], + "type": "text", + "content": "Unlabeled data enables poisoning by weak adversaries: Multiple researches have demonstrated the data poisoning threat to supervised ML [18, 28, 34, 36, 50, 44, 38]. However, as the training data in supervised ML undergo an extensive and careful inspection, these attacks assume strong adversaries with the knowledge of model parameters [28], training data [44, 4, 29], its distribution [50], or the ML algorithm. Such strong adversaries are important to evaluate worse-case security of a system, but are irrelevant in practice [38]. On the other hand, the key feature of SSL that makes it attractive to real-world applications is its ability to leverage large amounts of—raw, non-inspected—unlabeled data, e.g., the data scraped off the Internet. We argue that the use of non-inspected data by SSL presents a unique threat to its security, as it allows even the most naive adversaries (with no knowledge of training algorithm, data, etc.) to poison SSL models by simply fabricating malicious unlabeled data. Unfortunately, this ostensible threat is largely unexplored in the SSL literature." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "content": "To address this gap, in this paper, we take the first step towards understanding this threat by studying the possibility of backdoor attacks against SSL in real-world settings. Backdoor attacks aim to install a backdoor function in the target model, such that the backdoored target model will misclassify any test input to the adversary chosen target class when patched with a specific backdoor trigger, but" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4730" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 258, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 258, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 258, + 83 + ], + "type": "text", + "content": "will correctly classify test inputs without the trigger." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 204 + ], + "type": "text", + "content": "Existing backdoor attacks fail on SSL: There exist numerous backdoor attacks in the literature, however, except one attack—DeHiB [48], all of the prior attacks consider supervised ML. Our preliminary evaluations show that all of the existing state-of-the-art (SOTA) attacks, including DeHiB, completely fail against SSL under our realistic threat model (Section 3.1). Hence, to learn from these failures, we first systematically evaluate five SOTA backdoor attacks from three categories against five SOTA SSL algorithms, under our practical, unlabeled data poisoning threat model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 205, + 286, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 286, + 348 + ], + "type": "text", + "content": "Our systematic evaluation leads to the following three major lessons that not only guide our attack design, but can be useful building blocks for (future) backdoor attacks against SSL: (1) Backdoor attacks on SSL should be clean-label style attacks, i.e., poisoning data should be selected from the distribution of target class " + }, + { + "bbox": [ + 46, + 205, + 286, + 348 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 46, + 205, + 286, + 348 + ], + "type": "text", + "content": "; (2) Backdoor triggers should be of the same size as the poisoning sample, to circumvent strong augmentations, e.g., cutout [15], that all modern SSL algorithms use; (3) Backdoor triggers should be resistant to noise and with repetitive patterns1 to withstand large amounts of random noises due to strong augmentations, e.g., RandAugment [10], in SSL." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": "Our SSL-tailored backdoor method: The high-level intuition behind our backdoor attack is as follows. All modern SSL algorithms learn via a self-feedback mechanism, called pseudo-labeling, i.e., if current state of target model " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": " has high confidence prediction " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": " for an unlabeled sample " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": ", then they use " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}, \\tilde{y})" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": " as a labeled sample for further training. We exploit pseudo-labeling and design a clean-label attack that poisons unlabeled data only from the distribution of " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": ". Our attack patiently waits for " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": " to correctly label a poisoning sample " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "(\\mathbf{x} + T)" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": " is our pre-determined backdoor trigger. As " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": " trains further on " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "((\\mathbf{x} + T), y^{t})" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": ", our attack forces " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": " to associate features of our simple trigger " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": ", instead of the complex features of " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 46, + 348, + 286, + 515 + ], + "type": "text", + "content": ", thereby installing the backdoor in the target model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 516, + 286, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 286, + 612 + ], + "type": "text", + "content": "Note that, we consider the most challenging setting for designing attacks with the least capable and knowledgeable data poisoning adversary. Generally, trigger generation for data poisoning backdoor attacks is formalized as a bi-level optimization problem [29], however such attacks are well-known to be very expensive, and yet ineffective [29, 38]. Instead, our lessons lead us to a simple yet effective static, repetitive grid pattern backdoor trigger (Figure 2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 613, + 286, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 286, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 286, + 684 + ], + "type": "text", + "content": "**Evaluations:** We demonstrate the strength of our attack via an extensive evaluation against five SOTA SSL and one supervised ML algorithm, using four benchmark image classification tasks commonly used in the SSL literature. We note that our attack significantly outperforms prior attacks from both SSL and supervised ML literature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "text", + "content": "We measure success of our attacks using ASR metric: ASR measures the " + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "text", + "content": " of test inputs from non-target classes that the backdoored model classifies to the target class when patched with backdoor trigger. For the most combinations of algorithms and datasets, our attacks achieve high attack success rates (ASRs) " + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "(>80\\%)" + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "text", + "content": ", while poisoning just " + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "text", + "content": " of entire training data. For instance, our attacks have more than " + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "text", + "content": " ASR against CIFAR100 and more than " + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "text", + "content": " ASR against CIFAR10. For SVHN and STL10, our attack has more than " + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "text", + "content": " ASR with two exceptions each. While, under our practical threat model, DeHib attack achieves " + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "text", + "content": " ASR even with " + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "20\\times" + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "text", + "content": " more poisoning data. Through a systematic experiment design in Section 5.1.4, we show that our intuition aligns with the dynamics of our attacks and justify their strength. Our attack is highly stealthy, as (1) according to " + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 306, + 72, + 545, + 346 + ], + "type": "text", + "content": "-norm metric commonly used [50] for stealth measurement, it minimally perturbs the poisoning data and (2) it produces backdoored models which have high accuracy (close to non-backdoored models) on non-backdoored test inputs. We perform comprehensive ablation study (Section 5.2) to demonstrate the high efficacy of our attacks as we vary (1) size of labeled data, (2) backdoor target class, and (3) size of poisoning data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 348, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 443 + ], + "type": "text", + "content": "Finally, we show that our attack remains highly effective even when SSL is paired individually with five SOTA defenses against backdoor attacks that are agnostic to learning algorithms. To defend against such unlabeled data poisoning, we argue for SSL to depart from its philosophy of not inspecting unlabeled training data, and instead, pre-process/inspect the unlabeled data and/or design SSL algorithms that are robust-by-design to such poisoning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 455, + 486, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 486, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 486, + 467 + ], + "type": "text", + "content": "2. Preliminaries and Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 475, + 480, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 480, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 480, + 487 + ], + "type": "text", + "content": "2.1. Semi-supervised Learning (SSL)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "content": "Supervised ML requires completely labeled data, " + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "inline_equation", + "content": "D^{l}" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "content": ", which can be prohibitively expensive due to expensive manual labelling involved. SSL reduces this cost by using very few labeled " + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "inline_equation", + "content": "D^{l}" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "content": " and plenty of unlabeled data " + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "inline_equation", + "content": "D^{u}" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "content": ". SSL uses a convex combination of a supervised loss " + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{l}" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "inline_equation", + "content": "D^{l}" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "content": " and an unsupervised loss " + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{u}" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "inline_equation", + "content": "D^{u}" + }, + { + "bbox": [ + 304, + 493, + 545, + 590 + ], + "type": "text", + "content": ". Modern state-of-the-art SSL algorithms rely on two key building blocks: pseudolabeling [22] and consistency regularization [14, 35, 21]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "text", + "content": "Pseudo-labeling uses the current model, " + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "text", + "content": ", to obtain artificial pseudo-labels for " + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "inline_equation", + "content": "D^{u}" + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "text", + "content": " and only retains the data on which " + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "text", + "content": " has high confidence. Assume " + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "inline_equation", + "content": "q_{b} = f_{\\theta}(y|u_{b})" + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "text", + "content": " are the predictions of " + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "text", + "content": " on the batch " + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "inline_equation", + "content": "u_{b}" + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "text", + "content": " of unlabeled data. Then pseudo-labeling loss can be formalized as: " + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\frac{1}{|u_b|}\\sum_{b = 1}^{|u_b|}\\mathbb{1}(\\max (q_b)\\geq \\tau)H(\\hat{q}_b,q_b); \\hat{q}_b = \\operatorname {argmax}(q_b), H(. )" + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "text", + "content": " is cross-entropy and " + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 590, + 545, + 676 + ], + "type": "text", + "content": " is confidence threshold." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Consistency regularization trains " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " to output similar predictions for perturbed versions of the same input. It uses stochastic augmentations " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "a(\\mathbf{x}_u)" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " to perturb an unlabeled" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 713 + ], + "type": "text", + "content": "1Repetitive pixel patterns are the patterns on which if we zoom in on any part, we get similar pattern. For examples, check Figures 1 and 10." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "4731" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "sample " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_u" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " and forces " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " to have similar outputs on multiple " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "a(\\mathbf{x}_u)" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "'s using the following loss: " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "\\sum_{b=1}^{|u_b|} \\| f_{\\theta}(y | a(u_b)) - f_{\\theta}(y | a(u_b))\\|_2^2" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "a(.)" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " produces different output every time it is applied to a batch " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "u_b" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " of unlabeled data. Below we describe the five SOTA SSL algorithms we consider in this work." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 287, + 639 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 46, + 146, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 287, + 266 + ], + "type": "text", + "content": "(1) MixMatch [3] combines various prior semi-supervised learning techniques. For an unlabeled sample, MixMatch generates " + }, + { + "bbox": [ + 46, + 146, + 287, + 266 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 146, + 287, + 266 + ], + "type": "text", + "content": " weakly augmented versions of the unlabeled sample, computes outputs of the current model " + }, + { + "bbox": [ + 46, + 146, + 287, + 266 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 146, + 287, + 266 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 46, + 146, + 287, + 266 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 146, + 287, + 266 + ], + "type": "text", + "content": " versions, averages them, and sharpens the average prediction by raising all its probabilities by a power of 1/temperature and re-normalizing; it uses the sharpened prediction as the label of the unlabeled sample. Finally, it uses mixup regularization [53] on the combination of labeled and unlabeled data and trains the model using cross-entropy loss." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 266, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 350 + ], + "type": "text", + "content": "(2) Unsupervised data augmentation (UDA) [46] shows significant improvements in semi-supervised performances by just replacing the simple weak augmentations of MixMatch with a strong augmentation called Randaugment [10]. In an iteration, Randaugment randomly selects a few augmentations from a large set of augmentations and applies them to images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 482 + ], + "type": "text", + "content": "(3) ReMixMatch [3] builds on MixMatch by making multiple modifications, including 1) it replaces the simple weak augmentation in MixMatch with Autoagument [9], 2) it uses augmentation anchoring to improve consistency regularization, i.e., it uses the prediction on a weakly augmented version of unlabeled sample as the target prediction for a strongly augmented version of the unlabeled sample, and 3) it uses distribution alignment, i.e., it normalizes the new model predictions on unlabeled data using the running average of model predictions on unlabeled data. This significantly boosts the performance of resulting model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 482, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 553 + ], + "type": "text", + "content": "(4) FixMatch [40] simplifies the complex ReMixMatch algorithm by proposing to use a combination of Pseudolabeling and consistency regularization based on augmentation anchoring (discussed above). FixMatch significantly improves semi-supervised algorithms, especially in the low labeled data regimes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 554, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 639 + ], + "type": "text", + "content": "(5) FlexMatch [52] proposes curriculum pseudo labelling (CPL) approach to leverage unlabeled data according to model's learning status. The main idea behind CPL is to flexibly adjust the thresholds used for pseudo-labeling for different classes at each training iteration in order to select more information unlabeled data and their pseudo-labels. CPL can be combined with other algorithms, e.g., UDA." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 647, + 155, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 155, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 155, + 658 + ], + "type": "text", + "content": "2.2. Backdoor Attacks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "A backdoor adversary aims to implant a backdoor functionality into a target model. That is, given an input " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},y^{*})" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " with true label " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "y^{*}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ", the backdoored target model " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{b}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " should output an adversary-desired backdoor target label " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " for" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "the input patched with a pre-specified backdoor trigger " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ", but it should output the correct label for the benign input, i.e., " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{b}(\\mathbf{x} + T) \\mapsto y^{t}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{b}(\\mathbf{x}) \\mapsto y^{*}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ". There are two major types of backdoor attacks: (1) dirty-label attacks [19, 7, 36, 51, 33, 23] that poison both the features " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " and labels " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "y^{*}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " of benign, labeled data to obtain poisoning data " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "D^{p}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ". (2) On the other hand, clean-label attacks obtain " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "D^{p}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " by poisoning only the features " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " of benign data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 169, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 169, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 545, + 276 + ], + "type": "text", + "content": "Backdoor attacks on SSL, unfortunately, have not received significant attention from the scientific community. [48, 49, 16] study backdoor attacks against SSL. However, all of these attacks assume access to " + }, + { + "bbox": [ + 304, + 169, + 545, + 276 + ], + "type": "inline_equation", + "content": "D^{l}" + }, + { + "bbox": [ + 304, + 169, + 545, + 276 + ], + "type": "text", + "content": ", the labeled training data of SSL. This assumption renders their applicability questionable in realistic settings. For clarity of presentation, we discuss the details of these attacks in Appendix A.1, where we demonstrate (Table 2) and justify why these attacks fail to backdoor SSL." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 290, + 503, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 290, + 503, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 503, + 304 + ], + "type": "text", + "content": "3. Our Backdoor Attack Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 311, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 418 + ], + "type": "text", + "content": "We first discuss threat model of our attack, followed by our intuition behind backdoor attacks on SSL. Next, we note that, effectively backdoorsing semi-supervised learning (SSL) does not need a new backdoor attack, but requires careful adoption of existing backdoor attacks. We discuss three major lessons learned from systematically evaluating SOTA attacks under our threat model (Section 3.1). Finally, we detail our SOTA backdoor attack based on our intuition and the lessons learned." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 429, + 394, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 429, + 394, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 429, + 394, + 441 + ], + "type": "text", + "content": "3.1. Threat Model" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 484 + ], + "type": "text", + "content": "We consider a victim model trainer who collects data from multiple, potentially untrusted sources to train a ML model using SSL for a classification task with " + }, + { + "bbox": [ + 304, + 449, + 545, + 484 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 449, + 545, + 484 + ], + "type": "text", + "content": " classes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "text", + "content": "Adversary's goal: A backdoor adversary aims to install a backdoor function in the victim's target model. We denote the models without (benign) and with backdoor by " + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{b}" + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "text", + "content": ", respectively. Adversary's backdoor goal is to force " + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{b}" + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "text", + "content": " to incorrectly classify all the test inputs from non-target classes to the adversary-desired target class " + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "text", + "content": ", when they are patched with a pre-specified backdoor trigger, " + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "text", + "content": ". Adversary's stealth goal aims that " + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{b}" + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "text", + "content": " should correctly classify all the benign test inputs, i.e., any input without " + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 485, + 545, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Adversary's knowledge: As discussed in Section 1, we consider the most naive, real-world adversary with minimum knowledge of the SSL pipeline. More specifically, we assume that the adversary has no knowledge of the labeled or unlabeled training data and does not possess any data from true distribution; they just know the classification task, i.e., CIFAR10 or SVHN. Our adversary knows the details of the target SSL algorithm, but does not know model architecture, e.g., ResNet or VGG, i.e., our attacks are model architecture agnostic." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4732" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 133 + ], + "type": "text", + "content": "Adversary's capabilities: Due to our emphasis on practicality of our threat model, we consider a data poisoning adversary [38]. Specifically, our adversary can poison only the unlabeled data of SSL pipeline, and cannot poison or even access the model, code or the labeled data of SSL." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 139, + 246, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 139, + 246, + 150 + ], + "spans": [ + { + "bbox": [ + 47, + 139, + 246, + 150 + ], + "type": "text", + "content": "3.2. Intuition behind our backdoor attack" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 158, + 287, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 287, + 194 + ], + "type": "text", + "content": "For brevity, we discuss our intuition for FixMatch [40], but it applies to any SSL algorithms that use pseudolabeling and consistency regularization (Section 2.1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "content": "As explained in Section 2.1, FixMatch trains parameters " + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "content": " to learn a function " + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "content": " from the labeled data " + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "inline_equation", + "content": "D^{l}" + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "content": " and assigns a pseudo-label " + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "content": " to an unlabeled sample " + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in D^{u}" + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "content": ". Then it further trains " + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}, \\hat{y})" + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "content": " to improve " + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "content": ". As the training progresses, the confidence of " + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "content": " on the correct label of " + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "content": " increases which leads to better pseudo-labeling of " + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "inline_equation", + "content": "D^{u}" + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "content": " and further improvements in the accuracy of " + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 194, + 288, + 289 + ], + "type": "text", + "content": ". In other words, FixMatch learns via a self-feedback mechanism." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": "Recall that, our realistic data poisoning adversary cannot alter either the SSL training pipeline or the well-inspected labeled training data " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "D^{l}" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": ". Now, the first part of our intuition is that our attacks should be of clean-label type, i.e., we select unlabeled data " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "X^{y^t}" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": " from the target class, " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": ", and patch it with backdoor trigger " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "X^p = X^{y^t} + T" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": "; next section demonstrates the necessity of this condition. The second part of our intuition is that in initial part of training, FixMatch will assign the desired pseudo-labels " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": " due to the original features " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "X^{y^t}" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": ". However, due to the presence of backdoor trigger, " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": ", on all " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "X^p s" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": ", the model will be forced to eventually learn a much simpler task of associating " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 46, + 290, + 288, + 446 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": "To further understand this, consider three benign samples " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i\\in \\{1,2,3\\}}" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " with target class " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " as their true label. The adversary adds a trigger " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " to these samples to obtain " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{i\\in \\{1,2,3\\}} + T\\}" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " and inserts " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "D^u" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": ". Note that, initially during training, FixMatch learns the association " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "f_{\\theta}:\\mathcal{X}\\mapsto \\mathcal{V}" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " between feature and label spaces only through " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "D^l" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": ". And as our threat model assumes that " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "D^l" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " is benign (not poisoned), initially FixMatch focuses only on the benign features of " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": ", i.e., on " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i\\in \\{1,2,3\\}}" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " and assigns the correct label " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " to all " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " samples. This in turn forces FixMatch to learn from " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_{i\\in \\{1,2,3\\}} + T,y^t)" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " is present in all " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " samples, FixMatch incorrectly learns the simpler task of associating the static trigger " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": ", instead of the difficult task of associating the complex and dynamic benign features of " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i\\in \\{1,2,3\\}}" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 46, + 447, + 288, + 628 + ], + "type": "text", + "content": "; we very our intuition in Section 5.1.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 635, + 287, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 659 + ], + "type": "text", + "content": "3.3. Lessons from systematic evaluation of existing backdoor attacks against SSL" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "In Table 1, we categorize existing SOTA attacks on supervised ML and SSL in three types: dirty label, clean label small trigger and clean label adversarial samples. We evaluate representative attacks from each category and pro" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 539, + 159 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 539, + 159 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 539, + 159 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 539, + 159 + ], + "type": "image", + "image_path": "26b29a50c1ca59ee37af82cd221b8acd0fab762a56ab82a6eb5131b4be9b42a8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 164, + 529, + 176 + ], + "lines": [ + { + "bbox": [ + 321, + 164, + 529, + 176 + ], + "spans": [ + { + "bbox": [ + 321, + 164, + 529, + 176 + ], + "type": "text", + "content": "Figure 1: Our backdoor trigger and a poisoning sample." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 352, + 178, + 499, + 260 + ], + "blocks": [ + { + "bbox": [ + 352, + 178, + 499, + 260 + ], + "lines": [ + { + "bbox": [ + 352, + 178, + 499, + 260 + ], + "spans": [ + { + "bbox": [ + 352, + 178, + 499, + 260 + ], + "type": "image", + "image_path": "ef99f0d41cd3df2ea4ff3f12cb2e622154211804d1aaa4b4312fce5133f4c690.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 261, + 545, + 305 + ], + "lines": [ + { + "bbox": [ + 305, + 261, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 261, + 545, + 305 + ], + "type": "text", + "content": "Figure 2: Our backdoor trigger has three parameters: pixel intensity " + }, + { + "bbox": [ + 305, + 261, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 261, + 545, + 305 + ], + "type": "text", + "content": ", pixel gap " + }, + { + "bbox": [ + 305, + 261, + 545, + 305 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 261, + 545, + 305 + ], + "type": "text", + "content": ", and pixel width " + }, + { + "bbox": [ + 305, + 261, + 545, + 305 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 305, + 261, + 545, + 305 + ], + "type": "text", + "content": ". For presentation clarity, we use high pixel intensity here, but in experiments we use low intensities to ensure attack stealth." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 318, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 366 + ], + "type": "text", + "content": "vide justification for their failure against SSL, and the corresponding lesson that will guide future attack designs on SSL. Due to space limit, we present the key lessons here and defer detailed discussions to Appendix A." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 366, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 437 + ], + "type": "text", + "content": "Lesson-1: Backdoor attacks against semi-supervised learning should be clean-label type, i.e., the poisoning samples should be from the backdoor target class. Without this condition, model will be forced to learn to associate " + }, + { + "bbox": [ + 304, + 366, + 545, + 437 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 366, + 545, + 437 + ], + "type": "text", + "content": " with different classes (i.e., original classes of poisoning data " + }, + { + "bbox": [ + 304, + 366, + 545, + 437 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 304, + 366, + 545, + 437 + ], + "type": "text", + "content": "), and effectively, model will simply ignore " + }, + { + "bbox": [ + 304, + 366, + 545, + 437 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 366, + 545, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "content": "Lesson-2: The trigger should have same size as the entire sample (images in our case), to ensure that all the augmented instances of a poisoning sample contain majority of the trigger. This is necessary due to augmentations in SSL that occlude part of an image, e.g., cutout and cutmix, which all the modern SOTA SSL algorithms use." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 511, + 545, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 511, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 511, + 545, + 547 + ], + "type": "text", + "content": "Lesson-3: The trigger should be noise-resistant and with a repetitive pattern. This is necessary, again, to circumvent the occluding augmentations described above." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 548, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 596 + ], + "type": "text", + "content": "We believe that the above lessons give the minimum constraints to design backdoor attacks on SSL. But, they need not be exhaustive and may need modifications, e.g., based on different threat models and SSL algorithms." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 604, + 503, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 503, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 503, + 616 + ], + "type": "text", + "content": "3.4. Our State-of-the-art backdoor attack" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 624, + 545, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 684 + ], + "type": "text", + "content": "Based on our intuition and the three lessons detailed above, we develop a clean-label style backdoor attack using a specific static trigger pattern. Figure 1 depicts our static backdoor trigger and a corresponding poisoning image; we present more images for CIFAR, SVHN, and STL10" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "2Repetitive pixel patterns are the patterns on which if we zoom in on any part, we get similar pattern. For examples, check Figures 1 and 10." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4733" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 110, + 519, + 213 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "type": "text", + "content": "Table 1: The left column shows types of backdoor attacks based on specific characteristics, middle column lists existing attacks of each type. Right-most column presents lessons we learn from evaluating one/two representative attacks (in bold) of each type." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 110, + 519, + 213 + ], + "lines": [ + { + "bbox": [ + 66, + 110, + 519, + 213 + ], + "spans": [ + { + "bbox": [ + 66, + 110, + 519, + 213 + ], + "type": "table", + "html": "
Attack characteristic/typeExisting attacks of given typeLesson from evaluations
Dirty labelDeHiB [48], DL-Badnets [18], DL-Blend [7], Facehack [36]Attack should be a clean-label attack, i.e., poisoning samples should be from backdoor target class.
Clean-label small triggerCL-Badnets [50], CL-Blend [7]Trigger should span the entire sample/image to avoid cropping/covering by strong augmentations.
Clean-label adversarial samplesNarcissus [50], Label consistent [44], non-repeating trigger patterns, HTBA [34], SAA [42], Embedding [54]Trigger should be noise-resistant and its pattern should be repetitive so that even a part of trigger can install a backdoor.
", + "image_path": "a5ed83480c69aa2c0f019e8a2bf11e34fe65ab3488122419f4c7e61f68c51bb8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 136, + 255, + 449, + 326 + ], + "blocks": [ + { + "bbox": [ + 46, + 215, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 215, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 547, + 251 + ], + "type": "text", + "content": "Table 2: Impacts of existing backdoor attacks (Section 2.2) on various semi-supervised algorithms for CIFAR10 data. We poison " + }, + { + "bbox": [ + 46, + 215, + 547, + 251 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 46, + 215, + 547, + 251 + ], + "type": "text", + "content": " (100 samples) of all the training data. DeHib* is the original attack with the knowledge of labeled training data " + }, + { + "bbox": [ + 46, + 215, + 547, + 251 + ], + "type": "inline_equation", + "content": "D^l" + }, + { + "bbox": [ + 46, + 215, + 547, + 251 + ], + "type": "text", + "content": " while DeHib is the attack without the knowledge of " + }, + { + "bbox": [ + 46, + 215, + 547, + 251 + ], + "type": "inline_equation", + "content": "D^l" + }, + { + "bbox": [ + 46, + 215, + 547, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 136, + 255, + 449, + 326 + ], + "lines": [ + { + "bbox": [ + 136, + 255, + 449, + 326 + ], + "spans": [ + { + "bbox": [ + 136, + 255, + 449, + 326 + ], + "type": "table", + "html": "
AlgorithmDeHiB* ASR (%)DeHiB ASR (%)CL-Badnets ASR (%)LC ASR (%)Narcissus ASR (%)
Mixmatch [3]22.01.09.11.12.2
Remixmatch [2]10.90.90.00.00.0
UDA [46]21.21.25.10.00.0
Fixmatch [40]35.80.910.20.11.3
Flexmatch [52]16.91.29.10.11.1
", + "image_path": "9f02b90f6d6e381435a4d27f1c35ae7eaaf81aa579e1b9b89372cf83f46e33b5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "text", + "content": "datasets in Figures 14, 15, and 16 in Appendix C. Our backdoor trigger pattern has three parameters: intensity " + }, + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "text", + "content": ", gap " + }, + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "text", + "content": ", and width " + }, + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "text", + "content": " is the intensity of the bright pixels in the trigger and intensity of the rest of the pixel is 0; " + }, + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "text", + "content": " is the distance between two adjacent set of bright pixels and " + }, + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 332, + 288, + 428 + ], + "type": "text", + "content": " is the width of each set of bright pixels. Note that the size of our trigger is the same as that of the sample and has a fairly repetitive pattern, hence it satisfies both Lessons-2 and 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 429, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 513 + ], + "type": "text", + "content": "To summarize our attack: we select a set of samples from the target class (to satisfy Lesson-1, poison them by adding the trigger to them, and inject these poisoned samples into the unlabeled data. As we will show in Section 5.1.1 (Table 3), with poisoning just " + }, + { + "bbox": [ + 46, + 429, + 287, + 513 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 46, + 429, + 287, + 513 + ], + "type": "text", + "content": " of the entire training data, this simple backdoor method injects backdoors in SSL models with close to " + }, + { + "bbox": [ + 46, + 429, + 287, + 513 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 429, + 287, + 513 + ], + "type": "text", + "content": " accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 513, + 288, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 288, + 621 + ], + "type": "text", + "content": "Finally, it is worth mentioning that, there are many possible triggers that follow aforementioned lessons, but choice of our specific trigger is based on various triggers patterns we investigated in our initial explorations (Figure 10). Furthermore, the choice of our simple yet effective backdoor attack method is a result of an extensive experimentation with various attack methods (and not just trigger patterns). In Appendix C.1, we discuss alternate attack methods we explored but found them unsuccessful at backdoorsing SSL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 633, + 163, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 163, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 163, + 647 + ], + "type": "text", + "content": "4. Experimental setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "We evaluate our backdoor attacks using four datasets (CIFAR10, SVHN, CIFAR100 and STL10) commonly used to benchmark semi-supervised algorithms. Due to space limits, we defer the details of datasets, model architectures and hyperparameters to Appendix B." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 332, + 427, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 332, + 427, + 344 + ], + "spans": [ + { + "bbox": [ + 305, + 332, + 427, + 344 + ], + "type": "text", + "content": "4.1. Performance metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 350, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 409 + ], + "type": "text", + "content": "Clean accuracy (CA) [18] measures the accuracy of a model on clean test data without any backdoor trigger " + }, + { + "bbox": [ + 304, + 350, + 545, + 409 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 350, + 545, + 409 + ], + "type": "text", + "content": ". Backdoors models should have high CA to ensure that the backdoor attack does not impact their benign functionality to ensure the attack stealth." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "type": "text", + "content": "Backdoor attack success rate (ASR) [18] measures the accuracy of a model on the backdoored test data from the nontarget classes patched with " + }, + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 411, + 545, + 458 + ], + "type": "text", + "content": ". For a successful backdoor attack, backdoored model should have high ASR." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 459, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 506 + ], + "type": "text", + "content": "Target class accuracy (TA) [50] measures the accuracy of the clean test data from the target backdoor class which does not contain any " + }, + { + "bbox": [ + 304, + 459, + 545, + 506 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 459, + 545, + 506 + ], + "type": "text", + "content": ". For a backdoored model, TA should be high to ensure the attack stealth." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 514, + 413, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 514, + 413, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 514, + 413, + 526 + ], + "type": "text", + "content": "5. Empirical Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 528, + 504, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 528, + 504, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 528, + 504, + 540 + ], + "type": "text", + "content": "5.1. Our attacks effectively backdoor SSL" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 547, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 547, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 547, + 545, + 643 + ], + "type": "text", + "content": "In this section, we demonstrate the superiority of our backdoor attacks over various baseline attacks in terms of three metrics from Section 4.1. Note that, we poison at most " + }, + { + "bbox": [ + 304, + 547, + 545, + 643 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 304, + 547, + 545, + 643 + ], + "type": "text", + "content": " of the entire unlabeled data, which is significantly lower than what prior attacks use, e.g., " + }, + { + "bbox": [ + 304, + 547, + 545, + 643 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 547, + 545, + 643 + ], + "type": "text", + "content": " in DeHiB [48, 49]. Backdoor injection at such low poisoning percentages is extremely challenging as we aim to backdoor the entire test data and not just a single sample as in [5]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 658, + 471, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 658, + 471, + 670 + ], + "spans": [ + { + "bbox": [ + 305, + 658, + 471, + 670 + ], + "type": "text", + "content": "5.1.1 High attack success rate (ASR)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "text", + "content": "In Table 3, \"p%\" column shows poisoning percentage and ASR columns show the results. Our backdoor attacks outperform all the baseline backdoor attacks by very large" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4734" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 107, + 542, + 323 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 106 + ], + "type": "text", + "content": "Table 3: Impacts of backdoor attacks on various semi-supervised (SSL) algorithms (Section 2.1) under the unlabeled data poisoning threat model (Section 3.1). For all datasets, our attack (Section 3.4) significantly outperforms the baseline backdoor attack (DeHiB) against SSL and various clean-label attacks against supervised learning (Section 2.2). Best results are in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 107, + 542, + 323 + ], + "lines": [ + { + "bbox": [ + 51, + 107, + 542, + 323 + ], + "spans": [ + { + "bbox": [ + 51, + 107, + 542, + 323 + ], + "type": "table", + "html": "
CIFAR10AlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]92.20.093.50.292.115.394.291.10.094.991.11.492.13092.296.894.6
Remixmatch [2]91.30.094.90.291.01.195.091.30.095.990.82.194.83090.684.394.5
UDA [46]89.50.097.40.188.18.296.989.11.098.689.11.197.22089.681.596.7
Fixmatch [40]91.10.097.50.291.910.197.891.20.098.090.91.195.82093.588.197.6
Flexmatch [52]94.30.097.10.293.96.497.094.10.098.594.22.397.02093.887.996.9
SVHNAlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]94.40.095.40.294.55.493.894.50.096.194.43.295.03093.283.795.8
Remixmatch [2]87.60.095.50.288.01.295.487.10.095.988.11.795.93087.651.195.4
UDA [46]95.00.096.30.294.91.196.094.20.096.094.81.196.62094.995.595.8
Fixmatch [40]94.50.096.30.294.93.197.194.20.097.094.83.296.42094.597.193.9
Flexmatch [52]85.40.096.30.288.91.296.986.10.096.786.82.296.42083.950.196.6
STL10AlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]86.70.086.30.286.39.286.787.11.187.086.11.186.14086.486.287.9
Remixmatch [2]91.70.090.60.291.24.190.691.90.991.191.31.191.04091.282.291.4
UDA [46]88.10.077.50.288.15.577.189.00.177.988.51.777.43088.657.180.4
Fixmatch [40]92.10.086.10.292.213.186.692.10.086.992.02.286.23091.892.487.3
Flexmatch [52]88.10.088.80.288.16.588.188.40.988.087.81.787.93087.849.885.8
CIFAR100AlgorithmNo attackp%CL-BadnetsNarcissusDeHiBOur attack
CAASRTACAASRTACAASRTACAASRTAαCAASRTA
Mixmatch [3]71.60.067.20.271.930.167.572.01.568.372.31.168.13071.692.869.0
Remixmatch [2]73.30.059.10.273.318.959.373.21.160.273.20.559.93073.197.158.2
Fixmatch [40]71.30.049.30.270.622.049.871.41.150.171.42.349.81071.191.848.9
", + "image_path": "b67be20d300ab0d3a80b7dcd9131783c9be416b037fa72ca1ff37fc66625bfd2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 329, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 413 + ], + "type": "text", + "content": "margins for all the combinations of datasets and algorithms. More specifically, for various settings, ASRs of our attacks are at least " + }, + { + "bbox": [ + 46, + 329, + 287, + 413 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 329, + 287, + 413 + ], + "type": "text", + "content": " more than ASRs of Narcissus and DeHiB attacks, while they are at least " + }, + { + "bbox": [ + 46, + 329, + 287, + 413 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 329, + 287, + 413 + ], + "type": "text", + "content": " more than clean label (CL)-Badnets attacks. Due to space limits, we discuss all the baseline attacks in Appendix A. For UDA + CIFAR10, ASR is " + }, + { + "bbox": [ + 46, + 329, + 287, + 413 + ], + "type": "inline_equation", + "content": "81.5\\%" + }, + { + "bbox": [ + 46, + 329, + 287, + 413 + ], + "type": "text", + "content": " with poisoning just a " + }, + { + "bbox": [ + 46, + 329, + 287, + 413 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 46, + 329, + 287, + 413 + ], + "type": "text", + "content": " of training data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 413, + 288, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 288, + 521 + ], + "type": "text", + "content": "Narcissus and " + }, + { + "bbox": [ + 46, + 413, + 288, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{DeHiB}^3" + }, + { + "bbox": [ + 46, + 413, + 288, + 521 + ], + "type": "text", + "content": " attacks achieve close to " + }, + { + "bbox": [ + 46, + 413, + 288, + 521 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 46, + 413, + 288, + 521 + ], + "type": "text", + "content": " ASR for most combinations of datasets and SSL algorithms. As discussed in Appendix A.3, this is expected because all SSL algorithms use strong augmentations which easily obfuscate the dynamic backdoor triggers of these attacks. CL-Badnets attack exhibits relatively higher ASR performances, which is due to the static pattern of its triggers. However, the attack's ASRs remain below " + }, + { + "bbox": [ + 46, + 413, + 288, + 521 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 46, + 413, + 288, + 521 + ], + "type": "text", + "content": ", while ASRs of our attacks exceed " + }, + { + "bbox": [ + 46, + 413, + 288, + 521 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 413, + 288, + 521 + ], + "type": "text", + "content": " in all the cases." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 538, + 257, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 257, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 257, + 551 + ], + "type": "text", + "content": "5.1.2 Negligible impact on clean accuracy (CA)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "type": "text", + "content": "\"CA\" columns in Table 3 show the results. First note that, as Table 6 shows, we use significantly more labeled data for MixMatch than for the other semi-supervised algorithms, and therefore, for some datasets, MixMatch achieves higher accuracy than ReMixMatch or FixMatch. Note from Table 3 that our attacks are highly stealthy as they reduce CA by less than " + }, + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 557, + 287, + 665 + ], + "type": "text", + "content": ". Baseline attacks also reduce CA negligibly, but their ASRs are very low. Interestingly, for some combinations of dataset and algorithms, we observe an in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 329, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 329, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 329, + 545, + 425 + ], + "type": "text", + "content": "crease in CA when we mount our attacks, e.g., for CIFAR10 + FixMatch, CA increases from " + }, + { + "bbox": [ + 306, + 329, + 545, + 425 + ], + "type": "inline_equation", + "content": "91.1\\%" + }, + { + "bbox": [ + 306, + 329, + 545, + 425 + ], + "type": "text", + "content": " in the benign setting to " + }, + { + "bbox": [ + 306, + 329, + 545, + 425 + ], + "type": "inline_equation", + "content": "93.5\\%" + }, + { + "bbox": [ + 306, + 329, + 545, + 425 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 306, + 329, + 545, + 425 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 306, + 329, + 545, + 425 + ], + "type": "text", + "content": " absolute increase. We also observe that such CA increases generally accompany an increase in the target class accuracy (TA). This is because our attacks add a specific trigger to a subset of target class data and give the model an extra signal to better learn the target class. This improves the TA, and hence, also increases CA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 440, + 539, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 440, + 539, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 440, + 539, + 452 + ], + "type": "text", + "content": "5.1.3 Negligible impact on target class accuracy (TA)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 458, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 602 + ], + "type": "text", + "content": "\"TA\" columns in Table 3 show the results. Our attacks remain stealthy with respect to TA as well, as they incur negligible " + }, + { + "bbox": [ + 304, + 458, + 545, + 602 + ], + "type": "inline_equation", + "content": "(< 3\\%)" + }, + { + "bbox": [ + 304, + 458, + 545, + 602 + ], + "type": "text", + "content": " reduction in TA. The baseline attacks also do not reduce TAs, but their ASRs are very low. For STL10 + FlexMatch, we observe the maximum, " + }, + { + "bbox": [ + 304, + 458, + 545, + 602 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 458, + 545, + 602 + ], + "type": "text", + "content": ", reduction in TA. This is because the number of samples for a class that FlexMatch uses during training is inversely proportional to the confidence of the model on that class; the addition of backdoor trigger to the target class data increases the models' confidence on the target class and reduces the target class data that FlexMatch uses for training. On the other hand, in some cases, TA increases as discussed above." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 610, + 531, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 531, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 531, + 623 + ], + "type": "text", + "content": "5.1.4 Why and how our attacks work against SSL?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": "Below, we explain why and how our attack backdoors SSL. For brevity, we limit our discussions to FixMatch and ReMixMatch on CIFAR10 data with target class, " + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "inline_equation", + "content": "y^{t} = 7" + }, + { + "bbox": [ + 304, + 629, + 545, + 677 + ], + "type": "text", + "content": ", i.e., \"horse\", but the insights apply to other SSL algorithms." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "FixMatch: FixMatch (as discussed in Section 3.2) uses the current model, " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and assigns hard pseudo-labels to unlabeled data, " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "D^{u}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ", on which " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " has high confidence. Hence," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "text", + "content": "3Note that, the original DeHib attack makes an unrealistic assumption, i.e., access to the labeled portion, " + }, + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "inline_equation", + "content": "D^{l}" + }, + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "text", + "content": ", of the training data. Hence, for a fair comparison, instead of exact " + }, + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "inline_equation", + "content": "D^{l}" + }, + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "text", + "content": ", we assume that the attacker has some labeled data with same distribution as " + }, + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "inline_equation", + "content": "D^{l}" + }, + { + "bbox": [ + 46, + 673, + 287, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4735" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 42, + 72, + 289, + 187 + ], + "blocks": [ + { + "bbox": [ + 42, + 72, + 289, + 187 + ], + "lines": [ + { + "bbox": [ + 42, + 72, + 289, + 187 + ], + "spans": [ + { + "bbox": [ + 42, + 72, + 289, + 187 + ], + "type": "image", + "image_path": "851a6c262d43590641532fb4ab09a22e2e72da6a9cb639a2a9d81d4ad2348168.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 192, + 287, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 287, + 242 + ], + "type": "text", + "content": "Figure 3: Dynamics of our backdoor attacks: Initially, SSL assigns the backdoor target class " + }, + { + "bbox": [ + 46, + 192, + 287, + 242 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 46, + 192, + 287, + 242 + ], + "type": "text", + "content": " as pseudo-labels to poisoning data. Then, our attack forces the model to learn simpler task of associating the backdoor trigger to " + }, + { + "bbox": [ + 46, + 192, + 287, + 242 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 46, + 192, + 287, + 242 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 66, + 273, + 266, + 300 + ], + "blocks": [ + { + "bbox": [ + 46, + 248, + 287, + 272 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 287, + 272 + ], + "type": "text", + "content": "Table 4: Backdoor attacks' invisibility as " + }, + { + "bbox": [ + 46, + 248, + 287, + 272 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 46, + 248, + 287, + 272 + ], + "type": "text", + "content": "-norm of their trigger for CIFAR10. Stealthy attacks have small norms." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 273, + 266, + 300 + ], + "lines": [ + { + "bbox": [ + 66, + 273, + 266, + 300 + ], + "spans": [ + { + "bbox": [ + 66, + 273, + 266, + 300 + ], + "type": "table", + "html": "
CL-BadnetsNarcissusDeHiBOurs
At train time255/25532/25532/25530/255
At test time255/25532/25532/25530/255
", + "image_path": "67c1d28eb40587aaf17762b758f006f706979a4ce8f11fe906ae63b0848ee29d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "text", + "content": "to understand why and how our attacks work against FixMatch, in Figure 3-(left), we plot averages of hard pseudolabels of FixMatch on backdoored (poisoning) unlabeled data, " + }, + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "text", + "content": " and our attack's ASRs as SSL progresses. As training progresses, FixMatch assigns the " + }, + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "text", + "content": " to more and more of " + }, + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "text", + "content": ". This forces " + }, + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "text", + "content": " to shift its objective from learning the difficult salient features of the target class to learning much simpler backdoor trigger. Hence, backdoor ASR increases as the average pseudo-label shifts to " + }, + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "inline_equation", + "content": "y^t = 7" + }, + { + "bbox": [ + 46, + 311, + 287, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": "ReMixMatch: As detailed in Section 2.1, ReMixMatch averages predictions on a few augmented versions of an " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in X^p" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": " and then uses distribution alignment to compute a prediction vector that it uses as a soft label to train " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": ". Hence, to understand our backdoor attack on ReMixMatch, in Figure 3-(right), as the training progresses, we plot the average of " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": "'s confidence on " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": ", and backdoor ASRs. Initially ReMixMatch assigns low confidence to " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": " that is due to the distribution alignment, which ensures that ReMixMatch does not assign very high confidence to any single class. However, note from Figure 3 that, once " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": " learns the salient features of " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": " as true label), " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": " assigns very high confidence to " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": ". Next, similar to FixMatch, " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": " is forced to learn to associate trigger with " + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 46, + 419, + 287, + 586 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "text", + "content": "Summary: Our backdoor attacks exploit the high performance of modern SSL algorithms: As our intuition hypothesized in Section 3.2, once they achieve high confidences on " + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "text", + "content": ", our attack forces " + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "text", + "content": " to associate the simple trigger pattern of our attack with the target class, thereby installing the backdoor." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 216, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 216, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 216, + 681 + ], + "type": "text", + "content": "5.1.5 Additional effectiveness metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Visibility of backdoor trigger: The unlabeled data of SSL pipeline is never inspected, hence we believe that the visi" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "bility of our backdoor triggers is not a significant concern. Nevertheless, following [50], we measure the visibility of backdoor attack as the " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "-norm of their backdoor trigger " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": ". The lower the " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "-norm of a trigger, the more stealthier the backdoor attack. Table 4 shows the " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "-norms of " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " used for CIFAR10. We note that " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "-norm, i.e., visibility, of our " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": " is lower than that of all the baseline attacks. For many combinations of dataset and SSL algorithms, we need even lower " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "-norm triggers, e.g., to attack CIFAR10 with FixMatch, UDA and FlexMatch, we use " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "L_{\\infty} = 20 / 255" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": ", while attack on CIFAR100 with FixMatch uses " + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "inline_equation", + "content": "L_{\\infty} = 10 / 255" + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 205, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 546, + 445 + ], + "type": "text", + "content": "Efficacy against strong augmentations: In this section, we show that our attacks not only work against SSL, but generally perform well against strong augmentations (SA). To this end, we evaluate CL-Badnets, Narcissus and our attack against supervised ML (SML) with and without SAs (we use RandAugment [10]) and provide results in Table 5 for CIFAR10 and CIFAR100. We poison " + }, + { + "bbox": [ + 304, + 205, + 546, + 445 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 304, + 205, + 546, + 445 + ], + "type": "text", + "content": " of entire labeled training data for Narcissus and our attacks and " + }, + { + "bbox": [ + 304, + 205, + 546, + 445 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 205, + 546, + 445 + ], + "type": "text", + "content": " for CL-Badnets attack. We note that although CL-Badnets works well against SML without SAs, it completely fails when we use SAs. On the other hand, our attack works well against SML with and without SAs. Interestingly, Narcissus also works against SML with SAs, but completely fails against SSL (Section 5.1.1 and Appendix A.3). This is because, unlike in SSL, in SML, Narcissus already has the target label for its poisoning data. To summarize, our static pattern backdoor attack is a general attack against strong augmentations and can serve as a building block of backdoor attacks on numerous learning paradigms that use strong augmentations, e.g., self-supervised learning [6]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 454, + 397, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 397, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 397, + 467 + ], + "type": "text", + "content": "5.2. Ablation study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 473, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 533 + ], + "type": "text", + "content": "(1) Impact of sizes of labeled training data " + }, + { + "bbox": [ + 304, + 473, + 545, + 533 + ], + "type": "inline_equation", + "content": "(D^{l})" + }, + { + "bbox": [ + 304, + 473, + 545, + 533 + ], + "type": "text", + "content": ": Figure 4 plots ASR, CA and TA for our backdoor attacks when we vary " + }, + { + "bbox": [ + 304, + 473, + 545, + 533 + ], + "type": "inline_equation", + "content": "|D^{l}|" + }, + { + "bbox": [ + 304, + 473, + 545, + 533 + ], + "type": "text", + "content": ". Due to resource constraints, we experiment with a subset of combinations from Table 3 and use trigger intensities as reported in Table 3 for the combinations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "We note that ASRs remain above " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " in almost all cases, however we observe a dataset dependent pattern: increasing " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "|D^l|" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", ASRs first reduce and then increase for CIFAR10, but vice-versa for SVHN (Figure 11). We leave analyses of this phenomena to future work. For FixMatch, we observe that ASRs are almost always above " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ". This is because FixMatch has high TA and uses hard pseudo-labels, and hence, all poisoning data, " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", is correctly pseudolabeled as the backdoor target class, " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ". Consequently, the model learns to associate the trigger pattern with " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ". For CIFAR10 + ReMixMatch, we see that TAs are comparable to FixMatch but ASRs are lower. This is because ReMixMatch uses multiple regularizations, including mixup [53] that uses a convex combination of two randomly selected samples and their labels from training data to train the model," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4736" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 83, + 547, + 126 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 539, + 82 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 539, + 82 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 539, + 82 + ], + "type": "text", + "content": "Table 5: Efficacy of various backdoor attacks against supervised ML (SML) with and without strong augmentations (SA)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 83, + 547, + 126 + ], + "lines": [ + { + "bbox": [ + 48, + 83, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 547, + 126 + ], + "type": "table", + "html": "
AlgorithmCIFAR10CIFAR100
CL-BadnetsNarcissusOur attackCL-BadnetsNarcissusOur attack
CAASRTACAASRTACAASRTACAASRTACAASRTACAASRTA
SML94.783.495.794.6100.096.594.599.895.380.275.379.080.198.186.280.296.890.0
SML + SA94.40.096.794.499.596.894.488.994.980.40.076.080.092.184.380.280.289.0
", + "image_path": "9b4ce001142d5688b71459d0d1581cfadcc5612362b21b567426c60a12114674.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 138, + 172, + 248 + ], + "blocks": [ + { + "bbox": [ + 53, + 138, + 172, + 248 + ], + "lines": [ + { + "bbox": [ + 53, + 138, + 172, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 138, + 172, + 248 + ], + "type": "image", + "image_path": "c556492bb9f47abc1b1526f5673785b2e4942197ffc3f0b3bd3e6cab4771857b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 138, + 282, + 247 + ], + "blocks": [ + { + "bbox": [ + 175, + 138, + 282, + 247 + ], + "lines": [ + { + "bbox": [ + 175, + 138, + 282, + 247 + ], + "spans": [ + { + "bbox": [ + 175, + 138, + 282, + 247 + ], + "type": "image", + "image_path": "9b891b8133becfc0db657fb79d096063a757abe5732abb498f5d3b48bb78dd56.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 52, + 251, + 171, + 367 + ], + "blocks": [ + { + "bbox": [ + 52, + 251, + 171, + 367 + ], + "lines": [ + { + "bbox": [ + 52, + 251, + 171, + 367 + ], + "spans": [ + { + "bbox": [ + 52, + 251, + 171, + 367 + ], + "type": "image", + "image_path": "686fffba4f4b3010fd3b3b0ecab9c96222b38bd613ea2ecb5b46b531e4e5a20c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 373, + 286, + 409 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 286, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 286, + 409 + ], + "type": "text", + "content": "Figure 4: Impacts of varying labeled training data size, " + }, + { + "bbox": [ + 46, + 373, + 286, + 409 + ], + "type": "inline_equation", + "content": "|D^l|" + }, + { + "bbox": [ + 46, + 373, + 286, + 409 + ], + "type": "text", + "content": ", for CIFAR10 and {FixMatch, ReMixMatch} algorithms. Upper row shows ASRs and lower row shows CAs and TAs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 175, + 251, + 282, + 367 + ], + "blocks": [ + { + "bbox": [ + 175, + 251, + 282, + 367 + ], + "lines": [ + { + "bbox": [ + 175, + 251, + 282, + 367 + ], + "spans": [ + { + "bbox": [ + 175, + 251, + 282, + 367 + ], + "type": "image", + "image_path": "593a8e1d0b7fb14531477160b91c7024b9e811755a6550b896d59cdd4bc7bfa7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 422, + 286, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 286, + 458 + ], + "type": "text", + "content": "which reduces the effective trigger intensity and hence reduces the ASR. Due to space limits, we defer SVHN results and their discussion to Appendix C." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 460, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 46, + 460, + 286, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 508 + ], + "type": "text", + "content": "(2) Impact of backdoor target class " + }, + { + "bbox": [ + 46, + 460, + 286, + 508 + ], + "type": "inline_equation", + "content": "(y^{t})" + }, + { + "bbox": [ + 46, + 460, + 286, + 508 + ], + "type": "text", + "content": ": Figure 5 plots ASR, CA and TA of our backdoor attacks for different backdoor target classes, " + }, + { + "bbox": [ + 46, + 460, + 286, + 508 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 46, + 460, + 286, + 508 + ], + "type": "text", + "content": "; poisoning data " + }, + { + "bbox": [ + 46, + 460, + 286, + 508 + ], + "type": "inline_equation", + "content": "X^{p}" + }, + { + "bbox": [ + 46, + 460, + 286, + 508 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 46, + 460, + 286, + 508 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 46, + 460, + 286, + 508 + ], + "type": "text", + "content": " of the total training data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "text", + "content": "With two exceptions, we observe that lower TA for a target class leads to lower ASR. For instance, in CIFAR10 with FixMatch, when " + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "text", + "content": " is 2 and 3, TAs are " + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "inline_equation", + "content": "72\\%" + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "text", + "content": ", respectively. Due to low TAs, FixMatch assigns " + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "text", + "content": " to smaller proportions of " + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "inline_equation", + "content": "X^{p}" + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "text", + "content": ", which reduces ASRs. Note that, Carlini [5] also observed that targeted attacks are more effective against better performing SSL algorithms. We observe similar phenomena for CIFAR10 with ReMixMatch and " + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "inline_equation", + "content": "y^{t} \\in \\{3,5\\}" + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "text", + "content": ", and SVHN with FixMatch and " + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "inline_equation", + "content": "y^{t} \\in \\{3,5\\}" + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "text", + "content": ". However, we observe that for some classes, e.g., CIFAR10 with FixMatch and " + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "inline_equation", + "content": "y^{t} \\in \\{6,8\\}" + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "text", + "content": ", TAs are high but ASRs are close to " + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "text", + "content": ". We suspect that this is because features of these " + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "text", + "content": "s are too simple to learn, and hence, model correctly ignores the backdoor pattern. Finally, we note that for majority of classes, our attack's ASRs remain above " + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 509, + 286, + 687 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "(3) Varying the size of unlabeled poisoning data " + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "inline_equation", + "content": "(X^p)" + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": ": Figure 6 plots ASR, CA and TA for our backdoor attacks" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 301, + 133, + 422, + 245 + ], + "blocks": [ + { + "bbox": [ + 301, + 133, + 422, + 245 + ], + "lines": [ + { + "bbox": [ + 301, + 133, + 422, + 245 + ], + "spans": [ + { + "bbox": [ + 301, + 133, + 422, + 245 + ], + "type": "image", + "image_path": "0be57d5ab5db1fb3d56d79d5879f659c274e66e8a157ed3d695133d2bdb03968.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 424, + 133, + 544, + 246 + ], + "blocks": [ + { + "bbox": [ + 424, + 133, + 544, + 246 + ], + "lines": [ + { + "bbox": [ + 424, + 133, + 544, + 246 + ], + "spans": [ + { + "bbox": [ + 424, + 133, + 544, + 246 + ], + "type": "image", + "image_path": "99e92580986a1066b3d4b3f63ce7ece2227ca82f2e2834a06713f3a7fe84508d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 301, + 250, + 421, + 361 + ], + "blocks": [ + { + "bbox": [ + 301, + 250, + 421, + 361 + ], + "lines": [ + { + "bbox": [ + 301, + 250, + 421, + 361 + ], + "spans": [ + { + "bbox": [ + 301, + 250, + 421, + 361 + ], + "type": "image", + "image_path": "9a597b79ecd7fdb30d373f5fa2f2a396ae49561182a4d6c2a6c37894666d7dae.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 423, + 250, + 544, + 361 + ], + "blocks": [ + { + "bbox": [ + 423, + 250, + 544, + 361 + ], + "lines": [ + { + "bbox": [ + 423, + 250, + 544, + 361 + ], + "spans": [ + { + "bbox": [ + 423, + 250, + 544, + 361 + ], + "type": "image", + "image_path": "ebd75d7ad0c54c4515146a326c64e733d0d490cecc7002287aa23ae8f8bb3d54.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 299, + 365, + 421, + 485 + ], + "blocks": [ + { + "bbox": [ + 299, + 365, + 421, + 485 + ], + "lines": [ + { + "bbox": [ + 299, + 365, + 421, + 485 + ], + "spans": [ + { + "bbox": [ + 299, + 365, + 421, + 485 + ], + "type": "image", + "image_path": "210e2c335db9154670e0684ec6028bcabf9404cf64d5f2b2e62644380ff85af0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 491, + 545, + 516 + ], + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 516 + ], + "type": "text", + "content": "Figure 5: ASR, CA and TA of our backdoor attacks for different backdoor target classes, " + }, + { + "bbox": [ + 304, + 491, + 545, + 516 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 304, + 491, + 545, + 516 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 422, + 365, + 544, + 485 + ], + "blocks": [ + { + "bbox": [ + 422, + 365, + 544, + 485 + ], + "lines": [ + { + "bbox": [ + 422, + 365, + 544, + 485 + ], + "spans": [ + { + "bbox": [ + 422, + 365, + 544, + 485 + ], + "type": "image", + "image_path": "dbdb784c24485f6a60a050406d66859353979a355c08445886a04480c03a2d8e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": "with varying " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "|X^p|" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": ". More specifically, we vary " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "|X^p| \\in \\{0.1, 0.15, 0.2, 0.3, 0.4, 0.5\\}" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": " of the entire training data size. Here, we use labeled data sizes as in Table 6. For all three combinations of dataset and SSL algorithms that we study, we observe that having very small or very large " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "|X^p|" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": " leads to relatively ineffective backdoor attacks. This is because at low " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "|X^p|" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": ", although almost all of the " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": " get the target label, " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": ", they are not sufficient to install a backdoor in the target model. While, for large " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "|X^p|" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": ", not all of the " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": " samples get " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": " and some of them get arbitrary labels that are not " + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 304, + 520, + 545, + 675 + ], + "type": "text", + "content": ". This forces model to associate a single trigger pattern with multiple labels and effectively model completely ignores the trigger, which reduces backdoor ASR." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Throughout our evaluations, we found that our attacks have high performances (ASR>60%) for " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "|X^p| \\in [0.2, 0.4]\\%" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " of the entire training data size. Furthermore," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4737" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 42, + 68, + 157, + 177 + ], + "blocks": [ + { + "bbox": [ + 42, + 68, + 157, + 177 + ], + "lines": [ + { + "bbox": [ + 42, + 68, + 157, + 177 + ], + "spans": [ + { + "bbox": [ + 42, + 68, + 157, + 177 + ], + "type": "image", + "image_path": "9a147a2085a663be00db382217a38fe313450dc9a57e267783a26c8b1683263d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 161, + 68, + 278, + 176 + ], + "blocks": [ + { + "bbox": [ + 161, + 68, + 278, + 176 + ], + "lines": [ + { + "bbox": [ + 161, + 68, + 278, + 176 + ], + "spans": [ + { + "bbox": [ + 161, + 68, + 278, + 176 + ], + "type": "image", + "image_path": "1dbf1a152b9e8e668e0cad82fcb7a3ae3b71cd7b90e7b9e0d539d8b2907769a5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 42, + 178, + 157, + 289 + ], + "blocks": [ + { + "bbox": [ + 42, + 178, + 157, + 289 + ], + "lines": [ + { + "bbox": [ + 42, + 178, + 157, + 289 + ], + "spans": [ + { + "bbox": [ + 42, + 178, + 157, + 289 + ], + "type": "image", + "image_path": "2f7147a00c0c52ad864071dc5de1b018db5a2bef255af5f6f081892f6f957aba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 161, + 178, + 278, + 288 + ], + "blocks": [ + { + "bbox": [ + 161, + 178, + 278, + 288 + ], + "lines": [ + { + "bbox": [ + 161, + 178, + 278, + 288 + ], + "spans": [ + { + "bbox": [ + 161, + 178, + 278, + 288 + ], + "type": "image", + "image_path": "1361255578e11827fba38627728dd9ca9b653be8134564cfb91b8d3a3ec97700.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 42, + 290, + 157, + 407 + ], + "blocks": [ + { + "bbox": [ + 42, + 290, + 157, + 407 + ], + "lines": [ + { + "bbox": [ + 42, + 290, + 157, + 407 + ], + "spans": [ + { + "bbox": [ + 42, + 290, + 157, + 407 + ], + "type": "image", + "image_path": "cc3d8c174c3a8da6075d2c35b4ada8e2f949ea21084fc4364f91430969af9eb0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 411, + 287, + 435 + ], + "lines": [ + { + "bbox": [ + 47, + 411, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 287, + 435 + ], + "type": "text", + "content": "Figure 6: Impact of varying the sizes of poisoning data on ASR, CA and TA of our backdoor attacks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 162, + 290, + 278, + 407 + ], + "blocks": [ + { + "bbox": [ + 162, + 290, + 278, + 407 + ], + "lines": [ + { + "bbox": [ + 162, + 290, + 278, + 407 + ], + "spans": [ + { + "bbox": [ + 162, + 290, + 278, + 407 + ], + "type": "image", + "image_path": "37a68b8b80046fcdfd08140096b63333dc353231f68b7c0fd52d790017f0f47b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 441, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 466 + ], + "type": "text", + "content": "within these ranges, our attacks remain stealthy and do not significantly impact CA or TA of the backdoored models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 475, + 266, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 266, + 488 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 266, + 488 + ], + "type": "text", + "content": "5.3. Defenses against backdoor attacks on SSL" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 495, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 287, + 543 + ], + "type": "text", + "content": "Due to space limits, here we give highlights of our evaluations of efficacy of five existing backdoor defenses against our attack and defer detailed discussion to Appendix C.2; Table 7 and Figure 12 in Appendix C.2 show the results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 544, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 47, + 544, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 287, + 579 + ], + "type": "text", + "content": "(1) Fine-tuning (FT) and fine-pruning (FP) both can reduce ASR of our backdoor attack, however it comes at a significant reduction in CA of the resulting models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 580, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 287, + 652 + ], + "type": "text", + "content": "(2) Neural attention distillation (NAD) [24] performed best among the defenses we evaluate: for CIFAR10, NAD reduces ASR by " + }, + { + "bbox": [ + 47, + 580, + 287, + 652 + ], + "type": "inline_equation", + "content": "22.1\\%" + }, + { + "bbox": [ + 47, + 580, + 287, + 652 + ], + "type": "text", + "content": " for FixMatch and by " + }, + { + "bbox": [ + 47, + 580, + 287, + 652 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 47, + 580, + 287, + 652 + ], + "type": "text", + "content": " for ReMixMatch, but it does not perform as well for SVHN. Nonetheless, even with NAD, our attack still raises significant concerns as its ASR against NAD is always " + }, + { + "bbox": [ + 47, + 580, + 287, + 652 + ], + "type": "inline_equation", + "content": ">60\\%" + }, + { + "bbox": [ + 47, + 580, + 287, + 652 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": "(3) Strip [17] works well when backdoor ASR is very high but for moderate ASRs " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\in [80,90]\\%" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": " it fails to detect backdoor. For instance, Strip successfully identifies over " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": " of the backdoored test inputs, but it completely fails against CIFAR10 + FixMatch/ReMixMatch and SVHN +" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 353, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 353, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 353, + 83 + ], + "type": "text", + "content": "MixMatch." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 84, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 144 + ], + "type": "text", + "content": "(4) Anti-backdoor learning [25] (ABL) completely fails against SSL, because, SSL training extensively uses strong augmentations, and hence, the unsupervised loss on poisoning unlabeled data remains almost the same as that on benign unlabeled data (Figure 13 in Appendix C.2)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 144, + 546, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 546, + 289 + ], + "type": "text", + "content": "How to defend SSL from unlabeled data poisoning? We find that some of the SOTA post-processing (FT, FP, NAD, Strip) or in-processing (ABL) defenses cannot defend SSL from our attacks. In other words, current SSL practice of using non-inspected unlabeled training data makes it highly vulnerable to poisoning. Hence, we argue for SSL to depart from its philosophy of not at all inspecting its unlabeled training data and pre-process the unlabeled data to thwart poisoning attacks. Such pre-processing can be tailored to our attack, e.g., check for existence of patterns that follow our three lessons, or check for any abnormal frequency artifacts [51]. We leave pursuing this direction to future work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 300, + 383, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 383, + 313 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 383, + 313 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 321, + 545, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 451 + ], + "type": "text", + "content": "Semi-supervised learning (SSL) allows training on large unlabeled data without any inspection, thereby significantly reducing the cost of ML training. Unfortunately, as we show, this key feature can facilitate strong data poisoning attacks on SSL: a naive adversary, without any knowledge of training data or model architecture, can poison just " + }, + { + "bbox": [ + 304, + 321, + 545, + 451 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 304, + 321, + 545, + 451 + ], + "type": "text", + "content": " of the entire training data to install a strong backdoor functionality in SSL models. Our attack remains effective against various SSL algorithms and benchmark datasets, and even circumvents state-of-the-art defenses against backdoor attacks." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 453, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 536 + ], + "type": "text", + "content": "Note that, in contrast to numerous prior works [37, 48, 50, 44, 7, 4, 5, 29], in this work we considered a much weaker, hence more realistic, adversary. Due to our weak adversarial assumptions and simple attack methodology, all of the existing and future SSL applications can use our attack to measure and enhance their robustness against backdoor poisoning." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 537, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 545, + 621 + ], + "type": "text", + "content": "Backdoor attacks can have severe consequences in practice, e.g., gaining unauthorized access to a system [7] or denying services to minorities [38]. Hence, a major implication of our study is that real-world SSL applications cannot rely on non-inspected unlabeled data and must preprocess/inspect unlabeled training data and/or design SSL algorithms that are robust to unlabeled data poisoning." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 633, + 408, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 408, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 408, + 646 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in-part by the NSF grants 2131910 and 1953786 and by DARPA under Agreement No. HR00112190125. Approved for public release; distribution is unlimited. This work was also supported in-part by Sony AI." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4738" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In International conference on machine learning, pages 284-293. PMLR, 2018. 12" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 190 + ], + "type": "text", + "content": "[2] David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remix-match: Semi-supervised learning with distribution matching and augmentation anchoring. In International Conference on Learning Representations, 2019. 1, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 244 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 244 + ], + "type": "text", + "content": "[3] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in Neural Information Processing Systems, 32, 2019. 1, 3, 5, 6, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 280 + ], + "type": "text", + "content": "[4] B. Biggio, B. Nelson, and P. Laskov. Poisoning attacks against support vector machines. In Proceedings of 29th International Conference on Machine Learning, 2012. 1, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 287, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 322 + ], + "type": "text", + "content": "[5] Nicholas Carlini. Poisoning the unlabeled dataset of {Semi-Supervised} learning. In 30th USENIX Security Symposium (USENIX Security 21), pages 1577-1592, 2021. 5, 8, 9, 14, 15" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 368 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 412 + ], + "type": "text", + "content": "[7] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 3, 5, 9" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 468 + ], + "type": "text", + "content": "[8] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 215-223. JMLR Workshop and Conference Proceedings, 2011. 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 511 + ], + "type": "text", + "content": "[9] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 567 + ], + "type": "text", + "content": "[10] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 2, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 601 + ], + "type": "text", + "content": "[11] Aron Culotta and Andrew McCallum. Reducing labeling effort for structured prediction tasks. In AAAI, volume 5, pages 746-751, 2005. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 635 + ], + "type": "text", + "content": "[12] Jia Deng. A large-scale hierarchical image database. Proc. of IEEE Computer Vision and Pattern Recognition, 2009, 2009. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 680 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[14] Emily Denton, Sam Gross, and Rob Fergus. Semi-supervised learning with context-conditional generative adversarial networks. arXiv preprint arXiv:1611.06430, 2016. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 2, 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "text", + "content": "[16] Le Feng, Sheng Li, Zhenxing Qian, and Xinpeng Zhang. Unlabeled backdoor poisoning in semi-supervised learning. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "text", + "content": "[17] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 113–125, 2019. 9, 16" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "text", + "content": "[18] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 1, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "text", + "content": "[19] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 7:47230-47244, 2019. 3, 13" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 284, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 314 + ], + "type": "text", + "content": "[20] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 14" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 316, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 346 + ], + "type": "text", + "content": "[21] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 350, + 545, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 393 + ], + "type": "text", + "content": "[22] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, page 896, 2013. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 394, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 447 + ], + "type": "text", + "content": "[23] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 492 + ], + "type": "text", + "content": "[24] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. In International Conference on Learning Representations, 2020. 9, 16" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 494, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 494, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 494, + 545, + 536 + ], + "type": "text", + "content": "[25] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. Advances in Neural Information Processing Systems, 34:14900-14912, 2021. 9, 16" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 581 + ], + "type": "text", + "content": "[26] Yuncheng Li, Jianchao Yang, Yale Song, Liangliang Cao, Jiebo Luo, and Li-Jia Li. Learning from noisy labels with distillation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1910-1918, 2017. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "text", + "content": "[27] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273-294. Springer, 2018. 16" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 689 + ], + "type": "text", + "content": "[28] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In 25th Annual Network And Distributed System Security Symposium (NDSS 2018). Internet Soc, 2018. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[29] Luis Muñoz-González, Battista Biggio, Ambra Demontis, Andrea Paudice, Vasin Wongrassamee, Emil C Lupu, and" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4739" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "Fabio Roli. Towards poisoning of deep learning algorithms with back-gradient optimization. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pages 27-38. ACM, 2017. 1, 2, 9" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 139 + ], + "type": "text", + "content": "[30] Kevin P Murphy. Machine learning: a probabilistic perspective. MIT press, 2012. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[31] Nagarajan Natarajan, Inderjit S Dhillon, Pradeep K Ravikumar, and Ambuj Tewari. Learning with noisy labels. Advances in neural information processing systems, 26, 2013. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "text", + "content": "[32] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 274 + ], + "type": "text", + "content": "[33] Tuan Anh Nguyen and Anh Tran. Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems, 33:3454-3464, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 319 + ], + "type": "text", + "content": "[34] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 11957-11965, 2020. 1, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 322, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 322, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 287, + 365 + ], + "type": "text", + "content": "[35] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 287, + 409 + ], + "type": "text", + "content": "[36] Esha Sarkar, Hadjer Benkraouda, and Michail Maniatakos. Facehack: Triggering backdoored facial recognition systems using facial characteristics. arXiv preprint arXiv:2006.11623, 2020. 1, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 412, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 412, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 412, + 287, + 454 + ], + "type": "text", + "content": "[37] Virat Shejwalkar and Amir Houmansadr. Manipulating the byzantine: Optimizing model poisoning attacks and defenses for federated learning. In The Network and Distributed System Security Symposium (NDSS), 2021. 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "type": "text", + "content": "[38] V. Shejwalkar, A. Houmansadr, P. Kairouz, and D. Ramage. Back to the drawing board: A critical evaluation of poisoning attacks on production federated learning. In 2022 2022 IEEE Symposium on Security and Privacy (SP) (SP), pages 1117-1134, Los Alamitos, CA, USA, may 2022. IEEE Computer Society. 1, 2, 4, 9" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 524, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 287, + 555 + ], + "type": "text", + "content": "[39] Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1-48, 2019. 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 622 + ], + "type": "text", + "content": "[40] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in Neural Information Processing Systems, 33:596-608, 2020. 1, 3, 4, 5, 6, 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[41] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[42] Hossein Souri, Micah Goldblum, Liam Fowl, Rama Chellappa, and Tom Goldstein. Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. arXiv preprint arXiv:2106.08970, 2021. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 665 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[43] Luke Taylor and Geoff Nitschke. Improving deep learning with generic data augmentation. In 2018 IEEE Symposium Series on Computational Intelligence (SSCI), pages 1542-1547. IEEE, 2018. 12" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[44] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 1, 5, 9, 13" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[45] Yidong Wang, Hao Chen, Yue Fan, Hao Wu, Bowen Zhang, Wenxin Hou, Yuhao Chen, and Jindong Wang. Torchssl: A pytorch-based toolbox for semi-supervised learning. https://github.com/TorchSSL/TorchSSL, 2021. [Online; accessed 03-July-2022]. 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[46] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems, 33:6256-6268, 2020. 1, 3, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "text", + "content": "[47] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 373 + ], + "type": "text", + "content": "[48] Zhicong Yan, Gaolei Li, Yuan Tlan, Jun Wu, Shenghong Li, Mingzhe Chen, and H Vincent Poor. Dehib: Deep hidden backdoor attack on semi-supervised learning via adversarial perturbation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 10585-10593, 2021. 2, 3, 5, 9, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "type": "text", + "content": "[49] Zhicong Yan, Jun Wu, Gaolei Li, Shenghong Li, and Mohsen Guizani. Deep neural backdoor in semi-supervised learning: threats and countermeasures. IEEE Transactions on Information Forensics and Security, 16:4827-4842, 2021. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 420, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 545, + 464 + ], + "type": "text", + "content": "[50] Yi Zeng, Minzhou Pan, Hoang Anh Just, Lingjuan Lyu, Meikang Qiu, and Ruoxi Jia. Narcissus: A practical clean-label backdoor attack with limited information. arXiv preprint arXiv:2204.05255, 2022. 1, 2, 5, 7, 9, 12, 13" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 465, + 545, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 465, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 307, + 465, + 545, + 508 + ], + "type": "text", + "content": "[51] Yi Zeng, Won Park, Z Morley Mao, and Ruoxi Jia. Rethinking the backdoor attacks' triggers: A frequency perspective. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16473-16481, 2021. 3, 9" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 510, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 564 + ], + "type": "text", + "content": "[52] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 566, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 566, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 545, + 609 + ], + "type": "text", + "content": "[53] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 610, + 545, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 610, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 307, + 610, + 545, + 665 + ], + "type": "text", + "content": "[54] Haoti Zhong, Cong Liao, Anna Cinzia Squicciarini, Sencun Zhu, and David Miller. Backdoor embedding in convolutional neural network models via invisible perturbation. In Proceedings of the Tenth ACM Conference on Data and Application Security and Privacy, pages 97-108, 2020. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4740" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/3b9298a9-5799-4904-96da-bef6d287733c_content_list.json b/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/3b9298a9-5799-4904-96da-bef6d287733c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cbe6903fc7f3cc7b2d1937447fd8bfc0d417d423 --- /dev/null +++ b/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/3b9298a9-5799-4904-96da-bef6d287733c_content_list.json @@ -0,0 +1,1391 @@ +[ + { + "type": "text", + "text": "The Power of Sound (TPoS): Audio Reactive Video Generation with Stable Diffusion", + "text_level": 1, + "bbox": [ + 205, + 130, + 764, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yujin Jeong $^{1}$ , Wonjeong Ryoo $^{2}$ , Seunghyun Lee $^{2}$ , Dabin Seo $^{1}$ , Wonmin Byeon $^{3}$ , Sanggil Kim, $^{2,*}$ and Jinkyu Kim $^{1,*}$", + "bbox": [ + 236, + 202, + 730, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Computer Science and Engineering, Korea University, Seoul 02841, Korea \n $^{2}$ Department of Artificial Intelligence, Korea University, Seoul 02841, Korea \n $^{3}$ NVIDIA Research, Santa Clara 95050, USA", + "bbox": [ + 125, + 239, + 844, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Correspondences: S. Kim (spk7@korea.ac.kr) and J. Kim (jinkyukim@korea.ac.kr)", + "bbox": [ + 192, + 292, + 777, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b7c727e73eae2e0ef7588d2ee6ef4ab3623e5cc1fa50a4a3d64e8e9aca43c6c.jpg", + "image_caption": [ + "Figure 1: The Power of Sound (TPoS) is a novel framework that generates audio-reactive video sequences. Built upon the Stable Diffusion model, our model first generates an initial frame from a user-provided text prompt (e.g. \"a photo of a beautiful beach with a blue sky\"), then reactively manipulates the style of generated images corresponding to the sound inputs (e.g. an audio sequence of fireplace). Our model is indeed able to generate a frame conditioned on semantic information of the sound (see 1st and 2nd rows where images are manipulated driven by sound inputs such as fireplace or wave sound), while realistically dealing with temporal visual changes conditioned on changes of sound, e.g., increasing magnitude of sounds (see second row) or wave $\\rightarrow$ fireplace (see last row). TPoS creates visually compelling and contextually relevant video sequences in an open domain." + ], + "image_footnote": [], + "bbox": [ + 133, + 318, + 844, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 760, + 313, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, video generation has become a prominent generative tool and has drawn significant attention. However, there is little consideration in audio-to-video generation, though audio contains unique qualities like temporal semantics and magnitude. Hence, we propose The Power of Sound (TPoS) model to incorporate audio input that includes both changeable temporal semantics and mag-", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nitude. To generate video frames, TPoS utilizes a latent stable diffusion model with textual semantic information, which is then guided by the sequential audio embedding from our pretrained Audio Encoder. As a result, this method produces audio reactive video contents. We demonstrate the effectiveness of TPoS across various tasks and compare its results with current state-of-the-art techniques in the field of audio-to-video generation. More examples are available", + "bbox": [ + 496, + 762, + 893, + 883 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7822", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 135, + 209, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent generative models have demonstrated the potential to generate visuallyappealing video frames [32, 11, 10, 23, 36]. They often use a simple text prompt (e.g., \"a video of a person on the street on a rainy day\") to generate a video which is intuitive for end-users to drive a video generation. Text can effectively convey uni-modal object-wise guidance, such as \"a rainy day\" or \"a person\", but it may be challenging if users want to drive it into more complex sequential procedures, i.e., \"a video of a person on the street on a rainy day, but a rain suddenly stops, and a wind blows.\"", + "bbox": [ + 76, + 160, + 468, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we leverage sounds to guide the video generation models, i.e., sound-driven video generation. Audio is another modality that can complement texts by effectively providing sequential information (or temporal semantics): e.g., a continuous transition from the sound of light rain to the sound of heavy rain. There have been introduced sound-guided video generation approaches. However, existing sound-guided video generation approaches are limited to specific applications, such as face generation [15, 26, 21, 14, 21], where audio is used to provide a script for the avatar, or other simple synthetic motions (e.g., a video of musicians playing violin or a video of painting motions by an artist) [7, 13, 18, 4].", + "bbox": [ + 76, + 311, + 470, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Lee et al. [19] introduced a sound-guided landscape video generation model, leveraging the latent space of StyleGAN [16]. They focus on using audio only for semantic labels (i.e., a sound of the wind is simply encoded into a meaning of wind) but not temporal semantics – i.e. semantic information that changes over time. Thus, in this work, we focus on leveraging temporal semantics from audio inputs such that our video generator reactively manipulates video frames. Our model temporally aligns the latent space with the given audio sequence (e.g., continuous changes in audio, e.g., weak rain $\\rightarrow$ heavy rain, are reflected to generate corresponding video frames).", + "bbox": [ + 76, + 508, + 468, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work starts with Stable Diffusion [29], a text-driven image generator with advantages in generating high-resolution images based on the latent diffusion models. Its architectural advantages (i.e., attention mechanism and diffusion process) help leverage audio as a driving condition, generating temporally reactive and consistent video frames. Given the latent space of trained Stable Diffusion, we generate video frames temporally guided by audio sequences with regularizers to ensure temporal consistency (between generated consecutive frames) and correspondence with audio inputs.", + "bbox": [ + 76, + 689, + 467, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our model consists of two main modules: (i) Audio Encoder, which is designed to encode temporal semantics of audio sequences, producing a sequence of the latent vec", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt: \"A photo of a person on the street\"", + "bbox": [ + 506, + 90, + 687, + 99 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e6459603441b5731c5697bb6451a56ac13e1b78ee84b827291153dec493de081.jpg", + "image_caption": [ + "Figure 2: Limitation of text-driven image manipulation. Given a generated image by Stable Diffusion [29] with a text prompt, \"A photo of a person on the street,\" additional textual conditions of different semantic meanings (i.e. \"a little of rain\", \"rain\", and \"rain a lot\") produce similar images, failing to capture differences. Note that we apply SEGA [2]'s guidance to preserve content identity." + ], + "image_footnote": [], + "bbox": [ + 503, + 99, + 888, + 184 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tors. (ii) Audio Semantic Guidance Module, which uses the above-mentioned latent vectors as a condition in the diffusion process to generate corresponding image outputs. We apply identity regularizer to produce temporally consistent video frames, while we apply audio semantic guidance to generate audio-reactive video frames.", + "bbox": [ + 496, + 319, + 890, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, we observe that generating multifarious high-quality images solely from a sound input is challenging due to the lack of such a large-scale dataset to train a model. Instead, we first generate an initial frame using pre-trained Stable Diffusion model with a text prompt, then generate the following video frames conditioned on audio inputs. This frees a model from data dependency burdens and enables training with the current relatively small (than imagetext modality) audio dataset [19, 5], focusing on leveraging temporal semantics from audio. We summarize our contributions as follows:", + "bbox": [ + 496, + 410, + 892, + 575 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel sound-driven video generation method built upon Stable Diffusion [29] and can generate video frames reactively with audio sequence inputs.", + "- Our attention-based Audio Encoder produces temporally-aware latent vectors, which are consumed by Stable Diffusion as a per-time manipulation condition, producing audio-reactive video frames.", + "- Our model regularizes the latent features of diffusion models to produce temporally consistent video frames, preserving identity throughout the generated video.", + "- We demonstrate the effectiveness of our proposed model using a public dataset Landscape [19], generally outperforming other state-of-the-art sound-driven video generation approaches in terms of video quality metrics and human evaluation." + ], + "bbox": [ + 517, + 578, + 890, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 824, + 640, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Latent Diffusion Models. Recent success [29] suggests that the Latent Diffusion Models (LDM) improve the efficiency of the diffusion process, successfully generating", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "header", + "text": "at https://ku-vai.github.io/TPoS/", + "bbox": [ + 76, + 90, + 395, + 104 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7823", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/e65add0d7e1af0a8137060663caa20c029d1b79bb493eb70159f0c38dee6f0af.jpg", + "table_caption": [ + "Table 1: Comparison between existing state-of-the-art audio-driven video generation approaches in terms of whether they consider the following factors: temporal semantics, magnitude changes of sound, and target domains." + ], + "table_footnote": [], + "table_body": "
ModelTemporal SemanticsMagnitudeDomains (Audio Type)
Sound2Sight [4]-✓Closed
CCVS [18]-✓Closed (Music)
TräumerAI [13]-✓Closed (Music)
Lee et al. [19]✓-Closed (Nature)
Ours✓✓Open Domains
", + "bbox": [ + 78, + 157, + 468, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "high-quality images given a text prompt. One challenge in LDM is that the generation process is too sensitive to the condition, making it difficult to control semantics. Recently, there have been introduced to control semantics with LDM by a semantic mask [1] or by utilizing semantic information of the cross-attention layers [9]. Wu et al. [37] used linear combinations of text embeddings and Liu et al. [22] proposed composable diffusion models, but they still remained challenging to control fine-grained semantic changes. Recently, Semantic Guidance (SEGA) [2] computed a guidance vector in the latent space, enabling semantic control of diffusion models without further inputs. Inspired by SEGA, we also control the semantics in the latent space with temporally-encoded audio vector sequences.", + "bbox": [ + 75, + 268, + 468, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-driven Video Generation. Recent text-to-video generation tools, including Make-A-Video [32], Video Diffusion Models [11], Imagen video [10], and Phenaki [36] have shown promising performance in generating videos from textual descriptions. However, text-to-video generation has its limitations in terms of temporal coherence, which mostly leads to short video duration or a linear video change. Recent text-to-video generation methods, StyleGAN-V [34] and Dreamix [23], made progress addressing these issues. However, conditioning temporal semantics or complex scenarios is still challenging to be obtained from text inputs. Thus, in this paper, we want to explore conditioning a model with audio inputs, which inherently convey such temporal semantics.", + "bbox": [ + 75, + 486, + 468, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio-driven Video Generation. Leveraging temporal semantics was not seriously considered in previous audio-driven video generation approaches. Sound2Sight[4] and CCVS [18] generate video frames conditioned on the (nontemporal) context of the given audio, while TraumerAI [13] utilized the magnitude of the given audio. Recently, Lee et al. [19] explored a model that can consider audio semantics as a condition to drive a video generator. Also, their dependency on StyleGAN [34]-based embedding space makes it difficult for models to generate transitions in video. In this work, we focus on leveraging temporal semantics from audio inputs such that our generator reactively manipulates video frames.", + "bbox": [ + 75, + 703, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 89, + 589, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose a novel audio-driven video generation method, which generates video frames conditioned on audio sequences, ensuring temporal consistency between consecutive frames and temporal correspondence with audio inputs. As shown in Figure 3, our model consists of two main parts: (i) Audio Encoder, which encodes temporal semantics of audio sequences, producing a sequence of the latent vectors (Section 3.2). (ii) Audio Semantic Guidance Module, which uses the above-mentioned latent vectors as a condition in the diffusion process to generate corresponding image outputs, which are temporally consistent (by our identity regularizer) and audio-reactive (see Section 3.3).", + "bbox": [ + 496, + 114, + 893, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary: Latent Stable Diffusion", + "text_level": 1, + "bbox": [ + 498, + 306, + 821, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent Diffusion Models (LDMs) [29] are the method that uses an encoder to convert a noised latent vector $\\mathbf{z}^T$ to a denoised latent vector $\\mathbf{z} = \\mathbf{x} + \\epsilon$ , where $\\mathbf{z}$ is a latent vector of an input image $\\mathbf{x}$ and $\\epsilon$ is a noise. Stable Diffusion [29] is part of a conditional generation model that can synthesize an image given a condition $\\mathbf{y}$ . It uses U-Net [30] as denoising autoencoders represented by $\\epsilon_{\\theta}$ . To generate an output image, the autoencoder $\\epsilon_{\\theta}(\\mathbf{z}^{t}, t, \\tau_{\\theta}(\\mathbf{y}))$ takes three inputs: the noised latent vector $\\mathbf{z}^t$ , a sequence $t$ that is uniformly sampled from the set $1, \\dots, T$ , and a conditional input $\\mathbf{y}$ .", + "bbox": [ + 496, + 330, + 890, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conditional input $\\mathbf{y}$ is first transformed into a latent vector $\\mathbf{c}_p$ through a pretrained function $\\tau_{\\theta}$ and then it is fed into a cross-attention layer of the U-Net as the key and value, which is then combined with $\\mathbf{z}^t$ by an attention mechanism, where the query is the flattened intermediate representation of the U-Net. Next, the denoising autoencoder $\\epsilon_{\\theta}$ is used to denoise $\\mathbf{z}^t$ from $t = T$ to $t = 1$ sequentially. After $T$ denoising steps, the resulting denoised latent vector $\\mathbf{z}(= \\mathbf{z}^{1})$ is transmitted to the decoder, which produces the final output image $\\tilde{\\mathbf{x}}$ .", + "bbox": [ + 496, + 483, + 890, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Encoding Temporal Semantics from Audio", + "text_level": 1, + "bbox": [ + 498, + 643, + 862, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use an audio modality as a source of generating temporal conditions. We divide the audio mel-spectogram into uniform snippets (segments) to extract abundant audio features and feed this into the Audio Encoder to extract both temporal semantic features and intensity of audio in end-to-end manner. The figure of our training process is illustrated in Figure 4.", + "bbox": [ + 496, + 667, + 890, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio Feature Extraction. Audio inputs are first transformed into a mel-spectrogram representation, denoted as $\\mathbf{x}^a\\in \\mathbb{R}^{d\\times w}$ , where $d$ represents the number of mel-frequency bins and $w$ is the width of the spectrogram. To incorporate time information, the mel-spectrogram is divided into $N$ segments. Each segment, denoted as $\\mathbf{x}_n^a\\in \\mathbb{R}^{d\\times \\lceil \\frac{w}{N}\\rceil}$ where $n\\in \\{1,\\ldots ,N\\}$ , is then fed into a shared feature extraction module, i.e., the pre-trained ResNet18 [8]. The", + "bbox": [ + 496, + 779, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7824", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8bd59a99cce9913cd2bf813cf7788bc82765c429d1f8cd78bc500d99afda4e5b.jpg", + "image_caption": [ + "Figure 3: An overview of our proposed TPoS model. Our model consists of two main modules: (i) Audio Encoder, which produces a sequence of latent vectors, encoding temporal semantics of audio input by utilizing CLIP [27] space and highlighting the important temporal features and (ii) Audio Semantic Guidance Module, which is based on the diffusion process, generating video frames that are temporally consistent and audio-reactive." + ], + "image_footnote": [], + "bbox": [ + 122, + 90, + 851, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "feature extraction module $f_{a}(\\cdot)$ learns to extract low-level features from each audio segment regardless of its time dependency, i.e., $\\mathbf{w}_n = f_a(\\mathbf{x}_n^a)$", + "bbox": [ + 75, + 465, + 470, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LSTM-based Temporal Semantic Encoder. As our goal is to generate audio-reactive video frames, it is also important to encode temporal changes (or relations) of the given audio inputs. Similar to Lee et al. [19], given audio features $\\mathbf{w} \\in \\{\\mathbf{w}_1, \\mathbf{w}_2, \\dots, \\mathbf{w}_N\\}$ , which encodes per-segment disjoint audio representation, we apply the standard Long Short-Term Memory (LSTM) network [12] to encode temporal relations or changes between consecutive audio features $\\mathbf{w}$ . Formally, our LSTM takes the audio feature $\\mathbf{w}_{n-1}$ as input and updates its hidden state $\\mathbf{h}_n$ , producing an output $\\mathbf{s}_t$ : i.e. $(\\mathbf{s}_n, \\mathbf{h}_n) = \\mathrm{LSTM}(\\mathbf{h}_{n-1}, \\mathbf{w}_{n-1})$ . These outputs $\\mathbf{s} \\in \\{\\mathbf{s}_1, \\mathbf{s}_2, \\dots, \\mathbf{s}_N\\}$ are then fed into Temporal Attention Module (TAM) to further encode temporal semantics.", + "bbox": [ + 75, + 518, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Aligning Audio Semantics with Image-Text CLIP Joint Space. As we will use the output $s$ as a condition to manipulate video frames, it is important to ensure those audio features are well-aligned with other text and visual features in the CLIP [27]-based joint embedding space. Similar to Lee et al. [20], given the pre-trained image-text CLIP space, we apply the following loss $\\mathcal{L}_{\\mathrm{CLIP}}^{a \\leftrightarrow t}$ with the InfoNCE loss [24] $l_{sim}$ such that positive pairs (e.g. an audio of raining and a text prompt \"raining\") are pulled close to each other, while negative pairs are pushed farther away.", + "bbox": [ + 75, + 720, + 470, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C L I P}} ^ {a \\leftrightarrow t} = l _ {s i m} (\\mathbf {s} _ {N}, \\operatorname {C L I P} _ {t} (\\mathbf {t})) + l _ {s i m} (\\operatorname {C L I P} _ {t} (\\mathbf {t}), \\mathbf {s} _ {N}) \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/618663442c550a1cca5f593d6e0c47315cd9240b4ef34270ff1e8a8c5721ced2.jpg", + "image_caption": [ + "Figure 4: An overview of our Audio Encoder training process. Our model generates temporally-encoded audio embeddings with an LSTM [12] layer and Temporal Attention Module (TAM). Audio input is partitioned into $N$ segments, and each of these is encoded and used as a condition to manipulate audio-reactive video sequences (e.g. light rain → heavy rain). This is done by our Mapping Module (MAP), which maps the audio embedding to the latent space of Stable Diffusion." + ], + "image_footnote": [], + "bbox": [ + 504, + 465, + 885, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{CLIP}_t$ is a pre-trained CLIP-based text encoder, which takes a text prompt $\\mathbf{t}$ obtained from audio class labels as input, yielding an $d$ -dimensional feature. Note that", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7825", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we only apply this loss for the final output $\\mathbf{s}_N$ for efficient training. Given a set of positive pairs, we apply the following InfoNCE loss $l_{sim}(\\mathbf{a},\\mathbf{b})$ :", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {s i m} = - \\log \\frac {\\exp \\left(\\left\\langle \\mathbf {a} _ {i} , \\mathbf {b} _ {i} \\right\\rangle\\right) / \\tau}{\\sum_ {j} \\exp \\left(\\left\\langle \\mathbf {a} _ {i} , \\mathbf {b} _ {j} \\right\\rangle\\right) / \\tau} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 143, + 468, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\langle \\mathbf{a}_i,\\mathbf{b}_i\\rangle$ represents the cosine similarity with temperature $\\tau$ . Note that we set $\\tau$ to 0.07.", + "bbox": [ + 76, + 186, + 468, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Augmenting Audio Semantics. Audio data is often limited in volume and diversity; thus, augmentation techniques may be required to extract better-quality audio semantic features, preventing a representation collapse. We use SpecAugment [25] to apply random transformations (such as masking our certain frequency bands or time segments), yielding augmented audio inputs. We further add the InfoNCE loss [24] $\\mathcal{L}_{\\mathrm{CLIP}}^{a\\leftrightarrow a^{\\prime}}$ to pull augmented audio features together.", + "bbox": [ + 76, + 222, + 468, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C L I P}} ^ {a \\leftrightarrow a ^ {\\prime}} = l _ {s i m} \\left(\\mathbf {s} _ {N}, \\mathbf {s} _ {N} ^ {\\prime}\\right) + l _ {s i m} \\left(\\mathbf {s} _ {N} ^ {\\prime}, \\mathbf {s} _ {N}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 352, + 468, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where apostrophe indicates augmented view of an original audio data. Finally, we use the semantic loss $\\mathcal{L}_{\\mathrm{semantic}}$ : i.e. $\\mathcal{L}_{\\mathrm{semantic}} = \\mathcal{L}_{\\mathrm{CLIP}}^{a\\leftrightarrow t} + \\lambda_s\\mathcal{L}_{\\mathrm{CLIP}}^{a\\leftrightarrow a'}$ where $\\lambda_{s}$ is set to 0.6.", + "bbox": [ + 76, + 381, + 468, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Attention Module (TAM). As shown in Figure 5, we further use an attention-based module to encode temporal semantics from the audio inputs. We empirically observe that adding this module helps improve the quality of video frame generation, which is probably due to the fact that the model becomes flexible to focus on more important temporal information, improving the model's representation power. Formally, we first compute attention weight $\\alpha_{n}$ for a given audio feature $\\mathbf{s}_n$ by applying an MLP layer $f_{\\mathrm{proj}}$ followed by a softmax operation: i.e. $\\alpha_{n} = \\exp (f_{\\mathrm{proj}}(\\mathbf{s}_{n})) / \\sum_{n}\\exp (f_{\\mathrm{proj}}(\\mathbf{s}_{n}))$ such that $\\sum_{n}\\alpha_{n} = 1$ . We compute the weighted sum of audio features based on attention weights, yielding an attended audio feature $\\mathbf{o}^a = \\sum_n\\alpha_n\\mathbf{s}_n$ . We add another InfoNCE loss $\\mathcal{L}_{\\mathrm{temporal}}$ to align the audio features with text:", + "bbox": [ + 76, + 430, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t e m p o r a l}} = l _ {s i m} \\left(\\mathbf {o} ^ {a}, \\operatorname {C L I P} _ {t} (\\mathbf {t})\\right) + l _ {s i m} \\left(\\operatorname {C L I P} _ {t} (\\mathbf {t}), \\mathbf {o} ^ {a}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 666, + 468, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lastly, we also minimize the MSE loss between text embeddings (before the projection layer) and the projected audio feature $\\mathrm{MAP}(\\mathbf{o}^a)$ : $\\mathcal{L}_{\\mathrm{cond}} = ||\\mathbf{c}_t - \\mathrm{MAP}(\\mathbf{o}^a)||_2^2$ . Note that we exclude the feature for the $<\\mathrm{SOS}>$ token. More details are provided in the supplemental material.", + "bbox": [ + 76, + 696, + 468, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Total Loss. We train our Audio Encoder end-to-end by minimizing the following loss function $\\mathcal{L}$ :", + "bbox": [ + 76, + 777, + 468, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {s e m a n t i c}} + \\mathcal {L} _ {\\text {t e m p o r a l}} + \\mathcal {L} _ {\\text {c o n d}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 816, + 468, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Generating Video Frames with Stable Diffusion", + "text_level": 1, + "bbox": [ + 76, + 842, + 468, + 857 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initial Frame Generation from Text Prompt. Our model relies on leveraging combinational operations of denoising", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1e0b278d79c4a3a16687e44316599baa4b254eb4053bb14cc99223f74d5fc4a5.jpg", + "image_caption": [ + "Figure 5: Details of our Temporal Attention Module (TAM). Temporal Attention Module effectively captures the important audio segments (e.g. harsh raining sound) as attention weights, which are guided by CLIP [27]'s text embedding (e.g. \"raining\") in training phase and express the magnitude of audio in test phase." + ], + "image_footnote": [], + "bbox": [ + 509, + 88, + 883, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "process. The estimated noise space can be either random or manual by visual input with the help of the diffusion encoder. Based on this guided diffusion, our model first generates the initial frame with a text prompt (e.g. \"a photo of beautiful beach with blue sky\"). Given this generated image as content, we manipulate its styles according to audio inputs and generate corresponding video frames, i.e. given a series of latent vectors $\\{\\mathbf{c}_1,\\mathbf{c}_2,\\dots ,\\mathbf{c}_m\\}$ , we generate $m$ video frames. Note that the number of video frames is controllable by latent vector interpolation, as we will explain later. We follow the standard image generation process with the Stable Diffusion model [29], i.e., we compute a latent vector $\\mathbf{c}_p$ in a CLIP-based embedding space given a text prompt, conditioning it to generate an image.", + "bbox": [ + 496, + 311, + 892, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Audio Semantic Guidance. We employed the SEGA [2] framework to create video frames that incorporate sound style while preserving the main content identity. We first utilize the combination of attention $\\alpha_{n}$ and audio feature $\\mathbf{s}_n$ with normalization of the scale of output feature by multiplying $N$ to produce output $\\mathbf{c}_n$ : i.e., $\\mathbf{c}_n = N^k\\alpha_n^k\\mathbf{s}_n$ , where $k$ is hyper parameter that regulates attention ratio and is set to 1. The Audio Semantic Guidance module generates the $n$ -th video frame by taking audio condition $\\mathbf{c}_n$ as input to guide the diffusion models. The denoising autoencoder $\\epsilon_{\\theta}$ is executed $\\delta - 1$ times out of $T$ time to form incomplete noise along with the original text prompt meaning. From $t = \\delta$ , the audio semantic guidance operates through the following equation:", + "bbox": [ + 496, + 527, + 892, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {p}\\right) := \\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {\\varnothing}\\right) + g \\left(\\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {p}\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {\\varnothing}\\right)\\right) \\tag {6} \\\\ + \\lambda (\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {n}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 743, + 890, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{z}^{\\delta}$ is denoised latent vector at $t = \\delta$ , $g$ is the guidance scale of the text prompt, $\\lambda(\\mathbf{z}^{\\delta}, \\mathbf{c}_n)$ is the audio semantic guidance term, and $\\mathbf{c}_{\\emptyset}$ represents an unconditioned prompt that does not make any semantic difference. As a result, only the $\\lambda(\\mathbf{z}^{\\delta}, \\mathbf{c}_n)$ term has been added to the original denoising process from $t = \\delta$ to $t = 1$ . Note that $\\mathbf{z}^T$ is fixed through frames in one video. The audio semantic guidance", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7826", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5115286525a5b4ef32eb2f277fb5190dd65786aac2908ab750ea98a6bd89249e.jpg", + "table_caption": [ + "Table 2: Comparison of the quality of generated video frames with state-of-the-art audio-to-video generations in terms of IS [31], FVD [35], and CLIP [27]-based distances." + ], + "table_footnote": [], + "table_body": "
ModelInputIS↑FVD↓CLIP↑ (a ↔ v)CLIP↑ (t ↔ v)
Sound2Sight [4]1st Frame1.02 ± 0.02494.280.03640.2164
CCVS [18]1st Frame1.30 ± 0.20679.940.12510.2360
TräumerAI [13]-1.47 ± 0.19736.320.15890.1778
Sound-guided Video Generation [19]-1.16 ± 0.16544.090.11510.1702
Ours (w/o TAM)Latent Vector1.27 ± 0.23483.760.13420.2370
OursLatent Vector1.49 ± 0.38421.230.19640.2436
", + "bbox": [ + 140, + 130, + 828, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b1f2669aa345167bf55ac69a84dbf72eb48948130c331cbab85748f39e5c4c73.jpg", + "image_caption": [ + "Figure 6: Examples of generated video frames (given fire crackling and waterfall audio) by Sound2Sight [4], CCVS [18], TraumerAI [13], Lee et al. [19], and ours. Note that Sound2Sight and CCVS use an initial frame (highlighted in a red box)." + ], + "image_footnote": [], + "bbox": [ + 84, + 271, + 491, + 592 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/512f567d03735f7645ce65a57e475ef21a3b012b9c50a3bfe6f375118ce23eed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 271, + 877, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\lambda (\\mathbf{z}^{\\delta},\\mathbf{c}_n)$ is defined as follows:", + "bbox": [ + 76, + 652, + 287, + 670 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {n}\\right) = g _ {s} \\left(\\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {n}\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {z} _ {\\varnothing} ^ {\\delta}, \\mathbf {c} _ {\\varnothing}\\right)\\right) + \\sigma_ {m} \\Phi_ {m} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 691, + 468, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\sigma_{m} \\in [0,1]$ is the momentum hyper parameter that scales the momentum $\\Phi_{m}$ and $g_{s}$ focuses on the relevant dimensions to the audio manipulation task (More details can be found in the supplemental material or SEGA [2]). Different from SEGA framework, we only consider positive guidance of audio semantics, which is the semantic difference between the guidance provided by $\\mathbf{c}_n$ and the unconditioned guidance $\\mathbf{c}_{\\emptyset}$ .", + "bbox": [ + 75, + 719, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The generation process conditioned by $\\mathbf{c}_n$ is separately working through diffusion processes so that the different semantic meaning or magnitude which presents in $\\mathbf{c}_n$ can generate frames independently.", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Temporal Frame Interpolation. We use an interpolated latent vector to generate continuous video frames between two consecutive frames. Following the work by Ramesh et al. [28], we apply a spherical linear interpolation between all consecutive pairs of $\\mathbf{c}_n$ and $\\mathbf{c}_{n + 1}$ , yielding $k$ interpolated latent vectors. These vectors are then used as a condition for the diffusion models to generate temporally-interpolated video frames.", + "bbox": [ + 496, + 654, + 893, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 792, + 633, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We use two Audio-Video datasets to train our Audio Encoder: VGG-Sound [5] and Landscape [19]. VGG-Sound is an audio dataset with about 170,000 of 10-second clips of audio-video data, which consists of 309 classes. The dataset has numerous 'in the wild' audio data that spans", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7827", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d2afda1fe476ecb35dce1039f693b49994eed5bec975826a98400f58641bfea7.jpg", + "image_caption": [ + "Figure 7: Examples of our video generation conditioned on text prompt (top) and audio (bottom)." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a large number of challenging acoustic environments and real application noise characteristics. Since audio of nature sound is the perfect tool to stylize compared to the class such as people talking or sports, we add about 9,000 audio clips of Landscape audio dataset in the training process. For obtaining test sets, class-balanced sampling is applied to Landscape dataset.", + "bbox": [ + 75, + 277, + 468, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare our methods with existing audio-to-video generation methods, Sound2Sight [4], CCVS [18], and StyleGAN [34] based TraumerAI [13] and Sound-guided Video Generation [19]. We follow experiment methods by Lee et al. [19]. All baselines are trained or finetuned with Landscape dataset [19]. Since Sound2Sight and CCVS need first frame to generate the video, we randomly select first frame from the Landscape dataset at the inference task. For TraumerAI and Sound-guided Video Generation, we first pre-train StyleGAN [16] on high fidelity benchmark datasets (LHQ datasets [33]) and then train the model to navigate the latent space by Landscape dataset. Note that randomly initialized vector for TraumerAI [13] and Sound-guided Video Generation [19] are given. For our model, we initially generate image from random noise space and randomly sample the prompt related with landscape to generate the landscape like video for fair comparison.", + "bbox": [ + 75, + 392, + 468, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use the following two video quality metrics for our evaluations: (i) Fréchet Video Distance (FVD) [35] and (ii) Inception Score (IS) [31]. FVD is used to assess video quality by measuring the distribution gap between real and synthesized videos in the latent space. Additionally, IS score is commonly employed to evaluate the effectiveness of Generative Adversarial Networks (GANs) by computing the KL-divergence between the label distribution of each image and the marginal label distribution. To implement these, we fine-tune Inflated 3D ConvNet [3] with Landscape [19] dataset for FVD and used pre-trained InceptionNet [17] that is trained on ImageNet [6] dataset. We also measure CLIP [27]-based cosine similarity (CLIP score) between audio and image as well as text and image. To obtain the textual pivot feature, we fed the following prompt \"The photo of \" into CLIP text encoder.", + "bbox": [ + 75, + 657, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c4ef6f1b0b7e2a72cee27f6ec681d429b4a2754873d2080cc0f031cadd6e3bd4.jpg", + "image_caption": [ + "Figure 8: Generated video frames with and without our Temporal Attention Module. We generate video frames with splashing water sound where its amplitude changes over time (see top)." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 885, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Experiments. For fair comparison with other existing baselines such as TraumerAI [13] and Sound-guided Video Generation [19] which does not get a hint about what to generate, we use prompt that originally does not generate the style of sound. We set fps 20 and generated videos to extract images from all baselines. Table 2 shows that our approach produces the best quality results as video. Additionally, to ensure that the generated videos are semantically related to the sound, we compare the cosine similarity between text-audio and video embedding. Our methods show a superior performance in terms of multimodal semantics.", + "bbox": [ + 496, + 357, + 892, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Video Quality Comparison. We first evaluate the quality of generated video frames. In Figure 6, we provide typical examples of generated video frames by (from top) Sound2Sight [4], CCVS [18], TraumerAI [13], Lee et al. [19], and ours. Note that Sound2Sight [4] and CCVS [18] need the initial frame as input (see red boxes). With audio inputs, such as fire crackling and waterfall, as a condition, ours generally generate temporally-consistent audio-reactive video frames. We observe that Sound2Sight [4] and CCVS [18] often show blurring artifacts, while StyleGAN-based TraumerAI [13] and Sound-guided Video Generation [19] often fail to generate audio-reactive video frames, e.g., they produce landscape scenes that are not semantically aligned with a fire crackling audio. However, ours generate a scene of a waterfall or a fire on firewood, aligning well with audio inputs.", + "bbox": [ + 496, + 545, + 892, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison to Video Generation with Text. To analyze the benefit of audio modality, we conduct a qualitative experiment to compare the effect of audio and text modalities in generating visual content. As shown in Figure 7, we first generate an initial frame with the text prompt \"A photo of deep in the sea.\" Then we generate the next frames with text \"underwater bubbling\" (top) and underwater bubbling", + "bbox": [ + 496, + 794, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7828", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc8e924b9f3c71ba8146161b9ce4b01d892f87a41249f02e5b0cd779e84419f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 90, + 467, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a31f840981aad8318e349338506c53786970c0b37a5263a395b88c00d282b49e.jpg", + "image_caption": [ + "Figure 9: Generated video frames with and without our Audio Semantic Guidance. Edited parts by Audio Semantic Guidance are highlighted with red box (see second row).", + "Figure 10: Examples of generated video frames with a sound that changes over time (e.g. bird singing $\\rightarrow$ wind noise)." + ], + "image_footnote": [], + "bbox": [ + 76, + 277, + 468, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sound (bottom). It is difficult to make temporal changes conditioned on text unless we train our model with a text-video dataset. Thus, we instead change the noise scale to make temporal changes, preserving identity. However, as shown in Figure 7 (top), it generates distortions or a linear change, but this is not the case for audio. Our model with audio generates visually-appealing video frames.", + "bbox": [ + 75, + 527, + 468, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Temporal Attention Module. We use Temporal Attention Module (TAM) to improve the representation power to encode temporal semantics better. To analyze this, we perform an ablation study with and without TAM to see its effect on video frame generation. We observe in Figure 8 that our model is indeed able to generate video frames reactive to the audio changes over time (compare the changes along with the given splashing water sound).", + "bbox": [ + 75, + 638, + 468, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Audio Semantic Guidance. We demonstrate the effect of Audio Semantic Guidance with ablation study. First, we generate video frames without Audio Semantic Guidance by replacing audio semantic guidance term $\\lambda (\\mathbf{z}^{\\delta},\\mathbf{c}_n)$ with $\\epsilon_{\\theta}(\\mathbf{z}^{\\delta},\\mathbf{c}_n)$ (refer the notation in Section 3.3). To remove the effect of Audio Semantic Guidance, we also set $\\delta$ to $T$ . As illustrated in Figure 9 (top), without Audio Semantic Guidance, the model has a tendency to produce unnecessary changes (e.g. grass $\\rightarrow$ toil) and struggles to", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/14155dfec2472ad9982bf798bb225bb34574ba0c1f2134444c4083891903dc6a.jpg", + "image_caption": [ + "Figure 11: Our human evaluation results. We conduct a user study with 100 participants on Amazon Mechanical Turk (AMT). Participants are shown generated video frames and asked to evaluate them in terms of realistic, vividness, consistency, and relevance. The Likert scale is used (higher is better)." + ], + "image_footnote": [], + "bbox": [ + 504, + 92, + 880, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "achieve consistent alignment with audio (e.g. producing a road with a waterfall burbling sound). On the other hand, by leveraging Audio Semantic Guidance as in Figure 9 (bottom), we can generate sequential video frames that not only has consistent content but also represents natural temporal variations according to audio sound (e.g. producing wave of water with waterfall burbling sound), resulting in enhanced naturalness.", + "bbox": [ + 496, + 368, + 892, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experiments of Semantic Transition. In Figure 10, we provide examples where audio inputs are changed (e.g., bird singing $\\rightarrow$ wind noise, strong wind $\\rightarrow$ raining). As we observe in that figure, our model successfully adapts to the audio change, generating video frames accordingly. This may confirm that our model is indeed conditioned on the audio sequence and can generate audio-adaptive video frames.", + "bbox": [ + 496, + 494, + 892, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User study. Further, we conduct a human evaluation to evaluate the video quality by human judges. We recruit 100 participants from Amazon MTurk. Participants are shown video frames generated by five different audio-driven video generation models: Sound2Sight [4], CCVS [18], TraumerAI [13], Lee et al. [19], and ours. Participants are asked to evaluate the given video frames in terms of realism, vividness, consistency, and relevance. We use a five-point Likert scale where ideal video frames will get all five points. More details are provided as supplemental material. We observe in Figure 11 that our proposed method outperforms the other approaches in all categories. These results are consistent with our quantitative and qualitative results.", + "bbox": [ + 496, + 607, + 892, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Conditioning. Our model is capable of producing an initial frame using visual input (such as images) with the trained diffusion encoder. Figure 12 demonstrates that our framework leverages latent space of a diffusion model with a combination of visual input, text prompt and audio. We first randomly choose visual input (e.g., sun light photo) and", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7829", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/656558e23f2214679b30af002468d997bd4c1aa89f305ed7c485db50619ac55d.jpg", + "image_caption": [ + "Figure 12: Example of generated video with visual input, text prompt and audio sound. (e.g., 1st row: sun light photo conditioned by text \"A photo of beautiful garden with sky\" and splashing water sound.)" + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 447, + 224 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/aaec0af009da9ac6088f5700107fa276d1e594124cf93120e03698bf387fbb79.jpg", + "image_caption": [ + "Figure 13: Example of generated videos with audio-text joint condition (e.g., 2nd row: conditioned with text \"eruption\" and explosion sound)" + ], + "image_footnote": [], + "bbox": [ + 81, + 316, + 460, + 554 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "generate an initial frame with the text prompt (e.g., \"A photo of beautiful garden with sky\"). Then we incorporate an audio sound (e.g., splashing water) to generate videos. Shown in Figure 12, it can produce videos featuring a garden with water where the sun is centered (1st row) or where a black mountain in the background (2nd row).", + "bbox": [ + 75, + 630, + 468, + 720 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Text-Audio Joint Conditioning. As our model is built upon the Stable Diffusion model, it is also possible to use text and audio as a condition together. In Figure 13, we provide an example where we generate video frames conditioned on a sound of an explosion along with texts, such as \"eruption\", \"spew\", or \"cloud of ash.\" (see 2nd-4th rows) Preserving temporal semantics, our model successfully generates video frames guided by text as well.", + "bbox": [ + 75, + 727, + 470, + 849 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Face Generation. Our Audio Semantic Guidance enables detailed adjustment in certain areas, such as face generation. Figure 14 shows examples of human faces conditioned on", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e107fa8c28926bf5c18ed4d9ba4200263d258b82968d636d3c95ce00f328ecb7.jpg", + "image_caption": [ + "Figure 14: Examples of face generation with our methods. We use giggling and sobbing sound to manipulate face expressions of human face." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 887, + 281 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "giggling sound (1st and 3rd rows) and sobbing sound (2nd row). Our model excels in preserving the core facial attributes while manipulating emotional expressions.", + "bbox": [ + 498, + 354, + 890, + 400 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 412, + 617, + 428 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose The Power of Sound (TPoS), a novel audio-to-video generation with Stable Diffusion. Our work extends the usage of audio modality on generation models, and broaden the methods of using Stable Diffusion by generating realistic videos by our Audio Encoder. Superior performances are achieved over widely-used audio-to-video benchmarks, reflected by objective evaluations and User study, hence attributing towards the new formulation of video generation with audio modality.", + "bbox": [ + 496, + 438, + 890, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Research Foundation of Korea grant (NRF-2021R1C1C1009608, $10\\%$ ), Basic Science Research Program (NRF-2021R1A6A1A13044830, $5\\%$ ), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(2022-0-00043, $30\\%$ ), and the MSIT(Ministry of Science and ICT), Korea, under the ICT Creative Consilience program(IITP-2023-2020-0-01819, $5\\%$ ) supervised by the IITP. W. Ryoo, S. Lee, and S. Kim are supported by the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2019-0-00079, Artificial Intelligence Graduate School Program(Korea University), $30\\%$ ), the National Research Foundation of Korea grant (NRF-2022R1F1A1074334, $15\\%$ ), Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: 4D Content Generation and Copyright Protection with Artificial Intelligence, Project Number: R2022020068, $5\\%$ ), and supported by the Google Cloud Research Credits program(code-MU6X-FKAV-N3A3-1X0T).", + "bbox": [ + 496, + 579, + 892, + 881 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7830", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint arXiv:2206.02779, 2022. 3", + "[2] Manuel Brack, Patrick Schramowski, Felix Friedrich, Dominik Hintersdorf, and Kristian Kersting. The stable artist: Steering semantics in diffusion latent space. arXiv preprint arXiv:2212.06013, 2022. 2, 3, 5, 6", + "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7", + "[4] Moitreya Chatterjee and Anoop Cherian. Sound2sight: Generating visual dynamics from sound and context. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 701-719. Springer, 2020. 2, 3, 6, 7, 8", + "[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2, 6", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 7", + "[7] Sarah Gross, Xingxing Wei, and Jun Zhu. Automatic realistic music video generation from segments of youtube videos. arXiv preprint arXiv:1905.12245, 2019. 2", + "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3", + "[10] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3", + "[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 3", + "[12] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 4", + "[13] Dasaem Jeong, Seungheon Doh, and Taegyun Kwon. Traumerai: Dreaming music with stylegan. arXiv preprint arXiv:2102.04680, 2(4):10, 2021. 2, 3, 6, 7, 8", + "[14] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2", + "[15] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "emotional video portraits. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14080-14089, 2021. 2", + "[16] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2, 7", + "[17] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 7", + "[18] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Ccvs: context-aware controllable video synthesis. Advances in Neural Information Processing Systems, 34:14042-14055, 2021. 2, 3, 6, 7, 8", + "[19] Seung Hyun Lee, Gyeongrok Oh, Wonmin Byeon, Chanyoung Kim, Won Jeong Ryoo, Sang Ho Yoon, Hyunjun Cho, Jihyun Bae, Jinkyu Kim, and Sangpil Kim. Sound-guided semantic video generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 34-50. Springer, 2022. 2, 3, 4, 6, 7, 8", + "[20] Seung Hyun Lee, Wonseok Roh, Wonmin Byeon, Sang Ho Yoon, Chanyoung Kim, Jinkyu Kim, and Sangpil Kim. Sound-guided semantic image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3377-3386, 2022. 4", + "[21] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 2", + "[22] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 423-439. Springer, 2022. 3", + "[23] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 2, 3", + "[24] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4, 5", + "[25] Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le. Specaugment: A simple data augmentation method for automatic speech recognition. arXiv preprint arXiv:1904.08779, 2019. 5", + "[26] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7831", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 5, 6, 7", + "[28] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6", + "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 5", + "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3", + "[31] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 6, 7", + "[32] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 3", + "[33] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14144-14153, 2021. 7", + "[34] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022. 3, 7", + "[35] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6, 7", + "[36] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2, 3", + "[37] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. arXiv preprint arXiv:2212.08698, 2022. 3" + ], + "bbox": [ + 78, + 90, + 470, + 864 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7832", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/3b9298a9-5799-4904-96da-bef6d287733c_model.json b/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/3b9298a9-5799-4904-96da-bef6d287733c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..51bdd10fd2fdfa7e0d55c9469c945511a3c337e9 --- /dev/null +++ b/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/3b9298a9-5799-4904-96da-bef6d287733c_model.json @@ -0,0 +1,1916 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.207, + 0.131, + 0.765, + 0.175 + ], + "angle": 0, + "content": "The Power of Sound (TPoS): Audio Reactive Video Generation with Stable Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.203, + 0.732, + 0.24 + ], + "angle": 0, + "content": "Yujin Jeong\\(^{1}\\), Wonjeong Ryoo\\(^{2}\\), Seunghyun Lee\\(^{2}\\), Dabin Seo\\(^{1}\\), Wonmin Byeon\\(^{3}\\), Sanggil Kim,\\(^{2,*}\\) and Jinkyu Kim\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.24, + 0.845, + 0.292 + ], + "angle": 0, + "content": "\\(^{1}\\) Department of Computer Science and Engineering, Korea University, Seoul 02841, Korea \n\\(^{2}\\) Department of Artificial Intelligence, Korea University, Seoul 02841, Korea \n\\(^{3}\\) NVIDIA Research, Santa Clara 95050, USA" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.294, + 0.778, + 0.31 + ], + "angle": 0, + "content": "*Correspondences: S. Kim (spk7@korea.ac.kr) and J. Kim (jinkyukim@korea.ac.kr)" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.319, + 0.845, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.627, + 0.896, + 0.749 + ], + "angle": 0, + "content": "Figure 1: The Power of Sound (TPoS) is a novel framework that generates audio-reactive video sequences. Built upon the Stable Diffusion model, our model first generates an initial frame from a user-provided text prompt (e.g. \"a photo of a beautiful beach with a blue sky\"), then reactively manipulates the style of generated images corresponding to the sound inputs (e.g. an audio sequence of fireplace). Our model is indeed able to generate a frame conditioned on semantic information of the sound (see 1st and 2nd rows where images are manipulated driven by sound inputs such as fireplace or wave sound), while realistically dealing with temporal visual changes conditioned on changes of sound, e.g., increasing magnitude of sounds (see second row) or wave \\(\\rightarrow\\) fireplace (see last row). TPoS creates visually compelling and contextually relevant video sequences in an open domain." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.761, + 0.314, + 0.777 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In recent years, video generation has become a prominent generative tool and has drawn significant attention. However, there is little consideration in audio-to-video generation, though audio contains unique qualities like temporal semantics and magnitude. Hence, we propose The Power of Sound (TPoS) model to incorporate audio input that includes both changeable temporal semantics and mag-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.895, + 0.884 + ], + "angle": 0, + "content": "nitude. To generate video frames, TPoS utilizes a latent stable diffusion model with textual semantic information, which is then guided by the sequential audio embedding from our pretrained Audio Encoder. As a result, this method produces audio reactive video contents. We demonstrate the effectiveness of TPoS across various tasks and compare its results with current state-of-the-art techniques in the field of audio-to-video generation. More examples are available" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7822" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.078, + 0.092, + 0.396, + 0.106 + ], + "angle": 0, + "content": "at https://ku-vai.github.io/TPoS/" + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.136, + 0.21, + 0.151 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.161, + 0.47, + 0.312 + ], + "angle": 0, + "content": "Recent generative models have demonstrated the potential to generate visuallyappealing video frames [32, 11, 10, 23, 36]. They often use a simple text prompt (e.g., \"a video of a person on the street on a rainy day\") to generate a video which is intuitive for end-users to drive a video generation. Text can effectively convey uni-modal object-wise guidance, such as \"a rainy day\" or \"a person\", but it may be challenging if users want to drive it into more complex sequential procedures, i.e., \"a video of a person on the street on a rainy day, but a rain suddenly stops, and a wind blows.\"" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.313, + 0.471, + 0.508 + ], + "angle": 0, + "content": "In this paper, we leverage sounds to guide the video generation models, i.e., sound-driven video generation. Audio is another modality that can complement texts by effectively providing sequential information (or temporal semantics): e.g., a continuous transition from the sound of light rain to the sound of heavy rain. There have been introduced sound-guided video generation approaches. However, existing sound-guided video generation approaches are limited to specific applications, such as face generation [15, 26, 21, 14, 21], where audio is used to provide a script for the avatar, or other simple synthetic motions (e.g., a video of musicians playing violin or a video of painting motions by an artist) [7, 13, 18, 4]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.509, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Recently, Lee et al. [19] introduced a sound-guided landscape video generation model, leveraging the latent space of StyleGAN [16]. They focus on using audio only for semantic labels (i.e., a sound of the wind is simply encoded into a meaning of wind) but not temporal semantics – i.e. semantic information that changes over time. Thus, in this work, we focus on leveraging temporal semantics from audio inputs such that our video generator reactively manipulates video frames. Our model temporally aligns the latent space with the given audio sequence (e.g., continuous changes in audio, e.g., weak rain \\(\\rightarrow\\) heavy rain, are reflected to generate corresponding video frames)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.69, + 0.468, + 0.856 + ], + "angle": 0, + "content": "Our work starts with Stable Diffusion [29], a text-driven image generator with advantages in generating high-resolution images based on the latent diffusion models. Its architectural advantages (i.e., attention mechanism and diffusion process) help leverage audio as a driving condition, generating temporally reactive and consistent video frames. Given the latent space of trained Stable Diffusion, we generate video frames temporally guided by audio sequences with regularizers to ensure temporal consistency (between generated consecutive frames) and correspondence with audio inputs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our model consists of two main modules: (i) Audio Encoder, which is designed to encode temporal semantics of audio sequences, producing a sequence of the latent vec" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.091, + 0.688, + 0.1 + ], + "angle": 0, + "content": "Prompt: \"A photo of a person on the street\"" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.1, + 0.89, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.196, + 0.892, + 0.302 + ], + "angle": 0, + "content": "Figure 2: Limitation of text-driven image manipulation. Given a generated image by Stable Diffusion [29] with a text prompt, \"A photo of a person on the street,\" additional textual conditions of different semantic meanings (i.e. \"a little of rain\", \"rain\", and \"rain a lot\") produce similar images, failing to capture differences. Note that we apply SEGA [2]'s guidance to preserve content identity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.892, + 0.41 + ], + "angle": 0, + "content": "tors. (ii) Audio Semantic Guidance Module, which uses the above-mentioned latent vectors as a condition in the diffusion process to generate corresponding image outputs. We apply identity regularizer to produce temporally consistent video frames, while we apply audio semantic guidance to generate audio-reactive video frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.893, + 0.576 + ], + "angle": 0, + "content": "However, we observe that generating multifarious high-quality images solely from a sound input is challenging due to the lack of such a large-scale dataset to train a model. Instead, we first generate an initial frame using pre-trained Stable Diffusion model with a text prompt, then generate the following video frames conditioned on audio inputs. This frees a model from data dependency burdens and enables training with the current relatively small (than imagetext modality) audio dataset [19, 5], focusing on leveraging temporal semantics from audio. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.579, + 0.891, + 0.623 + ], + "angle": 0, + "content": "- We propose a novel sound-driven video generation method built upon Stable Diffusion [29] and can generate video frames reactively with audio sequence inputs." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.627, + 0.89, + 0.686 + ], + "angle": 0, + "content": "- Our attention-based Audio Encoder produces temporally-aware latent vectors, which are consumed by Stable Diffusion as a per-time manipulation condition, producing audio-reactive video frames." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.69, + 0.89, + 0.734 + ], + "angle": 0, + "content": "- Our model regularizes the latent features of diffusion models to produce temporally consistent video frames, preserving identity throughout the generated video." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.738, + 0.891, + 0.811 + ], + "angle": 0, + "content": "- We demonstrate the effectiveness of our proposed model using a public dataset Landscape [19], generally outperforming other state-of-the-art sound-driven video generation approaches in terms of video quality metrics and human evaluation." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.579, + 0.891, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.825, + 0.642, + 0.84 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Latent Diffusion Models. Recent success [29] suggests that the Latent Diffusion Models (LDM) improve the efficiency of the diffusion process, successfully generating" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7823" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.151 + ], + "angle": 0, + "content": "Table 1: Comparison between existing state-of-the-art audio-driven video generation approaches in terms of whether they consider the following factors: temporal semantics, magnitude changes of sound, and target domains." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.158, + 0.47, + 0.261 + ], + "angle": 0, + "content": "
ModelTemporal SemanticsMagnitudeDomains (Audio Type)
Sound2Sight [4]-✓Closed
CCVS [18]-✓Closed (Music)
TräumerAI [13]-✓Closed (Music)
Lee et al. [19]✓-Closed (Nature)
Ours✓✓Open Domains
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.27, + 0.47, + 0.482 + ], + "angle": 0, + "content": "high-quality images given a text prompt. One challenge in LDM is that the generation process is too sensitive to the condition, making it difficult to control semantics. Recently, there have been introduced to control semantics with LDM by a semantic mask [1] or by utilizing semantic information of the cross-attention layers [9]. Wu et al. [37] used linear combinations of text embeddings and Liu et al. [22] proposed composable diffusion models, but they still remained challenging to control fine-grained semantic changes. Recently, Semantic Guidance (SEGA) [2] computed a guidance vector in the latent space, enabling semantic control of diffusion models without further inputs. Inspired by SEGA, we also control the semantics in the latent space with temporally-encoded audio vector sequences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.699 + ], + "angle": 0, + "content": "Text-driven Video Generation. Recent text-to-video generation tools, including Make-A-Video [32], Video Diffusion Models [11], Imagen video [10], and Phenaki [36] have shown promising performance in generating videos from textual descriptions. However, text-to-video generation has its limitations in terms of temporal coherence, which mostly leads to short video duration or a linear video change. Recent text-to-video generation methods, StyleGAN-V [34] and Dreamix [23], made progress addressing these issues. However, conditioning temporal semantics or complex scenarios is still challenging to be obtained from text inputs. Thus, in this paper, we want to explore conditioning a model with audio inputs, which inherently convey such temporal semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Audio-driven Video Generation. Leveraging temporal semantics was not seriously considered in previous audio-driven video generation approaches. Sound2Sight[4] and CCVS [18] generate video frames conditioned on the (nontemporal) context of the given audio, while TraumerAI [13] utilized the magnitude of the given audio. Recently, Lee et al. [19] explored a model that can consider audio semantics as a condition to drive a video generator. Also, their dependency on StyleGAN [34]-based embedding space makes it difficult for models to generate transitions in video. In this work, we focus on leveraging temporal semantics from audio inputs such that our generator reactively manipulates video frames." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.591, + 0.107 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.895, + 0.299 + ], + "angle": 0, + "content": "In this work, we propose a novel audio-driven video generation method, which generates video frames conditioned on audio sequences, ensuring temporal consistency between consecutive frames and temporal correspondence with audio inputs. As shown in Figure 3, our model consists of two main parts: (i) Audio Encoder, which encodes temporal semantics of audio sequences, producing a sequence of the latent vectors (Section 3.2). (ii) Audio Semantic Guidance Module, which uses the above-mentioned latent vectors as a condition in the diffusion process to generate corresponding image outputs, which are temporally consistent (by our identity regularizer) and audio-reactive (see Section 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.308, + 0.822, + 0.324 + ], + "angle": 0, + "content": "3.1. Preliminary: Latent Stable Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.483 + ], + "angle": 0, + "content": "Latent Diffusion Models (LDMs) [29] are the method that uses an encoder to convert a noised latent vector \\(\\mathbf{z}^T\\) to a denoised latent vector \\(\\mathbf{z} = \\mathbf{x} + \\epsilon\\), where \\(\\mathbf{z}\\) is a latent vector of an input image \\(\\mathbf{x}\\) and \\(\\epsilon\\) is a noise. Stable Diffusion [29] is part of a conditional generation model that can synthesize an image given a condition \\(\\mathbf{y}\\). It uses U-Net [30] as denoising autoencoders represented by \\(\\epsilon_{\\theta}\\). To generate an output image, the autoencoder \\(\\epsilon_{\\theta}(\\mathbf{z}^{t}, t, \\tau_{\\theta}(\\mathbf{y}))\\) takes three inputs: the noised latent vector \\(\\mathbf{z}^t\\), a sequence \\(t\\) that is uniformly sampled from the set \\(1, \\dots, T\\), and a conditional input \\(\\mathbf{y}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.635 + ], + "angle": 0, + "content": "Conditional input \\(\\mathbf{y}\\) is first transformed into a latent vector \\(\\mathbf{c}_p\\) through a pretrained function \\(\\tau_{\\theta}\\) and then it is fed into a cross-attention layer of the U-Net as the key and value, which is then combined with \\(\\mathbf{z}^t\\) by an attention mechanism, where the query is the flattened intermediate representation of the U-Net. Next, the denoising autoencoder \\(\\epsilon_{\\theta}\\) is used to denoise \\(\\mathbf{z}^t\\) from \\(t = T\\) to \\(t = 1\\) sequentially. After \\(T\\) denoising steps, the resulting denoised latent vector \\(\\mathbf{z}(= \\mathbf{z}^{1})\\) is transmitted to the decoder, which produces the final output image \\(\\tilde{\\mathbf{x}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.645, + 0.864, + 0.663 + ], + "angle": 0, + "content": "3.2. Encoding Temporal Semantics from Audio" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.775 + ], + "angle": 0, + "content": "We use an audio modality as a source of generating temporal conditions. We divide the audio mel-spectogram into uniform snippets (segments) to extract abundant audio features and feed this into the Audio Encoder to extract both temporal semantic features and intensity of audio in end-to-end manner. The figure of our training process is illustrated in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Audio Feature Extraction. Audio inputs are first transformed into a mel-spectrogram representation, denoted as \\(\\mathbf{x}^a\\in \\mathbb{R}^{d\\times w}\\), where \\(d\\) represents the number of mel-frequency bins and \\(w\\) is the width of the spectrogram. To incorporate time information, the mel-spectrogram is divided into \\(N\\) segments. Each segment, denoted as \\(\\mathbf{x}_n^a\\in \\mathbb{R}^{d\\times \\lceil \\frac{w}{N}\\rceil}\\) where \\(n\\in \\{1,\\ldots ,N\\}\\), is then fed into a shared feature extraction module, i.e., the pre-trained ResNet18 [8]. The" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7824" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.092, + 0.852, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.392, + 0.895, + 0.456 + ], + "angle": 0, + "content": "Figure 3: An overview of our proposed TPoS model. Our model consists of two main modules: (i) Audio Encoder, which produces a sequence of latent vectors, encoding temporal semantics of audio input by utilizing CLIP [27] space and highlighting the important temporal features and (ii) Audio Semantic Guidance Module, which is based on the diffusion process, generating video frames that are temporally consistent and audio-reactive." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.471, + 0.515 + ], + "angle": 0, + "content": "feature extraction module \\( f_{a}(\\cdot) \\) learns to extract low-level features from each audio segment regardless of its time dependency, i.e., \\( \\mathbf{w}_n = f_a(\\mathbf{x}_n^a) \\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.716 + ], + "angle": 0, + "content": "LSTM-based Temporal Semantic Encoder. As our goal is to generate audio-reactive video frames, it is also important to encode temporal changes (or relations) of the given audio inputs. Similar to Lee et al. [19], given audio features \\(\\mathbf{w} \\in \\{\\mathbf{w}_1, \\mathbf{w}_2, \\dots, \\mathbf{w}_N\\}\\), which encodes per-segment disjoint audio representation, we apply the standard Long Short-Term Memory (LSTM) network [12] to encode temporal relations or changes between consecutive audio features \\(\\mathbf{w}\\). Formally, our LSTM takes the audio feature \\(\\mathbf{w}_{n-1}\\) as input and updates its hidden state \\(\\mathbf{h}_n\\), producing an output \\(\\mathbf{s}_t\\): i.e. \\((\\mathbf{s}_n, \\mathbf{h}_n) = \\mathrm{LSTM}(\\mathbf{h}_{n-1}, \\mathbf{w}_{n-1})\\). These outputs \\(\\mathbf{s} \\in \\{\\mathbf{s}_1, \\mathbf{s}_2, \\dots, \\mathbf{s}_N\\}\\) are then fed into Temporal Attention Module (TAM) to further encode temporal semantics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.721, + 0.471, + 0.873 + ], + "angle": 0, + "content": "Aligning Audio Semantics with Image-Text CLIP Joint Space. As we will use the output \\( s \\) as a condition to manipulate video frames, it is important to ensure those audio features are well-aligned with other text and visual features in the CLIP [27]-based joint embedding space. Similar to Lee et al. [20], given the pre-trained image-text CLIP space, we apply the following loss \\( \\mathcal{L}_{\\mathrm{CLIP}}^{a \\leftrightarrow t} \\) with the InfoNCE loss [24] \\( l_{sim} \\) such that positive pairs (e.g. an audio of raining and a text prompt \"raining\") are pulled close to each other, while negative pairs are pushed farther away." + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.884, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C L I P}} ^ {a \\leftrightarrow t} = l _ {s i m} (\\mathbf {s} _ {N}, \\operatorname {C L I P} _ {t} (\\mathbf {t})) + l _ {s i m} (\\operatorname {C L I P} _ {t} (\\mathbf {t}), \\mathbf {s} _ {N}) \\quad (1)\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.466, + 0.887, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Figure 4: An overview of our Audio Encoder training process. Our model generates temporally-encoded audio embeddings with an LSTM [12] layer and Temporal Attention Module (TAM). Audio input is partitioned into \\( N \\) segments, and each of these is encoded and used as a condition to manipulate audio-reactive video sequences (e.g. light rain → heavy rain). This is done by our Mapping Module (MAP), which maps the audio embedding to the latent space of Stable Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathrm{CLIP}_t\\) is a pre-trained CLIP-based text encoder, which takes a text prompt \\(\\mathbf{t}\\) obtained from audio class labels as input, yielding an \\(d\\)-dimensional feature. Note that" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7825" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "we only apply this loss for the final output \\(\\mathbf{s}_N\\) for efficient training. Given a set of positive pairs, we apply the following InfoNCE loss \\(l_{sim}(\\mathbf{a},\\mathbf{b})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.145, + 0.47, + 0.181 + ], + "angle": 0, + "content": "\\[\nl _ {s i m} = - \\log \\frac {\\exp \\left(\\left\\langle \\mathbf {a} _ {i} , \\mathbf {b} _ {i} \\right\\rangle\\right) / \\tau}{\\sum_ {j} \\exp \\left(\\left\\langle \\mathbf {a} _ {i} , \\mathbf {b} _ {j} \\right\\rangle\\right) / \\tau} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.188, + 0.469, + 0.218 + ], + "angle": 0, + "content": "where \\(\\langle \\mathbf{a}_i,\\mathbf{b}_i\\rangle\\) represents the cosine similarity with temperature \\(\\tau\\). Note that we set \\(\\tau\\) to 0.07." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.223, + 0.469, + 0.346 + ], + "angle": 0, + "content": "Augmenting Audio Semantics. Audio data is often limited in volume and diversity; thus, augmentation techniques may be required to extract better-quality audio semantic features, preventing a representation collapse. We use SpecAugment [25] to apply random transformations (such as masking our certain frequency bands or time segments), yielding augmented audio inputs. We further add the InfoNCE loss [24] \\(\\mathcal{L}_{\\mathrm{CLIP}}^{a\\leftrightarrow a^{\\prime}}\\) to pull augmented audio features together." + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.353, + 0.469, + 0.373 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C L I P}} ^ {a \\leftrightarrow a ^ {\\prime}} = l _ {s i m} \\left(\\mathbf {s} _ {N}, \\mathbf {s} _ {N} ^ {\\prime}\\right) + l _ {s i m} \\left(\\mathbf {s} _ {N} ^ {\\prime}, \\mathbf {s} _ {N}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.382, + 0.469, + 0.428 + ], + "angle": 0, + "content": "where apostrophe indicates augmented view of an original audio data. Finally, we use the semantic loss \\(\\mathcal{L}_{\\mathrm{semantic}}\\): i.e. \\(\\mathcal{L}_{\\mathrm{semantic}} = \\mathcal{L}_{\\mathrm{CLIP}}^{a\\leftrightarrow t} + \\lambda_s\\mathcal{L}_{\\mathrm{CLIP}}^{a\\leftrightarrow a'}\\) where \\(\\lambda_{s}\\) is set to 0.6." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.469, + 0.658 + ], + "angle": 0, + "content": "Temporal Attention Module (TAM). As shown in Figure 5, we further use an attention-based module to encode temporal semantics from the audio inputs. We empirically observe that adding this module helps improve the quality of video frame generation, which is probably due to the fact that the model becomes flexible to focus on more important temporal information, improving the model's representation power. Formally, we first compute attention weight \\(\\alpha_{n}\\) for a given audio feature \\(\\mathbf{s}_n\\) by applying an MLP layer \\(f_{\\mathrm{proj}}\\) followed by a softmax operation: i.e. \\(\\alpha_{n} = \\exp (f_{\\mathrm{proj}}(\\mathbf{s}_{n})) / \\sum_{n}\\exp (f_{\\mathrm{proj}}(\\mathbf{s}_{n}))\\) such that \\(\\sum_{n}\\alpha_{n} = 1\\). We compute the weighted sum of audio features based on attention weights, yielding an attended audio feature \\(\\mathbf{o}^a = \\sum_n\\alpha_n\\mathbf{s}_n\\). We add another InfoNCE loss \\(\\mathcal{L}_{\\mathrm{temporal}}\\) to align the audio features with text:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.667, + 0.469, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t e m p o r a l}} = l _ {s i m} \\left(\\mathbf {o} ^ {a}, \\operatorname {C L I P} _ {t} (\\mathbf {t})\\right) + l _ {s i m} \\left(\\operatorname {C L I P} _ {t} (\\mathbf {t}), \\mathbf {o} ^ {a}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.698, + 0.469, + 0.773 + ], + "angle": 0, + "content": "Lastly, we also minimize the MSE loss between text embeddings (before the projection layer) and the projected audio feature \\(\\mathrm{MAP}(\\mathbf{o}^a)\\): \\(\\mathcal{L}_{\\mathrm{cond}} = ||\\mathbf{c}_t - \\mathrm{MAP}(\\mathbf{o}^a)||_2^2\\). Note that we exclude the feature for the \\(<\\mathrm{SOS}>\\) token. More details are provided in the supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.778, + 0.469, + 0.809 + ], + "angle": 0, + "content": "Total Loss. We train our Audio Encoder end-to-end by minimizing the following loss function \\(\\mathcal{L}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.818, + 0.469, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {s e m a n t i c}} + \\mathcal {L} _ {\\text {t e m p o r a l}} + \\mathcal {L} _ {\\text {c o n d}} \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.843, + 0.469, + 0.858 + ], + "angle": 0, + "content": "3.3. Generating Video Frames with Stable Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Initial Frame Generation from Text Prompt. Our model relies on leveraging combinational operations of denoising" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.089, + 0.885, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.208, + 0.892, + 0.299 + ], + "angle": 0, + "content": "Figure 5: Details of our Temporal Attention Module (TAM). Temporal Attention Module effectively captures the important audio segments (e.g. harsh raining sound) as attention weights, which are guided by CLIP [27]'s text embedding (e.g. \"raining\") in training phase and express the magnitude of audio in test phase." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.893, + 0.525 + ], + "angle": 0, + "content": "process. The estimated noise space can be either random or manual by visual input with the help of the diffusion encoder. Based on this guided diffusion, our model first generates the initial frame with a text prompt (e.g. \"a photo of beautiful beach with blue sky\"). Given this generated image as content, we manipulate its styles according to audio inputs and generate corresponding video frames, i.e. given a series of latent vectors \\(\\{\\mathbf{c}_1,\\mathbf{c}_2,\\dots ,\\mathbf{c}_m\\}\\), we generate \\(m\\) video frames. Note that the number of video frames is controllable by latent vector interpolation, as we will explain later. We follow the standard image generation process with the Stable Diffusion model [29], i.e., we compute a latent vector \\(\\mathbf{c}_p\\) in a CLIP-based embedding space given a text prompt, conditioning it to generate an image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.893, + 0.74 + ], + "angle": 0, + "content": "Audio Semantic Guidance. We employed the SEGA [2] framework to create video frames that incorporate sound style while preserving the main content identity. We first utilize the combination of attention \\(\\alpha_{n}\\) and audio feature \\(\\mathbf{s}_n\\) with normalization of the scale of output feature by multiplying \\(N\\) to produce output \\(\\mathbf{c}_n\\): i.e., \\(\\mathbf{c}_n = N^k\\alpha_n^k\\mathbf{s}_n\\), where \\(k\\) is hyper parameter that regulates attention ratio and is set to 1. The Audio Semantic Guidance module generates the \\(n\\)-th video frame by taking audio condition \\(\\mathbf{c}_n\\) as input to guide the diffusion models. The denoising autoencoder \\(\\epsilon_{\\theta}\\) is executed \\(\\delta - 1\\) times out of \\(T\\) time to form incomplete noise along with the original text prompt meaning. From \\(t = \\delta\\), the audio semantic guidance operates through the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.744, + 0.891, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {p}\\right) := \\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {\\varnothing}\\right) + g \\left(\\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {p}\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {\\varnothing}\\right)\\right) \\tag {6} \\\\ + \\lambda (\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {n}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}^{\\delta}\\) is denoised latent vector at \\(t = \\delta\\), \\(g\\) is the guidance scale of the text prompt, \\(\\lambda(\\mathbf{z}^{\\delta}, \\mathbf{c}_n)\\) is the audio semantic guidance term, and \\(\\mathbf{c}_{\\emptyset}\\) represents an unconditioned prompt that does not make any semantic difference. As a result, only the \\(\\lambda(\\mathbf{z}^{\\delta}, \\mathbf{c}_n)\\) term has been added to the original denoising process from \\(t = \\delta\\) to \\(t = 1\\). Note that \\(\\mathbf{z}^T\\) is fixed through frames in one video. The audio semantic guidance" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7826" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.121 + ], + "angle": 0, + "content": "Table 2: Comparison of the quality of generated video frames with state-of-the-art audio-to-video generations in terms of IS [31], FVD [35], and CLIP [27]-based distances." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.131, + 0.83, + 0.269 + ], + "angle": 0, + "content": "
ModelInputIS↑FVD↓CLIP↑ (a ↔ v)CLIP↑ (t ↔ v)
Sound2Sight [4]1st Frame1.02 ± 0.02494.280.03640.2164
CCVS [18]1st Frame1.30 ± 0.20679.940.12510.2360
TräumerAI [13]-1.47 ± 0.19736.320.15890.1778
Sound-guided Video Generation [19]-1.16 ± 0.16544.090.11510.1702
Ours (w/o TAM)Latent Vector1.27 ± 0.23483.760.13420.2370
OursLatent Vector1.49 ± 0.38421.230.19640.2436
" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.272, + 0.493, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.272, + 0.878, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.603, + 0.893, + 0.637 + ], + "angle": 0, + "content": "Figure 6: Examples of generated video frames (given fire crackling and waterfall audio) by Sound2Sight [4], CCVS [18], TraumerAI [13], Lee et al. [19], and ours. Note that Sound2Sight and CCVS use an initial frame (highlighted in a red box)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.654, + 0.289, + 0.671 + ], + "angle": 0, + "content": "\\(\\lambda (\\mathbf{z}^{\\delta},\\mathbf{c}_n)\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.693, + 0.47, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\lambda \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {n}\\right) = g _ {s} \\left(\\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {n}\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {z} _ {\\varnothing} ^ {\\delta}, \\mathbf {c} _ {\\varnothing}\\right)\\right) + \\sigma_ {m} \\Phi_ {m} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.84 + ], + "angle": 0, + "content": "where \\(\\sigma_{m} \\in [0,1]\\) is the momentum hyper parameter that scales the momentum \\(\\Phi_{m}\\) and \\(g_{s}\\) focuses on the relevant dimensions to the audio manipulation task (More details can be found in the supplemental material or SEGA [2]). Different from SEGA framework, we only consider positive guidance of audio semantics, which is the semantic difference between the guidance provided by \\(\\mathbf{c}_n\\) and the unconditioned guidance \\(\\mathbf{c}_{\\emptyset}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The generation process conditioned by \\(\\mathbf{c}_n\\) is separately working through diffusion processes so that the different semantic meaning or magnitude which presents in \\(\\mathbf{c}_n\\) can generate frames independently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.895, + 0.777 + ], + "angle": 0, + "content": "Temporal Frame Interpolation. We use an interpolated latent vector to generate continuous video frames between two consecutive frames. Following the work by Ramesh et al. [28], we apply a spherical linear interpolation between all consecutive pairs of \\(\\mathbf{c}_n\\) and \\(\\mathbf{c}_{n + 1}\\), yielding \\(k\\) interpolated latent vectors. These vectors are then used as a condition for the diffusion models to generate temporally-interpolated video frames." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.793, + 0.634, + 0.811 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Datasets. We use two Audio-Video datasets to train our Audio Encoder: VGG-Sound [5] and Landscape [19]. VGG-Sound is an audio dataset with about 170,000 of 10-second clips of audio-video data, which consists of 309 classes. The dataset has numerous 'in the wild' audio data that spans" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7827" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.225, + 0.47, + 0.256 + ], + "angle": 0, + "content": "Figure 7: Examples of our video generation conditioned on text prompt (top) and audio (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.47, + 0.384 + ], + "angle": 0, + "content": "a large number of challenging acoustic environments and real application noise characteristics. Since audio of nature sound is the perfect tool to stylize compared to the class such as people talking or sports, we add about 9,000 audio clips of Landscape audio dataset in the training process. For obtaining test sets, class-balanced sampling is applied to Landscape dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.47, + 0.651 + ], + "angle": 0, + "content": "Baselines. We compare our methods with existing audio-to-video generation methods, Sound2Sight [4], CCVS [18], and StyleGAN [34] based TraumerAI [13] and Sound-guided Video Generation [19]. We follow experiment methods by Lee et al. [19]. All baselines are trained or finetuned with Landscape dataset [19]. Since Sound2Sight and CCVS need first frame to generate the video, we randomly select first frame from the Landscape dataset at the inference task. For TraumerAI and Sound-guided Video Generation, we first pre-train StyleGAN [16] on high fidelity benchmark datasets (LHQ datasets [33]) and then train the model to navigate the latent space by Landscape dataset. Note that randomly initialized vector for TraumerAI [13] and Sound-guided Video Generation [19] are given. For our model, we initially generate image from random noise space and randomly sample the prompt related with landscape to generate the landscape like video for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. We use the following two video quality metrics for our evaluations: (i) Fréchet Video Distance (FVD) [35] and (ii) Inception Score (IS) [31]. FVD is used to assess video quality by measuring the distribution gap between real and synthesized videos in the latent space. Additionally, IS score is commonly employed to evaluate the effectiveness of Generative Adversarial Networks (GANs) by computing the KL-divergence between the label distribution of each image and the marginal label distribution. To implement these, we fine-tune Inflated 3D ConvNet [3] with Landscape [19] dataset for FVD and used pre-trained InceptionNet [17] that is trained on ImageNet [6] dataset. We also measure CLIP [27]-based cosine similarity (CLIP score) between audio and image as well as text and image. To obtain the textual pivot feature, we fed the following prompt \"The photo of \" into CLIP text encoder." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.09, + 0.887, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Figure 8: Generated video frames with and without our Temporal Attention Module. We generate video frames with splashing water sound where its amplitude changes over time (see top)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.893, + 0.538 + ], + "angle": 0, + "content": "Quantitative Experiments. For fair comparison with other existing baselines such as TraumerAI [13] and Sound-guided Video Generation [19] which does not get a hint about what to generate, we use prompt that originally does not generate the style of sound. We set fps 20 and generated videos to extract images from all baselines. Table 2 shows that our approach produces the best quality results as video. Additionally, to ensure that the generated videos are semantically related to the sound, we compare the cosine similarity between text-audio and video embedding. Our methods show a superior performance in terms of multimodal semantics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.893, + 0.789 + ], + "angle": 0, + "content": "Qualitative Video Quality Comparison. We first evaluate the quality of generated video frames. In Figure 6, we provide typical examples of generated video frames by (from top) Sound2Sight [4], CCVS [18], TraumerAI [13], Lee et al. [19], and ours. Note that Sound2Sight [4] and CCVS [18] need the initial frame as input (see red boxes). With audio inputs, such as fire crackling and waterfall, as a condition, ours generally generate temporally-consistent audio-reactive video frames. We observe that Sound2Sight [4] and CCVS [18] often show blurring artifacts, while StyleGAN-based TraumerAI [13] and Sound-guided Video Generation [19] often fail to generate audio-reactive video frames, e.g., they produce landscape scenes that are not semantically aligned with a fire crackling audio. However, ours generate a scene of a waterfall or a fire on firewood, aligning well with audio inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Comparison to Video Generation with Text. To analyze the benefit of audio modality, we conduct a qualitative experiment to compare the effect of audio and text modalities in generating visual content. As shown in Figure 7, we first generate an initial frame with the text prompt \"A photo of deep in the sea.\" Then we generate the next frames with text \"underwater bubbling\" (top) and underwater bubbling" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7828" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.091, + 0.468, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.229, + 0.47, + 0.274 + ], + "angle": 0, + "content": "Figure 9: Generated video frames with and without our Audio Semantic Guidance. Edited parts by Audio Semantic Guidance are highlighted with red box (see second row)." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.279, + 0.47, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.466, + 0.47, + 0.512 + ], + "angle": 0, + "content": "Figure 10: Examples of generated video frames with a sound that changes over time (e.g. bird singing \\(\\rightarrow\\) wind noise)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.47, + 0.634 + ], + "angle": 0, + "content": "sound (bottom). It is difficult to make temporal changes conditioned on text unless we train our model with a text-video dataset. Thus, we instead change the noise scale to make temporal changes, preserving identity. However, as shown in Figure 7 (top), it generates distortions or a linear change, but this is not the case for audio. Our model with audio generates visually-appealing video frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.47, + 0.76 + ], + "angle": 0, + "content": "Effect of Temporal Attention Module. We use Temporal Attention Module (TAM) to improve the representation power to encode temporal semantics better. To analyze this, we perform an ablation study with and without TAM to see its effect on video frame generation. We observe in Figure 8 that our model is indeed able to generate video frames reactive to the audio changes over time (compare the changes along with the given splashing water sound)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Effect of Audio Semantic Guidance. We demonstrate the effect of Audio Semantic Guidance with ablation study. First, we generate video frames without Audio Semantic Guidance by replacing audio semantic guidance term \\(\\lambda (\\mathbf{z}^{\\delta},\\mathbf{c}_n)\\) with \\(\\epsilon_{\\theta}(\\mathbf{z}^{\\delta},\\mathbf{c}_n)\\) (refer the notation in Section 3.3). To remove the effect of Audio Semantic Guidance, we also set \\(\\delta\\) to \\(T\\). As illustrated in Figure 9 (top), without Audio Semantic Guidance, the model has a tendency to produce unnecessary changes (e.g. grass \\(\\rightarrow\\) toil) and struggles to" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.093, + 0.882, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.255, + 0.894, + 0.346 + ], + "angle": 0, + "content": "Figure 11: Our human evaluation results. We conduct a user study with 100 participants on Amazon Mechanical Turk (AMT). Participants are shown generated video frames and asked to evaluate them in terms of realistic, vividness, consistency, and relevance. The Likert scale is used (higher is better)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.893, + 0.49 + ], + "angle": 0, + "content": "achieve consistent alignment with audio (e.g. producing a road with a waterfall burbling sound). On the other hand, by leveraging Audio Semantic Guidance as in Figure 9 (bottom), we can generate sequential video frames that not only has consistent content but also represents natural temporal variations according to audio sound (e.g. producing wave of water with waterfall burbling sound), resulting in enhanced naturalness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.893, + 0.603 + ], + "angle": 0, + "content": "Experiments of Semantic Transition. In Figure 10, we provide examples where audio inputs are changed (e.g., bird singing \\(\\rightarrow\\) wind noise, strong wind \\(\\rightarrow\\) raining). As we observe in that figure, our model successfully adapts to the audio change, generating video frames accordingly. This may confirm that our model is indeed conditioned on the audio sequence and can generate audio-adaptive video frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.893, + 0.804 + ], + "angle": 0, + "content": "User study. Further, we conduct a human evaluation to evaluate the video quality by human judges. We recruit 100 participants from Amazon MTurk. Participants are shown video frames generated by five different audio-driven video generation models: Sound2Sight [4], CCVS [18], TraumerAI [13], Lee et al. [19], and ours. Participants are asked to evaluate the given video frames in terms of realism, vividness, consistency, and relevance. We use a five-point Likert scale where ideal video frames will get all five points. More details are provided as supplemental material. We observe in Figure 11 that our proposed method outperforms the other approaches in all categories. These results are consistent with our quantitative and qualitative results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Visual Conditioning. Our model is capable of producing an initial frame using visual input (such as images) with the trained diffusion encoder. Figure 12 demonstrates that our framework leverages latent space of a diffusion model with a combination of visual input, text prompt and audio. We first randomly choose visual input (e.g., sun light photo) and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7829" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.449, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.294 + ], + "angle": 0, + "content": "Figure 12: Example of generated video with visual input, text prompt and audio sound. (e.g., 1st row: sun light photo conditioned by text \"A photo of beautiful garden with sky\" and splashing water sound.)" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.318, + 0.462, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.613 + ], + "angle": 0, + "content": "Figure 13: Example of generated videos with audio-text joint condition (e.g., 2nd row: conditioned with text \"eruption\" and explosion sound)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.47, + 0.722 + ], + "angle": 0, + "content": "generate an initial frame with the text prompt (e.g., \"A photo of beautiful garden with sky\"). Then we incorporate an audio sound (e.g., splashing water) to generate videos. Shown in Figure 12, it can produce videos featuring a garden with water where the sun is centered (1st row) or where a black mountain in the background (2nd row)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.728, + 0.471, + 0.85 + ], + "angle": 0, + "content": "Text-Audio Joint Conditioning. As our model is built upon the Stable Diffusion model, it is also possible to use text and audio as a condition together. In Figure 13, we provide an example where we generate video frames conditioned on a sound of an explosion along with texts, such as \"eruption\", \"spew\", or \"cloud of ash.\" (see 2nd-4th rows) Preserving temporal semantics, our model successfully generates video frames guided by text as well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Face Generation. Our Audio Semantic Guidance enables detailed adjustment in certain areas, such as face generation. Figure 14 shows examples of human faces conditioned on" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.888, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.295, + 0.892, + 0.341 + ], + "angle": 0, + "content": "Figure 14: Examples of face generation with our methods. We use giggling and sobbing sound to manipulate face expressions of human face." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.355, + 0.892, + 0.401 + ], + "angle": 0, + "content": "giggling sound (1st and 3rd rows) and sobbing sound (2nd row). Our model excels in preserving the core facial attributes while manipulating emotional expressions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.414, + 0.619, + 0.429 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.576 + ], + "angle": 0, + "content": "In this paper, we propose The Power of Sound (TPoS), a novel audio-to-video generation with Stable Diffusion. Our work extends the usage of audio modality on generation models, and broaden the methods of using Stable Diffusion by generating realistic videos by our Audio Encoder. Superior performances are achieved over widely-used audio-to-video benchmarks, reflected by objective evaluations and User study, hence attributing towards the new formulation of video generation with audio modality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.893, + 0.882 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Research Foundation of Korea grant (NRF-2021R1C1C1009608, \\(10\\%\\)), Basic Science Research Program (NRF-2021R1A6A1A13044830, \\(5\\%\\)), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(2022-0-00043, \\(30\\%\\)), and the MSIT(Ministry of Science and ICT), Korea, under the ICT Creative Consilience program(IITP-2023-2020-0-01819, \\(5\\%\\)) supervised by the IITP. W. Ryoo, S. Lee, and S. Kim are supported by the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2019-0-00079, Artificial Intelligence Graduate School Program(Korea University), \\(30\\%\\)), the National Research Foundation of Korea grant (NRF-2022R1F1A1074334, \\(15\\%\\)), Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: 4D Content Generation and Copyright Protection with Artificial Intelligence, Project Number: R2022020068, \\(5\\%\\)), and supported by the Google Cloud Research Credits program(code-MU6X-FKAV-N3A3-1X0T)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7830" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint arXiv:2206.02779, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.144, + 0.469, + 0.198 + ], + "angle": 0, + "content": "[2] Manuel Brack, Patrick Schramowski, Felix Friedrich, Dominik Hintersdorf, and Kristian Kersting. The stable artist: Steering semantics in diffusion latent space. arXiv preprint arXiv:2212.06013, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.2, + 0.469, + 0.255 + ], + "angle": 0, + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.257, + 0.469, + 0.325 + ], + "angle": 0, + "content": "[4] Moitreya Chatterjee and Anoop Cherian. Sound2sight: Generating visual dynamics from sound and context. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 701-719. Springer, 2020. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.327, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.397, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.453, + 0.468, + 0.493 + ], + "angle": 0, + "content": "[7] Sarah Gross, Xingxing Wei, and Jun Zhu. Automatic realistic music video generation from segments of youtube videos. arXiv preprint arXiv:1905.12245, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.495, + 0.468, + 0.549 + ], + "angle": 0, + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.551, + 0.469, + 0.605 + ], + "angle": 0, + "content": "[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[10] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[12] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[13] Dasaem Jeong, Seungheon Doh, and Taegyun Kwon. Traumerai: Dreaming music with stylegan. arXiv preprint arXiv:2102.04680, 2(4):10, 2021. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[14] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "emotional video portraits. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14080-14089, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.137, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[16] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[17] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.265, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[18] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Ccvs: context-aware controllable video synthesis. Advances in Neural Information Processing Systems, 34:14042-14055, 2021. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.322, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[19] Seung Hyun Lee, Gyeongrok Oh, Wonmin Byeon, Chanyoung Kim, Won Jeong Ryoo, Sang Ho Yoon, Hyunjun Cho, Jihyun Bae, Jinkyu Kim, and Sangpil Kim. Sound-guided semantic video generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 34-50. Springer, 2022. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[20] Seung Hyun Lee, Wonseok Roh, Wonmin Byeon, Sang Ho Yoon, Chanyoung Kim, Jinkyu Kim, and Sangpil Kim. Sound-guided semantic image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3377-3386, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[21] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[22] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 423-439. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[23] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[24] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[25] Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le. Specaugment: A simple data augmentation method for automatic speech recognition. arXiv preprint arXiv:1904.08779, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7831" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[28] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.471, + 0.387 + ], + "angle": 0, + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.389, + 0.47, + 0.443 + ], + "angle": 0, + "content": "[31] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.445, + 0.469, + 0.513 + ], + "angle": 0, + "content": "[32] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.516, + 0.469, + 0.583 + ], + "angle": 0, + "content": "[33] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14144-14153, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.586, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[34] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.712 + ], + "angle": 0, + "content": "[35] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.469, + 0.794 + ], + "angle": 0, + "content": "[36] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.798, + 0.469, + 0.866 + ], + "angle": 0, + "content": "[37] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. arXiv preprint arXiv:2212.08698, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.866 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7832" + } + ] +] \ No newline at end of file diff --git a/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/3b9298a9-5799-4904-96da-bef6d287733c_origin.pdf b/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/3b9298a9-5799-4904-96da-bef6d287733c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e356e21052058af01f29c87ad701df21928e0be4 --- /dev/null +++ b/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/3b9298a9-5799-4904-96da-bef6d287733c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec9f5feb2d18e1b355032aa23e1efdc337e69bce25496afd4da57d4265a8891c +size 10481359 diff --git a/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/full.md b/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d0091f1ffc061c5d58612e0c9ffc8028408dc6c5 --- /dev/null +++ b/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/full.md @@ -0,0 +1,271 @@ +# The Power of Sound (TPoS): Audio Reactive Video Generation with Stable Diffusion + +Yujin Jeong $^{1}$ , Wonjeong Ryoo $^{2}$ , Seunghyun Lee $^{2}$ , Dabin Seo $^{1}$ , Wonmin Byeon $^{3}$ , Sanggil Kim, $^{2,*}$ and Jinkyu Kim $^{1,*}$ + +$^{1}$ Department of Computer Science and Engineering, Korea University, Seoul 02841, Korea + $^{2}$ Department of Artificial Intelligence, Korea University, Seoul 02841, Korea + $^{3}$ NVIDIA Research, Santa Clara 95050, USA + +*Correspondences: S. Kim (spk7@korea.ac.kr) and J. Kim (jinkyukim@korea.ac.kr) + +![](images/9b7c727e73eae2e0ef7588d2ee6ef4ab3623e5cc1fa50a4a3d64e8e9aca43c6c.jpg) +Figure 1: The Power of Sound (TPoS) is a novel framework that generates audio-reactive video sequences. Built upon the Stable Diffusion model, our model first generates an initial frame from a user-provided text prompt (e.g. "a photo of a beautiful beach with a blue sky"), then reactively manipulates the style of generated images corresponding to the sound inputs (e.g. an audio sequence of fireplace). Our model is indeed able to generate a frame conditioned on semantic information of the sound (see 1st and 2nd rows where images are manipulated driven by sound inputs such as fireplace or wave sound), while realistically dealing with temporal visual changes conditioned on changes of sound, e.g., increasing magnitude of sounds (see second row) or wave $\rightarrow$ fireplace (see last row). TPoS creates visually compelling and contextually relevant video sequences in an open domain. + +# Abstract + +In recent years, video generation has become a prominent generative tool and has drawn significant attention. However, there is little consideration in audio-to-video generation, though audio contains unique qualities like temporal semantics and magnitude. Hence, we propose The Power of Sound (TPoS) model to incorporate audio input that includes both changeable temporal semantics and mag- + +nitude. To generate video frames, TPoS utilizes a latent stable diffusion model with textual semantic information, which is then guided by the sequential audio embedding from our pretrained Audio Encoder. As a result, this method produces audio reactive video contents. We demonstrate the effectiveness of TPoS across various tasks and compare its results with current state-of-the-art techniques in the field of audio-to-video generation. More examples are available + +# 1. Introduction + +Recent generative models have demonstrated the potential to generate visuallyappealing video frames [32, 11, 10, 23, 36]. They often use a simple text prompt (e.g., "a video of a person on the street on a rainy day") to generate a video which is intuitive for end-users to drive a video generation. Text can effectively convey uni-modal object-wise guidance, such as "a rainy day" or "a person", but it may be challenging if users want to drive it into more complex sequential procedures, i.e., "a video of a person on the street on a rainy day, but a rain suddenly stops, and a wind blows." + +In this paper, we leverage sounds to guide the video generation models, i.e., sound-driven video generation. Audio is another modality that can complement texts by effectively providing sequential information (or temporal semantics): e.g., a continuous transition from the sound of light rain to the sound of heavy rain. There have been introduced sound-guided video generation approaches. However, existing sound-guided video generation approaches are limited to specific applications, such as face generation [15, 26, 21, 14, 21], where audio is used to provide a script for the avatar, or other simple synthetic motions (e.g., a video of musicians playing violin or a video of painting motions by an artist) [7, 13, 18, 4]. + +Recently, Lee et al. [19] introduced a sound-guided landscape video generation model, leveraging the latent space of StyleGAN [16]. They focus on using audio only for semantic labels (i.e., a sound of the wind is simply encoded into a meaning of wind) but not temporal semantics – i.e. semantic information that changes over time. Thus, in this work, we focus on leveraging temporal semantics from audio inputs such that our video generator reactively manipulates video frames. Our model temporally aligns the latent space with the given audio sequence (e.g., continuous changes in audio, e.g., weak rain $\rightarrow$ heavy rain, are reflected to generate corresponding video frames). + +Our work starts with Stable Diffusion [29], a text-driven image generator with advantages in generating high-resolution images based on the latent diffusion models. Its architectural advantages (i.e., attention mechanism and diffusion process) help leverage audio as a driving condition, generating temporally reactive and consistent video frames. Given the latent space of trained Stable Diffusion, we generate video frames temporally guided by audio sequences with regularizers to ensure temporal consistency (between generated consecutive frames) and correspondence with audio inputs. + +Our model consists of two main modules: (i) Audio Encoder, which is designed to encode temporal semantics of audio sequences, producing a sequence of the latent vec + +Prompt: "A photo of a person on the street" + +![](images/e6459603441b5731c5697bb6451a56ac13e1b78ee84b827291153dec493de081.jpg) +Figure 2: Limitation of text-driven image manipulation. Given a generated image by Stable Diffusion [29] with a text prompt, "A photo of a person on the street," additional textual conditions of different semantic meanings (i.e. "a little of rain", "rain", and "rain a lot") produce similar images, failing to capture differences. Note that we apply SEGA [2]'s guidance to preserve content identity. + +tors. (ii) Audio Semantic Guidance Module, which uses the above-mentioned latent vectors as a condition in the diffusion process to generate corresponding image outputs. We apply identity regularizer to produce temporally consistent video frames, while we apply audio semantic guidance to generate audio-reactive video frames. + +However, we observe that generating multifarious high-quality images solely from a sound input is challenging due to the lack of such a large-scale dataset to train a model. Instead, we first generate an initial frame using pre-trained Stable Diffusion model with a text prompt, then generate the following video frames conditioned on audio inputs. This frees a model from data dependency burdens and enables training with the current relatively small (than imagetext modality) audio dataset [19, 5], focusing on leveraging temporal semantics from audio. We summarize our contributions as follows: + +- We propose a novel sound-driven video generation method built upon Stable Diffusion [29] and can generate video frames reactively with audio sequence inputs. +- Our attention-based Audio Encoder produces temporally-aware latent vectors, which are consumed by Stable Diffusion as a per-time manipulation condition, producing audio-reactive video frames. +- Our model regularizes the latent features of diffusion models to produce temporally consistent video frames, preserving identity throughout the generated video. +- We demonstrate the effectiveness of our proposed model using a public dataset Landscape [19], generally outperforming other state-of-the-art sound-driven video generation approaches in terms of video quality metrics and human evaluation. + +# 2. Related Work + +Latent Diffusion Models. Recent success [29] suggests that the Latent Diffusion Models (LDM) improve the efficiency of the diffusion process, successfully generating + +Table 1: Comparison between existing state-of-the-art audio-driven video generation approaches in terms of whether they consider the following factors: temporal semantics, magnitude changes of sound, and target domains. + +
ModelTemporal SemanticsMagnitudeDomains (Audio Type)
Sound2Sight [4]-✓Closed
CCVS [18]-✓Closed (Music)
TräumerAI [13]-✓Closed (Music)
Lee et al. [19]✓-Closed (Nature)
Ours✓✓Open Domains
+ +high-quality images given a text prompt. One challenge in LDM is that the generation process is too sensitive to the condition, making it difficult to control semantics. Recently, there have been introduced to control semantics with LDM by a semantic mask [1] or by utilizing semantic information of the cross-attention layers [9]. Wu et al. [37] used linear combinations of text embeddings and Liu et al. [22] proposed composable diffusion models, but they still remained challenging to control fine-grained semantic changes. Recently, Semantic Guidance (SEGA) [2] computed a guidance vector in the latent space, enabling semantic control of diffusion models without further inputs. Inspired by SEGA, we also control the semantics in the latent space with temporally-encoded audio vector sequences. + +Text-driven Video Generation. Recent text-to-video generation tools, including Make-A-Video [32], Video Diffusion Models [11], Imagen video [10], and Phenaki [36] have shown promising performance in generating videos from textual descriptions. However, text-to-video generation has its limitations in terms of temporal coherence, which mostly leads to short video duration or a linear video change. Recent text-to-video generation methods, StyleGAN-V [34] and Dreamix [23], made progress addressing these issues. However, conditioning temporal semantics or complex scenarios is still challenging to be obtained from text inputs. Thus, in this paper, we want to explore conditioning a model with audio inputs, which inherently convey such temporal semantics. + +Audio-driven Video Generation. Leveraging temporal semantics was not seriously considered in previous audio-driven video generation approaches. Sound2Sight[4] and CCVS [18] generate video frames conditioned on the (nontemporal) context of the given audio, while TraumerAI [13] utilized the magnitude of the given audio. Recently, Lee et al. [19] explored a model that can consider audio semantics as a condition to drive a video generator. Also, their dependency on StyleGAN [34]-based embedding space makes it difficult for models to generate transitions in video. In this work, we focus on leveraging temporal semantics from audio inputs such that our generator reactively manipulates video frames. + +# 3. Method + +In this work, we propose a novel audio-driven video generation method, which generates video frames conditioned on audio sequences, ensuring temporal consistency between consecutive frames and temporal correspondence with audio inputs. As shown in Figure 3, our model consists of two main parts: (i) Audio Encoder, which encodes temporal semantics of audio sequences, producing a sequence of the latent vectors (Section 3.2). (ii) Audio Semantic Guidance Module, which uses the above-mentioned latent vectors as a condition in the diffusion process to generate corresponding image outputs, which are temporally consistent (by our identity regularizer) and audio-reactive (see Section 3.3). + +# 3.1. Preliminary: Latent Stable Diffusion + +Latent Diffusion Models (LDMs) [29] are the method that uses an encoder to convert a noised latent vector $\mathbf{z}^T$ to a denoised latent vector $\mathbf{z} = \mathbf{x} + \epsilon$ , where $\mathbf{z}$ is a latent vector of an input image $\mathbf{x}$ and $\epsilon$ is a noise. Stable Diffusion [29] is part of a conditional generation model that can synthesize an image given a condition $\mathbf{y}$ . It uses U-Net [30] as denoising autoencoders represented by $\epsilon_{\theta}$ . To generate an output image, the autoencoder $\epsilon_{\theta}(\mathbf{z}^{t}, t, \tau_{\theta}(\mathbf{y}))$ takes three inputs: the noised latent vector $\mathbf{z}^t$ , a sequence $t$ that is uniformly sampled from the set $1, \dots, T$ , and a conditional input $\mathbf{y}$ . + +Conditional input $\mathbf{y}$ is first transformed into a latent vector $\mathbf{c}_p$ through a pretrained function $\tau_{\theta}$ and then it is fed into a cross-attention layer of the U-Net as the key and value, which is then combined with $\mathbf{z}^t$ by an attention mechanism, where the query is the flattened intermediate representation of the U-Net. Next, the denoising autoencoder $\epsilon_{\theta}$ is used to denoise $\mathbf{z}^t$ from $t = T$ to $t = 1$ sequentially. After $T$ denoising steps, the resulting denoised latent vector $\mathbf{z}(= \mathbf{z}^{1})$ is transmitted to the decoder, which produces the final output image $\tilde{\mathbf{x}}$ . + +# 3.2. Encoding Temporal Semantics from Audio + +We use an audio modality as a source of generating temporal conditions. We divide the audio mel-spectogram into uniform snippets (segments) to extract abundant audio features and feed this into the Audio Encoder to extract both temporal semantic features and intensity of audio in end-to-end manner. The figure of our training process is illustrated in Figure 4. + +Audio Feature Extraction. Audio inputs are first transformed into a mel-spectrogram representation, denoted as $\mathbf{x}^a\in \mathbb{R}^{d\times w}$ , where $d$ represents the number of mel-frequency bins and $w$ is the width of the spectrogram. To incorporate time information, the mel-spectrogram is divided into $N$ segments. Each segment, denoted as $\mathbf{x}_n^a\in \mathbb{R}^{d\times \lceil \frac{w}{N}\rceil}$ where $n\in \{1,\ldots ,N\}$ , is then fed into a shared feature extraction module, i.e., the pre-trained ResNet18 [8]. The + +![](images/8bd59a99cce9913cd2bf813cf7788bc82765c429d1f8cd78bc500d99afda4e5b.jpg) +Figure 3: An overview of our proposed TPoS model. Our model consists of two main modules: (i) Audio Encoder, which produces a sequence of latent vectors, encoding temporal semantics of audio input by utilizing CLIP [27] space and highlighting the important temporal features and (ii) Audio Semantic Guidance Module, which is based on the diffusion process, generating video frames that are temporally consistent and audio-reactive. + +feature extraction module $f_{a}(\cdot)$ learns to extract low-level features from each audio segment regardless of its time dependency, i.e., $\mathbf{w}_n = f_a(\mathbf{x}_n^a)$ + +LSTM-based Temporal Semantic Encoder. As our goal is to generate audio-reactive video frames, it is also important to encode temporal changes (or relations) of the given audio inputs. Similar to Lee et al. [19], given audio features $\mathbf{w} \in \{\mathbf{w}_1, \mathbf{w}_2, \dots, \mathbf{w}_N\}$ , which encodes per-segment disjoint audio representation, we apply the standard Long Short-Term Memory (LSTM) network [12] to encode temporal relations or changes between consecutive audio features $\mathbf{w}$ . Formally, our LSTM takes the audio feature $\mathbf{w}_{n-1}$ as input and updates its hidden state $\mathbf{h}_n$ , producing an output $\mathbf{s}_t$ : i.e. $(\mathbf{s}_n, \mathbf{h}_n) = \mathrm{LSTM}(\mathbf{h}_{n-1}, \mathbf{w}_{n-1})$ . These outputs $\mathbf{s} \in \{\mathbf{s}_1, \mathbf{s}_2, \dots, \mathbf{s}_N\}$ are then fed into Temporal Attention Module (TAM) to further encode temporal semantics. + +Aligning Audio Semantics with Image-Text CLIP Joint Space. As we will use the output $s$ as a condition to manipulate video frames, it is important to ensure those audio features are well-aligned with other text and visual features in the CLIP [27]-based joint embedding space. Similar to Lee et al. [20], given the pre-trained image-text CLIP space, we apply the following loss $\mathcal{L}_{\mathrm{CLIP}}^{a \leftrightarrow t}$ with the InfoNCE loss [24] $l_{sim}$ such that positive pairs (e.g. an audio of raining and a text prompt "raining") are pulled close to each other, while negative pairs are pushed farther away. + +$$ +\mathcal {L} _ {\mathrm {C L I P}} ^ {a \leftrightarrow t} = l _ {s i m} (\mathbf {s} _ {N}, \operatorname {C L I P} _ {t} (\mathbf {t})) + l _ {s i m} (\operatorname {C L I P} _ {t} (\mathbf {t}), \mathbf {s} _ {N}) \quad (1) +$$ + +![](images/618663442c550a1cca5f593d6e0c47315cd9240b4ef34270ff1e8a8c5721ced2.jpg) +Figure 4: An overview of our Audio Encoder training process. Our model generates temporally-encoded audio embeddings with an LSTM [12] layer and Temporal Attention Module (TAM). Audio input is partitioned into $N$ segments, and each of these is encoded and used as a condition to manipulate audio-reactive video sequences (e.g. light rain → heavy rain). This is done by our Mapping Module (MAP), which maps the audio embedding to the latent space of Stable Diffusion. + +where $\mathrm{CLIP}_t$ is a pre-trained CLIP-based text encoder, which takes a text prompt $\mathbf{t}$ obtained from audio class labels as input, yielding an $d$ -dimensional feature. Note that + +we only apply this loss for the final output $\mathbf{s}_N$ for efficient training. Given a set of positive pairs, we apply the following InfoNCE loss $l_{sim}(\mathbf{a},\mathbf{b})$ : + +$$ +l _ {s i m} = - \log \frac {\exp \left(\left\langle \mathbf {a} _ {i} , \mathbf {b} _ {i} \right\rangle\right) / \tau}{\sum_ {j} \exp \left(\left\langle \mathbf {a} _ {i} , \mathbf {b} _ {j} \right\rangle\right) / \tau} \tag {2} +$$ + +where $\langle \mathbf{a}_i,\mathbf{b}_i\rangle$ represents the cosine similarity with temperature $\tau$ . Note that we set $\tau$ to 0.07. + +Augmenting Audio Semantics. Audio data is often limited in volume and diversity; thus, augmentation techniques may be required to extract better-quality audio semantic features, preventing a representation collapse. We use SpecAugment [25] to apply random transformations (such as masking our certain frequency bands or time segments), yielding augmented audio inputs. We further add the InfoNCE loss [24] $\mathcal{L}_{\mathrm{CLIP}}^{a\leftrightarrow a^{\prime}}$ to pull augmented audio features together. + +$$ +\mathcal {L} _ {\mathrm {C L I P}} ^ {a \leftrightarrow a ^ {\prime}} = l _ {s i m} \left(\mathbf {s} _ {N}, \mathbf {s} _ {N} ^ {\prime}\right) + l _ {s i m} \left(\mathbf {s} _ {N} ^ {\prime}, \mathbf {s} _ {N}\right) \tag {3} +$$ + +where apostrophe indicates augmented view of an original audio data. Finally, we use the semantic loss $\mathcal{L}_{\mathrm{semantic}}$ : i.e. $\mathcal{L}_{\mathrm{semantic}} = \mathcal{L}_{\mathrm{CLIP}}^{a\leftrightarrow t} + \lambda_s\mathcal{L}_{\mathrm{CLIP}}^{a\leftrightarrow a'}$ where $\lambda_{s}$ is set to 0.6. + +Temporal Attention Module (TAM). As shown in Figure 5, we further use an attention-based module to encode temporal semantics from the audio inputs. We empirically observe that adding this module helps improve the quality of video frame generation, which is probably due to the fact that the model becomes flexible to focus on more important temporal information, improving the model's representation power. Formally, we first compute attention weight $\alpha_{n}$ for a given audio feature $\mathbf{s}_n$ by applying an MLP layer $f_{\mathrm{proj}}$ followed by a softmax operation: i.e. $\alpha_{n} = \exp (f_{\mathrm{proj}}(\mathbf{s}_{n})) / \sum_{n}\exp (f_{\mathrm{proj}}(\mathbf{s}_{n}))$ such that $\sum_{n}\alpha_{n} = 1$ . We compute the weighted sum of audio features based on attention weights, yielding an attended audio feature $\mathbf{o}^a = \sum_n\alpha_n\mathbf{s}_n$ . We add another InfoNCE loss $\mathcal{L}_{\mathrm{temporal}}$ to align the audio features with text: + +$$ +\mathcal {L} _ {\text {t e m p o r a l}} = l _ {s i m} \left(\mathbf {o} ^ {a}, \operatorname {C L I P} _ {t} (\mathbf {t})\right) + l _ {s i m} \left(\operatorname {C L I P} _ {t} (\mathbf {t}), \mathbf {o} ^ {a}\right) \tag {4} +$$ + +Lastly, we also minimize the MSE loss between text embeddings (before the projection layer) and the projected audio feature $\mathrm{MAP}(\mathbf{o}^a)$ : $\mathcal{L}_{\mathrm{cond}} = ||\mathbf{c}_t - \mathrm{MAP}(\mathbf{o}^a)||_2^2$ . Note that we exclude the feature for the $<\mathrm{SOS}>$ token. More details are provided in the supplemental material. + +Total Loss. We train our Audio Encoder end-to-end by minimizing the following loss function $\mathcal{L}$ : + +$$ +\mathcal {L} = \mathcal {L} _ {\text {s e m a n t i c}} + \mathcal {L} _ {\text {t e m p o r a l}} + \mathcal {L} _ {\text {c o n d}} \tag {5} +$$ + +# 3.3. Generating Video Frames with Stable Diffusion + +Initial Frame Generation from Text Prompt. Our model relies on leveraging combinational operations of denoising + +![](images/1e0b278d79c4a3a16687e44316599baa4b254eb4053bb14cc99223f74d5fc4a5.jpg) +Figure 5: Details of our Temporal Attention Module (TAM). Temporal Attention Module effectively captures the important audio segments (e.g. harsh raining sound) as attention weights, which are guided by CLIP [27]'s text embedding (e.g. "raining") in training phase and express the magnitude of audio in test phase. + +process. The estimated noise space can be either random or manual by visual input with the help of the diffusion encoder. Based on this guided diffusion, our model first generates the initial frame with a text prompt (e.g. "a photo of beautiful beach with blue sky"). Given this generated image as content, we manipulate its styles according to audio inputs and generate corresponding video frames, i.e. given a series of latent vectors $\{\mathbf{c}_1,\mathbf{c}_2,\dots ,\mathbf{c}_m\}$ , we generate $m$ video frames. Note that the number of video frames is controllable by latent vector interpolation, as we will explain later. We follow the standard image generation process with the Stable Diffusion model [29], i.e., we compute a latent vector $\mathbf{c}_p$ in a CLIP-based embedding space given a text prompt, conditioning it to generate an image. + +Audio Semantic Guidance. We employed the SEGA [2] framework to create video frames that incorporate sound style while preserving the main content identity. We first utilize the combination of attention $\alpha_{n}$ and audio feature $\mathbf{s}_n$ with normalization of the scale of output feature by multiplying $N$ to produce output $\mathbf{c}_n$ : i.e., $\mathbf{c}_n = N^k\alpha_n^k\mathbf{s}_n$ , where $k$ is hyper parameter that regulates attention ratio and is set to 1. The Audio Semantic Guidance module generates the $n$ -th video frame by taking audio condition $\mathbf{c}_n$ as input to guide the diffusion models. The denoising autoencoder $\epsilon_{\theta}$ is executed $\delta - 1$ times out of $T$ time to form incomplete noise along with the original text prompt meaning. From $t = \delta$ , the audio semantic guidance operates through the following equation: + +$$ +\begin{array}{l} \tilde {\epsilon} _ {\theta} \left(\mathbf {z} ^ {\delta}, \mathbf {c} _ {p}\right) := \epsilon_ {\theta} \left(\mathbf {z} ^ {\delta}, \mathbf {c} _ {\varnothing}\right) + g \left(\epsilon_ {\theta} \left(\mathbf {z} ^ {\delta}, \mathbf {c} _ {p}\right) - \epsilon_ {\theta} \left(\mathbf {z} ^ {\delta}, \mathbf {c} _ {\varnothing}\right)\right) \tag {6} \\ + \lambda (\mathbf {z} ^ {\delta}, \mathbf {c} _ {n}) \\ \end{array} +$$ + +where $\mathbf{z}^{\delta}$ is denoised latent vector at $t = \delta$ , $g$ is the guidance scale of the text prompt, $\lambda(\mathbf{z}^{\delta}, \mathbf{c}_n)$ is the audio semantic guidance term, and $\mathbf{c}_{\emptyset}$ represents an unconditioned prompt that does not make any semantic difference. As a result, only the $\lambda(\mathbf{z}^{\delta}, \mathbf{c}_n)$ term has been added to the original denoising process from $t = \delta$ to $t = 1$ . Note that $\mathbf{z}^T$ is fixed through frames in one video. The audio semantic guidance + +Table 2: Comparison of the quality of generated video frames with state-of-the-art audio-to-video generations in terms of IS [31], FVD [35], and CLIP [27]-based distances. + +
ModelInputIS↑FVD↓CLIP↑ (a ↔ v)CLIP↑ (t ↔ v)
Sound2Sight [4]1st Frame1.02 ± 0.02494.280.03640.2164
CCVS [18]1st Frame1.30 ± 0.20679.940.12510.2360
TräumerAI [13]-1.47 ± 0.19736.320.15890.1778
Sound-guided Video Generation [19]-1.16 ± 0.16544.090.11510.1702
Ours (w/o TAM)Latent Vector1.27 ± 0.23483.760.13420.2370
OursLatent Vector1.49 ± 0.38421.230.19640.2436
+ +![](images/b1f2669aa345167bf55ac69a84dbf72eb48948130c331cbab85748f39e5c4c73.jpg) +Figure 6: Examples of generated video frames (given fire crackling and waterfall audio) by Sound2Sight [4], CCVS [18], TraumerAI [13], Lee et al. [19], and ours. Note that Sound2Sight and CCVS use an initial frame (highlighted in a red box). + +![](images/512f567d03735f7645ce65a57e475ef21a3b012b9c50a3bfe6f375118ce23eed.jpg) + +$\lambda (\mathbf{z}^{\delta},\mathbf{c}_n)$ is defined as follows: + +$$ +\lambda \left(\mathbf {z} ^ {\delta}, \mathbf {c} _ {n}\right) = g _ {s} \left(\epsilon_ {\theta} \left(\mathbf {z} ^ {\delta}, \mathbf {c} _ {n}\right) - \epsilon_ {\theta} \left(\mathbf {z} _ {\varnothing} ^ {\delta}, \mathbf {c} _ {\varnothing}\right)\right) + \sigma_ {m} \Phi_ {m} \tag {7} +$$ + +where $\sigma_{m} \in [0,1]$ is the momentum hyper parameter that scales the momentum $\Phi_{m}$ and $g_{s}$ focuses on the relevant dimensions to the audio manipulation task (More details can be found in the supplemental material or SEGA [2]). Different from SEGA framework, we only consider positive guidance of audio semantics, which is the semantic difference between the guidance provided by $\mathbf{c}_n$ and the unconditioned guidance $\mathbf{c}_{\emptyset}$ . + +The generation process conditioned by $\mathbf{c}_n$ is separately working through diffusion processes so that the different semantic meaning or magnitude which presents in $\mathbf{c}_n$ can generate frames independently. + +Temporal Frame Interpolation. We use an interpolated latent vector to generate continuous video frames between two consecutive frames. Following the work by Ramesh et al. [28], we apply a spherical linear interpolation between all consecutive pairs of $\mathbf{c}_n$ and $\mathbf{c}_{n + 1}$ , yielding $k$ interpolated latent vectors. These vectors are then used as a condition for the diffusion models to generate temporally-interpolated video frames. + +# 4. Experiments + +Datasets. We use two Audio-Video datasets to train our Audio Encoder: VGG-Sound [5] and Landscape [19]. VGG-Sound is an audio dataset with about 170,000 of 10-second clips of audio-video data, which consists of 309 classes. The dataset has numerous 'in the wild' audio data that spans + +![](images/d2afda1fe476ecb35dce1039f693b49994eed5bec975826a98400f58641bfea7.jpg) +Figure 7: Examples of our video generation conditioned on text prompt (top) and audio (bottom). + +a large number of challenging acoustic environments and real application noise characteristics. Since audio of nature sound is the perfect tool to stylize compared to the class such as people talking or sports, we add about 9,000 audio clips of Landscape audio dataset in the training process. For obtaining test sets, class-balanced sampling is applied to Landscape dataset. + +Baselines. We compare our methods with existing audio-to-video generation methods, Sound2Sight [4], CCVS [18], and StyleGAN [34] based TraumerAI [13] and Sound-guided Video Generation [19]. We follow experiment methods by Lee et al. [19]. All baselines are trained or finetuned with Landscape dataset [19]. Since Sound2Sight and CCVS need first frame to generate the video, we randomly select first frame from the Landscape dataset at the inference task. For TraumerAI and Sound-guided Video Generation, we first pre-train StyleGAN [16] on high fidelity benchmark datasets (LHQ datasets [33]) and then train the model to navigate the latent space by Landscape dataset. Note that randomly initialized vector for TraumerAI [13] and Sound-guided Video Generation [19] are given. For our model, we initially generate image from random noise space and randomly sample the prompt related with landscape to generate the landscape like video for fair comparison. + +Evaluation Metrics. We use the following two video quality metrics for our evaluations: (i) Fréchet Video Distance (FVD) [35] and (ii) Inception Score (IS) [31]. FVD is used to assess video quality by measuring the distribution gap between real and synthesized videos in the latent space. Additionally, IS score is commonly employed to evaluate the effectiveness of Generative Adversarial Networks (GANs) by computing the KL-divergence between the label distribution of each image and the marginal label distribution. To implement these, we fine-tune Inflated 3D ConvNet [3] with Landscape [19] dataset for FVD and used pre-trained InceptionNet [17] that is trained on ImageNet [6] dataset. We also measure CLIP [27]-based cosine similarity (CLIP score) between audio and image as well as text and image. To obtain the textual pivot feature, we fed the following prompt "The photo of " into CLIP text encoder. + +![](images/c4ef6f1b0b7e2a72cee27f6ec681d429b4a2754873d2080cc0f031cadd6e3bd4.jpg) +Figure 8: Generated video frames with and without our Temporal Attention Module. We generate video frames with splashing water sound where its amplitude changes over time (see top). + +Quantitative Experiments. For fair comparison with other existing baselines such as TraumerAI [13] and Sound-guided Video Generation [19] which does not get a hint about what to generate, we use prompt that originally does not generate the style of sound. We set fps 20 and generated videos to extract images from all baselines. Table 2 shows that our approach produces the best quality results as video. Additionally, to ensure that the generated videos are semantically related to the sound, we compare the cosine similarity between text-audio and video embedding. Our methods show a superior performance in terms of multimodal semantics. + +Qualitative Video Quality Comparison. We first evaluate the quality of generated video frames. In Figure 6, we provide typical examples of generated video frames by (from top) Sound2Sight [4], CCVS [18], TraumerAI [13], Lee et al. [19], and ours. Note that Sound2Sight [4] and CCVS [18] need the initial frame as input (see red boxes). With audio inputs, such as fire crackling and waterfall, as a condition, ours generally generate temporally-consistent audio-reactive video frames. We observe that Sound2Sight [4] and CCVS [18] often show blurring artifacts, while StyleGAN-based TraumerAI [13] and Sound-guided Video Generation [19] often fail to generate audio-reactive video frames, e.g., they produce landscape scenes that are not semantically aligned with a fire crackling audio. However, ours generate a scene of a waterfall or a fire on firewood, aligning well with audio inputs. + +Comparison to Video Generation with Text. To analyze the benefit of audio modality, we conduct a qualitative experiment to compare the effect of audio and text modalities in generating visual content. As shown in Figure 7, we first generate an initial frame with the text prompt "A photo of deep in the sea." Then we generate the next frames with text "underwater bubbling" (top) and underwater bubbling + +![](images/bc8e924b9f3c71ba8146161b9ce4b01d892f87a41249f02e5b0cd779e84419f5.jpg) + +![](images/a31f840981aad8318e349338506c53786970c0b37a5263a395b88c00d282b49e.jpg) +Figure 9: Generated video frames with and without our Audio Semantic Guidance. Edited parts by Audio Semantic Guidance are highlighted with red box (see second row). +Figure 10: Examples of generated video frames with a sound that changes over time (e.g. bird singing $\rightarrow$ wind noise). + +sound (bottom). It is difficult to make temporal changes conditioned on text unless we train our model with a text-video dataset. Thus, we instead change the noise scale to make temporal changes, preserving identity. However, as shown in Figure 7 (top), it generates distortions or a linear change, but this is not the case for audio. Our model with audio generates visually-appealing video frames. + +Effect of Temporal Attention Module. We use Temporal Attention Module (TAM) to improve the representation power to encode temporal semantics better. To analyze this, we perform an ablation study with and without TAM to see its effect on video frame generation. We observe in Figure 8 that our model is indeed able to generate video frames reactive to the audio changes over time (compare the changes along with the given splashing water sound). + +Effect of Audio Semantic Guidance. We demonstrate the effect of Audio Semantic Guidance with ablation study. First, we generate video frames without Audio Semantic Guidance by replacing audio semantic guidance term $\lambda (\mathbf{z}^{\delta},\mathbf{c}_n)$ with $\epsilon_{\theta}(\mathbf{z}^{\delta},\mathbf{c}_n)$ (refer the notation in Section 3.3). To remove the effect of Audio Semantic Guidance, we also set $\delta$ to $T$ . As illustrated in Figure 9 (top), without Audio Semantic Guidance, the model has a tendency to produce unnecessary changes (e.g. grass $\rightarrow$ toil) and struggles to + +![](images/14155dfec2472ad9982bf798bb225bb34574ba0c1f2134444c4083891903dc6a.jpg) +Figure 11: Our human evaluation results. We conduct a user study with 100 participants on Amazon Mechanical Turk (AMT). Participants are shown generated video frames and asked to evaluate them in terms of realistic, vividness, consistency, and relevance. The Likert scale is used (higher is better). + +achieve consistent alignment with audio (e.g. producing a road with a waterfall burbling sound). On the other hand, by leveraging Audio Semantic Guidance as in Figure 9 (bottom), we can generate sequential video frames that not only has consistent content but also represents natural temporal variations according to audio sound (e.g. producing wave of water with waterfall burbling sound), resulting in enhanced naturalness. + +Experiments of Semantic Transition. In Figure 10, we provide examples where audio inputs are changed (e.g., bird singing $\rightarrow$ wind noise, strong wind $\rightarrow$ raining). As we observe in that figure, our model successfully adapts to the audio change, generating video frames accordingly. This may confirm that our model is indeed conditioned on the audio sequence and can generate audio-adaptive video frames. + +User study. Further, we conduct a human evaluation to evaluate the video quality by human judges. We recruit 100 participants from Amazon MTurk. Participants are shown video frames generated by five different audio-driven video generation models: Sound2Sight [4], CCVS [18], TraumerAI [13], Lee et al. [19], and ours. Participants are asked to evaluate the given video frames in terms of realism, vividness, consistency, and relevance. We use a five-point Likert scale where ideal video frames will get all five points. More details are provided as supplemental material. We observe in Figure 11 that our proposed method outperforms the other approaches in all categories. These results are consistent with our quantitative and qualitative results. + +Visual Conditioning. Our model is capable of producing an initial frame using visual input (such as images) with the trained diffusion encoder. Figure 12 demonstrates that our framework leverages latent space of a diffusion model with a combination of visual input, text prompt and audio. We first randomly choose visual input (e.g., sun light photo) and + +![](images/656558e23f2214679b30af002468d997bd4c1aa89f305ed7c485db50619ac55d.jpg) +Figure 12: Example of generated video with visual input, text prompt and audio sound. (e.g., 1st row: sun light photo conditioned by text "A photo of beautiful garden with sky" and splashing water sound.) + +![](images/aaec0af009da9ac6088f5700107fa276d1e594124cf93120e03698bf387fbb79.jpg) +Figure 13: Example of generated videos with audio-text joint condition (e.g., 2nd row: conditioned with text "eruption" and explosion sound) + +generate an initial frame with the text prompt (e.g., "A photo of beautiful garden with sky"). Then we incorporate an audio sound (e.g., splashing water) to generate videos. Shown in Figure 12, it can produce videos featuring a garden with water where the sun is centered (1st row) or where a black mountain in the background (2nd row). + +Text-Audio Joint Conditioning. As our model is built upon the Stable Diffusion model, it is also possible to use text and audio as a condition together. In Figure 13, we provide an example where we generate video frames conditioned on a sound of an explosion along with texts, such as "eruption", "spew", or "cloud of ash." (see 2nd-4th rows) Preserving temporal semantics, our model successfully generates video frames guided by text as well. + +Face Generation. Our Audio Semantic Guidance enables detailed adjustment in certain areas, such as face generation. Figure 14 shows examples of human faces conditioned on + +![](images/e107fa8c28926bf5c18ed4d9ba4200263d258b82968d636d3c95ce00f328ecb7.jpg) +Figure 14: Examples of face generation with our methods. We use giggling and sobbing sound to manipulate face expressions of human face. + +giggling sound (1st and 3rd rows) and sobbing sound (2nd row). Our model excels in preserving the core facial attributes while manipulating emotional expressions. + +# 5. Conclusion + +In this paper, we propose The Power of Sound (TPoS), a novel audio-to-video generation with Stable Diffusion. Our work extends the usage of audio modality on generation models, and broaden the methods of using Stable Diffusion by generating realistic videos by our Audio Encoder. Superior performances are achieved over widely-used audio-to-video benchmarks, reflected by objective evaluations and User study, hence attributing towards the new formulation of video generation with audio modality. + +Acknowledgements. This work was supported by the National Research Foundation of Korea grant (NRF-2021R1C1C1009608, $10\%$ ), Basic Science Research Program (NRF-2021R1A6A1A13044830, $5\%$ ), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(2022-0-00043, $30\%$ ), and the MSIT(Ministry of Science and ICT), Korea, under the ICT Creative Consilience program(IITP-2023-2020-0-01819, $5\%$ ) supervised by the IITP. W. Ryoo, S. Lee, and S. Kim are supported by the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2019-0-00079, Artificial Intelligence Graduate School Program(Korea University), $30\%$ ), the National Research Foundation of Korea grant (NRF-2022R1F1A1074334, $15\%$ ), Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: 4D Content Generation and Copyright Protection with Artificial Intelligence, Project Number: R2022020068, $5\%$ ), and supported by the Google Cloud Research Credits program(code-MU6X-FKAV-N3A3-1X0T). + +# References + +[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint arXiv:2206.02779, 2022. 3 +[2] Manuel Brack, Patrick Schramowski, Felix Friedrich, Dominik Hintersdorf, and Kristian Kersting. The stable artist: Steering semantics in diffusion latent space. arXiv preprint arXiv:2212.06013, 2022. 2, 3, 5, 6 +[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7 +[4] Moitreya Chatterjee and Anoop Cherian. Sound2sight: Generating visual dynamics from sound and context. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 701-719. Springer, 2020. 2, 3, 6, 7, 8 +[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2, 6 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 7 +[7] Sarah Gross, Xingxing Wei, and Jun Zhu. Automatic realistic music video generation from segments of youtube videos. arXiv preprint arXiv:1905.12245, 2019. 2 +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3 +[10] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3 +[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 3 +[12] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 4 +[13] Dasaem Jeong, Seungheon Doh, and Taegyun Kwon. Traumerai: Dreaming music with stylegan. arXiv preprint arXiv:2102.04680, 2(4):10, 2021. 2, 3, 6, 7, 8 +[14] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2 +[15] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven + +emotional video portraits. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14080-14089, 2021. 2 +[16] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2, 7 +[17] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 7 +[18] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Ccvs: context-aware controllable video synthesis. Advances in Neural Information Processing Systems, 34:14042-14055, 2021. 2, 3, 6, 7, 8 +[19] Seung Hyun Lee, Gyeongrok Oh, Wonmin Byeon, Chanyoung Kim, Won Jeong Ryoo, Sang Ho Yoon, Hyunjun Cho, Jihyun Bae, Jinkyu Kim, and Sangpil Kim. Sound-guided semantic video generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 34-50. Springer, 2022. 2, 3, 4, 6, 7, 8 +[20] Seung Hyun Lee, Wonseok Roh, Wonmin Byeon, Sang Ho Yoon, Chanyoung Kim, Jinkyu Kim, and Sangpil Kim. Sound-guided semantic image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3377-3386, 2022. 4 +[21] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 2 +[22] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 423-439. Springer, 2022. 3 +[23] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 2, 3 +[24] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4, 5 +[25] Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le. Specaugment: A simple data augmentation method for automatic speech recognition. arXiv preprint arXiv:1904.08779, 2019. 5 +[26] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2 + +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 5, 6, 7 +[28] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6 +[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 5 +[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3 +[31] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 6, 7 +[32] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 3 +[33] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14144-14153, 2021. 7 +[34] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022. 3, 7 +[35] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6, 7 +[36] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2, 3 +[37] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. arXiv preprint arXiv:2212.08698, 2022. 3 \ No newline at end of file diff --git a/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/images.zip b/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c8f81f2cd705dc782fd65db4f27cb66c9218e530 --- /dev/null +++ b/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3509cedf6b9e48f3462f5a58fa802c72b5621d88d1d6a5f44ed446bdf309037f +size 967743 diff --git a/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/layout.json b/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f6b74f2ffb60b2e09613e38a4533c6904728937f --- /dev/null +++ b/2023/The Power of Sound (TPoS)_ Audio Reactive Video Generation with Stable Diffusion/layout.json @@ -0,0 +1,8440 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 126, + 103, + 468, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 103, + 468, + 138 + ], + "spans": [ + { + "bbox": [ + 126, + 103, + 468, + 138 + ], + "type": "text", + "content": "The Power of Sound (TPoS): Audio Reactive Video Generation with Stable Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "spans": [ + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "text", + "content": "Yujin Jeong" + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "text", + "content": ", Wonjeong Ryoo" + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "text", + "content": ", Seunghyun Lee" + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "text", + "content": ", Dabin Seo" + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "text", + "content": ", Wonmin Byeon" + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "text", + "content": ", Sanggil Kim," + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "text", + "content": " and Jinkyu Kim" + }, + { + "bbox": [ + 145, + 160, + 447, + 190 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 190, + 517, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 190, + 517, + 231 + ], + "spans": [ + { + "bbox": [ + 77, + 190, + 517, + 231 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 190, + 517, + 231 + ], + "type": "text", + "content": " Department of Computer Science and Engineering, Korea University, Seoul 02841, Korea \n" + }, + { + "bbox": [ + 77, + 190, + 517, + 231 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 190, + 517, + 231 + ], + "type": "text", + "content": " Department of Artificial Intelligence, Korea University, Seoul 02841, Korea \n" + }, + { + "bbox": [ + 77, + 190, + 517, + 231 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 190, + 517, + 231 + ], + "type": "text", + "content": " NVIDIA Research, Santa Clara 95050, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 232, + 476, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 232, + 476, + 245 + ], + "spans": [ + { + "bbox": [ + 118, + 232, + 476, + 245 + ], + "type": "text", + "content": "*Correspondences: S. Kim (spk7@korea.ac.kr) and J. Kim (jinkyukim@korea.ac.kr)" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 82, + 252, + 517, + 485 + ], + "blocks": [ + { + "bbox": [ + 82, + 252, + 517, + 485 + ], + "lines": [ + { + "bbox": [ + 82, + 252, + 517, + 485 + ], + "spans": [ + { + "bbox": [ + 82, + 252, + 517, + 485 + ], + "type": "image", + "image_path": "9b7c727e73eae2e0ef7588d2ee6ef4ab3623e5cc1fa50a4a3d64e8e9aca43c6c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 496, + 548, + 593 + ], + "lines": [ + { + "bbox": [ + 45, + 496, + 548, + 593 + ], + "spans": [ + { + "bbox": [ + 45, + 496, + 548, + 593 + ], + "type": "text", + "content": "Figure 1: The Power of Sound (TPoS) is a novel framework that generates audio-reactive video sequences. Built upon the Stable Diffusion model, our model first generates an initial frame from a user-provided text prompt (e.g. \"a photo of a beautiful beach with a blue sky\"), then reactively manipulates the style of generated images corresponding to the sound inputs (e.g. an audio sequence of fireplace). Our model is indeed able to generate a frame conditioned on semantic information of the sound (see 1st and 2nd rows where images are manipulated driven by sound inputs such as fireplace or wave sound), while realistically dealing with temporal visual changes conditioned on changes of sound, e.g., increasing magnitude of sounds (see second row) or wave " + }, + { + "bbox": [ + 45, + 496, + 548, + 593 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 45, + 496, + 548, + 593 + ], + "type": "text", + "content": " fireplace (see last row). TPoS creates visually compelling and contextually relevant video sequences in an open domain." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 602, + 192, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 602, + 192, + 615 + ], + "spans": [ + { + "bbox": [ + 143, + 602, + 192, + 615 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "In recent years, video generation has become a prominent generative tool and has drawn significant attention. However, there is little consideration in audio-to-video generation, though audio contains unique qualities like temporal semantics and magnitude. Hence, we propose The Power of Sound (TPoS) model to incorporate audio input that includes both changeable temporal semantics and mag-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 604, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 547, + 700 + ], + "type": "text", + "content": "nitude. To generate video frames, TPoS utilizes a latent stable diffusion model with textual semantic information, which is then guided by the sequential audio embedding from our pretrained Audio Encoder. As a result, this method produces audio reactive video contents. We demonstrate the effectiveness of TPoS across various tasks and compare its results with current state-of-the-art techniques in the field of audio-to-video generation. More examples are available" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7822" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 107, + 128, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 128, + 119 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 128, + 119 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 127, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 127, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 127, + 287, + 247 + ], + "type": "text", + "content": "Recent generative models have demonstrated the potential to generate visuallyappealing video frames [32, 11, 10, 23, 36]. They often use a simple text prompt (e.g., \"a video of a person on the street on a rainy day\") to generate a video which is intuitive for end-users to drive a video generation. Text can effectively convey uni-modal object-wise guidance, such as \"a rainy day\" or \"a person\", but it may be challenging if users want to drive it into more complex sequential procedures, i.e., \"a video of a person on the street on a rainy day, but a rain suddenly stops, and a wind blows.\"" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 247, + 288, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 288, + 402 + ], + "type": "text", + "content": "In this paper, we leverage sounds to guide the video generation models, i.e., sound-driven video generation. Audio is another modality that can complement texts by effectively providing sequential information (or temporal semantics): e.g., a continuous transition from the sound of light rain to the sound of heavy rain. There have been introduced sound-guided video generation approaches. However, existing sound-guided video generation approaches are limited to specific applications, such as face generation [15, 26, 21, 14, 21], where audio is used to provide a script for the avatar, or other simple synthetic motions (e.g., a video of musicians playing violin or a video of painting motions by an artist) [7, 13, 18, 4]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 403, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 287, + 545 + ], + "type": "text", + "content": "Recently, Lee et al. [19] introduced a sound-guided landscape video generation model, leveraging the latent space of StyleGAN [16]. They focus on using audio only for semantic labels (i.e., a sound of the wind is simply encoded into a meaning of wind) but not temporal semantics – i.e. semantic information that changes over time. Thus, in this work, we focus on leveraging temporal semantics from audio inputs such that our video generator reactively manipulates video frames. Our model temporally aligns the latent space with the given audio sequence (e.g., continuous changes in audio, e.g., weak rain " + }, + { + "bbox": [ + 47, + 403, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 47, + 403, + 287, + 545 + ], + "type": "text", + "content": " heavy rain, are reflected to generate corresponding video frames)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 546, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 286, + 677 + ], + "type": "text", + "content": "Our work starts with Stable Diffusion [29], a text-driven image generator with advantages in generating high-resolution images based on the latent diffusion models. Its architectural advantages (i.e., attention mechanism and diffusion process) help leverage audio as a driving condition, generating temporally reactive and consistent video frames. Given the latent space of trained Stable Diffusion, we generate video frames temporally guided by audio sequences with regularizers to ensure temporal consistency (between generated consecutive frames) and correspondence with audio inputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Our model consists of two main modules: (i) Audio Encoder, which is designed to encode temporal semantics of audio sequences, producing a sequence of the latent vec" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 310, + 72, + 421, + 79 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 72, + 421, + 79 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 421, + 79 + ], + "type": "text", + "content": "Prompt: \"A photo of a person on the street\"" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 79, + 544, + 146 + ], + "blocks": [ + { + "bbox": [ + 308, + 79, + 544, + 146 + ], + "lines": [ + { + "bbox": [ + 308, + 79, + 544, + 146 + ], + "spans": [ + { + "bbox": [ + 308, + 79, + 544, + 146 + ], + "type": "image", + "image_path": "e6459603441b5731c5697bb6451a56ac13e1b78ee84b827291153dec493de081.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 155, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 304, + 155, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 155, + 545, + 239 + ], + "type": "text", + "content": "Figure 2: Limitation of text-driven image manipulation. Given a generated image by Stable Diffusion [29] with a text prompt, \"A photo of a person on the street,\" additional textual conditions of different semantic meanings (i.e. \"a little of rain\", \"rain\", and \"rain a lot\") produce similar images, failing to capture differences. Note that we apply SEGA [2]'s guidance to preserve content identity." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 253, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 545, + 324 + ], + "type": "text", + "content": "tors. (ii) Audio Semantic Guidance Module, which uses the above-mentioned latent vectors as a condition in the diffusion process to generate corresponding image outputs. We apply identity regularizer to produce temporally consistent video frames, while we apply audio semantic guidance to generate audio-reactive video frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 325, + 546, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 546, + 456 + ], + "type": "text", + "content": "However, we observe that generating multifarious high-quality images solely from a sound input is challenging due to the lack of such a large-scale dataset to train a model. Instead, we first generate an initial frame using pre-trained Stable Diffusion model with a text prompt, then generate the following video frames conditioned on audio inputs. This frees a model from data dependency burdens and enables training with the current relatively small (than imagetext modality) audio dataset [19, 5], focusing on leveraging temporal semantics from audio. We summarize our contributions as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 458, + 545, + 642 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 317, + 458, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 458, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 317, + 458, + 545, + 493 + ], + "type": "text", + "content": "- We propose a novel sound-driven video generation method built upon Stable Diffusion [29] and can generate video frames reactively with audio sequence inputs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 496, + 544, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 496, + 544, + 543 + ], + "spans": [ + { + "bbox": [ + 317, + 496, + 544, + 543 + ], + "type": "text", + "content": "- Our attention-based Audio Encoder produces temporally-aware latent vectors, which are consumed by Stable Diffusion as a per-time manipulation condition, producing audio-reactive video frames." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 546, + 544, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 544, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 544, + 581 + ], + "type": "text", + "content": "- Our model regularizes the latent features of diffusion models to produce temporally consistent video frames, preserving identity throughout the generated video." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 584, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 584, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 317, + 584, + 545, + 642 + ], + "type": "text", + "content": "- We demonstrate the effectiveness of our proposed model using a public dataset Landscape [19], generally outperforming other state-of-the-art sound-driven video generation approaches in terms of video quality metrics and human evaluation." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 653, + 392, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 653, + 392, + 665 + ], + "spans": [ + { + "bbox": [ + 306, + 653, + 392, + 665 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Latent Diffusion Models. Recent success [29] suggests that the Latent Diffusion Models (LDM) improve the efficiency of the diffusion process, successfully generating" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 72, + 242, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 242, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 242, + 83 + ], + "type": "text", + "content": "at https://ku-vai.github.io/TPoS/" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7823" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 125, + 287, + 206 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 119 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 119 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 119 + ], + "type": "text", + "content": "Table 1: Comparison between existing state-of-the-art audio-driven video generation approaches in terms of whether they consider the following factors: temporal semantics, magnitude changes of sound, and target domains." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 125, + 287, + 206 + ], + "lines": [ + { + "bbox": [ + 48, + 125, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 125, + 287, + 206 + ], + "type": "table", + "html": "
ModelTemporal SemanticsMagnitudeDomains (Audio Type)
Sound2Sight [4]-✓Closed
CCVS [18]-✓Closed (Music)
TräumerAI [13]-✓Closed (Music)
Lee et al. [19]✓-Closed (Nature)
Ours✓✓Open Domains
", + "image_path": "e65add0d7e1af0a8137060663caa20c029d1b79bb493eb70159f0c38dee6f0af.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 213, + 287, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 287, + 381 + ], + "type": "text", + "content": "high-quality images given a text prompt. One challenge in LDM is that the generation process is too sensitive to the condition, making it difficult to control semantics. Recently, there have been introduced to control semantics with LDM by a semantic mask [1] or by utilizing semantic information of the cross-attention layers [9]. Wu et al. [37] used linear combinations of text embeddings and Liu et al. [22] proposed composable diffusion models, but they still remained challenging to control fine-grained semantic changes. Recently, Semantic Guidance (SEGA) [2] computed a guidance vector in the latent space, enabling semantic control of diffusion models without further inputs. Inspired by SEGA, we also control the semantics in the latent space with temporally-encoded audio vector sequences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 385, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 553 + ], + "type": "text", + "content": "Text-driven Video Generation. Recent text-to-video generation tools, including Make-A-Video [32], Video Diffusion Models [11], Imagen video [10], and Phenaki [36] have shown promising performance in generating videos from textual descriptions. However, text-to-video generation has its limitations in terms of temporal coherence, which mostly leads to short video duration or a linear video change. Recent text-to-video generation methods, StyleGAN-V [34] and Dreamix [23], made progress addressing these issues. However, conditioning temporal semantics or complex scenarios is still challenging to be obtained from text inputs. Thus, in this paper, we want to explore conditioning a model with audio inputs, which inherently convey such temporal semantics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": "Audio-driven Video Generation. Leveraging temporal semantics was not seriously considered in previous audio-driven video generation approaches. Sound2Sight[4] and CCVS [18] generate video frames conditioned on the (nontemporal) context of the given audio, while TraumerAI [13] utilized the magnitude of the given audio. Recently, Lee et al. [19] explored a model that can consider audio semantics as a condition to drive a video generator. Also, their dependency on StyleGAN [34]-based embedding space makes it difficult for models to generate transitions in video. In this work, we focus on leveraging temporal semantics from audio inputs such that our generator reactively manipulates video frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 71, + 361, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 361, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 361, + 84 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 236 + ], + "type": "text", + "content": "In this work, we propose a novel audio-driven video generation method, which generates video frames conditioned on audio sequences, ensuring temporal consistency between consecutive frames and temporal correspondence with audio inputs. As shown in Figure 3, our model consists of two main parts: (i) Audio Encoder, which encodes temporal semantics of audio sequences, producing a sequence of the latent vectors (Section 3.2). (ii) Audio Semantic Guidance Module, which uses the above-mentioned latent vectors as a condition in the diffusion process to generate corresponding image outputs, which are temporally consistent (by our identity regularizer) and audio-reactive (see Section 3.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 243, + 503, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 503, + 256 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 503, + 256 + ], + "type": "text", + "content": "3.1. Preliminary: Latent Stable Diffusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "content": "Latent Diffusion Models (LDMs) [29] are the method that uses an encoder to convert a noised latent vector " + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^T" + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "content": " to a denoised latent vector " + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\mathbf{x} + \\epsilon" + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "content": " is a latent vector of an input image " + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "content": " is a noise. Stable Diffusion [29] is part of a conditional generation model that can synthesize an image given a condition " + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "content": ". It uses U-Net [30] as denoising autoencoders represented by " + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "content": ". To generate an output image, the autoencoder " + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}^{t}, t, \\tau_{\\theta}(\\mathbf{y}))" + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "content": " takes three inputs: the noised latent vector " + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^t" + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "content": ", a sequence " + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "content": " that is uniformly sampled from the set " + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "inline_equation", + "content": "1, \\dots, T" + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "content": ", and a conditional input " + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 262, + 545, + 382 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "text", + "content": "Conditional input " + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "text", + "content": " is first transformed into a latent vector " + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_p" + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "text", + "content": " through a pretrained function " + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}" + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "text", + "content": " and then it is fed into a cross-attention layer of the U-Net as the key and value, which is then combined with " + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^t" + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "text", + "content": " by an attention mechanism, where the query is the flattened intermediate representation of the U-Net. Next, the denoising autoencoder " + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "text", + "content": " is used to denoise " + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^t" + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "text", + "content": " sequentially. After " + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "text", + "content": " denoising steps, the resulting denoised latent vector " + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{z}(= \\mathbf{z}^{1})" + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "text", + "content": " is transmitted to the decoder, which produces the final output image " + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 383, + 545, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 510, + 528, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 510, + 528, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 510, + 528, + 525 + ], + "type": "text", + "content": "3.2. Encoding Temporal Semantics from Audio" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 529, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 613 + ], + "type": "text", + "content": "We use an audio modality as a source of generating temporal conditions. We divide the audio mel-spectogram into uniform snippets (segments) to extract abundant audio features and feed this into the Audio Encoder to extract both temporal semantic features and intensity of audio in end-to-end manner. The figure of our training process is illustrated in Figure 4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": "Audio Feature Extraction. Audio inputs are first transformed into a mel-spectrogram representation, denoted as " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^a\\in \\mathbb{R}^{d\\times w}" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": " represents the number of mel-frequency bins and " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": " is the width of the spectrogram. To incorporate time information, the mel-spectrogram is divided into " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": " segments. Each segment, denoted as " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_n^a\\in \\mathbb{R}^{d\\times \\lceil \\frac{w}{N}\\rceil}" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "inline_equation", + "content": "n\\in \\{1,\\ldots ,N\\}" + }, + { + "bbox": [ + 304, + 617, + 545, + 714 + ], + "type": "text", + "content": ", is then fed into a shared feature extraction module, i.e., the pre-trained ResNet18 [8]. The" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7824" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 72, + 521, + 301 + ], + "blocks": [ + { + "bbox": [ + 75, + 72, + 521, + 301 + ], + "lines": [ + { + "bbox": [ + 75, + 72, + 521, + 301 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 521, + 301 + ], + "type": "image", + "image_path": "8bd59a99cce9913cd2bf813cf7788bc82765c429d1f8cd78bc500d99afda4e5b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 310, + 547, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 310, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 547, + 361 + ], + "type": "text", + "content": "Figure 3: An overview of our proposed TPoS model. Our model consists of two main modules: (i) Audio Encoder, which produces a sequence of latent vectors, encoding temporal semantics of audio input by utilizing CLIP [27] space and highlighting the important temporal features and (ii) Audio Semantic Guidance Module, which is based on the diffusion process, generating video frames that are temporally consistent and audio-reactive." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 369, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 288, + 407 + ], + "type": "text", + "content": "feature extraction module " + }, + { + "bbox": [ + 46, + 369, + 288, + 407 + ], + "type": "inline_equation", + "content": "f_{a}(\\cdot)" + }, + { + "bbox": [ + 46, + 369, + 288, + 407 + ], + "type": "text", + "content": " learns to extract low-level features from each audio segment regardless of its time dependency, i.e., " + }, + { + "bbox": [ + 46, + 369, + 288, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_n = f_a(\\mathbf{x}_n^a)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": "LSTM-based Temporal Semantic Encoder. As our goal is to generate audio-reactive video frames, it is also important to encode temporal changes (or relations) of the given audio inputs. Similar to Lee et al. [19], given audio features " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\{\\mathbf{w}_1, \\mathbf{w}_2, \\dots, \\mathbf{w}_N\\}" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": ", which encodes per-segment disjoint audio representation, we apply the standard Long Short-Term Memory (LSTM) network [12] to encode temporal relations or changes between consecutive audio features " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": ". Formally, our LSTM takes the audio feature " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{n-1}" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": " as input and updates its hidden state " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_n" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": ", producing an output " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": ": i.e. " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "(\\mathbf{s}_n, \\mathbf{h}_n) = \\mathrm{LSTM}(\\mathbf{h}_{n-1}, \\mathbf{w}_{n-1})" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": ". These outputs " + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\{\\mathbf{s}_1, \\mathbf{s}_2, \\dots, \\mathbf{s}_N\\}" + }, + { + "bbox": [ + 46, + 411, + 287, + 567 + ], + "type": "text", + "content": " are then fed into Temporal Attention Module (TAM) to further encode temporal semantics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 571, + 288, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 288, + 691 + ], + "type": "text", + "content": "Aligning Audio Semantics with Image-Text CLIP Joint Space. As we will use the output " + }, + { + "bbox": [ + 46, + 571, + 288, + 691 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 571, + 288, + 691 + ], + "type": "text", + "content": " as a condition to manipulate video frames, it is important to ensure those audio features are well-aligned with other text and visual features in the CLIP [27]-based joint embedding space. Similar to Lee et al. [20], given the pre-trained image-text CLIP space, we apply the following loss " + }, + { + "bbox": [ + 46, + 571, + 288, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CLIP}}^{a \\leftrightarrow t}" + }, + { + "bbox": [ + 46, + 571, + 288, + 691 + ], + "type": "text", + "content": " with the InfoNCE loss [24] " + }, + { + "bbox": [ + 46, + 571, + 288, + 691 + ], + "type": "inline_equation", + "content": "l_{sim}" + }, + { + "bbox": [ + 46, + 571, + 288, + 691 + ], + "type": "text", + "content": " such that positive pairs (e.g. an audio of raining and a text prompt \"raining\") are pulled close to each other, while negative pairs are pushed farther away." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C L I P}} ^ {a \\leftrightarrow t} = l _ {s i m} (\\mathbf {s} _ {N}, \\operatorname {C L I P} _ {t} (\\mathbf {t})) + l _ {s i m} (\\operatorname {C L I P} _ {t} (\\mathbf {t}), \\mathbf {s} _ {N}) \\quad (1)", + "image_path": "297969396136b54800ea4d538168cb808b93eb4771a95705a35b1eb75db10878.jpg" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 369, + 542, + 549 + ], + "blocks": [ + { + "bbox": [ + 309, + 369, + 542, + 549 + ], + "lines": [ + { + "bbox": [ + 309, + 369, + 542, + 549 + ], + "spans": [ + { + "bbox": [ + 309, + 369, + 542, + 549 + ], + "type": "image", + "image_path": "618663442c550a1cca5f593d6e0c47315cd9240b4ef34270ff1e8a8c5721ced2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "text", + "content": "Figure 4: An overview of our Audio Encoder training process. Our model generates temporally-encoded audio embeddings with an LSTM [12] layer and Temporal Attention Module (TAM). Audio input is partitioned into " + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 558, + 545, + 665 + ], + "type": "text", + "content": " segments, and each of these is encoded and used as a condition to manipulate audio-reactive video sequences (e.g. light rain → heavy rain). This is done by our Mapping Module (MAP), which maps the audio embedding to the latent space of Stable Diffusion." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}_t" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " is a pre-trained CLIP-based text encoder, which takes a text prompt " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " obtained from audio class labels as input, yielding an " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "-dimensional feature. Note that" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7825" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "we only apply this loss for the final output " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_N" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " for efficient training. Given a set of positive pairs, we apply the following InfoNCE loss " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "l_{sim}(\\mathbf{a},\\mathbf{b})" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 96, + 114, + 287, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 114, + 287, + 143 + ], + "spans": [ + { + "bbox": [ + 96, + 114, + 287, + 143 + ], + "type": "interline_equation", + "content": "l _ {s i m} = - \\log \\frac {\\exp \\left(\\left\\langle \\mathbf {a} _ {i} , \\mathbf {b} _ {i} \\right\\rangle\\right) / \\tau}{\\sum_ {j} \\exp \\left(\\left\\langle \\mathbf {a} _ {i} , \\mathbf {b} _ {j} \\right\\rangle\\right) / \\tau} \\tag {2}", + "image_path": "29b0c92254b3d354d1366f3647e42d80dafaf3847c1dc2c2f0e5aaf8be91a7ed.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 148, + 287, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 148, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 287, + 172 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 148, + 287, + 172 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{a}_i,\\mathbf{b}_i\\rangle" + }, + { + "bbox": [ + 47, + 148, + 287, + 172 + ], + "type": "text", + "content": " represents the cosine similarity with temperature " + }, + { + "bbox": [ + 47, + 148, + 287, + 172 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 148, + 287, + 172 + ], + "type": "text", + "content": ". Note that we set " + }, + { + "bbox": [ + 47, + 148, + 287, + 172 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 148, + 287, + 172 + ], + "type": "text", + "content": " to 0.07." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 176, + 287, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 176, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 287, + 274 + ], + "type": "text", + "content": "Augmenting Audio Semantics. Audio data is often limited in volume and diversity; thus, augmentation techniques may be required to extract better-quality audio semantic features, preventing a representation collapse. We use SpecAugment [25] to apply random transformations (such as masking our certain frequency bands or time segments), yielding augmented audio inputs. We further add the InfoNCE loss [24] " + }, + { + "bbox": [ + 47, + 176, + 287, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CLIP}}^{a\\leftrightarrow a^{\\prime}}" + }, + { + "bbox": [ + 47, + 176, + 287, + 274 + ], + "type": "text", + "content": " to pull augmented audio features together." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 279, + 287, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 279, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 85, + 279, + 287, + 295 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C L I P}} ^ {a \\leftrightarrow a ^ {\\prime}} = l _ {s i m} \\left(\\mathbf {s} _ {N}, \\mathbf {s} _ {N} ^ {\\prime}\\right) + l _ {s i m} \\left(\\mathbf {s} _ {N} ^ {\\prime}, \\mathbf {s} _ {N}\\right) \\tag {3}", + "image_path": "c3adf69a5a5ce350cf37bdf2ab6fcaebdb4d52c771cd2f8f15e1834d6ad1b4f7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 302, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 302, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 302, + 287, + 338 + ], + "type": "text", + "content": "where apostrophe indicates augmented view of an original audio data. Finally, we use the semantic loss " + }, + { + "bbox": [ + 47, + 302, + 287, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{semantic}}" + }, + { + "bbox": [ + 47, + 302, + 287, + 338 + ], + "type": "text", + "content": ": i.e. " + }, + { + "bbox": [ + 47, + 302, + 287, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{semantic}} = \\mathcal{L}_{\\mathrm{CLIP}}^{a\\leftrightarrow t} + \\lambda_s\\mathcal{L}_{\\mathrm{CLIP}}^{a\\leftrightarrow a'}" + }, + { + "bbox": [ + 47, + 302, + 287, + 338 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 302, + 287, + 338 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 47, + 302, + 287, + 338 + ], + "type": "text", + "content": " is set to 0.6." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "text", + "content": "Temporal Attention Module (TAM). As shown in Figure 5, we further use an attention-based module to encode temporal semantics from the audio inputs. We empirically observe that adding this module helps improve the quality of video frame generation, which is probably due to the fact that the model becomes flexible to focus on more important temporal information, improving the model's representation power. Formally, we first compute attention weight " + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\alpha_{n}" + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "text", + "content": " for a given audio feature " + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_n" + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "text", + "content": " by applying an MLP layer " + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{proj}}" + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "text", + "content": " followed by a softmax operation: i.e. " + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\alpha_{n} = \\exp (f_{\\mathrm{proj}}(\\mathbf{s}_{n})) / \\sum_{n}\\exp (f_{\\mathrm{proj}}(\\mathbf{s}_{n}))" + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\sum_{n}\\alpha_{n} = 1" + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "text", + "content": ". We compute the weighted sum of audio features based on attention weights, yielding an attended audio feature " + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^a = \\sum_n\\alpha_n\\mathbf{s}_n" + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "text", + "content": ". We add another InfoNCE loss " + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{temporal}}" + }, + { + "bbox": [ + 47, + 341, + 287, + 521 + ], + "type": "text", + "content": " to align the audio features with text:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 528, + 287, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 287, + 552 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t e m p o r a l}} = l _ {s i m} \\left(\\mathbf {o} ^ {a}, \\operatorname {C L I P} _ {t} (\\mathbf {t})\\right) + l _ {s i m} \\left(\\operatorname {C L I P} _ {t} (\\mathbf {t}), \\mathbf {o} ^ {a}\\right) \\tag {4}", + "image_path": "47242d58783f688db950d45154ac886f02065b4ede8a8abc21a7b55cd5edf1fa.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "text", + "content": "Lastly, we also minimize the MSE loss between text embeddings (before the projection layer) and the projected audio feature " + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "inline_equation", + "content": "\\mathrm{MAP}(\\mathbf{o}^a)" + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cond}} = ||\\mathbf{c}_t - \\mathrm{MAP}(\\mathbf{o}^a)||_2^2" + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "text", + "content": ". Note that we exclude the feature for the " + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "inline_equation", + "content": "<\\mathrm{SOS}>" + }, + { + "bbox": [ + 47, + 552, + 287, + 612 + ], + "type": "text", + "content": " token. More details are provided in the supplemental material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 616, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 287, + 640 + ], + "type": "text", + "content": "Total Loss. We train our Audio Encoder end-to-end by minimizing the following loss function " + }, + { + "bbox": [ + 47, + 616, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 47, + 616, + 287, + 640 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 100, + 647, + 287, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 647, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 100, + 647, + 287, + 661 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {s e m a n t i c}} + \\mathcal {L} _ {\\text {t e m p o r a l}} + \\mathcal {L} _ {\\text {c o n d}} \\tag {5}", + "image_path": "5b90af6a762210308ac15f45ba7057564bc01a768d5ecd12d408f5caf68ea908.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 667, + 287, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 667, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 47, + 667, + 287, + 679 + ], + "type": "text", + "content": "3.3. Generating Video Frames with Stable Diffusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Initial Frame Generation from Text Prompt. Our model relies on leveraging combinational operations of denoising" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 541, + 155 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 541, + 155 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 541, + 155 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 541, + 155 + ], + "type": "image", + "image_path": "1e0b278d79c4a3a16687e44316599baa4b254eb4053bb14cc99223f74d5fc4a5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 164, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 305, + 164, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 305, + 164, + 545, + 236 + ], + "type": "text", + "content": "Figure 5: Details of our Temporal Attention Module (TAM). Temporal Attention Module effectively captures the important audio segments (e.g. harsh raining sound) as attention weights, which are guided by CLIP [27]'s text embedding (e.g. \"raining\") in training phase and express the magnitude of audio in test phase." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 247, + 546, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 546, + 415 + ], + "type": "text", + "content": "process. The estimated noise space can be either random or manual by visual input with the help of the diffusion encoder. Based on this guided diffusion, our model first generates the initial frame with a text prompt (e.g. \"a photo of beautiful beach with blue sky\"). Given this generated image as content, we manipulate its styles according to audio inputs and generate corresponding video frames, i.e. given a series of latent vectors " + }, + { + "bbox": [ + 304, + 247, + 546, + 415 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_1,\\mathbf{c}_2,\\dots ,\\mathbf{c}_m\\}" + }, + { + "bbox": [ + 304, + 247, + 546, + 415 + ], + "type": "text", + "content": ", we generate " + }, + { + "bbox": [ + 304, + 247, + 546, + 415 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 247, + 546, + 415 + ], + "type": "text", + "content": " video frames. Note that the number of video frames is controllable by latent vector interpolation, as we will explain later. We follow the standard image generation process with the Stable Diffusion model [29], i.e., we compute a latent vector " + }, + { + "bbox": [ + 304, + 247, + 546, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_p" + }, + { + "bbox": [ + 304, + 247, + 546, + 415 + ], + "type": "text", + "content": " in a CLIP-based embedding space given a text prompt, conditioning it to generate an image." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "content": "Audio Semantic Guidance. We employed the SEGA [2] framework to create video frames that incorporate sound style while preserving the main content identity. We first utilize the combination of attention " + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "inline_equation", + "content": "\\alpha_{n}" + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "content": " and audio feature " + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_n" + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "content": " with normalization of the scale of output feature by multiplying " + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "content": " to produce output " + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_n" + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "content": ": i.e., " + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_n = N^k\\alpha_n^k\\mathbf{s}_n" + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "content": " is hyper parameter that regulates attention ratio and is set to 1. The Audio Semantic Guidance module generates the " + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "content": "-th video frame by taking audio condition " + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_n" + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "content": " as input to guide the diffusion models. The denoising autoencoder " + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "content": " is executed " + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "inline_equation", + "content": "\\delta - 1" + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "content": " times out of " + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "content": " time to form incomplete noise along with the original text prompt meaning. From " + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "inline_equation", + "content": "t = \\delta" + }, + { + "bbox": [ + 304, + 418, + 546, + 586 + ], + "type": "text", + "content": ", the audio semantic guidance operates through the following equation:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 589, + 545, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 589, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 310, + 589, + 545, + 619 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {p}\\right) := \\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {\\varnothing}\\right) + g \\left(\\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {p}\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {\\varnothing}\\right)\\right) \\tag {6} \\\\ + \\lambda (\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {n}) \\\\ \\end{array}", + "image_path": "f5371d52a32d00c6363057fc184bd2fabcd3ef06bb9af524494ecfefb47f5edc.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\delta}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is denoised latent vector at " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "t = \\delta" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is the guidance scale of the text prompt, " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda(\\mathbf{z}^{\\delta}, \\mathbf{c}_n)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is the audio semantic guidance term, and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\emptyset}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " represents an unconditioned prompt that does not make any semantic difference. As a result, only the " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda(\\mathbf{z}^{\\delta}, \\mathbf{c}_n)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " term has been added to the original denoising process from " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "t = \\delta" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^T" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " is fixed through frames in one video. The audio semantic guidance" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7826" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 103, + 507, + 213 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 2: Comparison of the quality of generated video frames with state-of-the-art audio-to-video generations in terms of IS [31], FVD [35], and CLIP [27]-based distances." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 103, + 507, + 213 + ], + "lines": [ + { + "bbox": [ + 86, + 103, + 507, + 213 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 507, + 213 + ], + "type": "table", + "html": "
ModelInputIS↑FVD↓CLIP↑ (a ↔ v)CLIP↑ (t ↔ v)
Sound2Sight [4]1st Frame1.02 ± 0.02494.280.03640.2164
CCVS [18]1st Frame1.30 ± 0.20679.940.12510.2360
TräumerAI [13]-1.47 ± 0.19736.320.15890.1778
Sound-guided Video Generation [19]-1.16 ± 0.16544.090.11510.1702
Ours (w/o TAM)Latent Vector1.27 ± 0.23483.760.13420.2370
OursLatent Vector1.49 ± 0.38421.230.19640.2436
", + "image_path": "5115286525a5b4ef32eb2f277fb5190dd65786aac2908ab750ea98a6bd89249e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 215, + 301, + 469 + ], + "blocks": [ + { + "bbox": [ + 52, + 215, + 301, + 469 + ], + "lines": [ + { + "bbox": [ + 52, + 215, + 301, + 469 + ], + "spans": [ + { + "bbox": [ + 52, + 215, + 301, + 469 + ], + "type": "image", + "image_path": "b1f2669aa345167bf55ac69a84dbf72eb48948130c331cbab85748f39e5c4c73.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 477, + 546, + 504 + ], + "lines": [ + { + "bbox": [ + 46, + 477, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 477, + 546, + 504 + ], + "type": "text", + "content": "Figure 6: Examples of generated video frames (given fire crackling and waterfall audio) by Sound2Sight [4], CCVS [18], TraumerAI [13], Lee et al. [19], and ours. Note that Sound2Sight and CCVS use an initial frame (highlighted in a red box)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 215, + 537, + 469 + ], + "blocks": [ + { + "bbox": [ + 302, + 215, + 537, + 469 + ], + "lines": [ + { + "bbox": [ + 302, + 215, + 537, + 469 + ], + "spans": [ + { + "bbox": [ + 302, + 215, + 537, + 469 + ], + "type": "image", + "image_path": "512f567d03735f7645ce65a57e475ef21a3b012b9c50a3bfe6f375118ce23eed.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 517, + 176, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 176, + 531 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 176, + 531 + ], + "type": "inline_equation", + "content": "\\lambda (\\mathbf{z}^{\\delta},\\mathbf{c}_n)" + }, + { + "bbox": [ + 47, + 517, + 176, + 531 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 548, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 287, + 563 + ], + "type": "interline_equation", + "content": "\\lambda \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {n}\\right) = g _ {s} \\left(\\epsilon_ {\\theta} \\left(\\mathbf {z} ^ {\\delta}, \\mathbf {c} _ {n}\\right) - \\epsilon_ {\\theta} \\left(\\mathbf {z} _ {\\varnothing} ^ {\\delta}, \\mathbf {c} _ {\\varnothing}\\right)\\right) + \\sigma_ {m} \\Phi_ {m} \\tag {7}", + "image_path": "31a45218f0c79fbc3d072fb487dee5267b178171b1623a8100fee8ff3ff502b9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\sigma_{m} \\in [0,1]" + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": " is the momentum hyper parameter that scales the momentum " + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\Phi_{m}" + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "inline_equation", + "content": "g_{s}" + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": " focuses on the relevant dimensions to the audio manipulation task (More details can be found in the supplemental material or SEGA [2]). Different from SEGA framework, we only consider positive guidance of audio semantics, which is the semantic difference between the guidance provided by " + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_n" + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": " and the unconditioned guidance " + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\emptyset}" + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "The generation process conditioned by " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_n" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is separately working through diffusion processes so that the different semantic meaning or magnitude which presents in " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_n" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " can generate frames independently." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 518, + 547, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 615 + ], + "type": "text", + "content": "Temporal Frame Interpolation. We use an interpolated latent vector to generate continuous video frames between two consecutive frames. Following the work by Ramesh et al. [28], we apply a spherical linear interpolation between all consecutive pairs of " + }, + { + "bbox": [ + 304, + 518, + 547, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_n" + }, + { + "bbox": [ + 304, + 518, + 547, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 518, + 547, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{n + 1}" + }, + { + "bbox": [ + 304, + 518, + 547, + 615 + ], + "type": "text", + "content": ", yielding " + }, + { + "bbox": [ + 304, + 518, + 547, + 615 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 518, + 547, + 615 + ], + "type": "text", + "content": " interpolated latent vectors. These vectors are then used as a condition for the diffusion models to generate temporally-interpolated video frames." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 628, + 388, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 628, + 388, + 642 + ], + "spans": [ + { + "bbox": [ + 306, + 628, + 388, + 642 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Datasets. We use two Audio-Video datasets to train our Audio Encoder: VGG-Sound [5] and Landscape [19]. VGG-Sound is an audio dataset with about 170,000 of 10-second clips of audio-video data, which consists of 309 classes. The dataset has numerous 'in the wild' audio data that spans" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7827" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 168 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 168 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 168 + ], + "type": "image", + "image_path": "d2afda1fe476ecb35dce1039f693b49994eed5bec975826a98400f58641bfea7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 287, + 202 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 287, + 202 + ], + "type": "text", + "content": "Figure 7: Examples of our video generation conditioned on text prompt (top) and audio (bottom)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 304 + ], + "type": "text", + "content": "a large number of challenging acoustic environments and real application noise characteristics. Since audio of nature sound is the perfect tool to stylize compared to the class such as people talking or sports, we add about 9,000 audio clips of Landscape audio dataset in the training process. For obtaining test sets, class-balanced sampling is applied to Landscape dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 311, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 287, + 515 + ], + "type": "text", + "content": "Baselines. We compare our methods with existing audio-to-video generation methods, Sound2Sight [4], CCVS [18], and StyleGAN [34] based TraumerAI [13] and Sound-guided Video Generation [19]. We follow experiment methods by Lee et al. [19]. All baselines are trained or finetuned with Landscape dataset [19]. Since Sound2Sight and CCVS need first frame to generate the video, we randomly select first frame from the Landscape dataset at the inference task. For TraumerAI and Sound-guided Video Generation, we first pre-train StyleGAN [16] on high fidelity benchmark datasets (LHQ datasets [33]) and then train the model to navigate the latent space by Landscape dataset. Note that randomly initialized vector for TraumerAI [13] and Sound-guided Video Generation [19] are given. For our model, we initially generate image from random noise space and randomly sample the prompt related with landscape to generate the landscape like video for fair comparison." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. We use the following two video quality metrics for our evaluations: (i) Fréchet Video Distance (FVD) [35] and (ii) Inception Score (IS) [31]. FVD is used to assess video quality by measuring the distribution gap between real and synthesized videos in the latent space. Additionally, IS score is commonly employed to evaluate the effectiveness of Generative Adversarial Networks (GANs) by computing the KL-divergence between the label distribution of each image and the marginal label distribution. To implement these, we fine-tune Inflated 3D ConvNet [3] with Landscape [19] dataset for FVD and used pre-trained InceptionNet [17] that is trained on ImageNet [6] dataset. We also measure CLIP [27]-based cosine similarity (CLIP score) between audio and image as well as text and image. To obtain the textual pivot feature, we fed the following prompt \"The photo of \" into CLIP text encoder." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 542, + 212 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 542, + 212 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 542, + 212 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 542, + 212 + ], + "type": "image", + "image_path": "c4ef6f1b0b7e2a72cee27f6ec681d429b4a2754873d2080cc0f031cadd6e3bd4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 269 + ], + "type": "text", + "content": "Figure 8: Generated video frames with and without our Temporal Attention Module. We generate video frames with splashing water sound where its amplitude changes over time (see top)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 283, + 546, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 546, + 426 + ], + "type": "text", + "content": "Quantitative Experiments. For fair comparison with other existing baselines such as TraumerAI [13] and Sound-guided Video Generation [19] which does not get a hint about what to generate, we use prompt that originally does not generate the style of sound. We set fps 20 and generated videos to extract images from all baselines. Table 2 shows that our approach produces the best quality results as video. Additionally, to ensure that the generated videos are semantically related to the sound, we compare the cosine similarity between text-audio and video embedding. Our methods show a superior performance in terms of multimodal semantics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 432, + 546, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 546, + 624 + ], + "type": "text", + "content": "Qualitative Video Quality Comparison. We first evaluate the quality of generated video frames. In Figure 6, we provide typical examples of generated video frames by (from top) Sound2Sight [4], CCVS [18], TraumerAI [13], Lee et al. [19], and ours. Note that Sound2Sight [4] and CCVS [18] need the initial frame as input (see red boxes). With audio inputs, such as fire crackling and waterfall, as a condition, ours generally generate temporally-consistent audio-reactive video frames. We observe that Sound2Sight [4] and CCVS [18] often show blurring artifacts, while StyleGAN-based TraumerAI [13] and Sound-guided Video Generation [19] often fail to generate audio-reactive video frames, e.g., they produce landscape scenes that are not semantically aligned with a fire crackling audio. However, ours generate a scene of a waterfall or a fire on firewood, aligning well with audio inputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "content": "Comparison to Video Generation with Text. To analyze the benefit of audio modality, we conduct a qualitative experiment to compare the effect of audio and text modalities in generating visual content. As shown in Figure 7, we first generate an initial frame with the text prompt \"A photo of deep in the sea.\" Then we generate the next frames with text \"underwater bubbling\" (top) and underwater bubbling" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7828" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 286, + 171 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 286, + 171 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 286, + 171 + ], + "type": "image", + "image_path": "bc8e924b9f3c71ba8146161b9ce4b01d892f87a41249f02e5b0cd779e84419f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 220, + 287, + 358 + ], + "blocks": [ + { + "bbox": [ + 46, + 181, + 287, + 217 + ], + "lines": [ + { + "bbox": [ + 46, + 181, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 287, + 217 + ], + "type": "text", + "content": "Figure 9: Generated video frames with and without our Audio Semantic Guidance. Edited parts by Audio Semantic Guidance are highlighted with red box (see second row)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 220, + 287, + 358 + ], + "lines": [ + { + "bbox": [ + 47, + 220, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 287, + 358 + ], + "type": "image", + "image_path": "a31f840981aad8318e349338506c53786970c0b37a5263a395b88c00d282b49e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 369, + 287, + 405 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 405 + ], + "type": "text", + "content": "Figure 10: Examples of generated video frames with a sound that changes over time (e.g. bird singing " + }, + { + "bbox": [ + 46, + 369, + 287, + 405 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 369, + 287, + 405 + ], + "type": "text", + "content": " wind noise)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "type": "text", + "content": "sound (bottom). It is difficult to make temporal changes conditioned on text unless we train our model with a text-video dataset. Thus, we instead change the noise scale to make temporal changes, preserving identity. However, as shown in Figure 7 (top), it generates distortions or a linear change, but this is not the case for audio. Our model with audio generates visually-appealing video frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "text", + "content": "Effect of Temporal Attention Module. We use Temporal Attention Module (TAM) to improve the representation power to encode temporal semantics better. To analyze this, we perform an ablation study with and without TAM to see its effect on video frame generation. We observe in Figure 8 that our model is indeed able to generate video frames reactive to the audio changes over time (compare the changes along with the given splashing water sound)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Effect of Audio Semantic Guidance. We demonstrate the effect of Audio Semantic Guidance with ablation study. First, we generate video frames without Audio Semantic Guidance by replacing audio semantic guidance term " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lambda (\\mathbf{z}^{\\delta},\\mathbf{c}_n)" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{z}^{\\delta},\\mathbf{c}_n)" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " (refer the notation in Section 3.3). To remove the effect of Audio Semantic Guidance, we also set " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": ". As illustrated in Figure 9 (top), without Audio Semantic Guidance, the model has a tendency to produce unnecessary changes (e.g. grass " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " toil) and struggles to" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 73, + 539, + 190 + ], + "blocks": [ + { + "bbox": [ + 309, + 73, + 539, + 190 + ], + "lines": [ + { + "bbox": [ + 309, + 73, + 539, + 190 + ], + "spans": [ + { + "bbox": [ + 309, + 73, + 539, + 190 + ], + "type": "image", + "image_path": "14155dfec2472ad9982bf798bb225bb34574ba0c1f2134444c4083891903dc6a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 201, + 547, + 274 + ], + "lines": [ + { + "bbox": [ + 304, + 201, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 547, + 274 + ], + "type": "text", + "content": "Figure 11: Our human evaluation results. We conduct a user study with 100 participants on Amazon Mechanical Turk (AMT). Participants are shown generated video frames and asked to evaluate them in terms of realistic, vividness, consistency, and relevance. The Likert scale is used (higher is better)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 292, + 546, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 546, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 546, + 388 + ], + "type": "text", + "content": "achieve consistent alignment with audio (e.g. producing a road with a waterfall burbling sound). On the other hand, by leveraging Audio Semantic Guidance as in Figure 9 (bottom), we can generate sequential video frames that not only has consistent content but also represents natural temporal variations according to audio sound (e.g. producing wave of water with waterfall burbling sound), resulting in enhanced naturalness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 392, + 546, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 546, + 477 + ], + "type": "text", + "content": "Experiments of Semantic Transition. In Figure 10, we provide examples where audio inputs are changed (e.g., bird singing " + }, + { + "bbox": [ + 304, + 392, + 546, + 477 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 392, + 546, + 477 + ], + "type": "text", + "content": " wind noise, strong wind " + }, + { + "bbox": [ + 304, + 392, + 546, + 477 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 392, + 546, + 477 + ], + "type": "text", + "content": " raining). As we observe in that figure, our model successfully adapts to the audio change, generating video frames accordingly. This may confirm that our model is indeed conditioned on the audio sequence and can generate audio-adaptive video frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 481, + 546, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 546, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 546, + 636 + ], + "type": "text", + "content": "User study. Further, we conduct a human evaluation to evaluate the video quality by human judges. We recruit 100 participants from Amazon MTurk. Participants are shown video frames generated by five different audio-driven video generation models: Sound2Sight [4], CCVS [18], TraumerAI [13], Lee et al. [19], and ours. Participants are asked to evaluate the given video frames in terms of realism, vividness, consistency, and relevance. We use a five-point Likert scale where ideal video frames will get all five points. More details are provided as supplemental material. We observe in Figure 11 that our proposed method outperforms the other approaches in all categories. These results are consistent with our quantitative and qualitative results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Visual Conditioning. Our model is capable of producing an initial frame using visual input (such as images) with the trained diffusion encoder. Figure 12 demonstrates that our framework leverages latent space of a diffusion model with a combination of visual input, text prompt and audio. We first randomly choose visual input (e.g., sun light photo) and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7829" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 274, + 178 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 274, + 178 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 274, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 274, + 178 + ], + "type": "image", + "image_path": "656558e23f2214679b30af002468d997bd4c1aa89f305ed7c485db50619ac55d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 183, + 287, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 287, + 232 + ], + "type": "text", + "content": "Figure 12: Example of generated video with visual input, text prompt and audio sound. (e.g., 1st row: sun light photo conditioned by text \"A photo of beautiful garden with sky\" and splashing water sound.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 251, + 282, + 439 + ], + "blocks": [ + { + "bbox": [ + 50, + 251, + 282, + 439 + ], + "lines": [ + { + "bbox": [ + 50, + 251, + 282, + 439 + ], + "spans": [ + { + "bbox": [ + 50, + 251, + 282, + 439 + ], + "type": "image", + "image_path": "aaec0af009da9ac6088f5700107fa276d1e594124cf93120e03698bf387fbb79.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 449, + 287, + 485 + ], + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 485 + ], + "type": "text", + "content": "Figure 13: Example of generated videos with audio-text joint condition (e.g., 2nd row: conditioned with text \"eruption\" and explosion sound)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 499, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 571 + ], + "type": "text", + "content": "generate an initial frame with the text prompt (e.g., \"A photo of beautiful garden with sky\"). Then we incorporate an audio sound (e.g., splashing water) to generate videos. Shown in Figure 12, it can produce videos featuring a garden with water where the sun is centered (1st row) or where a black mountain in the background (2nd row)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 576, + 288, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 576, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 576, + 288, + 673 + ], + "type": "text", + "content": "Text-Audio Joint Conditioning. As our model is built upon the Stable Diffusion model, it is also possible to use text and audio as a condition together. In Figure 13, we provide an example where we generate video frames conditioned on a sound of an explosion along with texts, such as \"eruption\", \"spew\", or \"cloud of ash.\" (see 2nd-4th rows) Preserving temporal semantics, our model successfully generates video frames guided by text as well." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Face Generation. Our Audio Semantic Guidance enables detailed adjustment in certain areas, such as face generation. Figure 14 shows examples of human faces conditioned on" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 543, + 223 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 543, + 223 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 543, + 223 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 543, + 223 + ], + "type": "image", + "image_path": "e107fa8c28926bf5c18ed4d9ba4200263d258b82968d636d3c95ce00f328ecb7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 233, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 305, + 233, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 233, + 545, + 270 + ], + "type": "text", + "content": "Figure 14: Examples of face generation with our methods. We use giggling and sobbing sound to manipulate face expressions of human face." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 281, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 317 + ], + "type": "text", + "content": "giggling sound (1st and 3rd rows) and sobbing sound (2nd row). Our model excels in preserving the core facial attributes while manipulating emotional expressions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 327, + 378, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 327, + 378, + 339 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 378, + 339 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 347, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 456 + ], + "type": "text", + "content": "In this paper, we propose The Power of Sound (TPoS), a novel audio-to-video generation with Stable Diffusion. Our work extends the usage of audio modality on generation models, and broaden the methods of using Stable Diffusion by generating realistic videos by our Audio Encoder. Superior performances are achieved over widely-used audio-to-video benchmarks, reflected by objective evaluations and User study, hence attributing towards the new formulation of video generation with audio modality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Research Foundation of Korea grant (NRF-2021R1C1C1009608, " + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "text", + "content": "), Basic Science Research Program (NRF-2021R1A6A1A13044830, " + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "text", + "content": "), Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(2022-0-00043, " + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "text", + "content": "), and the MSIT(Ministry of Science and ICT), Korea, under the ICT Creative Consilience program(IITP-2023-2020-0-01819, " + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "text", + "content": ") supervised by the IITP. W. Ryoo, S. Lee, and S. Kim are supported by the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No. 2019-0-00079, Artificial Intelligence Graduate School Program(Korea University), " + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "text", + "content": "), the National Research Foundation of Korea grant (NRF-2022R1F1A1074334, " + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "text", + "content": "), Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: 4D Content Generation and Copyright Protection with Artificial Intelligence, Project Number: R2022020068, " + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 459, + 546, + 698 + ], + "type": "text", + "content": "), and supported by the Google Cloud Research Credits program(code-MU6X-FKAV-N3A3-1X0T)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7830" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv preprint arXiv:2206.02779, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 156 + ], + "type": "text", + "content": "[2] Manuel Brack, Patrick Schramowski, Felix Friedrich, Dominik Hintersdorf, and Kristian Kersting. The stable artist: Steering semantics in diffusion latent space. arXiv preprint arXiv:2212.06013, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 287, + 201 + ], + "type": "text", + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 203, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 287, + 257 + ], + "type": "text", + "content": "[4] Moitreya Chatterjee and Anoop Cherian. Sound2sight: Generating visual dynamics from sound and context. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 701-719. Springer, 2020. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "text", + "content": "[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 286, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 286, + 390 + ], + "type": "text", + "content": "[7] Sarah Gross, Xingxing Wei, and Jun Zhu. Automatic realistic music video generation from segments of youtube videos. arXiv preprint arXiv:1905.12245, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "type": "text", + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "type": "text", + "content": "[9] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 480, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 534 + ], + "type": "text", + "content": "[10] Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 578 + ], + "type": "text", + "content": "[11] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 580, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 601 + ], + "type": "text", + "content": "[12] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "text", + "content": "[13] Dasaem Jeong, Seungheon Doh, and Taegyun Kwon. Traumerai: Dreaming music with stylegan. arXiv preprint arXiv:2102.04680, 2(4):10, 2021. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[14] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Qianyi Wu, Wayne Wu, Feng Xu, and Xun Cao. Eamm: One-shot emotional talking face via audio-based emotion-aware motion model. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "emotional video portraits. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14080-14089, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 108, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 108, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 308, + 108, + 545, + 151 + ], + "type": "text", + "content": "[16] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 153, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 545, + 206 + ], + "type": "text", + "content": "[17] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "text", + "content": "[18] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Ccvs: context-aware controllable video synthesis. Advances in Neural Information Processing Systems, 34:14042-14055, 2021. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 255, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 255, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 545, + 330 + ], + "type": "text", + "content": "[19] Seung Hyun Lee, Gyeongrok Oh, Wonmin Byeon, Chanyoung Kim, Won Jeong Ryoo, Sang Ho Yoon, Hyunjun Cho, Jihyun Bae, Jinkyu Kim, and Sangpil Kim. Sound-guided semantic video generation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 34-50. Springer, 2022. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 333, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 333, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 333, + 545, + 387 + ], + "type": "text", + "content": "[20] Seung Hyun Lee, Wonseok Roh, Wonmin Byeon, Sang Ho Yoon, Chanyoung Kim, Jinkyu Kim, and Sangpil Kim. Sound-guided semantic image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3377-3386, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 388, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 453 + ], + "type": "text", + "content": "[21] Borong Liang, Yan Pan, Zhizhi Guo, Hang Zhou, Zhibin Hong, Xiaoguang Han, Junyu Han, Jingtuo Liu, Errui Ding, and Jingdong Wang. Expressive talking head generation with granular audio-visual control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3387–3396, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 456, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 521 + ], + "type": "text", + "content": "[22] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 423-439. Springer, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 567 + ], + "type": "text", + "content": "[23] Eyal Molad, Eliahu Horwitz, Dani Valevski, Alex Rav Acha, Yossi Matias, Yael Pritch, Yaniv Leviathan, and Yedid Hoshen. Dreamix: Video diffusion models are general video editors. arXiv preprint arXiv:2302.01329, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 601 + ], + "type": "text", + "content": "[24] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 602, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 655 + ], + "type": "text", + "content": "[25] Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le. Specaugment: A simple data augmentation method for automatic speech recognition. arXiv preprint arXiv:1904.08779, 2019. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[26] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7831" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 685 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[28] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "text", + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 288, + 306 + ], + "type": "text", + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 350 + ], + "type": "text", + "content": "[31] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 352, + 287, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 352, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 48, + 352, + 287, + 406 + ], + "type": "text", + "content": "[32] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 408, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 408, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 287, + 461 + ], + "type": "text", + "content": "[33] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14144-14153, 2021. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 464, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 287, + 518 + ], + "type": "text", + "content": "[34] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "text", + "content": "[35] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 565, + 287, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 287, + 628 + ], + "type": "text", + "content": "[36] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description. arXiv preprint arXiv:2210.02399, 2022. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 632, + 287, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 632, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 48, + 632, + 287, + 685 + ], + "type": "text", + "content": "[37] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. arXiv preprint arXiv:2212.08698, 2022. 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7832" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/c2da89cd-17c8-420c-b57b-7af434285ab7_content_list.json b/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/c2da89cd-17c8-420c-b57b-7af434285ab7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ee8dda75fecaaa98fe8c85d041e4d939342dff8c --- /dev/null +++ b/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/c2da89cd-17c8-420c-b57b-7af434285ab7_content_list.json @@ -0,0 +1,1909 @@ +[ + { + "type": "text", + "text": "The Stable Signature: Rooting Watermarks in Latent Diffusion Models", + "text_level": 1, + "bbox": [ + 125, + 130, + 844, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pierre Fernandez $^{1,2}$ Guillaume Couairon $^{1,3}$ Hervé Jégou $^{1}$ Matthijs Douze $^{1}$ Teddy Furon $^{2*}$", + "bbox": [ + 96, + 179, + 879, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Meta AI 2Centre Inria de l'Université de Rennes 3Sorbonne University", + "bbox": [ + 214, + 215, + 756, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 253, + 313, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative image modeling enables a wide range of applications but raises ethical concerns about responsible deployment. We introduce an active content tracing method combining image watermarking and Latent Diffusion Models. The goal is for all generated images to conceal an invisible watermark allowing for future detection and/or identification. The method quickly fine-tunes the latent decoder of the image generator, conditioned on a binary signature. A pre-trained watermark extractor recovers the hidden signature from any generated image and a statistical test then determines whether it comes from the generative model. We evaluate the invisibility and robustness of the watermarks on a variety of generation tasks, showing that the Stable Signature is robust to image modifications. For instance, it detects the origin of an image generated from a text prompt, then cropped to keep $10\\%$ of the content, with $90 + \\%$ accuracy at a false positive rate below $10^{-6}$ .", + "bbox": [ + 75, + 287, + 473, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 561, + 209, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent progress in generative modeling and natural language processing enable easy creation and manipulation of photo-realistic images, such as with DALL-E 2 [64] or Stable Diffusion [68]. They have given birth to many image edition tools like ControlNet [104], Instruct-Pix2Pix [7], and others [13, 28, 71], that are becoming mainstream creative tools for artists, designers, and the general public.", + "bbox": [ + 75, + 589, + 468, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While this is a great step forward for generative AI, it also undermines confidence in the authenticity or veracity of photo-realistic images. Indeed, methods for photorealistic image edition existed before, but generative AI significantly lowers the barriers to convincing synthetic image generation and edition (e.g. a generated picture recently won an art competition [29]). This raises new risks like deep fakes, impersonation or copyright usurpation [8, 17]. A tool to determine that images are AI-generated would make it easier to ensure their compliance with ethical standards and to remove them from certain platforms.", + "bbox": [ + 75, + 696, + 470, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c7b5c2fde8d9e5922d8418a2133684ce4791ae3ac6f121e907c2d4e645869c8.jpg", + "image_caption": [ + "Figure 1. Overview. The latent decoder can be fine-tuned to preemptively embed a signature into all generated images." + ], + "image_footnote": [], + "bbox": [ + 522, + 276, + 875, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A baseline solution to identify generated images is forensics, i.e. methods to detect generated/manipulated images passively (the image is not modified for identification). An active baseline is to apply existing watermarking methods after the image generation. Watermarking invisibly embeds a secret message into the image, which can then be extracted and used to identify the image. This has several drawbacks. If the model leaks or is open-sourced, the post-generation watermarking can be removed trivially. The open source Stable Diffusion [70] is a case in point, since removing the watermark amounts to commenting out a single line in the source code.", + "bbox": [ + 496, + 566, + 892, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our Stable Signature method merges watermarking into the generation process itself, without any architectural change. It adjusts the pre-trained generative model such that all the images it produces conceal a given watermark. There are several advantages to this approach [46, 99]. It does not require additional processing of the generated image, which makes the watermarking computationally lighter, straightforward, and secure. Model providers could deploy their models to different user groups with a unique watermark, and monitor that they are used in a responsible manner.", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work supported by ANR / AID under Chaire SAIDA ANR-20-CHIA-0011. Correspondance to pfz@meta.com", + "bbox": [ + 76, + 875, + 477, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22466", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "They could give art platforms, news outlets and other sharing platforms the ability to detect when an image has been generated by their AI.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We focus on Latent Diffusion Models (LDM) [68] that can perform a wide range of generative tasks. We show that simply fine-tuning a small part of the generative model – the decoder that generates images from the latent vectors – is enough to natively embed a watermark into generated images. Stable Signature does not require an architectural change and does not modify the diffusion process. Hence it is to our knowledge compatible with all LDM-based generative methods [7, 13, 62, 71, 104]. The fine-tuning stage is performed by back-propagating a combination of a perceptual image loss and a message decoding loss from a watermark extractor back to the LDM decoder. We pre-train the extractor with a simplified version of the deep watermarking method HiDDeN [108].", + "bbox": [ + 75, + 136, + 468, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We create an evaluation benchmark close to real world situations where images may be edited. The tasks are: detection of AI generated images, tracing models from their generations. For instance, we detect $90\\%$ of images generated with the generative model, even if they are cropped to $10\\%$ of their original size, while flagging only one false positive every $10^{6}$ images. To ensure that the model's utility is not weakened, we show that the FID [34] score of the generation is not affected and that the generated images are perceptually indistinguishable from the ones produced by the original model. This is done over several tasks involving LDM (text-to-image, inpainting, edition, etc.).", + "bbox": [ + 75, + 347, + 470, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a summary, (1) we efficiently merge watermarking into the generation process of LDMs, in a way that is compatible with most of the LDM-based generative methods; (2) we demonstrate how it can be used to detect and trace generated images, through a real-world evaluation benchmark; (3) we compare to post-hoc watermarking methods, showing that it is competitive while being more secure and efficient, and (4) evaluate robustness to intentional attacks.", + "bbox": [ + 75, + 529, + 468, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 662, + 218, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image generation has long been dominated by GANs, still state-of-the-art on many datasets [37, 38, 39, 76, 87]. Transformers have also been successfully used for modeling image [66, 19] or video [79] distributions, providing higher diversity at the expense of increased inference time. Images are typically converted to token lists using vector-quantized architectures [20, 67, 96], relying on an image decoder.", + "bbox": [ + 75, + 688, + 468, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [18, 36, 58, 80] are a huge improvement in text-conditional image generation, current models synthesize high-resolution photo-realistic images for a wide variety of text prompts [3, 35, 65, 69, 74]. They can also perform conditional image generation – like inpainting or text-guided image editing – by fine-tuning the diffusion model with additional conditioning, e.g. masked input im", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "age, segmentation map, etc. [50, 73]. Because of their iterative denoising algorithm, diffusion models can also be adapted for image editing in a zero-shot fashion by guiding the generative process [13, 33, 40, 56, 85, 92]. All these methods, when combined with Stable Diffusion, operate in the latent rather than image space, requiring a latent decoder to produce an RGB image.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detection of AI-generated/manipulated images is notably active in the context of deep-fake detection [32, 107]. Many works focus on the detection of GAN-generated images [10, 31, 89, 106]. One approach is to detect inconsistencies, via lights, perspective or physical objects [21, 22, 45, 52, 88]. These approaches are restricted to photorealistic images or faces, artworks not intended to be physically correct are not covered.", + "bbox": [ + 496, + 208, + 890, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other approaches track traces left by the generators in the spatial [54, 97] or frequency [26, 106] domains. There are extensions to diffusion models in recent works [12, 77] that show encouraging results. However purely relying on forensics and passive detection is limiting, e.g. the best performing method to our knowledge [12] is able to detect $50\\%$ of generated images for a false positive rate around 1/100: if a user-generated content platform were to receive 1 billion images every day, it would wrongly flag 10 million images to detect only half of the generated images. Besides, passive techniques cannot trace images from different versions of the same model, in contrast with watermarking.", + "bbox": [ + 496, + 330, + 892, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image watermarking has long been studied in the context of tracing and intellectual property protection [14]. More recently, deep learning encoder/extractors like HiD-DeN [2, 44, 51, 100, 108] or iterative methods by Vukotic et al. [25, 42, 86] showed competitive results in terms of robustness to a wide range of transformations, namely geometric ones.", + "bbox": [ + 496, + 523, + 890, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the specific case of generative models, some works aim to watermark the training set on which the generative model is learned [98]. It is highly inefficient since every new message to embed requires a new training pipeline. Merging the watermarking and the generative process is a recent idea [23, 46, 59, 63, 93, 99, 102], that is closer to the model watermarking literature [84]. They suffer from two strong limitations. First, these methods only apply to GANs, while LDM are progressively replacing them for most applications. Second, watermarking is incorporated in the training process of the GAN from the start. This strategy is unsustainable because the generative model training is more and more costly1. Our work shows that a quick finetuning of the latent decoder part of the generative model is enough to achieve a good watermarking performance, provided that the watermark extractor is well chosen.", + "bbox": [ + 496, + 631, + 890, + 872 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 Stable Diffusion training costs \\(\\sim\\)\\\\(600k of cloud compute (Wikipedia).", + "bbox": [ + 517, + 886, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22467", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Problem Statement & Background", + "text_level": 1, + "bbox": [ + 76, + 89, + 392, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 1 shows a model provider Alice who deploys a latent diffusion model to users Bobs. Stable Signature embeds a binary signature into the generated images. This section derives how Alice can use this signature for two scenarios:", + "bbox": [ + 76, + 114, + 468, + 176 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Detection: \"Is it generated by my model?\" Alice detects if an image was generated by her model. Generated images should be flagged as reliably as possible, while controlling the probability of flagging a natural image.", + "- Identification: \"Who generated this image?\". Alice monitors who created each image, while avoiding to mistakenly identifying a Bob." + ], + "bbox": [ + 76, + 180, + 468, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Image watermarking for detection", + "text_level": 1, + "bbox": [ + 76, + 299, + 375, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alice embeds a $k$ -bit binary signature into the generated images. The watermark extractor then decodes messages from the images it receives and detects when the message is close to Alice's signature. An example application is to block AI-generated images on a content sharing platform.", + "bbox": [ + 75, + 323, + 468, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Statistical test Let $m \\in \\{0,1\\}^k$ be Alice's signature. We extract the message $m'$ from an image $x$ and compare it to $m$ . As done in previous works [46, 98], the detection test relies on the number of matching bits $M(m,m')$ : if", + "bbox": [ + 75, + 409, + 468, + 470 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM \\left(m, m ^ {\\prime}\\right) \\geq \\tau \\text {w h e r e} \\tau \\in \\{0, \\dots , k \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 473, + 468, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "then the image is flagged. This provides a level of robustness to imperfections of the watermarking.", + "bbox": [ + 75, + 492, + 467, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, we test the statistical hypothesis $H_{1}$ : \" $x$ was generated by Alice's model\" against the null hypothesis $H_{0}$ : \" $x$ was not generated by Alice's model\". Under $H_{0}$ (i.e. for vanilla images), we assume that bits $m_{1}', \\ldots, m_{k}'$ are (i.i.d.) Bernoulli random variables with parameter 0.5. Then $M(m, m')$ follows a binomial distribution with parameters $(k, 0.5)$ . We verify this assumption experimentally in App. B.5. The False Positive Rate (FPR) is the probability that $M(m, m')$ takes a value bigger than the threshold $\\tau$ . It is obtained from the CDF of the binomial distribution, and a closed-form can be written with the regularized incomplete beta function $I_{x}(a; b)$ :", + "bbox": [ + 75, + 523, + 468, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {F P R} (\\tau) = \\mathbb {P} (M > \\tau | H _ {0}) = I _ {1 / 2} (\\tau + 1, k - \\tau). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 710, + 468, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Image watermarking for identification", + "text_level": 1, + "bbox": [ + 500, + 90, + 830, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alice now embeds a signature $m^{(i)}$ drawn randomly from $\\{0,1\\}^k$ into the model distributed to $\\mathrm{Bob}^{(i)}$ (for $i = 1 \\cdots N$ , with $N$ the number of Bobs). Alice can trace any misuse of her model: generated images violating her policy (gore content, deepfakes) are linked back to the specific Bob by comparing the extracted message to Bobs' signatures.", + "bbox": [ + 496, + 113, + 890, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Statistical test We compare the message $m'$ from the watermark extractor to $(m^{(1)}, \\ldots, m^{(N)})$ . There are now $N$ detection hypotheses to test. If the $N$ hypotheses are rejected, we conclude that the image was not generated by any of the models. Otherwise, we attribute the image to $\\operatorname{argmax}_{i=1..N} M(m', m^{(i)})$ . With regards to the detection task, false positives are more likely since there are $N$ tests. The global FPR at a given threshold $\\tau$ is:", + "bbox": [ + 496, + 217, + 890, + 337 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {F P R} (\\tau , N) = 1 - (1 - \\operatorname {F P R} (\\tau)) ^ {N} \\approx N. \\operatorname {F P R} (\\tau). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 342, + 890, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Equation (3) (resp. (2)), is used reversely: we find threshold $\\tau$ to achieve a required FPR for identification (resp. detection). Note that these formulae hold only under the assumption of i.i.d. Bernoulli bits extracted from vanilla images. This condition is enforced in the next section.", + "bbox": [ + 496, + 364, + 890, + 440 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 500, + 455, + 589, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Stable Signature modifies the generative network so that the generated images have a given signature through a fixed watermark extractor. It is trained in two phases. First, we create the watermark extractor network $\\mathcal{W}$ . We then fine-tune the LDM decoder $\\mathcal{D}$ , such that all generated images yield a given signature through $\\mathcal{W}$ .", + "bbox": [ + 496, + 481, + 890, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Pre-training the watermark extractor", + "text_level": 1, + "bbox": [ + 498, + 582, + 823, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use HiDDeN [108], a classical method in the deep watermarking literature. It jointly optimizes the parameters of watermark encoder $\\mathcal{W}_E$ and extractor network $\\mathcal{W}$ to embed $k$ -bit messages into images, robustly to transformations that are applied during training. We discard $\\mathcal{W}_E$ after training, since only $\\mathcal{W}$ serves our purpose.", + "bbox": [ + 496, + 604, + 890, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, $\\mathcal{W}_E$ takes as inputs a cover image $x_o \\in \\mathbb{R}^{W \\times H \\times 3}$ and a $k$ -bit message $m \\in \\{0,1\\}^k$ . Similar to", + "bbox": [ + 500, + 696, + 890, + 727 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a8e372dd0b131cee24e7f98cf1a102573b2555b1a632bc6dc9f3c77ca79c9a7b.jpg", + "image_caption": [ + "(a) Pre-train watermark encoder/extractor" + ], + "image_footnote": [], + "bbox": [ + 89, + 762, + 398, + 859 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c9a2e0e6c5c67da3f73d1f1ed5f06268f224de6ad0bd88b0df88ac520db22afd.jpg", + "image_caption": [ + "(b) Fine-tune LDM decoder", + "(c) Generate", + "Figure 2. Steps of the method. (a) We pre-train a watermark encoder $\\mathcal{W}_E$ and extractor $\\mathcal{W}$ , to extract binary messages. (b) We fine-tune the decoder $\\mathcal{D}$ of the LDM's auto-encoder with a fixed signature $m$ such that all the generated images (c) lead to $m$ through $\\mathcal{W}$ ." + ], + "image_footnote": [], + "bbox": [ + 418, + 762, + 702, + 858 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/541d235c8465ce2d5220684d78dcd49854ea3ccb8613edfda496a461ad40b6e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 848, + 766, + 875, + 787 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/449466aa4f7c59bf5152aa317699ad2e0c6a8690bd875af411b24312fba4393e.jpg", + "image_caption": [ + "Text - Super-resolution Inpainting - Editing - ..." + ], + "image_footnote": [], + "bbox": [ + 720, + 791, + 854, + 849 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22468", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ReDMark [2], $\\mathcal{W}_E$ outputs a residual image $\\delta$ of the same size as $x_{o}$ , that is multiplied by a factor $\\alpha$ to produce watermarked image $x_w = x_o + \\alpha \\delta$ . At each optimization step an image transformation $T$ is sampled from a set $\\mathcal{T}$ that includes common image processing operations such as cropping and JPEG compression. A \"soft\" message is extracted from the transformed image: $m' = \\mathcal{W}(T(x_w))$ (at inference time, the decoded message is given by the signs of the components of $m'$ ). The message loss is the Binary Cross Entropy (BCE) between $m$ and the sigmoid $\\sigma(m')$ :", + "bbox": [ + 75, + 90, + 472, + 243 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m} = - \\sum_ {i = 1} ^ {k} m _ {i} \\cdot \\log \\sigma (m _ {i} ^ {\\prime}) + (1 - m _ {i}) \\cdot \\log (1 - \\sigma (m _ {i} ^ {\\prime})).\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 243, + 468, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The network architectures are kept simple to ease the LDM fine-tuning in the second phase. They are the same as HiDDeN [108] (see App. A.1) with two changes.", + "bbox": [ + 75, + 292, + 468, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, since $\\mathcal{W}_E$ is discarded, its perceptual quality is not as important, so the perceptual loss or the adversarial network are not needed. Instead, the distortion is constrained by a tanh function on output of $\\mathcal{W}_E$ and by the scaling factor $\\alpha$ . This improves the bit accuracy of the recovered message and makes it possible to increase its size $k$ .", + "bbox": [ + 75, + 339, + 468, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Second, we observed that $\\mathcal{W}$ 's output bits for vanilla images are correlated and highly biased, which violates the assumptions of Sec. 3.1. Therefore, to get closer to i.i.d. dimensions we remove the bias and decorrelate the outputs of $\\mathcal{W}$ by applying a PCA whitening transformation (more details in App. A.1).", + "bbox": [ + 75, + 429, + 468, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Fine-tuning the generative model", + "text_level": 1, + "bbox": [ + 76, + 529, + 369, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In LDM, the diffusion happens in the latent space of an auto-encoder. The latent vector $z$ obtained at the end of the diffusion is input to decoder $\\mathcal{D}$ to produce an image. Here we fine-tune $\\mathcal{D}$ such that the image contains a given message $m$ that can be extracted by $\\mathcal{W}$ . Stable Signature is compatible with many generative tasks, since modifying only $\\mathcal{D}$ does not affect the diffusion process.", + "bbox": [ + 75, + 551, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we fix the signature $m = (m_1,\\ldots ,m_k)\\in \\{0,1\\} ^k$ The fine-tuning of $\\mathcal{D}$ into $\\mathcal{D}_m$ is inspired by the original training of the auto-encoder in LDM [68].", + "bbox": [ + 75, + 657, + 468, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training image $x \\in \\mathbb{R}^{H \\times W \\times 3}$ is fed to the LDM encoder $\\mathcal{E}$ that outputs activation map $z = \\mathcal{E}(x) \\in \\mathbb{R}^{h \\times w \\times c}$ , downsampled by a power-of-two factor $f = H / h = W / w$ . The decoder reconstructs an image $x' = \\mathcal{D}_m(z)$ and the extractor recovers $m' = \\mathcal{W}(x')$ . The message loss is the BCE between $m'$ and the original $m$ : $\\mathcal{L}_m = \\mathrm{BCE}(\\sigma(m'), m)$ .", + "bbox": [ + 75, + 702, + 468, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, the original decoder $\\mathcal{D}$ reconstructs the image without watermark: $x_{o}^{\\prime} = \\mathcal{D}(z)$ . The image perceptual loss $\\mathcal{L}_{\\mathrm{i}}$ between $x^{\\prime}$ and $x_{o}^{\\prime}$ , controls the distortion. We use the Watson-VGG perceptual loss introduced by Czolbe et", + "bbox": [ + 75, + 794, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "al. [15], an improved version of LPIPS [105]. It is essential that the decoder learns luminance and contrast masking to add less visible watermarks. The weights of $\\mathcal{D}_m$ are optimized to minimize", + "bbox": [ + 500, + 90, + 890, + 148 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {m}} + \\lambda_ {\\mathrm {i}} \\mathcal {L} _ {\\mathrm {i}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 148, + 890, + 164 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This is done over 100 iterations with the AdamW optimizer [49] and batch of size 4, i.e. the fine-tuning sees fewer than 500 images and takes one minute on a single GPU. The learning rate follows a cosine annealing schedule with 20 iterations of linear warmup to $10^{-4}$ and decays to $10^{-6}$ . The factor $\\lambda_{\\mathrm{i}}$ in (4) is set to 0.2 by default.", + "bbox": [ + 496, + 167, + 893, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Text-to-Image Watermarking Performance", + "text_level": 1, + "bbox": [ + 498, + 271, + 883, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section discusses our method's detection and identification capability for images generated by a Stable-Diffusion-like model $[68]^3$ . We apply generative models watermarked with 48-bit signatures on prompts of the MS-COCO [47] validation set. We evaluate detection and identification on the outputs, as illustrated in Figure 1.", + "bbox": [ + 496, + 296, + 890, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate their robustness to different transformations applied to generated images: strong cropping (10% of the image remaining), brightness shift (strength factor 2.0), as well as a combination of crop 50%, brightness shift 1.5 and JPEG 80. This covers typical geometric and photometric edits (see Fig. 5 for visual examples).", + "bbox": [ + 496, + 387, + 890, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The detection rates are partly obtained from experiments and partly by extrapolating small-scale measurements.", + "bbox": [ + 498, + 478, + 890, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Detection results", + "text_level": 1, + "bbox": [ + 500, + 517, + 665, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For detection, we fine-tune the decoder of the LDM with a random key $m$ , generate 1000 images and use the test of Eq. (1). We report the tradeoff between True Positive Rate (TPR), i.e. the probability of flagging a generated image and the FPR, while varying $\\tau \\in \\{0,..,48\\}$ . For instance, for $\\tau = 0$ , we flag all images so $\\mathrm{FPR} = 1$ , and $\\mathrm{TPR} = 1$ . The TPR is measured directly, while the FPR is inferred from Eq. (2), because it would otherwise be too small to be measured on reasonably sized problems (this approximation is validated experimentally in App. B.6). The experiment is run on 10 random signatures and we report averaged results.", + "bbox": [ + 496, + 540, + 890, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 shows the tradeoff under image transformations. For example, when the generated images are not modified, Stable Signature detects $99\\%$ of them, while only 1 vanilla image out of $10^{9}$ is flagged. At the same FPR $= 10^{-9}$ , Stable Signature detects $84\\%$ of generated images for a crop that keeps $10\\%$ of the image, and $65\\%$ for a transformation that combines a crop, a color shift, and a JPEG compression. For comparison, we report results of a state-of-the-art passive method [12], applied on resized and", + "bbox": [ + 496, + 707, + 890, + 843 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2The transformation needs to be differentiable in pixel space. This is not the case for JPEG compression so we use the forward attack simulation layer introduced by Zhang et al. [101].", + "bbox": [ + 75, + 862, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3We refrain from experimenting with pre-existing third-party generative models, such as Stable Diffusion or LDMs, and instead use a large diffusion model (2.2B parameters) trained on an internal dataset of 330M licensed image-text pairs.", + "bbox": [ + 500, + 849, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22469", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "compressed images. As expected, we observe that these baseline results have orders of magnitudes larger FPR than Stable Signature, which actively marks the content.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Identification results", + "text_level": 1, + "bbox": [ + 76, + 151, + 272, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each Bob has its own copy of the generative model. Given an image, the goal is to find if any of the $N$ Bobs created it (detection) and if so, which one (identification). There are 3 types of error: false positive: flag a vanilla image; false negative: miss a generated image; false accusation: flag a generated image but identify the wrong Bob.", + "bbox": [ + 75, + 176, + 468, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For evaluation, we fine-tune $N' = 1000$ models with random signatures. Each model generates 100 images. For each of these 100k watermarked images, we extract the Stable Signature message, compute the matching score with all $N$ signatures and select the Bob with the highest score. The image is predicted to be generated by that Bob if this score is above threshold $\\tau$ . We determined $\\tau$ such that $\\mathrm{FPR} = 10^{-6}$ , see Eq. (3). For example, for $N = 1$ , $\\tau = 41$ and for $N = 1000$ , $\\tau = 44$ . Accuracy is extrapolated beyond the $N'$ Bobs by adding additional signatures and having $N > N'$ (e.g. Bobs that have not generated any images).", + "bbox": [ + 75, + 270, + 468, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 4 reports the per-transformation identification accuracy. For example, we identify a Bob among $N = 10^{5}$ with $98\\%$ accuracy when the image is not modified. Note that for the combined edit, this becomes $40\\%$ . This may still be dissuasive: if a Bob generates 3 images, he will be identified $80\\%$ of the time. We observe that at this scale, the false accusation rate is zero, i.e. we never identify the wrong Bob. This is because $\\tau$ is set high to avoid FPs, which also makes false accusations unlikely. We observe that the identification accuracy decreases when $N$ increases, because the threshold $\\tau$ required to avoid false positives is higher when $N$ increases, as pointed out by the approximation in (3). In a nutshell, by distributing more models, Alice trades some accuracy of detection against the ability to identify Bobs.", + "bbox": [ + 75, + 438, + 468, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 89, + 705, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The previous section showed how to leverage watermarks for detection and identification of images generated from text prompts. We now present more general results on robustness and image quality for different generative tasks. We also compare Stable Signature to post-generation watermarking algorithms.", + "bbox": [ + 498, + 114, + 890, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Tasks & evaluation metrics", + "text_level": 1, + "bbox": [ + 500, + 215, + 743, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our method only involves the LDM decoder, it is compatible with many generative tasks. We evaluate text-to-image generation and image edition on the validation set of MS-COCO [47], super-resolution and inpainting on the validation set of ImageNet [16] (all evaluation details are available in App. A.3).", + "bbox": [ + 498, + 238, + 890, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the image distortion with the Peak Signal-to-Noise Ratio (PSNR), which is defined as $\\mathrm{PSNR}(x, x') = -10 \\cdot \\log_{10}(\\mathrm{MSE}(x, x'))$ , for $x, x' \\in [0, 1]^{c \\times h \\times w}$ , as well as Structural Similarity score (SSIM) [90]. They compare images generated with and without watermark. On the other hand, we evaluate the diversity and quality of the generated images with the Fréchet Inception Distance (FID) [34]. The bit accuracy – the percentage of bits correctly decoded – evaluates the watermarks' robustness.", + "bbox": [ + 498, + 329, + 890, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.2. Image generation quality", + "text_level": 1, + "bbox": [ + 500, + 474, + 728, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 6 shows qualitative examples of how the image generation is altered by the latent decoder's fine-tuning. The difference is very hard to perceive even for a trained eye. This is surprising for such a low PSNR, especially since the watermark embedding is not constrained by any Human Visual System like in professional watermarking techniques. Most interestingly, the LDM decoder has indeed learned to add the watermark signal only over textured areas where the human eyes are not sensitive, while the uniform backgrounds are kept intact (see the pixel-wise difference).", + "bbox": [ + 498, + 498, + 890, + 650 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/830dc5ae205f38ce29636fc7375bf256df868f68501542c82e7659d8ce7b3dde.jpg", + "image_caption": [ + "Figure 3. Detection results. TPR/FPR curve of the detection under different transformations. Forensics† indicates passive detection (without watermark) [12]." + ], + "image_footnote": [], + "bbox": [ + 86, + 676, + 339, + 832 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1d8f6ca966033bc65ea96bef8580b6508dc40c9d42b1f302623a6907be82a3bf.jpg", + "image_caption": [ + "Figure 4. Identification results. Proportion of well-identified Bobs. Detection with $\\mathrm{FPR} = 10^{-6}$ is run beforehand, and we consider it an error if the image is not flagged." + ], + "image_footnote": [], + "bbox": [ + 375, + 675, + 629, + 832 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/69f4398b92a022160850a5cc639b8e5692c3e672d92cc1201911c36ec04f416f.jpg", + "image_caption": [ + "Crop $10\\%$" + ], + "image_footnote": [], + "bbox": [ + 669, + 676, + 772, + 766 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ef05b53d6b96c8e4e8ba8cf22fd9db4c9d054f624e914ca94c32d59c690d2017.jpg", + "image_caption": [ + "Combined" + ], + "image_footnote": [], + "bbox": [ + 779, + 676, + 880, + 766 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/96ddc50499967b5745d9f7384444a6efa2616b3c3d8db35831e9646b0807e0dc.jpg", + "image_caption": [ + "Figure 5. Transformations evaluated in Sec. 5 & 6. 'Combined' is made of crop $50\\%$ , brightness adjustment 1.5 and JPEG 80 compression." + ], + "image_footnote": [], + "bbox": [ + 794, + 780, + 866, + 835 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22470", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3b9784f55e700680f5b21d0aebf92e44d5f91161f654d1f6cf3546b5625e2066.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Generation quality and comparison to post-hoc watermarking on $512\\times 512$ images and 48-bit signatures. PSNR and SSIM are computed between generations of the original and watermarked generators. For FID, we show in (color) the difference with regards to original. Post-hoc watermarking is evaluated on text-generated images. (App. B.2 gives results on more transformations, and App. A gives more details on the evaluations.) Overall, Stable Signature has minimal impact on generation quality. It has comparable robustness to post-hoc methods while being rooted in the generation itself." + ], + "table_body": "
PSNR / SSIM ↑FID ↓Bit accuracy ↑ on:
NoneCropBrigh.Comb.
TasksText-to-ImageLDM [68]30.0 / 0.8919.6 (−0.3)0.990.950.970.92
Image EditionDiffEdit [13]31.2 / 0.9215.0 (−0.3)0.990.950.980.94
Inpainting - Full - Mask onlyGlide [57]31.1 / 0.9116.8 (+0.6)0.990.970.980.93
37.8 / 0.989.0 (+0.1)0.890.760.840.78
Super-ResolutionLDM [68]34.0 / 0.9411.6 (+0.0)0.980.930.960.92
WM MethodsPost generation
Dct-Dwt [14]0.14 (s/img)39.5 / 0.9719.5 (−0.4)0.860.520.510.51
SSL Watermark [25]0.45 (s/img)31.1 / 0.8620.6 (+0.7)1.000.730.930.66
FNNS [42]0.28 (s/img)32.1 / 0.9019.0 (−0.9)0.930.930.910.93
HiDDeN [108]0.11 (s/img)32.0 / 0.8819.7 (−0.2)0.990.970.990.98
Merged in generation
Stable Signature0.00 (s/img)30.0 / 0.8919.6 (−0.3)0.990.950.970.92
", + "bbox": [ + 81, + 88, + 637, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 presents a quantitative evaluation of image generation on the different tasks. We report the FID, and the average PSNR and SSIM that are computed between the images generated by the fine-tuned LDM and the original one. The results show that no matter the task, the watermarking has very small impact on the FID of the generation.", + "bbox": [ + 75, + 334, + 468, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The average PSNR is around 30 dB and SSIM around 0.9 between images generated by the original and a watermarked model. This is a bit low from a watermarking perspective, because we do not explicitly optimize for them. However, in a real world scenario, one would only have the watermarked version of the image. Therefore it is not as important to be extremely close to the original image, we just need to generate artifacts-free images. Without access to the image generated by the original LDM, it is very hard to tell whether a watermark is present or not.", + "bbox": [ + 75, + 430, + 470, + 583 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0891831bec1705ba992a3d421a68eaef704c641ffe4cdff967a39cc66271c124.jpg", + "image_caption": [ + "Figure 6. Images generated with Stable Diffusion. The PSNR is $35.4\\mathrm{dB}$ in the first row and $28.6\\mathrm{dB}$ in the second. Images generated with Stable Signature look natural because modified areas are located where the eye is not sensitive. More examples in App. C." + ], + "image_footnote": [], + "bbox": [ + 81, + 618, + 462, + 834 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7d6d7890e0308fd50ade6f5ba1a692af168cdcf1087ba81d35d5d6c5a8b75efa.jpg", + "table_caption": [ + "Table 2. Watermark robustness on image transformations applied before decoding, details of which are available in App. A.3. We report the bit accuracy, averaged over $10 \\times 1k$ images generated from COCO prompts with 10 different keys." + ], + "table_footnote": [], + "table_body": "
AttackBit acc.Comb.0.92Sharpness 2.00.99
None0.99Bright. 2.00.97Med. Filter k=70.94
Crop 0.10.95Cont. 2.00.98Resize 0.70.91
JPEG 500.88Sat. 2.00.99Text overlay0.99
", + "bbox": [ + 509, + 392, + 879, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3. Watermark robustness", + "text_level": 1, + "bbox": [ + 500, + 465, + 712, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the robustness of the watermark to different image transformations applied before extraction. For each task, we generate 1k images with 10 models fine-tuned for different messages, and report the average bit accuracy in Table 1. Additionally, Table 2 reports results on more image transformations for images generated from COCO prompts. The main transformations are presented in Fig. 5 (more evaluation details are available in App. A.3).", + "bbox": [ + 496, + 488, + 890, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We see that the watermark is indeed robust for several tasks and across transformations. The bit accuracy is always above 0.9, except for inpainting, when replacing only the masked region of the image (between $1 - 50\\%$ of the image, with an average of $27\\%$ across masks). Besides, the bit accuracy is not perfect even without edition, mainly because there are images that are hard to watermark (e.g. the ones that are very uniform, like the background in Fig. 6) and for which the accuracy is lower.", + "bbox": [ + 496, + 609, + 890, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that the robustness comes even without any transformation during the LDM fine-tuning phase: it is due to the watermark extractor. If the watermark embedding pipeline is learned to be robust against an augmentation, then the LDM learns how to produce watermarks that are robust against it during fine-tuning.", + "bbox": [ + 496, + 746, + 890, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.4. Comparison to post-hoc watermarking", + "text_level": 1, + "bbox": [ + 500, + 845, + 834, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An alternative way to watermark generated images is to process them after the generation (post-hoc). This may", + "bbox": [ + 498, + 869, + 890, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22471", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7812f07a1b7190e77daa4f222089cf561ef04d226d483963dd974fab8df327aa.jpg", + "table_caption": [ + "Table 3. Quality-robustness trade-off during fine-tuning." + ], + "table_footnote": [], + "table_body": "
λi for fine-tuning0.80.40.20.10.050.025
PSNR ↑31.430.629.728.526.824.6
Bit acc. ↑ on ‘comb.’0.850.880.900.920.940.95
", + "bbox": [ + 99, + 101, + 446, + 150 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/49beb7fda33255bc5b76b0d1e6649b632c7995855db434535282fb1b9471cd7d.jpg", + "table_caption": [ + "Table 4. Role of the attack simulation layer at pre-training." + ], + "table_footnote": [], + "table_body": "
Seen at W trainingBit accuracy ↑ at test time:
Crop 0.1Rot. 90JPEG 50Bright. 2.0Res. 0.7
X1.000.560.500.990.48
✓1.000.990.900.990.91
", + "bbox": [ + 91, + 176, + 455, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "be simpler, but less secure and efficient than Stable Signature. We compare our method to a frequency based method, DCT-DWT [14], iterative approaches (SSL Watermark [25] and FNNS [42]), and an encoder/decoder one like HiD-DeN [108]. We choose DCT-DWT since it is employed by the original open source release of Stable Diffusion [70], and the other methods because of their performance and their ability to handle arbitrary image sizes and number of bits. We use our implementations (see details in App. A.4).", + "bbox": [ + 75, + 257, + 467, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 compares the generation quality and the robustness over $5\\mathrm{k}$ generated images. Overall, Stable Signature achieves comparable results in terms of robustness. HiD-DeN's performance is a bit higher but its output bits are not i.i.d. meaning that it cannot be used with the same guarantees as the other methods. We also observe that post-hoc generation gives worse qualitative results, images tend to present artifacts (see Fig. 13 in the supplement). One explanation is that Stable Signature is merged into the high-quality generation process with the LDM auto-encoder model, which is able to modify images in a more subtle way.", + "bbox": [ + 75, + 393, + 467, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.5. Can we trade image quality for robustness?", + "text_level": 1, + "bbox": [ + 76, + 570, + 444, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can choose to maximize the image quality or the robustness of the watermark thanks to the weight $\\lambda_{\\mathrm{i}}$ of the perceptual loss in (4). We report the average PSNR of 1k generated images, as well as the bit accuracy obtained on the extracted message for the 'Combined' editing applied before detection (qualitative results are in App. B.1). A higher $\\lambda_{\\mathrm{i}}$ leads to an image closer to the original one, but to lower bit accuracies on the extracted message, see Table 3.", + "bbox": [ + 75, + 594, + 467, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.6. Attack simulation layer", + "text_level": 1, + "bbox": [ + 76, + 724, + 294, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Watermark robustness against image transformations depends solely on the watermark extractor. Here, we pretrain them with or without specific transformations in the simulation layer, on a shorter schedule of 50 epochs, with $128 \\times 128$ images and 16-bits messages. From there, we plug them in the LDM fine-tuning stage and we generate 1k images from text prompts. We report the bit accuracy of the extracted watermarks in Table 4. The extractor is naturally robust to some transformations, such as crops or brightness, without being trained with them, while others, like rotations", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "or JPEG, require simulation during training for the watermark to be recovered at test time. Empirically we observed that adding a transformation improves results for the latter, but makes training more challenging.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Attacks on Stable Signature's Watermarks", + "text_level": 1, + "bbox": [ + 498, + 167, + 880, + 184 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We examine the watermark's resistance to intentional tampering, as opposed to distortions that happen without bad intentions like crops or compression (discussed in Sec. 5). We consider two threat models: one is typical for many image watermarking methods [14] and operates at the image level, and another targets the generative model level. For image-level attacks, we evaluate on $5\\mathrm{k}$ images generated from COCO prompts. Full details on the following experiments can be found in Appendix A.5.", + "bbox": [ + 496, + 193, + 890, + 329 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.1. Image-level attacks", + "text_level": 1, + "bbox": [ + 498, + 340, + 684, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Watermark removal. Bob alters the image to remove the watermark with deep learning techniques, like methods used for adversarial purification [78, 94] or neural auto-encoders [1, 48]. Note that this kind of attacks has not been explored in the image watermarking literature to our knowledge. Figure 7 evaluates the robustness of the watermark against neural auto-encoders [4, 11, 20, 68] at different compression rates. To reduce the bit accuracy closer to random (50%), the image distortion needs to be strong (PSNR<26). However, assuming the attack is informed on the generative model, i.e. the auto-encoder is the same as the one used to generate the images, the attack becomes much more effective. It erases the watermark while achieving high quality (PSNR>29). This is because the image is modified precisely in the bandwidth where the watermark", + "bbox": [ + 496, + 364, + 890, + 592 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85e7692f5251de14abbe00d51d8887092e0d0405f0b822b74e9bab6f8d30f8df.jpg", + "image_caption": [ + "Figure 7. Removal attacks. $x_{o}$ is the image produced by the original generator, $x_{r}$ is the version produced by the watermarked generator and then attacked. Bit accuracy is on the watermark extracted from $x_{r}$ . Neural auto-encoders [4, 11, 20] follow the same trend, except for the one used by LDM ('KL-f8' for our LDM). When access to the watermark extractor is granted, adversarial attacks also remove the watermark at lower PSNR budget." + ], + "image_footnote": [], + "bbox": [ + 501, + 621, + 874, + 792 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22472", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is embedded. Note that this assumption is strong, because Alice does not need to distribute the original generator.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Watermark removal & embedding (white-box). To go further, we assume that the attack is informed on the watermark extractor - e.g. because it has leaked. Bob can use an adversarial attack to remove the watermark by optimizing the image under a PSNR constraint. The objective is to minimize the $\\ell_2$ distance between a random binary message sampled beforehand and the extractor's output, effectively replacing the original signature with a random one. It makes it possible to erase the watermark with a lower distortion budget, as seen in Fig. 7.", + "bbox": [ + 75, + 133, + 468, + 285 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instead of removing the watermark, an attacker could embed a signature into vanilla images (unauthorized embedding [14]) to impersonate another Bob of whom they have a generated image. It highlights the importance of keeping the watermark extractor private.", + "bbox": [ + 75, + 286, + 468, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.2. Network-level attacks", + "text_level": 1, + "bbox": [ + 76, + 372, + 281, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model purification. Bob gets Alice's generative model and uses a fine-tuning process akin to Sec. 4.2 to eliminate the watermark embedding – that we coin model purification. This involves removing the message loss $\\mathcal{L}_{\\mathrm{m}}$ , and shifting the focus to the perceptual loss $\\mathcal{L}_{\\mathrm{i}}$ between the original image and the one reconstructed by the LDM auto-encoder.", + "bbox": [ + 75, + 395, + 468, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 8 shows the results of this attack for the MSE loss. The PSNR between the watermarked and purified images is plotted at various stages of fine-tuning. Empirically, it is difficult to significantly reduce the bit accuracy without compromising the image quality: artifacts start to appear during the purification.", + "bbox": [ + 75, + 487, + 468, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model collusion. Users may collude by aggregating their models. For instance, $\\mathrm{Bob}^{(i)}$ and $\\mathrm{Bob}^{(j)}$ can average the weights of their models (like Model soups [91]) creating a new model to deceive identification. We found that the bit at position $\\ell$ output by the extractor will be 0 (resp. 1) when the $\\ell$ -th bits of $\\mathrm{Bob}^{(i)}$ and $\\mathrm{Bob}^{(j)}$ are both 0 (resp. 1), and that the extracted bit is random when their bits disagree. We", + "bbox": [ + 75, + 590, + 468, + 696 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fab0b6e04f8cc204629f9b8d0d333db1a6699340d819ef53d36ba1c99b0578a3.jpg", + "image_caption": [ + "Figure 8. Robustness to model purification, i.e. fine-tuning the model to remove watermarks. $x_w$ is the watermarked image, $x_r$ is generated with the purified model at different steps of the process." + ], + "image_footnote": [], + "bbox": [ + 106, + 724, + 441, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "show the distributions of the soft bits (before thresholding) output by the watermark extractor on images generated by the average model. The $\\ell$ -th output is labeled by bits of $\\mathrm{Bob}^{(i)}$ and $\\mathrm{Bob}^{(j)}$ ( $00$ means both have $0$ at position $\\ell$ ):", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/34f4af48989a935f75faa92f0873327c1aff8b0f6e920593dfb04870fb017f62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 161, + 883, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This so-called marking assumption plays a crucial role in traitor tracing literature [27, 55, 83]. Surprisingly, it holds even though our watermarking process is not explicitly designed for it. The study has room for improvement, such as creating user identifiers with more powerful traitor tracing codes [83] and using more powerful traitor accusation algorithms [27, 55]. Importantly, we found the precedent remarks also hold if the colluders operate at the image level.", + "bbox": [ + 498, + 238, + 890, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion & Discussion", + "text_level": 1, + "bbox": [ + 500, + 373, + 732, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By a quick fine-tuning of the decoder of Latent Diffusion Models, we can embed watermarks in all the images they generate. This does not alter the diffusion process, making it compatible with most of LDM-based generative models. These watermarks are robust, invisible to the human eye and can be employed to detect generated images and identify the user that generated it, with very high performance.", + "bbox": [ + 498, + 398, + 890, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The public release of image generative models has an important societal impact. With this work, we put to light the usefulness of using watermarking instead of relying on passive detection methods. We hope it will encourage researchers and practitioners to employ similar approaches before making their models publicly available.", + "bbox": [ + 498, + 505, + 890, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reproducibility Statement. Although the diffusion-based generative model has been trained on an internal dataset of licensed images, we use the KL auto-encoder from LDM [68] with compression factor $f = 8$ . This is the one used by open-source alternatives. Code is available at github.com/facebookresearch/stable_signature.", + "bbox": [ + 498, + 604, + 890, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Environmental Impact. We do not expect any environmental impact specific from this work. The cost of the experiments and the method is high, though order of magnitudes less than other computer vision fields. We roughly estimated that the total GPU-days used for running all our experiments to 2000, or $\\approx$ 50000 GPU-hours. This amounts to total emissions in the order of 10 tons of $\\mathrm{CO}_{2}\\mathrm{eq}$ . This is excluding the training of the generative model itself, since we did not perform that training. Estimations are conducted using the Machine Learning Impact calculator presented by Lacoste et al. [43]. We do not consider in this approximation: memory storage, CPU-hours, production cost of GPUs/ CPUs, etc.", + "bbox": [ + 498, + 704, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22473", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sahar Abdelnabi and Mario Fritz. Adversarial watermarking transformer: Towards tracing text provenance with data hiding. In 2021 IEEE Symposium on Security and Privacy (SP), pages 121-140. IEEE, 2021. 7", + "[2] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 2, 4", + "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2", + "[4] Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 7, 14", + "[5] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 17", + "[6] Jean Bégaint, Fabien Racapé, Simon Feltman, and Akshay Pushparaja. Compressai: a pytorch library and evaluation platform for end-to-end compression research. arXiv preprint arXiv:2011.03029, 2020. 14", + "[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 1, 2", + "[8] Miles Brundage, Shahar Avin, Jack Clark, Helen Toner, Peter Eckersley, Ben Garfinkel, Allan Dafoe, Paul Scharre, Thomas Zeitzoff, Bobby Filar, et al. The malicious use of artificial intelligence: Forecasting, prevention, and mitigation. arXiv preprint arXiv:1802.07228, 2018. 1", + "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV. IEEE, 2021. 14", + "[10] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In European conference on computer vision, pages 103-120. Springer, 2020. 2", + "[11] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7939-7948, 2020. 7, 14", + "[12] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. arXiv preprint arXiv:2211.00680, 2022. 2, 4, 5", + "[13] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 1, 2, 6, 13, 16", + "[14] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital watermarking and steganography. Morgan Kaufmann, 2007. 2, 6, 7, 8" + ], + "bbox": [ + 86, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Steffen Czolbe, Oswin Krause, Ingemar Cox, and Christian Igel. A loss function for generative neural networks based on watson's perceptual model. Advances in Neural Information Processing Systems, 33:2051-2061, 2020. 4, 15", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 14", + "[17] Emily Denton. Ethical considerations of generative ai. In AI for Content Creation Workshop CVPR. IEEE, 2021. 1", + "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2", + "[19] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 2", + "[20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 7, 14", + "[21] Hany Farid. Lighting (in) consistency of paint by text. arXiv preprint arXiv:2207.13744, 2022. 2", + "[22] Hany Farid. Perspective (in) consistency of paint by text. arXiv preprint arXiv:2206.14617, 2022. 2", + "[23] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Supervised gan watermarking for intellectual property protection. In 2022 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6. IEEE, 2022. 2", + "[24] Pierre Fernandez, Matthijs Douze, Hervé Jégou, and Teddy Furon. Active image indexing. arXiv preprint arXiv:2210.10620, 2022. 14", + "[25] Pierre Fernandez, Alexandre Sablayrolles, Teddy Furon, Hervé Jégou, and Matthijs Douze. Watermarking images in self-supervised latent spaces. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022. 2, 6, 7, 14, 15", + "[26] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 2", + "[27] Teddy Furon, Arnaud Guyader, and Frédéric Cérou. Decoding Fingerprinting Using the Markov Chain Monte Carlo Method. In WIFS - IEEE Workshop on Information Forensics and Security, Tenerife, Spain, Dec. 2012. IEEE. 8", + "[28] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1", + "[29] Matthew Gault. An ai-generated artwork won first place at a state fair fine arts competition, and artists are pissed. Vice," + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22474", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2022. https://www.vice.com/en/article/bvmvqm/an-ai-generated-artwork-won-first-place-at-a-state-fair-fine-arts-competition-and-artists-are-pissed (retrieved 2022-12-14). 1", + "[30] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. *ICLR*, 2014. 14", + "[31] Diego Gragnaniello, Davide Cozzolino, Francesco Marra, Giovanni Poggi, and Luisa Verdoliva. Are gan generated images easy to detect? a critical analysis of the state-of-the-art. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2", + "[32] Luca Guarnera, Oliver Giudice, and Sebastiano Battiato. Deepfake detection by analyzing convolutional traces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 666-667, 2020. 2", + "[33] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[34] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2, 5", + "[35] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[36] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2", + "[37] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 2", + "[38] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2", + "[39] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2", + "[40] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2", + "[41] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 15", + "[42] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2022. 2, 6, 7, 14" + ], + "bbox": [ + 86, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, and Thomas Dandres. Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700, 2019. 8", + "[44] Jae-Eun Lee, Young-Ho Seo, and Dong-Wook Kim. Convolutional neural network-based digital image watermarking adaptive to the resolution of image and watermark. Applied Sciences, 2020. 2", + "[45] Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. arXiv preprint arXiv:1811.00656, 2018. 2", + "[46] Dongdong Lin, Benedetta Tondi, Bin Li, and Mauro Barni. Cycleganwm: A cyclegan watermarking method for ownership verification. arXiv preprint arXiv:2211.13737, 2022. 1, 2, 3", + "[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4, 5, 13", + "[48] Wenqing Liu, Miaojing Shi, Teddy Furon, and Li Li. Defending adversarial examples via dnn bottleneck reinforcement. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1930-1938, 2020. 7", + "[49] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2018. 4", + "[50] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2", + "[51] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In CVPR. IEEE, 2020. 2", + "[52] Jingwei Ma, Lucy Chai, Minyoung Huh, Tongzhou Wang, Ser-Nam Lim, Phillip Isola, and Antonio Torralba. Totems: Physical objects for verifying visual integrity. In European Conference on Computer Vision, pages 164-180. Springer, 2022. 2", + "[53] Sébastien Marcel and Yann Rodriguez. Torchvision the machine-vision package of torch. In International Conference on Multimedia. ACM, 2010. 13", + "[54] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 2", + "[55] Peter Meerwald and Teddy Furon. Towards practical joint decoding of binary Tardos fingerprinting codes. IEEE Transactions on Information Forensics and Security, 7(4):1168-1180, Apr. 2012. 8", + "[56] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 2", + "[57] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22475", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 6, 13, 16", + "[58] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2", + "[59] Guangyu Nie, Changhoon Kim, Yezhou Yang, and Yi Ren. Attributing image generative models using latent fingerprints. arXiv preprint arXiv:2304.09752, 2023. 2", + "[60] Zoe Papakipos and Joanna Bitton. Augly: Data augmentations for robustness. arXiv preprint arXiv:2201.06494, 2022. 13", + "[61] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems 32, 2019. 16", + "[62] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022.2", + "[63] Tong Qiao, Yuyan Ma, Ning Zheng, Hanzhou Wu, Yanli Chen, Ming Xu, and Xiangyang Luo. A novel model watermarking for protecting generative adversarial network. Computers & Security, 127:103102, 2023. 2", + "[64] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[65] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 13", + "[66] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 13", + "[67] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2", + "[68] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2, 4, 6, 7, 8, 14, 16", + "[69] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 13", + "[70] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. Stable diffusion. Accessed Nov. 30 2022, 2022. 1, 7", + "[71] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + ], + "bbox": [ + 86, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 1, 2", + "[72] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Spreading vectors for similarity search. ICML, 2019. 17", + "[73] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2", + "[74] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 13", + "[75] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 14", + "[76] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2", + "[77] Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. Defake: Detection and attribution of fake images generated by text-to-image diffusion models. arXiv preprint arXiv:2210.06998, 2022. 2", + "[78] Changhao Shi, Chester Holtz, and Gal Mishne. Online adversarial purification based on self-supervision. arXiv preprint arXiv:2101.09387, 2021. 7", + "[79] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[80] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2", + "[81] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. *ICLR*, 2021. 13", + "[82] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2149–2159, 2022. 13, 20", + "[83] Gábor Tardos. Optimal probabilistic fingerprint codes. Journal of the ACM (JACM), 55(2):1-24, 2008. 8", + "[84] Yusuke Uchida, Yuki Nagai, Shigeyuki Sakazawa, and Shin'ichi Satoh. Embedding watermarks into deep neural networks. In Proceedings of the 2017 ACM on international conference on multimedia retrieval, pages 269-277, 2017. 2" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22476", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[85] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477, 2022. 2", + "[86] Vedran Vukotic, Vivien Chappelier, and Teddy Furon. Are deep neural networks good for blind image watermarking? In WIFS, 2018. 2", + "[87] Steven Walton, Ali Hassani, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Stylenat: Giving each head a new perspective. arXiv preprint arXiv:2211.05770, 2022. 2", + "[88] Sheng-Yu Wang, Oliver Wang, Andrew Owens, Richard Zhang, and Alexei A Efros. Detecting photoshopped faces by scripting photoshop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10072-10081, 2019. 2", + "[89] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 2", + "[90] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[91] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 8", + "[92] Chen Henry Wu and Fernando De la Torre. Unifying diffusion models' latent space, with applications to cycled diffusion and guidance. arXiv preprint arXiv:2210.05559, 2022. 2", + "[93] Hanzhou Wu, Gen Liu, Yuwei Yao, and Xinpeng Zhang. Watermarking neural networks with watermarked images. IEEE Transactions on Circuits and Systems for Video Technology, 31(7):2591-2601, 2020. 2", + "[94] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, pages 12062-12072. PMLR, 2021. 7", + "[95] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. In ICLR, 2020. 13", + "[96] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2", + "[97] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 2" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[98] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14448-14457, 2021. 2, 3", + "[99] Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. Responsible disclosure of generative models using scalable fingerprinting. In International Conference on Learning Representations (ICLR), 2022. 1, 2", + "[100] Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun, and In So Kweon. Udh: Universal deep hiding for steganography, watermarking, and light field messaging. NeurIPS, 2020. 2", + "[101] Chaoning Zhang, Adil Karjauv, Philipp Benz, and In So Kweon. Towards robust deep hiding under nondifferentiable distortions for practical blind watermarking. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5158-5166, 2021. 4, 13", + "[102] Jie Zhang, Dongdong Chen, Jing Liao, Han Fang, Weiming Zhang, Wenbo Zhou, Hao Cui, and Nenghai Yu. Model watermarking for image processing networks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12805-12812, 2020. 2", + "[103] Kevin Alex Zhang, Alfredo Cuesta-Infante, Lei Xu, and Kalyan Veeramachaneni. Steganogan: High capacity image steganography with gans. arXiv preprint arXiv:1901.03892, 2019. 14", + "[104] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 1, 2", + "[105] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR. IEEE, 2018. 4, 15", + "[106] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In 2019 IEEE international workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2019. 2", + "[107] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deep-fake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. 2", + "[108] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In ECCV, 2018. 2, 3, 4, 6, 7, 13" + ], + "bbox": [ + 503, + 92, + 893, + 739 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "22477", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/c2da89cd-17c8-420c-b57b-7af434285ab7_model.json b/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/c2da89cd-17c8-420c-b57b-7af434285ab7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b4c98c1f4b0acbf85f2d9ae71b752e8433e10aa9 --- /dev/null +++ b/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/c2da89cd-17c8-420c-b57b-7af434285ab7_model.json @@ -0,0 +1,3139 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.131, + 0.845, + 0.154 + ], + "angle": 0, + "content": "The Stable Signature: Rooting Watermarks in Latent Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.18, + 0.88, + 0.2 + ], + "angle": 0, + "content": "Pierre Fernandez\\(^{1,2}\\) Guillaume Couairon\\(^{1,3}\\) Hervé Jégou\\(^{1}\\) Matthijs Douze\\(^{1}\\) Teddy Furon\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.217, + 0.757, + 0.235 + ], + "angle": 0, + "content": "1Meta AI 2Centre Inria de l'Université de Rennes 3Sorbonne University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.255, + 0.314, + 0.272 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.474, + 0.546 + ], + "angle": 0, + "content": "Generative image modeling enables a wide range of applications but raises ethical concerns about responsible deployment. We introduce an active content tracing method combining image watermarking and Latent Diffusion Models. The goal is for all generated images to conceal an invisible watermark allowing for future detection and/or identification. The method quickly fine-tunes the latent decoder of the image generator, conditioned on a binary signature. A pre-trained watermark extractor recovers the hidden signature from any generated image and a statistical test then determines whether it comes from the generative model. We evaluate the invisibility and robustness of the watermarks on a variety of generation tasks, showing that the Stable Signature is robust to image modifications. For instance, it detects the origin of an image generated from a text prompt, then cropped to keep \\(10\\%\\) of the content, with \\(90 + \\%\\) accuracy at a false positive rate below \\(10^{-6}\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.563, + 0.21, + 0.579 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.47, + 0.696 + ], + "angle": 0, + "content": "Recent progress in generative modeling and natural language processing enable easy creation and manipulation of photo-realistic images, such as with DALL-E 2 [64] or Stable Diffusion [68]. They have given birth to many image edition tools like ControlNet [104], Instruct-Pix2Pix [7], and others [13, 28, 71], that are becoming mainstream creative tools for artists, designers, and the general public." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.471, + 0.864 + ], + "angle": 0, + "content": "While this is a great step forward for generative AI, it also undermines confidence in the authenticity or veracity of photo-realistic images. Indeed, methods for photorealistic image edition existed before, but generative AI significantly lowers the barriers to convincing synthetic image generation and edition (e.g. a generated picture recently won an art competition [29]). This raises new risks like deep fakes, impersonation or copyright usurpation [8, 17]. A tool to determine that images are AI-generated would make it easier to ensure their compliance with ethical standards and to remove them from certain platforms." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.277, + 0.877, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.53, + 0.892, + 0.558 + ], + "angle": 0, + "content": "Figure 1. Overview. The latent decoder can be fine-tuned to preemptively embed a signature into all generated images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.893, + 0.748 + ], + "angle": 0, + "content": "A baseline solution to identify generated images is forensics, i.e. methods to detect generated/manipulated images passively (the image is not modified for identification). An active baseline is to apply existing watermarking methods after the image generation. Watermarking invisibly embeds a secret message into the image, which can then be extracted and used to identify the image. This has several drawbacks. If the model leaks or is open-sourced, the post-generation watermarking can be removed trivially. The open source Stable Diffusion [70] is a case in point, since removing the watermark amounts to commenting out a single line in the source code." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Our Stable Signature method merges watermarking into the generation process itself, without any architectural change. It adjusts the pre-trained generative model such that all the images it produces conceal a given watermark. There are several advantages to this approach [46, 99]. It does not require additional processing of the generated image, which makes the watermarking computationally lighter, straightforward, and secure. Model providers could deploy their models to different user groups with a unique watermark, and monitor that they are used in a responsible manner." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.478, + 0.901 + ], + "angle": 0, + "content": "*Work supported by ANR / AID under Chaire SAIDA ANR-20-CHIA-0011. Correspondance to pfz@meta.com" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22466" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "They could give art platforms, news outlets and other sharing platforms the ability to detect when an image has been generated by their AI." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.348 + ], + "angle": 0, + "content": "We focus on Latent Diffusion Models (LDM) [68] that can perform a wide range of generative tasks. We show that simply fine-tuning a small part of the generative model – the decoder that generates images from the latent vectors – is enough to natively embed a watermark into generated images. Stable Signature does not require an architectural change and does not modify the diffusion process. Hence it is to our knowledge compatible with all LDM-based generative methods [7, 13, 62, 71, 104]. The fine-tuning stage is performed by back-propagating a combination of a perceptual image loss and a message decoding loss from a watermark extractor back to the LDM decoder. We pre-train the extractor with a simplified version of the deep watermarking method HiDDeN [108]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.471, + 0.529 + ], + "angle": 0, + "content": "We create an evaluation benchmark close to real world situations where images may be edited. The tasks are: detection of AI generated images, tracing models from their generations. For instance, we detect \\(90\\%\\) of images generated with the generative model, even if they are cropped to \\(10\\%\\) of their original size, while flagging only one false positive every \\(10^{6}\\) images. To ensure that the model's utility is not weakened, we show that the FID [34] score of the generation is not affected and that the generated images are perceptually indistinguishable from the ones produced by the original model. This is done over several tasks involving LDM (text-to-image, inpainting, edition, etc.)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.651 + ], + "angle": 0, + "content": "As a summary, (1) we efficiently merge watermarking into the generation process of LDMs, in a way that is compatible with most of the LDM-based generative methods; (2) we demonstrate how it can be used to detect and trace generated images, through a real-world evaluation benchmark; (3) we compare to post-hoc watermarking methods, showing that it is competitive while being more secure and efficient, and (4) evaluate robustness to intentional attacks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.664, + 0.22, + 0.68 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.795 + ], + "angle": 0, + "content": "Image generation has long been dominated by GANs, still state-of-the-art on many datasets [37, 38, 39, 76, 87]. Transformers have also been successfully used for modeling image [66, 19] or video [79] distributions, providing higher diversity at the expense of increased inference time. Images are typically converted to token lists using vector-quantized architectures [20, 67, 96], relying on an image decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Diffusion models [18, 36, 58, 80] are a huge improvement in text-conditional image generation, current models synthesize high-resolution photo-realistic images for a wide variety of text prompts [3, 35, 65, 69, 74]. They can also perform conditional image generation – like inpainting or text-guided image editing – by fine-tuning the diffusion model with additional conditioning, e.g. masked input im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "age, segmentation map, etc. [50, 73]. Because of their iterative denoising algorithm, diffusion models can also be adapted for image editing in a zero-shot fashion by guiding the generative process [13, 33, 40, 56, 85, 92]. All these methods, when combined with Stable Diffusion, operate in the latent rather than image space, requiring a latent decoder to produce an RGB image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.892, + 0.329 + ], + "angle": 0, + "content": "Detection of AI-generated/manipulated images is notably active in the context of deep-fake detection [32, 107]. Many works focus on the detection of GAN-generated images [10, 31, 89, 106]. One approach is to detect inconsistencies, via lights, perspective or physical objects [21, 22, 45, 52, 88]. These approaches are restricted to photorealistic images or faces, artworks not intended to be physically correct are not covered." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.893, + 0.513 + ], + "angle": 0, + "content": "Other approaches track traces left by the generators in the spatial [54, 97] or frequency [26, 106] domains. There are extensions to diffusion models in recent works [12, 77] that show encouraging results. However purely relying on forensics and passive detection is limiting, e.g. the best performing method to our knowledge [12] is able to detect \\(50\\%\\) of generated images for a false positive rate around 1/100: if a user-generated content platform were to receive 1 billion images every day, it would wrongly flag 10 million images to detect only half of the generated images. Besides, passive techniques cannot trace images from different versions of the same model, in contrast with watermarking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Image watermarking has long been studied in the context of tracing and intellectual property protection [14]. More recently, deep learning encoder/extractors like HiD-DeN [2, 44, 51, 100, 108] or iterative methods by Vukotic et al. [25, 42, 86] showed competitive results in terms of robustness to a wide range of transformations, namely geometric ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.892, + 0.873 + ], + "angle": 0, + "content": "In the specific case of generative models, some works aim to watermark the training set on which the generative model is learned [98]. It is highly inefficient since every new message to embed requires a new training pipeline. Merging the watermarking and the generative process is a recent idea [23, 46, 59, 63, 93, 99, 102], that is closer to the model watermarking literature [84]. They suffer from two strong limitations. First, these methods only apply to GANs, while LDM are progressively replacing them for most applications. Second, watermarking is incorporated in the training process of the GAN from the start. This strategy is unsustainable because the generative model training is more and more costly1. Our work shows that a quick finetuning of the latent decoder part of the generative model is enough to achieve a good watermarking performance, provided that the watermark extractor is well chosen." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.892, + 0.9 + ], + "angle": 0, + "content": "1 Stable Diffusion training costs \\(\\sim\\)\\\\(600k of cloud compute (Wikipedia)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22467" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.393, + 0.108 + ], + "angle": 0, + "content": "3. Problem Statement & Background" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.116, + 0.469, + 0.177 + ], + "angle": 0, + "content": "Figure 1 shows a model provider Alice who deploys a latent diffusion model to users Bobs. Stable Signature embeds a binary signature into the generated images. This section derives how Alice can use this signature for two scenarios:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.181, + 0.469, + 0.243 + ], + "angle": 0, + "content": "- Detection: \"Is it generated by my model?\" Alice detects if an image was generated by her model. Generated images should be flagged as reliably as possible, while controlling the probability of flagging a natural image." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.244, + 0.468, + 0.288 + ], + "angle": 0, + "content": "- Identification: \"Who generated this image?\". Alice monitors who created each image, while avoiding to mistakenly identifying a Bob." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.181, + 0.469, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.3, + 0.377, + 0.316 + ], + "angle": 0, + "content": "3.1. Image watermarking for detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.469, + 0.4 + ], + "angle": 0, + "content": "Alice embeds a \\( k \\)-bit binary signature into the generated images. The watermark extractor then decodes messages from the images it receives and detects when the message is close to Alice's signature. An example application is to block AI-generated images on a content sharing platform." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.469, + 0.471 + ], + "angle": 0, + "content": "Statistical test Let \\( m \\in \\{0,1\\}^k \\) be Alice's signature. We extract the message \\( m' \\) from an image \\( x \\) and compare it to \\( m \\). As done in previous works [46, 98], the detection test relies on the number of matching bits \\( M(m,m') \\): if" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.474, + 0.469, + 0.491 + ], + "angle": 0, + "content": "\\[\nM \\left(m, m ^ {\\prime}\\right) \\geq \\tau \\text {w h e r e} \\tau \\in \\{0, \\dots , k \\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.468, + 0.523 + ], + "angle": 0, + "content": "then the image is flagged. This provides a level of robustness to imperfections of the watermarking." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.469, + 0.706 + ], + "angle": 0, + "content": "Formally, we test the statistical hypothesis \\( H_{1} \\): \" \\( x \\) was generated by Alice's model\" against the null hypothesis \\( H_{0} \\): \" \\( x \\) was not generated by Alice's model\". Under \\( H_{0} \\) (i.e. for vanilla images), we assume that bits \\( m_{1}', \\ldots, m_{k}' \\) are (i.i.d.) Bernoulli random variables with parameter 0.5. Then \\( M(m, m') \\) follows a binomial distribution with parameters \\( (k, 0.5) \\). We verify this assumption experimentally in App. B.5. The False Positive Rate (FPR) is the probability that \\( M(m, m') \\) takes a value bigger than the threshold \\( \\tau \\). It is obtained from the CDF of the binomial distribution, and a closed-form can be written with the regularized incomplete beta function \\( I_{x}(a; b) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.711, + 0.469, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\operatorname {F P R} (\\tau) = \\mathbb {P} (M > \\tau | H _ {0}) = I _ {1 / 2} (\\tau + 1, k - \\tau). \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.831, + 0.107 + ], + "angle": 0, + "content": "3.2. Image watermarking for identification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.206 + ], + "angle": 0, + "content": "Alice now embeds a signature \\( m^{(i)} \\) drawn randomly from \\( \\{0,1\\}^k \\) into the model distributed to \\( \\mathrm{Bob}^{(i)} \\) (for \\( i = 1 \\cdots N \\), with \\( N \\) the number of Bobs). Alice can trace any misuse of her model: generated images violating her policy (gore content, deepfakes) are linked back to the specific Bob by comparing the extracted message to Bobs' signatures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.218, + 0.892, + 0.338 + ], + "angle": 0, + "content": "Statistical test We compare the message \\( m' \\) from the watermark extractor to \\( (m^{(1)}, \\ldots, m^{(N)}) \\). There are now \\( N \\) detection hypotheses to test. If the \\( N \\) hypotheses are rejected, we conclude that the image was not generated by any of the models. Otherwise, we attribute the image to \\( \\operatorname{argmax}_{i=1..N} M(m', m^{(i)}) \\). With regards to the detection task, false positives are more likely since there are \\( N \\) tests. The global FPR at a given threshold \\( \\tau \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.343, + 0.892, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\operatorname {F P R} (\\tau , N) = 1 - (1 - \\operatorname {F P R} (\\tau)) ^ {N} \\approx N. \\operatorname {F P R} (\\tau). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.892, + 0.441 + ], + "angle": 0, + "content": "Equation (3) (resp. (2)), is used reversely: we find threshold \\(\\tau\\) to achieve a required FPR for identification (resp. detection). Note that these formulae hold only under the assumption of i.i.d. Bernoulli bits extracted from vanilla images. This condition is enforced in the next section." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.456, + 0.59, + 0.471 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.572 + ], + "angle": 0, + "content": "Stable Signature modifies the generative network so that the generated images have a given signature through a fixed watermark extractor. It is trained in two phases. First, we create the watermark extractor network \\(\\mathcal{W}\\). We then fine-tune the LDM decoder \\(\\mathcal{D}\\), such that all generated images yield a given signature through \\(\\mathcal{W}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.583, + 0.825, + 0.599 + ], + "angle": 0, + "content": "4.1. Pre-training the watermark extractor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.891, + 0.697 + ], + "angle": 0, + "content": "We use HiDDeN [108], a classical method in the deep watermarking literature. It jointly optimizes the parameters of watermark encoder \\(\\mathcal{W}_E\\) and extractor network \\(\\mathcal{W}\\) to embed \\(k\\)-bit messages into images, robustly to transformations that are applied during training. We discard \\(\\mathcal{W}_E\\) after training, since only \\(\\mathcal{W}\\) serves our purpose." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.697, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Formally, \\(\\mathcal{W}_E\\) takes as inputs a cover image \\(x_o \\in \\mathbb{R}^{W \\times H \\times 3}\\) and a \\(k\\)-bit message \\(m \\in \\{0,1\\}^k\\). Similar to" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.747, + 0.316, + 0.76 + ], + "angle": 0, + "content": "(a) Pre-train watermark encoder/extractor" + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.763, + 0.4, + 0.86 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.748, + 0.555, + 0.76 + ], + "angle": 0, + "content": "(b) Fine-tune LDM decoder" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.763, + 0.703, + 0.859 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.749, + 0.788, + 0.76 + ], + "angle": 0, + "content": "(c) Generate" + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.767, + 0.832, + 0.788 + ], + "angle": 0, + "content": "Text - Super-resolution Inpainting - Editing - ..." + }, + { + "type": "image", + "bbox": [ + 0.849, + 0.767, + 0.877, + 0.788 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.792, + 0.855, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.87, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Figure 2. Steps of the method. (a) We pre-train a watermark encoder \\(\\mathcal{W}_E\\) and extractor \\(\\mathcal{W}\\), to extract binary messages. (b) We fine-tune the decoder \\(\\mathcal{D}\\) of the LDM's auto-encoder with a fixed signature \\(m\\) such that all the generated images (c) lead to \\(m\\) through \\(\\mathcal{W}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22468" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.244 + ], + "angle": 0, + "content": "ReDMark [2], \\(\\mathcal{W}_E\\) outputs a residual image \\(\\delta\\) of the same size as \\(x_{o}\\), that is multiplied by a factor \\(\\alpha\\) to produce watermarked image \\(x_w = x_o + \\alpha \\delta\\). At each optimization step an image transformation \\(T\\) is sampled from a set \\(\\mathcal{T}\\) that includes common image processing operations such as cropping and JPEG compression. A \"soft\" message is extracted from the transformed image: \\(m' = \\mathcal{W}(T(x_w))\\) (at inference time, the decoded message is given by the signs of the components of \\(m'\\)). The message loss is the Binary Cross Entropy (BCE) between \\(m\\) and the sigmoid \\(\\sigma(m')\\):" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.244, + 0.47, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m} = - \\sum_ {i = 1} ^ {k} m _ {i} \\cdot \\log \\sigma (m _ {i} ^ {\\prime}) + (1 - m _ {i}) \\cdot \\log (1 - \\sigma (m _ {i} ^ {\\prime})).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.47, + 0.339 + ], + "angle": 0, + "content": "The network architectures are kept simple to ease the LDM fine-tuning in the second phase. They are the same as HiDDeN [108] (see App. A.1) with two changes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.429 + ], + "angle": 0, + "content": "First, since \\(\\mathcal{W}_E\\) is discarded, its perceptual quality is not as important, so the perceptual loss or the adversarial network are not needed. Instead, the distortion is constrained by a tanh function on output of \\(\\mathcal{W}_E\\) and by the scaling factor \\(\\alpha\\). This improves the bit accuracy of the recovered message and makes it possible to increase its size \\(k\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.522 + ], + "angle": 0, + "content": "Second, we observed that \\(\\mathcal{W}\\)'s output bits for vanilla images are correlated and highly biased, which violates the assumptions of Sec. 3.1. Therefore, to get closer to i.i.d. dimensions we remove the bias and decorrelate the outputs of \\(\\mathcal{W}\\) by applying a PCA whitening transformation (more details in App. A.1)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.53, + 0.37, + 0.546 + ], + "angle": 0, + "content": "4.2. Fine-tuning the generative model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.658 + ], + "angle": 0, + "content": "In LDM, the diffusion happens in the latent space of an auto-encoder. The latent vector \\(z\\) obtained at the end of the diffusion is input to decoder \\(\\mathcal{D}\\) to produce an image. Here we fine-tune \\(\\mathcal{D}\\) such that the image contains a given message \\(m\\) that can be extracted by \\(\\mathcal{W}\\). Stable Signature is compatible with many generative tasks, since modifying only \\(\\mathcal{D}\\) does not affect the diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.469, + 0.703 + ], + "angle": 0, + "content": "First, we fix the signature \\(m = (m_1,\\ldots ,m_k)\\in \\{0,1\\} ^k\\) The fine-tuning of \\(\\mathcal{D}\\) into \\(\\mathcal{D}_m\\) is inspired by the original training of the auto-encoder in LDM [68]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.469, + 0.795 + ], + "angle": 0, + "content": "Training image \\(x \\in \\mathbb{R}^{H \\times W \\times 3}\\) is fed to the LDM encoder \\(\\mathcal{E}\\) that outputs activation map \\(z = \\mathcal{E}(x) \\in \\mathbb{R}^{h \\times w \\times c}\\), downsampled by a power-of-two factor \\(f = H / h = W / w\\). The decoder reconstructs an image \\(x' = \\mathcal{D}_m(z)\\) and the extractor recovers \\(m' = \\mathcal{W}(x')\\). The message loss is the BCE between \\(m'\\) and the original \\(m\\): \\(\\mathcal{L}_m = \\mathrm{BCE}(\\sigma(m'), m)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.855 + ], + "angle": 0, + "content": "In addition, the original decoder \\(\\mathcal{D}\\) reconstructs the image without watermark: \\(x_{o}^{\\prime} = \\mathcal{D}(z)\\). The image perceptual loss \\(\\mathcal{L}_{\\mathrm{i}}\\) between \\(x^{\\prime}\\) and \\(x_{o}^{\\prime}\\), controls the distortion. We use the Watson-VGG perceptual loss introduced by Czolbe et" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.091, + 0.892, + 0.15 + ], + "angle": 0, + "content": "al. [15], an improved version of LPIPS [105]. It is essential that the decoder learns luminance and contrast masking to add less visible watermarks. The weights of \\(\\mathcal{D}_m\\) are optimized to minimize" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.15, + 0.892, + 0.165 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {m}} + \\lambda_ {\\mathrm {i}} \\mathcal {L} _ {\\mathrm {i}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.894, + 0.259 + ], + "angle": 0, + "content": "This is done over 100 iterations with the AdamW optimizer [49] and batch of size 4, i.e. the fine-tuning sees fewer than 500 images and takes one minute on a single GPU. The learning rate follows a cosine annealing schedule with 20 iterations of linear warmup to \\(10^{-4}\\) and decays to \\(10^{-6}\\). The factor \\(\\lambda_{\\mathrm{i}}\\) in (4) is set to 0.2 by default." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.272, + 0.885, + 0.29 + ], + "angle": 0, + "content": "5. Text-to-Image Watermarking Performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.892, + 0.388 + ], + "angle": 0, + "content": "This section discusses our method's detection and identification capability for images generated by a Stable-Diffusion-like model \\([68]^3\\). We apply generative models watermarked with 48-bit signatures on prompts of the MS-COCO [47] validation set. We evaluate detection and identification on the outputs, as illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.479 + ], + "angle": 0, + "content": "We evaluate their robustness to different transformations applied to generated images: strong cropping (10% of the image remaining), brightness shift (strength factor 2.0), as well as a combination of crop 50%, brightness shift 1.5 and JPEG 80. This covers typical geometric and photometric edits (see Fig. 5 for visual examples)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.479, + 0.892, + 0.509 + ], + "angle": 0, + "content": "The detection rates are partly obtained from experiments and partly by extrapolating small-scale measurements." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.518, + 0.666, + 0.532 + ], + "angle": 0, + "content": "5.1. Detection results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.892, + 0.707 + ], + "angle": 0, + "content": "For detection, we fine-tune the decoder of the LDM with a random key \\( m \\), generate 1000 images and use the test of Eq. (1). We report the tradeoff between True Positive Rate (TPR), i.e. the probability of flagging a generated image and the FPR, while varying \\( \\tau \\in \\{0,..,48\\} \\). For instance, for \\( \\tau = 0 \\), we flag all images so \\( \\mathrm{FPR} = 1 \\), and \\( \\mathrm{TPR} = 1 \\). The TPR is measured directly, while the FPR is inferred from Eq. (2), because it would otherwise be too small to be measured on reasonably sized problems (this approximation is validated experimentally in App. B.6). The experiment is run on 10 random signatures and we report averaged results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.708, + 0.892, + 0.844 + ], + "angle": 0, + "content": "Figure 3 shows the tradeoff under image transformations. For example, when the generated images are not modified, Stable Signature detects \\(99\\%\\) of them, while only 1 vanilla image out of \\(10^{9}\\) is flagged. At the same FPR \\(= 10^{-9}\\), Stable Signature detects \\(84\\%\\) of generated images for a crop that keeps \\(10\\%\\) of the image, and \\(65\\%\\) for a transformation that combines a crop, a color shift, and a JPEG compression. For comparison, we report results of a state-of-the-art passive method [12], applied on resized and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.47, + 0.901 + ], + "angle": 0, + "content": "2The transformation needs to be differentiable in pixel space. This is not the case for JPEG compression so we use the forward attack simulation layer introduced by Zhang et al. [101]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.851, + 0.892, + 0.901 + ], + "angle": 0, + "content": "3We refrain from experimenting with pre-existing third-party generative models, such as Stable Diffusion or LDMs, and instead use a large diffusion model (2.2B parameters) trained on an internal dataset of 330M licensed image-text pairs." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22469" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "compressed images. As expected, we observe that these baseline results have orders of magnitudes larger FPR than Stable Signature, which actively marks the content." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.152, + 0.273, + 0.167 + ], + "angle": 0, + "content": "5.2. Identification results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.178, + 0.469, + 0.269 + ], + "angle": 0, + "content": "Each Bob has its own copy of the generative model. Given an image, the goal is to find if any of the \\( N \\) Bobs created it (detection) and if so, which one (identification). There are 3 types of error: false positive: flag a vanilla image; false negative: miss a generated image; false accusation: flag a generated image but identify the wrong Bob." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.271, + 0.469, + 0.437 + ], + "angle": 0, + "content": "For evaluation, we fine-tune \\( N' = 1000 \\) models with random signatures. Each model generates 100 images. For each of these 100k watermarked images, we extract the Stable Signature message, compute the matching score with all \\( N \\) signatures and select the Bob with the highest score. The image is predicted to be generated by that Bob if this score is above threshold \\( \\tau \\). We determined \\( \\tau \\) such that \\( \\mathrm{FPR} = 10^{-6} \\), see Eq. (3). For example, for \\( N = 1 \\), \\( \\tau = 41 \\) and for \\( N = 1000 \\), \\( \\tau = 44 \\). Accuracy is extrapolated beyond the \\( N' \\) Bobs by adding additional signatures and having \\( N > N' \\) (e.g. Bobs that have not generated any images)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.651 + ], + "angle": 0, + "content": "Figure 4 reports the per-transformation identification accuracy. For example, we identify a Bob among \\( N = 10^{5} \\) with \\( 98\\% \\) accuracy when the image is not modified. Note that for the combined edit, this becomes \\( 40\\% \\). This may still be dissuasive: if a Bob generates 3 images, he will be identified \\( 80\\% \\) of the time. We observe that at this scale, the false accusation rate is zero, i.e. we never identify the wrong Bob. This is because \\( \\tau \\) is set high to avoid FPs, which also makes false accusations unlikely. We observe that the identification accuracy decreases when \\( N \\) increases, because the threshold \\( \\tau \\) required to avoid false positives is higher when \\( N \\) increases, as pointed out by the approximation in (3). In a nutshell, by distributing more models, Alice trades some accuracy of detection against the ability to identify Bobs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.706, + 0.108 + ], + "angle": 0, + "content": "6. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.116, + 0.892, + 0.207 + ], + "angle": 0, + "content": "The previous section showed how to leverage watermarks for detection and identification of images generated from text prompts. We now present more general results on robustness and image quality for different generative tasks. We also compare Stable Signature to post-generation watermarking algorithms." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.216, + 0.744, + 0.231 + ], + "angle": 0, + "content": "6.1. Tasks & evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.239, + 0.892, + 0.33 + ], + "angle": 0, + "content": "Since our method only involves the LDM decoder, it is compatible with many generative tasks. We evaluate text-to-image generation and image edition on the validation set of MS-COCO [47], super-resolution and inpainting on the validation set of ImageNet [16] (all evaluation details are available in App. A.3)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.33, + 0.892, + 0.466 + ], + "angle": 0, + "content": "We evaluate the image distortion with the Peak Signal-to-Noise Ratio (PSNR), which is defined as \\(\\mathrm{PSNR}(x, x') = -10 \\cdot \\log_{10}(\\mathrm{MSE}(x, x'))\\), for \\(x, x' \\in [0, 1]^{c \\times h \\times w}\\), as well as Structural Similarity score (SSIM) [90]. They compare images generated with and without watermark. On the other hand, we evaluate the diversity and quality of the generated images with the Fréchet Inception Distance (FID) [34]. The bit accuracy – the percentage of bits correctly decoded – evaluates the watermarks' robustness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.476, + 0.729, + 0.492 + ], + "angle": 0, + "content": "6.2. Image generation quality" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.499, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Figure 6 shows qualitative examples of how the image generation is altered by the latent decoder's fine-tuning. The difference is very hard to perceive even for a trained eye. This is surprising for such a low PSNR, especially since the watermark embedding is not constrained by any Human Visual System like in professional watermarking techniques. Most interestingly, the LDM decoder has indeed learned to add the watermark signal only over textured areas where the human eyes are not sensitive, while the uniform backgrounds are kept intact (see the pixel-wise difference)." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.677, + 0.341, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.843, + 0.35, + 0.899 + ], + "angle": 0, + "content": "Figure 3. Detection results. TPR/FPR curve of the detection under different transformations. Forensics† indicates passive detection (without watermark) [12]." + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.676, + 0.63, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.843, + 0.64, + 0.899 + ], + "angle": 0, + "content": "Figure 4. Identification results. Proportion of well-identified Bobs. Detection with \\(\\mathrm{FPR} = 10^{-6}\\) is run beforehand, and we consider it an error if the image is not flagged." + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.677, + 0.773, + 0.767 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.769, + 0.73, + 0.781 + ], + "angle": 0, + "content": "Crop \\(10\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.677, + 0.882, + 0.767 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.769, + 0.843, + 0.78 + ], + "angle": 0, + "content": "Combined" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.781, + 0.867, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.84, + 0.892, + 0.896 + ], + "angle": 0, + "content": "Figure 5. Transformations evaluated in Sec. 5 & 6. 'Combined' is made of crop \\(50\\%\\), brightness adjustment 1.5 and JPEG 80 compression." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22470" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.638, + 0.306 + ], + "angle": 0, + "content": "
PSNR / SSIM ↑FID ↓Bit accuracy ↑ on:
NoneCropBrigh.Comb.
TasksText-to-ImageLDM [68]30.0 / 0.8919.6 (−0.3)0.990.950.970.92
Image EditionDiffEdit [13]31.2 / 0.9215.0 (−0.3)0.990.950.980.94
Inpainting - Full - Mask onlyGlide [57]31.1 / 0.9116.8 (+0.6)0.990.970.980.93
37.8 / 0.989.0 (+0.1)0.890.760.840.78
Super-ResolutionLDM [68]34.0 / 0.9411.6 (+0.0)0.980.930.960.92
WM MethodsPost generation
Dct-Dwt [14]0.14 (s/img)39.5 / 0.9719.5 (−0.4)0.860.520.510.51
SSL Watermark [25]0.45 (s/img)31.1 / 0.8620.6 (+0.7)1.000.730.930.66
FNNS [42]0.28 (s/img)32.1 / 0.9019.0 (−0.9)0.930.930.910.93
HiDDeN [108]0.11 (s/img)32.0 / 0.8819.7 (−0.2)0.990.970.990.98
Merged in generation
Stable Signature0.00 (s/img)30.0 / 0.8919.6 (−0.3)0.990.950.970.92
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.645, + 0.089, + 0.895, + 0.312 + ], + "angle": 0, + "content": "Table 1. Generation quality and comparison to post-hoc watermarking on \\(512\\times 512\\) images and 48-bit signatures. PSNR and SSIM are computed between generations of the original and watermarked generators. For FID, we show in (color) the difference with regards to original. Post-hoc watermarking is evaluated on text-generated images. (App. B.2 gives results on more transformations, and App. A gives more details on the evaluations.) Overall, Stable Signature has minimal impact on generation quality. It has comparable robustness to post-hoc methods while being rooted in the generation itself." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.47, + 0.426 + ], + "angle": 0, + "content": "Table 1 presents a quantitative evaluation of image generation on the different tasks. We report the FID, and the average PSNR and SSIM that are computed between the images generated by the fine-tuned LDM and the original one. The results show that no matter the task, the watermarking has very small impact on the FID of the generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.471, + 0.584 + ], + "angle": 0, + "content": "The average PSNR is around 30 dB and SSIM around 0.9 between images generated by the original and a watermarked model. This is a bit low from a watermarking perspective, because we do not explicitly optimize for them. However, in a real world scenario, one would only have the watermarked version of the image. Therefore it is not as important to be extremely close to the original image, we just need to generate artifacts-free images. Without access to the image generated by the original LDM, it is very hard to tell whether a watermark is present or not." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.619, + 0.464, + 0.835 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.843, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Figure 6. Images generated with Stable Diffusion. The PSNR is \\(35.4\\mathrm{dB}\\) in the first row and \\(28.6\\mathrm{dB}\\) in the second. Images generated with Stable Signature look natural because modified areas are located where the eye is not sensitive. More examples in App. C." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.331, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Table 2. Watermark robustness on image transformations applied before decoding, details of which are available in App. A.3. We report the bit accuracy, averaged over \\(10 \\times 1k\\) images generated from COCO prompts with 10 different keys." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.393, + 0.88, + 0.452 + ], + "angle": 0, + "content": "
AttackBit acc.Comb.0.92Sharpness 2.00.99
None0.99Bright. 2.00.97Med. Filter k=70.94
Crop 0.10.95Cont. 2.00.98Resize 0.70.91
JPEG 500.88Sat. 2.00.99Text overlay0.99
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.713, + 0.481 + ], + "angle": 0, + "content": "6.3. Watermark robustness" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.61 + ], + "angle": 0, + "content": "We evaluate the robustness of the watermark to different image transformations applied before extraction. For each task, we generate 1k images with 10 models fine-tuned for different messages, and report the average bit accuracy in Table 1. Additionally, Table 2 reports results on more image transformations for images generated from COCO prompts. The main transformations are presented in Fig. 5 (more evaluation details are available in App. A.3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.746 + ], + "angle": 0, + "content": "We see that the watermark is indeed robust for several tasks and across transformations. The bit accuracy is always above 0.9, except for inpainting, when replacing only the masked region of the image (between \\(1 - 50\\%\\) of the image, with an average of \\(27\\%\\) across masks). Besides, the bit accuracy is not perfect even without edition, mainly because there are images that are hard to watermark (e.g. the ones that are very uniform, like the background in Fig. 6) and for which the accuracy is lower." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Note that the robustness comes even without any transformation during the LDM fine-tuning phase: it is due to the watermark extractor. If the watermark embedding pipeline is learned to be robust against an augmentation, then the LDM learns how to produce watermarks that are robust against it during fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.836, + 0.865 + ], + "angle": 0, + "content": "6.4. Comparison to post-hoc watermarking" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "An alternative way to watermark generated images is to process them after the generation (post-hoc). This may" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22471" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.089, + 0.442, + 0.102 + ], + "angle": 0, + "content": "Table 3. Quality-robustness trade-off during fine-tuning." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.102, + 0.447, + 0.151 + ], + "angle": 0, + "content": "
λi for fine-tuning0.80.40.20.10.050.025
PSNR ↑31.430.629.728.526.824.6
Bit acc. ↑ on ‘comb.’0.850.880.900.920.940.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.165, + 0.446, + 0.178 + ], + "angle": 0, + "content": "Table 4. Role of the attack simulation layer at pre-training." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.178, + 0.457, + 0.243 + ], + "angle": 0, + "content": "
Seen at W trainingBit accuracy ↑ at test time:
Crop 0.1Rot. 90JPEG 50Bright. 2.0Res. 0.7
X1.000.560.500.990.48
✓1.000.990.900.990.91
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.468, + 0.394 + ], + "angle": 0, + "content": "be simpler, but less secure and efficient than Stable Signature. We compare our method to a frequency based method, DCT-DWT [14], iterative approaches (SSL Watermark [25] and FNNS [42]), and an encoder/decoder one like HiD-DeN [108]. We choose DCT-DWT since it is employed by the original open source release of Stable Diffusion [70], and the other methods because of their performance and their ability to handle arbitrary image sizes and number of bits. We use our implementations (see details in App. A.4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.468, + 0.561 + ], + "angle": 0, + "content": "Table 1 compares the generation quality and the robustness over \\(5\\mathrm{k}\\) generated images. Overall, Stable Signature achieves comparable results in terms of robustness. HiD-DeN's performance is a bit higher but its output bits are not i.i.d. meaning that it cannot be used with the same guarantees as the other methods. We also observe that post-hoc generation gives worse qualitative results, images tend to present artifacts (see Fig. 13 in the supplement). One explanation is that Stable Signature is merged into the high-quality generation process with the LDM auto-encoder model, which is able to modify images in a more subtle way." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.571, + 0.446, + 0.587 + ], + "angle": 0, + "content": "6.5. Can we trade image quality for robustness?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.468, + 0.716 + ], + "angle": 0, + "content": "We can choose to maximize the image quality or the robustness of the watermark thanks to the weight \\(\\lambda_{\\mathrm{i}}\\) of the perceptual loss in (4). We report the average PSNR of 1k generated images, as well as the bit accuracy obtained on the extracted message for the 'Combined' editing applied before detection (qualitative results are in App. B.1). A higher \\(\\lambda_{\\mathrm{i}}\\) leads to an image closer to the original one, but to lower bit accuracies on the extracted message, see Table 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.295, + 0.743 + ], + "angle": 0, + "content": "6.6. Attack simulation layer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Watermark robustness against image transformations depends solely on the watermark extractor. Here, we pretrain them with or without specific transformations in the simulation layer, on a shorter schedule of 50 epochs, with \\(128 \\times 128\\) images and 16-bits messages. From there, we plug them in the LDM fine-tuning stage and we generate 1k images from text prompts. We report the bit accuracy of the extracted watermarks in Table 4. The extractor is naturally robust to some transformations, such as crops or brightness, without being trained with them, while others, like rotations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "or JPEG, require simulation during training for the watermark to be recovered at test time. Empirically we observed that adding a transformation improves results for the latter, but makes training more challenging." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.168, + 0.882, + 0.185 + ], + "angle": 0, + "content": "7. Attacks on Stable Signature's Watermarks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.194, + 0.892, + 0.33 + ], + "angle": 0, + "content": "We examine the watermark's resistance to intentional tampering, as opposed to distortions that happen without bad intentions like crops or compression (discussed in Sec. 5). We consider two threat models: one is typical for many image watermarking methods [14] and operates at the image level, and another targets the generative model level. For image-level attacks, we evaluate on \\(5\\mathrm{k}\\) images generated from COCO prompts. Full details on the following experiments can be found in Appendix A.5." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.341, + 0.686, + 0.358 + ], + "angle": 0, + "content": "7.1. Image-level attacks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Watermark removal. Bob alters the image to remove the watermark with deep learning techniques, like methods used for adversarial purification [78, 94] or neural auto-encoders [1, 48]. Note that this kind of attacks has not been explored in the image watermarking literature to our knowledge. Figure 7 evaluates the robustness of the watermark against neural auto-encoders [4, 11, 20, 68] at different compression rates. To reduce the bit accuracy closer to random (50%), the image distortion needs to be strong (PSNR<26). However, assuming the attack is informed on the generative model, i.e. the auto-encoder is the same as the one used to generate the images, the attack becomes much more effective. It erases the watermark while achieving high quality (PSNR>29). This is because the image is modified precisely in the bandwidth where the watermark" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.622, + 0.875, + 0.793 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.801, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Figure 7. Removal attacks. \\( x_{o} \\) is the image produced by the original generator, \\( x_{r} \\) is the version produced by the watermarked generator and then attacked. Bit accuracy is on the watermark extracted from \\( x_{r} \\). Neural auto-encoders [4, 11, 20] follow the same trend, except for the one used by LDM ('KL-f8' for our LDM). When access to the watermark extractor is granted, adversarial attacks also remove the watermark at lower PSNR budget." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22472" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "is embedded. Note that this assumption is strong, because Alice does not need to distribute the original generator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.135, + 0.469, + 0.286 + ], + "angle": 0, + "content": "Watermark removal & embedding (white-box). To go further, we assume that the attack is informed on the watermark extractor - e.g. because it has leaked. Bob can use an adversarial attack to remove the watermark by optimizing the image under a PSNR constraint. The objective is to minimize the \\(\\ell_2\\) distance between a random binary message sampled beforehand and the extractor's output, effectively replacing the original signature with a random one. It makes it possible to erase the watermark with a lower distortion budget, as seen in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.47, + 0.362 + ], + "angle": 0, + "content": "Instead of removing the watermark, an attacker could embed a signature into vanilla images (unauthorized embedding [14]) to impersonate another Bob of whom they have a generated image. It highlights the importance of keeping the watermark extractor private." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.373, + 0.282, + 0.387 + ], + "angle": 0, + "content": "7.2. Network-level attacks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.469, + 0.487 + ], + "angle": 0, + "content": "Model purification. Bob gets Alice's generative model and uses a fine-tuning process akin to Sec. 4.2 to eliminate the watermark embedding – that we coin model purification. This involves removing the message loss \\( \\mathcal{L}_{\\mathrm{m}} \\), and shifting the focus to the perceptual loss \\( \\mathcal{L}_{\\mathrm{i}} \\) between the original image and the one reconstructed by the LDM auto-encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.469, + 0.578 + ], + "angle": 0, + "content": "Figure 8 shows the results of this attack for the MSE loss. The PSNR between the watermarked and purified images is plotted at various stages of fine-tuning. Empirically, it is difficult to significantly reduce the bit accuracy without compromising the image quality: artifacts start to appear during the purification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.469, + 0.698 + ], + "angle": 0, + "content": "Model collusion. Users may collude by aggregating their models. For instance, \\(\\mathrm{Bob}^{(i)}\\) and \\(\\mathrm{Bob}^{(j)}\\) can average the weights of their models (like Model soups [91]) creating a new model to deceive identification. We found that the bit at position \\(\\ell\\) output by the extractor will be 0 (resp. 1) when the \\(\\ell\\)-th bits of \\(\\mathrm{Bob}^{(i)}\\) and \\(\\mathrm{Bob}^{(j)}\\) are both 0 (resp. 1), and that the extracted bit is random when their bits disagree. We" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.726, + 0.442, + 0.848 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.857, + 0.47, + 0.899 + ], + "angle": 0, + "content": "Figure 8. Robustness to model purification, i.e. fine-tuning the model to remove watermarks. \\( x_w \\) is the watermarked image, \\( x_r \\) is generated with the purified model at different steps of the process." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "show the distributions of the soft bits (before thresholding) output by the watermark extractor on images generated by the average model. The \\(\\ell\\)-th output is labeled by bits of \\(\\mathrm{Bob}^{(i)}\\) and \\(\\mathrm{Bob}^{(j)}\\) (\\(00\\) means both have \\(0\\) at position \\(\\ell\\)):" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.162, + 0.885, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.239, + 0.892, + 0.361 + ], + "angle": 0, + "content": "This so-called marking assumption plays a crucial role in traitor tracing literature [27, 55, 83]. Surprisingly, it holds even though our watermarking process is not explicitly designed for it. The study has room for improvement, such as creating user identifiers with more powerful traitor tracing codes [83] and using more powerful traitor accusation algorithms [27, 55]. Importantly, we found the precedent remarks also hold if the colluders operate at the image level." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.374, + 0.733, + 0.39 + ], + "angle": 0, + "content": "8. Conclusion & Discussion" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.4, + 0.892, + 0.506 + ], + "angle": 0, + "content": "By a quick fine-tuning of the decoder of Latent Diffusion Models, we can embed watermarks in all the images they generate. This does not alter the diffusion process, making it compatible with most of LDM-based generative models. These watermarks are robust, invisible to the human eye and can be employed to detect generated images and identify the user that generated it, with very high performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.506, + 0.892, + 0.597 + ], + "angle": 0, + "content": "The public release of image generative models has an important societal impact. With this work, we put to light the usefulness of using watermarking instead of relying on passive detection methods. We hope it will encourage researchers and practitioners to employ similar approaches before making their models publicly available." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.605, + 0.892, + 0.697 + ], + "angle": 0, + "content": "Reproducibility Statement. Although the diffusion-based generative model has been trained on an internal dataset of licensed images, we use the KL auto-encoder from LDM [68] with compression factor \\( f = 8 \\). This is the one used by open-source alternatives. Code is available at github.com/facebookresearch/stable_signature." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.705, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Environmental Impact. We do not expect any environmental impact specific from this work. The cost of the experiments and the method is high, though order of magnitudes less than other computer vision fields. We roughly estimated that the total GPU-days used for running all our experiments to 2000, or \\(\\approx\\) 50000 GPU-hours. This amounts to total emissions in the order of 10 tons of \\(\\mathrm{CO}_{2}\\mathrm{eq}\\). This is excluding the training of the generative model itself, since we did not perform that training. Estimations are conducted using the Machine Learning Impact calculator presented by Lacoste et al. [43]. We do not consider in this approximation: memory storage, CPU-hours, production cost of GPUs/ CPUs, etc." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22473" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Sahar Abdelnabi and Mario Fritz. Adversarial watermarking transformer: Towards tracing text provenance with data hiding. In 2021 IEEE Symposium on Security and Privacy (SP), pages 121-140. IEEE, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.172, + 0.47, + 0.226 + ], + "angle": 0, + "content": "[2] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.228, + 0.469, + 0.296 + ], + "angle": 0, + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.298, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[4] Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.354, + 0.469, + 0.395 + ], + "angle": 0, + "content": "[5] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.397, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[6] Jean Bégaint, Fabien Racapé, Simon Feltman, and Akshay Pushparaja. Compressai: a pytorch library and evaluation platform for end-to-end compression research. arXiv preprint arXiv:2011.03029, 2020. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.453, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.495, + 0.469, + 0.563 + ], + "angle": 0, + "content": "[8] Miles Brundage, Shahar Avin, Jack Clark, Helen Toner, Peter Eckersley, Ben Garfinkel, Allan Dafoe, Paul Scharre, Thomas Zeitzoff, Bobby Filar, et al. The malicious use of artificial intelligence: Forecasting, prevention, and mitigation. arXiv preprint arXiv:1802.07228, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.565, + 0.469, + 0.619 + ], + "angle": 0, + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV. IEEE, 2021. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.621, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[10] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In European conference on computer vision, pages 103-120. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.677, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[11] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7939-7948, 2020. 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.747, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[12] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. arXiv preprint arXiv:2211.00680, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.804, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[13] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 1, 2, 6, 13, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital watermarking and steganography. Morgan Kaufmann, 2007. 2, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Steffen Czolbe, Oswin Krause, Ingemar Cox, and Christian Igel. A loss function for generative neural networks based on watson's perceptual model. Advances in Neural Information Processing Systems, 33:2051-2061, 2020. 4, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.151, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.209, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[17] Emily Denton. Ethical considerations of generative ai. In AI for Content Creation Workshop CVPR. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.239, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.282, + 0.892, + 0.35 + ], + "angle": 0, + "content": "[19] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.354, + 0.892, + 0.42 + ], + "angle": 0, + "content": "[20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.425, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[21] Hany Farid. Lighting (in) consistency of paint by text. arXiv preprint arXiv:2207.13744, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.456, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[22] Hany Farid. Perspective (in) consistency of paint by text. arXiv preprint arXiv:2206.14617, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.485, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[23] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Supervised gan watermarking for intellectual property protection. In 2022 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.557, + 0.892, + 0.597 + ], + "angle": 0, + "content": "[24] Pierre Fernandez, Matthijs Douze, Hervé Jégou, and Teddy Furon. Active image indexing. arXiv preprint arXiv:2210.10620, 2022. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.601, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[25] Pierre Fernandez, Alexandre Sablayrolles, Teddy Furon, Hervé Jégou, and Matthijs Douze. Watermarking images in self-supervised latent spaces. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022. 2, 6, 7, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.673, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[26] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.744, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[27] Teddy Furon, Arnaud Guyader, and Frédéric Cérou. Decoding Fingerprinting Using the Markov Chain Monte Carlo Method. In WIFS - IEEE Workshop on Information Forensics and Security, Tenerife, Spain, Dec. 2012. IEEE. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.802, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[28] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Matthew Gault. An ai-generated artwork won first place at a state fair fine arts competition, and artists are pissed. Vice," + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22474" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.468, + 0.146 + ], + "angle": 0, + "content": "2022. https://www.vice.com/en/article/bvmvqm/an-ai-generated-artwork-won-first-place-at-a-state-fair-fine-arts-competition-and-artists-are-pissed (retrieved 2022-12-14). 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.15, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[30] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. *ICLR*, 2014. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.193, + 0.468, + 0.262 + ], + "angle": 0, + "content": "[31] Diego Gragnaniello, Davide Cozzolino, Francesco Marra, Giovanni Poggi, and Luisa Verdoliva. Are gan generated images easy to detect? a critical analysis of the state-of-the-art. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.264, + 0.468, + 0.331 + ], + "angle": 0, + "content": "[32] Luca Guarnera, Oliver Giudice, and Sebastiano Battiato. Deepfake detection by analyzing convolutional traces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 666-667, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.334, + 0.468, + 0.389 + ], + "angle": 0, + "content": "[33] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.391, + 0.468, + 0.459 + ], + "angle": 0, + "content": "[34] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.462, + 0.468, + 0.531 + ], + "angle": 0, + "content": "[35] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.533, + 0.468, + 0.573 + ], + "angle": 0, + "content": "[36] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.576, + 0.468, + 0.63 + ], + "angle": 0, + "content": "[37] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.633, + 0.468, + 0.687 + ], + "angle": 0, + "content": "[38] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.689, + 0.468, + 0.758 + ], + "angle": 0, + "content": "[39] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.76, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[40] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.817, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[41] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[42] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2022. 2, 6, 7, 14" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[43] Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, and Thomas Dandres. Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.15, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[44] Jae-Eun Lee, Young-Ho Seo, and Dong-Wook Kim. Convolutional neural network-based digital image watermarking adaptive to the resolution of image and watermark. Applied Sciences, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.207, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[45] Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. arXiv preprint arXiv:1811.00656, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.25, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[46] Dongdong Lin, Benedetta Tondi, Bin Li, and Mauro Barni. Cycleganwm: A cyclegan watermarking method for ownership verification. arXiv preprint arXiv:2211.13737, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.307, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4, 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.377, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[48] Wenqing Liu, Miaojing Shi, Teddy Furon, and Li Li. Defending adversarial examples via dnn bottleneck reinforcement. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1930-1938, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.434, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[49] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.463, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[50] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.547, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[51] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In CVPR. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.591, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[52] Jingwei Ma, Lucy Chai, Minyoung Huh, Tongzhou Wang, Ser-Nam Lim, Phillip Isola, and Antonio Torralba. Totems: Physical objects for verifying visual integrity. In European Conference on Computer Vision, pages 164-180. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.661, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[53] Sébastien Marcel and Yann Rodriguez. Torchvision the machine-vision package of torch. In International Conference on Multimedia. ACM, 2010. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[54] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[55] Peter Meerwald and Teddy Furon. Towards practical joint decoding of binary Tardos fingerprinting codes. IEEE Transactions on Information Forensics and Security, 7(4):1168-1180, Apr. 2012. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[56] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[57] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22475" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 6, 13, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.135, + 0.471, + 0.189 + ], + "angle": 0, + "content": "[58] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.19, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[59] Guangyu Nie, Changhoon Kim, Yezhou Yang, and Yi Ren. Attributing image generative models using latent fingerprints. arXiv preprint arXiv:2304.09752, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.233, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[60] Zoe Papakipos and Joanna Bitton. Augly: Data augmentations for robustness. arXiv preprint arXiv:2201.06494, 2022. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.274, + 0.471, + 0.356 + ], + "angle": 0, + "content": "[61] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems 32, 2019. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.358, + 0.471, + 0.397 + ], + "angle": 0, + "content": "[62] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.4, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[63] Tong Qiao, Yuyan Ma, Ning Zheng, Hanzhou Wu, Yanli Chen, Ming Xu, and Xiangyang Luo. A novel model watermarking for protecting generative adversarial network. Computers & Security, 127:103102, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.455, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[64] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.511, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[65] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.567, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[66] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.637, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[67] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.679, + 0.471, + 0.76 + ], + "angle": 0, + "content": "[68] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2, 4, 6, 7, 8, 14, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.761, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[69] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.832, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[70] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. Stable diffusion. Accessed Nov. 30 2022, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.873, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[71] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[72] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Spreading vectors for similarity search. ICML, 2019. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[73] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.233, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[74] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.316, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[75] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.385, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[76] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.441, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[77] Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. Defake: Detection and attribution of fake images generated by text-to-image diffusion models. arXiv preprint arXiv:2210.06998, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.497, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[78] Changhao Shi, Chester Holtz, and Gal Mishne. Online adversarial purification based on self-supervision. arXiv preprint arXiv:2101.09387, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[79] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.608, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[80] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.65, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[81] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. *ICLR*, 2021. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.706, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[82] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2149–2159, 2022. 13, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.803, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[83] Gábor Tardos. Optimal probabilistic fingerprint codes. Journal of the ACM (JACM), 55(2):1-24, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[84] Yusuke Uchida, Yuki Nagai, Shigeyuki Sakazawa, and Shin'ichi Satoh. Embedding watermarks into deep neural networks. In Proceedings of the 2017 ACM on international conference on multimedia retrieval, pages 269-277, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22476" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[85] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.471, + 0.188 + ], + "angle": 0, + "content": "[86] Vedran Vukotic, Vivien Chappelier, and Teddy Furon. Are deep neural networks good for blind image watermarking? In WIFS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.19, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[87] Steven Walton, Ali Hassani, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Stylenat: Giving each head a new perspective. arXiv preprint arXiv:2211.05770, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.247, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[88] Sheng-Yu Wang, Oliver Wang, Andrew Owens, Richard Zhang, and Alexei A Efros. Detecting photoshopped faces by scripting photoshop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10072-10081, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.472, + 0.384 + ], + "angle": 0, + "content": "[89] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.386, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[90] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.442, + 0.471, + 0.537 + ], + "angle": 0, + "content": "[91] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.539, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[92] Chen Henry Wu and Fernando De la Torre. Unifying diffusion models' latent space, with applications to cycled diffusion and guidance. arXiv preprint arXiv:2210.05559, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.595, + 0.471, + 0.649 + ], + "angle": 0, + "content": "[93] Hanzhou Wu, Gen Liu, Yuwei Yao, and Xinpeng Zhang. Watermarking neural networks with watermarked images. IEEE Transactions on Circuits and Systems for Video Technology, 31(7):2591-2601, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.651, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[94] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, pages 12062-12072. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.707, + 0.471, + 0.773 + ], + "angle": 0, + "content": "[95] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. In ICLR, 2020. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.776, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[96] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[97] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[98] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14448-14457, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.164, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[99] Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. Responsible disclosure of generative models using scalable fingerprinting. In International Conference on Learning Representations (ICLR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.893, + 0.274 + ], + "angle": 0, + "content": "[100] Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun, and In So Kweon. Udh: Universal deep hiding for steganography, watermarking, and light field messaging. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.893, + 0.345 + ], + "angle": 0, + "content": "[101] Chaoning Zhang, Adil Karjauv, Philipp Benz, and In So Kweon. Towards robust deep hiding under nondifferentiable distortions for practical blind watermarking. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5158-5166, 2021. 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.893, + 0.415 + ], + "angle": 0, + "content": "[102] Jie Zhang, Dongdong Chen, Jing Liao, Han Fang, Weiming Zhang, Wenbo Zhou, Hao Cui, and Nenghai Yu. Model watermarking for image processing networks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12805-12812, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.418, + 0.893, + 0.471 + ], + "angle": 0, + "content": "[103] Kevin Alex Zhang, Alfredo Cuesta-Infante, Lei Xu, and Kalyan Veeramachaneni. Steganogan: High capacity image steganography with gans. arXiv preprint arXiv:1901.03892, 2019. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.893, + 0.514 + ], + "angle": 0, + "content": "[104] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.517, + 0.893, + 0.571 + ], + "angle": 0, + "content": "[105] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR. IEEE, 2018. 4, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.574, + 0.893, + 0.628 + ], + "angle": 0, + "content": "[106] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In 2019 IEEE international workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.631, + 0.893, + 0.697 + ], + "angle": 0, + "content": "[107] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deep-fake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.893, + 0.741 + ], + "angle": 0, + "content": "[108] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In ECCV, 2018. 2, 3, 4, 6, 7, 13" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22477" + } + ] +] \ No newline at end of file diff --git a/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/c2da89cd-17c8-420c-b57b-7af434285ab7_origin.pdf b/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/c2da89cd-17c8-420c-b57b-7af434285ab7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dd21ffc00f1438217e3aac03710bfdd79daa0812 --- /dev/null +++ b/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/c2da89cd-17c8-420c-b57b-7af434285ab7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a2e0d981865b5b3026b54b3ab5d63ef3540f25b8ffec2740c2869ee353ce801 +size 2868108 diff --git a/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/full.md b/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c8d2263bd59f2954461de59245e85a3d9579a3fc --- /dev/null +++ b/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/full.md @@ -0,0 +1,403 @@ +# The Stable Signature: Rooting Watermarks in Latent Diffusion Models + +Pierre Fernandez $^{1,2}$ Guillaume Couairon $^{1,3}$ Hervé Jégou $^{1}$ Matthijs Douze $^{1}$ Teddy Furon $^{2*}$ + +1Meta AI 2Centre Inria de l'Université de Rennes 3Sorbonne University + +# Abstract + +Generative image modeling enables a wide range of applications but raises ethical concerns about responsible deployment. We introduce an active content tracing method combining image watermarking and Latent Diffusion Models. The goal is for all generated images to conceal an invisible watermark allowing for future detection and/or identification. The method quickly fine-tunes the latent decoder of the image generator, conditioned on a binary signature. A pre-trained watermark extractor recovers the hidden signature from any generated image and a statistical test then determines whether it comes from the generative model. We evaluate the invisibility and robustness of the watermarks on a variety of generation tasks, showing that the Stable Signature is robust to image modifications. For instance, it detects the origin of an image generated from a text prompt, then cropped to keep $10\%$ of the content, with $90 + \%$ accuracy at a false positive rate below $10^{-6}$ . + +# 1. Introduction + +Recent progress in generative modeling and natural language processing enable easy creation and manipulation of photo-realistic images, such as with DALL-E 2 [64] or Stable Diffusion [68]. They have given birth to many image edition tools like ControlNet [104], Instruct-Pix2Pix [7], and others [13, 28, 71], that are becoming mainstream creative tools for artists, designers, and the general public. + +While this is a great step forward for generative AI, it also undermines confidence in the authenticity or veracity of photo-realistic images. Indeed, methods for photorealistic image edition existed before, but generative AI significantly lowers the barriers to convincing synthetic image generation and edition (e.g. a generated picture recently won an art competition [29]). This raises new risks like deep fakes, impersonation or copyright usurpation [8, 17]. A tool to determine that images are AI-generated would make it easier to ensure their compliance with ethical standards and to remove them from certain platforms. + +![](images/2c7b5c2fde8d9e5922d8418a2133684ce4791ae3ac6f121e907c2d4e645869c8.jpg) +Figure 1. Overview. The latent decoder can be fine-tuned to preemptively embed a signature into all generated images. + +A baseline solution to identify generated images is forensics, i.e. methods to detect generated/manipulated images passively (the image is not modified for identification). An active baseline is to apply existing watermarking methods after the image generation. Watermarking invisibly embeds a secret message into the image, which can then be extracted and used to identify the image. This has several drawbacks. If the model leaks or is open-sourced, the post-generation watermarking can be removed trivially. The open source Stable Diffusion [70] is a case in point, since removing the watermark amounts to commenting out a single line in the source code. + +Our Stable Signature method merges watermarking into the generation process itself, without any architectural change. It adjusts the pre-trained generative model such that all the images it produces conceal a given watermark. There are several advantages to this approach [46, 99]. It does not require additional processing of the generated image, which makes the watermarking computationally lighter, straightforward, and secure. Model providers could deploy their models to different user groups with a unique watermark, and monitor that they are used in a responsible manner. + +They could give art platforms, news outlets and other sharing platforms the ability to detect when an image has been generated by their AI. + +We focus on Latent Diffusion Models (LDM) [68] that can perform a wide range of generative tasks. We show that simply fine-tuning a small part of the generative model – the decoder that generates images from the latent vectors – is enough to natively embed a watermark into generated images. Stable Signature does not require an architectural change and does not modify the diffusion process. Hence it is to our knowledge compatible with all LDM-based generative methods [7, 13, 62, 71, 104]. The fine-tuning stage is performed by back-propagating a combination of a perceptual image loss and a message decoding loss from a watermark extractor back to the LDM decoder. We pre-train the extractor with a simplified version of the deep watermarking method HiDDeN [108]. + +We create an evaluation benchmark close to real world situations where images may be edited. The tasks are: detection of AI generated images, tracing models from their generations. For instance, we detect $90\%$ of images generated with the generative model, even if they are cropped to $10\%$ of their original size, while flagging only one false positive every $10^{6}$ images. To ensure that the model's utility is not weakened, we show that the FID [34] score of the generation is not affected and that the generated images are perceptually indistinguishable from the ones produced by the original model. This is done over several tasks involving LDM (text-to-image, inpainting, edition, etc.). + +As a summary, (1) we efficiently merge watermarking into the generation process of LDMs, in a way that is compatible with most of the LDM-based generative methods; (2) we demonstrate how it can be used to detect and trace generated images, through a real-world evaluation benchmark; (3) we compare to post-hoc watermarking methods, showing that it is competitive while being more secure and efficient, and (4) evaluate robustness to intentional attacks. + +# 2. Related Work + +Image generation has long been dominated by GANs, still state-of-the-art on many datasets [37, 38, 39, 76, 87]. Transformers have also been successfully used for modeling image [66, 19] or video [79] distributions, providing higher diversity at the expense of increased inference time. Images are typically converted to token lists using vector-quantized architectures [20, 67, 96], relying on an image decoder. + +Diffusion models [18, 36, 58, 80] are a huge improvement in text-conditional image generation, current models synthesize high-resolution photo-realistic images for a wide variety of text prompts [3, 35, 65, 69, 74]. They can also perform conditional image generation – like inpainting or text-guided image editing – by fine-tuning the diffusion model with additional conditioning, e.g. masked input im + +age, segmentation map, etc. [50, 73]. Because of their iterative denoising algorithm, diffusion models can also be adapted for image editing in a zero-shot fashion by guiding the generative process [13, 33, 40, 56, 85, 92]. All these methods, when combined with Stable Diffusion, operate in the latent rather than image space, requiring a latent decoder to produce an RGB image. + +Detection of AI-generated/manipulated images is notably active in the context of deep-fake detection [32, 107]. Many works focus on the detection of GAN-generated images [10, 31, 89, 106]. One approach is to detect inconsistencies, via lights, perspective or physical objects [21, 22, 45, 52, 88]. These approaches are restricted to photorealistic images or faces, artworks not intended to be physically correct are not covered. + +Other approaches track traces left by the generators in the spatial [54, 97] or frequency [26, 106] domains. There are extensions to diffusion models in recent works [12, 77] that show encouraging results. However purely relying on forensics and passive detection is limiting, e.g. the best performing method to our knowledge [12] is able to detect $50\%$ of generated images for a false positive rate around 1/100: if a user-generated content platform were to receive 1 billion images every day, it would wrongly flag 10 million images to detect only half of the generated images. Besides, passive techniques cannot trace images from different versions of the same model, in contrast with watermarking. + +Image watermarking has long been studied in the context of tracing and intellectual property protection [14]. More recently, deep learning encoder/extractors like HiD-DeN [2, 44, 51, 100, 108] or iterative methods by Vukotic et al. [25, 42, 86] showed competitive results in terms of robustness to a wide range of transformations, namely geometric ones. + +In the specific case of generative models, some works aim to watermark the training set on which the generative model is learned [98]. It is highly inefficient since every new message to embed requires a new training pipeline. Merging the watermarking and the generative process is a recent idea [23, 46, 59, 63, 93, 99, 102], that is closer to the model watermarking literature [84]. They suffer from two strong limitations. First, these methods only apply to GANs, while LDM are progressively replacing them for most applications. Second, watermarking is incorporated in the training process of the GAN from the start. This strategy is unsustainable because the generative model training is more and more costly1. Our work shows that a quick finetuning of the latent decoder part of the generative model is enough to achieve a good watermarking performance, provided that the watermark extractor is well chosen. + +# 3. Problem Statement & Background + +Figure 1 shows a model provider Alice who deploys a latent diffusion model to users Bobs. Stable Signature embeds a binary signature into the generated images. This section derives how Alice can use this signature for two scenarios: + +- Detection: "Is it generated by my model?" Alice detects if an image was generated by her model. Generated images should be flagged as reliably as possible, while controlling the probability of flagging a natural image. +- Identification: "Who generated this image?". Alice monitors who created each image, while avoiding to mistakenly identifying a Bob. + +# 3.1. Image watermarking for detection + +Alice embeds a $k$ -bit binary signature into the generated images. The watermark extractor then decodes messages from the images it receives and detects when the message is close to Alice's signature. An example application is to block AI-generated images on a content sharing platform. + +Statistical test Let $m \in \{0,1\}^k$ be Alice's signature. We extract the message $m'$ from an image $x$ and compare it to $m$ . As done in previous works [46, 98], the detection test relies on the number of matching bits $M(m,m')$ : if + +$$ +M \left(m, m ^ {\prime}\right) \geq \tau \text {w h e r e} \tau \in \{0, \dots , k \}, \tag {1} +$$ + +then the image is flagged. This provides a level of robustness to imperfections of the watermarking. + +Formally, we test the statistical hypothesis $H_{1}$ : " $x$ was generated by Alice's model" against the null hypothesis $H_{0}$ : " $x$ was not generated by Alice's model". Under $H_{0}$ (i.e. for vanilla images), we assume that bits $m_{1}', \ldots, m_{k}'$ are (i.i.d.) Bernoulli random variables with parameter 0.5. Then $M(m, m')$ follows a binomial distribution with parameters $(k, 0.5)$ . We verify this assumption experimentally in App. B.5. The False Positive Rate (FPR) is the probability that $M(m, m')$ takes a value bigger than the threshold $\tau$ . It is obtained from the CDF of the binomial distribution, and a closed-form can be written with the regularized incomplete beta function $I_{x}(a; b)$ : + +$$ +\operatorname {F P R} (\tau) = \mathbb {P} (M > \tau | H _ {0}) = I _ {1 / 2} (\tau + 1, k - \tau). \tag {2} +$$ + +# 3.2. Image watermarking for identification + +Alice now embeds a signature $m^{(i)}$ drawn randomly from $\{0,1\}^k$ into the model distributed to $\mathrm{Bob}^{(i)}$ (for $i = 1 \cdots N$ , with $N$ the number of Bobs). Alice can trace any misuse of her model: generated images violating her policy (gore content, deepfakes) are linked back to the specific Bob by comparing the extracted message to Bobs' signatures. + +Statistical test We compare the message $m'$ from the watermark extractor to $(m^{(1)}, \ldots, m^{(N)})$ . There are now $N$ detection hypotheses to test. If the $N$ hypotheses are rejected, we conclude that the image was not generated by any of the models. Otherwise, we attribute the image to $\operatorname{argmax}_{i=1..N} M(m', m^{(i)})$ . With regards to the detection task, false positives are more likely since there are $N$ tests. The global FPR at a given threshold $\tau$ is: + +$$ +\operatorname {F P R} (\tau , N) = 1 - (1 - \operatorname {F P R} (\tau)) ^ {N} \approx N. \operatorname {F P R} (\tau). \tag {3} +$$ + +Equation (3) (resp. (2)), is used reversely: we find threshold $\tau$ to achieve a required FPR for identification (resp. detection). Note that these formulae hold only under the assumption of i.i.d. Bernoulli bits extracted from vanilla images. This condition is enforced in the next section. + +# 4. Method + +Stable Signature modifies the generative network so that the generated images have a given signature through a fixed watermark extractor. It is trained in two phases. First, we create the watermark extractor network $\mathcal{W}$ . We then fine-tune the LDM decoder $\mathcal{D}$ , such that all generated images yield a given signature through $\mathcal{W}$ . + +# 4.1. Pre-training the watermark extractor + +We use HiDDeN [108], a classical method in the deep watermarking literature. It jointly optimizes the parameters of watermark encoder $\mathcal{W}_E$ and extractor network $\mathcal{W}$ to embed $k$ -bit messages into images, robustly to transformations that are applied during training. We discard $\mathcal{W}_E$ after training, since only $\mathcal{W}$ serves our purpose. + +Formally, $\mathcal{W}_E$ takes as inputs a cover image $x_o \in \mathbb{R}^{W \times H \times 3}$ and a $k$ -bit message $m \in \{0,1\}^k$ . Similar to + +![](images/a8e372dd0b131cee24e7f98cf1a102573b2555b1a632bc6dc9f3c77ca79c9a7b.jpg) +(a) Pre-train watermark encoder/extractor + +![](images/c9a2e0e6c5c67da3f73d1f1ed5f06268f224de6ad0bd88b0df88ac520db22afd.jpg) +(b) Fine-tune LDM decoder +(c) Generate +Figure 2. Steps of the method. (a) We pre-train a watermark encoder $\mathcal{W}_E$ and extractor $\mathcal{W}$ , to extract binary messages. (b) We fine-tune the decoder $\mathcal{D}$ of the LDM's auto-encoder with a fixed signature $m$ such that all the generated images (c) lead to $m$ through $\mathcal{W}$ . + +![](images/541d235c8465ce2d5220684d78dcd49854ea3ccb8613edfda496a461ad40b6e2.jpg) + +![](images/449466aa4f7c59bf5152aa317699ad2e0c6a8690bd875af411b24312fba4393e.jpg) +Text - Super-resolution Inpainting - Editing - ... + +ReDMark [2], $\mathcal{W}_E$ outputs a residual image $\delta$ of the same size as $x_{o}$ , that is multiplied by a factor $\alpha$ to produce watermarked image $x_w = x_o + \alpha \delta$ . At each optimization step an image transformation $T$ is sampled from a set $\mathcal{T}$ that includes common image processing operations such as cropping and JPEG compression. A "soft" message is extracted from the transformed image: $m' = \mathcal{W}(T(x_w))$ (at inference time, the decoded message is given by the signs of the components of $m'$ ). The message loss is the Binary Cross Entropy (BCE) between $m$ and the sigmoid $\sigma(m')$ : + +$$ +\mathcal {L} _ {m} = - \sum_ {i = 1} ^ {k} m _ {i} \cdot \log \sigma (m _ {i} ^ {\prime}) + (1 - m _ {i}) \cdot \log (1 - \sigma (m _ {i} ^ {\prime})). +$$ + +The network architectures are kept simple to ease the LDM fine-tuning in the second phase. They are the same as HiDDeN [108] (see App. A.1) with two changes. + +First, since $\mathcal{W}_E$ is discarded, its perceptual quality is not as important, so the perceptual loss or the adversarial network are not needed. Instead, the distortion is constrained by a tanh function on output of $\mathcal{W}_E$ and by the scaling factor $\alpha$ . This improves the bit accuracy of the recovered message and makes it possible to increase its size $k$ . + +Second, we observed that $\mathcal{W}$ 's output bits for vanilla images are correlated and highly biased, which violates the assumptions of Sec. 3.1. Therefore, to get closer to i.i.d. dimensions we remove the bias and decorrelate the outputs of $\mathcal{W}$ by applying a PCA whitening transformation (more details in App. A.1). + +# 4.2. Fine-tuning the generative model + +In LDM, the diffusion happens in the latent space of an auto-encoder. The latent vector $z$ obtained at the end of the diffusion is input to decoder $\mathcal{D}$ to produce an image. Here we fine-tune $\mathcal{D}$ such that the image contains a given message $m$ that can be extracted by $\mathcal{W}$ . Stable Signature is compatible with many generative tasks, since modifying only $\mathcal{D}$ does not affect the diffusion process. + +First, we fix the signature $m = (m_1,\ldots ,m_k)\in \{0,1\} ^k$ The fine-tuning of $\mathcal{D}$ into $\mathcal{D}_m$ is inspired by the original training of the auto-encoder in LDM [68]. + +Training image $x \in \mathbb{R}^{H \times W \times 3}$ is fed to the LDM encoder $\mathcal{E}$ that outputs activation map $z = \mathcal{E}(x) \in \mathbb{R}^{h \times w \times c}$ , downsampled by a power-of-two factor $f = H / h = W / w$ . The decoder reconstructs an image $x' = \mathcal{D}_m(z)$ and the extractor recovers $m' = \mathcal{W}(x')$ . The message loss is the BCE between $m'$ and the original $m$ : $\mathcal{L}_m = \mathrm{BCE}(\sigma(m'), m)$ . + +In addition, the original decoder $\mathcal{D}$ reconstructs the image without watermark: $x_{o}^{\prime} = \mathcal{D}(z)$ . The image perceptual loss $\mathcal{L}_{\mathrm{i}}$ between $x^{\prime}$ and $x_{o}^{\prime}$ , controls the distortion. We use the Watson-VGG perceptual loss introduced by Czolbe et + +al. [15], an improved version of LPIPS [105]. It is essential that the decoder learns luminance and contrast masking to add less visible watermarks. The weights of $\mathcal{D}_m$ are optimized to minimize + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {m}} + \lambda_ {\mathrm {i}} \mathcal {L} _ {\mathrm {i}}. \tag {4} +$$ + +This is done over 100 iterations with the AdamW optimizer [49] and batch of size 4, i.e. the fine-tuning sees fewer than 500 images and takes one minute on a single GPU. The learning rate follows a cosine annealing schedule with 20 iterations of linear warmup to $10^{-4}$ and decays to $10^{-6}$ . The factor $\lambda_{\mathrm{i}}$ in (4) is set to 0.2 by default. + +# 5. Text-to-Image Watermarking Performance + +This section discusses our method's detection and identification capability for images generated by a Stable-Diffusion-like model $[68]^3$ . We apply generative models watermarked with 48-bit signatures on prompts of the MS-COCO [47] validation set. We evaluate detection and identification on the outputs, as illustrated in Figure 1. + +We evaluate their robustness to different transformations applied to generated images: strong cropping (10% of the image remaining), brightness shift (strength factor 2.0), as well as a combination of crop 50%, brightness shift 1.5 and JPEG 80. This covers typical geometric and photometric edits (see Fig. 5 for visual examples). + +The detection rates are partly obtained from experiments and partly by extrapolating small-scale measurements. + +# 5.1. Detection results + +For detection, we fine-tune the decoder of the LDM with a random key $m$ , generate 1000 images and use the test of Eq. (1). We report the tradeoff between True Positive Rate (TPR), i.e. the probability of flagging a generated image and the FPR, while varying $\tau \in \{0,..,48\}$ . For instance, for $\tau = 0$ , we flag all images so $\mathrm{FPR} = 1$ , and $\mathrm{TPR} = 1$ . The TPR is measured directly, while the FPR is inferred from Eq. (2), because it would otherwise be too small to be measured on reasonably sized problems (this approximation is validated experimentally in App. B.6). The experiment is run on 10 random signatures and we report averaged results. + +Figure 3 shows the tradeoff under image transformations. For example, when the generated images are not modified, Stable Signature detects $99\%$ of them, while only 1 vanilla image out of $10^{9}$ is flagged. At the same FPR $= 10^{-9}$ , Stable Signature detects $84\%$ of generated images for a crop that keeps $10\%$ of the image, and $65\%$ for a transformation that combines a crop, a color shift, and a JPEG compression. For comparison, we report results of a state-of-the-art passive method [12], applied on resized and + +compressed images. As expected, we observe that these baseline results have orders of magnitudes larger FPR than Stable Signature, which actively marks the content. + +# 5.2. Identification results + +Each Bob has its own copy of the generative model. Given an image, the goal is to find if any of the $N$ Bobs created it (detection) and if so, which one (identification). There are 3 types of error: false positive: flag a vanilla image; false negative: miss a generated image; false accusation: flag a generated image but identify the wrong Bob. + +For evaluation, we fine-tune $N' = 1000$ models with random signatures. Each model generates 100 images. For each of these 100k watermarked images, we extract the Stable Signature message, compute the matching score with all $N$ signatures and select the Bob with the highest score. The image is predicted to be generated by that Bob if this score is above threshold $\tau$ . We determined $\tau$ such that $\mathrm{FPR} = 10^{-6}$ , see Eq. (3). For example, for $N = 1$ , $\tau = 41$ and for $N = 1000$ , $\tau = 44$ . Accuracy is extrapolated beyond the $N'$ Bobs by adding additional signatures and having $N > N'$ (e.g. Bobs that have not generated any images). + +Figure 4 reports the per-transformation identification accuracy. For example, we identify a Bob among $N = 10^{5}$ with $98\%$ accuracy when the image is not modified. Note that for the combined edit, this becomes $40\%$ . This may still be dissuasive: if a Bob generates 3 images, he will be identified $80\%$ of the time. We observe that at this scale, the false accusation rate is zero, i.e. we never identify the wrong Bob. This is because $\tau$ is set high to avoid FPs, which also makes false accusations unlikely. We observe that the identification accuracy decreases when $N$ increases, because the threshold $\tau$ required to avoid false positives is higher when $N$ increases, as pointed out by the approximation in (3). In a nutshell, by distributing more models, Alice trades some accuracy of detection against the ability to identify Bobs. + +# 6. Experimental Results + +The previous section showed how to leverage watermarks for detection and identification of images generated from text prompts. We now present more general results on robustness and image quality for different generative tasks. We also compare Stable Signature to post-generation watermarking algorithms. + +# 6.1. Tasks & evaluation metrics + +Since our method only involves the LDM decoder, it is compatible with many generative tasks. We evaluate text-to-image generation and image edition on the validation set of MS-COCO [47], super-resolution and inpainting on the validation set of ImageNet [16] (all evaluation details are available in App. A.3). + +We evaluate the image distortion with the Peak Signal-to-Noise Ratio (PSNR), which is defined as $\mathrm{PSNR}(x, x') = -10 \cdot \log_{10}(\mathrm{MSE}(x, x'))$ , for $x, x' \in [0, 1]^{c \times h \times w}$ , as well as Structural Similarity score (SSIM) [90]. They compare images generated with and without watermark. On the other hand, we evaluate the diversity and quality of the generated images with the Fréchet Inception Distance (FID) [34]. The bit accuracy – the percentage of bits correctly decoded – evaluates the watermarks' robustness. + +# 6.2. Image generation quality + +Figure 6 shows qualitative examples of how the image generation is altered by the latent decoder's fine-tuning. The difference is very hard to perceive even for a trained eye. This is surprising for such a low PSNR, especially since the watermark embedding is not constrained by any Human Visual System like in professional watermarking techniques. Most interestingly, the LDM decoder has indeed learned to add the watermark signal only over textured areas where the human eyes are not sensitive, while the uniform backgrounds are kept intact (see the pixel-wise difference). + +![](images/830dc5ae205f38ce29636fc7375bf256df868f68501542c82e7659d8ce7b3dde.jpg) +Figure 3. Detection results. TPR/FPR curve of the detection under different transformations. Forensics† indicates passive detection (without watermark) [12]. + +![](images/1d8f6ca966033bc65ea96bef8580b6508dc40c9d42b1f302623a6907be82a3bf.jpg) +Figure 4. Identification results. Proportion of well-identified Bobs. Detection with $\mathrm{FPR} = 10^{-6}$ is run beforehand, and we consider it an error if the image is not flagged. + +![](images/69f4398b92a022160850a5cc639b8e5692c3e672d92cc1201911c36ec04f416f.jpg) +Crop $10\%$ + +![](images/ef05b53d6b96c8e4e8ba8cf22fd9db4c9d054f624e914ca94c32d59c690d2017.jpg) +Combined + +![](images/96ddc50499967b5745d9f7384444a6efa2616b3c3d8db35831e9646b0807e0dc.jpg) +Figure 5. Transformations evaluated in Sec. 5 & 6. 'Combined' is made of crop $50\%$ , brightness adjustment 1.5 and JPEG 80 compression. + +
PSNR / SSIM ↑FID ↓Bit accuracy ↑ on:
NoneCropBrigh.Comb.
TasksText-to-ImageLDM [68]30.0 / 0.8919.6 (−0.3)0.990.950.970.92
Image EditionDiffEdit [13]31.2 / 0.9215.0 (−0.3)0.990.950.980.94
Inpainting - Full - Mask onlyGlide [57]31.1 / 0.9116.8 (+0.6)0.990.970.980.93
37.8 / 0.989.0 (+0.1)0.890.760.840.78
Super-ResolutionLDM [68]34.0 / 0.9411.6 (+0.0)0.980.930.960.92
WM MethodsPost generation
Dct-Dwt [14]0.14 (s/img)39.5 / 0.9719.5 (−0.4)0.860.520.510.51
SSL Watermark [25]0.45 (s/img)31.1 / 0.8620.6 (+0.7)1.000.730.930.66
FNNS [42]0.28 (s/img)32.1 / 0.9019.0 (−0.9)0.930.930.910.93
HiDDeN [108]0.11 (s/img)32.0 / 0.8819.7 (−0.2)0.990.970.990.98
Merged in generation
Stable Signature0.00 (s/img)30.0 / 0.8919.6 (−0.3)0.990.950.970.92
+ +Table 1. Generation quality and comparison to post-hoc watermarking on $512\times 512$ images and 48-bit signatures. PSNR and SSIM are computed between generations of the original and watermarked generators. For FID, we show in (color) the difference with regards to original. Post-hoc watermarking is evaluated on text-generated images. (App. B.2 gives results on more transformations, and App. A gives more details on the evaluations.) Overall, Stable Signature has minimal impact on generation quality. It has comparable robustness to post-hoc methods while being rooted in the generation itself. + +Table 1 presents a quantitative evaluation of image generation on the different tasks. We report the FID, and the average PSNR and SSIM that are computed between the images generated by the fine-tuned LDM and the original one. The results show that no matter the task, the watermarking has very small impact on the FID of the generation. + +The average PSNR is around 30 dB and SSIM around 0.9 between images generated by the original and a watermarked model. This is a bit low from a watermarking perspective, because we do not explicitly optimize for them. However, in a real world scenario, one would only have the watermarked version of the image. Therefore it is not as important to be extremely close to the original image, we just need to generate artifacts-free images. Without access to the image generated by the original LDM, it is very hard to tell whether a watermark is present or not. + +![](images/0891831bec1705ba992a3d421a68eaef704c641ffe4cdff967a39cc66271c124.jpg) +Figure 6. Images generated with Stable Diffusion. The PSNR is $35.4\mathrm{dB}$ in the first row and $28.6\mathrm{dB}$ in the second. Images generated with Stable Signature look natural because modified areas are located where the eye is not sensitive. More examples in App. C. + +Table 2. Watermark robustness on image transformations applied before decoding, details of which are available in App. A.3. We report the bit accuracy, averaged over $10 \times 1k$ images generated from COCO prompts with 10 different keys. + +
AttackBit acc.Comb.0.92Sharpness 2.00.99
None0.99Bright. 2.00.97Med. Filter k=70.94
Crop 0.10.95Cont. 2.00.98Resize 0.70.91
JPEG 500.88Sat. 2.00.99Text overlay0.99
+ +# 6.3. Watermark robustness + +We evaluate the robustness of the watermark to different image transformations applied before extraction. For each task, we generate 1k images with 10 models fine-tuned for different messages, and report the average bit accuracy in Table 1. Additionally, Table 2 reports results on more image transformations for images generated from COCO prompts. The main transformations are presented in Fig. 5 (more evaluation details are available in App. A.3). + +We see that the watermark is indeed robust for several tasks and across transformations. The bit accuracy is always above 0.9, except for inpainting, when replacing only the masked region of the image (between $1 - 50\%$ of the image, with an average of $27\%$ across masks). Besides, the bit accuracy is not perfect even without edition, mainly because there are images that are hard to watermark (e.g. the ones that are very uniform, like the background in Fig. 6) and for which the accuracy is lower. + +Note that the robustness comes even without any transformation during the LDM fine-tuning phase: it is due to the watermark extractor. If the watermark embedding pipeline is learned to be robust against an augmentation, then the LDM learns how to produce watermarks that are robust against it during fine-tuning. + +# 6.4. Comparison to post-hoc watermarking + +An alternative way to watermark generated images is to process them after the generation (post-hoc). This may + +Table 3. Quality-robustness trade-off during fine-tuning. + +
λi for fine-tuning0.80.40.20.10.050.025
PSNR ↑31.430.629.728.526.824.6
Bit acc. ↑ on ‘comb.’0.850.880.900.920.940.95
+ +Table 4. Role of the attack simulation layer at pre-training. + +
Seen at W trainingBit accuracy ↑ at test time:
Crop 0.1Rot. 90JPEG 50Bright. 2.0Res. 0.7
X1.000.560.500.990.48
✓1.000.990.900.990.91
+ +be simpler, but less secure and efficient than Stable Signature. We compare our method to a frequency based method, DCT-DWT [14], iterative approaches (SSL Watermark [25] and FNNS [42]), and an encoder/decoder one like HiD-DeN [108]. We choose DCT-DWT since it is employed by the original open source release of Stable Diffusion [70], and the other methods because of their performance and their ability to handle arbitrary image sizes and number of bits. We use our implementations (see details in App. A.4). + +Table 1 compares the generation quality and the robustness over $5\mathrm{k}$ generated images. Overall, Stable Signature achieves comparable results in terms of robustness. HiD-DeN's performance is a bit higher but its output bits are not i.i.d. meaning that it cannot be used with the same guarantees as the other methods. We also observe that post-hoc generation gives worse qualitative results, images tend to present artifacts (see Fig. 13 in the supplement). One explanation is that Stable Signature is merged into the high-quality generation process with the LDM auto-encoder model, which is able to modify images in a more subtle way. + +# 6.5. Can we trade image quality for robustness? + +We can choose to maximize the image quality or the robustness of the watermark thanks to the weight $\lambda_{\mathrm{i}}$ of the perceptual loss in (4). We report the average PSNR of 1k generated images, as well as the bit accuracy obtained on the extracted message for the 'Combined' editing applied before detection (qualitative results are in App. B.1). A higher $\lambda_{\mathrm{i}}$ leads to an image closer to the original one, but to lower bit accuracies on the extracted message, see Table 3. + +# 6.6. Attack simulation layer + +Watermark robustness against image transformations depends solely on the watermark extractor. Here, we pretrain them with or without specific transformations in the simulation layer, on a shorter schedule of 50 epochs, with $128 \times 128$ images and 16-bits messages. From there, we plug them in the LDM fine-tuning stage and we generate 1k images from text prompts. We report the bit accuracy of the extracted watermarks in Table 4. The extractor is naturally robust to some transformations, such as crops or brightness, without being trained with them, while others, like rotations + +or JPEG, require simulation during training for the watermark to be recovered at test time. Empirically we observed that adding a transformation improves results for the latter, but makes training more challenging. + +# 7. Attacks on Stable Signature's Watermarks + +We examine the watermark's resistance to intentional tampering, as opposed to distortions that happen without bad intentions like crops or compression (discussed in Sec. 5). We consider two threat models: one is typical for many image watermarking methods [14] and operates at the image level, and another targets the generative model level. For image-level attacks, we evaluate on $5\mathrm{k}$ images generated from COCO prompts. Full details on the following experiments can be found in Appendix A.5. + +# 7.1. Image-level attacks + +Watermark removal. Bob alters the image to remove the watermark with deep learning techniques, like methods used for adversarial purification [78, 94] or neural auto-encoders [1, 48]. Note that this kind of attacks has not been explored in the image watermarking literature to our knowledge. Figure 7 evaluates the robustness of the watermark against neural auto-encoders [4, 11, 20, 68] at different compression rates. To reduce the bit accuracy closer to random (50%), the image distortion needs to be strong (PSNR<26). However, assuming the attack is informed on the generative model, i.e. the auto-encoder is the same as the one used to generate the images, the attack becomes much more effective. It erases the watermark while achieving high quality (PSNR>29). This is because the image is modified precisely in the bandwidth where the watermark + +![](images/85e7692f5251de14abbe00d51d8887092e0d0405f0b822b74e9bab6f8d30f8df.jpg) +Figure 7. Removal attacks. $x_{o}$ is the image produced by the original generator, $x_{r}$ is the version produced by the watermarked generator and then attacked. Bit accuracy is on the watermark extracted from $x_{r}$ . Neural auto-encoders [4, 11, 20] follow the same trend, except for the one used by LDM ('KL-f8' for our LDM). When access to the watermark extractor is granted, adversarial attacks also remove the watermark at lower PSNR budget. + +is embedded. Note that this assumption is strong, because Alice does not need to distribute the original generator. + +Watermark removal & embedding (white-box). To go further, we assume that the attack is informed on the watermark extractor - e.g. because it has leaked. Bob can use an adversarial attack to remove the watermark by optimizing the image under a PSNR constraint. The objective is to minimize the $\ell_2$ distance between a random binary message sampled beforehand and the extractor's output, effectively replacing the original signature with a random one. It makes it possible to erase the watermark with a lower distortion budget, as seen in Fig. 7. + +Instead of removing the watermark, an attacker could embed a signature into vanilla images (unauthorized embedding [14]) to impersonate another Bob of whom they have a generated image. It highlights the importance of keeping the watermark extractor private. + +# 7.2. Network-level attacks + +Model purification. Bob gets Alice's generative model and uses a fine-tuning process akin to Sec. 4.2 to eliminate the watermark embedding – that we coin model purification. This involves removing the message loss $\mathcal{L}_{\mathrm{m}}$ , and shifting the focus to the perceptual loss $\mathcal{L}_{\mathrm{i}}$ between the original image and the one reconstructed by the LDM auto-encoder. + +Figure 8 shows the results of this attack for the MSE loss. The PSNR between the watermarked and purified images is plotted at various stages of fine-tuning. Empirically, it is difficult to significantly reduce the bit accuracy without compromising the image quality: artifacts start to appear during the purification. + +Model collusion. Users may collude by aggregating their models. For instance, $\mathrm{Bob}^{(i)}$ and $\mathrm{Bob}^{(j)}$ can average the weights of their models (like Model soups [91]) creating a new model to deceive identification. We found that the bit at position $\ell$ output by the extractor will be 0 (resp. 1) when the $\ell$ -th bits of $\mathrm{Bob}^{(i)}$ and $\mathrm{Bob}^{(j)}$ are both 0 (resp. 1), and that the extracted bit is random when their bits disagree. We + +![](images/fab0b6e04f8cc204629f9b8d0d333db1a6699340d819ef53d36ba1c99b0578a3.jpg) +Figure 8. Robustness to model purification, i.e. fine-tuning the model to remove watermarks. $x_w$ is the watermarked image, $x_r$ is generated with the purified model at different steps of the process. + +show the distributions of the soft bits (before thresholding) output by the watermark extractor on images generated by the average model. The $\ell$ -th output is labeled by bits of $\mathrm{Bob}^{(i)}$ and $\mathrm{Bob}^{(j)}$ ( $00$ means both have $0$ at position $\ell$ ): + +![](images/34f4af48989a935f75faa92f0873327c1aff8b0f6e920593dfb04870fb017f62.jpg) + +This so-called marking assumption plays a crucial role in traitor tracing literature [27, 55, 83]. Surprisingly, it holds even though our watermarking process is not explicitly designed for it. The study has room for improvement, such as creating user identifiers with more powerful traitor tracing codes [83] and using more powerful traitor accusation algorithms [27, 55]. Importantly, we found the precedent remarks also hold if the colluders operate at the image level. + +# 8. Conclusion & Discussion + +By a quick fine-tuning of the decoder of Latent Diffusion Models, we can embed watermarks in all the images they generate. This does not alter the diffusion process, making it compatible with most of LDM-based generative models. These watermarks are robust, invisible to the human eye and can be employed to detect generated images and identify the user that generated it, with very high performance. + +The public release of image generative models has an important societal impact. With this work, we put to light the usefulness of using watermarking instead of relying on passive detection methods. We hope it will encourage researchers and practitioners to employ similar approaches before making their models publicly available. + +Reproducibility Statement. Although the diffusion-based generative model has been trained on an internal dataset of licensed images, we use the KL auto-encoder from LDM [68] with compression factor $f = 8$ . This is the one used by open-source alternatives. Code is available at github.com/facebookresearch/stable_signature. + +Environmental Impact. We do not expect any environmental impact specific from this work. The cost of the experiments and the method is high, though order of magnitudes less than other computer vision fields. We roughly estimated that the total GPU-days used for running all our experiments to 2000, or $\approx$ 50000 GPU-hours. This amounts to total emissions in the order of 10 tons of $\mathrm{CO}_{2}\mathrm{eq}$ . This is excluding the training of the generative model itself, since we did not perform that training. Estimations are conducted using the Machine Learning Impact calculator presented by Lacoste et al. [43]. We do not consider in this approximation: memory storage, CPU-hours, production cost of GPUs/ CPUs, etc. + +# References + +[1] Sahar Abdelnabi and Mario Fritz. Adversarial watermarking transformer: Towards tracing text provenance with data hiding. In 2021 IEEE Symposium on Security and Privacy (SP), pages 121-140. IEEE, 2021. 7 +[2] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 2, 4 +[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2 +[4] Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 7, 14 +[5] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 17 +[6] Jean Bégaint, Fabien Racapé, Simon Feltman, and Akshay Pushparaja. Compressai: a pytorch library and evaluation platform for end-to-end compression research. arXiv preprint arXiv:2011.03029, 2020. 14 +[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 1, 2 +[8] Miles Brundage, Shahar Avin, Jack Clark, Helen Toner, Peter Eckersley, Ben Garfinkel, Allan Dafoe, Paul Scharre, Thomas Zeitzoff, Bobby Filar, et al. The malicious use of artificial intelligence: Forecasting, prevention, and mitigation. arXiv preprint arXiv:1802.07228, 2018. 1 +[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV. IEEE, 2021. 14 +[10] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In European conference on computer vision, pages 103-120. Springer, 2020. 2 +[11] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7939-7948, 2020. 7, 14 +[12] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. arXiv preprint arXiv:2211.00680, 2022. 2, 4, 5 +[13] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 1, 2, 6, 13, 16 +[14] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital watermarking and steganography. Morgan Kaufmann, 2007. 2, 6, 7, 8 + +[15] Steffen Czolbe, Oswin Krause, Ingemar Cox, and Christian Igel. A loss function for generative neural networks based on watson's perceptual model. Advances in Neural Information Processing Systems, 33:2051-2061, 2020. 4, 15 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 14 +[17] Emily Denton. Ethical considerations of generative ai. In AI for Content Creation Workshop CVPR. IEEE, 2021. 1 +[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2 +[19] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 2 +[20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 7, 14 +[21] Hany Farid. Lighting (in) consistency of paint by text. arXiv preprint arXiv:2207.13744, 2022. 2 +[22] Hany Farid. Perspective (in) consistency of paint by text. arXiv preprint arXiv:2206.14617, 2022. 2 +[23] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Supervised gan watermarking for intellectual property protection. In 2022 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6. IEEE, 2022. 2 +[24] Pierre Fernandez, Matthijs Douze, Hervé Jégou, and Teddy Furon. Active image indexing. arXiv preprint arXiv:2210.10620, 2022. 14 +[25] Pierre Fernandez, Alexandre Sablayrolles, Teddy Furon, Hervé Jégou, and Matthijs Douze. Watermarking images in self-supervised latent spaces. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022. 2, 6, 7, 14, 15 +[26] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 2 +[27] Teddy Furon, Arnaud Guyader, and Frédéric Cérou. Decoding Fingerprinting Using the Markov Chain Monte Carlo Method. In WIFS - IEEE Workshop on Information Forensics and Security, Tenerife, Spain, Dec. 2012. IEEE. 8 +[28] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1 +[29] Matthew Gault. An ai-generated artwork won first place at a state fair fine arts competition, and artists are pissed. Vice, + +2022. https://www.vice.com/en/article/bvmvqm/an-ai-generated-artwork-won-first-place-at-a-state-fair-fine-arts-competition-and-artists-are-pissed (retrieved 2022-12-14). 1 +[30] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. *ICLR*, 2014. 14 +[31] Diego Gragnaniello, Davide Cozzolino, Francesco Marra, Giovanni Poggi, and Luisa Verdoliva. Are gan generated images easy to detect? a critical analysis of the state-of-the-art. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2 +[32] Luca Guarnera, Oliver Giudice, and Sebastiano Battiato. Deepfake detection by analyzing convolutional traces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 666-667, 2020. 2 +[33] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[34] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2, 5 +[35] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[36] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2 +[37] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 2 +[38] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2 +[39] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2 +[40] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2 +[41] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 15 +[42] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2022. 2, 6, 7, 14 + +[43] Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, and Thomas Dandres. Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700, 2019. 8 +[44] Jae-Eun Lee, Young-Ho Seo, and Dong-Wook Kim. Convolutional neural network-based digital image watermarking adaptive to the resolution of image and watermark. Applied Sciences, 2020. 2 +[45] Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. arXiv preprint arXiv:1811.00656, 2018. 2 +[46] Dongdong Lin, Benedetta Tondi, Bin Li, and Mauro Barni. Cycleganwm: A cyclegan watermarking method for ownership verification. arXiv preprint arXiv:2211.13737, 2022. 1, 2, 3 +[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4, 5, 13 +[48] Wenqing Liu, Miaojing Shi, Teddy Furon, and Li Li. Defending adversarial examples via dnn bottleneck reinforcement. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1930-1938, 2020. 7 +[49] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2018. 4 +[50] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2 +[51] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In CVPR. IEEE, 2020. 2 +[52] Jingwei Ma, Lucy Chai, Minyoung Huh, Tongzhou Wang, Ser-Nam Lim, Phillip Isola, and Antonio Torralba. Totems: Physical objects for verifying visual integrity. In European Conference on Computer Vision, pages 164-180. Springer, 2022. 2 +[53] Sébastien Marcel and Yann Rodriguez. Torchvision the machine-vision package of torch. In International Conference on Multimedia. ACM, 2010. 13 +[54] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 2 +[55] Peter Meerwald and Teddy Furon. Towards practical joint decoding of binary Tardos fingerprinting codes. IEEE Transactions on Information Forensics and Security, 7(4):1168-1180, Apr. 2012. 8 +[56] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 2 +[57] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and + +Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 6, 13, 16 +[58] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2 +[59] Guangyu Nie, Changhoon Kim, Yezhou Yang, and Yi Ren. Attributing image generative models using latent fingerprints. arXiv preprint arXiv:2304.09752, 2023. 2 +[60] Zoe Papakipos and Joanna Bitton. Augly: Data augmentations for robustness. arXiv preprint arXiv:2201.06494, 2022. 13 +[61] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems 32, 2019. 16 +[62] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022.2 +[63] Tong Qiao, Yuyan Ma, Ning Zheng, Hanzhou Wu, Yanli Chen, Ming Xu, and Xiangyang Luo. A novel model watermarking for protecting generative adversarial network. Computers & Security, 127:103102, 2023. 2 +[64] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[65] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 13 +[66] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 13 +[67] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2 +[68] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2, 4, 6, 7, 8, 14, 16 +[69] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 13 +[70] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. Stable diffusion. Accessed Nov. 30 2022, 2022. 1, 7 +[71] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine + +tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 1, 2 +[72] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Spreading vectors for similarity search. ICML, 2019. 17 +[73] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2 +[74] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 13 +[75] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 14 +[76] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2 +[77] Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. Defake: Detection and attribution of fake images generated by text-to-image diffusion models. arXiv preprint arXiv:2210.06998, 2022. 2 +[78] Changhao Shi, Chester Holtz, and Gal Mishne. Online adversarial purification based on self-supervision. arXiv preprint arXiv:2101.09387, 2021. 7 +[79] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[80] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2 +[81] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. *ICLR*, 2021. 13 +[82] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2149–2159, 2022. 13, 20 +[83] Gábor Tardos. Optimal probabilistic fingerprint codes. Journal of the ACM (JACM), 55(2):1-24, 2008. 8 +[84] Yusuke Uchida, Yuki Nagai, Shigeyuki Sakazawa, and Shin'ichi Satoh. Embedding watermarks into deep neural networks. In Proceedings of the 2017 ACM on international conference on multimedia retrieval, pages 269-277, 2017. 2 + +[85] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477, 2022. 2 +[86] Vedran Vukotic, Vivien Chappelier, and Teddy Furon. Are deep neural networks good for blind image watermarking? In WIFS, 2018. 2 +[87] Steven Walton, Ali Hassani, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Stylenat: Giving each head a new perspective. arXiv preprint arXiv:2211.05770, 2022. 2 +[88] Sheng-Yu Wang, Oliver Wang, Andrew Owens, Richard Zhang, and Alexei A Efros. Detecting photoshopped faces by scripting photoshop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10072-10081, 2019. 2 +[89] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 2 +[90] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[91] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 8 +[92] Chen Henry Wu and Fernando De la Torre. Unifying diffusion models' latent space, with applications to cycled diffusion and guidance. arXiv preprint arXiv:2210.05559, 2022. 2 +[93] Hanzhou Wu, Gen Liu, Yuwei Yao, and Xinpeng Zhang. Watermarking neural networks with watermarked images. IEEE Transactions on Circuits and Systems for Video Technology, 31(7):2591-2601, 2020. 2 +[94] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, pages 12062-12072. PMLR, 2021. 7 +[95] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. In ICLR, 2020. 13 +[96] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2 +[97] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 2 + +[98] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14448-14457, 2021. 2, 3 +[99] Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. Responsible disclosure of generative models using scalable fingerprinting. In International Conference on Learning Representations (ICLR), 2022. 1, 2 +[100] Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun, and In So Kweon. Udh: Universal deep hiding for steganography, watermarking, and light field messaging. NeurIPS, 2020. 2 +[101] Chaoning Zhang, Adil Karjauv, Philipp Benz, and In So Kweon. Towards robust deep hiding under nondifferentiable distortions for practical blind watermarking. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5158-5166, 2021. 4, 13 +[102] Jie Zhang, Dongdong Chen, Jing Liao, Han Fang, Weiming Zhang, Wenbo Zhou, Hao Cui, and Nenghai Yu. Model watermarking for image processing networks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12805-12812, 2020. 2 +[103] Kevin Alex Zhang, Alfredo Cuesta-Infante, Lei Xu, and Kalyan Veeramachaneni. Steganogan: High capacity image steganography with gans. arXiv preprint arXiv:1901.03892, 2019. 14 +[104] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 1, 2 +[105] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR. IEEE, 2018. 4, 15 +[106] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In 2019 IEEE international workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2019. 2 +[107] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deep-fake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. 2 +[108] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In ECCV, 2018. 2, 3, 4, 6, 7, 13 \ No newline at end of file diff --git a/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/images.zip b/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..14616a8089e2c51c8732108b35e35b7b44589308 --- /dev/null +++ b/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9518348d5775d79afae4586d895192729bfecc03675524c1e43c705b2ed441ce +size 409505 diff --git a/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/layout.json b/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d75f38e13e748e1b1160d11096f9c1c57da21fef --- /dev/null +++ b/2023/The Stable Signature_ Rooting Watermarks in Latent Diffusion Models/layout.json @@ -0,0 +1,13011 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 517, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 517, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 517, + 121 + ], + "type": "text", + "content": "The Stable Signature: Rooting Watermarks in Latent Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 142, + 538, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 142, + 538, + 158 + ], + "spans": [ + { + "bbox": [ + 59, + 142, + 538, + 158 + ], + "type": "text", + "content": "Pierre Fernandez" + }, + { + "bbox": [ + 59, + 142, + 538, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 59, + 142, + 538, + 158 + ], + "type": "text", + "content": " Guillaume Couairon" + }, + { + "bbox": [ + 59, + 142, + 538, + 158 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 59, + 142, + 538, + 158 + ], + "type": "text", + "content": " Hervé Jégou" + }, + { + "bbox": [ + 59, + 142, + 538, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 59, + 142, + 538, + 158 + ], + "type": "text", + "content": " Matthijs Douze" + }, + { + "bbox": [ + 59, + 142, + 538, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 59, + 142, + 538, + 158 + ], + "type": "text", + "content": " Teddy Furon" + }, + { + "bbox": [ + 59, + 142, + 538, + 158 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 171, + 463, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 171, + 463, + 186 + ], + "spans": [ + { + "bbox": [ + 131, + 171, + 463, + 186 + ], + "type": "text", + "content": "1Meta AI 2Centre Inria de l'Université de Rennes 3Sorbonne University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 201, + 192, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 201, + 192, + 215 + ], + "spans": [ + { + "bbox": [ + 143, + 201, + 192, + 215 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "text", + "content": "Generative image modeling enables a wide range of applications but raises ethical concerns about responsible deployment. We introduce an active content tracing method combining image watermarking and Latent Diffusion Models. The goal is for all generated images to conceal an invisible watermark allowing for future detection and/or identification. The method quickly fine-tunes the latent decoder of the image generator, conditioned on a binary signature. A pre-trained watermark extractor recovers the hidden signature from any generated image and a statistical test then determines whether it comes from the generative model. We evaluate the invisibility and robustness of the watermarks on a variety of generation tasks, showing that the Stable Signature is robust to image modifications. For instance, it detects the origin of an image generated from a text prompt, then cropped to keep " + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "text", + "content": " of the content, with " + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "inline_equation", + "content": "90 + \\%" + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "text", + "content": " accuracy at a false positive rate below " + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 46, + 228, + 290, + 432 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 445, + 128, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 128, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 128, + 458 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 467, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 287, + 551 + ], + "type": "text", + "content": "Recent progress in generative modeling and natural language processing enable easy creation and manipulation of photo-realistic images, such as with DALL-E 2 [64] or Stable Diffusion [68]. They have given birth to many image edition tools like ControlNet [104], Instruct-Pix2Pix [7], and others [13, 28, 71], that are becoming mainstream creative tools for artists, designers, and the general public." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 552, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 288, + 684 + ], + "type": "text", + "content": "While this is a great step forward for generative AI, it also undermines confidence in the authenticity or veracity of photo-realistic images. Indeed, methods for photorealistic image edition existed before, but generative AI significantly lowers the barriers to convincing synthetic image generation and edition (e.g. a generated picture recently won an art competition [29]). This raises new risks like deep fakes, impersonation or copyright usurpation [8, 17]. A tool to determine that images are AI-generated would make it easier to ensure their compliance with ethical standards and to remove them from certain platforms." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 219, + 536, + 416 + ], + "blocks": [ + { + "bbox": [ + 320, + 219, + 536, + 416 + ], + "lines": [ + { + "bbox": [ + 320, + 219, + 536, + 416 + ], + "spans": [ + { + "bbox": [ + 320, + 219, + 536, + 416 + ], + "type": "image", + "image_path": "2c7b5c2fde8d9e5922d8418a2133684ce4791ae3ac6f121e907c2d4e645869c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 419, + 545, + 441 + ], + "lines": [ + { + "bbox": [ + 305, + 419, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 419, + 545, + 441 + ], + "type": "text", + "content": "Figure 1. Overview. The latent decoder can be fine-tuned to preemptively embed a signature into all generated images." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 546, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 592 + ], + "type": "text", + "content": "A baseline solution to identify generated images is forensics, i.e. methods to detect generated/manipulated images passively (the image is not modified for identification). An active baseline is to apply existing watermarking methods after the image generation. Watermarking invisibly embeds a secret message into the image, which can then be extracted and used to identify the image. This has several drawbacks. If the model leaks or is open-sourced, the post-generation watermarking can be removed trivially. The open source Stable Diffusion [70] is a case in point, since removing the watermark amounts to commenting out a single line in the source code." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Our Stable Signature method merges watermarking into the generation process itself, without any architectural change. It adjusts the pre-trained generative model such that all the images it produces conceal a given watermark. There are several advantages to this approach [46, 99]. It does not require additional processing of the generated image, which makes the watermarking computationally lighter, straightforward, and secure. Model providers could deploy their models to different user groups with a unique watermark, and monitor that they are used in a responsible manner." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 292, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 292, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 292, + 713 + ], + "type": "text", + "content": "*Work supported by ANR / AID under Chaire SAIDA ANR-20-CHIA-0011. Correspondance to pfz@meta.com" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22466" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "They could give art platforms, news outlets and other sharing platforms the ability to detect when an image has been generated by their AI." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 275 + ], + "type": "text", + "content": "We focus on Latent Diffusion Models (LDM) [68] that can perform a wide range of generative tasks. We show that simply fine-tuning a small part of the generative model – the decoder that generates images from the latent vectors – is enough to natively embed a watermark into generated images. Stable Signature does not require an architectural change and does not modify the diffusion process. Hence it is to our knowledge compatible with all LDM-based generative methods [7, 13, 62, 71, 104]. The fine-tuning stage is performed by back-propagating a combination of a perceptual image loss and a message decoding loss from a watermark extractor back to the LDM decoder. We pre-train the extractor with a simplified version of the deep watermarking method HiDDeN [108]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 275, + 288, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 288, + 418 + ], + "type": "text", + "content": "We create an evaluation benchmark close to real world situations where images may be edited. The tasks are: detection of AI generated images, tracing models from their generations. For instance, we detect " + }, + { + "bbox": [ + 46, + 275, + 288, + 418 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 275, + 288, + 418 + ], + "type": "text", + "content": " of images generated with the generative model, even if they are cropped to " + }, + { + "bbox": [ + 46, + 275, + 288, + 418 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 275, + 288, + 418 + ], + "type": "text", + "content": " of their original size, while flagging only one false positive every " + }, + { + "bbox": [ + 46, + 275, + 288, + 418 + ], + "type": "inline_equation", + "content": "10^{6}" + }, + { + "bbox": [ + 46, + 275, + 288, + 418 + ], + "type": "text", + "content": " images. To ensure that the model's utility is not weakened, we show that the FID [34] score of the generation is not affected and that the generated images are perceptually indistinguishable from the ones produced by the original model. This is done over several tasks involving LDM (text-to-image, inpainting, edition, etc.)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 419, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 515 + ], + "type": "text", + "content": "As a summary, (1) we efficiently merge watermarking into the generation process of LDMs, in a way that is compatible with most of the LDM-based generative methods; (2) we demonstrate how it can be used to detect and trace generated images, through a real-world evaluation benchmark; (3) we compare to post-hoc watermarking methods, showing that it is competitive while being more secure and efficient, and (4) evaluate robustness to intentional attacks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 525, + 134, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 134, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 134, + 538 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 629 + ], + "type": "text", + "content": "Image generation has long been dominated by GANs, still state-of-the-art on many datasets [37, 38, 39, 76, 87]. Transformers have also been successfully used for modeling image [66, 19] or video [79] distributions, providing higher diversity at the expense of increased inference time. Images are typically converted to token lists using vector-quantized architectures [20, 67, 96], relying on an image decoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Diffusion models [18, 36, 58, 80] are a huge improvement in text-conditional image generation, current models synthesize high-resolution photo-realistic images for a wide variety of text prompts [3, 35, 65, 69, 74]. They can also perform conditional image generation – like inpainting or text-guided image editing – by fine-tuning the diffusion model with additional conditioning, e.g. masked input im" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "age, segmentation map, etc. [50, 73]. Because of their iterative denoising algorithm, diffusion models can also be adapted for image editing in a zero-shot fashion by guiding the generative process [13, 33, 40, 56, 85, 92]. All these methods, when combined with Stable Diffusion, operate in the latent rather than image space, requiring a latent decoder to produce an RGB image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 165, + 545, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 545, + 260 + ], + "type": "text", + "content": "Detection of AI-generated/manipulated images is notably active in the context of deep-fake detection [32, 107]. Many works focus on the detection of GAN-generated images [10, 31, 89, 106]. One approach is to detect inconsistencies, via lights, perspective or physical objects [21, 22, 45, 52, 88]. These approaches are restricted to photorealistic images or faces, artworks not intended to be physically correct are not covered." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 262, + 546, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 546, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 546, + 406 + ], + "type": "text", + "content": "Other approaches track traces left by the generators in the spatial [54, 97] or frequency [26, 106] domains. There are extensions to diffusion models in recent works [12, 77] that show encouraging results. However purely relying on forensics and passive detection is limiting, e.g. the best performing method to our knowledge [12] is able to detect " + }, + { + "bbox": [ + 304, + 262, + 546, + 406 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 262, + 546, + 406 + ], + "type": "text", + "content": " of generated images for a false positive rate around 1/100: if a user-generated content platform were to receive 1 billion images every day, it would wrongly flag 10 million images to detect only half of the generated images. Besides, passive techniques cannot trace images from different versions of the same model, in contrast with watermarking." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 415, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 498 + ], + "type": "text", + "content": "Image watermarking has long been studied in the context of tracing and intellectual property protection [14]. More recently, deep learning encoder/extractors like HiD-DeN [2, 44, 51, 100, 108] or iterative methods by Vukotic et al. [25, 42, 86] showed competitive results in terms of robustness to a wide range of transformations, namely geometric ones." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 500, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 545, + 691 + ], + "type": "text", + "content": "In the specific case of generative models, some works aim to watermark the training set on which the generative model is learned [98]. It is highly inefficient since every new message to embed requires a new training pipeline. Merging the watermarking and the generative process is a recent idea [23, 46, 59, 63, 93, 99, 102], that is closer to the model watermarking literature [84]. They suffer from two strong limitations. First, these methods only apply to GANs, while LDM are progressively replacing them for most applications. Second, watermarking is incorporated in the training process of the GAN from the start. This strategy is unsustainable because the generative model training is more and more costly1. Our work shows that a quick finetuning of the latent decoder part of the generative model is enough to achieve a good watermarking performance, provided that the watermark extractor is well chosen." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 545, + 712 + ], + "type": "text", + "content": "1 Stable Diffusion training costs \\(\\sim\\)\\\\(600k of cloud compute (Wikipedia)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22467" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 240, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 240, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 240, + 85 + ], + "type": "text", + "content": "3. Problem Statement & Background" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 140 + ], + "type": "text", + "content": "Figure 1 shows a model provider Alice who deploys a latent diffusion model to users Bobs. Stable Signature embeds a binary signature into the generated images. This section derives how Alice can use this signature for two scenarios:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 143, + 287, + 228 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 47, + 143, + 287, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 143, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 143, + 287, + 192 + ], + "type": "text", + "content": "- Detection: \"Is it generated by my model?\" Alice detects if an image was generated by her model. Generated images should be flagged as reliably as possible, while controlling the probability of flagging a natural image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 193, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 193, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 286, + 228 + ], + "type": "text", + "content": "- Identification: \"Who generated this image?\". Alice monitors who created each image, while avoiding to mistakenly identifying a Bob." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 237, + 230, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 230, + 250 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 230, + 250 + ], + "type": "text", + "content": "3.1. Image watermarking for detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 256, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 316 + ], + "type": "text", + "content": "Alice embeds a " + }, + { + "bbox": [ + 46, + 256, + 287, + 316 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 256, + 287, + 316 + ], + "type": "text", + "content": "-bit binary signature into the generated images. The watermark extractor then decodes messages from the images it receives and detects when the message is close to Alice's signature. An example application is to block AI-generated images on a content sharing platform." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 324, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 373 + ], + "type": "text", + "content": "Statistical test Let " + }, + { + "bbox": [ + 46, + 324, + 287, + 373 + ], + "type": "inline_equation", + "content": "m \\in \\{0,1\\}^k" + }, + { + "bbox": [ + 46, + 324, + 287, + 373 + ], + "type": "text", + "content": " be Alice's signature. We extract the message " + }, + { + "bbox": [ + 46, + 324, + 287, + 373 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 46, + 324, + 287, + 373 + ], + "type": "text", + "content": " from an image " + }, + { + "bbox": [ + 46, + 324, + 287, + 373 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 324, + 287, + 373 + ], + "type": "text", + "content": " and compare it to " + }, + { + "bbox": [ + 46, + 324, + 287, + 373 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 324, + 287, + 373 + ], + "type": "text", + "content": ". As done in previous works [46, 98], the detection test relies on the number of matching bits " + }, + { + "bbox": [ + 46, + 324, + 287, + 373 + ], + "type": "inline_equation", + "content": "M(m,m')" + }, + { + "bbox": [ + 46, + 324, + 287, + 373 + ], + "type": "text", + "content": ": if" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 84, + 375, + 287, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 375, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 84, + 375, + 287, + 388 + ], + "type": "interline_equation", + "content": "M \\left(m, m ^ {\\prime}\\right) \\geq \\tau \\text {w h e r e} \\tau \\in \\{0, \\dots , k \\}, \\tag {1}", + "image_path": "f89db59f9871e48f8dd4bf627d3006e6a71f9367978546f26848afa317eeb1a5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 390, + 286, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 286, + 414 + ], + "type": "text", + "content": "then the image is flagged. This provides a level of robustness to imperfections of the watermarking." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "text", + "content": "Formally, we test the statistical hypothesis " + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "inline_equation", + "content": "H_{1}" + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "text", + "content": ": \" " + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "text", + "content": " was generated by Alice's model\" against the null hypothesis " + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "inline_equation", + "content": "H_{0}" + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "text", + "content": ": \" " + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "text", + "content": " was not generated by Alice's model\". Under " + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "inline_equation", + "content": "H_{0}" + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "text", + "content": " (i.e. for vanilla images), we assume that bits " + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "inline_equation", + "content": "m_{1}', \\ldots, m_{k}'" + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "text", + "content": " are (i.i.d.) Bernoulli random variables with parameter 0.5. Then " + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "inline_equation", + "content": "M(m, m')" + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "text", + "content": " follows a binomial distribution with parameters " + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "inline_equation", + "content": "(k, 0.5)" + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "text", + "content": ". We verify this assumption experimentally in App. B.5. The False Positive Rate (FPR) is the probability that " + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "inline_equation", + "content": "M(m, m')" + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "text", + "content": " takes a value bigger than the threshold " + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "text", + "content": ". It is obtained from the CDF of the binomial distribution, and a closed-form can be written with the regularized incomplete beta function " + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "inline_equation", + "content": "I_{x}(a; b)" + }, + { + "bbox": [ + 46, + 415, + 287, + 559 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 563, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 563, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 61, + 563, + 287, + 578 + ], + "type": "interline_equation", + "content": "\\operatorname {F P R} (\\tau) = \\mathbb {P} (M > \\tau | H _ {0}) = I _ {1 / 2} (\\tau + 1, k - \\tau). \\tag {2}", + "image_path": "4b1dcbc103f74a14d29df5d347789392c662f0b24f7dde1626326ac235b1ca8e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 508, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 508, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 508, + 84 + ], + "type": "text", + "content": "3.2. Image watermarking for identification" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 90, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 163 + ], + "type": "text", + "content": "Alice now embeds a signature " + }, + { + "bbox": [ + 304, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "m^{(i)}" + }, + { + "bbox": [ + 304, + 90, + 545, + 163 + ], + "type": "text", + "content": " drawn randomly from " + }, + { + "bbox": [ + 304, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\{0,1\\}^k" + }, + { + "bbox": [ + 304, + 90, + 545, + 163 + ], + "type": "text", + "content": " into the model distributed to " + }, + { + "bbox": [ + 304, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathrm{Bob}^{(i)}" + }, + { + "bbox": [ + 304, + 90, + 545, + 163 + ], + "type": "text", + "content": " (for " + }, + { + "bbox": [ + 304, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "i = 1 \\cdots N" + }, + { + "bbox": [ + 304, + 90, + 545, + 163 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 90, + 545, + 163 + ], + "type": "text", + "content": " the number of Bobs). Alice can trace any misuse of her model: generated images violating her policy (gore content, deepfakes) are linked back to the specific Bob by comparing the extracted message to Bobs' signatures." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "text", + "content": "Statistical test We compare the message " + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "text", + "content": " from the watermark extractor to " + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "inline_equation", + "content": "(m^{(1)}, \\ldots, m^{(N)})" + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "text", + "content": ". There are now " + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "text", + "content": " detection hypotheses to test. If the " + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "text", + "content": " hypotheses are rejected, we conclude that the image was not generated by any of the models. Otherwise, we attribute the image to " + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "inline_equation", + "content": "\\operatorname{argmax}_{i=1..N} M(m', m^{(i)})" + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "text", + "content": ". With regards to the detection task, false positives are more likely since there are " + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "text", + "content": " tests. The global FPR at a given threshold " + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 172, + 545, + 267 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 271, + 545, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 271, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 319, + 271, + 545, + 285 + ], + "type": "interline_equation", + "content": "\\operatorname {F P R} (\\tau , N) = 1 - (1 - \\operatorname {F P R} (\\tau)) ^ {N} \\approx N. \\operatorname {F P R} (\\tau). \\tag {3}", + "image_path": "5a763b2f95f718dfb9a34d6fa83e74c12295fe259740f4003c3a5007a010ce3c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 289, + 545, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 349 + ], + "type": "text", + "content": "Equation (3) (resp. (2)), is used reversely: we find threshold " + }, + { + "bbox": [ + 304, + 289, + 545, + 349 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 289, + 545, + 349 + ], + "type": "text", + "content": " to achieve a required FPR for identification (resp. detection). Note that these formulae hold only under the assumption of i.i.d. Bernoulli bits extracted from vanilla images. This condition is enforced in the next section." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 361, + 361, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 361, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 361, + 373 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 381, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 453 + ], + "type": "text", + "content": "Stable Signature modifies the generative network so that the generated images have a given signature through a fixed watermark extractor. It is trained in two phases. First, we create the watermark extractor network " + }, + { + "bbox": [ + 304, + 381, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 381, + 545, + 453 + ], + "type": "text", + "content": ". We then fine-tune the LDM decoder " + }, + { + "bbox": [ + 304, + 381, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 381, + 545, + 453 + ], + "type": "text", + "content": ", such that all generated images yield a given signature through " + }, + { + "bbox": [ + 304, + 381, + 545, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 381, + 545, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 461, + 504, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 461, + 504, + 474 + ], + "spans": [ + { + "bbox": [ + 305, + 461, + 504, + 474 + ], + "type": "text", + "content": "4.1. Pre-training the watermark extractor" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": "We use HiDDeN [108], a classical method in the deep watermarking literature. It jointly optimizes the parameters of watermark encoder " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_E" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": " and extractor network " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": " to embed " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": "-bit messages into images, robustly to transformations that are applied during training. We discard " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_E" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": " after training, since only " + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 479, + 545, + 552 + ], + "type": "text", + "content": " serves our purpose." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 552, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 545, + 576 + ], + "type": "text", + "content": "Formally, " + }, + { + "bbox": [ + 306, + 552, + 545, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_E" + }, + { + "bbox": [ + 306, + 552, + 545, + 576 + ], + "type": "text", + "content": " takes as inputs a cover image " + }, + { + "bbox": [ + 306, + 552, + 545, + 576 + ], + "type": "inline_equation", + "content": "x_o \\in \\mathbb{R}^{W \\times H \\times 3}" + }, + { + "bbox": [ + 306, + 552, + 545, + 576 + ], + "type": "text", + "content": " and a " + }, + { + "bbox": [ + 306, + 552, + 545, + 576 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 552, + 545, + 576 + ], + "type": "text", + "content": "-bit message " + }, + { + "bbox": [ + 306, + 552, + 545, + 576 + ], + "type": "inline_equation", + "content": "m \\in \\{0,1\\}^k" + }, + { + "bbox": [ + 306, + 552, + 545, + 576 + ], + "type": "text", + "content": ". Similar to" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 55, + 604, + 244, + 681 + ], + "blocks": [ + { + "bbox": [ + 55, + 591, + 193, + 601 + ], + "lines": [ + { + "bbox": [ + 55, + 591, + 193, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 193, + 601 + ], + "type": "text", + "content": "(a) Pre-train watermark encoder/extractor" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 604, + 244, + 681 + ], + "lines": [ + { + "bbox": [ + 55, + 604, + 244, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 244, + 681 + ], + "type": "image", + "image_path": "a8e372dd0b131cee24e7f98cf1a102573b2555b1a632bc6dc9f3c77ca79c9a7b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 256, + 604, + 430, + 680 + ], + "blocks": [ + { + "bbox": [ + 247, + 592, + 339, + 601 + ], + "lines": [ + { + "bbox": [ + 247, + 592, + 339, + 601 + ], + "spans": [ + { + "bbox": [ + 247, + 592, + 339, + 601 + ], + "type": "text", + "content": "(b) Fine-tune LDM decoder" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 256, + 604, + 430, + 680 + ], + "lines": [ + { + "bbox": [ + 256, + 604, + 430, + 680 + ], + "spans": [ + { + "bbox": [ + 256, + 604, + 430, + 680 + ], + "type": "image", + "image_path": "c9a2e0e6c5c67da3f73d1f1ed5f06268f224de6ad0bd88b0df88ac520db22afd.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 593, + 482, + 601 + ], + "lines": [ + { + "bbox": [ + 440, + 593, + 482, + 601 + ], + "spans": [ + { + "bbox": [ + 440, + 593, + 482, + 601 + ], + "type": "text", + "content": "(c) Generate" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "lines": [ + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "type": "text", + "content": "Figure 2. Steps of the method. (a) We pre-train a watermark encoder " + }, + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_E" + }, + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "type": "text", + "content": " and extractor " + }, + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "type": "text", + "content": ", to extract binary messages. (b) We fine-tune the decoder " + }, + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "type": "text", + "content": " of the LDM's auto-encoder with a fixed signature " + }, + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "type": "text", + "content": " such that all the generated images (c) lead to " + }, + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 45, + 689, + 545, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 519, + 607, + 536, + 624 + ], + "blocks": [ + { + "bbox": [ + 519, + 607, + 536, + 624 + ], + "lines": [ + { + "bbox": [ + 519, + 607, + 536, + 624 + ], + "spans": [ + { + "bbox": [ + 519, + 607, + 536, + 624 + ], + "type": "image", + "image_path": "541d235c8465ce2d5220684d78dcd49854ea3ccb8613edfda496a461ad40b6e2.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 441, + 627, + 523, + 673 + ], + "blocks": [ + { + "bbox": [ + 441, + 607, + 509, + 624 + ], + "lines": [ + { + "bbox": [ + 441, + 607, + 509, + 624 + ], + "spans": [ + { + "bbox": [ + 441, + 607, + 509, + 624 + ], + "type": "text", + "content": "Text - Super-resolution Inpainting - Editing - ..." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 441, + 627, + 523, + 673 + ], + "lines": [ + { + "bbox": [ + 441, + 627, + 523, + 673 + ], + "spans": [ + { + "bbox": [ + 441, + 627, + 523, + 673 + ], + "type": "image", + "image_path": "449466aa4f7c59bf5152aa317699ad2e0c6a8690bd875af411b24312fba4393e.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22468" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "ReDMark [2], " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_E" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " outputs a residual image " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " of the same size as " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "x_{o}" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": ", that is multiplied by a factor " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " to produce watermarked image " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "x_w = x_o + \\alpha \\delta" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": ". At each optimization step an image transformation " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " is sampled from a set " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " that includes common image processing operations such as cropping and JPEG compression. A \"soft\" message is extracted from the transformed image: " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "m' = \\mathcal{W}(T(x_w))" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " (at inference time, the decoded message is given by the signs of the components of " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": "). The message loss is the Binary Cross Entropy (BCE) between " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": " and the sigmoid " + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "inline_equation", + "content": "\\sigma(m')" + }, + { + "bbox": [ + 46, + 72, + 289, + 193 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 193, + 287, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 193, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 287, + 225 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m} = - \\sum_ {i = 1} ^ {k} m _ {i} \\cdot \\log \\sigma (m _ {i} ^ {\\prime}) + (1 - m _ {i}) \\cdot \\log (1 - \\sigma (m _ {i} ^ {\\prime})).", + "image_path": "bdfc25a9e525c6b480c0cda15fb77dc9ccff005564c7185edeba87757307e562.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 232, + 287, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 287, + 268 + ], + "type": "text", + "content": "The network architectures are kept simple to ease the LDM fine-tuning in the second phase. They are the same as HiDDeN [108] (see App. A.1) with two changes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "text", + "content": "First, since " + }, + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_E" + }, + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "text", + "content": " is discarded, its perceptual quality is not as important, so the perceptual loss or the adversarial network are not needed. Instead, the distortion is constrained by a tanh function on output of " + }, + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_E" + }, + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "text", + "content": " and by the scaling factor " + }, + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "text", + "content": ". This improves the bit accuracy of the recovered message and makes it possible to increase its size " + }, + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 269, + 287, + 339 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 340, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 413 + ], + "type": "text", + "content": "Second, we observed that " + }, + { + "bbox": [ + 46, + 340, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 340, + 287, + 413 + ], + "type": "text", + "content": "'s output bits for vanilla images are correlated and highly biased, which violates the assumptions of Sec. 3.1. Therefore, to get closer to i.i.d. dimensions we remove the bias and decorrelate the outputs of " + }, + { + "bbox": [ + 46, + 340, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 340, + 287, + 413 + ], + "type": "text", + "content": " by applying a PCA whitening transformation (more details in App. A.1)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 419, + 226, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 226, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 226, + 432 + ], + "type": "text", + "content": "4.2. Fine-tuning the generative model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "content": "In LDM, the diffusion happens in the latent space of an auto-encoder. The latent vector " + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "content": " obtained at the end of the diffusion is input to decoder " + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "content": " to produce an image. Here we fine-tune " + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "content": " such that the image contains a given message " + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "content": " that can be extracted by " + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "content": ". Stable Signature is compatible with many generative tasks, since modifying only " + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 437, + 287, + 521 + ], + "type": "text", + "content": " does not affect the diffusion process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "type": "text", + "content": "First, we fix the signature " + }, + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "type": "inline_equation", + "content": "m = (m_1,\\ldots ,m_k)\\in \\{0,1\\} ^k" + }, + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "type": "text", + "content": " The fine-tuning of " + }, + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m" + }, + { + "bbox": [ + 46, + 521, + 287, + 556 + ], + "type": "text", + "content": " is inspired by the original training of the auto-encoder in LDM [68]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "text", + "content": "Training image " + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "text", + "content": " is fed to the LDM encoder " + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "text", + "content": " that outputs activation map " + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(x) \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "text", + "content": ", downsampled by a power-of-two factor " + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "f = H / h = W / w" + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "text", + "content": ". The decoder reconstructs an image " + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "x' = \\mathcal{D}_m(z)" + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "text", + "content": " and the extractor recovers " + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "m' = \\mathcal{W}(x')" + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "text", + "content": ". The message loss is the BCE between " + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "text", + "content": " and the original " + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m = \\mathrm{BCE}(\\sigma(m'), m)" + }, + { + "bbox": [ + 46, + 556, + 287, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": "In addition, the original decoder " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": " reconstructs the image without watermark: " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "x_{o}^{\\prime} = \\mathcal{D}(z)" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": ". The image perceptual loss " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{i}}" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "x^{\\prime}" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "x_{o}^{\\prime}" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": ", controls the distortion. We use the Watson-VGG perceptual loss introduced by Czolbe et" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 545, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 118 + ], + "type": "text", + "content": "al. [15], an improved version of LPIPS [105]. It is essential that the decoder learns luminance and contrast masking to add less visible watermarks. The weights of " + }, + { + "bbox": [ + 306, + 72, + 545, + 118 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m" + }, + { + "bbox": [ + 306, + 72, + 545, + 118 + ], + "type": "text", + "content": " are optimized to minimize" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 389, + 118, + 545, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 118, + 545, + 130 + ], + "spans": [ + { + "bbox": [ + 389, + 118, + 545, + 130 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {m}} + \\lambda_ {\\mathrm {i}} \\mathcal {L} _ {\\mathrm {i}}. \\tag {4}", + "image_path": "5cfe796d8f56db8ba61a897b9f60bcb00c4ac9ff205eaae3daca57b4d931e3da.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 133, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 547, + 205 + ], + "type": "text", + "content": "This is done over 100 iterations with the AdamW optimizer [49] and batch of size 4, i.e. the fine-tuning sees fewer than 500 images and takes one minute on a single GPU. The learning rate follows a cosine annealing schedule with 20 iterations of linear warmup to " + }, + { + "bbox": [ + 304, + 133, + 547, + 205 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 133, + 547, + 205 + ], + "type": "text", + "content": " and decays to " + }, + { + "bbox": [ + 304, + 133, + 547, + 205 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 304, + 133, + 547, + 205 + ], + "type": "text", + "content": ". The factor " + }, + { + "bbox": [ + 304, + 133, + 547, + 205 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{i}}" + }, + { + "bbox": [ + 304, + 133, + 547, + 205 + ], + "type": "text", + "content": " in (4) is set to 0.2 by default." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 215, + 541, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 215, + 541, + 229 + ], + "spans": [ + { + "bbox": [ + 305, + 215, + 541, + 229 + ], + "type": "text", + "content": "5. Text-to-Image Watermarking Performance" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 235, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 545, + 307 + ], + "type": "text", + "content": "This section discusses our method's detection and identification capability for images generated by a Stable-Diffusion-like model " + }, + { + "bbox": [ + 304, + 235, + 545, + 307 + ], + "type": "inline_equation", + "content": "[68]^3" + }, + { + "bbox": [ + 304, + 235, + 545, + 307 + ], + "type": "text", + "content": ". We apply generative models watermarked with 48-bit signatures on prompts of the MS-COCO [47] validation set. We evaluate detection and identification on the outputs, as illustrated in Figure 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 307, + 545, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 379 + ], + "type": "text", + "content": "We evaluate their robustness to different transformations applied to generated images: strong cropping (10% of the image remaining), brightness shift (strength factor 2.0), as well as a combination of crop 50%, brightness shift 1.5 and JPEG 80. This covers typical geometric and photometric edits (see Fig. 5 for visual examples)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 379, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 379, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 305, + 379, + 545, + 403 + ], + "type": "text", + "content": "The detection rates are partly obtained from experiments and partly by extrapolating small-scale measurements." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 410, + 407, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 407, + 421 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 407, + 421 + ], + "type": "text", + "content": "5.1. Detection results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 428, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 545, + 559 + ], + "type": "text", + "content": "For detection, we fine-tune the decoder of the LDM with a random key " + }, + { + "bbox": [ + 304, + 428, + 545, + 559 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 428, + 545, + 559 + ], + "type": "text", + "content": ", generate 1000 images and use the test of Eq. (1). We report the tradeoff between True Positive Rate (TPR), i.e. the probability of flagging a generated image and the FPR, while varying " + }, + { + "bbox": [ + 304, + 428, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\tau \\in \\{0,..,48\\}" + }, + { + "bbox": [ + 304, + 428, + 545, + 559 + ], + "type": "text", + "content": ". For instance, for " + }, + { + "bbox": [ + 304, + 428, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\tau = 0" + }, + { + "bbox": [ + 304, + 428, + 545, + 559 + ], + "type": "text", + "content": ", we flag all images so " + }, + { + "bbox": [ + 304, + 428, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{FPR} = 1" + }, + { + "bbox": [ + 304, + 428, + 545, + 559 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 428, + 545, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{TPR} = 1" + }, + { + "bbox": [ + 304, + 428, + 545, + 559 + ], + "type": "text", + "content": ". The TPR is measured directly, while the FPR is inferred from Eq. (2), because it would otherwise be too small to be measured on reasonably sized problems (this approximation is validated experimentally in App. B.6). The experiment is run on 10 random signatures and we report averaged results." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "text", + "content": "Figure 3 shows the tradeoff under image transformations. For example, when the generated images are not modified, Stable Signature detects " + }, + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "text", + "content": " of them, while only 1 vanilla image out of " + }, + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "inline_equation", + "content": "10^{9}" + }, + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "text", + "content": " is flagged. At the same FPR " + }, + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "inline_equation", + "content": "= 10^{-9}" + }, + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "text", + "content": ", Stable Signature detects " + }, + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "inline_equation", + "content": "84\\%" + }, + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "text", + "content": " of generated images for a crop that keeps " + }, + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "text", + "content": " of the image, and " + }, + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 304, + 560, + 545, + 668 + ], + "type": "text", + "content": " for a transformation that combines a crop, a color shift, and a JPEG compression. For comparison, we report results of a state-of-the-art passive method [12], applied on resized and" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": "2The transformation needs to be differentiable in pixel space. This is not the case for JPEG compression so we use the forward attack simulation layer introduced by Zhang et al. [101]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 673, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 545, + 713 + ], + "type": "text", + "content": "3We refrain from experimenting with pre-existing third-party generative models, such as Stable Diffusion or LDMs, and instead use a large diffusion model (2.2B parameters) trained on an internal dataset of 330M licensed image-text pairs." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22469" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "compressed images. As expected, we observe that these baseline results have orders of magnitudes larger FPR than Stable Signature, which actively marks the content." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 120, + 167, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 120, + 167, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 120, + 167, + 132 + ], + "type": "text", + "content": "5.2. Identification results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 140, + 287, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 140, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 140, + 287, + 213 + ], + "type": "text", + "content": "Each Bob has its own copy of the generative model. Given an image, the goal is to find if any of the " + }, + { + "bbox": [ + 46, + 140, + 287, + 213 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 140, + 287, + 213 + ], + "type": "text", + "content": " Bobs created it (detection) and if so, which one (identification). There are 3 types of error: false positive: flag a vanilla image; false negative: miss a generated image; false accusation: flag a generated image but identify the wrong Bob." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "text", + "content": "For evaluation, we fine-tune " + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "inline_equation", + "content": "N' = 1000" + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "text", + "content": " models with random signatures. Each model generates 100 images. For each of these 100k watermarked images, we extract the Stable Signature message, compute the matching score with all " + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "text", + "content": " signatures and select the Bob with the highest score. The image is predicted to be generated by that Bob if this score is above threshold " + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "text", + "content": ". We determined " + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathrm{FPR} = 10^{-6}" + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "text", + "content": ", see Eq. (3). For example, for " + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "inline_equation", + "content": "N = 1" + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\tau = 41" + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "text", + "content": " and for " + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "inline_equation", + "content": "N = 1000" + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\tau = 44" + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "text", + "content": ". Accuracy is extrapolated beyond the " + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "inline_equation", + "content": "N'" + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "text", + "content": " Bobs by adding additional signatures and having " + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "inline_equation", + "content": "N > N'" + }, + { + "bbox": [ + 46, + 214, + 287, + 346 + ], + "type": "text", + "content": " (e.g. Bobs that have not generated any images)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "text", + "content": "Figure 4 reports the per-transformation identification accuracy. For example, we identify a Bob among " + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "inline_equation", + "content": "N = 10^{5}" + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "text", + "content": " accuracy when the image is not modified. Note that for the combined edit, this becomes " + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "text", + "content": ". This may still be dissuasive: if a Bob generates 3 images, he will be identified " + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "text", + "content": " of the time. We observe that at this scale, the false accusation rate is zero, i.e. we never identify the wrong Bob. This is because " + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "text", + "content": " is set high to avoid FPs, which also makes false accusations unlikely. We observe that the identification accuracy decreases when " + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "text", + "content": " increases, because the threshold " + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "text", + "content": " required to avoid false positives is higher when " + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 347, + 287, + 515 + ], + "type": "text", + "content": " increases, as pointed out by the approximation in (3). In a nutshell, by distributing more models, Alice trades some accuracy of detection against the ability to identify Bobs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 71, + 432, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 432, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 432, + 85 + ], + "type": "text", + "content": "6. Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 91, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 163 + ], + "type": "text", + "content": "The previous section showed how to leverage watermarks for detection and identification of images generated from text prompts. We now present more general results on robustness and image quality for different generative tasks. We also compare Stable Signature to post-generation watermarking algorithms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 171, + 455, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 171, + 455, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 455, + 182 + ], + "type": "text", + "content": "6.1. Tasks & evaluation metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 189, + 545, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 189, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 305, + 189, + 545, + 261 + ], + "type": "text", + "content": "Since our method only involves the LDM decoder, it is compatible with many generative tasks. We evaluate text-to-image generation and image edition on the validation set of MS-COCO [47], super-resolution and inpainting on the validation set of ImageNet [16] (all evaluation details are available in App. A.3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 261, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 261, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 261, + 545, + 369 + ], + "type": "text", + "content": "We evaluate the image distortion with the Peak Signal-to-Noise Ratio (PSNR), which is defined as " + }, + { + "bbox": [ + 305, + 261, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR}(x, x') = -10 \\cdot \\log_{10}(\\mathrm{MSE}(x, x'))" + }, + { + "bbox": [ + 305, + 261, + 545, + 369 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 305, + 261, + 545, + 369 + ], + "type": "inline_equation", + "content": "x, x' \\in [0, 1]^{c \\times h \\times w}" + }, + { + "bbox": [ + 305, + 261, + 545, + 369 + ], + "type": "text", + "content": ", as well as Structural Similarity score (SSIM) [90]. They compare images generated with and without watermark. On the other hand, we evaluate the diversity and quality of the generated images with the Fréchet Inception Distance (FID) [34]. The bit accuracy – the percentage of bits correctly decoded – evaluates the watermarks' robustness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 376, + 446, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 446, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 446, + 389 + ], + "type": "text", + "content": "6.2. Image generation quality" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 395, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 395, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 395, + 545, + 515 + ], + "type": "text", + "content": "Figure 6 shows qualitative examples of how the image generation is altered by the latent decoder's fine-tuning. The difference is very hard to perceive even for a trained eye. This is surprising for such a low PSNR, especially since the watermark embedding is not constrained by any Human Visual System like in professional watermarking techniques. Most interestingly, the LDM decoder has indeed learned to add the watermark signal only over textured areas where the human eyes are not sensitive, while the uniform backgrounds are kept intact (see the pixel-wise difference)." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 53, + 536, + 208, + 659 + ], + "blocks": [ + { + "bbox": [ + 53, + 536, + 208, + 659 + ], + "lines": [ + { + "bbox": [ + 53, + 536, + 208, + 659 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 208, + 659 + ], + "type": "image", + "image_path": "830dc5ae205f38ce29636fc7375bf256df868f68501542c82e7659d8ce7b3dde.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 667, + 214, + 712 + ], + "lines": [ + { + "bbox": [ + 46, + 667, + 214, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 667, + 214, + 712 + ], + "type": "text", + "content": "Figure 3. Detection results. TPR/FPR curve of the detection under different transformations. Forensics† indicates passive detection (without watermark) [12]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 230, + 535, + 385, + 659 + ], + "blocks": [ + { + "bbox": [ + 230, + 535, + 385, + 659 + ], + "lines": [ + { + "bbox": [ + 230, + 535, + 385, + 659 + ], + "spans": [ + { + "bbox": [ + 230, + 535, + 385, + 659 + ], + "type": "image", + "image_path": "1d8f6ca966033bc65ea96bef8580b6508dc40c9d42b1f302623a6907be82a3bf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 667, + 391, + 712 + ], + "lines": [ + { + "bbox": [ + 223, + 667, + 391, + 712 + ], + "spans": [ + { + "bbox": [ + 223, + 667, + 391, + 712 + ], + "type": "text", + "content": "Figure 4. Identification results. Proportion of well-identified Bobs. Detection with " + }, + { + "bbox": [ + 223, + 667, + 391, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{FPR} = 10^{-6}" + }, + { + "bbox": [ + 223, + 667, + 391, + 712 + ], + "type": "text", + "content": " is run beforehand, and we consider it an error if the image is not flagged." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 410, + 536, + 473, + 607 + ], + "blocks": [ + { + "bbox": [ + 410, + 536, + 473, + 607 + ], + "lines": [ + { + "bbox": [ + 410, + 536, + 473, + 607 + ], + "spans": [ + { + "bbox": [ + 410, + 536, + 473, + 607 + ], + "type": "image", + "image_path": "69f4398b92a022160850a5cc639b8e5692c3e672d92cc1201911c36ec04f416f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 609, + 446, + 618 + ], + "lines": [ + { + "bbox": [ + 413, + 609, + 446, + 618 + ], + "spans": [ + { + "bbox": [ + 413, + 609, + 446, + 618 + ], + "type": "text", + "content": "Crop " + }, + { + "bbox": [ + 413, + 609, + 446, + 618 + ], + "type": "inline_equation", + "content": "10\\%" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 477, + 536, + 539, + 607 + ], + "blocks": [ + { + "bbox": [ + 477, + 536, + 539, + 607 + ], + "lines": [ + { + "bbox": [ + 477, + 536, + 539, + 607 + ], + "spans": [ + { + "bbox": [ + 477, + 536, + 539, + 607 + ], + "type": "image", + "image_path": "ef05b53d6b96c8e4e8ba8cf22fd9db4c9d054f624e914ca94c32d59c690d2017.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 609, + 515, + 617 + ], + "lines": [ + { + "bbox": [ + 481, + 609, + 515, + 617 + ], + "spans": [ + { + "bbox": [ + 481, + 609, + 515, + 617 + ], + "type": "text", + "content": "Combined" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 486, + 618, + 530, + 662 + ], + "blocks": [ + { + "bbox": [ + 486, + 618, + 530, + 662 + ], + "lines": [ + { + "bbox": [ + 486, + 618, + 530, + 662 + ], + "spans": [ + { + "bbox": [ + 486, + 618, + 530, + 662 + ], + "type": "image", + "image_path": "96ddc50499967b5745d9f7384444a6efa2616b3c3d8db35831e9646b0807e0dc.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 665, + 545, + 709 + ], + "lines": [ + { + "bbox": [ + 402, + 665, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 402, + 665, + 545, + 709 + ], + "type": "text", + "content": "Figure 5. Transformations evaluated in Sec. 5 & 6. 'Combined' is made of crop " + }, + { + "bbox": [ + 402, + 665, + 545, + 709 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 402, + 665, + 545, + 709 + ], + "type": "text", + "content": ", brightness adjustment 1.5 and JPEG 80 compression." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22470" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 390, + 242 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 390, + 242 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 390, + 242 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 390, + 242 + ], + "type": "table", + "html": "
PSNR / SSIM ↑FID ↓Bit accuracy ↑ on:
NoneCropBrigh.Comb.
TasksText-to-ImageLDM [68]30.0 / 0.8919.6 (−0.3)0.990.950.970.92
Image EditionDiffEdit [13]31.2 / 0.9215.0 (−0.3)0.990.950.980.94
Inpainting - Full - Mask onlyGlide [57]31.1 / 0.9116.8 (+0.6)0.990.970.980.93
37.8 / 0.989.0 (+0.1)0.890.760.840.78
Super-ResolutionLDM [68]34.0 / 0.9411.6 (+0.0)0.980.930.960.92
WM MethodsPost generation
Dct-Dwt [14]0.14 (s/img)39.5 / 0.9719.5 (−0.4)0.860.520.510.51
SSL Watermark [25]0.45 (s/img)31.1 / 0.8620.6 (+0.7)1.000.730.930.66
FNNS [42]0.28 (s/img)32.1 / 0.9019.0 (−0.9)0.930.930.910.93
HiDDeN [108]0.11 (s/img)32.0 / 0.8819.7 (−0.2)0.990.970.990.98
Merged in generation
Stable Signature0.00 (s/img)30.0 / 0.8919.6 (−0.3)0.990.950.970.92
", + "image_path": "3b9784f55e700680f5b21d0aebf92e44d5f91161f654d1f6cf3546b5625e2066.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 394, + 70, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 394, + 70, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 394, + 70, + 547, + 247 + ], + "type": "text", + "content": "Table 1. Generation quality and comparison to post-hoc watermarking on " + }, + { + "bbox": [ + 394, + 70, + 547, + 247 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 394, + 70, + 547, + 247 + ], + "type": "text", + "content": " images and 48-bit signatures. PSNR and SSIM are computed between generations of the original and watermarked generators. For FID, we show in (color) the difference with regards to original. Post-hoc watermarking is evaluated on text-generated images. (App. B.2 gives results on more transformations, and App. A gives more details on the evaluations.) Overall, Stable Signature has minimal impact on generation quality. It has comparable robustness to post-hoc methods while being rooted in the generation itself." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 265, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 287, + 337 + ], + "type": "text", + "content": "Table 1 presents a quantitative evaluation of image generation on the different tasks. We report the FID, and the average PSNR and SSIM that are computed between the images generated by the fine-tuned LDM and the original one. The results show that no matter the task, the watermarking has very small impact on the FID of the generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 341, + 288, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 288, + 462 + ], + "type": "text", + "content": "The average PSNR is around 30 dB and SSIM around 0.9 between images generated by the original and a watermarked model. This is a bit low from a watermarking perspective, because we do not explicitly optimize for them. However, in a real world scenario, one would only have the watermarked version of the image. Therefore it is not as important to be extremely close to the original image, we just need to generate artifacts-free images. Without access to the image generated by the original LDM, it is very hard to tell whether a watermark is present or not." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 490, + 283, + 661 + ], + "blocks": [ + { + "bbox": [ + 50, + 490, + 283, + 661 + ], + "lines": [ + { + "bbox": [ + 50, + 490, + 283, + 661 + ], + "spans": [ + { + "bbox": [ + 50, + 490, + 283, + 661 + ], + "type": "image", + "image_path": "0891831bec1705ba992a3d421a68eaef704c641ffe4cdff967a39cc66271c124.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 667, + 287, + 712 + ], + "lines": [ + { + "bbox": [ + 46, + 667, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 667, + 287, + 712 + ], + "type": "text", + "content": "Figure 6. Images generated with Stable Diffusion. The PSNR is " + }, + { + "bbox": [ + 46, + 667, + 287, + 712 + ], + "type": "inline_equation", + "content": "35.4\\mathrm{dB}" + }, + { + "bbox": [ + 46, + 667, + 287, + 712 + ], + "type": "text", + "content": " in the first row and " + }, + { + "bbox": [ + 46, + 667, + 287, + 712 + ], + "type": "inline_equation", + "content": "28.6\\mathrm{dB}" + }, + { + "bbox": [ + 46, + 667, + 287, + 712 + ], + "type": "text", + "content": " in the second. Images generated with Stable Signature look natural because modified areas are located where the eye is not sensitive. More examples in App. C." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 312, + 311, + 538, + 357 + ], + "blocks": [ + { + "bbox": [ + 305, + 262, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 305, + 262, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 547, + 307 + ], + "type": "text", + "content": "Table 2. Watermark robustness on image transformations applied before decoding, details of which are available in App. A.3. We report the bit accuracy, averaged over " + }, + { + "bbox": [ + 305, + 262, + 547, + 307 + ], + "type": "inline_equation", + "content": "10 \\times 1k" + }, + { + "bbox": [ + 305, + 262, + 547, + 307 + ], + "type": "text", + "content": " images generated from COCO prompts with 10 different keys." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 311, + 538, + 357 + ], + "lines": [ + { + "bbox": [ + 312, + 311, + 538, + 357 + ], + "spans": [ + { + "bbox": [ + 312, + 311, + 538, + 357 + ], + "type": "table", + "html": "
AttackBit acc.Comb.0.92Sharpness 2.00.99
None0.99Bright. 2.00.97Med. Filter k=70.94
Crop 0.10.95Cont. 2.00.98Resize 0.70.91
JPEG 500.88Sat. 2.00.99Text overlay0.99
", + "image_path": "7d6d7890e0308fd50ade6f5ba1a692af168cdcf1087ba81d35d5d6c5a8b75efa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 369, + 436, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 436, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 436, + 380 + ], + "type": "text", + "content": "6.3. Watermark robustness" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": "We evaluate the robustness of the watermark to different image transformations applied before extraction. For each task, we generate 1k images with 10 models fine-tuned for different messages, and report the average bit accuracy in Table 1. Additionally, Table 2 reports results on more image transformations for images generated from COCO prompts. The main transformations are presented in Fig. 5 (more evaluation details are available in App. A.3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 483, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 590 + ], + "type": "text", + "content": "We see that the watermark is indeed robust for several tasks and across transformations. The bit accuracy is always above 0.9, except for inpainting, when replacing only the masked region of the image (between " + }, + { + "bbox": [ + 304, + 483, + 545, + 590 + ], + "type": "inline_equation", + "content": "1 - 50\\%" + }, + { + "bbox": [ + 304, + 483, + 545, + 590 + ], + "type": "text", + "content": " of the image, with an average of " + }, + { + "bbox": [ + 304, + 483, + 545, + 590 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 304, + 483, + 545, + 590 + ], + "type": "text", + "content": " across masks). Besides, the bit accuracy is not perfect even without edition, mainly because there are images that are hard to watermark (e.g. the ones that are very uniform, like the background in Fig. 6) and for which the accuracy is lower." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 591, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 662 + ], + "type": "text", + "content": "Note that the robustness comes even without any transformation during the LDM fine-tuning phase: it is due to the watermark extractor. If the watermark embedding pipeline is learned to be robust against an augmentation, then the LDM learns how to produce watermarks that are robust against it during fine-tuning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 670, + 511, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 511, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 511, + 685 + ], + "type": "text", + "content": "6.4. Comparison to post-hoc watermarking" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": "An alternative way to watermark generated images is to process them after the generation (post-hoc). This may" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22471" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 80, + 273, + 119 + ], + "blocks": [ + { + "bbox": [ + 64, + 70, + 270, + 80 + ], + "lines": [ + { + "bbox": [ + 64, + 70, + 270, + 80 + ], + "spans": [ + { + "bbox": [ + 64, + 70, + 270, + 80 + ], + "type": "text", + "content": "Table 3. Quality-robustness trade-off during fine-tuning." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 80, + 273, + 119 + ], + "lines": [ + { + "bbox": [ + 61, + 80, + 273, + 119 + ], + "spans": [ + { + "bbox": [ + 61, + 80, + 273, + 119 + ], + "type": "table", + "html": "
λi for fine-tuning0.80.40.20.10.050.025
PSNR ↑31.430.629.728.526.824.6
Bit acc. ↑ on ‘comb.’0.850.880.900.920.940.95
", + "image_path": "7812f07a1b7190e77daa4f222089cf561ef04d226d483963dd974fab8df327aa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 140, + 279, + 192 + ], + "blocks": [ + { + "bbox": [ + 61, + 130, + 272, + 140 + ], + "lines": [ + { + "bbox": [ + 61, + 130, + 272, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 130, + 272, + 140 + ], + "type": "text", + "content": "Table 4. Role of the attack simulation layer at pre-training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 140, + 279, + 192 + ], + "lines": [ + { + "bbox": [ + 56, + 140, + 279, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 279, + 192 + ], + "type": "table", + "html": "
Seen at W trainingBit accuracy ↑ at test time:
Crop 0.1Rot. 90JPEG 50Bright. 2.0Res. 0.7
X1.000.560.500.990.48
✓1.000.990.900.990.91
", + "image_path": "49beb7fda33255bc5b76b0d1e6649b632c7995855db434535282fb1b9471cd7d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 204, + 286, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 286, + 312 + ], + "type": "text", + "content": "be simpler, but less secure and efficient than Stable Signature. We compare our method to a frequency based method, DCT-DWT [14], iterative approaches (SSL Watermark [25] and FNNS [42]), and an encoder/decoder one like HiD-DeN [108]. We choose DCT-DWT since it is employed by the original open source release of Stable Diffusion [70], and the other methods because of their performance and their ability to handle arbitrary image sizes and number of bits. We use our implementations (see details in App. A.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 312, + 286, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 286, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 286, + 444 + ], + "type": "text", + "content": "Table 1 compares the generation quality and the robustness over " + }, + { + "bbox": [ + 46, + 312, + 286, + 444 + ], + "type": "inline_equation", + "content": "5\\mathrm{k}" + }, + { + "bbox": [ + 46, + 312, + 286, + 444 + ], + "type": "text", + "content": " generated images. Overall, Stable Signature achieves comparable results in terms of robustness. HiD-DeN's performance is a bit higher but its output bits are not i.i.d. meaning that it cannot be used with the same guarantees as the other methods. We also observe that post-hoc generation gives worse qualitative results, images tend to present artifacts (see Fig. 13 in the supplement). One explanation is that Stable Signature is merged into the high-quality generation process with the LDM auto-encoder model, which is able to modify images in a more subtle way." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 452, + 272, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 452, + 272, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 452, + 272, + 464 + ], + "type": "text", + "content": "6.5. Can we trade image quality for robustness?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "type": "text", + "content": "We can choose to maximize the image quality or the robustness of the watermark thanks to the weight " + }, + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{i}}" + }, + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "type": "text", + "content": " of the perceptual loss in (4). We report the average PSNR of 1k generated images, as well as the bit accuracy obtained on the extracted message for the 'Combined' editing applied before detection (qualitative results are in App. B.1). A higher " + }, + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{i}}" + }, + { + "bbox": [ + 46, + 471, + 286, + 567 + ], + "type": "text", + "content": " leads to an image closer to the original one, but to lower bit accuracies on the extracted message, see Table 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 574, + 180, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 180, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 180, + 588 + ], + "type": "text", + "content": "6.6. Attack simulation layer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Watermark robustness against image transformations depends solely on the watermark extractor. Here, we pretrain them with or without specific transformations in the simulation layer, on a shorter schedule of 50 epochs, with " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " images and 16-bits messages. From there, we plug them in the LDM fine-tuning stage and we generate 1k images from text prompts. We report the bit accuracy of the extracted watermarks in Table 4. The extractor is naturally robust to some transformations, such as crops or brightness, without being trained with them, while others, like rotations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "or JPEG, require simulation during training for the watermark to be recovered at test time. Empirically we observed that adding a transformation improves results for the latter, but makes training more challenging." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 133, + 539, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 539, + 146 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 539, + 146 + ], + "type": "text", + "content": "7. Attacks on Stable Signature's Watermarks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 153, + 545, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 153, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 304, + 153, + 545, + 261 + ], + "type": "text", + "content": "We examine the watermark's resistance to intentional tampering, as opposed to distortions that happen without bad intentions like crops or compression (discussed in Sec. 5). We consider two threat models: one is typical for many image watermarking methods [14] and operates at the image level, and another targets the generative model level. For image-level attacks, we evaluate on " + }, + { + "bbox": [ + 304, + 153, + 545, + 261 + ], + "type": "inline_equation", + "content": "5\\mathrm{k}" + }, + { + "bbox": [ + 304, + 153, + 545, + 261 + ], + "type": "text", + "content": " images generated from COCO prompts. Full details on the following experiments can be found in Appendix A.5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 270, + 419, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 419, + 283 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 419, + 283 + ], + "type": "text", + "content": "7.1. Image-level attacks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 289, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 469 + ], + "type": "text", + "content": "Watermark removal. Bob alters the image to remove the watermark with deep learning techniques, like methods used for adversarial purification [78, 94] or neural auto-encoders [1, 48]. Note that this kind of attacks has not been explored in the image watermarking literature to our knowledge. Figure 7 evaluates the robustness of the watermark against neural auto-encoders [4, 11, 20, 68] at different compression rates. To reduce the bit accuracy closer to random (50%), the image distortion needs to be strong (PSNR<26). However, assuming the attack is informed on the generative model, i.e. the auto-encoder is the same as the one used to generate the images, the attack becomes much more effective. It erases the watermark while achieving high quality (PSNR>29). This is because the image is modified precisely in the bandwidth where the watermark" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 492, + 535, + 628 + ], + "blocks": [ + { + "bbox": [ + 307, + 492, + 535, + 628 + ], + "lines": [ + { + "bbox": [ + 307, + 492, + 535, + 628 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 535, + 628 + ], + "type": "image", + "image_path": "85e7692f5251de14abbe00d51d8887092e0d0405f0b822b74e9bab6f8d30f8df.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 634, + 545, + 712 + ], + "lines": [ + { + "bbox": [ + 304, + 634, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 634, + 545, + 712 + ], + "type": "text", + "content": "Figure 7. Removal attacks. " + }, + { + "bbox": [ + 304, + 634, + 545, + 712 + ], + "type": "inline_equation", + "content": "x_{o}" + }, + { + "bbox": [ + 304, + 634, + 545, + 712 + ], + "type": "text", + "content": " is the image produced by the original generator, " + }, + { + "bbox": [ + 304, + 634, + 545, + 712 + ], + "type": "inline_equation", + "content": "x_{r}" + }, + { + "bbox": [ + 304, + 634, + 545, + 712 + ], + "type": "text", + "content": " is the version produced by the watermarked generator and then attacked. Bit accuracy is on the watermark extracted from " + }, + { + "bbox": [ + 304, + 634, + 545, + 712 + ], + "type": "inline_equation", + "content": "x_{r}" + }, + { + "bbox": [ + 304, + 634, + 545, + 712 + ], + "type": "text", + "content": ". Neural auto-encoders [4, 11, 20] follow the same trend, except for the one used by LDM ('KL-f8' for our LDM). When access to the watermark extractor is granted, adversarial attacks also remove the watermark at lower PSNR budget." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22472" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "is embedded. Note that this assumption is strong, because Alice does not need to distribute the original generator." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 106, + 287, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 106, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 106, + 287, + 226 + ], + "type": "text", + "content": "Watermark removal & embedding (white-box). To go further, we assume that the attack is informed on the watermark extractor - e.g. because it has leaked. Bob can use an adversarial attack to remove the watermark by optimizing the image under a PSNR constraint. The objective is to minimize the " + }, + { + "bbox": [ + 46, + 106, + 287, + 226 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 106, + 287, + 226 + ], + "type": "text", + "content": " distance between a random binary message sampled beforehand and the extractor's output, effectively replacing the original signature with a random one. It makes it possible to erase the watermark with a lower distortion budget, as seen in Fig. 7." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 227, + 287, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 287, + 286 + ], + "type": "text", + "content": "Instead of removing the watermark, an attacker could embed a signature into vanilla images (unauthorized embedding [14]) to impersonate another Bob of whom they have a generated image. It highlights the importance of keeping the watermark extractor private." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 295, + 172, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 295, + 172, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 172, + 306 + ], + "type": "text", + "content": "7.2. Network-level attacks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 313, + 287, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 287, + 385 + ], + "type": "text", + "content": "Model purification. Bob gets Alice's generative model and uses a fine-tuning process akin to Sec. 4.2 to eliminate the watermark embedding – that we coin model purification. This involves removing the message loss " + }, + { + "bbox": [ + 46, + 313, + 287, + 385 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{m}}" + }, + { + "bbox": [ + 46, + 313, + 287, + 385 + ], + "type": "text", + "content": ", and shifting the focus to the perceptual loss " + }, + { + "bbox": [ + 46, + 313, + 287, + 385 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{i}}" + }, + { + "bbox": [ + 46, + 313, + 287, + 385 + ], + "type": "text", + "content": " between the original image and the one reconstructed by the LDM auto-encoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 386, + 287, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 457 + ], + "type": "text", + "content": "Figure 8 shows the results of this attack for the MSE loss. The PSNR between the watermarked and purified images is plotted at various stages of fine-tuning. Empirically, it is difficult to significantly reduce the bit accuracy without compromising the image quality: artifacts start to appear during the purification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "content": "Model collusion. Users may collude by aggregating their models. For instance, " + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{Bob}^{(i)}" + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{Bob}^{(j)}" + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "content": " can average the weights of their models (like Model soups [91]) creating a new model to deceive identification. We found that the bit at position " + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "content": " output by the extractor will be 0 (resp. 1) when the " + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "content": "-th bits of " + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{Bob}^{(i)}" + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{Bob}^{(j)}" + }, + { + "bbox": [ + 46, + 468, + 287, + 552 + ], + "type": "text", + "content": " are both 0 (resp. 1), and that the extracted bit is random when their bits disagree. We" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 65, + 574, + 270, + 671 + ], + "blocks": [ + { + "bbox": [ + 65, + 574, + 270, + 671 + ], + "lines": [ + { + "bbox": [ + 65, + 574, + 270, + 671 + ], + "spans": [ + { + "bbox": [ + 65, + 574, + 270, + 671 + ], + "type": "image", + "image_path": "fab0b6e04f8cc204629f9b8d0d333db1a6699340d819ef53d36ba1c99b0578a3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 678, + 287, + 712 + ], + "lines": [ + { + "bbox": [ + 46, + 678, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 678, + 287, + 712 + ], + "type": "text", + "content": "Figure 8. Robustness to model purification, i.e. fine-tuning the model to remove watermarks. " + }, + { + "bbox": [ + 46, + 678, + 287, + 712 + ], + "type": "inline_equation", + "content": "x_w" + }, + { + "bbox": [ + 46, + 678, + 287, + 712 + ], + "type": "text", + "content": " is the watermarked image, " + }, + { + "bbox": [ + 46, + 678, + 287, + 712 + ], + "type": "inline_equation", + "content": "x_r" + }, + { + "bbox": [ + 46, + 678, + 287, + 712 + ], + "type": "text", + "content": " is generated with the purified model at different steps of the process." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "show the distributions of the soft bits (before thresholding) output by the watermark extractor on images generated by the average model. The " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "-th output is labeled by bits of " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathrm{Bob}^{(i)}" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathrm{Bob}^{(j)}" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "00" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " means both have " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "0" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " at position " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 128, + 541, + 178 + ], + "blocks": [ + { + "bbox": [ + 313, + 128, + 541, + 178 + ], + "lines": [ + { + "bbox": [ + 313, + 128, + 541, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 541, + 178 + ], + "type": "image", + "image_path": "34f4af48989a935f75faa92f0873327c1aff8b0f6e920593dfb04870fb017f62.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 189, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 189, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 189, + 545, + 285 + ], + "type": "text", + "content": "This so-called marking assumption plays a crucial role in traitor tracing literature [27, 55, 83]. Surprisingly, it holds even though our watermarking process is not explicitly designed for it. The study has room for improvement, such as creating user identifiers with more powerful traitor tracing codes [83] and using more powerful traitor accusation algorithms [27, 55]. Importantly, we found the precedent remarks also hold if the colluders operate at the image level." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 296, + 448, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 296, + 448, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 296, + 448, + 308 + ], + "type": "text", + "content": "8. Conclusion & Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 316, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 400 + ], + "type": "text", + "content": "By a quick fine-tuning of the decoder of Latent Diffusion Models, we can embed watermarks in all the images they generate. This does not alter the diffusion process, making it compatible with most of LDM-based generative models. These watermarks are robust, invisible to the human eye and can be employed to detect generated images and identify the user that generated it, with very high performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 400, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 400, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 400, + 545, + 472 + ], + "type": "text", + "content": "The public release of image generative models has an important societal impact. With this work, we put to light the usefulness of using watermarking instead of relying on passive detection methods. We hope it will encourage researchers and practitioners to employ similar approaches before making their models publicly available." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 479, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 545, + 552 + ], + "type": "text", + "content": "Reproducibility Statement. Although the diffusion-based generative model has been trained on an internal dataset of licensed images, we use the KL auto-encoder from LDM [68] with compression factor " + }, + { + "bbox": [ + 305, + 479, + 545, + 552 + ], + "type": "inline_equation", + "content": "f = 8" + }, + { + "bbox": [ + 305, + 479, + 545, + 552 + ], + "type": "text", + "content": ". This is the one used by open-source alternatives. Code is available at github.com/facebookresearch/stable_signature." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 558, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 558, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 558, + 546, + 712 + ], + "type": "text", + "content": "Environmental Impact. We do not expect any environmental impact specific from this work. The cost of the experiments and the method is high, though order of magnitudes less than other computer vision fields. We roughly estimated that the total GPU-days used for running all our experiments to 2000, or " + }, + { + "bbox": [ + 305, + 558, + 546, + 712 + ], + "type": "inline_equation", + "content": "\\approx" + }, + { + "bbox": [ + 305, + 558, + 546, + 712 + ], + "type": "text", + "content": " 50000 GPU-hours. This amounts to total emissions in the order of 10 tons of " + }, + { + "bbox": [ + 305, + 558, + 546, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{CO}_{2}\\mathrm{eq}" + }, + { + "bbox": [ + 305, + 558, + 546, + 712 + ], + "type": "text", + "content": ". This is excluding the training of the generative model itself, since we did not perform that training. Estimations are conducted using the Machine Learning Impact calculator presented by Lacoste et al. [43]. We do not consider in this approximation: memory storage, CPU-hours, production cost of GPUs/ CPUs, etc." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22473" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Sahar Abdelnabi and Mario Fritz. Adversarial watermarking transformer: Towards tracing text provenance with data hiding. In 2021 IEEE Symposium on Security and Privacy (SP), pages 121-140. IEEE, 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 136, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 136, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 136, + 287, + 178 + ], + "type": "text", + "content": "[2] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 180, + 287, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 180, + 287, + 234 + ], + "spans": [ + { + "bbox": [ + 58, + 180, + 287, + 234 + ], + "type": "text", + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 236, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 236, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 58, + 236, + 287, + 278 + ], + "type": "text", + "content": "[4] Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 7, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 280, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 280, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 58, + 280, + 287, + 312 + ], + "type": "text", + "content": "[5] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 17" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 314, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 314, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 58, + 314, + 287, + 357 + ], + "type": "text", + "content": "[6] Jean Bégaint, Fabien Racapé, Simon Feltman, and Akshay Pushparaja. Compressai: a pytorch library and evaluation platform for end-to-end compression research. arXiv preprint arXiv:2011.03029, 2020. 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 358, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 358, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 58, + 358, + 287, + 390 + ], + "type": "text", + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 392, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 392, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 58, + 392, + 287, + 445 + ], + "type": "text", + "content": "[8] Miles Brundage, Shahar Avin, Jack Clark, Helen Toner, Peter Eckersley, Ben Garfinkel, Allan Dafoe, Paul Scharre, Thomas Zeitzoff, Bobby Filar, et al. The malicious use of artificial intelligence: Forecasting, prevention, and mitigation. arXiv preprint arXiv:1802.07228, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 447, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 447, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 58, + 447, + 287, + 490 + ], + "type": "text", + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV. IEEE, 2021. 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 491, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 535 + ], + "type": "text", + "content": "[10] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In European conference on computer vision, pages 103-120. Springer, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 536, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 287, + 590 + ], + "type": "text", + "content": "[11] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7939-7948, 2020. 7, 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 591, + 287, + 635 + ], + "type": "text", + "content": "[12] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. arXiv preprint arXiv:2211.00680, 2022. 2, 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 53, + 636, + 287, + 679 + ], + "type": "text", + "content": "[13] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 1, 2, 6, 13, 16" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "type": "text", + "content": "[14] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital watermarking and steganography. Morgan Kaufmann, 2007. 2, 6, 7, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Steffen Czolbe, Oswin Krause, Ingemar Cox, and Christian Igel. A loss function for generative neural networks based on watson's perceptual model. Advances in Neural Information Processing Systems, 33:2051-2061, 2020. 4, 15" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 119, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 119, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 312, + 119, + 545, + 163 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 165, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 165, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 312, + 165, + 545, + 186 + ], + "type": "text", + "content": "[17] Emily Denton. Ethical considerations of generative ai. In AI for Content Creation Workshop CVPR. IEEE, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 189, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 189, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 312, + 189, + 545, + 220 + ], + "type": "text", + "content": "[18] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 223, + 545, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 223, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 312, + 223, + 545, + 277 + ], + "type": "text", + "content": "[19] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34:19822-19835, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 280, + 545, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 280, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 312, + 280, + 545, + 332 + ], + "type": "text", + "content": "[20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 7, 14" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 336, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 336, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 312, + 336, + 545, + 358 + ], + "type": "text", + "content": "[21] Hany Farid. Lighting (in) consistency of paint by text. arXiv preprint arXiv:2207.13744, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 361, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 361, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 312, + 361, + 545, + 381 + ], + "type": "text", + "content": "[22] Hany Farid. Perspective (in) consistency of paint by text. arXiv preprint arXiv:2206.14617, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 384, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 384, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 312, + 384, + 545, + 437 + ], + "type": "text", + "content": "[23] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Supervised gan watermarking for intellectual property protection. In 2022 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6. IEEE, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 441, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 441, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 312, + 441, + 545, + 472 + ], + "type": "text", + "content": "[24] Pierre Fernandez, Matthijs Douze, Hervé Jégou, and Teddy Furon. Active image indexing. arXiv preprint arXiv:2210.10620, 2022. 14" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 475, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 475, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 312, + 475, + 545, + 529 + ], + "type": "text", + "content": "[25] Pierre Fernandez, Alexandre Sablayrolles, Teddy Furon, Hervé Jégou, and Matthijs Douze. Watermarking images in self-supervised latent spaces. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022. 2, 6, 7, 14, 15" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 533, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 545, + 586 + ], + "type": "text", + "content": "[26] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 589, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 589, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 312, + 589, + 545, + 632 + ], + "type": "text", + "content": "[27] Teddy Furon, Arnaud Guyader, and Frédéric Cérou. Decoding Fingerprinting Using the Markov Chain Monte Carlo Method. In WIFS - IEEE Workshop on Information Forensics and Security, Tenerife, Spain, Dec. 2012. IEEE. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 545, + 689 + ], + "type": "text", + "content": "[28] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "text", + "content": "[29] Matthew Gault. An ai-generated artwork won first place at a state fair fine arts competition, and artists are pissed. Vice," + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22474" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 72, + 286, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 286, + 115 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 286, + 115 + ], + "type": "text", + "content": "2022. https://www.vice.com/en/article/bvmvqm/an-ai-generated-artwork-won-first-place-at-a-state-fair-fine-arts-competition-and-artists-are-pissed (retrieved 2022-12-14). 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 286, + 150 + ], + "type": "text", + "content": "[30] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. *ICLR*, 2014. 14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 152, + 286, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 152, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 53, + 152, + 286, + 207 + ], + "type": "text", + "content": "[31] Diego Gragnaniello, Davide Cozzolino, Francesco Marra, Giovanni Poggi, and Luisa Verdoliva. Are gan generated images easy to detect? a critical analysis of the state-of-the-art. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 209, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 286, + 262 + ], + "type": "text", + "content": "[32] Luca Guarnera, Oliver Giudice, and Sebastiano Battiato. Deepfake detection by analyzing convolutional traces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 666-667, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 264, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 286, + 308 + ], + "type": "text", + "content": "[33] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 309, + 286, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 309, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 309, + 286, + 363 + ], + "type": "text", + "content": "[34] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 365, + 286, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 286, + 420 + ], + "type": "text", + "content": "[35] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 422, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 422, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 422, + 286, + 453 + ], + "type": "text", + "content": "[36] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 456, + 286, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 286, + 498 + ], + "type": "text", + "content": "[37] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in Neural Information Processing Systems, 33:12104-12114, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 501, + 286, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 501, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 501, + 286, + 544 + ], + "type": "text", + "content": "[38] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 545, + 286, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 286, + 600 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 286, + 600 + ], + "type": "text", + "content": "[39] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 601, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 286, + 644 + ], + "type": "text", + "content": "[40] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 647, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 286, + 667 + ], + "type": "text", + "content": "[41] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 15" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 286, + 712 + ], + "type": "text", + "content": "[42] Varsha Kishore, Xiangyu Chen, Yan Wang, Boyi Li, and Kilian Q Weinberger. Fixed neural network steganography: Train the images, not the network. In International Conference on Learning Representations, 2022. 2, 6, 7, 14" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 115 + ], + "type": "text", + "content": "[43] Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, and Thomas Dandres. Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700, 2019. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 161 + ], + "type": "text", + "content": "[44] Jae-Eun Lee, Young-Ho Seo, and Dong-Wook Kim. Convolutional neural network-based digital image watermarking adaptive to the resolution of image and watermark. Applied Sciences, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 163, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 545, + 195 + ], + "type": "text", + "content": "[45] Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. arXiv preprint arXiv:1811.00656, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 198, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 198, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 312, + 198, + 545, + 239 + ], + "type": "text", + "content": "[46] Dongdong Lin, Benedetta Tondi, Bin Li, and Mauro Barni. Cycleganwm: A cyclegan watermarking method for ownership verification. arXiv preprint arXiv:2211.13737, 2022. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 243, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 243, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 312, + 243, + 545, + 297 + ], + "type": "text", + "content": "[47] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4, 5, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 298, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 298, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 312, + 298, + 545, + 342 + ], + "type": "text", + "content": "[48] Wenqing Liu, Miaojing Shi, Teddy Furon, and Li Li. Defending adversarial examples via dnn bottleneck reinforcement. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1930-1938, 2020. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 343, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 343, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 312, + 343, + 545, + 364 + ], + "type": "text", + "content": "[49] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2018. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 366, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 366, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 312, + 366, + 545, + 430 + ], + "type": "text", + "content": "[50] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 433, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 433, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 312, + 433, + 545, + 464 + ], + "type": "text", + "content": "[51] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In CVPR. IEEE, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 468, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 468, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 312, + 468, + 545, + 520 + ], + "type": "text", + "content": "[52] Jingwei Ma, Lucy Chai, Minyoung Huh, Tongzhou Wang, Ser-Nam Lim, Phillip Isola, and Antonio Torralba. Totems: Physical objects for verifying visual integrity. In European Conference on Computer Vision, pages 164-180. Springer, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 523, + 545, + 555 + ], + "type": "text", + "content": "[53] Sébastien Marcel and Yann Rodriguez. Torchvision the machine-vision package of torch. In International Conference on Multimedia. ACM, 2010. 13" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 545, + 600 + ], + "type": "text", + "content": "[54] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 312, + 601, + 545, + 644 + ], + "type": "text", + "content": "[55] Peter Meerwald and Teddy Furon. Towards practical joint decoding of binary Tardos fingerprinting codes. IEEE Transactions on Information Forensics and Security, 7(4):1168-1180, Apr. 2012. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 647, + 545, + 689 + ], + "type": "text", + "content": "[56] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. arXiv preprint arXiv:2211.09794, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 692, + 545, + 712 + ], + "type": "text", + "content": "[57] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22475" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 106 + ], + "type": "text", + "content": "Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 6, 13, 16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 288, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 288, + 149 + ], + "type": "text", + "content": "[58] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 150, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 288, + 183 + ], + "type": "text", + "content": "[59] Guangyu Nie, Changhoon Kim, Yezhou Yang, and Yi Ren. Attributing image generative models using latent fingerprints. arXiv preprint arXiv:2304.09752, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 184, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 288, + 216 + ], + "type": "text", + "content": "[60] Zoe Papakipos and Joanna Bitton. Augly: Data augmentations for robustness. arXiv preprint arXiv:2201.06494, 2022. 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 281 + ], + "type": "text", + "content": "[61] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in Neural Information Processing Systems 32, 2019. 16" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 283, + 288, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 288, + 314 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 288, + 314 + ], + "type": "text", + "content": "[62] William Peebles and Saining Xie. Scalable diffusion models with transformers. arXiv preprint arXiv:2212.09748, 2022.2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 316, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 316, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 316, + 288, + 360 + ], + "type": "text", + "content": "[63] Tong Qiao, Yuyan Ma, Ning Zheng, Hanzhou Wu, Yanli Chen, Ming Xu, and Xiangyang Luo. A novel model watermarking for protecting generative adversarial network. Computers & Security, 127:103102, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 360, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 360, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 360, + 288, + 403 + ], + "type": "text", + "content": "[64] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 404, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 288, + 448 + ], + "type": "text", + "content": "[65] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 449, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 449, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 449, + 288, + 502 + ], + "type": "text", + "content": "[66] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 13" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 504, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 504, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 53, + 504, + 288, + 536 + ], + "type": "text", + "content": "[67] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 537, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 537, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 537, + 288, + 601 + ], + "type": "text", + "content": "[68] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2, 4, 6, 7, 8, 14, 16" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 602, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 602, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 53, + 602, + 288, + 658 + ], + "type": "text", + "content": "[69] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 658, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 288, + 690 + ], + "type": "text", + "content": "[70] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. Stable diffusion. Accessed Nov. 30 2022, 2022. 1, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 691, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 288, + 712 + ], + "type": "text", + "content": "[71] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 312, + 95, + 545, + 128 + ], + "type": "text", + "content": "[72] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Spreading vectors for similarity search. ICML, 2019. 17" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 545, + 183 + ], + "type": "text", + "content": "[73] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 184, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 184, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 184, + 545, + 249 + ], + "type": "text", + "content": "[74] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 250, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 250, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 312, + 250, + 545, + 303 + ], + "type": "text", + "content": "[75] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 14" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 304, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 304, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 312, + 304, + 545, + 348 + ], + "type": "text", + "content": "[76] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 349, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 349, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 312, + 349, + 545, + 392 + ], + "type": "text", + "content": "[77] Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. Defake: Detection and attribution of fake images generated by text-to-image diffusion models. arXiv preprint arXiv:2210.06998, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 393, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 393, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 312, + 393, + 545, + 426 + ], + "type": "text", + "content": "[78] Changhao Shi, Chester Holtz, and Gal Mishne. Online adversarial purification based on self-supervision. arXiv preprint arXiv:2101.09387, 2021. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 545, + 480 + ], + "type": "text", + "content": "[79] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 481, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 481, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 481, + 545, + 514 + ], + "type": "text", + "content": "[80] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 514, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 514, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 312, + 514, + 545, + 558 + ], + "type": "text", + "content": "[81] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. *ICLR*, 2021. 13" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 559, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 559, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 559, + 545, + 635 + ], + "type": "text", + "content": "[82] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2149–2159, 2022. 13, 20" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 635, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 545, + 658 + ], + "type": "text", + "content": "[83] Gábor Tardos. Optimal probabilistic fingerprint codes. Journal of the ACM (JACM), 55(2):1-24, 2008. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "type": "text", + "content": "[84] Yusuke Uchida, Yuki Nagai, Shigeyuki Sakazawa, and Shin'ichi Satoh. Embedding watermarks into deep neural networks. In Proceedings of the 2017 ACM on international conference on multimedia retrieval, pages 269-277, 2017. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22476" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 116 + ], + "type": "text", + "content": "[85] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 288, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 288, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 288, + 148 + ], + "type": "text", + "content": "[86] Vedran Vukotic, Vivien Chappelier, and Teddy Furon. Are deep neural networks good for blind image watermarking? In WIFS, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 150, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 288, + 194 + ], + "type": "text", + "content": "[87] Steven Walton, Ali Hassani, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Stylenat: Giving each head a new perspective. arXiv preprint arXiv:2211.05770, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 195, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 288, + 249 + ], + "type": "text", + "content": "[88] Sheng-Yu Wang, Oliver Wang, Andrew Owens, Richard Zhang, and Alexei A Efros. Detecting photoshopped faces by scripting photoshop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10072-10081, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 251, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 304 + ], + "type": "text", + "content": "[89] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 305, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 305, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 305, + 288, + 348 + ], + "type": "text", + "content": "[90] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 350, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 288, + 425 + ], + "type": "text", + "content": "[91] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pages 23965-23998. PMLR, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 426, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 288, + 468 + ], + "type": "text", + "content": "[92] Chen Henry Wu and Fernando De la Torre. Unifying diffusion models' latent space, with applications to cycled diffusion and guidance. arXiv preprint arXiv:2210.05559, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 471, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 471, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 53, + 471, + 288, + 514 + ], + "type": "text", + "content": "[93] Hanzhou Wu, Gen Liu, Yuwei Yao, and Xinpeng Zhang. Watermarking neural networks with watermarked images. IEEE Transactions on Circuits and Systems for Video Technology, 31(7):2591-2601, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 515, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 515, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 515, + 288, + 557 + ], + "type": "text", + "content": "[94] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, pages 12062-12072. PMLR, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 559, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 288, + 612 + ], + "type": "text", + "content": "[95] Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. Large batch optimization for deep learning: Training bert in 76 minutes. In ICLR, 2020. 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 614, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 614, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 53, + 614, + 288, + 667 + ], + "type": "text", + "content": "[96] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "text", + "content": "[97] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 586 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 127 + ], + "type": "text", + "content": "[98] Ning Yu, Vladislav Skripniuk, Sahar Abdelnabi, and Mario Fritz. Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14448-14457, 2021. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 129, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 547, + 172 + ], + "type": "text", + "content": "[99] Ning Yu, Vladislav Skripniuk, Dingfan Chen, Larry Davis, and Mario Fritz. Responsible disclosure of generative models using scalable fingerprinting. In International Conference on Learning Representations (ICLR), 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 175, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 546, + 217 + ], + "type": "text", + "content": "[100] Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun, and In So Kweon. Udh: Universal deep hiding for steganography, watermarking, and light field messaging. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 219, + 546, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 546, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 546, + 273 + ], + "type": "text", + "content": "[101] Chaoning Zhang, Adil Karjauv, Philipp Benz, and In So Kweon. Towards robust deep hiding under nondifferentiable distortions for practical blind watermarking. In Proceedings of the 29th ACM International Conference on Multimedia, pages 5158-5166, 2021. 4, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 275, + 546, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 546, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 546, + 328 + ], + "type": "text", + "content": "[102] Jie Zhang, Dongdong Chen, Jing Liao, Han Fang, Weiming Zhang, Wenbo Zhou, Hao Cui, and Nenghai Yu. Model watermarking for image processing networks. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12805-12812, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 546, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 546, + 373 + ], + "type": "text", + "content": "[103] Kevin Alex Zhang, Alfredo Cuesta-Infante, Lei Xu, and Kalyan Veeramachaneni. Steganogan: High capacity image steganography with gans. arXiv preprint arXiv:1901.03892, 2019. 14" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 375, + 546, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 546, + 407 + ], + "type": "text", + "content": "[104] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 409, + 546, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 409, + 546, + 452 + ], + "spans": [ + { + "bbox": [ + 308, + 409, + 546, + 452 + ], + "type": "text", + "content": "[105] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR. IEEE, 2018. 4, 15" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 454, + 546, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 454, + 546, + 497 + ], + "spans": [ + { + "bbox": [ + 308, + 454, + 546, + 497 + ], + "type": "text", + "content": "[106] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In 2019 IEEE international workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 499, + 546, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 499, + 546, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 499, + 546, + 552 + ], + "type": "text", + "content": "[107] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deep-fake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 554, + 546, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 546, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 546, + 586 + ], + "type": "text", + "content": "[108] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In ECCV, 2018. 2, 3, 4, 6, 7, 13" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22477" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/c596fbdc-cd07-43b2-9a08-9301d722c491_content_list.json b/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/c596fbdc-cd07-43b2-9a08-9301d722c491_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ecbf93a08101e36def40e964a3bcc1ecbbeb0d1c --- /dev/null +++ b/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/c596fbdc-cd07-43b2-9a08-9301d722c491_content_list.json @@ -0,0 +1,1470 @@ +[ + { + "type": "text", + "text": "The Unreasonable Effectiveness of Large Language-Vision Models for Source-free Video Domain Adaptation", + "text_level": 1, + "bbox": [ + 151, + 130, + 818, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Giacomo Zara1*, Alessandro Conti1*, Subhankar Roy3, Stéphane Lathuilière3, Paolo Rota1, Elisa Ricci1,2", + "bbox": [ + 84, + 202, + 913, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Trento, Italy 2Fondazione Bruno Kessler, Italy", + "bbox": [ + 256, + 220, + 743, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ LTCI, Télécom Paris, Institut polytechnique de Paris, France", + "bbox": [ + 256, + 238, + 743, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{giacomo.zara, alessandro.conti-1}@unitn.it", + "bbox": [ + 313, + 258, + 684, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Source-Free Video Unsupervised Domain Adaptation (SFVUDA) task consists in adapting an action recognition model, trained on a labelled source dataset, to an unlabelled target dataset, without accessing the actual source data. The previous approaches have attempted to address SFVUDA by leveraging self-supervision (e.g., enforcing temporal consistency) derived from the target data itself. In this work, we take an orthogonal approach by exploiting \"web-supervision\" from Large Language-Vision Models (LLVMs), driven by the rationale that LLVMs contain a rich world prior surprisingly robust to domain-shift. We showcase the unreasonable effectiveness of integrating LLVMs for SFVUDA by devising an intuitive and parameter-efficient method, which we name Domain Adaptation with Large Language-Vision models (DALL-V), that distills the world prior and complementary source model information into a student network tailored for the target. Despite the simplicity, DALL-V achieves significant improvement over state-of-the-art SFVUDA methods.", + "bbox": [ + 75, + 340, + 472, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 672, + 209, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video analysis tasks, such as action recognition, have long been investigated in computer vision, due to the numerous applications, ranging from video surveillance to social robotics [50, 14, 29]. Major progress has been made in the last decade with the development of specialized deep architectures, such as 3D CNNs [4, 3] and Video Transformers [36], trained on large-scale annotated datasets. However, obtaining sufficient labelled training videos for real-world scenarios can be very costly and time-consuming. In order to alleviate the burden of annotating large scale datasets Video-based Unsupervised Domain", + "bbox": [ + 75, + 698, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c21b766958fb0a13750f29c7ec9c41d40ab534edbe973eea39260c5ae5a5069c.jpg", + "image_caption": [ + "Figure 1: Performance over time by various methods on the Daily-DA video benchmark. A pre-trained LLVM (e.g., CLIP [31]) is surprisingly better when compared with Unsupervised Domain Adaptation (UDA), Source Free UDA (SFUDA) and Source Free Video-based UDA (SFVUDA) methods. Our proposed DALL-V that is built on top of LLVM successfully outperforms all existing baselines." + ], + "image_footnote": [], + "bbox": [ + 509, + 314, + 883, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adaptation (VUDA) [1, 28, 2] have been introduced. The VUDA methods are derived from the common idea of transferring knowledge from a labelled source domain to an unlabelled target domain. While considering different strategies for adaptation, these approaches have all shown significant improvements in the robustness of learnt video models without requiring annotated target data.", + "bbox": [ + 496, + 627, + 893, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the last few years, the field of computer vision has also witnessed the emergence of a new generation of powerful deep architectures, trained on mammoth internet-scale image-text datasets [35]. These models, commonly known as foundation models [32, 37, 31, 10] or Large LanguageVision Models (LLVM), have achieved outstanding performance, and have become a cornerstone of modern computer vision research. In particular, recent works such as CLIP [31] or FLAVA [37] have shown that rich visual representations for images can be learned from natural language descriptions as a form of supervision. These pre-trained", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Giacomo Zara and Alessandro Conti contributed equally.", + "bbox": [ + 93, + 875, + 403, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code is available at https://github.com/giaczara/dallyv", + "bbox": [ + 93, + 887, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10307", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a6b9e8506112f283489cecbbdade7cbabf065c58eae3fd1a31eb2818d60f4e4.jpg", + "image_caption": [ + "(a) Traditional SFVUDA" + ], + "image_footnote": [], + "bbox": [ + 181, + 88, + 297, + 300 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd153a23f025253233b9242e0ce2ee4482b070514fc0f11da8ff0c03b29b8809.jpg", + "image_caption": [ + "(b) Zero Shot CLIP Inference" + ], + "image_footnote": [], + "bbox": [ + 321, + 87, + 545, + 300 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8fdc00db30f2dedfdbeb9b7fd781eea1fdbf3d8700cbc0336d5646e609037dfd.jpg", + "image_caption": [ + "(c) Our proposed SFVUDA", + "Figure 2: (a) In traditional SFVUDA a source-trained model is adapted to the unlabelled target dataset using self-supervision such as temporal consistency [44, 46] among frames. (b) LLVM (e.g., CLIP [31]) uses cosine similarity between the feature and language representation to predict the most probable class in a zero-shot (ZS) manner. (c) Our proposed SFVUDA solution distills the ZS-CLIP and source model predictions, while introducing very few learnable parameters, namely adapters. The $\\hat{\\mathbb{C}}$ means the network is trainable, while the $\\hat{\\mathbb{C}}$ means the network is frozen." + ], + "image_footnote": [], + "bbox": [ + 576, + 87, + 810, + 300 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLVMs are now publicly available and can be easily integrated into any recognition system. Lately, researchers have also shown that the LLVMs can be applied with success to the video domain and specifically to the supervised action recognition task [40, 41].", + "bbox": [ + 75, + 431, + 468, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the significant progress in the VUDA literature, to date, the role of LLVM in the context of action recognition has been overlooked. We argue that the LLVMs have a lot to offer to the progress of VUDA methods, which is currently untapped. Our work starts from a preliminary analysis demonstrating that current sophisticated VUDA methods can be largely outperformed by off-the-shelf publicly available LLVMs like CLIP, without needing explicit adaptation (see Fig. 1). This observation highlights the need for a paradigm shift among VUDA methods. The benefit of LLVM models is so significant that it raises a pertinent question: If explicit alignment methodologies, traditionally pursued in VUDA, is truly the way forward? Instead, the primary research question becomes: How can we efficiently integrate the prior knowledge derived from LLVMs to adapt a model to the target domain for VUDA?", + "bbox": [ + 75, + 507, + 468, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper presents the first approach in the VUDA literature, which attempts to address this question specifically. In particular, we consider the challenging scenario of Source-Free Video Unsupervised Domain Adaptation (SFVUDA) [44, 46], which consists in the task of adapting an action recognition model trained on a labelled source domain to an unlabelled target domain, without accessing the actual source data. The primary motivation for exploiting LLVMs in SFVUDA is that we expect the generalization capabilities of these models to effectively serve the pur", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pose of mitigating the effects of the domain shift existing between the data distributions of the source and the target domain. In particular, since in SFVUDA we do not have access to the source data but only the source pre-trained model, direct target adaptation may potentially lead to significantly poor performance, especially when the domain shift is large. On the contrary, the employment of LLVMs can efficiently counteract this negative effect owing to the wide range of visual domains observed during their training process. An additional motivation behind our approach lies in the fact that the rich visual representations derived from LLVMs can efficiently compensate when additional modalities (e.g., optical-flow), which are known to help reduce domain gap in VUDA frameworks [27, 33], are unavailable.", + "bbox": [ + 496, + 431, + 890, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach, which we name as Domain Adaptation with Large Language-Vision models (or DALL-V in short), is a simple yet effective technique to combine the knowledge from the visual representation of a pre-trained LLVM with that derived from the source model and the target data (see Fig. 2c). DALL-V involves a two-stage pipeline: (i) in the first stage pseudo-labels are extracted from the CLIP model and are subsequently used for adapting on the target dataset, and (ii) in the second stage an ensemble of CLIP, source and target models are used to distill information into a student network. DALL-V introduces very few trainable parameters on top of CLIP and is realized with the help of domain-specific adapters. Despite its simplicity, our approach outperforms SFVUDA and even VUDA methods by a significant margin (+11.8% w.r.t. the best competitor).", + "bbox": [ + 496, + 643, + 892, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are: (i) We show, for the first time in the literature that straightforward zero-shot", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10308", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "methods based on LLVMs vastly outperform the state-of-the-art approaches for SFVUDA. (ii) Building upon this observation, we propose DALL-V, a simple approach for SFVUDA which optimally integrates information derived from LLVMs, from a pretrained source model and from unlabelled videos of the target domain. (iii) We perform extensive experiments on a total of 20 domain adaptation settings, demonstrating the effectiveness of DALL-V in SFVUDA.", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 224, + 218, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Unsupervised Domain Adaptation. While domain adaptation techniques have been primarily studied in the context of image-level representations and fall under the umbrella of Unsupervised Domain Adaptation (UDA) methods [7, 47, 22, 39, 1, 42], addressing the domain shift problem is even more challenging in the case of videos, due to the additional temporal dimension. Researchers have proposed different strategies. For instance, Chen et al. introduced $\\mathrm{TA}^3\\mathrm{N}$ [1], a discrepancy-based method that aligns domains on the temporal axis by learning a temporal relationship across video sequences. TCoN was introduced in [28], a deep architecture which employs a cross-domain module to compute temporally-aligned source and target feature representations. CO2A [2] proposed an approach based on contrastive learning to align source and target video representations. One major limitation with both UDA and VUDA methods is the fact that they require access to source data. This limits their applicability in real-world scenarios where data may not be accessible due to privacy reasons. Differently, our approach focuses in the more realistic setting where we only have access to the source model.", + "bbox": [ + 75, + 250, + 468, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Source-Free Video Unsupervised Domain Adaptation. Due to the growing concerns related to privacy and data sharing, recently researchers have begun exploring methods Source-Free Unsupervised Domain Adaptation (SFUDA). One of the first works in this direction is SHOT [20], a deep architecture which employs an entropy loss and a classification loss on pseudo-labelled data to adapt the source pretrained network focusing solely on the target data. In a subsequent work [21], the authors introduced an additional auxiliary head to solve the relative rotation task, further improving prediction accuracy. Other works, such as 3C-GAN [18], SFIT [9], and SDDA [16], framed the problem of SFUDA as an image translation task.", + "bbox": [ + 75, + 568, + 468, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the literature, only a handful of works addressed the problem of Source-Free Video Unsupervised Domain Adaptation (SFVUDA). In particular, ATCoN [44] presented an approach which models temporal consistency across the video sequences. EXTERN [46] proposed to exploit mask-to-mix strategies and video-tailored regularizations for SFVUDA. Our approach radically departs from these previous works as we proposed to exploit LLVMs to adapt to the target video data.", + "bbox": [ + 75, + 765, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Language-Vision Models. The availability of vast web-scale datasets containing image-text pairs [35, 34, 37] have enabled the emergence of novel large multi-modal neural networks which learn joint visual-text embedding spaces [31, 10, 37]. These approaches, commonly referred as Large Language-Vision Models, utilise a separate encoder for each modality and employ a contrastive loss to align the data representations in the feature space [31, 10]. CLIP [31] is a prominent example of such an approach. Despite its simplicity, CLIP have been shown to achieve impressive zero-shot image recognition capabilities. More recently, LLVMs have been also extended with success to the video domains and particularly to the action recognition task [40]. However, we are not aware of previous works which have specifically used LLVMs to address the problem of domain shift in video action recognition.", + "bbox": [ + 496, + 90, + 890, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Unreasonable Effectiveness of LLVM", + "text_level": 1, + "bbox": [ + 498, + 345, + 872, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work our goal is to develop a Source-free Video Unsupervised Domain Adaptation (SFVUDA) method that can adapt a source trained model to a target domain of interest. Departing from the traditional SFVUDA approaches, where a multitude of self-supervised losses are optimized on the target domain [44], we pursue an orthogonal and unconventional approach. Our key idea is to leverage a LLVM (e.g., CLIP [31]), which has been pre-trained on web-scale image-text pairs, as a tool for bridging the domain gap.", + "bbox": [ + 496, + 372, + 890, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a part of a preliminary study we evaluate CLIP in a zero-shot manner (see Fig. 2b) on the target dataset. In other words, we simply run inference and evaluate the performance on the Daily-DA [44] benchmark using a pre-trained CLIP (see Sec. 4.2 for details on zero-shot inference), and compare with the recent state-of-the-art SFVUDA approaches. The evaluation is carried out on the four target domains of the Daily-DA, each of which has eight semantic categories. As showcased in Fig. 1, much to our surprise, CLIP (ViT-B/32) outperforms the best performing SFVUDA method EXTERN [46] by a huge margin of $+8.5\\%$ , despite CLIP never been explicitly fine-tuned on the target frames, let alone videos. These telling observations from our preliminary study compels us to explore and exploit the unreasonable effectiveness of the LLVMs, which is orthogonal to the existing SFVUDA literature.", + "bbox": [ + 496, + 508, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The implications of undertaking such a pathway to bridge the domain gap in SFVUDA can be several: (i) the LLVMs being publicly available, it will help democratize more efficient target adaptation, even if the source data is withheld due to privacy reasons, (ii) it disposes off the need to balance and tune complex training objectives, a de facto practice in SFVUDA (e.g., ATCoN [44] requires to jointly optimize up to 6 loss functions), (iii) it prevents the necessity of ad hoc and specialized network architectures and training objectives to solve the task at hand. Encouraged by", + "bbox": [ + 496, + 750, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10309", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the flexibility offered by LLVMs, we propose DALL-V for SFVUDA that can effectively exploit the world prior from the LLVMs and integrate it with complementary sources of information, such as the source trained model. Before we describe our proposed method, we formalize the SFVUDA task and introduce some preliminaries.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 76, + 194, + 176, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Problem definition and notations", + "text_level": 1, + "bbox": [ + 76, + 219, + 366, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In SFVUDA we are given an unlabelled target dataset $\\mathcal{D}^{\\mathrm{T}} = \\{\\mathbf{X}_i^{\\mathrm{T}}\\}_{i=1}^{m}$ , containing $m$ video sequences, and access to a source model $F^{\\mathrm{S}}(\\cdot)$ , trained on a labelled source dataset $\\mathcal{D}^{\\mathrm{S}} = \\{\\mathbf{X}_i^{\\mathrm{S}}, y_i^{\\mathrm{S}}\\}_{i=1}^{n}$ , which is not available during adaptation on the target. We assume that the target dataset $\\mathcal{D}^{\\mathrm{T}}$ contains video sequences of actions that share the same label space as the source, i.e., $\\mathcal{Y}^{\\mathrm{S}} = \\mathcal{Y}^{\\mathrm{T}} = \\{1, \\dots, C\\}$ , with $C$ being the number of semantic action categories. We also assume that the source and target marginal distributions are not the same, leading to the so-called domain-shift [44].", + "bbox": [ + 75, + 243, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of any SFVUDA method is to utilize the target dataset and the source-trained model to learn a mapping function (typically using a neural network) that can correctly predict the target samples, without needing any access to the source dataset.", + "bbox": [ + 75, + 393, + 470, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 478, + 218, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP [31], which stands for Contrastive Language-Image Pre-training, is composed of two encoders: a vision encoder $G_{\\mathrm{V}}(\\cdot)$ and a language (or text) encoder $G_{\\mathrm{L}}(\\cdot)$ (see Fig. 2b). To classify, it associates labels to visual inputs by computing their similarity with a set of textual descriptions. More precisely, given the names of the classes in a dataset (i.e., drinking, walking, etc.), we construct prompts containing each class name, e.g., \"a video of a person [CLS]\", where \"[CLS]\" denotes a class name.", + "bbox": [ + 75, + 501, + 468, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The language model $G_{\\mathrm{L}}(\\cdot)$ projects the class names into embeddings, denoted as $\\{z_c^{\\mathrm{L}}\\}_{c = 1}^{C}$ . On the other hand, assuming a test video of $K$ frames $\\mathbf{X} = \\{\\mathbf{x}_k\\}_{k = 1}^K$ , we extract the visual feature representation for each frame, i.e. $\\mathbf{Z}^{\\mathrm{V}} = \\{z_k^{\\mathrm{V}}\\}_{k = 1}^K$ using the vision encoder $G_{\\mathrm{V}}(\\cdot)$ . CLIP outputs a probability distribution, where the probability of the $k$ -th frame to belong to the $c$ -th class is given by:", + "bbox": [ + 75, + 638, + 468, + 743 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {c} (\\mathbf {x} _ {k}) = \\frac {\\exp \\left(< z _ {c} ^ {\\mathrm {L}} , z _ {k} ^ {\\mathrm {V}} > / \\tau\\right)}{\\sum_ {c ^ {\\prime} = 1} ^ {C} \\exp \\left(< z _ {c ^ {\\prime}} ^ {\\mathrm {L}}, z _ {k} ^ {\\mathrm {V}} > / \\tau\\right)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 753, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ is a temperature and $<\\cdot, \\cdot>$ is the inner-product (or cosine similarity) operator. The frame-wise predictions are then aggregated by simply averaging the class probabilities estimated for each frame.", + "bbox": [ + 75, + 795, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With respect to inference with CLIP, it is regarded as zero-shot inference if a given dataset has not been explicitly used for training the CLIP model. In other words,", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "zero-shot inference refers to the generalization to unseen datasets [31], and must not be confused with generalization to strictly unseen objects [17].", + "bbox": [ + 496, + 90, + 892, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Domain Adaptation with Large Language-Vision models (DALL-V)", + "text_level": 1, + "bbox": [ + 498, + 145, + 890, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we propose Domain Adaptation with Large Language-Vision models (DALL-V) as a SFVUDA method. As the name suggests, it leverages the rich world prior from the LLVMs, CLIP in particular, to help mitigate the domain gap. Recall that CLIP has been trained on image-text pairs and not video-caption pairs. On the other hand, the source model can offer complementary information as it has been trained supervisedly on the source videos. Thus, we aspire to jointly leverage these two reservoirs of information in a simple yet effective manner.", + "bbox": [ + 496, + 183, + 892, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our DALL-V mainly operates in two stages. The first stage, called Target adaptation (see Sec. 4.3.1), consists in pseudo-labelling the unlabelled target videos with CLIP and training a target-specific model with the pseudo-labelled data. In this way the target-specific model can benefit from the rich general purpose knowledge of CLIP and at the same time discover patterns that are inherent to the target domain. The second stage, called Ensemble distillation (see Sec. 4.3.2), focuses on ensembling the source model and target model into a smaller network, with knowledge distillation [8], which is then finally used during inference. Before we dive into the details of our proposed method, we describe how we carry out source training.", + "bbox": [ + 496, + 335, + 893, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Source Pre-Training. Opposed to the previous work AT-CoN, which fine-tunes all the source model weights using the source dataset, we propose an alternative approach. To prevent the source model to be biased towards the source dataset, we propose to initialize the source model with CLIP pre-trained weights, and fine-tune only a small set of parameters. To fine-tune CLIP on the source data, we introduce an adapter network [6], a small multi-layer perceptron network, that is appended on top of the vision encoder of CLIP. Source pre-training essentially consists in updating the weights of the adapter using a supervised training objective, while keeping the CLIP vision encoder frozen.", + "bbox": [ + 496, + 537, + 893, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, the vision encoder of CLIP $G_{\\mathrm{V}}(\\cdot)$ is extended with an adapter $A(\\cdot): \\mathcal{R}^d \\to \\mathcal{R}^d$ on top, where $d$ is the output dimension of $G_{\\mathrm{V}}(\\cdot)$ . We construct textual descriptions $t$ for the video sequences using a set of 16 manually-designed prompts [40], containing templates such as \"a video of [CLS]\", \"can you recognize the action of [CLS]?\", and so on, with $y$ being the class name [CLS] (refer to the Supplement for the complete list of prompts).", + "bbox": [ + 496, + 719, + 893, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt the training objective of ActionCLIP [40] to train the source model. In details, we first compute the ground-truth similarity scores $q(\\mathbf{x})$ for each input frame $\\mathbf{x}$ ,", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10310", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d372bd9eee848a026ea930442bbfba424ce52ae1df8bf5767cbb156631eb27c4.jpg", + "image_caption": [ + "(a) Source/target adapter $(A^{\\mathrm{S}} / A^{\\mathrm{T}})$ training. For the target domain, $q$ is constructed by pseudo-labelling with CLIP (ViT-B/32)." + ], + "image_footnote": [], + "bbox": [ + 109, + 140, + 428, + 301 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e0e869990c32cf734566767e346307e379a8c88f30c41eb7982f6ed555a85065.jpg", + "image_caption": [ + "(b) Ensemble distillation. We use $A^{\\mathrm{S}}, A^{\\mathrm{T}}$ , and CLIP (ViT-B/32) as teachers and train a student adapter $\\overline{A}$ on a CLIP (RN50) model.", + "Figure 3: Overview of the pipeline of DALL-V. The $\\mathbb{O}$ means the network is trainable, while the $\\text{念}$ means the network is frozen. We denote with bar the student components to distinguish them from the teacher." + ], + "image_footnote": [], + "bbox": [ + 532, + 101, + 846, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with 0 in the positions corresponding to negative pairs in the mini-batch and 1 for positive pairs, where a pair denotes the frame-prompt tuple. To train the network (essentially just the adapter $A^{\\mathrm{S}}$ ; language and vision encoders being frozen) we use the Kullback-Leibler (KL) divergence loss between $\\mathbf{p}(\\mathbf{X})$ (see Eq. (1)) and $\\mathbf{q}(\\mathbf{X})$ :", + "bbox": [ + 73, + 387, + 468, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a d p a t e r}} ^ {\\mathrm {S}} = \\mathbb {E} _ {(\\mathbf {X}, t) \\sim \\mathcal {D} ^ {\\mathrm {s}}} [ \\mathrm {K L} (\\mathbf {p} (\\mathbf {X}), \\mathbf {q} (\\mathbf {X})) ] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 491, + 468, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3.1 Target adaptation", + "text_level": 1, + "bbox": [ + 75, + 527, + 256, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To adapt CLIP on the target domain, we adopt a similar approach to that outlined for the source domain, with one key difference: the absence of ground-truth labels for the target domain. While it is a common practice among the source-free methods to utilize the source model to pseudolabel the unlabelled target data [20], we refrain from such an approach, and instead utilize the original CLIP model (i.e., without any source adapter $A^{\\mathrm{S}}$ ) for pseudo-labelling. Despite learning fewer parameters with the adapter, the source model can still be biased towards the source, and hence, yielding noisy pseudo-labels.", + "bbox": [ + 75, + 551, + 468, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To pseudo-label (PL), we utilize zero-shot CLIP inference (described in Sec. 4.2), with the same collection of textual descriptions used during the source fine-tuning stage, to assign a PL to each sample in the target domain. To this end, we use a confidence threshold for filtering out unreliable PLs (see Supplement for details). After the thresholding, we end up with a subset $\\tilde{\\mathcal{D}}^{\\mathrm{T}} = \\{\\mathbf{X}_i^{\\mathrm{T}}\\}_{i=1}^{m'} \\subset \\mathcal{D}^{\\mathrm{T}}$ , containing PLs as $\\{\\tilde{y}_i\\}_{i=1}^{m'}$ with $m'$ denoting the filtered samples.", + "bbox": [ + 75, + 718, + 468, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the source fine-tuning procedure, we append an adapter $A^{\\mathrm{T}}(\\cdot): \\mathcal{R}^{d} \\to \\mathcal{R}^{d}$ to the vision encoder of CLIP and minimize the following loss, while keeping CLIP frozen:", + "bbox": [ + 75, + 840, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a d a p t e r}} ^ {\\mathrm {T}} = \\mathbb {E} _ {(\\mathbf {X}, t) \\sim \\tilde {\\mathcal {D}} ^ {\\mathrm {T}}} [ \\operatorname {K L} (\\mathbf {p} (\\mathbf {X}), \\mathbf {q} (\\mathbf {X})) ] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 400, + 890, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3.2 Ensemble distillation", + "text_level": 1, + "bbox": [ + 498, + 434, + 700, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With three different models at our disposal: the source, target and the original CLIP, we can distill information from these three models into a single network. Knowledge Distillation [8] presents two important advantages: (i) allows to generate more reliable and informative pseudo-labels, and (ii) distillation into a smaller network reduces inference time without sacrificing much performance. For this second point, we explicitly use CLIP ViT-B/32 as teacher (and as backbone for the source and target adapters) and CLIP RN50 as student. Since we use an ensemble of three models to distill information, we call this step ensemble distillation. Now, with predictions available from three different models, the question is how to distill them together. A simple approach could be to average the three probability distributions as follows:", + "bbox": [ + 496, + 458, + 890, + 683 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} ^ {\\text {e n s}} (\\mathbf {X}) = \\frac {1}{3} \\left(\\underbrace {\\mathbf {p} (\\mathbf {X})} _ {\\text {Z S C L I P}} + \\underbrace {\\mathbf {p} ^ {\\mathrm {S}} (\\mathbf {X})} _ {\\text {S o u r c e a d p a t e r}} + \\underbrace {\\mathbf {p} ^ {\\mathrm {T}} (\\mathbf {X})} _ {\\text {T a r g e t a d p a t e r}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 695, + 890, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Whereas, another approach could consist in carrying out majority voting, where each of the three heads votes for a class, and the class with two or more votes becomes the new PL. Following [8], we utilize both kinds of PLs for distillation, using a standard multi-class classification loss (or cross-entropy loss $\\mathcal{L}_{\\mathrm{CE}}$ ) and a discrepancy loss. The discrepancy loss corresponding to the ensembled PL in Eq. (4) is given as:", + "bbox": [ + 496, + 742, + 890, + 863 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s c r e p a n c y}} = \\mathbb {E} _ {(\\mathbf {X}, t) \\sim \\mathcal {D} ^ {\\tau}} \\left[ \\mathrm {K L} \\left(\\frac {\\mathbf {p} ^ {\\text {s t u d e n t}} (\\mathbf {X})}{\\tau}, \\frac {\\mathbf {p} ^ {\\text {e n s}} (\\mathbf {X})}{\\tau}\\right) \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 873, + 890, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10311", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{p}^{\\mathrm{student}}$ denotes the student network's probability distribution, which is matched with that of $\\mathbf{p}^{\\mathrm{ens}}$ .", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then balance the contribution of the two losses as:", + "bbox": [ + 96, + 122, + 452, + 137 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t i l l}} = \\alpha * \\mathcal {L} _ {c e} + (1 - \\alpha) * \\mathcal {L} _ {d} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 156, + 468, + 171 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ is set according to the value proposed in [8]. Note that, we do not fine-tune all the student weights, but instead append an adapter on top of the frozen backbone and fine-tune only the adapter $\\overline{A}$ using Eq. (6).", + "bbox": [ + 76, + 183, + 468, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 260, + 209, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental setup", + "text_level": 1, + "bbox": [ + 76, + 285, + 264, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and settings. We present an extensive experimental evaluation on three standard benchmarks for SFVUDA. In particular, we report results on Daily-DA [44], which comprises 18,949 videos from 8 classes, and it is built from 4 original video action recognition datasets, namely HMDB51 [15], ARID [43], MIT [26] and Kinetics [12]. Additionally, we evaluate our framework on UCF-HMDBfull [1], a benchmark comprising 3,209 videos divided in 12 action categories from the HMDB51 [15] and UCF101 [38] action recognition datasets. Note that the first benchmark, Daily-DA, poses a more significant challenge when compared to the latter since it comprises videos with very different lighting conditions across domains. Last, we test our method on Sports-DA [45], a benchmark consisting of three datasets (i.e., UCF101 [38], Sports-1M [11], and Kinetics [12]), with 40,718 videos and 23 classes.", + "bbox": [ + 76, + 310, + 468, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We implement our adapters as two-layers perceptrons, following [6]. In accordance with [40], we use AdamW optimizer [24] with a learning rate of 0.01 and a weight decay of 0.2. The temperature $\\tau$ is set to 2.0 for all datasets. We trained all our models for 30 epochs. For a single experiment, we have used either 4 Tesla V100 or 2 RTX A6000 GPUs. Further implementation details can be found in the Supplement.", + "bbox": [ + 76, + 551, + 468, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and competitors. We compare our method with a selection of standard baseline methods for VUDA and SFVUDA, which include TRN [49], DANN [5], MK-MDD [23], $\\mathrm{TA}^3\\mathrm{N}$ [1], SFDA [13], SHOT [20], SHOT++ [21], MA [19], BAIT [48] and CPGA [30]. Additionally, we report the scores of the state-of-the-art competitors ATCoN [44] and EXTERN [46]. We also report the results for a set of CLIP-based baselines of our design to provide a more solid context for assessing our performance. In particular, CLIP (RN50) and CLIP (ViT-B/32) indicate the scores obtained in a zero-shot fashion with the CLIP model, with ResNet-50 and ViT-B/32 as backbones, respectively. Finally, we report the lower and upper bounds obtained by the source-only and target-supervised models, respectively.", + "bbox": [ + 76, + 674, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison with the State-of-the-Art", + "text_level": 1, + "bbox": [ + 500, + 90, + 825, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report our experimental validation scores in Tab. 1, Tab. 2, and Tab. 3, obtained on the Daily-DA, the UCF-HMDB $f_{ull}$ , and the Sports-DA benchmarks respectively. For all the benchmarks, the lower bound refers to the out-of-the-box performance of our source-trained model on the target data, i.e., the CLIP backbone and the source adapter $A^{\\mathrm{S}}$ . Analogously, the upper bound is the oracle performance obtained after training supervisedly on the target data, i.e., assuming knowledge of the target labels.", + "bbox": [ + 496, + 125, + 890, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regarding the more challenging Daily-DA dataset, it is possible to observe in Tab. 1 that our proposed distillation method DALL-V (RN50) obtains the best average score across the 12 settings among the competitors and baselines sharing the same backbone, with a $5\\%$ gain on the second best score (CLIP (RN50)). It is also visible that the best score is achieved in 9 out 12 total settings, proving our proposed distillation-based method to be significantly effective when addressing the SFVUDA task. Notably, it emerges that the scores achieved when addressing the ARID dataset $(\\rightarrow \\mathrm{A})$ as target domain are visibly lower than those reported for the other settings. We associate this behavior to the challenging visual features of this particular benchmark, characterized by being shot in a low illumination environment.", + "bbox": [ + 496, + 273, + 890, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As for the small-scale UCF-HMDBfull dataset, we can observe significantly higher absolute accuracy scores on both the involved SFVUDA directions, indicating that the benchmark is significantly closer to saturation for this task. Nonetheless, it emerges from the table that our proposed method achieves a state-of-the-art average accuracy score across the two datasets (+0.6% on the best competitor) and the best score for the HMDB→UCF direction (+1.2% on the second best).", + "bbox": [ + 496, + 496, + 890, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, as shown in Tab. 3, our method achieves comparable results with EXTERN [46] on Sports-DA, surpassing it on three settings, and achieving the same results on another. On the two settings with UCF101 as target (i.e., $\\mathrm{K} \\rightarrow \\mathrm{U}$ and $\\mathrm{S} \\rightarrow \\mathrm{U}$ ), our method falls behind by approximately 6 percentage points, resulting on an average score across the six scenarios slightly lower than the state-of-the-art (i.e., $-0.9\\%$ ). However, DALL-V still achieves an average accuracy of $+8.5\\%$ w.r.t. ATCoN, further demonstrating the effectiveness of our approach for SFVUDA. Interestingly, on Sports-DA, our lower bound already achieves competitive results with other SFVUDA methods, surpassing ATCoN by $+2.7\\%$ on average. The table additionally shows that CLIP zero-shot is an extremely strong baseline on the benchmark, achieving comparable results with EXTERN, i.e., $-2.0\\%$ with RN50, and $+2.6\\%$ with the ViT-B backbone.", + "bbox": [ + 496, + 643, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10312", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ae7dbdb9ecb1b7390949f908a7e53b832ba8b0e3e974c71ea2243a37edab2751.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAccuracy (%)
K→AK→HK→MM→AM→HM→KH→AH→MH→KA→HA→MA→KAvg.
ZSTRN [49]20.936.729.022.143.753.113.822.037.117.214.724.427.9
Lower bound15.647.935.734.744.661.617.525.545.114.615.517.831.3
CLIP (RN50) [31]30.550.042.230.550.062.930.542.262.950.042.262.946.4
CLIP (ViT-B/32) [31]31.349.646.031.349.665.431.346.065.449.646.065.448.1
UDADANN [5]21.237.521.722.843.358.814.229.538.220.119.727.029.5
MK-MDD [23]21.736.224.021.050.458.520.325.733.818.718.026.129.5
TA3N [1]19.937.731.521.643.055.514.425.738.414.915.623.428.5
SFUDASFDA [13]12.644.927.516.035.249.213.124.224.916.313.225.225.2
SHOT [20]12.044.629.515.336.751.013.624.221.217.114.024.325.3
SHOT++ [21]12.640.828.714.941.746.316.022.233.115.412.521.824.4
MA [19]12.845.830.017.737.453.512.925.022.216.715.224.326.1
BAIT [48]12.745.730.016.939.653.013.625.521.215.714.525.526.2
CPGA [30]13.146.030.718.139.255.113.126.225.519.216.526.726.5
SFVUDAATCoN [44]17.248.232.527.247.357.717.930.748.526.717.231.033.5
EXTERNAL [46]23.955.835.218.153.768.126.240.757.626.218.251.439.6
DALL-V (ours)24.052.547.024.065.478.124.047.076.757.945.775.051.4
Upper bound26.970.461.526.970.488.926.961.588.970.461.588.961.9
", + "bbox": [ + 80, + 89, + 893, + 366 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/50fe8b492463eed26452ccffb9a68c514698a10df3b7957068f89fe96f023b58.jpg", + "table_caption": [ + "Table 1: Validation accuracy for Daily-DA. Bold indicates best, underline represents best with same backbone as baseline (i.e. ResNet50). Lower bound indicates a source adapter and Upper bound a target adapter, both trained supervised on CLIP (RN50)." + ], + "table_footnote": [], + "table_body": "
MethodAccuracy (%)
H→UU→HAvg.
ZSTRN [49]72.872.172.4
Lower bound71.676.173.8
CLIP (RN50) [31]81.086.083.5
CLIP (ViT-B/32) [31]90.389.189.7
UDADANN [5]74.475.174.8
MK-MMD [23]74.779.777.2
TA3N [1]78.184.881.5
SFUDASFDA [13]69.875.072.4
SHOT [20]74.474.474.4
SHOT++ [21]71.168.169.6
MA [19]74.467.370.9
BAIT [48]75.376.375.8
CPGA [30]75.868.172.0
SFVUDAATCoN [44]85.379.782.5
EXTERNAL [46]91.988.990.4
DALL-V (ours)93.188.991.0
Upper bound93.791.492.6
", + "bbox": [ + 138, + 443, + 411, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Validation accuracy for UCF-HMDBfull. Bold indicates best, underline represents best with same backbone as baseline (i.e. ResNet50). Lower bound indicates a source adapter and Upper bound a target adapter, both trained supervised on CLIP (RN50).", + "bbox": [ + 75, + 696, + 468, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation analysis", + "text_level": 1, + "bbox": [ + 76, + 797, + 246, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report in Tab. 4 the intermediate accuracy scores achieved at different steps of our proposed source-free pipeline for the Daily-DA and UCF-HMDBfull. In particular, we report the performance of the source supervised model and the target unsupervised one, also reporting the", + "bbox": [ + 75, + 825, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "final model and the zero-shot version of CLIP for comparison. Finally, we report the scores obtained by assembling the predictions of the three aforementioned models. The $A^{\\mathrm{S}}$ and $A^{\\mathrm{T}}$ notation indicates the usage of the source and target adapters on top of the frozen CLIP model, that are trained on the respective domains as described in detail in 4.3. Next we discuss the individual contributions. Additional ablation analyses are available in the Supplement.", + "bbox": [ + 496, + 446, + 890, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of finetuning LLVM on the target. Tab. 4 clearly shows that fine-tuning our model in an unsupervised fashion on the target domain is significantly effective with respect to the target validation accuracy. The tables show indeed a gain of $2.0\\%$ on Daily-DA and $1.1\\%$ on UCF-HMDB $f_{ull}$ . Additionally, it is possible to observe a further gain on Daily-DA when ensembling the target model (which gains $+2.7\\%$ by itself) with CLIP, achieving a $+3.6\\%$ . From these results, it emerges that fine-tuning LLVMs on the target domain, even in an unsupervised manner, is effective for the SFVUDA task.", + "bbox": [ + 496, + 568, + 892, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of ensembles on LLVM and domain-specific networks. In the last line of Tab. 4, it is possible to observe that ensembling predictions from the three distinct models turns out to yield the best average target accuracy score on both the considered benchmarks, resulting in a gain of $+5.7\\%$ on Daily-DA and of $+2.9\\%$ on UCF-HMDBfull w.r.t. the CLIP baseline. These results and those reported in the previous two paragraphs represent further proof that the three models are effectively complementary, and each contributes with additional discriminative knowledge to the resulting model. Conversely, CLIP positively contributes to", + "bbox": [ + 496, + 734, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10313", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/35dec57ff1a25e72aac3f5f073bd98310347911e1281104e5d12bee3225e14df.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAccuracy (%)
K→UK→SS→US→KU→KU→SAvg.
ZSTRN [49]86.466.985.371.049.343.367.0
Lower bound85.479.584.478.267.264.376.5
CLIP (RN50) [31]83.479.983.480.480.479.981.2
CLIP (ViT-B/32) [31]90.082.490.085.185.182.485.8
UDADANN [5]88.075.085.773.465.955.173.8
MK-MMD [23]90.267.990.973.666.155.674.0
TA3N [1]90.368.693.072.663.654.173.7
SFUDASFDA [13]86.160.085.468.055.843.666.5
SHOT [20]91.264.988.872.053.943.669.1
SHOT++ [21]90.063.188.070.344.740.966.2
MA [19]91.065.987.871.960.739.469.5
BAIT [48]92.366.688.372.857.244.770.3
CPGA [30]89.466.386.572.555.244.569.1
SEVUDAATCoN [44]93.669.790.676.065.247.973.8
EXTERNAL [46]93.773.895.482.281.272.783.2
DALL-V (ours)88.077.788.882.381.275.982.3
Upper bound93.488.393.485.685.688.389.1
", + "bbox": [ + 243, + 88, + 733, + 361 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5b6884464e68b0ae3210e5d5eadc52909c5ffabae0408de9583d09950ce1cadb.jpg", + "table_caption": [ + "Table 3: Validation accuracy for Sports-DA. Bold indicates best, underline represents best with same backbone as baseline (i.e. ResNet50). Lower bound indicates a source adapter and Upper bound a target adapter, both trained supervised on CLIP (RN50)." + ], + "table_footnote": [], + "table_body": "
MethodAccuracy (%)
K→AnyM→AnyDaily-DAA→AnyAvg.UCF-HMDBfull
H→AnyA→AnyH→UU→HAvg.
CLIP42.348.747.653.748.190.389.189.7
AS40.0 [-2.3]55.7 [+7.0]35.3 [-12.4]35.0 [-18.7]41.5 [-6.6]91.0 [+0.7]80.5 [-8.6]85.8 [-3.9]
AT44.4 [+2.1]51.6 [+2.9]48.5 [+0.9]58.4 [+4.7]50.8 [+2.7]90.5 [+0.2]91.1 [+2.0]90.8 [+1.1]
CLIP + AS45.1 [+2.8]57.8 [+9.1]40.0 [-7.6]50.8 [-2.9]48.4 [+0.3]93.1 [+2.8]87.2 [-1.9]90.2 [+0.5]
CLIP + AT43.8 [+1.5]50.9 [+2.2]48.2 [+0.6]57.4 [+3.7]51.7 [+3.6]89.3 [-1.0]91.9 [+2.8]90.6 [+0.9]
CLIP + AS + AT46.2 [+3.9]57.3 [+8.6]49.5 [+1.9]62.1 [+8.4]53.8 [+5.7]94.9 [+4.6]90.3 [+1.2]92.6 [+2.9]
", + "bbox": [ + 80, + 433, + 890, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: Ablation on the relative improvement of the different modules of our proposed DALL-V framework w.r.t. CLIP (ViT-B/32) on both considered benchmarks. Bold indicates best. The + operator indicates assembling. Due to limited space, results are aggregated by source dataset.", + "bbox": [ + 75, + 590, + 893, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the final score thanks to its well-known ability to generalize to unseen domains in a zero-shot fashion; on the other hand, the tuning process of domain-specific adapters on the source and target domains, although not sufficient by itself, proves beneficial thanks to its ability to capture specific features of the involved domains and the associated label sets. Note that we still opted for distilling this knowledge to a ResNet-based model in order to be directly comparable with the competitors.", + "bbox": [ + 75, + 648, + 472, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of using multiple templates. As a further insight into the individual contributions of the different components of our framework, we propose in Fig. 5, a simple analysis on how the number of templates for the text input at inference time impacts on the accuracy score on the target validation set of the two considered benchmarks. We observe that increasing the number of templates from just", + "bbox": [ + 75, + 794, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a single one to 12 (following ActionCLIP [40]) is significantly beneficial with respect to the final accuracy score. For even higher number of templates, the performance of our model reaches a plateau. Note that for the choice of the number of template we opted for following [40].", + "bbox": [ + 496, + 648, + 893, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Modeling the feature space. We plot in Fig. 4 an UMAP [25] chart of the features produced by the zero-shot CLIP (RN50) model, the source model, and our DALL-V model, on the UCF-HMDBfull benchmark, showing the features belonging to different classes in different colors. The chart clearly highlights the ability of the proposed DALL-V method to produce well-separated clusters for different action categories, thanks to the effective combination of the complementary knowledge of intermediate models.", + "bbox": [ + 496, + 772, + 893, + 910 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10314", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/797cbb1017be164c8da2199bd4bc10da16b4260067b5c7730f71f69a737166cb.jpg", + "image_caption": [ + "(a) HMDB $\\rightarrow$ UCF" + ], + "image_footnote": [], + "bbox": [ + 98, + 104, + 467, + 208 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dc039be3a8c03a55a1e2965d8cf3f6b063ed9fc60f022905b4d1e49363428c43.jpg", + "image_caption": [ + "(b) UCF $\\rightarrow$ HMDB" + ], + "image_footnote": [], + "bbox": [ + 506, + 103, + 870, + 208 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e7ea954910da37e4eb4a11593610f5774c5305f10062a750494e0e89f3c21843.jpg", + "image_caption": [ + "Figure 4: UMAP visualisation of the feature space on $\\mathbf{UCF - HMDB}_{full}$ for the zero-shot CLIP (RN50), source and final DALL-V models. The chart shows the ability of our proposed DALL-V to efficiently cluster the action categories in the features space by combining knowledge from the other two models.", + "Figure 5: Sensitivity study on the number of templates used for the text input w.r.t. the accuracy score, on Daily-DA and UCF-HMDB $f_{ull}$ ." + ], + "image_footnote": [], + "bbox": [ + 104, + 289, + 444, + 460 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 76, + 535, + 181, + 550 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note that CLIP has been trained on web-crawled image-text pairs that are not made publicly available by OpenAI [35]. Due to its black-box nature, equally trusting CLIP predictions, as much as the source model, may elicit unpredictable behaviour, especially in safety critical applications. Thus, appropriate caution must be exercised. Moreover, ours being an empirical work, we can not provide any theoretical guarantee if using LLVMs, such as CLIP, will always result in better performance over traditional SFVUDA approaches. Thus, as future work, we plan to extensively evaluate using more open-source alternatives.", + "bbox": [ + 75, + 556, + 468, + 723 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 739, + 204, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We presented DALL-V, a SFVUDA method based on a simple but novel approach to SFVUDA driven by the intuition of combining the complementary information derived from domain-specific simple models and the powerful CLIP-based LLVMs trained on world knowledge. We provided motivation for such approach, and reported an extensive evaluation on three standard benchmarks for VUDA, pursued in our case for the source-free scenario. We compare our performance with existing methods, as well as with", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "a selection of CLIP-based baselines, showing that our proposed frameworks achieves state-of-the-art performance on both considered benchmarks.", + "bbox": [ + 498, + 286, + 892, + 330 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments. We acknowledge the support of the MUR PNRR project FAIR - Future AI Research (PE00000013) funded by the NextGenerationEU. E.R. is partially supported by the PRECRISIS, funded by the EU Internal Security Fund (ISFP-2022-TFI-AG-PROTECT-02-101100539), the EU project SPRING (No. 871245), and the by the PRIN project LEGO-AI (Prot. 2020TA3K9N). The work was carried out in the Vision and Learning joint laboratory of FBK and UNITN. This paper has been further supported by the French National Research Agency (ANR-20-CE23-0027), the EU Horizon 2020 Research and Innovation program under grant agreement No 957337.", + "bbox": [ + 498, + 349, + 893, + 531 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 541, + 596, + 556 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Min-Hung Chen, Zsolt Kira, Ghassan AlRegib, Jaekwon Yoo, Ruxin Chen, and Jian Zheng. Temporal attentive alignment for large-scale video domain adaptation. In ICCV, 2019. 1, 3, 6, 7, 8", + "[2] Victor G Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-head contrastive domain adaptation for video action recognition. In WACV, 2022. 1, 3", + "[3] Christoph Feichtenhofer, Axel Pinz, and Richard P Wildes. Spatiotemporal multiplier networks for video action recognition. In CVPR, 2017. 1", + "[4] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In CVPR, 2016. 1", + "[5] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation, 2015. 6, 7, 8", + "[6] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv, 2021. 4, 6", + "[7] Muhammad Ghifary, W Bastiaan Kleijn, Mengjie Zhang, David Balduzzi, and Wen Li. Deep reconstruction-classification networks for unsupervised domain adaptation. In ECCV, 2016. 3" + ], + "bbox": [ + 506, + 566, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10315", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[8] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv, 2015. 4, 5, 6", + "[9] Yunzhong Hou and Liang Zheng. Visualizing adapted knowledge in domain transfer. In CVPR, 2021. 3", + "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 3", + "[11] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, pages 1725–1732, 2014. 6", + "[12] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset, 2017. 6", + "[13] Youngeun Kim, Donghyeon Cho, Kyeongtak Han, Priyadarshini Panda, and Sungeun Hong. Domain adaptation without source data, 2020. 6, 7, 8", + "[14] Yu Kong and Yun Fu. Human action recognition and prediction: A survey. IJCV, 2022. 1", + "[15] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. Hmdb: A large video database for human motion recognition. In ICCV, 2011. 6", + "[16] Vinod K Kurmi, Venkatesh K Subramanian, and Vinay P Namboodiri. Domain impression: A source data free domain adaptation method. In WACV, 2021. 3", + "[17] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In CVPR, 2009. 4", + "[18] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 3", + "[19] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 6, 7, 8", + "[20] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 3, 5, 6, 7, 8", + "[21] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. TPAMI, 44(11):8602-8617, 2021. 3, 6, 7, 8", + "[22] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, 2015. 3", + "[23] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I. Jordan. Learning transferable features with deep adaptation networks, 2015. 6, 7, 8", + "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 6", + "[25] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction, 2018. 8" + ], + "bbox": [ + 80, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, and Aude Oliva. Moments in time dataset: one million videos for event understanding. TPAMI, 42(2):502-508, 2019. 6", + "[27] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In CVPR, 2020. 2", + "[28] Boxiao Pan, Zhangjie Cao, Ehsan Adeli, and Juan Carlos Niebles. Adversarial cross-domain action recognition with co-attention. In AAAI CAI, 2020. 1, 3", + "[29] Preksha Pareek and Ankit Thakkar. A survey on video-based human action recognition: recent updates, datasets, challenges, and applications. Artificial Intelligence Review, 2021. 1", + "[30] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation, 2021. 6, 7, 8", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 7, 8", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1", + "[33] Aadarsh Sahoo, Rutav Shah, Rameswar Panda, Kate Saenko, and Abir Das. Contrast and mix: Temporal contrastive video domain adaptation with background mixing. NeurIPS, 2021. 2", + "[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv, 2022. 3", + "[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv, 2021. 1, 3, 9", + "[36] Javier Selva, Anders S Johansen, Sergio Escalera, Kamal Nasrollahi, Thomas B Moeslund, and Albert Clapés. Video transformers: A survey. arXiv, 2022. 1", + "[37] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. 1, 3", + "[38] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild, 2012. 6", + "[39] Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. In Artificial Intelligence, volume 30, 2016. 3", + "[40] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv, 2021. 2, 3, 4, 6, 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10316", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Conference on Empirical Methods in Natural Language Processing, pages 6787-6800, 2021. 2", + "[42] Yuecong Xu, Haozhi Cao, Kezhi Mao, Zhenghua Chen, Lihua Xie, and Jianfei Yang. Aligning correlation information for domain adaptation in action recognition. Neural Networks and Learning Systems, 2022. 3", + "[43] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Arid: A new dataset for recognizing action in the dark. In International Workshop on Deep Learning for Human Activity Recognition, pages 70-84. Springer, 2021. 6", + "[44] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, and Zhenghua Chen. Source-free video domain adaptation by learning temporal consistency for action recognition. In ECCV, 2022. 2, 3, 4, 6, 7, 8", + "[45] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, Rui Zhao, and Zhenghua Chen. Multi-source video domain adaptation with temporal attentive moment alignment. arXiv, 2021. 6", + "[46] Yuecong Xu, Jianfei Yang, Haozhi Cao, Min Wu, Xiaoli Li, Lihua Xie, and Zhenghua Chen. Leveraging endo- and exotemporal regularization for black-box video domain adaptation, 2022. 2, 3, 6, 7, 8", + "[47] Jinyu Yang, Weizhi An, Sheng Wang, Xinliang Zhu, Chaochao Yan, and Junzhou Huang. Label-driven reconstruction for domain adaptation in semantic segmentation. In ECCV, 2020. 3", + "[48] Shiqi Yang, Yaxing Wang, Joost Weijer, and Luis Herranz. Unsupervised domain adaptation without source data by casting a bait, 2020. 6, 7, 8", + "[49] Bolei Zhou, Alex Andonian, Aude Oliva, and Antonio Torralba. Temporal relational reasoning in videos, 2018. 6, 7, 8", + "[50] Yi Zhu, Xinyu Li, Chunhui Liu, Mohammadreza Zolfaghari, Yuanjun Xiong, Chongruo Wu, Zhi Zhang, Joseph Tighe, R Manmatha, and Mu Li. A comprehensive study of deep video action recognition. arXiv, 2020. 1" + ], + "bbox": [ + 78, + 90, + 468, + 670 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10317", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/c596fbdc-cd07-43b2-9a08-9301d722c491_model.json b/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/c596fbdc-cd07-43b2-9a08-9301d722c491_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ed7a81e5e896c7765af9124d31cb5052d57f81cb --- /dev/null +++ b/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/c596fbdc-cd07-43b2-9a08-9301d722c491_model.json @@ -0,0 +1,2070 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.131, + 0.82, + 0.176 + ], + "angle": 0, + "content": "The Unreasonable Effectiveness of Large Language-Vision Models for Source-free Video Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.203, + 0.914, + 0.221 + ], + "angle": 0, + "content": "Giacomo Zara1*, Alessandro Conti1*, Subhankar Roy3, Stéphane Lathuilière3, Paolo Rota1, Elisa Ricci1,2" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.221, + 0.745, + 0.239 + ], + "angle": 0, + "content": "1University of Trento, Italy 2Fondazione Bruno Kessler, Italy" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.239, + 0.745, + 0.257 + ], + "angle": 0, + "content": "\\(^{3}\\)LTCI, Télécom Paris, Institut polytechnique de Paris, France" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.259, + 0.686, + 0.275 + ], + "angle": 0, + "content": "{giacomo.zara, alessandro.conti-1}@unitn.it" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.473, + 0.63 + ], + "angle": 0, + "content": "Source-Free Video Unsupervised Domain Adaptation (SFVUDA) task consists in adapting an action recognition model, trained on a labelled source dataset, to an unlabelled target dataset, without accessing the actual source data. The previous approaches have attempted to address SFVUDA by leveraging self-supervision (e.g., enforcing temporal consistency) derived from the target data itself. In this work, we take an orthogonal approach by exploiting \"web-supervision\" from Large Language-Vision Models (LLVMs), driven by the rationale that LLVMs contain a rich world prior surprisingly robust to domain-shift. We showcase the unreasonable effectiveness of integrating LLVMs for SFVUDA by devising an intuitive and parameter-efficient method, which we name Domain Adaptation with Large Language-Vision models (DALL-V), that distills the world prior and complementary source model information into a student network tailored for the target. Despite the simplicity, DALL-V achieves significant improvement over state-of-the-art SFVUDA methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.673, + 0.21, + 0.689 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Video analysis tasks, such as action recognition, have long been investigated in computer vision, due to the numerous applications, ranging from video surveillance to social robotics [50, 14, 29]. Major progress has been made in the last decade with the development of specialized deep architectures, such as 3D CNNs [4, 3] and Video Transformers [36], trained on large-scale annotated datasets. However, obtaining sufficient labelled training videos for real-world scenarios can be very costly and time-consuming. In order to alleviate the burden of annotating large scale datasets Video-based Unsupervised Domain" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.315, + 0.885, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.508, + 0.895, + 0.616 + ], + "angle": 0, + "content": "Figure 1: Performance over time by various methods on the Daily-DA video benchmark. A pre-trained LLVM (e.g., CLIP [31]) is surprisingly better when compared with Unsupervised Domain Adaptation (UDA), Source Free UDA (SFUDA) and Source Free Video-based UDA (SFVUDA) methods. Our proposed DALL-V that is built on top of LLVM successfully outperforms all existing baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.894, + 0.734 + ], + "angle": 0, + "content": "Adaptation (VUDA) [1, 28, 2] have been introduced. The VUDA methods are derived from the common idea of transferring knowledge from a labelled source domain to an unlabelled target domain. While considering different strategies for adaptation, these approaches have all shown significant improvements in the robustness of learnt video models without requiring annotated target data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In the last few years, the field of computer vision has also witnessed the emergence of a new generation of powerful deep architectures, trained on mammoth internet-scale image-text datasets [35]. These models, commonly known as foundation models [32, 37, 31, 10] or Large LanguageVision Models (LLVM), have achieved outstanding performance, and have become a cornerstone of modern computer vision research. In particular, recent works such as CLIP [31] or FLAVA [37] have shown that rich visual representations for images can be learned from natural language descriptions as a form of supervision. These pre-trained" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.404, + 0.888 + ], + "angle": 0, + "content": "*Giacomo Zara and Alessandro Conti contributed equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.888, + 0.469, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Code is available at https://github.com/giaczara/dallyv" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10307" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.089, + 0.298, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.318, + 0.309, + 0.332 + ], + "angle": 0, + "content": "(a) Traditional SFVUDA" + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.088, + 0.547, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.353, + 0.319, + 0.522, + 0.332 + ], + "angle": 0, + "content": "(b) Zero Shot CLIP Inference" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.088, + 0.811, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.319, + 0.78, + 0.332 + ], + "angle": 0, + "content": "(c) Our proposed SFVUDA" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.338, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Figure 2: (a) In traditional SFVUDA a source-trained model is adapted to the unlabelled target dataset using self-supervision such as temporal consistency [44, 46] among frames. (b) LLVM (e.g., CLIP [31]) uses cosine similarity between the feature and language representation to predict the most probable class in a zero-shot (ZS) manner. (c) Our proposed SFVUDA solution distills the ZS-CLIP and source model predictions, while introducing very few learnable parameters, namely adapters. The \\(\\hat{\\mathbb{C}}\\) means the network is trainable, while the \\(\\hat{\\mathbb{C}}\\) means the network is frozen." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.469, + 0.508 + ], + "angle": 0, + "content": "LLVMs are now publicly available and can be easily integrated into any recognition system. Lately, researchers have also shown that the LLVMs can be applied with success to the video domain and specifically to the supervised action recognition task [40, 41]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.469, + 0.75 + ], + "angle": 0, + "content": "Despite the significant progress in the VUDA literature, to date, the role of LLVM in the context of action recognition has been overlooked. We argue that the LLVMs have a lot to offer to the progress of VUDA methods, which is currently untapped. Our work starts from a preliminary analysis demonstrating that current sophisticated VUDA methods can be largely outperformed by off-the-shelf publicly available LLVMs like CLIP, without needing explicit adaptation (see Fig. 1). This observation highlights the need for a paradigm shift among VUDA methods. The benefit of LLVM models is so significant that it raises a pertinent question: If explicit alignment methodologies, traditionally pursued in VUDA, is truly the way forward? Instead, the primary research question becomes: How can we efficiently integrate the prior knowledge derived from LLVMs to adapt a model to the target domain for VUDA?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.901 + ], + "angle": 0, + "content": "This paper presents the first approach in the VUDA literature, which attempts to address this question specifically. In particular, we consider the challenging scenario of Source-Free Video Unsupervised Domain Adaptation (SFVUDA) [44, 46], which consists in the task of adapting an action recognition model trained on a labelled source domain to an unlabelled target domain, without accessing the actual source data. The primary motivation for exploiting LLVMs in SFVUDA is that we expect the generalization capabilities of these models to effectively serve the pur" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.644 + ], + "angle": 0, + "content": "pose of mitigating the effects of the domain shift existing between the data distributions of the source and the target domain. In particular, since in SFVUDA we do not have access to the source data but only the source pre-trained model, direct target adaptation may potentially lead to significantly poor performance, especially when the domain shift is large. On the contrary, the employment of LLVMs can efficiently counteract this negative effect owing to the wide range of visual domains observed during their training process. An additional motivation behind our approach lies in the fact that the rich visual representations derived from LLVMs can efficiently compensate when additional modalities (e.g., optical-flow), which are known to help reduce domain gap in VUDA frameworks [27, 33], are unavailable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.87 + ], + "angle": 0, + "content": "Our approach, which we name as Domain Adaptation with Large Language-Vision models (or DALL-V in short), is a simple yet effective technique to combine the knowledge from the visual representation of a pre-trained LLVM with that derived from the source model and the target data (see Fig. 2c). DALL-V involves a two-stage pipeline: (i) in the first stage pseudo-labels are extracted from the CLIP model and are subsequently used for adapting on the target dataset, and (ii) in the second stage an ensemble of CLIP, source and target models are used to distill information into a student network. DALL-V introduces very few trainable parameters on top of CLIP and is realized with the help of domain-specific adapters. Despite its simplicity, our approach outperforms SFVUDA and even VUDA methods by a significant margin (+11.8% w.r.t. the best competitor)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In summary, our contributions are: (i) We show, for the first time in the literature that straightforward zero-shot" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10308" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.212 + ], + "angle": 0, + "content": "methods based on LLVMs vastly outperform the state-of-the-art approaches for SFVUDA. (ii) Building upon this observation, we propose DALL-V, a simple approach for SFVUDA which optimally integrates information derived from LLVMs, from a pretrained source model and from unlabelled videos of the target domain. (iii) We perform extensive experiments on a total of 20 domain adaptation settings, demonstrating the effectiveness of DALL-V in SFVUDA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.226, + 0.22, + 0.243 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.251, + 0.47, + 0.568 + ], + "angle": 0, + "content": "Video Unsupervised Domain Adaptation. While domain adaptation techniques have been primarily studied in the context of image-level representations and fall under the umbrella of Unsupervised Domain Adaptation (UDA) methods [7, 47, 22, 39, 1, 42], addressing the domain shift problem is even more challenging in the case of videos, due to the additional temporal dimension. Researchers have proposed different strategies. For instance, Chen et al. introduced \\(\\mathrm{TA}^3\\mathrm{N}\\) [1], a discrepancy-based method that aligns domains on the temporal axis by learning a temporal relationship across video sequences. TCoN was introduced in [28], a deep architecture which employs a cross-domain module to compute temporally-aligned source and target feature representations. CO2A [2] proposed an approach based on contrastive learning to align source and target video representations. One major limitation with both UDA and VUDA methods is the fact that they require access to source data. This limits their applicability in real-world scenarios where data may not be accessible due to privacy reasons. Differently, our approach focuses in the more realistic setting where we only have access to the source model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Source-Free Video Unsupervised Domain Adaptation. Due to the growing concerns related to privacy and data sharing, recently researchers have begun exploring methods Source-Free Unsupervised Domain Adaptation (SFUDA). One of the first works in this direction is SHOT [20], a deep architecture which employs an entropy loss and a classification loss on pseudo-labelled data to adapt the source pretrained network focusing solely on the target data. In a subsequent work [21], the authors introduced an additional auxiliary head to solve the relative rotation task, further improving prediction accuracy. Other works, such as 3C-GAN [18], SFIT [9], and SDDA [16], framed the problem of SFUDA as an image translation task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In the literature, only a handful of works addressed the problem of Source-Free Video Unsupervised Domain Adaptation (SFVUDA). In particular, ATCoN [44] presented an approach which models temporal consistency across the video sequences. EXTERN [46] proposed to exploit mask-to-mix strategies and video-tailored regularizations for SFVUDA. Our approach radically departs from these previous works as we proposed to exploit LLVMs to adapt to the target video data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.333 + ], + "angle": 0, + "content": "Large Language-Vision Models. The availability of vast web-scale datasets containing image-text pairs [35, 34, 37] have enabled the emergence of novel large multi-modal neural networks which learn joint visual-text embedding spaces [31, 10, 37]. These approaches, commonly referred as Large Language-Vision Models, utilise a separate encoder for each modality and employ a contrastive loss to align the data representations in the feature space [31, 10]. CLIP [31] is a prominent example of such an approach. Despite its simplicity, CLIP have been shown to achieve impressive zero-shot image recognition capabilities. More recently, LLVMs have been also extended with success to the video domains and particularly to the action recognition task [40]. However, we are not aware of previous works which have specifically used LLVMs to address the problem of domain shift in video action recognition." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.347, + 0.874, + 0.364 + ], + "angle": 0, + "content": "3. The Unreasonable Effectiveness of LLVM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.508 + ], + "angle": 0, + "content": "In this work our goal is to develop a Source-free Video Unsupervised Domain Adaptation (SFVUDA) method that can adapt a source trained model to a target domain of interest. Departing from the traditional SFVUDA approaches, where a multitude of self-supervised losses are optimized on the target domain [44], we pursue an orthogonal and unconventional approach. Our key idea is to leverage a LLVM (e.g., CLIP [31]), which has been pre-trained on web-scale image-text pairs, as a tool for bridging the domain gap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.892, + 0.75 + ], + "angle": 0, + "content": "As a part of a preliminary study we evaluate CLIP in a zero-shot manner (see Fig. 2b) on the target dataset. In other words, we simply run inference and evaluate the performance on the Daily-DA [44] benchmark using a pre-trained CLIP (see Sec. 4.2 for details on zero-shot inference), and compare with the recent state-of-the-art SFVUDA approaches. The evaluation is carried out on the four target domains of the Daily-DA, each of which has eight semantic categories. As showcased in Fig. 1, much to our surprise, CLIP (ViT-B/32) outperforms the best performing SFVUDA method EXTERN [46] by a huge margin of \\(+8.5\\%\\), despite CLIP never been explicitly fine-tuned on the target frames, let alone videos. These telling observations from our preliminary study compels us to explore and exploit the unreasonable effectiveness of the LLVMs, which is orthogonal to the existing SFVUDA literature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The implications of undertaking such a pathway to bridge the domain gap in SFVUDA can be several: (i) the LLVMs being publicly available, it will help democratize more efficient target adaptation, even if the source data is withheld due to privacy reasons, (ii) it disposes off the need to balance and tune complex training objectives, a de facto practice in SFVUDA (e.g., ATCoN [44] requires to jointly optimize up to 6 loss functions), (iii) it prevents the necessity of ad hoc and specialized network architectures and training objectives to solve the task at hand. Encouraged by" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10309" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.184 + ], + "angle": 0, + "content": "the flexibility offered by LLVMs, we propose DALL-V for SFVUDA that can effectively exploit the world prior from the LLVMs and integrate it with complementary sources of information, such as the source trained model. Before we describe our proposed method, we formalize the SFVUDA task and introduce some preliminaries." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.195, + 0.178, + 0.211 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.22, + 0.367, + 0.236 + ], + "angle": 0, + "content": "4.1. Problem definition and notations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.47, + 0.395 + ], + "angle": 0, + "content": "In SFVUDA we are given an unlabelled target dataset \\(\\mathcal{D}^{\\mathrm{T}} = \\{\\mathbf{X}_i^{\\mathrm{T}}\\}_{i=1}^{m}\\), containing \\(m\\) video sequences, and access to a source model \\(F^{\\mathrm{S}}(\\cdot)\\), trained on a labelled source dataset \\(\\mathcal{D}^{\\mathrm{S}} = \\{\\mathbf{X}_i^{\\mathrm{S}}, y_i^{\\mathrm{S}}\\}_{i=1}^{n}\\), which is not available during adaptation on the target. We assume that the target dataset \\(\\mathcal{D}^{\\mathrm{T}}\\) contains video sequences of actions that share the same label space as the source, i.e., \\(\\mathcal{Y}^{\\mathrm{S}} = \\mathcal{Y}^{\\mathrm{T}} = \\{1, \\dots, C\\}\\), with \\(C\\) being the number of semantic action categories. We also assume that the source and target marginal distributions are not the same, leading to the so-called domain-shift [44]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.472, + 0.471 + ], + "angle": 0, + "content": "The goal of any SFVUDA method is to utilize the target dataset and the source-trained model to learn a mapping function (typically using a neural network) that can correctly predict the target samples, without needing any access to the source dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.479, + 0.22, + 0.495 + ], + "angle": 0, + "content": "4.2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.47, + 0.639 + ], + "angle": 0, + "content": "CLIP [31], which stands for Contrastive Language-Image Pre-training, is composed of two encoders: a vision encoder \\( G_{\\mathrm{V}}(\\cdot) \\) and a language (or text) encoder \\( G_{\\mathrm{L}}(\\cdot) \\) (see Fig. 2b). To classify, it associates labels to visual inputs by computing their similarity with a set of textual descriptions. More precisely, given the names of the classes in a dataset (i.e., drinking, walking, etc.), we construct prompts containing each class name, e.g., \"a video of a person [CLS]\", where \"[CLS]\" denotes a class name." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.47, + 0.744 + ], + "angle": 0, + "content": "The language model \\( G_{\\mathrm{L}}(\\cdot) \\) projects the class names into embeddings, denoted as \\( \\{z_c^{\\mathrm{L}}\\}_{c = 1}^{C} \\). On the other hand, assuming a test video of \\( K \\) frames \\( \\mathbf{X} = \\{\\mathbf{x}_k\\}_{k = 1}^K \\), we extract the visual feature representation for each frame, i.e. \\( \\mathbf{Z}^{\\mathrm{V}} = \\{z_k^{\\mathrm{V}}\\}_{k = 1}^K \\) using the vision encoder \\( G_{\\mathrm{V}}(\\cdot) \\). CLIP outputs a probability distribution, where the probability of the \\( k \\)-th frame to belong to the \\( c \\)-th class is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.754, + 0.47, + 0.793 + ], + "angle": 0, + "content": "\\[\np _ {c} (\\mathbf {x} _ {k}) = \\frac {\\exp \\left(< z _ {c} ^ {\\mathrm {L}} , z _ {k} ^ {\\mathrm {V}} > / \\tau\\right)}{\\sum_ {c ^ {\\prime} = 1} ^ {C} \\exp \\left(< z _ {c ^ {\\prime}} ^ {\\mathrm {L}}, z _ {k} ^ {\\mathrm {V}} > / \\tau\\right)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.855 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a temperature and \\(<\\cdot, \\cdot>\\) is the inner-product (or cosine similarity) operator. The frame-wise predictions are then aggregated by simply averaging the class probabilities estimated for each frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "With respect to inference with CLIP, it is regarded as zero-shot inference if a given dataset has not been explicitly used for training the CLIP model. In other words," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.138 + ], + "angle": 0, + "content": "zero-shot inference refers to the generalization to unseen datasets [31], and must not be confused with generalization to strictly unseen objects [17]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.146, + 0.892, + 0.178 + ], + "angle": 0, + "content": "4.3. Domain Adaptation with Large Language-Vision models (DALL-V)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.893, + 0.335 + ], + "angle": 0, + "content": "In this work, we propose Domain Adaptation with Large Language-Vision models (DALL-V) as a SFVUDA method. As the name suggests, it leverages the rich world prior from the LLVMs, CLIP in particular, to help mitigate the domain gap. Recall that CLIP has been trained on image-text pairs and not video-caption pairs. On the other hand, the source model can offer complementary information as it has been trained supervisedly on the source videos. Thus, we aspire to jointly leverage these two reservoirs of information in a simple yet effective manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.895, + 0.533 + ], + "angle": 0, + "content": "Our DALL-V mainly operates in two stages. The first stage, called Target adaptation (see Sec. 4.3.1), consists in pseudo-labelling the unlabelled target videos with CLIP and training a target-specific model with the pseudo-labelled data. In this way the target-specific model can benefit from the rich general purpose knowledge of CLIP and at the same time discover patterns that are inherent to the target domain. The second stage, called Ensemble distillation (see Sec. 4.3.2), focuses on ensembling the source model and target model into a smaller network, with knowledge distillation [8], which is then finally used during inference. Before we dive into the details of our proposed method, we describe how we carry out source training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.719 + ], + "angle": 0, + "content": "Source Pre-Training. Opposed to the previous work AT-CoN, which fine-tunes all the source model weights using the source dataset, we propose an alternative approach. To prevent the source model to be biased towards the source dataset, we propose to initialize the source model with CLIP pre-trained weights, and fine-tune only a small set of parameters. To fine-tune CLIP on the source data, we introduce an adapter network [6], a small multi-layer perceptron network, that is appended on top of the vision encoder of CLIP. Source pre-training essentially consists in updating the weights of the adapter using a supervised training objective, while keeping the CLIP vision encoder frozen." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.856 + ], + "angle": 0, + "content": "Formally, the vision encoder of CLIP \\( G_{\\mathrm{V}}(\\cdot) \\) is extended with an adapter \\( A(\\cdot): \\mathcal{R}^d \\to \\mathcal{R}^d \\) on top, where \\( d \\) is the output dimension of \\( G_{\\mathrm{V}}(\\cdot) \\). We construct textual descriptions \\( t \\) for the video sequences using a set of 16 manually-designed prompts [40], containing templates such as \"a video of [CLS]\", \"can you recognize the action of [CLS]?\", and so on, with \\( y \\) being the class name [CLS] (refer to the Supplement for the complete list of prompts)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We adopt the training objective of ActionCLIP [40] to train the source model. In details, we first compute the ground-truth similarity scores \\( q(\\mathbf{x}) \\) for each input frame \\( \\mathbf{x} \\)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "10310" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.141, + 0.429, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.309, + 0.468, + 0.336 + ], + "angle": 0, + "content": "(a) Source/target adapter \\((A^{\\mathrm{S}} / A^{\\mathrm{T}})\\) training. For the target domain, \\(q\\) is constructed by pseudo-labelling with CLIP (ViT-B/32)." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.102, + 0.847, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.309, + 0.857, + 0.335 + ], + "angle": 0, + "content": "(b) Ensemble distillation. We use \\(A^{\\mathrm{S}}, A^{\\mathrm{T}}\\), and CLIP (ViT-B/32) as teachers and train a student adapter \\(\\overline{A}\\) on a CLIP (RN50) model." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.892, + 0.377 + ], + "angle": 0, + "content": "Figure 3: Overview of the pipeline of DALL-V. The \\(\\mathbb{O}\\) means the network is trainable, while the \\(\\text{念}\\) means the network is frozen. We denote with bar the student components to distinguish them from the teacher." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.388, + 0.47, + 0.48 + ], + "angle": 0, + "content": "with 0 in the positions corresponding to negative pairs in the mini-batch and 1 for positive pairs, where a pair denotes the frame-prompt tuple. To train the network (essentially just the adapter \\( A^{\\mathrm{S}} \\); language and vision encoders being frozen) we use the Kullback-Leibler (KL) divergence loss between \\( \\mathbf{p}(\\mathbf{X}) \\) (see Eq. (1)) and \\( \\mathbf{q}(\\mathbf{X}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.492, + 0.47, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a d p a t e r}} ^ {\\mathrm {S}} = \\mathbb {E} _ {(\\mathbf {X}, t) \\sim \\mathcal {D} ^ {\\mathrm {s}}} [ \\mathrm {K L} (\\mathbf {p} (\\mathbf {X}), \\mathbf {q} (\\mathbf {X})) ] \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.529, + 0.257, + 0.545 + ], + "angle": 0, + "content": "4.3.1 Target adaptation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.718 + ], + "angle": 0, + "content": "To adapt CLIP on the target domain, we adopt a similar approach to that outlined for the source domain, with one key difference: the absence of ground-truth labels for the target domain. While it is a common practice among the source-free methods to utilize the source model to pseudolabel the unlabelled target data [20], we refrain from such an approach, and instead utilize the original CLIP model (i.e., without any source adapter \\( A^{\\mathrm{S}} \\)) for pseudo-labelling. Despite learning fewer parameters with the adapter, the source model can still be biased towards the source, and hence, yielding noisy pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.469, + 0.841 + ], + "angle": 0, + "content": "To pseudo-label (PL), we utilize zero-shot CLIP inference (described in Sec. 4.2), with the same collection of textual descriptions used during the source fine-tuning stage, to assign a PL to each sample in the target domain. To this end, we use a confidence threshold for filtering out unreliable PLs (see Supplement for details). After the thresholding, we end up with a subset \\(\\tilde{\\mathcal{D}}^{\\mathrm{T}} = \\{\\mathbf{X}_i^{\\mathrm{T}}\\}_{i=1}^{m'} \\subset \\mathcal{D}^{\\mathrm{T}}\\), containing PLs as \\(\\{\\tilde{y}_i\\}_{i=1}^{m'}\\) with \\(m'\\) denoting the filtered samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Following the source fine-tuning procedure, we append an adapter \\( A^{\\mathrm{T}}(\\cdot): \\mathcal{R}^{d} \\to \\mathcal{R}^{d} \\) to the vision encoder of CLIP and minimize the following loss, while keeping CLIP frozen:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.401, + 0.892, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a d a p t e r}} ^ {\\mathrm {T}} = \\mathbb {E} _ {(\\mathbf {X}, t) \\sim \\tilde {\\mathcal {D}} ^ {\\mathrm {T}}} [ \\operatorname {K L} (\\mathbf {p} (\\mathbf {X}), \\mathbf {q} (\\mathbf {X})) ] \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.435, + 0.702, + 0.449 + ], + "angle": 0, + "content": "4.3.2 Ensemble distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.684 + ], + "angle": 0, + "content": "With three different models at our disposal: the source, target and the original CLIP, we can distill information from these three models into a single network. Knowledge Distillation [8] presents two important advantages: (i) allows to generate more reliable and informative pseudo-labels, and (ii) distillation into a smaller network reduces inference time without sacrificing much performance. For this second point, we explicitly use CLIP ViT-B/32 as teacher (and as backbone for the source and target adapters) and CLIP RN50 as student. Since we use an ensemble of three models to distill information, we call this step ensemble distillation. Now, with predictions available from three different models, the question is how to distill them together. A simple approach could be to average the three probability distributions as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.696, + 0.892, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} ^ {\\text {e n s}} (\\mathbf {X}) = \\frac {1}{3} \\left(\\underbrace {\\mathbf {p} (\\mathbf {X})} _ {\\text {Z S C L I P}} + \\underbrace {\\mathbf {p} ^ {\\mathrm {S}} (\\mathbf {X})} _ {\\text {S o u r c e a d p a t e r}} + \\underbrace {\\mathbf {p} ^ {\\mathrm {T}} (\\mathbf {X})} _ {\\text {T a r g e t a d p a t e r}}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.892, + 0.864 + ], + "angle": 0, + "content": "Whereas, another approach could consist in carrying out majority voting, where each of the three heads votes for a class, and the class with two or more votes becomes the new PL. Following [8], we utilize both kinds of PLs for distillation, using a standard multi-class classification loss (or cross-entropy loss \\(\\mathcal{L}_{\\mathrm{CE}}\\)) and a discrepancy loss. The discrepancy loss corresponding to the ensembled PL in Eq. (4) is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.874, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s c r e p a n c y}} = \\mathbb {E} _ {(\\mathbf {X}, t) \\sim \\mathcal {D} ^ {\\tau}} \\left[ \\mathrm {K L} \\left(\\frac {\\mathbf {p} ^ {\\text {s t u d e n t}} (\\mathbf {X})}{\\tau}, \\frac {\\mathbf {p} ^ {\\text {e n s}} (\\mathbf {X})}{\\tau}\\right) \\right] \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10311" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.121 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}^{\\mathrm{student}}\\) denotes the student network's probability distribution, which is matched with that of \\(\\mathbf{p}^{\\mathrm{ens}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.123, + 0.453, + 0.138 + ], + "angle": 0, + "content": "We then balance the contribution of the two losses as:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.157, + 0.469, + 0.172 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t i l l}} = \\alpha * \\mathcal {L} _ {c e} + (1 - \\alpha) * \\mathcal {L} _ {d} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.184, + 0.47, + 0.245 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is set according to the value proposed in [8]. Note that, we do not fine-tune all the student weights, but instead append an adapter on top of the frozen backbone and fine-tune only the adapter \\(\\overline{A}\\) using Eq. (6)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.261, + 0.21, + 0.277 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.286, + 0.265, + 0.303 + ], + "angle": 0, + "content": "5.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.311, + 0.47, + 0.55 + ], + "angle": 0, + "content": "Datasets and settings. We present an extensive experimental evaluation on three standard benchmarks for SFVUDA. In particular, we report results on Daily-DA [44], which comprises 18,949 videos from 8 classes, and it is built from 4 original video action recognition datasets, namely HMDB51 [15], ARID [43], MIT [26] and Kinetics [12]. Additionally, we evaluate our framework on UCF-HMDBfull [1], a benchmark comprising 3,209 videos divided in 12 action categories from the HMDB51 [15] and UCF101 [38] action recognition datasets. Note that the first benchmark, Daily-DA, poses a more significant challenge when compared to the latter since it comprises videos with very different lighting conditions across domains. Last, we test our method on Sports-DA [45], a benchmark consisting of three datasets (i.e., UCF101 [38], Sports-1M [11], and Kinetics [12]), with 40,718 videos and 23 classes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.553, + 0.47, + 0.674 + ], + "angle": 0, + "content": "Implementation details. We implement our adapters as two-layers perceptrons, following [6]. In accordance with [40], we use AdamW optimizer [24] with a learning rate of 0.01 and a weight decay of 0.2. The temperature \\(\\tau\\) is set to 2.0 for all datasets. We trained all our models for 30 epochs. For a single experiment, we have used either 4 Tesla V100 or 2 RTX A6000 GPUs. Further implementation details can be found in the Supplement." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Baselines and competitors. We compare our method with a selection of standard baseline methods for VUDA and SFVUDA, which include TRN [49], DANN [5], MK-MDD [23], \\(\\mathrm{TA}^3\\mathrm{N}\\) [1], SFDA [13], SHOT [20], SHOT++ [21], MA [19], BAIT [48] and CPGA [30]. Additionally, we report the scores of the state-of-the-art competitors ATCoN [44] and EXTERN [46]. We also report the results for a set of CLIP-based baselines of our design to provide a more solid context for assessing our performance. In particular, CLIP (RN50) and CLIP (ViT-B/32) indicate the scores obtained in a zero-shot fashion with the CLIP model, with ResNet-50 and ViT-B/32 as backbones, respectively. Finally, we report the lower and upper bounds obtained by the source-only and target-supervised models, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.826, + 0.108 + ], + "angle": 0, + "content": "5.2. Comparison with the State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.892, + 0.263 + ], + "angle": 0, + "content": "We report our experimental validation scores in Tab. 1, Tab. 2, and Tab. 3, obtained on the Daily-DA, the UCF-HMDB \\(f_{ull}\\), and the Sports-DA benchmarks respectively. For all the benchmarks, the lower bound refers to the out-of-the-box performance of our source-trained model on the target data, i.e., the CLIP backbone and the source adapter \\(A^{\\mathrm{S}}\\). Analogously, the upper bound is the oracle performance obtained after training supervisedly on the target data, i.e., assuming knowledge of the target labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Regarding the more challenging Daily-DA dataset, it is possible to observe in Tab. 1 that our proposed distillation method DALL-V (RN50) obtains the best average score across the 12 settings among the competitors and baselines sharing the same backbone, with a \\(5\\%\\) gain on the second best score (CLIP (RN50)). It is also visible that the best score is achieved in 9 out 12 total settings, proving our proposed distillation-based method to be significantly effective when addressing the SFVUDA task. Notably, it emerges that the scores achieved when addressing the ARID dataset \\((\\rightarrow \\mathrm{A})\\) as target domain are visibly lower than those reported for the other settings. We associate this behavior to the challenging visual features of this particular benchmark, characterized by being shot in a low illumination environment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.892, + 0.633 + ], + "angle": 0, + "content": "As for the small-scale UCF-HMDBfull dataset, we can observe significantly higher absolute accuracy scores on both the involved SFVUDA directions, indicating that the benchmark is significantly closer to saturation for this task. Nonetheless, it emerges from the table that our proposed method achieves a state-of-the-art average accuracy score across the two datasets (+0.6% on the best competitor) and the best score for the HMDB→UCF direction (+1.2% on the second best)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Finally, as shown in Tab. 3, our method achieves comparable results with EXTERN [46] on Sports-DA, surpassing it on three settings, and achieving the same results on another. On the two settings with UCF101 as target (i.e., \\(\\mathrm{K} \\rightarrow \\mathrm{U}\\) and \\(\\mathrm{S} \\rightarrow \\mathrm{U}\\)), our method falls behind by approximately 6 percentage points, resulting on an average score across the six scenarios slightly lower than the state-of-the-art (i.e., \\(-0.9\\%\\)). However, DALL-V still achieves an average accuracy of \\(+8.5\\%\\) w.r.t. ATCoN, further demonstrating the effectiveness of our approach for SFVUDA. Interestingly, on Sports-DA, our lower bound already achieves competitive results with other SFVUDA methods, surpassing ATCoN by \\(+2.7\\%\\) on average. The table additionally shows that CLIP zero-shot is an extremely strong baseline on the benchmark, achieving comparable results with EXTERN, i.e., \\(-2.0\\%\\) with RN50, and \\(+2.6\\%\\) with the ViT-B backbone." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10312" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.09, + 0.894, + 0.367 + ], + "angle": 0, + "content": "
MethodAccuracy (%)
K→AK→HK→MM→AM→HM→KH→AH→MH→KA→HA→MA→KAvg.
ZSTRN [49]20.936.729.022.143.753.113.822.037.117.214.724.427.9
Lower bound15.647.935.734.744.661.617.525.545.114.615.517.831.3
CLIP (RN50) [31]30.550.042.230.550.062.930.542.262.950.042.262.946.4
CLIP (ViT-B/32) [31]31.349.646.031.349.665.431.346.065.449.646.065.448.1
UDADANN [5]21.237.521.722.843.358.814.229.538.220.119.727.029.5
MK-MDD [23]21.736.224.021.050.458.520.325.733.818.718.026.129.5
TA3N [1]19.937.731.521.643.055.514.425.738.414.915.623.428.5
SFUDASFDA [13]12.644.927.516.035.249.213.124.224.916.313.225.225.2
SHOT [20]12.044.629.515.336.751.013.624.221.217.114.024.325.3
SHOT++ [21]12.640.828.714.941.746.316.022.233.115.412.521.824.4
MA [19]12.845.830.017.737.453.512.925.022.216.715.224.326.1
BAIT [48]12.745.730.016.939.653.013.625.521.215.714.525.526.2
CPGA [30]13.146.030.718.139.255.113.126.225.519.216.526.726.5
SFVUDAATCoN [44]17.248.232.527.247.357.717.930.748.526.717.231.033.5
EXTERNAL [46]23.955.835.218.153.768.126.240.757.626.218.251.439.6
DALL-V (ours)24.052.547.024.065.478.124.047.076.757.945.775.051.4
Upper bound26.970.461.526.970.488.926.961.588.970.461.588.961.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.375, + 0.893, + 0.42 + ], + "angle": 0, + "content": "Table 1: Validation accuracy for Daily-DA. Bold indicates best, underline represents best with same backbone as baseline (i.e. ResNet50). Lower bound indicates a source adapter and Upper bound a target adapter, both trained supervised on CLIP (RN50)." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.444, + 0.413, + 0.695 + ], + "angle": 0, + "content": "
MethodAccuracy (%)
H→UU→HAvg.
ZSTRN [49]72.872.172.4
Lower bound71.676.173.8
CLIP (RN50) [31]81.086.083.5
CLIP (ViT-B/32) [31]90.389.189.7
UDADANN [5]74.475.174.8
MK-MMD [23]74.779.777.2
TA3N [1]78.184.881.5
SFUDASFDA [13]69.875.072.4
SHOT [20]74.474.474.4
SHOT++ [21]71.168.169.6
MA [19]74.467.370.9
BAIT [48]75.376.375.8
CPGA [30]75.868.172.0
SFVUDAATCoN [44]85.379.782.5
EXTERNAL [46]91.988.990.4
DALL-V (ours)93.188.991.0
Upper bound93.791.492.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.697, + 0.47, + 0.775 + ], + "angle": 0, + "content": "Table 2: Validation accuracy for UCF-HMDBfull. Bold indicates best, underline represents best with same backbone as baseline (i.e. ResNet50). Lower bound indicates a source adapter and Upper bound a target adapter, both trained supervised on CLIP (RN50)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.799, + 0.248, + 0.816 + ], + "angle": 0, + "content": "5.3. Ablation analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We report in Tab. 4 the intermediate accuracy scores achieved at different steps of our proposed source-free pipeline for the Daily-DA and UCF-HMDBfull. In particular, we report the performance of the source supervised model and the target unsupervised one, also reporting the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.568 + ], + "angle": 0, + "content": "final model and the zero-shot version of CLIP for comparison. Finally, we report the scores obtained by assembling the predictions of the three aforementioned models. The \\( A^{\\mathrm{S}} \\) and \\( A^{\\mathrm{T}} \\) notation indicates the usage of the source and target adapters on top of the frozen CLIP model, that are trained on the respective domains as described in detail in 4.3. Next we discuss the individual contributions. Additional ablation analyses are available in the Supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.893, + 0.734 + ], + "angle": 0, + "content": "Effectiveness of finetuning LLVM on the target. Tab. 4 clearly shows that fine-tuning our model in an unsupervised fashion on the target domain is significantly effective with respect to the target validation accuracy. The tables show indeed a gain of \\(2.0\\%\\) on Daily-DA and \\(1.1\\%\\) on UCF-HMDB \\(f_{ull}\\). Additionally, it is possible to observe a further gain on Daily-DA when ensembling the target model (which gains \\(+2.7\\%\\) by itself) with CLIP, achieving a \\(+3.6\\%\\). From these results, it emerges that fine-tuning LLVMs on the target domain, even in an unsupervised manner, is effective for the SFVUDA task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of ensembles on LLVM and domain-specific networks. In the last line of Tab. 4, it is possible to observe that ensembling predictions from the three distinct models turns out to yield the best average target accuracy score on both the considered benchmarks, resulting in a gain of \\(+5.7\\%\\) on Daily-DA and of \\(+2.9\\%\\) on UCF-HMDBfull w.r.t. the CLIP baseline. These results and those reported in the previous two paragraphs represent further proof that the three models are effectively complementary, and each contributes with additional discriminative knowledge to the resulting model. Conversely, CLIP positively contributes to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10313" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.244, + 0.089, + 0.734, + 0.362 + ], + "angle": 0, + "content": "
MethodAccuracy (%)
K→UK→SS→US→KU→KU→SAvg.
ZSTRN [49]86.466.985.371.049.343.367.0
Lower bound85.479.584.478.267.264.376.5
CLIP (RN50) [31]83.479.983.480.480.479.981.2
CLIP (ViT-B/32) [31]90.082.490.085.185.182.485.8
UDADANN [5]88.075.085.773.465.955.173.8
MK-MMD [23]90.267.990.973.666.155.674.0
TA3N [1]90.368.693.072.663.654.173.7
SFUDASFDA [13]86.160.085.468.055.843.666.5
SHOT [20]91.264.988.872.053.943.669.1
SHOT++ [21]90.063.188.070.344.740.966.2
MA [19]91.065.987.871.960.739.469.5
BAIT [48]92.366.688.372.857.244.770.3
CPGA [30]89.466.386.572.555.244.569.1
SEVUDAATCoN [44]93.669.790.676.065.247.973.8
EXTERNAL [46]93.773.895.482.281.272.783.2
DALL-V (ours)88.077.788.882.381.275.982.3
Upper bound93.488.393.485.685.688.389.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.372, + 0.895, + 0.418 + ], + "angle": 0, + "content": "Table 3: Validation accuracy for Sports-DA. Bold indicates best, underline represents best with same backbone as baseline (i.e. ResNet50). Lower bound indicates a source adapter and Upper bound a target adapter, both trained supervised on CLIP (RN50)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.434, + 0.892, + 0.582 + ], + "angle": 0, + "content": "
MethodAccuracy (%)
K→AnyM→AnyDaily-DAA→AnyAvg.UCF-HMDBfull
H→AnyA→AnyH→UU→HAvg.
CLIP42.348.747.653.748.190.389.189.7
AS40.0 [-2.3]55.7 [+7.0]35.3 [-12.4]35.0 [-18.7]41.5 [-6.6]91.0 [+0.7]80.5 [-8.6]85.8 [-3.9]
AT44.4 [+2.1]51.6 [+2.9]48.5 [+0.9]58.4 [+4.7]50.8 [+2.7]90.5 [+0.2]91.1 [+2.0]90.8 [+1.1]
CLIP + AS45.1 [+2.8]57.8 [+9.1]40.0 [-7.6]50.8 [-2.9]48.4 [+0.3]93.1 [+2.8]87.2 [-1.9]90.2 [+0.5]
CLIP + AT43.8 [+1.5]50.9 [+2.2]48.2 [+0.6]57.4 [+3.7]51.7 [+3.6]89.3 [-1.0]91.9 [+2.8]90.6 [+0.9]
CLIP + AS + AT46.2 [+3.9]57.3 [+8.6]49.5 [+1.9]62.1 [+8.4]53.8 [+5.7]94.9 [+4.6]90.3 [+1.2]92.6 [+2.9]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.591, + 0.895, + 0.639 + ], + "angle": 0, + "content": "Table 4: Ablation on the relative improvement of the different modules of our proposed DALL-V framework w.r.t. CLIP (ViT-B/32) on both considered benchmarks. Bold indicates best. The + operator indicates assembling. Due to limited space, results are aggregated by source dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.649, + 0.473, + 0.786 + ], + "angle": 0, + "content": "the final score thanks to its well-known ability to generalize to unseen domains in a zero-shot fashion; on the other hand, the tuning process of domain-specific adapters on the source and target domains, although not sufficient by itself, proves beneficial thanks to its ability to capture specific features of the involved domains and the associated label sets. Note that we still opted for distilling this knowledge to a ResNet-based model in order to be directly comparable with the competitors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Effectiveness of using multiple templates. As a further insight into the individual contributions of the different components of our framework, we propose in Fig. 5, a simple analysis on how the number of templates for the text input at inference time impacts on the accuracy score on the target validation set of the two considered benchmarks. We observe that increasing the number of templates from just" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.895, + 0.727 + ], + "angle": 0, + "content": "a single one to 12 (following ActionCLIP [40]) is significantly beneficial with respect to the final accuracy score. For even higher number of templates, the performance of our model reaches a plateau. Note that for the choice of the number of template we opted for following [40]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.773, + 0.895, + 0.911 + ], + "angle": 0, + "content": "Modeling the feature space. We plot in Fig. 4 an UMAP [25] chart of the features produced by the zero-shot CLIP (RN50) model, the source model, and our DALL-V model, on the UCF-HMDBfull benchmark, showing the features belonging to different classes in different colors. The chart clearly highlights the ability of the proposed DALL-V method to produce well-separated clusters for different action categories, thanks to the effective combination of the complementary knowledge of intermediate models." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "10314" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.106, + 0.468, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.215, + 0.332, + 0.227 + ], + "angle": 0, + "content": "(a) HMDB \\(\\rightarrow\\) UCF" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.104, + 0.872, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.215, + 0.741, + 0.227 + ], + "angle": 0, + "content": "(b) UCF \\(\\rightarrow\\) HMDB" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.895, + 0.279 + ], + "angle": 0, + "content": "Figure 4: UMAP visualisation of the feature space on \\( \\mathbf{UCF - HMDB}_{full} \\) for the zero-shot CLIP (RN50), source and final DALL-V models. The chart shows the ability of our proposed DALL-V to efficiently cluster the action categories in the features space by combining knowledge from the other two models." + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.29, + 0.446, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.471, + 0.472, + 0.517 + ], + "angle": 0, + "content": "Figure 5: Sensitivity study on the number of templates used for the text input w.r.t. the accuracy score, on Daily-DA and UCF-HMDB \\(f_{ull}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.536, + 0.182, + 0.551 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.47, + 0.724 + ], + "angle": 0, + "content": "Note that CLIP has been trained on web-crawled image-text pairs that are not made publicly available by OpenAI [35]. Due to its black-box nature, equally trusting CLIP predictions, as much as the source model, may elicit unpredictable behaviour, especially in safety critical applications. Thus, appropriate caution must be exercised. Moreover, ours being an empirical work, we can not provide any theoretical guarantee if using LLVMs, such as CLIP, will always result in better performance over traditional SFVUDA approaches. Thus, as future work, we plan to extensively evaluate using more open-source alternatives." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.205, + 0.755 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We presented DALL-V, a SFVUDA method based on a simple but novel approach to SFVUDA driven by the intuition of combining the complementary information derived from domain-specific simple models and the powerful CLIP-based LLVMs trained on world knowledge. We provided motivation for such approach, and reported an extensive evaluation on three standard benchmarks for VUDA, pursued in our case for the source-free scenario. We compare our performance with existing methods, as well as with" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.287, + 0.893, + 0.332 + ], + "angle": 0, + "content": "a selection of CLIP-based baselines, showing that our proposed frameworks achieves state-of-the-art performance on both considered benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.35, + 0.895, + 0.532 + ], + "angle": 0, + "content": "Acknowledgments. We acknowledge the support of the MUR PNRR project FAIR - Future AI Research (PE00000013) funded by the NextGenerationEU. E.R. is partially supported by the PRECRISIS, funded by the EU Internal Security Fund (ISFP-2022-TFI-AG-PROTECT-02-101100539), the EU project SPRING (No. 871245), and the by the PRIN project LEGO-AI (Prot. 2020TA3K9N). The work was carried out in the Vision and Learning joint laboratory of FBK and UNITN. This paper has been further supported by the French National Research Agency (ANR-20-CE23-0027), the EU Horizon 2020 Research and Innovation program under grant agreement No 957337." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.542, + 0.597, + 0.557 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.567, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[1] Min-Hung Chen, Zsolt Kira, Ghassan AlRegib, Jaekwon Yoo, Ruxin Chen, and Jian Zheng. Temporal attentive alignment for large-scale video domain adaptation. In ICCV, 2019. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.623, + 0.894, + 0.678 + ], + "angle": 0, + "content": "[2] Victor G Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-head contrastive domain adaptation for video action recognition. In WACV, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.679, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[3] Christoph Feichtenhofer, Axel Pinz, and Richard P Wildes. Spatiotemporal multiplier networks for video action recognition. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[4] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.762, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[5] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation, 2015. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[6] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[7] Muhammad Ghifary, W Bastiaan Kleijn, Mengjie Zhang, David Balduzzi, and Wen Li. Deep reconstruction-classification networks for unsupervised domain adaptation. In ECCV, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.567, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10315" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "[8] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv, 2015. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.122, + 0.468, + 0.149 + ], + "angle": 0, + "content": "[9] Yunzhong Hou and Liang Zheng. Visualizing adapted knowledge in domain transfer. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.152, + 0.468, + 0.207 + ], + "angle": 0, + "content": "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.21, + 0.468, + 0.264 + ], + "angle": 0, + "content": "[11] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, pages 1725–1732, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.267, + 0.468, + 0.334 + ], + "angle": 0, + "content": "[12] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.338, + 0.468, + 0.379 + ], + "angle": 0, + "content": "[13] Youngeun Kim, Donghyeon Cho, Kyeongtak Han, Priyadarshini Panda, and Sungeun Hong. Domain adaptation without source data, 2020. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.381, + 0.468, + 0.408 + ], + "angle": 0, + "content": "[14] Yu Kong and Yun Fu. Human action recognition and prediction: A survey. IJCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.411, + 0.468, + 0.45 + ], + "angle": 0, + "content": "[15] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. Hmdb: A large video database for human motion recognition. In ICCV, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.454, + 0.468, + 0.495 + ], + "angle": 0, + "content": "[16] Vinod K Kurmi, Venkatesh K Subramanian, and Vinay P Namboodiri. Domain impression: A source data free domain adaptation method. In WACV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.498, + 0.468, + 0.538 + ], + "angle": 0, + "content": "[17] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In CVPR, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.541, + 0.468, + 0.581 + ], + "angle": 0, + "content": "[18] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.585, + 0.468, + 0.625 + ], + "angle": 0, + "content": "[19] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.629, + 0.468, + 0.681 + ], + "angle": 0, + "content": "[20] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.686, + 0.468, + 0.739 + ], + "angle": 0, + "content": "[21] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. TPAMI, 44(11):8602-8617, 2021. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.743, + 0.468, + 0.783 + ], + "angle": 0, + "content": "[22] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.786, + 0.468, + 0.826 + ], + "angle": 0, + "content": "[23] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I. Jordan. Learning transferable features with deep adaptation networks, 2015. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.83, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.859, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[25] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction, 2018. 8" + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[26] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, and Aude Oliva. Moments in time dataset: one million videos for event understanding. TPAMI, 42(2):502-508, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[27] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[28] Boxiao Pan, Zhangjie Cao, Ehsan Adeli, and Juan Carlos Niebles. Adversarial cross-domain action recognition with co-attention. In AAAI CAI, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[29] Preksha Pareek and Ankit Thakkar. A survey on video-based human action recognition: recent updates, datasets, challenges, and applications. Artificial Intelligence Review, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[30] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[33] Aadarsh Sahoo, Rutav Shah, Rameswar Panda, Kate Saenko, and Abir Das. Contrast and mix: Temporal contrastive video domain adaptation with background mixing. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv, 2021. 1, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[36] Javier Selva, Anders S Johansen, Sergio Escalera, Kamal Nasrollahi, Thomas B Moeslund, and Albert Clapés. Video transformers: A survey. arXiv, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[37] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[38] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[39] Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. In Artificial Intelligence, volume 30, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[40] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv, 2021. 2, 3, 4, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10316" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[41] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Conference on Empirical Methods in Natural Language Processing, pages 6787-6800, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[42] Yuecong Xu, Haozhi Cao, Kezhi Mao, Zhenghua Chen, Lihua Xie, and Jianfei Yang. Aligning correlation information for domain adaptation in action recognition. Neural Networks and Learning Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.302 + ], + "angle": 0, + "content": "[43] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Arid: A new dataset for recognizing action in the dark. In International Workshop on Deep Learning for Human Activity Recognition, pages 70-84. Springer, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.47, + 0.358 + ], + "angle": 0, + "content": "[44] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, and Zhenghua Chen. Source-free video domain adaptation by learning temporal consistency for action recognition. In ECCV, 2022. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[45] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, Rui Zhao, and Zhenghua Chen. Multi-source video domain adaptation with temporal attentive moment alignment. arXiv, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[46] Yuecong Xu, Jianfei Yang, Haozhi Cao, Min Wu, Xiaoli Li, Lihua Xie, and Zhenghua Chen. Leveraging endo- and exotemporal regularization for black-box video domain adaptation, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.474, + 0.47, + 0.528 + ], + "angle": 0, + "content": "[47] Jinyu Yang, Weizhi An, Sheng Wang, Xinliang Zhu, Chaochao Yan, and Junzhou Huang. Label-driven reconstruction for domain adaptation in semantic segmentation. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[48] Shiqi Yang, Yaxing Wang, Joost Weijer, and Luis Herranz. Unsupervised domain adaptation without source data by casting a bait, 2020. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[49] Bolei Zhou, Alex Andonian, Aude Oliva, and Antonio Torralba. Temporal relational reasoning in videos, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[50] Yi Zhu, Xinyu Li, Chunhui Liu, Mohammadreza Zolfaghari, Yuanjun Xiong, Chongruo Wu, Zhi Zhang, Joseph Tighe, R Manmatha, and Mu Li. A comprehensive study of deep video action recognition. arXiv, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10317" + } + ] +] \ No newline at end of file diff --git a/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/c596fbdc-cd07-43b2-9a08-9301d722c491_origin.pdf b/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/c596fbdc-cd07-43b2-9a08-9301d722c491_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0d8c963c358313ee9ab84a8eed85ac55fba868f0 --- /dev/null +++ b/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/c596fbdc-cd07-43b2-9a08-9301d722c491_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24d56237d1934185816d71f16cb62e111ef3e5aff862742f41bb582cb80c7e71 +size 614799 diff --git a/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/full.md b/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a17a7ffced4b3ae810a9e35eb87193ff03f8f4ac --- /dev/null +++ b/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/full.md @@ -0,0 +1,285 @@ +# The Unreasonable Effectiveness of Large Language-Vision Models for Source-free Video Domain Adaptation + +Giacomo Zara1*, Alessandro Conti1*, Subhankar Roy3, Stéphane Lathuilière3, Paolo Rota1, Elisa Ricci1,2 + +1University of Trento, Italy 2Fondazione Bruno Kessler, Italy + +$^{3}$ LTCI, Télécom Paris, Institut polytechnique de Paris, France + +{giacomo.zara, alessandro.conti-1}@unitn.it + +# Abstract + +Source-Free Video Unsupervised Domain Adaptation (SFVUDA) task consists in adapting an action recognition model, trained on a labelled source dataset, to an unlabelled target dataset, without accessing the actual source data. The previous approaches have attempted to address SFVUDA by leveraging self-supervision (e.g., enforcing temporal consistency) derived from the target data itself. In this work, we take an orthogonal approach by exploiting "web-supervision" from Large Language-Vision Models (LLVMs), driven by the rationale that LLVMs contain a rich world prior surprisingly robust to domain-shift. We showcase the unreasonable effectiveness of integrating LLVMs for SFVUDA by devising an intuitive and parameter-efficient method, which we name Domain Adaptation with Large Language-Vision models (DALL-V), that distills the world prior and complementary source model information into a student network tailored for the target. Despite the simplicity, DALL-V achieves significant improvement over state-of-the-art SFVUDA methods. + +# 1. Introduction + +Video analysis tasks, such as action recognition, have long been investigated in computer vision, due to the numerous applications, ranging from video surveillance to social robotics [50, 14, 29]. Major progress has been made in the last decade with the development of specialized deep architectures, such as 3D CNNs [4, 3] and Video Transformers [36], trained on large-scale annotated datasets. However, obtaining sufficient labelled training videos for real-world scenarios can be very costly and time-consuming. In order to alleviate the burden of annotating large scale datasets Video-based Unsupervised Domain + +![](images/c21b766958fb0a13750f29c7ec9c41d40ab534edbe973eea39260c5ae5a5069c.jpg) +Figure 1: Performance over time by various methods on the Daily-DA video benchmark. A pre-trained LLVM (e.g., CLIP [31]) is surprisingly better when compared with Unsupervised Domain Adaptation (UDA), Source Free UDA (SFUDA) and Source Free Video-based UDA (SFVUDA) methods. Our proposed DALL-V that is built on top of LLVM successfully outperforms all existing baselines. + +Adaptation (VUDA) [1, 28, 2] have been introduced. The VUDA methods are derived from the common idea of transferring knowledge from a labelled source domain to an unlabelled target domain. While considering different strategies for adaptation, these approaches have all shown significant improvements in the robustness of learnt video models without requiring annotated target data. + +In the last few years, the field of computer vision has also witnessed the emergence of a new generation of powerful deep architectures, trained on mammoth internet-scale image-text datasets [35]. These models, commonly known as foundation models [32, 37, 31, 10] or Large LanguageVision Models (LLVM), have achieved outstanding performance, and have become a cornerstone of modern computer vision research. In particular, recent works such as CLIP [31] or FLAVA [37] have shown that rich visual representations for images can be learned from natural language descriptions as a form of supervision. These pre-trained + +![](images/7a6b9e8506112f283489cecbbdade7cbabf065c58eae3fd1a31eb2818d60f4e4.jpg) +(a) Traditional SFVUDA + +![](images/dd153a23f025253233b9242e0ce2ee4482b070514fc0f11da8ff0c03b29b8809.jpg) +(b) Zero Shot CLIP Inference + +![](images/8fdc00db30f2dedfdbeb9b7fd781eea1fdbf3d8700cbc0336d5646e609037dfd.jpg) +(c) Our proposed SFVUDA +Figure 2: (a) In traditional SFVUDA a source-trained model is adapted to the unlabelled target dataset using self-supervision such as temporal consistency [44, 46] among frames. (b) LLVM (e.g., CLIP [31]) uses cosine similarity between the feature and language representation to predict the most probable class in a zero-shot (ZS) manner. (c) Our proposed SFVUDA solution distills the ZS-CLIP and source model predictions, while introducing very few learnable parameters, namely adapters. The $\hat{\mathbb{C}}$ means the network is trainable, while the $\hat{\mathbb{C}}$ means the network is frozen. + +LLVMs are now publicly available and can be easily integrated into any recognition system. Lately, researchers have also shown that the LLVMs can be applied with success to the video domain and specifically to the supervised action recognition task [40, 41]. + +Despite the significant progress in the VUDA literature, to date, the role of LLVM in the context of action recognition has been overlooked. We argue that the LLVMs have a lot to offer to the progress of VUDA methods, which is currently untapped. Our work starts from a preliminary analysis demonstrating that current sophisticated VUDA methods can be largely outperformed by off-the-shelf publicly available LLVMs like CLIP, without needing explicit adaptation (see Fig. 1). This observation highlights the need for a paradigm shift among VUDA methods. The benefit of LLVM models is so significant that it raises a pertinent question: If explicit alignment methodologies, traditionally pursued in VUDA, is truly the way forward? Instead, the primary research question becomes: How can we efficiently integrate the prior knowledge derived from LLVMs to adapt a model to the target domain for VUDA? + +This paper presents the first approach in the VUDA literature, which attempts to address this question specifically. In particular, we consider the challenging scenario of Source-Free Video Unsupervised Domain Adaptation (SFVUDA) [44, 46], which consists in the task of adapting an action recognition model trained on a labelled source domain to an unlabelled target domain, without accessing the actual source data. The primary motivation for exploiting LLVMs in SFVUDA is that we expect the generalization capabilities of these models to effectively serve the pur + +pose of mitigating the effects of the domain shift existing between the data distributions of the source and the target domain. In particular, since in SFVUDA we do not have access to the source data but only the source pre-trained model, direct target adaptation may potentially lead to significantly poor performance, especially when the domain shift is large. On the contrary, the employment of LLVMs can efficiently counteract this negative effect owing to the wide range of visual domains observed during their training process. An additional motivation behind our approach lies in the fact that the rich visual representations derived from LLVMs can efficiently compensate when additional modalities (e.g., optical-flow), which are known to help reduce domain gap in VUDA frameworks [27, 33], are unavailable. + +Our approach, which we name as Domain Adaptation with Large Language-Vision models (or DALL-V in short), is a simple yet effective technique to combine the knowledge from the visual representation of a pre-trained LLVM with that derived from the source model and the target data (see Fig. 2c). DALL-V involves a two-stage pipeline: (i) in the first stage pseudo-labels are extracted from the CLIP model and are subsequently used for adapting on the target dataset, and (ii) in the second stage an ensemble of CLIP, source and target models are used to distill information into a student network. DALL-V introduces very few trainable parameters on top of CLIP and is realized with the help of domain-specific adapters. Despite its simplicity, our approach outperforms SFVUDA and even VUDA methods by a significant margin (+11.8% w.r.t. the best competitor). + +In summary, our contributions are: (i) We show, for the first time in the literature that straightforward zero-shot + +methods based on LLVMs vastly outperform the state-of-the-art approaches for SFVUDA. (ii) Building upon this observation, we propose DALL-V, a simple approach for SFVUDA which optimally integrates information derived from LLVMs, from a pretrained source model and from unlabelled videos of the target domain. (iii) We perform extensive experiments on a total of 20 domain adaptation settings, demonstrating the effectiveness of DALL-V in SFVUDA. + +# 2. Related Work + +Video Unsupervised Domain Adaptation. While domain adaptation techniques have been primarily studied in the context of image-level representations and fall under the umbrella of Unsupervised Domain Adaptation (UDA) methods [7, 47, 22, 39, 1, 42], addressing the domain shift problem is even more challenging in the case of videos, due to the additional temporal dimension. Researchers have proposed different strategies. For instance, Chen et al. introduced $\mathrm{TA}^3\mathrm{N}$ [1], a discrepancy-based method that aligns domains on the temporal axis by learning a temporal relationship across video sequences. TCoN was introduced in [28], a deep architecture which employs a cross-domain module to compute temporally-aligned source and target feature representations. CO2A [2] proposed an approach based on contrastive learning to align source and target video representations. One major limitation with both UDA and VUDA methods is the fact that they require access to source data. This limits their applicability in real-world scenarios where data may not be accessible due to privacy reasons. Differently, our approach focuses in the more realistic setting where we only have access to the source model. + +Source-Free Video Unsupervised Domain Adaptation. Due to the growing concerns related to privacy and data sharing, recently researchers have begun exploring methods Source-Free Unsupervised Domain Adaptation (SFUDA). One of the first works in this direction is SHOT [20], a deep architecture which employs an entropy loss and a classification loss on pseudo-labelled data to adapt the source pretrained network focusing solely on the target data. In a subsequent work [21], the authors introduced an additional auxiliary head to solve the relative rotation task, further improving prediction accuracy. Other works, such as 3C-GAN [18], SFIT [9], and SDDA [16], framed the problem of SFUDA as an image translation task. + +In the literature, only a handful of works addressed the problem of Source-Free Video Unsupervised Domain Adaptation (SFVUDA). In particular, ATCoN [44] presented an approach which models temporal consistency across the video sequences. EXTERN [46] proposed to exploit mask-to-mix strategies and video-tailored regularizations for SFVUDA. Our approach radically departs from these previous works as we proposed to exploit LLVMs to adapt to the target video data. + +Large Language-Vision Models. The availability of vast web-scale datasets containing image-text pairs [35, 34, 37] have enabled the emergence of novel large multi-modal neural networks which learn joint visual-text embedding spaces [31, 10, 37]. These approaches, commonly referred as Large Language-Vision Models, utilise a separate encoder for each modality and employ a contrastive loss to align the data representations in the feature space [31, 10]. CLIP [31] is a prominent example of such an approach. Despite its simplicity, CLIP have been shown to achieve impressive zero-shot image recognition capabilities. More recently, LLVMs have been also extended with success to the video domains and particularly to the action recognition task [40]. However, we are not aware of previous works which have specifically used LLVMs to address the problem of domain shift in video action recognition. + +# 3. The Unreasonable Effectiveness of LLVM + +In this work our goal is to develop a Source-free Video Unsupervised Domain Adaptation (SFVUDA) method that can adapt a source trained model to a target domain of interest. Departing from the traditional SFVUDA approaches, where a multitude of self-supervised losses are optimized on the target domain [44], we pursue an orthogonal and unconventional approach. Our key idea is to leverage a LLVM (e.g., CLIP [31]), which has been pre-trained on web-scale image-text pairs, as a tool for bridging the domain gap. + +As a part of a preliminary study we evaluate CLIP in a zero-shot manner (see Fig. 2b) on the target dataset. In other words, we simply run inference and evaluate the performance on the Daily-DA [44] benchmark using a pre-trained CLIP (see Sec. 4.2 for details on zero-shot inference), and compare with the recent state-of-the-art SFVUDA approaches. The evaluation is carried out on the four target domains of the Daily-DA, each of which has eight semantic categories. As showcased in Fig. 1, much to our surprise, CLIP (ViT-B/32) outperforms the best performing SFVUDA method EXTERN [46] by a huge margin of $+8.5\%$ , despite CLIP never been explicitly fine-tuned on the target frames, let alone videos. These telling observations from our preliminary study compels us to explore and exploit the unreasonable effectiveness of the LLVMs, which is orthogonal to the existing SFVUDA literature. + +The implications of undertaking such a pathway to bridge the domain gap in SFVUDA can be several: (i) the LLVMs being publicly available, it will help democratize more efficient target adaptation, even if the source data is withheld due to privacy reasons, (ii) it disposes off the need to balance and tune complex training objectives, a de facto practice in SFVUDA (e.g., ATCoN [44] requires to jointly optimize up to 6 loss functions), (iii) it prevents the necessity of ad hoc and specialized network architectures and training objectives to solve the task at hand. Encouraged by + +the flexibility offered by LLVMs, we propose DALL-V for SFVUDA that can effectively exploit the world prior from the LLVMs and integrate it with complementary sources of information, such as the source trained model. Before we describe our proposed method, we formalize the SFVUDA task and introduce some preliminaries. + +# 4. Methods + +# 4.1. Problem definition and notations + +In SFVUDA we are given an unlabelled target dataset $\mathcal{D}^{\mathrm{T}} = \{\mathbf{X}_i^{\mathrm{T}}\}_{i=1}^{m}$ , containing $m$ video sequences, and access to a source model $F^{\mathrm{S}}(\cdot)$ , trained on a labelled source dataset $\mathcal{D}^{\mathrm{S}} = \{\mathbf{X}_i^{\mathrm{S}}, y_i^{\mathrm{S}}\}_{i=1}^{n}$ , which is not available during adaptation on the target. We assume that the target dataset $\mathcal{D}^{\mathrm{T}}$ contains video sequences of actions that share the same label space as the source, i.e., $\mathcal{Y}^{\mathrm{S}} = \mathcal{Y}^{\mathrm{T}} = \{1, \dots, C\}$ , with $C$ being the number of semantic action categories. We also assume that the source and target marginal distributions are not the same, leading to the so-called domain-shift [44]. + +The goal of any SFVUDA method is to utilize the target dataset and the source-trained model to learn a mapping function (typically using a neural network) that can correctly predict the target samples, without needing any access to the source dataset. + +# 4.2. Preliminaries + +CLIP [31], which stands for Contrastive Language-Image Pre-training, is composed of two encoders: a vision encoder $G_{\mathrm{V}}(\cdot)$ and a language (or text) encoder $G_{\mathrm{L}}(\cdot)$ (see Fig. 2b). To classify, it associates labels to visual inputs by computing their similarity with a set of textual descriptions. More precisely, given the names of the classes in a dataset (i.e., drinking, walking, etc.), we construct prompts containing each class name, e.g., "a video of a person [CLS]", where "[CLS]" denotes a class name. + +The language model $G_{\mathrm{L}}(\cdot)$ projects the class names into embeddings, denoted as $\{z_c^{\mathrm{L}}\}_{c = 1}^{C}$ . On the other hand, assuming a test video of $K$ frames $\mathbf{X} = \{\mathbf{x}_k\}_{k = 1}^K$ , we extract the visual feature representation for each frame, i.e. $\mathbf{Z}^{\mathrm{V}} = \{z_k^{\mathrm{V}}\}_{k = 1}^K$ using the vision encoder $G_{\mathrm{V}}(\cdot)$ . CLIP outputs a probability distribution, where the probability of the $k$ -th frame to belong to the $c$ -th class is given by: + +$$ +p _ {c} (\mathbf {x} _ {k}) = \frac {\exp \left(< z _ {c} ^ {\mathrm {L}} , z _ {k} ^ {\mathrm {V}} > / \tau\right)}{\sum_ {c ^ {\prime} = 1} ^ {C} \exp \left(< z _ {c ^ {\prime}} ^ {\mathrm {L}}, z _ {k} ^ {\mathrm {V}} > / \tau\right)} \tag {1} +$$ + +where $\tau$ is a temperature and $<\cdot, \cdot>$ is the inner-product (or cosine similarity) operator. The frame-wise predictions are then aggregated by simply averaging the class probabilities estimated for each frame. + +With respect to inference with CLIP, it is regarded as zero-shot inference if a given dataset has not been explicitly used for training the CLIP model. In other words, + +zero-shot inference refers to the generalization to unseen datasets [31], and must not be confused with generalization to strictly unseen objects [17]. + +# 4.3. Domain Adaptation with Large Language-Vision models (DALL-V) + +In this work, we propose Domain Adaptation with Large Language-Vision models (DALL-V) as a SFVUDA method. As the name suggests, it leverages the rich world prior from the LLVMs, CLIP in particular, to help mitigate the domain gap. Recall that CLIP has been trained on image-text pairs and not video-caption pairs. On the other hand, the source model can offer complementary information as it has been trained supervisedly on the source videos. Thus, we aspire to jointly leverage these two reservoirs of information in a simple yet effective manner. + +Our DALL-V mainly operates in two stages. The first stage, called Target adaptation (see Sec. 4.3.1), consists in pseudo-labelling the unlabelled target videos with CLIP and training a target-specific model with the pseudo-labelled data. In this way the target-specific model can benefit from the rich general purpose knowledge of CLIP and at the same time discover patterns that are inherent to the target domain. The second stage, called Ensemble distillation (see Sec. 4.3.2), focuses on ensembling the source model and target model into a smaller network, with knowledge distillation [8], which is then finally used during inference. Before we dive into the details of our proposed method, we describe how we carry out source training. + +Source Pre-Training. Opposed to the previous work AT-CoN, which fine-tunes all the source model weights using the source dataset, we propose an alternative approach. To prevent the source model to be biased towards the source dataset, we propose to initialize the source model with CLIP pre-trained weights, and fine-tune only a small set of parameters. To fine-tune CLIP on the source data, we introduce an adapter network [6], a small multi-layer perceptron network, that is appended on top of the vision encoder of CLIP. Source pre-training essentially consists in updating the weights of the adapter using a supervised training objective, while keeping the CLIP vision encoder frozen. + +Formally, the vision encoder of CLIP $G_{\mathrm{V}}(\cdot)$ is extended with an adapter $A(\cdot): \mathcal{R}^d \to \mathcal{R}^d$ on top, where $d$ is the output dimension of $G_{\mathrm{V}}(\cdot)$ . We construct textual descriptions $t$ for the video sequences using a set of 16 manually-designed prompts [40], containing templates such as "a video of [CLS]", "can you recognize the action of [CLS]?", and so on, with $y$ being the class name [CLS] (refer to the Supplement for the complete list of prompts). + +We adopt the training objective of ActionCLIP [40] to train the source model. In details, we first compute the ground-truth similarity scores $q(\mathbf{x})$ for each input frame $\mathbf{x}$ , + +![](images/d372bd9eee848a026ea930442bbfba424ce52ae1df8bf5767cbb156631eb27c4.jpg) +(a) Source/target adapter $(A^{\mathrm{S}} / A^{\mathrm{T}})$ training. For the target domain, $q$ is constructed by pseudo-labelling with CLIP (ViT-B/32). + +![](images/e0e869990c32cf734566767e346307e379a8c88f30c41eb7982f6ed555a85065.jpg) +(b) Ensemble distillation. We use $A^{\mathrm{S}}, A^{\mathrm{T}}$ , and CLIP (ViT-B/32) as teachers and train a student adapter $\overline{A}$ on a CLIP (RN50) model. +Figure 3: Overview of the pipeline of DALL-V. The $\mathbb{O}$ means the network is trainable, while the $\text{念}$ means the network is frozen. We denote with bar the student components to distinguish them from the teacher. + +with 0 in the positions corresponding to negative pairs in the mini-batch and 1 for positive pairs, where a pair denotes the frame-prompt tuple. To train the network (essentially just the adapter $A^{\mathrm{S}}$ ; language and vision encoders being frozen) we use the Kullback-Leibler (KL) divergence loss between $\mathbf{p}(\mathbf{X})$ (see Eq. (1)) and $\mathbf{q}(\mathbf{X})$ : + +$$ +\mathcal {L} _ {\mathrm {a d p a t e r}} ^ {\mathrm {S}} = \mathbb {E} _ {(\mathbf {X}, t) \sim \mathcal {D} ^ {\mathrm {s}}} [ \mathrm {K L} (\mathbf {p} (\mathbf {X}), \mathbf {q} (\mathbf {X})) ] \tag {2} +$$ + +# 4.3.1 Target adaptation + +To adapt CLIP on the target domain, we adopt a similar approach to that outlined for the source domain, with one key difference: the absence of ground-truth labels for the target domain. While it is a common practice among the source-free methods to utilize the source model to pseudolabel the unlabelled target data [20], we refrain from such an approach, and instead utilize the original CLIP model (i.e., without any source adapter $A^{\mathrm{S}}$ ) for pseudo-labelling. Despite learning fewer parameters with the adapter, the source model can still be biased towards the source, and hence, yielding noisy pseudo-labels. + +To pseudo-label (PL), we utilize zero-shot CLIP inference (described in Sec. 4.2), with the same collection of textual descriptions used during the source fine-tuning stage, to assign a PL to each sample in the target domain. To this end, we use a confidence threshold for filtering out unreliable PLs (see Supplement for details). After the thresholding, we end up with a subset $\tilde{\mathcal{D}}^{\mathrm{T}} = \{\mathbf{X}_i^{\mathrm{T}}\}_{i=1}^{m'} \subset \mathcal{D}^{\mathrm{T}}$ , containing PLs as $\{\tilde{y}_i\}_{i=1}^{m'}$ with $m'$ denoting the filtered samples. + +Following the source fine-tuning procedure, we append an adapter $A^{\mathrm{T}}(\cdot): \mathcal{R}^{d} \to \mathcal{R}^{d}$ to the vision encoder of CLIP and minimize the following loss, while keeping CLIP frozen: + +$$ +\mathcal {L} _ {\text {a d a p t e r}} ^ {\mathrm {T}} = \mathbb {E} _ {(\mathbf {X}, t) \sim \tilde {\mathcal {D}} ^ {\mathrm {T}}} [ \operatorname {K L} (\mathbf {p} (\mathbf {X}), \mathbf {q} (\mathbf {X})) ] \tag {3} +$$ + +# 4.3.2 Ensemble distillation + +With three different models at our disposal: the source, target and the original CLIP, we can distill information from these three models into a single network. Knowledge Distillation [8] presents two important advantages: (i) allows to generate more reliable and informative pseudo-labels, and (ii) distillation into a smaller network reduces inference time without sacrificing much performance. For this second point, we explicitly use CLIP ViT-B/32 as teacher (and as backbone for the source and target adapters) and CLIP RN50 as student. Since we use an ensemble of three models to distill information, we call this step ensemble distillation. Now, with predictions available from three different models, the question is how to distill them together. A simple approach could be to average the three probability distributions as follows: + +$$ +\mathbf {p} ^ {\text {e n s}} (\mathbf {X}) = \frac {1}{3} \left(\underbrace {\mathbf {p} (\mathbf {X})} _ {\text {Z S C L I P}} + \underbrace {\mathbf {p} ^ {\mathrm {S}} (\mathbf {X})} _ {\text {S o u r c e a d p a t e r}} + \underbrace {\mathbf {p} ^ {\mathrm {T}} (\mathbf {X})} _ {\text {T a r g e t a d p a t e r}}\right) \tag {4} +$$ + +Whereas, another approach could consist in carrying out majority voting, where each of the three heads votes for a class, and the class with two or more votes becomes the new PL. Following [8], we utilize both kinds of PLs for distillation, using a standard multi-class classification loss (or cross-entropy loss $\mathcal{L}_{\mathrm{CE}}$ ) and a discrepancy loss. The discrepancy loss corresponding to the ensembled PL in Eq. (4) is given as: + +$$ +\mathcal {L} _ {\text {d i s c r e p a n c y}} = \mathbb {E} _ {(\mathbf {X}, t) \sim \mathcal {D} ^ {\tau}} \left[ \mathrm {K L} \left(\frac {\mathbf {p} ^ {\text {s t u d e n t}} (\mathbf {X})}{\tau}, \frac {\mathbf {p} ^ {\text {e n s}} (\mathbf {X})}{\tau}\right) \right] \tag {5} +$$ + +where $\mathbf{p}^{\mathrm{student}}$ denotes the student network's probability distribution, which is matched with that of $\mathbf{p}^{\mathrm{ens}}$ . + +We then balance the contribution of the two losses as: + +$$ +\mathcal {L} _ {\text {d i s t i l l}} = \alpha * \mathcal {L} _ {c e} + (1 - \alpha) * \mathcal {L} _ {d} \tag {6} +$$ + +where $\alpha$ is set according to the value proposed in [8]. Note that, we do not fine-tune all the student weights, but instead append an adapter on top of the frozen backbone and fine-tune only the adapter $\overline{A}$ using Eq. (6). + +# 5. Experiments + +# 5.1. Experimental setup + +Datasets and settings. We present an extensive experimental evaluation on three standard benchmarks for SFVUDA. In particular, we report results on Daily-DA [44], which comprises 18,949 videos from 8 classes, and it is built from 4 original video action recognition datasets, namely HMDB51 [15], ARID [43], MIT [26] and Kinetics [12]. Additionally, we evaluate our framework on UCF-HMDBfull [1], a benchmark comprising 3,209 videos divided in 12 action categories from the HMDB51 [15] and UCF101 [38] action recognition datasets. Note that the first benchmark, Daily-DA, poses a more significant challenge when compared to the latter since it comprises videos with very different lighting conditions across domains. Last, we test our method on Sports-DA [45], a benchmark consisting of three datasets (i.e., UCF101 [38], Sports-1M [11], and Kinetics [12]), with 40,718 videos and 23 classes. + +Implementation details. We implement our adapters as two-layers perceptrons, following [6]. In accordance with [40], we use AdamW optimizer [24] with a learning rate of 0.01 and a weight decay of 0.2. The temperature $\tau$ is set to 2.0 for all datasets. We trained all our models for 30 epochs. For a single experiment, we have used either 4 Tesla V100 or 2 RTX A6000 GPUs. Further implementation details can be found in the Supplement. + +Baselines and competitors. We compare our method with a selection of standard baseline methods for VUDA and SFVUDA, which include TRN [49], DANN [5], MK-MDD [23], $\mathrm{TA}^3\mathrm{N}$ [1], SFDA [13], SHOT [20], SHOT++ [21], MA [19], BAIT [48] and CPGA [30]. Additionally, we report the scores of the state-of-the-art competitors ATCoN [44] and EXTERN [46]. We also report the results for a set of CLIP-based baselines of our design to provide a more solid context for assessing our performance. In particular, CLIP (RN50) and CLIP (ViT-B/32) indicate the scores obtained in a zero-shot fashion with the CLIP model, with ResNet-50 and ViT-B/32 as backbones, respectively. Finally, we report the lower and upper bounds obtained by the source-only and target-supervised models, respectively. + +# 5.2. Comparison with the State-of-the-Art + +We report our experimental validation scores in Tab. 1, Tab. 2, and Tab. 3, obtained on the Daily-DA, the UCF-HMDB $f_{ull}$ , and the Sports-DA benchmarks respectively. For all the benchmarks, the lower bound refers to the out-of-the-box performance of our source-trained model on the target data, i.e., the CLIP backbone and the source adapter $A^{\mathrm{S}}$ . Analogously, the upper bound is the oracle performance obtained after training supervisedly on the target data, i.e., assuming knowledge of the target labels. + +Regarding the more challenging Daily-DA dataset, it is possible to observe in Tab. 1 that our proposed distillation method DALL-V (RN50) obtains the best average score across the 12 settings among the competitors and baselines sharing the same backbone, with a $5\%$ gain on the second best score (CLIP (RN50)). It is also visible that the best score is achieved in 9 out 12 total settings, proving our proposed distillation-based method to be significantly effective when addressing the SFVUDA task. Notably, it emerges that the scores achieved when addressing the ARID dataset $(\rightarrow \mathrm{A})$ as target domain are visibly lower than those reported for the other settings. We associate this behavior to the challenging visual features of this particular benchmark, characterized by being shot in a low illumination environment. + +As for the small-scale UCF-HMDBfull dataset, we can observe significantly higher absolute accuracy scores on both the involved SFVUDA directions, indicating that the benchmark is significantly closer to saturation for this task. Nonetheless, it emerges from the table that our proposed method achieves a state-of-the-art average accuracy score across the two datasets (+0.6% on the best competitor) and the best score for the HMDB→UCF direction (+1.2% on the second best). + +Finally, as shown in Tab. 3, our method achieves comparable results with EXTERN [46] on Sports-DA, surpassing it on three settings, and achieving the same results on another. On the two settings with UCF101 as target (i.e., $\mathrm{K} \rightarrow \mathrm{U}$ and $\mathrm{S} \rightarrow \mathrm{U}$ ), our method falls behind by approximately 6 percentage points, resulting on an average score across the six scenarios slightly lower than the state-of-the-art (i.e., $-0.9\%$ ). However, DALL-V still achieves an average accuracy of $+8.5\%$ w.r.t. ATCoN, further demonstrating the effectiveness of our approach for SFVUDA. Interestingly, on Sports-DA, our lower bound already achieves competitive results with other SFVUDA methods, surpassing ATCoN by $+2.7\%$ on average. The table additionally shows that CLIP zero-shot is an extremely strong baseline on the benchmark, achieving comparable results with EXTERN, i.e., $-2.0\%$ with RN50, and $+2.6\%$ with the ViT-B backbone. + +
MethodAccuracy (%)
K→AK→HK→MM→AM→HM→KH→AH→MH→KA→HA→MA→KAvg.
ZSTRN [49]20.936.729.022.143.753.113.822.037.117.214.724.427.9
Lower bound15.647.935.734.744.661.617.525.545.114.615.517.831.3
CLIP (RN50) [31]30.550.042.230.550.062.930.542.262.950.042.262.946.4
CLIP (ViT-B/32) [31]31.349.646.031.349.665.431.346.065.449.646.065.448.1
UDADANN [5]21.237.521.722.843.358.814.229.538.220.119.727.029.5
MK-MDD [23]21.736.224.021.050.458.520.325.733.818.718.026.129.5
TA3N [1]19.937.731.521.643.055.514.425.738.414.915.623.428.5
SFUDASFDA [13]12.644.927.516.035.249.213.124.224.916.313.225.225.2
SHOT [20]12.044.629.515.336.751.013.624.221.217.114.024.325.3
SHOT++ [21]12.640.828.714.941.746.316.022.233.115.412.521.824.4
MA [19]12.845.830.017.737.453.512.925.022.216.715.224.326.1
BAIT [48]12.745.730.016.939.653.013.625.521.215.714.525.526.2
CPGA [30]13.146.030.718.139.255.113.126.225.519.216.526.726.5
SFVUDAATCoN [44]17.248.232.527.247.357.717.930.748.526.717.231.033.5
EXTERNAL [46]23.955.835.218.153.768.126.240.757.626.218.251.439.6
DALL-V (ours)24.052.547.024.065.478.124.047.076.757.945.775.051.4
Upper bound26.970.461.526.970.488.926.961.588.970.461.588.961.9
+ +Table 1: Validation accuracy for Daily-DA. Bold indicates best, underline represents best with same backbone as baseline (i.e. ResNet50). Lower bound indicates a source adapter and Upper bound a target adapter, both trained supervised on CLIP (RN50). + +
MethodAccuracy (%)
H→UU→HAvg.
ZSTRN [49]72.872.172.4
Lower bound71.676.173.8
CLIP (RN50) [31]81.086.083.5
CLIP (ViT-B/32) [31]90.389.189.7
UDADANN [5]74.475.174.8
MK-MMD [23]74.779.777.2
TA3N [1]78.184.881.5
SFUDASFDA [13]69.875.072.4
SHOT [20]74.474.474.4
SHOT++ [21]71.168.169.6
MA [19]74.467.370.9
BAIT [48]75.376.375.8
CPGA [30]75.868.172.0
SFVUDAATCoN [44]85.379.782.5
EXTERNAL [46]91.988.990.4
DALL-V (ours)93.188.991.0
Upper bound93.791.492.6
+ +Table 2: Validation accuracy for UCF-HMDBfull. Bold indicates best, underline represents best with same backbone as baseline (i.e. ResNet50). Lower bound indicates a source adapter and Upper bound a target adapter, both trained supervised on CLIP (RN50). + +# 5.3. Ablation analysis + +We report in Tab. 4 the intermediate accuracy scores achieved at different steps of our proposed source-free pipeline for the Daily-DA and UCF-HMDBfull. In particular, we report the performance of the source supervised model and the target unsupervised one, also reporting the + +final model and the zero-shot version of CLIP for comparison. Finally, we report the scores obtained by assembling the predictions of the three aforementioned models. The $A^{\mathrm{S}}$ and $A^{\mathrm{T}}$ notation indicates the usage of the source and target adapters on top of the frozen CLIP model, that are trained on the respective domains as described in detail in 4.3. Next we discuss the individual contributions. Additional ablation analyses are available in the Supplement. + +Effectiveness of finetuning LLVM on the target. Tab. 4 clearly shows that fine-tuning our model in an unsupervised fashion on the target domain is significantly effective with respect to the target validation accuracy. The tables show indeed a gain of $2.0\%$ on Daily-DA and $1.1\%$ on UCF-HMDB $f_{ull}$ . Additionally, it is possible to observe a further gain on Daily-DA when ensembling the target model (which gains $+2.7\%$ by itself) with CLIP, achieving a $+3.6\%$ . From these results, it emerges that fine-tuning LLVMs on the target domain, even in an unsupervised manner, is effective for the SFVUDA task. + +Effectiveness of ensembles on LLVM and domain-specific networks. In the last line of Tab. 4, it is possible to observe that ensembling predictions from the three distinct models turns out to yield the best average target accuracy score on both the considered benchmarks, resulting in a gain of $+5.7\%$ on Daily-DA and of $+2.9\%$ on UCF-HMDBfull w.r.t. the CLIP baseline. These results and those reported in the previous two paragraphs represent further proof that the three models are effectively complementary, and each contributes with additional discriminative knowledge to the resulting model. Conversely, CLIP positively contributes to + +
MethodAccuracy (%)
K→UK→SS→US→KU→KU→SAvg.
ZSTRN [49]86.466.985.371.049.343.367.0
Lower bound85.479.584.478.267.264.376.5
CLIP (RN50) [31]83.479.983.480.480.479.981.2
CLIP (ViT-B/32) [31]90.082.490.085.185.182.485.8
UDADANN [5]88.075.085.773.465.955.173.8
MK-MMD [23]90.267.990.973.666.155.674.0
TA3N [1]90.368.693.072.663.654.173.7
SFUDASFDA [13]86.160.085.468.055.843.666.5
SHOT [20]91.264.988.872.053.943.669.1
SHOT++ [21]90.063.188.070.344.740.966.2
MA [19]91.065.987.871.960.739.469.5
BAIT [48]92.366.688.372.857.244.770.3
CPGA [30]89.466.386.572.555.244.569.1
SEVUDAATCoN [44]93.669.790.676.065.247.973.8
EXTERNAL [46]93.773.895.482.281.272.783.2
DALL-V (ours)88.077.788.882.381.275.982.3
Upper bound93.488.393.485.685.688.389.1
+ +Table 3: Validation accuracy for Sports-DA. Bold indicates best, underline represents best with same backbone as baseline (i.e. ResNet50). Lower bound indicates a source adapter and Upper bound a target adapter, both trained supervised on CLIP (RN50). + +
MethodAccuracy (%)
K→AnyM→AnyDaily-DAA→AnyAvg.UCF-HMDBfull
H→AnyA→AnyH→UU→HAvg.
CLIP42.348.747.653.748.190.389.189.7
AS40.0 [-2.3]55.7 [+7.0]35.3 [-12.4]35.0 [-18.7]41.5 [-6.6]91.0 [+0.7]80.5 [-8.6]85.8 [-3.9]
AT44.4 [+2.1]51.6 [+2.9]48.5 [+0.9]58.4 [+4.7]50.8 [+2.7]90.5 [+0.2]91.1 [+2.0]90.8 [+1.1]
CLIP + AS45.1 [+2.8]57.8 [+9.1]40.0 [-7.6]50.8 [-2.9]48.4 [+0.3]93.1 [+2.8]87.2 [-1.9]90.2 [+0.5]
CLIP + AT43.8 [+1.5]50.9 [+2.2]48.2 [+0.6]57.4 [+3.7]51.7 [+3.6]89.3 [-1.0]91.9 [+2.8]90.6 [+0.9]
CLIP + AS + AT46.2 [+3.9]57.3 [+8.6]49.5 [+1.9]62.1 [+8.4]53.8 [+5.7]94.9 [+4.6]90.3 [+1.2]92.6 [+2.9]
+ +Table 4: Ablation on the relative improvement of the different modules of our proposed DALL-V framework w.r.t. CLIP (ViT-B/32) on both considered benchmarks. Bold indicates best. The + operator indicates assembling. Due to limited space, results are aggregated by source dataset. + +the final score thanks to its well-known ability to generalize to unseen domains in a zero-shot fashion; on the other hand, the tuning process of domain-specific adapters on the source and target domains, although not sufficient by itself, proves beneficial thanks to its ability to capture specific features of the involved domains and the associated label sets. Note that we still opted for distilling this knowledge to a ResNet-based model in order to be directly comparable with the competitors. + +Effectiveness of using multiple templates. As a further insight into the individual contributions of the different components of our framework, we propose in Fig. 5, a simple analysis on how the number of templates for the text input at inference time impacts on the accuracy score on the target validation set of the two considered benchmarks. We observe that increasing the number of templates from just + +a single one to 12 (following ActionCLIP [40]) is significantly beneficial with respect to the final accuracy score. For even higher number of templates, the performance of our model reaches a plateau. Note that for the choice of the number of template we opted for following [40]. + +Modeling the feature space. We plot in Fig. 4 an UMAP [25] chart of the features produced by the zero-shot CLIP (RN50) model, the source model, and our DALL-V model, on the UCF-HMDBfull benchmark, showing the features belonging to different classes in different colors. The chart clearly highlights the ability of the proposed DALL-V method to produce well-separated clusters for different action categories, thanks to the effective combination of the complementary knowledge of intermediate models. + +![](images/797cbb1017be164c8da2199bd4bc10da16b4260067b5c7730f71f69a737166cb.jpg) +(a) HMDB $\rightarrow$ UCF + +![](images/dc039be3a8c03a55a1e2965d8cf3f6b063ed9fc60f022905b4d1e49363428c43.jpg) +(b) UCF $\rightarrow$ HMDB + +![](images/e7ea954910da37e4eb4a11593610f5774c5305f10062a750494e0e89f3c21843.jpg) +Figure 4: UMAP visualisation of the feature space on $\mathbf{UCF - HMDB}_{full}$ for the zero-shot CLIP (RN50), source and final DALL-V models. The chart shows the ability of our proposed DALL-V to efficiently cluster the action categories in the features space by combining knowledge from the other two models. +Figure 5: Sensitivity study on the number of templates used for the text input w.r.t. the accuracy score, on Daily-DA and UCF-HMDB $f_{ull}$ . + +# Limitations + +Note that CLIP has been trained on web-crawled image-text pairs that are not made publicly available by OpenAI [35]. Due to its black-box nature, equally trusting CLIP predictions, as much as the source model, may elicit unpredictable behaviour, especially in safety critical applications. Thus, appropriate caution must be exercised. Moreover, ours being an empirical work, we can not provide any theoretical guarantee if using LLVMs, such as CLIP, will always result in better performance over traditional SFVUDA approaches. Thus, as future work, we plan to extensively evaluate using more open-source alternatives. + +# 6. Conclusions + +We presented DALL-V, a SFVUDA method based on a simple but novel approach to SFVUDA driven by the intuition of combining the complementary information derived from domain-specific simple models and the powerful CLIP-based LLVMs trained on world knowledge. We provided motivation for such approach, and reported an extensive evaluation on three standard benchmarks for VUDA, pursued in our case for the source-free scenario. We compare our performance with existing methods, as well as with + +a selection of CLIP-based baselines, showing that our proposed frameworks achieves state-of-the-art performance on both considered benchmarks. + +Acknowledgments. We acknowledge the support of the MUR PNRR project FAIR - Future AI Research (PE00000013) funded by the NextGenerationEU. E.R. is partially supported by the PRECRISIS, funded by the EU Internal Security Fund (ISFP-2022-TFI-AG-PROTECT-02-101100539), the EU project SPRING (No. 871245), and the by the PRIN project LEGO-AI (Prot. 2020TA3K9N). The work was carried out in the Vision and Learning joint laboratory of FBK and UNITN. This paper has been further supported by the French National Research Agency (ANR-20-CE23-0027), the EU Horizon 2020 Research and Innovation program under grant agreement No 957337. + +# References + +[1] Min-Hung Chen, Zsolt Kira, Ghassan AlRegib, Jaekwon Yoo, Ruxin Chen, and Jian Zheng. Temporal attentive alignment for large-scale video domain adaptation. In ICCV, 2019. 1, 3, 6, 7, 8 +[2] Victor G Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-head contrastive domain adaptation for video action recognition. In WACV, 2022. 1, 3 +[3] Christoph Feichtenhofer, Axel Pinz, and Richard P Wildes. Spatiotemporal multiplier networks for video action recognition. In CVPR, 2017. 1 +[4] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In CVPR, 2016. 1 +[5] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation, 2015. 6, 7, 8 +[6] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv, 2021. 4, 6 +[7] Muhammad Ghifary, W Bastiaan Kleijn, Mengjie Zhang, David Balduzzi, and Wen Li. Deep reconstruction-classification networks for unsupervised domain adaptation. In ECCV, 2016. 3 + +[8] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv, 2015. 4, 5, 6 +[9] Yunzhong Hou and Liang Zheng. Visualizing adapted knowledge in domain transfer. In CVPR, 2021. 3 +[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 3 +[11] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, pages 1725–1732, 2014. 6 +[12] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset, 2017. 6 +[13] Youngeun Kim, Donghyeon Cho, Kyeongtak Han, Priyadarshini Panda, and Sungeun Hong. Domain adaptation without source data, 2020. 6, 7, 8 +[14] Yu Kong and Yun Fu. Human action recognition and prediction: A survey. IJCV, 2022. 1 +[15] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. Hmdb: A large video database for human motion recognition. In ICCV, 2011. 6 +[16] Vinod K Kurmi, Venkatesh K Subramanian, and Vinay P Namboodiri. Domain impression: A source data free domain adaptation method. In WACV, 2021. 3 +[17] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In CVPR, 2009. 4 +[18] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 3 +[19] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 6, 7, 8 +[20] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 3, 5, 6, 7, 8 +[21] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. TPAMI, 44(11):8602-8617, 2021. 3, 6, 7, 8 +[22] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, 2015. 3 +[23] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I. Jordan. Learning transferable features with deep adaptation networks, 2015. 6, 7, 8 +[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 6 +[25] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction, 2018. 8 + +[26] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, and Aude Oliva. Moments in time dataset: one million videos for event understanding. TPAMI, 42(2):502-508, 2019. 6 +[27] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In CVPR, 2020. 2 +[28] Boxiao Pan, Zhangjie Cao, Ehsan Adeli, and Juan Carlos Niebles. Adversarial cross-domain action recognition with co-attention. In AAAI CAI, 2020. 1, 3 +[29] Preksha Pareek and Ankit Thakkar. A survey on video-based human action recognition: recent updates, datasets, challenges, and applications. Artificial Intelligence Review, 2021. 1 +[30] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation, 2021. 6, 7, 8 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 7, 8 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1 +[33] Aadarsh Sahoo, Rutav Shah, Rameswar Panda, Kate Saenko, and Abir Das. Contrast and mix: Temporal contrastive video domain adaptation with background mixing. NeurIPS, 2021. 2 +[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv, 2022. 3 +[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv, 2021. 1, 3, 9 +[36] Javier Selva, Anders S Johansen, Sergio Escalera, Kamal Nasrollahi, Thomas B Moeslund, and Albert Clapés. Video transformers: A survey. arXiv, 2022. 1 +[37] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. 1, 3 +[38] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild, 2012. 6 +[39] Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. In Artificial Intelligence, volume 30, 2016. 3 +[40] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv, 2021. 2, 3, 4, 6, 8 + +[41] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Conference on Empirical Methods in Natural Language Processing, pages 6787-6800, 2021. 2 +[42] Yuecong Xu, Haozhi Cao, Kezhi Mao, Zhenghua Chen, Lihua Xie, and Jianfei Yang. Aligning correlation information for domain adaptation in action recognition. Neural Networks and Learning Systems, 2022. 3 +[43] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Arid: A new dataset for recognizing action in the dark. In International Workshop on Deep Learning for Human Activity Recognition, pages 70-84. Springer, 2021. 6 +[44] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, and Zhenghua Chen. Source-free video domain adaptation by learning temporal consistency for action recognition. In ECCV, 2022. 2, 3, 4, 6, 7, 8 +[45] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, Rui Zhao, and Zhenghua Chen. Multi-source video domain adaptation with temporal attentive moment alignment. arXiv, 2021. 6 +[46] Yuecong Xu, Jianfei Yang, Haozhi Cao, Min Wu, Xiaoli Li, Lihua Xie, and Zhenghua Chen. Leveraging endo- and exotemporal regularization for black-box video domain adaptation, 2022. 2, 3, 6, 7, 8 +[47] Jinyu Yang, Weizhi An, Sheng Wang, Xinliang Zhu, Chaochao Yan, and Junzhou Huang. Label-driven reconstruction for domain adaptation in semantic segmentation. In ECCV, 2020. 3 +[48] Shiqi Yang, Yaxing Wang, Joost Weijer, and Luis Herranz. Unsupervised domain adaptation without source data by casting a bait, 2020. 6, 7, 8 +[49] Bolei Zhou, Alex Andonian, Aude Oliva, and Antonio Torralba. Temporal relational reasoning in videos, 2018. 6, 7, 8 +[50] Yi Zhu, Xinyu Li, Chunhui Liu, Mohammadreza Zolfaghari, Yuanjun Xiong, Chongruo Wu, Zhi Zhang, Joseph Tighe, R Manmatha, and Mu Li. A comprehensive study of deep video action recognition. arXiv, 2020. 1 \ No newline at end of file diff --git a/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/images.zip b/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dfd5216122b639ef3bb3f64ed177fa16213c78fe --- /dev/null +++ b/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8aa1628c7140796cb1a2bcf540c596fcc8344ecf7334a367d2752cd3d421de24 +size 553897 diff --git a/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/layout.json b/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8092bf2727e86576de3b0e763f1348a79f3608ab --- /dev/null +++ b/2023/The Unreasonable Effectiveness of Large Language-Vision Models for Source-Free Video Domain Adaptation/layout.json @@ -0,0 +1,7862 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 501, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 501, + 139 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 501, + 139 + ], + "type": "text", + "content": "The Unreasonable Effectiveness of Large Language-Vision Models for Source-free Video Domain Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 160, + 559, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 160, + 559, + 175 + ], + "spans": [ + { + "bbox": [ + 52, + 160, + 559, + 175 + ], + "type": "text", + "content": "Giacomo Zara1*, Alessandro Conti1*, Subhankar Roy3, Stéphane Lathuilière3, Paolo Rota1, Elisa Ricci1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 175, + 455, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 175, + 455, + 189 + ], + "spans": [ + { + "bbox": [ + 157, + 175, + 455, + 189 + ], + "type": "text", + "content": "1University of Trento, Italy 2Fondazione Bruno Kessler, Italy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 157, + 189, + 455, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 189, + 455, + 203 + ], + "spans": [ + { + "bbox": [ + 157, + 189, + 455, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 157, + 189, + 455, + 203 + ], + "type": "text", + "content": "LTCI, Télécom Paris, Institut polytechnique de Paris, France" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 192, + 205, + 419, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 205, + 419, + 217 + ], + "spans": [ + { + "bbox": [ + 192, + 205, + 419, + 217 + ], + "type": "text", + "content": "{giacomo.zara, alessandro.conti-1}@unitn.it" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 270, + 289, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 498 + ], + "type": "text", + "content": "Source-Free Video Unsupervised Domain Adaptation (SFVUDA) task consists in adapting an action recognition model, trained on a labelled source dataset, to an unlabelled target dataset, without accessing the actual source data. The previous approaches have attempted to address SFVUDA by leveraging self-supervision (e.g., enforcing temporal consistency) derived from the target data itself. In this work, we take an orthogonal approach by exploiting \"web-supervision\" from Large Language-Vision Models (LLVMs), driven by the rationale that LLVMs contain a rich world prior surprisingly robust to domain-shift. We showcase the unreasonable effectiveness of integrating LLVMs for SFVUDA by devising an intuitive and parameter-efficient method, which we name Domain Adaptation with Large Language-Vision models (DALL-V), that distills the world prior and complementary source model information into a student network tailored for the target. Despite the simplicity, DALL-V achieves significant improvement over state-of-the-art SFVUDA methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 533, + 128, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 128, + 545 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 128, + 545 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": "Video analysis tasks, such as action recognition, have long been investigated in computer vision, due to the numerous applications, ranging from video surveillance to social robotics [50, 14, 29]. Major progress has been made in the last decade with the development of specialized deep architectures, such as 3D CNNs [4, 3] and Video Transformers [36], trained on large-scale annotated datasets. However, obtaining sufficient labelled training videos for real-world scenarios can be very costly and time-consuming. In order to alleviate the burden of annotating large scale datasets Video-based Unsupervised Domain" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 249, + 541, + 397 + ], + "blocks": [ + { + "bbox": [ + 312, + 249, + 541, + 397 + ], + "lines": [ + { + "bbox": [ + 312, + 249, + 541, + 397 + ], + "spans": [ + { + "bbox": [ + 312, + 249, + 541, + 397 + ], + "type": "image", + "image_path": "c21b766958fb0a13750f29c7ec9c41d40ab534edbe973eea39260c5ae5a5069c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 402, + 547, + 487 + ], + "lines": [ + { + "bbox": [ + 304, + 402, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 547, + 487 + ], + "type": "text", + "content": "Figure 1: Performance over time by various methods on the Daily-DA video benchmark. A pre-trained LLVM (e.g., CLIP [31]) is surprisingly better when compared with Unsupervised Domain Adaptation (UDA), Source Free UDA (SFUDA) and Source Free Video-based UDA (SFVUDA) methods. Our proposed DALL-V that is built on top of LLVM successfully outperforms all existing baselines." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 497, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 547, + 581 + ], + "type": "text", + "content": "Adaptation (VUDA) [1, 28, 2] have been introduced. The VUDA methods are derived from the common idea of transferring knowledge from a labelled source domain to an unlabelled target domain. While considering different strategies for adaptation, these approaches have all shown significant improvements in the robustness of learnt video models without requiring annotated target data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "In the last few years, the field of computer vision has also witnessed the emergence of a new generation of powerful deep architectures, trained on mammoth internet-scale image-text datasets [35]. These models, commonly known as foundation models [32, 37, 31, 10] or Large LanguageVision Models (LLVM), have achieved outstanding performance, and have become a cornerstone of modern computer vision research. In particular, recent works such as CLIP [31] or FLAVA [37] have shown that rich visual representations for images can be learned from natural language descriptions as a form of supervision. These pre-trained" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 247, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 247, + 703 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 247, + 703 + ], + "type": "text", + "content": "*Giacomo Zara and Alessandro Conti contributed equally." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 703, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 703, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 703, + 287, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 57, + 703, + 287, + 712 + ], + "type": "text", + "content": "Code is available at https://github.com/giaczara/dallyv" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10307" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 70, + 182, + 238 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 182, + 238 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 182, + 238 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 182, + 238 + ], + "type": "image", + "image_path": "7a6b9e8506112f283489cecbbdade7cbabf065c58eae3fd1a31eb2818d60f4e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 251, + 189, + 262 + ], + "lines": [ + { + "bbox": [ + 104, + 251, + 189, + 262 + ], + "spans": [ + { + "bbox": [ + 104, + 251, + 189, + 262 + ], + "type": "text", + "content": "(a) Traditional SFVUDA" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 197, + 69, + 334, + 238 + ], + "blocks": [ + { + "bbox": [ + 197, + 69, + 334, + 238 + ], + "lines": [ + { + "bbox": [ + 197, + 69, + 334, + 238 + ], + "spans": [ + { + "bbox": [ + 197, + 69, + 334, + 238 + ], + "type": "image", + "image_path": "dd153a23f025253233b9242e0ce2ee4482b070514fc0f11da8ff0c03b29b8809.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 252, + 319, + 262 + ], + "lines": [ + { + "bbox": [ + 216, + 252, + 319, + 262 + ], + "spans": [ + { + "bbox": [ + 216, + 252, + 319, + 262 + ], + "type": "text", + "content": "(b) Zero Shot CLIP Inference" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 353, + 69, + 496, + 238 + ], + "blocks": [ + { + "bbox": [ + 353, + 69, + 496, + 238 + ], + "lines": [ + { + "bbox": [ + 353, + 69, + 496, + 238 + ], + "spans": [ + { + "bbox": [ + 353, + 69, + 496, + 238 + ], + "type": "image", + "image_path": "8fdc00db30f2dedfdbeb9b7fd781eea1fdbf3d8700cbc0336d5646e609037dfd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 252, + 477, + 262 + ], + "lines": [ + { + "bbox": [ + 381, + 252, + 477, + 262 + ], + "spans": [ + { + "bbox": [ + 381, + 252, + 477, + 262 + ], + "type": "text", + "content": "(c) Our proposed SFVUDA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 267, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 267, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 545, + 329 + ], + "type": "text", + "content": "Figure 2: (a) In traditional SFVUDA a source-trained model is adapted to the unlabelled target dataset using self-supervision such as temporal consistency [44, 46] among frames. (b) LLVM (e.g., CLIP [31]) uses cosine similarity between the feature and language representation to predict the most probable class in a zero-shot (ZS) manner. (c) Our proposed SFVUDA solution distills the ZS-CLIP and source model predictions, while introducing very few learnable parameters, namely adapters. The " + }, + { + "bbox": [ + 46, + 267, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbb{C}}" + }, + { + "bbox": [ + 46, + 267, + 545, + 329 + ], + "type": "text", + "content": " means the network is trainable, while the " + }, + { + "bbox": [ + 46, + 267, + 545, + 329 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbb{C}}" + }, + { + "bbox": [ + 46, + 267, + 545, + 329 + ], + "type": "text", + "content": " means the network is frozen." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 342, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 402 + ], + "type": "text", + "content": "LLVMs are now publicly available and can be easily integrated into any recognition system. Lately, researchers have also shown that the LLVMs can be applied with success to the video domain and specifically to the supervised action recognition task [40, 41]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 402, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 594 + ], + "type": "text", + "content": "Despite the significant progress in the VUDA literature, to date, the role of LLVM in the context of action recognition has been overlooked. We argue that the LLVMs have a lot to offer to the progress of VUDA methods, which is currently untapped. Our work starts from a preliminary analysis demonstrating that current sophisticated VUDA methods can be largely outperformed by off-the-shelf publicly available LLVMs like CLIP, without needing explicit adaptation (see Fig. 1). This observation highlights the need for a paradigm shift among VUDA methods. The benefit of LLVM models is so significant that it raises a pertinent question: If explicit alignment methodologies, traditionally pursued in VUDA, is truly the way forward? Instead, the primary research question becomes: How can we efficiently integrate the prior knowledge derived from LLVMs to adapt a model to the target domain for VUDA?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "This paper presents the first approach in the VUDA literature, which attempts to address this question specifically. In particular, we consider the challenging scenario of Source-Free Video Unsupervised Domain Adaptation (SFVUDA) [44, 46], which consists in the task of adapting an action recognition model trained on a labelled source domain to an unlabelled target domain, without accessing the actual source data. The primary motivation for exploiting LLVMs in SFVUDA is that we expect the generalization capabilities of these models to effectively serve the pur" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "text", + "content": "pose of mitigating the effects of the domain shift existing between the data distributions of the source and the target domain. In particular, since in SFVUDA we do not have access to the source data but only the source pre-trained model, direct target adaptation may potentially lead to significantly poor performance, especially when the domain shift is large. On the contrary, the employment of LLVMs can efficiently counteract this negative effect owing to the wide range of visual domains observed during their training process. An additional motivation behind our approach lies in the fact that the rich visual representations derived from LLVMs can efficiently compensate when additional modalities (e.g., optical-flow), which are known to help reduce domain gap in VUDA frameworks [27, 33], are unavailable." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 689 + ], + "type": "text", + "content": "Our approach, which we name as Domain Adaptation with Large Language-Vision models (or DALL-V in short), is a simple yet effective technique to combine the knowledge from the visual representation of a pre-trained LLVM with that derived from the source model and the target data (see Fig. 2c). DALL-V involves a two-stage pipeline: (i) in the first stage pseudo-labels are extracted from the CLIP model and are subsequently used for adapting on the target dataset, and (ii) in the second stage an ensemble of CLIP, source and target models are used to distill information into a student network. DALL-V introduces very few trainable parameters on top of CLIP and is realized with the help of domain-specific adapters. Despite its simplicity, our approach outperforms SFVUDA and even VUDA methods by a significant margin (+11.8% w.r.t. the best competitor)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "In summary, our contributions are: (i) We show, for the first time in the literature that straightforward zero-shot" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10308" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "methods based on LLVMs vastly outperform the state-of-the-art approaches for SFVUDA. (ii) Building upon this observation, we propose DALL-V, a simple approach for SFVUDA which optimally integrates information derived from LLVMs, from a pretrained source model and from unlabelled videos of the target domain. (iii) We perform extensive experiments on a total of 20 domain adaptation settings, demonstrating the effectiveness of DALL-V in SFVUDA." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 178, + 134, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 134, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 134, + 192 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 449 + ], + "type": "text", + "content": "Video Unsupervised Domain Adaptation. While domain adaptation techniques have been primarily studied in the context of image-level representations and fall under the umbrella of Unsupervised Domain Adaptation (UDA) methods [7, 47, 22, 39, 1, 42], addressing the domain shift problem is even more challenging in the case of videos, due to the additional temporal dimension. Researchers have proposed different strategies. For instance, Chen et al. introduced " + }, + { + "bbox": [ + 46, + 198, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^3\\mathrm{N}" + }, + { + "bbox": [ + 46, + 198, + 287, + 449 + ], + "type": "text", + "content": " [1], a discrepancy-based method that aligns domains on the temporal axis by learning a temporal relationship across video sequences. TCoN was introduced in [28], a deep architecture which employs a cross-domain module to compute temporally-aligned source and target feature representations. CO2A [2] proposed an approach based on contrastive learning to align source and target video representations. One major limitation with both UDA and VUDA methods is the fact that they require access to source data. This limits their applicability in real-world scenarios where data may not be accessible due to privacy reasons. Differently, our approach focuses in the more realistic setting where we only have access to the source model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "text", + "content": "Source-Free Video Unsupervised Domain Adaptation. Due to the growing concerns related to privacy and data sharing, recently researchers have begun exploring methods Source-Free Unsupervised Domain Adaptation (SFUDA). One of the first works in this direction is SHOT [20], a deep architecture which employs an entropy loss and a classification loss on pseudo-labelled data to adapt the source pretrained network focusing solely on the target data. In a subsequent work [21], the authors introduced an additional auxiliary head to solve the relative rotation task, further improving prediction accuracy. Other works, such as 3C-GAN [18], SFIT [9], and SDDA [16], framed the problem of SFUDA as an image translation task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 287, + 714 + ], + "type": "text", + "content": "In the literature, only a handful of works addressed the problem of Source-Free Video Unsupervised Domain Adaptation (SFVUDA). In particular, ATCoN [44] presented an approach which models temporal consistency across the video sequences. EXTERN [46] proposed to exploit mask-to-mix strategies and video-tailored regularizations for SFVUDA. Our approach radically departs from these previous works as we proposed to exploit LLVMs to adapt to the target video data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "Large Language-Vision Models. The availability of vast web-scale datasets containing image-text pairs [35, 34, 37] have enabled the emergence of novel large multi-modal neural networks which learn joint visual-text embedding spaces [31, 10, 37]. These approaches, commonly referred as Large Language-Vision Models, utilise a separate encoder for each modality and employ a contrastive loss to align the data representations in the feature space [31, 10]. CLIP [31] is a prominent example of such an approach. Despite its simplicity, CLIP have been shown to achieve impressive zero-shot image recognition capabilities. More recently, LLVMs have been also extended with success to the video domains and particularly to the action recognition task [40]. However, we are not aware of previous works which have specifically used LLVMs to address the problem of domain shift in video action recognition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 274, + 534, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 274, + 534, + 288 + ], + "spans": [ + { + "bbox": [ + 305, + 274, + 534, + 288 + ], + "type": "text", + "content": "3. The Unreasonable Effectiveness of LLVM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 295, + 545, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 402 + ], + "type": "text", + "content": "In this work our goal is to develop a Source-free Video Unsupervised Domain Adaptation (SFVUDA) method that can adapt a source trained model to a target domain of interest. Departing from the traditional SFVUDA approaches, where a multitude of self-supervised losses are optimized on the target domain [44], we pursue an orthogonal and unconventional approach. Our key idea is to leverage a LLVM (e.g., CLIP [31]), which has been pre-trained on web-scale image-text pairs, as a tool for bridging the domain gap." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 403, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 594 + ], + "type": "text", + "content": "As a part of a preliminary study we evaluate CLIP in a zero-shot manner (see Fig. 2b) on the target dataset. In other words, we simply run inference and evaluate the performance on the Daily-DA [44] benchmark using a pre-trained CLIP (see Sec. 4.2 for details on zero-shot inference), and compare with the recent state-of-the-art SFVUDA approaches. The evaluation is carried out on the four target domains of the Daily-DA, each of which has eight semantic categories. As showcased in Fig. 1, much to our surprise, CLIP (ViT-B/32) outperforms the best performing SFVUDA method EXTERN [46] by a huge margin of " + }, + { + "bbox": [ + 304, + 403, + 545, + 594 + ], + "type": "inline_equation", + "content": "+8.5\\%" + }, + { + "bbox": [ + 304, + 403, + 545, + 594 + ], + "type": "text", + "content": ", despite CLIP never been explicitly fine-tuned on the target frames, let alone videos. These telling observations from our preliminary study compels us to explore and exploit the unreasonable effectiveness of the LLVMs, which is orthogonal to the existing SFVUDA literature." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": "The implications of undertaking such a pathway to bridge the domain gap in SFVUDA can be several: (i) the LLVMs being publicly available, it will help democratize more efficient target adaptation, even if the source data is withheld due to privacy reasons, (ii) it disposes off the need to balance and tune complex training objectives, a de facto practice in SFVUDA (e.g., ATCoN [44] requires to jointly optimize up to 6 loss functions), (iii) it prevents the necessity of ad hoc and specialized network architectures and training objectives to solve the task at hand. Encouraged by" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10309" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "the flexibility offered by LLVMs, we propose DALL-V for SFVUDA that can effectively exploit the world prior from the LLVMs and integrate it with complementary sources of information, such as the source trained model. Before we describe our proposed method, we formalize the SFVUDA task and introduce some preliminaries." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 154, + 108, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 154, + 108, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 108, + 167 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 174, + 224, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 174, + 224, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 174, + 224, + 186 + ], + "type": "text", + "content": "4.1. Problem definition and notations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "text", + "content": "In SFVUDA we are given an unlabelled target dataset " + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{T}} = \\{\\mathbf{X}_i^{\\mathrm{T}}\\}_{i=1}^{m}" + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "text", + "content": ", containing " + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "text", + "content": " video sequences, and access to a source model " + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{S}}(\\cdot)" + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "text", + "content": ", trained on a labelled source dataset " + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{S}} = \\{\\mathbf{X}_i^{\\mathrm{S}}, y_i^{\\mathrm{S}}\\}_{i=1}^{n}" + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "text", + "content": ", which is not available during adaptation on the target. We assume that the target dataset " + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "text", + "content": " contains video sequences of actions that share the same label space as the source, i.e., " + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^{\\mathrm{S}} = \\mathcal{Y}^{\\mathrm{T}} = \\{1, \\dots, C\\}" + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 193, + 287, + 312 + ], + "type": "text", + "content": " being the number of semantic action categories. We also assume that the source and target marginal distributions are not the same, leading to the so-called domain-shift [44]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 312, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 373 + ], + "type": "text", + "content": "The goal of any SFVUDA method is to utilize the target dataset and the source-trained model to learn a mapping function (typically using a neural network) that can correctly predict the target samples, without needing any access to the source dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 379, + 134, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 379, + 134, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 379, + 134, + 392 + ], + "type": "text", + "content": "4.2. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 397, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 506 + ], + "type": "text", + "content": "CLIP [31], which stands for Contrastive Language-Image Pre-training, is composed of two encoders: a vision encoder " + }, + { + "bbox": [ + 46, + 397, + 287, + 506 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{V}}(\\cdot)" + }, + { + "bbox": [ + 46, + 397, + 287, + 506 + ], + "type": "text", + "content": " and a language (or text) encoder " + }, + { + "bbox": [ + 46, + 397, + 287, + 506 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{L}}(\\cdot)" + }, + { + "bbox": [ + 46, + 397, + 287, + 506 + ], + "type": "text", + "content": " (see Fig. 2b). To classify, it associates labels to visual inputs by computing their similarity with a set of textual descriptions. More precisely, given the names of the classes in a dataset (i.e., drinking, walking, etc.), we construct prompts containing each class name, e.g., \"a video of a person [CLS]\", where \"[CLS]\" denotes a class name." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "text", + "content": "The language model " + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{L}}(\\cdot)" + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "text", + "content": " projects the class names into embeddings, denoted as " + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\{z_c^{\\mathrm{L}}\\}_{c = 1}^{C}" + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "text", + "content": ". On the other hand, assuming a test video of " + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\{\\mathbf{x}_k\\}_{k = 1}^K" + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "text", + "content": ", we extract the visual feature representation for each frame, i.e. " + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{\\mathrm{V}} = \\{z_k^{\\mathrm{V}}\\}_{k = 1}^K" + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "text", + "content": " using the vision encoder " + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{V}}(\\cdot)" + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "text", + "content": ". CLIP outputs a probability distribution, where the probability of the " + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "text", + "content": "-th frame to belong to the " + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 506, + 287, + 589 + ], + "type": "text", + "content": "-th class is given by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 597, + 287, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 597, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 85, + 597, + 287, + 628 + ], + "type": "interline_equation", + "content": "p _ {c} (\\mathbf {x} _ {k}) = \\frac {\\exp \\left(< z _ {c} ^ {\\mathrm {L}} , z _ {k} ^ {\\mathrm {V}} > / \\tau\\right)}{\\sum_ {c ^ {\\prime} = 1} ^ {C} \\exp \\left(< z _ {c ^ {\\prime}} ^ {\\mathrm {L}}, z _ {k} ^ {\\mathrm {V}} > / \\tau\\right)} \\tag {1}", + "image_path": "01ce5d6e32b5ae40f0607561bad502e8f258b06c24d1cbc9e0c4b2aa3f0c299e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 630, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 630, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 630, + 287, + 677 + ], + "type": "text", + "content": " is a temperature and " + }, + { + "bbox": [ + 46, + 630, + 287, + 677 + ], + "type": "inline_equation", + "content": "<\\cdot, \\cdot>" + }, + { + "bbox": [ + 46, + 630, + 287, + 677 + ], + "type": "text", + "content": " is the inner-product (or cosine similarity) operator. The frame-wise predictions are then aggregated by simply averaging the class probabilities estimated for each frame." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "With respect to inference with CLIP, it is regarded as zero-shot inference if a given dataset has not been explicitly used for training the CLIP model. In other words," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "content": "zero-shot inference refers to the generalization to unseen datasets [31], and must not be confused with generalization to strictly unseen objects [17]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 115, + 545, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 115, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 305, + 115, + 545, + 140 + ], + "type": "text", + "content": "4.3. Domain Adaptation with Large Language-Vision models (DALL-V)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 145, + 546, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 546, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 546, + 265 + ], + "type": "text", + "content": "In this work, we propose Domain Adaptation with Large Language-Vision models (DALL-V) as a SFVUDA method. As the name suggests, it leverages the rich world prior from the LLVMs, CLIP in particular, to help mitigate the domain gap. Recall that CLIP has been trained on image-text pairs and not video-caption pairs. On the other hand, the source model can offer complementary information as it has been trained supervisedly on the source videos. Thus, we aspire to jointly leverage these two reservoirs of information in a simple yet effective manner." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 266, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 547, + 422 + ], + "type": "text", + "content": "Our DALL-V mainly operates in two stages. The first stage, called Target adaptation (see Sec. 4.3.1), consists in pseudo-labelling the unlabelled target videos with CLIP and training a target-specific model with the pseudo-labelled data. In this way the target-specific model can benefit from the rich general purpose knowledge of CLIP and at the same time discover patterns that are inherent to the target domain. The second stage, called Ensemble distillation (see Sec. 4.3.2), focuses on ensembling the source model and target model into a smaller network, with knowledge distillation [8], which is then finally used during inference. Before we dive into the details of our proposed method, we describe how we carry out source training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 569 + ], + "type": "text", + "content": "Source Pre-Training. Opposed to the previous work AT-CoN, which fine-tunes all the source model weights using the source dataset, we propose an alternative approach. To prevent the source model to be biased towards the source dataset, we propose to initialize the source model with CLIP pre-trained weights, and fine-tune only a small set of parameters. To fine-tune CLIP on the source data, we introduce an adapter network [6], a small multi-layer perceptron network, that is appended on top of the vision encoder of CLIP. Source pre-training essentially consists in updating the weights of the adapter using a supervised training objective, while keeping the CLIP vision encoder frozen." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": "Formally, the vision encoder of CLIP " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{V}}(\\cdot)" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " is extended with an adapter " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "A(\\cdot): \\mathcal{R}^d \\to \\mathcal{R}^d" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " on top, where " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " is the output dimension of " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{V}}(\\cdot)" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": ". We construct textual descriptions " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " for the video sequences using a set of 16 manually-designed prompts [40], containing templates such as \"a video of [CLS]\", \"can you recognize the action of [CLS]?\", and so on, with " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " being the class name [CLS] (refer to the Supplement for the complete list of prompts)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "We adopt the training objective of ActionCLIP [40] to train the source model. In details, we first compute the ground-truth similarity scores " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " for each input frame " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10310" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 111, + 262, + 239 + ], + "blocks": [ + { + "bbox": [ + 67, + 111, + 262, + 239 + ], + "lines": [ + { + "bbox": [ + 67, + 111, + 262, + 239 + ], + "spans": [ + { + "bbox": [ + 67, + 111, + 262, + 239 + ], + "type": "image", + "image_path": "d372bd9eee848a026ea930442bbfba424ce52ae1df8bf5767cbb156631eb27c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 244, + 286, + 266 + ], + "lines": [ + { + "bbox": [ + 62, + 244, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 62, + 244, + 286, + 266 + ], + "type": "text", + "content": "(a) Source/target adapter " + }, + { + "bbox": [ + 62, + 244, + 286, + 266 + ], + "type": "inline_equation", + "content": "(A^{\\mathrm{S}} / A^{\\mathrm{T}})" + }, + { + "bbox": [ + 62, + 244, + 286, + 266 + ], + "type": "text", + "content": " training. For the target domain, " + }, + { + "bbox": [ + 62, + 244, + 286, + 266 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 62, + 244, + 286, + 266 + ], + "type": "text", + "content": " is constructed by pseudo-labelling with CLIP (ViT-B/32)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 326, + 80, + 518, + 240 + ], + "blocks": [ + { + "bbox": [ + 326, + 80, + 518, + 240 + ], + "lines": [ + { + "bbox": [ + 326, + 80, + 518, + 240 + ], + "spans": [ + { + "bbox": [ + 326, + 80, + 518, + 240 + ], + "type": "image", + "image_path": "e0e869990c32cf734566767e346307e379a8c88f30c41eb7982f6ed555a85065.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 244, + 524, + 265 + ], + "lines": [ + { + "bbox": [ + 312, + 244, + 524, + 265 + ], + "spans": [ + { + "bbox": [ + 312, + 244, + 524, + 265 + ], + "type": "text", + "content": "(b) Ensemble distillation. We use " + }, + { + "bbox": [ + 312, + 244, + 524, + 265 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{S}}, A^{\\mathrm{T}}" + }, + { + "bbox": [ + 312, + 244, + 524, + 265 + ], + "type": "text", + "content": ", and CLIP (ViT-B/32) as teachers and train a student adapter " + }, + { + "bbox": [ + 312, + 244, + 524, + 265 + ], + "type": "inline_equation", + "content": "\\overline{A}" + }, + { + "bbox": [ + 312, + 244, + 524, + 265 + ], + "type": "text", + "content": " on a CLIP (RN50) model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 270, + 545, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 545, + 298 + ], + "type": "text", + "content": "Figure 3: Overview of the pipeline of DALL-V. The " + }, + { + "bbox": [ + 46, + 270, + 545, + 298 + ], + "type": "inline_equation", + "content": "\\mathbb{O}" + }, + { + "bbox": [ + 46, + 270, + 545, + 298 + ], + "type": "text", + "content": " means the network is trainable, while the " + }, + { + "bbox": [ + 46, + 270, + 545, + 298 + ], + "type": "inline_equation", + "content": "\\text{念}" + }, + { + "bbox": [ + 46, + 270, + 545, + 298 + ], + "type": "text", + "content": " means the network is frozen. We denote with bar the student components to distinguish them from the teacher." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 307, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 307, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 45, + 307, + 287, + 380 + ], + "type": "text", + "content": "with 0 in the positions corresponding to negative pairs in the mini-batch and 1 for positive pairs, where a pair denotes the frame-prompt tuple. To train the network (essentially just the adapter " + }, + { + "bbox": [ + 45, + 307, + 287, + 380 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{S}}" + }, + { + "bbox": [ + 45, + 307, + 287, + 380 + ], + "type": "text", + "content": "; language and vision encoders being frozen) we use the Kullback-Leibler (KL) divergence loss between " + }, + { + "bbox": [ + 45, + 307, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{p}(\\mathbf{X})" + }, + { + "bbox": [ + 45, + 307, + 287, + 380 + ], + "type": "text", + "content": " (see Eq. (1)) and " + }, + { + "bbox": [ + 45, + 307, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{q}(\\mathbf{X})" + }, + { + "bbox": [ + 45, + 307, + 287, + 380 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 389, + 287, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 389, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 88, + 389, + 287, + 406 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a d p a t e r}} ^ {\\mathrm {S}} = \\mathbb {E} _ {(\\mathbf {X}, t) \\sim \\mathcal {D} ^ {\\mathrm {s}}} [ \\mathrm {K L} (\\mathbf {p} (\\mathbf {X}), \\mathbf {q} (\\mathbf {X})) ] \\tag {2}", + "image_path": "58ef10a14626e1adfb389bc0f98ecfcbf7380f66db9f6fe4aa449ea7f1b8bc9f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 418, + 157, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 157, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 157, + 431 + ], + "type": "text", + "content": "4.3.1 Target adaptation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 437, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 568 + ], + "type": "text", + "content": "To adapt CLIP on the target domain, we adopt a similar approach to that outlined for the source domain, with one key difference: the absence of ground-truth labels for the target domain. While it is a common practice among the source-free methods to utilize the source model to pseudolabel the unlabelled target data [20], we refrain from such an approach, and instead utilize the original CLIP model (i.e., without any source adapter " + }, + { + "bbox": [ + 46, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{S}}" + }, + { + "bbox": [ + 46, + 437, + 287, + 568 + ], + "type": "text", + "content": ") for pseudo-labelling. Despite learning fewer parameters with the adapter, the source model can still be biased towards the source, and hence, yielding noisy pseudo-labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 569, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 666 + ], + "type": "text", + "content": "To pseudo-label (PL), we utilize zero-shot CLIP inference (described in Sec. 4.2), with the same collection of textual descriptions used during the source fine-tuning stage, to assign a PL to each sample in the target domain. To this end, we use a confidence threshold for filtering out unreliable PLs (see Supplement for details). After the thresholding, we end up with a subset " + }, + { + "bbox": [ + 46, + 569, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}^{\\mathrm{T}} = \\{\\mathbf{X}_i^{\\mathrm{T}}\\}_{i=1}^{m'} \\subset \\mathcal{D}^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 569, + 287, + 666 + ], + "type": "text", + "content": ", containing PLs as " + }, + { + "bbox": [ + 46, + 569, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\{\\tilde{y}_i\\}_{i=1}^{m'}" + }, + { + "bbox": [ + 46, + 569, + 287, + 666 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 569, + 287, + 666 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 46, + 569, + 287, + 666 + ], + "type": "text", + "content": " denoting the filtered samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "content": "Following the source fine-tuning procedure, we append an adapter " + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{T}}(\\cdot): \\mathcal{R}^{d} \\to \\mathcal{R}^{d}" + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "content": " to the vision encoder of CLIP and minimize the following loss, while keeping CLIP frozen:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 347, + 317, + 545, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 317, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 347, + 317, + 545, + 333 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a d a p t e r}} ^ {\\mathrm {T}} = \\mathbb {E} _ {(\\mathbf {X}, t) \\sim \\tilde {\\mathcal {D}} ^ {\\mathrm {T}}} [ \\operatorname {K L} (\\mathbf {p} (\\mathbf {X}), \\mathbf {q} (\\mathbf {X})) ] \\tag {3}", + "image_path": "f1a50c8a44314bc138f93e59de325e2ba4210d489db8f0f591c392db9b3dd097.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 344, + 429, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 344, + 429, + 355 + ], + "spans": [ + { + "bbox": [ + 305, + 344, + 429, + 355 + ], + "type": "text", + "content": "4.3.2 Ensemble distillation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 363, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 541 + ], + "type": "text", + "content": "With three different models at our disposal: the source, target and the original CLIP, we can distill information from these three models into a single network. Knowledge Distillation [8] presents two important advantages: (i) allows to generate more reliable and informative pseudo-labels, and (ii) distillation into a smaller network reduces inference time without sacrificing much performance. For this second point, we explicitly use CLIP ViT-B/32 as teacher (and as backbone for the source and target adapters) and CLIP RN50 as student. Since we use an ensemble of three models to distill information, we call this step ensemble distillation. Now, with predictions available from three different models, the question is how to distill them together. A simple approach could be to average the three probability distributions as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 329, + 551, + 545, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 551, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 329, + 551, + 545, + 585 + ], + "type": "interline_equation", + "content": "\\mathbf {p} ^ {\\text {e n s}} (\\mathbf {X}) = \\frac {1}{3} \\left(\\underbrace {\\mathbf {p} (\\mathbf {X})} _ {\\text {Z S C L I P}} + \\underbrace {\\mathbf {p} ^ {\\mathrm {S}} (\\mathbf {X})} _ {\\text {S o u r c e a d p a t e r}} + \\underbrace {\\mathbf {p} ^ {\\mathrm {T}} (\\mathbf {X})} _ {\\text {T a r g e t a d p a t e r}}\\right) \\tag {4}", + "image_path": "44119ccb366ffcc56a65e0b69c31029a18b7022f675be3f89518556a256d1ff2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 588, + 545, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 545, + 684 + ], + "type": "text", + "content": "Whereas, another approach could consist in carrying out majority voting, where each of the three heads votes for a class, and the class with two or more votes becomes the new PL. Following [8], we utilize both kinds of PLs for distillation, using a standard multi-class classification loss (or cross-entropy loss " + }, + { + "bbox": [ + 304, + 588, + 545, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 304, + 588, + 545, + 684 + ], + "type": "text", + "content": ") and a discrepancy loss. The discrepancy loss corresponding to the ensembled PL in Eq. (4) is given as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 692, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 545, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s c r e p a n c y}} = \\mathbb {E} _ {(\\mathbf {X}, t) \\sim \\mathcal {D} ^ {\\tau}} \\left[ \\mathrm {K L} \\left(\\frac {\\mathbf {p} ^ {\\text {s t u d e n t}} (\\mathbf {X})}{\\tau}, \\frac {\\mathbf {p} ^ {\\text {e n s}} (\\mathbf {X})}{\\tau}\\right) \\right] \\tag {5}", + "image_path": "ace76c6060e65f3a7651ccca3381622e4e5de7ce89e707ab8e2345fa6942071c.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10311" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{\\mathrm{student}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": " denotes the student network's probability distribution, which is matched with that of " + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{\\mathrm{ens}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 97, + 277, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 97, + 277, + 109 + ], + "spans": [ + { + "bbox": [ + 59, + 97, + 277, + 109 + ], + "type": "text", + "content": "We then balance the contribution of the two losses as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 96, + 124, + 287, + 136 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 124, + 287, + 136 + ], + "spans": [ + { + "bbox": [ + 96, + 124, + 287, + 136 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t i l l}} = \\alpha * \\mathcal {L} _ {c e} + (1 - \\alpha) * \\mathcal {L} _ {d} \\tag {6}", + "image_path": "2b6a7cf661ea9e245a505321a61e431bd65025a73a515228ef77a1f667a4a339.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 145, + 287, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 145, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 145, + 287, + 194 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 145, + 287, + 194 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 145, + 287, + 194 + ], + "type": "text", + "content": " is set according to the value proposed in [8]. Note that, we do not fine-tune all the student weights, but instead append an adapter on top of the frozen backbone and fine-tune only the adapter " + }, + { + "bbox": [ + 47, + 145, + 287, + 194 + ], + "type": "inline_equation", + "content": "\\overline{A}" + }, + { + "bbox": [ + 47, + 145, + 287, + 194 + ], + "type": "text", + "content": " using Eq. (6)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 206, + 128, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 128, + 219 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 128, + 219 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 226, + 162, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 226, + 162, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 226, + 162, + 239 + ], + "type": "text", + "content": "5.1. Experimental setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 246, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 246, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 287, + 435 + ], + "type": "text", + "content": "Datasets and settings. We present an extensive experimental evaluation on three standard benchmarks for SFVUDA. In particular, we report results on Daily-DA [44], which comprises 18,949 videos from 8 classes, and it is built from 4 original video action recognition datasets, namely HMDB51 [15], ARID [43], MIT [26] and Kinetics [12]. Additionally, we evaluate our framework on UCF-HMDBfull [1], a benchmark comprising 3,209 videos divided in 12 action categories from the HMDB51 [15] and UCF101 [38] action recognition datasets. Note that the first benchmark, Daily-DA, poses a more significant challenge when compared to the latter since it comprises videos with very different lighting conditions across domains. Last, we test our method on Sports-DA [45], a benchmark consisting of three datasets (i.e., UCF101 [38], Sports-1M [11], and Kinetics [12]), with 40,718 videos and 23 classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "text", + "content": "Implementation details. We implement our adapters as two-layers perceptrons, following [6]. In accordance with [40], we use AdamW optimizer [24] with a learning rate of 0.01 and a weight decay of 0.2. The temperature " + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 437, + 287, + 533 + ], + "type": "text", + "content": " is set to 2.0 for all datasets. We trained all our models for 30 epochs. For a single experiment, we have used either 4 Tesla V100 or 2 RTX A6000 GPUs. Further implementation details can be found in the Supplement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 534, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 288, + 714 + ], + "type": "text", + "content": "Baselines and competitors. We compare our method with a selection of standard baseline methods for VUDA and SFVUDA, which include TRN [49], DANN [5], MK-MDD [23], " + }, + { + "bbox": [ + 47, + 534, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{TA}^3\\mathrm{N}" + }, + { + "bbox": [ + 47, + 534, + 288, + 714 + ], + "type": "text", + "content": " [1], SFDA [13], SHOT [20], SHOT++ [21], MA [19], BAIT [48] and CPGA [30]. Additionally, we report the scores of the state-of-the-art competitors ATCoN [44] and EXTERN [46]. We also report the results for a set of CLIP-based baselines of our design to provide a more solid context for assessing our performance. In particular, CLIP (RN50) and CLIP (ViT-B/32) indicate the scores obtained in a zero-shot fashion with the CLIP model, with ResNet-50 and ViT-B/32 as backbones, respectively. Finally, we report the lower and upper bounds obtained by the source-only and target-supervised models, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 505, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 505, + 85 + ], + "type": "text", + "content": "5.2. Comparison with the State-of-the-Art" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 99, + 545, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 99, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 304, + 99, + 545, + 208 + ], + "type": "text", + "content": "We report our experimental validation scores in Tab. 1, Tab. 2, and Tab. 3, obtained on the Daily-DA, the UCF-HMDB " + }, + { + "bbox": [ + 304, + 99, + 545, + 208 + ], + "type": "inline_equation", + "content": "f_{ull}" + }, + { + "bbox": [ + 304, + 99, + 545, + 208 + ], + "type": "text", + "content": ", and the Sports-DA benchmarks respectively. For all the benchmarks, the lower bound refers to the out-of-the-box performance of our source-trained model on the target data, i.e., the CLIP backbone and the source adapter " + }, + { + "bbox": [ + 304, + 99, + 545, + 208 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{S}}" + }, + { + "bbox": [ + 304, + 99, + 545, + 208 + ], + "type": "text", + "content": ". Analogously, the upper bound is the oracle performance obtained after training supervisedly on the target data, i.e., assuming knowledge of the target labels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 217, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 384 + ], + "type": "text", + "content": "Regarding the more challenging Daily-DA dataset, it is possible to observe in Tab. 1 that our proposed distillation method DALL-V (RN50) obtains the best average score across the 12 settings among the competitors and baselines sharing the same backbone, with a " + }, + { + "bbox": [ + 304, + 217, + 545, + 384 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 217, + 545, + 384 + ], + "type": "text", + "content": " gain on the second best score (CLIP (RN50)). It is also visible that the best score is achieved in 9 out 12 total settings, proving our proposed distillation-based method to be significantly effective when addressing the SFVUDA task. Notably, it emerges that the scores achieved when addressing the ARID dataset " + }, + { + "bbox": [ + 304, + 217, + 545, + 384 + ], + "type": "inline_equation", + "content": "(\\rightarrow \\mathrm{A})" + }, + { + "bbox": [ + 304, + 217, + 545, + 384 + ], + "type": "text", + "content": " as target domain are visibly lower than those reported for the other settings. We associate this behavior to the challenging visual features of this particular benchmark, characterized by being shot in a low illumination environment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 393, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 501 + ], + "type": "text", + "content": "As for the small-scale UCF-HMDBfull dataset, we can observe significantly higher absolute accuracy scores on both the involved SFVUDA directions, indicating that the benchmark is significantly closer to saturation for this task. Nonetheless, it emerges from the table that our proposed method achieves a state-of-the-art average accuracy score across the two datasets (+0.6% on the best competitor) and the best score for the HMDB→UCF direction (+1.2% on the second best)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "text", + "content": "Finally, as shown in Tab. 3, our method achieves comparable results with EXTERN [46] on Sports-DA, surpassing it on three settings, and achieving the same results on another. On the two settings with UCF101 as target (i.e., " + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{K} \\rightarrow \\mathrm{U}" + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{S} \\rightarrow \\mathrm{U}" + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "text", + "content": "), our method falls behind by approximately 6 percentage points, resulting on an average score across the six scenarios slightly lower than the state-of-the-art (i.e., " + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "inline_equation", + "content": "-0.9\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "text", + "content": "). However, DALL-V still achieves an average accuracy of " + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "inline_equation", + "content": "+8.5\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "text", + "content": " w.r.t. ATCoN, further demonstrating the effectiveness of our approach for SFVUDA. Interestingly, on Sports-DA, our lower bound already achieves competitive results with other SFVUDA methods, surpassing ATCoN by " + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "inline_equation", + "content": "+2.7\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "text", + "content": " on average. The table additionally shows that CLIP zero-shot is an extremely strong baseline on the benchmark, achieving comparable results with EXTERN, i.e., " + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "inline_equation", + "content": "-2.0\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "text", + "content": " with RN50, and " + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "inline_equation", + "content": "+2.6\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "text", + "content": " with the ViT-B backbone." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10312" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 71, + 547, + 290 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 547, + 290 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 547, + 290 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 547, + 290 + ], + "type": "table", + "html": "
MethodAccuracy (%)
K→AK→HK→MM→AM→HM→KH→AH→MH→KA→HA→MA→KAvg.
ZSTRN [49]20.936.729.022.143.753.113.822.037.117.214.724.427.9
Lower bound15.647.935.734.744.661.617.525.545.114.615.517.831.3
CLIP (RN50) [31]30.550.042.230.550.062.930.542.262.950.042.262.946.4
CLIP (ViT-B/32) [31]31.349.646.031.349.665.431.346.065.449.646.065.448.1
UDADANN [5]21.237.521.722.843.358.814.229.538.220.119.727.029.5
MK-MDD [23]21.736.224.021.050.458.520.325.733.818.718.026.129.5
TA3N [1]19.937.731.521.643.055.514.425.738.414.915.623.428.5
SFUDASFDA [13]12.644.927.516.035.249.213.124.224.916.313.225.225.2
SHOT [20]12.044.629.515.336.751.013.624.221.217.114.024.325.3
SHOT++ [21]12.640.828.714.941.746.316.022.233.115.412.521.824.4
MA [19]12.845.830.017.737.453.512.925.022.216.715.224.326.1
BAIT [48]12.745.730.016.939.653.013.625.521.215.714.525.526.2
CPGA [30]13.146.030.718.139.255.113.126.225.519.216.526.726.5
SFVUDAATCoN [44]17.248.232.527.247.357.717.930.748.526.717.231.033.5
EXTERNAL [46]23.955.835.218.153.768.126.240.757.626.218.251.439.6
DALL-V (ours)24.052.547.024.065.478.124.047.076.757.945.775.051.4
Upper bound26.970.461.526.970.488.926.961.588.970.461.588.961.9
", + "image_path": "ae7dbdb9ecb1b7390949f908a7e53b832ba8b0e3e974c71ea2243a37edab2751.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 85, + 351, + 252, + 550 + ], + "blocks": [ + { + "bbox": [ + 46, + 297, + 546, + 332 + ], + "lines": [ + { + "bbox": [ + 46, + 297, + 546, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 546, + 332 + ], + "type": "text", + "content": "Table 1: Validation accuracy for Daily-DA. Bold indicates best, underline represents best with same backbone as baseline (i.e. ResNet50). Lower bound indicates a source adapter and Upper bound a target adapter, both trained supervised on CLIP (RN50)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 351, + 252, + 550 + ], + "lines": [ + { + "bbox": [ + 85, + 351, + 252, + 550 + ], + "spans": [ + { + "bbox": [ + 85, + 351, + 252, + 550 + ], + "type": "table", + "html": "
MethodAccuracy (%)
H→UU→HAvg.
ZSTRN [49]72.872.172.4
Lower bound71.676.173.8
CLIP (RN50) [31]81.086.083.5
CLIP (ViT-B/32) [31]90.389.189.7
UDADANN [5]74.475.174.8
MK-MMD [23]74.779.777.2
TA3N [1]78.184.881.5
SFUDASFDA [13]69.875.072.4
SHOT [20]74.474.474.4
SHOT++ [21]71.168.169.6
MA [19]74.467.370.9
BAIT [48]75.376.375.8
CPGA [30]75.868.172.0
SFVUDAATCoN [44]85.379.782.5
EXTERNAL [46]91.988.990.4
DALL-V (ours)93.188.991.0
Upper bound93.791.492.6
", + "image_path": "50fe8b492463eed26452ccffb9a68c514698a10df3b7957068f89fe96f023b58.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 613 + ], + "type": "text", + "content": "Table 2: Validation accuracy for UCF-HMDBfull. Bold indicates best, underline represents best with same backbone as baseline (i.e. ResNet50). Lower bound indicates a source adapter and Upper bound a target adapter, both trained supervised on CLIP (RN50)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 632, + 151, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 151, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 151, + 646 + ], + "type": "text", + "content": "5.3. Ablation analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 715 + ], + "type": "text", + "content": "We report in Tab. 4 the intermediate accuracy scores achieved at different steps of our proposed source-free pipeline for the Daily-DA and UCF-HMDBfull. In particular, we report the performance of the source supervised model and the target unsupervised one, also reporting the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": "final model and the zero-shot version of CLIP for comparison. Finally, we report the scores obtained by assembling the predictions of the three aforementioned models. The " + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{S}}" + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "inline_equation", + "content": "A^{\\mathrm{T}}" + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": " notation indicates the usage of the source and target adapters on top of the frozen CLIP model, that are trained on the respective domains as described in detail in 4.3. Next we discuss the individual contributions. Additional ablation analyses are available in the Supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 450, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 546, + 581 + ], + "type": "text", + "content": "Effectiveness of finetuning LLVM on the target. Tab. 4 clearly shows that fine-tuning our model in an unsupervised fashion on the target domain is significantly effective with respect to the target validation accuracy. The tables show indeed a gain of " + }, + { + "bbox": [ + 304, + 450, + 546, + 581 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 304, + 450, + 546, + 581 + ], + "type": "text", + "content": " on Daily-DA and " + }, + { + "bbox": [ + 304, + 450, + 546, + 581 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 304, + 450, + 546, + 581 + ], + "type": "text", + "content": " on UCF-HMDB " + }, + { + "bbox": [ + 304, + 450, + 546, + 581 + ], + "type": "inline_equation", + "content": "f_{ull}" + }, + { + "bbox": [ + 304, + 450, + 546, + 581 + ], + "type": "text", + "content": ". Additionally, it is possible to observe a further gain on Daily-DA when ensembling the target model (which gains " + }, + { + "bbox": [ + 304, + 450, + 546, + 581 + ], + "type": "inline_equation", + "content": "+2.7\\%" + }, + { + "bbox": [ + 304, + 450, + 546, + 581 + ], + "type": "text", + "content": " by itself) with CLIP, achieving a " + }, + { + "bbox": [ + 304, + 450, + 546, + 581 + ], + "type": "inline_equation", + "content": "+3.6\\%" + }, + { + "bbox": [ + 304, + 450, + 546, + 581 + ], + "type": "text", + "content": ". From these results, it emerges that fine-tuning LLVMs on the target domain, even in an unsupervised manner, is effective for the SFVUDA task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": "Effectiveness of ensembles on LLVM and domain-specific networks. In the last line of Tab. 4, it is possible to observe that ensembling predictions from the three distinct models turns out to yield the best average target accuracy score on both the considered benchmarks, resulting in a gain of " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "+5.7\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": " on Daily-DA and of " + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "inline_equation", + "content": "+2.9\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": " on UCF-HMDBfull w.r.t. the CLIP baseline. These results and those reported in the previous two paragraphs represent further proof that the three models are effectively complementary, and each contributes with additional discriminative knowledge to the resulting model. Conversely, CLIP positively contributes to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10313" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 149, + 70, + 449, + 286 + ], + "blocks": [ + { + "bbox": [ + 149, + 70, + 449, + 286 + ], + "lines": [ + { + "bbox": [ + 149, + 70, + 449, + 286 + ], + "spans": [ + { + "bbox": [ + 149, + 70, + 449, + 286 + ], + "type": "table", + "html": "
MethodAccuracy (%)
K→UK→SS→US→KU→KU→SAvg.
ZSTRN [49]86.466.985.371.049.343.367.0
Lower bound85.479.584.478.267.264.376.5
CLIP (RN50) [31]83.479.983.480.480.479.981.2
CLIP (ViT-B/32) [31]90.082.490.085.185.182.485.8
UDADANN [5]88.075.085.773.465.955.173.8
MK-MMD [23]90.267.990.973.666.155.674.0
TA3N [1]90.368.693.072.663.654.173.7
SFUDASFDA [13]86.160.085.468.055.843.666.5
SHOT [20]91.264.988.872.053.943.669.1
SHOT++ [21]90.063.188.070.344.740.966.2
MA [19]91.065.987.871.960.739.469.5
BAIT [48]92.366.688.372.857.244.770.3
CPGA [30]89.466.386.572.555.244.569.1
SEVUDAATCoN [44]93.669.790.676.065.247.973.8
EXTERNAL [46]93.773.895.482.281.272.783.2
DALL-V (ours)88.077.788.882.381.275.982.3
Upper bound93.488.393.485.685.688.389.1
", + "image_path": "35dec57ff1a25e72aac3f5f073bd98310347911e1281104e5d12bee3225e14df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 343, + 545, + 460 + ], + "blocks": [ + { + "bbox": [ + 46, + 294, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 294, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 547, + 331 + ], + "type": "text", + "content": "Table 3: Validation accuracy for Sports-DA. Bold indicates best, underline represents best with same backbone as baseline (i.e. ResNet50). Lower bound indicates a source adapter and Upper bound a target adapter, both trained supervised on CLIP (RN50)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 343, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 49, + 343, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 545, + 460 + ], + "type": "table", + "html": "
MethodAccuracy (%)
K→AnyM→AnyDaily-DAA→AnyAvg.UCF-HMDBfull
H→AnyA→AnyH→UU→HAvg.
CLIP42.348.747.653.748.190.389.189.7
AS40.0 [-2.3]55.7 [+7.0]35.3 [-12.4]35.0 [-18.7]41.5 [-6.6]91.0 [+0.7]80.5 [-8.6]85.8 [-3.9]
AT44.4 [+2.1]51.6 [+2.9]48.5 [+0.9]58.4 [+4.7]50.8 [+2.7]90.5 [+0.2]91.1 [+2.0]90.8 [+1.1]
CLIP + AS45.1 [+2.8]57.8 [+9.1]40.0 [-7.6]50.8 [-2.9]48.4 [+0.3]93.1 [+2.8]87.2 [-1.9]90.2 [+0.5]
CLIP + AT43.8 [+1.5]50.9 [+2.2]48.2 [+0.6]57.4 [+3.7]51.7 [+3.6]89.3 [-1.0]91.9 [+2.8]90.6 [+0.9]
CLIP + AS + AT46.2 [+3.9]57.3 [+8.6]49.5 [+1.9]62.1 [+8.4]53.8 [+5.7]94.9 [+4.6]90.3 [+1.2]92.6 [+2.9]
", + "image_path": "5b6884464e68b0ae3210e5d5eadc52909c5ffabae0408de9583d09950ce1cadb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 468, + 547, + 506 + ], + "lines": [ + { + "bbox": [ + 46, + 468, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 547, + 506 + ], + "type": "text", + "content": "Table 4: Ablation on the relative improvement of the different modules of our proposed DALL-V framework w.r.t. CLIP (ViT-B/32) on both considered benchmarks. Bold indicates best. The + operator indicates assembling. Due to limited space, results are aggregated by source dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 514, + 289, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 289, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 289, + 622 + ], + "type": "text", + "content": "the final score thanks to its well-known ability to generalize to unseen domains in a zero-shot fashion; on the other hand, the tuning process of domain-specific adapters on the source and target domains, although not sufficient by itself, proves beneficial thanks to its ability to capture specific features of the involved domains and the associated label sets. Note that we still opted for distilling this knowledge to a ResNet-based model in order to be directly comparable with the competitors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": "Effectiveness of using multiple templates. As a further insight into the individual contributions of the different components of our framework, we propose in Fig. 5, a simple analysis on how the number of templates for the text input at inference time impacts on the accuracy score on the target validation set of the two considered benchmarks. We observe that increasing the number of templates from just" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 514, + 547, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 575 + ], + "type": "text", + "content": "a single one to 12 (following ActionCLIP [40]) is significantly beneficial with respect to the final accuracy score. For even higher number of templates, the performance of our model reaches a plateau. Note that for the choice of the number of template we opted for following [40]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 612, + 547, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 612, + 547, + 721 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 547, + 721 + ], + "type": "text", + "content": "Modeling the feature space. We plot in Fig. 4 an UMAP [25] chart of the features produced by the zero-shot CLIP (RN50) model, the source model, and our DALL-V model, on the UCF-HMDBfull benchmark, showing the features belonging to different classes in different colors. The chart clearly highlights the ability of the proposed DALL-V method to produce well-separated clusters for different action categories, thanks to the effective combination of the complementary knowledge of intermediate models." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "10314" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 83, + 286, + 165 + ], + "blocks": [ + { + "bbox": [ + 60, + 83, + 286, + 165 + ], + "lines": [ + { + "bbox": [ + 60, + 83, + 286, + 165 + ], + "spans": [ + { + "bbox": [ + 60, + 83, + 286, + 165 + ], + "type": "image", + "image_path": "797cbb1017be164c8da2199bd4bc10da16b4260067b5c7730f71f69a737166cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 170, + 203, + 179 + ], + "lines": [ + { + "bbox": [ + 141, + 170, + 203, + 179 + ], + "spans": [ + { + "bbox": [ + 141, + 170, + 203, + 179 + ], + "type": "text", + "content": "(a) HMDB " + }, + { + "bbox": [ + 141, + 170, + 203, + 179 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 141, + 170, + 203, + 179 + ], + "type": "text", + "content": " UCF" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 310, + 82, + 533, + 165 + ], + "blocks": [ + { + "bbox": [ + 310, + 82, + 533, + 165 + ], + "lines": [ + { + "bbox": [ + 310, + 82, + 533, + 165 + ], + "spans": [ + { + "bbox": [ + 310, + 82, + 533, + 165 + ], + "type": "image", + "image_path": "dc039be3a8c03a55a1e2965d8cf3f6b063ed9fc60f022905b4d1e49363428c43.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 170, + 453, + 179 + ], + "lines": [ + { + "bbox": [ + 392, + 170, + 453, + 179 + ], + "spans": [ + { + "bbox": [ + 392, + 170, + 453, + 179 + ], + "type": "text", + "content": "(b) UCF " + }, + { + "bbox": [ + 392, + 170, + 453, + 179 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 392, + 170, + 453, + 179 + ], + "type": "text", + "content": " HMDB" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 64, + 229, + 272, + 365 + ], + "blocks": [ + { + "bbox": [ + 46, + 183, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 547, + 220 + ], + "type": "text", + "content": "Figure 4: UMAP visualisation of the feature space on " + }, + { + "bbox": [ + 46, + 183, + 547, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{UCF - HMDB}_{full}" + }, + { + "bbox": [ + 46, + 183, + 547, + 220 + ], + "type": "text", + "content": " for the zero-shot CLIP (RN50), source and final DALL-V models. The chart shows the ability of our proposed DALL-V to efficiently cluster the action categories in the features space by combining knowledge from the other two models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 229, + 272, + 365 + ], + "lines": [ + { + "bbox": [ + 64, + 229, + 272, + 365 + ], + "spans": [ + { + "bbox": [ + 64, + 229, + 272, + 365 + ], + "type": "image", + "image_path": "e7ea954910da37e4eb4a11593610f5774c5305f10062a750494e0e89f3c21843.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 373, + 288, + 409 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 288, + 409 + ], + "type": "text", + "content": "Figure 5: Sensitivity study on the number of templates used for the text input w.r.t. the accuracy score, on Daily-DA and UCF-HMDB " + }, + { + "bbox": [ + 46, + 373, + 288, + 409 + ], + "type": "inline_equation", + "content": "f_{ull}" + }, + { + "bbox": [ + 46, + 373, + 288, + 409 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 424, + 111, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 111, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 111, + 436 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 441, + 287, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 573 + ], + "type": "text", + "content": "Note that CLIP has been trained on web-crawled image-text pairs that are not made publicly available by OpenAI [35]. Due to its black-box nature, equally trusting CLIP predictions, as much as the source model, may elicit unpredictable behaviour, especially in safety critical applications. Thus, appropriate caution must be exercised. Moreover, ours being an empirical work, we can not provide any theoretical guarantee if using LLVMs, such as CLIP, will always result in better performance over traditional SFVUDA approaches. Thus, as future work, we plan to extensively evaluate using more open-source alternatives." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 586, + 125, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 125, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 125, + 597 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "We presented DALL-V, a SFVUDA method based on a simple but novel approach to SFVUDA driven by the intuition of combining the complementary information derived from domain-specific simple models and the powerful CLIP-based LLVMs trained on world knowledge. We provided motivation for such approach, and reported an extensive evaluation on three standard benchmarks for VUDA, pursued in our case for the source-free scenario. We compare our performance with existing methods, as well as with" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 227, + 546, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 227, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 227, + 546, + 262 + ], + "type": "text", + "content": "a selection of CLIP-based baselines, showing that our proposed frameworks achieves state-of-the-art performance on both considered benchmarks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 277, + 547, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 277, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 547, + 421 + ], + "type": "text", + "content": "Acknowledgments. We acknowledge the support of the MUR PNRR project FAIR - Future AI Research (PE00000013) funded by the NextGenerationEU. E.R. is partially supported by the PRECRISIS, funded by the EU Internal Security Fund (ISFP-2022-TFI-AG-PROTECT-02-101100539), the EU project SPRING (No. 871245), and the by the PRIN project LEGO-AI (Prot. 2020TA3K9N). The work was carried out in the Vision and Learning joint laboratory of FBK and UNITN. This paper has been further supported by the French National Research Agency (ANR-20-CE23-0027), the EU Horizon 2020 Research and Innovation program under grant agreement No 957337." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 429, + 365, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 429, + 365, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 429, + 365, + 441 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 449, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 310, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 310, + 449, + 545, + 492 + ], + "type": "text", + "content": "[1] Min-Hung Chen, Zsolt Kira, Ghassan AlRegib, Jaekwon Yoo, Ruxin Chen, and Jian Zheng. Temporal attentive alignment for large-scale video domain adaptation. In ICCV, 2019. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 493, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 493, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 311, + 493, + 547, + 536 + ], + "type": "text", + "content": "[2] Victor G Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-head contrastive domain adaptation for video action recognition. In WACV, 2022. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 537, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 537, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 311, + 537, + 545, + 569 + ], + "type": "text", + "content": "[3] Christoph Feichtenhofer, Axel Pinz, and Richard P Wildes. Spatiotemporal multiplier networks for video action recognition. In CVPR, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 311, + 570, + 545, + 602 + ], + "type": "text", + "content": "[4] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In CVPR, 2016. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 603, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 603, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 311, + 603, + 545, + 624 + ], + "type": "text", + "content": "[5] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation, 2015. 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 311, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 311, + 625, + 545, + 669 + ], + "type": "text", + "content": "[6] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv, 2021. 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 311, + 670, + 545, + 712 + ], + "type": "text", + "content": "[7] Muhammad Ghifary, W Bastiaan Kleijn, Mengjie Zhang, David Balduzzi, and Wen Li. Deep reconstruction-classification networks for unsupervised domain adaptation. In ECCV, 2016. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10315" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 286, + 95 + ], + "type": "text", + "content": "[8] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv, 2015. 4, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 286, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 286, + 118 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 286, + 118 + ], + "type": "text", + "content": "[9] Yunzhong Hou and Liang Zheng. Visualizing adapted knowledge in domain transfer. In CVPR, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 120, + 286, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 120, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 120, + 286, + 163 + ], + "type": "text", + "content": "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 166, + 286, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 166, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 49, + 166, + 286, + 209 + ], + "type": "text", + "content": "[11] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, pages 1725–1732, 2014. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 211, + 286, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 211, + 286, + 264 + ], + "spans": [ + { + "bbox": [ + 49, + 211, + 286, + 264 + ], + "type": "text", + "content": "[12] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset, 2017. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 267, + 286, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 267, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 49, + 267, + 286, + 300 + ], + "type": "text", + "content": "[13] Youngeun Kim, Donghyeon Cho, Kyeongtak Han, Priyadarshini Panda, and Sungeun Hong. Domain adaptation without source data, 2020. 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 301, + 286, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 301, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 49, + 301, + 286, + 323 + ], + "type": "text", + "content": "[14] Yu Kong and Yun Fu. Human action recognition and prediction: A survey. IJCV, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 325, + 286, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 325, + 286, + 356 + ], + "spans": [ + { + "bbox": [ + 49, + 325, + 286, + 356 + ], + "type": "text", + "content": "[15] H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre. Hmdb: A large video database for human motion recognition. In ICCV, 2011. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 359, + 286, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 359, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 359, + 286, + 392 + ], + "type": "text", + "content": "[16] Vinod K Kurmi, Venkatesh K Subramanian, and Vinay P Namboodiri. Domain impression: A source data free domain adaptation method. In WACV, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 394, + 286, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 394, + 286, + 426 + ], + "spans": [ + { + "bbox": [ + 49, + 394, + 286, + 426 + ], + "type": "text", + "content": "[17] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In CVPR, 2009. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 428, + 286, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 428, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 49, + 428, + 286, + 460 + ], + "type": "text", + "content": "[18] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 463, + 286, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 463, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 49, + 463, + 286, + 495 + ], + "type": "text", + "content": "[19] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, 2020. 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 498, + 286, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 498, + 286, + 539 + ], + "spans": [ + { + "bbox": [ + 49, + 498, + 286, + 539 + ], + "type": "text", + "content": "[20] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, 2020. 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 543, + 286, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 543, + 286, + 585 + ], + "spans": [ + { + "bbox": [ + 49, + 543, + 286, + 585 + ], + "type": "text", + "content": "[21] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. TPAMI, 44(11):8602-8617, 2021. 3, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 588, + 286, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 588, + 286, + 620 + ], + "spans": [ + { + "bbox": [ + 49, + 588, + 286, + 620 + ], + "type": "text", + "content": "[22] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, 2015. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 622, + 286, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 622, + 286, + 654 + ], + "spans": [ + { + "bbox": [ + 49, + 622, + 286, + 654 + ], + "type": "text", + "content": "[23] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I. Jordan. Learning transferable features with deep adaptation networks, 2015. 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 657, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 657, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 49, + 657, + 286, + 677 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 49, + 680, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 680, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 680, + 286, + 712 + ], + "type": "text", + "content": "[25] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction, 2018. 8" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[26] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, and Aude Oliva. Moments in time dataset: one million videos for event understanding. TPAMI, 42(2):502-508, 2019. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 129, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 160 + ], + "type": "text", + "content": "[27] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In CVPR, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 163, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 194 + ], + "type": "text", + "content": "[28] Boxiao Pan, Zhangjie Cao, Ehsan Adeli, and Juan Carlos Niebles. Adversarial cross-domain action recognition with co-attention. In AAAI CAI, 2020. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[29] Preksha Pareek and Ankit Thakkar. A survey on video-based human action recognition: recent updates, datasets, challenges, and applications. Artificial Intelligence Review, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 242, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 284 + ], + "type": "text", + "content": "[30] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation, 2021. 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 287, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 341 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 343, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 374 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 376, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 418 + ], + "type": "text", + "content": "[33] Aadarsh Sahoo, Rutav Shah, Rameswar Panda, Kate Saenko, and Abir Das. Contrast and mix: Temporal contrastive video domain adaptation with background mixing. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 422, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 475 + ], + "type": "text", + "content": "[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 478, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 531 + ], + "type": "text", + "content": "[35] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv, 2021. 1, 3, 9" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 533, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 565 + ], + "type": "text", + "content": "[36] Javier Selva, Anders S Johansen, Sergio Escalera, Kamal Nasrollahi, Thomas B Moeslund, and Albert Clapés. Video transformers: A survey. arXiv, 2022. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 568, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 610 + ], + "type": "text", + "content": "[37] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "text", + "content": "[38] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild, 2012. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 647, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 677 + ], + "type": "text", + "content": "[39] Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. In Artificial Intelligence, volume 30, 2016. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[40] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv, 2021. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10316" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 531 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[41] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Conference on Empirical Methods in Natural Language Processing, pages 6787-6800, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[42] Yuecong Xu, Haozhi Cao, Kezhi Mao, Zhenghua Chen, Lihua Xie, and Jianfei Yang. Aligning correlation information for domain adaptation in action recognition. Neural Networks and Learning Systems, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 239 + ], + "type": "text", + "content": "[43] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Arid: A new dataset for recognizing action in the dark. In International Workshop on Deep Learning for Human Activity Recognition, pages 70-84. Springer, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 283 + ], + "type": "text", + "content": "[44] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, and Zhenghua Chen. Source-free video domain adaptation by learning temporal consistency for action recognition. In ECCV, 2022. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 327 + ], + "type": "text", + "content": "[45] Yuecong Xu, Jianfei Yang, Haozhi Cao, Keyu Wu, Min Wu, Rui Zhao, and Zhenghua Chen. Multi-source video domain adaptation with temporal attentive moment alignment. arXiv, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 373 + ], + "type": "text", + "content": "[46] Yuecong Xu, Jianfei Yang, Haozhi Cao, Min Wu, Xiaoli Li, Lihua Xie, and Zhenghua Chen. Leveraging endo- and exotemporal regularization for black-box video domain adaptation, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 375, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 375, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 287, + 418 + ], + "type": "text", + "content": "[47] Jinyu Yang, Weizhi An, Sheng Wang, Xinliang Zhu, Chaochao Yan, and Junzhou Huang. Label-driven reconstruction for domain adaptation in semantic segmentation. In ECCV, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 420, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 287, + 453 + ], + "type": "text", + "content": "[48] Shiqi Yang, Yaxing Wang, Joost Weijer, and Luis Herranz. Unsupervised domain adaptation without source data by casting a bait, 2020. 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 454, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 485 + ], + "type": "text", + "content": "[49] Bolei Zhou, Alex Andonian, Aude Oliva, and Antonio Torralba. Temporal relational reasoning in videos, 2018. 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 487, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 531 + ], + "type": "text", + "content": "[50] Yi Zhu, Xinyu Li, Chunhui Liu, Mohammadreza Zolfaghari, Yuanjun Xiong, Chongruo Wu, Zhi Zhang, Joseph Tighe, R Manmatha, and Mu Li. A comprehensive study of deep video action recognition. arXiv, 2020. 1" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10317" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/fdfec7d9-d3ec-4468-9f42-c0ec2550345d_content_list.json b/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/fdfec7d9-d3ec-4468-9f42-c0ec2550345d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..835d277c8c1df284285ac797640458a1d6951a4a --- /dev/null +++ b/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/fdfec7d9-d3ec-4468-9f42-c0ec2550345d_content_list.json @@ -0,0 +1,1483 @@ +[ + { + "type": "text", + "text": "The Victim and The Beneficiary: Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data", + "text_level": 1, + "bbox": [ + 86, + 128, + 883, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zixuan Zhu, Rui Wang*, Cong Zou, Lihua Jing \nSKLOIS, Institute of Information Engineering, CAS, Beijing, China \nSchool of Cyber Security, University of Chinese Academy of Sciences, Beijing, China \n{zhuzixuan, wangrui, zoucong, jinglihua}@iei.ac.cn", + "bbox": [ + 143, + 203, + 825, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, backdoor attacks have posed a serious security threat to the training process of deep neural networks (DNNs). The attacked model behaves normally on benign samples but outputs a specific result when the trigger is present. However, compared with the rocketing progress of backdoor attacks, existing defenses are difficult to deal with these threats effectively or require benign samples to work, which may be unavailable in real scenarios. In this paper, we find that the poisoned samples and benign samples can be distinguished with prediction entropy. This inspires us to propose a novel dual-network training framework: The Victim and The Beneficiary (V&B), which exploits a poisoned model to train a clean model without extra benign samples. Firstly, we sacrifice the Victim network to be a powerful poisoned sample detector by training on suspicious samples. Secondly, we train the Beneficiary network on the credible samples selected by the Victim to inhibit backdoor injection. Thirdly, a semi-supervised suppression strategy is adopted for erasing potential backdoors and improving model performance. Furthermore, to better inhibit missed poisoned samples, we propose a strong data augmentation method, AttentionMix, which works well with our proposed V&B framework. Extensive experiments on two widely used datasets against 6 state-of-the-art attacks demonstrate that our framework is effective in preventing backdoor injection and robust to various attacks while maintaining the performance on benign samples. Our code is available at https://github.com/Zixuan-Zhu/VaB.", + "bbox": [ + 76, + 340, + 472, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 777, + 209, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large amounts of training data and powerful computing resources are two key factors for the success of deep neural networks (DNNs). While in practical applications, obtaining enough samples for training is laborious, and training is likely to be resource-constrained. As a result, more and", + "bbox": [ + 76, + 801, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7272a10b706b0a05e6e2a427edef074ab60c8415e1eb25313c2d3eb57128a5f4.jpg", + "image_caption": [ + "Figure 1. The average prediction entropy of benign samples versus poisoned samples crafted by 6 backdoor attacks. We conduct the experiment on CIFAR-10 with ResNet-18 under poisoning rate $10\\%$ , where our special training strategy (described in section 3.3) was adopted." + ], + "image_footnote": [], + "bbox": [ + 537, + 306, + 852, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "more third-party data (e.g. data released on the Internet) and training platforms (e.g. Google Colab) are adopted to reduce the overhead, which brings new security risks.", + "bbox": [ + 498, + 580, + 892, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Backdoor attacks [5, 11, 24] pose a serious security threat to the training process of DNNs and can be easily executed through data poisoning. Specifically, attackers [7, 10, 22] inject the designed trigger (e.g. a small patch or random noise) to a few benign samples selected from the training set and change their labels to the attacker-defined target label. These poisoned samples will force models to learn the correlation between the trigger and the target label. In the reference process, the attacked model will behave normally on benign samples but output the target label when the trigger is present. Deploying attacked models can lead to severe consequences, even life-threatening in some scenarios (e.g. autonomous driving). Hence, a secure training framework is needed when using third-party data or training platforms.", + "bbox": [ + 496, + 627, + 893, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many researchers have devoted themselves to this topic and proposed feasible defenses. Li et al. [18] utilize local benign samples to erase existing backdoors in DNNs, and", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "155", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Borgnia et al. [3] employ strong data augmentations to inhibit backdoor injection during training. Li et al. [17] find that the training loss decreased faster for poisoned samples than benign samples and designed a loss function to separate them. Finally, they select a fixed portion of samples with the lowest loss to erase potential backdoors. In real scenarios, it is challenging to distinguish poisoned samples from benign ones in this way due to the unknown poisoning rate and close loss values. In experiments, we find the entropy of model prediction is a more discriminative property, and a fixed threshold could filter out most poisoned samples, as shown in Figure 1. During training, poisoned samples will be learned faster than benign samples due to the similar feature of their triggers [17]. Hence, the poisoned network can confidently predict poisoned samples as the target label in early epochs but hesitates about the benign samples.", + "bbox": [ + 80, + 90, + 467, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on this observation, we propose a novel secure training framework, The Victim and The Beneficiary (V&B), that can directly train clean models on poisoned data without resorting to benign samples. Firstly, in warming up stage, the training dataset is recomposed into a suspicious set and a credible set according to the prediction entropy is lower or higher than the predefined threshold. Then the Victim network is trained iteratively with the suspicious set to obtain a strong poisoned sample detector. Secondly, in the clean training stage, the Beneficiary network is trained with credible samples filtered by the Victim to inhibit backdoor injection. In later training epochs, the Victim network is fully exploited to provide its learned knowledge for the Beneficiary network to help the latter erase potential backdoors through semi-supervised learning. To diminish the threat of missed poisoned samples, we design a strong data augmentation AttentionMix, which mixes the image region with high activation values according to the attention map. Compared with existing augmentations [27, 30, 31], it has a stronger inhibition effect against stealthy backdoor attacks [23, 22, 26].", + "bbox": [ + 80, + 349, + 467, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are as follows: (1) We propose a novel dual-network secure training framework, V&B, which can train clean models on the poisoned dataset without resorting to benign samples. (2) We design a powerful data augmentation method called AttentionMix, which has a stronger inhibition effect against stealthy backdoor attacks. (3) Extensive experiments on two popular benchmark datasets demonstrate the effectiveness and robustness of our framework.", + "bbox": [ + 80, + 667, + 467, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 80, + 820, + 215, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Data Augmentation", + "text_level": 1, + "bbox": [ + 80, + 845, + 261, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data Augmentation is a common-used technique to improve the generalization capabilities of DNNs. In addition", + "bbox": [ + 80, + 871, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to some popular weak data augmentations (such as cropping, flipping, rotation, etc.), many strong data augmentation techniques [27, 30, 31] have been proposed to further improve model performance. Mixup [31] linearly interpolates any two images and their labels, then trains the model with generated samples. However, the generated samples tend to be unnatural, which will decrease model performances on localization and object detection tasks. Instead of linear interpolation, Cutmix [30] replaces a random patch with the same region from another image and mixes their labels according to the ratio of the patch area. Further, Attentive Cutmix [27] only pastes the influential region located by an attention mechanism to other images, effectively forcing models to learn discriminative details. Yet it also mixes labels based on the ratio of the region area, ignoring the importance of the region in two original images, and direct pasting may bring complete triggers into benign samples. Our AttentionMix takes into account the importance of the region in both images and blends the influential region with the same area in another image, which can destroy the completeness of triggers.", + "bbox": [ + 501, + 90, + 888, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Backdoor Attack and Defense", + "text_level": 1, + "bbox": [ + 503, + 421, + 763, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we only consider poisoning-based attacks toward image classification that can occur during the data preparation stage.", + "bbox": [ + 503, + 446, + 888, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Poisoned-based backdoor attack. In early studies, attackers usually adopt simple and visible patterns as the trigger, such as a black-white checkerboard [10]. To escape human inspections, some works use human-imperceptible deformations [22] as the trigger or design unique triggers [16, 23] for each poisoned sample. However, they randomly select benign samples to inject triggers and change their labels (called poison-label attacks), possibly causing the content of an image not to match its label. Instead of randomly selecting samples to poison, Turner et al. [26], Mauro et al. [1], and Liu et al. [21] only inject triggers into benign samples under the target label (called clean-label attacks). Although some triggers are imperceptible in the input space, they can be easily detected by defenses focused on latent space extracted by CNNs. Therefore, recent studies [33, 34] not only ensure the trigger's invisibility in the input space but also restrict the similarity between poison samples and corresponding benign samples in the latent space.", + "bbox": [ + 503, + 505, + 888, + 775 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backdoor Defense. For poisoning-based attacks, defense methods could be divided into three categories from their working stage, including (1) preprocessing before training [9, 12, 25], (2) poison suppression during training [14, 17, 20], and (3) backdoor erasing after training [18, 28, 32]. For preprocessing, Hayase et al. [12] and Chen et al. [4] filter the training set through statistics or clustering and", + "bbox": [ + 503, + 795, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "156", + "bbox": [ + 488, + 946, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "train the model with benign samples. For poison suppressing, Borgnia et al. [3] leverage strong data augmentation (e.g. Mixup, Cutmix) to alleviate poisoned samples' adverse effects. Huang et al. [14] design a decoupling training process, which first learns a feature extractor via SimCLR [6] to prevent poisoned samples from clustering in feature space. For backdoor erasing, Liu et al. [19] prune the low-activation neurons for validation samples (benign samples) and then fine-tune the model with these benign samples. While Li et al. [18] first fine-tune the poisoned model with a small portion of benign samples to obtain a teacher model, then erase the original model's backdoor through knowledge distillation.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Semi-supervised Learning", + "text_level": 1, + "bbox": [ + 76, + 297, + 316, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semi-supervised learning [8] aims to alleviate the dependence of deep learning on extensive labeled training data and jointly learns the model from a limited amount of labeled samples and a large number of unlabeled samples. Existing methods [2, 29, 35] generally assign each unlabeled sample with a pseudo label and design different loss functions for labeled and unlabeled samples. The weight of unlabeled sample loss will gradually increase during training to reduce the negative effect of low-quality pseudo labels in early epochs.", + "bbox": [ + 75, + 321, + 472, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 487, + 169, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 512, + 217, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pipeline of Poisoned-based Backdoor Attacks. Let $\\mathcal{D}_B = \\{(x_i, y_i)\\}_{i=1}^N$ denotes the benign training set containing $N$ samples, where $x_i \\in \\{0, 1, \\dots, 255\\}^{C \\times H \\times W}$ is the image, $y_i \\in \\{0, 1, \\dots, K\\}$ is its label, $K$ is the number of classes, $H, W$ and $C$ are the height, width and channels of an image. The attacker will select a subset of benign samples to poison with a designed injection function $G(\\cdot)$ . Let $\\mathcal{D}_G$ indicates the generated poisoned samples set and $\\mathcal{D}_R$ denotes the remaining benign samples set, where $\\mathcal{D}_G = \\{(x', y_t) | x' = G(x), (x, y) \\in \\mathcal{D}_B \\setminus \\mathcal{D}_R\\}$ , $y_t$ is the target label and $r = \\frac{\\left|\\mathcal{D}_G\\right|}{\\mathcal{D}_B}$ is the poisoning rate. The poisoned dataset $\\mathcal{D}_P = \\mathcal{D}_G \\cup \\mathcal{D}_R$ will be unknowingly adopted by victim users to train DNNs. At inference time, the attacker can manipulate the poisoned model's output through the injected backdoor.", + "bbox": [ + 75, + 536, + 472, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Threat Model and Defense Setting. In this paper, we focus on defending against poisoned-based backdoor attacks toward image classification. The attacker can arbitrarily modify the training set, but can not change other training components (e.g. model structure and hyperparameters) [1, 10, 22]. Such attacks may occur when using third-party datasets or training platforms that can poison users' clean data. We follow the common assumption [12, 14, 17] that", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "defenders have no information about the backdoor (including poisoning rate, trigger pattern, target label, etc.) and any local benign samples to use but have full control over the training process. The defense goal is to prevent the trained DNNs from predicting poisoned samples as the target label while maintaining its accuracy on benign samples.", + "bbox": [ + 496, + 90, + 890, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 500, + 191, + 609, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this subsection, we briefly introduce the pipeline of our framework. As Figure 2(a) illustrates, our framework consists of two networks, and the whole training process is divided into three stages — warming up, clean training and semi-supervised suppression training. In warming up, we aim to obtain a poisoned network that can roughly distinguish poisoned samples from benign ones according to the prediction entropy. Therefore we train the Victim network with only filtered suspicious samples (i.e. suspicious set) except for the first epoch. This training strategy remains unchanged in the following two stages. During clean training, we adopt credible samples (i.e. credible set) to train the Beneficiary network with our proposed AttentionMix data augmentation. The strong data augmentation can effectively prevent missed poisoned samples from creating backdoors. As the Victim network is trained, the entropy of its prediction for benign samples will also slowly decrease below the filtering threshold, leading to a reduction in the credible set size. So one credible set will be used for a few epochs to ensure the Beneficiary network gets enough training samples.", + "bbox": [ + 496, + 215, + 890, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To further improve the Beneficiary network performance and erase its potential backdoor, we regard the credible set as a labeled dataset and the suspicious set as an unlabeled dataset to train the Beneficiary network via semi-supervised learning. Following conventional practice [2, 15], each unlabeled sample will be assigned a pseudo label generated by the Beneficiary network, forming the relabeled set. However, if the Beneficiary network still contains a backdoor after stage 2, poisoned samples in the suspicious set will be relabeled as the target label, strengthening the backdoor. After specific training, the Victim network is sensitive to trigger patterns and thus can be utilized to suppress the component corresponding to the target class after relabeling. Furthermore, we also apply this suppression operation to the credible set to further weaken the impact of missed poisoned samples.", + "bbox": [ + 496, + 518, + 890, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Warming Up", + "text_level": 1, + "bbox": [ + 500, + 770, + 637, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At this stage, we aim to obtain a poisoned network that has only learned trigger patterns. In this way, the entropy of its prediction for poisoned samples will be significantly lower than that of benign samples. Then we can use a threshold to filter out poisoned samples.", + "bbox": [ + 496, + 794, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{D}$ denotes the training set and $F_{v}:\\{0,1,\\dots,255\\}^{C\\times H\\times W}\\to [0,1]^{K}$ indicates the Vic", + "bbox": [ + 500, + 869, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "157", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/666295a13902e0ab4279b53e1e326fc6bb0c5c928d306c2a0752f197d7d820e9.jpg", + "image_caption": [ + "(a) The pipeline of our training framework" + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 893, + 291 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eb3f649cbc22b76de502a6ef815778445d1c28d55a780d4a7f0c3cabe4910264.jpg", + "image_caption": [ + "Figure 2. (a) The pipeline of our training framework. Net V is the Victim network (i.e. poisoned network) that only learns from the filtered suspicious set, and Net B is the Beneficiary network (i.e. clean network) we wanted. Red triangles represent poisoned samples, and green circles are benign samples. The color of nets indicates their degree of poisoning: the darker the color, the deeper the poisoning, and vice versa. Deep blue areas are three training stages, and light blue areas are basic blocks that will be executed many times in the corresponding stage. (b) An example of AttentionMix. The bottom four images are the non-activation region of the cat image, the activation region of the cat image, the same activation region in the dog image, and the final generated image, respectively. And their labels are located below the images. (c) Suppression for poisoned samples and benign samples. The target label is assumed as the first label. Suppression can reduce the target label component of poisoned samples but have little influence on the ground-truth label of benign samples." + ], + "image_footnote": [], + "bbox": [ + 81, + 306, + 480, + 455 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fe32f823de59ee29057c059f09bfd60653437c7e5b3468c23fe139e29c8862b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 306, + 888, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tim network. We first use the whole training set to train the Victim network for one epoch with cross-entropy loss. Then the prediction entropy of a sample will be calculated with Equation 1, where $p$ is the prediction of the Victim network for that sample.", + "bbox": [ + 75, + 597, + 470, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne = - \\sum_ {k = 1} ^ {K} p _ {k} \\log p _ {k} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 680, + 468, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once all samples' prediction entropy is obtained, we normalize them to $[0,1]$ through mix-max normalization. Then a threshold $t_f$ is set to filter out suspicious samples $\\mathcal{D}_s = \\{(x_i,y_i)|(x_i,y_i)\\in \\mathcal{D},e_i < t_f\\}$ and the remaining samples are regarded as credible samples $\\mathcal{D}_c$ . To strengthen the learning of poisoned samples and delay the learning of benign samples, the Victim network will only be trained on suspicious sets in the future. The loss function is as follows:", + "bbox": [ + 76, + 729, + 470, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v} = \\mathbb {E} _ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {s}} \\left[ \\mathcal {L} _ {C E} \\left(F _ {v} (\\boldsymbol {x}), y\\right) \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 864, + 468, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{CE}$ denotes the cross-entropy loss.", + "bbox": [ + 76, + 886, + 362, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the first epoch, the prediction entropy of some poisoned samples may still be larger than that of most benign ones, although there is a large gap between their average values in Figure 1. If the filtering threshold $t_f$ is set too large, lots of benign samples will be discarded, which will affect the model performance on benign samples. So we set a smaller $t_f$ and wait for the prediction entropy of these poisoned samples to decrease. As training progresses, their prediction entropy will drop below $t_f$ , making the credible set cleaner. Simultaneously, the prediction entropy of benign samples will also decrease, resulting in a smaller credible set. We expect the Victim network to filter out a clean and relatively large credible set after this stage.", + "bbox": [ + 496, + 597, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Clean Training", + "text_level": 1, + "bbox": [ + 500, + 801, + 653, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At this stage, we begin to train the Beneficiary network with filtered credible samples $\\mathcal{D}_c$ , where the filtering strategy is the same as in stage 1. Because of the rough separation, a few poisoned samples may be included in the credible set. For strong attacks, a small number of poisoned sam", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "158", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ples is enough to create a backdoor. So we design a strong data augmentation called AttentionMix to inhibit the missed poisoned samples from working. Compared with existing augmentations [27, 30, 31], AttentionMix has a stronger inhibition effect on stealthy backdoor attacks.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The core idea of AttentionMix is to generate a new training sample $(\\tilde{\\boldsymbol{x}},\\tilde{\\boldsymbol{y}})$ by mixing two training samples $(x_{1},y_{1})$ and $(x_{2},y_{2})$ according to their attention maps, where $\\tilde{\\boldsymbol{y}}$ is a K-dimensional vector. Following [18], we adopt one sample's final feature maps to compute its attention map:", + "bbox": [ + 75, + 167, + 468, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} ^ {\\prime} = \\sum_ {i = 1} ^ {c} \\left| F _ {i} \\right| ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 252, + 468, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{F} \\in \\mathbb{R}^{c \\times h \\times w}$ indicates feature maps, $\\pmb{A}' \\in \\mathbb{R}^{h \\times w}$ is the corresponding attention map, $c, h$ and $w$ are the dimensions of channel, height and width. For each attention map, we first bilinearly interpolate it to the original image size $H \\times W$ , and then normalize its elements to $[0,1]$ by min-max normalization.", + "bbox": [ + 75, + 300, + 468, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After getting the attention map, we set a threshold $t_m$ to obtain the activation region that we want to mix:", + "bbox": [ + 75, + 392, + 468, + 421 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} _ {i, j} = \\left\\{ \\begin{array}{l l} 0 & \\boldsymbol {A} _ {i, j} < t _ {m} \\\\ 1 & \\boldsymbol {A} _ {i, j} \\geq t _ {m} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 431, + 468, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M \\in \\mathbb{R}^{H \\times W}$ is the binary mask indicating which pixels belong to the activation region and $\\pmb{A}$ is the attention map after interpolation and normalization.", + "bbox": [ + 75, + 473, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $A^1$ and $A^2$ denote the attention maps of $x_1$ and $x_2$ , $M^1$ denotes the first sample's mask. As Figure 2(b) shows, the mixed image $\\tilde{x}$ is composed of three parts: the non-activation region of $x_1$ , the activation region of $x_1$ and the same area in $x_2$ . During mixing, the non-activation region of $x_1$ remains unchanged, and the activation region is blended with the same area of $x_2$ :", + "bbox": [ + 75, + 518, + 468, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\boldsymbol {x}} = \\left(\\mathbf {1} - \\boldsymbol {M} ^ {1}\\right) \\odot \\boldsymbol {x} _ {1} + \\lambda_ {2} \\cdot \\left(\\boldsymbol {M} ^ {1} \\odot \\boldsymbol {x} _ {1}\\right) \\tag {5} \\\\ + \\lambda_ {3} \\cdot \\left(\\boldsymbol {M} ^ {1} \\odot \\boldsymbol {x} _ {2}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 633, + 468, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ is the Hadamard Product, $\\lambda_{2}$ and $\\lambda_{3}$ are the mixing weights calculated from the two attention maps. The calculation process is as follows:", + "bbox": [ + 75, + 681, + 468, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {1} = \\frac {\\operatorname {s u m} \\left(\\boldsymbol {M} ^ {1}\\right)}{\\operatorname {s u m} \\left(\\boldsymbol {A} ^ {1}\\right)}, w _ {2} = \\frac {\\operatorname {s u m} \\left(\\boldsymbol {M} ^ {1} \\odot \\boldsymbol {A} ^ {2}\\right)}{\\operatorname {s u m} \\left(\\boldsymbol {A} ^ {2}\\right)} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 734, + 468, + 771 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {1} = 1 - w _ {1}, \\lambda_ {2} = \\frac {w _ {1}}{w _ {1} + w _ {2}}, \\lambda_ {3} = \\frac {w _ {2}}{w _ {1} + w _ {2}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 781, + 468, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{sum}(\\cdot)$ is a summation function, $w_{1}$ and $w_{2}$ denote the importance of the activation region to the two images, and $\\lambda_{1}$ is the importance of the non-activation region to $x_{1}$ . The final label is also composed of three parts:", + "bbox": [ + 75, + 814, + 468, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {y}} = \\lambda_ {1} \\cdot \\boldsymbol {y} _ {1} + w _ {1} \\cdot \\left(\\lambda_ {2} \\cdot \\boldsymbol {y} _ {1} + \\lambda_ {3} \\cdot \\boldsymbol {y} _ {2}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 885, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{y}_1$ and $\\pmb{y}_2$ are the one-hot labels of the two samples. When the activation region of $\\pmb{x}_1$ corresponds to the background of $\\pmb{x}_2$ , $w_1$ will be significantly larger than $w_2$ , and $\\lambda_3$ will be a small value, which can reduce background interference during mixing.", + "bbox": [ + 498, + 90, + 890, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We denote the generated new credible set as $\\tilde{\\mathcal{D}}_c$ and train the Beneficiary network $F_{b}$ with the following loss function (i.e. Train with AttentionMix in Figure 2(a)):", + "bbox": [ + 498, + 166, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b} = \\mathbb {E} _ {(\\tilde {\\boldsymbol {x}}, \\tilde {\\boldsymbol {y}}) \\in \\tilde {\\mathcal {D}} _ {c}} [ \\mathcal {L} _ {C E} (F _ {b} (\\tilde {\\boldsymbol {x}}), \\tilde {\\boldsymbol {y}}) ] \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 224, + 890, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a poisoned network, the activation region of poisoned samples is more likely their trigger locations. Mixing the activation region can destroy the completeness of trigger patterns. While for benign samples, the mixing plays a regularization role that prevents the network from overfitting a particular subject and forces it to learn abundant features.", + "bbox": [ + 496, + 252, + 890, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Semi-supervised Suppression Training", + "text_level": 1, + "bbox": [ + 500, + 353, + 830, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Until now, we have got a poisoned Victim network that can filter out most of the poisoned samples and a Beneficiary network that is relatively clean but may still contain the backdoor. In the previous stage, the Beneficiary network is trained only with partial training samples (i.e. the credible set), which may harm its performance on benign samples. Thus we intend to enable the deprecated suspicious samples via semi-supervised learning at this stage to improve the Beneficiary network while erasing its potential backdoor.", + "bbox": [ + 496, + 376, + 890, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a credible set $\\mathcal{D}_c$ and a suspicious set $\\mathcal{D}_s$ , we use samples in $\\mathcal{D}_c$ as labeled samples and remove the labels of samples in $\\mathcal{D}_s$ to transform the problem into a semi-supervised learning setting. For each unlabeled sample, we average the predictions of the Beneficiary network over its multiple augmentations (e.g. random crop and random flip) and select the class with max probability as its pseudo-label, which is similar to [2] and [15]. The relabeled samples are denoted as $\\mathcal{D}_r = \\{(\\pmb{x}_i,\\hat{\\pmb{y}}_i)|(\\pmb{x}_i,\\pmb{y}_i)\\in \\mathcal{D}_s\\}$ .", + "bbox": [ + 496, + 527, + 890, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the poisoned samples in $\\mathcal{D}_s$ may be relabeled as the target label, i.e. $\\hat{y}_i = y_t$ , if the Beneficiary network still contains the backdoor. We can mitigate the impact of such mislabeling by exploiting the knowledge learned by the Victim network. Specifically, we adopt the Victim network's prediction to suppress the pseudo-label (i.e. Suppress in Figure 2(a)):", + "bbox": [ + 496, + 664, + 890, + 768 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {i} ^ {\\prime} = \\operatorname {c l i p} \\left(\\hat {\\boldsymbol {y}} _ {i} - \\operatorname {s o f t m a x} \\left(F _ {v} \\left(\\boldsymbol {x} _ {i}\\right)\\right), 0, 1\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 781, + 890, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(\\pmb{x}_i,\\hat{y}_i)\\in \\mathcal{D}_r$ $\\hat{\\pmb{y}}_i$ is the one-hot vector of $\\hat{y}_i$ and $\\mathrm{clip}(\\cdot ,0,1)$ is a function that will restrict its inputs to [0, 1] (i.e. inputs smaller than 0 became 0, inputs larger than 1 became 1, and the rest remain unchanged).", + "bbox": [ + 496, + 809, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use $\\mathcal{D}_r' = \\{(\\pmb{x}_i, \\pmb{y}_i') | (\\pmb{x}_i, \\hat{\\pmb{y}}_i) \\in \\mathcal{D}_r\\}$ to indicate the suspicious samples after relabeling and suppression. We", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "159", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1323a94e8b7af459a43b8cbcd3a11f911c7527ef6f35d7482dce7a62212a1729.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AttackBadNets [10]Blend [7]Dynamic [23]WaNet [22]SIG [1]CL-16 [26]CL-32 [26]
MetricBAASRBAASRBAASRBAASRBAASRBAASRBAASR
No Defense94.36%100.00%94.57%99.76%94.66%95.18%94.21%95.84%95.09%64.79%94.98%96.93%95.03%94.73%
FP [19]92.72%1.92%92.49%10.61%92.47%11.73%93.27%0.90%93.55%37.98%92.92%9.74%93.03%6.39%
NAD [18]88.83%1.83%87.72%1.38%88.30%2.32%91.00%0.99%90.42%3.52%88.92%1.64%90.19%2.60%
ABL [17]90.47%0.70%91.31%31.49%83.07%94.21%80.78%19.93%88.61%1.37%90.29%0.54%86.50%2.68%
DBD [14]92.00%3.06%91.62%3.78%90.98%17.56%89.13%0.11%89.42%1.61%92.85%91.08%93.24%14.88%
V&B (Ours)93.96%0.62%94.37%0.63%93.91%1.13%94.15%0.54%94.08%0.17%94.24%1.01%93.98%0.64%
", + "bbox": [ + 81, + 88, + 890, + 202 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/88ea00256f9633bfe7e2df3cbade1ad5d0e305477a66afc558b69326b1b2a51f.jpg", + "table_caption": [ + "Table 1. The performance of 5 defense methods against 7 backdoor attacks on CIFAR-10. The first 4 attacks are poison-label attacks and the last 3 attacks are clean-label attacks. Note that FP and NAD require local benign samples to work, while ABL, DBD and our method do not. The best results are marked in boldface and underlined items are the second-best results." + ], + "table_footnote": [], + "table_body": "
AttackBadNets [10]Blend [7]WaNet [22]SIG [1]
MetricBAASRBAASRBAASRBAASR
No Defense94.94%99.51%95.16%100.00%92.63%100.00%90.38%89.22%
FP [19]86.28%1.05%87.63%2.03%82.82%26.19%82.50%22.76%
NAD [18]85.83%3.82%86.73%2.90%83.93%3.88%80.71%3.88%
ABL [17]92.15%13.78%87.34%56.57%91.38%56.71%80.13%0.87%
DBD [14]93.08%0.31%93.81%0.87%92.06%1.00%84.47%98.47%
V&B (Ours)95.42%0.28%95.03%0.45%94.84%1.92%94.65%0.03%
", + "bbox": [ + 243, + 263, + 727, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The performance of 5 defense methods against 4 backdoor attacks on ImageNet subset.", + "bbox": [ + 199, + 380, + 767, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "also suppress the labels of credible samples in the same way to weaken the influence of poisoned samples that may be missed after filtering. As in stage 2, the suppressed credible samples are augmented with our AttentionMix, forming the training set $\\hat{\\mathcal{D}}_c'$ to further inhibit backdoor injection. We fine-tune the Beneficiary network with the following loss function:", + "bbox": [ + 75, + 405, + 470, + 511 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathbb {E} _ {(\\tilde {\\boldsymbol {x}}, \\tilde {\\boldsymbol {y}} ^ {\\prime}) \\in \\tilde {\\mathcal {D}} _ {c} ^ {\\prime}} [ \\mathcal {L} _ {C E} (F _ {b} (\\tilde {\\boldsymbol {x}}), \\tilde {\\boldsymbol {y}} ^ {\\prime}) ] + \\tag {11} \\\\ \\alpha \\cdot \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y} ^ {\\prime}) \\in \\mathcal {D} _ {r} ^ {\\prime}} [ \\mathcal {L} _ {C E} (F _ {b} (\\boldsymbol {x}), \\boldsymbol {y} ^ {\\prime}) ] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 515, + 468, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the Victim network is only trained with suspicious samples, it will confidently predict poisoned samples as the target label but give a wrong prediction for benign samples with high probability. Thus for a poisoned sample, the suppression can reduce the component corresponding to the target class in its label while having little influence on the component of a sample's ground-truth class, as shown in Figure 2(c).", + "bbox": [ + 75, + 561, + 468, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 695, + 209, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 719, + 284, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Models. We employ ResNet-18 [13] as our Victim and Beneficiary networks, a standard setting in previous studies [14]. For a fair comparison, we evaluate all defenses on two popular benchmark datasets, including CIFAR-10 and a subset of ImageNet containing 12 classes generated by Li et al. [17]. The details are summarized in Appendix A.1.", + "bbox": [ + 75, + 742, + 468, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attack Configures. We consider 6 representative backdoor attacks in our experiments, including four poison-label attacks: BadNets [10], Blend attack [7], WaNet attack [22],", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dynamic attack [23], and two clean-label attacks: Sinusoidal signal attack (SIG) [1] and Clean-label attack (CL) [26]. To achieve the best attack performance, we configure these attacks as close as possible to their released codes and original settings (e.g. trigger pattern and trigger size). While on the ImageNet subset, we implement these attacks based on the settings suggested by [14], since some attacks miss the corresponding experiments in their original papers. We uniformly set the poisoning ratio as 0.1 and select the first class as the target class in both datasets, i.e. $y_{t} = 0$ . More details about attacks are shown in Appendix A.2.", + "bbox": [ + 496, + 405, + 890, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Defense Configures and Training Details. We compare our V&B with 4 state-of-the-art defenses: Fine-pruning (FP) [19], Neural Attention Distillation (NAD) [18], Anti-Backdoor Learning (ABL) [17] and Decoupling (DBD) [14], in which FP and NAD require additional local benign samples to work. Similarly, we follow their original settings and released codes. For V&B, we warm up the Victim network for 3 epochs on CIFAR-10 and 6 epochs on the ImageNet subset, where the filtering threshold $t_f$ linearly decreases from 1 to 0.2 and from 1 to 0.4. The selection of these hyperparameters and an automatic warming-up strategy can be found in Appendix A.3. In the following stages, the filtering threshold keeps unchanged (i.e. 0.2 for CIFAR-10 and 0.4 for ImageNet subset). For fast convergence, we adopt the Adam optimizer to train the Victim network with a learning rate of 0.001 in the whole training process. We optimize the Beneficiary network for 200 epochs using an SGD optimizer with a momentum of 0.9 and a weight decay factor of $10^{-4}$ , where the first 100 epochs belong to stage 2 and the last 100 epochs belong to stage 3. The learning rate is initially set to 0.1 and is divided by 10 after every 50", + "bbox": [ + 496, + 583, + 892, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "160", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8a7d21e9e99b848c7786949fbd852c363a99505175a1332ca75b96de8fe7beef.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AttackBadNets [10]Blend [7]Dynamic [23]WaNet [22]CL-16 [26]
MetricBAASRBAASRBAASRBAASRBAASR
No Defense94.36%100%94.57%99.76%94.66%95.18%94.21%95.84%94.98%96.93%
Mixup [31]93.37%0.27%92.80%3.49%92.97%42.10%92.97%42.10%94.80%20.03%
Cutmix [30]93.52%0.40%93.40%1.07%91.40%53.23%91.40%53.23%94.54%1.16%
Attentive Cutmix [27]94.26%0.40%94.55%4.62%93.09%13.38%93.09%13.88%95.18%2.97%
AttentionMix (Ours)93.96%0.62%94.37%0.63%93.91%1.13%94.15%0.54%94.24%1.01%
", + "bbox": [ + 189, + 88, + 782, + 180 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. The inhibition effect of 4 strong data augmentations against 5 backdoor attacks on CIFAR-10.", + "bbox": [ + 179, + 189, + 784, + 203 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "epochs. In particular, the Victim network is optimized every 5 epochs at stage 2, i.e. one credible set is used to train the Beneficiary network for 5 epochs. We use a batch size of 128 for CIFAR-10 and 16 for the ImageNet subset, and set the mixing threshold $t_m$ to 0.1 in all cases. Following semi-supervised learning, we gradually increase the weight of suspicious sample loss in stage 3, simply setting $\\alpha$ to 0.1 in the first 50 epochs and 0.5 in the last 50 epochs.", + "bbox": [ + 75, + 214, + 470, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate the performance of defenses with two commonly used metrics: the model's classification accuracy on the benign test set, Benign Accuracy (BA), and the model's classification accuracy on the poison test set, Attack Success Rate (ASR). The lower the ASR and the higher BA, the stronger the defense. In particular, the poison test set removes the samples whose ground-truth label is the target label.", + "bbox": [ + 75, + 340, + 468, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Effectiveness of Our V&B Defense", + "text_level": 1, + "bbox": [ + 76, + 469, + 377, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To measure the effectiveness of our proposed V&B method, we compare its performance with four state-of-the-art defenses on CIFAR-10 and the ImageNet subset. The backdoor attack CL-16 and CL-32 are two versions of the Clean-label attack with different magnitudes of adversarial perturbations.", + "bbox": [ + 75, + 492, + 468, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 demonstrates the defense results on CIFAR-10, where No Defense is a baseline representing the capability of backdoor attacks. Considering real scenarios, we adopt common data augmentations (i.e. random crop and horizontal flipping) during training, which will slightly affect the attack success rate. It is obvious that our V&B achieves the best results against almost all attacks, even compared to defenses that need additional benign samples. Our method reduces the success rate of all attacks from above $95\\%$ (64.79% for SIG) to blow or close to $1\\%$ , while maintaining the accuracy on benign samples with only about $1\\%$ reduction. For WaNet and CL-16, V&B achieves the second-best performance in reducing the attack success rate, but the benign accuracy of V&B is significantly superior to the corresponding best defense, i.e. DBD and ABL. WaNet adopts a warping field as the trigger and adds a noisy training mode to interfere with detection. CL-16 only poisons the samples whose ground-truth label is the target label and adds adversarial perturbations to them. So it is relatively hard to distinguish the poisoned samples they craft from the benign ones.", + "bbox": [ + 75, + 583, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The defense performances on the ImageNet subset are shown in Table 2. Compared with other defenses, our method has the best accuracy on benign samples against all four attacks, even outperforming the baseline due to our AttentionMix data augmentation and semi-supervised suppression training. Semi-supervised learning can relabel poisoned samples to their ground-truth labels, improving benign accuracy. In terms of reducing the attack success rate, our method achieves the best results against BadNets, Blend, and SIG, and the second-best result against WaNet, which is comparable to the best result (1.92% vs. 1.00%).", + "bbox": [ + 496, + 214, + 890, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also validate the effectiveness of our method under different poisoning rates in Appendix A.4. Even with a poisoning rate of $50\\%$ , our method can still reduce the attack success rate of most attacks to below $1\\%$ , while maintaining a satisfactory benign accuracy.", + "bbox": [ + 496, + 380, + 890, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Effectiveness of AttentionMix", + "text_level": 1, + "bbox": [ + 500, + 465, + 764, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To illustrate the inhibition effect of our AttentionMix, we replace it with three other popular strong data augmentations in our framework, i.e. Mixup, Cutmix and Attentive Cutmix. As Table 3 shows, each data augmentation performs well against BadNets and Blend, while for more stealthy attacks, such as WaNet and Dynamic, our AttentionMix can better reduce the attack success rate. This is because most of poisoned samples crafted by BadNets and Blend can be filtered out before training the Beneficiary network due to the simple trigger patterns. But for stealthy attacks, the Victim network will only separate part of the poisoned samples after warming up, and the other three augmentations can not effectively inhibit the missed poisoned samples. For a poisoned sample, the poisoned network will focus on its trigger location, so blending the sample's activation region with other images can effectively destroy triggers and help to clean the network. Although Attentive Cutmix can locate the trigger region, direct pasting may leave the trigger intact in the generated image.", + "bbox": [ + 496, + 489, + 890, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 786, + 663, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Component Effects. As shown in Table 4, We first train the network without any defense as a baseline and then add our components one by one to investigate their effect. The variants are as follows: $(1) + F$ : filtering the training set and training on credible samples; $(2) + F + AT$ : using AttentionMix when training with credible samples; $(3) + F +$", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "161", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c7e29e7de4ddb338693d2ab6893a156e7152990b777318c93b0c0895e0d99345.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AttackBadNets [10]Blend [7]WaNet [22]Dynamic [23]CL-16 [26]
MetricBAASRBAASRBAASRBAASRBAASR
No Defense94.36%100.00%94.57%99.76%94.21%95.84%94.66%95.18%94.98%96.93%
+F93.50%10.01%93.47%76.70%92.92%90.97%93.80%93.04%93.63%2.70%
+F + AT92.27%0.42%92.72%1.42%91.58%25.16%93.08%50.01%93.09%2.09%
+F + AT + SST93.88%1.21%94.56%1.52%93.31%1.23%94.32%28.79%94.43%3.29%
+F + AT + SST + S (suspicious)94.11%1.00%94.49%0.73%93.49%0.61%94.61%1.43%94.00%1.06%
ALL93.96%0.62%94.37%0.63%94.15%0.54%93.91%1.13%94.24%1.01%
", + "bbox": [ + 81, + 89, + 893, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study of individual components in our framework on CIFAR-10.", + "bbox": [ + 236, + 214, + 725, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AT + SST: adding standard semi-supervised training [2] on the basis of the second variant; (4) + F + AT + SST + S (suspicious): suppressing suspicious samples during semi-supervised training on the basis of the third variant.", + "bbox": [ + 75, + 253, + 468, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As can be seen from Table 4, when defending against Blend, WaNet, and Dynamic, only filtering will fail because enough poisoned samples are missed to create backdoors. After using AttentionMix, although the attack success rate of most attacks is severely decreased, the benign accuracy also drops slightly. Thus semi-supervised training is adopted to improve the performance of the network on benign samples. For WaNet and Dynamic, semi-supervised training also reduces their attack success rate by relabeling poisoned samples as ground-truth labels. However, the Beneficiary network is not completely clean, so a small number of poisoned samples may be relabeled as the target label, causing the attack success rate to rebound (e.g. Bad-Nets, Blend and Cl-16). Suppression to suspicious samples can make up for the weakness of relabeling and adding suppression to credible samples will further reduce the attack success rate. We say that the complete framework is more secure when facing backdoor attacks, although its benign accuracy may be lower than other variants.", + "bbox": [ + 75, + 314, + 468, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further demonstrate the effect of stage 3, we compare the final results with those after stage 2 in Appendix A.5. In all cases, the attack success rate after stage 2 is higher than the final result, while the benign accuracy after stage 2 is lower than the final result, suggesting that our semi-supervised suppression training can both erasing existing backdoors and improve model performance.", + "bbox": [ + 75, + 601, + 468, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Selection of Mixing Threshold. As Figure 3 shows, we set the mixing threshold $t_m$ from 0.1 to 0.5 on CIFAR-10 and ImageNet subset to study its impact on model performance. It can be seen that the benign accuracy on CIFAR-10 starts to drop when the mixing threshold is large enough (e.g. 0.2 for BadNets and 0.3 for Blend), while the attack success rate only slightly fluctuates. The decline in benign accuracy is mainly caused by the unlabeled samples in stage 3. Due to the small image size $(32 \\times 32$ in CIFAR-10), the activation region will contain only a few clustered pixels or some separate pixels if the mixing threshold is too large. During mixing, these pixels will carry their labels but", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b76cc2a09db230eca8ca7be2fa9ff413cb89c389c917bde1ffc7ef3e84627789.jpg", + "image_caption": [ + "Figure 3. Ablation study of mixing threshold $t_m$ on CIFAR-10 (left) and ImageNet subset (right)." + ], + "image_footnote": [], + "bbox": [ + 501, + 250, + 697, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4abbd67143dc90a68c66a84737fa16bb6bc3f122a93dc899c8441f3ebaa020d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 250, + 890, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "can not provide the feature of corresponding classes, which is equivalent to introducing noise to labels. It hinders the learning of our Beneficiary network, leading to more and more suspicious samples being mislabeled in stage 3. This point can be confirmed on the ImageNet subset (image size is $224 \\times 224$ ), where the mixing threshold has little impact on our method.", + "bbox": [ + 498, + 436, + 890, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 556, + 617, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we show that poisoned samples could be easily distinguished by the entropy of the poisoned network's prediction. Based on this finding and the fact that strong data augmentation can inhibit backdoor injection, we propose a novel secure training framework consisting of two parallel networks to prevent a backdoor generation from three aspects: (1) filtering training data; (2) inhibiting backdoor injection; (3) erasing potential backdoors. Specifically, we sacrifice one network to exclusively learn backdoor triggers and filter credible samples according to the prediction entropy to train another network. During training, we design a strong data augmentation to inhibit missed poisoned samples from working and then introduced a semi-supervised learning method to erase potential backdoors with the help of the sacrifice network. Extensive experiments verify that our framework is effective against various backdoor attacks under different poisoning rates.", + "bbox": [ + 496, + 580, + 890, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work is supported in part by the National Natural Science Foundation of China Under Grants No.62176253 and No.U20B2066.", + "bbox": [ + 498, + 849, + 890, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "162", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A New Backdoor Attack in CNNS by Training Set Corruption Without Label Poisoning. In 2019 IEEE International Conference on Image Processing, pages 101-105. IEEE, 2019.", + "[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. MixMatch: A Holistic Approach to Semi-Supervised Learning. Advances in neural information processing systems, 32, 2019.", + "[3] Eitan Borgnia, Valerii Cherepanova, Liam Fowl, Amin Ghiasi, Jonas Geiping, Micah Goldblum, Tom Goldstein, and Arjun Gupta. Strong Data Augmentation Sanitizes Poisoning and Backdoor Attacks Without an Accuracy Tradeoff. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 3855-3859. IEEE, 2021.", + "[4] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian Molloy, and Biplav Srivastava. Detecting Backdoor Attacks on Deep Neural Networks by Activation Clustering. arXiv preprint arXiv:1811.03728, 2018.", + "[5] Huili Chen, Cheng Fu, Jishen Zhao, and Farinaz Koushanfar. ProFlip: Targeted Trojan Attack With Progressive Bit Flips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7718-7727, 2021.", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020.", + "[7] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted Backdoor Attacks on Deep Learning Systems Using Data Poisoning. arXiv preprint arXiv:1712.05526, 2017.", + "[8] Yanbei Chen, Massimiliano Mancini, Xiatian Zhu, and Zeynep Akata. Semi-Supervised and Unsupervised Deep Visual Learning: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.", + "[9] Bao Gia Doan, Ehsan Abbasnejad, and Damith C Ranasinghe. Februus: Input Purification Defense Against Trojan Attacks on Deep Neural Network Systems. In Annual Computer Security Applications Conference, pages 897-912, 2020.", + "[10] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain. arXiv preprint arXiv:1708.06733, 2017.", + "[11] Chuan Guo, Ruihan Wu, and Kilian Q Weinberger. Trojannet: Embedding Hidden Trojan Horse Models in Neural Networks. arXiv preprint arXiv:2002.10078, 2, 2020.", + "[12] Jonathan Hayase, Weihao Kong, Raghav Somani, and Sewoong Oh. SPECTRE: Defending Against Backdoor Attacks Using Robust Statistics. arXiv preprint arXiv:2104.11315, 2021.", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Kunzhe Huang, Yiming Li, Baoyuan Wu, Zhan Qin, and Kui Ren. Backdoor Defense via Decoupling the Training Process. In The Tenth International Conference on Learning Representations, 2022.", + "[15] Junnan Li, Richard Socher, and Steven CH Hoi. DIVIDEMIX: LEARNING WITH NOISY LABELS AS SEMI-SUPERVISED LEARNING. arXiv preprint arXiv:2002.07394, 2020.", + "[16] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible Backdoor Attack With Sample-Specific Triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021.", + "[17] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-Backdoor Learning: Training Clean Models on Poisoned Data. Advances in Neural Information Processing Systems, 34:14900–14912, 2021.", + "[18] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks. In International Conference on Learning Representations, 2021.", + "[19] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273–294. Springer, 2018.", + "[20] Xuankai Liu, Fengting Li, Bihan Wen, and Qi Li. Removing Backdoor-Based Watermarks in Neural Networks with Limited Data. In International Conference on Pattern Recognition, pages 10149-10156. IEEE, 2021.", + "[21] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection Backdoor: A Natural Backdoor Attack on Deep Neural Networks. In European Conference on Computer Vision, pages 182-199. Springer, 2020.", + "[22] Anh Nguyen and Anh Tran. WaNet-M Imperceptible Warping-based Backdoor Attack. arXiv preprint arXiv:2102.10369, 2021.", + "[23] Tuan Anh Nguyen and Anh Tran. Input-Aware Dynamic Backdoor Attack. Advances in Neural Information Processing Systems, 33:3454-3464, 2020.", + "[24] Xiangyu Qi, Tinghao Xie, Ruizhe Pan, Jifeng Zhu, Yong Yang, and Kai Bu. Towards Practical Deployment-Stage Backdoor Attack on Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13347-13357, 2022.", + "[25] Han Qiu, Yi Zeng, Shangwei Guo, Tianwei Zhang, Meikang Qiu, and Bhavani Thuraisingham. DeepSweep: An Evaluation Framework for Mitigating DNN Backdoor Attacks using Data Augmentation. In Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security, pages 363-377, 2021.", + "[26] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-Consistent Backdoor Attacks. arXiv preprint arXiv:1912.02771, 2019.", + "[27] Devesh Walawalkar, Zhiqiang Shen, Zechun Liu, and Marios Savvides. Attentive CutMix: An Enhanced Data Augmentation Approach for Deep Learning Based Image Classification. arXiv preprint arXiv:2003.13048, 2020." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "163", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural Cleans: Identifying and Mitigating Backdoor Attacks in Neural Networks. In 2019 IEEE Symposium on Security and Privacy, pages 707-723. IEEE, 2019.", + "[29] Jason Weston, Frédéric Ratle, and Ronan Collobert. Deep Learning via Semi-Supervised Embedding. In Proceedings of the 25th international conference on Machine learning, pages 1168-1175, 2008.", + "[30] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6023-6032, 2019.", + "[31] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. Mixup: Beyond Empirical Risk Minimization. In 6th International Conference on Learning Representations, 2018.", + "[32] Pu Zhao, Pin-Yu Chen, Payel Das, Karthikeyan Natesan Ramamurthy, and Xue Lin. Bridging Mode Connectivity in Loss Landscapes and Adversarial Robustness. 2020.", + "[33] Zhendong Zhao, Xiaojun Chen, Yuexin Xuan, Ye Dong, Dakui Wang, and Kaitai Liang. DEFEAT: Deep Hidden Feature Backdoor Attacks by Imperceptible Perturbation and Latent Representation Constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15213-15222, 2022.", + "[34] Nan Zhong, Zhenxing Qian, and Xinpeng Zhang. Imperceptible Backdoor Attack: From Input Space to Feature Representation. arXiv preprint arXiv:2205.03190, 2022.", + "[35] Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Scholkopf. Learning with Local and Global Consistency. Advances in neural information processing systems, 16, 2003." + ], + "bbox": [ + 78, + 90, + 468, + 571 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "164", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/fdfec7d9-d3ec-4468-9f42-c0ec2550345d_model.json b/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/fdfec7d9-d3ec-4468-9f42-c0ec2550345d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5583f64ca59ec1ada12c7ce1877dcc280b02db89 --- /dev/null +++ b/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/fdfec7d9-d3ec-4468-9f42-c0ec2550345d_model.json @@ -0,0 +1,1837 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.13, + 0.885, + 0.175 + ], + "angle": 0, + "content": "The Victim and The Beneficiary: Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.204, + 0.826, + 0.275 + ], + "angle": 0, + "content": "Zixuan Zhu, Rui Wang*, Cong Zou, Lihua Jing \nSKLOIS, Institute of Information Engineering, CAS, Beijing, China \nSchool of Cyber Security, University of Chinese Academy of Sciences, Beijing, China \n{zhuzixuan, wangrui, zoucong, jinglihua}@iei.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.341, + 0.473, + 0.765 + ], + "angle": 0, + "content": "Recently, backdoor attacks have posed a serious security threat to the training process of deep neural networks (DNNs). The attacked model behaves normally on benign samples but outputs a specific result when the trigger is present. However, compared with the rocketing progress of backdoor attacks, existing defenses are difficult to deal with these threats effectively or require benign samples to work, which may be unavailable in real scenarios. In this paper, we find that the poisoned samples and benign samples can be distinguished with prediction entropy. This inspires us to propose a novel dual-network training framework: The Victim and The Beneficiary (V&B), which exploits a poisoned model to train a clean model without extra benign samples. Firstly, we sacrifice the Victim network to be a powerful poisoned sample detector by training on suspicious samples. Secondly, we train the Beneficiary network on the credible samples selected by the Victim to inhibit backdoor injection. Thirdly, a semi-supervised suppression strategy is adopted for erasing potential backdoors and improving model performance. Furthermore, to better inhibit missed poisoned samples, we propose a strong data augmentation method, AttentionMix, which works well with our proposed V&B framework. Extensive experiments on two widely used datasets against 6 state-of-the-art attacks demonstrate that our framework is effective in preventing backdoor injection and robust to various attacks while maintaining the performance on benign samples. Our code is available at https://github.com/Zixuan-Zhu/VaB." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.778, + 0.21, + 0.794 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.803, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Large amounts of training data and powerful computing resources are two key factors for the success of deep neural networks (DNNs). While in practical applications, obtaining enough samples for training is laborious, and training is likely to be resource-constrained. As a result, more and" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.308, + 0.853, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.501, + 0.895, + 0.571 + ], + "angle": 0, + "content": "Figure 1. The average prediction entropy of benign samples versus poisoned samples crafted by 6 backdoor attacks. We conduct the experiment on CIFAR-10 with ResNet-18 under poisoning rate \\(10\\%\\), where our special training strategy (described in section 3.3) was adopted." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.581, + 0.893, + 0.627 + ], + "angle": 0, + "content": "more third-party data (e.g. data released on the Internet) and training platforms (e.g. Google Colab) are adopted to reduce the overhead, which brings new security risks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Backdoor attacks [5, 11, 24] pose a serious security threat to the training process of DNNs and can be easily executed through data poisoning. Specifically, attackers [7, 10, 22] inject the designed trigger (e.g. a small patch or random noise) to a few benign samples selected from the training set and change their labels to the attacker-defined target label. These poisoned samples will force models to learn the correlation between the trigger and the target label. In the reference process, the attacked model will behave normally on benign samples but output the target label when the trigger is present. Deploying attacked models can lead to severe consequences, even life-threatening in some scenarios (e.g. autonomous driving). Hence, a secure training framework is needed when using third-party data or training platforms." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Many researchers have devoted themselves to this topic and proposed feasible defenses. Li et al. [18] utilize local benign samples to erase existing backdoors in DNNs, and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "155" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.347 + ], + "angle": 0, + "content": "Borgnia et al. [3] employ strong data augmentations to inhibit backdoor injection during training. Li et al. [17] find that the training loss decreased faster for poisoned samples than benign samples and designed a loss function to separate them. Finally, they select a fixed portion of samples with the lowest loss to erase potential backdoors. In real scenarios, it is challenging to distinguish poisoned samples from benign ones in this way due to the unknown poisoning rate and close loss values. In experiments, we find the entropy of model prediction is a more discriminative property, and a fixed threshold could filter out most poisoned samples, as shown in Figure 1. During training, poisoned samples will be learned faster than benign samples due to the similar feature of their triggers [17]. Hence, the poisoned network can confidently predict poisoned samples as the target label in early epochs but hesitates about the benign samples." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.351, + 0.468, + 0.666 + ], + "angle": 0, + "content": "Based on this observation, we propose a novel secure training framework, The Victim and The Beneficiary (V&B), that can directly train clean models on poisoned data without resorting to benign samples. Firstly, in warming up stage, the training dataset is recomposed into a suspicious set and a credible set according to the prediction entropy is lower or higher than the predefined threshold. Then the Victim network is trained iteratively with the suspicious set to obtain a strong poisoned sample detector. Secondly, in the clean training stage, the Beneficiary network is trained with credible samples filtered by the Victim to inhibit backdoor injection. In later training epochs, the Victim network is fully exploited to provide its learned knowledge for the Beneficiary network to help the latter erase potential backdoors through semi-supervised learning. To diminish the threat of missed poisoned samples, we design a strong data augmentation AttentionMix, which mixes the image region with high activation values according to the attention map. Compared with existing augmentations [27, 30, 31], it has a stronger inhibition effect against stealthy backdoor attacks [23, 22, 26]." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.668, + 0.468, + 0.802 + ], + "angle": 0, + "content": "The main contributions of this paper are as follows: (1) We propose a novel dual-network secure training framework, V&B, which can train clean models on the poisoned dataset without resorting to benign samples. (2) We design a powerful data augmentation method called AttentionMix, which has a stronger inhibition effect against stealthy backdoor attacks. (3) Extensive experiments on two popular benchmark datasets demonstrate the effectiveness and robustness of our framework." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.821, + 0.216, + 0.835 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.847, + 0.262, + 0.862 + ], + "angle": 0, + "content": "2.1. Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.872, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Data Augmentation is a common-used technique to improve the generalization capabilities of DNNs. In addition" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.89, + 0.408 + ], + "angle": 0, + "content": "to some popular weak data augmentations (such as cropping, flipping, rotation, etc.), many strong data augmentation techniques [27, 30, 31] have been proposed to further improve model performance. Mixup [31] linearly interpolates any two images and their labels, then trains the model with generated samples. However, the generated samples tend to be unnatural, which will decrease model performances on localization and object detection tasks. Instead of linear interpolation, Cutmix [30] replaces a random patch with the same region from another image and mixes their labels according to the ratio of the patch area. Further, Attentive Cutmix [27] only pastes the influential region located by an attention mechanism to other images, effectively forcing models to learn discriminative details. Yet it also mixes labels based on the ratio of the region area, ignoring the importance of the region in two original images, and direct pasting may bring complete triggers into benign samples. Our AttentionMix takes into account the importance of the region in both images and blends the influential region with the same area in another image, which can destroy the completeness of triggers." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.422, + 0.764, + 0.436 + ], + "angle": 0, + "content": "2.2. Backdoor Attack and Defense" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.447, + 0.89, + 0.49 + ], + "angle": 0, + "content": "In this paper, we only consider poisoning-based attacks toward image classification that can occur during the data preparation stage." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.506, + 0.89, + 0.776 + ], + "angle": 0, + "content": "Poisoned-based backdoor attack. In early studies, attackers usually adopt simple and visible patterns as the trigger, such as a black-white checkerboard [10]. To escape human inspections, some works use human-imperceptible deformations [22] as the trigger or design unique triggers [16, 23] for each poisoned sample. However, they randomly select benign samples to inject triggers and change their labels (called poison-label attacks), possibly causing the content of an image not to match its label. Instead of randomly selecting samples to poison, Turner et al. [26], Mauro et al. [1], and Liu et al. [21] only inject triggers into benign samples under the target label (called clean-label attacks). Although some triggers are imperceptible in the input space, they can be easily detected by defenses focused on latent space extracted by CNNs. Therefore, recent studies [33, 34] not only ensure the trigger's invisibility in the input space but also restrict the similarity between poison samples and corresponding benign samples in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.796, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Backdoor Defense. For poisoning-based attacks, defense methods could be divided into three categories from their working stage, including (1) preprocessing before training [9, 12, 25], (2) poison suppression during training [14, 17, 20], and (3) backdoor erasing after training [18, 28, 32]. For preprocessing, Hayase et al. [12] and Chen et al. [4] filter the training set through statistics or clustering and" + }, + { + "type": "page_number", + "bbox": [ + 0.489, + 0.947, + 0.511, + 0.956 + ], + "angle": 0, + "content": "156" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "train the model with benign samples. For poison suppressing, Borgnia et al. [3] leverage strong data augmentation (e.g. Mixup, Cutmix) to alleviate poisoned samples' adverse effects. Huang et al. [14] design a decoupling training process, which first learns a feature extractor via SimCLR [6] to prevent poisoned samples from clustering in feature space. For backdoor erasing, Liu et al. [19] prune the low-activation neurons for validation samples (benign samples) and then fine-tune the model with these benign samples. While Li et al. [18] first fine-tune the poisoned model with a small portion of benign samples to obtain a teacher model, then erase the original model's backdoor through knowledge distillation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.298, + 0.317, + 0.315 + ], + "angle": 0, + "content": "2.3. Semi-supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.473, + 0.475 + ], + "angle": 0, + "content": "Semi-supervised learning [8] aims to alleviate the dependence of deep learning on extensive labeled training data and jointly learns the model from a limited amount of labeled samples and a large number of unlabeled samples. Existing methods [2, 29, 35] generally assign each unlabeled sample with a pseudo label and design different loss functions for labeled and unlabeled samples. The weight of unlabeled sample loss will gradually increase during training to reduce the negative effect of low-quality pseudo labels in early epochs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.488, + 0.17, + 0.503 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.218, + 0.529 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.473, + 0.766 + ], + "angle": 0, + "content": "Pipeline of Poisoned-based Backdoor Attacks. Let \\(\\mathcal{D}_B = \\{(x_i, y_i)\\}_{i=1}^N\\) denotes the benign training set containing \\(N\\) samples, where \\(x_i \\in \\{0, 1, \\dots, 255\\}^{C \\times H \\times W}\\) is the image, \\(y_i \\in \\{0, 1, \\dots, K\\}\\) is its label, \\(K\\) is the number of classes, \\(H, W\\) and \\(C\\) are the height, width and channels of an image. The attacker will select a subset of benign samples to poison with a designed injection function \\(G(\\cdot)\\). Let \\(\\mathcal{D}_G\\) indicates the generated poisoned samples set and \\(\\mathcal{D}_R\\) denotes the remaining benign samples set, where \\(\\mathcal{D}_G = \\{(x', y_t) | x' = G(x), (x, y) \\in \\mathcal{D}_B \\setminus \\mathcal{D}_R\\}\\), \\(y_t\\) is the target label and \\(r = \\frac{\\left|\\mathcal{D}_G\\right|}{\\mathcal{D}_B}\\) is the poisoning rate. The poisoned dataset \\(\\mathcal{D}_P = \\mathcal{D}_G \\cup \\mathcal{D}_R\\) will be unknowingly adopted by victim users to train DNNs. At inference time, the attacker can manipulate the poisoned model's output through the injected backdoor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Threat Model and Defense Setting. In this paper, we focus on defending against poisoned-based backdoor attacks toward image classification. The attacker can arbitrarily modify the training set, but can not change other training components (e.g. model structure and hyperparameters) [1, 10, 22]. Such attacks may occur when using third-party datasets or training platforms that can poison users' clean data. We follow the common assumption [12, 14, 17] that" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "defenders have no information about the backdoor (including poisoning rate, trigger pattern, target label, etc.) and any local benign samples to use but have full control over the training process. The defense goal is to prevent the trained DNNs from predicting poisoned samples as the target label while maintaining its accuracy on benign samples." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.193, + 0.611, + 0.208 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.519 + ], + "angle": 0, + "content": "In this subsection, we briefly introduce the pipeline of our framework. As Figure 2(a) illustrates, our framework consists of two networks, and the whole training process is divided into three stages — warming up, clean training and semi-supervised suppression training. In warming up, we aim to obtain a poisoned network that can roughly distinguish poisoned samples from benign ones according to the prediction entropy. Therefore we train the Victim network with only filtered suspicious samples (i.e. suspicious set) except for the first epoch. This training strategy remains unchanged in the following two stages. During clean training, we adopt credible samples (i.e. credible set) to train the Beneficiary network with our proposed AttentionMix data augmentation. The strong data augmentation can effectively prevent missed poisoned samples from creating backdoors. As the Victim network is trained, the entropy of its prediction for benign samples will also slowly decrease below the filtering threshold, leading to a reduction in the credible set size. So one credible set will be used for a few epochs to ensure the Beneficiary network gets enough training samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.762 + ], + "angle": 0, + "content": "To further improve the Beneficiary network performance and erase its potential backdoor, we regard the credible set as a labeled dataset and the suspicious set as an unlabeled dataset to train the Beneficiary network via semi-supervised learning. Following conventional practice [2, 15], each unlabeled sample will be assigned a pseudo label generated by the Beneficiary network, forming the relabeled set. However, if the Beneficiary network still contains a backdoor after stage 2, poisoned samples in the suspicious set will be relabeled as the target label, strengthening the backdoor. After specific training, the Victim network is sensitive to trigger patterns and thus can be utilized to suppress the component corresponding to the target class after relabeling. Furthermore, we also apply this suppression operation to the credible set to further weaken the impact of missed poisoned samples." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.638, + 0.788 + ], + "angle": 0, + "content": "3.3. Warming Up" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.87 + ], + "angle": 0, + "content": "At this stage, we aim to obtain a poisoned network that has only learned trigger patterns. In this way, the entropy of its prediction for poisoned samples will be significantly lower than that of benign samples. Then we can use a threshold to filter out poisoned samples." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Let \\(\\mathcal{D}\\) denotes the training set and \\(F_{v}:\\{0,1,\\dots,255\\}^{C\\times H\\times W}\\to [0,1]^{K}\\) indicates the Vic" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "157" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.293, + 0.586, + 0.305 + ], + "angle": 0, + "content": "(a) The pipeline of our training framework" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.307, + 0.482, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.307, + 0.89, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.472, + 0.895, + 0.584 + ], + "angle": 0, + "content": "Figure 2. (a) The pipeline of our training framework. Net V is the Victim network (i.e. poisoned network) that only learns from the filtered suspicious set, and Net B is the Beneficiary network (i.e. clean network) we wanted. Red triangles represent poisoned samples, and green circles are benign samples. The color of nets indicates their degree of poisoning: the darker the color, the deeper the poisoning, and vice versa. Deep blue areas are three training stages, and light blue areas are basic blocks that will be executed many times in the corresponding stage. (b) An example of AttentionMix. The bottom four images are the non-activation region of the cat image, the activation region of the cat image, the same activation region in the dog image, and the final generated image, respectively. And their labels are located below the images. (c) Suppression for poisoned samples and benign samples. The target label is assumed as the first label. Suppression can reduce the target label component of poisoned samples but have little influence on the ground-truth label of benign samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.471, + 0.675 + ], + "angle": 0, + "content": "tim network. We first use the whole training set to train the Victim network for one epoch with cross-entropy loss. Then the prediction entropy of a sample will be calculated with Equation 1, where \\( p \\) is the prediction of the Victim network for that sample." + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.681, + 0.47, + 0.723 + ], + "angle": 0, + "content": "\\[\ne = - \\sum_ {k = 1} ^ {K} p _ {k} \\log p _ {k} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.73, + 0.471, + 0.851 + ], + "angle": 0, + "content": "Once all samples' prediction entropy is obtained, we normalize them to \\([0,1]\\) through mix-max normalization. Then a threshold \\(t_f\\) is set to filter out suspicious samples \\(\\mathcal{D}_s = \\{(x_i,y_i)|(x_i,y_i)\\in \\mathcal{D},e_i < t_f\\}\\) and the remaining samples are regarded as credible samples \\(\\mathcal{D}_c\\). To strengthen the learning of poisoned samples and delay the learning of benign samples, the Victim network will only be trained on suspicious sets in the future. The loss function is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.865, + 0.469, + 0.883 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v} = \\mathbb {E} _ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {s}} \\left[ \\mathcal {L} _ {C E} \\left(F _ {v} (\\boldsymbol {x}), y\\right) \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.887, + 0.363, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{CE}\\) denotes the cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.795 + ], + "angle": 0, + "content": "After the first epoch, the prediction entropy of some poisoned samples may still be larger than that of most benign ones, although there is a large gap between their average values in Figure 1. If the filtering threshold \\( t_f \\) is set too large, lots of benign samples will be discarded, which will affect the model performance on benign samples. So we set a smaller \\( t_f \\) and wait for the prediction entropy of these poisoned samples to decrease. As training progresses, their prediction entropy will drop below \\( t_f \\), making the credible set cleaner. Simultaneously, the prediction entropy of benign samples will also decrease, resulting in a smaller credible set. We expect the Victim network to filter out a clean and relatively large credible set after this stage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.655, + 0.818 + ], + "angle": 0, + "content": "3.4. Clean Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "At this stage, we begin to train the Beneficiary network with filtered credible samples \\(\\mathcal{D}_c\\), where the filtering strategy is the same as in stage 1. Because of the rough separation, a few poisoned samples may be included in the credible set. For strong attacks, a small number of poisoned sam" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "158" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "ples is enough to create a backdoor. So we design a strong data augmentation called AttentionMix to inhibit the missed poisoned samples from working. Compared with existing augmentations [27, 30, 31], AttentionMix has a stronger inhibition effect on stealthy backdoor attacks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.243 + ], + "angle": 0, + "content": "The core idea of AttentionMix is to generate a new training sample \\((\\tilde{\\boldsymbol{x}},\\tilde{\\boldsymbol{y}})\\) by mixing two training samples \\((x_{1},y_{1})\\) and \\((x_{2},y_{2})\\) according to their attention maps, where \\(\\tilde{\\boldsymbol{y}}\\) is a K-dimensional vector. Following [18], we adopt one sample's final feature maps to compute its attention map:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.253, + 0.47, + 0.29 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} ^ {\\prime} = \\sum_ {i = 1} ^ {c} \\left| F _ {i} \\right| ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.469, + 0.391 + ], + "angle": 0, + "content": "where \\(\\pmb{F} \\in \\mathbb{R}^{c \\times h \\times w}\\) indicates feature maps, \\(\\pmb{A}' \\in \\mathbb{R}^{h \\times w}\\) is the corresponding attention map, \\(c, h\\) and \\(w\\) are the dimensions of channel, height and width. For each attention map, we first bilinearly interpolate it to the original image size \\(H \\times W\\), and then normalize its elements to \\([0,1]\\) by min-max normalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.469, + 0.422 + ], + "angle": 0, + "content": "After getting the attention map, we set a threshold \\( t_m \\) to obtain the activation region that we want to mix:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.432, + 0.469, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} _ {i, j} = \\left\\{ \\begin{array}{l l} 0 & \\boldsymbol {A} _ {i, j} < t _ {m} \\\\ 1 & \\boldsymbol {A} _ {i, j} \\geq t _ {m} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.469, + 0.519 + ], + "angle": 0, + "content": "where \\(M \\in \\mathbb{R}^{H \\times W}\\) is the binary mask indicating which pixels belong to the activation region and \\(\\pmb{A}\\) is the attention map after interpolation and normalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.469, + 0.625 + ], + "angle": 0, + "content": "Let \\( A^1 \\) and \\( A^2 \\) denote the attention maps of \\( x_1 \\) and \\( x_2 \\), \\( M^1 \\) denotes the first sample's mask. As Figure 2(b) shows, the mixed image \\( \\tilde{x} \\) is composed of three parts: the non-activation region of \\( x_1 \\), the activation region of \\( x_1 \\) and the same area in \\( x_2 \\). During mixing, the non-activation region of \\( x_1 \\) remains unchanged, and the activation region is blended with the same area of \\( x_2 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.634, + 0.469, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\boldsymbol {x}} = \\left(\\mathbf {1} - \\boldsymbol {M} ^ {1}\\right) \\odot \\boldsymbol {x} _ {1} + \\lambda_ {2} \\cdot \\left(\\boldsymbol {M} ^ {1} \\odot \\boldsymbol {x} _ {1}\\right) \\tag {5} \\\\ + \\lambda_ {3} \\cdot \\left(\\boldsymbol {M} ^ {1} \\odot \\boldsymbol {x} _ {2}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.469, + 0.726 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the Hadamard Product, \\(\\lambda_{2}\\) and \\(\\lambda_{3}\\) are the mixing weights calculated from the two attention maps. The calculation process is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.736, + 0.469, + 0.772 + ], + "angle": 0, + "content": "\\[\nw _ {1} = \\frac {\\operatorname {s u m} \\left(\\boldsymbol {M} ^ {1}\\right)}{\\operatorname {s u m} \\left(\\boldsymbol {A} ^ {1}\\right)}, w _ {2} = \\frac {\\operatorname {s u m} \\left(\\boldsymbol {M} ^ {1} \\odot \\boldsymbol {A} ^ {2}\\right)}{\\operatorname {s u m} \\left(\\boldsymbol {A} ^ {2}\\right)} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.782, + 0.469, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {1} = 1 - w _ {1}, \\lambda_ {2} = \\frac {w _ {1}}{w _ {1} + w _ {2}}, \\lambda_ {3} = \\frac {w _ {2}}{w _ {1} + w _ {2}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.815, + 0.469, + 0.875 + ], + "angle": 0, + "content": "where \\(\\mathrm{sum}(\\cdot)\\) is a summation function, \\(w_{1}\\) and \\(w_{2}\\) denote the importance of the activation region to the two images, and \\(\\lambda_{1}\\) is the importance of the non-activation region to \\(x_{1}\\). The final label is also composed of three parts:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.886, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {y}} = \\lambda_ {1} \\cdot \\boldsymbol {y} _ {1} + w _ {1} \\cdot \\left(\\lambda_ {2} \\cdot \\boldsymbol {y} _ {1} + \\lambda_ {3} \\cdot \\boldsymbol {y} _ {2}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "where \\( \\pmb{y}_1 \\) and \\( \\pmb{y}_2 \\) are the one-hot labels of the two samples. When the activation region of \\( \\pmb{x}_1 \\) corresponds to the background of \\( \\pmb{x}_2 \\), \\( w_1 \\) will be significantly larger than \\( w_2 \\), and \\( \\lambda_3 \\) will be a small value, which can reduce background interference during mixing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.167, + 0.892, + 0.213 + ], + "angle": 0, + "content": "We denote the generated new credible set as \\(\\tilde{\\mathcal{D}}_c\\) and train the Beneficiary network \\(F_{b}\\) with the following loss function (i.e. Train with AttentionMix in Figure 2(a)):" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.225, + 0.892, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b} = \\mathbb {E} _ {(\\tilde {\\boldsymbol {x}}, \\tilde {\\boldsymbol {y}}) \\in \\tilde {\\mathcal {D}} _ {c}} [ \\mathcal {L} _ {C E} (F _ {b} (\\tilde {\\boldsymbol {x}}), \\tilde {\\boldsymbol {y}}) ] \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.253, + 0.892, + 0.344 + ], + "angle": 0, + "content": "For a poisoned network, the activation region of poisoned samples is more likely their trigger locations. Mixing the activation region can destroy the completeness of trigger patterns. While for benign samples, the mixing plays a regularization role that prevents the network from overfitting a particular subject and forces it to learn abundant features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.354, + 0.831, + 0.37 + ], + "angle": 0, + "content": "3.5. Semi-supervised Suppression Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.527 + ], + "angle": 0, + "content": "Until now, we have got a poisoned Victim network that can filter out most of the poisoned samples and a Beneficiary network that is relatively clean but may still contain the backdoor. In the previous stage, the Beneficiary network is trained only with partial training samples (i.e. the credible set), which may harm its performance on benign samples. Thus we intend to enable the deprecated suspicious samples via semi-supervised learning at this stage to improve the Beneficiary network while erasing its potential backdoor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.664 + ], + "angle": 0, + "content": "Given a credible set \\(\\mathcal{D}_c\\) and a suspicious set \\(\\mathcal{D}_s\\), we use samples in \\(\\mathcal{D}_c\\) as labeled samples and remove the labels of samples in \\(\\mathcal{D}_s\\) to transform the problem into a semi-supervised learning setting. For each unlabeled sample, we average the predictions of the Beneficiary network over its multiple augmentations (e.g. random crop and random flip) and select the class with max probability as its pseudo-label, which is similar to [2] and [15]. The relabeled samples are denoted as \\(\\mathcal{D}_r = \\{(\\pmb{x}_i,\\hat{\\pmb{y}}_i)|(\\pmb{x}_i,\\pmb{y}_i)\\in \\mathcal{D}_s\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.77 + ], + "angle": 0, + "content": "However, the poisoned samples in \\(\\mathcal{D}_s\\) may be relabeled as the target label, i.e. \\(\\hat{y}_i = y_t\\), if the Beneficiary network still contains the backdoor. We can mitigate the impact of such mislabeling by exploiting the knowledge learned by the Victim network. Specifically, we adopt the Victim network's prediction to suppress the pseudo-label (i.e. Suppress in Figure 2(a)):" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.782, + 0.891, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {i} ^ {\\prime} = \\operatorname {c l i p} \\left(\\hat {\\boldsymbol {y}} _ {i} - \\operatorname {s o f t m a x} \\left(F _ {v} \\left(\\boldsymbol {x} _ {i}\\right)\\right), 0, 1\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.891, + 0.87 + ], + "angle": 0, + "content": "where \\((\\pmb{x}_i,\\hat{y}_i)\\in \\mathcal{D}_r\\) \\(\\hat{\\pmb{y}}_i\\) is the one-hot vector of \\(\\hat{y}_i\\) and \\(\\mathrm{clip}(\\cdot ,0,1)\\) is a function that will restrict its inputs to [0, 1] (i.e. inputs smaller than 0 became 0, inputs larger than 1 became 1, and the rest remain unchanged)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We use \\(\\mathcal{D}_r' = \\{(\\pmb{x}_i, \\pmb{y}_i') | (\\pmb{x}_i, \\hat{\\pmb{y}}_i) \\in \\mathcal{D}_r\\}\\) to indicate the suspicious samples after relabeling and suppression. We" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "159" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.203 + ], + "angle": 0, + "content": "
AttackBadNets [10]Blend [7]Dynamic [23]WaNet [22]SIG [1]CL-16 [26]CL-32 [26]
MetricBAASRBAASRBAASRBAASRBAASRBAASRBAASR
No Defense94.36%100.00%94.57%99.76%94.66%95.18%94.21%95.84%95.09%64.79%94.98%96.93%95.03%94.73%
FP [19]92.72%1.92%92.49%10.61%92.47%11.73%93.27%0.90%93.55%37.98%92.92%9.74%93.03%6.39%
NAD [18]88.83%1.83%87.72%1.38%88.30%2.32%91.00%0.99%90.42%3.52%88.92%1.64%90.19%2.60%
ABL [17]90.47%0.70%91.31%31.49%83.07%94.21%80.78%19.93%88.61%1.37%90.29%0.54%86.50%2.68%
DBD [14]92.00%3.06%91.62%3.78%90.98%17.56%89.13%0.11%89.42%1.61%92.85%91.08%93.24%14.88%
V&B (Ours)93.96%0.62%94.37%0.63%93.91%1.13%94.15%0.54%94.08%0.17%94.24%1.01%93.98%0.64%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.212, + 0.893, + 0.255 + ], + "angle": 0, + "content": "Table 1. The performance of 5 defense methods against 7 backdoor attacks on CIFAR-10. The first 4 attacks are poison-label attacks and the last 3 attacks are clean-label attacks. Note that FP and NAD require local benign samples to work, while ABL, DBD and our method do not. The best results are marked in boldface and underlined items are the second-best results." + }, + { + "type": "table", + "bbox": [ + 0.244, + 0.264, + 0.728, + 0.372 + ], + "angle": 0, + "content": "
AttackBadNets [10]Blend [7]WaNet [22]SIG [1]
MetricBAASRBAASRBAASRBAASR
No Defense94.94%99.51%95.16%100.00%92.63%100.00%90.38%89.22%
FP [19]86.28%1.05%87.63%2.03%82.82%26.19%82.50%22.76%
NAD [18]85.83%3.82%86.73%2.90%83.93%3.88%80.71%3.88%
ABL [17]92.15%13.78%87.34%56.57%91.38%56.71%80.13%0.87%
DBD [14]93.08%0.31%93.81%0.87%92.06%1.00%84.47%98.47%
V&B (Ours)95.42%0.28%95.03%0.45%94.84%1.92%94.65%0.03%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.2, + 0.381, + 0.768, + 0.396 + ], + "angle": 0, + "content": "Table 2. The performance of 5 defense methods against 4 backdoor attacks on ImageNet subset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.471, + 0.512 + ], + "angle": 0, + "content": "also suppress the labels of credible samples in the same way to weaken the influence of poisoned samples that may be missed after filtering. As in stage 2, the suppressed credible samples are augmented with our AttentionMix, forming the training set \\(\\hat{\\mathcal{D}}_c'\\) to further inhibit backdoor injection. We fine-tune the Beneficiary network with the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.516, + 0.47, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathbb {E} _ {(\\tilde {\\boldsymbol {x}}, \\tilde {\\boldsymbol {y}} ^ {\\prime}) \\in \\tilde {\\mathcal {D}} _ {c} ^ {\\prime}} [ \\mathcal {L} _ {C E} (F _ {b} (\\tilde {\\boldsymbol {x}}), \\tilde {\\boldsymbol {y}} ^ {\\prime}) ] + \\tag {11} \\\\ \\alpha \\cdot \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y} ^ {\\prime}) \\in \\mathcal {D} _ {r} ^ {\\prime}} [ \\mathcal {L} _ {C E} (F _ {b} (\\boldsymbol {x}), \\boldsymbol {y} ^ {\\prime}) ] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.683 + ], + "angle": 0, + "content": "Since the Victim network is only trained with suspicious samples, it will confidently predict poisoned samples as the target label but give a wrong prediction for benign samples with high probability. Thus for a poisoned sample, the suppression can reduce the component corresponding to the target class in its label while having little influence on the component of a sample's ground-truth class, as shown in Figure 2(c)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.21, + 0.712 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.72, + 0.285, + 0.736 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.47, + 0.849 + ], + "angle": 0, + "content": "Datasets and Models. We employ ResNet-18 [13] as our Victim and Beneficiary networks, a standard setting in previous studies [14]. For a fair comparison, we evaluate all defenses on two popular benchmark datasets, including CIFAR-10 and a subset of ImageNet containing 12 classes generated by Li et al. [17]. The details are summarized in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Attack Configures. We consider 6 representative backdoor attacks in our experiments, including four poison-label attacks: BadNets [10], Blend attack [7], WaNet attack [22]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.573 + ], + "angle": 0, + "content": "Dynamic attack [23], and two clean-label attacks: Sinusoidal signal attack (SIG) [1] and Clean-label attack (CL) [26]. To achieve the best attack performance, we configure these attacks as close as possible to their released codes and original settings (e.g. trigger pattern and trigger size). While on the ImageNet subset, we implement these attacks based on the settings suggested by [14], since some attacks miss the corresponding experiments in their original papers. We uniformly set the poisoning ratio as 0.1 and select the first class as the target class in both datasets, i.e. \\( y_{t} = 0 \\). More details about attacks are shown in Appendix A.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Defense Configures and Training Details. We compare our V&B with 4 state-of-the-art defenses: Fine-pruning (FP) [19], Neural Attention Distillation (NAD) [18], Anti-Backdoor Learning (ABL) [17] and Decoupling (DBD) [14], in which FP and NAD require additional local benign samples to work. Similarly, we follow their original settings and released codes. For V&B, we warm up the Victim network for 3 epochs on CIFAR-10 and 6 epochs on the ImageNet subset, where the filtering threshold \\( t_f \\) linearly decreases from 1 to 0.2 and from 1 to 0.4. The selection of these hyperparameters and an automatic warming-up strategy can be found in Appendix A.3. In the following stages, the filtering threshold keeps unchanged (i.e. 0.2 for CIFAR-10 and 0.4 for ImageNet subset). For fast convergence, we adopt the Adam optimizer to train the Victim network with a learning rate of 0.001 in the whole training process. We optimize the Beneficiary network for 200 epochs using an SGD optimizer with a momentum of 0.9 and a weight decay factor of \\( 10^{-4} \\), where the first 100 epochs belong to stage 2 and the last 100 epochs belong to stage 3. The learning rate is initially set to 0.1 and is divided by 10 after every 50" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "160" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.191, + 0.089, + 0.784, + 0.181 + ], + "angle": 0, + "content": "
AttackBadNets [10]Blend [7]Dynamic [23]WaNet [22]CL-16 [26]
MetricBAASRBAASRBAASRBAASRBAASR
No Defense94.36%100%94.57%99.76%94.66%95.18%94.21%95.84%94.98%96.93%
Mixup [31]93.37%0.27%92.80%3.49%92.97%42.10%92.97%42.10%94.80%20.03%
Cutmix [30]93.52%0.40%93.40%1.07%91.40%53.23%91.40%53.23%94.54%1.16%
Attentive Cutmix [27]94.26%0.40%94.55%4.62%93.09%13.38%93.09%13.88%95.18%2.97%
AttentionMix (Ours)93.96%0.62%94.37%0.63%93.91%1.13%94.15%0.54%94.24%1.01%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.18, + 0.19, + 0.785, + 0.204 + ], + "angle": 0, + "content": "Table 3. The inhibition effect of 4 strong data augmentations against 5 backdoor attacks on CIFAR-10." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.471, + 0.336 + ], + "angle": 0, + "content": "epochs. In particular, the Victim network is optimized every 5 epochs at stage 2, i.e. one credible set is used to train the Beneficiary network for 5 epochs. We use a batch size of 128 for CIFAR-10 and 16 for the ImageNet subset, and set the mixing threshold \\( t_m \\) to 0.1 in all cases. Following semi-supervised learning, we gradually increase the weight of suspicious sample loss in stage 3, simply setting \\( \\alpha \\) to 0.1 in the first 50 epochs and 0.5 in the last 50 epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.47, + 0.463 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate the performance of defenses with two commonly used metrics: the model's classification accuracy on the benign test set, Benign Accuracy (BA), and the model's classification accuracy on the poison test set, Attack Success Rate (ASR). The lower the ASR and the higher BA, the stronger the defense. In particular, the poison test set removes the samples whose ground-truth label is the target label." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.47, + 0.379, + 0.484 + ], + "angle": 0, + "content": "4.2. Effectiveness of Our V&B Defense" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.583 + ], + "angle": 0, + "content": "To measure the effectiveness of our proposed V&B method, we compare its performance with four state-of-the-art defenses on CIFAR-10 and the ImageNet subset. The backdoor attack CL-16 and CL-32 are two versions of the Clean-label attack with different magnitudes of adversarial perturbations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Table 1 demonstrates the defense results on CIFAR-10, where No Defense is a baseline representing the capability of backdoor attacks. Considering real scenarios, we adopt common data augmentations (i.e. random crop and horizontal flipping) during training, which will slightly affect the attack success rate. It is obvious that our V&B achieves the best results against almost all attacks, even compared to defenses that need additional benign samples. Our method reduces the success rate of all attacks from above \\(95\\%\\) (64.79% for SIG) to blow or close to \\(1\\%\\), while maintaining the accuracy on benign samples with only about \\(1\\%\\) reduction. For WaNet and CL-16, V&B achieves the second-best performance in reducing the attack success rate, but the benign accuracy of V&B is significantly superior to the corresponding best defense, i.e. DBD and ABL. WaNet adopts a warping field as the trigger and adds a noisy training mode to interfere with detection. CL-16 only poisons the samples whose ground-truth label is the target label and adds adversarial perturbations to them. So it is relatively hard to distinguish the poisoned samples they craft from the benign ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.892, + 0.381 + ], + "angle": 0, + "content": "The defense performances on the ImageNet subset are shown in Table 2. Compared with other defenses, our method has the best accuracy on benign samples against all four attacks, even outperforming the baseline due to our AttentionMix data augmentation and semi-supervised suppression training. Semi-supervised learning can relabel poisoned samples to their ground-truth labels, improving benign accuracy. In terms of reducing the attack success rate, our method achieves the best results against BadNets, Blend, and SIG, and the second-best result against WaNet, which is comparable to the best result (1.92% vs. 1.00%)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.457 + ], + "angle": 0, + "content": "We also validate the effectiveness of our method under different poisoning rates in Appendix A.4. Even with a poisoning rate of \\(50\\%\\), our method can still reduce the attack success rate of most attacks to below \\(1\\%\\), while maintaining a satisfactory benign accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.467, + 0.765, + 0.481 + ], + "angle": 0, + "content": "4.3. Effectiveness of AttentionMix" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.777 + ], + "angle": 0, + "content": "To illustrate the inhibition effect of our AttentionMix, we replace it with three other popular strong data augmentations in our framework, i.e. Mixup, Cutmix and Attentive Cutmix. As Table 3 shows, each data augmentation performs well against BadNets and Blend, while for more stealthy attacks, such as WaNet and Dynamic, our AttentionMix can better reduce the attack success rate. This is because most of poisoned samples crafted by BadNets and Blend can be filtered out before training the Beneficiary network due to the simple trigger patterns. But for stealthy attacks, the Victim network will only separate part of the poisoned samples after warming up, and the other three augmentations can not effectively inhibit the missed poisoned samples. For a poisoned sample, the poisoned network will focus on its trigger location, so blending the sample's activation region with other images can effectively destroy triggers and help to clean the network. Although Attentive Cutmix can locate the trigger region, direct pasting may leave the trigger intact in the generated image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.665, + 0.802 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Component Effects. As shown in Table 4, We first train the network without any defense as a baseline and then add our components one by one to investigate their effect. The variants are as follows: \\((1) + F\\): filtering the training set and training on credible samples; \\((2) + F + AT\\): using AttentionMix when training with credible samples; \\((3) + F +\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "161" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.894, + 0.205 + ], + "angle": 0, + "content": "
AttackBadNets [10]Blend [7]WaNet [22]Dynamic [23]CL-16 [26]
MetricBAASRBAASRBAASRBAASRBAASR
No Defense94.36%100.00%94.57%99.76%94.21%95.84%94.66%95.18%94.98%96.93%
+F93.50%10.01%93.47%76.70%92.92%90.97%93.80%93.04%93.63%2.70%
+F + AT92.27%0.42%92.72%1.42%91.58%25.16%93.08%50.01%93.09%2.09%
+F + AT + SST93.88%1.21%94.56%1.52%93.31%1.23%94.32%28.79%94.43%3.29%
+F + AT + SST + S (suspicious)94.11%1.00%94.49%0.73%93.49%0.61%94.61%1.43%94.00%1.06%
ALL93.96%0.62%94.37%0.63%94.15%0.54%93.91%1.13%94.24%1.01%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.238, + 0.215, + 0.726, + 0.228 + ], + "angle": 0, + "content": "Table 4. Ablation study of individual components in our framework on CIFAR-10." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.469, + 0.314 + ], + "angle": 0, + "content": "AT + SST: adding standard semi-supervised training [2] on the basis of the second variant; (4) + F + AT + SST + S (suspicious): suppressing suspicious samples during semi-supervised training on the basis of the third variant." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.469, + 0.6 + ], + "angle": 0, + "content": "As can be seen from Table 4, when defending against Blend, WaNet, and Dynamic, only filtering will fail because enough poisoned samples are missed to create backdoors. After using AttentionMix, although the attack success rate of most attacks is severely decreased, the benign accuracy also drops slightly. Thus semi-supervised training is adopted to improve the performance of the network on benign samples. For WaNet and Dynamic, semi-supervised training also reduces their attack success rate by relabeling poisoned samples as ground-truth labels. However, the Beneficiary network is not completely clean, so a small number of poisoned samples may be relabeled as the target label, causing the attack success rate to rebound (e.g. Bad-Nets, Blend and Cl-16). Suppression to suspicious samples can make up for the weakness of relabeling and adding suppression to credible samples will further reduce the attack success rate. We say that the complete framework is more secure when facing backdoor attacks, although its benign accuracy may be lower than other variants." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.469, + 0.707 + ], + "angle": 0, + "content": "To further demonstrate the effect of stage 3, we compare the final results with those after stage 2 in Appendix A.5. In all cases, the attack success rate after stage 2 is higher than the final result, while the benign accuracy after stage 2 is lower than the final result, suggesting that our semi-supervised suppression training can both erasing existing backdoors and improve model performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Selection of Mixing Threshold. As Figure 3 shows, we set the mixing threshold \\( t_m \\) from 0.1 to 0.5 on CIFAR-10 and ImageNet subset to study its impact on model performance. It can be seen that the benign accuracy on CIFAR-10 starts to drop when the mixing threshold is large enough (e.g. 0.2 for BadNets and 0.3 for Blend), while the attack success rate only slightly fluctuates. The decline in benign accuracy is mainly caused by the unlabeled samples in stage 3. Due to the small image size \\( (32 \\times 32 \\) in CIFAR-10), the activation region will contain only a few clustered pixels or some separate pixels if the mixing threshold is too large. During mixing, these pixels will carry their labels but" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.25, + 0.698, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.251, + 0.892, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.383, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Figure 3. Ablation study of mixing threshold \\( t_m \\) on CIFAR-10 (left) and ImageNet subset (right)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.437, + 0.892, + 0.542 + ], + "angle": 0, + "content": "can not provide the feature of corresponding classes, which is equivalent to introducing noise to labels. It hinders the learning of our Beneficiary network, leading to more and more suspicious samples being mislabeled in stage 3. This point can be confirmed on the ImageNet subset (image size is \\(224 \\times 224\\)), where the mixing threshold has little impact on our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.618, + 0.572 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.839 + ], + "angle": 0, + "content": "In this paper, we show that poisoned samples could be easily distinguished by the entropy of the poisoned network's prediction. Based on this finding and the fact that strong data augmentation can inhibit backdoor injection, we propose a novel secure training framework consisting of two parallel networks to prevent a backdoor generation from three aspects: (1) filtering training data; (2) inhibiting backdoor injection; (3) erasing potential backdoors. Specifically, we sacrifice one network to exclusively learn backdoor triggers and filter credible samples according to the prediction entropy to train another network. During training, we design a strong data augmentation to inhibit missed poisoned samples from working and then introduced a semi-supervised learning method to erase potential backdoors with the help of the sacrifice network. Extensive experiments verify that our framework is effective against various backdoor attacks under different poisoning rates." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.85, + 0.892, + 0.895 + ], + "angle": 0, + "content": "Acknowledgment. This work is supported in part by the National Natural Science Foundation of China Under Grants No.62176253 and No.U20B2066." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "162" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A New Backdoor Attack in CNNS by Training Set Corruption Without Label Poisoning. In 2019 IEEE International Conference on Image Processing, pages 101-105. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. MixMatch: A Holistic Approach to Semi-Supervised Learning. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.314 + ], + "angle": 0, + "content": "[3] Eitan Borgnia, Valerii Cherepanova, Liam Fowl, Amin Ghiasi, Jonas Geiping, Micah Goldblum, Tom Goldstein, and Arjun Gupta. Strong Data Augmentation Sanitizes Poisoning and Backdoor Attacks Without an Accuracy Tradeoff. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 3855-3859. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[4] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian Molloy, and Biplav Srivastava. Detecting Backdoor Attacks on Deep Neural Networks by Activation Clustering. arXiv preprint arXiv:1811.03728, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[5] Huili Chen, Cheng Fu, Jishen Zhao, and Farinaz Koushanfar. ProFlip: Targeted Trojan Attack With Progressive Bit Flips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7718-7727, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.47, + 0.556 + ], + "angle": 0, + "content": "[7] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted Backdoor Attacks on Deep Learning Systems Using Data Poisoning. arXiv preprint arXiv:1712.05526, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[8] Yanbei Chen, Massimiliano Mancini, Xiatian Zhu, and Zeynep Akata. Semi-Supervised and Unsupervised Deep Visual Learning: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.47, + 0.684 + ], + "angle": 0, + "content": "[9] Bao Gia Doan, Ehsan Abbasnejad, and Damith C Ranasinghe. Februus: Input Purification Defense Against Trojan Attacks on Deep Neural Network Systems. In Annual Computer Security Applications Conference, pages 897-912, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.741 + ], + "angle": 0, + "content": "[10] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain. arXiv preprint arXiv:1708.06733, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[11] Chuan Guo, Ruihan Wu, and Kilian Q Weinberger. Trojannet: Embedding Hidden Trojan Horse Models in Neural Networks. arXiv preprint arXiv:2002.10078, 2, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[12] Jonathan Hayase, Weihao Kong, Raghav Somani, and Sewoong Oh. SPECTRE: Defending Against Backdoor Attacks Using Robust Statistics. arXiv preprint arXiv:2104.11315, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Kunzhe Huang, Yiming Li, Baoyuan Wu, Zhan Qin, and Kui Ren. Backdoor Defense via Decoupling the Training Process. In The Tenth International Conference on Learning Representations, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[15] Junnan Li, Richard Socher, and Steven CH Hoi. DIVIDEMIX: LEARNING WITH NOISY LABELS AS SEMI-SUPERVISED LEARNING. arXiv preprint arXiv:2002.07394, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[16] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible Backdoor Attack With Sample-Specific Triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[17] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-Backdoor Learning: Training Clean Models on Poisoned Data. Advances in Neural Information Processing Systems, 34:14900–14912, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[18] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[19] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273–294. Springer, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[20] Xuankai Liu, Fengting Li, Bihan Wen, and Qi Li. Removing Backdoor-Based Watermarks in Neural Networks with Limited Data. In International Conference on Pattern Recognition, pages 10149-10156. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[21] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection Backdoor: A Natural Backdoor Attack on Deep Neural Networks. In European Conference on Computer Vision, pages 182-199. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[22] Anh Nguyen and Anh Tran. WaNet-M Imperceptible Warping-based Backdoor Attack. arXiv preprint arXiv:2102.10369, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[23] Tuan Anh Nguyen and Anh Tran. Input-Aware Dynamic Backdoor Attack. Advances in Neural Information Processing Systems, 33:3454-3464, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[24] Xiangyu Qi, Tinghao Xie, Ruizhe Pan, Jifeng Zhu, Yong Yang, and Kai Bu. Towards Practical Deployment-Stage Backdoor Attack on Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13347-13357, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[25] Han Qiu, Yi Zeng, Shangwei Guo, Tianwei Zhang, Meikang Qiu, and Bhavani Thuraisingham. DeepSweep: An Evaluation Framework for Mitigating DNN Backdoor Attacks using Data Augmentation. In Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security, pages 363-377, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-Consistent Backdoor Attacks. arXiv preprint arXiv:1912.02771, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Devesh Walawalkar, Zhiqiang Shen, Zechun Liu, and Marios Savvides. Attentive CutMix: An Enhanced Data Augmentation Approach for Deep Learning Based Image Classification. arXiv preprint arXiv:2003.13048, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "163" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural Cleans: Identifying and Mitigating Backdoor Attacks in Neural Networks. In 2019 IEEE Symposium on Security and Privacy, pages 707-723. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[29] Jason Weston, Frédéric Ratle, and Ronan Collobert. Deep Learning via Semi-Supervised Embedding. In Proceedings of the 25th international conference on Machine learning, pages 1168-1175, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[30] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6023-6032, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.345 + ], + "angle": 0, + "content": "[31] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. Mixup: Beyond Empirical Risk Minimization. In 6th International Conference on Learning Representations, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.47, + 0.388 + ], + "angle": 0, + "content": "[32] Pu Zhao, Pin-Yu Chen, Payel Das, Karthikeyan Natesan Ramamurthy, and Xue Lin. Bridging Mode Connectivity in Loss Landscapes and Adversarial Robustness. 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.47, + 0.472 + ], + "angle": 0, + "content": "[33] Zhendong Zhao, Xiaojun Chen, Yuexin Xuan, Ye Dong, Dakui Wang, and Kaitai Liang. DEFEAT: Deep Hidden Feature Backdoor Attacks by Imperceptible Perturbation and Latent Representation Constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15213-15222, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.474, + 0.47, + 0.515 + ], + "angle": 0, + "content": "[34] Nan Zhong, Zhenxing Qian, and Xinpeng Zhang. Imperceptible Backdoor Attack: From Input Space to Feature Representation. arXiv preprint arXiv:2205.03190, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[35] Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Scholkopf. Learning with Local and Global Consistency. Advances in neural information processing systems, 16, 2003." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "164" + } + ] +] \ No newline at end of file diff --git a/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/fdfec7d9-d3ec-4468-9f42-c0ec2550345d_origin.pdf b/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/fdfec7d9-d3ec-4468-9f42-c0ec2550345d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..25d53a6b01b8b43e630904261d35fd5d84dc4f89 --- /dev/null +++ b/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/fdfec7d9-d3ec-4468-9f42-c0ec2550345d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b05c73cc010e5c1a6c0b3b757f207371068217b721f95b50402b433de0d399a +size 1132873 diff --git a/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/full.md b/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2c8f4c3ebbb145b8dc2ad1679fb420eff6996b6a --- /dev/null +++ b/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/full.md @@ -0,0 +1,286 @@ +# The Victim and The Beneficiary: Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data + +Zixuan Zhu, Rui Wang*, Cong Zou, Lihua Jing +SKLOIS, Institute of Information Engineering, CAS, Beijing, China +School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China +{zhuzixuan, wangrui, zoucong, jinglihua}@iei.ac.cn + +# Abstract + +Recently, backdoor attacks have posed a serious security threat to the training process of deep neural networks (DNNs). The attacked model behaves normally on benign samples but outputs a specific result when the trigger is present. However, compared with the rocketing progress of backdoor attacks, existing defenses are difficult to deal with these threats effectively or require benign samples to work, which may be unavailable in real scenarios. In this paper, we find that the poisoned samples and benign samples can be distinguished with prediction entropy. This inspires us to propose a novel dual-network training framework: The Victim and The Beneficiary (V&B), which exploits a poisoned model to train a clean model without extra benign samples. Firstly, we sacrifice the Victim network to be a powerful poisoned sample detector by training on suspicious samples. Secondly, we train the Beneficiary network on the credible samples selected by the Victim to inhibit backdoor injection. Thirdly, a semi-supervised suppression strategy is adopted for erasing potential backdoors and improving model performance. Furthermore, to better inhibit missed poisoned samples, we propose a strong data augmentation method, AttentionMix, which works well with our proposed V&B framework. Extensive experiments on two widely used datasets against 6 state-of-the-art attacks demonstrate that our framework is effective in preventing backdoor injection and robust to various attacks while maintaining the performance on benign samples. Our code is available at https://github.com/Zixuan-Zhu/VaB. + +# 1. Introduction + +Large amounts of training data and powerful computing resources are two key factors for the success of deep neural networks (DNNs). While in practical applications, obtaining enough samples for training is laborious, and training is likely to be resource-constrained. As a result, more and + +![](images/7272a10b706b0a05e6e2a427edef074ab60c8415e1eb25313c2d3eb57128a5f4.jpg) +Figure 1. The average prediction entropy of benign samples versus poisoned samples crafted by 6 backdoor attacks. We conduct the experiment on CIFAR-10 with ResNet-18 under poisoning rate $10\%$ , where our special training strategy (described in section 3.3) was adopted. + +more third-party data (e.g. data released on the Internet) and training platforms (e.g. Google Colab) are adopted to reduce the overhead, which brings new security risks. + +Backdoor attacks [5, 11, 24] pose a serious security threat to the training process of DNNs and can be easily executed through data poisoning. Specifically, attackers [7, 10, 22] inject the designed trigger (e.g. a small patch or random noise) to a few benign samples selected from the training set and change their labels to the attacker-defined target label. These poisoned samples will force models to learn the correlation between the trigger and the target label. In the reference process, the attacked model will behave normally on benign samples but output the target label when the trigger is present. Deploying attacked models can lead to severe consequences, even life-threatening in some scenarios (e.g. autonomous driving). Hence, a secure training framework is needed when using third-party data or training platforms. + +Many researchers have devoted themselves to this topic and proposed feasible defenses. Li et al. [18] utilize local benign samples to erase existing backdoors in DNNs, and + +Borgnia et al. [3] employ strong data augmentations to inhibit backdoor injection during training. Li et al. [17] find that the training loss decreased faster for poisoned samples than benign samples and designed a loss function to separate them. Finally, they select a fixed portion of samples with the lowest loss to erase potential backdoors. In real scenarios, it is challenging to distinguish poisoned samples from benign ones in this way due to the unknown poisoning rate and close loss values. In experiments, we find the entropy of model prediction is a more discriminative property, and a fixed threshold could filter out most poisoned samples, as shown in Figure 1. During training, poisoned samples will be learned faster than benign samples due to the similar feature of their triggers [17]. Hence, the poisoned network can confidently predict poisoned samples as the target label in early epochs but hesitates about the benign samples. + +Based on this observation, we propose a novel secure training framework, The Victim and The Beneficiary (V&B), that can directly train clean models on poisoned data without resorting to benign samples. Firstly, in warming up stage, the training dataset is recomposed into a suspicious set and a credible set according to the prediction entropy is lower or higher than the predefined threshold. Then the Victim network is trained iteratively with the suspicious set to obtain a strong poisoned sample detector. Secondly, in the clean training stage, the Beneficiary network is trained with credible samples filtered by the Victim to inhibit backdoor injection. In later training epochs, the Victim network is fully exploited to provide its learned knowledge for the Beneficiary network to help the latter erase potential backdoors through semi-supervised learning. To diminish the threat of missed poisoned samples, we design a strong data augmentation AttentionMix, which mixes the image region with high activation values according to the attention map. Compared with existing augmentations [27, 30, 31], it has a stronger inhibition effect against stealthy backdoor attacks [23, 22, 26]. + +The main contributions of this paper are as follows: (1) We propose a novel dual-network secure training framework, V&B, which can train clean models on the poisoned dataset without resorting to benign samples. (2) We design a powerful data augmentation method called AttentionMix, which has a stronger inhibition effect against stealthy backdoor attacks. (3) Extensive experiments on two popular benchmark datasets demonstrate the effectiveness and robustness of our framework. + +# 2. Related Work + +# 2.1. Data Augmentation + +Data Augmentation is a common-used technique to improve the generalization capabilities of DNNs. In addition + +to some popular weak data augmentations (such as cropping, flipping, rotation, etc.), many strong data augmentation techniques [27, 30, 31] have been proposed to further improve model performance. Mixup [31] linearly interpolates any two images and their labels, then trains the model with generated samples. However, the generated samples tend to be unnatural, which will decrease model performances on localization and object detection tasks. Instead of linear interpolation, Cutmix [30] replaces a random patch with the same region from another image and mixes their labels according to the ratio of the patch area. Further, Attentive Cutmix [27] only pastes the influential region located by an attention mechanism to other images, effectively forcing models to learn discriminative details. Yet it also mixes labels based on the ratio of the region area, ignoring the importance of the region in two original images, and direct pasting may bring complete triggers into benign samples. Our AttentionMix takes into account the importance of the region in both images and blends the influential region with the same area in another image, which can destroy the completeness of triggers. + +# 2.2. Backdoor Attack and Defense + +In this paper, we only consider poisoning-based attacks toward image classification that can occur during the data preparation stage. + +Poisoned-based backdoor attack. In early studies, attackers usually adopt simple and visible patterns as the trigger, such as a black-white checkerboard [10]. To escape human inspections, some works use human-imperceptible deformations [22] as the trigger or design unique triggers [16, 23] for each poisoned sample. However, they randomly select benign samples to inject triggers and change their labels (called poison-label attacks), possibly causing the content of an image not to match its label. Instead of randomly selecting samples to poison, Turner et al. [26], Mauro et al. [1], and Liu et al. [21] only inject triggers into benign samples under the target label (called clean-label attacks). Although some triggers are imperceptible in the input space, they can be easily detected by defenses focused on latent space extracted by CNNs. Therefore, recent studies [33, 34] not only ensure the trigger's invisibility in the input space but also restrict the similarity between poison samples and corresponding benign samples in the latent space. + +Backdoor Defense. For poisoning-based attacks, defense methods could be divided into three categories from their working stage, including (1) preprocessing before training [9, 12, 25], (2) poison suppression during training [14, 17, 20], and (3) backdoor erasing after training [18, 28, 32]. For preprocessing, Hayase et al. [12] and Chen et al. [4] filter the training set through statistics or clustering and + +train the model with benign samples. For poison suppressing, Borgnia et al. [3] leverage strong data augmentation (e.g. Mixup, Cutmix) to alleviate poisoned samples' adverse effects. Huang et al. [14] design a decoupling training process, which first learns a feature extractor via SimCLR [6] to prevent poisoned samples from clustering in feature space. For backdoor erasing, Liu et al. [19] prune the low-activation neurons for validation samples (benign samples) and then fine-tune the model with these benign samples. While Li et al. [18] first fine-tune the poisoned model with a small portion of benign samples to obtain a teacher model, then erase the original model's backdoor through knowledge distillation. + +# 2.3. Semi-supervised Learning + +Semi-supervised learning [8] aims to alleviate the dependence of deep learning on extensive labeled training data and jointly learns the model from a limited amount of labeled samples and a large number of unlabeled samples. Existing methods [2, 29, 35] generally assign each unlabeled sample with a pseudo label and design different loss functions for labeled and unlabeled samples. The weight of unlabeled sample loss will gradually increase during training to reduce the negative effect of low-quality pseudo labels in early epochs. + +# 3. Method + +# 3.1. Preliminaries + +Pipeline of Poisoned-based Backdoor Attacks. Let $\mathcal{D}_B = \{(x_i, y_i)\}_{i=1}^N$ denotes the benign training set containing $N$ samples, where $x_i \in \{0, 1, \dots, 255\}^{C \times H \times W}$ is the image, $y_i \in \{0, 1, \dots, K\}$ is its label, $K$ is the number of classes, $H, W$ and $C$ are the height, width and channels of an image. The attacker will select a subset of benign samples to poison with a designed injection function $G(\cdot)$ . Let $\mathcal{D}_G$ indicates the generated poisoned samples set and $\mathcal{D}_R$ denotes the remaining benign samples set, where $\mathcal{D}_G = \{(x', y_t) | x' = G(x), (x, y) \in \mathcal{D}_B \setminus \mathcal{D}_R\}$ , $y_t$ is the target label and $r = \frac{\left|\mathcal{D}_G\right|}{\mathcal{D}_B}$ is the poisoning rate. The poisoned dataset $\mathcal{D}_P = \mathcal{D}_G \cup \mathcal{D}_R$ will be unknowingly adopted by victim users to train DNNs. At inference time, the attacker can manipulate the poisoned model's output through the injected backdoor. + +Threat Model and Defense Setting. In this paper, we focus on defending against poisoned-based backdoor attacks toward image classification. The attacker can arbitrarily modify the training set, but can not change other training components (e.g. model structure and hyperparameters) [1, 10, 22]. Such attacks may occur when using third-party datasets or training platforms that can poison users' clean data. We follow the common assumption [12, 14, 17] that + +defenders have no information about the backdoor (including poisoning rate, trigger pattern, target label, etc.) and any local benign samples to use but have full control over the training process. The defense goal is to prevent the trained DNNs from predicting poisoned samples as the target label while maintaining its accuracy on benign samples. + +# 3.2. Overview + +In this subsection, we briefly introduce the pipeline of our framework. As Figure 2(a) illustrates, our framework consists of two networks, and the whole training process is divided into three stages — warming up, clean training and semi-supervised suppression training. In warming up, we aim to obtain a poisoned network that can roughly distinguish poisoned samples from benign ones according to the prediction entropy. Therefore we train the Victim network with only filtered suspicious samples (i.e. suspicious set) except for the first epoch. This training strategy remains unchanged in the following two stages. During clean training, we adopt credible samples (i.e. credible set) to train the Beneficiary network with our proposed AttentionMix data augmentation. The strong data augmentation can effectively prevent missed poisoned samples from creating backdoors. As the Victim network is trained, the entropy of its prediction for benign samples will also slowly decrease below the filtering threshold, leading to a reduction in the credible set size. So one credible set will be used for a few epochs to ensure the Beneficiary network gets enough training samples. + +To further improve the Beneficiary network performance and erase its potential backdoor, we regard the credible set as a labeled dataset and the suspicious set as an unlabeled dataset to train the Beneficiary network via semi-supervised learning. Following conventional practice [2, 15], each unlabeled sample will be assigned a pseudo label generated by the Beneficiary network, forming the relabeled set. However, if the Beneficiary network still contains a backdoor after stage 2, poisoned samples in the suspicious set will be relabeled as the target label, strengthening the backdoor. After specific training, the Victim network is sensitive to trigger patterns and thus can be utilized to suppress the component corresponding to the target class after relabeling. Furthermore, we also apply this suppression operation to the credible set to further weaken the impact of missed poisoned samples. + +# 3.3. Warming Up + +At this stage, we aim to obtain a poisoned network that has only learned trigger patterns. In this way, the entropy of its prediction for poisoned samples will be significantly lower than that of benign samples. Then we can use a threshold to filter out poisoned samples. + +Let $\mathcal{D}$ denotes the training set and $F_{v}:\{0,1,\dots,255\}^{C\times H\times W}\to [0,1]^{K}$ indicates the Vic + +![](images/666295a13902e0ab4279b53e1e326fc6bb0c5c928d306c2a0752f197d7d820e9.jpg) +(a) The pipeline of our training framework + +![](images/eb3f649cbc22b76de502a6ef815778445d1c28d55a780d4a7f0c3cabe4910264.jpg) +Figure 2. (a) The pipeline of our training framework. Net V is the Victim network (i.e. poisoned network) that only learns from the filtered suspicious set, and Net B is the Beneficiary network (i.e. clean network) we wanted. Red triangles represent poisoned samples, and green circles are benign samples. The color of nets indicates their degree of poisoning: the darker the color, the deeper the poisoning, and vice versa. Deep blue areas are three training stages, and light blue areas are basic blocks that will be executed many times in the corresponding stage. (b) An example of AttentionMix. The bottom four images are the non-activation region of the cat image, the activation region of the cat image, the same activation region in the dog image, and the final generated image, respectively. And their labels are located below the images. (c) Suppression for poisoned samples and benign samples. The target label is assumed as the first label. Suppression can reduce the target label component of poisoned samples but have little influence on the ground-truth label of benign samples. + +![](images/fe32f823de59ee29057c059f09bfd60653437c7e5b3468c23fe139e29c8862b8.jpg) + +tim network. We first use the whole training set to train the Victim network for one epoch with cross-entropy loss. Then the prediction entropy of a sample will be calculated with Equation 1, where $p$ is the prediction of the Victim network for that sample. + +$$ +e = - \sum_ {k = 1} ^ {K} p _ {k} \log p _ {k} \tag {1} +$$ + +Once all samples' prediction entropy is obtained, we normalize them to $[0,1]$ through mix-max normalization. Then a threshold $t_f$ is set to filter out suspicious samples $\mathcal{D}_s = \{(x_i,y_i)|(x_i,y_i)\in \mathcal{D},e_i < t_f\}$ and the remaining samples are regarded as credible samples $\mathcal{D}_c$ . To strengthen the learning of poisoned samples and delay the learning of benign samples, the Victim network will only be trained on suspicious sets in the future. The loss function is as follows: + +$$ +\mathcal {L} _ {v} = \mathbb {E} _ {(\boldsymbol {x}, y) \in \mathcal {D} _ {s}} \left[ \mathcal {L} _ {C E} \left(F _ {v} (\boldsymbol {x}), y\right) \right] \tag {2} +$$ + +where $\mathcal{L}_{CE}$ denotes the cross-entropy loss. + +After the first epoch, the prediction entropy of some poisoned samples may still be larger than that of most benign ones, although there is a large gap between their average values in Figure 1. If the filtering threshold $t_f$ is set too large, lots of benign samples will be discarded, which will affect the model performance on benign samples. So we set a smaller $t_f$ and wait for the prediction entropy of these poisoned samples to decrease. As training progresses, their prediction entropy will drop below $t_f$ , making the credible set cleaner. Simultaneously, the prediction entropy of benign samples will also decrease, resulting in a smaller credible set. We expect the Victim network to filter out a clean and relatively large credible set after this stage. + +# 3.4. Clean Training + +At this stage, we begin to train the Beneficiary network with filtered credible samples $\mathcal{D}_c$ , where the filtering strategy is the same as in stage 1. Because of the rough separation, a few poisoned samples may be included in the credible set. For strong attacks, a small number of poisoned sam + +ples is enough to create a backdoor. So we design a strong data augmentation called AttentionMix to inhibit the missed poisoned samples from working. Compared with existing augmentations [27, 30, 31], AttentionMix has a stronger inhibition effect on stealthy backdoor attacks. + +The core idea of AttentionMix is to generate a new training sample $(\tilde{\boldsymbol{x}},\tilde{\boldsymbol{y}})$ by mixing two training samples $(x_{1},y_{1})$ and $(x_{2},y_{2})$ according to their attention maps, where $\tilde{\boldsymbol{y}}$ is a K-dimensional vector. Following [18], we adopt one sample's final feature maps to compute its attention map: + +$$ +\boldsymbol {A} ^ {\prime} = \sum_ {i = 1} ^ {c} \left| F _ {i} \right| ^ {2} \tag {3} +$$ + +where $\pmb{F} \in \mathbb{R}^{c \times h \times w}$ indicates feature maps, $\pmb{A}' \in \mathbb{R}^{h \times w}$ is the corresponding attention map, $c, h$ and $w$ are the dimensions of channel, height and width. For each attention map, we first bilinearly interpolate it to the original image size $H \times W$ , and then normalize its elements to $[0,1]$ by min-max normalization. + +After getting the attention map, we set a threshold $t_m$ to obtain the activation region that we want to mix: + +$$ +\boldsymbol {M} _ {i, j} = \left\{ \begin{array}{l l} 0 & \boldsymbol {A} _ {i, j} < t _ {m} \\ 1 & \boldsymbol {A} _ {i, j} \geq t _ {m} \end{array} \right. \tag {4} +$$ + +where $M \in \mathbb{R}^{H \times W}$ is the binary mask indicating which pixels belong to the activation region and $\pmb{A}$ is the attention map after interpolation and normalization. + +Let $A^1$ and $A^2$ denote the attention maps of $x_1$ and $x_2$ , $M^1$ denotes the first sample's mask. As Figure 2(b) shows, the mixed image $\tilde{x}$ is composed of three parts: the non-activation region of $x_1$ , the activation region of $x_1$ and the same area in $x_2$ . During mixing, the non-activation region of $x_1$ remains unchanged, and the activation region is blended with the same area of $x_2$ : + +$$ +\begin{array}{l} \tilde {\boldsymbol {x}} = \left(\mathbf {1} - \boldsymbol {M} ^ {1}\right) \odot \boldsymbol {x} _ {1} + \lambda_ {2} \cdot \left(\boldsymbol {M} ^ {1} \odot \boldsymbol {x} _ {1}\right) \tag {5} \\ + \lambda_ {3} \cdot \left(\boldsymbol {M} ^ {1} \odot \boldsymbol {x} _ {2}\right) \\ \end{array} +$$ + +where $\odot$ is the Hadamard Product, $\lambda_{2}$ and $\lambda_{3}$ are the mixing weights calculated from the two attention maps. The calculation process is as follows: + +$$ +w _ {1} = \frac {\operatorname {s u m} \left(\boldsymbol {M} ^ {1}\right)}{\operatorname {s u m} \left(\boldsymbol {A} ^ {1}\right)}, w _ {2} = \frac {\operatorname {s u m} \left(\boldsymbol {M} ^ {1} \odot \boldsymbol {A} ^ {2}\right)}{\operatorname {s u m} \left(\boldsymbol {A} ^ {2}\right)} \tag {6} +$$ + +$$ +\lambda_ {1} = 1 - w _ {1}, \lambda_ {2} = \frac {w _ {1}}{w _ {1} + w _ {2}}, \lambda_ {3} = \frac {w _ {2}}{w _ {1} + w _ {2}} \tag {7} +$$ + +where $\mathrm{sum}(\cdot)$ is a summation function, $w_{1}$ and $w_{2}$ denote the importance of the activation region to the two images, and $\lambda_{1}$ is the importance of the non-activation region to $x_{1}$ . The final label is also composed of three parts: + +$$ +\tilde {\boldsymbol {y}} = \lambda_ {1} \cdot \boldsymbol {y} _ {1} + w _ {1} \cdot \left(\lambda_ {2} \cdot \boldsymbol {y} _ {1} + \lambda_ {3} \cdot \boldsymbol {y} _ {2}\right) \tag {8} +$$ + +where $\pmb{y}_1$ and $\pmb{y}_2$ are the one-hot labels of the two samples. When the activation region of $\pmb{x}_1$ corresponds to the background of $\pmb{x}_2$ , $w_1$ will be significantly larger than $w_2$ , and $\lambda_3$ will be a small value, which can reduce background interference during mixing. + +We denote the generated new credible set as $\tilde{\mathcal{D}}_c$ and train the Beneficiary network $F_{b}$ with the following loss function (i.e. Train with AttentionMix in Figure 2(a)): + +$$ +\mathcal {L} _ {b} = \mathbb {E} _ {(\tilde {\boldsymbol {x}}, \tilde {\boldsymbol {y}}) \in \tilde {\mathcal {D}} _ {c}} [ \mathcal {L} _ {C E} (F _ {b} (\tilde {\boldsymbol {x}}), \tilde {\boldsymbol {y}}) ] \tag {9} +$$ + +For a poisoned network, the activation region of poisoned samples is more likely their trigger locations. Mixing the activation region can destroy the completeness of trigger patterns. While for benign samples, the mixing plays a regularization role that prevents the network from overfitting a particular subject and forces it to learn abundant features. + +# 3.5. Semi-supervised Suppression Training + +Until now, we have got a poisoned Victim network that can filter out most of the poisoned samples and a Beneficiary network that is relatively clean but may still contain the backdoor. In the previous stage, the Beneficiary network is trained only with partial training samples (i.e. the credible set), which may harm its performance on benign samples. Thus we intend to enable the deprecated suspicious samples via semi-supervised learning at this stage to improve the Beneficiary network while erasing its potential backdoor. + +Given a credible set $\mathcal{D}_c$ and a suspicious set $\mathcal{D}_s$ , we use samples in $\mathcal{D}_c$ as labeled samples and remove the labels of samples in $\mathcal{D}_s$ to transform the problem into a semi-supervised learning setting. For each unlabeled sample, we average the predictions of the Beneficiary network over its multiple augmentations (e.g. random crop and random flip) and select the class with max probability as its pseudo-label, which is similar to [2] and [15]. The relabeled samples are denoted as $\mathcal{D}_r = \{(\pmb{x}_i,\hat{\pmb{y}}_i)|(\pmb{x}_i,\pmb{y}_i)\in \mathcal{D}_s\}$ . + +However, the poisoned samples in $\mathcal{D}_s$ may be relabeled as the target label, i.e. $\hat{y}_i = y_t$ , if the Beneficiary network still contains the backdoor. We can mitigate the impact of such mislabeling by exploiting the knowledge learned by the Victim network. Specifically, we adopt the Victim network's prediction to suppress the pseudo-label (i.e. Suppress in Figure 2(a)): + +$$ +\boldsymbol {y} _ {i} ^ {\prime} = \operatorname {c l i p} \left(\hat {\boldsymbol {y}} _ {i} - \operatorname {s o f t m a x} \left(F _ {v} \left(\boldsymbol {x} _ {i}\right)\right), 0, 1\right) \tag {10} +$$ + +where $(\pmb{x}_i,\hat{y}_i)\in \mathcal{D}_r$ $\hat{\pmb{y}}_i$ is the one-hot vector of $\hat{y}_i$ and $\mathrm{clip}(\cdot ,0,1)$ is a function that will restrict its inputs to [0, 1] (i.e. inputs smaller than 0 became 0, inputs larger than 1 became 1, and the rest remain unchanged). + +We use $\mathcal{D}_r' = \{(\pmb{x}_i, \pmb{y}_i') | (\pmb{x}_i, \hat{\pmb{y}}_i) \in \mathcal{D}_r\}$ to indicate the suspicious samples after relabeling and suppression. We + +
AttackBadNets [10]Blend [7]Dynamic [23]WaNet [22]SIG [1]CL-16 [26]CL-32 [26]
MetricBAASRBAASRBAASRBAASRBAASRBAASRBAASR
No Defense94.36%100.00%94.57%99.76%94.66%95.18%94.21%95.84%95.09%64.79%94.98%96.93%95.03%94.73%
FP [19]92.72%1.92%92.49%10.61%92.47%11.73%93.27%0.90%93.55%37.98%92.92%9.74%93.03%6.39%
NAD [18]88.83%1.83%87.72%1.38%88.30%2.32%91.00%0.99%90.42%3.52%88.92%1.64%90.19%2.60%
ABL [17]90.47%0.70%91.31%31.49%83.07%94.21%80.78%19.93%88.61%1.37%90.29%0.54%86.50%2.68%
DBD [14]92.00%3.06%91.62%3.78%90.98%17.56%89.13%0.11%89.42%1.61%92.85%91.08%93.24%14.88%
V&B (Ours)93.96%0.62%94.37%0.63%93.91%1.13%94.15%0.54%94.08%0.17%94.24%1.01%93.98%0.64%
+ +Table 1. The performance of 5 defense methods against 7 backdoor attacks on CIFAR-10. The first 4 attacks are poison-label attacks and the last 3 attacks are clean-label attacks. Note that FP and NAD require local benign samples to work, while ABL, DBD and our method do not. The best results are marked in boldface and underlined items are the second-best results. + +
AttackBadNets [10]Blend [7]WaNet [22]SIG [1]
MetricBAASRBAASRBAASRBAASR
No Defense94.94%99.51%95.16%100.00%92.63%100.00%90.38%89.22%
FP [19]86.28%1.05%87.63%2.03%82.82%26.19%82.50%22.76%
NAD [18]85.83%3.82%86.73%2.90%83.93%3.88%80.71%3.88%
ABL [17]92.15%13.78%87.34%56.57%91.38%56.71%80.13%0.87%
DBD [14]93.08%0.31%93.81%0.87%92.06%1.00%84.47%98.47%
V&B (Ours)95.42%0.28%95.03%0.45%94.84%1.92%94.65%0.03%
+ +Table 2. The performance of 5 defense methods against 4 backdoor attacks on ImageNet subset. + +also suppress the labels of credible samples in the same way to weaken the influence of poisoned samples that may be missed after filtering. As in stage 2, the suppressed credible samples are augmented with our AttentionMix, forming the training set $\hat{\mathcal{D}}_c'$ to further inhibit backdoor injection. We fine-tune the Beneficiary network with the following loss function: + +$$ +\begin{array}{l} \mathcal {L} = \mathbb {E} _ {(\tilde {\boldsymbol {x}}, \tilde {\boldsymbol {y}} ^ {\prime}) \in \tilde {\mathcal {D}} _ {c} ^ {\prime}} [ \mathcal {L} _ {C E} (F _ {b} (\tilde {\boldsymbol {x}}), \tilde {\boldsymbol {y}} ^ {\prime}) ] + \tag {11} \\ \alpha \cdot \mathbb {E} _ {(\boldsymbol {x}, \boldsymbol {y} ^ {\prime}) \in \mathcal {D} _ {r} ^ {\prime}} [ \mathcal {L} _ {C E} (F _ {b} (\boldsymbol {x}), \boldsymbol {y} ^ {\prime}) ] \\ \end{array} +$$ + +Since the Victim network is only trained with suspicious samples, it will confidently predict poisoned samples as the target label but give a wrong prediction for benign samples with high probability. Thus for a poisoned sample, the suppression can reduce the component corresponding to the target class in its label while having little influence on the component of a sample's ground-truth class, as shown in Figure 2(c). + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets and Models. We employ ResNet-18 [13] as our Victim and Beneficiary networks, a standard setting in previous studies [14]. For a fair comparison, we evaluate all defenses on two popular benchmark datasets, including CIFAR-10 and a subset of ImageNet containing 12 classes generated by Li et al. [17]. The details are summarized in Appendix A.1. + +Attack Configures. We consider 6 representative backdoor attacks in our experiments, including four poison-label attacks: BadNets [10], Blend attack [7], WaNet attack [22], + +Dynamic attack [23], and two clean-label attacks: Sinusoidal signal attack (SIG) [1] and Clean-label attack (CL) [26]. To achieve the best attack performance, we configure these attacks as close as possible to their released codes and original settings (e.g. trigger pattern and trigger size). While on the ImageNet subset, we implement these attacks based on the settings suggested by [14], since some attacks miss the corresponding experiments in their original papers. We uniformly set the poisoning ratio as 0.1 and select the first class as the target class in both datasets, i.e. $y_{t} = 0$ . More details about attacks are shown in Appendix A.2. + +Defense Configures and Training Details. We compare our V&B with 4 state-of-the-art defenses: Fine-pruning (FP) [19], Neural Attention Distillation (NAD) [18], Anti-Backdoor Learning (ABL) [17] and Decoupling (DBD) [14], in which FP and NAD require additional local benign samples to work. Similarly, we follow their original settings and released codes. For V&B, we warm up the Victim network for 3 epochs on CIFAR-10 and 6 epochs on the ImageNet subset, where the filtering threshold $t_f$ linearly decreases from 1 to 0.2 and from 1 to 0.4. The selection of these hyperparameters and an automatic warming-up strategy can be found in Appendix A.3. In the following stages, the filtering threshold keeps unchanged (i.e. 0.2 for CIFAR-10 and 0.4 for ImageNet subset). For fast convergence, we adopt the Adam optimizer to train the Victim network with a learning rate of 0.001 in the whole training process. We optimize the Beneficiary network for 200 epochs using an SGD optimizer with a momentum of 0.9 and a weight decay factor of $10^{-4}$ , where the first 100 epochs belong to stage 2 and the last 100 epochs belong to stage 3. The learning rate is initially set to 0.1 and is divided by 10 after every 50 + +
AttackBadNets [10]Blend [7]Dynamic [23]WaNet [22]CL-16 [26]
MetricBAASRBAASRBAASRBAASRBAASR
No Defense94.36%100%94.57%99.76%94.66%95.18%94.21%95.84%94.98%96.93%
Mixup [31]93.37%0.27%92.80%3.49%92.97%42.10%92.97%42.10%94.80%20.03%
Cutmix [30]93.52%0.40%93.40%1.07%91.40%53.23%91.40%53.23%94.54%1.16%
Attentive Cutmix [27]94.26%0.40%94.55%4.62%93.09%13.38%93.09%13.88%95.18%2.97%
AttentionMix (Ours)93.96%0.62%94.37%0.63%93.91%1.13%94.15%0.54%94.24%1.01%
+ +Table 3. The inhibition effect of 4 strong data augmentations against 5 backdoor attacks on CIFAR-10. + +epochs. In particular, the Victim network is optimized every 5 epochs at stage 2, i.e. one credible set is used to train the Beneficiary network for 5 epochs. We use a batch size of 128 for CIFAR-10 and 16 for the ImageNet subset, and set the mixing threshold $t_m$ to 0.1 in all cases. Following semi-supervised learning, we gradually increase the weight of suspicious sample loss in stage 3, simply setting $\alpha$ to 0.1 in the first 50 epochs and 0.5 in the last 50 epochs. + +Evaluation Metrics. We evaluate the performance of defenses with two commonly used metrics: the model's classification accuracy on the benign test set, Benign Accuracy (BA), and the model's classification accuracy on the poison test set, Attack Success Rate (ASR). The lower the ASR and the higher BA, the stronger the defense. In particular, the poison test set removes the samples whose ground-truth label is the target label. + +# 4.2. Effectiveness of Our V&B Defense + +To measure the effectiveness of our proposed V&B method, we compare its performance with four state-of-the-art defenses on CIFAR-10 and the ImageNet subset. The backdoor attack CL-16 and CL-32 are two versions of the Clean-label attack with different magnitudes of adversarial perturbations. + +Table 1 demonstrates the defense results on CIFAR-10, where No Defense is a baseline representing the capability of backdoor attacks. Considering real scenarios, we adopt common data augmentations (i.e. random crop and horizontal flipping) during training, which will slightly affect the attack success rate. It is obvious that our V&B achieves the best results against almost all attacks, even compared to defenses that need additional benign samples. Our method reduces the success rate of all attacks from above $95\%$ (64.79% for SIG) to blow or close to $1\%$ , while maintaining the accuracy on benign samples with only about $1\%$ reduction. For WaNet and CL-16, V&B achieves the second-best performance in reducing the attack success rate, but the benign accuracy of V&B is significantly superior to the corresponding best defense, i.e. DBD and ABL. WaNet adopts a warping field as the trigger and adds a noisy training mode to interfere with detection. CL-16 only poisons the samples whose ground-truth label is the target label and adds adversarial perturbations to them. So it is relatively hard to distinguish the poisoned samples they craft from the benign ones. + +The defense performances on the ImageNet subset are shown in Table 2. Compared with other defenses, our method has the best accuracy on benign samples against all four attacks, even outperforming the baseline due to our AttentionMix data augmentation and semi-supervised suppression training. Semi-supervised learning can relabel poisoned samples to their ground-truth labels, improving benign accuracy. In terms of reducing the attack success rate, our method achieves the best results against BadNets, Blend, and SIG, and the second-best result against WaNet, which is comparable to the best result (1.92% vs. 1.00%). + +We also validate the effectiveness of our method under different poisoning rates in Appendix A.4. Even with a poisoning rate of $50\%$ , our method can still reduce the attack success rate of most attacks to below $1\%$ , while maintaining a satisfactory benign accuracy. + +# 4.3. Effectiveness of AttentionMix + +To illustrate the inhibition effect of our AttentionMix, we replace it with three other popular strong data augmentations in our framework, i.e. Mixup, Cutmix and Attentive Cutmix. As Table 3 shows, each data augmentation performs well against BadNets and Blend, while for more stealthy attacks, such as WaNet and Dynamic, our AttentionMix can better reduce the attack success rate. This is because most of poisoned samples crafted by BadNets and Blend can be filtered out before training the Beneficiary network due to the simple trigger patterns. But for stealthy attacks, the Victim network will only separate part of the poisoned samples after warming up, and the other three augmentations can not effectively inhibit the missed poisoned samples. For a poisoned sample, the poisoned network will focus on its trigger location, so blending the sample's activation region with other images can effectively destroy triggers and help to clean the network. Although Attentive Cutmix can locate the trigger region, direct pasting may leave the trigger intact in the generated image. + +# 4.4. Ablation Studies + +Component Effects. As shown in Table 4, We first train the network without any defense as a baseline and then add our components one by one to investigate their effect. The variants are as follows: $(1) + F$ : filtering the training set and training on credible samples; $(2) + F + AT$ : using AttentionMix when training with credible samples; $(3) + F +$ + +
AttackBadNets [10]Blend [7]WaNet [22]Dynamic [23]CL-16 [26]
MetricBAASRBAASRBAASRBAASRBAASR
No Defense94.36%100.00%94.57%99.76%94.21%95.84%94.66%95.18%94.98%96.93%
+F93.50%10.01%93.47%76.70%92.92%90.97%93.80%93.04%93.63%2.70%
+F + AT92.27%0.42%92.72%1.42%91.58%25.16%93.08%50.01%93.09%2.09%
+F + AT + SST93.88%1.21%94.56%1.52%93.31%1.23%94.32%28.79%94.43%3.29%
+F + AT + SST + S (suspicious)94.11%1.00%94.49%0.73%93.49%0.61%94.61%1.43%94.00%1.06%
ALL93.96%0.62%94.37%0.63%94.15%0.54%93.91%1.13%94.24%1.01%
+ +Table 4. Ablation study of individual components in our framework on CIFAR-10. + +AT + SST: adding standard semi-supervised training [2] on the basis of the second variant; (4) + F + AT + SST + S (suspicious): suppressing suspicious samples during semi-supervised training on the basis of the third variant. + +As can be seen from Table 4, when defending against Blend, WaNet, and Dynamic, only filtering will fail because enough poisoned samples are missed to create backdoors. After using AttentionMix, although the attack success rate of most attacks is severely decreased, the benign accuracy also drops slightly. Thus semi-supervised training is adopted to improve the performance of the network on benign samples. For WaNet and Dynamic, semi-supervised training also reduces their attack success rate by relabeling poisoned samples as ground-truth labels. However, the Beneficiary network is not completely clean, so a small number of poisoned samples may be relabeled as the target label, causing the attack success rate to rebound (e.g. Bad-Nets, Blend and Cl-16). Suppression to suspicious samples can make up for the weakness of relabeling and adding suppression to credible samples will further reduce the attack success rate. We say that the complete framework is more secure when facing backdoor attacks, although its benign accuracy may be lower than other variants. + +To further demonstrate the effect of stage 3, we compare the final results with those after stage 2 in Appendix A.5. In all cases, the attack success rate after stage 2 is higher than the final result, while the benign accuracy after stage 2 is lower than the final result, suggesting that our semi-supervised suppression training can both erasing existing backdoors and improve model performance. + +Selection of Mixing Threshold. As Figure 3 shows, we set the mixing threshold $t_m$ from 0.1 to 0.5 on CIFAR-10 and ImageNet subset to study its impact on model performance. It can be seen that the benign accuracy on CIFAR-10 starts to drop when the mixing threshold is large enough (e.g. 0.2 for BadNets and 0.3 for Blend), while the attack success rate only slightly fluctuates. The decline in benign accuracy is mainly caused by the unlabeled samples in stage 3. Due to the small image size $(32 \times 32$ in CIFAR-10), the activation region will contain only a few clustered pixels or some separate pixels if the mixing threshold is too large. During mixing, these pixels will carry their labels but + +![](images/b76cc2a09db230eca8ca7be2fa9ff413cb89c389c917bde1ffc7ef3e84627789.jpg) +Figure 3. Ablation study of mixing threshold $t_m$ on CIFAR-10 (left) and ImageNet subset (right). + +![](images/4abbd67143dc90a68c66a84737fa16bb6bc3f122a93dc899c8441f3ebaa020d8.jpg) + +can not provide the feature of corresponding classes, which is equivalent to introducing noise to labels. It hinders the learning of our Beneficiary network, leading to more and more suspicious samples being mislabeled in stage 3. This point can be confirmed on the ImageNet subset (image size is $224 \times 224$ ), where the mixing threshold has little impact on our method. + +# 5. Conclusion + +In this paper, we show that poisoned samples could be easily distinguished by the entropy of the poisoned network's prediction. Based on this finding and the fact that strong data augmentation can inhibit backdoor injection, we propose a novel secure training framework consisting of two parallel networks to prevent a backdoor generation from three aspects: (1) filtering training data; (2) inhibiting backdoor injection; (3) erasing potential backdoors. Specifically, we sacrifice one network to exclusively learn backdoor triggers and filter credible samples according to the prediction entropy to train another network. During training, we design a strong data augmentation to inhibit missed poisoned samples from working and then introduced a semi-supervised learning method to erase potential backdoors with the help of the sacrifice network. Extensive experiments verify that our framework is effective against various backdoor attacks under different poisoning rates. + +Acknowledgment. This work is supported in part by the National Natural Science Foundation of China Under Grants No.62176253 and No.U20B2066. + +# References + +[1] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A New Backdoor Attack in CNNS by Training Set Corruption Without Label Poisoning. In 2019 IEEE International Conference on Image Processing, pages 101-105. IEEE, 2019. +[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. MixMatch: A Holistic Approach to Semi-Supervised Learning. Advances in neural information processing systems, 32, 2019. +[3] Eitan Borgnia, Valerii Cherepanova, Liam Fowl, Amin Ghiasi, Jonas Geiping, Micah Goldblum, Tom Goldstein, and Arjun Gupta. Strong Data Augmentation Sanitizes Poisoning and Backdoor Attacks Without an Accuracy Tradeoff. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 3855-3859. IEEE, 2021. +[4] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian Molloy, and Biplav Srivastava. Detecting Backdoor Attacks on Deep Neural Networks by Activation Clustering. arXiv preprint arXiv:1811.03728, 2018. +[5] Huili Chen, Cheng Fu, Jishen Zhao, and Farinaz Koushanfar. ProFlip: Targeted Trojan Attack With Progressive Bit Flips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7718-7727, 2021. +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. +[7] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted Backdoor Attacks on Deep Learning Systems Using Data Poisoning. arXiv preprint arXiv:1712.05526, 2017. +[8] Yanbei Chen, Massimiliano Mancini, Xiatian Zhu, and Zeynep Akata. Semi-Supervised and Unsupervised Deep Visual Learning: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. +[9] Bao Gia Doan, Ehsan Abbasnejad, and Damith C Ranasinghe. Februus: Input Purification Defense Against Trojan Attacks on Deep Neural Network Systems. In Annual Computer Security Applications Conference, pages 897-912, 2020. +[10] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain. arXiv preprint arXiv:1708.06733, 2017. +[11] Chuan Guo, Ruihan Wu, and Kilian Q Weinberger. Trojannet: Embedding Hidden Trojan Horse Models in Neural Networks. arXiv preprint arXiv:2002.10078, 2, 2020. +[12] Jonathan Hayase, Weihao Kong, Raghav Somani, and Sewoong Oh. SPECTRE: Defending Against Backdoor Attacks Using Robust Statistics. arXiv preprint arXiv:2104.11315, 2021. +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. + +[14] Kunzhe Huang, Yiming Li, Baoyuan Wu, Zhan Qin, and Kui Ren. Backdoor Defense via Decoupling the Training Process. In The Tenth International Conference on Learning Representations, 2022. +[15] Junnan Li, Richard Socher, and Steven CH Hoi. DIVIDEMIX: LEARNING WITH NOISY LABELS AS SEMI-SUPERVISED LEARNING. arXiv preprint arXiv:2002.07394, 2020. +[16] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible Backdoor Attack With Sample-Specific Triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021. +[17] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-Backdoor Learning: Training Clean Models on Poisoned Data. Advances in Neural Information Processing Systems, 34:14900–14912, 2021. +[18] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks. In International Conference on Learning Representations, 2021. +[19] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273–294. Springer, 2018. +[20] Xuankai Liu, Fengting Li, Bihan Wen, and Qi Li. Removing Backdoor-Based Watermarks in Neural Networks with Limited Data. In International Conference on Pattern Recognition, pages 10149-10156. IEEE, 2021. +[21] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection Backdoor: A Natural Backdoor Attack on Deep Neural Networks. In European Conference on Computer Vision, pages 182-199. Springer, 2020. +[22] Anh Nguyen and Anh Tran. WaNet-M Imperceptible Warping-based Backdoor Attack. arXiv preprint arXiv:2102.10369, 2021. +[23] Tuan Anh Nguyen and Anh Tran. Input-Aware Dynamic Backdoor Attack. Advances in Neural Information Processing Systems, 33:3454-3464, 2020. +[24] Xiangyu Qi, Tinghao Xie, Ruizhe Pan, Jifeng Zhu, Yong Yang, and Kai Bu. Towards Practical Deployment-Stage Backdoor Attack on Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13347-13357, 2022. +[25] Han Qiu, Yi Zeng, Shangwei Guo, Tianwei Zhang, Meikang Qiu, and Bhavani Thuraisingham. DeepSweep: An Evaluation Framework for Mitigating DNN Backdoor Attacks using Data Augmentation. In Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security, pages 363-377, 2021. +[26] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-Consistent Backdoor Attacks. arXiv preprint arXiv:1912.02771, 2019. +[27] Devesh Walawalkar, Zhiqiang Shen, Zechun Liu, and Marios Savvides. Attentive CutMix: An Enhanced Data Augmentation Approach for Deep Learning Based Image Classification. arXiv preprint arXiv:2003.13048, 2020. + +[28] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural Cleans: Identifying and Mitigating Backdoor Attacks in Neural Networks. In 2019 IEEE Symposium on Security and Privacy, pages 707-723. IEEE, 2019. +[29] Jason Weston, Frédéric Ratle, and Ronan Collobert. Deep Learning via Semi-Supervised Embedding. In Proceedings of the 25th international conference on Machine learning, pages 1168-1175, 2008. +[30] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6023-6032, 2019. +[31] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. Mixup: Beyond Empirical Risk Minimization. In 6th International Conference on Learning Representations, 2018. +[32] Pu Zhao, Pin-Yu Chen, Payel Das, Karthikeyan Natesan Ramamurthy, and Xue Lin. Bridging Mode Connectivity in Loss Landscapes and Adversarial Robustness. 2020. +[33] Zhendong Zhao, Xiaojun Chen, Yuexin Xuan, Ye Dong, Dakui Wang, and Kaitai Liang. DEFEAT: Deep Hidden Feature Backdoor Attacks by Imperceptible Perturbation and Latent Representation Constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15213-15222, 2022. +[34] Nan Zhong, Zhenxing Qian, and Xinpeng Zhang. Imperceptible Backdoor Attack: From Input Space to Feature Representation. arXiv preprint arXiv:2205.03190, 2022. +[35] Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Scholkopf. Learning with Local and Global Consistency. Advances in neural information processing systems, 16, 2003. \ No newline at end of file diff --git a/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/images.zip b/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..efc1b84e6c50264b20ce21ce73e33db0c10f5258 --- /dev/null +++ b/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49e15063e24df7fb88e0dd803209d79e2fa41c8847f9b4ccc58397e5bcd0e830 +size 553165 diff --git a/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/layout.json b/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e2535b3c1e9fb458260d5c000801f11569ac6162 --- /dev/null +++ b/2023/The Victim and The Beneficiary_ Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data/layout.json @@ -0,0 +1,7579 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 102, + 541, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 102, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 541, + 138 + ], + "type": "text", + "content": "The Victim and The Beneficiary: Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 161, + 505, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 161, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 88, + 161, + 505, + 217 + ], + "type": "text", + "content": "Zixuan Zhu, Rui Wang*, Cong Zou, Lihua Jing \nSKLOIS, Institute of Information Engineering, CAS, Beijing, China \nSchool of Cyber Security, University of Chinese Academy of Sciences, Beijing, China \n{zhuzixuan, wangrui, zoucong, jinglihua}@iei.ac.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 270, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 289, + 605 + ], + "type": "text", + "content": "Recently, backdoor attacks have posed a serious security threat to the training process of deep neural networks (DNNs). The attacked model behaves normally on benign samples but outputs a specific result when the trigger is present. However, compared with the rocketing progress of backdoor attacks, existing defenses are difficult to deal with these threats effectively or require benign samples to work, which may be unavailable in real scenarios. In this paper, we find that the poisoned samples and benign samples can be distinguished with prediction entropy. This inspires us to propose a novel dual-network training framework: The Victim and The Beneficiary (V&B), which exploits a poisoned model to train a clean model without extra benign samples. Firstly, we sacrifice the Victim network to be a powerful poisoned sample detector by training on suspicious samples. Secondly, we train the Beneficiary network on the credible samples selected by the Victim to inhibit backdoor injection. Thirdly, a semi-supervised suppression strategy is adopted for erasing potential backdoors and improving model performance. Furthermore, to better inhibit missed poisoned samples, we propose a strong data augmentation method, AttentionMix, which works well with our proposed V&B framework. Extensive experiments on two widely used datasets against 6 state-of-the-art attacks demonstrate that our framework is effective in preventing backdoor injection and robust to various attacks while maintaining the performance on benign samples. Our code is available at https://github.com/Zixuan-Zhu/VaB." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "content": "Large amounts of training data and powerful computing resources are two key factors for the success of deep neural networks (DNNs). While in practical applications, obtaining enough samples for training is laborious, and training is likely to be resource-constrained. As a result, more and" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 329, + 243, + 522, + 392 + ], + "blocks": [ + { + "bbox": [ + 329, + 243, + 522, + 392 + ], + "lines": [ + { + "bbox": [ + 329, + 243, + 522, + 392 + ], + "spans": [ + { + "bbox": [ + 329, + 243, + 522, + 392 + ], + "type": "image", + "image_path": "7272a10b706b0a05e6e2a427edef074ab60c8415e1eb25313c2d3eb57128a5f4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 396, + 547, + 452 + ], + "lines": [ + { + "bbox": [ + 305, + 396, + 547, + 452 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 547, + 452 + ], + "type": "text", + "content": "Figure 1. The average prediction entropy of benign samples versus poisoned samples crafted by 6 backdoor attacks. We conduct the experiment on CIFAR-10 with ResNet-18 under poisoning rate " + }, + { + "bbox": [ + 305, + 396, + 547, + 452 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 305, + 396, + 547, + 452 + ], + "type": "text", + "content": ", where our special training strategy (described in section 3.3) was adopted." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 460, + 546, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 546, + 496 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 546, + 496 + ], + "type": "text", + "content": "more third-party data (e.g. data released on the Internet) and training platforms (e.g. Google Colab) are adopted to reduce the overhead, which brings new security risks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 497, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 547, + 676 + ], + "type": "text", + "content": "Backdoor attacks [5, 11, 24] pose a serious security threat to the training process of DNNs and can be easily executed through data poisoning. Specifically, attackers [7, 10, 22] inject the designed trigger (e.g. a small patch or random noise) to a few benign samples selected from the training set and change their labels to the attacker-defined target label. These poisoned samples will force models to learn the correlation between the trigger and the target label. In the reference process, the attacked model will behave normally on benign samples but output the target label when the trigger is present. Deploying attacked models can lead to severe consequences, even life-threatening in some scenarios (e.g. autonomous driving). Hence, a secure training framework is needed when using third-party data or training platforms." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Many researchers have devoted themselves to this topic and proposed feasible defenses. Li et al. [18] utilize local benign samples to erase existing backdoors in DNNs, and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "155" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 274 + ], + "type": "text", + "content": "Borgnia et al. [3] employ strong data augmentations to inhibit backdoor injection during training. Li et al. [17] find that the training loss decreased faster for poisoned samples than benign samples and designed a loss function to separate them. Finally, they select a fixed portion of samples with the lowest loss to erase potential backdoors. In real scenarios, it is challenging to distinguish poisoned samples from benign ones in this way due to the unknown poisoning rate and close loss values. In experiments, we find the entropy of model prediction is a more discriminative property, and a fixed threshold could filter out most poisoned samples, as shown in Figure 1. During training, poisoned samples will be learned faster than benign samples due to the similar feature of their triggers [17]. Hence, the poisoned network can confidently predict poisoned samples as the target label in early epochs but hesitates about the benign samples." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 277, + 286, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 277, + 286, + 527 + ], + "spans": [ + { + "bbox": [ + 49, + 277, + 286, + 527 + ], + "type": "text", + "content": "Based on this observation, we propose a novel secure training framework, The Victim and The Beneficiary (V&B), that can directly train clean models on poisoned data without resorting to benign samples. Firstly, in warming up stage, the training dataset is recomposed into a suspicious set and a credible set according to the prediction entropy is lower or higher than the predefined threshold. Then the Victim network is trained iteratively with the suspicious set to obtain a strong poisoned sample detector. Secondly, in the clean training stage, the Beneficiary network is trained with credible samples filtered by the Victim to inhibit backdoor injection. In later training epochs, the Victim network is fully exploited to provide its learned knowledge for the Beneficiary network to help the latter erase potential backdoors through semi-supervised learning. To diminish the threat of missed poisoned samples, we design a strong data augmentation AttentionMix, which mixes the image region with high activation values according to the attention map. Compared with existing augmentations [27, 30, 31], it has a stronger inhibition effect against stealthy backdoor attacks [23, 22, 26]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 529, + 286, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 529, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 49, + 529, + 286, + 635 + ], + "type": "text", + "content": "The main contributions of this paper are as follows: (1) We propose a novel dual-network secure training framework, V&B, which can train clean models on the poisoned dataset without resorting to benign samples. (2) We design a powerful data augmentation method called AttentionMix, which has a stronger inhibition effect against stealthy backdoor attacks. (3) Extensive experiments on two popular benchmark datasets demonstrate the effectiveness and robustness of our framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 650, + 132, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 650, + 132, + 661 + ], + "spans": [ + { + "bbox": [ + 49, + 650, + 132, + 661 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 670, + 160, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 160, + 682 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 160, + 682 + ], + "type": "text", + "content": "2.1. Data Augmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 690, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 690, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 690, + 286, + 712 + ], + "type": "text", + "content": "Data Augmentation is a common-used technique to improve the generalization capabilities of DNNs. In addition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 72, + 544, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 544, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 544, + 323 + ], + "type": "text", + "content": "to some popular weak data augmentations (such as cropping, flipping, rotation, etc.), many strong data augmentation techniques [27, 30, 31] have been proposed to further improve model performance. Mixup [31] linearly interpolates any two images and their labels, then trains the model with generated samples. However, the generated samples tend to be unnatural, which will decrease model performances on localization and object detection tasks. Instead of linear interpolation, Cutmix [30] replaces a random patch with the same region from another image and mixes their labels according to the ratio of the patch area. Further, Attentive Cutmix [27] only pastes the influential region located by an attention mechanism to other images, effectively forcing models to learn discriminative details. Yet it also mixes labels based on the ratio of the region area, ignoring the importance of the region in two original images, and direct pasting may bring complete triggers into benign samples. Our AttentionMix takes into account the importance of the region in both images and blends the influential region with the same area in another image, which can destroy the completeness of triggers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 334, + 467, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 334, + 467, + 345 + ], + "spans": [ + { + "bbox": [ + 308, + 334, + 467, + 345 + ], + "type": "text", + "content": "2.2. Backdoor Attack and Defense" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 354, + 544, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 544, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 544, + 388 + ], + "type": "text", + "content": "In this paper, we only consider poisoning-based attacks toward image classification that can occur during the data preparation stage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 400, + 544, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 400, + 544, + 614 + ], + "spans": [ + { + "bbox": [ + 308, + 400, + 544, + 614 + ], + "type": "text", + "content": "Poisoned-based backdoor attack. In early studies, attackers usually adopt simple and visible patterns as the trigger, such as a black-white checkerboard [10]. To escape human inspections, some works use human-imperceptible deformations [22] as the trigger or design unique triggers [16, 23] for each poisoned sample. However, they randomly select benign samples to inject triggers and change their labels (called poison-label attacks), possibly causing the content of an image not to match its label. Instead of randomly selecting samples to poison, Turner et al. [26], Mauro et al. [1], and Liu et al. [21] only inject triggers into benign samples under the target label (called clean-label attacks). Although some triggers are imperceptible in the input space, they can be easily detected by defenses focused on latent space extracted by CNNs. Therefore, recent studies [33, 34] not only ensure the trigger's invisibility in the input space but also restrict the similarity between poison samples and corresponding benign samples in the latent space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 630, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 630, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 630, + 544, + 712 + ], + "type": "text", + "content": "Backdoor Defense. For poisoning-based attacks, defense methods could be divided into three categories from their working stage, including (1) preprocessing before training [9, 12, 25], (2) poison suppression during training [14, 17, 20], and (3) backdoor erasing after training [18, 28, 32]. For preprocessing, Hayase et al. [12] and Chen et al. [4] filter the training set through statistics or clustering and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 750, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 757 + ], + "type": "text", + "content": "156" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "train the model with benign samples. For poison suppressing, Borgnia et al. [3] leverage strong data augmentation (e.g. Mixup, Cutmix) to alleviate poisoned samples' adverse effects. Huang et al. [14] design a decoupling training process, which first learns a feature extractor via SimCLR [6] to prevent poisoned samples from clustering in feature space. For backdoor erasing, Liu et al. [19] prune the low-activation neurons for validation samples (benign samples) and then fine-tune the model with these benign samples. While Li et al. [18] first fine-tune the poisoned model with a small portion of benign samples to obtain a teacher model, then erase the original model's backdoor through knowledge distillation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 236, + 194, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 194, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 194, + 249 + ], + "type": "text", + "content": "2.3. Semi-supervised Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 255, + 289, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 289, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 289, + 376 + ], + "type": "text", + "content": "Semi-supervised learning [8] aims to alleviate the dependence of deep learning on extensive labeled training data and jointly learns the model from a limited amount of labeled samples and a large number of unlabeled samples. Existing methods [2, 29, 35] generally assign each unlabeled sample with a pseudo label and design different loss functions for labeled and unlabeled samples. The weight of unlabeled sample loss will gradually increase during training to reduce the negative effect of low-quality pseudo labels in early epochs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 386, + 104, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 104, + 398 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 104, + 398 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 406, + 133, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 133, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 133, + 418 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": "Pipeline of Poisoned-based Backdoor Attacks. Let " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_B = \\{(x_i, y_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": " denotes the benign training set containing " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": " samples, where " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "x_i \\in \\{0, 1, \\dots, 255\\}^{C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": " is the image, " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "y_i \\in \\{0, 1, \\dots, K\\}" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": " is its label, " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": " is the number of classes, " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": " are the height, width and channels of an image. The attacker will select a subset of benign samples to poison with a designed injection function " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_G" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": " indicates the generated poisoned samples set and " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_R" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": " denotes the remaining benign samples set, where " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_G = \\{(x', y_t) | x' = G(x), (x, y) \\in \\mathcal{D}_B \\setminus \\mathcal{D}_R\\}" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "y_t" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": " is the target label and " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "r = \\frac{\\left|\\mathcal{D}_G\\right|}{\\mathcal{D}_B}" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": " is the poisoning rate. The poisoned dataset " + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_P = \\mathcal{D}_G \\cup \\mathcal{D}_R" + }, + { + "bbox": [ + 46, + 425, + 289, + 606 + ], + "type": "text", + "content": " will be unknowingly adopted by victim users to train DNNs. At inference time, the attacker can manipulate the poisoned model's output through the injected backdoor." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Threat Model and Defense Setting. In this paper, we focus on defending against poisoned-based backdoor attacks toward image classification. The attacker can arbitrarily modify the training set, but can not change other training components (e.g. model structure and hyperparameters) [1, 10, 22]. Such attacks may occur when using third-party datasets or training platforms that can poison users' clean data. We follow the common assumption [12, 14, 17] that" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "defenders have no information about the backdoor (including poisoning rate, trigger pattern, target label, etc.) and any local benign samples to use but have full control over the training process. The defense goal is to prevent the trained DNNs from predicting poisoned samples as the target label while maintaining its accuracy on benign samples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 152, + 373, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 152, + 373, + 164 + ], + "spans": [ + { + "bbox": [ + 306, + 152, + 373, + 164 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 171, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 411 + ], + "type": "text", + "content": "In this subsection, we briefly introduce the pipeline of our framework. As Figure 2(a) illustrates, our framework consists of two networks, and the whole training process is divided into three stages — warming up, clean training and semi-supervised suppression training. In warming up, we aim to obtain a poisoned network that can roughly distinguish poisoned samples from benign ones according to the prediction entropy. Therefore we train the Victim network with only filtered suspicious samples (i.e. suspicious set) except for the first epoch. This training strategy remains unchanged in the following two stages. During clean training, we adopt credible samples (i.e. credible set) to train the Beneficiary network with our proposed AttentionMix data augmentation. The strong data augmentation can effectively prevent missed poisoned samples from creating backdoors. As the Victim network is trained, the entropy of its prediction for benign samples will also slowly decrease below the filtering threshold, leading to a reduction in the credible set size. So one credible set will be used for a few epochs to ensure the Beneficiary network gets enough training samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 411, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 603 + ], + "type": "text", + "content": "To further improve the Beneficiary network performance and erase its potential backdoor, we regard the credible set as a labeled dataset and the suspicious set as an unlabeled dataset to train the Beneficiary network via semi-supervised learning. Following conventional practice [2, 15], each unlabeled sample will be assigned a pseudo label generated by the Beneficiary network, forming the relabeled set. However, if the Beneficiary network still contains a backdoor after stage 2, poisoned samples in the suspicious set will be relabeled as the target label, strengthening the backdoor. After specific training, the Victim network is sensitive to trigger patterns and thus can be utilized to suppress the component corresponding to the target class after relabeling. Furthermore, we also apply this suppression operation to the credible set to further weaken the impact of missed poisoned samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 610, + 390, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 390, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 390, + 624 + ], + "type": "text", + "content": "3.3. Warming Up" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": "At this stage, we aim to obtain a poisoned network that has only learned trigger patterns. In this way, the entropy of its prediction for poisoned samples will be significantly lower than that of benign samples. Then we can use a threshold to filter out poisoned samples." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "content": " denotes the training set and " + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "F_{v}:\\{0,1,\\dots,255\\}^{C\\times H\\times W}\\to [0,1]^{K}" + }, + { + "bbox": [ + 306, + 689, + 545, + 715 + ], + "type": "text", + "content": " indicates the Vic" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "157" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 547, + 231 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 231 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 231 + ], + "type": "image", + "image_path": "666295a13902e0ab4279b53e1e326fc6bb0c5c928d306c2a0752f197d7d820e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 232, + 358, + 241 + ], + "lines": [ + { + "bbox": [ + 213, + 232, + 358, + 241 + ], + "spans": [ + { + "bbox": [ + 213, + 232, + 358, + 241 + ], + "type": "text", + "content": "(a) The pipeline of our training framework" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 243, + 294, + 361 + ], + "blocks": [ + { + "bbox": [ + 50, + 243, + 294, + 361 + ], + "lines": [ + { + "bbox": [ + 50, + 243, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 50, + 243, + 294, + 361 + ], + "type": "image", + "image_path": "eb3f649cbc22b76de502a6ef815778445d1c28d55a780d4a7f0c3cabe4910264.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 373, + 547, + 462 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 547, + 462 + ], + "type": "text", + "content": "Figure 2. (a) The pipeline of our training framework. Net V is the Victim network (i.e. poisoned network) that only learns from the filtered suspicious set, and Net B is the Beneficiary network (i.e. clean network) we wanted. Red triangles represent poisoned samples, and green circles are benign samples. The color of nets indicates their degree of poisoning: the darker the color, the deeper the poisoning, and vice versa. Deep blue areas are three training stages, and light blue areas are basic blocks that will be executed many times in the corresponding stage. (b) An example of AttentionMix. The bottom four images are the non-activation region of the cat image, the activation region of the cat image, the same activation region in the dog image, and the final generated image, respectively. And their labels are located below the images. (c) Suppression for poisoned samples and benign samples. The target label is assumed as the first label. Suppression can reduce the target label component of poisoned samples but have little influence on the ground-truth label of benign samples." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 302, + 243, + 544, + 361 + ], + "blocks": [ + { + "bbox": [ + 302, + 243, + 544, + 361 + ], + "lines": [ + { + "bbox": [ + 302, + 243, + 544, + 361 + ], + "spans": [ + { + "bbox": [ + 302, + 243, + 544, + 361 + ], + "type": "image", + "image_path": "fe32f823de59ee29057c059f09bfd60653437c7e5b3468c23fe139e29c8862b8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 288, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 288, + 534 + ], + "type": "text", + "content": "tim network. We first use the whole training set to train the Victim network for one epoch with cross-entropy loss. Then the prediction entropy of a sample will be calculated with Equation 1, where " + }, + { + "bbox": [ + 46, + 473, + 288, + 534 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 473, + 288, + 534 + ], + "type": "text", + "content": " is the prediction of the Victim network for that sample." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 539, + 287, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 539, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 125, + 539, + 287, + 572 + ], + "type": "interline_equation", + "content": "e = - \\sum_ {k = 1} ^ {K} p _ {k} \\log p _ {k} \\tag {1}", + "image_path": "ec68cd75788060bb9bc1c54028c098f8d184df5a2199d148ec7f94c153b27cb2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 578, + 288, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 288, + 673 + ], + "type": "text", + "content": "Once all samples' prediction entropy is obtained, we normalize them to " + }, + { + "bbox": [ + 47, + 578, + 288, + 673 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 47, + 578, + 288, + 673 + ], + "type": "text", + "content": " through mix-max normalization. Then a threshold " + }, + { + "bbox": [ + 47, + 578, + 288, + 673 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 47, + 578, + 288, + 673 + ], + "type": "text", + "content": " is set to filter out suspicious samples " + }, + { + "bbox": [ + 47, + 578, + 288, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s = \\{(x_i,y_i)|(x_i,y_i)\\in \\mathcal{D},e_i < t_f\\}" + }, + { + "bbox": [ + 47, + 578, + 288, + 673 + ], + "type": "text", + "content": " and the remaining samples are regarded as credible samples " + }, + { + "bbox": [ + 47, + 578, + 288, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 47, + 578, + 288, + 673 + ], + "type": "text", + "content": ". To strengthen the learning of poisoned samples and delay the learning of benign samples, the Victim network will only be trained on suspicious sets in the future. The loss function is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 98, + 685, + 287, + 699 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 685, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 98, + 685, + 287, + 699 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v} = \\mathbb {E} _ {(\\boldsymbol {x}, y) \\in \\mathcal {D} _ {s}} \\left[ \\mathcal {L} _ {C E} \\left(F _ {v} (\\boldsymbol {x}), y\\right) \\right] \\tag {2}", + "image_path": "a265df12f22790fd764cf1ba53bb48cb979d6ef13decd9a21028666463add8a7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 702, + 222, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 222, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 222, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 702, + 222, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 47, + 702, + 222, + 714 + ], + "type": "text", + "content": " denotes the cross-entropy loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 473, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 629 + ], + "type": "text", + "content": "After the first epoch, the prediction entropy of some poisoned samples may still be larger than that of most benign ones, although there is a large gap between their average values in Figure 1. If the filtering threshold " + }, + { + "bbox": [ + 304, + 473, + 545, + 629 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 304, + 473, + 545, + 629 + ], + "type": "text", + "content": " is set too large, lots of benign samples will be discarded, which will affect the model performance on benign samples. So we set a smaller " + }, + { + "bbox": [ + 304, + 473, + 545, + 629 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 304, + 473, + 545, + 629 + ], + "type": "text", + "content": " and wait for the prediction entropy of these poisoned samples to decrease. As training progresses, their prediction entropy will drop below " + }, + { + "bbox": [ + 304, + 473, + 545, + 629 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 304, + 473, + 545, + 629 + ], + "type": "text", + "content": ", making the credible set cleaner. Simultaneously, the prediction entropy of benign samples will also decrease, resulting in a smaller credible set. We expect the Victim network to filter out a clean and relatively large credible set after this stage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 635, + 400, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 400, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 400, + 647 + ], + "type": "text", + "content": "3.4. Clean Training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "At this stage, we begin to train the Beneficiary network with filtered credible samples " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", where the filtering strategy is the same as in stage 1. Because of the rough separation, a few poisoned samples may be included in the credible set. For strong attacks, a small number of poisoned sam" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "158" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "ples is enough to create a backdoor. So we design a strong data augmentation called AttentionMix to inhibit the missed poisoned samples from working. Compared with existing augmentations [27, 30, 31], AttentionMix has a stronger inhibition effect on stealthy backdoor attacks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "type": "text", + "content": "The core idea of AttentionMix is to generate a new training sample " + }, + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "type": "inline_equation", + "content": "(\\tilde{\\boldsymbol{x}},\\tilde{\\boldsymbol{y}})" + }, + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "type": "text", + "content": " by mixing two training samples " + }, + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "type": "inline_equation", + "content": "(x_{1},y_{1})" + }, + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "type": "inline_equation", + "content": "(x_{2},y_{2})" + }, + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "type": "text", + "content": " according to their attention maps, where " + }, + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{y}}" + }, + { + "bbox": [ + 46, + 133, + 287, + 192 + ], + "type": "text", + "content": " is a K-dimensional vector. Following [18], we adopt one sample's final feature maps to compute its attention map:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 200, + 287, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 200, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 135, + 200, + 287, + 229 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} ^ {\\prime} = \\sum_ {i = 1} ^ {c} \\left| F _ {i} \\right| ^ {2} \\tag {3}", + "image_path": "8f795f5014eb975cf36f928c2b8eaff4aaf95280f0f4fd3d8fad044b6d6e4959.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "inline_equation", + "content": "\\pmb{F} \\in \\mathbb{R}^{c \\times h \\times w}" + }, + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "text", + "content": " indicates feature maps, " + }, + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "inline_equation", + "content": "\\pmb{A}' \\in \\mathbb{R}^{h \\times w}" + }, + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "text", + "content": " is the corresponding attention map, " + }, + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "inline_equation", + "content": "c, h" + }, + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "text", + "content": " are the dimensions of channel, height and width. For each attention map, we first bilinearly interpolate it to the original image size " + }, + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "text", + "content": ", and then normalize its elements to " + }, + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 46, + 238, + 287, + 309 + ], + "type": "text", + "content": " by min-max normalization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 311, + 287, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 287, + 334 + ], + "type": "text", + "content": "After getting the attention map, we set a threshold " + }, + { + "bbox": [ + 46, + 311, + 287, + 334 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 46, + 311, + 287, + 334 + ], + "type": "text", + "content": " to obtain the activation region that we want to mix:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 342, + 287, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 342, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 110, + 342, + 287, + 368 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} _ {i, j} = \\left\\{ \\begin{array}{l l} 0 & \\boldsymbol {A} _ {i, j} < t _ {m} \\\\ 1 & \\boldsymbol {A} _ {i, j} \\geq t _ {m} \\end{array} \\right. \\tag {4}", + "image_path": "6c88cce50452ff52cb9c1f58b35c49a44076f1df295cc3bd0a7c763422d71c14.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 375, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 375, + 287, + 411 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 46, + 375, + 287, + 411 + ], + "type": "text", + "content": " is the binary mask indicating which pixels belong to the activation region and " + }, + { + "bbox": [ + 46, + 375, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\pmb{A}" + }, + { + "bbox": [ + 46, + 375, + 287, + 411 + ], + "type": "text", + "content": " is the attention map after interpolation and normalization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "inline_equation", + "content": "A^1" + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "inline_equation", + "content": "A^2" + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "text", + "content": " denote the attention maps of " + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "inline_equation", + "content": "x_2" + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "inline_equation", + "content": "M^1" + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "text", + "content": " denotes the first sample's mask. As Figure 2(b) shows, the mixed image " + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "text", + "content": " is composed of three parts: the non-activation region of " + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "text", + "content": ", the activation region of " + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "text", + "content": " and the same area in " + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "inline_equation", + "content": "x_2" + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "text", + "content": ". During mixing, the non-activation region of " + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "text", + "content": " remains unchanged, and the activation region is blended with the same area of " + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "inline_equation", + "content": "x_2" + }, + { + "bbox": [ + 46, + 411, + 287, + 495 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 83, + 502, + 287, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 502, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 83, + 502, + 287, + 532 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\boldsymbol {x}} = \\left(\\mathbf {1} - \\boldsymbol {M} ^ {1}\\right) \\odot \\boldsymbol {x} _ {1} + \\lambda_ {2} \\cdot \\left(\\boldsymbol {M} ^ {1} \\odot \\boldsymbol {x} _ {1}\\right) \\tag {5} \\\\ + \\lambda_ {3} \\cdot \\left(\\boldsymbol {M} ^ {1} \\odot \\boldsymbol {x} _ {2}\\right) \\\\ \\end{array}", + "image_path": "b6135c069c5e58584304e068a09a4fe1bed2b377f7bd1c97721d0acc9e4dd52e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 540, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 540, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 540, + 287, + 574 + ], + "type": "text", + "content": " is the Hadamard Product, " + }, + { + "bbox": [ + 46, + 540, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 46, + 540, + 287, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 540, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\lambda_{3}" + }, + { + "bbox": [ + 46, + 540, + 287, + 574 + ], + "type": "text", + "content": " are the mixing weights calculated from the two attention maps. The calculation process is as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 80, + 582, + 287, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 582, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 80, + 582, + 287, + 611 + ], + "type": "interline_equation", + "content": "w _ {1} = \\frac {\\operatorname {s u m} \\left(\\boldsymbol {M} ^ {1}\\right)}{\\operatorname {s u m} \\left(\\boldsymbol {A} ^ {1}\\right)}, w _ {2} = \\frac {\\operatorname {s u m} \\left(\\boldsymbol {M} ^ {1} \\odot \\boldsymbol {A} ^ {2}\\right)}{\\operatorname {s u m} \\left(\\boldsymbol {A} ^ {2}\\right)} \\tag {6}", + "image_path": "012cd0642b583a832f80f66704230f5cbec6cebc7121f7cabd000f6be94ef13c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 73, + 619, + 287, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 619, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 73, + 619, + 287, + 642 + ], + "type": "interline_equation", + "content": "\\lambda_ {1} = 1 - w _ {1}, \\lambda_ {2} = \\frac {w _ {1}}{w _ {1} + w _ {2}}, \\lambda_ {3} = \\frac {w _ {2}}{w _ {1} + w _ {2}} \\tag {7}", + "image_path": "ebe049eec1c73d11227ff36f6d3777c2049cd9a92fa8bfc8192faa2a072d593d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "inline_equation", + "content": "\\mathrm{sum}(\\cdot)" + }, + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "text", + "content": " is a summation function, " + }, + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "inline_equation", + "content": "w_{1}" + }, + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "inline_equation", + "content": "w_{2}" + }, + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "text", + "content": " denote the importance of the activation region to the two images, and " + }, + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "text", + "content": " is the importance of the non-activation region to " + }, + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 46, + 645, + 287, + 693 + ], + "type": "text", + "content": ". The final label is also composed of three parts:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 87, + 701, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 701, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 87, + 701, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {y}} = \\lambda_ {1} \\cdot \\boldsymbol {y} _ {1} + w _ {1} \\cdot \\left(\\lambda_ {2} \\cdot \\boldsymbol {y} _ {1} + \\lambda_ {3} \\cdot \\boldsymbol {y} _ {2}\\right) \\tag {8}", + "image_path": "d25cdc859aa1cd716dc09eee8c1c6ccdb5452cee906ceb5ba7bcfa2d561132f9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "\\pmb{y}_1" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "\\pmb{y}_2" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " are the one-hot labels of the two samples. When the activation region of " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "\\pmb{x}_1" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " corresponds to the background of " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "\\pmb{x}_2" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "w_1" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " will be significantly larger than " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "w_2" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " will be a small value, which can reduce background interference during mixing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 132, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 132, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 132, + 545, + 168 + ], + "type": "text", + "content": "We denote the generated new credible set as " + }, + { + "bbox": [ + 305, + 132, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{D}}_c" + }, + { + "bbox": [ + 305, + 132, + 545, + 168 + ], + "type": "text", + "content": " and train the Beneficiary network " + }, + { + "bbox": [ + 305, + 132, + 545, + 168 + ], + "type": "inline_equation", + "content": "F_{b}" + }, + { + "bbox": [ + 305, + 132, + 545, + 168 + ], + "type": "text", + "content": " with the following loss function (i.e. Train with AttentionMix in Figure 2(a)):" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 357, + 178, + 545, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 178, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 357, + 178, + 545, + 192 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b} = \\mathbb {E} _ {(\\tilde {\\boldsymbol {x}}, \\tilde {\\boldsymbol {y}}) \\in \\tilde {\\mathcal {D}} _ {c}} [ \\mathcal {L} _ {C E} (F _ {b} (\\tilde {\\boldsymbol {x}}), \\tilde {\\boldsymbol {y}}) ] \\tag {9}", + "image_path": "33b01d9dca5b4c0fbb73750a9360f16b7f3b6307c1d9b33693e3700231766e3c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 200, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 200, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 545, + 272 + ], + "type": "text", + "content": "For a poisoned network, the activation region of poisoned samples is more likely their trigger locations. Mixing the activation region can destroy the completeness of trigger patterns. While for benign samples, the mixing plays a regularization role that prevents the network from overfitting a particular subject and forces it to learn abundant features." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 280, + 508, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 280, + 508, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 280, + 508, + 293 + ], + "type": "text", + "content": "3.5. Semi-supervised Suppression Training" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 298, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 545, + 417 + ], + "type": "text", + "content": "Until now, we have got a poisoned Victim network that can filter out most of the poisoned samples and a Beneficiary network that is relatively clean but may still contain the backdoor. In the previous stage, the Beneficiary network is trained only with partial training samples (i.e. the credible set), which may harm its performance on benign samples. Thus we intend to enable the deprecated suspicious samples via semi-supervised learning at this stage to improve the Beneficiary network while erasing its potential backdoor." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "text", + "content": "Given a credible set " + }, + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "text", + "content": " and a suspicious set " + }, + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "text", + "content": ", we use samples in " + }, + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "text", + "content": " as labeled samples and remove the labels of samples in " + }, + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "text", + "content": " to transform the problem into a semi-supervised learning setting. For each unlabeled sample, we average the predictions of the Beneficiary network over its multiple augmentations (e.g. random crop and random flip) and select the class with max probability as its pseudo-label, which is similar to [2] and [15]. The relabeled samples are denoted as " + }, + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r = \\{(\\pmb{x}_i,\\hat{\\pmb{y}}_i)|(\\pmb{x}_i,\\pmb{y}_i)\\in \\mathcal{D}_s\\}" + }, + { + "bbox": [ + 304, + 418, + 545, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "type": "text", + "content": "However, the poisoned samples in " + }, + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "type": "text", + "content": " may be relabeled as the target label, i.e. " + }, + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\hat{y}_i = y_t" + }, + { + "bbox": [ + 304, + 526, + 545, + 609 + ], + "type": "text", + "content": ", if the Beneficiary network still contains the backdoor. We can mitigate the impact of such mislabeling by exploiting the knowledge learned by the Victim network. Specifically, we adopt the Victim network's prediction to suppress the pseudo-label (i.e. Suppress in Figure 2(a)):" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 345, + 619, + 545, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 619, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 345, + 619, + 545, + 632 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {i} ^ {\\prime} = \\operatorname {c l i p} \\left(\\hat {\\boldsymbol {y}} _ {i} - \\operatorname {s o f t m a x} \\left(F _ {v} \\left(\\boldsymbol {x} _ {i}\\right)\\right), 0, 1\\right) \\tag {10}", + "image_path": "5fb3ac42d8a8198110e6d47bf4d822d32d0053753149c2fb246d420f3a8650cd.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_i,\\hat{y}_i)\\in \\mathcal{D}_r" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_i" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": " is the one-hot vector of " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{clip}(\\cdot ,0,1)" + }, + { + "bbox": [ + 304, + 641, + 545, + 689 + ], + "type": "text", + "content": " is a function that will restrict its inputs to [0, 1] (i.e. inputs smaller than 0 became 0, inputs larger than 1 became 1, and the rest remain unchanged)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r' = \\{(\\pmb{x}_i, \\pmb{y}_i') | (\\pmb{x}_i, \\hat{\\pmb{y}}_i) \\in \\mathcal{D}_r\\}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " to indicate the suspicious samples after relabeling and suppression. We" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "159" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 160 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 160 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 160 + ], + "type": "table", + "html": "
AttackBadNets [10]Blend [7]Dynamic [23]WaNet [22]SIG [1]CL-16 [26]CL-32 [26]
MetricBAASRBAASRBAASRBAASRBAASRBAASRBAASR
No Defense94.36%100.00%94.57%99.76%94.66%95.18%94.21%95.84%95.09%64.79%94.98%96.93%95.03%94.73%
FP [19]92.72%1.92%92.49%10.61%92.47%11.73%93.27%0.90%93.55%37.98%92.92%9.74%93.03%6.39%
NAD [18]88.83%1.83%87.72%1.38%88.30%2.32%91.00%0.99%90.42%3.52%88.92%1.64%90.19%2.60%
ABL [17]90.47%0.70%91.31%31.49%83.07%94.21%80.78%19.93%88.61%1.37%90.29%0.54%86.50%2.68%
DBD [14]92.00%3.06%91.62%3.78%90.98%17.56%89.13%0.11%89.42%1.61%92.85%91.08%93.24%14.88%
V&B (Ours)93.96%0.62%94.37%0.63%93.91%1.13%94.15%0.54%94.08%0.17%94.24%1.01%93.98%0.64%
", + "image_path": "1323a94e8b7af459a43b8cbcd3a11f911c7527ef6f35d7482dce7a62212a1729.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 149, + 209, + 445, + 294 + ], + "blocks": [ + { + "bbox": [ + 46, + 167, + 546, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 546, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 546, + 201 + ], + "type": "text", + "content": "Table 1. The performance of 5 defense methods against 7 backdoor attacks on CIFAR-10. The first 4 attacks are poison-label attacks and the last 3 attacks are clean-label attacks. Note that FP and NAD require local benign samples to work, while ABL, DBD and our method do not. The best results are marked in boldface and underlined items are the second-best results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 149, + 209, + 445, + 294 + ], + "lines": [ + { + "bbox": [ + 149, + 209, + 445, + 294 + ], + "spans": [ + { + "bbox": [ + 149, + 209, + 445, + 294 + ], + "type": "table", + "html": "
AttackBadNets [10]Blend [7]WaNet [22]SIG [1]
MetricBAASRBAASRBAASRBAASR
No Defense94.94%99.51%95.16%100.00%92.63%100.00%90.38%89.22%
FP [19]86.28%1.05%87.63%2.03%82.82%26.19%82.50%22.76%
NAD [18]85.83%3.82%86.73%2.90%83.93%3.88%80.71%3.88%
ABL [17]92.15%13.78%87.34%56.57%91.38%56.71%80.13%0.87%
DBD [14]93.08%0.31%93.81%0.87%92.06%1.00%84.47%98.47%
V&B (Ours)95.42%0.28%95.03%0.45%94.84%1.92%94.65%0.03%
", + "image_path": "88ea00256f9633bfe7e2df3cbade1ad5d0e305477a66afc558b69326b1b2a51f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 301, + 470, + 313 + ], + "lines": [ + { + "bbox": [ + 122, + 301, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 122, + 301, + 470, + 313 + ], + "type": "text", + "content": "Table 2. The performance of 5 defense methods against 4 backdoor attacks on ImageNet subset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 321, + 288, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 288, + 405 + ], + "type": "text", + "content": "also suppress the labels of credible samples in the same way to weaken the influence of poisoned samples that may be missed after filtering. As in stage 2, the suppressed credible samples are augmented with our AttentionMix, forming the training set " + }, + { + "bbox": [ + 46, + 321, + 288, + 405 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}_c'" + }, + { + "bbox": [ + 46, + 321, + 288, + 405 + ], + "type": "text", + "content": " to further inhibit backdoor injection. We fine-tune the Beneficiary network with the following loss function:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 408, + 287, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 408, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 93, + 408, + 287, + 441 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathbb {E} _ {(\\tilde {\\boldsymbol {x}}, \\tilde {\\boldsymbol {y}} ^ {\\prime}) \\in \\tilde {\\mathcal {D}} _ {c} ^ {\\prime}} [ \\mathcal {L} _ {C E} (F _ {b} (\\tilde {\\boldsymbol {x}}), \\tilde {\\boldsymbol {y}} ^ {\\prime}) ] + \\tag {11} \\\\ \\alpha \\cdot \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y} ^ {\\prime}) \\in \\mathcal {D} _ {r} ^ {\\prime}} [ \\mathcal {L} _ {C E} (F _ {b} (\\boldsymbol {x}), \\boldsymbol {y} ^ {\\prime}) ] \\\\ \\end{array}", + "image_path": "2a16f73ed71e202a507efd6c719d7c6571b60da7ca1f9d6f0dc2b3e7aeefaf62.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 445, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 540 + ], + "type": "text", + "content": "Since the Victim network is only trained with suspicious samples, it will confidently predict poisoned samples as the target label but give a wrong prediction for benign samples with high probability. Thus for a poisoned sample, the suppression can reduce the component corresponding to the target class in its label while having little influence on the component of a sample's ground-truth class, as shown in Figure 2(c)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 551, + 128, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 128, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 128, + 563 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 570, + 174, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 174, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 174, + 582 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 588, + 287, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 287, + 672 + ], + "type": "text", + "content": "Datasets and Models. We employ ResNet-18 [13] as our Victim and Beneficiary networks, a standard setting in previous studies [14]. For a fair comparison, we evaluate all defenses on two popular benchmark datasets, including CIFAR-10 and a subset of ImageNet containing 12 classes generated by Li et al. [17]. The details are summarized in Appendix A.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Attack Configures. We consider 6 representative backdoor attacks in our experiments, including four poison-label attacks: BadNets [10], Blend attack [7], WaNet attack [22]," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 321, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 453 + ], + "type": "text", + "content": "Dynamic attack [23], and two clean-label attacks: Sinusoidal signal attack (SIG) [1] and Clean-label attack (CL) [26]. To achieve the best attack performance, we configure these attacks as close as possible to their released codes and original settings (e.g. trigger pattern and trigger size). While on the ImageNet subset, we implement these attacks based on the settings suggested by [14], since some attacks miss the corresponding experiments in their original papers. We uniformly set the poisoning ratio as 0.1 and select the first class as the target class in both datasets, i.e. " + }, + { + "bbox": [ + 304, + 321, + 545, + 453 + ], + "type": "inline_equation", + "content": "y_{t} = 0" + }, + { + "bbox": [ + 304, + 321, + 545, + 453 + ], + "type": "text", + "content": ". More details about attacks are shown in Appendix A.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 715 + ], + "type": "text", + "content": "Defense Configures and Training Details. We compare our V&B with 4 state-of-the-art defenses: Fine-pruning (FP) [19], Neural Attention Distillation (NAD) [18], Anti-Backdoor Learning (ABL) [17] and Decoupling (DBD) [14], in which FP and NAD require additional local benign samples to work. Similarly, we follow their original settings and released codes. For V&B, we warm up the Victim network for 3 epochs on CIFAR-10 and 6 epochs on the ImageNet subset, where the filtering threshold " + }, + { + "bbox": [ + 304, + 462, + 546, + 715 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 304, + 462, + 546, + 715 + ], + "type": "text", + "content": " linearly decreases from 1 to 0.2 and from 1 to 0.4. The selection of these hyperparameters and an automatic warming-up strategy can be found in Appendix A.3. In the following stages, the filtering threshold keeps unchanged (i.e. 0.2 for CIFAR-10 and 0.4 for ImageNet subset). For fast convergence, we adopt the Adam optimizer to train the Victim network with a learning rate of 0.001 in the whole training process. We optimize the Beneficiary network for 200 epochs using an SGD optimizer with a momentum of 0.9 and a weight decay factor of " + }, + { + "bbox": [ + 304, + 462, + 546, + 715 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 462, + 546, + 715 + ], + "type": "text", + "content": ", where the first 100 epochs belong to stage 2 and the last 100 epochs belong to stage 3. The learning rate is initially set to 0.1 and is divided by 10 after every 50" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "160" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 116, + 70, + 479, + 143 + ], + "blocks": [ + { + "bbox": [ + 116, + 70, + 479, + 143 + ], + "lines": [ + { + "bbox": [ + 116, + 70, + 479, + 143 + ], + "spans": [ + { + "bbox": [ + 116, + 70, + 479, + 143 + ], + "type": "table", + "html": "
AttackBadNets [10]Blend [7]Dynamic [23]WaNet [22]CL-16 [26]
MetricBAASRBAASRBAASRBAASRBAASR
No Defense94.36%100%94.57%99.76%94.66%95.18%94.21%95.84%94.98%96.93%
Mixup [31]93.37%0.27%92.80%3.49%92.97%42.10%92.97%42.10%94.80%20.03%
Cutmix [30]93.52%0.40%93.40%1.07%91.40%53.23%91.40%53.23%94.54%1.16%
Attentive Cutmix [27]94.26%0.40%94.55%4.62%93.09%13.38%93.09%13.88%95.18%2.97%
AttentionMix (Ours)93.96%0.62%94.37%0.63%93.91%1.13%94.15%0.54%94.24%1.01%
", + "image_path": "8a7d21e9e99b848c7786949fbd852c363a99505175a1332ca75b96de8fe7beef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 150, + 480, + 161 + ], + "lines": [ + { + "bbox": [ + 110, + 150, + 480, + 161 + ], + "spans": [ + { + "bbox": [ + 110, + 150, + 480, + 161 + ], + "type": "text", + "content": "Table 3. The inhibition effect of 4 strong data augmentations against 5 backdoor attacks on CIFAR-10." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 170, + 288, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 288, + 266 + ], + "type": "text", + "content": "epochs. In particular, the Victim network is optimized every 5 epochs at stage 2, i.e. one credible set is used to train the Beneficiary network for 5 epochs. We use a batch size of 128 for CIFAR-10 and 16 for the ImageNet subset, and set the mixing threshold " + }, + { + "bbox": [ + 46, + 170, + 288, + 266 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 46, + 170, + 288, + 266 + ], + "type": "text", + "content": " to 0.1 in all cases. Following semi-supervised learning, we gradually increase the weight of suspicious sample loss in stage 3, simply setting " + }, + { + "bbox": [ + 46, + 170, + 288, + 266 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 170, + 288, + 266 + ], + "type": "text", + "content": " to 0.1 in the first 50 epochs and 0.5 in the last 50 epochs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 270, + 287, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 366 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate the performance of defenses with two commonly used metrics: the model's classification accuracy on the benign test set, Benign Accuracy (BA), and the model's classification accuracy on the poison test set, Attack Success Rate (ASR). The lower the ASR and the higher BA, the stronger the defense. In particular, the poison test set removes the samples whose ground-truth label is the target label." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 372, + 231, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 231, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 231, + 383 + ], + "type": "text", + "content": "4.2. Effectiveness of Our V&B Defense" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 390, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 461 + ], + "type": "text", + "content": "To measure the effectiveness of our proposed V&B method, we compare its performance with four state-of-the-art defenses on CIFAR-10 and the ImageNet subset. The backdoor attack CL-16 and CL-32 are two versions of the Clean-label attack with different magnitudes of adversarial perturbations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": "Table 1 demonstrates the defense results on CIFAR-10, where No Defense is a baseline representing the capability of backdoor attacks. Considering real scenarios, we adopt common data augmentations (i.e. random crop and horizontal flipping) during training, which will slightly affect the attack success rate. It is obvious that our V&B achieves the best results against almost all attacks, even compared to defenses that need additional benign samples. Our method reduces the success rate of all attacks from above " + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": " (64.79% for SIG) to blow or close to " + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": ", while maintaining the accuracy on benign samples with only about " + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 462, + 288, + 713 + ], + "type": "text", + "content": " reduction. For WaNet and CL-16, V&B achieves the second-best performance in reducing the attack success rate, but the benign accuracy of V&B is significantly superior to the corresponding best defense, i.e. DBD and ABL. WaNet adopts a warping field as the trigger and adds a noisy training mode to interfere with detection. CL-16 only poisons the samples whose ground-truth label is the target label and adds adversarial perturbations to them. So it is relatively hard to distinguish the poisoned samples they craft from the benign ones." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 170, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 545, + 301 + ], + "type": "text", + "content": "The defense performances on the ImageNet subset are shown in Table 2. Compared with other defenses, our method has the best accuracy on benign samples against all four attacks, even outperforming the baseline due to our AttentionMix data augmentation and semi-supervised suppression training. Semi-supervised learning can relabel poisoned samples to their ground-truth labels, improving benign accuracy. In terms of reducing the attack success rate, our method achieves the best results against BadNets, Blend, and SIG, and the second-best result against WaNet, which is comparable to the best result (1.92% vs. 1.00%)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "text", + "content": "We also validate the effectiveness of our method under different poisoning rates in Appendix A.4. Even with a poisoning rate of " + }, + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "text", + "content": ", our method can still reduce the attack success rate of most attacks to below " + }, + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "text", + "content": ", while maintaining a satisfactory benign accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 369, + 468, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 468, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 468, + 380 + ], + "type": "text", + "content": "4.3. Effectiveness of AttentionMix" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 388, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 615 + ], + "type": "text", + "content": "To illustrate the inhibition effect of our AttentionMix, we replace it with three other popular strong data augmentations in our framework, i.e. Mixup, Cutmix and Attentive Cutmix. As Table 3 shows, each data augmentation performs well against BadNets and Blend, while for more stealthy attacks, such as WaNet and Dynamic, our AttentionMix can better reduce the attack success rate. This is because most of poisoned samples crafted by BadNets and Blend can be filtered out before training the Beneficiary network due to the simple trigger patterns. But for stealthy attacks, the Victim network will only separate part of the poisoned samples after warming up, and the other three augmentations can not effectively inhibit the missed poisoned samples. For a poisoned sample, the poisoned network will focus on its trigger location, so blending the sample's activation region with other images can effectively destroy triggers and help to clean the network. Although Attentive Cutmix can locate the trigger region, direct pasting may leave the trigger intact in the generated image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 623, + 406, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 406, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 406, + 635 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Component Effects. As shown in Table 4, We first train the network without any defense as a baseline and then add our components one by one to investigate their effect. The variants are as follows: " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "(1) + F" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": ": filtering the training set and training on credible samples; " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "(2) + F + AT" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": ": using AttentionMix when training with credible samples; " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "(3) + F +" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "161" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 547, + 162 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 547, + 162 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 547, + 162 + ], + "type": "table", + "html": "
AttackBadNets [10]Blend [7]WaNet [22]Dynamic [23]CL-16 [26]
MetricBAASRBAASRBAASRBAASRBAASR
No Defense94.36%100.00%94.57%99.76%94.21%95.84%94.66%95.18%94.98%96.93%
+F93.50%10.01%93.47%76.70%92.92%90.97%93.80%93.04%93.63%2.70%
+F + AT92.27%0.42%92.72%1.42%91.58%25.16%93.08%50.01%93.09%2.09%
+F + AT + SST93.88%1.21%94.56%1.52%93.31%1.23%94.32%28.79%94.43%3.29%
+F + AT + SST + S (suspicious)94.11%1.00%94.49%0.73%93.49%0.61%94.61%1.43%94.00%1.06%
ALL93.96%0.62%94.37%0.63%94.15%0.54%93.91%1.13%94.24%1.01%
", + "image_path": "c7e29e7de4ddb338693d2ab6893a156e7152990b777318c93b0c0895e0d99345.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 170, + 444, + 180 + ], + "lines": [ + { + "bbox": [ + 145, + 170, + 444, + 180 + ], + "spans": [ + { + "bbox": [ + 145, + 170, + 444, + 180 + ], + "type": "text", + "content": "Table 4. Ablation study of individual components in our framework on CIFAR-10." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 201, + 287, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 248 + ], + "type": "text", + "content": "AT + SST: adding standard semi-supervised training [2] on the basis of the second variant; (4) + F + AT + SST + S (suspicious): suppressing suspicious samples during semi-supervised training on the basis of the third variant." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 249, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 287, + 475 + ], + "type": "text", + "content": "As can be seen from Table 4, when defending against Blend, WaNet, and Dynamic, only filtering will fail because enough poisoned samples are missed to create backdoors. After using AttentionMix, although the attack success rate of most attacks is severely decreased, the benign accuracy also drops slightly. Thus semi-supervised training is adopted to improve the performance of the network on benign samples. For WaNet and Dynamic, semi-supervised training also reduces their attack success rate by relabeling poisoned samples as ground-truth labels. However, the Beneficiary network is not completely clean, so a small number of poisoned samples may be relabeled as the target label, causing the attack success rate to rebound (e.g. Bad-Nets, Blend and Cl-16). Suppression to suspicious samples can make up for the weakness of relabeling and adding suppression to credible samples will further reduce the attack success rate. We say that the complete framework is more secure when facing backdoor attacks, although its benign accuracy may be lower than other variants." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 476, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 287, + 559 + ], + "type": "text", + "content": "To further demonstrate the effect of stage 3, we compare the final results with those after stage 2 in Appendix A.5. In all cases, the attack success rate after stage 2 is higher than the final result, while the benign accuracy after stage 2 is lower than the final result, suggesting that our semi-supervised suppression training can both erasing existing backdoors and improve model performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Selection of Mixing Threshold. As Figure 3 shows, we set the mixing threshold " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " from 0.1 to 0.5 on CIFAR-10 and ImageNet subset to study its impact on model performance. It can be seen that the benign accuracy on CIFAR-10 starts to drop when the mixing threshold is large enough (e.g. 0.2 for BadNets and 0.3 for Blend), while the attack success rate only slightly fluctuates. The decline in benign accuracy is mainly caused by the unlabeled samples in stage 3. Due to the small image size " + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "inline_equation", + "content": "(32 \\times 32" + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": " in CIFAR-10), the activation region will contain only a few clustered pixels or some separate pixels if the mixing threshold is too large. During mixing, these pixels will carry their labels but" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 198, + 427, + 291 + ], + "blocks": [ + { + "bbox": [ + 307, + 198, + 427, + 291 + ], + "lines": [ + { + "bbox": [ + 307, + 198, + 427, + 291 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 427, + 291 + ], + "type": "image", + "image_path": "b76cc2a09db230eca8ca7be2fa9ff413cb89c389c917bde1ffc7ef3e84627789.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 303, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 306, + 303, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 306, + 303, + 545, + 326 + ], + "type": "text", + "content": "Figure 3. Ablation study of mixing threshold " + }, + { + "bbox": [ + 306, + 303, + 545, + 326 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 306, + 303, + 545, + 326 + ], + "type": "text", + "content": " on CIFAR-10 (left) and ImageNet subset (right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 198, + 545, + 291 + ], + "blocks": [ + { + "bbox": [ + 427, + 198, + 545, + 291 + ], + "lines": [ + { + "bbox": [ + 427, + 198, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 427, + 198, + 545, + 291 + ], + "type": "image", + "image_path": "4abbd67143dc90a68c66a84737fa16bb6bc3f122a93dc899c8441f3ebaa020d8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 346, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 545, + 429 + ], + "type": "text", + "content": "can not provide the feature of corresponding classes, which is equivalent to introducing noise to labels. It hinders the learning of our Beneficiary network, leading to more and more suspicious samples being mislabeled in stage 3. This point can be confirmed on the ImageNet subset (image size is " + }, + { + "bbox": [ + 305, + 346, + 545, + 429 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 305, + 346, + 545, + 429 + ], + "type": "text", + "content": "), where the mixing threshold has little impact on our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 664 + ], + "type": "text", + "content": "In this paper, we show that poisoned samples could be easily distinguished by the entropy of the poisoned network's prediction. Based on this finding and the fact that strong data augmentation can inhibit backdoor injection, we propose a novel secure training framework consisting of two parallel networks to prevent a backdoor generation from three aspects: (1) filtering training data; (2) inhibiting backdoor injection; (3) erasing potential backdoors. Specifically, we sacrifice one network to exclusively learn backdoor triggers and filter credible samples according to the prediction entropy to train another network. During training, we design a strong data augmentation to inhibit missed poisoned samples from working and then introduced a semi-supervised learning method to erase potential backdoors with the help of the sacrifice network. Extensive experiments verify that our framework is effective against various backdoor attacks under different poisoning rates." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 673, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 673, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 305, + 673, + 545, + 708 + ], + "type": "text", + "content": "Acknowledgment. This work is supported in part by the National Natural Science Foundation of China Under Grants No.62176253 and No.U20B2066." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "162" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Mauro Barni, Kassem Kallas, and Benedetta Tondi. A New Backdoor Attack in CNNS by Training Set Corruption Without Label Poisoning. In 2019 IEEE International Conference on Image Processing, pages 101-105. IEEE, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "text", + "content": "[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. MixMatch: A Holistic Approach to Semi-Supervised Learning. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 248 + ], + "type": "text", + "content": "[3] Eitan Borgnia, Valerii Cherepanova, Liam Fowl, Amin Ghiasi, Jonas Geiping, Micah Goldblum, Tom Goldstein, and Arjun Gupta. Strong Data Augmentation Sanitizes Poisoning and Backdoor Attacks Without an Accuracy Tradeoff. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 3855-3859. IEEE, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 304 + ], + "type": "text", + "content": "[4] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian Molloy, and Biplav Srivastava. Detecting Backdoor Attacks on Deep Neural Networks by Activation Clustering. arXiv preprint arXiv:1811.03728, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 350 + ], + "type": "text", + "content": "[5] Huili Chen, Cheng Fu, Jishen Zhao, and Farinaz Koushanfar. ProFlip: Targeted Trojan Attack With Progressive Bit Flips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7718-7727, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 396 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 440 + ], + "type": "text", + "content": "[7] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted Backdoor Attacks on Deep Learning Systems Using Data Poisoning. arXiv preprint arXiv:1712.05526, 2017." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 485 + ], + "type": "text", + "content": "[8] Yanbei Chen, Massimiliano Mancini, Xiatian Zhu, and Zeynep Akata. Semi-Supervised and Unsupervised Deep Visual Learning: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 541 + ], + "type": "text", + "content": "[9] Bao Gia Doan, Ehsan Abbasnejad, and Damith C Ranasinghe. Februus: Input Purification Defense Against Trojan Attacks on Deep Neural Network Systems. In Annual Computer Security Applications Conference, pages 897-912, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 586 + ], + "type": "text", + "content": "[10] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain. arXiv preprint arXiv:1708.06733, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "text", + "content": "[11] Chuan Guo, Ruihan Wu, and Kilian Q Weinberger. Trojannet: Embedding Hidden Trojan Horse Models in Neural Networks. arXiv preprint arXiv:2002.10078, 2, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[12] Jonathan Hayase, Weihao Kong, Raghav Somani, and Sewoong Oh. SPECTRE: Defending Against Backdoor Attacks Using Robust Statistics. arXiv preprint arXiv:2104.11315, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Kunzhe Huang, Yiming Li, Baoyuan Wu, Zhan Qin, and Kui Ren. Backdoor Defense via Decoupling the Training Process. In The Tenth International Conference on Learning Representations, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[15] Junnan Li, Richard Socher, and Steven CH Hoi. DIVIDEMIX: LEARNING WITH NOISY LABELS AS SEMI-SUPERVISED LEARNING. arXiv preprint arXiv:2002.07394, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "text", + "content": "[16] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible Backdoor Attack With Sample-Specific Triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "text", + "content": "[17] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-Backdoor Learning: Training Clean Models on Poisoned Data. Advances in Neural Information Processing Systems, 34:14900–14912, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[18] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 305, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 360 + ], + "type": "text", + "content": "[19] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273–294. Springer, 2018." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 403 + ], + "type": "text", + "content": "[20] Xuankai Liu, Fengting Li, Bihan Wen, and Qi Li. Removing Backdoor-Based Watermarks in Neural Networks with Limited Data. In International Conference on Pattern Recognition, pages 10149-10156. IEEE, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "text", + "content": "[21] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection Backdoor: A Natural Backdoor Attack on Deep Neural Networks. In European Conference on Computer Vision, pages 182-199. Springer, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "text", + "content": "[22] Anh Nguyen and Anh Tran. WaNet-M Imperceptible Warping-based Backdoor Attack. arXiv preprint arXiv:2102.10369, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[23] Tuan Anh Nguyen and Anh Tran. Input-Aware Dynamic Backdoor Attack. Advances in Neural Information Processing Systems, 33:3454-3464, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 569 + ], + "type": "text", + "content": "[24] Xiangyu Qi, Tinghao Xie, Ruizhe Pan, Jifeng Zhu, Yong Yang, and Kai Bu. Towards Practical Deployment-Stage Backdoor Attack on Deep Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13347-13357, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 570, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 635 + ], + "type": "text", + "content": "[25] Han Qiu, Yi Zeng, Shangwei Guo, Tianwei Zhang, Meikang Qiu, and Bhavani Thuraisingham. DeepSweep: An Evaluation Framework for Mitigating DNN Backdoor Attacks using Data Augmentation. In Proceedings of the 2021 ACM Asia Conference on Computer and Communications Security, pages 363-377, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 636, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 668 + ], + "type": "text", + "content": "[26] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-Consistent Backdoor Attacks. arXiv preprint arXiv:1912.02771, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[27] Devesh Walawalkar, Zhiqiang Shen, Zechun Liu, and Marios Savvides. Attentive CutMix: An Enhanced Data Augmentation Approach for Deep Learning Based Image Classification. arXiv preprint arXiv:2003.13048, 2020." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "163" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 453 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural Cleans: Identifying and Mitigating Backdoor Attacks in Neural Networks. In 2019 IEEE Symposium on Security and Privacy, pages 707-723. IEEE, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[29] Jason Weston, Frédéric Ratle, and Ronan Collobert. Deep Learning via Semi-Supervised Embedding. In Proceedings of the 25th international conference on Machine learning, pages 1168-1175, 2008." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "text", + "content": "[30] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6023-6032, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 273 + ], + "type": "text", + "content": "[31] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. Mixup: Beyond Empirical Risk Minimization. In 6th International Conference on Learning Representations, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 307 + ], + "type": "text", + "content": "[32] Pu Zhao, Pin-Yu Chen, Payel Das, Karthikeyan Natesan Ramamurthy, and Xue Lin. Bridging Mode Connectivity in Loss Landscapes and Adversarial Robustness. 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 373 + ], + "type": "text", + "content": "[33] Zhendong Zhao, Xiaojun Chen, Yuexin Xuan, Ye Dong, Dakui Wang, and Kaitai Liang. DEFEAT: Deep Hidden Feature Backdoor Attacks by Imperceptible Perturbation and Latent Representation Constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15213-15222, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 375, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 375, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 287, + 407 + ], + "type": "text", + "content": "[34] Nan Zhong, Zhenxing Qian, and Xinpeng Zhang. Imperceptible Backdoor Attack: From Input Space to Feature Representation. arXiv preprint arXiv:2205.03190, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 453 + ], + "type": "text", + "content": "[35] Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Scholkopf. Learning with Local and Global Consistency. Advances in neural information processing systems, 16, 2003." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "164" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/5fbfedc8-4ebc-4e53-8cf0-a4029a8a49ba_content_list.json b/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/5fbfedc8-4ebc-4e53-8cf0-a4029a8a49ba_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..afa0273d69a766d185a52104fcbd2b4f0d9d1d00 --- /dev/null +++ b/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/5fbfedc8-4ebc-4e53-8cf0-a4029a8a49ba_content_list.json @@ -0,0 +1,2143 @@ +[ + { + "type": "text", + "text": "Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences", + "text_level": 1, + "bbox": [ + 86, + 130, + 883, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Felix Rydell", + "bbox": [ + 243, + 203, + 346, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "KTH Royal Institute of Technology", + "bbox": [ + 153, + 222, + 437, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lindstedtsvagen 25, Stockholm, Sweden", + "bbox": [ + 135, + 239, + 457, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "felixry@kth.se", + "bbox": [ + 233, + 258, + 359, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Elima Shehu", + "bbox": [ + 614, + 204, + 720, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Osnabrück and MPI MiS", + "bbox": [ + 514, + 220, + 821, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Albrechtstrasse 28a, Osnabrück, Germany", + "bbox": [ + 501, + 239, + 834, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inselstrasse 22, Leipzig, Germany", + "bbox": [ + 532, + 257, + 803, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "elima.shehu@mis.mpg.de", + "bbox": [ + 568, + 277, + 764, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Angélica Torres", + "bbox": [ + 419, + 299, + 550, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Centre de Recerca Matemàtica", + "bbox": [ + 362, + 318, + 607, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Edifici C, Campus UAB, Bellaterra, Spain", + "bbox": [ + 318, + 335, + 651, + 352 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "atorres@crm.cat", + "bbox": [ + 416, + 354, + 552, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 404, + 313, + 420 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We study the joint image of lines incident to points, meaning the set of image tuples obtained from fixed cameras observing a varying 3D point-line incidence. We prove a formula for the number of complex critical points of the triangulation problem that aims to compute a 3D point-line incidence from noisy images. Our formula works for an arbitrary number of images and measures the intrinsic difficulty of this triangulation. Additionally, we conduct numerical experiments using homotopy continuation methods, comparing different approaches of triangulation of such incidences. In our setup, exploiting the incidence relations gives a notably faster point reconstruction with comparable accuracy.", + "bbox": [ + 75, + 436, + 472, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Introduction", + "text_level": 1, + "bbox": [ + 76, + 662, + 189, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Structure-from-Motion pipeline aims to estimate the camera positions and the relative position of the objects present in a given set of images. The process starts by identifying point and line features in one image that are recognizable as the same points or lines in another. These matched features, that we call correspondences, are used to estimate the camera position and orientation, which then allow for the triangulation step where the position of the points and lines is estimated.", + "bbox": [ + 75, + 688, + 468, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we study the triangulation of points and lines satisfying a certain incidence relation. Specifically, we focus on the triangulation of points contained in a single line; see Figure 1. We note that our methods can also be used when multiple lines going through a fixed point. We", + "bbox": [ + 75, + 825, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f9dcca4617d849064605759b8dc46da8cce53217a254be158992f305c8ec7154.jpg", + "image_caption": [ + "Figure 1: The illustration of two different types of incidence relations: (L1) represents the scenario where multiple points are incident to a line, while (P1) represents the scenario where multiple lines are incident to a point." + ], + "image_footnote": [], + "bbox": [ + 576, + 407, + 816, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "develop the theory for the triangulation of these problems for an arbitrary number of pinhole cameras assuming complete visibility.", + "bbox": [ + 496, + 595, + 890, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The triangulation of points is well understood and, in practice, it is efficiently implemented. However, to the best of our knowledge, incidence relations are not considered in these implementations. The inclusion of line features and incidence relations in the triangulation process could give a more accurate estimation of scenes as lines are more robust to noise than points, and incidence relations appear frequently in interior scenes and man-made scenarios; see [34], [24], [30]. Additionally, in some real data sets standard feature detection algorithms fail due to a lack of point correspondences but succeed when line correspondences are taken into account [15].", + "bbox": [ + 496, + 642, + 892, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The main tools in our work are (algebraic) varieties, that is, the vanishing sets of systems of polynomial equations. Algebraic varieties have been used extensively to study triangulation of point correspondences [1, 17, 31], line correspondences [7, 16, 29] and minimal problems [12, 13] to", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3748", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "name a few. In these works, algebraic varieties arise naturally due to the algebraic nature of pinhole cameras.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A pinhole camera, is modeled by a (complex) projective linear map $C: \\mathbb{P}^3 \\dashrightarrow \\mathbb{P}^2$ defined by a $3 \\times 4$ matrix $C$ of full rank that takes a point $X \\in \\mathbb{P}^3$ and sends it to $CX \\in \\mathbb{P}^2$ . A camera arrangement with $m \\geq 2$ cameras is denoted by $\\mathcal{C} = (C_1, \\ldots, C_m)$ , and the joint camera map", + "bbox": [ + 75, + 121, + 468, + 196 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Phi_ {\\mathcal {C}}: \\mathbb {P} ^ {3} \\dashrightarrow (\\mathbb {P} ^ {2}) ^ {m}, \\tag {1} \\\\ X \\mapsto (C _ {1} X, \\dots , C _ {m} X), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 208, + 468, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models the process of taking the images of a point $X$ in homogeneous coordinates with $m$ cameras. For fixed cameras, the (point) multiview variety $\\mathcal{M}_{\\mathcal{C}}$ is the smallest variety that contains all point correspondences. An extensive account of the pinhole cameras is given by [22], and a survey of the multiview variety is found in [36]. The joint camera map can be extended from the space of points $\\mathbb{P}^3$ to the space of lines, denoted by $\\mathrm{Gr}(1,\\mathbb{P}^3)$ , where each line is parametrized as the span of two points. The map", + "bbox": [ + 75, + 256, + 470, + 393 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Upsilon_ {\\mathcal {C}}: \\operatorname {G r} (1, \\mathbb {P} ^ {3}) \\dashrightarrow (\\mathbb {P} ^ {2}) ^ {m}, \\tag {2} \\\\ L \\mapsto (C _ {1} \\cdot L, \\dots , C _ {m} \\cdot L). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 402, + 468, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models the image of a line $L$ taken by the cameras of $\\mathcal{C}$ . To be precise, if $u, v$ span the line $L$ in $\\mathbb{P}^3$ , then $C \\cdot L$ is defined as $\\ell = Cu \\times Cv \\in \\mathbb{P}^2$ , where $\\times$ is the cross product. Observe that $\\{x \\in \\mathbb{P}^2 : \\ell^T x = 0\\}$ equals $\\operatorname{span}\\{Cu, Cv\\}$ . Recently, in [7], the authors study the line multiview variety referred to as $\\mathcal{L}_{\\mathcal{C}}$ , which is the smallest variety containing all line correspondences.", + "bbox": [ + 75, + 450, + 468, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contribution is the definition and study of the anchored point and line multiview varieties. Given a fixed line $L$ in $\\mathbb{P}^3$ , the anchored point multiview variety, $\\mathcal{M}_C^L$ , is defined as the smallest variety containing all point correspondences coming from points in $L$ ; and the anchored line multiview variety, $\\mathcal{L}_C^X$ , is the smallest variety containing the line correspondences coming from lines passing through a given point $X \\in \\mathbb{P}^3$ .", + "bbox": [ + 75, + 556, + 468, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For these new varieties we prove formulas for their Euclidean distance degree (EDD), which are a measurement of the complexity for error correction using exact algebraic methods [11]. Specifically, we prove the following theorem:", + "bbox": [ + 75, + 676, + 468, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Theorem 0.1. Let $\\mathcal{C}$ be a generic arrangement of $m$ cameras.", + "bbox": [ + 75, + 748, + 468, + 777 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. $\\operatorname{EDD}(\\mathcal{M}_{\\mathcal{C}}^{L}) = 3m - 2$", + "2. If $m \\geq 3$ , then $\\mathrm{EDD}(\\mathcal{L}_{\\mathcal{C}}^{X}) = \\frac{9}{2} m^{2} - \\frac{19}{2} m + 3$ ." + ], + "bbox": [ + 86, + 787, + 423, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We provide a precise definition of this degree in Section 1.2. Previous work on EDD for multiview varieties includes [19, 31] and [7, Section 5]. The EDD of the point multiview variety is $\\mathrm{EDD}(\\mathcal{M}_{\\mathcal{C}}) = \\frac{9}{2} m^3 -$", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$\\frac{21}{2} m^2 + 8m - 4$ , according to [31]. The fact that the EDDs of the anchored multiview varieties are polynomials of smaller degrees suggests that they are less complex for the purpose of data correction. This conclusion is backed by the results of our numerical experiments using HomotopyContinuation.jl [9], presented in Section 2.", + "bbox": [ + 498, + 90, + 890, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another contribution of our work is the numerical simulations for the triangulation of points contained in a line using the anchored multiview varieties. We present different approaches to triangulating data of type (L1) that differ from the traditional method of fitting point correspondences to the multiview variety $\\mathcal{M}_{\\mathcal{C}}$ . For $m = 2$ views, our implementation is notably faster than the traditional triangulation of points described above, while the accuracy is comparable. For $m = 3$ views we get both higher accuracy and faster speed by using $\\mathcal{L}_{\\mathcal{C}}$ , compared to usual point triangulation. All of our proposed methods outperform the traditional approaches in terms of run-time and give a comparably accurate result. In practice, special software and hardware are used for triangulation, and based on our experiments, we believe that these approaches could be implemented efficiently with good results.", + "bbox": [ + 496, + 196, + 892, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Related work", + "text_level": 1, + "bbox": [ + 500, + 448, + 609, + 463 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Algebra. Algebraic geometry, whose connection to computer vision is well established, is our main tool for the theoretical study of the triangulation of problem (L1). Fundamental theory regarding the point multiview variety is found in [17, 20, 22], in particular for two and three views. The paper [36] by Trager et. al. serves as a survey for the algebraic properties of this multiview variety. For applications, finding a good set of polynomial constraints satisfied by the point correspondences is important, this is precisely the work of [1] by Agarwal et al.", + "bbox": [ + 496, + 472, + 890, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regarding the algebra of lines in a computer vision setting, Kileel presented in [29, Definition 3.9, Theorem 3.10] several types of multiview varieties with respect to a point-line incidence relation in $\\mathbb{P}^3$ in 3 views, and their basic properties. One of those types is called the LLL-multiview variety and constitutes a special case of the more recent work by Breiding et. al. [7], where they study algebraic properties of the line multiview variety. Furthermore, a recent manuscript [6] studies the polynomial constraints satisfied by line correspondences.", + "bbox": [ + 496, + 625, + 890, + 775 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Projective Reconstruction. Different approaches to the triangulation of points have been considered in the literature, in particular for two views [3, 4, 21, 26-28]. Hartley and Sturm compare many different approaches in [21], including the \"midpoint\" method. The midpoint method inputs a point correspondence in two views and outputs the midpoint of the line segment determined by the points on", + "bbox": [ + 496, + 794, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3749", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the two back-projected lines that are closest to each other. In other words, it finds the midpoint of the common perpendicular to the two back-projected lines. This is a natural approach, but it has several downsides, also pointed out in [3, 4]. Importantly, it is not projectively invariant. A projectively invariant reconstruction has the property that acting on the cameras by a global projective transformation induces the same action on the triangulated point or line. Instead, Hartley and Sturm propose minimizing the reprojection error as the optimal triangulation method, which is accepted as a standard formulation [35] and it is projectively invariant. In [35], the focus is on triangulation via minimizing reprojections errors in three views and the authors point out that three views often lead to greater stability and stronger disambiguation compared to two views.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implementation of Line Reconstruction. From the algorithmic point of view, the simultaneous reconstruction of point and line features have been studied specially with the goal of reconstructing line segments. For example, [2] provides a thorough overview of the structure-from-motion pipeline using lines, going through different methods for line parameterization and error correction depending on such parameterizations. In [33], Quan and Takeo study the algebraic structure of line correspondences with uncalibrated affine cameras. They reduce the problem of reconstructing affine lines to the reconstruction of projective points in a lower-dimensional projective camera. In [20], Hartley and coauthors provide an algorithm for the reconstruction of point and line features where 3D lines are parameterized by their projections in 2 views (as the intersection of two back-projected planes). Micusik and Wildenauer reconstruct lines to estimate line segments, but only use incident points for error correction [32]. Furthermore in [23], the authors propose a technique for 3D reconstruction incorporating line segments. They assert that this method surpasses traditional approaches, especially in efficiency while getting accurate results. Finally, in [14], the authors present a framework for the computation of the relative motion between two images using a triplet of lines.", + "bbox": [ + 75, + 347, + 472, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper is structured as follows. In the first part of Section 1 we formally introduce anchored multiview varieties and study some properties. In the second part of Section 1, we study their smoothness and find their Euclidean Distance Degree. Finally, in Section 2, we provide a numerical analysis of different approaches to reconstructing point correspondences incident to a line correspondence. The proofs of all our results are included in the Supplementary Material together with a small background on the concepts of Algebraic Geometry used for these results and the code for the numerical experiments.", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. Anchored Multiview Varieties", + "text_level": 1, + "bbox": [ + 500, + 89, + 774, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A camera refers to a full-rank $3 \\times 4$ matrix. A camera arrangement is a collection $\\mathcal{C} = (C_1, \\ldots, C_m)$ of $m \\geq 2$ cameras whose centers, i.e. kernels, are all distinct.", + "bbox": [ + 496, + 114, + 893, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A variety is the solution set to a system of polynomial equations. The Zariski closure of a set $U$ is the smallest variety containing $U$ . We work in complex projective space, denoted $\\mathbb{P}^n$ . This is defined as $(\\mathbb{C}^n \\setminus \\{0\\}) / \\sim$ , where $x \\sim y$ if $x$ and $y$ differ by a non-zero constant. The set of lines in $\\mathbb{P}^n$ is denoted as $\\operatorname{Gr}(1, \\mathbb{P}^n)$ . In the language of algebraic geometry, it is called the Grassmanian of lines in $\\mathbb{P}^n$ . In general we denote with upper case letters world objects and with lower case letters image objects. For a rational map $\\varphi: X \\dashrightarrow Y$ , we denote $\\varphi|_A$ the restriction of $\\varphi$ to $A \\subseteq X$ .", + "bbox": [ + 496, + 160, + 893, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1.1. Let $X \\in \\mathbb{P}^3$ be a point distinct from all camera centers, and $L$ a line in $\\mathbb{P}^3$ containing none of the camera centers.", + "bbox": [ + 496, + 316, + 893, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. The anchored point multiview variety, denoted $\\mathcal{M}_C^L$ , is defined as the Zariski closure of the image of the map", + "bbox": [ + 511, + 367, + 890, + 398 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c} \\Phi_ {\\mathcal {C}} | _ {L}: L \\dashrightarrow (\\mathbb {P} ^ {2}) ^ {m}, \\\\ X \\longmapsto (C _ {1} X, \\dots , C _ {m} X). \\end{array} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 402, + 890, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. The anchored line multiview variety, denoted $\\mathcal{L}_C^X$ , is defined as the Zariski closure of the image of the map", + "bbox": [ + 511, + 450, + 890, + 482 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c} \\Upsilon_ {\\mathcal {C}} | _ {\\Lambda (X)}: \\Lambda (X) \\dashrightarrow (\\mathbb {P} ^ {2}) ^ {m}, \\\\ L \\longmapsto \\left(C _ {1} \\cdot L, \\dots , C _ {m} \\cdot L\\right), \\end{array} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 486, + 890, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Lambda(X)$ denotes the set of lines in $\\mathbb{P}^3$ that contain $X$ and $\\ell_i = C_i \\cdot L$ is defined as in the introduction.", + "bbox": [ + 531, + 529, + 890, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The anchored point multiview variety $\\mathcal{M}_{\\mathcal{C}}^{L}$ is the smallest variety that contains all point correspondences $\\Phi_{\\mathcal{C}}(X)$ for $X \\in L$ . Similarly, the anchored line multiview variety $\\mathcal{L}_{\\mathcal{C}}^{X}$ is the smallest variety that contains all line correspondences $\\Upsilon_{\\mathcal{C}}(L)$ for $L$ meeting $X$ and no center. We highlight that our definition of anchored multiview varieties is different from the anchored features in [25] for monocular EKF-SLAM.", + "bbox": [ + 496, + 565, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1.2. Consider an arrangement of $m$ cameras $\\mathcal{C} = (C_1, \\ldots, C_m)$ , a point $X \\in \\mathbb{P}^3$ and a line $L$ in $\\mathbb{P}^3$ satisfying the conditions of Definition 1.1.", + "bbox": [ + 496, + 691, + 890, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. If there are two different camera centers $c_{i}$ and $c_{j}$ such that the span of $\\{c_{i}, c_{j}, L\\}$ is $\\mathbb{P}^3$ , then", + "bbox": [ + 511, + 743, + 890, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {\\mathcal {C}} ^ {L} = \\left\\{\\left(x _ {1}, \\dots , x _ {m}\\right) \\in \\mathcal {M} _ {\\mathcal {C}}: x _ {i} \\in C _ {i} \\cdot L \\right\\}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 777, + 890, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. If for each camera center $c_i$ , the line spanned by $c_i$ and $X$ does not contain any other camera center, then", + "bbox": [ + 511, + 806, + 890, + 837 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {C}} ^ {X} = \\left\\{\\left(\\ell_ {1}, \\dots , \\ell_ {m}\\right) \\in \\mathcal {L} _ {\\mathcal {C}}: C _ {i} X \\in \\ell_ {i} \\right\\}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 842, + 890, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proofs for this and all our results are found in the Supplementary Material.", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3750", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1.1. Properties of Anchored Multiview Varieties", + "text_level": 1, + "bbox": [ + 76, + 90, + 444, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A fundamental property of the anchored multiview varieties is that they are linearly isomorphic to multiview varieties arising from projections $\\mathbb{P}^1\\longrightarrow \\mathbb{P}^1$ and $\\mathbb{P}^2\\longrightarrow \\mathbb{P}^1$ respectively. A linear isomorphism is a linear map (given by a matrix) with an inverse that is also linear.", + "bbox": [ + 75, + 113, + 467, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consider an arrangement $\\widetilde{\\mathcal{C}}$ of $m$ full-rank $2\\times 2$ matrices, and an arrangement $\\widehat{\\mathcal{C}}$ of $m$ full-rank $2\\times 3$ matrices. Also here we assume $m\\geq 2$ . We define $\\mathcal{M}_{\\widetilde{\\mathcal{C}}}^{1,1}$ , and $\\mathcal{M}_{\\widehat{\\mathcal{C}}}^{2,1}$ respectively as the Zariski closure of the image of the maps", + "bbox": [ + 75, + 188, + 467, + 250 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {\\tilde {C}}: \\mathbb {P} ^ {1} \\longrightarrow (\\mathbb {P} ^ {1}) ^ {m}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 268, + 467, + 294 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX \\longmapsto (\\widetilde {C} _ {1} X, \\dots , \\widetilde {C} _ {m} X)\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 290, + 377, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and", + "bbox": [ + 76, + 315, + 106, + 327 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {\\widehat {C}}: \\mathbb {P} ^ {2} \\dashrightarrow (\\mathbb {P} ^ {1}) ^ {m}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 334, + 467, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX \\longmapsto (\\widehat {C} _ {1} X, \\dots \\widehat {C} _ {m} X).\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 354, + 375, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1.3.", + "text_level": 1, + "bbox": [ + 76, + 380, + 174, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1. Let $\\phi_L: L \\to \\mathbb{P}^1$ and $\\psi_{\\mathcal{C},i}: C_i \\cdot L \\to \\mathbb{P}^1$ be any choices of linear isomorphisms. Let $\\tilde{\\mathcal{C}}$ denote the arrangement of matrices $C_i := \\psi_{\\mathcal{C},i} \\circ C_i \\circ \\phi_L^{-1}$ . Then", + "bbox": [ + 89, + 401, + 468, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {\\mathcal {C}, L} := \\left(\\psi_ {\\mathcal {C}, 1}, \\dots , \\psi_ {\\mathcal {C}, m}\\right): \\mathcal {M} _ {\\mathcal {C}} ^ {L} \\rightarrow \\mathcal {M} _ {\\tilde {\\mathcal {C}}} ^ {1, 1} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 470, + 467, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is a linear isomorphism.", + "bbox": [ + 109, + 498, + 272, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2. Let $\\phi_X: \\Lambda(X) \\to \\mathbb{P}^2$ and $\\psi_{\\mathcal{C},i}: \\Lambda(C_iX) \\to \\mathbb{P}^1$ be any choices of linear isomorphisms. Let $\\widehat{\\mathcal{C}}$ denote the arrangement of matrices $\\widehat{C}_i \\coloneqq \\psi_{\\mathcal{C},i} \\circ C_i \\circ \\phi_X^{-1}$ . Then", + "bbox": [ + 88, + 520, + 468, + 570 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {\\mathcal {C}, X} := \\left(\\psi_ {\\mathcal {C}, 1}, \\dots , \\psi_ {\\mathcal {C}, m}\\right): \\mathcal {L} _ {\\mathcal {C}} ^ {X} \\rightarrow \\mathcal {M} _ {\\widehat {\\mathcal {C}}} ^ {2, 1} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 589, + 467, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is a linear isomorphism.", + "bbox": [ + 109, + 616, + 272, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Section 2.1.1, we exploit these linear isomorphisms to improve the speed of triangulation for our setting. A consequence of this theorem is also that many results for the anchored multiview varieties translate into results about $\\mathcal{M}_{\\hat{\\mathcal{C}}}^{1,1}$ and $\\mathcal{M}_{\\hat{\\mathcal{C}}}^{2,1}$ .", + "bbox": [ + 75, + 637, + 467, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1.4. $\\mathcal{M}_C^L$ and $\\mathcal{L}_C^X$ are irreducible. Further,", + "bbox": [ + 76, + 726, + 447, + 743 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. $\\mathcal{M}_C^L$ is isomorphic to $\\mathbb{P}^1$ . In particular, $\\dim \\mathcal{M}_C^L = 1$ .", + "2. If the span of the centers $c_{i}$ and the point $X$ are not collinear, then $\\dim \\mathcal{L}_{\\mathcal{C}}^{X} = 2$ ." + ], + "bbox": [ + 88, + 748, + 468, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using the fundamental matrices of $C_i$ and $C_j$ , denoted as $F^{ij}$ , we introduce a set of polynomial constraints that must be fulfilled by the correspondences of points or lines in the anchored multiview varieties. These constraints imply that some of the equations obtained from Proposition 1.2 are redundant in the generic case.", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1.5. For a point $X \\in \\mathbb{P}^3$ and line $L$ in $\\mathbb{P}^3$ , let $\\mathcal{C}$ be a generic (random) camera arrangement of $m$ cameras.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. $x \\in \\mathcal{M}_{\\mathcal{C}}^{L}$ if and only if $x_{1}^{T}F^{1j}x_{j} = 0$ for every $j = 2, \\ldots, m$ and $x_{i}^{T}C_{i} \\cdot L = 0$ for every $i = 1, \\ldots, m$ .", + "2. $\\ell \\in \\mathcal{L}_C^X$ if and only if" + ], + "bbox": [ + 511, + 130, + 890, + 188 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\det \\left[ \\begin{array}{l l l} C _ {1} ^ {T} \\ell_ {1} & C _ {2} ^ {T} \\ell_ {2} & C _ {i} ^ {T} \\ell_ {i} \\end{array} \\right] = 0, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 199, + 890, + 224 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\det \\left[ \\begin{array}{c c c} C _ {1} ^ {T} \\ell_ {1} & C _ {3} ^ {T} \\ell_ {3} & C _ {i} ^ {T} \\ell_ {i} \\end{array} \\right] = 0\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 219, + 818, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for $i = 3,\\ldots ,m$ and $\\ell_i^T C_iX = 0$ for every $i = 1,\\dots ,m$ .", + "bbox": [ + 527, + 250, + 890, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The determinantal constraints described in the second item correspond to the constraints that are satisfied by elements of the line multiview variety, presented in [7].", + "bbox": [ + 496, + 291, + 890, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Supplementary Material Section B we define the multidegree of a variety in a product of projective spaces, determine it for the anchored multiview varieties and explain its relevance for computer vision.", + "bbox": [ + 496, + 337, + 890, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1.2. The Euclidean Distance problem", + "text_level": 1, + "bbox": [ + 500, + 407, + 785, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a variety $\\mathcal{X} \\subseteq \\mathbb{R}^n$ and a point $u \\in \\mathbb{R}^n$ outside the variety, a natural problem is to find the closest point on $\\mathcal{X}$ to $u$ , which corresponds to the optimization problem", + "bbox": [ + 496, + 430, + 890, + 476 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {m i n i m i z e} \\quad \\sum_ {i = 1} ^ {n} \\left(u _ {i} - x _ {i}\\right) ^ {2} \\quad \\text {s u b j e c t t o} \\quad x \\in \\mathcal {X}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 488, + 890, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Equation (12) is called the Euclidean distance problem and models the process of error correction and fitting noisy data to a mathematical model $\\mathcal{X}$ . In the case of a smooth variety, defined below, the Euclidean distance degree (EDD) is the number of complex solutions to the critical equations of (12) [11]. The EDD is an estimate of how difficult it is to solve this problem by exact algebraic methods. For a variety $\\mathcal{X}$ in a product of projective spaces $\\mathbb{P}^{n_1} \\times \\dots \\times \\mathbb{P}^{n_m}$ , the EDD is the EDD of $\\mathcal{X} \\cap U_1 \\times \\dots \\times U_m \\subseteq \\mathbb{R}^{n_1 + \\dots + n_m}$ , where $U_i$ is a generic affine patch of $\\mathbb{P}^{h_i}$ for each $i$ . An affine patch $U$ of $\\mathbb{P}^n$ is a subset defined by an affine equation with non-zero constant part, for instance $x_0 = 1$ . We have $U \\cong \\mathbb{R}^n$ over the real numbers.", + "bbox": [ + 496, + 537, + 890, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A variety $\\mathcal{X}$ is smooth at a point $x$ if $\\mathcal{X}$ locally around $x$ looks like Euclidean space. $\\mathcal{X}$ is smooth if all its points are smooth. The singular locus of a variety is the set of non-smooth points. See [18] for more details. The singular locus of the point multiview variety is well-understood, see for instance [36], and it is mostly understood for the line multiview variety, see [7, Section 3]. The proposition below guarantees that the anchored multiview varieties are smooth for generic camera arrangements, which helps us to compute its EDD.", + "bbox": [ + 496, + 734, + 890, + 883 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3751", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 1.6.", + "text_level": 1, + "bbox": [ + 76, + 90, + 189, + 104 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. $\\mathcal{M}_C^L$ smooth.", + "2. If there are exactly two cameras, or the centers together with the point $X$ span $\\mathbb{P}^3$ , then $\\mathcal{L}_C^X$ is smooth." + ], + "bbox": [ + 88, + 114, + 467, + 171 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the EDD to be relevant in a particular setting, the Euclidean distance has to be a good measurement of distance. This is naturally the case for points in $\\mathbb{R}^2$ , i.e. an affine patch of $\\mathbb{P}^2$ . This extends to an affine patch of $\\mathcal{M}_{\\mathcal{C}}$ . It is perhaps less clear how to best measure distances between lines. Recall that we regard $\\mathcal{L}_{\\mathcal{C}}$ as a subset of $(\\mathbb{P}^2)^m$ . Indeed, we identify image lines with the point that defines its normal vector. We choose the Euclidean distance between these normal vectors (in affine patches of $\\mathbb{P}^2$ ) as our distance between lines.", + "bbox": [ + 75, + 180, + 467, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The theorem below presents our computations of the EDDs of the anchored multiview varieties. Additionally, our numerical computations using the HomotopyContinuation.jl [8] package in the julia [5] programming language have confirmed the accuracy of these formulas for $m \\leq 10$ .", + "bbox": [ + 76, + 330, + 467, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1.7. Let $\\mathcal{C}$ be a generic arrangement of $m$ cameras.", + "bbox": [ + 76, + 430, + 467, + 459 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. $\\operatorname{EDD}(\\mathcal{M}_C^L) = 3m - 2$ .", + "2. If $m \\geq 3$ , then $\\mathrm{EDD}(\\mathcal{L}_{\\mathcal{C}}^{X}) = \\frac{9}{2} m^{2} - \\frac{19}{2} m + 3$ ." + ], + "bbox": [ + 88, + 467, + 421, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We end with an implication of Theorem 1.3: There is a direct correspondence between the EDDs of the anchored multiview varieties and $\\mathcal{M}_{\\widetilde{\\mathcal{C}}}^{1,1}, \\mathcal{M}_{\\widehat{\\mathcal{C}}}^{2,1}$ .", + "bbox": [ + 76, + 515, + 467, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1.8. Let $\\hat{C}$ and $\\hat{C}$ be generic arrangements of cardinality $m$ .", + "bbox": [ + 76, + 573, + 467, + 603 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. $\\operatorname{EDD}(\\mathcal{M}_{\\widetilde{c}}^{1,1}) = 3m - 2.$", + "2. If $m \\geq 3$ , then $\\mathrm{EDD}(\\mathcal{M}_{\\hat{\\mathcal{C}}}^{2,1}) = \\frac{9}{2} m^2 - \\frac{19}{2} m + 3$ ." + ], + "bbox": [ + 88, + 609, + 434, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2. Numerical Experiments", + "text_level": 1, + "bbox": [ + 76, + 667, + 300, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section we conduct numerical experiments of the triangulation process of incident point and lines. All the code can be found in the Supplementary Material.", + "bbox": [ + 76, + 694, + 467, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By (L1) we refer to point-line arrangements in $\\mathbb{P}^3$ of one line incident to $p$ points, as depicted in Figure 1. Our goal is to reconstruct such arrangements, given point correspondences incident to a single line correspondence across $m$ views. The following is a list of natural approaches for such triangulation given a camera arrangement $\\mathcal{C}$ . It is not necessarily a complete list.", + "bbox": [ + 75, + 739, + 467, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(L1).0 Triangulate each point correspondence by fitting it to the point multiview variety $\\mathcal{M}_{\\mathcal{C}}$ , i.e. find the closest point correspondence in $\\mathcal{M}_{\\mathcal{C}}$ ;", + "bbox": [ + 84, + 854, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/348ea6aca160ede2bc865b7669ebe94dc2efe1b17bfe45052a37721cf20985a9.jpg", + "image_caption": [ + "Figure 2: Comparison of (L1).0 and (L1).1 triangulation approaches. (L1).0 expects non-collinear points in the reconstruction, while (L1).1 produces collinear points." + ], + "image_footnote": [], + "bbox": [ + 529, + 92, + 694, + 203 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cda082f93cda9cec7bd913c67990e93c00e05227df19c28c051063128f0822b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 94, + 862, + 203 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(L1).1 Reconstruct the 3D line $L$ by back-projecting the image lines from two views. Triangulate the point correspondences by fitting them to the anchored multiview variety $\\mathcal{M}_{\\mathcal{C}}^{L}$ ;", + "(L1).2 Triangulate two point correspondences by fitting them to $\\mathcal{M}_C$ to get 3D points $X$ and $Y$ . Let $L$ be the line they span in $\\mathbb{R}^3$ . Triangulate the remaining point correspondences by fitting them to $\\mathcal{M}_C^L$ ;", + "(L1).3 Triangulate one point correspondence by fitting it to $\\mathcal{M}_{\\mathcal{C}}$ to get a 3D point $X$ . Reconstruct the 3D line $L$ by fitting the line correspondence to the anchored variety $\\mathcal{L}_{\\mathcal{C}}^{X}$ . Triangulate the remaining point correspondences by fitting them to $\\mathcal{M}_{\\mathcal{C}}^{L}$ ;", + "(L1).4 Reconstruct the 3D line $L$ by fitting the line correspondence to the line multiview variety $\\mathcal{L}_{\\mathcal{C}}$ . Triangulate the point correspondences by fitting them to the anchored multiview variety $\\mathcal{M}_{\\mathcal{C}}^{L}$ ." + ], + "bbox": [ + 506, + 281, + 890, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Approaches (L1).1-4 take the incidence relations into account in the triangulation, whereas (L1).0 does not. Therefore, in contrast to (L1).1-4, the resulting triangulation of (L1).0 does not preserve the point-line incidences; see Figure 2.", + "bbox": [ + 498, + 582, + 890, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.1. Implementation", + "text_level": 1, + "bbox": [ + 500, + 667, + 658, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our experiments are implemented in HomotopyContinuation.jl [8] in Julia [5]. Here we explain them in some detail. The full explanation can be found in Supplementary Material. We ran the code on a Intel(R) Core(TM) i5-8300H CPU running at 2.30GHz.", + "bbox": [ + 498, + 691, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.1.1 Reducing the number of parameters", + "text_level": 1, + "bbox": [ + 500, + 800, + 805, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We reduce the number of parameters involved in the problem of fitting data to $\\mathcal{M}_{\\mathcal{C}}^{L}$ , respectively $\\mathcal{L}_{\\mathcal{C}}^{X}$ , by translating it to and solving it for $\\mathcal{M}_{\\widetilde{\\mathcal{C}}}^{1,1}$ , respectively $\\mathcal{M}_{\\widehat{\\mathcal{C}}}^{2,1}$ . Details are found in Supplementary Material Section B. This translation corresponds to a reduction of parameters because an", + "bbox": [ + 498, + 823, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3752", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "affine patch of $\\mathcal{M}_{\\widehat{\\mathcal{C}}}^{1,1}$ lives in $(\\mathbb{R}^1)^m$ , while an affine patch of $\\mathcal{M}_{\\mathcal{C}}^{L}$ lives in $(\\mathbb{R}^2)^m$ . The analogous is true for $\\mathcal{M}_{\\widehat{\\mathcal{C}}}^{2,1}$ and $\\mathcal{L}_{\\mathcal{C}}^{X}$ .", + "bbox": [ + 75, + 89, + 467, + 143 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We call the methods that do not use this translation standard and for instance, write (L1).1 std. to denote it. By just writing (L1).1, we denote the implementation of this translation.", + "bbox": [ + 75, + 143, + 467, + 203 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2.1.2 Evaluating error", + "text_level": 1, + "bbox": [ + 76, + 226, + 246, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiments, each iteration starts by randomly generating a 3D line $L$ and $p$ points $X_{i}$ on this line. This line and these points are projected by randomly generated cameras to line and point correspondences in $(\\mathbb{R}^2)^m$ . For fixed $\\epsilon = 10^{-12}$ , randomly generated noise vectors $\\sigma (\\epsilon)$ of length $\\epsilon$ are added in each factor. Our approaches then find points $Y_{i}\\in \\mathbb{R}^{3}$ that match the noisy correspondences. We measure the accuracy by taking the logarithm after averaging the relative error of reconstructed points:", + "bbox": [ + 75, + 251, + 468, + 387 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ne = \\log_ {1 0} \\left(\\frac {\\sum \\| Y _ {i} - X _ {i} \\|}{p \\epsilon}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 398, + 468, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\| Y_{i} - X_{i}\\|$ denotes the Euclidean distance. The interpretation of this number is that the error in the data gets amplified by $10^{e}$ during the triangulation.", + "bbox": [ + 75, + 445, + 468, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2.1.3 Algorithms", + "text_level": 1, + "bbox": [ + 76, + 512, + 209, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The pseudocodes for all approaches are included in the Supplementary Material. The pseudocode for approach (L1).1 is presented in Algorithm 1. We use the notation that for a column vector $X \\in \\mathbb{R}^n$ , $[X; 1] \\in \\mathbb{R}^{n+1}$ is the vector we get by adding a 1 as the last coordinate. Let $\\iota$ be the function that scales a vector such that its last coordinate is 1, and then removes that coordinate. In lines 1 and 2 of the algorithm, we generate noisy data by introducing randomly generated noise $\\sigma(\\epsilon)$ as explained above, where $\\epsilon = 10^{-12}$ is a fixed parameter for our experiments. Lines 3 and 4 correspond to the triangulation of the line and the point correspondences using the anchored point multiview variety at the line $L_0$ . Note that we solve the closest point problem in line 4 by computing the zeros of a system of polynomial (critical) equations using the solve function in HomotopyContinuation.jl [9]. Finally, line 5 compares the points obtained in the previous step with the original starting points, measuring the logarithmic average relative error of the $p$ points.", + "bbox": [ + 75, + 537, + 468, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Most other pseudocodes are similar to Algorithm 1. For instance, approach (L1).0 does not take into account the line, so it corresponds deleting lines 1 and 3 from Algorithm 1 and modifying the minimization domain in line 4, replacing $L_{0}$ by $\\mathcal{M}_{\\mathcal{C}}$ . In approach (L1).2, line 3 is modified", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "so that $L_{0}$ is obtained by the span of two triangulated points instead of by intersecting two back-projected planes.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1: One iteration of the (L1).1 std. method given a randomly generated camera arrangement $\\mathcal{C}$ of $3\\times 4$ matrices, a projective line $L$ spanned by two vectors of $\\mathbb{R}^4$ , and $p$ points $X_{i}\\in \\mathbb{R}^{3}$ such that $[X_i;1]$ lie on $L$ .", + "bbox": [ + 506, + 140, + 848, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input : $\\mathcal{C} = (C_1, \\dots, C_m)$ , $L, X_1, \\dots, X_p$", + "bbox": [ + 519, + 219, + 813, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Output: The log of the average relative error", + "bbox": [ + 521, + 236, + 818, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1 for $j$ from 1 to $m$ do", + "bbox": [ + 508, + 251, + 661, + 265 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\text {\\bf 2} \\\\ \\text {\\bf 3} \\\\ \\text {\\bf 4} \\end{array} \\left\\lfloor \\begin{array}{l} \\text {\\bf f o r} i \\text {f r o m 1 t o p d o} \\\\ \\text {\\bf \\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\c h i}}}}}}}}}}}}}}}}}}}}}} q _ {i, j} \\leftarrow \\iota (C _ {j} [ X _ {i}; 1 ]) + \\sigma (\\epsilon); \\\\ u _ {j} \\leftarrow \\iota (C _ {j} \\cdot L) + \\sigma (\\epsilon); \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 265, + 759, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 $L_0\\gets$ nullspace $\\left[C_1^T [u_1;1]\\quad C_2^T [u_2;1]\\right]^T;$", + "bbox": [ + 506, + 319, + 816, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6 for $i$ from 1 to $p$ do", + "bbox": [ + 508, + 340, + 651, + 353 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c}Y_{i}\\leftarrow \\\\ \\underset {X\\in \\mathbb{R}^{3}:[X;1]\\in L_{0}}{\\operatorname *{argmin}}\\sum_{j = 1}^{m}(q_{i,j} - \\iota (C_{j}[X;1]))^{2}; \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 353, + 843, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "8 $e\\gets \\log_{10}\\left(\\frac{1}{p\\epsilon}\\sum_{i = 1}^{p}\\| Y_i - X_i\\|\\right)$", + "bbox": [ + 506, + 398, + 750, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Return: $e$", + "bbox": [ + 521, + 422, + 591, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2.2. Numerical results", + "text_level": 1, + "bbox": [ + 500, + 467, + 671, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The main results of our numerical experiments are presented in Figures 3 to 5 and Tables 1 to 3. We compare the performance of the five different triangulation approaches for $p = 5$ and different number of $m$ cameras. In the tables, we present the median, mean and standard deviation $\\sigma$ of the logarithmic average relative error and time. The results are displayed in histograms and tables created with Plots.jl [10].", + "bbox": [ + 496, + 489, + 890, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that for two cameras, (L1).0 and (L1).2 are the exact same, and (L1).1 and (L1).4 are essentially the same. In Figure 3 and Table 1 we present results for $m = 2$ and therefore only include (L1).0, (L1).1 and (L1).3. We also compare with the standard implementations of (L1).1 and (L1).3 that do not use the linear isomorphisms of Theorem 1.3 to improve computation time as described in Section 2.1.1. This simulation is iterated 1000 times.", + "bbox": [ + 496, + 612, + 890, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 4 and Table 2 compare the different reconstruction approaches for $m = 3$ cameras. The simulations are again iterated 1000 times.", + "bbox": [ + 496, + 733, + 890, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 5 and Table 3 compare the different reconstruction approaches for $m = 4$ cameras. The simulations are iterated only 100 times, due to the longer run-time of the simulations.", + "bbox": [ + 496, + 777, + 890, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our theoretical studies in Section 1, specifically Theorem 1.7, guarantee that methods such as (L1).1, (L1).2 and (L1).3 are less complex from the algebraic point of view, than (L1).0 for triangulation of points incident to a line. Ad-", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3753", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0175f04a2e365855a322993fdc8857b162070b3e7ac30ab52c9620d79dd58e21.jpg", + "image_caption": [ + "Accuracy: 2 cameras, 5 point correspondences", + "Speed: 2 cameras, 5 point correspondences" + ], + "image_footnote": [], + "bbox": [ + 130, + 99, + 411, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/22e3360de14aec2cce14ca0de0569b1e13fc38f5c61b988d4374fdfaeac40563.jpg", + "image_caption": [ + "Figure 3: Triangulation of $p = 5$ point correspondences incident to a line for $m = 2$ views with complete visibility. 1000 iterations. The histograms show the frequency of the average relative error and the running time of the depicted triangulation methods." + ], + "image_footnote": [], + "bbox": [ + 130, + 247, + 411, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ditionally, using Section 2.1.1, we were able to improve the computation speed while retaining comparable accuracy, which is exemplified by the two cases (L1).1 vs. (L1).1 std. and (L1).3 vs. (L1).3 std. in Table 1.", + "bbox": [ + 75, + 482, + 467, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, our numerical simulations show that in our HomotopyContinuation.jl implementation, the (L1).1 approach is the fastest across all number of views, but least accurate. In the case of $m = 3$ , (L1).4 is the most accurate and faster than the traditional (L1).0 method. Even for $m = 4$ , (L1).4 is the most accurate but is notably slower than (L1).0. Increasing the number of cameras $m$ improves accuracy and reduces speed, but the exact impact depends on the approach. Among (L1).1-4, (L1).1 is least affected in both regards, and (L1).4 is most affected in both regards. The speed can be thought of as an affine function in the number of point correspondences $p$ . Reconstructing the line correspondence takes some fixed amount of time independent of $p$ and then all $p$ correspondences are independently reconstructed.", + "bbox": [ + 75, + 542, + 468, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 76, + 784, + 346, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This work studied the triangulation of incident points and lines. We introduced the anchored line multiview varieties and showed that they correspond to multiview varieties arising from projections $\\mathbb{P}^2\\longrightarrow \\mathbb{P}^1$ and $\\mathbb{P}^1\\longrightarrow \\mathbb{P}^1$ . In addition, we proved that they are less complex for reconstruction via critical points. These theoretical results are also", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5d3eeaf745c4f1f65d24cd75a99ba01ed53637bff1baf89464b47fcfddfef9a8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Accuracymedianmeanσ
(L1).01.0311.2431.213
(L1).11.4891.7021.085
(L1).1 std.1.3741.5480.990
(L1).31.2501.4490.964
(L1).3 std.1.1471.3931.205
", + "bbox": [ + 550, + 88, + 841, + 185 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4908226caef855744712075ca15cd2a24e36369a2898b0d7d765fedd1062f072.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Speedmedianmeanσ
(L1).00.1700.2010.220
(L1).10.04300.04700.0207
(L1).1 std.0.1220.1410.202
(L1).30.08380.09410.0464
(L1).3 std.0.1670.1880.0764
", + "bbox": [ + 550, + 186, + 841, + 284 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1: Triangulation of $p = 5$ point correspondences incident to a line for $m = 2$ views with complete visibility. 1000 iterations. The tables show the accuracy and speed of the depicted triangulation methods in terms of median, mean and standard deviation.", + "bbox": [ + 500, + 306, + 893, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7844462ac936ab3829315c0261a2a06e2d7b0ed717a40c4d5da5c6411dee73e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 369, + 836, + 516 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b3799ef3cda454409458f2ed5ead66ac62d356c3bc81e6219458d6fd99f4a625.jpg", + "image_caption": [ + "Speed: 3 cameras, 5 point correspondences", + "Figure 4: Triangulation of $p = 5$ point correspondences incident to a line for $m = 3$ views with complete visibility. 1000 iterations. The histograms show the frequency of the average relative error and the running time of the depicted triangulation methods." + ], + "image_footnote": [], + "bbox": [ + 553, + 527, + 834, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "aligned with the numerical experiments we conducted. In particular, the proposed methods in Section 2 compare different methods for triangulating a set of points incident to a line. We highlight that according to our experiments, the use of the theoretical results in Section 1 allows for a notably faster triangulation while preserving the accuracy of the traditional method for point reconstruction.", + "bbox": [ + 496, + 763, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We hope this work motivates the use of new algebraic constraints in the triangulation of incident points and lines.", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3754", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/65a0941c07334f78e5414d734115351d955cefb899082633161ae5eb2bdee88d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Accuracymedianmeanσ
(L1).00.6200.8171.023
(L1).11.5621.7811.128
(L1).20.8781.1250.998
(L1).31.0111.2431.078
(L1).40.4070.4990.901
Speedmedianmeanσ
(L1).00.9841.0600.514
(L1).10.08060.08580.0318
(L1).20.4310.4560.155
(L1).30.3040.3310.134
(L1).40.6160.7630.500
", + "bbox": [ + 130, + 88, + 415, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/416a17a83418e470fa4023c56f763904a3ecee623e65ca56a26a88ab5fcab6fa.jpg", + "image_caption": [ + "Speed: 4 cameras, 5 point correspondences" + ], + "image_footnote": [], + "bbox": [ + 130, + 369, + 413, + 516 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ce8bbba55b9911187364903051bca646c121f6452cdcab3358a919daf52771d.jpg", + "image_caption": [ + "Figure 5: Triangulation of $p = 5$ point correspondences incident to a line for $m = 4$ views with complete visibility. 100 iterations. The histograms show the frequency of the average relative error and the running time of the depicted triangulation methods." + ], + "image_footnote": [], + "bbox": [ + 130, + 529, + 410, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Some questions that arose in the development of this work, and that we believe are worth studying, are included below.", + "bbox": [ + 76, + 763, + 468, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- A triangulation approach inspired by [2] consists of reconstructing the 3D line $L$ by fitting it to $\\mathcal{L}_{\\mathcal{C}}$ solving the following optimization problem", + "bbox": [ + 94, + 806, + 468, + 852 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\text {m i n i m i z e} \\quad \\sum_ {i j} d \\left(x _ {i j}, \\ell_ {i}\\right) ^ {2} \\text {s u b j e c t t o} \\quad \\ell \\in \\mathcal {L} _ {\\mathcal {C}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 864, + 468, + 896 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f8fe7031949babb374f6c37d5097beacf08aef7df78d76dd17c816b360d720d4.jpg", + "table_caption": [ + "Table 2: Triangulation of $p = 5$ point correspondences incident to a line for $m = 3$ views with complete visibility. 1000 iterations. The tables show the accuracy and speed of the depicted triangulation methods in terms of median, mean and standard deviation." + ], + "table_footnote": [], + "table_body": "
Accuracymedianmeanσ
(L1).00.3850.5510.776
(L1).11.4371.7621.249
(L1).21.0951.0770.845
(L1).30.8171.1101.094
(L1).40.2030.4211.324
Speedmedianmeanσ
(L1).04.3654.6751.576
(L1).10.1410.1490.0431
(L1).21.8091.8630.323
(L1).31.1071.1570.234
(L1).46.5997.3913.068
", + "bbox": [ + 558, + 88, + 834, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: Triangulation of $p = 5$ point correspondences incident to a line for $m = 4$ views with complete visibility. 100 iterations. The tables show the accuracy and speed of the depicted triangulation methods in terms of median, mean and standard deviation.", + "bbox": [ + 500, + 306, + 890, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $x_{ij}$ denotes the $j = 1,\\dots ,p$ point correspondences across $i = 1,\\ldots ,m$ cameras, and $d(x_{ij},\\ell_i)$ is the affine distance from a line in $\\mathbb{R}^2$ to the point, meaning", + "bbox": [ + 529, + 378, + 890, + 439 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nd \\left(x _ {i j}, \\ell_ {i}\\right) = \\frac {\\left| 1 + \\left(\\ell_ {i}\\right) _ {1} \\left(x _ {i j}\\right) _ {1} + \\left(\\ell_ {i}\\right) _ {2} \\left(x _ {i j}\\right) _ {2} \\right|}{\\sqrt {\\left(\\ell_ {i}\\right) _ {1} ^ {2} + \\left(\\ell_ {i}\\right) _ {2} ^ {2}}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 446, + 890, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We believe that this method is more accurate than (L1).4, but with longer running time, due to the appearance of fractions.", + "bbox": [ + 529, + 489, + 890, + 532 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Consider the point-line problem (P1) defined by one point incident to $l$ lines, illustrated in Figure 1. Can similar approaches to (L1).1-4 be formulated for the triangulation of the (P1) setting? Can they be used to improve accuracy and/or the speed of triangulation?", + "- Our implementation for the numerical experiments does not directly correspond to the specialized software and hardware used for triangulation in practice. However, our results motivate that patterns displayed in our figures and tables could translate to such settings. This could improve the efficiency of current specialized solvers." + ], + "bbox": [ + 517, + 542, + 890, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. The authors thank Kathlen Kohn for helpful discussions, Lukas Gustafsson for the initial insight of Theorem 1.3, and Viktor Larsson for pointing out relevant references at the start of this project. Felix Rydell and Angelica Torres were supported by the Knut and Alice Wallenberg Foundation within their WASP (Wallenberg AI, Autonomous Systems and Software Program) AI/Math initiative. Elima Shehu is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), Projektnummer 445466444. Angelica Torres is currently supported by the Spanish State Research Agency, through", + "bbox": [ + 498, + 750, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3755", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the Severo Ochoa and María de Maeztu Program for Centers and Units of Excellence in R&D.", + "bbox": [ + 76, + 92, + 468, + 119 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 138, + 171, + 152 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sameer Agarwal, Andrew Pryhuber, and Rekha R Thomas. Ideals of the multiview variety. IEEE transactions on pattern analysis and machine intelligence, 2019. 1, 2", + "[2] Adrien Bartoli and Peter Sturm. Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer Vision and Image Understanding, 100(3):416-441, 2005. 3, 8", + "[3] Paul A Beardsley, Andrew Zisserman, and David William Murray. Navigation using affine structure from motion. Lecture Notes in Computer Science, 801:85-96, 1994. 2, 3", + "[4] Paul A Beardsley, Andrew Zisserman, and David William Murray. Sequential updating of projective and affine structure from motion. International journal of computer vision, 23:235-259, 1997. 2, 3", + "[5] Jeff Bezanson, Stefan Karpinski, Viral B Shah, and Alan Edelman. Julia: A fast dynamic language for technical computing. arXiv preprint arXiv:1209.5145, 2012. 5", + "[6] Paul Breiding, Timothy Duff, Lukas Gustafsson, Felix Rydell, and Elima Shehu. Line multiview ideals. arXiv preprint arXiv:2303.02066, 2023. 2", + "[7] Paul Breiding, Felix Rydell, Elima Shehu, and Angélica Torres. Line multiview varieties. SIAM Journal on Applied Algebra and Geometry, 7(2):470-504, 2023. 1, 2, 4", + "[8] Paul Breiding and Sascha Timme. Homotopycontinuation. jl: A package for homotopy continuation in julia. In International Congress on Mathematical Software, pages 458-465.Springer,2018.5", + "[9] Paul Breiding and Sascha Timme. HomotopyContinuation.jl: A Package for Homotopy Continuation in Julia. In Mathematical Software - ICMS 2018, pages 458-465, Cham, 2018. Springer International Publishing. 2, 6", + "[10] Tom Breloff and other contributors. JuliaPlots/Plots.jl. 6", + "[11] Jan Draisma, Emil Horobet, Giorgio Ottaviani, Bernd Sturmfels, and Rekha R Thomas. The euclidean distance degree of an algebraic variety. Foundations of computational mathematics, 16(1):99-149, 2016. 2, 4", + "[12] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajdla. Plmp-point-line minimal problems in complete multi-view visibility. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1675-1684, 2019. 1", + "[13] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajda. PI\\$\\_1\\$\\$p - point-line minimal problems under partial visibility in three views. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 175–192, Cham, 2020. Springer International Publishing. 1", + "[14] A. Elqursh and A. Elgammal. Line-based relative pose estimation. In Conference on Computer Vision and Pattern Recognition, 2011. 3" + ], + "bbox": [ + 78, + 164, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Ricardo Fabbri, Timothy Duff, Hongyi Fan, Margaret H Regan, David da Costa de Pinho, Elias Tsigaridas, Charles W Wampler, Jonathan D Hauenstein, Peter J Giblin, Benjamin Kimia, et al. Trplp-trifocal relative pose from lines at points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12073-12083, 2020. 1", + "[16] Olivier Faugeras. Three-dimensional computer vision: A geometric viewpoint. MIT press, Cambridge, MA, 1993. 1", + "[17] Olivier Faugeras and Bernard Mourrain. On the geometry and algebra of the point and line correspondences between n images. In Proceedings of IEEE International Conference on Computer Vision, pages 951-956. IEEE, 1995. 1, 2", + "[18] Andreas Gathmann. Algebraic geometry, 2019/20. Class Notes TU Kaiserslautern. Available at https://www.mathematik.uni-kl.de/~gathmann/de/alggeom.php.4", + "[19] Corey Harris and Daniel Lowengrub. The Chern-mather class of the multiview variety. Communications in Algebra, 46(6):2488-2499, 2018. 2", + "[20] Richard I. Hartley. Lines and points in three views and the trifocal tensor. International Journal of Computer Vision, 22(2):125-140, 1997. 2, 3", + "[21] Richard I Hartley and Peter Sturm. Triangulation. Computer vision and image understanding, 68(2):146-157, 1997. 2", + "[22] Richard I. Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 2", + "[23] Manuel Hofer, Andreas Wendel, and Horst Bischof. Incremental line-based 3d reconstruction using geometric constraints. The British Machine Vision Association and Society for Pattern Recognition, 2013. 3", + "[24] Kun Huang, Yifan Wang, Zihan Zhou, Tianjiao Ding, Shenghua Gao, and Yi Ma. Learning to parse wireframes in images of man-made environments. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 626-635, 2018. 1", + "[25] Javier Civera Joan Solà, Teresa Vidal-Calleja and José María Martínez Montiel. Impact of landmark parametrization on monocular ekf-slam with points and lines. International Journal of Computer Vision, 97(3):339-368, 2012. 3, 9", + "[26] Kenichi Kanatani. Statistical optimization for geometric computation: theory and practice. Courier Corporation, 2005. 2", + "[27] Kenichi Kanatani, Yasuyuki Sugaya, and Hirotaka Niitsuma. Triangulation from two views revisited: Hartley-sturm vs. optimal correction. practice, 4(5):99, 2008. 2", + "[28] Yasushi Kanazawa and Kenichi Kanatani. Reliability of 3- d reconstruction by stereo vision. IEICE TRANSACTIONS on Information and Systems, 78(10):1301-1306, 1995. 2", + "[29] Joseph David Kileel. Algebraic Geometry for Computer Vision. ProQuest LLC, Ann Arbor, MI, 2017. Thesis (Ph.D.)-University of California, Berkeley. 1, 2", + "[30] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21445-21455, June 2023. 1" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3756", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Laurentiu G. Maxim, Jose I. Rodriguez, and Botong Wang. Euclidean distance degree of the multiview variety. SIAM Journal on Applied Algebra and Geometry, 4(1):28-48, 2020. 1, 2", + "[32] Branislav Micusik and Horst Wildenauer. Structure from motion with line segments under relaxed endpoint constraints. In 2014 2nd International Conference on 3D Vision, volume 1, pages 13-19, 2014. 3", + "[33] Long Quan and T. Kanade. Affine structure from line correspondences with uncalibrated affine cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(8):834-845, 1997. 3", + "[34] Grant Schindler, Panchapagesan Krishnamurthy, and Frank Dellaert. Line-based structure from motion for urban environments. Third International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT'06), pages 846-853, 2006. 1", + "[35] Henrik Stewénius, Frederik Schaffalitzky, and David Nister. How hard is 3-view triangulation really? In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 1, pages 686-693. IEEE, 2005. 3", + "[36] Matthew Trager, Martial Hebert, and Jean Ponce. The joint image handbook. In Proceedings of the IEEE international conference on computer vision, pages 909-917, 2015. 2, 4" + ], + "bbox": [ + 78, + 90, + 470, + 431 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3757", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/5fbfedc8-4ebc-4e53-8cf0-a4029a8a49ba_model.json b/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/5fbfedc8-4ebc-4e53-8cf0-a4029a8a49ba_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e899d13a824325a3ab1dd36fe412b835f7b07933 --- /dev/null +++ b/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/5fbfedc8-4ebc-4e53-8cf0-a4029a8a49ba_model.json @@ -0,0 +1,2706 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.087, + 0.131, + 0.885, + 0.175 + ], + "angle": 0, + "content": "Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.204, + 0.348, + 0.222 + ], + "angle": 0, + "content": "Felix Rydell" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.223, + 0.439, + 0.241 + ], + "angle": 0, + "content": "KTH Royal Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.241, + 0.458, + 0.257 + ], + "angle": 0, + "content": "Lindstedtsvagen 25, Stockholm, Sweden" + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.26, + 0.361, + 0.274 + ], + "angle": 0, + "content": "felixry@kth.se" + }, + { + "type": "text", + "bbox": [ + 0.616, + 0.205, + 0.722, + 0.221 + ], + "angle": 0, + "content": "Elima Shehu" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.222, + 0.822, + 0.239 + ], + "angle": 0, + "content": "University of Osnabrück and MPI MiS" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.24, + 0.835, + 0.257 + ], + "angle": 0, + "content": "Albrechtstrasse 28a, Osnabrück, Germany" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.258, + 0.804, + 0.276 + ], + "angle": 0, + "content": "Inselstrasse 22, Leipzig, Germany" + }, + { + "type": "text", + "bbox": [ + 0.57, + 0.278, + 0.766, + 0.292 + ], + "angle": 0, + "content": "elima.shehu@mis.mpg.de" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.3, + 0.551, + 0.318 + ], + "angle": 0, + "content": "Angélica Torres" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.319, + 0.608, + 0.334 + ], + "angle": 0, + "content": "Centre de Recerca Matemàtica" + }, + { + "type": "text", + "bbox": [ + 0.319, + 0.336, + 0.653, + 0.353 + ], + "angle": 0, + "content": "Edifici C, Campus UAB, Bellaterra, Spain" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.356, + 0.553, + 0.368 + ], + "angle": 0, + "content": "atorres@crm.cat" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.405, + 0.314, + 0.421 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.473, + 0.635 + ], + "angle": 0, + "content": "We study the joint image of lines incident to points, meaning the set of image tuples obtained from fixed cameras observing a varying 3D point-line incidence. We prove a formula for the number of complex critical points of the triangulation problem that aims to compute a 3D point-line incidence from noisy images. Our formula works for an arbitrary number of images and measures the intrinsic difficulty of this triangulation. Additionally, we conduct numerical experiments using homotopy continuation methods, comparing different approaches of triangulation of such incidences. In our setup, exploiting the incidence relations gives a notably faster point reconstruction with comparable accuracy." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.664, + 0.19, + 0.68 + ], + "angle": 0, + "content": "Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.825 + ], + "angle": 0, + "content": "The Structure-from-Motion pipeline aims to estimate the camera positions and the relative position of the objects present in a given set of images. The process starts by identifying point and line features in one image that are recognizable as the same points or lines in another. These matched features, that we call correspondences, are used to estimate the camera position and orientation, which then allow for the triangulation step where the position of the points and lines is estimated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this work, we study the triangulation of points and lines satisfying a certain incidence relation. Specifically, we focus on the triangulation of points contained in a single line; see Figure 1. We note that our methods can also be used when multiple lines going through a fixed point. We" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.408, + 0.818, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.518, + 0.895, + 0.568 + ], + "angle": 0, + "content": "Figure 1: The illustration of two different types of incidence relations: (L1) represents the scenario where multiple points are incident to a line, while (P1) represents the scenario where multiple lines are incident to a point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.642 + ], + "angle": 0, + "content": "develop the theory for the triangulation of these problems for an arbitrary number of pinhole cameras assuming complete visibility." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.893, + 0.824 + ], + "angle": 0, + "content": "The triangulation of points is well understood and, in practice, it is efficiently implemented. However, to the best of our knowledge, incidence relations are not considered in these implementations. The inclusion of line features and incidence relations in the triangulation process could give a more accurate estimation of scenes as lines are more robust to noise than points, and incidence relations appear frequently in interior scenes and man-made scenarios; see [34], [24], [30]. Additionally, in some real data sets standard feature detection algorithms fail due to a lack of point correspondences but succeed when line correspondences are taken into account [15]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The main tools in our work are (algebraic) varieties, that is, the vanishing sets of systems of polynomial equations. Algebraic varieties have been used extensively to study triangulation of point correspondences [1, 17, 31], line correspondences [7, 16, 29] and minimal problems [12, 13] to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3748" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "name a few. In these works, algebraic varieties arise naturally due to the algebraic nature of pinhole cameras." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.47, + 0.198 + ], + "angle": 0, + "content": "A pinhole camera, is modeled by a (complex) projective linear map \\(C: \\mathbb{P}^3 \\dashrightarrow \\mathbb{P}^2\\) defined by a \\(3 \\times 4\\) matrix \\(C\\) of full rank that takes a point \\(X \\in \\mathbb{P}^3\\) and sends it to \\(CX \\in \\mathbb{P}^2\\). A camera arrangement with \\(m \\geq 2\\) cameras is denoted by \\(\\mathcal{C} = (C_1, \\ldots, C_m)\\), and the joint camera map" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.209, + 0.47, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Phi_ {\\mathcal {C}}: \\mathbb {P} ^ {3} \\dashrightarrow (\\mathbb {P} ^ {2}) ^ {m}, \\tag {1} \\\\ X \\mapsto (C _ {1} X, \\dots , C _ {m} X), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.257, + 0.471, + 0.394 + ], + "angle": 0, + "content": "models the process of taking the images of a point \\(X\\) in homogeneous coordinates with \\(m\\) cameras. For fixed cameras, the (point) multiview variety \\(\\mathcal{M}_{\\mathcal{C}}\\) is the smallest variety that contains all point correspondences. An extensive account of the pinhole cameras is given by [22], and a survey of the multiview variety is found in [36]. The joint camera map can be extended from the space of points \\(\\mathbb{P}^3\\) to the space of lines, denoted by \\(\\mathrm{Gr}(1,\\mathbb{P}^3)\\), where each line is parametrized as the span of two points. The map" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.403, + 0.47, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Upsilon_ {\\mathcal {C}}: \\operatorname {G r} (1, \\mathbb {P} ^ {3}) \\dashrightarrow (\\mathbb {P} ^ {2}) ^ {m}, \\tag {2} \\\\ L \\mapsto (C _ {1} \\cdot L, \\dots , C _ {m} \\cdot L). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.47, + 0.557 + ], + "angle": 0, + "content": "models the image of a line \\(L\\) taken by the cameras of \\(\\mathcal{C}\\). To be precise, if \\(u, v\\) span the line \\(L\\) in \\(\\mathbb{P}^3\\), then \\(C \\cdot L\\) is defined as \\(\\ell = Cu \\times Cv \\in \\mathbb{P}^2\\), where \\(\\times\\) is the cross product. Observe that \\(\\{x \\in \\mathbb{P}^2 : \\ell^T x = 0\\}\\) equals \\(\\operatorname{span}\\{Cu, Cv\\}\\). Recently, in [7], the authors study the line multiview variety referred to as \\(\\mathcal{L}_{\\mathcal{C}}\\), which is the smallest variety containing all line correspondences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.47, + 0.678 + ], + "angle": 0, + "content": "Our main contribution is the definition and study of the anchored point and line multiview varieties. Given a fixed line \\(L\\) in \\(\\mathbb{P}^3\\), the anchored point multiview variety, \\(\\mathcal{M}_C^L\\), is defined as the smallest variety containing all point correspondences coming from points in \\(L\\); and the anchored line multiview variety, \\(\\mathcal{L}_C^X\\), is the smallest variety containing the line correspondences coming from lines passing through a given point \\(X \\in \\mathbb{P}^3\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.47, + 0.739 + ], + "angle": 0, + "content": "For these new varieties we prove formulas for their Euclidean distance degree (EDD), which are a measurement of the complexity for error correction using exact algebraic methods [11]. Specifically, we prove the following theorem:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.779 + ], + "angle": 0, + "content": "Theorem 0.1. Let \\(\\mathcal{C}\\) be a generic arrangement of \\(m\\) cameras." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.27, + 0.806 + ], + "angle": 0, + "content": "1. \\(\\operatorname{EDD}(\\mathcal{M}_{\\mathcal{C}}^{L}) = 3m - 2\\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.813, + 0.424, + 0.833 + ], + "angle": 0, + "content": "2. If \\(m \\geq 3\\), then \\(\\mathrm{EDD}(\\mathcal{L}_{\\mathcal{C}}^{X}) = \\frac{9}{2} m^{2} - \\frac{19}{2} m + 3\\)." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.788, + 0.424, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We provide a precise definition of this degree in Section 1.2. Previous work on EDD for multiview varieties includes [19, 31] and [7, Section 5]. The EDD of the point multiview variety is \\(\\mathrm{EDD}(\\mathcal{M}_{\\mathcal{C}}) = \\frac{9}{2} m^3 -\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.196 + ], + "angle": 0, + "content": "\\(\\frac{21}{2} m^2 + 8m - 4\\), according to [31]. The fact that the EDDs of the anchored multiview varieties are polynomials of smaller degrees suggests that they are less complex for the purpose of data correction. This conclusion is backed by the results of our numerical experiments using HomotopyContinuation.jl [9], presented in Section 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.893, + 0.44 + ], + "angle": 0, + "content": "Another contribution of our work is the numerical simulations for the triangulation of points contained in a line using the anchored multiview varieties. We present different approaches to triangulating data of type (L1) that differ from the traditional method of fitting point correspondences to the multiview variety \\(\\mathcal{M}_{\\mathcal{C}}\\). For \\(m = 2\\) views, our implementation is notably faster than the traditional triangulation of points described above, while the accuracy is comparable. For \\(m = 3\\) views we get both higher accuracy and faster speed by using \\(\\mathcal{L}_{\\mathcal{C}}\\), compared to usual point triangulation. All of our proposed methods outperform the traditional approaches in terms of run-time and give a comparably accurate result. In practice, special software and hardware are used for triangulation, and based on our experiments, we believe that these approaches could be implemented efficiently with good results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.449, + 0.61, + 0.464 + ], + "angle": 0, + "content": "Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.624 + ], + "angle": 0, + "content": "Algebra. Algebraic geometry, whose connection to computer vision is well established, is our main tool for the theoretical study of the triangulation of problem (L1). Fundamental theory regarding the point multiview variety is found in [17, 20, 22], in particular for two and three views. The paper [36] by Trager et. al. serves as a survey for the algebraic properties of this multiview variety. For applications, finding a good set of polynomial constraints satisfied by the point correspondences is important, this is precisely the work of [1] by Agarwal et al." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.776 + ], + "angle": 0, + "content": "Regarding the algebra of lines in a computer vision setting, Kileel presented in [29, Definition 3.9, Theorem 3.10] several types of multiview varieties with respect to a point-line incidence relation in \\(\\mathbb{P}^3\\) in 3 views, and their basic properties. One of those types is called the LLL-multiview variety and constitutes a special case of the more recent work by Breiding et. al. [7], where they study algebraic properties of the line multiview variety. Furthermore, a recent manuscript [6] studies the polynomial constraints satisfied by line correspondences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Projective Reconstruction. Different approaches to the triangulation of points have been considered in the literature, in particular for two views [3, 4, 21, 26-28]. Hartley and Sturm compare many different approaches in [21], including the \"midpoint\" method. The midpoint method inputs a point correspondence in two views and outputs the midpoint of the line segment determined by the points on" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3749" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "the two back-projected lines that are closest to each other. In other words, it finds the midpoint of the common perpendicular to the two back-projected lines. This is a natural approach, but it has several downsides, also pointed out in [3, 4]. Importantly, it is not projectively invariant. A projectively invariant reconstruction has the property that acting on the cameras by a global projective transformation induces the same action on the triangulated point or line. Instead, Hartley and Sturm propose minimizing the reprojection error as the optimal triangulation method, which is accepted as a standard formulation [35] and it is projectively invariant. In [35], the focus is on triangulation via minimizing reprojections errors in three views and the authors point out that three views often lead to greater stability and stronger disambiguation compared to two views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.71 + ], + "angle": 0, + "content": "Implementation of Line Reconstruction. From the algorithmic point of view, the simultaneous reconstruction of point and line features have been studied specially with the goal of reconstructing line segments. For example, [2] provides a thorough overview of the structure-from-motion pipeline using lines, going through different methods for line parameterization and error correction depending on such parameterizations. In [33], Quan and Takeo study the algebraic structure of line correspondences with uncalibrated affine cameras. They reduce the problem of reconstructing affine lines to the reconstruction of projective points in a lower-dimensional projective camera. In [20], Hartley and coauthors provide an algorithm for the reconstruction of point and line features where 3D lines are parameterized by their projections in 2 views (as the intersection of two back-projected planes). Micusik and Wildenauer reconstruct lines to estimate line segments, but only use incident points for error correction [32]. Furthermore in [23], the authors propose a technique for 3D reconstruction incorporating line segments. They assert that this method surpasses traditional approaches, especially in efficiency while getting accurate results. Finally, in [14], the authors present a framework for the computation of the relative motion between two images using a triplet of lines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "This paper is structured as follows. In the first part of Section 1 we formally introduce anchored multiview varieties and study some properties. In the second part of Section 1, we study their smoothness and find their Euclidean Distance Degree. Finally, in Section 2, we provide a numerical analysis of different approaches to reconstructing point correspondences incident to a line correspondence. The proofs of all our results are included in the Supplementary Material together with a small background on the concepts of Algebraic Geometry used for these results and the code for the numerical experiments." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.776, + 0.107 + ], + "angle": 0, + "content": "1. Anchored Multiview Varieties" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.894, + 0.16 + ], + "angle": 0, + "content": "A camera refers to a full-rank \\(3 \\times 4\\) matrix. A camera arrangement is a collection \\(\\mathcal{C} = (C_1, \\ldots, C_m)\\) of \\(m \\geq 2\\) cameras whose centers, i.e. kernels, are all distinct." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.161, + 0.895, + 0.314 + ], + "angle": 0, + "content": "A variety is the solution set to a system of polynomial equations. The Zariski closure of a set \\(U\\) is the smallest variety containing \\(U\\). We work in complex projective space, denoted \\(\\mathbb{P}^n\\). This is defined as \\((\\mathbb{C}^n \\setminus \\{0\\}) / \\sim\\), where \\(x \\sim y\\) if \\(x\\) and \\(y\\) differ by a non-zero constant. The set of lines in \\(\\mathbb{P}^n\\) is denoted as \\(\\operatorname{Gr}(1, \\mathbb{P}^n)\\). In the language of algebraic geometry, it is called the Grassmanian of lines in \\(\\mathbb{P}^n\\). In general we denote with upper case letters world objects and with lower case letters image objects. For a rational map \\(\\varphi: X \\dashrightarrow Y\\), we denote \\(\\varphi|_A\\) the restriction of \\(\\varphi\\) to \\(A \\subseteq X\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.894, + 0.362 + ], + "angle": 0, + "content": "Definition 1.1. Let \\( X \\in \\mathbb{P}^3 \\) be a point distinct from all camera centers, and \\( L \\) a line in \\( \\mathbb{P}^3 \\) containing none of the camera centers." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.368, + 0.892, + 0.4 + ], + "angle": 0, + "content": "1. The anchored point multiview variety, denoted \\(\\mathcal{M}_C^L\\), is defined as the Zariski closure of the image of the map" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.404, + 0.892, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c} \\Phi_ {\\mathcal {C}} | _ {L}: L \\dashrightarrow (\\mathbb {P} ^ {2}) ^ {m}, \\\\ X \\longmapsto (C _ {1} X, \\dots , C _ {m} X). \\end{array} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.892, + 0.483 + ], + "angle": 0, + "content": "2. The anchored line multiview variety, denoted \\(\\mathcal{L}_C^X\\), is defined as the Zariski closure of the image of the map" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.487, + 0.891, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c} \\Upsilon_ {\\mathcal {C}} | _ {\\Lambda (X)}: \\Lambda (X) \\dashrightarrow (\\mathbb {P} ^ {2}) ^ {m}, \\\\ L \\longmapsto \\left(C _ {1} \\cdot L, \\dots , C _ {m} \\cdot L\\right), \\end{array} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.53, + 0.892, + 0.561 + ], + "angle": 0, + "content": "where \\(\\Lambda(X)\\) denotes the set of lines in \\(\\mathbb{P}^3\\) that contain \\(X\\) and \\(\\ell_i = C_i \\cdot L\\) is defined as in the introduction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.687 + ], + "angle": 0, + "content": "The anchored point multiview variety \\(\\mathcal{M}_{\\mathcal{C}}^{L}\\) is the smallest variety that contains all point correspondences \\(\\Phi_{\\mathcal{C}}(X)\\) for \\(X \\in L\\). Similarly, the anchored line multiview variety \\(\\mathcal{L}_{\\mathcal{C}}^{X}\\) is the smallest variety that contains all line correspondences \\(\\Upsilon_{\\mathcal{C}}(L)\\) for \\(L\\) meeting \\(X\\) and no center. We highlight that our definition of anchored multiview varieties is different from the anchored features in [25] for monocular EKF-SLAM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.693, + 0.892, + 0.739 + ], + "angle": 0, + "content": "Proposition 1.2. Consider an arrangement of \\(m\\) cameras \\(\\mathcal{C} = (C_1, \\ldots, C_m)\\), a point \\(X \\in \\mathbb{P}^3\\) and a line \\(L\\) in \\(\\mathbb{P}^3\\) satisfying the conditions of Definition 1.1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.744, + 0.892, + 0.775 + ], + "angle": 0, + "content": "1. If there are two different camera centers \\( c_{i} \\) and \\( c_{j} \\) such that the span of \\( \\{c_{i}, c_{j}, L\\} \\) is \\( \\mathbb{P}^3 \\), then" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.779, + 0.892, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {\\mathcal {C}} ^ {L} = \\left\\{\\left(x _ {1}, \\dots , x _ {m}\\right) \\in \\mathcal {M} _ {\\mathcal {C}}: x _ {i} \\in C _ {i} \\cdot L \\right\\}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.892, + 0.838 + ], + "angle": 0, + "content": "2. If for each camera center \\( c_i \\), the line spanned by \\( c_i \\) and \\( X \\) does not contain any other camera center, then" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.843, + 0.892, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {C}} ^ {X} = \\left\\{\\left(\\ell_ {1}, \\dots , \\ell_ {m}\\right) \\in \\mathcal {L} _ {\\mathcal {C}}: C _ {i} X \\in \\ell_ {i} \\right\\}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The proofs for this and all our results are found in the Supplementary Material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3750" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.446, + 0.107 + ], + "angle": 0, + "content": "1.1. Properties of Anchored Multiview Varieties" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.468, + 0.189 + ], + "angle": 0, + "content": "A fundamental property of the anchored multiview varieties is that they are linearly isomorphic to multiview varieties arising from projections \\(\\mathbb{P}^1\\longrightarrow \\mathbb{P}^1\\) and \\(\\mathbb{P}^2\\longrightarrow \\mathbb{P}^1\\) respectively. A linear isomorphism is a linear map (given by a matrix) with an inverse that is also linear." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.189, + 0.468, + 0.251 + ], + "angle": 0, + "content": "Consider an arrangement \\(\\widetilde{\\mathcal{C}}\\) of \\(m\\) full-rank \\(2\\times 2\\) matrices, and an arrangement \\(\\widehat{\\mathcal{C}}\\) of \\(m\\) full-rank \\(2\\times 3\\) matrices. Also here we assume \\(m\\geq 2\\). We define \\(\\mathcal{M}_{\\widetilde{\\mathcal{C}}}^{1,1}\\), and \\(\\mathcal{M}_{\\widehat{\\mathcal{C}}}^{2,1}\\) respectively as the Zariski closure of the image of the maps" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.269, + 0.468, + 0.295 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {\\tilde {C}}: \\mathbb {P} ^ {1} \\longrightarrow (\\mathbb {P} ^ {1}) ^ {m}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.291, + 0.378, + 0.309 + ], + "angle": 0, + "content": "\\[\nX \\longmapsto (\\widetilde {C} _ {1} X, \\dots , \\widetilde {C} _ {m} X)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.316, + 0.107, + 0.328 + ], + "angle": 0, + "content": "and" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.335, + 0.468, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {\\widehat {C}}: \\mathbb {P} ^ {2} \\dashrightarrow (\\mathbb {P} ^ {1}) ^ {m}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.356, + 0.376, + 0.374 + ], + "angle": 0, + "content": "\\[\nX \\longmapsto (\\widehat {C} _ {1} X, \\dots \\widehat {C} _ {m} X).\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.381, + 0.175, + 0.395 + ], + "angle": 0, + "content": "Theorem 1.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.47, + 0.452 + ], + "angle": 0, + "content": "1. Let \\(\\phi_L: L \\to \\mathbb{P}^1\\) and \\(\\psi_{\\mathcal{C},i}: C_i \\cdot L \\to \\mathbb{P}^1\\) be any choices of linear isomorphisms. Let \\(\\tilde{\\mathcal{C}}\\) denote the arrangement of matrices \\(C_i := \\psi_{\\mathcal{C},i} \\circ C_i \\circ \\phi_L^{-1}\\). Then" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.471, + 0.468, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\psi_ {\\mathcal {C}, L} := \\left(\\psi_ {\\mathcal {C}, 1}, \\dots , \\psi_ {\\mathcal {C}, m}\\right): \\mathcal {M} _ {\\mathcal {C}} ^ {L} \\rightarrow \\mathcal {M} _ {\\tilde {\\mathcal {C}}} ^ {1, 1} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.499, + 0.273, + 0.513 + ], + "angle": 0, + "content": "is a linear isomorphism." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.521, + 0.469, + 0.571 + ], + "angle": 0, + "content": "2. Let \\(\\phi_X: \\Lambda(X) \\to \\mathbb{P}^2\\) and \\(\\psi_{\\mathcal{C},i}: \\Lambda(C_iX) \\to \\mathbb{P}^1\\) be any choices of linear isomorphisms. Let \\(\\widehat{\\mathcal{C}}\\) denote the arrangement of matrices \\(\\widehat{C}_i \\coloneqq \\psi_{\\mathcal{C},i} \\circ C_i \\circ \\phi_X^{-1}\\). Then" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.59, + 0.468, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\psi_ {\\mathcal {C}, X} := \\left(\\psi_ {\\mathcal {C}, 1}, \\dots , \\psi_ {\\mathcal {C}, m}\\right): \\mathcal {L} _ {\\mathcal {C}} ^ {X} \\rightarrow \\mathcal {M} _ {\\widehat {\\mathcal {C}}} ^ {2, 1} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.617, + 0.273, + 0.632 + ], + "angle": 0, + "content": "is a linear isomorphism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.468, + 0.721 + ], + "angle": 0, + "content": "In Section 2.1.1, we exploit these linear isomorphisms to improve the speed of triangulation for our setting. A consequence of this theorem is also that many results for the anchored multiview varieties translate into results about \\(\\mathcal{M}_{\\hat{\\mathcal{C}}}^{1,1}\\) and \\(\\mathcal{M}_{\\hat{\\mathcal{C}}}^{2,1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.727, + 0.448, + 0.744 + ], + "angle": 0, + "content": "Proposition 1.4. \\(\\mathcal{M}_C^L\\) and \\(\\mathcal{L}_C^X\\) are irreducible. Further," + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.466, + 0.766 + ], + "angle": 0, + "content": "1. \\(\\mathcal{M}_C^L\\) is isomorphic to \\(\\mathbb{P}^1\\). In particular, \\(\\dim \\mathcal{M}_C^L = 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.773, + 0.469, + 0.804 + ], + "angle": 0, + "content": "2. If the span of the centers \\( c_{i} \\) and the point \\( X \\) are not collinear, then \\( \\dim \\mathcal{L}_{\\mathcal{C}}^{X} = 2 \\)." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.749, + 0.469, + 0.804 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Using the fundamental matrices of \\( C_i \\) and \\( C_j \\), denoted as \\( F^{ij} \\), we introduce a set of polynomial constraints that must be fulfilled by the correspondences of points or lines in the anchored multiview varieties. These constraints imply that some of the equations obtained from Proposition 1.2 are redundant in the generic case." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.122 + ], + "angle": 0, + "content": "Proposition 1.5. For a point \\(X \\in \\mathbb{P}^3\\) and line \\(L\\) in \\(\\mathbb{P}^3\\), let \\(\\mathcal{C}\\) be a generic (random) camera arrangement of \\(m\\) cameras." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.131, + 0.892, + 0.164 + ], + "angle": 0, + "content": "1. \\(x \\in \\mathcal{M}_{\\mathcal{C}}^{L}\\) if and only if \\(x_{1}^{T}F^{1j}x_{j} = 0\\) for every \\(j = 2, \\ldots, m\\) and \\(x_{i}^{T}C_{i} \\cdot L = 0\\) for every \\(i = 1, \\ldots, m\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.172, + 0.679, + 0.189 + ], + "angle": 0, + "content": "2. \\(\\ell \\in \\mathcal{L}_C^X\\) if and only if" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.131, + 0.892, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.2, + 0.892, + 0.225 + ], + "angle": 0, + "content": "\\[\n\\det \\left[ \\begin{array}{l l l} C _ {1} ^ {T} \\ell_ {1} & C _ {2} ^ {T} \\ell_ {2} & C _ {i} ^ {T} \\ell_ {i} \\end{array} \\right] = 0, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.22, + 0.82, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\det \\left[ \\begin{array}{c c c} C _ {1} ^ {T} \\ell_ {1} & C _ {3} ^ {T} \\ell_ {3} & C _ {i} ^ {T} \\ell_ {i} \\end{array} \\right] = 0\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.25, + 0.892, + 0.281 + ], + "angle": 0, + "content": "for \\(i = 3,\\ldots ,m\\) and \\(\\ell_i^T C_iX = 0\\) for every \\(i = 1,\\dots ,m\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.337 + ], + "angle": 0, + "content": "The determinantal constraints described in the second item correspond to the constraints that are satisfied by elements of the line multiview variety, presented in [7]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.398 + ], + "angle": 0, + "content": "In Supplementary Material Section B we define the multidegree of a variety in a product of projective spaces, determine it for the anchored multiview varieties and explain its relevance for computer vision." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.408, + 0.787, + 0.424 + ], + "angle": 0, + "content": "1.2. The Euclidean Distance problem" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.477 + ], + "angle": 0, + "content": "For a variety \\(\\mathcal{X} \\subseteq \\mathbb{R}^n\\) and a point \\(u \\in \\mathbb{R}^n\\) outside the variety, a natural problem is to find the closest point on \\(\\mathcal{X}\\) to \\(u\\), which corresponds to the optimization problem" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.489, + 0.891, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\text {m i n i m i z e} \\quad \\sum_ {i = 1} ^ {n} \\left(u _ {i} - x _ {i}\\right) ^ {2} \\quad \\text {s u b j e c t t o} \\quad x \\in \\mathcal {X}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Equation (12) is called the Euclidean distance problem and models the process of error correction and fitting noisy data to a mathematical model \\(\\mathcal{X}\\). In the case of a smooth variety, defined below, the Euclidean distance degree (EDD) is the number of complex solutions to the critical equations of (12) [11]. The EDD is an estimate of how difficult it is to solve this problem by exact algebraic methods. For a variety \\(\\mathcal{X}\\) in a product of projective spaces \\(\\mathbb{P}^{n_1} \\times \\dots \\times \\mathbb{P}^{n_m}\\), the EDD is the EDD of \\(\\mathcal{X} \\cap U_1 \\times \\dots \\times U_m \\subseteq \\mathbb{R}^{n_1 + \\dots + n_m}\\), where \\(U_i\\) is a generic affine patch of \\(\\mathbb{P}^{h_i}\\) for each \\(i\\). An affine patch \\(U\\) of \\(\\mathbb{P}^n\\) is a subset defined by an affine equation with non-zero constant part, for instance \\(x_0 = 1\\). We have \\(U \\cong \\mathbb{R}^n\\) over the real numbers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.885 + ], + "angle": 0, + "content": "A variety \\(\\mathcal{X}\\) is smooth at a point \\(x\\) if \\(\\mathcal{X}\\) locally around \\(x\\) looks like Euclidean space. \\(\\mathcal{X}\\) is smooth if all its points are smooth. The singular locus of a variety is the set of non-smooth points. See [18] for more details. The singular locus of the point multiview variety is well-understood, see for instance [36], and it is mostly understood for the line multiview variety, see [7, Section 3]. The proposition below guarantees that the anchored multiview varieties are smooth for generic camera arrangements, which helps us to compute its EDD." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3751" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.092, + 0.191, + 0.106 + ], + "angle": 0, + "content": "Proposition 1.6." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.199, + 0.133 + ], + "angle": 0, + "content": "1. \\(\\mathcal{M}_C^L\\) smooth." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.141, + 0.468, + 0.172 + ], + "angle": 0, + "content": "2. If there are exactly two cameras, or the centers together with the point \\(X\\) span \\(\\mathbb{P}^3\\), then \\(\\mathcal{L}_C^X\\) is smooth." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.116, + 0.468, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.181, + 0.468, + 0.331 + ], + "angle": 0, + "content": "For the EDD to be relevant in a particular setting, the Euclidean distance has to be a good measurement of distance. This is naturally the case for points in \\(\\mathbb{R}^2\\), i.e. an affine patch of \\(\\mathbb{P}^2\\). This extends to an affine patch of \\(\\mathcal{M}_{\\mathcal{C}}\\). It is perhaps less clear how to best measure distances between lines. Recall that we regard \\(\\mathcal{L}_{\\mathcal{C}}\\) as a subset of \\((\\mathbb{P}^2)^m\\). Indeed, we identify image lines with the point that defines its normal vector. We choose the Euclidean distance between these normal vectors (in affine patches of \\(\\mathbb{P}^2\\)) as our distance between lines." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.332, + 0.468, + 0.422 + ], + "angle": 0, + "content": "The theorem below presents our computations of the EDDs of the anchored multiview varieties. Additionally, our numerical computations using the HomotopyContinuation.jl [8] package in the julia [5] programming language have confirmed the accuracy of these formulas for \\( m \\leq 10 \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.468, + 0.46 + ], + "angle": 0, + "content": "Theorem 1.7. Let \\( \\mathcal{C} \\) be a generic arrangement of \\( m \\) cameras." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.269, + 0.485 + ], + "angle": 0, + "content": "1. \\(\\operatorname{EDD}(\\mathcal{M}_C^L) = 3m - 2\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.492, + 0.423, + 0.51 + ], + "angle": 0, + "content": "2. If \\(m \\geq 3\\), then \\(\\mathrm{EDD}(\\mathcal{L}_{\\mathcal{C}}^{X}) = \\frac{9}{2} m^{2} - \\frac{19}{2} m + 3\\)." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.468, + 0.423, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.516, + 0.468, + 0.566 + ], + "angle": 0, + "content": "We end with an implication of Theorem 1.3: There is a direct correspondence between the EDDs of the anchored multiview varieties and \\(\\mathcal{M}_{\\widetilde{\\mathcal{C}}}^{1,1}, \\mathcal{M}_{\\widehat{\\mathcal{C}}}^{2,1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.574, + 0.468, + 0.604 + ], + "angle": 0, + "content": "Theorem 1.8. Let \\(\\hat{C}\\) and \\(\\hat{C}\\) be generic arrangements of cardinality \\(m\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.277, + 0.631 + ], + "angle": 0, + "content": "1. \\(\\operatorname{EDD}(\\mathcal{M}_{\\widetilde{c}}^{1,1}) = 3m - 2.\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.639, + 0.435, + 0.66 + ], + "angle": 0, + "content": "2. If \\(m \\geq 3\\), then \\(\\mathrm{EDD}(\\mathcal{M}_{\\hat{\\mathcal{C}}}^{2,1}) = \\frac{9}{2} m^2 - \\frac{19}{2} m + 3\\)." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.611, + 0.435, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.669, + 0.302, + 0.687 + ], + "angle": 0, + "content": "2. Numerical Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.695, + 0.468, + 0.739 + ], + "angle": 0, + "content": "In this section we conduct numerical experiments of the triangulation process of incident point and lines. All the code can be found in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.74, + 0.468, + 0.845 + ], + "angle": 0, + "content": "By (L1) we refer to point-line arrangements in \\(\\mathbb{P}^3\\) of one line incident to \\(p\\) points, as depicted in Figure 1. Our goal is to reconstruct such arrangements, given point correspondences incident to a single line correspondence across \\(m\\) views. The following is a list of natural approaches for such triangulation given a camera arrangement \\(\\mathcal{C}\\). It is not necessarily a complete list." + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "(L1).0 Triangulate each point correspondence by fitting it to the point multiview variety \\(\\mathcal{M}_{\\mathcal{C}}\\), i.e. find the closest point correspondence in \\(\\mathcal{M}_{\\mathcal{C}}\\);" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.093, + 0.695, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.095, + 0.864, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.219, + 0.892, + 0.256 + ], + "angle": 0, + "content": "Figure 2: Comparison of (L1).0 and (L1).1 triangulation approaches. (L1).0 expects non-collinear points in the reconstruction, while (L1).1 produces collinear points." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.282, + 0.892, + 0.343 + ], + "angle": 0, + "content": "(L1).1 Reconstruct the 3D line \\(L\\) by back-projecting the image lines from two views. Triangulate the point correspondences by fitting them to the anchored multiview variety \\(\\mathcal{M}_{\\mathcal{C}}^{L}\\);" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.353, + 0.892, + 0.413 + ], + "angle": 0, + "content": "(L1).2 Triangulate two point correspondences by fitting them to \\(\\mathcal{M}_C\\) to get 3D points \\(X\\) and \\(Y\\). Let \\(L\\) be the line they span in \\(\\mathbb{R}^3\\). Triangulate the remaining point correspondences by fitting them to \\(\\mathcal{M}_C^L\\);" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.424, + 0.892, + 0.5 + ], + "angle": 0, + "content": "(L1).3 Triangulate one point correspondence by fitting it to \\(\\mathcal{M}_{\\mathcal{C}}\\) to get a 3D point \\(X\\). Reconstruct the 3D line \\(L\\) by fitting the line correspondence to the anchored variety \\(\\mathcal{L}_{\\mathcal{C}}^{X}\\). Triangulate the remaining point correspondences by fitting them to \\(\\mathcal{M}_{\\mathcal{C}}^{L}\\);" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.51, + 0.892, + 0.571 + ], + "angle": 0, + "content": "(L1).4 Reconstruct the 3D line \\(L\\) by fitting the line correspondence to the line multiview variety \\(\\mathcal{L}_{\\mathcal{C}}\\). Triangulate the point correspondences by fitting them to the anchored multiview variety \\(\\mathcal{M}_{\\mathcal{C}}^{L}\\)." + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.282, + 0.892, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.583, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Approaches (L1).1-4 take the incidence relations into account in the triangulation, whereas (L1).0 does not. Therefore, in contrast to (L1).1-4, the resulting triangulation of (L1).0 does not preserve the point-line incidences; see Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.668, + 0.66, + 0.684 + ], + "angle": 0, + "content": "2.1. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.692, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Our experiments are implemented in HomotopyContinuation.jl [8] in Julia [5]. Here we explain them in some detail. The full explanation can be found in Supplementary Material. We ran the code on a Intel(R) Core(TM) i5-8300H CPU running at 2.30GHz." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.807, + 0.815 + ], + "angle": 0, + "content": "2.1.1 Reducing the number of parameters" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.824, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We reduce the number of parameters involved in the problem of fitting data to \\(\\mathcal{M}_{\\mathcal{C}}^{L}\\), respectively \\(\\mathcal{L}_{\\mathcal{C}}^{X}\\), by translating it to and solving it for \\(\\mathcal{M}_{\\widetilde{\\mathcal{C}}}^{1,1}\\), respectively \\(\\mathcal{M}_{\\widehat{\\mathcal{C}}}^{2,1}\\). Details are found in Supplementary Material Section B. This translation corresponds to a reduction of parameters because an" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3752" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.468, + 0.145 + ], + "angle": 0, + "content": "affine patch of \\(\\mathcal{M}_{\\widehat{\\mathcal{C}}}^{1,1}\\) lives in \\((\\mathbb{R}^1)^m\\), while an affine patch of \\(\\mathcal{M}_{\\mathcal{C}}^{L}\\) lives in \\((\\mathbb{R}^2)^m\\). The analogous is true for \\(\\mathcal{M}_{\\widehat{\\mathcal{C}}}^{2,1}\\) and \\(\\mathcal{L}_{\\mathcal{C}}^{X}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.145, + 0.468, + 0.204 + ], + "angle": 0, + "content": "We call the methods that do not use this translation standard and for instance, write (L1).1 std. to denote it. By just writing (L1).1, we denote the implementation of this translation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.227, + 0.248, + 0.243 + ], + "angle": 0, + "content": "2.1.2 Evaluating error" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.252, + 0.47, + 0.388 + ], + "angle": 0, + "content": "In our experiments, each iteration starts by randomly generating a 3D line \\(L\\) and \\(p\\) points \\(X_{i}\\) on this line. This line and these points are projected by randomly generated cameras to line and point correspondences in \\((\\mathbb{R}^2)^m\\). For fixed \\(\\epsilon = 10^{-12}\\), randomly generated noise vectors \\(\\sigma (\\epsilon)\\) of length \\(\\epsilon\\) are added in each factor. Our approaches then find points \\(Y_{i}\\in \\mathbb{R}^{3}\\) that match the noisy correspondences. We measure the accuracy by taking the logarithm after averaging the relative error of reconstructed points:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.4, + 0.469, + 0.434 + ], + "angle": 0, + "content": "\\[\ne = \\log_ {1 0} \\left(\\frac {\\sum \\| Y _ {i} - X _ {i} \\|}{p \\epsilon}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.469, + 0.491 + ], + "angle": 0, + "content": "where \\( \\| Y_{i} - X_{i}\\| \\) denotes the Euclidean distance. The interpretation of this number is that the error in the data gets amplified by \\( 10^{e} \\) during the triangulation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.21, + 0.529 + ], + "angle": 0, + "content": "2.1.3 Algorithms" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.469, + 0.825 + ], + "angle": 0, + "content": "The pseudocodes for all approaches are included in the Supplementary Material. The pseudocode for approach (L1).1 is presented in Algorithm 1. We use the notation that for a column vector \\( X \\in \\mathbb{R}^n \\), \\([X; 1] \\in \\mathbb{R}^{n+1}\\) is the vector we get by adding a 1 as the last coordinate. Let \\(\\iota\\) be the function that scales a vector such that its last coordinate is 1, and then removes that coordinate. In lines 1 and 2 of the algorithm, we generate noisy data by introducing randomly generated noise \\(\\sigma(\\epsilon)\\) as explained above, where \\(\\epsilon = 10^{-12}\\) is a fixed parameter for our experiments. Lines 3 and 4 correspond to the triangulation of the line and the point correspondences using the anchored point multiview variety at the line \\( L_0 \\). Note that we solve the closest point problem in line 4 by computing the zeros of a system of polynomial (critical) equations using the solve function in HomotopyContinuation.jl [9]. Finally, line 5 compares the points obtained in the previous step with the original starting points, measuring the logarithmic average relative error of the \\( p \\) points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Most other pseudocodes are similar to Algorithm 1. For instance, approach (L1).0 does not take into account the line, so it corresponds deleting lines 1 and 3 from Algorithm 1 and modifying the minimization domain in line 4, replacing \\( L_{0} \\) by \\( \\mathcal{M}_{\\mathcal{C}} \\). In approach (L1).2, line 3 is modified" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "so that \\( L_{0} \\) is obtained by the span of two triangulated points instead of by intersecting two back-projected planes." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.141, + 0.849, + 0.218 + ], + "angle": 0, + "content": "Algorithm 1: One iteration of the (L1).1 std. method given a randomly generated camera arrangement \\(\\mathcal{C}\\) of \\(3\\times 4\\) matrices, a projective line \\(L\\) spanned by two vectors of \\(\\mathbb{R}^4\\), and \\(p\\) points \\(X_{i}\\in \\mathbb{R}^{3}\\) such that \\([X_i;1]\\) lie on \\(L\\)." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.22, + 0.815, + 0.237 + ], + "angle": 0, + "content": "Input : \\( \\mathcal{C} = (C_1, \\dots, C_m) \\), \\( L, X_1, \\dots, X_p \\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.237, + 0.819, + 0.251 + ], + "angle": 0, + "content": "Output: The log of the average relative error" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.252, + 0.662, + 0.266 + ], + "angle": 0, + "content": "1 for \\(j\\) from 1 to \\(m\\) do" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.266, + 0.761, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\text {\\bf 2} \\\\ \\text {\\bf 3} \\\\ \\text {\\bf 4} \\end{array} \\left\\lfloor \\begin{array}{l} \\text {\\bf f o r} i \\text {f r o m 1 t o p d o} \\\\ \\text {\\bf \\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\c h i}}}}}}}}}}}}}}}}}}}}}} q _ {i, j} \\leftarrow \\iota (C _ {j} [ X _ {i}; 1 ]) + \\sigma (\\epsilon); \\\\ u _ {j} \\leftarrow \\iota (C _ {j} \\cdot L) + \\sigma (\\epsilon); \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.32, + 0.818, + 0.34 + ], + "angle": 0, + "content": "5 \\(L_0\\gets\\) nullspace \\(\\left[C_1^T [u_1;1]\\quad C_2^T [u_2;1]\\right]^T;\\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.341, + 0.653, + 0.354 + ], + "angle": 0, + "content": "6 for \\(i\\) from 1 to \\(p\\) do" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.354, + 0.844, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c}Y_{i}\\leftarrow \\\\ \\underset {X\\in \\mathbb{R}^{3}:[X;1]\\in L_{0}}{\\operatorname *{argmin}}\\sum_{j = 1}^{m}(q_{i,j} - \\iota (C_{j}[X;1]))^{2}; \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.399, + 0.75, + 0.423 + ], + "angle": 0, + "content": "8 \\(e\\gets \\log_{10}\\left(\\frac{1}{p\\epsilon}\\sum_{i = 1}^{p}\\| Y_i - X_i\\|\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.423, + 0.593, + 0.434 + ], + "angle": 0, + "content": "Return: \\( e \\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.672, + 0.482 + ], + "angle": 0, + "content": "2.2. Numerical results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.612 + ], + "angle": 0, + "content": "The main results of our numerical experiments are presented in Figures 3 to 5 and Tables 1 to 3. We compare the performance of the five different triangulation approaches for \\( p = 5 \\) and different number of \\( m \\) cameras. In the tables, we present the median, mean and standard deviation \\( \\sigma \\) of the logarithmic average relative error and time. The results are displayed in histograms and tables created with Plots.jl [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Note that for two cameras, (L1).0 and (L1).2 are the exact same, and (L1).1 and (L1).4 are essentially the same. In Figure 3 and Table 1 we present results for \\( m = 2 \\) and therefore only include (L1).0, (L1).1 and (L1).3. We also compare with the standard implementations of (L1).1 and (L1).3 that do not use the linear isomorphisms of Theorem 1.3 to improve computation time as described in Section 2.1.1. This simulation is iterated 1000 times." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.891, + 0.778 + ], + "angle": 0, + "content": "Figure 4 and Table 2 compare the different reconstruction approaches for \\( m = 3 \\) cameras. The simulations are again iterated 1000 times." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.891, + 0.839 + ], + "angle": 0, + "content": "Figure 5 and Table 3 compare the different reconstruction approaches for \\( m = 4 \\) cameras. The simulations are iterated only 100 times, due to the longer run-time of the simulations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our theoretical studies in Section 1, specifically Theorem 1.7, guarantee that methods such as (L1).1, (L1).2 and (L1).3 are less complex from the algebraic point of view, than (L1).0 for triangulation of points incident to a line. Ad-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3753" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.089, + 0.4, + 0.1 + ], + "angle": 0, + "content": "Accuracy: 2 cameras, 5 point correspondences" + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.101, + 0.412, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.236, + 0.394, + 0.247 + ], + "angle": 0, + "content": "Speed: 2 cameras, 5 point correspondences" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.248, + 0.412, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.395, + 0.471, + 0.445 + ], + "angle": 0, + "content": "Figure 3: Triangulation of \\( p = 5 \\) point correspondences incident to a line for \\( m = 2 \\) views with complete visibility. 1000 iterations. The histograms show the frequency of the average relative error and the running time of the depicted triangulation methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.468, + 0.543 + ], + "angle": 0, + "content": "ditionally, using Section 2.1.1, we were able to improve the computation speed while retaining comparable accuracy, which is exemplified by the two cases (L1).1 vs. (L1).1 std. and (L1).3 vs. (L1).3 std. in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.47, + 0.77 + ], + "angle": 0, + "content": "Finally, our numerical simulations show that in our HomotopyContinuation.jl implementation, the (L1).1 approach is the fastest across all number of views, but least accurate. In the case of \\( m = 3 \\), (L1).4 is the most accurate and faster than the traditional (L1).0 method. Even for \\( m = 4 \\), (L1).4 is the most accurate but is notably slower than (L1).0. Increasing the number of cameras \\( m \\) improves accuracy and reduces speed, but the exact impact depends on the approach. Among (L1).1-4, (L1).1 is least affected in both regards, and (L1).4 is most affected in both regards. The speed can be thought of as an affine function in the number of point correspondences \\( p \\). Reconstructing the line correspondence takes some fixed amount of time independent of \\( p \\) and then all \\( p \\) correspondences are independently reconstructed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.347, + 0.801 + ], + "angle": 0, + "content": "3. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "This work studied the triangulation of incident points and lines. We introduced the anchored line multiview varieties and showed that they correspond to multiview varieties arising from projections \\(\\mathbb{P}^2\\longrightarrow \\mathbb{P}^1\\) and \\(\\mathbb{P}^1\\longrightarrow \\mathbb{P}^1\\). In addition, we proved that they are less complex for reconstruction via critical points. These theoretical results are also" + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.089, + 0.842, + 0.186 + ], + "angle": 0, + "content": "
Accuracymedianmeanσ
(L1).01.0311.2431.213
(L1).11.4891.7021.085
(L1).1 std.1.3741.5480.990
(L1).31.2501.4490.964
(L1).3 std.1.1471.3931.205
" + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.187, + 0.842, + 0.285 + ], + "angle": 0, + "content": "
Speedmedianmeanσ
(L1).00.1700.2010.220
(L1).10.04300.04700.0207
(L1).1 std.0.1220.1410.202
(L1).30.08380.09410.0464
(L1).3 std.0.1670.1880.0764
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.307, + 0.894, + 0.356 + ], + "angle": 0, + "content": "Table 1: Triangulation of \\( p = 5 \\) point correspondences incident to a line for \\( m = 2 \\) views with complete visibility. 1000 iterations. The tables show the accuracy and speed of the depicted triangulation methods in terms of median, mean and standard deviation." + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.37, + 0.837, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.518, + 0.817, + 0.528 + ], + "angle": 0, + "content": "Speed: 3 cameras, 5 point correspondences" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.529, + 0.836, + 0.664 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.676, + 0.892, + 0.726 + ], + "angle": 0, + "content": "Figure 4: Triangulation of \\( p = 5 \\) point correspondences incident to a line for \\( m = 3 \\) views with complete visibility. 1000 iterations. The histograms show the frequency of the average relative error and the running time of the depicted triangulation methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.87 + ], + "angle": 0, + "content": "aligned with the numerical experiments we conducted. In particular, the proposed methods in Section 2 compare different methods for triangulating a set of points incident to a line. We highlight that according to our experiments, the use of the theoretical results in Section 1 allows for a notably faster triangulation while preserving the accuracy of the traditional method for point reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We hope this work motivates the use of new algebraic constraints in the triangulation of incident points and lines." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3754" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.132, + 0.089, + 0.416, + 0.284 + ], + "angle": 0, + "content": "
Accuracymedianmeanσ
(L1).00.6200.8171.023
(L1).11.5621.7811.128
(L1).20.8781.1250.998
(L1).31.0111.2431.078
(L1).40.4070.4990.901
Speedmedianmeanσ
(L1).00.9841.0600.514
(L1).10.08060.08580.0318
(L1).20.4310.4560.155
(L1).30.3040.3310.134
(L1).40.6160.7630.500
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.307, + 0.471, + 0.356 + ], + "angle": 0, + "content": "Table 2: Triangulation of \\( p = 5 \\) point correspondences incident to a line for \\( m = 3 \\) views with complete visibility. 1000 iterations. The tables show the accuracy and speed of the depicted triangulation methods in terms of median, mean and standard deviation." + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.37, + 0.415, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.518, + 0.393, + 0.529 + ], + "angle": 0, + "content": "Speed: 4 cameras, 5 point correspondences" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.53, + 0.411, + 0.664 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.676, + 0.47, + 0.726 + ], + "angle": 0, + "content": "Figure 5: Triangulation of \\( p = 5 \\) point correspondences incident to a line for \\( m = 4 \\) views with complete visibility. 100 iterations. The histograms show the frequency of the average relative error and the running time of the depicted triangulation methods." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.469, + 0.796 + ], + "angle": 0, + "content": "Some questions that arose in the development of this work, and that we believe are worth studying, are included below." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.808, + 0.469, + 0.853 + ], + "angle": 0, + "content": "- A triangulation approach inspired by [2] consists of reconstructing the 3D line \\(L\\) by fitting it to \\(\\mathcal{L}_{\\mathcal{C}}\\) solving the following optimization problem" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.865, + 0.469, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\text {m i n i m i z e} \\quad \\sum_ {i j} d \\left(x _ {i j}, \\ell_ {i}\\right) ^ {2} \\text {s u b j e c t t o} \\quad \\ell \\in \\mathcal {L} _ {\\mathcal {C}}. \\tag {14}\n\\]" + }, + { + "type": "table", + "bbox": [ + 0.559, + 0.089, + 0.835, + 0.284 + ], + "angle": 0, + "content": "
Accuracymedianmeanσ
(L1).00.3850.5510.776
(L1).11.4371.7621.249
(L1).21.0951.0770.845
(L1).30.8171.1101.094
(L1).40.2030.4211.324
Speedmedianmeanσ
(L1).04.3654.6751.576
(L1).10.1410.1490.0431
(L1).21.8091.8630.323
(L1).31.1071.1570.234
(L1).46.5997.3913.068
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.307, + 0.892, + 0.356 + ], + "angle": 0, + "content": "Table 3: Triangulation of \\( p = 5 \\) point correspondences incident to a line for \\( m = 4 \\) views with complete visibility. 100 iterations. The tables show the accuracy and speed of the depicted triangulation methods in terms of median, mean and standard deviation." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.38, + 0.892, + 0.44 + ], + "angle": 0, + "content": "Here, \\( x_{ij} \\) denotes the \\( j = 1,\\dots ,p \\) point correspondences across \\( i = 1,\\ldots ,m \\) cameras, and \\( d(x_{ij},\\ell_i) \\) is the affine distance from a line in \\( \\mathbb{R}^2 \\) to the point, meaning" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.448, + 0.891, + 0.484 + ], + "angle": 0, + "content": "\\[\nd \\left(x _ {i j}, \\ell_ {i}\\right) = \\frac {\\left| 1 + \\left(\\ell_ {i}\\right) _ {1} \\left(x _ {i j}\\right) _ {1} + \\left(\\ell_ {i}\\right) _ {2} \\left(x _ {i j}\\right) _ {2} \\right|}{\\sqrt {\\left(\\ell_ {i}\\right) _ {1} ^ {2} + \\left(\\ell_ {i}\\right) _ {2} ^ {2}}}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.49, + 0.891, + 0.534 + ], + "angle": 0, + "content": "We believe that this method is more accurate than (L1).4, but with longer running time, due to the appearance of fractions." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.544, + 0.891, + 0.62 + ], + "angle": 0, + "content": "- Consider the point-line problem (P1) defined by one point incident to \\(l\\) lines, illustrated in Figure 1. Can similar approaches to (L1).1-4 be formulated for the triangulation of the (P1) setting? Can they be used to improve accuracy and/or the speed of triangulation?" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.628, + 0.892, + 0.733 + ], + "angle": 0, + "content": "- Our implementation for the numerical experiments does not directly correspond to the specialized software and hardware used for triangulation in practice. However, our results motivate that patterns displayed in our figures and tables could translate to such settings. This could improve the efficiency of current specialized solvers." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.544, + 0.892, + 0.733 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.751, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements. The authors thank Kathlen Kohn for helpful discussions, Lukas Gustafsson for the initial insight of Theorem 1.3, and Viktor Larsson for pointing out relevant references at the start of this project. Felix Rydell and Angelica Torres were supported by the Knut and Alice Wallenberg Foundation within their WASP (Wallenberg AI, Autonomous Systems and Software Program) AI/Math initiative. Elima Shehu is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), Projektnummer 445466444. Angelica Torres is currently supported by the Spanish State Research Agency, through" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3755" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.093, + 0.47, + 0.121 + ], + "angle": 0, + "content": "the Severo Ochoa and María de Maeztu Program for Centers and Units of Excellence in R&D." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.139, + 0.173, + 0.154 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.165, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[1] Sameer Agarwal, Andrew Pryhuber, and Rekha R Thomas. Ideals of the multiview variety. IEEE transactions on pattern analysis and machine intelligence, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.209, + 0.469, + 0.263 + ], + "angle": 0, + "content": "[2] Adrien Bartoli and Peter Sturm. Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer Vision and Image Understanding, 100(3):416-441, 2005. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.266, + 0.469, + 0.32 + ], + "angle": 0, + "content": "[3] Paul A Beardsley, Andrew Zisserman, and David William Murray. Navigation using affine structure from motion. Lecture Notes in Computer Science, 801:85-96, 1994. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.324, + 0.469, + 0.378 + ], + "angle": 0, + "content": "[4] Paul A Beardsley, Andrew Zisserman, and David William Murray. Sequential updating of projective and affine structure from motion. International journal of computer vision, 23:235-259, 1997. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[5] Jeff Bezanson, Stefan Karpinski, Viral B Shah, and Alan Edelman. Julia: A fast dynamic language for technical computing. arXiv preprint arXiv:1209.5145, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.469, + 0.466 + ], + "angle": 0, + "content": "[6] Paul Breiding, Timothy Duff, Lukas Gustafsson, Felix Rydell, and Elima Shehu. Line multiview ideals. arXiv preprint arXiv:2303.02066, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.469, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[7] Paul Breiding, Felix Rydell, Elima Shehu, and Angélica Torres. Line multiview varieties. SIAM Journal on Applied Algebra and Geometry, 7(2):470-504, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.513, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[8] Paul Breiding and Sascha Timme. Homotopycontinuation. jl: A package for homotopy continuation in julia. In International Congress on Mathematical Software, pages 458-465.Springer,2018.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.571, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[9] Paul Breiding and Sascha Timme. HomotopyContinuation.jl: A Package for Homotopy Continuation in Julia. In Mathematical Software - ICMS 2018, pages 458-465, Cham, 2018. Springer International Publishing. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.449, + 0.643 + ], + "angle": 0, + "content": "[10] Tom Breloff and other contributors. JuliaPlots/Plots.jl. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[11] Jan Draisma, Emil Horobet, Giorgio Ottaviani, Bernd Sturmfels, and Rekha R Thomas. The euclidean distance degree of an algebraic variety. Foundations of computational mathematics, 16(1):99-149, 2016. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[12] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajdla. Plmp-point-line minimal problems in complete multi-view visibility. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1675-1684, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[13] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajda. PI\\$\\_1\\$\\$p - point-line minimal problems under partial visibility in three views. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 175–192, Cham, 2020. Springer International Publishing. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] A. Elqursh and A. Elgammal. Line-based relative pose estimation. In Conference on Computer Vision and Pattern Recognition, 2011. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[15] Ricardo Fabbri, Timothy Duff, Hongyi Fan, Margaret H Regan, David da Costa de Pinho, Elias Tsigaridas, Charles W Wampler, Jonathan D Hauenstein, Peter J Giblin, Benjamin Kimia, et al. Trplp-trifocal relative pose from lines at points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12073-12083, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[16] Olivier Faugeras. Three-dimensional computer vision: A geometric viewpoint. MIT press, Cambridge, MA, 1993. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[17] Olivier Faugeras and Bernard Mourrain. On the geometry and algebra of the point and line correspondences between n images. In Proceedings of IEEE International Conference on Computer Vision, pages 951-956. IEEE, 1995. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[18] Andreas Gathmann. Algebraic geometry, 2019/20. Class Notes TU Kaiserslautern. Available at https://www.mathematik.uni-kl.de/~gathmann/de/alggeom.php.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[19] Corey Harris and Daniel Lowengrub. The Chern-mather class of the multiview variety. Communications in Algebra, 46(6):2488-2499, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[20] Richard I. Hartley. Lines and points in three views and the trifocal tensor. International Journal of Computer Vision, 22(2):125-140, 1997. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[21] Richard I Hartley and Peter Sturm. Triangulation. Computer vision and image understanding, 68(2):146-157, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[22] Richard I. Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[23] Manuel Hofer, Andreas Wendel, and Horst Bischof. Incremental line-based 3d reconstruction using geometric constraints. The British Machine Vision Association and Society for Pattern Recognition, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[24] Kun Huang, Yifan Wang, Zihan Zhou, Tianjiao Ding, Shenghua Gao, and Yi Ma. Learning to parse wireframes in images of man-made environments. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 626-635, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[25] Javier Civera Joan Solà, Teresa Vidal-Calleja and José María Martínez Montiel. Impact of landmark parametrization on monocular ekf-slam with points and lines. International Journal of Computer Vision, 97(3):339-368, 2012. 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.678, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[26] Kenichi Kanatani. Statistical optimization for geometric computation: theory and practice. Courier Corporation, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[27] Kenichi Kanatani, Yasuyuki Sugaya, and Hirotaka Niitsuma. Triangulation from two views revisited: Hartley-sturm vs. optimal correction. practice, 4(5):99, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[28] Yasushi Kanazawa and Kenichi Kanatani. Reliability of 3- d reconstruction by stereo vision. IEICE TRANSACTIONS on Information and Systems, 78(10):1301-1306, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[29] Joseph David Kileel. Algebraic Geometry for Computer Vision. ProQuest LLC, Ann Arbor, MI, 2017. Thesis (Ph.D.)-University of California, Berkeley. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21445-21455, June 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3756" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] Laurentiu G. Maxim, Jose I. Rodriguez, and Botong Wang. Euclidean distance degree of the multiview variety. SIAM Journal on Applied Algebra and Geometry, 4(1):28-48, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[32] Branislav Micusik and Horst Wildenauer. Structure from motion with line segments under relaxed endpoint constraints. In 2014 2nd International Conference on 3D Vision, volume 1, pages 13-19, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[33] Long Quan and T. Kanade. Affine structure from line correspondences with uncalibrated affine cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(8):834-845, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.471, + 0.331 + ], + "angle": 0, + "content": "[34] Grant Schindler, Panchapagesan Krishnamurthy, and Frank Dellaert. Line-based structure from motion for urban environments. Third International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT'06), pages 846-853, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.333, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[35] Henrik Stewénius, Frederik Schaffalitzky, and David Nister. How hard is 3-view triangulation really? In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 1, pages 686-693. IEEE, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.39, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[36] Matthew Trager, Martial Hebert, and Jean Ponce. The joint image handbook. In Proceedings of the IEEE international conference on computer vision, pages 909-917, 2015. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3757" + } + ] +] \ No newline at end of file diff --git a/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/5fbfedc8-4ebc-4e53-8cf0-a4029a8a49ba_origin.pdf b/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/5fbfedc8-4ebc-4e53-8cf0-a4029a8a49ba_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fe69364cbcb40635f20e1a6dd01bd3ca62deee34 --- /dev/null +++ b/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/5fbfedc8-4ebc-4e53-8cf0-a4029a8a49ba_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2c4b084a48103cf497b9a2fcedc9a734bfae2b3de9e0ce9f6560599a8fc303c +size 928072 diff --git a/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/full.md b/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/full.md new file mode 100644 index 0000000000000000000000000000000000000000..301f7f70214dc2c5bc9d62efcaec231355625c81 --- /dev/null +++ b/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/full.md @@ -0,0 +1,429 @@ +# Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences + +Felix Rydell + +KTH Royal Institute of Technology + +Lindstedtsvagen 25, Stockholm, Sweden + +felixry@kth.se + +Elima Shehu + +University of Osnabrück and MPI MiS + +Albrechtstrasse 28a, Osnabrück, Germany + +Inselstrasse 22, Leipzig, Germany + +elima.shehu@mis.mpg.de + +Angélica Torres + +Centre de Recerca Matemàtica + +Edifici C, Campus UAB, Bellaterra, Spain + +atorres@crm.cat + +# Abstract + +We study the joint image of lines incident to points, meaning the set of image tuples obtained from fixed cameras observing a varying 3D point-line incidence. We prove a formula for the number of complex critical points of the triangulation problem that aims to compute a 3D point-line incidence from noisy images. Our formula works for an arbitrary number of images and measures the intrinsic difficulty of this triangulation. Additionally, we conduct numerical experiments using homotopy continuation methods, comparing different approaches of triangulation of such incidences. In our setup, exploiting the incidence relations gives a notably faster point reconstruction with comparable accuracy. + +# Introduction + +The Structure-from-Motion pipeline aims to estimate the camera positions and the relative position of the objects present in a given set of images. The process starts by identifying point and line features in one image that are recognizable as the same points or lines in another. These matched features, that we call correspondences, are used to estimate the camera position and orientation, which then allow for the triangulation step where the position of the points and lines is estimated. + +In this work, we study the triangulation of points and lines satisfying a certain incidence relation. Specifically, we focus on the triangulation of points contained in a single line; see Figure 1. We note that our methods can also be used when multiple lines going through a fixed point. We + +![](images/f9dcca4617d849064605759b8dc46da8cce53217a254be158992f305c8ec7154.jpg) +Figure 1: The illustration of two different types of incidence relations: (L1) represents the scenario where multiple points are incident to a line, while (P1) represents the scenario where multiple lines are incident to a point. + +develop the theory for the triangulation of these problems for an arbitrary number of pinhole cameras assuming complete visibility. + +The triangulation of points is well understood and, in practice, it is efficiently implemented. However, to the best of our knowledge, incidence relations are not considered in these implementations. The inclusion of line features and incidence relations in the triangulation process could give a more accurate estimation of scenes as lines are more robust to noise than points, and incidence relations appear frequently in interior scenes and man-made scenarios; see [34], [24], [30]. Additionally, in some real data sets standard feature detection algorithms fail due to a lack of point correspondences but succeed when line correspondences are taken into account [15]. + +The main tools in our work are (algebraic) varieties, that is, the vanishing sets of systems of polynomial equations. Algebraic varieties have been used extensively to study triangulation of point correspondences [1, 17, 31], line correspondences [7, 16, 29] and minimal problems [12, 13] to + +name a few. In these works, algebraic varieties arise naturally due to the algebraic nature of pinhole cameras. + +A pinhole camera, is modeled by a (complex) projective linear map $C: \mathbb{P}^3 \dashrightarrow \mathbb{P}^2$ defined by a $3 \times 4$ matrix $C$ of full rank that takes a point $X \in \mathbb{P}^3$ and sends it to $CX \in \mathbb{P}^2$ . A camera arrangement with $m \geq 2$ cameras is denoted by $\mathcal{C} = (C_1, \ldots, C_m)$ , and the joint camera map + +$$ +\begin{array}{l} \Phi_ {\mathcal {C}}: \mathbb {P} ^ {3} \dashrightarrow (\mathbb {P} ^ {2}) ^ {m}, \tag {1} \\ X \mapsto (C _ {1} X, \dots , C _ {m} X), \\ \end{array} +$$ + +models the process of taking the images of a point $X$ in homogeneous coordinates with $m$ cameras. For fixed cameras, the (point) multiview variety $\mathcal{M}_{\mathcal{C}}$ is the smallest variety that contains all point correspondences. An extensive account of the pinhole cameras is given by [22], and a survey of the multiview variety is found in [36]. The joint camera map can be extended from the space of points $\mathbb{P}^3$ to the space of lines, denoted by $\mathrm{Gr}(1,\mathbb{P}^3)$ , where each line is parametrized as the span of two points. The map + +$$ +\begin{array}{l} \Upsilon_ {\mathcal {C}}: \operatorname {G r} (1, \mathbb {P} ^ {3}) \dashrightarrow (\mathbb {P} ^ {2}) ^ {m}, \tag {2} \\ L \mapsto (C _ {1} \cdot L, \dots , C _ {m} \cdot L). \\ \end{array} +$$ + +models the image of a line $L$ taken by the cameras of $\mathcal{C}$ . To be precise, if $u, v$ span the line $L$ in $\mathbb{P}^3$ , then $C \cdot L$ is defined as $\ell = Cu \times Cv \in \mathbb{P}^2$ , where $\times$ is the cross product. Observe that $\{x \in \mathbb{P}^2 : \ell^T x = 0\}$ equals $\operatorname{span}\{Cu, Cv\}$ . Recently, in [7], the authors study the line multiview variety referred to as $\mathcal{L}_{\mathcal{C}}$ , which is the smallest variety containing all line correspondences. + +Our main contribution is the definition and study of the anchored point and line multiview varieties. Given a fixed line $L$ in $\mathbb{P}^3$ , the anchored point multiview variety, $\mathcal{M}_C^L$ , is defined as the smallest variety containing all point correspondences coming from points in $L$ ; and the anchored line multiview variety, $\mathcal{L}_C^X$ , is the smallest variety containing the line correspondences coming from lines passing through a given point $X \in \mathbb{P}^3$ . + +For these new varieties we prove formulas for their Euclidean distance degree (EDD), which are a measurement of the complexity for error correction using exact algebraic methods [11]. Specifically, we prove the following theorem: + +Theorem 0.1. Let $\mathcal{C}$ be a generic arrangement of $m$ cameras. + +1. $\operatorname{EDD}(\mathcal{M}_{\mathcal{C}}^{L}) = 3m - 2$ +2. If $m \geq 3$ , then $\mathrm{EDD}(\mathcal{L}_{\mathcal{C}}^{X}) = \frac{9}{2} m^{2} - \frac{19}{2} m + 3$ . + +We provide a precise definition of this degree in Section 1.2. Previous work on EDD for multiview varieties includes [19, 31] and [7, Section 5]. The EDD of the point multiview variety is $\mathrm{EDD}(\mathcal{M}_{\mathcal{C}}) = \frac{9}{2} m^3 -$ + +$\frac{21}{2} m^2 + 8m - 4$ , according to [31]. The fact that the EDDs of the anchored multiview varieties are polynomials of smaller degrees suggests that they are less complex for the purpose of data correction. This conclusion is backed by the results of our numerical experiments using HomotopyContinuation.jl [9], presented in Section 2. + +Another contribution of our work is the numerical simulations for the triangulation of points contained in a line using the anchored multiview varieties. We present different approaches to triangulating data of type (L1) that differ from the traditional method of fitting point correspondences to the multiview variety $\mathcal{M}_{\mathcal{C}}$ . For $m = 2$ views, our implementation is notably faster than the traditional triangulation of points described above, while the accuracy is comparable. For $m = 3$ views we get both higher accuracy and faster speed by using $\mathcal{L}_{\mathcal{C}}$ , compared to usual point triangulation. All of our proposed methods outperform the traditional approaches in terms of run-time and give a comparably accurate result. In practice, special software and hardware are used for triangulation, and based on our experiments, we believe that these approaches could be implemented efficiently with good results. + +# Related work + +Algebra. Algebraic geometry, whose connection to computer vision is well established, is our main tool for the theoretical study of the triangulation of problem (L1). Fundamental theory regarding the point multiview variety is found in [17, 20, 22], in particular for two and three views. The paper [36] by Trager et. al. serves as a survey for the algebraic properties of this multiview variety. For applications, finding a good set of polynomial constraints satisfied by the point correspondences is important, this is precisely the work of [1] by Agarwal et al. + +Regarding the algebra of lines in a computer vision setting, Kileel presented in [29, Definition 3.9, Theorem 3.10] several types of multiview varieties with respect to a point-line incidence relation in $\mathbb{P}^3$ in 3 views, and their basic properties. One of those types is called the LLL-multiview variety and constitutes a special case of the more recent work by Breiding et. al. [7], where they study algebraic properties of the line multiview variety. Furthermore, a recent manuscript [6] studies the polynomial constraints satisfied by line correspondences. + +Projective Reconstruction. Different approaches to the triangulation of points have been considered in the literature, in particular for two views [3, 4, 21, 26-28]. Hartley and Sturm compare many different approaches in [21], including the "midpoint" method. The midpoint method inputs a point correspondence in two views and outputs the midpoint of the line segment determined by the points on + +the two back-projected lines that are closest to each other. In other words, it finds the midpoint of the common perpendicular to the two back-projected lines. This is a natural approach, but it has several downsides, also pointed out in [3, 4]. Importantly, it is not projectively invariant. A projectively invariant reconstruction has the property that acting on the cameras by a global projective transformation induces the same action on the triangulated point or line. Instead, Hartley and Sturm propose minimizing the reprojection error as the optimal triangulation method, which is accepted as a standard formulation [35] and it is projectively invariant. In [35], the focus is on triangulation via minimizing reprojections errors in three views and the authors point out that three views often lead to greater stability and stronger disambiguation compared to two views. + +Implementation of Line Reconstruction. From the algorithmic point of view, the simultaneous reconstruction of point and line features have been studied specially with the goal of reconstructing line segments. For example, [2] provides a thorough overview of the structure-from-motion pipeline using lines, going through different methods for line parameterization and error correction depending on such parameterizations. In [33], Quan and Takeo study the algebraic structure of line correspondences with uncalibrated affine cameras. They reduce the problem of reconstructing affine lines to the reconstruction of projective points in a lower-dimensional projective camera. In [20], Hartley and coauthors provide an algorithm for the reconstruction of point and line features where 3D lines are parameterized by their projections in 2 views (as the intersection of two back-projected planes). Micusik and Wildenauer reconstruct lines to estimate line segments, but only use incident points for error correction [32]. Furthermore in [23], the authors propose a technique for 3D reconstruction incorporating line segments. They assert that this method surpasses traditional approaches, especially in efficiency while getting accurate results. Finally, in [14], the authors present a framework for the computation of the relative motion between two images using a triplet of lines. + +This paper is structured as follows. In the first part of Section 1 we formally introduce anchored multiview varieties and study some properties. In the second part of Section 1, we study their smoothness and find their Euclidean Distance Degree. Finally, in Section 2, we provide a numerical analysis of different approaches to reconstructing point correspondences incident to a line correspondence. The proofs of all our results are included in the Supplementary Material together with a small background on the concepts of Algebraic Geometry used for these results and the code for the numerical experiments. + +# 1. Anchored Multiview Varieties + +A camera refers to a full-rank $3 \times 4$ matrix. A camera arrangement is a collection $\mathcal{C} = (C_1, \ldots, C_m)$ of $m \geq 2$ cameras whose centers, i.e. kernels, are all distinct. + +A variety is the solution set to a system of polynomial equations. The Zariski closure of a set $U$ is the smallest variety containing $U$ . We work in complex projective space, denoted $\mathbb{P}^n$ . This is defined as $(\mathbb{C}^n \setminus \{0\}) / \sim$ , where $x \sim y$ if $x$ and $y$ differ by a non-zero constant. The set of lines in $\mathbb{P}^n$ is denoted as $\operatorname{Gr}(1, \mathbb{P}^n)$ . In the language of algebraic geometry, it is called the Grassmanian of lines in $\mathbb{P}^n$ . In general we denote with upper case letters world objects and with lower case letters image objects. For a rational map $\varphi: X \dashrightarrow Y$ , we denote $\varphi|_A$ the restriction of $\varphi$ to $A \subseteq X$ . + +Definition 1.1. Let $X \in \mathbb{P}^3$ be a point distinct from all camera centers, and $L$ a line in $\mathbb{P}^3$ containing none of the camera centers. + +1. The anchored point multiview variety, denoted $\mathcal{M}_C^L$ , is defined as the Zariski closure of the image of the map + +$$ +\begin{array}{c} \Phi_ {\mathcal {C}} | _ {L}: L \dashrightarrow (\mathbb {P} ^ {2}) ^ {m}, \\ X \longmapsto (C _ {1} X, \dots , C _ {m} X). \end{array} \tag {3} +$$ + +2. The anchored line multiview variety, denoted $\mathcal{L}_C^X$ , is defined as the Zariski closure of the image of the map + +$$ +\begin{array}{c} \Upsilon_ {\mathcal {C}} | _ {\Lambda (X)}: \Lambda (X) \dashrightarrow (\mathbb {P} ^ {2}) ^ {m}, \\ L \longmapsto \left(C _ {1} \cdot L, \dots , C _ {m} \cdot L\right), \end{array} \tag {4} +$$ + +where $\Lambda(X)$ denotes the set of lines in $\mathbb{P}^3$ that contain $X$ and $\ell_i = C_i \cdot L$ is defined as in the introduction. + +The anchored point multiview variety $\mathcal{M}_{\mathcal{C}}^{L}$ is the smallest variety that contains all point correspondences $\Phi_{\mathcal{C}}(X)$ for $X \in L$ . Similarly, the anchored line multiview variety $\mathcal{L}_{\mathcal{C}}^{X}$ is the smallest variety that contains all line correspondences $\Upsilon_{\mathcal{C}}(L)$ for $L$ meeting $X$ and no center. We highlight that our definition of anchored multiview varieties is different from the anchored features in [25] for monocular EKF-SLAM. + +Proposition 1.2. Consider an arrangement of $m$ cameras $\mathcal{C} = (C_1, \ldots, C_m)$ , a point $X \in \mathbb{P}^3$ and a line $L$ in $\mathbb{P}^3$ satisfying the conditions of Definition 1.1. + +1. If there are two different camera centers $c_{i}$ and $c_{j}$ such that the span of $\{c_{i}, c_{j}, L\}$ is $\mathbb{P}^3$ , then + +$$ +\mathcal {M} _ {\mathcal {C}} ^ {L} = \left\{\left(x _ {1}, \dots , x _ {m}\right) \in \mathcal {M} _ {\mathcal {C}}: x _ {i} \in C _ {i} \cdot L \right\}. \tag {5} +$$ + +2. If for each camera center $c_i$ , the line spanned by $c_i$ and $X$ does not contain any other camera center, then + +$$ +\mathcal {L} _ {\mathcal {C}} ^ {X} = \left\{\left(\ell_ {1}, \dots , \ell_ {m}\right) \in \mathcal {L} _ {\mathcal {C}}: C _ {i} X \in \ell_ {i} \right\}. \tag {6} +$$ + +The proofs for this and all our results are found in the Supplementary Material. + +# 1.1. Properties of Anchored Multiview Varieties + +A fundamental property of the anchored multiview varieties is that they are linearly isomorphic to multiview varieties arising from projections $\mathbb{P}^1\longrightarrow \mathbb{P}^1$ and $\mathbb{P}^2\longrightarrow \mathbb{P}^1$ respectively. A linear isomorphism is a linear map (given by a matrix) with an inverse that is also linear. + +Consider an arrangement $\widetilde{\mathcal{C}}$ of $m$ full-rank $2\times 2$ matrices, and an arrangement $\widehat{\mathcal{C}}$ of $m$ full-rank $2\times 3$ matrices. Also here we assume $m\geq 2$ . We define $\mathcal{M}_{\widetilde{\mathcal{C}}}^{1,1}$ , and $\mathcal{M}_{\widehat{\mathcal{C}}}^{2,1}$ respectively as the Zariski closure of the image of the maps + +$$ +\Phi_ {\tilde {C}}: \mathbb {P} ^ {1} \longrightarrow (\mathbb {P} ^ {1}) ^ {m}, \tag {7} +$$ + +$$ +X \longmapsto (\widetilde {C} _ {1} X, \dots , \widetilde {C} _ {m} X) +$$ + +and + +$$ +\Phi_ {\widehat {C}}: \mathbb {P} ^ {2} \dashrightarrow (\mathbb {P} ^ {1}) ^ {m}, \tag {8} +$$ + +$$ +X \longmapsto (\widehat {C} _ {1} X, \dots \widehat {C} _ {m} X). +$$ + +# Theorem 1.3. + +1. Let $\phi_L: L \to \mathbb{P}^1$ and $\psi_{\mathcal{C},i}: C_i \cdot L \to \mathbb{P}^1$ be any choices of linear isomorphisms. Let $\tilde{\mathcal{C}}$ denote the arrangement of matrices $C_i := \psi_{\mathcal{C},i} \circ C_i \circ \phi_L^{-1}$ . Then + +$$ +\psi_ {\mathcal {C}, L} := \left(\psi_ {\mathcal {C}, 1}, \dots , \psi_ {\mathcal {C}, m}\right): \mathcal {M} _ {\mathcal {C}} ^ {L} \rightarrow \mathcal {M} _ {\tilde {\mathcal {C}}} ^ {1, 1} \tag {9} +$$ + +is a linear isomorphism. + +2. Let $\phi_X: \Lambda(X) \to \mathbb{P}^2$ and $\psi_{\mathcal{C},i}: \Lambda(C_iX) \to \mathbb{P}^1$ be any choices of linear isomorphisms. Let $\widehat{\mathcal{C}}$ denote the arrangement of matrices $\widehat{C}_i \coloneqq \psi_{\mathcal{C},i} \circ C_i \circ \phi_X^{-1}$ . Then + +$$ +\psi_ {\mathcal {C}, X} := \left(\psi_ {\mathcal {C}, 1}, \dots , \psi_ {\mathcal {C}, m}\right): \mathcal {L} _ {\mathcal {C}} ^ {X} \rightarrow \mathcal {M} _ {\widehat {\mathcal {C}}} ^ {2, 1} \tag {10} +$$ + +is a linear isomorphism. + +In Section 2.1.1, we exploit these linear isomorphisms to improve the speed of triangulation for our setting. A consequence of this theorem is also that many results for the anchored multiview varieties translate into results about $\mathcal{M}_{\hat{\mathcal{C}}}^{1,1}$ and $\mathcal{M}_{\hat{\mathcal{C}}}^{2,1}$ . + +Proposition 1.4. $\mathcal{M}_C^L$ and $\mathcal{L}_C^X$ are irreducible. Further, + +1. $\mathcal{M}_C^L$ is isomorphic to $\mathbb{P}^1$ . In particular, $\dim \mathcal{M}_C^L = 1$ . +2. If the span of the centers $c_{i}$ and the point $X$ are not collinear, then $\dim \mathcal{L}_{\mathcal{C}}^{X} = 2$ . + +Using the fundamental matrices of $C_i$ and $C_j$ , denoted as $F^{ij}$ , we introduce a set of polynomial constraints that must be fulfilled by the correspondences of points or lines in the anchored multiview varieties. These constraints imply that some of the equations obtained from Proposition 1.2 are redundant in the generic case. + +Proposition 1.5. For a point $X \in \mathbb{P}^3$ and line $L$ in $\mathbb{P}^3$ , let $\mathcal{C}$ be a generic (random) camera arrangement of $m$ cameras. + +1. $x \in \mathcal{M}_{\mathcal{C}}^{L}$ if and only if $x_{1}^{T}F^{1j}x_{j} = 0$ for every $j = 2, \ldots, m$ and $x_{i}^{T}C_{i} \cdot L = 0$ for every $i = 1, \ldots, m$ . +2. $\ell \in \mathcal{L}_C^X$ if and only if + +$$ +\det \left[ \begin{array}{l l l} C _ {1} ^ {T} \ell_ {1} & C _ {2} ^ {T} \ell_ {2} & C _ {i} ^ {T} \ell_ {i} \end{array} \right] = 0, \tag {11} +$$ + +$$ +\det \left[ \begin{array}{c c c} C _ {1} ^ {T} \ell_ {1} & C _ {3} ^ {T} \ell_ {3} & C _ {i} ^ {T} \ell_ {i} \end{array} \right] = 0 +$$ + +for $i = 3,\ldots ,m$ and $\ell_i^T C_iX = 0$ for every $i = 1,\dots ,m$ . + +The determinantal constraints described in the second item correspond to the constraints that are satisfied by elements of the line multiview variety, presented in [7]. + +In Supplementary Material Section B we define the multidegree of a variety in a product of projective spaces, determine it for the anchored multiview varieties and explain its relevance for computer vision. + +# 1.2. The Euclidean Distance problem + +For a variety $\mathcal{X} \subseteq \mathbb{R}^n$ and a point $u \in \mathbb{R}^n$ outside the variety, a natural problem is to find the closest point on $\mathcal{X}$ to $u$ , which corresponds to the optimization problem + +$$ +\text {m i n i m i z e} \quad \sum_ {i = 1} ^ {n} \left(u _ {i} - x _ {i}\right) ^ {2} \quad \text {s u b j e c t t o} \quad x \in \mathcal {X}. \tag {12} +$$ + +Equation (12) is called the Euclidean distance problem and models the process of error correction and fitting noisy data to a mathematical model $\mathcal{X}$ . In the case of a smooth variety, defined below, the Euclidean distance degree (EDD) is the number of complex solutions to the critical equations of (12) [11]. The EDD is an estimate of how difficult it is to solve this problem by exact algebraic methods. For a variety $\mathcal{X}$ in a product of projective spaces $\mathbb{P}^{n_1} \times \dots \times \mathbb{P}^{n_m}$ , the EDD is the EDD of $\mathcal{X} \cap U_1 \times \dots \times U_m \subseteq \mathbb{R}^{n_1 + \dots + n_m}$ , where $U_i$ is a generic affine patch of $\mathbb{P}^{h_i}$ for each $i$ . An affine patch $U$ of $\mathbb{P}^n$ is a subset defined by an affine equation with non-zero constant part, for instance $x_0 = 1$ . We have $U \cong \mathbb{R}^n$ over the real numbers. + +A variety $\mathcal{X}$ is smooth at a point $x$ if $\mathcal{X}$ locally around $x$ looks like Euclidean space. $\mathcal{X}$ is smooth if all its points are smooth. The singular locus of a variety is the set of non-smooth points. See [18] for more details. The singular locus of the point multiview variety is well-understood, see for instance [36], and it is mostly understood for the line multiview variety, see [7, Section 3]. The proposition below guarantees that the anchored multiview varieties are smooth for generic camera arrangements, which helps us to compute its EDD. + +# Proposition 1.6. + +1. $\mathcal{M}_C^L$ smooth. +2. If there are exactly two cameras, or the centers together with the point $X$ span $\mathbb{P}^3$ , then $\mathcal{L}_C^X$ is smooth. + +For the EDD to be relevant in a particular setting, the Euclidean distance has to be a good measurement of distance. This is naturally the case for points in $\mathbb{R}^2$ , i.e. an affine patch of $\mathbb{P}^2$ . This extends to an affine patch of $\mathcal{M}_{\mathcal{C}}$ . It is perhaps less clear how to best measure distances between lines. Recall that we regard $\mathcal{L}_{\mathcal{C}}$ as a subset of $(\mathbb{P}^2)^m$ . Indeed, we identify image lines with the point that defines its normal vector. We choose the Euclidean distance between these normal vectors (in affine patches of $\mathbb{P}^2$ ) as our distance between lines. + +The theorem below presents our computations of the EDDs of the anchored multiview varieties. Additionally, our numerical computations using the HomotopyContinuation.jl [8] package in the julia [5] programming language have confirmed the accuracy of these formulas for $m \leq 10$ . + +Theorem 1.7. Let $\mathcal{C}$ be a generic arrangement of $m$ cameras. + +1. $\operatorname{EDD}(\mathcal{M}_C^L) = 3m - 2$ . +2. If $m \geq 3$ , then $\mathrm{EDD}(\mathcal{L}_{\mathcal{C}}^{X}) = \frac{9}{2} m^{2} - \frac{19}{2} m + 3$ . + +We end with an implication of Theorem 1.3: There is a direct correspondence between the EDDs of the anchored multiview varieties and $\mathcal{M}_{\widetilde{\mathcal{C}}}^{1,1}, \mathcal{M}_{\widehat{\mathcal{C}}}^{2,1}$ . + +Theorem 1.8. Let $\hat{C}$ and $\hat{C}$ be generic arrangements of cardinality $m$ . + +1. $\operatorname{EDD}(\mathcal{M}_{\widetilde{c}}^{1,1}) = 3m - 2.$ +2. If $m \geq 3$ , then $\mathrm{EDD}(\mathcal{M}_{\hat{\mathcal{C}}}^{2,1}) = \frac{9}{2} m^2 - \frac{19}{2} m + 3$ . + +# 2. Numerical Experiments + +In this section we conduct numerical experiments of the triangulation process of incident point and lines. All the code can be found in the Supplementary Material. + +By (L1) we refer to point-line arrangements in $\mathbb{P}^3$ of one line incident to $p$ points, as depicted in Figure 1. Our goal is to reconstruct such arrangements, given point correspondences incident to a single line correspondence across $m$ views. The following is a list of natural approaches for such triangulation given a camera arrangement $\mathcal{C}$ . It is not necessarily a complete list. + +(L1).0 Triangulate each point correspondence by fitting it to the point multiview variety $\mathcal{M}_{\mathcal{C}}$ , i.e. find the closest point correspondence in $\mathcal{M}_{\mathcal{C}}$ ; + +![](images/348ea6aca160ede2bc865b7669ebe94dc2efe1b17bfe45052a37721cf20985a9.jpg) +Figure 2: Comparison of (L1).0 and (L1).1 triangulation approaches. (L1).0 expects non-collinear points in the reconstruction, while (L1).1 produces collinear points. + +![](images/cda082f93cda9cec7bd913c67990e93c00e05227df19c28c051063128f0822b8.jpg) + +(L1).1 Reconstruct the 3D line $L$ by back-projecting the image lines from two views. Triangulate the point correspondences by fitting them to the anchored multiview variety $\mathcal{M}_{\mathcal{C}}^{L}$ ; +(L1).2 Triangulate two point correspondences by fitting them to $\mathcal{M}_C$ to get 3D points $X$ and $Y$ . Let $L$ be the line they span in $\mathbb{R}^3$ . Triangulate the remaining point correspondences by fitting them to $\mathcal{M}_C^L$ ; +(L1).3 Triangulate one point correspondence by fitting it to $\mathcal{M}_{\mathcal{C}}$ to get a 3D point $X$ . Reconstruct the 3D line $L$ by fitting the line correspondence to the anchored variety $\mathcal{L}_{\mathcal{C}}^{X}$ . Triangulate the remaining point correspondences by fitting them to $\mathcal{M}_{\mathcal{C}}^{L}$ ; +(L1).4 Reconstruct the 3D line $L$ by fitting the line correspondence to the line multiview variety $\mathcal{L}_{\mathcal{C}}$ . Triangulate the point correspondences by fitting them to the anchored multiview variety $\mathcal{M}_{\mathcal{C}}^{L}$ . + +Approaches (L1).1-4 take the incidence relations into account in the triangulation, whereas (L1).0 does not. Therefore, in contrast to (L1).1-4, the resulting triangulation of (L1).0 does not preserve the point-line incidences; see Figure 2. + +# 2.1. Implementation + +Our experiments are implemented in HomotopyContinuation.jl [8] in Julia [5]. Here we explain them in some detail. The full explanation can be found in Supplementary Material. We ran the code on a Intel(R) Core(TM) i5-8300H CPU running at 2.30GHz. + +# 2.1.1 Reducing the number of parameters + +We reduce the number of parameters involved in the problem of fitting data to $\mathcal{M}_{\mathcal{C}}^{L}$ , respectively $\mathcal{L}_{\mathcal{C}}^{X}$ , by translating it to and solving it for $\mathcal{M}_{\widetilde{\mathcal{C}}}^{1,1}$ , respectively $\mathcal{M}_{\widehat{\mathcal{C}}}^{2,1}$ . Details are found in Supplementary Material Section B. This translation corresponds to a reduction of parameters because an + +affine patch of $\mathcal{M}_{\widehat{\mathcal{C}}}^{1,1}$ lives in $(\mathbb{R}^1)^m$ , while an affine patch of $\mathcal{M}_{\mathcal{C}}^{L}$ lives in $(\mathbb{R}^2)^m$ . The analogous is true for $\mathcal{M}_{\widehat{\mathcal{C}}}^{2,1}$ and $\mathcal{L}_{\mathcal{C}}^{X}$ . + +We call the methods that do not use this translation standard and for instance, write (L1).1 std. to denote it. By just writing (L1).1, we denote the implementation of this translation. + +# 2.1.2 Evaluating error + +In our experiments, each iteration starts by randomly generating a 3D line $L$ and $p$ points $X_{i}$ on this line. This line and these points are projected by randomly generated cameras to line and point correspondences in $(\mathbb{R}^2)^m$ . For fixed $\epsilon = 10^{-12}$ , randomly generated noise vectors $\sigma (\epsilon)$ of length $\epsilon$ are added in each factor. Our approaches then find points $Y_{i}\in \mathbb{R}^{3}$ that match the noisy correspondences. We measure the accuracy by taking the logarithm after averaging the relative error of reconstructed points: + +$$ +e = \log_ {1 0} \left(\frac {\sum \| Y _ {i} - X _ {i} \|}{p \epsilon}\right), \tag {13} +$$ + +where $\| Y_{i} - X_{i}\|$ denotes the Euclidean distance. The interpretation of this number is that the error in the data gets amplified by $10^{e}$ during the triangulation. + +# 2.1.3 Algorithms + +The pseudocodes for all approaches are included in the Supplementary Material. The pseudocode for approach (L1).1 is presented in Algorithm 1. We use the notation that for a column vector $X \in \mathbb{R}^n$ , $[X; 1] \in \mathbb{R}^{n+1}$ is the vector we get by adding a 1 as the last coordinate. Let $\iota$ be the function that scales a vector such that its last coordinate is 1, and then removes that coordinate. In lines 1 and 2 of the algorithm, we generate noisy data by introducing randomly generated noise $\sigma(\epsilon)$ as explained above, where $\epsilon = 10^{-12}$ is a fixed parameter for our experiments. Lines 3 and 4 correspond to the triangulation of the line and the point correspondences using the anchored point multiview variety at the line $L_0$ . Note that we solve the closest point problem in line 4 by computing the zeros of a system of polynomial (critical) equations using the solve function in HomotopyContinuation.jl [9]. Finally, line 5 compares the points obtained in the previous step with the original starting points, measuring the logarithmic average relative error of the $p$ points. + +Most other pseudocodes are similar to Algorithm 1. For instance, approach (L1).0 does not take into account the line, so it corresponds deleting lines 1 and 3 from Algorithm 1 and modifying the minimization domain in line 4, replacing $L_{0}$ by $\mathcal{M}_{\mathcal{C}}$ . In approach (L1).2, line 3 is modified + +so that $L_{0}$ is obtained by the span of two triangulated points instead of by intersecting two back-projected planes. + +Algorithm 1: One iteration of the (L1).1 std. method given a randomly generated camera arrangement $\mathcal{C}$ of $3\times 4$ matrices, a projective line $L$ spanned by two vectors of $\mathbb{R}^4$ , and $p$ points $X_{i}\in \mathbb{R}^{3}$ such that $[X_i;1]$ lie on $L$ . + +Input : $\mathcal{C} = (C_1, \dots, C_m)$ , $L, X_1, \dots, X_p$ + +Output: The log of the average relative error + +1 for $j$ from 1 to $m$ do + +$$ +\begin{array}{l} \text {\bf 2} \\ \text {\bf 3} \\ \text {\bf 4} \end{array} \left\lfloor \begin{array}{l} \text {\bf f o r} i \text {f r o m 1 t o p d o} \\ \text {\bf \hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\hbox {\c h i}}}}}}}}}}}}}}}}}}}}}} q _ {i, j} \leftarrow \iota (C _ {j} [ X _ {i}; 1 ]) + \sigma (\epsilon); \\ u _ {j} \leftarrow \iota (C _ {j} \cdot L) + \sigma (\epsilon); \end{array} \right. +$$ + +5 $L_0\gets$ nullspace $\left[C_1^T [u_1;1]\quad C_2^T [u_2;1]\right]^T;$ + +6 for $i$ from 1 to $p$ do + +$$ +\begin{array}{c}Y_{i}\leftarrow \\ \underset {X\in \mathbb{R}^{3}:[X;1]\in L_{0}}{\operatorname *{argmin}}\sum_{j = 1}^{m}(q_{i,j} - \iota (C_{j}[X;1]))^{2}; \end{array} +$$ + +8 $e\gets \log_{10}\left(\frac{1}{p\epsilon}\sum_{i = 1}^{p}\| Y_i - X_i\|\right)$ + +Return: $e$ + +# 2.2. Numerical results + +The main results of our numerical experiments are presented in Figures 3 to 5 and Tables 1 to 3. We compare the performance of the five different triangulation approaches for $p = 5$ and different number of $m$ cameras. In the tables, we present the median, mean and standard deviation $\sigma$ of the logarithmic average relative error and time. The results are displayed in histograms and tables created with Plots.jl [10]. + +Note that for two cameras, (L1).0 and (L1).2 are the exact same, and (L1).1 and (L1).4 are essentially the same. In Figure 3 and Table 1 we present results for $m = 2$ and therefore only include (L1).0, (L1).1 and (L1).3. We also compare with the standard implementations of (L1).1 and (L1).3 that do not use the linear isomorphisms of Theorem 1.3 to improve computation time as described in Section 2.1.1. This simulation is iterated 1000 times. + +Figure 4 and Table 2 compare the different reconstruction approaches for $m = 3$ cameras. The simulations are again iterated 1000 times. + +Figure 5 and Table 3 compare the different reconstruction approaches for $m = 4$ cameras. The simulations are iterated only 100 times, due to the longer run-time of the simulations. + +Our theoretical studies in Section 1, specifically Theorem 1.7, guarantee that methods such as (L1).1, (L1).2 and (L1).3 are less complex from the algebraic point of view, than (L1).0 for triangulation of points incident to a line. Ad- + +![](images/0175f04a2e365855a322993fdc8857b162070b3e7ac30ab52c9620d79dd58e21.jpg) +Accuracy: 2 cameras, 5 point correspondences +Speed: 2 cameras, 5 point correspondences + +![](images/22e3360de14aec2cce14ca0de0569b1e13fc38f5c61b988d4374fdfaeac40563.jpg) +Figure 3: Triangulation of $p = 5$ point correspondences incident to a line for $m = 2$ views with complete visibility. 1000 iterations. The histograms show the frequency of the average relative error and the running time of the depicted triangulation methods. + +ditionally, using Section 2.1.1, we were able to improve the computation speed while retaining comparable accuracy, which is exemplified by the two cases (L1).1 vs. (L1).1 std. and (L1).3 vs. (L1).3 std. in Table 1. + +Finally, our numerical simulations show that in our HomotopyContinuation.jl implementation, the (L1).1 approach is the fastest across all number of views, but least accurate. In the case of $m = 3$ , (L1).4 is the most accurate and faster than the traditional (L1).0 method. Even for $m = 4$ , (L1).4 is the most accurate but is notably slower than (L1).0. Increasing the number of cameras $m$ improves accuracy and reduces speed, but the exact impact depends on the approach. Among (L1).1-4, (L1).1 is least affected in both regards, and (L1).4 is most affected in both regards. The speed can be thought of as an affine function in the number of point correspondences $p$ . Reconstructing the line correspondence takes some fixed amount of time independent of $p$ and then all $p$ correspondences are independently reconstructed. + +# 3. Conclusion and Future Work + +This work studied the triangulation of incident points and lines. We introduced the anchored line multiview varieties and showed that they correspond to multiview varieties arising from projections $\mathbb{P}^2\longrightarrow \mathbb{P}^1$ and $\mathbb{P}^1\longrightarrow \mathbb{P}^1$ . In addition, we proved that they are less complex for reconstruction via critical points. These theoretical results are also + +
Accuracymedianmeanσ
(L1).01.0311.2431.213
(L1).11.4891.7021.085
(L1).1 std.1.3741.5480.990
(L1).31.2501.4490.964
(L1).3 std.1.1471.3931.205
+ +
Speedmedianmeanσ
(L1).00.1700.2010.220
(L1).10.04300.04700.0207
(L1).1 std.0.1220.1410.202
(L1).30.08380.09410.0464
(L1).3 std.0.1670.1880.0764
+ +Table 1: Triangulation of $p = 5$ point correspondences incident to a line for $m = 2$ views with complete visibility. 1000 iterations. The tables show the accuracy and speed of the depicted triangulation methods in terms of median, mean and standard deviation. + +![](images/d7844462ac936ab3829315c0261a2a06e2d7b0ed717a40c4d5da5c6411dee73e.jpg) + +![](images/b3799ef3cda454409458f2ed5ead66ac62d356c3bc81e6219458d6fd99f4a625.jpg) +Speed: 3 cameras, 5 point correspondences +Figure 4: Triangulation of $p = 5$ point correspondences incident to a line for $m = 3$ views with complete visibility. 1000 iterations. The histograms show the frequency of the average relative error and the running time of the depicted triangulation methods. + +aligned with the numerical experiments we conducted. In particular, the proposed methods in Section 2 compare different methods for triangulating a set of points incident to a line. We highlight that according to our experiments, the use of the theoretical results in Section 1 allows for a notably faster triangulation while preserving the accuracy of the traditional method for point reconstruction. + +We hope this work motivates the use of new algebraic constraints in the triangulation of incident points and lines. + +
Accuracymedianmeanσ
(L1).00.6200.8171.023
(L1).11.5621.7811.128
(L1).20.8781.1250.998
(L1).31.0111.2431.078
(L1).40.4070.4990.901
Speedmedianmeanσ
(L1).00.9841.0600.514
(L1).10.08060.08580.0318
(L1).20.4310.4560.155
(L1).30.3040.3310.134
(L1).40.6160.7630.500
+ +![](images/416a17a83418e470fa4023c56f763904a3ecee623e65ca56a26a88ab5fcab6fa.jpg) +Speed: 4 cameras, 5 point correspondences + +![](images/0ce8bbba55b9911187364903051bca646c121f6452cdcab3358a919daf52771d.jpg) +Figure 5: Triangulation of $p = 5$ point correspondences incident to a line for $m = 4$ views with complete visibility. 100 iterations. The histograms show the frequency of the average relative error and the running time of the depicted triangulation methods. + +Some questions that arose in the development of this work, and that we believe are worth studying, are included below. + +- A triangulation approach inspired by [2] consists of reconstructing the 3D line $L$ by fitting it to $\mathcal{L}_{\mathcal{C}}$ solving the following optimization problem + +$$ +\text {m i n i m i z e} \quad \sum_ {i j} d \left(x _ {i j}, \ell_ {i}\right) ^ {2} \text {s u b j e c t t o} \quad \ell \in \mathcal {L} _ {\mathcal {C}}. \tag {14} +$$ + +Table 2: Triangulation of $p = 5$ point correspondences incident to a line for $m = 3$ views with complete visibility. 1000 iterations. The tables show the accuracy and speed of the depicted triangulation methods in terms of median, mean and standard deviation. + +
Accuracymedianmeanσ
(L1).00.3850.5510.776
(L1).11.4371.7621.249
(L1).21.0951.0770.845
(L1).30.8171.1101.094
(L1).40.2030.4211.324
Speedmedianmeanσ
(L1).04.3654.6751.576
(L1).10.1410.1490.0431
(L1).21.8091.8630.323
(L1).31.1071.1570.234
(L1).46.5997.3913.068
+ +Table 3: Triangulation of $p = 5$ point correspondences incident to a line for $m = 4$ views with complete visibility. 100 iterations. The tables show the accuracy and speed of the depicted triangulation methods in terms of median, mean and standard deviation. + +Here, $x_{ij}$ denotes the $j = 1,\dots ,p$ point correspondences across $i = 1,\ldots ,m$ cameras, and $d(x_{ij},\ell_i)$ is the affine distance from a line in $\mathbb{R}^2$ to the point, meaning + +$$ +d \left(x _ {i j}, \ell_ {i}\right) = \frac {\left| 1 + \left(\ell_ {i}\right) _ {1} \left(x _ {i j}\right) _ {1} + \left(\ell_ {i}\right) _ {2} \left(x _ {i j}\right) _ {2} \right|}{\sqrt {\left(\ell_ {i}\right) _ {1} ^ {2} + \left(\ell_ {i}\right) _ {2} ^ {2}}}. \tag {15} +$$ + +We believe that this method is more accurate than (L1).4, but with longer running time, due to the appearance of fractions. + +- Consider the point-line problem (P1) defined by one point incident to $l$ lines, illustrated in Figure 1. Can similar approaches to (L1).1-4 be formulated for the triangulation of the (P1) setting? Can they be used to improve accuracy and/or the speed of triangulation? +- Our implementation for the numerical experiments does not directly correspond to the specialized software and hardware used for triangulation in practice. However, our results motivate that patterns displayed in our figures and tables could translate to such settings. This could improve the efficiency of current specialized solvers. + +Acknowledgements. The authors thank Kathlen Kohn for helpful discussions, Lukas Gustafsson for the initial insight of Theorem 1.3, and Viktor Larsson for pointing out relevant references at the start of this project. Felix Rydell and Angelica Torres were supported by the Knut and Alice Wallenberg Foundation within their WASP (Wallenberg AI, Autonomous Systems and Software Program) AI/Math initiative. Elima Shehu is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), Projektnummer 445466444. Angelica Torres is currently supported by the Spanish State Research Agency, through + +the Severo Ochoa and María de Maeztu Program for Centers and Units of Excellence in R&D. + +# References + +[1] Sameer Agarwal, Andrew Pryhuber, and Rekha R Thomas. Ideals of the multiview variety. IEEE transactions on pattern analysis and machine intelligence, 2019. 1, 2 +[2] Adrien Bartoli and Peter Sturm. Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer Vision and Image Understanding, 100(3):416-441, 2005. 3, 8 +[3] Paul A Beardsley, Andrew Zisserman, and David William Murray. Navigation using affine structure from motion. Lecture Notes in Computer Science, 801:85-96, 1994. 2, 3 +[4] Paul A Beardsley, Andrew Zisserman, and David William Murray. Sequential updating of projective and affine structure from motion. International journal of computer vision, 23:235-259, 1997. 2, 3 +[5] Jeff Bezanson, Stefan Karpinski, Viral B Shah, and Alan Edelman. Julia: A fast dynamic language for technical computing. arXiv preprint arXiv:1209.5145, 2012. 5 +[6] Paul Breiding, Timothy Duff, Lukas Gustafsson, Felix Rydell, and Elima Shehu. Line multiview ideals. arXiv preprint arXiv:2303.02066, 2023. 2 +[7] Paul Breiding, Felix Rydell, Elima Shehu, and Angélica Torres. Line multiview varieties. SIAM Journal on Applied Algebra and Geometry, 7(2):470-504, 2023. 1, 2, 4 +[8] Paul Breiding and Sascha Timme. Homotopycontinuation. jl: A package for homotopy continuation in julia. In International Congress on Mathematical Software, pages 458-465.Springer,2018.5 +[9] Paul Breiding and Sascha Timme. HomotopyContinuation.jl: A Package for Homotopy Continuation in Julia. In Mathematical Software - ICMS 2018, pages 458-465, Cham, 2018. Springer International Publishing. 2, 6 +[10] Tom Breloff and other contributors. JuliaPlots/Plots.jl. 6 +[11] Jan Draisma, Emil Horobet, Giorgio Ottaviani, Bernd Sturmfels, and Rekha R Thomas. The euclidean distance degree of an algebraic variety. Foundations of computational mathematics, 16(1):99-149, 2016. 2, 4 +[12] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajdla. Plmp-point-line minimal problems in complete multi-view visibility. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1675-1684, 2019. 1 +[13] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajda. PI\$\_1\$\$p - point-line minimal problems under partial visibility in three views. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 175–192, Cham, 2020. Springer International Publishing. 1 +[14] A. Elqursh and A. Elgammal. Line-based relative pose estimation. In Conference on Computer Vision and Pattern Recognition, 2011. 3 + +[15] Ricardo Fabbri, Timothy Duff, Hongyi Fan, Margaret H Regan, David da Costa de Pinho, Elias Tsigaridas, Charles W Wampler, Jonathan D Hauenstein, Peter J Giblin, Benjamin Kimia, et al. Trplp-trifocal relative pose from lines at points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12073-12083, 2020. 1 +[16] Olivier Faugeras. Three-dimensional computer vision: A geometric viewpoint. MIT press, Cambridge, MA, 1993. 1 +[17] Olivier Faugeras and Bernard Mourrain. On the geometry and algebra of the point and line correspondences between n images. In Proceedings of IEEE International Conference on Computer Vision, pages 951-956. IEEE, 1995. 1, 2 +[18] Andreas Gathmann. Algebraic geometry, 2019/20. Class Notes TU Kaiserslautern. Available at https://www.mathematik.uni-kl.de/~gathmann/de/alggeom.php.4 +[19] Corey Harris and Daniel Lowengrub. The Chern-mather class of the multiview variety. Communications in Algebra, 46(6):2488-2499, 2018. 2 +[20] Richard I. Hartley. Lines and points in three views and the trifocal tensor. International Journal of Computer Vision, 22(2):125-140, 1997. 2, 3 +[21] Richard I Hartley and Peter Sturm. Triangulation. Computer vision and image understanding, 68(2):146-157, 1997. 2 +[22] Richard I. Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 2 +[23] Manuel Hofer, Andreas Wendel, and Horst Bischof. Incremental line-based 3d reconstruction using geometric constraints. The British Machine Vision Association and Society for Pattern Recognition, 2013. 3 +[24] Kun Huang, Yifan Wang, Zihan Zhou, Tianjiao Ding, Shenghua Gao, and Yi Ma. Learning to parse wireframes in images of man-made environments. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 626-635, 2018. 1 +[25] Javier Civera Joan Solà, Teresa Vidal-Calleja and José María Martínez Montiel. Impact of landmark parametrization on monocular ekf-slam with points and lines. International Journal of Computer Vision, 97(3):339-368, 2012. 3, 9 +[26] Kenichi Kanatani. Statistical optimization for geometric computation: theory and practice. Courier Corporation, 2005. 2 +[27] Kenichi Kanatani, Yasuyuki Sugaya, and Hirotaka Niitsuma. Triangulation from two views revisited: Hartley-sturm vs. optimal correction. practice, 4(5):99, 2008. 2 +[28] Yasushi Kanazawa and Kenichi Kanatani. Reliability of 3- d reconstruction by stereo vision. IEICE TRANSACTIONS on Information and Systems, 78(10):1301-1306, 1995. 2 +[29] Joseph David Kileel. Algebraic Geometry for Computer Vision. ProQuest LLC, Ann Arbor, MI, 2017. Thesis (Ph.D.)-University of California, Berkeley. 1, 2 +[30] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21445-21455, June 2023. 1 + +[31] Laurentiu G. Maxim, Jose I. Rodriguez, and Botong Wang. Euclidean distance degree of the multiview variety. SIAM Journal on Applied Algebra and Geometry, 4(1):28-48, 2020. 1, 2 +[32] Branislav Micusik and Horst Wildenauer. Structure from motion with line segments under relaxed endpoint constraints. In 2014 2nd International Conference on 3D Vision, volume 1, pages 13-19, 2014. 3 +[33] Long Quan and T. Kanade. Affine structure from line correspondences with uncalibrated affine cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(8):834-845, 1997. 3 +[34] Grant Schindler, Panchapagesan Krishnamurthy, and Frank Dellaert. Line-based structure from motion for urban environments. Third International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT'06), pages 846-853, 2006. 1 +[35] Henrik Stewénius, Frederik Schaffalitzky, and David Nister. How hard is 3-view triangulation really? In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 1, pages 686-693. IEEE, 2005. 3 +[36] Matthew Trager, Martial Hebert, and Jean Ponce. The joint image handbook. In Proceedings of the IEEE international conference on computer vision, pages 909-917, 2015. 2, 4 \ No newline at end of file diff --git a/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/images.zip b/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..15fdef2d560df774afaaa6c31583a9f9fd908ef1 --- /dev/null +++ b/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bae8243eff0793391904aeda28d01dafa775209df031073313875801c94d2061 +size 330633 diff --git a/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/layout.json b/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9855ec6c9d94dbcc4ffb2c4ea3f9d488885d284e --- /dev/null +++ b/2023/Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences/layout.json @@ -0,0 +1,13639 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 103, + 541, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 541, + 138 + ], + "type": "text", + "content": "Theoretical and Numerical Analysis of 3D Reconstruction Using Point and Line Incidences" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 161, + 212, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 161, + 212, + 175 + ], + "spans": [ + { + "bbox": [ + 149, + 161, + 212, + 175 + ], + "type": "text", + "content": "Felix Rydell" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 176, + 268, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 176, + 268, + 190 + ], + "spans": [ + { + "bbox": [ + 94, + 176, + 268, + 190 + ], + "type": "text", + "content": "KTH Royal Institute of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 190, + 280, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 190, + 280, + 203 + ], + "spans": [ + { + "bbox": [ + 83, + 190, + 280, + 203 + ], + "type": "text", + "content": "Lindstedtsvagen 25, Stockholm, Sweden" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 205, + 220, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 205, + 220, + 217 + ], + "spans": [ + { + "bbox": [ + 143, + 205, + 220, + 217 + ], + "type": "text", + "content": "felixry@kth.se" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 376, + 162, + 441, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 162, + 441, + 175 + ], + "spans": [ + { + "bbox": [ + 376, + 162, + 441, + 175 + ], + "type": "text", + "content": "Elima Shehu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 175, + 503, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 175, + 503, + 189 + ], + "spans": [ + { + "bbox": [ + 315, + 175, + 503, + 189 + ], + "type": "text", + "content": "University of Osnabrück and MPI MiS" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 190, + 511, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 190, + 511, + 203 + ], + "spans": [ + { + "bbox": [ + 307, + 190, + 511, + 203 + ], + "type": "text", + "content": "Albrechtstrasse 28a, Osnabrück, Germany" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 204, + 492, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 204, + 492, + 218 + ], + "spans": [ + { + "bbox": [ + 326, + 204, + 492, + 218 + ], + "type": "text", + "content": "Inselstrasse 22, Leipzig, Germany" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 348, + 220, + 468, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 220, + 468, + 231 + ], + "spans": [ + { + "bbox": [ + 348, + 220, + 468, + 231 + ], + "type": "text", + "content": "elima.shehu@mis.mpg.de" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 257, + 237, + 337, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 237, + 337, + 251 + ], + "spans": [ + { + "bbox": [ + 257, + 237, + 337, + 251 + ], + "type": "text", + "content": "Angélica Torres" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 222, + 252, + 372, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 252, + 372, + 264 + ], + "spans": [ + { + "bbox": [ + 222, + 252, + 372, + 264 + ], + "type": "text", + "content": "Centre de Recerca Matemàtica" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 195, + 266, + 399, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 266, + 399, + 279 + ], + "spans": [ + { + "bbox": [ + 195, + 266, + 399, + 279 + ], + "type": "text", + "content": "Edifici C, Campus UAB, Bellaterra, Spain" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 255, + 281, + 338, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 281, + 338, + 291 + ], + "spans": [ + { + "bbox": [ + 255, + 281, + 338, + 291 + ], + "type": "text", + "content": "atorres@crm.cat" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 143, + 320, + 192, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 320, + 192, + 333 + ], + "spans": [ + { + "bbox": [ + 143, + 320, + 192, + 333 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 346, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 289, + 502 + ], + "type": "text", + "content": "We study the joint image of lines incident to points, meaning the set of image tuples obtained from fixed cameras observing a varying 3D point-line incidence. We prove a formula for the number of complex critical points of the triangulation problem that aims to compute a 3D point-line incidence from noisy images. Our formula works for an arbitrary number of images and measures the intrinsic difficulty of this triangulation. Additionally, we conduct numerical experiments using homotopy continuation methods, comparing different approaches of triangulation of such incidences. In our setup, exploiting the incidence relations gives a notably faster point reconstruction with comparable accuracy." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 525, + 116, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 116, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 116, + 538 + ], + "type": "text", + "content": "Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 545, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 653 + ], + "type": "text", + "content": "The Structure-from-Motion pipeline aims to estimate the camera positions and the relative position of the objects present in a given set of images. The process starts by identifying point and line features in one image that are recognizable as the same points or lines in another. These matched features, that we call correspondences, are used to estimate the camera position and orientation, which then allow for the triangulation step where the position of the points and lines is estimated." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 288, + 714 + ], + "type": "text", + "content": "In this work, we study the triangulation of points and lines satisfying a certain incidence relation. Specifically, we focus on the triangulation of points contained in a single line; see Figure 1. We note that our methods can also be used when multiple lines going through a fixed point. We" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 353, + 323, + 500, + 399 + ], + "blocks": [ + { + "bbox": [ + 353, + 323, + 500, + 399 + ], + "lines": [ + { + "bbox": [ + 353, + 323, + 500, + 399 + ], + "spans": [ + { + "bbox": [ + 353, + 323, + 500, + 399 + ], + "type": "image", + "image_path": "f9dcca4617d849064605759b8dc46da8cce53217a254be158992f305c8ec7154.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 410, + 547, + 449 + ], + "lines": [ + { + "bbox": [ + 305, + 410, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 547, + 449 + ], + "type": "text", + "content": "Figure 1: The illustration of two different types of incidence relations: (L1) represents the scenario where multiple points are incident to a line, while (P1) represents the scenario where multiple lines are incident to a point." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 472, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 508 + ], + "type": "text", + "content": "develop the theory for the triangulation of these problems for an arbitrary number of pinhole cameras assuming complete visibility." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 509, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 546, + 652 + ], + "type": "text", + "content": "The triangulation of points is well understood and, in practice, it is efficiently implemented. However, to the best of our knowledge, incidence relations are not considered in these implementations. The inclusion of line features and incidence relations in the triangulation process could give a more accurate estimation of scenes as lines are more robust to noise than points, and incidence relations appear frequently in interior scenes and man-made scenarios; see [34], [24], [30]. Additionally, in some real data sets standard feature detection algorithms fail due to a lack of point correspondences but succeed when line correspondences are taken into account [15]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "The main tools in our work are (algebraic) varieties, that is, the vanishing sets of systems of polynomial equations. Algebraic varieties have been used extensively to study triangulation of point correspondences [1, 17, 31], line correspondences [7, 16, 29] and minimal problems [12, 13] to" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3748" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "name a few. In these works, algebraic varieties arise naturally due to the algebraic nature of pinhole cameras." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "text", + "content": "A pinhole camera, is modeled by a (complex) projective linear map " + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "inline_equation", + "content": "C: \\mathbb{P}^3 \\dashrightarrow \\mathbb{P}^2" + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "text", + "content": " defined by a " + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "inline_equation", + "content": "3 \\times 4" + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "text", + "content": " matrix " + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "text", + "content": " of full rank that takes a point " + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{P}^3" + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "text", + "content": " and sends it to " + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "inline_equation", + "content": "CX \\in \\mathbb{P}^2" + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "text", + "content": ". A camera arrangement with " + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "inline_equation", + "content": "m \\geq 2" + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "text", + "content": " cameras is denoted by " + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = (C_1, \\ldots, C_m)" + }, + { + "bbox": [ + 46, + 96, + 287, + 156 + ], + "type": "text", + "content": ", and the joint camera map" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 102, + 165, + 287, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 165, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 102, + 165, + 287, + 194 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Phi_ {\\mathcal {C}}: \\mathbb {P} ^ {3} \\dashrightarrow (\\mathbb {P} ^ {2}) ^ {m}, \\tag {1} \\\\ X \\mapsto (C _ {1} X, \\dots , C _ {m} X), \\\\ \\end{array}", + "image_path": "8c88e6dedd080fa58b759a4df3654fc0ff88e082b53246f62ddf5c4e9114d82d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 203, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 203, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 288, + 312 + ], + "type": "text", + "content": "models the process of taking the images of a point " + }, + { + "bbox": [ + 46, + 203, + 288, + 312 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 203, + 288, + 312 + ], + "type": "text", + "content": " in homogeneous coordinates with " + }, + { + "bbox": [ + 46, + 203, + 288, + 312 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 203, + 288, + 312 + ], + "type": "text", + "content": " cameras. For fixed cameras, the (point) multiview variety " + }, + { + "bbox": [ + 46, + 203, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 203, + 288, + 312 + ], + "type": "text", + "content": " is the smallest variety that contains all point correspondences. An extensive account of the pinhole cameras is given by [22], and a survey of the multiview variety is found in [36]. The joint camera map can be extended from the space of points " + }, + { + "bbox": [ + 46, + 203, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^3" + }, + { + "bbox": [ + 46, + 203, + 288, + 312 + ], + "type": "text", + "content": " to the space of lines, denoted by " + }, + { + "bbox": [ + 46, + 203, + 288, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{Gr}(1,\\mathbb{P}^3)" + }, + { + "bbox": [ + 46, + 203, + 288, + 312 + ], + "type": "text", + "content": ", where each line is parametrized as the span of two points. The map" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 319, + 287, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 319, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 83, + 319, + 287, + 348 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Upsilon_ {\\mathcal {C}}: \\operatorname {G r} (1, \\mathbb {P} ^ {3}) \\dashrightarrow (\\mathbb {P} ^ {2}) ^ {m}, \\tag {2} \\\\ L \\mapsto (C _ {1} \\cdot L, \\dots , C _ {m} \\cdot L). \\\\ \\end{array}", + "image_path": "52d2bdde049ed7a86f63c6e64c6f068ba0d5195bb9ca491aa049f5a8f41d33b1.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": "models the image of a line " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": " taken by the cameras of " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": ". To be precise, if " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "u, v" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": " span the line " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^3" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "C \\cdot L" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\ell = Cu \\times Cv \\in \\mathbb{P}^2" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": " is the cross product. Observe that " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\{x \\in \\mathbb{P}^2 : \\ell^T x = 0\\}" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": " equals " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\operatorname{span}\\{Cu, Cv\\}" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": ". Recently, in [7], the authors study the line multiview variety referred to as " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": ", which is the smallest variety containing all line correspondences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "text", + "content": "Our main contribution is the definition and study of the anchored point and line multiview varieties. Given a fixed line " + }, + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^3" + }, + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "text", + "content": ", the anchored point multiview variety, " + }, + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_C^L" + }, + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "text", + "content": ", is defined as the smallest variety containing all point correspondences coming from points in " + }, + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "text", + "content": "; and the anchored line multiview variety, " + }, + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C^X" + }, + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "text", + "content": ", is the smallest variety containing the line correspondences coming from lines passing through a given point " + }, + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{P}^3" + }, + { + "bbox": [ + 46, + 441, + 287, + 536 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 536, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 585 + ], + "type": "text", + "content": "For these new varieties we prove formulas for their Euclidean distance degree (EDD), which are a measurement of the complexity for error correction using exact algebraic methods [11]. Specifically, we prove the following theorem:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "content": "Theorem 0.1. Let " + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "content": " be a generic arrangement of " + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 593, + 287, + 616 + ], + "type": "text", + "content": " cameras." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 624, + 259, + 659 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 624, + 165, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 165, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 165, + 638 + ], + "type": "text", + "content": "1. " + }, + { + "bbox": [ + 55, + 624, + 165, + 638 + ], + "type": "inline_equation", + "content": "\\operatorname{EDD}(\\mathcal{M}_{\\mathcal{C}}^{L}) = 3m - 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 643, + 259, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 643, + 259, + 659 + ], + "spans": [ + { + "bbox": [ + 53, + 643, + 259, + 659 + ], + "type": "text", + "content": "2. If " + }, + { + "bbox": [ + 53, + 643, + 259, + 659 + ], + "type": "inline_equation", + "content": "m \\geq 3" + }, + { + "bbox": [ + 53, + 643, + 259, + 659 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 53, + 643, + 259, + 659 + ], + "type": "inline_equation", + "content": "\\mathrm{EDD}(\\mathcal{L}_{\\mathcal{C}}^{X}) = \\frac{9}{2} m^{2} - \\frac{19}{2} m + 3" + }, + { + "bbox": [ + 53, + 643, + 259, + 659 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "We provide a precise definition of this degree in Section 1.2. Previous work on EDD for multiview varieties includes [19, 31] and [7, Section 5]. The EDD of the point multiview variety is " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{EDD}(\\mathcal{M}_{\\mathcal{C}}) = \\frac{9}{2} m^3 -" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 155 + ], + "type": "inline_equation", + "content": "\\frac{21}{2} m^2 + 8m - 4" + }, + { + "bbox": [ + 305, + 72, + 545, + 155 + ], + "type": "text", + "content": ", according to [31]. The fact that the EDDs of the anchored multiview varieties are polynomials of smaller degrees suggests that they are less complex for the purpose of data correction. This conclusion is backed by the results of our numerical experiments using HomotopyContinuation.jl [9], presented in Section 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "content": "Another contribution of our work is the numerical simulations for the triangulation of points contained in a line using the anchored multiview varieties. We present different approaches to triangulating data of type (L1) that differ from the traditional method of fitting point correspondences to the multiview variety " + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C}}" + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "inline_equation", + "content": "m = 2" + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "content": " views, our implementation is notably faster than the traditional triangulation of points described above, while the accuracy is comparable. For " + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "inline_equation", + "content": "m = 3" + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "content": " views we get both higher accuracy and faster speed by using " + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{C}}" + }, + { + "bbox": [ + 304, + 156, + 546, + 348 + ], + "type": "text", + "content": ", compared to usual point triangulation. All of our proposed methods outperform the traditional approaches in terms of run-time and give a comparably accurate result. In practice, special software and hardware are used for triangulation, and based on our experiments, we believe that these approaches could be implemented efficiently with good results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 355, + 373, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 355, + 373, + 367 + ], + "spans": [ + { + "bbox": [ + 306, + 355, + 373, + 367 + ], + "type": "text", + "content": "Related work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 374, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 494 + ], + "type": "text", + "content": "Algebra. Algebraic geometry, whose connection to computer vision is well established, is our main tool for the theoretical study of the triangulation of problem (L1). Fundamental theory regarding the point multiview variety is found in [17, 20, 22], in particular for two and three views. The paper [36] by Trager et. al. serves as a survey for the algebraic properties of this multiview variety. For applications, finding a good set of polynomial constraints satisfied by the point correspondences is important, this is precisely the work of [1] by Agarwal et al." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 495, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 614 + ], + "type": "text", + "content": "Regarding the algebra of lines in a computer vision setting, Kileel presented in [29, Definition 3.9, Theorem 3.10] several types of multiview varieties with respect to a point-line incidence relation in " + }, + { + "bbox": [ + 304, + 495, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^3" + }, + { + "bbox": [ + 304, + 495, + 545, + 614 + ], + "type": "text", + "content": " in 3 views, and their basic properties. One of those types is called the LLL-multiview variety and constitutes a special case of the more recent work by Breiding et. al. [7], where they study algebraic properties of the line multiview variety. Furthermore, a recent manuscript [6] studies the polynomial constraints satisfied by line correspondences." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "content": "Projective Reconstruction. Different approaches to the triangulation of points have been considered in the literature, in particular for two views [3, 4, 21, 26-28]. Hartley and Sturm compare many different approaches in [21], including the \"midpoint\" method. The midpoint method inputs a point correspondence in two views and outputs the midpoint of the line segment determined by the points on" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3749" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "the two back-projected lines that are closest to each other. In other words, it finds the midpoint of the common perpendicular to the two back-projected lines. This is a natural approach, but it has several downsides, also pointed out in [3, 4]. Importantly, it is not projectively invariant. A projectively invariant reconstruction has the property that acting on the cameras by a global projective transformation induces the same action on the triangulated point or line. Instead, Hartley and Sturm propose minimizing the reprojection error as the optimal triangulation method, which is accepted as a standard formulation [35] and it is projectively invariant. In [35], the focus is on triangulation via minimizing reprojections errors in three views and the authors point out that three views often lead to greater stability and stronger disambiguation compared to two views." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 289, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 562 + ], + "type": "text", + "content": "Implementation of Line Reconstruction. From the algorithmic point of view, the simultaneous reconstruction of point and line features have been studied specially with the goal of reconstructing line segments. For example, [2] provides a thorough overview of the structure-from-motion pipeline using lines, going through different methods for line parameterization and error correction depending on such parameterizations. In [33], Quan and Takeo study the algebraic structure of line correspondences with uncalibrated affine cameras. They reduce the problem of reconstructing affine lines to the reconstruction of projective points in a lower-dimensional projective camera. In [20], Hartley and coauthors provide an algorithm for the reconstruction of point and line features where 3D lines are parameterized by their projections in 2 views (as the intersection of two back-projected planes). Micusik and Wildenauer reconstruct lines to estimate line segments, but only use incident points for error correction [32]. Furthermore in [23], the authors propose a technique for 3D reconstruction incorporating line segments. They assert that this method surpasses traditional approaches, especially in efficiency while getting accurate results. Finally, in [14], the authors present a framework for the computation of the relative motion between two images using a triplet of lines." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "This paper is structured as follows. In the first part of Section 1 we formally introduce anchored multiview varieties and study some properties. In the second part of Section 1, we study their smoothness and find their Euclidean Distance Degree. Finally, in Section 2, we provide a numerical analysis of different approaches to reconstructing point correspondences incident to a line correspondence. The proofs of all our results are included in the Supplementary Material together with a small background on the concepts of Algebraic Geometry used for these results and the code for the numerical experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 71, + 474, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 474, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 474, + 84 + ], + "type": "text", + "content": "1. Anchored Multiview Varieties" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 91, + 547, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 126 + ], + "type": "text", + "content": "A camera refers to a full-rank " + }, + { + "bbox": [ + 304, + 91, + 547, + 126 + ], + "type": "inline_equation", + "content": "3 \\times 4" + }, + { + "bbox": [ + 304, + 91, + 547, + 126 + ], + "type": "text", + "content": " matrix. A camera arrangement is a collection " + }, + { + "bbox": [ + 304, + 91, + 547, + 126 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = (C_1, \\ldots, C_m)" + }, + { + "bbox": [ + 304, + 91, + 547, + 126 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 91, + 547, + 126 + ], + "type": "inline_equation", + "content": "m \\geq 2" + }, + { + "bbox": [ + 304, + 91, + 547, + 126 + ], + "type": "text", + "content": " cameras whose centers, i.e. kernels, are all distinct." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": "A variety is the solution set to a system of polynomial equations. The Zariski closure of a set " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": " is the smallest variety containing " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": ". We work in complex projective space, denoted " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^n" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": ". This is defined as " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "(\\mathbb{C}^n \\setminus \\{0\\}) / \\sim" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "x \\sim y" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": " differ by a non-zero constant. The set of lines in " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^n" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\operatorname{Gr}(1, \\mathbb{P}^n)" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": ". In the language of algebraic geometry, it is called the Grassmanian of lines in " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^n" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": ". In general we denote with upper case letters world objects and with lower case letters image objects. For a rational map " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\varphi: X \\dashrightarrow Y" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": ", we denote " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\varphi|_A" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": " the restriction of " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "inline_equation", + "content": "A \\subseteq X" + }, + { + "bbox": [ + 304, + 127, + 547, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 251, + 547, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 547, + 286 + ], + "type": "text", + "content": "Definition 1.1. Let " + }, + { + "bbox": [ + 304, + 251, + 547, + 286 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{P}^3" + }, + { + "bbox": [ + 304, + 251, + 547, + 286 + ], + "type": "text", + "content": " be a point distinct from all camera centers, and " + }, + { + "bbox": [ + 304, + 251, + 547, + 286 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 251, + 547, + 286 + ], + "type": "text", + "content": " a line in " + }, + { + "bbox": [ + 304, + 251, + 547, + 286 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^3" + }, + { + "bbox": [ + 304, + 251, + 547, + 286 + ], + "type": "text", + "content": " containing none of the camera centers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 291, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 545, + 316 + ], + "type": "text", + "content": "1. The anchored point multiview variety, denoted " + }, + { + "bbox": [ + 313, + 291, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_C^L" + }, + { + "bbox": [ + 313, + 291, + 545, + 316 + ], + "type": "text", + "content": ", is defined as the Zariski closure of the image of the map" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 367, + 319, + 545, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 319, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 367, + 319, + 545, + 346 + ], + "type": "interline_equation", + "content": "\\begin{array}{c} \\Phi_ {\\mathcal {C}} | _ {L}: L \\dashrightarrow (\\mathbb {P} ^ {2}) ^ {m}, \\\\ X \\longmapsto (C _ {1} X, \\dots , C _ {m} X). \\end{array} \\tag {3}", + "image_path": "57882ebda4fdf46acdb024e357d4421387febb80bcccba816c408a5b5b1bad90.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 357, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 545, + 382 + ], + "type": "text", + "content": "2. The anchored line multiview variety, denoted " + }, + { + "bbox": [ + 313, + 357, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C^X" + }, + { + "bbox": [ + 313, + 357, + 545, + 382 + ], + "type": "text", + "content": ", is defined as the Zariski closure of the image of the map" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 385, + 545, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 385, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 340, + 385, + 545, + 415 + ], + "type": "interline_equation", + "content": "\\begin{array}{c} \\Upsilon_ {\\mathcal {C}} | _ {\\Lambda (X)}: \\Lambda (X) \\dashrightarrow (\\mathbb {P} ^ {2}) ^ {m}, \\\\ L \\longmapsto \\left(C _ {1} \\cdot L, \\dots , C _ {m} \\cdot L\\right), \\end{array} \\tag {4}", + "image_path": "75b3a9bb77a69d9f6a075394b62cceda43a7daea12e2973112c2d6b05fa63578.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 419, + 545, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 419, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 325, + 419, + 545, + 444 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 325, + 419, + 545, + 444 + ], + "type": "inline_equation", + "content": "\\Lambda(X)" + }, + { + "bbox": [ + 325, + 419, + 545, + 444 + ], + "type": "text", + "content": " denotes the set of lines in " + }, + { + "bbox": [ + 325, + 419, + 545, + 444 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^3" + }, + { + "bbox": [ + 325, + 419, + 545, + 444 + ], + "type": "text", + "content": " that contain " + }, + { + "bbox": [ + 325, + 419, + 545, + 444 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 325, + 419, + 545, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 325, + 419, + 545, + 444 + ], + "type": "inline_equation", + "content": "\\ell_i = C_i \\cdot L" + }, + { + "bbox": [ + 325, + 419, + 545, + 444 + ], + "type": "text", + "content": " is defined as in the introduction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "text", + "content": "The anchored point multiview variety " + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C}}^{L}" + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "text", + "content": " is the smallest variety that contains all point correspondences " + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathcal{C}}(X)" + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "inline_equation", + "content": "X \\in L" + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "text", + "content": ". Similarly, the anchored line multiview variety " + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{C}}^{X}" + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "text", + "content": " is the smallest variety that contains all line correspondences " + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\Upsilon_{\\mathcal{C}}(L)" + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "text", + "content": " meeting " + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 448, + 545, + 544 + ], + "type": "text", + "content": " and no center. We highlight that our definition of anchored multiview varieties is different from the anchored features in [25] for monocular EKF-SLAM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 548, + 545, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 585 + ], + "type": "text", + "content": "Proposition 1.2. Consider an arrangement of " + }, + { + "bbox": [ + 304, + 548, + 545, + 585 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 548, + 545, + 585 + ], + "type": "text", + "content": " cameras " + }, + { + "bbox": [ + 304, + 548, + 545, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = (C_1, \\ldots, C_m)" + }, + { + "bbox": [ + 304, + 548, + 545, + 585 + ], + "type": "text", + "content": ", a point " + }, + { + "bbox": [ + 304, + 548, + 545, + 585 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{P}^3" + }, + { + "bbox": [ + 304, + 548, + 545, + 585 + ], + "type": "text", + "content": " and a line " + }, + { + "bbox": [ + 304, + 548, + 545, + 585 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 548, + 545, + 585 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 548, + 545, + 585 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^3" + }, + { + "bbox": [ + 304, + 548, + 545, + 585 + ], + "type": "text", + "content": " satisfying the conditions of Definition 1.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 589, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 545, + 613 + ], + "type": "text", + "content": "1. If there are two different camera centers " + }, + { + "bbox": [ + 313, + 589, + 545, + 613 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 589, + 545, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 589, + 545, + 613 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 313, + 589, + 545, + 613 + ], + "type": "text", + "content": " such that the span of " + }, + { + "bbox": [ + 313, + 589, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\{c_{i}, c_{j}, L\\}" + }, + { + "bbox": [ + 313, + 589, + 545, + 613 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 589, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^3" + }, + { + "bbox": [ + 313, + 589, + 545, + 613 + ], + "type": "text", + "content": ", then" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 337, + 616, + 545, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 616, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 337, + 616, + 545, + 632 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {\\mathcal {C}} ^ {L} = \\left\\{\\left(x _ {1}, \\dots , x _ {m}\\right) \\in \\mathcal {M} _ {\\mathcal {C}}: x _ {i} \\in C _ {i} \\cdot L \\right\\}. \\tag {5}", + "image_path": "95d72329b9b5d934cc48a5ba3d1e8c63b414a6b6abcd46641bdb8fa0d318e35d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 639, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 545, + 663 + ], + "type": "text", + "content": "2. If for each camera center " + }, + { + "bbox": [ + 313, + 639, + 545, + 663 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 639, + 545, + 663 + ], + "type": "text", + "content": ", the line spanned by " + }, + { + "bbox": [ + 313, + 639, + 545, + 663 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 639, + 545, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 639, + 545, + 663 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 639, + 545, + 663 + ], + "type": "text", + "content": " does not contain any other camera center, then" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 352, + 667, + 545, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 667, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 352, + 667, + 545, + 682 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {C}} ^ {X} = \\left\\{\\left(\\ell_ {1}, \\dots , \\ell_ {m}\\right) \\in \\mathcal {L} _ {\\mathcal {C}}: C _ {i} X \\in \\ell_ {i} \\right\\}. \\tag {6}", + "image_path": "245e1e2547e942134f298ba4cc7eaaf5d0d2194cf1cdc77a150a8c73ce2a109b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "The proofs for this and all our results are found in the Supplementary Material." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3750" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 272, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 272, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 272, + 84 + ], + "type": "text", + "content": "1.1. Properties of Anchored Multiview Varieties" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 286, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 286, + 149 + ], + "type": "text", + "content": "A fundamental property of the anchored multiview varieties is that they are linearly isomorphic to multiview varieties arising from projections " + }, + { + "bbox": [ + 46, + 90, + 286, + 149 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^1\\longrightarrow \\mathbb{P}^1" + }, + { + "bbox": [ + 46, + 90, + 286, + 149 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 90, + 286, + 149 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^2\\longrightarrow \\mathbb{P}^1" + }, + { + "bbox": [ + 46, + 90, + 286, + 149 + ], + "type": "text", + "content": " respectively. A linear isomorphism is a linear map (given by a matrix) with an inverse that is also linear." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "text", + "content": "Consider an arrangement " + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "text", + "content": " full-rank " + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "inline_equation", + "content": "2\\times 2" + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "text", + "content": " matrices, and an arrangement " + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "text", + "content": " full-rank " + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "inline_equation", + "content": "2\\times 3" + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "text", + "content": " matrices. Also here we assume " + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "inline_equation", + "content": "m\\geq 2" + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "text", + "content": ". We define " + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\widetilde{\\mathcal{C}}}^{1,1}" + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\widehat{\\mathcal{C}}}^{2,1}" + }, + { + "bbox": [ + 46, + 149, + 286, + 198 + ], + "type": "text", + "content": " respectively as the Zariski closure of the image of the maps" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 213, + 286, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 213, + 286, + 233 + ], + "spans": [ + { + "bbox": [ + 101, + 213, + 286, + 233 + ], + "type": "interline_equation", + "content": "\\Phi_ {\\tilde {C}}: \\mathbb {P} ^ {1} \\longrightarrow (\\mathbb {P} ^ {1}) ^ {m}, \\tag {7}", + "image_path": "b9566ac60aa39d782c47c5f7082738d0f357a2e152f92268c34c1d2d5b7b0330.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 230, + 231, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 230, + 231, + 244 + ], + "spans": [ + { + "bbox": [ + 125, + 230, + 231, + 244 + ], + "type": "interline_equation", + "content": "X \\longmapsto (\\widetilde {C} _ {1} X, \\dots , \\widetilde {C} _ {m} X)", + "image_path": "d34175290c2a09fa1b9e44ce8922e7e89caf2bacc5e8cb735bf1220616a847df.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 250, + 65, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 250, + 65, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 250, + 65, + 259 + ], + "type": "text", + "content": "and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 265, + 286, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 265, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 102, + 265, + 286, + 285 + ], + "type": "interline_equation", + "content": "\\Phi_ {\\widehat {C}}: \\mathbb {P} ^ {2} \\dashrightarrow (\\mathbb {P} ^ {1}) ^ {m}, \\tag {8}", + "image_path": "1c84a3f710243ec475dc8d5b92507bfce678507eeb8e9bb7532e51f0ec54761f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 281, + 230, + 296 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 281, + 230, + 296 + ], + "spans": [ + { + "bbox": [ + 126, + 281, + 230, + 296 + ], + "type": "interline_equation", + "content": "X \\longmapsto (\\widehat {C} _ {1} X, \\dots \\widehat {C} _ {m} X).", + "image_path": "05990bd134aae672ed761b214a6df8ed241c61fb13a548c4aaf0c192d2453b56.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 301, + 107, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 107, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 107, + 312 + ], + "type": "text", + "content": "Theorem 1.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 318, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 287, + 357 + ], + "type": "text", + "content": "1. Let " + }, + { + "bbox": [ + 55, + 318, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\phi_L: L \\to \\mathbb{P}^1" + }, + { + "bbox": [ + 55, + 318, + 287, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 318, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{C},i}: C_i \\cdot L \\to \\mathbb{P}^1" + }, + { + "bbox": [ + 55, + 318, + 287, + 357 + ], + "type": "text", + "content": " be any choices of linear isomorphisms. Let " + }, + { + "bbox": [ + 55, + 318, + 287, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{C}}" + }, + { + "bbox": [ + 55, + 318, + 287, + 357 + ], + "type": "text", + "content": " denote the arrangement of matrices " + }, + { + "bbox": [ + 55, + 318, + 287, + 357 + ], + "type": "inline_equation", + "content": "C_i := \\psi_{\\mathcal{C},i} \\circ C_i \\circ \\phi_L^{-1}" + }, + { + "bbox": [ + 55, + 318, + 287, + 357 + ], + "type": "text", + "content": ". Then" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 90, + 373, + 286, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 373, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 90, + 373, + 286, + 390 + ], + "type": "interline_equation", + "content": "\\psi_ {\\mathcal {C}, L} := \\left(\\psi_ {\\mathcal {C}, 1}, \\dots , \\psi_ {\\mathcal {C}, m}\\right): \\mathcal {M} _ {\\mathcal {C}} ^ {L} \\rightarrow \\mathcal {M} _ {\\tilde {\\mathcal {C}}} ^ {1, 1} \\tag {9}", + "image_path": "1843b6b72e49516378beed836187fb027556b46beba62bc3e78abe6125b3a2fb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 395, + 167, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 395, + 167, + 406 + ], + "spans": [ + { + "bbox": [ + 67, + 395, + 167, + 406 + ], + "type": "text", + "content": "is a linear isomorphism." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 412, + 287, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 412, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 54, + 412, + 287, + 452 + ], + "type": "text", + "content": "2. Let " + }, + { + "bbox": [ + 54, + 412, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\phi_X: \\Lambda(X) \\to \\mathbb{P}^2" + }, + { + "bbox": [ + 54, + 412, + 287, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 412, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\psi_{\\mathcal{C},i}: \\Lambda(C_iX) \\to \\mathbb{P}^1" + }, + { + "bbox": [ + 54, + 412, + 287, + 452 + ], + "type": "text", + "content": " be any choices of linear isomorphisms. Let " + }, + { + "bbox": [ + 54, + 412, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{C}}" + }, + { + "bbox": [ + 54, + 412, + 287, + 452 + ], + "type": "text", + "content": " denote the arrangement of matrices " + }, + { + "bbox": [ + 54, + 412, + 287, + 452 + ], + "type": "inline_equation", + "content": "\\widehat{C}_i \\coloneqq \\psi_{\\mathcal{C},i} \\circ C_i \\circ \\phi_X^{-1}" + }, + { + "bbox": [ + 54, + 412, + 287, + 452 + ], + "type": "text", + "content": ". Then" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 83, + 467, + 286, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 467, + 286, + 484 + ], + "spans": [ + { + "bbox": [ + 83, + 467, + 286, + 484 + ], + "type": "interline_equation", + "content": "\\psi_ {\\mathcal {C}, X} := \\left(\\psi_ {\\mathcal {C}, 1}, \\dots , \\psi_ {\\mathcal {C}, m}\\right): \\mathcal {L} _ {\\mathcal {C}} ^ {X} \\rightarrow \\mathcal {M} _ {\\widehat {\\mathcal {C}}} ^ {2, 1} \\tag {10}", + "image_path": "42634cd9aef884853ddb7e25444d61d07a4f9b74db937a9a41590a1d90e17e1e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 488, + 167, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 488, + 167, + 500 + ], + "spans": [ + { + "bbox": [ + 67, + 488, + 167, + 500 + ], + "type": "text", + "content": "is a linear isomorphism." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 505, + 286, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 286, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 286, + 571 + ], + "type": "text", + "content": "In Section 2.1.1, we exploit these linear isomorphisms to improve the speed of triangulation for our setting. A consequence of this theorem is also that many results for the anchored multiview varieties translate into results about " + }, + { + "bbox": [ + 46, + 505, + 286, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\hat{\\mathcal{C}}}^{1,1}" + }, + { + "bbox": [ + 46, + 505, + 286, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 505, + 286, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\hat{\\mathcal{C}}}^{2,1}" + }, + { + "bbox": [ + 46, + 505, + 286, + 571 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 575, + 274, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 274, + 589 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 274, + 589 + ], + "type": "text", + "content": "Proposition 1.4. " + }, + { + "bbox": [ + 47, + 575, + 274, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_C^L" + }, + { + "bbox": [ + 47, + 575, + 274, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 575, + 274, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C^X" + }, + { + "bbox": [ + 47, + 575, + 274, + 589 + ], + "type": "text", + "content": " are irreducible. Further," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 593, + 287, + 636 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 55, + 593, + 285, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 285, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 285, + 606 + ], + "type": "text", + "content": "1. " + }, + { + "bbox": [ + 55, + 593, + 285, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_C^L" + }, + { + "bbox": [ + 55, + 593, + 285, + 606 + ], + "type": "text", + "content": " is isomorphic to " + }, + { + "bbox": [ + 55, + 593, + 285, + 606 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^1" + }, + { + "bbox": [ + 55, + 593, + 285, + 606 + ], + "type": "text", + "content": ". In particular, " + }, + { + "bbox": [ + 55, + 593, + 285, + 606 + ], + "type": "inline_equation", + "content": "\\dim \\mathcal{M}_C^L = 1" + }, + { + "bbox": [ + 55, + 593, + 285, + 606 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 54, + 612, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 612, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 54, + 612, + 287, + 636 + ], + "type": "text", + "content": "2. If the span of the centers " + }, + { + "bbox": [ + 54, + 612, + 287, + 636 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 54, + 612, + 287, + 636 + ], + "type": "text", + "content": " and the point " + }, + { + "bbox": [ + 54, + 612, + 287, + 636 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 612, + 287, + 636 + ], + "type": "text", + "content": " are not collinear, then " + }, + { + "bbox": [ + 54, + 612, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\dim \\mathcal{L}_{\\mathcal{C}}^{X} = 2" + }, + { + "bbox": [ + 54, + 612, + 287, + 636 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "Using the fundamental matrices of " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "inline_equation", + "content": "F^{ij}" + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": ", we introduce a set of polynomial constraints that must be fulfilled by the correspondences of points or lines in the anchored multiview varieties. These constraints imply that some of the equations obtained from Proposition 1.2 are redundant in the generic case." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "Proposition 1.5. For a point " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{P}^3" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " and line " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^3" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " be a generic (random) camera arrangement of " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " cameras." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 103, + 545, + 149 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 313, + 103, + 545, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 103, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 103, + 545, + 129 + ], + "type": "text", + "content": "1. " + }, + { + "bbox": [ + 313, + 103, + 545, + 129 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{M}_{\\mathcal{C}}^{L}" + }, + { + "bbox": [ + 313, + 103, + 545, + 129 + ], + "type": "text", + "content": " if and only if " + }, + { + "bbox": [ + 313, + 103, + 545, + 129 + ], + "type": "inline_equation", + "content": "x_{1}^{T}F^{1j}x_{j} = 0" + }, + { + "bbox": [ + 313, + 103, + 545, + 129 + ], + "type": "text", + "content": " for every " + }, + { + "bbox": [ + 313, + 103, + 545, + 129 + ], + "type": "inline_equation", + "content": "j = 2, \\ldots, m" + }, + { + "bbox": [ + 313, + 103, + 545, + 129 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 103, + 545, + 129 + ], + "type": "inline_equation", + "content": "x_{i}^{T}C_{i} \\cdot L = 0" + }, + { + "bbox": [ + 313, + 103, + 545, + 129 + ], + "type": "text", + "content": " for every " + }, + { + "bbox": [ + 313, + 103, + 545, + 129 + ], + "type": "inline_equation", + "content": "i = 1, \\ldots, m" + }, + { + "bbox": [ + 313, + 103, + 545, + 129 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 136, + 415, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 136, + 415, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 136, + 415, + 149 + ], + "type": "text", + "content": "2. " + }, + { + "bbox": [ + 313, + 136, + 415, + 149 + ], + "type": "inline_equation", + "content": "\\ell \\in \\mathcal{L}_C^X" + }, + { + "bbox": [ + 313, + 136, + 415, + 149 + ], + "type": "text", + "content": " if and only if" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 368, + 158, + 545, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 158, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 368, + 158, + 545, + 178 + ], + "type": "interline_equation", + "content": "\\det \\left[ \\begin{array}{l l l} C _ {1} ^ {T} \\ell_ {1} & C _ {2} ^ {T} \\ell_ {2} & C _ {i} ^ {T} \\ell_ {i} \\end{array} \\right] = 0, \\tag {11}", + "image_path": "fbb8f9c9957b9c5413f4cdb109280300b0cf5e156683940829538b851ad8acad.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 370, + 174, + 501, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 174, + 501, + 189 + ], + "spans": [ + { + "bbox": [ + 370, + 174, + 501, + 189 + ], + "type": "interline_equation", + "content": "\\det \\left[ \\begin{array}{c c c} C _ {1} ^ {T} \\ell_ {1} & C _ {3} ^ {T} \\ell_ {3} & C _ {i} ^ {T} \\ell_ {i} \\end{array} \\right] = 0", + "image_path": "749a88e093f9842aa9c9b06acdd2447b75cce4f37dcd6759899ee0741bb3498b.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 323, + 198, + 545, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 198, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 323, + 198, + 545, + 222 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 323, + 198, + 545, + 222 + ], + "type": "inline_equation", + "content": "i = 3,\\ldots ,m" + }, + { + "bbox": [ + 323, + 198, + 545, + 222 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 323, + 198, + 545, + 222 + ], + "type": "inline_equation", + "content": "\\ell_i^T C_iX = 0" + }, + { + "bbox": [ + 323, + 198, + 545, + 222 + ], + "type": "text", + "content": " for every " + }, + { + "bbox": [ + 323, + 198, + 545, + 222 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,m" + }, + { + "bbox": [ + 323, + 198, + 545, + 222 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 231, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 266 + ], + "type": "text", + "content": "The determinantal constraints described in the second item correspond to the constraints that are satisfied by elements of the line multiview variety, presented in [7]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 267, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 315 + ], + "type": "text", + "content": "In Supplementary Material Section B we define the multidegree of a variety in a product of projective spaces, determine it for the anchored multiview varieties and explain its relevance for computer vision." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 323, + 481, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 481, + 335 + ], + "type": "text", + "content": "1.2. The Euclidean Distance problem" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 341, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 377 + ], + "type": "text", + "content": "For a variety " + }, + { + "bbox": [ + 304, + 341, + 545, + 377 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\subseteq \\mathbb{R}^n" + }, + { + "bbox": [ + 304, + 341, + 545, + 377 + ], + "type": "text", + "content": " and a point " + }, + { + "bbox": [ + 304, + 341, + 545, + 377 + ], + "type": "inline_equation", + "content": "u \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 304, + 341, + 545, + 377 + ], + "type": "text", + "content": " outside the variety, a natural problem is to find the closest point on " + }, + { + "bbox": [ + 304, + 341, + 545, + 377 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 341, + 545, + 377 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 341, + 545, + 377 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 341, + 545, + 377 + ], + "type": "text", + "content": ", which corresponds to the optimization problem" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 318, + 387, + 545, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 387, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 318, + 387, + 545, + 417 + ], + "type": "interline_equation", + "content": "\\text {m i n i m i z e} \\quad \\sum_ {i = 1} ^ {n} \\left(u _ {i} - x _ {i}\\right) ^ {2} \\quad \\text {s u b j e c t t o} \\quad x \\in \\mathcal {X}. \\tag {12}", + "image_path": "15e47067b2de17c4c0d4edef374095932d18f478da92852a6e7b5734d59fcd0e.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": "Equation (12) is called the Euclidean distance problem and models the process of error correction and fitting noisy data to a mathematical model " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": ". In the case of a smooth variety, defined below, the Euclidean distance degree (EDD) is the number of complex solutions to the critical equations of (12) [11]. The EDD is an estimate of how difficult it is to solve this problem by exact algebraic methods. For a variety " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": " in a product of projective spaces " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^{n_1} \\times \\dots \\times \\mathbb{P}^{n_m}" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": ", the EDD is the EDD of " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\cap U_1 \\times \\dots \\times U_m \\subseteq \\mathbb{R}^{n_1 + \\dots + n_m}" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "U_i" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": " is a generic affine patch of " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^{h_i}" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": ". An affine patch " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^n" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": " is a subset defined by an affine equation with non-zero constant part, for instance " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "x_0 = 1" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": ". We have " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "U \\cong \\mathbb{R}^n" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": " over the real numbers." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "text", + "content": "A variety " + }, + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "text", + "content": " is smooth at a point " + }, + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "text", + "content": " locally around " + }, + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "text", + "content": " looks like Euclidean space. " + }, + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 582, + 545, + 700 + ], + "type": "text", + "content": " is smooth if all its points are smooth. The singular locus of a variety is the set of non-smooth points. See [18] for more details. The singular locus of the point multiview variety is well-understood, see for instance [36], and it is mostly understood for the line multiview variety, see [7, Section 3]. The proposition below guarantees that the anchored multiview varieties are smooth for generic camera arrangements, which helps us to compute its EDD." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3751" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 116, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 116, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 116, + 83 + ], + "type": "text", + "content": "Proposition 1.6." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 91, + 286, + 136 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 91, + 121, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 121, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 121, + 105 + ], + "type": "text", + "content": "1. " + }, + { + "bbox": [ + 55, + 91, + 121, + 105 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_C^L" + }, + { + "bbox": [ + 55, + 91, + 121, + 105 + ], + "type": "text", + "content": " smooth." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 111, + 286, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 111, + 286, + 136 + ], + "spans": [ + { + "bbox": [ + 54, + 111, + 286, + 136 + ], + "type": "text", + "content": "2. If there are exactly two cameras, or the centers together with the point " + }, + { + "bbox": [ + 54, + 111, + 286, + 136 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 111, + 286, + 136 + ], + "type": "text", + "content": " span " + }, + { + "bbox": [ + 54, + 111, + 286, + 136 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^3" + }, + { + "bbox": [ + 54, + 111, + 286, + 136 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 54, + 111, + 286, + 136 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C^X" + }, + { + "bbox": [ + 54, + 111, + 286, + 136 + ], + "type": "text", + "content": " is smooth." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "text", + "content": "For the EDD to be relevant in a particular setting, the Euclidean distance has to be a good measurement of distance. This is naturally the case for points in " + }, + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "text", + "content": ", i.e. an affine patch of " + }, + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^2" + }, + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "text", + "content": ". This extends to an affine patch of " + }, + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "text", + "content": ". It is perhaps less clear how to best measure distances between lines. Recall that we regard " + }, + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "text", + "content": " as a subset of " + }, + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "inline_equation", + "content": "(\\mathbb{P}^2)^m" + }, + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "text", + "content": ". Indeed, we identify image lines with the point that defines its normal vector. We choose the Euclidean distance between these normal vectors (in affine patches of " + }, + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^2" + }, + { + "bbox": [ + 46, + 143, + 286, + 262 + ], + "type": "text", + "content": ") as our distance between lines." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 262, + 286, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 286, + 334 + ], + "type": "text", + "content": "The theorem below presents our computations of the EDDs of the anchored multiview varieties. Additionally, our numerical computations using the HomotopyContinuation.jl [8] package in the julia [5] programming language have confirmed the accuracy of these formulas for " + }, + { + "bbox": [ + 47, + 262, + 286, + 334 + ], + "type": "inline_equation", + "content": "m \\leq 10" + }, + { + "bbox": [ + 47, + 262, + 286, + 334 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 341, + 286, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 286, + 364 + ], + "type": "text", + "content": "Theorem 1.7. Let " + }, + { + "bbox": [ + 47, + 341, + 286, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 47, + 341, + 286, + 364 + ], + "type": "text", + "content": " be a generic arrangement of " + }, + { + "bbox": [ + 47, + 341, + 286, + 364 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 341, + 286, + 364 + ], + "type": "text", + "content": " cameras." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 370, + 258, + 403 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 370, + 164, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 164, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 164, + 384 + ], + "type": "text", + "content": "1. " + }, + { + "bbox": [ + 55, + 370, + 164, + 384 + ], + "type": "inline_equation", + "content": "\\operatorname{EDD}(\\mathcal{M}_C^L) = 3m - 2" + }, + { + "bbox": [ + 55, + 370, + 164, + 384 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 389, + 258, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 389, + 258, + 403 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 258, + 403 + ], + "type": "text", + "content": "2. If " + }, + { + "bbox": [ + 54, + 389, + 258, + 403 + ], + "type": "inline_equation", + "content": "m \\geq 3" + }, + { + "bbox": [ + 54, + 389, + 258, + 403 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 54, + 389, + 258, + 403 + ], + "type": "inline_equation", + "content": "\\mathrm{EDD}(\\mathcal{L}_{\\mathcal{C}}^{X}) = \\frac{9}{2} m^{2} - \\frac{19}{2} m + 3" + }, + { + "bbox": [ + 54, + 389, + 258, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 408, + 286, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 286, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 286, + 448 + ], + "type": "text", + "content": "We end with an implication of Theorem 1.3: There is a direct correspondence between the EDDs of the anchored multiview varieties and " + }, + { + "bbox": [ + 47, + 408, + 286, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\widetilde{\\mathcal{C}}}^{1,1}, \\mathcal{M}_{\\widehat{\\mathcal{C}}}^{2,1}" + }, + { + "bbox": [ + 47, + 408, + 286, + 448 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 454, + 286, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 286, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 286, + 478 + ], + "type": "text", + "content": "Theorem 1.8. Let " + }, + { + "bbox": [ + 47, + 454, + 286, + 478 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 47, + 454, + 286, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 454, + 286, + 478 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 47, + 454, + 286, + 478 + ], + "type": "text", + "content": " be generic arrangements of cardinality " + }, + { + "bbox": [ + 47, + 454, + 286, + 478 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 454, + 286, + 478 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 483, + 266, + 522 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 55, + 483, + 169, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 169, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 169, + 499 + ], + "type": "text", + "content": "1. " + }, + { + "bbox": [ + 55, + 483, + 169, + 499 + ], + "type": "inline_equation", + "content": "\\operatorname{EDD}(\\mathcal{M}_{\\widetilde{c}}^{1,1}) = 3m - 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 506, + 266, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 506, + 266, + 522 + ], + "spans": [ + { + "bbox": [ + 54, + 506, + 266, + 522 + ], + "type": "text", + "content": "2. If " + }, + { + "bbox": [ + 54, + 506, + 266, + 522 + ], + "type": "inline_equation", + "content": "m \\geq 3" + }, + { + "bbox": [ + 54, + 506, + 266, + 522 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 54, + 506, + 266, + 522 + ], + "type": "inline_equation", + "content": "\\mathrm{EDD}(\\mathcal{M}_{\\hat{\\mathcal{C}}}^{2,1}) = \\frac{9}{2} m^2 - \\frac{19}{2} m + 3" + }, + { + "bbox": [ + 54, + 506, + 266, + 522 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 529, + 184, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 184, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 184, + 544 + ], + "type": "text", + "content": "2. Numerical Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 550, + 286, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 286, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 286, + 585 + ], + "type": "text", + "content": "In this section we conduct numerical experiments of the triangulation process of incident point and lines. All the code can be found in the Supplementary Material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 586, + 286, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 286, + 669 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 286, + 669 + ], + "type": "text", + "content": "By (L1) we refer to point-line arrangements in " + }, + { + "bbox": [ + 46, + 586, + 286, + 669 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^3" + }, + { + "bbox": [ + 46, + 586, + 286, + 669 + ], + "type": "text", + "content": " of one line incident to " + }, + { + "bbox": [ + 46, + 586, + 286, + 669 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 586, + 286, + 669 + ], + "type": "text", + "content": " points, as depicted in Figure 1. Our goal is to reconstruct such arrangements, given point correspondences incident to a single line correspondence across " + }, + { + "bbox": [ + 46, + 586, + 286, + 669 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 586, + 286, + 669 + ], + "type": "text", + "content": " views. The following is a list of natural approaches for such triangulation given a camera arrangement " + }, + { + "bbox": [ + 46, + 586, + 286, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 586, + 286, + 669 + ], + "type": "text", + "content": ". It is not necessarily a complete list." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 52, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 52, + 677, + 286, + 713 + ], + "type": "text", + "content": "(L1).0 Triangulate each point correspondence by fitting it to the point multiview variety " + }, + { + "bbox": [ + 52, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C}}" + }, + { + "bbox": [ + 52, + 677, + 286, + 713 + ], + "type": "text", + "content": ", i.e. find the closest point correspondence in " + }, + { + "bbox": [ + 52, + 677, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C}}" + }, + { + "bbox": [ + 52, + 677, + 286, + 713 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 324, + 73, + 425, + 161 + ], + "blocks": [ + { + "bbox": [ + 324, + 73, + 425, + 161 + ], + "lines": [ + { + "bbox": [ + 324, + 73, + 425, + 161 + ], + "spans": [ + { + "bbox": [ + 324, + 73, + 425, + 161 + ], + "type": "image", + "image_path": "348ea6aca160ede2bc865b7669ebe94dc2efe1b17bfe45052a37721cf20985a9.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 173, + 545, + 202 + ], + "lines": [ + { + "bbox": [ + 305, + 173, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 305, + 173, + 545, + 202 + ], + "type": "text", + "content": "Figure 2: Comparison of (L1).0 and (L1).1 triangulation approaches. (L1).0 expects non-collinear points in the reconstruction, while (L1).1 produces collinear points." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 425, + 75, + 528, + 161 + ], + "blocks": [ + { + "bbox": [ + 425, + 75, + 528, + 161 + ], + "lines": [ + { + "bbox": [ + 425, + 75, + 528, + 161 + ], + "spans": [ + { + "bbox": [ + 425, + 75, + 528, + 161 + ], + "type": "image", + "image_path": "cda082f93cda9cec7bd913c67990e93c00e05227df19c28c051063128f0822b8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 310, + 223, + 545, + 452 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 310, + 223, + 545, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 223, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 310, + 223, + 545, + 271 + ], + "type": "text", + "content": "(L1).1 Reconstruct the 3D line " + }, + { + "bbox": [ + 310, + 223, + 545, + 271 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 310, + 223, + 545, + 271 + ], + "type": "text", + "content": " by back-projecting the image lines from two views. Triangulate the point correspondences by fitting them to the anchored multiview variety " + }, + { + "bbox": [ + 310, + 223, + 545, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C}}^{L}" + }, + { + "bbox": [ + 310, + 223, + 545, + 271 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "text", + "content": "(L1).2 Triangulate two point correspondences by fitting them to " + }, + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_C" + }, + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "text", + "content": " to get 3D points " + }, + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "text", + "content": " be the line they span in " + }, + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "text", + "content": ". Triangulate the remaining point correspondences by fitting them to " + }, + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_C^L" + }, + { + "bbox": [ + 310, + 279, + 545, + 327 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 310, + 335, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 335, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 310, + 335, + 545, + 396 + ], + "type": "text", + "content": "(L1).3 Triangulate one point correspondence by fitting it to " + }, + { + "bbox": [ + 310, + 335, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C}}" + }, + { + "bbox": [ + 310, + 335, + 545, + 396 + ], + "type": "text", + "content": " to get a 3D point " + }, + { + "bbox": [ + 310, + 335, + 545, + 396 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 310, + 335, + 545, + 396 + ], + "type": "text", + "content": ". Reconstruct the 3D line " + }, + { + "bbox": [ + 310, + 335, + 545, + 396 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 310, + 335, + 545, + 396 + ], + "type": "text", + "content": " by fitting the line correspondence to the anchored variety " + }, + { + "bbox": [ + 310, + 335, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{C}}^{X}" + }, + { + "bbox": [ + 310, + 335, + 545, + 396 + ], + "type": "text", + "content": ". Triangulate the remaining point correspondences by fitting them to " + }, + { + "bbox": [ + 310, + 335, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C}}^{L}" + }, + { + "bbox": [ + 310, + 335, + 545, + 396 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 310, + 403, + 545, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 403, + 545, + 452 + ], + "spans": [ + { + "bbox": [ + 310, + 403, + 545, + 452 + ], + "type": "text", + "content": "(L1).4 Reconstruct the 3D line " + }, + { + "bbox": [ + 310, + 403, + 545, + 452 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 310, + 403, + 545, + 452 + ], + "type": "text", + "content": " by fitting the line correspondence to the line multiview variety " + }, + { + "bbox": [ + 310, + 403, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{C}}" + }, + { + "bbox": [ + 310, + 403, + 545, + 452 + ], + "type": "text", + "content": ". Triangulate the point correspondences by fitting them to the anchored multiview variety " + }, + { + "bbox": [ + 310, + 403, + 545, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C}}^{L}" + }, + { + "bbox": [ + 310, + 403, + 545, + 452 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 461, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 461, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 305, + 461, + 545, + 520 + ], + "type": "text", + "content": "Approaches (L1).1-4 take the incidence relations into account in the triangulation, whereas (L1).0 does not. Therefore, in contrast to (L1).1-4, the resulting triangulation of (L1).0 does not preserve the point-line incidences; see Figure 2." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 529, + 403, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 403, + 541 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 403, + 541 + ], + "type": "text", + "content": "2.1. Implementation" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 548, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 548, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 305, + 548, + 545, + 617 + ], + "type": "text", + "content": "Our experiments are implemented in HomotopyContinuation.jl [8] in Julia [5]. Here we explain them in some detail. The full explanation can be found in Supplementary Material. We ran the code on a Intel(R) Core(TM) i5-8300H CPU running at 2.30GHz." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 634, + 493, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 493, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 493, + 645 + ], + "type": "text", + "content": "2.1.1 Reducing the number of parameters" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 652, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 652, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 652, + 545, + 713 + ], + "type": "text", + "content": "We reduce the number of parameters involved in the problem of fitting data to " + }, + { + "bbox": [ + 305, + 652, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C}}^{L}" + }, + { + "bbox": [ + 305, + 652, + 545, + 713 + ], + "type": "text", + "content": ", respectively " + }, + { + "bbox": [ + 305, + 652, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{C}}^{X}" + }, + { + "bbox": [ + 305, + 652, + 545, + 713 + ], + "type": "text", + "content": ", by translating it to and solving it for " + }, + { + "bbox": [ + 305, + 652, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\widetilde{\\mathcal{C}}}^{1,1}" + }, + { + "bbox": [ + 305, + 652, + 545, + 713 + ], + "type": "text", + "content": ", respectively " + }, + { + "bbox": [ + 305, + 652, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\widehat{\\mathcal{C}}}^{2,1}" + }, + { + "bbox": [ + 305, + 652, + 545, + 713 + ], + "type": "text", + "content": ". Details are found in Supplementary Material Section B. This translation corresponds to a reduction of parameters because an" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3752" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "text", + "content": "affine patch of " + }, + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\widehat{\\mathcal{C}}}^{1,1}" + }, + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "text", + "content": " lives in " + }, + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^1)^m" + }, + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "text", + "content": ", while an affine patch of " + }, + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C}}^{L}" + }, + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "text", + "content": " lives in " + }, + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^2)^m" + }, + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "text", + "content": ". The analogous is true for " + }, + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\widehat{\\mathcal{C}}}^{2,1}" + }, + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{C}}^{X}" + }, + { + "bbox": [ + 46, + 71, + 286, + 114 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 114, + 286, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 114, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 46, + 114, + 286, + 161 + ], + "type": "text", + "content": "We call the methods that do not use this translation standard and for instance, write (L1).1 std. to denote it. By just writing (L1).1, we denote the implementation of this translation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 179, + 151, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 179, + 151, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 179, + 151, + 192 + ], + "type": "text", + "content": "2.1.2 Evaluating error" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "text", + "content": "In our experiments, each iteration starts by randomly generating a 3D line " + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "text", + "content": " on this line. This line and these points are projected by randomly generated cameras to line and point correspondences in " + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^2)^m" + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "text", + "content": ". For fixed " + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\epsilon = 10^{-12}" + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "text", + "content": ", randomly generated noise vectors " + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\sigma (\\epsilon)" + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "text", + "content": " are added in each factor. Our approaches then find points " + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "inline_equation", + "content": "Y_{i}\\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 46, + 199, + 287, + 307 + ], + "type": "text", + "content": " that match the noisy correspondences. We measure the accuracy by taking the logarithm after averaging the relative error of reconstructed points:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 316, + 287, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 316, + 287, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 287, + 343 + ], + "type": "interline_equation", + "content": "e = \\log_ {1 0} \\left(\\frac {\\sum \\| Y _ {i} - X _ {i} \\|}{p \\epsilon}\\right), \\tag {13}", + "image_path": "f6a20009e2baebeb694f37a523e606f6cac94a124122855dca9ed8b6268660ab.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 353, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 353, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\| Y_{i} - X_{i}\\|" + }, + { + "bbox": [ + 46, + 353, + 287, + 388 + ], + "type": "text", + "content": " denotes the Euclidean distance. The interpretation of this number is that the error in the data gets amplified by " + }, + { + "bbox": [ + 46, + 353, + 287, + 388 + ], + "type": "inline_equation", + "content": "10^{e}" + }, + { + "bbox": [ + 46, + 353, + 287, + 388 + ], + "type": "text", + "content": " during the triangulation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 406, + 128, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 128, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 128, + 418 + ], + "type": "text", + "content": "2.1.3 Algorithms" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "text", + "content": "The pseudocodes for all approaches are included in the Supplementary Material. The pseudocode for approach (L1).1 is presented in Algorithm 1. We use the notation that for a column vector " + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "inline_equation", + "content": "[X; 1] \\in \\mathbb{R}^{n+1}" + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "text", + "content": " is the vector we get by adding a 1 as the last coordinate. Let " + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\iota" + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "text", + "content": " be the function that scales a vector such that its last coordinate is 1, and then removes that coordinate. In lines 1 and 2 of the algorithm, we generate noisy data by introducing randomly generated noise " + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\sigma(\\epsilon)" + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "text", + "content": " as explained above, where " + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\epsilon = 10^{-12}" + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "text", + "content": " is a fixed parameter for our experiments. Lines 3 and 4 correspond to the triangulation of the line and the point correspondences using the anchored point multiview variety at the line " + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "inline_equation", + "content": "L_0" + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "text", + "content": ". Note that we solve the closest point problem in line 4 by computing the zeros of a system of polynomial (critical) equations using the solve function in HomotopyContinuation.jl [9]. Finally, line 5 compares the points obtained in the previous step with the original starting points, measuring the logarithmic average relative error of the " + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 426, + 287, + 653 + ], + "type": "text", + "content": " points." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Most other pseudocodes are similar to Algorithm 1. For instance, approach (L1).0 does not take into account the line, so it corresponds deleting lines 1 and 3 from Algorithm 1 and modifying the minimization domain in line 4, replacing " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{0}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\mathcal{C}}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": ". In approach (L1).2, line 3 is modified" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "so that " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "L_{0}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " is obtained by the span of two triangulated points instead of by intersecting two back-projected planes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "spans": [ + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "text", + "content": "Algorithm 1: One iteration of the (L1).1 std. method given a randomly generated camera arrangement " + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "inline_equation", + "content": "3\\times 4" + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "text", + "content": " matrices, a projective line " + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "text", + "content": " spanned by two vectors of " + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^4" + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "inline_equation", + "content": "X_{i}\\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "inline_equation", + "content": "[X_i;1]" + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "text", + "content": " lie on " + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 310, + 111, + 519, + 172 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 174, + 498, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 174, + 498, + 187 + ], + "spans": [ + { + "bbox": [ + 318, + 174, + 498, + 187 + ], + "type": "text", + "content": "Input : " + }, + { + "bbox": [ + 318, + 174, + 498, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = (C_1, \\dots, C_m)" + }, + { + "bbox": [ + 318, + 174, + 498, + 187 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 318, + 174, + 498, + 187 + ], + "type": "inline_equation", + "content": "L, X_1, \\dots, X_p" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 187, + 501, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 187, + 501, + 198 + ], + "spans": [ + { + "bbox": [ + 319, + 187, + 501, + 198 + ], + "type": "text", + "content": "Output: The log of the average relative error" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 199, + 405, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 199, + 405, + 210 + ], + "spans": [ + { + "bbox": [ + 311, + 199, + 405, + 210 + ], + "type": "text", + "content": "1 for " + }, + { + "bbox": [ + 311, + 199, + 405, + 210 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 311, + 199, + 405, + 210 + ], + "type": "text", + "content": " from 1 to " + }, + { + "bbox": [ + 311, + 199, + 405, + 210 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 311, + 199, + 405, + 210 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 210, + 465, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 210, + 465, + 251 + ], + "spans": [ + { + "bbox": [ + 311, + 210, + 465, + 251 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\text {\\bf 2} \\\\ \\text {\\bf 3} \\\\ \\text {\\bf 4} \\end{array} \\left\\lfloor \\begin{array}{l} \\text {\\bf f o r} i \\text {f r o m 1 t o p d o} \\\\ \\text {\\bf \\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\hbox {\\c h i}}}}}}}}}}}}}}}}}}}}}} q _ {i, j} \\leftarrow \\iota (C _ {j} [ X _ {i}; 1 ]) + \\sigma (\\epsilon); \\\\ u _ {j} \\leftarrow \\iota (C _ {j} \\cdot L) + \\sigma (\\epsilon); \\end{array} \\right.", + "image_path": "069e24d1671a025f09260dd39be33b1fdfc0bb32c77805a52776c056983ba543.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 253, + 500, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 253, + 500, + 269 + ], + "spans": [ + { + "bbox": [ + 310, + 253, + 500, + 269 + ], + "type": "text", + "content": "5 " + }, + { + "bbox": [ + 310, + 253, + 500, + 269 + ], + "type": "inline_equation", + "content": "L_0\\gets" + }, + { + "bbox": [ + 310, + 253, + 500, + 269 + ], + "type": "text", + "content": " nullspace " + }, + { + "bbox": [ + 310, + 253, + 500, + 269 + ], + "type": "inline_equation", + "content": "\\left[C_1^T [u_1;1]\\quad C_2^T [u_2;1]\\right]^T;" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 270, + 399, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 270, + 399, + 280 + ], + "spans": [ + { + "bbox": [ + 311, + 270, + 399, + 280 + ], + "type": "text", + "content": "6 for " + }, + { + "bbox": [ + 311, + 270, + 399, + 280 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 311, + 270, + 399, + 280 + ], + "type": "text", + "content": " from 1 to " + }, + { + "bbox": [ + 311, + 270, + 399, + 280 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 311, + 270, + 399, + 280 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 280, + 516, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 280, + 516, + 313 + ], + "spans": [ + { + "bbox": [ + 310, + 280, + 516, + 313 + ], + "type": "interline_equation", + "content": "\\begin{array}{c}Y_{i}\\leftarrow \\\\ \\underset {X\\in \\mathbb{R}^{3}:[X;1]\\in L_{0}}{\\operatorname *{argmin}}\\sum_{j = 1}^{m}(q_{i,j} - \\iota (C_{j}[X;1]))^{2}; \\end{array}", + "image_path": "29d74e13ff3d63d06ba8c01c8c61e5d6e31d15289ea54f605bb2117e27ac94b6.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 310, + 316, + 459, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 316, + 459, + 335 + ], + "spans": [ + { + "bbox": [ + 310, + 316, + 459, + 335 + ], + "type": "text", + "content": "8 " + }, + { + "bbox": [ + 310, + 316, + 459, + 335 + ], + "type": "inline_equation", + "content": "e\\gets \\log_{10}\\left(\\frac{1}{p\\epsilon}\\sum_{i = 1}^{p}\\| Y_i - X_i\\|\\right)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 319, + 335, + 362, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 335, + 362, + 343 + ], + "spans": [ + { + "bbox": [ + 319, + 335, + 362, + 343 + ], + "type": "text", + "content": "Return: " + }, + { + "bbox": [ + 319, + 335, + 362, + 343 + ], + "type": "inline_equation", + "content": "e" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 370, + 411, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 411, + 381 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 411, + 381 + ], + "type": "text", + "content": "2.2. Numerical results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": "The main results of our numerical experiments are presented in Figures 3 to 5 and Tables 1 to 3. We compare the performance of the five different triangulation approaches for " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "p = 5" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": " and different number of " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": " cameras. In the tables, we present the median, mean and standard deviation " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": " of the logarithmic average relative error and time. The results are displayed in histograms and tables created with Plots.jl [10]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 485, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 579 + ], + "type": "text", + "content": "Note that for two cameras, (L1).0 and (L1).2 are the exact same, and (L1).1 and (L1).4 are essentially the same. In Figure 3 and Table 1 we present results for " + }, + { + "bbox": [ + 304, + 485, + 545, + 579 + ], + "type": "inline_equation", + "content": "m = 2" + }, + { + "bbox": [ + 304, + 485, + 545, + 579 + ], + "type": "text", + "content": " and therefore only include (L1).0, (L1).1 and (L1).3. We also compare with the standard implementations of (L1).1 and (L1).3 that do not use the linear isomorphisms of Theorem 1.3 to improve computation time as described in Section 2.1.1. This simulation is iterated 1000 times." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "type": "text", + "content": "Figure 4 and Table 2 compare the different reconstruction approaches for " + }, + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "type": "inline_equation", + "content": "m = 3" + }, + { + "bbox": [ + 304, + 581, + 545, + 616 + ], + "type": "text", + "content": " cameras. The simulations are again iterated 1000 times." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "text", + "content": "Figure 5 and Table 3 compare the different reconstruction approaches for " + }, + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "inline_equation", + "content": "m = 4" + }, + { + "bbox": [ + 304, + 616, + 545, + 664 + ], + "type": "text", + "content": " cameras. The simulations are iterated only 100 times, due to the longer run-time of the simulations." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Our theoretical studies in Section 1, specifically Theorem 1.7, guarantee that methods such as (L1).1, (L1).2 and (L1).3 are less complex from the algebraic point of view, than (L1).0 for triangulation of points incident to a line. Ad-" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3753" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 79, + 252, + 186 + ], + "blocks": [ + { + "bbox": [ + 100, + 70, + 244, + 79 + ], + "lines": [ + { + "bbox": [ + 100, + 70, + 244, + 79 + ], + "spans": [ + { + "bbox": [ + 100, + 70, + 244, + 79 + ], + "type": "text", + "content": "Accuracy: 2 cameras, 5 point correspondences" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 79, + 252, + 186 + ], + "lines": [ + { + "bbox": [ + 80, + 79, + 252, + 186 + ], + "spans": [ + { + "bbox": [ + 80, + 79, + 252, + 186 + ], + "type": "image", + "image_path": "0175f04a2e365855a322993fdc8857b162070b3e7ac30ab52c9620d79dd58e21.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 186, + 241, + 195 + ], + "lines": [ + { + "bbox": [ + 107, + 186, + 241, + 195 + ], + "spans": [ + { + "bbox": [ + 107, + 186, + 241, + 195 + ], + "type": "text", + "content": "Speed: 2 cameras, 5 point correspondences" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 80, + 196, + 252, + 303 + ], + "blocks": [ + { + "bbox": [ + 80, + 196, + 252, + 303 + ], + "lines": [ + { + "bbox": [ + 80, + 196, + 252, + 303 + ], + "spans": [ + { + "bbox": [ + 80, + 196, + 252, + 303 + ], + "type": "image", + "image_path": "22e3360de14aec2cce14ca0de0569b1e13fc38f5c61b988d4374fdfaeac40563.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 312, + 288, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 352 + ], + "type": "text", + "content": "Figure 3: Triangulation of " + }, + { + "bbox": [ + 46, + 312, + 288, + 352 + ], + "type": "inline_equation", + "content": "p = 5" + }, + { + "bbox": [ + 46, + 312, + 288, + 352 + ], + "type": "text", + "content": " point correspondences incident to a line for " + }, + { + "bbox": [ + 46, + 312, + 288, + 352 + ], + "type": "inline_equation", + "content": "m = 2" + }, + { + "bbox": [ + 46, + 312, + 288, + 352 + ], + "type": "text", + "content": " views with complete visibility. 1000 iterations. The histograms show the frequency of the average relative error and the running time of the depicted triangulation methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 382, + 286, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 286, + 430 + ], + "type": "text", + "content": "ditionally, using Section 2.1.1, we were able to improve the computation speed while retaining comparable accuracy, which is exemplified by the two cases (L1).1 vs. (L1).1 std. and (L1).3 vs. (L1).3 std. in Table 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "text", + "content": "Finally, our numerical simulations show that in our HomotopyContinuation.jl implementation, the (L1).1 approach is the fastest across all number of views, but least accurate. In the case of " + }, + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "inline_equation", + "content": "m = 3" + }, + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "text", + "content": ", (L1).4 is the most accurate and faster than the traditional (L1).0 method. Even for " + }, + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "inline_equation", + "content": "m = 4" + }, + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "text", + "content": ", (L1).4 is the most accurate but is notably slower than (L1).0. Increasing the number of cameras " + }, + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "text", + "content": " improves accuracy and reduces speed, but the exact impact depends on the approach. Among (L1).1-4, (L1).1 is least affected in both regards, and (L1).4 is most affected in both regards. The speed can be thought of as an affine function in the number of point correspondences " + }, + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "text", + "content": ". Reconstructing the line correspondence takes some fixed amount of time independent of " + }, + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "text", + "content": " and then all " + }, + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 430, + 287, + 609 + ], + "type": "text", + "content": " correspondences are independently reconstructed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 621, + 212, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 212, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 212, + 634 + ], + "type": "text", + "content": "3. Conclusion and Future Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "This work studied the triangulation of incident points and lines. We introduced the anchored line multiview varieties and showed that they correspond to multiview varieties arising from projections " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^2\\longrightarrow \\mathbb{P}^1" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^1\\longrightarrow \\mathbb{P}^1" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": ". In addition, we proved that they are less complex for reconstruction via critical points. These theoretical results are also" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 337, + 70, + 515, + 147 + ], + "blocks": [ + { + "bbox": [ + 337, + 70, + 515, + 147 + ], + "lines": [ + { + "bbox": [ + 337, + 70, + 515, + 147 + ], + "spans": [ + { + "bbox": [ + 337, + 70, + 515, + 147 + ], + "type": "table", + "html": "
Accuracymedianmeanσ
(L1).01.0311.2431.213
(L1).11.4891.7021.085
(L1).1 std.1.3741.5480.990
(L1).31.2501.4490.964
(L1).3 std.1.1471.3931.205
", + "image_path": "5d3eeaf745c4f1f65d24cd75a99ba01ed53637bff1baf89464b47fcfddfef9a8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 337, + 148, + 515, + 225 + ], + "blocks": [ + { + "bbox": [ + 337, + 148, + 515, + 225 + ], + "lines": [ + { + "bbox": [ + 337, + 148, + 515, + 225 + ], + "spans": [ + { + "bbox": [ + 337, + 148, + 515, + 225 + ], + "type": "table", + "html": "
Speedmedianmeanσ
(L1).00.1700.2010.220
(L1).10.04300.04700.0207
(L1).1 std.0.1220.1410.202
(L1).30.08380.09410.0464
(L1).3 std.0.1670.1880.0764
", + "image_path": "4908226caef855744712075ca15cd2a24e36369a2898b0d7d765fedd1062f072.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 243, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 306, + 243, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 547, + 281 + ], + "type": "text", + "content": "Table 1: Triangulation of " + }, + { + "bbox": [ + 306, + 243, + 547, + 281 + ], + "type": "inline_equation", + "content": "p = 5" + }, + { + "bbox": [ + 306, + 243, + 547, + 281 + ], + "type": "text", + "content": " point correspondences incident to a line for " + }, + { + "bbox": [ + 306, + 243, + 547, + 281 + ], + "type": "inline_equation", + "content": "m = 2" + }, + { + "bbox": [ + 306, + 243, + 547, + 281 + ], + "type": "text", + "content": " views with complete visibility. 1000 iterations. The tables show the accuracy and speed of the depicted triangulation methods in terms of median, mean and standard deviation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 339, + 293, + 512, + 409 + ], + "blocks": [ + { + "bbox": [ + 339, + 293, + 512, + 409 + ], + "lines": [ + { + "bbox": [ + 339, + 293, + 512, + 409 + ], + "spans": [ + { + "bbox": [ + 339, + 293, + 512, + 409 + ], + "type": "image", + "image_path": "d7844462ac936ab3829315c0261a2a06e2d7b0ed717a40c4d5da5c6411dee73e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 339, + 418, + 511, + 525 + ], + "blocks": [ + { + "bbox": [ + 366, + 410, + 500, + 418 + ], + "lines": [ + { + "bbox": [ + 366, + 410, + 500, + 418 + ], + "spans": [ + { + "bbox": [ + 366, + 410, + 500, + 418 + ], + "type": "text", + "content": "Speed: 3 cameras, 5 point correspondences" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 339, + 418, + 511, + 525 + ], + "lines": [ + { + "bbox": [ + 339, + 418, + 511, + 525 + ], + "spans": [ + { + "bbox": [ + 339, + 418, + 511, + 525 + ], + "type": "image", + "image_path": "b3799ef3cda454409458f2ed5ead66ac62d356c3bc81e6219458d6fd99f4a625.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 535, + 545, + 574 + ], + "lines": [ + { + "bbox": [ + 306, + 535, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 535, + 545, + 574 + ], + "type": "text", + "content": "Figure 4: Triangulation of " + }, + { + "bbox": [ + 306, + 535, + 545, + 574 + ], + "type": "inline_equation", + "content": "p = 5" + }, + { + "bbox": [ + 306, + 535, + 545, + 574 + ], + "type": "text", + "content": " point correspondences incident to a line for " + }, + { + "bbox": [ + 306, + 535, + 545, + 574 + ], + "type": "inline_equation", + "content": "m = 3" + }, + { + "bbox": [ + 306, + 535, + 545, + 574 + ], + "type": "text", + "content": " views with complete visibility. 1000 iterations. The histograms show the frequency of the average relative error and the running time of the depicted triangulation methods." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 689 + ], + "type": "text", + "content": "aligned with the numerical experiments we conducted. In particular, the proposed methods in Section 2 compare different methods for triangulating a set of points incident to a line. We highlight that according to our experiments, the use of the theoretical results in Section 1 allows for a notably faster triangulation while preserving the accuracy of the traditional method for point reconstruction." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "We hope this work motivates the use of new algebraic constraints in the triangulation of incident points and lines." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3754" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 254, + 224 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 254, + 224 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 254, + 224 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 254, + 224 + ], + "type": "table", + "html": "
Accuracymedianmeanσ
(L1).00.6200.8171.023
(L1).11.5621.7811.128
(L1).20.8781.1250.998
(L1).31.0111.2431.078
(L1).40.4070.4990.901
Speedmedianmeanσ
(L1).00.9841.0600.514
(L1).10.08060.08580.0318
(L1).20.4310.4560.155
(L1).30.3040.3310.134
(L1).40.6160.7630.500
", + "image_path": "65a0941c07334f78e5414d734115351d955cefb899082633161ae5eb2bdee88d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 80, + 293, + 253, + 409 + ], + "blocks": [ + { + "bbox": [ + 80, + 293, + 253, + 409 + ], + "lines": [ + { + "bbox": [ + 80, + 293, + 253, + 409 + ], + "spans": [ + { + "bbox": [ + 80, + 293, + 253, + 409 + ], + "type": "image", + "image_path": "416a17a83418e470fa4023c56f763904a3ecee623e65ca56a26a88ab5fcab6fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 106, + 410, + 240, + 418 + ], + "lines": [ + { + "bbox": [ + 106, + 410, + 240, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 240, + 418 + ], + "type": "text", + "content": "Speed: 4 cameras, 5 point correspondences" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 80, + 419, + 251, + 525 + ], + "blocks": [ + { + "bbox": [ + 80, + 419, + 251, + 525 + ], + "lines": [ + { + "bbox": [ + 80, + 419, + 251, + 525 + ], + "spans": [ + { + "bbox": [ + 80, + 419, + 251, + 525 + ], + "type": "image", + "image_path": "0ce8bbba55b9911187364903051bca646c121f6452cdcab3358a919daf52771d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 535, + 287, + 574 + ], + "lines": [ + { + "bbox": [ + 47, + 535, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 287, + 574 + ], + "type": "text", + "content": "Figure 5: Triangulation of " + }, + { + "bbox": [ + 47, + 535, + 287, + 574 + ], + "type": "inline_equation", + "content": "p = 5" + }, + { + "bbox": [ + 47, + 535, + 287, + 574 + ], + "type": "text", + "content": " point correspondences incident to a line for " + }, + { + "bbox": [ + 47, + 535, + 287, + 574 + ], + "type": "inline_equation", + "content": "m = 4" + }, + { + "bbox": [ + 47, + 535, + 287, + 574 + ], + "type": "text", + "content": " views with complete visibility. 100 iterations. The histograms show the frequency of the average relative error and the running time of the depicted triangulation methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 630 + ], + "type": "text", + "content": "Some questions that arose in the development of this work, and that we believe are worth studying, are included below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 639, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 639, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 58, + 639, + 287, + 675 + ], + "type": "text", + "content": "- A triangulation approach inspired by [2] consists of reconstructing the 3D line " + }, + { + "bbox": [ + 58, + 639, + 287, + 675 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 58, + 639, + 287, + 675 + ], + "type": "text", + "content": " by fitting it to " + }, + { + "bbox": [ + 58, + 639, + 287, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{C}}" + }, + { + "bbox": [ + 58, + 639, + 287, + 675 + ], + "type": "text", + "content": " solving the following optimization problem" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 72, + 685, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 685, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 72, + 685, + 287, + 710 + ], + "type": "interline_equation", + "content": "\\text {m i n i m i z e} \\quad \\sum_ {i j} d \\left(x _ {i j}, \\ell_ {i}\\right) ^ {2} \\text {s u b j e c t t o} \\quad \\ell \\in \\mathcal {L} _ {\\mathcal {C}}. \\tag {14}", + "image_path": "b7088513d7ac96e39ff875b37354ef9a45b8e27e8d0b911323fb523587effac8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 342, + 70, + 511, + 224 + ], + "blocks": [ + { + "bbox": [ + 47, + 243, + 288, + 281 + ], + "lines": [ + { + "bbox": [ + 47, + 243, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 288, + 281 + ], + "type": "text", + "content": "Table 2: Triangulation of " + }, + { + "bbox": [ + 47, + 243, + 288, + 281 + ], + "type": "inline_equation", + "content": "p = 5" + }, + { + "bbox": [ + 47, + 243, + 288, + 281 + ], + "type": "text", + "content": " point correspondences incident to a line for " + }, + { + "bbox": [ + 47, + 243, + 288, + 281 + ], + "type": "inline_equation", + "content": "m = 3" + }, + { + "bbox": [ + 47, + 243, + 288, + 281 + ], + "type": "text", + "content": " views with complete visibility. 1000 iterations. The tables show the accuracy and speed of the depicted triangulation methods in terms of median, mean and standard deviation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 70, + 511, + 224 + ], + "lines": [ + { + "bbox": [ + 342, + 70, + 511, + 224 + ], + "spans": [ + { + "bbox": [ + 342, + 70, + 511, + 224 + ], + "type": "table", + "html": "
Accuracymedianmeanσ
(L1).00.3850.5510.776
(L1).11.4371.7621.249
(L1).21.0951.0770.845
(L1).30.8171.1101.094
(L1).40.2030.4211.324
Speedmedianmeanσ
(L1).04.3654.6751.576
(L1).10.1410.1490.0431
(L1).21.8091.8630.323
(L1).31.1071.1570.234
(L1).46.5997.3913.068
", + "image_path": "f8fe7031949babb374f6c37d5097beacf08aef7df78d76dd17c816b360d720d4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 243, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 306, + 243, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 545, + 281 + ], + "type": "text", + "content": "Table 3: Triangulation of " + }, + { + "bbox": [ + 306, + 243, + 545, + 281 + ], + "type": "inline_equation", + "content": "p = 5" + }, + { + "bbox": [ + 306, + 243, + 545, + 281 + ], + "type": "text", + "content": " point correspondences incident to a line for " + }, + { + "bbox": [ + 306, + 243, + 545, + 281 + ], + "type": "inline_equation", + "content": "m = 4" + }, + { + "bbox": [ + 306, + 243, + 545, + 281 + ], + "type": "text", + "content": " views with complete visibility. 100 iterations. The tables show the accuracy and speed of the depicted triangulation methods in terms of median, mean and standard deviation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 324, + 300, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 300, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 324, + 300, + 545, + 348 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 324, + 300, + 545, + 348 + ], + "type": "inline_equation", + "content": "x_{ij}" + }, + { + "bbox": [ + 324, + 300, + 545, + 348 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 324, + 300, + 545, + 348 + ], + "type": "inline_equation", + "content": "j = 1,\\dots ,p" + }, + { + "bbox": [ + 324, + 300, + 545, + 348 + ], + "type": "text", + "content": " point correspondences across " + }, + { + "bbox": [ + 324, + 300, + 545, + 348 + ], + "type": "inline_equation", + "content": "i = 1,\\ldots ,m" + }, + { + "bbox": [ + 324, + 300, + 545, + 348 + ], + "type": "text", + "content": " cameras, and " + }, + { + "bbox": [ + 324, + 300, + 545, + 348 + ], + "type": "inline_equation", + "content": "d(x_{ij},\\ell_i)" + }, + { + "bbox": [ + 324, + 300, + 545, + 348 + ], + "type": "text", + "content": " is the affine distance from a line in " + }, + { + "bbox": [ + 324, + 300, + 545, + 348 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 324, + 300, + 545, + 348 + ], + "type": "text", + "content": " to the point, meaning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 354, + 545, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 354, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 335, + 354, + 545, + 383 + ], + "type": "interline_equation", + "content": "d \\left(x _ {i j}, \\ell_ {i}\\right) = \\frac {\\left| 1 + \\left(\\ell_ {i}\\right) _ {1} \\left(x _ {i j}\\right) _ {1} + \\left(\\ell_ {i}\\right) _ {2} \\left(x _ {i j}\\right) _ {2} \\right|}{\\sqrt {\\left(\\ell_ {i}\\right) _ {1} ^ {2} + \\left(\\ell_ {i}\\right) _ {2} ^ {2}}}. \\tag {15}", + "image_path": "3839eae56d45f72b6005583daa32312315dc07381766f3782a8da20b3463e13d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 388, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 388, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 324, + 388, + 545, + 422 + ], + "type": "text", + "content": "We believe that this method is more accurate than (L1).4, but with longer running time, due to the appearance of fractions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 430, + 545, + 580 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 317, + 430, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 430, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 430, + 545, + 491 + ], + "type": "text", + "content": "- Consider the point-line problem (P1) defined by one point incident to " + }, + { + "bbox": [ + 317, + 430, + 545, + 491 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 317, + 430, + 545, + 491 + ], + "type": "text", + "content": " lines, illustrated in Figure 1. Can similar approaches to (L1).1-4 be formulated for the triangulation of the (P1) setting? Can they be used to improve accuracy and/or the speed of triangulation?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 497, + 545, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 497, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 497, + 545, + 580 + ], + "type": "text", + "content": "- Our implementation for the numerical experiments does not directly correspond to the specialized software and hardware used for triangulation in practice. However, our results motivate that patterns displayed in our figures and tables could translate to such settings. This could improve the efficiency of current specialized solvers." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 594, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgements. The authors thank Kathlen Kohn for helpful discussions, Lukas Gustafsson for the initial insight of Theorem 1.3, and Viktor Larsson for pointing out relevant references at the start of this project. Felix Rydell and Angelica Torres were supported by the Knut and Alice Wallenberg Foundation within their WASP (Wallenberg AI, Autonomous Systems and Software Program) AI/Math initiative. Elima Shehu is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation), Projektnummer 445466444. Angelica Torres is currently supported by the Spanish State Research Agency, through" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3755" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 73, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 73, + 287, + 95 + ], + "type": "text", + "content": "the Severo Ochoa and María de Maeztu Program for Centers and Units of Excellence in R&D." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 110, + 105, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 110, + 105, + 121 + ], + "spans": [ + { + "bbox": [ + 48, + 110, + 105, + 121 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 130, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 130, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 53, + 130, + 287, + 163 + ], + "type": "text", + "content": "[1] Sameer Agarwal, Andrew Pryhuber, and Rekha R Thomas. Ideals of the multiview variety. IEEE transactions on pattern analysis and machine intelligence, 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 165, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 165, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 53, + 165, + 287, + 208 + ], + "type": "text", + "content": "[2] Adrien Bartoli and Peter Sturm. Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer Vision and Image Understanding, 100(3):416-441, 2005. 3, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 210, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 210, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 53, + 210, + 287, + 253 + ], + "type": "text", + "content": "[3] Paul A Beardsley, Andrew Zisserman, and David William Murray. Navigation using affine structure from motion. Lecture Notes in Computer Science, 801:85-96, 1994. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 256, + 287, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 256, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 287, + 299 + ], + "type": "text", + "content": "[4] Paul A Beardsley, Andrew Zisserman, and David William Murray. Sequential updating of projective and affine structure from motion. International journal of computer vision, 23:235-259, 1997. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 335 + ], + "type": "text", + "content": "[5] Jeff Bezanson, Stefan Karpinski, Viral B Shah, and Alan Edelman. Julia: A fast dynamic language for technical computing. arXiv preprint arXiv:1209.5145, 2012. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 337, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 287, + 369 + ], + "type": "text", + "content": "[6] Paul Breiding, Timothy Duff, Lukas Gustafsson, Felix Rydell, and Elima Shehu. Line multiview ideals. arXiv preprint arXiv:2303.02066, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 371, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 371, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 53, + 371, + 287, + 404 + ], + "type": "text", + "content": "[7] Paul Breiding, Felix Rydell, Elima Shehu, and Angélica Torres. Line multiview varieties. SIAM Journal on Applied Algebra and Geometry, 7(2):470-504, 2023. 1, 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 406, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 406, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 53, + 406, + 287, + 449 + ], + "type": "text", + "content": "[8] Paul Breiding and Sascha Timme. Homotopycontinuation. jl: A package for homotopy continuation in julia. In International Congress on Mathematical Software, pages 458-465.Springer,2018.5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 452, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 452, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 53, + 452, + 287, + 495 + ], + "type": "text", + "content": "[9] Paul Breiding and Sascha Timme. HomotopyContinuation.jl: A Package for Homotopy Continuation in Julia. In Mathematical Software - ICMS 2018, pages 458-465, Cham, 2018. Springer International Publishing. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 498, + 274, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 274, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 274, + 509 + ], + "type": "text", + "content": "[10] Tom Breloff and other contributors. JuliaPlots/Plots.jl. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "text", + "content": "[11] Jan Draisma, Emil Horobet, Giorgio Ottaviani, Bernd Sturmfels, and Rekha R Thomas. The euclidean distance degree of an algebraic variety. Foundations of computational mathematics, 16(1):99-149, 2016. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 556, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 609 + ], + "type": "text", + "content": "[12] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajdla. Plmp-point-line minimal problems in complete multi-view visibility. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1675-1684, 2019. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "text", + "content": "[13] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajda. PI\\$\\_1\\$\\$p - point-line minimal problems under partial visibility in three views. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 175–192, Cham, 2020. Springer International Publishing. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[14] A. Elqursh and A. Elgammal. Line-based relative pose estimation. In Conference on Computer Vision and Pattern Recognition, 2011. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "type": "text", + "content": "[15] Ricardo Fabbri, Timothy Duff, Hongyi Fan, Margaret H Regan, David da Costa de Pinho, Elias Tsigaridas, Charles W Wampler, Jonathan D Hauenstein, Peter J Giblin, Benjamin Kimia, et al. Trplp-trifocal relative pose from lines at points. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12073-12083, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 150, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 172 + ], + "type": "text", + "content": "[16] Olivier Faugeras. Three-dimensional computer vision: A geometric viewpoint. MIT press, Cambridge, MA, 1993. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "text", + "content": "[17] Olivier Faugeras and Bernard Mourrain. On the geometry and algebra of the point and line correspondences between n images. In Proceedings of IEEE International Conference on Computer Vision, pages 951-956. IEEE, 1995. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 259 + ], + "type": "text", + "content": "[18] Andreas Gathmann. Algebraic geometry, 2019/20. Class Notes TU Kaiserslautern. Available at https://www.mathematik.uni-kl.de/~gathmann/de/alggeom.php.4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "text", + "content": "[19] Corey Harris and Daniel Lowengrub. The Chern-mather class of the multiview variety. Communications in Algebra, 46(6):2488-2499, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "type": "text", + "content": "[20] Richard I. Hartley. Lines and points in three views and the trifocal tensor. International Journal of Computer Vision, 22(2):125-140, 1997. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 327, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 349 + ], + "type": "text", + "content": "[21] Richard I Hartley and Peter Sturm. Triangulation. Computer vision and image understanding, 68(2):146-157, 1997. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "type": "text", + "content": "[22] Richard I. Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "type": "text", + "content": "[23] Manuel Hofer, Andreas Wendel, and Horst Bischof. Incremental line-based 3d reconstruction using geometric constraints. The British Machine Vision Association and Society for Pattern Recognition, 2013. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 480 + ], + "type": "text", + "content": "[24] Kun Huang, Yifan Wang, Zihan Zhou, Tianjiao Ding, Shenghua Gao, and Yi Ma. Learning to parse wireframes in images of man-made environments. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 626-635, 2018. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 482, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 535 + ], + "type": "text", + "content": "[25] Javier Civera Joan Solà, Teresa Vidal-Calleja and José María Martínez Montiel. Impact of landmark parametrization on monocular ekf-slam with points and lines. International Journal of Computer Vision, 97(3):339-368, 2012. 3, 9" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 536, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 536, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 545, + 568 + ], + "type": "text", + "content": "[26] Kenichi Kanatani. Statistical optimization for geometric computation: theory and practice. Courier Corporation, 2005. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "text", + "content": "[27] Kenichi Kanatani, Yasuyuki Sugaya, and Hirotaka Niitsuma. Triangulation from two views revisited: Hartley-sturm vs. optimal correction. practice, 4(5):99, 2008. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 603, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 635 + ], + "type": "text", + "content": "[28] Yasushi Kanazawa and Kenichi Kanatani. Reliability of 3- d reconstruction by stereo vision. IEICE TRANSACTIONS on Information and Systems, 78(10):1301-1306, 1995. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 636, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 669 + ], + "type": "text", + "content": "[29] Joseph David Kileel. Algebraic Geometry for Computer Vision. ProQuest LLC, Ann Arbor, MI, 2017. Thesis (Ph.D.)-University of California, Berkeley. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 713 + ], + "type": "text", + "content": "[30] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21445-21455, June 2023. 1" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3756" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 342 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[31] Laurentiu G. Maxim, Jose I. Rodriguez, and Botong Wang. Euclidean distance degree of the multiview variety. SIAM Journal on Applied Algebra and Geometry, 4(1):28-48, 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[32] Branislav Micusik and Horst Wildenauer. Structure from motion with line segments under relaxed endpoint constraints. In 2014 2nd International Conference on 3D Vision, volume 1, pages 13-19, 2014. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 288, + 206 + ], + "type": "text", + "content": "[33] Long Quan and T. Kanade. Affine structure from line correspondences with uncalibrated affine cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(8):834-845, 1997. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 288, + 262 + ], + "type": "text", + "content": "[34] Grant Schindler, Panchapagesan Krishnamurthy, and Frank Dellaert. Line-based structure from motion for urban environments. Third International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT'06), pages 846-853, 2006. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 263, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 263, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 263, + 288, + 308 + ], + "type": "text", + "content": "[35] Henrik Stewénius, Frederik Schaffalitzky, and David Nister. How hard is 3-view triangulation really? In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, volume 1, pages 686-693. IEEE, 2005. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 308, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 308, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 308, + 287, + 342 + ], + "type": "text", + "content": "[36] Matthew Trager, Martial Hebert, and Jean Ponce. The joint image handbook. In Proceedings of the IEEE international conference on computer vision, pages 909-917, 2015. 2, 4" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3757" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/42e3e8c9-44a4-4bef-b505-a2c5b22ff20f_content_list.json b/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/42e3e8c9-44a4-4bef-b505-a2c5b22ff20f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eea39ee229503411d34fc81c4913d65a064d5353 --- /dev/null +++ b/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/42e3e8c9-44a4-4bef-b505-a2c5b22ff20f_content_list.json @@ -0,0 +1,1742 @@ +[ + { + "type": "text", + "text": "Thinking Image Color Aesthetics Assessment: Models, Datasets and Benchmarks", + "text_level": 1, + "bbox": [ + 78, + 130, + 893, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuai He† Anlong Ming†,* Yaqi Li Jinyuan Sun ShunTian Zheng Huadong Ma Beijing University of Posts and Telecommunications", + "bbox": [ + 138, + 180, + 833, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hs19951021, mal, lyq645146943, jinyuan, zhengshuntian, mhd}@bupt.edu.cn", + "bbox": [ + 166, + 218, + 803, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/woshidandan/Image-Color-Aesthetics-Assessment", + "bbox": [ + 197, + 236, + 769, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a comprehensive study on a new task named image color aesthetics assessment (ICAA), which aims to assess color aesthetics based on human perception. ICAA is important for various applications such as imaging measurement and image analysis. However, due to the highly diverse aesthetic preferences and numerous color combinations, ICAA presents more challenges than conventional image quality assessment tasks. To advance ICAA research, 1) we propose a baseline model called the Delegate Transformer, which not only deploys deformable transformers to adaptively allocate interest points, but also learns human color space segmentation behavior by the dedicated module. 2) We elaborately build a color-oriented dataset, ICAA17K, containing 17K images, covering 30 popular color combinations, 80 devices and 50 scenes, with each image densely annotated by more than 1,500 people. Moreover, we develop a large-scale benchmark of 15 methods, the most comprehensive one thus far based on two datasets, SPAQ and ICAA17K. Our work, not only achieves state-of-the-art performance, but more importantly offers the community a roadmap to explore solutions for ICAA. Code and dataset are available in here.", + "bbox": [ + 75, + 318, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 696, + 209, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Color is known as having a higher degree of discriminability and correlation compared to other visual features [1, 2, 3], because the human eye can directly perceive light of different wavelengths and convert them into different color signals. As digital photography expands rapidly, ICAA has become one of the most important criteria to automatically assess whether the image meets users' aesthetic preferences [4, 5, 6, 7]. It is also an essential step in imaging measurements among manufacturers to evaluate the performance of smartphones and cameras [8, 9, 10].", + "bbox": [ + 75, + 723, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/332ce92dfe3feb0f9441c4d51d6d75ae2f5714a88c9ab2d3b39afb151f2a3599.jpg", + "image_caption": [ + "Figure 1. The results of different methods on two similar images with tiny color adjustment, ground-truth (red) and predicted (black) scores are shown below. (a) Similar images. (b) Color histogram [16] is unable to differentiate and quantify the color aesthetics. The saliency maps of (c) MaxViT [17], (d) NIMA [18] and (e) our method, whereas our method achieves a fine-grained perception of color aesthetics." + ], + "image_footnote": [], + "bbox": [ + 504, + 285, + 890, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although ICAA is an important branch of IAA (image aesthetics assessment), they deal with different tasks. IAA evaluates the holistic aesthetics of an image, which implicitly depends on color and other attributes (brightness, sharpness, etc.) [11, 12, 8]. ICAA focuses more on evaluating the impact of color harmony, color combinations and other factors on color perception.", + "bbox": [ + 496, + 535, + 893, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ICAA is confronted with two major challenges. First, the types and combinations of colors are intrinsically complex, thus the color aesthetics are determined by specific colors and their relative positions in color spaces [13]. However, biological studies show that humans can perceive only up to 8 colors due to the human eye and perception limits [14]. Therefore, extracting dominant colors from an image is a crucial process and important prior knowledge [6, 15]. Second, human color preferences are subjective and vary from person to person based on factors such as age and culture that may change over time.", + "bbox": [ + 496, + 642, + 893, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most traditional quantization methods for ICAA rely on objective rules such as color histogram [16] or color wheel [19] theories. However, the color information obtained from pixel statistics is limited to capture the nuances of color aesthetics (Fig. 1 (b)), mainly due to these methods cannot perceive the spatial information of colors and the ef-", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal Contribution; \\*Contact Author.", + "bbox": [ + 101, + 886, + 308, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21838", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fect of image content and semantics. In addition, they are suited to qualitative analysis instead of directly quantifying the aesthetics of colors.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "IAA methods, especially leaning-based IAA approaches such as [18, 20, 17], are not specifically designed for ICAA and then lack of prior color knowledge, which result in struggling to achieve a fine-grained perception of the importance of different color spaces, and being susceptible to interference from multiple visual elements (Fig. 1 (c)(d)). Furthermore, to our knowledge, there are no datasets specifically designed for ICAA, which results in most data-driven methods not generalizing well to deal with subjective ICAA tasks. In addition, the community of image quality assessment lacks of a fair ICAA benchmark for reliable comparisons.", + "bbox": [ + 75, + 136, + 468, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of our paper lie in:", + "bbox": [ + 96, + 316, + 341, + 332 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The limits of human perception and pixel-based extraction methods for dominant colors are revealed, which guides us to propose a baseline model, called the Delegate (Deformable Gate) Transformer, to adaptively allocate interest points for dominant colors and simulate human behavior of color space segmentation.", + "- To verify our method more convincingly, a dataset, named ICAA17K, is exclusively designed for ICAA. Specifically, it is a color-oriented dataset with 17K images and the richest annotations thus far. Furthermore, dedicated preassessment strategy is adopted to alleviate long-tailed distributions of dataset.", + "- Based on ICAA17K, 15 state-of-the-art baselines (see Table 1) are evaluated, which takes our benchmark as the most complete one for ICAA thus far. Our work, not only gains state-of-the art performance, but more importantly serves as a potential catalyst for promoting large-scale model comparisons in future ICAA research." + ], + "bbox": [ + 76, + 335, + 468, + 616 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 641, + 218, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Color Aesthetics Assessment", + "text_level": 1, + "bbox": [ + 76, + 665, + 383, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional quantization methods. To mitigate the gap between subjective and objective evaluations, early methods usually extracted hand-crafted features from brightness, hue, and saturation or applied color difference formulas. For instance, Yang et al. [11] employed the CIEDE2000 color difference formula to align objective evaluation results with subjective visual perception. Shi et al. [30] utilized hue as the primary color information and proposed the structure and hue similarity model (SHSIM) to assess image color quality. Some other works believed that aesthetically pleasing colors typically conform to certain established rules. O'Donovan et al. [6, 31] studied color compatibility theories and color aesthetics based on color pattern (templates). Cohen et al. [19] evaluated color harmony", + "bbox": [ + 75, + 688, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "using a color wheel, while Nishiyama et al. [32] applied a color harmony model to evaluate the distributions of hue, lightness, and chroma.", + "bbox": [ + 496, + 90, + 890, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the above-mentioned methods are based on statistical quantitative information of image pixels, ignoring how spatial and semantic content affect color aesthetics. Although these methods can give qualitative analysis results, they cannot quantify the aesthetic differences brought about by a tiny change in color, as shown in Fig. 1.", + "bbox": [ + 496, + 137, + 890, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data-driven methods. In the deep learning era, some data-driven methods are proposed to fit aesthetic information in images (Table 1). For example, the works of [18, 20, 28, 29] introduced additional context information, e.g., theme and layout information to improve the performance, while others developed models of personalized image aesthetics assessment (PIAA) [22, 33]. The latest works have explored adopting the transformer structure to extract more powerful aesthetic features [17, 27].", + "bbox": [ + 496, + 229, + 890, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, these methods typically extract holistic aesthetic features and lack of prior color knowledge, which take themselves harder to perceive the spatial distribution and composition of different colors in an image, then leading to diffuse attention against perceiving color space. On the other hand, they cannot assign different attention weights based on color importance, which leads to in poor fine-grained perception for color (Fig. 5).", + "bbox": [ + 496, + 364, + 890, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As revealed in Section I, although images are naturally composed of distinct color spaces, dominant and secondary colors have different effects on color perception. The proposed Delegate Transformer learns to segment color space from dedicated deformable attention rather than static pixel values, and thus captures spatial information of color. Furthermore, different color spaces are assigned different levels of attention by the Delegate Transformer, which exactly matches human behavior for color space segmentation.", + "bbox": [ + 496, + 487, + 890, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Image Aesthetics Assessment Datasets", + "text_level": 1, + "bbox": [ + 500, + 633, + 826, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, several datasets have been developed for IAA, as shown in Table 2. Murray et al. [34] created the AVA dataset, a large-scale IAA dataset containing nearly 255,000 images that has become one of the most popular IAA datasets. However, this dataset does not provide annotations for color. Kong et al. [12] introduced the Aesthetics and Attributes Database (AADB), which includes subjective aesthetic scores and attributes annotated by individual users and provides two binary labels for color harmony. Yu et al. [35] developed the Photo Critique Captioning Dataset (PCCD) to address the problem of generating captions for photos based on their aesthetics, which includes evaluations of the color & lighting attributes. Recently, Fang et al. [8] presented the Smartphone Photography Attribute and Quality (SPAQ) dataset with colorfulness attribute scores.", + "bbox": [ + 496, + 657, + 890, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The above datasets primarily focus on evaluating the", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21839", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/a06d686bb95f9d1529fa9c9b15cdd58600999bf69ca45e1904e443daf0a998e9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
No.ModelYearPub.BasicPlatform
1RAPID [21]2014ACMMMincorporate heterogeneousPython&Lua
2AADB [12]2016ECCVsampling strategy, ranking lossMATLAB
3PAM [22]2017ICCVresidual-based, active learningCaffe
4A-Lamp [23]2017CVPRlayout-aware, multi-patchScipy
5NIMA [18]2017TIPpredict distributionTensorflow
6MPada [24]2018ACMMMattention, multi-patchTensorflow
7MLSP [20]2019CVPRstaged training,multi-level featuresTensorflow
8MT-A [8]2020CVPRmulti-task, high-level semantics interactPyTorch
9BIAA [25]2020TCYBmeta-learning, bilevel optimizationPyTorch
10UIAA [26]2020TIPunified probabilistic formulationMATLAB
11MUSIQ [27]2021ICCVmulti-scale representationPyTorch
12HGCN [28]2021CVPRgraph convolution networksJittor
13TANet [29]2022IJCAIfuse theme adaptivelyPyTorch
14MaxViT [17]2022ECCVmulti-axis attentionPyTorch
15Ours2023ICCVdelegate attention, color space segmentationPyTorch
", + "bbox": [ + 80, + 88, + 888, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "holistic quality of images and lack of detailed color annotations, with limited color types or combinations. Specifically, these datasets exhibit serious selection bias. e.g., about $50\\%$ images of the AVA dataset are \"black and white\" images, which outnumber other colors by 10 to 100 times, and the PCCD and SPAQ datasets have few images of \"pink\" and \"violet\" colors. Therefore, these IAA datasets are not suitable for ICAA tasks and cannot support the generalization of ICAA models well.", + "bbox": [ + 75, + 410, + 468, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the above issue, we try to develop a specialized and color-oriented dataset the first time. To the best of our knowledge, our ICAA17K dataset is the largest, as well as most densely annotated ICAA dataset with a diverse range of color types and image collection devices (see Table 2). It also has a dedicated dataset construction strategy to avoid selection bias and long-tailed distributions, which ensures it a high-quality and robust dataset.", + "bbox": [ + 75, + 546, + 468, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 76, + 679, + 267, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dominant or primary colors tend to draw attention quickly and are often the focal point of visual design. Instead of using traditional methods that rely on the number of pixels to differentiate between dominant and secondary colors, we directly employ attention mechanism to distinguish them in a visual context. First, a sparse, data-related, and deformable attention mechanism are introduced in our baseline model to allocate the attention by setting sparse interest points. Second, the color space segmentation module is utilized to group these interest points into different color spaces, and determine the importance of each space.", + "bbox": [ + 75, + 704, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall architecture. An overview of the architecture is presented in Fig. 2 (a). Our implementation begins with", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/655cbee5905f1a2ca3ad6e6080fa46cae90685fc71360460485701d2e7a72767.jpg", + "table_caption": [ + "Table 1. Summary of 15 existing representative IAA or PIAA models and the proposed methods. Code: available code links are marked in red, and our code is provided in the supplementary material." + ], + "table_footnote": [], + "table_body": "
DatasetYearPub.ImagesRating
DP Challenge [36]2008ICIP16,509100
CUHK-PQ [37]2011ICCV17,67310
AVA [34]2012CVPR255,530250
AADB [12]2016ECCV10,0005
PCCD [35]2017ICCV4,2357
DPC-Captions [38]2019ACMMM154,3846
SPAQ [8]2020CVPR11,125600
TAD66K [29]2022IJCAI66,3271,200
ICAA17K (Ours)2023ICCV17,7261,500
", + "bbox": [ + 506, + 407, + 883, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 2. Summary of the related datasets.", + "bbox": [ + 566, + 601, + 818, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "embedding an input image through a $4 \\times 4$ convolution operation. The first and second stages consist of two Swin Transformer blocks [39]. At these stages, the feature maps are processed by shifting window attention to locally aggregate aesthetic information. Subsequently, several proposed Delegate Transformer blocks, which form Stage III and IV, are utilized to process these feature maps globally and model the long-range relationships. Finally, these feature points are sent to a dedicated module to divide the color space and then sent to an output head to predict the color aesthetics score.", + "bbox": [ + 496, + 646, + 890, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Delegate Transformer for Allocating Attention", + "text_level": 1, + "bbox": [ + 500, + 829, + 890, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Utilizing the projection matrices to generate queries $Q$ , keys $K$ , and values $V$ from the original feature map $X$ , the classical multihead self-attention (MSA) can be calculated", + "bbox": [ + 496, + 854, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21840", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/30a509a3bb7aa2c24b457d68e857d8b65ba8405d298fa9a5b0c350c54c7209f5.jpg", + "image_caption": [ + "(a) Overall Architecture" + ], + "image_footnote": [], + "bbox": [ + 122, + 89, + 851, + 209 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/324f949bf5ff669c06fde313683e5391e3c6775a4aa4a616e26af98753ce4007.jpg", + "image_caption": [ + "(b) Delegate Attention Module", + "(c) Delegate Transformer Block" + ], + "image_footnote": [], + "bbox": [ + 119, + 219, + 545, + 377 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a5b6663445b21f4bb1c97b440d9da9dc69a4f34f336b08cb00c704d8dcb036fc.jpg", + "image_caption": [ + "Figure 2. The proposed model. (a) Overall architecture. (b) Feature processing flow. The input features are sent to learn the original interest points and adjusted by offsets, then filtered by the gate module. (d) The entire block. (d) The color space segmentation module, which is modeled as a stack of CNN layers." + ], + "image_footnote": [], + "bbox": [ + 558, + 215, + 705, + 375 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/457adef5a92b09d0c99ffe3070eeebea9be91bf487a145979531dd5ae98b278e.jpg", + "image_caption": [ + "(d) Color Space Segmentation Module" + ], + "image_footnote": [], + "bbox": [ + 717, + 217, + 849, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as:", + "bbox": [ + 76, + 470, + 99, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M S A} = \\operatorname {S o f t m a x} \\left(\\mathbf {Q} ^ {(m)} \\mathbf {K} ^ {(m) T} / \\sqrt {D}\\right) \\mathbf {V} ^ {(m) T}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 493, + 468, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D$ is the query/key dimension, $m$ denotes the $m$ -th attention head. However, the calculation based on the entire feature map significantly increases the computational complexity. The output attention map also requires a large number of calculations when sent to the next block for color space segmentation. Instead of the calculation for the entire feature map, in this paper, the corresponding calculation is delegated to interest points for achieving fewer computations and higher flexibility (Fig. 2(b)).", + "bbox": [ + 75, + 530, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The feature map $X$ is fed into a lightweight subnetwork $Net_{i}$ to generate the initial interest points. Then, another subnetwork $Net_{o}$ is adopted to calculate the offset from $Q$ to adjust the positions of the interest points, which is formulated as:", + "bbox": [ + 76, + 667, + 470, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ = X W _ {Q}, K = \\widetilde {X} W _ {k}, V = \\widetilde {X} W _ {v}, w i t h\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 753, + 413, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {i} = \\emptyset [ X; G * (\\underbrace {\\operatorname {N e t} _ {i} (X)} _ {\\text {i n t e r s t p o i n t s}} + \\underbrace {\\operatorname {N e t} _ {o} (Q)}) ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 776, + 468, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where interest points and offsets are both represented by two-dimensional coordinates that are normalized to the range $[-1, +1]$ , where $(-1, -1)$ represents the top-left corner and $(+1, +1)$ represents the bottom-right corner. A gating module, $G$ , is employed to determine activation of the", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "current adjusted interest point coordinates, reducing redundancy further in the space and preventing outliers and overlapping interest points. It can be expressed as:", + "bbox": [ + 498, + 468, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG = \\operatorname {S i g m o i d} \\left(\\operatorname {N e t} _ {g} (X) \\cdot \\alpha\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 523, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is a value large enough to ensure that the gate matrix is either 0 or 1. After obtaining the adjusted and filtered interest points $\\widetilde{X}$ , the bilinear interpolation is applied to sample the output feature from $\\widetilde{X}$ , which can be represented as", + "bbox": [ + 498, + 551, + 890, + 627 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\varnothing [ X; (p _ {x}, p _ {y}) ] = \\sum_ {(q _ {x}, q _ {y})} M (p _ {x}, p _ {y}) M (q _ {x}, q _ {y}) X (q _ {x}, q _ {y}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 638, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(p_x, p_y)$ represents one of the coordinates of the interest points, $M(a, b) = \\max(0, 1 - |a - b|)$ , and $(q_x, q_y)$ is non-zero on only the 4 integral points in $X$ closest to $(p_x, p_y)$ . Finally, the standard multihead attention mechanism of formula (1) is applied to calculate the output.", + "bbox": [ + 498, + 686, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Module for Color Space Segmentation", + "text_level": 1, + "bbox": [ + 500, + 771, + 828, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The selection of dominant and secondary colors is based on human understanding of the relationship between image content and color, which is highly subjective and leads to uncertainty in the size and segmentation of color spaces. To address this issue, we propose a learning-based approach to determine the size and the segmentation of color spaces. After obtaining the color features of the interest points, 1)", + "bbox": [ + 496, + 794, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21841", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fc88593a5f26dc4b45d01096f6ed17a432f0de7c6fd8f7b15010e03297cc3348.jpg", + "image_caption": [ + "Figure 3. The proposed ICAA17K dataset. (a) Visualization of images with different color types; (b) quantification of 15 complementary/polychromatic colors; (c) quantification of 15 monochromatic colors; (d) examples in the proposed ICAA17K dataset with the ground truth (and confidence) scores shown below each image." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 351, + 284 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cc9ba2ca66102fc0a5dc1ca4a1fe5e85127535e489dc48618e8f5a4feea3b385.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 336, + 89, + 580, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/329bd31d8bd452db0f8029c151a1ecd8aaa5b299a729010e4d7e31e171b327ed.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 580, + 93, + 707, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ac10bd085bfa1fcd0779b2774a81a5183f53bb43b98646c65b9ce98c2ba397cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 99, + 779, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/912345f7027d8e16417efead5266ace274730cdef96037a66bcc43a4c2b05325.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 99, + 859, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5051098053cdec903071ed06e3ec1867a4b70eef6c0d77fb950cff303b4d3fbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 862, + 99, + 885, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/00d6ec8c0eb28a688933aa156118e178889f83b833c1d0dcf78dea25b516b7aa.jpg", + "image_caption": [ + "3.66 (0.74)", + "4.63(0.63) 4.89", + "5.46(0.66)" + ], + "image_footnote": [], + "bbox": [ + 709, + 146, + 795, + 184 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6a9be4e5ae2a3269f8e820e1d56a35a43077cdc7f00c0d02e5d9e3f5b3ad278c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 146, + 821, + 184 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2026a41866eb13b283d42ca7dd39d448ca78cc67412e74ca8b8709bc2a940077.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 185, + 821, + 226 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/926c7dfaeb5827647b94a7e77b2ee8061d7160a42c09a5ddb201fdfb71ebf747.jpg", + "image_caption": [ + "(0.76)" + ], + "image_footnote": [], + "bbox": [ + 821, + 146, + 885, + 184 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ec70bebb05f6dc34d702d98a8942b5b642e95399f4c374ca6749cb2b246d762c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 185, + 885, + 226 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fbf03364f7cb87b333816d3297eb73514bfee9e8195cce7e89d2d5835f4a214e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 226, + 795, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cb8fc96b94bc8ec7582996bbcc7698816f0daa34f63250d960feed446db441f9.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 797, + 227, + 885, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0d18ec308a943c77d22b9e1aac09494ee258b6f06ca2a1044e5f3ff2b9f39827.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 354, + 316, + 441 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8e49bcfc7b6e4fcf8395256bdaf18846ccb5905ec48b7159116ecc11309a59f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 354, + 465, + 422 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0bc106784b3c2928c7a947f5df6a1e8c883885ca9b9e2bf53350774216accf23.jpg", + "image_caption": [ + "(b)", + "Figure 4. An example of color space segmentation with 4 color spaces $(\\kappa = 4)$ . (a) Three sample points. (b) The weights of the three points in each space calculated through formula (5). (c) The weights of each point in each space." + ], + "image_footnote": [], + "bbox": [ + 81, + 441, + 313, + 563 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fd4007517604df7d0eec30f11c7a000e9444347869e0fbdce0834329f0a9ef70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 431, + 464, + 491 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/40190d9223b99351e8d237fa7cb47e42375c60904a9b87efad9263a712058aaa.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 320, + 500, + 464, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the interest points are grouped into several color spaces ( $\\kappa$ pre-defined types). 2) Each color space is mapped to the $[0, +1]$ interval, where each interval can be determined by its widths and centers. 3) By counting the weight distribution of points in each color space, it is easy to distinguish the attention of each color space, thus dividing the dominant and secondary color space.", + "bbox": [ + 75, + 669, + 468, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, using fixed bin widths and centers is not differentiable, and the parameters cannot be updated through backpropagation. To resolve this issue, the widths and centers are adaptively generated using two lightweight networks, $Net_{w}$ and $Net_{c}$ , which can be learned during training. To calculate the uncertainty in color space segmentation, our model learns a probability distribution for each point $p_j$ to indicate its weight in $\\kappa$ color spaces, denoted as", + "bbox": [ + 75, + 779, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$S(p_{j}) = [s_{1}(p_{j}),\\dots,s_{\\kappa}(p_{j})]$ , the weight $s_i()$ of $p_j$ allocated to the i-th space is determined by the piecewise linear basis function.", + "bbox": [ + 500, + 354, + 890, + 400 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} \\left(p _ {j}\\right) = \\max \\left\\{0, 1 - w _ {i} \\times \\left| p _ {j} - c _ {i} \\right| \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 415, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{i}$ and $c_{i}$ are the widths and centers of the i-th color space. If an interest point falls into the i-th bin, e.g., the interval of $[c_i - w_i,c_i + w_i]$ , it votes for the i-th bin with a positive weight $s_i(p_j)$ . Fig. 4 shows an example with $\\kappa = 4$ color spaces.", + "bbox": [ + 496, + 445, + 890, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The operation in formula (5) can be designed into a lightweight color space partition block (Fig. 2 (d)), where $p_j - c_i$ is equivalent to convolving the point feature maps with a fixed $1 \\times 1$ kernel and a learnable bias $c_i$ . Then, after taking the absolute value, the remaining operation is completed using another learnable weight, a fixed $1 \\times 1$ bias, and a ReLU activation function.", + "bbox": [ + 496, + 521, + 890, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Proposed Dataset", + "text_level": 1, + "bbox": [ + 500, + 640, + 669, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Image Collection", + "text_level": 1, + "bbox": [ + 500, + 665, + 666, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "How to avoid selection bias by building a comprehensive ICAA dataset. Selection bias occurs when some types of the intended color, scenes and devices are less likely to be included than others, it can affect the validity and generalizability of training models. To solve this issue, 1) we analyzed the most frequently uploaded colors on the Flickr and COLOURLovers website from 2009 to 2022. Finally, 15 complementary/polychromatic colors and 15 monochromatic colors are selected as the basic color types, which ensures that the dataset includes a diverse range of color types (Fig. 3). 2) To improve the dataset's generalizability across different scenes, the images in ICAA17K cover more than 50 scene categories, e.g., animals, plants, mountains, landscapes, and night scenes. 3) We collected images", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21842", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3ee87521a80e9bd1ad858d0d399f8e03e6ecb2d0239ffce8b212c599f4123e7b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Metric [ICAA]2014-20192020-2022
RAPID [21]AADB [12]PAM [22]A-Lamp [23]NIMA [18]\\( MP_{ada} \\) [24]MLSP [20]MT-A [8]BIAA [25]UIAA [26]MUSIQ [27]HGCN [28]TANet [29]MaxViT [17]Ours *1Ours *2
SRCC↑.759.788.796.810.809.810.826.821.820.817.835.826.829.855.873.887
LCC↑.771.793.804.818.815.819.837.830.829.830.841.831.836.874.890.901
Accu↑.836.887.899.908.906.906.921.908.913.910.918.909.903.920.957.965
Epoch↓80606050605050601006080601102002540
", + "bbox": [ + 81, + 89, + 890, + 205 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/447afd4ef330ad70fc4be740a81e994c8833b037f67f6cfcf5c73a0615b20258.jpg", + "table_caption": [ + "Table 3. Comparisons of 15 models on our ICAA17K dataset. We retrained the models for the best performance with the recommended parameter and dataset settings. See subsection 5.1 for details regarding the metrics. \"Ours *1\" indicates that, similar to other methods, only the color annotation scores from the ICAA17K dataset are used during training. \"Ours *2\" indicates that the holistic aesthetics score is jointly predicted when using multiple tasks." + ], + "table_footnote": [], + "table_body": "
Metric [SPAQ]2014-20192020-2022
RAPID [21]AADB [12]PAM [22]A-Lamp [23]NIMA [18]\\( MP_{ada} \\) [24]MLSP [20]MT-A [8]BIAA [25]UIAA [26]MUSIQ [27]HGCN [28]TANet [29]MaxViT [17]Ours * 1
SRCC↑.721.725.732.747.730.746.753.751.748.750.753.755.750.753.770
LCC↑.730.739.744.760.746.757.773.763.759.763.770.771.761.770.791
Epoch↓95757065556555308570904511015025
", + "bbox": [ + 91, + 282, + 879, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Comparisons of 15 models on the SPAQ dataset [8]. All models used only the colorfulness score in the dataset for training, and were retrained for the best performance using the recommended parameters and dataset settings.", + "bbox": [ + 75, + 393, + 892, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "taken by more than 80 different types of devices, including mobile phones (e.g., iPhone X, Huawei Mate 10, and Samsung Galaxy S10) and digital cameras (e.g., Canon, Nikon, Sony, Panasonic), to encompass a wide range of devices.", + "bbox": [ + 75, + 446, + 468, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "How to alleviate long-tailed distributions by performing a preassessment strategy. The annotated label in the most aesthetic datasets have long-tailed distributions [34, 8], which bias models trained on imbalanced data toward majority examples. To avoid correcting the long-tailed distributions after label annotation, we propose a preassessment strategy during data collection to maximize label diversity.", + "bbox": [ + 75, + 507, + 468, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, the color harmony [19] is applied to evaluate objective color quality by measuring the distance between the tested image and templates on the hue wheel. Given a harmonic template $T(\\beta)$ (one of the predefined type of templates $T$ ) with an associated orientation $\\beta$ , the color harmony $f_{obj}$ of a tested image $C$ could be calculated as follows:", + "bbox": [ + 75, + 628, + 468, + 731 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {o b j} (C) = \\sum_ {l \\in C} \\| H (l) - E _ {T (\\beta)} (l) \\| \\cdot Y (l), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 739, + 468, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $l$ is the hue of a pixel, $E_{T(\\beta)}$ is the sector border hue of harmonic template $T(\\beta)$ , $H$ and $Y$ mean the hue and the saturation channels, respectively, $\\|\\cdot\\|$ denotes the arc-length distance on the hue wheel (measured in radians).", + "bbox": [ + 75, + 779, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, we crawl images from Flickr and record their metadata, including \"views\" (number of visits $VI$ ) and \"faves\" (number of clicks that favor image $FA$ ). Intuitively, an image with a higher ratio of faves to views is more pop-", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ular and beautiful. Specifically, considering the affect of user's herd mentality, we adopt the following formula to approximate the subjective perception:", + "bbox": [ + 496, + 446, + 892, + 492 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {s u b j} (C) = \\frac {F A (C)}{\\sqrt {V I (C)}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 503, + 890, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we normalize the $f_{subj}$ and $f_{obj}$ , and add them together as the final preassessment score of the image (not shown to annotators), and extract balanced samples from each score range for the next stage of manual evaluation.", + "bbox": [ + 496, + 549, + 890, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Image Annotation", + "text_level": 1, + "bbox": [ + 500, + 619, + 674, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "How to avoid unreliablely by improving annotation quality. The traditional annotation process suffers from a lack of a baseline or reference for evaluating aesthetics, which makes subjective responses unreliable in the early stages. To mitigate this problem, 1) we first provided annotators simple guidelines based on photography theories to guide their judgment of image color aesthetics. 2) We provided anchor images with different rating scales as references. Experts rated the anchor images based on their rich photography experience, allowing annotators to understand the differences between rating scales. Before annotating, annotators had to review the anchor images and guidelines thoroughly. The average value was then calculated as the ground truth. 3) Annotators could provide a confidence score (from 0 to 1) to indicate their certainty in each rating. This information can be useful for studying the noise and subjectivity in the dataset. 4) Finally, each image has", + "bbox": [ + 496, + 643, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21843", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/689db2d5d58389a81be6a6170f33d7557dd7823f72c581f3a1f44d26689e2a94.jpg", + "image_caption": [ + "Figure 5. Saliency maps of the 15 models. The attention of our method is segmented into distinct regions based on the color distribution." + ], + "image_footnote": [], + "bbox": [ + 163, + 89, + 803, + 338 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/abad8b514bc70e63ac3fc6ffd9aae3c3ef8a0efab3b5f454756d67d6dbfecaec.jpg", + "image_caption": [ + "Figure 6. Analysis of kernel sizes in $Net_{i}$ and the number of color spaces $\\kappa$ . Larger kernel sizes result in fewer interest points." + ], + "image_footnote": [], + "bbox": [ + 119, + 380, + 429, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "an annotation score from 1 to 10, which represents color aesthetic perception from lowest to highest, we selected images with an average confidence score above 0.5 as samples in the dataset. We collected about 1,500 opinions for each image. Some examples are shown in Fig. 3 (d).", + "bbox": [ + 76, + 627, + 468, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 741, + 282, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benchmark Datasets. We performed model evaluations on two datasets, the proposed ICAA17K and SPAQ datasets [8]. Although the SPAQ dataset focuses on image quality rather than aesthetics, it provides an annotation on colorfulness, which can also afford testing the ability of various baseline models to extract color features to some extent.", + "bbox": [ + 76, + 763, + 467, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benchmark Models and Training Protocols. To the best of our knowledge, there is no publicly available learning-based ICAA model. Therefore, we selected 14 deep learn", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing baselines (Table. 1) according to the following criteria: 1) classical architectures with available code and 2) SOTA in a specific field, e.g., personalized IAA. These baselines are trained with the recommended parameter settings (e.g., optimizer and batch size), using the same training and testing settings. The mean squared error (MSE) is adopted as the loss function across all methods.", + "bbox": [ + 498, + 382, + 890, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt two popular evaluation metrics, the Spearman's rank correlation coefficient (SRCC) and the linear correlation coefficient (LCC), to evaluate the model's ability. The metrics for the ICAA17K dataset also include binary classification accuracy (color aesthetically negative or positive). Since the ground-truth annotations in the ICAA17K dataset are based on a 5-point boundary line, if an annotator tends to give a score greater than 5, it is considered aesthetically positive; otherwise, it is considered negative. Therefore, this metric can evaluate the model's ability to perform coarse-grained evaluations.", + "bbox": [ + 498, + 488, + 890, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results and Dataset Analysis", + "text_level": 1, + "bbox": [ + 500, + 665, + 756, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ICAA17K Dataset. Table 3 lists the results of 15 models on the ICAA dataset. The proposed baseline method (marked by \"Ours *1\") achieves the best performance on all metrics and has a higher training speed. This is mainly due to its sparse attention and adaptive color space partitioning. Furthermore, our ICAA17K dataset also provides additional annotations regarding holistic aesthetics. Therefore, we incorporate this additional supervision information into the multitask training (marked by \"Ours *2\"), which improves performance but increases the training time.", + "bbox": [ + 498, + 688, + 890, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SPAQ Dataset. Based on the results of the 15 methods reported in Table 4, our method still achieves the top performance on all metrics with fewer training epochs, showing that it is a promising solution for the IQA problem.", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21844", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b308b498a137da42536feeaeb0ca2401664b9205abff113abb75053adcb5947c.jpg", + "image_caption": [ + "Figure 7. Visualization of our attention mechanism. (a) Sampled interest points adjusted by offsets. (b) Interest points filtered by the gating module. (c) The radius of the circles represents the attention weights at multiple heads, with a larger radius indicating a higher weight. (d) The color space segmentation of each point, where the numbers represent the primary and secondary relationship of the color space, e.g., \"1\" represents the dominant color." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 271, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea74892bb8a1944a3a5caa49c8434988ed8bd79d74b666bb183a1ac9de88fd52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 90, + 444, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dataset Analysis. Compared to the SPAQ dataset, when using the ICAA17K dataset for training, most networks converge faster and have higher SRCC and LCC metrics, which indicates that models trained on the ICAA17K dataset more accurately fit human subjective color perception. It also suggests that introducing more annotators (SPAQ 600 vs ICAA17K 1500) and providing more annotation references for annotators can alleviate noise caused by subjective annotation, which is a promising direction toward improving the quality of IQA or IAA datasets.", + "bbox": [ + 75, + 383, + 468, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 544, + 230, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Delegate vs Shift/Deformable Transformers. In Table 5, we replace the Delegate Transformer blocks with the shift attention mechanism of Swin Transformer [39] and deformable attention of DAT Transformer [40] in the last stage. In this case, our model has better performance and lower computation, which outperforms Swin by $5.2\\%$ in the SRCC with $41.7\\%$ FLOPs and outperforms DAT by $4.4\\%$ in the SRCC with $41.7\\%$ FLOPs. Additionally, the stacking way also has a significant impact on the performance. The Delegate Transformer blocks may result in information loss if it is employed in the early stages due to the sparse attention mechanism.", + "bbox": [ + 75, + 568, + 468, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Number of Interest Points and Color Space. The number of interest points mainly depends on the kernel size of the subnetwork $Net_{i}$ . In Fig. 6, the results show that an inadequate or excessive number of interest points leads to degradation in the model's performance. Moreover, selecting the color space hyperparameter $\\kappa$ is also crucial for the performance of our model. Our experiments reveal that the optimal performance is achieved when $\\kappa = 7$ . Additionally, a larger value of $\\kappa$ leads to an increase in computation and memory requirements. In contrast, a", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2e52d80c05a16be86ce230ce62efd927261c5f5b007b719d05a9d629b21f75f7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#StagesFLOPs↓SRCC↑LCC↑
1234
1SSSS240G0.8300.843
2SSDD240G0.8360.848
3SSGG140G0.8730.890
4GGSS210G0.8280.835
5SGGS140G0.8380.850
6GSSG210G0.8250.833
7GGGG130G0.8370.846
", + "bbox": [ + 516, + 89, + 879, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablations on the application of different attention mechanisms in different stages, where $S$ , $D$ and $G$ represent the application of Swin, DAT and our blocks, respectively.", + "bbox": [ + 498, + 247, + 890, + 289 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "smaller value of $\\kappa$ may not provide a sufficient characterization of the color composition in the input image. Moreover, the effectiveness of different modules are shown in the Appendix.", + "bbox": [ + 498, + 314, + 890, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Visualization", + "text_level": 1, + "bbox": [ + 500, + 383, + 635, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The GradCAM [41] method is applied to visualize the saliency maps in Fig. 5. Our model differs from other methods because it segments the image into distinct regions based on the color distribution, and the region that typically receives the greatest attention is the region where the dominant color is located. The reason is further explained in Fig. 7. First, our model generates interest points via $\\text{Net}_i$ and adjusts their positions by $\\text{Net}_o$ . Second, some redundant points are eliminated via $\\text{Net}_g$ and weights are calculated using multihead attention. Third, the color space segmentation module and formula (5) are employed to divide the different spaces and direct different attention toward them.", + "bbox": [ + 498, + 407, + 890, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 602, + 625, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, a comprehensive study on ICAA has been conducted by developing a Delegate Transformer model as a baseline, constructing a color-oriented dataset benchmark, and establishing the largest-scale benchmarks to date. Compared with traditional or data-driven models, our model achieves SOTA performance and produces visually consistent results. We hope our work should contribute to the community's research on ICAA. However, ICAA remains far from being solved, we will optimize our approach by incorporating more color prior knowledge in the future.", + "bbox": [ + 496, + 627, + 890, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 792, + 679, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the Funds for Creative Research Groups of China under Grant 61921003, the national key R & D program intergovernmental international science and technology innovation cooperation project 2021YFE0101600.", + "bbox": [ + 496, + 816, + 890, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21845", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ahmed Talib, Massudi Mahmuddin, Husniza Husni, and Loay E George. A weighted dominant color descriptor for content-based image retrieval. Journal of Visual Communication and Image Representation, 24(3):345-360, 2013.", + "[2] Yining Deng, BS Manjunath, Charles Kenney, Michael S Moore, and Hyundoo Shin. An efficient color representation for image retrieval. IEEE Transactions on image processing, 10(1):140-147, 2001.", + "[3] Bangalore S Manjunath, J-R Ohm, Vinod V Vasudevan, and Akio Yamada. Color and texture descriptors. IEEE Transactions on circuits and systems for video technology, 11(6):703-715, 2001.", + "[4] Luca Marchesotti, Naila Murray, and Florent Perronnin. Discovering beautiful attributes for aesthetic image analysis. International journal of computer vision, 113:246-266, 2015.", + "[5] Andrew J Elliot and Markus A Maier. Color psychology: Effects of perceiving color on psychological functioning in humans. Annual review of psychology, 65:95-120, 2014.", + "[6] Peter O'Donovan, Aseem Agarwala, and Aaron Hertzmann. Color compatibility from large datasets. In ACM SIGGRAPH 2011 papers, pages 1-12. 2011.", + "[7] Joon-Young Lee, Kalyan Sunkavalli, Zhe Lin, Xiaohui Shen, and In So Kweon. Automatic content-aware color and tone stylization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2470–2478, 2016.", + "[8] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In CVPR, pages 3677-3686, 2020.", + "[9] Emilie Baudin, François-Xavier Bucher, Laurent Chanas, and Frédéric Guichard. Dxomark objective video quality measurements. *Electronic Imaging*, 2020(9):166-1, 2020.", + "[10] Sabine E Susstrunk and Stefan Winkler. Color image quality on the internet. In _Internet imaging V_, volume 5304, pages 118-131. SPIE, 2003.", + "[11] Yang Yang, Jun Ming, and Nenghai Yu. Color image quality assessment based on ciede2000. Advances in Multimedia, 2012, 2012.", + "[12] Shu Kong, Xiaohui Shen, Zhe Lin, Radomir Mech, and Charless Fowlkes. Photo aesthetics ranking network with attributes and content adaptation. In ECCV, pages 662-679. Springer, 2016.", + "[13] Xuan Xu and Xin Li. Scan: Spatial color attention networks for real single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019.", + "[14] Aleksandra Mojsilovic, H Hu, and Emina Soljanin. Extraction of perceptually important colors and similarity measurement for image matching, retrieval and analysis. IEEE Transactions on Image Processing, 11(11):1238-1248, 2002.", + "[15] Stephen E Palmer, Karen B Schloss, and Jonathan Sammartino. Visual aesthetics and human preference. Annual review of psychology, 64:77-107, 2013." + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Michael J Swain and Dana H Ballard. Color indexing. International journal of computer vision, 7(1):11-32, 1991.", + "[17] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. ECCV, 2022.", + "[18] Hossein Talebi and Peyman Milanfar. NIMA: Neural image assessment. TIP, 27(8):3998-4011, 2018.", + "[19] Daniel Cohen-Or, Olga Sorkine, Ran Gal, Tommer Leyvand, and Ying-Qing Xu. Color harmonization. In ACM SIGGRAPH 2006 Papers, pages 624-630. 2006.", + "[20] Vlad Hosu, Bastian Goldlucke, and Dietmar Saupe. Effective aesthetics prediction with multi-level spatially pooled features. In CVPR, 2019.", + "[21] Xin Lu, Zhe Lin, Hailin Jin, Jianchao Yang, and James Z. Wang. RAPID: Rating pictorial aesthetics using deep learning. In ACMMM, pages 457-466, 2014.", + "[22] Jian Ren, Xiaohui Shen, Zhe Lin, Radomir Mech, and David J Foran. Personalized image aesthetics. In ICCV, pages 638-647, 2017.", + "[23] Shuang Ma, Jing Liu, and Chang Wen Chen. A-lamp: Adaptive layout-aware multi-patch deep convolutional neural network for photo aesthetic assessment. In CVPR, pages 722-731, 2017.", + "[24] Kekai Sheng, Weiming Dong, Chongyang Ma, Xing Mei, Feiyue Huang, and Bao-Gang Hu. Attention-based multipatch aggregation for image aesthetic assessment. In ACMMM, 2018.", + "[25] Hancheng Zhu, Leida Li, Jinjian Wu, Sicheng Zhao, Guiguang Ding, and Guangming Shi. Personalized image aesthetics assessment via meta-learning with bilevel gradient optimization. TCYB, 2020.", + "[26] Hui Zeng, Zisheng Cao, Lei Zhang, and Alan C Bovik. A unified probabilistic formulation of image aesthetic assessment. TIP, 29:1548-1561, 2019.", + "[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, pages 5148-5157, 2021.", + "[28] Dongyu She, Yu-Kun Lai, Gaoxiong Yi, and Kun Xu. Hierarchical layout-aware graph convolutional network for unified aesthetics assessment. In CVPR, pages 8475-8484, 2021.", + "[29] Shuai He, Yongchang Zhang, Rui Xie, Dongxiang Jiang, and Anlong Ming. Rethinking image aesthetics assessment: Models, datasets and benchmarks. *IJCAI*, 2022.", + "[30] Yunyu Shi, Youdong Ding, Ranran Zhang, and Jun Li. Structure and hue similarity for color image quality assessment. In 2009 International Conference on Electronic Computer Technology, pages 329-333. IEEE, 2009.", + "[31] Peter O'Donovan, Aseem Agarwala, and Aaron Hertzmann. Collaborative filtering of color aesthetics. In Proceedings of the Workshop on Computational Aesthetics, pages 33-40, 2014.", + "[32] Masashi Nishiyama, Takahiro Okabe, Imari Sato, and Yoichi Sato. Aesthetic quality classification of photographs based on color harmony. In CVPR 2011, pages 33-40. IEEE, 2011." + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21846", + "bbox": [ + 478, + 926, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Pei Lv, Jianqi Fan, Xixi Nie, Weiming Dong, Xiaoheng Jiang, Bing Zhou, Mingliang Xu, and Changsheng Xu. User-guided personalized image aesthetic assessment based on deep reinforcement learning. arXiv:2106.07488, 2021.", + "[34] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In CVPR. IEEE, 2012.", + "[35] Kuang-Yu Chang, Kung-Hung Lu, and Chu-Song Chen. Aesthetic critiques generation for photos. In ICCV, pages 3514-3523, 2017.", + "[36] Ritendra Datta, Jia Li, and James Z Wang. Algorithmic infor-ferencing of aesthetics and emotion in natural images: An exposition. In ICIP. IEEE, 2008.", + "[37] Wei Luo, Xiaogang Wang, and Xiaou Tang. Content-based photo quality assessment. In ICCV, pages 2206-2213. IEEE, 2011.", + "[38] Xin Jin, Le Wu, Geng Zhao, Xiaodong Li, Xiaokun Zhang, Shiming Ge, Dongqing Zou, Bin Zhou, and Xinghui Zhou. Aesthetic attributes assessment of images. In ACMMM, pages 311-319, 2019.", + "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv:2103.14030, 2021.", + "[40] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4794-4803, 2022.", + "[41] Jacob Gildenblat and contributors. Pytorch library for cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021." + ], + "bbox": [ + 78, + 90, + 468, + 561 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21847", + "bbox": [ + 478, + 926, + 517, + 936 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/42e3e8c9-44a4-4bef-b505-a2c5b22ff20f_model.json b/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/42e3e8c9-44a4-4bef-b505-a2c5b22ff20f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f7b74ded192ad562785ab3a870a0aa167c9b3251 --- /dev/null +++ b/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/42e3e8c9-44a4-4bef-b505-a2c5b22ff20f_model.json @@ -0,0 +1,2299 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.895, + 0.155 + ], + "angle": 0, + "content": "Thinking Image Color Aesthetics Assessment: Models, Datasets and Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.181, + 0.834, + 0.218 + ], + "angle": 0, + "content": "Shuai He† Anlong Ming†,* Yaqi Li Jinyuan Sun ShunTian Zheng Huadong Ma Beijing University of Posts and Telecommunications" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.219, + 0.804, + 0.235 + ], + "angle": 0, + "content": "{hs19951021, mal, lyq645146943, jinyuan, zhengshuntian, mhd}@bupt.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.237, + 0.771, + 0.251 + ], + "angle": 0, + "content": "https://github.com/woshidandan/Image-Color-Aesthetics-Assessment" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.651 + ], + "angle": 0, + "content": "We present a comprehensive study on a new task named image color aesthetics assessment (ICAA), which aims to assess color aesthetics based on human perception. ICAA is important for various applications such as imaging measurement and image analysis. However, due to the highly diverse aesthetic preferences and numerous color combinations, ICAA presents more challenges than conventional image quality assessment tasks. To advance ICAA research, 1) we propose a baseline model called the Delegate Transformer, which not only deploys deformable transformers to adaptively allocate interest points, but also learns human color space segmentation behavior by the dedicated module. 2) We elaborately build a color-oriented dataset, ICAA17K, containing 17K images, covering 30 popular color combinations, 80 devices and 50 scenes, with each image densely annotated by more than 1,500 people. Moreover, we develop a large-scale benchmark of 15 methods, the most comprehensive one thus far based on two datasets, SPAQ and ICAA17K. Our work, not only achieves state-of-the-art performance, but more importantly offers the community a roadmap to explore solutions for ICAA. Code and dataset are available in here." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.698, + 0.21, + 0.714 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Color is known as having a higher degree of discriminability and correlation compared to other visual features [1, 2, 3], because the human eye can directly perceive light of different wavelengths and convert them into different color signals. As digital photography expands rapidly, ICAA has become one of the most important criteria to automatically assess whether the image meets users' aesthetic preferences [4, 5, 6, 7]. It is also an essential step in imaging measurements among manufacturers to evaluate the performance of smartphones and cameras [8, 9, 10]." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.286, + 0.891, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.411, + 0.895, + 0.509 + ], + "angle": 0, + "content": "Figure 1. The results of different methods on two similar images with tiny color adjustment, ground-truth (red) and predicted (black) scores are shown below. (a) Similar images. (b) Color histogram [16] is unable to differentiate and quantify the color aesthetics. The saliency maps of (c) MaxViT [17], (d) NIMA [18] and (e) our method, whereas our method achieves a fine-grained perception of color aesthetics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.894, + 0.643 + ], + "angle": 0, + "content": "Although ICAA is an important branch of IAA (image aesthetics assessment), they deal with different tasks. IAA evaluates the holistic aesthetics of an image, which implicitly depends on color and other attributes (brightness, sharpness, etc.) [11, 12, 8]. ICAA focuses more on evaluating the impact of color harmony, color combinations and other factors on color perception." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.895, + 0.809 + ], + "angle": 0, + "content": "ICAA is confronted with two major challenges. First, the types and combinations of colors are intrinsically complex, thus the color aesthetics are determined by specific colors and their relative positions in color spaces [13]. However, biological studies show that humans can perceive only up to 8 colors due to the human eye and perception limits [14]. Therefore, extracting dominant colors from an image is a crucial process and important prior knowledge [6, 15]. Second, human color preferences are subjective and vary from person to person based on factors such as age and culture that may change over time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Most traditional quantization methods for ICAA rely on objective rules such as color histogram [16] or color wheel [19] theories. However, the color information obtained from pixel statistics is limited to capture the nuances of color aesthetics (Fig. 1 (b)), mainly due to these methods cannot perceive the spatial information of colors and the ef-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.887, + 0.31, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal Contribution; \\*Contact Author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21838" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "fect of image content and semantics. In addition, they are suited to qualitative analysis instead of directly quantifying the aesthetics of colors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.317 + ], + "angle": 0, + "content": "IAA methods, especially leaning-based IAA approaches such as [18, 20, 17], are not specifically designed for ICAA and then lack of prior color knowledge, which result in struggling to achieve a fine-grained perception of the importance of different color spaces, and being susceptible to interference from multiple visual elements (Fig. 1 (c)(d)). Furthermore, to our knowledge, there are no datasets specifically designed for ICAA, which results in most data-driven methods not generalizing well to deal with subjective ICAA tasks. In addition, the community of image quality assessment lacks of a fair ICAA benchmark for reliable comparisons." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.318, + 0.343, + 0.333 + ], + "angle": 0, + "content": "The contributions of our paper lie in:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.337, + 0.469, + 0.429 + ], + "angle": 0, + "content": "- The limits of human perception and pixel-based extraction methods for dominant colors are revealed, which guides us to propose a baseline model, called the Delegate (Deformable Gate) Transformer, to adaptively allocate interest points for dominant colors and simulate human behavior of color space segmentation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.432, + 0.469, + 0.522 + ], + "angle": 0, + "content": "- To verify our method more convincingly, a dataset, named ICAA17K, is exclusively designed for ICAA. Specifically, it is a color-oriented dataset with 17K images and the richest annotations thus far. Furthermore, dedicated preassessment strategy is adopted to alleviate long-tailed distributions of dataset." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.527, + 0.469, + 0.617 + ], + "angle": 0, + "content": "- Based on ICAA17K, 15 state-of-the-art baselines (see Table 1) are evaluated, which takes our benchmark as the most complete one for ICAA thus far. Our work, not only gains state-of-the art performance, but more importantly serves as a potential catalyst for promoting large-scale model comparisons in future ICAA research." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.337, + 0.469, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.642, + 0.219, + 0.657 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.385, + 0.683 + ], + "angle": 0, + "content": "2.1. Image Color Aesthetics Assessment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Traditional quantization methods. To mitigate the gap between subjective and objective evaluations, early methods usually extracted hand-crafted features from brightness, hue, and saturation or applied color difference formulas. For instance, Yang et al. [11] employed the CIEDE2000 color difference formula to align objective evaluation results with subjective visual perception. Shi et al. [30] utilized hue as the primary color information and proposed the structure and hue similarity model (SHSIM) to assess image color quality. Some other works believed that aesthetically pleasing colors typically conform to certain established rules. O'Donovan et al. [6, 31] studied color compatibility theories and color aesthetics based on color pattern (templates). Cohen et al. [19] evaluated color harmony" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "using a color wheel, while Nishiyama et al. [32] applied a color harmony model to evaluate the distributions of hue, lightness, and chroma." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.228 + ], + "angle": 0, + "content": "However, the above-mentioned methods are based on statistical quantitative information of image pixels, ignoring how spatial and semantic content affect color aesthetics. Although these methods can give qualitative analysis results, they cannot quantify the aesthetic differences brought about by a tiny change in color, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.892, + 0.365 + ], + "angle": 0, + "content": "Data-driven methods. In the deep learning era, some data-driven methods are proposed to fit aesthetic information in images (Table 1). For example, the works of [18, 20, 28, 29] introduced additional context information, e.g., theme and layout information to improve the performance, while others developed models of personalized image aesthetics assessment (PIAA) [22, 33]. The latest works have explored adopting the transformer structure to extract more powerful aesthetic features [17, 27]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.892, + 0.487 + ], + "angle": 0, + "content": "Nevertheless, these methods typically extract holistic aesthetic features and lack of prior color knowledge, which take themselves harder to perceive the spatial distribution and composition of different colors in an image, then leading to diffuse attention against perceiving color space. On the other hand, they cannot assign different attention weights based on color importance, which leads to in poor fine-grained perception for color (Fig. 5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.623 + ], + "angle": 0, + "content": "As revealed in Section I, although images are naturally composed of distinct color spaces, dominant and secondary colors have different effects on color perception. The proposed Delegate Transformer learns to segment color space from dedicated deformable attention rather than static pixel values, and thus captures spatial information of color. Furthermore, different color spaces are assigned different levels of attention by the Delegate Transformer, which exactly matches human behavior for color space segmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.828, + 0.65 + ], + "angle": 0, + "content": "2.2. Image Aesthetics Assessment Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.884 + ], + "angle": 0, + "content": "In recent years, several datasets have been developed for IAA, as shown in Table 2. Murray et al. [34] created the AVA dataset, a large-scale IAA dataset containing nearly 255,000 images that has become one of the most popular IAA datasets. However, this dataset does not provide annotations for color. Kong et al. [12] introduced the Aesthetics and Attributes Database (AADB), which includes subjective aesthetic scores and attributes annotated by individual users and provides two binary labels for color harmony. Yu et al. [35] developed the Photo Critique Captioning Dataset (PCCD) to address the problem of generating captions for photos based on their aesthetics, which includes evaluations of the color & lighting attributes. Recently, Fang et al. [8] presented the Smartphone Photography Attribute and Quality (SPAQ) dataset with colorfulness attribute scores." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The above datasets primarily focus on evaluating the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "21839" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.89, + 0.348 + ], + "angle": 0, + "content": "
No.ModelYearPub.BasicPlatform
1RAPID [21]2014ACMMMincorporate heterogeneousPython&Lua
2AADB [12]2016ECCVsampling strategy, ranking lossMATLAB
3PAM [22]2017ICCVresidual-based, active learningCaffe
4A-Lamp [23]2017CVPRlayout-aware, multi-patchScipy
5NIMA [18]2017TIPpredict distributionTensorflow
6MPada [24]2018ACMMMattention, multi-patchTensorflow
7MLSP [20]2019CVPRstaged training,multi-level featuresTensorflow
8MT-A [8]2020CVPRmulti-task, high-level semantics interactPyTorch
9BIAA [25]2020TCYBmeta-learning, bilevel optimizationPyTorch
10UIAA [26]2020TIPunified probabilistic formulationMATLAB
11MUSIQ [27]2021ICCVmulti-scale representationPyTorch
12HGCN [28]2021CVPRgraph convolution networksJittor
13TANet [29]2022IJCAIfuse theme adaptivelyPyTorch
14MaxViT [17]2022ECCVmulti-axis attentionPyTorch
15Ours2023ICCVdelegate attention, color space segmentationPyTorch
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.358, + 0.893, + 0.389 + ], + "angle": 0, + "content": "Table 1. Summary of 15 existing representative IAA or PIAA models and the proposed methods. Code: available code links are marked in red, and our code is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.47, + 0.546 + ], + "angle": 0, + "content": "holistic quality of images and lack of detailed color annotations, with limited color types or combinations. Specifically, these datasets exhibit serious selection bias. e.g., about \\(50\\%\\) images of the AVA dataset are \"black and white\" images, which outnumber other colors by 10 to 100 times, and the PCCD and SPAQ datasets have few images of \"pink\" and \"violet\" colors. Therefore, these IAA datasets are not suitable for ICAA tasks and cannot support the generalization of ICAA models well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.669 + ], + "angle": 0, + "content": "To address the above issue, we try to develop a specialized and color-oriented dataset the first time. To the best of our knowledge, our ICAA17K dataset is the largest, as well as most densely annotated ICAA dataset with a diverse range of color types and image collection devices (see Table 2). It also has a dedicated dataset construction strategy to avoid selection bias and long-tailed distributions, which ensures it a high-quality and robust dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.268, + 0.697 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Dominant or primary colors tend to draw attention quickly and are often the focal point of visual design. Instead of using traditional methods that rely on the number of pixels to differentiate between dominant and secondary colors, we directly employ attention mechanism to distinguish them in a visual context. First, a sparse, data-related, and deformable attention mechanism are introduced in our baseline model to allocate the attention by setting sparse interest points. Second, the color space segmentation module is utilized to group these interest points into different color spaces, and determine the importance of each space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Overall architecture. An overview of the architecture is presented in Fig. 2 (a). Our implementation begins with" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.409, + 0.885, + 0.593 + ], + "angle": 0, + "content": "
DatasetYearPub.ImagesRating
DP Challenge [36]2008ICIP16,509100
CUHK-PQ [37]2011ICCV17,67310
AVA [34]2012CVPR255,530250
AADB [12]2016ECCV10,0005
PCCD [35]2017ICCV4,2357
DPC-Captions [38]2019ACMMM154,3846
SPAQ [8]2020CVPR11,125600
TAD66K [29]2022IJCAI66,3271,200
ICAA17K (Ours)2023ICCV17,7261,500
" + }, + { + "type": "table_caption", + "bbox": [ + 0.567, + 0.602, + 0.82, + 0.616 + ], + "angle": 0, + "content": "Table 2. Summary of the related datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.892, + 0.812 + ], + "angle": 0, + "content": "embedding an input image through a \\(4 \\times 4\\) convolution operation. The first and second stages consist of two Swin Transformer blocks [39]. At these stages, the feature maps are processed by shifting window attention to locally aggregate aesthetic information. Subsequently, several proposed Delegate Transformer blocks, which form Stage III and IV, are utilized to process these feature maps globally and model the long-range relationships. Finally, these feature points are sent to a dedicated module to divide the color space and then sent to an output head to predict the color aesthetics score." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.892, + 0.846 + ], + "angle": 0, + "content": "3.1. Delegate Transformer for Allocating Attention" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Utilizing the projection matrices to generate queries \\( Q \\), keys \\( K \\), and values \\( V \\) from the original feature map \\( X \\), the classical multihead self-attention (MSA) can be calculated" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "21840" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.09, + 0.852, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.21, + 0.548, + 0.22 + ], + "angle": 0, + "content": "(a) Overall Architecture" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.22, + 0.546, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.38, + 0.413, + 0.393 + ], + "angle": 0, + "content": "(b) Delegate Attention Module" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.216, + 0.706, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.38, + 0.707, + 0.393 + ], + "angle": 0, + "content": "(c) Delegate Transformer Block" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.218, + 0.85, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.376, + 0.84, + 0.398 + ], + "angle": 0, + "content": "(d) Color Space Segmentation Module" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.402, + 0.895, + 0.445 + ], + "angle": 0, + "content": "Figure 2. The proposed model. (a) Overall architecture. (b) Feature processing flow. The input features are sent to learn the original interest points and adjusted by offsets, then filtered by the gate module. (d) The entire block. (d) The color space segmentation module, which is modeled as a stack of CNN layers." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.471, + 0.101, + 0.482 + ], + "angle": 0, + "content": "as:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.494, + 0.47, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M S A} = \\operatorname {S o f t m a x} \\left(\\mathbf {Q} ^ {(m)} \\mathbf {K} ^ {(m) T} / \\sqrt {D}\\right) \\mathbf {V} ^ {(m) T}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.47, + 0.668 + ], + "angle": 0, + "content": "where \\(D\\) is the query/key dimension, \\(m\\) denotes the \\(m\\)-th attention head. However, the calculation based on the entire feature map significantly increases the computational complexity. The output attention map also requires a large number of calculations when sent to the next block for color space segmentation. Instead of the calculation for the entire feature map, in this paper, the corresponding calculation is delegated to interest points for achieving fewer computations and higher flexibility (Fig. 2(b))." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.669, + 0.471, + 0.743 + ], + "angle": 0, + "content": "The feature map \\( X \\) is fed into a lightweight subnetwork \\( Net_{i} \\) to generate the initial interest points. Then, another subnetwork \\( Net_{o} \\) is adopted to calculate the offset from \\( Q \\) to adjust the positions of the interest points, which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.755, + 0.415, + 0.775 + ], + "angle": 0, + "content": "\\[\nQ = X W _ {Q}, K = \\widetilde {X} W _ {k}, V = \\widetilde {X} W _ {v}, w i t h\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.777, + 0.469, + 0.813 + ], + "angle": 0, + "content": "\\[\nX _ {i} = \\emptyset [ X; G * (\\underbrace {\\operatorname {N e t} _ {i} (X)} _ {\\text {i n t e r s t p o i n t s}} + \\underbrace {\\operatorname {N e t} _ {o} (Q)}) ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where interest points and offsets are both represented by two-dimensional coordinates that are normalized to the range \\([-1, +1]\\), where \\((-1, -1)\\) represents the top-left corner and \\((+1, +1)\\) represents the bottom-right corner. A gating module, \\(G\\), is employed to determine activation of the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.469, + 0.892, + 0.515 + ], + "angle": 0, + "content": "current adjusted interest point coordinates, reducing redundancy further in the space and preventing outliers and overlapping interest points. It can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.525, + 0.892, + 0.543 + ], + "angle": 0, + "content": "\\[\nG = \\operatorname {S i g m o i d} \\left(\\operatorname {N e t} _ {g} (X) \\cdot \\alpha\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.553, + 0.892, + 0.628 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a value large enough to ensure that the gate matrix is either 0 or 1. After obtaining the adjusted and filtered interest points \\(\\widetilde{X}\\), the bilinear interpolation is applied to sample the output feature from \\(\\widetilde{X}\\), which can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.64, + 0.892, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\varnothing [ X; (p _ {x}, p _ {y}) ] = \\sum_ {(q _ {x}, q _ {y})} M (p _ {x}, p _ {y}) M (q _ {x}, q _ {y}) X (q _ {x}, q _ {y}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.688, + 0.892, + 0.764 + ], + "angle": 0, + "content": "where \\((p_x, p_y)\\) represents one of the coordinates of the interest points, \\(M(a, b) = \\max(0, 1 - |a - b|)\\), and \\((q_x, q_y)\\) is non-zero on only the 4 integral points in \\(X\\) closest to \\((p_x, p_y)\\). Finally, the standard multihead attention mechanism of formula (1) is applied to calculate the output." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.83, + 0.789 + ], + "angle": 0, + "content": "3.2. Module for Color Space Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The selection of dominant and secondary colors is based on human understanding of the relationship between image content and color, which is highly subjective and leads to uncertainty in the size and segmentation of color spaces. To address this issue, we propose a learning-based approach to determine the size and the segmentation of color spaces. After obtaining the color features of the interest points, 1)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "21841" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.352, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.09, + 0.581, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.277, + 0.467, + 0.286 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.094, + 0.708, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.277, + 0.658, + 0.286 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.101, + 0.781, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.14, + 0.77, + 0.147 + ], + "angle": 0, + "content": "3.66 (0.74)" + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.101, + 0.861, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.14, + 0.857, + 0.147 + ], + "angle": 0, + "content": "4.63(0.63) 4.89" + }, + { + "type": "image", + "bbox": [ + 0.864, + 0.101, + 0.886, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.864, + 0.14, + 0.885, + 0.147 + ], + "angle": 0, + "content": "(0.76)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.147, + 0.797, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.186, + 0.767, + 0.192 + ], + "angle": 0, + "content": "5.46(0.66)" + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.147, + 0.822, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.186, + 0.822, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.823, + 0.147, + 0.886, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.823, + 0.186, + 0.886, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.227, + 0.797, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.228, + 0.886, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.277, + 0.814, + 0.285 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.894, + 0.332 + ], + "angle": 0, + "content": "Figure 3. The proposed ICAA17K dataset. (a) Visualization of images with different color types; (b) quantification of 15 complementary/polychromatic colors; (c) quantification of 15 monochromatic colors; (d) examples in the proposed ICAA17K dataset with the ground truth (and confidence) scores shown below each image." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.355, + 0.317, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.356, + 0.466, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.442, + 0.315, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.566, + 0.208, + 0.578 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.432, + 0.465, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.5, + 0.465, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.566, + 0.402, + 0.578 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.638 + ], + "angle": 0, + "content": "Figure 4. An example of color space segmentation with 4 color spaces \\((\\kappa = 4)\\). (a) Three sample points. (b) The weights of the three points in each space calculated through formula (5). (c) The weights of each point in each space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.469, + 0.776 + ], + "angle": 0, + "content": "the interest points are grouped into several color spaces (\\(\\kappa\\) pre-defined types). 2) Each color space is mapped to the \\([0, +1]\\) interval, where each interval can be determined by its widths and centers. 3) By counting the weight distribution of points in each color space, it is easy to distinguish the attention of each color space, thus dividing the dominant and secondary color space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.902 + ], + "angle": 0, + "content": "However, using fixed bin widths and centers is not differentiable, and the parameters cannot be updated through backpropagation. To resolve this issue, the widths and centers are adaptively generated using two lightweight networks, \\(Net_{w}\\) and \\(Net_{c}\\), which can be learned during training. To calculate the uncertainty in color space segmentation, our model learns a probability distribution for each point \\(p_j\\) to indicate its weight in \\(\\kappa\\) color spaces, denoted as" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.356, + 0.892, + 0.401 + ], + "angle": 0, + "content": "\\(S(p_{j}) = [s_{1}(p_{j}),\\dots,s_{\\kappa}(p_{j})]\\), the weight \\(s_i()\\) of \\(p_j\\) allocated to the i-th space is determined by the piecewise linear basis function." + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.416, + 0.892, + 0.433 + ], + "angle": 0, + "content": "\\[\ns _ {i} \\left(p _ {j}\\right) = \\max \\left\\{0, 1 - w _ {i} \\times \\left| p _ {j} - c _ {i} \\right| \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.52 + ], + "angle": 0, + "content": "where \\(w_{i}\\) and \\(c_{i}\\) are the widths and centers of the i-th color space. If an interest point falls into the i-th bin, e.g., the interval of \\([c_i - w_i,c_i + w_i]\\), it votes for the i-th bin with a positive weight \\(s_i(p_j)\\). Fig. 4 shows an example with \\(\\kappa = 4\\) color spaces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.627 + ], + "angle": 0, + "content": "The operation in formula (5) can be designed into a lightweight color space partition block (Fig. 2 (d)), where \\( p_j - c_i \\) is equivalent to convolving the point feature maps with a fixed \\( 1 \\times 1 \\) kernel and a learnable bias \\( c_i \\). Then, after taking the absolute value, the remaining operation is completed using another learnable weight, a fixed \\( 1 \\times 1 \\) bias, and a ReLU activation function." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.641, + 0.671, + 0.658 + ], + "angle": 0, + "content": "4. Proposed Dataset" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.667, + 0.682 + ], + "angle": 0, + "content": "4.1. Image Collection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "How to avoid selection bias by building a comprehensive ICAA dataset. Selection bias occurs when some types of the intended color, scenes and devices are less likely to be included than others, it can affect the validity and generalizability of training models. To solve this issue, 1) we analyzed the most frequently uploaded colors on the Flickr and COLOURLovers website from 2009 to 2022. Finally, 15 complementary/polychromatic colors and 15 monochromatic colors are selected as the basic color types, which ensures that the dataset includes a diverse range of color types (Fig. 3). 2) To improve the dataset's generalizability across different scenes, the images in ICAA17K cover more than 50 scene categories, e.g., animals, plants, mountains, landscapes, and night scenes. 3) We collected images" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.52, + 0.938 + ], + "angle": 0, + "content": "21842" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.207 + ], + "angle": 0, + "content": "
Metric [ICAA]2014-20192020-2022
RAPID [21]AADB [12]PAM [22]A-Lamp [23]NIMA [18]\\( MP_{ada} \\) [24]MLSP [20]MT-A [8]BIAA [25]UIAA [26]MUSIQ [27]HGCN [28]TANet [29]MaxViT [17]Ours *1Ours *2
SRCC↑.759.788.796.810.809.810.826.821.820.817.835.826.829.855.873.887
LCC↑.771.793.804.818.815.819.837.830.829.830.841.831.836.874.890.901
Accu↑.836.887.899.908.906.906.921.908.913.910.918.909.903.920.957.965
Epoch↓80606050605050601006080601102002540
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.216, + 0.893, + 0.272 + ], + "angle": 0, + "content": "Table 3. Comparisons of 15 models on our ICAA17K dataset. We retrained the models for the best performance with the recommended parameter and dataset settings. See subsection 5.1 for details regarding the metrics. \"Ours *1\" indicates that, similar to other methods, only the color annotation scores from the ICAA17K dataset are used during training. \"Ours *2\" indicates that the holistic aesthetics score is jointly predicted when using multiple tasks." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.284, + 0.88, + 0.386 + ], + "angle": 0, + "content": "
Metric [SPAQ]2014-20192020-2022
RAPID [21]AADB [12]PAM [22]A-Lamp [23]NIMA [18]\\( MP_{ada} \\) [24]MLSP [20]MT-A [8]BIAA [25]UIAA [26]MUSIQ [27]HGCN [28]TANet [29]MaxViT [17]Ours * 1
SRCC↑.721.725.732.747.730.746.753.751.748.750.753.755.750.753.770
LCC↑.730.739.744.760.746.757.773.763.759.763.770.771.761.770.791
Epoch↓95757065556555308570904511015025
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.394, + 0.893, + 0.423 + ], + "angle": 0, + "content": "Table 4. Comparisons of 15 models on the SPAQ dataset [8]. All models used only the colorfulness score in the dataset for training, and were retrained for the best performance using the recommended parameters and dataset settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.469, + 0.508 + ], + "angle": 0, + "content": "taken by more than 80 different types of devices, including mobile phones (e.g., iPhone X, Huawei Mate 10, and Samsung Galaxy S10) and digital cameras (e.g., Canon, Nikon, Sony, Panasonic), to encompass a wide range of devices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.469, + 0.628 + ], + "angle": 0, + "content": "How to alleviate long-tailed distributions by performing a preassessment strategy. The annotated label in the most aesthetic datasets have long-tailed distributions [34, 8], which bias models trained on imbalanced data toward majority examples. To avoid correcting the long-tailed distributions after label annotation, we propose a preassessment strategy during data collection to maximize label diversity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.732 + ], + "angle": 0, + "content": "First, the color harmony [19] is applied to evaluate objective color quality by measuring the distance between the tested image and templates on the hue wheel. Given a harmonic template \\( T(\\beta) \\) (one of the predefined type of templates \\( T \\)) with an associated orientation \\( \\beta \\), the color harmony \\( f_{obj} \\) of a tested image \\( C \\) could be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.74, + 0.469, + 0.772 + ], + "angle": 0, + "content": "\\[\nf _ {o b j} (C) = \\sum_ {l \\in C} \\| H (l) - E _ {T (\\beta)} (l) \\| \\cdot Y (l), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.84 + ], + "angle": 0, + "content": "where \\(l\\) is the hue of a pixel, \\(E_{T(\\beta)}\\) is the sector border hue of harmonic template \\(T(\\beta)\\), \\(H\\) and \\(Y\\) mean the hue and the saturation channels, respectively, \\(\\|\\cdot\\|\\) denotes the arc-length distance on the hue wheel (measured in radians)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Second, we crawl images from Flickr and record their metadata, including \"views\" (number of visits \\(VI\\)) and \"faves\" (number of clicks that favor image \\(FA\\)). Intuitively, an image with a higher ratio of faves to views is more pop-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.893, + 0.493 + ], + "angle": 0, + "content": "ular and beautiful. Specifically, considering the affect of user's herd mentality, we adopt the following formula to approximate the subjective perception:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.504, + 0.892, + 0.54 + ], + "angle": 0, + "content": "\\[\nf _ {s u b j} (C) = \\frac {F A (C)}{\\sqrt {V I (C)}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Finally, we normalize the \\( f_{subj} \\) and \\( f_{obj} \\), and add them together as the final preassessment score of the image (not shown to annotators), and extract balanced samples from each score range for the next stage of manual evaluation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.676, + 0.637 + ], + "angle": 0, + "content": "4.2. Image Annotation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.901 + ], + "angle": 0, + "content": "How to avoid unreliablely by improving annotation quality. The traditional annotation process suffers from a lack of a baseline or reference for evaluating aesthetics, which makes subjective responses unreliable in the early stages. To mitigate this problem, 1) we first provided annotators simple guidelines based on photography theories to guide their judgment of image color aesthetics. 2) We provided anchor images with different rating scales as references. Experts rated the anchor images based on their rich photography experience, allowing annotators to understand the differences between rating scales. Before annotating, annotators had to review the anchor images and guidelines thoroughly. The average value was then calculated as the ground truth. 3) Annotators could provide a confidence score (from 0 to 1) to indicate their certainty in each rating. This information can be useful for studying the noise and subjectivity in the dataset. 4) Finally, each image has" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "21843" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.165, + 0.09, + 0.805, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.344, + 0.882, + 0.358 + ], + "angle": 0, + "content": "Figure 5. Saliency maps of the 15 models. The attention of our method is segmented into distinct regions based on the color distribution." + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.381, + 0.431, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.574, + 0.47, + 0.603 + ], + "angle": 0, + "content": "Figure 6. Analysis of kernel sizes in \\(Net_{i}\\) and the number of color spaces \\(\\kappa\\). Larger kernel sizes result in fewer interest points." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.628, + 0.47, + 0.703 + ], + "angle": 0, + "content": "an annotation score from 1 to 10, which represents color aesthetic perception from lowest to highest, we selected images with an average confidence score above 0.5 as samples in the dataset. We collected about 1,500 opinions for each image. Some examples are shown in Fig. 3 (d)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.717, + 0.21, + 0.734 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.742, + 0.284, + 0.758 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.468, + 0.854 + ], + "angle": 0, + "content": "Benchmark Datasets. We performed model evaluations on two datasets, the proposed ICAA17K and SPAQ datasets [8]. Although the SPAQ dataset focuses on image quality rather than aesthetics, it provides an annotation on colorfulness, which can also afford testing the ability of various baseline models to extract color features to some extent." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Benchmark Models and Training Protocols. To the best of our knowledge, there is no publicly available learning-based ICAA model. Therefore, we selected 14 deep learn" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.383, + 0.892, + 0.487 + ], + "angle": 0, + "content": "ing baselines (Table. 1) according to the following criteria: 1) classical architectures with available code and 2) SOTA in a specific field, e.g., personalized IAA. These baselines are trained with the recommended parameter settings (e.g., optimizer and batch size), using the same training and testing settings. The mean squared error (MSE) is adopted as the loss function across all methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.489, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt two popular evaluation metrics, the Spearman's rank correlation coefficient (SRCC) and the linear correlation coefficient (LCC), to evaluate the model's ability. The metrics for the ICAA17K dataset also include binary classification accuracy (color aesthetically negative or positive). Since the ground-truth annotations in the ICAA17K dataset are based on a 5-point boundary line, if an annotator tends to give a score greater than 5, it is considered aesthetically positive; otherwise, it is considered negative. Therefore, this metric can evaluate the model's ability to perform coarse-grained evaluations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.757, + 0.682 + ], + "angle": 0, + "content": "5.2. Results and Dataset Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.689, + 0.892, + 0.84 + ], + "angle": 0, + "content": "ICAA17K Dataset. Table 3 lists the results of 15 models on the ICAA dataset. The proposed baseline method (marked by \"Ours *1\") achieves the best performance on all metrics and has a higher training speed. This is mainly due to its sparse attention and adaptive color space partitioning. Furthermore, our ICAA17K dataset also provides additional annotations regarding holistic aesthetics. Therefore, we incorporate this additional supervision information into the multitask training (marked by \"Ours *2\"), which improves performance but increases the training time." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "SPAQ Dataset. Based on the results of the 15 methods reported in Table 4, our method still achieves the top performance on all metrics with fewer training epochs, showing that it is a promising solution for the IQA problem." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "21844" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.272, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.091, + 0.446, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.262, + 0.473, + 0.359 + ], + "angle": 0, + "content": "Figure 7. Visualization of our attention mechanism. (a) Sampled interest points adjusted by offsets. (b) Interest points filtered by the gating module. (c) The radius of the circles represents the attention weights at multiple heads, with a larger radius indicating a higher weight. (d) The color space segmentation of each point, where the numbers represent the primary and secondary relationship of the color space, e.g., \"1\" represents the dominant color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.47, + 0.536 + ], + "angle": 0, + "content": "Dataset Analysis. Compared to the SPAQ dataset, when using the ICAA17K dataset for training, most networks converge faster and have higher SRCC and LCC metrics, which indicates that models trained on the ICAA17K dataset more accurately fit human subjective color perception. It also suggests that introducing more annotators (SPAQ 600 vs ICAA17K 1500) and providing more annotation references for annotators can alleviate noise caused by subjective annotation, which is a promising direction toward improving the quality of IQA or IAA datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.232, + 0.563 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Delegate vs Shift/Deformable Transformers. In Table 5, we replace the Delegate Transformer blocks with the shift attention mechanism of Swin Transformer [39] and deformable attention of DAT Transformer [40] in the last stage. In this case, our model has better performance and lower computation, which outperforms Swin by \\(5.2\\%\\) in the SRCC with \\(41.7\\%\\) FLOPs and outperforms DAT by \\(4.4\\%\\) in the SRCC with \\(41.7\\%\\) FLOPs. Additionally, the stacking way also has a significant impact on the performance. The Delegate Transformer blocks may result in information loss if it is employed in the early stages due to the sparse attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The Number of Interest Points and Color Space. The number of interest points mainly depends on the kernel size of the subnetwork \\(Net_{i}\\). In Fig. 6, the results show that an inadequate or excessive number of interest points leads to degradation in the model's performance. Moreover, selecting the color space hyperparameter \\(\\kappa\\) is also crucial for the performance of our model. Our experiments reveal that the optimal performance is achieved when \\(\\kappa = 7\\). Additionally, a larger value of \\(\\kappa\\) leads to an increase in computation and memory requirements. In contrast, a" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.88, + 0.238 + ], + "angle": 0, + "content": "
#StagesFLOPs↓SRCC↑LCC↑
1234
1SSSS240G0.8300.843
2SSDD240G0.8360.848
3SSGG140G0.8730.890
4GGSS210G0.8280.835
5SGGS140G0.8380.850
6GSSG210G0.8250.833
7GGGG130G0.8370.846
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.248, + 0.892, + 0.29 + ], + "angle": 0, + "content": "Table 5. Ablations on the application of different attention mechanisms in different stages, where \\( S \\), \\( D \\) and \\( G \\) represent the application of Swin, DAT and our blocks, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.315, + 0.892, + 0.375 + ], + "angle": 0, + "content": "smaller value of \\(\\kappa\\) may not provide a sufficient characterization of the color composition in the input image. Moreover, the effectiveness of different modules are shown in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.385, + 0.637, + 0.4 + ], + "angle": 0, + "content": "5.4. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.408, + 0.892, + 0.59 + ], + "angle": 0, + "content": "The GradCAM [41] method is applied to visualize the saliency maps in Fig. 5. Our model differs from other methods because it segments the image into distinct regions based on the color distribution, and the region that typically receives the greatest attention is the region where the dominant color is located. The reason is further explained in Fig. 7. First, our model generates interest points via \\( \\text{Net}_i \\) and adjusts their positions by \\( \\text{Net}_o \\). Second, some redundant points are eliminated via \\( \\text{Net}_g \\) and weights are calculated using multihead attention. Third, the color space segmentation module and formula (5) are employed to divide the different spaces and direct different attention toward them." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.603, + 0.627, + 0.618 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.78 + ], + "angle": 0, + "content": "In this paper, a comprehensive study on ICAA has been conducted by developing a Delegate Transformer model as a baseline, constructing a color-oriented dataset benchmark, and establishing the largest-scale benchmarks to date. Compared with traditional or data-driven models, our model achieves SOTA performance and produces visually consistent results. We hope our work should contribute to the community's research on ICAA. However, ICAA remains far from being solved, we will optimize our approach by incorporating more color prior knowledge in the future." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.793, + 0.68, + 0.81 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.818, + 0.892, + 0.894 + ], + "angle": 0, + "content": "This work was supported by the Funds for Creative Research Groups of China under Grant 61921003, the national key R & D program intergovernmental international science and technology innovation cooperation project 2021YFE0101600." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "21845" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Ahmed Talib, Massudi Mahmuddin, Husniza Husni, and Loay E George. A weighted dominant color descriptor for content-based image retrieval. Journal of Visual Communication and Image Representation, 24(3):345-360, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.174, + 0.472, + 0.231 + ], + "angle": 0, + "content": "[2] Yining Deng, BS Manjunath, Charles Kenney, Michael S Moore, and Hyundoo Shin. An efficient color representation for image retrieval. IEEE Transactions on image processing, 10(1):140-147, 2001." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.233, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[3] Bangalore S Manjunath, J-R Ohm, Vinod V Vasudevan, and Akio Yamada. Color and texture descriptors. IEEE Transactions on circuits and systems for video technology, 11(6):703-715, 2001." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.293, + 0.47, + 0.335 + ], + "angle": 0, + "content": "[4] Luca Marchesotti, Naila Murray, and Florent Perronnin. Discovering beautiful attributes for aesthetic image analysis. International journal of computer vision, 113:246-266, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.338, + 0.468, + 0.38 + ], + "angle": 0, + "content": "[5] Andrew J Elliot and Markus A Maier. Color psychology: Effects of perceiving color on psychological functioning in humans. Annual review of psychology, 65:95-120, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.468, + 0.425 + ], + "angle": 0, + "content": "[6] Peter O'Donovan, Aseem Agarwala, and Aaron Hertzmann. Color compatibility from large datasets. In ACM SIGGRAPH 2011 papers, pages 1-12. 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.428, + 0.468, + 0.497 + ], + "angle": 0, + "content": "[7] Joon-Young Lee, Kalyan Sunkavalli, Zhe Lin, Xiaohui Shen, and In So Kweon. Automatic content-aware color and tone stylization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2470–2478, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.468, + 0.543 + ], + "angle": 0, + "content": "[8] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In CVPR, pages 3677-3686, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.547, + 0.468, + 0.589 + ], + "angle": 0, + "content": "[9] Emilie Baudin, François-Xavier Bucher, Laurent Chanas, and Frédéric Guichard. Dxomark objective video quality measurements. *Electronic Imaging*, 2020(9):166-1, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.592, + 0.468, + 0.633 + ], + "angle": 0, + "content": "[10] Sabine E Susstrunk and Stefan Winkler. Color image quality on the internet. In _Internet imaging V_, volume 5304, pages 118-131. SPIE, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.468, + 0.678 + ], + "angle": 0, + "content": "[11] Yang Yang, Jun Ming, and Nenghai Yu. Color image quality assessment based on ciede2000. Advances in Multimedia, 2012, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.682, + 0.468, + 0.738 + ], + "angle": 0, + "content": "[12] Shu Kong, Xiaohui Shen, Zhe Lin, Radomir Mech, and Charless Fowlkes. Photo aesthetics ranking network with attributes and content adaptation. In ECCV, pages 662-679. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.741, + 0.47, + 0.797 + ], + "angle": 0, + "content": "[13] Xuan Xu and Xin Li. Scan: Spatial color attention networks for real single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.801, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[14] Aleksandra Mojsilovic, H Hu, and Emina Soljanin. Extraction of perceptually important colors and similarity measurement for image matching, retrieval and analysis. IEEE Transactions on Image Processing, 11(11):1238-1248, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.902 + ], + "angle": 0, + "content": "[15] Stephen E Palmer, Karen B Schloss, and Jonathan Sammartino. Visual aesthetics and human preference. Annual review of psychology, 64:77-107, 2013." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "[16] Michael J Swain and Dana H Ballard. Color indexing. International journal of computer vision, 7(1):11-32, 1991." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.124, + 0.892, + 0.166 + ], + "angle": 0, + "content": "[17] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.169, + 0.892, + 0.197 + ], + "angle": 0, + "content": "[18] Hossein Talebi and Peyman Milanfar. NIMA: Neural image assessment. TIP, 27(8):3998-4011, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.201, + 0.892, + 0.243 + ], + "angle": 0, + "content": "[19] Daniel Cohen-Or, Olga Sorkine, Ran Gal, Tommer Leyvand, and Ying-Qing Xu. Color harmonization. In ACM SIGGRAPH 2006 Papers, pages 624-630. 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[20] Vlad Hosu, Bastian Goldlucke, and Dietmar Saupe. Effective aesthetics prediction with multi-level spatially pooled features. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.334 + ], + "angle": 0, + "content": "[21] Xin Lu, Zhe Lin, Hailin Jin, Jianchao Yang, and James Z. Wang. RAPID: Rating pictorial aesthetics using deep learning. In ACMMM, pages 457-466, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.337, + 0.892, + 0.379 + ], + "angle": 0, + "content": "[22] Jian Ren, Xiaohui Shen, Zhe Lin, Radomir Mech, and David J Foran. Personalized image aesthetics. In ICCV, pages 638-647, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.382, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[23] Shuang Ma, Jing Liu, and Chang Wen Chen. A-lamp: Adaptive layout-aware multi-patch deep convolutional neural network for photo aesthetic assessment. In CVPR, pages 722-731, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[24] Kekai Sheng, Weiming Dong, Chongyang Ma, Xing Mei, Feiyue Huang, and Bao-Gang Hu. Attention-based multipatch aggregation for image aesthetic assessment. In ACMMM, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.5, + 0.892, + 0.556 + ], + "angle": 0, + "content": "[25] Hancheng Zhu, Leida Li, Jinjian Wu, Sicheng Zhao, Guiguang Ding, and Guangming Shi. Personalized image aesthetics assessment via meta-learning with bilevel gradient optimization. TCYB, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.559, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[26] Hui Zeng, Zisheng Cao, Lei Zhang, and Alan C Bovik. A unified probabilistic formulation of image aesthetic assessment. TIP, 29:1548-1561, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.647 + ], + "angle": 0, + "content": "[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, pages 5148-5157, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.892, + 0.693 + ], + "angle": 0, + "content": "[28] Dongyu She, Yu-Kun Lai, Gaoxiong Yi, and Kun Xu. Hierarchical layout-aware graph convolutional network for unified aesthetics assessment. In CVPR, pages 8475-8484, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.696, + 0.892, + 0.738 + ], + "angle": 0, + "content": "[29] Shuai He, Yongchang Zhang, Rui Xie, Dongxiang Jiang, and Anlong Ming. Rethinking image aesthetics assessment: Models, datasets and benchmarks. *IJCAI*, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.741, + 0.892, + 0.797 + ], + "angle": 0, + "content": "[30] Yunyu Shi, Youdong Ding, Ranran Zhang, and Jun Li. Structure and hue similarity for color image quality assessment. In 2009 International Conference on Electronic Computer Technology, pages 329-333. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.8, + 0.892, + 0.855 + ], + "angle": 0, + "content": "[31] Peter O'Donovan, Aseem Agarwala, and Aaron Hertzmann. Collaborative filtering of color aesthetics. In Proceedings of the Workshop on Computational Aesthetics, pages 33-40, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[32] Masashi Nishiyama, Takahiro Okabe, Imari Sato, and Yoichi Sato. Aesthetic quality classification of photographs based on color harmony. In CVPR 2011, pages 33-40. IEEE, 2011." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.927, + 0.521, + 0.938 + ], + "angle": 0, + "content": "21846" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[33] Pei Lv, Jianqi Fan, Xixi Nie, Weiming Dong, Xiaoheng Jiang, Bing Zhou, Mingliang Xu, and Changsheng Xu. User-guided personalized image aesthetic assessment based on deep reinforcement learning. arXiv:2106.07488, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.153, + 0.47, + 0.194 + ], + "angle": 0, + "content": "[34] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In CVPR. IEEE, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.2, + 0.469, + 0.24 + ], + "angle": 0, + "content": "[35] Kuang-Yu Chang, Kung-Hung Lu, and Chu-Song Chen. Aesthetic critiques generation for photos. In ICCV, pages 3514-3523, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[36] Ritendra Datta, Jia Li, and James Z Wang. Algorithmic infor-ferencing of aesthetics and emotion in natural images: An exposition. In ICIP. IEEE, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[37] Wei Luo, Xiaogang Wang, and Xiaou Tang. Content-based photo quality assessment. In ICCV, pages 2206-2213. IEEE, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.339, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[38] Xin Jin, Le Wu, Geng Zhao, Xiaodong Li, Xiaokun Zhang, Shiming Ge, Dongqing Zou, Bin Zhou, and Xinghui Zhou. Aesthetic attributes assessment of images. In ACMMM, pages 311-319, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.454 + ], + "angle": 0, + "content": "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv:2103.14030, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[40] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4794-4803, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.521, + 0.469, + 0.562 + ], + "angle": 0, + "content": "[41] Jacob Gildenblat and contributors. Pytorch library for cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "21847" + } + ] +] \ No newline at end of file diff --git a/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/42e3e8c9-44a4-4bef-b505-a2c5b22ff20f_origin.pdf b/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/42e3e8c9-44a4-4bef-b505-a2c5b22ff20f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0c2d40dbfd27cd7585aaa949e2a2a50f1d3b7f12 --- /dev/null +++ b/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/42e3e8c9-44a4-4bef-b505-a2c5b22ff20f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18f5e1916dac1f0c006bda8c4ed8fc7ae4f5c2d758ffb07f438cee95b4e07376 +size 6149287 diff --git a/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/full.md b/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e923cb3f1fbd4db5dadc2faf369d8774ec75f91f --- /dev/null +++ b/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/full.md @@ -0,0 +1,335 @@ +# Thinking Image Color Aesthetics Assessment: Models, Datasets and Benchmarks + +Shuai He† Anlong Ming†,* Yaqi Li Jinyuan Sun ShunTian Zheng Huadong Ma Beijing University of Posts and Telecommunications + +{hs19951021, mal, lyq645146943, jinyuan, zhengshuntian, mhd}@bupt.edu.cn + +https://github.com/woshidandan/Image-Color-Aesthetics-Assessment + +# Abstract + +We present a comprehensive study on a new task named image color aesthetics assessment (ICAA), which aims to assess color aesthetics based on human perception. ICAA is important for various applications such as imaging measurement and image analysis. However, due to the highly diverse aesthetic preferences and numerous color combinations, ICAA presents more challenges than conventional image quality assessment tasks. To advance ICAA research, 1) we propose a baseline model called the Delegate Transformer, which not only deploys deformable transformers to adaptively allocate interest points, but also learns human color space segmentation behavior by the dedicated module. 2) We elaborately build a color-oriented dataset, ICAA17K, containing 17K images, covering 30 popular color combinations, 80 devices and 50 scenes, with each image densely annotated by more than 1,500 people. Moreover, we develop a large-scale benchmark of 15 methods, the most comprehensive one thus far based on two datasets, SPAQ and ICAA17K. Our work, not only achieves state-of-the-art performance, but more importantly offers the community a roadmap to explore solutions for ICAA. Code and dataset are available in here. + +# 1. Introduction + +Color is known as having a higher degree of discriminability and correlation compared to other visual features [1, 2, 3], because the human eye can directly perceive light of different wavelengths and convert them into different color signals. As digital photography expands rapidly, ICAA has become one of the most important criteria to automatically assess whether the image meets users' aesthetic preferences [4, 5, 6, 7]. It is also an essential step in imaging measurements among manufacturers to evaluate the performance of smartphones and cameras [8, 9, 10]. + +![](images/332ce92dfe3feb0f9441c4d51d6d75ae2f5714a88c9ab2d3b39afb151f2a3599.jpg) +Figure 1. The results of different methods on two similar images with tiny color adjustment, ground-truth (red) and predicted (black) scores are shown below. (a) Similar images. (b) Color histogram [16] is unable to differentiate and quantify the color aesthetics. The saliency maps of (c) MaxViT [17], (d) NIMA [18] and (e) our method, whereas our method achieves a fine-grained perception of color aesthetics. + +Although ICAA is an important branch of IAA (image aesthetics assessment), they deal with different tasks. IAA evaluates the holistic aesthetics of an image, which implicitly depends on color and other attributes (brightness, sharpness, etc.) [11, 12, 8]. ICAA focuses more on evaluating the impact of color harmony, color combinations and other factors on color perception. + +ICAA is confronted with two major challenges. First, the types and combinations of colors are intrinsically complex, thus the color aesthetics are determined by specific colors and their relative positions in color spaces [13]. However, biological studies show that humans can perceive only up to 8 colors due to the human eye and perception limits [14]. Therefore, extracting dominant colors from an image is a crucial process and important prior knowledge [6, 15]. Second, human color preferences are subjective and vary from person to person based on factors such as age and culture that may change over time. + +Most traditional quantization methods for ICAA rely on objective rules such as color histogram [16] or color wheel [19] theories. However, the color information obtained from pixel statistics is limited to capture the nuances of color aesthetics (Fig. 1 (b)), mainly due to these methods cannot perceive the spatial information of colors and the ef- + +fect of image content and semantics. In addition, they are suited to qualitative analysis instead of directly quantifying the aesthetics of colors. + +IAA methods, especially leaning-based IAA approaches such as [18, 20, 17], are not specifically designed for ICAA and then lack of prior color knowledge, which result in struggling to achieve a fine-grained perception of the importance of different color spaces, and being susceptible to interference from multiple visual elements (Fig. 1 (c)(d)). Furthermore, to our knowledge, there are no datasets specifically designed for ICAA, which results in most data-driven methods not generalizing well to deal with subjective ICAA tasks. In addition, the community of image quality assessment lacks of a fair ICAA benchmark for reliable comparisons. + +The contributions of our paper lie in: + +- The limits of human perception and pixel-based extraction methods for dominant colors are revealed, which guides us to propose a baseline model, called the Delegate (Deformable Gate) Transformer, to adaptively allocate interest points for dominant colors and simulate human behavior of color space segmentation. +- To verify our method more convincingly, a dataset, named ICAA17K, is exclusively designed for ICAA. Specifically, it is a color-oriented dataset with 17K images and the richest annotations thus far. Furthermore, dedicated preassessment strategy is adopted to alleviate long-tailed distributions of dataset. +- Based on ICAA17K, 15 state-of-the-art baselines (see Table 1) are evaluated, which takes our benchmark as the most complete one for ICAA thus far. Our work, not only gains state-of-the art performance, but more importantly serves as a potential catalyst for promoting large-scale model comparisons in future ICAA research. + +# 2. Related Work + +# 2.1. Image Color Aesthetics Assessment + +Traditional quantization methods. To mitigate the gap between subjective and objective evaluations, early methods usually extracted hand-crafted features from brightness, hue, and saturation or applied color difference formulas. For instance, Yang et al. [11] employed the CIEDE2000 color difference formula to align objective evaluation results with subjective visual perception. Shi et al. [30] utilized hue as the primary color information and proposed the structure and hue similarity model (SHSIM) to assess image color quality. Some other works believed that aesthetically pleasing colors typically conform to certain established rules. O'Donovan et al. [6, 31] studied color compatibility theories and color aesthetics based on color pattern (templates). Cohen et al. [19] evaluated color harmony + +using a color wheel, while Nishiyama et al. [32] applied a color harmony model to evaluate the distributions of hue, lightness, and chroma. + +However, the above-mentioned methods are based on statistical quantitative information of image pixels, ignoring how spatial and semantic content affect color aesthetics. Although these methods can give qualitative analysis results, they cannot quantify the aesthetic differences brought about by a tiny change in color, as shown in Fig. 1. + +Data-driven methods. In the deep learning era, some data-driven methods are proposed to fit aesthetic information in images (Table 1). For example, the works of [18, 20, 28, 29] introduced additional context information, e.g., theme and layout information to improve the performance, while others developed models of personalized image aesthetics assessment (PIAA) [22, 33]. The latest works have explored adopting the transformer structure to extract more powerful aesthetic features [17, 27]. + +Nevertheless, these methods typically extract holistic aesthetic features and lack of prior color knowledge, which take themselves harder to perceive the spatial distribution and composition of different colors in an image, then leading to diffuse attention against perceiving color space. On the other hand, they cannot assign different attention weights based on color importance, which leads to in poor fine-grained perception for color (Fig. 5). + +As revealed in Section I, although images are naturally composed of distinct color spaces, dominant and secondary colors have different effects on color perception. The proposed Delegate Transformer learns to segment color space from dedicated deformable attention rather than static pixel values, and thus captures spatial information of color. Furthermore, different color spaces are assigned different levels of attention by the Delegate Transformer, which exactly matches human behavior for color space segmentation. + +# 2.2. Image Aesthetics Assessment Datasets + +In recent years, several datasets have been developed for IAA, as shown in Table 2. Murray et al. [34] created the AVA dataset, a large-scale IAA dataset containing nearly 255,000 images that has become one of the most popular IAA datasets. However, this dataset does not provide annotations for color. Kong et al. [12] introduced the Aesthetics and Attributes Database (AADB), which includes subjective aesthetic scores and attributes annotated by individual users and provides two binary labels for color harmony. Yu et al. [35] developed the Photo Critique Captioning Dataset (PCCD) to address the problem of generating captions for photos based on their aesthetics, which includes evaluations of the color & lighting attributes. Recently, Fang et al. [8] presented the Smartphone Photography Attribute and Quality (SPAQ) dataset with colorfulness attribute scores. + +The above datasets primarily focus on evaluating the + +
No.ModelYearPub.BasicPlatform
1RAPID [21]2014ACMMMincorporate heterogeneousPython&Lua
2AADB [12]2016ECCVsampling strategy, ranking lossMATLAB
3PAM [22]2017ICCVresidual-based, active learningCaffe
4A-Lamp [23]2017CVPRlayout-aware, multi-patchScipy
5NIMA [18]2017TIPpredict distributionTensorflow
6MPada [24]2018ACMMMattention, multi-patchTensorflow
7MLSP [20]2019CVPRstaged training,multi-level featuresTensorflow
8MT-A [8]2020CVPRmulti-task, high-level semantics interactPyTorch
9BIAA [25]2020TCYBmeta-learning, bilevel optimizationPyTorch
10UIAA [26]2020TIPunified probabilistic formulationMATLAB
11MUSIQ [27]2021ICCVmulti-scale representationPyTorch
12HGCN [28]2021CVPRgraph convolution networksJittor
13TANet [29]2022IJCAIfuse theme adaptivelyPyTorch
14MaxViT [17]2022ECCVmulti-axis attentionPyTorch
15Ours2023ICCVdelegate attention, color space segmentationPyTorch
+ +holistic quality of images and lack of detailed color annotations, with limited color types or combinations. Specifically, these datasets exhibit serious selection bias. e.g., about $50\%$ images of the AVA dataset are "black and white" images, which outnumber other colors by 10 to 100 times, and the PCCD and SPAQ datasets have few images of "pink" and "violet" colors. Therefore, these IAA datasets are not suitable for ICAA tasks and cannot support the generalization of ICAA models well. + +To address the above issue, we try to develop a specialized and color-oriented dataset the first time. To the best of our knowledge, our ICAA17K dataset is the largest, as well as most densely annotated ICAA dataset with a diverse range of color types and image collection devices (see Table 2). It also has a dedicated dataset construction strategy to avoid selection bias and long-tailed distributions, which ensures it a high-quality and robust dataset. + +# 3. Proposed Approach + +Dominant or primary colors tend to draw attention quickly and are often the focal point of visual design. Instead of using traditional methods that rely on the number of pixels to differentiate between dominant and secondary colors, we directly employ attention mechanism to distinguish them in a visual context. First, a sparse, data-related, and deformable attention mechanism are introduced in our baseline model to allocate the attention by setting sparse interest points. Second, the color space segmentation module is utilized to group these interest points into different color spaces, and determine the importance of each space. + +Overall architecture. An overview of the architecture is presented in Fig. 2 (a). Our implementation begins with + +Table 1. Summary of 15 existing representative IAA or PIAA models and the proposed methods. Code: available code links are marked in red, and our code is provided in the supplementary material. + +
DatasetYearPub.ImagesRating
DP Challenge [36]2008ICIP16,509100
CUHK-PQ [37]2011ICCV17,67310
AVA [34]2012CVPR255,530250
AADB [12]2016ECCV10,0005
PCCD [35]2017ICCV4,2357
DPC-Captions [38]2019ACMMM154,3846
SPAQ [8]2020CVPR11,125600
TAD66K [29]2022IJCAI66,3271,200
ICAA17K (Ours)2023ICCV17,7261,500
+ +Table 2. Summary of the related datasets. + +embedding an input image through a $4 \times 4$ convolution operation. The first and second stages consist of two Swin Transformer blocks [39]. At these stages, the feature maps are processed by shifting window attention to locally aggregate aesthetic information. Subsequently, several proposed Delegate Transformer blocks, which form Stage III and IV, are utilized to process these feature maps globally and model the long-range relationships. Finally, these feature points are sent to a dedicated module to divide the color space and then sent to an output head to predict the color aesthetics score. + +# 3.1. Delegate Transformer for Allocating Attention + +Utilizing the projection matrices to generate queries $Q$ , keys $K$ , and values $V$ from the original feature map $X$ , the classical multihead self-attention (MSA) can be calculated + +![](images/30a509a3bb7aa2c24b457d68e857d8b65ba8405d298fa9a5b0c350c54c7209f5.jpg) +(a) Overall Architecture + +![](images/324f949bf5ff669c06fde313683e5391e3c6775a4aa4a616e26af98753ce4007.jpg) +(b) Delegate Attention Module +(c) Delegate Transformer Block + +![](images/a5b6663445b21f4bb1c97b440d9da9dc69a4f34f336b08cb00c704d8dcb036fc.jpg) +Figure 2. The proposed model. (a) Overall architecture. (b) Feature processing flow. The input features are sent to learn the original interest points and adjusted by offsets, then filtered by the gate module. (d) The entire block. (d) The color space segmentation module, which is modeled as a stack of CNN layers. + +![](images/457adef5a92b09d0c99ffe3070eeebea9be91bf487a145979531dd5ae98b278e.jpg) +(d) Color Space Segmentation Module + +as: + +$$ +\mathrm {M S A} = \operatorname {S o f t m a x} \left(\mathbf {Q} ^ {(m)} \mathbf {K} ^ {(m) T} / \sqrt {D}\right) \mathbf {V} ^ {(m) T}, \tag {1} +$$ + +where $D$ is the query/key dimension, $m$ denotes the $m$ -th attention head. However, the calculation based on the entire feature map significantly increases the computational complexity. The output attention map also requires a large number of calculations when sent to the next block for color space segmentation. Instead of the calculation for the entire feature map, in this paper, the corresponding calculation is delegated to interest points for achieving fewer computations and higher flexibility (Fig. 2(b)). + +The feature map $X$ is fed into a lightweight subnetwork $Net_{i}$ to generate the initial interest points. Then, another subnetwork $Net_{o}$ is adopted to calculate the offset from $Q$ to adjust the positions of the interest points, which is formulated as: + +$$ +Q = X W _ {Q}, K = \widetilde {X} W _ {k}, V = \widetilde {X} W _ {v}, w i t h +$$ + +$$ +X _ {i} = \emptyset [ X; G * (\underbrace {\operatorname {N e t} _ {i} (X)} _ {\text {i n t e r s t p o i n t s}} + \underbrace {\operatorname {N e t} _ {o} (Q)}) ], \tag {2} +$$ + +where interest points and offsets are both represented by two-dimensional coordinates that are normalized to the range $[-1, +1]$ , where $(-1, -1)$ represents the top-left corner and $(+1, +1)$ represents the bottom-right corner. A gating module, $G$ , is employed to determine activation of the + +current adjusted interest point coordinates, reducing redundancy further in the space and preventing outliers and overlapping interest points. It can be expressed as: + +$$ +G = \operatorname {S i g m o i d} \left(\operatorname {N e t} _ {g} (X) \cdot \alpha\right), \tag {3} +$$ + +where $\alpha$ is a value large enough to ensure that the gate matrix is either 0 or 1. After obtaining the adjusted and filtered interest points $\widetilde{X}$ , the bilinear interpolation is applied to sample the output feature from $\widetilde{X}$ , which can be represented as + +$$ +\varnothing [ X; (p _ {x}, p _ {y}) ] = \sum_ {(q _ {x}, q _ {y})} M (p _ {x}, p _ {y}) M (q _ {x}, q _ {y}) X (q _ {x}, q _ {y}), \tag {4} +$$ + +where $(p_x, p_y)$ represents one of the coordinates of the interest points, $M(a, b) = \max(0, 1 - |a - b|)$ , and $(q_x, q_y)$ is non-zero on only the 4 integral points in $X$ closest to $(p_x, p_y)$ . Finally, the standard multihead attention mechanism of formula (1) is applied to calculate the output. + +# 3.2. Module for Color Space Segmentation + +The selection of dominant and secondary colors is based on human understanding of the relationship between image content and color, which is highly subjective and leads to uncertainty in the size and segmentation of color spaces. To address this issue, we propose a learning-based approach to determine the size and the segmentation of color spaces. After obtaining the color features of the interest points, 1) + +![](images/fc88593a5f26dc4b45d01096f6ed17a432f0de7c6fd8f7b15010e03297cc3348.jpg) +Figure 3. The proposed ICAA17K dataset. (a) Visualization of images with different color types; (b) quantification of 15 complementary/polychromatic colors; (c) quantification of 15 monochromatic colors; (d) examples in the proposed ICAA17K dataset with the ground truth (and confidence) scores shown below each image. + +![](images/cc9ba2ca66102fc0a5dc1ca4a1fe5e85127535e489dc48618e8f5a4feea3b385.jpg) +(b) + +![](images/329bd31d8bd452db0f8029c151a1ecd8aaa5b299a729010e4d7e31e171b327ed.jpg) +(c) + +![](images/ac10bd085bfa1fcd0779b2774a81a5183f53bb43b98646c65b9ce98c2ba397cd.jpg) + +![](images/912345f7027d8e16417efead5266ace274730cdef96037a66bcc43a4c2b05325.jpg) + +![](images/5051098053cdec903071ed06e3ec1867a4b70eef6c0d77fb950cff303b4d3fbc.jpg) + +![](images/00d6ec8c0eb28a688933aa156118e178889f83b833c1d0dcf78dea25b516b7aa.jpg) +3.66 (0.74) +4.63(0.63) 4.89 +5.46(0.66) + +![](images/6a9be4e5ae2a3269f8e820e1d56a35a43077cdc7f00c0d02e5d9e3f5b3ad278c.jpg) + +![](images/2026a41866eb13b283d42ca7dd39d448ca78cc67412e74ca8b8709bc2a940077.jpg) + +![](images/926c7dfaeb5827647b94a7e77b2ee8061d7160a42c09a5ddb201fdfb71ebf747.jpg) +(0.76) + +![](images/ec70bebb05f6dc34d702d98a8942b5b642e95399f4c374ca6749cb2b246d762c.jpg) + +![](images/fbf03364f7cb87b333816d3297eb73514bfee9e8195cce7e89d2d5835f4a214e.jpg) + +![](images/cb8fc96b94bc8ec7582996bbcc7698816f0daa34f63250d960feed446db441f9.jpg) +(d) + +![](images/0d18ec308a943c77d22b9e1aac09494ee258b6f06ca2a1044e5f3ff2b9f39827.jpg) + +![](images/8e49bcfc7b6e4fcf8395256bdaf18846ccb5905ec48b7159116ecc11309a59f5.jpg) + +![](images/0bc106784b3c2928c7a947f5df6a1e8c883885ca9b9e2bf53350774216accf23.jpg) +(b) +Figure 4. An example of color space segmentation with 4 color spaces $(\kappa = 4)$ . (a) Three sample points. (b) The weights of the three points in each space calculated through formula (5). (c) The weights of each point in each space. + +![](images/fd4007517604df7d0eec30f11c7a000e9444347869e0fbdce0834329f0a9ef70.jpg) + +![](images/40190d9223b99351e8d237fa7cb47e42375c60904a9b87efad9263a712058aaa.jpg) +(c) + +the interest points are grouped into several color spaces ( $\kappa$ pre-defined types). 2) Each color space is mapped to the $[0, +1]$ interval, where each interval can be determined by its widths and centers. 3) By counting the weight distribution of points in each color space, it is easy to distinguish the attention of each color space, thus dividing the dominant and secondary color space. + +However, using fixed bin widths and centers is not differentiable, and the parameters cannot be updated through backpropagation. To resolve this issue, the widths and centers are adaptively generated using two lightweight networks, $Net_{w}$ and $Net_{c}$ , which can be learned during training. To calculate the uncertainty in color space segmentation, our model learns a probability distribution for each point $p_j$ to indicate its weight in $\kappa$ color spaces, denoted as + +$S(p_{j}) = [s_{1}(p_{j}),\dots,s_{\kappa}(p_{j})]$ , the weight $s_i()$ of $p_j$ allocated to the i-th space is determined by the piecewise linear basis function. + +$$ +s _ {i} \left(p _ {j}\right) = \max \left\{0, 1 - w _ {i} \times \left| p _ {j} - c _ {i} \right| \right\}, \tag {5} +$$ + +where $w_{i}$ and $c_{i}$ are the widths and centers of the i-th color space. If an interest point falls into the i-th bin, e.g., the interval of $[c_i - w_i,c_i + w_i]$ , it votes for the i-th bin with a positive weight $s_i(p_j)$ . Fig. 4 shows an example with $\kappa = 4$ color spaces. + +The operation in formula (5) can be designed into a lightweight color space partition block (Fig. 2 (d)), where $p_j - c_i$ is equivalent to convolving the point feature maps with a fixed $1 \times 1$ kernel and a learnable bias $c_i$ . Then, after taking the absolute value, the remaining operation is completed using another learnable weight, a fixed $1 \times 1$ bias, and a ReLU activation function. + +# 4. Proposed Dataset + +# 4.1. Image Collection + +How to avoid selection bias by building a comprehensive ICAA dataset. Selection bias occurs when some types of the intended color, scenes and devices are less likely to be included than others, it can affect the validity and generalizability of training models. To solve this issue, 1) we analyzed the most frequently uploaded colors on the Flickr and COLOURLovers website from 2009 to 2022. Finally, 15 complementary/polychromatic colors and 15 monochromatic colors are selected as the basic color types, which ensures that the dataset includes a diverse range of color types (Fig. 3). 2) To improve the dataset's generalizability across different scenes, the images in ICAA17K cover more than 50 scene categories, e.g., animals, plants, mountains, landscapes, and night scenes. 3) We collected images + +
Metric [ICAA]2014-20192020-2022
RAPID [21]AADB [12]PAM [22]A-Lamp [23]NIMA [18]\( MP_{ada} \) [24]MLSP [20]MT-A [8]BIAA [25]UIAA [26]MUSIQ [27]HGCN [28]TANet [29]MaxViT [17]Ours *1Ours *2
SRCC↑.759.788.796.810.809.810.826.821.820.817.835.826.829.855.873.887
LCC↑.771.793.804.818.815.819.837.830.829.830.841.831.836.874.890.901
Accu↑.836.887.899.908.906.906.921.908.913.910.918.909.903.920.957.965
Epoch↓80606050605050601006080601102002540
+ +Table 3. Comparisons of 15 models on our ICAA17K dataset. We retrained the models for the best performance with the recommended parameter and dataset settings. See subsection 5.1 for details regarding the metrics. "Ours *1" indicates that, similar to other methods, only the color annotation scores from the ICAA17K dataset are used during training. "Ours *2" indicates that the holistic aesthetics score is jointly predicted when using multiple tasks. + +
Metric [SPAQ]2014-20192020-2022
RAPID [21]AADB [12]PAM [22]A-Lamp [23]NIMA [18]\( MP_{ada} \) [24]MLSP [20]MT-A [8]BIAA [25]UIAA [26]MUSIQ [27]HGCN [28]TANet [29]MaxViT [17]Ours * 1
SRCC↑.721.725.732.747.730.746.753.751.748.750.753.755.750.753.770
LCC↑.730.739.744.760.746.757.773.763.759.763.770.771.761.770.791
Epoch↓95757065556555308570904511015025
+ +Table 4. Comparisons of 15 models on the SPAQ dataset [8]. All models used only the colorfulness score in the dataset for training, and were retrained for the best performance using the recommended parameters and dataset settings. + +taken by more than 80 different types of devices, including mobile phones (e.g., iPhone X, Huawei Mate 10, and Samsung Galaxy S10) and digital cameras (e.g., Canon, Nikon, Sony, Panasonic), to encompass a wide range of devices. + +How to alleviate long-tailed distributions by performing a preassessment strategy. The annotated label in the most aesthetic datasets have long-tailed distributions [34, 8], which bias models trained on imbalanced data toward majority examples. To avoid correcting the long-tailed distributions after label annotation, we propose a preassessment strategy during data collection to maximize label diversity. + +First, the color harmony [19] is applied to evaluate objective color quality by measuring the distance between the tested image and templates on the hue wheel. Given a harmonic template $T(\beta)$ (one of the predefined type of templates $T$ ) with an associated orientation $\beta$ , the color harmony $f_{obj}$ of a tested image $C$ could be calculated as follows: + +$$ +f _ {o b j} (C) = \sum_ {l \in C} \| H (l) - E _ {T (\beta)} (l) \| \cdot Y (l), \tag {6} +$$ + +where $l$ is the hue of a pixel, $E_{T(\beta)}$ is the sector border hue of harmonic template $T(\beta)$ , $H$ and $Y$ mean the hue and the saturation channels, respectively, $\|\cdot\|$ denotes the arc-length distance on the hue wheel (measured in radians). + +Second, we crawl images from Flickr and record their metadata, including "views" (number of visits $VI$ ) and "faves" (number of clicks that favor image $FA$ ). Intuitively, an image with a higher ratio of faves to views is more pop- + +ular and beautiful. Specifically, considering the affect of user's herd mentality, we adopt the following formula to approximate the subjective perception: + +$$ +f _ {s u b j} (C) = \frac {F A (C)}{\sqrt {V I (C)}}. \tag {7} +$$ + +Finally, we normalize the $f_{subj}$ and $f_{obj}$ , and add them together as the final preassessment score of the image (not shown to annotators), and extract balanced samples from each score range for the next stage of manual evaluation. + +# 4.2. Image Annotation + +How to avoid unreliablely by improving annotation quality. The traditional annotation process suffers from a lack of a baseline or reference for evaluating aesthetics, which makes subjective responses unreliable in the early stages. To mitigate this problem, 1) we first provided annotators simple guidelines based on photography theories to guide their judgment of image color aesthetics. 2) We provided anchor images with different rating scales as references. Experts rated the anchor images based on their rich photography experience, allowing annotators to understand the differences between rating scales. Before annotating, annotators had to review the anchor images and guidelines thoroughly. The average value was then calculated as the ground truth. 3) Annotators could provide a confidence score (from 0 to 1) to indicate their certainty in each rating. This information can be useful for studying the noise and subjectivity in the dataset. 4) Finally, each image has + +![](images/689db2d5d58389a81be6a6170f33d7557dd7823f72c581f3a1f44d26689e2a94.jpg) +Figure 5. Saliency maps of the 15 models. The attention of our method is segmented into distinct regions based on the color distribution. + +![](images/abad8b514bc70e63ac3fc6ffd9aae3c3ef8a0efab3b5f454756d67d6dbfecaec.jpg) +Figure 6. Analysis of kernel sizes in $Net_{i}$ and the number of color spaces $\kappa$ . Larger kernel sizes result in fewer interest points. + +an annotation score from 1 to 10, which represents color aesthetic perception from lowest to highest, we selected images with an average confidence score above 0.5 as samples in the dataset. We collected about 1,500 opinions for each image. Some examples are shown in Fig. 3 (d). + +# 5. Experiments + +# 5.1. Experimental Settings + +Benchmark Datasets. We performed model evaluations on two datasets, the proposed ICAA17K and SPAQ datasets [8]. Although the SPAQ dataset focuses on image quality rather than aesthetics, it provides an annotation on colorfulness, which can also afford testing the ability of various baseline models to extract color features to some extent. + +Benchmark Models and Training Protocols. To the best of our knowledge, there is no publicly available learning-based ICAA model. Therefore, we selected 14 deep learn + +ing baselines (Table. 1) according to the following criteria: 1) classical architectures with available code and 2) SOTA in a specific field, e.g., personalized IAA. These baselines are trained with the recommended parameter settings (e.g., optimizer and batch size), using the same training and testing settings. The mean squared error (MSE) is adopted as the loss function across all methods. + +Evaluation Metrics. We adopt two popular evaluation metrics, the Spearman's rank correlation coefficient (SRCC) and the linear correlation coefficient (LCC), to evaluate the model's ability. The metrics for the ICAA17K dataset also include binary classification accuracy (color aesthetically negative or positive). Since the ground-truth annotations in the ICAA17K dataset are based on a 5-point boundary line, if an annotator tends to give a score greater than 5, it is considered aesthetically positive; otherwise, it is considered negative. Therefore, this metric can evaluate the model's ability to perform coarse-grained evaluations. + +# 5.2. Results and Dataset Analysis + +ICAA17K Dataset. Table 3 lists the results of 15 models on the ICAA dataset. The proposed baseline method (marked by "Ours *1") achieves the best performance on all metrics and has a higher training speed. This is mainly due to its sparse attention and adaptive color space partitioning. Furthermore, our ICAA17K dataset also provides additional annotations regarding holistic aesthetics. Therefore, we incorporate this additional supervision information into the multitask training (marked by "Ours *2"), which improves performance but increases the training time. + +SPAQ Dataset. Based on the results of the 15 methods reported in Table 4, our method still achieves the top performance on all metrics with fewer training epochs, showing that it is a promising solution for the IQA problem. + +![](images/b308b498a137da42536feeaeb0ca2401664b9205abff113abb75053adcb5947c.jpg) +Figure 7. Visualization of our attention mechanism. (a) Sampled interest points adjusted by offsets. (b) Interest points filtered by the gating module. (c) The radius of the circles represents the attention weights at multiple heads, with a larger radius indicating a higher weight. (d) The color space segmentation of each point, where the numbers represent the primary and secondary relationship of the color space, e.g., "1" represents the dominant color. + +![](images/ea74892bb8a1944a3a5caa49c8434988ed8bd79d74b666bb183a1ac9de88fd52.jpg) + +Dataset Analysis. Compared to the SPAQ dataset, when using the ICAA17K dataset for training, most networks converge faster and have higher SRCC and LCC metrics, which indicates that models trained on the ICAA17K dataset more accurately fit human subjective color perception. It also suggests that introducing more annotators (SPAQ 600 vs ICAA17K 1500) and providing more annotation references for annotators can alleviate noise caused by subjective annotation, which is a promising direction toward improving the quality of IQA or IAA datasets. + +# 5.3. Ablation Study + +Delegate vs Shift/Deformable Transformers. In Table 5, we replace the Delegate Transformer blocks with the shift attention mechanism of Swin Transformer [39] and deformable attention of DAT Transformer [40] in the last stage. In this case, our model has better performance and lower computation, which outperforms Swin by $5.2\%$ in the SRCC with $41.7\%$ FLOPs and outperforms DAT by $4.4\%$ in the SRCC with $41.7\%$ FLOPs. Additionally, the stacking way also has a significant impact on the performance. The Delegate Transformer blocks may result in information loss if it is employed in the early stages due to the sparse attention mechanism. + +The Number of Interest Points and Color Space. The number of interest points mainly depends on the kernel size of the subnetwork $Net_{i}$ . In Fig. 6, the results show that an inadequate or excessive number of interest points leads to degradation in the model's performance. Moreover, selecting the color space hyperparameter $\kappa$ is also crucial for the performance of our model. Our experiments reveal that the optimal performance is achieved when $\kappa = 7$ . Additionally, a larger value of $\kappa$ leads to an increase in computation and memory requirements. In contrast, a + +
#StagesFLOPs↓SRCC↑LCC↑
1234
1SSSS240G0.8300.843
2SSDD240G0.8360.848
3SSGG140G0.8730.890
4GGSS210G0.8280.835
5SGGS140G0.8380.850
6GSSG210G0.8250.833
7GGGG130G0.8370.846
+ +Table 5. Ablations on the application of different attention mechanisms in different stages, where $S$ , $D$ and $G$ represent the application of Swin, DAT and our blocks, respectively. + +smaller value of $\kappa$ may not provide a sufficient characterization of the color composition in the input image. Moreover, the effectiveness of different modules are shown in the Appendix. + +# 5.4. Visualization + +The GradCAM [41] method is applied to visualize the saliency maps in Fig. 5. Our model differs from other methods because it segments the image into distinct regions based on the color distribution, and the region that typically receives the greatest attention is the region where the dominant color is located. The reason is further explained in Fig. 7. First, our model generates interest points via $\text{Net}_i$ and adjusts their positions by $\text{Net}_o$ . Second, some redundant points are eliminated via $\text{Net}_g$ and weights are calculated using multihead attention. Third, the color space segmentation module and formula (5) are employed to divide the different spaces and direct different attention toward them. + +# 6. Conclusions + +In this paper, a comprehensive study on ICAA has been conducted by developing a Delegate Transformer model as a baseline, constructing a color-oriented dataset benchmark, and establishing the largest-scale benchmarks to date. Compared with traditional or data-driven models, our model achieves SOTA performance and produces visually consistent results. We hope our work should contribute to the community's research on ICAA. However, ICAA remains far from being solved, we will optimize our approach by incorporating more color prior knowledge in the future. + +# 7. Acknowledgments + +This work was supported by the Funds for Creative Research Groups of China under Grant 61921003, the national key R & D program intergovernmental international science and technology innovation cooperation project 2021YFE0101600. + +# References + +[1] Ahmed Talib, Massudi Mahmuddin, Husniza Husni, and Loay E George. A weighted dominant color descriptor for content-based image retrieval. Journal of Visual Communication and Image Representation, 24(3):345-360, 2013. +[2] Yining Deng, BS Manjunath, Charles Kenney, Michael S Moore, and Hyundoo Shin. An efficient color representation for image retrieval. IEEE Transactions on image processing, 10(1):140-147, 2001. +[3] Bangalore S Manjunath, J-R Ohm, Vinod V Vasudevan, and Akio Yamada. Color and texture descriptors. IEEE Transactions on circuits and systems for video technology, 11(6):703-715, 2001. +[4] Luca Marchesotti, Naila Murray, and Florent Perronnin. Discovering beautiful attributes for aesthetic image analysis. International journal of computer vision, 113:246-266, 2015. +[5] Andrew J Elliot and Markus A Maier. Color psychology: Effects of perceiving color on psychological functioning in humans. Annual review of psychology, 65:95-120, 2014. +[6] Peter O'Donovan, Aseem Agarwala, and Aaron Hertzmann. Color compatibility from large datasets. In ACM SIGGRAPH 2011 papers, pages 1-12. 2011. +[7] Joon-Young Lee, Kalyan Sunkavalli, Zhe Lin, Xiaohui Shen, and In So Kweon. Automatic content-aware color and tone stylization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2470–2478, 2016. +[8] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In CVPR, pages 3677-3686, 2020. +[9] Emilie Baudin, François-Xavier Bucher, Laurent Chanas, and Frédéric Guichard. Dxomark objective video quality measurements. *Electronic Imaging*, 2020(9):166-1, 2020. +[10] Sabine E Susstrunk and Stefan Winkler. Color image quality on the internet. In _Internet imaging V_, volume 5304, pages 118-131. SPIE, 2003. +[11] Yang Yang, Jun Ming, and Nenghai Yu. Color image quality assessment based on ciede2000. Advances in Multimedia, 2012, 2012. +[12] Shu Kong, Xiaohui Shen, Zhe Lin, Radomir Mech, and Charless Fowlkes. Photo aesthetics ranking network with attributes and content adaptation. In ECCV, pages 662-679. Springer, 2016. +[13] Xuan Xu and Xin Li. Scan: Spatial color attention networks for real single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. +[14] Aleksandra Mojsilovic, H Hu, and Emina Soljanin. Extraction of perceptually important colors and similarity measurement for image matching, retrieval and analysis. IEEE Transactions on Image Processing, 11(11):1238-1248, 2002. +[15] Stephen E Palmer, Karen B Schloss, and Jonathan Sammartino. Visual aesthetics and human preference. Annual review of psychology, 64:77-107, 2013. + +[16] Michael J Swain and Dana H Ballard. Color indexing. International journal of computer vision, 7(1):11-32, 1991. +[17] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. ECCV, 2022. +[18] Hossein Talebi and Peyman Milanfar. NIMA: Neural image assessment. TIP, 27(8):3998-4011, 2018. +[19] Daniel Cohen-Or, Olga Sorkine, Ran Gal, Tommer Leyvand, and Ying-Qing Xu. Color harmonization. In ACM SIGGRAPH 2006 Papers, pages 624-630. 2006. +[20] Vlad Hosu, Bastian Goldlucke, and Dietmar Saupe. Effective aesthetics prediction with multi-level spatially pooled features. In CVPR, 2019. +[21] Xin Lu, Zhe Lin, Hailin Jin, Jianchao Yang, and James Z. Wang. RAPID: Rating pictorial aesthetics using deep learning. In ACMMM, pages 457-466, 2014. +[22] Jian Ren, Xiaohui Shen, Zhe Lin, Radomir Mech, and David J Foran. Personalized image aesthetics. In ICCV, pages 638-647, 2017. +[23] Shuang Ma, Jing Liu, and Chang Wen Chen. A-lamp: Adaptive layout-aware multi-patch deep convolutional neural network for photo aesthetic assessment. In CVPR, pages 722-731, 2017. +[24] Kekai Sheng, Weiming Dong, Chongyang Ma, Xing Mei, Feiyue Huang, and Bao-Gang Hu. Attention-based multipatch aggregation for image aesthetic assessment. In ACMMM, 2018. +[25] Hancheng Zhu, Leida Li, Jinjian Wu, Sicheng Zhao, Guiguang Ding, and Guangming Shi. Personalized image aesthetics assessment via meta-learning with bilevel gradient optimization. TCYB, 2020. +[26] Hui Zeng, Zisheng Cao, Lei Zhang, and Alan C Bovik. A unified probabilistic formulation of image aesthetic assessment. TIP, 29:1548-1561, 2019. +[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, pages 5148-5157, 2021. +[28] Dongyu She, Yu-Kun Lai, Gaoxiong Yi, and Kun Xu. Hierarchical layout-aware graph convolutional network for unified aesthetics assessment. In CVPR, pages 8475-8484, 2021. +[29] Shuai He, Yongchang Zhang, Rui Xie, Dongxiang Jiang, and Anlong Ming. Rethinking image aesthetics assessment: Models, datasets and benchmarks. *IJCAI*, 2022. +[30] Yunyu Shi, Youdong Ding, Ranran Zhang, and Jun Li. Structure and hue similarity for color image quality assessment. In 2009 International Conference on Electronic Computer Technology, pages 329-333. IEEE, 2009. +[31] Peter O'Donovan, Aseem Agarwala, and Aaron Hertzmann. Collaborative filtering of color aesthetics. In Proceedings of the Workshop on Computational Aesthetics, pages 33-40, 2014. +[32] Masashi Nishiyama, Takahiro Okabe, Imari Sato, and Yoichi Sato. Aesthetic quality classification of photographs based on color harmony. In CVPR 2011, pages 33-40. IEEE, 2011. + +[33] Pei Lv, Jianqi Fan, Xixi Nie, Weiming Dong, Xiaoheng Jiang, Bing Zhou, Mingliang Xu, and Changsheng Xu. User-guided personalized image aesthetic assessment based on deep reinforcement learning. arXiv:2106.07488, 2021. +[34] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In CVPR. IEEE, 2012. +[35] Kuang-Yu Chang, Kung-Hung Lu, and Chu-Song Chen. Aesthetic critiques generation for photos. In ICCV, pages 3514-3523, 2017. +[36] Ritendra Datta, Jia Li, and James Z Wang. Algorithmic infor-ferencing of aesthetics and emotion in natural images: An exposition. In ICIP. IEEE, 2008. +[37] Wei Luo, Xiaogang Wang, and Xiaou Tang. Content-based photo quality assessment. In ICCV, pages 2206-2213. IEEE, 2011. +[38] Xin Jin, Le Wu, Geng Zhao, Xiaodong Li, Xiaokun Zhang, Shiming Ge, Dongqing Zou, Bin Zhou, and Xinghui Zhou. Aesthetic attributes assessment of images. In ACMMM, pages 311-319, 2019. +[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv:2103.14030, 2021. +[40] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4794-4803, 2022. +[41] Jacob Gildenblat and contributors. Pytorch library for cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021. \ No newline at end of file diff --git a/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/images.zip b/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e3e941bce251d37cbd24a8b5dc12ad4620452ad9 --- /dev/null +++ b/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84a2176a8d5b44399dfdf41d53bd711af7b5a65536b81dd1ff7efb551c011e63 +size 841163 diff --git a/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/layout.json b/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a9ab234fc3b48e8b63332454c94ef22875382f38 --- /dev/null +++ b/2023/Thinking Image Color Aesthetics Assessment_ Models, Datasets and Benchmarks/layout.json @@ -0,0 +1,8788 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 547, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 547, + 122 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 547, + 122 + ], + "type": "text", + "content": "Thinking Image Color Aesthetics Assessment: Models, Datasets and Benchmarks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 143, + 510, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 143, + 510, + 172 + ], + "spans": [ + { + "bbox": [ + 85, + 143, + 510, + 172 + ], + "type": "text", + "content": "Shuai He† Anlong Ming†,* Yaqi Li Jinyuan Sun ShunTian Zheng Huadong Ma Beijing University of Posts and Telecommunications" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 173, + 492, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 173, + 492, + 186 + ], + "spans": [ + { + "bbox": [ + 102, + 173, + 492, + 186 + ], + "type": "text", + "content": "{hs19951021, mal, lyq645146943, jinyuan, zhengshuntian, mhd}@bupt.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 187, + 471, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 187, + 471, + 198 + ], + "spans": [ + { + "bbox": [ + 121, + 187, + 471, + 198 + ], + "type": "text", + "content": "https://github.com/woshidandan/Image-Color-Aesthetics-Assessment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 252, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 515 + ], + "type": "text", + "content": "We present a comprehensive study on a new task named image color aesthetics assessment (ICAA), which aims to assess color aesthetics based on human perception. ICAA is important for various applications such as imaging measurement and image analysis. However, due to the highly diverse aesthetic preferences and numerous color combinations, ICAA presents more challenges than conventional image quality assessment tasks. To advance ICAA research, 1) we propose a baseline model called the Delegate Transformer, which not only deploys deformable transformers to adaptively allocate interest points, but also learns human color space segmentation behavior by the dedicated module. 2) We elaborately build a color-oriented dataset, ICAA17K, containing 17K images, covering 30 popular color combinations, 80 devices and 50 scenes, with each image densely annotated by more than 1,500 people. Moreover, we develop a large-scale benchmark of 15 methods, the most comprehensive one thus far based on two datasets, SPAQ and ICAA17K. Our work, not only achieves state-of-the-art performance, but more importantly offers the community a roadmap to explore solutions for ICAA. Code and dataset are available in here." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 128, + 565 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 573, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 287, + 693 + ], + "type": "text", + "content": "Color is known as having a higher degree of discriminability and correlation compared to other visual features [1, 2, 3], because the human eye can directly perceive light of different wavelengths and convert them into different color signals. As digital photography expands rapidly, ICAA has become one of the most important criteria to automatically assess whether the image meets users' aesthetic preferences [4, 5, 6, 7]. It is also an essential step in imaging measurements among manufacturers to evaluate the performance of smartphones and cameras [8, 9, 10]." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 226, + 545, + 323 + ], + "blocks": [ + { + "bbox": [ + 309, + 226, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 309, + 226, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 309, + 226, + 545, + 323 + ], + "type": "image", + "image_path": "332ce92dfe3feb0f9441c4d51d6d75ae2f5714a88c9ab2d3b39afb151f2a3599.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 325, + 547, + 403 + ], + "lines": [ + { + "bbox": [ + 305, + 325, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 547, + 403 + ], + "type": "text", + "content": "Figure 1. The results of different methods on two similar images with tiny color adjustment, ground-truth (red) and predicted (black) scores are shown below. (a) Similar images. (b) Color histogram [16] is unable to differentiate and quantify the color aesthetics. The saliency maps of (c) MaxViT [17], (d) NIMA [18] and (e) our method, whereas our method achieves a fine-grained perception of color aesthetics." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 424, + 547, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 547, + 509 + ], + "type": "text", + "content": "Although ICAA is an important branch of IAA (image aesthetics assessment), they deal with different tasks. IAA evaluates the holistic aesthetics of an image, which implicitly depends on color and other attributes (brightness, sharpness, etc.) [11, 12, 8]. ICAA focuses more on evaluating the impact of color harmony, color combinations and other factors on color perception." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 509, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 640 + ], + "type": "text", + "content": "ICAA is confronted with two major challenges. First, the types and combinations of colors are intrinsically complex, thus the color aesthetics are determined by specific colors and their relative positions in color spaces [13]. However, biological studies show that humans can perceive only up to 8 colors due to the human eye and perception limits [14]. Therefore, extracting dominant colors from an image is a crucial process and important prior knowledge [6, 15]. Second, human color preferences are subjective and vary from person to person based on factors such as age and culture that may change over time." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Most traditional quantization methods for ICAA rely on objective rules such as color histogram [16] or color wheel [19] theories. However, the color information obtained from pixel statistics is limited to capture the nuances of color aesthetics (Fig. 1 (b)), mainly due to these methods cannot perceive the spatial information of colors and the ef-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 702, + 189, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 702, + 189, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 702, + 189, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 62, + 702, + 189, + 713 + ], + "type": "text", + "content": " Equal Contribution; \\*Contact Author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21838" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "fect of image content and semantics. In addition, they are suited to qualitative analysis instead of directly quantifying the aesthetics of colors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 251 + ], + "type": "text", + "content": "IAA methods, especially leaning-based IAA approaches such as [18, 20, 17], are not specifically designed for ICAA and then lack of prior color knowledge, which result in struggling to achieve a fine-grained perception of the importance of different color spaces, and being susceptible to interference from multiple visual elements (Fig. 1 (c)(d)). Furthermore, to our knowledge, there are no datasets specifically designed for ICAA, which results in most data-driven methods not generalizing well to deal with subjective ICAA tasks. In addition, the community of image quality assessment lacks of a fair ICAA benchmark for reliable comparisons." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 251, + 209, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 251, + 209, + 263 + ], + "spans": [ + { + "bbox": [ + 59, + 251, + 209, + 263 + ], + "type": "text", + "content": "The contributions of our paper lie in:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 266, + 287, + 488 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 266, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 287, + 339 + ], + "type": "text", + "content": "- The limits of human perception and pixel-based extraction methods for dominant colors are revealed, which guides us to propose a baseline model, called the Delegate (Deformable Gate) Transformer, to adaptively allocate interest points for dominant colors and simulate human behavior of color space segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 342, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 287, + 413 + ], + "type": "text", + "content": "- To verify our method more convincingly, a dataset, named ICAA17K, is exclusively designed for ICAA. Specifically, it is a color-oriented dataset with 17K images and the richest annotations thus far. Furthermore, dedicated preassessment strategy is adopted to alleviate long-tailed distributions of dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 417, + 287, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 417, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 47, + 417, + 287, + 488 + ], + "type": "text", + "content": "- Based on ICAA17K, 15 state-of-the-art baselines (see Table 1) are evaluated, which takes our benchmark as the most complete one for ICAA thus far. Our work, not only gains state-of-the art performance, but more importantly serves as a potential catalyst for promoting large-scale model comparisons in future ICAA research." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 508, + 134, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 134, + 520 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 134, + 520 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 527, + 235, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 235, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 235, + 540 + ], + "type": "text", + "content": "2.1. Image Color Aesthetics Assessment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": "Traditional quantization methods. To mitigate the gap between subjective and objective evaluations, early methods usually extracted hand-crafted features from brightness, hue, and saturation or applied color difference formulas. For instance, Yang et al. [11] employed the CIEDE2000 color difference formula to align objective evaluation results with subjective visual perception. Shi et al. [30] utilized hue as the primary color information and proposed the structure and hue similarity model (SHSIM) to assess image color quality. Some other works believed that aesthetically pleasing colors typically conform to certain established rules. O'Donovan et al. [6, 31] studied color compatibility theories and color aesthetics based on color pattern (templates). Cohen et al. [19] evaluated color harmony" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": "using a color wheel, while Nishiyama et al. [32] applied a color harmony model to evaluate the distributions of hue, lightness, and chroma." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 109, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 180 + ], + "type": "text", + "content": "However, the above-mentioned methods are based on statistical quantitative information of image pixels, ignoring how spatial and semantic content affect color aesthetics. Although these methods can give qualitative analysis results, they cannot quantify the aesthetic differences brought about by a tiny change in color, as shown in Fig. 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 545, + 289 + ], + "type": "text", + "content": "Data-driven methods. In the deep learning era, some data-driven methods are proposed to fit aesthetic information in images (Table 1). For example, the works of [18, 20, 28, 29] introduced additional context information, e.g., theme and layout information to improve the performance, while others developed models of personalized image aesthetics assessment (PIAA) [22, 33]. The latest works have explored adopting the transformer structure to extract more powerful aesthetic features [17, 27]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 289, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 385 + ], + "type": "text", + "content": "Nevertheless, these methods typically extract holistic aesthetic features and lack of prior color knowledge, which take themselves harder to perceive the spatial distribution and composition of different colors in an image, then leading to diffuse attention against perceiving color space. On the other hand, they cannot assign different attention weights based on color importance, which leads to in poor fine-grained perception for color (Fig. 5)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 386, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 493 + ], + "type": "text", + "content": "As revealed in Section I, although images are naturally composed of distinct color spaces, dominant and secondary colors have different effects on color perception. The proposed Delegate Transformer learns to segment color space from dedicated deformable attention rather than static pixel values, and thus captures spatial information of color. Furthermore, different color spaces are assigned different levels of attention by the Delegate Transformer, which exactly matches human behavior for color space segmentation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 502, + 506, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 506, + 514 + ], + "type": "text", + "content": "2.2. Image Aesthetics Assessment Datasets" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 700 + ], + "type": "text", + "content": "In recent years, several datasets have been developed for IAA, as shown in Table 2. Murray et al. [34] created the AVA dataset, a large-scale IAA dataset containing nearly 255,000 images that has become one of the most popular IAA datasets. However, this dataset does not provide annotations for color. Kong et al. [12] introduced the Aesthetics and Attributes Database (AADB), which includes subjective aesthetic scores and attributes annotated by individual users and provides two binary labels for color harmony. Yu et al. [35] developed the Photo Critique Captioning Dataset (PCCD) to address the problem of generating captions for photos based on their aesthetics, which includes evaluations of the color & lighting attributes. Recently, Fang et al. [8] presented the Smartphone Photography Attribute and Quality (SPAQ) dataset with colorfulness attribute scores." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "The above datasets primarily focus on evaluating the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "21839" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 544, + 275 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 544, + 275 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 544, + 275 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 544, + 275 + ], + "type": "table", + "html": "
No.ModelYearPub.BasicPlatform
1RAPID [21]2014ACMMMincorporate heterogeneousPython&Lua
2AADB [12]2016ECCVsampling strategy, ranking lossMATLAB
3PAM [22]2017ICCVresidual-based, active learningCaffe
4A-Lamp [23]2017CVPRlayout-aware, multi-patchScipy
5NIMA [18]2017TIPpredict distributionTensorflow
6MPada [24]2018ACMMMattention, multi-patchTensorflow
7MLSP [20]2019CVPRstaged training,multi-level featuresTensorflow
8MT-A [8]2020CVPRmulti-task, high-level semantics interactPyTorch
9BIAA [25]2020TCYBmeta-learning, bilevel optimizationPyTorch
10UIAA [26]2020TIPunified probabilistic formulationMATLAB
11MUSIQ [27]2021ICCVmulti-scale representationPyTorch
12HGCN [28]2021CVPRgraph convolution networksJittor
13TANet [29]2022IJCAIfuse theme adaptivelyPyTorch
14MaxViT [17]2022ECCVmulti-axis attentionPyTorch
15Ours2023ICCVdelegate attention, color space segmentationPyTorch
", + "image_path": "a06d686bb95f9d1529fa9c9b15cdd58600999bf69ca45e1904e443daf0a998e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 325, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 432 + ], + "type": "text", + "content": "holistic quality of images and lack of detailed color annotations, with limited color types or combinations. Specifically, these datasets exhibit serious selection bias. e.g., about " + }, + { + "bbox": [ + 46, + 325, + 287, + 432 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 325, + 287, + 432 + ], + "type": "text", + "content": " images of the AVA dataset are \"black and white\" images, which outnumber other colors by 10 to 100 times, and the PCCD and SPAQ datasets have few images of \"pink\" and \"violet\" colors. Therefore, these IAA datasets are not suitable for ICAA tasks and cannot support the generalization of ICAA models well." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 433, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 529 + ], + "type": "text", + "content": "To address the above issue, we try to develop a specialized and color-oriented dataset the first time. To the best of our knowledge, our ICAA17K dataset is the largest, as well as most densely annotated ICAA dataset with a diverse range of color types and image collection devices (see Table 2). It also has a dedicated dataset construction strategy to avoid selection bias and long-tailed distributions, which ensures it a high-quality and robust dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 538, + 164, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 164, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 164, + 552 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 689 + ], + "type": "text", + "content": "Dominant or primary colors tend to draw attention quickly and are often the focal point of visual design. Instead of using traditional methods that rely on the number of pixels to differentiate between dominant and secondary colors, we directly employ attention mechanism to distinguish them in a visual context. First, a sparse, data-related, and deformable attention mechanism are introduced in our baseline model to allocate the attention by setting sparse interest points. Second, the color space segmentation module is utilized to group these interest points into different color spaces, and determine the importance of each space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Overall architecture. An overview of the architecture is presented in Fig. 2 (a). Our implementation begins with" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 323, + 541, + 469 + ], + "blocks": [ + { + "bbox": [ + 46, + 283, + 546, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 283, + 546, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 546, + 308 + ], + "type": "text", + "content": "Table 1. Summary of 15 existing representative IAA or PIAA models and the proposed methods. Code: available code links are marked in red, and our code is provided in the supplementary material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 323, + 541, + 469 + ], + "lines": [ + { + "bbox": [ + 310, + 323, + 541, + 469 + ], + "spans": [ + { + "bbox": [ + 310, + 323, + 541, + 469 + ], + "type": "table", + "html": "
DatasetYearPub.ImagesRating
DP Challenge [36]2008ICIP16,509100
CUHK-PQ [37]2011ICCV17,67310
AVA [34]2012CVPR255,530250
AADB [12]2016ECCV10,0005
PCCD [35]2017ICCV4,2357
DPC-Captions [38]2019ACMMM154,3846
SPAQ [8]2020CVPR11,125600
TAD66K [29]2022IJCAI66,3271,200
ICAA17K (Ours)2023ICCV17,7261,500
", + "image_path": "655cbee5905f1a2ca3ad6e6080fa46cae90685fc71360460485701d2e7a72767.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 347, + 476, + 501, + 487 + ], + "lines": [ + { + "bbox": [ + 347, + 476, + 501, + 487 + ], + "spans": [ + { + "bbox": [ + 347, + 476, + 501, + 487 + ], + "type": "text", + "content": "Table 2. Summary of the related datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 512, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 545, + 643 + ], + "type": "text", + "content": "embedding an input image through a " + }, + { + "bbox": [ + 304, + 512, + 545, + 643 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 304, + 512, + 545, + 643 + ], + "type": "text", + "content": " convolution operation. The first and second stages consist of two Swin Transformer blocks [39]. At these stages, the feature maps are processed by shifting window attention to locally aggregate aesthetic information. Subsequently, several proposed Delegate Transformer blocks, which form Stage III and IV, are utilized to process these feature maps globally and model the long-range relationships. Finally, these feature points are sent to a dedicated module to divide the color space and then sent to an output head to predict the color aesthetics score." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 657, + 545, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 545, + 670 + ], + "type": "text", + "content": "3.1. Delegate Transformer for Allocating Attention" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": "Utilizing the projection matrices to generate queries " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": ", keys " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": ", and values " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": " from the original feature map " + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 677, + 546, + 713 + ], + "type": "text", + "content": ", the classical multihead self-attention (MSA) can be calculated" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "21840" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 71, + 521, + 166 + ], + "blocks": [ + { + "bbox": [ + 75, + 71, + 521, + 166 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 521, + 166 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 521, + 166 + ], + "type": "image", + "image_path": "30a509a3bb7aa2c24b457d68e857d8b65ba8405d298fa9a5b0c350c54c7209f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 166, + 335, + 174 + ], + "lines": [ + { + "bbox": [ + 261, + 166, + 335, + 174 + ], + "spans": [ + { + "bbox": [ + 261, + 166, + 335, + 174 + ], + "type": "text", + "content": "(a) Overall Architecture" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 73, + 174, + 334, + 299 + ], + "blocks": [ + { + "bbox": [ + 73, + 174, + 334, + 299 + ], + "lines": [ + { + "bbox": [ + 73, + 174, + 334, + 299 + ], + "spans": [ + { + "bbox": [ + 73, + 174, + 334, + 299 + ], + "type": "image", + "image_path": "324f949bf5ff669c06fde313683e5391e3c6775a4aa4a616e26af98753ce4007.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 300, + 252, + 311 + ], + "lines": [ + { + "bbox": [ + 156, + 300, + 252, + 311 + ], + "spans": [ + { + "bbox": [ + 156, + 300, + 252, + 311 + ], + "type": "text", + "content": "(b) Delegate Attention Module" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 333, + 300, + 432, + 311 + ], + "lines": [ + { + "bbox": [ + 333, + 300, + 432, + 311 + ], + "spans": [ + { + "bbox": [ + 333, + 300, + 432, + 311 + ], + "type": "text", + "content": "(c) Delegate Transformer Block" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 342, + 171, + 432, + 297 + ], + "blocks": [ + { + "bbox": [ + 342, + 171, + 432, + 297 + ], + "lines": [ + { + "bbox": [ + 342, + 171, + 432, + 297 + ], + "spans": [ + { + "bbox": [ + 342, + 171, + 432, + 297 + ], + "type": "image", + "image_path": "a5b6663445b21f4bb1c97b440d9da9dc69a4f34f336b08cb00c704d8dcb036fc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 318, + 547, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 318, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 547, + 352 + ], + "type": "text", + "content": "Figure 2. The proposed model. (a) Overall architecture. (b) Feature processing flow. The input features are sent to learn the original interest points and adjusted by offsets, then filtered by the gate module. (d) The entire block. (d) The color space segmentation module, which is modeled as a stack of CNN layers." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 439, + 172, + 520, + 297 + ], + "blocks": [ + { + "bbox": [ + 439, + 172, + 520, + 297 + ], + "lines": [ + { + "bbox": [ + 439, + 172, + 520, + 297 + ], + "spans": [ + { + "bbox": [ + 439, + 172, + 520, + 297 + ], + "type": "image", + "image_path": "457adef5a92b09d0c99ffe3070eeebea9be91bf487a145979531dd5ae98b278e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 297, + 514, + 315 + ], + "lines": [ + { + "bbox": [ + 443, + 297, + 514, + 315 + ], + "spans": [ + { + "bbox": [ + 443, + 297, + 514, + 315 + ], + "type": "text", + "content": "(d) Color Space Segmentation Module" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 373, + 61, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 61, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 61, + 381 + ], + "type": "text", + "content": "as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 391, + 287, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 391, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 64, + 391, + 287, + 411 + ], + "type": "interline_equation", + "content": "\\mathrm {M S A} = \\operatorname {S o f t m a x} \\left(\\mathbf {Q} ^ {(m)} \\mathbf {K} ^ {(m) T} / \\sqrt {D}\\right) \\mathbf {V} ^ {(m) T}, \\tag {1}", + "image_path": "42b5080794ca133b685098e47e6d8ac434dad50f937211a5f51388dbabd02d95.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "text", + "content": " is the query/key dimension, " + }, + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 420, + 287, + 529 + ], + "type": "text", + "content": "-th attention head. However, the calculation based on the entire feature map significantly increases the computational complexity. The output attention map also requires a large number of calculations when sent to the next block for color space segmentation. Instead of the calculation for the entire feature map, in this paper, the corresponding calculation is delegated to interest points for achieving fewer computations and higher flexibility (Fig. 2(b))." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 529, + 288, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 288, + 588 + ], + "type": "text", + "content": "The feature map " + }, + { + "bbox": [ + 47, + 529, + 288, + 588 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 529, + 288, + 588 + ], + "type": "text", + "content": " is fed into a lightweight subnetwork " + }, + { + "bbox": [ + 47, + 529, + 288, + 588 + ], + "type": "inline_equation", + "content": "Net_{i}" + }, + { + "bbox": [ + 47, + 529, + 288, + 588 + ], + "type": "text", + "content": " to generate the initial interest points. Then, another subnetwork " + }, + { + "bbox": [ + 47, + 529, + 288, + 588 + ], + "type": "inline_equation", + "content": "Net_{o}" + }, + { + "bbox": [ + 47, + 529, + 288, + 588 + ], + "type": "text", + "content": " is adopted to calculate the offset from " + }, + { + "bbox": [ + 47, + 529, + 288, + 588 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 47, + 529, + 288, + 588 + ], + "type": "text", + "content": " to adjust the positions of the interest points, which is formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 80, + 597, + 253, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 597, + 253, + 613 + ], + "spans": [ + { + "bbox": [ + 80, + 597, + 253, + 613 + ], + "type": "interline_equation", + "content": "Q = X W _ {Q}, K = \\widetilde {X} W _ {k}, V = \\widetilde {X} W _ {v}, w i t h", + "image_path": "ad1757dfb7ec712964e5d82951a2f140b62ee55bb6ada602d11e0190a67d9c97.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 82, + 615, + 287, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 615, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 82, + 615, + 287, + 643 + ], + "type": "interline_equation", + "content": "X _ {i} = \\emptyset [ X; G * (\\underbrace {\\operatorname {N e t} _ {i} (X)} _ {\\text {i n t e r s t p o i n t s}} + \\underbrace {\\operatorname {N e t} _ {o} (Q)}) ], \\tag {2}", + "image_path": "2b2b6afc18cf15003623a16eb47ee6477a9ce624a97f4b620dc985f1fa81934e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "where interest points and offsets are both represented by two-dimensional coordinates that are normalized to the range " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "[-1, +1]" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "(-1, -1)" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " represents the top-left corner and " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "(+1, +1)" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " represents the bottom-right corner. A gating module, " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": ", is employed to determine activation of the" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 371, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 371, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 545, + 407 + ], + "type": "text", + "content": "current adjusted interest point coordinates, reducing redundancy further in the space and preventing outliers and overlapping interest points. It can be expressed as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 363, + 415, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 415, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 363, + 415, + 545, + 430 + ], + "type": "interline_equation", + "content": "G = \\operatorname {S i g m o i d} \\left(\\operatorname {N e t} _ {g} (X) \\cdot \\alpha\\right), \\tag {3}", + "image_path": "b763a24d94161a2009be882696b6a7bb5fbb6577290b0c8acae3124f76e92523.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 437, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 497 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 437, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 437, + 545, + 497 + ], + "type": "text", + "content": " is a value large enough to ensure that the gate matrix is either 0 or 1. After obtaining the adjusted and filtered interest points " + }, + { + "bbox": [ + 305, + 437, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\widetilde{X}" + }, + { + "bbox": [ + 305, + 437, + 545, + 497 + ], + "type": "text", + "content": ", the bilinear interpolation is applied to sample the output feature from " + }, + { + "bbox": [ + 305, + 437, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\widetilde{X}" + }, + { + "bbox": [ + 305, + 437, + 545, + 497 + ], + "type": "text", + "content": ", which can be represented as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 310, + 506, + 545, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 506, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 310, + 506, + 545, + 544 + ], + "type": "interline_equation", + "content": "\\varnothing [ X; (p _ {x}, p _ {y}) ] = \\sum_ {(q _ {x}, q _ {y})} M (p _ {x}, p _ {y}) M (q _ {x}, q _ {y}) X (q _ {x}, q _ {y}), \\tag {4}", + "image_path": "112188252d1583b9a14214b275ae49c9cd0f452b9e9e048bcc7b126812798125.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 544, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 545, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "(p_x, p_y)" + }, + { + "bbox": [ + 305, + 544, + 545, + 605 + ], + "type": "text", + "content": " represents one of the coordinates of the interest points, " + }, + { + "bbox": [ + 305, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "M(a, b) = \\max(0, 1 - |a - b|)" + }, + { + "bbox": [ + 305, + 544, + 545, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "(q_x, q_y)" + }, + { + "bbox": [ + 305, + 544, + 545, + 605 + ], + "type": "text", + "content": " is non-zero on only the 4 integral points in " + }, + { + "bbox": [ + 305, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 544, + 545, + 605 + ], + "type": "text", + "content": " closest to " + }, + { + "bbox": [ + 305, + 544, + 545, + 605 + ], + "type": "inline_equation", + "content": "(p_x, p_y)" + }, + { + "bbox": [ + 305, + 544, + 545, + 605 + ], + "type": "text", + "content": ". Finally, the standard multihead attention mechanism of formula (1) is applied to calculate the output." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 611, + 507, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 507, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 507, + 624 + ], + "type": "text", + "content": "3.2. Module for Color Space Segmentation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 714 + ], + "type": "text", + "content": "The selection of dominant and secondary colors is based on human understanding of the relationship between image content and color, which is highly subjective and leads to uncertainty in the size and segmentation of color spaces. To address this issue, we propose a learning-based approach to determine the size and the segmentation of color spaces. After obtaining the color features of the interest points, 1)" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "21841" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 215, + 225 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 215, + 225 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 215, + 225 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 215, + 225 + ], + "type": "image", + "image_path": "fc88593a5f26dc4b45d01096f6ed17a432f0de7c6fd8f7b15010e03297cc3348.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 547, + 262 + ], + "type": "text", + "content": "Figure 3. The proposed ICAA17K dataset. (a) Visualization of images with different color types; (b) quantification of 15 complementary/polychromatic colors; (c) quantification of 15 monochromatic colors; (d) examples in the proposed ICAA17K dataset with the ground truth (and confidence) scores shown below each image." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 206, + 71, + 355, + 217 + ], + "blocks": [ + { + "bbox": [ + 206, + 71, + 355, + 217 + ], + "lines": [ + { + "bbox": [ + 206, + 71, + 355, + 217 + ], + "spans": [ + { + "bbox": [ + 206, + 71, + 355, + 217 + ], + "type": "image", + "image_path": "cc9ba2ca66102fc0a5dc1ca4a1fe5e85127535e489dc48618e8f5a4feea3b385.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 219, + 285, + 226 + ], + "lines": [ + { + "bbox": [ + 277, + 219, + 285, + 226 + ], + "spans": [ + { + "bbox": [ + 277, + 219, + 285, + 226 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 355, + 74, + 433, + 217 + ], + "blocks": [ + { + "bbox": [ + 355, + 74, + 433, + 217 + ], + "lines": [ + { + "bbox": [ + 355, + 74, + 433, + 217 + ], + "spans": [ + { + "bbox": [ + 355, + 74, + 433, + 217 + ], + "type": "image", + "image_path": "329bd31d8bd452db0f8029c151a1ecd8aaa5b299a729010e4d7e31e171b327ed.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 219, + 402, + 226 + ], + "lines": [ + { + "bbox": [ + 393, + 219, + 402, + 226 + ], + "spans": [ + { + "bbox": [ + 393, + 219, + 402, + 226 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 433, + 79, + 477, + 109 + ], + "blocks": [ + { + "bbox": [ + 433, + 79, + 477, + 109 + ], + "lines": [ + { + "bbox": [ + 433, + 79, + 477, + 109 + ], + "spans": [ + { + "bbox": [ + 433, + 79, + 477, + 109 + ], + "type": "image", + "image_path": "ac10bd085bfa1fcd0779b2774a81a5183f53bb43b98646c65b9ce98c2ba397cd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 480, + 79, + 526, + 109 + ], + "blocks": [ + { + "bbox": [ + 480, + 79, + 526, + 109 + ], + "lines": [ + { + "bbox": [ + 480, + 79, + 526, + 109 + ], + "spans": [ + { + "bbox": [ + 480, + 79, + 526, + 109 + ], + "type": "image", + "image_path": "912345f7027d8e16417efead5266ace274730cdef96037a66bcc43a4c2b05325.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 528, + 79, + 542, + 109 + ], + "blocks": [ + { + "bbox": [ + 528, + 79, + 542, + 109 + ], + "lines": [ + { + "bbox": [ + 528, + 79, + 542, + 109 + ], + "spans": [ + { + "bbox": [ + 528, + 79, + 542, + 109 + ], + "type": "image", + "image_path": "5051098053cdec903071ed06e3ec1867a4b70eef6c0d77fb950cff303b4d3fbc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 116, + 487, + 146 + ], + "blocks": [ + { + "bbox": [ + 443, + 110, + 471, + 116 + ], + "lines": [ + { + "bbox": [ + 443, + 110, + 471, + 116 + ], + "spans": [ + { + "bbox": [ + 443, + 110, + 471, + 116 + ], + "type": "text", + "content": "3.66 (0.74)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 110, + 524, + 116 + ], + "lines": [ + { + "bbox": [ + 481, + 110, + 524, + 116 + ], + "spans": [ + { + "bbox": [ + 481, + 110, + 524, + 116 + ], + "type": "text", + "content": "4.63(0.63) 4.89" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 434, + 116, + 487, + 146 + ], + "lines": [ + { + "bbox": [ + 434, + 116, + 487, + 146 + ], + "spans": [ + { + "bbox": [ + 434, + 116, + 487, + 146 + ], + "type": "image", + "image_path": "00d6ec8c0eb28a688933aa156118e178889f83b833c1d0dcf78dea25b516b7aa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 147, + 469, + 152 + ], + "lines": [ + { + "bbox": [ + 436, + 147, + 469, + 152 + ], + "spans": [ + { + "bbox": [ + 436, + 147, + 469, + 152 + ], + "type": "text", + "content": "5.46(0.66)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 471, + 116, + 503, + 146 + ], + "blocks": [ + { + "bbox": [ + 471, + 116, + 503, + 146 + ], + "lines": [ + { + "bbox": [ + 471, + 116, + 503, + 146 + ], + "spans": [ + { + "bbox": [ + 471, + 116, + 503, + 146 + ], + "type": "image", + "image_path": "6a9be4e5ae2a3269f8e820e1d56a35a43077cdc7f00c0d02e5d9e3f5b3ad278c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 471, + 147, + 503, + 179 + ], + "blocks": [ + { + "bbox": [ + 471, + 147, + 503, + 179 + ], + "lines": [ + { + "bbox": [ + 471, + 147, + 503, + 179 + ], + "spans": [ + { + "bbox": [ + 471, + 147, + 503, + 179 + ], + "type": "image", + "image_path": "2026a41866eb13b283d42ca7dd39d448ca78cc67412e74ca8b8709bc2a940077.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 503, + 116, + 542, + 146 + ], + "blocks": [ + { + "bbox": [ + 528, + 110, + 541, + 116 + ], + "lines": [ + { + "bbox": [ + 528, + 110, + 541, + 116 + ], + "spans": [ + { + "bbox": [ + 528, + 110, + 541, + 116 + ], + "type": "text", + "content": "(0.76)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 503, + 116, + 542, + 146 + ], + "lines": [ + { + "bbox": [ + 503, + 116, + 542, + 146 + ], + "spans": [ + { + "bbox": [ + 503, + 116, + 542, + 146 + ], + "type": "image", + "image_path": "926c7dfaeb5827647b94a7e77b2ee8061d7160a42c09a5ddb201fdfb71ebf747.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 503, + 147, + 542, + 179 + ], + "blocks": [ + { + "bbox": [ + 503, + 147, + 542, + 179 + ], + "lines": [ + { + "bbox": [ + 503, + 147, + 542, + 179 + ], + "spans": [ + { + "bbox": [ + 503, + 147, + 542, + 179 + ], + "type": "image", + "image_path": "ec70bebb05f6dc34d702d98a8942b5b642e95399f4c374ca6749cb2b246d762c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 433, + 179, + 487, + 217 + ], + "blocks": [ + { + "bbox": [ + 433, + 179, + 487, + 217 + ], + "lines": [ + { + "bbox": [ + 433, + 179, + 487, + 217 + ], + "spans": [ + { + "bbox": [ + 433, + 179, + 487, + 217 + ], + "type": "image", + "image_path": "fbf03364f7cb87b333816d3297eb73514bfee9e8195cce7e89d2d5835f4a214e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 488, + 180, + 542, + 217 + ], + "blocks": [ + { + "bbox": [ + 488, + 180, + 542, + 217 + ], + "lines": [ + { + "bbox": [ + 488, + 180, + 542, + 217 + ], + "spans": [ + { + "bbox": [ + 488, + 180, + 542, + 217 + ], + "type": "image", + "image_path": "cb8fc96b94bc8ec7582996bbcc7698816f0daa34f63250d960feed446db441f9.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 219, + 498, + 225 + ], + "lines": [ + { + "bbox": [ + 488, + 219, + 498, + 225 + ], + "spans": [ + { + "bbox": [ + 488, + 219, + 498, + 225 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 52, + 281, + 194, + 350 + ], + "blocks": [ + { + "bbox": [ + 52, + 281, + 194, + 350 + ], + "lines": [ + { + "bbox": [ + 52, + 281, + 194, + 350 + ], + "spans": [ + { + "bbox": [ + 52, + 281, + 194, + 350 + ], + "type": "image", + "image_path": "0d18ec308a943c77d22b9e1aac09494ee258b6f06ca2a1044e5f3ff2b9f39827.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 195, + 281, + 285, + 335 + ], + "blocks": [ + { + "bbox": [ + 195, + 281, + 285, + 335 + ], + "lines": [ + { + "bbox": [ + 195, + 281, + 285, + 335 + ], + "spans": [ + { + "bbox": [ + 195, + 281, + 285, + 335 + ], + "type": "image", + "image_path": "8e49bcfc7b6e4fcf8395256bdaf18846ccb5905ec48b7159116ecc11309a59f5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 50, + 350, + 192, + 446 + ], + "blocks": [ + { + "bbox": [ + 50, + 350, + 192, + 446 + ], + "lines": [ + { + "bbox": [ + 50, + 350, + 192, + 446 + ], + "spans": [ + { + "bbox": [ + 50, + 350, + 192, + 446 + ], + "type": "image", + "image_path": "0bc106784b3c2928c7a947f5df6a1e8c883885ca9b9e2bf53350774216accf23.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 448, + 127, + 457 + ], + "lines": [ + { + "bbox": [ + 115, + 448, + 127, + 457 + ], + "spans": [ + { + "bbox": [ + 115, + 448, + 127, + 457 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 460, + 287, + 505 + ], + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 505 + ], + "type": "text", + "content": "Figure 4. An example of color space segmentation with 4 color spaces " + }, + { + "bbox": [ + 46, + 460, + 287, + 505 + ], + "type": "inline_equation", + "content": "(\\kappa = 4)" + }, + { + "bbox": [ + 46, + 460, + 287, + 505 + ], + "type": "text", + "content": ". (a) Three sample points. (b) The weights of the three points in each space calculated through formula (5). (c) The weights of each point in each space." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 195, + 342, + 284, + 389 + ], + "blocks": [ + { + "bbox": [ + 195, + 342, + 284, + 389 + ], + "lines": [ + { + "bbox": [ + 195, + 342, + 284, + 389 + ], + "spans": [ + { + "bbox": [ + 195, + 342, + 284, + 389 + ], + "type": "image", + "image_path": "fd4007517604df7d0eec30f11c7a000e9444347869e0fbdce0834329f0a9ef70.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 196, + 396, + 284, + 445 + ], + "blocks": [ + { + "bbox": [ + 196, + 396, + 284, + 445 + ], + "lines": [ + { + "bbox": [ + 196, + 396, + 284, + 445 + ], + "spans": [ + { + "bbox": [ + 196, + 396, + 284, + 445 + ], + "type": "image", + "image_path": "40190d9223b99351e8d237fa7cb47e42375c60904a9b87efad9263a712058aaa.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 448, + 246, + 457 + ], + "lines": [ + { + "bbox": [ + 234, + 448, + 246, + 457 + ], + "spans": [ + { + "bbox": [ + 234, + 448, + 246, + 457 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 530, + 287, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 614 + ], + "type": "text", + "content": "the interest points are grouped into several color spaces (" + }, + { + "bbox": [ + 46, + 530, + 287, + 614 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 530, + 287, + 614 + ], + "type": "text", + "content": " pre-defined types). 2) Each color space is mapped to the " + }, + { + "bbox": [ + 46, + 530, + 287, + 614 + ], + "type": "inline_equation", + "content": "[0, +1]" + }, + { + "bbox": [ + 46, + 530, + 287, + 614 + ], + "type": "text", + "content": " interval, where each interval can be determined by its widths and centers. 3) By counting the weight distribution of points in each color space, it is easy to distinguish the attention of each color space, thus dividing the dominant and secondary color space." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": "However, using fixed bin widths and centers is not differentiable, and the parameters cannot be updated through backpropagation. To resolve this issue, the widths and centers are adaptively generated using two lightweight networks, " + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "inline_equation", + "content": "Net_{w}" + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "inline_equation", + "content": "Net_{c}" + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": ", which can be learned during training. To calculate the uncertainty in color space segmentation, our model learns a probability distribution for each point " + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": " to indicate its weight in " + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 617, + 287, + 714 + ], + "type": "text", + "content": " color spaces, denoted as" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 281, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 545, + 317 + ], + "type": "inline_equation", + "content": "S(p_{j}) = [s_{1}(p_{j}),\\dots,s_{\\kappa}(p_{j})]" + }, + { + "bbox": [ + 306, + 281, + 545, + 317 + ], + "type": "text", + "content": ", the weight " + }, + { + "bbox": [ + 306, + 281, + 545, + 317 + ], + "type": "inline_equation", + "content": "s_i()" + }, + { + "bbox": [ + 306, + 281, + 545, + 317 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 306, + 281, + 545, + 317 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 306, + 281, + 545, + 317 + ], + "type": "text", + "content": " allocated to the i-th space is determined by the piecewise linear basis function." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 345, + 329, + 545, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 329, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 345, + 329, + 545, + 342 + ], + "type": "interline_equation", + "content": "s _ {i} \\left(p _ {j}\\right) = \\max \\left\\{0, 1 - w _ {i} \\times \\left| p _ {j} - c _ {i} \\right| \\right\\}, \\tag {5}", + "image_path": "00df60a1478abe67d81bdf087c514505ac66e70c4d5e2baadb2af6570307a3c9.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "text", + "content": " are the widths and centers of the i-th color space. If an interest point falls into the i-th bin, e.g., the interval of " + }, + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "inline_equation", + "content": "[c_i - w_i,c_i + w_i]" + }, + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "text", + "content": ", it votes for the i-th bin with a positive weight " + }, + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "inline_equation", + "content": "s_i(p_j)" + }, + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "text", + "content": ". Fig. 4 shows an example with " + }, + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\kappa = 4" + }, + { + "bbox": [ + 304, + 353, + 545, + 411 + ], + "type": "text", + "content": " color spaces." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "type": "text", + "content": "The operation in formula (5) can be designed into a lightweight color space partition block (Fig. 2 (d)), where " + }, + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "type": "inline_equation", + "content": "p_j - c_i" + }, + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "type": "text", + "content": " is equivalent to convolving the point feature maps with a fixed " + }, + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "type": "text", + "content": " kernel and a learnable bias " + }, + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "type": "text", + "content": ". Then, after taking the absolute value, the remaining operation is completed using another learnable weight, a fixed " + }, + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "type": "text", + "content": " bias, and a ReLU activation function." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 306, + 507, + 410, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 507, + 410, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 507, + 410, + 521 + ], + "type": "text", + "content": "4. Proposed Dataset" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 527, + 408, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 408, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 408, + 540 + ], + "type": "text", + "content": "4.1. Image Collection" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "How to avoid selection bias by building a comprehensive ICAA dataset. Selection bias occurs when some types of the intended color, scenes and devices are less likely to be included than others, it can affect the validity and generalizability of training models. To solve this issue, 1) we analyzed the most frequently uploaded colors on the Flickr and COLOURLovers website from 2009 to 2022. Finally, 15 complementary/polychromatic colors and 15 monochromatic colors are selected as the basic color types, which ensures that the dataset includes a diverse range of color types (Fig. 3). 2) To improve the dataset's generalizability across different scenes, the images in ICAA17K cover more than 50 scene categories, e.g., animals, plants, mountains, landscapes, and night scenes. 3) We collected images" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "21842" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 545, + 163 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 163 + ], + "type": "table", + "html": "
Metric [ICAA]2014-20192020-2022
RAPID [21]AADB [12]PAM [22]A-Lamp [23]NIMA [18]\\( MP_{ada} \\) [24]MLSP [20]MT-A [8]BIAA [25]UIAA [26]MUSIQ [27]HGCN [28]TANet [29]MaxViT [17]Ours *1Ours *2
SRCC↑.759.788.796.810.809.810.826.821.820.817.835.826.829.855.873.887
LCC↑.771.793.804.818.815.819.837.830.829.830.841.831.836.874.890.901
Accu↑.836.887.899.908.906.906.921.908.913.910.918.909.903.920.957.965
Epoch↓80606050605050601006080601102002540
", + "image_path": "3ee87521a80e9bd1ad858d0d399f8e03e6ecb2d0239ffce8b212c599f4123e7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 224, + 538, + 305 + ], + "blocks": [ + { + "bbox": [ + 46, + 171, + 546, + 215 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 546, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 546, + 215 + ], + "type": "text", + "content": "Table 3. Comparisons of 15 models on our ICAA17K dataset. We retrained the models for the best performance with the recommended parameter and dataset settings. See subsection 5.1 for details regarding the metrics. \"Ours *1\" indicates that, similar to other methods, only the color annotation scores from the ICAA17K dataset are used during training. \"Ours *2\" indicates that the holistic aesthetics score is jointly predicted when using multiple tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 224, + 538, + 305 + ], + "lines": [ + { + "bbox": [ + 56, + 224, + 538, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 224, + 538, + 305 + ], + "type": "table", + "html": "
Metric [SPAQ]2014-20192020-2022
RAPID [21]AADB [12]PAM [22]A-Lamp [23]NIMA [18]\\( MP_{ada} \\) [24]MLSP [20]MT-A [8]BIAA [25]UIAA [26]MUSIQ [27]HGCN [28]TANet [29]MaxViT [17]Ours * 1
SRCC↑.721.725.732.747.730.746.753.751.748.750.753.755.750.753.770
LCC↑.730.739.744.760.746.757.773.763.759.763.770.771.761.770.791
Epoch↓95757065556555308570904511015025
", + "image_path": "447afd4ef330ad70fc4be740a81e994c8833b037f67f6cfcf5c73a0615b20258.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 312, + 546, + 335 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 546, + 335 + ], + "type": "text", + "content": "Table 4. Comparisons of 15 models on the SPAQ dataset [8]. All models used only the colorfulness score in the dataset for training, and were retrained for the best performance using the recommended parameters and dataset settings." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 354, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 402 + ], + "type": "text", + "content": "taken by more than 80 different types of devices, including mobile phones (e.g., iPhone X, Huawei Mate 10, and Samsung Galaxy S10) and digital cameras (e.g., Canon, Nikon, Sony, Panasonic), to encompass a wide range of devices." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 497 + ], + "type": "text", + "content": "How to alleviate long-tailed distributions by performing a preassessment strategy. The annotated label in the most aesthetic datasets have long-tailed distributions [34, 8], which bias models trained on imbalanced data toward majority examples. To avoid correcting the long-tailed distributions after label annotation, we propose a preassessment strategy during data collection to maximize label diversity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 579 + ], + "type": "text", + "content": "First, the color harmony [19] is applied to evaluate objective color quality by measuring the distance between the tested image and templates on the hue wheel. Given a harmonic template " + }, + { + "bbox": [ + 46, + 498, + 287, + 579 + ], + "type": "inline_equation", + "content": "T(\\beta)" + }, + { + "bbox": [ + 46, + 498, + 287, + 579 + ], + "type": "text", + "content": " (one of the predefined type of templates " + }, + { + "bbox": [ + 46, + 498, + 287, + 579 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 498, + 287, + 579 + ], + "type": "text", + "content": ") with an associated orientation " + }, + { + "bbox": [ + 46, + 498, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 498, + 287, + 579 + ], + "type": "text", + "content": ", the color harmony " + }, + { + "bbox": [ + 46, + 498, + 287, + 579 + ], + "type": "inline_equation", + "content": "f_{obj}" + }, + { + "bbox": [ + 46, + 498, + 287, + 579 + ], + "type": "text", + "content": " of a tested image " + }, + { + "bbox": [ + 46, + 498, + 287, + 579 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 498, + 287, + 579 + ], + "type": "text", + "content": " could be calculated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 586, + 287, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 586, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 80, + 586, + 287, + 611 + ], + "type": "interline_equation", + "content": "f _ {o b j} (C) = \\sum_ {l \\in C} \\| H (l) - E _ {T (\\beta)} (l) \\| \\cdot Y (l), \\tag {6}", + "image_path": "38f0fa068f4701539ebe30c0a0a3385db3ed73b29feff58db21f0c88bff5a555.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": " is the hue of a pixel, " + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "inline_equation", + "content": "E_{T(\\beta)}" + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": " is the sector border hue of harmonic template " + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "inline_equation", + "content": "T(\\beta)" + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": " mean the hue and the saturation channels, respectively, " + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|" + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": " denotes the arc-length distance on the hue wheel (measured in radians)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Second, we crawl images from Flickr and record their metadata, including \"views\" (number of visits " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "VI" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": ") and \"faves\" (number of clicks that favor image " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "FA" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "). Intuitively, an image with a higher ratio of faves to views is more pop-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 354, + 546, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 390 + ], + "type": "text", + "content": "ular and beautiful. Specifically, considering the affect of user's herd mentality, we adopt the following formula to approximate the subjective perception:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 376, + 399, + 545, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 399, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 376, + 399, + 545, + 427 + ], + "type": "interline_equation", + "content": "f _ {s u b j} (C) = \\frac {F A (C)}{\\sqrt {V I (C)}}. \\tag {7}", + "image_path": "4bac91a0bcc4e23723d2f4a63f9288ef37a6d8f01e1eced8f71bf06caf098e10.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 435, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 483 + ], + "type": "text", + "content": "Finally, we normalize the " + }, + { + "bbox": [ + 304, + 435, + 545, + 483 + ], + "type": "inline_equation", + "content": "f_{subj}" + }, + { + "bbox": [ + 304, + 435, + 545, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 435, + 545, + 483 + ], + "type": "inline_equation", + "content": "f_{obj}" + }, + { + "bbox": [ + 304, + 435, + 545, + 483 + ], + "type": "text", + "content": ", and add them together as the final preassessment score of the image (not shown to annotators), and extract balanced samples from each score range for the next stage of manual evaluation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 491, + 413, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 413, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 413, + 504 + ], + "type": "text", + "content": "4.2. Image Annotation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": "How to avoid unreliablely by improving annotation quality. The traditional annotation process suffers from a lack of a baseline or reference for evaluating aesthetics, which makes subjective responses unreliable in the early stages. To mitigate this problem, 1) we first provided annotators simple guidelines based on photography theories to guide their judgment of image color aesthetics. 2) We provided anchor images with different rating scales as references. Experts rated the anchor images based on their rich photography experience, allowing annotators to understand the differences between rating scales. Before annotating, annotators had to review the anchor images and guidelines thoroughly. The average value was then calculated as the ground truth. 3) Annotators could provide a confidence score (from 0 to 1) to indicate their certainty in each rating. This information can be useful for studying the noise and subjectivity in the dataset. 4) Finally, each image has" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "21843" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 71, + 492, + 268 + ], + "blocks": [ + { + "bbox": [ + 100, + 71, + 492, + 268 + ], + "lines": [ + { + "bbox": [ + 100, + 71, + 492, + 268 + ], + "spans": [ + { + "bbox": [ + 100, + 71, + 492, + 268 + ], + "type": "image", + "image_path": "689db2d5d58389a81be6a6170f33d7557dd7823f72c581f3a1f44d26689e2a94.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 272, + 539, + 283 + ], + "lines": [ + { + "bbox": [ + 47, + 272, + 539, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 539, + 283 + ], + "type": "text", + "content": "Figure 5. Saliency maps of the 15 models. The attention of our method is segmented into distinct regions based on the color distribution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 73, + 301, + 263, + 450 + ], + "blocks": [ + { + "bbox": [ + 73, + 301, + 263, + 450 + ], + "lines": [ + { + "bbox": [ + 73, + 301, + 263, + 450 + ], + "spans": [ + { + "bbox": [ + 73, + 301, + 263, + 450 + ], + "type": "image", + "image_path": "abad8b514bc70e63ac3fc6ffd9aae3c3ef8a0efab3b5f454756d67d6dbfecaec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 454, + 287, + 477 + ], + "lines": [ + { + "bbox": [ + 47, + 454, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 287, + 477 + ], + "type": "text", + "content": "Figure 6. Analysis of kernel sizes in " + }, + { + "bbox": [ + 47, + 454, + 287, + 477 + ], + "type": "inline_equation", + "content": "Net_{i}" + }, + { + "bbox": [ + 47, + 454, + 287, + 477 + ], + "type": "text", + "content": " and the number of color spaces " + }, + { + "bbox": [ + 47, + 454, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 47, + 454, + 287, + 477 + ], + "type": "text", + "content": ". Larger kernel sizes result in fewer interest points." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 497, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 287, + 556 + ], + "type": "text", + "content": "an annotation score from 1 to 10, which represents color aesthetic perception from lowest to highest, we selected images with an average confidence score above 0.5 as samples in the dataset. We collected about 1,500 opinions for each image. Some examples are shown in Fig. 3 (d)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 567, + 128, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 581 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 587, + 173, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 173, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 173, + 600 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 605, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 286, + 676 + ], + "type": "text", + "content": "Benchmark Datasets. We performed model evaluations on two datasets, the proposed ICAA17K and SPAQ datasets [8]. Although the SPAQ dataset focuses on image quality rather than aesthetics, it provides an annotation on colorfulness, which can also afford testing the ability of various baseline models to extract color features to some extent." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "Benchmark Models and Training Protocols. To the best of our knowledge, there is no publicly available learning-based ICAA model. Therefore, we selected 14 deep learn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 303, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 303, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 545, + 385 + ], + "type": "text", + "content": "ing baselines (Table. 1) according to the following criteria: 1) classical architectures with available code and 2) SOTA in a specific field, e.g., personalized IAA. These baselines are trained with the recommended parameter settings (e.g., optimizer and batch size), using the same training and testing settings. The mean squared error (MSE) is adopted as the loss function across all methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 387, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 545, + 518 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt two popular evaluation metrics, the Spearman's rank correlation coefficient (SRCC) and the linear correlation coefficient (LCC), to evaluate the model's ability. The metrics for the ICAA17K dataset also include binary classification accuracy (color aesthetically negative or positive). Since the ground-truth annotations in the ICAA17K dataset are based on a 5-point boundary line, if an annotator tends to give a score greater than 5, it is considered aesthetically positive; otherwise, it is considered negative. Therefore, this metric can evaluate the model's ability to perform coarse-grained evaluations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 527, + 463, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 463, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 463, + 540 + ], + "type": "text", + "content": "5.2. Results and Dataset Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 545, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 545, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 545, + 545, + 665 + ], + "type": "text", + "content": "ICAA17K Dataset. Table 3 lists the results of 15 models on the ICAA dataset. The proposed baseline method (marked by \"Ours *1\") achieves the best performance on all metrics and has a higher training speed. This is mainly due to its sparse attention and adaptive color space partitioning. Furthermore, our ICAA17K dataset also provides additional annotations regarding holistic aesthetics. Therefore, we incorporate this additional supervision information into the multitask training (marked by \"Ours *2\"), which improves performance but increases the training time." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "SPAQ Dataset. Based on the results of the 15 methods reported in Table 4, our method still achieves the top performance on all metrics with fewer training epochs, showing that it is a promising solution for the IQA problem." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "21844" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 166, + 203 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 166, + 203 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 166, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 166, + 203 + ], + "type": "image", + "image_path": "b308b498a137da42536feeaeb0ca2401664b9205abff113abb75053adcb5947c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 207, + 289, + 284 + ], + "lines": [ + { + "bbox": [ + 46, + 207, + 289, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 289, + 284 + ], + "type": "text", + "content": "Figure 7. Visualization of our attention mechanism. (a) Sampled interest points adjusted by offsets. (b) Interest points filtered by the gating module. (c) The radius of the circles represents the attention weights at multiple heads, with a larger radius indicating a higher weight. (d) The color space segmentation of each point, where the numbers represent the primary and secondary relationship of the color space, e.g., \"1\" represents the dominant color." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 72, + 272, + 204 + ], + "blocks": [ + { + "bbox": [ + 167, + 72, + 272, + 204 + ], + "lines": [ + { + "bbox": [ + 167, + 72, + 272, + 204 + ], + "spans": [ + { + "bbox": [ + 167, + 72, + 272, + 204 + ], + "type": "image", + "image_path": "ea74892bb8a1944a3a5caa49c8434988ed8bd79d74b666bb183a1ac9de88fd52.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 304, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 424 + ], + "type": "text", + "content": "Dataset Analysis. Compared to the SPAQ dataset, when using the ICAA17K dataset for training, most networks converge faster and have higher SRCC and LCC metrics, which indicates that models trained on the ICAA17K dataset more accurately fit human subjective color perception. It also suggests that introducing more annotators (SPAQ 600 vs ICAA17K 1500) and providing more annotation references for annotators can alleviate noise caused by subjective annotation, which is a promising direction toward improving the quality of IQA or IAA datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 431, + 141, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 141, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 141, + 445 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 593 + ], + "type": "text", + "content": "Delegate vs Shift/Deformable Transformers. In Table 5, we replace the Delegate Transformer blocks with the shift attention mechanism of Swin Transformer [39] and deformable attention of DAT Transformer [40] in the last stage. In this case, our model has better performance and lower computation, which outperforms Swin by " + }, + { + "bbox": [ + 46, + 450, + 287, + 593 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 46, + 450, + 287, + 593 + ], + "type": "text", + "content": " in the SRCC with " + }, + { + "bbox": [ + 46, + 450, + 287, + 593 + ], + "type": "inline_equation", + "content": "41.7\\%" + }, + { + "bbox": [ + 46, + 450, + 287, + 593 + ], + "type": "text", + "content": " FLOPs and outperforms DAT by " + }, + { + "bbox": [ + 46, + 450, + 287, + 593 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 46, + 450, + 287, + 593 + ], + "type": "text", + "content": " in the SRCC with " + }, + { + "bbox": [ + 46, + 450, + 287, + 593 + ], + "type": "inline_equation", + "content": "41.7\\%" + }, + { + "bbox": [ + 46, + 450, + 287, + 593 + ], + "type": "text", + "content": " FLOPs. Additionally, the stacking way also has a significant impact on the performance. The Delegate Transformer blocks may result in information loss if it is employed in the early stages due to the sparse attention mechanism." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "The Number of Interest Points and Color Space. The number of interest points mainly depends on the kernel size of the subnetwork " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "Net_{i}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". In Fig. 6, the results show that an inadequate or excessive number of interest points leads to degradation in the model's performance. Moreover, selecting the color space hyperparameter " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " is also crucial for the performance of our model. Our experiments reveal that the optimal performance is achieved when " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\kappa = 7" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": ". Additionally, a larger value of " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " leads to an increase in computation and memory requirements. In contrast, a" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 538, + 188 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 538, + 188 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 538, + 188 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 538, + 188 + ], + "type": "table", + "html": "
#StagesFLOPs↓SRCC↑LCC↑
1234
1SSSS240G0.8300.843
2SSDD240G0.8360.848
3SSGG140G0.8730.890
4GGSS210G0.8280.835
5SGGS140G0.8380.850
6GSSG210G0.8250.833
7GGGG130G0.8370.846
", + "image_path": "2e52d80c05a16be86ce230ce62efd927261c5f5b007b719d05a9d629b21f75f7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 196, + 545, + 229 + ], + "lines": [ + { + "bbox": [ + 305, + 196, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 305, + 196, + 545, + 229 + ], + "type": "text", + "content": "Table 5. Ablations on the application of different attention mechanisms in different stages, where " + }, + { + "bbox": [ + 305, + 196, + 545, + 229 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 196, + 545, + 229 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 196, + 545, + 229 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 196, + 545, + 229 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 196, + 545, + 229 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 196, + 545, + 229 + ], + "type": "text", + "content": " represent the application of Swin, DAT and our blocks, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 249, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 249, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 249, + 545, + 297 + ], + "type": "text", + "content": "smaller value of " + }, + { + "bbox": [ + 305, + 249, + 545, + 297 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 305, + 249, + 545, + 297 + ], + "type": "text", + "content": " may not provide a sufficient characterization of the color composition in the input image. Moreover, the effectiveness of different modules are shown in the Appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 304, + 389, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 389, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 389, + 316 + ], + "type": "text", + "content": "5.4. Visualization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 323, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 467 + ], + "type": "text", + "content": "The GradCAM [41] method is applied to visualize the saliency maps in Fig. 5. Our model differs from other methods because it segments the image into distinct regions based on the color distribution, and the region that typically receives the greatest attention is the region where the dominant color is located. The reason is further explained in Fig. 7. First, our model generates interest points via " + }, + { + "bbox": [ + 305, + 323, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\text{Net}_i" + }, + { + "bbox": [ + 305, + 323, + 545, + 467 + ], + "type": "text", + "content": " and adjusts their positions by " + }, + { + "bbox": [ + 305, + 323, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\text{Net}_o" + }, + { + "bbox": [ + 305, + 323, + 545, + 467 + ], + "type": "text", + "content": ". Second, some redundant points are eliminated via " + }, + { + "bbox": [ + 305, + 323, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\text{Net}_g" + }, + { + "bbox": [ + 305, + 323, + 545, + 467 + ], + "type": "text", + "content": " and weights are calculated using multihead attention. Third, the color space segmentation module and formula (5) are employed to divide the different spaces and direct different attention toward them." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 477, + 383, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 383, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 383, + 489 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": "In this paper, a comprehensive study on ICAA has been conducted by developing a Delegate Transformer model as a baseline, constructing a color-oriented dataset benchmark, and establishing the largest-scale benchmarks to date. Compared with traditional or data-driven models, our model achieves SOTA performance and produces visually consistent results. We hope our work should contribute to the community's research on ICAA. However, ICAA remains far from being solved, we will optimize our approach by incorporating more color prior knowledge in the future." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 628, + 416, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 628, + 416, + 641 + ], + "spans": [ + { + "bbox": [ + 306, + 628, + 416, + 641 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 647, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 708 + ], + "type": "text", + "content": "This work was supported by the Funds for Creative Research Groups of China under Grant 61921003, the national key R & D program intergovernmental international science and technology innovation cooperation project 2021YFE0101600." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "21845" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Ahmed Talib, Massudi Mahmuddin, Husniza Husni, and Loay E George. A weighted dominant color descriptor for content-based image retrieval. Journal of Visual Communication and Image Representation, 24(3):345-360, 2013." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 182 + ], + "type": "text", + "content": "[2] Yining Deng, BS Manjunath, Charles Kenney, Michael S Moore, and Hyundoo Shin. An efficient color representation for image retrieval. IEEE Transactions on image processing, 10(1):140-147, 2001." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 184, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 287, + 228 + ], + "type": "text", + "content": "[3] Bangalore S Manjunath, J-R Ohm, Vinod V Vasudevan, and Akio Yamada. Color and texture descriptors. IEEE Transactions on circuits and systems for video technology, 11(6):703-715, 2001." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 232, + 287, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 232, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 53, + 232, + 287, + 265 + ], + "type": "text", + "content": "[4] Luca Marchesotti, Naila Murray, and Florent Perronnin. Discovering beautiful attributes for aesthetic image analysis. International journal of computer vision, 113:246-266, 2015." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 267, + 286, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 267, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 267, + 286, + 300 + ], + "type": "text", + "content": "[5] Andrew J Elliot and Markus A Maier. Color psychology: Effects of perceiving color on psychological functioning in humans. Annual review of psychology, 65:95-120, 2014." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 286, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 286, + 336 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 286, + 336 + ], + "type": "text", + "content": "[6] Peter O'Donovan, Aseem Agarwala, and Aaron Hertzmann. Color compatibility from large datasets. In ACM SIGGRAPH 2011 papers, pages 1-12. 2011." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 338, + 286, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 286, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 286, + 393 + ], + "type": "text", + "content": "[7] Joon-Young Lee, Kalyan Sunkavalli, Zhe Lin, Xiaohui Shen, and In So Kweon. Automatic content-aware color and tone stylization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2470–2478, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 396, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 286, + 430 + ], + "type": "text", + "content": "[8] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In CVPR, pages 3677-3686, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 433, + 286, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 433, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 53, + 433, + 286, + 466 + ], + "type": "text", + "content": "[9] Emilie Baudin, François-Xavier Bucher, Laurent Chanas, and Frédéric Guichard. Dxomark objective video quality measurements. *Electronic Imaging*, 2020(9):166-1, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 468, + 286, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 286, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 286, + 501 + ], + "type": "text", + "content": "[10] Sabine E Susstrunk and Stefan Winkler. Color image quality on the internet. In _Internet imaging V_, volume 5304, pages 118-131. SPIE, 2003." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 504, + 286, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 286, + 536 + ], + "type": "text", + "content": "[11] Yang Yang, Jun Ming, and Nenghai Yu. Color image quality assessment based on ciede2000. Advances in Multimedia, 2012, 2012." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 540, + 286, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 540, + 286, + 584 + ], + "spans": [ + { + "bbox": [ + 48, + 540, + 286, + 584 + ], + "type": "text", + "content": "[12] Shu Kong, Xiaohui Shen, Zhe Lin, Radomir Mech, and Charless Fowlkes. Photo aesthetics ranking network with attributes and content adaptation. In ECCV, pages 662-679. Springer, 2016." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 586, + 287, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 586, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 48, + 586, + 287, + 631 + ], + "type": "text", + "content": "[13] Xuan Xu and Xin Li. Scan: Spatial color attention networks for real single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 634, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 634, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 634, + 286, + 677 + ], + "type": "text", + "content": "[14] Aleksandra Mojsilovic, H Hu, and Emina Soljanin. Extraction of perceptually important colors and similarity measurement for image matching, retrieval and analysis. IEEE Transactions on Image Processing, 11(11):1238-1248, 2002." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 286, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 714 + ], + "type": "text", + "content": "[15] Stephen E Palmer, Karen B Schloss, and Jonathan Sammartino. Visual aesthetics and human preference. Annual review of psychology, 64:77-107, 2013." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[16] Michael J Swain and Dana H Ballard. Color indexing. International journal of computer vision, 7(1):11-32, 1991." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 98, + 545, + 131 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 98, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 307, + 98, + 545, + 131 + ], + "type": "text", + "content": "[17] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. ECCV, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 133, + 545, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 133, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 307, + 133, + 545, + 156 + ], + "type": "text", + "content": "[18] Hossein Talebi and Peyman Milanfar. NIMA: Neural image assessment. TIP, 27(8):3998-4011, 2018." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 159, + 545, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 159, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 545, + 192 + ], + "type": "text", + "content": "[19] Daniel Cohen-Or, Olga Sorkine, Ran Gal, Tommer Leyvand, and Ying-Qing Xu. Color harmonization. In ACM SIGGRAPH 2006 Papers, pages 624-630. 2006." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 227 + ], + "type": "text", + "content": "[20] Vlad Hosu, Bastian Goldlucke, and Dietmar Saupe. Effective aesthetics prediction with multi-level spatially pooled features. In CVPR, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 230, + 545, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 264 + ], + "type": "text", + "content": "[21] Xin Lu, Zhe Lin, Hailin Jin, Jianchao Yang, and James Z. Wang. RAPID: Rating pictorial aesthetics using deep learning. In ACMMM, pages 457-466, 2014." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 266, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 300 + ], + "type": "text", + "content": "[22] Jian Ren, Xiaohui Shen, Zhe Lin, Radomir Mech, and David J Foran. Personalized image aesthetics. In ICCV, pages 638-647, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 545, + 346 + ], + "type": "text", + "content": "[23] Shuang Ma, Jing Liu, and Chang Wen Chen. A-lamp: Adaptive layout-aware multi-patch deep convolutional neural network for photo aesthetic assessment. In CVPR, pages 722-731, 2017." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 349, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 392 + ], + "type": "text", + "content": "[24] Kekai Sheng, Weiming Dong, Chongyang Ma, Xing Mei, Feiyue Huang, and Bao-Gang Hu. Attention-based multipatch aggregation for image aesthetic assessment. In ACMMM, 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 396, + 545, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 396, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 545, + 440 + ], + "type": "text", + "content": "[25] Hancheng Zhu, Leida Li, Jinjian Wu, Sicheng Zhao, Guiguang Ding, and Guangming Shi. Personalized image aesthetics assessment via meta-learning with bilevel gradient optimization. TCYB, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 442, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 442, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 442, + 545, + 475 + ], + "type": "text", + "content": "[26] Hui Zeng, Zisheng Cao, Lei Zhang, and Alan C Bovik. A unified probabilistic formulation of image aesthetic assessment. TIP, 29:1548-1561, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 479, + 545, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 512 + ], + "type": "text", + "content": "[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, pages 5148-5157, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 514, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 548 + ], + "type": "text", + "content": "[28] Dongyu She, Yu-Kun Lai, Gaoxiong Yi, and Kun Xu. Hierarchical layout-aware graph convolutional network for unified aesthetics assessment. In CVPR, pages 8475-8484, 2021." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 551, + 545, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 551, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 307, + 551, + 545, + 584 + ], + "type": "text", + "content": "[29] Shuai He, Yongchang Zhang, Rui Xie, Dongxiang Jiang, and Anlong Ming. Rethinking image aesthetics assessment: Models, datasets and benchmarks. *IJCAI*, 2022." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 586, + 545, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 586, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 307, + 586, + 545, + 631 + ], + "type": "text", + "content": "[30] Yunyu Shi, Youdong Ding, Ranran Zhang, and Jun Li. Structure and hue similarity for color image quality assessment. In 2009 International Conference on Electronic Computer Technology, pages 329-333. IEEE, 2009." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "type": "text", + "content": "[31] Peter O'Donovan, Aseem Agarwala, and Aaron Hertzmann. Collaborative filtering of color aesthetics. In Proceedings of the Workshop on Computational Aesthetics, pages 33-40, 2014." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "type": "text", + "content": "[32] Masashi Nishiyama, Takahiro Okabe, Imari Sato, and Yoichi Sato. Aesthetic quality classification of photographs based on color harmony. In CVPR 2011, pages 33-40. IEEE, 2011." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "text", + "content": "21846" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 445 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[33] Pei Lv, Jianqi Fan, Xixi Nie, Weiming Dong, Xiaoheng Jiang, Bing Zhou, Mingliang Xu, and Changsheng Xu. User-guided personalized image aesthetic assessment based on deep reinforcement learning. arXiv:2106.07488, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 121, + 287, + 153 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 121, + 287, + 153 + ], + "spans": [ + { + "bbox": [ + 48, + 121, + 287, + 153 + ], + "type": "text", + "content": "[34] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In CVPR. IEEE, 2012." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 158, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 158, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 48, + 158, + 287, + 190 + ], + "type": "text", + "content": "[35] Kuang-Yu Chang, Kung-Hung Lu, and Chu-Song Chen. Aesthetic critiques generation for photos. In ICCV, pages 3514-3523, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 227 + ], + "type": "text", + "content": "[36] Ritendra Datta, Jia Li, and James Z Wang. Algorithmic infor-ferencing of aesthetics and emotion in natural images: An exposition. In ICIP. IEEE, 2008." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 263 + ], + "type": "text", + "content": "[37] Wei Luo, Xiaogang Wang, and Xiaou Tang. Content-based photo quality assessment. In ICCV, pages 2206-2213. IEEE, 2011." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "type": "text", + "content": "[38] Xin Jin, Le Wu, Geng Zhao, Xiaodong Li, Xiaokun Zhang, Shiming Ge, Dongqing Zou, Bin Zhou, and Xinghui Zhou. Aesthetic attributes assessment of images. In ACMMM, pages 311-319, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 287, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 359 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 359 + ], + "type": "text", + "content": "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv:2103.14030, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "type": "text", + "content": "[40] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4794-4803, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 412, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 412, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 412, + 287, + 445 + ], + "type": "text", + "content": "[41] Jacob Gildenblat and contributors. Pytorch library for cam methods. https://github.com/jacobgil/pytorch-grad-cam, 2021." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "text", + "content": "21847" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TiDAL_ Learning Training Dynamics for Active Learning/cbb37e36-1eb0-42b7-b0b1-3b2793b016b9_content_list.json b/2023/TiDAL_ Learning Training Dynamics for Active Learning/cbb37e36-1eb0-42b7-b0b1-3b2793b016b9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b8cfb61fb6510aa4fa9d9c7c18bb583ee7114295 --- /dev/null +++ b/2023/TiDAL_ Learning Training Dynamics for Active Learning/cbb37e36-1eb0-42b7-b0b1-3b2793b016b9_content_list.json @@ -0,0 +1,1661 @@ +[ + { + "type": "text", + "text": "TiDAL: Learning Training Dynamics for Active Learning", + "text_level": 1, + "bbox": [ + 194, + 130, + 777, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seong Min Kye $^{1,\\dagger}$ Kwanghee Choi $^{2,\\dagger}$ Hyeongmin Byun $^{1}$ Buru Chang $^{3,*}$ $^{1}$ Hyperconnect $^{2}$ Carnegie Mellon University $^{3}$ Sogang University", + "bbox": [ + 187, + 179, + 785, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{harris,hyeongmin.byun}@hpcnt.com kwanghec@andrew.cmu.edu buru@sogang.ac.kr", + "bbox": [ + 143, + 219, + 825, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Active learning (AL) aims to select the most useful data samples from an unlabeled data pool and annotate them to expand the labeled dataset under a limited budget. Especially, uncertainty-based methods choose the most uncertain samples, which are known to be effective in improving model performance. However, previous methods often overlook training dynamics (TD), defined as the ever-changing model behavior during optimization via stochastic gradient descent, even though other research areas have empirically shown that TD provides important clues for measuring the data uncertainty. In this paper, we first provide theoretical and empirical evidence to argue the usefulness of utilizing the ever-changing model behavior rather than the fully trained model snapshot. We then propose a novel AL method, Training Dynamics for Active Learning (TiDAL), which efficiently predicts the training dynamics of unlabeled data to estimate their uncertainty. Experimental results show that our TiDAL achieves better or comparable performance on both balanced and imbalanced benchmark datasets compared to state-of-the-art AL methods, which estimate data uncertainty using only static information after model training.", + "bbox": [ + 75, + 300, + 473, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 661, + 209, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"There is a tide in the affairs of men. Which taken at the flood, leads on to fortune.\" — William Shakespeare", + "bbox": [ + 73, + 686, + 468, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Active learning (AL) [5, 31] aims to solve the real-world problem of selecting the most useful data samples from large-scale unlabeled data pools and annotating them to expand labeled data under a limited budget. Since the current deep neural networks are data-hungry, AL has increasingly gained attention in recent years. Existing AL methods can be divided into two mainstream categories: diversity- and uncertainty-based methods. Diversity-based methods [42, 14] focus on constructing a subset that fol", + "bbox": [ + 75, + 717, + 470, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/604e19a021cbe9bfdf754ce80ffa26e7b8dc303a11ccf08b5fcdb6a885d1b41e.jpg", + "image_caption": [ + "Figure 1: Our proposed TiDAL. TD of training samples $x$ may differ even if they converge to the same final predicted probability $p(y^{*}|x)$ (Upper row). Hence, we are motivated to utilize the readily available rich information generated during training, i.e., leveraging TD. We estimate TD of large-scale unlabeled data using a prediction module instead of tracking the actual TD of all the unlabeled samples to avoid the computational overhead (Lower row)." + ], + "image_footnote": [], + "bbox": [ + 500, + 267, + 893, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ows the target data distribution. Uncertainty-based methods [13, 6, 52] choose the most uncertain samples, which are known to be effective in improving model performance. Hence, the most critical question for the latter becomes, \"How can we quantify the data uncertainty?\"", + "bbox": [ + 496, + 703, + 890, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we leverage training dynamics (TD) to quantify data uncertainty. TD is defined as the ever-changing model behavior on each data sample during optimization via stochastic gradient descent. Recent studies [9, 29, 48, 47] have provided empirical evidence that TD provides important clues for measuring the contribution of each data sample to model performance improvement. Inspired by these studies, we argue that the data uncertainty", + "bbox": [ + 496, + 780, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution.", + "bbox": [ + 93, + 862, + 205, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 93, + 876, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was done while all authors were affiliated with Hyperconnect.", + "bbox": [ + 78, + 887, + 460, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22335", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of unlabeled data can be estimated with TD. However, most uncertainty-based methods quantify data uncertainty based on static information (e.g., loss [52] or predicted probability [45]) from a fully-trained model \"snapshot,\" neglecting the valuable information generated during training. We further argue that TD is more effective in separating uncertain and certain data than static information from a model snapshot captured after model training. In §3, we provide both theoretical and empirical evidence to support our argument that TD is a valuable tool for quantifying data uncertainty.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite its huge potential, TD is not yet actively explored in the domain of AL. This is because AL assumes a massive unlabeled data pool. Previous studies track TD only for the training data every epoch as it can be recorded easily during model optimization. On the other hand, AL targets a large number of unlabeled data, where tracking the TD for each unlabeled sample requires an impractical amount of computation (e.g., inference all the unlabeled samples every training epoch).", + "bbox": [ + 75, + 242, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we propose TiDAL (Training Dynamics for Active Learning), a novel AL method that efficiently quantifies the uncertainty of unlabeled data by estimating their TD. We avoid tracking the TD of large-scale unlabeled data every epoch by predicting the TD of unlabeled samples with a TD prediction module. The module is trained with the TD of labeled data, which is readily available during model optimization. During the data selection phase, we predict the TD of unlabeled data with the trained module to quantify their uncertainties. We efficiently obtain TD using the module, which avoids inferring all the unlabeled samples every epoch. Experimental results demonstrate that our TiDAL achieves better or comparable performance to existing AL methods on both balanced and imbalanced datasets. Additional analyses show that our prediction module successfully predicts TD, and the predicted TD is useful in estimating uncertainties of unlabeled data. Our proposed method are illustrated in Figure 1.", + "bbox": [ + 75, + 378, + 472, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions of our study: (1) We bridge the concept of training dynamics and active learning with the theoretical and experimental evidence that training dynamics is effective in estimating data uncertainty. (2) We propose a new method that efficiently predicts the training dynamics of unlabeled data to estimate their uncertainty. (3) Our proposed method achieves better or comparable performance on both balanced and imbalanced benchmark datasets compared to existing active learning methods. For reproducibility, we release the source code1.", + "bbox": [ + 75, + 651, + 472, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 818, + 217, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To better understand our proposed method, we first summarize key concepts, including uncertainty-based active", + "bbox": [ + 76, + 844, + 468, + 875 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learning, quantification of uncertainty, and training dynamics.", + "bbox": [ + 500, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Uncertainty-based active learning. In this work, we focus on uncertainty-based AL for multi-class classification problems. We define the predicted probabilities of the given sample $x$ for $C$ classes as:", + "bbox": [ + 498, + 142, + 893, + 203 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} = \\left[ p (1 | x), p (2 | x), \\dots , p (C | x) \\right] ^ {T} \\in [ 0, 1 ] ^ {C}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 214, + 890, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where we denote the true label of $x$ as $y$ and the classifier as $f$ . $\\mathcal{D}$ and $\\mathcal{D}_u$ denote a labeled dataset and an unlabeled data pool, respectively. The general cycle of uncertainty-based AL is in two steps: (1) train the target classifier $f$ on the labeled dataset $\\mathcal{D}$ and (2) select top- $k$ uncertain data samples from the unlabeled data pool $\\mathcal{D}_u$ . Selected samples are then given to the human annotators to expand the labeled dataset $\\mathcal{D}$ , cycling back to the first step.", + "bbox": [ + 498, + 244, + 893, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Quantifying uncertainty. The objective of this study is to establish a connection between the concept of TD and the field of AL. In order to clearly demonstrate the effectiveness of utilizing TD to quantify data uncertainty, we have employed two of the most prevalent and straightforward estimators, entropy [43] and margin [41], to measure data uncertainty in this paper. Entropy $H$ is defined as follows:", + "bbox": [ + 498, + 387, + 890, + 494 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nH (\\boldsymbol {p}) = - \\sum_ {c = 1} ^ {C} p (c | x) \\log p (c | x), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 505, + 890, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where the sample $x$ is from the unlabeled data pool $\\mathcal{D}_u$ . Entropy concentrates on the level of the model's confidence on the given sample $x$ and gets bigger when the prediction across the classes becomes uniform (i.e., uncertain). Margin $M$ measures the difference between the probability of the true label and the maximum of the others:", + "bbox": [ + 498, + 542, + 893, + 633 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nM (\\boldsymbol {p}) = p (y | x) - \\max _ {c \\neq \\hat {y}} p (c | x), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 647, + 890, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $y$ denotes the true label. The smaller the margin, the lower the model's confidence in the sample, so it can be considered uncertain. Both entropy and margin are computed with the predicted probabilities $p$ of the fully trained classifier $f$ , only taking the snapshot of $f$ into account.", + "bbox": [ + 498, + 681, + 890, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Defining training dynamics. Our TiDAL targets to leverage TD of unlabeled data to estimate their uncertainties. TD can be defined as any model behavior during optimization, such as the area under the margin between logit values of the target class and the other largest class [39] or the variance of the predicted probabilities generated at each epoch [47]. In this work, we define the TD $\\bar{p}^{(t)}$ as the area under the predicted probabilities of each data sample $x$ obtained", + "bbox": [ + 498, + 779, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1https://github.com/hyperconnect/TiDAL", + "bbox": [ + 94, + 886, + 395, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22336", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eac8b6fa1fbad8f46dffbf5a2823b54aa7e8300a9a3b224b270ac33614f68f33.jpg", + "image_caption": [ + "(a) Entropy Distribution", + "Figure 2: Score distribution after long-tailed training. We plot the marginal distributions using kernel density estimation (KDE). It is difficult to separate major (certain) and minor (uncertain) samples by the model snapshot-based scores (horizontal), unlike the TD-driven scores (vertical) that enable clearly separating the certain and uncertain samples." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 272, + 238 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/23a522c3282f3f5d58ab0d0f03fdf43768b9ed46a90b35ceb90b3f9da570de87.jpg", + "image_caption": [ + "(b) Margin Distribution" + ], + "image_footnote": [], + "bbox": [ + 277, + 92, + 467, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "during the $t$ time steps of optimizing the target classifier $f$ :", + "bbox": [ + 76, + 383, + 465, + 401 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} ^ {(i)} = \\left[ p ^ {(i)} (1 | x), p ^ {(i)} (2 | x), \\dots , p ^ {(i)} (C | x) \\right] ^ {T}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 410, + 468, + 429 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\bar {\\boldsymbol {p}} ^ {(t)} = \\left[ \\bar {p} ^ {(t)} (1 | x), \\bar {p} ^ {(t)} (2 | x), \\dots , \\bar {p} ^ {(t)} (C | x) \\right] ^ {T} \\\\ = \\sum_ {\\tau} \\boldsymbol {p} ^ {(\\tau)} \\Delta \\tau \\simeq \\sum_ {i = 1} ^ {t} \\boldsymbol {p} ^ {(i)} / t, \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 431, + 468, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{p}^{(i)}$ is the predicted probabilities of a target classifier $f$ at the $i$ -th time step. $\\Delta \\tau$ is the unit time step to normalize the predicted probabilities. For simplicity, we record $\\pmb{p}^{(i)}$ every epoch and choose $\\Delta \\tau = 1 / t$ , namely, averaging the predicted probabilities during $t$ epochs [47, 46]. The TD $\\bar{\\pmb{p}}^{(t)}$ takes all the predicted probabilities during model optimization into account. Hence, it encapsulates the overall tendency of the model during $t$ epochs of optimization, avoiding being solely biased towards the snapshot of $\\pmb{p}^{(t)}$ in the final epoch $t$ .", + "bbox": [ + 75, + 492, + 468, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Is TD Useful for Quantifying Uncertainty?", + "text_level": 1, + "bbox": [ + 76, + 657, + 455, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we provide empirical and theoretical evidence to support our argument: $TD$ is more effective in separating uncertain data from certain data than the model snapshot, where the latter is often utilized to quantify data uncertainty in previous works [52, 45].", + "bbox": [ + 75, + 684, + 468, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivating Observation", + "text_level": 1, + "bbox": [ + 76, + 770, + 297, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Settings. We aim to observe and compare the behavior of TD and the model snapshot for different sample difficulties. However, it is nontrivial to directly measure sample-wise difficulty, inhibiting the quantitative analysis of data uncertainty. To avoid this, we borrow the theoretical and empirical results of long-tailed visual recognition [33, 8, 19]: it is hard for the deep neural network-based model to train with", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fewer samples. Hence, we regard major and minor class samples to contain many certain and uncertain samples for the model, respectively. We train the target classifier $f$ on the long-tailed dataset during $T$ epochs to obtain the TD and the model snapshot. We apply both approaches to the common estimators, entropy and margin. We denote entropy and margin scores from the model snapshot as $H$ and $M$ . In opposition, we denote the TD-driven scores as $\\bar{H}$ and $\\bar{M}$ . More details and discussions are described in Appendix B.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Results. Figure 2 shows the distribution the scores calculated with TD $(x$ -axis) and model snapshot $(y$ -axis). We can observe that scores from TD $(\\bar{H}, \\bar{M})$ successfully separate the major and the minor class samples, whereas scores from the model snapshot $(H, M)$ fail to do so. We conclude that compared to model snapshots, TD is more helpful in separating uncertain samples from certain samples.", + "bbox": [ + 496, + 244, + 890, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Theoretical Evidence", + "text_level": 1, + "bbox": [ + 500, + 357, + 697, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1. (Informal) Under the LE-SDE framework [54], with the assumption of local elasticity [17], certain samples and uncertain samples reveal different $TD$ ; especially, certain samples converge quickly than uncertain samples.", + "bbox": [ + 496, + 380, + 890, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The above theorem discusses different model behaviors depending on the difficulty of the sample. Compared to the uncertain sample, the certain sample has the same class samples nearby, which is the fundamental idea of level set estimation [22] and nearest neighbor [36] literature. We suspect that, due to the local elasticity of deep nets, samples close by have a bigger impact on the certain sample, hence changing its predicted probability more rapidly. As the certain sample is quicker to converge, its TD is larger than that of the uncertain sample. Intuitively, slower to train, struggling the classifier is to learn, hence TD capturing the uncertainty in the classifier's perspective.", + "bbox": [ + 496, + 460, + 890, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 2. (Informal) Estimators such as Entropy (Equation 10) and Margin (Equation 11) successfully capture the difference of TD between easy and hard samples even for the case where it cannot be distinguished via the predicted probabilities of the model snapshot.", + "bbox": [ + 496, + 647, + 890, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The above theorem discusses the validity of entropy and margin on whether they can successfully differentiate between two samples of different TD but with the same final prediction. With Theorem 1, one can conclude that the common estimators' scores calculated with TD are effective in capturing the data uncertainty. Due to the space constraints, we provide the details of the above results in Appendix A.", + "bbox": [ + 496, + 727, + 890, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Utilizing TD for Active Learning", + "text_level": 1, + "bbox": [ + 500, + 844, + 797, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As tracking the TD of all the unlabeled data is computationally infeasible, we devise an efficient method to es", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22337", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "estimate the TD of unlabeled samples. We train the module that directly predicts the TD of each sample by feeding the training samples, where its TD are freely available during training. Then, based on the predicted TD of each unlabeled sample, we use the common estimators, entropy or margin, to determine which sample is the most uncertain so that human annotators can label it. Hence, in this section, we describe the details of the module that estimates TD (§4.1) and how to train the module (§4.2). Finally, calculating the uncertainties using the module predictions for active learning is illustrated (§4.3).", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Training Dynamics Prediction Module", + "text_level": 1, + "bbox": [ + 76, + 263, + 410, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned, it is not computationally feasible to track TD for the large-scale unlabeled data as it requires model inference on all the unlabeled data every training epoch. Thus, we propose the TD prediction module $m$ to efficiently predict the TD of unlabeled data at the $t$ -th epoch. Being influenced by the previous studies [11, 52, 45, 25] that use additional modules to predict useful values such as loss or confidence by the target model outputs, multi-scale feature maps are aggregated and passed into our TD prediction module. The module produces the $C$ -dimensional predictions:", + "bbox": [ + 75, + 287, + 468, + 438 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {p}} _ {m} ^ {(t)} = \\left[ \\tilde {p} _ {m} ^ {(t)} (1 | x), \\dots , \\tilde {p} _ {m} ^ {(t)} (C | x) \\right] ^ {T} \\in [ 0, 1 ] ^ {C} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 441, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "estimating the actual TD $\\bar{\\pmb{p}}^{(t)}$ of the given sample $x$ in Equation 5. TD prediction module is jointly trained with the target classifier using a handful of parameters, having a negligible computational cost during training. The detailed architecture of the module is described in Appendix C.", + "bbox": [ + 75, + 468, + 468, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Even though the architecture is similar to previous works [52, 45, 25], we observed that ours were much more stable during optimization and easier to train. We suspect that it is due to the target task difference; previous works trained the module that outputs only a single value via regression, whereas our module outputs $C$ -dimensional probability distribution, which is similar to the main task of classifying images.", + "bbox": [ + 75, + 545, + 468, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Training Objectives", + "text_level": 1, + "bbox": [ + 76, + 671, + 266, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train the target classifier $f$ at the $t$ -th epoch, we use the cross-entropy loss function $\\mathcal{L}_{\\mathrm{target}}$ on the predicted probability $\\pmb{p}^{(t)}$ and a one-hot encoded vector $\\pmb{y} \\in \\{0,1\\}^C$ of the true label $y$ :", + "bbox": [ + 75, + 695, + 468, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t a r g e t}} = \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {p} ^ {(t)}, \\boldsymbol {y}\\right) = - \\log p ^ {(t)} (y | x). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 760, + 468, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meanwhile, the prediction module $m$ learns the TD of a sample $x$ by minimizing the Kullback-Leibler (KL) divergence between the predicted TD $\\tilde{p}_m^{(t)}$ and the actual TD $\\bar{p}^{(t)}$ :", + "bbox": [ + 76, + 784, + 468, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {m o d u l e}} = \\mathcal {L} _ {\\mathrm {K L}} (\\bar {\\boldsymbol {p}} ^ {(t)} | | \\tilde {\\boldsymbol {p}} _ {m} ^ {(t)}) \\\\ = \\sum_ {c = 1} ^ {C} \\bar {p} ^ {(t)} (c | x) \\log \\left(\\frac {\\bar {p} ^ {(t)} (c | x)}{\\tilde {p} _ {m} ^ {(t)} (c | x)}\\right). \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 835, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final objective function of our proposed method is defined as follows:", + "bbox": [ + 500, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {t a r g e t}} + \\lambda \\mathcal {L} _ {\\text {m o d u l e}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 133, + 890, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is a balancing factor to control the effect of $\\mathcal{L}_{\\mathrm{module}}$ during model training.", + "bbox": [ + 498, + 159, + 890, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Quantifying Uncertainty with TD", + "text_level": 1, + "bbox": [ + 500, + 198, + 794, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We argue that uncertain samples can be effectively distinguished from unlabeled data using the predicted TD. To verify the effectiveness of leveraging TD, we feed the predicted TD to entropy and margin (§2) by replacing snapshot probability $\\pmb{p}$ with the predicted TD $\\bar{\\pmb{p}}$ . We choose these estimators as they are widely used for quantifying uncertainty. We feed $\\bar{\\pmb{p}}$ , replacing $\\pmb{p}$ , to the entropy $\\bar{H}$ :", + "bbox": [ + 498, + 222, + 890, + 328 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {H} (\\bar {\\boldsymbol {p}}) = - \\sum_ {c = 1} ^ {C} \\bar {p} (c | x) \\log \\bar {p} (c | x). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 337, + 890, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Entropy $\\bar{H}$ is maximized when $\\bar{p}$ is uniform, i.e., the sample is uncertain for the target classifier. Margin $\\bar{M}$ is also similarly employed:", + "bbox": [ + 498, + 375, + 890, + 421 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {M} (\\bar {\\boldsymbol {p}}) = \\bar {p} (\\hat {y} | x) - \\max _ {c \\neq \\hat {y}} \\bar {p} (c | x). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 431, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since we do not have true labels of unlabeled samples, we use the predicted labels $\\hat{y}$ of the target classifier instead of the true labels. There are several possible variants of $\\bar{M}$ depending on the definition of $\\hat{y}$ . We conduct experiments to compare $\\bar{M}$ with its variants. The experimental details and results are in Appendix D.4.", + "bbox": [ + 496, + 465, + 890, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At the data selection phase, we use the predicted TD $\\tilde{\\pmb{p}}_m^{(T)}$ instead of the actual TD $\\bar{\\pmb{p}}^{(T)}$ as in Equation 10 & 11 to estimate the TD-driven uncertainties of the unlabeled sample $x$ at the final epoch $T$ . By using the estimated uncertainty with the predicted TD, we select the most informative samples for model training.", + "bbox": [ + 498, + 556, + 890, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 661, + 633, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we experimentally verify the effectiveness of our method, TiDAL, which utilizes the estimated training dynamics from the prediction module to discern uncertain samples from unlabeled data. We describe the detailed settings and the baseline methods for our experiments (§5.1) and show the results on both balanced (§5.2) and imbalanced datasets (§5.3). We further analyze whether the TD prediction module is effective for AL performance and can successfully estimate the TD (§5.4). We end the section by discussing the potential limitations of our method (§5.5).", + "bbox": [ + 496, + 686, + 890, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 847, + 689, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. To assess the performance of our proposed method and baseline methods, we conduct experiments on", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22338", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d9f583e97a79654ab095d0371233dc233538617bb324361ac5209f944045735e.jpg", + "image_caption": [ + "Figure 3: Averaged relative accuracy improvement curves and their $95\\%$ confidence interval (shaded) of AL methods over the number of labeled samples on balanced datasets. TiDAL $(\\bar{H})$ and TiDAL $(\\bar{M})$ denote the performance of TiDAL when with entropy $\\bar{H}$ and margin $\\bar{M}$ as the data uncertainty estimation strategy, respectively." + ], + "image_footnote": [], + "bbox": [ + 107, + 89, + 359, + 281 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8034f319a656da524e8075fec458aec9bfba223f71708b8a784ab8590b040a56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 89, + 609, + 281 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b20e847e82013244bf3d032003e0fee0f58fbce49a4f5b44e908178f9c452e72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 90, + 859, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the following five datasets: CIFAR10/100 [27], FashionMNIST [51], SVHN [34], and iNaturalist2018 [50]. Since CIFAR and FashionMNIST are both balanced, we further modify them to simulate the data imbalance in the real world, following the previous long-tail visual recognition studies [8, 33, 56, 19]. The imbalance ratio is defined as $N_{\\mathrm{max}} / N_{\\mathrm{min}}$ where $N$ is the number of samples in each class. We make two variants with data imbalance ratios 10 and 100 for each dataset. Unlike the above, SVHN and iNaturalist18 are already imbalanced. Especially, iNaturalist2018 is commonly chosen to demonstrate how methods work in imbalanced real-world settings. The dataset statistics are summarized in Appendix D.", + "bbox": [ + 75, + 353, + 472, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. For a fair comparison, we compare our TiDAL with the following baselines which train a target classifier with only labeled data. Random sampling: a simple baseline that randomly selects data samples from the unlabeled dataset. Entropy sampling [43]: an uncertainty-based method that selects data samples based on the maximum entropy. BALD [13]: an uncertainty-based method that selects data samples based on the mutual information between the model prediction and the posterior. CoreSet [42]: a diversity-based method that selects representative data samples covering all data through a minimum radius. LLoss [52]: an uncertainty-based method that learns to estimate the errors of the predictions (loss) made by the learner and select data samples based on the predicted loss. CAL [55]: recent work on using TD, gathering sample-wise TD information on whether the classifier was consistently correct or not during training. CAL splits the samples into two classes by applying a heuristic threshold to the TD information to train a binary classifier that outputs uncertainty score. To verify the effectiveness of TiDAL, we further compare it with the two semi-supervised AL methods, VAAL [45] and TA-VAAL [25] in Appendix D.3. Note that", + "bbox": [ + 75, + 568, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "these methods further utilize unlabeled data for training the selection module, thus it is unfair for our TiDAL.", + "bbox": [ + 498, + 353, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Active learning setting. We follow the same setting from [6, 52] for the detailed AL settings. For the initial step, we randomly select initial samples to be annotated from the unlabeled dataset, where we use them to train the initial target classifier. Then, we obtain a random subset from the unlabeled data pool $\\mathcal{D}_u$ to choose the top- $k$ samples based on the criterion of each method, where those samples will be annotated. We repeat the above cycle, training a classifier from scratch from the continuously expanding labeled set.", + "bbox": [ + 498, + 404, + 892, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. For a fair comparison, we use the same backbone network ResNet-18 [18] except for iNaturalist2018, where we use ResNet-50 [18] pretrained on ImageNet [12]. All models are trained with SGD optimizer with momentum 0.9, weight decay $5 \\cdot 10^{-4}$ , and learning rate (LR) decay of 0.1. For CIFAR10/100 and SVHN, we train the model for 200 epochs with an initial LR of 0.1 and decay at epoch 160. For FashionMNIST, 100 epochs with an initial LR of 0.1 and decay at epoch 80. For iNaturalist2018, 50 epochs with an initial LR of 0.01 and decay at epoch 40. For CIFAR10/100, SVHN and FashionMNIST, we set the batch size and the unlabeled subset size to be 128 and $10^{4}$ , respectively. For iNaturalist2018, which is much larger than other datasets, we set the batch size and the unlabeled subset size to 256 and $10^{6}$ , respectively. We set the balancing factor to 1.0.", + "bbox": [ + 496, + 561, + 893, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation details. To compare with other state-of-the-art baselines, we show the average accuracy and $95\\%$ confidence interval with three trials. We mainly compare the model performances with relative accuracy improvement to random sampling, demonstrating how much it improves", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22339", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1181e6314028dd29509618f6ab64eb40c97a6c7a6371fd6d3f7e069ee71d9386.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 88, + 361, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a4ea931fe8bad4459808ffbaddfa44848cb21a44fd233cc67f9951f974108930.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 88, + 609, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7501908db5eb3fc363adb5751b189283cae593b9f0b1da9b67f4dc63629bda90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 89, + 861, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f87e300182c62cb525f75ed87fbd59409b5395407b9d904ffb59196eaadf43f7.jpg", + "image_caption": [ + "Figure 4: Averaged relative accuracy improvement curves and their $95\\%$ confidence interval (shaded) of AL methods over the number of labeled samples on synthetically imbalanced datasets. We use the imbalance ratio (IR) of 10 and 100 on CIFAR10, CIFAR100, and FashionMNIST." + ], + "image_footnote": [], + "bbox": [ + 106, + 285, + 357, + 476 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/27c6d797034dd96d42bc087dd208c14a39ff9acb99b61c61347020cba1712be1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 285, + 609, + 476 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b8758b58439a1df2c3596da18174104d72945c87fb13f0f2a81b9f952175895a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 285, + 859, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "upon the naive approach on each cycle. Additionally, absolute accuracy is also plotted in Appendix D.", + "bbox": [ + 75, + 547, + 468, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results on Balanced Datasets", + "text_level": 1, + "bbox": [ + 76, + 588, + 339, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3 and 10 compare our TiDAL against the state-of-the-art methods on various balanced datasets: CIFAR10, CIFAR100, and FashionMNIST. For all the datasets, the two variants of TiDAL outperform all the baselines at all AL cycles except for LLoss, which shows better improvement than TiDAL $(\\bar{M})$ on CIFAR10 with an imbalance ratio of 100. Nonetheless, our TiDAL achieves the best final performance compared to all the baselines. CAL, which uses training dynamics, generally underperforms compared to others. We suspect that CAL is sensitive to its threshold hyperparameter.", + "bbox": [ + 75, + 611, + 470, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Results on Imbalanced Datasets", + "text_level": 1, + "bbox": [ + 76, + 786, + 357, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetically imbalanced datasets. Similar to the above, Figure 4, 9, and 11 shows the performance improvements on the synthetically imbalanced datasets with the two imbalance ratios, 10 and 100. Except for the CIFAR10 with an imbalance ratio of 100, our methods show superb performance across all the imbalanced settings. TiDAL per", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "forms especially well with a small variance in imbalanced CIFAR100, where the number of classes is the largest. In imbalanced FashionMNIST, the performance quickly rises to 2.5k labeled images and then saturates. This implies that FashionMNIST is easier than other datasets, and needs to focus more on the early training steps to compare with other models. TiDAL also shows overall better performance on FashionMNIST, especially in the early steps.", + "bbox": [ + 496, + 547, + 893, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real-world imbalanced datasets. Figure 5 and 10 shows evaluation results on real-world imbalanced datasets. For iNaturalist2018, which is the large-scale long-tailed classification dataset, TiDAL shows outstanding performance compared to other methods. For SVHN, TiDAL shows the best improvements with low variance as the number of labeled images increases except for the initial stage. LLoss shows outstanding performance only in the initial stage, where we presume that the loss prediction module of LLoss acts as a regularizer during model optimization.", + "bbox": [ + 496, + 686, + 893, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Analysis on the TD Prediction Module", + "text_level": 1, + "bbox": [ + 498, + 847, + 831, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of the TD prediction module. In order to verify the efficacy of using the predicted TD $\\tilde{p}_m$ , we con", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22340", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e1b6e5833c32890dc326cc2e42ed900d4986818fbc0b271c334d6ff9c0932fb8.jpg", + "image_caption": [ + "Figure 5: Averaged relative accuracy improvement curves and its $95\\%$ confidence interval (shaded) of AL methods over the number of labeled samples on real-world imbalanced datasets: iNaturalist2018 and SVHN. For SVHN, LLoss shows a substantial improvement of $20.02\\% \\pm 6.77\\%$ at the initial phase (1k), but we clip the plot to show the performance afterward more clearly." + ], + "image_footnote": [], + "bbox": [ + 217, + 89, + 483, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2686bdbb634a9f14026082c31130e56d5441311031f45a813b9c7f9978bf7e43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 89, + 751, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8d4ce1cc952aeb5a31d048225d4897d73ce8aa1770257ff510568f117e73fb59.jpg", + "image_caption": [ + "Figure 6: Ablation test results. $\\bar{H} (\\tilde{\\pmb{p}}_m)$ and $\\bar{M} (\\tilde{\\pmb{p}}_m)$ use the predicted TD $\\tilde{\\pmb{p}}_m$ of the prediction module $m$ . In contrast, $H(\\pmb {p})$ and $M(\\pmb {p})$ use the predicted probability of the model snapshot $\\pmb{p}$ . TD shows better performance than the model snapshot, implying that TD is better at quantifying data uncertainty." + ], + "image_footnote": [], + "bbox": [ + 81, + 375, + 318, + 578 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/90395acbbb13988562b801aed1ebebd2b906b556b710cc159dab0d0d644bd9b4.jpg", + "image_caption": [ + "Figure 7: KL divergence scores of the actual TD $\\bar{\\pmb{p}}^{(T)}$ with the predicted TD $\\tilde{\\pmb{p}}_m^{(t)}$ and the predicted probability of the model snapshot $\\pmb{p}^{(t)}$ , respectively, during model optimization. Our predicted TD can accurately approximate the actual TD." + ], + "image_footnote": [], + "bbox": [ + 321, + 375, + 552, + 577 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a0a8388e1446991ed393ec952424c5f38a90460fcd543e84cf9adb9e0c90bfb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 386, + 875, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "duct an ablation test that compares the performance between when using and not using the TD prediction module $m$ . Figure 6 shows the results on balanced CIFAR10/100. We observe that $\\bar{H}(\\tilde{\\pmb{p}}_m)$ and $\\bar{M}(\\tilde{\\pmb{p}}_m)$ using the predicted TD $\\tilde{\\pmb{p}}_m$ to estimate the data uncertainty significantly outperform the methods $H(\\pmb{p})$ and $M(\\pmb{p})$ that use only the final predicted probabilities $\\pmb{p}$ of the target classifier $f$ , showing better performance in the whole training cycle. Even $M(\\pmb{p})$ shows temporary improvement in earlier steps on CIFAR100, $\\bar{H}(\\tilde{\\pmb{p}}_m)$ and $\\bar{M}(\\tilde{\\pmb{p}}_m)$ maintain stable improvement, eventually winning over $M(\\pmb{p})$ . This indicates that the predicted TD $\\tilde{\\pmb{p}}_m$ of the TD prediction module $m$ produces better data uncertainty estimation than the predicted", + "bbox": [ + 75, + 691, + 472, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "probability $\\pmb{p}$ of the target classifier $f$ .", + "bbox": [ + 500, + 691, + 751, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Predictive performance of the TD prediction module. We verify whether the TD prediction module $m$ accurately predicts the actual TD $\\bar{p}$ . Its prediction performance is crucial as we use the predicted TD $\\tilde{\\boldsymbol{p}}_m$ of the module $m$ to quantify uncertainties of unlabeled data. Using the KL divergence $\\mathcal{L}_{\\mathrm{KL}}$ , we analyze that the predicted TD $\\tilde{\\boldsymbol{p}}_m$ converges to the actual TD $\\bar{p}$ at the data selection phase. We calculate $\\mathcal{L}_{\\mathrm{KL}}(\\bar{\\boldsymbol{p}}^{(T)}||\\tilde{\\boldsymbol{p}}_m^{(t)})$ and compare it with $\\mathcal{L}_{\\mathrm{KL}}(\\bar{\\boldsymbol{p}}^{(T)}||\\boldsymbol{p}^{(t)})$ which is set as a baseline computed with the actual TD $\\bar{p}$ and the predicted probabilities $\\boldsymbol{p}$ (snapshot) of the target classifier $f$ . In this analysis, we use the bal", + "bbox": [ + 496, + 731, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22341", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "anced CIFAR10 where the sample-wise averaged KL divergence scores are computed on the test set. Figure 7 shows that the final predicted TD successfully approximates the actual TD, while the predicted probability is highly different from the actual TD. We conclude that the TD prediction module $m$ can produce the TD efficiently, leading to performance improvement, and the predicted TD acts as a better approximation of the actual TD than the predicted probability of a model snapshot captured at each epoch.", + "bbox": [ + 75, + 90, + 470, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Limitations", + "text_level": 1, + "bbox": [ + 76, + 238, + 202, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We found two potential limitations of our TiDAL derived from the fact that it relies on the outputs of the target classifier to compute the TD. First, TiDAL is designed only for classification tasks, and thus it cannot be applied to AL targeting other tasks, such as regression [10, 15]. Second, TiDAL is highly influenced by the performance of the target classifier, especially when the target classifier wrongly classifies the hard negative samples with a high confidence during model optimization. These samples can be treated as certain samples (i.e. will not be selected for annotation) because they have low estimated uncertainties from the predicted TD, even though the target classifier fails to predict the true label of the samples correctly. As a future work, we will study extending our TiDAL in the task-agnostic ways with a safeguard combating the wrongly classified samples.", + "bbox": [ + 75, + 262, + 470, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Related Work", + "text_level": 1, + "bbox": [ + 76, + 503, + 218, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1. Active Learning", + "text_level": 1, + "bbox": [ + 76, + 529, + 238, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AL methods target to construct a dataset with the most useful samples based on the assumption that each sample has different importance in model training [40]. Two mainstream AL approaches exist for efficiently querying the unlabeled data: pool-based methods [31, 52, 45] use various ways to extract samples from an unlabeled data pool effectively, and synthesis-based methods [1, 58, 49] generate informative samples for the model. Pool-based methods can be roughly divided based on query strategies: uncertainty-based [13, 52, 45, 20] and diversity-based [42, 14, 38] methods, where some methods use the hybrid of both [4, 44, 25]. Uncertainty-based methods focus on finding which samples would be the most uncertain for the model, whereas diversity-based methods aim to construct a subset of representative samples of the input distribution. Our proposed method, TiDAL, lies in uncertainty-based methods. The significant difference between TiDAL and previous uncertainty-based methods is that TiDAL estimates data uncertainty using TD that contains additional hints generated during model training. In contrast, the previous methods leverage only static information (e.g., loss [52, 20] and predicted probabilities [13, 45, 25]) obtained by a model snapshot at the data selection phase.", + "bbox": [ + 75, + 553, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Training Dynamics", + "text_level": 1, + "bbox": [ + 500, + 90, + 684, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TD focuses on how deep neural networks are optimized under back-propagation-based stepwise weight updates. Many studies try to understand how gradient descent can effectively obtain the global minimum by analyzing the loss landscape of neural networks [24, 32] or its loss trajectory [3]. Some also import alternative models that are more mathematically approachable to analyze, such as neural tangent kernels [21], deep Gaussian processes [30], or stochastic differential equations [54]. On the other hand, the phenomenological and practical viewpoint of TD also exists. [48] coin the term Forgetting Dynamics to assert that unforgettable samples are often less helpful, and [9] show that the model could prefer samples that are often wrongly predicted throughout model training. TD is also commonly used in noisy label literature to find potential noisy labels as they tend to fit later on model training [2, 39] or locate samples that can be relabeled correctly [46]. Furthermore, [57] calculate the Dynamic Instance Hardness score by monitoring losses of each sample or whether the prediction gets flipped so that higher scored samples can be prioritized for curriculum learning, and [23] feed the loss history to the auxiliary neural network to mediate the curriculum for training. [29] also introduce temporal ensembling for semi-supervised learning, where the model fits towards averaged probability outputs. [47, 37] devise Data Maps to inspect datasets with two TD measures; confidence and its variability across epochs on the true class prediction. [55] further extend the Data Maps for AL, whether the target classifier was consistently correct or not during training. The proposed method splits the labeled samples by applying a heuristic threshold on the level of consistency to train a binary classifier that is trained to discern uncertain samples. Even though the work, similar to ours, also utilizes TD, it relies on empirical observations and heuristic choices to separate the certain and uncertain samples. In this study, we link the concept of TD to AL with both empirical and theoretical results to estimate the uncertainty of unlabeled samples, which is often neglected in previous TD studies.", + "bbox": [ + 496, + 114, + 893, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 707, + 617, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel active learning method, Training Dynamics for Active Learning (TiDAL), by linking the concept of training dynamics to active learning. We provide motivating observations and theoretical evidence for using training dynamics to estimate the uncertainty of unlabeled data. Since tracking the training dynamics of large-scale unlabeled data is infeasible, TiDAL utilizes a training dynamics prediction module to efficiently predict the training dynamics of the unlabeled data. Based on the predicted training dynamics, TiDAL quantifies data uncertainty using the common uncertainty estimators: entropy and mar", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22342", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gin. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness of our method, surpassing the existing state-of-the-art active learning methods. We further analyze that using our training dynamics prediction module is effective and the module successfully predicts the TD of unlabeled data.", + "bbox": [ + 75, + 90, + 472, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 198, + 173, + 214 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dana Angluin. Queries and concept learning. Machine learning, 2(4):319-342, 1988. 8", + "[2] Eric Arazo, Diego Ortego, Paul Albert, Noel O'Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In International conference on machine learning, pages 312-321. PMLR, 2019. 8", + "[3] Sanjeev Arora, Nadav Cohen, Noah Golowich, and Wei Hu. A convergence analysis of gradient descent for deep linear neural networks. In International Conference on Learning Representations, 2018. 8", + "[4] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671, 2019.8", + "[5] Les E Atlas, David A Cohn, and Richard E Ladner. Training connectionist networks with queries and selective sampling. In Advances in neural information processing systems, pages 566-573. CiteSeer, 1990. 1", + "[6] William H Beluch, Tim Genewein, Andreas Nurnberger, and Jan M Kohler. The power of ensembles for active learning in image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9368-9377, 2018. 1, 5", + "[7] Javad Zolfaghari Bengar, Joost van de Weijer, Laura Lopez Fuentes, and Bogdan Raducanu. Class-balanced active learning for image classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1536-1545, 2022. 15", + "[8] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in Neural Information Processing Systems, 32:1567-1578, 2019. 3, 5, 15", + "[9] Haw-Shiuan Chang, Erik Learned-Miller, and Andrew McCallum. Active bias: Training more accurate neural networks by emphasizing high variance samples. Advances in Neural Information Processing Systems, 30, 2017. 1, 8", + "[10] David Cohn, Les Atlas, and Richard Ladner. Improving generalization with active learning. Machine learning, 15(2):201-221, 1994. 8", + "[11] Charles Corbière, Nicolas Thome, Avner Bar-Hen, Matthieu Cord, and Patrick Pérez. Addressing failure prediction by learning model confidence. Advances in Neural Information Processing Systems, 32, 2019. 4", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + ], + "bbox": [ + 78, + 223, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International Conference on Machine Learning, pages 1183-1192. PMLR, 2017. 1, 5, 8", + "[14] Daniel Gissin and Shai Shalev-Shwartz. Discriminative active learning. arXiv preprint arXiv:1907.06347, 2019. 1, 8", + "[15] Jia Gong, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Meta agent teaming active learning for pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079-11089, 2022. 8", + "[16] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 15", + "[17] Hangfeng He and Weijie Su. The local elasticity of neural networks. In International Conference on Learning Representations, 2019. 3, 12", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 15", + "[19] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6626-6636, 2021. 3, 5, 15", + "[20] Siyu Huang, Tianyang Wang, Haoyi Xiong, Jun Huan, and Dejing Dou. Semi-supervised active learning with temporal output discrepancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3447-3456, 2021. 8", + "[21] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 8", + "[22] Heinrich Jiang, Been Kim, Melody Guan, and Maya Gupta. To trust or not to trust a classifier. Advances in neural information processing systems, 31, 2018. 3, 12", + "[23] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International Conference on Machine Learning, pages 2304–2313. PMLR, 2018. 8", + "[24] Kenji Kawaguchi. Deep learning without poor local minima. Advances in neural information processing systems, 29, 2016. 8", + "[25] Kwanyoung Kim, Dongwon Park, Kwang In Kim, and Se Young Chun. Task-aware variational adversarial active learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8166-8175, 2021. 4, 5, 8, 16", + "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 15", + "[27] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 5" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22343", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Alex Krizhevsky, Vinod Nair, and Geofrey Hinton. The CIFar-10 dataset. http://www.cs.toronto.edu/kriz/cifar.html, 55(5), 2014. 15", + "[29] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In International Conference on Learning Representations (ICLR). OpenReview.net, 2017. 1, 8", + "[30] Jaehoon Lee, Yasaman Bahri, Roman Novak, Samuel S Schoenholz, Jeffrey Pennington, and Jascha Sohl-Dickstein. Deep neural networks as gaussian processes. In International Conference on Learning Representations, 2018. 8", + "[31] David D Lewis and William A Gale. A sequential algorithm for training text classifiers. In Proceedings of the 17th annual international ACM SIGIR conference on Research and development in information retrieval, pages 3-12, 1994. 1, 8", + "[32] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018. 8", + "[33] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 3, 5, 15", + "[34] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 5", + "[35] Curtis Northcutt, Lu Jiang, and Isaac Chuang. Confident learning: Estimating uncertainty in dataset labels. Journal of Artificial Intelligence Research, 70:1373-1411, 2021. 15", + "[36] Nicolas Papernot and Patrick McDaniel. Deep k-nearest neighbors: Towards confident, interpretable and robust deep learning. arXiv preprint arXiv:1803.04765, 2018. 3, 12", + "[37] Seo Yeon Park and Cornelia Caragea. A data cartography based mixup for pre-trained language models. arXiv preprint arXiv:2205.03403, 2022. 8", + "[38] Amin Parvaneh, Ehsan Abbasnejad, Damien Teney, Gholam-reza Reza Haffari, Anton van den Hengel, and Javen Qinfeng Shi. Active learning by feature mixing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12237-12246, 2022. 8", + "[39] Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. Identifying unlabeled data using the area under the margin ranking. Advances in Neural Information Processing Systems, 33:17044-17056, 2020. 2, 8, 15", + "[40] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning. ACM Computing Surveys (CSUR), 54(9):1-40, 2021. 8", + "[41] Dan Roth and Kevin Small. Margin-based active learning for structured output spaces. In European Conference on Machine Learning, pages 413-424. Springer, 2006. 2", + "[42] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 1, 5, 8" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 2, 5", + "[44] Changjian Shui, Fan Zhou, Christian Gagné, and Boyu Wang. Deep active learning: Unified and principled method for query and training. In International Conference on Artificial Intelligence and Statistics, pages 1308-1318. PMLR, 2020. 8", + "[45] Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell. Variational adversarial active learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5972-5981, 2019. 2, 3, 4, 5, 8, 16", + "[46] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In International Conference on Machine Learning, pages 5907-5915. PMLR, 2019. 3, 8", + "[47] Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A Smith, and Yejin Choi. Dataset cartography: Mapping and diagnosing datasets with training dynamics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9275-9293, 2020. 1, 2, 3, 8", + "[48] Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J Gordon. An empirical study of example forgetting during deep neural network learning. In International Conference on Learning Representations, 2018. 1, 8, 15", + "[49] Toan Tran, Thanh-Toan Do, Ian Reid, and Gustavo Carneiro. Bayesian generative active deep learning. In International Conference on Machine Learning, pages 6295-6304. PMLR, 2019. 8", + "[50] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 5", + "[51] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 5", + "[52] Donggeun Yoo and In So Kweon. Learning loss for active learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 93-102, 2019. 1, 2, 3, 4, 5, 8, 15, 16", + "[53] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 64(3):107-115, 2021. 12", + "[54] Jiayao Zhang, Hua Wang, and Weijie Su. Imitating deep learning dynamics via locally elastic stochastic differential equations. Advances in Neural Information Processing Systems, 34, 2021. 3, 8, 12", + "[55] Mike Zhang and Barbara Plank. Cartography active learning. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 395-406, 2021. 5, 8" + ], + "bbox": [ + 501, + 90, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22344", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, 2020. 5, 15", + "[57] Tianyi Zhou, Shengjie Wang, and Jeffrey Bilmes. Curriculum learning by dynamic instance hardness. Advances in Neural Information Processing Systems, 33:8602-8613, 2020. 8", + "[58] Jia-Jie Zhu and José Bento. Generative adversarial active learning. arXiv preprint arXiv:1702.07956, 2017. 8" + ], + "bbox": [ + 78, + 90, + 470, + 247 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22345", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TiDAL_ Learning Training Dynamics for Active Learning/cbb37e36-1eb0-42b7-b0b1-3b2793b016b9_model.json b/2023/TiDAL_ Learning Training Dynamics for Active Learning/cbb37e36-1eb0-42b7-b0b1-3b2793b016b9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..56b323d8ef5aa071cbc6ae819fbd898cbd6a32ad --- /dev/null +++ b/2023/TiDAL_ Learning Training Dynamics for Active Learning/cbb37e36-1eb0-42b7-b0b1-3b2793b016b9_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.195, + 0.131, + 0.778, + 0.154 + ], + "angle": 0, + "content": "TiDAL: Learning Training Dynamics for Active Learning" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.18, + 0.786, + 0.22 + ], + "angle": 0, + "content": "Seong Min Kye\\(^{1,\\dagger}\\) Kwanghee Choi\\(^{2,\\dagger}\\) Hyeongmin Byun\\(^{1}\\) Buru Chang\\(^{3,*}\\) \n\\(^{1}\\)Hyperconnect \\(^{2}\\)Carnegie Mellon University \\(^{3}\\)Sogang University" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.22, + 0.826, + 0.234 + ], + "angle": 0, + "content": "{harris,hyeongmin.byun}@hpcnt.com kwanghec@andrew.cmu.edu buru@sogang.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.635 + ], + "angle": 0, + "content": "Active learning (AL) aims to select the most useful data samples from an unlabeled data pool and annotate them to expand the labeled dataset under a limited budget. Especially, uncertainty-based methods choose the most uncertain samples, which are known to be effective in improving model performance. However, previous methods often overlook training dynamics (TD), defined as the ever-changing model behavior during optimization via stochastic gradient descent, even though other research areas have empirically shown that TD provides important clues for measuring the data uncertainty. In this paper, we first provide theoretical and empirical evidence to argue the usefulness of utilizing the ever-changing model behavior rather than the fully trained model snapshot. We then propose a novel AL method, Training Dynamics for Active Learning (TiDAL), which efficiently predicts the training dynamics of unlabeled data to estimate their uncertainty. Experimental results show that our TiDAL achieves better or comparable performance on both balanced and imbalanced benchmark datasets compared to state-of-the-art AL methods, which estimate data uncertainty using only static information after model training." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.662, + 0.21, + 0.679 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.688, + 0.47, + 0.717 + ], + "angle": 0, + "content": "\"There is a tide in the affairs of men. Which taken at the flood, leads on to fortune.\" — William Shakespeare" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.471, + 0.854 + ], + "angle": 0, + "content": "Active learning (AL) [5, 31] aims to solve the real-world problem of selecting the most useful data samples from large-scale unlabeled data pools and annotating them to expand labeled data under a limited budget. Since the current deep neural networks are data-hungry, AL has increasingly gained attention in recent years. Existing AL methods can be divided into two mainstream categories: diversity- and uncertainty-based methods. Diversity-based methods [42, 14] focus on constructing a subset that fol" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.268, + 0.895, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.553, + 0.895, + 0.676 + ], + "angle": 0, + "content": "Figure 1: Our proposed TiDAL. TD of training samples \\( x \\) may differ even if they converge to the same final predicted probability \\( p(y^{*}|x) \\) (Upper row). Hence, we are motivated to utilize the readily available rich information generated during training, i.e., leveraging TD. We estimate TD of large-scale unlabeled data using a prediction module instead of tracking the actual TD of all the unlabeled samples to avoid the computational overhead (Lower row)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.78 + ], + "angle": 0, + "content": "ows the target data distribution. Uncertainty-based methods [13, 6, 52] choose the most uncertain samples, which are known to be effective in improving model performance. Hence, the most critical question for the latter becomes, \"How can we quantify the data uncertainty?\"" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.893, + 0.902 + ], + "angle": 0, + "content": "In this paper, we leverage training dynamics (TD) to quantify data uncertainty. TD is defined as the ever-changing model behavior on each data sample during optimization via stochastic gradient descent. Recent studies [9, 29, 48, 47] have provided empirical evidence that TD provides important clues for measuring the contribution of each data sample to model performance improvement. Inspired by these studies, we argue that the data uncertainty" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.863, + 0.207, + 0.876 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.877, + 0.223, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.079, + 0.888, + 0.462, + 0.901 + ], + "angle": 0, + "content": "This work was done while all authors were affiliated with Hyperconnect." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.863, + 0.462, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22335" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "of unlabeled data can be estimated with TD. However, most uncertainty-based methods quantify data uncertainty based on static information (e.g., loss [52] or predicted probability [45]) from a fully-trained model \"snapshot,\" neglecting the valuable information generated during training. We further argue that TD is more effective in separating uncertain and certain data than static information from a model snapshot captured after model training. In §3, we provide both theoretical and empirical evidence to support our argument that TD is a valuable tool for quantifying data uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.379 + ], + "angle": 0, + "content": "Despite its huge potential, TD is not yet actively explored in the domain of AL. This is because AL assumes a massive unlabeled data pool. Previous studies track TD only for the training data every epoch as it can be recorded easily during model optimization. On the other hand, AL targets a large number of unlabeled data, where tracking the TD for each unlabeled sample requires an impractical amount of computation (e.g., inference all the unlabeled samples every training epoch)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.473, + 0.652 + ], + "angle": 0, + "content": "Therefore, we propose TiDAL (Training Dynamics for Active Learning), a novel AL method that efficiently quantifies the uncertainty of unlabeled data by estimating their TD. We avoid tracking the TD of large-scale unlabeled data every epoch by predicting the TD of unlabeled samples with a TD prediction module. The module is trained with the TD of labeled data, which is readily available during model optimization. During the data selection phase, we predict the TD of unlabeled data with the trained module to quantify their uncertainties. We efficiently obtain TD using the module, which avoids inferring all the unlabeled samples every epoch. Experimental results demonstrate that our TiDAL achieves better or comparable performance to existing AL methods on both balanced and imbalanced datasets. Additional analyses show that our prediction module successfully predicts TD, and the predicted TD is useful in estimating uncertainties of unlabeled data. Our proposed method are illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.473, + 0.803 + ], + "angle": 0, + "content": "Contributions of our study: (1) We bridge the concept of training dynamics and active learning with the theoretical and experimental evidence that training dynamics is effective in estimating data uncertainty. (2) We propose a new method that efficiently predicts the training dynamics of unlabeled data to estimate their uncertainty. (3) Our proposed method achieves better or comparable performance on both balanced and imbalanced benchmark datasets compared to existing active learning methods. For reproducibility, we release the source code1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.819, + 0.218, + 0.835 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.845, + 0.47, + 0.876 + ], + "angle": 0, + "content": "To better understand our proposed method, we first summarize key concepts, including uncertainty-based active" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "learning, quantification of uncertainty, and training dynamics." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.143, + 0.894, + 0.204 + ], + "angle": 0, + "content": "Uncertainty-based active learning. In this work, we focus on uncertainty-based AL for multi-class classification problems. We define the predicted probabilities of the given sample \\( x \\) for \\( C \\) classes as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.215, + 0.892, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} = \\left[ p (1 | x), p (2 | x), \\dots , p (C | x) \\right] ^ {T} \\in [ 0, 1 ] ^ {C}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.246, + 0.895, + 0.368 + ], + "angle": 0, + "content": "where we denote the true label of \\( x \\) as \\( y \\) and the classifier as \\( f \\). \\( \\mathcal{D} \\) and \\( \\mathcal{D}_u \\) denote a labeled dataset and an unlabeled data pool, respectively. The general cycle of uncertainty-based AL is in two steps: (1) train the target classifier \\( f \\) on the labeled dataset \\( \\mathcal{D} \\) and (2) select top- \\( k \\) uncertain data samples from the unlabeled data pool \\( \\mathcal{D}_u \\). Selected samples are then given to the human annotators to expand the labeled dataset \\( \\mathcal{D} \\), cycling back to the first step." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.388, + 0.892, + 0.495 + ], + "angle": 0, + "content": "Quantifying uncertainty. The objective of this study is to establish a connection between the concept of TD and the field of AL. In order to clearly demonstrate the effectiveness of utilizing TD to quantify data uncertainty, we have employed two of the most prevalent and straightforward estimators, entropy [43] and margin [41], to measure data uncertainty in this paper. Entropy \\( H \\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.506, + 0.892, + 0.534 + ], + "angle": 0, + "content": "\\[\nH (\\boldsymbol {p}) = - \\sum_ {c = 1} ^ {C} p (c | x) \\log p (c | x), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.544, + 0.894, + 0.635 + ], + "angle": 0, + "content": "where the sample \\( x \\) is from the unlabeled data pool \\( \\mathcal{D}_u \\). Entropy concentrates on the level of the model's confidence on the given sample \\( x \\) and gets bigger when the prediction across the classes becomes uniform (i.e., uncertain). Margin \\( M \\) measures the difference between the probability of the true label and the maximum of the others:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.648, + 0.892, + 0.673 + ], + "angle": 0, + "content": "\\[\nM (\\boldsymbol {p}) = p (y | x) - \\max _ {c \\neq \\hat {y}} p (c | x), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.683, + 0.892, + 0.76 + ], + "angle": 0, + "content": "where \\( y \\) denotes the true label. The smaller the margin, the lower the model's confidence in the sample, so it can be considered uncertain. Both entropy and margin are computed with the predicted probabilities \\( p \\) of the fully trained classifier \\( f \\), only taking the snapshot of \\( f \\) into account." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Defining training dynamics. Our TiDAL targets to leverage TD of unlabeled data to estimate their uncertainties. TD can be defined as any model behavior during optimization, such as the area under the margin between logit values of the target class and the other largest class [39] or the variance of the predicted probabilities generated at each epoch [47]. In this work, we define the TD \\(\\bar{p}^{(t)}\\) as the area under the predicted probabilities of each data sample \\(x\\) obtained" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.396, + 0.901 + ], + "angle": 0, + "content": "1https://github.com/hyperconnect/TiDAL" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22336" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.273, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.248, + 0.251, + 0.261 + ], + "angle": 0, + "content": "(a) Entropy Distribution" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.093, + 0.468, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.248, + 0.443, + 0.261 + ], + "angle": 0, + "content": "(b) Margin Distribution" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.266, + 0.47, + 0.357 + ], + "angle": 0, + "content": "Figure 2: Score distribution after long-tailed training. We plot the marginal distributions using kernel density estimation (KDE). It is difficult to separate major (certain) and minor (uncertain) samples by the model snapshot-based scores (horizontal), unlike the TD-driven scores (vertical) that enable clearly separating the certain and uncertain samples." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.385, + 0.467, + 0.402 + ], + "angle": 0, + "content": "during the \\(t\\) time steps of optimizing the target classifier \\(f\\):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.411, + 0.469, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} ^ {(i)} = \\left[ p ^ {(i)} (1 | x), p ^ {(i)} (2 | x), \\dots , p ^ {(i)} (C | x) \\right] ^ {T}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.432, + 0.469, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\bar {\\boldsymbol {p}} ^ {(t)} = \\left[ \\bar {p} ^ {(t)} (1 | x), \\bar {p} ^ {(t)} (2 | x), \\dots , \\bar {p} ^ {(t)} (C | x) \\right] ^ {T} \\\\ = \\sum_ {\\tau} \\boldsymbol {p} ^ {(\\tau)} \\Delta \\tau \\simeq \\sum_ {i = 1} ^ {t} \\boldsymbol {p} ^ {(i)} / t, \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.645 + ], + "angle": 0, + "content": "where \\(\\pmb{p}^{(i)}\\) is the predicted probabilities of a target classifier \\(f\\) at the \\(i\\)-th time step. \\(\\Delta \\tau\\) is the unit time step to normalize the predicted probabilities. For simplicity, we record \\(\\pmb{p}^{(i)}\\) every epoch and choose \\(\\Delta \\tau = 1 / t\\), namely, averaging the predicted probabilities during \\(t\\) epochs [47, 46]. The TD \\(\\bar{\\pmb{p}}^{(t)}\\) takes all the predicted probabilities during model optimization into account. Hence, it encapsulates the overall tendency of the model during \\(t\\) epochs of optimization, avoiding being solely biased towards the snapshot of \\(\\pmb{p}^{(t)}\\) in the final epoch \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.659, + 0.457, + 0.678 + ], + "angle": 0, + "content": "3. Is TD Useful for Quantifying Uncertainty?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.47, + 0.762 + ], + "angle": 0, + "content": "In this section, we provide empirical and theoretical evidence to support our argument: \\( TD \\) is more effective in separating uncertain data from certain data than the model snapshot, where the latter is often utilized to quantify data uncertainty in previous works [52, 45]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.298, + 0.789 + ], + "angle": 0, + "content": "3.1. Motivating Observation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Settings. We aim to observe and compare the behavior of TD and the model snapshot for different sample difficulties. However, it is nontrivial to directly measure sample-wise difficulty, inhibiting the quantitative analysis of data uncertainty. To avoid this, we borrow the theoretical and empirical results of long-tailed visual recognition [33, 8, 19]: it is hard for the deep neural network-based model to train with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "fewer samples. Hence, we regard major and minor class samples to contain many certain and uncertain samples for the model, respectively. We train the target classifier \\( f \\) on the long-tailed dataset during \\( T \\) epochs to obtain the TD and the model snapshot. We apply both approaches to the common estimators, entropy and margin. We denote entropy and margin scores from the model snapshot as \\( H \\) and \\( M \\). In opposition, we denote the TD-driven scores as \\( \\bar{H} \\) and \\( \\bar{M} \\). More details and discussions are described in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Results. Figure 2 shows the distribution the scores calculated with TD \\((x\\)-axis) and model snapshot \\((y\\)-axis). We can observe that scores from TD \\((\\bar{H}, \\bar{M})\\) successfully separate the major and the minor class samples, whereas scores from the model snapshot \\((H, M)\\) fail to do so. We conclude that compared to model snapshots, TD is more helpful in separating uncertain samples from certain samples." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.358, + 0.699, + 0.373 + ], + "angle": 0, + "content": "3.2. Theoretical Evidence" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Theorem 1. (Informal) Under the LE-SDE framework [54], with the assumption of local elasticity [17], certain samples and uncertain samples reveal different \\(TD\\); especially, certain samples converge quickly than uncertain samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.644 + ], + "angle": 0, + "content": "The above theorem discusses different model behaviors depending on the difficulty of the sample. Compared to the uncertain sample, the certain sample has the same class samples nearby, which is the fundamental idea of level set estimation [22] and nearest neighbor [36] literature. We suspect that, due to the local elasticity of deep nets, samples close by have a bigger impact on the certain sample, hence changing its predicted probability more rapidly. As the certain sample is quicker to converge, its TD is larger than that of the uncertain sample. Intuitively, slower to train, struggling the classifier is to learn, hence TD capturing the uncertainty in the classifier's perspective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.892, + 0.724 + ], + "angle": 0, + "content": "Theorem 2. (Informal) Estimators such as Entropy (Equation 10) and Margin (Equation 11) successfully capture the difference of TD between easy and hard samples even for the case where it cannot be distinguished via the predicted probabilities of the model snapshot." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.892, + 0.835 + ], + "angle": 0, + "content": "The above theorem discusses the validity of entropy and margin on whether they can successfully differentiate between two samples of different TD but with the same final prediction. With Theorem 1, one can conclude that the common estimators' scores calculated with TD are effective in capturing the data uncertainty. Due to the space constraints, we provide the details of the above results in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.799, + 0.864 + ], + "angle": 0, + "content": "4. Utilizing TD for Active Learning" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As tracking the TD of all the unlabeled data is computationally infeasible, we devise an efficient method to es" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22337" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "estimate the TD of unlabeled samples. We train the module that directly predicts the TD of each sample by feeding the training samples, where its TD are freely available during training. Then, based on the predicted TD of each unlabeled sample, we use the common estimators, entropy or margin, to determine which sample is the most uncertain so that human annotators can label it. Hence, in this section, we describe the details of the module that estimates TD (§4.1) and how to train the module (§4.2). Finally, calculating the uncertainties using the module predictions for active learning is illustrated (§4.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.265, + 0.411, + 0.281 + ], + "angle": 0, + "content": "4.1. Training Dynamics Prediction Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.47, + 0.439 + ], + "angle": 0, + "content": "As mentioned, it is not computationally feasible to track TD for the large-scale unlabeled data as it requires model inference on all the unlabeled data every training epoch. Thus, we propose the TD prediction module \\( m \\) to efficiently predict the TD of unlabeled data at the \\( t \\)-th epoch. Being influenced by the previous studies [11, 52, 45, 25] that use additional modules to predict useful values such as loss or confidence by the target model outputs, multi-scale feature maps are aggregated and passed into our TD prediction module. The module produces the \\( C \\)-dimensional predictions:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.443, + 0.469, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {p}} _ {m} ^ {(t)} = \\left[ \\tilde {p} _ {m} ^ {(t)} (1 | x), \\dots , \\tilde {p} _ {m} ^ {(t)} (C | x) \\right] ^ {T} \\in [ 0, 1 ] ^ {C} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.469, + 0.545 + ], + "angle": 0, + "content": "estimating the actual TD \\(\\bar{\\pmb{p}}^{(t)}\\) of the given sample \\(x\\) in Equation 5. TD prediction module is jointly trained with the target classifier using a handful of parameters, having a negligible computational cost during training. The detailed architecture of the module is described in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.666 + ], + "angle": 0, + "content": "Even though the architecture is similar to previous works [52, 45, 25], we observed that ours were much more stable during optimization and easier to train. We suspect that it is due to the target task difference; previous works trained the module that outputs only a single value via regression, whereas our module outputs \\( C \\)-dimensional probability distribution, which is similar to the main task of classifying images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.672, + 0.267, + 0.689 + ], + "angle": 0, + "content": "4.2. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.47, + 0.756 + ], + "angle": 0, + "content": "To train the target classifier \\( f \\) at the \\( t \\)-th epoch, we use the cross-entropy loss function \\( \\mathcal{L}_{\\mathrm{target}} \\) on the predicted probability \\( \\pmb{p}^{(t)} \\) and a one-hot encoded vector \\( \\pmb{y} \\in \\{0,1\\}^C \\) of the true label \\( y \\):" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.761, + 0.469, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t a r g e t}} = \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {p} ^ {(t)}, \\boldsymbol {y}\\right) = - \\log p ^ {(t)} (y | x). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.785, + 0.47, + 0.833 + ], + "angle": 0, + "content": "Meanwhile, the prediction module \\(m\\) learns the TD of a sample \\(x\\) by minimizing the Kullback-Leibler (KL) divergence between the predicted TD \\(\\tilde{p}_m^{(t)}\\) and the actual TD \\(\\bar{p}^{(t)}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.836, + 0.47, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {m o d u l e}} = \\mathcal {L} _ {\\mathrm {K L}} (\\bar {\\boldsymbol {p}} ^ {(t)} | | \\tilde {\\boldsymbol {p}} _ {m} ^ {(t)}) \\\\ = \\sum_ {c = 1} ^ {C} \\bar {p} ^ {(t)} (c | x) \\log \\left(\\frac {\\bar {p} ^ {(t)} (c | x)}{\\tilde {p} _ {m} ^ {(t)} (c | x)}\\right). \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "The final objective function of our proposed method is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.134, + 0.892, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {t a r g e t}} + \\lambda \\mathcal {L} _ {\\text {m o d u l e}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.16, + 0.892, + 0.191 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a balancing factor to control the effect of \\(\\mathcal{L}_{\\mathrm{module}}\\) during model training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.199, + 0.795, + 0.216 + ], + "angle": 0, + "content": "4.3. Quantifying Uncertainty with TD" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.223, + 0.892, + 0.329 + ], + "angle": 0, + "content": "We argue that uncertain samples can be effectively distinguished from unlabeled data using the predicted TD. To verify the effectiveness of leveraging TD, we feed the predicted TD to entropy and margin (§2) by replacing snapshot probability \\( \\pmb{p} \\) with the predicted TD \\( \\bar{\\pmb{p}} \\). We choose these estimators as they are widely used for quantifying uncertainty. We feed \\( \\bar{\\pmb{p}} \\), replacing \\( \\pmb{p} \\), to the entropy \\( \\bar{H} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.338, + 0.892, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\bar {H} (\\bar {\\boldsymbol {p}}) = - \\sum_ {c = 1} ^ {C} \\bar {p} (c | x) \\log \\bar {p} (c | x). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.375, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Entropy \\(\\bar{H}\\) is maximized when \\(\\bar{p}\\) is uniform, i.e., the sample is uncertain for the target classifier. Margin \\(\\bar{M}\\) is also similarly employed:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.433, + 0.892, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\bar {M} (\\bar {\\boldsymbol {p}}) = \\bar {p} (\\hat {y} | x) - \\max _ {c \\neq \\hat {y}} \\bar {p} (c | x). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.892, + 0.557 + ], + "angle": 0, + "content": "Since we do not have true labels of unlabeled samples, we use the predicted labels \\(\\hat{y}\\) of the target classifier instead of the true labels. There are several possible variants of \\(\\bar{M}\\) depending on the definition of \\(\\hat{y}\\). We conduct experiments to compare \\(\\bar{M}\\) with its variants. The experimental details and results are in Appendix D.4." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.557, + 0.892, + 0.65 + ], + "angle": 0, + "content": "At the data selection phase, we use the predicted TD \\(\\tilde{\\pmb{p}}_m^{(T)}\\) instead of the actual TD \\(\\bar{\\pmb{p}}^{(T)}\\) as in Equation 10 & 11 to estimate the TD-driven uncertainties of the unlabeled sample \\(x\\) at the final epoch \\(T\\). By using the estimated uncertainty with the predicted TD, we select the most informative samples for model training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.662, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.839 + ], + "angle": 0, + "content": "In this section, we experimentally verify the effectiveness of our method, TiDAL, which utilizes the estimated training dynamics from the prediction module to discern uncertain samples from unlabeled data. We describe the detailed settings and the baseline methods for our experiments (§5.1) and show the results on both balanced (§5.2) and imbalanced datasets (§5.3). We further analyze whether the TD prediction module is effective for AL performance and can successfully estimate the TD (§5.4). We end the section by discussing the potential limitations of our method (§5.5)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.691, + 0.864 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets. To assess the performance of our proposed method and baseline methods, we conduct experiments on" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22338" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.09, + 0.36, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.09, + 0.611, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.091, + 0.861, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.895, + 0.337 + ], + "angle": 0, + "content": "Figure 3: Averaged relative accuracy improvement curves and their \\(95\\%\\) confidence interval (shaded) of AL methods over the number of labeled samples on balanced datasets. TiDAL \\((\\bar{H})\\) and TiDAL \\((\\bar{M})\\) denote the performance of TiDAL when with entropy \\(\\bar{H}\\) and margin \\(\\bar{M}\\) as the data uncertainty estimation strategy, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.473, + 0.552 + ], + "angle": 0, + "content": "the following five datasets: CIFAR10/100 [27], FashionMNIST [51], SVHN [34], and iNaturalist2018 [50]. Since CIFAR and FashionMNIST are both balanced, we further modify them to simulate the data imbalance in the real world, following the previous long-tail visual recognition studies [8, 33, 56, 19]. The imbalance ratio is defined as \\( N_{\\mathrm{max}} / N_{\\mathrm{min}} \\) where \\( N \\) is the number of samples in each class. We make two variants with data imbalance ratios 10 and 100 for each dataset. Unlike the above, SVHN and iNaturalist18 are already imbalanced. Especially, iNaturalist2018 is commonly chosen to demonstrate how methods work in imbalanced real-world settings. The dataset statistics are summarized in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Baselines. For a fair comparison, we compare our TiDAL with the following baselines which train a target classifier with only labeled data. Random sampling: a simple baseline that randomly selects data samples from the unlabeled dataset. Entropy sampling [43]: an uncertainty-based method that selects data samples based on the maximum entropy. BALD [13]: an uncertainty-based method that selects data samples based on the mutual information between the model prediction and the posterior. CoreSet [42]: a diversity-based method that selects representative data samples covering all data through a minimum radius. LLoss [52]: an uncertainty-based method that learns to estimate the errors of the predictions (loss) made by the learner and select data samples based on the predicted loss. CAL [55]: recent work on using TD, gathering sample-wise TD information on whether the classifier was consistently correct or not during training. CAL splits the samples into two classes by applying a heuristic threshold to the TD information to train a binary classifier that outputs uncertainty score. To verify the effectiveness of TiDAL, we further compare it with the two semi-supervised AL methods, VAAL [45] and TA-VAAL [25] in Appendix D.3. Note that" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.354, + 0.892, + 0.385 + ], + "angle": 0, + "content": "these methods further utilize unlabeled data for training the selection module, thus it is unfair for our TiDAL." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.405, + 0.893, + 0.542 + ], + "angle": 0, + "content": "Active learning setting. We follow the same setting from [6, 52] for the detailed AL settings. For the initial step, we randomly select initial samples to be annotated from the unlabeled dataset, where we use them to train the initial target classifier. Then, we obtain a random subset from the unlabeled data pool \\( \\mathcal{D}_u \\) to choose the top- \\( k \\) samples based on the criterion of each method, where those samples will be annotated. We repeat the above cycle, training a classifier from scratch from the continuously expanding labeled set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.895, + 0.804 + ], + "angle": 0, + "content": "Implementation details. For a fair comparison, we use the same backbone network ResNet-18 [18] except for iNaturalist2018, where we use ResNet-50 [18] pretrained on ImageNet [12]. All models are trained with SGD optimizer with momentum 0.9, weight decay \\(5 \\cdot 10^{-4}\\), and learning rate (LR) decay of 0.1. For CIFAR10/100 and SVHN, we train the model for 200 epochs with an initial LR of 0.1 and decay at epoch 160. For FashionMNIST, 100 epochs with an initial LR of 0.1 and decay at epoch 80. For iNaturalist2018, 50 epochs with an initial LR of 0.01 and decay at epoch 40. For CIFAR10/100, SVHN and FashionMNIST, we set the batch size and the unlabeled subset size to be 128 and \\(10^{4}\\), respectively. For iNaturalist2018, which is much larger than other datasets, we set the batch size and the unlabeled subset size to 256 and \\(10^{6}\\), respectively. We set the balancing factor to 1.0." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Evaluation details. To compare with other state-of-the-art baselines, we show the average accuracy and \\(95\\%\\) confidence interval with three trials. We mainly compare the model performances with relative accuracy improvement to random sampling, demonstrating how much it improves" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22339" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.362, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.089, + 0.611, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.09, + 0.862, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.286, + 0.359, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.286, + 0.61, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.286, + 0.861, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.483, + 0.894, + 0.529 + ], + "angle": 0, + "content": "Figure 4: Averaged relative accuracy improvement curves and their \\(95\\%\\) confidence interval (shaded) of AL methods over the number of labeled samples on synthetically imbalanced datasets. We use the imbalance ratio (IR) of 10 and 100 on CIFAR10, CIFAR100, and FashionMNIST." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.47, + 0.581 + ], + "angle": 0, + "content": "upon the naive approach on each cycle. Additionally, absolute accuracy is also plotted in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.34, + 0.604 + ], + "angle": 0, + "content": "5.2. Results on Balanced Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.471, + 0.779 + ], + "angle": 0, + "content": "Figure 3 and 10 compare our TiDAL against the state-of-the-art methods on various balanced datasets: CIFAR10, CIFAR100, and FashionMNIST. For all the datasets, the two variants of TiDAL outperform all the baselines at all AL cycles except for LLoss, which shows better improvement than TiDAL \\((\\bar{M})\\) on CIFAR10 with an imbalance ratio of 100. Nonetheless, our TiDAL achieves the best final performance compared to all the baselines. CAL, which uses training dynamics, generally underperforms compared to others. We suspect that CAL is sensitive to its threshold hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.359, + 0.803 + ], + "angle": 0, + "content": "5.3. Results on Imbalanced Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Synthetically imbalanced datasets. Similar to the above, Figure 4, 9, and 11 shows the performance improvements on the synthetically imbalanced datasets with the two imbalance ratios, 10 and 100. Except for the CIFAR10 with an imbalance ratio of 100, our methods show superb performance across all the imbalanced settings. TiDAL per" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.895, + 0.673 + ], + "angle": 0, + "content": "forms especially well with a small variance in imbalanced CIFAR100, where the number of classes is the largest. In imbalanced FashionMNIST, the performance quickly rises to 2.5k labeled images and then saturates. This implies that FashionMNIST is easier than other datasets, and needs to focus more on the early training steps to compare with other models. TiDAL also shows overall better performance on FashionMNIST, especially in the early steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.895, + 0.84 + ], + "angle": 0, + "content": "Real-world imbalanced datasets. Figure 5 and 10 shows evaluation results on real-world imbalanced datasets. For iNaturalist2018, which is the large-scale long-tailed classification dataset, TiDAL shows outstanding performance compared to other methods. For SVHN, TiDAL shows the best improvements with low variance as the number of labeled images increases except for the initial stage. LLoss shows outstanding performance only in the initial stage, where we presume that the loss prediction module of LLoss acts as a regularizer during model optimization." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.848, + 0.833, + 0.864 + ], + "angle": 0, + "content": "5.4. Analysis on the TD Prediction Module" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of the TD prediction module. In order to verify the efficacy of using the predicted TD \\(\\tilde{p}_m\\), we con" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22340" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.218, + 0.09, + 0.484, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.09, + 0.752, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.308, + 0.895, + 0.37 + ], + "angle": 0, + "content": "Figure 5: Averaged relative accuracy improvement curves and its \\(95\\%\\) confidence interval (shaded) of AL methods over the number of labeled samples on real-world imbalanced datasets: iNaturalist2018 and SVHN. For SVHN, LLoss shows a substantial improvement of \\(20.02\\% \\pm 6.77\\%\\) at the initial phase (1k), but we clip the plot to show the performance afterward more clearly." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.376, + 0.319, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.376, + 0.553, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.08, + 0.59, + 0.558, + 0.666 + ], + "angle": 0, + "content": "Figure 6: Ablation test results. \\(\\bar{H} (\\tilde{\\pmb{p}}_m)\\) and \\(\\bar{M} (\\tilde{\\pmb{p}}_m)\\) use the predicted TD \\(\\tilde{\\pmb{p}}_m\\) of the prediction module \\(m\\). In contrast, \\(H(\\pmb {p})\\) and \\(M(\\pmb {p})\\) use the predicted probability of the model snapshot \\(\\pmb{p}\\). TD shows better performance than the model snapshot, implying that TD is better at quantifying data uncertainty." + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.387, + 0.876, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.57, + 0.889, + 0.661 + ], + "angle": 0, + "content": "Figure 7: KL divergence scores of the actual TD \\(\\bar{\\pmb{p}}^{(T)}\\) with the predicted TD \\(\\tilde{\\pmb{p}}_m^{(t)}\\) and the predicted probability of the model snapshot \\(\\pmb{p}^{(t)}\\), respectively, during model optimization. Our predicted TD can accurately approximate the actual TD." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.473, + 0.89 + ], + "angle": 0, + "content": "duct an ablation test that compares the performance between when using and not using the TD prediction module \\( m \\). Figure 6 shows the results on balanced CIFAR10/100. We observe that \\( \\bar{H}(\\tilde{\\pmb{p}}_m) \\) and \\( \\bar{M}(\\tilde{\\pmb{p}}_m) \\) using the predicted TD \\( \\tilde{\\pmb{p}}_m \\) to estimate the data uncertainty significantly outperform the methods \\( H(\\pmb{p}) \\) and \\( M(\\pmb{p}) \\) that use only the final predicted probabilities \\( \\pmb{p} \\) of the target classifier \\( f \\), showing better performance in the whole training cycle. Even \\( M(\\pmb{p}) \\) shows temporary improvement in earlier steps on CIFAR100, \\( \\bar{H}(\\tilde{\\pmb{p}}_m) \\) and \\( \\bar{M}(\\tilde{\\pmb{p}}_m) \\) maintain stable improvement, eventually winning over \\( M(\\pmb{p}) \\). This indicates that the predicted TD \\( \\tilde{\\pmb{p}}_m \\) of the TD prediction module \\( m \\) produces better data uncertainty estimation than the predicted" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.693, + 0.752, + 0.709 + ], + "angle": 0, + "content": "probability \\( \\pmb{p} \\) of the target classifier \\( f \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Predictive performance of the TD prediction module. We verify whether the TD prediction module \\(m\\) accurately predicts the actual TD \\(\\bar{p}\\). Its prediction performance is crucial as we use the predicted TD \\(\\tilde{\\boldsymbol{p}}_m\\) of the module \\(m\\) to quantify uncertainties of unlabeled data. Using the KL divergence \\(\\mathcal{L}_{\\mathrm{KL}}\\), we analyze that the predicted TD \\(\\tilde{\\boldsymbol{p}}_m\\) converges to the actual TD \\(\\bar{p}\\) at the data selection phase. We calculate \\(\\mathcal{L}_{\\mathrm{KL}}(\\bar{\\boldsymbol{p}}^{(T)}||\\tilde{\\boldsymbol{p}}_m^{(t)})\\) and compare it with \\(\\mathcal{L}_{\\mathrm{KL}}(\\bar{\\boldsymbol{p}}^{(T)}||\\boldsymbol{p}^{(t)})\\) which is set as a baseline computed with the actual TD \\(\\bar{p}\\) and the predicted probabilities \\(\\boldsymbol{p}\\) (snapshot) of the target classifier \\(f\\). In this analysis, we use the bal" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22341" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.228 + ], + "angle": 0, + "content": "anced CIFAR10 where the sample-wise averaged KL divergence scores are computed on the test set. Figure 7 shows that the final predicted TD successfully approximates the actual TD, while the predicted probability is highly different from the actual TD. We conclude that the TD prediction module \\( m \\) can produce the TD efficiently, leading to performance improvement, and the predicted TD acts as a better approximation of the actual TD than the predicted probability of a model snapshot captured at each epoch." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.239, + 0.204, + 0.254 + ], + "angle": 0, + "content": "5.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.471, + 0.491 + ], + "angle": 0, + "content": "We found two potential limitations of our TiDAL derived from the fact that it relies on the outputs of the target classifier to compute the TD. First, TiDAL is designed only for classification tasks, and thus it cannot be applied to AL targeting other tasks, such as regression [10, 15]. Second, TiDAL is highly influenced by the performance of the target classifier, especially when the target classifier wrongly classifies the hard negative samples with a high confidence during model optimization. These samples can be treated as certain samples (i.e. will not be selected for annotation) because they have low estimated uncertainties from the predicted TD, even though the target classifier fails to predict the true label of the samples correctly. As a future work, we will study extending our TiDAL in the task-agnostic ways with a safeguard combating the wrongly classified samples." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.505, + 0.22, + 0.52 + ], + "angle": 0, + "content": "6. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.53, + 0.239, + 0.546 + ], + "angle": 0, + "content": "6.1. Active Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.903 + ], + "angle": 0, + "content": "AL methods target to construct a dataset with the most useful samples based on the assumption that each sample has different importance in model training [40]. Two mainstream AL approaches exist for efficiently querying the unlabeled data: pool-based methods [31, 52, 45] use various ways to extract samples from an unlabeled data pool effectively, and synthesis-based methods [1, 58, 49] generate informative samples for the model. Pool-based methods can be roughly divided based on query strategies: uncertainty-based [13, 52, 45, 20] and diversity-based [42, 14, 38] methods, where some methods use the hybrid of both [4, 44, 25]. Uncertainty-based methods focus on finding which samples would be the most uncertain for the model, whereas diversity-based methods aim to construct a subset of representative samples of the input distribution. Our proposed method, TiDAL, lies in uncertainty-based methods. The significant difference between TiDAL and previous uncertainty-based methods is that TiDAL estimates data uncertainty using TD that contains additional hints generated during model training. In contrast, the previous methods leverage only static information (e.g., loss [52, 20] and predicted probabilities [13, 45, 25]) obtained by a model snapshot at the data selection phase." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.685, + 0.108 + ], + "angle": 0, + "content": "6.2. Training Dynamics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.895, + 0.691 + ], + "angle": 0, + "content": "TD focuses on how deep neural networks are optimized under back-propagation-based stepwise weight updates. Many studies try to understand how gradient descent can effectively obtain the global minimum by analyzing the loss landscape of neural networks [24, 32] or its loss trajectory [3]. Some also import alternative models that are more mathematically approachable to analyze, such as neural tangent kernels [21], deep Gaussian processes [30], or stochastic differential equations [54]. On the other hand, the phenomenological and practical viewpoint of TD also exists. [48] coin the term Forgetting Dynamics to assert that unforgettable samples are often less helpful, and [9] show that the model could prefer samples that are often wrongly predicted throughout model training. TD is also commonly used in noisy label literature to find potential noisy labels as they tend to fit later on model training [2, 39] or locate samples that can be relabeled correctly [46]. Furthermore, [57] calculate the Dynamic Instance Hardness score by monitoring losses of each sample or whether the prediction gets flipped so that higher scored samples can be prioritized for curriculum learning, and [23] feed the loss history to the auxiliary neural network to mediate the curriculum for training. [29] also introduce temporal ensembling for semi-supervised learning, where the model fits towards averaged probability outputs. [47, 37] devise Data Maps to inspect datasets with two TD measures; confidence and its variability across epochs on the true class prediction. [55] further extend the Data Maps for AL, whether the target classifier was consistently correct or not during training. The proposed method splits the labeled samples by applying a heuristic threshold on the level of consistency to train a binary classifier that is trained to discern uncertain samples. Even though the work, similar to ours, also utilizes TD, it relies on empirical observations and heuristic choices to separate the certain and uncertain samples. In this study, we link the concept of TD to AL with both empirical and theoretical results to estimate the uncertainty of unlabeled samples, which is often neglected in previous TD studies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.708, + 0.619, + 0.724 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We propose a novel active learning method, Training Dynamics for Active Learning (TiDAL), by linking the concept of training dynamics to active learning. We provide motivating observations and theoretical evidence for using training dynamics to estimate the uncertainty of unlabeled data. Since tracking the training dynamics of large-scale unlabeled data is infeasible, TiDAL utilizes a training dynamics prediction module to efficiently predict the training dynamics of the unlabeled data. Based on the predicted training dynamics, TiDAL quantifies data uncertainty using the common uncertainty estimators: entropy and mar" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22342" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.182 + ], + "angle": 0, + "content": "gin. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness of our method, surpassing the existing state-of-the-art active learning methods. We further analyze that using our training dynamics prediction module is effective and the module successfully predicts the TD of unlabeled data." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.199, + 0.174, + 0.215 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.224, + 0.47, + 0.252 + ], + "angle": 0, + "content": "[1] Dana Angluin. Queries and concept learning. Machine learning, 2(4):319-342, 1988. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.255, + 0.472, + 0.31 + ], + "angle": 0, + "content": "[2] Eric Arazo, Diego Ortego, Paul Albert, Noel O'Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In International conference on machine learning, pages 312-321. PMLR, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.47, + 0.368 + ], + "angle": 0, + "content": "[3] Sanjeev Arora, Nadav Cohen, Noah Golowich, and Wei Hu. A convergence analysis of gradient descent for deep linear neural networks. In International Conference on Learning Representations, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.371, + 0.469, + 0.425 + ], + "angle": 0, + "content": "[4] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671, 2019.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.428, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[5] Les E Atlas, David A Cohn, and Richard E Ladner. Training connectionist networks with queries and selective sampling. In Advances in neural information processing systems, pages 566-573. CiteSeer, 1990. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.486, + 0.469, + 0.554 + ], + "angle": 0, + "content": "[6] William H Beluch, Tim Genewein, Andreas Nurnberger, and Jan M Kohler. The power of ensembles for active learning in image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9368-9377, 2018. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.557, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[7] Javad Zolfaghari Bengar, Joost van de Weijer, Laura Lopez Fuentes, and Bogdan Raducanu. Class-balanced active learning for image classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1536-1545, 2022. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.629, + 0.469, + 0.683 + ], + "angle": 0, + "content": "[8] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in Neural Information Processing Systems, 32:1567-1578, 2019. 3, 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.687, + 0.469, + 0.741 + ], + "angle": 0, + "content": "[9] Haw-Shiuan Chang, Erik Learned-Miller, and Andrew McCallum. Active bias: Training more accurate neural networks by emphasizing high variance samples. Advances in Neural Information Processing Systems, 30, 2017. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[10] David Cohn, Les Atlas, and Richard Ladner. Improving generalization with active learning. Machine learning, 15(2):201-221, 1994. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[11] Charles Corbière, Nicolas Thome, Avner Bar-Hen, Matthieu Cord, and Patrick Pérez. Addressing failure prediction by learning model confidence. Advances in Neural Information Processing Systems, 32, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.224, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[13] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International Conference on Machine Learning, pages 1183-1192. PMLR, 2017. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[14] Daniel Gissin and Shai Shalev-Shwartz. Discriminative active learning. arXiv preprint arXiv:1907.06347, 2019. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[15] Jia Gong, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Meta agent teaming active learning for pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079-11089, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.299 + ], + "angle": 0, + "content": "[16] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[17] Hangfeng He and Weijie Su. The local elasticity of neural networks. In International Conference on Learning Representations, 2019. 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[19] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6626-6636, 2021. 3, 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.536 + ], + "angle": 0, + "content": "[20] Siyu Huang, Tianyang Wang, Haoyi Xiong, Jun Huan, and Dejing Dou. Semi-supervised active learning with temporal output discrepancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3447-3456, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[21] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[22] Heinrich Jiang, Been Kim, Melody Guan, and Maya Gupta. To trust or not to trust a classifier. Advances in neural information processing systems, 31, 2018. 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[23] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International Conference on Machine Learning, pages 2304–2313. PMLR, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[24] Kenji Kawaguchi. Deep learning without poor local minima. Advances in neural information processing systems, 29, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[25] Kwanyoung Kim, Dongwon Park, Kwang In Kim, and Se Young Chun. Task-aware variational adversarial active learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8166-8175, 2021. 4, 5, 8, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22343" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[28] Alex Krizhevsky, Vinod Nair, and Geofrey Hinton. The CIFar-10 dataset. http://www.cs.toronto.edu/kriz/cifar.html, 55(5), 2014. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.148, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[29] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In International Conference on Learning Representations (ICLR). OpenReview.net, 2017. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.471, + 0.246 + ], + "angle": 0, + "content": "[30] Jaehoon Lee, Yasaman Bahri, Roman Novak, Samuel S Schoenholz, Jeffrey Pennington, and Jascha Sohl-Dickstein. Deep neural networks as gaussian processes. In International Conference on Learning Representations, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[31] David D Lewis and William A Gale. A sequential algorithm for training text classifiers. In Proceedings of the 17th annual international ACM SIGIR conference on Research and development in information retrieval, pages 3-12, 1994. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[32] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.357, + 0.471, + 0.427 + ], + "angle": 0, + "content": "[33] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 3, 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.471, + 0.496 + ], + "angle": 0, + "content": "[34] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.496, + 0.471, + 0.539 + ], + "angle": 0, + "content": "[35] Curtis Northcutt, Lu Jiang, and Isaac Chuang. Confident learning: Estimating uncertainty in dataset labels. Journal of Artificial Intelligence Research, 70:1373-1411, 2021. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.581 + ], + "angle": 0, + "content": "[36] Nicolas Papernot and Patrick McDaniel. Deep k-nearest neighbors: Towards confident, interpretable and robust deep learning. arXiv preprint arXiv:1803.04765, 2018. 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.471, + 0.622 + ], + "angle": 0, + "content": "[37] Seo Yeon Park and Cornelia Caragea. A data cartography based mixup for pre-trained language models. arXiv preprint arXiv:2205.03403, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.471, + 0.692 + ], + "angle": 0, + "content": "[38] Amin Parvaneh, Ehsan Abbasnejad, Damien Teney, Gholam-reza Reza Haffari, Anton van den Hengel, and Javen Qinfeng Shi. Active learning by feature mixing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12237-12246, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.748 + ], + "angle": 0, + "content": "[39] Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. Identifying unlabeled data using the area under the margin ranking. Advances in Neural Information Processing Systems, 33:17044-17056, 2020. 2, 8, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[40] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning. ACM Computing Surveys (CSUR), 54(9):1-40, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[41] Dan Roth and Kevin Small. Margin-based active learning for structured output spaces. In European Conference on Machine Learning, pages 413-424. Springer, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[42] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 1, 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[43] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[44] Changjian Shui, Fan Zhou, Christian Gagné, and Boyu Wang. Deep active learning: Unified and principled method for query and training. In International Conference on Artificial Intelligence and Statistics, pages 1308-1318. PMLR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.265 + ], + "angle": 0, + "content": "[45] Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell. Variational adversarial active learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5972-5981, 2019. 2, 3, 4, 5, 8, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[46] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In International Conference on Machine Learning, pages 5907-5915. PMLR, 2019. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[47] Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A Smith, and Yejin Choi. Dataset cartography: Mapping and diagnosing datasets with training dynamics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9275-9293, 2020. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.41, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[48] Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J Gordon. An empirical study of example forgetting during deep neural network learning. In International Conference on Learning Representations, 2018. 1, 8, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.482, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[49] Toan Tran, Thanh-Toan Do, Ian Reid, and Gustavo Carneiro. Bayesian generative active deep learning. In International Conference on Machine Learning, pages 6295-6304. PMLR, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.624 + ], + "angle": 0, + "content": "[50] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.625, + 0.892, + 0.681 + ], + "angle": 0, + "content": "[51] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.684, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[52] Donggeun Yoo and In So Kweon. Learning loss for active learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 93-102, 2019. 1, 2, 3, 4, 5, 8, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.741, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[53] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 64(3):107-115, 2021. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.8, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[54] Jiayao Zhang, Hua Wang, and Weijie Su. Imitating deep learning dynamics via locally elastic stochastic differential equations. Advances in Neural Information Processing Systems, 34, 2021. 3, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.857, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[55] Mike Zhang and Barbara Plank. Cartography active learning. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 395-406, 2021. 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22344" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.161 + ], + "angle": 0, + "content": "[56] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, 2020. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[57] Tianyi Zhou, Shengjie Wang, and Jeffrey Bilmes. Curriculum learning by dynamic instance hardness. Advances in Neural Information Processing Systems, 33:8602-8613, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[58] Jia-Jie Zhu and José Bento. Generative adversarial active learning. arXiv preprint arXiv:1702.07956, 2017. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "22345" + } + ] +] \ No newline at end of file diff --git a/2023/TiDAL_ Learning Training Dynamics for Active Learning/cbb37e36-1eb0-42b7-b0b1-3b2793b016b9_origin.pdf b/2023/TiDAL_ Learning Training Dynamics for Active Learning/cbb37e36-1eb0-42b7-b0b1-3b2793b016b9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ad8289956438781e34690f3b4c71293dd127a6ea --- /dev/null +++ b/2023/TiDAL_ Learning Training Dynamics for Active Learning/cbb37e36-1eb0-42b7-b0b1-3b2793b016b9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa94e500d3fa8ab2b0583ecb31ba9d793ffca7d69b2d74f90b3b977da9f78be8 +size 1447093 diff --git a/2023/TiDAL_ Learning Training Dynamics for Active Learning/full.md b/2023/TiDAL_ Learning Training Dynamics for Active Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..efc6c43f0dd7aaec99cfeb66375c9002ac985544 --- /dev/null +++ b/2023/TiDAL_ Learning Training Dynamics for Active Learning/full.md @@ -0,0 +1,323 @@ +# TiDAL: Learning Training Dynamics for Active Learning + +Seong Min Kye $^{1,\dagger}$ Kwanghee Choi $^{2,\dagger}$ Hyeongmin Byun $^{1}$ Buru Chang $^{3,*}$ $^{1}$ Hyperconnect $^{2}$ Carnegie Mellon University $^{3}$ Sogang University + +{harris,hyeongmin.byun}@hpcnt.com kwanghec@andrew.cmu.edu buru@sogang.ac.kr + +# Abstract + +Active learning (AL) aims to select the most useful data samples from an unlabeled data pool and annotate them to expand the labeled dataset under a limited budget. Especially, uncertainty-based methods choose the most uncertain samples, which are known to be effective in improving model performance. However, previous methods often overlook training dynamics (TD), defined as the ever-changing model behavior during optimization via stochastic gradient descent, even though other research areas have empirically shown that TD provides important clues for measuring the data uncertainty. In this paper, we first provide theoretical and empirical evidence to argue the usefulness of utilizing the ever-changing model behavior rather than the fully trained model snapshot. We then propose a novel AL method, Training Dynamics for Active Learning (TiDAL), which efficiently predicts the training dynamics of unlabeled data to estimate their uncertainty. Experimental results show that our TiDAL achieves better or comparable performance on both balanced and imbalanced benchmark datasets compared to state-of-the-art AL methods, which estimate data uncertainty using only static information after model training. + +# 1. Introduction + +"There is a tide in the affairs of men. Which taken at the flood, leads on to fortune." — William Shakespeare + +Active learning (AL) [5, 31] aims to solve the real-world problem of selecting the most useful data samples from large-scale unlabeled data pools and annotating them to expand labeled data under a limited budget. Since the current deep neural networks are data-hungry, AL has increasingly gained attention in recent years. Existing AL methods can be divided into two mainstream categories: diversity- and uncertainty-based methods. Diversity-based methods [42, 14] focus on constructing a subset that fol + +![](images/604e19a021cbe9bfdf754ce80ffa26e7b8dc303a11ccf08b5fcdb6a885d1b41e.jpg) +Figure 1: Our proposed TiDAL. TD of training samples $x$ may differ even if they converge to the same final predicted probability $p(y^{*}|x)$ (Upper row). Hence, we are motivated to utilize the readily available rich information generated during training, i.e., leveraging TD. We estimate TD of large-scale unlabeled data using a prediction module instead of tracking the actual TD of all the unlabeled samples to avoid the computational overhead (Lower row). + +ows the target data distribution. Uncertainty-based methods [13, 6, 52] choose the most uncertain samples, which are known to be effective in improving model performance. Hence, the most critical question for the latter becomes, "How can we quantify the data uncertainty?" + +In this paper, we leverage training dynamics (TD) to quantify data uncertainty. TD is defined as the ever-changing model behavior on each data sample during optimization via stochastic gradient descent. Recent studies [9, 29, 48, 47] have provided empirical evidence that TD provides important clues for measuring the contribution of each data sample to model performance improvement. Inspired by these studies, we argue that the data uncertainty + +of unlabeled data can be estimated with TD. However, most uncertainty-based methods quantify data uncertainty based on static information (e.g., loss [52] or predicted probability [45]) from a fully-trained model "snapshot," neglecting the valuable information generated during training. We further argue that TD is more effective in separating uncertain and certain data than static information from a model snapshot captured after model training. In §3, we provide both theoretical and empirical evidence to support our argument that TD is a valuable tool for quantifying data uncertainty. + +Despite its huge potential, TD is not yet actively explored in the domain of AL. This is because AL assumes a massive unlabeled data pool. Previous studies track TD only for the training data every epoch as it can be recorded easily during model optimization. On the other hand, AL targets a large number of unlabeled data, where tracking the TD for each unlabeled sample requires an impractical amount of computation (e.g., inference all the unlabeled samples every training epoch). + +Therefore, we propose TiDAL (Training Dynamics for Active Learning), a novel AL method that efficiently quantifies the uncertainty of unlabeled data by estimating their TD. We avoid tracking the TD of large-scale unlabeled data every epoch by predicting the TD of unlabeled samples with a TD prediction module. The module is trained with the TD of labeled data, which is readily available during model optimization. During the data selection phase, we predict the TD of unlabeled data with the trained module to quantify their uncertainties. We efficiently obtain TD using the module, which avoids inferring all the unlabeled samples every epoch. Experimental results demonstrate that our TiDAL achieves better or comparable performance to existing AL methods on both balanced and imbalanced datasets. Additional analyses show that our prediction module successfully predicts TD, and the predicted TD is useful in estimating uncertainties of unlabeled data. Our proposed method are illustrated in Figure 1. + +Contributions of our study: (1) We bridge the concept of training dynamics and active learning with the theoretical and experimental evidence that training dynamics is effective in estimating data uncertainty. (2) We propose a new method that efficiently predicts the training dynamics of unlabeled data to estimate their uncertainty. (3) Our proposed method achieves better or comparable performance on both balanced and imbalanced benchmark datasets compared to existing active learning methods. For reproducibility, we release the source code1. + +# 2. Preliminaries + +To better understand our proposed method, we first summarize key concepts, including uncertainty-based active + +learning, quantification of uncertainty, and training dynamics. + +Uncertainty-based active learning. In this work, we focus on uncertainty-based AL for multi-class classification problems. We define the predicted probabilities of the given sample $x$ for $C$ classes as: + +$$ +\boldsymbol {p} = \left[ p (1 | x), p (2 | x), \dots , p (C | x) \right] ^ {T} \in [ 0, 1 ] ^ {C}, \tag {1} +$$ + +where we denote the true label of $x$ as $y$ and the classifier as $f$ . $\mathcal{D}$ and $\mathcal{D}_u$ denote a labeled dataset and an unlabeled data pool, respectively. The general cycle of uncertainty-based AL is in two steps: (1) train the target classifier $f$ on the labeled dataset $\mathcal{D}$ and (2) select top- $k$ uncertain data samples from the unlabeled data pool $\mathcal{D}_u$ . Selected samples are then given to the human annotators to expand the labeled dataset $\mathcal{D}$ , cycling back to the first step. + +Quantifying uncertainty. The objective of this study is to establish a connection between the concept of TD and the field of AL. In order to clearly demonstrate the effectiveness of utilizing TD to quantify data uncertainty, we have employed two of the most prevalent and straightforward estimators, entropy [43] and margin [41], to measure data uncertainty in this paper. Entropy $H$ is defined as follows: + +$$ +H (\boldsymbol {p}) = - \sum_ {c = 1} ^ {C} p (c | x) \log p (c | x), \tag {2} +$$ + +where the sample $x$ is from the unlabeled data pool $\mathcal{D}_u$ . Entropy concentrates on the level of the model's confidence on the given sample $x$ and gets bigger when the prediction across the classes becomes uniform (i.e., uncertain). Margin $M$ measures the difference between the probability of the true label and the maximum of the others: + +$$ +M (\boldsymbol {p}) = p (y | x) - \max _ {c \neq \hat {y}} p (c | x), \tag {3} +$$ + +where $y$ denotes the true label. The smaller the margin, the lower the model's confidence in the sample, so it can be considered uncertain. Both entropy and margin are computed with the predicted probabilities $p$ of the fully trained classifier $f$ , only taking the snapshot of $f$ into account. + +Defining training dynamics. Our TiDAL targets to leverage TD of unlabeled data to estimate their uncertainties. TD can be defined as any model behavior during optimization, such as the area under the margin between logit values of the target class and the other largest class [39] or the variance of the predicted probabilities generated at each epoch [47]. In this work, we define the TD $\bar{p}^{(t)}$ as the area under the predicted probabilities of each data sample $x$ obtained + +![](images/eac8b6fa1fbad8f46dffbf5a2823b54aa7e8300a9a3b224b270ac33614f68f33.jpg) +(a) Entropy Distribution +Figure 2: Score distribution after long-tailed training. We plot the marginal distributions using kernel density estimation (KDE). It is difficult to separate major (certain) and minor (uncertain) samples by the model snapshot-based scores (horizontal), unlike the TD-driven scores (vertical) that enable clearly separating the certain and uncertain samples. + +![](images/23a522c3282f3f5d58ab0d0f03fdf43768b9ed46a90b35ceb90b3f9da570de87.jpg) +(b) Margin Distribution + +during the $t$ time steps of optimizing the target classifier $f$ : + +$$ +\boldsymbol {p} ^ {(i)} = \left[ p ^ {(i)} (1 | x), p ^ {(i)} (2 | x), \dots , p ^ {(i)} (C | x) \right] ^ {T}, \tag {4} +$$ + +$$ +\begin{array}{l} \bar {\boldsymbol {p}} ^ {(t)} = \left[ \bar {p} ^ {(t)} (1 | x), \bar {p} ^ {(t)} (2 | x), \dots , \bar {p} ^ {(t)} (C | x) \right] ^ {T} \\ = \sum_ {\tau} \boldsymbol {p} ^ {(\tau)} \Delta \tau \simeq \sum_ {i = 1} ^ {t} \boldsymbol {p} ^ {(i)} / t, \tag {5} \\ \end{array} +$$ + +where $\pmb{p}^{(i)}$ is the predicted probabilities of a target classifier $f$ at the $i$ -th time step. $\Delta \tau$ is the unit time step to normalize the predicted probabilities. For simplicity, we record $\pmb{p}^{(i)}$ every epoch and choose $\Delta \tau = 1 / t$ , namely, averaging the predicted probabilities during $t$ epochs [47, 46]. The TD $\bar{\pmb{p}}^{(t)}$ takes all the predicted probabilities during model optimization into account. Hence, it encapsulates the overall tendency of the model during $t$ epochs of optimization, avoiding being solely biased towards the snapshot of $\pmb{p}^{(t)}$ in the final epoch $t$ . + +# 3. Is TD Useful for Quantifying Uncertainty? + +In this section, we provide empirical and theoretical evidence to support our argument: $TD$ is more effective in separating uncertain data from certain data than the model snapshot, where the latter is often utilized to quantify data uncertainty in previous works [52, 45]. + +# 3.1. Motivating Observation + +Settings. We aim to observe and compare the behavior of TD and the model snapshot for different sample difficulties. However, it is nontrivial to directly measure sample-wise difficulty, inhibiting the quantitative analysis of data uncertainty. To avoid this, we borrow the theoretical and empirical results of long-tailed visual recognition [33, 8, 19]: it is hard for the deep neural network-based model to train with + +fewer samples. Hence, we regard major and minor class samples to contain many certain and uncertain samples for the model, respectively. We train the target classifier $f$ on the long-tailed dataset during $T$ epochs to obtain the TD and the model snapshot. We apply both approaches to the common estimators, entropy and margin. We denote entropy and margin scores from the model snapshot as $H$ and $M$ . In opposition, we denote the TD-driven scores as $\bar{H}$ and $\bar{M}$ . More details and discussions are described in Appendix B. + +Results. Figure 2 shows the distribution the scores calculated with TD $(x$ -axis) and model snapshot $(y$ -axis). We can observe that scores from TD $(\bar{H}, \bar{M})$ successfully separate the major and the minor class samples, whereas scores from the model snapshot $(H, M)$ fail to do so. We conclude that compared to model snapshots, TD is more helpful in separating uncertain samples from certain samples. + +# 3.2. Theoretical Evidence + +Theorem 1. (Informal) Under the LE-SDE framework [54], with the assumption of local elasticity [17], certain samples and uncertain samples reveal different $TD$ ; especially, certain samples converge quickly than uncertain samples. + +The above theorem discusses different model behaviors depending on the difficulty of the sample. Compared to the uncertain sample, the certain sample has the same class samples nearby, which is the fundamental idea of level set estimation [22] and nearest neighbor [36] literature. We suspect that, due to the local elasticity of deep nets, samples close by have a bigger impact on the certain sample, hence changing its predicted probability more rapidly. As the certain sample is quicker to converge, its TD is larger than that of the uncertain sample. Intuitively, slower to train, struggling the classifier is to learn, hence TD capturing the uncertainty in the classifier's perspective. + +Theorem 2. (Informal) Estimators such as Entropy (Equation 10) and Margin (Equation 11) successfully capture the difference of TD between easy and hard samples even for the case where it cannot be distinguished via the predicted probabilities of the model snapshot. + +The above theorem discusses the validity of entropy and margin on whether they can successfully differentiate between two samples of different TD but with the same final prediction. With Theorem 1, one can conclude that the common estimators' scores calculated with TD are effective in capturing the data uncertainty. Due to the space constraints, we provide the details of the above results in Appendix A. + +# 4. Utilizing TD for Active Learning + +As tracking the TD of all the unlabeled data is computationally infeasible, we devise an efficient method to es + +estimate the TD of unlabeled samples. We train the module that directly predicts the TD of each sample by feeding the training samples, where its TD are freely available during training. Then, based on the predicted TD of each unlabeled sample, we use the common estimators, entropy or margin, to determine which sample is the most uncertain so that human annotators can label it. Hence, in this section, we describe the details of the module that estimates TD (§4.1) and how to train the module (§4.2). Finally, calculating the uncertainties using the module predictions for active learning is illustrated (§4.3). + +# 4.1. Training Dynamics Prediction Module + +As mentioned, it is not computationally feasible to track TD for the large-scale unlabeled data as it requires model inference on all the unlabeled data every training epoch. Thus, we propose the TD prediction module $m$ to efficiently predict the TD of unlabeled data at the $t$ -th epoch. Being influenced by the previous studies [11, 52, 45, 25] that use additional modules to predict useful values such as loss or confidence by the target model outputs, multi-scale feature maps are aggregated and passed into our TD prediction module. The module produces the $C$ -dimensional predictions: + +$$ +\tilde {\boldsymbol {p}} _ {m} ^ {(t)} = \left[ \tilde {p} _ {m} ^ {(t)} (1 | x), \dots , \tilde {p} _ {m} ^ {(t)} (C | x) \right] ^ {T} \in [ 0, 1 ] ^ {C} \tag {6} +$$ + +estimating the actual TD $\bar{\pmb{p}}^{(t)}$ of the given sample $x$ in Equation 5. TD prediction module is jointly trained with the target classifier using a handful of parameters, having a negligible computational cost during training. The detailed architecture of the module is described in Appendix C. + +Even though the architecture is similar to previous works [52, 45, 25], we observed that ours were much more stable during optimization and easier to train. We suspect that it is due to the target task difference; previous works trained the module that outputs only a single value via regression, whereas our module outputs $C$ -dimensional probability distribution, which is similar to the main task of classifying images. + +# 4.2. Training Objectives + +To train the target classifier $f$ at the $t$ -th epoch, we use the cross-entropy loss function $\mathcal{L}_{\mathrm{target}}$ on the predicted probability $\pmb{p}^{(t)}$ and a one-hot encoded vector $\pmb{y} \in \{0,1\}^C$ of the true label $y$ : + +$$ +\mathcal {L} _ {\text {t a r g e t}} = \mathcal {L} _ {\mathrm {C E}} \left(\boldsymbol {p} ^ {(t)}, \boldsymbol {y}\right) = - \log p ^ {(t)} (y | x). \tag {7} +$$ + +Meanwhile, the prediction module $m$ learns the TD of a sample $x$ by minimizing the Kullback-Leibler (KL) divergence between the predicted TD $\tilde{p}_m^{(t)}$ and the actual TD $\bar{p}^{(t)}$ : + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {m o d u l e}} = \mathcal {L} _ {\mathrm {K L}} (\bar {\boldsymbol {p}} ^ {(t)} | | \tilde {\boldsymbol {p}} _ {m} ^ {(t)}) \\ = \sum_ {c = 1} ^ {C} \bar {p} ^ {(t)} (c | x) \log \left(\frac {\bar {p} ^ {(t)} (c | x)}{\tilde {p} _ {m} ^ {(t)} (c | x)}\right). \tag {8} \\ \end{array} +$$ + +The final objective function of our proposed method is defined as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {t a r g e t}} + \lambda \mathcal {L} _ {\text {m o d u l e}} \tag {9} +$$ + +where $\lambda$ is a balancing factor to control the effect of $\mathcal{L}_{\mathrm{module}}$ during model training. + +# 4.3. Quantifying Uncertainty with TD + +We argue that uncertain samples can be effectively distinguished from unlabeled data using the predicted TD. To verify the effectiveness of leveraging TD, we feed the predicted TD to entropy and margin (§2) by replacing snapshot probability $\pmb{p}$ with the predicted TD $\bar{\pmb{p}}$ . We choose these estimators as they are widely used for quantifying uncertainty. We feed $\bar{\pmb{p}}$ , replacing $\pmb{p}$ , to the entropy $\bar{H}$ : + +$$ +\bar {H} (\bar {\boldsymbol {p}}) = - \sum_ {c = 1} ^ {C} \bar {p} (c | x) \log \bar {p} (c | x). \tag {10} +$$ + +Entropy $\bar{H}$ is maximized when $\bar{p}$ is uniform, i.e., the sample is uncertain for the target classifier. Margin $\bar{M}$ is also similarly employed: + +$$ +\bar {M} (\bar {\boldsymbol {p}}) = \bar {p} (\hat {y} | x) - \max _ {c \neq \hat {y}} \bar {p} (c | x). \tag {11} +$$ + +Since we do not have true labels of unlabeled samples, we use the predicted labels $\hat{y}$ of the target classifier instead of the true labels. There are several possible variants of $\bar{M}$ depending on the definition of $\hat{y}$ . We conduct experiments to compare $\bar{M}$ with its variants. The experimental details and results are in Appendix D.4. + +At the data selection phase, we use the predicted TD $\tilde{\pmb{p}}_m^{(T)}$ instead of the actual TD $\bar{\pmb{p}}^{(T)}$ as in Equation 10 & 11 to estimate the TD-driven uncertainties of the unlabeled sample $x$ at the final epoch $T$ . By using the estimated uncertainty with the predicted TD, we select the most informative samples for model training. + +# 5. Experiments + +In this section, we experimentally verify the effectiveness of our method, TiDAL, which utilizes the estimated training dynamics from the prediction module to discern uncertain samples from unlabeled data. We describe the detailed settings and the baseline methods for our experiments (§5.1) and show the results on both balanced (§5.2) and imbalanced datasets (§5.3). We further analyze whether the TD prediction module is effective for AL performance and can successfully estimate the TD (§5.4). We end the section by discussing the potential limitations of our method (§5.5). + +# 5.1. Experimental Setup + +Datasets. To assess the performance of our proposed method and baseline methods, we conduct experiments on + +![](images/d9f583e97a79654ab095d0371233dc233538617bb324361ac5209f944045735e.jpg) +Figure 3: Averaged relative accuracy improvement curves and their $95\%$ confidence interval (shaded) of AL methods over the number of labeled samples on balanced datasets. TiDAL $(\bar{H})$ and TiDAL $(\bar{M})$ denote the performance of TiDAL when with entropy $\bar{H}$ and margin $\bar{M}$ as the data uncertainty estimation strategy, respectively. + +![](images/8034f319a656da524e8075fec458aec9bfba223f71708b8a784ab8590b040a56.jpg) + +![](images/b20e847e82013244bf3d032003e0fee0f58fbce49a4f5b44e908178f9c452e72.jpg) + +the following five datasets: CIFAR10/100 [27], FashionMNIST [51], SVHN [34], and iNaturalist2018 [50]. Since CIFAR and FashionMNIST are both balanced, we further modify them to simulate the data imbalance in the real world, following the previous long-tail visual recognition studies [8, 33, 56, 19]. The imbalance ratio is defined as $N_{\mathrm{max}} / N_{\mathrm{min}}$ where $N$ is the number of samples in each class. We make two variants with data imbalance ratios 10 and 100 for each dataset. Unlike the above, SVHN and iNaturalist18 are already imbalanced. Especially, iNaturalist2018 is commonly chosen to demonstrate how methods work in imbalanced real-world settings. The dataset statistics are summarized in Appendix D. + +Baselines. For a fair comparison, we compare our TiDAL with the following baselines which train a target classifier with only labeled data. Random sampling: a simple baseline that randomly selects data samples from the unlabeled dataset. Entropy sampling [43]: an uncertainty-based method that selects data samples based on the maximum entropy. BALD [13]: an uncertainty-based method that selects data samples based on the mutual information between the model prediction and the posterior. CoreSet [42]: a diversity-based method that selects representative data samples covering all data through a minimum radius. LLoss [52]: an uncertainty-based method that learns to estimate the errors of the predictions (loss) made by the learner and select data samples based on the predicted loss. CAL [55]: recent work on using TD, gathering sample-wise TD information on whether the classifier was consistently correct or not during training. CAL splits the samples into two classes by applying a heuristic threshold to the TD information to train a binary classifier that outputs uncertainty score. To verify the effectiveness of TiDAL, we further compare it with the two semi-supervised AL methods, VAAL [45] and TA-VAAL [25] in Appendix D.3. Note that + +these methods further utilize unlabeled data for training the selection module, thus it is unfair for our TiDAL. + +Active learning setting. We follow the same setting from [6, 52] for the detailed AL settings. For the initial step, we randomly select initial samples to be annotated from the unlabeled dataset, where we use them to train the initial target classifier. Then, we obtain a random subset from the unlabeled data pool $\mathcal{D}_u$ to choose the top- $k$ samples based on the criterion of each method, where those samples will be annotated. We repeat the above cycle, training a classifier from scratch from the continuously expanding labeled set. + +Implementation details. For a fair comparison, we use the same backbone network ResNet-18 [18] except for iNaturalist2018, where we use ResNet-50 [18] pretrained on ImageNet [12]. All models are trained with SGD optimizer with momentum 0.9, weight decay $5 \cdot 10^{-4}$ , and learning rate (LR) decay of 0.1. For CIFAR10/100 and SVHN, we train the model for 200 epochs with an initial LR of 0.1 and decay at epoch 160. For FashionMNIST, 100 epochs with an initial LR of 0.1 and decay at epoch 80. For iNaturalist2018, 50 epochs with an initial LR of 0.01 and decay at epoch 40. For CIFAR10/100, SVHN and FashionMNIST, we set the batch size and the unlabeled subset size to be 128 and $10^{4}$ , respectively. For iNaturalist2018, which is much larger than other datasets, we set the batch size and the unlabeled subset size to 256 and $10^{6}$ , respectively. We set the balancing factor to 1.0. + +Evaluation details. To compare with other state-of-the-art baselines, we show the average accuracy and $95\%$ confidence interval with three trials. We mainly compare the model performances with relative accuracy improvement to random sampling, demonstrating how much it improves + +![](images/1181e6314028dd29509618f6ab64eb40c97a6c7a6371fd6d3f7e069ee71d9386.jpg) + +![](images/a4ea931fe8bad4459808ffbaddfa44848cb21a44fd233cc67f9951f974108930.jpg) + +![](images/7501908db5eb3fc363adb5751b189283cae593b9f0b1da9b67f4dc63629bda90.jpg) + +![](images/f87e300182c62cb525f75ed87fbd59409b5395407b9d904ffb59196eaadf43f7.jpg) +Figure 4: Averaged relative accuracy improvement curves and their $95\%$ confidence interval (shaded) of AL methods over the number of labeled samples on synthetically imbalanced datasets. We use the imbalance ratio (IR) of 10 and 100 on CIFAR10, CIFAR100, and FashionMNIST. + +![](images/27c6d797034dd96d42bc087dd208c14a39ff9acb99b61c61347020cba1712be1.jpg) + +![](images/b8758b58439a1df2c3596da18174104d72945c87fb13f0f2a81b9f952175895a.jpg) + +upon the naive approach on each cycle. Additionally, absolute accuracy is also plotted in Appendix D. + +# 5.2. Results on Balanced Datasets + +Figure 3 and 10 compare our TiDAL against the state-of-the-art methods on various balanced datasets: CIFAR10, CIFAR100, and FashionMNIST. For all the datasets, the two variants of TiDAL outperform all the baselines at all AL cycles except for LLoss, which shows better improvement than TiDAL $(\bar{M})$ on CIFAR10 with an imbalance ratio of 100. Nonetheless, our TiDAL achieves the best final performance compared to all the baselines. CAL, which uses training dynamics, generally underperforms compared to others. We suspect that CAL is sensitive to its threshold hyperparameter. + +# 5.3. Results on Imbalanced Datasets + +Synthetically imbalanced datasets. Similar to the above, Figure 4, 9, and 11 shows the performance improvements on the synthetically imbalanced datasets with the two imbalance ratios, 10 and 100. Except for the CIFAR10 with an imbalance ratio of 100, our methods show superb performance across all the imbalanced settings. TiDAL per + +forms especially well with a small variance in imbalanced CIFAR100, where the number of classes is the largest. In imbalanced FashionMNIST, the performance quickly rises to 2.5k labeled images and then saturates. This implies that FashionMNIST is easier than other datasets, and needs to focus more on the early training steps to compare with other models. TiDAL also shows overall better performance on FashionMNIST, especially in the early steps. + +Real-world imbalanced datasets. Figure 5 and 10 shows evaluation results on real-world imbalanced datasets. For iNaturalist2018, which is the large-scale long-tailed classification dataset, TiDAL shows outstanding performance compared to other methods. For SVHN, TiDAL shows the best improvements with low variance as the number of labeled images increases except for the initial stage. LLoss shows outstanding performance only in the initial stage, where we presume that the loss prediction module of LLoss acts as a regularizer during model optimization. + +# 5.4. Analysis on the TD Prediction Module + +Effectiveness of the TD prediction module. In order to verify the efficacy of using the predicted TD $\tilde{p}_m$ , we con + +![](images/e1b6e5833c32890dc326cc2e42ed900d4986818fbc0b271c334d6ff9c0932fb8.jpg) +Figure 5: Averaged relative accuracy improvement curves and its $95\%$ confidence interval (shaded) of AL methods over the number of labeled samples on real-world imbalanced datasets: iNaturalist2018 and SVHN. For SVHN, LLoss shows a substantial improvement of $20.02\% \pm 6.77\%$ at the initial phase (1k), but we clip the plot to show the performance afterward more clearly. + +![](images/2686bdbb634a9f14026082c31130e56d5441311031f45a813b9c7f9978bf7e43.jpg) + +![](images/8d4ce1cc952aeb5a31d048225d4897d73ce8aa1770257ff510568f117e73fb59.jpg) +Figure 6: Ablation test results. $\bar{H} (\tilde{\pmb{p}}_m)$ and $\bar{M} (\tilde{\pmb{p}}_m)$ use the predicted TD $\tilde{\pmb{p}}_m$ of the prediction module $m$ . In contrast, $H(\pmb {p})$ and $M(\pmb {p})$ use the predicted probability of the model snapshot $\pmb{p}$ . TD shows better performance than the model snapshot, implying that TD is better at quantifying data uncertainty. + +![](images/90395acbbb13988562b801aed1ebebd2b906b556b710cc159dab0d0d644bd9b4.jpg) +Figure 7: KL divergence scores of the actual TD $\bar{\pmb{p}}^{(T)}$ with the predicted TD $\tilde{\pmb{p}}_m^{(t)}$ and the predicted probability of the model snapshot $\pmb{p}^{(t)}$ , respectively, during model optimization. Our predicted TD can accurately approximate the actual TD. + +![](images/0a0a8388e1446991ed393ec952424c5f38a90460fcd543e84cf9adb9e0c90bfb.jpg) + +duct an ablation test that compares the performance between when using and not using the TD prediction module $m$ . Figure 6 shows the results on balanced CIFAR10/100. We observe that $\bar{H}(\tilde{\pmb{p}}_m)$ and $\bar{M}(\tilde{\pmb{p}}_m)$ using the predicted TD $\tilde{\pmb{p}}_m$ to estimate the data uncertainty significantly outperform the methods $H(\pmb{p})$ and $M(\pmb{p})$ that use only the final predicted probabilities $\pmb{p}$ of the target classifier $f$ , showing better performance in the whole training cycle. Even $M(\pmb{p})$ shows temporary improvement in earlier steps on CIFAR100, $\bar{H}(\tilde{\pmb{p}}_m)$ and $\bar{M}(\tilde{\pmb{p}}_m)$ maintain stable improvement, eventually winning over $M(\pmb{p})$ . This indicates that the predicted TD $\tilde{\pmb{p}}_m$ of the TD prediction module $m$ produces better data uncertainty estimation than the predicted + +probability $\pmb{p}$ of the target classifier $f$ . + +Predictive performance of the TD prediction module. We verify whether the TD prediction module $m$ accurately predicts the actual TD $\bar{p}$ . Its prediction performance is crucial as we use the predicted TD $\tilde{\boldsymbol{p}}_m$ of the module $m$ to quantify uncertainties of unlabeled data. Using the KL divergence $\mathcal{L}_{\mathrm{KL}}$ , we analyze that the predicted TD $\tilde{\boldsymbol{p}}_m$ converges to the actual TD $\bar{p}$ at the data selection phase. We calculate $\mathcal{L}_{\mathrm{KL}}(\bar{\boldsymbol{p}}^{(T)}||\tilde{\boldsymbol{p}}_m^{(t)})$ and compare it with $\mathcal{L}_{\mathrm{KL}}(\bar{\boldsymbol{p}}^{(T)}||\boldsymbol{p}^{(t)})$ which is set as a baseline computed with the actual TD $\bar{p}$ and the predicted probabilities $\boldsymbol{p}$ (snapshot) of the target classifier $f$ . In this analysis, we use the bal + +anced CIFAR10 where the sample-wise averaged KL divergence scores are computed on the test set. Figure 7 shows that the final predicted TD successfully approximates the actual TD, while the predicted probability is highly different from the actual TD. We conclude that the TD prediction module $m$ can produce the TD efficiently, leading to performance improvement, and the predicted TD acts as a better approximation of the actual TD than the predicted probability of a model snapshot captured at each epoch. + +# 5.5. Limitations + +We found two potential limitations of our TiDAL derived from the fact that it relies on the outputs of the target classifier to compute the TD. First, TiDAL is designed only for classification tasks, and thus it cannot be applied to AL targeting other tasks, such as regression [10, 15]. Second, TiDAL is highly influenced by the performance of the target classifier, especially when the target classifier wrongly classifies the hard negative samples with a high confidence during model optimization. These samples can be treated as certain samples (i.e. will not be selected for annotation) because they have low estimated uncertainties from the predicted TD, even though the target classifier fails to predict the true label of the samples correctly. As a future work, we will study extending our TiDAL in the task-agnostic ways with a safeguard combating the wrongly classified samples. + +# 6. Related Work + +# 6.1. Active Learning + +AL methods target to construct a dataset with the most useful samples based on the assumption that each sample has different importance in model training [40]. Two mainstream AL approaches exist for efficiently querying the unlabeled data: pool-based methods [31, 52, 45] use various ways to extract samples from an unlabeled data pool effectively, and synthesis-based methods [1, 58, 49] generate informative samples for the model. Pool-based methods can be roughly divided based on query strategies: uncertainty-based [13, 52, 45, 20] and diversity-based [42, 14, 38] methods, where some methods use the hybrid of both [4, 44, 25]. Uncertainty-based methods focus on finding which samples would be the most uncertain for the model, whereas diversity-based methods aim to construct a subset of representative samples of the input distribution. Our proposed method, TiDAL, lies in uncertainty-based methods. The significant difference between TiDAL and previous uncertainty-based methods is that TiDAL estimates data uncertainty using TD that contains additional hints generated during model training. In contrast, the previous methods leverage only static information (e.g., loss [52, 20] and predicted probabilities [13, 45, 25]) obtained by a model snapshot at the data selection phase. + +# 6.2. Training Dynamics + +TD focuses on how deep neural networks are optimized under back-propagation-based stepwise weight updates. Many studies try to understand how gradient descent can effectively obtain the global minimum by analyzing the loss landscape of neural networks [24, 32] or its loss trajectory [3]. Some also import alternative models that are more mathematically approachable to analyze, such as neural tangent kernels [21], deep Gaussian processes [30], or stochastic differential equations [54]. On the other hand, the phenomenological and practical viewpoint of TD also exists. [48] coin the term Forgetting Dynamics to assert that unforgettable samples are often less helpful, and [9] show that the model could prefer samples that are often wrongly predicted throughout model training. TD is also commonly used in noisy label literature to find potential noisy labels as they tend to fit later on model training [2, 39] or locate samples that can be relabeled correctly [46]. Furthermore, [57] calculate the Dynamic Instance Hardness score by monitoring losses of each sample or whether the prediction gets flipped so that higher scored samples can be prioritized for curriculum learning, and [23] feed the loss history to the auxiliary neural network to mediate the curriculum for training. [29] also introduce temporal ensembling for semi-supervised learning, where the model fits towards averaged probability outputs. [47, 37] devise Data Maps to inspect datasets with two TD measures; confidence and its variability across epochs on the true class prediction. [55] further extend the Data Maps for AL, whether the target classifier was consistently correct or not during training. The proposed method splits the labeled samples by applying a heuristic threshold on the level of consistency to train a binary classifier that is trained to discern uncertain samples. Even though the work, similar to ours, also utilizes TD, it relies on empirical observations and heuristic choices to separate the certain and uncertain samples. In this study, we link the concept of TD to AL with both empirical and theoretical results to estimate the uncertainty of unlabeled samples, which is often neglected in previous TD studies. + +# 7. Conclusion + +We propose a novel active learning method, Training Dynamics for Active Learning (TiDAL), by linking the concept of training dynamics to active learning. We provide motivating observations and theoretical evidence for using training dynamics to estimate the uncertainty of unlabeled data. Since tracking the training dynamics of large-scale unlabeled data is infeasible, TiDAL utilizes a training dynamics prediction module to efficiently predict the training dynamics of the unlabeled data. Based on the predicted training dynamics, TiDAL quantifies data uncertainty using the common uncertainty estimators: entropy and mar + +gin. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness of our method, surpassing the existing state-of-the-art active learning methods. We further analyze that using our training dynamics prediction module is effective and the module successfully predicts the TD of unlabeled data. + +# References + +[1] Dana Angluin. Queries and concept learning. Machine learning, 2(4):319-342, 1988. 8 +[2] Eric Arazo, Diego Ortego, Paul Albert, Noel O'Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In International conference on machine learning, pages 312-321. PMLR, 2019. 8 +[3] Sanjeev Arora, Nadav Cohen, Noah Golowich, and Wei Hu. A convergence analysis of gradient descent for deep linear neural networks. In International Conference on Learning Representations, 2018. 8 +[4] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671, 2019.8 +[5] Les E Atlas, David A Cohn, and Richard E Ladner. Training connectionist networks with queries and selective sampling. In Advances in neural information processing systems, pages 566-573. CiteSeer, 1990. 1 +[6] William H Beluch, Tim Genewein, Andreas Nurnberger, and Jan M Kohler. The power of ensembles for active learning in image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9368-9377, 2018. 1, 5 +[7] Javad Zolfaghari Bengar, Joost van de Weijer, Laura Lopez Fuentes, and Bogdan Raducanu. Class-balanced active learning for image classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1536-1545, 2022. 15 +[8] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in Neural Information Processing Systems, 32:1567-1578, 2019. 3, 5, 15 +[9] Haw-Shiuan Chang, Erik Learned-Miller, and Andrew McCallum. Active bias: Training more accurate neural networks by emphasizing high variance samples. Advances in Neural Information Processing Systems, 30, 2017. 1, 8 +[10] David Cohn, Les Atlas, and Richard Ladner. Improving generalization with active learning. Machine learning, 15(2):201-221, 1994. 8 +[11] Charles Corbière, Nicolas Thome, Avner Bar-Hen, Matthieu Cord, and Patrick Pérez. Addressing failure prediction by learning model confidence. Advances in Neural Information Processing Systems, 32, 2019. 4 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 + +[13] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International Conference on Machine Learning, pages 1183-1192. PMLR, 2017. 1, 5, 8 +[14] Daniel Gissin and Shai Shalev-Shwartz. Discriminative active learning. arXiv preprint arXiv:1907.06347, 2019. 1, 8 +[15] Jia Gong, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Meta agent teaming active learning for pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079-11089, 2022. 8 +[16] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 15 +[17] Hangfeng He and Weijie Su. The local elasticity of neural networks. In International Conference on Learning Representations, 2019. 3, 12 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 15 +[19] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6626-6636, 2021. 3, 5, 15 +[20] Siyu Huang, Tianyang Wang, Haoyi Xiong, Jun Huan, and Dejing Dou. Semi-supervised active learning with temporal output discrepancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3447-3456, 2021. 8 +[21] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 8 +[22] Heinrich Jiang, Been Kim, Melody Guan, and Maya Gupta. To trust or not to trust a classifier. Advances in neural information processing systems, 31, 2018. 3, 12 +[23] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International Conference on Machine Learning, pages 2304–2313. PMLR, 2018. 8 +[24] Kenji Kawaguchi. Deep learning without poor local minima. Advances in neural information processing systems, 29, 2016. 8 +[25] Kwanyoung Kim, Dongwon Park, Kwang In Kim, and Se Young Chun. Task-aware variational adversarial active learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8166-8175, 2021. 4, 5, 8, 16 +[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 15 +[27] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 5 + +[28] Alex Krizhevsky, Vinod Nair, and Geofrey Hinton. The CIFar-10 dataset. http://www.cs.toronto.edu/kriz/cifar.html, 55(5), 2014. 15 +[29] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In International Conference on Learning Representations (ICLR). OpenReview.net, 2017. 1, 8 +[30] Jaehoon Lee, Yasaman Bahri, Roman Novak, Samuel S Schoenholz, Jeffrey Pennington, and Jascha Sohl-Dickstein. Deep neural networks as gaussian processes. In International Conference on Learning Representations, 2018. 8 +[31] David D Lewis and William A Gale. A sequential algorithm for training text classifiers. In Proceedings of the 17th annual international ACM SIGIR conference on Research and development in information retrieval, pages 3-12, 1994. 1, 8 +[32] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018. 8 +[33] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 3, 5, 15 +[34] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 5 +[35] Curtis Northcutt, Lu Jiang, and Isaac Chuang. Confident learning: Estimating uncertainty in dataset labels. Journal of Artificial Intelligence Research, 70:1373-1411, 2021. 15 +[36] Nicolas Papernot and Patrick McDaniel. Deep k-nearest neighbors: Towards confident, interpretable and robust deep learning. arXiv preprint arXiv:1803.04765, 2018. 3, 12 +[37] Seo Yeon Park and Cornelia Caragea. A data cartography based mixup for pre-trained language models. arXiv preprint arXiv:2205.03403, 2022. 8 +[38] Amin Parvaneh, Ehsan Abbasnejad, Damien Teney, Gholam-reza Reza Haffari, Anton van den Hengel, and Javen Qinfeng Shi. Active learning by feature mixing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12237-12246, 2022. 8 +[39] Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. Identifying unlabeled data using the area under the margin ranking. Advances in Neural Information Processing Systems, 33:17044-17056, 2020. 2, 8, 15 +[40] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning. ACM Computing Surveys (CSUR), 54(9):1-40, 2021. 8 +[41] Dan Roth and Kevin Small. Margin-based active learning for structured output spaces. In European Conference on Machine Learning, pages 413-424. Springer, 2006. 2 +[42] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 1, 5, 8 + +[43] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 2, 5 +[44] Changjian Shui, Fan Zhou, Christian Gagné, and Boyu Wang. Deep active learning: Unified and principled method for query and training. In International Conference on Artificial Intelligence and Statistics, pages 1308-1318. PMLR, 2020. 8 +[45] Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell. Variational adversarial active learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5972-5981, 2019. 2, 3, 4, 5, 8, 16 +[46] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In International Conference on Machine Learning, pages 5907-5915. PMLR, 2019. 3, 8 +[47] Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A Smith, and Yejin Choi. Dataset cartography: Mapping and diagnosing datasets with training dynamics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9275-9293, 2020. 1, 2, 3, 8 +[48] Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J Gordon. An empirical study of example forgetting during deep neural network learning. In International Conference on Learning Representations, 2018. 1, 8, 15 +[49] Toan Tran, Thanh-Toan Do, Ian Reid, and Gustavo Carneiro. Bayesian generative active deep learning. In International Conference on Machine Learning, pages 6295-6304. PMLR, 2019. 8 +[50] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 5 +[51] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 5 +[52] Donggeun Yoo and In So Kweon. Learning loss for active learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 93-102, 2019. 1, 2, 3, 4, 5, 8, 15, 16 +[53] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 64(3):107-115, 2021. 12 +[54] Jiayao Zhang, Hua Wang, and Weijie Su. Imitating deep learning dynamics via locally elastic stochastic differential equations. Advances in Neural Information Processing Systems, 34, 2021. 3, 8, 12 +[55] Mike Zhang and Barbara Plank. Cartography active learning. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 395-406, 2021. 5, 8 + +[56] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, 2020. 5, 15 +[57] Tianyi Zhou, Shengjie Wang, and Jeffrey Bilmes. Curriculum learning by dynamic instance hardness. Advances in Neural Information Processing Systems, 33:8602-8613, 2020. 8 +[58] Jia-Jie Zhu and José Bento. Generative adversarial active learning. arXiv preprint arXiv:1702.07956, 2017. 8 \ No newline at end of file diff --git a/2023/TiDAL_ Learning Training Dynamics for Active Learning/images.zip b/2023/TiDAL_ Learning Training Dynamics for Active Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..da92bc9c1076ced098b4d6241c283244668b8aa6 --- /dev/null +++ b/2023/TiDAL_ Learning Training Dynamics for Active Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:969478e0d62c865de231f0ae9fe76ef3d85cc0f24cec3e443959ca8ac524c31f +size 591052 diff --git a/2023/TiDAL_ Learning Training Dynamics for Active Learning/layout.json b/2023/TiDAL_ Learning Training Dynamics for Active Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f9078aa5344f4a2e3433e8941b3bd78cdb60c228 --- /dev/null +++ b/2023/TiDAL_ Learning Training Dynamics for Active Learning/layout.json @@ -0,0 +1,9939 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 476, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 476, + 121 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 476, + 121 + ], + "type": "text", + "content": "TiDAL: Learning Training Dynamics for Active Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "text", + "content": "Seong Min Kye" + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "text", + "content": " Kwanghee Choi" + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "text", + "content": " Hyeongmin Byun" + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "text", + "content": " Buru Chang" + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "inline_equation", + "content": "^{3,*}" + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "text", + "content": "Hyperconnect " + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "text", + "content": "Carnegie Mellon University " + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 115, + 142, + 481, + 174 + ], + "type": "text", + "content": "Sogang University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 174, + 505, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 174, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 88, + 174, + 505, + 185 + ], + "type": "text", + "content": "{harris,hyeongmin.byun}@hpcnt.com kwanghec@andrew.cmu.edu buru@sogang.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "type": "text", + "content": "Active learning (AL) aims to select the most useful data samples from an unlabeled data pool and annotate them to expand the labeled dataset under a limited budget. Especially, uncertainty-based methods choose the most uncertain samples, which are known to be effective in improving model performance. However, previous methods often overlook training dynamics (TD), defined as the ever-changing model behavior during optimization via stochastic gradient descent, even though other research areas have empirically shown that TD provides important clues for measuring the data uncertainty. In this paper, we first provide theoretical and empirical evidence to argue the usefulness of utilizing the ever-changing model behavior rather than the fully trained model snapshot. We then propose a novel AL method, Training Dynamics for Active Learning (TiDAL), which efficiently predicts the training dynamics of unlabeled data to estimate their uncertainty. Experimental results show that our TiDAL achieves better or comparable performance on both balanced and imbalanced benchmark datasets compared to state-of-the-art AL methods, which estimate data uncertainty using only static information after model training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 524, + 128, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 524, + 128, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 524, + 128, + 537 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 544, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 544, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 45, + 544, + 287, + 567 + ], + "type": "text", + "content": "\"There is a tide in the affairs of men. Which taken at the flood, leads on to fortune.\" — William Shakespeare" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 568, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 288, + 676 + ], + "type": "text", + "content": "Active learning (AL) [5, 31] aims to solve the real-world problem of selecting the most useful data samples from large-scale unlabeled data pools and annotating them to expand labeled data under a limited budget. Since the current deep neural networks are data-hungry, AL has increasingly gained attention in recent years. Existing AL methods can be divided into two mainstream categories: diversity- and uncertainty-based methods. Diversity-based methods [42, 14] focus on constructing a subset that fol" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 212, + 547, + 430 + ], + "blocks": [ + { + "bbox": [ + 306, + 212, + 547, + 430 + ], + "lines": [ + { + "bbox": [ + 306, + 212, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 547, + 430 + ], + "type": "image", + "image_path": "604e19a021cbe9bfdf754ce80ffa26e7b8dc303a11ccf08b5fcdb6a885d1b41e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 437, + 547, + 535 + ], + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 535 + ], + "type": "text", + "content": "Figure 1: Our proposed TiDAL. TD of training samples " + }, + { + "bbox": [ + 304, + 437, + 547, + 535 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 437, + 547, + 535 + ], + "type": "text", + "content": " may differ even if they converge to the same final predicted probability " + }, + { + "bbox": [ + 304, + 437, + 547, + 535 + ], + "type": "inline_equation", + "content": "p(y^{*}|x)" + }, + { + "bbox": [ + 304, + 437, + 547, + 535 + ], + "type": "text", + "content": " (Upper row). Hence, we are motivated to utilize the readily available rich information generated during training, i.e., leveraging TD. We estimate TD of large-scale unlabeled data using a prediction module instead of tracking the actual TD of all the unlabeled samples to avoid the computational overhead (Lower row)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "text", + "content": "ows the target data distribution. Uncertainty-based methods [13, 6, 52] choose the most uncertain samples, which are known to be effective in improving model performance. Hence, the most critical question for the latter becomes, \"How can we quantify the data uncertainty?\"" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": "In this paper, we leverage training dynamics (TD) to quantify data uncertainty. TD is defined as the ever-changing model behavior on each data sample during optimization via stochastic gradient descent. Recent studies [9, 29, 48, 47] have provided empirical evidence that TD provides important clues for measuring the contribution of each data sample to model performance improvement. Inspired by these studies, we argue that the data uncertainty" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 683, + 126, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 683, + 126, + 693 + ], + "spans": [ + { + "bbox": [ + 57, + 683, + 126, + 693 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 57, + 683, + 126, + 693 + ], + "type": "text", + "content": " Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 694, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 694, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 57, + 694, + 136, + 703 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 703, + 282, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 703, + 282, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 703, + 282, + 713 + ], + "type": "text", + "content": "This work was done while all authors were affiliated with Hyperconnect." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22335" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "of unlabeled data can be estimated with TD. However, most uncertainty-based methods quantify data uncertainty based on static information (e.g., loss [52] or predicted probability [45]) from a fully-trained model \"snapshot,\" neglecting the valuable information generated during training. We further argue that TD is more effective in separating uncertain and certain data than static information from a model snapshot captured after model training. In §3, we provide both theoretical and empirical evidence to support our argument that TD is a valuable tool for quantifying data uncertainty." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 300 + ], + "type": "text", + "content": "Despite its huge potential, TD is not yet actively explored in the domain of AL. This is because AL assumes a massive unlabeled data pool. Previous studies track TD only for the training data every epoch as it can be recorded easily during model optimization. On the other hand, AL targets a large number of unlabeled data, where tracking the TD for each unlabeled sample requires an impractical amount of computation (e.g., inference all the unlabeled samples every training epoch)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 300, + 289, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 516 + ], + "type": "text", + "content": "Therefore, we propose TiDAL (Training Dynamics for Active Learning), a novel AL method that efficiently quantifies the uncertainty of unlabeled data by estimating their TD. We avoid tracking the TD of large-scale unlabeled data every epoch by predicting the TD of unlabeled samples with a TD prediction module. The module is trained with the TD of labeled data, which is readily available during model optimization. During the data selection phase, we predict the TD of unlabeled data with the trained module to quantify their uncertainties. We efficiently obtain TD using the module, which avoids inferring all the unlabeled samples every epoch. Experimental results demonstrate that our TiDAL achieves better or comparable performance to existing AL methods on both balanced and imbalanced datasets. Additional analyses show that our prediction module successfully predicts TD, and the predicted TD is useful in estimating uncertainties of unlabeled data. Our proposed method are illustrated in Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 516, + 289, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 289, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 289, + 635 + ], + "type": "text", + "content": "Contributions of our study: (1) We bridge the concept of training dynamics and active learning with the theoretical and experimental evidence that training dynamics is effective in estimating data uncertainty. (2) We propose a new method that efficiently predicts the training dynamics of unlabeled data to estimate their uncertainty. (3) Our proposed method achieves better or comparable performance on both balanced and imbalanced benchmark datasets compared to existing active learning methods. For reproducibility, we release the source code1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 648, + 133, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 648, + 133, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 648, + 133, + 661 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 287, + 693 + ], + "type": "text", + "content": "To better understand our proposed method, we first summarize key concepts, including uncertainty-based active" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 96 + ], + "type": "text", + "content": "learning, quantification of uncertainty, and training dynamics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 113, + 547, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 113, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 305, + 113, + 547, + 161 + ], + "type": "text", + "content": "Uncertainty-based active learning. In this work, we focus on uncertainty-based AL for multi-class classification problems. We define the predicted probabilities of the given sample " + }, + { + "bbox": [ + 305, + 113, + 547, + 161 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 113, + 547, + 161 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 305, + 113, + 547, + 161 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 113, + 547, + 161 + ], + "type": "text", + "content": " classes as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 170, + 545, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 170, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 318, + 170, + 545, + 186 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} = \\left[ p (1 | x), p (2 | x), \\dots , p (C | x) \\right] ^ {T} \\in [ 0, 1 ] ^ {C}, \\tag {1}", + "image_path": "00a3cb7ce39a8be8940ea0f87f5f6a77cc258d466a15c024a9f16ad57d2fc302.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "text", + "content": "where we denote the true label of " + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "text", + "content": " and the classifier as " + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "text", + "content": " denote a labeled dataset and an unlabeled data pool, respectively. The general cycle of uncertainty-based AL is in two steps: (1) train the target classifier " + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "text", + "content": " on the labeled dataset " + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "text", + "content": " and (2) select top- " + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "text", + "content": " uncertain data samples from the unlabeled data pool " + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "text", + "content": ". Selected samples are then given to the human annotators to expand the labeled dataset " + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 194, + 547, + 291 + ], + "type": "text", + "content": ", cycling back to the first step." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 307, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 307, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 545, + 392 + ], + "type": "text", + "content": "Quantifying uncertainty. The objective of this study is to establish a connection between the concept of TD and the field of AL. In order to clearly demonstrate the effectiveness of utilizing TD to quantify data uncertainty, we have employed two of the most prevalent and straightforward estimators, entropy [43] and margin [41], to measure data uncertainty in this paper. Entropy " + }, + { + "bbox": [ + 305, + 307, + 545, + 392 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 305, + 307, + 545, + 392 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 400, + 545, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 400, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 351, + 400, + 545, + 422 + ], + "type": "interline_equation", + "content": "H (\\boldsymbol {p}) = - \\sum_ {c = 1} ^ {C} p (c | x) \\log p (c | x), \\tag {2}", + "image_path": "b74b3a25befc2ec0277f5d765de37cde57b7158e9f608b65c5f73138472b5708.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 430, + 547, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 430, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 547, + 502 + ], + "type": "text", + "content": "where the sample " + }, + { + "bbox": [ + 305, + 430, + 547, + 502 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 430, + 547, + 502 + ], + "type": "text", + "content": " is from the unlabeled data pool " + }, + { + "bbox": [ + 305, + 430, + 547, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 305, + 430, + 547, + 502 + ], + "type": "text", + "content": ". Entropy concentrates on the level of the model's confidence on the given sample " + }, + { + "bbox": [ + 305, + 430, + 547, + 502 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 430, + 547, + 502 + ], + "type": "text", + "content": " and gets bigger when the prediction across the classes becomes uniform (i.e., uncertain). Margin " + }, + { + "bbox": [ + 305, + 430, + 547, + 502 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 430, + 547, + 502 + ], + "type": "text", + "content": " measures the difference between the probability of the true label and the maximum of the others:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 513, + 545, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 513, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 361, + 513, + 545, + 533 + ], + "type": "interline_equation", + "content": "M (\\boldsymbol {p}) = p (y | x) - \\max _ {c \\neq \\hat {y}} p (c | x), \\tag {3}", + "image_path": "4bbbe362c80f240c9591005f74bda010331ad0bbb51d89da666afed94ec8cce2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 540, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 540, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 540, + 545, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 540, + 545, + 601 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 540, + 545, + 601 + ], + "type": "text", + "content": " denotes the true label. The smaller the margin, the lower the model's confidence in the sample, so it can be considered uncertain. Both entropy and margin are computed with the predicted probabilities " + }, + { + "bbox": [ + 305, + 540, + 545, + 601 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 540, + 545, + 601 + ], + "type": "text", + "content": " of the fully trained classifier " + }, + { + "bbox": [ + 305, + 540, + 545, + 601 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 540, + 545, + 601 + ], + "type": "text", + "content": ", only taking the snapshot of " + }, + { + "bbox": [ + 305, + 540, + 545, + 601 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 540, + 545, + 601 + ], + "type": "text", + "content": " into account." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 547, + 715 + ], + "type": "text", + "content": "Defining training dynamics. Our TiDAL targets to leverage TD of unlabeled data to estimate their uncertainties. TD can be defined as any model behavior during optimization, such as the area under the margin between logit values of the target class and the other largest class [39] or the variance of the predicted probabilities generated at each epoch [47]. In this work, we define the TD " + }, + { + "bbox": [ + 305, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\bar{p}^{(t)}" + }, + { + "bbox": [ + 305, + 617, + 547, + 715 + ], + "type": "text", + "content": " as the area under the predicted probabilities of each data sample " + }, + { + "bbox": [ + 305, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 617, + 547, + 715 + ], + "type": "text", + "content": " obtained" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 242, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 242, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 242, + 713 + ], + "type": "text", + "content": "1https://github.com/hyperconnect/TiDAL" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22336" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 167, + 189 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 167, + 189 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 167, + 189 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 167, + 189 + ], + "type": "image", + "image_path": "eac8b6fa1fbad8f46dffbf5a2823b54aa7e8300a9a3b224b270ac33614f68f33.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 196, + 153, + 206 + ], + "lines": [ + { + "bbox": [ + 64, + 196, + 153, + 206 + ], + "spans": [ + { + "bbox": [ + 64, + 196, + 153, + 206 + ], + "type": "text", + "content": "(a) Entropy Distribution" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 287, + 282 + ], + "type": "text", + "content": "Figure 2: Score distribution after long-tailed training. We plot the marginal distributions using kernel density estimation (KDE). It is difficult to separate major (certain) and minor (uncertain) samples by the model snapshot-based scores (horizontal), unlike the TD-driven scores (vertical) that enable clearly separating the certain and uncertain samples." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 73, + 286, + 189 + ], + "blocks": [ + { + "bbox": [ + 170, + 73, + 286, + 189 + ], + "lines": [ + { + "bbox": [ + 170, + 73, + 286, + 189 + ], + "spans": [ + { + "bbox": [ + 170, + 73, + 286, + 189 + ], + "type": "image", + "image_path": "23a522c3282f3f5d58ab0d0f03fdf43768b9ed46a90b35ceb90b3f9da570de87.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 196, + 271, + 206 + ], + "lines": [ + { + "bbox": [ + 185, + 196, + 271, + 206 + ], + "spans": [ + { + "bbox": [ + 185, + 196, + 271, + 206 + ], + "type": "text", + "content": "(b) Margin Distribution" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 304, + 285, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 285, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 285, + 318 + ], + "type": "text", + "content": "during the " + }, + { + "bbox": [ + 47, + 304, + 285, + 318 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 304, + 285, + 318 + ], + "type": "text", + "content": " time steps of optimizing the target classifier " + }, + { + "bbox": [ + 47, + 304, + 285, + 318 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 304, + 285, + 318 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 325, + 287, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 325, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 61, + 325, + 287, + 340 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} ^ {(i)} = \\left[ p ^ {(i)} (1 | x), p ^ {(i)} (2 | x), \\dots , p ^ {(i)} (C | x) \\right] ^ {T}, \\tag {4}", + "image_path": "b70038a7a618e357304fb05555a7903c197acf0f7c420348d8bc9f41f21b5d06.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 342, + 287, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 342, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 61, + 342, + 287, + 380 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\bar {\\boldsymbol {p}} ^ {(t)} = \\left[ \\bar {p} ^ {(t)} (1 | x), \\bar {p} ^ {(t)} (2 | x), \\dots , \\bar {p} ^ {(t)} (C | x) \\right] ^ {T} \\\\ = \\sum_ {\\tau} \\boldsymbol {p} ^ {(\\tau)} \\Delta \\tau \\simeq \\sum_ {i = 1} ^ {t} \\boldsymbol {p} ^ {(i)} / t, \\tag {5} \\\\ \\end{array}", + "image_path": "ae6ce4c1d99d603abd02b5d62e92ac21d8b73e5d16c29cbb507afcd73a28228e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{(i)}" + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": " is the predicted probabilities of a target classifier " + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": "-th time step. " + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\Delta \\tau" + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": " is the unit time step to normalize the predicted probabilities. For simplicity, we record " + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{(i)}" + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": " every epoch and choose " + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\Delta \\tau = 1 / t" + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": ", namely, averaging the predicted probabilities during " + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": " epochs [47, 46]. The TD " + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{p}}^{(t)}" + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": " takes all the predicted probabilities during model optimization into account. Hence, it encapsulates the overall tendency of the model during " + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": " epochs of optimization, avoiding being solely biased towards the snapshot of " + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{(t)}" + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": " in the final epoch " + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 390, + 287, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 521, + 279, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 279, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 279, + 536 + ], + "type": "text", + "content": "3. Is TD Useful for Quantifying Uncertainty?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 542, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 603 + ], + "type": "text", + "content": "In this section, we provide empirical and theoretical evidence to support our argument: " + }, + { + "bbox": [ + 46, + 542, + 287, + 603 + ], + "type": "inline_equation", + "content": "TD" + }, + { + "bbox": [ + 46, + 542, + 287, + 603 + ], + "type": "text", + "content": " is more effective in separating uncertain data from certain data than the model snapshot, where the latter is often utilized to quantify data uncertainty in previous works [52, 45]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 610, + 182, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 182, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 182, + 624 + ], + "type": "text", + "content": "3.1. Motivating Observation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "Settings. We aim to observe and compare the behavior of TD and the model snapshot for different sample difficulties. However, it is nontrivial to directly measure sample-wise difficulty, inhibiting the quantitative analysis of data uncertainty. To avoid this, we borrow the theoretical and empirical results of long-tailed visual recognition [33, 8, 19]: it is hard for the deep neural network-based model to train with" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "fewer samples. Hence, we regard major and minor class samples to contain many certain and uncertain samples for the model, respectively. We train the target classifier " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " on the long-tailed dataset during " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " epochs to obtain the TD and the model snapshot. We apply both approaches to the common estimators, entropy and margin. We denote entropy and margin scores from the model snapshot as " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ". In opposition, we denote the TD-driven scores as " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\bar{H}" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\bar{M}" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ". More details and discussions are described in Appendix B." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "text", + "content": "Results. Figure 2 shows the distribution the scores calculated with TD " + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "inline_equation", + "content": "(x" + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "text", + "content": "-axis) and model snapshot " + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "inline_equation", + "content": "(y" + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "text", + "content": "-axis). We can observe that scores from TD " + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "inline_equation", + "content": "(\\bar{H}, \\bar{M})" + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "text", + "content": " successfully separate the major and the minor class samples, whereas scores from the model snapshot " + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "inline_equation", + "content": "(H, M)" + }, + { + "bbox": [ + 304, + 194, + 545, + 277 + ], + "type": "text", + "content": " fail to do so. We conclude that compared to model snapshots, TD is more helpful in separating uncertain samples from certain samples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 283, + 427, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 283, + 427, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 427, + 295 + ], + "type": "text", + "content": "3.2. Theoretical Evidence" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "text", + "content": "Theorem 1. (Informal) Under the LE-SDE framework [54], with the assumption of local elasticity [17], certain samples and uncertain samples reveal different " + }, + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "inline_equation", + "content": "TD" + }, + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "text", + "content": "; especially, certain samples converge quickly than uncertain samples." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 365, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 510 + ], + "type": "text", + "content": "The above theorem discusses different model behaviors depending on the difficulty of the sample. Compared to the uncertain sample, the certain sample has the same class samples nearby, which is the fundamental idea of level set estimation [22] and nearest neighbor [36] literature. We suspect that, due to the local elasticity of deep nets, samples close by have a bigger impact on the certain sample, hence changing its predicted probability more rapidly. As the certain sample is quicker to converge, its TD is larger than that of the uncertain sample. Intuitively, slower to train, struggling the classifier is to learn, hence TD capturing the uncertainty in the classifier's perspective." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 513, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 545, + 573 + ], + "type": "text", + "content": "Theorem 2. (Informal) Estimators such as Entropy (Equation 10) and Margin (Equation 11) successfully capture the difference of TD between easy and hard samples even for the case where it cannot be distinguished via the predicted probabilities of the model snapshot." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 576, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 576, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 576, + 545, + 661 + ], + "type": "text", + "content": "The above theorem discusses the validity of entropy and margin on whether they can successfully differentiate between two samples of different TD but with the same final prediction. With Theorem 1, one can conclude that the common estimators' scores calculated with TD are effective in capturing the data uncertainty. Due to the space constraints, we provide the details of the above results in Appendix A." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 669, + 488, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 488, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 488, + 684 + ], + "type": "text", + "content": "4. Utilizing TD for Active Learning" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "As tracking the TD of all the unlabeled data is computationally infeasible, we devise an efficient method to es" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22337" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "estimate the TD of unlabeled samples. We train the module that directly predicts the TD of each sample by feeding the training samples, where its TD are freely available during training. Then, based on the predicted TD of each unlabeled sample, we use the common estimators, entropy or margin, to determine which sample is the most uncertain so that human annotators can label it. Hence, in this section, we describe the details of the module that estimates TD (§4.1) and how to train the module (§4.2). Finally, calculating the uncertainties using the module predictions for active learning is illustrated (§4.3)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 209, + 251, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 251, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 251, + 222 + ], + "type": "text", + "content": "4.1. Training Dynamics Prediction Module" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 228, + 287, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 347 + ], + "type": "text", + "content": "As mentioned, it is not computationally feasible to track TD for the large-scale unlabeled data as it requires model inference on all the unlabeled data every training epoch. Thus, we propose the TD prediction module " + }, + { + "bbox": [ + 46, + 228, + 287, + 347 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 228, + 287, + 347 + ], + "type": "text", + "content": " to efficiently predict the TD of unlabeled data at the " + }, + { + "bbox": [ + 46, + 228, + 287, + 347 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 228, + 287, + 347 + ], + "type": "text", + "content": "-th epoch. Being influenced by the previous studies [11, 52, 45, 25] that use additional modules to predict useful values such as loss or confidence by the target model outputs, multi-scale feature maps are aggregated and passed into our TD prediction module. The module produces the " + }, + { + "bbox": [ + 46, + 228, + 287, + 347 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 228, + 287, + 347 + ], + "type": "text", + "content": "-dimensional predictions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 350, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 350, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 70, + 350, + 287, + 365 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {p}} _ {m} ^ {(t)} = \\left[ \\tilde {p} _ {m} ^ {(t)} (1 | x), \\dots , \\tilde {p} _ {m} ^ {(t)} (C | x) \\right] ^ {T} \\in [ 0, 1 ] ^ {C} \\tag {6}", + "image_path": "9b51482b5479473925da659fdbd78380bda16dd89e800cb4793e420ac3ac4023.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 371, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 431 + ], + "type": "text", + "content": "estimating the actual TD " + }, + { + "bbox": [ + 46, + 371, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{p}}^{(t)}" + }, + { + "bbox": [ + 46, + 371, + 287, + 431 + ], + "type": "text", + "content": " of the given sample " + }, + { + "bbox": [ + 46, + 371, + 287, + 431 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 371, + 287, + 431 + ], + "type": "text", + "content": " in Equation 5. TD prediction module is jointly trained with the target classifier using a handful of parameters, having a negligible computational cost during training. The detailed architecture of the module is described in Appendix C." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 432, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 527 + ], + "type": "text", + "content": "Even though the architecture is similar to previous works [52, 45, 25], we observed that ours were much more stable during optimization and easier to train. We suspect that it is due to the target task difference; previous works trained the module that outputs only a single value via regression, whereas our module outputs " + }, + { + "bbox": [ + 46, + 432, + 287, + 527 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 432, + 287, + 527 + ], + "type": "text", + "content": "-dimensional probability distribution, which is similar to the main task of classifying images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 532, + 163, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 163, + 545 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 163, + 545 + ], + "type": "text", + "content": "4.2. Training Objectives" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "text", + "content": "To train the target classifier " + }, + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "text", + "content": "-th epoch, we use the cross-entropy loss function " + }, + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{target}}" + }, + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "text", + "content": " on the predicted probability " + }, + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{(t)}" + }, + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "text", + "content": " and a one-hot encoded vector " + }, + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\pmb{y} \\in \\{0,1\\}^C" + }, + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "text", + "content": " of the true label " + }, + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 551, + 287, + 598 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 80, + 602, + 287, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 602, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 80, + 602, + 287, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t a r g e t}} = \\mathcal {L} _ {\\mathrm {C E}} \\left(\\boldsymbol {p} ^ {(t)}, \\boldsymbol {y}\\right) = - \\log p ^ {(t)} (y | x). \\tag {7}", + "image_path": "30f6b00fd3b8578ebc724cfe6b2cca039022d0b62a979a33aab3e302b1735a0c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 621, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 287, + 659 + ], + "type": "text", + "content": "Meanwhile, the prediction module " + }, + { + "bbox": [ + 47, + 621, + 287, + 659 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 621, + 287, + 659 + ], + "type": "text", + "content": " learns the TD of a sample " + }, + { + "bbox": [ + 47, + 621, + 287, + 659 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 621, + 287, + 659 + ], + "type": "text", + "content": " by minimizing the Kullback-Leibler (KL) divergence between the predicted TD " + }, + { + "bbox": [ + 47, + 621, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\tilde{p}_m^{(t)}" + }, + { + "bbox": [ + 47, + 621, + 287, + 659 + ], + "type": "text", + "content": " and the actual TD " + }, + { + "bbox": [ + 47, + 621, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\bar{p}^{(t)}" + }, + { + "bbox": [ + 47, + 621, + 287, + 659 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 72, + 662, + 287, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 662, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 72, + 662, + 287, + 712 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {m o d u l e}} = \\mathcal {L} _ {\\mathrm {K L}} (\\bar {\\boldsymbol {p}} ^ {(t)} | | \\tilde {\\boldsymbol {p}} _ {m} ^ {(t)}) \\\\ = \\sum_ {c = 1} ^ {C} \\bar {p} ^ {(t)} (c | x) \\log \\left(\\frac {\\bar {p} ^ {(t)} (c | x)}{\\tilde {p} _ {m} ^ {(t)} (c | x)}\\right). \\tag {8} \\\\ \\end{array}", + "image_path": "181fa4c32e35216d987b8d94851517d0a8538c3af2f149e9fa3d1023bff69e10.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "content": "The final objective function of our proposed method is defined as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 375, + 106, + 545, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 106, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 375, + 106, + 545, + 118 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {t a r g e t}} + \\lambda \\mathcal {L} _ {\\text {m o d u l e}} \\tag {9}", + "image_path": "52d7248d07c0bfbdb77d15d1e5abaee12bcdf6bc51dfb2bea6bb655d43d042f2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 126, + 545, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 126, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 305, + 126, + 545, + 151 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 126, + 545, + 151 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 126, + 545, + 151 + ], + "type": "text", + "content": " is a balancing factor to control the effect of " + }, + { + "bbox": [ + 305, + 126, + 545, + 151 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{module}}" + }, + { + "bbox": [ + 305, + 126, + 545, + 151 + ], + "type": "text", + "content": " during model training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 157, + 486, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 157, + 486, + 171 + ], + "spans": [ + { + "bbox": [ + 306, + 157, + 486, + 171 + ], + "type": "text", + "content": "4.3. Quantifying Uncertainty with TD" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 176, + 545, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 176, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 305, + 176, + 545, + 260 + ], + "type": "text", + "content": "We argue that uncertain samples can be effectively distinguished from unlabeled data using the predicted TD. To verify the effectiveness of leveraging TD, we feed the predicted TD to entropy and margin (§2) by replacing snapshot probability " + }, + { + "bbox": [ + 305, + 176, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 305, + 176, + 545, + 260 + ], + "type": "text", + "content": " with the predicted TD " + }, + { + "bbox": [ + 305, + 176, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{p}}" + }, + { + "bbox": [ + 305, + 176, + 545, + 260 + ], + "type": "text", + "content": ". We choose these estimators as they are widely used for quantifying uncertainty. We feed " + }, + { + "bbox": [ + 305, + 176, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{p}}" + }, + { + "bbox": [ + 305, + 176, + 545, + 260 + ], + "type": "text", + "content": ", replacing " + }, + { + "bbox": [ + 305, + 176, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 305, + 176, + 545, + 260 + ], + "type": "text", + "content": ", to the entropy " + }, + { + "bbox": [ + 305, + 176, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\bar{H}" + }, + { + "bbox": [ + 305, + 176, + 545, + 260 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 351, + 267, + 545, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 267, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 351, + 267, + 545, + 289 + ], + "type": "interline_equation", + "content": "\\bar {H} (\\bar {\\boldsymbol {p}}) = - \\sum_ {c = 1} ^ {C} \\bar {p} (c | x) \\log \\bar {p} (c | x). \\tag {10}", + "image_path": "1295e38c9f87a7180272c70a03b15b70f9ab83e96ef4b4923fbf1edcb556de36.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 297, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 545, + 334 + ], + "type": "text", + "content": "Entropy " + }, + { + "bbox": [ + 305, + 297, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\bar{H}" + }, + { + "bbox": [ + 305, + 297, + 545, + 334 + ], + "type": "text", + "content": " is maximized when " + }, + { + "bbox": [ + 305, + 297, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\bar{p}" + }, + { + "bbox": [ + 305, + 297, + 545, + 334 + ], + "type": "text", + "content": " is uniform, i.e., the sample is uncertain for the target classifier. Margin " + }, + { + "bbox": [ + 305, + 297, + 545, + 334 + ], + "type": "inline_equation", + "content": "\\bar{M}" + }, + { + "bbox": [ + 305, + 297, + 545, + 334 + ], + "type": "text", + "content": " is also similarly employed:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 342, + 545, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 342, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 361, + 342, + 545, + 361 + ], + "type": "interline_equation", + "content": "\\bar {M} (\\bar {\\boldsymbol {p}}) = \\bar {p} (\\hat {y} | x) - \\max _ {c \\neq \\hat {y}} \\bar {p} (c | x). \\tag {11}", + "image_path": "4b7e28b96b9be7aa1ca155fcff2c1a4b07c88f9fc2d797c5fa5ed6301c8b9449.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 369, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 441 + ], + "type": "text", + "content": "Since we do not have true labels of unlabeled samples, we use the predicted labels " + }, + { + "bbox": [ + 304, + 369, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 369, + 545, + 441 + ], + "type": "text", + "content": " of the target classifier instead of the true labels. There are several possible variants of " + }, + { + "bbox": [ + 304, + 369, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\bar{M}" + }, + { + "bbox": [ + 304, + 369, + 545, + 441 + ], + "type": "text", + "content": " depending on the definition of " + }, + { + "bbox": [ + 304, + 369, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 369, + 545, + 441 + ], + "type": "text", + "content": ". We conduct experiments to compare " + }, + { + "bbox": [ + 304, + 369, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\bar{M}" + }, + { + "bbox": [ + 304, + 369, + 545, + 441 + ], + "type": "text", + "content": " with its variants. The experimental details and results are in Appendix D.4." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "text", + "content": "At the data selection phase, we use the predicted TD " + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_m^{(T)}" + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "text", + "content": " instead of the actual TD " + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{p}}^{(T)}" + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "text", + "content": " as in Equation 10 & 11 to estimate the TD-driven uncertainties of the unlabeled sample " + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "text", + "content": " at the final epoch " + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 441, + 545, + 514 + ], + "type": "text", + "content": ". By using the estimated uncertainty with the predicted TD, we select the most informative samples for model training." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 524, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 524, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 524, + 388, + 538 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 544, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 664 + ], + "type": "text", + "content": "In this section, we experimentally verify the effectiveness of our method, TiDAL, which utilizes the estimated training dynamics from the prediction module to discern uncertain samples from unlabeled data. We describe the detailed settings and the baseline methods for our experiments (§5.1) and show the results on both balanced (§5.2) and imbalanced datasets (§5.3). We further analyze whether the TD prediction module is effective for AL performance and can successfully estimate the TD (§5.4). We end the section by discussing the potential limitations of our method (§5.5)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 671, + 422, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 422, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 422, + 684 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Datasets. To assess the performance of our proposed method and baseline methods, we conduct experiments on" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22338" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 71, + 220, + 223 + ], + "blocks": [ + { + "bbox": [ + 66, + 71, + 220, + 223 + ], + "lines": [ + { + "bbox": [ + 66, + 71, + 220, + 223 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 220, + 223 + ], + "type": "image", + "image_path": "d9f583e97a79654ab095d0371233dc233538617bb324361ac5209f944045735e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 266 + ], + "type": "text", + "content": "Figure 3: Averaged relative accuracy improvement curves and their " + }, + { + "bbox": [ + 46, + 228, + 547, + 266 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 228, + 547, + 266 + ], + "type": "text", + "content": " confidence interval (shaded) of AL methods over the number of labeled samples on balanced datasets. TiDAL " + }, + { + "bbox": [ + 46, + 228, + 547, + 266 + ], + "type": "inline_equation", + "content": "(\\bar{H})" + }, + { + "bbox": [ + 46, + 228, + 547, + 266 + ], + "type": "text", + "content": " and TiDAL " + }, + { + "bbox": [ + 46, + 228, + 547, + 266 + ], + "type": "inline_equation", + "content": "(\\bar{M})" + }, + { + "bbox": [ + 46, + 228, + 547, + 266 + ], + "type": "text", + "content": " denote the performance of TiDAL when with entropy " + }, + { + "bbox": [ + 46, + 228, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\bar{H}" + }, + { + "bbox": [ + 46, + 228, + 547, + 266 + ], + "type": "text", + "content": " and margin " + }, + { + "bbox": [ + 46, + 228, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\bar{M}" + }, + { + "bbox": [ + 46, + 228, + 547, + 266 + ], + "type": "text", + "content": " as the data uncertainty estimation strategy, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 222, + 71, + 373, + 223 + ], + "blocks": [ + { + "bbox": [ + 222, + 71, + 373, + 223 + ], + "lines": [ + { + "bbox": [ + 222, + 71, + 373, + 223 + ], + "spans": [ + { + "bbox": [ + 222, + 71, + 373, + 223 + ], + "type": "image", + "image_path": "8034f319a656da524e8075fec458aec9bfba223f71708b8a784ab8590b040a56.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 376, + 72, + 526, + 223 + ], + "blocks": [ + { + "bbox": [ + 376, + 72, + 526, + 223 + ], + "lines": [ + { + "bbox": [ + 376, + 72, + 526, + 223 + ], + "spans": [ + { + "bbox": [ + 376, + 72, + 526, + 223 + ], + "type": "image", + "image_path": "b20e847e82013244bf3d032003e0fee0f58fbce49a4f5b44e908178f9c452e72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 289, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 289, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 289, + 437 + ], + "type": "text", + "content": "the following five datasets: CIFAR10/100 [27], FashionMNIST [51], SVHN [34], and iNaturalist2018 [50]. Since CIFAR and FashionMNIST are both balanced, we further modify them to simulate the data imbalance in the real world, following the previous long-tail visual recognition studies [8, 33, 56, 19]. The imbalance ratio is defined as " + }, + { + "bbox": [ + 46, + 280, + 289, + 437 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{max}} / N_{\\mathrm{min}}" + }, + { + "bbox": [ + 46, + 280, + 289, + 437 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 280, + 289, + 437 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 280, + 289, + 437 + ], + "type": "text", + "content": " is the number of samples in each class. We make two variants with data imbalance ratios 10 and 100 for each dataset. Unlike the above, SVHN and iNaturalist18 are already imbalanced. Especially, iNaturalist2018 is commonly chosen to demonstrate how methods work in imbalanced real-world settings. The dataset statistics are summarized in Appendix D." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": "Baselines. For a fair comparison, we compare our TiDAL with the following baselines which train a target classifier with only labeled data. Random sampling: a simple baseline that randomly selects data samples from the unlabeled dataset. Entropy sampling [43]: an uncertainty-based method that selects data samples based on the maximum entropy. BALD [13]: an uncertainty-based method that selects data samples based on the mutual information between the model prediction and the posterior. CoreSet [42]: a diversity-based method that selects representative data samples covering all data through a minimum radius. LLoss [52]: an uncertainty-based method that learns to estimate the errors of the predictions (loss) made by the learner and select data samples based on the predicted loss. CAL [55]: recent work on using TD, gathering sample-wise TD information on whether the classifier was consistently correct or not during training. CAL splits the samples into two classes by applying a heuristic threshold to the TD information to train a binary classifier that outputs uncertainty score. To verify the effectiveness of TiDAL, we further compare it with the two semi-supervised AL methods, VAAL [45] and TA-VAAL [25] in Appendix D.3. Note that" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 280, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 280, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 280, + 545, + 304 + ], + "type": "text", + "content": "these methods further utilize unlabeled data for training the selection module, thus it is unfair for our TiDAL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 320, + 546, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 320, + 546, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 320, + 546, + 429 + ], + "type": "text", + "content": "Active learning setting. We follow the same setting from [6, 52] for the detailed AL settings. For the initial step, we randomly select initial samples to be annotated from the unlabeled dataset, where we use them to train the initial target classifier. Then, we obtain a random subset from the unlabeled data pool " + }, + { + "bbox": [ + 305, + 320, + 546, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 305, + 320, + 546, + 429 + ], + "type": "text", + "content": " to choose the top- " + }, + { + "bbox": [ + 305, + 320, + 546, + 429 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 320, + 546, + 429 + ], + "type": "text", + "content": " samples based on the criterion of each method, where those samples will be annotated. We repeat the above cycle, training a classifier from scratch from the continuously expanding labeled set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 445, + 547, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 636 + ], + "type": "text", + "content": "Implementation details. For a fair comparison, we use the same backbone network ResNet-18 [18] except for iNaturalist2018, where we use ResNet-50 [18] pretrained on ImageNet [12]. All models are trained with SGD optimizer with momentum 0.9, weight decay " + }, + { + "bbox": [ + 304, + 445, + 547, + 636 + ], + "type": "inline_equation", + "content": "5 \\cdot 10^{-4}" + }, + { + "bbox": [ + 304, + 445, + 547, + 636 + ], + "type": "text", + "content": ", and learning rate (LR) decay of 0.1. For CIFAR10/100 and SVHN, we train the model for 200 epochs with an initial LR of 0.1 and decay at epoch 160. For FashionMNIST, 100 epochs with an initial LR of 0.1 and decay at epoch 80. For iNaturalist2018, 50 epochs with an initial LR of 0.01 and decay at epoch 40. For CIFAR10/100, SVHN and FashionMNIST, we set the batch size and the unlabeled subset size to be 128 and " + }, + { + "bbox": [ + 304, + 445, + 547, + 636 + ], + "type": "inline_equation", + "content": "10^{4}" + }, + { + "bbox": [ + 304, + 445, + 547, + 636 + ], + "type": "text", + "content": ", respectively. For iNaturalist2018, which is much larger than other datasets, we set the batch size and the unlabeled subset size to 256 and " + }, + { + "bbox": [ + 304, + 445, + 547, + 636 + ], + "type": "inline_equation", + "content": "10^{6}" + }, + { + "bbox": [ + 304, + 445, + 547, + 636 + ], + "type": "text", + "content": ", respectively. We set the balancing factor to 1.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Evaluation details. To compare with other state-of-the-art baselines, we show the average accuracy and " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " confidence interval with three trials. We mainly compare the model performances with relative accuracy improvement to random sampling, demonstrating how much it improves" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22339" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 221, + 223 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 221, + 223 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 221, + 223 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 221, + 223 + ], + "type": "image", + "image_path": "1181e6314028dd29509618f6ab64eb40c97a6c7a6371fd6d3f7e069ee71d9386.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 223, + 70, + 373, + 223 + ], + "blocks": [ + { + "bbox": [ + 223, + 70, + 373, + 223 + ], + "lines": [ + { + "bbox": [ + 223, + 70, + 373, + 223 + ], + "spans": [ + { + "bbox": [ + 223, + 70, + 373, + 223 + ], + "type": "image", + "image_path": "a4ea931fe8bad4459808ffbaddfa44848cb21a44fd233cc67f9951f974108930.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 375, + 71, + 527, + 223 + ], + "blocks": [ + { + "bbox": [ + 375, + 71, + 527, + 223 + ], + "lines": [ + { + "bbox": [ + 375, + 71, + 527, + 223 + ], + "spans": [ + { + "bbox": [ + 375, + 71, + 527, + 223 + ], + "type": "image", + "image_path": "7501908db5eb3fc363adb5751b189283cae593b9f0b1da9b67f4dc63629bda90.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 65, + 226, + 219, + 377 + ], + "blocks": [ + { + "bbox": [ + 65, + 226, + 219, + 377 + ], + "lines": [ + { + "bbox": [ + 65, + 226, + 219, + 377 + ], + "spans": [ + { + "bbox": [ + 65, + 226, + 219, + 377 + ], + "type": "image", + "image_path": "f87e300182c62cb525f75ed87fbd59409b5395407b9d904ffb59196eaadf43f7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 382, + 547, + 418 + ], + "lines": [ + { + "bbox": [ + 45, + 382, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 45, + 382, + 547, + 418 + ], + "type": "text", + "content": "Figure 4: Averaged relative accuracy improvement curves and their " + }, + { + "bbox": [ + 45, + 382, + 547, + 418 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 45, + 382, + 547, + 418 + ], + "type": "text", + "content": " confidence interval (shaded) of AL methods over the number of labeled samples on synthetically imbalanced datasets. We use the imbalance ratio (IR) of 10 and 100 on CIFAR10, CIFAR100, and FashionMNIST." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 221, + 226, + 373, + 377 + ], + "blocks": [ + { + "bbox": [ + 221, + 226, + 373, + 377 + ], + "lines": [ + { + "bbox": [ + 221, + 226, + 373, + 377 + ], + "spans": [ + { + "bbox": [ + 221, + 226, + 373, + 377 + ], + "type": "image", + "image_path": "27c6d797034dd96d42bc087dd208c14a39ff9acb99b61c61347020cba1712be1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 375, + 226, + 526, + 377 + ], + "blocks": [ + { + "bbox": [ + 375, + 226, + 526, + 377 + ], + "lines": [ + { + "bbox": [ + 375, + 226, + 526, + 377 + ], + "spans": [ + { + "bbox": [ + 375, + 226, + 526, + 377 + ], + "type": "image", + "image_path": "b8758b58439a1df2c3596da18174104d72945c87fb13f0f2a81b9f952175895a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 434, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 460 + ], + "type": "text", + "content": "upon the naive approach on each cycle. Additionally, absolute accuracy is also plotted in Appendix D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 466, + 208, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 208, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 208, + 478 + ], + "type": "text", + "content": "5.2. Results on Balanced Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 484, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 288, + 616 + ], + "type": "text", + "content": "Figure 3 and 10 compare our TiDAL against the state-of-the-art methods on various balanced datasets: CIFAR10, CIFAR100, and FashionMNIST. For all the datasets, the two variants of TiDAL outperform all the baselines at all AL cycles except for LLoss, which shows better improvement than TiDAL " + }, + { + "bbox": [ + 46, + 484, + 288, + 616 + ], + "type": "inline_equation", + "content": "(\\bar{M})" + }, + { + "bbox": [ + 46, + 484, + 288, + 616 + ], + "type": "text", + "content": " on CIFAR10 with an imbalance ratio of 100. Nonetheless, our TiDAL achieves the best final performance compared to all the baselines. CAL, which uses training dynamics, generally underperforms compared to others. We suspect that CAL is sensitive to its threshold hyperparameter." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 623, + 219, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 219, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 219, + 635 + ], + "type": "text", + "content": "5.3. Results on Imbalanced Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Synthetically imbalanced datasets. Similar to the above, Figure 4, 9, and 11 shows the performance improvements on the synthetically imbalanced datasets with the two imbalance ratios, 10 and 100. Except for the CIFAR10 with an imbalance ratio of 100, our methods show superb performance across all the imbalanced settings. TiDAL per" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 434, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 533 + ], + "type": "text", + "content": "forms especially well with a small variance in imbalanced CIFAR100, where the number of classes is the largest. In imbalanced FashionMNIST, the performance quickly rises to 2.5k labeled images and then saturates. This implies that FashionMNIST is easier than other datasets, and needs to focus more on the early training steps to compare with other models. TiDAL also shows overall better performance on FashionMNIST, especially in the early steps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 544, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 665 + ], + "type": "text", + "content": "Real-world imbalanced datasets. Figure 5 and 10 shows evaluation results on real-world imbalanced datasets. For iNaturalist2018, which is the large-scale long-tailed classification dataset, TiDAL shows outstanding performance compared to other methods. For SVHN, TiDAL shows the best improvements with low variance as the number of labeled images increases except for the initial stage. LLoss shows outstanding performance only in the initial stage, where we presume that the loss prediction module of LLoss acts as a regularizer during model optimization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 671, + 509, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 671, + 509, + 684 + ], + "spans": [ + { + "bbox": [ + 305, + 671, + 509, + 684 + ], + "type": "text", + "content": "5.4. Analysis on the TD Prediction Module" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "Effectiveness of the TD prediction module. In order to verify the efficacy of using the predicted TD " + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{p}_m" + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": ", we con" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22340" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 133, + 71, + 296, + 233 + ], + "blocks": [ + { + "bbox": [ + 133, + 71, + 296, + 233 + ], + "lines": [ + { + "bbox": [ + 133, + 71, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 133, + 71, + 296, + 233 + ], + "type": "image", + "image_path": "e1b6e5833c32890dc326cc2e42ed900d4986818fbc0b271c334d6ff9c0932fb8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 293 + ], + "type": "text", + "content": "Figure 5: Averaged relative accuracy improvement curves and its " + }, + { + "bbox": [ + 46, + 243, + 547, + 293 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 243, + 547, + 293 + ], + "type": "text", + "content": " confidence interval (shaded) of AL methods over the number of labeled samples on real-world imbalanced datasets: iNaturalist2018 and SVHN. For SVHN, LLoss shows a substantial improvement of " + }, + { + "bbox": [ + 46, + 243, + 547, + 293 + ], + "type": "inline_equation", + "content": "20.02\\% \\pm 6.77\\%" + }, + { + "bbox": [ + 46, + 243, + 547, + 293 + ], + "type": "text", + "content": " at the initial phase (1k), but we clip the plot to show the performance afterward more clearly." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 71, + 460, + 233 + ], + "blocks": [ + { + "bbox": [ + 299, + 71, + 460, + 233 + ], + "lines": [ + { + "bbox": [ + 299, + 71, + 460, + 233 + ], + "spans": [ + { + "bbox": [ + 299, + 71, + 460, + 233 + ], + "type": "image", + "image_path": "2686bdbb634a9f14026082c31130e56d5441311031f45a813b9c7f9978bf7e43.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 297, + 195, + 458 + ], + "blocks": [ + { + "bbox": [ + 50, + 297, + 195, + 458 + ], + "lines": [ + { + "bbox": [ + 50, + 297, + 195, + 458 + ], + "spans": [ + { + "bbox": [ + 50, + 297, + 195, + 458 + ], + "type": "image", + "image_path": "8d4ce1cc952aeb5a31d048225d4897d73ce8aa1770257ff510568f117e73fb59.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "lines": [ + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "text", + "content": "Figure 6: Ablation test results. " + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "inline_equation", + "content": "\\bar{H} (\\tilde{\\pmb{p}}_m)" + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "inline_equation", + "content": "\\bar{M} (\\tilde{\\pmb{p}}_m)" + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "text", + "content": " use the predicted TD " + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_m" + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "text", + "content": " of the prediction module " + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "text", + "content": ". In contrast, " + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "inline_equation", + "content": "H(\\pmb {p})" + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "inline_equation", + "content": "M(\\pmb {p})" + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "text", + "content": " use the predicted probability of the model snapshot " + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 48, + 467, + 341, + 527 + ], + "type": "text", + "content": ". TD shows better performance than the model snapshot, implying that TD is better at quantifying data uncertainty." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 197, + 297, + 338, + 457 + ], + "blocks": [ + { + "bbox": [ + 197, + 297, + 338, + 457 + ], + "lines": [ + { + "bbox": [ + 197, + 297, + 338, + 457 + ], + "spans": [ + { + "bbox": [ + 197, + 297, + 338, + 457 + ], + "type": "image", + "image_path": "90395acbbb13988562b801aed1ebebd2b906b556b710cc159dab0d0d644bd9b4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 451, + 544, + 523 + ], + "lines": [ + { + "bbox": [ + 351, + 451, + 544, + 523 + ], + "spans": [ + { + "bbox": [ + 351, + 451, + 544, + 523 + ], + "type": "text", + "content": "Figure 7: KL divergence scores of the actual TD " + }, + { + "bbox": [ + 351, + 451, + 544, + 523 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{p}}^{(T)}" + }, + { + "bbox": [ + 351, + 451, + 544, + 523 + ], + "type": "text", + "content": " with the predicted TD " + }, + { + "bbox": [ + 351, + 451, + 544, + 523 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_m^{(t)}" + }, + { + "bbox": [ + 351, + 451, + 544, + 523 + ], + "type": "text", + "content": " and the predicted probability of the model snapshot " + }, + { + "bbox": [ + 351, + 451, + 544, + 523 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{(t)}" + }, + { + "bbox": [ + 351, + 451, + 544, + 523 + ], + "type": "text", + "content": ", respectively, during model optimization. Our predicted TD can accurately approximate the actual TD." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 359, + 306, + 536, + 441 + ], + "blocks": [ + { + "bbox": [ + 359, + 306, + 536, + 441 + ], + "lines": [ + { + "bbox": [ + 359, + 306, + 536, + 441 + ], + "spans": [ + { + "bbox": [ + 359, + 306, + 536, + 441 + ], + "type": "image", + "image_path": "0a0a8388e1446991ed393ec952424c5f38a90460fcd543e84cf9adb9e0c90bfb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": "duct an ablation test that compares the performance between when using and not using the TD prediction module " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": ". Figure 6 shows the results on balanced CIFAR10/100. We observe that " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "\\bar{H}(\\tilde{\\pmb{p}}_m)" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "\\bar{M}(\\tilde{\\pmb{p}}_m)" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": " using the predicted TD " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_m" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": " to estimate the data uncertainty significantly outperform the methods " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "H(\\pmb{p})" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "M(\\pmb{p})" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": " that use only the final predicted probabilities " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": " of the target classifier " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": ", showing better performance in the whole training cycle. Even " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "M(\\pmb{p})" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": " shows temporary improvement in earlier steps on CIFAR100, " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "\\bar{H}(\\tilde{\\pmb{p}}_m)" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "\\bar{M}(\\tilde{\\pmb{p}}_m)" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": " maintain stable improvement, eventually winning over " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "M(\\pmb{p})" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": ". This indicates that the predicted TD " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_m" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": " of the TD prediction module " + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 548, + 289, + 704 + ], + "type": "text", + "content": " produces better data uncertainty estimation than the predicted" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 548, + 460, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 460, + 561 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 460, + 561 + ], + "type": "text", + "content": "probability " + }, + { + "bbox": [ + 306, + 548, + 460, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 306, + 548, + 460, + 561 + ], + "type": "text", + "content": " of the target classifier " + }, + { + "bbox": [ + 306, + 548, + 460, + 561 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 306, + 548, + 460, + 561 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "content": "Predictive performance of the TD prediction module. We verify whether the TD prediction module " + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "content": " accurately predicts the actual TD " + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\bar{p}" + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "content": ". Its prediction performance is crucial as we use the predicted TD " + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{p}}_m" + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "content": " of the module " + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "content": " to quantify uncertainties of unlabeled data. Using the KL divergence " + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}" + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "content": ", we analyze that the predicted TD " + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{p}}_m" + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "content": " converges to the actual TD " + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\bar{p}" + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "content": " at the data selection phase. We calculate " + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}(\\bar{\\boldsymbol{p}}^{(T)}||\\tilde{\\boldsymbol{p}}_m^{(t)})" + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "content": " and compare it with " + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}(\\bar{\\boldsymbol{p}}^{(T)}||\\boldsymbol{p}^{(t)})" + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "content": " which is set as a baseline computed with the actual TD " + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\bar{p}" + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "content": " and the predicted probabilities " + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{p}" + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "content": " (snapshot) of the target classifier " + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 579, + 545, + 713 + ], + "type": "text", + "content": ". In this analysis, we use the bal" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22341" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 180 + ], + "type": "text", + "content": "anced CIFAR10 where the sample-wise averaged KL divergence scores are computed on the test set. Figure 7 shows that the final predicted TD successfully approximates the actual TD, while the predicted probability is highly different from the actual TD. We conclude that the TD prediction module " + }, + { + "bbox": [ + 46, + 72, + 288, + 180 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 72, + 288, + 180 + ], + "type": "text", + "content": " can produce the TD efficiently, leading to performance improvement, and the predicted TD acts as a better approximation of the actual TD than the predicted probability of a model snapshot captured at each epoch." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 189, + 124, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 189, + 124, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 189, + 124, + 201 + ], + "type": "text", + "content": "5.5. Limitations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 208, + 288, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 288, + 388 + ], + "type": "text", + "content": "We found two potential limitations of our TiDAL derived from the fact that it relies on the outputs of the target classifier to compute the TD. First, TiDAL is designed only for classification tasks, and thus it cannot be applied to AL targeting other tasks, such as regression [10, 15]. Second, TiDAL is highly influenced by the performance of the target classifier, especially when the target classifier wrongly classifies the hard negative samples with a high confidence during model optimization. These samples can be treated as certain samples (i.e. will not be selected for annotation) because they have low estimated uncertainties from the predicted TD, even though the target classifier fails to predict the true label of the samples correctly. As a future work, we will study extending our TiDAL in the task-agnostic ways with a safeguard combating the wrongly classified samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 399, + 134, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 134, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 134, + 411 + ], + "type": "text", + "content": "6. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 419, + 146, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 146, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 146, + 432 + ], + "type": "text", + "content": "6.1. Active Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 715 + ], + "type": "text", + "content": "AL methods target to construct a dataset with the most useful samples based on the assumption that each sample has different importance in model training [40]. Two mainstream AL approaches exist for efficiently querying the unlabeled data: pool-based methods [31, 52, 45] use various ways to extract samples from an unlabeled data pool effectively, and synthesis-based methods [1, 58, 49] generate informative samples for the model. Pool-based methods can be roughly divided based on query strategies: uncertainty-based [13, 52, 45, 20] and diversity-based [42, 14, 38] methods, where some methods use the hybrid of both [4, 44, 25]. Uncertainty-based methods focus on finding which samples would be the most uncertain for the model, whereas diversity-based methods aim to construct a subset of representative samples of the input distribution. Our proposed method, TiDAL, lies in uncertainty-based methods. The significant difference between TiDAL and previous uncertainty-based methods is that TiDAL estimates data uncertainty using TD that contains additional hints generated during model training. In contrast, the previous methods leverage only static information (e.g., loss [52, 20] and predicted probabilities [13, 45, 25]) obtained by a model snapshot at the data selection phase." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 419, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 419, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 419, + 85 + ], + "type": "text", + "content": "6.2. Training Dynamics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 91, + 547, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 547 + ], + "type": "text", + "content": "TD focuses on how deep neural networks are optimized under back-propagation-based stepwise weight updates. Many studies try to understand how gradient descent can effectively obtain the global minimum by analyzing the loss landscape of neural networks [24, 32] or its loss trajectory [3]. Some also import alternative models that are more mathematically approachable to analyze, such as neural tangent kernels [21], deep Gaussian processes [30], or stochastic differential equations [54]. On the other hand, the phenomenological and practical viewpoint of TD also exists. [48] coin the term Forgetting Dynamics to assert that unforgettable samples are often less helpful, and [9] show that the model could prefer samples that are often wrongly predicted throughout model training. TD is also commonly used in noisy label literature to find potential noisy labels as they tend to fit later on model training [2, 39] or locate samples that can be relabeled correctly [46]. Furthermore, [57] calculate the Dynamic Instance Hardness score by monitoring losses of each sample or whether the prediction gets flipped so that higher scored samples can be prioritized for curriculum learning, and [23] feed the loss history to the auxiliary neural network to mediate the curriculum for training. [29] also introduce temporal ensembling for semi-supervised learning, where the model fits towards averaged probability outputs. [47, 37] devise Data Maps to inspect datasets with two TD measures; confidence and its variability across epochs on the true class prediction. [55] further extend the Data Maps for AL, whether the target classifier was consistently correct or not during training. The proposed method splits the labeled samples by applying a heuristic threshold on the level of consistency to train a binary classifier that is trained to discern uncertain samples. Even though the work, similar to ours, also utilizes TD, it relies on empirical observations and heuristic choices to separate the certain and uncertain samples. In this study, we link the concept of TD to AL with both empirical and theoretical results to estimate the uncertainty of unlabeled samples, which is often neglected in previous TD studies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 560, + 378, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 560, + 378, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 560, + 378, + 573 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "We propose a novel active learning method, Training Dynamics for Active Learning (TiDAL), by linking the concept of training dynamics to active learning. We provide motivating observations and theoretical evidence for using training dynamics to estimate the uncertainty of unlabeled data. Since tracking the training dynamics of large-scale unlabeled data is infeasible, TiDAL utilizes a training dynamics prediction module to efficiently predict the training dynamics of the unlabeled data. Based on the predicted training dynamics, TiDAL quantifies data uncertainty using the common uncertainty estimators: entropy and mar" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22342" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": "gin. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness of our method, surpassing the existing state-of-the-art active learning methods. We further analyze that using our training dynamics prediction module is effective and the module successfully predicts the TD of unlabeled data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 157, + 106, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 157, + 106, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 157, + 106, + 170 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 177, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 177, + 287, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 177, + 287, + 199 + ], + "spans": [ + { + "bbox": [ + 53, + 177, + 287, + 199 + ], + "type": "text", + "content": "[1] Dana Angluin. Queries and concept learning. Machine learning, 2(4):319-342, 1988. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 201, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 201, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 288, + 245 + ], + "type": "text", + "content": "[2] Eric Arazo, Diego Ortego, Paul Albert, Noel O'Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In International conference on machine learning, pages 312-321. PMLR, 2019. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 287, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 291 + ], + "type": "text", + "content": "[3] Sanjeev Arora, Nadav Cohen, Noah Golowich, and Wei Hu. A convergence analysis of gradient descent for deep linear neural networks. In International Conference on Learning Representations, 2018. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 293, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 287, + 336 + ], + "type": "text", + "content": "[4] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671, 2019.8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 338, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 287, + 381 + ], + "type": "text", + "content": "[5] Les E Atlas, David A Cohn, and Richard E Ladner. Training connectionist networks with queries and selective sampling. In Advances in neural information processing systems, pages 566-573. CiteSeer, 1990. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 384, + 287, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 287, + 438 + ], + "type": "text", + "content": "[6] William H Beluch, Tim Genewein, Andreas Nurnberger, and Jan M Kohler. The power of ensembles for active learning in image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9368-9377, 2018. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 441, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 287, + 495 + ], + "type": "text", + "content": "[7] Javad Zolfaghari Bengar, Joost van de Weijer, Laura Lopez Fuentes, and Bogdan Raducanu. Class-balanced active learning for image classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1536-1545, 2022. 15" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 287, + 540 + ], + "type": "text", + "content": "[8] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in Neural Information Processing Systems, 32:1567-1578, 2019. 3, 5, 15" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 544, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 287, + 586 + ], + "type": "text", + "content": "[9] Haw-Shiuan Chang, Erik Learned-Miller, and Andrew McCallum. Active bias: Training more accurate neural networks by emphasizing high variance samples. Advances in Neural Information Processing Systems, 30, 2017. 1, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 589, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 620 + ], + "type": "text", + "content": "[10] David Cohn, Les Atlas, and Richard Ladner. Improving generalization with active learning. Machine learning, 15(2):201-221, 1994. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "text", + "content": "[11] Charles Corbière, Nicolas Thome, Avner Bar-Hen, Matthieu Cord, and Patrick Pérez. Addressing failure prediction by learning model confidence. Advances in Neural Information Processing Systems, 32, 2019. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[13] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International Conference on Machine Learning, pages 1183-1192. PMLR, 2017. 1, 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 138 + ], + "type": "text", + "content": "[14] Daniel Gissin and Shai Shalev-Shwartz. Discriminative active learning. arXiv preprint arXiv:1907.06347, 2019. 1, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "type": "text", + "content": "[15] Jia Gong, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Meta agent teaming active learning for pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079-11089, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 195, + 545, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 236 + ], + "type": "text", + "content": "[16] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 15" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 270 + ], + "type": "text", + "content": "[17] Hangfeng He and Weijie Su. The local elasticity of neural networks. In International Conference on Learning Representations, 2019. 3, 12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 315 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 15" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 317, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 370 + ], + "type": "text", + "content": "[19] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6626-6636, 2021. 3, 5, 15" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 372, + 545, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 424 + ], + "type": "text", + "content": "[20] Siyu Huang, Tianyang Wang, Haoyi Xiong, Jun Huan, and Dejing Dou. Semi-supervised active learning with temporal output discrepancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3447-3456, 2021. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "text", + "content": "[21] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "text", + "content": "[22] Heinrich Jiang, Been Kim, Melody Guan, and Maya Gupta. To trust or not to trust a classifier. Advances in neural information processing systems, 31, 2018. 3, 12" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 505, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 557 + ], + "type": "text", + "content": "[23] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International Conference on Machine Learning, pages 2304–2313. PMLR, 2018. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 590 + ], + "type": "text", + "content": "[24] Kenji Kawaguchi. Deep learning without poor local minima. Advances in neural information processing systems, 29, 2016. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 592, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 645 + ], + "type": "text", + "content": "[25] Kwanyoung Kim, Dongwon Park, Kwang In Kim, and Se Young Chun. Task-aware variational adversarial active learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8166-8175, 2021. 4, 5, 8, 16" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 679 + ], + "type": "text", + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 15" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[27] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22343" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[28] Alex Krizhevsky, Vinod Nair, and Geofrey Hinton. The CIFar-10 dataset. http://www.cs.toronto.edu/kriz/cifar.html, 55(5), 2014. 15" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 117, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 117, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 117, + 287, + 150 + ], + "type": "text", + "content": "[29] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In International Conference on Learning Representations (ICLR). OpenReview.net, 2017. 1, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 288, + 194 + ], + "type": "text", + "content": "[30] Jaehoon Lee, Yasaman Bahri, Roman Novak, Samuel S Schoenholz, Jeffrey Pennington, and Jascha Sohl-Dickstein. Deep neural networks as gaussian processes. In International Conference on Learning Representations, 2018. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 239 + ], + "type": "text", + "content": "[31] David D Lewis and William A Gale. A sequential algorithm for training text classifiers. In Proceedings of the 17th annual international ACM SIGIR conference on Research and development in information retrieval, pages 3-12, 1994. 1, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 282 + ], + "type": "text", + "content": "[32] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 282, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 282, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 282, + 288, + 338 + ], + "type": "text", + "content": "[33] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 3, 5, 15" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 392 + ], + "type": "text", + "content": "[34] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y. Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 392, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 392, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 392, + 288, + 426 + ], + "type": "text", + "content": "[35] Curtis Northcutt, Lu Jiang, and Isaac Chuang. Confident learning: Estimating uncertainty in dataset labels. Journal of Artificial Intelligence Research, 70:1373-1411, 2021. 15" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 288, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 460 + ], + "type": "text", + "content": "[36] Nicolas Papernot and Patrick McDaniel. Deep k-nearest neighbors: Towards confident, interpretable and robust deep learning. arXiv preprint arXiv:1803.04765, 2018. 3, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 288, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 492 + ], + "type": "text", + "content": "[37] Seo Yeon Park and Cornelia Caragea. A data cartography based mixup for pre-trained language models. arXiv preprint arXiv:2205.03403, 2022. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 492, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 548 + ], + "type": "text", + "content": "[38] Amin Parvaneh, Ehsan Abbasnejad, Damien Teney, Gholam-reza Reza Haffari, Anton van den Hengel, and Javen Qinfeng Shi. Active learning by feature mixing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12237-12246, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "type": "text", + "content": "[39] Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. Identifying unlabeled data using the area under the margin ranking. Advances in Neural Information Processing Systems, 33:17044-17056, 2020. 2, 8, 15" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 635 + ], + "type": "text", + "content": "[40] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning. ACM Computing Surveys (CSUR), 54(9):1-40, 2021. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 669 + ], + "type": "text", + "content": "[41] Dan Roth and Kevin Small. Margin-based active learning for structured output spaces. In European Conference on Machine Learning, pages 413-424. Springer, 2006. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 712 + ], + "type": "text", + "content": "[42] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 1, 5, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[43] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 163 + ], + "type": "text", + "content": "[44] Changjian Shui, Fan Zhou, Christian Gagné, and Boyu Wang. Deep active learning: Unified and principled method for query and training. In International Conference on Artificial Intelligence and Statistics, pages 1308-1318. PMLR, 2020. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 209 + ], + "type": "text", + "content": "[45] Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell. Variational adversarial active learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5972-5981, 2019. 2, 3, 4, 5, 8, 16" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 209, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 255 + ], + "type": "text", + "content": "[46] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In International Conference on Machine Learning, pages 5907-5915. PMLR, 2019. 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 256, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 545, + 323 + ], + "type": "text", + "content": "[47] Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A Smith, and Yejin Choi. Dataset cartography: Mapping and diagnosing datasets with training dynamics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9275-9293, 2020. 1, 2, 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 324, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 324, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 324, + 545, + 380 + ], + "type": "text", + "content": "[48] Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J Gordon. An empirical study of example forgetting during deep neural network learning. In International Conference on Learning Representations, 2018. 1, 8, 15" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 381, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 381, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 381, + 545, + 425 + ], + "type": "text", + "content": "[49] Toan Tran, Thanh-Toan Do, Ian Reid, and Gustavo Carneiro. Bayesian generative active deep learning. In International Conference on Machine Learning, pages 6295-6304. PMLR, 2019. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 426, + 545, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 494 + ], + "type": "text", + "content": "[50] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 495, + 545, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 495, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 307, + 495, + 545, + 539 + ], + "type": "text", + "content": "[51] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 541, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 541, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 541, + 545, + 586 + ], + "type": "text", + "content": "[52] Donggeun Yoo and In So Kweon. Learning loss for active learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 93-102, 2019. 1, 2, 3, 4, 5, 8, 15, 16" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 586, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 586, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 586, + 545, + 632 + ], + "type": "text", + "content": "[53] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 64(3):107-115, 2021. 12" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 633, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 633, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 633, + 545, + 678 + ], + "type": "text", + "content": "[54] Jiayao Zhang, Hua Wang, and Weijie Su. Imitating deep learning dynamics via locally elastic stochastic differential equations. Advances in Neural Information Processing Systems, 34, 2021. 3, 8, 12" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 678, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 678, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 678, + 545, + 714 + ], + "type": "text", + "content": "[55] Mike Zhang and Barbara Plank. Cartography active learning. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 395-406, 2021. 5, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22344" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 196 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[56] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, 2020. 5, 15" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 171 + ], + "type": "text", + "content": "[57] Tianyi Zhou, Shengjie Wang, and Jeffrey Bilmes. Curriculum learning by dynamic instance hardness. Advances in Neural Information Processing Systems, 33:8602-8613, 2020. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 196 + ], + "type": "text", + "content": "[58] Jia-Jie Zhu and José Bento. Generative adversarial active learning. arXiv preprint arXiv:1702.07956, 2017. 8" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22345" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/0fbb2c88-7b44-409c-9798-461b2922052e_content_list.json b/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/0fbb2c88-7b44-409c-9798-461b2922052e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d39d4d96dc7e35074a74da5b70fe7b34b5fa6083 --- /dev/null +++ b/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/0fbb2c88-7b44-409c-9798-461b2922052e_content_list.json @@ -0,0 +1,2007 @@ +[ + { + "type": "text", + "text": "TiDy-PSFs: Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions", + "text_level": 1, + "bbox": [ + 148, + 128, + 821, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sachin Shah*", + "bbox": [ + 256, + 203, + 366, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Maryland, College Park", + "bbox": [ + 161, + 220, + 460, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shah2022@umd.edu", + "bbox": [ + 238, + 242, + 383, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sakshum Kulshrestha*", + "bbox": [ + 566, + 203, + 746, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Maryland, College Park", + "bbox": [ + 511, + 222, + 807, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sakshumk@umd.edu", + "bbox": [ + 586, + 242, + 730, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christopher A. Metzler \nUniversity of Maryland, College Park", + "text_level": 1, + "bbox": [ + 334, + 263, + 633, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "metzler@umd.edu", + "bbox": [ + 416, + 303, + 552, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 351, + 313, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point-spread-function (PSF) engineering is a powerful computational imaging technique wherein a custom phase mask is integrated into an optical system to encode additional information into captured images. Used in combination with deep learning, such systems now offer state-of-the-art performance at monocular depth estimation, extended depth-of-field imaging, lensless imaging, and other tasks. Inspired by recent advances in spatial light modulator (SLM) technology, this paper answers a natural question: Can one encode additional information and achieve superior performance by changing a phase mask dynamically over time? We first prove that the set of PSFs described by static phase masks is non-convex and that, as a result, time-averaged PSFs generated by dynamic phase masks are fundamentally more expressive. We then demonstrate, in simulation, that time-averaged dynamic (TiDy) phase masks can leverage this increased expressiveness to offer substantially improved monocular depth estimation and extended depth-of-field imaging performance.", + "bbox": [ + 75, + 383, + 473, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 699, + 209, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Extracting depth information from an image is a critical task across a range of applications including autonomous driving [29, 33], robotics [23, 34], microscopy [7, 19], and augmented reality [31, 14]. To this end, researchers have developed engineered phase masks and apertures which serve to encode depth information into an image [12, 25]. To optimize these phase masks, recent works have exploited deep learning: By simultaneously optimizing a phase mask and a reconstruction algorithm \"end-to-end learning\" is able to dramatically improve system performance [32, 26].", + "bbox": [ + 75, + 724, + 468, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6ef10940b8d16c587641143d9f8434ec8872535f1b4381367784e089ac36ac7.jpg", + "image_caption": [ + "Figure 1. Time-averaged Dynamic PSFs Top: Phase mask sequence that was optimized to perform simultaneous extended depth-of-field imaging and monocular depth estimation. Middle: Proposed TiDy PSFs at specific depths. Bottom left: Depth estimation and all-in-focus imaging performance improve as one averages over more phase masks. Bottom right: Depth-encoded image and reconstructed depth map." + ], + "image_footnote": [], + "bbox": [ + 514, + 354, + 880, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing works have focused on learning or optimizing a single phase mask for passive depth perception. We conjecture that this restriction leaves much room for improvement. Perhaps by using an SLM to introduce a sequence of phase masks over time, one could do much better.", + "bbox": [ + 496, + 664, + 890, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Supporting this idea is the fact, which we prove in Theorem 3, that the set of PSFs described by a single phase mask is non-convex. This implies that time-averaged PSFs, which span the convex hull of this set, can be significantly more expressive. In this work, we exploit the PSF nonconvexity by developing a multi-phase mask end-to-end optimization approach for learning a sequence of phase masks whose PSFs are averaged over time.", + "bbox": [ + 496, + 739, + 890, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work's central contributions are as follows:", + "bbox": [ + 517, + 859, + 836, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- We prove the set of PSFs generated by a single phase", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work", + "bbox": [ + 94, + 886, + 346, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10657", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mask is non-convex. Thus, dynamic phase-masks offer a fundamentally larger design space.", + "bbox": [ + 107, + 90, + 468, + 122 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We extend the end-to-end learning optics and algorithm design framework to design a dynamic set of phase masks.", + "- We demonstrate, in simulation, that time-averaged PSFs can achieve superior monocular depth estimation and extended depth-of-field imaging performance." + ], + "bbox": [ + 96, + 132, + 468, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 76, + 247, + 205, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Formation Model. One can simulate the formation of an image in a camera by discretizing an RGB image by depth, convolving each depth with it's the corresponding PSF, and compositing the outputs to form the signal on the sensor. This process can be represented by the equation", + "bbox": [ + 76, + 272, + 468, + 349 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nI = \\sum_ {d = 1} ^ {D} O _ {d} \\left(L * h _ {d}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 359, + 468, + 401 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $L$ represents all-in-focus image, $\\{1,\\dots ,D\\}$ represent a set of discrete depth layers, $O_{d}$ is the occlusion mask at depth $d$ , and the set $\\{h_1,\\dots ,h_D\\}$ represent the depth-dependent PSF, i.e., the cameras response to point sources at various depths [9]. Other works assume no depth discontinuities [26] or add additional computation to improve blurring at depth boundaries [10]. Our model is similar to those used in [32, 3].", + "bbox": [ + 76, + 412, + 468, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PSF Formation Model. A PSF $h_d$ can be formed as a function of distance $d$ and phase modulation $\\phi^M$ caused by height variation on a phase mask.", + "bbox": [ + 76, + 553, + 468, + 599 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nh _ {d} = \\left| \\mathcal {F} \\left[ A \\exp \\left(i \\phi^ {D F} (d) + i \\phi^ {M}\\right) \\right] \\right| ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 609, + 468, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\phi^{DF}(d)$ is the defocus aberration due to the distance $d$ between the focus point and the depth plane. Note that because this PSF depends on depth, it can be used to encode depth information into $I$ [8].", + "bbox": [ + 76, + 640, + 468, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key idea behind PSF-engineering and end-to-end learning is that one can use the aforementioned relationships to encode additional information into a captured image $I$ by selecting a particularly effective mask $\\phi^M$ .", + "bbox": [ + 76, + 715, + 468, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Related Work", + "text_level": 1, + "bbox": [ + 76, + 791, + 218, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Computational Optics for Depth Tasks", + "text_level": 1, + "bbox": [ + 76, + 816, + 408, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optics based approaches for depth estimation use sensors and optical setups to encode and recover depth information. Many methods have used the depth-dependent blur induced by the imaging system to estimate the depth", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of pixels in an image [20]. These approaches compare the blur at different ranges to the expected blur caused by an aperture focused at a fixed distance. Groups improved on this idea by implementing coded apertures, retaining more high frequency information about the scene to disambiguate depths [12]. Similar to depth estimation tasks, static phase masks have been used to produce tailored PSFs more invariant to depth, allowing for extended depth-of-field imaging [6]. However, these optically driven approaches with numerical analysis have been passed in performance by modern deep neural networks, allowing for joint optimization of optical elements and neural reconstruction networks.", + "bbox": [ + 496, + 90, + 890, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2. Deep Optics for Depth Tasks", + "text_level": 1, + "bbox": [ + 500, + 280, + 756, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Many methods have engineered phase masks with specific depth qualities. By maximizing Fisher information for depth, the coded image theoretically will have the most amount of depth cues as possible [24] and by minimizing Fisher information, one may achieve an extended depth-of-field image [6]. Deep learning techniques can be used to jointly train the optical parameters and neural network based estimation methods. The idea is that one can \"code\" an image to retain additional information about a scene, and then use a deep neural network to produce reconstructions. By using a differentiable model for light propagation, back-propagation can be used to update phase mask values simultaneously with neural network parameters. This approach was demonstrated for extended depth-of-field imaging [28, 26, 10, 13], depth estimation [32, 3, 10], high-dynamic-range imaging [16, 27], and holography [5, 4]. While these previous approaches successfully improved performance, they focused on enhancing a single phase mask. We build on these works by simultaneously optimizing multiple phase masks, which allows us to search over a larger space of PSFs.", + "bbox": [ + 496, + 303, + 890, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4. Theory", + "text_level": 1, + "bbox": [ + 500, + 633, + 586, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Micro-electromechanical system (MEMS) based SLMs offer high frameters but have limited phase precision due to heavy quantization [1]. As [4] noted, intensity averaging of multiple frames can improve quality by increasing effective precision to overcome quantization. Our key insight is that even as SLM technology improves, intensity averaging yields a more expressive design space than a single phase mask. This is supported by the claim that the set of PSFs that can be generated by a single phase mask is non-convex. We provide a rigorous proof for the claim as follows.", + "bbox": [ + 496, + 657, + 890, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Theorem 1. The set of PSFs that can be generated by a phase mask with an infinite aperture ( $A(x) = 1$ for all $x \\in \\mathbb{R}^2$ ) is non-convex.", + "bbox": [ + 496, + 816, + 890, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Proof. Consider two tilt shift phase masks, $M_1(x) = ax$ and $M_2(x) = -ax$ , with respect to a coordinate $x$ with", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10658", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "non-zero $a$ . The corresponding averaged PSF is $\\frac{1}{2}\\delta(a) + \\frac{1}{2}\\delta(-a)$ , where $\\delta$ denotes a Dirac delta function. To realize this PSF with a single phase mask we would need the field at the aperture to satisfy", + "bbox": [ + 76, + 90, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\mathcal {F} (E) \\right| ^ {2} = \\frac {1}{2} \\delta (a) + \\frac {1}{2} \\delta (- a). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 162, + 468, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This implies", + "bbox": [ + 76, + 203, + 163, + 218 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} E (x) = \\frac {1}{\\sqrt {2}} e ^ {- i a x} e ^ {i \\gamma_ {1}} + \\frac {1}{\\sqrt {2}} e ^ {i a x} e ^ {i \\gamma_ {2}}, (4) \\\\ = \\frac {2}{\\sqrt {2}} e ^ {i \\frac {\\gamma_ {1} + \\gamma_ {2}}{2}} \\cos \\left(a x + \\frac {\\gamma_ {1} - \\gamma_ {2}}{2}\\right) (5) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 228, + 468, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for some $\\gamma_1, \\gamma_2 \\in \\mathbb{R}$ . $E(x)$ 's amplitude varies according to a cosine and thus cannot be realized by a phase-only mask.", + "bbox": [ + 76, + 308, + 468, + 339 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5da6facc43e2d419a210f5266d8e6cd006331b7f7a07aeeb88273598254b276b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 339, + 468, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now demonstrate the non-convexity claim holds for finite apertures that lie on a discrete grid.", + "bbox": [ + 76, + 368, + 468, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1. $A \\in \\{0,1\\}^{N \\times N}$ is some valid aperture with a non-zero region $S$ such that there exists lines $L_{1}$ and $L_{2}$ where $S$ can be contained between them, and $L_{1} \\parallel L_{2}$ and $u = S \\cap L_{1}$ and $v = S \\cap L_{2}$ are single points (Figure 2).", + "bbox": [ + 76, + 409, + 468, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This definition of $A$ supports most commonly used apertures including but not limited to circles, squares, and $n$ -sided regular polygons. See supplement for proof for all shapes.", + "bbox": [ + 76, + 481, + 468, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2. Let $T_A(N)$ be the set of $N \\times N$ matrices in $\\mathbb{T}^{N \\times N}$ with non-zero support $A$ , i.e., the matrix is supported only where $A = 1$ , where $\\mathbb{T}$ is the complex unit circle.", + "bbox": [ + 76, + 554, + 468, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The PSF induced by a phase mask $M$ can be modeled as the squared magnitude of the Fourier transform of the pupil function $f$ [32].", + "bbox": [ + 76, + 611, + 468, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3. Let $f: \\mathbb{R}^{N \\times N} \\to T_A(N)$ be defined by", + "bbox": [ + 76, + 667, + 433, + 684 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf (M) = A \\odot \\exp (i D + i c M) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 696, + 468, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\odot$ denotes entry-wise multiplication, and $D \\in \\mathbb{R}^{N \\times N}$ and $c \\in \\mathbb{R} - \\{0\\}$ (the reals except for 0) are fixed constants.", + "bbox": [ + 76, + 726, + 468, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 4. Let $g: T_A(N) \\to \\mathbb{R}^{N \\times N}$ be defined by", + "bbox": [ + 76, + 781, + 431, + 799 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng (X) = \\frac {\\left| \\mathcal {F} (X) \\right| \\odot \\left| \\mathcal {F} (X) \\right|}{\\left\\| \\mathcal {F} (X) \\right\\| _ {F} ^ {2}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 809, + 468, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{F}$ denotes the discrete Fourier Transform with sufficient zero-padding, $|\\cdot|$ denotes entry-wise absolute value, and $\\| \\cdot \\|_F$ denotes the Frobenius norm.", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/59e333fd386162a54fcd57030d467ca16227cb5f33a9e1295209df2ae792316b.jpg", + "image_caption": [ + "Figure 2. Example aperture that satisfies constraints on A. The aperture is fitted between parallel lines $L1$ and $L2$ , which only intersect the aperture at one point each. Common aperture shapes fit into these constraints." + ], + "image_footnote": [], + "bbox": [ + 578, + 87, + 816, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lemma 2. From fourier optics theory [8], any single phase mask's PSF at a specific depth can be written as", + "bbox": [ + 498, + 369, + 890, + 398 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP S F = g \\circ f.\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 407, + 745, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 3. The range of PSF is not a convex set.", + "bbox": [ + 500, + 441, + 833, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof. $f$ is clearly surjective, so it suffices to argue the range of $g$ is not convex. Assume by way of contradiction that the range of $g$ is convex. Then, for all $X^{(1)}, \\ldots, X^{(k)} \\in T_A(N)$ there exists $Y \\in T_A(N)$ such that $g(Y) = \\frac{1}{k} \\sum_{i=1}^{k} g(X^{(i)})$ . By Parseval's Theorem,", + "bbox": [ + 498, + 464, + 890, + 544 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\| \\mathcal {F} (X) \\| _ {F} ^ {2} = N ^ {2} \\| X \\| _ {F} ^ {2} = N ^ {2} \\sum_ {i = 0} ^ {N} \\sum_ {j = 0} ^ {N} A _ {i, j} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 551, + 890, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "so the condition is", + "bbox": [ + 500, + 604, + 624, + 617 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n| \\mathcal {F} (Y) | \\odot | \\mathcal {F} (Y) | = \\frac {1}{k} \\sum_ {i = 1} ^ {k} | \\mathcal {F} \\left(X ^ {(i)}\\right) | \\odot | \\mathcal {F} \\left(X ^ {(i)}\\right) | \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 625, + 890, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "or equivalently", + "bbox": [ + 500, + 674, + 602, + 689 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} (Y) \\odot \\overline {{\\mathcal {F} (Y)}} = \\frac {1}{k} \\sum_ {i = 1} ^ {k} \\mathcal {F} \\left(X ^ {(i)}\\right) \\odot \\overline {{\\mathcal {F} \\left(X ^ {(i)}\\right)}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 696, + 890, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then the cross-correlation theorem reduces it to", + "bbox": [ + 500, + 747, + 816, + 760 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} (Y \\star Y) = \\frac {1}{k} \\sum_ {i = 1} ^ {k} \\mathcal {F} \\left(X ^ {(i)} \\star X ^ {(i)}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 768, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\star$ denotes cross-correlation. Because the Fourier Transform is linear we finally have", + "bbox": [ + 500, + 816, + 890, + 847 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nY \\star Y = \\frac {1}{k} \\sum_ {i = 1} ^ {k} X ^ {(i)} \\star X ^ {(i)}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 856, + 890, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10659", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7ea44b005c5551a31dc3c8080f2c86574e5447acf106b291a30d55c6fc17f078.jpg", + "image_caption": [ + "Figure 3. Geometric interpretation of correlation $(\\mathbf{X} \\star \\mathbf{X})_{\\mathbf{v} - \\mathbf{u}}$ . The figure represents the correlation step when the shift is $v - u$ . Notice that only $u$ and $v$ overlap once the shift is applied." + ], + "image_footnote": [], + "bbox": [ + 148, + 88, + 405, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, the convexity of the range of $g$ is equivalent to the convexity of the set $\\{X \\star X : X \\in T_A(N)\\}$ . We will show the set's projection onto a particular coordinate is not convex.", + "bbox": [ + 76, + 330, + 467, + 390 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(X \\star X) _ {s, r} = \\sum_ {i = 0} ^ {N} \\sum_ {j = 0} ^ {N} X _ {i, j} \\overline {{X _ {i + s , j + r}}} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 398, + 468, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we adopt the convention that $X_{s,r} = 0$ when $s, r > N$ or $s, r < 0$ . Take the points $u$ and $v$ from the definition of $A$ (1). Also observe that correlation can be represented geometrically as shifting $\\overline{X}$ over $X$ . In this representation, notice that as the shift $(s, r)$ approaches $v - u$ , the non-zero overlap between $X$ and $\\overline{X}$ shifted by $(s, r)$ approaches 1 by construction. That is, when $L_1$ is shifted to overlap $L_2$ , $u$ and $v$ will be the only non-zero overlaps between the shifted and original non-zero points (Figure 3). No other non-zero points can overlap above or below $L_2$ by definition of $S$ . Therefore, $(X \\star X)_{v-u}$ becomes", + "bbox": [ + 76, + 449, + 468, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {u} \\bar {X _ {v}} + \\sum_ {i = 1} ^ {N ^ {2} - 1} 0. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 625, + 468, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Because $X_{u}\\overline{X_{v}}\\in \\mathbb{T}$ , $(X\\star X)_{v - u}\\in \\mathbb{T}$ which is a nonconvex set. Therefore, the set of correlation's of values on the complex unit circle masked by $A$ is also not convex, and so is $PSF$ .", + "bbox": [ + 76, + 676, + 468, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Time-averaged PSFs span the convex hull of the set of static-mask PSFs, meaning there exists some PSFs achievable only through intensity averaging PSFs from a sequence of phase masks.", + "bbox": [ + 76, + 750, + 468, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Multi-Phase Mask Optimization", + "text_level": 1, + "bbox": [ + 76, + 821, + 370, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Optical Forward Model", + "text_level": 1, + "bbox": [ + 76, + 847, + 295, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to PhaseCam3D [32], we model light propagation using Fourier optics theory [8]. In contrast to previous", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "work, we compute the forward model (1) for multiple phase masks, producing a stack of output images, which form our coded image when averaged. This coded image simulates the recorded signal from imaging a scene using a sequence of phase masks in a single exposure (Figure 4).", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.2. Specialized Networks", + "text_level": 1, + "bbox": [ + 500, + 174, + 700, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the monocular depth estimation task, we use the MiDaS Small network [22]. This is a well known convolutional monocular depth estimation network designed to take in natural images and output relative depth maps. The network is trained end-to-end with the phase masks. A mean-squared error (MSE) loss term is defined in terms of the depth reconstruction prediction, $\\hat{D}$ and the ground truth depth map $D$ ,", + "bbox": [ + 498, + 198, + 890, + 319 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {D e p t h} = \\frac {1}{N} \\| D - \\hat {D} \\| _ {2} ^ {2} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 327, + 890, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ is the number of pixels. This process allows for the simultaneous optimization of the phase masks as well as fine tuning MiDaS to reconstruct from our coded images.", + "bbox": [ + 498, + 363, + 890, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the extended depth-of-field task, we use an Attention U-Net [18] to reconstruct all-in-focus images. The network is optimized jointly with the phase mask sequence. To learn a reconstruction $\\hat{I}$ to be similar to the all-in-focus ground truth image $I$ , we define the loss term using MSE error", + "bbox": [ + 498, + 409, + 890, + 483 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {A i F} = \\frac {1}{N} \\| I - \\hat {I} \\| _ {2} ^ {2} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 491, + 890, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ is the number of pixels.", + "bbox": [ + 500, + 527, + 718, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.3. Joint Task Optimization", + "text_level": 1, + "bbox": [ + 500, + 551, + 720, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also present an alternative to the specialized networks: a single network jointly trained for monocular depth estimation and extended depth-of-field using a sequence of phase masks. This network has a basic Attention U-Net architecture outputting 4 channels representing depth maps as well as all-in-focus images. Similar to prior works, we use a combined loss function, adding a coefficient to weight the losses for each individual task:", + "bbox": [ + 498, + 575, + 890, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = \\lambda_ {D e p t h} L _ {D e p t h} + \\lambda_ {A i F} L _ {A i F}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 705, + 890, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6. Experimental Details", + "text_level": 1, + "bbox": [ + 500, + 731, + 700, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6.1. Training Details", + "text_level": 1, + "bbox": [ + 500, + 756, + 660, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the FlyingThings3D from Scene Flow Datasets [15], which uses synthetic data generation to obtain all-in-focus RGB images and disparity maps. We use the cropped $278 \\times 278$ all-in-focus images from [32]. In total, we use 5077 training patches and 419 test patches.", + "bbox": [ + 498, + 779, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Both the optical layer and reconstruction networks are differentiable, so the phase mask sequence and neural network can be optimized through back-propagation. Each", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10660", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d822209196e8b2ba69d77c5c5ca65ccf34ceffe3a5a271695078321606299f63.jpg", + "image_caption": [ + "Figure 4. Multi-phase mask forward model overview. A sequence of phase masks are used to generate a sequence of depth-dependent PSFs. These PSFs are convolved with depth masked clean images to simulate depth dependent convolution. The images produced by each phase mask are averaged to create a coded image which is fed into an attention U-Net. The reconstruction loss is back-propagated end-to-end through the network and the optical model to design phase masks and algorithms capable of performing monocular depth estimation and extended depth-of-field simultaneously." + ], + "image_footnote": [], + "bbox": [ + 122, + 89, + 851, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "part is implemented in PyTorch. During training, we use the Adam [11] optimizer with parameters $\\beta_{1} = 0.99$ and $\\beta_{2} = 0.999$ . The learning rate for the phase masks is $10^{-8}$ and for the reconstruction network it is $10^{-4}$ , and the batch size was 32. Finally, training and testing were performed on NVIDIA Quadro P6000 GPUs.", + "bbox": [ + 75, + 319, + 470, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We parameterize $23 \\times 23$ phase masks pixel-wise as [13] found pixel-wise parameterization to produce the best overall performance. The monocular depth estimation task uses the MiDaS Small architecture pretrained weights for monocular depth estimation downloadable from PyTorch [22]. The extended depth-of-field task pretrains an Attention U-Net with a fixed Fresnel lens for 300 epochs. For the joint task, we set $\\lambda_{Depth} = \\lambda_{AiF} = 1$ to balance overall performance, and we pretrain the Attention U-Net for 300 epochs with a fixed Fresnel lens. In simulation, the red, blue, and green channels are approximated by discretized wavelengths, $610~\\mathrm{nm}$ , $530~\\mathrm{nm}$ , and $470~\\mathrm{nm}$ respectively. Additionally, the depth range is discretized into 21 bins on the interval $[-20, 20]$ , which is larger than previous works.", + "bbox": [ + 73, + 414, + 468, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.2. Evaluation Details", + "text_level": 1, + "bbox": [ + 75, + 647, + 253, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For ablation studies on our method, we used the testing split of the FlyingThings3D set for both monocular depth estimation and extended depth-of-field imaging [15]. For comparisons to existing work, we also tested our monocular depth estimation network on the labeled NYU Depth v2 set [17]. The ground truth depth maps were translated to layered masks for the clean images by bucketing the depth values into 21 bins, allowing us to convolve each depth in an image with the required PSF. We use root mean squared error (RMSE) between ground truth and estimated depth maps for depth estimation and RMSE between ground truth and reconstructed all-in-focus images for extended depth-of-field imaging. We also use peak signal-to-noise ratio (PSNR) and structural similarity index [30] (SSIM) for extended depth-of-field imaging.", + "bbox": [ + 75, + 674, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7. Results", + "text_level": 1, + "bbox": [ + 500, + 318, + 584, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our time averaged dynamic PSF method to the state-of-the-art methods for both extended depth-of-field imaging and monocular depth estimation. The relevant works we compare to are as follows:", + "bbox": [ + 498, + 343, + 890, + 405 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. PhaseCam3D [32] used a $23 \\times 23$ phase mask based on 55 Zernike coefficients. The phase mask parameters were then end-to-end optimized with a U-Net reconstruction network to perform depth estimation.", + "2. Chang et al. [3] used a singlet lens introducing chromatic aberrations with radially symmetric PSFs. Similar to [32], the lens parameters were also then end-to-end optimized.", + "3. Ikoma et al. [10] used a radially symmetric diffractive optical element (DOE). The blurred image was preconditioned with an approximate inverse of the PSF depth dependent blur. The RGB image stack was fed into a U-Net to produce both an all-in-focus image and a depth map. The DOE and U-Net parameters were optimized in an end-to-end fashion.", + "4. Liu et al. [13] used various phase mask parameterizations with the same U-Net architecture as [10]. One method used pixel-wise height maps (PW) and the other introduced orbital angular momentum (OAM).", + "5. Sitzmann et al. [26] implements a single DOE based on Zernike coefficients, and solves the Tikhonov-regularized least-squares problem to reconstruct an all-in-focus image.", + "6. MiDaS [21] and ZoeDepth [2] are state of the art single shot monocular depth estimation methods with all-in-focus images as inputs." + ], + "bbox": [ + 511, + 414, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Because both [10] and [13] simultaneously learn all-in-focus images and depth maps, when comparing against our", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10661", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/04b1c7ada2d7d33f401511f06deb7988e985c073d008eb4abd9a34fbc6c1d180.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFlyingThings3DNYUv2
PhaseCam3D [32]0.5210.382
Chang et al. [3]0.4900.433
Ikoma et al. [10]0.184-
MiDaS [21]-0.357
ZoeDepth [2]-0.277
TiDy (1)0.0260.259
TiDy (5)0.0190.175
", + "bbox": [ + 102, + 88, + 444, + 215 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e44d0b58183f698292444d785603f279ca67f86e75c55e688af7655c1eff69ff.jpg", + "table_caption": [ + "Table 1. RMSE comparison of monocular depth estimation methods. We present quantitative results on two datasets to compare to state of the art optical and single shot monocular depth estimation methods. Our method performs best with our 5 phase mask system achieving the lowest error on both datasets." + ], + "table_footnote": [], + "table_body": "
MethodRMSE↓PSNR↑SSIM↑
Liu et al. [13]-29.80-
Ikoma et al. [10]0.132731.880.905
Sitzmann et al. [26]-32.44-
TiDy (1)0.014837.330.968
TiDy (5)0.009241.110.989
", + "bbox": [ + 93, + 318, + 452, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "specialized methods, we take their best performing weighting of each task.", + "bbox": [ + 75, + 518, + 467, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Individual Tasks. For monocular depth estimation, our specialized method using a sequence of 5 phase masks trained for 300 epochs outperforms prior work on FlyingThings3D (Table 1). Additionally, our approach performs significantly better and achieves lower error than previous methods on NYUv2 without any additional fine tuning. For extended depth-of-field, our specialized method using a sequence of 5 phase masks outperforms prior work on FlyingThings3D (Table 2). This demonstrates the benefit of multi-phase mask learning on computational imaging tasks.", + "bbox": [ + 75, + 595, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-Objective Optimization. We also evaluate our method against other joint all-in-focus and depth map learning approaches. This problem is challenging because good depth cues to produce depth maps is antithetical to producing an all-in-focus image. Our combined 5 phase mask trained for 300 epochs approach outperforms prior jointly trained approaches (Table 3).", + "bbox": [ + 75, + 794, + 467, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a2b151d2cbf234f1609cda4a6ad48e24681c41ab9b0df7b1cefa8160b17580ae.jpg", + "table_caption": [ + "Table 2. Comparison of extended depth-of-field imaging methods. We present quantitative results on FlyingThings3D to compare to state-of-the-art. Our method performs best with our 5 phase mask system achieving the best PSNR." + ], + "table_footnote": [], + "table_body": "
MethodAll-in-focus PSNR↑Depth RMSE↓
Ikoma et al. [10]31.880.191
Liu et al. [13] - PW29.800.056
Liu et al. [13] - OAMt25.860.053
TiDy (1)31.200.052
TiDy (5)34.790.034
", + "bbox": [ + 526, + 88, + 867, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison of multi-objective optimization of extended depth-of-field imaging and depth estimation methods. We compare quantitative results on FlyingThings3D to the state-of-the-art. Our method performs best with our 5 phase mask system achieving the best balance between objectives.", + "bbox": [ + 498, + 210, + 890, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d3dec9c3c45e64e15eab15de2277d24bd6d465a930b012baf5d35fe175bbbacc.jpg", + "image_caption": [ + "Figure 5. RMSE for specialized tasks for each phase mask sequence length. RMSE decreases with respect to phase mask sequence length for both specialized extended depth-of-field imaging and monocular depth estimation tasks. 0 phase masks refers to a reconstruction neural network with a fixed Fresnel lens." + ], + "image_footnote": [], + "bbox": [ + 547, + 314, + 844, + 473 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f261afc545da41913c04fdd4882016163dff55fc2d92c59376d68c71eafe23b.jpg", + "image_caption": [ + "Figure 6. RMSE for joint optimization of monocular depth estimation and extended depth-of-field imaging for each phase mask sequence length. RMSE decreases with respect to phase mask sequence length for this complex joint task, demonstrating the benefit of multi-phase mask learning. 0 phase masks refers to a reconstruction neural network with a fixed Fresnel lens." + ], + "image_footnote": [], + "bbox": [ + 547, + 595, + 843, + 753 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10662", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4127ab4795135aec3fcf921155eeb96c0baf098ed5a0911b4629988756e632a9.jpg", + "image_caption": [ + "Figure 7. All-in-focus imaging RMSE distribution for each phase mask length without a reconstruction network. The best RMSE for each phase mask count has low correlation with respect to phase mask sequence length, but the variance of RMSE decreases." + ], + "image_footnote": [], + "bbox": [ + 107, + 95, + 444, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "8. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 364, + 243, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "8.1. Effect of Phase Mask Sequence Length", + "text_level": 1, + "bbox": [ + 76, + 388, + 413, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For both all-in-focus imaging and depth estimation, we vary the phase mask count that the end-to-end system is trained with to gauge the benefits of using multiple phase masks. The forward model and initial phase masks were held standard while the phase mask count was varied. The resulting networks were evaluated at convergence. For the extended depth-of-field task, the masks were all initialized with random noise uniform from 0 to $1.2 \\times 10^{-6}$ . For the depth estimation task, the masks were initialized with the Fisher mask with added Gaussian noise parameterized by a $5.35 \\times 10^{-7}$ mean and $3.05 \\times 10^{-7}$ standard deviation.", + "bbox": [ + 75, + 412, + 468, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "End-to-end optimization on each task with a specialized network yielded improved performance as the phase mask count increased, visualized in Figure 5. This result implies that sequences of phase masks are successful in making the PSF space more expressive. Additionally, even for the more complex joint task, learning a system that can produce both all-in-focus images and depth maps, error decreases with phase mask count until a plateau is reached (Figure 6).", + "bbox": [ + 75, + 580, + 468, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "8.2. All-in-Focus without Reconstruction Networks", + "text_level": 1, + "bbox": [ + 76, + 710, + 468, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A phase mask generating a PSF of the unit impulse function at every depth would be ideal for extended depth-of-field as each depth would be in focus. If possible, this phase mask would not require any digital processing. We optimize phase mask sequences of varying lengths to produce an averaged PSF close to the unit impulse function for all depths. For each sequence length, phase masks are optimized using MSE loss between the unit impulse function and the averaged PSF at each depth until convergence. We ran 1000 trials of random phase mask initialization for each length. Observe that a side-effect of longer phase masks is training", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49323f82872848c595d84c3a8a46b5a1c987e376054c3cedba130284ae30de34.jpg", + "image_caption": [ + "Figure 8. Visualization of phase mask initializations. Each row represents a different initial phase mask sequence." + ], + "image_footnote": [], + "bbox": [ + 550, + 93, + 844, + 320 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "stability. The range of RMSE between the simulated capture image and ground truth all-in-focus image decreases as the sequence length increases (Figure 7). This indicates training longer sequences is more resilient to initialization.", + "bbox": [ + 498, + 397, + 892, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "8.3. Phase Mask Initialization for Depth Perception", + "text_level": 1, + "bbox": [ + 500, + 467, + 890, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Deep optics for depth perception can be very dependent on the initialization of optical parameters before training [32]. To find the extent of the effect of mask initialization on performance, we varied the the initial phase masks while keeping number of masks, the optical model, and duration of training fixed. We trained for 200 epochs. We tested four initializations of sequences of 5 phase masks as shown in Figure 8. The first was uniformly distributed noise from 0 to $1.2 \\times 10^{-6}$ . The second was the first mask in the sequence set to a Fisher mask while the rest are uniform noise. The third is setting each mask to a rotation of the Fisher mask and adding Gaussian noise parameterized by a $5.35 \\times 10^{-7}$ mean and $3.05 \\times 10^{-7}$ standard deviation to 4 masks. Lastly, we set each mask to a rotation of the Fisher mask and added noise to only the last two masks in the sequence. Of the four initializations, it is clear that the 3 Fisher masks and 2 Fisher masks with noise performed the best (Table 4).", + "bbox": [ + 498, + 489, + 892, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "8.4. Modeling SLM Imperfections", + "text_level": 1, + "bbox": [ + 500, + 771, + 764, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "State Switching. Our optical forward model assumes an SLM can swap between two phase patterns instantly. In practice, however, some light will be captured during the intermediate states between phase patterns. These phase patterns, in the worst case, could be random phase patterns, effectively adding noise to our coded images. We model these intermediate states by averaging output images pro", + "bbox": [ + 498, + 794, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10663", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/89e839f6d49bc7b31543d56c5cb4bb700ece369b602fc33d41348329c4a0641d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
InitializationRMSE↓
1 Fisher + All noise0.0329
1 Fisher + Fisher w/ Noise0.0271
All noise0.0254
3 Fisher + Fisher w/ Noise0.0207
", + "bbox": [ + 138, + 88, + 408, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative evaluation of phase mask initializations. Four sequence initializations are evaluated on the monocular depth estimation task. Ultimately, 3 Fisher masks and 2 noisy Fisher masks have the best performance after training.", + "bbox": [ + 76, + 180, + 470, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04900e05ca70c5b7d5c4adf067137e62e360c1650799f3189c36cb2cdefa3be9.jpg", + "image_caption": [ + "Figure 9. Qualitative results of a specialized network on extended depth-of-field imaging. Both 1 and 5 phase mask systems are evaluated on FlyingThings3D. Error is computed pixel wise between the ground truth all-in-focus image and the reconstructed output and is boosted by a factor of 3. Notice that the 5 phase mask system introduces minimal error." + ], + "image_footnote": [], + "bbox": [ + 124, + 253, + 423, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "duced by phase masks and the randomized phase patterns weighted by the time that they are displayed for. We model the total exposure time as $100\\mathrm{ms}$ , with various durations of switching times from 1 to $16\\mathrm{ms}$ per swap. We evaluate our joint optimized network on these new, more noisy, coded images without any additional training (Figure 12). Observe that because the 5 phase mask system includes more swaps, performance degrades faster than fewer phase mask systems. However, for short switching times, 5 phase masks still outperform the others without needing any fine tuning.", + "bbox": [ + 75, + 704, + 470, + 856 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantization. Current high-speed MEMS based SLM technology experience heavy quantization [1]. We model", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fbd13953e34a1d415320526170cd157d1f065042d76dfb0a66e4f27a2e4f469d.jpg", + "image_caption": [ + "Figure 10. Qualitative results of a specialized networks on monocular depth estimation. Performance using the five phase mask method outperforms one phase mask on both datasets." + ], + "image_footnote": [], + "bbox": [ + 526, + 89, + 869, + 372 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/329bff1a5117f54e4ae67c656179621f630f238f55f3c15612a2bc8a63bfaf2f.jpg", + "image_caption": [ + "Figure 11. Qualitative results of a joint optimized system for extended depth-of-field imagining and monocular depth estimation. Both one and five phase mask networks are evaluated on the FlyingThings3D datasets. Notice that five masks has fewer artifacts than a single mask." + ], + "image_footnote": [], + "bbox": [ + 542, + 458, + 836, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the effects of quantization error by adding varying amounts of noise during inference. As demonstrated in Figure 13,", + "bbox": [ + 498, + 869, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10664", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73b4aa753f6b63d6c7513ad1788aa7affc98808f340aba7129a338178d155d19.jpg", + "image_caption": [ + "Figure 12. Effect of switching time on joint system performance. Reconstruction error across phase mask counts as a function of switching time with $100\\mathrm{ms}$ overall exposure. Performance of the jointly optimized system degrades as the switching time between phase masks increases, as expected. Our system still performs well when the time spent switching is less than $25\\%$ of the overall exposure." + ], + "image_footnote": [], + "bbox": [ + 104, + 90, + 444, + 239 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dde2d180287f3dcf84e733d47386b05c20afb33db3f39c2485d97dcc2c9c7242.jpg", + "image_caption": [ + "Figure 13. Effect of quantization on joint system performance. Performance of the jointly optimized system degrades as the number of effective bits decrease; however, even under heavy quantization, our system outperforms single mask setups." + ], + "image_footnote": [], + "bbox": [ + 102, + 375, + 444, + 503 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "time-averaged PSFs are robust to quantization noise: 5 masks quantized to 4-bits outperform one mask quantized to 32-bits. This is intuitive as multiple masks can leverage averaging to achieve better precision and remove noise [4].", + "bbox": [ + 75, + 604, + 467, + 666 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "8.5. A Path Towards Better Single-Mask Designs", + "text_level": 1, + "bbox": [ + 76, + 678, + 452, + 694 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The performance of end-to-end designed phase masks is highly sensitive to how they are initialized [32]. Can we use our 5-mask design to provide a better initialization for a 1-mask design?", + "bbox": [ + 76, + 703, + 467, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Starting from random initializations, we optimized 200 single phase masks to minimize the mean squared error between each of their PSFs and the time-averaged PSF formed by our optimized 5-mask system. We then took the single mask whose PSF was closest to the 5-mask design and fine-tuned our reconstruction network using its PSF. This static design improved performance over a randomly initialized end-to-end trained 1-mask system; +2.67 PSNR for all-in-focus imaging and -0.003 RMSE for depth estima", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "tion. However, its performance was still strictly worse than a 5-mask system.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "8.6. Modulating Amplitude and Phase", + "text_level": 1, + "bbox": [ + 500, + 130, + 795, + 145 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As illustrated in our proof of Theorem 1, \"complex\" masks which modulate both amplitude and phase are more expressive than phase-only masks. Using end-to-end learning, we optimized a pixel wise complex mask (initialized with a random phase mask and all ones amplitude mask) on the joint depth estimation and all-in-focus imaging task. The complex mask offered substationally improved performance; $+2.49$ PSNR for all-in-focus imaging and $-0.015$ RMSE for depth estimation. However, its performance too was still strictly worse than a 5-mask system. We conjecture this is due to improved training stability of the 5-mask system.", + "bbox": [ + 496, + 152, + 890, + 333 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "9. Limitations", + "text_level": 1, + "bbox": [ + 500, + 347, + 620, + 363 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While we were successful in learning dynamic phase masks to improve state-of-the-art performance on imaging tasks, our method carries some limitations. First, our optical model assumes perfect switching between phase masks during training. While evaluation with non-zero switching times showed little degradation of performance, accounting for state switching while training could produce phase masks that are more performant. Our optical model also simulates depths as layered masks over an image, which does not account for blending at depth boundaries. Additionally, our method assumes that scenes are static for the duration of a single exposure. Lastly, though their prices are falling, SLMs are still quite expensive and bulky.", + "bbox": [ + 496, + 372, + 890, + 569 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "10. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 575, + 627, + 590 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is founded upon the insight that the set of PSFs that can be described by a single phase mask is nonconvex and that, as a result, time-averaged PSFs are fundamentally more expressive. We demonstrate that one can learn a sequence of phase masks that, when one dynamically switches among them over time, can substantially improve computational imaging performance across a range of tasks, including depth estimation and all-in-focus imaging. Our work unlocks an exciting new direction for PSF engineering and computational imaging system design.", + "bbox": [ + 496, + 601, + 890, + 752 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 765, + 666, + 782 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by AFOSR Young Investigator Program Award FA9550-22-1-0208.", + "bbox": [ + 500, + 790, + 888, + 820 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 833, + 596, + 848 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "[1] Terry A. Bartlett, William C. McDonald, and James N. Hall. Adapting Texas Instruments DLP technology to demonstrate a phase spatial light modulator. In Michael R. Douglass, John", + "bbox": [ + 506, + 858, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10665", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ehmke, and Benjamin L. Lee, editors, Emerging Digital Micromirror Device Based Systems and Applications XI, volume 10932, page 109320S. International Society for Optics and Photonics, SPIE, 2019. 2, 8", + "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 5, 6", + "[3] Julie Chang and Gordon Wetzstein. Deep optics for monocular depth estimation and 3d object detection. In Proc. IEEE ICCV, 2019. 2, 5, 6", + "[4] Suyeon Choi, Manu Gopakumar, Yifan Peng, Jonghyun Kim, Matthew O'Toole, and Gordon Wetzstein. Time-multiplexed neural holography: A flexible framework for holographic near-eye displays with fast heavily-quantized spatial light modulators. In Proceedings of the ACM SIGGRAPH, page 1-9, 2022. 2, 9", + "[5] Suyeon Choi, Manu Gopakumar, Yifan Peng, Jonghyun Kim, and Gordon Wetzstein. Neural 3d holography: Learning accurate wave propagation models for 3d holographic virtual and augmented reality displays. ACM Trans. Graph., 40(6), December 2021. 2", + "[6] Edward R. Dowski and W. Thomas Cathey. Extended depth of field through wave-front coding. Appl. Opt., 34(11):1859-1866, April 1995. 2", + "[7] Robert Fischer, Yicong Wu, Pakorn Kanchanawong, Hari Shroff, and Clare Waterman-Storer. Microscopy in 3d: A biologist's toolbox. Trends in cell biology, 21:682-91, October 2011. 1", + "[8] Joseph W. Goodman. Introduction to fourier optics. Freeman, 2017. 2, 3, 4", + "[9] Samuel W. Hasinoff and Kiriakos N. Kutulakos. A layer-based restoration framework for variable-aperture photography. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8, 2007. 2", + "[10] Hayato Ikoma, Cindy M. Nguyen, Christopher A. Metzler, Yifan Peng, and Gordon Wetzstein. Depth from defocus with learned optics for imaging and occlusion-aware depth estimation. IEEE International Conference on Computational Photography (ICCP), 2021. 2, 5, 6", + "[11] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 5", + "[12] Anat Levin, Rob Fergus, Frédo Durand, and William T. Freeman. Image and depth from a conventional camera with a coded aperture. In ACM SIGGRAPH '2007 Papers, SIGGRAPH '07, page 70-es, New York, NY, USA, 2007. Association for Computing Machinery. 1, 2", + "[13] Xin Liu, Linpei Li, Xu Liu, Xiang Hao, and Yifan Peng. Investigating deep optics model representation in affecting resolved all-in-focus image quality and depth estimation fidelity. Opt. Express, 30(20):36973-36984, September 2022. 2, 5, 6", + "[14] Yawen Lu, Sophia Kourian, Carl Salvaggio, Chenliang Xu, and Guoyu Lu. Single image 3d vehicle pose estimation for augmented reality. In 2019 IEEE Global Conference on Signal and Information Processing (GlobalSIP), pages 1-5, 2019. 1", + "[15] Nikolaus Mayer, Eddy Ilg, Philip Häusser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016. 4, 5", + "[16] Christopher A Metzler, Hayato Ikoma, Yifan Peng, and Gordon Wetzstein. Deep optics for single-shot high-dynamic-range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1375-1385, 2020. 2", + "[17] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 5", + "[18] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert. Attention u-net: Learning where to look for the pancreas. In Medical Imaging with Deep Learning, 2018. 4", + "[19] Luca Palmieri, Gabriele Scrofani, Nicolò Incardona, Genaro Saavedra, Manuel Martínez-Corral, and Reinhard Koch. Robust depth estimation for light field microscopy. Sensors, 19(3), 2019. 1", + "[20] Alex Paul Pentland. A new sense for depth of field. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-9(4):523-531, 1987. 2", + "[21] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 5, 6", + "[22] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1623-1637, 2022. 4, 5", + "[23] Anupa Sabnis and Leena Vachhani. Single image based depth estimation for robotic applications. In 2011 IEEE Recent Advances in Intelligent Computational Systems, pages 102-106, 2011. 1", + "[24] Yoav Shechtman, Steffen J. Sahl, Adam S. Backer, and W. E. Moerner. Optimal point spread function design for 3d imaging. Phys. Rev. Lett., 113:133902, September 2014. 2", + "[25] Yoav Shechtman, Lucien Weiss, Adam Backer, Steffen Sahl, and William Moerner. Precise three-dimensional scan-free multiple-particle tracking over large axial ranges with tetrapod point spread functions. *Nano letters*, 15, May 2015. 1", + "[26] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Trans. Graph., 37(4), July 2018. 1, 2, 5, 6", + "[27] Qilin Sun, Ethan Tseng, Qiang Fu, Wolfgang Heidrich, and Felix Heide. Learning rank-1 diffractive optics for single-shot high dynamic range imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1386-1396, 2020. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10666", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Ashok Veeraraghavan, Ramesh Raskar, Amit Agrawal, Ankit Mohan, and Jack Tumblin. Dappled photography: Mask enhanced cameras for heterodyned light fields and coded aperture refocusing. ACM Trans. Graph., 26(3):69-es, jul 2007. 2", + "[29] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1", + "[30] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5", + "[31] Woontack Woo, Wonwoo Lee, and Nohyoung Park. Depth-assisted real-time 3d object detection for augmented reality. In International Conference on Artificial Reality and Teleexistence (ICAT), 2011. 1", + "[32] Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin Sankaranarayanan, and Ashok Veeraraghavan. Phasecam3d — learning phase masks for passive single view depth estimation. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2019. 1, 2, 3, 4, 5, 6, 7, 9", + "[33] Feng Xue, Guirong Zhuo, Ziyuan Huang, Wufei Fu, Zhuoyue Wu, and Marcelo H Ang. Toward hierarchical self-supervised monocular absolute depth estimation for autonomous driving applications. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2330-2337. IEEE, 2020. 1", + "[34] Menglong Ye, Edward Johns, Ankur Handa, Lin Zhang, Philip Pratt, and Guang-Zhong Yang. Self-supervised siamese learning on stereo image pairs for depth estimation in robotic surgery, 2017. 1" + ], + "bbox": [ + 78, + 90, + 468, + 583 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10667", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/0fbb2c88-7b44-409c-9798-461b2922052e_model.json b/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/0fbb2c88-7b44-409c-9798-461b2922052e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..013561e5740013e371e07352810247633015b1d5 --- /dev/null +++ b/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/0fbb2c88-7b44-409c-9798-461b2922052e_model.json @@ -0,0 +1,2499 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.13, + 0.823, + 0.176 + ], + "angle": 0, + "content": "TiDy-PSFs: Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.204, + 0.367, + 0.221 + ], + "angle": 0, + "content": "Sachin Shah*" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.222, + 0.462, + 0.24 + ], + "angle": 0, + "content": "University of Maryland, College Park" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.243, + 0.384, + 0.255 + ], + "angle": 0, + "content": "shah2022@umd.edu" + }, + { + "type": "text", + "bbox": [ + 0.567, + 0.204, + 0.748, + 0.221 + ], + "angle": 0, + "content": "Sakshum Kulshrestha*" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.223, + 0.808, + 0.24 + ], + "angle": 0, + "content": "University of Maryland, College Park" + }, + { + "type": "text", + "bbox": [ + 0.587, + 0.243, + 0.731, + 0.255 + ], + "angle": 0, + "content": "sakshumk@umd.edu" + }, + { + "type": "title", + "bbox": [ + 0.336, + 0.265, + 0.635, + 0.3 + ], + "angle": 0, + "content": "Christopher A. Metzler \nUniversity of Maryland, College Park" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.304, + 0.553, + 0.316 + ], + "angle": 0, + "content": "metzler@umd.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.352, + 0.314, + 0.368 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.474, + 0.673 + ], + "angle": 0, + "content": "Point-spread-function (PSF) engineering is a powerful computational imaging technique wherein a custom phase mask is integrated into an optical system to encode additional information into captured images. Used in combination with deep learning, such systems now offer state-of-the-art performance at monocular depth estimation, extended depth-of-field imaging, lensless imaging, and other tasks. Inspired by recent advances in spatial light modulator (SLM) technology, this paper answers a natural question: Can one encode additional information and achieve superior performance by changing a phase mask dynamically over time? We first prove that the set of PSFs described by static phase masks is non-convex and that, as a result, time-averaged PSFs generated by dynamic phase masks are fundamentally more expressive. We then demonstrate, in simulation, that time-averaged dynamic (TiDy) phase masks can leverage this increased expressiveness to offer substantially improved monocular depth estimation and extended depth-of-field imaging performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.7, + 0.21, + 0.716 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.47, + 0.878 + ], + "angle": 0, + "content": "Extracting depth information from an image is a critical task across a range of applications including autonomous driving [29, 33], robotics [23, 34], microscopy [7, 19], and augmented reality [31, 14]. To this end, researchers have developed engineered phase masks and apertures which serve to encode depth information into an image [12, 25]. To optimize these phase masks, recent works have exploited deep learning: By simultaneously optimizing a phase mask and a reconstruction algorithm \"end-to-end learning\" is able to dramatically improve system performance [32, 26]." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.355, + 0.882, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.549, + 0.895, + 0.648 + ], + "angle": 0, + "content": "Figure 1. Time-averaged Dynamic PSFs Top: Phase mask sequence that was optimized to perform simultaneous extended depth-of-field imaging and monocular depth estimation. Middle: Proposed TiDy PSFs at specific depths. Bottom left: Depth estimation and all-in-focus imaging performance improve as one averages over more phase masks. Bottom right: Depth-encoded image and reconstructed depth map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.74 + ], + "angle": 0, + "content": "Most existing works have focused on learning or optimizing a single phase mask for passive depth perception. We conjecture that this restriction leaves much room for improvement. Perhaps by using an SLM to introduce a sequence of phase masks over time, one could do much better." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.741, + 0.892, + 0.861 + ], + "angle": 0, + "content": "Supporting this idea is the fact, which we prove in Theorem 3, that the set of PSFs described by a single phase mask is non-convex. This implies that time-averaged PSFs, which span the convex hull of this set, can be significantly more expressive. In this work, we exploit the PSF nonconvexity by developing a multi-phase mask end-to-end optimization approach for learning a sequence of phase masks whose PSFs are averaged over time." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.861, + 0.838, + 0.875 + ], + "angle": 0, + "content": "This work's central contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "- We prove the set of PSFs generated by a single phase" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.348, + 0.9 + ], + "angle": 0, + "content": "*These authors contributed equally to this work" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10657" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "mask is non-convex. Thus, dynamic phase-masks offer a fundamentally larger design space." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.133, + 0.47, + 0.178 + ], + "angle": 0, + "content": "- We extend the end-to-end learning optics and algorithm design framework to design a dynamic set of phase masks." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.188, + 0.468, + 0.234 + ], + "angle": 0, + "content": "- We demonstrate, in simulation, that time-averaged PSFs can achieve superior monocular depth estimation and extended depth-of-field imaging performance." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.133, + 0.47, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.248, + 0.206, + 0.265 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.273, + 0.469, + 0.35 + ], + "angle": 0, + "content": "Image Formation Model. One can simulate the formation of an image in a camera by discretizing an RGB image by depth, convolving each depth with it's the corresponding PSF, and compositing the outputs to form the signal on the sensor. This process can be represented by the equation" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.361, + 0.469, + 0.402 + ], + "angle": 0, + "content": "\\[\nI = \\sum_ {d = 1} ^ {D} O _ {d} \\left(L * h _ {d}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.414, + 0.469, + 0.535 + ], + "angle": 0, + "content": "where \\(L\\) represents all-in-focus image, \\(\\{1,\\dots ,D\\}\\) represent a set of discrete depth layers, \\(O_{d}\\) is the occlusion mask at depth \\(d\\), and the set \\(\\{h_1,\\dots ,h_D\\}\\) represent the depth-dependent PSF, i.e., the cameras response to point sources at various depths [9]. Other works assume no depth discontinuities [26] or add additional computation to improve blurring at depth boundaries [10]. Our model is similar to those used in [32, 3]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.554, + 0.469, + 0.6 + ], + "angle": 0, + "content": "PSF Formation Model. A PSF \\( h_d \\) can be formed as a function of distance \\( d \\) and phase modulation \\( \\phi^M \\) caused by height variation on a phase mask." + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.61, + 0.469, + 0.629 + ], + "angle": 0, + "content": "\\[\nh _ {d} = \\left| \\mathcal {F} \\left[ A \\exp \\left(i \\phi^ {D F} (d) + i \\phi^ {M}\\right) \\right] \\right| ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.641, + 0.469, + 0.703 + ], + "angle": 0, + "content": "where \\(\\phi^{DF}(d)\\) is the defocus aberration due to the distance \\(d\\) between the focus point and the depth plane. Note that because this PSF depends on depth, it can be used to encode depth information into \\(I\\) [8]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.717, + 0.469, + 0.779 + ], + "angle": 0, + "content": "The key idea behind PSF-engineering and end-to-end learning is that one can use the aforementioned relationships to encode additional information into a captured image \\(I\\) by selecting a particularly effective mask \\(\\phi^M\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.792, + 0.219, + 0.807 + ], + "angle": 0, + "content": "3. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.817, + 0.409, + 0.833 + ], + "angle": 0, + "content": "3.1. Computational Optics for Depth Tasks" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Optics based approaches for depth estimation use sensors and optical setups to encode and recover depth information. Many methods have used the depth-dependent blur induced by the imaging system to estimate the depth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.273 + ], + "angle": 0, + "content": "of pixels in an image [20]. These approaches compare the blur at different ranges to the expected blur caused by an aperture focused at a fixed distance. Groups improved on this idea by implementing coded apertures, retaining more high frequency information about the scene to disambiguate depths [12]. Similar to depth estimation tasks, static phase masks have been used to produce tailored PSFs more invariant to depth, allowing for extended depth-of-field imaging [6]. However, these optically driven approaches with numerical analysis have been passed in performance by modern deep neural networks, allowing for joint optimization of optical elements and neural reconstruction networks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.281, + 0.757, + 0.297 + ], + "angle": 0, + "content": "3.2. Deep Optics for Depth Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.622 + ], + "angle": 0, + "content": "Many methods have engineered phase masks with specific depth qualities. By maximizing Fisher information for depth, the coded image theoretically will have the most amount of depth cues as possible [24] and by minimizing Fisher information, one may achieve an extended depth-of-field image [6]. Deep learning techniques can be used to jointly train the optical parameters and neural network based estimation methods. The idea is that one can \"code\" an image to retain additional information about a scene, and then use a deep neural network to produce reconstructions. By using a differentiable model for light propagation, back-propagation can be used to update phase mask values simultaneously with neural network parameters. This approach was demonstrated for extended depth-of-field imaging [28, 26, 10, 13], depth estimation [32, 3, 10], high-dynamic-range imaging [16, 27], and holography [5, 4]. While these previous approaches successfully improved performance, they focused on enhancing a single phase mask. We build on these works by simultaneously optimizing multiple phase masks, which allows us to search over a larger space of PSFs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.587, + 0.651 + ], + "angle": 0, + "content": "4. Theory" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Micro-electromechanical system (MEMS) based SLMs offer high frameters but have limited phase precision due to heavy quantization [1]. As [4] noted, intensity averaging of multiple frames can improve quality by increasing effective precision to overcome quantization. Our key insight is that even as SLM technology improves, intensity averaging yields a more expressive design space than a single phase mask. This is supported by the claim that the set of PSFs that can be generated by a single phase mask is non-convex. We provide a rigorous proof for the claim as follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.818, + 0.892, + 0.862 + ], + "angle": 0, + "content": "Theorem 1. The set of PSFs that can be generated by a phase mask with an infinite aperture (\\(A(x) = 1\\) for all \\(x \\in \\mathbb{R}^2\\)) is non-convex." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Proof. Consider two tilt shift phase masks, \\( M_1(x) = ax \\) and \\( M_2(x) = -ax \\), with respect to a coordinate \\( x \\) with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10658" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.152 + ], + "angle": 0, + "content": "non-zero \\( a \\). The corresponding averaged PSF is \\( \\frac{1}{2}\\delta(a) + \\frac{1}{2}\\delta(-a) \\), where \\( \\delta \\) denotes a Dirac delta function. To realize this PSF with a single phase mask we would need the field at the aperture to satisfy" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.164, + 0.469, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\left| \\mathcal {F} (E) \\right| ^ {2} = \\frac {1}{2} \\delta (a) + \\frac {1}{2} \\delta (- a). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.204, + 0.165, + 0.219 + ], + "angle": 0, + "content": "This implies" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.229, + 0.469, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} E (x) = \\frac {1}{\\sqrt {2}} e ^ {- i a x} e ^ {i \\gamma_ {1}} + \\frac {1}{\\sqrt {2}} e ^ {i a x} e ^ {i \\gamma_ {2}}, (4) \\\\ = \\frac {2}{\\sqrt {2}} e ^ {i \\frac {\\gamma_ {1} + \\gamma_ {2}}{2}} \\cos \\left(a x + \\frac {\\gamma_ {1} - \\gamma_ {2}}{2}\\right) (5) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.309, + 0.469, + 0.34 + ], + "angle": 0, + "content": "for some \\(\\gamma_1, \\gamma_2 \\in \\mathbb{R}\\). \\(E(x)\\)'s amplitude varies according to a cosine and thus cannot be realized by a phase-only mask." + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.34, + 0.469, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.369, + 0.469, + 0.4 + ], + "angle": 0, + "content": "We now demonstrate the non-convexity claim holds for finite apertures that lie on a discrete grid." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.41, + 0.469, + 0.472 + ], + "angle": 0, + "content": "Definition 1. \\(A \\in \\{0,1\\}^{N \\times N}\\) is some valid aperture with a non-zero region \\(S\\) such that there exists lines \\(L_{1}\\) and \\(L_{2}\\) where \\(S\\) can be contained between them, and \\(L_{1} \\parallel L_{2}\\) and \\(u = S \\cap L_{1}\\) and \\(v = S \\cap L_{2}\\) are single points (Figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.482, + 0.469, + 0.544 + ], + "angle": 0, + "content": "This definition of \\( A \\) supports most commonly used apertures including but not limited to circles, squares, and \\( n \\)-sided regular polygons. See supplement for proof for all shapes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.555, + 0.469, + 0.601 + ], + "angle": 0, + "content": "Definition 2. Let \\( T_A(N) \\) be the set of \\( N \\times N \\) matrices in \\( \\mathbb{T}^{N \\times N} \\) with non-zero support \\( A \\), i.e., the matrix is supported only where \\( A = 1 \\), where \\( \\mathbb{T} \\) is the complex unit circle." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.612, + 0.469, + 0.658 + ], + "angle": 0, + "content": "The PSF induced by a phase mask \\(M\\) can be modeled as the squared magnitude of the Fourier transform of the pupil function \\(f\\) [32]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.668, + 0.434, + 0.685 + ], + "angle": 0, + "content": "Definition 3. Let \\( f: \\mathbb{R}^{N \\times N} \\to T_A(N) \\) be defined by" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.697, + 0.469, + 0.715 + ], + "angle": 0, + "content": "\\[\nf (M) = A \\odot \\exp (i D + i c M) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.727, + 0.469, + 0.771 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes entry-wise multiplication, and \\(D \\in \\mathbb{R}^{N \\times N}\\) and \\(c \\in \\mathbb{R} - \\{0\\}\\) (the reals except for 0) are fixed constants." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.782, + 0.432, + 0.8 + ], + "angle": 0, + "content": "Definition 4. Let \\( g: T_A(N) \\to \\mathbb{R}^{N \\times N} \\) be defined by" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.81, + 0.469, + 0.845 + ], + "angle": 0, + "content": "\\[\ng (X) = \\frac {\\left| \\mathcal {F} (X) \\right| \\odot \\left| \\mathcal {F} (X) \\right|}{\\left\\| \\mathcal {F} (X) \\right\\| _ {F} ^ {2}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}\\) denotes the discrete Fourier Transform with sufficient zero-padding, \\(|\\cdot|\\) denotes entry-wise absolute value, and \\(\\| \\cdot \\|_F\\) denotes the Frobenius norm." + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.088, + 0.817, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.293, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Example aperture that satisfies constraints on A. The aperture is fitted between parallel lines \\( L1 \\) and \\( L2 \\), which only intersect the aperture at one point each. Common aperture shapes fit into these constraints." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.37, + 0.892, + 0.4 + ], + "angle": 0, + "content": "Lemma 2. From fourier optics theory [8], any single phase mask's PSF at a specific depth can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.646, + 0.409, + 0.746, + 0.425 + ], + "angle": 0, + "content": "\\[\nP S F = g \\circ f.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.443, + 0.834, + 0.458 + ], + "angle": 0, + "content": "Theorem 3. The range of PSF is not a convex set." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.465, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Proof. \\( f \\) is clearly surjective, so it suffices to argue the range of \\( g \\) is not convex. Assume by way of contradiction that the range of \\( g \\) is convex. Then, for all \\( X^{(1)}, \\ldots, X^{(k)} \\in T_A(N) \\) there exists \\( Y \\in T_A(N) \\) such that \\( g(Y) = \\frac{1}{k} \\sum_{i=1}^{k} g(X^{(i)}) \\). By Parseval's Theorem," + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.553, + 0.891, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\| \\mathcal {F} (X) \\| _ {F} ^ {2} = N ^ {2} \\| X \\| _ {F} ^ {2} = N ^ {2} \\sum_ {i = 0} ^ {N} \\sum_ {j = 0} ^ {N} A _ {i, j} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.605, + 0.625, + 0.618 + ], + "angle": 0, + "content": "so the condition is" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.625, + 0.891, + 0.666 + ], + "angle": 0, + "content": "\\[\n| \\mathcal {F} (Y) | \\odot | \\mathcal {F} (Y) | = \\frac {1}{k} \\sum_ {i = 1} ^ {k} | \\mathcal {F} \\left(X ^ {(i)}\\right) | \\odot | \\mathcal {F} \\left(X ^ {(i)}\\right) | \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.675, + 0.604, + 0.69 + ], + "angle": 0, + "content": "or equivalently" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.698, + 0.891, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} (Y) \\odot \\overline {{\\mathcal {F} (Y)}} = \\frac {1}{k} \\sum_ {i = 1} ^ {k} \\mathcal {F} \\left(X ^ {(i)}\\right) \\odot \\overline {{\\mathcal {F} \\left(X ^ {(i)}\\right)}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.748, + 0.818, + 0.761 + ], + "angle": 0, + "content": "Then the cross-correlation theorem reduces it to" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.769, + 0.891, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} (Y \\star Y) = \\frac {1}{k} \\sum_ {i = 1} ^ {k} \\mathcal {F} \\left(X ^ {(i)} \\star X ^ {(i)}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.818, + 0.892, + 0.848 + ], + "angle": 0, + "content": "where \\(\\star\\) denotes cross-correlation. Because the Fourier Transform is linear we finally have" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.857, + 0.891, + 0.898 + ], + "angle": 0, + "content": "\\[\nY \\star Y = \\frac {1}{k} \\sum_ {i = 1} ^ {k} X ^ {(i)} \\star X ^ {(i)}. \\tag {12}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10659" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.089, + 0.406, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.266, + 0.47, + 0.309 + ], + "angle": 0, + "content": "Figure 3. Geometric interpretation of correlation \\((\\mathbf{X} \\star \\mathbf{X})_{\\mathbf{v} - \\mathbf{u}}\\). The figure represents the correlation step when the shift is \\( v - u \\). Notice that only \\( u \\) and \\( v \\) overlap once the shift is applied." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.332, + 0.468, + 0.391 + ], + "angle": 0, + "content": "Therefore, the convexity of the range of \\( g \\) is equivalent to the convexity of the set \\( \\{X \\star X : X \\in T_A(N)\\} \\). We will show the set's projection onto a particular coordinate is not convex." + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.399, + 0.469, + 0.441 + ], + "angle": 0, + "content": "\\[\n(X \\star X) _ {s, r} = \\sum_ {i = 0} ^ {N} \\sum_ {j = 0} ^ {N} X _ {i, j} \\overline {{X _ {i + s , j + r}}} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.45, + 0.469, + 0.616 + ], + "angle": 0, + "content": "where we adopt the convention that \\( X_{s,r} = 0 \\) when \\( s, r > N \\) or \\( s, r < 0 \\). Take the points \\( u \\) and \\( v \\) from the definition of \\( A \\) (1). Also observe that correlation can be represented geometrically as shifting \\( \\overline{X} \\) over \\( X \\). In this representation, notice that as the shift \\( (s, r) \\) approaches \\( v - u \\), the non-zero overlap between \\( X \\) and \\( \\overline{X} \\) shifted by \\( (s, r) \\) approaches 1 by construction. That is, when \\( L_1 \\) is shifted to overlap \\( L_2 \\), \\( u \\) and \\( v \\) will be the only non-zero overlaps between the shifted and original non-zero points (Figure 3). No other non-zero points can overlap above or below \\( L_2 \\) by definition of \\( S \\). Therefore, \\( (X \\star X)_{v-u} \\) becomes" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.626, + 0.469, + 0.668 + ], + "angle": 0, + "content": "\\[\nX _ {u} \\bar {X _ {v}} + \\sum_ {i = 1} ^ {N ^ {2} - 1} 0. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.678, + 0.469, + 0.739 + ], + "angle": 0, + "content": "Because \\( X_{u}\\overline{X_{v}}\\in \\mathbb{T} \\), \\( (X\\star X)_{v - u}\\in \\mathbb{T} \\) which is a nonconvex set. Therefore, the set of correlation's of values on the complex unit circle masked by \\( A \\) is also not convex, and so is \\( PSF \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.75, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Time-averaged PSFs span the convex hull of the set of static-mask PSFs, meaning there exists some PSFs achievable only through intensity averaging PSFs from a sequence of phase masks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.822, + 0.371, + 0.839 + ], + "angle": 0, + "content": "5. Multi-Phase Mask Optimization" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.296, + 0.863 + ], + "angle": 0, + "content": "5.1. Optical Forward Model" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Similar to PhaseCam3D [32], we model light propagation using Fourier optics theory [8]. In contrast to previous" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "work, we compute the forward model (1) for multiple phase masks, producing a stack of output images, which form our coded image when averaged. This coded image simulates the recorded signal from imaging a scene using a sequence of phase masks in a single exposure (Figure 4)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.175, + 0.701, + 0.191 + ], + "angle": 0, + "content": "5.2. Specialized Networks" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.199, + 0.892, + 0.32 + ], + "angle": 0, + "content": "For the monocular depth estimation task, we use the MiDaS Small network [22]. This is a well known convolutional monocular depth estimation network designed to take in natural images and output relative depth maps. The network is trained end-to-end with the phase masks. A mean-squared error (MSE) loss term is defined in terms of the depth reconstruction prediction, \\(\\hat{D}\\) and the ground truth depth map \\(D\\)," + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.328, + 0.892, + 0.356 + ], + "angle": 0, + "content": "\\[\nL _ {D e p t h} = \\frac {1}{N} \\| D - \\hat {D} \\| _ {2} ^ {2} \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.364, + 0.892, + 0.409 + ], + "angle": 0, + "content": "where \\(N\\) is the number of pixels. This process allows for the simultaneous optimization of the phase masks as well as fine tuning MiDaS to reconstruct from our coded images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.484 + ], + "angle": 0, + "content": "For the extended depth-of-field task, we use an Attention U-Net [18] to reconstruct all-in-focus images. The network is optimized jointly with the phase mask sequence. To learn a reconstruction \\(\\hat{I}\\) to be similar to the all-in-focus ground truth image \\(I\\), we define the loss term using MSE error" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.492, + 0.891, + 0.521 + ], + "angle": 0, + "content": "\\[\nL _ {A i F} = \\frac {1}{N} \\| I - \\hat {I} \\| _ {2} ^ {2} \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.529, + 0.72, + 0.544 + ], + "angle": 0, + "content": "where \\(N\\) is the number of pixels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.552, + 0.722, + 0.568 + ], + "angle": 0, + "content": "5.3. Joint Task Optimization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.576, + 0.892, + 0.695 + ], + "angle": 0, + "content": "We also present an alternative to the specialized networks: a single network jointly trained for monocular depth estimation and extended depth-of-field using a sequence of phase masks. This network has a basic Attention U-Net architecture outputting 4 channels representing depth maps as well as all-in-focus images. Similar to prior works, we use a combined loss function, adding a coefficient to weight the losses for each individual task:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.706, + 0.891, + 0.722 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = \\lambda_ {D e p t h} L _ {D e p t h} + \\lambda_ {A i F} L _ {A i F}. \\tag {17}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.732, + 0.701, + 0.749 + ], + "angle": 0, + "content": "6. Experimental Details" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.661, + 0.773 + ], + "angle": 0, + "content": "6.1. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.892, + 0.856 + ], + "angle": 0, + "content": "We use the FlyingThings3D from Scene Flow Datasets [15], which uses synthetic data generation to obtain all-in-focus RGB images and disparity maps. We use the cropped \\(278 \\times 278\\) all-in-focus images from [32]. In total, we use 5077 training patches and 419 test patches." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Both the optical layer and reconstruction networks are differentiable, so the phase mask sequence and neural network can be optimized through back-propagation. Each" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10660" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.09, + 0.852, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.245, + 0.892, + 0.316 + ], + "angle": 0, + "content": "Figure 4. Multi-phase mask forward model overview. A sequence of phase masks are used to generate a sequence of depth-dependent PSFs. These PSFs are convolved with depth masked clean images to simulate depth dependent convolution. The images produced by each phase mask are averaged to create a coded image which is fed into an attention U-Net. The reconstruction loss is back-propagated end-to-end through the network and the optical model to design phase masks and algorithms capable of performing monocular depth estimation and extended depth-of-field simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.471, + 0.412 + ], + "angle": 0, + "content": "part is implemented in PyTorch. During training, we use the Adam [11] optimizer with parameters \\(\\beta_{1} = 0.99\\) and \\(\\beta_{2} = 0.999\\). The learning rate for the phase masks is \\(10^{-8}\\) and for the reconstruction network it is \\(10^{-4}\\), and the batch size was 32. Finally, training and testing were performed on NVIDIA Quadro P6000 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.415, + 0.47, + 0.627 + ], + "angle": 0, + "content": "We parameterize \\(23 \\times 23\\) phase masks pixel-wise as [13] found pixel-wise parameterization to produce the best overall performance. The monocular depth estimation task uses the MiDaS Small architecture pretrained weights for monocular depth estimation downloadable from PyTorch [22]. The extended depth-of-field task pretrains an Attention U-Net with a fixed Fresnel lens for 300 epochs. For the joint task, we set \\(\\lambda_{Depth} = \\lambda_{AiF} = 1\\) to balance overall performance, and we pretrain the Attention U-Net for 300 epochs with a fixed Fresnel lens. In simulation, the red, blue, and green channels are approximated by discretized wavelengths, \\(610~\\mathrm{nm}\\), \\(530~\\mathrm{nm}\\), and \\(470~\\mathrm{nm}\\) respectively. Additionally, the depth range is discretized into 21 bins on the interval \\([-20, 20]\\), which is larger than previous works." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.648, + 0.254, + 0.663 + ], + "angle": 0, + "content": "6.2. Evaluation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.903 + ], + "angle": 0, + "content": "For ablation studies on our method, we used the testing split of the FlyingThings3D set for both monocular depth estimation and extended depth-of-field imaging [15]. For comparisons to existing work, we also tested our monocular depth estimation network on the labeled NYU Depth v2 set [17]. The ground truth depth maps were translated to layered masks for the clean images by bucketing the depth values into 21 bins, allowing us to convolve each depth in an image with the required PSF. We use root mean squared error (RMSE) between ground truth and estimated depth maps for depth estimation and RMSE between ground truth and reconstructed all-in-focus images for extended depth-of-field imaging. We also use peak signal-to-noise ratio (PSNR) and structural similarity index [30] (SSIM) for extended depth-of-field imaging." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.319, + 0.586, + 0.334 + ], + "angle": 0, + "content": "7. Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.344, + 0.892, + 0.406 + ], + "angle": 0, + "content": "We compare our time averaged dynamic PSF method to the state-of-the-art methods for both extended depth-of-field imaging and monocular depth estimation. The relevant works we compare to are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.415, + 0.892, + 0.477 + ], + "angle": 0, + "content": "1. PhaseCam3D [32] used a \\(23 \\times 23\\) phase mask based on 55 Zernike coefficients. The phase mask parameters were then end-to-end optimized with a U-Net reconstruction network to perform depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.892, + 0.547 + ], + "angle": 0, + "content": "2. Chang et al. [3] used a singlet lens introducing chromatic aberrations with radially symmetric PSFs. Similar to [32], the lens parameters were also then end-to-end optimized." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.892, + 0.661 + ], + "angle": 0, + "content": "3. Ikoma et al. [10] used a radially symmetric diffractive optical element (DOE). The blurred image was preconditioned with an approximate inverse of the PSF depth dependent blur. The RGB image stack was fed into a U-Net to produce both an all-in-focus image and a depth map. The DOE and U-Net parameters were optimized in an end-to-end fashion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.892, + 0.734 + ], + "angle": 0, + "content": "4. Liu et al. [13] used various phase mask parameterizations with the same U-Net architecture as [10]. One method used pixel-wise height maps (PW) and the other introduced orbital angular momentum (OAM)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.892, + 0.805 + ], + "angle": 0, + "content": "5. Sitzmann et al. [26] implements a single DOE based on Zernike coefficients, and solves the Tikhonov-regularized least-squares problem to reconstruct an all-in-focus image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.815, + 0.892, + 0.86 + ], + "angle": 0, + "content": "6. MiDaS [21] and ZoeDepth [2] are state of the art single shot monocular depth estimation methods with all-in-focus images as inputs." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.415, + 0.892, + 0.86 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Because both [10] and [13] simultaneously learn all-in-focus images and depth maps, when comparing against our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10661" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.103, + 0.089, + 0.445, + 0.217 + ], + "angle": 0, + "content": "
MethodFlyingThings3DNYUv2
PhaseCam3D [32]0.5210.382
Chang et al. [3]0.4900.433
Ikoma et al. [10]0.184-
MiDaS [21]-0.357
ZoeDepth [2]-0.277
TiDy (1)0.0260.259
TiDy (5)0.0190.175
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.227, + 0.47, + 0.298 + ], + "angle": 0, + "content": "Table 1. RMSE comparison of monocular depth estimation methods. We present quantitative results on two datasets to compare to state of the art optical and single shot monocular depth estimation methods. Our method performs best with our 5 phase mask system achieving the lowest error on both datasets." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.319, + 0.454, + 0.418 + ], + "angle": 0, + "content": "
MethodRMSE↓PSNR↑SSIM↑
Liu et al. [13]-29.80-
Ikoma et al. [10]0.132731.880.905
Sitzmann et al. [26]-32.44-
TiDy (1)0.014837.330.968
TiDy (5)0.009241.110.989
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.428, + 0.47, + 0.484 + ], + "angle": 0, + "content": "Table 2. Comparison of extended depth-of-field imaging methods. We present quantitative results on FlyingThings3D to compare to state-of-the-art. Our method performs best with our 5 phase mask system achieving the best PSNR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.468, + 0.551 + ], + "angle": 0, + "content": "specialized methods, we take their best performing weighting of each task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Individual Tasks. For monocular depth estimation, our specialized method using a sequence of 5 phase masks trained for 300 epochs outperforms prior work on FlyingThings3D (Table 1). Additionally, our approach performs significantly better and achieves lower error than previous methods on NYUv2 without any additional fine tuning. For extended depth-of-field, our specialized method using a sequence of 5 phase masks outperforms prior work on FlyingThings3D (Table 2). This demonstrates the benefit of multi-phase mask learning on computational imaging tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.903 + ], + "angle": 0, + "content": "Multi-Objective Optimization. We also evaluate our method against other joint all-in-focus and depth map learning approaches. This problem is challenging because good depth cues to produce depth maps is antithetical to producing an all-in-focus image. Our combined 5 phase mask trained for 300 epochs approach outperforms prior jointly trained approaches (Table 3)." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.089, + 0.868, + 0.201 + ], + "angle": 0, + "content": "
MethodAll-in-focus PSNR↑Depth RMSE↓
Ikoma et al. [10]31.880.191
Liu et al. [13] - PW29.800.056
Liu et al. [13] - OAMt25.860.053
TiDy (1)31.200.052
TiDy (5)34.790.034
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.212, + 0.892, + 0.282 + ], + "angle": 0, + "content": "Table 3. Comparison of multi-objective optimization of extended depth-of-field imaging and depth estimation methods. We compare quantitative results on FlyingThings3D to the state-of-the-art. Our method performs best with our 5 phase mask system achieving the best balance between objectives." + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.315, + 0.845, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.496, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Figure 5. RMSE for specialized tasks for each phase mask sequence length. RMSE decreases with respect to phase mask sequence length for both specialized extended depth-of-field imaging and monocular depth estimation tasks. 0 phase masks refers to a reconstruction neural network with a fixed Fresnel lens." + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.597, + 0.844, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.778, + 0.892, + 0.861 + ], + "angle": 0, + "content": "Figure 6. RMSE for joint optimization of monocular depth estimation and extended depth-of-field imaging for each phase mask sequence length. RMSE decreases with respect to phase mask sequence length for this complex joint task, demonstrating the benefit of multi-phase mask learning. 0 phase masks refers to a reconstruction neural network with a fixed Fresnel lens." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10662" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.096, + 0.446, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.343 + ], + "angle": 0, + "content": "Figure 7. All-in-focus imaging RMSE distribution for each phase mask length without a reconstruction network. The best RMSE for each phase mask count has low correlation with respect to phase mask sequence length, but the variance of RMSE decreases." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.365, + 0.245, + 0.382 + ], + "angle": 0, + "content": "8. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.39, + 0.414, + 0.408 + ], + "angle": 0, + "content": "8.1. Effect of Phase Mask Sequence Length" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.47, + 0.579 + ], + "angle": 0, + "content": "For both all-in-focus imaging and depth estimation, we vary the phase mask count that the end-to-end system is trained with to gauge the benefits of using multiple phase masks. The forward model and initial phase masks were held standard while the phase mask count was varied. The resulting networks were evaluated at convergence. For the extended depth-of-field task, the masks were all initialized with random noise uniform from 0 to \\(1.2 \\times 10^{-6}\\). For the depth estimation task, the masks were initialized with the Fisher mask with added Gaussian noise parameterized by a \\(5.35 \\times 10^{-7}\\) mean and \\(3.05 \\times 10^{-7}\\) standard deviation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.703 + ], + "angle": 0, + "content": "End-to-end optimization on each task with a specialized network yielded improved performance as the phase mask count increased, visualized in Figure 5. This result implies that sequences of phase masks are successful in making the PSF space more expressive. Additionally, even for the more complex joint task, learning a system that can produce both all-in-focus images and depth maps, error decreases with phase mask count until a plateau is reached (Figure 6)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.47, + 0.728 + ], + "angle": 0, + "content": "8.2. All-in-Focus without Reconstruction Networks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "A phase mask generating a PSF of the unit impulse function at every depth would be ideal for extended depth-of-field as each depth would be in focus. If possible, this phase mask would not require any digital processing. We optimize phase mask sequences of varying lengths to produce an averaged PSF close to the unit impulse function for all depths. For each sequence length, phase masks are optimized using MSE loss between the unit impulse function and the averaged PSF at each depth until convergence. We ran 1000 trials of random phase mask initialization for each length. Observe that a side-effect of longer phase masks is training" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.094, + 0.845, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.345, + 0.892, + 0.374 + ], + "angle": 0, + "content": "Figure 8. Visualization of phase mask initializations. Each row represents a different initial phase mask sequence." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.398, + 0.893, + 0.459 + ], + "angle": 0, + "content": "stability. The range of RMSE between the simulated capture image and ground truth all-in-focus image decreases as the sequence length increases (Figure 7). This indicates training longer sequences is more resilient to initialization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.892, + 0.485 + ], + "angle": 0, + "content": "8.3. Phase Mask Initialization for Depth Perception" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.491, + 0.893, + 0.763 + ], + "angle": 0, + "content": "Deep optics for depth perception can be very dependent on the initialization of optical parameters before training [32]. To find the extent of the effect of mask initialization on performance, we varied the the initial phase masks while keeping number of masks, the optical model, and duration of training fixed. We trained for 200 epochs. We tested four initializations of sequences of 5 phase masks as shown in Figure 8. The first was uniformly distributed noise from 0 to \\(1.2 \\times 10^{-6}\\). The second was the first mask in the sequence set to a Fisher mask while the rest are uniform noise. The third is setting each mask to a rotation of the Fisher mask and adding Gaussian noise parameterized by a \\(5.35 \\times 10^{-7}\\) mean and \\(3.05 \\times 10^{-7}\\) standard deviation to 4 masks. Lastly, we set each mask to a rotation of the Fisher mask and added noise to only the last two masks in the sequence. Of the four initializations, it is clear that the 3 Fisher masks and 2 Fisher masks with noise performed the best (Table 4)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.766, + 0.789 + ], + "angle": 0, + "content": "8.4. Modeling SLM Imperfections" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "State Switching. Our optical forward model assumes an SLM can swap between two phase patterns instantly. In practice, however, some light will be captured during the intermediate states between phase patterns. These phase patterns, in the worst case, could be random phase patterns, effectively adding noise to our coded images. We model these intermediate states by averaging output images pro" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10663" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.14, + 0.089, + 0.409, + 0.171 + ], + "angle": 0, + "content": "
InitializationRMSE↓
1 Fisher + All noise0.0329
1 Fisher + Fisher w/ Noise0.0271
All noise0.0254
3 Fisher + Fisher w/ Noise0.0207
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.181, + 0.471, + 0.237 + ], + "angle": 0, + "content": "Table 4. Quantitative evaluation of phase mask initializations. Four sequence initializations are evaluated on the monocular depth estimation task. Ultimately, 3 Fisher masks and 2 noisy Fisher masks have the best performance after training." + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.255, + 0.424, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.679 + ], + "angle": 0, + "content": "Figure 9. Qualitative results of a specialized network on extended depth-of-field imaging. Both 1 and 5 phase mask systems are evaluated on FlyingThings3D. Error is computed pixel wise between the ground truth all-in-focus image and the reconstructed output and is boosted by a factor of 3. Notice that the 5 phase mask system introduces minimal error." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.857 + ], + "angle": 0, + "content": "duced by phase masks and the randomized phase patterns weighted by the time that they are displayed for. We model the total exposure time as \\(100\\mathrm{ms}\\), with various durations of switching times from 1 to \\(16\\mathrm{ms}\\) per swap. We evaluate our joint optimized network on these new, more noisy, coded images without any additional training (Figure 12). Observe that because the 5 phase mask system includes more swaps, performance degrades faster than fewer phase mask systems. However, for short switching times, 5 phase masks still outperform the others without needing any fine tuning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Quantization. Current high-speed MEMS based SLM technology experience heavy quantization [1]. We model" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.09, + 0.87, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.394, + 0.892, + 0.437 + ], + "angle": 0, + "content": "Figure 10. Qualitative results of a specialized networks on monocular depth estimation. Performance using the five phase mask method outperforms one phase mask on both datasets." + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.459, + 0.837, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.77, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Figure 11. Qualitative results of a joint optimized system for extended depth-of-field imagining and monocular depth estimation. Both one and five phase mask networks are evaluated on the FlyingThings3D datasets. Notice that five masks has fewer artifacts than a single mask." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "the effects of quantization error by adding varying amounts of noise during inference. As demonstrated in Figure 13," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10664" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.105, + 0.092, + 0.445, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.26, + 0.47, + 0.357 + ], + "angle": 0, + "content": "Figure 12. Effect of switching time on joint system performance. Reconstruction error across phase mask counts as a function of switching time with \\(100\\mathrm{ms}\\) overall exposure. Performance of the jointly optimized system degrades as the switching time between phase masks increases, as expected. Our system still performs well when the time spent switching is less than \\(25\\%\\) of the overall exposure." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.375, + 0.446, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.522, + 0.468, + 0.579 + ], + "angle": 0, + "content": "Figure 13. Effect of quantization on joint system performance. Performance of the jointly optimized system degrades as the number of effective bits decrease; however, even under heavy quantization, our system outperforms single mask setups." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.468, + 0.667 + ], + "angle": 0, + "content": "time-averaged PSFs are robust to quantization noise: 5 masks quantized to 4-bits outperform one mask quantized to 32-bits. This is intuitive as multiple masks can leverage averaging to achieve better precision and remove noise [4]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.453, + 0.695 + ], + "angle": 0, + "content": "8.5. A Path Towards Better Single-Mask Designs" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.704, + 0.468, + 0.763 + ], + "angle": 0, + "content": "The performance of end-to-end designed phase masks is highly sensitive to how they are initialized [32]. Can we use our 5-mask design to provide a better initialization for a 1-mask design?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Starting from random initializations, we optimized 200 single phase masks to minimize the mean squared error between each of their PSFs and the time-averaged PSF formed by our optimized 5-mask system. We then took the single mask whose PSF was closest to the 5-mask design and fine-tuned our reconstruction network using its PSF. This static design improved performance over a randomly initialized end-to-end trained 1-mask system; +2.67 PSNR for all-in-focus imaging and -0.003 RMSE for depth estima" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "tion. However, its performance was still strictly worse than a 5-mask system." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.131, + 0.796, + 0.146 + ], + "angle": 0, + "content": "8.6. Modulating Amplitude and Phase" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.154, + 0.892, + 0.334 + ], + "angle": 0, + "content": "As illustrated in our proof of Theorem 1, \"complex\" masks which modulate both amplitude and phase are more expressive than phase-only masks. Using end-to-end learning, we optimized a pixel wise complex mask (initialized with a random phase mask and all ones amplitude mask) on the joint depth estimation and all-in-focus imaging task. The complex mask offered substationally improved performance; \\(+2.49\\) PSNR for all-in-focus imaging and \\(-0.015\\) RMSE for depth estimation. However, its performance too was still strictly worse than a 5-mask system. We conjecture this is due to improved training stability of the 5-mask system." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.348, + 0.621, + 0.364 + ], + "angle": 0, + "content": "9. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.57 + ], + "angle": 0, + "content": "While we were successful in learning dynamic phase masks to improve state-of-the-art performance on imaging tasks, our method carries some limitations. First, our optical model assumes perfect switching between phase masks during training. While evaluation with non-zero switching times showed little degradation of performance, accounting for state switching while training could produce phase masks that are more performant. Our optical model also simulates depths as layered masks over an image, which does not account for blending at depth boundaries. Additionally, our method assumes that scenes are static for the duration of a single exposure. Lastly, though their prices are falling, SLMs are still quite expensive and bulky." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.576, + 0.629, + 0.592 + ], + "angle": 0, + "content": "10. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.892, + 0.753 + ], + "angle": 0, + "content": "This work is founded upon the insight that the set of PSFs that can be described by a single phase mask is nonconvex and that, as a result, time-averaged PSFs are fundamentally more expressive. We demonstrate that one can learn a sequence of phase masks that, when one dynamically switches among them over time, can substantially improve computational imaging performance across a range of tasks, including depth estimation and all-in-focus imaging. Our work unlocks an exciting new direction for PSF engineering and computational imaging system design." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.766, + 0.667, + 0.783 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.791, + 0.89, + 0.821 + ], + "angle": 0, + "content": "This work was partially supported by AFOSR Young Investigator Program Award FA9550-22-1-0208." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.834, + 0.597, + 0.849 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[1] Terry A. Bartlett, William C. McDonald, and James N. Hall. Adapting Texas Instruments DLP technology to demonstrate a phase spatial light modulator. In Michael R. Douglass, John" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10665" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Ehmke, and Benjamin L. Lee, editors, Emerging Digital Micromirror Device Based Systems and Applications XI, volume 10932, page 109320S. International Society for Optics and Photonics, SPIE, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.472, + 0.19 + ], + "angle": 0, + "content": "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.191, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[3] Julie Chang and Gordon Wetzstein. Deep optics for monocular depth estimation and 3d object detection. In Proc. IEEE ICCV, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.233, + 0.472, + 0.315 + ], + "angle": 0, + "content": "[4] Suyeon Choi, Manu Gopakumar, Yifan Peng, Jonghyun Kim, Matthew O'Toole, and Gordon Wetzstein. Time-multiplexed neural holography: A flexible framework for holographic near-eye displays with fast heavily-quantized spatial light modulators. In Proceedings of the ACM SIGGRAPH, page 1-9, 2022. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.472, + 0.384 + ], + "angle": 0, + "content": "[5] Suyeon Choi, Manu Gopakumar, Yifan Peng, Jonghyun Kim, and Gordon Wetzstein. Neural 3d holography: Learning accurate wave propagation models for 3d holographic virtual and augmented reality displays. ACM Trans. Graph., 40(6), December 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.472, + 0.426 + ], + "angle": 0, + "content": "[6] Edward R. Dowski and W. Thomas Cathey. Extended depth of field through wave-front coding. Appl. Opt., 34(11):1859-1866, April 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.427, + 0.472, + 0.482 + ], + "angle": 0, + "content": "[7] Robert Fischer, Yicong Wu, Pakorn Kanchanawong, Hari Shroff, and Clare Waterman-Storer. Microscopy in 3d: A biologist's toolbox. Trends in cell biology, 21:682-91, October 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.484, + 0.472, + 0.51 + ], + "angle": 0, + "content": "[8] Joseph W. Goodman. Introduction to fourier optics. Freeman, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.472, + 0.567 + ], + "angle": 0, + "content": "[9] Samuel W. Hasinoff and Kiriakos N. Kutulakos. A layer-based restoration framework for variable-aperture photography. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.568, + 0.472, + 0.636 + ], + "angle": 0, + "content": "[10] Hayato Ikoma, Cindy M. Nguyen, Christopher A. Metzler, Yifan Peng, and Gordon Wetzstein. Depth from defocus with learned optics for imaging and occlusion-aware depth estimation. IEEE International Conference on Computational Photography (ICCP), 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.637, + 0.472, + 0.664 + ], + "angle": 0, + "content": "[11] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.665, + 0.472, + 0.734 + ], + "angle": 0, + "content": "[12] Anat Levin, Rob Fergus, Frédo Durand, and William T. Freeman. Image and depth from a conventional camera with a coded aperture. In ACM SIGGRAPH '2007 Papers, SIGGRAPH '07, page 70-es, New York, NY, USA, 2007. Association for Computing Machinery. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.735, + 0.472, + 0.802 + ], + "angle": 0, + "content": "[13] Xin Liu, Linpei Li, Xu Liu, Xiang Hao, and Yifan Peng. Investigating deep optics model representation in affecting resolved all-in-focus image quality and depth estimation fidelity. Opt. Express, 30(20):36973-36984, September 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.804, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[14] Yawen Lu, Sophia Kourian, Carl Salvaggio, Chenliang Xu, and Guoyu Lu. Single image 3d vehicle pose estimation for augmented reality. In 2019 IEEE Global Conference on Signal and Information Processing (GlobalSIP), pages 1-5, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[15] Nikolaus Mayer, Eddy Ilg, Philip Häusser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.149 + ], + "angle": 0, + "content": "large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[16] Christopher A Metzler, Hayato Ikoma, Yifan Peng, and Gordon Wetzstein. Deep optics for single-shot high-dynamic-range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1375-1385, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[17] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[18] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert. Attention u-net: Learning where to look for the pancreas. In Medical Imaging with Deep Learning, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[19] Luca Palmieri, Gabriele Scrofani, Nicolò Incardona, Genaro Saavedra, Manuel Martínez-Corral, and Reinhard Koch. Robust depth estimation for light field microscopy. Sensors, 19(3), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[20] Alex Paul Pentland. A new sense for depth of field. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-9(4):523-531, 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[21] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[22] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1623-1637, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[23] Anupa Sabnis and Leena Vachhani. Single image based depth estimation for robotic applications. In 2011 IEEE Recent Advances in Intelligent Computational Systems, pages 102-106, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[24] Yoav Shechtman, Steffen J. Sahl, Adam S. Backer, and W. E. Moerner. Optimal point spread function design for 3d imaging. Phys. Rev. Lett., 113:133902, September 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[25] Yoav Shechtman, Lucien Weiss, Adam Backer, Steffen Sahl, and William Moerner. Precise three-dimensional scan-free multiple-particle tracking over large axial ranges with tetrapod point spread functions. *Nano letters*, 15, May 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[26] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Trans. Graph., 37(4), July 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[27] Qilin Sun, Ethan Tseng, Qiang Fu, Wolfgang Heidrich, and Felix Heide. Learning rank-1 diffractive optics for single-shot high dynamic range imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1386-1396, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10666" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Ashok Veeraraghavan, Ramesh Raskar, Amit Agrawal, Ankit Mohan, and Jack Tumblin. Dappled photography: Mask enhanced cameras for heterodyned light fields and coded aperture refocusing. ACM Trans. Graph., 26(3):69-es, jul 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.164, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[29] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.248, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[30] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.469, + 0.358 + ], + "angle": 0, + "content": "[31] Woontack Woo, Wonwoo Lee, and Nohyoung Park. Depth-assisted real-time 3d object detection for augmented reality. In International Conference on Artificial Reality and Teleexistence (ICAT), 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[32] Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin Sankaranarayanan, and Ashok Veeraraghavan. Phasecam3d — learning phase masks for passive single view depth estimation. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2019. 1, 2, 3, 4, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.445, + 0.469, + 0.528 + ], + "angle": 0, + "content": "[33] Feng Xue, Guirong Zhuo, Ziyuan Huang, Wufei Fu, Zhuoyue Wu, and Marcelo H Ang. Toward hierarchical self-supervised monocular absolute depth estimation for autonomous driving applications. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2330-2337. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.53, + 0.469, + 0.584 + ], + "angle": 0, + "content": "[34] Menglong Ye, Edward Johns, Ankur Handa, Lin Zhang, Philip Pratt, and Guang-Zhong Yang. Self-supervised siamese learning on stereo image pairs for depth estimation in robotic surgery, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10667" + } + ] +] \ No newline at end of file diff --git a/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/0fbb2c88-7b44-409c-9798-461b2922052e_origin.pdf b/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/0fbb2c88-7b44-409c-9798-461b2922052e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5462134239fdafbaf83b140e39002b685d8b8315 --- /dev/null +++ b/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/0fbb2c88-7b44-409c-9798-461b2922052e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66d25e7b5be8b84c52f40e197ade7ba9f60c3f6a8c66f7a845dc339a0679a5d7 +size 5314258 diff --git a/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/full.md b/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0a2d692e6f28a9412a730265751b3e5427c7e4f4 --- /dev/null +++ b/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/full.md @@ -0,0 +1,395 @@ +# TiDy-PSFs: Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions + +Sachin Shah* + +University of Maryland, College Park + +shah2022@umd.edu + +Sakshum Kulshrestha* + +University of Maryland, College Park + +sakshumk@umd.edu + +# Christopher A. Metzler +University of Maryland, College Park + +metzler@umd.edu + +# Abstract + +Point-spread-function (PSF) engineering is a powerful computational imaging technique wherein a custom phase mask is integrated into an optical system to encode additional information into captured images. Used in combination with deep learning, such systems now offer state-of-the-art performance at monocular depth estimation, extended depth-of-field imaging, lensless imaging, and other tasks. Inspired by recent advances in spatial light modulator (SLM) technology, this paper answers a natural question: Can one encode additional information and achieve superior performance by changing a phase mask dynamically over time? We first prove that the set of PSFs described by static phase masks is non-convex and that, as a result, time-averaged PSFs generated by dynamic phase masks are fundamentally more expressive. We then demonstrate, in simulation, that time-averaged dynamic (TiDy) phase masks can leverage this increased expressiveness to offer substantially improved monocular depth estimation and extended depth-of-field imaging performance. + +# 1. Introduction + +Extracting depth information from an image is a critical task across a range of applications including autonomous driving [29, 33], robotics [23, 34], microscopy [7, 19], and augmented reality [31, 14]. To this end, researchers have developed engineered phase masks and apertures which serve to encode depth information into an image [12, 25]. To optimize these phase masks, recent works have exploited deep learning: By simultaneously optimizing a phase mask and a reconstruction algorithm "end-to-end learning" is able to dramatically improve system performance [32, 26]. + +![](images/d6ef10940b8d16c587641143d9f8434ec8872535f1b4381367784e089ac36ac7.jpg) +Figure 1. Time-averaged Dynamic PSFs Top: Phase mask sequence that was optimized to perform simultaneous extended depth-of-field imaging and monocular depth estimation. Middle: Proposed TiDy PSFs at specific depths. Bottom left: Depth estimation and all-in-focus imaging performance improve as one averages over more phase masks. Bottom right: Depth-encoded image and reconstructed depth map. + +Most existing works have focused on learning or optimizing a single phase mask for passive depth perception. We conjecture that this restriction leaves much room for improvement. Perhaps by using an SLM to introduce a sequence of phase masks over time, one could do much better. + +Supporting this idea is the fact, which we prove in Theorem 3, that the set of PSFs described by a single phase mask is non-convex. This implies that time-averaged PSFs, which span the convex hull of this set, can be significantly more expressive. In this work, we exploit the PSF nonconvexity by developing a multi-phase mask end-to-end optimization approach for learning a sequence of phase masks whose PSFs are averaged over time. + +This work's central contributions are as follows: + +- We prove the set of PSFs generated by a single phase + +mask is non-convex. Thus, dynamic phase-masks offer a fundamentally larger design space. + +- We extend the end-to-end learning optics and algorithm design framework to design a dynamic set of phase masks. +- We demonstrate, in simulation, that time-averaged PSFs can achieve superior monocular depth estimation and extended depth-of-field imaging performance. + +# 2. Background + +Image Formation Model. One can simulate the formation of an image in a camera by discretizing an RGB image by depth, convolving each depth with it's the corresponding PSF, and compositing the outputs to form the signal on the sensor. This process can be represented by the equation + +$$ +I = \sum_ {d = 1} ^ {D} O _ {d} \left(L * h _ {d}\right), \tag {1} +$$ + +where $L$ represents all-in-focus image, $\{1,\dots ,D\}$ represent a set of discrete depth layers, $O_{d}$ is the occlusion mask at depth $d$ , and the set $\{h_1,\dots ,h_D\}$ represent the depth-dependent PSF, i.e., the cameras response to point sources at various depths [9]. Other works assume no depth discontinuities [26] or add additional computation to improve blurring at depth boundaries [10]. Our model is similar to those used in [32, 3]. + +PSF Formation Model. A PSF $h_d$ can be formed as a function of distance $d$ and phase modulation $\phi^M$ caused by height variation on a phase mask. + +$$ +h _ {d} = \left| \mathcal {F} \left[ A \exp \left(i \phi^ {D F} (d) + i \phi^ {M}\right) \right] \right| ^ {2} \tag {2} +$$ + +where $\phi^{DF}(d)$ is the defocus aberration due to the distance $d$ between the focus point and the depth plane. Note that because this PSF depends on depth, it can be used to encode depth information into $I$ [8]. + +The key idea behind PSF-engineering and end-to-end learning is that one can use the aforementioned relationships to encode additional information into a captured image $I$ by selecting a particularly effective mask $\phi^M$ . + +# 3. Related Work + +# 3.1. Computational Optics for Depth Tasks + +Optics based approaches for depth estimation use sensors and optical setups to encode and recover depth information. Many methods have used the depth-dependent blur induced by the imaging system to estimate the depth + +of pixels in an image [20]. These approaches compare the blur at different ranges to the expected blur caused by an aperture focused at a fixed distance. Groups improved on this idea by implementing coded apertures, retaining more high frequency information about the scene to disambiguate depths [12]. Similar to depth estimation tasks, static phase masks have been used to produce tailored PSFs more invariant to depth, allowing for extended depth-of-field imaging [6]. However, these optically driven approaches with numerical analysis have been passed in performance by modern deep neural networks, allowing for joint optimization of optical elements and neural reconstruction networks. + +# 3.2. Deep Optics for Depth Tasks + +Many methods have engineered phase masks with specific depth qualities. By maximizing Fisher information for depth, the coded image theoretically will have the most amount of depth cues as possible [24] and by minimizing Fisher information, one may achieve an extended depth-of-field image [6]. Deep learning techniques can be used to jointly train the optical parameters and neural network based estimation methods. The idea is that one can "code" an image to retain additional information about a scene, and then use a deep neural network to produce reconstructions. By using a differentiable model for light propagation, back-propagation can be used to update phase mask values simultaneously with neural network parameters. This approach was demonstrated for extended depth-of-field imaging [28, 26, 10, 13], depth estimation [32, 3, 10], high-dynamic-range imaging [16, 27], and holography [5, 4]. While these previous approaches successfully improved performance, they focused on enhancing a single phase mask. We build on these works by simultaneously optimizing multiple phase masks, which allows us to search over a larger space of PSFs. + +# 4. Theory + +Micro-electromechanical system (MEMS) based SLMs offer high frameters but have limited phase precision due to heavy quantization [1]. As [4] noted, intensity averaging of multiple frames can improve quality by increasing effective precision to overcome quantization. Our key insight is that even as SLM technology improves, intensity averaging yields a more expressive design space than a single phase mask. This is supported by the claim that the set of PSFs that can be generated by a single phase mask is non-convex. We provide a rigorous proof for the claim as follows. + +Theorem 1. The set of PSFs that can be generated by a phase mask with an infinite aperture ( $A(x) = 1$ for all $x \in \mathbb{R}^2$ ) is non-convex. + +Proof. Consider two tilt shift phase masks, $M_1(x) = ax$ and $M_2(x) = -ax$ , with respect to a coordinate $x$ with + +non-zero $a$ . The corresponding averaged PSF is $\frac{1}{2}\delta(a) + \frac{1}{2}\delta(-a)$ , where $\delta$ denotes a Dirac delta function. To realize this PSF with a single phase mask we would need the field at the aperture to satisfy + +$$ +\left| \mathcal {F} (E) \right| ^ {2} = \frac {1}{2} \delta (a) + \frac {1}{2} \delta (- a). \tag {3} +$$ + +This implies + +$$ +\begin{array}{l} E (x) = \frac {1}{\sqrt {2}} e ^ {- i a x} e ^ {i \gamma_ {1}} + \frac {1}{\sqrt {2}} e ^ {i a x} e ^ {i \gamma_ {2}}, (4) \\ = \frac {2}{\sqrt {2}} e ^ {i \frac {\gamma_ {1} + \gamma_ {2}}{2}} \cos \left(a x + \frac {\gamma_ {1} - \gamma_ {2}}{2}\right) (5) \\ \end{array} +$$ + +for some $\gamma_1, \gamma_2 \in \mathbb{R}$ . $E(x)$ 's amplitude varies according to a cosine and thus cannot be realized by a phase-only mask. + +![](images/5da6facc43e2d419a210f5266d8e6cd006331b7f7a07aeeb88273598254b276b.jpg) + +We now demonstrate the non-convexity claim holds for finite apertures that lie on a discrete grid. + +Definition 1. $A \in \{0,1\}^{N \times N}$ is some valid aperture with a non-zero region $S$ such that there exists lines $L_{1}$ and $L_{2}$ where $S$ can be contained between them, and $L_{1} \parallel L_{2}$ and $u = S \cap L_{1}$ and $v = S \cap L_{2}$ are single points (Figure 2). + +This definition of $A$ supports most commonly used apertures including but not limited to circles, squares, and $n$ -sided regular polygons. See supplement for proof for all shapes. + +Definition 2. Let $T_A(N)$ be the set of $N \times N$ matrices in $\mathbb{T}^{N \times N}$ with non-zero support $A$ , i.e., the matrix is supported only where $A = 1$ , where $\mathbb{T}$ is the complex unit circle. + +The PSF induced by a phase mask $M$ can be modeled as the squared magnitude of the Fourier transform of the pupil function $f$ [32]. + +Definition 3. Let $f: \mathbb{R}^{N \times N} \to T_A(N)$ be defined by + +$$ +f (M) = A \odot \exp (i D + i c M) \tag {6} +$$ + +where $\odot$ denotes entry-wise multiplication, and $D \in \mathbb{R}^{N \times N}$ and $c \in \mathbb{R} - \{0\}$ (the reals except for 0) are fixed constants. + +Definition 4. Let $g: T_A(N) \to \mathbb{R}^{N \times N}$ be defined by + +$$ +g (X) = \frac {\left| \mathcal {F} (X) \right| \odot \left| \mathcal {F} (X) \right|}{\left\| \mathcal {F} (X) \right\| _ {F} ^ {2}} \tag {7} +$$ + +where $\mathcal{F}$ denotes the discrete Fourier Transform with sufficient zero-padding, $|\cdot|$ denotes entry-wise absolute value, and $\| \cdot \|_F$ denotes the Frobenius norm. + +![](images/59e333fd386162a54fcd57030d467ca16227cb5f33a9e1295209df2ae792316b.jpg) +Figure 2. Example aperture that satisfies constraints on A. The aperture is fitted between parallel lines $L1$ and $L2$ , which only intersect the aperture at one point each. Common aperture shapes fit into these constraints. + +Lemma 2. From fourier optics theory [8], any single phase mask's PSF at a specific depth can be written as + +$$ +P S F = g \circ f. +$$ + +Theorem 3. The range of PSF is not a convex set. + +Proof. $f$ is clearly surjective, so it suffices to argue the range of $g$ is not convex. Assume by way of contradiction that the range of $g$ is convex. Then, for all $X^{(1)}, \ldots, X^{(k)} \in T_A(N)$ there exists $Y \in T_A(N)$ such that $g(Y) = \frac{1}{k} \sum_{i=1}^{k} g(X^{(i)})$ . By Parseval's Theorem, + +$$ +\| \mathcal {F} (X) \| _ {F} ^ {2} = N ^ {2} \| X \| _ {F} ^ {2} = N ^ {2} \sum_ {i = 0} ^ {N} \sum_ {j = 0} ^ {N} A _ {i, j} \tag {8} +$$ + +so the condition is + +$$ +| \mathcal {F} (Y) | \odot | \mathcal {F} (Y) | = \frac {1}{k} \sum_ {i = 1} ^ {k} | \mathcal {F} \left(X ^ {(i)}\right) | \odot | \mathcal {F} \left(X ^ {(i)}\right) | \tag {9} +$$ + +or equivalently + +$$ +\mathcal {F} (Y) \odot \overline {{\mathcal {F} (Y)}} = \frac {1}{k} \sum_ {i = 1} ^ {k} \mathcal {F} \left(X ^ {(i)}\right) \odot \overline {{\mathcal {F} \left(X ^ {(i)}\right)}}. \tag {10} +$$ + +Then the cross-correlation theorem reduces it to + +$$ +\mathcal {F} (Y \star Y) = \frac {1}{k} \sum_ {i = 1} ^ {k} \mathcal {F} \left(X ^ {(i)} \star X ^ {(i)}\right) \tag {11} +$$ + +where $\star$ denotes cross-correlation. Because the Fourier Transform is linear we finally have + +$$ +Y \star Y = \frac {1}{k} \sum_ {i = 1} ^ {k} X ^ {(i)} \star X ^ {(i)}. \tag {12} +$$ + +![](images/7ea44b005c5551a31dc3c8080f2c86574e5447acf106b291a30d55c6fc17f078.jpg) +Figure 3. Geometric interpretation of correlation $(\mathbf{X} \star \mathbf{X})_{\mathbf{v} - \mathbf{u}}$ . The figure represents the correlation step when the shift is $v - u$ . Notice that only $u$ and $v$ overlap once the shift is applied. + +Therefore, the convexity of the range of $g$ is equivalent to the convexity of the set $\{X \star X : X \in T_A(N)\}$ . We will show the set's projection onto a particular coordinate is not convex. + +$$ +(X \star X) _ {s, r} = \sum_ {i = 0} ^ {N} \sum_ {j = 0} ^ {N} X _ {i, j} \overline {{X _ {i + s , j + r}}} \tag {13} +$$ + +where we adopt the convention that $X_{s,r} = 0$ when $s, r > N$ or $s, r < 0$ . Take the points $u$ and $v$ from the definition of $A$ (1). Also observe that correlation can be represented geometrically as shifting $\overline{X}$ over $X$ . In this representation, notice that as the shift $(s, r)$ approaches $v - u$ , the non-zero overlap between $X$ and $\overline{X}$ shifted by $(s, r)$ approaches 1 by construction. That is, when $L_1$ is shifted to overlap $L_2$ , $u$ and $v$ will be the only non-zero overlaps between the shifted and original non-zero points (Figure 3). No other non-zero points can overlap above or below $L_2$ by definition of $S$ . Therefore, $(X \star X)_{v-u}$ becomes + +$$ +X _ {u} \bar {X _ {v}} + \sum_ {i = 1} ^ {N ^ {2} - 1} 0. \tag {14} +$$ + +Because $X_{u}\overline{X_{v}}\in \mathbb{T}$ , $(X\star X)_{v - u}\in \mathbb{T}$ which is a nonconvex set. Therefore, the set of correlation's of values on the complex unit circle masked by $A$ is also not convex, and so is $PSF$ . + +Time-averaged PSFs span the convex hull of the set of static-mask PSFs, meaning there exists some PSFs achievable only through intensity averaging PSFs from a sequence of phase masks. + +# 5. Multi-Phase Mask Optimization + +# 5.1. Optical Forward Model + +Similar to PhaseCam3D [32], we model light propagation using Fourier optics theory [8]. In contrast to previous + +work, we compute the forward model (1) for multiple phase masks, producing a stack of output images, which form our coded image when averaged. This coded image simulates the recorded signal from imaging a scene using a sequence of phase masks in a single exposure (Figure 4). + +# 5.2. Specialized Networks + +For the monocular depth estimation task, we use the MiDaS Small network [22]. This is a well known convolutional monocular depth estimation network designed to take in natural images and output relative depth maps. The network is trained end-to-end with the phase masks. A mean-squared error (MSE) loss term is defined in terms of the depth reconstruction prediction, $\hat{D}$ and the ground truth depth map $D$ , + +$$ +L _ {D e p t h} = \frac {1}{N} \| D - \hat {D} \| _ {2} ^ {2} \tag {15} +$$ + +where $N$ is the number of pixels. This process allows for the simultaneous optimization of the phase masks as well as fine tuning MiDaS to reconstruct from our coded images. + +For the extended depth-of-field task, we use an Attention U-Net [18] to reconstruct all-in-focus images. The network is optimized jointly with the phase mask sequence. To learn a reconstruction $\hat{I}$ to be similar to the all-in-focus ground truth image $I$ , we define the loss term using MSE error + +$$ +L _ {A i F} = \frac {1}{N} \| I - \hat {I} \| _ {2} ^ {2} \tag {16} +$$ + +where $N$ is the number of pixels. + +# 5.3. Joint Task Optimization + +We also present an alternative to the specialized networks: a single network jointly trained for monocular depth estimation and extended depth-of-field using a sequence of phase masks. This network has a basic Attention U-Net architecture outputting 4 channels representing depth maps as well as all-in-focus images. Similar to prior works, we use a combined loss function, adding a coefficient to weight the losses for each individual task: + +$$ +L _ {t o t a l} = \lambda_ {D e p t h} L _ {D e p t h} + \lambda_ {A i F} L _ {A i F}. \tag {17} +$$ + +# 6. Experimental Details + +# 6.1. Training Details + +We use the FlyingThings3D from Scene Flow Datasets [15], which uses synthetic data generation to obtain all-in-focus RGB images and disparity maps. We use the cropped $278 \times 278$ all-in-focus images from [32]. In total, we use 5077 training patches and 419 test patches. + +Both the optical layer and reconstruction networks are differentiable, so the phase mask sequence and neural network can be optimized through back-propagation. Each + +![](images/d822209196e8b2ba69d77c5c5ca65ccf34ceffe3a5a271695078321606299f63.jpg) +Figure 4. Multi-phase mask forward model overview. A sequence of phase masks are used to generate a sequence of depth-dependent PSFs. These PSFs are convolved with depth masked clean images to simulate depth dependent convolution. The images produced by each phase mask are averaged to create a coded image which is fed into an attention U-Net. The reconstruction loss is back-propagated end-to-end through the network and the optical model to design phase masks and algorithms capable of performing monocular depth estimation and extended depth-of-field simultaneously. + +part is implemented in PyTorch. During training, we use the Adam [11] optimizer with parameters $\beta_{1} = 0.99$ and $\beta_{2} = 0.999$ . The learning rate for the phase masks is $10^{-8}$ and for the reconstruction network it is $10^{-4}$ , and the batch size was 32. Finally, training and testing were performed on NVIDIA Quadro P6000 GPUs. + +We parameterize $23 \times 23$ phase masks pixel-wise as [13] found pixel-wise parameterization to produce the best overall performance. The monocular depth estimation task uses the MiDaS Small architecture pretrained weights for monocular depth estimation downloadable from PyTorch [22]. The extended depth-of-field task pretrains an Attention U-Net with a fixed Fresnel lens for 300 epochs. For the joint task, we set $\lambda_{Depth} = \lambda_{AiF} = 1$ to balance overall performance, and we pretrain the Attention U-Net for 300 epochs with a fixed Fresnel lens. In simulation, the red, blue, and green channels are approximated by discretized wavelengths, $610~\mathrm{nm}$ , $530~\mathrm{nm}$ , and $470~\mathrm{nm}$ respectively. Additionally, the depth range is discretized into 21 bins on the interval $[-20, 20]$ , which is larger than previous works. + +# 6.2. Evaluation Details + +For ablation studies on our method, we used the testing split of the FlyingThings3D set for both monocular depth estimation and extended depth-of-field imaging [15]. For comparisons to existing work, we also tested our monocular depth estimation network on the labeled NYU Depth v2 set [17]. The ground truth depth maps were translated to layered masks for the clean images by bucketing the depth values into 21 bins, allowing us to convolve each depth in an image with the required PSF. We use root mean squared error (RMSE) between ground truth and estimated depth maps for depth estimation and RMSE between ground truth and reconstructed all-in-focus images for extended depth-of-field imaging. We also use peak signal-to-noise ratio (PSNR) and structural similarity index [30] (SSIM) for extended depth-of-field imaging. + +# 7. Results + +We compare our time averaged dynamic PSF method to the state-of-the-art methods for both extended depth-of-field imaging and monocular depth estimation. The relevant works we compare to are as follows: + +1. PhaseCam3D [32] used a $23 \times 23$ phase mask based on 55 Zernike coefficients. The phase mask parameters were then end-to-end optimized with a U-Net reconstruction network to perform depth estimation. +2. Chang et al. [3] used a singlet lens introducing chromatic aberrations with radially symmetric PSFs. Similar to [32], the lens parameters were also then end-to-end optimized. +3. Ikoma et al. [10] used a radially symmetric diffractive optical element (DOE). The blurred image was preconditioned with an approximate inverse of the PSF depth dependent blur. The RGB image stack was fed into a U-Net to produce both an all-in-focus image and a depth map. The DOE and U-Net parameters were optimized in an end-to-end fashion. +4. Liu et al. [13] used various phase mask parameterizations with the same U-Net architecture as [10]. One method used pixel-wise height maps (PW) and the other introduced orbital angular momentum (OAM). +5. Sitzmann et al. [26] implements a single DOE based on Zernike coefficients, and solves the Tikhonov-regularized least-squares problem to reconstruct an all-in-focus image. +6. MiDaS [21] and ZoeDepth [2] are state of the art single shot monocular depth estimation methods with all-in-focus images as inputs. + +Because both [10] and [13] simultaneously learn all-in-focus images and depth maps, when comparing against our + +
MethodFlyingThings3DNYUv2
PhaseCam3D [32]0.5210.382
Chang et al. [3]0.4900.433
Ikoma et al. [10]0.184-
MiDaS [21]-0.357
ZoeDepth [2]-0.277
TiDy (1)0.0260.259
TiDy (5)0.0190.175
+ +Table 1. RMSE comparison of monocular depth estimation methods. We present quantitative results on two datasets to compare to state of the art optical and single shot monocular depth estimation methods. Our method performs best with our 5 phase mask system achieving the lowest error on both datasets. + +
MethodRMSE↓PSNR↑SSIM↑
Liu et al. [13]-29.80-
Ikoma et al. [10]0.132731.880.905
Sitzmann et al. [26]-32.44-
TiDy (1)0.014837.330.968
TiDy (5)0.009241.110.989
+ +specialized methods, we take their best performing weighting of each task. + +Individual Tasks. For monocular depth estimation, our specialized method using a sequence of 5 phase masks trained for 300 epochs outperforms prior work on FlyingThings3D (Table 1). Additionally, our approach performs significantly better and achieves lower error than previous methods on NYUv2 without any additional fine tuning. For extended depth-of-field, our specialized method using a sequence of 5 phase masks outperforms prior work on FlyingThings3D (Table 2). This demonstrates the benefit of multi-phase mask learning on computational imaging tasks. + +Multi-Objective Optimization. We also evaluate our method against other joint all-in-focus and depth map learning approaches. This problem is challenging because good depth cues to produce depth maps is antithetical to producing an all-in-focus image. Our combined 5 phase mask trained for 300 epochs approach outperforms prior jointly trained approaches (Table 3). + +Table 2. Comparison of extended depth-of-field imaging methods. We present quantitative results on FlyingThings3D to compare to state-of-the-art. Our method performs best with our 5 phase mask system achieving the best PSNR. + +
MethodAll-in-focus PSNR↑Depth RMSE↓
Ikoma et al. [10]31.880.191
Liu et al. [13] - PW29.800.056
Liu et al. [13] - OAMt25.860.053
TiDy (1)31.200.052
TiDy (5)34.790.034
+ +Table 3. Comparison of multi-objective optimization of extended depth-of-field imaging and depth estimation methods. We compare quantitative results on FlyingThings3D to the state-of-the-art. Our method performs best with our 5 phase mask system achieving the best balance between objectives. + +![](images/d3dec9c3c45e64e15eab15de2277d24bd6d465a930b012baf5d35fe175bbbacc.jpg) +Figure 5. RMSE for specialized tasks for each phase mask sequence length. RMSE decreases with respect to phase mask sequence length for both specialized extended depth-of-field imaging and monocular depth estimation tasks. 0 phase masks refers to a reconstruction neural network with a fixed Fresnel lens. + +![](images/4f261afc545da41913c04fdd4882016163dff55fc2d92c59376d68c71eafe23b.jpg) +Figure 6. RMSE for joint optimization of monocular depth estimation and extended depth-of-field imaging for each phase mask sequence length. RMSE decreases with respect to phase mask sequence length for this complex joint task, demonstrating the benefit of multi-phase mask learning. 0 phase masks refers to a reconstruction neural network with a fixed Fresnel lens. + +![](images/4127ab4795135aec3fcf921155eeb96c0baf098ed5a0911b4629988756e632a9.jpg) +Figure 7. All-in-focus imaging RMSE distribution for each phase mask length without a reconstruction network. The best RMSE for each phase mask count has low correlation with respect to phase mask sequence length, but the variance of RMSE decreases. + +# 8. Ablation Studies + +# 8.1. Effect of Phase Mask Sequence Length + +For both all-in-focus imaging and depth estimation, we vary the phase mask count that the end-to-end system is trained with to gauge the benefits of using multiple phase masks. The forward model and initial phase masks were held standard while the phase mask count was varied. The resulting networks were evaluated at convergence. For the extended depth-of-field task, the masks were all initialized with random noise uniform from 0 to $1.2 \times 10^{-6}$ . For the depth estimation task, the masks were initialized with the Fisher mask with added Gaussian noise parameterized by a $5.35 \times 10^{-7}$ mean and $3.05 \times 10^{-7}$ standard deviation. + +End-to-end optimization on each task with a specialized network yielded improved performance as the phase mask count increased, visualized in Figure 5. This result implies that sequences of phase masks are successful in making the PSF space more expressive. Additionally, even for the more complex joint task, learning a system that can produce both all-in-focus images and depth maps, error decreases with phase mask count until a plateau is reached (Figure 6). + +# 8.2. All-in-Focus without Reconstruction Networks + +A phase mask generating a PSF of the unit impulse function at every depth would be ideal for extended depth-of-field as each depth would be in focus. If possible, this phase mask would not require any digital processing. We optimize phase mask sequences of varying lengths to produce an averaged PSF close to the unit impulse function for all depths. For each sequence length, phase masks are optimized using MSE loss between the unit impulse function and the averaged PSF at each depth until convergence. We ran 1000 trials of random phase mask initialization for each length. Observe that a side-effect of longer phase masks is training + +![](images/49323f82872848c595d84c3a8a46b5a1c987e376054c3cedba130284ae30de34.jpg) +Figure 8. Visualization of phase mask initializations. Each row represents a different initial phase mask sequence. + +stability. The range of RMSE between the simulated capture image and ground truth all-in-focus image decreases as the sequence length increases (Figure 7). This indicates training longer sequences is more resilient to initialization. + +# 8.3. Phase Mask Initialization for Depth Perception + +Deep optics for depth perception can be very dependent on the initialization of optical parameters before training [32]. To find the extent of the effect of mask initialization on performance, we varied the the initial phase masks while keeping number of masks, the optical model, and duration of training fixed. We trained for 200 epochs. We tested four initializations of sequences of 5 phase masks as shown in Figure 8. The first was uniformly distributed noise from 0 to $1.2 \times 10^{-6}$ . The second was the first mask in the sequence set to a Fisher mask while the rest are uniform noise. The third is setting each mask to a rotation of the Fisher mask and adding Gaussian noise parameterized by a $5.35 \times 10^{-7}$ mean and $3.05 \times 10^{-7}$ standard deviation to 4 masks. Lastly, we set each mask to a rotation of the Fisher mask and added noise to only the last two masks in the sequence. Of the four initializations, it is clear that the 3 Fisher masks and 2 Fisher masks with noise performed the best (Table 4). + +# 8.4. Modeling SLM Imperfections + +State Switching. Our optical forward model assumes an SLM can swap between two phase patterns instantly. In practice, however, some light will be captured during the intermediate states between phase patterns. These phase patterns, in the worst case, could be random phase patterns, effectively adding noise to our coded images. We model these intermediate states by averaging output images pro + +
InitializationRMSE↓
1 Fisher + All noise0.0329
1 Fisher + Fisher w/ Noise0.0271
All noise0.0254
3 Fisher + Fisher w/ Noise0.0207
+ +Table 4. Quantitative evaluation of phase mask initializations. Four sequence initializations are evaluated on the monocular depth estimation task. Ultimately, 3 Fisher masks and 2 noisy Fisher masks have the best performance after training. + +![](images/04900e05ca70c5b7d5c4adf067137e62e360c1650799f3189c36cb2cdefa3be9.jpg) +Figure 9. Qualitative results of a specialized network on extended depth-of-field imaging. Both 1 and 5 phase mask systems are evaluated on FlyingThings3D. Error is computed pixel wise between the ground truth all-in-focus image and the reconstructed output and is boosted by a factor of 3. Notice that the 5 phase mask system introduces minimal error. + +duced by phase masks and the randomized phase patterns weighted by the time that they are displayed for. We model the total exposure time as $100\mathrm{ms}$ , with various durations of switching times from 1 to $16\mathrm{ms}$ per swap. We evaluate our joint optimized network on these new, more noisy, coded images without any additional training (Figure 12). Observe that because the 5 phase mask system includes more swaps, performance degrades faster than fewer phase mask systems. However, for short switching times, 5 phase masks still outperform the others without needing any fine tuning. + +Quantization. Current high-speed MEMS based SLM technology experience heavy quantization [1]. We model + +![](images/fbd13953e34a1d415320526170cd157d1f065042d76dfb0a66e4f27a2e4f469d.jpg) +Figure 10. Qualitative results of a specialized networks on monocular depth estimation. Performance using the five phase mask method outperforms one phase mask on both datasets. + +![](images/329bff1a5117f54e4ae67c656179621f630f238f55f3c15612a2bc8a63bfaf2f.jpg) +Figure 11. Qualitative results of a joint optimized system for extended depth-of-field imagining and monocular depth estimation. Both one and five phase mask networks are evaluated on the FlyingThings3D datasets. Notice that five masks has fewer artifacts than a single mask. + +the effects of quantization error by adding varying amounts of noise during inference. As demonstrated in Figure 13, + +![](images/73b4aa753f6b63d6c7513ad1788aa7affc98808f340aba7129a338178d155d19.jpg) +Figure 12. Effect of switching time on joint system performance. Reconstruction error across phase mask counts as a function of switching time with $100\mathrm{ms}$ overall exposure. Performance of the jointly optimized system degrades as the switching time between phase masks increases, as expected. Our system still performs well when the time spent switching is less than $25\%$ of the overall exposure. + +![](images/dde2d180287f3dcf84e733d47386b05c20afb33db3f39c2485d97dcc2c9c7242.jpg) +Figure 13. Effect of quantization on joint system performance. Performance of the jointly optimized system degrades as the number of effective bits decrease; however, even under heavy quantization, our system outperforms single mask setups. + +time-averaged PSFs are robust to quantization noise: 5 masks quantized to 4-bits outperform one mask quantized to 32-bits. This is intuitive as multiple masks can leverage averaging to achieve better precision and remove noise [4]. + +# 8.5. A Path Towards Better Single-Mask Designs + +The performance of end-to-end designed phase masks is highly sensitive to how they are initialized [32]. Can we use our 5-mask design to provide a better initialization for a 1-mask design? + +Starting from random initializations, we optimized 200 single phase masks to minimize the mean squared error between each of their PSFs and the time-averaged PSF formed by our optimized 5-mask system. We then took the single mask whose PSF was closest to the 5-mask design and fine-tuned our reconstruction network using its PSF. This static design improved performance over a randomly initialized end-to-end trained 1-mask system; +2.67 PSNR for all-in-focus imaging and -0.003 RMSE for depth estima + +tion. However, its performance was still strictly worse than a 5-mask system. + +# 8.6. Modulating Amplitude and Phase + +As illustrated in our proof of Theorem 1, "complex" masks which modulate both amplitude and phase are more expressive than phase-only masks. Using end-to-end learning, we optimized a pixel wise complex mask (initialized with a random phase mask and all ones amplitude mask) on the joint depth estimation and all-in-focus imaging task. The complex mask offered substationally improved performance; $+2.49$ PSNR for all-in-focus imaging and $-0.015$ RMSE for depth estimation. However, its performance too was still strictly worse than a 5-mask system. We conjecture this is due to improved training stability of the 5-mask system. + +# 9. Limitations + +While we were successful in learning dynamic phase masks to improve state-of-the-art performance on imaging tasks, our method carries some limitations. First, our optical model assumes perfect switching between phase masks during training. While evaluation with non-zero switching times showed little degradation of performance, accounting for state switching while training could produce phase masks that are more performant. Our optical model also simulates depths as layered masks over an image, which does not account for blending at depth boundaries. Additionally, our method assumes that scenes are static for the duration of a single exposure. Lastly, though their prices are falling, SLMs are still quite expensive and bulky. + +# 10. Conclusion + +This work is founded upon the insight that the set of PSFs that can be described by a single phase mask is nonconvex and that, as a result, time-averaged PSFs are fundamentally more expressive. We demonstrate that one can learn a sequence of phase masks that, when one dynamically switches among them over time, can substantially improve computational imaging performance across a range of tasks, including depth estimation and all-in-focus imaging. Our work unlocks an exciting new direction for PSF engineering and computational imaging system design. + +# Acknowledgements + +This work was partially supported by AFOSR Young Investigator Program Award FA9550-22-1-0208. + +# References + +[1] Terry A. Bartlett, William C. McDonald, and James N. Hall. Adapting Texas Instruments DLP technology to demonstrate a phase spatial light modulator. In Michael R. Douglass, John + +Ehmke, and Benjamin L. Lee, editors, Emerging Digital Micromirror Device Based Systems and Applications XI, volume 10932, page 109320S. International Society for Optics and Photonics, SPIE, 2019. 2, 8 +[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 5, 6 +[3] Julie Chang and Gordon Wetzstein. Deep optics for monocular depth estimation and 3d object detection. In Proc. IEEE ICCV, 2019. 2, 5, 6 +[4] Suyeon Choi, Manu Gopakumar, Yifan Peng, Jonghyun Kim, Matthew O'Toole, and Gordon Wetzstein. Time-multiplexed neural holography: A flexible framework for holographic near-eye displays with fast heavily-quantized spatial light modulators. In Proceedings of the ACM SIGGRAPH, page 1-9, 2022. 2, 9 +[5] Suyeon Choi, Manu Gopakumar, Yifan Peng, Jonghyun Kim, and Gordon Wetzstein. Neural 3d holography: Learning accurate wave propagation models for 3d holographic virtual and augmented reality displays. ACM Trans. Graph., 40(6), December 2021. 2 +[6] Edward R. Dowski and W. Thomas Cathey. Extended depth of field through wave-front coding. Appl. Opt., 34(11):1859-1866, April 1995. 2 +[7] Robert Fischer, Yicong Wu, Pakorn Kanchanawong, Hari Shroff, and Clare Waterman-Storer. Microscopy in 3d: A biologist's toolbox. Trends in cell biology, 21:682-91, October 2011. 1 +[8] Joseph W. Goodman. Introduction to fourier optics. Freeman, 2017. 2, 3, 4 +[9] Samuel W. Hasinoff and Kiriakos N. Kutulakos. A layer-based restoration framework for variable-aperture photography. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8, 2007. 2 +[10] Hayato Ikoma, Cindy M. Nguyen, Christopher A. Metzler, Yifan Peng, and Gordon Wetzstein. Depth from defocus with learned optics for imaging and occlusion-aware depth estimation. IEEE International Conference on Computational Photography (ICCP), 2021. 2, 5, 6 +[11] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 5 +[12] Anat Levin, Rob Fergus, Frédo Durand, and William T. Freeman. Image and depth from a conventional camera with a coded aperture. In ACM SIGGRAPH '2007 Papers, SIGGRAPH '07, page 70-es, New York, NY, USA, 2007. Association for Computing Machinery. 1, 2 +[13] Xin Liu, Linpei Li, Xu Liu, Xiang Hao, and Yifan Peng. Investigating deep optics model representation in affecting resolved all-in-focus image quality and depth estimation fidelity. Opt. Express, 30(20):36973-36984, September 2022. 2, 5, 6 +[14] Yawen Lu, Sophia Kourian, Carl Salvaggio, Chenliang Xu, and Guoyu Lu. Single image 3d vehicle pose estimation for augmented reality. In 2019 IEEE Global Conference on Signal and Information Processing (GlobalSIP), pages 1-5, 2019. 1 +[15] Nikolaus Mayer, Eddy Ilg, Philip Häusser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A + +large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016. 4, 5 +[16] Christopher A Metzler, Hayato Ikoma, Yifan Peng, and Gordon Wetzstein. Deep optics for single-shot high-dynamic-range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1375-1385, 2020. 2 +[17] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 5 +[18] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert. Attention u-net: Learning where to look for the pancreas. In Medical Imaging with Deep Learning, 2018. 4 +[19] Luca Palmieri, Gabriele Scrofani, Nicolò Incardona, Genaro Saavedra, Manuel Martínez-Corral, and Reinhard Koch. Robust depth estimation for light field microscopy. Sensors, 19(3), 2019. 1 +[20] Alex Paul Pentland. A new sense for depth of field. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-9(4):523-531, 1987. 2 +[21] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 5, 6 +[22] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1623-1637, 2022. 4, 5 +[23] Anupa Sabnis and Leena Vachhani. Single image based depth estimation for robotic applications. In 2011 IEEE Recent Advances in Intelligent Computational Systems, pages 102-106, 2011. 1 +[24] Yoav Shechtman, Steffen J. Sahl, Adam S. Backer, and W. E. Moerner. Optimal point spread function design for 3d imaging. Phys. Rev. Lett., 113:133902, September 2014. 2 +[25] Yoav Shechtman, Lucien Weiss, Adam Backer, Steffen Sahl, and William Moerner. Precise three-dimensional scan-free multiple-particle tracking over large axial ranges with tetrapod point spread functions. *Nano letters*, 15, May 2015. 1 +[26] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Trans. Graph., 37(4), July 2018. 1, 2, 5, 6 +[27] Qilin Sun, Ethan Tseng, Qiang Fu, Wolfgang Heidrich, and Felix Heide. Learning rank-1 diffractive optics for single-shot high dynamic range imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1386-1396, 2020. 2 + +[28] Ashok Veeraraghavan, Ramesh Raskar, Amit Agrawal, Ankit Mohan, and Jack Tumblin. Dappled photography: Mask enhanced cameras for heterodyned light fields and coded aperture refocusing. ACM Trans. Graph., 26(3):69-es, jul 2007. 2 +[29] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1 +[30] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5 +[31] Woontack Woo, Wonwoo Lee, and Nohyoung Park. Depth-assisted real-time 3d object detection for augmented reality. In International Conference on Artificial Reality and Teleexistence (ICAT), 2011. 1 +[32] Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin Sankaranarayanan, and Ashok Veeraraghavan. Phasecam3d — learning phase masks for passive single view depth estimation. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2019. 1, 2, 3, 4, 5, 6, 7, 9 +[33] Feng Xue, Guirong Zhuo, Ziyuan Huang, Wufei Fu, Zhuoyue Wu, and Marcelo H Ang. Toward hierarchical self-supervised monocular absolute depth estimation for autonomous driving applications. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2330-2337. IEEE, 2020. 1 +[34] Menglong Ye, Edward Johns, Ankur Handa, Lin Zhang, Philip Pratt, and Guang-Zhong Yang. Self-supervised siamese learning on stereo image pairs for depth estimation in robotic surgery, 2017. 1 \ No newline at end of file diff --git a/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/images.zip b/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1275d94d4ed86cc44a89445ddf89dedbbdc16a62 --- /dev/null +++ b/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b3fa72a56eb6e0bd4cbbd69a5daa6d247583450542653671e429be42b20d711 +size 514773 diff --git a/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/layout.json b/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..48fa338fcf69a42a38a1963dcbf3f000285ecf06 --- /dev/null +++ b/2023/TiDy-PSFs_ Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions/layout.json @@ -0,0 +1,10054 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 102, + 503, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 102, + 503, + 139 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 503, + 139 + ], + "type": "text", + "content": "TiDy-PSFs: Computational Imaging with Time-Averaged Dynamic Point-Spread-Functions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 161, + 224, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 161, + 224, + 175 + ], + "spans": [ + { + "bbox": [ + 157, + 161, + 224, + 175 + ], + "type": "text", + "content": "Sachin Shah*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 175, + 282, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 175, + 282, + 190 + ], + "spans": [ + { + "bbox": [ + 99, + 175, + 282, + 190 + ], + "type": "text", + "content": "University of Maryland, College Park" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 192, + 235, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 192, + 235, + 201 + ], + "spans": [ + { + "bbox": [ + 146, + 192, + 235, + 201 + ], + "type": "text", + "content": "shah2022@umd.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 347, + 161, + 457, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 161, + 457, + 175 + ], + "spans": [ + { + "bbox": [ + 347, + 161, + 457, + 175 + ], + "type": "text", + "content": "Sakshum Kulshrestha*" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 176, + 494, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 494, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 494, + 190 + ], + "type": "text", + "content": "University of Maryland, College Park" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 359, + 192, + 447, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 192, + 447, + 201 + ], + "spans": [ + { + "bbox": [ + 359, + 192, + 447, + 201 + ], + "type": "text", + "content": "sakshumk@umd.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 205, + 209, + 388, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 209, + 388, + 237 + ], + "spans": [ + { + "bbox": [ + 205, + 209, + 388, + 237 + ], + "type": "text", + "content": "Christopher A. Metzler \nUniversity of Maryland, College Park" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 255, + 240, + 338, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 240, + 338, + 250 + ], + "spans": [ + { + "bbox": [ + 255, + 240, + 338, + 250 + ], + "type": "text", + "content": "metzler@umd.edu" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 278, + 192, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 278, + 192, + 291 + ], + "spans": [ + { + "bbox": [ + 143, + 278, + 192, + 291 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 304, + 290, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 290, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 290, + 533 + ], + "type": "text", + "content": "Point-spread-function (PSF) engineering is a powerful computational imaging technique wherein a custom phase mask is integrated into an optical system to encode additional information into captured images. Used in combination with deep learning, such systems now offer state-of-the-art performance at monocular depth estimation, extended depth-of-field imaging, lensless imaging, and other tasks. Inspired by recent advances in spatial light modulator (SLM) technology, this paper answers a natural question: Can one encode additional information and achieve superior performance by changing a phase mask dynamically over time? We first prove that the set of PSFs described by static phase masks is non-convex and that, as a result, time-averaged PSFs generated by dynamic phase masks are fundamentally more expressive. We then demonstrate, in simulation, that time-averaged dynamic (TiDy) phase masks can leverage this increased expressiveness to offer substantially improved monocular depth estimation and extended depth-of-field imaging performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 554, + 128, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 128, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 128, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 574, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 695 + ], + "type": "text", + "content": "Extracting depth information from an image is a critical task across a range of applications including autonomous driving [29, 33], robotics [23, 34], microscopy [7, 19], and augmented reality [31, 14]. To this end, researchers have developed engineered phase masks and apertures which serve to encode depth information into an image [12, 25]. To optimize these phase masks, recent works have exploited deep learning: By simultaneously optimizing a phase mask and a reconstruction algorithm \"end-to-end learning\" is able to dramatically improve system performance [32, 26]." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 315, + 281, + 539, + 423 + ], + "blocks": [ + { + "bbox": [ + 315, + 281, + 539, + 423 + ], + "lines": [ + { + "bbox": [ + 315, + 281, + 539, + 423 + ], + "spans": [ + { + "bbox": [ + 315, + 281, + 539, + 423 + ], + "type": "image", + "image_path": "d6ef10940b8d16c587641143d9f8434ec8872535f1b4381367784e089ac36ac7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 434, + 547, + 513 + ], + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 513 + ], + "type": "text", + "content": "Figure 1. Time-averaged Dynamic PSFs Top: Phase mask sequence that was optimized to perform simultaneous extended depth-of-field imaging and monocular depth estimation. Middle: Proposed TiDy PSFs at specific depths. Bottom left: Depth estimation and all-in-focus imaging performance improve as one averages over more phase masks. Bottom right: Depth-encoded image and reconstructed depth map." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 526, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 586 + ], + "type": "text", + "content": "Most existing works have focused on learning or optimizing a single phase mask for passive depth perception. We conjecture that this restriction leaves much room for improvement. Perhaps by using an SLM to introduce a sequence of phase masks over time, one could do much better." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 681 + ], + "type": "text", + "content": "Supporting this idea is the fact, which we prove in Theorem 3, that the set of PSFs described by a single phase mask is non-convex. This implies that time-averaged PSFs, which span the convex hull of this set, can be significantly more expressive. In this work, we exploit the PSF nonconvexity by developing a multi-phase mask end-to-end optimization approach for learning a sequence of phase masks whose PSFs are averaged over time." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 681, + 512, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 512, + 693 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 512, + 693 + ], + "type": "text", + "content": "This work's central contributions are as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "- We prove the set of PSFs generated by a single phase" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 212, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 212, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 212, + 712 + ], + "type": "text", + "content": "*These authors contributed equally to this work" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10657" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 97 + ], + "type": "text", + "content": "mask is non-convex. Thus, dynamic phase-masks offer a fundamentally larger design space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 105, + 287, + 185 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 59, + 105, + 287, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 105, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 59, + 105, + 287, + 140 + ], + "type": "text", + "content": "- We extend the end-to-end learning optics and algorithm design framework to design a dynamic set of phase masks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 148, + 286, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 148, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 59, + 148, + 286, + 185 + ], + "type": "text", + "content": "- We demonstrate, in simulation, that time-averaged PSFs can achieve superior monocular depth estimation and extended depth-of-field imaging performance." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 196, + 126, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 196, + 126, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 196, + 126, + 209 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 216, + 287, + 277 + ], + "type": "text", + "content": "Image Formation Model. One can simulate the formation of an image in a camera by discretizing an RGB image by depth, convolving each depth with it's the corresponding PSF, and compositing the outputs to form the signal on the sensor. This process can be represented by the equation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 285, + 287, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 285, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 121, + 285, + 287, + 318 + ], + "type": "interline_equation", + "content": "I = \\sum_ {d = 1} ^ {D} O _ {d} \\left(L * h _ {d}\\right), \\tag {1}", + "image_path": "9c4c847201d8458a4579dc2b08c27c98333e16504992ef590114cfcb8e5ae5a3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 327, + 287, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 287, + 423 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 327, + 287, + 423 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 327, + 287, + 423 + ], + "type": "text", + "content": " represents all-in-focus image, " + }, + { + "bbox": [ + 47, + 327, + 287, + 423 + ], + "type": "inline_equation", + "content": "\\{1,\\dots ,D\\}" + }, + { + "bbox": [ + 47, + 327, + 287, + 423 + ], + "type": "text", + "content": " represent a set of discrete depth layers, " + }, + { + "bbox": [ + 47, + 327, + 287, + 423 + ], + "type": "inline_equation", + "content": "O_{d}" + }, + { + "bbox": [ + 47, + 327, + 287, + 423 + ], + "type": "text", + "content": " is the occlusion mask at depth " + }, + { + "bbox": [ + 47, + 327, + 287, + 423 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 327, + 287, + 423 + ], + "type": "text", + "content": ", and the set " + }, + { + "bbox": [ + 47, + 327, + 287, + 423 + ], + "type": "inline_equation", + "content": "\\{h_1,\\dots ,h_D\\}" + }, + { + "bbox": [ + 47, + 327, + 287, + 423 + ], + "type": "text", + "content": " represent the depth-dependent PSF, i.e., the cameras response to point sources at various depths [9]. Other works assume no depth discontinuities [26] or add additional computation to improve blurring at depth boundaries [10]. Our model is similar to those used in [32, 3]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "type": "text", + "content": "PSF Formation Model. A PSF " + }, + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "type": "inline_equation", + "content": "h_d" + }, + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "type": "text", + "content": " can be formed as a function of distance " + }, + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "type": "text", + "content": " and phase modulation " + }, + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\phi^M" + }, + { + "bbox": [ + 47, + 438, + 287, + 475 + ], + "type": "text", + "content": " caused by height variation on a phase mask." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 92, + 483, + 287, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 483, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 92, + 483, + 287, + 498 + ], + "type": "interline_equation", + "content": "h _ {d} = \\left| \\mathcal {F} \\left[ A \\exp \\left(i \\phi^ {D F} (d) + i \\phi^ {M}\\right) \\right] \\right| ^ {2} \\tag {2}", + "image_path": "733dd358b855defc97f3c8e44b6d0a6ef8bb8bf7b7718dd94799f0144f5ce341.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 507, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 507, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 287, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 507, + 287, + 556 + ], + "type": "inline_equation", + "content": "\\phi^{DF}(d)" + }, + { + "bbox": [ + 47, + 507, + 287, + 556 + ], + "type": "text", + "content": " is the defocus aberration due to the distance " + }, + { + "bbox": [ + 47, + 507, + 287, + 556 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 507, + 287, + 556 + ], + "type": "text", + "content": " between the focus point and the depth plane. Note that because this PSF depends on depth, it can be used to encode depth information into " + }, + { + "bbox": [ + 47, + 507, + 287, + 556 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 507, + 287, + 556 + ], + "type": "text", + "content": " [8]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 567, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 287, + 616 + ], + "type": "text", + "content": "The key idea behind PSF-engineering and end-to-end learning is that one can use the aforementioned relationships to encode additional information into a captured image " + }, + { + "bbox": [ + 47, + 567, + 287, + 616 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 567, + 287, + 616 + ], + "type": "text", + "content": " by selecting a particularly effective mask " + }, + { + "bbox": [ + 47, + 567, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\phi^M" + }, + { + "bbox": [ + 47, + 567, + 287, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 627, + 134, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 134, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 134, + 639 + ], + "type": "text", + "content": "3. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 647, + 250, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 250, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 250, + 659 + ], + "type": "text", + "content": "3.1. Computational Optics for Depth Tasks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "Optics based approaches for depth estimation use sensors and optical setups to encode and recover depth information. Many methods have used the depth-dependent blur induced by the imaging system to estimate the depth" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 216 + ], + "type": "text", + "content": "of pixels in an image [20]. These approaches compare the blur at different ranges to the expected blur caused by an aperture focused at a fixed distance. Groups improved on this idea by implementing coded apertures, retaining more high frequency information about the scene to disambiguate depths [12]. Similar to depth estimation tasks, static phase masks have been used to produce tailored PSFs more invariant to depth, allowing for extended depth-of-field imaging [6]. However, these optically driven approaches with numerical analysis have been passed in performance by modern deep neural networks, allowing for joint optimization of optical elements and neural reconstruction networks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 222, + 463, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 222, + 463, + 235 + ], + "spans": [ + { + "bbox": [ + 306, + 222, + 463, + 235 + ], + "type": "text", + "content": "3.2. Deep Optics for Depth Tasks" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 240, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 492 + ], + "type": "text", + "content": "Many methods have engineered phase masks with specific depth qualities. By maximizing Fisher information for depth, the coded image theoretically will have the most amount of depth cues as possible [24] and by minimizing Fisher information, one may achieve an extended depth-of-field image [6]. Deep learning techniques can be used to jointly train the optical parameters and neural network based estimation methods. The idea is that one can \"code\" an image to retain additional information about a scene, and then use a deep neural network to produce reconstructions. By using a differentiable model for light propagation, back-propagation can be used to update phase mask values simultaneously with neural network parameters. This approach was demonstrated for extended depth-of-field imaging [28, 26, 10, 13], depth estimation [32, 3, 10], high-dynamic-range imaging [16, 27], and holography [5, 4]. While these previous approaches successfully improved performance, they focused on enhancing a single phase mask. We build on these works by simultaneously optimizing multiple phase masks, which allows us to search over a larger space of PSFs." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 502, + 359, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 359, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 359, + 515 + ], + "type": "text", + "content": "4. Theory" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "content": "Micro-electromechanical system (MEMS) based SLMs offer high frameters but have limited phase precision due to heavy quantization [1]. As [4] noted, intensity averaging of multiple frames can improve quality by increasing effective precision to overcome quantization. Our key insight is that even as SLM technology improves, intensity averaging yields a more expressive design space than a single phase mask. This is supported by the claim that the set of PSFs that can be generated by a single phase mask is non-convex. We provide a rigorous proof for the claim as follows." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "text", + "content": "Theorem 1. The set of PSFs that can be generated by a phase mask with an infinite aperture (" + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "inline_equation", + "content": "A(x) = 1" + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 647, + 545, + 682 + ], + "type": "text", + "content": ") is non-convex." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Proof. Consider two tilt shift phase masks, " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "M_1(x) = ax" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "M_2(x) = -ax" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": ", with respect to a coordinate " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10658" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": "non-zero " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": ". The corresponding averaged PSF is " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}\\delta(a) + \\frac{1}{2}\\delta(-a)" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": " denotes a Dirac delta function. To realize this PSF with a single phase mask we would need the field at the aperture to satisfy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 129, + 287, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 129, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 287, + 152 + ], + "type": "interline_equation", + "content": "\\left| \\mathcal {F} (E) \\right| ^ {2} = \\frac {1}{2} \\delta (a) + \\frac {1}{2} \\delta (- a). \\tag {3}", + "image_path": "de6b4e421bea44dc4e9ff49d6af2d220a89a9e5b126782c34af3fdb0ce570f64.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 161, + 100, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 161, + 100, + 173 + ], + "spans": [ + { + "bbox": [ + 47, + 161, + 100, + 173 + ], + "type": "text", + "content": "This implies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 181, + 287, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 181, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 81, + 181, + 287, + 235 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} E (x) = \\frac {1}{\\sqrt {2}} e ^ {- i a x} e ^ {i \\gamma_ {1}} + \\frac {1}{\\sqrt {2}} e ^ {i a x} e ^ {i \\gamma_ {2}}, (4) \\\\ = \\frac {2}{\\sqrt {2}} e ^ {i \\frac {\\gamma_ {1} + \\gamma_ {2}}{2}} \\cos \\left(a x + \\frac {\\gamma_ {1} - \\gamma_ {2}}{2}\\right) (5) \\\\ \\end{array}", + "image_path": "04ab7e90017653308d654c420d6bafde1186a5370b1b641f0d513ec853088a46.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 244, + 287, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 244, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 244, + 287, + 269 + ], + "type": "text", + "content": "for some " + }, + { + "bbox": [ + 47, + 244, + 287, + 269 + ], + "type": "inline_equation", + "content": "\\gamma_1, \\gamma_2 \\in \\mathbb{R}" + }, + { + "bbox": [ + 47, + 244, + 287, + 269 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 244, + 287, + 269 + ], + "type": "inline_equation", + "content": "E(x)" + }, + { + "bbox": [ + 47, + 244, + 287, + 269 + ], + "type": "text", + "content": "'s amplitude varies according to a cosine and thus cannot be realized by a phase-only mask." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 276, + 269, + 287, + 278 + ], + "blocks": [ + { + "bbox": [ + 276, + 269, + 287, + 278 + ], + "lines": [ + { + "bbox": [ + 276, + 269, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 276, + 269, + 287, + 278 + ], + "type": "image", + "image_path": "5da6facc43e2d419a210f5266d8e6cd006331b7f7a07aeeb88273598254b276b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 287, + 316 + ], + "type": "text", + "content": "We now demonstrate the non-convexity claim holds for finite apertures that lie on a discrete grid." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "text", + "content": "Definition 1. " + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "inline_equation", + "content": "A \\in \\{0,1\\}^{N \\times N}" + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "text", + "content": " is some valid aperture with a non-zero region " + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "text", + "content": " such that there exists lines " + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "text", + "content": " can be contained between them, and " + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "inline_equation", + "content": "L_{1} \\parallel L_{2}" + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "inline_equation", + "content": "u = S \\cap L_{1}" + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "inline_equation", + "content": "v = S \\cap L_{2}" + }, + { + "bbox": [ + 47, + 324, + 287, + 373 + ], + "type": "text", + "content": " are single points (Figure 2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 381, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 287, + 430 + ], + "type": "text", + "content": "This definition of " + }, + { + "bbox": [ + 47, + 381, + 287, + 430 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 381, + 287, + 430 + ], + "type": "text", + "content": " supports most commonly used apertures including but not limited to circles, squares, and " + }, + { + "bbox": [ + 47, + 381, + 287, + 430 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 381, + 287, + 430 + ], + "type": "text", + "content": "-sided regular polygons. See supplement for proof for all shapes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "text", + "content": "Definition 2. Let " + }, + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "inline_equation", + "content": "T_A(N)" + }, + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "text", + "content": " be the set of " + }, + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "text", + "content": " matrices in " + }, + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathbb{T}^{N \\times N}" + }, + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "text", + "content": " with non-zero support " + }, + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "text", + "content": ", i.e., the matrix is supported only where " + }, + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "inline_equation", + "content": "A = 1" + }, + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 47, + 439, + 287, + 475 + ], + "type": "text", + "content": " is the complex unit circle." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 484, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 287, + 521 + ], + "type": "text", + "content": "The PSF induced by a phase mask " + }, + { + "bbox": [ + 47, + 484, + 287, + 521 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 484, + 287, + 521 + ], + "type": "text", + "content": " can be modeled as the squared magnitude of the Fourier transform of the pupil function " + }, + { + "bbox": [ + 47, + 484, + 287, + 521 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 484, + 287, + 521 + ], + "type": "text", + "content": " [32]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 529, + 265, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 265, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 265, + 542 + ], + "type": "text", + "content": "Definition 3. Let " + }, + { + "bbox": [ + 47, + 529, + 265, + 542 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^{N \\times N} \\to T_A(N)" + }, + { + "bbox": [ + 47, + 529, + 265, + 542 + ], + "type": "text", + "content": " be defined by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 552, + 287, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 552, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 104, + 552, + 287, + 566 + ], + "type": "interline_equation", + "content": "f (M) = A \\odot \\exp (i D + i c M) \\tag {6}", + "image_path": "43646da9a2b9ffa80b2901805a567e418bf91b95978c356cd26525a32173dcab.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 575, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 287, + 610 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 575, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 575, + 287, + 610 + ], + "type": "text", + "content": " denotes entry-wise multiplication, and " + }, + { + "bbox": [ + 47, + 575, + 287, + 610 + ], + "type": "inline_equation", + "content": "D \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 47, + 575, + 287, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 575, + 287, + 610 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R} - \\{0\\}" + }, + { + "bbox": [ + 47, + 575, + 287, + 610 + ], + "type": "text", + "content": " (the reals except for 0) are fixed constants." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 619, + 264, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 264, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 264, + 633 + ], + "type": "text", + "content": "Definition 4. Let " + }, + { + "bbox": [ + 47, + 619, + 264, + 633 + ], + "type": "inline_equation", + "content": "g: T_A(N) \\to \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 47, + 619, + 264, + 633 + ], + "type": "text", + "content": " be defined by" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 110, + 641, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 641, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 110, + 641, + 287, + 669 + ], + "type": "interline_equation", + "content": "g (X) = \\frac {\\left| \\mathcal {F} (X) \\right| \\odot \\left| \\mathcal {F} (X) \\right|}{\\left\\| \\mathcal {F} (X) \\right\\| _ {F} ^ {2}} \\tag {7}", + "image_path": "1292f7c0f8f0096603a52c83c6373bcd8490b2d3bc83261a21ba8d2f155c8f9f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " denotes the discrete Fourier Transform with sufficient zero-padding, " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " denotes entry-wise absolute value, and " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_F" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " denotes the Frobenius norm." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 354, + 69, + 500, + 228 + ], + "blocks": [ + { + "bbox": [ + 354, + 69, + 500, + 228 + ], + "lines": [ + { + "bbox": [ + 354, + 69, + 500, + 228 + ], + "spans": [ + { + "bbox": [ + 354, + 69, + 500, + 228 + ], + "type": "image", + "image_path": "59e333fd386162a54fcd57030d467ca16227cb5f33a9e1295209df2ae792316b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 232, + 545, + 275 + ], + "lines": [ + { + "bbox": [ + 305, + 232, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 545, + 275 + ], + "type": "text", + "content": "Figure 2. Example aperture that satisfies constraints on A. The aperture is fitted between parallel lines " + }, + { + "bbox": [ + 305, + 232, + 545, + 275 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 305, + 232, + 545, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 232, + 545, + 275 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 305, + 232, + 545, + 275 + ], + "type": "text", + "content": ", which only intersect the aperture at one point each. Common aperture shapes fit into these constraints." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 316 + ], + "type": "text", + "content": "Lemma 2. From fourier optics theory [8], any single phase mask's PSF at a specific depth can be written as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 395, + 323, + 456, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 323, + 456, + 336 + ], + "spans": [ + { + "bbox": [ + 395, + 323, + 456, + 336 + ], + "type": "interline_equation", + "content": "P S F = g \\circ f.", + "image_path": "22308d4cc770291948209dd347bcd684f9bf51469dc2a751985bc249df9aabf6.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 350, + 510, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 510, + 362 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 510, + 362 + ], + "type": "text", + "content": "Theorem 3. The range of PSF is not a convex set." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "text", + "content": "Proof. " + }, + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "text", + "content": " is clearly surjective, so it suffices to argue the range of " + }, + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "text", + "content": " is not convex. Assume by way of contradiction that the range of " + }, + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "text", + "content": " is convex. Then, for all " + }, + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "inline_equation", + "content": "X^{(1)}, \\ldots, X^{(k)} \\in T_A(N)" + }, + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "text", + "content": " there exists " + }, + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "inline_equation", + "content": "Y \\in T_A(N)" + }, + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "inline_equation", + "content": "g(Y) = \\frac{1}{k} \\sum_{i=1}^{k} g(X^{(i)})" + }, + { + "bbox": [ + 305, + 368, + 545, + 431 + ], + "type": "text", + "content": ". By Parseval's Theorem," + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 339, + 437, + 545, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 437, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 339, + 437, + 545, + 472 + ], + "type": "interline_equation", + "content": "\\| \\mathcal {F} (X) \\| _ {F} ^ {2} = N ^ {2} \\| X \\| _ {F} ^ {2} = N ^ {2} \\sum_ {i = 0} ^ {N} \\sum_ {j = 0} ^ {N} A _ {i, j} \\tag {8}", + "image_path": "47677be1cfd83be70e1ec61f599f13f6e86c7b84f49f3540fff045a3cefede59.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 479, + 382, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 382, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 382, + 489 + ], + "type": "text", + "content": "so the condition is" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 495, + 545, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 495, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 317, + 495, + 545, + 527 + ], + "type": "interline_equation", + "content": "| \\mathcal {F} (Y) | \\odot | \\mathcal {F} (Y) | = \\frac {1}{k} \\sum_ {i = 1} ^ {k} | \\mathcal {F} \\left(X ^ {(i)}\\right) | \\odot | \\mathcal {F} \\left(X ^ {(i)}\\right) | \\tag {9}", + "image_path": "ac28217131f45c0197d64afd9d650bb16f4bb97cd9023f72ff47843b8c75ab67.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 534, + 369, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 534, + 369, + 546 + ], + "spans": [ + { + "bbox": [ + 306, + 534, + 369, + 546 + ], + "type": "text", + "content": "or equivalently" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 324, + 552, + 545, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 552, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 324, + 552, + 545, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {F} (Y) \\odot \\overline {{\\mathcal {F} (Y)}} = \\frac {1}{k} \\sum_ {i = 1} ^ {k} \\mathcal {F} \\left(X ^ {(i)}\\right) \\odot \\overline {{\\mathcal {F} \\left(X ^ {(i)}\\right)}}. \\tag {10}", + "image_path": "e28f26d66a3cce2776ba292146b6f877a7f27fe1ecca0a74f101c18ae8fedabe.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 592, + 500, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 592, + 500, + 602 + ], + "spans": [ + { + "bbox": [ + 306, + 592, + 500, + 602 + ], + "type": "text", + "content": "Then the cross-correlation theorem reduces it to" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 352, + 609, + 545, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 609, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 352, + 609, + 545, + 641 + ], + "type": "interline_equation", + "content": "\\mathcal {F} (Y \\star Y) = \\frac {1}{k} \\sum_ {i = 1} ^ {k} \\mathcal {F} \\left(X ^ {(i)} \\star X ^ {(i)}\\right) \\tag {11}", + "image_path": "ef2fd75b0e91320d84ffd66029f91026228d14cad0ebff04804db1c639218286.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 647, + 545, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 545, + 671 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 647, + 545, + 671 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 306, + 647, + 545, + 671 + ], + "type": "text", + "content": " denotes cross-correlation. Because the Fourier Transform is linear we finally have" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 367, + 678, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 678, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 367, + 678, + 545, + 711 + ], + "type": "interline_equation", + "content": "Y \\star Y = \\frac {1}{k} \\sum_ {i = 1} ^ {k} X ^ {(i)} \\star X ^ {(i)}. \\tag {12}", + "image_path": "bf23933716144dc2e0118a861888d15025e9628bb15f888429a376b751d647ea.jpg" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10659" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 70, + 248, + 208 + ], + "blocks": [ + { + "bbox": [ + 91, + 70, + 248, + 208 + ], + "lines": [ + { + "bbox": [ + 91, + 70, + 248, + 208 + ], + "spans": [ + { + "bbox": [ + 91, + 70, + 248, + 208 + ], + "type": "image", + "image_path": "7ea44b005c5551a31dc3c8080f2c86574e5447acf106b291a30d55c6fc17f078.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 210, + 287, + 244 + ], + "lines": [ + { + "bbox": [ + 47, + 210, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 47, + 210, + 287, + 244 + ], + "type": "text", + "content": "Figure 3. Geometric interpretation of correlation " + }, + { + "bbox": [ + 47, + 210, + 287, + 244 + ], + "type": "inline_equation", + "content": "(\\mathbf{X} \\star \\mathbf{X})_{\\mathbf{v} - \\mathbf{u}}" + }, + { + "bbox": [ + 47, + 210, + 287, + 244 + ], + "type": "text", + "content": ". The figure represents the correlation step when the shift is " + }, + { + "bbox": [ + 47, + 210, + 287, + 244 + ], + "type": "inline_equation", + "content": "v - u" + }, + { + "bbox": [ + 47, + 210, + 287, + 244 + ], + "type": "text", + "content": ". Notice that only " + }, + { + "bbox": [ + 47, + 210, + 287, + 244 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 47, + 210, + 287, + 244 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 210, + 287, + 244 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 210, + 287, + 244 + ], + "type": "text", + "content": " overlap once the shift is applied." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 262, + 286, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 286, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 286, + 309 + ], + "type": "text", + "content": "Therefore, the convexity of the range of " + }, + { + "bbox": [ + 47, + 262, + 286, + 309 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 47, + 262, + 286, + 309 + ], + "type": "text", + "content": " is equivalent to the convexity of the set " + }, + { + "bbox": [ + 47, + 262, + 286, + 309 + ], + "type": "inline_equation", + "content": "\\{X \\star X : X \\in T_A(N)\\}" + }, + { + "bbox": [ + 47, + 262, + 286, + 309 + ], + "type": "text", + "content": ". We will show the set's projection onto a particular coordinate is not convex." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 316, + 287, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 316, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 92, + 316, + 287, + 349 + ], + "type": "interline_equation", + "content": "(X \\star X) _ {s, r} = \\sum_ {i = 0} ^ {N} \\sum_ {j = 0} ^ {N} X _ {i, j} \\overline {{X _ {i + s , j + r}}} \\tag {13}", + "image_path": "740f3307959277c7a2cab27955296b7ef56f864b3e9d7488069316ba1490d0cb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": "where we adopt the convention that " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "X_{s,r} = 0" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "s, r > N" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "s, r < 0" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": ". Take the points " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " from the definition of " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " (1). Also observe that correlation can be represented geometrically as shifting " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\overline{X}" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": ". In this representation, notice that as the shift " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "(s, r)" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " approaches " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "v - u" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": ", the non-zero overlap between " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\overline{X}" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " shifted by " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "(s, r)" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " approaches 1 by construction. That is, when " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " is shifted to overlap " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " will be the only non-zero overlaps between the shifted and original non-zero points (Figure 3). No other non-zero points can overlap above or below " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " by definition of " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": ". Therefore, " + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "inline_equation", + "content": "(X \\star X)_{v-u}" + }, + { + "bbox": [ + 47, + 356, + 287, + 487 + ], + "type": "text", + "content": " becomes" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 495, + 287, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 495, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 130, + 495, + 287, + 529 + ], + "type": "interline_equation", + "content": "X _ {u} \\bar {X _ {v}} + \\sum_ {i = 1} ^ {N ^ {2} - 1} 0. \\tag {14}", + "image_path": "bba8f2053c219dbdf80e2765ff414b7215514d10b3c407ef9d846645016d4617.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": "Because " + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "inline_equation", + "content": "X_{u}\\overline{X_{v}}\\in \\mathbb{T}" + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "inline_equation", + "content": "(X\\star X)_{v - u}\\in \\mathbb{T}" + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": " which is a nonconvex set. Therefore, the set of correlation's of values on the complex unit circle masked by " + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": " is also not convex, and so is " + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "inline_equation", + "content": "PSF" + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 594, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 641 + ], + "type": "text", + "content": "Time-averaged PSFs span the convex hull of the set of static-mask PSFs, meaning there exists some PSFs achievable only through intensity averaging PSFs from a sequence of phase masks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 651, + 227, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 227, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 227, + 664 + ], + "type": "text", + "content": "5. Multi-Phase Mask Optimization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 181, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 181, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 181, + 683 + ], + "type": "text", + "content": "5.1. Optical Forward Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Similar to PhaseCam3D [32], we model light propagation using Fourier optics theory [8]. In contrast to previous" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "work, we compute the forward model (1) for multiple phase masks, producing a stack of output images, which form our coded image when averaged. This coded image simulates the recorded signal from imaging a scene using a sequence of phase masks in a single exposure (Figure 4)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 138, + 429, + 151 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 138, + 429, + 151 + ], + "spans": [ + { + "bbox": [ + 306, + 138, + 429, + 151 + ], + "type": "text", + "content": "5.2. Specialized Networks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 157, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 157, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 305, + 157, + 545, + 253 + ], + "type": "text", + "content": "For the monocular depth estimation task, we use the MiDaS Small network [22]. This is a well known convolutional monocular depth estimation network designed to take in natural images and output relative depth maps. The network is trained end-to-end with the phase masks. A mean-squared error (MSE) loss term is defined in terms of the depth reconstruction prediction, " + }, + { + "bbox": [ + 305, + 157, + 545, + 253 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 305, + 157, + 545, + 253 + ], + "type": "text", + "content": " and the ground truth depth map " + }, + { + "bbox": [ + 305, + 157, + 545, + 253 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 157, + 545, + 253 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 375, + 259, + 545, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 259, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 375, + 259, + 545, + 281 + ], + "type": "interline_equation", + "content": "L _ {D e p t h} = \\frac {1}{N} \\| D - \\hat {D} \\| _ {2} ^ {2} \\tag {15}", + "image_path": "e26178a16f1d7eaa2c29541b41b361cc4cb0aa07f09271029d1eb371d92765da.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 288, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 288, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 545, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 288, + 545, + 323 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 288, + 545, + 323 + ], + "type": "text", + "content": " is the number of pixels. This process allows for the simultaneous optimization of the phase masks as well as fine tuning MiDaS to reconstruct from our coded images." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 324, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 383 + ], + "type": "text", + "content": "For the extended depth-of-field task, we use an Attention U-Net [18] to reconstruct all-in-focus images. The network is optimized jointly with the phase mask sequence. To learn a reconstruction " + }, + { + "bbox": [ + 305, + 324, + 545, + 383 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 305, + 324, + 545, + 383 + ], + "type": "text", + "content": " to be similar to the all-in-focus ground truth image " + }, + { + "bbox": [ + 305, + 324, + 545, + 383 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 305, + 324, + 545, + 383 + ], + "type": "text", + "content": ", we define the loss term using MSE error" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 381, + 389, + 545, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 389, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 381, + 389, + 545, + 412 + ], + "type": "interline_equation", + "content": "L _ {A i F} = \\frac {1}{N} \\| I - \\hat {I} \\| _ {2} ^ {2} \\tag {16}", + "image_path": "b303f4ba8072d6b9e89f0c88fb619c47bbde56d9e069c16778545f25c1c26d4b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 418, + 440, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 440, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 440, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 418, + 440, + 430 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 418, + 440, + 430 + ], + "type": "text", + "content": " is the number of pixels." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 437, + 441, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 441, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 441, + 449 + ], + "type": "text", + "content": "5.3. Joint Task Optimization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 456, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 456, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 305, + 456, + 545, + 550 + ], + "type": "text", + "content": "We also present an alternative to the specialized networks: a single network jointly trained for monocular depth estimation and extended depth-of-field using a sequence of phase masks. This network has a basic Attention U-Net architecture outputting 4 channels representing depth maps as well as all-in-focus images. Similar to prior works, we use a combined loss function, adding a coefficient to weight the losses for each individual task:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 347, + 559, + 545, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 559, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 347, + 559, + 545, + 571 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = \\lambda_ {D e p t h} L _ {D e p t h} + \\lambda_ {A i F} L _ {A i F}. \\tag {17}", + "image_path": "f272daf2515b8449b67135c5ec96909d24e9f17c88d370ca3573397804d5d49b.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 579, + 429, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 579, + 429, + 593 + ], + "spans": [ + { + "bbox": [ + 306, + 579, + 429, + 593 + ], + "type": "text", + "content": "6. Experimental Details" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 599, + 404, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 404, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 404, + 612 + ], + "type": "text", + "content": "6.1. Training Details" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 617, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 545, + 677 + ], + "type": "text", + "content": "We use the FlyingThings3D from Scene Flow Datasets [15], which uses synthetic data generation to obtain all-in-focus RGB images and disparity maps. We use the cropped " + }, + { + "bbox": [ + 305, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "278 \\times 278" + }, + { + "bbox": [ + 305, + 617, + 545, + 677 + ], + "type": "text", + "content": " all-in-focus images from [32]. In total, we use 5077 training patches and 419 test patches." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Both the optical layer and reconstruction networks are differentiable, so the phase mask sequence and neural network can be optimized through back-propagation. Each" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10660" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 71, + 521, + 180 + ], + "blocks": [ + { + "bbox": [ + 75, + 71, + 521, + 180 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 521, + 180 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 521, + 180 + ], + "type": "image", + "image_path": "d822209196e8b2ba69d77c5c5ca65ccf34ceffe3a5a271695078321606299f63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 194, + 545, + 250 + ], + "lines": [ + { + "bbox": [ + 45, + 194, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 45, + 194, + 545, + 250 + ], + "type": "text", + "content": "Figure 4. Multi-phase mask forward model overview. A sequence of phase masks are used to generate a sequence of depth-dependent PSFs. These PSFs are convolved with depth masked clean images to simulate depth dependent convolution. The images produced by each phase mask are averaged to create a coded image which is fed into an attention U-Net. The reconstruction loss is back-propagated end-to-end through the network and the optical model to design phase masks and algorithms capable of performing monocular depth estimation and extended depth-of-field simultaneously." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 253, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 288, + 326 + ], + "type": "text", + "content": "part is implemented in PyTorch. During training, we use the Adam [11] optimizer with parameters " + }, + { + "bbox": [ + 46, + 253, + 288, + 326 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.99" + }, + { + "bbox": [ + 46, + 253, + 288, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 253, + 288, + 326 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 46, + 253, + 288, + 326 + ], + "type": "text", + "content": ". The learning rate for the phase masks is " + }, + { + "bbox": [ + 46, + 253, + 288, + 326 + ], + "type": "inline_equation", + "content": "10^{-8}" + }, + { + "bbox": [ + 46, + 253, + 288, + 326 + ], + "type": "text", + "content": " and for the reconstruction network it is " + }, + { + "bbox": [ + 46, + 253, + 288, + 326 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 253, + 288, + 326 + ], + "type": "text", + "content": ", and the batch size was 32. Finally, training and testing were performed on NVIDIA Quadro P6000 GPUs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "text", + "content": "We parameterize " + }, + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "inline_equation", + "content": "23 \\times 23" + }, + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "text", + "content": " phase masks pixel-wise as [13] found pixel-wise parameterization to produce the best overall performance. The monocular depth estimation task uses the MiDaS Small architecture pretrained weights for monocular depth estimation downloadable from PyTorch [22]. The extended depth-of-field task pretrains an Attention U-Net with a fixed Fresnel lens for 300 epochs. For the joint task, we set " + }, + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\lambda_{Depth} = \\lambda_{AiF} = 1" + }, + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "text", + "content": " to balance overall performance, and we pretrain the Attention U-Net for 300 epochs with a fixed Fresnel lens. In simulation, the red, blue, and green channels are approximated by discretized wavelengths, " + }, + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "inline_equation", + "content": "610~\\mathrm{nm}" + }, + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "inline_equation", + "content": "530~\\mathrm{nm}" + }, + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "inline_equation", + "content": "470~\\mathrm{nm}" + }, + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "text", + "content": " respectively. Additionally, the depth range is discretized into 21 bins on the interval " + }, + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "inline_equation", + "content": "[-20, 20]" + }, + { + "bbox": [ + 45, + 328, + 287, + 496 + ], + "type": "text", + "content": ", which is larger than previous works." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 513, + 155, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 155, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 155, + 525 + ], + "type": "text", + "content": "6.2. Evaluation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": "For ablation studies on our method, we used the testing split of the FlyingThings3D set for both monocular depth estimation and extended depth-of-field imaging [15]. For comparisons to existing work, we also tested our monocular depth estimation network on the labeled NYU Depth v2 set [17]. The ground truth depth maps were translated to layered masks for the clean images by bucketing the depth values into 21 bins, allowing us to convolve each depth in an image with the required PSF. We use root mean squared error (RMSE) between ground truth and estimated depth maps for depth estimation and RMSE between ground truth and reconstructed all-in-focus images for extended depth-of-field imaging. We also use peak signal-to-noise ratio (PSNR) and structural similarity index [30] (SSIM) for extended depth-of-field imaging." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 252, + 358, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 252, + 358, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 252, + 358, + 264 + ], + "type": "text", + "content": "7. Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 272, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 272, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 305, + 272, + 545, + 321 + ], + "type": "text", + "content": "We compare our time averaged dynamic PSF method to the state-of-the-art methods for both extended depth-of-field imaging and monocular depth estimation. The relevant works we compare to are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 328, + 545, + 681 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 313, + 328, + 545, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 545, + 377 + ], + "type": "text", + "content": "1. PhaseCam3D [32] used a " + }, + { + "bbox": [ + 313, + 328, + 545, + 377 + ], + "type": "inline_equation", + "content": "23 \\times 23" + }, + { + "bbox": [ + 313, + 328, + 545, + 377 + ], + "type": "text", + "content": " phase mask based on 55 Zernike coefficients. The phase mask parameters were then end-to-end optimized with a U-Net reconstruction network to perform depth estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 384, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 545, + 433 + ], + "type": "text", + "content": "2. Chang et al. [3] used a singlet lens introducing chromatic aberrations with radially symmetric PSFs. Similar to [32], the lens parameters were also then end-to-end optimized." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 441, + 545, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 545, + 523 + ], + "type": "text", + "content": "3. Ikoma et al. [10] used a radially symmetric diffractive optical element (DOE). The blurred image was preconditioned with an approximate inverse of the PSF depth dependent blur. The RGB image stack was fed into a U-Net to produce both an all-in-focus image and a depth map. The DOE and U-Net parameters were optimized in an end-to-end fashion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 545, + 581 + ], + "type": "text", + "content": "4. Liu et al. [13] used various phase mask parameterizations with the same U-Net architecture as [10]. One method used pixel-wise height maps (PW) and the other introduced orbital angular momentum (OAM)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 589, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 545, + 637 + ], + "type": "text", + "content": "5. Sitzmann et al. [26] implements a single DOE based on Zernike coefficients, and solves the Tikhonov-regularized least-squares problem to reconstruct an all-in-focus image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 645, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 545, + 681 + ], + "type": "text", + "content": "6. MiDaS [21] and ZoeDepth [2] are state of the art single shot monocular depth estimation methods with all-in-focus images as inputs." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Because both [10] and [13] simultaneously learn all-in-focus images and depth maps, when comparing against our" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10661" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 272, + 171 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 272, + 171 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 272, + 171 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 272, + 171 + ], + "type": "table", + "html": "
MethodFlyingThings3DNYUv2
PhaseCam3D [32]0.5210.382
Chang et al. [3]0.4900.433
Ikoma et al. [10]0.184-
MiDaS [21]-0.357
ZoeDepth [2]-0.277
TiDy (1)0.0260.259
TiDy (5)0.0190.175
", + "image_path": "04b1c7ada2d7d33f401511f06deb7988e985c073d008eb4abd9a34fbc6c1d180.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 252, + 277, + 331 + ], + "blocks": [ + { + "bbox": [ + 46, + 179, + 287, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 179, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 287, + 236 + ], + "type": "text", + "content": "Table 1. RMSE comparison of monocular depth estimation methods. We present quantitative results on two datasets to compare to state of the art optical and single shot monocular depth estimation methods. Our method performs best with our 5 phase mask system achieving the lowest error on both datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 252, + 277, + 331 + ], + "lines": [ + { + "bbox": [ + 57, + 252, + 277, + 331 + ], + "spans": [ + { + "bbox": [ + 57, + 252, + 277, + 331 + ], + "type": "table", + "html": "
MethodRMSE↓PSNR↑SSIM↑
Liu et al. [13]-29.80-
Ikoma et al. [10]0.132731.880.905
Sitzmann et al. [26]-32.44-
TiDy (1)0.014837.330.968
TiDy (5)0.009241.110.989
", + "image_path": "e44d0b58183f698292444d785603f279ca67f86e75c55e688af7655c1eff69ff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 411, + 286, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 286, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 286, + 436 + ], + "type": "text", + "content": "specialized methods, we take their best performing weighting of each task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 593 + ], + "type": "text", + "content": "Individual Tasks. For monocular depth estimation, our specialized method using a sequence of 5 phase masks trained for 300 epochs outperforms prior work on FlyingThings3D (Table 1). Additionally, our approach performs significantly better and achieves lower error than previous methods on NYUv2 without any additional fine tuning. For extended depth-of-field, our specialized method using a sequence of 5 phase masks outperforms prior work on FlyingThings3D (Table 2). This demonstrates the benefit of multi-phase mask learning on computational imaging tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 286, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 715 + ], + "type": "text", + "content": "Multi-Objective Optimization. We also evaluate our method against other joint all-in-focus and depth map learning approaches. This problem is challenging because good depth cues to produce depth maps is antithetical to producing an all-in-focus image. Our combined 5 phase mask trained for 300 epochs approach outperforms prior jointly trained approaches (Table 3)." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 322, + 70, + 531, + 159 + ], + "blocks": [ + { + "bbox": [ + 46, + 338, + 287, + 383 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 383 + ], + "type": "text", + "content": "Table 2. Comparison of extended depth-of-field imaging methods. We present quantitative results on FlyingThings3D to compare to state-of-the-art. Our method performs best with our 5 phase mask system achieving the best PSNR." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 70, + 531, + 159 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 531, + 159 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 531, + 159 + ], + "type": "table", + "html": "
MethodAll-in-focus PSNR↑Depth RMSE↓
Ikoma et al. [10]31.880.191
Liu et al. [13] - PW29.800.056
Liu et al. [13] - OAMt25.860.053
TiDy (1)31.200.052
TiDy (5)34.790.034
", + "image_path": "a2b151d2cbf234f1609cda4a6ad48e24681c41ab9b0df7b1cefa8160b17580ae.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 167, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 305, + 167, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 545, + 223 + ], + "type": "text", + "content": "Table 3. Comparison of multi-objective optimization of extended depth-of-field imaging and depth estimation methods. We compare quantitative results on FlyingThings3D to the state-of-the-art. Our method performs best with our 5 phase mask system achieving the best balance between objectives." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 335, + 249, + 517, + 375 + ], + "blocks": [ + { + "bbox": [ + 335, + 249, + 517, + 375 + ], + "lines": [ + { + "bbox": [ + 335, + 249, + 517, + 375 + ], + "spans": [ + { + "bbox": [ + 335, + 249, + 517, + 375 + ], + "type": "image", + "image_path": "d3dec9c3c45e64e15eab15de2277d24bd6d465a930b012baf5d35fe175bbbacc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 392, + 545, + 448 + ], + "lines": [ + { + "bbox": [ + 305, + 392, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 545, + 448 + ], + "type": "text", + "content": "Figure 5. RMSE for specialized tasks for each phase mask sequence length. RMSE decreases with respect to phase mask sequence length for both specialized extended depth-of-field imaging and monocular depth estimation tasks. 0 phase masks refers to a reconstruction neural network with a fixed Fresnel lens." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 335, + 472, + 516, + 597 + ], + "blocks": [ + { + "bbox": [ + 335, + 472, + 516, + 597 + ], + "lines": [ + { + "bbox": [ + 335, + 472, + 516, + 597 + ], + "spans": [ + { + "bbox": [ + 335, + 472, + 516, + 597 + ], + "type": "image", + "image_path": "4f261afc545da41913c04fdd4882016163dff55fc2d92c59376d68c71eafe23b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 616, + 545, + 681 + ], + "lines": [ + { + "bbox": [ + 305, + 616, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 305, + 616, + 545, + 681 + ], + "type": "text", + "content": "Figure 6. RMSE for joint optimization of monocular depth estimation and extended depth-of-field imaging for each phase mask sequence length. RMSE decreases with respect to phase mask sequence length for this complex joint task, demonstrating the benefit of multi-phase mask learning. 0 phase masks refers to a reconstruction neural network with a fixed Fresnel lens." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10662" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 76, + 272, + 201 + ], + "blocks": [ + { + "bbox": [ + 66, + 76, + 272, + 201 + ], + "lines": [ + { + "bbox": [ + 66, + 76, + 272, + 201 + ], + "spans": [ + { + "bbox": [ + 66, + 76, + 272, + 201 + ], + "type": "image", + "image_path": "4127ab4795135aec3fcf921155eeb96c0baf098ed5a0911b4629988756e632a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 289, + 271 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 271 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 271 + ], + "type": "text", + "content": "Figure 7. All-in-focus imaging RMSE distribution for each phase mask length without a reconstruction network. The best RMSE for each phase mask count has low correlation with respect to phase mask sequence length, but the variance of RMSE decreases." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 289, + 149, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 149, + 302 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 149, + 302 + ], + "type": "text", + "content": "8. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 308, + 253, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 253, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 253, + 323 + ], + "type": "text", + "content": "8.1. Effect of Phase Mask Sequence Length" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 327, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 458 + ], + "type": "text", + "content": "For both all-in-focus imaging and depth estimation, we vary the phase mask count that the end-to-end system is trained with to gauge the benefits of using multiple phase masks. The forward model and initial phase masks were held standard while the phase mask count was varied. The resulting networks were evaluated at convergence. For the extended depth-of-field task, the masks were all initialized with random noise uniform from 0 to " + }, + { + "bbox": [ + 46, + 327, + 287, + 458 + ], + "type": "inline_equation", + "content": "1.2 \\times 10^{-6}" + }, + { + "bbox": [ + 46, + 327, + 287, + 458 + ], + "type": "text", + "content": ". For the depth estimation task, the masks were initialized with the Fisher mask with added Gaussian noise parameterized by a " + }, + { + "bbox": [ + 46, + 327, + 287, + 458 + ], + "type": "inline_equation", + "content": "5.35 \\times 10^{-7}" + }, + { + "bbox": [ + 46, + 327, + 287, + 458 + ], + "type": "text", + "content": " mean and " + }, + { + "bbox": [ + 46, + 327, + 287, + 458 + ], + "type": "inline_equation", + "content": "3.05 \\times 10^{-7}" + }, + { + "bbox": [ + 46, + 327, + 287, + 458 + ], + "type": "text", + "content": " standard deviation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 556 + ], + "type": "text", + "content": "End-to-end optimization on each task with a specialized network yielded improved performance as the phase mask count increased, visualized in Figure 5. This result implies that sequences of phase masks are successful in making the PSF space more expressive. Additionally, even for the more complex joint task, learning a system that can produce both all-in-focus images and depth maps, error decreases with phase mask count until a plateau is reached (Figure 6)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 563, + 287, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 287, + 576 + ], + "type": "text", + "content": "8.2. All-in-Focus without Reconstruction Networks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "A phase mask generating a PSF of the unit impulse function at every depth would be ideal for extended depth-of-field as each depth would be in focus. If possible, this phase mask would not require any digital processing. We optimize phase mask sequences of varying lengths to produce an averaged PSF close to the unit impulse function for all depths. For each sequence length, phase masks are optimized using MSE loss between the unit impulse function and the averaged PSF at each depth until convergence. We ran 1000 trials of random phase mask initialization for each length. Observe that a side-effect of longer phase masks is training" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 337, + 74, + 517, + 254 + ], + "blocks": [ + { + "bbox": [ + 337, + 74, + 517, + 254 + ], + "lines": [ + { + "bbox": [ + 337, + 74, + 517, + 254 + ], + "spans": [ + { + "bbox": [ + 337, + 74, + 517, + 254 + ], + "type": "image", + "image_path": "49323f82872848c595d84c3a8a46b5a1c987e376054c3cedba130284ae30de34.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 273, + 545, + 296 + ], + "lines": [ + { + "bbox": [ + 306, + 273, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 273, + 545, + 296 + ], + "type": "text", + "content": "Figure 8. Visualization of phase mask initializations. Each row represents a different initial phase mask sequence." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 315, + 546, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 315, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 305, + 315, + 546, + 363 + ], + "type": "text", + "content": "stability. The range of RMSE between the simulated capture image and ground truth all-in-focus image decreases as the sequence length increases (Figure 7). This indicates training longer sequences is more resilient to initialization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 370, + 545, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 545, + 384 + ], + "type": "text", + "content": "8.3. Phase Mask Initialization for Depth Perception" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 388, + 546, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 546, + 604 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 546, + 604 + ], + "type": "text", + "content": "Deep optics for depth perception can be very dependent on the initialization of optical parameters before training [32]. To find the extent of the effect of mask initialization on performance, we varied the the initial phase masks while keeping number of masks, the optical model, and duration of training fixed. We trained for 200 epochs. We tested four initializations of sequences of 5 phase masks as shown in Figure 8. The first was uniformly distributed noise from 0 to " + }, + { + "bbox": [ + 305, + 388, + 546, + 604 + ], + "type": "inline_equation", + "content": "1.2 \\times 10^{-6}" + }, + { + "bbox": [ + 305, + 388, + 546, + 604 + ], + "type": "text", + "content": ". The second was the first mask in the sequence set to a Fisher mask while the rest are uniform noise. The third is setting each mask to a rotation of the Fisher mask and adding Gaussian noise parameterized by a " + }, + { + "bbox": [ + 305, + 388, + 546, + 604 + ], + "type": "inline_equation", + "content": "5.35 \\times 10^{-7}" + }, + { + "bbox": [ + 305, + 388, + 546, + 604 + ], + "type": "text", + "content": " mean and " + }, + { + "bbox": [ + 305, + 388, + 546, + 604 + ], + "type": "inline_equation", + "content": "3.05 \\times 10^{-7}" + }, + { + "bbox": [ + 305, + 388, + 546, + 604 + ], + "type": "text", + "content": " standard deviation to 4 masks. Lastly, we set each mask to a rotation of the Fisher mask and added noise to only the last two masks in the sequence. Of the four initializations, it is clear that the 3 Fisher masks and 2 Fisher masks with noise performed the best (Table 4)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 611, + 468, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 468, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 468, + 624 + ], + "type": "text", + "content": "8.4. Modeling SLM Imperfections" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 715 + ], + "type": "text", + "content": "State Switching. Our optical forward model assumes an SLM can swap between two phase patterns instantly. In practice, however, some light will be captured during the intermediate states between phase patterns. These phase patterns, in the worst case, could be random phase patterns, effectively adding noise to our coded images. We model these intermediate states by averaging output images pro" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10663" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 70, + 250, + 135 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 250, + 135 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 250, + 135 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 250, + 135 + ], + "type": "table", + "html": "
InitializationRMSE↓
1 Fisher + All noise0.0329
1 Fisher + Fisher w/ Noise0.0271
All noise0.0254
3 Fisher + Fisher w/ Noise0.0207
", + "image_path": "89e839f6d49bc7b31543d56c5cb4bb700ece369b602fc33d41348329c4a0641d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 143, + 288, + 187 + ], + "lines": [ + { + "bbox": [ + 47, + 143, + 288, + 187 + ], + "spans": [ + { + "bbox": [ + 47, + 143, + 288, + 187 + ], + "type": "text", + "content": "Table 4. Quantitative evaluation of phase mask initializations. Four sequence initializations are evaluated on the monocular depth estimation task. Ultimately, 3 Fisher masks and 2 noisy Fisher masks have the best performance after training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 76, + 201, + 259, + 456 + ], + "blocks": [ + { + "bbox": [ + 76, + 201, + 259, + 456 + ], + "lines": [ + { + "bbox": [ + 76, + 201, + 259, + 456 + ], + "spans": [ + { + "bbox": [ + 76, + 201, + 259, + 456 + ], + "type": "image", + "image_path": "04900e05ca70c5b7d5c4adf067137e62e360c1650799f3189c36cb2cdefa3be9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 471, + 287, + 537 + ], + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 537 + ], + "type": "text", + "content": "Figure 9. Qualitative results of a specialized network on extended depth-of-field imaging. Both 1 and 5 phase mask systems are evaluated on FlyingThings3D. Error is computed pixel wise between the ground truth all-in-focus image and the reconstructed output and is boosted by a factor of 3. Notice that the 5 phase mask system introduces minimal error." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 558, + 288, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 678 + ], + "type": "text", + "content": "duced by phase masks and the randomized phase patterns weighted by the time that they are displayed for. We model the total exposure time as " + }, + { + "bbox": [ + 46, + 558, + 288, + 678 + ], + "type": "inline_equation", + "content": "100\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 558, + 288, + 678 + ], + "type": "text", + "content": ", with various durations of switching times from 1 to " + }, + { + "bbox": [ + 46, + 558, + 288, + 678 + ], + "type": "inline_equation", + "content": "16\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 558, + 288, + 678 + ], + "type": "text", + "content": " per swap. We evaluate our joint optimized network on these new, more noisy, coded images without any additional training (Figure 12). Observe that because the 5 phase mask system includes more swaps, performance degrades faster than fewer phase mask systems. However, for short switching times, 5 phase masks still outperform the others without needing any fine tuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Quantization. Current high-speed MEMS based SLM technology experience heavy quantization [1]. We model" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 322, + 71, + 532, + 295 + ], + "blocks": [ + { + "bbox": [ + 322, + 71, + 532, + 295 + ], + "lines": [ + { + "bbox": [ + 322, + 71, + 532, + 295 + ], + "spans": [ + { + "bbox": [ + 322, + 71, + 532, + 295 + ], + "type": "image", + "image_path": "fbd13953e34a1d415320526170cd157d1f065042d76dfb0a66e4f27a2e4f469d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 312, + 545, + 346 + ], + "lines": [ + { + "bbox": [ + 306, + 312, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 545, + 346 + ], + "type": "text", + "content": "Figure 10. Qualitative results of a specialized networks on monocular depth estimation. Performance using the five phase mask method outperforms one phase mask on both datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 332, + 363, + 512, + 594 + ], + "blocks": [ + { + "bbox": [ + 332, + 363, + 512, + 594 + ], + "lines": [ + { + "bbox": [ + 332, + 363, + 512, + 594 + ], + "spans": [ + { + "bbox": [ + 332, + 363, + 512, + 594 + ], + "type": "image", + "image_path": "329bff1a5117f54e4ae67c656179621f630f238f55f3c15612a2bc8a63bfaf2f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 609, + 545, + 665 + ], + "lines": [ + { + "bbox": [ + 305, + 609, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 545, + 665 + ], + "type": "text", + "content": "Figure 11. Qualitative results of a joint optimized system for extended depth-of-field imagining and monocular depth estimation. Both one and five phase mask networks are evaluated on the FlyingThings3D datasets. Notice that five masks has fewer artifacts than a single mask." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": "the effects of quantization error by adding varying amounts of noise during inference. As demonstrated in Figure 13," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10664" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 72, + 272, + 190 + ], + "blocks": [ + { + "bbox": [ + 64, + 72, + 272, + 190 + ], + "lines": [ + { + "bbox": [ + 64, + 72, + 272, + 190 + ], + "spans": [ + { + "bbox": [ + 64, + 72, + 272, + 190 + ], + "type": "image", + "image_path": "73b4aa753f6b63d6c7513ad1788aa7affc98808f340aba7129a338178d155d19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 287, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 282 + ], + "type": "text", + "content": "Figure 12. Effect of switching time on joint system performance. Reconstruction error across phase mask counts as a function of switching time with " + }, + { + "bbox": [ + 46, + 205, + 287, + 282 + ], + "type": "inline_equation", + "content": "100\\mathrm{ms}" + }, + { + "bbox": [ + 46, + 205, + 287, + 282 + ], + "type": "text", + "content": " overall exposure. Performance of the jointly optimized system degrades as the switching time between phase masks increases, as expected. Our system still performs well when the time spent switching is less than " + }, + { + "bbox": [ + 46, + 205, + 287, + 282 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 205, + 287, + 282 + ], + "type": "text", + "content": " of the overall exposure." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 297, + 272, + 399 + ], + "blocks": [ + { + "bbox": [ + 63, + 297, + 272, + 399 + ], + "lines": [ + { + "bbox": [ + 63, + 297, + 272, + 399 + ], + "spans": [ + { + "bbox": [ + 63, + 297, + 272, + 399 + ], + "type": "image", + "image_path": "dde2d180287f3dcf84e733d47386b05c20afb33db3f39c2485d97dcc2c9c7242.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 413, + 286, + 458 + ], + "lines": [ + { + "bbox": [ + 46, + 413, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 286, + 458 + ], + "type": "text", + "content": "Figure 13. Effect of quantization on joint system performance. Performance of the jointly optimized system degrades as the number of effective bits decrease; however, even under heavy quantization, our system outperforms single mask setups." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 479, + 286, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 286, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 286, + 528 + ], + "type": "text", + "content": "time-averaged PSFs are robust to quantization noise: 5 masks quantized to 4-bits outperform one mask quantized to 32-bits. This is intuitive as multiple masks can leverage averaging to achieve better precision and remove noise [4]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 537, + 277, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 277, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 277, + 550 + ], + "type": "text", + "content": "8.5. A Path Towards Better Single-Mask Designs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 557, + 286, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 557, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 557, + 286, + 604 + ], + "type": "text", + "content": "The performance of end-to-end designed phase masks is highly sensitive to how they are initialized [32]. Can we use our 5-mask design to provide a better initialization for a 1-mask design?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "Starting from random initializations, we optimized 200 single phase masks to minimize the mean squared error between each of their PSFs and the time-averaged PSF formed by our optimized 5-mask system. We then took the single mask whose PSF was closest to the 5-mask design and fine-tuned our reconstruction network using its PSF. This static design improved performance over a randomly initialized end-to-end trained 1-mask system; +2.67 PSNR for all-in-focus imaging and -0.003 RMSE for depth estima" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "tion. However, its performance was still strictly worse than a 5-mask system." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 103, + 487, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 103, + 487, + 115 + ], + "spans": [ + { + "bbox": [ + 306, + 103, + 487, + 115 + ], + "type": "text", + "content": "8.6. Modulating Amplitude and Phase" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 121, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 264 + ], + "type": "text", + "content": "As illustrated in our proof of Theorem 1, \"complex\" masks which modulate both amplitude and phase are more expressive than phase-only masks. Using end-to-end learning, we optimized a pixel wise complex mask (initialized with a random phase mask and all ones amplitude mask) on the joint depth estimation and all-in-focus imaging task. The complex mask offered substationally improved performance; " + }, + { + "bbox": [ + 304, + 121, + 545, + 264 + ], + "type": "inline_equation", + "content": "+2.49" + }, + { + "bbox": [ + 304, + 121, + 545, + 264 + ], + "type": "text", + "content": " PSNR for all-in-focus imaging and " + }, + { + "bbox": [ + 304, + 121, + 545, + 264 + ], + "type": "inline_equation", + "content": "-0.015" + }, + { + "bbox": [ + 304, + 121, + 545, + 264 + ], + "type": "text", + "content": " RMSE for depth estimation. However, its performance too was still strictly worse than a 5-mask system. We conjecture this is due to improved training stability of the 5-mask system." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 275, + 380, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 275, + 380, + 288 + ], + "spans": [ + { + "bbox": [ + 306, + 275, + 380, + 288 + ], + "type": "text", + "content": "9. Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 295, + 545, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 451 + ], + "type": "text", + "content": "While we were successful in learning dynamic phase masks to improve state-of-the-art performance on imaging tasks, our method carries some limitations. First, our optical model assumes perfect switching between phase masks during training. While evaluation with non-zero switching times showed little degradation of performance, accounting for state switching while training could produce phase masks that are more performant. Our optical model also simulates depths as layered masks over an image, which does not account for blending at depth boundaries. Additionally, our method assumes that scenes are static for the duration of a single exposure. Lastly, though their prices are falling, SLMs are still quite expensive and bulky." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 456, + 384, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 384, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 384, + 468 + ], + "type": "text", + "content": "10. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 476, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 545, + 596 + ], + "type": "text", + "content": "This work is founded upon the insight that the set of PSFs that can be described by a single phase mask is nonconvex and that, as a result, time-averaged PSFs are fundamentally more expressive. We demonstrate that one can learn a sequence of phase masks that, when one dynamically switches among them over time, can substantially improve computational imaging performance across a range of tasks, including depth estimation and all-in-focus imaging. Our work unlocks an exciting new direction for PSF engineering and computational imaging system design." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 606, + 408, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 606, + 408, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 606, + 408, + 620 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 626, + 544, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 626, + 544, + 650 + ], + "spans": [ + { + "bbox": [ + 306, + 626, + 544, + 650 + ], + "type": "text", + "content": "This work was partially supported by AFOSR Young Investigator Program Award FA9550-22-1-0208." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 660, + 365, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 660, + 365, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 660, + 365, + 672 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 310, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 310, + 680, + 545, + 713 + ], + "type": "text", + "content": "[1] Terry A. Bartlett, William C. McDonald, and James N. Hall. Adapting Texas Instruments DLP technology to demonstrate a phase spatial light modulator. In Michael R. Douglass, John" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10665" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "Ehmke, and Benjamin L. Lee, editors, Emerging Digital Micromirror Device Based Systems and Applications XI, volume 10932, page 109320S. International Society for Optics and Photonics, SPIE, 2019. 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 288, + 150 + ], + "type": "text", + "content": "[2] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 151, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 151, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 151, + 288, + 183 + ], + "type": "text", + "content": "[3] Julie Chang and Gordon Wetzstein. Deep optics for monocular depth estimation and 3d object detection. In Proc. IEEE ICCV, 2019. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 184, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 288, + 249 + ], + "type": "text", + "content": "[4] Suyeon Choi, Manu Gopakumar, Yifan Peng, Jonghyun Kim, Matthew O'Toole, and Gordon Wetzstein. Time-multiplexed neural holography: A flexible framework for holographic near-eye displays with fast heavily-quantized spatial light modulators. In Proceedings of the ACM SIGGRAPH, page 1-9, 2022. 2, 9" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 304 + ], + "type": "text", + "content": "[5] Suyeon Choi, Manu Gopakumar, Yifan Peng, Jonghyun Kim, and Gordon Wetzstein. Neural 3d holography: Learning accurate wave propagation models for 3d holographic virtual and augmented reality displays. ACM Trans. Graph., 40(6), December 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 304, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 288, + 337 + ], + "type": "text", + "content": "[6] Edward R. Dowski and W. Thomas Cathey. Extended depth of field through wave-front coding. Appl. Opt., 34(11):1859-1866, April 1995. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 338, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 288, + 381 + ], + "type": "text", + "content": "[7] Robert Fischer, Yicong Wu, Pakorn Kanchanawong, Hari Shroff, and Clare Waterman-Storer. Microscopy in 3d: A biologist's toolbox. Trends in cell biology, 21:682-91, October 2011. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 383, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 383, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 383, + 288, + 403 + ], + "type": "text", + "content": "[8] Joseph W. Goodman. Introduction to fourier optics. Freeman, 2017. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 404, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 288, + 449 + ], + "type": "text", + "content": "[9] Samuel W. Hasinoff and Kiriakos N. Kutulakos. A layer-based restoration framework for variable-aperture photography. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8, 2007. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "type": "text", + "content": "[10] Hayato Ikoma, Cindy M. Nguyen, Christopher A. Metzler, Yifan Peng, and Gordon Wetzstein. Depth from defocus with learned optics for imaging and occlusion-aware depth estimation. IEEE International Conference on Computational Photography (ICCP), 2021. 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 504, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 288, + 525 + ], + "type": "text", + "content": "[11] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 526, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 288, + 581 + ], + "type": "text", + "content": "[12] Anat Levin, Rob Fergus, Frédo Durand, and William T. Freeman. Image and depth from a conventional camera with a coded aperture. In ACM SIGGRAPH '2007 Papers, SIGGRAPH '07, page 70-es, New York, NY, USA, 2007. Association for Computing Machinery. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 582, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 635 + ], + "type": "text", + "content": "[13] Xin Liu, Linpei Li, Xu Liu, Xiang Hao, and Yifan Peng. Investigating deep optics model representation in affecting resolved all-in-focus image quality and depth estimation fidelity. Opt. Express, 30(20):36973-36984, September 2022. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 689 + ], + "type": "text", + "content": "[14] Yawen Lu, Sophia Kourian, Carl Salvaggio, Chenliang Xu, and Guoyu Lu. Single image 3d vehicle pose estimation for augmented reality. In 2019 IEEE Global Conference on Signal and Information Processing (GlobalSIP), pages 1-5, 2019. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[15] Nikolaus Mayer, Eddy Ilg, Philip Häusser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 118 + ], + "type": "text", + "content": "large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016. 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": "[16] Christopher A Metzler, Hayato Ikoma, Yifan Peng, and Gordon Wetzstein. Deep optics for single-shot high-dynamic-range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1375-1385, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[17] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 274 + ], + "type": "text", + "content": "[18] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, Ben Glocker, and Daniel Rueckert. Attention u-net: Learning where to look for the pancreas. In Medical Imaging with Deep Learning, 2018. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[19] Luca Palmieri, Gabriele Scrofani, Nicolò Incardona, Genaro Saavedra, Manuel Martínez-Corral, and Reinhard Koch. Robust depth estimation for light field microscopy. Sensors, 19(3), 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "text", + "content": "[20] Alex Paul Pentland. A new sense for depth of field. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-9(4):523-531, 1987. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 355, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 409 + ], + "type": "text", + "content": "[21] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 411, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 465 + ], + "type": "text", + "content": "[22] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3):1623-1637, 2022. 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "type": "text", + "content": "[23] Anupa Sabnis and Leena Vachhani. Single image based depth estimation for robotic applications. In 2011 IEEE Recent Advances in Intelligent Computational Systems, pages 102-106, 2011. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "text", + "content": "[24] Yoav Shechtman, Steffen J. Sahl, Adam S. Backer, and W. E. Moerner. Optimal point spread function design for 3d imaging. Phys. Rev. Lett., 113:133902, September 2014. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "type": "text", + "content": "[25] Yoav Shechtman, Lucien Weiss, Adam Backer, Steffen Sahl, and William Moerner. Precise three-dimensional scan-free multiple-particle tracking over large axial ranges with tetrapod point spread functions. *Nano letters*, 15, May 2015. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 591, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 656 + ], + "type": "text", + "content": "[26] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Trans. Graph., 37(4), July 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "text", + "content": "[27] Qilin Sun, Ethan Tseng, Qiang Fu, Wolfgang Heidrich, and Felix Heide. Learning rank-1 diffractive optics for single-shot high dynamic range imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1386-1396, 2020. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "10666" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 462 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Ashok Veeraraghavan, Ramesh Raskar, Amit Agrawal, Ankit Mohan, and Jack Tumblin. Dappled photography: Mask enhanced cameras for heterodyned light fields and coded aperture refocusing. ACM Trans. Graph., 26(3):69-es, jul 2007. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 287, + 194 + ], + "type": "text", + "content": "[29] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "type": "text", + "content": "[30] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 283 + ], + "type": "text", + "content": "[31] Woontack Woo, Wonwoo Lee, and Nohyoung Park. Depth-assisted real-time 3d object detection for augmented reality. In International Conference on Artificial Reality and Teleexistence (ICAT), 2011. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 350 + ], + "type": "text", + "content": "[32] Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin Sankaranarayanan, and Ashok Veeraraghavan. Phasecam3d — learning phase masks for passive single view depth estimation. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2019. 1, 2, 3, 4, 5, 6, 7, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 352, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 352, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 352, + 287, + 418 + ], + "type": "text", + "content": "[33] Feng Xue, Guirong Zhuo, Ziyuan Huang, Wufei Fu, Zhuoyue Wu, and Marcelo H Ang. Toward hierarchical self-supervised monocular absolute depth estimation for autonomous driving applications. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2330-2337. IEEE, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 419, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 419, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 48, + 419, + 287, + 462 + ], + "type": "text", + "content": "[34] Menglong Ye, Edward Johns, Ankur Handa, Lin Zhang, Philip Pratt, and Guang-Zhong Yang. Self-supervised siamese learning on stereo image pairs for depth estimation in robotic surgery, 2017. 1" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10667" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Tiled Multiplane Images for Practical 3D Photography/332ad0e6-c530-4dba-82ae-058961745db1_content_list.json b/2023/Tiled Multiplane Images for Practical 3D Photography/332ad0e6-c530-4dba-82ae-058961745db1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b06ec1232b16344f8613acf27e5d74a484deb28d --- /dev/null +++ b/2023/Tiled Multiplane Images for Practical 3D Photography/332ad0e6-c530-4dba-82ae-058961745db1_content_list.json @@ -0,0 +1,1789 @@ +[ + { + "type": "text", + "text": "Tiled Multiplane Images for Practical 3D Photography", + "text_level": 1, + "bbox": [ + 207, + 130, + 759, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Numair Khan", + "bbox": [ + 210, + 181, + 323, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "numairkhan@meta.com", + "bbox": [ + 181, + 202, + 352, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lei Xiao", + "bbox": [ + 444, + 181, + 517, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lei.xiao@meta.com", + "bbox": [ + 403, + 202, + 557, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Douglas Lanman", + "bbox": [ + 629, + 181, + 766, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "doug.lanman@meta.com", + "bbox": [ + 607, + 202, + 787, + 215 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc5342a98343c8a0d9d0a150e3e00703918263f40220310233be39068a3d9b6e.jpg", + "image_caption": [ + "Figure 1: An overview of our 3D photography method. Left: Using a single RGB image and an estimated monocular depth map, the scene is recreated as a tiled grid of many small MPs. Middle: A visualization of a $2 \\times 2$ tiled grid of MPs, each with three RGBA layers. Right: Novel views are rendered by warping and compositing each MPI tile into the target camera's frustum. The $768 \\times 1152$ pixel image shown is generated using a $7 \\times 11$ grid of 4-layer MPs." + ], + "image_footnote": [], + "bbox": [ + 102, + 250, + 870, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 530, + 313, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The task of synthesizing novel views from a single image has useful applications in virtual reality and mobile computing, and a number of approaches to the problem have been proposed in recent years. A Multiplane Image (MPI) estimates the scene as a stack of RGBA layers, and can model complex appearance effects, anti-alias depth errors and synthesize soft edges better than methods that use textured meshes or layered depth images. And unlike neural radiance fields, an MPI can be efficiently rendered on graphics hardware. However, MPIs are highly redundant and require a large number of depth layers to achieve plausible results. Based on the observation that the depth complexity in local image regions is lower than that over the entire image, we split an MPI into many small, tiled regions, each with only a few depth planes. We call this representation a Tiled Multiplane Image (TMPI). We propose a method for generating a TMPI with adaptive depth planes for single-view 3D photography in the wild. Our synthesized results are comparable to state-of-the-art single-view MPI methods while having lower computational overhead.", + "bbox": [ + 75, + 569, + 473, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 530, + 630, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The novel view synthesis (NVS) problem involves using a set of input images to generate views from new and unseen camera positions, allowing three dimensional interaction with photos. This is a long-studied problem, with early work relying on interpolation within dense structured image sets [20, 11, 8]. The specialized rigs commonly required to capture the large number of images restricted these methods to lab settings [46]. However, the potential applications offered by novel view synthesis on modern mobile and VR devices has kindled wide interest in the problem, and encouraged researchers to seek methods that make the technology more accessible. The term 3D photography refers to the use of novel view synthesis in everyday capture settings, often from a single image.", + "bbox": [ + 496, + 561, + 890, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past few years a number of proposed scene representations have leveraged the great strides being made in learning-based techniques to achieve more accurate synthesis with fewer constraints. The most recent of these are neural radiance fields (NeRFs) [26, 50] which represent the scene as multi-layer perceptrons. Their results define the high bar of novel view synthesis. However, this high quality has a significant data and computational cost.", + "bbox": [ + 496, + 780, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10454", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An alternate representation, a multiplane image (MPI), defines the scene as a stack of fronto-parallel RGBA planes that can be warped and rendered into novel viewpoints [53, 26, 9]. An MPI offers the advantage of rendering speed, and suffers from less aliasing than mesh or point-based methods [38, 47, 29]. The latter characteristic is important for applications that require temporal stability. However, an MPI is a highly redundant scene representation: the number of RGBA planes required to capture and reconstruct all the depth variation in a scene can be quite high. Since most scenes have a larger amount of free space than occupied, most of the planes in an MPI are very sparse. This makes them inefficient as a representation [2], and expensive to generate, transmit, and store.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper we propose to address these shortcomings of multiplane images and develop a lightweight solution to the 3D photography problem that can be practically implemented on mobile and VR devices. Some examples of the applications we envision are 3D video conferencing, telepresence, and VR passthrough [49]. We show how subdividing the image plane into many small MPs with only a few planes in each, provides a more efficient representation from a computational and memory perspective. However, the naive approach of directly using existing MPI methods with tiles creates boundary artifacts in the novel views. This happens because the commonly used fixed spacing of MPI planes fails to capture the full depth range of a tile when the number of planes is small. Furthermore, it is sensitive to outliers in small regions. We propose a clustering-based approach using learnt confidence weights to predict per-tile MPI planes that better represents local depth features. Our method is lightweight and generates results comparable to the state-of-the-art in MPI-based 3D photography.", + "bbox": [ + 75, + 301, + 472, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contribution of this work are,", + "bbox": [ + 96, + 590, + 437, + 604 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. The demonstration of tiled multiplane images as a practical representation for view synthesis tasks.", + "2. A learning framework for generating tiled multiplane images from a single RGB input for 3D photography.", + "3. A novel approach to adaptive MPI plane positioning." + ], + "bbox": [ + 88, + 614, + 468, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 724, + 218, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The progenitors of current 3D photography were the early works on image-based rendering (IBR). These methods usually relied on interpolation within the convex hull of a large set of images to generate novel views. Levoy et al. [20] and Gortler et al. [11] proposed the canonical two-plane parameterization of light fields that renders novel views by quadrilinear interpolation. Gortler's method also provided an early demonstration of the use of geometric proxies to improve rendering quality. Davis et al.'s [8] work extended interpolation-based view synthesis to unstructured", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "images. The excellent analysis of plenoptic sampling done by Chai et al. [5] proved, however, that for large distances the number of images required for view synthesis by interpolation was impractically high. Consequently, the large majority of recent view synthesis methods have relied on learnt priors to overcome the high sampling requirements.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One of the corollaries of Chai et al.'s analysis was that the sampling requirements for view-interpolation are inversely related to the geometric information of the scene. Thus, many subsequent methods have relied on coarse geometric proxies to improve view synthesis quality [34, 35, 13]. In Mildenhall et al.'s [26] method this proxy takes the form of a multiplane image (MPI), which they use to achieve interpolation-based view synthesis that overcomes Chai et al.'s sampling limits. MPIs were first proposed by Zhou et al. [53] who used them for extrapolating novel views outside the convex hull of the input stereo cameras (Tucker and Snavely [42] observe that an MPI can be considered as an instance of Szeliski and Golland's [40] earlier \"stack of acetates\" volumetric model). Mildenhall et al. [26] and Srinivasan et al. [39], respectively, provide a theoretical analysis of the limits of view interpolation and extrapolation using MPIs. Flynn et al. [9] use learnt gradient descent to generate MPIs from multiple views. Attal et al. [1] and Broxton et al. [3] extend the MPI concept to concentric RGBA spheres that can be used for view synthesis in 360 degrees. Wizadwongsa et al. [48] and Li et al. [22] replace the discrete RGBA planes of MPIs with continuous neural surfaces to achieve higher quality results. Tucker and Snavely [42] use a scale-invariant method that allows them to learn strong data priors that can generate MPIs from a single view. Their approach is in a line of recent work dubbed 3D photography that aims for novel views from in-the-wild, single-view images. Li et al. [23], Han et al. [12], and Luvizon et al.'s [25] methods are MPI-based examples of this approach. An MPI, however, is an over-parameterized scene representation. Recent methods [24, 10, 25, 12] have sought to overcome this shortcoming to some extent through a more judicious placement of depth planes. Nonetheless, their high level of redundancy dilates their memory and computational footprint, and limits their wider adoption in mobile and AR/VR applications.", + "bbox": [ + 496, + 185, + 892, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For such use cases, a more efficient approach to novel view synthesis involves depth-based warping [6, 49], often followed by inpainting [18, 19]. Shih et al. [38] propose to guide inpainting in disoccluded regions of the warped view using a layered depth representation. Li et al. [21] apply this general approach to 360-degree input. Wiles et al. [47] use a depth map to generate a point cloud of neural features which can be projected and rendered in novel views using a generative network. Choi et al. [7] estimate a probability volume instead of a single depth map to handle uncertainty in difficult regions. Niklaus et al. [29] use seg", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10455", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/67bbec6e0ad24ec0274d2d47657b11ebb87b122c7d2e4541c1bd0a1b90225258.jpg", + "image_caption": [ + "Figure 2: Tiled multiplane image generation for single-view 3D photography. Given an RGB image and depth from a monocular estimator, our method first generates a pixel-wise confidence map and a pre-processed depth map. These are used to predict $n$ depth planes per tile through weighted $k$ -means clustering (with $k = n$ ). The layer masks defined by pixel labels are used to peel RGBA layers, yielding a rough initial MPI. This is concatenated $(\\bigoplus)$ to the depth and passed to a refinement network that generates the RGBA images of the final per-tile MPI." + ], + "image_footnote": [], + "bbox": [ + 89, + 90, + 883, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mentation to remove the geometric and semantic distortions from depth that often impair the rendering quality of this approach. Nonetheless, depth-based warping suffers from hard boundaries and is over-sensitive to errors in the depth estimate. While Jampani et al. [14] propose soft layering with alpha mattes to address this shortcoming, MPIs are inherently capable of handling such artifacts via blending.", + "bbox": [ + 75, + 402, + 468, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, our review of view synthesis would not be complete without mentioning neural radiance fields (NeRFs) [27, 50] which have recently沦ued in popularity. While great strides are being made in improving the time and data requirements of NeRFs [51, 28], they remain expensive for interactive applications. Some recent work [33, 41, 43] shows that decomposing a single large radiance field into smaller sub-components can improve efficiency. This is similar to the approach we adopt for multiplane images.", + "bbox": [ + 75, + 508, + 470, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 656, + 168, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A traditional MPI represents the scene as a set of $N$ fronto-parallel planes in the camera frustum of a reference view $\\mathcal{I}$ . Each plane is associated with an RGBA image. While it is possible to place the planes at any depth, they are usually arranged linearly in disparity (inverse depth) [53, 9, 26, 42]. A novel view $\\mathcal{I}_t$ is rendered by warping the planes into the target camera's image space via a homography, and compositing them front-to-back using the over [31] operator:", + "bbox": [ + 75, + 681, + 468, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {t} = \\sum_ {i = 1} ^ {N} \\left(\\alpha_ {i} c _ {i} \\prod_ {j = i + 1} ^ {N} (1 - \\alpha_ {i})\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 823, + 468, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ and $c_{i}$ are the warped alpha and color channels, respectively, of the $i^{th}$ plane.", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both warping and compositing can be done very efficiently on graphics hardware, allowing real-time rendering of novel views [26]. Moreover, the alpha channel at each plane allows MPs to represent soft edges and anti-alias any errors in the scene reconstruction, leading to fewer perceptually objectionable artifacts than depth-based warping methods. However, the number of planes $N$ required to capture all the depth variation in a scene is usually large, even though most planes are very sparse. We propose to overcome this shortcoming by representing the reference image $\\mathcal{I}$ as a tiled grid of many small MPs (Figure 3). Given $\\mathcal{I}$ and its depth map from a monocular depth estimator, our method predicts the placement of $n \\ll N$ depth planes within each tiled region and uses this prediction to generate the RGBA images of the MPI in a single forward pass.", + "bbox": [ + 496, + 402, + 890, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Tiled Multiplane Image Representation", + "text_level": 1, + "bbox": [ + 498, + 637, + 838, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our scene representation is based on a set of $m$ tiles, each representing a square sliding block of size $h$ at 2D pixel locations $\\mathbf{x}_1, \\mathbf{x}_2, \\dots, \\mathbf{x}_m$ in the source image $\\mathcal{I}$ . The locations $\\mathbf{x}_k$ lie on a regular grid with spacing determined by some stride $r$ . Each tile consists of $n$ front-parallel RGBA planes, the depth placement of which is not fixed but varies across tiles. We let $\\alpha_j^i$ , $c_j^i$ and $d_j^i$ denote, respectively, the alpha channel, the color channel, and depth of the $i$ -th plane in the $j$ -th tile. Then the tiled multiplane image (TMPI) representation $\\Gamma(\\mathcal{I})$ of the image is defined as:", + "bbox": [ + 496, + 660, + 890, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma (\\mathcal {I}) = \\left\\{\\left(\\alpha_ {j} ^ {i}, c _ {j} ^ {i}, d _ {j} ^ {i}, \\mathbf {x} _ {j}\\right) \\right\\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 816, + 890, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for all $i = 1,2,\\dots,n$ and $j = 1,2,\\dots,m$ . Further, we define an ordering on this set of 4-tuplets as,", + "bbox": [ + 496, + 842, + 890, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\Gamma (\\mathcal {I}), \\leq\\right) = \\left\\{\\left(\\alpha , c, d, \\mathbf {x}\\right) _ {k = 1, \\dots , m n} \\mid d _ {k} \\leq d _ {k + 1} \\right\\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 878, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10456", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9fca864d2e7945f78dac19a524f93a86a109c0601dc0dee4f6de92529a24fd88.jpg", + "image_caption": [ + "Figure 3: (a) A traditional MPI uses five planes per pixel to represent this toy scene, even though no region has more than two overlapping objects. (b) A TMPI exploits the low local depth complexity and has only two planes per pixel." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 472, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As with traditional MPIs, novel views are rendered in a differentiable manner by warping all planes into the target camera's image space. However, as the depth of the planes varies across the tiles and, hence, across the pixels of $\\mathcal{I}$ , a planar inverse warp on the target image plane cannot be directly computed via a homography. Instead, each plane must be warped in tile space via a homography computed using a shifted intrinsic matrix, and all the warped planes composited sequentially in the target view at their respective tile locations (Algorithm 1). While this makes the rendering of tiled multiplane images less efficient and somewhat less elegant than MPs (Equation 1), this is only true during training when differentiability is required. At inference, the tiles can be rendered as textured quads using hardware-accelerated rasterization making this stage of the pipeline as efficient as traditional MPs and with lower texture memory requirements. Thus, their compact form makes TMPIs well-suited for rendering over networks, or on mobile devices and VR headsets. The blending of many MPs locally is related to Mildenhall et al.'s [26] light field fusion. However, their method renders each MPI separately before blending the results with scalar weights. Our method operates at a much finer scale, and composites the planes of all MPs together to synthesize a novel view.", + "bbox": [ + 75, + 330, + 472, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Single-View 3D Photography", + "text_level": 1, + "bbox": [ + 76, + 709, + 334, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now describe our approach to generating tiled multiplane images from a single RGB input (Figure 2). Broadly, our method splits the image plane into a regular tiled grid of learnt confidence and outlier-corrected depth. For each tile, the placement of $n$ fronto-parallel depth planes is determined by clustering pixel depth values weighted by the predicted confidence estimates. This latter step is motivated by the fact that with a small depth plane budget $n$ , the commonly used equal spacing in disparity [53, 9, 26, 42] is wasteful. Thus, the goal is to predict the planes that optimally represent all depth variation within a tile. Using the", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1: Differentiable view synthesis using TMPIs. Angled brackets denote pixel indexing.", + "bbox": [ + 506, + 94, + 849, + 125 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "RenderTMPI $(\\Gamma (\\mathcal{I}),\\leq)$ $\\pmb {R},t,\\pmb {K})$ Input: $(\\Gamma (\\mathcal{I}),\\leq)$ : ordered TMPI planes R: relative rotation of novel view t: relative translation K: camera intrinsics Output: Novel view $\\mathcal{I}_t\\in \\mathbb{R}^{3\\times H\\times W}$ $\\mathcal{I}_t\\gets \\mathbb{0}^{3\\times H\\times W};\\mathcal{T}_t\\gets \\mathbb{1}^{H\\times W};$ foreach $(\\alpha ,c,d,\\pmb {x})\\in \\left(\\Gamma (\\mathcal{I}),\\leq\\right)$ do $\\hat{\\mathbf{K}}\\gets \\mathbf{K} - \\left[ \\begin{array}{cc}\\mathbf{I} & \\mathbf{x}\\\\ 0 & 1 \\end{array} \\right];$ foreach $u\\in [1,\\dots,s]\\times [1,\\dots,s]$ do n $\\left[\\begin{array}{l}\\mathbf{u}_s\\\\ 1\\end{array}\\right]\\gets \\hat{\\mathbf{K}} (\\mathbf{R} - \\mathbf{tn}^T /d)\\hat{\\mathbf{K}}^{-1}\\left[\\begin{array}{l}\\mathbf{u}\\\\ 1\\end{array}\\right];$ $w\\gets \\alpha \\langle \\mathbf{u}_s\\rangle \\mathcal{T}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle ;$ $\\mathcal{I}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle \\gets \\mathcal{I}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle +w c\\langle \\mathbf{u}_s\\rangle ;$ $\\mathcal{T}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle \\gets \\mathcal{T}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle (1 - \\alpha \\langle \\mathbf{u}_s\\rangle);$ end end return I", + "bbox": [ + 519, + 127, + 851, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "predicted planes, a fully convolutional network generates the $n$ RGBA images that constitute a per-tile MPI. Unlike Han et al.'s [12] adaptive plane method, we generate the RGBA images in a single forward pass. The resulting TMPI is rendered as a set of textured quads using the rasterization pipeline of graphics hardware.", + "bbox": [ + 496, + 503, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In more detail, given the source image $\\mathcal{I}$ , we first obtain a depth map $\\mathcal{Z}$ for it using a monocular depth estimator. A two-headed U-Net $\\Theta(\\cdot)$ then predicts a confidence map $\\mathcal{C}$ along with denoised depth $\\mathcal{D}$ . The goal is to learn a representation that ameliorates the sensitivity of the subsequent $k$ -means clustering step to outliers. The joint prediction of confidence and depth is similar to the depth-routing of Weder et al. [45] and the aleotoric uncertainty estimation of [15]. The predicted depth and confidence, and the original color image are then unfolded into a set of $m$ square sliding blocks of size $h$ and stride $r$ : $\\{(\\mathcal{D}^i,\\mathcal{C}^i,\\mathcal{I}^i)_{i=1,\\dots,m}\\}$ .", + "bbox": [ + 496, + 595, + 892, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Running $\\Theta(\\cdot)$ on $\\mathcal{I}$ and $\\mathcal{Z}$ rather than individual tiles allows it to consider non-local features and avoid undesirable tiling artifacts. Additionally, setting $r < h$ in the unfolding step allows neighboring tiles to overlap. This prevents gaps along tile boundaries and also regularizes per-tile operations across neighbors. However, it also increases the total number of tiles and, thus, the computational requirements. We empirically determine a stride value of $r = h - h/8$ for a good balance between quality and computational efficiency.", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10457", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Per-Tile Planes Prediction: Next, we predict the $n$ depth planes $\\{z_{j=1,\\dots,n}^{i}\\}$ that optimally represent the features of the $i$ -th tile. The common approach of spacing the planes linearly in disparity grows inaccurate as $n$ becomes small. Luvizon et al. [25] place the planes at depth discontinuities identified via the histogram of depth values. However, their approach is sensitive to parameter settings and fails for smooth surfaces which have no discontinuities. A learning-based approach is adopted by Han et al. [12] and Li et al. [23]. The former use multi-headed self-attention to adjust a linear placement. While capable of modeling interplane interactions, their method is computationally expensive (Table 2). The latter uses a CNN to directly predict the planes. However, we found that without strong regularization a direct approach lacks topological order and has a strong bias towards a fixed placement. An adversarial loss helps improve this but makes the training more unstable.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We observe that as depth is known, plane positioning can be posed as a simple clustering problem. Thus, we predict $\\{z_j^i\\}$ using $k$ -means clustering on the depth in each tile $\\mathcal{D}^i$ .", + "bbox": [ + 76, + 347, + 468, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Standard $k$ -means is sensitive to outliers and can generate significantly different plane predictions across neighboring tiles causing artifacts in novel views. Advantageously, along with the $n$ depth planes, clustering also assigns a label to each input pixel, thereby generating a label map that represents discretized depth. Furthermore, the cluster centers of $k$ -means are differentiable with respect to the input samples. Thus, we address the outlier problem by training $\\Theta(\\cdot)$ to filter the input through a self-supervised reconstruction loss on the discretized depth map generated by a weighted $k$ -means. In weighted $k$ -means, the cluster centers are updated each iteration using the confidence-weighted mean of the constituent samples. Since we do not directly supervise the depth output of $\\Theta(\\cdot)$ , the network can go beyond outlier-filtering to learn any modifications that improves the discrete reconstruction, and consequently optimizes the placement of the $n$ depth planes within each tile.", + "bbox": [ + 75, + 393, + 472, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Per-Tile MPI Generation: Given $\\{z_j^i\\}$ and the discrete depth map, we estimate a preliminary MPI per tile by peeling RGBA layers from $\\mathcal{I}^i$ using the discrete labels as an alpha mask. The masked RGB regions of each plane are inpainted by upsampling valid values from a Gaussian pyramid. A second network $\\Psi (\\cdot)$ then refines these estimates to generate the final $n$ RGBA images of the MPI for each tile. Following Zhou et al. [53] and Tucker and Snavely [42], we represent the RGBA output as a pixel-wise blend of the input image $\\mathcal{I}^i$ and a learnt background. However, unlike these works we predict a background image $\\mathcal{B}_j^i$ per plane:", + "bbox": [ + 75, + 648, + 472, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} _ {j} ^ {i} = \\prod_ {k > j} \\left(1 - \\alpha_ {k} ^ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 823, + 468, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {j} ^ {i} = \\mathcal {W} _ {j} ^ {i} \\mathcal {I} ^ {i} + (1 - \\mathcal {W} _ {j} ^ {i}) \\mathcal {B} _ {j} ^ {i} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 859, + 468, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $\\alpha_{k}^{i}$ is the predicted alpha value for each plane.", + "bbox": [ + 76, + 883, + 431, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/12a5be94f7060e2e2c845982942c1dab146f2ebc695cbbea2b934f4d8969998a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Spaces Dataset
MethodPSNR↑SSIM↑LPIPS↓L1↓
SVMPI [42]25.420.7480.2100.040
VMPI [23]22.370.6360.2680.057
MINE [22]24.020.7020.2290.048
AdaMPI [12]26.170.7030.2290.047
Ours24.930.7500.1750.037
Tanks & Temples Dataset
MethodPSNR↑SSIM↑LPIPS↓L1↓
SVMPI17.850.5300.3700.082
VMPI16.320.4630.3950.103
MINE17.230.5060.3660.088
AdaMPI18.620.5650.2700.073
Ours18.690.5690.2670.073
", + "bbox": [ + 509, + 88, + 883, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Quantitative evaluation of view synthesis results on the Spaces and Tanks and Temples multi-view datasets. SVMPI, MINE and AdaMPI use 32 MPI depth planes; VMPI uses 8; our approach uses 4 planes per image tile.", + "bbox": [ + 496, + 364, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Training Procedure", + "text_level": 1, + "bbox": [ + 500, + 450, + 689, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train $\\Theta(\\cdot)$ in a self-supervised manner by minimizing the L1 loss between the input depth $\\mathcal{Z}$ and the folded discrete depth maps produced per tile by weighted $k$ -means clustering. Then, we freeze $\\Theta(\\cdot)$ and train the MPI generation network $\\Psi(\\cdot)$ on a novel view synthesis task by following Han et al.'s [12] warp-back strategy to generate pseudo ground truth multi-view training data. This involves warping single-view images into a target camera using monocular depth, and inpainting disocclusion holes with a specially trained network. The view-synthesis training objective is a combination of VGG, structural similarity [44], and L1 losses on the synthesized color image $\\mathcal{I}_t$ , weighed as 0.1, 0.25, and 1.0 respectively. For both networks, we use the 111K images of the COCO dataset [4] and the monocular depth method of Ranftl et al. [32].", + "bbox": [ + 496, + 476, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 715, + 632, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation", + "text_level": 1, + "bbox": [ + 500, + 739, + 660, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method is implemented in PyTorch and trained on eight Nvidia Tesla V100 GPUs. We use $256 \\times 384$ images and accumulate gradients across eight mini-batches of 16 samples each. For both networks $\\Theta(\\cdot)$ and $\\Psi(\\cdot)$ , we use the Adam optimizer with a learning rate of $1 \\times 10^{-3}$ and a cosine annealing schedule with restarts every 200 epochs. A vectorized implementation of the differentiable TMPI renderer (Algorithm 1) runs at $\\sim 100\\mathrm{ms} / \\mathrm{mini}$ -batch allowing efficient parallel computation.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10458", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ca38c01f4a607ac82084cc96d2befaf9e16b4767dfe9ed406d2e6244e739a24a.jpg", + "table_caption": [], + "table_footnote": [ + "*Monocular depth method. † Method uses monocular depth.", + "Table 2: Run-time, memory and complexity evaluation of all methods with a single $350 \\times 630$ RGB image on an NVIDIA GeForce RTX 3080 GPU. The choice of resolution is dictated by the baseline methods which run out of memory for larger images on the specified hardware. GMACS are Giga Multiply-Accumulate ops./second." + ], + "table_body": "
Params.↓(M)GMAC↓Runtime (ms)Peak (GB)Space (MB)↓
DPT*123.0110314.39-
SVMPI43.558.01114.5326.9
MINE38.12501104.91107
VMPI†4.3152.396.04.5713.5
AdaMPI†19.02883505.94107
Ours†6.4357.091.63.205.25
", + "bbox": [ + 78, + 89, + 467, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Baselines", + "text_level": 1, + "bbox": [ + 76, + 387, + 184, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our approach to four state-of-the-art single-view 3D photography methods based on multiplane images: Tucker et al. [42] (SVMPI), Li et al. [23] (VMPI), Li et al. [22] (MINE) and Han et al. [12] (AdaMPI). Like us, VMPI and AdaMPI use a monocular estimator to recover depth as the first step of their pipeline. The input to all methods is the same, however — a single unconstrained RGB image — and so we evaluate them as end-to-end 3D photography approaches. Nonetheless, we do note the additional depth estimation step when evaluating computational and memory performance (Table 2). We use $N = 32$ planes for all baselines except VMPI, which is designed for $N = 8$ .", + "bbox": [ + 75, + 411, + 468, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We do not compare to single-view methods based on neural radiance fields [51] or the recent work of Nicklaus et al. [29] and Shih et al. [38] as these are too computationally intensive for our intended use cases on mobile and VR devices. While recent work on NeRFs has demonstrated impressive rendering speeds, training remains expensive, and further, requires a large number of input views and a static scene. Jampani et al.'s [14] method, though not MPI-based, is related. But the authors have not released their code.", + "bbox": [ + 75, + 593, + 468, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Testing Datasets", + "text_level": 1, + "bbox": [ + 76, + 739, + 238, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We test all methods on the Spaces [9], and Tanks and Temples [17] datasets. Spaces consists of 100 indoor and outdoor scenes captured using a purpose-built, 16-camera rig. For Tanks and Temples we use the Intermediate split which has uniformly sampled frames from eight high-resolution videos of more challenging outdoor environments. We compute camera poses and depth maps for all scenes using COLMAP [37, 36]. The depth maps are required to resolve the scale ambiguity of monocular depth", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6f72d671c8552ddaf56415b7e9da58bbb5a2b7a0707f7b95a36fc9d3d054f95f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3: Evaluating the reconstruction error of different depth discretization approaches. All values are $\\times 10^{-3}$ ." + ], + "table_body": "
Spaces Dataset
VariantMAE ↓MSE ↓Q25 ↓
Vanilla k-means35.32.0014.0
Linear plane spacing48.73.8019.3
Ours27.11.2010.6
Tanks& Temples Dataset
VariantMAE ↓MSE ↓Q25 ↓
Vanilla k-means35.72.0014.2
Linear plane spacing46.03.4018.1
Ours27.01.2010.6
", + "bbox": [ + 532, + 88, + 859, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for correct reprojection to target views. We randomly select 1000 source views from each dataset scaled to $350 \\times 630$ and use the next image in capture sequence as the target for view synthesis. This choice of resolution is dictated by the baseline methods which run out of memory for larger images on the specified hardware. We present high resolution results for our method on the Davis dataset [30] in Figure 5.", + "bbox": [ + 496, + 351, + 890, + 457 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 465, + 683, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the rendered views quantitatively on four metrics: Peak-Signal to Noise Ratio (PSNR), Structural Similarity [44] (SSIM), Learned Perceptual Image Patch Similarity [52] (LPIPS) with a VGG-16 backbone, and the mean absolute error (L1). Following previous work [12, 42, 23], we crop $15\\%$ of the image around the edge to account for disocclusions. Further, if more than $15\\%$ of the remaining pixels are blank in a view synthesized by any method, we discard the result across all baselines.", + "bbox": [ + 496, + 489, + 890, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.5. Results", + "text_level": 1, + "bbox": [ + 500, + 635, + 591, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 presents quantitative evaluation of all methods on the two test sets. Qualitative results are shown in Figure 4 for Tanks and Temples and Figure 9 for Spaces. Our approach uses $n = 4$ depth planes with a tile size of $h = 64$ . The competitive performance of our method despite having much fewer planes per tile can be attributed to the adaptive placement of depth planes in each tile which allows it to effectively cover a larger depth range than the monolithic MPIs of the baselines, Memory and computational performance is evaluated in Table 2. VMPI, AdaMPI and our method have the additional overhead of monocular depth estimation using Ranftl et al.'s [32] DPT. Our approach has lower runtime, peak memory and space requirements than AdaMPI while achieving similar quality results.", + "bbox": [ + 496, + 657, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our plane placement strategy in Table 3. We measure the reconstruction quality of the discretized depth", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10459", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d5d3341a5294991bb73f76a7a5c258140aacb5621ea241c723965860eec4661c.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 83, + 87, + 357, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84783fce5dc6a37c49c093bfb3727840eefa4334eb2703fb85d3ce54f377b690.jpg", + "image_caption": [ + "SVMPI" + ], + "image_footnote": [], + "bbox": [ + 359, + 88, + 632, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2523570da249fc8e6dfbf319c70a860aea132d5c9e2c9b22b938e5bc6e82b5b.jpg", + "image_caption": [ + "VMPI" + ], + "image_footnote": [], + "bbox": [ + 635, + 88, + 890, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/097f682ea8c12c38c1f1ffaaf7ff2abe8e6df5fe69322eee873526f5fe2030a3.jpg", + "image_caption": [ + "MINE" + ], + "image_footnote": [], + "bbox": [ + 84, + 218, + 357, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/572db4012fa1f463b941f2423b3fb8ee7ba6321dfdaad11c6ddca1db04b9fd48.jpg", + "image_caption": [ + "AdaMPI" + ], + "image_footnote": [], + "bbox": [ + 359, + 218, + 632, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f7f617268f73475a6729d1458798073f875fb9cad7d8ee9a58c1131fbb82dc4.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 635, + 218, + 890, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49d27162c44df8a204fde1b953248a24e7d390ba9198728f093af29363924e00.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 84, + 359, + 357, + 478 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/58b248577670ae1c2b6b7862f78fc86844984b1aa9fb4ceecda86b34370eea78.jpg", + "image_caption": [ + "SVMPI" + ], + "image_footnote": [], + "bbox": [ + 359, + 359, + 632, + 478 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f03703d0453e85b3317cb5c982a503706458ef246fcd6fee10d9fd2c11edc4c7.jpg", + "image_caption": [ + "VMPI" + ], + "image_footnote": [], + "bbox": [ + 635, + 359, + 890, + 478 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f8ae46a32fcf89b1b05befaa4eada23aa99779cc367bcbe6700d3b758e2162e.jpg", + "image_caption": [ + "MINE" + ], + "image_footnote": [], + "bbox": [ + 84, + 494, + 357, + 612 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/70e328d37e970b678d4f405488b6f162a2f105b71215f5c6a30520439afc33e7.jpg", + "image_caption": [ + "AdaMPI" + ], + "image_footnote": [], + "bbox": [ + 359, + 494, + 632, + 612 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/590fbd5d7287e30b57857b1fb04bc3a2ecc07feb23b30d2db0f68f588117d990.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 635, + 494, + 890, + 612 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/689381e5f19e89df23b908c7ccddd5b904e32871b61b9581cd9a31d9db6e94cc.jpg", + "image_caption": [ + "Figure 4: Comparing the novel view synthesis results of the baseline methods and our approach on the Tanks and Temples dataset. Our results are better than SVMPI, VMPI and MINE, and comparable to AdaMPI while using far fewer depth planes." + ], + "image_footnote": [], + "bbox": [ + 78, + 676, + 282, + 767 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7743671019847d52e6b4c0901f604c37e8777d4d572fde853624d0b48945484d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 676, + 485, + 767 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/747e0e95f70c4b24c39ae960a3fec7b1abd5d337150ac8e644193b88fc1b6ca5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 676, + 687, + 767 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a1e224c48122c66cab594aad955fb886ee2f8e9aa8621147ce4906e856080cb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 676, + 890, + 767 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db4e8bdeca117ce1f8e31f77529e7fc98b35aa624312fe66075f296e564357a5.jpg", + "image_caption": [ + "Figure 5: View synthesis results of our method on the HD $(1080\\times 1920)$ Davis [30] dataset. Original views are inset." + ], + "image_footnote": [], + "bbox": [ + 78, + 767, + 282, + 856 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87313071b9aa01d92b46cfffe7b11be87cf454a1811ca01230e2a377f471f99c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 767, + 485, + 856 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/22fda26b6807b906f95131b175a6a454e553e29346b806d7be7534aed23d9909.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 767, + 687, + 856 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/672b2781e7fe6008b7dcf44f62781d90d422d4e05df0dca7a75897d6660c3ffd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 767, + 890, + 856 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10460", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c7a3055412f0e0e4a50f6494ad45f6c09826bf106612bdb1c08205ec67d651c9.jpg", + "image_caption": [ + "Figure 6: Top to bottom: Input image (Davis dataset); inverse monocular depth from DPT [32]; naive discretization by spacing four planes linearly in the inverse-depth range of each tile; our weighted clustering-based discretization with four planes. Our representation shows finer variation on receding surfaces and suffers fewer tiling artifacts." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 472, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "map defined by the planes in each tile on a monocular depth input that is perturbed by a small amount of Gaussian noise $(\\lambda = 0, \\sigma^2 = 1 \\times 10^{-3})$ . We compare our approach to naive linear spacing of depth planes in inverse disparity space, and to vanilla unweighted $k$ -means. Our method is robust to outliers and yields a much better reconstruction. Figure 6 provides qualitative comparison with linear spacing on samples from the Davis dataset [30].", + "bbox": [ + 75, + 551, + 468, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7 evaluates the effect of tile size and number of planes on the quality of view synthesis. In general, the results uphold the intuition that small tiles and a large number of planes improve quality. This trend is less clear, however, for number of planes $>8$ . This would seem to support the observation of Khakhulin et al. [16] and Hu et al. [13] that reductive models are unable to handle redundant geometry effectively. Moreover, the model's complexity increases proportionately with the number of depth planes, leading to slower convergence for the same number of training steps.", + "bbox": [ + 75, + 674, + 470, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 76, + 844, + 200, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As previously observed, a tiled multiplane image cannot exploit the elegant warping and compositing equations of a", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3cbb07f10df1902889d721b0156c6606976423c2c71d286fba9e3ac079739cf8.jpg", + "image_caption": [ + "Figure 7: Evaluating the effect of tile size and number of depth planes on view synthesis quality. In general, the results uphold the intuition that small tiles and a large number of planes improve quality. The trend, however, is less clear when the number of planes $>8$ , indicating that the model may be unable to handle redundant geometry effectively." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 867, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/15577a5f4af11bb15bc43cb95cc81aaaa245cdc049108abd1a6601c5ac8611ff.jpg", + "image_caption": [ + "Figure 8: Failure cases of our approach: in some cases, fine features are inconsistently reconstructed across tiles." + ], + "image_footnote": [], + "bbox": [ + 501, + 414, + 890, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "traditional MPI for differentiable rendering during training. Furthermore, in some cases our method fails to reconstruct thin features consistently across tiles (Figure 8).", + "bbox": [ + 496, + 577, + 890, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 635, + 619, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a method for estimating tiled multiplane images from a single RGB input for 3D photography. This includes a novel approach to adaptively spacing a small number of depth planes within an MPI tile to better represent local features. Our method is lightweight, and points a path to realizing novel view synthesis on mobile and VR devices.", + "bbox": [ + 496, + 660, + 890, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 763, + 596, + 779 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Benjamin Attal, Selena Ling, Aaron Gokaslan, Christian Richardt, and James Tompkin. MatryODShka: Real-time 6DoF video view synthesis using multi-sphere images. In European Conference on Computer Vision (ECCV), Aug. 2020. 2", + "[2] Horace B Barlow et al. Possible principles underlying the transformation of sensory messages. Sensory communication, 1(01):217-233, 1961. 2" + ], + "bbox": [ + 506, + 787, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10461", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/304d47c8828685d112135b487dee026ef179927a7a178b1630a2c9d8474a8208.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 124, + 88, + 372, + 186 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c346e1eaabb8454f35f6b5c2f928dcfb1619cf8e5707566d03a81d939be7a4d2.jpg", + "image_caption": [ + "SVMPI" + ], + "image_footnote": [], + "bbox": [ + 375, + 88, + 622, + 186 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/145c469ab0aa27995c281690b5e73f0eca8040170466179c0bfdc49955b90b31.jpg", + "image_caption": [ + "VMPI" + ], + "image_footnote": [], + "bbox": [ + 625, + 88, + 849, + 186 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/da7f5e229080bd2bd42eca9888f86ba10cfa1b00ddfaf5468aed200e39d31b82.jpg", + "image_caption": [ + "MINE" + ], + "image_footnote": [], + "bbox": [ + 122, + 202, + 372, + 310 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/da4a4744a9b4f9ec02700b1b92a3fd32c253a48196921c6df42b3b020f4be215.jpg", + "image_caption": [ + "AdaMPI" + ], + "image_footnote": [], + "bbox": [ + 374, + 202, + 622, + 310 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e8edcc08156e8c9157aa5a9b35e23755173db93e55d4b061f0190da0efe545c6.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 625, + 202, + 849, + 310 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/a21dcc29f2aeceaf5dfd84fc8c60f4e7abd6892ab6a7b0dc4f7127b670c1a5b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 338, + 372, + 446 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/d148ab71e9ab70ccfebd50a85b601424bbea3601713f8e77d19df476f35f9b7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 338, + 622, + 446 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8dda61d1d1ecb50c79cc281fc8316b79ebb16f2226d8f1e07937f0ba18a4de9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 338, + 849, + 446 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/052b27b31fa575bee61292c07bb2f1593ec8aa7cf5b60a0871f081fa55d21377.jpg", + "image_caption": [ + "Ground Truth", + "MINE" + ], + "image_footnote": [], + "bbox": [ + 122, + 464, + 372, + 571 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/fec7b93b2942740b0578375aeaaa6c182bbaa7314c61111576cc8ba8aca8b804.jpg", + "image_caption": [ + "SVMPI", + "AdaMPI" + ], + "image_footnote": [], + "bbox": [ + 374, + 464, + 622, + 571 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/5e642b2d76297fdc0fd52d4f4e689ab590fbc490d218aecee2431ce2c27eac6e.jpg", + "image_caption": [ + "VMPI", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 625, + 464, + 849, + 571 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/0a8340d8bf2e211fc04a3befeb94c5110340b884fa9530ad3c7e356086857c87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 599, + 372, + 705 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/1fa0d4d74591c473e12bfffb405f7499600af16bcef976a026b8bb87b39b27df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 599, + 624, + 705 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/078ca2fa2a9593a01fb37b7386f95bb6572b5cee085333145b4d107a175fad70.jpg", + "image_caption": [ + "VMPI" + ], + "image_footnote": [], + "bbox": [ + 625, + 599, + 849, + 705 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ec0528a6027fa37d44d330f9d2388455b92e00a0359a1302e65d1ae6ff807dee.jpg", + "image_caption": [ + "Ground Truth", + "MINE", + "Figure 9: Comparing the novel view synthesis results of the baseline methods and our approach on the Spaces dataset." + ], + "image_footnote": [], + "bbox": [ + 124, + 722, + 372, + 830 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8dde621a480905b77b0e49619bf4ca9d7d539d6bc400447d6c53bf250a3dfdf1.jpg", + "image_caption": [ + "SVMPI", + "AdaMPI" + ], + "image_footnote": [], + "bbox": [ + 374, + 722, + 622, + 830 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/310c22732e27657398b0997f1f70ffc7e9b0367a74428e9295ab31022c7cd7cf.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 625, + 722, + 849, + 830 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10462", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[3] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. 2", + "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 5", + "[5] Jin-Xiang Chai, Xin Tong, Shing-Chow Chan, and Heung-Yeung Shum. Plenoptic sampling. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 307-318, 2000. 2", + "[6] Gaurav Chaurasia, Arthur Nieuwoudt, Alexandru-Eugen Ichim, Richard Szeliski, and Alexander Sorkine-Hornung. Passthrough+ real-time stereoscopic view synthesis for mobile mixed reality. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1-17, 2020. 2", + "[7] Inchang Choi, Orazio Gallo, Alejandro Troccoli, Min H Kim, and Jan Kautz. Extreme view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7781-7790, 2019. 2", + "[8] Abe Davis, Marc Levoy, and Fredo Durand. Unstructured light fields. In Computer Graphics Forum, volume 31, pages 305-314. Wiley Online Library, 2012. 1, 2", + "[9] John Flynn, Michael Broxton, Paul Debevec, Matthew DuVall, Graham Fyffe, Ryan Overbeck, Noah Snavely, and Richard Tucker. Deepview: View synthesis with learned gradient descent. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2367-2376, 2019. 2, 3, 4, 6", + "[10] Sushobhan Ghosh, Zhaoyang Lv, Nathan Matsuda, Lei Xiao, Andrew Berkovich, and Oliver Cossairt. Liveview: dynamic target-centered mpi for view synthesis. arXiv preprint arXiv:2107.05113, 2021. 2", + "[11] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 43-54, 1996. 1, 2", + "[12] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-view synthesis in the wild with learned adaptive multi-plane images. In ACM SIGGRAPH, 2022. 2, 4, 5, 6", + "[13] Ronghang Hu, Nikhila Ravi, Alexander C Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12528-12537, 2021. 2, 8", + "[14] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, and Ce Liu. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 3, 6", + "[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 4", + "[16] Taras Khakhulin, Denis Korzhenkov, Pavel Solovev, Gleb Sterkin, Andrei-Timotei Ardelean, and Victor Lempitsky." + ], + "bbox": [ + 78, + 92, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Stereo magnification with multi-layer images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8687-8696, 2022. 8", + "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 6", + "[18] Johannes Kopf, Suhib Alsisan, Francis Ge, Yangming Chong, Kevin Matzen, Ocean Quigley, Josh Patterson, Jossie Tirado, Shu Wu, and Michael F Cohen. Practical 3d photography. In Proceedings of CVPR Workshops, volume 1, 2019. 2", + "[19] Johannes Kopf, Kevin Matzen, Suhib Alsisan, Ocean Quigley, Francis Ge, Yangming Chong, Josh Patterson, Jan-Michael Frahm, Shu Wu, Matthew Yu, et al. One shot 3d photography. ACM Transactions on Graphics (TOG), 39(4):76-1, 2020. 2", + "[20] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996. 1, 2", + "[21] David Li, Yinda Zhang, Christian Hane, Danhang Tang, Amitabh Varshney, and Ruofei Du. Omnisyn: Synthesizing 360 videos with wide-baseline panoramas. In 2022 IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops (VRW), pages 670-671. IEEE, 2022. 2", + "[22] Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. MINE: Towards continuous depth spi with nerf for novel view synthesis. In ICCV, 2021. 2, 5, 6", + "[23] Qinbo Li and Nima Khademi Kalantari. Synthesizing light field from a single image with variable mpi and two network fusion. ACM Trans. Graph., 39(6):229-1, 2020. 2, 5, 6", + "[24] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4521-4530, 2019. 2", + "[25] Diogo C Luvizon, Gustavo Sutter P Carvalho, Andreza A dos Santos, Jhonatas S Conceicao, Jose L Flores-Campana, Luis GL Decker, Marcos R Souza, Helio Pedrini, Antonio Joia, and Otavio AB Penatti. Adaptive multiplane image generation from a single internet picture. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2556-2565, 2021. 2, 5", + "[26] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 1, 2, 3, 4", + "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a mul" + ], + "bbox": [ + 503, + 92, + 892, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10463", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3", + "[29] Simon Niklaus, Long Mai, Jimei Yang, and Feng Liu. 3d ken burns effect from a single image. ACM Transactions on Graphics (ToG), 38(6):1-15, 2019. 2, 6", + "[30] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 6, 7, 8", + "[31] Thomas Porter and Tom Duff. Compositing digital images. In Proceedings of the 11th annual conference on Computer graphics and interactive techniques, pages 253-259, 1984. 3", + "[32] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 5, 6, 8", + "[33] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 3", + "[34] Gernot Riegler and Vladlen Koltun. Free view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 623-640. Springer, 2020. 2", + "[35] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12216-12225, 2021. 2", + "[36] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6", + "[37] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 6", + "[38] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6", + "[39] Pratul P Srinivasan, Richard Tucker, Jonathan T Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snively. Pushing the boundaries of view extrapolation with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 175-184, 2019. 2", + "[40] Richard Szeliski and Polina Golland. Stereo matching with transparency and matting. In Sixth International Conference on Computer Vision (IEEE Cat. No. 98CH36271), pages 517-524. IEEE, 1998. 2", + "[41] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 3", + "[42] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In The IEEE Conference" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 3, 4, 5, 6", + "[43] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 3", + "[44] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5, 6", + "[45] Silvan Weder, Johannes Schonberger, Marc Pollefeys, and Martin R Oswald. Routedfusion: Learning real-time depth map fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4887-4897, 2020. 4", + "[46] Bennett Wilburn, Neel Joshi, Vaibhav Vaish, Eino-Ville Talvala, Emilio Antunez, Adam Barth, Andrew Adams, Mark Horowitz, and Marc Levoy. High performance imaging using large camera arrays. In ACM SIGGRAPH 2005 Papers, pages 765-776. 2005. 1", + "[47] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 2", + "[48] Suttisak Wizadwongsa, Pakkapon Phongthawee, Jiraphon Yenphraphai, and Supasorn Suwajanakorn. Nex: Real-time view synthesis with neural basis expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8534-8543, 2021. 2", + "[49] Lei Xiao, Salah Nouri, Joel Hegland, Alberto Garcia Garcia, and Douglas Lanman. Neuralpassthrough: Learned real-time view synthesis for vr. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2", + "[50] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 1, 3", + "[51] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 3, 6", + "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[53] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018. 2, 3, 4, 5" + ], + "bbox": [ + 501, + 92, + 893, + 838 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10464", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Tiled Multiplane Images for Practical 3D Photography/332ad0e6-c530-4dba-82ae-058961745db1_model.json b/2023/Tiled Multiplane Images for Practical 3D Photography/332ad0e6-c530-4dba-82ae-058961745db1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9f0f3a0291e83f9dc85b592eed513bb012879545 --- /dev/null +++ b/2023/Tiled Multiplane Images for Practical 3D Photography/332ad0e6-c530-4dba-82ae-058961745db1_model.json @@ -0,0 +1,2664 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.761, + 0.154 + ], + "angle": 0, + "content": "Tiled Multiplane Images for Practical 3D Photography" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.182, + 0.324, + 0.198 + ], + "angle": 0, + "content": "Numair Khan" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.203, + 0.354, + 0.216 + ], + "angle": 0, + "content": "numairkhan@meta.com" + }, + { + "type": "text", + "bbox": [ + 0.445, + 0.182, + 0.518, + 0.198 + ], + "angle": 0, + "content": "Lei Xiao" + }, + { + "type": "text", + "bbox": [ + 0.405, + 0.203, + 0.558, + 0.215 + ], + "angle": 0, + "content": "lei.xiao@meta.com" + }, + { + "type": "text", + "bbox": [ + 0.63, + 0.182, + 0.767, + 0.199 + ], + "angle": 0, + "content": "Douglas Lanman" + }, + { + "type": "text", + "bbox": [ + 0.609, + 0.203, + 0.789, + 0.216 + ], + "angle": 0, + "content": "doug.lanman@meta.com" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.251, + 0.871, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.457, + 0.893, + 0.517 + ], + "angle": 0, + "content": "Figure 1: An overview of our 3D photography method. Left: Using a single RGB image and an estimated monocular depth map, the scene is recreated as a tiled grid of many small MPs. Middle: A visualization of a \\(2 \\times 2\\) tiled grid of MPs, each with three RGBA layers. Right: Novel views are rendered by warping and compositing each MPI tile into the target camera's frustum. The \\(768 \\times 1152\\) pixel image shown is generated using a \\(7 \\times 11\\) grid of 4-layer MPs." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.531, + 0.314, + 0.545 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.474, + 0.873 + ], + "angle": 0, + "content": "The task of synthesizing novel views from a single image has useful applications in virtual reality and mobile computing, and a number of approaches to the problem have been proposed in recent years. A Multiplane Image (MPI) estimates the scene as a stack of RGBA layers, and can model complex appearance effects, anti-alias depth errors and synthesize soft edges better than methods that use textured meshes or layered depth images. And unlike neural radiance fields, an MPI can be efficiently rendered on graphics hardware. However, MPIs are highly redundant and require a large number of depth layers to achieve plausible results. Based on the observation that the depth complexity in local image regions is lower than that over the entire image, we split an MPI into many small, tiled regions, each with only a few depth planes. We call this representation a Tiled Multiplane Image (TMPI). We propose a method for generating a TMPI with adaptive depth planes for single-view 3D photography in the wild. Our synthesized results are comparable to state-of-the-art single-view MPI methods while having lower computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.531, + 0.631, + 0.545 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.775 + ], + "angle": 0, + "content": "The novel view synthesis (NVS) problem involves using a set of input images to generate views from new and unseen camera positions, allowing three dimensional interaction with photos. This is a long-studied problem, with early work relying on interpolation within dense structured image sets [20, 11, 8]. The specialized rigs commonly required to capture the large number of images restricted these methods to lab settings [46]. However, the potential applications offered by novel view synthesis on modern mobile and VR devices has kindled wide interest in the problem, and encouraged researchers to seek methods that make the technology more accessible. The term 3D photography refers to the use of novel view synthesis in everyday capture settings, often from a single image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Over the past few years a number of proposed scene representations have leveraged the great strides being made in learning-based techniques to achieve more accurate synthesis with fewer constraints. The most recent of these are neural radiance fields (NeRFs) [26, 50] which represent the scene as multi-layer perceptrons. Their results define the high bar of novel view synthesis. However, this high quality has a significant data and computational cost." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10454" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "An alternate representation, a multiplane image (MPI), defines the scene as a stack of fronto-parallel RGBA planes that can be warped and rendered into novel viewpoints [53, 26, 9]. An MPI offers the advantage of rendering speed, and suffers from less aliasing than mesh or point-based methods [38, 47, 29]. The latter characteristic is important for applications that require temporal stability. However, an MPI is a highly redundant scene representation: the number of RGBA planes required to capture and reconstruct all the depth variation in a scene can be quite high. Since most scenes have a larger amount of free space than occupied, most of the planes in an MPI are very sparse. This makes them inefficient as a representation [2], and expensive to generate, transmit, and store." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.473, + 0.59 + ], + "angle": 0, + "content": "In this paper we propose to address these shortcomings of multiplane images and develop a lightweight solution to the 3D photography problem that can be practically implemented on mobile and VR devices. Some examples of the applications we envision are 3D video conferencing, telepresence, and VR passthrough [49]. We show how subdividing the image plane into many small MPs with only a few planes in each, provides a more efficient representation from a computational and memory perspective. However, the naive approach of directly using existing MPI methods with tiles creates boundary artifacts in the novel views. This happens because the commonly used fixed spacing of MPI planes fails to capture the full depth range of a tile when the number of planes is small. Furthermore, it is sensitive to outliers in small regions. We propose a clustering-based approach using learnt confidence weights to predict per-tile MPI planes that better represents local depth features. Our method is lightweight and generates results comparable to the state-of-the-art in MPI-based 3D photography." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.591, + 0.438, + 0.605 + ], + "angle": 0, + "content": "In summary, the main contribution of this work are," + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.615, + 0.47, + 0.647 + ], + "angle": 0, + "content": "1. The demonstration of tiled multiplane images as a practical representation for view synthesis tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.655, + 0.468, + 0.687 + ], + "angle": 0, + "content": "2. A learning framework for generating tiled multiplane images from a single RGB input for 3D photography." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.696, + 0.458, + 0.712 + ], + "angle": 0, + "content": "3. A novel approach to adaptive MPI plane positioning." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.615, + 0.47, + 0.712 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.725, + 0.22, + 0.74 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The progenitors of current 3D photography were the early works on image-based rendering (IBR). These methods usually relied on interpolation within the convex hull of a large set of images to generate novel views. Levoy et al. [20] and Gortler et al. [11] proposed the canonical two-plane parameterization of light fields that renders novel views by quadrilinear interpolation. Gortler's method also provided an early demonstration of the use of geometric proxies to improve rendering quality. Davis et al.'s [8] work extended interpolation-based view synthesis to unstructured" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "images. The excellent analysis of plenoptic sampling done by Chai et al. [5] proved, however, that for large distances the number of images required for view synthesis by interpolation was impractically high. Consequently, the large majority of recent view synthesis methods have relied on learnt priors to overcome the high sampling requirements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.186, + 0.893, + 0.732 + ], + "angle": 0, + "content": "One of the corollaries of Chai et al.'s analysis was that the sampling requirements for view-interpolation are inversely related to the geometric information of the scene. Thus, many subsequent methods have relied on coarse geometric proxies to improve view synthesis quality [34, 35, 13]. In Mildenhall et al.'s [26] method this proxy takes the form of a multiplane image (MPI), which they use to achieve interpolation-based view synthesis that overcomes Chai et al.'s sampling limits. MPIs were first proposed by Zhou et al. [53] who used them for extrapolating novel views outside the convex hull of the input stereo cameras (Tucker and Snavely [42] observe that an MPI can be considered as an instance of Szeliski and Golland's [40] earlier \"stack of acetates\" volumetric model). Mildenhall et al. [26] and Srinivasan et al. [39], respectively, provide a theoretical analysis of the limits of view interpolation and extrapolation using MPIs. Flynn et al. [9] use learnt gradient descent to generate MPIs from multiple views. Attal et al. [1] and Broxton et al. [3] extend the MPI concept to concentric RGBA spheres that can be used for view synthesis in 360 degrees. Wizadwongsa et al. [48] and Li et al. [22] replace the discrete RGBA planes of MPIs with continuous neural surfaces to achieve higher quality results. Tucker and Snavely [42] use a scale-invariant method that allows them to learn strong data priors that can generate MPIs from a single view. Their approach is in a line of recent work dubbed 3D photography that aims for novel views from in-the-wild, single-view images. Li et al. [23], Han et al. [12], and Luvizon et al.'s [25] methods are MPI-based examples of this approach. An MPI, however, is an over-parameterized scene representation. Recent methods [24, 10, 25, 12] have sought to overcome this shortcoming to some extent through a more judicious placement of depth planes. Nonetheless, their high level of redundancy dilates their memory and computational footprint, and limits their wider adoption in mobile and AR/VR applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "For such use cases, a more efficient approach to novel view synthesis involves depth-based warping [6, 49], often followed by inpainting [18, 19]. Shih et al. [38] propose to guide inpainting in disoccluded regions of the warped view using a layered depth representation. Li et al. [21] apply this general approach to 360-degree input. Wiles et al. [47] use a depth map to generate a point cloud of neural features which can be projected and rendered in novel views using a generative network. Choi et al. [7] estimate a probability volume instead of a single depth map to handle uncertainty in difficult regions. Niklaus et al. [29] use seg" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10455" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.092, + 0.885, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.301, + 0.895, + 0.379 + ], + "angle": 0, + "content": "Figure 2: Tiled multiplane image generation for single-view 3D photography. Given an RGB image and depth from a monocular estimator, our method first generates a pixel-wise confidence map and a pre-processed depth map. These are used to predict \\( n \\) depth planes per tile through weighted \\( k \\)-means clustering (with \\( k = n \\)). The layer masks defined by pixel labels are used to peel RGBA layers, yielding a rough initial MPI. This is concatenated \\( (\\bigoplus) \\) to the depth and passed to a refinement network that generates the RGBA images of the final per-tile MPI." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.47, + 0.509 + ], + "angle": 0, + "content": "mentation to remove the geometric and semantic distortions from depth that often impair the rendering quality of this approach. Nonetheless, depth-based warping suffers from hard boundaries and is over-sensitive to errors in the depth estimate. While Jampani et al. [14] propose soft layering with alpha mattes to address this shortcoming, MPIs are inherently capable of handling such artifacts via blending." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.471, + 0.646 + ], + "angle": 0, + "content": "Finally, our review of view synthesis would not be complete without mentioning neural radiance fields (NeRFs) [27, 50] which have recently沦ued in popularity. While great strides are being made in improving the time and data requirements of NeRFs [51, 28], they remain expensive for interactive applications. Some recent work [33, 41, 43] shows that decomposing a single large radiance field into smaller sub-components can improve efficiency. This is similar to the approach we adopt for multiplane images." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.657, + 0.169, + 0.672 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.818 + ], + "angle": 0, + "content": "A traditional MPI represents the scene as a set of \\(N\\) fronto-parallel planes in the camera frustum of a reference view \\(\\mathcal{I}\\). Each plane is associated with an RGBA image. While it is possible to place the planes at any depth, they are usually arranged linearly in disparity (inverse depth) [53, 9, 26, 42]. A novel view \\(\\mathcal{I}_t\\) is rendered by warping the planes into the target camera's image space via a homography, and compositing them front-to-back using the over [31] operator:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.824, + 0.47, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {t} = \\sum_ {i = 1} ^ {N} \\left(\\alpha_ {i} c _ {i} \\prod_ {j = i + 1} ^ {N} (1 - \\alpha_ {i})\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) and \\(c_{i}\\) are the warped alpha and color channels, respectively, of the \\(i^{th}\\) plane." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.892, + 0.631 + ], + "angle": 0, + "content": "Both warping and compositing can be done very efficiently on graphics hardware, allowing real-time rendering of novel views [26]. Moreover, the alpha channel at each plane allows MPs to represent soft edges and anti-alias any errors in the scene reconstruction, leading to fewer perceptually objectionable artifacts than depth-based warping methods. However, the number of planes \\( N \\) required to capture all the depth variation in a scene is usually large, even though most planes are very sparse. We propose to overcome this shortcoming by representing the reference image \\( \\mathcal{I} \\) as a tiled grid of many small MPs (Figure 3). Given \\( \\mathcal{I} \\) and its depth map from a monocular depth estimator, our method predicts the placement of \\( n \\ll N \\) depth planes within each tiled region and uses this prediction to generate the RGBA images of the MPI in a single forward pass." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.638, + 0.839, + 0.656 + ], + "angle": 0, + "content": "3.1. Tiled Multiplane Image Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.661, + 0.892, + 0.812 + ], + "angle": 0, + "content": "Our scene representation is based on a set of \\(m\\) tiles, each representing a square sliding block of size \\(h\\) at 2D pixel locations \\(\\mathbf{x}_1, \\mathbf{x}_2, \\dots, \\mathbf{x}_m\\) in the source image \\(\\mathcal{I}\\). The locations \\(\\mathbf{x}_k\\) lie on a regular grid with spacing determined by some stride \\(r\\). Each tile consists of \\(n\\) front-parallel RGBA planes, the depth placement of which is not fixed but varies across tiles. We let \\(\\alpha_j^i\\), \\(c_j^i\\) and \\(d_j^i\\) denote, respectively, the alpha channel, the color channel, and depth of the \\(i\\)-th plane in the \\(j\\)-th tile. Then the tiled multiplane image (TMPI) representation \\(\\Gamma(\\mathcal{I})\\) of the image is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.818, + 0.892, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\Gamma (\\mathcal {I}) = \\left\\{\\left(\\alpha_ {j} ^ {i}, c _ {j} ^ {i}, d _ {j} ^ {i}, \\mathbf {x} _ {j}\\right) \\right\\} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.843, + 0.892, + 0.873 + ], + "angle": 0, + "content": "for all \\(i = 1,2,\\dots,n\\) and \\(j = 1,2,\\dots,m\\). Further, we define an ordering on this set of 4-tuplets as," + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.88, + 0.892, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\left(\\Gamma (\\mathcal {I}), \\leq\\right) = \\left\\{\\left(\\alpha , c, d, \\mathbf {x}\\right) _ {k = 1, \\dots , m n} \\mid d _ {k} \\leq d _ {k + 1} \\right\\} \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10456" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.473, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.47, + 0.301 + ], + "angle": 0, + "content": "Figure 3: (a) A traditional MPI uses five planes per pixel to represent this toy scene, even though no region has more than two overlapping objects. (b) A TMPI exploits the low local depth complexity and has only two planes per pixel." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.473, + 0.695 + ], + "angle": 0, + "content": "As with traditional MPIs, novel views are rendered in a differentiable manner by warping all planes into the target camera's image space. However, as the depth of the planes varies across the tiles and, hence, across the pixels of \\(\\mathcal{I}\\), a planar inverse warp on the target image plane cannot be directly computed via a homography. Instead, each plane must be warped in tile space via a homography computed using a shifted intrinsic matrix, and all the warped planes composited sequentially in the target view at their respective tile locations (Algorithm 1). While this makes the rendering of tiled multiplane images less efficient and somewhat less elegant than MPs (Equation 1), this is only true during training when differentiability is required. At inference, the tiles can be rendered as textured quads using hardware-accelerated rasterization making this stage of the pipeline as efficient as traditional MPs and with lower texture memory requirements. Thus, their compact form makes TMPIs well-suited for rendering over networks, or on mobile devices and VR headsets. The blending of many MPs locally is related to Mildenhall et al.'s [26] light field fusion. However, their method renders each MPI separately before blending the results with scalar weights. Our method operates at a much finer scale, and composites the planes of all MPs together to synthesize a novel view." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.336, + 0.726 + ], + "angle": 0, + "content": "3.2. Single-View 3D Photography" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We now describe our approach to generating tiled multiplane images from a single RGB input (Figure 2). Broadly, our method splits the image plane into a regular tiled grid of learnt confidence and outlier-corrected depth. For each tile, the placement of \\( n \\) fronto-parallel depth planes is determined by clustering pixel depth values weighted by the predicted confidence estimates. This latter step is motivated by the fact that with a small depth plane budget \\( n \\), the commonly used equal spacing in disparity [53, 9, 26, 42] is wasteful. Thus, the goal is to predict the planes that optimally represent all depth variation within a tile. Using the" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.095, + 0.85, + 0.126 + ], + "angle": 0, + "content": "Algorithm 1: Differentiable view synthesis using TMPIs. Angled brackets denote pixel indexing." + }, + { + "type": "algorithm", + "bbox": [ + 0.52, + 0.128, + 0.852, + 0.467 + ], + "angle": 0, + "content": "RenderTMPI \\((\\Gamma (\\mathcal{I}),\\leq)\\) \\(\\pmb {R},t,\\pmb {K})\\) Input: \\((\\Gamma (\\mathcal{I}),\\leq)\\) : ordered TMPI planes R: relative rotation of novel view t: relative translation K: camera intrinsics Output: Novel view \\(\\mathcal{I}_t\\in \\mathbb{R}^{3\\times H\\times W}\\) \\(\\mathcal{I}_t\\gets \\mathbb{0}^{3\\times H\\times W};\\mathcal{T}_t\\gets \\mathbb{1}^{H\\times W};\\) foreach \\((\\alpha ,c,d,\\pmb {x})\\in \\left(\\Gamma (\\mathcal{I}),\\leq\\right)\\) do \\(\\hat{\\mathbf{K}}\\gets \\mathbf{K} - \\left[ \\begin{array}{cc}\\mathbf{I} & \\mathbf{x}\\\\ 0 & 1 \\end{array} \\right];\\) foreach \\(u\\in [1,\\dots,s]\\times [1,\\dots,s]\\) do n \\(\\left[\\begin{array}{l}\\mathbf{u}_s\\\\ 1\\end{array}\\right]\\gets \\hat{\\mathbf{K}} (\\mathbf{R} - \\mathbf{tn}^T /d)\\hat{\\mathbf{K}}^{-1}\\left[\\begin{array}{l}\\mathbf{u}\\\\ 1\\end{array}\\right];\\) \\(w\\gets \\alpha \\langle \\mathbf{u}_s\\rangle \\mathcal{T}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle ;\\) \\(\\mathcal{I}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle \\gets \\mathcal{I}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle +w c\\langle \\mathbf{u}_s\\rangle ;\\) \\(\\mathcal{T}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle \\gets \\mathcal{T}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle (1 - \\alpha \\langle \\mathbf{u}_s\\rangle);\\) end end return I" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.596 + ], + "angle": 0, + "content": "predicted planes, a fully convolutional network generates the \\( n \\) RGBA images that constitute a per-tile MPI. Unlike Han et al.'s [12] adaptive plane method, we generate the RGBA images in a single forward pass. The resulting TMPI is rendered as a set of textured quads using the rasterization pipeline of graphics hardware." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.893, + 0.765 + ], + "angle": 0, + "content": "In more detail, given the source image \\(\\mathcal{I}\\), we first obtain a depth map \\(\\mathcal{Z}\\) for it using a monocular depth estimator. A two-headed U-Net \\(\\Theta(\\cdot)\\) then predicts a confidence map \\(\\mathcal{C}\\) along with denoised depth \\(\\mathcal{D}\\). The goal is to learn a representation that ameliorates the sensitivity of the subsequent \\(k\\)-means clustering step to outliers. The joint prediction of confidence and depth is similar to the depth-routing of Weder et al. [45] and the aleotoric uncertainty estimation of [15]. The predicted depth and confidence, and the original color image are then unfolded into a set of \\(m\\) square sliding blocks of size \\(h\\) and stride \\(r\\): \\(\\{(\\mathcal{D}^i,\\mathcal{C}^i,\\mathcal{I}^i)_{i=1,\\dots,m}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Running \\(\\Theta(\\cdot)\\) on \\(\\mathcal{I}\\) and \\(\\mathcal{Z}\\) rather than individual tiles allows it to consider non-local features and avoid undesirable tiling artifacts. Additionally, setting \\(r < h\\) in the unfolding step allows neighboring tiles to overlap. This prevents gaps along tile boundaries and also regularizes per-tile operations across neighbors. However, it also increases the total number of tiles and, thus, the computational requirements. We empirically determine a stride value of \\(r = h - h/8\\) for a good balance between quality and computational efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10457" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.348 + ], + "angle": 0, + "content": "Per-Tile Planes Prediction: Next, we predict the \\( n \\) depth planes \\( \\{z_{j=1,\\dots,n}^{i}\\} \\) that optimally represent the features of the \\( i \\)-th tile. The common approach of spacing the planes linearly in disparity grows inaccurate as \\( n \\) becomes small. Luvizon et al. [25] place the planes at depth discontinuities identified via the histogram of depth values. However, their approach is sensitive to parameter settings and fails for smooth surfaces which have no discontinuities. A learning-based approach is adopted by Han et al. [12] and Li et al. [23]. The former use multi-headed self-attention to adjust a linear placement. While capable of modeling interplane interactions, their method is computationally expensive (Table 2). The latter uses a CNN to directly predict the planes. However, we found that without strong regularization a direct approach lacks topological order and has a strong bias towards a fixed placement. An adversarial loss helps improve this but makes the training more unstable." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.348, + 0.47, + 0.395 + ], + "angle": 0, + "content": "We observe that as depth is known, plane positioning can be posed as a simple clustering problem. Thus, we predict \\(\\{z_j^i\\}\\) using \\(k\\)-means clustering on the depth in each tile \\(\\mathcal{D}^i\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.473, + 0.65 + ], + "angle": 0, + "content": "Standard \\(k\\)-means is sensitive to outliers and can generate significantly different plane predictions across neighboring tiles causing artifacts in novel views. Advantageously, along with the \\(n\\) depth planes, clustering also assigns a label to each input pixel, thereby generating a label map that represents discretized depth. Furthermore, the cluster centers of \\(k\\)-means are differentiable with respect to the input samples. Thus, we address the outlier problem by training \\(\\Theta(\\cdot)\\) to filter the input through a self-supervised reconstruction loss on the discretized depth map generated by a weighted \\(k\\)-means. In weighted \\(k\\)-means, the cluster centers are updated each iteration using the confidence-weighted mean of the constituent samples. Since we do not directly supervise the depth output of \\(\\Theta(\\cdot)\\), the network can go beyond outlier-filtering to learn any modifications that improves the discrete reconstruction, and consequently optimizes the placement of the \\(n\\) depth planes within each tile." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.473, + 0.818 + ], + "angle": 0, + "content": "Per-Tile MPI Generation: Given \\(\\{z_j^i\\}\\) and the discrete depth map, we estimate a preliminary MPI per tile by peeling RGBA layers from \\(\\mathcal{I}^i\\) using the discrete labels as an alpha mask. The masked RGB regions of each plane are inpainted by upsampling valid values from a Gaussian pyramid. A second network \\(\\Psi (\\cdot)\\) then refines these estimates to generate the final \\(n\\) RGBA images of the MPI for each tile. Following Zhou et al. [53] and Tucker and Snavely [42], we represent the RGBA output as a pixel-wise blend of the input image \\(\\mathcal{I}^i\\) and a learnt background. However, unlike these works we predict a background image \\(\\mathcal{B}_j^i\\) per plane:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.824, + 0.47, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} _ {j} ^ {i} = \\prod_ {k > j} \\left(1 - \\alpha_ {k} ^ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.86, + 0.469, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {j} ^ {i} = \\mathcal {W} _ {j} ^ {i} \\mathcal {I} ^ {i} + (1 - \\mathcal {W} _ {j} ^ {i}) \\mathcal {B} _ {j} ^ {i} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.885, + 0.432, + 0.903 + ], + "angle": 0, + "content": "Where \\(\\alpha_{k}^{i}\\) is the predicted alpha value for each plane." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.089, + 0.885, + 0.344 + ], + "angle": 0, + "content": "
Spaces Dataset
MethodPSNR↑SSIM↑LPIPS↓L1↓
SVMPI [42]25.420.7480.2100.040
VMPI [23]22.370.6360.2680.057
MINE [22]24.020.7020.2290.048
AdaMPI [12]26.170.7030.2290.047
Ours24.930.7500.1750.037
Tanks & Temples Dataset
MethodPSNR↑SSIM↑LPIPS↓L1↓
SVMPI17.850.5300.3700.082
VMPI16.320.4630.3950.103
MINE17.230.5060.3660.088
AdaMPI18.620.5650.2700.073
Ours18.690.5690.2670.073
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.427 + ], + "angle": 0, + "content": "Table 1: Quantitative evaluation of view synthesis results on the Spaces and Tanks and Temples multi-view datasets. SVMPI, MINE and AdaMPI use 32 MPI depth planes; VMPI uses 8; our approach uses 4 planes per image tile." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.452, + 0.691, + 0.469 + ], + "angle": 0, + "content": "4. Training Procedure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.704 + ], + "angle": 0, + "content": "We train \\(\\Theta(\\cdot)\\) in a self-supervised manner by minimizing the L1 loss between the input depth \\(\\mathcal{Z}\\) and the folded discrete depth maps produced per tile by weighted \\(k\\)-means clustering. Then, we freeze \\(\\Theta(\\cdot)\\) and train the MPI generation network \\(\\Psi(\\cdot)\\) on a novel view synthesis task by following Han et al.'s [12] warp-back strategy to generate pseudo ground truth multi-view training data. This involves warping single-view images into a target camera using monocular depth, and inpainting disocclusion holes with a specially trained network. The view-synthesis training objective is a combination of VGG, structural similarity [44], and L1 losses on the synthesized color image \\(\\mathcal{I}_t\\), weighed as 0.1, 0.25, and 1.0 respectively. For both networks, we use the 111K images of the COCO dataset [4] and the monocular depth method of Ranftl et al. [32]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.633, + 0.734 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.661, + 0.758 + ], + "angle": 0, + "content": "5.1. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our method is implemented in PyTorch and trained on eight Nvidia Tesla V100 GPUs. We use \\(256 \\times 384\\) images and accumulate gradients across eight mini-batches of 16 samples each. For both networks \\(\\Theta(\\cdot)\\) and \\(\\Psi(\\cdot)\\), we use the Adam optimizer with a learning rate of \\(1 \\times 10^{-3}\\) and a cosine annealing schedule with restarts every 200 epochs. A vectorized implementation of the differentiable TMPI renderer (Algorithm 1) runs at \\(\\sim 100\\mathrm{ms} / \\mathrm{mini}\\)-batch allowing efficient parallel computation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10458" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.468, + 0.23 + ], + "angle": 0, + "content": "
Params.↓(M)GMAC↓Runtime (ms)Peak (GB)Space (MB)↓
DPT*123.0110314.39-
SVMPI43.558.01114.5326.9
MINE38.12501104.91107
VMPI†4.3152.396.04.5713.5
AdaMPI†19.02883505.94107
Ours†6.4357.091.63.205.25
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.081, + 0.232, + 0.405, + 0.245 + ], + "angle": 0, + "content": "*Monocular depth method. † Method uses monocular depth." + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.27, + 0.47, + 0.362 + ], + "angle": 0, + "content": "Table 2: Run-time, memory and complexity evaluation of all methods with a single \\(350 \\times 630\\) RGB image on an NVIDIA GeForce RTX 3080 GPU. The choice of resolution is dictated by the baseline methods which run out of memory for larger images on the specified hardware. GMACS are Giga Multiply-Accumulate ops./second." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.388, + 0.186, + 0.403 + ], + "angle": 0, + "content": "5.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.47, + 0.593 + ], + "angle": 0, + "content": "We compare our approach to four state-of-the-art single-view 3D photography methods based on multiplane images: Tucker et al. [42] (SVMPI), Li et al. [23] (VMPI), Li et al. [22] (MINE) and Han et al. [12] (AdaMPI). Like us, VMPI and AdaMPI use a monocular estimator to recover depth as the first step of their pipeline. The input to all methods is the same, however — a single unconstrained RGB image — and so we evaluate them as end-to-end 3D photography approaches. Nonetheless, we do note the additional depth estimation step when evaluating computational and memory performance (Table 2). We use \\( N = 32 \\) planes for all baselines except VMPI, which is designed for \\( N = 8 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.594, + 0.47, + 0.73 + ], + "angle": 0, + "content": "We do not compare to single-view methods based on neural radiance fields [51] or the recent work of Nicklaus et al. [29] and Shih et al. [38] as these are too computationally intensive for our intended use cases on mobile and VR devices. While recent work on NeRFs has demonstrated impressive rendering speeds, training remains expensive, and further, requires a large number of input views and a static scene. Jampani et al.'s [14] method, though not MPI-based, is related. But the authors have not released their code." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.24, + 0.758 + ], + "angle": 0, + "content": "5.3. Testing Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We test all methods on the Spaces [9], and Tanks and Temples [17] datasets. Spaces consists of 100 indoor and outdoor scenes captured using a purpose-built, 16-camera rig. For Tanks and Temples we use the Intermediate split which has uniformly sampled frames from eight high-resolution videos of more challenging outdoor environments. We compute camera poses and depth maps for all scenes using COLMAP [37, 36]. The depth maps are required to resolve the scale ambiguity of monocular depth" + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.089, + 0.86, + 0.282 + ], + "angle": 0, + "content": "
Spaces Dataset
VariantMAE ↓MSE ↓Q25 ↓
Vanilla k-means35.32.0014.0
Linear plane spacing48.73.8019.3
Ours27.11.2010.6
Tanks& Temples Dataset
VariantMAE ↓MSE ↓Q25 ↓
Vanilla k-means35.72.0014.2
Linear plane spacing46.03.4018.1
Ours27.01.2010.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.305, + 0.892, + 0.336 + ], + "angle": 0, + "content": "Table 3: Evaluating the reconstruction error of different depth discretization approaches. All values are \\(\\times 10^{-3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.458 + ], + "angle": 0, + "content": "for correct reprojection to target views. We randomly select 1000 source views from each dataset scaled to \\(350 \\times 630\\) and use the next image in capture sequence as the target for view synthesis. This choice of resolution is dictated by the baseline methods which run out of memory for larger images on the specified hardware. We present high resolution results for our method on the Davis dataset [30] in Figure 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.467, + 0.684, + 0.482 + ], + "angle": 0, + "content": "5.4. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.626 + ], + "angle": 0, + "content": "We evaluate the rendered views quantitatively on four metrics: Peak-Signal to Noise Ratio (PSNR), Structural Similarity [44] (SSIM), Learned Perceptual Image Patch Similarity [52] (LPIPS) with a VGG-16 backbone, and the mean absolute error (L1). Following previous work [12, 42, 23], we crop \\(15\\%\\) of the image around the edge to account for disocclusions. Further, if more than \\(15\\%\\) of the remaining pixels are blank in a view synthesized by any method, we discard the result across all baselines." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.593, + 0.651 + ], + "angle": 0, + "content": "5.5. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Table 1 presents quantitative evaluation of all methods on the two test sets. Qualitative results are shown in Figure 4 for Tanks and Temples and Figure 9 for Spaces. Our approach uses \\( n = 4 \\) depth planes with a tile size of \\( h = 64 \\). The competitive performance of our method despite having much fewer planes per tile can be attributed to the adaptive placement of depth planes in each tile which allows it to effectively cover a larger depth range than the monolithic MPIs of the baselines, Memory and computational performance is evaluated in Table 2. VMPI, AdaMPI and our method have the additional overhead of monocular depth estimation using Ranftl et al.'s [32] DPT. Our approach has lower runtime, peak memory and space requirements than AdaMPI while achieving similar quality results." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We evaluate our plane placement strategy in Table 3. We measure the reconstruction quality of the discretized depth" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10459" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.088, + 0.359, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.206, + 0.266, + 0.217 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.089, + 0.633, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.206, + 0.516, + 0.217 + ], + "angle": 0, + "content": "SVMPI" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.892, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.206, + 0.796, + 0.217 + ], + "angle": 0, + "content": "VMPI" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.219, + 0.358, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.341, + 0.242, + 0.352 + ], + "angle": 0, + "content": "MINE" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.219, + 0.633, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.34, + 0.52, + 0.35 + ], + "angle": 0, + "content": "AdaMPI" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.219, + 0.892, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.34, + 0.796, + 0.35 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.361, + 0.358, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.48, + 0.265, + 0.491 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.361, + 0.633, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.48, + 0.516, + 0.49 + ], + "angle": 0, + "content": "SVMPI" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.361, + 0.892, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.48, + 0.796, + 0.49 + ], + "angle": 0, + "content": "VMPI" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.495, + 0.358, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.615, + 0.241, + 0.625 + ], + "angle": 0, + "content": "MINE" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.495, + 0.633, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.613, + 0.519, + 0.623 + ], + "angle": 0, + "content": "AdaMPI" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.495, + 0.892, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.614, + 0.797, + 0.624 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.636, + 0.895, + 0.667 + ], + "angle": 0, + "content": "Figure 4: Comparing the novel view synthesis results of the baseline methods and our approach on the Tanks and Temples dataset. Our results are better than SVMPI, VMPI and MINE, and comparable to AdaMPI while using far fewer depth planes." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.677, + 0.284, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.677, + 0.486, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.677, + 0.688, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.677, + 0.892, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.768, + 0.283, + 0.857 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.768, + 0.486, + 0.857 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.768, + 0.688, + 0.857 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.768, + 0.892, + 0.857 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.876, + 0.867, + 0.892 + ], + "angle": 0, + "content": "Figure 5: View synthesis results of our method on the HD \\((1080\\times 1920)\\) Davis [30] dataset. Original views are inset." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10460" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.473, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.439, + 0.473, + 0.531 + ], + "angle": 0, + "content": "Figure 6: Top to bottom: Input image (Davis dataset); inverse monocular depth from DPT [32]; naive discretization by spacing four planes linearly in the inverse-depth range of each tile; our weighted clustering-based discretization with four planes. Our representation shows finer variation on receding surfaces and suffers fewer tiling artifacts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.674 + ], + "angle": 0, + "content": "map defined by the planes in each tile on a monocular depth input that is perturbed by a small amount of Gaussian noise \\((\\lambda = 0, \\sigma^2 = 1 \\times 10^{-3})\\). We compare our approach to naive linear spacing of depth planes in inverse disparity space, and to vanilla unweighted \\(k\\)-means. Our method is robust to outliers and yields a much better reconstruction. Figure 6 provides qualitative comparison with linear spacing on samples from the Davis dataset [30]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.829 + ], + "angle": 0, + "content": "Figure 7 evaluates the effect of tile size and number of planes on the quality of view synthesis. In general, the results uphold the intuition that small tiles and a large number of planes improve quality. This trend is less clear, however, for number of planes \\(>8\\). This would seem to support the observation of Khakhulin et al. [16] and Hu et al. [13] that reductive models are unable to handle redundant geometry effectively. Moreover, the model's complexity increases proportionately with the number of depth planes, leading to slower convergence for the same number of training steps." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.201, + 0.86 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As previously observed, a tiled multiplane image cannot exploit the elegant warping and compositing equations of a" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.091, + 0.868, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.404 + ], + "angle": 0, + "content": "Figure 7: Evaluating the effect of tile size and number of depth planes on view synthesis quality. In general, the results uphold the intuition that small tiles and a large number of planes improve quality. The trend, however, is less clear when the number of planes \\(>8\\), indicating that the model may be unable to handle redundant geometry effectively." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.415, + 0.892, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.523, + 0.892, + 0.555 + ], + "angle": 0, + "content": "Figure 8: Failure cases of our approach: in some cases, fine features are inconsistently reconstructed across tiles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.625 + ], + "angle": 0, + "content": "traditional MPI for differentiable rendering during training. Furthermore, in some cases our method fails to reconstruct thin features consistently across tiles (Figure 8)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.62, + 0.652 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.661, + 0.892, + 0.753 + ], + "angle": 0, + "content": "We present a method for estimating tiled multiplane images from a single RGB input for 3D photography. This includes a novel approach to adaptively spacing a small number of depth planes within an MPI tile to better represent local features. Our method is lightweight, and points a path to realizing novel view synthesis on mobile and VR devices." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.765, + 0.597, + 0.78 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.789, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[1] Benjamin Attal, Selena Ling, Aaron Gokaslan, Christian Richardt, and James Tompkin. MatryODShka: Real-time 6DoF video view synthesis using multi-sphere images. In European Conference on Computer Vision (ECCV), Aug. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.859, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[2] Horace B Barlow et al. Possible principles underlying the transformation of sensory messages. Sensory communication, 1(01):217-233, 1961. 2" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.789, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10461" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.089, + 0.373, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.188, + 0.313, + 0.201 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.089, + 0.624, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.188, + 0.522, + 0.201 + ], + "angle": 0, + "content": "SVMPI" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.089, + 0.851, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.188, + 0.77, + 0.2 + ], + "angle": 0, + "content": "VMPI" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.203, + 0.373, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.312, + 0.282, + 0.324 + ], + "angle": 0, + "content": "MINE" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.203, + 0.624, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.312, + 0.528, + 0.324 + ], + "angle": 0, + "content": "AdaMPI" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.203, + 0.851, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.312, + 0.768, + 0.324 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.339, + 0.374, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.451, + 0.313, + 0.464 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.339, + 0.624, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.451, + 0.522, + 0.463 + ], + "angle": 0, + "content": "SVMPI" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.339, + 0.851, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.451, + 0.77, + 0.463 + ], + "angle": 0, + "content": "VMPI" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.465, + 0.373, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.576, + 0.282, + 0.588 + ], + "angle": 0, + "content": "MINE" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.465, + 0.624, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.576, + 0.528, + 0.588 + ], + "angle": 0, + "content": "AdaMPI" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.465, + 0.851, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.576, + 0.768, + 0.587 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.601, + 0.373, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.709, + 0.313, + 0.721 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.601, + 0.625, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.709, + 0.522, + 0.721 + ], + "angle": 0, + "content": "SVMPI" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.601, + 0.851, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.708, + 0.77, + 0.72 + ], + "angle": 0, + "content": "VMPI" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.723, + 0.373, + 0.831 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.833, + 0.282, + 0.845 + ], + "angle": 0, + "content": "MINE" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.723, + 0.624, + 0.831 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.833, + 0.528, + 0.845 + ], + "angle": 0, + "content": "AdaMPI" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.723, + 0.851, + 0.831 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.833, + 0.768, + 0.845 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.865, + 0.872, + 0.881 + ], + "angle": 0, + "content": "Figure 9: Comparing the novel view synthesis results of the baseline methods and our approach on the Spaces dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10462" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[3] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.162, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[5] Jin-Xiang Chai, Xin Tong, Shing-Chow Chan, and Heung-Yeung Shum. Plenoptic sampling. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 307-318, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.471, + 0.344 + ], + "angle": 0, + "content": "[6] Gaurav Chaurasia, Arthur Nieuwoudt, Alexandru-Eugen Ichim, Richard Szeliski, and Alexander Sorkine-Hornung. Passthrough+ real-time stereoscopic view synthesis for mobile mixed reality. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1-17, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[7] Inchang Choi, Orazio Gallo, Alejandro Troccoli, Min H Kim, and Jan Kautz. Extreme view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7781-7790, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.4, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[8] Abe Davis, Marc Levoy, and Fredo Durand. Unstructured light fields. In Computer Graphics Forum, volume 31, pages 305-314. Wiley Online Library, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.441, + 0.471, + 0.523 + ], + "angle": 0, + "content": "[9] John Flynn, Michael Broxton, Paul Debevec, Matthew DuVall, Graham Fyffe, Ryan Overbeck, Noah Snavely, and Richard Tucker. Deepview: View synthesis with learned gradient descent. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2367-2376, 2019. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.525, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[10] Sushobhan Ghosh, Zhaoyang Lv, Nathan Matsuda, Lei Xiao, Andrew Berkovich, and Oliver Cossairt. Liveview: dynamic target-centered mpi for view synthesis. arXiv preprint arXiv:2107.05113, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.581, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[11] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 43-54, 1996. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.636, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[12] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-view synthesis in the wild with learned adaptive multi-plane images. In ACM SIGGRAPH, 2022. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.679, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[13] Ronghang Hu, Nikhila Ravi, Alexander C Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12528-12537, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.748, + 0.471, + 0.83 + ], + "angle": 0, + "content": "[14] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, and Ce Liu. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.831, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[16] Taras Khakhulin, Denis Korzhenkov, Pavel Solovev, Gleb Sterkin, Andrei-Timotei Ardelean, and Victor Lempitsky." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.093, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.135 + ], + "angle": 0, + "content": "Stereo magnification with multi-layer images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8687-8696, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.893, + 0.19 + ], + "angle": 0, + "content": "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.893, + 0.261 + ], + "angle": 0, + "content": "[18] Johannes Kopf, Suhib Alsisan, Francis Ge, Yangming Chong, Kevin Matzen, Ocean Quigley, Josh Patterson, Jossie Tirado, Shu Wu, and Michael F Cohen. Practical 3d photography. In Proceedings of CVPR Workshops, volume 1, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.893, + 0.332 + ], + "angle": 0, + "content": "[19] Johannes Kopf, Kevin Matzen, Suhib Alsisan, Ocean Quigley, Francis Ge, Yangming Chong, Josh Patterson, Jan-Michael Frahm, Shu Wu, Matthew Yu, et al. One shot 3d photography. ACM Transactions on Graphics (TOG), 39(4):76-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.893, + 0.376 + ], + "angle": 0, + "content": "[20] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.893, + 0.459 + ], + "angle": 0, + "content": "[21] David Li, Yinda Zhang, Christian Hane, Danhang Tang, Amitabh Varshney, and Ruofei Du. Omnisyn: Synthesizing 360 videos with wide-baseline panoramas. In 2022 IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops (VRW), pages 670-671. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.461, + 0.893, + 0.516 + ], + "angle": 0, + "content": "[22] Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. MINE: Towards continuous depth spi with nerf for novel view synthesis. In ICCV, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.893, + 0.561 + ], + "angle": 0, + "content": "[23] Qinbo Li and Nima Khademi Kalantari. Synthesizing light field from a single image with variable mpi and two network fusion. ACM Trans. Graph., 39(6):229-1, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.893, + 0.631 + ], + "angle": 0, + "content": "[24] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4521-4530, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.893, + 0.73 + ], + "angle": 0, + "content": "[25] Diogo C Luvizon, Gustavo Sutter P Carvalho, Andreza A dos Santos, Jhonatas S Conceicao, Jose L Flores-Campana, Luis GL Decker, Marcos R Souza, Helio Pedrini, Antonio Joia, and Otavio AB Penatti. Adaptive multiplane image generation from a single internet picture. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2556-2565, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.893, + 0.801 + ], + "angle": 0, + "content": "[26] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.893, + 0.87 + ], + "angle": 0, + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.872, + 0.893, + 0.903 + ], + "angle": 0, + "content": "[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a mul" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.893, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10463" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "tiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.163 + ], + "angle": 0, + "content": "[29] Simon Niklaus, Long Mai, Jimei Yang, and Feng Liu. 3d ken burns effect from a single image. ACM Transactions on Graphics (ToG), 38(6):1-15, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[30] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.263 + ], + "angle": 0, + "content": "[31] Thomas Porter and Tom Duff. Compositing digital images. In Proceedings of the 11th annual conference on Computer graphics and interactive techniques, pages 253-259, 1984. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.471, + 0.32 + ], + "angle": 0, + "content": "[32] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[33] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[34] Gernot Riegler and Vladlen Koltun. Free view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 623-640. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[35] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12216-12225, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.469, + 0.547 + ], + "angle": 0, + "content": "[36] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.549, + 0.469, + 0.603 + ], + "angle": 0, + "content": "[37] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[38] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.662, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[39] Pratul P Srinivasan, Richard Tucker, Jonathan T Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snively. Pushing the boundaries of view extrapolation with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 175-184, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[40] Richard Szeliski and Polina Golland. Stereo matching with transparency and matting. In Sixth International Conference on Computer Vision (IEEE Cat. No. 98CH36271), pages 517-524. IEEE, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[41] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[42] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In The IEEE Conference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[43] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[44] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[45] Silvan Weder, Johannes Schonberger, Marc Pollefeys, and Martin R Oswald. Routedfusion: Learning real-time depth map fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4887-4897, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.319, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[46] Bennett Wilburn, Neel Joshi, Vaibhav Vaish, Eino-Ville Talvala, Emilio Antunez, Adam Barth, Andrew Adams, Mark Horowitz, and Marc Levoy. High performance imaging using large camera arrays. In ACM SIGGRAPH 2005 Papers, pages 765-776. 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.458 + ], + "angle": 0, + "content": "[47] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.461, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[48] Suttisak Wizadwongsa, Pakkapon Phongthawee, Jiraphon Yenphraphai, and Supasorn Suwajanakorn. Nex: Real-time view synthesis with neural basis expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8534-8543, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.531, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[49] Lei Xiao, Salah Nouri, Joel Hegland, Alberto Garcia Garcia, and Douglas Lanman. Neuralpassthrough: Learned real-time view synthesis for vr. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.588, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[50] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.672, + 0.892, + 0.739 + ], + "angle": 0, + "content": "[51] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.742, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.786, + 0.892, + 0.839 + ], + "angle": 0, + "content": "[53] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018. 2, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10464" + } + ] +] \ No newline at end of file diff --git a/2023/Tiled Multiplane Images for Practical 3D Photography/332ad0e6-c530-4dba-82ae-058961745db1_origin.pdf b/2023/Tiled Multiplane Images for Practical 3D Photography/332ad0e6-c530-4dba-82ae-058961745db1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c4d428e4dc59f868fc128245b2817b4cc849d4c3 --- /dev/null +++ b/2023/Tiled Multiplane Images for Practical 3D Photography/332ad0e6-c530-4dba-82ae-058961745db1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2939f0973b4cd38173e16f7275efd025c0691db8d9441e0d6af4a61f2a2e930b +size 3113716 diff --git a/2023/Tiled Multiplane Images for Practical 3D Photography/full.md b/2023/Tiled Multiplane Images for Practical 3D Photography/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4b83dbb2b03d659a365a679487e875cd74c81345 --- /dev/null +++ b/2023/Tiled Multiplane Images for Practical 3D Photography/full.md @@ -0,0 +1,356 @@ +# Tiled Multiplane Images for Practical 3D Photography + +Numair Khan + +numairkhan@meta.com + +Lei Xiao + +lei.xiao@meta.com + +Douglas Lanman + +doug.lanman@meta.com + +![](images/fc5342a98343c8a0d9d0a150e3e00703918263f40220310233be39068a3d9b6e.jpg) +Figure 1: An overview of our 3D photography method. Left: Using a single RGB image and an estimated monocular depth map, the scene is recreated as a tiled grid of many small MPs. Middle: A visualization of a $2 \times 2$ tiled grid of MPs, each with three RGBA layers. Right: Novel views are rendered by warping and compositing each MPI tile into the target camera's frustum. The $768 \times 1152$ pixel image shown is generated using a $7 \times 11$ grid of 4-layer MPs. + +# Abstract + +The task of synthesizing novel views from a single image has useful applications in virtual reality and mobile computing, and a number of approaches to the problem have been proposed in recent years. A Multiplane Image (MPI) estimates the scene as a stack of RGBA layers, and can model complex appearance effects, anti-alias depth errors and synthesize soft edges better than methods that use textured meshes or layered depth images. And unlike neural radiance fields, an MPI can be efficiently rendered on graphics hardware. However, MPIs are highly redundant and require a large number of depth layers to achieve plausible results. Based on the observation that the depth complexity in local image regions is lower than that over the entire image, we split an MPI into many small, tiled regions, each with only a few depth planes. We call this representation a Tiled Multiplane Image (TMPI). We propose a method for generating a TMPI with adaptive depth planes for single-view 3D photography in the wild. Our synthesized results are comparable to state-of-the-art single-view MPI methods while having lower computational overhead. + +# 1. Introduction + +The novel view synthesis (NVS) problem involves using a set of input images to generate views from new and unseen camera positions, allowing three dimensional interaction with photos. This is a long-studied problem, with early work relying on interpolation within dense structured image sets [20, 11, 8]. The specialized rigs commonly required to capture the large number of images restricted these methods to lab settings [46]. However, the potential applications offered by novel view synthesis on modern mobile and VR devices has kindled wide interest in the problem, and encouraged researchers to seek methods that make the technology more accessible. The term 3D photography refers to the use of novel view synthesis in everyday capture settings, often from a single image. + +Over the past few years a number of proposed scene representations have leveraged the great strides being made in learning-based techniques to achieve more accurate synthesis with fewer constraints. The most recent of these are neural radiance fields (NeRFs) [26, 50] which represent the scene as multi-layer perceptrons. Their results define the high bar of novel view synthesis. However, this high quality has a significant data and computational cost. + +An alternate representation, a multiplane image (MPI), defines the scene as a stack of fronto-parallel RGBA planes that can be warped and rendered into novel viewpoints [53, 26, 9]. An MPI offers the advantage of rendering speed, and suffers from less aliasing than mesh or point-based methods [38, 47, 29]. The latter characteristic is important for applications that require temporal stability. However, an MPI is a highly redundant scene representation: the number of RGBA planes required to capture and reconstruct all the depth variation in a scene can be quite high. Since most scenes have a larger amount of free space than occupied, most of the planes in an MPI are very sparse. This makes them inefficient as a representation [2], and expensive to generate, transmit, and store. + +In this paper we propose to address these shortcomings of multiplane images and develop a lightweight solution to the 3D photography problem that can be practically implemented on mobile and VR devices. Some examples of the applications we envision are 3D video conferencing, telepresence, and VR passthrough [49]. We show how subdividing the image plane into many small MPs with only a few planes in each, provides a more efficient representation from a computational and memory perspective. However, the naive approach of directly using existing MPI methods with tiles creates boundary artifacts in the novel views. This happens because the commonly used fixed spacing of MPI planes fails to capture the full depth range of a tile when the number of planes is small. Furthermore, it is sensitive to outliers in small regions. We propose a clustering-based approach using learnt confidence weights to predict per-tile MPI planes that better represents local depth features. Our method is lightweight and generates results comparable to the state-of-the-art in MPI-based 3D photography. + +In summary, the main contribution of this work are, + +1. The demonstration of tiled multiplane images as a practical representation for view synthesis tasks. +2. A learning framework for generating tiled multiplane images from a single RGB input for 3D photography. +3. A novel approach to adaptive MPI plane positioning. + +# 2. Related Work + +The progenitors of current 3D photography were the early works on image-based rendering (IBR). These methods usually relied on interpolation within the convex hull of a large set of images to generate novel views. Levoy et al. [20] and Gortler et al. [11] proposed the canonical two-plane parameterization of light fields that renders novel views by quadrilinear interpolation. Gortler's method also provided an early demonstration of the use of geometric proxies to improve rendering quality. Davis et al.'s [8] work extended interpolation-based view synthesis to unstructured + +images. The excellent analysis of plenoptic sampling done by Chai et al. [5] proved, however, that for large distances the number of images required for view synthesis by interpolation was impractically high. Consequently, the large majority of recent view synthesis methods have relied on learnt priors to overcome the high sampling requirements. + +One of the corollaries of Chai et al.'s analysis was that the sampling requirements for view-interpolation are inversely related to the geometric information of the scene. Thus, many subsequent methods have relied on coarse geometric proxies to improve view synthesis quality [34, 35, 13]. In Mildenhall et al.'s [26] method this proxy takes the form of a multiplane image (MPI), which they use to achieve interpolation-based view synthesis that overcomes Chai et al.'s sampling limits. MPIs were first proposed by Zhou et al. [53] who used them for extrapolating novel views outside the convex hull of the input stereo cameras (Tucker and Snavely [42] observe that an MPI can be considered as an instance of Szeliski and Golland's [40] earlier "stack of acetates" volumetric model). Mildenhall et al. [26] and Srinivasan et al. [39], respectively, provide a theoretical analysis of the limits of view interpolation and extrapolation using MPIs. Flynn et al. [9] use learnt gradient descent to generate MPIs from multiple views. Attal et al. [1] and Broxton et al. [3] extend the MPI concept to concentric RGBA spheres that can be used for view synthesis in 360 degrees. Wizadwongsa et al. [48] and Li et al. [22] replace the discrete RGBA planes of MPIs with continuous neural surfaces to achieve higher quality results. Tucker and Snavely [42] use a scale-invariant method that allows them to learn strong data priors that can generate MPIs from a single view. Their approach is in a line of recent work dubbed 3D photography that aims for novel views from in-the-wild, single-view images. Li et al. [23], Han et al. [12], and Luvizon et al.'s [25] methods are MPI-based examples of this approach. An MPI, however, is an over-parameterized scene representation. Recent methods [24, 10, 25, 12] have sought to overcome this shortcoming to some extent through a more judicious placement of depth planes. Nonetheless, their high level of redundancy dilates their memory and computational footprint, and limits their wider adoption in mobile and AR/VR applications. + +For such use cases, a more efficient approach to novel view synthesis involves depth-based warping [6, 49], often followed by inpainting [18, 19]. Shih et al. [38] propose to guide inpainting in disoccluded regions of the warped view using a layered depth representation. Li et al. [21] apply this general approach to 360-degree input. Wiles et al. [47] use a depth map to generate a point cloud of neural features which can be projected and rendered in novel views using a generative network. Choi et al. [7] estimate a probability volume instead of a single depth map to handle uncertainty in difficult regions. Niklaus et al. [29] use seg + +![](images/67bbec6e0ad24ec0274d2d47657b11ebb87b122c7d2e4541c1bd0a1b90225258.jpg) +Figure 2: Tiled multiplane image generation for single-view 3D photography. Given an RGB image and depth from a monocular estimator, our method first generates a pixel-wise confidence map and a pre-processed depth map. These are used to predict $n$ depth planes per tile through weighted $k$ -means clustering (with $k = n$ ). The layer masks defined by pixel labels are used to peel RGBA layers, yielding a rough initial MPI. This is concatenated $(\bigoplus)$ to the depth and passed to a refinement network that generates the RGBA images of the final per-tile MPI. + +mentation to remove the geometric and semantic distortions from depth that often impair the rendering quality of this approach. Nonetheless, depth-based warping suffers from hard boundaries and is over-sensitive to errors in the depth estimate. While Jampani et al. [14] propose soft layering with alpha mattes to address this shortcoming, MPIs are inherently capable of handling such artifacts via blending. + +Finally, our review of view synthesis would not be complete without mentioning neural radiance fields (NeRFs) [27, 50] which have recently沦ued in popularity. While great strides are being made in improving the time and data requirements of NeRFs [51, 28], they remain expensive for interactive applications. Some recent work [33, 41, 43] shows that decomposing a single large radiance field into smaller sub-components can improve efficiency. This is similar to the approach we adopt for multiplane images. + +# 3. Method + +A traditional MPI represents the scene as a set of $N$ fronto-parallel planes in the camera frustum of a reference view $\mathcal{I}$ . Each plane is associated with an RGBA image. While it is possible to place the planes at any depth, they are usually arranged linearly in disparity (inverse depth) [53, 9, 26, 42]. A novel view $\mathcal{I}_t$ is rendered by warping the planes into the target camera's image space via a homography, and compositing them front-to-back using the over [31] operator: + +$$ +\mathcal {I} _ {t} = \sum_ {i = 1} ^ {N} \left(\alpha_ {i} c _ {i} \prod_ {j = i + 1} ^ {N} (1 - \alpha_ {i})\right) \tag {1} +$$ + +where $\alpha_{i}$ and $c_{i}$ are the warped alpha and color channels, respectively, of the $i^{th}$ plane. + +Both warping and compositing can be done very efficiently on graphics hardware, allowing real-time rendering of novel views [26]. Moreover, the alpha channel at each plane allows MPs to represent soft edges and anti-alias any errors in the scene reconstruction, leading to fewer perceptually objectionable artifacts than depth-based warping methods. However, the number of planes $N$ required to capture all the depth variation in a scene is usually large, even though most planes are very sparse. We propose to overcome this shortcoming by representing the reference image $\mathcal{I}$ as a tiled grid of many small MPs (Figure 3). Given $\mathcal{I}$ and its depth map from a monocular depth estimator, our method predicts the placement of $n \ll N$ depth planes within each tiled region and uses this prediction to generate the RGBA images of the MPI in a single forward pass. + +# 3.1. Tiled Multiplane Image Representation + +Our scene representation is based on a set of $m$ tiles, each representing a square sliding block of size $h$ at 2D pixel locations $\mathbf{x}_1, \mathbf{x}_2, \dots, \mathbf{x}_m$ in the source image $\mathcal{I}$ . The locations $\mathbf{x}_k$ lie on a regular grid with spacing determined by some stride $r$ . Each tile consists of $n$ front-parallel RGBA planes, the depth placement of which is not fixed but varies across tiles. We let $\alpha_j^i$ , $c_j^i$ and $d_j^i$ denote, respectively, the alpha channel, the color channel, and depth of the $i$ -th plane in the $j$ -th tile. Then the tiled multiplane image (TMPI) representation $\Gamma(\mathcal{I})$ of the image is defined as: + +$$ +\Gamma (\mathcal {I}) = \left\{\left(\alpha_ {j} ^ {i}, c _ {j} ^ {i}, d _ {j} ^ {i}, \mathbf {x} _ {j}\right) \right\} \tag {2} +$$ + +for all $i = 1,2,\dots,n$ and $j = 1,2,\dots,m$ . Further, we define an ordering on this set of 4-tuplets as, + +$$ +\left(\Gamma (\mathcal {I}), \leq\right) = \left\{\left(\alpha , c, d, \mathbf {x}\right) _ {k = 1, \dots , m n} \mid d _ {k} \leq d _ {k + 1} \right\} \tag {3} +$$ + +![](images/9fca864d2e7945f78dac19a524f93a86a109c0601dc0dee4f6de92529a24fd88.jpg) +Figure 3: (a) A traditional MPI uses five planes per pixel to represent this toy scene, even though no region has more than two overlapping objects. (b) A TMPI exploits the low local depth complexity and has only two planes per pixel. + +As with traditional MPIs, novel views are rendered in a differentiable manner by warping all planes into the target camera's image space. However, as the depth of the planes varies across the tiles and, hence, across the pixels of $\mathcal{I}$ , a planar inverse warp on the target image plane cannot be directly computed via a homography. Instead, each plane must be warped in tile space via a homography computed using a shifted intrinsic matrix, and all the warped planes composited sequentially in the target view at their respective tile locations (Algorithm 1). While this makes the rendering of tiled multiplane images less efficient and somewhat less elegant than MPs (Equation 1), this is only true during training when differentiability is required. At inference, the tiles can be rendered as textured quads using hardware-accelerated rasterization making this stage of the pipeline as efficient as traditional MPs and with lower texture memory requirements. Thus, their compact form makes TMPIs well-suited for rendering over networks, or on mobile devices and VR headsets. The blending of many MPs locally is related to Mildenhall et al.'s [26] light field fusion. However, their method renders each MPI separately before blending the results with scalar weights. Our method operates at a much finer scale, and composites the planes of all MPs together to synthesize a novel view. + +# 3.2. Single-View 3D Photography + +We now describe our approach to generating tiled multiplane images from a single RGB input (Figure 2). Broadly, our method splits the image plane into a regular tiled grid of learnt confidence and outlier-corrected depth. For each tile, the placement of $n$ fronto-parallel depth planes is determined by clustering pixel depth values weighted by the predicted confidence estimates. This latter step is motivated by the fact that with a small depth plane budget $n$ , the commonly used equal spacing in disparity [53, 9, 26, 42] is wasteful. Thus, the goal is to predict the planes that optimally represent all depth variation within a tile. Using the + +Algorithm 1: Differentiable view synthesis using TMPIs. Angled brackets denote pixel indexing. + +RenderTMPI $(\Gamma (\mathcal{I}),\leq)$ $\pmb {R},t,\pmb {K})$ Input: $(\Gamma (\mathcal{I}),\leq)$ : ordered TMPI planes R: relative rotation of novel view t: relative translation K: camera intrinsics Output: Novel view $\mathcal{I}_t\in \mathbb{R}^{3\times H\times W}$ $\mathcal{I}_t\gets \mathbb{0}^{3\times H\times W};\mathcal{T}_t\gets \mathbb{1}^{H\times W};$ foreach $(\alpha ,c,d,\pmb {x})\in \left(\Gamma (\mathcal{I}),\leq\right)$ do $\hat{\mathbf{K}}\gets \mathbf{K} - \left[ \begin{array}{cc}\mathbf{I} & \mathbf{x}\\ 0 & 1 \end{array} \right];$ foreach $u\in [1,\dots,s]\times [1,\dots,s]$ do n $\left[\begin{array}{l}\mathbf{u}_s\\ 1\end{array}\right]\gets \hat{\mathbf{K}} (\mathbf{R} - \mathbf{tn}^T /d)\hat{\mathbf{K}}^{-1}\left[\begin{array}{l}\mathbf{u}\\ 1\end{array}\right];$ $w\gets \alpha \langle \mathbf{u}_s\rangle \mathcal{T}_t\langle \mathbf{u} + \mathbf{x}\rangle ;$ $\mathcal{I}_t\langle \mathbf{u} + \mathbf{x}\rangle \gets \mathcal{I}_t\langle \mathbf{u} + \mathbf{x}\rangle +w c\langle \mathbf{u}_s\rangle ;$ $\mathcal{T}_t\langle \mathbf{u} + \mathbf{x}\rangle \gets \mathcal{T}_t\langle \mathbf{u} + \mathbf{x}\rangle (1 - \alpha \langle \mathbf{u}_s\rangle);$ end end return I + +predicted planes, a fully convolutional network generates the $n$ RGBA images that constitute a per-tile MPI. Unlike Han et al.'s [12] adaptive plane method, we generate the RGBA images in a single forward pass. The resulting TMPI is rendered as a set of textured quads using the rasterization pipeline of graphics hardware. + +In more detail, given the source image $\mathcal{I}$ , we first obtain a depth map $\mathcal{Z}$ for it using a monocular depth estimator. A two-headed U-Net $\Theta(\cdot)$ then predicts a confidence map $\mathcal{C}$ along with denoised depth $\mathcal{D}$ . The goal is to learn a representation that ameliorates the sensitivity of the subsequent $k$ -means clustering step to outliers. The joint prediction of confidence and depth is similar to the depth-routing of Weder et al. [45] and the aleotoric uncertainty estimation of [15]. The predicted depth and confidence, and the original color image are then unfolded into a set of $m$ square sliding blocks of size $h$ and stride $r$ : $\{(\mathcal{D}^i,\mathcal{C}^i,\mathcal{I}^i)_{i=1,\dots,m}\}$ . + +Running $\Theta(\cdot)$ on $\mathcal{I}$ and $\mathcal{Z}$ rather than individual tiles allows it to consider non-local features and avoid undesirable tiling artifacts. Additionally, setting $r < h$ in the unfolding step allows neighboring tiles to overlap. This prevents gaps along tile boundaries and also regularizes per-tile operations across neighbors. However, it also increases the total number of tiles and, thus, the computational requirements. We empirically determine a stride value of $r = h - h/8$ for a good balance between quality and computational efficiency. + +Per-Tile Planes Prediction: Next, we predict the $n$ depth planes $\{z_{j=1,\dots,n}^{i}\}$ that optimally represent the features of the $i$ -th tile. The common approach of spacing the planes linearly in disparity grows inaccurate as $n$ becomes small. Luvizon et al. [25] place the planes at depth discontinuities identified via the histogram of depth values. However, their approach is sensitive to parameter settings and fails for smooth surfaces which have no discontinuities. A learning-based approach is adopted by Han et al. [12] and Li et al. [23]. The former use multi-headed self-attention to adjust a linear placement. While capable of modeling interplane interactions, their method is computationally expensive (Table 2). The latter uses a CNN to directly predict the planes. However, we found that without strong regularization a direct approach lacks topological order and has a strong bias towards a fixed placement. An adversarial loss helps improve this but makes the training more unstable. + +We observe that as depth is known, plane positioning can be posed as a simple clustering problem. Thus, we predict $\{z_j^i\}$ using $k$ -means clustering on the depth in each tile $\mathcal{D}^i$ . + +Standard $k$ -means is sensitive to outliers and can generate significantly different plane predictions across neighboring tiles causing artifacts in novel views. Advantageously, along with the $n$ depth planes, clustering also assigns a label to each input pixel, thereby generating a label map that represents discretized depth. Furthermore, the cluster centers of $k$ -means are differentiable with respect to the input samples. Thus, we address the outlier problem by training $\Theta(\cdot)$ to filter the input through a self-supervised reconstruction loss on the discretized depth map generated by a weighted $k$ -means. In weighted $k$ -means, the cluster centers are updated each iteration using the confidence-weighted mean of the constituent samples. Since we do not directly supervise the depth output of $\Theta(\cdot)$ , the network can go beyond outlier-filtering to learn any modifications that improves the discrete reconstruction, and consequently optimizes the placement of the $n$ depth planes within each tile. + +Per-Tile MPI Generation: Given $\{z_j^i\}$ and the discrete depth map, we estimate a preliminary MPI per tile by peeling RGBA layers from $\mathcal{I}^i$ using the discrete labels as an alpha mask. The masked RGB regions of each plane are inpainted by upsampling valid values from a Gaussian pyramid. A second network $\Psi (\cdot)$ then refines these estimates to generate the final $n$ RGBA images of the MPI for each tile. Following Zhou et al. [53] and Tucker and Snavely [42], we represent the RGBA output as a pixel-wise blend of the input image $\mathcal{I}^i$ and a learnt background. However, unlike these works we predict a background image $\mathcal{B}_j^i$ per plane: + +$$ +\mathcal {W} _ {j} ^ {i} = \prod_ {k > j} \left(1 - \alpha_ {k} ^ {i}\right), \tag {4} +$$ + +$$ +\mathcal {I} _ {j} ^ {i} = \mathcal {W} _ {j} ^ {i} \mathcal {I} ^ {i} + (1 - \mathcal {W} _ {j} ^ {i}) \mathcal {B} _ {j} ^ {i} \tag {5} +$$ + +Where $\alpha_{k}^{i}$ is the predicted alpha value for each plane. + +
Spaces Dataset
MethodPSNR↑SSIM↑LPIPS↓L1↓
SVMPI [42]25.420.7480.2100.040
VMPI [23]22.370.6360.2680.057
MINE [22]24.020.7020.2290.048
AdaMPI [12]26.170.7030.2290.047
Ours24.930.7500.1750.037
Tanks & Temples Dataset
MethodPSNR↑SSIM↑LPIPS↓L1↓
SVMPI17.850.5300.3700.082
VMPI16.320.4630.3950.103
MINE17.230.5060.3660.088
AdaMPI18.620.5650.2700.073
Ours18.690.5690.2670.073
+ +Table 1: Quantitative evaluation of view synthesis results on the Spaces and Tanks and Temples multi-view datasets. SVMPI, MINE and AdaMPI use 32 MPI depth planes; VMPI uses 8; our approach uses 4 planes per image tile. + +# 4. Training Procedure + +We train $\Theta(\cdot)$ in a self-supervised manner by minimizing the L1 loss between the input depth $\mathcal{Z}$ and the folded discrete depth maps produced per tile by weighted $k$ -means clustering. Then, we freeze $\Theta(\cdot)$ and train the MPI generation network $\Psi(\cdot)$ on a novel view synthesis task by following Han et al.'s [12] warp-back strategy to generate pseudo ground truth multi-view training data. This involves warping single-view images into a target camera using monocular depth, and inpainting disocclusion holes with a specially trained network. The view-synthesis training objective is a combination of VGG, structural similarity [44], and L1 losses on the synthesized color image $\mathcal{I}_t$ , weighed as 0.1, 0.25, and 1.0 respectively. For both networks, we use the 111K images of the COCO dataset [4] and the monocular depth method of Ranftl et al. [32]. + +# 5. Experiments + +# 5.1. Implementation + +Our method is implemented in PyTorch and trained on eight Nvidia Tesla V100 GPUs. We use $256 \times 384$ images and accumulate gradients across eight mini-batches of 16 samples each. For both networks $\Theta(\cdot)$ and $\Psi(\cdot)$ , we use the Adam optimizer with a learning rate of $1 \times 10^{-3}$ and a cosine annealing schedule with restarts every 200 epochs. A vectorized implementation of the differentiable TMPI renderer (Algorithm 1) runs at $\sim 100\mathrm{ms} / \mathrm{mini}$ -batch allowing efficient parallel computation. + +
Params.↓(M)GMAC↓Runtime (ms)Peak (GB)Space (MB)↓
DPT*123.0110314.39-
SVMPI43.558.01114.5326.9
MINE38.12501104.91107
VMPI†4.3152.396.04.5713.5
AdaMPI†19.02883505.94107
Ours†6.4357.091.63.205.25
+ +*Monocular depth method. † Method uses monocular depth. +Table 2: Run-time, memory and complexity evaluation of all methods with a single $350 \times 630$ RGB image on an NVIDIA GeForce RTX 3080 GPU. The choice of resolution is dictated by the baseline methods which run out of memory for larger images on the specified hardware. GMACS are Giga Multiply-Accumulate ops./second. + +# 5.2. Baselines + +We compare our approach to four state-of-the-art single-view 3D photography methods based on multiplane images: Tucker et al. [42] (SVMPI), Li et al. [23] (VMPI), Li et al. [22] (MINE) and Han et al. [12] (AdaMPI). Like us, VMPI and AdaMPI use a monocular estimator to recover depth as the first step of their pipeline. The input to all methods is the same, however — a single unconstrained RGB image — and so we evaluate them as end-to-end 3D photography approaches. Nonetheless, we do note the additional depth estimation step when evaluating computational and memory performance (Table 2). We use $N = 32$ planes for all baselines except VMPI, which is designed for $N = 8$ . + +We do not compare to single-view methods based on neural radiance fields [51] or the recent work of Nicklaus et al. [29] and Shih et al. [38] as these are too computationally intensive for our intended use cases on mobile and VR devices. While recent work on NeRFs has demonstrated impressive rendering speeds, training remains expensive, and further, requires a large number of input views and a static scene. Jampani et al.'s [14] method, though not MPI-based, is related. But the authors have not released their code. + +# 5.3. Testing Datasets + +We test all methods on the Spaces [9], and Tanks and Temples [17] datasets. Spaces consists of 100 indoor and outdoor scenes captured using a purpose-built, 16-camera rig. For Tanks and Temples we use the Intermediate split which has uniformly sampled frames from eight high-resolution videos of more challenging outdoor environments. We compute camera poses and depth maps for all scenes using COLMAP [37, 36]. The depth maps are required to resolve the scale ambiguity of monocular depth + +
Spaces Dataset
VariantMAE ↓MSE ↓Q25 ↓
Vanilla k-means35.32.0014.0
Linear plane spacing48.73.8019.3
Ours27.11.2010.6
Tanks& Temples Dataset
VariantMAE ↓MSE ↓Q25 ↓
Vanilla k-means35.72.0014.2
Linear plane spacing46.03.4018.1
Ours27.01.2010.6
+ +Table 3: Evaluating the reconstruction error of different depth discretization approaches. All values are $\times 10^{-3}$ . + +for correct reprojection to target views. We randomly select 1000 source views from each dataset scaled to $350 \times 630$ and use the next image in capture sequence as the target for view synthesis. This choice of resolution is dictated by the baseline methods which run out of memory for larger images on the specified hardware. We present high resolution results for our method on the Davis dataset [30] in Figure 5. + +# 5.4. Evaluation Metrics + +We evaluate the rendered views quantitatively on four metrics: Peak-Signal to Noise Ratio (PSNR), Structural Similarity [44] (SSIM), Learned Perceptual Image Patch Similarity [52] (LPIPS) with a VGG-16 backbone, and the mean absolute error (L1). Following previous work [12, 42, 23], we crop $15\%$ of the image around the edge to account for disocclusions. Further, if more than $15\%$ of the remaining pixels are blank in a view synthesized by any method, we discard the result across all baselines. + +# 5.5. Results + +Table 1 presents quantitative evaluation of all methods on the two test sets. Qualitative results are shown in Figure 4 for Tanks and Temples and Figure 9 for Spaces. Our approach uses $n = 4$ depth planes with a tile size of $h = 64$ . The competitive performance of our method despite having much fewer planes per tile can be attributed to the adaptive placement of depth planes in each tile which allows it to effectively cover a larger depth range than the monolithic MPIs of the baselines, Memory and computational performance is evaluated in Table 2. VMPI, AdaMPI and our method have the additional overhead of monocular depth estimation using Ranftl et al.'s [32] DPT. Our approach has lower runtime, peak memory and space requirements than AdaMPI while achieving similar quality results. + +We evaluate our plane placement strategy in Table 3. We measure the reconstruction quality of the discretized depth + +![](images/d5d3341a5294991bb73f76a7a5c258140aacb5621ea241c723965860eec4661c.jpg) +Ground Truth + +![](images/84783fce5dc6a37c49c093bfb3727840eefa4334eb2703fb85d3ce54f377b690.jpg) +SVMPI + +![](images/e2523570da249fc8e6dfbf319c70a860aea132d5c9e2c9b22b938e5bc6e82b5b.jpg) +VMPI + +![](images/097f682ea8c12c38c1f1ffaaf7ff2abe8e6df5fe69322eee873526f5fe2030a3.jpg) +MINE + +![](images/572db4012fa1f463b941f2423b3fb8ee7ba6321dfdaad11c6ddca1db04b9fd48.jpg) +AdaMPI + +![](images/5f7f617268f73475a6729d1458798073f875fb9cad7d8ee9a58c1131fbb82dc4.jpg) +Ours + +![](images/49d27162c44df8a204fde1b953248a24e7d390ba9198728f093af29363924e00.jpg) +Ground Truth + +![](images/58b248577670ae1c2b6b7862f78fc86844984b1aa9fb4ceecda86b34370eea78.jpg) +SVMPI + +![](images/f03703d0453e85b3317cb5c982a503706458ef246fcd6fee10d9fd2c11edc4c7.jpg) +VMPI + +![](images/9f8ae46a32fcf89b1b05befaa4eada23aa99779cc367bcbe6700d3b758e2162e.jpg) +MINE + +![](images/70e328d37e970b678d4f405488b6f162a2f105b71215f5c6a30520439afc33e7.jpg) +AdaMPI + +![](images/590fbd5d7287e30b57857b1fb04bc3a2ecc07feb23b30d2db0f68f588117d990.jpg) +Ours + +![](images/689381e5f19e89df23b908c7ccddd5b904e32871b61b9581cd9a31d9db6e94cc.jpg) +Figure 4: Comparing the novel view synthesis results of the baseline methods and our approach on the Tanks and Temples dataset. Our results are better than SVMPI, VMPI and MINE, and comparable to AdaMPI while using far fewer depth planes. + +![](images/7743671019847d52e6b4c0901f604c37e8777d4d572fde853624d0b48945484d.jpg) + +![](images/747e0e95f70c4b24c39ae960a3fec7b1abd5d337150ac8e644193b88fc1b6ca5.jpg) + +![](images/a1e224c48122c66cab594aad955fb886ee2f8e9aa8621147ce4906e856080cb5.jpg) + +![](images/db4e8bdeca117ce1f8e31f77529e7fc98b35aa624312fe66075f296e564357a5.jpg) +Figure 5: View synthesis results of our method on the HD $(1080\times 1920)$ Davis [30] dataset. Original views are inset. + +![](images/87313071b9aa01d92b46cfffe7b11be87cf454a1811ca01230e2a377f471f99c.jpg) + +![](images/22fda26b6807b906f95131b175a6a454e553e29346b806d7be7534aed23d9909.jpg) + +![](images/672b2781e7fe6008b7dcf44f62781d90d422d4e05df0dca7a75897d6660c3ffd.jpg) + +![](images/c7a3055412f0e0e4a50f6494ad45f6c09826bf106612bdb1c08205ec67d651c9.jpg) +Figure 6: Top to bottom: Input image (Davis dataset); inverse monocular depth from DPT [32]; naive discretization by spacing four planes linearly in the inverse-depth range of each tile; our weighted clustering-based discretization with four planes. Our representation shows finer variation on receding surfaces and suffers fewer tiling artifacts. + +map defined by the planes in each tile on a monocular depth input that is perturbed by a small amount of Gaussian noise $(\lambda = 0, \sigma^2 = 1 \times 10^{-3})$ . We compare our approach to naive linear spacing of depth planes in inverse disparity space, and to vanilla unweighted $k$ -means. Our method is robust to outliers and yields a much better reconstruction. Figure 6 provides qualitative comparison with linear spacing on samples from the Davis dataset [30]. + +Figure 7 evaluates the effect of tile size and number of planes on the quality of view synthesis. In general, the results uphold the intuition that small tiles and a large number of planes improve quality. This trend is less clear, however, for number of planes $>8$ . This would seem to support the observation of Khakhulin et al. [16] and Hu et al. [13] that reductive models are unable to handle redundant geometry effectively. Moreover, the model's complexity increases proportionately with the number of depth planes, leading to slower convergence for the same number of training steps. + +# 6. Limitations + +As previously observed, a tiled multiplane image cannot exploit the elegant warping and compositing equations of a + +![](images/3cbb07f10df1902889d721b0156c6606976423c2c71d286fba9e3ac079739cf8.jpg) +Figure 7: Evaluating the effect of tile size and number of depth planes on view synthesis quality. In general, the results uphold the intuition that small tiles and a large number of planes improve quality. The trend, however, is less clear when the number of planes $>8$ , indicating that the model may be unable to handle redundant geometry effectively. + +![](images/15577a5f4af11bb15bc43cb95cc81aaaa245cdc049108abd1a6601c5ac8611ff.jpg) +Figure 8: Failure cases of our approach: in some cases, fine features are inconsistently reconstructed across tiles. + +traditional MPI for differentiable rendering during training. Furthermore, in some cases our method fails to reconstruct thin features consistently across tiles (Figure 8). + +# 7. Conclusion + +We present a method for estimating tiled multiplane images from a single RGB input for 3D photography. This includes a novel approach to adaptively spacing a small number of depth planes within an MPI tile to better represent local features. Our method is lightweight, and points a path to realizing novel view synthesis on mobile and VR devices. + +# References + +[1] Benjamin Attal, Selena Ling, Aaron Gokaslan, Christian Richardt, and James Tompkin. MatryODShka: Real-time 6DoF video view synthesis using multi-sphere images. In European Conference on Computer Vision (ECCV), Aug. 2020. 2 +[2] Horace B Barlow et al. Possible principles underlying the transformation of sensory messages. Sensory communication, 1(01):217-233, 1961. 2 + +![](images/304d47c8828685d112135b487dee026ef179927a7a178b1630a2c9d8474a8208.jpg) +Ground Truth + +![](images/c346e1eaabb8454f35f6b5c2f928dcfb1619cf8e5707566d03a81d939be7a4d2.jpg) +SVMPI + +![](images/145c469ab0aa27995c281690b5e73f0eca8040170466179c0bfdc49955b90b31.jpg) +VMPI + +![](images/da7f5e229080bd2bd42eca9888f86ba10cfa1b00ddfaf5468aed200e39d31b82.jpg) +MINE + +![](images/da4a4744a9b4f9ec02700b1b92a3fd32c253a48196921c6df42b3b020f4be215.jpg) +AdaMPI + +![](images/e8edcc08156e8c9157aa5a9b35e23755173db93e55d4b061f0190da0efe545c6.jpg) +Ours + +![](images/a21dcc29f2aeceaf5dfd84fc8c60f4e7abd6892ab6a7b0dc4f7127b670c1a5b9.jpg) + +![](images/d148ab71e9ab70ccfebd50a85b601424bbea3601713f8e77d19df476f35f9b7e.jpg) + +![](images/8dda61d1d1ecb50c79cc281fc8316b79ebb16f2226d8f1e07937f0ba18a4de9b.jpg) + +![](images/052b27b31fa575bee61292c07bb2f1593ec8aa7cf5b60a0871f081fa55d21377.jpg) +Ground Truth +MINE + +![](images/fec7b93b2942740b0578375aeaaa6c182bbaa7314c61111576cc8ba8aca8b804.jpg) +SVMPI +AdaMPI + +![](images/5e642b2d76297fdc0fd52d4f4e689ab590fbc490d218aecee2431ce2c27eac6e.jpg) +VMPI +Ours + +![](images/0a8340d8bf2e211fc04a3befeb94c5110340b884fa9530ad3c7e356086857c87.jpg) + +![](images/1fa0d4d74591c473e12bfffb405f7499600af16bcef976a026b8bb87b39b27df.jpg) + +![](images/078ca2fa2a9593a01fb37b7386f95bb6572b5cee085333145b4d107a175fad70.jpg) +VMPI + +![](images/ec0528a6027fa37d44d330f9d2388455b92e00a0359a1302e65d1ae6ff807dee.jpg) +Ground Truth +MINE +Figure 9: Comparing the novel view synthesis results of the baseline methods and our approach on the Spaces dataset. + +![](images/8dde621a480905b77b0e49619bf4ca9d7d539d6bc400447d6c53bf250a3dfdf1.jpg) +SVMPI +AdaMPI + +![](images/310c22732e27657398b0997f1f70ffc7e9b0367a74428e9295ab31022c7cd7cf.jpg) +Ours + +[3] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. 2 +[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 5 +[5] Jin-Xiang Chai, Xin Tong, Shing-Chow Chan, and Heung-Yeung Shum. Plenoptic sampling. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 307-318, 2000. 2 +[6] Gaurav Chaurasia, Arthur Nieuwoudt, Alexandru-Eugen Ichim, Richard Szeliski, and Alexander Sorkine-Hornung. Passthrough+ real-time stereoscopic view synthesis for mobile mixed reality. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1-17, 2020. 2 +[7] Inchang Choi, Orazio Gallo, Alejandro Troccoli, Min H Kim, and Jan Kautz. Extreme view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7781-7790, 2019. 2 +[8] Abe Davis, Marc Levoy, and Fredo Durand. Unstructured light fields. In Computer Graphics Forum, volume 31, pages 305-314. Wiley Online Library, 2012. 1, 2 +[9] John Flynn, Michael Broxton, Paul Debevec, Matthew DuVall, Graham Fyffe, Ryan Overbeck, Noah Snavely, and Richard Tucker. Deepview: View synthesis with learned gradient descent. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2367-2376, 2019. 2, 3, 4, 6 +[10] Sushobhan Ghosh, Zhaoyang Lv, Nathan Matsuda, Lei Xiao, Andrew Berkovich, and Oliver Cossairt. Liveview: dynamic target-centered mpi for view synthesis. arXiv preprint arXiv:2107.05113, 2021. 2 +[11] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 43-54, 1996. 1, 2 +[12] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-view synthesis in the wild with learned adaptive multi-plane images. In ACM SIGGRAPH, 2022. 2, 4, 5, 6 +[13] Ronghang Hu, Nikhila Ravi, Alexander C Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12528-12537, 2021. 2, 8 +[14] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, and Ce Liu. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 3, 6 +[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 4 +[16] Taras Khakhulin, Denis Korzhenkov, Pavel Solovev, Gleb Sterkin, Andrei-Timotei Ardelean, and Victor Lempitsky. + +Stereo magnification with multi-layer images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8687-8696, 2022. 8 +[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 6 +[18] Johannes Kopf, Suhib Alsisan, Francis Ge, Yangming Chong, Kevin Matzen, Ocean Quigley, Josh Patterson, Jossie Tirado, Shu Wu, and Michael F Cohen. Practical 3d photography. In Proceedings of CVPR Workshops, volume 1, 2019. 2 +[19] Johannes Kopf, Kevin Matzen, Suhib Alsisan, Ocean Quigley, Francis Ge, Yangming Chong, Josh Patterson, Jan-Michael Frahm, Shu Wu, Matthew Yu, et al. One shot 3d photography. ACM Transactions on Graphics (TOG), 39(4):76-1, 2020. 2 +[20] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996. 1, 2 +[21] David Li, Yinda Zhang, Christian Hane, Danhang Tang, Amitabh Varshney, and Ruofei Du. Omnisyn: Synthesizing 360 videos with wide-baseline panoramas. In 2022 IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops (VRW), pages 670-671. IEEE, 2022. 2 +[22] Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. MINE: Towards continuous depth spi with nerf for novel view synthesis. In ICCV, 2021. 2, 5, 6 +[23] Qinbo Li and Nima Khademi Kalantari. Synthesizing light field from a single image with variable mpi and two network fusion. ACM Trans. Graph., 39(6):229-1, 2020. 2, 5, 6 +[24] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4521-4530, 2019. 2 +[25] Diogo C Luvizon, Gustavo Sutter P Carvalho, Andreza A dos Santos, Jhonatas S Conceicao, Jose L Flores-Campana, Luis GL Decker, Marcos R Souza, Helio Pedrini, Antonio Joia, and Otavio AB Penatti. Adaptive multiplane image generation from a single internet picture. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2556-2565, 2021. 2, 5 +[26] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 1, 2, 3, 4 +[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a mul + +tiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3 +[29] Simon Niklaus, Long Mai, Jimei Yang, and Feng Liu. 3d ken burns effect from a single image. ACM Transactions on Graphics (ToG), 38(6):1-15, 2019. 2, 6 +[30] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 6, 7, 8 +[31] Thomas Porter and Tom Duff. Compositing digital images. In Proceedings of the 11th annual conference on Computer graphics and interactive techniques, pages 253-259, 1984. 3 +[32] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 5, 6, 8 +[33] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 3 +[34] Gernot Riegler and Vladlen Koltun. Free view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 623-640. Springer, 2020. 2 +[35] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12216-12225, 2021. 2 +[36] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6 +[37] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 6 +[38] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6 +[39] Pratul P Srinivasan, Richard Tucker, Jonathan T Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snively. Pushing the boundaries of view extrapolation with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 175-184, 2019. 2 +[40] Richard Szeliski and Polina Golland. Stereo matching with transparency and matting. In Sixth International Conference on Computer Vision (IEEE Cat. No. 98CH36271), pages 517-524. IEEE, 1998. 2 +[41] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 3 +[42] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In The IEEE Conference + +on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 3, 4, 5, 6 +[43] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 3 +[44] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5, 6 +[45] Silvan Weder, Johannes Schonberger, Marc Pollefeys, and Martin R Oswald. Routedfusion: Learning real-time depth map fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4887-4897, 2020. 4 +[46] Bennett Wilburn, Neel Joshi, Vaibhav Vaish, Eino-Ville Talvala, Emilio Antunez, Adam Barth, Andrew Adams, Mark Horowitz, and Marc Levoy. High performance imaging using large camera arrays. In ACM SIGGRAPH 2005 Papers, pages 765-776. 2005. 1 +[47] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 2 +[48] Suttisak Wizadwongsa, Pakkapon Phongthawee, Jiraphon Yenphraphai, and Supasorn Suwajanakorn. Nex: Real-time view synthesis with neural basis expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8534-8543, 2021. 2 +[49] Lei Xiao, Salah Nouri, Joel Hegland, Alberto Garcia Garcia, and Douglas Lanman. Neuralpassthrough: Learned real-time view synthesis for vr. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2 +[50] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 1, 3 +[51] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 3, 6 +[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[53] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018. 2, 3, 4, 5 \ No newline at end of file diff --git a/2023/Tiled Multiplane Images for Practical 3D Photography/images.zip b/2023/Tiled Multiplane Images for Practical 3D Photography/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9bc072493c9b4091d608b8ac2b32df18048c2f6b --- /dev/null +++ b/2023/Tiled Multiplane Images for Practical 3D Photography/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bb18d47d9fc611a41c6ef383fb0bcd2e01de751e0d9dc1287143e9d376ee878 +size 1067392 diff --git a/2023/Tiled Multiplane Images for Practical 3D Photography/layout.json b/2023/Tiled Multiplane Images for Practical 3D Photography/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5fc0f3b1cf4a6eb60f694b8ec7c27e1a872a5ab9 --- /dev/null +++ b/2023/Tiled Multiplane Images for Practical 3D Photography/layout.json @@ -0,0 +1,10635 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 465, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 465, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 465, + 121 + ], + "type": "text", + "content": "Tiled Multiplane Images for Practical 3D Photography" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 144, + 198, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 144, + 198, + 156 + ], + "spans": [ + { + "bbox": [ + 129, + 144, + 198, + 156 + ], + "type": "text", + "content": "Numair Khan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 160, + 216, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 160, + 216, + 171 + ], + "spans": [ + { + "bbox": [ + 111, + 160, + 216, + 171 + ], + "type": "text", + "content": "numairkhan@meta.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 272, + 144, + 317, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 144, + 317, + 156 + ], + "spans": [ + { + "bbox": [ + 272, + 144, + 317, + 156 + ], + "type": "text", + "content": "Lei Xiao" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 247, + 160, + 341, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 160, + 341, + 170 + ], + "spans": [ + { + "bbox": [ + 247, + 160, + 341, + 170 + ], + "type": "text", + "content": "lei.xiao@meta.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 385, + 144, + 469, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 144, + 469, + 157 + ], + "spans": [ + { + "bbox": [ + 385, + 144, + 469, + 157 + ], + "type": "text", + "content": "Douglas Lanman" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 372, + 160, + 482, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 160, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 372, + 160, + 482, + 171 + ], + "type": "text", + "content": "doug.lanman@meta.com" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 63, + 198, + 533, + 350 + ], + "blocks": [ + { + "bbox": [ + 63, + 198, + 533, + 350 + ], + "lines": [ + { + "bbox": [ + 63, + 198, + 533, + 350 + ], + "spans": [ + { + "bbox": [ + 63, + 198, + 533, + 350 + ], + "type": "image", + "image_path": "fc5342a98343c8a0d9d0a150e3e00703918263f40220310233be39068a3d9b6e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 361, + 546, + 409 + ], + "lines": [ + { + "bbox": [ + 46, + 361, + 546, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 546, + 409 + ], + "type": "text", + "content": "Figure 1: An overview of our 3D photography method. Left: Using a single RGB image and an estimated monocular depth map, the scene is recreated as a tiled grid of many small MPs. Middle: A visualization of a " + }, + { + "bbox": [ + 46, + 361, + 546, + 409 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 46, + 361, + 546, + 409 + ], + "type": "text", + "content": " tiled grid of MPs, each with three RGBA layers. Right: Novel views are rendered by warping and compositing each MPI tile into the target camera's frustum. The " + }, + { + "bbox": [ + 46, + 361, + 546, + 409 + ], + "type": "inline_equation", + "content": "768 \\times 1152" + }, + { + "bbox": [ + 46, + 361, + 546, + 409 + ], + "type": "text", + "content": " pixel image shown is generated using a " + }, + { + "bbox": [ + 46, + 361, + 546, + 409 + ], + "type": "inline_equation", + "content": "7 \\times 11" + }, + { + "bbox": [ + 46, + 361, + 546, + 409 + ], + "type": "text", + "content": " grid of 4-layer MPs." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 420, + 192, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 420, + 192, + 431 + ], + "spans": [ + { + "bbox": [ + 143, + 420, + 192, + 431 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 451, + 290, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 290, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 290, + 691 + ], + "type": "text", + "content": "The task of synthesizing novel views from a single image has useful applications in virtual reality and mobile computing, and a number of approaches to the problem have been proposed in recent years. A Multiplane Image (MPI) estimates the scene as a stack of RGBA layers, and can model complex appearance effects, anti-alias depth errors and synthesize soft edges better than methods that use textured meshes or layered depth images. And unlike neural radiance fields, an MPI can be efficiently rendered on graphics hardware. However, MPIs are highly redundant and require a large number of depth layers to achieve plausible results. Based on the observation that the depth complexity in local image regions is lower than that over the entire image, we split an MPI into many small, tiled regions, each with only a few depth planes. We call this representation a Tiled Multiplane Image (TMPI). We propose a method for generating a TMPI with adaptive depth planes for single-view 3D photography in the wild. Our synthesized results are comparable to state-of-the-art single-view MPI methods while having lower computational overhead." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 420, + 386, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 386, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 386, + 431 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 445, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 613 + ], + "type": "text", + "content": "The novel view synthesis (NVS) problem involves using a set of input images to generate views from new and unseen camera positions, allowing three dimensional interaction with photos. This is a long-studied problem, with early work relying on interpolation within dense structured image sets [20, 11, 8]. The specialized rigs commonly required to capture the large number of images restricted these methods to lab settings [46]. However, the potential applications offered by novel view synthesis on modern mobile and VR devices has kindled wide interest in the problem, and encouraged researchers to seek methods that make the technology more accessible. The term 3D photography refers to the use of novel view synthesis in everyday capture settings, often from a single image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": "Over the past few years a number of proposed scene representations have leveraged the great strides being made in learning-based techniques to achieve more accurate synthesis with fewer constraints. The most recent of these are neural radiance fields (NeRFs) [26, 50] which represent the scene as multi-layer perceptrons. Their results define the high bar of novel view synthesis. However, this high quality has a significant data and computational cost." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10454" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "An alternate representation, a multiplane image (MPI), defines the scene as a stack of fronto-parallel RGBA planes that can be warped and rendered into novel viewpoints [53, 26, 9]. An MPI offers the advantage of rendering speed, and suffers from less aliasing than mesh or point-based methods [38, 47, 29]. The latter characteristic is important for applications that require temporal stability. However, an MPI is a highly redundant scene representation: the number of RGBA planes required to capture and reconstruct all the depth variation in a scene can be quite high. Since most scenes have a larger amount of free space than occupied, most of the planes in an MPI are very sparse. This makes them inefficient as a representation [2], and expensive to generate, transmit, and store." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 467 + ], + "type": "text", + "content": "In this paper we propose to address these shortcomings of multiplane images and develop a lightweight solution to the 3D photography problem that can be practically implemented on mobile and VR devices. Some examples of the applications we envision are 3D video conferencing, telepresence, and VR passthrough [49]. We show how subdividing the image plane into many small MPs with only a few planes in each, provides a more efficient representation from a computational and memory perspective. However, the naive approach of directly using existing MPI methods with tiles creates boundary artifacts in the novel views. This happens because the commonly used fixed spacing of MPI planes fails to capture the full depth range of a tile when the number of planes is small. Furthermore, it is sensitive to outliers in small regions. We propose a clustering-based approach using learnt confidence weights to predict per-tile MPI planes that better represents local depth features. Our method is lightweight and generates results comparable to the state-of-the-art in MPI-based 3D photography." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 468, + 268, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 468, + 268, + 479 + ], + "spans": [ + { + "bbox": [ + 59, + 468, + 268, + 479 + ], + "type": "text", + "content": "In summary, the main contribution of this work are," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 487, + 287, + 563 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 487, + 287, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 287, + 512 + ], + "type": "text", + "content": "1. The demonstration of tiled multiplane images as a practical representation for view synthesis tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 518, + 286, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 518, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 518, + 286, + 544 + ], + "type": "text", + "content": "2. A learning framework for generating tiled multiplane images from a single RGB input for 3D photography." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 551, + 280, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 551, + 280, + 563 + ], + "spans": [ + { + "bbox": [ + 54, + 551, + 280, + 563 + ], + "type": "text", + "content": "3. A novel approach to adaptive MPI plane positioning." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 574, + 134, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 134, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 134, + 586 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "The progenitors of current 3D photography were the early works on image-based rendering (IBR). These methods usually relied on interpolation within the convex hull of a large set of images to generate novel views. Levoy et al. [20] and Gortler et al. [11] proposed the canonical two-plane parameterization of light fields that renders novel views by quadrilinear interpolation. Gortler's method also provided an early demonstration of the use of geometric proxies to improve rendering quality. Davis et al.'s [8] work extended interpolation-based view synthesis to unstructured" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "images. The excellent analysis of plenoptic sampling done by Chai et al. [5] proved, however, that for large distances the number of images required for view synthesis by interpolation was impractically high. Consequently, the large majority of recent view synthesis methods have relied on learnt priors to overcome the high sampling requirements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 147, + 546, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 546, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 546, + 579 + ], + "type": "text", + "content": "One of the corollaries of Chai et al.'s analysis was that the sampling requirements for view-interpolation are inversely related to the geometric information of the scene. Thus, many subsequent methods have relied on coarse geometric proxies to improve view synthesis quality [34, 35, 13]. In Mildenhall et al.'s [26] method this proxy takes the form of a multiplane image (MPI), which they use to achieve interpolation-based view synthesis that overcomes Chai et al.'s sampling limits. MPIs were first proposed by Zhou et al. [53] who used them for extrapolating novel views outside the convex hull of the input stereo cameras (Tucker and Snavely [42] observe that an MPI can be considered as an instance of Szeliski and Golland's [40] earlier \"stack of acetates\" volumetric model). Mildenhall et al. [26] and Srinivasan et al. [39], respectively, provide a theoretical analysis of the limits of view interpolation and extrapolation using MPIs. Flynn et al. [9] use learnt gradient descent to generate MPIs from multiple views. Attal et al. [1] and Broxton et al. [3] extend the MPI concept to concentric RGBA spheres that can be used for view synthesis in 360 degrees. Wizadwongsa et al. [48] and Li et al. [22] replace the discrete RGBA planes of MPIs with continuous neural surfaces to achieve higher quality results. Tucker and Snavely [42] use a scale-invariant method that allows them to learn strong data priors that can generate MPIs from a single view. Their approach is in a line of recent work dubbed 3D photography that aims for novel views from in-the-wild, single-view images. Li et al. [23], Han et al. [12], and Luvizon et al.'s [25] methods are MPI-based examples of this approach. An MPI, however, is an over-parameterized scene representation. Recent methods [24, 10, 25, 12] have sought to overcome this shortcoming to some extent through a more judicious placement of depth planes. Nonetheless, their high level of redundancy dilates their memory and computational footprint, and limits their wider adoption in mobile and AR/VR applications." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "For such use cases, a more efficient approach to novel view synthesis involves depth-based warping [6, 49], often followed by inpainting [18, 19]. Shih et al. [38] propose to guide inpainting in disoccluded regions of the warped view using a layered depth representation. Li et al. [21] apply this general approach to 360-degree input. Wiles et al. [47] use a depth map to generate a point cloud of neural features which can be projected and rendered in novel views using a generative network. Choi et al. [7] estimate a probability volume instead of a single depth map to handle uncertainty in difficult regions. Niklaus et al. [29] use seg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10455" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 72, + 541, + 225 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 541, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 541, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 541, + 225 + ], + "type": "image", + "image_path": "67bbec6e0ad24ec0274d2d47657b11ebb87b122c7d2e4541c1bd0a1b90225258.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 238, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 547, + 300 + ], + "type": "text", + "content": "Figure 2: Tiled multiplane image generation for single-view 3D photography. Given an RGB image and depth from a monocular estimator, our method first generates a pixel-wise confidence map and a pre-processed depth map. These are used to predict " + }, + { + "bbox": [ + 46, + 238, + 547, + 300 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 238, + 547, + 300 + ], + "type": "text", + "content": " depth planes per tile through weighted " + }, + { + "bbox": [ + 46, + 238, + 547, + 300 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 238, + 547, + 300 + ], + "type": "text", + "content": "-means clustering (with " + }, + { + "bbox": [ + 46, + 238, + 547, + 300 + ], + "type": "inline_equation", + "content": "k = n" + }, + { + "bbox": [ + 46, + 238, + 547, + 300 + ], + "type": "text", + "content": "). The layer masks defined by pixel labels are used to peel RGBA layers, yielding a rough initial MPI. This is concatenated " + }, + { + "bbox": [ + 46, + 238, + 547, + 300 + ], + "type": "inline_equation", + "content": "(\\bigoplus)" + }, + { + "bbox": [ + 46, + 238, + 547, + 300 + ], + "type": "text", + "content": " to the depth and passed to a refinement network that generates the RGBA images of the final per-tile MPI." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 319, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 403 + ], + "type": "text", + "content": "mentation to remove the geometric and semantic distortions from depth that often impair the rendering quality of this approach. Nonetheless, depth-based warping suffers from hard boundaries and is over-sensitive to errors in the depth estimate. While Jampani et al. [14] propose soft layering with alpha mattes to address this shortcoming, MPIs are inherently capable of handling such artifacts via blending." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 403, + 288, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 288, + 511 + ], + "type": "text", + "content": "Finally, our review of view synthesis would not be complete without mentioning neural radiance fields (NeRFs) [27, 50] which have recently沦ued in popularity. While great strides are being made in improving the time and data requirements of NeRFs [51, 28], they remain expensive for interactive applications. Some recent work [33, 41, 43] shows that decomposing a single large radiance field into smaller sub-components can improve efficiency. This is similar to the approach we adopt for multiplane images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 520, + 103, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 103, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 103, + 532 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "text", + "content": "A traditional MPI represents the scene as a set of " + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "text", + "content": " fronto-parallel planes in the camera frustum of a reference view " + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "text", + "content": ". Each plane is associated with an RGBA image. While it is possible to place the planes at any depth, they are usually arranged linearly in disparity (inverse depth) [53, 9, 26, 42]. A novel view " + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t" + }, + { + "bbox": [ + 46, + 540, + 287, + 647 + ], + "type": "text", + "content": " is rendered by warping the planes into the target camera's image space via a homography, and compositing them front-to-back using the over [31] operator:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 652, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 652, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 102, + 652, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {t} = \\sum_ {i = 1} ^ {N} \\left(\\alpha_ {i} c _ {i} \\prod_ {j = i + 1} ^ {N} (1 - \\alpha_ {i})\\right) \\tag {1}", + "image_path": "4bd9e52cf17814af6c34495fdeedb96868883126d59eadfca3c413cbe377e6cc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " are the warped alpha and color channels, respectively, of the " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " plane." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 319, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 499 + ], + "type": "text", + "content": "Both warping and compositing can be done very efficiently on graphics hardware, allowing real-time rendering of novel views [26]. Moreover, the alpha channel at each plane allows MPs to represent soft edges and anti-alias any errors in the scene reconstruction, leading to fewer perceptually objectionable artifacts than depth-based warping methods. However, the number of planes " + }, + { + "bbox": [ + 304, + 319, + 545, + 499 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 319, + 545, + 499 + ], + "type": "text", + "content": " required to capture all the depth variation in a scene is usually large, even though most planes are very sparse. We propose to overcome this shortcoming by representing the reference image " + }, + { + "bbox": [ + 304, + 319, + 545, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 319, + 545, + 499 + ], + "type": "text", + "content": " as a tiled grid of many small MPs (Figure 3). Given " + }, + { + "bbox": [ + 304, + 319, + 545, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 319, + 545, + 499 + ], + "type": "text", + "content": " and its depth map from a monocular depth estimator, our method predicts the placement of " + }, + { + "bbox": [ + 304, + 319, + 545, + 499 + ], + "type": "inline_equation", + "content": "n \\ll N" + }, + { + "bbox": [ + 304, + 319, + 545, + 499 + ], + "type": "text", + "content": " depth planes within each tiled region and uses this prediction to generate the RGBA images of the MPI in a single forward pass." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 505, + 513, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 505, + 513, + 519 + ], + "spans": [ + { + "bbox": [ + 305, + 505, + 513, + 519 + ], + "type": "text", + "content": "3.1. Tiled Multiplane Image Representation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": "Our scene representation is based on a set of " + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": " tiles, each representing a square sliding block of size " + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": " at 2D pixel locations " + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1, \\mathbf{x}_2, \\dots, \\mathbf{x}_m" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": " in the source image " + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": ". The locations " + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": " lie on a regular grid with spacing determined by some stride " + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": ". Each tile consists of " + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": " front-parallel RGBA planes, the depth placement of which is not fixed but varies across tiles. We let " + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\alpha_j^i" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "inline_equation", + "content": "c_j^i" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "inline_equation", + "content": "d_j^i" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": " denote, respectively, the alpha channel, the color channel, and depth of the " + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": "-th plane in the " + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": "-th tile. Then the tiled multiplane image (TMPI) representation " + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "inline_equation", + "content": "\\Gamma(\\mathcal{I})" + }, + { + "bbox": [ + 304, + 523, + 545, + 643 + ], + "type": "text", + "content": " of the image is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 372, + 647, + 545, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 647, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 372, + 647, + 545, + 663 + ], + "type": "interline_equation", + "content": "\\Gamma (\\mathcal {I}) = \\left\\{\\left(\\alpha_ {j} ^ {i}, c _ {j} ^ {i}, d _ {j} ^ {i}, \\mathbf {x} _ {j}\\right) \\right\\} \\tag {2}", + "image_path": "377f8f9105b7269fa74c91051c5c80422fbf07db9820d4f06085417c56cf9a64.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 667, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 667, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 667, + 545, + 691 + ], + "type": "text", + "content": "for all " + }, + { + "bbox": [ + 304, + 667, + 545, + 691 + ], + "type": "inline_equation", + "content": "i = 1,2,\\dots,n" + }, + { + "bbox": [ + 304, + 667, + 545, + 691 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 667, + 545, + 691 + ], + "type": "inline_equation", + "content": "j = 1,2,\\dots,m" + }, + { + "bbox": [ + 304, + 667, + 545, + 691 + ], + "type": "text", + "content": ". Further, we define an ordering on this set of 4-tuplets as," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 696, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 696, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 696, + 545, + 712 + ], + "type": "interline_equation", + "content": "\\left(\\Gamma (\\mathcal {I}), \\leq\\right) = \\left\\{\\left(\\alpha , c, d, \\mathbf {x}\\right) _ {k = 1, \\dots , m n} \\mid d _ {k} \\leq d _ {k + 1} \\right\\} \\tag {3}", + "image_path": "919b9a5cc4c1c6fc9223a3888aef78874dae93e5f837271494b908814c4c8153.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10456" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 289, + 182 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 289, + 182 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 289, + 182 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 289, + 182 + ], + "type": "image", + "image_path": "9fca864d2e7945f78dac19a524f93a86a109c0601dc0dee4f6de92529a24fd88.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 287, + 238 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 287, + 238 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 287, + 238 + ], + "type": "text", + "content": "Figure 3: (a) A traditional MPI uses five planes per pixel to represent this toy scene, even though no region has more than two overlapping objects. (b) A TMPI exploits the low local depth complexity and has only two planes per pixel." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 262, + 289, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 289, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 289, + 550 + ], + "type": "text", + "content": "As with traditional MPIs, novel views are rendered in a differentiable manner by warping all planes into the target camera's image space. However, as the depth of the planes varies across the tiles and, hence, across the pixels of " + }, + { + "bbox": [ + 46, + 262, + 289, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 262, + 289, + 550 + ], + "type": "text", + "content": ", a planar inverse warp on the target image plane cannot be directly computed via a homography. Instead, each plane must be warped in tile space via a homography computed using a shifted intrinsic matrix, and all the warped planes composited sequentially in the target view at their respective tile locations (Algorithm 1). While this makes the rendering of tiled multiplane images less efficient and somewhat less elegant than MPs (Equation 1), this is only true during training when differentiability is required. At inference, the tiles can be rendered as textured quads using hardware-accelerated rasterization making this stage of the pipeline as efficient as traditional MPs and with lower texture memory requirements. Thus, their compact form makes TMPIs well-suited for rendering over networks, or on mobile devices and VR headsets. The blending of many MPs locally is related to Mildenhall et al.'s [26] light field fusion. However, their method renders each MPI separately before blending the results with scalar weights. Our method operates at a much finer scale, and composites the planes of all MPs together to synthesize a novel view." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 562, + 205, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 205, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 205, + 574 + ], + "type": "text", + "content": "3.2. Single-View 3D Photography" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "We now describe our approach to generating tiled multiplane images from a single RGB input (Figure 2). Broadly, our method splits the image plane into a regular tiled grid of learnt confidence and outlier-corrected depth. For each tile, the placement of " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": " fronto-parallel depth planes is determined by clustering pixel depth values weighted by the predicted confidence estimates. This latter step is motivated by the fact that with a small depth plane budget " + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": ", the commonly used equal spacing in disparity [53, 9, 26, 42] is wasteful. Thus, the goal is to predict the planes that optimally represent all depth variation within a tile. Using the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 310, + 75, + 520, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 75, + 520, + 99 + ], + "spans": [ + { + "bbox": [ + 310, + 75, + 520, + 99 + ], + "type": "text", + "content": "Algorithm 1: Differentiable view synthesis using TMPIs. Angled brackets denote pixel indexing." + } + ] + } + ], + "index": 5 + }, + { + "type": "code", + "bbox": [ + 318, + 101, + 521, + 369 + ], + "blocks": [ + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "lines": [ + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "spans": [ + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "text", + "content": "RenderTMPI " + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "inline_equation", + "content": "(\\Gamma (\\mathcal{I}),\\leq)" + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "inline_equation", + "content": "\\pmb {R},t,\\pmb {K})" + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "text", + "content": " Input: " + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "inline_equation", + "content": "(\\Gamma (\\mathcal{I}),\\leq)" + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "text", + "content": " : ordered TMPI planes R: relative rotation of novel view t: relative translation K: camera intrinsics Output: Novel view " + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t\\in \\mathbb{R}^{3\\times H\\times W}" + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t\\gets \\mathbb{0}^{3\\times H\\times W};\\mathcal{T}_t\\gets \\mathbb{1}^{H\\times W};" + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "text", + "content": " foreach " + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "inline_equation", + "content": "(\\alpha ,c,d,\\pmb {x})\\in \\left(\\Gamma (\\mathcal{I}),\\leq\\right)" + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{K}}\\gets \\mathbf{K} - \\left[ \\begin{array}{cc}\\mathbf{I} & \\mathbf{x}\\\\ 0 & 1 \\end{array} \\right];" + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "text", + "content": " foreach " + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "inline_equation", + "content": "u\\in [1,\\dots,s]\\times [1,\\dots,s]" + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "text", + "content": " do n " + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "inline_equation", + "content": "\\left[\\begin{array}{l}\\mathbf{u}_s\\\\ 1\\end{array}\\right]\\gets \\hat{\\mathbf{K}} (\\mathbf{R} - \\mathbf{tn}^T /d)\\hat{\\mathbf{K}}^{-1}\\left[\\begin{array}{l}\\mathbf{u}\\\\ 1\\end{array}\\right];" + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "inline_equation", + "content": "w\\gets \\alpha \\langle \\mathbf{u}_s\\rangle \\mathcal{T}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle ;" + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle \\gets \\mathcal{I}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle +w c\\langle \\mathbf{u}_s\\rangle ;" + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle \\gets \\mathcal{T}_t\\langle \\mathbf{u} + \\mathbf{x}\\rangle (1 - \\alpha \\langle \\mathbf{u}_s\\rangle);" + }, + { + "bbox": [ + 318, + 101, + 521, + 369 + ], + "type": "text", + "content": " end end return I" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_body" + } + ], + "index": 6, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": "predicted planes, a fully convolutional network generates the " + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": " RGBA images that constitute a per-tile MPI. Unlike Han et al.'s [12] adaptive plane method, we generate the RGBA images in a single forward pass. The resulting TMPI is rendered as a set of textured quads using the rasterization pipeline of graphics hardware." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "text", + "content": "In more detail, given the source image " + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "text", + "content": ", we first obtain a depth map " + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "text", + "content": " for it using a monocular depth estimator. A two-headed U-Net " + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "text", + "content": " then predicts a confidence map " + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "text", + "content": " along with denoised depth " + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "text", + "content": ". The goal is to learn a representation that ameliorates the sensitivity of the subsequent " + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "text", + "content": "-means clustering step to outliers. The joint prediction of confidence and depth is similar to the depth-routing of Weder et al. [45] and the aleotoric uncertainty estimation of [15]. The predicted depth and confidence, and the original color image are then unfolded into a set of " + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "text", + "content": " square sliding blocks of size " + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "text", + "content": " and stride " + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "inline_equation", + "content": "\\{(\\mathcal{D}^i,\\mathcal{C}^i,\\mathcal{I}^i)_{i=1,\\dots,m}\\}" + }, + { + "bbox": [ + 304, + 472, + 546, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Running " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " rather than individual tiles allows it to consider non-local features and avoid undesirable tiling artifacts. Additionally, setting " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "r < h" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " in the unfolding step allows neighboring tiles to overlap. This prevents gaps along tile boundaries and also regularizes per-tile operations across neighbors. However, it also increases the total number of tiles and, thus, the computational requirements. We empirically determine a stride value of " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "r = h - h/8" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " for a good balance between quality and computational efficiency." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10457" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "Per-Tile Planes Prediction: Next, we predict the " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": " depth planes " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "\\{z_{j=1,\\dots,n}^{i}\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": " that optimally represent the features of the " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "-th tile. The common approach of spacing the planes linearly in disparity grows inaccurate as " + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": " becomes small. Luvizon et al. [25] place the planes at depth discontinuities identified via the histogram of depth values. However, their approach is sensitive to parameter settings and fails for smooth surfaces which have no discontinuities. A learning-based approach is adopted by Han et al. [12] and Li et al. [23]. The former use multi-headed self-attention to adjust a linear placement. While capable of modeling interplane interactions, their method is computationally expensive (Table 2). The latter uses a CNN to directly predict the planes. However, we found that without strong regularization a direct approach lacks topological order and has a strong bias towards a fixed placement. An adversarial loss helps improve this but makes the training more unstable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 275, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 287, + 312 + ], + "type": "text", + "content": "We observe that as depth is known, plane positioning can be posed as a simple clustering problem. Thus, we predict " + }, + { + "bbox": [ + 47, + 275, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\{z_j^i\\}" + }, + { + "bbox": [ + 47, + 275, + 287, + 312 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 47, + 275, + 287, + 312 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 275, + 287, + 312 + ], + "type": "text", + "content": "-means clustering on the depth in each tile " + }, + { + "bbox": [ + 47, + 275, + 287, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^i" + }, + { + "bbox": [ + 47, + 275, + 287, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "text", + "content": "Standard " + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "text", + "content": "-means is sensitive to outliers and can generate significantly different plane predictions across neighboring tiles causing artifacts in novel views. Advantageously, along with the " + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "text", + "content": " depth planes, clustering also assigns a label to each input pixel, thereby generating a label map that represents discretized depth. Furthermore, the cluster centers of " + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "text", + "content": "-means are differentiable with respect to the input samples. Thus, we address the outlier problem by training " + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "text", + "content": " to filter the input through a self-supervised reconstruction loss on the discretized depth map generated by a weighted " + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "text", + "content": "-means. In weighted " + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "text", + "content": "-means, the cluster centers are updated each iteration using the confidence-weighted mean of the constituent samples. Since we do not directly supervise the depth output of " + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "text", + "content": ", the network can go beyond outlier-filtering to learn any modifications that improves the discrete reconstruction, and consequently optimizes the placement of the " + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 312, + 289, + 514 + ], + "type": "text", + "content": " depth planes within each tile." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "text", + "content": "Per-Tile MPI Generation: Given " + }, + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "inline_equation", + "content": "\\{z_j^i\\}" + }, + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "text", + "content": " and the discrete depth map, we estimate a preliminary MPI per tile by peeling RGBA layers from " + }, + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^i" + }, + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "text", + "content": " using the discrete labels as an alpha mask. The masked RGB regions of each plane are inpainted by upsampling valid values from a Gaussian pyramid. A second network " + }, + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "inline_equation", + "content": "\\Psi (\\cdot)" + }, + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "text", + "content": " then refines these estimates to generate the final " + }, + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "text", + "content": " RGBA images of the MPI for each tile. Following Zhou et al. [53] and Tucker and Snavely [42], we represent the RGBA output as a pixel-wise blend of the input image " + }, + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^i" + }, + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "text", + "content": " and a learnt background. However, unlike these works we predict a background image " + }, + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_j^i" + }, + { + "bbox": [ + 46, + 514, + 289, + 647 + ], + "type": "text", + "content": " per plane:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 652, + 287, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 652, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 108, + 652, + 287, + 679 + ], + "type": "interline_equation", + "content": "\\mathcal {W} _ {j} ^ {i} = \\prod_ {k > j} \\left(1 - \\alpha_ {k} ^ {i}\\right), \\tag {4}", + "image_path": "01735fe1ccf3680ce412a9ea4c346543d9b89012a8f276afc3d62bbd35fd78db.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 681, + 287, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 681, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 113, + 681, + 287, + 696 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {j} ^ {i} = \\mathcal {W} _ {j} ^ {i} \\mathcal {I} ^ {i} + (1 - \\mathcal {W} _ {j} ^ {i}) \\mathcal {B} _ {j} ^ {i} \\tag {5}", + "image_path": "6e47ffa2827826aa58cfbf5dfa9b915e48178635c3ced4e25a043751c24a5ae6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 700, + 264, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 700, + 264, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 700, + 264, + 715 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 47, + 700, + 264, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{k}^{i}" + }, + { + "bbox": [ + 47, + 700, + 264, + 715 + ], + "type": "text", + "content": " is the predicted alpha value for each plane." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 312, + 70, + 541, + 272 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 541, + 272 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 541, + 272 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 541, + 272 + ], + "type": "table", + "html": "
Spaces Dataset
MethodPSNR↑SSIM↑LPIPS↓L1↓
SVMPI [42]25.420.7480.2100.040
VMPI [23]22.370.6360.2680.057
MINE [22]24.020.7020.2290.048
AdaMPI [12]26.170.7030.2290.047
Ours24.930.7500.1750.037
Tanks & Temples Dataset
MethodPSNR↑SSIM↑LPIPS↓L1↓
SVMPI17.850.5300.3700.082
VMPI16.320.4630.3950.103
MINE17.230.5060.3660.088
AdaMPI18.620.5650.2700.073
Ours18.690.5690.2670.073
", + "image_path": "12a5be94f7060e2e2c845982942c1dab146f2ebc695cbbea2b934f4d8969998a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 289, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 338 + ], + "type": "text", + "content": "Table 1: Quantitative evaluation of view synthesis results on the Spaces and Tanks and Temples multi-view datasets. SVMPI, MINE and AdaMPI use 32 MPI depth planes; VMPI uses 8; our approach uses 4 planes per image tile." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 357, + 422, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 357, + 422, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 357, + 422, + 371 + ], + "type": "text", + "content": "4. Training Procedure" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "text", + "content": "We train " + }, + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "text", + "content": " in a self-supervised manner by minimizing the L1 loss between the input depth " + }, + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "text", + "content": " and the folded discrete depth maps produced per tile by weighted " + }, + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "text", + "content": "-means clustering. Then, we freeze " + }, + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "text", + "content": " and train the MPI generation network " + }, + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\Psi(\\cdot)" + }, + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "text", + "content": " on a novel view synthesis task by following Han et al.'s [12] warp-back strategy to generate pseudo ground truth multi-view training data. This involves warping single-view images into a target camera using monocular depth, and inpainting disocclusion holes with a specially trained network. The view-synthesis training objective is a combination of VGG, structural similarity [44], and L1 losses on the synthesized color image " + }, + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t" + }, + { + "bbox": [ + 304, + 377, + 545, + 557 + ], + "type": "text", + "content": ", weighed as 0.1, 0.25, and 1.0 respectively. For both networks, we use the 111K images of the COCO dataset [4] and the monocular depth method of Ranftl et al. [32]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 567, + 387, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 387, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 387, + 581 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 586, + 404, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 404, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 404, + 600 + ], + "type": "text", + "content": "5.1. Implementation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Our method is implemented in PyTorch and trained on eight Nvidia Tesla V100 GPUs. We use " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 384" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " images and accumulate gradients across eight mini-batches of 16 samples each. For both networks " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Theta(\\cdot)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Psi(\\cdot)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", we use the Adam optimizer with a learning rate of " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and a cosine annealing schedule with restarts every 200 epochs. A vectorized implementation of the differentiable TMPI renderer (Algorithm 1) runs at " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sim 100\\mathrm{ms} / \\mathrm{mini}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "-batch allowing efficient parallel computation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10458" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 286, + 182 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 286, + 182 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 286, + 182 + ], + "type": "table", + "html": "
Params.↓(M)GMAC↓Runtime (ms)Peak (GB)Space (MB)↓
DPT*123.0110314.39-
SVMPI43.558.01114.5326.9
MINE38.12501104.91107
VMPI†4.3152.396.04.5713.5
AdaMPI†19.02883505.94107
Ours†6.4357.091.63.205.25
", + "image_path": "ca38c01f4a607ac82084cc96d2befaf9e16b4767dfe9ed406d2e6244e739a24a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 49, + 183, + 247, + 194 + ], + "lines": [ + { + "bbox": [ + 49, + 183, + 247, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 183, + 247, + 194 + ], + "type": "text", + "content": "*Monocular depth method. † Method uses monocular depth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 46, + 213, + 287, + 286 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 287, + 286 + ], + "type": "text", + "content": "Table 2: Run-time, memory and complexity evaluation of all methods with a single " + }, + { + "bbox": [ + 46, + 213, + 287, + 286 + ], + "type": "inline_equation", + "content": "350 \\times 630" + }, + { + "bbox": [ + 46, + 213, + 287, + 286 + ], + "type": "text", + "content": " RGB image on an NVIDIA GeForce RTX 3080 GPU. The choice of resolution is dictated by the baseline methods which run out of memory for larger images on the specified hardware. GMACS are Giga Multiply-Accumulate ops./second." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 307, + 113, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 113, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 113, + 319 + ], + "type": "text", + "content": "5.2. Baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 326, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 287, + 469 + ], + "type": "text", + "content": "We compare our approach to four state-of-the-art single-view 3D photography methods based on multiplane images: Tucker et al. [42] (SVMPI), Li et al. [23] (VMPI), Li et al. [22] (MINE) and Han et al. [12] (AdaMPI). Like us, VMPI and AdaMPI use a monocular estimator to recover depth as the first step of their pipeline. The input to all methods is the same, however — a single unconstrained RGB image — and so we evaluate them as end-to-end 3D photography approaches. Nonetheless, we do note the additional depth estimation step when evaluating computational and memory performance (Table 2). We use " + }, + { + "bbox": [ + 46, + 326, + 287, + 469 + ], + "type": "inline_equation", + "content": "N = 32" + }, + { + "bbox": [ + 46, + 326, + 287, + 469 + ], + "type": "text", + "content": " planes for all baselines except VMPI, which is designed for " + }, + { + "bbox": [ + 46, + 326, + 287, + 469 + ], + "type": "inline_equation", + "content": "N = 8" + }, + { + "bbox": [ + 46, + 326, + 287, + 469 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 470, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 470, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 470, + 287, + 578 + ], + "type": "text", + "content": "We do not compare to single-view methods based on neural radiance fields [51] or the recent work of Nicklaus et al. [29] and Shih et al. [38] as these are too computationally intensive for our intended use cases on mobile and VR devices. While recent work on NeRFs has demonstrated impressive rendering speeds, training remains expensive, and further, requires a large number of input views and a static scene. Jampani et al.'s [14] method, though not MPI-based, is related. But the authors have not released their code." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 586, + 146, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 146, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 146, + 600 + ], + "type": "text", + "content": "5.3. Testing Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "We test all methods on the Spaces [9], and Tanks and Temples [17] datasets. Spaces consists of 100 indoor and outdoor scenes captured using a purpose-built, 16-camera rig. For Tanks and Temples we use the Intermediate split which has uniformly sampled frames from eight high-resolution videos of more challenging outdoor environments. We compute camera poses and depth maps for all scenes using COLMAP [37, 36]. The depth maps are required to resolve the scale ambiguity of monocular depth" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 326, + 70, + 526, + 223 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 526, + 223 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 526, + 223 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 526, + 223 + ], + "type": "table", + "html": "
Spaces Dataset
VariantMAE ↓MSE ↓Q25 ↓
Vanilla k-means35.32.0014.0
Linear plane spacing48.73.8019.3
Ours27.11.2010.6
Tanks& Temples Dataset
VariantMAE ↓MSE ↓Q25 ↓
Vanilla k-means35.72.0014.2
Linear plane spacing46.03.4018.1
Ours27.01.2010.6
", + "image_path": "6f72d671c8552ddaf56415b7e9da58bbb5a2b7a0707f7b95a36fc9d3d054f95f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 241, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 305, + 241, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 241, + 545, + 266 + ], + "type": "text", + "content": "Table 3: Evaluating the reconstruction error of different depth discretization approaches. All values are " + }, + { + "bbox": [ + 305, + 241, + 545, + 266 + ], + "type": "inline_equation", + "content": "\\times 10^{-3}" + }, + { + "bbox": [ + 305, + 241, + 545, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 278, + 545, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 362 + ], + "type": "text", + "content": "for correct reprojection to target views. We randomly select 1000 source views from each dataset scaled to " + }, + { + "bbox": [ + 304, + 278, + 545, + 362 + ], + "type": "inline_equation", + "content": "350 \\times 630" + }, + { + "bbox": [ + 304, + 278, + 545, + 362 + ], + "type": "text", + "content": " and use the next image in capture sequence as the target for view synthesis. This choice of resolution is dictated by the baseline methods which run out of memory for larger images on the specified hardware. We present high resolution results for our method on the Davis dataset [30] in Figure 5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 369, + 418, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 418, + 381 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 418, + 381 + ], + "type": "text", + "content": "5.4. Evaluation Metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 388, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 495 + ], + "type": "text", + "content": "We evaluate the rendered views quantitatively on four metrics: Peak-Signal to Noise Ratio (PSNR), Structural Similarity [44] (SSIM), Learned Perceptual Image Patch Similarity [52] (LPIPS) with a VGG-16 backbone, and the mean absolute error (L1). Following previous work [12, 42, 23], we crop " + }, + { + "bbox": [ + 304, + 388, + 545, + 495 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 388, + 545, + 495 + ], + "type": "text", + "content": " of the image around the edge to account for disocclusions. Further, if more than " + }, + { + "bbox": [ + 304, + 388, + 545, + 495 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 388, + 545, + 495 + ], + "type": "text", + "content": " of the remaining pixels are blank in a view synthesized by any method, we discard the result across all baselines." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 503, + 362, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 362, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 362, + 515 + ], + "type": "text", + "content": "5.5. Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 689 + ], + "type": "text", + "content": "Table 1 presents quantitative evaluation of all methods on the two test sets. Qualitative results are shown in Figure 4 for Tanks and Temples and Figure 9 for Spaces. Our approach uses " + }, + { + "bbox": [ + 304, + 521, + 545, + 689 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 304, + 521, + 545, + 689 + ], + "type": "text", + "content": " depth planes with a tile size of " + }, + { + "bbox": [ + 304, + 521, + 545, + 689 + ], + "type": "inline_equation", + "content": "h = 64" + }, + { + "bbox": [ + 304, + 521, + 545, + 689 + ], + "type": "text", + "content": ". The competitive performance of our method despite having much fewer planes per tile can be attributed to the adaptive placement of depth planes in each tile which allows it to effectively cover a larger depth range than the monolithic MPIs of the baselines, Memory and computational performance is evaluated in Table 2. VMPI, AdaMPI and our method have the additional overhead of monocular depth estimation using Ranftl et al.'s [32] DPT. Our approach has lower runtime, peak memory and space requirements than AdaMPI while achieving similar quality results." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We evaluate our plane placement strategy in Table 3. We measure the reconstruction quality of the discretized depth" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10459" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 69, + 219, + 162 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 219, + 162 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 219, + 162 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 219, + 162 + ], + "type": "image", + "image_path": "d5d3341a5294991bb73f76a7a5c258140aacb5621ea241c723965860eec4661c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 163, + 162, + 171 + ], + "lines": [ + { + "bbox": [ + 112, + 163, + 162, + 171 + ], + "spans": [ + { + "bbox": [ + 112, + 163, + 162, + 171 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 70, + 387, + 162 + ], + "blocks": [ + { + "bbox": [ + 220, + 70, + 387, + 162 + ], + "lines": [ + { + "bbox": [ + 220, + 70, + 387, + 162 + ], + "spans": [ + { + "bbox": [ + 220, + 70, + 387, + 162 + ], + "type": "image", + "image_path": "84783fce5dc6a37c49c093bfb3727840eefa4334eb2703fb85d3ce54f377b690.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 163, + 315, + 171 + ], + "lines": [ + { + "bbox": [ + 290, + 163, + 315, + 171 + ], + "spans": [ + { + "bbox": [ + 290, + 163, + 315, + 171 + ], + "type": "text", + "content": "SVMPI" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 545, + 162 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 545, + 162 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 545, + 162 + ], + "type": "image", + "image_path": "e2523570da249fc8e6dfbf319c70a860aea132d5c9e2c9b22b938e5bc6e82b5b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 163, + 487, + 171 + ], + "lines": [ + { + "bbox": [ + 466, + 163, + 487, + 171 + ], + "spans": [ + { + "bbox": [ + 466, + 163, + 487, + 171 + ], + "type": "text", + "content": "VMPI" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 52, + 173, + 219, + 267 + ], + "blocks": [ + { + "bbox": [ + 52, + 173, + 219, + 267 + ], + "lines": [ + { + "bbox": [ + 52, + 173, + 219, + 267 + ], + "spans": [ + { + "bbox": [ + 52, + 173, + 219, + 267 + ], + "type": "image", + "image_path": "097f682ea8c12c38c1f1ffaaf7ff2abe8e6df5fe69322eee873526f5fe2030a3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 270, + 148, + 278 + ], + "lines": [ + { + "bbox": [ + 127, + 270, + 148, + 278 + ], + "spans": [ + { + "bbox": [ + 127, + 270, + 148, + 278 + ], + "type": "text", + "content": "MINE" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 173, + 387, + 267 + ], + "blocks": [ + { + "bbox": [ + 220, + 173, + 387, + 267 + ], + "lines": [ + { + "bbox": [ + 220, + 173, + 387, + 267 + ], + "spans": [ + { + "bbox": [ + 220, + 173, + 387, + 267 + ], + "type": "image", + "image_path": "572db4012fa1f463b941f2423b3fb8ee7ba6321dfdaad11c6ddca1db04b9fd48.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 269, + 318, + 277 + ], + "lines": [ + { + "bbox": [ + 288, + 269, + 318, + 277 + ], + "spans": [ + { + "bbox": [ + 288, + 269, + 318, + 277 + ], + "type": "text", + "content": "AdaMPI" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 389, + 173, + 545, + 267 + ], + "blocks": [ + { + "bbox": [ + 389, + 173, + 545, + 267 + ], + "lines": [ + { + "bbox": [ + 389, + 173, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 389, + 173, + 545, + 267 + ], + "type": "image", + "image_path": "5f7f617268f73475a6729d1458798073f875fb9cad7d8ee9a58c1131fbb82dc4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 269, + 487, + 277 + ], + "lines": [ + { + "bbox": [ + 468, + 269, + 487, + 277 + ], + "spans": [ + { + "bbox": [ + 468, + 269, + 487, + 277 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 52, + 285, + 219, + 379 + ], + "blocks": [ + { + "bbox": [ + 52, + 285, + 219, + 379 + ], + "lines": [ + { + "bbox": [ + 52, + 285, + 219, + 379 + ], + "spans": [ + { + "bbox": [ + 52, + 285, + 219, + 379 + ], + "type": "image", + "image_path": "49d27162c44df8a204fde1b953248a24e7d390ba9198728f093af29363924e00.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 380, + 162, + 388 + ], + "lines": [ + { + "bbox": [ + 112, + 380, + 162, + 388 + ], + "spans": [ + { + "bbox": [ + 112, + 380, + 162, + 388 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 220, + 285, + 387, + 379 + ], + "blocks": [ + { + "bbox": [ + 220, + 285, + 387, + 379 + ], + "lines": [ + { + "bbox": [ + 220, + 285, + 387, + 379 + ], + "spans": [ + { + "bbox": [ + 220, + 285, + 387, + 379 + ], + "type": "image", + "image_path": "58b248577670ae1c2b6b7862f78fc86844984b1aa9fb4ceecda86b34370eea78.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 380, + 315, + 388 + ], + "lines": [ + { + "bbox": [ + 290, + 380, + 315, + 388 + ], + "spans": [ + { + "bbox": [ + 290, + 380, + 315, + 388 + ], + "type": "text", + "content": "SVMPI" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 389, + 285, + 545, + 379 + ], + "blocks": [ + { + "bbox": [ + 389, + 285, + 545, + 379 + ], + "lines": [ + { + "bbox": [ + 389, + 285, + 545, + 379 + ], + "spans": [ + { + "bbox": [ + 389, + 285, + 545, + 379 + ], + "type": "image", + "image_path": "f03703d0453e85b3317cb5c982a503706458ef246fcd6fee10d9fd2c11edc4c7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 380, + 487, + 388 + ], + "lines": [ + { + "bbox": [ + 466, + 380, + 487, + 388 + ], + "spans": [ + { + "bbox": [ + 466, + 380, + 487, + 388 + ], + "type": "text", + "content": "VMPI" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 52, + 392, + 219, + 485 + ], + "blocks": [ + { + "bbox": [ + 52, + 392, + 219, + 485 + ], + "lines": [ + { + "bbox": [ + 52, + 392, + 219, + 485 + ], + "spans": [ + { + "bbox": [ + 52, + 392, + 219, + 485 + ], + "type": "image", + "image_path": "9f8ae46a32fcf89b1b05befaa4eada23aa99779cc367bcbe6700d3b758e2162e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 487, + 147, + 495 + ], + "lines": [ + { + "bbox": [ + 127, + 487, + 147, + 495 + ], + "spans": [ + { + "bbox": [ + 127, + 487, + 147, + 495 + ], + "type": "text", + "content": "MINE" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 220, + 392, + 387, + 485 + ], + "blocks": [ + { + "bbox": [ + 220, + 392, + 387, + 485 + ], + "lines": [ + { + "bbox": [ + 220, + 392, + 387, + 485 + ], + "spans": [ + { + "bbox": [ + 220, + 392, + 387, + 485 + ], + "type": "image", + "image_path": "70e328d37e970b678d4f405488b6f162a2f105b71215f5c6a30520439afc33e7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 485, + 317, + 493 + ], + "lines": [ + { + "bbox": [ + 288, + 485, + 317, + 493 + ], + "spans": [ + { + "bbox": [ + 288, + 485, + 317, + 493 + ], + "type": "text", + "content": "AdaMPI" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 389, + 392, + 545, + 485 + ], + "blocks": [ + { + "bbox": [ + 389, + 392, + 545, + 485 + ], + "lines": [ + { + "bbox": [ + 389, + 392, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 389, + 392, + 545, + 485 + ], + "type": "image", + "image_path": "590fbd5d7287e30b57857b1fb04bc3a2ecc07feb23b30d2db0f68f588117d990.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 486, + 487, + 494 + ], + "lines": [ + { + "bbox": [ + 469, + 486, + 487, + 494 + ], + "spans": [ + { + "bbox": [ + 469, + 486, + 487, + 494 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 48, + 536, + 173, + 608 + ], + "blocks": [ + { + "bbox": [ + 46, + 503, + 547, + 528 + ], + "lines": [ + { + "bbox": [ + 46, + 503, + 547, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 547, + 528 + ], + "type": "text", + "content": "Figure 4: Comparing the novel view synthesis results of the baseline methods and our approach on the Tanks and Temples dataset. Our results are better than SVMPI, VMPI and MINE, and comparable to AdaMPI while using far fewer depth planes." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 536, + 173, + 608 + ], + "lines": [ + { + "bbox": [ + 48, + 536, + 173, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 173, + 608 + ], + "type": "image", + "image_path": "689381e5f19e89df23b908c7ccddd5b904e32871b61b9581cd9a31d9db6e94cc.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 173, + 536, + 297, + 608 + ], + "blocks": [ + { + "bbox": [ + 173, + 536, + 297, + 608 + ], + "lines": [ + { + "bbox": [ + 173, + 536, + 297, + 608 + ], + "spans": [ + { + "bbox": [ + 173, + 536, + 297, + 608 + ], + "type": "image", + "image_path": "7743671019847d52e6b4c0901f604c37e8777d4d572fde853624d0b48945484d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 297, + 536, + 421, + 608 + ], + "blocks": [ + { + "bbox": [ + 297, + 536, + 421, + 608 + ], + "lines": [ + { + "bbox": [ + 297, + 536, + 421, + 608 + ], + "spans": [ + { + "bbox": [ + 297, + 536, + 421, + 608 + ], + "type": "image", + "image_path": "747e0e95f70c4b24c39ae960a3fec7b1abd5d337150ac8e644193b88fc1b6ca5.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 421, + 536, + 545, + 608 + ], + "blocks": [ + { + "bbox": [ + 421, + 536, + 545, + 608 + ], + "lines": [ + { + "bbox": [ + 421, + 536, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 421, + 536, + 545, + 608 + ], + "type": "image", + "image_path": "a1e224c48122c66cab594aad955fb886ee2f8e9aa8621147ce4906e856080cb5.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 48, + 608, + 173, + 678 + ], + "blocks": [ + { + "bbox": [ + 48, + 608, + 173, + 678 + ], + "lines": [ + { + "bbox": [ + 48, + 608, + 173, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 608, + 173, + 678 + ], + "type": "image", + "image_path": "db4e8bdeca117ce1f8e31f77529e7fc98b35aa624312fe66075f296e564357a5.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 693, + 530, + 706 + ], + "lines": [ + { + "bbox": [ + 61, + 693, + 530, + 706 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 530, + 706 + ], + "type": "text", + "content": "Figure 5: View synthesis results of our method on the HD " + }, + { + "bbox": [ + 61, + 693, + 530, + 706 + ], + "type": "inline_equation", + "content": "(1080\\times 1920)" + }, + { + "bbox": [ + 61, + 693, + 530, + 706 + ], + "type": "text", + "content": " Davis [30] dataset. Original views are inset." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 173, + 608, + 297, + 678 + ], + "blocks": [ + { + "bbox": [ + 173, + 608, + 297, + 678 + ], + "lines": [ + { + "bbox": [ + 173, + 608, + 297, + 678 + ], + "spans": [ + { + "bbox": [ + 173, + 608, + 297, + 678 + ], + "type": "image", + "image_path": "87313071b9aa01d92b46cfffe7b11be87cf454a1811ca01230e2a377f471f99c.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 297, + 608, + 421, + 678 + ], + "blocks": [ + { + "bbox": [ + 297, + 608, + 421, + 678 + ], + "lines": [ + { + "bbox": [ + 297, + 608, + 421, + 678 + ], + "spans": [ + { + "bbox": [ + 297, + 608, + 421, + 678 + ], + "type": "image", + "image_path": "22fda26b6807b906f95131b175a6a454e553e29346b806d7be7534aed23d9909.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 421, + 608, + 545, + 678 + ], + "blocks": [ + { + "bbox": [ + 421, + 608, + 545, + 678 + ], + "lines": [ + { + "bbox": [ + 421, + 608, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 421, + 608, + 545, + 678 + ], + "type": "image", + "image_path": "672b2781e7fe6008b7dcf44f62781d90d422d4e05df0dca7a75897d6660c3ffd.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10460" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 289, + 340 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 289, + 340 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 289, + 340 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 289, + 340 + ], + "type": "image", + "image_path": "c7a3055412f0e0e4a50f6494ad45f6c09826bf106612bdb1c08205ec67d651c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 347, + 289, + 420 + ], + "lines": [ + { + "bbox": [ + 46, + 347, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 289, + 420 + ], + "type": "text", + "content": "Figure 6: Top to bottom: Input image (Davis dataset); inverse monocular depth from DPT [32]; naive discretization by spacing four planes linearly in the inverse-depth range of each tile; our weighted clustering-based discretization with four planes. Our representation shows finer variation on receding surfaces and suffers fewer tiling artifacts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "content": "map defined by the planes in each tile on a monocular depth input that is perturbed by a small amount of Gaussian noise " + }, + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "inline_equation", + "content": "(\\lambda = 0, \\sigma^2 = 1 \\times 10^{-3})" + }, + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "content": ". We compare our approach to naive linear spacing of depth planes in inverse disparity space, and to vanilla unweighted " + }, + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "content": "-means. Our method is robust to outliers and yields a much better reconstruction. Figure 6 provides qualitative comparison with linear spacing on samples from the Davis dataset [30]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 534, + 288, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 656 + ], + "type": "text", + "content": "Figure 7 evaluates the effect of tile size and number of planes on the quality of view synthesis. In general, the results uphold the intuition that small tiles and a large number of planes improve quality. This trend is less clear, however, for number of planes " + }, + { + "bbox": [ + 46, + 534, + 288, + 656 + ], + "type": "inline_equation", + "content": ">8" + }, + { + "bbox": [ + 46, + 534, + 288, + 656 + ], + "type": "text", + "content": ". This would seem to support the observation of Khakhulin et al. [16] and Hu et al. [13] that reductive models are unable to handle redundant geometry effectively. Moreover, the model's complexity increases proportionately with the number of depth planes, leading to slower convergence for the same number of training steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 669, + 123, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 123, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 123, + 681 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "As previously observed, a tiled multiplane image cannot exploit the elegant warping and compositing equations of a" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 531, + 233 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 531, + 233 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 531, + 233 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 531, + 233 + ], + "type": "image", + "image_path": "3cbb07f10df1902889d721b0156c6606976423c2c71d286fba9e3ac079739cf8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 246, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 319 + ], + "type": "text", + "content": "Figure 7: Evaluating the effect of tile size and number of depth planes on view synthesis quality. In general, the results uphold the intuition that small tiles and a large number of planes improve quality. The trend, however, is less clear when the number of planes " + }, + { + "bbox": [ + 304, + 246, + 545, + 319 + ], + "type": "inline_equation", + "content": ">8" + }, + { + "bbox": [ + 304, + 246, + 545, + 319 + ], + "type": "text", + "content": ", indicating that the model may be unable to handle redundant geometry effectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 328, + 545, + 406 + ], + "blocks": [ + { + "bbox": [ + 307, + 328, + 545, + 406 + ], + "lines": [ + { + "bbox": [ + 307, + 328, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 307, + 328, + 545, + 406 + ], + "type": "image", + "image_path": "15577a5f4af11bb15bc43cb95cc81aaaa245cdc049108abd1a6601c5ac8611ff.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 414, + 545, + 439 + ], + "lines": [ + { + "bbox": [ + 306, + 414, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 545, + 439 + ], + "type": "text", + "content": "Figure 8: Failure cases of our approach: in some cases, fine features are inconsistently reconstructed across tiles." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 457, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 495 + ], + "type": "text", + "content": "traditional MPI for differentiable rendering during training. Furthermore, in some cases our method fails to reconstruct thin features consistently across tiles (Figure 8)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 503, + 379, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 379, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 379, + 516 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 523, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 523, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 545, + 596 + ], + "type": "text", + "content": "We present a method for estimating tiled multiplane images from a single RGB input for 3D photography. This includes a novel approach to adaptively spacing a small number of depth planes within an MPI tile to better represent local features. Our method is lightweight, and points a path to realizing novel view synthesis on mobile and VR devices." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 605, + 365, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 365, + 617 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 365, + 617 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 624, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 310, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 310, + 624, + 545, + 677 + ], + "type": "text", + "content": "[1] Benjamin Attal, Selena Ling, Aaron Gokaslan, Christian Richardt, and James Tompkin. MatryODShka: Real-time 6DoF video view synthesis using multi-sphere images. In European Conference on Computer Vision (ECCV), Aug. 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 680, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 680, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 310, + 680, + 547, + 712 + ], + "type": "text", + "content": "[2] Horace B Barlow et al. Possible principles underlying the transformation of sensory messages. Sensory communication, 1(01):217-233, 1961. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10461" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 70, + 228, + 148 + ], + "blocks": [ + { + "bbox": [ + 76, + 70, + 228, + 148 + ], + "lines": [ + { + "bbox": [ + 76, + 70, + 228, + 148 + ], + "spans": [ + { + "bbox": [ + 76, + 70, + 228, + 148 + ], + "type": "image", + "image_path": "304d47c8828685d112135b487dee026ef179927a7a178b1630a2c9d8474a8208.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 148, + 191, + 159 + ], + "lines": [ + { + "bbox": [ + 127, + 148, + 191, + 159 + ], + "spans": [ + { + "bbox": [ + 127, + 148, + 191, + 159 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 70, + 381, + 148 + ], + "blocks": [ + { + "bbox": [ + 230, + 70, + 381, + 148 + ], + "lines": [ + { + "bbox": [ + 230, + 70, + 381, + 148 + ], + "spans": [ + { + "bbox": [ + 230, + 70, + 381, + 148 + ], + "type": "image", + "image_path": "c346e1eaabb8454f35f6b5c2f928dcfb1619cf8e5707566d03a81d939be7a4d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 148, + 319, + 159 + ], + "lines": [ + { + "bbox": [ + 288, + 148, + 319, + 159 + ], + "spans": [ + { + "bbox": [ + 288, + 148, + 319, + 159 + ], + "type": "text", + "content": "SVMPI" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 383, + 70, + 520, + 148 + ], + "blocks": [ + { + "bbox": [ + 383, + 70, + 520, + 148 + ], + "lines": [ + { + "bbox": [ + 383, + 70, + 520, + 148 + ], + "spans": [ + { + "bbox": [ + 383, + 70, + 520, + 148 + ], + "type": "image", + "image_path": "145c469ab0aa27995c281690b5e73f0eca8040170466179c0bfdc49955b90b31.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 148, + 471, + 158 + ], + "lines": [ + { + "bbox": [ + 444, + 148, + 471, + 158 + ], + "spans": [ + { + "bbox": [ + 444, + 148, + 471, + 158 + ], + "type": "text", + "content": "VMPI" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 75, + 160, + 228, + 246 + ], + "blocks": [ + { + "bbox": [ + 75, + 160, + 228, + 246 + ], + "lines": [ + { + "bbox": [ + 75, + 160, + 228, + 246 + ], + "spans": [ + { + "bbox": [ + 75, + 160, + 228, + 246 + ], + "type": "image", + "image_path": "da7f5e229080bd2bd42eca9888f86ba10cfa1b00ddfaf5468aed200e39d31b82.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 247, + 172, + 256 + ], + "lines": [ + { + "bbox": [ + 146, + 247, + 172, + 256 + ], + "spans": [ + { + "bbox": [ + 146, + 247, + 172, + 256 + ], + "type": "text", + "content": "MINE" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 229, + 160, + 381, + 246 + ], + "blocks": [ + { + "bbox": [ + 229, + 160, + 381, + 246 + ], + "lines": [ + { + "bbox": [ + 229, + 160, + 381, + 246 + ], + "spans": [ + { + "bbox": [ + 229, + 160, + 381, + 246 + ], + "type": "image", + "image_path": "da4a4744a9b4f9ec02700b1b92a3fd32c253a48196921c6df42b3b020f4be215.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 247, + 323, + 256 + ], + "lines": [ + { + "bbox": [ + 285, + 247, + 323, + 256 + ], + "spans": [ + { + "bbox": [ + 285, + 247, + 323, + 256 + ], + "type": "text", + "content": "AdaMPI" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 383, + 160, + 520, + 246 + ], + "blocks": [ + { + "bbox": [ + 383, + 160, + 520, + 246 + ], + "lines": [ + { + "bbox": [ + 383, + 160, + 520, + 246 + ], + "spans": [ + { + "bbox": [ + 383, + 160, + 520, + 246 + ], + "type": "image", + "image_path": "e8edcc08156e8c9157aa5a9b35e23755173db93e55d4b061f0190da0efe545c6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 247, + 470, + 256 + ], + "lines": [ + { + "bbox": [ + 446, + 247, + 470, + 256 + ], + "spans": [ + { + "bbox": [ + 446, + 247, + 470, + 256 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 76, + 268, + 228, + 354 + ], + "blocks": [ + { + "bbox": [ + 76, + 268, + 228, + 354 + ], + "lines": [ + { + "bbox": [ + 76, + 268, + 228, + 354 + ], + "spans": [ + { + "bbox": [ + 76, + 268, + 228, + 354 + ], + "type": "image", + "image_path": "a21dcc29f2aeceaf5dfd84fc8c60f4e7abd6892ab6a7b0dc4f7127b670c1a5b9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 229, + 268, + 381, + 354 + ], + "blocks": [ + { + "bbox": [ + 229, + 268, + 381, + 354 + ], + "lines": [ + { + "bbox": [ + 229, + 268, + 381, + 354 + ], + "spans": [ + { + "bbox": [ + 229, + 268, + 381, + 354 + ], + "type": "image", + "image_path": "d148ab71e9ab70ccfebd50a85b601424bbea3601713f8e77d19df476f35f9b7e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 383, + 268, + 520, + 354 + ], + "blocks": [ + { + "bbox": [ + 383, + 268, + 520, + 354 + ], + "lines": [ + { + "bbox": [ + 383, + 268, + 520, + 354 + ], + "spans": [ + { + "bbox": [ + 383, + 268, + 520, + 354 + ], + "type": "image", + "image_path": "8dda61d1d1ecb50c79cc281fc8316b79ebb16f2226d8f1e07937f0ba18a4de9b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 75, + 368, + 228, + 453 + ], + "blocks": [ + { + "bbox": [ + 127, + 357, + 191, + 367 + ], + "lines": [ + { + "bbox": [ + 127, + 357, + 191, + 367 + ], + "spans": [ + { + "bbox": [ + 127, + 357, + 191, + 367 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 368, + 228, + 453 + ], + "lines": [ + { + "bbox": [ + 75, + 368, + 228, + 453 + ], + "spans": [ + { + "bbox": [ + 75, + 368, + 228, + 453 + ], + "type": "image", + "image_path": "052b27b31fa575bee61292c07bb2f1593ec8aa7cf5b60a0871f081fa55d21377.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 456, + 172, + 465 + ], + "lines": [ + { + "bbox": [ + 146, + 456, + 172, + 465 + ], + "spans": [ + { + "bbox": [ + 146, + 456, + 172, + 465 + ], + "type": "text", + "content": "MINE" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 229, + 368, + 381, + 453 + ], + "blocks": [ + { + "bbox": [ + 288, + 357, + 319, + 366 + ], + "lines": [ + { + "bbox": [ + 288, + 357, + 319, + 366 + ], + "spans": [ + { + "bbox": [ + 288, + 357, + 319, + 366 + ], + "type": "text", + "content": "SVMPI" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 229, + 368, + 381, + 453 + ], + "lines": [ + { + "bbox": [ + 229, + 368, + 381, + 453 + ], + "spans": [ + { + "bbox": [ + 229, + 368, + 381, + 453 + ], + "type": "image", + "image_path": "fec7b93b2942740b0578375aeaaa6c182bbaa7314c61111576cc8ba8aca8b804.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 456, + 323, + 465 + ], + "lines": [ + { + "bbox": [ + 285, + 456, + 323, + 465 + ], + "spans": [ + { + "bbox": [ + 285, + 456, + 323, + 465 + ], + "type": "text", + "content": "AdaMPI" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 383, + 368, + 520, + 453 + ], + "blocks": [ + { + "bbox": [ + 444, + 357, + 471, + 366 + ], + "lines": [ + { + "bbox": [ + 444, + 357, + 471, + 366 + ], + "spans": [ + { + "bbox": [ + 444, + 357, + 471, + 366 + ], + "type": "text", + "content": "VMPI" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 368, + 520, + 453 + ], + "lines": [ + { + "bbox": [ + 383, + 368, + 520, + 453 + ], + "spans": [ + { + "bbox": [ + 383, + 368, + 520, + 453 + ], + "type": "image", + "image_path": "5e642b2d76297fdc0fd52d4f4e689ab590fbc490d218aecee2431ce2c27eac6e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 456, + 470, + 464 + ], + "lines": [ + { + "bbox": [ + 446, + 456, + 470, + 464 + ], + "spans": [ + { + "bbox": [ + 446, + 456, + 470, + 464 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 76, + 475, + 228, + 559 + ], + "blocks": [ + { + "bbox": [ + 76, + 475, + 228, + 559 + ], + "lines": [ + { + "bbox": [ + 76, + 475, + 228, + 559 + ], + "spans": [ + { + "bbox": [ + 76, + 475, + 228, + 559 + ], + "type": "image", + "image_path": "0a8340d8bf2e211fc04a3befeb94c5110340b884fa9530ad3c7e356086857c87.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 230, + 475, + 382, + 559 + ], + "blocks": [ + { + "bbox": [ + 230, + 475, + 382, + 559 + ], + "lines": [ + { + "bbox": [ + 230, + 475, + 382, + 559 + ], + "spans": [ + { + "bbox": [ + 230, + 475, + 382, + 559 + ], + "type": "image", + "image_path": "1fa0d4d74591c473e12bfffb405f7499600af16bcef976a026b8bb87b39b27df.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 383, + 475, + 520, + 559 + ], + "blocks": [ + { + "bbox": [ + 383, + 475, + 520, + 559 + ], + "lines": [ + { + "bbox": [ + 383, + 475, + 520, + 559 + ], + "spans": [ + { + "bbox": [ + 383, + 475, + 520, + 559 + ], + "type": "image", + "image_path": "078ca2fa2a9593a01fb37b7386f95bb6572b5cee085333145b4d107a175fad70.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 560, + 471, + 570 + ], + "lines": [ + { + "bbox": [ + 444, + 560, + 471, + 570 + ], + "spans": [ + { + "bbox": [ + 444, + 560, + 471, + 570 + ], + "type": "text", + "content": "VMPI" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 76, + 572, + 228, + 658 + ], + "blocks": [ + { + "bbox": [ + 128, + 561, + 191, + 571 + ], + "lines": [ + { + "bbox": [ + 128, + 561, + 191, + 571 + ], + "spans": [ + { + "bbox": [ + 128, + 561, + 191, + 571 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 572, + 228, + 658 + ], + "lines": [ + { + "bbox": [ + 76, + 572, + 228, + 658 + ], + "spans": [ + { + "bbox": [ + 76, + 572, + 228, + 658 + ], + "type": "image", + "image_path": "ec0528a6027fa37d44d330f9d2388455b92e00a0359a1302e65d1ae6ff807dee.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 659, + 172, + 669 + ], + "lines": [ + { + "bbox": [ + 146, + 659, + 172, + 669 + ], + "spans": [ + { + "bbox": [ + 146, + 659, + 172, + 669 + ], + "type": "text", + "content": "MINE" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 685, + 533, + 697 + ], + "lines": [ + { + "bbox": [ + 58, + 685, + 533, + 697 + ], + "spans": [ + { + "bbox": [ + 58, + 685, + 533, + 697 + ], + "type": "text", + "content": "Figure 9: Comparing the novel view synthesis results of the baseline methods and our approach on the Spaces dataset." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 229, + 572, + 381, + 658 + ], + "blocks": [ + { + "bbox": [ + 288, + 561, + 319, + 571 + ], + "lines": [ + { + "bbox": [ + 288, + 561, + 319, + 571 + ], + "spans": [ + { + "bbox": [ + 288, + 561, + 319, + 571 + ], + "type": "text", + "content": "SVMPI" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 229, + 572, + 381, + 658 + ], + "lines": [ + { + "bbox": [ + 229, + 572, + 381, + 658 + ], + "spans": [ + { + "bbox": [ + 229, + 572, + 381, + 658 + ], + "type": "image", + "image_path": "8dde621a480905b77b0e49619bf4ca9d7d539d6bc400447d6c53bf250a3dfdf1.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 659, + 323, + 669 + ], + "lines": [ + { + "bbox": [ + 285, + 659, + 323, + 669 + ], + "spans": [ + { + "bbox": [ + 285, + 659, + 323, + 669 + ], + "type": "text", + "content": "AdaMPI" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 383, + 572, + 520, + 658 + ], + "blocks": [ + { + "bbox": [ + 383, + 572, + 520, + 658 + ], + "lines": [ + { + "bbox": [ + 383, + 572, + 520, + 658 + ], + "spans": [ + { + "bbox": [ + 383, + 572, + 520, + 658 + ], + "type": "image", + "image_path": "310c22732e27657398b0997f1f70ffc7e9b0367a74428e9295ab31022c7cd7cf.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 659, + 470, + 669 + ], + "lines": [ + { + "bbox": [ + 446, + 659, + 470, + 669 + ], + "spans": [ + { + "bbox": [ + 446, + 659, + 470, + 669 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10462" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 127 + ], + "type": "text", + "content": "[3] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4):86-1, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 128, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 128, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 128, + 288, + 172 + ], + "type": "text", + "content": "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocosstuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 172, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 288, + 216 + ], + "type": "text", + "content": "[5] Jin-Xiang Chai, Xin Tong, Shing-Chow Chan, and Heung-Yeung Shum. Plenoptic sampling. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 307-318, 2000. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 217, + 288, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 272 + ], + "type": "text", + "content": "[6] Gaurav Chaurasia, Arthur Nieuwoudt, Alexandru-Eugen Ichim, Richard Szeliski, and Alexander Sorkine-Hornung. Passthrough+ real-time stereoscopic view synthesis for mobile mixed reality. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 3(1):1-17, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "text", + "content": "[7] Inchang Choi, Orazio Gallo, Alejandro Troccoli, Min H Kim, and Jan Kautz. Extreme view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7781-7790, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 316, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 316, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 316, + 288, + 348 + ], + "type": "text", + "content": "[8] Abe Davis, Marc Levoy, and Fredo Durand. Unstructured light fields. In Computer Graphics Forum, volume 31, pages 305-314. Wiley Online Library, 2012. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 349, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 349, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 349, + 288, + 414 + ], + "type": "text", + "content": "[9] John Flynn, Michael Broxton, Paul Debevec, Matthew DuVall, Graham Fyffe, Ryan Overbeck, Noah Snavely, and Richard Tucker. Deepview: View synthesis with learned gradient descent. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2367-2376, 2019. 2, 3, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "text", + "content": "[10] Sushobhan Ghosh, Zhaoyang Lv, Nathan Matsuda, Lei Xiao, Andrew Berkovich, and Oliver Cossairt. Liveview: dynamic target-centered mpi for view synthesis. arXiv preprint arXiv:2107.05113, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 502 + ], + "type": "text", + "content": "[11] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 43-54, 1996. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 536 + ], + "type": "text", + "content": "[12] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-view synthesis in the wild with learned adaptive multi-plane images. In ACM SIGGRAPH, 2022. 2, 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 591 + ], + "type": "text", + "content": "[13] Ronghang Hu, Nikhila Ravi, Alexander C Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12528-12537, 2021. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 657 + ], + "type": "text", + "content": "[14] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, and Ce Liu. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 691 + ], + "type": "text", + "content": "[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[16] Taras Khakhulin, Denis Korzhenkov, Pavel Solovev, Gleb Sterkin, Andrei-Timotei Ardelean, and Victor Lempitsky." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 546, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 106 + ], + "type": "text", + "content": "Stereo magnification with multi-layer images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8687-8696, 2022. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 107, + 546, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 546, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 546, + 150 + ], + "type": "text", + "content": "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 152, + 546, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 546, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 546, + 206 + ], + "type": "text", + "content": "[18] Johannes Kopf, Suhib Alsisan, Francis Ge, Yangming Chong, Kevin Matzen, Ocean Quigley, Josh Patterson, Jossie Tirado, Shu Wu, and Michael F Cohen. Practical 3d photography. In Proceedings of CVPR Workshops, volume 1, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 208, + 546, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 546, + 262 + ], + "type": "text", + "content": "[19] Johannes Kopf, Kevin Matzen, Suhib Alsisan, Ocean Quigley, Francis Ge, Yangming Chong, Josh Patterson, Jan-Michael Frahm, Shu Wu, Matthew Yu, et al. One shot 3d photography. ACM Transactions on Graphics (TOG), 39(4):76-1, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 546, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 546, + 297 + ], + "type": "text", + "content": "[20] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 31-42, 1996. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 298, + 546, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 546, + 363 + ], + "type": "text", + "content": "[21] David Li, Yinda Zhang, Christian Hane, Danhang Tang, Amitabh Varshney, and Ruofei Du. Omnisyn: Synthesizing 360 videos with wide-baseline panoramas. In 2022 IEEE Conference on Virtual Reality and 3D User Interfaces Abstracts and Workshops (VRW), pages 670-671. IEEE, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 546, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 546, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 546, + 408 + ], + "type": "text", + "content": "[22] Jiaxin Li, Zijian Feng, Qi She, Henghui Ding, Changhu Wang, and Gim Hee Lee. MINE: Towards continuous depth spi with nerf for novel view synthesis. In ICCV, 2021. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 411, + 546, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 546, + 444 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 546, + 444 + ], + "type": "text", + "content": "[23] Qinbo Li and Nima Khademi Kalantari. Synthesizing light field from a single image with variable mpi and two network fusion. ACM Trans. Graph., 39(6):229-1, 2020. 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 445, + 546, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 546, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 546, + 499 + ], + "type": "text", + "content": "[24] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4521-4530, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 500, + 546, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 546, + 578 + ], + "type": "text", + "content": "[25] Diogo C Luvizon, Gustavo Sutter P Carvalho, Andreza A dos Santos, Jhonatas S Conceicao, Jose L Flores-Campana, Luis GL Decker, Marcos R Souza, Helio Pedrini, Antonio Joia, and Otavio AB Penatti. Adaptive multiplane image generation from a single internet picture. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2556-2565, 2021. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 579, + 546, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 546, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 546, + 634 + ], + "type": "text", + "content": "[26] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 1, 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 635, + 546, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 546, + 689 + ], + "type": "text", + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 690, + 546, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 690, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 308, + 690, + 546, + 715 + ], + "type": "text", + "content": "[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a mul" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10463" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "tiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[29] Simon Niklaus, Long Mai, Jimei Yang, and Feng Liu. 3d ken burns effect from a single image. ACM Transactions on Graphics (ToG), 38(6):1-15, 2019. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 174 + ], + "type": "text", + "content": "[30] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv:1704.00675, 2017. 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 208 + ], + "type": "text", + "content": "[31] Thomas Porter and Tom Duff. Compositing digital images. In Proceedings of the 11th annual conference on Computer graphics and interactive techniques, pages 253-259, 1984. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 288, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 253 + ], + "type": "text", + "content": "[32] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 5, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 255, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 308 + ], + "type": "text", + "content": "[33] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "type": "text", + "content": "[34] Gernot Riegler and Vladlen Koltun. Free view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 623-640. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "type": "text", + "content": "[35] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12216-12225, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 400, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 287, + 433 + ], + "type": "text", + "content": "[36] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 477 + ], + "type": "text", + "content": "[37] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 521 + ], + "type": "text", + "content": "[38] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 524, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 287, + 578 + ], + "type": "text", + "content": "[39] Pratul P Srinivasan, Richard Tucker, Jonathan T Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snively. Pushing the boundaries of view extrapolation with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 175-184, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 580, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 622 + ], + "type": "text", + "content": "[40] Richard Szeliski and Polina Golland. Stereo matching with transparency and matting. In Sixth International Conference on Computer Vision (IEEE Cat. No. 98CH36271), pages 517-524. IEEE, 1998. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "text", + "content": "[41] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[42] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In The IEEE Conference" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 664 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition (CVPR), June 2020. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[43] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12922–12931, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[44] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[45] Silvan Weder, Johannes Schonberger, Marc Pollefeys, and Martin R Oswald. Routedfusion: Learning real-time depth map fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4887-4897, 2020. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 252, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 545, + 307 + ], + "type": "text", + "content": "[46] Bennett Wilburn, Neel Joshi, Vaibhav Vaish, Eino-Ville Talvala, Emilio Antunez, Adam Barth, Andrew Adams, Mark Horowitz, and Marc Levoy. High performance imaging using large camera arrays. In ACM SIGGRAPH 2005 Papers, pages 765-776. 2005. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 308, + 545, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 362 + ], + "type": "text", + "content": "[47] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 365, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 419 + ], + "type": "text", + "content": "[48] Suttisak Wizadwongsa, Pakkapon Phongthawee, Jiraphon Yenphraphai, and Supasorn Suwajanakorn. Nex: Real-time view synthesis with neural basis expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8534-8543, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 420, + 545, + 464 + ], + "type": "text", + "content": "[49] Lei Xiao, Salah Nouri, Joel Hegland, Alberto Garcia Garcia, and Douglas Lanman. Neuralpassthrough: Learned real-time view synthesis for vr. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 465, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 465, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 465, + 545, + 529 + ], + "type": "text", + "content": "[50] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 532, + 545, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 532, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 308, + 532, + 545, + 585 + ], + "type": "text", + "content": "[51] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 587, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 587, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 308, + 587, + 545, + 620 + ], + "type": "text", + "content": "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 622, + 545, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 622, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 308, + 622, + 545, + 664 + ], + "type": "text", + "content": "[53] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018. 2, 3, 4, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10464" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/1b69a328-51e9-4f52-9cd0-ceb8e6de4113_content_list.json b/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/1b69a328-51e9-4f52-9cd0-ceb8e6de4113_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..306572a30d1d36b02814382a0f8a419115e7bfbb --- /dev/null +++ b/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/1b69a328-51e9-4f52-9cd0-ceb8e6de4113_content_list.json @@ -0,0 +1,1667 @@ +[ + { + "type": "text", + "text": "Time Does Tell: Self-Supervised Time-Tuning of Dense Image Representations", + "text_level": 1, + "bbox": [ + 91, + 130, + 879, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mohammadreza Salehi, Efstratios Gavves, Cees G. M. Snoek, Yuki M. Asano \nQUVA Lab, University of Amsterdam", + "bbox": [ + 176, + 180, + 792, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(s-salehidehnavi, e.gavves, c.g.m.snoek, y.m.asano)@uva.nl", + "bbox": [ + 228, + 219, + 743, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spatially dense self-supervised learning is a rapidly growing problem domain with promising applications for unsupervised segmentation and pretraining for dense downstream tasks. Despite the abundance of temporal data in the form of videos, this information-rich source has been largely overlooked. Our paper aims to address this gap by proposing a novel approach that incorporates temporal consistency in dense self-supervised learning. While methods designed solely for images face difficulties in achieving even the same performance on videos, our method improves not only the representation quality for videos – but also images. Our approach, which we call time-tuning, starts from image-pretrained models and fine-tunes them with a novel self-supervised temporal-alignment clustering loss on unlabeled videos. This effectively facilitates the transfer of high-level information from videos to image representations. Time-tuning improves the state-of-the-art by $8 - 10\\%$ for unsupervised semantic segmentation on videos and matches it for images. We believe this method paves the way for further self-supervised scaling by leveraging the abundant availability of videos. The implementation can be found here: https://github.com/SMSD75/Timetuning", + "bbox": [ + 75, + 300, + 473, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 662, + 209, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dense self-supervised learning, whereby meaningful deep features for each pixel or patch of input are learned in an unsupervised manner, has recently received increasing attention [74, 24, 57, 51]. By learning spatially consistent features for different views of an input, strong gains in unsupervised semantic segmentation have been achieved using unlabeled images. However, so far, an even more information-rich source for unsupervised training has been largely overlooked: videos. With their additional time dimension and being the most rapidly growing form of digital content, they are well-suited to scaling dense self-supervised learning even further.", + "bbox": [ + 75, + 688, + 470, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some efforts have already been made to learn from the video domain by using different frames from a video as aug-", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d6a21dd569aa53cf12c8bad541e03655ae4cf6be65c2e2b8b02480bded49a00.jpg", + "image_caption": [ + "Figure 1: Time-tuning compared to previous methods. Unlike existing methods that ignore or utilize expensive 3D models to implicitly model temporal information, the proposed method explicitly incorporates temporal consistency in dense feature representations using a temporal self-supervised loss. The method starts with a 2D encoder pretrained on images and fine-tunes it using unlabeled videos. This approach leads to improved performance not only for videos but also for images." + ], + "image_footnote": [], + "bbox": [ + 501, + 271, + 883, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mentations [18, 63, 49] or by involving temporal correspondence [71, 58]; however, they mostly did it in a supervised way [29, 73, 19, 40, 65, 38, 31, 28, 41, 37, 54, 44], which is not scalable specifically for dense tasks where the number of targets can increase significantly as the number of pixels grows. To this end, self-supervised learning approaches offer a solution by reducing the need for supervision. However, these methods typically rely on the notion of \"views\", which involves learning similar features for corresponding locations over time. This usually leads to a chicken-and-egg problem, where the correspondences are required for learning dense features - which in turn enable good correspondences [30].", + "bbox": [ + 496, + 626, + 893, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In images, the challenge is trivially solved by considering the correspondence function based on the augmentation function. For instance, In the case of color augmentations, this correspondence is simply given by the identity. However, shifts in time cannot be viewed as mere augmenta", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16536", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions. As we demonstrate through a new evaluation protocol, using image models on frames alone is not nearly as effective. A similar finding has also been reported, albeit for non-dense works [20, 34], which assumed the passage of time as an image augmentation. These works have generally reported reduced performances, even when compared to simple image-based pretraining methods. Similarly, video-level tasks [17, 18, 55] assume sufficiently similar semantics between different frames. This is also not true for dense tasks, as static features can only be assumed where nothing is moving – which is rare due to possible object, camera, and background motion between the frames.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we propose to model the additional time-dimension explicitly to identify which pixels should retain similar embeddings and which should not. We propose two separate modules to tackle the correspondence and the dense learning, respectively. For the former, we introduce the Feature-Forwarder (FF) module, which breaks the mentioned chicken-and-egg loop by leveraging the good tracking performance of pretrained image models, and allows an approximate second \"view\" that can then be treated as a target for the further dense self-supervised loss. On top of this, we introduce a spatio-temporally dense clustering module, which learns unsupervised clusters across samples, locations and time. Using these two components and starting from image-pretrained features, our proposed method allows time-tuning (TIMET) the dense representation in a self-supervised manner, see Figure 1.", + "bbox": [ + 75, + 273, + 472, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we demonstrate that TIMET paves the way for further scaling of self-supervised learning by leveraging the abundant availability of video datasets and transferring their knowledge to the image domain. This results in consistently achieving state-of-the-art performances not only for the task of unsupervised semantic segmentation of videos, but also for unsupervised image semantic segmentation, a feat previously out of reach for methods trained on videos.", + "bbox": [ + 75, + 515, + 472, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, this paper makes the following contributions:", + "bbox": [ + 96, + 636, + 454, + 652 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We show that image-based unsupervised dense segmentation models applied to videos exhibit degraded performance and lack temporal consistency in their segmentation maps.", + "- Building on this observation, we propose a novel dense self-supervised learning method that utilizes temporal consistency as a learning signal.", + "- We demonstrate that our method enables the scaling of self-supervised learning by leveraging abundant video datasets and effectively transferring knowledge to the image domain. Our approach consistently achieves state-of-the-art performance for both images and videos, opening up new opportunities in the field." + ], + "bbox": [ + 94, + 664, + 470, + 887 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 89, + 650, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dense self-supervised learning. These methods build upon image-level self-supervised representation learning by incorporating existing losses to enhance the spatial features, demonstrating a commendable advancement. DenseCL [61] works on spatial features by constructing dense correspondences across views using the contrastive objective given in MoCo [23], while PixPro [62] utilizes the augmentation wrapper to get the spatial correspondence of the pixel intersection between two views. Similarly, MaskContrast [57], Leopart [74], DetCon [24] and Odin [25] also ensure spatial feature similarities via contrastive learning and spatial correspondences. Self-Patch [67] treats the spatial neighbors of the patch as positive examples for learning more semantically meaningful relations among patches. Inspired by SelfPatch, ADCLR [69] proposes patch-level contrasting via query crop and cross-attention mechanism. Pursuing the same objective through none end-to-end approaches. Both [68] and [59] propose an unsupervised salient object segmentation pipeline that extracts noisy object masks from the inputs and fine-tunes a specifically designed object segmentation head with several self-training steps to make an unsupervised semantic segmentation model. Unlike the existing image-based approaches, we propose a method that improves the dense prediction performance of a pretrained encoder by explicitly modeling the temporal dimension and learning from diverse natural dynamics and variations found in videos.", + "bbox": [ + 496, + 114, + 893, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video to image knowledge transfer. Learning from videos instead of images has recently received attention since they contain far more information than still images and hold the potential for learning rich representations of the visual world. To this end, several non-dense works have been released [46, 4, 27]. VITO [46] shows that naively applying image domain self-supervised learning methods on videos can lead to a performance drop coming from the distribution shift, which by applying data processing techniques is relieved but not fully solved. Following the same way, [27] trains masked autoencoders [22] with contrastive learning [10] on video datasets and shows a decent performance on both image and video tasks. While video to image knowledge transfer is the end goal of such methods and our paper; our dense task is different and is as of yet unaddressed.", + "bbox": [ + 496, + 545, + 893, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised video object segmentation. Unsupervised video object segmentation methods do not require any manual annotations. However, they are only designed to tackle a foreground/background segmentation task, which refers to the segmentation of the most prominent, general objects in video sequences [66, 43, 50, 1, 39]. For instance, [66]", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16537", + "bbox": [ + 480, + 924, + 519, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/28984550502a23d9386dd0e8146f9d52eb8369dfc1db154caa89a6f52ebad4ca.jpg", + "image_caption": [ + "Figure 2: Time-Tuning overview. We conduct self-supervised video semantic segmentation by tuning image-pretrained models with temporal information from videos. (a): Our general pipeline of adapting an image-pretrained ViT transformer on video data using our dense clustering loss. The encoder is specified by $\\Psi$ and kept frozen except for the last two layers. Also, $z_{t}$ shows the output of Sinkhorn-Knopp algorithm, which is forwarded to time-step $T$ to be compared with the features that are obtained using the last time-step. (b): Detailed view into our Feature-Forwarder module that is used for aligning cluster-maps from past frames to logit-features of future ones." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 467, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fc73b7e624c145655f8448ce3e80b31cca95cd290dd4ee16fc8e3c875e621a2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 88, + 870, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "uses an AutoEncoder-based architecture based on slot attention so that pixels that show the same motion flows based on optical-flow are grouped together. In this way, dominant moving objects can be detected and separated from the background. In our experiments, we benchmark two representative instances of this line of work, i.e., [66, 1], to provide better insights into the differences of unsupervised object detection methods and our proposed benchmark for video semantic segmentation.", + "bbox": [ + 75, + 387, + 468, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To conclude, we propose the first self-supervised video semantic segmentation method, which tunes image-trained models such as [9, 72] based on the temporal information that exists in unlabeled videos to enhance their effectiveness in dense prediction tasks.", + "bbox": [ + 75, + 523, + 468, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 613, + 169, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method works by densely clustering features in a manner that is consistent with time. At a high level, it works by transforming features from a past time $t' = t$ to the current frame $t' = T$ and forcing these to be consistent with the currently observed ones at $t' = T$ . The overall learning signal comes from the dense temporal clustering of forwarded features and allows \"timetuning\" image-pretrained models to learn from videos. Figure 2 shows the proposed method's architecture. In the following, we describe each component in detail.", + "bbox": [ + 75, + 638, + 470, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Feature Forwarding", + "text_level": 1, + "bbox": [ + 76, + 801, + 269, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of this component, the feature-forwarder (FF), is to propagate the dense features in time as accurately as possible, given some RGB input frames. More formally, let $f$ be the feature forwarding function, $I_{t}$ be a frame at time $t$ for a given video, and $z_{t}$ the features corresponding to $I_{t}$ ,", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "then the task is to propagate $z_{t}$ to a future time $T$ ,", + "bbox": [ + 500, + 387, + 830, + 402 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(\\left[ I _ {t}, \\dots , I _ {T} \\right], z _ {t}\\right)\\rightarrow \\hat {z} _ {T}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 414, + 890, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that $\\hat{z}_T$ and $z_{T}$ are not necessarily the same for $T > t$ ; since $\\hat{z}_T$ represents an approximation of $z_{T}$ using the previous features. Obviously, the further $T$ is from $t$ , the more variance is seen and therefore a better learning signal might be provided; however, for such larger intervals, the idea of feature forwarding encounters two main challenges: error accumulation and object non-permanence.", + "bbox": [ + 496, + 443, + 893, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenge 1: Error accumulation. While computing how features have changed simply based on the start and end frames is the most straightforward (i.e., $I_{t}$ and $I_{T}$ ), it does not fully leverage the knowledge contained in the intermediate frames. Instead, it is common [30, 60, 36] to forward features for every time-step $\\delta t$ , i.e.,", + "bbox": [ + 496, + 568, + 893, + 659 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf ([ I _ {t},.., I _ {T} ], z _ {t}) = \\bigcirc_ {t ^ {\\prime} = t + \\delta t} ^ {T} f ([ I _ {t ^ {\\prime} - \\delta t}, I _ {t ^ {\\prime}} ], z _ {t} ]), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 669, + 890, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bigcirc$ is a composition operator such as element-wise multiplication. However, this results in the accumulation of errors over time.", + "bbox": [ + 496, + 699, + 893, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenge 2: Object non-permanence. One difficulty that arises when using videos as training data is the fact that objects sometimes simply disappear from the screen. Compared to images, which carry a heavy photographer's bias, videos are more prone to natural variation due to camera and object motion resulting in temporary object non-permanence in a particular video clip. A method that simply assumes an object (or a feature) has to be present in any given frame simply because it was present in the previous", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16538", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "one might therefore easily fail when going towards videos that exist in-the-wild. A case in point are simple occlusions that arise from an object being fully- or partially-covered by another one further in the background. Taking longer training clips has been attempted to resolve the issue. However, this increases the training complexity, especially for dense-tasks that necessitate a higher number of predictions for each input.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We address these challenges using a novel, yet simple component, which we call the Stabilizing Feature-Forwarder. The first choice in designing this module is the function and data used for composing information across time. For instance, a simple approach would be to use pixel-wise optical flow and aggregate the flow at the feature level to predict how the features change over time. However, as we also show in our experiments, optical flow is prone to accumulating error with time and also cannot easily recover from occlusions [6]. Instead, we utilize the fact that we wish to pretrain the visual encoder and recycle part of it for the purpose of feature forwarding. This not only speeds up the process, as no additional encoder/modality is required, but also produces a positive feedback loop of better forwarding and better feature learning.", + "bbox": [ + 75, + 224, + 472, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, we first compute the L2-normalised dense features of a pretrained visual encoder $\\Phi$ to compute spatial similarities across time (the ' in $t^\\prime$ left out for clarity):", + "bbox": [ + 76, + 450, + 470, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i j} ^ {t} = \\left\\langle \\Phi \\left(I _ {t}\\right) _ {i}, \\Phi \\left(I _ {T}\\right) _ {j} \\right\\rangle / \\tau , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 502, + 468, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which yields matrices $F^{\\{t,\\dots T - \\delta t\\}} \\in R^{N\\times N}$ with values between [0,1] indicating semantic similarities between the $N$ spatial features that are sharpened by a temperature $\\tau$ . Note that because we compare each frame with the final frame, the effect of object non-permanence can be minimized: even if the object is only present in every other frame, there is enough signal to forward its features.", + "bbox": [ + 75, + 526, + 468, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next to propagate from time $t$ to $T$ , these similarities are stacked and normalised along time, as follows:", + "bbox": [ + 76, + 633, + 468, + 664 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {F} _ {i j} ^ {t} = \\exp \\left(\\mathcal {N} \\left(F _ {i j} ^ {t}\\right)\\right) / \\sum_ {i, t ^ {\\prime}} \\exp \\left(\\mathcal {N} \\left(F _ {i j} ^ {t ^ {\\prime}}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 681, + 468, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t \\leq t' \\leq T$ and $\\mathcal{N}$ is a neighborhood thresholding method which forces short-term spatial-smoothness, i.e., $\\mathcal{N}(\\tilde{F}_{ij}^t) = 0$ if $i$ is not within a local window around $j$ with the size $k$ . Finally, the propagated feature are computed $\\hat{z}_T$ as:", + "bbox": [ + 76, + 720, + 468, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {T} (j) = \\sum_ {t ^ {\\prime}, i} \\tilde {F} _ {i j} ^ {t ^ {\\prime}} \\hat {z} _ {t ^ {\\prime}} (i), \\quad j \\in \\{1,.., N \\}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 816, + 468, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This means that for arriving at the target feature $\\hat{z}_T$ , we not only use the previous frame as the source, but instead use the past $(T - t) / \\delta t$ frames and aggregate these.", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relation to mask-propagation methods. While previous methods such as DINO [9] and STC [30] have utilized a similar technique for propagating ground-truth masks for evaluating, for example on the DAVIS dataset, there are three key differences to our forwarding method. First, instead of propagating binary maps of foreground-vs-background, our feature-forwarder has as inputs soft, noisy and multi-label self-supervised segmentation maps, which require the forwarder to tolerate overlapping of different object probabilities throughout training. Second, previous methods have used this approach mainly for inference; we, however, use this module as a trainable component and show how our loss improves this forwarder with training time. Finally, we do not follow a typical re-normalizing step across feature dimensions (typically done after Eq. (5)) as this harms the scale of logits that are being propagated and leads to heavily diluted target distributions.", + "bbox": [ + 496, + 90, + 893, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Self-supervised dense clustering", + "text_level": 1, + "bbox": [ + 500, + 357, + 782, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the Feature-Forwarder produces target features that include information about the dynamics with time, its computation utilizes $\\Phi(I_T)$ and could lead to trivial solutions, i.e., $f([I_t,..,I_T],z_t) = \\Phi(I_T)$ . To counteract this, we propose a self-supervised clustering task across views in time. For this, we utilize the basic online clustering algorithm based optimal-transport [12], utilized in works such as SeLa [2], SwAV [8] and DINOv2 [45]. In particular, let $\\Psi$ be a visual encoder with a clustering head $g$ that yields a $K$ dimensional output, then the clustering loss is given by:", + "bbox": [ + 496, + 380, + 893, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(x _ {i}\\right) = - \\tilde {y} _ {i} \\log (g (\\Psi (x) _ {i})), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 542, + 890, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which is a standard cross-entropy loss with regards to self-supervised pseudo-label $\\tilde{y}_i$ . These labels, in turn, are generated by solving an entropy-regularised optimal transport problem on the batch $\\mathcal{B}$ [2]:", + "bbox": [ + 496, + 571, + 893, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\tilde {y}} \\langle \\tilde {y}, - \\log g (\\Psi (x)) \\rangle + \\frac {1}{\\lambda} \\operatorname {K L} (\\tilde {y} \\| r c ^ {\\top}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 642, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w i t h} r = \\frac {1}{K} \\cdot \\mathbb {1}, \\quad c = \\frac {1}{| \\mathcal {B} |} \\cdot \\mathbb {1}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 674, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\lambda$ controls the entropy regularisation and $r, c$ are the marginals for the prototypes and the batch, respectively. Note that solving this problem can be done extremely quickly on the GPU and yields soft pseudo-labels $\\tilde{Y}$ , such that $\\operatorname{argmax}(\\tilde{Y}) = \\tilde{y}$ .", + "bbox": [ + 496, + 715, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "n, these might yield sufficiently similar semantics [17] this is not true for dense tasks, as static features can only be assumed for videos where nothing is moving, which is rare.", + "bbox": [ + 496, + 792, + 890, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Overall TIMET loss", + "text_level": 1, + "bbox": [ + 500, + 847, + 679, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We combine the previous two modules to arrive at our full method (as shown in Fig. 2): First, self-supervised", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16539", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sinkhorn-Knopp clustering is conducted on early features $g(\\Psi(I_t))$ , yielding soft pseudo-labels $\\mathrm{SK}(g(\\Psi(I_t))) = \\tilde{Y}_t$ . These are then forwarded in time using our Feature-Forwarder to arrive at dense targets $\\mathrm{FF}(\\tilde{Y}_t)$ , which are used in the final loss. Compactly:", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T I M E T}} \\left(I _ {T}\\right) = - \\sum_ {i, j} \\operatorname {F F} \\left(\\tilde {Y} _ {t}\\right) \\log \\left(g \\left(\\Psi \\left(I _ {T}\\right)\\right)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 181, + 468, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 231, + 209, + 248 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 76, + 257, + 158, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We train our method and baselines on YTVOS [64], one of the largest video segmentation datasets available, and evaluate on DAVIS17 [48] and YTVOS. For YTVOS the ground truth masks are only available for the first frames of the test and validation sets, and therefore, a fixed random $20\\%$ of the training set is used for testing. For transfer learning experiments, we use the validation set of Pascal VOC 2012 [16], As the dataset has been commonly used as a main reference for recent works in dense self-supervised image segmentation [74, 57, 61]. For completeness, we also report the performance on egocentric datasets that have less object-centric bias and are prevalent in real-world scenarios. For egocentric experiments, we train on EPIC-KITCHENS-100 [13] and evaluate on VISOR [14]. Further details are provided in Appendix A.", + "bbox": [ + 75, + 282, + 470, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models and baselines. Currently, there is a lack of available unsupervised semantic video semantic segmentation methods. Nevertheless, we have included a comprehensive comparison of our method with state-of-the-art techniques in both image and video domains for unsupervised image semantic segmentation and unsupervised video object segmentation. To evaluate the image-based models, we utilized either official reported numbers or provided pretrained models of STEGO [21] and Leopart [74]. We have taken every measure to ensure fairness in our comparison. To do so, all the used pretrained models have the same pretraining dataset (ImageNet-1k), and the number of their backbone parameters is roughly similar. For those models that use extra datasets, for instance, Leopart, we select the pretrained backbones that closely match the specification of YTVOS training dataset. Additionally, we trained image-based models on the same video datasets, where we converted video data into their corresponding image data and reported their performance wherever possible. To ensure comprehensive analysis, we have included the recent concurrent work Flowdino [71], which utilizes optical flow to refine DINO features on unlabeled videos, in our comparison benchmarks as well. This is done only in cases where there is a shared experimental setup.", + "bbox": [ + 75, + 537, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation procedure. Despite unsupervised object segmentation being a well-established evaluation in the image domain [57, 74], evaluating unsupervised video multi-label object segmentation is challenging due to the absence of an established evaluation protocol for video object semantic segmentation without supervision. In this regard, we propose a set of evaluation protocols for unsupervised video multi-label object segmentation, which exploits existing video object segmentation datasets for evaluation purposes (see details in Appendix B). In our experiments, we discard any projection head used during training and evaluate ViT's spatial tokens directly, similar to [57, 32, 61], using four methods: classification with a linear head, classification with an FCN head, overclustering, and clustering with as many clusters as the number of ground truth objects. To fine-tune different heads on top of the frozen spatial tokens, we follow [57]. For unsupervised semantic segmentation, we apply $K$ -Means on all spatial tokens, where $K$ is chosen to be higher or equal to the ground-truth number of classes, following the common protocol in image clustering [32, 56]. For grouping clusters to ground-truth classes, we match them either by pixel-wise precision or Hungarian matching on merged cluster maps [35]. For video datasets, we allow the matching to be per-frame, per-video, and per-dataset. The evaluation metric is permutation-invariant, following [32], and the results are averaged over five different seeds. Clustering evaluations are preferred as they require less supervision and fewer hyperparameters than training a linear classifier and operate directly on the learned embedding spaces. We report our results in mean Intersection over Union (mIoU) unless otherwise specified.", + "bbox": [ + 496, + 90, + 893, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model training. We train a ViT-Small with patch size 16 and initialized from ImageNet-pretrained DINO weights [9] using the proposed self-supervised loss function. For the ablations and the main experiments, our models are trained for 12 and 30 epochs respectively. Further training details are provided in Appendix A. Code will be made available.", + "bbox": [ + 496, + 577, + 893, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Ablations", + "text_level": 1, + "bbox": [ + 500, + 676, + 609, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first examine the essential parameters of our method by training TIMET on YTVOS and assessing its ability to perform semantic segmentation on Pascal VOC. In addition to presenting the results through clustering and overclustering, we also demonstrate linear classification outcomes.", + "bbox": [ + 496, + 700, + 893, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Number of prototypes. We first ablate the influence of the number of prototypes on downstream segmentation performance. Results in Table 1a indicate a sharp increase in performance with a rise in the number of prototypes, but once a sufficiently large number is reached, performance stabilizes. We observe peak performance with a moderate number of 200 prototypes. The stability of our method over", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16540", + "bbox": [ + 480, + 924, + 519, + 936 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2c273b3c3448037f728c81648f31d6d37bd9d7f8eaac1f43ac9157f1e9a86e0e.jpg", + "table_caption": [ + "(a) Ablating # prototypes $K$" + ], + "table_footnote": [], + "table_body": "
PLCK=21K=500
1037.56.124.1
5058.28.240.5
20059.79.242.8
30059.49.042.3
", + "bbox": [ + 112, + 111, + 292, + 198 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c089f25c179b8536a01c8198f91c144c25f0c1b8e554ca862f0b50cfea02b71c.jpg", + "table_caption": [ + "(b) Ablating time interval $\\delta T$" + ], + "table_footnote": [], + "table_body": "
#framesδTLCK=21K=500
10s50.75.626.2
40.2s56.57.436.5
40.5s59.79.242.8
41.0s57.18.338.1
", + "bbox": [ + 361, + 112, + 625, + 196 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9e5d6606ba88abed857c01df1dc919420cbbfac6f8c0153d4decd4c125da6648.jpg", + "table_caption": [ + "(c) Ablating number of frames used." + ], + "table_footnote": [], + "table_body": "
#framesTLCK=21K=500
10s50.75.626.2
22.0s52.66.237.1
42.0s59.79.242.8
82.0s59.79.042.3
", + "bbox": [ + 643, + 112, + 890, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a range of prototypes (50-300) suggests that our approach is robust to the change of this parameter.", + "bbox": [ + 75, + 295, + 468, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Understanding FF. Next, we ablate the temporal dimensions that influence the working of the Feature-Forwarder. In Table 1b, we vary the time interval $\\delta T$ between frames whilst keeping a fixed number of four frames. We generally find an increase in performance when increasing the time-interval, with performance peaking at 0.5s. One of the most critical observations is the clear difference in clustering performance from $26.2\\%$ to $42.8\\%$ between training on single frames (first row) vs. training with multiple frames. This clearly indicates the significant benefit of utilizing temporal information from the Feature-Forwarder during training.", + "bbox": [ + 75, + 348, + 468, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, in Table 1c we vary the number of frames given a fixed clip duration of 2s. Again, we find that a moderate number of frames between the source and target time is most favorable. While additional frames do not degrade the performance much, they do add computational complexity, so we prefer using 4 frames in our main method.", + "bbox": [ + 75, + 515, + 468, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Choice of propagation feature. Finally, in Table 2, we vary the type of feature that we forward to a future time. In the first row, we report the performance when not using any feature forwarding – thus solely training on single-frame inputs (i.e., the same as row 1 in Table 1c). In the second row, \"Identity\" shows the performance that is obtained when the feature map from the source frame is simply forwarded to the future without any changes, which shows an increase in the performance compared to the first row. This shows that often, training videos are not very dynamic, and a static assumption can already lead to some gains. However, compared to forwarding the Sinkhorn-Knopp (SK) clustered features, these gains are small (+4.6% vs. +17%). Importantly, our Feature-Forwarder relies on forwarding SK-sharpened feature maps and does not work when simply forwarding the network's output logits $\\Phi(x)$ , followed by a clustering step. The reason for obtaining considerably lower numbers here can be traced back to using an un-entropy", + "bbox": [ + 75, + 628, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/31ae8e1ac86b230dfabf37b28584290f01d65f60b1085c048a0ff06f165cd623.jpg", + "table_caption": [ + "Table 1: Ablations of the key parameters of our method. The model is trained for 12 epochs on Pascal VOC, and results for unsupervised segmentation with clustering $(K = 21)$ , overclustering $(K = 500)$ , and linear pixel-wise classification (LC) are shown. The stability of our method over a range of prototypes (50-300), inter-frame time intervals $(\\delta T \\in [0.5s - 1.0s])$ , and the number of training frames (4-8) at a fixed clip duration $(T)$ shows the robustness of the method." + ], + "table_footnote": [], + "table_body": "
FF typeLCK=21K=500
None50.75.626.2
Identity55.37.436.1
Ψ(x)53.16.635.5
SK(Ψ(x))59.79.242.8
", + "bbox": [ + 563, + 292, + 831, + 378 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/22f5c6a4515553f90f64b74380e5e3ee582b3be313b7f5544b899dac2ebb3297.jpg", + "table_caption": [ + "Table 2: Propagating different features in FF on Pascal VOC. We find that our Sinkhorn-Knopp (SK) based module outperforms static training ('None' or 'Identity'). Moreover, propagating logits $\\Psi(x)$ does not work as well as SK-regularised features." + ], + "table_footnote": [], + "table_body": "
PretrainBackboneAt Init+TIMET
K=500LCK=500LC
MSN [3]ViT-S/1626.055.448.3↑22.367.2↑11.8
iBOT [72]ViT-S/1632.162.147.1↑15.067.1↑5.0
DINO [9]ViT-S/822.555.853.9↑31.469.5↑13.7
DINO [9]ViT-B/1628.959.152.7↑23.870.2↑11.1
", + "bbox": [ + 506, + 487, + 888, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3: Applying TIMET to different pretrainings on Pascal VOC. TIMET can boost $(\\uparrow)$ the performance of different backbones with different initialization by a considerable margin, showing the generality of our approach.", + "bbox": [ + 496, + 597, + 890, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "regularized clustering algorithm. In this case, the logits tend to cause a few prototypes to dominate the cluster centers, exacerbated after the propagation, resulting in highly noisy and uninformative propagated logits. This shows that careful design of the Feature-Forwarder is indeed required.", + "bbox": [ + 496, + 693, + 890, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Applying TIMET to different models. As shown in Table 3, our method is generalisable to different backbones and self-supervised learning initializations, enabling it to be an effective method to transfer the knowledge of the videos to images in an unsupervised way, which reduces the cost of labeling for different downstream tasks.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16541", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2bf554d9064aafcabf706009fc36cda2f1000c3051dd8771930d0a7cc3b5ea4.jpg", + "image_caption": [ + "Pascal" + ], + "image_footnote": [], + "bbox": [ + 80, + 98, + 171, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87d163b4a5d1599b1025b5b3114a31d5585e00ce5c7897190285587edc754ef0.jpg", + "image_caption": [ + "Figure 3: TIMET unsupervised segmentations. As we regularize DINO's backbone to be consistent across time on YTVOS, it obtains strong performance on both image and video segmentation datasets, yielding high class consistencies (indicated by the segmentation colors) and tight borders. We provide more qualitative results in Appendix D." + ], + "image_footnote": [], + "bbox": [ + 173, + 98, + 264, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a25781e63c584f4090988165ab0be9a3e4d80f26c37e766cdf2efdfcb782b9aa.jpg", + "image_caption": [ + "DAVIS" + ], + "image_footnote": [], + "bbox": [ + 285, + 98, + 377, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a9fb08e3310fa4ecc3051380b0d2343b339d9bce864880c6e3d1b075cb98613.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 98, + 470, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1fdbc1e8b8ec9f21d7cc2d4f2c49bd9bf71a4563ab27d404be004d8f82587bc1.jpg", + "image_caption": [ + "YTVOS" + ], + "image_footnote": [], + "bbox": [ + 496, + 98, + 588, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/384f2e894c87e57b41bbd3aced56eb59e5711d58807fa4efdcf74d44e6cb137e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 98, + 681, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b30519f2e15c0745bb711cec09f0c67d48d038dfb6f34eeebb7112227728083c.jpg", + "image_caption": [ + "VISOR" + ], + "image_footnote": [], + "bbox": [ + 707, + 98, + 800, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d172b6c77e28e038a2a8e85d0f7a0e96b0476e5fd2e22447bfd052647fd3e10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 98, + 890, + 170 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b583fd95eb9a09d8657786af735083df84c1fe0001089e5fe96a56ce8aecea9d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ClusteringOverclustering
YTVOSDAVISYTVOSDAVIS
FCDFCDFCDFCD
Trained on Images
Resnet5044.043.41.739.337.44.255.652.83.146.644.28.4
SwAV [8]39.538.23.232.029.67.359.858.15.850.550.125.7
DINO [9]39.137.91.930.231.01.666.265.44.056.954.917.9
Leopart [74]39.237.911.730.330.216.564.562.815.554.954.426.7
Trained on Videos
STEGO*41.540.32.031.931.03.258.154.35.147.646.310.4
DINO*37.236.11.229.329.22.453.150.91.345.444.08.6
Leopart*41.540.57.737.536.512.660.859.86.853.753.116.8
TIMET(ours)52.551.313.353.753.020.568.666.815.859.861.531.8
", + "bbox": [ + 94, + 267, + 877, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4: Unsupervised video semantic segmentation results in mIoU for clustering (K=GT) and overclustering. TIMET not only gets better numbers on YTVOS, but also achieves considerably better results on DAVIS. This shows the better quality of the learned features at transferability to other video datasets. The clip-length is set to 16 and 4 for DAVIS and YTVOS, respectively. For clustering, the Hungarian algorithm [35] matches the unsupervised segmentation clusters (K) with the ground truth (GT) per frame $(F)$ , clip $(C)$ or across the whole dataset $(D)$ . For overclustering, we use K=10 for the frames-wise $(F)$ and clip-wise $(C)$ evaluations, and for dataset-wise $(D)$ evaluation, K=200 for DAVIS and K=500 for YTVOS. *denotes finetuning pretrained models on the same video frames as input as our method. The matching protocol is greedy many-to-one, see Appendix B for details.", + "bbox": [ + 75, + 500, + 893, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Large-scale experiments", + "text_level": 1, + "bbox": [ + 76, + 646, + 302, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unsupervised video semantic segmentation. In this section, we train our TIMET method on the YTVOS dataset and evaluate it for unsupervised video object segmentation on both DAVIS and YTVOS. The results are shown in Table 4. It should be noted that the number of prototypes we analyzed in Table 1a is not the number of clusters used to report accuracies for Table 4, as the actual number is usually unknown in the case of unsupervised learning. For the clustering experiment, we set K to the number of ground truth objects, which makes evaluation metrics easier to interpret and is commonly done in image clustering and segmentation. This evaluation is repeated but for larger numbers of clusters, in the \"overclustering\" scenario. Note that in dense self-supervised learning, overclustering can be particularly important because the learned representations are", + "bbox": [ + 75, + 674, + 472, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "used as a feature extractor for downstream tasks such as semantic segmentation or object detection. By using more fine-grained representations, the network may be able to extract more discriminative features such as object parts [74], which can lead to better performances.", + "bbox": [ + 496, + 647, + 890, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Table 4, we observe a clear trend: our method, trained on YTVOS not only achieves superior performances on YTVOS, but also beats existing image-trained models on DAVIS by a large margin. In particular, the state-of-the-art self-supervised clustering method, Leopart, has $4\\%$ lower performance on per dataset DAVIS evaluation for $\\mathrm{K = GT}$ and $1.6\\%$ lower performance on YTVOS. The gap even becomes larger when Leopart is trained on the same video dataset with $8\\%$ to $16\\%$ lower numbers in per dataset and per clip numbers across different datasets. This means that when grouping objects of the same class over time and", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16542", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the whole dataset, the image-based self-supervised methods have trouble generalising to videos, where objects are not centered in the frame, can appear in varying poses, and are more difficult in general. A strong contender is Leopart, which matches the performance of DINO for per frame and per clip clustering, however, improves considerably with the per dataset clustering. We attribute this to the fact that their dense learning objective improves mainly the generalisation capabilities of the learned representation, however falling short of generalising to temporal variations. We make similar observations for overclustering, improving the DINO baseline which we utilize as initialisation by $12\\%$ to $14\\%$ across different datasets in the per dataset metric.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conclude that the proposed method outperforms all other methods in learning robust and discriminative representations for dense video understanding, and image-based self-supervised learning may not have sufficient generalisation capacity. This shows that if used right, time is a particularly strong inductive bias for self-supervised learning. Figure 3 shows segmentations returned by the proposed model trained on YTVOS and tested on YTVOS, DAVIS, and VISOR respectively. The proposed method groups objects accurately, and importantly, the frame segmentations are considerably consistent over time (see more examples in Appendix D). This highlights the importance and relevance of temporal fine-tuning, not just for higher accuracy, but crucially for consistency and robustness.", + "bbox": [ + 75, + 287, + 472, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transfer from video training to images. Despite the common belief that features learned from videos perform worse when transferred to images [34, 20], the results in Table 5 demonstrate our method achieves high performance that match state-of-the-art methods directly trained on images, specifically for the $\\mathrm{K} = 500$ , FCN, and LC metrics. We also compare our performance gains in videos against models designed for unsupervised image segmentation. The results show that models highly biased towards image datasets cannot provide the same performance when trained on videos, lagging behind our approach by $7\\%$ to $30\\%$ . Our results demonstrate that achieving high transfer performance on challenging tasks through video-based self-supervised learning is not only feasible but can also maintain high performance across modalities. These findings suggest that our method can drive further advances in self-supervised learning and inspire new directions for research in this field. Figure 3 shows the qualitative results on Pascal VOC. For more visualisations we refer to Appendix D.", + "bbox": [ + 75, + 518, + 472, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Salient object segmentation. In Table 6, we compare foreground masks obtained with various DINO ViT based methods. We use the cluster-based foreground extraction protocol from [74] (details provided in Appendix C). First, we find that our method outperforms the original DINO at", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/03d0bcb406c2fe9194d8f827b8d4e5ee182440535ee9cdcfad6d70141308dc30.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pascal VOC
K=21K=500LCFCN
Trained on Images
ResNet-504.536.553.8-
DINO [9]5.517.450.660.6
SwAV [8]11.635.750.7-
MaskContrast [57]35.045.449.2-
DenseCL [61]-43.649.069.4
STEGO [21]7.019.559.163.5
CrOC [52]20.6-61.6-
Leopard [74]36.650.568.070.1
Trained on Videos
STEGO*4.015.551.155.5
Leopard*14.921.253.263.2
Flowdino†[70]--59.4-
TIMET (ours)34.553.268.070.6
", + "bbox": [ + 511, + 88, + 883, + 353 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/edb4ade60672b3830cce1206c636d9117fcb1a5d10dc8e5c4bba2ecd9752717c.jpg", + "table_caption": [ + "Table 5: Transfer from video training to images. Numbers taken from [74, 52, 70]. *: finetuning pretrained models on the same video frames as input as our method. †: uses a 40% larger superset of our train data. Details on the different datasets and methods are provided in the Appendix." + ], + "table_footnote": [], + "table_body": "
Pascal VOCDAVISYTVOS
DINO [9]52.134.532.1
Leopard [74]59.637.338.6
STEGO [21]49.130.432.1
TIMET (ours)63.944.543.5
", + "bbox": [ + 524, + 460, + 867, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6: Salient object segmentation. We report performance using the Jaccard score [9] and use the official pretrained models for evaluation.", + "bbox": [ + 496, + 561, + 890, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tention maps consistently by $10 - 11\\%$ . We also surpass the results of Leopart [74] and STEGO [21], works that rely on the same pretrained backbone. Even for the evaluation on Pascal VOC, our method achieves higher performances despite the domain shift of having trained on videos.", + "bbox": [ + 496, + 643, + 893, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalisation to egocentric datasets. We train our method on EPIC-Kitchens-100 [13] and evaluate on VISOR [14]. We use the official code to convert VISOR to a DAVIS-like structure, in which we can report our numbers for per frame and per clip evaluation. Note that, after conversion, the object IDs do not maintain global consistency in the whole dataset; therefore, we cannot report the per dataset number. As Table 7 shows, TIMET outperforms image-based competitors on egocentric datasets as well.", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16543", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b5f3f9c0c93dbe1fb5fffc6c78060c98b50f5d65cef2185b35d846a7d617b163.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VISOR
FC
DINO [9]24.818.7
Leopard [74]24.118.5
TIMET (ours)26.521.5
", + "bbox": [ + 173, + 88, + 374, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Visual In-Context Learning evaluation. Here, we contrast our approach with a recently introduced benchmark, which assesses the in-context reasoning capabilities of models within the vision domain. Unlike linear or FCN classification methods, visual in-context learning evaluation obviates the need for fine-tuning or end-to-end training. Instead, it constructs validation segmentation maps using patch-level, feature-wise nearest neighbor similarities between validation images (referred to as \"queries\") and training samples (termed \"keys\"). This approach mirrors strategies in the NLP domain, aiming to evaluate the proficiency of models in learning tasks from limited examples presented as demonstrations. The results are shown by Table 8. Given that the vast majority of extant models in the domain utilize ViT-S16 as their backbone, we conducted a re-evaluation of their checkpoints to furnish a consistent and directly comparable evaluation table. Subsequently, we re-trained our model employing ViT-B16 and benchmarked it against other baselines as presented by [5]. The results, as depicted in the table, indicate that even though our model was exclusively trained on videos, it registers performance metrics in line with Leopart [74]. Furthermore, it surpasses the results of the leading method, CrOC [52], which was trained on images. Contrary to Hummingbird [5], TIMET is not custom-fitted to the specific evaluation setup and boasts superior computational efficiency. Notably, it requires only a single GPU for training, in contrast to the 16 TPUs demanded by Hummingbird.", + "bbox": [ + 76, + 297, + 472, + 720 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 738, + 194, + 753 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper has aimed to learn dense representations that are learned from videos; yet can be generalised to the image domain as well. As video content is growing rapidly and contains more information compared to images, learning generalisable knowledge from them facilitates further scaling of self-supervised learning methods. To this effect, we have proposed a self-supervised clustering loss to encourage temporally consistent features between different frames of a video. To efficiently find corresponding views between", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1810de54920a10b0c11a61eebf044bef2181a6d40ff5c492ce1980e2eb4a35ba.jpg", + "table_caption": [ + "Table 7: Generalisation to egocentric datasets. Unsupervised semantic segmentation results in mIoU for clustering $(\\mathrm{K} = \\mathrm{GT})$ on VISOR [14] after training onn EPIC-Kitchens-100 [13]. The clip-length is set to 4. Our method gets better results on egocentric datasets as well." + ], + "table_footnote": [], + "table_body": "
EncoderParamsmIoU
Trained on Images
SupervisedViT-S1621M35.1
MoCo-v3* [11]ViT-S1621M19.5
DINO* [9]ViT-S1621M47.9
CrOC* [52]ViT-S1621M50.0
Leopard* [74]ViT-S1621M63.6
DINO [9]ViT-B1686M55.9
MoCo-v3 [11]ViT-B1686M37.2
MAE [22]ViT-B1686M6.6
LOCA [7]ViT-B1686M57.5
Hummingbird [5]ViT-B1686M70.5
Trained on Videos
TIMET (ours)ViT-S1621M61.6
TIMET (ours)ViT-B1686M65.5
", + "bbox": [ + 524, + 88, + 870, + 345 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 8: Visual In-Context Learning evaluation. Numbers are taken from [5]. *: numbers are produced by this paper. This approach is equivalent to a non-parametric nearest-neighbor based evaluation on Pascal VOC. The details of the evaluation benchmark can be found in the provided implementation codes.", + "bbox": [ + 496, + 354, + 890, + 445 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "clip frames, we have proposed to recycle pretrained transformer features to leverage their natural tracking ability at each clip. Our empirical results indicate that this method achieves significant gains on the challenging task of video object segmentation across three different evaluation protocols and three datasets. Moreover, by transferring the learned model to the image domain, we have demonstrated the generalisability of the learned features by surpassing or matching the state-of-the-art for unsupervised image segmentation.", + "bbox": [ + 496, + 474, + 892, + 625 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 642, + 679, + 659 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is financially supported by Qualcomm Technologies Inc., the University of Amsterdam and the allowance Top consortia for Knowledge and Innovation (TKIs) from the Netherlands Ministry of Economic Affairs and Climate Policy.", + "bbox": [ + 496, + 667, + 890, + 744 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 761, + 596, + 776 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nikita Araslanov, Simone Schaub-Meyer, and Stefan Roth. Dense unsupervised learning for video segmentation. Advances in Neural Information Processing Systems, 34:25308-25319, 2021. 2, 3, 15", + "[2] Yuki M. Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In International Conference on Learning Representations (ICLR), 2020. 4" + ], + "bbox": [ + 506, + 786, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16544", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI, pages 456-473. Springer, 2022. 6", + "[4] Arthur Aubret, Markus R. Ernst, Céline Teulière, and Jochen Triesch. Time to augment self-supervised visual representation learning. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[5] Ivana Balăžević, David Steiner, Nikhil Parthasarathy, Relja Arandjelović, and Olivier J Henaff. Towards in-context scene understanding. arXiv preprint arXiv:2306.01667, 2023. 9", + "[6] Thomas Brox and Jitendra Malik. Large displacement optical flow: Descriptor matching in variational motion estimation. IEEE TPAMI, 33(3):500-513, 2011. 4", + "[7] Mathilde Caron, Neil Houlsby, and Cordelia Schmid. Location-aware self-supervised transformers. arXiv preprint arXiv:2212.02400, 2022. 9", + "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 4, 7, 8", + "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 3, 4, 5, 6, 7, 8, 9, 15", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[11] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. ICCV, 2021. 9", + "[12] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural Information Processing Systems, volume 26, 2013. 4", + "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 5, 8, 9", + "[14] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2022. 5, 8, 9", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl-" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 13", + "[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 5, 13, 14", + "[17] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3299-3309, 2021. 2, 4", + "[18] Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould. Self-supervised video representation learning with odd-one-out networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3636-3645, 2017. 1, 2", + "[19] Raghudeep Gadde, Varun Jampani, and Peter V Gehler. Semantic video cnns through representation warping. In Proceedings of the IEEE International Conference on Computer Vision, pages 4453-4462, 2017. 1", + "[20] Daniel Gordon, Kiana Ehsani, Dieter Fox, and Ali Farhadi. Watching the world go by: Representation learning from unlabeled videos. arXiv preprint arXiv:2003.07990, 2020. 2, 8", + "[21] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 5, 8", + "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 9", + "[23] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2", + "[24] Olivier J Henaff, Skanda Koppula, Jean-Baptiste Alayrac, Aaron Van den Oord, Oriol Vinyals, and João Carreira. Efficient visual pretraining with contrastive detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10086-10096, 2021. 1, 2", + "[25] Olivier J Hénaff, Skanda Koppula, Evan Shelhamer, Daniel Zoran, Andrew Jaegle, Andrew Zisserman, João Carreira, and Relja Arandjelović. Object discovery and representation networks. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 123-143. Springer, 2022. 2", + "[26] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 13", + "[27] Jefferson Hernandez, Ruben Villegas, and Vicente Ordonez. Visual representation learning from unlabeled video using contrastive masked autoencoders. arXiv preprint arXiv:2303.12001, 2023. 2", + "[28] Ping Hu, Fabian Caba, Oliver Wang, Zhe Lin, Stan Sclaroff, and Federico Perazzi. Temporally distributed networks" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16545", + "bbox": [ + 480, + 925, + 519, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for fast video semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8827, 2020. 1", + "[29] Junhwa Hur and Stefan Roth. Joint optical flow and temporally consistent semantic segmentation. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part I 14, pages 163-177. Springer, 2016. 1", + "[30] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 1, 3, 4", + "[31] Samvit Jain, Xin Wang, and Joseph E Gonzalez. Accel: A corrective fusion network for efficient semantic segmentation on video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8866-8875, 2019. 1", + "[32] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9865-9874, 2019. 5", + "[33] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 13", + "[34] Thomas Kipf, Gamaleldin F Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. arXiv preprint arXiv:2111.12594, 2021. 2, 8", + "[35] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5, 7", + "[36] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[37] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Video k-net: A simple, strong, and unified baseline for video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18847-18857, 2022. 1", + "[38] Yule Li, Jianping Shi, and Dahua Lin. Low-latency video semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5997-6005, 2018. 1", + "[39] Runtao Liu, Zhirong Wu, Stella Yu, and Stephen Lin. The emergence of objectness: Learning zero-shot segmentation from videos. Advances in Neural Information Processing Systems, 34:13137-13152, 2021. 2", + "[40] Si Liu, Changhu Wang, Ruihe Qian, Han Yu, Renda Bao, and Yao Sun. Surveillance video parsing with single frame supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 413-421, 2017. 1", + "[41] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "inference. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 352-368. Springer, 2020. 1", + "[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 13", + "[43] Xiankai Lu, Wenguan Wang, Jianbing Shen, Yu-Wing Tai, David J Crandall, and Steven CH Hoi. Learning video object segmentation from unlabeled videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8960-8970, 2020. 2", + "[44] Yuval Nirkin, Lior Wolf, and Tal Hassner. Hyperseg: Patchwise hypernetwork for real-time semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4061-4070, 2021. 1", + "[45] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4", + "[46] Nikhil Parthasarathy, SM Eslami, João Carreira, and Olivier J Henaff. Self-supervised video pretraining yields strong image representations. arXiv preprint arXiv:2210.06433, 2022. 2", + "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 13", + "[48] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5, 13", + "[49] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael S Ryoo. Self-supervised video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2874-2884, 2022. 1", + "[50] Sucheng Ren, Wenxi Liu, Yongtuo Liu, Haoxin Chen, Guoqiang Han, and Shengfeng He. Reciprocal transformations for unsupervised video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15455-15464, 2021. 2", + "[51] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860, 2022. 1", + "[52] Thomas Stegmüller, Tim Lebailly, Behzad Bozorgtabar, Tinne Tuytelaars, and Jean-Philippe Thiran. Croc: Cross-view online clustering for dense visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7000–7009, 2023. 8, 9", + "[53] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE/CVF conference on" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16546", + "bbox": [ + 480, + 925, + 519, + 936 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "computer vision and pattern recognition, pages 8934-8943, 2018. 15", + "[54] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3126-3137, 2022. 1", + "[55] Fida Mohammad Thoker, Hazel Doughty, Piyush Bagad, and Cees GM Snoek. How severe is benchmark-sensitivity in video self-supervised learning? In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 632-652. Springer, 2022. 2", + "[56] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X, pages 268-285. Springer, 2020. 5", + "[57] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10052-10062, 2021. 1, 2, 5, 8, 13", + "[58] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Chuanxin Tang, Xiyang Dai, Yucheng Zhao, Yujia Xie, Lu Yuan, and Yu-Gang Jiang. Look before you match: Instance understanding matters in video object segmentation. arXiv preprint arXiv:2212.06826, 2022. 1", + "[59] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. arXiv preprint arXiv:2301.11320, 2023. 2", + "[60] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2566-2576, 2019. 3", + "[61] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 2, 5, 8, 13", + "[62] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 2", + "[63] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10334–10343, 2019. 1", + "[64] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 5, 13", + "[65] Yu-Syuan Xu, Tsu-Jui Fu, Hsuan-Kung Yang, and Chun-Yi Lee. Dynamic video segmentation network. In Proceed-" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE conference on computer vision and pattern recognition, pages 6556-6565, 2018. 1", + "[66] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7177-7188, 2021. 2, 3, 15", + "[67] Sukmin Yun, Hankook Lee, Jaehyung Kim, and Jinwoo Shin. Patch-level representation learning for self-supervised vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8354-8363, 2022. 2", + "[68] Andrii Zadaianchuk, Matthaeus Kleindessner, Yi Zhu, Francesco Locatello, and Thomas Brox. Unsupervised semantic segmentation with self-supervised object-centric representations. arXiv preprint arXiv:2207.05027, 2022. 2", + "[69] Shaofeng Zhang, Feng Zhu, Rui Zhao, and Junchi Yan. Patch-level contrasting without patch correspondence for accurate and dense contrastive representation learning. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[70] Xinyu Zhang and Abdeslam Boullarias. Optical flow boosts unsupervised localization and segmentation. arXiv preprint arXiv:2307.13640, 2023. 8", + "[71] Yurong Zhang, Liulei Li, Wenguan Wang, Rong Xie, Li Song, and Wenjun Zhang. Boosting video object segmentation via space-time correspondence learning. arXiv preprint arXiv:2304.06211, 2023. 1, 5", + "[72] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. ICLR, 2022. 3, 6", + "[73] Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, and Yichen Wei. Deep feature flow for video recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2349-2358, 2017. 1", + "[74] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022. 1, 2, 5, 7, 8, 9, 13, 14, 15, 16, 18" + ], + "bbox": [ + 501, + 92, + 893, + 656 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "16547", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/1b69a328-51e9-4f52-9cd0-ceb8e6de4113_model.json b/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/1b69a328-51e9-4f52-9cd0-ceb8e6de4113_model.json new file mode 100644 index 0000000000000000000000000000000000000000..87dcee8a7b6033dce80d2b9c692104ff82129fbb --- /dev/null +++ b/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/1b69a328-51e9-4f52-9cd0-ceb8e6de4113_model.json @@ -0,0 +1,2545 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.88, + 0.155 + ], + "angle": 0, + "content": "Time Does Tell: Self-Supervised Time-Tuning of Dense Image Representations" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.181, + 0.794, + 0.217 + ], + "angle": 0, + "content": "Mohammadreza Salehi, Efstratios Gavves, Cees G. M. Snoek, Yuki M. Asano \nQUVA Lab, University of Amsterdam" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.22, + 0.744, + 0.234 + ], + "angle": 0, + "content": "(s-salehidehnavi, e.gavves, c.g.m.snoek, y.m.asano)@uva.nl" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.635 + ], + "angle": 0, + "content": "Spatially dense self-supervised learning is a rapidly growing problem domain with promising applications for unsupervised segmentation and pretraining for dense downstream tasks. Despite the abundance of temporal data in the form of videos, this information-rich source has been largely overlooked. Our paper aims to address this gap by proposing a novel approach that incorporates temporal consistency in dense self-supervised learning. While methods designed solely for images face difficulties in achieving even the same performance on videos, our method improves not only the representation quality for videos – but also images. Our approach, which we call time-tuning, starts from image-pretrained models and fine-tunes them with a novel self-supervised temporal-alignment clustering loss on unlabeled videos. This effectively facilitates the transfer of high-level information from videos to image representations. Time-tuning improves the state-of-the-art by \\(8 - 10\\%\\) for unsupervised semantic segmentation on videos and matches it for images. We believe this method paves the way for further self-supervised scaling by leveraging the abundant availability of videos. The implementation can be found here: https://github.com/SMSD75/Timetuning" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.664, + 0.21, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.87 + ], + "angle": 0, + "content": "Dense self-supervised learning, whereby meaningful deep features for each pixel or patch of input are learned in an unsupervised manner, has recently received increasing attention [74, 24, 57, 51]. By learning spatially consistent features for different views of an input, strong gains in unsupervised semantic segmentation have been achieved using unlabeled images. However, so far, an even more information-rich source for unsupervised training has been largely overlooked: videos. With their additional time dimension and being the most rapidly growing form of digital content, they are well-suited to scaling dense self-supervised learning even further." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Some efforts have already been made to learn from the video domain by using different frames from a video as aug-" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.272, + 0.885, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.46, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Figure 1: Time-tuning compared to previous methods. Unlike existing methods that ignore or utilize expensive 3D models to implicitly model temporal information, the proposed method explicitly incorporates temporal consistency in dense feature representations using a temporal self-supervised loss. The method starts with a 2D encoder pretrained on images and fine-tunes it using unlabeled videos. This approach leads to improved performance not only for videos but also for images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.895, + 0.824 + ], + "angle": 0, + "content": "mentations [18, 63, 49] or by involving temporal correspondence [71, 58]; however, they mostly did it in a supervised way [29, 73, 19, 40, 65, 38, 31, 28, 41, 37, 54, 44], which is not scalable specifically for dense tasks where the number of targets can increase significantly as the number of pixels grows. To this end, self-supervised learning approaches offer a solution by reducing the need for supervision. However, these methods typically rely on the notion of \"views\", which involves learning similar features for corresponding locations over time. This usually leads to a chicken-and-egg problem, where the correspondences are required for learning dense features - which in turn enable good correspondences [30]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In images, the challenge is trivially solved by considering the correspondence function based on the augmentation function. For instance, In the case of color augmentations, this correspondence is simply given by the identity. However, shifts in time cannot be viewed as mere augmenta" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "16536" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "tions. As we demonstrate through a new evaluation protocol, using image models on frames alone is not nearly as effective. A similar finding has also been reported, albeit for non-dense works [20, 34], which assumed the passage of time as an image augmentation. These works have generally reported reduced performances, even when compared to simple image-based pretraining methods. Similarly, video-level tasks [17, 18, 55] assume sufficiently similar semantics between different frames. This is also not true for dense tasks, as static features can only be assumed where nothing is moving – which is rare due to possible object, camera, and background motion between the frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.515 + ], + "angle": 0, + "content": "To address this challenge, we propose to model the additional time-dimension explicitly to identify which pixels should retain similar embeddings and which should not. We propose two separate modules to tackle the correspondence and the dense learning, respectively. For the former, we introduce the Feature-Forwarder (FF) module, which breaks the mentioned chicken-and-egg loop by leveraging the good tracking performance of pretrained image models, and allows an approximate second \"view\" that can then be treated as a target for the further dense self-supervised loss. On top of this, we introduce a spatio-temporally dense clustering module, which learns unsupervised clusters across samples, locations and time. Using these two components and starting from image-pretrained features, our proposed method allows time-tuning (TIMET) the dense representation in a self-supervised manner, see Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.473, + 0.636 + ], + "angle": 0, + "content": "Finally, we demonstrate that TIMET paves the way for further scaling of self-supervised learning by leveraging the abundant availability of video datasets and transferring their knowledge to the image domain. This results in consistently achieving state-of-the-art performances not only for the task of unsupervised semantic segmentation of videos, but also for unsupervised image semantic segmentation, a feat previously out of reach for methods trained on videos." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.637, + 0.455, + 0.654 + ], + "angle": 0, + "content": "Overall, this paper makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.665, + 0.47, + 0.727 + ], + "angle": 0, + "content": "- We show that image-based unsupervised dense segmentation models applied to videos exhibit degraded performance and lack temporal consistency in their segmentation maps." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.738, + 0.47, + 0.784 + ], + "angle": 0, + "content": "- Building on this observation, we propose a novel dense self-supervised learning method that utilizes temporal consistency as a learning signal." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.796, + 0.471, + 0.888 + ], + "angle": 0, + "content": "- We demonstrate that our method enables the scaling of self-supervised learning by leveraging abundant video datasets and effectively transferring knowledge to the image domain. Our approach consistently achieves state-of-the-art performance for both images and videos, opening up new opportunities in the field." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.665, + 0.471, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.651, + 0.107 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.895, + 0.525 + ], + "angle": 0, + "content": "Dense self-supervised learning. These methods build upon image-level self-supervised representation learning by incorporating existing losses to enhance the spatial features, demonstrating a commendable advancement. DenseCL [61] works on spatial features by constructing dense correspondences across views using the contrastive objective given in MoCo [23], while PixPro [62] utilizes the augmentation wrapper to get the spatial correspondence of the pixel intersection between two views. Similarly, MaskContrast [57], Leopart [74], DetCon [24] and Odin [25] also ensure spatial feature similarities via contrastive learning and spatial correspondences. Self-Patch [67] treats the spatial neighbors of the patch as positive examples for learning more semantically meaningful relations among patches. Inspired by SelfPatch, ADCLR [69] proposes patch-level contrasting via query crop and cross-attention mechanism. Pursuing the same objective through none end-to-end approaches. Both [68] and [59] propose an unsupervised salient object segmentation pipeline that extracts noisy object masks from the inputs and fine-tunes a specifically designed object segmentation head with several self-training steps to make an unsupervised semantic segmentation model. Unlike the existing image-based approaches, we propose a method that improves the dense prediction performance of a pretrained encoder by explicitly modeling the temporal dimension and learning from diverse natural dynamics and variations found in videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.895, + 0.789 + ], + "angle": 0, + "content": "Video to image knowledge transfer. Learning from videos instead of images has recently received attention since they contain far more information than still images and hold the potential for learning rich representations of the visual world. To this end, several non-dense works have been released [46, 4, 27]. VITO [46] shows that naively applying image domain self-supervised learning methods on videos can lead to a performance drop coming from the distribution shift, which by applying data processing techniques is relieved but not fully solved. Following the same way, [27] trains masked autoencoders [22] with contrastive learning [10] on video datasets and shows a decent performance on both image and video tasks. While video to image knowledge transfer is the end goal of such methods and our paper; our dense task is different and is as of yet unaddressed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Unsupervised video object segmentation. Unsupervised video object segmentation methods do not require any manual annotations. However, they are only designed to tackle a foreground/background segmentation task, which refers to the segmentation of the most prominent, general objects in video sequences [66, 43, 50, 1, 39]. For instance, [66]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.52, + 0.938 + ], + "angle": 0, + "content": "16537" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.088, + 0.468, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.089, + 0.872, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.27, + 0.895, + 0.364 + ], + "angle": 0, + "content": "Figure 2: Time-Tuning overview. We conduct self-supervised video semantic segmentation by tuning image-pretrained models with temporal information from videos. (a): Our general pipeline of adapting an image-pretrained ViT transformer on video data using our dense clustering loss. The encoder is specified by \\(\\Psi\\) and kept frozen except for the last two layers. Also, \\(z_{t}\\) shows the output of Sinkhorn-Knopp algorithm, which is forwarded to time-step \\(T\\) to be compared with the features that are obtained using the last time-step. (b): Detailed view into our Feature-Forwarder module that is used for aligning cluster-maps from past frames to logit-features of future ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.47, + 0.523 + ], + "angle": 0, + "content": "uses an AutoEncoder-based architecture based on slot attention so that pixels that show the same motion flows based on optical-flow are grouped together. In this way, dominant moving objects can be detected and separated from the background. In our experiments, we benchmark two representative instances of this line of work, i.e., [66, 1], to provide better insights into the differences of unsupervised object detection methods and our proposed benchmark for video semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.6 + ], + "angle": 0, + "content": "To conclude, we propose the first self-supervised video semantic segmentation method, which tunes image-trained models such as [9, 72] based on the temporal information that exists in unlabeled videos to enhance their effectiveness in dense prediction tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.614, + 0.17, + 0.63 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.471, + 0.79 + ], + "angle": 0, + "content": "Our method works by densely clustering features in a manner that is consistent with time. At a high level, it works by transforming features from a past time \\( t' = t \\) to the current frame \\( t' = T \\) and forcing these to be consistent with the currently observed ones at \\( t' = T \\). The overall learning signal comes from the dense temporal clustering of forwarded features and allows \"timetuning\" image-pretrained models to learn from videos. Figure 2 shows the proposed method's architecture. In the following, we describe each component in detail." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.27, + 0.818 + ], + "angle": 0, + "content": "3.1. Feature Forwarding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The goal of this component, the feature-forwarder (FF), is to propagate the dense features in time as accurately as possible, given some RGB input frames. More formally, let \\( f \\) be the feature forwarding function, \\( I_{t} \\) be a frame at time \\( t \\) for a given video, and \\( z_{t} \\) the features corresponding to \\( I_{t} \\)," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.388, + 0.831, + 0.404 + ], + "angle": 0, + "content": "then the task is to propagate \\( z_{t} \\) to a future time \\( T \\)," + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.415, + 0.892, + 0.433 + ], + "angle": 0, + "content": "\\[\nf \\left(\\left[ I _ {t}, \\dots , I _ {T} \\right], z _ {t}\\right)\\rightarrow \\hat {z} _ {T}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.894, + 0.55 + ], + "angle": 0, + "content": "Note that \\(\\hat{z}_T\\) and \\(z_{T}\\) are not necessarily the same for \\(T > t\\); since \\(\\hat{z}_T\\) represents an approximation of \\(z_{T}\\) using the previous features. Obviously, the further \\(T\\) is from \\(t\\), the more variance is seen and therefore a better learning signal might be provided; however, for such larger intervals, the idea of feature forwarding encounters two main challenges: error accumulation and object non-permanence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.894, + 0.66 + ], + "angle": 0, + "content": "Challenge 1: Error accumulation. While computing how features have changed simply based on the start and end frames is the most straightforward (i.e., \\( I_{t} \\) and \\( I_{T} \\)), it does not fully leverage the knowledge contained in the intermediate frames. Instead, it is common [30, 60, 36] to forward features for every time-step \\( \\delta t \\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.67, + 0.892, + 0.69 + ], + "angle": 0, + "content": "\\[\nf ([ I _ {t},.., I _ {T} ], z _ {t}) = \\bigcirc_ {t ^ {\\prime} = t + \\delta t} ^ {T} f ([ I _ {t ^ {\\prime} - \\delta t}, I _ {t ^ {\\prime}} ], z _ {t} ]), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.894, + 0.745 + ], + "angle": 0, + "content": "where \\( \\bigcirc \\) is a composition operator such as element-wise multiplication. However, this results in the accumulation of errors over time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Challenge 2: Object non-permanence. One difficulty that arises when using videos as training data is the fact that objects sometimes simply disappear from the screen. Compared to images, which carry a heavy photographer's bias, videos are more prone to natural variation due to camera and object motion resulting in temporary object non-permanence in a particular video clip. A method that simply assumes an object (or a feature) has to be present in any given frame simply because it was present in the previous" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "16538" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "one might therefore easily fail when going towards videos that exist in-the-wild. A case in point are simple occlusions that arise from an object being fully- or partially-covered by another one further in the background. Taking longer training clips has been attempted to resolve the issue. However, this increases the training complexity, especially for dense-tasks that necessitate a higher number of predictions for each input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.225, + 0.473, + 0.451 + ], + "angle": 0, + "content": "We address these challenges using a novel, yet simple component, which we call the Stabilizing Feature-Forwarder. The first choice in designing this module is the function and data used for composing information across time. For instance, a simple approach would be to use pixel-wise optical flow and aggregate the flow at the feature level to predict how the features change over time. However, as we also show in our experiments, optical flow is prone to accumulating error with time and also cannot easily recover from occlusions [6]. Instead, we utilize the fact that we wish to pretrain the visual encoder and recycle part of it for the purpose of feature forwarding. This not only speeds up the process, as no additional encoder/modality is required, but also produces a positive feedback loop of better forwarding and better feature learning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.452, + 0.472, + 0.497 + ], + "angle": 0, + "content": "Concretely, we first compute the L2-normalised dense features of a pretrained visual encoder \\(\\Phi\\) to compute spatial similarities across time (the ' in \\(t^\\prime\\) left out for clarity):" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.503, + 0.47, + 0.521 + ], + "angle": 0, + "content": "\\[\nF _ {i j} ^ {t} = \\left\\langle \\Phi \\left(I _ {t}\\right) _ {i}, \\Phi \\left(I _ {T}\\right) _ {j} \\right\\rangle / \\tau , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.527, + 0.47, + 0.634 + ], + "angle": 0, + "content": "which yields matrices \\( F^{\\{t,\\dots T - \\delta t\\}} \\in R^{N\\times N} \\) with values between [0,1] indicating semantic similarities between the \\( N \\) spatial features that are sharpened by a temperature \\( \\tau \\). Note that because we compare each frame with the final frame, the effect of object non-permanence can be minimized: even if the object is only present in every other frame, there is enough signal to forward its features." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.635, + 0.47, + 0.665 + ], + "angle": 0, + "content": "Next to propagate from time \\( t \\) to \\( T \\), these similarities are stacked and normalised along time, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.683, + 0.47, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\tilde {F} _ {i j} ^ {t} = \\exp \\left(\\mathcal {N} \\left(F _ {i j} ^ {t}\\right)\\right) / \\sum_ {i, t ^ {\\prime}} \\exp \\left(\\mathcal {N} \\left(F _ {i j} ^ {t ^ {\\prime}}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.722, + 0.47, + 0.798 + ], + "angle": 0, + "content": "where \\(t \\leq t' \\leq T\\) and \\(\\mathcal{N}\\) is a neighborhood thresholding method which forces short-term spatial-smoothness, i.e., \\(\\mathcal{N}(\\tilde{F}_{ij}^t) = 0\\) if \\(i\\) is not within a local window around \\(j\\) with the size \\(k\\). Finally, the propagated feature are computed \\(\\hat{z}_T\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.817, + 0.47, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {T} (j) = \\sum_ {t ^ {\\prime}, i} \\tilde {F} _ {i j} ^ {t ^ {\\prime}} \\hat {z} _ {t ^ {\\prime}} (i), \\quad j \\in \\{1,.., N \\}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "This means that for arriving at the target feature \\(\\hat{z}_T\\), we not only use the previous frame as the source, but instead use the past \\((T - t) / \\delta t\\) frames and aggregate these." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.35 + ], + "angle": 0, + "content": "Relation to mask-propagation methods. While previous methods such as DINO [9] and STC [30] have utilized a similar technique for propagating ground-truth masks for evaluating, for example on the DAVIS dataset, there are three key differences to our forwarding method. First, instead of propagating binary maps of foreground-vs-background, our feature-forwarder has as inputs soft, noisy and multi-label self-supervised segmentation maps, which require the forwarder to tolerate overlapping of different object probabilities throughout training. Second, previous methods have used this approach mainly for inference; we, however, use this module as a trainable component and show how our loss improves this forwarder with training time. Finally, we do not follow a typical re-normalizing step across feature dimensions (typically done after Eq. (5)) as this harms the scale of logits that are being propagated and leads to heavily diluted target distributions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.358, + 0.783, + 0.374 + ], + "angle": 0, + "content": "3.2. Self-supervised dense clustering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.895, + 0.534 + ], + "angle": 0, + "content": "While the Feature-Forwarder produces target features that include information about the dynamics with time, its computation utilizes \\(\\Phi(I_T)\\) and could lead to trivial solutions, i.e., \\(f([I_t,..,I_T],z_t) = \\Phi(I_T)\\). To counteract this, we propose a self-supervised clustering task across views in time. For this, we utilize the basic online clustering algorithm based optimal-transport [12], utilized in works such as SeLa [2], SwAV [8] and DINOv2 [45]. In particular, let \\(\\Psi\\) be a visual encoder with a clustering head \\(g\\) that yields a \\(K\\) dimensional output, then the clustering loss is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.544, + 0.892, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(x _ {i}\\right) = - \\tilde {y} _ {i} \\log (g (\\Psi (x) _ {i})), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.894, + 0.634 + ], + "angle": 0, + "content": "which is a standard cross-entropy loss with regards to self-supervised pseudo-label \\(\\tilde{y}_i\\). These labels, in turn, are generated by solving an entropy-regularised optimal transport problem on the batch \\(\\mathcal{B}\\) [2]:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.643, + 0.892, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\tilde {y}} \\langle \\tilde {y}, - \\log g (\\Psi (x)) \\rangle + \\frac {1}{\\lambda} \\operatorname {K L} (\\tilde {y} \\| r c ^ {\\top}), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.675, + 0.892, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\text {w i t h} r = \\frac {1}{K} \\cdot \\mathbb {1}, \\quad c = \\frac {1}{| \\mathcal {B} |} \\cdot \\mathbb {1}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Here \\(\\lambda\\) controls the entropy regularisation and \\(r, c\\) are the marginals for the prototypes and the batch, respectively. Note that solving this problem can be done extremely quickly on the GPU and yields soft pseudo-labels \\(\\tilde{Y}\\), such that \\(\\operatorname{argmax}(\\tilde{Y}) = \\tilde{y}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.793, + 0.892, + 0.839 + ], + "angle": 0, + "content": "n, these might yield sufficiently similar semantics [17] this is not true for dense tasks, as static features can only be assumed for videos where nothing is moving, which is rare." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.681, + 0.862 + ], + "angle": 0, + "content": "3.3. Overall TIMET loss" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "We combine the previous two modules to arrive at our full method (as shown in Fig. 2): First, self-supervised" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "16539" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "Sinkhorn-Knopp clustering is conducted on early features \\( g(\\Psi(I_t)) \\), yielding soft pseudo-labels \\( \\mathrm{SK}(g(\\Psi(I_t))) = \\tilde{Y}_t \\). These are then forwarded in time using our Feature-Forwarder to arrive at dense targets \\( \\mathrm{FF}(\\tilde{Y}_t) \\), which are used in the final loss. Compactly:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.182, + 0.47, + 0.215 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T I M E T}} \\left(I _ {T}\\right) = - \\sum_ {i, j} \\operatorname {F F} \\left(\\tilde {Y} _ {t}\\right) \\log \\left(g \\left(\\Psi \\left(I _ {T}\\right)\\right)\\right). \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.232, + 0.21, + 0.249 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.258, + 0.16, + 0.275 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.471, + 0.511 + ], + "angle": 0, + "content": "Datasets. We train our method and baselines on YTVOS [64], one of the largest video segmentation datasets available, and evaluate on DAVIS17 [48] and YTVOS. For YTVOS the ground truth masks are only available for the first frames of the test and validation sets, and therefore, a fixed random \\(20\\%\\) of the training set is used for testing. For transfer learning experiments, we use the validation set of Pascal VOC 2012 [16], As the dataset has been commonly used as a main reference for recent works in dense self-supervised image segmentation [74, 57, 61]. For completeness, we also report the performance on egocentric datasets that have less object-centric bias and are prevalent in real-world scenarios. For egocentric experiments, we train on EPIC-KITCHENS-100 [13] and evaluate on VISOR [14]. Further details are provided in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Models and baselines. Currently, there is a lack of available unsupervised semantic video semantic segmentation methods. Nevertheless, we have included a comprehensive comparison of our method with state-of-the-art techniques in both image and video domains for unsupervised image semantic segmentation and unsupervised video object segmentation. To evaluate the image-based models, we utilized either official reported numbers or provided pretrained models of STEGO [21] and Leopart [74]. We have taken every measure to ensure fairness in our comparison. To do so, all the used pretrained models have the same pretraining dataset (ImageNet-1k), and the number of their backbone parameters is roughly similar. For those models that use extra datasets, for instance, Leopart, we select the pretrained backbones that closely match the specification of YTVOS training dataset. Additionally, we trained image-based models on the same video datasets, where we converted video data into their corresponding image data and reported their performance wherever possible. To ensure comprehensive analysis, we have included the recent concurrent work Flowdino [71], which utilizes optical flow to refine DINO features on unlabeled videos, in our comparison benchmarks as well. This is done only in cases where there is a shared experimental setup." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.562 + ], + "angle": 0, + "content": "Evaluation procedure. Despite unsupervised object segmentation being a well-established evaluation in the image domain [57, 74], evaluating unsupervised video multi-label object segmentation is challenging due to the absence of an established evaluation protocol for video object semantic segmentation without supervision. In this regard, we propose a set of evaluation protocols for unsupervised video multi-label object segmentation, which exploits existing video object segmentation datasets for evaluation purposes (see details in Appendix B). In our experiments, we discard any projection head used during training and evaluate ViT's spatial tokens directly, similar to [57, 32, 61], using four methods: classification with a linear head, classification with an FCN head, overclustering, and clustering with as many clusters as the number of ground truth objects. To fine-tune different heads on top of the frozen spatial tokens, we follow [57]. For unsupervised semantic segmentation, we apply \\( K \\)-Means on all spatial tokens, where \\( K \\) is chosen to be higher or equal to the ground-truth number of classes, following the common protocol in image clustering [32, 56]. For grouping clusters to ground-truth classes, we match them either by pixel-wise precision or Hungarian matching on merged cluster maps [35]. For video datasets, we allow the matching to be per-frame, per-video, and per-dataset. The evaluation metric is permutation-invariant, following [32], and the results are averaged over five different seeds. Clustering evaluations are preferred as they require less supervision and fewer hyperparameters than training a linear classifier and operate directly on the learned embedding spaces. We report our results in mean Intersection over Union (mIoU) unless otherwise specified." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.894, + 0.671 + ], + "angle": 0, + "content": "Model training. We train a ViT-Small with patch size 16 and initialized from ImageNet-pretrained DINO weights [9] using the proposed self-supervised loss function. For the ablations and the main experiments, our models are trained for 12 and 30 epochs respectively. Further training details are provided in Appendix A. Code will be made available." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.611, + 0.693 + ], + "angle": 0, + "content": "4.2. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.894, + 0.778 + ], + "angle": 0, + "content": "We first examine the essential parameters of our method by training TIMET on YTVOS and assessing its ability to perform semantic segmentation on Pascal VOC. In addition to presenting the results through clustering and overclustering, we also demonstrate linear classification outcomes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Number of prototypes. We first ablate the influence of the number of prototypes on downstream segmentation performance. Results in Table 1a indicate a sharp increase in performance with a rise in the number of prototypes, but once a sufficiently large number is reached, performance stabilizes. We observe peak performance with a moderate number of 200 prototypes. The stability of our method over" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "16540" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.097, + 0.289, + 0.112 + ], + "angle": 0, + "content": "(a) Ablating # prototypes \\(K\\)" + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.112, + 0.293, + 0.199 + ], + "angle": 0, + "content": "
PLCK=21K=500
1037.56.124.1
5058.28.240.5
20059.79.242.8
30059.49.042.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.396, + 0.098, + 0.574, + 0.112 + ], + "angle": 0, + "content": "(b) Ablating time interval \\(\\delta T\\)" + }, + { + "type": "table", + "bbox": [ + 0.362, + 0.113, + 0.626, + 0.198 + ], + "angle": 0, + "content": "
#framesδTLCK=21K=500
10s50.75.626.2
40.2s56.57.436.5
40.5s59.79.242.8
41.0s57.18.338.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.661, + 0.098, + 0.876, + 0.112 + ], + "angle": 0, + "content": "(c) Ablating number of frames used." + }, + { + "type": "table", + "bbox": [ + 0.645, + 0.113, + 0.892, + 0.198 + ], + "angle": 0, + "content": "
#framesTLCK=21K=500
10s50.75.626.2
22.0s52.66.237.1
42.0s59.79.242.8
82.0s59.79.042.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.208, + 0.895, + 0.271 + ], + "angle": 0, + "content": "Table 1: Ablations of the key parameters of our method. The model is trained for 12 epochs on Pascal VOC, and results for unsupervised segmentation with clustering \\((K = 21)\\), overclustering \\((K = 500)\\), and linear pixel-wise classification (LC) are shown. The stability of our method over a range of prototypes (50-300), inter-frame time intervals \\((\\delta T \\in [0.5s - 1.0s])\\), and the number of training frames (4-8) at a fixed clip duration \\((T)\\) shows the robustness of the method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.47, + 0.327 + ], + "angle": 0, + "content": "a range of prototypes (50-300) suggests that our approach is robust to the change of this parameter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.47, + 0.516 + ], + "angle": 0, + "content": "Understanding FF. Next, we ablate the temporal dimensions that influence the working of the Feature-Forwarder. In Table 1b, we vary the time interval \\(\\delta T\\) between frames whilst keeping a fixed number of four frames. We generally find an increase in performance when increasing the time-interval, with performance peaking at 0.5s. One of the most critical observations is the clear difference in clustering performance from \\(26.2\\%\\) to \\(42.8\\%\\) between training on single frames (first row) vs. training with multiple frames. This clearly indicates the significant benefit of utilizing temporal information from the Feature-Forwarder during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.47, + 0.607 + ], + "angle": 0, + "content": "Next, in Table 1c we vary the number of frames given a fixed clip duration of 2s. Again, we find that a moderate number of frames between the source and target time is most favorable. While additional frames do not degrade the performance much, they do add computational complexity, so we prefer using 4 frames in our main method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Choice of propagation feature. Finally, in Table 2, we vary the type of feature that we forward to a future time. In the first row, we report the performance when not using any feature forwarding – thus solely training on single-frame inputs (i.e., the same as row 1 in Table 1c). In the second row, \"Identity\" shows the performance that is obtained when the feature map from the source frame is simply forwarded to the future without any changes, which shows an increase in the performance compared to the first row. This shows that often, training videos are not very dynamic, and a static assumption can already lead to some gains. However, compared to forwarding the Sinkhorn-Knopp (SK) clustered features, these gains are small (+4.6% vs. +17%). Importantly, our Feature-Forwarder relies on forwarding SK-sharpened feature maps and does not work when simply forwarding the network's output logits \\(\\Phi(x)\\), followed by a clustering step. The reason for obtaining considerably lower numbers here can be traced back to using an un-entropy" + }, + { + "type": "table", + "bbox": [ + 0.564, + 0.293, + 0.832, + 0.38 + ], + "angle": 0, + "content": "
FF typeLCK=21K=500
None50.75.626.2
Identity55.37.436.1
Ψ(x)53.16.635.5
SK(Ψ(x))59.79.242.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.466 + ], + "angle": 0, + "content": "Table 2: Propagating different features in FF on Pascal VOC. We find that our Sinkhorn-Knopp (SK) based module outperforms static training ('None' or 'Identity'). Moreover, propagating logits \\(\\Psi(x)\\) does not work as well as SK-regularised features." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.488, + 0.89, + 0.589 + ], + "angle": 0, + "content": "
PretrainBackboneAt Init+TIMET
K=500LCK=500LC
MSN [3]ViT-S/1626.055.448.3↑22.367.2↑11.8
iBOT [72]ViT-S/1632.162.147.1↑15.067.1↑5.0
DINO [9]ViT-S/822.555.853.9↑31.469.5↑13.7
DINO [9]ViT-B/1628.959.152.7↑23.870.2↑11.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.66 + ], + "angle": 0, + "content": "Table 3: Applying TIMET to different pretrainings on Pascal VOC. TIMET can boost \\((\\uparrow)\\) the performance of different backbones with different initialization by a considerable margin, showing the generality of our approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.892, + 0.77 + ], + "angle": 0, + "content": "regularized clustering algorithm. In this case, the logits tend to cause a few prototypes to dominate the cluster centers, exacerbated after the propagation, resulting in highly noisy and uninformative propagated logits. This shows that careful design of the Feature-Forwarder is indeed required." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Applying TIMET to different models. As shown in Table 3, our method is generalisable to different backbones and self-supervised learning initializations, enabling it to be an effective method to transfer the knowledge of the videos to images in an unsupervised way, which reduces the cost of labeling for different downstream tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "16541" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.099, + 0.172, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.174, + 0.205, + 0.187 + ], + "angle": 0, + "content": "Pascal" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.099, + 0.266, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.099, + 0.379, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.099, + 0.472, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.174, + 0.407, + 0.187 + ], + "angle": 0, + "content": "DAVIS" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.099, + 0.589, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.099, + 0.682, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.174, + 0.613, + 0.187 + ], + "angle": 0, + "content": "YTVOS" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.099, + 0.802, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.099, + 0.892, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.174, + 0.828, + 0.187 + ], + "angle": 0, + "content": "VISOR" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.209, + 0.894, + 0.256 + ], + "angle": 0, + "content": "Figure 3: TIMET unsupervised segmentations. As we regularize DINO's backbone to be consistent across time on YTVOS, it obtains strong performance on both image and video segmentation datasets, yielding high class consistencies (indicated by the segmentation colors) and tight borders. We provide more qualitative results in Appendix D." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.268, + 0.879, + 0.49 + ], + "angle": 0, + "content": "
ClusteringOverclustering
YTVOSDAVISYTVOSDAVIS
FCDFCDFCDFCD
Trained on Images
Resnet5044.043.41.739.337.44.255.652.83.146.644.28.4
SwAV [8]39.538.23.232.029.67.359.858.15.850.550.125.7
DINO [9]39.137.91.930.231.01.666.265.44.056.954.917.9
Leopart [74]39.237.911.730.330.216.564.562.815.554.954.426.7
Trained on Videos
STEGO*41.540.32.031.931.03.258.154.35.147.646.310.4
DINO*37.236.11.229.329.22.453.150.91.345.444.08.6
Leopart*41.540.57.737.536.512.660.859.86.853.753.116.8
TIMET(ours)52.551.313.353.753.020.568.666.815.859.861.531.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.5, + 0.895, + 0.621 + ], + "angle": 0, + "content": "Table 4: Unsupervised video semantic segmentation results in mIoU for clustering (K=GT) and overclustering. TIMET not only gets better numbers on YTVOS, but also achieves considerably better results on DAVIS. This shows the better quality of the learned features at transferability to other video datasets. The clip-length is set to 16 and 4 for DAVIS and YTVOS, respectively. For clustering, the Hungarian algorithm [35] matches the unsupervised segmentation clusters (K) with the ground truth (GT) per frame \\((F)\\), clip \\((C)\\) or across the whole dataset \\((D)\\). For overclustering, we use K=10 for the frames-wise \\((F)\\) and clip-wise \\((C)\\) evaluations, and for dataset-wise \\((D)\\) evaluation, K=200 for DAVIS and K=500 for YTVOS. *denotes finetuning pretrained models on the same video frames as input as our method. The matching protocol is greedy many-to-one, see Appendix B for details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.647, + 0.303, + 0.664 + ], + "angle": 0, + "content": "4.3. Large-scale experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Unsupervised video semantic segmentation. In this section, we train our TIMET method on the YTVOS dataset and evaluate it for unsupervised video object segmentation on both DAVIS and YTVOS. The results are shown in Table 4. It should be noted that the number of prototypes we analyzed in Table 1a is not the number of clusters used to report accuracies for Table 4, as the actual number is usually unknown in the case of unsupervised learning. For the clustering experiment, we set K to the number of ground truth objects, which makes evaluation metrics easier to interpret and is commonly done in image clustering and segmentation. This evaluation is repeated but for larger numbers of clusters, in the \"overclustering\" scenario. Note that in dense self-supervised learning, overclustering can be particularly important because the learned representations are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.648, + 0.892, + 0.724 + ], + "angle": 0, + "content": "used as a feature extractor for downstream tasks such as semantic segmentation or object detection. By using more fine-grained representations, the network may be able to extract more discriminative features such as object parts [74], which can lead to better performances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "From Table 4, we observe a clear trend: our method, trained on YTVOS not only achieves superior performances on YTVOS, but also beats existing image-trained models on DAVIS by a large margin. In particular, the state-of-the-art self-supervised clustering method, Leopart, has \\(4\\%\\) lower performance on per dataset DAVIS evaluation for \\(\\mathrm{K = GT}\\) and \\(1.6\\%\\) lower performance on YTVOS. The gap even becomes larger when Leopart is trained on the same video dataset with \\(8\\%\\) to \\(16\\%\\) lower numbers in per dataset and per clip numbers across different datasets. This means that when grouping objects of the same class over time and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "16542" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "the whole dataset, the image-based self-supervised methods have trouble generalising to videos, where objects are not centered in the frame, can appear in varying poses, and are more difficult in general. A strong contender is Leopart, which matches the performance of DINO for per frame and per clip clustering, however, improves considerably with the per dataset clustering. We attribute this to the fact that their dense learning objective improves mainly the generalisation capabilities of the learned representation, however falling short of generalising to temporal variations. We make similar observations for overclustering, improving the DINO baseline which we utilize as initialisation by \\(12\\%\\) to \\(14\\%\\) across different datasets in the per dataset metric." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.473, + 0.501 + ], + "angle": 0, + "content": "We conclude that the proposed method outperforms all other methods in learning robust and discriminative representations for dense video understanding, and image-based self-supervised learning may not have sufficient generalisation capacity. This shows that if used right, time is a particularly strong inductive bias for self-supervised learning. Figure 3 shows segmentations returned by the proposed model trained on YTVOS and tested on YTVOS, DAVIS, and VISOR respectively. The proposed method groups objects accurately, and importantly, the frame segmentations are considerably consistent over time (see more examples in Appendix D). This highlights the importance and relevance of temporal fine-tuning, not just for higher accuracy, but crucially for consistency and robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.473, + 0.807 + ], + "angle": 0, + "content": "Transfer from video training to images. Despite the common belief that features learned from videos perform worse when transferred to images [34, 20], the results in Table 5 demonstrate our method achieves high performance that match state-of-the-art methods directly trained on images, specifically for the \\(\\mathrm{K} = 500\\), FCN, and LC metrics. We also compare our performance gains in videos against models designed for unsupervised image segmentation. The results show that models highly biased towards image datasets cannot provide the same performance when trained on videos, lagging behind our approach by \\(7\\%\\) to \\(30\\%\\). Our results demonstrate that achieving high transfer performance on challenging tasks through video-based self-supervised learning is not only feasible but can also maintain high performance across modalities. These findings suggest that our method can drive further advances in self-supervised learning and inspire new directions for research in this field. Figure 3 shows the qualitative results on Pascal VOC. For more visualisations we refer to Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Salient object segmentation. In Table 6, we compare foreground masks obtained with various DINO ViT based methods. We use the cluster-based foreground extraction protocol from [74] (details provided in Appendix C). First, we find that our method outperforms the original DINO at" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.089, + 0.885, + 0.354 + ], + "angle": 0, + "content": "
Pascal VOC
K=21K=500LCFCN
Trained on Images
ResNet-504.536.553.8-
DINO [9]5.517.450.660.6
SwAV [8]11.635.750.7-
MaskContrast [57]35.045.449.2-
DenseCL [61]-43.649.069.4
STEGO [21]7.019.559.163.5
CrOC [52]20.6-61.6-
Leopard [74]36.650.568.070.1
Trained on Videos
STEGO*4.015.551.155.5
Leopard*14.921.253.263.2
Flowdino†[70]--59.4-
TIMET (ours)34.553.268.070.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.362, + 0.895, + 0.44 + ], + "angle": 0, + "content": "Table 5: Transfer from video training to images. Numbers taken from [74, 52, 70]. *: finetuning pretrained models on the same video frames as input as our method. †: uses a 40% larger superset of our train data. Details on the different datasets and methods are provided in the Appendix." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.461, + 0.869, + 0.553 + ], + "angle": 0, + "content": "
Pascal VOCDAVISYTVOS
DINO [9]52.134.532.1
Leopard [74]59.637.338.6
STEGO [21]49.130.432.1
TIMET (ours)63.944.543.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.608 + ], + "angle": 0, + "content": "Table 6: Salient object segmentation. We report performance using the Jaccard score [9] and use the official pretrained models for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.721 + ], + "angle": 0, + "content": "tention maps consistently by \\(10 - 11\\%\\). We also surpass the results of Leopart [74] and STEGO [21], works that rely on the same pretrained backbone. Even for the evaluation on Pascal VOC, our method achieves higher performances despite the domain shift of having trained on videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Generalisation to egocentric datasets. We train our method on EPIC-Kitchens-100 [13] and evaluate on VISOR [14]. We use the official code to convert VISOR to a DAVIS-like structure, in which we can report our numbers for per frame and per clip evaluation. Note that, after conversion, the object IDs do not maintain global consistency in the whole dataset; therefore, we cannot report the per dataset number. As Table 7 shows, TIMET outperforms image-based competitors on egocentric datasets as well." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.521, + 0.938 + ], + "angle": 0, + "content": "16543" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.174, + 0.089, + 0.375, + 0.182 + ], + "angle": 0, + "content": "
VISOR
FC
DINO [9]24.818.7
Leopard [74]24.118.5
TIMET (ours)26.521.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.192, + 0.472, + 0.268 + ], + "angle": 0, + "content": "Table 7: Generalisation to egocentric datasets. Unsupervised semantic segmentation results in mIoU for clustering \\((\\mathrm{K} = \\mathrm{GT})\\) on VISOR [14] after training onn EPIC-Kitchens-100 [13]. The clip-length is set to 4. Our method gets better results on egocentric datasets as well." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.298, + 0.473, + 0.721 + ], + "angle": 0, + "content": "Visual In-Context Learning evaluation. Here, we contrast our approach with a recently introduced benchmark, which assesses the in-context reasoning capabilities of models within the vision domain. Unlike linear or FCN classification methods, visual in-context learning evaluation obviates the need for fine-tuning or end-to-end training. Instead, it constructs validation segmentation maps using patch-level, feature-wise nearest neighbor similarities between validation images (referred to as \"queries\") and training samples (termed \"keys\"). This approach mirrors strategies in the NLP domain, aiming to evaluate the proficiency of models in learning tasks from limited examples presented as demonstrations. The results are shown by Table 8. Given that the vast majority of extant models in the domain utilize ViT-S16 as their backbone, we conducted a re-evaluation of their checkpoints to furnish a consistent and directly comparable evaluation table. Subsequently, we re-trained our model employing ViT-B16 and benchmarked it against other baselines as presented by [5]. The results, as depicted in the table, indicate that even though our model was exclusively trained on videos, it registers performance metrics in line with Leopart [74]. Furthermore, it surpasses the results of the leading method, CrOC [52], which was trained on images. Contrary to Hummingbird [5], TIMET is not custom-fitted to the specific evaluation setup and boasts superior computational efficiency. Notably, it requires only a single GPU for training, in contrast to the 16 TPUs demanded by Hummingbird." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.739, + 0.196, + 0.754 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "This paper has aimed to learn dense representations that are learned from videos; yet can be generalised to the image domain as well. As video content is growing rapidly and contains more information compared to images, learning generalisable knowledge from them facilitates further scaling of self-supervised learning methods. To this effect, we have proposed a self-supervised clustering loss to encourage temporally consistent features between different frames of a video. To efficiently find corresponding views between" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.089, + 0.872, + 0.346 + ], + "angle": 0, + "content": "
EncoderParamsmIoU
Trained on Images
SupervisedViT-S1621M35.1
MoCo-v3* [11]ViT-S1621M19.5
DINO* [9]ViT-S1621M47.9
CrOC* [52]ViT-S1621M50.0
Leopard* [74]ViT-S1621M63.6
DINO [9]ViT-B1686M55.9
MoCo-v3 [11]ViT-B1686M37.2
MAE [22]ViT-B1686M6.6
LOCA [7]ViT-B1686M57.5
Hummingbird [5]ViT-B1686M70.5
Trained on Videos
TIMET (ours)ViT-S1621M61.6
TIMET (ours)ViT-B1686M65.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Table 8: Visual In-Context Learning evaluation. Numbers are taken from [5]. *: numbers are produced by this paper. This approach is equivalent to a non-parametric nearest-neighbor based evaluation on Pascal VOC. The details of the evaluation benchmark can be found in the provided implementation codes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.893, + 0.626 + ], + "angle": 0, + "content": "clip frames, we have proposed to recycle pretrained transformer features to leverage their natural tracking ability at each clip. Our empirical results indicate that this method achieves significant gains on the challenging task of video object segmentation across three different evaluation protocols and three datasets. Moreover, by transferring the learned model to the image domain, we have demonstrated the generalisability of the learned features by surpassing or matching the state-of-the-art for unsupervised image segmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.643, + 0.681, + 0.66 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.892, + 0.745 + ], + "angle": 0, + "content": "This work is financially supported by Qualcomm Technologies Inc., the University of Amsterdam and the allowance Top consortia for Knowledge and Innovation (TKIs) from the Netherlands Ministry of Economic Affairs and Climate Policy." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.762, + 0.597, + 0.777 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.787, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[1] Nikita Araslanov, Simone Schaub-Meyer, and Stefan Roth. Dense unsupervised learning for video segmentation. Advances in Neural Information Processing Systems, 34:25308-25319, 2021. 2, 3, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.845, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[2] Yuki M. Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In International Conference on Learning Representations (ICLR), 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.787, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "16544" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI, pages 456-473. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.19, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[4] Arthur Aubret, Markus R. Ernst, Céline Teulière, and Jochen Triesch. Time to augment self-supervised visual representation learning. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.249, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[5] Ivana Balăžević, David Steiner, Nikhil Parthasarathy, Relja Arandjelović, and Olivier J Henaff. Towards in-context scene understanding. arXiv preprint arXiv:2306.01667, 2023. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.292, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[6] Thomas Brox and Jitendra Malik. Large displacement optical flow: Descriptor matching in variational motion estimation. IEEE TPAMI, 33(3):500-513, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Neil Houlsby, and Cordelia Schmid. Location-aware self-supervised transformers. arXiv preprint arXiv:2212.02400, 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.377, + 0.471, + 0.446 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.449, + 0.471, + 0.53 + ], + "angle": 0, + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 3, 4, 5, 6, 7, 8, 9, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[11] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. ICCV, 2021. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.47, + 0.673 + ], + "angle": 0, + "content": "[12] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural Information Processing Systems, volume 26, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 5, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[14] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2022. 5, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "vain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 5, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.894, + 0.259 + ], + "angle": 0, + "content": "[17] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3299-3309, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[18] Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould. Self-supervised video representation learning with odd-one-out networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3636-3645, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[19] Raghudeep Gadde, Varun Jampani, and Peter V Gehler. Semantic video cnns through representation warping. In Proceedings of the IEEE International Conference on Computer Vision, pages 4453-4462, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[20] Daniel Gordon, Kiana Ehsani, Dieter Fox, and Ali Farhadi. Watching the world go by: Representation learning from unlabeled videos. arXiv preprint arXiv:2003.07990, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[21] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[23] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[24] Olivier J Henaff, Skanda Koppula, Jean-Baptiste Alayrac, Aaron Van den Oord, Oriol Vinyals, and João Carreira. Efficient visual pretraining with contrastive detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10086-10096, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[25] Olivier J Hénaff, Skanda Koppula, Evan Shelhamer, Daniel Zoran, Andrew Jaegle, Andrew Zisserman, João Carreira, and Relja Arandjelović. Object discovery and representation networks. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 123-143. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[26] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[27] Jefferson Hernandez, Ruben Villegas, and Vicente Ordonez. Visual representation learning from unlabeled video using contrastive masked autoencoders. arXiv preprint arXiv:2303.12001, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Ping Hu, Fabian Caba, Oliver Wang, Zhe Lin, Stan Sclaroff, and Federico Perazzi. Temporally distributed networks" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.52, + 0.937 + ], + "angle": 0, + "content": "16545" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.471, + 0.134 + ], + "angle": 0, + "content": "for fast video semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8827, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[29] Junhwa Hur and Stefan Roth. Joint optical flow and temporally consistent semantic segmentation. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part I 14, pages 163-177. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.26 + ], + "angle": 0, + "content": "[30] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[31] Samvit Jain, Xin Wang, and Joseph E Gonzalez. Accel: A corrective fusion network for efficient semantic segmentation on video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8866-8875, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[32] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9865-9874, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[33] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[34] Thomas Kipf, Gamaleldin F Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. arXiv preprint arXiv:2111.12594, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[35] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[36] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[37] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Video k-net: A simple, strong, and unified baseline for video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18847-18857, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[38] Yule Li, Jianping Shi, and Dahua Lin. Low-latency video semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5997-6005, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[39] Runtao Liu, Zhirong Wu, Stella Yu, and Stephen Lin. The emergence of objectness: Learning zero-shot segmentation from videos. Advances in Neural Information Processing Systems, 34:13137-13152, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[40] Si Liu, Changhu Wang, Ruihe Qian, Han Yu, Renda Bao, and Yao Sun. Surveillance video parsing with single frame supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 413-421, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "inference. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 352-368. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[43] Xiankai Lu, Wenguan Wang, Jianbing Shen, Yu-Wing Tai, David J Crandall, and Steven CH Hoi. Learning video object segmentation from unlabeled videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8960-8970, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[44] Yuval Nirkin, Lior Wolf, and Tal Hassner. Hyperseg: Patchwise hypernetwork for real-time semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4061-4070, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.294, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[45] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[46] Nikhil Parthasarathy, SM Eslami, João Carreira, and Olivier J Henaff. Self-supervised video pretraining yields strong image representations. arXiv preprint arXiv:2210.06433, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[48] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[49] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael S Ryoo. Self-supervised video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2874-2884, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[50] Sucheng Ren, Wenxi Liu, Yongtuo Liu, Haoxin Chen, Guoqiang Han, and Shengfeng He. Reciprocal transformations for unsupervised video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15455-15464, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[51] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[52] Thomas Stegmüller, Tim Lebailly, Behzad Bozorgtabar, Tinne Tuytelaars, and Jean-Philippe Thiran. Croc: Cross-view online clustering for dense visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7000–7009, 2023. 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[53] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE/CVF conference on" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.52, + 0.937 + ], + "angle": 0, + "content": "16546" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.468, + 0.119 + ], + "angle": 0, + "content": "computer vision and pattern recognition, pages 8934-8943, 2018. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[54] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3126-3137, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.471, + 0.274 + ], + "angle": 0, + "content": "[55] Fida Mohammad Thoker, Hazel Doughty, Piyush Bagad, and Cees GM Snoek. How severe is benchmark-sensitivity in video self-supervised learning? In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 632-652. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[56] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X, pages 268-285. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.471, + 0.427 + ], + "angle": 0, + "content": "[57] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10052-10062, 2021. 1, 2, 5, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.471, + 0.496 + ], + "angle": 0, + "content": "[58] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Chuanxin Tang, Xiyang Dai, Yucheng Zhao, Yujia Xie, Lu Yuan, and Yu-Gang Jiang. Look before you match: Instance understanding matters in video object segmentation. arXiv preprint arXiv:2212.06826, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[59] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. arXiv preprint arXiv:2301.11320, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[60] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2566-2576, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[61] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 2, 5, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.746 + ], + "angle": 0, + "content": "[62] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.817 + ], + "angle": 0, + "content": "[63] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10334–10343, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[64] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[65] Yu-Syuan Xu, Tsu-Jui Fu, Hsuan-Kung Yang, and Chun-Yi Lee. Dynamic video segmentation network. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 6556-6565, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[66] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7177-7188, 2021. 2, 3, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[67] Sukmin Yun, Hankook Lee, Jaehyung Kim, and Jinwoo Shin. Patch-level representation learning for self-supervised vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8354-8363, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[68] Andrii Zadaianchuk, Matthaeus Kleindessner, Yi Zhu, Francesco Locatello, and Thomas Brox. Unsupervised semantic segmentation with self-supervised object-centric representations. arXiv preprint arXiv:2207.05027, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[69] Shaofeng Zhang, Feng Zhu, Rui Zhao, and Junchi Yan. Patch-level contrasting without patch correspondence for accurate and dense contrastive representation learning. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[70] Xinyu Zhang and Abdeslam Boullarias. Optical flow boosts unsupervised localization and segmentation. arXiv preprint arXiv:2307.13640, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.432, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[71] Yurong Zhang, Liulei Li, Wenguan Wang, Rong Xie, Li Song, and Wenjun Zhang. Boosting video object segmentation via space-time correspondence learning. arXiv preprint arXiv:2304.06211, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[72] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. ICLR, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[73] Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, and Yichen Wei. Deep feature flow for video recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2349-2358, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.588, + 0.894, + 0.657 + ], + "angle": 0, + "content": "[74] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022. 1, 2, 5, 7, 8, 9, 13, 14, 15, 16, 18" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "16547" + } + ] +] \ No newline at end of file diff --git a/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/1b69a328-51e9-4f52-9cd0-ceb8e6de4113_origin.pdf b/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/1b69a328-51e9-4f52-9cd0-ceb8e6de4113_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b8ca9d6036d15e39ea2990d5fc69881ab6a2f2c8 --- /dev/null +++ b/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/1b69a328-51e9-4f52-9cd0-ceb8e6de4113_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d27f9bd7d9a0ba9fb41665d9f8f35f728b672b0a70b23e17099b30cf15d0e2d +size 732657 diff --git a/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/full.md b/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a89bea5fb71d869298bf392689f559d80a909280 --- /dev/null +++ b/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/full.md @@ -0,0 +1,351 @@ +# Time Does Tell: Self-Supervised Time-Tuning of Dense Image Representations + +Mohammadreza Salehi, Efstratios Gavves, Cees G. M. Snoek, Yuki M. Asano +QUVA Lab, University of Amsterdam + +(s-salehidehnavi, e.gavves, c.g.m.snoek, y.m.asano)@uva.nl + +# Abstract + +Spatially dense self-supervised learning is a rapidly growing problem domain with promising applications for unsupervised segmentation and pretraining for dense downstream tasks. Despite the abundance of temporal data in the form of videos, this information-rich source has been largely overlooked. Our paper aims to address this gap by proposing a novel approach that incorporates temporal consistency in dense self-supervised learning. While methods designed solely for images face difficulties in achieving even the same performance on videos, our method improves not only the representation quality for videos – but also images. Our approach, which we call time-tuning, starts from image-pretrained models and fine-tunes them with a novel self-supervised temporal-alignment clustering loss on unlabeled videos. This effectively facilitates the transfer of high-level information from videos to image representations. Time-tuning improves the state-of-the-art by $8 - 10\%$ for unsupervised semantic segmentation on videos and matches it for images. We believe this method paves the way for further self-supervised scaling by leveraging the abundant availability of videos. The implementation can be found here: https://github.com/SMSD75/Timetuning + +# 1. Introduction + +Dense self-supervised learning, whereby meaningful deep features for each pixel or patch of input are learned in an unsupervised manner, has recently received increasing attention [74, 24, 57, 51]. By learning spatially consistent features for different views of an input, strong gains in unsupervised semantic segmentation have been achieved using unlabeled images. However, so far, an even more information-rich source for unsupervised training has been largely overlooked: videos. With their additional time dimension and being the most rapidly growing form of digital content, they are well-suited to scaling dense self-supervised learning even further. + +Some efforts have already been made to learn from the video domain by using different frames from a video as aug- + +![](images/6d6a21dd569aa53cf12c8bad541e03655ae4cf6be65c2e2b8b02480bded49a00.jpg) +Figure 1: Time-tuning compared to previous methods. Unlike existing methods that ignore or utilize expensive 3D models to implicitly model temporal information, the proposed method explicitly incorporates temporal consistency in dense feature representations using a temporal self-supervised loss. The method starts with a 2D encoder pretrained on images and fine-tunes it using unlabeled videos. This approach leads to improved performance not only for videos but also for images. + +mentations [18, 63, 49] or by involving temporal correspondence [71, 58]; however, they mostly did it in a supervised way [29, 73, 19, 40, 65, 38, 31, 28, 41, 37, 54, 44], which is not scalable specifically for dense tasks where the number of targets can increase significantly as the number of pixels grows. To this end, self-supervised learning approaches offer a solution by reducing the need for supervision. However, these methods typically rely on the notion of "views", which involves learning similar features for corresponding locations over time. This usually leads to a chicken-and-egg problem, where the correspondences are required for learning dense features - which in turn enable good correspondences [30]. + +In images, the challenge is trivially solved by considering the correspondence function based on the augmentation function. For instance, In the case of color augmentations, this correspondence is simply given by the identity. However, shifts in time cannot be viewed as mere augmenta + +tions. As we demonstrate through a new evaluation protocol, using image models on frames alone is not nearly as effective. A similar finding has also been reported, albeit for non-dense works [20, 34], which assumed the passage of time as an image augmentation. These works have generally reported reduced performances, even when compared to simple image-based pretraining methods. Similarly, video-level tasks [17, 18, 55] assume sufficiently similar semantics between different frames. This is also not true for dense tasks, as static features can only be assumed where nothing is moving – which is rare due to possible object, camera, and background motion between the frames. + +To address this challenge, we propose to model the additional time-dimension explicitly to identify which pixels should retain similar embeddings and which should not. We propose two separate modules to tackle the correspondence and the dense learning, respectively. For the former, we introduce the Feature-Forwarder (FF) module, which breaks the mentioned chicken-and-egg loop by leveraging the good tracking performance of pretrained image models, and allows an approximate second "view" that can then be treated as a target for the further dense self-supervised loss. On top of this, we introduce a spatio-temporally dense clustering module, which learns unsupervised clusters across samples, locations and time. Using these two components and starting from image-pretrained features, our proposed method allows time-tuning (TIMET) the dense representation in a self-supervised manner, see Figure 1. + +Finally, we demonstrate that TIMET paves the way for further scaling of self-supervised learning by leveraging the abundant availability of video datasets and transferring their knowledge to the image domain. This results in consistently achieving state-of-the-art performances not only for the task of unsupervised semantic segmentation of videos, but also for unsupervised image semantic segmentation, a feat previously out of reach for methods trained on videos. + +Overall, this paper makes the following contributions: + +- We show that image-based unsupervised dense segmentation models applied to videos exhibit degraded performance and lack temporal consistency in their segmentation maps. +- Building on this observation, we propose a novel dense self-supervised learning method that utilizes temporal consistency as a learning signal. +- We demonstrate that our method enables the scaling of self-supervised learning by leveraging abundant video datasets and effectively transferring knowledge to the image domain. Our approach consistently achieves state-of-the-art performance for both images and videos, opening up new opportunities in the field. + +# 2. Related Works + +Dense self-supervised learning. These methods build upon image-level self-supervised representation learning by incorporating existing losses to enhance the spatial features, demonstrating a commendable advancement. DenseCL [61] works on spatial features by constructing dense correspondences across views using the contrastive objective given in MoCo [23], while PixPro [62] utilizes the augmentation wrapper to get the spatial correspondence of the pixel intersection between two views. Similarly, MaskContrast [57], Leopart [74], DetCon [24] and Odin [25] also ensure spatial feature similarities via contrastive learning and spatial correspondences. Self-Patch [67] treats the spatial neighbors of the patch as positive examples for learning more semantically meaningful relations among patches. Inspired by SelfPatch, ADCLR [69] proposes patch-level contrasting via query crop and cross-attention mechanism. Pursuing the same objective through none end-to-end approaches. Both [68] and [59] propose an unsupervised salient object segmentation pipeline that extracts noisy object masks from the inputs and fine-tunes a specifically designed object segmentation head with several self-training steps to make an unsupervised semantic segmentation model. Unlike the existing image-based approaches, we propose a method that improves the dense prediction performance of a pretrained encoder by explicitly modeling the temporal dimension and learning from diverse natural dynamics and variations found in videos. + +Video to image knowledge transfer. Learning from videos instead of images has recently received attention since they contain far more information than still images and hold the potential for learning rich representations of the visual world. To this end, several non-dense works have been released [46, 4, 27]. VITO [46] shows that naively applying image domain self-supervised learning methods on videos can lead to a performance drop coming from the distribution shift, which by applying data processing techniques is relieved but not fully solved. Following the same way, [27] trains masked autoencoders [22] with contrastive learning [10] on video datasets and shows a decent performance on both image and video tasks. While video to image knowledge transfer is the end goal of such methods and our paper; our dense task is different and is as of yet unaddressed. + +Unsupervised video object segmentation. Unsupervised video object segmentation methods do not require any manual annotations. However, they are only designed to tackle a foreground/background segmentation task, which refers to the segmentation of the most prominent, general objects in video sequences [66, 43, 50, 1, 39]. For instance, [66] + +![](images/28984550502a23d9386dd0e8146f9d52eb8369dfc1db154caa89a6f52ebad4ca.jpg) +Figure 2: Time-Tuning overview. We conduct self-supervised video semantic segmentation by tuning image-pretrained models with temporal information from videos. (a): Our general pipeline of adapting an image-pretrained ViT transformer on video data using our dense clustering loss. The encoder is specified by $\Psi$ and kept frozen except for the last two layers. Also, $z_{t}$ shows the output of Sinkhorn-Knopp algorithm, which is forwarded to time-step $T$ to be compared with the features that are obtained using the last time-step. (b): Detailed view into our Feature-Forwarder module that is used for aligning cluster-maps from past frames to logit-features of future ones. + +![](images/fc73b7e624c145655f8448ce3e80b31cca95cd290dd4ee16fc8e3c875e621a2b.jpg) + +uses an AutoEncoder-based architecture based on slot attention so that pixels that show the same motion flows based on optical-flow are grouped together. In this way, dominant moving objects can be detected and separated from the background. In our experiments, we benchmark two representative instances of this line of work, i.e., [66, 1], to provide better insights into the differences of unsupervised object detection methods and our proposed benchmark for video semantic segmentation. + +To conclude, we propose the first self-supervised video semantic segmentation method, which tunes image-trained models such as [9, 72] based on the temporal information that exists in unlabeled videos to enhance their effectiveness in dense prediction tasks. + +# 3. Method + +Our method works by densely clustering features in a manner that is consistent with time. At a high level, it works by transforming features from a past time $t' = t$ to the current frame $t' = T$ and forcing these to be consistent with the currently observed ones at $t' = T$ . The overall learning signal comes from the dense temporal clustering of forwarded features and allows "timetuning" image-pretrained models to learn from videos. Figure 2 shows the proposed method's architecture. In the following, we describe each component in detail. + +# 3.1. Feature Forwarding + +The goal of this component, the feature-forwarder (FF), is to propagate the dense features in time as accurately as possible, given some RGB input frames. More formally, let $f$ be the feature forwarding function, $I_{t}$ be a frame at time $t$ for a given video, and $z_{t}$ the features corresponding to $I_{t}$ , + +then the task is to propagate $z_{t}$ to a future time $T$ , + +$$ +f \left(\left[ I _ {t}, \dots , I _ {T} \right], z _ {t}\right)\rightarrow \hat {z} _ {T}. \tag {1} +$$ + +Note that $\hat{z}_T$ and $z_{T}$ are not necessarily the same for $T > t$ ; since $\hat{z}_T$ represents an approximation of $z_{T}$ using the previous features. Obviously, the further $T$ is from $t$ , the more variance is seen and therefore a better learning signal might be provided; however, for such larger intervals, the idea of feature forwarding encounters two main challenges: error accumulation and object non-permanence. + +Challenge 1: Error accumulation. While computing how features have changed simply based on the start and end frames is the most straightforward (i.e., $I_{t}$ and $I_{T}$ ), it does not fully leverage the knowledge contained in the intermediate frames. Instead, it is common [30, 60, 36] to forward features for every time-step $\delta t$ , i.e., + +$$ +f ([ I _ {t},.., I _ {T} ], z _ {t}) = \bigcirc_ {t ^ {\prime} = t + \delta t} ^ {T} f ([ I _ {t ^ {\prime} - \delta t}, I _ {t ^ {\prime}} ], z _ {t} ]), \tag {2} +$$ + +where $\bigcirc$ is a composition operator such as element-wise multiplication. However, this results in the accumulation of errors over time. + +Challenge 2: Object non-permanence. One difficulty that arises when using videos as training data is the fact that objects sometimes simply disappear from the screen. Compared to images, which carry a heavy photographer's bias, videos are more prone to natural variation due to camera and object motion resulting in temporary object non-permanence in a particular video clip. A method that simply assumes an object (or a feature) has to be present in any given frame simply because it was present in the previous + +one might therefore easily fail when going towards videos that exist in-the-wild. A case in point are simple occlusions that arise from an object being fully- or partially-covered by another one further in the background. Taking longer training clips has been attempted to resolve the issue. However, this increases the training complexity, especially for dense-tasks that necessitate a higher number of predictions for each input. + +We address these challenges using a novel, yet simple component, which we call the Stabilizing Feature-Forwarder. The first choice in designing this module is the function and data used for composing information across time. For instance, a simple approach would be to use pixel-wise optical flow and aggregate the flow at the feature level to predict how the features change over time. However, as we also show in our experiments, optical flow is prone to accumulating error with time and also cannot easily recover from occlusions [6]. Instead, we utilize the fact that we wish to pretrain the visual encoder and recycle part of it for the purpose of feature forwarding. This not only speeds up the process, as no additional encoder/modality is required, but also produces a positive feedback loop of better forwarding and better feature learning. + +Concretely, we first compute the L2-normalised dense features of a pretrained visual encoder $\Phi$ to compute spatial similarities across time (the ' in $t^\prime$ left out for clarity): + +$$ +F _ {i j} ^ {t} = \left\langle \Phi \left(I _ {t}\right) _ {i}, \Phi \left(I _ {T}\right) _ {j} \right\rangle / \tau , \tag {3} +$$ + +which yields matrices $F^{\{t,\dots T - \delta t\}} \in R^{N\times N}$ with values between [0,1] indicating semantic similarities between the $N$ spatial features that are sharpened by a temperature $\tau$ . Note that because we compare each frame with the final frame, the effect of object non-permanence can be minimized: even if the object is only present in every other frame, there is enough signal to forward its features. + +Next to propagate from time $t$ to $T$ , these similarities are stacked and normalised along time, as follows: + +$$ +\tilde {F} _ {i j} ^ {t} = \exp \left(\mathcal {N} \left(F _ {i j} ^ {t}\right)\right) / \sum_ {i, t ^ {\prime}} \exp \left(\mathcal {N} \left(F _ {i j} ^ {t ^ {\prime}}\right)\right), \tag {4} +$$ + +where $t \leq t' \leq T$ and $\mathcal{N}$ is a neighborhood thresholding method which forces short-term spatial-smoothness, i.e., $\mathcal{N}(\tilde{F}_{ij}^t) = 0$ if $i$ is not within a local window around $j$ with the size $k$ . Finally, the propagated feature are computed $\hat{z}_T$ as: + +$$ +\hat {z} _ {T} (j) = \sum_ {t ^ {\prime}, i} \tilde {F} _ {i j} ^ {t ^ {\prime}} \hat {z} _ {t ^ {\prime}} (i), \quad j \in \{1,.., N \}. \tag {5} +$$ + +This means that for arriving at the target feature $\hat{z}_T$ , we not only use the previous frame as the source, but instead use the past $(T - t) / \delta t$ frames and aggregate these. + +Relation to mask-propagation methods. While previous methods such as DINO [9] and STC [30] have utilized a similar technique for propagating ground-truth masks for evaluating, for example on the DAVIS dataset, there are three key differences to our forwarding method. First, instead of propagating binary maps of foreground-vs-background, our feature-forwarder has as inputs soft, noisy and multi-label self-supervised segmentation maps, which require the forwarder to tolerate overlapping of different object probabilities throughout training. Second, previous methods have used this approach mainly for inference; we, however, use this module as a trainable component and show how our loss improves this forwarder with training time. Finally, we do not follow a typical re-normalizing step across feature dimensions (typically done after Eq. (5)) as this harms the scale of logits that are being propagated and leads to heavily diluted target distributions. + +# 3.2. Self-supervised dense clustering + +While the Feature-Forwarder produces target features that include information about the dynamics with time, its computation utilizes $\Phi(I_T)$ and could lead to trivial solutions, i.e., $f([I_t,..,I_T],z_t) = \Phi(I_T)$ . To counteract this, we propose a self-supervised clustering task across views in time. For this, we utilize the basic online clustering algorithm based optimal-transport [12], utilized in works such as SeLa [2], SwAV [8] and DINOv2 [45]. In particular, let $\Psi$ be a visual encoder with a clustering head $g$ that yields a $K$ dimensional output, then the clustering loss is given by: + +$$ +\mathcal {L} \left(x _ {i}\right) = - \tilde {y} _ {i} \log (g (\Psi (x) _ {i})), \tag {6} +$$ + +which is a standard cross-entropy loss with regards to self-supervised pseudo-label $\tilde{y}_i$ . These labels, in turn, are generated by solving an entropy-regularised optimal transport problem on the batch $\mathcal{B}$ [2]: + +$$ +\min _ {\tilde {y}} \langle \tilde {y}, - \log g (\Psi (x)) \rangle + \frac {1}{\lambda} \operatorname {K L} (\tilde {y} \| r c ^ {\top}), \tag {7} +$$ + +$$ +\text {w i t h} r = \frac {1}{K} \cdot \mathbb {1}, \quad c = \frac {1}{| \mathcal {B} |} \cdot \mathbb {1}. \tag {8} +$$ + +Here $\lambda$ controls the entropy regularisation and $r, c$ are the marginals for the prototypes and the batch, respectively. Note that solving this problem can be done extremely quickly on the GPU and yields soft pseudo-labels $\tilde{Y}$ , such that $\operatorname{argmax}(\tilde{Y}) = \tilde{y}$ . + +n, these might yield sufficiently similar semantics [17] this is not true for dense tasks, as static features can only be assumed for videos where nothing is moving, which is rare. + +# 3.3. Overall TIMET loss + +We combine the previous two modules to arrive at our full method (as shown in Fig. 2): First, self-supervised + +Sinkhorn-Knopp clustering is conducted on early features $g(\Psi(I_t))$ , yielding soft pseudo-labels $\mathrm{SK}(g(\Psi(I_t))) = \tilde{Y}_t$ . These are then forwarded in time using our Feature-Forwarder to arrive at dense targets $\mathrm{FF}(\tilde{Y}_t)$ , which are used in the final loss. Compactly: + +$$ +\mathcal {L} _ {\mathrm {T I M E T}} \left(I _ {T}\right) = - \sum_ {i, j} \operatorname {F F} \left(\tilde {Y} _ {t}\right) \log \left(g \left(\Psi \left(I _ {T}\right)\right)\right). \tag {9} +$$ + +# 4. Experiments + +# 4.1. Setup + +Datasets. We train our method and baselines on YTVOS [64], one of the largest video segmentation datasets available, and evaluate on DAVIS17 [48] and YTVOS. For YTVOS the ground truth masks are only available for the first frames of the test and validation sets, and therefore, a fixed random $20\%$ of the training set is used for testing. For transfer learning experiments, we use the validation set of Pascal VOC 2012 [16], As the dataset has been commonly used as a main reference for recent works in dense self-supervised image segmentation [74, 57, 61]. For completeness, we also report the performance on egocentric datasets that have less object-centric bias and are prevalent in real-world scenarios. For egocentric experiments, we train on EPIC-KITCHENS-100 [13] and evaluate on VISOR [14]. Further details are provided in Appendix A. + +Models and baselines. Currently, there is a lack of available unsupervised semantic video semantic segmentation methods. Nevertheless, we have included a comprehensive comparison of our method with state-of-the-art techniques in both image and video domains for unsupervised image semantic segmentation and unsupervised video object segmentation. To evaluate the image-based models, we utilized either official reported numbers or provided pretrained models of STEGO [21] and Leopart [74]. We have taken every measure to ensure fairness in our comparison. To do so, all the used pretrained models have the same pretraining dataset (ImageNet-1k), and the number of their backbone parameters is roughly similar. For those models that use extra datasets, for instance, Leopart, we select the pretrained backbones that closely match the specification of YTVOS training dataset. Additionally, we trained image-based models on the same video datasets, where we converted video data into their corresponding image data and reported their performance wherever possible. To ensure comprehensive analysis, we have included the recent concurrent work Flowdino [71], which utilizes optical flow to refine DINO features on unlabeled videos, in our comparison benchmarks as well. This is done only in cases where there is a shared experimental setup. + +Evaluation procedure. Despite unsupervised object segmentation being a well-established evaluation in the image domain [57, 74], evaluating unsupervised video multi-label object segmentation is challenging due to the absence of an established evaluation protocol for video object semantic segmentation without supervision. In this regard, we propose a set of evaluation protocols for unsupervised video multi-label object segmentation, which exploits existing video object segmentation datasets for evaluation purposes (see details in Appendix B). In our experiments, we discard any projection head used during training and evaluate ViT's spatial tokens directly, similar to [57, 32, 61], using four methods: classification with a linear head, classification with an FCN head, overclustering, and clustering with as many clusters as the number of ground truth objects. To fine-tune different heads on top of the frozen spatial tokens, we follow [57]. For unsupervised semantic segmentation, we apply $K$ -Means on all spatial tokens, where $K$ is chosen to be higher or equal to the ground-truth number of classes, following the common protocol in image clustering [32, 56]. For grouping clusters to ground-truth classes, we match them either by pixel-wise precision or Hungarian matching on merged cluster maps [35]. For video datasets, we allow the matching to be per-frame, per-video, and per-dataset. The evaluation metric is permutation-invariant, following [32], and the results are averaged over five different seeds. Clustering evaluations are preferred as they require less supervision and fewer hyperparameters than training a linear classifier and operate directly on the learned embedding spaces. We report our results in mean Intersection over Union (mIoU) unless otherwise specified. + +Model training. We train a ViT-Small with patch size 16 and initialized from ImageNet-pretrained DINO weights [9] using the proposed self-supervised loss function. For the ablations and the main experiments, our models are trained for 12 and 30 epochs respectively. Further training details are provided in Appendix A. Code will be made available. + +# 4.2. Ablations + +We first examine the essential parameters of our method by training TIMET on YTVOS and assessing its ability to perform semantic segmentation on Pascal VOC. In addition to presenting the results through clustering and overclustering, we also demonstrate linear classification outcomes. + +Number of prototypes. We first ablate the influence of the number of prototypes on downstream segmentation performance. Results in Table 1a indicate a sharp increase in performance with a rise in the number of prototypes, but once a sufficiently large number is reached, performance stabilizes. We observe peak performance with a moderate number of 200 prototypes. The stability of our method over + +(a) Ablating # prototypes $K$ + +
PLCK=21K=500
1037.56.124.1
5058.28.240.5
20059.79.242.8
30059.49.042.3
+ +(b) Ablating time interval $\delta T$ + +
#framesδTLCK=21K=500
10s50.75.626.2
40.2s56.57.436.5
40.5s59.79.242.8
41.0s57.18.338.1
+ +(c) Ablating number of frames used. + +
#framesTLCK=21K=500
10s50.75.626.2
22.0s52.66.237.1
42.0s59.79.242.8
82.0s59.79.042.3
+ +a range of prototypes (50-300) suggests that our approach is robust to the change of this parameter. + +Understanding FF. Next, we ablate the temporal dimensions that influence the working of the Feature-Forwarder. In Table 1b, we vary the time interval $\delta T$ between frames whilst keeping a fixed number of four frames. We generally find an increase in performance when increasing the time-interval, with performance peaking at 0.5s. One of the most critical observations is the clear difference in clustering performance from $26.2\%$ to $42.8\%$ between training on single frames (first row) vs. training with multiple frames. This clearly indicates the significant benefit of utilizing temporal information from the Feature-Forwarder during training. + +Next, in Table 1c we vary the number of frames given a fixed clip duration of 2s. Again, we find that a moderate number of frames between the source and target time is most favorable. While additional frames do not degrade the performance much, they do add computational complexity, so we prefer using 4 frames in our main method. + +Choice of propagation feature. Finally, in Table 2, we vary the type of feature that we forward to a future time. In the first row, we report the performance when not using any feature forwarding – thus solely training on single-frame inputs (i.e., the same as row 1 in Table 1c). In the second row, "Identity" shows the performance that is obtained when the feature map from the source frame is simply forwarded to the future without any changes, which shows an increase in the performance compared to the first row. This shows that often, training videos are not very dynamic, and a static assumption can already lead to some gains. However, compared to forwarding the Sinkhorn-Knopp (SK) clustered features, these gains are small (+4.6% vs. +17%). Importantly, our Feature-Forwarder relies on forwarding SK-sharpened feature maps and does not work when simply forwarding the network's output logits $\Phi(x)$ , followed by a clustering step. The reason for obtaining considerably lower numbers here can be traced back to using an un-entropy + +Table 1: Ablations of the key parameters of our method. The model is trained for 12 epochs on Pascal VOC, and results for unsupervised segmentation with clustering $(K = 21)$ , overclustering $(K = 500)$ , and linear pixel-wise classification (LC) are shown. The stability of our method over a range of prototypes (50-300), inter-frame time intervals $(\delta T \in [0.5s - 1.0s])$ , and the number of training frames (4-8) at a fixed clip duration $(T)$ shows the robustness of the method. + +
FF typeLCK=21K=500
None50.75.626.2
Identity55.37.436.1
Ψ(x)53.16.635.5
SK(Ψ(x))59.79.242.8
+ +Table 2: Propagating different features in FF on Pascal VOC. We find that our Sinkhorn-Knopp (SK) based module outperforms static training ('None' or 'Identity'). Moreover, propagating logits $\Psi(x)$ does not work as well as SK-regularised features. + +
PretrainBackboneAt Init+TIMET
K=500LCK=500LC
MSN [3]ViT-S/1626.055.448.3↑22.367.2↑11.8
iBOT [72]ViT-S/1632.162.147.1↑15.067.1↑5.0
DINO [9]ViT-S/822.555.853.9↑31.469.5↑13.7
DINO [9]ViT-B/1628.959.152.7↑23.870.2↑11.1
+ +Table 3: Applying TIMET to different pretrainings on Pascal VOC. TIMET can boost $(\uparrow)$ the performance of different backbones with different initialization by a considerable margin, showing the generality of our approach. + +regularized clustering algorithm. In this case, the logits tend to cause a few prototypes to dominate the cluster centers, exacerbated after the propagation, resulting in highly noisy and uninformative propagated logits. This shows that careful design of the Feature-Forwarder is indeed required. + +Applying TIMET to different models. As shown in Table 3, our method is generalisable to different backbones and self-supervised learning initializations, enabling it to be an effective method to transfer the knowledge of the videos to images in an unsupervised way, which reduces the cost of labeling for different downstream tasks. + +![](images/e2bf554d9064aafcabf706009fc36cda2f1000c3051dd8771930d0a7cc3b5ea4.jpg) +Pascal + +![](images/87d163b4a5d1599b1025b5b3114a31d5585e00ce5c7897190285587edc754ef0.jpg) +Figure 3: TIMET unsupervised segmentations. As we regularize DINO's backbone to be consistent across time on YTVOS, it obtains strong performance on both image and video segmentation datasets, yielding high class consistencies (indicated by the segmentation colors) and tight borders. We provide more qualitative results in Appendix D. + +![](images/a25781e63c584f4090988165ab0be9a3e4d80f26c37e766cdf2efdfcb782b9aa.jpg) +DAVIS + +![](images/9a9fb08e3310fa4ecc3051380b0d2343b339d9bce864880c6e3d1b075cb98613.jpg) + +![](images/1fdbc1e8b8ec9f21d7cc2d4f2c49bd9bf71a4563ab27d404be004d8f82587bc1.jpg) +YTVOS + +![](images/384f2e894c87e57b41bbd3aced56eb59e5711d58807fa4efdcf74d44e6cb137e.jpg) + +![](images/b30519f2e15c0745bb711cec09f0c67d48d038dfb6f34eeebb7112227728083c.jpg) +VISOR + +![](images/0d172b6c77e28e038a2a8e85d0f7a0e96b0476e5fd2e22447bfd052647fd3e10.jpg) + +
ClusteringOverclustering
YTVOSDAVISYTVOSDAVIS
FCDFCDFCDFCD
Trained on Images
Resnet5044.043.41.739.337.44.255.652.83.146.644.28.4
SwAV [8]39.538.23.232.029.67.359.858.15.850.550.125.7
DINO [9]39.137.91.930.231.01.666.265.44.056.954.917.9
Leopart [74]39.237.911.730.330.216.564.562.815.554.954.426.7
Trained on Videos
STEGO*41.540.32.031.931.03.258.154.35.147.646.310.4
DINO*37.236.11.229.329.22.453.150.91.345.444.08.6
Leopart*41.540.57.737.536.512.660.859.86.853.753.116.8
TIMET(ours)52.551.313.353.753.020.568.666.815.859.861.531.8
+ +Table 4: Unsupervised video semantic segmentation results in mIoU for clustering (K=GT) and overclustering. TIMET not only gets better numbers on YTVOS, but also achieves considerably better results on DAVIS. This shows the better quality of the learned features at transferability to other video datasets. The clip-length is set to 16 and 4 for DAVIS and YTVOS, respectively. For clustering, the Hungarian algorithm [35] matches the unsupervised segmentation clusters (K) with the ground truth (GT) per frame $(F)$ , clip $(C)$ or across the whole dataset $(D)$ . For overclustering, we use K=10 for the frames-wise $(F)$ and clip-wise $(C)$ evaluations, and for dataset-wise $(D)$ evaluation, K=200 for DAVIS and K=500 for YTVOS. *denotes finetuning pretrained models on the same video frames as input as our method. The matching protocol is greedy many-to-one, see Appendix B for details. + +# 4.3. Large-scale experiments + +Unsupervised video semantic segmentation. In this section, we train our TIMET method on the YTVOS dataset and evaluate it for unsupervised video object segmentation on both DAVIS and YTVOS. The results are shown in Table 4. It should be noted that the number of prototypes we analyzed in Table 1a is not the number of clusters used to report accuracies for Table 4, as the actual number is usually unknown in the case of unsupervised learning. For the clustering experiment, we set K to the number of ground truth objects, which makes evaluation metrics easier to interpret and is commonly done in image clustering and segmentation. This evaluation is repeated but for larger numbers of clusters, in the "overclustering" scenario. Note that in dense self-supervised learning, overclustering can be particularly important because the learned representations are + +used as a feature extractor for downstream tasks such as semantic segmentation or object detection. By using more fine-grained representations, the network may be able to extract more discriminative features such as object parts [74], which can lead to better performances. + +From Table 4, we observe a clear trend: our method, trained on YTVOS not only achieves superior performances on YTVOS, but also beats existing image-trained models on DAVIS by a large margin. In particular, the state-of-the-art self-supervised clustering method, Leopart, has $4\%$ lower performance on per dataset DAVIS evaluation for $\mathrm{K = GT}$ and $1.6\%$ lower performance on YTVOS. The gap even becomes larger when Leopart is trained on the same video dataset with $8\%$ to $16\%$ lower numbers in per dataset and per clip numbers across different datasets. This means that when grouping objects of the same class over time and + +the whole dataset, the image-based self-supervised methods have trouble generalising to videos, where objects are not centered in the frame, can appear in varying poses, and are more difficult in general. A strong contender is Leopart, which matches the performance of DINO for per frame and per clip clustering, however, improves considerably with the per dataset clustering. We attribute this to the fact that their dense learning objective improves mainly the generalisation capabilities of the learned representation, however falling short of generalising to temporal variations. We make similar observations for overclustering, improving the DINO baseline which we utilize as initialisation by $12\%$ to $14\%$ across different datasets in the per dataset metric. + +We conclude that the proposed method outperforms all other methods in learning robust and discriminative representations for dense video understanding, and image-based self-supervised learning may not have sufficient generalisation capacity. This shows that if used right, time is a particularly strong inductive bias for self-supervised learning. Figure 3 shows segmentations returned by the proposed model trained on YTVOS and tested on YTVOS, DAVIS, and VISOR respectively. The proposed method groups objects accurately, and importantly, the frame segmentations are considerably consistent over time (see more examples in Appendix D). This highlights the importance and relevance of temporal fine-tuning, not just for higher accuracy, but crucially for consistency and robustness. + +Transfer from video training to images. Despite the common belief that features learned from videos perform worse when transferred to images [34, 20], the results in Table 5 demonstrate our method achieves high performance that match state-of-the-art methods directly trained on images, specifically for the $\mathrm{K} = 500$ , FCN, and LC metrics. We also compare our performance gains in videos against models designed for unsupervised image segmentation. The results show that models highly biased towards image datasets cannot provide the same performance when trained on videos, lagging behind our approach by $7\%$ to $30\%$ . Our results demonstrate that achieving high transfer performance on challenging tasks through video-based self-supervised learning is not only feasible but can also maintain high performance across modalities. These findings suggest that our method can drive further advances in self-supervised learning and inspire new directions for research in this field. Figure 3 shows the qualitative results on Pascal VOC. For more visualisations we refer to Appendix D. + +Salient object segmentation. In Table 6, we compare foreground masks obtained with various DINO ViT based methods. We use the cluster-based foreground extraction protocol from [74] (details provided in Appendix C). First, we find that our method outperforms the original DINO at + +
Pascal VOC
K=21K=500LCFCN
Trained on Images
ResNet-504.536.553.8-
DINO [9]5.517.450.660.6
SwAV [8]11.635.750.7-
MaskContrast [57]35.045.449.2-
DenseCL [61]-43.649.069.4
STEGO [21]7.019.559.163.5
CrOC [52]20.6-61.6-
Leopard [74]36.650.568.070.1
Trained on Videos
STEGO*4.015.551.155.5
Leopard*14.921.253.263.2
Flowdino†[70]--59.4-
TIMET (ours)34.553.268.070.6
+ +Table 5: Transfer from video training to images. Numbers taken from [74, 52, 70]. *: finetuning pretrained models on the same video frames as input as our method. †: uses a 40% larger superset of our train data. Details on the different datasets and methods are provided in the Appendix. + +
Pascal VOCDAVISYTVOS
DINO [9]52.134.532.1
Leopard [74]59.637.338.6
STEGO [21]49.130.432.1
TIMET (ours)63.944.543.5
+ +Table 6: Salient object segmentation. We report performance using the Jaccard score [9] and use the official pretrained models for evaluation. + +tention maps consistently by $10 - 11\%$ . We also surpass the results of Leopart [74] and STEGO [21], works that rely on the same pretrained backbone. Even for the evaluation on Pascal VOC, our method achieves higher performances despite the domain shift of having trained on videos. + +Generalisation to egocentric datasets. We train our method on EPIC-Kitchens-100 [13] and evaluate on VISOR [14]. We use the official code to convert VISOR to a DAVIS-like structure, in which we can report our numbers for per frame and per clip evaluation. Note that, after conversion, the object IDs do not maintain global consistency in the whole dataset; therefore, we cannot report the per dataset number. As Table 7 shows, TIMET outperforms image-based competitors on egocentric datasets as well. + +
VISOR
FC
DINO [9]24.818.7
Leopard [74]24.118.5
TIMET (ours)26.521.5
+ +Visual In-Context Learning evaluation. Here, we contrast our approach with a recently introduced benchmark, which assesses the in-context reasoning capabilities of models within the vision domain. Unlike linear or FCN classification methods, visual in-context learning evaluation obviates the need for fine-tuning or end-to-end training. Instead, it constructs validation segmentation maps using patch-level, feature-wise nearest neighbor similarities between validation images (referred to as "queries") and training samples (termed "keys"). This approach mirrors strategies in the NLP domain, aiming to evaluate the proficiency of models in learning tasks from limited examples presented as demonstrations. The results are shown by Table 8. Given that the vast majority of extant models in the domain utilize ViT-S16 as their backbone, we conducted a re-evaluation of their checkpoints to furnish a consistent and directly comparable evaluation table. Subsequently, we re-trained our model employing ViT-B16 and benchmarked it against other baselines as presented by [5]. The results, as depicted in the table, indicate that even though our model was exclusively trained on videos, it registers performance metrics in line with Leopart [74]. Furthermore, it surpasses the results of the leading method, CrOC [52], which was trained on images. Contrary to Hummingbird [5], TIMET is not custom-fitted to the specific evaluation setup and boasts superior computational efficiency. Notably, it requires only a single GPU for training, in contrast to the 16 TPUs demanded by Hummingbird. + +# 5. Conclusion + +This paper has aimed to learn dense representations that are learned from videos; yet can be generalised to the image domain as well. As video content is growing rapidly and contains more information compared to images, learning generalisable knowledge from them facilitates further scaling of self-supervised learning methods. To this effect, we have proposed a self-supervised clustering loss to encourage temporally consistent features between different frames of a video. To efficiently find corresponding views between + +Table 7: Generalisation to egocentric datasets. Unsupervised semantic segmentation results in mIoU for clustering $(\mathrm{K} = \mathrm{GT})$ on VISOR [14] after training onn EPIC-Kitchens-100 [13]. The clip-length is set to 4. Our method gets better results on egocentric datasets as well. + +
EncoderParamsmIoU
Trained on Images
SupervisedViT-S1621M35.1
MoCo-v3* [11]ViT-S1621M19.5
DINO* [9]ViT-S1621M47.9
CrOC* [52]ViT-S1621M50.0
Leopard* [74]ViT-S1621M63.6
DINO [9]ViT-B1686M55.9
MoCo-v3 [11]ViT-B1686M37.2
MAE [22]ViT-B1686M6.6
LOCA [7]ViT-B1686M57.5
Hummingbird [5]ViT-B1686M70.5
Trained on Videos
TIMET (ours)ViT-S1621M61.6
TIMET (ours)ViT-B1686M65.5
+ +Table 8: Visual In-Context Learning evaluation. Numbers are taken from [5]. *: numbers are produced by this paper. This approach is equivalent to a non-parametric nearest-neighbor based evaluation on Pascal VOC. The details of the evaluation benchmark can be found in the provided implementation codes. + +clip frames, we have proposed to recycle pretrained transformer features to leverage their natural tracking ability at each clip. Our empirical results indicate that this method achieves significant gains on the challenging task of video object segmentation across three different evaluation protocols and three datasets. Moreover, by transferring the learned model to the image domain, we have demonstrated the generalisability of the learned features by surpassing or matching the state-of-the-art for unsupervised image segmentation. + +# 6. Acknowledgement + +This work is financially supported by Qualcomm Technologies Inc., the University of Amsterdam and the allowance Top consortia for Knowledge and Innovation (TKIs) from the Netherlands Ministry of Economic Affairs and Climate Policy. + +# References + +[1] Nikita Araslanov, Simone Schaub-Meyer, and Stefan Roth. Dense unsupervised learning for video segmentation. Advances in Neural Information Processing Systems, 34:25308-25319, 2021. 2, 3, 15 +[2] Yuki M. Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In International Conference on Learning Representations (ICLR), 2020. 4 + +[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI, pages 456-473. Springer, 2022. 6 +[4] Arthur Aubret, Markus R. Ernst, Céline Teulière, and Jochen Triesch. Time to augment self-supervised visual representation learning. In The Eleventh International Conference on Learning Representations, 2023. 2 +[5] Ivana Balăžević, David Steiner, Nikhil Parthasarathy, Relja Arandjelović, and Olivier J Henaff. Towards in-context scene understanding. arXiv preprint arXiv:2306.01667, 2023. 9 +[6] Thomas Brox and Jitendra Malik. Large displacement optical flow: Descriptor matching in variational motion estimation. IEEE TPAMI, 33(3):500-513, 2011. 4 +[7] Mathilde Caron, Neil Houlsby, and Cordelia Schmid. Location-aware self-supervised transformers. arXiv preprint arXiv:2212.02400, 2022. 9 +[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 4, 7, 8 +[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 3, 4, 5, 6, 7, 8, 9, 15 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[11] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. ICCV, 2021. 9 +[12] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural Information Processing Systems, volume 26, 2013. 4 +[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 5, 8, 9 +[14] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2022. 5, 8, 9 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- + +vain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 13 +[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 5, 13, 14 +[17] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3299-3309, 2021. 2, 4 +[18] Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould. Self-supervised video representation learning with odd-one-out networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3636-3645, 2017. 1, 2 +[19] Raghudeep Gadde, Varun Jampani, and Peter V Gehler. Semantic video cnns through representation warping. In Proceedings of the IEEE International Conference on Computer Vision, pages 4453-4462, 2017. 1 +[20] Daniel Gordon, Kiana Ehsani, Dieter Fox, and Ali Farhadi. Watching the world go by: Representation learning from unlabeled videos. arXiv preprint arXiv:2003.07990, 2020. 2, 8 +[21] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 5, 8 +[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 9 +[23] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[24] Olivier J Henaff, Skanda Koppula, Jean-Baptiste Alayrac, Aaron Van den Oord, Oriol Vinyals, and João Carreira. Efficient visual pretraining with contrastive detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10086-10096, 2021. 1, 2 +[25] Olivier J Hénaff, Skanda Koppula, Evan Shelhamer, Daniel Zoran, Andrew Jaegle, Andrew Zisserman, João Carreira, and Relja Arandjelović. Object discovery and representation networks. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 123-143. Springer, 2022. 2 +[26] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 13 +[27] Jefferson Hernandez, Ruben Villegas, and Vicente Ordonez. Visual representation learning from unlabeled video using contrastive masked autoencoders. arXiv preprint arXiv:2303.12001, 2023. 2 +[28] Ping Hu, Fabian Caba, Oliver Wang, Zhe Lin, Stan Sclaroff, and Federico Perazzi. Temporally distributed networks + +for fast video semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8827, 2020. 1 +[29] Junhwa Hur and Stefan Roth. Joint optical flow and temporally consistent semantic segmentation. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part I 14, pages 163-177. Springer, 2016. 1 +[30] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 1, 3, 4 +[31] Samvit Jain, Xin Wang, and Joseph E Gonzalez. Accel: A corrective fusion network for efficient semantic segmentation on video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8866-8875, 2019. 1 +[32] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9865-9874, 2019. 5 +[33] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 13 +[34] Thomas Kipf, Gamaleldin F Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. arXiv preprint arXiv:2111.12594, 2021. 2, 8 +[35] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5, 7 +[36] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019. 3 +[37] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Video k-net: A simple, strong, and unified baseline for video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18847-18857, 2022. 1 +[38] Yule Li, Jianping Shi, and Dahua Lin. Low-latency video semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5997-6005, 2018. 1 +[39] Runtao Liu, Zhirong Wu, Stella Yu, and Stephen Lin. The emergence of objectness: Learning zero-shot segmentation from videos. Advances in Neural Information Processing Systems, 34:13137-13152, 2021. 2 +[40] Si Liu, Changhu Wang, Ruihe Qian, Han Yu, Renda Bao, and Yao Sun. Surveillance video parsing with single frame supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 413-421, 2017. 1 +[41] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame + +inference. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 352-368. Springer, 2020. 1 +[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 13 +[43] Xiankai Lu, Wenguan Wang, Jianbing Shen, Yu-Wing Tai, David J Crandall, and Steven CH Hoi. Learning video object segmentation from unlabeled videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8960-8970, 2020. 2 +[44] Yuval Nirkin, Lior Wolf, and Tal Hassner. Hyperseg: Patchwise hypernetwork for real-time semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4061-4070, 2021. 1 +[45] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4 +[46] Nikhil Parthasarathy, SM Eslami, João Carreira, and Olivier J Henaff. Self-supervised video pretraining yields strong image representations. arXiv preprint arXiv:2210.06433, 2022. 2 +[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 13 +[48] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5, 13 +[49] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael S Ryoo. Self-supervised video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2874-2884, 2022. 1 +[50] Sucheng Ren, Wenxi Liu, Yongtuo Liu, Haoxin Chen, Guoqiang Han, and Shengfeng He. Reciprocal transformations for unsupervised video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15455-15464, 2021. 2 +[51] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860, 2022. 1 +[52] Thomas Stegmüller, Tim Lebailly, Behzad Bozorgtabar, Tinne Tuytelaars, and Jean-Philippe Thiran. Croc: Cross-view online clustering for dense visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7000–7009, 2023. 8, 9 +[53] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE/CVF conference on + +computer vision and pattern recognition, pages 8934-8943, 2018. 15 +[54] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3126-3137, 2022. 1 +[55] Fida Mohammad Thoker, Hazel Doughty, Piyush Bagad, and Cees GM Snoek. How severe is benchmark-sensitivity in video self-supervised learning? In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 632-652. Springer, 2022. 2 +[56] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X, pages 268-285. Springer, 2020. 5 +[57] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10052-10062, 2021. 1, 2, 5, 8, 13 +[58] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Chuanxin Tang, Xiyang Dai, Yucheng Zhao, Yujia Xie, Lu Yuan, and Yu-Gang Jiang. Look before you match: Instance understanding matters in video object segmentation. arXiv preprint arXiv:2212.06826, 2022. 1 +[59] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. arXiv preprint arXiv:2301.11320, 2023. 2 +[60] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2566-2576, 2019. 3 +[61] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 2, 5, 8, 13 +[62] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 2 +[63] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10334–10343, 2019. 1 +[64] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 5, 13 +[65] Yu-Syuan Xu, Tsu-Jui Fu, Hsuan-Kung Yang, and Chun-Yi Lee. Dynamic video segmentation network. In Proceed- + +ings of the IEEE conference on computer vision and pattern recognition, pages 6556-6565, 2018. 1 +[66] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7177-7188, 2021. 2, 3, 15 +[67] Sukmin Yun, Hankook Lee, Jaehyung Kim, and Jinwoo Shin. Patch-level representation learning for self-supervised vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8354-8363, 2022. 2 +[68] Andrii Zadaianchuk, Matthaeus Kleindessner, Yi Zhu, Francesco Locatello, and Thomas Brox. Unsupervised semantic segmentation with self-supervised object-centric representations. arXiv preprint arXiv:2207.05027, 2022. 2 +[69] Shaofeng Zhang, Feng Zhu, Rui Zhao, and Junchi Yan. Patch-level contrasting without patch correspondence for accurate and dense contrastive representation learning. In The Eleventh International Conference on Learning Representations, 2023. 2 +[70] Xinyu Zhang and Abdeslam Boullarias. Optical flow boosts unsupervised localization and segmentation. arXiv preprint arXiv:2307.13640, 2023. 8 +[71] Yurong Zhang, Liulei Li, Wenguan Wang, Rong Xie, Li Song, and Wenjun Zhang. Boosting video object segmentation via space-time correspondence learning. arXiv preprint arXiv:2304.06211, 2023. 1, 5 +[72] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. ICLR, 2022. 3, 6 +[73] Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, and Yichen Wei. Deep feature flow for video recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2349-2358, 2017. 1 +[74] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022. 1, 2, 5, 7, 8, 9, 13, 14, 15, 16, 18 \ No newline at end of file diff --git a/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/images.zip b/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..957a122d24ee01d5a1900533a6da42b548246739 --- /dev/null +++ b/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b37eefb79ed6ecd94da6f96fb87bbb3e573b4d725328446de624a240b80e898b +size 446742 diff --git a/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/layout.json b/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..10a19d73ca331217a4ef158700a7377d37915394 --- /dev/null +++ b/2023/Time Does Tell_ Self-Supervised Time-Tuning of Dense Image Representations/layout.json @@ -0,0 +1,9554 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 538, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 538, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 538, + 122 + ], + "type": "text", + "content": "Time Does Tell: Self-Supervised Time-Tuning of Dense Image Representations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 143, + 485, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 143, + 485, + 171 + ], + "spans": [ + { + "bbox": [ + 108, + 143, + 485, + 171 + ], + "type": "text", + "content": "Mohammadreza Salehi, Efstratios Gavves, Cees G. M. Snoek, Yuki M. Asano \nQUVA Lab, University of Amsterdam" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 174, + 455, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 174, + 455, + 185 + ], + "spans": [ + { + "bbox": [ + 140, + 174, + 455, + 185 + ], + "type": "text", + "content": "(s-salehidehnavi, e.gavves, c.g.m.snoek, y.m.asano)@uva.nl" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "type": "text", + "content": "Spatially dense self-supervised learning is a rapidly growing problem domain with promising applications for unsupervised segmentation and pretraining for dense downstream tasks. Despite the abundance of temporal data in the form of videos, this information-rich source has been largely overlooked. Our paper aims to address this gap by proposing a novel approach that incorporates temporal consistency in dense self-supervised learning. While methods designed solely for images face difficulties in achieving even the same performance on videos, our method improves not only the representation quality for videos – but also images. Our approach, which we call time-tuning, starts from image-pretrained models and fine-tunes them with a novel self-supervised temporal-alignment clustering loss on unlabeled videos. This effectively facilitates the transfer of high-level information from videos to image representations. Time-tuning improves the state-of-the-art by " + }, + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "type": "inline_equation", + "content": "8 - 10\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "type": "text", + "content": " for unsupervised semantic segmentation on videos and matches it for images. We believe this method paves the way for further self-supervised scaling by leveraging the abundant availability of videos. The implementation can be found here: https://github.com/SMSD75/Timetuning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 689 + ], + "type": "text", + "content": "Dense self-supervised learning, whereby meaningful deep features for each pixel or patch of input are learned in an unsupervised manner, has recently received increasing attention [74, 24, 57, 51]. By learning spatially consistent features for different views of an input, strong gains in unsupervised semantic segmentation have been achieved using unlabeled images. However, so far, an even more information-rich source for unsupervised training has been largely overlooked: videos. With their additional time dimension and being the most rapidly growing form of digital content, they are well-suited to scaling dense self-supervised learning even further." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Some efforts have already been made to learn from the video domain by using different frames from a video as aug-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 215, + 541, + 361 + ], + "blocks": [ + { + "bbox": [ + 307, + 215, + 541, + 361 + ], + "lines": [ + { + "bbox": [ + 307, + 215, + 541, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 215, + 541, + 361 + ], + "type": "image", + "image_path": "6d6a21dd569aa53cf12c8bad541e03655ae4cf6be65c2e2b8b02480bded49a00.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 364, + 545, + 472 + ], + "lines": [ + { + "bbox": [ + 305, + 364, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 545, + 472 + ], + "type": "text", + "content": "Figure 1: Time-tuning compared to previous methods. Unlike existing methods that ignore or utilize expensive 3D models to implicitly model temporal information, the proposed method explicitly incorporates temporal consistency in dense feature representations using a temporal self-supervised loss. The method starts with a 2D encoder pretrained on images and fine-tunes it using unlabeled videos. This approach leads to improved performance not only for videos but also for images." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 496, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 652 + ], + "type": "text", + "content": "mentations [18, 63, 49] or by involving temporal correspondence [71, 58]; however, they mostly did it in a supervised way [29, 73, 19, 40, 65, 38, 31, 28, 41, 37, 54, 44], which is not scalable specifically for dense tasks where the number of targets can increase significantly as the number of pixels grows. To this end, self-supervised learning approaches offer a solution by reducing the need for supervision. However, these methods typically rely on the notion of \"views\", which involves learning similar features for corresponding locations over time. This usually leads to a chicken-and-egg problem, where the correspondences are required for learning dense features - which in turn enable good correspondences [30]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "In images, the challenge is trivially solved by considering the correspondence function based on the augmentation function. For instance, In the case of color augmentations, this correspondence is simply given by the identity. However, shifts in time cannot be viewed as mere augmenta" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "16536" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "tions. As we demonstrate through a new evaluation protocol, using image models on frames alone is not nearly as effective. A similar finding has also been reported, albeit for non-dense works [20, 34], which assumed the passage of time as an image augmentation. These works have generally reported reduced performances, even when compared to simple image-based pretraining methods. Similarly, video-level tasks [17, 18, 55] assume sufficiently similar semantics between different frames. This is also not true for dense tasks, as static features can only be assumed where nothing is moving – which is rare due to possible object, camera, and background motion between the frames." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 407 + ], + "type": "text", + "content": "To address this challenge, we propose to model the additional time-dimension explicitly to identify which pixels should retain similar embeddings and which should not. We propose two separate modules to tackle the correspondence and the dense learning, respectively. For the former, we introduce the Feature-Forwarder (FF) module, which breaks the mentioned chicken-and-egg loop by leveraging the good tracking performance of pretrained image models, and allows an approximate second \"view\" that can then be treated as a target for the further dense self-supervised loss. On top of this, we introduce a spatio-temporally dense clustering module, which learns unsupervised clusters across samples, locations and time. Using these two components and starting from image-pretrained features, our proposed method allows time-tuning (TIMET) the dense representation in a self-supervised manner, see Figure 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 408, + 289, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 289, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 289, + 503 + ], + "type": "text", + "content": "Finally, we demonstrate that TIMET paves the way for further scaling of self-supervised learning by leveraging the abundant availability of video datasets and transferring their knowledge to the image domain. This results in consistently achieving state-of-the-art performances not only for the task of unsupervised semantic segmentation of videos, but also for unsupervised image semantic segmentation, a feat previously out of reach for methods trained on videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 504, + 278, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 504, + 278, + 517 + ], + "spans": [ + { + "bbox": [ + 59, + 504, + 278, + 517 + ], + "type": "text", + "content": "Overall, this paper makes the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 526, + 288, + 703 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 526, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 526, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 59, + 526, + 287, + 575 + ], + "type": "text", + "content": "- We show that image-based unsupervised dense segmentation models applied to videos exhibit degraded performance and lack temporal consistency in their segmentation maps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 584, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 584, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 58, + 584, + 287, + 620 + ], + "type": "text", + "content": "- Building on this observation, we propose a novel dense self-supervised learning method that utilizes temporal consistency as a learning signal." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 630, + 288, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 630, + 288, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 630, + 288, + 703 + ], + "type": "text", + "content": "- We demonstrate that our method enables the scaling of self-supervised learning by leveraging abundant video datasets and effectively transferring knowledge to the image domain. Our approach consistently achieves state-of-the-art performance for both images and videos, opening up new opportunities in the field." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 415 + ], + "type": "text", + "content": "Dense self-supervised learning. These methods build upon image-level self-supervised representation learning by incorporating existing losses to enhance the spatial features, demonstrating a commendable advancement. DenseCL [61] works on spatial features by constructing dense correspondences across views using the contrastive objective given in MoCo [23], while PixPro [62] utilizes the augmentation wrapper to get the spatial correspondence of the pixel intersection between two views. Similarly, MaskContrast [57], Leopart [74], DetCon [24] and Odin [25] also ensure spatial feature similarities via contrastive learning and spatial correspondences. Self-Patch [67] treats the spatial neighbors of the patch as positive examples for learning more semantically meaningful relations among patches. Inspired by SelfPatch, ADCLR [69] proposes patch-level contrasting via query crop and cross-attention mechanism. Pursuing the same objective through none end-to-end approaches. Both [68] and [59] propose an unsupervised salient object segmentation pipeline that extracts noisy object masks from the inputs and fine-tunes a specifically designed object segmentation head with several self-training steps to make an unsupervised semantic segmentation model. Unlike the existing image-based approaches, we propose a method that improves the dense prediction performance of a pretrained encoder by explicitly modeling the temporal dimension and learning from diverse natural dynamics and variations found in videos." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 432, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 547, + 624 + ], + "type": "text", + "content": "Video to image knowledge transfer. Learning from videos instead of images has recently received attention since they contain far more information than still images and hold the potential for learning rich representations of the visual world. To this end, several non-dense works have been released [46, 4, 27]. VITO [46] shows that naively applying image domain self-supervised learning methods on videos can lead to a performance drop coming from the distribution shift, which by applying data processing techniques is relieved but not fully solved. Following the same way, [27] trains masked autoencoders [22] with contrastive learning [10] on video datasets and shows a decent performance on both image and video tasks. While video to image knowledge transfer is the end goal of such methods and our paper; our dense task is different and is as of yet unaddressed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Unsupervised video object segmentation. Unsupervised video object segmentation methods do not require any manual annotations. However, they are only designed to tackle a foreground/background segmentation task, which refers to the segmentation of the most prominent, general objects in video sequences [66, 43, 50, 1, 39]. For instance, [66]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "text", + "content": "16537" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 286, + 202 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 286, + 202 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 286, + 202 + ], + "type": "image", + "image_path": "28984550502a23d9386dd0e8146f9d52eb8369dfc1db154caa89a6f52ebad4ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 547, + 288 + ], + "type": "text", + "content": "Figure 2: Time-Tuning overview. We conduct self-supervised video semantic segmentation by tuning image-pretrained models with temporal information from videos. (a): Our general pipeline of adapting an image-pretrained ViT transformer on video data using our dense clustering loss. The encoder is specified by " + }, + { + "bbox": [ + 46, + 213, + 547, + 288 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 46, + 213, + 547, + 288 + ], + "type": "text", + "content": " and kept frozen except for the last two layers. Also, " + }, + { + "bbox": [ + 46, + 213, + 547, + 288 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 213, + 547, + 288 + ], + "type": "text", + "content": " shows the output of Sinkhorn-Knopp algorithm, which is forwarded to time-step " + }, + { + "bbox": [ + 46, + 213, + 547, + 288 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 213, + 547, + 288 + ], + "type": "text", + "content": " to be compared with the features that are obtained using the last time-step. (b): Detailed view into our Feature-Forwarder module that is used for aligning cluster-maps from past frames to logit-features of future ones." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 288, + 70, + 533, + 204 + ], + "blocks": [ + { + "bbox": [ + 288, + 70, + 533, + 204 + ], + "lines": [ + { + "bbox": [ + 288, + 70, + 533, + 204 + ], + "spans": [ + { + "bbox": [ + 288, + 70, + 533, + 204 + ], + "type": "image", + "image_path": "fc73b7e624c145655f8448ce3e80b31cca95cd290dd4ee16fc8e3c875e621a2b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 307, + 287, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 414 + ], + "type": "text", + "content": "uses an AutoEncoder-based architecture based on slot attention so that pixels that show the same motion flows based on optical-flow are grouped together. In this way, dominant moving objects can be detected and separated from the background. In our experiments, we benchmark two representative instances of this line of work, i.e., [66, 1], to provide better insights into the differences of unsupervised object detection methods and our proposed benchmark for video semantic segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "text", + "content": "To conclude, we propose the first self-supervised video semantic segmentation method, which tunes image-trained models such as [9, 72] based on the temporal information that exists in unlabeled videos to enhance their effectiveness in dense prediction tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 486, + 104, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 104, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 104, + 498 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 506, + 288, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 288, + 625 + ], + "type": "text", + "content": "Our method works by densely clustering features in a manner that is consistent with time. At a high level, it works by transforming features from a past time " + }, + { + "bbox": [ + 46, + 506, + 288, + 625 + ], + "type": "inline_equation", + "content": "t' = t" + }, + { + "bbox": [ + 46, + 506, + 288, + 625 + ], + "type": "text", + "content": " to the current frame " + }, + { + "bbox": [ + 46, + 506, + 288, + 625 + ], + "type": "inline_equation", + "content": "t' = T" + }, + { + "bbox": [ + 46, + 506, + 288, + 625 + ], + "type": "text", + "content": " and forcing these to be consistent with the currently observed ones at " + }, + { + "bbox": [ + 46, + 506, + 288, + 625 + ], + "type": "inline_equation", + "content": "t' = T" + }, + { + "bbox": [ + 46, + 506, + 288, + 625 + ], + "type": "text", + "content": ". The overall learning signal comes from the dense temporal clustering of forwarded features and allows \"timetuning\" image-pretrained models to learn from videos. Figure 2 shows the proposed method's architecture. In the following, we describe each component in detail." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 635, + 165, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 165, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 165, + 647 + ], + "type": "text", + "content": "3.1. Feature Forwarding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "The goal of this component, the feature-forwarder (FF), is to propagate the dense features in time as accurately as possible, given some RGB input frames. More formally, let " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " be the feature forwarding function, " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " be a frame at time " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " for a given video, and " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " the features corresponding to " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 307, + 508, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 307, + 508, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 508, + 319 + ], + "type": "text", + "content": "then the task is to propagate " + }, + { + "bbox": [ + 306, + 307, + 508, + 319 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 306, + 307, + 508, + 319 + ], + "type": "text", + "content": " to a future time " + }, + { + "bbox": [ + 306, + 307, + 508, + 319 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 307, + 508, + 319 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 373, + 328, + 545, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 328, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 373, + 328, + 545, + 342 + ], + "type": "interline_equation", + "content": "f \\left(\\left[ I _ {t}, \\dots , I _ {T} \\right], z _ {t}\\right)\\rightarrow \\hat {z} _ {T}. \\tag {1}", + "image_path": "5e5dafc92c03705251fda97d44ebad265e2a78c7e4dd01c2ad88db607535a211.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "inline_equation", + "content": "\\hat{z}_T" + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "text", + "content": " are not necessarily the same for " + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "inline_equation", + "content": "T > t" + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "text", + "content": "; since " + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "inline_equation", + "content": "\\hat{z}_T" + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "text", + "content": " represents an approximation of " + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "text", + "content": " using the previous features. Obviously, the further " + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "text", + "content": " is from " + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 351, + 547, + 435 + ], + "type": "text", + "content": ", the more variance is seen and therefore a better learning signal might be provided; however, for such larger intervals, the idea of feature forwarding encounters two main challenges: error accumulation and object non-permanence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 450, + 547, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 522 + ], + "type": "text", + "content": "Challenge 1: Error accumulation. While computing how features have changed simply based on the start and end frames is the most straightforward (i.e., " + }, + { + "bbox": [ + 304, + 450, + 547, + 522 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 450, + 547, + 522 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 450, + 547, + 522 + ], + "type": "inline_equation", + "content": "I_{T}" + }, + { + "bbox": [ + 304, + 450, + 547, + 522 + ], + "type": "text", + "content": "), it does not fully leverage the knowledge contained in the intermediate frames. Instead, it is common [30, 60, 36] to forward features for every time-step " + }, + { + "bbox": [ + 304, + 450, + 547, + 522 + ], + "type": "inline_equation", + "content": "\\delta t" + }, + { + "bbox": [ + 304, + 450, + 547, + 522 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 530, + 545, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 530, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 321, + 530, + 545, + 546 + ], + "type": "interline_equation", + "content": "f ([ I _ {t},.., I _ {T} ], z _ {t}) = \\bigcirc_ {t ^ {\\prime} = t + \\delta t} ^ {T} f ([ I _ {t ^ {\\prime} - \\delta t}, I _ {t ^ {\\prime}} ], z _ {t} ]), \\tag {2}", + "image_path": "7435b41ddbb617125f986bd39e89d75c59f649309f0ea4fc209a93fd6270de50.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 554, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 547, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 554, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 304, + 554, + 547, + 590 + ], + "type": "text", + "content": " is a composition operator such as element-wise multiplication. However, this results in the accumulation of errors over time." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "Challenge 2: Object non-permanence. One difficulty that arises when using videos as training data is the fact that objects sometimes simply disappear from the screen. Compared to images, which carry a heavy photographer's bias, videos are more prone to natural variation due to camera and object motion resulting in temporary object non-permanence in a particular video clip. A method that simply assumes an object (or a feature) has to be present in any given frame simply because it was present in the previous" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "16538" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "one might therefore easily fail when going towards videos that exist in-the-wild. A case in point are simple occlusions that arise from an object being fully- or partially-covered by another one further in the background. Taking longer training clips has been attempted to resolve the issue. However, this increases the training complexity, especially for dense-tasks that necessitate a higher number of predictions for each input." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 178, + 289, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 178, + 289, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 289, + 357 + ], + "type": "text", + "content": "We address these challenges using a novel, yet simple component, which we call the Stabilizing Feature-Forwarder. The first choice in designing this module is the function and data used for composing information across time. For instance, a simple approach would be to use pixel-wise optical flow and aggregate the flow at the feature level to predict how the features change over time. However, as we also show in our experiments, optical flow is prone to accumulating error with time and also cannot easily recover from occlusions [6]. Instead, we utilize the fact that we wish to pretrain the visual encoder and recycle part of it for the purpose of feature forwarding. This not only speeds up the process, as no additional encoder/modality is required, but also produces a positive feedback loop of better forwarding and better feature learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 357, + 288, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 288, + 393 + ], + "type": "text", + "content": "Concretely, we first compute the L2-normalised dense features of a pretrained visual encoder " + }, + { + "bbox": [ + 47, + 357, + 288, + 393 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 47, + 357, + 288, + 393 + ], + "type": "text", + "content": " to compute spatial similarities across time (the ' in " + }, + { + "bbox": [ + 47, + 357, + 288, + 393 + ], + "type": "inline_equation", + "content": "t^\\prime" + }, + { + "bbox": [ + 47, + 357, + 288, + 393 + ], + "type": "text", + "content": " left out for clarity):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 398, + 287, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 398, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 111, + 398, + 287, + 412 + ], + "type": "interline_equation", + "content": "F _ {i j} ^ {t} = \\left\\langle \\Phi \\left(I _ {t}\\right) _ {i}, \\Phi \\left(I _ {T}\\right) _ {j} \\right\\rangle / \\tau , \\tag {3}", + "image_path": "8a7946a7f4c530f9096b693465f6dbbfe798e59951e92bf61a46b1fe641b70fa.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 417, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 417, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 417, + 287, + 502 + ], + "type": "text", + "content": "which yields matrices " + }, + { + "bbox": [ + 46, + 417, + 287, + 502 + ], + "type": "inline_equation", + "content": "F^{\\{t,\\dots T - \\delta t\\}} \\in R^{N\\times N}" + }, + { + "bbox": [ + 46, + 417, + 287, + 502 + ], + "type": "text", + "content": " with values between [0,1] indicating semantic similarities between the " + }, + { + "bbox": [ + 46, + 417, + 287, + 502 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 417, + 287, + 502 + ], + "type": "text", + "content": " spatial features that are sharpened by a temperature " + }, + { + "bbox": [ + 46, + 417, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 417, + 287, + 502 + ], + "type": "text", + "content": ". Note that because we compare each frame with the final frame, the effect of object non-permanence can be minimized: even if the object is only present in every other frame, there is enough signal to forward its features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 502, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 287, + 526 + ], + "type": "text", + "content": "Next to propagate from time " + }, + { + "bbox": [ + 47, + 502, + 287, + 526 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 502, + 287, + 526 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 502, + 287, + 526 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 502, + 287, + 526 + ], + "type": "text", + "content": ", these similarities are stacked and normalised along time, as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 540, + 287, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 540, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 85, + 540, + 287, + 566 + ], + "type": "interline_equation", + "content": "\\tilde {F} _ {i j} ^ {t} = \\exp \\left(\\mathcal {N} \\left(F _ {i j} ^ {t}\\right)\\right) / \\sum_ {i, t ^ {\\prime}} \\exp \\left(\\mathcal {N} \\left(F _ {i j} ^ {t ^ {\\prime}}\\right)\\right), \\tag {4}", + "image_path": "1015ce61730d78f6de21941cbbcd982ed6643902ec43581c6dcfc5e002f315c8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "inline_equation", + "content": "t \\leq t' \\leq T" + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "text", + "content": " is a neighborhood thresholding method which forces short-term spatial-smoothness, i.e., " + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\tilde{F}_{ij}^t) = 0" + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "text", + "content": " is not within a local window around " + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "text", + "content": " with the size " + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "text", + "content": ". Finally, the propagated feature are computed " + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "inline_equation", + "content": "\\hat{z}_T" + }, + { + "bbox": [ + 47, + 571, + 287, + 632 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 647, + 287, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 647, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 85, + 647, + 287, + 672 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {T} (j) = \\sum_ {t ^ {\\prime}, i} \\tilde {F} _ {i j} ^ {t ^ {\\prime}} \\hat {z} _ {t ^ {\\prime}} (i), \\quad j \\in \\{1,.., N \\}. \\tag {5}", + "image_path": "45abb494fd5d5909663de928c47b8c213a68b7796cc07e680babc786c6b3d8ca.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "This means that for arriving at the target feature " + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{z}_T" + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": ", we not only use the previous frame as the source, but instead use the past " + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "(T - t) / \\delta t" + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": " frames and aggregate these." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 277 + ], + "type": "text", + "content": "Relation to mask-propagation methods. While previous methods such as DINO [9] and STC [30] have utilized a similar technique for propagating ground-truth masks for evaluating, for example on the DAVIS dataset, there are three key differences to our forwarding method. First, instead of propagating binary maps of foreground-vs-background, our feature-forwarder has as inputs soft, noisy and multi-label self-supervised segmentation maps, which require the forwarder to tolerate overlapping of different object probabilities throughout training. Second, previous methods have used this approach mainly for inference; we, however, use this module as a trainable component and show how our loss improves this forwarder with training time. Finally, we do not follow a typical re-normalizing step across feature dimensions (typically done after Eq. (5)) as this harms the scale of logits that are being propagated and leads to heavily diluted target distributions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 283, + 479, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 283, + 479, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 479, + 296 + ], + "type": "text", + "content": "3.2. Self-supervised dense clustering" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 301, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 547, + 422 + ], + "type": "text", + "content": "While the Feature-Forwarder produces target features that include information about the dynamics with time, its computation utilizes " + }, + { + "bbox": [ + 304, + 301, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\Phi(I_T)" + }, + { + "bbox": [ + 304, + 301, + 547, + 422 + ], + "type": "text", + "content": " and could lead to trivial solutions, i.e., " + }, + { + "bbox": [ + 304, + 301, + 547, + 422 + ], + "type": "inline_equation", + "content": "f([I_t,..,I_T],z_t) = \\Phi(I_T)" + }, + { + "bbox": [ + 304, + 301, + 547, + 422 + ], + "type": "text", + "content": ". To counteract this, we propose a self-supervised clustering task across views in time. For this, we utilize the basic online clustering algorithm based optimal-transport [12], utilized in works such as SeLa [2], SwAV [8] and DINOv2 [45]. In particular, let " + }, + { + "bbox": [ + 304, + 301, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 301, + 547, + 422 + ], + "type": "text", + "content": " be a visual encoder with a clustering head " + }, + { + "bbox": [ + 304, + 301, + 547, + 422 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 301, + 547, + 422 + ], + "type": "text", + "content": " that yields a " + }, + { + "bbox": [ + 304, + 301, + 547, + 422 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 301, + 547, + 422 + ], + "type": "text", + "content": " dimensional output, then the clustering loss is given by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 366, + 430, + 545, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 430, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 366, + 430, + 545, + 444 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(x _ {i}\\right) = - \\tilde {y} _ {i} \\log (g (\\Psi (x) _ {i})), \\tag {6}", + "image_path": "d212a9375ab84c2ad31762967a7c750ecb8af844591fce0fb01fedc4e43af837.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 453, + 547, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 547, + 502 + ], + "type": "text", + "content": "which is a standard cross-entropy loss with regards to self-supervised pseudo-label " + }, + { + "bbox": [ + 304, + 453, + 547, + 502 + ], + "type": "inline_equation", + "content": "\\tilde{y}_i" + }, + { + "bbox": [ + 304, + 453, + 547, + 502 + ], + "type": "text", + "content": ". These labels, in turn, are generated by solving an entropy-regularised optimal transport problem on the batch " + }, + { + "bbox": [ + 304, + 453, + 547, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 453, + 547, + 502 + ], + "type": "text", + "content": " [2]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 509, + 545, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 509, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 341, + 509, + 545, + 533 + ], + "type": "interline_equation", + "content": "\\min _ {\\tilde {y}} \\langle \\tilde {y}, - \\log g (\\Psi (x)) \\rangle + \\frac {1}{\\lambda} \\operatorname {K L} (\\tilde {y} \\| r c ^ {\\top}), \\tag {7}", + "image_path": "c3f1c9ddd0c7c2bf86c80361717d5ad76185c825804ca09ecf6781421493886f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 379, + 534, + 545, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 534, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 379, + 534, + 545, + 561 + ], + "type": "interline_equation", + "content": "\\text {w i t h} r = \\frac {1}{K} \\cdot \\mathbb {1}, \\quad c = \\frac {1}{| \\mathcal {B} |} \\cdot \\mathbb {1}. \\tag {8}", + "image_path": "aeb28d406cc8fa9afbb0081f7eed9775ea751cea58deac9f4ebbc164c03bbc0d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": " controls the entropy regularisation and " + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "r, c" + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": " are the marginals for the prototypes and the batch, respectively. Note that solving this problem can be done extremely quickly on the GPU and yields soft pseudo-labels " + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\tilde{Y}" + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\operatorname{argmax}(\\tilde{Y}) = \\tilde{y}" + }, + { + "bbox": [ + 304, + 567, + 545, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 628, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 664 + ], + "type": "text", + "content": "n, these might yield sufficiently similar semantics [17] this is not true for dense tasks, as static features can only be assumed for videos where nothing is moving, which is rare." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 671, + 416, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 416, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 416, + 682 + ], + "type": "text", + "content": "3.3. Overall TIMET loss" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "We combine the previous two modules to arrive at our full method (as shown in Fig. 2): First, self-supervised" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "16539" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "Sinkhorn-Knopp clustering is conducted on early features " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "g(\\Psi(I_t))" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": ", yielding soft pseudo-labels " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{SK}(g(\\Psi(I_t))) = \\tilde{Y}_t" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": ". These are then forwarded in time using our Feature-Forwarder to arrive at dense targets " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}(\\tilde{Y}_t)" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": ", which are used in the final loss. Compactly:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 144, + 287, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 144, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 76, + 144, + 287, + 170 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T I M E T}} \\left(I _ {T}\\right) = - \\sum_ {i, j} \\operatorname {F F} \\left(\\tilde {Y} _ {t}\\right) \\log \\left(g \\left(\\Psi \\left(I _ {T}\\right)\\right)\\right). \\tag {9}", + "image_path": "192999b28a742ae16e55f178cd6d6417c21c2354d259ffa0f81202b2870e5022.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 183, + 128, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 183, + 128, + 197 + ], + "spans": [ + { + "bbox": [ + 47, + 183, + 128, + 197 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 204, + 97, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 204, + 97, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 97, + 217 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 224, + 288, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 288, + 404 + ], + "type": "text", + "content": "Datasets. We train our method and baselines on YTVOS [64], one of the largest video segmentation datasets available, and evaluate on DAVIS17 [48] and YTVOS. For YTVOS the ground truth masks are only available for the first frames of the test and validation sets, and therefore, a fixed random " + }, + { + "bbox": [ + 46, + 224, + 288, + 404 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 224, + 288, + 404 + ], + "type": "text", + "content": " of the training set is used for testing. For transfer learning experiments, we use the validation set of Pascal VOC 2012 [16], As the dataset has been commonly used as a main reference for recent works in dense self-supervised image segmentation [74, 57, 61]. For completeness, we also report the performance on egocentric datasets that have less object-centric bias and are prevalent in real-world scenarios. For egocentric experiments, we train on EPIC-KITCHENS-100 [13] and evaluate on VISOR [14]. Further details are provided in Appendix A." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 714 + ], + "type": "text", + "content": "Models and baselines. Currently, there is a lack of available unsupervised semantic video semantic segmentation methods. Nevertheless, we have included a comprehensive comparison of our method with state-of-the-art techniques in both image and video domains for unsupervised image semantic segmentation and unsupervised video object segmentation. To evaluate the image-based models, we utilized either official reported numbers or provided pretrained models of STEGO [21] and Leopart [74]. We have taken every measure to ensure fairness in our comparison. To do so, all the used pretrained models have the same pretraining dataset (ImageNet-1k), and the number of their backbone parameters is roughly similar. For those models that use extra datasets, for instance, Leopart, we select the pretrained backbones that closely match the specification of YTVOS training dataset. Additionally, we trained image-based models on the same video datasets, where we converted video data into their corresponding image data and reported their performance wherever possible. To ensure comprehensive analysis, we have included the recent concurrent work Flowdino [71], which utilizes optical flow to refine DINO features on unlabeled videos, in our comparison benchmarks as well. This is done only in cases where there is a shared experimental setup." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 445 + ], + "type": "text", + "content": "Evaluation procedure. Despite unsupervised object segmentation being a well-established evaluation in the image domain [57, 74], evaluating unsupervised video multi-label object segmentation is challenging due to the absence of an established evaluation protocol for video object semantic segmentation without supervision. In this regard, we propose a set of evaluation protocols for unsupervised video multi-label object segmentation, which exploits existing video object segmentation datasets for evaluation purposes (see details in Appendix B). In our experiments, we discard any projection head used during training and evaluate ViT's spatial tokens directly, similar to [57, 32, 61], using four methods: classification with a linear head, classification with an FCN head, overclustering, and clustering with as many clusters as the number of ground truth objects. To fine-tune different heads on top of the frozen spatial tokens, we follow [57]. For unsupervised semantic segmentation, we apply " + }, + { + "bbox": [ + 304, + 72, + 547, + 445 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 547, + 445 + ], + "type": "text", + "content": "-Means on all spatial tokens, where " + }, + { + "bbox": [ + 304, + 72, + 547, + 445 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 547, + 445 + ], + "type": "text", + "content": " is chosen to be higher or equal to the ground-truth number of classes, following the common protocol in image clustering [32, 56]. For grouping clusters to ground-truth classes, we match them either by pixel-wise precision or Hungarian matching on merged cluster maps [35]. For video datasets, we allow the matching to be per-frame, per-video, and per-dataset. The evaluation metric is permutation-invariant, following [32], and the results are averaged over five different seeds. Clustering evaluations are preferred as they require less supervision and fewer hyperparameters than training a linear classifier and operate directly on the learned embedding spaces. We report our results in mean Intersection over Union (mIoU) unless otherwise specified." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 457, + 547, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 547, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 547, + 531 + ], + "type": "text", + "content": "Model training. We train a ViT-Small with patch size 16 and initialized from ImageNet-pretrained DINO weights [9] using the proposed self-supervised loss function. For the ablations and the main experiments, our models are trained for 12 and 30 epochs respectively. Further training details are provided in Appendix A. Code will be made available." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 536, + 373, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 373, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 373, + 548 + ], + "type": "text", + "content": "4.2. Ablations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 555, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 547, + 616 + ], + "type": "text", + "content": "We first examine the essential parameters of our method by training TIMET on YTVOS and assessing its ability to perform semantic segmentation on Pascal VOC. In addition to presenting the results through clustering and overclustering, we also demonstrate linear classification outcomes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Number of prototypes. We first ablate the influence of the number of prototypes on downstream segmentation performance. Results in Table 1a indicate a sharp increase in performance with a rise in the number of prototypes, but once a sufficiently large number is reached, performance stabilizes. We observe peak performance with a moderate number of 200 prototypes. The stability of our method over" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "text", + "content": "16540" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 88, + 179, + 157 + ], + "blocks": [ + { + "bbox": [ + 70, + 76, + 176, + 88 + ], + "lines": [ + { + "bbox": [ + 70, + 76, + 176, + 88 + ], + "spans": [ + { + "bbox": [ + 70, + 76, + 176, + 88 + ], + "type": "text", + "content": "(a) Ablating # prototypes " + }, + { + "bbox": [ + 70, + 76, + 176, + 88 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 88, + 179, + 157 + ], + "lines": [ + { + "bbox": [ + 69, + 88, + 179, + 157 + ], + "spans": [ + { + "bbox": [ + 69, + 88, + 179, + 157 + ], + "type": "table", + "html": "
PLCK=21K=500
1037.56.124.1
5058.28.240.5
20059.79.242.8
30059.49.042.3
", + "image_path": "2c273b3c3448037f728c81648f31d6d37bd9d7f8eaac1f43ac9157f1e9a86e0e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 221, + 89, + 383, + 156 + ], + "blocks": [ + { + "bbox": [ + 242, + 77, + 351, + 88 + ], + "lines": [ + { + "bbox": [ + 242, + 77, + 351, + 88 + ], + "spans": [ + { + "bbox": [ + 242, + 77, + 351, + 88 + ], + "type": "text", + "content": "(b) Ablating time interval " + }, + { + "bbox": [ + 242, + 77, + 351, + 88 + ], + "type": "inline_equation", + "content": "\\delta T" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 221, + 89, + 383, + 156 + ], + "lines": [ + { + "bbox": [ + 221, + 89, + 383, + 156 + ], + "spans": [ + { + "bbox": [ + 221, + 89, + 383, + 156 + ], + "type": "table", + "html": "
#framesδTLCK=21K=500
10s50.75.626.2
40.2s56.57.436.5
40.5s59.79.242.8
41.0s57.18.338.1
", + "image_path": "c089f25c179b8536a01c8198f91c144c25f0c1b8e554ca862f0b50cfea02b71c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 394, + 89, + 545, + 156 + ], + "blocks": [ + { + "bbox": [ + 404, + 77, + 536, + 88 + ], + "lines": [ + { + "bbox": [ + 404, + 77, + 536, + 88 + ], + "spans": [ + { + "bbox": [ + 404, + 77, + 536, + 88 + ], + "type": "text", + "content": "(c) Ablating number of frames used." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 394, + 89, + 545, + 156 + ], + "lines": [ + { + "bbox": [ + 394, + 89, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 394, + 89, + 545, + 156 + ], + "type": "table", + "html": "
#framesTLCK=21K=500
10s50.75.626.2
22.0s52.66.237.1
42.0s59.79.242.8
82.0s59.79.042.3
", + "image_path": "9e5d6606ba88abed857c01df1dc919420cbbfac6f8c0153d4decd4c125da6648.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 234, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 287, + 258 + ], + "type": "text", + "content": "a range of prototypes (50-300) suggests that our approach is robust to the change of this parameter." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 276, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 408 + ], + "type": "text", + "content": "Understanding FF. Next, we ablate the temporal dimensions that influence the working of the Feature-Forwarder. In Table 1b, we vary the time interval " + }, + { + "bbox": [ + 46, + 276, + 287, + 408 + ], + "type": "inline_equation", + "content": "\\delta T" + }, + { + "bbox": [ + 46, + 276, + 287, + 408 + ], + "type": "text", + "content": " between frames whilst keeping a fixed number of four frames. We generally find an increase in performance when increasing the time-interval, with performance peaking at 0.5s. One of the most critical observations is the clear difference in clustering performance from " + }, + { + "bbox": [ + 46, + 276, + 287, + 408 + ], + "type": "inline_equation", + "content": "26.2\\%" + }, + { + "bbox": [ + 46, + 276, + 287, + 408 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 276, + 287, + 408 + ], + "type": "inline_equation", + "content": "42.8\\%" + }, + { + "bbox": [ + 46, + 276, + 287, + 408 + ], + "type": "text", + "content": " between training on single frames (first row) vs. training with multiple frames. This clearly indicates the significant benefit of utilizing temporal information from the Feature-Forwarder during training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 408, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 480 + ], + "type": "text", + "content": "Next, in Table 1c we vary the number of frames given a fixed clip duration of 2s. Again, we find that a moderate number of frames between the source and target time is most favorable. While additional frames do not degrade the performance much, they do add computational complexity, so we prefer using 4 frames in our main method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": "Choice of propagation feature. Finally, in Table 2, we vary the type of feature that we forward to a future time. In the first row, we report the performance when not using any feature forwarding – thus solely training on single-frame inputs (i.e., the same as row 1 in Table 1c). In the second row, \"Identity\" shows the performance that is obtained when the feature map from the source frame is simply forwarded to the future without any changes, which shows an increase in the performance compared to the first row. This shows that often, training videos are not very dynamic, and a static assumption can already lead to some gains. However, compared to forwarding the Sinkhorn-Knopp (SK) clustered features, these gains are small (+4.6% vs. +17%). Importantly, our Feature-Forwarder relies on forwarding SK-sharpened feature maps and does not work when simply forwarding the network's output logits " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Phi(x)" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": ", followed by a clustering step. The reason for obtaining considerably lower numbers here can be traced back to using an un-entropy" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 345, + 232, + 509, + 300 + ], + "blocks": [ + { + "bbox": [ + 46, + 164, + 547, + 214 + ], + "lines": [ + { + "bbox": [ + 46, + 164, + 547, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 164, + 547, + 214 + ], + "type": "text", + "content": "Table 1: Ablations of the key parameters of our method. The model is trained for 12 epochs on Pascal VOC, and results for unsupervised segmentation with clustering " + }, + { + "bbox": [ + 46, + 164, + 547, + 214 + ], + "type": "inline_equation", + "content": "(K = 21)" + }, + { + "bbox": [ + 46, + 164, + 547, + 214 + ], + "type": "text", + "content": ", overclustering " + }, + { + "bbox": [ + 46, + 164, + 547, + 214 + ], + "type": "inline_equation", + "content": "(K = 500)" + }, + { + "bbox": [ + 46, + 164, + 547, + 214 + ], + "type": "text", + "content": ", and linear pixel-wise classification (LC) are shown. The stability of our method over a range of prototypes (50-300), inter-frame time intervals " + }, + { + "bbox": [ + 46, + 164, + 547, + 214 + ], + "type": "inline_equation", + "content": "(\\delta T \\in [0.5s - 1.0s])" + }, + { + "bbox": [ + 46, + 164, + 547, + 214 + ], + "type": "text", + "content": ", and the number of training frames (4-8) at a fixed clip duration " + }, + { + "bbox": [ + 46, + 164, + 547, + 214 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 46, + 164, + 547, + 214 + ], + "type": "text", + "content": " shows the robustness of the method." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 232, + 509, + 300 + ], + "lines": [ + { + "bbox": [ + 345, + 232, + 509, + 300 + ], + "spans": [ + { + "bbox": [ + 345, + 232, + 509, + 300 + ], + "type": "table", + "html": "
FF typeLCK=21K=500
None50.75.626.2
Identity55.37.436.1
Ψ(x)53.16.635.5
SK(Ψ(x))59.79.242.8
", + "image_path": "31ae8e1ac86b230dfabf37b28584290f01d65f60b1085c048a0ff06f165cd623.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 310, + 386, + 544, + 466 + ], + "blocks": [ + { + "bbox": [ + 304, + 308, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 369 + ], + "type": "text", + "content": "Table 2: Propagating different features in FF on Pascal VOC. We find that our Sinkhorn-Knopp (SK) based module outperforms static training ('None' or 'Identity'). Moreover, propagating logits " + }, + { + "bbox": [ + 304, + 308, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\Psi(x)" + }, + { + "bbox": [ + 304, + 308, + 545, + 369 + ], + "type": "text", + "content": " does not work as well as SK-regularised features." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 386, + 544, + 466 + ], + "lines": [ + { + "bbox": [ + 310, + 386, + 544, + 466 + ], + "spans": [ + { + "bbox": [ + 310, + 386, + 544, + 466 + ], + "type": "table", + "html": "
PretrainBackboneAt Init+TIMET
K=500LCK=500LC
MSN [3]ViT-S/1626.055.448.3↑22.367.2↑11.8
iBOT [72]ViT-S/1632.162.147.1↑15.067.1↑5.0
DINO [9]ViT-S/822.555.853.9↑31.469.5↑13.7
DINO [9]ViT-B/1628.959.152.7↑23.870.2↑11.1
", + "image_path": "22f5c6a4515553f90f64b74380e5e3ee582b3be313b7f5544b899dac2ebb3297.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 473, + 545, + 522 + ], + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 522 + ], + "type": "text", + "content": "Table 3: Applying TIMET to different pretrainings on Pascal VOC. TIMET can boost " + }, + { + "bbox": [ + 304, + 473, + 545, + 522 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 304, + 473, + 545, + 522 + ], + "type": "text", + "content": " the performance of different backbones with different initialization by a considerable margin, showing the generality of our approach." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 545, + 609 + ], + "type": "text", + "content": "regularized clustering algorithm. In this case, the logits tend to cause a few prototypes to dominate the cluster centers, exacerbated after the propagation, resulting in highly noisy and uninformative propagated logits. This shows that careful design of the Feature-Forwarder is indeed required." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Applying TIMET to different models. As shown in Table 3, our method is generalisable to different backbones and self-supervised learning initializations, enabling it to be an effective method to transfer the knowledge of the videos to images in an unsupervised way, which reduces the cost of labeling for different downstream tasks." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "16541" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 78, + 105, + 135 + ], + "blocks": [ + { + "bbox": [ + 49, + 78, + 105, + 135 + ], + "lines": [ + { + "bbox": [ + 49, + 78, + 105, + 135 + ], + "spans": [ + { + "bbox": [ + 49, + 78, + 105, + 135 + ], + "type": "image", + "image_path": "e2bf554d9064aafcabf706009fc36cda2f1000c3051dd8771930d0a7cc3b5ea4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 137, + 125, + 148 + ], + "lines": [ + { + "bbox": [ + 94, + 137, + 125, + 148 + ], + "spans": [ + { + "bbox": [ + 94, + 137, + 125, + 148 + ], + "type": "text", + "content": "Pascal" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 106, + 78, + 162, + 135 + ], + "blocks": [ + { + "bbox": [ + 106, + 78, + 162, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 162, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 162, + 135 + ], + "type": "image", + "image_path": "87d163b4a5d1599b1025b5b3114a31d5585e00ce5c7897190285587edc754ef0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 165, + 547, + 202 + ], + "lines": [ + { + "bbox": [ + 46, + 165, + 547, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 165, + 547, + 202 + ], + "type": "text", + "content": "Figure 3: TIMET unsupervised segmentations. As we regularize DINO's backbone to be consistent across time on YTVOS, it obtains strong performance on both image and video segmentation datasets, yielding high class consistencies (indicated by the segmentation colors) and tight borders. We provide more qualitative results in Appendix D." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 78, + 231, + 135 + ], + "blocks": [ + { + "bbox": [ + 175, + 78, + 231, + 135 + ], + "lines": [ + { + "bbox": [ + 175, + 78, + 231, + 135 + ], + "spans": [ + { + "bbox": [ + 175, + 78, + 231, + 135 + ], + "type": "image", + "image_path": "a25781e63c584f4090988165ab0be9a3e4d80f26c37e766cdf2efdfcb782b9aa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 137, + 249, + 148 + ], + "lines": [ + { + "bbox": [ + 221, + 137, + 249, + 148 + ], + "spans": [ + { + "bbox": [ + 221, + 137, + 249, + 148 + ], + "type": "text", + "content": "DAVIS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 232, + 78, + 288, + 135 + ], + "blocks": [ + { + "bbox": [ + 232, + 78, + 288, + 135 + ], + "lines": [ + { + "bbox": [ + 232, + 78, + 288, + 135 + ], + "spans": [ + { + "bbox": [ + 232, + 78, + 288, + 135 + ], + "type": "image", + "image_path": "9a9fb08e3310fa4ecc3051380b0d2343b339d9bce864880c6e3d1b075cb98613.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 78, + 360, + 135 + ], + "blocks": [ + { + "bbox": [ + 304, + 78, + 360, + 135 + ], + "lines": [ + { + "bbox": [ + 304, + 78, + 360, + 135 + ], + "spans": [ + { + "bbox": [ + 304, + 78, + 360, + 135 + ], + "type": "image", + "image_path": "1fdbc1e8b8ec9f21d7cc2d4f2c49bd9bf71a4563ab27d404be004d8f82587bc1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 137, + 375, + 148 + ], + "lines": [ + { + "bbox": [ + 345, + 137, + 375, + 148 + ], + "spans": [ + { + "bbox": [ + 345, + 137, + 375, + 148 + ], + "type": "text", + "content": "YTVOS" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 361, + 78, + 417, + 135 + ], + "blocks": [ + { + "bbox": [ + 361, + 78, + 417, + 135 + ], + "lines": [ + { + "bbox": [ + 361, + 78, + 417, + 135 + ], + "spans": [ + { + "bbox": [ + 361, + 78, + 417, + 135 + ], + "type": "image", + "image_path": "384f2e894c87e57b41bbd3aced56eb59e5711d58807fa4efdcf74d44e6cb137e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 433, + 78, + 490, + 135 + ], + "blocks": [ + { + "bbox": [ + 433, + 78, + 490, + 135 + ], + "lines": [ + { + "bbox": [ + 433, + 78, + 490, + 135 + ], + "spans": [ + { + "bbox": [ + 433, + 78, + 490, + 135 + ], + "type": "image", + "image_path": "b30519f2e15c0745bb711cec09f0c67d48d038dfb6f34eeebb7112227728083c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 137, + 506, + 148 + ], + "lines": [ + { + "bbox": [ + 478, + 137, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 478, + 137, + 506, + 148 + ], + "type": "text", + "content": "VISOR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 492, + 78, + 545, + 135 + ], + "blocks": [ + { + "bbox": [ + 492, + 78, + 545, + 135 + ], + "lines": [ + { + "bbox": [ + 492, + 78, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 492, + 78, + 545, + 135 + ], + "type": "image", + "image_path": "0d172b6c77e28e038a2a8e85d0f7a0e96b0476e5fd2e22447bfd052647fd3e10.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 58, + 212, + 537, + 388 + ], + "blocks": [ + { + "bbox": [ + 58, + 212, + 537, + 388 + ], + "lines": [ + { + "bbox": [ + 58, + 212, + 537, + 388 + ], + "spans": [ + { + "bbox": [ + 58, + 212, + 537, + 388 + ], + "type": "table", + "html": "
ClusteringOverclustering
YTVOSDAVISYTVOSDAVIS
FCDFCDFCDFCD
Trained on Images
Resnet5044.043.41.739.337.44.255.652.83.146.644.28.4
SwAV [8]39.538.23.232.029.67.359.858.15.850.550.125.7
DINO [9]39.137.91.930.231.01.666.265.44.056.954.917.9
Leopart [74]39.237.911.730.330.216.564.562.815.554.954.426.7
Trained on Videos
STEGO*41.540.32.031.931.03.258.154.35.147.646.310.4
DINO*37.236.11.229.329.22.453.150.91.345.444.08.6
Leopart*41.540.57.737.536.512.660.859.86.853.753.116.8
TIMET(ours)52.551.313.353.753.020.568.666.815.859.861.531.8
", + "image_path": "b583fd95eb9a09d8657786af735083df84c1fe0001089e5fe96a56ce8aecea9d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "type": "text", + "content": "Table 4: Unsupervised video semantic segmentation results in mIoU for clustering (K=GT) and overclustering. TIMET not only gets better numbers on YTVOS, but also achieves considerably better results on DAVIS. This shows the better quality of the learned features at transferability to other video datasets. The clip-length is set to 16 and 4 for DAVIS and YTVOS, respectively. For clustering, the Hungarian algorithm [35] matches the unsupervised segmentation clusters (K) with the ground truth (GT) per frame " + }, + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "type": "inline_equation", + "content": "(F)" + }, + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "type": "text", + "content": ", clip " + }, + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "type": "inline_equation", + "content": "(C)" + }, + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "type": "text", + "content": " or across the whole dataset " + }, + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "type": "text", + "content": ". For overclustering, we use K=10 for the frames-wise " + }, + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "type": "inline_equation", + "content": "(F)" + }, + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "type": "text", + "content": " and clip-wise " + }, + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "type": "inline_equation", + "content": "(C)" + }, + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "type": "text", + "content": " evaluations, and for dataset-wise " + }, + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 46, + 396, + 547, + 491 + ], + "type": "text", + "content": " evaluation, K=200 for DAVIS and K=500 for YTVOS. *denotes finetuning pretrained models on the same video frames as input as our method. The matching protocol is greedy many-to-one, see Appendix B for details." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 512, + 185, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 185, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 185, + 525 + ], + "type": "text", + "content": "4.3. Large-scale experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": "Unsupervised video semantic segmentation. In this section, we train our TIMET method on the YTVOS dataset and evaluate it for unsupervised video object segmentation on both DAVIS and YTVOS. The results are shown in Table 4. It should be noted that the number of prototypes we analyzed in Table 1a is not the number of clusters used to report accuracies for Table 4, as the actual number is usually unknown in the case of unsupervised learning. For the clustering experiment, we set K to the number of ground truth objects, which makes evaluation metrics easier to interpret and is commonly done in image clustering and segmentation. This evaluation is repeated but for larger numbers of clusters, in the \"overclustering\" scenario. Note that in dense self-supervised learning, overclustering can be particularly important because the learned representations are" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 513, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 545, + 573 + ], + "type": "text", + "content": "used as a feature extractor for downstream tasks such as semantic segmentation or object detection. By using more fine-grained representations, the network may be able to extract more discriminative features such as object parts [74], which can lead to better performances." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "From Table 4, we observe a clear trend: our method, trained on YTVOS not only achieves superior performances on YTVOS, but also beats existing image-trained models on DAVIS by a large margin. In particular, the state-of-the-art self-supervised clustering method, Leopart, has " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " lower performance on per dataset DAVIS evaluation for " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{K = GT}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " lower performance on YTVOS. The gap even becomes larger when Leopart is trained on the same video dataset with " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " lower numbers in per dataset and per clip numbers across different datasets. This means that when grouping objects of the same class over time and" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "16542" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "the whole dataset, the image-based self-supervised methods have trouble generalising to videos, where objects are not centered in the frame, can appear in varying poses, and are more difficult in general. A strong contender is Leopart, which matches the performance of DINO for per frame and per clip clustering, however, improves considerably with the per dataset clustering. We attribute this to the fact that their dense learning objective improves mainly the generalisation capabilities of the learned representation, however falling short of generalising to temporal variations. We make similar observations for overclustering, improving the DINO baseline which we utilize as initialisation by " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " across different datasets in the per dataset metric." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 396 + ], + "type": "text", + "content": "We conclude that the proposed method outperforms all other methods in learning robust and discriminative representations for dense video understanding, and image-based self-supervised learning may not have sufficient generalisation capacity. This shows that if used right, time is a particularly strong inductive bias for self-supervised learning. Figure 3 shows segmentations returned by the proposed model trained on YTVOS and tested on YTVOS, DAVIS, and VISOR respectively. The proposed method groups objects accurately, and importantly, the frame segmentations are considerably consistent over time (see more examples in Appendix D). This highlights the importance and relevance of temporal fine-tuning, not just for higher accuracy, but crucially for consistency and robustness." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "text", + "content": "Transfer from video training to images. Despite the common belief that features learned from videos perform worse when transferred to images [34, 20], the results in Table 5 demonstrate our method achieves high performance that match state-of-the-art methods directly trained on images, specifically for the " + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 500" + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "text", + "content": ", FCN, and LC metrics. We also compare our performance gains in videos against models designed for unsupervised image segmentation. The results show that models highly biased towards image datasets cannot provide the same performance when trained on videos, lagging behind our approach by " + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 411, + 289, + 639 + ], + "type": "text", + "content": ". Our results demonstrate that achieving high transfer performance on challenging tasks through video-based self-supervised learning is not only feasible but can also maintain high performance across modalities. These findings suggest that our method can drive further advances in self-supervised learning and inspire new directions for research in this field. Figure 3 shows the qualitative results on Pascal VOC. For more visualisations we refer to Appendix D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Salient object segmentation. In Table 6, we compare foreground masks obtained with various DINO ViT based methods. We use the cluster-based foreground extraction protocol from [74] (details provided in Appendix C). First, we find that our method outperforms the original DINO at" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 313, + 70, + 541, + 280 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 541, + 280 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 541, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 541, + 280 + ], + "type": "table", + "html": "
Pascal VOC
K=21K=500LCFCN
Trained on Images
ResNet-504.536.553.8-
DINO [9]5.517.450.660.6
SwAV [8]11.635.750.7-
MaskContrast [57]35.045.449.2-
DenseCL [61]-43.649.069.4
STEGO [21]7.019.559.163.5
CrOC [52]20.6-61.6-
Leopard [74]36.650.568.070.1
Trained on Videos
STEGO*4.015.551.155.5
Leopard*14.921.253.263.2
Flowdino†[70]--59.4-
TIMET (ours)34.553.268.070.6
", + "image_path": "03d0bcb406c2fe9194d8f827b8d4e5ee182440535ee9cdcfad6d70141308dc30.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 321, + 365, + 531, + 437 + ], + "blocks": [ + { + "bbox": [ + 304, + 286, + 547, + 348 + ], + "lines": [ + { + "bbox": [ + 304, + 286, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 547, + 348 + ], + "type": "text", + "content": "Table 5: Transfer from video training to images. Numbers taken from [74, 52, 70]. *: finetuning pretrained models on the same video frames as input as our method. †: uses a 40% larger superset of our train data. Details on the different datasets and methods are provided in the Appendix." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 365, + 531, + 437 + ], + "lines": [ + { + "bbox": [ + 321, + 365, + 531, + 437 + ], + "spans": [ + { + "bbox": [ + 321, + 365, + 531, + 437 + ], + "type": "table", + "html": "
Pascal VOCDAVISYTVOS
DINO [9]52.134.532.1
Leopard [74]59.637.338.6
STEGO [21]49.130.432.1
TIMET (ours)63.944.543.5
", + "image_path": "edb4ade60672b3830cce1206c636d9117fcb1a5d10dc8e5c4bba2ecd9752717c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 445, + 545, + 481 + ], + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 481 + ], + "type": "text", + "content": "Table 6: Salient object segmentation. We report performance using the Jaccard score [9] and use the official pretrained models for evaluation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 510, + 547, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 571 + ], + "type": "text", + "content": "tention maps consistently by " + }, + { + "bbox": [ + 304, + 510, + 547, + 571 + ], + "type": "inline_equation", + "content": "10 - 11\\%" + }, + { + "bbox": [ + 304, + 510, + 547, + 571 + ], + "type": "text", + "content": ". We also surpass the results of Leopart [74] and STEGO [21], works that rely on the same pretrained backbone. Even for the evaluation on Pascal VOC, our method achieves higher performances despite the domain shift of having trained on videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Generalisation to egocentric datasets. We train our method on EPIC-Kitchens-100 [13] and evaluate on VISOR [14]. We use the official code to convert VISOR to a DAVIS-like structure, in which we can report our numbers for per frame and per clip evaluation. Note that, after conversion, the object IDs do not maintain global consistency in the whole dataset; therefore, we cannot report the per dataset number. As Table 7 shows, TIMET outperforms image-based competitors on egocentric datasets as well." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "16543" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 70, + 229, + 144 + ], + "blocks": [ + { + "bbox": [ + 106, + 70, + 229, + 144 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 229, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 229, + 144 + ], + "type": "table", + "html": "
VISOR
FC
DINO [9]24.818.7
Leopard [74]24.118.5
TIMET (ours)26.521.5
", + "image_path": "b5f3f9c0c93dbe1fb5fffc6c78060c98b50f5d65cef2185b35d846a7d617b163.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 236, + 289, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 236, + 289, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 236, + 289, + 571 + ], + "type": "text", + "content": "Visual In-Context Learning evaluation. Here, we contrast our approach with a recently introduced benchmark, which assesses the in-context reasoning capabilities of models within the vision domain. Unlike linear or FCN classification methods, visual in-context learning evaluation obviates the need for fine-tuning or end-to-end training. Instead, it constructs validation segmentation maps using patch-level, feature-wise nearest neighbor similarities between validation images (referred to as \"queries\") and training samples (termed \"keys\"). This approach mirrors strategies in the NLP domain, aiming to evaluate the proficiency of models in learning tasks from limited examples presented as demonstrations. The results are shown by Table 8. Given that the vast majority of extant models in the domain utilize ViT-S16 as their backbone, we conducted a re-evaluation of their checkpoints to furnish a consistent and directly comparable evaluation table. Subsequently, we re-trained our model employing ViT-B16 and benchmarked it against other baselines as presented by [5]. The results, as depicted in the table, indicate that even though our model was exclusively trained on videos, it registers performance metrics in line with Leopart [74]. Furthermore, it surpasses the results of the leading method, CrOC [52], which was trained on images. Contrary to Hummingbird [5], TIMET is not custom-fitted to the specific evaluation setup and boasts superior computational efficiency. Notably, it requires only a single GPU for training, in contrast to the 16 TPUs demanded by Hummingbird." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 585, + 119, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 119, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 119, + 597 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "This paper has aimed to learn dense representations that are learned from videos; yet can be generalised to the image domain as well. As video content is growing rapidly and contains more information compared to images, learning generalisable knowledge from them facilitates further scaling of self-supervised learning methods. To this effect, we have proposed a self-supervised clustering loss to encourage temporally consistent features between different frames of a video. To efficiently find corresponding views between" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 321, + 70, + 533, + 274 + ], + "blocks": [ + { + "bbox": [ + 47, + 152, + 288, + 212 + ], + "lines": [ + { + "bbox": [ + 47, + 152, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 288, + 212 + ], + "type": "text", + "content": "Table 7: Generalisation to egocentric datasets. Unsupervised semantic segmentation results in mIoU for clustering " + }, + { + "bbox": [ + 47, + 152, + 288, + 212 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} = \\mathrm{GT})" + }, + { + "bbox": [ + 47, + 152, + 288, + 212 + ], + "type": "text", + "content": " on VISOR [14] after training onn EPIC-Kitchens-100 [13]. The clip-length is set to 4. Our method gets better results on egocentric datasets as well." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 70, + 533, + 274 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 533, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 533, + 274 + ], + "type": "table", + "html": "
EncoderParamsmIoU
Trained on Images
SupervisedViT-S1621M35.1
MoCo-v3* [11]ViT-S1621M19.5
DINO* [9]ViT-S1621M47.9
CrOC* [52]ViT-S1621M50.0
Leopard* [74]ViT-S1621M63.6
DINO [9]ViT-B1686M55.9
MoCo-v3 [11]ViT-B1686M37.2
MAE [22]ViT-B1686M6.6
LOCA [7]ViT-B1686M57.5
Hummingbird [5]ViT-B1686M70.5
Trained on Videos
TIMET (ours)ViT-S1621M61.6
TIMET (ours)ViT-B1686M65.5
", + "image_path": "1810de54920a10b0c11a61eebf044bef2181a6d40ff5c492ce1980e2eb4a35ba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 281, + 545, + 353 + ], + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 353 + ], + "type": "text", + "content": "Table 8: Visual In-Context Learning evaluation. Numbers are taken from [5]. *: numbers are produced by this paper. This approach is equivalent to a non-parametric nearest-neighbor based evaluation on Pascal VOC. The details of the evaluation benchmark can be found in the provided implementation codes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 376, + 546, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 495 + ], + "type": "text", + "content": "clip frames, we have proposed to recycle pretrained transformer features to leverage their natural tracking ability at each clip. Our empirical results indicate that this method achieves significant gains on the challenging task of video object segmentation across three different evaluation protocols and three datasets. Moreover, by transferring the learned model to the image domain, we have demonstrated the generalisability of the learned features by surpassing or matching the state-of-the-art for unsupervised image segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 509, + 416, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 416, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 416, + 522 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 529, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 545, + 590 + ], + "type": "text", + "content": "This work is financially supported by Qualcomm Technologies Inc., the University of Amsterdam and the allowance Top consortia for Knowledge and Innovation (TKIs) from the Netherlands Ministry of Economic Affairs and Climate Policy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 603, + 365, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 603, + 365, + 615 + ], + "spans": [ + { + "bbox": [ + 306, + 603, + 365, + 615 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 623, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 310, + 623, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 623, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 310, + 623, + 545, + 666 + ], + "type": "text", + "content": "[1] Nikita Araslanov, Simone Schaub-Meyer, and Stefan Roth. Dense unsupervised learning for video segmentation. Advances in Neural Information Processing Systems, 34:25308-25319, 2021. 2, 3, 15" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 669, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 669, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 311, + 669, + 547, + 713 + ], + "type": "text", + "content": "[2] Yuki M. Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In International Conference on Learning Representations (ICLR), 2020. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "16544" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 148 + ], + "type": "text", + "content": "[3] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI, pages 456-473. Springer, 2022. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 150, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 288, + 195 + ], + "type": "text", + "content": "[4] Arthur Aubret, Markus R. Ernst, Céline Teulière, and Jochen Triesch. Time to augment self-supervised visual representation learning. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 197, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 288, + 229 + ], + "type": "text", + "content": "[5] Ivana Balăžević, David Steiner, Nikhil Parthasarathy, Relja Arandjelović, and Olivier J Henaff. Towards in-context scene understanding. arXiv preprint arXiv:2306.01667, 2023. 9" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 231, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 231, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 231, + 288, + 262 + ], + "type": "text", + "content": "[6] Thomas Brox and Jitendra Malik. Large displacement optical flow: Descriptor matching in variational motion estimation. IEEE TPAMI, 33(3):500-513, 2011. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 264, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 288, + 297 + ], + "type": "text", + "content": "[7] Mathilde Caron, Neil Houlsby, and Cordelia Schmid. Location-aware self-supervised transformers. arXiv preprint arXiv:2212.02400, 2022. 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 298, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 288, + 353 + ], + "type": "text", + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 33:9912-9924, 2020. 4, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 355, + 288, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 355, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 355, + 288, + 419 + ], + "type": "text", + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 3, 4, 5, 6, 7, 8, 9, 15" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 498 + ], + "type": "text", + "content": "[11] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. ICCV, 2021. 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 533 + ], + "type": "text", + "content": "[12] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural Information Processing Systems, volume 26, 2013. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 610 + ], + "type": "text", + "content": "[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 5, 8, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 677 + ], + "type": "text", + "content": "[14] Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. Epic-kitchens visor benchmark: Video segmentations and object relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2022. 5, 8, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "text", + "content": "vain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 13" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 149 + ], + "type": "text", + "content": "[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 5, 13, 14" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 150, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 547, + 205 + ], + "type": "text", + "content": "[17] Christoph Feichtenhofer, Haoqi Fan, Bo Xiong, Ross Girshick, and Kaiming He. A large-scale study on unsupervised spatiotemporal representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3299-3309, 2021. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "text", + "content": "[18] Basura Fernando, Hakan Bilen, Efstratios Gavves, and Stephen Gould. Self-supervised video representation learning with odd-one-out networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3636-3645, 2017. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[19] Raghudeep Gadde, Varun Jampani, and Peter V Gehler. Semantic video cnns through representation warping. In Proceedings of the IEEE International Conference on Computer Vision, pages 4453-4462, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 305, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 346 + ], + "type": "text", + "content": "[20] Daniel Gordon, Kiana Ehsani, Dieter Fox, and Ali Farhadi. Watching the world go by: Representation learning from unlabeled videos. arXiv preprint arXiv:2003.07990, 2020. 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "type": "text", + "content": "[21] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022. 5, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 394, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 447 + ], + "type": "text", + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 2, 9" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[23] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "text", + "content": "[24] Olivier J Henaff, Skanda Koppula, Jean-Baptiste Alayrac, Aaron Van den Oord, Oriol Vinyals, and João Carreira. Efficient visual pretraining with contrastive detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10086-10096, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "type": "text", + "content": "[25] Olivier J Hénaff, Skanda Koppula, Evan Shelhamer, Daniel Zoran, Andrew Jaegle, Andrew Zisserman, João Carreira, and Relja Arandjelović. Object discovery and representation networks. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 123-143. Springer, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 625, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 646 + ], + "type": "text", + "content": "[26] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 13" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[27] Jefferson Hernandez, Ruben Villegas, and Vicente Ordonez. Visual representation learning from unlabeled video using contrastive masked autoencoders. arXiv preprint arXiv:2303.12001, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[28] Ping Hu, Fabian Caba, Oliver Wang, Zhe Lin, Stan Sclaroff, and Federico Perazzi. Temporally distributed networks" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "text", + "content": "16545" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 73, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 288, + 106 + ], + "type": "text", + "content": "for fast video semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8818-8827, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[29] Junhwa Hur and Stefan Roth. Joint optical flow and temporally consistent semantic segmentation. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part I 14, pages 163-177. Springer, 2016. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 205 + ], + "type": "text", + "content": "[30] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 1, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[31] Samvit Jain, Xin Wang, and Joseph E Gonzalez. Accel: A corrective fusion network for efficient semantic segmentation on video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8866-8875, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 317 + ], + "type": "text", + "content": "[32] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9865-9874, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 352 + ], + "type": "text", + "content": "[33] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547, 2019. 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "text", + "content": "[34] Thomas Kipf, Gamaleldin F Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. arXiv preprint arXiv:2111.12594, 2021. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 441 + ], + "type": "text", + "content": "[35] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 487 + ], + "type": "text", + "content": "[36] Xueting Li, Sifei Liu, Shalini De Mello, Xiaolong Wang, Jan Kautz, and Ming-Hsuan Yang. Joint-task self-supervised learning for temporal correspondence. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 553 + ], + "type": "text", + "content": "[37] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Video k-net: A simple, strong, and unified baseline for video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18847-18857, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 599 + ], + "type": "text", + "content": "[38] Yule Li, Jianping Shi, and Dahua Lin. Low-latency video semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5997-6005, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[39] Runtao Liu, Zhirong Wu, Stella Yu, and Stephen Lin. The emergence of objectness: Learning zero-shot segmentation from videos. Advances in Neural Information Processing Systems, 34:13137-13152, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[40] Si Liu, Changhu Wang, Ruihe Qian, Han Yu, Renda Bao, and Yao Sun. Surveillance video parsing with single frame supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 413-421, 2017. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[41] Yifan Liu, Chunhua Shen, Changqian Yu, and Jingdong Wang. Efficient semantic video segmentation with per-frame" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "inference. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 352-368. Springer, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 129 + ], + "type": "text", + "content": "[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 13" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 131, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 131, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 547, + 185 + ], + "type": "text", + "content": "[43] Xiankai Lu, Wenguan Wang, Jianbing Shen, Yu-Wing Tai, David J Crandall, and Steven CH Hoi. Learning video object segmentation from unlabeled videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8960-8970, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "text", + "content": "[44] Yuval Nirkin, Lior Wolf, and Tal Hassner. Hyperseg: Patchwise hypernetwork for real-time semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4061-4070, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 285 + ], + "type": "text", + "content": "[45] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 288, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 545, + 330 + ], + "type": "text", + "content": "[46] Nikhil Parthasarathy, SM Eslami, João Carreira, and Olivier J Henaff. Self-supervised video pretraining yields strong image representations. arXiv preprint arXiv:2210.06433, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 333, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 333, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 333, + 545, + 397 + ], + "type": "text", + "content": "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 13" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "text", + "content": "[48] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5, 13" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 498 + ], + "type": "text", + "content": "[49] Kanchana Ranasinghe, Muzammal Naseer, Salman Khan, Fahad Shahbaz Khan, and Michael S Ryoo. Self-supervised video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2874-2884, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "text", + "content": "[50] Sucheng Ren, Wenxi Liu, Yongtuo Liu, Haoxin Chen, Guoqiang Han, and Shengfeng He. Reciprocal transformations for unsupervised video object segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15455-15464, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "text", + "content": "[51] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 614, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 677 + ], + "type": "text", + "content": "[52] Thomas Stegmüller, Tim Lebailly, Behzad Bozorgtabar, Tinne Tuytelaars, and Jean-Philippe Thiran. Croc: Cross-view online clustering for dense visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7000–7009, 2023. 8, 9" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[53] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In Proceedings of the IEEE/CVF conference on" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "text", + "content": "16546" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 94 + ], + "type": "text", + "content": "computer vision and pattern recognition, pages 8934-8943, 2018. 15" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "text", + "content": "[54] Guolei Sun, Yun Liu, Henghui Ding, Thomas Probst, and Luc Van Gool. Coarse-to-fine feature mining for video semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3126-3137, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 288, + 217 + ], + "type": "text", + "content": "[55] Fida Mohammad Thoker, Hazel Doughty, Piyush Bagad, and Cees GM Snoek. How severe is benchmark-sensitivity in video self-supervised learning? In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 632-652. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 282 + ], + "type": "text", + "content": "[56] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X, pages 268-285. Springer, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 288, + 338 + ], + "type": "text", + "content": "[57] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10052-10062, 2021. 1, 2, 5, 8, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 392 + ], + "type": "text", + "content": "[58] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Chuanxin Tang, Xiyang Dai, Yucheng Zhao, Yujia Xie, Lu Yuan, and Yu-Gang Jiang. Look before you match: Instance understanding matters in video object segmentation. arXiv preprint arXiv:2212.06826, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 393, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 288, + 426 + ], + "type": "text", + "content": "[59] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. arXiv preprint arXiv:2301.11320, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "text", + "content": "[60] Xiaolong Wang, Allan Jabri, and Alexei A Efros. Learning correspondence from the cycle-consistency of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2566-2576, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "text", + "content": "[61] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024-3033, 2021. 2, 5, 8, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 590 + ], + "type": "text", + "content": "[62] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684-16693, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 647 + ], + "type": "text", + "content": "[63] Dejing Xu, Jun Xiao, Zhou Zhao, Jian Shao, Di Xie, and Yueting Zhuang. Self-supervised spatiotemporal learning via video clip order prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10334–10343, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "text", + "content": "[64] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas Huang. Youtube-vos: A large-scale video object segmentation benchmark. arXiv preprint arXiv:1809.03327, 2018. 5, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[65] Yu-Syuan Xu, Tsu-Jui Fu, Hsuan-Kung Yang, and Chun-Yi Lee. Dynamic video segmentation network. In Proceed-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 520 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 6556-6565, 2018. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[66] Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7177-7188, 2021. 2, 3, 15" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[67] Sukmin Yun, Hankook Lee, Jaehyung Kim, and Jinwoo Shin. Patch-level representation learning for self-supervised vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8354-8363, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 251 + ], + "type": "text", + "content": "[68] Andrii Zadaianchuk, Matthaeus Kleindessner, Yi Zhu, Francesco Locatello, and Thomas Brox. Unsupervised semantic segmentation with self-supervised object-centric representations. arXiv preprint arXiv:2207.05027, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 545, + 306 + ], + "type": "text", + "content": "[69] Shaofeng Zhang, Feng Zhu, Rui Zhao, and Junchi Yan. Patch-level contrasting without patch correspondence for accurate and dense contrastive representation learning. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "type": "text", + "content": "[70] Xinyu Zhang and Abdeslam Boullarias. Optical flow boosts unsupervised localization and segmentation. arXiv preprint arXiv:2307.13640, 2023. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "text", + "content": "[71] Yurong Zhang, Liulei Li, Wenguan Wang, Rong Xie, Li Song, and Wenjun Zhang. Boosting video object segmentation via space-time correspondence learning. arXiv preprint arXiv:2304.06211, 2023. 1, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 387, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 419 + ], + "type": "text", + "content": "[72] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. ICLR, 2022. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 420, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 545, + 464 + ], + "type": "text", + "content": "[73] Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, and Yichen Wei. Deep feature flow for video recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2349-2358, 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 465, + 547, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 465, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 465, + 547, + 520 + ], + "type": "text", + "content": "[74] Adrian Ziegler and Yuki M Asano. Self-supervised learning of object parts for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14502-14511, 2022. 1, 2, 5, 7, 8, 9, 13, 14, 15, 16, 18" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "16547" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/0f2857ef-2537-43f4-b49d-a3319a025266_content_list.json b/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/0f2857ef-2537-43f4-b49d-a3319a025266_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..797bbe96d7c1118b0bc6b1a96d112269b5d240bf --- /dev/null +++ b/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/0f2857ef-2537-43f4-b49d-a3319a025266_content_list.json @@ -0,0 +1,1795 @@ +[ + { + "type": "text", + "text": "Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera", + "text_level": 1, + "bbox": [ + 102, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Urbano Miguel Nunes1, Laurent Udo Perrinet2 and Sio-Hoi Ieng1", + "bbox": [ + 223, + 202, + 743, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Sorbonne University, $^{2}$ Aix Marseille Univ, CNRS, INT, Institut de Neurosciences de la Timone urbano.goncalves-nunes@inserm.fr, laurent.perrinet@univ-amu.fr, sio-hoi.ieng@upmc.fr", + "bbox": [ + 111, + 223, + 854, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras asynchronously report brightness changes with a temporal resolution in the order of microseconds, which makes them inherently suitable to address problems that involve rapid motion perception. In this paper, we address the problem of time-to-contact (TTC) estimation using a single event camera. This problem is typically addressed by estimating a single global TTC measure, which explicitly assumes that the surface/obstacle is planar and fronto-parallel. We relax this assumption by proposing an incremental event-based method to estimate the TTC that jointly estimates the (up-to scale) inverse depth and global motion using a single event camera. The proposed method is reliable and fast while asynchronously maintaining a TTC map (TTCM), which provides per-pixel TTC estimates. As a side product, the proposed method can also estimate per-event optical flow. We achieve state-of-the-art performances on TTC estimation in terms of accuracy and runtime per event while achieving competitive performance on optical flow estimation.", + "bbox": [ + 75, + 325, + 473, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 646, + 209, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras differ from standard frame-based cameras, which capture visual data at a fixed rate and independently of the observing scene. Instead, event cameras respond asynchronously to pixel-wise brightness changes by generating events [6, 25]. Event cameras are thus data-driven sensors that offer several advantages, including high temporal resolution in the order of microseconds, low latency, low power consumption, and high dynamic range. These properties place event cameras as suitable candidates to address vision-based problems that involve (high-speed) motion, e.g., optical flow estimation [4, 27, 51], ego-motion estimation [15, 21, 22, 34], motion segmentation [33, 44], and obstacle avoidance [10, 12]. Due to the distinct visual sensing paradigm, however, new methods are necessary to fully exploit the potential of event cameras [13, 23].", + "bbox": [ + 75, + 674, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cba9e1000800efaf48a731f8b951e7fde7238357dc06ad4e3210505788e1f80e.jpg", + "image_caption": [ + "Figure 1. Runtime vs. accuracy comparison for TTC estimation methods. Average results on the Ventral Landing benchmark [28]." + ], + "image_footnote": [], + "bbox": [ + 509, + 292, + 882, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras have been used to address the problem of fast TTC estimation, which comes up often in the vision-based obstacle avoidance [10, 12, 29, 39] and ventral landing [28, 36] literature. The TTC is the time that would elapse before a camera reaches an obstacle/surface, assuming the current relative motion between them remains constant [10]. Previous methods that use a single event camera have focused on estimating a single global TTC measure, which assumes that the surface is planar and fronto-parallel. To overcome this limitation, other methods use additional sensing, e.g., depth frames, to build a dense TTCM [20, 47].", + "bbox": [ + 496, + 534, + 892, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We instead propose to extend the Dispersion Minimization (DMin) framework [34] to estimate the TTC for each incoming event using a single event camera. The proposed method jointly estimates the relative global motion and per-event (up-to scale) inverse depth. We can then asynchronously maintain a semi-dense TTCM which provides per-pixel TTC estimates or compute a global TTC measure with greater accuracy by averaging over the TTC estimates. Since there is at least one scaling degree of freedom (DOF), we also propose an effective strategy to mitigate event collapse [40, 41]. The proposed method is also computationally fast, reaching $\\sim 1$ microsecond processing time per event on a standard laptop. Fig. 1 compares the", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23653", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "runtime vs. accuracy for TTC estimation methods, whereby the proposed method achieves state-of-the-art performance. We also estimate the per-event optical flow as a side product and achieve competitive performance compared to state-of-the-art optical flow methods that use events.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Main contributions:", + "text_level": 1, + "bbox": [ + 76, + 167, + 222, + 181 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. First event-based method that explicitly estimates the TTC for each event and maintains a semi-dense TTCM using a single event camera.", + "2. DMin framework [34] extension to jointly handle local and global estimates, i.e., inverse depth and global motion, respectively.", + "3. Effective approach that mitigates event collapse [40, 41] for incremental event-based estimation." + ], + "bbox": [ + 88, + 183, + 468, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1. Time-to-Contact", + "text_level": 1, + "bbox": [ + 76, + 315, + 240, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consider a freely moving camera with angular and linear velocities $\\omega(t) = (\\omega_x(t), \\omega_y(t), \\omega_z(t))^{\\top}$ and $\\pmb{\\nu}(t) = (\\nu_x(t), \\nu_y(t), \\nu_z(t))^{\\top}$ , respectively, that is observing a point $\\alpha$ , with 3D coordinates $\\pmb{\\alpha}(t) = (X(t), Y(t), Z(t))^{\\top}$ relative to the camera, as shown in Fig. 2. $Z(t)$ is also referred as the depth of point $\\alpha$ relative to the camera. The instantaneous TTC between the moving camera and point $\\alpha$ is thus given by:", + "bbox": [ + 75, + 339, + 468, + 465 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\tau (t) := - \\frac {Z (t)}{\\frac {d Z (t)}{d t}} = \\frac {Z (t)}{\\nu_ {z} (t)}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 465, + 468, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The minus sign disappears because we define the linear velocity w.r.t. the camera's frame of reference, not w.r.t. the point's frame of reference, i.e., $\\nu_{z}(t) = -dZ(t) / dt$ . Based on Eq. (1), the exact values of depth and relative approaching motion do not need to be estimated, only the ratio between them.", + "bbox": [ + 75, + 510, + 468, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 617, + 218, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We review recent related works on the following topics: TTC, global motion and optical flow estimation. We refer to [13] for a detailed survey.", + "bbox": [ + 75, + 643, + 468, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Time-to-Contact Estimation. The first work on event-based TTC using a single event camera relied solely on the estimation of visual motion flows [10], whereby the motion flows were computed by fitting a local plane to the time surface [4]. Other works that followed were geared towards two main use cases, namely obstacle avoidance [12, 29, 39] and ventral landing [28, 36, 43]. Event-based obstacle avoidance methods are built to be fast reacting and, although they come from either bio-inspired [29, 39] or mathematically grounded principles [12], they typically rely on empirically-validated heuristics to speed-up computations. Existing event-based ventral landing approaches only compute a single TTC estimate, which assumes that the surface is planar and fronto-parallel. To overcome this assumption,", + "bbox": [ + 75, + 688, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/605e5596d26923c5bebe65252c2d25c2c3fb8001614de17f6958618fdca4b29e.jpg", + "image_caption": [ + "Figure 2. Camera coordinate system." + ], + "image_footnote": [], + "bbox": [ + 531, + 87, + 864, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "other works fuse events with additional sensory information, e.g. depth [47]. The proposed approach uses a single event camera while being mathematically grounded and computationally fast.", + "bbox": [ + 496, + 265, + 890, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Global Motion Estimation. Also denoted by ego-motion estimation [13], it refers to estimating the parameters that explain the triggered events according to some global motion model. These methods can be broadly characterized by whether they rely on key-frame registration [8, 9, 14, 21, 22, 38] or perform the estimation without relying on any key-frame [15, 34]. The former methods are reminiscent of the frame-based paradigm and also include methods based on artificial neural networks (ANN) [16, 52], whereby an intermediate frame-based representation is needed for the estimation. The latter methods tend towards a more event-based processing paradigm and include methods based on spiking neural networks (SNN) [18, 37], whereby events are either processed on an event-by-event basis [34] or in batches [15]. Although both approaches have advantages and disadvantages, methods that rely on key-frame registration typically require an intermediate frame-based representation, which is still an open problem in the event-based community. Similarly to [22], the proposed method jointly estimates up-to scale inverse depth and global motion. By building on the DMin framework [34], which also allows processing events on an event-by-event basis, our method does not rely on key-frame registration or background inverse depth regularization to improve convergence.", + "bbox": [ + 496, + 325, + 892, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optical Flow Estimation. Several model-based methods have been previously proposed, which can be further divided into: frame-based [2, 7, 26, 27], batch-based [42, 49], and event-based [1, 3, 4]. By selecting the most relevant events, i.e., typically the most recent, frame-based methods build frames from which the optical flow is computed using techniques from standard image-based optical flow, e.g., Lucas-Kanade [5]. Since each event does not carry much information on its own, batch-based methods aggregate the most recent events by forming batches but perform the computations directly on the events. Event-based methods follow the most event-driven paradigm by performing event-by-event processing, typically being the computationally fastest. However, event-based methods tend to suffer more", + "bbox": [ + 496, + 688, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23654", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from the aperture problem since all the computations are performed locally, and thus frame-based and model-based methods achieve currently better accuracy. In terms of accuracy, learning-based ANN methods [11, 35, 46, 52] generally achieve state-of-the-art performance. Besides the need to convert events into frames for more efficient processing, these methods are known to be very data hungry, sensitive to the training data [48], and consume large amounts of energy [27]. Another line of research in learning-based methods has been to use SNN [19, 24], which combine the event-based processing and learning paradigms and thus do not require an intermediate frame-based representation. However, it is not trivial to train SNN, and the empirical validation is still not on par with ANN methods. Although it is not the primary objective of this work, the proposed method can provide per-event flow estimates while still being competitive in terms of accuracy w.r.t. state-of-the-art methods.", + "bbox": [ + 75, + 90, + 472, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 359, + 169, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the proposed incremental event-based method for TTCM estimation. We first briefly review the event cameras' working principle, and the DMin framework [34], based on which we develop the proposed method. Refer to the supplementary material for the full mathematical derivations and additional details.", + "bbox": [ + 75, + 385, + 470, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Event Cameras and Dispersion Minimization", + "text_level": 1, + "bbox": [ + 76, + 482, + 460, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event cameras output a stream of asynchronous temporal contrast events $\\{e_i\\}, i \\in \\mathbb{N}$ . Each event $e_i$ represents a spatio-temporal asynchronous brightness change, being defined as a tuple $e_i := (\\mathbf{x}_i, t_i, p_i)$ , where $\\mathbf{x}_i = (x_i, y_i)$ are the pixel coordinates, $t_i$ is the timestamp at which the event was generated, and $p_i \\in \\{-1, +1\\}$ is its polarity. An event $e_i$ is generated when the change in log-brightness $\\log \\mathbf{I}_{x,y}(t) := \\bar{\\mathbf{I}}_{x,y}(t)$ is above a threshold $L$", + "bbox": [ + 75, + 506, + 468, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\Delta \\bar {\\mathbf {I}} _ {x _ {i}, y _ {i}} \\left(t _ {i}\\right) \\right| = \\left| \\bar {\\mathbf {I}} _ {x _ {i}, y _ {i}} \\left(t _ {i}\\right) - \\bar {\\mathbf {I}} _ {x _ {i}, y _ {i}} \\left(t _ {i} - \\Delta t _ {i}\\right) \\right| \\geq L, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 632, + 468, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Delta t_{i}$ is the time since the last event at the same pixel.", + "bbox": [ + 76, + 657, + 467, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The DMin framework [34] estimates the parameters $\\pmb{\\theta}$ of a transformation model $\\mathcal{M}$ from the stream of events $\\mathcal{E} = \\{e_i\\}_{i=1}^{N_e}$ by minimizing a dispersion measure of the transformed events $\\mathbf{f}_i = \\mathcal{M}(e_i; t_{\\mathrm{ref}}, \\pmb{\\theta})$ . We consider the Potential measure with a Gaussian kernel $\\mathcal{N}(\\mathbf{x}; \\pmb{\\mu}, \\pmb{\\Sigma})$ :", + "bbox": [ + 76, + 671, + 468, + 748 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP (\\mathcal {E}; \\boldsymbol {\\theta}) = - \\sum_ {i, j} ^ {N _ {e}} \\mathcal {N} \\left(\\mathbf {f} _ {i}; \\mathbf {f} _ {j}, \\mathbf {I}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 753, + 468, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{I}$ is the identity matrix. A key distinction of the DMin framework is that it allows to incrementally estimate the model parameters $\\theta$ on an event-by-event basis, whereby the model parameters $\\theta$ can be iteratively solved by linearizing the transformation model $\\mathcal{M}$ , such that:", + "bbox": [ + 76, + 801, + 468, + 878 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {i} = \\mathcal {M} \\left(e _ {i}; t _ {\\mathrm {r e f}}, \\boldsymbol {\\theta}\\right) = \\tilde {\\mathbf {f}} _ {i} + \\Delta t _ {i, \\mathrm {r e f}} \\mathbf {B} _ {i} \\boldsymbol {\\theta}, \\mathbf {B} _ {i} := \\mathbf {B} \\left(e _ {i}\\right), (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 883, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{B}_i$ is the model-dependent linearization matrix. The parameters $\\theta^{*}$ that minimize Eq. (3) are thus given by:", + "bbox": [ + 498, + 90, + 890, + 122 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} ^ {*} = - \\left(\\sum_ {i, j} ^ {N _ {e}} \\mathbf {W} _ {i, j} \\Delta \\mathbf {C} _ {i, j}\\right) ^ {- 1} \\left(\\sum_ {i, j} ^ {N _ {e}} \\mathbf {W} _ {i, j} \\Delta \\tilde {\\mathbf {f}} _ {i, j}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 128, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Delta \\mathbf{C}_{i,j} = \\mathbf{C}_i - \\mathbf{C}_j$ , $\\Delta \\tilde{\\mathbf{f}}_{i,j} = \\tilde{\\mathbf{f}}_i - \\tilde{\\mathbf{f}}_j$ , $\\mathbf{W}_{i,j} = w_{i,j}\\partial \\Delta \\mathbf{f}_{i,j}^{\\top} / \\partial \\pmb {\\theta}$ , $\\mathbf{C}_i = \\Delta t_{i,\\mathrm{ref}}\\mathbf{B}_i$ and $w_{i,j} = \\mathcal{N}(\\mathbf{f}_i;\\mathbf{f}_j,\\mathbf{I})$ .", + "bbox": [ + 498, + 186, + 893, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Adapted Dispersion Minimization", + "text_level": 1, + "bbox": [ + 500, + 227, + 797, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the DMin method [34] provides a general framework for global incremental event-based model estimation, it can also be adapted to jointly estimate global and local measures, i.e., global angular and linear velocities and local inverse depth. However, the DMin framework may encounter estimation issues when the global model has at least one scaling DOF, as noted in [34, 40, 41], known as event collapse. Event collapse occurs when the events are transformed into a single point, which minimizes the events' dispersion or maximizes the image contrast while the parameters' estimates diverge. So far, to the best of our knowledge, the mitigation discussion in the literature has been on how to constrain the optimization loss to discourage divergent estimates by analyzing the effects of the scaling transformations on the event-based data. Several mitigation strategies have thus been proposed on the events [34, 40] and parameters level [41] by adding terms to regularize the objective measure. While these strategies generally prevent event collapse, they increase the complexity of the optimization framework and introduce additional parameters to tune.", + "bbox": [ + 496, + 250, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We instead observe that event collapse fundamentally stems from the event transformation to a common time reference $t_{\\mathrm{ref}}$ , e.g., given by Eq. (4), by identifying two problems: 1) the constant velocity assumption may not hold depending on the time difference $\\Delta t_{i,\\mathrm{ref}}$ , and 2) there is no built-in constraint on the magnitude of the model parameters, e.g., such that the difference between transformed events $\\mathbf{f}_i - \\mathbf{f}_j$ explicitly penalizes divergent estimates; although, according to Eq. (4), $\\mathbf{f}_i$ and $\\mathbf{f}_j$ individually diverge if the model parameters also diverge, their difference $\\mathbf{f}_i - \\mathbf{f}_j$ is not guaranteed to diverge and thus penalize divergent estimates. The first problem is typically addressed by heuristically making the time difference as short as possible, and its effects are of limited significance in practice. The second problem, however, is intrinsically linked to batch-based processing since the alignment of the events in a batch needs to be measured in some common time reference [15, 32].", + "bbox": [ + 496, + 553, + 890, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, for incremental event-based processing, the event transformation and, consequently, the dispersion measure can be modified without loss of generality such that the event collapse is prevented by implicitly addressing the two problems identified. Fig. 3 depicts the idea whereby we only transform the event neighbors to the current event's", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23655", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1b248b34a5449923deacc9fef9846fd2b30875d43fadf59b6a6f2e9d15c7a1a7.jpg", + "image_caption": [ + "Figure 3. Incremental event collapse mitigation. Instead of transforming all the events to some time reference (left), we locally transform the events to the current event's timestamp (right)." + ], + "image_footnote": [], + "bbox": [ + 112, + 93, + 310, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/929f61df89c36391bdca5a91846302bf95f20a7e025a504bc707f0ebfa479b2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 93, + 446, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "timestamp instead of transforming all the events to a common time reference, including the current event. Formally, instead of transforming the events according to Eq. (4), only the neighboring events $e_j$ of the current event $e_i$ are transformed according to", + "bbox": [ + 75, + 239, + 467, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {j} ^ {\\prime} = \\mathbf {x} _ {j} + \\Delta t _ {i, j} \\mathbf {B} _ {j} \\boldsymbol {\\theta} = \\mathbf {x} _ {j} + \\mathbf {C} _ {i, j} \\boldsymbol {\\theta}, e _ {j} \\in \\operatorname {n e i g h} (e _ {i}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 325, + 468, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where 'neigh' is shorthand for neighborhood, and the dependency of $\\mathbf{x}_j^\\prime$ on $\\pmb{\\theta}$ was omitted for brevity. From the resultant residual $\\mathbf{r}_{i,j} = \\mathbf{x}_i - \\mathbf{x}_j^\\prime = \\Delta \\mathbf{x}_{i,j} - \\mathbf{C}_{i,j}\\pmb {\\theta}$ , we see that the proposed modification to the DMin framework addresses both identified problems. First, the constant velocity assumption is better held since the time difference satisfies $\\Delta t_{i,j}\\leq \\Delta t_{i,\\mathrm{ref}}$ . Second, if the model parameters $\\pmb{\\theta}$ diverge, then the residual $\\mathbf{r}_{i,j}$ also diverges, which effectively penalizes divergent estimates. We highlight that the proposed adaptation only works for incremental event-based processing: it is not suitable for batch-based processing since the proposed transformation only works locally.", + "bbox": [ + 75, + 353, + 468, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Potential measure is modified by computing the difference between the current event's coordinates $\\mathbf{x}_i$ and the neighboring events' transformed coordinates $\\mathbf{x}_j'$ :", + "bbox": [ + 75, + 535, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(\\mathcal {E}; \\boldsymbol {\\theta}\\right) = - \\sum_ {i, j} \\mathcal {N} \\left(\\mathbf {x} _ {j} ^ {\\prime}; \\mathbf {x} _ {i}, \\mathbf {I}\\right), e _ {j} \\in \\operatorname {n e i g h} \\left(e _ {i}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 592, + 468, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By minimizing Eq. (7) by linearizing the residual according to Taylor's formula $\\mathbf{r}_{i,j}(\\pmb {\\theta} + \\Delta \\pmb {\\theta})\\approx \\mathbf{r}_{i,j}(\\pmb {\\theta}) + \\mathbf{J}_{i,j}(\\pmb {\\theta})\\Delta \\pmb{\\theta}$ , the optimized model parameters $\\pmb{\\theta}^{*}$ are iteratively updated:", + "bbox": [ + 75, + 635, + 468, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} ^ {*} \\leftarrow \\boldsymbol {\\theta} ^ {*} + \\Delta \\boldsymbol {\\theta}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 705, + 468, + 723 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\boldsymbol {\\theta} = - \\underbrace {\\left(\\sum_ {j} w _ {i , j} \\mathbf {J} _ {i , j} ^ {\\top} \\mathbf {J} _ {i , j}\\right)} _ {\\boldsymbol {\\Psi}} ^ {- 1} \\underbrace {\\left(\\sum_ {j} w _ {i , j} \\mathbf {J} _ {i , j} ^ {\\top} \\mathbf {r} _ {i , j}\\right)} _ {\\boldsymbol {\\psi}},\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 726, + 447, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{i,j} = \\mathcal{N}\\left(\\mathbf{x}_j';\\mathbf{x}_i,\\mathbf{I}\\right)$ , $\\mathbf{J}_{i,j} = \\partial \\mathbf{r}_{i,j} / \\partial \\pmb{\\theta}$ and $e_j\\in$ $\\mathrm{neigh}(e_i)$ .", + "bbox": [ + 75, + 806, + 467, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Inverse Depth and Global Motion Model", + "text_level": 1, + "bbox": [ + 76, + 847, + 423, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider that a calibrated event camera can freely move and whose global motion is parameterized by the 3D", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/69a437785dceb881ab23e23df94a55bf5358cba59b3f598fc4af52bcd562db5d.jpg", + "image_caption": [ + "Figure 4. Typical per-event estimation of optical flow and depth." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 885, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "angular and linear velocities, as defined in Sec. 1.1, $\\pmb{\\theta} = (\\pmb{\\nu}^{\\mathsf{T}},\\pmb{\\omega}^{\\mathsf{T}})^{\\mathsf{T}}$ . For each event $e_i$ , we estimate its inverse depth $\\rho_{i}\\coloneqq 1 / Z_{i}$ , based on the well-known expression for the apparent velocity on the image plane:", + "bbox": [ + 498, + 250, + 890, + 314 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {i} = \\left( \\begin{array}{l l} \\rho_ {i} \\mathbf {V} _ {i} & \\boldsymbol {\\Omega} _ {i} \\end{array} \\right) \\binom {\\boldsymbol {\\nu}} {\\omega} = \\rho_ {i} \\mathbf {V} _ {i} \\boldsymbol {\\nu} + \\boldsymbol {\\Omega} _ {i} \\boldsymbol {\\omega} := \\mathbf {B} _ {i} \\boldsymbol {\\theta}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 323, + 890, + 356 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} _ {i} = \\left( \\begin{array}{c c c} - f _ {x} & 0 & x _ {i} - c _ {x} \\\\ 0 & - f _ {y} & y _ {i} - c _ {y} \\end{array} \\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 358, + 705, + 391 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Omega} _ {i} = \\left( \\begin{array}{l l l} \\frac {(x _ {i} - c _ {x}) (y _ {i} - c _ {y})}{f _ {y}} & - f _ {x} - \\frac {(x _ {i} - c _ {x}) ^ {2}}{f _ {x}} & (y _ {i} - c _ {y}) \\frac {f _ {x}}{f _ {y}} \\\\ f _ {y} + \\frac {(y _ {i} - c _ {y}) ^ {2}}{f _ {y}} & - \\frac {(x _ {i} - c _ {x}) (y _ {i} - c _ {y})}{f _ {x}} & - (x _ {i} - c _ {x}) \\frac {f _ {y}}{f _ {x}} \\end{array} \\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 393, + 885, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we make explicit the dependency on the camera intrinsic parameters, namely the horizontal and vertical focal lengths $f_{x}$ and $f_{y}$ , respectively, and the horizontal and vertical focal center coordinates $c_{x}$ and $c_{y}$ , respectively. In this paper, we assume that the focal center coordinates represent the focus of expansion (FOE). Fig. 4 shows the typical per-event estimation of optical flow and depth using the proposed method.", + "bbox": [ + 496, + 446, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Time-to-Contact Map", + "text_level": 1, + "bbox": [ + 500, + 575, + 702, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each event $e_i$ , we estimate its inverse depth $\\rho_{i}$ and update the global motion parameters $\\pmb{\\theta} = (\\pmb{\\nu}^{\\mathsf{T}},\\pmb{\\omega}^{\\mathsf{T}})^{\\mathsf{T}}$ . The model parameters $\\gamma_{i}$ are formed by stacking the motion parameters and inverse depth $\\gamma_{i} = (\\pmb{\\theta}^{\\mathsf{T}},\\rho_{i})^{\\mathsf{T}}$ . We impose a smoothness constraint to Eq. (6), so that neighboring inverse depth estimates are assumed to be equal to $\\rho_{i}$ :", + "bbox": [ + 496, + 599, + 890, + 693 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {j} ^ {\\prime} = \\mathbf {x} _ {j} + \\Delta t _ {i, j} \\mathbf {B} _ {i, j} \\boldsymbol {\\theta}, e _ {j} \\in \\operatorname {n e i g h} \\left(e _ {i}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 702, + 890, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{B}_{i,j} = (\\rho_i\\mathbf{V}_j\\boldsymbol {\\Omega}_j)$ . The iterative update $\\Delta \\gamma_{i}$ is given by Eq. (8), where $\\mathbf{J}_{i,j} = -\\Delta t_{i,j}\\left(\\mathbf{B}_{i,j}\\mathbf{V}_j\\boldsymbol {\\nu}\\right)$ . We maintain the TTCM by computing the TTC for each event $e_i$ based on Eq. (1): $\\tau_{i} = 1 / (\\rho_{i}\\nu_{z})$ . We can also estimate the global TTC by averaging over the values maintained in the TTCM. Fig. 5 shows the typical per-event estimation of optical flow and TTCM using the proposed method.", + "bbox": [ + 496, + 729, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Initialization", + "text_level": 1, + "bbox": [ + 500, + 847, + 635, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The motion parameters $\\theta$ are global measures which are estimated by aggregating events, while each inverse depth", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23656", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/77d3fbbe11908f2e76c824e565ee59d80421fef507c962bd274b8dfc39da7c2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 93, + 312, + 118 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/38c2954c7bc377b2656f9f1a925c5cad92a8768ab8419ba5d93100c5ed65650b.jpg", + "image_caption": [ + "indoor_flying1 [50]", + "Accumulated events", + "Figure 5. Typical per-event estimation of optical flow and TTCM." + ], + "image_footnote": [], + "bbox": [ + 84, + 119, + 207, + 193 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dbb593d55e20fc4aee47a2ccfe3b8472d6f709d90c2dddb7ceeaa10131e6edc4.jpg", + "image_caption": [ + "1sec 10sec", + "Optical flow" + ], + "image_footnote": [], + "bbox": [ + 210, + 119, + 367, + 193 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/382f43152d0220e9fcdc3a81ecf1955d3901ebf212e147e7f6e39c9066eb23f8.jpg", + "image_caption": [ + "TTCM" + ], + "image_footnote": [], + "bbox": [ + 367, + 119, + 464, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\rho_{i}$ value corresponds to the estimate of a single event's inverse depth. The initialization of the motion parameters $\\theta$ can thus be almost arbitrarily set, e.g., typically set to 0, and it is only performed once at the beginning. However, a more careful initialization procedure must be considered for the case of inverse depth since it needs to be performed for each event. Based on recent advances in event-based global time decay [31], we perform a weighted average based on the previous neighboring events' inverse depth estimates as the initialization procedure. Each neighboring event's weight $w_{i}(t)$ is given by:", + "bbox": [ + 75, + 252, + 468, + 419 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} (t) = \\frac {1}{1 + a (t) \\left(t - t _ {i}\\right)}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 430, + 468, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $a(t)$ is the global event activity. If there are no previous neighboring events' inverse depth estimates, which should only occur at the start of the estimation, the initial inverse depth estimate is set to 1.", + "bbox": [ + 75, + 473, + 468, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Practical Considerations", + "text_level": 1, + "bbox": [ + 76, + 544, + 302, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned in Sec. 1.1, only the ratio between the depth and relative motion is required to compute the TTC. Since the depth and linear velocities estimates are obtained up-to a scale factor due to the monocular ambiguity, we constrain the linear velocities $\\pmb{\\nu}$ to have at most unit norm, i.e., $|\\pmb{\\nu}|_2 \\leq 1$ , which is useful to improve the method's computational stability by bounding the allowed estimates' values. This is not related to event collapse; rather, it stems from the ratio given by Eq. (1), whereby we can introduce an arbitrary non-zero multiplicative scalar to the numerator and denominator and have the same TTC estimate.", + "bbox": [ + 75, + 566, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the events are generated by 3D points in the field of view (FOV) of the camera and should have positive depth values, we constrain the depth values to be strictly positive by introducing a parameterization variable $\\lambda \\in \\mathbb{R}$ such that $\\rho (\\lambda) = e^{\\lambda} > 0$ .", + "bbox": [ + 75, + 734, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The iterative update given by Eq. (8) can be quite noisy since it is only computed in a small neighborhood $\\mathrm{neigh}(e_i) = \\{e_k:|\\mathbf{x}_k - \\mathbf{x}_i|_{\\infty}\\leq s\\}$ . We thus introduce two prior parameters for the global motion parameters and the local inverse depth estimates $l_{\\theta}$ and $l_{\\rho}$ , respectively. The resultant iterative update of the model parameters is given", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5cf9c4eef45a08c841164aed3effaac431fadf5a15a79ab6eb868ae02d6c4a7f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Neighborhood size s\\( {w}_{\\text{thresh }} \\)\\( {l}_{\\theta } \\)\\( {l}_{\\rho } \\)
30.01\\( \\propto O \\)(camera resolution)1
", + "bbox": [ + 503, + 89, + 890, + 128 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Hyper-parameters used across the experiments.", + "bbox": [ + 527, + 138, + 861, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by $\\Delta \\pmb{\\gamma}_i = (\\pmb{\\Psi} + \\mathbf{L})^{-1}\\left(-\\pmb{\\psi} + \\mathbf{L}(\\pmb{\\gamma}_{\\mathrm{prev}} - \\pmb{\\gamma}_i^*)\\right)$ , where $\\mathbf{L} = \\mathrm{diag}(l_{\\pmb{\\theta}},\\dots ,l_{\\pmb{\\theta}},l_{\\rho})$ , and $\\pmb{\\gamma}_{\\mathrm{prev}}$ are the parameters' estimates from the previous event $e_{i-1}$ . We also weigh each event $e_j$ contribution according to the corresponding weight $w_j(t)$ given by Eq. (11), and discard any event whose weight is below a threshold $w_{\\mathrm{thresh}}$ [31].", + "bbox": [ + 496, + 176, + 890, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Evaluation", + "text_level": 1, + "bbox": [ + 498, + 282, + 733, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the proposed method in TTC estimation and optical flow estimation, given that the apparent velocity on the image plane can be estimated according to Eq. (9) and due to the lack of event datasets dedicated to TTC estimation. The optical flow benchmark also provides a common ground to compare the proposed method with other state-of-the-art methods that estimate optical flow. Tab. 1 presents the hyper-parameters that were used across the experiments. Since the proposed method computes per-event estimates, we only evaluate on the respective pixel locations1.", + "bbox": [ + 496, + 308, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 498, + 470, + 697, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VL Dataset $[28]^2$ . It consists of 7 real event sequences observing planar prints of landing surfaces and 1 real event sequence observing the 3D print of a landing surface. Each sequence has 15sec of duration, totaling 120sec of dataset duration. The events were recorded by a Prophesee event camera with $1280 \\times 720$ resolution, and the ground truth (GT) depth measurements were recorded with an Intel RealSense camera. However, only a global GT depth measurement is provided per timestamp.", + "bbox": [ + 496, + 492, + 890, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To comply with the evaluation reported [28], we assess the proposed method at certain timestamps that correspond to event batches of 0.5sec. The comparison metrics are the divergence REE (\\%) and runtime per event (microsec). The divergence $\\mu$ is the inverse of the TTC, i.e., $\\mu = 1 / \\tau$ , and the REE is given by:", + "bbox": [ + 496, + 630, + 890, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R E E} = 1 0 0 \\frac {\\left| \\hat {\\mu} - \\mu_ {\\mathrm {g t}} \\right|}{\\left| \\mu_ {\\mathrm {g t}} \\right|}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 731, + 890, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mu}$ is the estimate and $\\mu_{\\mathrm{gt}}$ is the GT. The global motion prior $l_{\\theta}$ was set to 1000.", + "bbox": [ + 496, + 776, + 890, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MVSEC Dataset $[50]^3$ . It consists of several real indoor and outdoor sequences, providing events, standard grayscale frames, IMU data, camera poses, and scene depth.", + "bbox": [ + 496, + 806, + 890, + 852 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1 All the code will be made available upon acceptance.", + "bbox": [ + 514, + 862, + 805, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/s-mcleod/ventral-landing-event-dataset.", + "bbox": [ + 517, + 875, + 833, + 887 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3 https://daniilidis-group.github.io/mvsec.", + "bbox": [ + 519, + 887, + 736, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23657", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7fa15892468d64f7891765992992d5cb92140aee79571fa6f2d6aafb6350a99d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method2D-12D-22D-32D-42D-52D-62D-73DAvg.
REENCECMD [28]13.487.416.1112.1910.415.123.7212.908.92
||Proposed Scaling7.198.799.1312.184.987.794.9211.348.29
SProposed Translation5.037.026.175.254.614.346.186.545.64
3Proposed 6-DOF6.7713.7314.977.506.9112.1512.776.1010.11
3Proposed Scaling14.9915.1911.6921.6911.1915.4411.0119.4315.08
||Proposed Translation11.157.017.9813.806.3711.174.5512.469.31
SProposed 6-DOF6.516.817.256.776.065.596.186.196.42
RuntimeNCECMD [28]63.3762.5151.2141.3692.6938.9725.5731.5350.90
||Proposed Scaling0.820.780.800.810.700.840.780.700.78
SProposed Translation1.101.191.221.081.021.211.221.161.15
3Proposed 6-DOF1.401.431.451.441.371.541.601.471.46
3Proposed Scaling1.181.231.251.261.121.291.261.751.29
||Proposed Translation1.771.701.801.841.681.711.711.701.74
SProposed 6-DOF2.432.402.362.372.482.342.232.272.36
", + "bbox": [ + 158, + 88, + 812, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The events were recorded by a DAVIS [6] with $346 \\times 260$ resolution. The evaluated sequences span approximately 265sec. The optical flow GT is also provided [51], generated from the scene depth and camera velocity. We generate GT TTCM's by applying Eq. (1) given the GT depth maps and camera velocity.", + "bbox": [ + 75, + 378, + 468, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess optical flow accuracy, we use the following metrics: average endpoint error (AEE) (in pixel/frame, as is conventional in the literature [42, 51])", + "bbox": [ + 75, + 469, + 468, + 513 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A E E} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| \\hat {\\mathbf {m}} _ {i} - \\mathbf {m} _ {\\mathrm {g t}, i} \\right| _ {2}, \\mathbf {m} = \\Delta t _ {\\text {f r a m e}} \\mathbf {v}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 523, + 468, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "outliers (Out) as the percentage of pixels with AEE greater than 3, average relative endpoint error (AREE) $(\\%)$ [27]", + "bbox": [ + 75, + 575, + 468, + 607 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A R E E} = \\frac {1 0 0}{N} \\sum_ {i = 1} ^ {N} \\frac {\\left| \\hat {\\mathbf {m}} _ {i} - \\mathbf {m} _ {\\mathrm {g t} , i} \\right| _ {2}}{\\left| \\mathbf {m} _ {\\mathrm {g t} , i} \\right| _ {2}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 617, + 468, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and average angular error (AAE) $(^{\\circ})$ [27]", + "bbox": [ + 76, + 669, + 352, + 686 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A A E} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\operatorname {a r c c o s} \\frac {\\hat {\\mathbf {m}} _ {i} ^ {\\top} \\mathbf {m} _ {\\mathrm {g t} , i}}{| \\hat {\\mathbf {m}} _ {i} | _ {2} | \\mathbf {m} _ {\\mathrm {g t} , i} | _ {2}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 695, + 468, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess TTCM accuracy, we use the AREE $(\\%)$ between the divergence estimate $\\hat{\\mu}_i$ and corresponding GT $\\mu_{\\mathrm{gt},i}$ . The global motion prior $l_{\\theta}$ was set to 100.", + "bbox": [ + 75, + 747, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Other Global Motion Models", + "text_level": 1, + "bbox": [ + 76, + 801, + 336, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the general 6-DOF global motion model described in Sec. 3.3, we can consider other more constrained motion models depending on the application, as follows.", + "bbox": [ + 75, + 824, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Translation. This model is parameterized by the 3D linear velocities $\\pmb{\\nu}$ . Thus, $\\mathbf{B}_i = \\rho_i\\mathbf{V}_i$ , being $\\mathbf{V}_i$ given by Eq. (9).", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5e9337a16cfcb329198ac3b7c644b760b8cd8e72f3714e8f1018b43d4dae9e34.jpg", + "table_caption": [ + "Table 2. Global divergence estimation. Quantitative results on the VL dataset [28], in terms of REE (\\%) and runtime per event (microsec) averaged over 100 trials. Lower is better." + ], + "table_footnote": [], + "table_body": "
Modelindoor_flying1indoor_flying2outdoor_day1
Driving153.3084.8047.27
Translation19.8228.2345.78
6-DOF53.6556.1256.66
", + "bbox": [ + 504, + 375, + 888, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Divergence estimation. Quantitative results on the MVSEC dataset [50], in terms of AREE (\\%). Lower is better.", + "bbox": [ + 498, + 449, + 892, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Driving. This model is parameterized by the most significant DOF's when driving a car, namely the angular velocity around the camera $y$ -axis $\\omega_{y}$ and the linear velocity in the $z$ -axis $\\nu_{z}$ (see Fig. 2). Hence, $\\mathbf{B}_i = \\left( \\begin{array}{cc} \\rho_i(x_i - c_x) & -f_x - \\frac{(x_i - c_x)^2}{f_x} \\\\ \\rho_i(y_i - c_y) & -\\frac{(x_i - c_x)(y_i - c_y)}{f_x} \\end{array} \\right)$ . When using this motion model, we impose $|\\nu_z|_2 = 1$ to ensure that the depth is properly estimated.", + "bbox": [ + 498, + 505, + 890, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scaling. This model is parameterized by the linear velocity in the $z$ -axis $\\nu_{z}$ . Hence, $\\mathbf{B}_i = \\rho_i\\left( \\begin{array}{c}x_i - c_x\\\\ y_i - c_y \\end{array} \\right)$ . It is only considered since the global motion model used in [28] is the 1-DOF scaling, which assumes that the scene is planar and it does not estimate the depth.", + "bbox": [ + 498, + 636, + 890, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results", + "text_level": 1, + "bbox": [ + 500, + 739, + 591, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Time-to-Contact. Tab. 2 reports the results on global divergence estimation on the VL benchmark [28]. The proposed method using the Translation model with neighboring size $s = 2$ achieves the best accuracy, outperforming ECMD [28] by $36.77\\%$ on average. The proposed method using the Scaling and full 6-DOF models also outperform ECMD [28] in terms of average accuracy with neighboring size $s = 2$ and $s = 3$ , respectively. The results indicate that considering a smaller neighborhood benefits global mod", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23658", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e30d9f8e74d5418c6363cae011c3b9dee6db5549a71b141cf0bcb14d28b08e67.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodindoor_flying1indoor_flying2indoor_flying3outdoor_day1
AEEOutAEEOutAEEOutAEEOut
LBEV-FlowNet (MB) [52]0.580.001.024.000.873.000.320.00
EV-FlowNet (HQF) [46]0.561.000.661.000.591.000.681.00
Ding et al. [11]0.570.100.791.600.721.300.420.00
ConvGRU-EV-FlowNet [19]0.600.511.178.060.935.640.470.25
MBFBBrebion et al. [7]0.520.100.985.500.712.100.530.20
BBShiba et al. [42]0.420.100.600.590.500.280.300.10
EBAung et al. [3]----2.31-1.26-
ARMS [1]1.52-1.59-1.89-2.75-
Proposed Driving1.407.502.4226.322.1426.050.390.12
Proposed Translation0.630.410.942.830.791.650.550.97
Proposed 6-DOF0.881.971.6012.021.389.990.812.26
", + "bbox": [ + 173, + 88, + 802, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "els with fewer parameters; conversely, a larger neighborhood benefits models with more parameters. Even though the motion for all sequences is predominantly dominated by just 1 scaling DOF, the results suggest that considering motion models with additional DOFs, e.g., Translation and 6-DOF, is beneficial. The extra DOFs may explain other small motions, whereas these small motions would just be considered noise for the Scaling model.", + "bbox": [ + 75, + 344, + 470, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In terms of runtime per event, the proposed method outperforms ECMD [28] by between $95.36\\%$ and $98.47\\%$ , achieving real-time processing for all the sequences in the VL dataset [28], being capable of processing between 420k and 1.28M events per second. The results indicate that the runtime increases with the neighboring size $s$ and the number of parameters of the global motion model.", + "bbox": [ + 75, + 469, + 468, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 3 reports the results on divergence estimation on the MVSEC benchmark [50]. The Translation model achieves the best performance overall, indicating that the corresponding DOFs are sufficient to explain the perceived motions in the sequences evaluated while minimizing the optimization complexity. As mentioned in Sec. 4.2, the Driving model is tuned to the outdoor_day1 sequence, whose performance is on par with the Translation model. However, it performs poorly on the indoor_flying3 sequence since it can not handle more complex types of motions that are present. Being the most general model, the 6-DOF motion model achieves similar performance for all the sequences. Its performance is worse than the Translation model due to the increased optimization complexity, while the additional DOFs do not contribute to improving the accuracy. The proposed method achieves real-time processing in the MVSEC dataset [50] since, on average, the sequences exhibit a maximum of around 400k events per second.", + "bbox": [ + 75, + 579, + 470, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optical Flow. Since it is difficult to compare directly the results reported in Tab. 3, we evaluate the proposed method on optical flow estimation, and compare the results with other", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/89e651a1eadbae73151fc9f5d4f1fd66b2f3d480f5ebc0d2a64ea2df147edf02.jpg", + "table_caption": [ + "Table 4. Optical flow estimation. Quantitative results on the MVSEC dataset [50], in terms of AEE (pixel/frame) and Out (\\%). Lower is better. Legend: learning-based (LB), model-based (MB), frame-based (FB), batch-based (BB), event-based (EB)." + ], + "table_footnote": [], + "table_body": "
Methodindoor_flying3outdoor_day1
AREEAAEAREEAAE
EV-FlowNet [51]63.6132.5565.5422.88
Aung et al. [3]95.6469.35122.1069.41
ABMOF [26]52.6916.5790.1333.76
EDFLOW [27]37.5212.1569.4023.30
Proposed Driving104.0560.3650.2513.31
Proposed Translation44.4611.3872.6721.77
Proposed 6-DOF95.2321.35122.0431.68
", + "bbox": [ + 514, + 340, + 879, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Optical flow estimation. Quantitative results on the MVSEC dataset [50], in terms of AREE (\\%) and AAE (\\%). Lower is better.", + "bbox": [ + 498, + 481, + 893, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods. Tab. 4 reports the results on optical flow estimation on the MVSEC benchmark [50] in terms of AEE and Out, as is commonly found in the literature. The proposed method achieves state-of-the-art performance over the EB methods, on par performance with the FB method, and competitive performance overall.", + "bbox": [ + 496, + 556, + 890, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 5 reports the results on optical flow estimation on the MVSEC benchmark [50] in terms of AREE and AAE. The proposed method achieves state-of-the-art performance on the outdoor_day1 sequence using the Driving model and on par performance on the indoor_flying3 sequence using the Translation model. The results on Tab. 5 indicate that the proposed method is comparatively more accurate in estimating the direction of the flow, i.e., compared with the other methods, the proposed method achieves lower values of AAE overall. In terms of accuracy, this suggests that more improvements can be achieved by improving the method's per-event inverse depth estimation since inverse depth estimates mainly contribute to the flow magnitude.", + "bbox": [ + 496, + 652, + 892, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 6 presents qualitative results on sequences of the MVSEC dataset [50]. The estimated TTCM and optical flow resemble the GT ones.", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23659", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d20894387040a86101b5cf6f7fc906801d12351d33524eb5ce5d8e3592aa7dd5.jpg", + "image_caption": [ + "Figure 6. Qualitative results on the MVSEC dataset [50]." + ], + "image_footnote": [], + "bbox": [ + 112, + 92, + 854, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitations", + "text_level": 1, + "bbox": [ + 76, + 513, + 204, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similarly to other event-based methods [15, 22, 35, 52], the proposed method also inherits the brightness assumption from the DMin framework [34]. It can thus provide wrongful estimates for events that are not caused by motion, e.g., due to flickering lights. While this limitation is somewhat mitigated when estimating global quantities, it can struggle to reliably estimate the inverse depth of events that are not caused by motion. Introducing probabilistic uncertainties to the estimates could alleviate this issue while also improving the inverse depth's initialization procedure.", + "bbox": [ + 75, + 537, + 468, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Also related, although we explicitly impose a local smoothness constraint to the inverse depth estimates, given by Eq. (10), the proposed method can still provide inverse depth estimates that differ significantly from the neighboring inverse depth estimates. This typically occurs for events that are generated by noise. Carefully tuning the event camera biases and filtering out outliers could improve the overall inverse depth estimation. Having a back-end inverse depth regularizer [22] could also help to mitigate this issue.", + "bbox": [ + 75, + 688, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed method can only estimate one global motion. Thus, it can not adequately handle more than one motion simultaneously, e.g., due to cars moving [50]. Considering a multi-scale approach [1, 42] or explicitly modeling more than one possible motion occurring simultane", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ously [33, 44] are possible avenues for future research.", + "bbox": [ + 498, + 513, + 861, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 542, + 619, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed a novel method that estimates the TTCM using a single event camera. The proposed method builds on the DMin framework to incrementally estimate local and global quantities, i.e., inverse depth, and global motion, respectively. We have also proposed an approach that effectively prevents event collapse for incremental event-based estimation without introducing regularizers or additional hyper-parameters. The proposed method also achieves state-of-the-art performance in TTC estimation in terms of accuracy and computational runtime while achieving competitive performance in optical flow estimation. Broadly, the proposed work further builds on the increasing amount of evidence that event cameras are especially suited to address visual motion-based problems; in particular, it further shows that incremental event-based processing can provide a flexible and general methodology to consider when using event cameras, which avoids issues introduced when converting events to other representations, e.g., batches and frames.", + "bbox": [ + 496, + 568, + 890, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This research received funding from the French National Research Agency (ANR), under grant agreement $\\mathrm{N}^{\\circ}$ ANR-20-CE23-0021, \"AgileNeuRobot\".", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23660", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Himanshu Akolkar, Sio Hoi Ieng, and Ryad Benosman. Real-time High Speed Motion Prediction Using Fast Aperture-Robust Event-Driven Visual Flow. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–1, 2020. 2, 7, 8", + "[2] Mohammed Almatrafi, Raymond Baldwin, Kiyoharu Aizawa, and Keigo Hirakawa. Distance Surface for Event-Based Optical Flow. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1547-1556, 2020. 2", + "[3] Myo Tun Aung, Rodney Teo, and Garrick Orchard. Event-based Plane-fitting Optical Flow for Dynamic Vision Sensors in FPGA. In IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-5, 2018. 2, 7", + "[4] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-Based Visual Flow. IEEE Transactions on Neural Networks and Learning Systems, 25(2):407-417, 2014. 1, 2", + "[5] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous Frameless Event-based Optical Flow. Neural Networks, 27:32-37, 2012. 2", + "[6] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \\times 180$ dB 3 μs Latency Global Shutter Spatiotemporal Vision Sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1, 6", + "[7] Vincent Brebion, Julien Moreau, and Franck Davoine. Real-Time Optical Flow for Vehicular Perception With Low- and High-Resolution Event Cameras. IEEE Transactions on Intelligent Transportation Systems, 23(9):15066-15078, 2022. 2, 7", + "[8] William Chamorro, Juan Andrade-Cetto, and Joan Solà. High-Speed Event-Based Camera Tracking. In British Machine Vision Conference (BMVC), 2020. 2", + "[9] William Chamorro, Joan Solà, and Juan Andrade-Cetto. Event-Based Line SLAM in Real-Time. IEEE Robotics and Automation Letters, 7(3):8146-8153, July 2022. 2", + "[10] Xavier Clady, Charles Clercq, Sio-Hoi Ieng, Fouzhan Houseini, Marco Randazzo, Lorenzo Natale, Chiara Bartolozzi, and Ryad Benosman. Asynchronous Visual Event-based Time-to-Contact. Frontiers in Neuroscience, 8, 2014. 1, 2", + "[11] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-Temporal Recurrent Networks for Event-Based Optical Flow Estimation. In AAAI Conference on Artificial Intelligence, volume 36, pages 525-533, 2022. 3, 7", + "[12] Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dynamic obstacle avoidance for quadrotors with event cameras. Science Robotics, 5(40):eaaz9712, 2020. 1, 2", + "[13] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-Based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):154-180, 2022. 1, 2", + "[14] Guillermo Gallego, Jon E.A. Lund, Elias Mueggler, Henri Rebecq, Tobi Delbruck, and Davide Scaramuzza. Event-" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Based, 6-DOF Camera Tracking from Photometric Depth Maps. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(10):2402-2412, 2018. 2", + "[15] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A Unifying Contrast Maximization Framework for Event Cameras, with Applications to Motion, Depth, and Optical Flow Estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3867-3876, 2018. 1, 2, 3, 8", + "[16] Daniel Gehrig, Antonio Loquercio, Konstantinos Derpanis, and Davide Scaramuzza. End-to-End Learning of Representations for Asynchronous Event-Based Data. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5632–5642, 2019. 2", + "[17] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense Optical Flow from Event Cameras. In International Conference of 3D Vision (3DV), pages 197–206, 2021. 1", + "[18] Germain Haessig, Xavier Berthelon, Sio-Hoi Ieng, and Ryad Benosman. A Spiking Neural Network Model of Depth from Defocus for Event-based Neuromorphic Vision. Scientific Reports, 9(1):3744, 2019. 2", + "[19] Jesse Hagenaars, Federico Paredes-Valles, and Guido de Croon. Self-Supervised Learning of Event-Based Optical Flow with Spiking Neural Networks. In Advances in Neural Information Processing Systems (NIPS), volume 34, pages 7167–7179, 2021. 3, 7", + "[20] Botao He, Haojia Li, Siyuan Wu, Dong Wang, Zhiwei Zhang, Qianli Dong, Chao Xu, and Fei Gao. FAST-Dynamic-Vision: Detection and Tracking Dynamic Objects with Event and Depth Sensing. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3071-3078, 2021. 1", + "[21] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew Davison. Simultaneous Mosaicing and Tracking with an Event Camera. In British Machine Vision Conference (BMVC). British Machine Vision Association (BMVA), 2014. 1, 2", + "[22] Hanme Kim, Stefan Leutenegger, and Andrew J. Davison. Real-Time 3D Reconstruction and 6-DoF Tracking with an Event Camera. In Computer Vision – ECCV, Lecture Notes in Computer Science, pages 349–364. Springer International Publishing, 2016. 1, 2, 8", + "[23] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E. Shi, and Ryad B. Benosman. HOTS: A Hierarchy of Event-Based Time-Surfaces for Pattern Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(7):1346-1359, July 2017. 1", + "[24] Chankyu Lee, Adarsh Kumar Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, and Kaushik Roy. Spike-FlowNet: Event-Based Optical Flow Estimation with Energy-Efficient Hybrid Neural Networks. In Computer Vision - ECCV, Lecture Notes in Computer Science, pages 366-382. Springer International Publishing, 2020. 3", + "[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \\times 128$ dB 15 μs Latency Asynchronous Temporal Contrast Vision Sensor. IEEE Journal of Solid-State Circuits, 43(2):566-576, 2008. 1" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23661", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Min Liu and T. Delbruck. Adaptive Time-Slice Block-Matching Optical Flow Algorithm for Dynamic Vision Sensors. In British Machine Vision Conference (BMVC), 2018, 2, 7", + "[27] Min Liu and Tobi Delbruck. EDFLOW: Event Driven Optical Flow Camera With Keypoint Detection and Adaptive Block Matching. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5776-5789, 2022. 1, 2, 3, 6, 7", + "[28] Sofia McLeod, Gabriele Meoni, Dario Izzo, Anne Mergy, Daqi Liu, Yasir Latif, Ian Reid, and Tat-Jun Chin. Globally Optimal Event-Based Divergence Estimation for Ventral Landing, 2022. 1, 2, 5, 6, 7, 9, 10", + "[29] Moritz B. Milde, Olivier J.N. Bertrand, Ryad Benosmanz, Martin Egelhaaf, and Elisabetta Chicca. Bioinspired event-driven collision avoidance algorithm based on optic flow. In International Conference on Event-based Control, Communication, and Signal Processing (EBCCSP), pages 1-7, 2015. 1, 2", + "[30] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The Event-Camera Dataset and Simulator: Event-based Data for Pose Estimation, Visual Odometry, and SLAM. The International Journal of Robotics Research, 36(2):142-149, 2017. 4", + "[31] Urbano Miguel Nunes, Ryad Benosman, and Sio-Hoi Ieng. Adaptive Global Decay Process for Event Cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9771-9780, 2023. 5", + "[32] Urbano Miguel Nunes and Yiannis Demiris. Entropy Minimisation Framework for Event-based Vision Model Estimation. In Computer Vision - ECCV, Lecture Notes in Computer Science, pages 161-176. Springer International Publishing, 2020. 3", + "[33] Urbano Miguel Nunes and Yiannis Demiris. Kinematic Structure Estimation of Arbitrary Articulated Rigid Objects for Event Cameras. In International Conference on Robotics and Automation (ICRA), pages 508-514, 2022. 1, 8", + "[34] Urbano Miguel Nunes and Yiannis Demiris. Robust Event-Based Vision Model Estimation by Dispersion Minimisation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9561-9573, 2022. 1, 2, 3, 8", + "[35] Federico Paredes-Vallés and Guido C. H. E. de Croon. Back to Event Basics: Self-Supervised Learning of Image Reconstruction for Event Cameras via Photometric Constancy. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3445-3454, 2021. 3, 8", + "[36] Bas J. Pijnacker Hordijk, Kirk Y. W. Scheper, and Guido C. H. E. de Croon. Vertical Landing for Micro Air Vehicles using Event-based Optical Flow. Journal of Field Robotics, 35(1):69-90, 2018. 1, 2", + "[37] Christoph Posch, Teresa Serrano-Gotarredona, Bernabe Linares-Barranco, and Tobi Delbruck. Retinomorphic Event-Based Vision Sensors: Bioinspired Cameras With Spiking Output. Proceedings of the IEEE, 102(10):1470-1484, 2014. 2", + "[38] Henri Rebecq, Timo Horstschaefer, Guillermo Gallego, and Davide Scaramuzza. EVO: A Geometric Approach to Event-Based 6-DOF Parallel Tracking and Mapping in Real Time." + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Robotics and Automation Letters, 2(2):593-600, 2017. 2", + "[39] Juan Pablo Rodríguez-Gómez, Raul Tapia, Maria del Mar Guzmán Garcia, Jose Ramiro Martínez-de Dios, and Anibal Ollero. Free as a Bird: Event-Based Dynamic Sense-and-Avoid for Ornithopter Robot Flight. IEEE Robotics and Automation Letters, 7(2):5413-5420, 2022. 1, 2", + "[40] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event Collapse in Contrast Maximization Frameworks. Sensors, 22(14):5190, 2022. 1, 2, 3", + "[41] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. A Fast Geometric Regularizer to Mitigate Event Collapse in the Contrast Maximization Framework. Advanced Intelligent Systems, 2022. 1, 2, 3", + "[42] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of Event-Based Optical Flow. In Computer Vision – ECCV, Lecture Notes in Computer Science, pages 628–645. Springer Nature Switzerland, 2022. 2, 6, 7, 8", + "[43] Olaf Sikorski, Dario Izzo, and Gabriele Meoni. Event-based Spacecraft Landing using Time-to-Contact. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1941–1950, 2021. ISSN: 2160-7516. 1, 2", + "[44] Timo Stoffregen, Guillermo Gallego, Tom Drummond, Lindsay Kleeman, and Davide Scaramuzza. Event-Based Motion Segmentation by Motion Compensation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 7243-7252, 2019. 1, 8", + "[45] Timo Stoffregen and Lindsay Kleeman. Simultaneous Optical Flow and Segmentation (SOFAS) using Dynamic Vision Sensor. In Australasian Conference on Robotics and Automation (ACRA), pages 52-61. Australian Robotics and Automation Association (ARAA), 2017. 1", + "[46] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the Sim-to-Real Gap for Event Cameras. In Computer Vision - ECCV, Lecture Notes in Computer Science, pages 534-549. Springer International Publishing, 2020. 3, 7", + "[47] Celyn Walters and Simon Hadfield. EVReflex: Dense Timeto-Impact Prediction for Event-based Obstacle Avoidance. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1304-1309, 2021. 1, 2", + "[48] Chengxi Ye, Anton Mitrokhin, Cornelia Fermüller, James A. Yorke, and Yiannis Aloimonos. Unsupervised Learning of Dense Optical Flow, Depth and Egomotion with Event-Based Sensors. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5831-5838, 2020. 3", + "[49] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based Feature Tracking with Probabilistic Data Association. In IEEE International Conference on Robotics and Automation (ICRA), pages 4465-4470, 2017. 2", + "[50] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The Multivehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018. 5, 6, 7, 8" + ], + "bbox": [ + 501, + 90, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23662", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. EV-FlowNet: Self-Supervised Optical Flow Estimation for Event-based Cameras. In Proceedings of Robotics: Science and Systems, 2018. 1, 6, 7", + "[52] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised Event-Based Learning of Optical Flow, Depth, and Egomotion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 989-997, 2019. 2, 3, 7, 8" + ], + "bbox": [ + 78, + 90, + 470, + 218 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23663", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/0f2857ef-2537-43f4-b49d-a3319a025266_model.json b/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/0f2857ef-2537-43f4-b49d-a3319a025266_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a9ee6b1c20335bcad555c06e1d26ba473159c7a5 --- /dev/null +++ b/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/0f2857ef-2537-43f4-b49d-a3319a025266_model.json @@ -0,0 +1,2433 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.13, + 0.868, + 0.177 + ], + "angle": 0, + "content": "Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.203, + 0.744, + 0.223 + ], + "angle": 0, + "content": "Urbano Miguel Nunes1, Laurent Udo Perrinet2 and Sio-Hoi Ieng1" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.224, + 0.856, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Sorbonne University, \\(^{2}\\)Aix Marseille Univ, CNRS, INT, Institut de Neurosciences de la Timone urbano.goncalves-nunes@inserm.fr, laurent.perrinet@univ-amu.fr, sio-hoi.ieng@upmc.fr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.614 + ], + "angle": 0, + "content": "Event cameras asynchronously report brightness changes with a temporal resolution in the order of microseconds, which makes them inherently suitable to address problems that involve rapid motion perception. In this paper, we address the problem of time-to-contact (TTC) estimation using a single event camera. This problem is typically addressed by estimating a single global TTC measure, which explicitly assumes that the surface/obstacle is planar and fronto-parallel. We relax this assumption by proposing an incremental event-based method to estimate the TTC that jointly estimates the (up-to scale) inverse depth and global motion using a single event camera. The proposed method is reliable and fast while asynchronously maintaining a TTC map (TTCM), which provides per-pixel TTC estimates. As a side product, the proposed method can also estimate per-event optical flow. We achieve state-of-the-art performances on TTC estimation in terms of accuracy and runtime per event while achieving competitive performance on optical flow estimation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.647, + 0.21, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Event cameras differ from standard frame-based cameras, which capture visual data at a fixed rate and independently of the observing scene. Instead, event cameras respond asynchronously to pixel-wise brightness changes by generating events [6, 25]. Event cameras are thus data-driven sensors that offer several advantages, including high temporal resolution in the order of microseconds, low latency, low power consumption, and high dynamic range. These properties place event cameras as suitable candidates to address vision-based problems that involve (high-speed) motion, e.g., optical flow estimation [4, 27, 51], ego-motion estimation [15, 21, 22, 34], motion segmentation [33, 44], and obstacle avoidance [10, 12]. Due to the distinct visual sensing paradigm, however, new methods are necessary to fully exploit the potential of event cameras [13, 23]." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.293, + 0.883, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.475, + 0.893, + 0.504 + ], + "angle": 0, + "content": "Figure 1. Runtime vs. accuracy comparison for TTC estimation methods. Average results on the Ventral Landing benchmark [28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.893, + 0.702 + ], + "angle": 0, + "content": "Event cameras have been used to address the problem of fast TTC estimation, which comes up often in the vision-based obstacle avoidance [10, 12, 29, 39] and ventral landing [28, 36] literature. The TTC is the time that would elapse before a camera reaches an obstacle/surface, assuming the current relative motion between them remains constant [10]. Previous methods that use a single event camera have focused on estimating a single global TTC measure, which assumes that the surface is planar and fronto-parallel. To overcome this limitation, other methods use additional sensing, e.g., depth frames, to build a dense TTCM [20, 47]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We instead propose to extend the Dispersion Minimization (DMin) framework [34] to estimate the TTC for each incoming event using a single event camera. The proposed method jointly estimates the relative global motion and per-event (up-to scale) inverse depth. We can then asynchronously maintain a semi-dense TTCM which provides per-pixel TTC estimates or compute a global TTC measure with greater accuracy by averaging over the TTC estimates. Since there is at least one scaling degree of freedom (DOF), we also propose an effective strategy to mitigate event collapse [40, 41]. The proposed method is also computationally fast, reaching \\(\\sim 1\\) microsecond processing time per event on a standard laptop. Fig. 1 compares the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23653" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "runtime vs. accuracy for TTC estimation methods, whereby the proposed method achieves state-of-the-art performance. We also estimate the per-event optical flow as a side product and achieve competitive performance compared to state-of-the-art optical flow methods that use events." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.168, + 0.223, + 0.182 + ], + "angle": 0, + "content": "Main contributions:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.469, + 0.228 + ], + "angle": 0, + "content": "1. First event-based method that explicitly estimates the TTC for each event and maintains a semi-dense TTCM using a single event camera." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.229, + 0.468, + 0.273 + ], + "angle": 0, + "content": "2. DMin framework [34] extension to jointly handle local and global estimates, i.e., inverse depth and global motion, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.275, + 0.468, + 0.304 + ], + "angle": 0, + "content": "3. Effective approach that mitigates event collapse [40, 41] for incremental event-based estimation." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.184, + 0.469, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.316, + 0.241, + 0.331 + ], + "angle": 0, + "content": "1.1. Time-to-Contact" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.466 + ], + "angle": 0, + "content": "Consider a freely moving camera with angular and linear velocities \\(\\omega(t) = (\\omega_x(t), \\omega_y(t), \\omega_z(t))^{\\top}\\) and \\(\\pmb{\\nu}(t) = (\\nu_x(t), \\nu_y(t), \\nu_z(t))^{\\top}\\), respectively, that is observing a point \\(\\alpha\\), with 3D coordinates \\(\\pmb{\\alpha}(t) = (X(t), Y(t), Z(t))^{\\top}\\) relative to the camera, as shown in Fig. 2. \\(Z(t)\\) is also referred as the depth of point \\(\\alpha\\) relative to the camera. The instantaneous TTC between the moving camera and point \\(\\alpha\\) is thus given by:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.467, + 0.469, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\tau (t) := - \\frac {Z (t)}{\\frac {d Z (t)}{d t}} = \\frac {Z (t)}{\\nu_ {z} (t)}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.469, + 0.601 + ], + "angle": 0, + "content": "The minus sign disappears because we define the linear velocity w.r.t. the camera's frame of reference, not w.r.t. the point's frame of reference, i.e., \\(\\nu_{z}(t) = -dZ(t) / dt\\). Based on Eq. (1), the exact values of depth and relative approaching motion do not need to be estimated, only the ratio between them." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.618, + 0.219, + 0.633 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.688 + ], + "angle": 0, + "content": "We review recent related works on the following topics: TTC, global motion and optical flow estimation. We refer to [13] for a detailed survey." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Time-to-Contact Estimation. The first work on event-based TTC using a single event camera relied solely on the estimation of visual motion flows [10], whereby the motion flows were computed by fitting a local plane to the time surface [4]. Other works that followed were geared towards two main use cases, namely obstacle avoidance [12, 29, 39] and ventral landing [28, 36, 43]. Event-based obstacle avoidance methods are built to be fast reacting and, although they come from either bio-inspired [29, 39] or mathematically grounded principles [12], they typically rely on empirically-validated heuristics to speed-up computations. Existing event-based ventral landing approaches only compute a single TTC estimate, which assumes that the surface is planar and fronto-parallel. To overcome this assumption," + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.088, + 0.865, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.225, + 0.807, + 0.24 + ], + "angle": 0, + "content": "Figure 2. Camera coordinate system." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.326 + ], + "angle": 0, + "content": "other works fuse events with additional sensory information, e.g. depth [47]. The proposed approach uses a single event camera while being mathematically grounded and computationally fast." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.893, + 0.689 + ], + "angle": 0, + "content": "Global Motion Estimation. Also denoted by ego-motion estimation [13], it refers to estimating the parameters that explain the triggered events according to some global motion model. These methods can be broadly characterized by whether they rely on key-frame registration [8, 9, 14, 21, 22, 38] or perform the estimation without relying on any key-frame [15, 34]. The former methods are reminiscent of the frame-based paradigm and also include methods based on artificial neural networks (ANN) [16, 52], whereby an intermediate frame-based representation is needed for the estimation. The latter methods tend towards a more event-based processing paradigm and include methods based on spiking neural networks (SNN) [18, 37], whereby events are either processed on an event-by-event basis [34] or in batches [15]. Although both approaches have advantages and disadvantages, methods that rely on key-frame registration typically require an intermediate frame-based representation, which is still an open problem in the event-based community. Similarly to [22], the proposed method jointly estimates up-to scale inverse depth and global motion. By building on the DMin framework [34], which also allows processing events on an event-by-event basis, our method does not rely on key-frame registration or background inverse depth regularization to improve convergence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Optical Flow Estimation. Several model-based methods have been previously proposed, which can be further divided into: frame-based [2, 7, 26, 27], batch-based [42, 49], and event-based [1, 3, 4]. By selecting the most relevant events, i.e., typically the most recent, frame-based methods build frames from which the optical flow is computed using techniques from standard image-based optical flow, e.g., Lucas-Kanade [5]. Since each event does not carry much information on its own, batch-based methods aggregate the most recent events by forming batches but perform the computations directly on the events. Event-based methods follow the most event-driven paradigm by performing event-by-event processing, typically being the computationally fastest. However, event-based methods tend to suffer more" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23654" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.351 + ], + "angle": 0, + "content": "from the aperture problem since all the computations are performed locally, and thus frame-based and model-based methods achieve currently better accuracy. In terms of accuracy, learning-based ANN methods [11, 35, 46, 52] generally achieve state-of-the-art performance. Besides the need to convert events into frames for more efficient processing, these methods are known to be very data hungry, sensitive to the training data [48], and consume large amounts of energy [27]. Another line of research in learning-based methods has been to use SNN [19, 24], which combine the event-based processing and learning paradigms and thus do not require an intermediate frame-based representation. However, it is not trivial to train SNN, and the empirical validation is still not on par with ANN methods. Although it is not the primary objective of this work, the proposed method can provide per-event flow estimates while still being competitive in terms of accuracy w.r.t. state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.36, + 0.17, + 0.377 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.471, + 0.476 + ], + "angle": 0, + "content": "In this section, we describe the proposed incremental event-based method for TTCM estimation. We first briefly review the event cameras' working principle, and the DMin framework [34], based on which we develop the proposed method. Refer to the supplementary material for the full mathematical derivations and additional details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.483, + 0.462, + 0.5 + ], + "angle": 0, + "content": "3.1. Event Cameras and Dispersion Minimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Event cameras output a stream of asynchronous temporal contrast events \\(\\{e_i\\}, i \\in \\mathbb{N}\\). Each event \\(e_i\\) represents a spatio-temporal asynchronous brightness change, being defined as a tuple \\(e_i := (\\mathbf{x}_i, t_i, p_i)\\), where \\(\\mathbf{x}_i = (x_i, y_i)\\) are the pixel coordinates, \\(t_i\\) is the timestamp at which the event was generated, and \\(p_i \\in \\{-1, +1\\}\\) is its polarity. An event \\(e_i\\) is generated when the change in log-brightness \\(\\log \\mathbf{I}_{x,y}(t) := \\bar{\\mathbf{I}}_{x,y}(t)\\) is above a threshold \\(L\\)" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.633, + 0.47, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\left| \\Delta \\bar {\\mathbf {I}} _ {x _ {i}, y _ {i}} \\left(t _ {i}\\right) \\right| = \\left| \\bar {\\mathbf {I}} _ {x _ {i}, y _ {i}} \\left(t _ {i}\\right) - \\bar {\\mathbf {I}} _ {x _ {i}, y _ {i}} \\left(t _ {i} - \\Delta t _ {i}\\right) \\right| \\geq L, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.658, + 0.468, + 0.672 + ], + "angle": 0, + "content": "where \\(\\Delta t_{i}\\) is the time since the last event at the same pixel." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.672, + 0.47, + 0.749 + ], + "angle": 0, + "content": "The DMin framework [34] estimates the parameters \\(\\pmb{\\theta}\\) of a transformation model \\(\\mathcal{M}\\) from the stream of events \\(\\mathcal{E} = \\{e_i\\}_{i=1}^{N_e}\\) by minimizing a dispersion measure of the transformed events \\(\\mathbf{f}_i = \\mathcal{M}(e_i; t_{\\mathrm{ref}}, \\pmb{\\theta})\\). We consider the Potential measure with a Gaussian kernel \\(\\mathcal{N}(\\mathbf{x}; \\pmb{\\mu}, \\pmb{\\Sigma})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.755, + 0.469, + 0.798 + ], + "angle": 0, + "content": "\\[\nP (\\mathcal {E}; \\boldsymbol {\\theta}) = - \\sum_ {i, j} ^ {N _ {e}} \\mathcal {N} \\left(\\mathbf {f} _ {i}; \\mathbf {f} _ {j}, \\mathbf {I}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.803, + 0.47, + 0.879 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}\\) is the identity matrix. A key distinction of the DMin framework is that it allows to incrementally estimate the model parameters \\(\\theta\\) on an event-by-event basis, whereby the model parameters \\(\\theta\\) can be iteratively solved by linearizing the transformation model \\(\\mathcal{M}\\), such that:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.884, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {i} = \\mathcal {M} \\left(e _ {i}; t _ {\\mathrm {r e f}}, \\boldsymbol {\\theta}\\right) = \\tilde {\\mathbf {f}} _ {i} + \\Delta t _ {i, \\mathrm {r e f}} \\mathbf {B} _ {i} \\boldsymbol {\\theta}, \\mathbf {B} _ {i} := \\mathbf {B} \\left(e _ {i}\\right), (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.123 + ], + "angle": 0, + "content": "where \\(\\mathbf{B}_i\\) is the model-dependent linearization matrix. The parameters \\(\\theta^{*}\\) that minimize Eq. (3) are thus given by:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.129, + 0.892, + 0.182 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} ^ {*} = - \\left(\\sum_ {i, j} ^ {N _ {e}} \\mathbf {W} _ {i, j} \\Delta \\mathbf {C} _ {i, j}\\right) ^ {- 1} \\left(\\sum_ {i, j} ^ {N _ {e}} \\mathbf {W} _ {i, j} \\Delta \\tilde {\\mathbf {f}} _ {i, j}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.188, + 0.894, + 0.223 + ], + "angle": 0, + "content": "where \\(\\Delta \\mathbf{C}_{i,j} = \\mathbf{C}_i - \\mathbf{C}_j\\), \\(\\Delta \\tilde{\\mathbf{f}}_{i,j} = \\tilde{\\mathbf{f}}_i - \\tilde{\\mathbf{f}}_j\\), \\(\\mathbf{W}_{i,j} = w_{i,j}\\partial \\Delta \\mathbf{f}_{i,j}^{\\top} / \\partial \\pmb {\\theta}\\), \\(\\mathbf{C}_i = \\Delta t_{i,\\mathrm{ref}}\\mathbf{B}_i\\) and \\(w_{i,j} = \\mathcal{N}(\\mathbf{f}_i;\\mathbf{f}_j,\\mathbf{I})\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.228, + 0.799, + 0.244 + ], + "angle": 0, + "content": "3.2. Adapted Dispersion Minimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.553 + ], + "angle": 0, + "content": "While the DMin method [34] provides a general framework for global incremental event-based model estimation, it can also be adapted to jointly estimate global and local measures, i.e., global angular and linear velocities and local inverse depth. However, the DMin framework may encounter estimation issues when the global model has at least one scaling DOF, as noted in [34, 40, 41], known as event collapse. Event collapse occurs when the events are transformed into a single point, which minimizes the events' dispersion or maximizes the image contrast while the parameters' estimates diverge. So far, to the best of our knowledge, the mitigation discussion in the literature has been on how to constrain the optimization loss to discourage divergent estimates by analyzing the effects of the scaling transformations on the event-based data. Several mitigation strategies have thus been proposed on the events [34, 40] and parameters level [41] by adding terms to regularize the objective measure. While these strategies generally prevent event collapse, they increase the complexity of the optimization framework and introduce additional parameters to tune." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.81 + ], + "angle": 0, + "content": "We instead observe that event collapse fundamentally stems from the event transformation to a common time reference \\( t_{\\mathrm{ref}} \\), e.g., given by Eq. (4), by identifying two problems: 1) the constant velocity assumption may not hold depending on the time difference \\( \\Delta t_{i,\\mathrm{ref}} \\), and 2) there is no built-in constraint on the magnitude of the model parameters, e.g., such that the difference between transformed events \\( \\mathbf{f}_i - \\mathbf{f}_j \\) explicitly penalizes divergent estimates; although, according to Eq. (4), \\( \\mathbf{f}_i \\) and \\( \\mathbf{f}_j \\) individually diverge if the model parameters also diverge, their difference \\( \\mathbf{f}_i - \\mathbf{f}_j \\) is not guaranteed to diverge and thus penalize divergent estimates. The first problem is typically addressed by heuristically making the time difference as short as possible, and its effects are of limited significance in practice. The second problem, however, is intrinsically linked to batch-based processing since the alignment of the events in a batch needs to be measured in some common time reference [15, 32]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, for incremental event-based processing, the event transformation and, consequently, the dispersion measure can be modified without loss of generality such that the event collapse is prevented by implicitly addressing the two problems identified. Fig. 3 depicts the idea whereby we only transform the event neighbors to the current event's" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "23655" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.094, + 0.312, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.094, + 0.447, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.174, + 0.47, + 0.217 + ], + "angle": 0, + "content": "Figure 3. Incremental event collapse mitigation. Instead of transforming all the events to some time reference (left), we locally transform the events to the current event's timestamp (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.241, + 0.468, + 0.317 + ], + "angle": 0, + "content": "timestamp instead of transforming all the events to a common time reference, including the current event. Formally, instead of transforming the events according to Eq. (4), only the neighboring events \\( e_j \\) of the current event \\( e_i \\) are transformed according to" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.326, + 0.469, + 0.346 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {j} ^ {\\prime} = \\mathbf {x} _ {j} + \\Delta t _ {i, j} \\mathbf {B} _ {j} \\boldsymbol {\\theta} = \\mathbf {x} _ {j} + \\mathbf {C} _ {i, j} \\boldsymbol {\\theta}, e _ {j} \\in \\operatorname {n e i g h} (e _ {i}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.469, + 0.536 + ], + "angle": 0, + "content": "where 'neigh' is shorthand for neighborhood, and the dependency of \\(\\mathbf{x}_j^\\prime\\) on \\(\\pmb{\\theta}\\) was omitted for brevity. From the resultant residual \\(\\mathbf{r}_{i,j} = \\mathbf{x}_i - \\mathbf{x}_j^\\prime = \\Delta \\mathbf{x}_{i,j} - \\mathbf{C}_{i,j}\\pmb {\\theta}\\), we see that the proposed modification to the DMin framework addresses both identified problems. First, the constant velocity assumption is better held since the time difference satisfies \\(\\Delta t_{i,j}\\leq \\Delta t_{i,\\mathrm{ref}}\\). Second, if the model parameters \\(\\pmb{\\theta}\\) diverge, then the residual \\(\\mathbf{r}_{i,j}\\) also diverges, which effectively penalizes divergent estimates. We highlight that the proposed adaptation only works for incremental event-based processing: it is not suitable for batch-based processing since the proposed transformation only works locally." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.469, + 0.582 + ], + "angle": 0, + "content": "The Potential measure is modified by computing the difference between the current event's coordinates \\(\\mathbf{x}_i\\) and the neighboring events' transformed coordinates \\(\\mathbf{x}_j'\\):" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.593, + 0.469, + 0.627 + ], + "angle": 0, + "content": "\\[\nP \\left(\\mathcal {E}; \\boldsymbol {\\theta}\\right) = - \\sum_ {i, j} \\mathcal {N} \\left(\\mathbf {x} _ {j} ^ {\\prime}; \\mathbf {x} _ {i}, \\mathbf {I}\\right), e _ {j} \\in \\operatorname {n e i g h} \\left(e _ {i}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.469, + 0.682 + ], + "angle": 0, + "content": "By minimizing Eq. (7) by linearizing the residual according to Taylor's formula \\(\\mathbf{r}_{i,j}(\\pmb {\\theta} + \\Delta \\pmb {\\theta})\\approx \\mathbf{r}_{i,j}(\\pmb {\\theta}) + \\mathbf{J}_{i,j}(\\pmb {\\theta})\\Delta \\pmb{\\theta}\\), the optimized model parameters \\(\\pmb{\\theta}^{*}\\) are iteratively updated:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.707, + 0.469, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} ^ {*} \\leftarrow \\boldsymbol {\\theta} ^ {*} + \\Delta \\boldsymbol {\\theta}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.727, + 0.449, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\Delta \\boldsymbol {\\theta} = - \\underbrace {\\left(\\sum_ {j} w _ {i , j} \\mathbf {J} _ {i , j} ^ {\\top} \\mathbf {J} _ {i , j}\\right)} _ {\\boldsymbol {\\Psi}} ^ {- 1} \\underbrace {\\left(\\sum_ {j} w _ {i , j} \\mathbf {J} _ {i , j} ^ {\\top} \\mathbf {r} _ {i , j}\\right)} _ {\\boldsymbol {\\psi}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.807, + 0.468, + 0.84 + ], + "angle": 0, + "content": "where \\(w_{i,j} = \\mathcal{N}\\left(\\mathbf{x}_j';\\mathbf{x}_i,\\mathbf{I}\\right)\\), \\(\\mathbf{J}_{i,j} = \\partial \\mathbf{r}_{i,j} / \\partial \\pmb{\\theta}\\) and \\(e_j\\in\\) \\(\\mathrm{neigh}(e_i)\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.424, + 0.864 + ], + "angle": 0, + "content": "3.3. Inverse Depth and Global Motion Model" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We consider that a calibrated event camera can freely move and whose global motion is parameterized by the 3D" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.093, + 0.886, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.213, + 0.881, + 0.228 + ], + "angle": 0, + "content": "Figure 4. Typical per-event estimation of optical flow and depth." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.251, + 0.892, + 0.315 + ], + "angle": 0, + "content": "angular and linear velocities, as defined in Sec. 1.1, \\(\\pmb{\\theta} = (\\pmb{\\nu}^{\\mathsf{T}},\\pmb{\\omega}^{\\mathsf{T}})^{\\mathsf{T}}\\). For each event \\(e_i\\), we estimate its inverse depth \\(\\rho_{i}\\coloneqq 1 / Z_{i}\\), based on the well-known expression for the apparent velocity on the image plane:" + }, + { + "type": "equation", + "bbox": [ + 0.505, + 0.324, + 0.892, + 0.357 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {i} = \\left( \\begin{array}{l l} \\rho_ {i} \\mathbf {V} _ {i} & \\boldsymbol {\\Omega} _ {i} \\end{array} \\right) \\binom {\\boldsymbol {\\nu}} {\\omega} = \\rho_ {i} \\mathbf {V} _ {i} \\boldsymbol {\\nu} + \\boldsymbol {\\Omega} _ {i} \\boldsymbol {\\omega} := \\mathbf {B} _ {i} \\boldsymbol {\\theta}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.359, + 0.706, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} _ {i} = \\left( \\begin{array}{c c c} - f _ {x} & 0 & x _ {i} - c _ {x} \\\\ 0 & - f _ {y} & y _ {i} - c _ {y} \\end{array} \\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.394, + 0.887, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Omega} _ {i} = \\left( \\begin{array}{l l l} \\frac {(x _ {i} - c _ {x}) (y _ {i} - c _ {y})}{f _ {y}} & - f _ {x} - \\frac {(x _ {i} - c _ {x}) ^ {2}}{f _ {x}} & (y _ {i} - c _ {y}) \\frac {f _ {x}}{f _ {y}} \\\\ f _ {y} + \\frac {(y _ {i} - c _ {y}) ^ {2}}{f _ {y}} & - \\frac {(x _ {i} - c _ {x}) (y _ {i} - c _ {y})}{f _ {x}} & - (x _ {i} - c _ {x}) \\frac {f _ {y}}{f _ {x}} \\end{array} \\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.568 + ], + "angle": 0, + "content": "where we make explicit the dependency on the camera intrinsic parameters, namely the horizontal and vertical focal lengths \\( f_{x} \\) and \\( f_{y} \\), respectively, and the horizontal and vertical focal center coordinates \\( c_{x} \\) and \\( c_{y} \\), respectively. In this paper, we assume that the focal center coordinates represent the focus of expansion (FOE). Fig. 4 shows the typical per-event estimation of optical flow and depth using the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.577, + 0.704, + 0.593 + ], + "angle": 0, + "content": "3.4. Time-to-Contact Map" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.694 + ], + "angle": 0, + "content": "For each event \\(e_i\\), we estimate its inverse depth \\(\\rho_{i}\\) and update the global motion parameters \\(\\pmb{\\theta} = (\\pmb{\\nu}^{\\mathsf{T}},\\pmb{\\omega}^{\\mathsf{T}})^{\\mathsf{T}}\\). The model parameters \\(\\gamma_{i}\\) are formed by stacking the motion parameters and inverse depth \\(\\gamma_{i} = (\\pmb{\\theta}^{\\mathsf{T}},\\rho_{i})^{\\mathsf{T}}\\). We impose a smoothness constraint to Eq. (6), so that neighboring inverse depth estimates are assumed to be equal to \\(\\rho_{i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.703, + 0.891, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {j} ^ {\\prime} = \\mathbf {x} _ {j} + \\Delta t _ {i, j} \\mathbf {B} _ {i, j} \\boldsymbol {\\theta}, e _ {j} \\in \\operatorname {n e i g h} \\left(e _ {i}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.892, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathbf{B}_{i,j} = (\\rho_i\\mathbf{V}_j\\boldsymbol {\\Omega}_j)\\). The iterative update \\(\\Delta \\gamma_{i}\\) is given by Eq. (8), where \\(\\mathbf{J}_{i,j} = -\\Delta t_{i,j}\\left(\\mathbf{B}_{i,j}\\mathbf{V}_j\\boldsymbol {\\nu}\\right)\\). We maintain the TTCM by computing the TTC for each event \\(e_i\\) based on Eq. (1): \\(\\tau_{i} = 1 / (\\rho_{i}\\nu_{z})\\). We can also estimate the global TTC by averaging over the values maintained in the TTCM. Fig. 5 shows the typical per-event estimation of optical flow and TTCM using the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.636, + 0.862 + ], + "angle": 0, + "content": "3.5. Initialization" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The motion parameters \\(\\theta\\) are global measures which are estimated by aggregating events, while each inverse depth" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23656" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.098, + 0.205, + 0.114 + ], + "angle": 0, + "content": "indoor_flying1 [50]" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.094, + 0.313, + 0.119 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.103, + 0.46, + 0.113 + ], + "angle": 0, + "content": "1sec 10sec" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.121, + 0.209, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.195, + 0.201, + 0.205 + ], + "angle": 0, + "content": "Accumulated events" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.121, + 0.368, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.195, + 0.307, + 0.206 + ], + "angle": 0, + "content": "Optical flow" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.121, + 0.465, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.195, + 0.422, + 0.206 + ], + "angle": 0, + "content": "TTCM" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.213, + 0.468, + 0.228 + ], + "angle": 0, + "content": "Figure 5. Typical per-event estimation of optical flow and TTCM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.253, + 0.47, + 0.42 + ], + "angle": 0, + "content": "\\(\\rho_{i}\\) value corresponds to the estimate of a single event's inverse depth. The initialization of the motion parameters \\(\\theta\\) can thus be almost arbitrarily set, e.g., typically set to 0, and it is only performed once at the beginning. However, a more careful initialization procedure must be considered for the case of inverse depth since it needs to be performed for each event. Based on recent advances in event-based global time decay [31], we perform a weighted average based on the previous neighboring events' inverse depth estimates as the initialization procedure. Each neighboring event's weight \\(w_{i}(t)\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.431, + 0.469, + 0.464 + ], + "angle": 0, + "content": "\\[\nw _ {i} (t) = \\frac {1}{1 + a (t) \\left(t - t _ {i}\\right)}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.469, + 0.535 + ], + "angle": 0, + "content": "where \\( a(t) \\) is the global event activity. If there are no previous neighboring events' inverse depth estimates, which should only occur at the start of the estimation, the initial inverse depth estimate is set to 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.303, + 0.559 + ], + "angle": 0, + "content": "3.6. Practical Considerations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.469, + 0.734 + ], + "angle": 0, + "content": "As mentioned in Sec. 1.1, only the ratio between the depth and relative motion is required to compute the TTC. Since the depth and linear velocities estimates are obtained up-to a scale factor due to the monocular ambiguity, we constrain the linear velocities \\(\\pmb{\\nu}\\) to have at most unit norm, i.e., \\(|\\pmb{\\nu}|_2 \\leq 1\\), which is useful to improve the method's computational stability by bounding the allowed estimates' values. This is not related to event collapse; rather, it stems from the ratio given by Eq. (1), whereby we can introduce an arbitrary non-zero multiplicative scalar to the numerator and denominator and have the same TTC estimate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.81 + ], + "angle": 0, + "content": "Since the events are generated by 3D points in the field of view (FOV) of the camera and should have positive depth values, we constrain the depth values to be strictly positive by introducing a parameterization variable \\(\\lambda \\in \\mathbb{R}\\) such that \\(\\rho (\\lambda) = e^{\\lambda} > 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The iterative update given by Eq. (8) can be quite noisy since it is only computed in a small neighborhood \\(\\mathrm{neigh}(e_i) = \\{e_k:|\\mathbf{x}_k - \\mathbf{x}_i|_{\\infty}\\leq s\\}\\). We thus introduce two prior parameters for the global motion parameters and the local inverse depth estimates \\(l_{\\theta}\\) and \\(l_{\\rho}\\), respectively. The resultant iterative update of the model parameters is given" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.09, + 0.891, + 0.129 + ], + "angle": 0, + "content": "
Neighborhood size s\\( {w}_{\\text{thresh }} \\)\\( {l}_{\\theta } \\)\\( {l}_{\\rho } \\)
30.01\\( \\propto O \\)(camera resolution)1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.529, + 0.139, + 0.862, + 0.153 + ], + "angle": 0, + "content": "Table 1. Hyper-parameters used across the experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.177, + 0.892, + 0.27 + ], + "angle": 0, + "content": "by \\(\\Delta \\pmb{\\gamma}_i = (\\pmb{\\Psi} + \\mathbf{L})^{-1}\\left(-\\pmb{\\psi} + \\mathbf{L}(\\pmb{\\gamma}_{\\mathrm{prev}} - \\pmb{\\gamma}_i^*)\\right)\\), where \\(\\mathbf{L} = \\mathrm{diag}(l_{\\pmb{\\theta}},\\dots ,l_{\\pmb{\\theta}},l_{\\rho})\\), and \\(\\pmb{\\gamma}_{\\mathrm{prev}}\\) are the parameters' estimates from the previous event \\(e_{i-1}\\). We also weigh each event \\(e_j\\) contribution according to the corresponding weight \\(w_j(t)\\) given by Eq. (11), and discard any event whose weight is below a threshold \\(w_{\\mathrm{thresh}}\\) [31]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.284, + 0.735, + 0.301 + ], + "angle": 0, + "content": "4. Experimental Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.892, + 0.462 + ], + "angle": 0, + "content": "We evaluate the proposed method in TTC estimation and optical flow estimation, given that the apparent velocity on the image plane can be estimated according to Eq. (9) and due to the lack of event datasets dedicated to TTC estimation. The optical flow benchmark also provides a common ground to compare the proposed method with other state-of-the-art methods that estimate optical flow. Tab. 1 presents the hyper-parameters that were used across the experiments. Since the proposed method computes per-event estimates, we only evaluate on the respective pixel locations1." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.471, + 0.698, + 0.485 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.63 + ], + "angle": 0, + "content": "VL Dataset \\([28]^2\\). It consists of 7 real event sequences observing planar prints of landing surfaces and 1 real event sequence observing the 3D print of a landing surface. Each sequence has 15sec of duration, totaling 120sec of dataset duration. The events were recorded by a Prophesee event camera with \\(1280 \\times 720\\) resolution, and the ground truth (GT) depth measurements were recorded with an Intel RealSense camera. However, only a global GT depth measurement is provided per timestamp." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.892, + 0.722 + ], + "angle": 0, + "content": "To comply with the evaluation reported [28], we assess the proposed method at certain timestamps that correspond to event batches of 0.5sec. The comparison metrics are the divergence REE (\\%) and runtime per event (microsec). The divergence \\(\\mu\\) is the inverse of the TTC, i.e., \\(\\mu = 1 / \\tau\\), and the REE is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.732, + 0.892, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R E E} = 1 0 0 \\frac {\\left| \\hat {\\mu} - \\mu_ {\\mathrm {g t}} \\right|}{\\left| \\mu_ {\\mathrm {g t}} \\right|}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.777, + 0.891, + 0.806 + ], + "angle": 0, + "content": "where \\(\\hat{\\mu}\\) is the estimate and \\(\\mu_{\\mathrm{gt}}\\) is the GT. The global motion prior \\(l_{\\theta}\\) was set to 1000." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.807, + 0.892, + 0.853 + ], + "angle": 0, + "content": "MVSEC Dataset \\([50]^3\\). It consists of several real indoor and outdoor sequences, providing events, standard grayscale frames, IMU data, camera poses, and scene depth." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.863, + 0.807, + 0.876 + ], + "angle": 0, + "content": "1 All the code will be made available upon acceptance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.519, + 0.876, + 0.834, + 0.888 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/s-mcleod/ventral-landing-event-dataset." + }, + { + "type": "page_footnote", + "bbox": [ + 0.52, + 0.888, + 0.737, + 0.9 + ], + "angle": 0, + "content": "3 https://daniilidis-group.github.io/mvsec." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.863, + 0.834, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23657" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.16, + 0.089, + 0.813, + 0.317 + ], + "angle": 0, + "content": "
Method2D-12D-22D-32D-42D-52D-62D-73DAvg.
REENCECMD [28]13.487.416.1112.1910.415.123.7212.908.92
||Proposed Scaling7.198.799.1312.184.987.794.9211.348.29
SProposed Translation5.037.026.175.254.614.346.186.545.64
3Proposed 6-DOF6.7713.7314.977.506.9112.1512.776.1010.11
3Proposed Scaling14.9915.1911.6921.6911.1915.4411.0119.4315.08
||Proposed Translation11.157.017.9813.806.3711.174.5512.469.31
SProposed 6-DOF6.516.817.256.776.065.596.186.196.42
RuntimeNCECMD [28]63.3762.5151.2141.3692.6938.9725.5731.5350.90
||Proposed Scaling0.820.780.800.810.700.840.780.700.78
SProposed Translation1.101.191.221.081.021.211.221.161.15
3Proposed 6-DOF1.401.431.451.441.371.541.601.471.46
3Proposed Scaling1.181.231.251.261.121.291.261.751.29
||Proposed Translation1.771.701.801.841.681.711.711.701.74
SProposed 6-DOF2.432.402.362.372.482.342.232.272.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.326, + 0.894, + 0.355 + ], + "angle": 0, + "content": "Table 2. Global divergence estimation. Quantitative results on the VL dataset [28], in terms of REE (\\%) and runtime per event (microsec) averaged over 100 trials. Lower is better." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.469, + 0.469 + ], + "angle": 0, + "content": "The events were recorded by a DAVIS [6] with \\(346 \\times 260\\) resolution. The evaluated sequences span approximately 265sec. The optical flow GT is also provided [51], generated from the scene depth and camera velocity. We generate GT TTCM's by applying Eq. (1) given the GT depth maps and camera velocity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.515 + ], + "angle": 0, + "content": "To assess optical flow accuracy, we use the following metrics: average endpoint error (AEE) (in pixel/frame, as is conventional in the literature [42, 51])" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.525, + 0.47, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A E E} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| \\hat {\\mathbf {m}} _ {i} - \\mathbf {m} _ {\\mathrm {g t}, i} \\right| _ {2}, \\mathbf {m} = \\Delta t _ {\\text {f r a m e}} \\mathbf {v}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.608 + ], + "angle": 0, + "content": "outliers (Out) as the percentage of pixels with AEE greater than 3, average relative endpoint error (AREE) \\((\\%)\\) [27]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.618, + 0.47, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A R E E} = \\frac {1 0 0}{N} \\sum_ {i = 1} ^ {N} \\frac {\\left| \\hat {\\mathbf {m}} _ {i} - \\mathbf {m} _ {\\mathrm {g t} , i} \\right| _ {2}}{\\left| \\mathbf {m} _ {\\mathrm {g t} , i} \\right| _ {2}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.67, + 0.354, + 0.687 + ], + "angle": 0, + "content": "and average angular error (AAE) \\((^{\\circ})\\) [27]" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.696, + 0.47, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A A E} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\operatorname {a r c c o s} \\frac {\\hat {\\mathbf {m}} _ {i} ^ {\\top} \\mathbf {m} _ {\\mathrm {g t} , i}}{| \\hat {\\mathbf {m}} _ {i} | _ {2} | \\mathbf {m} _ {\\mathrm {g t} , i} | _ {2}}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.47, + 0.794 + ], + "angle": 0, + "content": "To assess TTCM accuracy, we use the AREE \\((\\%)\\) between the divergence estimate \\(\\hat{\\mu}_i\\) and corresponding GT \\(\\mu_{\\mathrm{gt},i}\\). The global motion prior \\(l_{\\theta}\\) was set to 100." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.338, + 0.817 + ], + "angle": 0, + "content": "4.2. Other Global Motion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.871 + ], + "angle": 0, + "content": "Based on the general 6-DOF global motion model described in Sec. 3.3, we can consider other more constrained motion models depending on the application, as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Translation. This model is parameterized by the 3D linear velocities \\(\\pmb{\\nu}\\). Thus, \\(\\mathbf{B}_i = \\rho_i\\mathbf{V}_i\\), being \\(\\mathbf{V}_i\\) given by Eq. (9)." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.376, + 0.89, + 0.442 + ], + "angle": 0, + "content": "
Modelindoor_flying1indoor_flying2outdoor_day1
Driving153.3084.8047.27
Translation19.8228.2345.78
6-DOF53.6556.1256.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.45, + 0.893, + 0.478 + ], + "angle": 0, + "content": "Table 3. Divergence estimation. Quantitative results on the MVSEC dataset [50], in terms of AREE (\\%). Lower is better." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.506, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Driving. This model is parameterized by the most significant DOF's when driving a car, namely the angular velocity around the camera \\(y\\)-axis \\(\\omega_{y}\\) and the linear velocity in the \\(z\\)-axis \\(\\nu_{z}\\) (see Fig. 2). Hence, \\(\\mathbf{B}_i = \\left( \\begin{array}{cc} \\rho_i(x_i - c_x) & -f_x - \\frac{(x_i - c_x)^2}{f_x} \\\\ \\rho_i(y_i - c_y) & -\\frac{(x_i - c_x)(y_i - c_y)}{f_x} \\end{array} \\right)\\). When using this motion model, we impose \\(|\\nu_z|_2 = 1\\) to ensure that the depth is properly estimated." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.637, + 0.892, + 0.727 + ], + "angle": 0, + "content": "Scaling. This model is parameterized by the linear velocity in the \\(z\\)-axis \\(\\nu_{z}\\). Hence, \\(\\mathbf{B}_i = \\rho_i\\left( \\begin{array}{c}x_i - c_x\\\\ y_i - c_y \\end{array} \\right)\\). It is only considered since the global motion model used in [28] is the 1-DOF scaling, which assumes that the scene is planar and it does not estimate the depth." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.593, + 0.755 + ], + "angle": 0, + "content": "4.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Time-to-Contact. Tab. 2 reports the results on global divergence estimation on the VL benchmark [28]. The proposed method using the Translation model with neighboring size \\( s = 2 \\) achieves the best accuracy, outperforming ECMD [28] by \\( 36.77\\% \\) on average. The proposed method using the Scaling and full 6-DOF models also outperform ECMD [28] in terms of average accuracy with neighboring size \\( s = 2 \\) and \\( s = 3 \\), respectively. The results indicate that considering a smaller neighborhood benefits global mod" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23658" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.174, + 0.089, + 0.803, + 0.284 + ], + "angle": 0, + "content": "
Methodindoor_flying1indoor_flying2indoor_flying3outdoor_day1
AEEOutAEEOutAEEOutAEEOut
LBEV-FlowNet (MB) [52]0.580.001.024.000.873.000.320.00
EV-FlowNet (HQF) [46]0.561.000.661.000.591.000.681.00
Ding et al. [11]0.570.100.791.600.721.300.420.00
ConvGRU-EV-FlowNet [19]0.600.511.178.060.935.640.470.25
MBFBBrebion et al. [7]0.520.100.985.500.712.100.530.20
BBShiba et al. [42]0.420.100.600.590.500.280.300.10
EBAung et al. [3]----2.31-1.26-
ARMS [1]1.52-1.59-1.89-2.75-
Proposed Driving1.407.502.4226.322.1426.050.390.12
Proposed Translation0.630.410.942.830.791.650.550.97
Proposed 6-DOF0.881.971.6012.021.389.990.812.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.292, + 0.895, + 0.321 + ], + "angle": 0, + "content": "Table 4. Optical flow estimation. Quantitative results on the MVSEC dataset [50], in terms of AEE (pixel/frame) and Out (\\%). Lower is better. Legend: learning-based (LB), model-based (MB), frame-based (FB), batch-based (BB), event-based (EB)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.471, + 0.466 + ], + "angle": 0, + "content": "els with fewer parameters; conversely, a larger neighborhood benefits models with more parameters. Even though the motion for all sequences is predominantly dominated by just 1 scaling DOF, the results suggest that considering motion models with additional DOFs, e.g., Translation and 6-DOF, is beneficial. The extra DOFs may explain other small motions, whereas these small motions would just be considered noise for the Scaling model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.576 + ], + "angle": 0, + "content": "In terms of runtime per event, the proposed method outperforms ECMD [28] by between \\(95.36\\%\\) and \\(98.47\\%\\), achieving real-time processing for all the sequences in the VL dataset [28], being capable of processing between 420k and 1.28M events per second. The results indicate that the runtime increases with the neighboring size \\(s\\) and the number of parameters of the global motion model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.471, + 0.852 + ], + "angle": 0, + "content": "Tab. 3 reports the results on divergence estimation on the MVSEC benchmark [50]. The Translation model achieves the best performance overall, indicating that the corresponding DOFs are sufficient to explain the perceived motions in the sequences evaluated while minimizing the optimization complexity. As mentioned in Sec. 4.2, the Driving model is tuned to the outdoor_day1 sequence, whose performance is on par with the Translation model. However, it performs poorly on the indoor_flying3 sequence since it can not handle more complex types of motions that are present. Being the most general model, the 6-DOF motion model achieves similar performance for all the sequences. Its performance is worse than the Translation model due to the increased optimization complexity, while the additional DOFs do not contribute to improving the accuracy. The proposed method achieves real-time processing in the MVSEC dataset [50] since, on average, the sequences exhibit a maximum of around 400k events per second." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Optical Flow. Since it is difficult to compare directly the results reported in Tab. 3, we evaluate the proposed method on optical flow estimation, and compare the results with other" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.342, + 0.88, + 0.473 + ], + "angle": 0, + "content": "
Methodindoor_flying3outdoor_day1
AREEAAEAREEAAE
EV-FlowNet [51]63.6132.5565.5422.88
Aung et al. [3]95.6469.35122.1069.41
ABMOF [26]52.6916.5790.1333.76
EDFLOW [27]37.5212.1569.4023.30
Proposed Driving104.0560.3650.2513.31
Proposed Translation44.4611.3872.6721.77
Proposed 6-DOF95.2321.35122.0431.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.482, + 0.895, + 0.524 + ], + "angle": 0, + "content": "Table 5. Optical flow estimation. Quantitative results on the MVSEC dataset [50], in terms of AREE (\\%) and AAE (\\%). Lower is better." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.649 + ], + "angle": 0, + "content": "methods. Tab. 4 reports the results on optical flow estimation on the MVSEC benchmark [50] in terms of AEE and Out, as is commonly found in the literature. The proposed method achieves state-of-the-art performance over the EB methods, on par performance with the FB method, and competitive performance overall." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.654, + 0.893, + 0.851 + ], + "angle": 0, + "content": "Tab. 5 reports the results on optical flow estimation on the MVSEC benchmark [50] in terms of AREE and AAE. The proposed method achieves state-of-the-art performance on the outdoor_day1 sequence using the Driving model and on par performance on the indoor_flying3 sequence using the Translation model. The results on Tab. 5 indicate that the proposed method is comparatively more accurate in estimating the direction of the flow, i.e., compared with the other methods, the proposed method achieves lower values of AAE overall. In terms of accuracy, this suggests that more improvements can be achieved by improving the method's per-event inverse depth estimation since inverse depth estimates mainly contribute to the flow magnitude." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 6 presents qualitative results on sequences of the MVSEC dataset [50]. The estimated TTCM and optical flow resemble the GT ones." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23659" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.093, + 0.856, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.476, + 0.655, + 0.492 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on the MVSEC dataset [50]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.205, + 0.53 + ], + "angle": 0, + "content": "4.4. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Similarly to other event-based methods [15, 22, 35, 52], the proposed method also inherits the brightness assumption from the DMin framework [34]. It can thus provide wrongful estimates for events that are not caused by motion, e.g., due to flickering lights. While this limitation is somewhat mitigated when estimating global quantities, it can struggle to reliably estimate the inverse depth of events that are not caused by motion. Introducing probabilistic uncertainties to the estimates could alleviate this issue while also improving the inverse depth's initialization procedure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Also related, although we explicitly impose a local smoothness constraint to the inverse depth estimates, given by Eq. (10), the proposed method can still provide inverse depth estimates that differ significantly from the neighboring inverse depth estimates. This typically occurs for events that are generated by noise. Carefully tuning the event camera biases and filtering out outliers could improve the overall inverse depth estimation. Having a back-end inverse depth regularizer [22] could also help to mitigate this issue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The proposed method can only estimate one global motion. Thus, it can not adequately handle more than one motion simultaneously, e.g., due to cars moving [50]. Considering a multi-scale approach [1, 42] or explicitly modeling more than one possible motion occurring simultane" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.515, + 0.862, + 0.532 + ], + "angle": 0, + "content": "ously [33, 44] are possible avenues for future research." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.543, + 0.62, + 0.559 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.854 + ], + "angle": 0, + "content": "We have proposed a novel method that estimates the TTCM using a single event camera. The proposed method builds on the DMin framework to incrementally estimate local and global quantities, i.e., inverse depth, and global motion, respectively. We have also proposed an approach that effectively prevents event collapse for incremental event-based estimation without introducing regularizers or additional hyper-parameters. The proposed method also achieves state-of-the-art performance in TTC estimation in terms of accuracy and computational runtime while achieving competitive performance in optical flow estimation. Broadly, the proposed work further builds on the increasing amount of evidence that event cameras are especially suited to address visual motion-based problems; in particular, it further shows that incremental event-based processing can provide a flexible and general methodology to consider when using event cameras, which avoids issues introduced when converting events to other representations, e.g., batches and frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This research received funding from the French National Research Agency (ANR), under grant agreement \\(\\mathrm{N}^{\\circ}\\) ANR-20-CE23-0021, \"AgileNeuRobot\"." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23660" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Himanshu Akolkar, Sio Hoi Ieng, and Ryad Benosman. Real-time High Speed Motion Prediction Using Fast Aperture-Robust Event-Driven Visual Flow. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–1, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Mohammed Almatrafi, Raymond Baldwin, Kiyoharu Aizawa, and Keigo Hirakawa. Distance Surface for Event-Based Optical Flow. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1547-1556, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Myo Tun Aung, Rodney Teo, and Garrick Orchard. Event-based Plane-fitting Optical Flow for Dynamic Vision Sensors in FPGA. In IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-5, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[4] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-Based Visual Flow. IEEE Transactions on Neural Networks and Learning Systems, 25(2):407-417, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.409 + ], + "angle": 0, + "content": "[5] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous Frameless Event-based Optical Flow. Neural Networks, 27:32-37, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.466 + ], + "angle": 0, + "content": "[6] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A \\(240 \\times 180\\) dB 3 μs Latency Global Shutter Spatiotemporal Vision Sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[7] Vincent Brebion, Julien Moreau, and Franck Davoine. Real-Time Optical Flow for Vehicular Perception With Low- and High-Resolution Event Cameras. IEEE Transactions on Intelligent Transportation Systems, 23(9):15066-15078, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[8] William Chamorro, Juan Andrade-Cetto, and Joan Solà. High-Speed Event-Based Camera Tracking. In British Machine Vision Conference (BMVC), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.471, + 0.62 + ], + "angle": 0, + "content": "[9] William Chamorro, Joan Solà, and Juan Andrade-Cetto. Event-Based Line SLAM in Real-Time. IEEE Robotics and Automation Letters, 7(3):8146-8153, July 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.621, + 0.471, + 0.676 + ], + "angle": 0, + "content": "[10] Xavier Clady, Charles Clercq, Sio-Hoi Ieng, Fouzhan Houseini, Marco Randazzo, Lorenzo Natale, Chiara Bartolozzi, and Ryad Benosman. Asynchronous Visual Event-based Time-to-Contact. Frontiers in Neuroscience, 8, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.677, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[11] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-Temporal Recurrent Networks for Event-Based Optical Flow Estimation. In AAAI Conference on Artificial Intelligence, volume 36, pages 525-533, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.747, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[12] Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dynamic obstacle avoidance for quadrotors with event cameras. Science Robotics, 5(40):eaaz9712, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[13] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-Based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):154-180, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.872, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[14] Guillermo Gallego, Jon E.A. Lund, Elias Mueggler, Henri Rebecq, Tobi Delbruck, and Davide Scaramuzza. Event-" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Based, 6-DOF Camera Tracking from Photometric Depth Maps. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(10):2402-2412, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[15] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A Unifying Contrast Maximization Framework for Event Cameras, with Applications to Motion, Depth, and Optical Flow Estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3867-3876, 2018. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[16] Daniel Gehrig, Antonio Loquercio, Konstantinos Derpanis, and Davide Scaramuzza. End-to-End Learning of Representations for Asynchronous Event-Based Data. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5632–5642, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[17] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense Optical Flow from Event Cameras. In International Conference of 3D Vision (3DV), pages 197–206, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[18] Germain Haessig, Xavier Berthelon, Sio-Hoi Ieng, and Ryad Benosman. A Spiking Neural Network Model of Depth from Defocus for Event-based Neuromorphic Vision. Scientific Reports, 9(1):3744, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[19] Jesse Hagenaars, Federico Paredes-Valles, and Guido de Croon. Self-Supervised Learning of Event-Based Optical Flow with Spiking Neural Networks. In Advances in Neural Information Processing Systems (NIPS), volume 34, pages 7167–7179, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[20] Botao He, Haojia Li, Siyuan Wu, Dong Wang, Zhiwei Zhang, Qianli Dong, Chao Xu, and Fei Gao. FAST-Dynamic-Vision: Detection and Tracking Dynamic Objects with Event and Depth Sensing. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3071-3078, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[21] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew Davison. Simultaneous Mosaicing and Tracking with an Event Camera. In British Machine Vision Conference (BMVC). British Machine Vision Association (BMVA), 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[22] Hanme Kim, Stefan Leutenegger, and Andrew J. Davison. Real-Time 3D Reconstruction and 6-DoF Tracking with an Event Camera. In Computer Vision – ECCV, Lecture Notes in Computer Science, pages 349–364. Springer International Publishing, 2016. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[23] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E. Shi, and Ryad B. Benosman. HOTS: A Hierarchy of Event-Based Time-Surfaces for Pattern Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(7):1346-1359, July 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Chankyu Lee, Adarsh Kumar Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, and Kaushik Roy. Spike-FlowNet: Event-Based Optical Flow Estimation with Energy-Efficient Hybrid Neural Networks. In Computer Vision - ECCV, Lecture Notes in Computer Science, pages 366-382. Springer International Publishing, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A \\(128 \\times 128\\) dB 15 μs Latency Asynchronous Temporal Contrast Vision Sensor. IEEE Journal of Solid-State Circuits, 43(2):566-576, 2008. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23661" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[26] Min Liu and T. Delbruck. Adaptive Time-Slice Block-Matching Optical Flow Algorithm for Dynamic Vision Sensors. In British Machine Vision Conference (BMVC), 2018, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.471, + 0.217 + ], + "angle": 0, + "content": "[27] Min Liu and Tobi Delbruck. EDFLOW: Event Driven Optical Flow Camera With Keypoint Detection and Adaptive Block Matching. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5776-5789, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[28] Sofia McLeod, Gabriele Meoni, Dario Izzo, Anne Mergy, Daqi Liu, Yasir Latif, Ian Reid, and Tat-Jun Chin. Globally Optimal Event-Based Divergence Estimation for Ventral Landing, 2022. 1, 2, 5, 6, 7, 9, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.471, + 0.356 + ], + "angle": 0, + "content": "[29] Moritz B. Milde, Olivier J.N. Bertrand, Ryad Benosmanz, Martin Egelhaaf, and Elisabetta Chicca. Bioinspired event-driven collision avoidance algorithm based on optic flow. In International Conference on Event-based Control, Communication, and Signal Processing (EBCCSP), pages 1-7, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.357, + 0.471, + 0.426 + ], + "angle": 0, + "content": "[30] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The Event-Camera Dataset and Simulator: Event-based Data for Pose Estimation, Visual Odometry, and SLAM. The International Journal of Robotics Research, 36(2):142-149, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[31] Urbano Miguel Nunes, Ryad Benosman, and Sio-Hoi Ieng. Adaptive Global Decay Process for Event Cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9771-9780, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.471, + 0.552 + ], + "angle": 0, + "content": "[32] Urbano Miguel Nunes and Yiannis Demiris. Entropy Minimisation Framework for Event-based Vision Model Estimation. In Computer Vision - ECCV, Lecture Notes in Computer Science, pages 161-176. Springer International Publishing, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.471, + 0.608 + ], + "angle": 0, + "content": "[33] Urbano Miguel Nunes and Yiannis Demiris. Kinematic Structure Estimation of Arbitrary Articulated Rigid Objects for Event Cameras. In International Conference on Robotics and Automation (ICRA), pages 508-514, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[34] Urbano Miguel Nunes and Yiannis Demiris. Robust Event-Based Vision Model Estimation by Dispersion Minimisation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9561-9573, 2022. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.734 + ], + "angle": 0, + "content": "[35] Federico Paredes-Vallés and Guido C. H. E. de Croon. Back to Event Basics: Self-Supervised Learning of Image Reconstruction for Event Cameras via Photometric Constancy. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3445-3454, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[36] Bas J. Pijnacker Hordijk, Kirk Y. W. Scheper, and Guido C. H. E. de Croon. Vertical Landing for Micro Air Vehicles using Event-based Optical Flow. Journal of Field Robotics, 35(1):69-90, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[37] Christoph Posch, Teresa Serrano-Gotarredona, Bernabe Linares-Barranco, and Tobi Delbruck. Retinomorphic Event-Based Vision Sensors: Bioinspired Cameras With Spiking Output. Proceedings of the IEEE, 102(10):1470-1484, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.858, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[38] Henri Rebecq, Timo Horstschaefer, Guillermo Gallego, and Davide Scaramuzza. EVO: A Geometric Approach to Event-Based 6-DOF Parallel Tracking and Mapping in Real Time." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.092, + 0.892, + 0.119 + ], + "angle": 0, + "content": "IEEE Robotics and Automation Letters, 2(2):593-600, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[39] Juan Pablo Rodríguez-Gómez, Raul Tapia, Maria del Mar Guzmán Garcia, Jose Ramiro Martínez-de Dios, and Anibal Ollero. Free as a Bird: Event-Based Dynamic Sense-and-Avoid for Ornithopter Robot Flight. IEEE Robotics and Automation Letters, 7(2):5413-5420, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[40] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event Collapse in Contrast Maximization Frameworks. Sensors, 22(14):5190, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[41] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. A Fast Geometric Regularizer to Mitigate Event Collapse in the Contrast Maximization Framework. Advanced Intelligent Systems, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[42] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of Event-Based Optical Flow. In Computer Vision – ECCV, Lecture Notes in Computer Science, pages 628–645. Springer Nature Switzerland, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[43] Olaf Sikorski, Dario Izzo, and Gabriele Meoni. Event-based Spacecraft Landing using Time-to-Contact. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1941–1950, 2021. ISSN: 2160-7516. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[44] Timo Stoffregen, Guillermo Gallego, Tom Drummond, Lindsay Kleeman, and Davide Scaramuzza. Event-Based Motion Segmentation by Motion Compensation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 7243-7252, 2019. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[45] Timo Stoffregen and Lindsay Kleeman. Simultaneous Optical Flow and Segmentation (SOFAS) using Dynamic Vision Sensor. In Australasian Conference on Robotics and Automation (ACRA), pages 52-61. Australian Robotics and Automation Association (ARAA), 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[46] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the Sim-to-Real Gap for Event Cameras. In Computer Vision - ECCV, Lecture Notes in Computer Science, pages 534-549. Springer International Publishing, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[47] Celyn Walters and Simon Hadfield. EVReflex: Dense Timeto-Impact Prediction for Event-based Obstacle Avoidance. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1304-1309, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[48] Chengxi Ye, Anton Mitrokhin, Cornelia Fermüller, James A. Yorke, and Yiannis Aloimonos. Unsupervised Learning of Dense Optical Flow, Depth and Egomotion with Event-Based Sensors. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5831-5838, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[49] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based Feature Tracking with Probabilistic Data Association. In IEEE International Conference on Robotics and Automation (ICRA), pages 4465-4470, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[50] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The Multivehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018. 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23662" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[51] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. EV-FlowNet: Self-Supervised Optical Flow Estimation for Event-based Cameras. In Proceedings of Robotics: Science and Systems, 2018. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.471, + 0.219 + ], + "angle": 0, + "content": "[52] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised Event-Based Learning of Optical Flow, Depth, and Egomotion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 989-997, 2019. 2, 3, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.471, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "23663" + } + ] +] \ No newline at end of file diff --git a/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/0f2857ef-2537-43f4-b49d-a3319a025266_origin.pdf b/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/0f2857ef-2537-43f4-b49d-a3319a025266_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7af1d9f5eb52dff37fba59629ba42b7a8edd7430 --- /dev/null +++ b/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/0f2857ef-2537-43f4-b49d-a3319a025266_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d222116663c4f28b7e86a9dc6eb0cb53528e4cf58b4d9dd37f4ef97af9e5361 +size 1894410 diff --git a/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/full.md b/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0c3f4f3f0fa5e5328315147747ceb8c423ed0e8a --- /dev/null +++ b/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/full.md @@ -0,0 +1,365 @@ +# Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera + +Urbano Miguel Nunes1, Laurent Udo Perrinet2 and Sio-Hoi Ieng1 + +$^{1}$ Sorbonne University, $^{2}$ Aix Marseille Univ, CNRS, INT, Institut de Neurosciences de la Timone urbano.goncalves-nunes@inserm.fr, laurent.perrinet@univ-amu.fr, sio-hoi.ieng@upmc.fr + +# Abstract + +Event cameras asynchronously report brightness changes with a temporal resolution in the order of microseconds, which makes them inherently suitable to address problems that involve rapid motion perception. In this paper, we address the problem of time-to-contact (TTC) estimation using a single event camera. This problem is typically addressed by estimating a single global TTC measure, which explicitly assumes that the surface/obstacle is planar and fronto-parallel. We relax this assumption by proposing an incremental event-based method to estimate the TTC that jointly estimates the (up-to scale) inverse depth and global motion using a single event camera. The proposed method is reliable and fast while asynchronously maintaining a TTC map (TTCM), which provides per-pixel TTC estimates. As a side product, the proposed method can also estimate per-event optical flow. We achieve state-of-the-art performances on TTC estimation in terms of accuracy and runtime per event while achieving competitive performance on optical flow estimation. + +# 1. Introduction + +Event cameras differ from standard frame-based cameras, which capture visual data at a fixed rate and independently of the observing scene. Instead, event cameras respond asynchronously to pixel-wise brightness changes by generating events [6, 25]. Event cameras are thus data-driven sensors that offer several advantages, including high temporal resolution in the order of microseconds, low latency, low power consumption, and high dynamic range. These properties place event cameras as suitable candidates to address vision-based problems that involve (high-speed) motion, e.g., optical flow estimation [4, 27, 51], ego-motion estimation [15, 21, 22, 34], motion segmentation [33, 44], and obstacle avoidance [10, 12]. Due to the distinct visual sensing paradigm, however, new methods are necessary to fully exploit the potential of event cameras [13, 23]. + +![](images/cba9e1000800efaf48a731f8b951e7fde7238357dc06ad4e3210505788e1f80e.jpg) +Figure 1. Runtime vs. accuracy comparison for TTC estimation methods. Average results on the Ventral Landing benchmark [28]. + +Event cameras have been used to address the problem of fast TTC estimation, which comes up often in the vision-based obstacle avoidance [10, 12, 29, 39] and ventral landing [28, 36] literature. The TTC is the time that would elapse before a camera reaches an obstacle/surface, assuming the current relative motion between them remains constant [10]. Previous methods that use a single event camera have focused on estimating a single global TTC measure, which assumes that the surface is planar and fronto-parallel. To overcome this limitation, other methods use additional sensing, e.g., depth frames, to build a dense TTCM [20, 47]. + +We instead propose to extend the Dispersion Minimization (DMin) framework [34] to estimate the TTC for each incoming event using a single event camera. The proposed method jointly estimates the relative global motion and per-event (up-to scale) inverse depth. We can then asynchronously maintain a semi-dense TTCM which provides per-pixel TTC estimates or compute a global TTC measure with greater accuracy by averaging over the TTC estimates. Since there is at least one scaling degree of freedom (DOF), we also propose an effective strategy to mitigate event collapse [40, 41]. The proposed method is also computationally fast, reaching $\sim 1$ microsecond processing time per event on a standard laptop. Fig. 1 compares the + +runtime vs. accuracy for TTC estimation methods, whereby the proposed method achieves state-of-the-art performance. We also estimate the per-event optical flow as a side product and achieve competitive performance compared to state-of-the-art optical flow methods that use events. + +# Main contributions: + +1. First event-based method that explicitly estimates the TTC for each event and maintains a semi-dense TTCM using a single event camera. +2. DMin framework [34] extension to jointly handle local and global estimates, i.e., inverse depth and global motion, respectively. +3. Effective approach that mitigates event collapse [40, 41] for incremental event-based estimation. + +# 1.1. Time-to-Contact + +Consider a freely moving camera with angular and linear velocities $\omega(t) = (\omega_x(t), \omega_y(t), \omega_z(t))^{\top}$ and $\pmb{\nu}(t) = (\nu_x(t), \nu_y(t), \nu_z(t))^{\top}$ , respectively, that is observing a point $\alpha$ , with 3D coordinates $\pmb{\alpha}(t) = (X(t), Y(t), Z(t))^{\top}$ relative to the camera, as shown in Fig. 2. $Z(t)$ is also referred as the depth of point $\alpha$ relative to the camera. The instantaneous TTC between the moving camera and point $\alpha$ is thus given by: + +$$ +\tau (t) := - \frac {Z (t)}{\frac {d Z (t)}{d t}} = \frac {Z (t)}{\nu_ {z} (t)}. \tag {1} +$$ + +The minus sign disappears because we define the linear velocity w.r.t. the camera's frame of reference, not w.r.t. the point's frame of reference, i.e., $\nu_{z}(t) = -dZ(t) / dt$ . Based on Eq. (1), the exact values of depth and relative approaching motion do not need to be estimated, only the ratio between them. + +# 2. Related Work + +We review recent related works on the following topics: TTC, global motion and optical flow estimation. We refer to [13] for a detailed survey. + +Time-to-Contact Estimation. The first work on event-based TTC using a single event camera relied solely on the estimation of visual motion flows [10], whereby the motion flows were computed by fitting a local plane to the time surface [4]. Other works that followed were geared towards two main use cases, namely obstacle avoidance [12, 29, 39] and ventral landing [28, 36, 43]. Event-based obstacle avoidance methods are built to be fast reacting and, although they come from either bio-inspired [29, 39] or mathematically grounded principles [12], they typically rely on empirically-validated heuristics to speed-up computations. Existing event-based ventral landing approaches only compute a single TTC estimate, which assumes that the surface is planar and fronto-parallel. To overcome this assumption, + +![](images/605e5596d26923c5bebe65252c2d25c2c3fb8001614de17f6958618fdca4b29e.jpg) +Figure 2. Camera coordinate system. + +other works fuse events with additional sensory information, e.g. depth [47]. The proposed approach uses a single event camera while being mathematically grounded and computationally fast. + +Global Motion Estimation. Also denoted by ego-motion estimation [13], it refers to estimating the parameters that explain the triggered events according to some global motion model. These methods can be broadly characterized by whether they rely on key-frame registration [8, 9, 14, 21, 22, 38] or perform the estimation without relying on any key-frame [15, 34]. The former methods are reminiscent of the frame-based paradigm and also include methods based on artificial neural networks (ANN) [16, 52], whereby an intermediate frame-based representation is needed for the estimation. The latter methods tend towards a more event-based processing paradigm and include methods based on spiking neural networks (SNN) [18, 37], whereby events are either processed on an event-by-event basis [34] or in batches [15]. Although both approaches have advantages and disadvantages, methods that rely on key-frame registration typically require an intermediate frame-based representation, which is still an open problem in the event-based community. Similarly to [22], the proposed method jointly estimates up-to scale inverse depth and global motion. By building on the DMin framework [34], which also allows processing events on an event-by-event basis, our method does not rely on key-frame registration or background inverse depth regularization to improve convergence. + +Optical Flow Estimation. Several model-based methods have been previously proposed, which can be further divided into: frame-based [2, 7, 26, 27], batch-based [42, 49], and event-based [1, 3, 4]. By selecting the most relevant events, i.e., typically the most recent, frame-based methods build frames from which the optical flow is computed using techniques from standard image-based optical flow, e.g., Lucas-Kanade [5]. Since each event does not carry much information on its own, batch-based methods aggregate the most recent events by forming batches but perform the computations directly on the events. Event-based methods follow the most event-driven paradigm by performing event-by-event processing, typically being the computationally fastest. However, event-based methods tend to suffer more + +from the aperture problem since all the computations are performed locally, and thus frame-based and model-based methods achieve currently better accuracy. In terms of accuracy, learning-based ANN methods [11, 35, 46, 52] generally achieve state-of-the-art performance. Besides the need to convert events into frames for more efficient processing, these methods are known to be very data hungry, sensitive to the training data [48], and consume large amounts of energy [27]. Another line of research in learning-based methods has been to use SNN [19, 24], which combine the event-based processing and learning paradigms and thus do not require an intermediate frame-based representation. However, it is not trivial to train SNN, and the empirical validation is still not on par with ANN methods. Although it is not the primary objective of this work, the proposed method can provide per-event flow estimates while still being competitive in terms of accuracy w.r.t. state-of-the-art methods. + +# 3. Method + +In this section, we describe the proposed incremental event-based method for TTCM estimation. We first briefly review the event cameras' working principle, and the DMin framework [34], based on which we develop the proposed method. Refer to the supplementary material for the full mathematical derivations and additional details. + +# 3.1. Event Cameras and Dispersion Minimization + +Event cameras output a stream of asynchronous temporal contrast events $\{e_i\}, i \in \mathbb{N}$ . Each event $e_i$ represents a spatio-temporal asynchronous brightness change, being defined as a tuple $e_i := (\mathbf{x}_i, t_i, p_i)$ , where $\mathbf{x}_i = (x_i, y_i)$ are the pixel coordinates, $t_i$ is the timestamp at which the event was generated, and $p_i \in \{-1, +1\}$ is its polarity. An event $e_i$ is generated when the change in log-brightness $\log \mathbf{I}_{x,y}(t) := \bar{\mathbf{I}}_{x,y}(t)$ is above a threshold $L$ + +$$ +\left| \Delta \bar {\mathbf {I}} _ {x _ {i}, y _ {i}} \left(t _ {i}\right) \right| = \left| \bar {\mathbf {I}} _ {x _ {i}, y _ {i}} \left(t _ {i}\right) - \bar {\mathbf {I}} _ {x _ {i}, y _ {i}} \left(t _ {i} - \Delta t _ {i}\right) \right| \geq L, \tag {2} +$$ + +where $\Delta t_{i}$ is the time since the last event at the same pixel. + +The DMin framework [34] estimates the parameters $\pmb{\theta}$ of a transformation model $\mathcal{M}$ from the stream of events $\mathcal{E} = \{e_i\}_{i=1}^{N_e}$ by minimizing a dispersion measure of the transformed events $\mathbf{f}_i = \mathcal{M}(e_i; t_{\mathrm{ref}}, \pmb{\theta})$ . We consider the Potential measure with a Gaussian kernel $\mathcal{N}(\mathbf{x}; \pmb{\mu}, \pmb{\Sigma})$ : + +$$ +P (\mathcal {E}; \boldsymbol {\theta}) = - \sum_ {i, j} ^ {N _ {e}} \mathcal {N} \left(\mathbf {f} _ {i}; \mathbf {f} _ {j}, \mathbf {I}\right), \tag {3} +$$ + +where $\mathbf{I}$ is the identity matrix. A key distinction of the DMin framework is that it allows to incrementally estimate the model parameters $\theta$ on an event-by-event basis, whereby the model parameters $\theta$ can be iteratively solved by linearizing the transformation model $\mathcal{M}$ , such that: + +$$ +\mathbf {f} _ {i} = \mathcal {M} \left(e _ {i}; t _ {\mathrm {r e f}}, \boldsymbol {\theta}\right) = \tilde {\mathbf {f}} _ {i} + \Delta t _ {i, \mathrm {r e f}} \mathbf {B} _ {i} \boldsymbol {\theta}, \mathbf {B} _ {i} := \mathbf {B} \left(e _ {i}\right), (4) +$$ + +where $\mathbf{B}_i$ is the model-dependent linearization matrix. The parameters $\theta^{*}$ that minimize Eq. (3) are thus given by: + +$$ +\boldsymbol {\theta} ^ {*} = - \left(\sum_ {i, j} ^ {N _ {e}} \mathbf {W} _ {i, j} \Delta \mathbf {C} _ {i, j}\right) ^ {- 1} \left(\sum_ {i, j} ^ {N _ {e}} \mathbf {W} _ {i, j} \Delta \tilde {\mathbf {f}} _ {i, j}\right), \tag {5} +$$ + +where $\Delta \mathbf{C}_{i,j} = \mathbf{C}_i - \mathbf{C}_j$ , $\Delta \tilde{\mathbf{f}}_{i,j} = \tilde{\mathbf{f}}_i - \tilde{\mathbf{f}}_j$ , $\mathbf{W}_{i,j} = w_{i,j}\partial \Delta \mathbf{f}_{i,j}^{\top} / \partial \pmb {\theta}$ , $\mathbf{C}_i = \Delta t_{i,\mathrm{ref}}\mathbf{B}_i$ and $w_{i,j} = \mathcal{N}(\mathbf{f}_i;\mathbf{f}_j,\mathbf{I})$ . + +# 3.2. Adapted Dispersion Minimization + +While the DMin method [34] provides a general framework for global incremental event-based model estimation, it can also be adapted to jointly estimate global and local measures, i.e., global angular and linear velocities and local inverse depth. However, the DMin framework may encounter estimation issues when the global model has at least one scaling DOF, as noted in [34, 40, 41], known as event collapse. Event collapse occurs when the events are transformed into a single point, which minimizes the events' dispersion or maximizes the image contrast while the parameters' estimates diverge. So far, to the best of our knowledge, the mitigation discussion in the literature has been on how to constrain the optimization loss to discourage divergent estimates by analyzing the effects of the scaling transformations on the event-based data. Several mitigation strategies have thus been proposed on the events [34, 40] and parameters level [41] by adding terms to regularize the objective measure. While these strategies generally prevent event collapse, they increase the complexity of the optimization framework and introduce additional parameters to tune. + +We instead observe that event collapse fundamentally stems from the event transformation to a common time reference $t_{\mathrm{ref}}$ , e.g., given by Eq. (4), by identifying two problems: 1) the constant velocity assumption may not hold depending on the time difference $\Delta t_{i,\mathrm{ref}}$ , and 2) there is no built-in constraint on the magnitude of the model parameters, e.g., such that the difference between transformed events $\mathbf{f}_i - \mathbf{f}_j$ explicitly penalizes divergent estimates; although, according to Eq. (4), $\mathbf{f}_i$ and $\mathbf{f}_j$ individually diverge if the model parameters also diverge, their difference $\mathbf{f}_i - \mathbf{f}_j$ is not guaranteed to diverge and thus penalize divergent estimates. The first problem is typically addressed by heuristically making the time difference as short as possible, and its effects are of limited significance in practice. The second problem, however, is intrinsically linked to batch-based processing since the alignment of the events in a batch needs to be measured in some common time reference [15, 32]. + +However, for incremental event-based processing, the event transformation and, consequently, the dispersion measure can be modified without loss of generality such that the event collapse is prevented by implicitly addressing the two problems identified. Fig. 3 depicts the idea whereby we only transform the event neighbors to the current event's + +![](images/1b248b34a5449923deacc9fef9846fd2b30875d43fadf59b6a6f2e9d15c7a1a7.jpg) +Figure 3. Incremental event collapse mitigation. Instead of transforming all the events to some time reference (left), we locally transform the events to the current event's timestamp (right). + +![](images/929f61df89c36391bdca5a91846302bf95f20a7e025a504bc707f0ebfa479b2b.jpg) + +timestamp instead of transforming all the events to a common time reference, including the current event. Formally, instead of transforming the events according to Eq. (4), only the neighboring events $e_j$ of the current event $e_i$ are transformed according to + +$$ +\mathbf {x} _ {j} ^ {\prime} = \mathbf {x} _ {j} + \Delta t _ {i, j} \mathbf {B} _ {j} \boldsymbol {\theta} = \mathbf {x} _ {j} + \mathbf {C} _ {i, j} \boldsymbol {\theta}, e _ {j} \in \operatorname {n e i g h} (e _ {i}), \tag {6} +$$ + +where 'neigh' is shorthand for neighborhood, and the dependency of $\mathbf{x}_j^\prime$ on $\pmb{\theta}$ was omitted for brevity. From the resultant residual $\mathbf{r}_{i,j} = \mathbf{x}_i - \mathbf{x}_j^\prime = \Delta \mathbf{x}_{i,j} - \mathbf{C}_{i,j}\pmb {\theta}$ , we see that the proposed modification to the DMin framework addresses both identified problems. First, the constant velocity assumption is better held since the time difference satisfies $\Delta t_{i,j}\leq \Delta t_{i,\mathrm{ref}}$ . Second, if the model parameters $\pmb{\theta}$ diverge, then the residual $\mathbf{r}_{i,j}$ also diverges, which effectively penalizes divergent estimates. We highlight that the proposed adaptation only works for incremental event-based processing: it is not suitable for batch-based processing since the proposed transformation only works locally. + +The Potential measure is modified by computing the difference between the current event's coordinates $\mathbf{x}_i$ and the neighboring events' transformed coordinates $\mathbf{x}_j'$ : + +$$ +P \left(\mathcal {E}; \boldsymbol {\theta}\right) = - \sum_ {i, j} \mathcal {N} \left(\mathbf {x} _ {j} ^ {\prime}; \mathbf {x} _ {i}, \mathbf {I}\right), e _ {j} \in \operatorname {n e i g h} \left(e _ {i}\right). \tag {7} +$$ + +By minimizing Eq. (7) by linearizing the residual according to Taylor's formula $\mathbf{r}_{i,j}(\pmb {\theta} + \Delta \pmb {\theta})\approx \mathbf{r}_{i,j}(\pmb {\theta}) + \mathbf{J}_{i,j}(\pmb {\theta})\Delta \pmb{\theta}$ , the optimized model parameters $\pmb{\theta}^{*}$ are iteratively updated: + +$$ +\boldsymbol {\theta} ^ {*} \leftarrow \boldsymbol {\theta} ^ {*} + \Delta \boldsymbol {\theta}, \tag {8} +$$ + +$$ +\Delta \boldsymbol {\theta} = - \underbrace {\left(\sum_ {j} w _ {i , j} \mathbf {J} _ {i , j} ^ {\top} \mathbf {J} _ {i , j}\right)} _ {\boldsymbol {\Psi}} ^ {- 1} \underbrace {\left(\sum_ {j} w _ {i , j} \mathbf {J} _ {i , j} ^ {\top} \mathbf {r} _ {i , j}\right)} _ {\boldsymbol {\psi}}, +$$ + +where $w_{i,j} = \mathcal{N}\left(\mathbf{x}_j';\mathbf{x}_i,\mathbf{I}\right)$ , $\mathbf{J}_{i,j} = \partial \mathbf{r}_{i,j} / \partial \pmb{\theta}$ and $e_j\in$ $\mathrm{neigh}(e_i)$ . + +# 3.3. Inverse Depth and Global Motion Model + +We consider that a calibrated event camera can freely move and whose global motion is parameterized by the 3D + +![](images/69a437785dceb881ab23e23df94a55bf5358cba59b3f598fc4af52bcd562db5d.jpg) +Figure 4. Typical per-event estimation of optical flow and depth. + +angular and linear velocities, as defined in Sec. 1.1, $\pmb{\theta} = (\pmb{\nu}^{\mathsf{T}},\pmb{\omega}^{\mathsf{T}})^{\mathsf{T}}$ . For each event $e_i$ , we estimate its inverse depth $\rho_{i}\coloneqq 1 / Z_{i}$ , based on the well-known expression for the apparent velocity on the image plane: + +$$ +\mathbf {v} _ {i} = \left( \begin{array}{l l} \rho_ {i} \mathbf {V} _ {i} & \boldsymbol {\Omega} _ {i} \end{array} \right) \binom {\boldsymbol {\nu}} {\omega} = \rho_ {i} \mathbf {V} _ {i} \boldsymbol {\nu} + \boldsymbol {\Omega} _ {i} \boldsymbol {\omega} := \mathbf {B} _ {i} \boldsymbol {\theta}, \tag {9} +$$ + +$$ +\mathbf {V} _ {i} = \left( \begin{array}{c c c} - f _ {x} & 0 & x _ {i} - c _ {x} \\ 0 & - f _ {y} & y _ {i} - c _ {y} \end{array} \right), +$$ + +$$ +\boldsymbol {\Omega} _ {i} = \left( \begin{array}{l l l} \frac {(x _ {i} - c _ {x}) (y _ {i} - c _ {y})}{f _ {y}} & - f _ {x} - \frac {(x _ {i} - c _ {x}) ^ {2}}{f _ {x}} & (y _ {i} - c _ {y}) \frac {f _ {x}}{f _ {y}} \\ f _ {y} + \frac {(y _ {i} - c _ {y}) ^ {2}}{f _ {y}} & - \frac {(x _ {i} - c _ {x}) (y _ {i} - c _ {y})}{f _ {x}} & - (x _ {i} - c _ {x}) \frac {f _ {y}}{f _ {x}} \end{array} \right) +$$ + +where we make explicit the dependency on the camera intrinsic parameters, namely the horizontal and vertical focal lengths $f_{x}$ and $f_{y}$ , respectively, and the horizontal and vertical focal center coordinates $c_{x}$ and $c_{y}$ , respectively. In this paper, we assume that the focal center coordinates represent the focus of expansion (FOE). Fig. 4 shows the typical per-event estimation of optical flow and depth using the proposed method. + +# 3.4. Time-to-Contact Map + +For each event $e_i$ , we estimate its inverse depth $\rho_{i}$ and update the global motion parameters $\pmb{\theta} = (\pmb{\nu}^{\mathsf{T}},\pmb{\omega}^{\mathsf{T}})^{\mathsf{T}}$ . The model parameters $\gamma_{i}$ are formed by stacking the motion parameters and inverse depth $\gamma_{i} = (\pmb{\theta}^{\mathsf{T}},\rho_{i})^{\mathsf{T}}$ . We impose a smoothness constraint to Eq. (6), so that neighboring inverse depth estimates are assumed to be equal to $\rho_{i}$ : + +$$ +\mathbf {x} _ {j} ^ {\prime} = \mathbf {x} _ {j} + \Delta t _ {i, j} \mathbf {B} _ {i, j} \boldsymbol {\theta}, e _ {j} \in \operatorname {n e i g h} \left(e _ {i}\right), \tag {10} +$$ + +where $\mathbf{B}_{i,j} = (\rho_i\mathbf{V}_j\boldsymbol {\Omega}_j)$ . The iterative update $\Delta \gamma_{i}$ is given by Eq. (8), where $\mathbf{J}_{i,j} = -\Delta t_{i,j}\left(\mathbf{B}_{i,j}\mathbf{V}_j\boldsymbol {\nu}\right)$ . We maintain the TTCM by computing the TTC for each event $e_i$ based on Eq. (1): $\tau_{i} = 1 / (\rho_{i}\nu_{z})$ . We can also estimate the global TTC by averaging over the values maintained in the TTCM. Fig. 5 shows the typical per-event estimation of optical flow and TTCM using the proposed method. + +# 3.5. Initialization + +The motion parameters $\theta$ are global measures which are estimated by aggregating events, while each inverse depth + +![](images/77d3fbbe11908f2e76c824e565ee59d80421fef507c962bd274b8dfc39da7c2a.jpg) + +![](images/38c2954c7bc377b2656f9f1a925c5cad92a8768ab8419ba5d93100c5ed65650b.jpg) +indoor_flying1 [50] +Accumulated events +Figure 5. Typical per-event estimation of optical flow and TTCM. + +![](images/dbb593d55e20fc4aee47a2ccfe3b8472d6f709d90c2dddb7ceeaa10131e6edc4.jpg) +1sec 10sec +Optical flow + +![](images/382f43152d0220e9fcdc3a81ecf1955d3901ebf212e147e7f6e39c9066eb23f8.jpg) +TTCM + +$\rho_{i}$ value corresponds to the estimate of a single event's inverse depth. The initialization of the motion parameters $\theta$ can thus be almost arbitrarily set, e.g., typically set to 0, and it is only performed once at the beginning. However, a more careful initialization procedure must be considered for the case of inverse depth since it needs to be performed for each event. Based on recent advances in event-based global time decay [31], we perform a weighted average based on the previous neighboring events' inverse depth estimates as the initialization procedure. Each neighboring event's weight $w_{i}(t)$ is given by: + +$$ +w _ {i} (t) = \frac {1}{1 + a (t) \left(t - t _ {i}\right)}, \tag {11} +$$ + +where $a(t)$ is the global event activity. If there are no previous neighboring events' inverse depth estimates, which should only occur at the start of the estimation, the initial inverse depth estimate is set to 1. + +# 3.6. Practical Considerations + +As mentioned in Sec. 1.1, only the ratio between the depth and relative motion is required to compute the TTC. Since the depth and linear velocities estimates are obtained up-to a scale factor due to the monocular ambiguity, we constrain the linear velocities $\pmb{\nu}$ to have at most unit norm, i.e., $|\pmb{\nu}|_2 \leq 1$ , which is useful to improve the method's computational stability by bounding the allowed estimates' values. This is not related to event collapse; rather, it stems from the ratio given by Eq. (1), whereby we can introduce an arbitrary non-zero multiplicative scalar to the numerator and denominator and have the same TTC estimate. + +Since the events are generated by 3D points in the field of view (FOV) of the camera and should have positive depth values, we constrain the depth values to be strictly positive by introducing a parameterization variable $\lambda \in \mathbb{R}$ such that $\rho (\lambda) = e^{\lambda} > 0$ . + +The iterative update given by Eq. (8) can be quite noisy since it is only computed in a small neighborhood $\mathrm{neigh}(e_i) = \{e_k:|\mathbf{x}_k - \mathbf{x}_i|_{\infty}\leq s\}$ . We thus introduce two prior parameters for the global motion parameters and the local inverse depth estimates $l_{\theta}$ and $l_{\rho}$ , respectively. The resultant iterative update of the model parameters is given + +
Neighborhood size s\( {w}_{\text{thresh }} \)\( {l}_{\theta } \)\( {l}_{\rho } \)
30.01\( \propto O \)(camera resolution)1
+ +Table 1. Hyper-parameters used across the experiments. + +by $\Delta \pmb{\gamma}_i = (\pmb{\Psi} + \mathbf{L})^{-1}\left(-\pmb{\psi} + \mathbf{L}(\pmb{\gamma}_{\mathrm{prev}} - \pmb{\gamma}_i^*)\right)$ , where $\mathbf{L} = \mathrm{diag}(l_{\pmb{\theta}},\dots ,l_{\pmb{\theta}},l_{\rho})$ , and $\pmb{\gamma}_{\mathrm{prev}}$ are the parameters' estimates from the previous event $e_{i-1}$ . We also weigh each event $e_j$ contribution according to the corresponding weight $w_j(t)$ given by Eq. (11), and discard any event whose weight is below a threshold $w_{\mathrm{thresh}}$ [31]. + +# 4. Experimental Evaluation + +We evaluate the proposed method in TTC estimation and optical flow estimation, given that the apparent velocity on the image plane can be estimated according to Eq. (9) and due to the lack of event datasets dedicated to TTC estimation. The optical flow benchmark also provides a common ground to compare the proposed method with other state-of-the-art methods that estimate optical flow. Tab. 1 presents the hyper-parameters that were used across the experiments. Since the proposed method computes per-event estimates, we only evaluate on the respective pixel locations1. + +# 4.1. Datasets and Metrics + +VL Dataset $[28]^2$ . It consists of 7 real event sequences observing planar prints of landing surfaces and 1 real event sequence observing the 3D print of a landing surface. Each sequence has 15sec of duration, totaling 120sec of dataset duration. The events were recorded by a Prophesee event camera with $1280 \times 720$ resolution, and the ground truth (GT) depth measurements were recorded with an Intel RealSense camera. However, only a global GT depth measurement is provided per timestamp. + +To comply with the evaluation reported [28], we assess the proposed method at certain timestamps that correspond to event batches of 0.5sec. The comparison metrics are the divergence REE (\%) and runtime per event (microsec). The divergence $\mu$ is the inverse of the TTC, i.e., $\mu = 1 / \tau$ , and the REE is given by: + +$$ +\mathrm {R E E} = 1 0 0 \frac {\left| \hat {\mu} - \mu_ {\mathrm {g t}} \right|}{\left| \mu_ {\mathrm {g t}} \right|}, \tag {12} +$$ + +where $\hat{\mu}$ is the estimate and $\mu_{\mathrm{gt}}$ is the GT. The global motion prior $l_{\theta}$ was set to 1000. + +MVSEC Dataset $[50]^3$ . It consists of several real indoor and outdoor sequences, providing events, standard grayscale frames, IMU data, camera poses, and scene depth. + +
Method2D-12D-22D-32D-42D-52D-62D-73DAvg.
REENCECMD [28]13.487.416.1112.1910.415.123.7212.908.92
||Proposed Scaling7.198.799.1312.184.987.794.9211.348.29
SProposed Translation5.037.026.175.254.614.346.186.545.64
3Proposed 6-DOF6.7713.7314.977.506.9112.1512.776.1010.11
3Proposed Scaling14.9915.1911.6921.6911.1915.4411.0119.4315.08
||Proposed Translation11.157.017.9813.806.3711.174.5512.469.31
SProposed 6-DOF6.516.817.256.776.065.596.186.196.42
RuntimeNCECMD [28]63.3762.5151.2141.3692.6938.9725.5731.5350.90
||Proposed Scaling0.820.780.800.810.700.840.780.700.78
SProposed Translation1.101.191.221.081.021.211.221.161.15
3Proposed 6-DOF1.401.431.451.441.371.541.601.471.46
3Proposed Scaling1.181.231.251.261.121.291.261.751.29
||Proposed Translation1.771.701.801.841.681.711.711.701.74
SProposed 6-DOF2.432.402.362.372.482.342.232.272.36
+ +The events were recorded by a DAVIS [6] with $346 \times 260$ resolution. The evaluated sequences span approximately 265sec. The optical flow GT is also provided [51], generated from the scene depth and camera velocity. We generate GT TTCM's by applying Eq. (1) given the GT depth maps and camera velocity. + +To assess optical flow accuracy, we use the following metrics: average endpoint error (AEE) (in pixel/frame, as is conventional in the literature [42, 51]) + +$$ +\mathrm {A E E} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left| \hat {\mathbf {m}} _ {i} - \mathbf {m} _ {\mathrm {g t}, i} \right| _ {2}, \mathbf {m} = \Delta t _ {\text {f r a m e}} \mathbf {v}, \tag {13} +$$ + +outliers (Out) as the percentage of pixels with AEE greater than 3, average relative endpoint error (AREE) $(\%)$ [27] + +$$ +\operatorname {A R E E} = \frac {1 0 0}{N} \sum_ {i = 1} ^ {N} \frac {\left| \hat {\mathbf {m}} _ {i} - \mathbf {m} _ {\mathrm {g t} , i} \right| _ {2}}{\left| \mathbf {m} _ {\mathrm {g t} , i} \right| _ {2}}, \tag {14} +$$ + +and average angular error (AAE) $(^{\circ})$ [27] + +$$ +\mathrm {A A E} = \frac {1}{N} \sum_ {i = 1} ^ {N} \operatorname {a r c c o s} \frac {\hat {\mathbf {m}} _ {i} ^ {\top} \mathbf {m} _ {\mathrm {g t} , i}}{| \hat {\mathbf {m}} _ {i} | _ {2} | \mathbf {m} _ {\mathrm {g t} , i} | _ {2}}. \tag {15} +$$ + +To assess TTCM accuracy, we use the AREE $(\%)$ between the divergence estimate $\hat{\mu}_i$ and corresponding GT $\mu_{\mathrm{gt},i}$ . The global motion prior $l_{\theta}$ was set to 100. + +# 4.2. Other Global Motion Models + +Based on the general 6-DOF global motion model described in Sec. 3.3, we can consider other more constrained motion models depending on the application, as follows. + +Translation. This model is parameterized by the 3D linear velocities $\pmb{\nu}$ . Thus, $\mathbf{B}_i = \rho_i\mathbf{V}_i$ , being $\mathbf{V}_i$ given by Eq. (9). + +Table 2. Global divergence estimation. Quantitative results on the VL dataset [28], in terms of REE (\%) and runtime per event (microsec) averaged over 100 trials. Lower is better. + +
Modelindoor_flying1indoor_flying2outdoor_day1
Driving153.3084.8047.27
Translation19.8228.2345.78
6-DOF53.6556.1256.66
+ +Table 3. Divergence estimation. Quantitative results on the MVSEC dataset [50], in terms of AREE (\%). Lower is better. + +Driving. This model is parameterized by the most significant DOF's when driving a car, namely the angular velocity around the camera $y$ -axis $\omega_{y}$ and the linear velocity in the $z$ -axis $\nu_{z}$ (see Fig. 2). Hence, $\mathbf{B}_i = \left( \begin{array}{cc} \rho_i(x_i - c_x) & -f_x - \frac{(x_i - c_x)^2}{f_x} \\ \rho_i(y_i - c_y) & -\frac{(x_i - c_x)(y_i - c_y)}{f_x} \end{array} \right)$ . When using this motion model, we impose $|\nu_z|_2 = 1$ to ensure that the depth is properly estimated. + +Scaling. This model is parameterized by the linear velocity in the $z$ -axis $\nu_{z}$ . Hence, $\mathbf{B}_i = \rho_i\left( \begin{array}{c}x_i - c_x\\ y_i - c_y \end{array} \right)$ . It is only considered since the global motion model used in [28] is the 1-DOF scaling, which assumes that the scene is planar and it does not estimate the depth. + +# 4.3. Results + +Time-to-Contact. Tab. 2 reports the results on global divergence estimation on the VL benchmark [28]. The proposed method using the Translation model with neighboring size $s = 2$ achieves the best accuracy, outperforming ECMD [28] by $36.77\%$ on average. The proposed method using the Scaling and full 6-DOF models also outperform ECMD [28] in terms of average accuracy with neighboring size $s = 2$ and $s = 3$ , respectively. The results indicate that considering a smaller neighborhood benefits global mod + +
Methodindoor_flying1indoor_flying2indoor_flying3outdoor_day1
AEEOutAEEOutAEEOutAEEOut
LBEV-FlowNet (MB) [52]0.580.001.024.000.873.000.320.00
EV-FlowNet (HQF) [46]0.561.000.661.000.591.000.681.00
Ding et al. [11]0.570.100.791.600.721.300.420.00
ConvGRU-EV-FlowNet [19]0.600.511.178.060.935.640.470.25
MBFBBrebion et al. [7]0.520.100.985.500.712.100.530.20
BBShiba et al. [42]0.420.100.600.590.500.280.300.10
EBAung et al. [3]----2.31-1.26-
ARMS [1]1.52-1.59-1.89-2.75-
Proposed Driving1.407.502.4226.322.1426.050.390.12
Proposed Translation0.630.410.942.830.791.650.550.97
Proposed 6-DOF0.881.971.6012.021.389.990.812.26
+ +els with fewer parameters; conversely, a larger neighborhood benefits models with more parameters. Even though the motion for all sequences is predominantly dominated by just 1 scaling DOF, the results suggest that considering motion models with additional DOFs, e.g., Translation and 6-DOF, is beneficial. The extra DOFs may explain other small motions, whereas these small motions would just be considered noise for the Scaling model. + +In terms of runtime per event, the proposed method outperforms ECMD [28] by between $95.36\%$ and $98.47\%$ , achieving real-time processing for all the sequences in the VL dataset [28], being capable of processing between 420k and 1.28M events per second. The results indicate that the runtime increases with the neighboring size $s$ and the number of parameters of the global motion model. + +Tab. 3 reports the results on divergence estimation on the MVSEC benchmark [50]. The Translation model achieves the best performance overall, indicating that the corresponding DOFs are sufficient to explain the perceived motions in the sequences evaluated while minimizing the optimization complexity. As mentioned in Sec. 4.2, the Driving model is tuned to the outdoor_day1 sequence, whose performance is on par with the Translation model. However, it performs poorly on the indoor_flying3 sequence since it can not handle more complex types of motions that are present. Being the most general model, the 6-DOF motion model achieves similar performance for all the sequences. Its performance is worse than the Translation model due to the increased optimization complexity, while the additional DOFs do not contribute to improving the accuracy. The proposed method achieves real-time processing in the MVSEC dataset [50] since, on average, the sequences exhibit a maximum of around 400k events per second. + +Optical Flow. Since it is difficult to compare directly the results reported in Tab. 3, we evaluate the proposed method on optical flow estimation, and compare the results with other + +Table 4. Optical flow estimation. Quantitative results on the MVSEC dataset [50], in terms of AEE (pixel/frame) and Out (\%). Lower is better. Legend: learning-based (LB), model-based (MB), frame-based (FB), batch-based (BB), event-based (EB). + +
Methodindoor_flying3outdoor_day1
AREEAAEAREEAAE
EV-FlowNet [51]63.6132.5565.5422.88
Aung et al. [3]95.6469.35122.1069.41
ABMOF [26]52.6916.5790.1333.76
EDFLOW [27]37.5212.1569.4023.30
Proposed Driving104.0560.3650.2513.31
Proposed Translation44.4611.3872.6721.77
Proposed 6-DOF95.2321.35122.0431.68
+ +Table 5. Optical flow estimation. Quantitative results on the MVSEC dataset [50], in terms of AREE (\%) and AAE (\%). Lower is better. + +methods. Tab. 4 reports the results on optical flow estimation on the MVSEC benchmark [50] in terms of AEE and Out, as is commonly found in the literature. The proposed method achieves state-of-the-art performance over the EB methods, on par performance with the FB method, and competitive performance overall. + +Tab. 5 reports the results on optical flow estimation on the MVSEC benchmark [50] in terms of AREE and AAE. The proposed method achieves state-of-the-art performance on the outdoor_day1 sequence using the Driving model and on par performance on the indoor_flying3 sequence using the Translation model. The results on Tab. 5 indicate that the proposed method is comparatively more accurate in estimating the direction of the flow, i.e., compared with the other methods, the proposed method achieves lower values of AAE overall. In terms of accuracy, this suggests that more improvements can be achieved by improving the method's per-event inverse depth estimation since inverse depth estimates mainly contribute to the flow magnitude. + +Qualitative Results. Fig. 6 presents qualitative results on sequences of the MVSEC dataset [50]. The estimated TTCM and optical flow resemble the GT ones. + +![](images/d20894387040a86101b5cf6f7fc906801d12351d33524eb5ce5d8e3592aa7dd5.jpg) +Figure 6. Qualitative results on the MVSEC dataset [50]. + +# 4.4. Limitations + +Similarly to other event-based methods [15, 22, 35, 52], the proposed method also inherits the brightness assumption from the DMin framework [34]. It can thus provide wrongful estimates for events that are not caused by motion, e.g., due to flickering lights. While this limitation is somewhat mitigated when estimating global quantities, it can struggle to reliably estimate the inverse depth of events that are not caused by motion. Introducing probabilistic uncertainties to the estimates could alleviate this issue while also improving the inverse depth's initialization procedure. + +Also related, although we explicitly impose a local smoothness constraint to the inverse depth estimates, given by Eq. (10), the proposed method can still provide inverse depth estimates that differ significantly from the neighboring inverse depth estimates. This typically occurs for events that are generated by noise. Carefully tuning the event camera biases and filtering out outliers could improve the overall inverse depth estimation. Having a back-end inverse depth regularizer [22] could also help to mitigate this issue. + +The proposed method can only estimate one global motion. Thus, it can not adequately handle more than one motion simultaneously, e.g., due to cars moving [50]. Considering a multi-scale approach [1, 42] or explicitly modeling more than one possible motion occurring simultane + +ously [33, 44] are possible avenues for future research. + +# 5. Conclusion + +We have proposed a novel method that estimates the TTCM using a single event camera. The proposed method builds on the DMin framework to incrementally estimate local and global quantities, i.e., inverse depth, and global motion, respectively. We have also proposed an approach that effectively prevents event collapse for incremental event-based estimation without introducing regularizers or additional hyper-parameters. The proposed method also achieves state-of-the-art performance in TTC estimation in terms of accuracy and computational runtime while achieving competitive performance in optical flow estimation. Broadly, the proposed work further builds on the increasing amount of evidence that event cameras are especially suited to address visual motion-based problems; in particular, it further shows that incremental event-based processing can provide a flexible and general methodology to consider when using event cameras, which avoids issues introduced when converting events to other representations, e.g., batches and frames. + +Acknowledgment. This research received funding from the French National Research Agency (ANR), under grant agreement $\mathrm{N}^{\circ}$ ANR-20-CE23-0021, "AgileNeuRobot". + +# References + +[1] Himanshu Akolkar, Sio Hoi Ieng, and Ryad Benosman. Real-time High Speed Motion Prediction Using Fast Aperture-Robust Event-Driven Visual Flow. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–1, 2020. 2, 7, 8 +[2] Mohammed Almatrafi, Raymond Baldwin, Kiyoharu Aizawa, and Keigo Hirakawa. Distance Surface for Event-Based Optical Flow. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1547-1556, 2020. 2 +[3] Myo Tun Aung, Rodney Teo, and Garrick Orchard. Event-based Plane-fitting Optical Flow for Dynamic Vision Sensors in FPGA. In IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-5, 2018. 2, 7 +[4] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-Based Visual Flow. IEEE Transactions on Neural Networks and Learning Systems, 25(2):407-417, 2014. 1, 2 +[5] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous Frameless Event-based Optical Flow. Neural Networks, 27:32-37, 2012. 2 +[6] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \times 180$ dB 3 μs Latency Global Shutter Spatiotemporal Vision Sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1, 6 +[7] Vincent Brebion, Julien Moreau, and Franck Davoine. Real-Time Optical Flow for Vehicular Perception With Low- and High-Resolution Event Cameras. IEEE Transactions on Intelligent Transportation Systems, 23(9):15066-15078, 2022. 2, 7 +[8] William Chamorro, Juan Andrade-Cetto, and Joan Solà. High-Speed Event-Based Camera Tracking. In British Machine Vision Conference (BMVC), 2020. 2 +[9] William Chamorro, Joan Solà, and Juan Andrade-Cetto. Event-Based Line SLAM in Real-Time. IEEE Robotics and Automation Letters, 7(3):8146-8153, July 2022. 2 +[10] Xavier Clady, Charles Clercq, Sio-Hoi Ieng, Fouzhan Houseini, Marco Randazzo, Lorenzo Natale, Chiara Bartolozzi, and Ryad Benosman. Asynchronous Visual Event-based Time-to-Contact. Frontiers in Neuroscience, 8, 2014. 1, 2 +[11] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-Temporal Recurrent Networks for Event-Based Optical Flow Estimation. In AAAI Conference on Artificial Intelligence, volume 36, pages 525-533, 2022. 3, 7 +[12] Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dynamic obstacle avoidance for quadrotors with event cameras. Science Robotics, 5(40):eaaz9712, 2020. 1, 2 +[13] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-Based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):154-180, 2022. 1, 2 +[14] Guillermo Gallego, Jon E.A. Lund, Elias Mueggler, Henri Rebecq, Tobi Delbruck, and Davide Scaramuzza. Event- + +Based, 6-DOF Camera Tracking from Photometric Depth Maps. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(10):2402-2412, 2018. 2 +[15] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A Unifying Contrast Maximization Framework for Event Cameras, with Applications to Motion, Depth, and Optical Flow Estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3867-3876, 2018. 1, 2, 3, 8 +[16] Daniel Gehrig, Antonio Loquercio, Konstantinos Derpanis, and Davide Scaramuzza. End-to-End Learning of Representations for Asynchronous Event-Based Data. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5632–5642, 2019. 2 +[17] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense Optical Flow from Event Cameras. In International Conference of 3D Vision (3DV), pages 197–206, 2021. 1 +[18] Germain Haessig, Xavier Berthelon, Sio-Hoi Ieng, and Ryad Benosman. A Spiking Neural Network Model of Depth from Defocus for Event-based Neuromorphic Vision. Scientific Reports, 9(1):3744, 2019. 2 +[19] Jesse Hagenaars, Federico Paredes-Valles, and Guido de Croon. Self-Supervised Learning of Event-Based Optical Flow with Spiking Neural Networks. In Advances in Neural Information Processing Systems (NIPS), volume 34, pages 7167–7179, 2021. 3, 7 +[20] Botao He, Haojia Li, Siyuan Wu, Dong Wang, Zhiwei Zhang, Qianli Dong, Chao Xu, and Fei Gao. FAST-Dynamic-Vision: Detection and Tracking Dynamic Objects with Event and Depth Sensing. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3071-3078, 2021. 1 +[21] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew Davison. Simultaneous Mosaicing and Tracking with an Event Camera. In British Machine Vision Conference (BMVC). British Machine Vision Association (BMVA), 2014. 1, 2 +[22] Hanme Kim, Stefan Leutenegger, and Andrew J. Davison. Real-Time 3D Reconstruction and 6-DoF Tracking with an Event Camera. In Computer Vision – ECCV, Lecture Notes in Computer Science, pages 349–364. Springer International Publishing, 2016. 1, 2, 8 +[23] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E. Shi, and Ryad B. Benosman. HOTS: A Hierarchy of Event-Based Time-Surfaces for Pattern Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(7):1346-1359, July 2017. 1 +[24] Chankyu Lee, Adarsh Kumar Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, and Kaushik Roy. Spike-FlowNet: Event-Based Optical Flow Estimation with Energy-Efficient Hybrid Neural Networks. In Computer Vision - ECCV, Lecture Notes in Computer Science, pages 366-382. Springer International Publishing, 2020. 3 +[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \times 128$ dB 15 μs Latency Asynchronous Temporal Contrast Vision Sensor. IEEE Journal of Solid-State Circuits, 43(2):566-576, 2008. 1 + +[26] Min Liu and T. Delbruck. Adaptive Time-Slice Block-Matching Optical Flow Algorithm for Dynamic Vision Sensors. In British Machine Vision Conference (BMVC), 2018, 2, 7 +[27] Min Liu and Tobi Delbruck. EDFLOW: Event Driven Optical Flow Camera With Keypoint Detection and Adaptive Block Matching. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5776-5789, 2022. 1, 2, 3, 6, 7 +[28] Sofia McLeod, Gabriele Meoni, Dario Izzo, Anne Mergy, Daqi Liu, Yasir Latif, Ian Reid, and Tat-Jun Chin. Globally Optimal Event-Based Divergence Estimation for Ventral Landing, 2022. 1, 2, 5, 6, 7, 9, 10 +[29] Moritz B. Milde, Olivier J.N. Bertrand, Ryad Benosmanz, Martin Egelhaaf, and Elisabetta Chicca. Bioinspired event-driven collision avoidance algorithm based on optic flow. In International Conference on Event-based Control, Communication, and Signal Processing (EBCCSP), pages 1-7, 2015. 1, 2 +[30] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The Event-Camera Dataset and Simulator: Event-based Data for Pose Estimation, Visual Odometry, and SLAM. The International Journal of Robotics Research, 36(2):142-149, 2017. 4 +[31] Urbano Miguel Nunes, Ryad Benosman, and Sio-Hoi Ieng. Adaptive Global Decay Process for Event Cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9771-9780, 2023. 5 +[32] Urbano Miguel Nunes and Yiannis Demiris. Entropy Minimisation Framework for Event-based Vision Model Estimation. In Computer Vision - ECCV, Lecture Notes in Computer Science, pages 161-176. Springer International Publishing, 2020. 3 +[33] Urbano Miguel Nunes and Yiannis Demiris. Kinematic Structure Estimation of Arbitrary Articulated Rigid Objects for Event Cameras. In International Conference on Robotics and Automation (ICRA), pages 508-514, 2022. 1, 8 +[34] Urbano Miguel Nunes and Yiannis Demiris. Robust Event-Based Vision Model Estimation by Dispersion Minimisation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9561-9573, 2022. 1, 2, 3, 8 +[35] Federico Paredes-Vallés and Guido C. H. E. de Croon. Back to Event Basics: Self-Supervised Learning of Image Reconstruction for Event Cameras via Photometric Constancy. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3445-3454, 2021. 3, 8 +[36] Bas J. Pijnacker Hordijk, Kirk Y. W. Scheper, and Guido C. H. E. de Croon. Vertical Landing for Micro Air Vehicles using Event-based Optical Flow. Journal of Field Robotics, 35(1):69-90, 2018. 1, 2 +[37] Christoph Posch, Teresa Serrano-Gotarredona, Bernabe Linares-Barranco, and Tobi Delbruck. Retinomorphic Event-Based Vision Sensors: Bioinspired Cameras With Spiking Output. Proceedings of the IEEE, 102(10):1470-1484, 2014. 2 +[38] Henri Rebecq, Timo Horstschaefer, Guillermo Gallego, and Davide Scaramuzza. EVO: A Geometric Approach to Event-Based 6-DOF Parallel Tracking and Mapping in Real Time. + +IEEE Robotics and Automation Letters, 2(2):593-600, 2017. 2 +[39] Juan Pablo Rodríguez-Gómez, Raul Tapia, Maria del Mar Guzmán Garcia, Jose Ramiro Martínez-de Dios, and Anibal Ollero. Free as a Bird: Event-Based Dynamic Sense-and-Avoid for Ornithopter Robot Flight. IEEE Robotics and Automation Letters, 7(2):5413-5420, 2022. 1, 2 +[40] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event Collapse in Contrast Maximization Frameworks. Sensors, 22(14):5190, 2022. 1, 2, 3 +[41] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. A Fast Geometric Regularizer to Mitigate Event Collapse in the Contrast Maximization Framework. Advanced Intelligent Systems, 2022. 1, 2, 3 +[42] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of Event-Based Optical Flow. In Computer Vision – ECCV, Lecture Notes in Computer Science, pages 628–645. Springer Nature Switzerland, 2022. 2, 6, 7, 8 +[43] Olaf Sikorski, Dario Izzo, and Gabriele Meoni. Event-based Spacecraft Landing using Time-to-Contact. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1941–1950, 2021. ISSN: 2160-7516. 1, 2 +[44] Timo Stoffregen, Guillermo Gallego, Tom Drummond, Lindsay Kleeman, and Davide Scaramuzza. Event-Based Motion Segmentation by Motion Compensation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 7243-7252, 2019. 1, 8 +[45] Timo Stoffregen and Lindsay Kleeman. Simultaneous Optical Flow and Segmentation (SOFAS) using Dynamic Vision Sensor. In Australasian Conference on Robotics and Automation (ACRA), pages 52-61. Australian Robotics and Automation Association (ARAA), 2017. 1 +[46] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the Sim-to-Real Gap for Event Cameras. In Computer Vision - ECCV, Lecture Notes in Computer Science, pages 534-549. Springer International Publishing, 2020. 3, 7 +[47] Celyn Walters and Simon Hadfield. EVReflex: Dense Timeto-Impact Prediction for Event-based Obstacle Avoidance. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1304-1309, 2021. 1, 2 +[48] Chengxi Ye, Anton Mitrokhin, Cornelia Fermüller, James A. Yorke, and Yiannis Aloimonos. Unsupervised Learning of Dense Optical Flow, Depth and Egomotion with Event-Based Sensors. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5831-5838, 2020. 3 +[49] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based Feature Tracking with Probabilistic Data Association. In IEEE International Conference on Robotics and Automation (ICRA), pages 4465-4470, 2017. 2 +[50] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The Multivehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018. 5, 6, 7, 8 + +[51] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. EV-FlowNet: Self-Supervised Optical Flow Estimation for Event-based Cameras. In Proceedings of Robotics: Science and Systems, 2018. 1, 6, 7 +[52] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised Event-Based Learning of Optical Flow, Depth, and Egomotion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 989-997, 2019. 2, 3, 7, 8 \ No newline at end of file diff --git a/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/images.zip b/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9aab6fcad0772598a13055383593bca5537291f6 --- /dev/null +++ b/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc21a8c84d16a2a998aaa8b5b16533657bae4e3f2facd6fe6f4e9bb4309edb72 +size 619212 diff --git a/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/layout.json b/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6e111c7dd0247e560db75c4b25fec343cd0301ab --- /dev/null +++ b/2023/Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera/layout.json @@ -0,0 +1,10040 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 531, + 140 + ], + "type": "text", + "content": "Time-to-Contact Map by Joint Estimation of Up-to-Scale Inverse Depth and Global Motion using a Single Event Camera" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 160, + 455, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 160, + 455, + 176 + ], + "spans": [ + { + "bbox": [ + 137, + 160, + 455, + 176 + ], + "type": "text", + "content": "Urbano Miguel Nunes1, Laurent Udo Perrinet2 and Sio-Hoi Ieng1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 177, + 523, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 177, + 523, + 203 + ], + "spans": [ + { + "bbox": [ + 68, + 177, + 523, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 68, + 177, + 523, + 203 + ], + "type": "text", + "content": "Sorbonne University, " + }, + { + "bbox": [ + 68, + 177, + 523, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 68, + 177, + 523, + 203 + ], + "type": "text", + "content": "Aix Marseille Univ, CNRS, INT, Institut de Neurosciences de la Timone urbano.goncalves-nunes@inserm.fr, laurent.perrinet@univ-amu.fr, sio-hoi.ieng@upmc.fr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 258, + 290, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 486 + ], + "type": "text", + "content": "Event cameras asynchronously report brightness changes with a temporal resolution in the order of microseconds, which makes them inherently suitable to address problems that involve rapid motion perception. In this paper, we address the problem of time-to-contact (TTC) estimation using a single event camera. This problem is typically addressed by estimating a single global TTC measure, which explicitly assumes that the surface/obstacle is planar and fronto-parallel. We relax this assumption by proposing an incremental event-based method to estimate the TTC that jointly estimates the (up-to scale) inverse depth and global motion using a single event camera. The proposed method is reliable and fast while asynchronously maintaining a TTC map (TTCM), which provides per-pixel TTC estimates. As a side product, the proposed method can also estimate per-event optical flow. We achieve state-of-the-art performances on TTC estimation in terms of accuracy and runtime per event while achieving competitive performance on optical flow estimation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 128, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": "Event cameras differ from standard frame-based cameras, which capture visual data at a fixed rate and independently of the observing scene. Instead, event cameras respond asynchronously to pixel-wise brightness changes by generating events [6, 25]. Event cameras are thus data-driven sensors that offer several advantages, including high temporal resolution in the order of microseconds, low latency, low power consumption, and high dynamic range. These properties place event cameras as suitable candidates to address vision-based problems that involve (high-speed) motion, e.g., optical flow estimation [4, 27, 51], ego-motion estimation [15, 21, 22, 34], motion segmentation [33, 44], and obstacle avoidance [10, 12]. Due to the distinct visual sensing paradigm, however, new methods are necessary to fully exploit the potential of event cameras [13, 23]." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 232, + 540, + 373 + ], + "blocks": [ + { + "bbox": [ + 312, + 232, + 540, + 373 + ], + "lines": [ + { + "bbox": [ + 312, + 232, + 540, + 373 + ], + "spans": [ + { + "bbox": [ + 312, + 232, + 540, + 373 + ], + "type": "image", + "image_path": "cba9e1000800efaf48a731f8b951e7fde7238357dc06ad4e3210505788e1f80e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 376, + 546, + 399 + ], + "lines": [ + { + "bbox": [ + 305, + 376, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 546, + 399 + ], + "type": "text", + "content": "Figure 1. Runtime vs. accuracy comparison for TTC estimation methods. Average results on the Ventral Landing benchmark [28]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 546, + 555 + ], + "type": "text", + "content": "Event cameras have been used to address the problem of fast TTC estimation, which comes up often in the vision-based obstacle avoidance [10, 12, 29, 39] and ventral landing [28, 36] literature. The TTC is the time that would elapse before a camera reaches an obstacle/surface, assuming the current relative motion between them remains constant [10]. Previous methods that use a single event camera have focused on estimating a single global TTC measure, which assumes that the surface is planar and fronto-parallel. To overcome this limitation, other methods use additional sensing, e.g., depth frames, to build a dense TTCM [20, 47]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "We instead propose to extend the Dispersion Minimization (DMin) framework [34] to estimate the TTC for each incoming event using a single event camera. The proposed method jointly estimates the relative global motion and per-event (up-to scale) inverse depth. We can then asynchronously maintain a semi-dense TTCM which provides per-pixel TTC estimates or compute a global TTC measure with greater accuracy by averaging over the TTC estimates. Since there is at least one scaling degree of freedom (DOF), we also propose an effective strategy to mitigate event collapse [40, 41]. The proposed method is also computationally fast, reaching " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\sim 1" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": " microsecond processing time per event on a standard laptop. Fig. 1 compares the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23653" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "runtime vs. accuracy for TTC estimation methods, whereby the proposed method achieves state-of-the-art performance. We also estimate the per-event optical flow as a side product and achieve competitive performance compared to state-of-the-art optical flow methods that use events." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 133, + 136, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 136, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 136, + 144 + ], + "type": "text", + "content": "Main contributions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 145, + 287, + 240 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 145, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 287, + 180 + ], + "type": "text", + "content": "1. First event-based method that explicitly estimates the TTC for each event and maintains a semi-dense TTCM using a single event camera." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 181, + 286, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 181, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 54, + 181, + 286, + 216 + ], + "type": "text", + "content": "2. DMin framework [34] extension to jointly handle local and global estimates, i.e., inverse depth and global motion, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 217, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 217, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 286, + 240 + ], + "type": "text", + "content": "3. Effective approach that mitigates event collapse [40, 41] for incremental event-based estimation." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 250, + 147, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 250, + 147, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 250, + 147, + 262 + ], + "type": "text", + "content": "1.1. Time-to-Contact" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "text", + "content": "Consider a freely moving camera with angular and linear velocities " + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\omega(t) = (\\omega_x(t), \\omega_y(t), \\omega_z(t))^{\\top}" + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{\\nu}(t) = (\\nu_x(t), \\nu_y(t), \\nu_z(t))^{\\top}" + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "text", + "content": ", respectively, that is observing a point " + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "text", + "content": ", with 3D coordinates " + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\pmb{\\alpha}(t) = (X(t), Y(t), Z(t))^{\\top}" + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "text", + "content": " relative to the camera, as shown in Fig. 2. " + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "inline_equation", + "content": "Z(t)" + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "text", + "content": " is also referred as the depth of point " + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "text", + "content": " relative to the camera. The instantaneous TTC between the moving camera and point " + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 269, + 287, + 369 + ], + "type": "text", + "content": " is thus given by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 369, + 287, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 369, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 113, + 369, + 287, + 399 + ], + "type": "interline_equation", + "content": "\\tau (t) := - \\frac {Z (t)}{\\frac {d Z (t)}{d t}} = \\frac {Z (t)}{\\nu_ {z} (t)}. \\tag {1}", + "image_path": "69626488b8d0f52ba3bad451d5760c82068f8078bdd2eb7d6f755fac918ad627.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 404, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 475 + ], + "type": "text", + "content": "The minus sign disappears because we define the linear velocity w.r.t. the camera's frame of reference, not w.r.t. the point's frame of reference, i.e., " + }, + { + "bbox": [ + 46, + 404, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\nu_{z}(t) = -dZ(t) / dt" + }, + { + "bbox": [ + 46, + 404, + 287, + 475 + ], + "type": "text", + "content": ". Based on Eq. (1), the exact values of depth and relative approaching motion do not need to be estimated, only the ratio between them." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 489, + 134, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 134, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 134, + 501 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 510, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 544 + ], + "type": "text", + "content": "We review recent related works on the following topics: TTC, global motion and optical flow estimation. We refer to [13] for a detailed survey." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 713 + ], + "type": "text", + "content": "Time-to-Contact Estimation. The first work on event-based TTC using a single event camera relied solely on the estimation of visual motion flows [10], whereby the motion flows were computed by fitting a local plane to the time surface [4]. Other works that followed were geared towards two main use cases, namely obstacle avoidance [12, 29, 39] and ventral landing [28, 36, 43]. Event-based obstacle avoidance methods are built to be fast reacting and, although they come from either bio-inspired [29, 39] or mathematically grounded principles [12], they typically rely on empirically-validated heuristics to speed-up computations. Existing event-based ventral landing approaches only compute a single TTC estimate, which assumes that the surface is planar and fronto-parallel. To overcome this assumption," + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 325, + 69, + 529, + 173 + ], + "blocks": [ + { + "bbox": [ + 325, + 69, + 529, + 173 + ], + "lines": [ + { + "bbox": [ + 325, + 69, + 529, + 173 + ], + "spans": [ + { + "bbox": [ + 325, + 69, + 529, + 173 + ], + "type": "image", + "image_path": "605e5596d26923c5bebe65252c2d25c2c3fb8001614de17f6958618fdca4b29e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 178, + 493, + 190 + ], + "lines": [ + { + "bbox": [ + 358, + 178, + 493, + 190 + ], + "spans": [ + { + "bbox": [ + 358, + 178, + 493, + 190 + ], + "type": "text", + "content": "Figure 2. Camera coordinate system." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 210, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 258 + ], + "type": "text", + "content": "other works fuse events with additional sensory information, e.g. depth [47]. The proposed approach uses a single event camera while being mathematically grounded and computationally fast." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 258, + 546, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 546, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 546, + 545 + ], + "type": "text", + "content": "Global Motion Estimation. Also denoted by ego-motion estimation [13], it refers to estimating the parameters that explain the triggered events according to some global motion model. These methods can be broadly characterized by whether they rely on key-frame registration [8, 9, 14, 21, 22, 38] or perform the estimation without relying on any key-frame [15, 34]. The former methods are reminiscent of the frame-based paradigm and also include methods based on artificial neural networks (ANN) [16, 52], whereby an intermediate frame-based representation is needed for the estimation. The latter methods tend towards a more event-based processing paradigm and include methods based on spiking neural networks (SNN) [18, 37], whereby events are either processed on an event-by-event basis [34] or in batches [15]. Although both approaches have advantages and disadvantages, methods that rely on key-frame registration typically require an intermediate frame-based representation, which is still an open problem in the event-based community. Similarly to [22], the proposed method jointly estimates up-to scale inverse depth and global motion. By building on the DMin framework [34], which also allows processing events on an event-by-event basis, our method does not rely on key-frame registration or background inverse depth regularization to improve convergence." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 546, + 713 + ], + "type": "text", + "content": "Optical Flow Estimation. Several model-based methods have been previously proposed, which can be further divided into: frame-based [2, 7, 26, 27], batch-based [42, 49], and event-based [1, 3, 4]. By selecting the most relevant events, i.e., typically the most recent, frame-based methods build frames from which the optical flow is computed using techniques from standard image-based optical flow, e.g., Lucas-Kanade [5]. Since each event does not carry much information on its own, batch-based methods aggregate the most recent events by forming batches but perform the computations directly on the events. Event-based methods follow the most event-driven paradigm by performing event-by-event processing, typically being the computationally fastest. However, event-based methods tend to suffer more" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23654" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": "from the aperture problem since all the computations are performed locally, and thus frame-based and model-based methods achieve currently better accuracy. In terms of accuracy, learning-based ANN methods [11, 35, 46, 52] generally achieve state-of-the-art performance. Besides the need to convert events into frames for more efficient processing, these methods are known to be very data hungry, sensitive to the training data [48], and consume large amounts of energy [27]. Another line of research in learning-based methods has been to use SNN [19, 24], which combine the event-based processing and learning paradigms and thus do not require an intermediate frame-based representation. However, it is not trivial to train SNN, and the empirical validation is still not on par with ANN methods. Although it is not the primary objective of this work, the proposed method can provide per-event flow estimates while still being competitive in terms of accuracy w.r.t. state-of-the-art methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 285, + 104, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 104, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 104, + 298 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 305, + 288, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 288, + 376 + ], + "type": "text", + "content": "In this section, we describe the proposed incremental event-based method for TTCM estimation. We first briefly review the event cameras' working principle, and the DMin framework [34], based on which we develop the proposed method. Refer to the supplementary material for the full mathematical derivations and additional details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 382, + 282, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 282, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 282, + 396 + ], + "type": "text", + "content": "3.1. Event Cameras and Dispersion Minimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "content": "Event cameras output a stream of asynchronous temporal contrast events " + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\{e_i\\}, i \\in \\mathbb{N}" + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "content": ". Each event " + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "content": " represents a spatio-temporal asynchronous brightness change, being defined as a tuple " + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "inline_equation", + "content": "e_i := (\\mathbf{x}_i, t_i, p_i)" + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = (x_i, y_i)" + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "content": " are the pixel coordinates, " + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "content": " is the timestamp at which the event was generated, and " + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "inline_equation", + "content": "p_i \\in \\{-1, +1\\}" + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "content": " is its polarity. An event " + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "content": " is generated when the change in log-brightness " + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\log \\mathbf{I}_{x,y}(t) := \\bar{\\mathbf{I}}_{x,y}(t)" + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "content": " is above a threshold " + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "inline_equation", + "content": "L" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 501, + 287, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 287, + 517 + ], + "type": "interline_equation", + "content": "\\left| \\Delta \\bar {\\mathbf {I}} _ {x _ {i}, y _ {i}} \\left(t _ {i}\\right) \\right| = \\left| \\bar {\\mathbf {I}} _ {x _ {i}, y _ {i}} \\left(t _ {i}\\right) - \\bar {\\mathbf {I}} _ {x _ {i}, y _ {i}} \\left(t _ {i} - \\Delta t _ {i}\\right) \\right| \\geq L, \\tag {2}", + "image_path": "7f6de42661512bd6f2b2da121944dede91dd9b762c8ed90e4b65a93544cc6ef3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 521, + 286, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 286, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 286, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 521, + 286, + 532 + ], + "type": "inline_equation", + "content": "\\Delta t_{i}" + }, + { + "bbox": [ + 47, + 521, + 286, + 532 + ], + "type": "text", + "content": " is the time since the last event at the same pixel." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 532, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 287, + 593 + ], + "type": "text", + "content": "The DMin framework [34] estimates the parameters " + }, + { + "bbox": [ + 47, + 532, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 47, + 532, + 287, + 593 + ], + "type": "text", + "content": " of a transformation model " + }, + { + "bbox": [ + 47, + 532, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 532, + 287, + 593 + ], + "type": "text", + "content": " from the stream of events " + }, + { + "bbox": [ + 47, + 532, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e_i\\}_{i=1}^{N_e}" + }, + { + "bbox": [ + 47, + 532, + 287, + 593 + ], + "type": "text", + "content": " by minimizing a dispersion measure of the transformed events " + }, + { + "bbox": [ + 47, + 532, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i = \\mathcal{M}(e_i; t_{\\mathrm{ref}}, \\pmb{\\theta})" + }, + { + "bbox": [ + 47, + 532, + 287, + 593 + ], + "type": "text", + "content": ". We consider the Potential measure with a Gaussian kernel " + }, + { + "bbox": [ + 47, + 532, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{x}; \\pmb{\\mu}, \\pmb{\\Sigma})" + }, + { + "bbox": [ + 47, + 532, + 287, + 593 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 103, + 597, + 287, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 597, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 103, + 597, + 287, + 632 + ], + "type": "interline_equation", + "content": "P (\\mathcal {E}; \\boldsymbol {\\theta}) = - \\sum_ {i, j} ^ {N _ {e}} \\mathcal {N} \\left(\\mathbf {f} _ {i}; \\mathbf {f} _ {j}, \\mathbf {I}\\right), \\tag {3}", + "image_path": "62c5435cde7e71aed3c060680e3558586c05a7103af2b51c699fb7eaf29d7818.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "content": " is the identity matrix. A key distinction of the DMin framework is that it allows to incrementally estimate the model parameters " + }, + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "content": " on an event-by-event basis, whereby the model parameters " + }, + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "content": " can be iteratively solved by linearizing the transformation model " + }, + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 635, + 287, + 696 + ], + "type": "text", + "content": ", such that:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 52, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {i} = \\mathcal {M} \\left(e _ {i}; t _ {\\mathrm {r e f}}, \\boldsymbol {\\theta}\\right) = \\tilde {\\mathbf {f}} _ {i} + \\Delta t _ {i, \\mathrm {r e f}} \\mathbf {B} _ {i} \\boldsymbol {\\theta}, \\mathbf {B} _ {i} := \\mathbf {B} \\left(e _ {i}\\right), (4)", + "image_path": "1c2d923052fd255140c86d94ef1fc5589fbeee408c1f3d994f32aba53a491b46.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " is the model-dependent linearization matrix. The parameters " + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 305, + 72, + 545, + 97 + ], + "type": "text", + "content": " that minimize Eq. (3) are thus given by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 102, + 545, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 102, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 102, + 545, + 144 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} ^ {*} = - \\left(\\sum_ {i, j} ^ {N _ {e}} \\mathbf {W} _ {i, j} \\Delta \\mathbf {C} _ {i, j}\\right) ^ {- 1} \\left(\\sum_ {i, j} ^ {N _ {e}} \\mathbf {W} _ {i, j} \\Delta \\tilde {\\mathbf {f}} _ {i, j}\\right), \\tag {5}", + "image_path": "25d2c7786647dc9d3a7a6f6ff067322338ea31fe24c2ba8653ba1ed13644c26c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 148, + 547, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 148, + 547, + 176 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 547, + 176 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 148, + 547, + 176 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{C}_{i,j} = \\mathbf{C}_i - \\mathbf{C}_j" + }, + { + "bbox": [ + 305, + 148, + 547, + 176 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 148, + 547, + 176 + ], + "type": "inline_equation", + "content": "\\Delta \\tilde{\\mathbf{f}}_{i,j} = \\tilde{\\mathbf{f}}_i - \\tilde{\\mathbf{f}}_j" + }, + { + "bbox": [ + 305, + 148, + 547, + 176 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 148, + 547, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{i,j} = w_{i,j}\\partial \\Delta \\mathbf{f}_{i,j}^{\\top} / \\partial \\pmb {\\theta}" + }, + { + "bbox": [ + 305, + 148, + 547, + 176 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 148, + 547, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i = \\Delta t_{i,\\mathrm{ref}}\\mathbf{B}_i" + }, + { + "bbox": [ + 305, + 148, + 547, + 176 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 148, + 547, + 176 + ], + "type": "inline_equation", + "content": "w_{i,j} = \\mathcal{N}(\\mathbf{f}_i;\\mathbf{f}_j,\\mathbf{I})" + }, + { + "bbox": [ + 305, + 148, + 547, + 176 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 180, + 488, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 488, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 488, + 193 + ], + "type": "text", + "content": "3.2. Adapted Dispersion Minimization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 198, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 437 + ], + "type": "text", + "content": "While the DMin method [34] provides a general framework for global incremental event-based model estimation, it can also be adapted to jointly estimate global and local measures, i.e., global angular and linear velocities and local inverse depth. However, the DMin framework may encounter estimation issues when the global model has at least one scaling DOF, as noted in [34, 40, 41], known as event collapse. Event collapse occurs when the events are transformed into a single point, which minimizes the events' dispersion or maximizes the image contrast while the parameters' estimates diverge. So far, to the best of our knowledge, the mitigation discussion in the literature has been on how to constrain the optimization loss to discourage divergent estimates by analyzing the effects of the scaling transformations on the event-based data. Several mitigation strategies have thus been proposed on the events [34, 40] and parameters level [41] by adding terms to regularize the objective measure. While these strategies generally prevent event collapse, they increase the complexity of the optimization framework and introduce additional parameters to tune." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "text", + "content": "We instead observe that event collapse fundamentally stems from the event transformation to a common time reference " + }, + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "text", + "content": ", e.g., given by Eq. (4), by identifying two problems: 1) the constant velocity assumption may not hold depending on the time difference " + }, + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\Delta t_{i,\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "text", + "content": ", and 2) there is no built-in constraint on the magnitude of the model parameters, e.g., such that the difference between transformed events " + }, + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i - \\mathbf{f}_j" + }, + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "text", + "content": " explicitly penalizes divergent estimates; although, according to Eq. (4), " + }, + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_j" + }, + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "text", + "content": " individually diverge if the model parameters also diverge, their difference " + }, + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i - \\mathbf{f}_j" + }, + { + "bbox": [ + 304, + 438, + 545, + 641 + ], + "type": "text", + "content": " is not guaranteed to diverge and thus penalize divergent estimates. The first problem is typically addressed by heuristically making the time difference as short as possible, and its effects are of limited significance in practice. The second problem, however, is intrinsically linked to batch-based processing since the alignment of the events in a batch needs to be measured in some common time reference [15, 32]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "However, for incremental event-based processing, the event transformation and, consequently, the dispersion measure can be modified without loss of generality such that the event collapse is prevented by implicitly addressing the two problems identified. Fig. 3 depicts the idea whereby we only transform the event neighbors to the current event's" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "23655" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 74, + 190, + 133 + ], + "blocks": [ + { + "bbox": [ + 69, + 74, + 190, + 133 + ], + "lines": [ + { + "bbox": [ + 69, + 74, + 190, + 133 + ], + "spans": [ + { + "bbox": [ + 69, + 74, + 190, + 133 + ], + "type": "image", + "image_path": "1b248b34a5449923deacc9fef9846fd2b30875d43fadf59b6a6f2e9d15c7a1a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 137, + 287, + 171 + ], + "lines": [ + { + "bbox": [ + 46, + 137, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 46, + 137, + 287, + 171 + ], + "type": "text", + "content": "Figure 3. Incremental event collapse mitigation. Instead of transforming all the events to some time reference (left), we locally transform the events to the current event's timestamp (right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 206, + 74, + 273, + 132 + ], + "blocks": [ + { + "bbox": [ + 206, + 74, + 273, + 132 + ], + "lines": [ + { + "bbox": [ + 206, + 74, + 273, + 132 + ], + "spans": [ + { + "bbox": [ + 206, + 74, + 273, + 132 + ], + "type": "image", + "image_path": "929f61df89c36391bdca5a91846302bf95f20a7e025a504bc707f0ebfa479b2b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 190, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 190, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 286, + 251 + ], + "type": "text", + "content": "timestamp instead of transforming all the events to a common time reference, including the current event. Formally, instead of transforming the events according to Eq. (4), only the neighboring events " + }, + { + "bbox": [ + 46, + 190, + 286, + 251 + ], + "type": "inline_equation", + "content": "e_j" + }, + { + "bbox": [ + 46, + 190, + 286, + 251 + ], + "type": "text", + "content": " of the current event " + }, + { + "bbox": [ + 46, + 190, + 286, + 251 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 46, + 190, + 286, + 251 + ], + "type": "text", + "content": " are transformed according to" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 258, + 287, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 258, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 52, + 258, + 287, + 274 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {j} ^ {\\prime} = \\mathbf {x} _ {j} + \\Delta t _ {i, j} \\mathbf {B} _ {j} \\boldsymbol {\\theta} = \\mathbf {x} _ {j} + \\mathbf {C} _ {i, j} \\boldsymbol {\\theta}, e _ {j} \\in \\operatorname {n e i g h} (e _ {i}), \\tag {6}", + "image_path": "877fc3580e21f365aaec0daec7eb02af2cdad972564d255ca8d1c76d978fb90c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "text", + "content": "where 'neigh' is shorthand for neighborhood, and the dependency of " + }, + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j^\\prime" + }, + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "text", + "content": " was omitted for brevity. From the resultant residual " + }, + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{i,j} = \\mathbf{x}_i - \\mathbf{x}_j^\\prime = \\Delta \\mathbf{x}_{i,j} - \\mathbf{C}_{i,j}\\pmb {\\theta}" + }, + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "text", + "content": ", we see that the proposed modification to the DMin framework addresses both identified problems. First, the constant velocity assumption is better held since the time difference satisfies " + }, + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\Delta t_{i,j}\\leq \\Delta t_{i,\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "text", + "content": ". Second, if the model parameters " + }, + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "text", + "content": " diverge, then the residual " + }, + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{i,j}" + }, + { + "bbox": [ + 46, + 280, + 287, + 424 + ], + "type": "text", + "content": " also diverges, which effectively penalizes divergent estimates. We highlight that the proposed adaptation only works for incremental event-based processing: it is not suitable for batch-based processing since the proposed transformation only works locally." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 424, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 460 + ], + "type": "text", + "content": "The Potential measure is modified by computing the difference between the current event's coordinates " + }, + { + "bbox": [ + 46, + 424, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 424, + 287, + 460 + ], + "type": "text", + "content": " and the neighboring events' transformed coordinates " + }, + { + "bbox": [ + 46, + 424, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j'" + }, + { + "bbox": [ + 46, + 424, + 287, + 460 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 469, + 287, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 469, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 61, + 469, + 287, + 496 + ], + "type": "interline_equation", + "content": "P \\left(\\mathcal {E}; \\boldsymbol {\\theta}\\right) = - \\sum_ {i, j} \\mathcal {N} \\left(\\mathbf {x} _ {j} ^ {\\prime}; \\mathbf {x} _ {i}, \\mathbf {I}\\right), e _ {j} \\in \\operatorname {n e i g h} \\left(e _ {i}\\right). \\tag {7}", + "image_path": "32658ee4b4098ce26ba1c267f3751f6e0e853eff811db59c767351e4317eeb3a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 503, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 540 + ], + "type": "text", + "content": "By minimizing Eq. (7) by linearizing the residual according to Taylor's formula " + }, + { + "bbox": [ + 46, + 503, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{i,j}(\\pmb {\\theta} + \\Delta \\pmb {\\theta})\\approx \\mathbf{r}_{i,j}(\\pmb {\\theta}) + \\mathbf{J}_{i,j}(\\pmb {\\theta})\\Delta \\pmb{\\theta}" + }, + { + "bbox": [ + 46, + 503, + 287, + 540 + ], + "type": "text", + "content": ", the optimized model parameters " + }, + { + "bbox": [ + 46, + 503, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}^{*}" + }, + { + "bbox": [ + 46, + 503, + 287, + 540 + ], + "type": "text", + "content": " are iteratively updated:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 559, + 287, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 559, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 64, + 559, + 287, + 573 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} ^ {*} \\leftarrow \\boldsymbol {\\theta} ^ {*} + \\Delta \\boldsymbol {\\theta}, \\tag {8}", + "image_path": "37cd91da76849d823a1394ca982493ddacbe7c0f8aea2bcf56624cfa4abad6ae.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 60, + 575, + 274, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 575, + 274, + 631 + ], + "spans": [ + { + "bbox": [ + 60, + 575, + 274, + 631 + ], + "type": "interline_equation", + "content": "\\Delta \\boldsymbol {\\theta} = - \\underbrace {\\left(\\sum_ {j} w _ {i , j} \\mathbf {J} _ {i , j} ^ {\\top} \\mathbf {J} _ {i , j}\\right)} _ {\\boldsymbol {\\Psi}} ^ {- 1} \\underbrace {\\left(\\sum_ {j} w _ {i , j} \\mathbf {J} _ {i , j} ^ {\\top} \\mathbf {r} _ {i , j}\\right)} _ {\\boldsymbol {\\psi}},", + "image_path": "e38c2edd55b56ddd616c7f850afc69f7ca768416b5c4cbc88de0465ffd9f5a01.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 639, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 286, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 639, + 286, + 665 + ], + "type": "inline_equation", + "content": "w_{i,j} = \\mathcal{N}\\left(\\mathbf{x}_j';\\mathbf{x}_i,\\mathbf{I}\\right)" + }, + { + "bbox": [ + 46, + 639, + 286, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 639, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{i,j} = \\partial \\mathbf{r}_{i,j} / \\partial \\pmb{\\theta}" + }, + { + "bbox": [ + 46, + 639, + 286, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 639, + 286, + 665 + ], + "type": "inline_equation", + "content": "e_j\\in" + }, + { + "bbox": [ + 46, + 639, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{neigh}(e_i)" + }, + { + "bbox": [ + 46, + 639, + 286, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 671, + 259, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 259, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 259, + 684 + ], + "type": "text", + "content": "3.3. Inverse Depth and Global Motion Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We consider that a calibrated event camera can freely move and whose global motion is parameterized by the 3D" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 542, + 163 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 542, + 163 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 542, + 163 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 542, + 163 + ], + "type": "image", + "image_path": "69a437785dceb881ab23e23df94a55bf5358cba59b3f598fc4af52bcd562db5d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 168, + 539, + 180 + ], + "lines": [ + { + "bbox": [ + 306, + 168, + 539, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 539, + 180 + ], + "type": "text", + "content": "Figure 4. Typical per-event estimation of optical flow and depth." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 198, + 545, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 249 + ], + "type": "text", + "content": "angular and linear velocities, as defined in Sec. 1.1, " + }, + { + "bbox": [ + 305, + 198, + 545, + 249 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} = (\\pmb{\\nu}^{\\mathsf{T}},\\pmb{\\omega}^{\\mathsf{T}})^{\\mathsf{T}}" + }, + { + "bbox": [ + 305, + 198, + 545, + 249 + ], + "type": "text", + "content": ". For each event " + }, + { + "bbox": [ + 305, + 198, + 545, + 249 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 305, + 198, + 545, + 249 + ], + "type": "text", + "content": ", we estimate its inverse depth " + }, + { + "bbox": [ + 305, + 198, + 545, + 249 + ], + "type": "inline_equation", + "content": "\\rho_{i}\\coloneqq 1 / Z_{i}" + }, + { + "bbox": [ + 305, + 198, + 545, + 249 + ], + "type": "text", + "content": ", based on the well-known expression for the apparent velocity on the image plane:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 309, + 256, + 545, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 256, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 309, + 256, + 545, + 282 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {i} = \\left( \\begin{array}{l l} \\rho_ {i} \\mathbf {V} _ {i} & \\boldsymbol {\\Omega} _ {i} \\end{array} \\right) \\binom {\\boldsymbol {\\nu}} {\\omega} = \\rho_ {i} \\mathbf {V} _ {i} \\boldsymbol {\\nu} + \\boldsymbol {\\Omega} _ {i} \\boldsymbol {\\omega} := \\mathbf {B} _ {i} \\boldsymbol {\\theta}, \\tag {9}", + "image_path": "0ec8dadd78d8fda4456e8674704383bbd0472a047ca318047aa2ebe4dc133157.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 284, + 432, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 284, + 432, + 310 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 432, + 310 + ], + "type": "interline_equation", + "content": "\\mathbf {V} _ {i} = \\left( \\begin{array}{c c c} - f _ {x} & 0 & x _ {i} - c _ {x} \\\\ 0 & - f _ {y} & y _ {i} - c _ {y} \\end{array} \\right),", + "image_path": "f48077edaa0c387d89504e2d373e591e16fcb4547ca2e8831abd70718c464701.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 312, + 542, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 312, + 542, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 542, + 348 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Omega} _ {i} = \\left( \\begin{array}{l l l} \\frac {(x _ {i} - c _ {x}) (y _ {i} - c _ {y})}{f _ {y}} & - f _ {x} - \\frac {(x _ {i} - c _ {x}) ^ {2}}{f _ {x}} & (y _ {i} - c _ {y}) \\frac {f _ {x}}{f _ {y}} \\\\ f _ {y} + \\frac {(y _ {i} - c _ {y}) ^ {2}}{f _ {y}} & - \\frac {(x _ {i} - c _ {x}) (y _ {i} - c _ {y})}{f _ {x}} & - (x _ {i} - c _ {x}) \\frac {f _ {y}}{f _ {x}} \\end{array} \\right)", + "image_path": "0e82d1490253090068cd5c4f3c44a73fbc20b9cd91b1fb9b440291f37658606c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": "where we make explicit the dependency on the camera intrinsic parameters, namely the horizontal and vertical focal lengths " + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "inline_equation", + "content": "f_{y}" + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": ", respectively, and the horizontal and vertical focal center coordinates " + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "inline_equation", + "content": "c_{x}" + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "inline_equation", + "content": "c_{y}" + }, + { + "bbox": [ + 304, + 354, + 545, + 449 + ], + "type": "text", + "content": ", respectively. In this paper, we assume that the focal center coordinates represent the focus of expansion (FOE). Fig. 4 shows the typical per-event estimation of optical flow and depth using the proposed method." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 456, + 430, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 430, + 469 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 430, + 469 + ], + "type": "text", + "content": "3.4. Time-to-Contact Map" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "text", + "content": "For each event " + }, + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "text", + "content": ", we estimate its inverse depth " + }, + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\rho_{i}" + }, + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "text", + "content": " and update the global motion parameters " + }, + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} = (\\pmb{\\nu}^{\\mathsf{T}},\\pmb{\\omega}^{\\mathsf{T}})^{\\mathsf{T}}" + }, + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "text", + "content": ". The model parameters " + }, + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\gamma_{i}" + }, + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "text", + "content": " are formed by stacking the motion parameters and inverse depth " + }, + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\gamma_{i} = (\\pmb{\\theta}^{\\mathsf{T}},\\rho_{i})^{\\mathsf{T}}" + }, + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "text", + "content": ". We impose a smoothness constraint to Eq. (6), so that neighboring inverse depth estimates are assumed to be equal to " + }, + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\rho_{i}" + }, + { + "bbox": [ + 304, + 475, + 545, + 549 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 343, + 556, + 545, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 556, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 343, + 556, + 545, + 571 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {j} ^ {\\prime} = \\mathbf {x} _ {j} + \\Delta t _ {i, j} \\mathbf {B} _ {i, j} \\boldsymbol {\\theta}, e _ {j} \\in \\operatorname {n e i g h} \\left(e _ {i}\\right), \\tag {10}", + "image_path": "0b86c379575b29e5101191137e38299a89830428a16588e89bb5acb68b7ac517.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 578, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 578, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{i,j} = (\\rho_i\\mathbf{V}_j\\boldsymbol {\\Omega}_j)" + }, + { + "bbox": [ + 304, + 578, + 545, + 665 + ], + "type": "text", + "content": ". The iterative update " + }, + { + "bbox": [ + 304, + 578, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\Delta \\gamma_{i}" + }, + { + "bbox": [ + 304, + 578, + 545, + 665 + ], + "type": "text", + "content": " is given by Eq. (8), where " + }, + { + "bbox": [ + 304, + 578, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{i,j} = -\\Delta t_{i,j}\\left(\\mathbf{B}_{i,j}\\mathbf{V}_j\\boldsymbol {\\nu}\\right)" + }, + { + "bbox": [ + 304, + 578, + 545, + 665 + ], + "type": "text", + "content": ". We maintain the TTCM by computing the TTC for each event " + }, + { + "bbox": [ + 304, + 578, + 545, + 665 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 578, + 545, + 665 + ], + "type": "text", + "content": " based on Eq. (1): " + }, + { + "bbox": [ + 304, + 578, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\tau_{i} = 1 / (\\rho_{i}\\nu_{z})" + }, + { + "bbox": [ + 304, + 578, + 545, + 665 + ], + "type": "text", + "content": ". We can also estimate the global TTC by averaging over the values maintained in the TTCM. Fig. 5 shows the typical per-event estimation of optical flow and TTCM using the proposed method." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 671, + 389, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 389, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 389, + 682 + ], + "type": "text", + "content": "3.5. Initialization" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "The motion parameters " + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": " are global measures which are estimated by aggregating events, while each inverse depth" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23656" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 170, + 74, + 191, + 94 + ], + "blocks": [ + { + "bbox": [ + 170, + 74, + 191, + 94 + ], + "lines": [ + { + "bbox": [ + 170, + 74, + 191, + 94 + ], + "spans": [ + { + "bbox": [ + 170, + 74, + 191, + 94 + ], + "type": "image", + "image_path": "77d3fbbe11908f2e76c824e565ee59d80421fef507c962bd274b8dfc39da7c2a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 52, + 95, + 127, + 153 + ], + "blocks": [ + { + "bbox": [ + 52, + 77, + 125, + 90 + ], + "lines": [ + { + "bbox": [ + 52, + 77, + 125, + 90 + ], + "spans": [ + { + "bbox": [ + 52, + 77, + 125, + 90 + ], + "type": "text", + "content": "indoor_flying1 [50]" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 95, + 127, + 153 + ], + "lines": [ + { + "bbox": [ + 52, + 95, + 127, + 153 + ], + "spans": [ + { + "bbox": [ + 52, + 95, + 127, + 153 + ], + "type": "image", + "image_path": "38c2954c7bc377b2656f9f1a925c5cad92a8768ab8419ba5d93100c5ed65650b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 154, + 123, + 162 + ], + "lines": [ + { + "bbox": [ + 56, + 154, + 123, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 123, + 162 + ], + "type": "text", + "content": "Accumulated events" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 168, + 286, + 180 + ], + "lines": [ + { + "bbox": [ + 47, + 168, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 286, + 180 + ], + "type": "text", + "content": "Figure 5. Typical per-event estimation of optical flow and TTCM." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 129, + 95, + 225, + 153 + ], + "blocks": [ + { + "bbox": [ + 214, + 81, + 281, + 89 + ], + "lines": [ + { + "bbox": [ + 214, + 81, + 281, + 89 + ], + "spans": [ + { + "bbox": [ + 214, + 81, + 281, + 89 + ], + "type": "text", + "content": "1sec 10sec" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 95, + 225, + 153 + ], + "lines": [ + { + "bbox": [ + 129, + 95, + 225, + 153 + ], + "spans": [ + { + "bbox": [ + 129, + 95, + 225, + 153 + ], + "type": "image", + "image_path": "dbb593d55e20fc4aee47a2ccfe3b8472d6f709d90c2dddb7ceeaa10131e6edc4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 154, + 187, + 163 + ], + "lines": [ + { + "bbox": [ + 147, + 154, + 187, + 163 + ], + "spans": [ + { + "bbox": [ + 147, + 154, + 187, + 163 + ], + "type": "text", + "content": "Optical flow" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 225, + 95, + 284, + 153 + ], + "blocks": [ + { + "bbox": [ + 225, + 95, + 284, + 153 + ], + "lines": [ + { + "bbox": [ + 225, + 95, + 284, + 153 + ], + "spans": [ + { + "bbox": [ + 225, + 95, + 284, + 153 + ], + "type": "image", + "image_path": "382f43152d0220e9fcdc3a81ecf1955d3901ebf212e147e7f6e39c9066eb23f8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 154, + 258, + 163 + ], + "lines": [ + { + "bbox": [ + 234, + 154, + 258, + 163 + ], + "spans": [ + { + "bbox": [ + 234, + 154, + 258, + 163 + ], + "type": "text", + "content": "TTCM" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\rho_{i}" + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "text", + "content": " value corresponds to the estimate of a single event's inverse depth. The initialization of the motion parameters " + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "text", + "content": " can thus be almost arbitrarily set, e.g., typically set to 0, and it is only performed once at the beginning. However, a more careful initialization procedure must be considered for the case of inverse depth since it needs to be performed for each event. Based on recent advances in event-based global time decay [31], we perform a weighted average based on the previous neighboring events' inverse depth estimates as the initialization procedure. Each neighboring event's weight " + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "inline_equation", + "content": "w_{i}(t)" + }, + { + "bbox": [ + 46, + 200, + 287, + 332 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 341, + 287, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 341, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 113, + 341, + 287, + 367 + ], + "type": "interline_equation", + "content": "w _ {i} (t) = \\frac {1}{1 + a (t) \\left(t - t _ {i}\\right)}, \\tag {11}", + "image_path": "caa200ef4beba63305a9df6a6eec25fa83e435d46ca0c3100ff1b091af53ab35.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 375, + 287, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 423 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 375, + 287, + 423 + ], + "type": "inline_equation", + "content": "a(t)" + }, + { + "bbox": [ + 46, + 375, + 287, + 423 + ], + "type": "text", + "content": " is the global event activity. If there are no previous neighboring events' inverse depth estimates, which should only occur at the start of the estimation, the initial inverse depth estimate is set to 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 431, + 185, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 185, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 185, + 442 + ], + "type": "text", + "content": "3.6. Practical Considerations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": "As mentioned in Sec. 1.1, only the ratio between the depth and relative motion is required to compute the TTC. Since the depth and linear velocities estimates are obtained up-to a scale factor due to the monocular ambiguity, we constrain the linear velocities " + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\pmb{\\nu}" + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": " to have at most unit norm, i.e., " + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "inline_equation", + "content": "|\\pmb{\\nu}|_2 \\leq 1" + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": ", which is useful to improve the method's computational stability by bounding the allowed estimates' values. This is not related to event collapse; rather, it stems from the ratio given by Eq. (1), whereby we can introduce an arbitrary non-zero multiplicative scalar to the numerator and denominator and have the same TTC estimate." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "content": "Since the events are generated by 3D points in the field of view (FOV) of the camera and should have positive depth values, we constrain the depth values to be strictly positive by introducing a parameterization variable " + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\rho (\\lambda) = e^{\\lambda} > 0" + }, + { + "bbox": [ + 46, + 582, + 287, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "The iterative update given by Eq. (8) can be quite noisy since it is only computed in a small neighborhood " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{neigh}(e_i) = \\{e_k:|\\mathbf{x}_k - \\mathbf{x}_i|_{\\infty}\\leq s\\}" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": ". We thus introduce two prior parameters for the global motion parameters and the local inverse depth estimates " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "l_{\\theta}" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "inline_equation", + "content": "l_{\\rho}" + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": ", respectively. The resultant iterative update of the model parameters is given" + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 308, + 71, + 545, + 102 + ], + "blocks": [ + { + "bbox": [ + 308, + 71, + 545, + 102 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 545, + 102 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 545, + 102 + ], + "type": "table", + "html": "
Neighborhood size s\\( {w}_{\\text{thresh }} \\)\\( {l}_{\\theta } \\)\\( {l}_{\\rho } \\)
30.01\\( \\propto O \\)(camera resolution)1
", + "image_path": "5cf9c4eef45a08c841164aed3effaac431fadf5a15a79ab6eb868ae02d6c4a7f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 323, + 110, + 527, + 121 + ], + "lines": [ + { + "bbox": [ + 323, + 110, + 527, + 121 + ], + "spans": [ + { + "bbox": [ + 323, + 110, + 527, + 121 + ], + "type": "text", + "content": "Table 1. Hyper-parameters used across the experiments." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "text", + "content": "by " + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "inline_equation", + "content": "\\Delta \\pmb{\\gamma}_i = (\\pmb{\\Psi} + \\mathbf{L})^{-1}\\left(-\\pmb{\\psi} + \\mathbf{L}(\\pmb{\\gamma}_{\\mathrm{prev}} - \\pmb{\\gamma}_i^*)\\right)" + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{L} = \\mathrm{diag}(l_{\\pmb{\\theta}},\\dots ,l_{\\pmb{\\theta}},l_{\\rho})" + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "inline_equation", + "content": "\\pmb{\\gamma}_{\\mathrm{prev}}" + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "text", + "content": " are the parameters' estimates from the previous event " + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "inline_equation", + "content": "e_{i-1}" + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "text", + "content": ". We also weigh each event " + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "inline_equation", + "content": "e_j" + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "text", + "content": " contribution according to the corresponding weight " + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "inline_equation", + "content": "w_j(t)" + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "text", + "content": " given by Eq. (11), and discard any event whose weight is below a threshold " + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{thresh}}" + }, + { + "bbox": [ + 304, + 140, + 545, + 213 + ], + "type": "text", + "content": " [31]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 224, + 449, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 449, + 238 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 449, + 238 + ], + "type": "text", + "content": "4. Experimental Evaluation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 244, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 545, + 365 + ], + "type": "text", + "content": "We evaluate the proposed method in TTC estimation and optical flow estimation, given that the apparent velocity on the image plane can be estimated according to Eq. (9) and due to the lack of event datasets dedicated to TTC estimation. The optical flow benchmark also provides a common ground to compare the proposed method with other state-of-the-art methods that estimate optical flow. Tab. 1 presents the hyper-parameters that were used across the experiments. Since the proposed method computes per-event estimates, we only evaluate on the respective pixel locations1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 373, + 427, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 427, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 427, + 384 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "type": "text", + "content": "VL Dataset " + }, + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "type": "inline_equation", + "content": "[28]^2" + }, + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "type": "text", + "content": ". It consists of 7 real event sequences observing planar prints of landing surfaces and 1 real event sequence observing the 3D print of a landing surface. Each sequence has 15sec of duration, totaling 120sec of dataset duration. The events were recorded by a Prophesee event camera with " + }, + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 304, + 390, + 545, + 498 + ], + "type": "text", + "content": " resolution, and the ground truth (GT) depth measurements were recorded with an Intel RealSense camera. However, only a global GT depth measurement is provided per timestamp." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 499, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 545, + 571 + ], + "type": "text", + "content": "To comply with the evaluation reported [28], we assess the proposed method at certain timestamps that correspond to event batches of 0.5sec. The comparison metrics are the divergence REE (\\%) and runtime per event (microsec). The divergence " + }, + { + "bbox": [ + 304, + 499, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 499, + 545, + 571 + ], + "type": "text", + "content": " is the inverse of the TTC, i.e., " + }, + { + "bbox": [ + 304, + 499, + 545, + 571 + ], + "type": "inline_equation", + "content": "\\mu = 1 / \\tau" + }, + { + "bbox": [ + 304, + 499, + 545, + 571 + ], + "type": "text", + "content": ", and the REE is given by:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 380, + 579, + 545, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 579, + 545, + 607 + ], + "spans": [ + { + "bbox": [ + 380, + 579, + 545, + 607 + ], + "type": "interline_equation", + "content": "\\mathrm {R E E} = 1 0 0 \\frac {\\left| \\hat {\\mu} - \\mu_ {\\mathrm {g t}} \\right|}{\\left| \\mu_ {\\mathrm {g t}} \\right|}, \\tag {12}", + "image_path": "354461b0722b4e89c2136c75df3b8760ac3cad780272c2c24c5f095fdaa97b94.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 615, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 615, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 615, + 545, + 638 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 615, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}" + }, + { + "bbox": [ + 304, + 615, + 545, + 638 + ], + "type": "text", + "content": " is the estimate and " + }, + { + "bbox": [ + 304, + 615, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{gt}}" + }, + { + "bbox": [ + 304, + 615, + 545, + 638 + ], + "type": "text", + "content": " is the GT. The global motion prior " + }, + { + "bbox": [ + 304, + 615, + 545, + 638 + ], + "type": "inline_equation", + "content": "l_{\\theta}" + }, + { + "bbox": [ + 304, + 615, + 545, + 638 + ], + "type": "text", + "content": " was set to 1000." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 639, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 675 + ], + "type": "text", + "content": "MVSEC Dataset " + }, + { + "bbox": [ + 304, + 639, + 545, + 675 + ], + "type": "inline_equation", + "content": "[50]^3" + }, + { + "bbox": [ + 304, + 639, + 545, + 675 + ], + "type": "text", + "content": ". It consists of several real indoor and outdoor sequences, providing events, standard grayscale frames, IMU data, camera poses, and scene depth." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 683, + 493, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 683, + 493, + 693 + ], + "spans": [ + { + "bbox": [ + 315, + 683, + 493, + 693 + ], + "type": "text", + "content": "1 All the code will be made available upon acceptance." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 693, + 510, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 693, + 510, + 703 + ], + "spans": [ + { + "bbox": [ + 317, + 693, + 510, + 703 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 317, + 693, + 510, + 703 + ], + "type": "text", + "content": "https://github.com/s-mcleod/ventral-landing-event-dataset." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 318, + 703, + 451, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 703, + 451, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 703, + 451, + 712 + ], + "type": "text", + "content": "3 https://daniilidis-group.github.io/mvsec." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23657" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 497, + 251 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 497, + 251 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 497, + 251 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 497, + 251 + ], + "type": "table", + "html": "
Method2D-12D-22D-32D-42D-52D-62D-73DAvg.
REENCECMD [28]13.487.416.1112.1910.415.123.7212.908.92
||Proposed Scaling7.198.799.1312.184.987.794.9211.348.29
SProposed Translation5.037.026.175.254.614.346.186.545.64
3Proposed 6-DOF6.7713.7314.977.506.9112.1512.776.1010.11
3Proposed Scaling14.9915.1911.6921.6911.1915.4411.0119.4315.08
||Proposed Translation11.157.017.9813.806.3711.174.5512.469.31
SProposed 6-DOF6.516.817.256.776.065.596.186.196.42
RuntimeNCECMD [28]63.3762.5151.2141.3692.6938.9725.5731.5350.90
||Proposed Scaling0.820.780.800.810.700.840.780.700.78
SProposed Translation1.101.191.221.081.021.211.221.161.15
3Proposed 6-DOF1.401.431.451.441.371.541.601.471.46
3Proposed Scaling1.181.231.251.261.121.291.261.751.29
||Proposed Translation1.771.701.801.841.681.711.711.701.74
SProposed 6-DOF2.432.402.362.372.482.342.232.272.36
", + "image_path": "7fa15892468d64f7891765992992d5cb92140aee79571fa6f2d6aafb6350a99d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 287, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 371 + ], + "type": "text", + "content": "The events were recorded by a DAVIS [6] with " + }, + { + "bbox": [ + 46, + 300, + 287, + 371 + ], + "type": "inline_equation", + "content": "346 \\times 260" + }, + { + "bbox": [ + 46, + 300, + 287, + 371 + ], + "type": "text", + "content": " resolution. The evaluated sequences span approximately 265sec. The optical flow GT is also provided [51], generated from the scene depth and camera velocity. We generate GT TTCM's by applying Eq. (1) given the GT depth maps and camera velocity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 407 + ], + "type": "text", + "content": "To assess optical flow accuracy, we use the following metrics: average endpoint error (AEE) (in pixel/frame, as is conventional in the literature [42, 51])" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 415, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 415, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 63, + 415, + 287, + 449 + ], + "type": "interline_equation", + "content": "\\mathrm {A E E} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| \\hat {\\mathbf {m}} _ {i} - \\mathbf {m} _ {\\mathrm {g t}, i} \\right| _ {2}, \\mathbf {m} = \\Delta t _ {\\text {f r a m e}} \\mathbf {v}, \\tag {13}", + "image_path": "49b5c404da9fef89acbfb0bd683866d1dab2e5287ea59f174c1bdeaf91215988.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 481 + ], + "type": "text", + "content": "outliers (Out) as the percentage of pixels with AEE greater than 3, average relative endpoint error (AREE) " + }, + { + "bbox": [ + 46, + 456, + 287, + 481 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 456, + 287, + 481 + ], + "type": "text", + "content": " [27]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 489, + 287, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 489, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 97, + 489, + 287, + 522 + ], + "type": "interline_equation", + "content": "\\operatorname {A R E E} = \\frac {1 0 0}{N} \\sum_ {i = 1} ^ {N} \\frac {\\left| \\hat {\\mathbf {m}} _ {i} - \\mathbf {m} _ {\\mathrm {g t} , i} \\right| _ {2}}{\\left| \\mathbf {m} _ {\\mathrm {g t} , i} \\right| _ {2}}, \\tag {14}", + "image_path": "ba9c8bc976a4c4794d33552824020ecf6c4f7a6757bb00013989caef9f83ab40.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 530, + 216, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 216, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 216, + 544 + ], + "type": "text", + "content": "and average angular error (AAE) " + }, + { + "bbox": [ + 47, + 530, + 216, + 544 + ], + "type": "inline_equation", + "content": "(^{\\circ})" + }, + { + "bbox": [ + 47, + 530, + 216, + 544 + ], + "type": "text", + "content": " [27]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 90, + 551, + 287, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 551, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 90, + 551, + 287, + 583 + ], + "type": "interline_equation", + "content": "\\mathrm {A A E} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\operatorname {a r c c o s} \\frac {\\hat {\\mathbf {m}} _ {i} ^ {\\top} \\mathbf {m} _ {\\mathrm {g t} , i}}{| \\hat {\\mathbf {m}} _ {i} | _ {2} | \\mathbf {m} _ {\\mathrm {g t} , i} | _ {2}}. \\tag {15}", + "image_path": "44c67b46dd714796deff174af9e7620a8ad49ff8642b8500cb434a689c8e6f16.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "text", + "content": "To assess TTCM accuracy, we use the AREE " + }, + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "text", + "content": " between the divergence estimate " + }, + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}_i" + }, + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "text", + "content": " and corresponding GT " + }, + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{gt},i}" + }, + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "text", + "content": ". The global motion prior " + }, + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "inline_equation", + "content": "l_{\\theta}" + }, + { + "bbox": [ + 46, + 592, + 287, + 628 + ], + "type": "text", + "content": " was set to 100." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 206, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 206, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 206, + 647 + ], + "type": "text", + "content": "4.2. Other Global Motion Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": "Based on the general 6-DOF global motion model described in Sec. 3.3, we can consider other more constrained motion models depending on the application, as follows." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Translation. This model is parameterized by the 3D linear velocities " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{\\nu}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ". Thus, " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i = \\rho_i\\mathbf{V}_i" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ", being " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " given by Eq. (9)." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 297, + 544, + 350 + ], + "blocks": [ + { + "bbox": [ + 46, + 258, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 547, + 281 + ], + "type": "text", + "content": "Table 2. Global divergence estimation. Quantitative results on the VL dataset [28], in terms of REE (\\%) and runtime per event (microsec) averaged over 100 trials. Lower is better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 297, + 544, + 350 + ], + "lines": [ + { + "bbox": [ + 309, + 297, + 544, + 350 + ], + "spans": [ + { + "bbox": [ + 309, + 297, + 544, + 350 + ], + "type": "table", + "html": "
Modelindoor_flying1indoor_flying2outdoor_day1
Driving153.3084.8047.27
Translation19.8228.2345.78
6-DOF53.6556.1256.66
", + "image_path": "5e9337a16cfcb329198ac3b7c644b760b8cd8e72f3714e8f1018b43d4dae9e34.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 356, + 546, + 378 + ], + "lines": [ + { + "bbox": [ + 305, + 356, + 546, + 378 + ], + "spans": [ + { + "bbox": [ + 305, + 356, + 546, + 378 + ], + "type": "text", + "content": "Table 3. Divergence estimation. Quantitative results on the MVSEC dataset [50], in terms of AREE (\\%). Lower is better." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "text", + "content": "Driving. This model is parameterized by the most significant DOF's when driving a car, namely the angular velocity around the camera " + }, + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "text", + "content": "-axis " + }, + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\omega_{y}" + }, + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "text", + "content": " and the linear velocity in the " + }, + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "text", + "content": "-axis " + }, + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\nu_{z}" + }, + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "text", + "content": " (see Fig. 2). Hence, " + }, + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i = \\left( \\begin{array}{cc} \\rho_i(x_i - c_x) & -f_x - \\frac{(x_i - c_x)^2}{f_x} \\\\ \\rho_i(y_i - c_y) & -\\frac{(x_i - c_x)(y_i - c_y)}{f_x} \\end{array} \\right)" + }, + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "text", + "content": ". When using this motion model, we impose " + }, + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "inline_equation", + "content": "|\\nu_z|_2 = 1" + }, + { + "bbox": [ + 305, + 400, + 545, + 503 + ], + "type": "text", + "content": " to ensure that the depth is properly estimated." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 504, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 504, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 305, + 504, + 545, + 575 + ], + "type": "text", + "content": "Scaling. This model is parameterized by the linear velocity in the " + }, + { + "bbox": [ + 305, + 504, + 545, + 575 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 504, + 545, + 575 + ], + "type": "text", + "content": "-axis " + }, + { + "bbox": [ + 305, + 504, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\nu_{z}" + }, + { + "bbox": [ + 305, + 504, + 545, + 575 + ], + "type": "text", + "content": ". Hence, " + }, + { + "bbox": [ + 305, + 504, + 545, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i = \\rho_i\\left( \\begin{array}{c}x_i - c_x\\\\ y_i - c_y \\end{array} \\right)" + }, + { + "bbox": [ + 305, + 504, + 545, + 575 + ], + "type": "text", + "content": ". It is only considered since the global motion model used in [28] is the 1-DOF scaling, which assumes that the scene is planar and it does not estimate the depth." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 586, + 362, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 362, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 362, + 597 + ], + "type": "text", + "content": "4.3. Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "Time-to-Contact. Tab. 2 reports the results on global divergence estimation on the VL benchmark [28]. The proposed method using the Translation model with neighboring size " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "s = 2" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " achieves the best accuracy, outperforming ECMD [28] by " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "36.77\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " on average. The proposed method using the Scaling and full 6-DOF models also outperform ECMD [28] in terms of average accuracy with neighboring size " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "s = 2" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "s = 3" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": ", respectively. The results indicate that considering a smaller neighborhood benefits global mod" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23658" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 70, + 491, + 224 + ], + "blocks": [ + { + "bbox": [ + 106, + 70, + 491, + 224 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 491, + 224 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 491, + 224 + ], + "type": "table", + "html": "
Methodindoor_flying1indoor_flying2indoor_flying3outdoor_day1
AEEOutAEEOutAEEOutAEEOut
LBEV-FlowNet (MB) [52]0.580.001.024.000.873.000.320.00
EV-FlowNet (HQF) [46]0.561.000.661.000.591.000.681.00
Ding et al. [11]0.570.100.791.600.721.300.420.00
ConvGRU-EV-FlowNet [19]0.600.511.178.060.935.640.470.25
MBFBBrebion et al. [7]0.520.100.985.500.712.100.530.20
BBShiba et al. [42]0.420.100.600.590.500.280.300.10
EBAung et al. [3]----2.31-1.26-
ARMS [1]1.52-1.59-1.89-2.75-
Proposed Driving1.407.502.4226.322.1426.050.390.12
Proposed Translation0.630.410.942.830.791.650.550.97
Proposed 6-DOF0.881.971.6012.021.389.990.812.26
", + "image_path": "e30d9f8e74d5418c6363cae011c3b9dee6db5549a71b141cf0bcb14d28b08e67.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 273, + 288, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 288, + 369 + ], + "type": "text", + "content": "els with fewer parameters; conversely, a larger neighborhood benefits models with more parameters. Even though the motion for all sequences is predominantly dominated by just 1 scaling DOF, the results suggest that considering motion models with additional DOFs, e.g., Translation and 6-DOF, is beneficial. The extra DOFs may explain other small motions, whereas these small motions would just be considered noise for the Scaling model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "content": "In terms of runtime per event, the proposed method outperforms ECMD [28] by between " + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "inline_equation", + "content": "95.36\\%" + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "inline_equation", + "content": "98.47\\%" + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "content": ", achieving real-time processing for all the sequences in the VL dataset [28], being capable of processing between 420k and 1.28M events per second. The results indicate that the runtime increases with the neighboring size " + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 372, + 287, + 456 + ], + "type": "text", + "content": " and the number of parameters of the global motion model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 459, + 288, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 288, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 288, + 674 + ], + "type": "text", + "content": "Tab. 3 reports the results on divergence estimation on the MVSEC benchmark [50]. The Translation model achieves the best performance overall, indicating that the corresponding DOFs are sufficient to explain the perceived motions in the sequences evaluated while minimizing the optimization complexity. As mentioned in Sec. 4.2, the Driving model is tuned to the outdoor_day1 sequence, whose performance is on par with the Translation model. However, it performs poorly on the indoor_flying3 sequence since it can not handle more complex types of motions that are present. Being the most general model, the 6-DOF motion model achieves similar performance for all the sequences. Its performance is worse than the Translation model due to the increased optimization complexity, while the additional DOFs do not contribute to improving the accuracy. The proposed method achieves real-time processing in the MVSEC dataset [50] since, on average, the sequences exhibit a maximum of around 400k events per second." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Optical Flow. Since it is difficult to compare directly the results reported in Tab. 3, we evaluate the proposed method on optical flow estimation, and compare the results with other" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 270, + 538, + 374 + ], + "blocks": [ + { + "bbox": [ + 46, + 231, + 547, + 254 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 547, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 547, + 254 + ], + "type": "text", + "content": "Table 4. Optical flow estimation. Quantitative results on the MVSEC dataset [50], in terms of AEE (pixel/frame) and Out (\\%). Lower is better. Legend: learning-based (LB), model-based (MB), frame-based (FB), batch-based (BB), event-based (EB)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 270, + 538, + 374 + ], + "lines": [ + { + "bbox": [ + 315, + 270, + 538, + 374 + ], + "spans": [ + { + "bbox": [ + 315, + 270, + 538, + 374 + ], + "type": "table", + "html": "
Methodindoor_flying3outdoor_day1
AREEAAEAREEAAE
EV-FlowNet [51]63.6132.5565.5422.88
Aung et al. [3]95.6469.35122.1069.41
ABMOF [26]52.6916.5790.1333.76
EDFLOW [27]37.5212.1569.4023.30
Proposed Driving104.0560.3650.2513.31
Proposed Translation44.4611.3872.6721.77
Proposed 6-DOF95.2321.35122.0431.68
", + "image_path": "89e651a1eadbae73151fc9f5d4f1fd66b2f3d480f5ebc0d2a64ea2df147edf02.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 381, + 547, + 415 + ], + "lines": [ + { + "bbox": [ + 305, + 381, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 381, + 547, + 415 + ], + "type": "text", + "content": "Table 5. Optical flow estimation. Quantitative results on the MVSEC dataset [50], in terms of AREE (\\%) and AAE (\\%). Lower is better." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 441, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 514 + ], + "type": "text", + "content": "methods. Tab. 4 reports the results on optical flow estimation on the MVSEC benchmark [50] in terms of AEE and Out, as is commonly found in the literature. The proposed method achieves state-of-the-art performance over the EB methods, on par performance with the FB method, and competitive performance overall." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 517, + 546, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 546, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 546, + 673 + ], + "type": "text", + "content": "Tab. 5 reports the results on optical flow estimation on the MVSEC benchmark [50] in terms of AREE and AAE. The proposed method achieves state-of-the-art performance on the outdoor_day1 sequence using the Driving model and on par performance on the indoor_flying3 sequence using the Translation model. The results on Tab. 5 indicate that the proposed method is comparatively more accurate in estimating the direction of the flow, i.e., compared with the other methods, the proposed method achieves lower values of AAE overall. In terms of accuracy, this suggests that more improvements can be achieved by improving the method's per-event inverse depth estimation since inverse depth estimates mainly contribute to the flow magnitude." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Qualitative Results. Fig. 6 presents qualitative results on sequences of the MVSEC dataset [50]. The estimated TTCM and optical flow resemble the GT ones." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23659" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 73, + 523, + 371 + ], + "blocks": [ + { + "bbox": [ + 69, + 73, + 523, + 371 + ], + "lines": [ + { + "bbox": [ + 69, + 73, + 523, + 371 + ], + "spans": [ + { + "bbox": [ + 69, + 73, + 523, + 371 + ], + "type": "image", + "image_path": "d20894387040a86101b5cf6f7fc906801d12351d33524eb5ce5d8e3592aa7dd5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 376, + 400, + 389 + ], + "lines": [ + { + "bbox": [ + 192, + 376, + 400, + 389 + ], + "spans": [ + { + "bbox": [ + 192, + 376, + 400, + 389 + ], + "type": "text", + "content": "Figure 6. Qualitative results on the MVSEC dataset [50]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 407, + 125, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 125, + 419 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 125, + 419 + ], + "type": "text", + "content": "4.4. Limitations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 545 + ], + "type": "text", + "content": "Similarly to other event-based methods [15, 22, 35, 52], the proposed method also inherits the brightness assumption from the DMin framework [34]. It can thus provide wrongful estimates for events that are not caused by motion, e.g., due to flickering lights. While this limitation is somewhat mitigated when estimating global quantities, it can struggle to reliably estimate the inverse depth of events that are not caused by motion. Introducing probabilistic uncertainties to the estimates could alleviate this issue while also improving the inverse depth's initialization procedure." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 653 + ], + "type": "text", + "content": "Also related, although we explicitly impose a local smoothness constraint to the inverse depth estimates, given by Eq. (10), the proposed method can still provide inverse depth estimates that differ significantly from the neighboring inverse depth estimates. This typically occurs for events that are generated by noise. Carefully tuning the event camera biases and filtering out outliers could improve the overall inverse depth estimation. Having a back-end inverse depth regularizer [22] could also help to mitigate this issue." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "The proposed method can only estimate one global motion. Thus, it can not adequately handle more than one motion simultaneously, e.g., due to cars moving [50]. Considering a multi-scale approach [1, 42] or explicitly modeling more than one possible motion occurring simultane" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 407, + 527, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 527, + 421 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 527, + 421 + ], + "type": "text", + "content": "ously [33, 44] are possible avenues for future research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 430, + 379, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 430, + 379, + 442 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 379, + 442 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 450, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 676 + ], + "type": "text", + "content": "We have proposed a novel method that estimates the TTCM using a single event camera. The proposed method builds on the DMin framework to incrementally estimate local and global quantities, i.e., inverse depth, and global motion, respectively. We have also proposed an approach that effectively prevents event collapse for incremental event-based estimation without introducing regularizers or additional hyper-parameters. The proposed method also achieves state-of-the-art performance in TTC estimation in terms of accuracy and computational runtime while achieving competitive performance in optical flow estimation. Broadly, the proposed work further builds on the increasing amount of evidence that event cameras are especially suited to address visual motion-based problems; in particular, it further shows that incremental event-based processing can provide a flexible and general methodology to consider when using event cameras, which avoids issues introduced when converting events to other representations, e.g., batches and frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgment. This research received funding from the French National Research Agency (ANR), under grant agreement " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{N}^{\\circ}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " ANR-20-CE23-0021, \"AgileNeuRobot\"." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23660" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Himanshu Akolkar, Sio Hoi Ieng, and Ryad Benosman. Real-time High Speed Motion Prediction Using Fast Aperture-Robust Event-Driven Visual Flow. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–1, 2020. 2, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 191 + ], + "type": "text", + "content": "[2] Mohammed Almatrafi, Raymond Baldwin, Kiyoharu Aizawa, and Keigo Hirakawa. Distance Surface for Event-Based Optical Flow. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(7):1547-1556, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 235 + ], + "type": "text", + "content": "[3] Myo Tun Aung, Rodney Teo, and Garrick Orchard. Event-based Plane-fitting Optical Flow for Dynamic Vision Sensors in FPGA. In IEEE International Symposium on Circuits and Systems (ISCAS), pages 1-5, 2018. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "text", + "content": "[4] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-Based Visual Flow. IEEE Transactions on Neural Networks and Learning Systems, 25(2):407-417, 2014. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 323 + ], + "type": "text", + "content": "[5] Ryad Benosman, Sio-Hoi Ieng, Charles Clercq, Chiara Bartolozzi, and Mandyam Srinivasan. Asynchronous Frameless Event-based Optical Flow. Neural Networks, 27:32-37, 2012. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 369 + ], + "type": "text", + "content": "[6] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A " + }, + { + "bbox": [ + 53, + 324, + 288, + 369 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 53, + 324, + 288, + 369 + ], + "type": "text", + "content": " dB 3 μs Latency Global Shutter Spatiotemporal Vision Sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "text", + "content": "[7] Vincent Brebion, Julien Moreau, and Franck Davoine. Real-Time Optical Flow for Vehicular Perception With Low- and High-Resolution Event Cameras. IEEE Transactions on Intelligent Transportation Systems, 23(9):15066-15078, 2022. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 424, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 458 + ], + "type": "text", + "content": "[8] William Chamorro, Juan Andrade-Cetto, and Joan Solà. High-Speed Event-Based Camera Tracking. In British Machine Vision Conference (BMVC), 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 458, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 491 + ], + "type": "text", + "content": "[9] William Chamorro, Joan Solà, and Juan Andrade-Cetto. Event-Based Line SLAM in Real-Time. IEEE Robotics and Automation Letters, 7(3):8146-8153, July 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 535 + ], + "type": "text", + "content": "[10] Xavier Clady, Charles Clercq, Sio-Hoi Ieng, Fouzhan Houseini, Marco Randazzo, Lorenzo Natale, Chiara Bartolozzi, and Ryad Benosman. Asynchronous Visual Event-based Time-to-Contact. Frontiers in Neuroscience, 8, 2014. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 591 + ], + "type": "text", + "content": "[11] Ziluo Ding, Rui Zhao, Jiyuan Zhang, Tianxiao Gao, Ruiqin Xiong, Zhaofei Yu, and Tiejun Huang. Spatio-Temporal Recurrent Networks for Event-Based Optical Flow Estimation. In AAAI Conference on Artificial Intelligence, volume 36, pages 525-533, 2022. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "type": "text", + "content": "[12] Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dynamic obstacle avoidance for quadrotors with event cameras. Science Robotics, 5(40):eaaz9712, 2020. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 690 + ], + "type": "text", + "content": "[13] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-Based Vision: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):154-180, 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 690, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 715 + ], + "type": "text", + "content": "[14] Guillermo Gallego, Jon E.A. Lund, Elias Mueggler, Henri Rebecq, Tobi Delbruck, and Davide Scaramuzza. Event-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Based, 6-DOF Camera Tracking from Photometric Depth Maps. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(10):2402-2412, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 107, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 547, + 172 + ], + "type": "text", + "content": "[15] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A Unifying Contrast Maximization Framework for Event Cameras, with Applications to Motion, Depth, and Optical Flow Estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3867-3876, 2018. 1, 2, 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 173, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 228 + ], + "type": "text", + "content": "[16] Daniel Gehrig, Antonio Loquercio, Konstantinos Derpanis, and Davide Scaramuzza. End-to-End Learning of Representations for Asynchronous Event-Based Data. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 5632–5642, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 272 + ], + "type": "text", + "content": "[17] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense Optical Flow from Event Cameras. In International Conference of 3D Vision (3DV), pages 197–206, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "type": "text", + "content": "[18] Germain Haessig, Xavier Berthelon, Sio-Hoi Ieng, and Ryad Benosman. A Spiking Neural Network Model of Depth from Defocus for Event-based Neuromorphic Vision. Scientific Reports, 9(1):3744, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 316, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 371 + ], + "type": "text", + "content": "[19] Jesse Hagenaars, Federico Paredes-Valles, and Guido de Croon. Self-Supervised Learning of Event-Based Optical Flow with Spiking Neural Networks. In Advances in Neural Information Processing Systems (NIPS), volume 34, pages 7167–7179, 2021. 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 371, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 437 + ], + "type": "text", + "content": "[20] Botao He, Haojia Li, Siyuan Wu, Dong Wang, Zhiwei Zhang, Qianli Dong, Chao Xu, and Fei Gao. FAST-Dynamic-Vision: Detection and Tracking Dynamic Objects with Event and Depth Sensing. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3071-3078, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 437, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 492 + ], + "type": "text", + "content": "[21] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew Davison. Simultaneous Mosaicing and Tracking with an Event Camera. In British Machine Vision Conference (BMVC). British Machine Vision Association (BMVA), 2014. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 492, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 548 + ], + "type": "text", + "content": "[22] Hanme Kim, Stefan Leutenegger, and Andrew J. Davison. Real-Time 3D Reconstruction and 6-DoF Tracking with an Event Camera. In Computer Vision – ECCV, Lecture Notes in Computer Science, pages 349–364. Springer International Publishing, 2016. 1, 2, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 603 + ], + "type": "text", + "content": "[23] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E. Shi, and Ryad B. Benosman. HOTS: A Hierarchy of Event-Based Time-Surfaces for Pattern Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(7):1346-1359, July 2017. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 603, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 669 + ], + "type": "text", + "content": "[24] Chankyu Lee, Adarsh Kumar Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, and Kaushik Roy. Spike-FlowNet: Event-Based Optical Flow Estimation with Energy-Efficient Hybrid Neural Networks. In Computer Vision - ECCV, Lecture Notes in Computer Science, pages 366-382. Springer International Publishing, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A " + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": " dB 15 μs Latency Asynchronous Temporal Contrast Vision Sensor. IEEE Journal of Solid-State Circuits, 43(2):566-576, 2008. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23661" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[26] Min Liu and T. Delbruck. Adaptive Time-Slice Block-Matching Optical Flow Algorithm for Dynamic Vision Sensors. In British Machine Vision Conference (BMVC), 2018, 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 171 + ], + "type": "text", + "content": "[27] Min Liu and Tobi Delbruck. EDFLOW: Event Driven Optical Flow Camera With Keypoint Detection and Adaptive Block Matching. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5776-5789, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "text", + "content": "[28] Sofia McLeod, Gabriele Meoni, Dario Izzo, Anne Mergy, Daqi Liu, Yasir Latif, Ian Reid, and Tat-Jun Chin. Globally Optimal Event-Based Divergence Estimation for Ventral Landing, 2022. 1, 2, 5, 6, 7, 9, 10" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 281 + ], + "type": "text", + "content": "[29] Moritz B. Milde, Olivier J.N. Bertrand, Ryad Benosmanz, Martin Egelhaaf, and Elisabetta Chicca. Bioinspired event-driven collision avoidance algorithm based on optic flow. In International Conference on Event-based Control, Communication, and Signal Processing (EBCCSP), pages 1-7, 2015. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 282, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 282, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 282, + 288, + 337 + ], + "type": "text", + "content": "[30] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The Event-Camera Dataset and Simulator: Event-based Data for Pose Estimation, Visual Odometry, and SLAM. The International Journal of Robotics Research, 36(2):142-149, 2017. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "text", + "content": "[31] Urbano Miguel Nunes, Ryad Benosman, and Sio-Hoi Ieng. Adaptive Global Decay Process for Event Cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9771-9780, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 382, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 288, + 437 + ], + "type": "text", + "content": "[32] Urbano Miguel Nunes and Yiannis Demiris. Entropy Minimisation Framework for Event-based Vision Model Estimation. In Computer Vision - ECCV, Lecture Notes in Computer Science, pages 161-176. Springer International Publishing, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 481 + ], + "type": "text", + "content": "[33] Urbano Miguel Nunes and Yiannis Demiris. Kinematic Structure Estimation of Arbitrary Articulated Rigid Objects for Event Cameras. In International Conference on Robotics and Automation (ICRA), pages 508-514, 2022. 1, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "type": "text", + "content": "[34] Urbano Miguel Nunes and Yiannis Demiris. Robust Event-Based Vision Model Estimation by Dispersion Minimisation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9561-9573, 2022. 1, 2, 3, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "type": "text", + "content": "[35] Federico Paredes-Vallés and Guido C. H. E. de Croon. Back to Event Basics: Self-Supervised Learning of Image Reconstruction for Event Cameras via Photometric Constancy. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3445-3454, 2021. 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 624 + ], + "type": "text", + "content": "[36] Bas J. Pijnacker Hordijk, Kirk Y. W. Scheper, and Guido C. H. E. de Croon. Vertical Landing for Micro Air Vehicles using Event-based Optical Flow. Journal of Field Robotics, 35(1):69-90, 2018. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 678 + ], + "type": "text", + "content": "[37] Christoph Posch, Teresa Serrano-Gotarredona, Bernabe Linares-Barranco, and Tobi Delbruck. Retinomorphic Event-Based Vision Sensors: Bioinspired Cameras With Spiking Output. Proceedings of the IEEE, 102(10):1470-1484, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 679, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 679, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 679, + 288, + 714 + ], + "type": "text", + "content": "[38] Henri Rebecq, Timo Horstschaefer, Guillermo Gallego, and Davide Scaramuzza. EVO: A Geometric Approach to Event-Based 6-DOF Parallel Tracking and Mapping in Real Time." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 72, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 72, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 545, + 94 + ], + "type": "text", + "content": "IEEE Robotics and Automation Letters, 2(2):593-600, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 149 + ], + "type": "text", + "content": "[39] Juan Pablo Rodríguez-Gómez, Raul Tapia, Maria del Mar Guzmán Garcia, Jose Ramiro Martínez-de Dios, and Anibal Ollero. Free as a Bird: Event-Based Dynamic Sense-and-Avoid for Ornithopter Robot Flight. IEEE Robotics and Automation Letters, 7(2):5413-5420, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 150, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 183 + ], + "type": "text", + "content": "[40] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event Collapse in Contrast Maximization Frameworks. Sensors, 22(14):5190, 2022. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "type": "text", + "content": "[41] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. A Fast Geometric Regularizer to Mitigate Event Collapse in the Contrast Maximization Framework. Advanced Intelligent Systems, 2022. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 271 + ], + "type": "text", + "content": "[42] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of Event-Based Optical Flow. In Computer Vision – ECCV, Lecture Notes in Computer Science, pages 628–645. Springer Nature Switzerland, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "type": "text", + "content": "[43] Olaf Sikorski, Dario Izzo, and Gabriele Meoni. Event-based Spacecraft Landing using Time-to-Contact. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1941–1950, 2021. ISSN: 2160-7516. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 382 + ], + "type": "text", + "content": "[44] Timo Stoffregen, Guillermo Gallego, Tom Drummond, Lindsay Kleeman, and Davide Scaramuzza. Event-Based Motion Segmentation by Motion Compensation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 7243-7252, 2019. 1, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "type": "text", + "content": "[45] Timo Stoffregen and Lindsay Kleeman. Simultaneous Optical Flow and Segmentation (SOFAS) using Dynamic Vision Sensor. In Australasian Conference on Robotics and Automation (ACRA), pages 52-61. Australian Robotics and Automation Association (ARAA), 2017. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 437, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 503 + ], + "type": "text", + "content": "[46] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the Sim-to-Real Gap for Event Cameras. In Computer Vision - ECCV, Lecture Notes in Computer Science, pages 534-549. Springer International Publishing, 2020. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 503, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 545, + 548 + ], + "type": "text", + "content": "[47] Celyn Walters and Simon Hadfield. EVReflex: Dense Timeto-Impact Prediction for Event-based Obstacle Avoidance. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1304-1309, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 548, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 613 + ], + "type": "text", + "content": "[48] Chengxi Ye, Anton Mitrokhin, Cornelia Fermüller, James A. Yorke, and Yiannis Aloimonos. Unsupervised Learning of Dense Optical Flow, Depth and Egomotion with Event-Based Sensors. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5831-5838, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 658 + ], + "type": "text", + "content": "[49] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based Feature Tracking with Probabilistic Data Association. In IEEE International Conference on Robotics and Automation (ICRA), pages 4465-4470, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 713 + ], + "type": "text", + "content": "[50] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The Multivehicle Stereo Event Camera Dataset: An Event Camera Dataset for 3D Perception. IEEE Robotics and Automation Letters, 3(3):2032-2039, 2018. 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23662" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 173 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[51] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. EV-FlowNet: Self-Supervised Optical Flow Estimation for Event-based Cameras. In Proceedings of Robotics: Science and Systems, 2018. 1, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 173 + ], + "type": "text", + "content": "[52] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised Event-Based Learning of Optical Flow, Depth, and Egomotion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 989-997, 2019. 2, 3, 7, 8" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23663" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/23a6116f-5b95-4dcd-97e5-95528a6d1fff_content_list.json b/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/23a6116f-5b95-4dcd-97e5-95528a6d1fff_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..40ad66e3d1ff7226d110a81b414b884b6c3e03b9 --- /dev/null +++ b/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/23a6116f-5b95-4dcd-97e5-95528a6d1fff_content_list.json @@ -0,0 +1,1830 @@ +[ + { + "type": "text", + "text": "Tiny Updater: Towards Efficient Neural Network-Driven Software Updating", + "text_level": 1, + "bbox": [ + 264, + 130, + 704, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Linfeng Zhang, Kaisheng Ma*", + "bbox": [ + 364, + 203, + 609, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Institute for Interdisciplinary Information Sciences, Tsinghua University", + "bbox": [ + 199, + 220, + 771, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhanglinfeng1997@outlook.com, kaisheng@mail.tsinghua.edu.cn", + "bbox": [ + 222, + 241, + 746, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Significant advancements have been accomplished with deep neural networks in diverse visual tasks, which have substantially elevated their deployment in edge device software. However, during the update of neural network-based software, users are required to download all the parameters of the neural network anew, which harms the user experience. Motivated by previous progress in model compression, we propose a novel training methodology named Tiny Updater to address this issue. Specifically, by adopting the variant of pruning and knowledge distillation methods, Tiny Updater can update the neural network-based software by only downloading a few parameters $(10\\% \\sim 20\\%)$ instead of all the parameters in the neural network. Experiments on eleven datasets of three tasks, including image classification, image-to-image translation, and video recognition have demonstrated its effectiveness. Codes have been released in https://github.com/ArchipLab-LinfengZhang/TinyUpdater.", + "bbox": [ + 75, + 321, + 473, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 625, + 209, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the availability of large-scale datasets [15, 16, 51] and high-performance computing platforms, deep neural networks have achieved remarkable achievements in various visual tasks such as image classification [18, 28, 75, 80], segmentation [53, 62], and object detection [50, 67, 68, 76]. Encouraged by their impressive performance, numerous software developers have effectively integrated neural networks into their software products and deployed them on edge devices such as mobile phones and tablet computers.", + "bbox": [ + 75, + 650, + 468, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Typically, the development roadmap for a neural network-based software follows the paradigm illustrate in Figure 1(a). Initially, users install the software by downloading all the neural network parameters from a cloud platform. Subsequently, as the software interacts with its users, an abundance of new training data and requirements can be", + "bbox": [ + 75, + 786, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/23d4b0134c1e078a5ef075e3a0efaec3351dc53b29617eba4311d559be81d6f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 292, + 883, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a41bc52d9b777ef4eab555d464f5060e5eab4385a97c04334fbd915434a9ba3c.jpg", + "image_caption": [ + "Figure 1. Comparison between (a) the traditional model updating scheme and (b) the proposed efficient model updating. In efficient model updating, only around $10\\%$ parameters in the neural network are actually changed and thus users only require to download these changed parameters for updating." + ], + "image_footnote": [], + "bbox": [ + 509, + 419, + 885, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "collected. Then, software developers may retrain the neural network with the gathered data and update their software to enhance its performance. However, during the retraining phase, all the neural network parameters are typically changed compared to their values before updating. Consequently, users are compelled to download all the parameters of the neural network again from the cloud platform, seriously impairing their experience. While recent research has explored text prompts and adapter layers to fine-tune a large-scale pre-trained model at a low training cost [52, 79], there have been no prior efforts to reduce the download cost during model updating, which has a more direct impact on users with edge devices.", + "bbox": [ + 496, + 657, + 892, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes the challenge of efficient model updating with the objective of reducing the download overhead of neural network-based software during updating. As", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23447", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "depicted in Figure 1(b), during the retraining phase, efficient model updating introduces an additional constraint limiting the change of only a small subset (e.g. $10\\%$ ) of the neural network parameters compared to the pre-updating model. Consequently, users are only required to download a few parameters that have actually changed instead of all the parameters in the neural network. In general, efficient model updating raises two questions: how to find the optimal parameters that should be changed in the neural network, and how to achieve comparable accuracy with the fully-updated model (i.e., the model which has all the parameters changed during updating).", + "bbox": [ + 73, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this challenge, we propose a novel neural network training framework named Tiny Updater. Motivated by previous works in model compression, Tiny Updater is composed of the variants of two typical model compression techniques - neural network pruning and knowledge distillation. Firstly, to determine which channels or layers in neural networks should be modified during updating, Tiny Updater applies the pruning technique. As shown in Figure 2, it iteratively calculates the $L_{1}$ -norm distance between the pre-updating model and the post-updating model parameters. The channels with smaller distances are deemed unnecessary for updating and are pruned to their original values before updating. Conversely, the channels with larger distances are considered essential for updating, and thus they are actually changed. Secondly, during the retraining period, we propose to improve the performance of the partially-updated model by distilling the knowledge from a fully-updated teacher model. The partially-updated student model is trained to give predictions that are similar to those of the fully-updated teacher model by optimizing the knowledge distillation loss. This ensures that the partially-updated model achieves comparable performance with the fully-updated model.", + "bbox": [ + 73, + 273, + 472, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments have been conducted on eleven datasets and three different tasks, including, image classification, video classification, and image-to-image translation, using seven different neural networks. Experimental results demonstrate that Tiny Updater can update the model by changing only $10\\%$ to $20\\%$ of the parameters with minimal performance degradation. In summary, the main contribution of this paper can be outlined as follows.", + "bbox": [ + 75, + 622, + 470, + 744 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we first propose the challenge of efficient model updating to reduce the communication cost of downloading neural network parameters to the edge devices for software updating.", + "- To tackle this challenge, we propose Tiny Updater, which prunes the fully-updated model to the pre-updating model and distills the knowledge from the fully-updated model to the partially-updated model." + ], + "bbox": [ + 94, + 762, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments have been conducted on three tasks with eleven datasets, including image classification, image-to-image translation and video recognition. Experimental results demonstrate that Tiny Updater can reduce the training overhead by $80\\% \\sim 90\\%$ with almost no performance degradation.", + "bbox": [ + 517, + 90, + 890, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 196, + 640, + 213 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Neural Network Pruning", + "text_level": 1, + "bbox": [ + 500, + 222, + 728, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural network pruning is one of the most effective methods for deep neural network compression by deleting the unimportant neurons [42, 47], filters [29, 38], channels [30, 90] and layers [8, 69]. LeCun et al. and Hassibi et al. first propose to prune neural networks depending on the Hessian matrix [27, 41]. Recently, Han et al. propose deep compression, which finds the important connection in neural networks with its absolute value and then iteratively prunes them [24]. Liu et al. train a meta-network to generate the weights of neural networks and then apply it to search the best-pruned architecture [58]. Ding et al. propose to apply a LSTM to learn the hierarchical characteristics of deep neural networks and then generates the corresponding pruning scheme [17]. Liu et al. propose joint multi-dimension pruning to prune the channels, layers and resolutions at the same time [59]. Frankle et al. propose the lottery ticket hypothesis and find that a standard pruning technique naturally uncovers subnetworks whose initializations made them capable of training effectively [21]. Besides classification, neural network pruning has also be utilized in image-to-image translation [37, 42], unconditional image generation [56], object detection [22, 66], video retrieval [10] and pre-trained language models [14].", + "bbox": [ + 496, + 246, + 890, + 593 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, Tiny Updater introduces neural network pruning methods to find the parameters which are actually valuable for model updating. Its main difference compared with previous pruning methods is that Tiny Updater prunes the parameters to their value before updating instead of zero, which makes it easier to be solved than traditional pruning. Our experiments also show that Tiny Updater can be utilized on the pruned models.", + "bbox": [ + 496, + 594, + 890, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Knowledge Distillation", + "text_level": 1, + "bbox": [ + 500, + 724, + 710, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge distillation, also known as student-teacher learning, has become one of the most popular deep learning techniques in various domains [6, 49, 72, 85, 86]. It firstly trains an over-parameterized teacher model and then distills teacher knowledge to a lightweight student. By training the student to mimic the prediction results of the teacher, the student can inherit the knowledge learned by the teacher. Thus, it can achieve better performance than traditional training. The concept of knowledge distillation is first proposed by Bucilua et al. [7] to compress ensemble", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23448", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3fe7b5607e3165fab270d9f037b829b89b04319d209b9fe818443eb1c5596641.jpg", + "image_caption": [ + "Figure 2. The details of Tiny Updater. (1) Tiny Updater re-trains the model with the new data collected from users (named as fully-updated models) by changing all the parameters. (2) Secondly, Tiny Updater finds the channels in the fully-updated models which have similar value to the pre-updating model, and prunes these channels to their value before updating (the obtained models are named as partially-updated models). (3) Then, Tiny Updater re-trains the unpruned weights. During re-training, the fully-updated model is utilized as the teacher model for the partially-updated model with knowledge distillation (KD) loss to improve its performance. (4) Tiny Updater repeats step-2 and step-3 until most channels have been pruned to their value before pruning. (5) Finally, users only require to download very few parameters which are actually changed for updating." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 893, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of neural networks, and then extended by Hinton et al. with a temperature hyper-parameter in the softmax function [32]. Following their success, abundant methods have been proposed to distill teacher knowledge in their backbone features [42, 70], spatial attention [20, 84], task-oriented information [87], pixel-wise relation [20, 44, 49, 86], sample-wise relation [63, 65, 78], prediction residual [43] and so on.", + "bbox": [ + 75, + 332, + 468, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous knowledge distillation methods usually transfer teacher knowledge in the categorical probability distribution. Then, abundant methods have been introduced to distill the knowledge in the teacher feature [70] and its variants, such as attention [84, 86], relational information [63, 78], self-supervised knowledge [81] and task-oriented information [87]. Besides distilling knowledge from a large teacher to a tiny student, there have also been fruitful knowledge distillation methods that distill knowledge from deeper layers to shallow layers [88, 89], from all the channels to partial channels [83], from ensemble classifiers to single classifier [40], from multiple frames to few frames [5], from RGB images to RGB-D images [23]. In this paper, we introduce knowledge distillation by distilling the knowledge from a teacher model, which has all the parameters changed during updating to a student model, which is only updated by changing a few parameters.", + "bbox": [ + 75, + 439, + 470, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Incremental Learning", + "text_level": 1, + "bbox": [ + 76, + 709, + 282, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the observation that human beings can incrementally learn knowledge about new tasks and categories without forgetting the old knowledge, incremental learning (a.k.a. lifelong learning) is proposed to give the neural networks the similar ability. Elastic weight consolidation (EWC) is proposed to maintain the knowledge of old tasks and categories by first estimating the importance of each neuron with Bays estimation or Fisher information matrix and then training them with a consolidation constraint [39, 55, 77]. Then, memory aware synapse [1] is proposed to compute the importance score of neural net", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "work parameters in an unsupervised and online manner, which is also similar to the Hebbian learning in biological systems [31]. Besides these consolidation-based methods, knowledge distillation methods have also been utilized in incremental learning to maintain the old knowledge by learning its response to new tasks [13, 48, 92]. Besides image classification, abundant recent research has applied incremental learning to the other visual tasks, such as object detection [11, 26], action recognition [60] and imaged-to-image translation [74].", + "bbox": [ + 496, + 333, + 890, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both the proposed efficient model updating and incremental learning expect the neural network to maintain its knowledge of old parameters. Their main difference is that: (a) Incremental learning does not limit the change in old parameters strictly, while efficient model updating has a direct constraint on the number of changed parameters to reduce the download overhead for edge devices. (b) Incremental learning usually assumes that the data for old knowledge is not available while efficient model updating can still access data, which is more practical to industrial applications. (c) The target of incremental learning is to build human-like artificial intelligence which can incrementally learn various tasks, while the target of Efficient Model Updating is to reduce the number of parameters updated when new data for the same task is collected.", + "bbox": [ + 496, + 484, + 892, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Efficient Pretrained Model Finetuning", + "text_level": 1, + "bbox": [ + 500, + 724, + 831, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large-scale pretrained models recently have shown powerful representation ability in both natural language processing and computer vision. However, since these pretrained models usually have billions of parameters, directly finetuning them on downstream tasks usually suffers from massive training overhead, making them impractical in real-world applications. Recently, prompts methods have been proposed to address this problem by designing specific input templates for the specific downstream tasks [4,25,45,52, 54,79,91]. Moreover, adapter methods have also been intro-", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23449", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "duced to finetune several additional trainable layers instead of the whole pre-trained model [3, 33]. These methods are firstly proposed for pre-trained language models and then extended to language-vision multi-modal models and vision models [12, 36, 82]. Both these efficient finetuning methods and the proposed efficient model updating aim to freeze the parameters of neural networks. Their main difference is that: (a) Efficient finetuning is usually applied to the models which are firstly pre-trained on large-scale datasets to learn task-unbiased knowledge and then finetuned for specific downstream tasks. Instead, the pre-updating models in Efficient Model Updating are trained with very few training samples (e.g. only $20\\%$ in our experiments) and then updated with the new data collected from users. (b) The target of efficient finetuning is to reduce the training overhead of pre-trained models and improve their performance in downstream tasks. In contrast, the target of Efficient Model Updating is to reduce the download overhead during software updating. (c) Besides, adapter-based methods can be considered as a special case in Tiny Updater where all the parameters of backbone layers are frozen and all the adapter layers are updated. Our experiments also demonstrate that Tiny Updater achieves clearly better performance than directly applying efficient model updating methods. Please refer to Appendix A for a more detailed comparison among the proposed efficient model updating, incremental learning, and efficient pretrained model finetuning.", + "bbox": [ + 76, + 90, + 472, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 517, + 212, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Efficient Model Updating", + "text_level": 1, + "bbox": [ + 76, + 542, + 308, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a training dataset $\\mathcal{D} = \\{(x_1, y_1), \\dots, (x_n, y_n)\\}$ , the software developers firstly train a deep neural network $\\mathcal{F}$ with parameter $\\Theta$ for their software. After deploying $\\mathcal{F}$ on the edge devices, abundant new training samples $\\{(x_{n+1}, y_{n+1}), \\dots, (x_{n+m}, y_{n+m})\\}$ can be collected from users and the dataset can be extended as $\\mathcal{D}^+ = \\mathcal{D} \\cup \\{(x_{n+1}, y_{n+1}), \\dots, (x_{n+m}, y_{n+m})\\}$ . Then, software developers can re-train $\\mathcal{F}$ on $\\mathcal{D}^+$ to obtain better performance, whose parameters can be denoted as $\\Phi$ . Usually, during this re-training step, all the values in $\\theta$ can be totally different from that in $\\Phi$ . Thus, the users have to download all the parameters of $\\Phi$ to update the software, which harms the user experience and limits the frequently updating of neural network-driven software. In such kind of paradigm, we name the model before updating as the pre-updating model, the model after updating as the post-updating model, the model with all the parameters updated as the fully-updated model, and the model with partial parameters updated as the partially updated model. The target of efficient model updating is to obtain a partially updated model which has most of the parameters unchanged while achieving similar performance to the fully-updated model.", + "bbox": [ + 76, + 568, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 The proposed Tiny Updater.", + "text_level": 1, + "bbox": [ + 500, + 90, + 777, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Dataset $\\mathcal{D}^{+} = \\{(x_{1},y_{1}),\\dots,(x_{n + m},y_{n + m})\\}$ , the pre-updating model $\\mathcal{F}$ with parameter $\\Theta$ , an expected ratio of updated parameters $\\tau$ .", + "bbox": [ + 498, + 108, + 890, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: The partially-updated model $\\mathcal{F}$ with parameter $\\Phi$ . // Update the model with $\\mathcal{D}^+$ .", + "bbox": [ + 498, + 154, + 888, + 184 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initialize the partially-updated model $\\mathcal{F}$ with parameter $\\Phi$ . while $\\mathcal{F}_{\\Phi}$ is not converged do:", + "bbox": [ + 500, + 185, + 888, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sample a batch of data $\\mathcal{X}$ from $\\mathcal{D}^+$", + "bbox": [ + 540, + 214, + 772, + 228 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compute $\\hat{y} \\coloneqq \\mathcal{F}_{\\Phi}(\\mathcal{X})$", + "bbox": [ + 540, + 229, + 692, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compute the task loss between $y$ and $\\hat{y}$", + "bbox": [ + 540, + 244, + 802, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Back propagate gradients and update $\\Phi$ .", + "bbox": [ + 540, + 261, + 803, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "//Pruning, and re-training with KD loss.", + "bbox": [ + 500, + 276, + 885, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initialize the parameters of the teacher $\\mathcal{T} \\coloneqq \\Phi$ , an index set of pruned weights as $\\mathcal{I} = \\{\\}$ .", + "bbox": [ + 500, + 290, + 890, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "while $\\frac{Card(\\mathcal{T})}{Card(\\Theta)} < 1 - \\tau$ do:", + "bbox": [ + 500, + 321, + 678, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compute the $L_{1}$ -norm distance between $\\Phi$ and $\\theta$ .", + "bbox": [ + 540, + 339, + 864, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Append the indices of channels which have relatively smaller distance to $\\mathcal{I}$ .", + "bbox": [ + 540, + 354, + 890, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\Phi [\\mathcal{I}]\\coloneqq \\Theta [\\mathcal{I}].$ //Prune $\\Phi$ to $\\Theta$", + "bbox": [ + 540, + 383, + 818, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "while $\\mathcal{F}_{\\Phi}$ is not converged do:", + "bbox": [ + 540, + 400, + 750, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sample a batch of data $\\mathcal{X}$ from $\\mathcal{D}^+$", + "bbox": [ + 580, + 415, + 812, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compute $\\hat{y} \\coloneqq \\mathcal{F}_{\\Phi}(\\mathcal{X})$ and $y_{t} \\coloneqq \\mathcal{F}_{\\mathcal{T}}(\\mathcal{X})$ .", + "bbox": [ + 580, + 429, + 852, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compute the task loss between $y$ and $\\hat{y}$ , and KD loss between $y$ and $y_{t}$ .", + "bbox": [ + 580, + 445, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Back propagate gradients and update the parameters of $\\Phi$ which are not in $\\mathcal{I}$ .", + "bbox": [ + 580, + 474, + 846, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Return The partially-updated model $\\mathcal{F}$ with parameter $\\Phi$ .", + "bbox": [ + 500, + 505, + 880, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Tiny Updater", + "text_level": 1, + "bbox": [ + 500, + 545, + 640, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fruitful previous works in model compression have successfully proven that even a very tiny neural network can have powerful representation ability, which motivates us to propose to learn the knowledge in the collected training data $\\mathcal{D}^{+}$ with only a few parameters instead of all the parameters. The optimization objective of Tiny Updater can be formulated as", + "bbox": [ + 496, + 569, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\arg \\min _ {\\Phi} \\frac {1}{n + m} \\sum_ {i = 1} ^ {n + m} \\mathcal {L} _ {\\text {t a s k}} \\left(x _ {i}, y _ {i}\\right) \\tag {1} \\\\ \\text {s u b j e c t} \\quad \\frac {\\left| \\Theta - \\Phi \\right| _ {0}}{\\operatorname {C a r d} (\\Theta)} < \\tau \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 681, + 890, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{task}}$ indicates the original task-specific loss function, such as cross-entropy loss for image classification. $|\\cdot|_0$ indicates the $L_0$ -norm, which measures the number of non-zero elements in a tensor. $\\operatorname{Card}(\\cdot)$ denotes the cardinality, which describes the number of parameters in a tensor. $\\tau$ is a ratio threshold that determines how many parameters should be changed during updating. It is observed that when $\\theta$ in Equation (1) becomes zero, the optimization objective is similar to another deep learning technique - network pruning.", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23450", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b36fefbdb4b5e217ed12a10095a4d2a9a10435740ee707ed084c273a3e73c389.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 133, + 97, + 313, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b290562004e043f34a63d1a885afa74aadc369dcca042e093c3928fc18272e6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 97, + 498, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9f8ff91c52461d6f7c1ff64c941ffe7ee8d3f2d9d1e43472e0fe3c9e1d36c0af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 97, + 676, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/aa3dd3e6487db150fe7d07f893ea9e74a0e125e65a686fb26510f79ae2bb2cdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 98, + 864, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/897b7e72e728fc446de4342db63d21a94b4f8dcec205374df982c35aa04d41da.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 129, + 220, + 313, + 345 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/36c1f9f04cc5c2bd3d34850338c8a67f6663b499c237529e1bf9a6b06854abd8.jpg", + "image_caption": [ + "Figure 3. Experiments on four fine-grained image classification datasets with ResNet50. The pre-updating and the post-updating models are trained with $25\\%$ and $100\\%$ training data, respectively. In subfigure (a), the newly collected images come from all categories uniformly. In subfigure (b), the newly collected images come from only the categories that are not available before updating." + ], + "image_footnote": [], + "bbox": [ + 330, + 220, + 496, + 345 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c1ed06d9070221be0875fc09117adf5422bbfa3a0602904b55af8b75d6943d44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 220, + 681, + 345 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/daf44520f2e6b57347a837ee249401eefd5dfc9e4a10e89029570eeff551cca2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 220, + 861, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing. Thus, we propose to solve this problem by introducing iterative pruning methods. Besides, knowledge distillation is also introduced to improve the model performance with only a few parameters changed during updating with the teacher from a model with all the parameters changed. Please note that Tiny Updater is a framework that applies pruning and knowledge distillation methods to tackle the challenge of efficient model updating and it does not propose any new pruning and knowledge distillation method.", + "bbox": [ + 75, + 407, + 467, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The details of Tiny Updater are shown in Algorithm 1. With the extended dataset $\\mathcal{D}^+$ , it first trains a fully-updated model $\\mathcal{F}_{\\Phi}$ without any constraints and copies its parameters to another neural network $\\mathcal{F}_{\\mathcal{T}}$ , which is considered as the teacher network. Then, channel pruning is utilized to find the parameters in $\\Phi$ (e.g. channels in convolutional layers) whose value is not significantly changed compared with the pre-updating model $\\Theta$ . Then, these parameters are reset to their value in $\\Theta$ and are not further trained in the later re-training. During the re-training period, the partially-updated model is trained with not only the origin task loss but also the knowledge distillation loss from the teacher of the fully-updated model $\\mathcal{F}_{\\mathcal{T}}$ . Please note that the proposed Tiny Updater is a framework that employs pruning and knowledge distillation to tackle the challenge of efficient model updating and it does not depend on any specific pruning and knowledge distillation methods.", + "bbox": [ + 75, + 545, + 467, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 819, + 200, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Image Classification", + "text_level": 1, + "bbox": [ + 76, + 845, + 267, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiment Setting Our method has mainly been evaluated on seven image classification datasets, including CI-", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3eebe9df91662db65fc75f5342fd761d67879a6bd554412520aeccf187d35fa0.jpg", + "table_caption": [ + "Table 1. The performance of Tiny Updater on ImageNet. Models are firstly trained with $25\\%$ data and then updated with all the data. \"Updated Ratio\" indicates the ratio of updated parameters." + ], + "table_footnote": [], + "table_body": "
ModelUpdated Ratio (%)Top-1 AccuracyTop-5 Accuracy
ResNet5000.0068.86 -6.4488.74 -3.46
21.5474.76 -0.5491.70 -0.50
42.4275.12 -0.1891.93 -0.27
65.6976.25 +0.9593.14 -0.94
100.075.3092.20
MobileNetv200.0062.23 -3.0884.25 -2.29
24.3164.54 -0.7785.53 -1.01
31.6665.23 -0.0886.42 -0.12
66.5566.30 +0.9987.19 +0.65
87.6066.86 +1.5587.50 +0.96
100.065.3186.54
Swin Transformer00.0072.04 -9.2689.75 -5.85
18.5679.91 -1.3994.33 -1.27
36.3180.87 -0.4394.53 -1.07
80.2681.73 +0.4395.74 +0.14
100.081.3095.60
", + "bbox": [ + 501, + 450, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FAR10, CIFAR100, and ImageNet for general image classification, Stanford Car, Oxford Flower, Stanford Dog, FGVC Aircraft for fine-grained image classification. Please refer to Appendix B for the details of these datasets. Note that on the fine-grained image classification datasets, models are initialized with backbone weights pre-trained on ImageNet. In ImageNet experiments, we adopt the basic training policy from PyTorch [64] with task-oriented feature distillation [87]. On the other datasets, each model is trained by 200 epochs with the naive logit and feature knowledge distillation loss [32, 70].", + "bbox": [ + 496, + 696, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Main Results The performance of Tiny Updater on ImageNet, CIFAR10&100, and four fine-grained image classifi-", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23451", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cb65524168a6ad9ded1610690e13f54cc0a722db902d46be70af15da0dc69208.jpg", + "image_caption": [ + "Figure 4. Experiments on CIFAR with ResNet50. The pre-updating models and fully-updated models are trained with $25\\%$ and $100\\%$ training data, respectively. Tiny Updater with random pruning indicates pruning the randomly selected channels." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 454, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cation are shown in Table 1, Figure 4, Figure 3, respectively. It is observed that: (i) Tiny Updater achieves consistent effectiveness on all seven datasets. By changing around $20\\%$ parameters during updating, the neural networks trained with Tiny Updater achieve a similar performance with fully-updated models on all these datasets. (ii) On ImageNet, Tiny Updater achieves significant performance on the regular convolutional network such as ResNet [28], the efficient convolutional networks such as MobileNetv2 [71], and Swin Transformer [57]. (iii) With the proposed Tiny Updater, when a large ratio of parameters is changed, it even leads to higher performance than the fully-updated model. For instance, on Oxford Flowers datasets in Figure 3, the $84\\%$ updated model trained with Tiny Updater has $0.25\\%$ higher accuracy than the fully-updated model. We suggest this accuracy benefit is caused by the knowledge distillation loss in Tiny Updater. As pointed out by some previous research [2, 61], even if the student and the teacher have similar performance, knowledge distillation can still lead to consistent accuracy benefits.", + "bbox": [ + 75, + 285, + 472, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation Study Tiny Updater is mainly composed of two modules - pruning and knowledge distillation. Ablation studies on CIFAR10 and CIFAR100 are shown in Figure 4. It is observed that: (i) Compared with Tiny Updater with random pruning, Tiny Updater with $L_{1}$ -norm based pruning leads to consistently higher accuracy (around $6\\%$ ) especially when only a low ratio ( $0 \\sim 50\\%$ ) of parameters are changed, indicating $L_{1}$ -norm is an effective metric to find which channels are actually important for model updating. (ii) Significant accuracy boosts can be observed by applying knowledge distillation during re-training. For instance, on CIFAR100, when around $10\\%$ and $20\\%$ parameters are changed, $1.79\\%$ and $1.15\\%$ accuracy boost can be observed with knowledge distillation, respectively. These observations indicate that the pruning and knowledge distillation in Tiny Updater have their individual effectiveness.", + "bbox": [ + 75, + 599, + 468, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Categorical Incremental Updating Figure 3(b) shows the effectiveness of Tiny Updater in the categorical incremental settings, where the pre-updating models are trained", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6a2a06b3d9b663ad5b6aa4b406f336cab67aebff51b587193133caea518124dc.jpg", + "image_caption": [ + "Figure 5. Experiments on CIFAR10 with the pre-updating models trained with different ratios of data." + ], + "image_footnote": [], + "bbox": [ + 611, + 90, + 774, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with $25\\%$ training data belonging to $25\\%$ categories. During the updating period, the model is further trained with data of all the categories. It is observed that Tiny Updater can still achieve significant performance by only changing around $10\\%$ parameters, indicating that Tiny Updater is also effective on the categorical incremental updating setting.", + "bbox": [ + 496, + 250, + 890, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Influence from Pre-updating Model In previous experiments, the pre-updating models are trained with $25\\%$ data. With less training data, the pre-updating models tend to have less representative ability, and thus the performance of Tiny Updater tends to be reduced. In this subsection, we study how the performance of the pre-updating model influence Tiny Updater on CIFAR10 in Figure 3(a). It is observed that when most of the parameters $(>50\\%)$ are changed during updating, the accuracy gap between different pre-updating models is not significant $(< 0.2\\%)$ . In contrast, when only a few parameters $(10\\% \\sim 20\\%)$ are changed, the ratio of data for training pre-updating models has a more significant influence. For example, compared with the pre-updating model trained with $25\\%$ data, the pre-updating model trained with only $5\\%$ data leads to around $2.5\\%$ accuracy loss when $20\\%$ parameters are changed during updating, indicating that the performance of the pre-updating model has a direct influence to Tiny Updater. Specifically, when the pre-updating model is not trained with any data, it can be considered as a model with all parameters initialized with zero matrices, and thus Tiny Updater in this case degenerates to the common neural network pruning problem.", + "bbox": [ + 496, + 351, + 892, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-step Model Updating Previous experiments mainly show the result of one-step updating from using $25\\%$ to $100\\%$ training data. In this subsection, we show the performance of Tiny Updater in a multi-step updating, where the model is first trained with $5\\%$ data and then updated with $10\\%$ , $15\\%$ , $20\\%$ , $25\\%$ , and $100\\%$ training data, successively. As shown in Table 4, there is almost no accuracy loss when around $20\\%$ parameters are changed for each update. When there are around $10\\%$ parameters changed for each update, accuracy loss becomes significant with the increment on the update iterations (from $-0.44\\%$ to $-2.33\\%$ ). This observation indicates that the accuracy loss tends to be accumulated in multi-step updating when an extremely low ratio of parameters is changed for updating.", + "bbox": [ + 496, + 693, + 893, + 905 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23452", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a9892a5843b6366d0ca6cde2a9b351e9c72cc6a5f78f27771b4e85b0a3cf4993.jpg", + "table_caption": [ + "Table 2. Experiments of multi-step updating on CIFAR10. Models are firstly trained with $5\\%$ training data and then updated by five times with $10\\%$ , $15\\%$ , $20\\%$ , $25\\%$ , and $100\\%$ training data by changing all, $10\\%$ and $20\\%$ parameters." + ], + "table_footnote": [], + "table_body": "
Data Ratio \nParams. Ratio5%10%15%20%25%100%
Fully-Updated46.564.278.884.687.895.3
10% Params. Updated46.563.878.383.285.992.9
20% Params. Updated46.564.379.084.287.594.8
", + "bbox": [ + 78, + 148, + 468, + 217 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tiny Updater on Pruned Models In this subsection, we study whether the proposed Tiny Updater can be utilized on a pruned model. Concretely, we first train a ResNet18 model with $25\\%$ training data on CIFAR10, which achieves $87.43\\%$ top-1 accuracy. Then, we apply the $L_{1}$ -norm pruning technique to prune $53\\%$ neurons at the expense of $0.37\\%$ accuracy loss $(87.43\\% \\rightarrow 87.06\\%)$ . Thirdly, the proposed Tiny Updater is applied to update this model with $100\\%$ training data, which achieves $94.22\\%$ at the expense of changing $24.3\\%$ parameters $(87.06\\% \\rightarrow 94.22\\%)$ . In contrast, a pruned fully-updated model with the same pruning ratio achieves $94.87\\%$ accuracy, which is only $0.65\\%$ higher than the pruned model with Tiny Updater $(87.06\\% \\rightarrow 94.87\\%)$ . These observations indicate that the proposed Tiny Updater can be utilized on pruned models.", + "bbox": [ + 75, + 244, + 470, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Finetuning and Adapter Finetuning and adapter-based methods are two well-known efficient training methods and the parameters changed in these two methods are also much fewer than global finetuning. Figure 4 gives the comparison between Tiny Updater and these two methods on CIFAR10, CIFAR100 with ResNet18. It is observed that when the same number of parameters are changed, Tiny Updater outperforms these two methods by a clear margin, indicating that Tiny Updater is more effective than directly applying previous methods.", + "bbox": [ + 75, + 486, + 468, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Image-to-Image Translation with GANs", + "text_level": 1, + "bbox": [ + 76, + 648, + 419, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiment Setting We mainly evaluate Tiny Up- dater on image-to-image translation with CycleGAN for Horse $\\rightarrow$ Zebra translation, and Pix2Pix for Edge $\\rightarrow$ Shoe translation. CycleGAN is a typical unpaired image-toimage translation model, which has a ResNet-based generator trained with GAN loss and cycle consistency [94]. Pix2Pix is a typical paired image-to-image translation model, which has a U-Net architecture generator trained with conditional GAN loss and the $L_{1}$ -norm loss [35]. Horse $\\rightarrow$ Zebra is an unpaired dataset that translates natural images of horses to zebras and vice versa. It consists of 1,187 horse images and 1,474 zebra [93]. Edge $\\rightarrow$ Shoe is a paired data that translates the edges of shoes to their corresponding natural images [34]. Fréchet Inception Distance (FID), which measures the distance between the feature dis", + "bbox": [ + 75, + 672, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fda6bd8009aee50fb7cdd3009802e081d435b992fca883c2ee4f7fac4763dbd4.jpg", + "image_caption": [ + "Figure 6. Qualitative analysis on Horse $\\rightarrow$ Zebra and Zebra $\\rightarrow$ Horse with CycleGAN. The partially-updated model trained with Tiny Updater has around $20\\%$ parameters changed. The pre-updating models are trained with $25\\%$ data." + ], + "image_footnote": [], + "bbox": [ + 500, + 87, + 883, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0abb3c0873994993d3a63401fe5271041001d7e4038ecbdbbf9983b9ab5268c.jpg", + "image_caption": [ + "Figure 7. Experiments of image-to-image translation on Horse $\\rightarrow$ Zebra and Zebra $\\rightarrow$ Horse with CycleGAN, and Edge $\\rightarrow$ Shoe with Pix2Pix. In all these experiments, the pre-updating models are trained with $25\\%$ training data." + ], + "image_footnote": [], + "bbox": [ + 504, + 339, + 887, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tribution of the real and the generated images, is utilized as the metric for both datasets. A lower FID indicates the synthetic images have better quality.", + "bbox": [ + 498, + 507, + 890, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Main Results Experiments of Tiny Updater on image-to-image translation are shown in Figure 7. It is observed that on the three image-to-image translation tasks, Tiny Updater is able to improve model performance by $5\\sim 10$ FID when around $60\\%$ parameters are changed during updating. Besides, Tiny Updater can update the models by changing $40\\%$ parameters with no performance loss or by changing only $20\\%$ parameters with a slight FID increment. Qualitative analysis in Figure 6 shows that the pre-updating models can not transform the whole body of horses into the zebras, or remove all the stripes of zebras. In contrast, with Tiny Updater, the partially-updated model with only $20\\%$ parameters changed can address this problem and achieve comparable performance with the fully-updated models, indicating that Tiny Updater enables the neural network to be successfully updated by changing only a few parameters.", + "bbox": [ + 496, + 564, + 892, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Video Recognition", + "text_level": 1, + "bbox": [ + 500, + 816, + 676, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experiment Setting We evaluate Tiny Updater in video recognition datasets including UCF-101 [73] and Diving-48 [46] with video recognition models including SlowOnly [19] and Inception3D [9]. UCF-101 is an ac", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23453", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/358257611ad7f6a95e1513dcf69c6a0b721ec66d24f8dac003ab9d5b85dd59ce.jpg", + "image_caption": [ + "Figure 8. Experimental results of Tiny Updater of video recognition on UCF-101 and Diving-48 with SlowOnly and Inception3D (I3D). The pre-updating models are trained with $25\\%$ training data." + ], + "image_footnote": [], + "bbox": [ + 98, + 98, + 285, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10b2c5672046302d5922cdd9e9519eae47a6b7901e9b1485cdb37980c09cc7e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 98, + 486, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea2c9945c955e96f6d29edd0a67747d2c05253749bd49b9c617434b727415a5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 98, + 694, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cfed865dc1be1ce7088535073fd1272c1bb3045060b205b4306a76d23097890e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 98, + 897, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion recognition dataset with 101 action classes over 13,000 video clips [73]. Diving-48 is a fine-grained video dataset of competitive diving. It has around 18,000 video clips belonging to 48 dive sequences [9]. Both top-1 and top-5 accuracy are reported.", + "bbox": [ + 75, + 289, + 470, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Main Results Figure 8 shows the performance of Tiny Updater on video recognition. It is observed that on both SlowOnly and Inception3D (I3D), UCF-101 and Diving-48, the model with only $20\\%$ parameters changed trained by Tiny Updater achieves comparable and even higher performance than the fully-updated models. Moreover, the model with only $10\\%$ parameters changed during updating leads to only around $0.5\\%$ accuracy loss.", + "bbox": [ + 75, + 375, + 468, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 508, + 191, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Choice of Pruning Granularity", + "text_level": 1, + "bbox": [ + 76, + 534, + 349, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Existing network pruning methods can be mainly divided into regular pruning (e.g. channel-wise pruning) and irregular pruning (e.g. element-wise pruning). Usually, irregular pruning can achieve a higher compression rate, while regular pruning is more friendly to hardware. In this paper, we prefer to use regular pruning in Tiny Updater and suggest that using irregular pruning may cause the problem of index overhead and memory access overhead.", + "bbox": [ + 75, + 556, + 468, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Index Overhead When Tiny Updater is utilized for updating neural network-based software, users should download not only the parameters which are changed during updating but also an index file that records the corresponding position of each changed parameter in the neural network. Given a neural network with $N$ parameters, when irregular pruning is utilized, for each updated parameter, $\\log_2N$ bits storage space is required to record which layer, filter, and position in the kernel this parameter should be. Thus, an ignorable additional downloading overhead is caused. In contrast, the index overhead for channel-wise pruning is $\\log_2\\frac{N}{C\\times K^2}$ bits, which is small enough to be almost ignorable. Here $C$ and $K$ indicate the channel number and the kernel size, respectively.", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Memory Access Overhead In Tiny Updater, irregular pruning can result in the updated parameters being distributed across various neural layers, channels, and filters, leading to their storage in different data blocks of the file system. Consequently, when updating the model with the downloaded parameters, almost all data blocks of the edge device need to undergo a writing operation, resulting in a significant memory access overhead. In contrast, regular pruning ensures that the updated weights are channel-wise and, therefore, can be stored in the same data blocks. Consequently, only these data blocks for the updated parameters require memory access, potentially reducing the updating overhead on edge devices.", + "bbox": [ + 496, + 289, + 893, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 516, + 617, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces the challenge of efficient model updating to reduce the download overhead during neural network-based software updating. To this end, we propose Tiny Updater, which aims to change only a small subset of the parameters during model retraining, thereby reducing the number of parameters that need to be downloaded by users. Tiny Updater comprises an iterative pruning technique that prunes the parameters of the fully-updated models to the pre-updating models and a knowledge distillation technique that treats the fully-updated models as teachers and the partially-updated models as students.", + "bbox": [ + 496, + 547, + 890, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extensive experiments were conducted on eleven datasets covering general image classification, fine-grained image classification, video recognition, paired image-to-image translation, and unpaired image-to-image translation. The results demonstrate that Tiny Updater can update neural network-based software by changing only $10\\%$ to $20\\%$ of the parameters with almost no loss in accuracy. In addition, the experimental results show the effectiveness of Tiny Updater in various scenarios, including categorical updating, multi-step updating, and pruning. We believe that this paper will encourage further research to address the challenge of efficient model updating on edge devices.", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23454", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ff230d1a3bb6ed6ca90137e6553675ff4ba50e7ea48b0cb924b5c6d5fa2e5590.jpg", + "table_caption": [ + "Table 3. Comparison among incremental learning, efficient finetuning, and the proposed efficient model updating." + ], + "table_footnote": [], + "table_body": "
ProblemTargetPre / Post TasksUsage of Training DataConstraints on Parameters
Incremental LearningThe human-like ability of incrementally learning new tasks without forgetting old tasks.Different.A small ratio of data is used for training pre-updating models. More data is used for updating. Data for pre-training data can not be accessed during updating (learning new tasks).No direct constraints. Reducing the number of changed parameters sometimes lead to positive influence.
Efficient FinetuningFinetuning pre-trained models for downstream tasks with low training overhead, and achieving better downstream performance.Different.Pre-updating models are pretrained on large-scale datasets, and then finetuned on much less data for downstream tasks. Usually do not access data for pre-training during finetuning.No direct constraints. Do not finetune the whole model usually leads to no changes in pre-trained weights. Additional parameters in adapter layers and prompts are required.
Efficient Model UpdatingUpdate models with new data by changing only a few parameters to reduce communicat-ion overhead.Same.A small ratio of data is used for pre-training models, and these data can still be accessed during updating.Having direct constraints, the changed parameters should be as few as possible.
", + "bbox": [ + 133, + 101, + 843, + 363 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 398, + 264, + 416 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was partially supported by National Key R&D Program of China (2022YFB2804103), Key Research and Development Program of Shaanxi (2021ZDLGY01-05), Tsinghua University Dushi Program, National Natural Science Foundation of China (20211710187), and Tsinghua University Talent Program.", + "bbox": [ + 75, + 417, + 468, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A. Detailed Comparison", + "text_level": 1, + "bbox": [ + 76, + 523, + 284, + 540 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A detailed comparison among incremental learning, efficient finetuning and the proposed efficient model updating is shown in Table 3.", + "bbox": [ + 75, + 545, + 468, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "B. Detailed Experimental Setting", + "text_level": 1, + "bbox": [ + 76, + 599, + 357, + 618 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CIFAR10/100 are two datasets for low resolution general image classification with 50,000 images and 1,000 images in the training and validating set, respectively. ImageNet is a large-scale dataset for general image classification with images belonging to 1,000 categories. Stanford Car is a dataset for fine-grained image classification with car images belonging to 197 classes. Oxford Flower is a dataset for fine-grained image classification which consists of 102 different categories of flowers common to the UK. Stanford Dog is a dataset for fine-grained image classification which contains 20,580 dog images of 120 breeds of dogs from around the world. FGVC Aircraft is a dataset for fine-grained image classifier on which contains 10,000 images of aircrafts spanning 100 aircraft models.", + "bbox": [ + 75, + 622, + 470, + 834 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C. Experiments on More Datasets and Tasks", + "text_level": 1, + "bbox": [ + 76, + 844, + 452, + 862 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Experimental results on more kinds of tasks and datasets are shown in Table 4 and Table 5. It is observed that", + "bbox": [ + 76, + 869, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/8ce8861994936bad622594c321b40d825ca9459c0a36485c0ac3cf48070a5366.jpg", + "table_caption": [ + "Table 4. Comparison between our method with previous efficient finetuning methods on ImageNet with Swin Transformer. Models are firstly trained with $25\\%$ data and then updated with all the data." + ], + "table_footnote": [], + "table_body": "
Updating MethodRatio of Updated Params.(%)Top-1 Acc (%)
Pre-updating Model0.0%72.04
Global Finetuning100.0%81.30
Linear Finetuning1.1672.23
Non-linear Finetuning234.675.53
Non-linear Finetuning366.378.80
Adapter13.572.40
Adapter217.274.39
Visual Prompt Tuning≤ 1%72.32
Ours136.3180.87
Ours218.5679.91
", + "bbox": [ + 501, + 450, + 890, + 621 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/d3e2102386fcef66ba4bfb3e11f0e326cecfc9f24ff4a63bc6543c67a34c04bd.jpg", + "table_caption": [ + "Table 5. Experimental results on more tasks." + ], + "table_footnote": [], + "table_body": "
TaskDatasetModelUpdated Params (%)Acc. / mAP / PSNR
Cross Domain Image ClassificationOffice-31ResNet500.0092.3
18.798.3
10.998.0
100.098.5
Object DetectionMS COCOFaster RCNN0.0030.5
27.336.2
14.635.6
100.036.4
Single Image Super-ResolutionDIV2KESRGAN0.0033.90
22.434.60
13.034.56
100.034.61
", + "bbox": [ + 501, + 664, + 890, + 797 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "our method leads to consistent good performance on cross-domain image classification, object detection and image super-resolution. Besides, our method outperforms previous efficient finetuning methods on ImageNet.", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23455", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 3", + "[2] Zeyuan Allen-Zhu and Yuanzhi Li. Towards understanding ensemble, knowledge distillation and self-distillation in deep learning. CoRR, abs/2012.09816, 2020. 6", + "[3] Ankur Bapna, Naveen Arivazhagan, and Orhan First. Simple, scalable adaptation for neural machine translation. arXiv preprint arXiv:1909.08478, 2019. 4", + "[4] Eyal Ben-David, Nadav Oved, and Roi Reichart. Pada: A prompt-based autoregressive approach for adaptation to unseen domains. arXiv preprint arXiv:2102.12206, 2021. 3", + "[5] Shweta Bhardwaj, Mukundhan Srinivasan, and Mitesh M Khapra. Efficient video classification using fewer frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 354-363, 2019. 3", + "[6] Cunling Bian, Wei Feng, Liang Wan, and Song Wang. Structural knowledge distillation for efficient skeleton-based action recognition. IEEE Transactions on Image Processing, 30:2963-2976, 2021. 2", + "[7] Cristian Bucilua, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 535-541. ACM, 2006. 2", + "[8] Efe Camci, Manas Gupta, Min Wu, and Jie Lin. Qlp: Deep q-learning for pruning deep neural networks. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 2", + "[9] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7, 8", + "[10] Hyun Sung Chang, Sanghoon Sull, and Sang Uk Lee. Efficient video indexing scheme for content-based retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 9(8):1269-1279, 1999. 2", + "[11] Li Chen, Chunyan Yu, and Lvcai Chen. A new knowledge distillation for incremental object detection. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2019. 3", + "[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 4", + "[13] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2534-2543, 2021. 3", + "[14] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 248-255, 2009. 1", + "[17] Guiguang Ding, Shuo Zhang, Zizhou Jia, Jing Zhong, and Jungong Han. Where to prune: Using lstm to guide data-dependent soft pruning. IEEE Transactions on Image Processing, 30:293-304, 2021. 2", + "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1", + "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 7", + "[20] Yingchao Feng, Xian Sun, Wenhui Dao, Jihao Li, and Xin Gao. Double similarity distillation for semantic image segmentation. IEEE Transactions on Image Processing, 30:5363-5376, 2021. 3", + "[21] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635, 2018. 2", + "[22] Sharath Girish, Shishira R Maiya, Kamal Gupta, Hao Chen, Larry S Davis, and Abhinav Shrivastava. The lottery ticket hypothesis for object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 762-771, 2021. 2", + "[23] Saurabh Gupta, Judy Hoffman, and Jitendra Malik. Cross modal distillation for supervision transfer. In CVPR, pages 2827-2836, 2016. 3", + "[24] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In International Conference on Learning Representations, 2016. 2", + "[25] Xu Han, Weilin Zhao, Ning Ding, Zhiyuan Liu, and Maosong Sun. *Ptr: Prompt tuning with rules for text classification.* arXiv preprint arXiv:2105.11259, 2021. 3", + "[26] Yu Hao, Yanwei Fu, Yu-Gang Jiang, and Qi Tian. An end-to-end architecture for class-incremental object detection with knowledge distillation. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2019. 3", + "[27] Babak Hassibi and David G Stork. Second order derivatives for network pruning: Optimal brain surgeon. In Advances in neural information processing systems, pages 164-171, 1993. 2", + "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23456", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6", + "[29] Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang. Soft filter pruning for accelerating deep convolutional neural networks. arXiv preprint arXiv:1808.06866, 2018. 2", + "[30] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 1389-1397, 2017. 2", + "[31] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology Press, 2005. 3", + "[32] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. In NeurIPS, 2014. 3, 5", + "[33] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 4", + "[34] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Pix2pix datasets, http://efrosgans.eecs.berkeley.edu/pix2pix/datasets/.7", + "[35] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 7", + "[36] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 4", + "[37] Qing Jin, Jian Ren, Oliver J Woodford, Jiazhuo Wang, Geng Yuan, Yanzhi Wang, and Sergey Tulyakov. Teachers do more than teach: Compressing image-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13600-13611, 2021. 2", + "[38] Hyeong-Ju Kang. Accelerator-aware pruning for convolutional neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 30(7):2093-2103, 2020. 2", + "[39] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3", + "[40] Xu Lan, Xiatian Zhu, and Shaogang Gong. Knowledge distillation by on-the-fly native ensemble. arXiv preprint arXiv:1806.04606, 2018. 3", + "[41] Yann LeCun, John S Denker, and Sara A Solla. Optimal brain damage. In Advances in neural information processing systems, pages 598-605, 1990. 2", + "[42] Muyang Li, Ji Lin, Yaoyao Ding, Zhijian Liu, Jun-Yan Zhu, and Song Han. Gan compression: Efficient architectures for interactive conditional gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5284-5294, 2020. 2, 3", + "[43] Xuewei Li, Songyuan Li, Bourahla Omar, Fei Wu, and Xi Li. Reskd: Residual-guided knowledge distillation. IEEE Transactions on Image Processing, 30:4735-4746, 2021. 3" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Xiaojie Li, Jianlong Wu, Hongyu Fang, Yue Liao, Fei Wang, and Chen Qian. Local correlation consistency for knowledge distillation. In European Conference on Computer Vision, pages 18-33. Springer, 2020. 3", + "[45] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3", + "[46] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 7", + "[47] Yun Li, Zechun Liu, Weiqun Wu, Haotian Yao, Xiangyu Zhang, Chi Zhang, and Baoqun Yin. Weight-dependent gates for network pruning. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 2", + "[48] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 3", + "[49] Zeqi Li, Ruowei Jiang, and Parham Aarabi. Semantic relation preserving knowledge distillation for image-to-image translation. In European Conference on Computer Vision, pages 648-663. Springer, 2020. 2, 3", + "[50] Tsung-Yi Lin, Priyal Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE transactions on pattern analysis and machine intelligence, 2018. 1", + "[51] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1", + "[52] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint arXiv:2107.13586, 2021. 1, 3", + "[53] Weide Liu, Guosheng Lin, Tianyi Zhang, and Zichuan Liu. Guided co-segmentation network for fast video object segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 31(4):1607-1617, 2021. 1", + "[54] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602, 2021. 3", + "[55] Xialei Liu, Marc Masana, Luis Herranz, Joost Van de Weijer, Antonio M Lopez, and Andrew D Bagdanov. Rotate your networks: Better weight consolidation and less catastrophic forgetting. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 2262-2268. IEEE, 2018. 3", + "[56] Yuchen Liu, Zhixin Shu, Yijun Li, Zhe Lin, Federico Peruzzi, and Sun-Yuan Kung. Content-aware gan compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12156-12166, 2021. 2", + "[57] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23457", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6", + "[58] Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting Cheng, and Jian Sun. Metapruning: Meta learning for automatic neural network channel pruning. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 2", + "[59] Zechun Liu, Xiangyu Zhang, Zhiqiang Shen, Yichen Wei, Kwang-Ting Cheng, and Jian Sun. Joint multi-dimension pruning via numerical gradient update. IEEE Transactions on Image Processing, 30:8034-8045, 2021. 2", + "[60] Rashid Minhas, Abdul Adeel Mohammed, and Q. M. Jonathan Wu. Incremental learning in human action recognition based on snippets. IEEE Transactions on Circuits and Systems for Video Technology, 22(11):1529-1541, 2012. 3", + "[61] Hossein Mobahi, Mehrdad Farajtabar, and Peter L Bartlett. Self-distillation amplifies regularization in hilbert space. arXiv preprint arXiv:2002.05715, 2020. 6", + "[62] Diego Ortego, Juan C. Sanmiguel, and José M. Martínez. Hierarchical improvement of foreground segmentation masks in background subtraction. IEEE Transactions on Circuits and Systems for Video Technology, 29(6):1645-1658, 2019. 1", + "[63] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 3", + "[64] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5", + "[65] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In Proceedings of the IEEE International Conference on Computer Vision, pages 5007-5016, 2019. 3", + "[66] Zheng Qin, Zeming Li, Zhaoning Zhang, Yiping Bao, Gang Yu, Yuxing Peng, and Jian Sun. Thundernet: Towards real-time generic object detection on mobile devices. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 2", + "[67] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1", + "[68] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91-99, 2015. 1", + "[69] Youngmin Ro and Jin Young Choi. Layer-wise pruning and auto-tuning of layer-wise learning rates in fine-tuning of deep networks. arXiv preprint arXiv:2002.06048, 2020. 2" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In International Conference on Learning Representations, 2015. 3, 5", + "[71] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 6", + "[72] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 2", + "[73] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7, 8", + "[74] Daniel Stanley Tan, Yong-Xiang Lin, and Kai-Lung Hua. Incremental learning of multi-domain image-to-image translations. IEEE Transactions on Circuits and Systems for Video Technology, 31(4):1526-1539, 2021. 3", + "[75] Mingxing Tan and Quoc V Le. Efficientnet: Rethinking model scaling for convolutional neural networks. arXiv preprint arXiv:1905.11946, 2019. 1", + "[76] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020. 1", + "[77] Brian Thompson, Jeremy Gwinnup, Huda Khayrallah, Kevin Duh, and Philipp Koehn. Overcoming catastrophic forgetting during domain adaptation of neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2062–2068, 2019. 3", + "[78] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1365-1374, 2019. 3", + "[79] Ruize Wang, Duyu Tang, Nan Duan, Zhongyu Wei, Xuanjing Huang, Guihong Cao, Daxin Jiang, Ming Zhou, et al. K-adapter: Infusing knowledge into pre-trained models with adapters. arXiv preprint arXiv:2002.01808, 2020. 1, 3", + "[80] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5987-5995, 2017. 1", + "[81] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In European Conference on Computer Vision, pages 588-604. Springer, 2020. 3", + "[82] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021.4", + "[83] Jiahui Yu and Thomas S. Huang. Universally slimmable networks and improved training techniques. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "23458", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[84] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In International Conference on Learning Representations, 2017. 3", + "[85] Linfeng Zhang, Xin Chen, Xiaobing Tu, Pengfei Wan, Ning Xu, and Kaiheng Ma. Wavelet knowledge distillation: Towards efficient image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12464-12474, June 2022. 2", + "[86] Linfeng Zhang and Ma Kaisheng. Improve object detection with feature-based knowledge distillation: Towards accurate and efficient detectors. In International Conference on Learning Representations, 2021. 2, 3", + "[87] Linfeng Zhang, Yukang Shi, Zuoqiang Shi, Kaisheng Ma, and Chenglong Bao. Task-oriented feature distillation. In NeurIPS, 2020. 3, 5", + "[88] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In arXiv preprint: 1905.08094, 2019. 3", + "[89] Linfeng Zhang, Zhanhong Tan, Jiebo Song, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Scan: A scalable neural networks framework towards compact and efficient models. ArXiv, abs/1906.03951, 2019. 3", + "[90] Mingxin Zhao, Junbo Peng, Shuangming Yu, Liyuan Liu, and Nanjian Wu. Exploring structural sparsity in cnn via selective penalty. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1658-1666, 2022. 2", + "[91] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. arXiv preprint arXiv:2203.05557, 2022. 3", + "[92] Peng Zhou, Long Mai, Jianming Zhang, Ning Xu, Zuxuan Wu, and Larry S Davis. M2kd: Multi-model and multilevel knowledge distillation for incremental learning. arXiv preprint arXiv:1904.01769, 2019. 3", + "[93] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Pix2pix datasets, http://efrosgans.eecs.berkeley.edu/cyclegan/datasets.7", + "[94] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 7" + ], + "bbox": [ + 78, + 90, + 468, + 712 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "23459", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/23a6116f-5b95-4dcd-97e5-95528a6d1fff_model.json b/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/23a6116f-5b95-4dcd-97e5-95528a6d1fff_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3dc848982c00f9714ed34189de1a885e16b7bdb6 --- /dev/null +++ b/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/23a6116f-5b95-4dcd-97e5-95528a6d1fff_model.json @@ -0,0 +1,2877 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.266, + 0.131, + 0.705, + 0.177 + ], + "angle": 0, + "content": "Tiny Updater: Towards Efficient Neural Network-Driven Software Updating" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.204, + 0.61, + 0.222 + ], + "angle": 0, + "content": "Linfeng Zhang, Kaisheng Ma*" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.222, + 0.772, + 0.24 + ], + "angle": 0, + "content": "Institute for Interdisciplinary Information Sciences, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.242, + 0.747, + 0.256 + ], + "angle": 0, + "content": "zhanglinfeng1997@outlook.com, kaisheng@mail.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.474, + 0.596 + ], + "angle": 0, + "content": "Significant advancements have been accomplished with deep neural networks in diverse visual tasks, which have substantially elevated their deployment in edge device software. However, during the update of neural network-based software, users are required to download all the parameters of the neural network anew, which harms the user experience. Motivated by previous progress in model compression, we propose a novel training methodology named Tiny Updater to address this issue. Specifically, by adopting the variant of pruning and knowledge distillation methods, Tiny Updater can update the neural network-based software by only downloading a few parameters \\((10\\% \\sim 20\\%)\\) instead of all the parameters in the neural network. Experiments on eleven datasets of three tasks, including image classification, image-to-image translation, and video recognition have demonstrated its effectiveness. Codes have been released in https://github.com/ArchipLab-LinfengZhang/TinyUpdater." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.625, + 0.21, + 0.641 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.786 + ], + "angle": 0, + "content": "With the availability of large-scale datasets [15, 16, 51] and high-performance computing platforms, deep neural networks have achieved remarkable achievements in various visual tasks such as image classification [18, 28, 75, 80], segmentation [53, 62], and object detection [50, 67, 68, 76]. Encouraged by their impressive performance, numerous software developers have effectively integrated neural networks into their software products and deployed them on edge devices such as mobile phones and tablet computers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.471, + 0.879 + ], + "angle": 0, + "content": "Typically, the development roadmap for a neural network-based software follows the paradigm illustrate in Figure 1(a). Initially, users install the software by downloading all the neural network parameters from a cloud platform. Subsequently, as the software interacts with its users, an abundance of new training data and requirements can be" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.294, + 0.885, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.42, + 0.887, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.563, + 0.895, + 0.634 + ], + "angle": 0, + "content": "Figure 1. Comparison between (a) the traditional model updating scheme and (b) the proposed efficient model updating. In efficient model updating, only around \\(10\\%\\) parameters in the neural network are actually changed and thus users only require to download these changed parameters for updating." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.854 + ], + "angle": 0, + "content": "collected. Then, software developers may retrain the neural network with the gathered data and update their software to enhance its performance. However, during the retraining phase, all the neural network parameters are typically changed compared to their values before updating. Consequently, users are compelled to download all the parameters of the neural network again from the cloud platform, seriously impairing their experience. While recent research has explored text prompts and adapter layers to fine-tune a large-scale pre-trained model at a low training cost [52, 79], there have been no prior efforts to reduce the download cost during model updating, which has a more direct impact on users with edge devices." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "This paper proposes the challenge of efficient model updating with the objective of reducing the download overhead of neural network-based software during updating. As" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "23447" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.075, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "depicted in Figure 1(b), during the retraining phase, efficient model updating introduces an additional constraint limiting the change of only a small subset (e.g. \\(10\\%\\)) of the neural network parameters compared to the pre-updating model. Consequently, users are only required to download a few parameters that have actually changed instead of all the parameters in the neural network. In general, efficient model updating raises two questions: how to find the optimal parameters that should be changed in the neural network, and how to achieve comparable accuracy with the fully-updated model (i.e., the model which has all the parameters changed during updating)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.274, + 0.473, + 0.622 + ], + "angle": 0, + "content": "To tackle this challenge, we propose a novel neural network training framework named Tiny Updater. Motivated by previous works in model compression, Tiny Updater is composed of the variants of two typical model compression techniques - neural network pruning and knowledge distillation. Firstly, to determine which channels or layers in neural networks should be modified during updating, Tiny Updater applies the pruning technique. As shown in Figure 2, it iteratively calculates the \\( L_{1} \\)-norm distance between the pre-updating model and the post-updating model parameters. The channels with smaller distances are deemed unnecessary for updating and are pruned to their original values before updating. Conversely, the channels with larger distances are considered essential for updating, and thus they are actually changed. Secondly, during the retraining period, we propose to improve the performance of the partially-updated model by distilling the knowledge from a fully-updated teacher model. The partially-updated student model is trained to give predictions that are similar to those of the fully-updated teacher model by optimizing the knowledge distillation loss. This ensures that the partially-updated model achieves comparable performance with the fully-updated model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.471, + 0.745 + ], + "angle": 0, + "content": "Extensive experiments have been conducted on eleven datasets and three different tasks, including, image classification, video classification, and image-to-image translation, using seven different neural networks. Experimental results demonstrate that Tiny Updater can update the model by changing only \\(10\\%\\) to \\(20\\%\\) of the parameters with minimal performance degradation. In summary, the main contribution of this paper can be outlined as follows." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.763, + 0.47, + 0.825 + ], + "angle": 0, + "content": "- To the best of our knowledge, we first propose the challenge of efficient model updating to reduce the communication cost of downloading neural network parameters to the edge devices for software updating." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- To tackle this challenge, we propose Tiny Updater, which prunes the fully-updated model to the pre-updating model and distills the knowledge from the fully-updated model to the partially-updated model." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.763, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "- Extensive experiments have been conducted on three tasks with eleven datasets, including image classification, image-to-image translation and video recognition. Experimental results demonstrate that Tiny Updater can reduce the training overhead by \\(80\\% \\sim 90\\%\\) with almost no performance degradation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.198, + 0.642, + 0.214 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.223, + 0.729, + 0.24 + ], + "angle": 0, + "content": "2.1. Neural Network Pruning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.247, + 0.892, + 0.594 + ], + "angle": 0, + "content": "Neural network pruning is one of the most effective methods for deep neural network compression by deleting the unimportant neurons [42, 47], filters [29, 38], channels [30, 90] and layers [8, 69]. LeCun et al. and Hassibi et al. first propose to prune neural networks depending on the Hessian matrix [27, 41]. Recently, Han et al. propose deep compression, which finds the important connection in neural networks with its absolute value and then iteratively prunes them [24]. Liu et al. train a meta-network to generate the weights of neural networks and then apply it to search the best-pruned architecture [58]. Ding et al. propose to apply a LSTM to learn the hierarchical characteristics of deep neural networks and then generates the corresponding pruning scheme [17]. Liu et al. propose joint multi-dimension pruning to prune the channels, layers and resolutions at the same time [59]. Frankle et al. propose the lottery ticket hypothesis and find that a standard pruning technique naturally uncovers subnetworks whose initializations made them capable of training effectively [21]. Besides classification, neural network pruning has also be utilized in image-to-image translation [37, 42], unconditional image generation [56], object detection [22, 66], video retrieval [10] and pre-trained language models [14]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.716 + ], + "angle": 0, + "content": "In this paper, Tiny Updater introduces neural network pruning methods to find the parameters which are actually valuable for model updating. Its main difference compared with previous pruning methods is that Tiny Updater prunes the parameters to their value before updating instead of zero, which makes it easier to be solved than traditional pruning. Our experiments also show that Tiny Updater can be utilized on the pruned models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.712, + 0.743 + ], + "angle": 0, + "content": "2.2. Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Knowledge distillation, also known as student-teacher learning, has become one of the most popular deep learning techniques in various domains [6, 49, 72, 85, 86]. It firstly trains an over-parameterized teacher model and then distills teacher knowledge to a lightweight student. By training the student to mimic the prediction results of the teacher, the student can inherit the knowledge learned by the teacher. Thus, it can achieve better performance than traditional training. The concept of knowledge distillation is first proposed by Bucilua et al. [7] to compress ensemble" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "23448" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.895, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.218, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 2. The details of Tiny Updater. (1) Tiny Updater re-trains the model with the new data collected from users (named as fully-updated models) by changing all the parameters. (2) Secondly, Tiny Updater finds the channels in the fully-updated models which have similar value to the pre-updating model, and prunes these channels to their value before updating (the obtained models are named as partially-updated models). (3) Then, Tiny Updater re-trains the unpruned weights. During re-training, the fully-updated model is utilized as the teacher model for the partially-updated model with knowledge distillation (KD) loss to improve its performance. (4) Tiny Updater repeats step-2 and step-3 until most channels have been pruned to their value before pruning. (5) Finally, users only require to download very few parameters which are actually changed for updating." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.47, + 0.44 + ], + "angle": 0, + "content": "of neural networks, and then extended by Hinton et al. with a temperature hyper-parameter in the softmax function [32]. Following their success, abundant methods have been proposed to distill teacher knowledge in their backbone features [42, 70], spatial attention [20, 84], task-oriented information [87], pixel-wise relation [20, 44, 49, 86], sample-wise relation [63, 65, 78], prediction residual [43] and so on." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.471, + 0.698 + ], + "angle": 0, + "content": "Previous knowledge distillation methods usually transfer teacher knowledge in the categorical probability distribution. Then, abundant methods have been introduced to distill the knowledge in the teacher feature [70] and its variants, such as attention [84, 86], relational information [63, 78], self-supervised knowledge [81] and task-oriented information [87]. Besides distilling knowledge from a large teacher to a tiny student, there have also been fruitful knowledge distillation methods that distill knowledge from deeper layers to shallow layers [88, 89], from all the channels to partial channels [83], from ensemble classifiers to single classifier [40], from multiple frames to few frames [5], from RGB images to RGB-D images [23]. In this paper, we introduce knowledge distillation by distilling the knowledge from a teacher model, which has all the parameters changed during updating to a student model, which is only updated by changing a few parameters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.283, + 0.727 + ], + "angle": 0, + "content": "2.3. Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Inspired by the observation that human beings can incrementally learn knowledge about new tasks and categories without forgetting the old knowledge, incremental learning (a.k.a. lifelong learning) is proposed to give the neural networks the similar ability. Elastic weight consolidation (EWC) is proposed to maintain the knowledge of old tasks and categories by first estimating the importance of each neuron with Bays estimation or Fisher information matrix and then training them with a consolidation constraint [39, 55, 77]. Then, memory aware synapse [1] is proposed to compute the importance score of neural net" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.485 + ], + "angle": 0, + "content": "work parameters in an unsupervised and online manner, which is also similar to the Hebbian learning in biological systems [31]. Besides these consolidation-based methods, knowledge distillation methods have also been utilized in incremental learning to maintain the old knowledge by learning its response to new tasks [13, 48, 92]. Besides image classification, abundant recent research has applied incremental learning to the other visual tasks, such as object detection [11, 26], action recognition [60] and imaged-to-image translation [74]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.893, + 0.713 + ], + "angle": 0, + "content": "Both the proposed efficient model updating and incremental learning expect the neural network to maintain its knowledge of old parameters. Their main difference is that: (a) Incremental learning does not limit the change in old parameters strictly, while efficient model updating has a direct constraint on the number of changed parameters to reduce the download overhead for edge devices. (b) Incremental learning usually assumes that the data for old knowledge is not available while efficient model updating can still access data, which is more practical to industrial applications. (c) The target of incremental learning is to build human-like artificial intelligence which can incrementally learn various tasks, while the target of Efficient Model Updating is to reduce the number of parameters updated when new data for the same task is collected." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.833, + 0.743 + ], + "angle": 0, + "content": "2.4. Efficient Pretrained Model Finetuning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Large-scale pretrained models recently have shown powerful representation ability in both natural language processing and computer vision. However, since these pretrained models usually have billions of parameters, directly finetuning them on downstream tasks usually suffers from massive training overhead, making them impractical in real-world applications. Recently, prompts methods have been proposed to address this problem by designing specific input templates for the specific downstream tasks [4,25,45,52, 54,79,91]. Moreover, adapter methods have also been intro-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "23449" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.501 + ], + "angle": 0, + "content": "duced to finetune several additional trainable layers instead of the whole pre-trained model [3, 33]. These methods are firstly proposed for pre-trained language models and then extended to language-vision multi-modal models and vision models [12, 36, 82]. Both these efficient finetuning methods and the proposed efficient model updating aim to freeze the parameters of neural networks. Their main difference is that: (a) Efficient finetuning is usually applied to the models which are firstly pre-trained on large-scale datasets to learn task-unbiased knowledge and then finetuned for specific downstream tasks. Instead, the pre-updating models in Efficient Model Updating are trained with very few training samples (e.g. only \\(20\\%\\) in our experiments) and then updated with the new data collected from users. (b) The target of efficient finetuning is to reduce the training overhead of pre-trained models and improve their performance in downstream tasks. In contrast, the target of Efficient Model Updating is to reduce the download overhead during software updating. (c) Besides, adapter-based methods can be considered as a special case in Tiny Updater where all the parameters of backbone layers are frozen and all the adapter layers are updated. Our experiments also demonstrate that Tiny Updater achieves clearly better performance than directly applying efficient model updating methods. Please refer to Appendix A for a more detailed comparison among the proposed efficient model updating, incremental learning, and efficient pretrained model finetuning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.518, + 0.214, + 0.535 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.544, + 0.31, + 0.56 + ], + "angle": 0, + "content": "3.1. Efficient Model Updating" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.569, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Given a training dataset \\(\\mathcal{D} = \\{(x_1, y_1), \\dots, (x_n, y_n)\\}\\), the software developers firstly train a deep neural network \\(\\mathcal{F}\\) with parameter \\(\\Theta\\) for their software. After deploying \\(\\mathcal{F}\\) on the edge devices, abundant new training samples \\(\\{(x_{n+1}, y_{n+1}), \\dots, (x_{n+m}, y_{n+m})\\}\\) can be collected from users and the dataset can be extended as \\(\\mathcal{D}^+ = \\mathcal{D} \\cup \\{(x_{n+1}, y_{n+1}), \\dots, (x_{n+m}, y_{n+m})\\}\\). Then, software developers can re-train \\(\\mathcal{F}\\) on \\(\\mathcal{D}^+\\) to obtain better performance, whose parameters can be denoted as \\(\\Phi\\). Usually, during this re-training step, all the values in \\(\\theta\\) can be totally different from that in \\(\\Phi\\). Thus, the users have to download all the parameters of \\(\\Phi\\) to update the software, which harms the user experience and limits the frequently updating of neural network-driven software. In such kind of paradigm, we name the model before updating as the pre-updating model, the model after updating as the post-updating model, the model with all the parameters updated as the fully-updated model, and the model with partial parameters updated as the partially updated model. The target of efficient model updating is to obtain a partially updated model which has most of the parameters unchanged while achieving similar performance to the fully-updated model." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.092, + 0.778, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 The proposed Tiny Updater." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.109, + 0.892, + 0.154 + ], + "angle": 0, + "content": "Input: Dataset \\(\\mathcal{D}^{+} = \\{(x_{1},y_{1}),\\dots,(x_{n + m},y_{n + m})\\}\\), the pre-updating model \\(\\mathcal{F}\\) with parameter \\(\\Theta\\), an expected ratio of updated parameters \\(\\tau\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.155, + 0.89, + 0.185 + ], + "angle": 0, + "content": "Output: The partially-updated model \\(\\mathcal{F}\\) with parameter \\(\\Phi\\). // Update the model with \\(\\mathcal{D}^+\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.186, + 0.89, + 0.215 + ], + "angle": 0, + "content": "Initialize the partially-updated model \\(\\mathcal{F}\\) with parameter \\(\\Phi\\). while \\(\\mathcal{F}_{\\Phi}\\) is not converged do:" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.215, + 0.773, + 0.229 + ], + "angle": 0, + "content": "Sample a batch of data \\(\\mathcal{X}\\) from \\(\\mathcal{D}^+\\)" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.23, + 0.694, + 0.246 + ], + "angle": 0, + "content": "Compute \\(\\hat{y} \\coloneqq \\mathcal{F}_{\\Phi}(\\mathcal{X})\\)" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.246, + 0.803, + 0.261 + ], + "angle": 0, + "content": "Compute the task loss between \\(y\\) and \\(\\hat{y}\\)" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.262, + 0.804, + 0.276 + ], + "angle": 0, + "content": "Back propagate gradients and update \\(\\Phi\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.277, + 0.887, + 0.29 + ], + "angle": 0, + "content": "//Pruning, and re-training with KD loss." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.291, + 0.891, + 0.321 + ], + "angle": 0, + "content": "Initialize the parameters of the teacher \\(\\mathcal{T} \\coloneqq \\Phi\\), an index set of pruned weights as \\(\\mathcal{I} = \\{\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.322, + 0.679, + 0.34 + ], + "angle": 0, + "content": "while \\(\\frac{Card(\\mathcal{T})}{Card(\\Theta)} < 1 - \\tau\\) do:" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.34, + 0.866, + 0.354 + ], + "angle": 0, + "content": "Compute the \\(L_{1}\\)-norm distance between \\(\\Phi\\) and \\(\\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.355, + 0.892, + 0.383 + ], + "angle": 0, + "content": "Append the indices of channels which have relatively smaller distance to \\(\\mathcal{I}\\)." + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.385, + 0.819, + 0.4 + ], + "angle": 0, + "content": "\\(\\Phi [\\mathcal{I}]\\coloneqq \\Theta [\\mathcal{I}].\\) //Prune \\(\\Phi\\) to \\(\\Theta\\)" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.401, + 0.751, + 0.415 + ], + "angle": 0, + "content": "while \\(\\mathcal{F}_{\\Phi}\\) is not converged do:" + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.416, + 0.813, + 0.429 + ], + "angle": 0, + "content": "Sample a batch of data \\(\\mathcal{X}\\) from \\(\\mathcal{D}^+\\)" + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.43, + 0.854, + 0.445 + ], + "angle": 0, + "content": "Compute \\(\\hat{y} \\coloneqq \\mathcal{F}_{\\Phi}(\\mathcal{X})\\) and \\(y_{t} \\coloneqq \\mathcal{F}_{\\mathcal{T}}(\\mathcal{X})\\)." + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.446, + 0.891, + 0.475 + ], + "angle": 0, + "content": "Compute the task loss between \\(y\\) and \\(\\hat{y}\\), and KD loss between \\(y\\) and \\(y_{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.476, + 0.848, + 0.505 + ], + "angle": 0, + "content": "Back propagate gradients and update the parameters of \\(\\Phi\\) which are not in \\(\\mathcal{I}\\)." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.506, + 0.882, + 0.521 + ], + "angle": 0, + "content": "Return The partially-updated model \\(\\mathcal{F}\\) with parameter \\(\\Phi\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.546, + 0.641, + 0.563 + ], + "angle": 0, + "content": "3.2. Tiny Updater" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Fruitful previous works in model compression have successfully proven that even a very tiny neural network can have powerful representation ability, which motivates us to propose to learn the knowledge in the collected training data \\(\\mathcal{D}^{+}\\) with only a few parameters instead of all the parameters. The optimization objective of Tiny Updater can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.683, + 0.892, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\arg \\min _ {\\Phi} \\frac {1}{n + m} \\sum_ {i = 1} ^ {n + m} \\mathcal {L} _ {\\text {t a s k}} \\left(x _ {i}, y _ {i}\\right) \\tag {1} \\\\ \\text {s u b j e c t} \\quad \\frac {\\left| \\Theta - \\Phi \\right| _ {0}}{\\operatorname {C a r d} (\\Theta)} < \\tau \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{task}}\\) indicates the original task-specific loss function, such as cross-entropy loss for image classification. \\(|\\cdot|_0\\) indicates the \\(L_0\\)-norm, which measures the number of non-zero elements in a tensor. \\(\\operatorname{Card}(\\cdot)\\) denotes the cardinality, which describes the number of parameters in a tensor. \\(\\tau\\) is a ratio threshold that determines how many parameters should be changed during updating. It is observed that when \\(\\theta\\) in Equation (1) becomes zero, the optimization objective is similar to another deep learning technique - network pruning." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "23450" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.147, + 0.128, + 0.16 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.098, + 0.315, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.098, + 0.499, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.098, + 0.677, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.099, + 0.865, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.269, + 0.128, + 0.283 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.13, + 0.221, + 0.314, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.221, + 0.498, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.221, + 0.682, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.221, + 0.862, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.353, + 0.892, + 0.396 + ], + "angle": 0, + "content": "Figure 3. Experiments on four fine-grained image classification datasets with ResNet50. The pre-updating and the post-updating models are trained with \\(25\\%\\) and \\(100\\%\\) training data, respectively. In subfigure (a), the newly collected images come from all categories uniformly. In subfigure (b), the newly collected images come from only the categories that are not available before updating." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.468, + 0.546 + ], + "angle": 0, + "content": "ing. Thus, we propose to solve this problem by introducing iterative pruning methods. Besides, knowledge distillation is also introduced to improve the model performance with only a few parameters changed during updating with the teacher from a model with all the parameters changed. Please note that Tiny Updater is a framework that applies pruning and knowledge distillation methods to tackle the challenge of efficient model updating and it does not propose any new pruning and knowledge distillation method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.468, + 0.803 + ], + "angle": 0, + "content": "The details of Tiny Updater are shown in Algorithm 1. With the extended dataset \\(\\mathcal{D}^+\\), it first trains a fully-updated model \\(\\mathcal{F}_{\\Phi}\\) without any constraints and copies its parameters to another neural network \\(\\mathcal{F}_{\\mathcal{T}}\\), which is considered as the teacher network. Then, channel pruning is utilized to find the parameters in \\(\\Phi\\) (e.g. channels in convolutional layers) whose value is not significantly changed compared with the pre-updating model \\(\\Theta\\). Then, these parameters are reset to their value in \\(\\Theta\\) and are not further trained in the later re-training. During the re-training period, the partially-updated model is trained with not only the origin task loss but also the knowledge distillation loss from the teacher of the fully-updated model \\(\\mathcal{F}_{\\mathcal{T}}\\). Please note that the proposed Tiny Updater is a framework that employs pruning and knowledge distillation to tackle the challenge of efficient model updating and it does not depend on any specific pruning and knowledge distillation methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.82, + 0.202, + 0.837 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.268, + 0.862 + ], + "angle": 0, + "content": "4.1. Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Experiment Setting Our method has mainly been evaluated on seven image classification datasets, including CI-" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.405, + 0.892, + 0.447 + ], + "angle": 0, + "content": "Table 1. The performance of Tiny Updater on ImageNet. Models are firstly trained with \\(25\\%\\) data and then updated with all the data. \"Updated Ratio\" indicates the ratio of updated parameters." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.451, + 0.892, + 0.673 + ], + "angle": 0, + "content": "
ModelUpdated Ratio (%)Top-1 AccuracyTop-5 Accuracy
ResNet5000.0068.86 -6.4488.74 -3.46
21.5474.76 -0.5491.70 -0.50
42.4275.12 -0.1891.93 -0.27
65.6976.25 +0.9593.14 -0.94
100.075.3092.20
MobileNetv200.0062.23 -3.0884.25 -2.29
24.3164.54 -0.7785.53 -1.01
31.6665.23 -0.0886.42 -0.12
66.5566.30 +0.9987.19 +0.65
87.6066.86 +1.5587.50 +0.96
100.065.3186.54
Swin Transformer00.0072.04 -9.2689.75 -5.85
18.5679.91 -1.3994.33 -1.27
36.3180.87 -0.4394.53 -1.07
80.2681.73 +0.4395.74 +0.14
100.081.3095.60
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.892, + 0.865 + ], + "angle": 0, + "content": "FAR10, CIFAR100, and ImageNet for general image classification, Stanford Car, Oxford Flower, Stanford Dog, FGVC Aircraft for fine-grained image classification. Please refer to Appendix B for the details of these datasets. Note that on the fine-grained image classification datasets, models are initialized with backbone weights pre-trained on ImageNet. In ImageNet experiments, we adopt the basic training policy from PyTorch [64] with task-oriented feature distillation [87]. On the other datasets, each model is trained by 200 epochs with the naive logit and feature knowledge distillation loss [32, 70]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Main Results The performance of Tiny Updater on ImageNet, CIFAR10&100, and four fine-grained image classifi-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "23451" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.455, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.205, + 0.47, + 0.262 + ], + "angle": 0, + "content": "Figure 4. Experiments on CIFAR with ResNet50. The pre-updating models and fully-updated models are trained with \\(25\\%\\) and \\(100\\%\\) training data, respectively. Tiny Updater with random pruning indicates pruning the randomly selected channels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.473, + 0.59 + ], + "angle": 0, + "content": "cation are shown in Table 1, Figure 4, Figure 3, respectively. It is observed that: (i) Tiny Updater achieves consistent effectiveness on all seven datasets. By changing around \\(20\\%\\) parameters during updating, the neural networks trained with Tiny Updater achieve a similar performance with fully-updated models on all these datasets. (ii) On ImageNet, Tiny Updater achieves significant performance on the regular convolutional network such as ResNet [28], the efficient convolutional networks such as MobileNetv2 [71], and Swin Transformer [57]. (iii) With the proposed Tiny Updater, when a large ratio of parameters is changed, it even leads to higher performance than the fully-updated model. For instance, on Oxford Flowers datasets in Figure 3, the \\(84\\%\\) updated model trained with Tiny Updater has \\(0.25\\%\\) higher accuracy than the fully-updated model. We suggest this accuracy benefit is caused by the knowledge distillation loss in Tiny Updater. As pointed out by some previous research [2, 61], even if the student and the teacher have similar performance, knowledge distillation can still lead to consistent accuracy benefits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.47, + 0.844 + ], + "angle": 0, + "content": "Ablation Study Tiny Updater is mainly composed of two modules - pruning and knowledge distillation. Ablation studies on CIFAR10 and CIFAR100 are shown in Figure 4. It is observed that: (i) Compared with Tiny Updater with random pruning, Tiny Updater with \\( L_{1} \\)-norm based pruning leads to consistently higher accuracy (around \\( 6\\% \\)) especially when only a low ratio (\\( 0 \\sim 50\\% \\)) of parameters are changed, indicating \\( L_{1} \\)-norm is an effective metric to find which channels are actually important for model updating. (ii) Significant accuracy boosts can be observed by applying knowledge distillation during re-training. For instance, on CIFAR100, when around \\( 10\\% \\) and \\( 20\\% \\) parameters are changed, \\( 1.79\\% \\) and \\( 1.15\\% \\) accuracy boost can be observed with knowledge distillation, respectively. These observations indicate that the pruning and knowledge distillation in Tiny Updater have their individual effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Categorical Incremental Updating Figure 3(b) shows the effectiveness of Tiny Updater in the categorical incremental settings, where the pre-updating models are trained" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.092, + 0.776, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.2, + 0.892, + 0.227 + ], + "angle": 0, + "content": "Figure 5. Experiments on CIFAR10 with the pre-updating models trained with different ratios of data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.342 + ], + "angle": 0, + "content": "with \\(25\\%\\) training data belonging to \\(25\\%\\) categories. During the updating period, the model is further trained with data of all the categories. It is observed that Tiny Updater can still achieve significant performance by only changing around \\(10\\%\\) parameters, indicating that Tiny Updater is also effective on the categorical incremental updating setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.893, + 0.685 + ], + "angle": 0, + "content": "Influence from Pre-updating Model In previous experiments, the pre-updating models are trained with \\(25\\%\\) data. With less training data, the pre-updating models tend to have less representative ability, and thus the performance of Tiny Updater tends to be reduced. In this subsection, we study how the performance of the pre-updating model influence Tiny Updater on CIFAR10 in Figure 3(a). It is observed that when most of the parameters \\((>50\\%)\\) are changed during updating, the accuracy gap between different pre-updating models is not significant \\((< 0.2\\%)\\). In contrast, when only a few parameters \\((10\\% \\sim 20\\%)\\) are changed, the ratio of data for training pre-updating models has a more significant influence. For example, compared with the pre-updating model trained with \\(25\\%\\) data, the pre-updating model trained with only \\(5\\%\\) data leads to around \\(2.5\\%\\) accuracy loss when \\(20\\%\\) parameters are changed during updating, indicating that the performance of the pre-updating model has a direct influence to Tiny Updater. Specifically, when the pre-updating model is not trained with any data, it can be considered as a model with all parameters initialized with zero matrices, and thus Tiny Updater in this case degenerates to the common neural network pruning problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.895, + 0.906 + ], + "angle": 0, + "content": "Multi-step Model Updating Previous experiments mainly show the result of one-step updating from using \\(25\\%\\) to \\(100\\%\\) training data. In this subsection, we show the performance of Tiny Updater in a multi-step updating, where the model is first trained with \\(5\\%\\) data and then updated with \\(10\\%\\), \\(15\\%\\), \\(20\\%\\), \\(25\\%\\), and \\(100\\%\\) training data, successively. As shown in Table 4, there is almost no accuracy loss when around \\(20\\%\\) parameters are changed for each update. When there are around \\(10\\%\\) parameters changed for each update, accuracy loss becomes significant with the increment on the update iterations (from \\(-0.44\\%\\) to \\(-2.33\\%\\)). This observation indicates that the accuracy loss tends to be accumulated in multi-step updating when an extremely low ratio of parameters is changed for updating." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "23452" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.47, + 0.143 + ], + "angle": 0, + "content": "Table 2. Experiments of multi-step updating on CIFAR10. Models are firstly trained with \\(5\\%\\) training data and then updated by five times with \\(10\\%\\), \\(15\\%\\), \\(20\\%\\), \\(25\\%\\), and \\(100\\%\\) training data by changing all, \\(10\\%\\) and \\(20\\%\\) parameters." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.15, + 0.47, + 0.218 + ], + "angle": 0, + "content": "
Data Ratio \nParams. Ratio5%10%15%20%25%100%
Fully-Updated46.564.278.884.687.895.3
10% Params. Updated46.563.878.383.285.992.9
20% Params. Updated46.564.379.084.287.594.8
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.471, + 0.471 + ], + "angle": 0, + "content": "Tiny Updater on Pruned Models In this subsection, we study whether the proposed Tiny Updater can be utilized on a pruned model. Concretely, we first train a ResNet18 model with \\(25\\%\\) training data on CIFAR10, which achieves \\(87.43\\%\\) top-1 accuracy. Then, we apply the \\(L_{1}\\)-norm pruning technique to prune \\(53\\%\\) neurons at the expense of \\(0.37\\%\\) accuracy loss \\((87.43\\% \\rightarrow 87.06\\%)\\). Thirdly, the proposed Tiny Updater is applied to update this model with \\(100\\%\\) training data, which achieves \\(94.22\\%\\) at the expense of changing \\(24.3\\%\\) parameters \\((87.06\\% \\rightarrow 94.22\\%)\\). In contrast, a pruned fully-updated model with the same pruning ratio achieves \\(94.87\\%\\) accuracy, which is only \\(0.65\\%\\) higher than the pruned model with Tiny Updater \\((87.06\\% \\rightarrow 94.87\\%)\\). These observations indicate that the proposed Tiny Updater can be utilized on pruned models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.639 + ], + "angle": 0, + "content": "Comparison with Finetuning and Adapter Finetuning and adapter-based methods are two well-known efficient training methods and the parameters changed in these two methods are also much fewer than global finetuning. Figure 4 gives the comparison between Tiny Updater and these two methods on CIFAR10, CIFAR100 with ResNet18. It is observed that when the same number of parameters are changed, Tiny Updater outperforms these two methods by a clear margin, indicating that Tiny Updater is more effective than directly applying previous methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.42, + 0.667 + ], + "angle": 0, + "content": "4.2. Image-to-Image Translation with GANs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Experiment Setting We mainly evaluate Tiny Up- dater on image-to-image translation with CycleGAN for Horse \\(\\rightarrow\\) Zebra translation, and Pix2Pix for Edge \\(\\rightarrow\\) Shoe translation. CycleGAN is a typical unpaired image-toimage translation model, which has a ResNet-based generator trained with GAN loss and cycle consistency [94]. Pix2Pix is a typical paired image-to-image translation model, which has a U-Net architecture generator trained with conditional GAN loss and the \\(L_{1}\\) -norm loss [35]. Horse \\(\\rightarrow\\) Zebra is an unpaired dataset that translates natural images of horses to zebras and vice versa. It consists of 1,187 horse images and 1,474 zebra [93]. Edge \\(\\rightarrow\\) Shoe is a paired data that translates the edges of shoes to their corresponding natural images [34]. Fréchet Inception Distance (FID), which measures the distance between the feature dis" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.088, + 0.885, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.274, + 0.892, + 0.329 + ], + "angle": 0, + "content": "Figure 6. Qualitative analysis on Horse \\(\\rightarrow\\) Zebra and Zebra \\(\\rightarrow\\) Horse with CycleGAN. The partially-updated model trained with Tiny Updater has around \\(20\\%\\) parameters changed. The pre-updating models are trained with \\(25\\%\\) data." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.34, + 0.888, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.426, + 0.892, + 0.483 + ], + "angle": 0, + "content": "Figure 7. Experiments of image-to-image translation on Horse \\(\\rightarrow\\) Zebra and Zebra \\(\\rightarrow\\) Horse with CycleGAN, and Edge \\(\\rightarrow\\) Shoe with Pix2Pix. In all these experiments, the pre-updating models are trained with \\(25\\%\\) training data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.892, + 0.553 + ], + "angle": 0, + "content": "tribution of the real and the generated images, is utilized as the metric for both datasets. A lower FID indicates the synthetic images have better quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.893, + 0.807 + ], + "angle": 0, + "content": "Main Results Experiments of Tiny Updater on image-to-image translation are shown in Figure 7. It is observed that on the three image-to-image translation tasks, Tiny Updater is able to improve model performance by \\(5\\sim 10\\) FID when around \\(60\\%\\) parameters are changed during updating. Besides, Tiny Updater can update the models by changing \\(40\\%\\) parameters with no performance loss or by changing only \\(20\\%\\) parameters with a slight FID increment. Qualitative analysis in Figure 6 shows that the pre-updating models can not transform the whole body of horses into the zebras, or remove all the stripes of zebras. In contrast, with Tiny Updater, the partially-updated model with only \\(20\\%\\) parameters changed can address this problem and achieve comparable performance with the fully-updated models, indicating that Tiny Updater enables the neural network to be successfully updated by changing only a few parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.678, + 0.833 + ], + "angle": 0, + "content": "4.3. Video Recognition" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Experiment Setting We evaluate Tiny Updater in video recognition datasets including UCF-101 [73] and Diving-48 [46] with video recognition models including SlowOnly [19] and Inception3D [9]. UCF-101 is an ac" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "23453" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.099, + 0.287, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.099, + 0.488, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.099, + 0.695, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.099, + 0.898, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.237, + 0.892, + 0.266 + ], + "angle": 0, + "content": "Figure 8. Experimental results of Tiny Updater of video recognition on UCF-101 and Diving-48 with SlowOnly and Inception3D (I3D). The pre-updating models are trained with \\(25\\%\\) training data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.471, + 0.368 + ], + "angle": 0, + "content": "tion recognition dataset with 101 action classes over 13,000 video clips [73]. Diving-48 is a fine-grained video dataset of competitive diving. It has around 18,000 video clips belonging to 48 dive sequences [9]. Both top-1 and top-5 accuracy are reported." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.47, + 0.496 + ], + "angle": 0, + "content": "Main Results Figure 8 shows the performance of Tiny Updater on video recognition. It is observed that on both SlowOnly and Inception3D (I3D), UCF-101 and Diving-48, the model with only \\(20\\%\\) parameters changed trained by Tiny Updater achieves comparable and even higher performance than the fully-updated models. Moreover, the model with only \\(10\\%\\) parameters changed during updating leads to only around \\(0.5\\%\\) accuracy loss." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.51, + 0.192, + 0.525 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.351, + 0.551 + ], + "angle": 0, + "content": "5.1. Choice of Pruning Granularity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.47, + 0.679 + ], + "angle": 0, + "content": "Existing network pruning methods can be mainly divided into regular pruning (e.g. channel-wise pruning) and irregular pruning (e.g. element-wise pruning). Usually, irregular pruning can achieve a higher compression rate, while regular pruning is more friendly to hardware. In this paper, we prefer to use regular pruning in Tiny Updater and suggest that using irregular pruning may cause the problem of index overhead and memory access overhead." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Index Overhead When Tiny Updater is utilized for updating neural network-based software, users should download not only the parameters which are changed during updating but also an index file that records the corresponding position of each changed parameter in the neural network. Given a neural network with \\( N \\) parameters, when irregular pruning is utilized, for each updated parameter, \\( \\log_2N \\) bits storage space is required to record which layer, filter, and position in the kernel this parameter should be. Thus, an ignorable additional downloading overhead is caused. In contrast, the index overhead for channel-wise pruning is \\( \\log_2\\frac{N}{C\\times K^2} \\) bits, which is small enough to be almost ignorable. Here \\( C \\) and \\( K \\) indicate the channel number and the kernel size, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.895, + 0.488 + ], + "angle": 0, + "content": "Memory Access Overhead In Tiny Updater, irregular pruning can result in the updated parameters being distributed across various neural layers, channels, and filters, leading to their storage in different data blocks of the file system. Consequently, when updating the model with the downloaded parameters, almost all data blocks of the edge device need to undergo a writing operation, resulting in a significant memory access overhead. In contrast, regular pruning ensures that the updated weights are channel-wise and, therefore, can be stored in the same data blocks. Consequently, only these data blocks for the updated parameters require memory access, potentially reducing the updating overhead on edge devices." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.517, + 0.619, + 0.533 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.715 + ], + "angle": 0, + "content": "This paper introduces the challenge of efficient model updating to reduce the download overhead during neural network-based software updating. To this end, we propose Tiny Updater, which aims to change only a small subset of the parameters during model retraining, thereby reducing the number of parameters that need to be downloaded by users. Tiny Updater comprises an iterative pruning technique that prunes the parameters of the fully-updated models to the pre-updating models and a knowledge distillation technique that treats the fully-updated models as teachers and the partially-updated models as students." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Extensive experiments were conducted on eleven datasets covering general image classification, fine-grained image classification, video recognition, paired image-to-image translation, and unpaired image-to-image translation. The results demonstrate that Tiny Updater can update neural network-based software by changing only \\(10\\%\\) to \\(20\\%\\) of the parameters with almost no loss in accuracy. In addition, the experimental results show the effectiveness of Tiny Updater in various scenarios, including categorical updating, multi-step updating, and pruning. We believe that this paper will encourage further research to address the challenge of efficient model updating on edge devices." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "23454" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.089, + 0.819, + 0.102 + ], + "angle": 0, + "content": "Table 3. Comparison among incremental learning, efficient finetuning, and the proposed efficient model updating." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.102, + 0.844, + 0.364 + ], + "angle": 0, + "content": "
ProblemTargetPre / Post TasksUsage of Training DataConstraints on Parameters
Incremental LearningThe human-like ability of incrementally learning new tasks without forgetting old tasks.Different.A small ratio of data is used for training pre-updating models. More data is used for updating. Data for pre-training data can not be accessed during updating (learning new tasks).No direct constraints. Reducing the number of changed parameters sometimes lead to positive influence.
Efficient FinetuningFinetuning pre-trained models for downstream tasks with low training overhead, and achieving better downstream performance.Different.Pre-updating models are pretrained on large-scale datasets, and then finetuned on much less data for downstream tasks. Usually do not access data for pre-training during finetuning.No direct constraints. Do not finetune the whole model usually leads to no changes in pre-trained weights. Additional parameters in adapter layers and prompts are required.
Efficient Model UpdatingUpdate models with new data by changing only a few parameters to reduce communicat-ion overhead.Same.A small ratio of data is used for pre-training models, and these data can still be accessed during updating.Having direct constraints, the changed parameters should be as few as possible.
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.4, + 0.265, + 0.417 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.51 + ], + "angle": 0, + "content": "This research was partially supported by National Key R&D Program of China (2022YFB2804103), Key Research and Development Program of Shaanxi (2021ZDLGY01-05), Tsinghua University Dushi Program, National Natural Science Foundation of China (20211710187), and Tsinghua University Talent Program." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.524, + 0.285, + 0.541 + ], + "angle": 0, + "content": "A. Detailed Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.59 + ], + "angle": 0, + "content": "A detailed comparison among incremental learning, efficient finetuning and the proposed efficient model updating is shown in Table 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.358, + 0.619 + ], + "angle": 0, + "content": "B. Detailed Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.471, + 0.835 + ], + "angle": 0, + "content": "CIFAR10/100 are two datasets for low resolution general image classification with 50,000 images and 1,000 images in the training and validating set, respectively. ImageNet is a large-scale dataset for general image classification with images belonging to 1,000 categories. Stanford Car is a dataset for fine-grained image classification with car images belonging to 197 classes. Oxford Flower is a dataset for fine-grained image classification which consists of 102 different categories of flowers common to the UK. Stanford Dog is a dataset for fine-grained image classification which contains 20,580 dog images of 120 breeds of dogs from around the world. FGVC Aircraft is a dataset for fine-grained image classifier on which contains 10,000 images of aircrafts spanning 100 aircraft models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.454, + 0.863 + ], + "angle": 0, + "content": "C. Experiments on More Datasets and Tasks" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Experimental results on more kinds of tasks and datasets are shown in Table 4 and Table 5. It is observed that" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.398, + 0.892, + 0.44 + ], + "angle": 0, + "content": "Table 4. Comparison between our method with previous efficient finetuning methods on ImageNet with Swin Transformer. Models are firstly trained with \\(25\\%\\) data and then updated with all the data." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.452, + 0.892, + 0.622 + ], + "angle": 0, + "content": "
Updating MethodRatio of Updated Params.(%)Top-1 Acc (%)
Pre-updating Model0.0%72.04
Global Finetuning100.0%81.30
Linear Finetuning1.1672.23
Non-linear Finetuning234.675.53
Non-linear Finetuning366.378.80
Adapter13.572.40
Adapter217.274.39
Visual Prompt Tuning≤ 1%72.32
Ours136.3180.87
Ours218.5679.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.564, + 0.651, + 0.828, + 0.664 + ], + "angle": 0, + "content": "Table 5. Experimental results on more tasks." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.665, + 0.892, + 0.799 + ], + "angle": 0, + "content": "
TaskDatasetModelUpdated Params (%)Acc. / mAP / PSNR
Cross Domain Image ClassificationOffice-31ResNet500.0092.3
18.798.3
10.998.0
100.098.5
Object DetectionMS COCOFaster RCNN0.0030.5
27.336.2
14.635.6
100.036.4
Single Image Super-ResolutionDIV2KESRGAN0.0033.90
22.434.60
13.034.56
100.034.61
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "our method leads to consistent good performance on cross-domain image classification, object detection and image super-resolution. Besides, our method outperforms previous efficient finetuning methods on ImageNet." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "23455" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Zeyuan Allen-Zhu and Yuanzhi Li. Towards understanding ensemble, knowledge distillation and self-distillation in deep learning. CoRR, abs/2012.09816, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[3] Ankur Bapna, Naveen Arivazhagan, and Orhan First. Simple, scalable adaptation for neural machine translation. arXiv preprint arXiv:1909.08478, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.471, + 0.316 + ], + "angle": 0, + "content": "[4] Eyal Ben-David, Nadav Oved, and Roi Reichart. Pada: A prompt-based autoregressive approach for adaptation to unseen domains. arXiv preprint arXiv:2102.12206, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.471, + 0.373 + ], + "angle": 0, + "content": "[5] Shweta Bhardwaj, Mukundhan Srinivasan, and Mitesh M Khapra. Efficient video classification using fewer frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 354-363, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.471, + 0.429 + ], + "angle": 0, + "content": "[6] Cunling Bian, Wei Feng, Liang Wan, and Song Wang. Structural knowledge distillation for efficient skeleton-based action recognition. IEEE Transactions on Image Processing, 30:2963-2976, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.471, + 0.487 + ], + "angle": 0, + "content": "[7] Cristian Bucilua, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 535-541. ACM, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[8] Efe Camci, Manas Gupta, Min Wu, and Jie Lin. Qlp: Deep q-learning for pruning deep neural networks. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.471, + 0.602 + ], + "angle": 0, + "content": "[9] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.47, + 0.658 + ], + "angle": 0, + "content": "[10] Hyun Sung Chang, Sanghoon Sull, and Sang Uk Lee. Efficient video indexing scheme for content-based retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 9(8):1269-1279, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[11] Li Chen, Chunyan Yu, and Lvcai Chen. A new knowledge distillation for incremental object detection. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[13] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2534-2543, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[14] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 248-255, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[17] Guiguang Ding, Shuo Zhang, Zizhou Jia, Jing Zhong, and Jungong Han. Where to prune: Using lstm to guide data-dependent soft pruning. IEEE Transactions on Image Processing, 30:293-304, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[20] Yingchao Feng, Xian Sun, Wenhui Dao, Jihao Li, and Xin Gao. Double similarity distillation for semantic image segmentation. IEEE Transactions on Image Processing, 30:5363-5376, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[21] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[22] Sharath Girish, Shishira R Maiya, Kamal Gupta, Hao Chen, Larry S Davis, and Abhinav Shrivastava. The lottery ticket hypothesis for object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 762-771, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[23] Saurabh Gupta, Judy Hoffman, and Jitendra Malik. Cross modal distillation for supervision transfer. In CVPR, pages 2827-2836, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[24] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In International Conference on Learning Representations, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[25] Xu Han, Weilin Zhao, Ning Ding, Zhiyuan Liu, and Maosong Sun. *Ptr: Prompt tuning with rules for text classification.* arXiv preprint arXiv:2105.11259, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[26] Yu Hao, Yanwei Fu, Yu-Gang Jiang, and Qi Tian. An end-to-end architecture for class-incremental object detection with knowledge distillation. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[27] Babak Hassibi and David G Stork. Second order derivatives for network pruning: Optimal brain surgeon. In Advances in neural information processing systems, pages 164-171, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "23456" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.162 + ], + "angle": 0, + "content": "[29] Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang. Soft filter pruning for accelerating deep convolutional neural networks. arXiv preprint arXiv:1808.06866, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[30] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 1389-1397, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[31] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology Press, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.248, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[32] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. In NeurIPS, 2014. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.275, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[33] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[34] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Pix2pix datasets, http://efrosgans.eecs.berkeley.edu/pix2pix/datasets/.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[35] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[36] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[37] Qing Jin, Jian Ren, Oliver J Woodford, Jiazhuo Wang, Geng Yuan, Yanzhi Wang, and Sergey Tulyakov. Teachers do more than teach: Compressing image-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13600-13611, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.622 + ], + "angle": 0, + "content": "[38] Hyeong-Ju Kang. Accelerator-aware pruning for convolutional neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 30(7):2093-2103, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[39] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[40] Xu Lan, Xiatian Zhu, and Shaogang Gong. Knowledge distillation by on-the-fly native ensemble. arXiv preprint arXiv:1806.04606, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[41] Yann LeCun, John S Denker, and Sara A Solla. Optimal brain damage. In Advances in neural information processing systems, pages 598-605, 1990. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[42] Muyang Li, Ji Lin, Yaoyao Ding, Zhijian Liu, Jun-Yan Zhu, and Song Han. Gan compression: Efficient architectures for interactive conditional gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5284-5294, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[43] Xuewei Li, Songyuan Li, Bourahla Omar, Fei Wu, and Xi Li. Reskd: Residual-guided knowledge distillation. IEEE Transactions on Image Processing, 30:4735-4746, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.149 + ], + "angle": 0, + "content": "[44] Xiaojie Li, Jianlong Wu, Hongyu Fang, Yue Liao, Fei Wang, and Chen Qian. Local correlation consistency for knowledge distillation. In European Conference on Computer Vision, pages 18-33. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[45] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[46] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[47] Yun Li, Zechun Liu, Weiqun Wu, Haotian Yao, Xiangyu Zhang, Chi Zhang, and Baoqun Yin. Weight-dependent gates for network pruning. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[48] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[49] Zeqi Li, Ruowei Jiang, and Parham Aarabi. Semantic relation preserving knowledge distillation for image-to-image translation. In European Conference on Computer Vision, pages 648-663. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[50] Tsung-Yi Lin, Priyal Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE transactions on pattern analysis and machine intelligence, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[51] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[52] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint arXiv:2107.13586, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[53] Weide Liu, Guosheng Lin, Tianyi Zhang, and Zichuan Liu. Guided co-segmentation network for fast video object segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 31(4):1607-1617, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[54] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[55] Xialei Liu, Marc Masana, Luis Herranz, Joost Van de Weijer, Antonio M Lopez, and Andrew D Bagdanov. Rotate your networks: Better weight consolidation and less catastrophic forgetting. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 2262-2268. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[56] Yuchen Liu, Zhixin Shu, Yijun Li, Zhe Lin, Federico Peruzzi, and Sun-Yuan Kung. Content-aware gan compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12156-12166, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.858, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[57] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "23457" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[58] Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting Cheng, and Jian Sun. Metapruning: Meta learning for automatic neural network channel pruning. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[59] Zechun Liu, Xiangyu Zhang, Zhiqiang Shen, Yichen Wei, Kwang-Ting Cheng, and Jian Sun. Joint multi-dimension pruning via numerical gradient update. IEEE Transactions on Image Processing, 30:8034-8045, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[60] Rashid Minhas, Abdul Adeel Mohammed, and Q. M. Jonathan Wu. Incremental learning in human action recognition based on snippets. IEEE Transactions on Circuits and Systems for Video Technology, 22(11):1529-1541, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.468, + 0.362 + ], + "angle": 0, + "content": "[61] Hossein Mobahi, Mehrdad Farajtabar, and Peter L Bartlett. Self-distillation amplifies regularization in hilbert space. arXiv preprint arXiv:2002.05715, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[62] Diego Ortego, Juan C. Sanmiguel, and José M. Martínez. Hierarchical improvement of foreground segmentation masks in background subtraction. IEEE Transactions on Circuits and Systems for Video Technology, 29(6):1645-1658, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.469, + 0.49 + ], + "angle": 0, + "content": "[63] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.492, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[64] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[65] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In Proceedings of the IEEE International Conference on Computer Vision, pages 5007-5016, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[66] Zheng Qin, Zeming Li, Zhaoning Zhang, Yiping Bao, Gang Yu, Yuxing Peng, and Jian Sun. Thundernet: Towards real-time generic object detection on mobile devices. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[67] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[68] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91-99, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[69] Youngmin Ro and Jin Young Choi. Layer-wise pruning and auto-tuning of layer-wise learning rates in fine-tuning of deep networks. arXiv preprint arXiv:2002.06048, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[70] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In International Conference on Learning Representations, 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[71] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[72] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[73] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[74] Daniel Stanley Tan, Yong-Xiang Lin, and Kai-Lung Hua. Incremental learning of multi-domain image-to-image translations. IEEE Transactions on Circuits and Systems for Video Technology, 31(4):1526-1539, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[75] Mingxing Tan and Quoc V Le. Efficientnet: Rethinking model scaling for convolutional neural networks. arXiv preprint arXiv:1905.11946, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[76] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[77] Brian Thompson, Jeremy Gwinnup, Huda Khayrallah, Kevin Duh, and Philipp Koehn. Overcoming catastrophic forgetting during domain adaptation of neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2062–2068, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[78] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1365-1374, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[79] Ruize Wang, Duyu Tang, Nan Duan, Zhongyu Wei, Xuanjing Huang, Guihong Cao, Daxin Jiang, Ming Zhou, et al. K-adapter: Infusing knowledge into pre-trained models with adapters. arXiv preprint arXiv:2002.01808, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[80] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5987-5995, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[81] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In European Conference on Computer Vision, pages 588-604. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[82] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[83] Jiahui Yu and Thomas S. Huang. Universally slimmable networks and improved training techniques. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "23458" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[84] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In International Conference on Learning Representations, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[85] Linfeng Zhang, Xin Chen, Xiaobing Tu, Pengfei Wan, Ning Xu, and Kaiheng Ma. Wavelet knowledge distillation: Towards efficient image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12464-12474, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[86] Linfeng Zhang and Ma Kaisheng. Improve object detection with feature-based knowledge distillation: Towards accurate and efficient detectors. In International Conference on Learning Representations, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.331 + ], + "angle": 0, + "content": "[87] Linfeng Zhang, Yukang Shi, Zuoqiang Shi, Kaisheng Ma, and Chenglong Bao. Task-oriented feature distillation. In NeurIPS, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[88] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In arXiv preprint: 1905.08094, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.47, + 0.444 + ], + "angle": 0, + "content": "[89] Linfeng Zhang, Zhanhong Tan, Jiebo Song, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Scan: A scalable neural networks framework towards compact and efficient models. ArXiv, abs/1906.03951, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.47, + 0.502 + ], + "angle": 0, + "content": "[90] Mingxin Zhao, Junbo Peng, Shuangming Yu, Liyuan Liu, and Nanjian Wu. Exploring structural sparsity in cnn via selective penalty. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1658-1666, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.503, + 0.47, + 0.544 + ], + "angle": 0, + "content": "[91] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. arXiv preprint arXiv:2203.05557, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.47, + 0.601 + ], + "angle": 0, + "content": "[92] Peng Zhou, Long Mai, Jianming Zhang, Ning Xu, Zuxuan Wu, and Larry S Davis. M2kd: Multi-model and multilevel knowledge distillation for incremental learning. arXiv preprint arXiv:1904.01769, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.47, + 0.644 + ], + "angle": 0, + "content": "[93] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Pix2pix datasets, http://efrosgans.eecs.berkeley.edu/cyclegan/datasets.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[94] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "23459" + } + ] +] \ No newline at end of file diff --git a/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/23a6116f-5b95-4dcd-97e5-95528a6d1fff_origin.pdf b/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/23a6116f-5b95-4dcd-97e5-95528a6d1fff_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ccba8d0c2753685e388ba5a7ac35aaddb300046e --- /dev/null +++ b/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/23a6116f-5b95-4dcd-97e5-95528a6d1fff_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b43eeec9366be7e9604f713ef00134bd82e4f0fac9310ed2867d6e6faf065279 +size 3585547 diff --git a/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/full.md b/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6d3d8004553b48515234a5f6e6050ee93c1cad1d --- /dev/null +++ b/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/full.md @@ -0,0 +1,372 @@ +# Tiny Updater: Towards Efficient Neural Network-Driven Software Updating + +Linfeng Zhang, Kaisheng Ma* + +Institute for Interdisciplinary Information Sciences, Tsinghua University + +zhanglinfeng1997@outlook.com, kaisheng@mail.tsinghua.edu.cn + +# Abstract + +Significant advancements have been accomplished with deep neural networks in diverse visual tasks, which have substantially elevated their deployment in edge device software. However, during the update of neural network-based software, users are required to download all the parameters of the neural network anew, which harms the user experience. Motivated by previous progress in model compression, we propose a novel training methodology named Tiny Updater to address this issue. Specifically, by adopting the variant of pruning and knowledge distillation methods, Tiny Updater can update the neural network-based software by only downloading a few parameters $(10\% \sim 20\%)$ instead of all the parameters in the neural network. Experiments on eleven datasets of three tasks, including image classification, image-to-image translation, and video recognition have demonstrated its effectiveness. Codes have been released in https://github.com/ArchipLab-LinfengZhang/TinyUpdater. + +# 1. Introduction + +With the availability of large-scale datasets [15, 16, 51] and high-performance computing platforms, deep neural networks have achieved remarkable achievements in various visual tasks such as image classification [18, 28, 75, 80], segmentation [53, 62], and object detection [50, 67, 68, 76]. Encouraged by their impressive performance, numerous software developers have effectively integrated neural networks into their software products and deployed them on edge devices such as mobile phones and tablet computers. + +Typically, the development roadmap for a neural network-based software follows the paradigm illustrate in Figure 1(a). Initially, users install the software by downloading all the neural network parameters from a cloud platform. Subsequently, as the software interacts with its users, an abundance of new training data and requirements can be + +![](images/23d4b0134c1e078a5ef075e3a0efaec3351dc53b29617eba4311d559be81d6f1.jpg) + +![](images/a41bc52d9b777ef4eab555d464f5060e5eab4385a97c04334fbd915434a9ba3c.jpg) +Figure 1. Comparison between (a) the traditional model updating scheme and (b) the proposed efficient model updating. In efficient model updating, only around $10\%$ parameters in the neural network are actually changed and thus users only require to download these changed parameters for updating. + +collected. Then, software developers may retrain the neural network with the gathered data and update their software to enhance its performance. However, during the retraining phase, all the neural network parameters are typically changed compared to their values before updating. Consequently, users are compelled to download all the parameters of the neural network again from the cloud platform, seriously impairing their experience. While recent research has explored text prompts and adapter layers to fine-tune a large-scale pre-trained model at a low training cost [52, 79], there have been no prior efforts to reduce the download cost during model updating, which has a more direct impact on users with edge devices. + +This paper proposes the challenge of efficient model updating with the objective of reducing the download overhead of neural network-based software during updating. As + +depicted in Figure 1(b), during the retraining phase, efficient model updating introduces an additional constraint limiting the change of only a small subset (e.g. $10\%$ ) of the neural network parameters compared to the pre-updating model. Consequently, users are only required to download a few parameters that have actually changed instead of all the parameters in the neural network. In general, efficient model updating raises two questions: how to find the optimal parameters that should be changed in the neural network, and how to achieve comparable accuracy with the fully-updated model (i.e., the model which has all the parameters changed during updating). + +To tackle this challenge, we propose a novel neural network training framework named Tiny Updater. Motivated by previous works in model compression, Tiny Updater is composed of the variants of two typical model compression techniques - neural network pruning and knowledge distillation. Firstly, to determine which channels or layers in neural networks should be modified during updating, Tiny Updater applies the pruning technique. As shown in Figure 2, it iteratively calculates the $L_{1}$ -norm distance between the pre-updating model and the post-updating model parameters. The channels with smaller distances are deemed unnecessary for updating and are pruned to their original values before updating. Conversely, the channels with larger distances are considered essential for updating, and thus they are actually changed. Secondly, during the retraining period, we propose to improve the performance of the partially-updated model by distilling the knowledge from a fully-updated teacher model. The partially-updated student model is trained to give predictions that are similar to those of the fully-updated teacher model by optimizing the knowledge distillation loss. This ensures that the partially-updated model achieves comparable performance with the fully-updated model. + +Extensive experiments have been conducted on eleven datasets and three different tasks, including, image classification, video classification, and image-to-image translation, using seven different neural networks. Experimental results demonstrate that Tiny Updater can update the model by changing only $10\%$ to $20\%$ of the parameters with minimal performance degradation. In summary, the main contribution of this paper can be outlined as follows. + +- To the best of our knowledge, we first propose the challenge of efficient model updating to reduce the communication cost of downloading neural network parameters to the edge devices for software updating. +- To tackle this challenge, we propose Tiny Updater, which prunes the fully-updated model to the pre-updating model and distills the knowledge from the fully-updated model to the partially-updated model. + +- Extensive experiments have been conducted on three tasks with eleven datasets, including image classification, image-to-image translation and video recognition. Experimental results demonstrate that Tiny Updater can reduce the training overhead by $80\% \sim 90\%$ with almost no performance degradation. + +# 2. Related Work + +# 2.1. Neural Network Pruning + +Neural network pruning is one of the most effective methods for deep neural network compression by deleting the unimportant neurons [42, 47], filters [29, 38], channels [30, 90] and layers [8, 69]. LeCun et al. and Hassibi et al. first propose to prune neural networks depending on the Hessian matrix [27, 41]. Recently, Han et al. propose deep compression, which finds the important connection in neural networks with its absolute value and then iteratively prunes them [24]. Liu et al. train a meta-network to generate the weights of neural networks and then apply it to search the best-pruned architecture [58]. Ding et al. propose to apply a LSTM to learn the hierarchical characteristics of deep neural networks and then generates the corresponding pruning scheme [17]. Liu et al. propose joint multi-dimension pruning to prune the channels, layers and resolutions at the same time [59]. Frankle et al. propose the lottery ticket hypothesis and find that a standard pruning technique naturally uncovers subnetworks whose initializations made them capable of training effectively [21]. Besides classification, neural network pruning has also be utilized in image-to-image translation [37, 42], unconditional image generation [56], object detection [22, 66], video retrieval [10] and pre-trained language models [14]. + +In this paper, Tiny Updater introduces neural network pruning methods to find the parameters which are actually valuable for model updating. Its main difference compared with previous pruning methods is that Tiny Updater prunes the parameters to their value before updating instead of zero, which makes it easier to be solved than traditional pruning. Our experiments also show that Tiny Updater can be utilized on the pruned models. + +# 2.2. Knowledge Distillation + +Knowledge distillation, also known as student-teacher learning, has become one of the most popular deep learning techniques in various domains [6, 49, 72, 85, 86]. It firstly trains an over-parameterized teacher model and then distills teacher knowledge to a lightweight student. By training the student to mimic the prediction results of the teacher, the student can inherit the knowledge learned by the teacher. Thus, it can achieve better performance than traditional training. The concept of knowledge distillation is first proposed by Bucilua et al. [7] to compress ensemble + +![](images/3fe7b5607e3165fab270d9f037b829b89b04319d209b9fe818443eb1c5596641.jpg) +Figure 2. The details of Tiny Updater. (1) Tiny Updater re-trains the model with the new data collected from users (named as fully-updated models) by changing all the parameters. (2) Secondly, Tiny Updater finds the channels in the fully-updated models which have similar value to the pre-updating model, and prunes these channels to their value before updating (the obtained models are named as partially-updated models). (3) Then, Tiny Updater re-trains the unpruned weights. During re-training, the fully-updated model is utilized as the teacher model for the partially-updated model with knowledge distillation (KD) loss to improve its performance. (4) Tiny Updater repeats step-2 and step-3 until most channels have been pruned to their value before pruning. (5) Finally, users only require to download very few parameters which are actually changed for updating. + +of neural networks, and then extended by Hinton et al. with a temperature hyper-parameter in the softmax function [32]. Following their success, abundant methods have been proposed to distill teacher knowledge in their backbone features [42, 70], spatial attention [20, 84], task-oriented information [87], pixel-wise relation [20, 44, 49, 86], sample-wise relation [63, 65, 78], prediction residual [43] and so on. + +Previous knowledge distillation methods usually transfer teacher knowledge in the categorical probability distribution. Then, abundant methods have been introduced to distill the knowledge in the teacher feature [70] and its variants, such as attention [84, 86], relational information [63, 78], self-supervised knowledge [81] and task-oriented information [87]. Besides distilling knowledge from a large teacher to a tiny student, there have also been fruitful knowledge distillation methods that distill knowledge from deeper layers to shallow layers [88, 89], from all the channels to partial channels [83], from ensemble classifiers to single classifier [40], from multiple frames to few frames [5], from RGB images to RGB-D images [23]. In this paper, we introduce knowledge distillation by distilling the knowledge from a teacher model, which has all the parameters changed during updating to a student model, which is only updated by changing a few parameters. + +# 2.3. Incremental Learning + +Inspired by the observation that human beings can incrementally learn knowledge about new tasks and categories without forgetting the old knowledge, incremental learning (a.k.a. lifelong learning) is proposed to give the neural networks the similar ability. Elastic weight consolidation (EWC) is proposed to maintain the knowledge of old tasks and categories by first estimating the importance of each neuron with Bays estimation or Fisher information matrix and then training them with a consolidation constraint [39, 55, 77]. Then, memory aware synapse [1] is proposed to compute the importance score of neural net + +work parameters in an unsupervised and online manner, which is also similar to the Hebbian learning in biological systems [31]. Besides these consolidation-based methods, knowledge distillation methods have also been utilized in incremental learning to maintain the old knowledge by learning its response to new tasks [13, 48, 92]. Besides image classification, abundant recent research has applied incremental learning to the other visual tasks, such as object detection [11, 26], action recognition [60] and imaged-to-image translation [74]. + +Both the proposed efficient model updating and incremental learning expect the neural network to maintain its knowledge of old parameters. Their main difference is that: (a) Incremental learning does not limit the change in old parameters strictly, while efficient model updating has a direct constraint on the number of changed parameters to reduce the download overhead for edge devices. (b) Incremental learning usually assumes that the data for old knowledge is not available while efficient model updating can still access data, which is more practical to industrial applications. (c) The target of incremental learning is to build human-like artificial intelligence which can incrementally learn various tasks, while the target of Efficient Model Updating is to reduce the number of parameters updated when new data for the same task is collected. + +# 2.4. Efficient Pretrained Model Finetuning + +Large-scale pretrained models recently have shown powerful representation ability in both natural language processing and computer vision. However, since these pretrained models usually have billions of parameters, directly finetuning them on downstream tasks usually suffers from massive training overhead, making them impractical in real-world applications. Recently, prompts methods have been proposed to address this problem by designing specific input templates for the specific downstream tasks [4,25,45,52, 54,79,91]. Moreover, adapter methods have also been intro- + +duced to finetune several additional trainable layers instead of the whole pre-trained model [3, 33]. These methods are firstly proposed for pre-trained language models and then extended to language-vision multi-modal models and vision models [12, 36, 82]. Both these efficient finetuning methods and the proposed efficient model updating aim to freeze the parameters of neural networks. Their main difference is that: (a) Efficient finetuning is usually applied to the models which are firstly pre-trained on large-scale datasets to learn task-unbiased knowledge and then finetuned for specific downstream tasks. Instead, the pre-updating models in Efficient Model Updating are trained with very few training samples (e.g. only $20\%$ in our experiments) and then updated with the new data collected from users. (b) The target of efficient finetuning is to reduce the training overhead of pre-trained models and improve their performance in downstream tasks. In contrast, the target of Efficient Model Updating is to reduce the download overhead during software updating. (c) Besides, adapter-based methods can be considered as a special case in Tiny Updater where all the parameters of backbone layers are frozen and all the adapter layers are updated. Our experiments also demonstrate that Tiny Updater achieves clearly better performance than directly applying efficient model updating methods. Please refer to Appendix A for a more detailed comparison among the proposed efficient model updating, incremental learning, and efficient pretrained model finetuning. + +# 3. Methodology + +# 3.1. Efficient Model Updating + +Given a training dataset $\mathcal{D} = \{(x_1, y_1), \dots, (x_n, y_n)\}$ , the software developers firstly train a deep neural network $\mathcal{F}$ with parameter $\Theta$ for their software. After deploying $\mathcal{F}$ on the edge devices, abundant new training samples $\{(x_{n+1}, y_{n+1}), \dots, (x_{n+m}, y_{n+m})\}$ can be collected from users and the dataset can be extended as $\mathcal{D}^+ = \mathcal{D} \cup \{(x_{n+1}, y_{n+1}), \dots, (x_{n+m}, y_{n+m})\}$ . Then, software developers can re-train $\mathcal{F}$ on $\mathcal{D}^+$ to obtain better performance, whose parameters can be denoted as $\Phi$ . Usually, during this re-training step, all the values in $\theta$ can be totally different from that in $\Phi$ . Thus, the users have to download all the parameters of $\Phi$ to update the software, which harms the user experience and limits the frequently updating of neural network-driven software. In such kind of paradigm, we name the model before updating as the pre-updating model, the model after updating as the post-updating model, the model with all the parameters updated as the fully-updated model, and the model with partial parameters updated as the partially updated model. The target of efficient model updating is to obtain a partially updated model which has most of the parameters unchanged while achieving similar performance to the fully-updated model. + +# Algorithm 1 The proposed Tiny Updater. + +Input: Dataset $\mathcal{D}^{+} = \{(x_{1},y_{1}),\dots,(x_{n + m},y_{n + m})\}$ , the pre-updating model $\mathcal{F}$ with parameter $\Theta$ , an expected ratio of updated parameters $\tau$ . + +Output: The partially-updated model $\mathcal{F}$ with parameter $\Phi$ . // Update the model with $\mathcal{D}^+$ . + +Initialize the partially-updated model $\mathcal{F}$ with parameter $\Phi$ . while $\mathcal{F}_{\Phi}$ is not converged do: + +Sample a batch of data $\mathcal{X}$ from $\mathcal{D}^+$ + +Compute $\hat{y} \coloneqq \mathcal{F}_{\Phi}(\mathcal{X})$ + +Compute the task loss between $y$ and $\hat{y}$ + +Back propagate gradients and update $\Phi$ . + +//Pruning, and re-training with KD loss. + +Initialize the parameters of the teacher $\mathcal{T} \coloneqq \Phi$ , an index set of pruned weights as $\mathcal{I} = \{\}$ . + +while $\frac{Card(\mathcal{T})}{Card(\Theta)} < 1 - \tau$ do: + +Compute the $L_{1}$ -norm distance between $\Phi$ and $\theta$ . + +Append the indices of channels which have relatively smaller distance to $\mathcal{I}$ . + +$\Phi [\mathcal{I}]\coloneqq \Theta [\mathcal{I}].$ //Prune $\Phi$ to $\Theta$ + +while $\mathcal{F}_{\Phi}$ is not converged do: + +Sample a batch of data $\mathcal{X}$ from $\mathcal{D}^+$ + +Compute $\hat{y} \coloneqq \mathcal{F}_{\Phi}(\mathcal{X})$ and $y_{t} \coloneqq \mathcal{F}_{\mathcal{T}}(\mathcal{X})$ . + +Compute the task loss between $y$ and $\hat{y}$ , and KD loss between $y$ and $y_{t}$ . + +Back propagate gradients and update the parameters of $\Phi$ which are not in $\mathcal{I}$ . + +Return The partially-updated model $\mathcal{F}$ with parameter $\Phi$ . + +# 3.2. Tiny Updater + +Fruitful previous works in model compression have successfully proven that even a very tiny neural network can have powerful representation ability, which motivates us to propose to learn the knowledge in the collected training data $\mathcal{D}^{+}$ with only a few parameters instead of all the parameters. The optimization objective of Tiny Updater can be formulated as + +$$ +\begin{array}{l} \arg \min _ {\Phi} \frac {1}{n + m} \sum_ {i = 1} ^ {n + m} \mathcal {L} _ {\text {t a s k}} \left(x _ {i}, y _ {i}\right) \tag {1} \\ \text {s u b j e c t} \quad \frac {\left| \Theta - \Phi \right| _ {0}}{\operatorname {C a r d} (\Theta)} < \tau \\ \end{array} +$$ + +where $\mathcal{L}_{\mathrm{task}}$ indicates the original task-specific loss function, such as cross-entropy loss for image classification. $|\cdot|_0$ indicates the $L_0$ -norm, which measures the number of non-zero elements in a tensor. $\operatorname{Card}(\cdot)$ denotes the cardinality, which describes the number of parameters in a tensor. $\tau$ is a ratio threshold that determines how many parameters should be changed during updating. It is observed that when $\theta$ in Equation (1) becomes zero, the optimization objective is similar to another deep learning technique - network pruning. + +![](images/b36fefbdb4b5e217ed12a10095a4d2a9a10435740ee707ed084c273a3e73c389.jpg) +(a) + +![](images/b290562004e043f34a63d1a885afa74aadc369dcca042e093c3928fc18272e6d.jpg) + +![](images/9f8ff91c52461d6f7c1ff64c941ffe7ee8d3f2d9d1e43472e0fe3c9e1d36c0af.jpg) + +![](images/aa3dd3e6487db150fe7d07f893ea9e74a0e125e65a686fb26510f79ae2bb2cdd.jpg) + +![](images/897b7e72e728fc446de4342db63d21a94b4f8dcec205374df982c35aa04d41da.jpg) +(b) + +![](images/36c1f9f04cc5c2bd3d34850338c8a67f6663b499c237529e1bf9a6b06854abd8.jpg) +Figure 3. Experiments on four fine-grained image classification datasets with ResNet50. The pre-updating and the post-updating models are trained with $25\%$ and $100\%$ training data, respectively. In subfigure (a), the newly collected images come from all categories uniformly. In subfigure (b), the newly collected images come from only the categories that are not available before updating. + +![](images/c1ed06d9070221be0875fc09117adf5422bbfa3a0602904b55af8b75d6943d44.jpg) + +![](images/daf44520f2e6b57347a837ee249401eefd5dfc9e4a10e89029570eeff551cca2.jpg) + +ing. Thus, we propose to solve this problem by introducing iterative pruning methods. Besides, knowledge distillation is also introduced to improve the model performance with only a few parameters changed during updating with the teacher from a model with all the parameters changed. Please note that Tiny Updater is a framework that applies pruning and knowledge distillation methods to tackle the challenge of efficient model updating and it does not propose any new pruning and knowledge distillation method. + +The details of Tiny Updater are shown in Algorithm 1. With the extended dataset $\mathcal{D}^+$ , it first trains a fully-updated model $\mathcal{F}_{\Phi}$ without any constraints and copies its parameters to another neural network $\mathcal{F}_{\mathcal{T}}$ , which is considered as the teacher network. Then, channel pruning is utilized to find the parameters in $\Phi$ (e.g. channels in convolutional layers) whose value is not significantly changed compared with the pre-updating model $\Theta$ . Then, these parameters are reset to their value in $\Theta$ and are not further trained in the later re-training. During the re-training period, the partially-updated model is trained with not only the origin task loss but also the knowledge distillation loss from the teacher of the fully-updated model $\mathcal{F}_{\mathcal{T}}$ . Please note that the proposed Tiny Updater is a framework that employs pruning and knowledge distillation to tackle the challenge of efficient model updating and it does not depend on any specific pruning and knowledge distillation methods. + +# 4. Experiment + +# 4.1. Image Classification + +Experiment Setting Our method has mainly been evaluated on seven image classification datasets, including CI- + +Table 1. The performance of Tiny Updater on ImageNet. Models are firstly trained with $25\%$ data and then updated with all the data. "Updated Ratio" indicates the ratio of updated parameters. + +
ModelUpdated Ratio (%)Top-1 AccuracyTop-5 Accuracy
ResNet5000.0068.86 -6.4488.74 -3.46
21.5474.76 -0.5491.70 -0.50
42.4275.12 -0.1891.93 -0.27
65.6976.25 +0.9593.14 -0.94
100.075.3092.20
MobileNetv200.0062.23 -3.0884.25 -2.29
24.3164.54 -0.7785.53 -1.01
31.6665.23 -0.0886.42 -0.12
66.5566.30 +0.9987.19 +0.65
87.6066.86 +1.5587.50 +0.96
100.065.3186.54
Swin Transformer00.0072.04 -9.2689.75 -5.85
18.5679.91 -1.3994.33 -1.27
36.3180.87 -0.4394.53 -1.07
80.2681.73 +0.4395.74 +0.14
100.081.3095.60
+ +FAR10, CIFAR100, and ImageNet for general image classification, Stanford Car, Oxford Flower, Stanford Dog, FGVC Aircraft for fine-grained image classification. Please refer to Appendix B for the details of these datasets. Note that on the fine-grained image classification datasets, models are initialized with backbone weights pre-trained on ImageNet. In ImageNet experiments, we adopt the basic training policy from PyTorch [64] with task-oriented feature distillation [87]. On the other datasets, each model is trained by 200 epochs with the naive logit and feature knowledge distillation loss [32, 70]. + +Main Results The performance of Tiny Updater on ImageNet, CIFAR10&100, and four fine-grained image classifi- + +![](images/cb65524168a6ad9ded1610690e13f54cc0a722db902d46be70af15da0dc69208.jpg) +Figure 4. Experiments on CIFAR with ResNet50. The pre-updating models and fully-updated models are trained with $25\%$ and $100\%$ training data, respectively. Tiny Updater with random pruning indicates pruning the randomly selected channels. + +cation are shown in Table 1, Figure 4, Figure 3, respectively. It is observed that: (i) Tiny Updater achieves consistent effectiveness on all seven datasets. By changing around $20\%$ parameters during updating, the neural networks trained with Tiny Updater achieve a similar performance with fully-updated models on all these datasets. (ii) On ImageNet, Tiny Updater achieves significant performance on the regular convolutional network such as ResNet [28], the efficient convolutional networks such as MobileNetv2 [71], and Swin Transformer [57]. (iii) With the proposed Tiny Updater, when a large ratio of parameters is changed, it even leads to higher performance than the fully-updated model. For instance, on Oxford Flowers datasets in Figure 3, the $84\%$ updated model trained with Tiny Updater has $0.25\%$ higher accuracy than the fully-updated model. We suggest this accuracy benefit is caused by the knowledge distillation loss in Tiny Updater. As pointed out by some previous research [2, 61], even if the student and the teacher have similar performance, knowledge distillation can still lead to consistent accuracy benefits. + +Ablation Study Tiny Updater is mainly composed of two modules - pruning and knowledge distillation. Ablation studies on CIFAR10 and CIFAR100 are shown in Figure 4. It is observed that: (i) Compared with Tiny Updater with random pruning, Tiny Updater with $L_{1}$ -norm based pruning leads to consistently higher accuracy (around $6\%$ ) especially when only a low ratio ( $0 \sim 50\%$ ) of parameters are changed, indicating $L_{1}$ -norm is an effective metric to find which channels are actually important for model updating. (ii) Significant accuracy boosts can be observed by applying knowledge distillation during re-training. For instance, on CIFAR100, when around $10\%$ and $20\%$ parameters are changed, $1.79\%$ and $1.15\%$ accuracy boost can be observed with knowledge distillation, respectively. These observations indicate that the pruning and knowledge distillation in Tiny Updater have their individual effectiveness. + +Categorical Incremental Updating Figure 3(b) shows the effectiveness of Tiny Updater in the categorical incremental settings, where the pre-updating models are trained + +![](images/6a2a06b3d9b663ad5b6aa4b406f336cab67aebff51b587193133caea518124dc.jpg) +Figure 5. Experiments on CIFAR10 with the pre-updating models trained with different ratios of data. + +with $25\%$ training data belonging to $25\%$ categories. During the updating period, the model is further trained with data of all the categories. It is observed that Tiny Updater can still achieve significant performance by only changing around $10\%$ parameters, indicating that Tiny Updater is also effective on the categorical incremental updating setting. + +Influence from Pre-updating Model In previous experiments, the pre-updating models are trained with $25\%$ data. With less training data, the pre-updating models tend to have less representative ability, and thus the performance of Tiny Updater tends to be reduced. In this subsection, we study how the performance of the pre-updating model influence Tiny Updater on CIFAR10 in Figure 3(a). It is observed that when most of the parameters $(>50\%)$ are changed during updating, the accuracy gap between different pre-updating models is not significant $(< 0.2\%)$ . In contrast, when only a few parameters $(10\% \sim 20\%)$ are changed, the ratio of data for training pre-updating models has a more significant influence. For example, compared with the pre-updating model trained with $25\%$ data, the pre-updating model trained with only $5\%$ data leads to around $2.5\%$ accuracy loss when $20\%$ parameters are changed during updating, indicating that the performance of the pre-updating model has a direct influence to Tiny Updater. Specifically, when the pre-updating model is not trained with any data, it can be considered as a model with all parameters initialized with zero matrices, and thus Tiny Updater in this case degenerates to the common neural network pruning problem. + +Multi-step Model Updating Previous experiments mainly show the result of one-step updating from using $25\%$ to $100\%$ training data. In this subsection, we show the performance of Tiny Updater in a multi-step updating, where the model is first trained with $5\%$ data and then updated with $10\%$ , $15\%$ , $20\%$ , $25\%$ , and $100\%$ training data, successively. As shown in Table 4, there is almost no accuracy loss when around $20\%$ parameters are changed for each update. When there are around $10\%$ parameters changed for each update, accuracy loss becomes significant with the increment on the update iterations (from $-0.44\%$ to $-2.33\%$ ). This observation indicates that the accuracy loss tends to be accumulated in multi-step updating when an extremely low ratio of parameters is changed for updating. + +Table 2. Experiments of multi-step updating on CIFAR10. Models are firstly trained with $5\%$ training data and then updated by five times with $10\%$ , $15\%$ , $20\%$ , $25\%$ , and $100\%$ training data by changing all, $10\%$ and $20\%$ parameters. + +
Data Ratio +Params. Ratio5%10%15%20%25%100%
Fully-Updated46.564.278.884.687.895.3
10% Params. Updated46.563.878.383.285.992.9
20% Params. Updated46.564.379.084.287.594.8
+ +Tiny Updater on Pruned Models In this subsection, we study whether the proposed Tiny Updater can be utilized on a pruned model. Concretely, we first train a ResNet18 model with $25\%$ training data on CIFAR10, which achieves $87.43\%$ top-1 accuracy. Then, we apply the $L_{1}$ -norm pruning technique to prune $53\%$ neurons at the expense of $0.37\%$ accuracy loss $(87.43\% \rightarrow 87.06\%)$ . Thirdly, the proposed Tiny Updater is applied to update this model with $100\%$ training data, which achieves $94.22\%$ at the expense of changing $24.3\%$ parameters $(87.06\% \rightarrow 94.22\%)$ . In contrast, a pruned fully-updated model with the same pruning ratio achieves $94.87\%$ accuracy, which is only $0.65\%$ higher than the pruned model with Tiny Updater $(87.06\% \rightarrow 94.87\%)$ . These observations indicate that the proposed Tiny Updater can be utilized on pruned models. + +Comparison with Finetuning and Adapter Finetuning and adapter-based methods are two well-known efficient training methods and the parameters changed in these two methods are also much fewer than global finetuning. Figure 4 gives the comparison between Tiny Updater and these two methods on CIFAR10, CIFAR100 with ResNet18. It is observed that when the same number of parameters are changed, Tiny Updater outperforms these two methods by a clear margin, indicating that Tiny Updater is more effective than directly applying previous methods. + +# 4.2. Image-to-Image Translation with GANs + +Experiment Setting We mainly evaluate Tiny Up- dater on image-to-image translation with CycleGAN for Horse $\rightarrow$ Zebra translation, and Pix2Pix for Edge $\rightarrow$ Shoe translation. CycleGAN is a typical unpaired image-toimage translation model, which has a ResNet-based generator trained with GAN loss and cycle consistency [94]. Pix2Pix is a typical paired image-to-image translation model, which has a U-Net architecture generator trained with conditional GAN loss and the $L_{1}$ -norm loss [35]. Horse $\rightarrow$ Zebra is an unpaired dataset that translates natural images of horses to zebras and vice versa. It consists of 1,187 horse images and 1,474 zebra [93]. Edge $\rightarrow$ Shoe is a paired data that translates the edges of shoes to their corresponding natural images [34]. Fréchet Inception Distance (FID), which measures the distance between the feature dis + +![](images/fda6bd8009aee50fb7cdd3009802e081d435b992fca883c2ee4f7fac4763dbd4.jpg) +Figure 6. Qualitative analysis on Horse $\rightarrow$ Zebra and Zebra $\rightarrow$ Horse with CycleGAN. The partially-updated model trained with Tiny Updater has around $20\%$ parameters changed. The pre-updating models are trained with $25\%$ data. + +![](images/a0abb3c0873994993d3a63401fe5271041001d7e4038ecbdbbf9983b9ab5268c.jpg) +Figure 7. Experiments of image-to-image translation on Horse $\rightarrow$ Zebra and Zebra $\rightarrow$ Horse with CycleGAN, and Edge $\rightarrow$ Shoe with Pix2Pix. In all these experiments, the pre-updating models are trained with $25\%$ training data. + +tribution of the real and the generated images, is utilized as the metric for both datasets. A lower FID indicates the synthetic images have better quality. + +Main Results Experiments of Tiny Updater on image-to-image translation are shown in Figure 7. It is observed that on the three image-to-image translation tasks, Tiny Updater is able to improve model performance by $5\sim 10$ FID when around $60\%$ parameters are changed during updating. Besides, Tiny Updater can update the models by changing $40\%$ parameters with no performance loss or by changing only $20\%$ parameters with a slight FID increment. Qualitative analysis in Figure 6 shows that the pre-updating models can not transform the whole body of horses into the zebras, or remove all the stripes of zebras. In contrast, with Tiny Updater, the partially-updated model with only $20\%$ parameters changed can address this problem and achieve comparable performance with the fully-updated models, indicating that Tiny Updater enables the neural network to be successfully updated by changing only a few parameters. + +# 4.3. Video Recognition + +Experiment Setting We evaluate Tiny Updater in video recognition datasets including UCF-101 [73] and Diving-48 [46] with video recognition models including SlowOnly [19] and Inception3D [9]. UCF-101 is an ac + +![](images/358257611ad7f6a95e1513dcf69c6a0b721ec66d24f8dac003ab9d5b85dd59ce.jpg) +Figure 8. Experimental results of Tiny Updater of video recognition on UCF-101 and Diving-48 with SlowOnly and Inception3D (I3D). The pre-updating models are trained with $25\%$ training data. + +![](images/10b2c5672046302d5922cdd9e9519eae47a6b7901e9b1485cdb37980c09cc7e3.jpg) + +![](images/ea2c9945c955e96f6d29edd0a67747d2c05253749bd49b9c617434b727415a5d.jpg) + +![](images/cfed865dc1be1ce7088535073fd1272c1bb3045060b205b4306a76d23097890e.jpg) + +tion recognition dataset with 101 action classes over 13,000 video clips [73]. Diving-48 is a fine-grained video dataset of competitive diving. It has around 18,000 video clips belonging to 48 dive sequences [9]. Both top-1 and top-5 accuracy are reported. + +Main Results Figure 8 shows the performance of Tiny Updater on video recognition. It is observed that on both SlowOnly and Inception3D (I3D), UCF-101 and Diving-48, the model with only $20\%$ parameters changed trained by Tiny Updater achieves comparable and even higher performance than the fully-updated models. Moreover, the model with only $10\%$ parameters changed during updating leads to only around $0.5\%$ accuracy loss. + +# 5. Discussion + +# 5.1. Choice of Pruning Granularity + +Existing network pruning methods can be mainly divided into regular pruning (e.g. channel-wise pruning) and irregular pruning (e.g. element-wise pruning). Usually, irregular pruning can achieve a higher compression rate, while regular pruning is more friendly to hardware. In this paper, we prefer to use regular pruning in Tiny Updater and suggest that using irregular pruning may cause the problem of index overhead and memory access overhead. + +Index Overhead When Tiny Updater is utilized for updating neural network-based software, users should download not only the parameters which are changed during updating but also an index file that records the corresponding position of each changed parameter in the neural network. Given a neural network with $N$ parameters, when irregular pruning is utilized, for each updated parameter, $\log_2N$ bits storage space is required to record which layer, filter, and position in the kernel this parameter should be. Thus, an ignorable additional downloading overhead is caused. In contrast, the index overhead for channel-wise pruning is $\log_2\frac{N}{C\times K^2}$ bits, which is small enough to be almost ignorable. Here $C$ and $K$ indicate the channel number and the kernel size, respectively. + +Memory Access Overhead In Tiny Updater, irregular pruning can result in the updated parameters being distributed across various neural layers, channels, and filters, leading to their storage in different data blocks of the file system. Consequently, when updating the model with the downloaded parameters, almost all data blocks of the edge device need to undergo a writing operation, resulting in a significant memory access overhead. In contrast, regular pruning ensures that the updated weights are channel-wise and, therefore, can be stored in the same data blocks. Consequently, only these data blocks for the updated parameters require memory access, potentially reducing the updating overhead on edge devices. + +# 6. Conclusion + +This paper introduces the challenge of efficient model updating to reduce the download overhead during neural network-based software updating. To this end, we propose Tiny Updater, which aims to change only a small subset of the parameters during model retraining, thereby reducing the number of parameters that need to be downloaded by users. Tiny Updater comprises an iterative pruning technique that prunes the parameters of the fully-updated models to the pre-updating models and a knowledge distillation technique that treats the fully-updated models as teachers and the partially-updated models as students. + +Extensive experiments were conducted on eleven datasets covering general image classification, fine-grained image classification, video recognition, paired image-to-image translation, and unpaired image-to-image translation. The results demonstrate that Tiny Updater can update neural network-based software by changing only $10\%$ to $20\%$ of the parameters with almost no loss in accuracy. In addition, the experimental results show the effectiveness of Tiny Updater in various scenarios, including categorical updating, multi-step updating, and pruning. We believe that this paper will encourage further research to address the challenge of efficient model updating on edge devices. + +Table 3. Comparison among incremental learning, efficient finetuning, and the proposed efficient model updating. + +
ProblemTargetPre / Post TasksUsage of Training DataConstraints on Parameters
Incremental LearningThe human-like ability of incrementally learning new tasks without forgetting old tasks.Different.A small ratio of data is used for training pre-updating models. More data is used for updating. Data for pre-training data can not be accessed during updating (learning new tasks).No direct constraints. Reducing the number of changed parameters sometimes lead to positive influence.
Efficient FinetuningFinetuning pre-trained models for downstream tasks with low training overhead, and achieving better downstream performance.Different.Pre-updating models are pretrained on large-scale datasets, and then finetuned on much less data for downstream tasks. Usually do not access data for pre-training during finetuning.No direct constraints. Do not finetune the whole model usually leads to no changes in pre-trained weights. Additional parameters in adapter layers and prompts are required.
Efficient Model UpdatingUpdate models with new data by changing only a few parameters to reduce communicat-ion overhead.Same.A small ratio of data is used for pre-training models, and these data can still be accessed during updating.Having direct constraints, the changed parameters should be as few as possible.
+ +# 7. Acknowledgements + +This research was partially supported by National Key R&D Program of China (2022YFB2804103), Key Research and Development Program of Shaanxi (2021ZDLGY01-05), Tsinghua University Dushi Program, National Natural Science Foundation of China (20211710187), and Tsinghua University Talent Program. + +# A. Detailed Comparison + +A detailed comparison among incremental learning, efficient finetuning and the proposed efficient model updating is shown in Table 3. + +# B. Detailed Experimental Setting + +CIFAR10/100 are two datasets for low resolution general image classification with 50,000 images and 1,000 images in the training and validating set, respectively. ImageNet is a large-scale dataset for general image classification with images belonging to 1,000 categories. Stanford Car is a dataset for fine-grained image classification with car images belonging to 197 classes. Oxford Flower is a dataset for fine-grained image classification which consists of 102 different categories of flowers common to the UK. Stanford Dog is a dataset for fine-grained image classification which contains 20,580 dog images of 120 breeds of dogs from around the world. FGVC Aircraft is a dataset for fine-grained image classifier on which contains 10,000 images of aircrafts spanning 100 aircraft models. + +# C. Experiments on More Datasets and Tasks + +Experimental results on more kinds of tasks and datasets are shown in Table 4 and Table 5. It is observed that + +Table 4. Comparison between our method with previous efficient finetuning methods on ImageNet with Swin Transformer. Models are firstly trained with $25\%$ data and then updated with all the data. + +
Updating MethodRatio of Updated Params.(%)Top-1 Acc (%)
Pre-updating Model0.0%72.04
Global Finetuning100.0%81.30
Linear Finetuning1.1672.23
Non-linear Finetuning234.675.53
Non-linear Finetuning366.378.80
Adapter13.572.40
Adapter217.274.39
Visual Prompt Tuning≤ 1%72.32
Ours136.3180.87
Ours218.5679.91
+ +Table 5. Experimental results on more tasks. + +
TaskDatasetModelUpdated Params (%)Acc. / mAP / PSNR
Cross Domain Image ClassificationOffice-31ResNet500.0092.3
18.798.3
10.998.0
100.098.5
Object DetectionMS COCOFaster RCNN0.0030.5
27.336.2
14.635.6
100.036.4
Single Image Super-ResolutionDIV2KESRGAN0.0033.90
22.434.60
13.034.56
100.034.61
+ +our method leads to consistent good performance on cross-domain image classification, object detection and image super-resolution. Besides, our method outperforms previous efficient finetuning methods on ImageNet. + +# References + +[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 3 +[2] Zeyuan Allen-Zhu and Yuanzhi Li. Towards understanding ensemble, knowledge distillation and self-distillation in deep learning. CoRR, abs/2012.09816, 2020. 6 +[3] Ankur Bapna, Naveen Arivazhagan, and Orhan First. Simple, scalable adaptation for neural machine translation. arXiv preprint arXiv:1909.08478, 2019. 4 +[4] Eyal Ben-David, Nadav Oved, and Roi Reichart. Pada: A prompt-based autoregressive approach for adaptation to unseen domains. arXiv preprint arXiv:2102.12206, 2021. 3 +[5] Shweta Bhardwaj, Mukundhan Srinivasan, and Mitesh M Khapra. Efficient video classification using fewer frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 354-363, 2019. 3 +[6] Cunling Bian, Wei Feng, Liang Wan, and Song Wang. Structural knowledge distillation for efficient skeleton-based action recognition. IEEE Transactions on Image Processing, 30:2963-2976, 2021. 2 +[7] Cristian Bucilua, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 535-541. ACM, 2006. 2 +[8] Efe Camci, Manas Gupta, Min Wu, and Jie Lin. Qlp: Deep q-learning for pruning deep neural networks. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 2 +[9] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7, 8 +[10] Hyun Sung Chang, Sanghoon Sull, and Sang Uk Lee. Efficient video indexing scheme for content-based retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 9(8):1269-1279, 1999. 2 +[11] Li Chen, Chunyan Yu, and Lvcai Chen. A new knowledge distillation for incremental object detection. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2019. 3 +[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 4 +[13] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2534-2543, 2021. 3 +[14] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2 + +[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 248-255, 2009. 1 +[17] Guiguang Ding, Shuo Zhang, Zizhou Jia, Jing Zhong, and Jungong Han. Where to prune: Using lstm to guide data-dependent soft pruning. IEEE Transactions on Image Processing, 30:293-304, 2021. 2 +[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1 +[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 7 +[20] Yingchao Feng, Xian Sun, Wenhui Dao, Jihao Li, and Xin Gao. Double similarity distillation for semantic image segmentation. IEEE Transactions on Image Processing, 30:5363-5376, 2021. 3 +[21] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635, 2018. 2 +[22] Sharath Girish, Shishira R Maiya, Kamal Gupta, Hao Chen, Larry S Davis, and Abhinav Shrivastava. The lottery ticket hypothesis for object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 762-771, 2021. 2 +[23] Saurabh Gupta, Judy Hoffman, and Jitendra Malik. Cross modal distillation for supervision transfer. In CVPR, pages 2827-2836, 2016. 3 +[24] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In International Conference on Learning Representations, 2016. 2 +[25] Xu Han, Weilin Zhao, Ning Ding, Zhiyuan Liu, and Maosong Sun. *Ptr: Prompt tuning with rules for text classification.* arXiv preprint arXiv:2105.11259, 2021. 3 +[26] Yu Hao, Yanwei Fu, Yu-Gang Jiang, and Qi Tian. An end-to-end architecture for class-incremental object detection with knowledge distillation. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2019. 3 +[27] Babak Hassibi and David G Stork. Second order derivatives for network pruning: Optimal brain surgeon. In Advances in neural information processing systems, pages 164-171, 1993. 2 +[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- + +ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6 +[29] Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang. Soft filter pruning for accelerating deep convolutional neural networks. arXiv preprint arXiv:1808.06866, 2018. 2 +[30] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 1389-1397, 2017. 2 +[31] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology Press, 2005. 3 +[32] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. In NeurIPS, 2014. 3, 5 +[33] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 4 +[34] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Pix2pix datasets, http://efrosgans.eecs.berkeley.edu/pix2pix/datasets/.7 +[35] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 7 +[36] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 4 +[37] Qing Jin, Jian Ren, Oliver J Woodford, Jiazhuo Wang, Geng Yuan, Yanzhi Wang, and Sergey Tulyakov. Teachers do more than teach: Compressing image-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13600-13611, 2021. 2 +[38] Hyeong-Ju Kang. Accelerator-aware pruning for convolutional neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 30(7):2093-2103, 2020. 2 +[39] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3 +[40] Xu Lan, Xiatian Zhu, and Shaogang Gong. Knowledge distillation by on-the-fly native ensemble. arXiv preprint arXiv:1806.04606, 2018. 3 +[41] Yann LeCun, John S Denker, and Sara A Solla. Optimal brain damage. In Advances in neural information processing systems, pages 598-605, 1990. 2 +[42] Muyang Li, Ji Lin, Yaoyao Ding, Zhijian Liu, Jun-Yan Zhu, and Song Han. Gan compression: Efficient architectures for interactive conditional gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5284-5294, 2020. 2, 3 +[43] Xuewei Li, Songyuan Li, Bourahla Omar, Fei Wu, and Xi Li. Reskd: Residual-guided knowledge distillation. IEEE Transactions on Image Processing, 30:4735-4746, 2021. 3 + +[44] Xiaojie Li, Jianlong Wu, Hongyu Fang, Yue Liao, Fei Wang, and Chen Qian. Local correlation consistency for knowledge distillation. In European Conference on Computer Vision, pages 18-33. Springer, 2020. 3 +[45] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3 +[46] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 7 +[47] Yun Li, Zechun Liu, Weiqun Wu, Haotian Yao, Xiangyu Zhang, Chi Zhang, and Baoqun Yin. Weight-dependent gates for network pruning. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 2 +[48] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 3 +[49] Zeqi Li, Ruowei Jiang, and Parham Aarabi. Semantic relation preserving knowledge distillation for image-to-image translation. In European Conference on Computer Vision, pages 648-663. Springer, 2020. 2, 3 +[50] Tsung-Yi Lin, Priyal Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE transactions on pattern analysis and machine intelligence, 2018. 1 +[51] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1 +[52] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint arXiv:2107.13586, 2021. 1, 3 +[53] Weide Liu, Guosheng Lin, Tianyi Zhang, and Zichuan Liu. Guided co-segmentation network for fast video object segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 31(4):1607-1617, 2021. 1 +[54] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602, 2021. 3 +[55] Xialei Liu, Marc Masana, Luis Herranz, Joost Van de Weijer, Antonio M Lopez, and Andrew D Bagdanov. Rotate your networks: Better weight consolidation and less catastrophic forgetting. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 2262-2268. IEEE, 2018. 3 +[56] Yuchen Liu, Zhixin Shu, Yijun Li, Zhe Lin, Federico Peruzzi, and Sun-Yuan Kung. Content-aware gan compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12156-12166, 2021. 2 +[57] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In + +Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6 +[58] Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting Cheng, and Jian Sun. Metapruning: Meta learning for automatic neural network channel pruning. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 2 +[59] Zechun Liu, Xiangyu Zhang, Zhiqiang Shen, Yichen Wei, Kwang-Ting Cheng, and Jian Sun. Joint multi-dimension pruning via numerical gradient update. IEEE Transactions on Image Processing, 30:8034-8045, 2021. 2 +[60] Rashid Minhas, Abdul Adeel Mohammed, and Q. M. Jonathan Wu. Incremental learning in human action recognition based on snippets. IEEE Transactions on Circuits and Systems for Video Technology, 22(11):1529-1541, 2012. 3 +[61] Hossein Mobahi, Mehrdad Farajtabar, and Peter L Bartlett. Self-distillation amplifies regularization in hilbert space. arXiv preprint arXiv:2002.05715, 2020. 6 +[62] Diego Ortego, Juan C. Sanmiguel, and José M. Martínez. Hierarchical improvement of foreground segmentation masks in background subtraction. IEEE Transactions on Circuits and Systems for Video Technology, 29(6):1645-1658, 2019. 1 +[63] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 3 +[64] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5 +[65] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In Proceedings of the IEEE International Conference on Computer Vision, pages 5007-5016, 2019. 3 +[66] Zheng Qin, Zeming Li, Zhaoning Zhang, Yiping Bao, Gang Yu, Yuxing Peng, and Jian Sun. Thundernet: Towards real-time generic object detection on mobile devices. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 2 +[67] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1 +[68] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91-99, 2015. 1 +[69] Youngmin Ro and Jin Young Choi. Layer-wise pruning and auto-tuning of layer-wise learning rates in fine-tuning of deep networks. arXiv preprint arXiv:2002.06048, 2020. 2 + +[70] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In International Conference on Learning Representations, 2015. 3, 5 +[71] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 6 +[72] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 2 +[73] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7, 8 +[74] Daniel Stanley Tan, Yong-Xiang Lin, and Kai-Lung Hua. Incremental learning of multi-domain image-to-image translations. IEEE Transactions on Circuits and Systems for Video Technology, 31(4):1526-1539, 2021. 3 +[75] Mingxing Tan and Quoc V Le. Efficientnet: Rethinking model scaling for convolutional neural networks. arXiv preprint arXiv:1905.11946, 2019. 1 +[76] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020. 1 +[77] Brian Thompson, Jeremy Gwinnup, Huda Khayrallah, Kevin Duh, and Philipp Koehn. Overcoming catastrophic forgetting during domain adaptation of neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2062–2068, 2019. 3 +[78] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1365-1374, 2019. 3 +[79] Ruize Wang, Duyu Tang, Nan Duan, Zhongyu Wei, Xuanjing Huang, Guihong Cao, Daxin Jiang, Ming Zhou, et al. K-adapter: Infusing knowledge into pre-trained models with adapters. arXiv preprint arXiv:2002.01808, 2020. 1, 3 +[80] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5987-5995, 2017. 1 +[81] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In European Conference on Computer Vision, pages 588-604. Springer, 2020. 3 +[82] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021.4 +[83] Jiahui Yu and Thomas S. Huang. Universally slimmable networks and improved training techniques. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 3 + +[84] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In International Conference on Learning Representations, 2017. 3 +[85] Linfeng Zhang, Xin Chen, Xiaobing Tu, Pengfei Wan, Ning Xu, and Kaiheng Ma. Wavelet knowledge distillation: Towards efficient image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12464-12474, June 2022. 2 +[86] Linfeng Zhang and Ma Kaisheng. Improve object detection with feature-based knowledge distillation: Towards accurate and efficient detectors. In International Conference on Learning Representations, 2021. 2, 3 +[87] Linfeng Zhang, Yukang Shi, Zuoqiang Shi, Kaisheng Ma, and Chenglong Bao. Task-oriented feature distillation. In NeurIPS, 2020. 3, 5 +[88] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In arXiv preprint: 1905.08094, 2019. 3 +[89] Linfeng Zhang, Zhanhong Tan, Jiebo Song, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Scan: A scalable neural networks framework towards compact and efficient models. ArXiv, abs/1906.03951, 2019. 3 +[90] Mingxin Zhao, Junbo Peng, Shuangming Yu, Liyuan Liu, and Nanjian Wu. Exploring structural sparsity in cnn via selective penalty. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1658-1666, 2022. 2 +[91] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. arXiv preprint arXiv:2203.05557, 2022. 3 +[92] Peng Zhou, Long Mai, Jianming Zhang, Ning Xu, Zuxuan Wu, and Larry S Davis. M2kd: Multi-model and multilevel knowledge distillation for incremental learning. arXiv preprint arXiv:1904.01769, 2019. 3 +[93] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Pix2pix datasets, http://efrosgans.eecs.berkeley.edu/cyclegan/datasets.7 +[94] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 7 \ No newline at end of file diff --git a/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/images.zip b/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b2cb2c868e4c436b2aed9f4ae49c7d3d00589add --- /dev/null +++ b/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc2079034c89023d41647bc2b0e60d06e3c51a1d9cf1f9e7419b51044f33b3e9 +size 658977 diff --git a/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/layout.json b/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..475d225fd9dd699054a5deb22978d2c3b0628f73 --- /dev/null +++ b/2023/Tiny Updater_ Towards Efficient Neural Network-Driven Software Updating/layout.json @@ -0,0 +1,12038 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 162, + 103, + 431, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 103, + 431, + 140 + ], + "spans": [ + { + "bbox": [ + 162, + 103, + 431, + 140 + ], + "type": "text", + "content": "Tiny Updater: Towards Efficient Neural Network-Driven Software Updating" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 161, + 373, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 161, + 373, + 175 + ], + "spans": [ + { + "bbox": [ + 223, + 161, + 373, + 175 + ], + "type": "text", + "content": "Linfeng Zhang, Kaisheng Ma*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 122, + 175, + 472, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 175, + 472, + 190 + ], + "spans": [ + { + "bbox": [ + 122, + 175, + 472, + 190 + ], + "type": "text", + "content": "Institute for Interdisciplinary Information Sciences, Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 191, + 457, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 191, + 457, + 202 + ], + "spans": [ + { + "bbox": [ + 136, + 191, + 457, + 202 + ], + "type": "text", + "content": "zhanglinfeng1997@outlook.com, kaisheng@mail.tsinghua.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 255, + 290, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 290, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 290, + 472 + ], + "type": "text", + "content": "Significant advancements have been accomplished with deep neural networks in diverse visual tasks, which have substantially elevated their deployment in edge device software. However, during the update of neural network-based software, users are required to download all the parameters of the neural network anew, which harms the user experience. Motivated by previous progress in model compression, we propose a novel training methodology named Tiny Updater to address this issue. Specifically, by adopting the variant of pruning and knowledge distillation methods, Tiny Updater can update the neural network-based software by only downloading a few parameters " + }, + { + "bbox": [ + 46, + 255, + 290, + 472 + ], + "type": "inline_equation", + "content": "(10\\% \\sim 20\\%)" + }, + { + "bbox": [ + 46, + 255, + 290, + 472 + ], + "type": "text", + "content": " instead of all the parameters in the neural network. Experiments on eleven datasets of three tasks, including image classification, image-to-image translation, and video recognition have demonstrated its effectiveness. Codes have been released in https://github.com/ArchipLab-LinfengZhang/TinyUpdater." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 495, + 128, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 128, + 507 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 128, + 507 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "type": "text", + "content": "With the availability of large-scale datasets [15, 16, 51] and high-performance computing platforms, deep neural networks have achieved remarkable achievements in various visual tasks such as image classification [18, 28, 75, 80], segmentation [53, 62], and object detection [50, 67, 68, 76]. Encouraged by their impressive performance, numerous software developers have effectively integrated neural networks into their software products and deployed them on edge devices such as mobile phones and tablet computers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 288, + 696 + ], + "type": "text", + "content": "Typically, the development roadmap for a neural network-based software follows the paradigm illustrate in Figure 1(a). Initially, users install the software by downloading all the neural network parameters from a cloud platform. Subsequently, as the software interacts with its users, an abundance of new training data and requirements can be" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 232, + 541, + 329 + ], + "blocks": [ + { + "bbox": [ + 312, + 232, + 541, + 329 + ], + "lines": [ + { + "bbox": [ + 312, + 232, + 541, + 329 + ], + "spans": [ + { + "bbox": [ + 312, + 232, + 541, + 329 + ], + "type": "image", + "image_path": "23d4b0134c1e078a5ef075e3a0efaec3351dc53b29617eba4311d559be81d6f1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 332, + 542, + 434 + ], + "blocks": [ + { + "bbox": [ + 312, + 332, + 542, + 434 + ], + "lines": [ + { + "bbox": [ + 312, + 332, + 542, + 434 + ], + "spans": [ + { + "bbox": [ + 312, + 332, + 542, + 434 + ], + "type": "image", + "image_path": "a41bc52d9b777ef4eab555d464f5060e5eab4385a97c04334fbd915434a9ba3c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 445, + 547, + 502 + ], + "lines": [ + { + "bbox": [ + 305, + 445, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 547, + 502 + ], + "type": "text", + "content": "Figure 1. Comparison between (a) the traditional model updating scheme and (b) the proposed efficient model updating. In efficient model updating, only around " + }, + { + "bbox": [ + 305, + 445, + 547, + 502 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 305, + 445, + 547, + 502 + ], + "type": "text", + "content": " parameters in the neural network are actually changed and thus users only require to download these changed parameters for updating." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 546, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 676 + ], + "type": "text", + "content": "collected. Then, software developers may retrain the neural network with the gathered data and update their software to enhance its performance. However, during the retraining phase, all the neural network parameters are typically changed compared to their values before updating. Consequently, users are compelled to download all the parameters of the neural network again from the cloud platform, seriously impairing their experience. While recent research has explored text prompts and adapter layers to fine-tune a large-scale pre-trained model at a low training cost [52, 79], there have been no prior efforts to reduce the download cost during model updating, which has a more direct impact on users with edge devices." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "This paper proposes the challenge of efficient model updating with the objective of reducing the download overhead of neural network-based software during updating. As" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23447" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 45, + 72, + 289, + 216 + ], + "type": "text", + "content": "depicted in Figure 1(b), during the retraining phase, efficient model updating introduces an additional constraint limiting the change of only a small subset (e.g. " + }, + { + "bbox": [ + 45, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 45, + 72, + 289, + 216 + ], + "type": "text", + "content": ") of the neural network parameters compared to the pre-updating model. Consequently, users are only required to download a few parameters that have actually changed instead of all the parameters in the neural network. In general, efficient model updating raises two questions: how to find the optimal parameters that should be changed in the neural network, and how to achieve comparable accuracy with the fully-updated model (i.e., the model which has all the parameters changed during updating)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 217, + 289, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 217, + 289, + 492 + ], + "spans": [ + { + "bbox": [ + 45, + 217, + 289, + 492 + ], + "type": "text", + "content": "To tackle this challenge, we propose a novel neural network training framework named Tiny Updater. Motivated by previous works in model compression, Tiny Updater is composed of the variants of two typical model compression techniques - neural network pruning and knowledge distillation. Firstly, to determine which channels or layers in neural networks should be modified during updating, Tiny Updater applies the pruning technique. As shown in Figure 2, it iteratively calculates the " + }, + { + "bbox": [ + 45, + 217, + 289, + 492 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 45, + 217, + 289, + 492 + ], + "type": "text", + "content": "-norm distance between the pre-updating model and the post-updating model parameters. The channels with smaller distances are deemed unnecessary for updating and are pruned to their original values before updating. Conversely, the channels with larger distances are considered essential for updating, and thus they are actually changed. Secondly, during the retraining period, we propose to improve the performance of the partially-updated model by distilling the knowledge from a fully-updated teacher model. The partially-updated student model is trained to give predictions that are similar to those of the fully-updated teacher model by optimizing the knowledge distillation loss. This ensures that the partially-updated model achieves comparable performance with the fully-updated model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 493, + 288, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 288, + 590 + ], + "type": "text", + "content": "Extensive experiments have been conducted on eleven datasets and three different tasks, including, image classification, video classification, and image-to-image translation, using seven different neural networks. Experimental results demonstrate that Tiny Updater can update the model by changing only " + }, + { + "bbox": [ + 46, + 493, + 288, + 590 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 493, + 288, + 590 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 493, + 288, + 590 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 493, + 288, + 590 + ], + "type": "text", + "content": " of the parameters with minimal performance degradation. In summary, the main contribution of this paper can be outlined as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 604, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 59, + 604, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 604, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 59, + 604, + 287, + 653 + ], + "type": "text", + "content": "- To the best of our knowledge, we first propose the challenge of efficient model updating to reduce the communication cost of downloading neural network parameters to the edge devices for software updating." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 665, + 288, + 714 + ], + "type": "text", + "content": "- To tackle this challenge, we propose Tiny Updater, which prunes the fully-updated model to the pre-updating model and distills the knowledge from the fully-updated model to the partially-updated model." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 145 + ], + "type": "text", + "content": "- Extensive experiments have been conducted on three tasks with eleven datasets, including image classification, image-to-image translation and video recognition. Experimental results demonstrate that Tiny Updater can reduce the training overhead by " + }, + { + "bbox": [ + 317, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "80\\% \\sim 90\\%" + }, + { + "bbox": [ + 317, + 72, + 545, + 145 + ], + "type": "text", + "content": " with almost no performance degradation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 156, + 392, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 392, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 392, + 169 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 176, + 446, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 176, + 446, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 176, + 446, + 190 + ], + "type": "text", + "content": "2.1. Neural Network Pruning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 195, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 195, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 195, + 545, + 470 + ], + "type": "text", + "content": "Neural network pruning is one of the most effective methods for deep neural network compression by deleting the unimportant neurons [42, 47], filters [29, 38], channels [30, 90] and layers [8, 69]. LeCun et al. and Hassibi et al. first propose to prune neural networks depending on the Hessian matrix [27, 41]. Recently, Han et al. propose deep compression, which finds the important connection in neural networks with its absolute value and then iteratively prunes them [24]. Liu et al. train a meta-network to generate the weights of neural networks and then apply it to search the best-pruned architecture [58]. Ding et al. propose to apply a LSTM to learn the hierarchical characteristics of deep neural networks and then generates the corresponding pruning scheme [17]. Liu et al. propose joint multi-dimension pruning to prune the channels, layers and resolutions at the same time [59]. Frankle et al. propose the lottery ticket hypothesis and find that a standard pruning technique naturally uncovers subnetworks whose initializations made them capable of training effectively [21]. Besides classification, neural network pruning has also be utilized in image-to-image translation [37, 42], unconditional image generation [56], object detection [22, 66], video retrieval [10] and pre-trained language models [14]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 471, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 567 + ], + "type": "text", + "content": "In this paper, Tiny Updater introduces neural network pruning methods to find the parameters which are actually valuable for model updating. Its main difference compared with previous pruning methods is that Tiny Updater prunes the parameters to their value before updating instead of zero, which makes it easier to be solved than traditional pruning. Our experiments also show that Tiny Updater can be utilized on the pruned models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 574, + 435, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 435, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 435, + 588 + ], + "type": "text", + "content": "2.2. Knowledge Distillation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Knowledge distillation, also known as student-teacher learning, has become one of the most popular deep learning techniques in various domains [6, 49, 72, 85, 86]. It firstly trains an over-parameterized teacher model and then distills teacher knowledge to a lightweight student. By training the student to mimic the prediction results of the teacher, the student can inherit the knowledge learned by the teacher. Thus, it can achieve better performance than traditional training. The concept of knowledge distillation is first proposed by Bucilua et al. [7] to compress ensemble" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23448" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 547, + 163 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 547, + 163 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 547, + 163 + ], + "type": "image", + "image_path": "3fe7b5607e3165fab270d9f037b829b89b04319d209b9fe818443eb1c5596641.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 172, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 172, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 172, + 547, + 251 + ], + "type": "text", + "content": "Figure 2. The details of Tiny Updater. (1) Tiny Updater re-trains the model with the new data collected from users (named as fully-updated models) by changing all the parameters. (2) Secondly, Tiny Updater finds the channels in the fully-updated models which have similar value to the pre-updating model, and prunes these channels to their value before updating (the obtained models are named as partially-updated models). (3) Then, Tiny Updater re-trains the unpruned weights. During re-training, the fully-updated model is utilized as the teacher model for the partially-updated model with knowledge distillation (KD) loss to improve its performance. (4) Tiny Updater repeats step-2 and step-3 until most channels have been pruned to their value before pruning. (5) Finally, users only require to download very few parameters which are actually changed for updating." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 263, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 348 + ], + "type": "text", + "content": "of neural networks, and then extended by Hinton et al. with a temperature hyper-parameter in the softmax function [32]. Following their success, abundant methods have been proposed to distill teacher knowledge in their backbone features [42, 70], spatial attention [20, 84], task-oriented information [87], pixel-wise relation [20, 44, 49, 86], sample-wise relation [63, 65, 78], prediction residual [43] and so on." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 348, + 288, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 288, + 552 + ], + "type": "text", + "content": "Previous knowledge distillation methods usually transfer teacher knowledge in the categorical probability distribution. Then, abundant methods have been introduced to distill the knowledge in the teacher feature [70] and its variants, such as attention [84, 86], relational information [63, 78], self-supervised knowledge [81] and task-oriented information [87]. Besides distilling knowledge from a large teacher to a tiny student, there have also been fruitful knowledge distillation methods that distill knowledge from deeper layers to shallow layers [88, 89], from all the channels to partial channels [83], from ensemble classifiers to single classifier [40], from multiple frames to few frames [5], from RGB images to RGB-D images [23]. In this paper, we introduce knowledge distillation by distilling the knowledge from a teacher model, which has all the parameters changed during updating to a student model, which is only updated by changing a few parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 562, + 173, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 173, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 173, + 575 + ], + "type": "text", + "content": "2.3. Incremental Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Inspired by the observation that human beings can incrementally learn knowledge about new tasks and categories without forgetting the old knowledge, incremental learning (a.k.a. lifelong learning) is proposed to give the neural networks the similar ability. Elastic weight consolidation (EWC) is proposed to maintain the knowledge of old tasks and categories by first estimating the importance of each neuron with Bays estimation or Fisher information matrix and then training them with a consolidation constraint [39, 55, 77]. Then, memory aware synapse [1] is proposed to compute the importance score of neural net" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 264, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 384 + ], + "type": "text", + "content": "work parameters in an unsupervised and online manner, which is also similar to the Hebbian learning in biological systems [31]. Besides these consolidation-based methods, knowledge distillation methods have also been utilized in incremental learning to maintain the old knowledge by learning its response to new tasks [13, 48, 92]. Besides image classification, abundant recent research has applied incremental learning to the other visual tasks, such as object detection [11, 26], action recognition [60] and imaged-to-image translation [74]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 384, + 546, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 546, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 546, + 564 + ], + "type": "text", + "content": "Both the proposed efficient model updating and incremental learning expect the neural network to maintain its knowledge of old parameters. Their main difference is that: (a) Incremental learning does not limit the change in old parameters strictly, while efficient model updating has a direct constraint on the number of changed parameters to reduce the download overhead for edge devices. (b) Incremental learning usually assumes that the data for old knowledge is not available while efficient model updating can still access data, which is more practical to industrial applications. (c) The target of incremental learning is to build human-like artificial intelligence which can incrementally learn various tasks, while the target of Efficient Model Updating is to reduce the number of parameters updated when new data for the same task is collected." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 574, + 509, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 509, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 509, + 588 + ], + "type": "text", + "content": "2.4. Efficient Pretrained Model Finetuning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Large-scale pretrained models recently have shown powerful representation ability in both natural language processing and computer vision. However, since these pretrained models usually have billions of parameters, directly finetuning them on downstream tasks usually suffers from massive training overhead, making them impractical in real-world applications. Recently, prompts methods have been proposed to address this problem by designing specific input templates for the specific downstream tasks [4,25,45,52, 54,79,91]. Moreover, adapter methods have also been intro-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23449" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": "duced to finetune several additional trainable layers instead of the whole pre-trained model [3, 33]. These methods are firstly proposed for pre-trained language models and then extended to language-vision multi-modal models and vision models [12, 36, 82]. Both these efficient finetuning methods and the proposed efficient model updating aim to freeze the parameters of neural networks. Their main difference is that: (a) Efficient finetuning is usually applied to the models which are firstly pre-trained on large-scale datasets to learn task-unbiased knowledge and then finetuned for specific downstream tasks. Instead, the pre-updating models in Efficient Model Updating are trained with very few training samples (e.g. only " + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 47, + 72, + 289, + 396 + ], + "type": "text", + "content": " in our experiments) and then updated with the new data collected from users. (b) The target of efficient finetuning is to reduce the training overhead of pre-trained models and improve their performance in downstream tasks. In contrast, the target of Efficient Model Updating is to reduce the download overhead during software updating. (c) Besides, adapter-based methods can be considered as a special case in Tiny Updater where all the parameters of backbone layers are frozen and all the adapter layers are updated. Our experiments also demonstrate that Tiny Updater achieves clearly better performance than directly applying efficient model updating methods. Please refer to Appendix A for a more detailed comparison among the proposed efficient model updating, incremental learning, and efficient pretrained model finetuning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 410, + 130, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 130, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 130, + 423 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 430, + 189, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 189, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 189, + 443 + ], + "type": "text", + "content": "3.1. Efficient Model Updating" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": "Given a training dataset " + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x_1, y_1), \\dots, (x_n, y_n)\\}" + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": ", the software developers firstly train a deep neural network " + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": " for their software. After deploying " + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": " on the edge devices, abundant new training samples " + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{(x_{n+1}, y_{n+1}), \\dots, (x_{n+m}, y_{n+m})\\}" + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": " can be collected from users and the dataset can be extended as " + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^+ = \\mathcal{D} \\cup \\{(x_{n+1}, y_{n+1}), \\dots, (x_{n+m}, y_{n+m})\\}" + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": ". Then, software developers can re-train " + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^+" + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": " to obtain better performance, whose parameters can be denoted as " + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": ". Usually, during this re-training step, all the values in " + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": " can be totally different from that in " + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": ". Thus, the users have to download all the parameters of " + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 47, + 450, + 289, + 715 + ], + "type": "text", + "content": " to update the software, which harms the user experience and limits the frequently updating of neural network-driven software. In such kind of paradigm, we name the model before updating as the pre-updating model, the model after updating as the post-updating model, the model with all the parameters updated as the fully-updated model, and the model with partial parameters updated as the partially updated model. The target of efficient model updating is to obtain a partially updated model which has most of the parameters unchanged while achieving similar performance to the fully-updated model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 72, + 476, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 476, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 476, + 84 + ], + "type": "text", + "content": "Algorithm 1 The proposed Tiny Updater." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 86, + 545, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 86, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 305, + 86, + 545, + 121 + ], + "type": "text", + "content": "Input: Dataset " + }, + { + "bbox": [ + 305, + 86, + 545, + 121 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{+} = \\{(x_{1},y_{1}),\\dots,(x_{n + m},y_{n + m})\\}" + }, + { + "bbox": [ + 305, + 86, + 545, + 121 + ], + "type": "text", + "content": ", the pre-updating model " + }, + { + "bbox": [ + 305, + 86, + 545, + 121 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 86, + 545, + 121 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 305, + 86, + 545, + 121 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 305, + 86, + 545, + 121 + ], + "type": "text", + "content": ", an expected ratio of updated parameters " + }, + { + "bbox": [ + 305, + 86, + 545, + 121 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 86, + 545, + 121 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 122, + 544, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 122, + 544, + 146 + ], + "spans": [ + { + "bbox": [ + 305, + 122, + 544, + 146 + ], + "type": "text", + "content": "Output: The partially-updated model " + }, + { + "bbox": [ + 305, + 122, + 544, + 146 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 122, + 544, + 146 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 305, + 122, + 544, + 146 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 305, + 122, + 544, + 146 + ], + "type": "text", + "content": ". // Update the model with " + }, + { + "bbox": [ + 305, + 122, + 544, + 146 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^+" + }, + { + "bbox": [ + 305, + 122, + 544, + 146 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 147, + 544, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 147, + 544, + 170 + ], + "spans": [ + { + "bbox": [ + 306, + 147, + 544, + 170 + ], + "type": "text", + "content": "Initialize the partially-updated model " + }, + { + "bbox": [ + 306, + 147, + 544, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 306, + 147, + 544, + 170 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 306, + 147, + 544, + 170 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 306, + 147, + 544, + 170 + ], + "type": "text", + "content": ". while " + }, + { + "bbox": [ + 306, + 147, + 544, + 170 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\Phi}" + }, + { + "bbox": [ + 306, + 147, + 544, + 170 + ], + "type": "text", + "content": " is not converged do:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 331, + 170, + 473, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 170, + 473, + 181 + ], + "spans": [ + { + "bbox": [ + 331, + 170, + 473, + 181 + ], + "type": "text", + "content": "Sample a batch of data " + }, + { + "bbox": [ + 331, + 170, + 473, + 181 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 331, + 170, + 473, + 181 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 331, + 170, + 473, + 181 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^+" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 331, + 182, + 424, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 182, + 424, + 194 + ], + "spans": [ + { + "bbox": [ + 331, + 182, + 424, + 194 + ], + "type": "text", + "content": "Compute " + }, + { + "bbox": [ + 331, + 182, + 424, + 194 + ], + "type": "inline_equation", + "content": "\\hat{y} \\coloneqq \\mathcal{F}_{\\Phi}(\\mathcal{X})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 331, + 194, + 491, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 194, + 491, + 206 + ], + "spans": [ + { + "bbox": [ + 331, + 194, + 491, + 206 + ], + "type": "text", + "content": "Compute the task loss between " + }, + { + "bbox": [ + 331, + 194, + 491, + 206 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 331, + 194, + 491, + 206 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 331, + 194, + 491, + 206 + ], + "type": "inline_equation", + "content": "\\hat{y}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 207, + 492, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 207, + 492, + 218 + ], + "spans": [ + { + "bbox": [ + 331, + 207, + 492, + 218 + ], + "type": "text", + "content": "Back propagate gradients and update " + }, + { + "bbox": [ + 331, + 207, + 492, + 218 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 331, + 207, + 492, + 218 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 219, + 542, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 219, + 542, + 229 + ], + "spans": [ + { + "bbox": [ + 306, + 219, + 542, + 229 + ], + "type": "text", + "content": "//Pruning, and re-training with KD loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 230, + 545, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 230, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 306, + 230, + 545, + 254 + ], + "type": "text", + "content": "Initialize the parameters of the teacher " + }, + { + "bbox": [ + 306, + 230, + 545, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\coloneqq \\Phi" + }, + { + "bbox": [ + 306, + 230, + 545, + 254 + ], + "type": "text", + "content": ", an index set of pruned weights as " + }, + { + "bbox": [ + 306, + 230, + 545, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{\\}" + }, + { + "bbox": [ + 306, + 230, + 545, + 254 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 255, + 415, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 255, + 415, + 269 + ], + "spans": [ + { + "bbox": [ + 306, + 255, + 415, + 269 + ], + "type": "text", + "content": "while " + }, + { + "bbox": [ + 306, + 255, + 415, + 269 + ], + "type": "inline_equation", + "content": "\\frac{Card(\\mathcal{T})}{Card(\\Theta)} < 1 - \\tau" + }, + { + "bbox": [ + 306, + 255, + 415, + 269 + ], + "type": "text", + "content": " do:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 331, + 269, + 529, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 269, + 529, + 280 + ], + "spans": [ + { + "bbox": [ + 331, + 269, + 529, + 280 + ], + "type": "text", + "content": "Compute the " + }, + { + "bbox": [ + 331, + 269, + 529, + 280 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 331, + 269, + 529, + 280 + ], + "type": "text", + "content": "-norm distance between " + }, + { + "bbox": [ + 331, + 269, + 529, + 280 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 331, + 269, + 529, + 280 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 331, + 269, + 529, + 280 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 331, + 269, + 529, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 331, + 281, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 281, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 331, + 281, + 545, + 303 + ], + "type": "text", + "content": "Append the indices of channels which have relatively smaller distance to " + }, + { + "bbox": [ + 331, + 281, + 545, + 303 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 331, + 281, + 545, + 303 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 331, + 304, + 501, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 304, + 501, + 316 + ], + "spans": [ + { + "bbox": [ + 331, + 304, + 501, + 316 + ], + "type": "inline_equation", + "content": "\\Phi [\\mathcal{I}]\\coloneqq \\Theta [\\mathcal{I}]." + }, + { + "bbox": [ + 331, + 304, + 501, + 316 + ], + "type": "text", + "content": " //Prune " + }, + { + "bbox": [ + 331, + 304, + 501, + 316 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 331, + 304, + 501, + 316 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 331, + 304, + 501, + 316 + ], + "type": "inline_equation", + "content": "\\Theta" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 331, + 317, + 459, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 317, + 459, + 328 + ], + "spans": [ + { + "bbox": [ + 331, + 317, + 459, + 328 + ], + "type": "text", + "content": "while " + }, + { + "bbox": [ + 331, + 317, + 459, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\Phi}" + }, + { + "bbox": [ + 331, + 317, + 459, + 328 + ], + "type": "text", + "content": " is not converged do:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 355, + 329, + 497, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 329, + 497, + 339 + ], + "spans": [ + { + "bbox": [ + 355, + 329, + 497, + 339 + ], + "type": "text", + "content": "Sample a batch of data " + }, + { + "bbox": [ + 355, + 329, + 497, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 355, + 329, + 497, + 339 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 355, + 329, + 497, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^+" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 355, + 340, + 522, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 340, + 522, + 352 + ], + "spans": [ + { + "bbox": [ + 355, + 340, + 522, + 352 + ], + "type": "text", + "content": "Compute " + }, + { + "bbox": [ + 355, + 340, + 522, + 352 + ], + "type": "inline_equation", + "content": "\\hat{y} \\coloneqq \\mathcal{F}_{\\Phi}(\\mathcal{X})" + }, + { + "bbox": [ + 355, + 340, + 522, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 355, + 340, + 522, + 352 + ], + "type": "inline_equation", + "content": "y_{t} \\coloneqq \\mathcal{F}_{\\mathcal{T}}(\\mathcal{X})" + }, + { + "bbox": [ + 355, + 340, + 522, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 355, + 353, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 353, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 355, + 353, + 545, + 376 + ], + "type": "text", + "content": "Compute the task loss between " + }, + { + "bbox": [ + 355, + 353, + 545, + 376 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 355, + 353, + 545, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 355, + 353, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 355, + 353, + 545, + 376 + ], + "type": "text", + "content": ", and KD loss between " + }, + { + "bbox": [ + 355, + 353, + 545, + 376 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 355, + 353, + 545, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 355, + 353, + 545, + 376 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 355, + 353, + 545, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 355, + 376, + 518, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 376, + 518, + 399 + ], + "spans": [ + { + "bbox": [ + 355, + 376, + 518, + 399 + ], + "type": "text", + "content": "Back propagate gradients and update the parameters of " + }, + { + "bbox": [ + 355, + 376, + 518, + 399 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 355, + 376, + 518, + 399 + ], + "type": "text", + "content": " which are not in " + }, + { + "bbox": [ + 355, + 376, + 518, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 355, + 376, + 518, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 400, + 539, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 400, + 539, + 412 + ], + "spans": [ + { + "bbox": [ + 306, + 400, + 539, + 412 + ], + "type": "text", + "content": "Return The partially-updated model " + }, + { + "bbox": [ + 306, + 400, + 539, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 306, + 400, + 539, + 412 + ], + "type": "text", + "content": " with parameter " + }, + { + "bbox": [ + 306, + 400, + 539, + 412 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 306, + 400, + 539, + 412 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 432, + 392, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 392, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 392, + 445 + ], + "type": "text", + "content": "3.2. Tiny Updater" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 451, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 533 + ], + "type": "text", + "content": "Fruitful previous works in model compression have successfully proven that even a very tiny neural network can have powerful representation ability, which motivates us to propose to learn the knowledge in the collected training data " + }, + { + "bbox": [ + 304, + 451, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{+}" + }, + { + "bbox": [ + 304, + 451, + 545, + 533 + ], + "type": "text", + "content": " with only a few parameters instead of all the parameters. The optimization objective of Tiny Updater can be formulated as" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 358, + 540, + 545, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 540, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 358, + 540, + 545, + 600 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\arg \\min _ {\\Phi} \\frac {1}{n + m} \\sum_ {i = 1} ^ {n + m} \\mathcal {L} _ {\\text {t a s k}} \\left(x _ {i}, y _ {i}\\right) \\tag {1} \\\\ \\text {s u b j e c t} \\quad \\frac {\\left| \\Theta - \\Phi \\right| _ {0}}{\\operatorname {C a r d} (\\Theta)} < \\tau \\\\ \\end{array}", + "image_path": "5de2d9971af2fc77358e9dbc0984237a99fac1cb9f3edee18fe0d8e271b80db7.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{task}}" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " indicates the original task-specific loss function, such as cross-entropy loss for image classification. " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "|\\cdot|_0" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " indicates the " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "L_0" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "-norm, which measures the number of non-zero elements in a tensor. " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\operatorname{Card}(\\cdot)" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " denotes the cardinality, which describes the number of parameters in a tensor. " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " is a ratio threshold that determines how many parameters should be changed during updating. It is observed that when " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " in Equation (1) becomes zero, the optimization objective is similar to another deep learning technique - network pruning." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23450" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 77, + 192, + 167 + ], + "blocks": [ + { + "bbox": [ + 62, + 116, + 78, + 126 + ], + "lines": [ + { + "bbox": [ + 62, + 116, + 78, + 126 + ], + "spans": [ + { + "bbox": [ + 62, + 116, + 78, + 126 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 77, + 192, + 167 + ], + "lines": [ + { + "bbox": [ + 82, + 77, + 192, + 167 + ], + "spans": [ + { + "bbox": [ + 82, + 77, + 192, + 167 + ], + "type": "image", + "image_path": "b36fefbdb4b5e217ed12a10095a4d2a9a10435740ee707ed084c273a3e73c389.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 198, + 77, + 305, + 167 + ], + "blocks": [ + { + "bbox": [ + 198, + 77, + 305, + 167 + ], + "lines": [ + { + "bbox": [ + 198, + 77, + 305, + 167 + ], + "spans": [ + { + "bbox": [ + 198, + 77, + 305, + 167 + ], + "type": "image", + "image_path": "b290562004e043f34a63d1a885afa74aadc369dcca042e093c3928fc18272e6d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 77, + 414, + 167 + ], + "blocks": [ + { + "bbox": [ + 307, + 77, + 414, + 167 + ], + "lines": [ + { + "bbox": [ + 307, + 77, + 414, + 167 + ], + "spans": [ + { + "bbox": [ + 307, + 77, + 414, + 167 + ], + "type": "image", + "image_path": "9f8ff91c52461d6f7c1ff64c941ffe7ee8d3f2d9d1e43472e0fe3c9e1d36c0af.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 417, + 78, + 529, + 167 + ], + "blocks": [ + { + "bbox": [ + 417, + 78, + 529, + 167 + ], + "lines": [ + { + "bbox": [ + 417, + 78, + 529, + 167 + ], + "spans": [ + { + "bbox": [ + 417, + 78, + 529, + 167 + ], + "type": "image", + "image_path": "aa3dd3e6487db150fe7d07f893ea9e74a0e125e65a686fb26510f79ae2bb2cdd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 79, + 175, + 192, + 274 + ], + "blocks": [ + { + "bbox": [ + 62, + 213, + 78, + 224 + ], + "lines": [ + { + "bbox": [ + 62, + 213, + 78, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 78, + 224 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 79, + 175, + 192, + 274 + ], + "lines": [ + { + "bbox": [ + 79, + 175, + 192, + 274 + ], + "spans": [ + { + "bbox": [ + 79, + 175, + 192, + 274 + ], + "type": "image", + "image_path": "897b7e72e728fc446de4342db63d21a94b4f8dcec205374df982c35aa04d41da.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 202, + 175, + 304, + 274 + ], + "blocks": [ + { + "bbox": [ + 202, + 175, + 304, + 274 + ], + "lines": [ + { + "bbox": [ + 202, + 175, + 304, + 274 + ], + "spans": [ + { + "bbox": [ + 202, + 175, + 304, + 274 + ], + "type": "image", + "image_path": "36c1f9f04cc5c2bd3d34850338c8a67f6663b499c237529e1bf9a6b06854abd8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 279, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 279, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 545, + 313 + ], + "type": "text", + "content": "Figure 3. Experiments on four fine-grained image classification datasets with ResNet50. The pre-updating and the post-updating models are trained with " + }, + { + "bbox": [ + 46, + 279, + 545, + 313 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 279, + 545, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 279, + 545, + 313 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 279, + 545, + 313 + ], + "type": "text", + "content": " training data, respectively. In subfigure (a), the newly collected images come from all categories uniformly. In subfigure (b), the newly collected images come from only the categories that are not available before updating." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 175, + 417, + 274 + ], + "blocks": [ + { + "bbox": [ + 315, + 175, + 417, + 274 + ], + "lines": [ + { + "bbox": [ + 315, + 175, + 417, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 175, + 417, + 274 + ], + "type": "image", + "image_path": "c1ed06d9070221be0875fc09117adf5422bbfa3a0602904b55af8b75d6943d44.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 425, + 175, + 527, + 274 + ], + "blocks": [ + { + "bbox": [ + 425, + 175, + 527, + 274 + ], + "lines": [ + { + "bbox": [ + 425, + 175, + 527, + 274 + ], + "spans": [ + { + "bbox": [ + 425, + 175, + 527, + 274 + ], + "type": "image", + "image_path": "daf44520f2e6b57347a837ee249401eefd5dfc9e4a10e89029570eeff551cca2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 323, + 286, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 286, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 286, + 432 + ], + "type": "text", + "content": "ing. Thus, we propose to solve this problem by introducing iterative pruning methods. Besides, knowledge distillation is also introduced to improve the model performance with only a few parameters changed during updating with the teacher from a model with all the parameters changed. Please note that Tiny Updater is a framework that applies pruning and knowledge distillation methods to tackle the challenge of efficient model updating and it does not propose any new pruning and knowledge distillation method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "text", + "content": "The details of Tiny Updater are shown in Algorithm 1. With the extended dataset " + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^+" + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "text", + "content": ", it first trains a fully-updated model " + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\Phi}" + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "text", + "content": " without any constraints and copies its parameters to another neural network " + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{T}}" + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "text", + "content": ", which is considered as the teacher network. Then, channel pruning is utilized to find the parameters in " + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "text", + "content": " (e.g. channels in convolutional layers) whose value is not significantly changed compared with the pre-updating model " + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "text", + "content": ". Then, these parameters are reset to their value in " + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "text", + "content": " and are not further trained in the later re-training. During the re-training period, the partially-updated model is trained with not only the origin task loss but also the knowledge distillation loss from the teacher of the fully-updated model " + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{T}}" + }, + { + "bbox": [ + 46, + 432, + 286, + 635 + ], + "type": "text", + "content": ". Please note that the proposed Tiny Updater is a framework that employs pruning and knowledge distillation to tackle the challenge of efficient model updating and it does not depend on any specific pruning and knowledge distillation methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 649, + 123, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 649, + 123, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 649, + 123, + 662 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 670, + 164, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 164, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 164, + 682 + ], + "type": "text", + "content": "4.1. Image Classification" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Experiment Setting Our method has mainly been evaluated on seven image classification datasets, including CI-" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 307, + 357, + 545, + 533 + ], + "blocks": [ + { + "bbox": [ + 305, + 320, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 305, + 320, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 320, + 545, + 354 + ], + "type": "text", + "content": "Table 1. The performance of Tiny Updater on ImageNet. Models are firstly trained with " + }, + { + "bbox": [ + 305, + 320, + 545, + 354 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 305, + 320, + 545, + 354 + ], + "type": "text", + "content": " data and then updated with all the data. \"Updated Ratio\" indicates the ratio of updated parameters." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 357, + 545, + 533 + ], + "lines": [ + { + "bbox": [ + 307, + 357, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 357, + 545, + 533 + ], + "type": "table", + "html": "
ModelUpdated Ratio (%)Top-1 AccuracyTop-5 Accuracy
ResNet5000.0068.86 -6.4488.74 -3.46
21.5474.76 -0.5491.70 -0.50
42.4275.12 -0.1891.93 -0.27
65.6976.25 +0.9593.14 -0.94
100.075.3092.20
MobileNetv200.0062.23 -3.0884.25 -2.29
24.3164.54 -0.7785.53 -1.01
31.6665.23 -0.0886.42 -0.12
66.5566.30 +0.9987.19 +0.65
87.6066.86 +1.5587.50 +0.96
100.065.3186.54
Swin Transformer00.0072.04 -9.2689.75 -5.85
18.5679.91 -1.3994.33 -1.27
36.3180.87 -0.4394.53 -1.07
80.2681.73 +0.4395.74 +0.14
100.081.3095.60
", + "image_path": "3eebe9df91662db65fc75f5342fd761d67879a6bd554412520aeccf187d35fa0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 552, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 685 + ], + "type": "text", + "content": "FAR10, CIFAR100, and ImageNet for general image classification, Stanford Car, Oxford Flower, Stanford Dog, FGVC Aircraft for fine-grained image classification. Please refer to Appendix B for the details of these datasets. Note that on the fine-grained image classification datasets, models are initialized with backbone weights pre-trained on ImageNet. In ImageNet experiments, we adopt the basic training policy from PyTorch [64] with task-oriented feature distillation [87]. On the other datasets, each model is trained by 200 epochs with the naive logit and feature knowledge distillation loss [32, 70]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Main Results The performance of Tiny Updater on ImageNet, CIFAR10&100, and four fine-grained image classifi-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23451" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 278, + 145 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 278, + 145 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 278, + 145 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 278, + 145 + ], + "type": "image", + "image_path": "cb65524168a6ad9ded1610690e13f54cc0a722db902d46be70af15da0dc69208.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 162, + 287, + 207 + ], + "lines": [ + { + "bbox": [ + 46, + 162, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 287, + 207 + ], + "type": "text", + "content": "Figure 4. Experiments on CIFAR with ResNet50. The pre-updating models and fully-updated models are trained with " + }, + { + "bbox": [ + 46, + 162, + 287, + 207 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 162, + 287, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 162, + 287, + 207 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 162, + 287, + 207 + ], + "type": "text", + "content": " training data, respectively. Tiny Updater with random pruning indicates pruning the randomly selected channels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 226, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 289, + 467 + ], + "type": "text", + "content": "cation are shown in Table 1, Figure 4, Figure 3, respectively. It is observed that: (i) Tiny Updater achieves consistent effectiveness on all seven datasets. By changing around " + }, + { + "bbox": [ + 46, + 226, + 289, + 467 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 226, + 289, + 467 + ], + "type": "text", + "content": " parameters during updating, the neural networks trained with Tiny Updater achieve a similar performance with fully-updated models on all these datasets. (ii) On ImageNet, Tiny Updater achieves significant performance on the regular convolutional network such as ResNet [28], the efficient convolutional networks such as MobileNetv2 [71], and Swin Transformer [57]. (iii) With the proposed Tiny Updater, when a large ratio of parameters is changed, it even leads to higher performance than the fully-updated model. For instance, on Oxford Flowers datasets in Figure 3, the " + }, + { + "bbox": [ + 46, + 226, + 289, + 467 + ], + "type": "inline_equation", + "content": "84\\%" + }, + { + "bbox": [ + 46, + 226, + 289, + 467 + ], + "type": "text", + "content": " updated model trained with Tiny Updater has " + }, + { + "bbox": [ + 46, + 226, + 289, + 467 + ], + "type": "inline_equation", + "content": "0.25\\%" + }, + { + "bbox": [ + 46, + 226, + 289, + 467 + ], + "type": "text", + "content": " higher accuracy than the fully-updated model. We suggest this accuracy benefit is caused by the knowledge distillation loss in Tiny Updater. As pointed out by some previous research [2, 61], even if the student and the teacher have similar performance, knowledge distillation can still lead to consistent accuracy benefits." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "text", + "content": "Ablation Study Tiny Updater is mainly composed of two modules - pruning and knowledge distillation. Ablation studies on CIFAR10 and CIFAR100 are shown in Figure 4. It is observed that: (i) Compared with Tiny Updater with random pruning, Tiny Updater with " + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "text", + "content": "-norm based pruning leads to consistently higher accuracy (around " + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "text", + "content": ") especially when only a low ratio (" + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "inline_equation", + "content": "0 \\sim 50\\%" + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "text", + "content": ") of parameters are changed, indicating " + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "text", + "content": "-norm is an effective metric to find which channels are actually important for model updating. (ii) Significant accuracy boosts can be observed by applying knowledge distillation during re-training. For instance, on CIFAR100, when around " + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "text", + "content": " parameters are changed, " + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "inline_equation", + "content": "1.79\\%" + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "inline_equation", + "content": "1.15\\%" + }, + { + "bbox": [ + 46, + 475, + 287, + 668 + ], + "type": "text", + "content": " accuracy boost can be observed with knowledge distillation, respectively. These observations indicate that the pruning and knowledge distillation in Tiny Updater have their individual effectiveness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Categorical Incremental Updating Figure 3(b) shows the effectiveness of Tiny Updater in the categorical incremental settings, where the pre-updating models are trained" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 374, + 72, + 474, + 152 + ], + "blocks": [ + { + "bbox": [ + 374, + 72, + 474, + 152 + ], + "lines": [ + { + "bbox": [ + 374, + 72, + 474, + 152 + ], + "spans": [ + { + "bbox": [ + 374, + 72, + 474, + 152 + ], + "type": "image", + "image_path": "6a2a06b3d9b663ad5b6aa4b406f336cab67aebff51b587193133caea518124dc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 158, + 545, + 179 + ], + "lines": [ + { + "bbox": [ + 305, + 158, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 305, + 158, + 545, + 179 + ], + "type": "text", + "content": "Figure 5. Experiments on CIFAR10 with the pre-updating models trained with different ratios of data." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 198, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 270 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 304, + 198, + 545, + 270 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 198, + 545, + 270 + ], + "type": "text", + "content": " training data belonging to " + }, + { + "bbox": [ + 304, + 198, + 545, + 270 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 198, + 545, + 270 + ], + "type": "text", + "content": " categories. During the updating period, the model is further trained with data of all the categories. It is observed that Tiny Updater can still achieve significant performance by only changing around " + }, + { + "bbox": [ + 304, + 198, + 545, + 270 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 198, + 545, + 270 + ], + "type": "text", + "content": " parameters, indicating that Tiny Updater is also effective on the categorical incremental updating setting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "text", + "content": "Influence from Pre-updating Model In previous experiments, the pre-updating models are trained with " + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "text", + "content": " data. With less training data, the pre-updating models tend to have less representative ability, and thus the performance of Tiny Updater tends to be reduced. In this subsection, we study how the performance of the pre-updating model influence Tiny Updater on CIFAR10 in Figure 3(a). It is observed that when most of the parameters " + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "inline_equation", + "content": "(>50\\%)" + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "text", + "content": " are changed during updating, the accuracy gap between different pre-updating models is not significant " + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "inline_equation", + "content": "(< 0.2\\%)" + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "text", + "content": ". In contrast, when only a few parameters " + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "inline_equation", + "content": "(10\\% \\sim 20\\%)" + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "text", + "content": " are changed, the ratio of data for training pre-updating models has a more significant influence. For example, compared with the pre-updating model trained with " + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "text", + "content": " data, the pre-updating model trained with only " + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "text", + "content": " data leads to around " + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "text", + "content": " accuracy loss when " + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 278, + 546, + 542 + ], + "type": "text", + "content": " parameters are changed during updating, indicating that the performance of the pre-updating model has a direct influence to Tiny Updater. Specifically, when the pre-updating model is not trained with any data, it can be considered as a model with all parameters initialized with zero matrices, and thus Tiny Updater in this case degenerates to the common neural network pruning problem." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "content": "Multi-step Model Updating Previous experiments mainly show the result of one-step updating from using " + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "content": " training data. In this subsection, we show the performance of Tiny Updater in a multi-step updating, where the model is first trained with " + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "content": " data and then updated with " + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "content": " training data, successively. As shown in Table 4, there is almost no accuracy loss when around " + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "content": " parameters are changed for each update. When there are around " + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "content": " parameters changed for each update, accuracy loss becomes significant with the increment on the update iterations (from " + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "inline_equation", + "content": "-0.44\\%" + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "inline_equation", + "content": "-2.33\\%" + }, + { + "bbox": [ + 304, + 549, + 547, + 717 + ], + "type": "text", + "content": "). This observation indicates that the accuracy loss tends to be accumulated in multi-step updating when an extremely low ratio of parameters is changed for updating." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23452" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 118, + 287, + 172 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "text", + "content": "Table 2. Experiments of multi-step updating on CIFAR10. Models are firstly trained with " + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "text", + "content": " training data and then updated by five times with " + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "text", + "content": " training data by changing all, " + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 47, + 69, + 287, + 113 + ], + "type": "text", + "content": " parameters." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "table", + "html": "
Data Ratio \nParams. Ratio5%10%15%20%25%100%
Fully-Updated46.564.278.884.687.895.3
10% Params. Updated46.563.878.383.285.992.9
20% Params. Updated46.564.379.084.287.594.8
", + "image_path": "a9892a5843b6366d0ca6cde2a9b351e9c72cc6a5f78f27771b4e85b0a3cf4993.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": "Tiny Updater on Pruned Models In this subsection, we study whether the proposed Tiny Updater can be utilized on a pruned model. Concretely, we first train a ResNet18 model with " + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": " training data on CIFAR10, which achieves " + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "inline_equation", + "content": "87.43\\%" + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": " top-1 accuracy. Then, we apply the " + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": "-norm pruning technique to prune " + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "inline_equation", + "content": "53\\%" + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": " neurons at the expense of " + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "inline_equation", + "content": "0.37\\%" + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": " accuracy loss " + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "inline_equation", + "content": "(87.43\\% \\rightarrow 87.06\\%)" + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": ". Thirdly, the proposed Tiny Updater is applied to update this model with " + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": " training data, which achieves " + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "inline_equation", + "content": "94.22\\%" + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": " at the expense of changing " + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "inline_equation", + "content": "24.3\\%" + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": " parameters " + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "inline_equation", + "content": "(87.06\\% \\rightarrow 94.22\\%)" + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": ". In contrast, a pruned fully-updated model with the same pruning ratio achieves " + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "inline_equation", + "content": "94.87\\%" + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": " accuracy, which is only " + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "inline_equation", + "content": "0.65\\%" + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": " higher than the pruned model with Tiny Updater " + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "inline_equation", + "content": "(87.06\\% \\rightarrow 94.87\\%)" + }, + { + "bbox": [ + 46, + 194, + 288, + 373 + ], + "type": "text", + "content": ". These observations indicate that the proposed Tiny Updater can be utilized on pruned models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 385, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 506 + ], + "type": "text", + "content": "Comparison with Finetuning and Adapter Finetuning and adapter-based methods are two well-known efficient training methods and the parameters changed in these two methods are also much fewer than global finetuning. Figure 4 gives the comparison between Tiny Updater and these two methods on CIFAR10, CIFAR100 with ResNet18. It is observed that when the same number of parameters are changed, Tiny Updater outperforms these two methods by a clear margin, indicating that Tiny Updater is more effective than directly applying previous methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 514, + 257, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 257, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 257, + 528 + ], + "type": "text", + "content": "4.2. Image-to-Image Translation with GANs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": "Experiment Setting We mainly evaluate Tiny Up- dater on image-to-image translation with CycleGAN for Horse " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " Zebra translation, and Pix2Pix for Edge " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " Shoe translation. CycleGAN is a typical unpaired image-toimage translation model, which has a ResNet-based generator trained with GAN loss and cycle consistency [94]. Pix2Pix is a typical paired image-to-image translation model, which has a U-Net architecture generator trained with conditional GAN loss and the " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " -norm loss [35]. Horse " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " Zebra is an unpaired dataset that translates natural images of horses to zebras and vice versa. It consists of 1,187 horse images and 1,474 zebra [93]. Edge " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " Shoe is a paired data that translates the edges of shoes to their corresponding natural images [34]. Fréchet Inception Distance (FID), which measures the distance between the feature dis" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 69, + 541, + 209 + ], + "blocks": [ + { + "bbox": [ + 306, + 69, + 541, + 209 + ], + "lines": [ + { + "bbox": [ + 306, + 69, + 541, + 209 + ], + "spans": [ + { + "bbox": [ + 306, + 69, + 541, + 209 + ], + "type": "image", + "image_path": "fda6bd8009aee50fb7cdd3009802e081d435b992fca883c2ee4f7fac4763dbd4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 217, + 545, + 260 + ], + "lines": [ + { + "bbox": [ + 305, + 217, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 545, + 260 + ], + "type": "text", + "content": "Figure 6. Qualitative analysis on Horse " + }, + { + "bbox": [ + 305, + 217, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 305, + 217, + 545, + 260 + ], + "type": "text", + "content": " Zebra and Zebra " + }, + { + "bbox": [ + 305, + 217, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 305, + 217, + 545, + 260 + ], + "type": "text", + "content": " Horse with CycleGAN. The partially-updated model trained with Tiny Updater has around " + }, + { + "bbox": [ + 305, + 217, + 545, + 260 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 305, + 217, + 545, + 260 + ], + "type": "text", + "content": " parameters changed. The pre-updating models are trained with " + }, + { + "bbox": [ + 305, + 217, + 545, + 260 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 305, + 217, + 545, + 260 + ], + "type": "text", + "content": " data." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 269, + 543, + 334 + ], + "blocks": [ + { + "bbox": [ + 309, + 269, + 543, + 334 + ], + "lines": [ + { + "bbox": [ + 309, + 269, + 543, + 334 + ], + "spans": [ + { + "bbox": [ + 309, + 269, + 543, + 334 + ], + "type": "image", + "image_path": "a0abb3c0873994993d3a63401fe5271041001d7e4038ecbdbbf9983b9ab5268c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 337, + 545, + 382 + ], + "lines": [ + { + "bbox": [ + 305, + 337, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 305, + 337, + 545, + 382 + ], + "type": "text", + "content": "Figure 7. Experiments of image-to-image translation on Horse " + }, + { + "bbox": [ + 305, + 337, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 305, + 337, + 545, + 382 + ], + "type": "text", + "content": " Zebra and Zebra " + }, + { + "bbox": [ + 305, + 337, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 305, + 337, + 545, + 382 + ], + "type": "text", + "content": " Horse with CycleGAN, and Edge " + }, + { + "bbox": [ + 305, + 337, + 545, + 382 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 305, + 337, + 545, + 382 + ], + "type": "text", + "content": " Shoe with Pix2Pix. In all these experiments, the pre-updating models are trained with " + }, + { + "bbox": [ + 305, + 337, + 545, + 382 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 305, + 337, + 545, + 382 + ], + "type": "text", + "content": " training data." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 402, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 545, + 437 + ], + "type": "text", + "content": "tribution of the real and the generated images, is utilized as the metric for both datasets. A lower FID indicates the synthetic images have better quality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 447, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 546, + 639 + ], + "type": "text", + "content": "Main Results Experiments of Tiny Updater on image-to-image translation are shown in Figure 7. It is observed that on the three image-to-image translation tasks, Tiny Updater is able to improve model performance by " + }, + { + "bbox": [ + 304, + 447, + 546, + 639 + ], + "type": "inline_equation", + "content": "5\\sim 10" + }, + { + "bbox": [ + 304, + 447, + 546, + 639 + ], + "type": "text", + "content": " FID when around " + }, + { + "bbox": [ + 304, + 447, + 546, + 639 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 447, + 546, + 639 + ], + "type": "text", + "content": " parameters are changed during updating. Besides, Tiny Updater can update the models by changing " + }, + { + "bbox": [ + 304, + 447, + 546, + 639 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 447, + 546, + 639 + ], + "type": "text", + "content": " parameters with no performance loss or by changing only " + }, + { + "bbox": [ + 304, + 447, + 546, + 639 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 447, + 546, + 639 + ], + "type": "text", + "content": " parameters with a slight FID increment. Qualitative analysis in Figure 6 shows that the pre-updating models can not transform the whole body of horses into the zebras, or remove all the stripes of zebras. In contrast, with Tiny Updater, the partially-updated model with only " + }, + { + "bbox": [ + 304, + 447, + 546, + 639 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 447, + 546, + 639 + ], + "type": "text", + "content": " parameters changed can address this problem and achieve comparable performance with the fully-updated models, indicating that Tiny Updater enables the neural network to be successfully updated by changing only a few parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 647, + 414, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 414, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 414, + 659 + ], + "type": "text", + "content": "4.3. Video Recognition" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "Experiment Setting We evaluate Tiny Updater in video recognition datasets including UCF-101 [73] and Diving-48 [46] with video recognition models including SlowOnly [19] and Inception3D [9]. UCF-101 is an ac" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23453" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 78, + 175, + 173 + ], + "blocks": [ + { + "bbox": [ + 60, + 78, + 175, + 173 + ], + "lines": [ + { + "bbox": [ + 60, + 78, + 175, + 173 + ], + "spans": [ + { + "bbox": [ + 60, + 78, + 175, + 173 + ], + "type": "image", + "image_path": "358257611ad7f6a95e1513dcf69c6a0b721ec66d24f8dac003ab9d5b85dd59ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 187, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 46, + 187, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 545, + 210 + ], + "type": "text", + "content": "Figure 8. Experimental results of Tiny Updater of video recognition on UCF-101 and Diving-48 with SlowOnly and Inception3D (I3D). The pre-updating models are trained with " + }, + { + "bbox": [ + 46, + 187, + 545, + 210 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 187, + 545, + 210 + ], + "type": "text", + "content": " training data." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 78, + 298, + 173 + ], + "blocks": [ + { + "bbox": [ + 184, + 78, + 298, + 173 + ], + "lines": [ + { + "bbox": [ + 184, + 78, + 298, + 173 + ], + "spans": [ + { + "bbox": [ + 184, + 78, + 298, + 173 + ], + "type": "image", + "image_path": "10b2c5672046302d5922cdd9e9519eae47a6b7901e9b1485cdb37980c09cc7e3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 310, + 78, + 425, + 173 + ], + "blocks": [ + { + "bbox": [ + 310, + 78, + 425, + 173 + ], + "lines": [ + { + "bbox": [ + 310, + 78, + 425, + 173 + ], + "spans": [ + { + "bbox": [ + 310, + 78, + 425, + 173 + ], + "type": "image", + "image_path": "ea2c9945c955e96f6d29edd0a67747d2c05253749bd49b9c617434b727415a5d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 433, + 78, + 549, + 172 + ], + "blocks": [ + { + "bbox": [ + 433, + 78, + 549, + 172 + ], + "lines": [ + { + "bbox": [ + 433, + 78, + 549, + 172 + ], + "spans": [ + { + "bbox": [ + 433, + 78, + 549, + 172 + ], + "type": "image", + "image_path": "cfed865dc1be1ce7088535073fd1272c1bb3045060b205b4306a76d23097890e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 229, + 288, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 288, + 291 + ], + "type": "text", + "content": "tion recognition dataset with 101 action classes over 13,000 video clips [73]. Diving-48 is a fine-grained video dataset of competitive diving. It has around 18,000 video clips belonging to 48 dive sequences [9]. Both top-1 and top-5 accuracy are reported." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 297, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 287, + 392 + ], + "type": "text", + "content": "Main Results Figure 8 shows the performance of Tiny Updater on video recognition. It is observed that on both SlowOnly and Inception3D (I3D), UCF-101 and Diving-48, the model with only " + }, + { + "bbox": [ + 46, + 297, + 287, + 392 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 297, + 287, + 392 + ], + "type": "text", + "content": " parameters changed trained by Tiny Updater achieves comparable and even higher performance than the fully-updated models. Moreover, the model with only " + }, + { + "bbox": [ + 46, + 297, + 287, + 392 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 297, + 287, + 392 + ], + "type": "text", + "content": " parameters changed during updating leads to only around " + }, + { + "bbox": [ + 46, + 297, + 287, + 392 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 297, + 287, + 392 + ], + "type": "text", + "content": " accuracy loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 403, + 117, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 117, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 117, + 415 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 423, + 214, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 214, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 214, + 436 + ], + "type": "text", + "content": "5.1. Choice of Pruning Granularity" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 441, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 287, + 537 + ], + "type": "text", + "content": "Existing network pruning methods can be mainly divided into regular pruning (e.g. channel-wise pruning) and irregular pruning (e.g. element-wise pruning). Usually, irregular pruning can achieve a higher compression rate, while regular pruning is more friendly to hardware. In this paper, we prefer to use regular pruning in Tiny Updater and suggest that using irregular pruning may cause the problem of index overhead and memory access overhead." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Index Overhead When Tiny Updater is utilized for updating neural network-based software, users should download not only the parameters which are changed during updating but also an index file that records the corresponding position of each changed parameter in the neural network. Given a neural network with " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " parameters, when irregular pruning is utilized, for each updated parameter, " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\log_2N" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " bits storage space is required to record which layer, filter, and position in the kernel this parameter should be. Thus, an ignorable additional downloading overhead is caused. In contrast, the index overhead for channel-wise pruning is " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\log_2\\frac{N}{C\\times K^2}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " bits, which is small enough to be almost ignorable. Here " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " indicate the channel number and the kernel size, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 229, + 547, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 547, + 386 + ], + "type": "text", + "content": "Memory Access Overhead In Tiny Updater, irregular pruning can result in the updated parameters being distributed across various neural layers, channels, and filters, leading to their storage in different data blocks of the file system. Consequently, when updating the model with the downloaded parameters, almost all data blocks of the edge device need to undergo a writing operation, resulting in a significant memory access overhead. In contrast, regular pruning ensures that the updated weights are channel-wise and, therefore, can be stored in the same data blocks. Consequently, only these data blocks for the updated parameters require memory access, potentially reducing the updating overhead on edge devices." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 409, + 378, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 409, + 378, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 409, + 378, + 422 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 566 + ], + "type": "text", + "content": "This paper introduces the challenge of efficient model updating to reduce the download overhead during neural network-based software updating. To this end, we propose Tiny Updater, which aims to change only a small subset of the parameters during model retraining, thereby reducing the number of parameters that need to be downloaded by users. Tiny Updater comprises an iterative pruning technique that prunes the parameters of the fully-updated models to the pre-updating models and a knowledge distillation technique that treats the fully-updated models as teachers and the partially-updated models as students." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Extensive experiments were conducted on eleven datasets covering general image classification, fine-grained image classification, video recognition, paired image-to-image translation, and unpaired image-to-image translation. The results demonstrate that Tiny Updater can update neural network-based software by changing only " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " of the parameters with almost no loss in accuracy. In addition, the experimental results show the effectiveness of Tiny Updater in various scenarios, including categorical updating, multi-step updating, and pruning. We believe that this paper will encourage further research to address the challenge of efficient model updating on edge devices." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23454" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 80, + 516, + 288 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 501, + 80 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 501, + 80 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 501, + 80 + ], + "type": "text", + "content": "Table 3. Comparison among incremental learning, efficient finetuning, and the proposed efficient model updating." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 80, + 516, + 288 + ], + "lines": [ + { + "bbox": [ + 82, + 80, + 516, + 288 + ], + "spans": [ + { + "bbox": [ + 82, + 80, + 516, + 288 + ], + "type": "table", + "html": "
ProblemTargetPre / Post TasksUsage of Training DataConstraints on Parameters
Incremental LearningThe human-like ability of incrementally learning new tasks without forgetting old tasks.Different.A small ratio of data is used for training pre-updating models. More data is used for updating. Data for pre-training data can not be accessed during updating (learning new tasks).No direct constraints. Reducing the number of changed parameters sometimes lead to positive influence.
Efficient FinetuningFinetuning pre-trained models for downstream tasks with low training overhead, and achieving better downstream performance.Different.Pre-updating models are pretrained on large-scale datasets, and then finetuned on much less data for downstream tasks. Usually do not access data for pre-training during finetuning.No direct constraints. Do not finetune the whole model usually leads to no changes in pre-trained weights. Additional parameters in adapter layers and prompts are required.
Efficient Model UpdatingUpdate models with new data by changing only a few parameters to reduce communicat-ion overhead.Same.A small ratio of data is used for pre-training models, and these data can still be accessed during updating.Having direct constraints, the changed parameters should be as few as possible.
", + "image_path": "ff230d1a3bb6ed6ca90137e6553675ff4ba50e7ea48b0cb924b5c6d5fa2e5590.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 316, + 162, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 162, + 330 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 162, + 330 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 403 + ], + "type": "text", + "content": "This research was partially supported by National Key R&D Program of China (2022YFB2804103), Key Research and Development Program of Shaanxi (2021ZDLGY01-05), Tsinghua University Dushi Program, National Natural Science Foundation of China (20211710187), and Tsinghua University Talent Program." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 415, + 174, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 174, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 174, + 428 + ], + "type": "text", + "content": "A. Detailed Comparison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 432, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 467 + ], + "type": "text", + "content": "A detailed comparison among incremental learning, efficient finetuning and the proposed efficient model updating is shown in Table 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 475, + 219, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 219, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 219, + 490 + ], + "type": "text", + "content": "B. Detailed Experimental Setting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 493, + 288, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 288, + 661 + ], + "type": "text", + "content": "CIFAR10/100 are two datasets for low resolution general image classification with 50,000 images and 1,000 images in the training and validating set, respectively. ImageNet is a large-scale dataset for general image classification with images belonging to 1,000 categories. Stanford Car is a dataset for fine-grained image classification with car images belonging to 197 classes. Oxford Flower is a dataset for fine-grained image classification which consists of 102 different categories of flowers common to the UK. Stanford Dog is a dataset for fine-grained image classification which contains 20,580 dog images of 120 breeds of dogs from around the world. FGVC Aircraft is a dataset for fine-grained image classifier on which contains 10,000 images of aircrafts spanning 100 aircraft models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 669, + 277, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 277, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 277, + 683 + ], + "type": "text", + "content": "C. Experiments on More Datasets and Tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 712 + ], + "type": "text", + "content": "Experimental results on more kinds of tasks and datasets are shown in Table 4 and Table 5. It is observed that" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 357, + 545, + 492 + ], + "blocks": [ + { + "bbox": [ + 305, + 315, + 545, + 348 + ], + "lines": [ + { + "bbox": [ + 305, + 315, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 305, + 315, + 545, + 348 + ], + "type": "text", + "content": "Table 4. Comparison between our method with previous efficient finetuning methods on ImageNet with Swin Transformer. Models are firstly trained with " + }, + { + "bbox": [ + 305, + 315, + 545, + 348 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 305, + 315, + 545, + 348 + ], + "type": "text", + "content": " data and then updated with all the data." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 357, + 545, + 492 + ], + "lines": [ + { + "bbox": [ + 307, + 357, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 357, + 545, + 492 + ], + "type": "table", + "html": "
Updating MethodRatio of Updated Params.(%)Top-1 Acc (%)
Pre-updating Model0.0%72.04
Global Finetuning100.0%81.30
Linear Finetuning1.1672.23
Non-linear Finetuning234.675.53
Non-linear Finetuning366.378.80
Adapter13.572.40
Adapter217.274.39
Visual Prompt Tuning≤ 1%72.32
Ours136.3180.87
Ours218.5679.91
", + "image_path": "8ce8861994936bad622594c321b40d825ca9459c0a36485c0ac3cf48070a5366.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 307, + 526, + 545, + 632 + ], + "blocks": [ + { + "bbox": [ + 345, + 515, + 506, + 525 + ], + "lines": [ + { + "bbox": [ + 345, + 515, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 345, + 515, + 506, + 525 + ], + "type": "text", + "content": "Table 5. Experimental results on more tasks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 526, + 545, + 632 + ], + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 632 + ], + "type": "table", + "html": "
TaskDatasetModelUpdated Params (%)Acc. / mAP / PSNR
Cross Domain Image ClassificationOffice-31ResNet500.0092.3
18.798.3
10.998.0
100.098.5
Object DetectionMS COCOFaster RCNN0.0030.5
27.336.2
14.635.6
100.036.4
Single Image Super-ResolutionDIV2KESRGAN0.0033.90
22.434.60
13.034.56
100.034.61
", + "image_path": "d3e2102386fcef66ba4bfb3e11f0e326cecfc9f24ff4a63bc6543c67a34c04bd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "our method leads to consistent good performance on cross-domain image classification, object detection and image super-resolution. Besides, our method outperforms previous efficient finetuning methods on ImageNet." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23455" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 147 + ], + "type": "text", + "content": "[1] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European Conference on Computer Vision (ECCV), pages 139-154, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "type": "text", + "content": "[2] Zeyuan Allen-Zhu and Yuanzhi Li. Towards understanding ensemble, knowledge distillation and self-distillation in deep learning. CoRR, abs/2012.09816, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "text", + "content": "[3] Ankur Bapna, Naveen Arivazhagan, and Orhan First. Simple, scalable adaptation for neural machine translation. arXiv preprint arXiv:1909.08478, 2019. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 250 + ], + "type": "text", + "content": "[4] Eyal Ben-David, Nadav Oved, and Roi Reichart. Pada: A prompt-based autoregressive approach for adaptation to unseen domains. arXiv preprint arXiv:2102.12206, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 295 + ], + "type": "text", + "content": "[5] Shweta Bhardwaj, Mukundhan Srinivasan, and Mitesh M Khapra. Efficient video classification using fewer frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 354-363, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 297, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 288, + 339 + ], + "type": "text", + "content": "[6] Cunling Bian, Wei Feng, Liang Wan, and Song Wang. Structural knowledge distillation for efficient skeleton-based action recognition. IEEE Transactions on Image Processing, 30:2963-2976, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 342, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 288, + 385 + ], + "type": "text", + "content": "[7] Cristian Bucilua, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 535-541. ACM, 2006. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 387, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 288, + 430 + ], + "type": "text", + "content": "[8] Efe Camci, Manas Gupta, Min Wu, and Jie Lin. Qlp: Deep q-learning for pruning deep neural networks. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 432, + 288, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 288, + 476 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 288, + 476 + ], + "type": "text", + "content": "[9] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "type": "text", + "content": "[10] Hyun Sung Chang, Sanghoon Sull, and Sang Uk Lee. Efficient video indexing scheme for content-based retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 9(8):1269-1279, 1999. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 566 + ], + "type": "text", + "content": "[11] Li Chen, Chunyan Yu, and Lvcai Chen. A new knowledge distillation for incremental object detection. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 610 + ], + "type": "text", + "content": "[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 679 + ], + "type": "text", + "content": "[13] Ali Cheraghian, Shafin Rahman, Pengfei Fang, Soumava Kumar Roy, Lars Petersson, and Mehrtash Harandi. Semantic-aware knowledge distillation for few-shot class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2534-2543, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "type": "text", + "content": "[14] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "text", + "content": "[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 248-255, 2009. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[17] Guiguang Ding, Shuo Zhang, Zizhou Jia, Jing Zhong, and Jungong Han. Where to prune: Using lstm to guide data-dependent soft pruning. IEEE Transactions on Image Processing, 30:293-304, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 230, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 295 + ], + "type": "text", + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 341 + ], + "type": "text", + "content": "[19] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 385 + ], + "type": "text", + "content": "[20] Yingchao Feng, Xian Sun, Wenhui Dao, Jihao Li, and Xin Gao. Double similarity distillation for semantic image segmentation. IEEE Transactions on Image Processing, 30:5363-5376, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "text", + "content": "[21] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 422, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 475 + ], + "type": "text", + "content": "[22] Sharath Girish, Shishira R Maiya, Kamal Gupta, Hao Chen, Larry S Davis, and Abhinav Shrivastava. The lottery ticket hypothesis for object recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 762-771, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 478, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 509 + ], + "type": "text", + "content": "[23] Saurabh Gupta, Judy Hoffman, and Jitendra Malik. Cross modal distillation for supervision transfer. In CVPR, pages 2827-2836, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 555 + ], + "type": "text", + "content": "[24] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In International Conference on Learning Representations, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 557, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 589 + ], + "type": "text", + "content": "[25] Xu Han, Weilin Zhao, Ning Ding, Zhiyuan Liu, and Maosong Sun. *Ptr: Prompt tuning with rules for text classification.* arXiv preprint arXiv:2105.11259, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 591, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 644 + ], + "type": "text", + "content": "[26] Yu Hao, Yanwei Fu, Yu-Gang Jiang, and Qi Tian. An end-to-end architecture for class-incremental object detection with knowledge distillation. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[27] Babak Hassibi and David G Stork. Second order derivatives for network pruning: Optimal brain surgeon. In Advances in neural information processing systems, pages 164-171, 1993. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 692, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 714 + ], + "type": "text", + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23456" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "text", + "content": "[29] Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang. Soft filter pruning for accelerating deep convolutional neural networks. arXiv preprint arXiv:1808.06866, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "text", + "content": "[30] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 1389-1397, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 195 + ], + "type": "text", + "content": "[31] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology Press, 2005. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 196, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 196, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 196, + 287, + 217 + ], + "type": "text", + "content": "[32] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. In NeurIPS, 2014. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 271 + ], + "type": "text", + "content": "[33] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "text", + "content": "[34] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Pix2pix datasets, http://efrosgans.eecs.berkeley.edu/pix2pix/datasets/.7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "type": "text", + "content": "[35] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 360, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 404 + ], + "type": "text", + "content": "[36] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "text", + "content": "[37] Qing Jin, Jian Ren, Oliver J Woodford, Jiazhuo Wang, Geng Yuan, Yanzhi Wang, and Sergey Tulyakov. Teachers do more than teach: Compressing image-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13600-13611, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 492 + ], + "type": "text", + "content": "[38] Hyeong-Ju Kang. Accelerator-aware pruning for convolutional neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 30(7):2093-2103, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 492, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 558 + ], + "type": "text", + "content": "[39] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 591 + ], + "type": "text", + "content": "[40] Xu Lan, Xiatian Zhu, and Shaogang Gong. Knowledge distillation by on-the-fly native ensemble. arXiv preprint arXiv:1806.04606, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "text", + "content": "[41] Yann LeCun, John S Denker, and Sara A Solla. Optimal brain damage. In Advances in neural information processing systems, pages 598-605, 1990. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "text", + "content": "[42] Muyang Li, Ji Lin, Yaoyao Ding, Zhijian Liu, Jun-Yan Zhu, and Song Han. Gan compression: Efficient architectures for interactive conditional gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5284-5294, 2020. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[43] Xuewei Li, Songyuan Li, Bourahla Omar, Fei Wu, and Xi Li. Reskd: Residual-guided knowledge distillation. IEEE Transactions on Image Processing, 30:4735-4746, 2021. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "type": "text", + "content": "[44] Xiaojie Li, Jianlong Wu, Hongyu Fang, Yue Liao, Fei Wang, and Chen Qian. Local correlation consistency for knowledge distillation. In European Conference on Computer Vision, pages 18-33. Springer, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[45] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[46] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In Proceedings of the European Conference on Computer Vision (ECCV), pages 513-528, 2018. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 196, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 545, + 241 + ], + "type": "text", + "content": "[47] Yun Li, Zechun Liu, Weiqun Wu, Haotian Yao, Xiangyu Zhang, Chi Zhang, and Baoqun Yin. Weight-dependent gates for network pruning. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "type": "text", + "content": "[48] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 276, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 320 + ], + "type": "text", + "content": "[49] Zeqi Li, Ruowei Jiang, and Parham Aarabi. Semantic relation preserving knowledge distillation for image-to-image translation. In European Conference on Computer Vision, pages 648-663. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "type": "text", + "content": "[50] Tsung-Yi Lin, Priyal Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE transactions on pattern analysis and machine intelligence, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 365, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 421 + ], + "type": "text", + "content": "[51] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "type": "text", + "content": "[52] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv preprint arXiv:2107.13586, 2021. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 545, + 521 + ], + "type": "text", + "content": "[53] Weide Liu, Guosheng Lin, Tianyi Zhang, and Zichuan Liu. Guided co-segmentation network for fast video object segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 31(4):1607-1617, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 522, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 567 + ], + "type": "text", + "content": "[54] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "type": "text", + "content": "[55] Xialei Liu, Marc Masana, Luis Herranz, Joost Van de Weijer, Antonio M Lopez, and Andrew D Bagdanov. Rotate your networks: Better weight consolidation and less catastrophic forgetting. In 2018 24th International Conference on Pattern Recognition (ICPR), pages 2262-2268. IEEE, 2018. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "text", + "content": "[56] Yuchen Liu, Zhixin Shu, Yijun Li, Zhe Lin, Federico Peruzzi, and Sun-Yuan Kung. Content-aware gan compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12156-12166, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 679, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 679, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 679, + 545, + 713 + ], + "type": "text", + "content": "[57] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23457" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 150 + ], + "type": "text", + "content": "[58] Zechun Liu, Haoyuan Mu, Xiangyu Zhang, Zichao Guo, Xin Yang, Kwang-Ting Cheng, and Jian Sun. Metapruning: Meta learning for automatic neural network channel pruning. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "text", + "content": "[59] Zechun Liu, Xiangyu Zhang, Zhiqiang Shen, Yichen Wei, Kwang-Ting Cheng, and Jian Sun. Joint multi-dimension pruning via numerical gradient update. IEEE Transactions on Image Processing, 30:8034-8045, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "type": "text", + "content": "[60] Rashid Minhas, Abdul Adeel Mohammed, and Q. M. Jonathan Wu. Incremental learning in human action recognition based on snippets. IEEE Transactions on Circuits and Systems for Video Technology, 22(11):1529-1541, 2012. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 286, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 286, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 286, + 286 + ], + "type": "text", + "content": "[61] Hossein Mobahi, Mehrdad Farajtabar, and Peter L Bartlett. Self-distillation amplifies regularization in hilbert space. arXiv preprint arXiv:2002.05715, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 287, + 342 + ], + "type": "text", + "content": "[62] Diego Ortego, Juan C. Sanmiguel, and José M. Martínez. Hierarchical improvement of foreground segmentation masks in background subtraction. IEEE Transactions on Circuits and Systems for Video Technology, 29(6):1645-1658, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 388 + ], + "type": "text", + "content": "[63] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 389, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 389, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 389, + 287, + 487 + ], + "type": "text", + "content": "[64] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "type": "text", + "content": "[65] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In Proceedings of the IEEE International Conference on Computer Vision, pages 5007-5016, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 545, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 599 + ], + "type": "text", + "content": "[66] Zheng Qin, Zeming Li, Zhaoning Zhang, Yiping Bao, Gang Yu, Yuxing Peng, and Jian Sun. Thundernet: Towards real-time generic object detection on mobile devices. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "type": "text", + "content": "[67] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[68] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91-99, 2015. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[69] Youngmin Ro and Jin Young Choi. Layer-wise pruning and auto-tuning of layer-wise learning rates in fine-tuning of deep networks. arXiv preprint arXiv:2002.06048, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[70] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In International Conference on Learning Representations, 2015. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[71] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4510-4520, 2018. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 215 + ], + "type": "text", + "content": "[72] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 249 + ], + "type": "text", + "content": "[73] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 250, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 293 + ], + "type": "text", + "content": "[74] Daniel Stanley Tan, Yong-Xiang Lin, and Kai-Lung Hua. Incremental learning of multi-domain image-to-image translations. IEEE Transactions on Circuits and Systems for Video Technology, 31(4):1526-1539, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 294, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 327 + ], + "type": "text", + "content": "[75] Mingxing Tan and Quoc V Le. Efficientnet: Rethinking model scaling for convolutional neural networks. arXiv preprint arXiv:1905.11946, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "type": "text", + "content": "[76] Mingxing Tan, Ruoming Pang, and Quoc V Le. Efficientdet: Scalable and efficient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10781-10790, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 372, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 448 + ], + "type": "text", + "content": "[77] Brian Thompson, Jeremy Gwinnup, Huda Khayrallah, Kevin Duh, and Philipp Koehn. Overcoming catastrophic forgetting during domain adaptation of neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2062–2068, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 449, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 481 + ], + "type": "text", + "content": "[78] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1365-1374, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "text", + "content": "[79] Ruize Wang, Duyu Tang, Nan Duan, Zhongyu Wei, Xuanjing Huang, Guihong Cao, Daxin Jiang, Ming Zhou, et al. K-adapter: Infusing knowledge into pre-trained models with adapters. arXiv preprint arXiv:2002.01808, 2020. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 526, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 580 + ], + "type": "text", + "content": "[80] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5987-5995, 2017. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 581, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 545, + 624 + ], + "type": "text", + "content": "[81] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In European Conference on Computer Vision, pages 588-604. Springer, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "text", + "content": "[82] Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, TatSeng Chua, and Maosong Sun. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797, 2021.4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[83] Jiahui Yu and Thomas S. Huang. Universally slimmable networks and improved training techniques. In The IEEE International Conference on Computer Vision (ICCV), October 2019. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "23458" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 564 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[84] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In International Conference on Learning Representations, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 183 + ], + "type": "text", + "content": "[85] Linfeng Zhang, Xin Chen, Xiaobing Tu, Pengfei Wan, Ning Xu, and Kaiheng Ma. Wavelet knowledge distillation: Towards efficient image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12464-12474, June 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "type": "text", + "content": "[86] Linfeng Zhang and Ma Kaisheng. Improve object detection with feature-based knowledge distillation: Towards accurate and efficient detectors. In International Conference on Learning Representations, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 262 + ], + "type": "text", + "content": "[87] Linfeng Zhang, Yukang Shi, Zuoqiang Shi, Kaisheng Ma, and Chenglong Bao. Task-oriented feature distillation. In NeurIPS, 2020. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 308 + ], + "type": "text", + "content": "[88] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In arXiv preprint: 1905.08094, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 351 + ], + "type": "text", + "content": "[89] Linfeng Zhang, Zhanhong Tan, Jiebo Song, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Scan: A scalable neural networks framework towards compact and efficient models. ArXiv, abs/1906.03951, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 397 + ], + "type": "text", + "content": "[90] Mingxin Zhao, Junbo Peng, Shuangming Yu, Liyuan Liu, and Nanjian Wu. Exploring structural sparsity in cnn via selective penalty. IEEE Transactions on Circuits and Systems for Video Technology, 32(3):1658-1666, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 398, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 398, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 398, + 287, + 430 + ], + "type": "text", + "content": "[91] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. arXiv preprint arXiv:2203.05557, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 432, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 287, + 475 + ], + "type": "text", + "content": "[92] Peng Zhou, Long Mai, Jianming Zhang, Ning Xu, Zuxuan Wu, and Larry S Davis. M2kd: Multi-model and multilevel knowledge distillation for incremental learning. arXiv preprint arXiv:1904.01769, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 476, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 510 + ], + "type": "text", + "content": "[93] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Pix2pix datasets, http://efrosgans.eecs.berkeley.edu/cyclegan/datasets.7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 564 + ], + "type": "text", + "content": "[94] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 7" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "23459" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/7a0fee57-111a-443f-8a63-68f0c3d1c0b8_content_list.json b/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/7a0fee57-111a-443f-8a63-68f0c3d1c0b8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c625553eb0924e67fc2454b5944d185a8d6f4409 --- /dev/null +++ b/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/7a0fee57-111a-443f-8a63-68f0c3d1c0b8_content_list.json @@ -0,0 +1,1602 @@ +[ + { + "type": "text", + "text": "TinyCLIP: CLIP Distillation via Affinity Mimicking and Weight Inheritance", + "text_level": 1, + "bbox": [ + 99, + 128, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kan Wu $^{1,\\ast}$ , Houwen Peng $^{3,\\ast,\\dagger}$ , Zhenghong Zhou $^{2,\\ast}$ , Bin Xiao $^{3}$ , Mengchen Liu $^{3}$ , Lu Yuan $^{3}$ , Hong Xuan $^{3}$ , Michael Valenzuela $^{3}$ , Xi (Stephen) Chen $^{3}$ , Xinggang Wang $^{2}$ , Hongyang Chao $^{1}$ , Han Hu $^{3}$ , Sun Yat-sen University, Huazhong University of Science & Technology, Microsoft", + "bbox": [ + 84, + 186, + 895, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 285, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a novel cross-modal distillation method, called TinyCLIP, for large-scale language-image pre-trained models. The method introduces two core techniques: affinity mimicking and weight inheritance. Affinity mimicking explores the interaction between modalities during distillation, enabling student models to mimic teachers' behavior of learning cross-modal feature alignment in a visual-linguistic affinity space. Weight inheritance transmits the pre-trained weights from the teacher models to their student counterparts to improve distillation efficiency. Moreover, we extend the method into a multi-stage progressive distillation to mitigate the loss of informative weights during extreme compression. Comprehensive experiments demonstrate the efficacy of TinyCLIP, showing that it can reduce the size of the pre-trained CLIP ViT-B/32 by $50\\%$ , while maintaining comparable zero-shot performance. While aiming for comparable performance, distillation with weight inheritance can speed up the training by $1.4 - 7.8 \\times$ compared to training from scratch. Moreover, our TinyCLIP ViT-8M/16, trained on YFCC-15M, achieves an impressive zero-shot top-1 accuracy of $41.1\\%$ on ImageNet, surpassing the original CLIP ViT-B/16 by $3.5\\%$ while utilizing only $8.9\\%$ parameters. Finally, we demonstrate the good transferability of TinyCLIP in various downstream tasks. Code and models will be open-sourced at aka.ms/tinyclip.", + "bbox": [ + 76, + 316, + 473, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 732, + 209, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale language-image pretraining, e.g., CLIP [46], has recently gained significant attention due to its remarkable zero-shot transfer capability [46] and unprecedented performance in text-to-image generation [47]. Due to complex nature of vision and language, current approaches often resort to utilizing huge amounts of parameters to endow models with cross-modal capabilities [46, 21, 71, 72, 43, 39,", + "bbox": [ + 75, + 758, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc00e8caee1bbc668dcf0890656304b70bb3c1286c8df0fcf45e41f05d538a4d.jpg", + "image_caption": [ + "Figure 1. Comparison of the OpenCLIP [20] and TinyCLIP. TinyCLIP is pre-trained and distilled on LAION-400M [50] using OpenCLIP ViT-B/32 [20] as the teacher, whose zero-shot top-1 accuracy is $65.6\\%$ ." + ], + "image_footnote": [], + "bbox": [ + 506, + 262, + 704, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/34580367f76f388d597d70d6ebe129118186435bb8b4ae9e0432eeaee12e712a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 263, + 859, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1]. This in turn leads to high costs in terms of storage, memory, and computation time for these models, which motivates the need for model compression to make them smaller and faster for real-world applications [64, 13].", + "bbox": [ + 496, + 536, + 893, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a core compression technique, knowledge distillation has been extensively studied and applied in single-modal settings [19, 16]. However, its potential for multi-modality remains underexplored. Unlike single-modal models, the distillation of language-image cross-modal models poses distinct challenges. First, CLIP-like language-image models commonly consist of two branches: an image encoder and a text encoder [21, 72, 52, 73]. When distilling such multi-branch models, it is crucial to consider the interaction of information across the different modality branches in both the teacher and student models. Second, the original CLIP [46] models are pre-trained on 400 million image-text pairs for 32 epochs, which requires thousands of GPU days, making distillation a significant challenge when computational resources are limited. Is there any way to reduce the cost of CLIP distillation?", + "bbox": [ + 496, + 598, + 895, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle these challenges, we present a novel cross-modal distillation method dubbed TinyCLIP, which introduces two key techniques: affinity mimicking and weight inheritance. In contrast to the methods that rely on either", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. Kan and Zhenghong were interns of Microsoft.", + "bbox": [ + 99, + 875, + 467, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding: houwen.peng@microsoft.com", + "bbox": [ + 99, + 887, + 357, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21970", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image or text features for distillation, we empirically show that distilling knowledge in an image-text affinity space is more effective. Specifically, we leverage the cosine similarity of the image and text embeddings in the teacher model to facilitate the distillation of the student model, allowing the student to mimic the teacher's visual-linguistic feature alignment. We refer to this process as affinity mimicking.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To improve distillation efficiency, we introduce weight inheritance, a technique transferring the pre-trained weights from teacher models to their student counterparts. Since inheriting weights provides a good initialization for the student models, the distillation progress can be largely accelerated. The key challenge of weight inheritance lies in determining which weights are more advantageous. To address this issue, we introduce two solutions: manual and automatic inheritance. We surprisingly found that a simple manual selection of $k$ -dimension or $k$ -layer weights from the teacher model can yield satisfactory results for CLIP distillation. On the other hand, we also introduce learnable masks to automatically identify the most important weights from the teacher model. The masks are imposed independently on the vision and language branches, enabling them to capture the differences across modalities.", + "bbox": [ + 75, + 198, + 470, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, we extend the proposed weight inheritance to a multi-stage progressive procedure, where each subsequent stage automatically inherits the important weights from the preceding stages. We observed that when the teacher model exhibits higher performance and shares a similar architecture with the student, weight inheritance can provide better results. This is because significant architecture differences may undermine the learned weights when transmitting from the teacher to the student. Therefore, we break the inheritance into multiple stages, allowing the student model in each stage to share a more similar structure with the predecessor teacher and inherit the weights progressively.", + "bbox": [ + 75, + 439, + 468, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments show that TinyCLIP delivers competitive models at all levels of speedups and model sizes in ImageNet zero-shot evaluation and various downstream tasks. As shown in Fig. 1, pre-trained on the same LAION-400M dataset [50], TinyCLIP-ViT using 63M parameters achieves $61.4\\%$ zero-shot top-1 accuracy on ImageNet [7], being $2.0 \\times$ smaller and $1.5 \\times$ faster than the OpenCLIP [20] model (62.9% accuracy with 126M parameters). Meanwhile, TinyCLIP-ResNet with 38M parameters obtains $56.4\\%$ zero-shot top-1 accuracy, being $2.1 \\times$ smaller and $2.0 \\times$ faster than CLIP ResNet-50 (59.6% accuracy with 76M parameters). Moreover, our method can speed up the original OpenCLIP training by $1.4 \\times -7.8 \\times$ while achieving similar performance. Also, TinyCLIP demonstrates good transfer capacities in downstream scenarios. In summary, the contributions of this work are two-fold:", + "bbox": [ + 75, + 619, + 470, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a new cross-modal distillation approach to unleash the capacity of small CLIP models, fully lever", + "bbox": [ + 84, + 869, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "aging large-scale models as well as pre-training data. To our best knowledge, this is the first work exploring the pre-training distillation of language-image models.", + "bbox": [ + 519, + 90, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present state-of-the-art language-image pre-trained models at small scale, striking the best trade-off between speed and accuracy. Extensive experiments demonstrate the superiority and good generalization ability of the small models in various downstream tasks.", + "bbox": [ + 506, + 148, + 893, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 238, + 640, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language-Image Pre-training has achieved remarkable progress over the past few years [10, 4]. In particular, contrastive language-image pre-training demonstrates very impressive zero-shot transfer and generalization capacities [45, 46, 21, 72, 61, 43, 74]. One of the most representative works is CLIP [46]. A large amount of follow-up works have been proposed to optimize the pre-training framework [27, 76, 15, 68, 62, 14, 28, 67]. Meanwhile, a line of works leverage the pre-trained models for downstream tasks, such as open-vocabulary detection and segmentation [29, 75, 66, 32], video recognition [56, 23], and text-to-image generation [47, 41].", + "bbox": [ + 496, + 265, + 890, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most recently, there are a few works attempting to scale language-image pre-trained models. BEiT-3 [57] uses a multiway transformer to scales up multi-modal pre-trained models to 1.9B parameters. FLIP [30] scales image-text pre-training via masked modeling, enabling CLIP to enjoy faster training speed while getting better performance. On the other hand, there are also few works on scaling down CLIP models [51]. However, they all focus on specific tasks, such as generation [5, 51] and video content understanding [42]. In contrast, our work concentrates on language-image pre-training, which is the first work on CLIP compression.", + "bbox": [ + 496, + 446, + 892, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge Distillation in a teacher-student framework [19] has been widely used to transfer knowledge from large models to small ones. Distillation in single modality, such as vision [54, 64, 37] or language [48, 22], has been extensively studied. However, the exploration of cross-modality distillation is relatively limited. A few early works exploring cross-modal distillation only focus on specific tasks, such as VQA and image caption [5, 11, 59], limiting the methodologies to be applied in general downstream tasks. In contrast, this work pays attention to general cross-modal pre-training distillation, in which the distilled models can be transferred to various downstream scenarios.", + "bbox": [ + 496, + 628, + 890, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is worth noting that our proposed weight inheritance strategy shares similar spirits to model pruning techniques [17, 65, 51]. Both approaches involve the identification of important weights while removing redundant ones. However, they have three fundamental differences. 1) Previous pruning methods predominantly focus on single-modality", + "bbox": [ + 496, + 810, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21971", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models [17, 65, 3], whereas our approach shifts attention to multi-modality. Since different modalities exhibit varying redundancy characteristics, it is necessary to consider these differences when selecting important weights from each modality. 2) Unlike previous pruning methods which have emphasized the importance of the pruned architecture [12, 34, 58], we argue that the weights inherited from the original models are also highly beneficial, particularly when compressing CLIP-like language-image models. 3) Our method leverages a progressive multi-stage process, which enables the inherited weights to accelerate the convergence of small models during training.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 285, + 168, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we propose TinyCLIP, a simple and effective method for distilling large-scale language-image pretraining models, such as CLIP. It consists of three components: affinity mimicking (§3.1), weight inheritance (§3.2) and multi-stage progressive distillation (§3.3).", + "bbox": [ + 75, + 310, + 468, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Distillation with Affinity Mimicking", + "text_level": 1, + "bbox": [ + 76, + 395, + 390, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A language-image pre-training model often consists of two branches, an image encoder to extract visual representations and a text encoder to extract textual representations. The visual and textual representations are linearly projected into a cross-modal common space by minimizing a contrastive loss $\\mathcal{L}_0$ . The supervision signal can be viewed as an identity matrix, in which diagonal values are set to 1 for positive image-text pairs, and all others are set to 0 for negative pairs. This signal disregards the similarity between negative pairs, which prevents the model from acquiring a nuanced understanding of the intricate relationships among negative pairs. Therefore, we introduce affinity mimicking to enable student models to learn similarities between negative pairs from teacher models.", + "bbox": [ + 75, + 417, + 468, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More specifically, as shown in Fig. 2, we consider two types of affinity distillation losses: image-to-language loss $\\mathcal{L}_{I2T}$ and language-to-image loss $\\mathcal{L}_{T2I}$ . The former learns the alignment between the teacher and student models based on the image-to-language affinity $A_{I2T}$ , which represents the affinity scores of an image with all the text descriptions in a batch (the yellow row highlighted in Fig. 2). The latter refers to the alignment of language-to-image affinity $A_{T2I}$ , which compares a text description with all the images to obtain a matching score (the blue column in Fig. 2). The combination of these two losses formulates our affinity mimicking distillation, which is represented as:", + "bbox": [ + 75, + 630, + 468, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i s t i l l}} = \\mathcal {L} _ {I 2 T} + \\mathcal {L} _ {T 2 I}, \\tag {1} \\\\ = C E \\left(A _ {I 2 T} ^ {s}, A _ {I 2 T} ^ {t}\\right) + C E \\left(A _ {T 2 I} ^ {s}, A _ {T 2 I} ^ {t}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 823, + 467, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $CE$ represents the cross-entropy loss. The superscripts $s$ and $t$ denote the student and teacher models, re", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/070cab7b775b02144ec4257e65c466a94e5563f728d1555c61ff45401e06aae3.jpg", + "image_caption": [ + "Figure 2. Affinity mimicking for language-image models. The loss includes image-to-text loss (yellow) and text-to-image loss (blue)." + ], + "image_footnote": [], + "bbox": [ + 527, + 88, + 864, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "spectively. The elements in $A_{I2T}$ and $A_{T2I}$ are defined as:", + "bbox": [ + 498, + 353, + 890, + 369 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA _ {I 2 T} (i, j) = \\frac {\\exp \\left(I _ {i} \\cdot T _ {j} / \\tau\\right)}{\\sum_ {k \\in \\mathcal {B}} \\exp \\left(I _ {i} \\cdot T _ {k} / \\tau\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 391, + 890, + 426 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA _ {T 2 I} (i, j) = \\frac {\\exp \\left(I _ {i} \\cdot T _ {j} / \\tau\\right)}{\\sum_ {k \\in \\mathcal {B}} \\exp \\left(I _ {k} \\cdot T _ {j} / \\tau\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 428, + 890, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is a temperature parameter, $I_{i}$ is the feature embedding of $i$ -th image in the batch $\\mathcal{B}$ , $T_{j}$ is the feature of $j$ -th text description in $\\mathcal{B}$ . Affinity mimicking allows student models to imitate the behavior of the large model in learning visual-linguistic alignment. In contrast to previous methods that rely on either image or text features for distillation, we show that distilling knowledge in this image-text affinity space is more effective.", + "bbox": [ + 496, + 473, + 892, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Distillation with Weight Inheritance", + "text_level": 1, + "bbox": [ + 500, + 604, + 812, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The original CLIP models are pre-trained on 400 million image-text pairs for 32 epochs, taking thousands of GPU days. This presents a significant cost challenge for distillation. To improve the training efficiency, we introduce weight inheritance, a technique that inherits the important weights from the well-trained large teacher models to smaller student models. The key challenge of inheriting weights is identifying important weights from the large amount of weights of the teacher. We propose two approaches to select the important weights: manual weight inheritance and automatic weight inheritance.", + "bbox": [ + 496, + 628, + 890, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Manual Inheritance. For manual weight inheritance, we first analyze the redundancy of existing CLIP pre-trained models. Fig. 4 shows that the text encoder displays more redundancy in depth (layer-wise), while the image encoder exhibits more redundancy in width (channel-wise). Based upon this discovery, we adopt the approach of uniformly selecting $k$ layers of the text branch and directly taking the", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21972", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "front $k$ channels of the image branch to select the important weights from large teacher models. These weights will serve as the initialization for small models during distillation. We surprisingly found that such a simple manual selected weights can largely accelerate CLIP distillation.", + "bbox": [ + 75, + 90, + 468, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Automatic Inheritance. Although manual inheritance can be considered an effective approach for model compression, it does present the drawback of necessitating prior knowledge in order to determine which weights to inherit. This limitation may render the technique impractical for application to diverse models. To solve this issue, we present an automatic weight inheritance scheme. Inspired by structured pruning in large language models [65], we introduce learnable masks to identify weight importance. Considering the difference across modalities, the learnable masks are imposed on vision and language branches independently, as visualized in Fig. 3. An overall sparsity constraint is introduced to guarantee the selected number of important weights to meet our compression requirements. Without loss of generality, here we use the transformer architecture as an example to introduce the detailed procedure of weight inheritance. Specifically, a standard transformer block contains a multi-head attention (MHA) layer and a feed-forward network (FFN). To capture the importance of weights in a fine-grained level, we introduce two mask variables $\\mathbf{m}_{\\mathrm{head}}$ , $\\mathbf{m}_{\\mathrm{int}} \\in \\{0,1\\}$ to identify the redundant attention heads in MHA and neurons in FFN respectively, while keeping the important ones. These two kinds of masks are imposed on the activation of attention heads and the intermediate layer of FFN, which is formulated as", + "bbox": [ + 75, + 167, + 470, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M H A} (\\mathbf {X}) = \\sum_ {h = 1} ^ {N _ {H}} \\mathbf {m} _ {\\text {h e a d}} ^ {h} \\cdot \\operatorname {A t t n} _ {\\mathbf {W} _ {Q} ^ {h}, \\mathbf {W} _ {K} ^ {h}, \\mathbf {W} _ {V} ^ {h}, \\mathbf {W} _ {O} ^ {h}} (\\mathbf {X}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 549, + 468, + 589 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {F F N} (\\mathbf {X}) = \\operatorname {G e L U} \\left(\\mathbf {X} \\mathbf {W} _ {U}\\right) \\cdot \\operatorname {d i a g} \\left(\\mathbf {m} _ {\\text {i n t}}\\right) \\cdot \\mathbf {W} _ {D}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 593, + 468, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{X}$ is the layer input, $\\mathbf{W}_Q^h$ , $\\mathbf{W}_K^h$ , $\\mathbf{W}_V^h$ , $\\mathbf{W}_O^h \\in \\mathbb{R}^{d \\times d_h}$ denote the query, key, value and output matrices in MHA respectively, and $\\mathbf{W}_U \\in \\mathbb{R}^{d \\times d_f}$ and $\\mathbf{W}_D \\in \\mathbb{R}^{d_f \\times d}$ represent the parameters of FFN layers. Here, $d$ denotes the hidden dimension (e.g., 768), and $d_h = d / N_H$ denotes the output dimension of each head (e.g., 64), where $N_H$ is the total number of heads. The trainable masks $\\mathbf{m}_{\\mathrm{head}}$ and $\\mathbf{m}_{\\mathrm{int}}$ serve as importance indicators for MHA and FFN layers. Moreover, to further learn the importance of embedding dimensions in transformer, we introduce an additional mask $\\mathbf{m}_{\\mathrm{embed}} \\in \\{0,1\\}$ . This mask is shared across all layers because each dimension in the hidden representation is connected to the corresponding dimension in the subsequent layer through a residual connection.", + "bbox": [ + 75, + 616, + 468, + 827 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During automatic inheritance, there are two losses used for optimizing the masks and the model jointly, i.e., a sparsity loss and a distillation loss defined in Eq. (1):", + "bbox": [ + 75, + 829, + 468, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {d i s t i l l}} + \\mathcal {L} _ {\\text {s p a r s i t y}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 881, + 468, + 897 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a4bf8109bb3809a89651d98aff1e26bcba7e2318a3d6dafdeb9dde1fceb76366.jpg", + "image_caption": [ + "10 Learnable mask/manual mask", + "Figure 3. Weight inheritance of CLIP. Given a large pre-trained CLIP model, the inherited weights are selected by masks. The weights whose mask is 0 will be removed. For the manual way, the masks are set to 1 in the first k-dimension or uniform k-layer, the rest is set to 0. For the automatic way, the masks are learned jointly with the model under the sparsity constraint." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 705, + 265 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/74d8b9cee14fdbba392dde2ac887ef215c07794afc6610d703dd83cecda7939d.jpg", + "image_caption": [ + "Model sparsity $=$" + ], + "image_footnote": [], + "bbox": [ + 707, + 89, + 870, + 265 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s p a r s i t y}} = \\lambda \\cdot (p - q) + \\beta \\cdot (p - q) ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 401, + 890, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda, \\beta$ are learnable multipliers that guarantee $p = q[60, 65]$ . Here, $q$ is the target compression rate, $p$ is the overall compression rate of learnable masks for the model, including image encoder and text encoder:", + "bbox": [ + 496, + 429, + 890, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p = \\frac {1}{M _ {t} + M _ {t}} \\sum_ {i m g, t x t} (4 \\cdot d _ {h} \\cdot \\sum_ {i} ^ {L} \\sum_ {j} ^ {N _ {H}} \\sum_ {k} ^ {d} \\mathbf {m} _ {\\text {h e a d}} ^ {i, j} \\cdot \\mathbf {m} _ {\\text {e m b e d}} ^ {k} \\\\ + 2 \\cdot \\sum_ {i} ^ {L} \\sum_ {j} ^ {d _ {f}} \\sum_ {k} ^ {d} \\mathbf {m} _ {\\text {i n t}} ^ {i, j} \\cdot \\mathbf {m} _ {\\text {e m b e d}} ^ {k}), \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 498, + 890, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_{i}$ and $M_{t}$ represent the full model size of the image encoder and text encoder, respectively, $L$ is the total number of transformer layers.", + "bbox": [ + 496, + 550, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Progressive Multi-Stage Distillation", + "text_level": 1, + "bbox": [ + 500, + 604, + 807, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When attempting to achieve a high target sparsity, i.e., $>70\\%$ , compressing the model in a single stage can lead to a significant reduction in accuracy and even result in convergence failure. This is due to the fact that most weights of the large model are directly discarded, including those that are important for ensuring model quality and convergence. As a solution, we propose a multi-stage progressive distillation method to achieve a high compression rate without seriously sacrificing accuracy. In each stage, we use a modest degree of compression, e.g., $25\\%$ , to avoid large loss of performance and make training stable.", + "bbox": [ + 496, + 628, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, each stage includes weight inheritance and affinity mimicking, as described in Algorithm 1. During the weight inheritance phase, the weights of the large model are gradually reduced until the target sparsity and key weights are retained for the small model. After that, the small model is distilled with affinity mimicking, which involves transferring the visual-linguistic affinity knowledge from", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21973", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: TinyCLIP: Cross-Modal Distillation" + ], + "code_body": "Input: a pre-trained CLIP model $f(x;\\theta_0)$ with weight $\\theta_0$ , target compression rate $q$ total iteration steps $N$ 1 Initialize the number of stages $G$ , each stage steps $L$ , target compression rate $q_{i}$ of stage i, weight inheritance steps $L_{M}$ input $x$ , learnable mask $M = \\{\\mathbf{m}_{\\mathrm{head}},\\mathbf{m}_{\\mathrm{int}},\\mathbf{m}_{\\mathrm{embed}}\\} ;^1$ \n2 for i in [1,...,G] do \n3 // Increase mask compression rate, retain important weights \n4 for j in [1,...,LM] do \n5 Update target compression rate $^2 q$ of current step based on $q_{i}$ and step $j$ . \n6 Calculate distillation loss between origin teacher model $f(x;\\theta_0)$ and masked student model $f(x;\\theta_{i - 1}\\odot M_i)$ . \n7 Calculate sparsity loss between learnable mask compression rate $^3 p$ and target compression rate $q$ . \n8 Optimize $\\theta_{i - 1}$ and $M_{i}$ . \n9 end \n10 $\\theta_{i}\\gets \\theta_{i - 1}\\odot M_{i}$ //Inherit weight with masks \n11 // Cross-Modal distillation \n12 for j in $[L_M + 1,\\dots ,L]$ do \n13 Calculate distillation loss between origin teacher model $f(x;\\theta_0)$ and compressed model $f(x;\\theta_i)$ . \n14 Optimize $\\theta_{i}$ . \n15 end \n16 end \n17 Return compressed model $f(x;\\theta_G)$ at compression rate $q$ \n1 Each stage compress $25\\%$ $q_{i} = 1 - i\\cdot 25\\%$ $G = \\lceil (1 - q) / 0.25\\rceil$ $L = N / G,L_{M} = 3000,M$ is initialized to 1, $q_{i}$ can be adjusted according to different models. \n2 Target compression rate $q$ increases linearly from O to $q_{i}$ with step $j$ . \n $p$ is calculated by masks $M$ in Eq. (8).", + "bbox": [ + 81, + 111, + 468, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the teacher to the student. This procedure is repeated until the target sparsity level is achieved.", + "bbox": [ + 75, + 491, + 468, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 532, + 209, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 551, + 294, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture. The original CLIP employs a transformer-based model as its text encoder, while its image encoder has two variations, i.e., ViT [9] and modified ResNet [18], covering both transformer-based and CNN-based architectures. For weight inheritance, we inherited the weights from two model variants: OpenCLIP ViT-B/32 or ViT-B/16 [20] and CLIP ResNet-50 [46], which are pre-trained on LAION-2B [49] and WIT-400M [46], respectively1. For distillation, we use OpenCLIP ViT-B/32 [20] pre-trained on LAION-2B [49] as the teacher model, since it achieves a high zero-shot accuracy of $65.6\\%$ on ImageNet [7] with a high inference throughput. Besides, we also extend our method to DaViT [8], which is a hybrid architecture. We re-produce Florence-DaViT-5M [72, 8] on LAION-400M as a baseline. Then we distill the same model using affinity mimicking and weight inheritance. The teacher model is Florence-DaViT-D3, which is pre-trained on FLD-900M [72] and achieves a top-1 accuracy of $78.0\\%$ on ImageNet [7] with 128M parameters.", + "bbox": [ + 75, + 575, + 472, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Affinity Mimicking. As defined in Tab. 2, we explore", + "bbox": [ + 94, + 863, + 468, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "different interaction schemes between teacher and student models across modalities. The contrastive loss $\\mathcal{L}_0$ and affinity mimicking $\\mathcal{L}_1$ (i.e., $\\mathcal{L}_{\\text{distill}}$ ) serve as the basic ones, which can be combined with other losses. The weight of each loss is set to 1 when combined. The temperature parameter $\\tau$ in Eq. (3) is set to $\\frac{1}{50}$ by default. The sensitivity analysis of $\\tau$ is presented in the supplementary material.", + "bbox": [ + 500, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Weight Inheritance. We compress OpenCLIP ViT-B [20] and CLIP ResNet-50 [46] using the proposed multi-stage progressive distillation. In each stage, we apply manual or automatic weight inheritance. In manual inheritance, we compress the image encoder and the text encoder separately, by reducing the width of image encoder and the depth of text encoder. In automatic inheritance, we set the target sparsity for the whole model. In each stage, the trainable masks are initialized to 1 and updated in the first 3,000 training iterations, where the optimizer is AdamW [35, 24] with a constant learning rate of 0.01 and no weight decay. The learnable multipliers $\\lambda, \\beta$ defined in Eq. (7) are both initialized to 0.01.", + "bbox": [ + 496, + 198, + 892, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Settings. We train our models on two public datasets, namely LAION-400M [50] and YFCC-15M [53]. On LAION-400M, the models are compressed in 3 stages, including $100\\%$ to $75\\%$ parameters for 6 epochs, $75\\%$ to $50\\%$ parameters for 16 epochs, and $50\\%$ to $25\\%$ parameters for 16 epochs. On YFCC-15M, it contains 2 compression stages, where the training epochs are both 25 from $100\\%$ to $50\\%$ parameters, and $50\\%$ to $10\\%$ . We follow the hyper-parameter of CLIP [46] except that the learning rate is set to $10^{-4}$ when using weight inheritance. The details are shown in the supplementary material. Unless otherwise specified, our ablation studies utilize the same settings, except for training the models only for 1 epoch with a learning rate of $5 \\times 10^{-4}$ . All models are trained on 32 Nvidia V100 or A100, and implemented with PyTorch [40], OpenCLIP [20], gradient cache [36], and Timm library [63].", + "bbox": [ + 496, + 395, + 892, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Settings. The models are evaluated on multiple benchmarks. In zero-shot transfer evaluation and robustness evaluation, we follow the same prompt engineering in CLIP [46], where 80 text templates per class are used for ImageNet [7]. In linear probe, we use Elevater toolkit [26] for evaluation, where a classification head is trained for 50 epochs with searched hyper-parameters. Note that we do not count the number of parameters in the text embedding layer. It is a look-up table whose parameter size is the same as the models with the same hidden dimension and vocabulary size. The inference throughput is measured on Nvidia V100 with CUDA 11 and PyTorch v1.12 [40], where the batch size is 1,024.", + "bbox": [ + 496, + 638, + 893, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art Models", + "text_level": 1, + "bbox": [ + 500, + 845, + 856, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our TinyCLIP with state-of-the-art models in Tab. 1. The zero-shot evaluation on ImageNet-1K [7],", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1The models can be downloaded from OpenCLIP and OpenAI.", + "bbox": [ + 94, + 886, + 428, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21974", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d11a2489ce94c466333d7ce98d971982d9e21178964ad457aa6ed5f749e0f2c8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodImageEncoderText Encoder#Params (M)Image+TextMACs(G)Throughput(pairs/s)TrainingdatasetsIN-1Ktop1 acc (%)Flickr30kMSCOCO
depthwidthI→T@1T→I@1I→T@1T→I@1
Training data: 15M / 20M
RILS [69]ViT-B/161251286 + 3820.5818LAION-20M45.045.134.932.225.5
MaskCLIP [77]ViT-B/161251286 + 3820.5818LAION-20M46.664.948.138.524.8
SLIP [38]ViT-B/161251286 + 3820.5818YFCC-15M42.857.640.131.120.3
CLIP [46, 38]ViT-B/161251286 + 3820.5818YFCC-15M37.651.632.226.517.1
TinyCLIP (Ours)ViT-39M/16651239+19(2.1×)9.5(2.2×)1,469(1.8×)YFCC-15M63.584.466.754.938.9
TinyCLIP (Ours)ViT-8M/1632568+3(11.3×)2.0(10.3×)4,150(5.1×)YFCC-15M41.162.342.336.221.5
Training data: 400M
Florence [72]DaViT-5M62565+51.12,980LAION-400M45.061.241.536.220.9
TinyCLIP (Ours)DaViT-5M62565+51.12,980LAION-400M50.066.248.540.724.5
CLIP [46]ResNet-1011251256 + 3812.81,161WIT-400M62.278.159.249.330.7
CLIP [46]ResNet-501251238 + 389.11,549WIT-400M59.681.258.250.828.3
TinyCLIP (Ours)ResNet-30M951230+29(1.3×)6.9(1.3×)1,811(1.2×)LAION-400M59.181.161.252.733.9
TinyCLIP (Ours)ResNet-19M651219+19(2.0×)4.4(2.1×)3,024(2.0×)LAION-400M56.476.258.348.930.9
OpenCLIP [20]ViT-B/321251288 + 387.42,452LAION-2B65.784.766.856.939.3
CLIP [46]ViT-B/321251288 + 387.42,452WIT-400M63.280.159.851.230.6
OpenCLIP [20]ViT-B/321251288 + 387.42,452LAION-400M62.979.362.053.335.4
TinyCLIP (Ours)ViT-61M/32951261+29(1.4×)5.3(1.4×)3,191(1.3×)LAION-400M62.178.663.353.935.9
TinyCLIP (Ours)ViT-40M/32651240+19(2.1×)3.5(2.1×)4,641(1.9×)LAION-400M59.777.358.949.833.1
TinyCLIP (Ours)ViT-63M/32autoauto63+31(1.3×)5.6(1.3×)2,905(1.2×)LAION-400M63.983.264.455.537.6
TinyCLIP (Ours)ViT-45M/32autoauto45+18(2.0×)3.7(2.0×)3,682(1.5×)LAION-400M61.480.962.252.834.7
TinyCLIP (Ours)ViT-22M/32autoauto22+10(3.9×)1.9(3.9×)5,504(2.2×)LAION-400M53.771.352.044.428.3
TinyCLIP (Ours)ViT-63M/32autoauto63+31(1.3×)5.6(1.3×)2,909(1.2×)LAION+YFCC-400M64.584.966.056.938.5
TinyCLIP (Ours)ViT-45M/32autoauto45+18(2.0×)3.7(2.0×)3,685(1.5×)LAION+YFCC-400M62.780.363.954.036.7
", + "bbox": [ + 78, + 75, + 890, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with the state-of-the-art methods. The architecture of text encoder is Transformer [55]. \"auto\" denotes automatic weight inheritance, and our other models use manual weight inheritance.", + "bbox": [ + 76, + 357, + 892, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Flickr30k [44] and MSCOCO retrieval [33] are reported.", + "bbox": [ + 76, + 410, + 450, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When performing training on YFCC-15M [53], we use affinity mimicking and manual weight inheritance to distill OpenCLIP ViT-B/16 [20], whose zero-shot performance on ImageNet is $70.2\\%$ . By inheriting half of parameters, our TinyCLIP ViT-39M/16 can achieve a zero-shot accuracy of $63.5\\%$ on ImageNet, surpassing the original CLIP ViT-B/16 [46] by $25.9\\%$ . Furthermore, when compressing the image encoder to 10 layers with 256 dimensions and the text encoder to 3 layers with 256 dimensions, TinyCLIP ViT-8M/16 still surpasses CLIP ViT-B/16 [46] by $3.5\\%$ while using $11.3\\times$ fewer parameters and running $5.1\\times$ faster.", + "bbox": [ + 75, + 426, + 470, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When conducting training on LAION-400M [50], we distill three different models, i.e., Florence DaViT [72, 8], CLIP ResNet [46], and OpenCLIP ViT-B/32 [20]. The results in Tab. 1 demonstrate the efficacy of our method. More specifically, our TinyCLIP DaViT-5M achieves up to $5.0\\%$ improvements over the original model on ImageNet, and $5.0 / 4.5\\%$ gains in image-to-text retrieval on Flicker and MSCOCO, respectively. It is worth noting that this is a tiny language-image pre-trained model using only 10M parameters to achieve $50.0\\%$ accuracy on ImageNet.", + "bbox": [ + 75, + 595, + 468, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For CLIP ResNet, TinyCLIP-30M is only trained for 4 epochs in the first stage, inheriting $75\\%$ parameters from the original model. The results indicate that there is $0.5\\%$ slight decrease in top-1 accuracy on ImageNet, however, has $\\sim 2\\%$ improvements in retrieval tasks. TinyCLIP ResNet-19M inherits the weights from the 30M model, and it is trained for 12 epochs. It reduces the parameters by $50\\%$ while getting $2\\times$ inference speedup. But it gets $2.7\\%$ performance drops on ImageNet due to the large reduction in parameters.", + "bbox": [ + 75, + 747, + 470, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For OpenCLIP ViT-B/32, we distill it using our affin", + "bbox": [ + 96, + 883, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ity mimicking method with manual or automatic weight inheritance. We design a three-stage progressive distillation in which each stage compresses the model by $25\\%$ . The three stages are trained for 6, 16, and 16 epochs, respectively. In manual inheritance, TinyCLIP ViT-61M/32 obtains comparable performance compared to OpenCLIP ViT-B/32 [20], while reducing the parameters by $29\\%$ . On the other hand, automatic inheritance obtains better results than the manual one. For example, TinyCLIP ViT-63M/32 outperforms the manually inherited ViT-61M/32 model by more than $\\sim 1\\%$ on both classification and retrieval tasks. When auto-compressed the model by $3.9\\times$ , TinyCLIP ViT-22M/32 obtains the highest inference throughput of 5,504 and reaches $53.7\\%$ top-1 accuracy. Since YFCC-15M contains high-quality image-text pairs, if using it to replace parts of LAION data, our method can obtain additional improvements, especially for retrieval tasks.", + "bbox": [ + 496, + 410, + 893, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 679, + 653, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of affinity mmicking. As shown in Tab. 2, we present four different interaction modes. The affinity mimicking $\\mathcal{L}_1$ (i.e., $\\mathcal{L}_{distill}$ ) outperforms the contrastive loss $\\mathcal{L}_0$ by $2.1\\%$ in terms of top-1 accuracy on ImageNet [7]. The alignment in a visual-linguistic affinity space brings more similarity information. The accuracy of cross modalities interaction is lower than that of affinity mimicking by $0.2\\%$ . In this mode, the embedding features of student image and text do not have any interaction, but they are aligned to the teacher's embedding space. Single modality interaction has lower accuracy due to missing a loss to align student's image and text features in common embedding space.", + "bbox": [ + 496, + 703, + 893, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of weight inheritance. We verify the efficacy of", + "bbox": [ + 517, + 885, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21975", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f0383c934ae31221e4d5fbbe1701183c6ee60c37b95d2c3a72c277cd119f90d3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Interaction ModeLoss FormulaTop-1 Acc
Contrastive loss [46]L0=CE(<Is,Ts>,I)53.4 %
Affinity mimickingL1=CE(<Is,Ts>,<It,Tt>)55.5 %
Cross modalitiesL2=CE(<Is,Tt>,<It,Tt>)55.3 %
L3=CE(<It,Ts>,<It,Tt>)
Single modalityL4=CE(<Is,I>.I)19.2 %
L5=CE(<Ts,Tt>,I)
", + "bbox": [ + 81, + 88, + 467, + 179 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2094b009db3040487fccea5795494dd3488ccf20ac0fe53ab637b079f617025f.jpg", + "image_caption": [ + "Figure 4. The cosine similarity of the input and output embedding of each layer, and the zero-shot accuracy on ImageNet [7] after removing a layer for OpenCLIP ViT-B-32 [20, 46] pretrained on LAION-2B [49]." + ], + "image_footnote": [], + "bbox": [ + 86, + 306, + 267, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c41e5c3551a412d1d91981f688f44bca3fd4144f14f686a9202939f77b8359ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 306, + 460, + 398 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3d5b23431a66cd2bf8ebfc47429b3d831da05da80a12e7a9e3f61e3314dbd9be.jpg", + "table_caption": [ + "Table 2. The interaction of information across image $I$ and text $T$ modalities in student and teacher models. The subscripts $s$ and $t$ denote student and teacher, respectively. $\\mathcal{I}$ is an identity matrix. $CE$ is cross entropy function. TinyCLIP-ViT-40M/32 is trained with an interaction mode and manual weight inheritance on LAION-400M [50] for 1 epoch. The zero-shot accuracy on ImageNet [7] is reported." + ], + "table_footnote": [], + "table_body": "
Weight InheritanceIN-1K top1 acc (%)Flickr30kMSCOCO
I→T@1T→I@1I→T@1T→I@1
None37.850.233.429.716.6
Manual54.070.452.946.529.3
Automatic54.973.555.947.930.8
", + "bbox": [ + 78, + 479, + 470, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the proposed two kinds of weight inheritance in Tab. 3. Compared to the model without using weight inheritance, manual inheritance brings $16.2\\%$ accuracy improvements on ImageNet [7]. Automatic weight inheritance can further improve the accuracy by $0.9\\%$ . On image-text retrievals, the improvements of automatic inheritance is relatively higher, i.e., $1.4 - 3.1\\%$ over the manual method.", + "bbox": [ + 75, + 609, + 468, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The redundancy of language-image models. We analyze the redundancy of image encoder and text encoder by similar strategy [6, 2, 25] and zero-shot accuracy by removing a layer. As shown in Fig. 4, we observe a high correlation between cosine similarity and accuracy. Comparing with image encoder, text encoder has higher cosine similarity. It shows the redundancy of the text encoder is higher than that of the image encoder. Consequently, compressing the text encoder along the depth dimension can be achieved without significant performance degradation.", + "bbox": [ + 75, + 718, + 468, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further explore the redundancy by the proposed automatic weight inheritance. As shown in Fig. 5, for width", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a5866aefa3bf45f4445ee87701ace50ceb2739df2218dfcfdae2cb611321a2e.jpg", + "image_caption": [ + "Figure 5. Visualization of learnable masks at $50\\%$ sparsity on hidden dimension, attention heads and MLP intermediate dimension. The blocks in red color are removed. The original model is OpenCLIP ViT-B/32 [20]." + ], + "image_footnote": [], + "bbox": [ + 514, + 90, + 880, + 316 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fc701be2b2d27a0e1c144ab3dc3f5062fd2685e15e912a1a1e861b59664ec540.jpg", + "table_caption": [ + "Table 3. Manual inheritance vs. Automatic inheritance. The model OpenCLIP ViT-B/32 [20] is compressed by $50\\%$ with distillation on LAION-400M [50] for 1 epoch." + ], + "table_footnote": [], + "table_body": "
Pretrained Teacher ModelTeacher Acc.(%)Inherited RatioStudent Acc.(%)
w/o weight inheritance-036.2
CLIP ViT-B/32 [46]63.259 / 12652.4(+16.2)
OpenCLIP ViT-B/32 [20]62.959 / 12653.5(+17.3)
OpenCLIP ViT-B/16 [20]67.159 / 12452.8(+16.6)
OpenCLIP ViT-L/14 [20]75.359 / 39045.1(+8.9)
OpenCLIP ViT-H/14 [20]78.059 / 93541.1(+4.9)
", + "bbox": [ + 532, + 388, + 846, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study on inherited teacher model. The student model TinyCLIP ViT-40M/32 with 59M parameters is inherited from a teacher model, then trained without distillation on LAION-400M [50] for 1 epoch. The top-1 zero-shot accuracy on ImageNet [7] is reported.", + "bbox": [ + 498, + 488, + 890, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "analysis, it is observed that the embedding channels of the text encoder experience only a marginal reduction (e.g., 512 $\\rightarrow$ 511), whereas the embedding channels of the image encoder are reduced to a greater extent (e.g., $768\\to 526$ ). Regarding depth analysis, it was observed that a significant proportion of MHA layers and FFN channels could be removed from the text encoder, indicating redundancy within it. This finding suggests that the image encoder can be compressed along the width dimension, while the text encoder can be compressed along the depth dimension.", + "bbox": [ + 496, + 582, + 890, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of teacher models. We investigate which model is better for weight inheritance. As analyzed in Tab. 4, a strong teacher does not necessarily guarantee better results for weight inheritance. Instead, teacher models with similar architectures and higher performance offer a better choice. For instance, although ViT-H/14 ranks as the highest-performing teacher model, it lags behind other models in terms of weight inheritance.", + "bbox": [ + 496, + 734, + 890, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of multi-stage progressive distillation. Multi-stage progressive distillation shrinks the gap between the compressed model and the inherited model in each stage.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21976", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/700152c4b5b97d09ed66e1e251fef22500307f11cf798b3c7dbbe5993cbb1de7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodImage EncoderFood101CIFAR10CIFAR100SUN397Stanford CarsFGYC AircraftVOC207DTDOxford PetsCaltech101Flowers102MNISTFER2013STL10EuroSATRESISC45GTSRBKITTICountry211PCamHateful MemesRendered SST2ImageNet
Zero-shot performance
DeCLIP [31]ViT-B/3276.390.067.963.550.68.980.545.384.489.383.212.816.397.540.048.410.335.311.554.752.950.166.1
FILIP [70]ViT-B/3282.886.965.569.155.557.2-49.388.191.985.3----49.9------68.8
CLIP [46]ViT-B/3284.491.365.163.259.421.283.144.587.087.966.751.947.397.249.460.332.239.417.858.457.659.663.2
OpenCLIP [20]ViT-B/3280.990.770.666.879.316.682.254.486.590.166.037.442.395.651.657.642.031.614.850.152.952.362.9
TinyCLIP (Ours)ViT-45M/3278.392.271.866.079.923.083.253.885.388.760.762.545.296.252.261.841.116.916.551.553.552.661.2
Linear probe performance
CLIP [46]ViT-B/3288.895.180.576.681.852.087.776.590.093.096.999.069.298.397.090.585.366.227.883.966.770.876.1
TinyCLIP (Ours)ViT-45M/3284.097.486.068.985.439.085.471.289.993.482.199.566.997.697.588.198.579.716.190.458.462.573.2
", + "bbox": [ + 81, + 75, + 890, + 255 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dcdf0d9ad944d524a0da6d52e351cd32fde894c79f158c9b0d0789395a903396.jpg", + "table_caption": [ + "Table 5. Zero-shot and linear-probe classification top-1 accuracy on 23 datasets." + ], + "table_footnote": [], + "table_body": "
Number of StagesModel SizeEpochImageNet top1 acc (%)MSCOCO
I→TR@1T→IR@1
1100% → 25%247.039.423.1
2100% → 50%154.046.529.3
50% → 25%149.040.225.3
", + "bbox": [ + 78, + 292, + 465, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6035fb3cd23eef3dc6f6b2890861ce5b6c7fe80d3d2ac9fb61b2272097f96e64.jpg", + "image_caption": [ + "Figure 6. Training efficiency. OpenCLIP ViT-B/32 [20] is trained from scratch on LAION-400M [50] for 32 epochs. TinyCLIP ViT/32 is trained for 6+16 epochs using the proposed 2-stage progressive distillation, where the teacher is OpenCLIP ViT-B/32." + ], + "image_footnote": [], + "bbox": [ + 114, + 417, + 431, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare it with single-stage in Tab. 6. Using the same training epochs, the model trained with 2-stage surpasses that with 1-stage by $2.0\\%$ in terms of zero-shot accuracy on ImageNet. It demonstrates that multi-stage progressive compression is superior when transferring the knowledge of the pre-trained model to the compressed small model.", + "bbox": [ + 76, + 595, + 468, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training cost. We also study the distillation efficiency. As shown in Fig. 6, the original OpenCLIP ViT-B/32 [20] takes around 287 GPU days when trained on LAION-400M for 32 epochs. Our models with $75\\%$ and $50\\%$ parameters reach the similar accuracy as OpenCLIP counterparts, getting a speed up of $7.8\\times$ and $1.4\\times$ , respectively. The underlying reason is that our training benefits more from the affinity mimicking supervision and the good initialization using weight inheritance. This also proves that weight inheritance is capable of accelerating cross-modal distillation.", + "bbox": [ + 75, + 686, + 468, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Transfer Learning Results", + "text_level": 1, + "bbox": [ + 76, + 847, + 316, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Classification on 23 datasets. We also evaluate the performance of zero-shot and linear-probe classification on 23", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/97f99dcbe92ebafa254d62cfa85b6e17085aa98cadbcaf15cb8d1b4ef278a8a1.jpg", + "table_caption": [ + "Table 6. Ablation study on multi-stage progressive distillation. The model OpenCLIP ViT-B/32 [20] is compressed to $25\\%$ model size on LAION-400M [50] for 2 epochs." + ], + "table_footnote": [], + "table_body": "
ModelIN-V2IN-AIN-RObjectNetIN-Sketch
CLIP ViT-B/32 [46]56.031.669.429.942.3
OpenCLIP ViT-B/32 [20]55.121.773.428.949.4
TinyCLIP ViT-63M/32(Ours)55.722.874.131.250.8
TinyCLIP ViT-45M/32(Ours)52.619.871.529.148.8
", + "bbox": [ + 501, + 292, + 890, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Zero-shot robustness evaluation. CLIP [46] is trained on WIT-400M, OpenCLIP [20] and TinyCLIP is trained on LAION-400M [50]. Top-1 accuracy is reported.", + "bbox": [ + 498, + 367, + 890, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "datasets by the toolkit Elevater [26]. As shown in Tab. 5, our model wins on 7 datasets in zero-shot benchmark, and 9 datasets on linear-probe benchmark. The performance of our model is close to OpenCLIP ViT-B/32 [20], where the two models are both trained on Laion-400M [50].", + "bbox": [ + 496, + 424, + 890, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Zero-shot robustness evaluation. We verify the robustness of our models on 5 datasets, which are out of ImageNet distribution. As shown in Tab. 7, our TinyCLIP ViT-63M/32 outperforms OpenCLIP ViT-B/32 [20], and wins 3 datasets when compared with CLIP ViT-B/32 [46]. Our TinyCLIP ViT-45M/32 uses $50\\%$ fewer parameters yet still wins on IN-R and IN-Sketch, compared to CLIP ViT-B/32. This demonstrates the robustness of the distilled TinyCLIP models and the efficacy of our proposed distillation method.", + "bbox": [ + 496, + 501, + 892, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 643, + 625, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose TinyCLIP, a method for distilling large-scale vision-language pre-trained models. It mainly introduced two core techniques named affinity mimicking and weight inheritance. Extensive experiments and ablation studies have demonstrated the effectiveness of TinyCLIP, showing that it can largely reduce model size while maintaining competitive performance. In the future, we will explore ways to further improve cross-modal distillation efficiency on extremely small models.", + "bbox": [ + 496, + 670, + 890, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 814, + 658, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prof. Hongyang Chao was partially supported by NSFC (U1611461, 61672548) and NSFC (U22A2095). Prof. Xinggang Wang was partly supported by NSFC (No. 62276108).", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21977", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1", + "[2] Shaojie Bai, J Zico Kolter, and Vladlen Koltun. Deep equilibrium models. NeurIPS, 32, 2019. 7", + "[3] Arnav Chavan, Zhiqiang Shen, Zhuang Liu, Zechun Liu, Kwang-Ting Cheng, and Eric P Xing. Vision transformer slimming: Multi-dimension searching in continuous optimization space. In CVPR, 2022. 3", + "[4] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 2023. 2", + "[5] Wenliang Dai, Lu Hou, Lifeng Shang, Xin Jiang, Qun Liu, and Pascale Fung. Enabling multimodal generation on clip via vision-language knowledge distillation. In ACL, 2022. 2", + "[6] Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Lukasz Kaiser. Universal transformers. In ICLR, 2019. 7", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5, 6, 7", + "[8] Mingyu Ding, Bin Xiao, Noel Codella, Ping Luo, Jingdong Wang, and Lu Yuan. Davit: Dual attention vision transformers. In ECCV, 2022. 5, 6", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5", + "[10] Yifan Du, Zikang Liu, Junyi Li, and Wayne Xin Zhao. A survey of vision-language pre-trained models. In *IJCAI*, 2022. 2", + "[11] Zhiyuan Fang, Jianfeng Wang, Xiaowei Hu, Lijuan Wang, Yezhou Yang, and Zicheng Liu. Compressing visual-linguistic model via knowledge distillation. In ICCV, 2021. 2", + "[12] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In ICLR, 2019. 3", + "[13] Prakhar Ganesh, Yao Chen, Xin Lou, Mohammad Ali Khan, Yin Yang, Hassan Sajjad, Preslav Nakov, Deming Chen, and Marianne Winslett. Compressing large-scale transformer-based models: A case study on bert. ACL, 9, 2021. 1", + "[14] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. In NeurIPS, 2022. 2", + "[15] Yunhao Ge, Jie Ren, Ming-Hsuan Yang, Yuxiao Wang, Andrew Gallagher, Hartwig Adam, Laurent Itti, Balaji Lakshminarayanan, and Jiaping Zhao. Improving zero-shot generalization and robustness of multi-modal models. In NeurIPS ML Safety Workshop, 2022. 2", + "[16] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. IJCV, 129:1789-1819, 2021. 1" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Manish Gupta and Puneet Agrawal. Compression of deep learning models for text: A survey. TKDD, 2022. 2, 3", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5", + "[19] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. arXiv, 2015. 1, 2", + "[20] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. 1, 2, 5, 6, 7, 8", + "[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 2", + "[22] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding. In EMNLP, 2020. 2", + "[23] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 2", + "[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, ICLR, 2015. 5", + "[25] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. ICLR, 2020. 7", + "[26] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, and Jianfeng Gao. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. NeurIPS, 2022. 5, 8", + "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2", + "[28] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2", + "[29] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 2", + "[30] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. arXiv preprint arXiv:2212.00794, 2022. 2", + "[31] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 8", + "[32] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21978", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. arXiv preprint arXiv:2210.04150, 2022. 2", + "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6", + "[34] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. In ICLR, 2019. 3", + "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2018. 5", + "[36] Jiawei Han Jamie Callan Luyu Gao, Yunyi Zhang. Scaling deep contrastive learning batch size under memory limited setup. In RepL4NLP, 2021. 5", + "[37] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 5191-5198, 2020. 2", + "[38] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. 6", + "[39] Basil Mustafa, Carlos Riquelme Ruiz, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts. In NeurIPS, 2022. 1", + "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5", + "[41] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 2", + "[42] Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, and Youliang Yan. Clipping: Distilling clip-based models for video-language understanding. OpenReview, 2023. 2", + "[43] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, et al. Combined scaling for open-vocabulary image classification. arXiv, 2021. 1, 2", + "[44] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015, 6", + "[45] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020. 2", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5, 6, 7, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2", + "[48] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 2", + "[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 5, 7", + "[50] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2, 5, 6, 7, 8", + "[51] Dachuan Shi, Chaofan Tao, Ying Jin, Zhendong Yang, Chun Yuan, and Jiaqi Wang. Upop: Unified and progressive pruning for compressing vision-language transformers. arXiv preprint arXiv:2301.13741, 2023. 2", + "[52] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. 1", + "[53] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016. 5, 6", + "[54] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2", + "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6", + "[56] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 2", + "[57] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2", + "[58] Yulong Wang, Xiaolu Zhang, Lingxi Xie, Jun Zhou, Hang Su, Bo Zhang, and Xiaolin Hu. Pruning from scratch. In AAAI, 2020. 3", + "[59] Zhecan Wang, Noel Codella, Yen-Chun Chen, Luowei Zhou, Jianwei Yang, Xiyang Dai, Bin Xiao, Haoxuan You, Shih-Fu Chang, and Lu Yuan. Clip-td: Clip targeted distillation for vision-language tasks. arXiv preprint arXiv:2201.05729, 2022. 2", + "[60] Ziheng Wang, Jeremy Wohlwend, and Tao Lei. Structured pruning of large language models. In EMNLP, 2020. 4" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21979", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. In ICLR, 2021. 2", + "[62] Yixuan Wei, Yue Cao, Zheng Zhang, Zhuliang Yao, Zhenda Xie, Han Hu, and Baining Guo. icar: Bridging image classification and image-text alignment for visual recognition. In NeurIPS, 2022. 2", + "[63] Ross Wightman. Pytorch image models, 2019. 5", + "[64] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In ECCV, 2022. 1, 2", + "[65] Mengzhou Xia, Zexuan Zhong, and Danqi Chen. Structured pruning learns compact and accurate models. In ACL, 2022. 2, 3, 4", + "[66] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In ECCV, 2022. 2", + "[67] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. 2022. 2", + "[68] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In CVPR, 2022. 2", + "[69] Shusheng Yang, Yixiao Ge, Kun Yi, Dian Li, Ying Shan, Xiaohu Qie, and Xinggang Wang. Masked visual reconstruction in language semantic space. arXiv preprint arXiv:2301.06958, 2023. 6", + "[70] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2022. 8", + "[71] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. In NeurIPS, 2022. 1", + "[72] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1, 2, 5, 6", + "[73] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 1", + "[74] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In CVPR, pages 18123-18133, 2022. 2", + "[75] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. In NeurIPS, 2022. 2", + "[76] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 2", + "[77] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 6" + ], + "bbox": [ + 501, + 92, + 890, + 148 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21980", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/7a0fee57-111a-443f-8a63-68f0c3d1c0b8_model.json b/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/7a0fee57-111a-443f-8a63-68f0c3d1c0b8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bcd45522275cc2dfaa3ef2968e654825efea9f8c --- /dev/null +++ b/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/7a0fee57-111a-443f-8a63-68f0c3d1c0b8_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.13, + 0.872, + 0.154 + ], + "angle": 0, + "content": "TinyCLIP: CLIP Distillation via Affinity Mimicking and Weight Inheritance" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.187, + 0.897, + 0.245 + ], + "angle": 0, + "content": "Kan Wu\\(^{1,\\ast}\\), Houwen Peng\\(^{3,\\ast,\\dagger}\\), Zhenghong Zhou\\(^{2,\\ast}\\), Bin Xiao\\(^{3}\\), Mengchen Liu\\(^{3}\\), Lu Yuan\\(^{3}\\), Hong Xuan\\(^{3}\\), Michael Valenzuela\\(^{3}\\), Xi (Stephen) Chen\\(^{3}\\), Xinggang Wang\\(^{2}\\), Hongyang Chao\\(^{1}\\), Han Hu\\(^{3}\\), Sun Yat-sen University, Huazhong University of Science & Technology, Microsoft" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.286, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.474, + 0.712 + ], + "angle": 0, + "content": "In this paper, we propose a novel cross-modal distillation method, called TinyCLIP, for large-scale language-image pre-trained models. The method introduces two core techniques: affinity mimicking and weight inheritance. Affinity mimicking explores the interaction between modalities during distillation, enabling student models to mimic teachers' behavior of learning cross-modal feature alignment in a visual-linguistic affinity space. Weight inheritance transmits the pre-trained weights from the teacher models to their student counterparts to improve distillation efficiency. Moreover, we extend the method into a multi-stage progressive distillation to mitigate the loss of informative weights during extreme compression. Comprehensive experiments demonstrate the efficacy of TinyCLIP, showing that it can reduce the size of the pre-trained CLIP ViT-B/32 by \\(50\\%\\), while maintaining comparable zero-shot performance. While aiming for comparable performance, distillation with weight inheritance can speed up the training by \\(1.4 - 7.8 \\times\\) compared to training from scratch. Moreover, our TinyCLIP ViT-8M/16, trained on YFCC-15M, achieves an impressive zero-shot top-1 accuracy of \\(41.1\\%\\) on ImageNet, surpassing the original CLIP ViT-B/16 by \\(3.5\\%\\) while utilizing only \\(8.9\\%\\) parameters. Finally, we demonstrate the good transferability of TinyCLIP in various downstream tasks. Code and models will be open-sourced at aka.ms/tinyclip." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.733, + 0.21, + 0.75 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Large-scale language-image pretraining, e.g., CLIP [46], has recently gained significant attention due to its remarkable zero-shot transfer capability [46] and unprecedented performance in text-to-image generation [47]. Due to complex nature of vision and language, current approaches often resort to utilizing huge amounts of parameters to endow models with cross-modal capabilities [46, 21, 71, 72, 43, 39," + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.263, + 0.705, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.264, + 0.861, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.53 + ], + "angle": 0, + "content": "Figure 1. Comparison of the OpenCLIP [20] and TinyCLIP. TinyCLIP is pre-trained and distilled on LAION-400M [50] using OpenCLIP ViT-B/32 [20] as the teacher, whose zero-shot top-1 accuracy is \\(65.6\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.895, + 0.599 + ], + "angle": 0, + "content": "1]. This in turn leads to high costs in terms of storage, memory, and computation time for these models, which motivates the need for model compression to make them smaller and faster for real-world applications [64, 13]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.896, + 0.839 + ], + "angle": 0, + "content": "As a core compression technique, knowledge distillation has been extensively studied and applied in single-modal settings [19, 16]. However, its potential for multi-modality remains underexplored. Unlike single-modal models, the distillation of language-image cross-modal models poses distinct challenges. First, CLIP-like language-image models commonly consist of two branches: an image encoder and a text encoder [21, 72, 52, 73]. When distilling such multi-branch models, it is crucial to consider the interaction of information across the different modality branches in both the teacher and student models. Second, the original CLIP [46] models are pre-trained on 400 million image-text pairs for 32 epochs, which requires thousands of GPU days, making distillation a significant challenge when computational resources are limited. Is there any way to reduce the cost of CLIP distillation?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To tackle these challenges, we present a novel cross-modal distillation method dubbed TinyCLIP, which introduces two key techniques: affinity mimicking and weight inheritance. In contrast to the methods that rely on either" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.468, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution. Kan and Zhenghong were interns of Microsoft." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.888, + 0.358, + 0.9 + ], + "angle": 0, + "content": "† Corresponding: houwen.peng@microsoft.com" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21970" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "image or text features for distillation, we empirically show that distilling knowledge in an image-text affinity space is more effective. Specifically, we leverage the cosine similarity of the image and text embeddings in the teacher model to facilitate the distillation of the student model, allowing the student to mimic the teacher's visual-linguistic feature alignment. We refer to this process as affinity mimicking." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.438 + ], + "angle": 0, + "content": "To improve distillation efficiency, we introduce weight inheritance, a technique transferring the pre-trained weights from teacher models to their student counterparts. Since inheriting weights provides a good initialization for the student models, the distillation progress can be largely accelerated. The key challenge of weight inheritance lies in determining which weights are more advantageous. To address this issue, we introduce two solutions: manual and automatic inheritance. We surprisingly found that a simple manual selection of \\( k \\)-dimension or \\( k \\)-layer weights from the teacher model can yield satisfactory results for CLIP distillation. On the other hand, we also introduce learnable masks to automatically identify the most important weights from the teacher model. The masks are imposed independently on the vision and language branches, enabling them to capture the differences across modalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.62 + ], + "angle": 0, + "content": "Moreover, we extend the proposed weight inheritance to a multi-stage progressive procedure, where each subsequent stage automatically inherits the important weights from the preceding stages. We observed that when the teacher model exhibits higher performance and shares a similar architecture with the student, weight inheritance can provide better results. This is because significant architecture differences may undermine the learned weights when transmitting from the teacher to the student. Therefore, we break the inheritance into multiple stages, allowing the student model in each stage to share a more similar structure with the predecessor teacher and inherit the weights progressively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.471, + 0.862 + ], + "angle": 0, + "content": "Our experiments show that TinyCLIP delivers competitive models at all levels of speedups and model sizes in ImageNet zero-shot evaluation and various downstream tasks. As shown in Fig. 1, pre-trained on the same LAION-400M dataset [50], TinyCLIP-ViT using 63M parameters achieves \\(61.4\\%\\) zero-shot top-1 accuracy on ImageNet [7], being \\(2.0 \\times\\) smaller and \\(1.5 \\times\\) faster than the OpenCLIP [20] model (62.9% accuracy with 126M parameters). Meanwhile, TinyCLIP-ResNet with 38M parameters obtains \\(56.4\\%\\) zero-shot top-1 accuracy, being \\(2.1 \\times\\) smaller and \\(2.0 \\times\\) faster than CLIP ResNet-50 (59.6% accuracy with 76M parameters). Moreover, our method can speed up the original OpenCLIP training by \\(1.4 \\times -7.8 \\times\\) while achieving similar performance. Also, TinyCLIP demonstrates good transfer capacities in downstream scenarios. In summary, the contributions of this work are two-fold:" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We propose a new cross-modal distillation approach to unleash the capacity of small CLIP models, fully lever" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "aging large-scale models as well as pre-training data. To our best knowledge, this is the first work exploring the pre-training distillation of language-image models." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.149, + 0.894, + 0.224 + ], + "angle": 0, + "content": "- We present state-of-the-art language-image pre-trained models at small scale, striking the best trade-off between speed and accuracy. Extensive experiments demonstrate the superiority and good generalization ability of the small models in various downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.239, + 0.642, + 0.256 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Language-Image Pre-training has achieved remarkable progress over the past few years [10, 4]. In particular, contrastive language-image pre-training demonstrates very impressive zero-shot transfer and generalization capacities [45, 46, 21, 72, 61, 43, 74]. One of the most representative works is CLIP [46]. A large amount of follow-up works have been proposed to optimize the pre-training framework [27, 76, 15, 68, 62, 14, 28, 67]. Meanwhile, a line of works leverage the pre-trained models for downstream tasks, such as open-vocabulary detection and segmentation [29, 75, 66, 32], video recognition [56, 23], and text-to-image generation [47, 41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.893, + 0.628 + ], + "angle": 0, + "content": "Most recently, there are a few works attempting to scale language-image pre-trained models. BEiT-3 [57] uses a multiway transformer to scales up multi-modal pre-trained models to 1.9B parameters. FLIP [30] scales image-text pre-training via masked modeling, enabling CLIP to enjoy faster training speed while getting better performance. On the other hand, there are also few works on scaling down CLIP models [51]. However, they all focus on specific tasks, such as generation [5, 51] and video content understanding [42]. In contrast, our work concentrates on language-image pre-training, which is the first work on CLIP compression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Knowledge Distillation in a teacher-student framework [19] has been widely used to transfer knowledge from large models to small ones. Distillation in single modality, such as vision [54, 64, 37] or language [48, 22], has been extensively studied. However, the exploration of cross-modality distillation is relatively limited. A few early works exploring cross-modal distillation only focus on specific tasks, such as VQA and image caption [5, 11, 59], limiting the methodologies to be applied in general downstream tasks. In contrast, this work pays attention to general cross-modal pre-training distillation, in which the distilled models can be transferred to various downstream scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.902 + ], + "angle": 0, + "content": "It is worth noting that our proposed weight inheritance strategy shares similar spirits to model pruning techniques [17, 65, 51]. Both approaches involve the identification of important weights while removing redundant ones. However, they have three fundamental differences. 1) Previous pruning methods predominantly focus on single-modality" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21971" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.273 + ], + "angle": 0, + "content": "models [17, 65, 3], whereas our approach shifts attention to multi-modality. Since different modalities exhibit varying redundancy characteristics, it is necessary to consider these differences when selecting important weights from each modality. 2) Unlike previous pruning methods which have emphasized the importance of the pruned architecture [12, 34, 58], we argue that the weights inherited from the original models are also highly beneficial, particularly when compressing CLIP-like language-image models. 3) Our method leverages a progressive multi-stage process, which enables the inherited weights to accelerate the convergence of small models during training." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.286, + 0.169, + 0.303 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.47, + 0.388 + ], + "angle": 0, + "content": "In this section, we propose TinyCLIP, a simple and effective method for distilling large-scale language-image pretraining models, such as CLIP. It consists of three components: affinity mimicking (§3.1), weight inheritance (§3.2) and multi-stage progressive distillation (§3.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.396, + 0.391, + 0.413 + ], + "angle": 0, + "content": "3.1. Distillation with Affinity Mimicking" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.629 + ], + "angle": 0, + "content": "A language-image pre-training model often consists of two branches, an image encoder to extract visual representations and a text encoder to extract textual representations. The visual and textual representations are linearly projected into a cross-modal common space by minimizing a contrastive loss \\(\\mathcal{L}_0\\). The supervision signal can be viewed as an identity matrix, in which diagonal values are set to 1 for positive image-text pairs, and all others are set to 0 for negative pairs. This signal disregards the similarity between negative pairs, which prevents the model from acquiring a nuanced understanding of the intricate relationships among negative pairs. Therefore, we introduce affinity mimicking to enable student models to learn similarities between negative pairs from teacher models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.47, + 0.812 + ], + "angle": 0, + "content": "More specifically, as shown in Fig. 2, we consider two types of affinity distillation losses: image-to-language loss \\(\\mathcal{L}_{I2T}\\) and language-to-image loss \\(\\mathcal{L}_{T2I}\\). The former learns the alignment between the teacher and student models based on the image-to-language affinity \\(A_{I2T}\\), which represents the affinity scores of an image with all the text descriptions in a batch (the yellow row highlighted in Fig. 2). The latter refers to the alignment of language-to-image affinity \\(A_{T2I}\\), which compares a text description with all the images to obtain a matching score (the blue column in Fig. 2). The combination of these two losses formulates our affinity mimicking distillation, which is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.824, + 0.468, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {d i s t i l l}} = \\mathcal {L} _ {I 2 T} + \\mathcal {L} _ {T 2 I}, \\tag {1} \\\\ = C E \\left(A _ {I 2 T} ^ {s}, A _ {I 2 T} ^ {t}\\right) + C E \\left(A _ {T 2 I} ^ {s}, A _ {T 2 I} ^ {t}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Here, \\( CE \\) represents the cross-entropy loss. The superscripts \\( s \\) and \\( t \\) denote the student and teacher models, re" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.089, + 0.865, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.3, + 0.892, + 0.328 + ], + "angle": 0, + "content": "Figure 2. Affinity mimicking for language-image models. The loss includes image-to-text loss (yellow) and text-to-image loss (blue)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.354, + 0.892, + 0.37 + ], + "angle": 0, + "content": "spectively. The elements in \\(A_{I2T}\\) and \\(A_{T2I}\\) are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.392, + 0.892, + 0.427 + ], + "angle": 0, + "content": "\\[\nA _ {I 2 T} (i, j) = \\frac {\\exp \\left(I _ {i} \\cdot T _ {j} / \\tau\\right)}{\\sum_ {k \\in \\mathcal {B}} \\exp \\left(I _ {i} \\cdot T _ {k} / \\tau\\right)}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.429, + 0.892, + 0.464 + ], + "angle": 0, + "content": "\\[\nA _ {T 2 I} (i, j) = \\frac {\\exp \\left(I _ {i} \\cdot T _ {j} / \\tau\\right)}{\\sum_ {k \\in \\mathcal {B}} \\exp \\left(I _ {k} \\cdot T _ {j} / \\tau\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.893, + 0.595 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a temperature parameter, \\(I_{i}\\) is the feature embedding of \\(i\\)-th image in the batch \\(\\mathcal{B}\\), \\(T_{j}\\) is the feature of \\(j\\)-th text description in \\(\\mathcal{B}\\). Affinity mimicking allows student models to imitate the behavior of the large model in learning visual-linguistic alignment. In contrast to previous methods that rely on either image or text features for distillation, we show that distilling knowledge in this image-text affinity space is more effective." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.813, + 0.621 + ], + "angle": 0, + "content": "3.2. Distillation with Weight Inheritance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.795 + ], + "angle": 0, + "content": "The original CLIP models are pre-trained on 400 million image-text pairs for 32 epochs, taking thousands of GPU days. This presents a significant cost challenge for distillation. To improve the training efficiency, we introduce weight inheritance, a technique that inherits the important weights from the well-trained large teacher models to smaller student models. The key challenge of inheriting weights is identifying important weights from the large amount of weights of the teacher. We propose two approaches to select the important weights: manual weight inheritance and automatic weight inheritance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Manual Inheritance. For manual weight inheritance, we first analyze the redundancy of existing CLIP pre-trained models. Fig. 4 shows that the text encoder displays more redundancy in depth (layer-wise), while the image encoder exhibits more redundancy in width (channel-wise). Based upon this discovery, we adopt the approach of uniformly selecting \\( k \\) layers of the text branch and directly taking the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21972" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.167 + ], + "angle": 0, + "content": "front \\(k\\) channels of the image branch to select the important weights from large teacher models. These weights will serve as the initialization for small models during distillation. We surprisingly found that such a simple manual selected weights can largely accelerate CLIP distillation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.545 + ], + "angle": 0, + "content": "Automatic Inheritance. Although manual inheritance can be considered an effective approach for model compression, it does present the drawback of necessitating prior knowledge in order to determine which weights to inherit. This limitation may render the technique impractical for application to diverse models. To solve this issue, we present an automatic weight inheritance scheme. Inspired by structured pruning in large language models [65], we introduce learnable masks to identify weight importance. Considering the difference across modalities, the learnable masks are imposed on vision and language branches independently, as visualized in Fig. 3. An overall sparsity constraint is introduced to guarantee the selected number of important weights to meet our compression requirements. Without loss of generality, here we use the transformer architecture as an example to introduce the detailed procedure of weight inheritance. Specifically, a standard transformer block contains a multi-head attention (MHA) layer and a feed-forward network (FFN). To capture the importance of weights in a fine-grained level, we introduce two mask variables \\(\\mathbf{m}_{\\mathrm{head}}\\), \\(\\mathbf{m}_{\\mathrm{int}} \\in \\{0,1\\}\\) to identify the redundant attention heads in MHA and neurons in FFN respectively, while keeping the important ones. These two kinds of masks are imposed on the activation of attention heads and the intermediate layer of FFN, which is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.55, + 0.47, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M H A} (\\mathbf {X}) = \\sum_ {h = 1} ^ {N _ {H}} \\mathbf {m} _ {\\text {h e a d}} ^ {h} \\cdot \\operatorname {A t t n} _ {\\mathbf {W} _ {Q} ^ {h}, \\mathbf {W} _ {K} ^ {h}, \\mathbf {W} _ {V} ^ {h}, \\mathbf {W} _ {O} ^ {h}} (\\mathbf {X}), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.594, + 0.469, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\operatorname {F F N} (\\mathbf {X}) = \\operatorname {G e L U} \\left(\\mathbf {X} \\mathbf {W} _ {U}\\right) \\cdot \\operatorname {d i a g} \\left(\\mathbf {m} _ {\\text {i n t}}\\right) \\cdot \\mathbf {W} _ {D}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.469, + 0.828 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}\\) is the layer input, \\(\\mathbf{W}_Q^h\\), \\(\\mathbf{W}_K^h\\), \\(\\mathbf{W}_V^h\\), \\(\\mathbf{W}_O^h \\in \\mathbb{R}^{d \\times d_h}\\) denote the query, key, value and output matrices in MHA respectively, and \\(\\mathbf{W}_U \\in \\mathbb{R}^{d \\times d_f}\\) and \\(\\mathbf{W}_D \\in \\mathbb{R}^{d_f \\times d}\\) represent the parameters of FFN layers. Here, \\(d\\) denotes the hidden dimension (e.g., 768), and \\(d_h = d / N_H\\) denotes the output dimension of each head (e.g., 64), where \\(N_H\\) is the total number of heads. The trainable masks \\(\\mathbf{m}_{\\mathrm{head}}\\) and \\(\\mathbf{m}_{\\mathrm{int}}\\) serve as importance indicators for MHA and FFN layers. Moreover, to further learn the importance of embedding dimensions in transformer, we introduce an additional mask \\(\\mathbf{m}_{\\mathrm{embed}} \\in \\{0,1\\}\\). This mask is shared across all layers because each dimension in the hidden representation is connected to the corresponding dimension in the subsequent layer through a residual connection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.83, + 0.469, + 0.875 + ], + "angle": 0, + "content": "During automatic inheritance, there are two losses used for optimizing the masks and the model jointly, i.e., a sparsity loss and a distillation loss defined in Eq. (1):" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.882, + 0.469, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {d i s t i l l}} + \\mathcal {L} _ {\\text {s p a r s i t y}}, \\tag {6}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.706, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.27, + 0.713, + 0.285 + ], + "angle": 0, + "content": "10 Learnable mask/manual mask" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.09, + 0.871, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.271, + 0.868, + 0.285 + ], + "angle": 0, + "content": "Model sparsity \\(=\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.29, + 0.892, + 0.374 + ], + "angle": 0, + "content": "Figure 3. Weight inheritance of CLIP. Given a large pre-trained CLIP model, the inherited weights are selected by masks. The weights whose mask is 0 will be removed. For the manual way, the masks are set to 1 in the first k-dimension or uniform k-layer, the rest is set to 0. For the automatic way, the masks are learned jointly with the model under the sparsity constraint." + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.402, + 0.891, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s p a r s i t y}} = \\lambda \\cdot (p - q) + \\beta \\cdot (p - q) ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.491 + ], + "angle": 0, + "content": "where \\(\\lambda, \\beta\\) are learnable multipliers that guarantee \\(p = q[60, 65]\\). Here, \\(q\\) is the target compression rate, \\(p\\) is the overall compression rate of learnable masks for the model, including image encoder and text encoder:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.499, + 0.891, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p = \\frac {1}{M _ {t} + M _ {t}} \\sum_ {i m g, t x t} (4 \\cdot d _ {h} \\cdot \\sum_ {i} ^ {L} \\sum_ {j} ^ {N _ {H}} \\sum_ {k} ^ {d} \\mathbf {m} _ {\\text {h e a d}} ^ {i, j} \\cdot \\mathbf {m} _ {\\text {e m b e d}} ^ {k} \\\\ + 2 \\cdot \\sum_ {i} ^ {L} \\sum_ {j} ^ {d _ {f}} \\sum_ {k} ^ {d} \\mathbf {m} _ {\\text {i n t}} ^ {i, j} \\cdot \\mathbf {m} _ {\\text {e m b e d}} ^ {k}), \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.597 + ], + "angle": 0, + "content": "where \\( M_{i} \\) and \\( M_{t} \\) represent the full model size of the image encoder and text encoder, respectively, \\( L \\) is the total number of transformer layers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.808, + 0.623 + ], + "angle": 0, + "content": "3.3. Progressive Multi-Stage Distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.795 + ], + "angle": 0, + "content": "When attempting to achieve a high target sparsity, i.e., \\(>70\\%\\), compressing the model in a single stage can lead to a significant reduction in accuracy and even result in convergence failure. This is due to the fact that most weights of the large model are directly discarded, including those that are important for ensuring model quality and convergence. As a solution, we propose a multi-stage progressive distillation method to achieve a high compression rate without seriously sacrificing accuracy. In each stage, we use a modest degree of compression, e.g., \\(25\\%\\), to avoid large loss of performance and make training stable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Specifically, each stage includes weight inheritance and affinity mimicking, as described in Algorithm 1. During the weight inheritance phase, the weights of the large model are gradually reduced until the target sparsity and key weights are retained for the small model. After that, the small model is distilled with affinity mimicking, which involves transferring the visual-linguistic affinity knowledge from" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21973" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.088, + 0.096, + 0.4, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: TinyCLIP: Cross-Modal Distillation" + }, + { + "type": "algorithm", + "bbox": [ + 0.082, + 0.112, + 0.47, + 0.462 + ], + "angle": 0, + "content": "Input: a pre-trained CLIP model \\(f(x;\\theta_0)\\) with weight \\(\\theta_0\\) , target compression rate \\(q\\) total iteration steps \\(N\\) 1 Initialize the number of stages \\(G\\) , each stage steps \\(L\\) , target compression rate \\(q_{i}\\) of stage i, weight inheritance steps \\(L_{M}\\) input \\(x\\) , learnable mask \\(M = \\{\\mathbf{m}_{\\mathrm{head}},\\mathbf{m}_{\\mathrm{int}},\\mathbf{m}_{\\mathrm{embed}}\\} ;^1\\) \n2 for i in [1,...,G] do \n3 // Increase mask compression rate, retain important weights \n4 for j in [1,...,LM] do \n5 Update target compression rate \\(^2 q\\) of current step based on \\(q_{i}\\) and step \\(j\\) . \n6 Calculate distillation loss between origin teacher model \\(f(x;\\theta_0)\\) and masked student model \\(f(x;\\theta_{i - 1}\\odot M_i)\\) . \n7 Calculate sparsity loss between learnable mask compression rate \\(^3 p\\) and target compression rate \\(q\\) . \n8 Optimize \\(\\theta_{i - 1}\\) and \\(M_{i}\\) . \n9 end \n10 \\(\\theta_{i}\\gets \\theta_{i - 1}\\odot M_{i}\\) //Inherit weight with masks \n11 // Cross-Modal distillation \n12 for j in \\([L_M + 1,\\dots ,L]\\) do \n13 Calculate distillation loss between origin teacher model \\(f(x;\\theta_0)\\) and compressed model \\(f(x;\\theta_i)\\) . \n14 Optimize \\(\\theta_{i}\\) . \n15 end \n16 end \n17 Return compressed model \\(f(x;\\theta_G)\\) at compression rate \\(q\\) \n1 Each stage compress \\(25\\%\\) \\(q_{i} = 1 - i\\cdot 25\\%\\) \\(G = \\lceil (1 - q) / 0.25\\rceil\\) \\(L = N / G,L_{M} = 3000,M\\) is initialized to 1, \\(q_{i}\\) can be adjusted according to different models. \n2 Target compression rate \\(q\\) increases linearly from O to \\(q_{i}\\) with step \\(j\\) . \n\\(p\\) is calculated by masks \\(M\\) in Eq. (8)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.523 + ], + "angle": 0, + "content": "the teacher to the student. This procedure is repeated until the target sparsity level is achieved." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.533, + 0.21, + 0.55 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.553, + 0.295, + 0.57 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.473, + 0.863 + ], + "angle": 0, + "content": "Architecture. The original CLIP employs a transformer-based model as its text encoder, while its image encoder has two variations, i.e., ViT [9] and modified ResNet [18], covering both transformer-based and CNN-based architectures. For weight inheritance, we inherited the weights from two model variants: OpenCLIP ViT-B/32 or ViT-B/16 [20] and CLIP ResNet-50 [46], which are pre-trained on LAION-2B [49] and WIT-400M [46], respectively1. For distillation, we use OpenCLIP ViT-B/32 [20] pre-trained on LAION-2B [49] as the teacher model, since it achieves a high zero-shot accuracy of \\(65.6\\%\\) on ImageNet [7] with a high inference throughput. Besides, we also extend our method to DaViT [8], which is a hybrid architecture. We re-produce Florence-DaViT-5M [72, 8] on LAION-400M as a baseline. Then we distill the same model using affinity mimicking and weight inheritance. The teacher model is Florence-DaViT-D3, which is pre-trained on FLD-900M [72] and achieves a top-1 accuracy of \\(78.0\\%\\) on ImageNet [7] with 128M parameters." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.864, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Affinity Mimicking. As defined in Tab. 2, we explore" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "different interaction schemes between teacher and student models across modalities. The contrastive loss \\(\\mathcal{L}_0\\) and affinity mimicking \\(\\mathcal{L}_1\\) (i.e., \\(\\mathcal{L}_{\\text{distill}}\\)) serve as the basic ones, which can be combined with other losses. The weight of each loss is set to 1 when combined. The temperature parameter \\(\\tau\\) in Eq. (3) is set to \\(\\frac{1}{50}\\) by default. The sensitivity analysis of \\(\\tau\\) is presented in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.893, + 0.394 + ], + "angle": 0, + "content": "Weight Inheritance. We compress OpenCLIP ViT-B [20] and CLIP ResNet-50 [46] using the proposed multi-stage progressive distillation. In each stage, we apply manual or automatic weight inheritance. In manual inheritance, we compress the image encoder and the text encoder separately, by reducing the width of image encoder and the depth of text encoder. In automatic inheritance, we set the target sparsity for the whole model. In each stage, the trainable masks are initialized to 1 and updated in the first 3,000 training iterations, where the optimizer is AdamW [35, 24] with a constant learning rate of 0.01 and no weight decay. The learnable multipliers \\(\\lambda, \\beta\\) defined in Eq. (7) are both initialized to 0.01." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.893, + 0.637 + ], + "angle": 0, + "content": "Training Settings. We train our models on two public datasets, namely LAION-400M [50] and YFCC-15M [53]. On LAION-400M, the models are compressed in 3 stages, including \\(100\\%\\) to \\(75\\%\\) parameters for 6 epochs, \\(75\\%\\) to \\(50\\%\\) parameters for 16 epochs, and \\(50\\%\\) to \\(25\\%\\) parameters for 16 epochs. On YFCC-15M, it contains 2 compression stages, where the training epochs are both 25 from \\(100\\%\\) to \\(50\\%\\) parameters, and \\(50\\%\\) to \\(10\\%\\). We follow the hyper-parameter of CLIP [46] except that the learning rate is set to \\(10^{-4}\\) when using weight inheritance. The details are shown in the supplementary material. Unless otherwise specified, our ablation studies utilize the same settings, except for training the models only for 1 epoch with a learning rate of \\(5 \\times 10^{-4}\\). All models are trained on 32 Nvidia V100 or A100, and implemented with PyTorch [40], OpenCLIP [20], gradient cache [36], and Timm library [63]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.895, + 0.834 + ], + "angle": 0, + "content": "Evaluation Settings. The models are evaluated on multiple benchmarks. In zero-shot transfer evaluation and robustness evaluation, we follow the same prompt engineering in CLIP [46], where 80 text templates per class are used for ImageNet [7]. In linear probe, we use Elevater toolkit [26] for evaluation, where a classification head is trained for 50 epochs with searched hyper-parameters. Note that we do not count the number of parameters in the text embedding layer. It is a look-up table whose parameter size is the same as the models with the same hidden dimension and vocabulary size. The inference throughput is measured on Nvidia V100 with CUDA 11 and PyTorch v1.12 [40], where the batch size is 1,024." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.857, + 0.863 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art Models" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We compare our TinyCLIP with state-of-the-art models in Tab. 1. The zero-shot evaluation on ImageNet-1K [7]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.429, + 0.901 + ], + "angle": 0, + "content": "1The models can be downloaded from OpenCLIP and OpenAI." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21974" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.077, + 0.891, + 0.354 + ], + "angle": 0, + "content": "
MethodImageEncoderText Encoder#Params (M)Image+TextMACs(G)Throughput(pairs/s)TrainingdatasetsIN-1Ktop1 acc (%)Flickr30kMSCOCO
depthwidthI→T@1T→I@1I→T@1T→I@1
Training data: 15M / 20M
RILS [69]ViT-B/161251286 + 3820.5818LAION-20M45.045.134.932.225.5
MaskCLIP [77]ViT-B/161251286 + 3820.5818LAION-20M46.664.948.138.524.8
SLIP [38]ViT-B/161251286 + 3820.5818YFCC-15M42.857.640.131.120.3
CLIP [46, 38]ViT-B/161251286 + 3820.5818YFCC-15M37.651.632.226.517.1
TinyCLIP (Ours)ViT-39M/16651239+19(2.1×)9.5(2.2×)1,469(1.8×)YFCC-15M63.584.466.754.938.9
TinyCLIP (Ours)ViT-8M/1632568+3(11.3×)2.0(10.3×)4,150(5.1×)YFCC-15M41.162.342.336.221.5
Training data: 400M
Florence [72]DaViT-5M62565+51.12,980LAION-400M45.061.241.536.220.9
TinyCLIP (Ours)DaViT-5M62565+51.12,980LAION-400M50.066.248.540.724.5
CLIP [46]ResNet-1011251256 + 3812.81,161WIT-400M62.278.159.249.330.7
CLIP [46]ResNet-501251238 + 389.11,549WIT-400M59.681.258.250.828.3
TinyCLIP (Ours)ResNet-30M951230+29(1.3×)6.9(1.3×)1,811(1.2×)LAION-400M59.181.161.252.733.9
TinyCLIP (Ours)ResNet-19M651219+19(2.0×)4.4(2.1×)3,024(2.0×)LAION-400M56.476.258.348.930.9
OpenCLIP [20]ViT-B/321251288 + 387.42,452LAION-2B65.784.766.856.939.3
CLIP [46]ViT-B/321251288 + 387.42,452WIT-400M63.280.159.851.230.6
OpenCLIP [20]ViT-B/321251288 + 387.42,452LAION-400M62.979.362.053.335.4
TinyCLIP (Ours)ViT-61M/32951261+29(1.4×)5.3(1.4×)3,191(1.3×)LAION-400M62.178.663.353.935.9
TinyCLIP (Ours)ViT-40M/32651240+19(2.1×)3.5(2.1×)4,641(1.9×)LAION-400M59.777.358.949.833.1
TinyCLIP (Ours)ViT-63M/32autoauto63+31(1.3×)5.6(1.3×)2,905(1.2×)LAION-400M63.983.264.455.537.6
TinyCLIP (Ours)ViT-45M/32autoauto45+18(2.0×)3.7(2.0×)3,682(1.5×)LAION-400M61.480.962.252.834.7
TinyCLIP (Ours)ViT-22M/32autoauto22+10(3.9×)1.9(3.9×)5,504(2.2×)LAION-400M53.771.352.044.428.3
TinyCLIP (Ours)ViT-63M/32autoauto63+31(1.3×)5.6(1.3×)2,909(1.2×)LAION+YFCC-400M64.584.966.056.938.5
TinyCLIP (Ours)ViT-45M/32autoauto45+18(2.0×)3.7(2.0×)3,685(1.5×)LAION+YFCC-400M62.780.363.954.036.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.358, + 0.893, + 0.387 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-art methods. The architecture of text encoder is Transformer [55]. \"auto\" denotes automatic weight inheritance, and our other models use manual weight inheritance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.411, + 0.452, + 0.426 + ], + "angle": 0, + "content": "Flickr30k [44] and MSCOCO retrieval [33] are reported." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.471, + 0.594 + ], + "angle": 0, + "content": "When performing training on YFCC-15M [53], we use affinity mimicking and manual weight inheritance to distill OpenCLIP ViT-B/16 [20], whose zero-shot performance on ImageNet is \\(70.2\\%\\). By inheriting half of parameters, our TinyCLIP ViT-39M/16 can achieve a zero-shot accuracy of \\(63.5\\%\\) on ImageNet, surpassing the original CLIP ViT-B/16 [46] by \\(25.9\\%\\). Furthermore, when compressing the image encoder to 10 layers with 256 dimensions and the text encoder to 3 layers with 256 dimensions, TinyCLIP ViT-8M/16 still surpasses CLIP ViT-B/16 [46] by \\(3.5\\%\\) while using \\(11.3\\times\\) fewer parameters and running \\(5.1\\times\\) faster." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.746 + ], + "angle": 0, + "content": "When conducting training on LAION-400M [50], we distill three different models, i.e., Florence DaViT [72, 8], CLIP ResNet [46], and OpenCLIP ViT-B/32 [20]. The results in Tab. 1 demonstrate the efficacy of our method. More specifically, our TinyCLIP DaViT-5M achieves up to \\(5.0\\%\\) improvements over the original model on ImageNet, and \\(5.0 / 4.5\\%\\) gains in image-to-text retrieval on Flicker and MSCOCO, respectively. It is worth noting that this is a tiny language-image pre-trained model using only 10M parameters to achieve \\(50.0\\%\\) accuracy on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.471, + 0.884 + ], + "angle": 0, + "content": "For CLIP ResNet, TinyCLIP-30M is only trained for 4 epochs in the first stage, inheriting \\(75\\%\\) parameters from the original model. The results indicate that there is \\(0.5\\%\\) slight decrease in top-1 accuracy on ImageNet, however, has \\(\\sim 2\\%\\) improvements in retrieval tasks. TinyCLIP ResNet-19M inherits the weights from the 30M model, and it is trained for 12 epochs. It reduces the parameters by \\(50\\%\\) while getting \\(2\\times\\) inference speedup. But it gets \\(2.7\\%\\) performance drops on ImageNet due to the large reduction in parameters." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.885, + 0.47, + 0.901 + ], + "angle": 0, + "content": "For OpenCLIP ViT-B/32, we distill it using our affin" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.895, + 0.669 + ], + "angle": 0, + "content": "ity mimicking method with manual or automatic weight inheritance. We design a three-stage progressive distillation in which each stage compresses the model by \\(25\\%\\). The three stages are trained for 6, 16, and 16 epochs, respectively. In manual inheritance, TinyCLIP ViT-61M/32 obtains comparable performance compared to OpenCLIP ViT-B/32 [20], while reducing the parameters by \\(29\\%\\). On the other hand, automatic inheritance obtains better results than the manual one. For example, TinyCLIP ViT-63M/32 outperforms the manually inherited ViT-61M/32 model by more than \\(\\sim 1\\%\\) on both classification and retrieval tasks. When auto-compressed the model by \\(3.9\\times\\), TinyCLIP ViT-22M/32 obtains the highest inference throughput of 5,504 and reaches \\(53.7\\%\\) top-1 accuracy. Since YFCC-15M contains high-quality image-text pairs, if using it to replace parts of LAION data, our method can obtain additional improvements, especially for retrieval tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.655, + 0.696 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.885 + ], + "angle": 0, + "content": "Impact of affinity mmicking. As shown in Tab. 2, we present four different interaction modes. The affinity mimicking \\(\\mathcal{L}_1\\) (i.e., \\(\\mathcal{L}_{distill}\\)) outperforms the contrastive loss \\(\\mathcal{L}_0\\) by \\(2.1\\%\\) in terms of top-1 accuracy on ImageNet [7]. The alignment in a visual-linguistic affinity space brings more similarity information. The accuracy of cross modalities interaction is lower than that of affinity mimicking by \\(0.2\\%\\). In this mode, the embedding features of student image and text do not have any interaction, but they are aligned to the teacher's embedding space. Single modality interaction has lower accuracy due to missing a loss to align student's image and text features in common embedding space." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Impact of weight inheritance. We verify the efficacy of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21975" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.18 + ], + "angle": 0, + "content": "
Interaction ModeLoss FormulaTop-1 Acc
Contrastive loss [46]L0=CE(<Is,Ts>,I)53.4 %
Affinity mimickingL1=CE(<Is,Ts>,<It,Tt>)55.5 %
Cross modalitiesL2=CE(<Is,Tt>,<It,Tt>)55.3 %
L3=CE(<It,Ts>,<It,Tt>)
Single modalityL4=CE(<Is,I>.I)19.2 %
L5=CE(<Ts,Tt>,I)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.19, + 0.471, + 0.288 + ], + "angle": 0, + "content": "Table 2. The interaction of information across image \\( I \\) and text \\( T \\) modalities in student and teacher models. The subscripts \\( s \\) and \\( t \\) denote student and teacher, respectively. \\( \\mathcal{I} \\) is an identity matrix. \\( CE \\) is cross entropy function. TinyCLIP-ViT-40M/32 is trained with an interaction mode and manual weight inheritance on LAION-400M [50] for 1 epoch. The zero-shot accuracy on ImageNet [7] is reported." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.307, + 0.268, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.307, + 0.462, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.466 + ], + "angle": 0, + "content": "Figure 4. The cosine similarity of the input and output embedding of each layer, and the zero-shot accuracy on ImageNet [7] after removing a layer for OpenCLIP ViT-B-32 [20, 46] pretrained on LAION-2B [49]." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.48, + 0.472, + 0.549 + ], + "angle": 0, + "content": "
Weight InheritanceIN-1K top1 acc (%)Flickr30kMSCOCO
I→T@1T→I@1I→T@1T→I@1
None37.850.233.429.716.6
Manual54.070.452.946.529.3
Automatic54.973.555.947.930.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.592 + ], + "angle": 0, + "content": "Table 3. Manual inheritance vs. Automatic inheritance. The model OpenCLIP ViT-B/32 [20] is compressed by \\(50\\%\\) with distillation on LAION-400M [50] for 1 epoch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.47, + 0.716 + ], + "angle": 0, + "content": "the proposed two kinds of weight inheritance in Tab. 3. Compared to the model without using weight inheritance, manual inheritance brings \\(16.2\\%\\) accuracy improvements on ImageNet [7]. Automatic weight inheritance can further improve the accuracy by \\(0.9\\%\\). On image-text retrievals, the improvements of automatic inheritance is relatively higher, i.e., \\(1.4 - 3.1\\%\\) over the manual method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.869 + ], + "angle": 0, + "content": "The redundancy of language-image models. We analyze the redundancy of image encoder and text encoder by similar strategy [6, 2, 25] and zero-shot accuracy by removing a layer. As shown in Fig. 4, we observe a high correlation between cosine similarity and accuracy. Comparing with image encoder, text encoder has higher cosine similarity. It shows the redundancy of the text encoder is higher than that of the image encoder. Consequently, compressing the text encoder along the depth dimension can be achieved without significant performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We further explore the redundancy by the proposed automatic weight inheritance. As shown in Fig. 5, for width" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.091, + 0.882, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.323, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Figure 5. Visualization of learnable masks at \\(50\\%\\) sparsity on hidden dimension, attention heads and MLP intermediate dimension. The blocks in red color are removed. The original model is OpenCLIP ViT-B/32 [20]." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.389, + 0.847, + 0.485 + ], + "angle": 0, + "content": "
Pretrained Teacher ModelTeacher Acc.(%)Inherited RatioStudent Acc.(%)
w/o weight inheritance-036.2
CLIP ViT-B/32 [46]63.259 / 12652.4(+16.2)
OpenCLIP ViT-B/32 [20]62.959 / 12653.5(+17.3)
OpenCLIP ViT-B/16 [20]67.159 / 12452.8(+16.6)
OpenCLIP ViT-L/14 [20]75.359 / 39045.1(+8.9)
OpenCLIP ViT-H/14 [20]78.059 / 93541.1(+4.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.489, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Table 4. Ablation study on inherited teacher model. The student model TinyCLIP ViT-40M/32 with 59M parameters is inherited from a teacher model, then trained without distillation on LAION-400M [50] for 1 epoch. The top-1 zero-shot accuracy on ImageNet [7] is reported." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.734 + ], + "angle": 0, + "content": "analysis, it is observed that the embedding channels of the text encoder experience only a marginal reduction (e.g., 512 \\(\\rightarrow\\) 511), whereas the embedding channels of the image encoder are reduced to a greater extent (e.g., \\(768\\to 526\\)). Regarding depth analysis, it was observed that a significant proportion of MHA layers and FFN channels could be removed from the text encoder, indicating redundancy within it. This finding suggests that the image encoder can be compressed along the width dimension, while the text encoder can be compressed along the depth dimension." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Impact of teacher models. We investigate which model is better for weight inheritance. As analyzed in Tab. 4, a strong teacher does not necessarily guarantee better results for weight inheritance. Instead, teacher models with similar architectures and higher performance offer a better choice. For instance, although ViT-H/14 ranks as the highest-performing teacher model, it lags behind other models in terms of weight inheritance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Impact of multi-stage progressive distillation. Multi-stage progressive distillation shrinks the gap between the compressed model and the inherited model in each stage." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21976" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.077, + 0.891, + 0.256 + ], + "angle": 0, + "content": "
MethodImage EncoderFood101CIFAR10CIFAR100SUN397Stanford CarsFGYC AircraftVOC207DTDOxford PetsCaltech101Flowers102MNISTFER2013STL10EuroSATRESISC45GTSRBKITTICountry211PCamHateful MemesRendered SST2ImageNet
Zero-shot performance
DeCLIP [31]ViT-B/3276.390.067.963.550.68.980.545.384.489.383.212.816.397.540.048.410.335.311.554.752.950.166.1
FILIP [70]ViT-B/3282.886.965.569.155.557.2-49.388.191.985.3----49.9------68.8
CLIP [46]ViT-B/3284.491.365.163.259.421.283.144.587.087.966.751.947.397.249.460.332.239.417.858.457.659.663.2
OpenCLIP [20]ViT-B/3280.990.770.666.879.316.682.254.486.590.166.037.442.395.651.657.642.031.614.850.152.952.362.9
TinyCLIP (Ours)ViT-45M/3278.392.271.866.079.923.083.253.885.388.760.762.545.296.252.261.841.116.916.551.553.552.661.2
Linear probe performance
CLIP [46]ViT-B/3288.895.180.576.681.852.087.776.590.093.096.999.069.298.397.090.585.366.227.883.966.770.876.1
TinyCLIP (Ours)ViT-45M/3284.097.486.068.985.439.085.471.289.993.482.199.566.997.697.588.198.579.716.190.458.462.573.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.248, + 0.265, + 0.721, + 0.278 + ], + "angle": 0, + "content": "Table 5. Zero-shot and linear-probe classification top-1 accuracy on 23 datasets." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.294, + 0.466, + 0.368 + ], + "angle": 0, + "content": "
Number of StagesModel SizeEpochImageNet top1 acc (%)MSCOCO
I→TR@1T→IR@1
1100% → 25%247.039.423.1
2100% → 50%154.046.529.3
50% → 25%149.040.225.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.37, + 0.469, + 0.412 + ], + "angle": 0, + "content": "Table 6. Ablation study on multi-stage progressive distillation. The model OpenCLIP ViT-B/32 [20] is compressed to \\(25\\%\\) model size on LAION-400M [50] for 2 epochs." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.418, + 0.432, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.526, + 0.47, + 0.582 + ], + "angle": 0, + "content": "Figure 6. Training efficiency. OpenCLIP ViT-B/32 [20] is trained from scratch on LAION-400M [50] for 32 epochs. TinyCLIP ViT/32 is trained for 6+16 epochs using the proposed 2-stage progressive distillation, where the teacher is OpenCLIP ViT-B/32." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.596, + 0.469, + 0.687 + ], + "angle": 0, + "content": "We compare it with single-stage in Tab. 6. Using the same training epochs, the model trained with 2-stage surpasses that with 1-stage by \\(2.0\\%\\) in terms of zero-shot accuracy on ImageNet. It demonstrates that multi-stage progressive compression is superior when transferring the knowledge of the pre-trained model to the compressed small model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.469, + 0.839 + ], + "angle": 0, + "content": "Training cost. We also study the distillation efficiency. As shown in Fig. 6, the original OpenCLIP ViT-B/32 [20] takes around 287 GPU days when trained on LAION-400M for 32 epochs. Our models with \\(75\\%\\) and \\(50\\%\\) parameters reach the similar accuracy as OpenCLIP counterparts, getting a speed up of \\(7.8\\times\\) and \\(1.4\\times\\), respectively. The underlying reason is that our training benefits more from the affinity mimicking supervision and the good initialization using weight inheritance. This also proves that weight inheritance is capable of accelerating cross-modal distillation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.317, + 0.864 + ], + "angle": 0, + "content": "4.4. Transfer Learning Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Classification on 23 datasets. We also evaluate the performance of zero-shot and linear-probe classification on 23" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.293, + 0.891, + 0.358 + ], + "angle": 0, + "content": "
ModelIN-V2IN-AIN-RObjectNetIN-Sketch
CLIP ViT-B/32 [46]56.031.669.429.942.3
OpenCLIP ViT-B/32 [20]55.121.773.428.949.4
TinyCLIP ViT-63M/32(Ours)55.722.874.131.250.8
TinyCLIP ViT-45M/32(Ours)52.619.871.529.148.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.368, + 0.892, + 0.411 + ], + "angle": 0, + "content": "Table 7. Zero-shot robustness evaluation. CLIP [46] is trained on WIT-400M, OpenCLIP [20] and TinyCLIP is trained on LAION-400M [50]. Top-1 accuracy is reported." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.501 + ], + "angle": 0, + "content": "datasets by the toolkit Elevater [26]. As shown in Tab. 5, our model wins on 7 datasets in zero-shot benchmark, and 9 datasets on linear-probe benchmark. The performance of our model is close to OpenCLIP ViT-B/32 [20], where the two models are both trained on Laion-400M [50]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.893, + 0.638 + ], + "angle": 0, + "content": "Zero-shot robustness evaluation. We verify the robustness of our models on 5 datasets, which are out of ImageNet distribution. As shown in Tab. 7, our TinyCLIP ViT-63M/32 outperforms OpenCLIP ViT-B/32 [20], and wins 3 datasets when compared with CLIP ViT-B/32 [46]. Our TinyCLIP ViT-45M/32 uses \\(50\\%\\) fewer parameters yet still wins on IN-R and IN-Sketch, compared to CLIP ViT-B/32. This demonstrates the robustness of the distilled TinyCLIP models and the efficacy of our proposed distillation method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.645, + 0.627, + 0.66 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.807 + ], + "angle": 0, + "content": "In this paper, we propose TinyCLIP, a method for distilling large-scale vision-language pre-trained models. It mainly introduced two core techniques named affinity mimicking and weight inheritance. Extensive experiments and ablation studies have demonstrated the effectiveness of TinyCLIP, showing that it can largely reduce model size while maintaining competitive performance. In the future, we will explore ways to further improve cross-modal distillation efficiency on extremely small models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.66, + 0.832 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Prof. Hongyang Chao was partially supported by NSFC (U1611461, 61672548) and NSFC (U22A2095). Prof. Xinggang Wang was partly supported by NSFC (No. 62276108)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21977" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.171 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.469, + 0.198 + ], + "angle": 0, + "content": "[2] Shaojie Bai, J Zico Kolter, and Vladlen Koltun. Deep equilibrium models. NeurIPS, 32, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.469, + 0.255 + ], + "angle": 0, + "content": "[3] Arnav Chavan, Zhiqiang Shen, Zhuang Liu, Zechun Liu, Kwang-Ting Cheng, and Eric P Xing. Vision transformer slimming: Multi-dimension searching in continuous optimization space. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.469, + 0.311 + ], + "angle": 0, + "content": "[4] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.469, + 0.354 + ], + "angle": 0, + "content": "[5] Wenliang Dai, Lu Hou, Lifeng Shang, Xin Jiang, Qun Liu, and Pascale Fung. Enabling multimodal generation on clip via vision-language knowledge distillation. In ACL, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.469, + 0.395 + ], + "angle": 0, + "content": "[6] Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Lukasz Kaiser. Universal transformers. In ICLR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.469, + 0.479 + ], + "angle": 0, + "content": "[8] Mingyu Ding, Bin Xiao, Noel Codella, Ping Luo, Jingdong Wang, and Lu Yuan. Davit: Dual attention vision transformers. In ECCV, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.469, + 0.59 + ], + "angle": 0, + "content": "[10] Yifan Du, Zikang Liu, Junyi Li, and Wayne Xin Zhao. A survey of vision-language pre-trained models. In *IJCAI*, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.593, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[11] Zhiyuan Fang, Jianfeng Wang, Xiaowei Hu, Lijuan Wang, Yezhou Yang, and Zicheng Liu. Compressing visual-linguistic model via knowledge distillation. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[12] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In ICLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[13] Prakhar Ganesh, Yao Chen, Xin Lou, Mohammad Ali Khan, Yin Yang, Hassan Sajjad, Preslav Nakov, Deming Chen, and Marianne Winslett. Compressing large-scale transformer-based models: A case study on bert. ACL, 9, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[14] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[15] Yunhao Ge, Jie Ren, Ming-Hsuan Yang, Yuxiao Wang, Andrew Gallagher, Hartwig Adam, Laurent Itti, Balaji Lakshminarayanan, and Jiaping Zhao. Improving zero-shot generalization and robustness of multi-modal models. In NeurIPS ML Safety Workshop, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[16] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. IJCV, 129:1789-1819, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[17] Manish Gupta and Puneet Agrawal. Compression of deep learning models for text: A survey. TKDD, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[19] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. arXiv, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[20] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[22] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding. In EMNLP, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[23] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, ICLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[25] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. ICLR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[26] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, and Jianfeng Gao. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. NeurIPS, 2022. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[28] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[29] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[30] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. arXiv preprint arXiv:2212.00794, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[31] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21978" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. arXiv preprint arXiv:2210.04150, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[34] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. In ICLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.468, + 0.262 + ], + "angle": 0, + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[36] Jiawei Han Jamie Callan Luyu Gao, Yunyi Zhang. Scaling deep contrastive learning batch size under memory limited setup. In RepL4NLP, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.376 + ], + "angle": 0, + "content": "[37] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 5191-5198, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.469, + 0.419 + ], + "angle": 0, + "content": "[38] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.476 + ], + "angle": 0, + "content": "[39] Basil Mustafa, Carlos Riquelme Ruiz, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.589 + ], + "angle": 0, + "content": "[41] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[42] Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, and Youliang Yan. Clipping: Distilling clip-based models for video-language understanding. OpenReview, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[43] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, et al. Combined scaling for open-vocabulary image classification. arXiv, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[44] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[45] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[47] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[48] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[50] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[51] Dachuan Shi, Chaofan Tao, Ying Jin, Zhendong Yang, Chun Yuan, and Jiaqi Wang. Upop: Unified and progressive pruning for compressing vision-language transformers. arXiv preprint arXiv:2301.13741, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[52] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[53] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[54] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[56] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[57] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[58] Yulong Wang, Xiaolu Zhang, Lingxi Xie, Jun Zhou, Hang Su, Bo Zhang, and Xiaolin Hu. Pruning from scratch. In AAAI, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[59] Zhecan Wang, Noel Codella, Yen-Chun Chen, Luowei Zhou, Jianwei Yang, Xiyang Dai, Bin Xiao, Haoxuan You, Shih-Fu Chang, and Lu Yuan. Clip-td: Clip targeted distillation for vision-language tasks. arXiv preprint arXiv:2201.05729, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[60] Ziheng Wang, Jeremy Wohlwend, and Tao Lei. Structured pruning of large language models. In EMNLP, 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21979" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[61] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[62] Yixuan Wei, Yue Cao, Zheng Zhang, Zhuliang Yao, Zhenda Xie, Han Hu, and Baining Guo. icar: Bridging image classification and image-text alignment for visual recognition. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.402, + 0.206 + ], + "angle": 0, + "content": "[63] Ross Wightman. Pytorch image models, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[64] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.304 + ], + "angle": 0, + "content": "[65] Mengzhou Xia, Zexuan Zhong, and Danqi Chen. Structured pruning learns compact and accurate models. In ACL, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.362 + ], + "angle": 0, + "content": "[66] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[67] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.462 + ], + "angle": 0, + "content": "[68] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.464, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[69] Shusheng Yang, Yixiao Ge, Kun Yi, Dian Li, Ying Shan, Xiaohu Qie, and Xinggang Wang. Masked visual reconstruction in language semantic space. arXiv preprint arXiv:2301.06958, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[70] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[71] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[72] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[73] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[74] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In CVPR, pages 18123-18133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[75] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[76] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.149 + ], + "angle": 0, + "content": "[77] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21980" + } + ] +] \ No newline at end of file diff --git a/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/7a0fee57-111a-443f-8a63-68f0c3d1c0b8_origin.pdf b/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/7a0fee57-111a-443f-8a63-68f0c3d1c0b8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7890b958cbe6c98bb6f2f80d049ea27690608856 --- /dev/null +++ b/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/7a0fee57-111a-443f-8a63-68f0c3d1c0b8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd18181da6a7798d46898a6224bf06da3847c191410b87be70c3ee9cac27b46d +size 818159 diff --git a/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/full.md b/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0d3cacf02184a6c31792762a5d719bbba928eb31 --- /dev/null +++ b/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/full.md @@ -0,0 +1,350 @@ +# TinyCLIP: CLIP Distillation via Affinity Mimicking and Weight Inheritance + +Kan Wu $^{1,\ast}$ , Houwen Peng $^{3,\ast,\dagger}$ , Zhenghong Zhou $^{2,\ast}$ , Bin Xiao $^{3}$ , Mengchen Liu $^{3}$ , Lu Yuan $^{3}$ , Hong Xuan $^{3}$ , Michael Valenzuela $^{3}$ , Xi (Stephen) Chen $^{3}$ , Xinggang Wang $^{2}$ , Hongyang Chao $^{1}$ , Han Hu $^{3}$ , Sun Yat-sen University, Huazhong University of Science & Technology, Microsoft + +# Abstract + +In this paper, we propose a novel cross-modal distillation method, called TinyCLIP, for large-scale language-image pre-trained models. The method introduces two core techniques: affinity mimicking and weight inheritance. Affinity mimicking explores the interaction between modalities during distillation, enabling student models to mimic teachers' behavior of learning cross-modal feature alignment in a visual-linguistic affinity space. Weight inheritance transmits the pre-trained weights from the teacher models to their student counterparts to improve distillation efficiency. Moreover, we extend the method into a multi-stage progressive distillation to mitigate the loss of informative weights during extreme compression. Comprehensive experiments demonstrate the efficacy of TinyCLIP, showing that it can reduce the size of the pre-trained CLIP ViT-B/32 by $50\%$ , while maintaining comparable zero-shot performance. While aiming for comparable performance, distillation with weight inheritance can speed up the training by $1.4 - 7.8 \times$ compared to training from scratch. Moreover, our TinyCLIP ViT-8M/16, trained on YFCC-15M, achieves an impressive zero-shot top-1 accuracy of $41.1\%$ on ImageNet, surpassing the original CLIP ViT-B/16 by $3.5\%$ while utilizing only $8.9\%$ parameters. Finally, we demonstrate the good transferability of TinyCLIP in various downstream tasks. Code and models will be open-sourced at aka.ms/tinyclip. + +# 1. Introduction + +Large-scale language-image pretraining, e.g., CLIP [46], has recently gained significant attention due to its remarkable zero-shot transfer capability [46] and unprecedented performance in text-to-image generation [47]. Due to complex nature of vision and language, current approaches often resort to utilizing huge amounts of parameters to endow models with cross-modal capabilities [46, 21, 71, 72, 43, 39, + +![](images/bc00e8caee1bbc668dcf0890656304b70bb3c1286c8df0fcf45e41f05d538a4d.jpg) +Figure 1. Comparison of the OpenCLIP [20] and TinyCLIP. TinyCLIP is pre-trained and distilled on LAION-400M [50] using OpenCLIP ViT-B/32 [20] as the teacher, whose zero-shot top-1 accuracy is $65.6\%$ . + +![](images/34580367f76f388d597d70d6ebe129118186435bb8b4ae9e0432eeaee12e712a.jpg) + +1]. This in turn leads to high costs in terms of storage, memory, and computation time for these models, which motivates the need for model compression to make them smaller and faster for real-world applications [64, 13]. + +As a core compression technique, knowledge distillation has been extensively studied and applied in single-modal settings [19, 16]. However, its potential for multi-modality remains underexplored. Unlike single-modal models, the distillation of language-image cross-modal models poses distinct challenges. First, CLIP-like language-image models commonly consist of two branches: an image encoder and a text encoder [21, 72, 52, 73]. When distilling such multi-branch models, it is crucial to consider the interaction of information across the different modality branches in both the teacher and student models. Second, the original CLIP [46] models are pre-trained on 400 million image-text pairs for 32 epochs, which requires thousands of GPU days, making distillation a significant challenge when computational resources are limited. Is there any way to reduce the cost of CLIP distillation? + +To tackle these challenges, we present a novel cross-modal distillation method dubbed TinyCLIP, which introduces two key techniques: affinity mimicking and weight inheritance. In contrast to the methods that rely on either + +image or text features for distillation, we empirically show that distilling knowledge in an image-text affinity space is more effective. Specifically, we leverage the cosine similarity of the image and text embeddings in the teacher model to facilitate the distillation of the student model, allowing the student to mimic the teacher's visual-linguistic feature alignment. We refer to this process as affinity mimicking. + +To improve distillation efficiency, we introduce weight inheritance, a technique transferring the pre-trained weights from teacher models to their student counterparts. Since inheriting weights provides a good initialization for the student models, the distillation progress can be largely accelerated. The key challenge of weight inheritance lies in determining which weights are more advantageous. To address this issue, we introduce two solutions: manual and automatic inheritance. We surprisingly found that a simple manual selection of $k$ -dimension or $k$ -layer weights from the teacher model can yield satisfactory results for CLIP distillation. On the other hand, we also introduce learnable masks to automatically identify the most important weights from the teacher model. The masks are imposed independently on the vision and language branches, enabling them to capture the differences across modalities. + +Moreover, we extend the proposed weight inheritance to a multi-stage progressive procedure, where each subsequent stage automatically inherits the important weights from the preceding stages. We observed that when the teacher model exhibits higher performance and shares a similar architecture with the student, weight inheritance can provide better results. This is because significant architecture differences may undermine the learned weights when transmitting from the teacher to the student. Therefore, we break the inheritance into multiple stages, allowing the student model in each stage to share a more similar structure with the predecessor teacher and inherit the weights progressively. + +Our experiments show that TinyCLIP delivers competitive models at all levels of speedups and model sizes in ImageNet zero-shot evaluation and various downstream tasks. As shown in Fig. 1, pre-trained on the same LAION-400M dataset [50], TinyCLIP-ViT using 63M parameters achieves $61.4\%$ zero-shot top-1 accuracy on ImageNet [7], being $2.0 \times$ smaller and $1.5 \times$ faster than the OpenCLIP [20] model (62.9% accuracy with 126M parameters). Meanwhile, TinyCLIP-ResNet with 38M parameters obtains $56.4\%$ zero-shot top-1 accuracy, being $2.1 \times$ smaller and $2.0 \times$ faster than CLIP ResNet-50 (59.6% accuracy with 76M parameters). Moreover, our method can speed up the original OpenCLIP training by $1.4 \times -7.8 \times$ while achieving similar performance. Also, TinyCLIP demonstrates good transfer capacities in downstream scenarios. In summary, the contributions of this work are two-fold: + +- We propose a new cross-modal distillation approach to unleash the capacity of small CLIP models, fully lever + +aging large-scale models as well as pre-training data. To our best knowledge, this is the first work exploring the pre-training distillation of language-image models. + +- We present state-of-the-art language-image pre-trained models at small scale, striking the best trade-off between speed and accuracy. Extensive experiments demonstrate the superiority and good generalization ability of the small models in various downstream tasks. + +# 2. Related Work + +Language-Image Pre-training has achieved remarkable progress over the past few years [10, 4]. In particular, contrastive language-image pre-training demonstrates very impressive zero-shot transfer and generalization capacities [45, 46, 21, 72, 61, 43, 74]. One of the most representative works is CLIP [46]. A large amount of follow-up works have been proposed to optimize the pre-training framework [27, 76, 15, 68, 62, 14, 28, 67]. Meanwhile, a line of works leverage the pre-trained models for downstream tasks, such as open-vocabulary detection and segmentation [29, 75, 66, 32], video recognition [56, 23], and text-to-image generation [47, 41]. + +Most recently, there are a few works attempting to scale language-image pre-trained models. BEiT-3 [57] uses a multiway transformer to scales up multi-modal pre-trained models to 1.9B parameters. FLIP [30] scales image-text pre-training via masked modeling, enabling CLIP to enjoy faster training speed while getting better performance. On the other hand, there are also few works on scaling down CLIP models [51]. However, they all focus on specific tasks, such as generation [5, 51] and video content understanding [42]. In contrast, our work concentrates on language-image pre-training, which is the first work on CLIP compression. + +Knowledge Distillation in a teacher-student framework [19] has been widely used to transfer knowledge from large models to small ones. Distillation in single modality, such as vision [54, 64, 37] or language [48, 22], has been extensively studied. However, the exploration of cross-modality distillation is relatively limited. A few early works exploring cross-modal distillation only focus on specific tasks, such as VQA and image caption [5, 11, 59], limiting the methodologies to be applied in general downstream tasks. In contrast, this work pays attention to general cross-modal pre-training distillation, in which the distilled models can be transferred to various downstream scenarios. + +It is worth noting that our proposed weight inheritance strategy shares similar spirits to model pruning techniques [17, 65, 51]. Both approaches involve the identification of important weights while removing redundant ones. However, they have three fundamental differences. 1) Previous pruning methods predominantly focus on single-modality + +models [17, 65, 3], whereas our approach shifts attention to multi-modality. Since different modalities exhibit varying redundancy characteristics, it is necessary to consider these differences when selecting important weights from each modality. 2) Unlike previous pruning methods which have emphasized the importance of the pruned architecture [12, 34, 58], we argue that the weights inherited from the original models are also highly beneficial, particularly when compressing CLIP-like language-image models. 3) Our method leverages a progressive multi-stage process, which enables the inherited weights to accelerate the convergence of small models during training. + +# 3. Method + +In this section, we propose TinyCLIP, a simple and effective method for distilling large-scale language-image pretraining models, such as CLIP. It consists of three components: affinity mimicking (§3.1), weight inheritance (§3.2) and multi-stage progressive distillation (§3.3). + +# 3.1. Distillation with Affinity Mimicking + +A language-image pre-training model often consists of two branches, an image encoder to extract visual representations and a text encoder to extract textual representations. The visual and textual representations are linearly projected into a cross-modal common space by minimizing a contrastive loss $\mathcal{L}_0$ . The supervision signal can be viewed as an identity matrix, in which diagonal values are set to 1 for positive image-text pairs, and all others are set to 0 for negative pairs. This signal disregards the similarity between negative pairs, which prevents the model from acquiring a nuanced understanding of the intricate relationships among negative pairs. Therefore, we introduce affinity mimicking to enable student models to learn similarities between negative pairs from teacher models. + +More specifically, as shown in Fig. 2, we consider two types of affinity distillation losses: image-to-language loss $\mathcal{L}_{I2T}$ and language-to-image loss $\mathcal{L}_{T2I}$ . The former learns the alignment between the teacher and student models based on the image-to-language affinity $A_{I2T}$ , which represents the affinity scores of an image with all the text descriptions in a batch (the yellow row highlighted in Fig. 2). The latter refers to the alignment of language-to-image affinity $A_{T2I}$ , which compares a text description with all the images to obtain a matching score (the blue column in Fig. 2). The combination of these two losses formulates our affinity mimicking distillation, which is represented as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {d i s t i l l}} = \mathcal {L} _ {I 2 T} + \mathcal {L} _ {T 2 I}, \tag {1} \\ = C E \left(A _ {I 2 T} ^ {s}, A _ {I 2 T} ^ {t}\right) + C E \left(A _ {T 2 I} ^ {s}, A _ {T 2 I} ^ {t}\right). \\ \end{array} +$$ + +Here, $CE$ represents the cross-entropy loss. The superscripts $s$ and $t$ denote the student and teacher models, re + +![](images/070cab7b775b02144ec4257e65c466a94e5563f728d1555c61ff45401e06aae3.jpg) +Figure 2. Affinity mimicking for language-image models. The loss includes image-to-text loss (yellow) and text-to-image loss (blue). + +spectively. The elements in $A_{I2T}$ and $A_{T2I}$ are defined as: + +$$ +A _ {I 2 T} (i, j) = \frac {\exp \left(I _ {i} \cdot T _ {j} / \tau\right)}{\sum_ {k \in \mathcal {B}} \exp \left(I _ {i} \cdot T _ {k} / \tau\right)}, \tag {2} +$$ + +$$ +A _ {T 2 I} (i, j) = \frac {\exp \left(I _ {i} \cdot T _ {j} / \tau\right)}{\sum_ {k \in \mathcal {B}} \exp \left(I _ {k} \cdot T _ {j} / \tau\right)}, \tag {3} +$$ + +where $\tau$ is a temperature parameter, $I_{i}$ is the feature embedding of $i$ -th image in the batch $\mathcal{B}$ , $T_{j}$ is the feature of $j$ -th text description in $\mathcal{B}$ . Affinity mimicking allows student models to imitate the behavior of the large model in learning visual-linguistic alignment. In contrast to previous methods that rely on either image or text features for distillation, we show that distilling knowledge in this image-text affinity space is more effective. + +# 3.2. Distillation with Weight Inheritance + +The original CLIP models are pre-trained on 400 million image-text pairs for 32 epochs, taking thousands of GPU days. This presents a significant cost challenge for distillation. To improve the training efficiency, we introduce weight inheritance, a technique that inherits the important weights from the well-trained large teacher models to smaller student models. The key challenge of inheriting weights is identifying important weights from the large amount of weights of the teacher. We propose two approaches to select the important weights: manual weight inheritance and automatic weight inheritance. + +Manual Inheritance. For manual weight inheritance, we first analyze the redundancy of existing CLIP pre-trained models. Fig. 4 shows that the text encoder displays more redundancy in depth (layer-wise), while the image encoder exhibits more redundancy in width (channel-wise). Based upon this discovery, we adopt the approach of uniformly selecting $k$ layers of the text branch and directly taking the + +front $k$ channels of the image branch to select the important weights from large teacher models. These weights will serve as the initialization for small models during distillation. We surprisingly found that such a simple manual selected weights can largely accelerate CLIP distillation. + +Automatic Inheritance. Although manual inheritance can be considered an effective approach for model compression, it does present the drawback of necessitating prior knowledge in order to determine which weights to inherit. This limitation may render the technique impractical for application to diverse models. To solve this issue, we present an automatic weight inheritance scheme. Inspired by structured pruning in large language models [65], we introduce learnable masks to identify weight importance. Considering the difference across modalities, the learnable masks are imposed on vision and language branches independently, as visualized in Fig. 3. An overall sparsity constraint is introduced to guarantee the selected number of important weights to meet our compression requirements. Without loss of generality, here we use the transformer architecture as an example to introduce the detailed procedure of weight inheritance. Specifically, a standard transformer block contains a multi-head attention (MHA) layer and a feed-forward network (FFN). To capture the importance of weights in a fine-grained level, we introduce two mask variables $\mathbf{m}_{\mathrm{head}}$ , $\mathbf{m}_{\mathrm{int}} \in \{0,1\}$ to identify the redundant attention heads in MHA and neurons in FFN respectively, while keeping the important ones. These two kinds of masks are imposed on the activation of attention heads and the intermediate layer of FFN, which is formulated as + +$$ +\operatorname {M H A} (\mathbf {X}) = \sum_ {h = 1} ^ {N _ {H}} \mathbf {m} _ {\text {h e a d}} ^ {h} \cdot \operatorname {A t t n} _ {\mathbf {W} _ {Q} ^ {h}, \mathbf {W} _ {K} ^ {h}, \mathbf {W} _ {V} ^ {h}, \mathbf {W} _ {O} ^ {h}} (\mathbf {X}), \tag {4} +$$ + +$$ +\operatorname {F F N} (\mathbf {X}) = \operatorname {G e L U} \left(\mathbf {X} \mathbf {W} _ {U}\right) \cdot \operatorname {d i a g} \left(\mathbf {m} _ {\text {i n t}}\right) \cdot \mathbf {W} _ {D}, \tag {5} +$$ + +where $\mathbf{X}$ is the layer input, $\mathbf{W}_Q^h$ , $\mathbf{W}_K^h$ , $\mathbf{W}_V^h$ , $\mathbf{W}_O^h \in \mathbb{R}^{d \times d_h}$ denote the query, key, value and output matrices in MHA respectively, and $\mathbf{W}_U \in \mathbb{R}^{d \times d_f}$ and $\mathbf{W}_D \in \mathbb{R}^{d_f \times d}$ represent the parameters of FFN layers. Here, $d$ denotes the hidden dimension (e.g., 768), and $d_h = d / N_H$ denotes the output dimension of each head (e.g., 64), where $N_H$ is the total number of heads. The trainable masks $\mathbf{m}_{\mathrm{head}}$ and $\mathbf{m}_{\mathrm{int}}$ serve as importance indicators for MHA and FFN layers. Moreover, to further learn the importance of embedding dimensions in transformer, we introduce an additional mask $\mathbf{m}_{\mathrm{embed}} \in \{0,1\}$ . This mask is shared across all layers because each dimension in the hidden representation is connected to the corresponding dimension in the subsequent layer through a residual connection. + +During automatic inheritance, there are two losses used for optimizing the masks and the model jointly, i.e., a sparsity loss and a distillation loss defined in Eq. (1): + +$$ +\mathcal {L} = \mathcal {L} _ {\text {d i s t i l l}} + \mathcal {L} _ {\text {s p a r s i t y}}, \tag {6} +$$ + +![](images/a4bf8109bb3809a89651d98aff1e26bcba7e2318a3d6dafdeb9dde1fceb76366.jpg) +10 Learnable mask/manual mask +Figure 3. Weight inheritance of CLIP. Given a large pre-trained CLIP model, the inherited weights are selected by masks. The weights whose mask is 0 will be removed. For the manual way, the masks are set to 1 in the first k-dimension or uniform k-layer, the rest is set to 0. For the automatic way, the masks are learned jointly with the model under the sparsity constraint. + +![](images/74d8b9cee14fdbba392dde2ac887ef215c07794afc6610d703dd83cecda7939d.jpg) +Model sparsity $=$ + +$$ +\mathcal {L} _ {\text {s p a r s i t y}} = \lambda \cdot (p - q) + \beta \cdot (p - q) ^ {2}, \tag {7} +$$ + +where $\lambda, \beta$ are learnable multipliers that guarantee $p = q[60, 65]$ . Here, $q$ is the target compression rate, $p$ is the overall compression rate of learnable masks for the model, including image encoder and text encoder: + +$$ +\begin{array}{l} p = \frac {1}{M _ {t} + M _ {t}} \sum_ {i m g, t x t} (4 \cdot d _ {h} \cdot \sum_ {i} ^ {L} \sum_ {j} ^ {N _ {H}} \sum_ {k} ^ {d} \mathbf {m} _ {\text {h e a d}} ^ {i, j} \cdot \mathbf {m} _ {\text {e m b e d}} ^ {k} \\ + 2 \cdot \sum_ {i} ^ {L} \sum_ {j} ^ {d _ {f}} \sum_ {k} ^ {d} \mathbf {m} _ {\text {i n t}} ^ {i, j} \cdot \mathbf {m} _ {\text {e m b e d}} ^ {k}), \tag {8} \\ \end{array} +$$ + +where $M_{i}$ and $M_{t}$ represent the full model size of the image encoder and text encoder, respectively, $L$ is the total number of transformer layers. + +# 3.3. Progressive Multi-Stage Distillation + +When attempting to achieve a high target sparsity, i.e., $>70\%$ , compressing the model in a single stage can lead to a significant reduction in accuracy and even result in convergence failure. This is due to the fact that most weights of the large model are directly discarded, including those that are important for ensuring model quality and convergence. As a solution, we propose a multi-stage progressive distillation method to achieve a high compression rate without seriously sacrificing accuracy. In each stage, we use a modest degree of compression, e.g., $25\%$ , to avoid large loss of performance and make training stable. + +Specifically, each stage includes weight inheritance and affinity mimicking, as described in Algorithm 1. During the weight inheritance phase, the weights of the large model are gradually reduced until the target sparsity and key weights are retained for the small model. After that, the small model is distilled with affinity mimicking, which involves transferring the visual-linguistic affinity knowledge from + +Algorithm 1: TinyCLIP: Cross-Modal Distillation +Input: a pre-trained CLIP model $f(x;\theta_0)$ with weight $\theta_0$ , target compression rate $q$ total iteration steps $N$ 1 Initialize the number of stages $G$ , each stage steps $L$ , target compression rate $q_{i}$ of stage i, weight inheritance steps $L_{M}$ input $x$ , learnable mask $M = \{\mathbf{m}_{\mathrm{head}},\mathbf{m}_{\mathrm{int}},\mathbf{m}_{\mathrm{embed}}\} ;^1$ +2 for i in [1,...,G] do +3 // Increase mask compression rate, retain important weights +4 for j in [1,...,LM] do +5 Update target compression rate $^2 q$ of current step based on $q_{i}$ and step $j$ . +6 Calculate distillation loss between origin teacher model $f(x;\theta_0)$ and masked student model $f(x;\theta_{i - 1}\odot M_i)$ . +7 Calculate sparsity loss between learnable mask compression rate $^3 p$ and target compression rate $q$ . +8 Optimize $\theta_{i - 1}$ and $M_{i}$ . +9 end +10 $\theta_{i}\gets \theta_{i - 1}\odot M_{i}$ //Inherit weight with masks +11 // Cross-Modal distillation +12 for j in $[L_M + 1,\dots ,L]$ do +13 Calculate distillation loss between origin teacher model $f(x;\theta_0)$ and compressed model $f(x;\theta_i)$ . +14 Optimize $\theta_{i}$ . +15 end +16 end +17 Return compressed model $f(x;\theta_G)$ at compression rate $q$ +1 Each stage compress $25\%$ $q_{i} = 1 - i\cdot 25\%$ $G = \lceil (1 - q) / 0.25\rceil$ $L = N / G,L_{M} = 3000,M$ is initialized to 1, $q_{i}$ can be adjusted according to different models. +2 Target compression rate $q$ increases linearly from O to $q_{i}$ with step $j$ . + $p$ is calculated by masks $M$ in Eq. (8). + +the teacher to the student. This procedure is repeated until the target sparsity level is achieved. + +# 4. Experiments + +# 4.1. Implementation Details + +Architecture. The original CLIP employs a transformer-based model as its text encoder, while its image encoder has two variations, i.e., ViT [9] and modified ResNet [18], covering both transformer-based and CNN-based architectures. For weight inheritance, we inherited the weights from two model variants: OpenCLIP ViT-B/32 or ViT-B/16 [20] and CLIP ResNet-50 [46], which are pre-trained on LAION-2B [49] and WIT-400M [46], respectively1. For distillation, we use OpenCLIP ViT-B/32 [20] pre-trained on LAION-2B [49] as the teacher model, since it achieves a high zero-shot accuracy of $65.6\%$ on ImageNet [7] with a high inference throughput. Besides, we also extend our method to DaViT [8], which is a hybrid architecture. We re-produce Florence-DaViT-5M [72, 8] on LAION-400M as a baseline. Then we distill the same model using affinity mimicking and weight inheritance. The teacher model is Florence-DaViT-D3, which is pre-trained on FLD-900M [72] and achieves a top-1 accuracy of $78.0\%$ on ImageNet [7] with 128M parameters. + +Affinity Mimicking. As defined in Tab. 2, we explore + +different interaction schemes between teacher and student models across modalities. The contrastive loss $\mathcal{L}_0$ and affinity mimicking $\mathcal{L}_1$ (i.e., $\mathcal{L}_{\text{distill}}$ ) serve as the basic ones, which can be combined with other losses. The weight of each loss is set to 1 when combined. The temperature parameter $\tau$ in Eq. (3) is set to $\frac{1}{50}$ by default. The sensitivity analysis of $\tau$ is presented in the supplementary material. + +Weight Inheritance. We compress OpenCLIP ViT-B [20] and CLIP ResNet-50 [46] using the proposed multi-stage progressive distillation. In each stage, we apply manual or automatic weight inheritance. In manual inheritance, we compress the image encoder and the text encoder separately, by reducing the width of image encoder and the depth of text encoder. In automatic inheritance, we set the target sparsity for the whole model. In each stage, the trainable masks are initialized to 1 and updated in the first 3,000 training iterations, where the optimizer is AdamW [35, 24] with a constant learning rate of 0.01 and no weight decay. The learnable multipliers $\lambda, \beta$ defined in Eq. (7) are both initialized to 0.01. + +Training Settings. We train our models on two public datasets, namely LAION-400M [50] and YFCC-15M [53]. On LAION-400M, the models are compressed in 3 stages, including $100\%$ to $75\%$ parameters for 6 epochs, $75\%$ to $50\%$ parameters for 16 epochs, and $50\%$ to $25\%$ parameters for 16 epochs. On YFCC-15M, it contains 2 compression stages, where the training epochs are both 25 from $100\%$ to $50\%$ parameters, and $50\%$ to $10\%$ . We follow the hyper-parameter of CLIP [46] except that the learning rate is set to $10^{-4}$ when using weight inheritance. The details are shown in the supplementary material. Unless otherwise specified, our ablation studies utilize the same settings, except for training the models only for 1 epoch with a learning rate of $5 \times 10^{-4}$ . All models are trained on 32 Nvidia V100 or A100, and implemented with PyTorch [40], OpenCLIP [20], gradient cache [36], and Timm library [63]. + +Evaluation Settings. The models are evaluated on multiple benchmarks. In zero-shot transfer evaluation and robustness evaluation, we follow the same prompt engineering in CLIP [46], where 80 text templates per class are used for ImageNet [7]. In linear probe, we use Elevater toolkit [26] for evaluation, where a classification head is trained for 50 epochs with searched hyper-parameters. Note that we do not count the number of parameters in the text embedding layer. It is a look-up table whose parameter size is the same as the models with the same hidden dimension and vocabulary size. The inference throughput is measured on Nvidia V100 with CUDA 11 and PyTorch v1.12 [40], where the batch size is 1,024. + +# 4.2. Comparison with State-of-the-Art Models + +We compare our TinyCLIP with state-of-the-art models in Tab. 1. The zero-shot evaluation on ImageNet-1K [7], + +
MethodImageEncoderText Encoder#Params (M)Image+TextMACs(G)Throughput(pairs/s)TrainingdatasetsIN-1Ktop1 acc (%)Flickr30kMSCOCO
depthwidthI→T@1T→I@1I→T@1T→I@1
Training data: 15M / 20M
RILS [69]ViT-B/161251286 + 3820.5818LAION-20M45.045.134.932.225.5
MaskCLIP [77]ViT-B/161251286 + 3820.5818LAION-20M46.664.948.138.524.8
SLIP [38]ViT-B/161251286 + 3820.5818YFCC-15M42.857.640.131.120.3
CLIP [46, 38]ViT-B/161251286 + 3820.5818YFCC-15M37.651.632.226.517.1
TinyCLIP (Ours)ViT-39M/16651239+19(2.1×)9.5(2.2×)1,469(1.8×)YFCC-15M63.584.466.754.938.9
TinyCLIP (Ours)ViT-8M/1632568+3(11.3×)2.0(10.3×)4,150(5.1×)YFCC-15M41.162.342.336.221.5
Training data: 400M
Florence [72]DaViT-5M62565+51.12,980LAION-400M45.061.241.536.220.9
TinyCLIP (Ours)DaViT-5M62565+51.12,980LAION-400M50.066.248.540.724.5
CLIP [46]ResNet-1011251256 + 3812.81,161WIT-400M62.278.159.249.330.7
CLIP [46]ResNet-501251238 + 389.11,549WIT-400M59.681.258.250.828.3
TinyCLIP (Ours)ResNet-30M951230+29(1.3×)6.9(1.3×)1,811(1.2×)LAION-400M59.181.161.252.733.9
TinyCLIP (Ours)ResNet-19M651219+19(2.0×)4.4(2.1×)3,024(2.0×)LAION-400M56.476.258.348.930.9
OpenCLIP [20]ViT-B/321251288 + 387.42,452LAION-2B65.784.766.856.939.3
CLIP [46]ViT-B/321251288 + 387.42,452WIT-400M63.280.159.851.230.6
OpenCLIP [20]ViT-B/321251288 + 387.42,452LAION-400M62.979.362.053.335.4
TinyCLIP (Ours)ViT-61M/32951261+29(1.4×)5.3(1.4×)3,191(1.3×)LAION-400M62.178.663.353.935.9
TinyCLIP (Ours)ViT-40M/32651240+19(2.1×)3.5(2.1×)4,641(1.9×)LAION-400M59.777.358.949.833.1
TinyCLIP (Ours)ViT-63M/32autoauto63+31(1.3×)5.6(1.3×)2,905(1.2×)LAION-400M63.983.264.455.537.6
TinyCLIP (Ours)ViT-45M/32autoauto45+18(2.0×)3.7(2.0×)3,682(1.5×)LAION-400M61.480.962.252.834.7
TinyCLIP (Ours)ViT-22M/32autoauto22+10(3.9×)1.9(3.9×)5,504(2.2×)LAION-400M53.771.352.044.428.3
TinyCLIP (Ours)ViT-63M/32autoauto63+31(1.3×)5.6(1.3×)2,909(1.2×)LAION+YFCC-400M64.584.966.056.938.5
TinyCLIP (Ours)ViT-45M/32autoauto45+18(2.0×)3.7(2.0×)3,685(1.5×)LAION+YFCC-400M62.780.363.954.036.7
+ +Table 1. Comparison with the state-of-the-art methods. The architecture of text encoder is Transformer [55]. "auto" denotes automatic weight inheritance, and our other models use manual weight inheritance. + +Flickr30k [44] and MSCOCO retrieval [33] are reported. + +When performing training on YFCC-15M [53], we use affinity mimicking and manual weight inheritance to distill OpenCLIP ViT-B/16 [20], whose zero-shot performance on ImageNet is $70.2\%$ . By inheriting half of parameters, our TinyCLIP ViT-39M/16 can achieve a zero-shot accuracy of $63.5\%$ on ImageNet, surpassing the original CLIP ViT-B/16 [46] by $25.9\%$ . Furthermore, when compressing the image encoder to 10 layers with 256 dimensions and the text encoder to 3 layers with 256 dimensions, TinyCLIP ViT-8M/16 still surpasses CLIP ViT-B/16 [46] by $3.5\%$ while using $11.3\times$ fewer parameters and running $5.1\times$ faster. + +When conducting training on LAION-400M [50], we distill three different models, i.e., Florence DaViT [72, 8], CLIP ResNet [46], and OpenCLIP ViT-B/32 [20]. The results in Tab. 1 demonstrate the efficacy of our method. More specifically, our TinyCLIP DaViT-5M achieves up to $5.0\%$ improvements over the original model on ImageNet, and $5.0 / 4.5\%$ gains in image-to-text retrieval on Flicker and MSCOCO, respectively. It is worth noting that this is a tiny language-image pre-trained model using only 10M parameters to achieve $50.0\%$ accuracy on ImageNet. + +For CLIP ResNet, TinyCLIP-30M is only trained for 4 epochs in the first stage, inheriting $75\%$ parameters from the original model. The results indicate that there is $0.5\%$ slight decrease in top-1 accuracy on ImageNet, however, has $\sim 2\%$ improvements in retrieval tasks. TinyCLIP ResNet-19M inherits the weights from the 30M model, and it is trained for 12 epochs. It reduces the parameters by $50\%$ while getting $2\times$ inference speedup. But it gets $2.7\%$ performance drops on ImageNet due to the large reduction in parameters. + +For OpenCLIP ViT-B/32, we distill it using our affin + +ity mimicking method with manual or automatic weight inheritance. We design a three-stage progressive distillation in which each stage compresses the model by $25\%$ . The three stages are trained for 6, 16, and 16 epochs, respectively. In manual inheritance, TinyCLIP ViT-61M/32 obtains comparable performance compared to OpenCLIP ViT-B/32 [20], while reducing the parameters by $29\%$ . On the other hand, automatic inheritance obtains better results than the manual one. For example, TinyCLIP ViT-63M/32 outperforms the manually inherited ViT-61M/32 model by more than $\sim 1\%$ on both classification and retrieval tasks. When auto-compressed the model by $3.9\times$ , TinyCLIP ViT-22M/32 obtains the highest inference throughput of 5,504 and reaches $53.7\%$ top-1 accuracy. Since YFCC-15M contains high-quality image-text pairs, if using it to replace parts of LAION data, our method can obtain additional improvements, especially for retrieval tasks. + +# 4.3. Ablation Study + +Impact of affinity mmicking. As shown in Tab. 2, we present four different interaction modes. The affinity mimicking $\mathcal{L}_1$ (i.e., $\mathcal{L}_{distill}$ ) outperforms the contrastive loss $\mathcal{L}_0$ by $2.1\%$ in terms of top-1 accuracy on ImageNet [7]. The alignment in a visual-linguistic affinity space brings more similarity information. The accuracy of cross modalities interaction is lower than that of affinity mimicking by $0.2\%$ . In this mode, the embedding features of student image and text do not have any interaction, but they are aligned to the teacher's embedding space. Single modality interaction has lower accuracy due to missing a loss to align student's image and text features in common embedding space. + +Impact of weight inheritance. We verify the efficacy of + +
Interaction ModeLoss FormulaTop-1 Acc
Contrastive loss [46]L0=CE(<Is,Ts>,I)53.4 %
Affinity mimickingL1=CE(<Is,Ts>,<It,Tt>)55.5 %
Cross modalitiesL2=CE(<Is,Tt>,<It,Tt>)55.3 %
L3=CE(<It,Ts>,<It,Tt>)
Single modalityL4=CE(<Is,I>.I)19.2 %
L5=CE(<Ts,Tt>,I)
+ +![](images/2094b009db3040487fccea5795494dd3488ccf20ac0fe53ab637b079f617025f.jpg) +Figure 4. The cosine similarity of the input and output embedding of each layer, and the zero-shot accuracy on ImageNet [7] after removing a layer for OpenCLIP ViT-B-32 [20, 46] pretrained on LAION-2B [49]. + +![](images/c41e5c3551a412d1d91981f688f44bca3fd4144f14f686a9202939f77b8359ae.jpg) + +Table 2. The interaction of information across image $I$ and text $T$ modalities in student and teacher models. The subscripts $s$ and $t$ denote student and teacher, respectively. $\mathcal{I}$ is an identity matrix. $CE$ is cross entropy function. TinyCLIP-ViT-40M/32 is trained with an interaction mode and manual weight inheritance on LAION-400M [50] for 1 epoch. The zero-shot accuracy on ImageNet [7] is reported. + +
Weight InheritanceIN-1K top1 acc (%)Flickr30kMSCOCO
I→T@1T→I@1I→T@1T→I@1
None37.850.233.429.716.6
Manual54.070.452.946.529.3
Automatic54.973.555.947.930.8
+ +the proposed two kinds of weight inheritance in Tab. 3. Compared to the model without using weight inheritance, manual inheritance brings $16.2\%$ accuracy improvements on ImageNet [7]. Automatic weight inheritance can further improve the accuracy by $0.9\%$ . On image-text retrievals, the improvements of automatic inheritance is relatively higher, i.e., $1.4 - 3.1\%$ over the manual method. + +The redundancy of language-image models. We analyze the redundancy of image encoder and text encoder by similar strategy [6, 2, 25] and zero-shot accuracy by removing a layer. As shown in Fig. 4, we observe a high correlation between cosine similarity and accuracy. Comparing with image encoder, text encoder has higher cosine similarity. It shows the redundancy of the text encoder is higher than that of the image encoder. Consequently, compressing the text encoder along the depth dimension can be achieved without significant performance degradation. + +We further explore the redundancy by the proposed automatic weight inheritance. As shown in Fig. 5, for width + +![](images/5a5866aefa3bf45f4445ee87701ace50ceb2739df2218dfcfdae2cb611321a2e.jpg) +Figure 5. Visualization of learnable masks at $50\%$ sparsity on hidden dimension, attention heads and MLP intermediate dimension. The blocks in red color are removed. The original model is OpenCLIP ViT-B/32 [20]. + +Table 3. Manual inheritance vs. Automatic inheritance. The model OpenCLIP ViT-B/32 [20] is compressed by $50\%$ with distillation on LAION-400M [50] for 1 epoch. + +
Pretrained Teacher ModelTeacher Acc.(%)Inherited RatioStudent Acc.(%)
w/o weight inheritance-036.2
CLIP ViT-B/32 [46]63.259 / 12652.4(+16.2)
OpenCLIP ViT-B/32 [20]62.959 / 12653.5(+17.3)
OpenCLIP ViT-B/16 [20]67.159 / 12452.8(+16.6)
OpenCLIP ViT-L/14 [20]75.359 / 39045.1(+8.9)
OpenCLIP ViT-H/14 [20]78.059 / 93541.1(+4.9)
+ +Table 4. Ablation study on inherited teacher model. The student model TinyCLIP ViT-40M/32 with 59M parameters is inherited from a teacher model, then trained without distillation on LAION-400M [50] for 1 epoch. The top-1 zero-shot accuracy on ImageNet [7] is reported. + +analysis, it is observed that the embedding channels of the text encoder experience only a marginal reduction (e.g., 512 $\rightarrow$ 511), whereas the embedding channels of the image encoder are reduced to a greater extent (e.g., $768\to 526$ ). Regarding depth analysis, it was observed that a significant proportion of MHA layers and FFN channels could be removed from the text encoder, indicating redundancy within it. This finding suggests that the image encoder can be compressed along the width dimension, while the text encoder can be compressed along the depth dimension. + +Impact of teacher models. We investigate which model is better for weight inheritance. As analyzed in Tab. 4, a strong teacher does not necessarily guarantee better results for weight inheritance. Instead, teacher models with similar architectures and higher performance offer a better choice. For instance, although ViT-H/14 ranks as the highest-performing teacher model, it lags behind other models in terms of weight inheritance. + +Impact of multi-stage progressive distillation. Multi-stage progressive distillation shrinks the gap between the compressed model and the inherited model in each stage. + +
MethodImage EncoderFood101CIFAR10CIFAR100SUN397Stanford CarsFGYC AircraftVOC207DTDOxford PetsCaltech101Flowers102MNISTFER2013STL10EuroSATRESISC45GTSRBKITTICountry211PCamHateful MemesRendered SST2ImageNet
Zero-shot performance
DeCLIP [31]ViT-B/3276.390.067.963.550.68.980.545.384.489.383.212.816.397.540.048.410.335.311.554.752.950.166.1
FILIP [70]ViT-B/3282.886.965.569.155.557.2-49.388.191.985.3----49.9------68.8
CLIP [46]ViT-B/3284.491.365.163.259.421.283.144.587.087.966.751.947.397.249.460.332.239.417.858.457.659.663.2
OpenCLIP [20]ViT-B/3280.990.770.666.879.316.682.254.486.590.166.037.442.395.651.657.642.031.614.850.152.952.362.9
TinyCLIP (Ours)ViT-45M/3278.392.271.866.079.923.083.253.885.388.760.762.545.296.252.261.841.116.916.551.553.552.661.2
Linear probe performance
CLIP [46]ViT-B/3288.895.180.576.681.852.087.776.590.093.096.999.069.298.397.090.585.366.227.883.966.770.876.1
TinyCLIP (Ours)ViT-45M/3284.097.486.068.985.439.085.471.289.993.482.199.566.997.697.588.198.579.716.190.458.462.573.2
+ +Table 5. Zero-shot and linear-probe classification top-1 accuracy on 23 datasets. + +
Number of StagesModel SizeEpochImageNet top1 acc (%)MSCOCO
I→TR@1T→IR@1
1100% → 25%247.039.423.1
2100% → 50%154.046.529.3
50% → 25%149.040.225.3
+ +![](images/6035fb3cd23eef3dc6f6b2890861ce5b6c7fe80d3d2ac9fb61b2272097f96e64.jpg) +Figure 6. Training efficiency. OpenCLIP ViT-B/32 [20] is trained from scratch on LAION-400M [50] for 32 epochs. TinyCLIP ViT/32 is trained for 6+16 epochs using the proposed 2-stage progressive distillation, where the teacher is OpenCLIP ViT-B/32. + +We compare it with single-stage in Tab. 6. Using the same training epochs, the model trained with 2-stage surpasses that with 1-stage by $2.0\%$ in terms of zero-shot accuracy on ImageNet. It demonstrates that multi-stage progressive compression is superior when transferring the knowledge of the pre-trained model to the compressed small model. + +Training cost. We also study the distillation efficiency. As shown in Fig. 6, the original OpenCLIP ViT-B/32 [20] takes around 287 GPU days when trained on LAION-400M for 32 epochs. Our models with $75\%$ and $50\%$ parameters reach the similar accuracy as OpenCLIP counterparts, getting a speed up of $7.8\times$ and $1.4\times$ , respectively. The underlying reason is that our training benefits more from the affinity mimicking supervision and the good initialization using weight inheritance. This also proves that weight inheritance is capable of accelerating cross-modal distillation. + +# 4.4. Transfer Learning Results + +Classification on 23 datasets. We also evaluate the performance of zero-shot and linear-probe classification on 23 + +Table 6. Ablation study on multi-stage progressive distillation. The model OpenCLIP ViT-B/32 [20] is compressed to $25\%$ model size on LAION-400M [50] for 2 epochs. + +
ModelIN-V2IN-AIN-RObjectNetIN-Sketch
CLIP ViT-B/32 [46]56.031.669.429.942.3
OpenCLIP ViT-B/32 [20]55.121.773.428.949.4
TinyCLIP ViT-63M/32(Ours)55.722.874.131.250.8
TinyCLIP ViT-45M/32(Ours)52.619.871.529.148.8
+ +Table 7. Zero-shot robustness evaluation. CLIP [46] is trained on WIT-400M, OpenCLIP [20] and TinyCLIP is trained on LAION-400M [50]. Top-1 accuracy is reported. + +datasets by the toolkit Elevater [26]. As shown in Tab. 5, our model wins on 7 datasets in zero-shot benchmark, and 9 datasets on linear-probe benchmark. The performance of our model is close to OpenCLIP ViT-B/32 [20], where the two models are both trained on Laion-400M [50]. + +Zero-shot robustness evaluation. We verify the robustness of our models on 5 datasets, which are out of ImageNet distribution. As shown in Tab. 7, our TinyCLIP ViT-63M/32 outperforms OpenCLIP ViT-B/32 [20], and wins 3 datasets when compared with CLIP ViT-B/32 [46]. Our TinyCLIP ViT-45M/32 uses $50\%$ fewer parameters yet still wins on IN-R and IN-Sketch, compared to CLIP ViT-B/32. This demonstrates the robustness of the distilled TinyCLIP models and the efficacy of our proposed distillation method. + +# 5. Conclusions + +In this paper, we propose TinyCLIP, a method for distilling large-scale vision-language pre-trained models. It mainly introduced two core techniques named affinity mimicking and weight inheritance. Extensive experiments and ablation studies have demonstrated the effectiveness of TinyCLIP, showing that it can largely reduce model size while maintaining competitive performance. In the future, we will explore ways to further improve cross-modal distillation efficiency on extremely small models. + +# Acknowledgement + +Prof. Hongyang Chao was partially supported by NSFC (U1611461, 61672548) and NSFC (U22A2095). Prof. Xinggang Wang was partly supported by NSFC (No. 62276108). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1 +[2] Shaojie Bai, J Zico Kolter, and Vladlen Koltun. Deep equilibrium models. NeurIPS, 32, 2019. 7 +[3] Arnav Chavan, Zhiqiang Shen, Zhuang Liu, Zechun Liu, Kwang-Ting Cheng, and Eric P Xing. Vision transformer slimming: Multi-dimension searching in continuous optimization space. In CVPR, 2022. 3 +[4] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 2023. 2 +[5] Wenliang Dai, Lu Hou, Lifeng Shang, Xin Jiang, Qun Liu, and Pascale Fung. Enabling multimodal generation on clip via vision-language knowledge distillation. In ACL, 2022. 2 +[6] Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Lukasz Kaiser. Universal transformers. In ICLR, 2019. 7 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5, 6, 7 +[8] Mingyu Ding, Bin Xiao, Noel Codella, Ping Luo, Jingdong Wang, and Lu Yuan. Davit: Dual attention vision transformers. In ECCV, 2022. 5, 6 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5 +[10] Yifan Du, Zikang Liu, Junyi Li, and Wayne Xin Zhao. A survey of vision-language pre-trained models. In *IJCAI*, 2022. 2 +[11] Zhiyuan Fang, Jianfeng Wang, Xiaowei Hu, Lijuan Wang, Yezhou Yang, and Zicheng Liu. Compressing visual-linguistic model via knowledge distillation. In ICCV, 2021. 2 +[12] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In ICLR, 2019. 3 +[13] Prakhar Ganesh, Yao Chen, Xin Lou, Mohammad Ali Khan, Yin Yang, Hassan Sajjad, Preslav Nakov, Deming Chen, and Marianne Winslett. Compressing large-scale transformer-based models: A case study on bert. ACL, 9, 2021. 1 +[14] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. In NeurIPS, 2022. 2 +[15] Yunhao Ge, Jie Ren, Ming-Hsuan Yang, Yuxiao Wang, Andrew Gallagher, Hartwig Adam, Laurent Itti, Balaji Lakshminarayanan, and Jiaping Zhao. Improving zero-shot generalization and robustness of multi-modal models. In NeurIPS ML Safety Workshop, 2022. 2 +[16] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. IJCV, 129:1789-1819, 2021. 1 + +[17] Manish Gupta and Puneet Agrawal. Compression of deep learning models for text: A survey. TKDD, 2022. 2, 3 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5 +[19] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. arXiv, 2015. 1, 2 +[20] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. 1, 2, 5, 6, 7, 8 +[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 2 +[22] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding. In EMNLP, 2020. 2 +[23] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 2 +[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, ICLR, 2015. 5 +[25] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. ICLR, 2020. 7 +[26] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, and Jianfeng Gao. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. NeurIPS, 2022. 5, 8 +[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2 +[28] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2 +[29] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 2 +[30] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. arXiv preprint arXiv:2212.00794, 2022. 2 +[31] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 8 +[32] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana + +Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. arXiv preprint arXiv:2210.04150, 2022. 2 +[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6 +[34] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. In ICLR, 2019. 3 +[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2018. 5 +[36] Jiawei Han Jamie Callan Luyu Gao, Yunyi Zhang. Scaling deep contrastive learning batch size under memory limited setup. In RepL4NLP, 2021. 5 +[37] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 5191-5198, 2020. 2 +[38] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. 6 +[39] Basil Mustafa, Carlos Riquelme Ruiz, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts. In NeurIPS, 2022. 1 +[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5 +[41] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 2 +[42] Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, and Youliang Yan. Clipping: Distilling clip-based models for video-language understanding. OpenReview, 2023. 2 +[43] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, et al. Combined scaling for open-vocabulary image classification. arXiv, 2021. 1, 2 +[44] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015, 6 +[45] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020. 2 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5, 6, 7, 8 + +[47] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2 +[48] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 2 +[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 5, 7 +[50] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2, 5, 6, 7, 8 +[51] Dachuan Shi, Chaofan Tao, Ying Jin, Zhendong Yang, Chun Yuan, and Jiaqi Wang. Upop: Unified and progressive pruning for compressing vision-language transformers. arXiv preprint arXiv:2301.13741, 2023. 2 +[52] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. 1 +[53] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016. 5, 6 +[54] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2 +[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6 +[56] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 2 +[57] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2 +[58] Yulong Wang, Xiaolu Zhang, Lingxi Xie, Jun Zhou, Hang Su, Bo Zhang, and Xiaolin Hu. Pruning from scratch. In AAAI, 2020. 3 +[59] Zhecan Wang, Noel Codella, Yen-Chun Chen, Luowei Zhou, Jianwei Yang, Xiyang Dai, Bin Xiao, Haoxuan You, Shih-Fu Chang, and Lu Yuan. Clip-td: Clip targeted distillation for vision-language tasks. arXiv preprint arXiv:2201.05729, 2022. 2 +[60] Ziheng Wang, Jeremy Wohlwend, and Tao Lei. Structured pruning of large language models. In EMNLP, 2020. 4 + +[61] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. In ICLR, 2021. 2 +[62] Yixuan Wei, Yue Cao, Zheng Zhang, Zhuliang Yao, Zhenda Xie, Han Hu, and Baining Guo. icar: Bridging image classification and image-text alignment for visual recognition. In NeurIPS, 2022. 2 +[63] Ross Wightman. Pytorch image models, 2019. 5 +[64] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In ECCV, 2022. 1, 2 +[65] Mengzhou Xia, Zexuan Zhong, and Danqi Chen. Structured pruning learns compact and accurate models. In ACL, 2022. 2, 3, 4 +[66] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In ECCV, 2022. 2 +[67] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. 2022. 2 +[68] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In CVPR, 2022. 2 +[69] Shusheng Yang, Yixiao Ge, Kun Yi, Dian Li, Ying Shan, Xiaohu Qie, and Xinggang Wang. Masked visual reconstruction in language semantic space. arXiv preprint arXiv:2301.06958, 2023. 6 +[70] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2022. 8 +[71] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. In NeurIPS, 2022. 1 +[72] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1, 2, 5, 6 +[73] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 1 +[74] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In CVPR, pages 18123-18133, 2022. 2 +[75] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. In NeurIPS, 2022. 2 +[76] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang + +Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 2 +[77] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 6 \ No newline at end of file diff --git a/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/images.zip b/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..52cbc7093344f8cb87b01e98aff2e54f70fa5dc3 --- /dev/null +++ b/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94bca63e930a625a417211c464be2b277b994381b2ea50a1010c50f06bf835a8 +size 654334 diff --git a/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/layout.json b/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f7533cbed102b9097b07d7b247172b5fa5c300eb --- /dev/null +++ b/2023/TinyCLIP_ CLIP Distillation via Affinity Mimicking and Weight Inheritance/layout.json @@ -0,0 +1,10946 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 102, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 533, + 121 + ], + "type": "text", + "content": "TinyCLIP: CLIP Distillation via Affinity Mimicking and Weight Inheritance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "spans": [ + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "content": "Kan Wu" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "content": ", Houwen Peng" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "inline_equation", + "content": "^{3,\\ast,\\dagger}" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "content": ", Zhenghong Zhou" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "inline_equation", + "content": "^{2,\\ast}" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "content": ", Bin Xiao" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "content": ", Mengchen Liu" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "content": ", Lu Yuan" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "content": ", Hong Xuan" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "content": ", Michael Valenzuela" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "content": ", Xi (Stephen) Chen" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "content": ", Xinggang Wang" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "content": ", Hongyang Chao" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "content": ", Han Hu" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 148, + 548, + 194 + ], + "type": "text", + "content": ", Sun Yat-sen University, Huazhong University of Science & Technology, Microsoft" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "text", + "content": "In this paper, we propose a novel cross-modal distillation method, called TinyCLIP, for large-scale language-image pre-trained models. The method introduces two core techniques: affinity mimicking and weight inheritance. Affinity mimicking explores the interaction between modalities during distillation, enabling student models to mimic teachers' behavior of learning cross-modal feature alignment in a visual-linguistic affinity space. Weight inheritance transmits the pre-trained weights from the teacher models to their student counterparts to improve distillation efficiency. Moreover, we extend the method into a multi-stage progressive distillation to mitigate the loss of informative weights during extreme compression. Comprehensive experiments demonstrate the efficacy of TinyCLIP, showing that it can reduce the size of the pre-trained CLIP ViT-B/32 by " + }, + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "text", + "content": ", while maintaining comparable zero-shot performance. While aiming for comparable performance, distillation with weight inheritance can speed up the training by " + }, + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "inline_equation", + "content": "1.4 - 7.8 \\times" + }, + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "text", + "content": " compared to training from scratch. Moreover, our TinyCLIP ViT-8M/16, trained on YFCC-15M, achieves an impressive zero-shot top-1 accuracy of " + }, + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "inline_equation", + "content": "41.1\\%" + }, + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "text", + "content": " on ImageNet, surpassing the original CLIP ViT-B/16 by " + }, + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "text", + "content": " while utilizing only " + }, + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 47, + 251, + 290, + 563 + ], + "type": "text", + "content": " parameters. Finally, we demonstrate the good transferability of TinyCLIP in various downstream tasks. Code and models will be open-sourced at aka.ms/tinyclip." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 580, + 128, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 580, + 128, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 580, + 128, + 594 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 601, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 685 + ], + "type": "text", + "content": "Large-scale language-image pretraining, e.g., CLIP [46], has recently gained significant attention due to its remarkable zero-shot transfer capability [46] and unprecedented performance in text-to-image generation [47]. Due to complex nature of vision and language, current approaches often resort to utilizing huge amounts of parameters to endow models with cross-modal capabilities [46, 21, 71, 72, 43, 39," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 208, + 431, + 370 + ], + "blocks": [ + { + "bbox": [ + 310, + 208, + 431, + 370 + ], + "lines": [ + { + "bbox": [ + 310, + 208, + 431, + 370 + ], + "spans": [ + { + "bbox": [ + 310, + 208, + 431, + 370 + ], + "type": "image", + "image_path": "bc00e8caee1bbc668dcf0890656304b70bb3c1286c8df0fcf45e41f05d538a4d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 374, + 547, + 419 + ], + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 419 + ], + "type": "text", + "content": "Figure 1. Comparison of the OpenCLIP [20] and TinyCLIP. TinyCLIP is pre-trained and distilled on LAION-400M [50] using OpenCLIP ViT-B/32 [20] as the teacher, whose zero-shot top-1 accuracy is " + }, + { + "bbox": [ + 304, + 374, + 547, + 419 + ], + "type": "inline_equation", + "content": "65.6\\%" + }, + { + "bbox": [ + 304, + 374, + 547, + 419 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 432, + 209, + 526, + 369 + ], + "blocks": [ + { + "bbox": [ + 432, + 209, + 526, + 369 + ], + "lines": [ + { + "bbox": [ + 432, + 209, + 526, + 369 + ], + "spans": [ + { + "bbox": [ + 432, + 209, + 526, + 369 + ], + "type": "image", + "image_path": "34580367f76f388d597d70d6ebe129118186435bb8b4ae9e0432eeaee12e712a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 425, + 547, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 547, + 474 + ], + "type": "text", + "content": "1]. This in turn leads to high costs in terms of storage, memory, and computation time for these models, which motivates the need for model compression to make them smaller and faster for real-world applications [64, 13]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 474, + 548, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 548, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 548, + 664 + ], + "type": "text", + "content": "As a core compression technique, knowledge distillation has been extensively studied and applied in single-modal settings [19, 16]. However, its potential for multi-modality remains underexplored. Unlike single-modal models, the distillation of language-image cross-modal models poses distinct challenges. First, CLIP-like language-image models commonly consist of two branches: an image encoder and a text encoder [21, 72, 52, 73]. When distilling such multi-branch models, it is crucial to consider the interaction of information across the different modality branches in both the teacher and student models. Second, the original CLIP [46] models are pre-trained on 400 million image-text pairs for 32 epochs, which requires thousands of GPU days, making distillation a significant challenge when computational resources are limited. Is there any way to reduce the cost of CLIP distillation?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "To tackle these challenges, we present a novel cross-modal distillation method dubbed TinyCLIP, which introduces two key techniques: affinity mimicking and weight inheritance. In contrast to the methods that rely on either" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 286, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 286, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 286, + 703 + ], + "type": "text", + "content": "* Equal contribution. Kan and Zhenghong were interns of Microsoft." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 61, + 703, + 219, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 219, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 219, + 712 + ], + "type": "text", + "content": "† Corresponding: houwen.peng@microsoft.com" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21970" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "image or text features for distillation, we empirically show that distilling knowledge in an image-text affinity space is more effective. Specifically, we leverage the cosine similarity of the image and text embeddings in the teacher model to facilitate the distillation of the student model, allowing the student to mimic the teacher's visual-linguistic feature alignment. We refer to this process as affinity mimicking." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 346 + ], + "type": "text", + "content": "To improve distillation efficiency, we introduce weight inheritance, a technique transferring the pre-trained weights from teacher models to their student counterparts. Since inheriting weights provides a good initialization for the student models, the distillation progress can be largely accelerated. The key challenge of weight inheritance lies in determining which weights are more advantageous. To address this issue, we introduce two solutions: manual and automatic inheritance. We surprisingly found that a simple manual selection of " + }, + { + "bbox": [ + 46, + 157, + 288, + 346 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 157, + 288, + 346 + ], + "type": "text", + "content": "-dimension or " + }, + { + "bbox": [ + 46, + 157, + 288, + 346 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 157, + 288, + 346 + ], + "type": "text", + "content": "-layer weights from the teacher model can yield satisfactory results for CLIP distillation. On the other hand, we also introduce learnable masks to automatically identify the most important weights from the teacher model. The masks are imposed independently on the vision and language branches, enabling them to capture the differences across modalities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 348, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 491 + ], + "type": "text", + "content": "Moreover, we extend the proposed weight inheritance to a multi-stage progressive procedure, where each subsequent stage automatically inherits the important weights from the preceding stages. We observed that when the teacher model exhibits higher performance and shares a similar architecture with the student, weight inheritance can provide better results. This is because significant architecture differences may undermine the learned weights when transmitting from the teacher to the student. Therefore, we break the inheritance into multiple stages, allowing the student model in each stage to share a more similar structure with the predecessor teacher and inherit the weights progressively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "text", + "content": "Our experiments show that TinyCLIP delivers competitive models at all levels of speedups and model sizes in ImageNet zero-shot evaluation and various downstream tasks. As shown in Fig. 1, pre-trained on the same LAION-400M dataset [50], TinyCLIP-ViT using 63M parameters achieves " + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "inline_equation", + "content": "61.4\\%" + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "text", + "content": " zero-shot top-1 accuracy on ImageNet [7], being " + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "inline_equation", + "content": "2.0 \\times" + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "text", + "content": " smaller and " + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "inline_equation", + "content": "1.5 \\times" + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "text", + "content": " faster than the OpenCLIP [20] model (62.9% accuracy with 126M parameters). Meanwhile, TinyCLIP-ResNet with 38M parameters obtains " + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "inline_equation", + "content": "56.4\\%" + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "text", + "content": " zero-shot top-1 accuracy, being " + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "inline_equation", + "content": "2.1 \\times" + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "text", + "content": " smaller and " + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "inline_equation", + "content": "2.0 \\times" + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "text", + "content": " faster than CLIP ResNet-50 (59.6% accuracy with 76M parameters). Moreover, our method can speed up the original OpenCLIP training by " + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "inline_equation", + "content": "1.4 \\times -7.8 \\times" + }, + { + "bbox": [ + 46, + 491, + 288, + 682 + ], + "type": "text", + "content": " while achieving similar performance. Also, TinyCLIP demonstrates good transfer capacities in downstream scenarios. In summary, the contributions of this work are two-fold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 52, + 689, + 287, + 714 + ], + "type": "text", + "content": "- We propose a new cross-modal distillation approach to unleash the capacity of small CLIP models, fully lever" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 318, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 545, + 109 + ], + "type": "text", + "content": "aging large-scale models as well as pre-training data. To our best knowledge, this is the first work exploring the pre-training distillation of language-image models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 310, + 118, + 547, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 118, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 310, + 118, + 547, + 177 + ], + "type": "text", + "content": "- We present state-of-the-art language-image pre-trained models at small scale, striking the best trade-off between speed and accuracy. Extensive experiments demonstrate the superiority and good generalization ability of the small models in various downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 189, + 392, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 189, + 392, + 202 + ], + "spans": [ + { + "bbox": [ + 306, + 189, + 392, + 202 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 353 + ], + "type": "text", + "content": "Language-Image Pre-training has achieved remarkable progress over the past few years [10, 4]. In particular, contrastive language-image pre-training demonstrates very impressive zero-shot transfer and generalization capacities [45, 46, 21, 72, 61, 43, 74]. One of the most representative works is CLIP [46]. A large amount of follow-up works have been proposed to optimize the pre-training framework [27, 76, 15, 68, 62, 14, 28, 67]. Meanwhile, a line of works leverage the pre-trained models for downstream tasks, such as open-vocabulary detection and segmentation [29, 75, 66, 32], video recognition [56, 23], and text-to-image generation [47, 41]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 354, + 546, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 497 + ], + "type": "text", + "content": "Most recently, there are a few works attempting to scale language-image pre-trained models. BEiT-3 [57] uses a multiway transformer to scales up multi-modal pre-trained models to 1.9B parameters. FLIP [30] scales image-text pre-training via masked modeling, enabling CLIP to enjoy faster training speed while getting better performance. On the other hand, there are also few works on scaling down CLIP models [51]. However, they all focus on specific tasks, such as generation [5, 51] and video content understanding [42]. In contrast, our work concentrates on language-image pre-training, which is the first work on CLIP compression." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 640 + ], + "type": "text", + "content": "Knowledge Distillation in a teacher-student framework [19] has been widely used to transfer knowledge from large models to small ones. Distillation in single modality, such as vision [54, 64, 37] or language [48, 22], has been extensively studied. However, the exploration of cross-modality distillation is relatively limited. A few early works exploring cross-modal distillation only focus on specific tasks, such as VQA and image caption [5, 11, 59], limiting the methodologies to be applied in general downstream tasks. In contrast, this work pays attention to general cross-modal pre-training distillation, in which the distilled models can be transferred to various downstream scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "content": "It is worth noting that our proposed weight inheritance strategy shares similar spirits to model pruning techniques [17, 65, 51]. Both approaches involve the identification of important weights while removing redundant ones. However, they have three fundamental differences. 1) Previous pruning methods predominantly focus on single-modality" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21971" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "models [17, 65, 3], whereas our approach shifts attention to multi-modality. Since different modalities exhibit varying redundancy characteristics, it is necessary to consider these differences when selecting important weights from each modality. 2) Unlike previous pruning methods which have emphasized the importance of the pruned architecture [12, 34, 58], we argue that the weights inherited from the original models are also highly beneficial, particularly when compressing CLIP-like language-image models. 3) Our method leverages a progressive multi-stage process, which enables the inherited weights to accelerate the convergence of small models during training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 226, + 103, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 226, + 103, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 226, + 103, + 239 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 246, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 307 + ], + "type": "text", + "content": "In this section, we propose TinyCLIP, a simple and effective method for distilling large-scale language-image pretraining models, such as CLIP. It consists of three components: affinity mimicking (§3.1), weight inheritance (§3.2) and multi-stage progressive distillation (§3.3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 313, + 239, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 239, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 239, + 327 + ], + "type": "text", + "content": "3.1. Distillation with Affinity Mimicking" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 331, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 498 + ], + "type": "text", + "content": "A language-image pre-training model often consists of two branches, an image encoder to extract visual representations and a text encoder to extract textual representations. The visual and textual representations are linearly projected into a cross-modal common space by minimizing a contrastive loss " + }, + { + "bbox": [ + 46, + 331, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_0" + }, + { + "bbox": [ + 46, + 331, + 287, + 498 + ], + "type": "text", + "content": ". The supervision signal can be viewed as an identity matrix, in which diagonal values are set to 1 for positive image-text pairs, and all others are set to 0 for negative pairs. This signal disregards the similarity between negative pairs, which prevents the model from acquiring a nuanced understanding of the intricate relationships among negative pairs. Therefore, we introduce affinity mimicking to enable student models to learn similarities between negative pairs from teacher models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 499, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 643 + ], + "type": "text", + "content": "More specifically, as shown in Fig. 2, we consider two types of affinity distillation losses: image-to-language loss " + }, + { + "bbox": [ + 46, + 499, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{I2T}" + }, + { + "bbox": [ + 46, + 499, + 287, + 643 + ], + "type": "text", + "content": " and language-to-image loss " + }, + { + "bbox": [ + 46, + 499, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{T2I}" + }, + { + "bbox": [ + 46, + 499, + 287, + 643 + ], + "type": "text", + "content": ". The former learns the alignment between the teacher and student models based on the image-to-language affinity " + }, + { + "bbox": [ + 46, + 499, + 287, + 643 + ], + "type": "inline_equation", + "content": "A_{I2T}" + }, + { + "bbox": [ + 46, + 499, + 287, + 643 + ], + "type": "text", + "content": ", which represents the affinity scores of an image with all the text descriptions in a batch (the yellow row highlighted in Fig. 2). The latter refers to the alignment of language-to-image affinity " + }, + { + "bbox": [ + 46, + 499, + 287, + 643 + ], + "type": "inline_equation", + "content": "A_{T2I}" + }, + { + "bbox": [ + 46, + 499, + 287, + 643 + ], + "type": "text", + "content": ", which compares a text description with all the images to obtain a matching score (the blue column in Fig. 2). The combination of these two losses formulates our affinity mimicking distillation, which is represented as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 652, + 286, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 652, + 286, + 681 + ], + "spans": [ + { + "bbox": [ + 64, + 652, + 286, + 681 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {d i s t i l l}} = \\mathcal {L} _ {I 2 T} + \\mathcal {L} _ {T 2 I}, \\tag {1} \\\\ = C E \\left(A _ {I 2 T} ^ {s}, A _ {I 2 T} ^ {t}\\right) + C E \\left(A _ {T 2 I} ^ {s}, A _ {T 2 I} ^ {t}\\right). \\\\ \\end{array}", + "image_path": "9e2a95b71649394204f9edaf7e65fa3ff96f96e052e2df29f708e940cc77c508.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "CE" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " represents the cross-entropy loss. The superscripts " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " denote the student and teacher models, re" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 323, + 70, + 529, + 231 + ], + "blocks": [ + { + "bbox": [ + 323, + 70, + 529, + 231 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 529, + 231 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 529, + 231 + ], + "type": "image", + "image_path": "070cab7b775b02144ec4257e65c466a94e5563f728d1555c61ff45401e06aae3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 237, + 545, + 259 + ], + "lines": [ + { + "bbox": [ + 305, + 237, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 305, + 237, + 545, + 259 + ], + "type": "text", + "content": "Figure 2. Affinity mimicking for language-image models. The loss includes image-to-text loss (yellow) and text-to-image loss (blue)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 280, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 280, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 280, + 545, + 293 + ], + "type": "text", + "content": "spectively. The elements in " + }, + { + "bbox": [ + 305, + 280, + 545, + 293 + ], + "type": "inline_equation", + "content": "A_{I2T}" + }, + { + "bbox": [ + 305, + 280, + 545, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 280, + 545, + 293 + ], + "type": "inline_equation", + "content": "A_{T2I}" + }, + { + "bbox": [ + 305, + 280, + 545, + 293 + ], + "type": "text", + "content": " are defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 348, + 310, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 310, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 348, + 310, + 545, + 338 + ], + "type": "interline_equation", + "content": "A _ {I 2 T} (i, j) = \\frac {\\exp \\left(I _ {i} \\cdot T _ {j} / \\tau\\right)}{\\sum_ {k \\in \\mathcal {B}} \\exp \\left(I _ {i} \\cdot T _ {k} / \\tau\\right)}, \\tag {2}", + "image_path": "841d2872290085a8c15c41935ab1b84d3011ece30a0da308307750c7062bf3b1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 339, + 545, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 339, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 348, + 339, + 545, + 367 + ], + "type": "interline_equation", + "content": "A _ {T 2 I} (i, j) = \\frac {\\exp \\left(I _ {i} \\cdot T _ {j} / \\tau\\right)}{\\sum_ {k \\in \\mathcal {B}} \\exp \\left(I _ {k} \\cdot T _ {j} / \\tau\\right)}, \\tag {3}", + "image_path": "37a9742b3f3d17e292d303d7186bab03a201ba8d7784983c8dc5cdd3f00721b0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "text", + "content": " is a temperature parameter, " + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "text", + "content": " is the feature embedding of " + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "text", + "content": "-th image in the batch " + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "text", + "content": " is the feature of " + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "text", + "content": "-th text description in " + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 375, + 546, + 471 + ], + "type": "text", + "content": ". Affinity mimicking allows student models to imitate the behavior of the large model in learning visual-linguistic alignment. In contrast to previous methods that rely on either image or text features for distillation, we show that distilling knowledge in this image-text affinity space is more effective." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 479, + 497, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 497, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 497, + 491 + ], + "type": "text", + "content": "3.2. Distillation with Weight Inheritance" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "text", + "content": "The original CLIP models are pre-trained on 400 million image-text pairs for 32 epochs, taking thousands of GPU days. This presents a significant cost challenge for distillation. To improve the training efficiency, we introduce weight inheritance, a technique that inherits the important weights from the well-trained large teacher models to smaller student models. The key challenge of inheriting weights is identifying important weights from the large amount of weights of the teacher. We propose two approaches to select the important weights: manual weight inheritance and automatic weight inheritance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Manual Inheritance. For manual weight inheritance, we first analyze the redundancy of existing CLIP pre-trained models. Fig. 4 shows that the text encoder displays more redundancy in depth (layer-wise), while the image encoder exhibits more redundancy in width (channel-wise). Based upon this discovery, we adopt the approach of uniformly selecting " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " layers of the text branch and directly taking the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21972" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "content": "front " + }, + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "content": " channels of the image branch to select the important weights from large teacher models. These weights will serve as the initialization for small models during distillation. We surprisingly found that such a simple manual selected weights can largely accelerate CLIP distillation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 431 + ], + "type": "text", + "content": "Automatic Inheritance. Although manual inheritance can be considered an effective approach for model compression, it does present the drawback of necessitating prior knowledge in order to determine which weights to inherit. This limitation may render the technique impractical for application to diverse models. To solve this issue, we present an automatic weight inheritance scheme. Inspired by structured pruning in large language models [65], we introduce learnable masks to identify weight importance. Considering the difference across modalities, the learnable masks are imposed on vision and language branches independently, as visualized in Fig. 3. An overall sparsity constraint is introduced to guarantee the selected number of important weights to meet our compression requirements. Without loss of generality, here we use the transformer architecture as an example to introduce the detailed procedure of weight inheritance. Specifically, a standard transformer block contains a multi-head attention (MHA) layer and a feed-forward network (FFN). To capture the importance of weights in a fine-grained level, we introduce two mask variables " + }, + { + "bbox": [ + 46, + 133, + 288, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\mathrm{head}}" + }, + { + "bbox": [ + 46, + 133, + 288, + 431 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 133, + 288, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\mathrm{int}} \\in \\{0,1\\}" + }, + { + "bbox": [ + 46, + 133, + 288, + 431 + ], + "type": "text", + "content": " to identify the redundant attention heads in MHA and neurons in FFN respectively, while keeping the important ones. These two kinds of masks are imposed on the activation of attention heads and the intermediate layer of FFN, which is formulated as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 435, + 287, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 287, + 467 + ], + "type": "interline_equation", + "content": "\\operatorname {M H A} (\\mathbf {X}) = \\sum_ {h = 1} ^ {N _ {H}} \\mathbf {m} _ {\\text {h e a d}} ^ {h} \\cdot \\operatorname {A t t n} _ {\\mathbf {W} _ {Q} ^ {h}, \\mathbf {W} _ {K} ^ {h}, \\mathbf {W} _ {V} ^ {h}, \\mathbf {W} _ {O} ^ {h}} (\\mathbf {X}), \\tag {4}", + "image_path": "7d4c96a183f40ee1d36ab8a7ad38a120d75a9619e158c2ea280e66d4ac96d058.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 470, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 470, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 58, + 470, + 287, + 483 + ], + "type": "interline_equation", + "content": "\\operatorname {F F N} (\\mathbf {X}) = \\operatorname {G e L U} \\left(\\mathbf {X} \\mathbf {W} _ {U}\\right) \\cdot \\operatorname {d i a g} \\left(\\mathbf {m} _ {\\text {i n t}}\\right) \\cdot \\mathbf {W} _ {D}, \\tag {5}", + "image_path": "c905801518ee2560d09198979e6babc987f52bfa0bf091444f57fd4c3bfe60b2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": " is the layer input, " + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_Q^h" + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_K^h" + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_V^h" + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_O^h \\in \\mathbb{R}^{d \\times d_h}" + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": " denote the query, key, value and output matrices in MHA respectively, and " + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_U \\in \\mathbb{R}^{d \\times d_f}" + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_D \\in \\mathbb{R}^{d_f \\times d}" + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": " represent the parameters of FFN layers. Here, " + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": " denotes the hidden dimension (e.g., 768), and " + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "inline_equation", + "content": "d_h = d / N_H" + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": " denotes the output dimension of each head (e.g., 64), where " + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "inline_equation", + "content": "N_H" + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": " is the total number of heads. The trainable masks " + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\mathrm{head}}" + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\mathrm{int}}" + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": " serve as importance indicators for MHA and FFN layers. Moreover, to further learn the importance of embedding dimensions in transformer, we introduce an additional mask " + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_{\\mathrm{embed}} \\in \\{0,1\\}" + }, + { + "bbox": [ + 46, + 488, + 287, + 655 + ], + "type": "text", + "content": ". This mask is shared across all layers because each dimension in the hidden representation is connected to the corresponding dimension in the subsequent layer through a residual connection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 657, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 657, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 657, + 287, + 693 + ], + "type": "text", + "content": "During automatic inheritance, there are two losses used for optimizing the masks and the model jointly, i.e., a sparsity loss and a distillation loss defined in Eq. (1):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 698, + 287, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 698, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 698, + 287, + 711 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {d i s t i l l}} + \\mathcal {L} _ {\\text {s p a r s i t y}}, \\tag {6}", + "image_path": "9fdcb04d4bb5f2c0f713a7bdca141015536206d28a74b79c55476ba6eefb6d9a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 432, + 210 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 432, + 210 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 432, + 210 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 432, + 210 + ], + "type": "image", + "image_path": "a4bf8109bb3809a89651d98aff1e26bcba7e2318a3d6dafdeb9dde1fceb76366.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 213, + 436, + 225 + ], + "lines": [ + { + "bbox": [ + 321, + 213, + 436, + 225 + ], + "spans": [ + { + "bbox": [ + 321, + 213, + 436, + 225 + ], + "type": "text", + "content": "10 Learnable mask/manual mask" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 229, + 545, + 296 + ], + "lines": [ + { + "bbox": [ + 304, + 229, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 545, + 296 + ], + "type": "text", + "content": "Figure 3. Weight inheritance of CLIP. Given a large pre-trained CLIP model, the inherited weights are selected by masks. The weights whose mask is 0 will be removed. For the manual way, the masks are set to 1 in the first k-dimension or uniform k-layer, the rest is set to 0. For the automatic way, the masks are learned jointly with the model under the sparsity constraint." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 433, + 71, + 533, + 210 + ], + "blocks": [ + { + "bbox": [ + 433, + 71, + 533, + 210 + ], + "lines": [ + { + "bbox": [ + 433, + 71, + 533, + 210 + ], + "spans": [ + { + "bbox": [ + 433, + 71, + 533, + 210 + ], + "type": "image", + "image_path": "74d8b9cee14fdbba392dde2ac887ef215c07794afc6610d703dd83cecda7939d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 214, + 531, + 225 + ], + "lines": [ + { + "bbox": [ + 444, + 214, + 531, + 225 + ], + "spans": [ + { + "bbox": [ + 444, + 214, + 531, + 225 + ], + "type": "text", + "content": "Model sparsity " + }, + { + "bbox": [ + 444, + 214, + 531, + 225 + ], + "type": "inline_equation", + "content": "=" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 346, + 318, + 545, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 318, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 346, + 318, + 545, + 332 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s p a r s i t y}} = \\lambda \\cdot (p - q) + \\beta \\cdot (p - q) ^ {2}, \\tag {7}", + "image_path": "dc7df61619cb272eab5b4598f4b4c1a72a85c00f25614d939edf436235612d17.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 340, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 340, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\lambda, \\beta" + }, + { + "bbox": [ + 304, + 340, + 545, + 388 + ], + "type": "text", + "content": " are learnable multipliers that guarantee " + }, + { + "bbox": [ + 304, + 340, + 545, + 388 + ], + "type": "inline_equation", + "content": "p = q[60, 65]" + }, + { + "bbox": [ + 304, + 340, + 545, + 388 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 340, + 545, + 388 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 340, + 545, + 388 + ], + "type": "text", + "content": " is the target compression rate, " + }, + { + "bbox": [ + 304, + 340, + 545, + 388 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 340, + 545, + 388 + ], + "type": "text", + "content": " is the overall compression rate of learnable masks for the model, including image encoder and text encoder:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 395, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 545, + 430 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p = \\frac {1}{M _ {t} + M _ {t}} \\sum_ {i m g, t x t} (4 \\cdot d _ {h} \\cdot \\sum_ {i} ^ {L} \\sum_ {j} ^ {N _ {H}} \\sum_ {k} ^ {d} \\mathbf {m} _ {\\text {h e a d}} ^ {i, j} \\cdot \\mathbf {m} _ {\\text {e m b e d}} ^ {k} \\\\ + 2 \\cdot \\sum_ {i} ^ {L} \\sum_ {j} ^ {d _ {f}} \\sum_ {k} ^ {d} \\mathbf {m} _ {\\text {i n t}} ^ {i, j} \\cdot \\mathbf {m} _ {\\text {e m b e d}} ^ {k}), \\tag {8} \\\\ \\end{array}", + "image_path": "8e8f9d5d0b3732f89aa438ab02b5bfc5a4a2aba15ccd608fd8c2f8f9373c2403.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 436, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 436, + 545, + 472 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 304, + 436, + 545, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 436, + 545, + 472 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 304, + 436, + 545, + 472 + ], + "type": "text", + "content": " represent the full model size of the image encoder and text encoder, respectively, " + }, + { + "bbox": [ + 304, + 436, + 545, + 472 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 436, + 545, + 472 + ], + "type": "text", + "content": " is the total number of transformer layers." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 479, + 494, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 494, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 494, + 493 + ], + "type": "text", + "content": "3.3. Progressive Multi-Stage Distillation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "text", + "content": "When attempting to achieve a high target sparsity, i.e., " + }, + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "inline_equation", + "content": ">70\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "text", + "content": ", compressing the model in a single stage can lead to a significant reduction in accuracy and even result in convergence failure. This is due to the fact that most weights of the large model are directly discarded, including those that are important for ensuring model quality and convergence. As a solution, we propose a multi-stage progressive distillation method to achieve a high compression rate without seriously sacrificing accuracy. In each stage, we use a modest degree of compression, e.g., " + }, + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 498, + 545, + 629 + ], + "type": "text", + "content": ", to avoid large loss of performance and make training stable." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Specifically, each stage includes weight inheritance and affinity mimicking, as described in Algorithm 1. During the weight inheritance phase, the weights of the large model are gradually reduced until the target sparsity and key weights are retained for the small model. After that, the small model is distilled with affinity mimicking, which involves transferring the visual-linguistic affinity knowledge from" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21973" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 50, + 88, + 287, + 365 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 244, + 87 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 244, + 87 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 244, + 87 + ], + "type": "text", + "content": "Algorithm 1: TinyCLIP: Cross-Modal Distillation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "lines": [ + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": "Input: a pre-trained CLIP model " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "f(x;\\theta_0)" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " with weight " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " , target compression rate " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " total iteration steps " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " 1 Initialize the number of stages " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " , each stage steps " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " , target compression rate " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " of stage i, weight inheritance steps " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "L_{M}" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " input " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " , learnable mask " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "M = \\{\\mathbf{m}_{\\mathrm{head}},\\mathbf{m}_{\\mathrm{int}},\\mathbf{m}_{\\mathrm{embed}}\\} ;^1" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " \n2 for i in [1,...,G] do \n3 // Increase mask compression rate, retain important weights \n4 for j in [1,...,LM] do \n5 Update target compression rate " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "^2 q" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " of current step based on " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " and step " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " . \n6 Calculate distillation loss between origin teacher model " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "f(x;\\theta_0)" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " and masked student model " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "f(x;\\theta_{i - 1}\\odot M_i)" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " . \n7 Calculate sparsity loss between learnable mask compression rate " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "^3 p" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " and target compression rate " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " . \n8 Optimize " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\theta_{i - 1}" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " . \n9 end \n10 " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\theta_{i}\\gets \\theta_{i - 1}\\odot M_{i}" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " //Inherit weight with masks \n11 // Cross-Modal distillation \n12 for j in " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "[L_M + 1,\\dots ,L]" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " do \n13 Calculate distillation loss between origin teacher model " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "f(x;\\theta_0)" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " and compressed model " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "f(x;\\theta_i)" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " . \n14 Optimize " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " . \n15 end \n16 end \n17 Return compressed model " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "f(x;\\theta_G)" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " at compression rate " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " \n1 Each stage compress " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "q_{i} = 1 - i\\cdot 25\\%" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "G = \\lceil (1 - q) / 0.25\\rceil" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "L = N / G,L_{M} = 3000,M" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " is initialized to 1, " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " can be adjusted according to different models. \n2 Target compression rate " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " increases linearly from O to " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " with step " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " . \n" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " is calculated by masks " + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 50, + 88, + 287, + 365 + ], + "type": "text", + "content": " in Eq. (8)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 389, + 287, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 414 + ], + "type": "text", + "content": "the teacher to the student. This procedure is repeated until the target sparsity level is achieved." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 422, + 128, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 128, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 128, + 435 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 437, + 180, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 180, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 180, + 451 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 289, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 289, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 289, + 683 + ], + "type": "text", + "content": "Architecture. The original CLIP employs a transformer-based model as its text encoder, while its image encoder has two variations, i.e., ViT [9] and modified ResNet [18], covering both transformer-based and CNN-based architectures. For weight inheritance, we inherited the weights from two model variants: OpenCLIP ViT-B/32 or ViT-B/16 [20] and CLIP ResNet-50 [46], which are pre-trained on LAION-2B [49] and WIT-400M [46], respectively1. For distillation, we use OpenCLIP ViT-B/32 [20] pre-trained on LAION-2B [49] as the teacher model, since it achieves a high zero-shot accuracy of " + }, + { + "bbox": [ + 46, + 456, + 289, + 683 + ], + "type": "inline_equation", + "content": "65.6\\%" + }, + { + "bbox": [ + 46, + 456, + 289, + 683 + ], + "type": "text", + "content": " on ImageNet [7] with a high inference throughput. Besides, we also extend our method to DaViT [8], which is a hybrid architecture. We re-produce Florence-DaViT-5M [72, 8] on LAION-400M as a baseline. Then we distill the same model using affinity mimicking and weight inheritance. The teacher model is Florence-DaViT-D3, which is pre-trained on FLD-900M [72] and achieves a top-1 accuracy of " + }, + { + "bbox": [ + 46, + 456, + 289, + 683 + ], + "type": "inline_equation", + "content": "78.0\\%" + }, + { + "bbox": [ + 46, + 456, + 289, + 683 + ], + "type": "text", + "content": " on ImageNet [7] with 128M parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 684, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 684, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 58, + 684, + 287, + 696 + ], + "type": "text", + "content": "Affinity Mimicking. As defined in Tab. 2, we explore" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "content": "different interaction schemes between teacher and student models across modalities. The contrastive loss " + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_0" + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "content": " and affinity mimicking " + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{distill}}" + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "content": ") serve as the basic ones, which can be combined with other losses. The weight of each loss is set to 1 when combined. The temperature parameter " + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "content": " in Eq. (3) is set to " + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\frac{1}{50}" + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "content": " by default. The sensitivity analysis of " + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "content": " is presented in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 157, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 546, + 312 + ], + "type": "text", + "content": "Weight Inheritance. We compress OpenCLIP ViT-B [20] and CLIP ResNet-50 [46] using the proposed multi-stage progressive distillation. In each stage, we apply manual or automatic weight inheritance. In manual inheritance, we compress the image encoder and the text encoder separately, by reducing the width of image encoder and the depth of text encoder. In automatic inheritance, we set the target sparsity for the whole model. In each stage, the trainable masks are initialized to 1 and updated in the first 3,000 training iterations, where the optimizer is AdamW [35, 24] with a constant learning rate of 0.01 and no weight decay. The learnable multipliers " + }, + { + "bbox": [ + 304, + 157, + 546, + 312 + ], + "type": "inline_equation", + "content": "\\lambda, \\beta" + }, + { + "bbox": [ + 304, + 157, + 546, + 312 + ], + "type": "text", + "content": " defined in Eq. (7) are both initialized to 0.01." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "content": "Training Settings. We train our models on two public datasets, namely LAION-400M [50] and YFCC-15M [53]. On LAION-400M, the models are compressed in 3 stages, including " + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "content": " parameters for 6 epochs, " + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "content": " parameters for 16 epochs, and " + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "content": " parameters for 16 epochs. On YFCC-15M, it contains 2 compression stages, where the training epochs are both 25 from " + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "content": " parameters, and " + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "content": ". We follow the hyper-parameter of CLIP [46] except that the learning rate is set to " + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "content": " when using weight inheritance. The details are shown in the supplementary material. Unless otherwise specified, our ablation studies utilize the same settings, except for training the models only for 1 epoch with a learning rate of " + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 313, + 546, + 504 + ], + "type": "text", + "content": ". All models are trained on 32 Nvidia V100 or A100, and implemented with PyTorch [40], OpenCLIP [20], gradient cache [36], and Timm library [63]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 506, + 547, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 660 + ], + "type": "text", + "content": "Evaluation Settings. The models are evaluated on multiple benchmarks. In zero-shot transfer evaluation and robustness evaluation, we follow the same prompt engineering in CLIP [46], where 80 text templates per class are used for ImageNet [7]. In linear probe, we use Elevater toolkit [26] for evaluation, where a classification head is trained for 50 epochs with searched hyper-parameters. Note that we do not count the number of parameters in the text embedding layer. It is a look-up table whose parameter size is the same as the models with the same hidden dimension and vocabulary size. The inference throughput is measured on Nvidia V100 with CUDA 11 and PyTorch v1.12 [40], where the batch size is 1,024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 670, + 524, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 524, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 524, + 683 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We compare our TinyCLIP with state-of-the-art models in Tab. 1. The zero-shot evaluation on ImageNet-1K [7]," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 262, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 262, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 262, + 713 + ], + "type": "text", + "content": "1The models can be downloaded from OpenCLIP and OpenAI." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21974" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 60, + 545, + 280 + ], + "blocks": [ + { + "bbox": [ + 48, + 60, + 545, + 280 + ], + "lines": [ + { + "bbox": [ + 48, + 60, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 48, + 60, + 545, + 280 + ], + "type": "table", + "html": "
MethodImageEncoderText Encoder#Params (M)Image+TextMACs(G)Throughput(pairs/s)TrainingdatasetsIN-1Ktop1 acc (%)Flickr30kMSCOCO
depthwidthI→T@1T→I@1I→T@1T→I@1
Training data: 15M / 20M
RILS [69]ViT-B/161251286 + 3820.5818LAION-20M45.045.134.932.225.5
MaskCLIP [77]ViT-B/161251286 + 3820.5818LAION-20M46.664.948.138.524.8
SLIP [38]ViT-B/161251286 + 3820.5818YFCC-15M42.857.640.131.120.3
CLIP [46, 38]ViT-B/161251286 + 3820.5818YFCC-15M37.651.632.226.517.1
TinyCLIP (Ours)ViT-39M/16651239+19(2.1×)9.5(2.2×)1,469(1.8×)YFCC-15M63.584.466.754.938.9
TinyCLIP (Ours)ViT-8M/1632568+3(11.3×)2.0(10.3×)4,150(5.1×)YFCC-15M41.162.342.336.221.5
Training data: 400M
Florence [72]DaViT-5M62565+51.12,980LAION-400M45.061.241.536.220.9
TinyCLIP (Ours)DaViT-5M62565+51.12,980LAION-400M50.066.248.540.724.5
CLIP [46]ResNet-1011251256 + 3812.81,161WIT-400M62.278.159.249.330.7
CLIP [46]ResNet-501251238 + 389.11,549WIT-400M59.681.258.250.828.3
TinyCLIP (Ours)ResNet-30M951230+29(1.3×)6.9(1.3×)1,811(1.2×)LAION-400M59.181.161.252.733.9
TinyCLIP (Ours)ResNet-19M651219+19(2.0×)4.4(2.1×)3,024(2.0×)LAION-400M56.476.258.348.930.9
OpenCLIP [20]ViT-B/321251288 + 387.42,452LAION-2B65.784.766.856.939.3
CLIP [46]ViT-B/321251288 + 387.42,452WIT-400M63.280.159.851.230.6
OpenCLIP [20]ViT-B/321251288 + 387.42,452LAION-400M62.979.362.053.335.4
TinyCLIP (Ours)ViT-61M/32951261+29(1.4×)5.3(1.4×)3,191(1.3×)LAION-400M62.178.663.353.935.9
TinyCLIP (Ours)ViT-40M/32651240+19(2.1×)3.5(2.1×)4,641(1.9×)LAION-400M59.777.358.949.833.1
TinyCLIP (Ours)ViT-63M/32autoauto63+31(1.3×)5.6(1.3×)2,905(1.2×)LAION-400M63.983.264.455.537.6
TinyCLIP (Ours)ViT-45M/32autoauto45+18(2.0×)3.7(2.0×)3,682(1.5×)LAION-400M61.480.962.252.834.7
TinyCLIP (Ours)ViT-22M/32autoauto22+10(3.9×)1.9(3.9×)5,504(2.2×)LAION-400M53.771.352.044.428.3
TinyCLIP (Ours)ViT-63M/32autoauto63+31(1.3×)5.6(1.3×)2,909(1.2×)LAION+YFCC-400M64.584.966.056.938.5
TinyCLIP (Ours)ViT-45M/32autoauto45+18(2.0×)3.7(2.0×)3,685(1.5×)LAION+YFCC-400M62.780.363.954.036.7
", + "image_path": "d11a2489ce94c466333d7ce98d971982d9e21178964ad457aa6ed5f749e0f2c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 283, + 546, + 306 + ], + "lines": [ + { + "bbox": [ + 47, + 283, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 546, + 306 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-art methods. The architecture of text encoder is Transformer [55]. \"auto\" denotes automatic weight inheritance, and our other models use manual weight inheritance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 325, + 276, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 276, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 276, + 337 + ], + "type": "text", + "content": "Flickr30k [44] and MSCOCO retrieval [33] are reported." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "text", + "content": "When performing training on YFCC-15M [53], we use affinity mimicking and manual weight inheritance to distill OpenCLIP ViT-B/16 [20], whose zero-shot performance on ImageNet is " + }, + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "inline_equation", + "content": "70.2\\%" + }, + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "text", + "content": ". By inheriting half of parameters, our TinyCLIP ViT-39M/16 can achieve a zero-shot accuracy of " + }, + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "inline_equation", + "content": "63.5\\%" + }, + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "text", + "content": " on ImageNet, surpassing the original CLIP ViT-B/16 [46] by " + }, + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "inline_equation", + "content": "25.9\\%" + }, + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "text", + "content": ". Furthermore, when compressing the image encoder to 10 layers with 256 dimensions and the text encoder to 3 layers with 256 dimensions, TinyCLIP ViT-8M/16 still surpasses CLIP ViT-B/16 [46] by " + }, + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "text", + "content": " while using " + }, + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "inline_equation", + "content": "11.3\\times" + }, + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "text", + "content": " fewer parameters and running " + }, + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "inline_equation", + "content": "5.1\\times" + }, + { + "bbox": [ + 46, + 338, + 288, + 470 + ], + "type": "text", + "content": " faster." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 590 + ], + "type": "text", + "content": "When conducting training on LAION-400M [50], we distill three different models, i.e., Florence DaViT [72, 8], CLIP ResNet [46], and OpenCLIP ViT-B/32 [20]. The results in Tab. 1 demonstrate the efficacy of our method. More specifically, our TinyCLIP DaViT-5M achieves up to " + }, + { + "bbox": [ + 46, + 472, + 287, + 590 + ], + "type": "inline_equation", + "content": "5.0\\%" + }, + { + "bbox": [ + 46, + 472, + 287, + 590 + ], + "type": "text", + "content": " improvements over the original model on ImageNet, and " + }, + { + "bbox": [ + 46, + 472, + 287, + 590 + ], + "type": "inline_equation", + "content": "5.0 / 4.5\\%" + }, + { + "bbox": [ + 46, + 472, + 287, + 590 + ], + "type": "text", + "content": " gains in image-to-text retrieval on Flicker and MSCOCO, respectively. It is worth noting that this is a tiny language-image pre-trained model using only 10M parameters to achieve " + }, + { + "bbox": [ + 46, + 472, + 287, + 590 + ], + "type": "inline_equation", + "content": "50.0\\%" + }, + { + "bbox": [ + 46, + 472, + 287, + 590 + ], + "type": "text", + "content": " accuracy on ImageNet." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "text", + "content": "For CLIP ResNet, TinyCLIP-30M is only trained for 4 epochs in the first stage, inheriting " + }, + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "text", + "content": " parameters from the original model. The results indicate that there is " + }, + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "text", + "content": " slight decrease in top-1 accuracy on ImageNet, however, has " + }, + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\sim 2\\%" + }, + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "text", + "content": " improvements in retrieval tasks. TinyCLIP ResNet-19M inherits the weights from the 30M model, and it is trained for 12 epochs. It reduces the parameters by " + }, + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "text", + "content": " while getting " + }, + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "text", + "content": " inference speedup. But it gets " + }, + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 46, + 592, + 288, + 700 + ], + "type": "text", + "content": " performance drops on ImageNet due to the large reduction in parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 700, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 700, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 700, + 287, + 713 + ], + "type": "text", + "content": "For OpenCLIP ViT-B/32, we distill it using our affin" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 325, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 547, + 529 + ], + "type": "text", + "content": "ity mimicking method with manual or automatic weight inheritance. We design a three-stage progressive distillation in which each stage compresses the model by " + }, + { + "bbox": [ + 304, + 325, + 547, + 529 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 325, + 547, + 529 + ], + "type": "text", + "content": ". The three stages are trained for 6, 16, and 16 epochs, respectively. In manual inheritance, TinyCLIP ViT-61M/32 obtains comparable performance compared to OpenCLIP ViT-B/32 [20], while reducing the parameters by " + }, + { + "bbox": [ + 304, + 325, + 547, + 529 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 304, + 325, + 547, + 529 + ], + "type": "text", + "content": ". On the other hand, automatic inheritance obtains better results than the manual one. For example, TinyCLIP ViT-63M/32 outperforms the manually inherited ViT-61M/32 model by more than " + }, + { + "bbox": [ + 304, + 325, + 547, + 529 + ], + "type": "inline_equation", + "content": "\\sim 1\\%" + }, + { + "bbox": [ + 304, + 325, + 547, + 529 + ], + "type": "text", + "content": " on both classification and retrieval tasks. When auto-compressed the model by " + }, + { + "bbox": [ + 304, + 325, + 547, + 529 + ], + "type": "inline_equation", + "content": "3.9\\times" + }, + { + "bbox": [ + 304, + 325, + 547, + 529 + ], + "type": "text", + "content": ", TinyCLIP ViT-22M/32 obtains the highest inference throughput of 5,504 and reaches " + }, + { + "bbox": [ + 304, + 325, + 547, + 529 + ], + "type": "inline_equation", + "content": "53.7\\%" + }, + { + "bbox": [ + 304, + 325, + 547, + 529 + ], + "type": "text", + "content": " top-1 accuracy. Since YFCC-15M contains high-quality image-text pairs, if using it to replace parts of LAION data, our method can obtain additional improvements, especially for retrieval tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 538, + 400, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 400, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 400, + 551 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "text", + "content": "Impact of affinity mmicking. As shown in Tab. 2, we present four different interaction modes. The affinity mimicking " + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{distill}" + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "text", + "content": ") outperforms the contrastive loss " + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_0" + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "text", + "content": " in terms of top-1 accuracy on ImageNet [7]. The alignment in a visual-linguistic affinity space brings more similarity information. The accuracy of cross modalities interaction is lower than that of affinity mimicking by " + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "text", + "content": ". In this mode, the embedding features of student image and text do not have any interaction, but they are aligned to the teacher's embedding space. Single modality interaction has lower accuracy due to missing a loss to align student's image and text features in common embedding space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 701, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 547, + 714 + ], + "type": "text", + "content": "Impact of weight inheritance. We verify the efficacy of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21975" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 142 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 142 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 142 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 142 + ], + "type": "table", + "html": "
Interaction ModeLoss FormulaTop-1 Acc
Contrastive loss [46]L0=CE(<Is,Ts>,I)53.4 %
Affinity mimickingL1=CE(<Is,Ts>,<It,Tt>)55.5 %
Cross modalitiesL2=CE(<Is,Tt>,<It,Tt>)55.3 %
L3=CE(<It,Ts>,<It,Tt>)
Single modalityL4=CE(<Is,I>.I)19.2 %
L5=CE(<Ts,Tt>,I)
", + "image_path": "f0383c934ae31221e4d5fbbe1701183c6ee60c37b95d2c3a72c277cd119f90d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 243, + 164, + 316 + ], + "blocks": [ + { + "bbox": [ + 53, + 243, + 164, + 316 + ], + "lines": [ + { + "bbox": [ + 53, + 243, + 164, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 164, + 316 + ], + "type": "image", + "image_path": "2094b009db3040487fccea5795494dd3488ccf20ac0fe53ab637b079f617025f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 324, + 287, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 369 + ], + "type": "text", + "content": "Figure 4. The cosine similarity of the input and output embedding of each layer, and the zero-shot accuracy on ImageNet [7] after removing a layer for OpenCLIP ViT-B-32 [20, 46] pretrained on LAION-2B [49]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 243, + 282, + 316 + ], + "blocks": [ + { + "bbox": [ + 173, + 243, + 282, + 316 + ], + "lines": [ + { + "bbox": [ + 173, + 243, + 282, + 316 + ], + "spans": [ + { + "bbox": [ + 173, + 243, + 282, + 316 + ], + "type": "image", + "image_path": "c41e5c3551a412d1d91981f688f44bca3fd4144f14f686a9202939f77b8359ae.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 380, + 288, + 434 + ], + "blocks": [ + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "type": "text", + "content": "Table 2. The interaction of information across image " + }, + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "type": "text", + "content": " modalities in student and teacher models. The subscripts " + }, + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "type": "text", + "content": " denote student and teacher, respectively. " + }, + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "type": "text", + "content": " is an identity matrix. " + }, + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "type": "inline_equation", + "content": "CE" + }, + { + "bbox": [ + 46, + 150, + 288, + 228 + ], + "type": "text", + "content": " is cross entropy function. TinyCLIP-ViT-40M/32 is trained with an interaction mode and manual weight inheritance on LAION-400M [50] for 1 epoch. The zero-shot accuracy on ImageNet [7] is reported." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 380, + 288, + 434 + ], + "lines": [ + { + "bbox": [ + 48, + 380, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 48, + 380, + 288, + 434 + ], + "type": "table", + "html": "
Weight InheritanceIN-1K top1 acc (%)Flickr30kMSCOCO
I→T@1T→I@1I→T@1T→I@1
None37.850.233.429.716.6
Manual54.070.452.946.529.3
Automatic54.973.555.947.930.8
", + "image_path": "3d5b23431a66cd2bf8ebfc47429b3d831da05da80a12e7a9e3f61e3314dbd9be.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": "the proposed two kinds of weight inheritance in Tab. 3. Compared to the model without using weight inheritance, manual inheritance brings " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "16.2\\%" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": " accuracy improvements on ImageNet [7]. Automatic weight inheritance can further improve the accuracy by " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": ". On image-text retrievals, the improvements of automatic inheritance is relatively higher, i.e., " + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "inline_equation", + "content": "1.4 - 3.1\\%" + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": " over the manual method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 569, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 688 + ], + "type": "text", + "content": "The redundancy of language-image models. We analyze the redundancy of image encoder and text encoder by similar strategy [6, 2, 25] and zero-shot accuracy by removing a layer. As shown in Fig. 4, we observe a high correlation between cosine similarity and accuracy. Comparing with image encoder, text encoder has higher cosine similarity. It shows the redundancy of the text encoder is higher than that of the image encoder. Consequently, compressing the text encoder along the depth dimension can be achieved without significant performance degradation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "We further explore the redundancy by the proposed automatic weight inheritance. As shown in Fig. 5, for width" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 539, + 251 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 539, + 251 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 539, + 251 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 539, + 251 + ], + "type": "image", + "image_path": "5a5866aefa3bf45f4445ee87701ace50ceb2739df2218dfcfdae2cb611321a2e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 255, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 305, + 255, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 545, + 300 + ], + "type": "text", + "content": "Figure 5. Visualization of learnable masks at " + }, + { + "bbox": [ + 305, + 255, + 545, + 300 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 255, + 545, + 300 + ], + "type": "text", + "content": " sparsity on hidden dimension, attention heads and MLP intermediate dimension. The blocks in red color are removed. The original model is OpenCLIP ViT-B/32 [20]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 326, + 308, + 518, + 384 + ], + "blocks": [ + { + "bbox": [ + 46, + 435, + 287, + 468 + ], + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 468 + ], + "type": "text", + "content": "Table 3. Manual inheritance vs. Automatic inheritance. The model OpenCLIP ViT-B/32 [20] is compressed by " + }, + { + "bbox": [ + 46, + 435, + 287, + 468 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 435, + 287, + 468 + ], + "type": "text", + "content": " with distillation on LAION-400M [50] for 1 epoch." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 308, + 518, + 384 + ], + "lines": [ + { + "bbox": [ + 326, + 308, + 518, + 384 + ], + "spans": [ + { + "bbox": [ + 326, + 308, + 518, + 384 + ], + "type": "table", + "html": "
Pretrained Teacher ModelTeacher Acc.(%)Inherited RatioStudent Acc.(%)
w/o weight inheritance-036.2
CLIP ViT-B/32 [46]63.259 / 12652.4(+16.2)
OpenCLIP ViT-B/32 [20]62.959 / 12653.5(+17.3)
OpenCLIP ViT-B/16 [20]67.159 / 12452.8(+16.6)
OpenCLIP ViT-L/14 [20]75.359 / 39045.1(+8.9)
OpenCLIP ViT-H/14 [20]78.059 / 93541.1(+4.9)
", + "image_path": "fc701be2b2d27a0e1c144ab3dc3f5062fd2685e15e912a1a1e861b59664ec540.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 387, + 545, + 443 + ], + "lines": [ + { + "bbox": [ + 305, + 387, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 545, + 443 + ], + "type": "text", + "content": "Table 4. Ablation study on inherited teacher model. The student model TinyCLIP ViT-40M/32 with 59M parameters is inherited from a teacher model, then trained without distillation on LAION-400M [50] for 1 epoch. The top-1 zero-shot accuracy on ImageNet [7] is reported." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "type": "text", + "content": "analysis, it is observed that the embedding channels of the text encoder experience only a marginal reduction (e.g., 512 " + }, + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "type": "text", + "content": " 511), whereas the embedding channels of the image encoder are reduced to a greater extent (e.g., " + }, + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "type": "inline_equation", + "content": "768\\to 526" + }, + { + "bbox": [ + 304, + 461, + 545, + 581 + ], + "type": "text", + "content": "). Regarding depth analysis, it was observed that a significant proportion of MHA layers and FFN channels could be removed from the text encoder, indicating redundancy within it. This finding suggests that the image encoder can be compressed along the width dimension, while the text encoder can be compressed along the depth dimension." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 677 + ], + "type": "text", + "content": "Impact of teacher models. We investigate which model is better for weight inheritance. As analyzed in Tab. 4, a strong teacher does not necessarily guarantee better results for weight inheritance. Instead, teacher models with similar architectures and higher performance offer a better choice. For instance, although ViT-H/14 ranks as the highest-performing teacher model, it lags behind other models in terms of weight inheritance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Impact of multi-stage progressive distillation. Multi-stage progressive distillation shrinks the gap between the compressed model and the inherited model in each stage." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21976" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 60, + 545, + 202 + ], + "blocks": [ + { + "bbox": [ + 50, + 60, + 545, + 202 + ], + "lines": [ + { + "bbox": [ + 50, + 60, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 50, + 60, + 545, + 202 + ], + "type": "table", + "html": "
MethodImage EncoderFood101CIFAR10CIFAR100SUN397Stanford CarsFGYC AircraftVOC207DTDOxford PetsCaltech101Flowers102MNISTFER2013STL10EuroSATRESISC45GTSRBKITTICountry211PCamHateful MemesRendered SST2ImageNet
Zero-shot performance
DeCLIP [31]ViT-B/3276.390.067.963.550.68.980.545.384.489.383.212.816.397.540.048.410.335.311.554.752.950.166.1
FILIP [70]ViT-B/3282.886.965.569.155.557.2-49.388.191.985.3----49.9------68.8
CLIP [46]ViT-B/3284.491.365.163.259.421.283.144.587.087.966.751.947.397.249.460.332.239.417.858.457.659.663.2
OpenCLIP [20]ViT-B/3280.990.770.666.879.316.682.254.486.590.166.037.442.395.651.657.642.031.614.850.152.952.362.9
TinyCLIP (Ours)ViT-45M/3278.392.271.866.079.923.083.253.885.388.760.762.545.296.252.261.841.116.916.551.553.552.661.2
Linear probe performance
CLIP [46]ViT-B/3288.895.180.576.681.852.087.776.590.093.096.999.069.298.397.090.585.366.227.883.966.770.876.1
TinyCLIP (Ours)ViT-45M/3284.097.486.068.985.439.085.471.289.993.482.199.566.997.697.588.198.579.716.190.458.462.573.2
", + "image_path": "700152c4b5b97d09ed66e1e251fef22500307f11cf798b3c7dbbe5993cbb1de7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 232, + 285, + 291 + ], + "blocks": [ + { + "bbox": [ + 151, + 209, + 441, + 220 + ], + "lines": [ + { + "bbox": [ + 151, + 209, + 441, + 220 + ], + "spans": [ + { + "bbox": [ + 151, + 209, + 441, + 220 + ], + "type": "text", + "content": "Table 5. Zero-shot and linear-probe classification top-1 accuracy on 23 datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 232, + 285, + 291 + ], + "lines": [ + { + "bbox": [ + 48, + 232, + 285, + 291 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 285, + 291 + ], + "type": "table", + "html": "
Number of StagesModel SizeEpochImageNet top1 acc (%)MSCOCO
I→TR@1T→IR@1
1100% → 25%247.039.423.1
2100% → 50%154.046.529.3
50% → 25%149.040.225.3
", + "image_path": "dcdf0d9ad944d524a0da6d52e351cd32fde894c79f158c9b0d0789395a903396.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 331, + 264, + 410 + ], + "blocks": [ + { + "bbox": [ + 70, + 331, + 264, + 410 + ], + "lines": [ + { + "bbox": [ + 70, + 331, + 264, + 410 + ], + "spans": [ + { + "bbox": [ + 70, + 331, + 264, + 410 + ], + "type": "image", + "image_path": "6035fb3cd23eef3dc6f6b2890861ce5b6c7fe80d3d2ac9fb61b2272097f96e64.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 416, + 287, + 460 + ], + "lines": [ + { + "bbox": [ + 47, + 416, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 287, + 460 + ], + "type": "text", + "content": "Figure 6. Training efficiency. OpenCLIP ViT-B/32 [20] is trained from scratch on LAION-400M [50] for 32 epochs. TinyCLIP ViT/32 is trained for 6+16 epochs using the proposed 2-stage progressive distillation, where the teacher is OpenCLIP ViT-B/32." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 472, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 287, + 544 + ], + "type": "text", + "content": "We compare it with single-stage in Tab. 6. Using the same training epochs, the model trained with 2-stage surpasses that with 1-stage by " + }, + { + "bbox": [ + 47, + 472, + 287, + 544 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 47, + 472, + 287, + 544 + ], + "type": "text", + "content": " in terms of zero-shot accuracy on ImageNet. It demonstrates that multi-stage progressive compression is superior when transferring the knowledge of the pre-trained model to the compressed small model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "text", + "content": "Training cost. We also study the distillation efficiency. As shown in Fig. 6, the original OpenCLIP ViT-B/32 [20] takes around 287 GPU days when trained on LAION-400M for 32 epochs. Our models with " + }, + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "text", + "content": " parameters reach the similar accuracy as OpenCLIP counterparts, getting a speed up of " + }, + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "inline_equation", + "content": "7.8\\times" + }, + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "inline_equation", + "content": "1.4\\times" + }, + { + "bbox": [ + 46, + 544, + 287, + 664 + ], + "type": "text", + "content": ", respectively. The underlying reason is that our training benefits more from the affinity mimicking supervision and the good initialization using weight inheritance. This also proves that weight inheritance is capable of accelerating cross-modal distillation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 671, + 194, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 194, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 194, + 684 + ], + "type": "text", + "content": "4.4. Transfer Learning Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Classification on 23 datasets. We also evaluate the performance of zero-shot and linear-probe classification on 23" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 232, + 545, + 283 + ], + "blocks": [ + { + "bbox": [ + 47, + 293, + 287, + 326 + ], + "lines": [ + { + "bbox": [ + 47, + 293, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 287, + 326 + ], + "type": "text", + "content": "Table 6. Ablation study on multi-stage progressive distillation. The model OpenCLIP ViT-B/32 [20] is compressed to " + }, + { + "bbox": [ + 47, + 293, + 287, + 326 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 47, + 293, + 287, + 326 + ], + "type": "text", + "content": " model size on LAION-400M [50] for 2 epochs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 232, + 545, + 283 + ], + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 283 + ], + "type": "table", + "html": "
ModelIN-V2IN-AIN-RObjectNetIN-Sketch
CLIP ViT-B/32 [46]56.031.669.429.942.3
OpenCLIP ViT-B/32 [20]55.121.773.428.949.4
TinyCLIP ViT-63M/32(Ours)55.722.874.131.250.8
TinyCLIP ViT-45M/32(Ours)52.619.871.529.148.8
", + "image_path": "97f99dcbe92ebafa254d62cfa85b6e17085aa98cadbcaf15cb8d1b4ef278a8a1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 291, + 545, + 325 + ], + "lines": [ + { + "bbox": [ + 305, + 291, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 305, + 291, + 545, + 325 + ], + "type": "text", + "content": "Table 7. Zero-shot robustness evaluation. CLIP [46] is trained on WIT-400M, OpenCLIP [20] and TinyCLIP is trained on LAION-400M [50]. Top-1 accuracy is reported." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 336, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 396 + ], + "type": "text", + "content": "datasets by the toolkit Elevater [26]. As shown in Tab. 5, our model wins on 7 datasets in zero-shot benchmark, and 9 datasets on linear-probe benchmark. The performance of our model is close to OpenCLIP ViT-B/32 [20], where the two models are both trained on Laion-400M [50]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 397, + 546, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 546, + 505 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 546, + 505 + ], + "type": "text", + "content": "Zero-shot robustness evaluation. We verify the robustness of our models on 5 datasets, which are out of ImageNet distribution. As shown in Tab. 7, our TinyCLIP ViT-63M/32 outperforms OpenCLIP ViT-B/32 [20], and wins 3 datasets when compared with CLIP ViT-B/32 [46]. Our TinyCLIP ViT-45M/32 uses " + }, + { + "bbox": [ + 304, + 397, + 546, + 505 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 397, + 546, + 505 + ], + "type": "text", + "content": " fewer parameters yet still wins on IN-R and IN-Sketch, compared to CLIP ViT-B/32. This demonstrates the robustness of the distilled TinyCLIP models and the efficacy of our proposed distillation method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 510, + 383, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 383, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 383, + 522 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 639 + ], + "type": "text", + "content": "In this paper, we propose TinyCLIP, a method for distilling large-scale vision-language pre-trained models. It mainly introduced two core techniques named affinity mimicking and weight inheritance. Extensive experiments and ablation studies have demonstrated the effectiveness of TinyCLIP, showing that it can largely reduce model size while maintaining competitive performance. In the future, we will explore ways to further improve cross-modal distillation efficiency on extremely small models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 645, + 403, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 403, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 403, + 658 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Prof. Hongyang Chao was partially supported by NSFC (U1611461, 61672548) and NSFC (U22A2095). Prof. Xinggang Wang was partly supported by NSFC (No. 62276108)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21977" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 156 + ], + "type": "text", + "content": "[2] Shaojie Bai, J Zico Kolter, and Vladlen Koltun. Deep equilibrium models. NeurIPS, 32, 2019. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 287, + 201 + ], + "type": "text", + "content": "[3] Arnav Chavan, Zhiqiang Shen, Zhuang Liu, Zechun Liu, Kwang-Ting Cheng, and Eric P Xing. Vision transformer slimming: Multi-dimension searching in continuous optimization space. In CVPR, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 203, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 287, + 246 + ], + "type": "text", + "content": "[4] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 280 + ], + "type": "text", + "content": "[5] Wenliang Dai, Lu Hou, Lifeng Shang, Xin Jiang, Qun Liu, and Pascale Fung. Enabling multimodal generation on clip via vision-language knowledge distillation. In ACL, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 281, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 287, + 312 + ], + "type": "text", + "content": "[6] Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Lukasz Kaiser. Universal transformers. In ICLR, 2019. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 314, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 346 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 347, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 379 + ], + "type": "text", + "content": "[8] Mingyu Ding, Bin Xiao, Noel Codella, Ping Luo, Jingdong Wang, and Lu Yuan. Davit: Dual attention vision transformers. In ECCV, 2022. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 435 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 287, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 287, + 467 + ], + "type": "text", + "content": "[10] Yifan Du, Zikang Liu, Junyi Li, and Wayne Xin Zhao. A survey of vision-language pre-trained models. In *IJCAI*, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 469, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 469, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 469, + 287, + 512 + ], + "type": "text", + "content": "[11] Zhiyuan Fang, Jianfeng Wang, Xiaowei Hu, Lijuan Wang, Yezhou Yang, and Zicheng Liu. Compressing visual-linguistic model via knowledge distillation. In ICCV, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "type": "text", + "content": "[12] Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In ICLR, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "type": "text", + "content": "[13] Prakhar Ganesh, Yao Chen, Xin Lou, Mohammad Ali Khan, Yin Yang, Hassan Sajjad, Preslav Nakov, Deming Chen, and Marianne Winslett. Compressing large-scale transformer-based models: A case study on bert. ACL, 9, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 624 + ], + "type": "text", + "content": "[14] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "text", + "content": "[15] Yunhao Ge, Jie Ren, Ming-Hsuan Yang, Yuxiao Wang, Andrew Gallagher, Hartwig Adam, Laurent Itti, Balaji Lakshminarayanan, and Jiaping Zhao. Improving zero-shot generalization and robustness of multi-modal models. In NeurIPS ML Safety Workshop, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[16] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. IJCV, 129:1789-1819, 2021. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[17] Manish Gupta and Puneet Agrawal. Compression of deep learning models for text: A survey. TKDD, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 130, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 152 + ], + "type": "text", + "content": "[19] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. arXiv, 2015. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 153, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 207 + ], + "type": "text", + "content": "[20] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 252 + ], + "type": "text", + "content": "[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 254, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 296 + ], + "type": "text", + "content": "[22] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. Tinybert: Distilling bert for natural language understanding. In EMNLP, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "type": "text", + "content": "[23] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In ECCV, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "type": "text", + "content": "[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors, ICLR, 2015. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 366, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 545, + 409 + ], + "type": "text", + "content": "[25] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. ICLR, 2020. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "text", + "content": "[26] Chunyuan Li, Haotian Liu, Liunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, and Jianfeng Gao. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. NeurIPS, 2022. 5, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "type": "text", + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "text", + "content": "[28] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Deepak Gotmare, Shafiq Joty, Caiming Xiong, and Steven Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 557, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 610 + ], + "type": "text", + "content": "[29] Liunian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "type": "text", + "content": "[30] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. arXiv preprint arXiv:2212.00794, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "text", + "content": "[31] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[32] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21978" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "text", + "content": "Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. arXiv preprint arXiv:2210.04150, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 184 + ], + "type": "text", + "content": "[34] Zhuang Liu, Mingjie Sun, Tinghui Zhou, Gao Huang, and Trevor Darrell. Rethinking the value of network pruning. In ICLR, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 286, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 286, + 207 + ], + "type": "text", + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 241 + ], + "type": "text", + "content": "[36] Jiawei Han Jamie Callan Luyu Gao, Yunyi Zhang. Scaling deep contrastive learning batch size under memory limited setup. In RepL4NLP, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "type": "text", + "content": "[37] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 5191-5198, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 299, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 331 + ], + "type": "text", + "content": "[38] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 333, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 376 + ], + "type": "text", + "content": "[39] Basil Mustafa, Carlos Riquelme Ruiz, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 378, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 432 + ], + "type": "text", + "content": "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 466 + ], + "type": "text", + "content": "[41] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In ICCV, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "type": "text", + "content": "[42] Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, and Youliang Yan. Clipping: Distilling clip-based models for video-language understanding. OpenReview, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 513, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 556 + ], + "type": "text", + "content": "[43] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, et al. Combined scaling for open-vocabulary image classification. arXiv, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 558, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 610 + ], + "type": "text", + "content": "[44] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[45] Di Qi, Lin Su, Jia Song, Edward Cui, Taroon Bharti, and Arun Sacheti. Imagebert: Cross-modal pre-training with large-scale weak-supervised image-text data. arXiv preprint arXiv:2001.07966, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[47] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[48] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 239 + ], + "type": "text", + "content": "[49] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "text", + "content": "[50] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "text", + "content": "[51] Dachuan Shi, Chaofan Tao, Ying Jin, Zhendong Yang, Chun Yuan, and Jiaqi Wang. Upop: Unified and progressive pruning for compressing vision-language transformers. arXiv preprint arXiv:2301.13741, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 385 + ], + "type": "text", + "content": "[52] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "text", + "content": "[53] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 2016. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "type": "text", + "content": "[54] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 477, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 545, + 510 + ], + "type": "text", + "content": "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 511, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 543 + ], + "type": "text", + "content": "[56] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 545, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 600 + ], + "type": "text", + "content": "[57] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "text", + "content": "[58] Yulong Wang, Xiaolu Zhang, Lingxi Xie, Jun Zhou, Hang Su, Bo Zhang, and Xiaolin Hu. Pruning from scratch. In AAAI, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[59] Zhecan Wang, Noel Codella, Yen-Chun Chen, Luowei Zhou, Jianwei Yang, Xiyang Dai, Bin Xiao, Haoxuan You, Shih-Fu Chang, and Lu Yuan. Clip-td: Clip targeted distillation for vision-language tasks. arXiv preprint arXiv:2201.05729, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[60] Ziheng Wang, Jeremy Wohlwend, and Tao Lei. Structured pruning of large language models. In EMNLP, 2020. 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21979" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[61] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. In ICLR, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[62] Yixuan Wei, Yue Cao, Zheng Zhang, Zhuliang Yao, Zhenda Xie, Han Hu, and Baining Guo. icar: Bridging image classification and image-text alignment for visual recognition. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 246, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 246, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 246, + 163 + ], + "type": "text", + "content": "[63] Ross Wightman. Pytorch image models, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 164, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 206 + ], + "type": "text", + "content": "[64] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "text", + "content": "[65] Mengzhou Xia, Zexuan Zhong, and Danqi Chen. Structured pruning learns compact and accurate models. In ACL, 2022. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 286 + ], + "type": "text", + "content": "[66] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In ECCV, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 287, + 331 + ], + "type": "text", + "content": "[67] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 365 + ], + "type": "text", + "content": "[68] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In CVPR, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 367, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 367, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 367, + 287, + 409 + ], + "type": "text", + "content": "[69] Shusheng Yang, Yixiao Ge, Kun Yi, Dian Li, Ying Shan, Xiaohu Qie, and Xinggang Wang. Masked visual reconstruction in language semantic space. arXiv preprint arXiv:2301.06958, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 455 + ], + "type": "text", + "content": "[70] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2022. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "type": "text", + "content": "[71] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "text", + "content": "[72] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "type": "text", + "content": "[73] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. arXiv preprint arXiv:2111.08276, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 591, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 634 + ], + "type": "text", + "content": "[74] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In CVPR, pages 18123-18133, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[75] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[76] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "type": "text", + "content": "[77] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In ECCV, 2022. 6" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21980" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/d5cc5a6e-449b-4085-98b6-42b0071a0627_content_list.json b/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/d5cc5a6e-449b-4085-98b6-42b0071a0627_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..91957163f7b921b3d1a0c8ec07af55e98b42d0e6 --- /dev/null +++ b/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/d5cc5a6e-449b-4085-98b6-42b0071a0627_content_list.json @@ -0,0 +1,1480 @@ +[ + { + "type": "text", + "text": "To Adapt or Not to Adapt?", + "text_level": 1, + "bbox": [ + 348, + 130, + 622, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Real-Time Adaptation for Semantic Segmentation", + "text_level": 1, + "bbox": [ + 232, + 154, + 738, + 176 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94a51354b0137b899c5ed9b46e74d97ea543d6355691a13a73f671aadecada51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 200, + 875, + 320 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b531da2a21f0adb6eccf38eac398688d2049a0530162967e18a2c900ed1113b.jpg", + "image_caption": [ + "Figure 1. Real-time adaptation with HAMLET. Online adaptation to continuous and unforeseeable domain shifts is hard and computationally expensive. HAMLET can deal with it at almost 30FPS outperforming much slower online methods - e.g. OnDA and CoTTA." + ], + "image_footnote": [], + "bbox": [ + 78, + 337, + 890, + 422 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 483, + 313, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The goal of Online Domain Adaptation for semantic segmentation is to handle unforeseeable domain changes that occur during deployment, like sudden weather events. However, the high computational costs associated with brute-force adaptation make this paradigm unfeasible for real-world applications. In this paper we propose HAMLET, a Hardware-Aware Modular Least Expensive Training framework for real-time domain adaptation. Our approach includes a hardware-aware back-propagation orchestration agent (HAMT) and a dedicated domain-shift detector that enables active control over when and how the model is adapted (LT). Thanks to these advancements, our approach is capable of performing semantic segmentation while simultaneously adapting at more than 29FPS on a single consumer-grade GPU. Our framework's encouraging accuracy and speed trade-off is demonstrated on OnDA and SHIFT benchmarks through experimental results.", + "bbox": [ + 73, + 520, + 473, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 823, + 209, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation aims at classifying an image at a pixel level, based on the local and global context, to enable a higher level of understanding of the depicted", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4bc99e80e8f50d2a7405896a09971660fe857c9fdcbe03c4b1d93eb72c00b6c4.jpg", + "image_caption": [ + "Figure 2: Online adaptation methods on the Increasing Storm. We plot mIoUs achieved on single domains. Colors from colder to warmer encode slower to faster methods." + ], + "image_footnote": [], + "bbox": [ + 555, + 493, + 872, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scene. In recent years, deep learning has become the dominant paradigm to tackle this task effectively employing CNNs [5, 69, 4] or, more recently, transformers [65], at the expense of requiring large quantities of annotated images for training. Specifically, annotating for this task needs per-pixel labeling, which is an expensive and time-consuming task, severely limiting the availability of training data.", + "bbox": [ + 496, + 722, + 892, + 827 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The use of simulations and graphics engines [42] to generate annotated frames enabled a marked decrease in the time and cost necessary to gather labeled data thanks", + "bbox": [ + 496, + 828, + 893, + 875 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Joint first authorship † Part of the work done while at Univres", + "bbox": [ + 522, + 886, + 883, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16548", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to the availability of the ground truth. However, despite the increasing quality in data realism [47], there is a substantial difference between simulated data generated by graphics engines and real-world images, such that leveraging these data for real-world applications requires adapting over a significant domain shift. The promise of unlocking this cheap and plentiful source of training data has provided a major impulse behind the development of a large body of work on Unsupervised Domain Adaptation (UDA) techniques [74, 61, 18, 15, 55], consisting of training semantic segmentation networks on labelled synthetic frames – the source domain – and then adapting the network to operate on real images, representing the target domain, without requiring human annotation. However, the synthetic-to-real shift represents only one of many possible domain transitions; specifically, when dealing with real-world deployment, domain shifts can occur from various causes, from different camera placements to different lighting, weather conditions, urban scenario, or any possible combination of the above. Because of the combinatorial nature of the problem, it is simply impossible to evenly represent all possible deployment domains in a dataset. This curse of dimensionality prevents having generalized robust performances [41, 45]. However, the recent advent of online domain adaptation [41] potentially allows us to face continuous and unpredictable domain shifts at deployment time, without requiring data associated with such domain shifts beforehand. Nonetheless, despite its potential, several severe limitations still hamper the online adaptation paradigm. In particular, continuously performing backpropagation on a frame-by-frame schedule [41] incurs a high computational cost, which negatively affects the performance of the network, dropping its overall fraterate to accommodate the need for continuous adaptation. Various factors are involved in this matter: first, the severity of this overhead is proportional to the complexity of the network itself – the larger the number of parameters, the heavier the adaptation process becomes; second, we argue that frame-by-frame optimization is an excessive process for the adaptation itself – not only the network might need much fewer optimization steps to effectively counter domain shifts, but also such an intense adaptation definitely increases the likelihood of catastrophic forgetting over previous domains [26, 45]. In summary, a practical solution for online domain adaptation in semantic segmentation that can effectively operate in real-world environments and applications still seems to be a distant goal.", + "bbox": [ + 76, + 90, + 472, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel framework aimed at overcoming these issues and thus allowing for real-time, online domain adaptation:", + "bbox": [ + 76, + 800, + 468, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We address the problem of online training by designing an automatic lightweight mechanism capable of significantly reducing back-propagation complex", + "bbox": [ + 94, + 854, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ity. We exploit the model modularity to automatically choose to train the network subset which yields the highest improvement for the allocated optimisation time. This approach reduces back-propagation FLOPS by $34\\%$ while minimizing the impact on accuracy.", + "bbox": [ + 529, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- In an orthogonal fashion to the previous contribution, we introduce a lightweight domain detector. This allows us to design principled strategies to activate training only when it really matters as well as setting hyperparameters to maximize adaptation speed. Overall, these strategies increase our speed by over $5 \\times$ while sacrificing less than $2.6\\%$ in mIoU.", + "- We evaluate our method on multiple online domain adaptation benchmarks both fully synthetic [45] and semi-synthetic CityScapes domain sequences [41], showing superior accuracy and speed compared to other test-time adaptation strategies." + ], + "bbox": [ + 517, + 176, + 890, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 1 demonstrates the superior real-time adaptation performance of HAMLET compared to slower methods such as CoTTA [57], which experience significant drops in performance when forced to maintain a similar framerate by adapting only once every 50 frames. In contrast, HAMLET achieves an impressive 29 FPS while maintaining high accuracy. Additionally, Fig. 2 offers a glimpse of HAMLET's performance on the Increasing Storm benchmark [41], further highlighting its favorable accuracy-speed trade-off.", + "bbox": [ + 498, + 378, + 890, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 527, + 640, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We review the literature relevant to our work, about semantic segmentation and UDA, with particular attention to continuous and online methodologies.", + "bbox": [ + 498, + 553, + 890, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic Segmentation. Very much like classification, deep learning plays a fundamental role in semantic segmentation. Fully Convolutional Network (FCN) [36] represents the pivotal step in this field, adapting common networks by means of learned upsample operators (deconvolutions). Several works aimed at improving FCN both in terms of speed [68, 38] and accuracy [5, 6, 7], with a large body of literature focusing on the latter. Major improvements have been achieved by enlarging the receptive field [72, 66, 5, 6, 7], introducing refinement modules [14, 73, 17], exploiting boundary cues [3, 10, 46] or using attention mechanisms in different flavors [13, 31, 58, 64]. The recent spread of Transformers in computer vision [11] reached semantic segmentation as well [64, 69, 65], with SegFormer [65] representing the state-of-the-art in the field and being the object of studies in the domain adaptation literature as well [20].", + "bbox": [ + 496, + 599, + 892, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Domain Adaptation (UDA). This body of research aims at adapting a network trained on a source, labeled domain to a target, unlabeled one. Early approaches rely on the notion of \"style\" and learn how to transfer it", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16549", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "across domains [74, 61, 18, 32, 12, 67]. Common strategies consist of learning domain-invariant features [15, 25], often using adversarial learning in the process [15, 55, 8, 19, 51]. A popular trend in UDA is Self-Training. These methods rely on self-supervision to learn from unlabelled data. In UDA, a successful strategy consists of leveraging target-curated pseudo-labels. Popular approaches for this purpose make use of confidence [77, 37, 76], try to balance the class predictions [75, 20], or use prototypes [2, 71, 70] to improve the quality of the pseudo-labels. Among many domain shifts, the synthetic-to-real one is the most studied, since the earliest works [74, 61, 18] to the latest [60, 30, 21, 28, 16, 40, 24]. However, this shift is one of a kind since it occurs only once after training, and without the requirement of avoiding forgetting the source domain.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Continuous/Test-Time UDA. This family of approaches marries UDA with continuous learning, thus dealing with the catastrophic forgetting issue ignored in the synthetic-to-real case. Most continuous UDA approaches deal with it by introducing a Replay Buffer [1, 29, 27], while additional strategies make use of style transfer [62], contrastive [44, 53] or adversarial learning [63]. Despite the definition, continuous UDA often deals with offline adaptation, with well-defined target domains over which to adapt. Conceptually similar to it, is the branch of test-time adaptation, or source-free UDA, although tackling the problem in deployment rather than offline - i.e. with no access to the data from the source domain [43]. Popular strategies to deal with it consist of generating pseudo-source data to avoid forgetting [35], freezing the final layers in the model [33], aligning features [34], batch norm retraining through entropy minimization [54] or prototypes adaptation [22].", + "bbox": [ + 75, + 320, + 472, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Online UDA. Although similar in principle to test-time adaptation, online UDA [45, 41, 52] aims to tackle multiple domain shifts, occurring unpredictably during deployment in real applications and without clear boundaries between them. On this track, the SHIFT dataset [45] provides a synthetic benchmark specifically thought for this scenario, while OASIS [52] proposes a novel protocol to evaluate UDA approaches, considering an online setting and constraining the evaluated methods to deal with frame-by-frame sequences. As for methods, OnDA [41] implements self-training as the orchestration of a static and a dynamic teacher to achieve effective online adaptation while avoiding forgetting, yet introducing massive overhead.", + "bbox": [ + 75, + 580, + 472, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Real-time performance is an essential aspect of online adaptation, particularly in applications such as autonomous driving where slow models are impractical. A slow adaptation process not only limits the practicality of real-world applications but also fails to provide high accuracy until the adaptation is complete, thereby defeating the original purpose. Therefore, accelerating the adaptation process is crucial for achieving high accuracy in real-time scenarios.", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 498, + 89, + 599, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces HAMLET, a framework for Hardware-Aware Modular Least Expensive Training. The framework aims to solve the problem of online domain adaptation with real-time performance through several synergistic strategies. First, we introduce a Hardware-Aware Modular Training (HAMT) agent able to optimize online a trade-off between model accuracy and adaptation time. HAMT allows us to significantly reduce online training time and GFLOPS. Nevertheless, the cheapest training consists of no training at all. Therefore, as the second strategy, we introduce a formal geometric model for online domain shifts that enable reliable domain shift detection and domain estimator signals (Adaptive Domain Detection, Sec. 3.3.1). These can be easily integrated to activate the adaptation process only at specific times, as least as possible. Moreover, we can further leverage these signals by designing adaptive training policies that dynamically adapt domain-sensitive hyperparameters. We refer to these as Active Training Modulations. We present an overview of HAMLET in Fig. 3.", + "bbox": [ + 496, + 114, + 893, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Setup", + "text_level": 1, + "bbox": [ + 500, + 409, + 635, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach builds on the recent progress in unsupervised domain adaptation and segmentation networks. We start with DAFormer [20], a state-of-the-art UDA method, and adopt SegFormer [65] as our segmentation backbone due to its strong generalization capacity. We use three instances of the backbone, all pre-trained on the source domain: a student, a teacher, and a static (i.e. frozen) teacher. During training, the student receives a mix of target and source images [49] and is supervised with a \"mixed-sample\" cross-entropy loss, $\\mathcal{L}_T$ (represented by green, blue and red dashed lines, in Fig. 3). This loss is computed by mixing the teacher's pseudo-labels and source annotations. To improve training stability, the teacher is updated as the exponential moving average (EMA) of the student. To further regularize the student, we use source samples stored in a replay buffer and apply two additional losses (blue lines in Fig. 3). First, we minimize the feature distance (Euclidean) between the student and the static teacher's encoder, $\\mathcal{L}_{FD}$ . Then, we employ a supervised cross-entropy task loss $\\mathcal{L}_S$ . Our complete objective is $\\mathcal{L} = \\mathcal{L}_S + \\mathcal{L}_T + \\lambda_{FD}\\mathcal{L}_{FD}$ , with $\\lambda_{FD}$ being a weight factor. During inference on the target domain, only the student is used (red lines in Fig. 3).", + "bbox": [ + 496, + 431, + 893, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hardware-Aware Modular Training (HAMT)", + "text_level": 1, + "bbox": [ + 500, + 771, + 883, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Online adaptation requires updating the parameters during deployment time. However, back-propagation is computationally expensive and hence too slow to be continuously applied on a deployed agent. Opting for a partial weight update, for example by finetuning the last module of the network, would enable much more efficient training time. However, domain shifts can manifest as changes", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16550", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/622101d847939fac3dfcfdabded3d0f986e12affbefee80ea1b3898f02dd95a3.jpg", + "image_caption": [ + "Figure 3: HAMLET framework. We employ a student-teacher model with an EMA and a static teacher. HAMT orchestrates the back-propagation over the student restricting it to a network subsection. The Active Training Modulation instead controls the adaptation process by selectively enabling it only when necessary as well as tweaking sensitive training parameters." + ], + "image_footnote": [], + "bbox": [ + 99, + 93, + 869, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in both the data input distribution (such as attributes of the images, e.g. day/night) and the output distribution (e.g. class priors). This information could be encoded in different parts of the network, therefore just updating the very last segment might not suffice. This motivates the need for orchestrating the training process, to ensure sufficient training while minimizing the computational overhead. Inspired by reward-punishment [48] and reinforcement learning [56] policies, we introduce an orchestration agent in charge of deciding how deeply the network shall be finetuned through a trade-off between the pseudo-loss minimization rate and the computational time. In contrast to previous efficient back-propagation approaches [59, 23, 9], our model is pre-trained on the task and thus requires smaller updates to adapt. Let us start by modeling the problem. Our model backbone, $f$ , is composed of four different modules: $f = m_4 \\circ m_3 \\circ m_2 \\circ m_1$ . This defines our action space $\\mathcal{A} = \\{\\mathrm{T}_1, \\mathrm{T}_2, \\mathrm{T}_3, \\mathrm{T}_4\\}$ where $\\mathrm{T}_4$ corresponds to training just the last module of the network, $m_4$ , while $\\mathrm{T}_3$ the last two modules, i.e. $m_4 \\circ m_3$ , $\\mathrm{T}_2$ the last three, i.e. $m_4 \\circ m_3 \\circ m_2$ , and $\\mathrm{T}_1$ the whole network $f$ . We also define a continuous state space $S = \\{\\mathrm{R}, \\mathrm{V}\\}$ where $\\mathrm{R}$ is the second derivative of the EMA teacher pseudo-loss, $l_t$ , over time, hence $R_t = -\\frac{\\Delta^2 l}{(\\Delta t)^2}$ , computed in discrete form as $R_t = -(l_t - 2l_{t-1} + l_{t-2})$ . V represents a cumulative vector with the same dimension as the action space $\\mathcal{A}$ , initialized at zero. Now we have everything in place to employ an expected-improvement based decision model. At each time-step $t$ , action $T_j$ is selected for $j = \\operatorname{argmax} V_t$ . During training step $t$ , $V[j]$ is updated as:", + "bbox": [ + 75, + 402, + 472, + 861 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {V} [ j ] _ {t + 1} = \\alpha R _ {t} + (1 - \\alpha) \\mathrm {V} [ j ] _ {t} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is a smoothing factor, e.g. 0.1. i.e. $\\mathrm{V_t}$ hold a discrete exponential moving average of $R_{t}$ . Therefore, our policy can be seen as a greedy module selection based on the highest expected loss improvement over its linear approximation. A notable drawback of this policy is that we will inevitably converge towards picking more rewarding, yet expensive, actions i.e. $\\mathrm{T_1,T_2}$ compared to more efficient but potentially less effective actions i.e. $\\mathrm{T}_3,\\mathrm{T}_4$ . However, our goal is not to maximize $-\\frac{\\Delta^2l}{(\\Delta t)^2}$ where $\\Delta t$ is the number of updates, our goal is instead to maximize $-\\frac{\\Delta^2l}{(\\Delta\\tau)^2}$ where $\\Delta \\tau$ is a real-time interval. Therefore, we have to introduce in the optimization policy some notion of the actual training cost of each action in $\\mathcal{A}$ on the target device. To start with, we measure the training time associated with each action, obtaining $\\omega_{T} = \\{\\omega_{T_{1}},\\omega_{T_{2}},\\omega_{T_{3}},\\omega_{T_{4}}\\}$ . With this we can compute the time-conditioning vector $\\gamma$ as", + "bbox": [ + 496, + 402, + 893, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {j} = \\frac {e ^ {\\frac {1}{\\beta \\omega_ {T _ {j}}}}}{\\sum_ {k = 1} ^ {K} e ^ {\\frac {1}{\\beta \\omega_ {T _ {k}}}}} \\quad \\text {f o r} j = 1, \\dots , K \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 648, + 890, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ is the softmax temperature, and $K$ the number of actions, i.e. 4 in our model. We modify our update policy to favor less computationally expensive modules by scaling the updates with $\\gamma$ , replacing Eq. 1 with:", + "bbox": [ + 496, + 705, + 893, + 767 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {V} [ j ] _ {t + 1} = \\left\\{ \\begin{array}{l l} \\gamma_ {j} \\alpha R _ {t} + (1 - \\alpha) \\mathrm {V} [ j ] _ {t} & \\text {i f} R _ {t} \\geq 0 \\\\ (1 - \\gamma_ {j}) \\alpha R _ {t} + (1 - \\alpha) \\mathrm {V} [ j ] _ {t} & \\text {i f} R _ {t} < 0 \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 791, + 890, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This policy makes it so that more expensive actions receive smaller rewards and larger punishments. Despite its simplicity, this leads to a significant reduction in FLOPS for an average back-propagation $\\beta$ , i.e. $-30\\%$ with $\\beta = 2.75$", + "bbox": [ + 496, + 840, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16551", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "or $-43\\%$ with $\\beta = 1$ . We finally choose $\\beta = 1.75$ to obtain a FLOPS reduction of $-34\\%$ . Exhaustive ablations on HAMT are presented in the supplementary material.", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Active Training Modulation", + "text_level": 1, + "bbox": [ + 76, + 152, + 328, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Continuous and test-time adaptation methods tackle online learning as a continuous and constant process carried out on the data stream. Nevertheless, this approach presents several shortcomings when it comes to real-world deployments. Performing adaptation when the deployment domain is unchanged does not lead to further performance improvements on the current domain; instead, it might cause significant forgetting on previous domains, hence hindering model generalization (we present evidence of this in the supplementary material). Even if mitigated by HAMT, online training remains a computationally expensive procedure, also due to several teachers' necessary forward passes. However, knowing when and what kind of adaptation is needed is not a trivial task. We tackle this by introducing an Adaptive Domain Detection mechanism, in Sec. 3.3.1, and then a set of strategies to reduce the training time while optimizing the learning rate accordingly, in Sec. 3.3.2.", + "bbox": [ + 75, + 178, + 472, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Adaptive Domain Detection", + "text_level": 1, + "bbox": [ + 76, + 465, + 326, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A key element of an online adaptation system consists of acquiring awareness of the trajectory in the data distribution space, i.e. domains, traveled by the student model during deployment. We can model the problem by setting the trajectory origin in the source domain. With high dimensional data, the data distribution is not tractable, therefore the trajectory cannot be described in closed form. Recent work [41] introduced the notion of distance between the current deployed domain and source by approximating it with the confidence drop of a source pre-trained model. This approach heavily relies on the assumption that the pretrained model is well-calibrated. While this might hold for domains close to source, the calibration quickly degrades in farther domains [45, 41]. This myopic behavior dampen the simple use of confidence for domain detection. Furthermore, the additional forward pass increases the computational cost during deployment. We tackle these limitations with an equivalently simple, yet more robust, approach. We modify the backbone of the static teacher $f^{\\mathrm{st}}$ used for the feature distance loss $\\mathcal{L}_{FD}$ by connecting a lightweight segmentation head, $d_1^{\\mathrm{st}}$ , after the first encoder module $m_1^{\\mathrm{st}}$ : $h_1^{\\mathrm{st}} = d_1^{\\mathrm{st}} \\circ m_1^{\\mathrm{st}}$ . This additional decoder, $h_1^{\\mathrm{st}}$ , is trained offline, on source data, without propagating gradients in the backbone ( $m_1^{\\mathrm{st}}$ is frozen). Given a target sample $x_T$ , we propose to compute the cross-entropy between the one-hot encoded student prediction $p(x_T) = 1_{\\arg \\max(f(x_T))}$ and the lightweight decoder prediction $g(x_T) = h_1^{\\mathrm{st}}(x_T)$ as", + "bbox": [ + 75, + 493, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {T} ^ {(i)} = - \\sum_ {p = 1} ^ {H \\times W} \\sum_ {c = 1} ^ {C} p \\left(x _ {T} ^ {(i)}\\right) \\log g \\left(x _ {T} ^ {(i)}\\right) \\Big | _ {p, c} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 112, + 893, + 156 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thanks to the student model's higher generalization capability (both due to a larger number of parameters and the unsupervised adaptation process), it will always outperform the lightweight decoder head. Nevertheless, since now the distance is measured in the prediction space, we are not subjected to model miscalibration. Furthermore, since the student model is in constant adaptation, the domain distance accuracy actually improves over time, leading to better results. We present evidence of these claims in the supplementary material. We now define a denoised signal by using bin-averaging $A_T^{(i)} = \\sum_{j=m}^{m(i+1)-1} \\frac{H_T^{(j)}}{m}$ where $m$ is the bin size. Domains are modeled as discrete steps of $A_T^{(i)}$", + "bbox": [ + 496, + 165, + 893, + 358 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nB _ {0} = A _ {0} \\quad B _ {i} = \\left\\{ \\begin{array}{l l} A _ {i} & \\text {i f} | B _ {i - 1} - A _ {i} | > z \\\\ B _ {i - 1} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 375, + 892, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B$ is the discretized signal and $z$ is the minimum distance used to identify new domains. Finally, we refer to the signed amplitude of domain shifts as $\\Delta B_{i} = B_{i} - B_{i - 1}$ and a domain change is detected whenever $|\\Delta B_i| > z$ .", + "bbox": [ + 496, + 419, + 893, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Least Training and Adaptive Learning Rate", + "text_level": 1, + "bbox": [ + 498, + 498, + 859, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The definitions of $B$ allow us to customize the training process. To this end, we adopt a Least Training (LT) strategy and trigger adaptation only when facing a new domain, which occurs when $|\\Delta B_i| > z$ . Effective online learning performance depends heavily on the choice of hyperparameters such as the learning rate $\\eta$ and learning rate decay rate. Therefore, we can adjust these parameters to facilitate adaptation according to the nature and intensity of domain shifts we encounter, we refer to this orchestration as Adaptive Learning Rate (ALR). For example, the larger the domain shift (i.e. $|\\Delta B_i|$ ), the more we need to adapt to counteract its effect. This can be achieved by either running more optimization steps or using a higher learning rate. Whenever a domain shift is detected, we compute the number of adaptation iterations $L = K_l\\frac{|\\Delta B_i|}{z}$ , hence proportionally to the amplitude of the shift $|\\Delta B_i|$ relative to the threshold $z$ . $K_l$ is a multiplicative factor representing the minimum adaptation iterations. If a new domain shift takes place before the adaptation process completes, we accumulate the required optimization steps. Then, we can play on two further parameters: $K_l$ and the learning rate schedule. We argue that proper scheduling is crucial for attaining a smoother adaptation. The learning rate, $\\eta$ , is linearly decayed until the adaptation is concluded – the smaller the domain shift, the faster the decay. While the initial learning rate, $K_{\\eta}$ , should", + "bbox": [ + 496, + 521, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16552", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/44f7b326ae0dbd48758eddb9e71d95fc184ee4ecd72b7456baac2e6512710f22.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
HAMTLTALRDCMRCS200mm (mIoU)All-domains (mIoU)FPSAverage GFLOPSAdaptation GFLOPS
TotalFwd.Bwd.Fwd.Bwd.
(A)-----62.2 ± 0.969.5 ± 0.35.9 ± 0.0125.2 ± 0.094.4 ± 0.030.8 ± 0.056.6 ± 0.030.8 ± 0.0
(B)✓----60.2 ± 0.568.7 ± 0.37.0 ± 0.1114.7 ± 0.094.4 ± 0.020.3 ± 0.056.6 ± 0.020.3 ± 0.0
(C)✓✓---51.8 ± 0.565.7 ± 0.229.5 ± 0.644.4 ± 0.542.6 ± 0.41.8 ± 0.256.6 ± 0.020.2 ± 0.2
(D)✓✓✓--54.1 ± 1.265.9 ± 0.229.5 ± 0.544.4 ± 0.342.7 ± 0.21.8 ± 0.156.6 ± 0.020.3 ± 0.1
(E)✓✓✓✓-56.6 ± 0.866.3 ± 0.128.9 ± 0.344.7 ± 0.242.9 ± 0.21.8 ± 0.156.6 ± 0.020.2 ± 0.0
(F)✓✓✓-✓55.8 ± 1.066.3 ± 0.229.1 ± 1.145.2 ± 0.143.2 ± 0.12.0 ± 0.056.6 ± 0.020.3 ± 0.0
(G)✓✓✓✓✓58.2 ± 0.866.9 ± 0.329.7 ± 0.645.7 ± 0.343.6 ± 0.22.1 ± 0.156.6 ± 0.020.2 ± 0.1
(a)
clear 1200mmclear 2100mmclear 375mmclear 4clear h-meantarget h-meantotal h-meanFPSGFLOPS
(A)72.952.273.664.273.067.673.473.260.667.25.6125.2
(B)73.050.473.462.173.067.373.273.159.166.46.8114.7
(C)73.446.073.561.573.666.173.873.656.565.17.2100.0
(G)73.453.673.165.273.568.273.273.361.667.89.182.2
(b)
", + "bbox": [ + 98, + 87, + 898, + 228 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Ablation studies - HAMLET components. Top: Increasing Storm (8925 frames per domain) [41], bottom: Fast Storm C [41] (2975 frames per domain). For each configuration, we report mIoU, framerate, and GFLOPS.", + "bbox": [ + 75, + 232, + 890, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "be higher when the domain shift is triggered in domains farther from the source", + "bbox": [ + 75, + 277, + 467, + 306 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nK _ {\\eta} = K _ {\\eta , \\min } + \\frac {\\left(B _ {\\mathrm {i}} - B _ {\\text {s o u r c e}}\\right) \\left(K _ {\\eta , \\max } - K _ {\\eta , \\min }\\right)}{B _ {\\text {h a r d}} - B _ {\\text {s o u r c e}}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 329, + 468, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $B_{\\mathrm{source}}$ (resp. $B_{\\mathrm{hard}}$ ) is an estimate of $B$ when the network is close to (resp. far from) the source domain; and $K_{\\eta, \\min}$ (resp. $K_{\\eta, \\max}$ ) is the value of $K_{\\eta}$ assigned when the network is close to (resp. far away from) the source. Concerning $K_{l}$ , we posit that moving towards the source requires less adaptation than going towards harder domains: the model shows good recalling of previously explored domains and thanks to the employed regularization strategies", + "bbox": [ + 76, + 367, + 468, + 488 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nK _ {l} = \\left\\{ \\begin{array}{l l} K _ {l, \\max } & \\text {i f} \\Delta B _ {i} \\geq 0 \\\\ K _ {1, \\min } + \\frac {\\left(B _ {\\mathrm {i}} - B _ {\\text {s o u r c e}}\\right) \\left(K _ {1 , \\max } - K _ {1 , \\min }\\right)}{B _ {\\text {h a r d}} - B _ {\\text {s o u r c e}}} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 511, + 468, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $K_{l,\\min}$ (resp. $K_{l,\\max}$ ) is the value of $K_{l}$ assigned when the model is close to (resp. far away from) the source domain. Extensive ablations in the supplementary material will highlight how the orchestration of the adaptation hyperparameters improves the accuracy-speed trade-off.", + "bbox": [ + 76, + 556, + 468, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.3 Dynamic ClassMix (DCM)", + "text_level": 1, + "bbox": [ + 76, + 654, + 316, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ClassMix [39] provides a simple mechanism for data augmentation by mixing classes from the source dataset into target images. Usually $50\\%$ of the classes in the source dataset are selected, however we notice that this percentage is a highly sensitive hyperparameter in online domain adaptation. Injecting a significant portion of source classes has a beneficial impact when adapting to domains closer to the source domain, whereas when adapting to domains further from the source the opposite effect can be observed, as it effectively slows down the adaptation process. We therefore exploit once more the deployment domain awareness to control the mixing augmentation:", + "bbox": [ + 75, + 679, + 468, + 859 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nK _ {\\mathrm {C M}} = K _ {\\mathrm {C M}, \\min } + \\frac {\\left(B _ {\\mathrm {i}} - B _ {\\text {s o u r c e}}\\right) \\left(K _ {\\mathrm {C M} , \\max } - K _ {\\mathrm {C M} , \\min }\\right)}{B _ {\\text {h a r d}} - B _ {\\text {s o u r c e}}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 876, + 468, + 905 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $K_{\\mathrm{CM}}$ is the percentage of source classes used during adaptation; and $K_{\\mathrm{CM,min}}$ (resp. $K_{\\mathrm{CM,max}}$ ) is the value of $K_{\\mathrm{CM}}$ assigned when the network is close to (resp. far away from) the source domain.", + "bbox": [ + 500, + 277, + 890, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.4 Buffer Sampling", + "text_level": 1, + "bbox": [ + 500, + 354, + 669, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [41], to simulate real deployment, we limit our access to the source domain by using a replay buffer. Additionally, instead of initializing at random (with a uniform prior), we apply Rare Class Sampling (RCS) (skewed priors) as in [20]. This incentives a more balanced class distribution over the buffer, ultimately leading to better accuracy.", + "bbox": [ + 496, + 378, + 890, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 481, + 704, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experiments are carried out on (a) the OnDA benchmarks [41] and (b) the SHIFT dataset [45]. (a) is a semi-syntheptic benchmark, as it applies synthetic rain and fog [50] over 4 different intensities profiles. The main benchmark, Increasing Storm, presents a storm with a pyramidal intensity profile; see Fig. 4. In contrast, (b) is a purely synthetic dataset, where both the underlying image and the weather are synthetically generated and thus domain change is fully controllable. All models are evaluated using mIoU: following [41], we report the harmonic mean over domains to present the overall adaptation performance. All experiments were carried out using an Nvidia™ RTX 3090 GPU. We refer to supplementary material for further details.", + "bbox": [ + 496, + 507, + 890, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 710, + 663, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1 we study the impact of each contribution to adaptation performance, both in terms of accuracy and efficiency. For each configuration, we report mIoU over different portions of the sequence, the framerate and the amount of GFLOPS - respectively averages of: total, forward and backward passes, and dedicated adaptation only, also divided in forward (Fwd) and backward (Bwd). Tab. 1 (a) shows results on the Increasing Storm scenario [41]. Here, we show mIoU over the $200\\mathrm{mm}$ domain, i.e. the hardest in the sequence, as well as the mIoU averaged over forward and backward adaptation, i.e., from clear to $200\\mathrm{mm}$ rain", + "bbox": [ + 496, + 734, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16553", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6c6f9b71ed359d9f1d07076d2643823f75623ee47b1119997bca1ceaf559c2b5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
clear25mm50mm75mm100mm200mmh-meanFPSGFLOPS
FBFBFBFBFBFBFBT
(A)DeepLabV2 (no adaptation)64.5-57.1-48.7-41.5-34.4-18.537.3--39.4-
(B)DeepLabV2 fully supervised (oracle)64.5-64.1-63.7-63.0-62.4-58.262.6--39.4-
(C)OnDA64.564.860.457.157.354.554.852.252.049.142.254.255.1-1.3-
(D)SegFormer MiT-B1 (no adaptation)73.4-68.8-64.2-58.0-51.8-31.257.8--48.434.9
(E)SegFormer MiT-B5 (no adaptation)77.6-73.9-71.0-67.2-62.6-46.764.7--11.5240.4
(F)SegFormer MiT-B1 fully supervised (oracle)72.9-72.4-72.1-71.5-70.7-68.671.3--48.434.9
(G)TENT73.072.868.568.664.564.859.760.254.554.835.956.263.659.910.0-
(H)TENT + Replay Buffer73.072.868.568.664.564.859.760.254.454.735.856.163.659.97.8-
(I)CoTTA72.574.469.570.965.968.266.164.764.663.557.265.668.166.80.6593.8
(J)CoTTA real-time73.375.470.370.666.966.462.561.457.656.939.759.265.562.327.041.7
(K)HAMLET (ours)73.471.070.168.867.767.566.666.465.564.659.266.867.667.229.145.7
", + "bbox": [ + 96, + 88, + 870, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Comparison against other models – Increasing storm scenario. (A-C) methods built over DeepLabv2, (D-E) SegFormer variants trained on source, (F) oracle, (G-K) models adapted online. We report mIoU, framerate, and GFLOPS.", + "bbox": [ + 75, + 246, + 890, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and backward. Results are averaged over 3 runs with different seeds, with standard deviation being reported. (A) reports the results achieved by naively performing full adaptation of the model. HAMT can increase the framerate by roughly $15\\%$ by reducing the Bwd GFLOPS of $34\\%$ , at the expense of as few as $0.7\\mathrm{mIoU}$ on average, i.e., about 2 points on the $200\\mathrm{mm}$ domain. The main boost in terms of speed is obviously given by LT (C), which inhibits the training in absence of detected domain shifts. LT increases the framerate by approximately $4\\times$ by decimating the total GFLOPS, yet not affecting the adaptation Bwd GFLOPS. This comes with a price in terms of mIoU, dropping by about 4 points on average and more than 10 points on $200\\mathrm{mm}$ - not a moderate drop anymore. LT impact highly depends on the domain sequence experienced during deployment: frequent domain changes could prevent training inhibition, thus neglecting LT gains in terms of efficiency, as we will appreciate later. The loss in accuracy is progressively regained by adding ALR (D), with further improvements yielded by one between DCM (E) and RCS (F), or both together (G) leading to the full HAMLET configuration. The three together allow for reducing the gap to 2.5 points mIoU - 4 over the $200\\mathrm{mm}$ domain - without sacrificing any efficiency. Tab. 1 (b) shows further results, on a faster version of Storm C [41]. This represents a much more challenging scenario, with harsher and $3\\times$ more frequent domain shifts. Here we show the single domains mIoU, as well as harmonic mean on source and target domains, and all frames. As expected, in this benchmark, LT alone (C) results much less effective than before, with a much lower gain in FPS and GFLOPS. Here, the synergy between the HAMT, LT, and the other components (G) allows for the best accuracy and speedup - even outperforming the full training variant (A) - highlighting their complementarity. Further ablations are in the supplementary material.", + "bbox": [ + 75, + 303, + 472, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results on Increasing Storm", + "text_level": 1, + "bbox": [ + 76, + 845, + 331, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 shows a direct comparison between HAMLET and relevant approaches. The presented test-time adaptation", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f4433bc49d15129f594d5c8ec125e604e8568ee8e10b04e24fd6d40576a4f56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 310, + 849, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d24ecd2323e20c9b68c67191586e436a765e6a0dbea0a39967845645c400b4bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 349, + 848, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79e0895ecd0d4a25ec03c1c4836f78786d7d64dc4e4bb9e31e46c9d83370760d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 552, + 420, + 846, + 496 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fda9ff3d1862761f0f33b02a91c509e2a6e7eae557fe386105aad3905486660b.jpg", + "image_caption": [ + "Figure 4: HAMLET on the Increasing Storm. We show rain intensity (in millimetres), mIoU over active (bold) and inactive (dashed) domains, learning rate and FPS." + ], + "image_footnote": [], + "bbox": [ + 547, + 500, + 846, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "strategies namely - TENT and CoTTA - were revised to handle the online setting and be fairly compared with HAMLET. All methods start with the same exact initial weights - with HAMLET requiring the additional lightweight decoder, not needed by TENT and CoTTA - using SegFormer MiT-B1 as the backbone, since it is $4 \\times$ faster than SegFormer MiT-B5 and thus better suited to keep real-time performance even during adaptation. We report results achieved by DeepLabv2 trained on source data only (A), an oracle model trained with full supervision (B), as well as OnDA [41] (C) as a reference. Then, we report SegFormer models trained on the source domain only (D) and (E). In (F) we show the performance achieved by an oracle SegFormer, trained on all domains fully supervised. Following [41], columns \"F\" concern forward adaptation from clear to $200\\mathrm{mm}$ , while columns \"B\" show backward adap", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16554", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c3e17f1a06f9bbc9db60b7e986743fd76d51f1b6498ce31fe1661a731e2e38f0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
clear750m375m150m75mh-meanFPSGFLOPS
FBFBFBFBFBFBT
OnDA64.965.863.362.360.758.851.649.142.155.154.1-1.3-
SegFormer MiT-B1 (no adaptation)71.1-70.0-67.5-58.8-46.961.3--48.434.9
Full training71.572.172.974.771.973.167.668.161.368.771.970.35.6125.2
HAMLET (ours)71.171.670.370.868.869.264.364.357.065.968.967.424.850.7
", + "bbox": [ + 109, + 88, + 859, + 148 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cde8bd9b6f9bd3c966bb0af02c355124e705c5f8361ff5cd23d622250eda5156.jpg", + "table_caption": [ + "Table 3: Results on foggy domains. Comparison between OnDA, Source SegFormer, full training adaptation, and HAMLET." + ], + "table_footnote": [], + "table_body": "
ClearCloudyOvercastSmall rainMid rainHeavy rainh-meanFPSGFLOPS
FBFBFBFBFBFBFBT
SegFormer MiT-B1 fully supervised (oracle)80.1-79.9-79.8-78.9-78.7-77.179.1--48.434.93
SegFormer MiT-B1 (no adaptation)79.6-77.1-75.4-73.4-71.4-66.773.7--48.434.93
Full training78.979.376.776.876.877.974.874.876.376.574.076.277.076.65.0125.1
HAMLET (ours)79.678.976.976.676.177.473.374.374.276.074.275.776.676.126.843.9
", + "bbox": [ + 84, + 167, + 885, + 228 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: Results on SHIFT dataset [45]. Comparison between Source SegFormer, full training adaptation, and HAMLET.", + "bbox": [ + 80, + 229, + 887, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tation from $200\\mathrm{mm}$ to clear, while the h-mean T refers to the overall harmonic mean. We can notice how SegFormer results are much more robust to domain changes with respect to DeepLabv2. Indeed, SegFormer MiT-B5 (E), without any adaptation, results more accurate than DeepLabv2 oracle (B), as well as better and faster than OnDA (C). The faster variant (D) outperforms OnDA both in speed and accuracy, reaching 48 FPS. Nevertheless, domain changes still dampen the full potential of SegFormer. Indeed, the oracle (F) outperforms (D) by about $+14$ mIoU. However, this is not meaningful for real deployment experiencing unpredictable domain shifts, as it assumes to have data available in advance. Concerning test-time models, TENT starts adapting properly only beyond $50\\mathrm{mm}$ , both with (G) and without (H) frame buffer, while it loses some accuracy on $25\\mathrm{mm}$ . This makes its overall forward adaptation performance slightly worse compared to the pre-trained model (D), while being better at backward adaptation. Despite outperforming SegFormer MiT-B1, TENT is both slower and less accurate than SegFormer MiT-B5 running without any adaptation, further suggesting the robustness of the latter and making TENT not suitable for real-world deployment. On the contrary, CoTTA (I) outperforms both SegFormer models trained on source only, at the expense of dropping the framerate below 1FPS. It is worth mentioning that these metrics were collected after each domain was completed by each model individually. In an evaluation setup imposing a shared time frame, slower models would present much lower metrics, since their adaptation process would result constantly lagged. In fact, forcing CoTTA to run in real-time, at nearly 30FPS – i.e. by training once every 50 frames – dramatically reduces the effectiveness of the adaptation process (J), with drastic drops in the hardest domains. Finally, HAMLET (K) succeeds on any fronts, improving the baseline (D) by about 10 points with only a cost of $25\\%$ in terms of speed, while outperforming SegFormer MiT-B5 (E) both on accuracy $(+2.5$ mIoU) and speed $(3\\times$ faster) – being the only method achieving this, and thus the only suitable choice for real-time applications. Fig. 4 shows the overall behavior of HAMLET while adapting over the Increasing Storm. In addition to the rain intensity and the", + "bbox": [ + 76, + 271, + 472, + 895 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b8e0c1eeba7fae64b64798ae555b3b1cccde62d7dee9dc25e223452ccce2c5d.jpg", + "image_caption": [ + "Figure 5: HAMLET on the SHIFT benchmark. We show mIoU over active (bold) and inactive (dashed) domains, learning rate and FPS." + ], + "image_footnote": [], + "bbox": [ + 540, + 277, + 851, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mIoU achieved on each domain - active (bold) or inactive (dashed), i.e. respectively the mIoU on the domain being currently faced during deployment, and how the current adaptation affects the performance on the other domains to highlight the robustness to forgetting - we also report how the learning rate is modulated in correspondence of detected domain shifts, with a consequent drop in FPS due to the short training process taking place. For further experiments on harsher and sudden adaptation cycles, we include results of Storms A, B, C [41] in the supplementary material.", + "bbox": [ + 496, + 623, + 890, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Additional Results: Fog and SHIFT", + "text_level": 1, + "bbox": [ + 500, + 786, + 810, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fog. In Tab. 3, we investigate adaptation on the Increasing Fog scenario in the OnDA benchmark [41]. Crucially, for this experiment, we keep the same hyperparameters used for the Increasing Storm, since in both cases the starting SegFormer model is trained on the same source domain. This allows for validating how the proposed setting general-", + "bbox": [ + 496, + 810, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16555", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/754b478f08cbb8f9ef25a77d64555c04cb1d29a05f6758abea85409722214ff3.jpg", + "image_caption": [ + "Figure 6: Qualitative results - HAMLET in action. From left to right, we show frames from clean, $50\\mathrm{mm}$ , $100\\mathrm{mm}$ , and $200\\mathrm{m}$ domains. From top to bottom: input image, prediction by SegFormer trained on source domain and HAMLET." + ], + "image_footnote": [], + "bbox": [ + 137, + 88, + 834, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "izes at dealing with different kind of domain shifts, beyond those considered in the main experiments. We effectively use Increasing Fog as test set, and compare against Seg-Former trained on source (no adaptation) and a model that has been adapted by means of full online training optimization (configuration (A) of Table 1). HAMLET is able to adapt almost as well as the full online training model, with less than a $3\\mathrm{mIoU}$ gap, while enjoying real-time adaptation at nearly $5\\times$ the speed using just $40\\%$ of the FLOPS.", + "bbox": [ + 75, + 373, + 472, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "SHIFT. We further test HAMLET on the SHIFT dataset [45]. Tab. 4 collects the results achieved by Seg-Former trained on source, full online training and HAMLET respectively, both at forward and backward adaptation across Clear, Cloudy, Overcast, Small rain, Mid rain and Heavy rain domains. Here HAMLET results highly competitive with the full training regime, with only 0.5 drop in average mIoU, while being more than $5 \\times$ faster. Fig. 5 depicts, from top to bottom, the rain intensity characterizing any domain encountered on SHIFT, the mIoU achieved both on current (bold) and inactive (dashed) domains, the learning rate changes based on the domain shift detection, and the framerate achieved at any step. We refer to the supplementary material for a deeper analysis.", + "bbox": [ + 75, + 508, + 468, + 720 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Qualitative results. To conclude, Fig. 6 shows some qualitative examples from CityScapes. We can notice how SegFormer accuracy (second tow) drops with severe rain, whereas HAMLET (third row) is capable of keeping the same segmentation quality across the storm.", + "bbox": [ + 75, + 722, + 468, + 799 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 813, + 191, + 829 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Orthogonality. HAMT and LT act independently. Indeed, by strongly constraining the adaptation periods through LT, HAMT has a limited margin of action. The impact of HAMT also depends on the backbone and by care", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "fully crafting modular architectures, one can achieve further optimization. Nevertheless, in a deployment environment where domain shifts occur at high frequencies (e.g., Storm C), LT is ineffective, while HAMT thrives.", + "bbox": [ + 496, + 373, + 890, + 434 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Measuring forgetting. An interesting topic we have not investigated consists of introducing an explicit awareness of which domains have been explored and how well we can recall them, expanding the distance $B$ to multiple dimensions.", + "bbox": [ + 496, + 435, + 890, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Safety. We believe dynamic adaptation has the potential to enhance safety, but we acknowledge the necessity for rigorous testing and verification to safeguard against drift or catastrophic forgetting. This mandates a comprehensive effort from academia, industry, and certification authorities for ensuring the integrity of dynamically adapting models.", + "bbox": [ + 496, + 496, + 890, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6. Summary & Conclusion", + "text_level": 1, + "bbox": [ + 498, + 604, + 725, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have presented HAMLET, a framework for real-time adaptation for semantic segmentation that achieves state-of-the-art performance on established benchmarks with continuous domain changes. Our approach combines a hardware-aware backpropagation orchestrator and a specialized domain-shift detector to enable active control over the model's adaptation, resulting in high framereates on a consumer-grade GPU. These advancements enable HAMLET to be a promising solution for in-the-wild deployment, making it a valuable tool for applications that require robust performance in the face of unforeseen domain changes.", + "bbox": [ + 496, + 630, + 890, + 797 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. The authors thank Gianluca Villani for the insightful discussion on reward-punishment policies, Leonardo Ravaglia for his expertise on hardware-aware training, and Lorenzo Andraghetti for exceptional technical support throughout the project. Their assistance was invaluable in the completion of this work.", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16556", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andreea Bobu, Judy Hoffman, Eric Tzeng, and Trevor Darrell. Adapting to continuously shifting domains. In ICLR 2018 Workshop Program Chairs, 2018. 00000.", + "[2] Chaoqi Chen, Weiping Xie, Wenbing Huang, Yu Rong, Xinghao Ding, Yue Huang, Tingyang Xu, and Junzhou Huang. Progressive feature alignment for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 627-636, 2019.", + "[3] Liang-Chieh Chen, Jonathan T Barron, George Papandreou, Kevin Murphy, and Alan L Yuille. Semantic image segmentation with task-specific edge detection using cnns and a discriminatively trained domain transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4545-4554, 2016.", + "[4] Liang-Chieh Chen, Raphael Gontijo Lopes, Bowen Cheng, Maxwell D Collins, Ekin D Cubuk, Barret Zoph, Hartwig Adam, and Jonathon Shlens. Naive-student: Leveraging semi-supervised learning in video sequences for urban scene segmentation. In European Conference on Computer Vision, pages 695-714. Springer, 2020.", + "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017.", + "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(4):834-848, Apr 2018.", + "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018.", + "[8] Yi-Hsin Chen, Wei-Yu Chen, Yu-Ting Chen, Bo-Cheng Tsai, Yu-Chiang Frank Wang, and Min Sun. No more discrimination: Cross city adaptation of road scene segmenters. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2011-2020. IEEE, 2017. 00000.", + "[9] Feng Cheng, Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Li, and Wei Xia. Stochastic backpropagation: A memory efficient strategy for training video models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8301-8310, 2022.", + "[10] Henghui Ding, Xudong Jiang, Ai Qun Liu, Nadia Magnenat Thalmann, and Gang Wang. Boundary-aware feature propagation for scene segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6819-6829, 2019.", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[12] A. Dundar, M. Y. Liu, Z. Yu, T. C. Wang, J. Zedlewski, and J. Kautz. Domain stylization: A fast covariance matching framework towards domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2020.", + "[13] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3146-3154, 2019.", + "[14] Jun Fu, Jing Liu, Yuhang Wang, Yong Li, Yongjun Bao, Jinhui Tang, and Hanqing Lu. Adaptive context network for scene parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6748-6757, 2019.", + "[15] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, Jan. 2016.", + "[16] Rui Gong, Martin Danelljan, Dengxin Dai, Danda Pani Paudel, Ajad Chhatkuli, Fisher Yu, and Luc Van Gool. Tacs: Taxonomy adaptive cross-domain semantic segmentation. In European Conference on Computer Vision, pages 19-35. Springer, 2022.", + "[17] Junjun He, Zhongying Deng, Lei Zhou, Yali Wang, and Yu Qiao. Adaptive pyramid context network for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7519-7528, 2019.", + "[18] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. CyCADA: Cycle-consistent adversarial domain adaptation. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 1989-1998, Stockholm, Sweden, 10-15 Jul 2018. PMLR.", + "[19] Judy Hoffman, Dequan Wang, Fisher Yu, and Trevor Darryll. FCNs in the wild: Pixel-level adversarial and constraint-based adaptation. CoRR, 2016. 00000.", + "[20] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. arXiv preprint arXiv:2111.14887, 2021.", + "[21] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. *Hrda: Context-aware high-resolution domain-adaptive semantic segmentation*. In European Conference on Computer Vision (ECCV), 2022.", + "[22] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021.", + "[23] Angela H Jiang, Daniel L-K Wong, Giulio Zhou, David G Andersen, Jeffrey Dean, Gregory R Ganger, Gauri Joshi," + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16557", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Michael Kaminsky, Michael Kozuch, Zachary C Lipton, et al. Accelerating deep learning by focusing on the biggest losers. arXiv preprint arXiv:1910.00762, 2019.", + "[24] Zhengkai Jiang, Yuxi Li, Ceyuan Yang, Peng Gao, Yabiao Wang, Ying Tai, and Chengjie Wang. Prototypical contrast adaptation for domain adaptive semantic segmentation. In European Conference on Computer Vision, pages 36-54. Springer, 2022.", + "[25] Myeongjin Kim and Hyeran Byun. Learning texture invariant representation for domain adaptation of semantic segmentation. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2020.", + "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114, 12 2016.", + "[27] Yevhen Kuznietsov, Marc Proesmans, and Luc Van Gool. Towards unsupervised online domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 261-271, 2022.", + "[28] Xin Lai, Zhuotao Tian, Xiaogang Xu, Yingcong Chen, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Decouplenet: Decoupled network for domain adaptive semantic segmentation. In European Conference on Computer Vision, pages 369-387. Springer, 2022.", + "[29] Qicheng Lao, Xiang Jiang, Mohammad Havaei, and Yoshua Bengio. Continuous domain adaptation with variational domain-agnostic feature replay. arXiv preprint arXiv:2003.04382, 2020.", + "[30] Geon Lee, Chanho Eom, Wonkyung Lee, Hyekang Park, and Bumsub Ham. Bi-directional contrastive learning for domain adaptive semantic segmentation. In European Conference on Computer Vision, pages 38-55. Springer, 2022.", + "[31] Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention networks for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9167-9176, 2019.", + "[32] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2019.", + "[33] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. CoRR, 2020.", + "[34] Yuejiang Liu, Parth Kothari, Bastien Germain van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++: When does self-supervised test-time training fail or thrive? In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021.", + "[35] Yang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. 2021." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Jonathan Long, Evan Shelhamer, and Trevor Darrel. Fully convolutional networks for semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.", + "[37] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. Lecture Notes in Computer Science, page 415-430, 2020.", + "[38] Vladimir Nekrasov, Chunhua Shen, and Ian Reid. Lightweight refinenet for real-time semantic segmentation. In British Conference on Computer Vision (BMVC), 2018.", + "[39] Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1369-1378, 2021.", + "[40] Fei Pan, Sungsu Hur, Seokju Lee, Junsik Kim, and In So Kweon. Ml-bpm: Multi-teacher learning with bidirectional photometric mixing for open compound domain adaptation in semantic segmentation. In European Conference on Computer Vision, pages 236–251. Springer, 2022.", + "[41] Theodoros Panagiotakopoulos, Pier Luigi Dovesi, Linus Harenstam-Nielsen, and Matteo Poggi. Online domain adaptation for semantic segmentation in ever-changing conditions. In European Conference on Computer Vision (ECCV), 2022.", + "[42] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In European conference on computer vision, pages 102-118. Springer, 2016.", + "[43] Serban Stan and Mohammad Rostami. Unsupervised model adaptation for continual semantic segmentation. In AAAI, 2021.", + "[44] Peng Su, Shixiang Tang, Peng Gao, Di Qiu, Ni Zhao, and Xiaogang Wang. Gradient regularized contrastive learning for continual domain adaptation. 2020. 00000.", + "[45] Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. SHIFT: a synthetic driving dataset for continuous multi-task domain adaptation. In Computer Vision and Pattern Recognition, 2022.", + "[46] Towaki Takikawa, David Acuna, Varun Jampani, and Sanja Fidler. Gated-scnn: Gated shape cnns for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5229-5238, 2019.", + "[47] Phillip Thomas, Lars Pandikow, Alex Kim, Michael Stanley, and James Grieve. Open synthetic dataset for improving cyclist detection, Nov 2021.", + "[48] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 195-204, 2019.", + "[49] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1379-1389, January 2021." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16558", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Maxime Tremblay, Shirsendu S. Halder, Raoul de Charette, and Jean-François Lalonde. Rain rendering for evaluating and improving robustness to bad weather. International Journal of Computer Vision, 2020.", + "[51] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Jun 2018.", + "[52] Riccardo Volpi, Pau de Jorge, Diane Larlus, and Gabriela Csurka. On the road to online adaptation for semantic image segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022.", + "[53] Vibashan VS, Poojan Oza, and Vishal M. Patel. Towards online domain adaptive object detection. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Jan 2023.", + "[54] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021.", + "[55] Haoran Wang, Tong Shen, Wei Zhang, Lingyu Duan, and Tao Mei. Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In The European Conference on Computer Vision (ECCV), August 2020.", + "[56] Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. Haq: Hardware-aware automated quantization with mixed precision, 2018.", + "[57] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation, 2022.", + "[58] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018.", + "[59] Bingzhen Wei, Xu Sun, Xuancheng Ren, and Jingjing Xu. Minimal effort back propagation for convolutional neural networks. arXiv preprint arXiv:1709.05804, 2017.", + "[60] Tsung-Han Wu, Yi-Syuan Liou, Shao-Ji Yuan, Hsin-Ying Lee, Tung-I Chen, Kuan-Chih Huang, and Winston H Hsu. D2ada: Dynamic density-aware active domain adaptation for semantic segmentation. In European Conference on Computer Vision (ECCV), 2022.", + "[61] Zuxuan Wu, Xintong Han, Yen-Liang Lin, Mustafa Gokhan Uzunbas, Tom Goldstein, Ser Nam Lim, and Larry S. Davis. Dcan: Dual channel-wise alignment networks for unsupervised scene adaptation. Lecture Notes in Computer Science, page 535-552, 2018.", + "[62] Zuxuan Wu, Xin Wang, Joseph Gonzalez, Tom Goldstein, and Larry Davis. ACE: Adapting to changing environments for semantic segmentation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2121-2130. IEEE, 2019.", + "[63] Markus Wulfmeier, Alex Bewley, and Ingmar Posner. Incremental adversarial domain adaptation for continually changing environments. 2018. 00000.", + "[64] Enze Xie, Wenjia Wang, Wenhai Wang, Peize Sun, Hang Xu, Ding Liang, and Ping Luo. Segmenting transparent objects in the wild with transformer. In IJCAI, 2021." + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021.", + "[66] Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic segmentation in street scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3684-3692, 2018.", + "[67] Yanchao Yang and Stefano Soatto. FDA: Fourier domain adaptation for semantic segmentation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4084-4094. IEEE, 2020.", + "[68] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 325-341, 2018.", + "[69] Yuhui Yuan, Xiaokang Chen, Xilin Chen, and Jingdong Wang. Segmentation transformer: Object-contextual representations for semantic segmentation. In European Conference on Computer Vision (ECCV), 2020.", + "[70] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. 2021.", + "[71] Qiming Zhang, Jing Zhang, Wei Liu, and Dacheng Tao. Category anchor-guided unsupervised domain adaptation for semantic segmentation. Advances in Neural Information Processing Systems, 2019.", + "[72] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017.", + "[73] Yizhou Zhou, Xiaoyan Sun, Zheng-Jun Zha, and Wenjun Zeng. Context-reinforced semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4046-4055, 2019.", + "[74] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), Oct 2017.", + "[75] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018.", + "[76] Yang Zou, Zhiding Yu, Xiaofeng Liu, BVK Kumar, and Jinsong Wang. Confidence regularized self-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5982-5991, 2019.", + "[77] Yang Zou, Zhiding Yu, Xiaofeng Liu, B. V. K. Vijaya Kumar, and Jinsong Wang. Confidence regularized self-training. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019." + ], + "bbox": [ + 501, + 90, + 892, + 853 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "16559", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/d5cc5a6e-449b-4085-98b6-42b0071a0627_model.json b/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/d5cc5a6e-449b-4085-98b6-42b0071a0627_model.json new file mode 100644 index 0000000000000000000000000000000000000000..25d7b702ca5e059ea0ed2b18e5db0a5a03f7c9fa --- /dev/null +++ b/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/d5cc5a6e-449b-4085-98b6-42b0071a0627_model.json @@ -0,0 +1,2303 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.349, + 0.131, + 0.623, + 0.153 + ], + "angle": 0, + "content": "To Adapt or Not to Adapt?" + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.155, + 0.739, + 0.177 + ], + "angle": 0, + "content": "Real-Time Adaptation for Semantic Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.202, + 0.877, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.338, + 0.891, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.438, + 0.893, + 0.469 + ], + "angle": 0, + "content": "Figure 1. Real-time adaptation with HAMLET. Online adaptation to continuous and unforeseeable domain shifts is hard and computationally expensive. HAMLET can deal with it at almost 30FPS outperforming much slower online methods - e.g. OnDA and CoTTA." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.484, + 0.314, + 0.5 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.521, + 0.474, + 0.78 + ], + "angle": 0, + "content": "The goal of Online Domain Adaptation for semantic segmentation is to handle unforeseeable domain changes that occur during deployment, like sudden weather events. However, the high computational costs associated with brute-force adaptation make this paradigm unfeasible for real-world applications. In this paper we propose HAMLET, a Hardware-Aware Modular Least Expensive Training framework for real-time domain adaptation. Our approach includes a hardware-aware back-propagation orchestration agent (HAMT) and a dedicated domain-shift detector that enables active control over when and how the model is adapted (LT). Thanks to these advancements, our approach is capable of performing semantic segmentation while simultaneously adapting at more than 29FPS on a single consumer-grade GPU. Our framework's encouraging accuracy and speed trade-off is demonstrated on OnDA and SHIFT benchmarks through experimental results." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.824, + 0.21, + 0.84 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Semantic segmentation aims at classifying an image at a pixel level, based on the local and global context, to enable a higher level of understanding of the depicted" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.494, + 0.873, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.649, + 0.894, + 0.694 + ], + "angle": 0, + "content": "Figure 2: Online adaptation methods on the Increasing Storm. We plot mIoUs achieved on single domains. Colors from colder to warmer encode slower to faster methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.723, + 0.893, + 0.828 + ], + "angle": 0, + "content": "scene. In recent years, deep learning has become the dominant paradigm to tackle this task effectively employing CNNs [5, 69, 4] or, more recently, transformers [65], at the expense of requiring large quantities of annotated images for training. Specifically, annotating for this task needs per-pixel labeling, which is an expensive and time-consuming task, severely limiting the availability of training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.829, + 0.894, + 0.875 + ], + "angle": 0, + "content": "The use of simulations and graphics engines [42] to generate annotated frames enabled a marked decrease in the time and cost necessary to gather labeled data thanks" + }, + { + "type": "page_footnote", + "bbox": [ + 0.524, + 0.887, + 0.884, + 0.901 + ], + "angle": 0, + "content": "* Joint first authorship † Part of the work done while at Univres" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16548" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.8 + ], + "angle": 0, + "content": "to the availability of the ground truth. However, despite the increasing quality in data realism [47], there is a substantial difference between simulated data generated by graphics engines and real-world images, such that leveraging these data for real-world applications requires adapting over a significant domain shift. The promise of unlocking this cheap and plentiful source of training data has provided a major impulse behind the development of a large body of work on Unsupervised Domain Adaptation (UDA) techniques [74, 61, 18, 15, 55], consisting of training semantic segmentation networks on labelled synthetic frames – the source domain – and then adapting the network to operate on real images, representing the target domain, without requiring human annotation. However, the synthetic-to-real shift represents only one of many possible domain transitions; specifically, when dealing with real-world deployment, domain shifts can occur from various causes, from different camera placements to different lighting, weather conditions, urban scenario, or any possible combination of the above. Because of the combinatorial nature of the problem, it is simply impossible to evenly represent all possible deployment domains in a dataset. This curse of dimensionality prevents having generalized robust performances [41, 45]. However, the recent advent of online domain adaptation [41] potentially allows us to face continuous and unpredictable domain shifts at deployment time, without requiring data associated with such domain shifts beforehand. Nonetheless, despite its potential, several severe limitations still hamper the online adaptation paradigm. In particular, continuously performing backpropagation on a frame-by-frame schedule [41] incurs a high computational cost, which negatively affects the performance of the network, dropping its overall fraterate to accommodate the need for continuous adaptation. Various factors are involved in this matter: first, the severity of this overhead is proportional to the complexity of the network itself – the larger the number of parameters, the heavier the adaptation process becomes; second, we argue that frame-by-frame optimization is an excessive process for the adaptation itself – not only the network might need much fewer optimization steps to effectively counter domain shifts, but also such an intense adaptation definitely increases the likelihood of catastrophic forgetting over previous domains [26, 45]. In summary, a practical solution for online domain adaptation in semantic segmentation that can effectively operate in real-world environments and applications still seems to be a distant goal." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.801, + 0.47, + 0.849 + ], + "angle": 0, + "content": "In this paper, we propose a novel framework aimed at overcoming these issues and thus allowing for real-time, online domain adaptation:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "- We address the problem of online training by designing an automatic lightweight mechanism capable of significantly reducing back-propagation complex" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "ity. We exploit the model modularity to automatically choose to train the network subset which yields the highest improvement for the allocated optimisation time. This approach reduces back-propagation FLOPS by \\(34\\%\\) while minimizing the impact on accuracy." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.177, + 0.892, + 0.283 + ], + "angle": 0, + "content": "- In an orthogonal fashion to the previous contribution, we introduce a lightweight domain detector. This allows us to design principled strategies to activate training only when it really matters as well as setting hyperparameters to maximize adaptation speed. Overall, these strategies increase our speed by over \\(5 \\times\\) while sacrificing less than \\(2.6\\%\\) in mIoU." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.293, + 0.892, + 0.369 + ], + "angle": 0, + "content": "- We evaluate our method on multiple online domain adaptation benchmarks both fully synthetic [45] and semi-synthetic CityScapes domain sequences [41], showing superior accuracy and speed compared to other test-time adaptation strategies." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.177, + 0.892, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.379, + 0.892, + 0.516 + ], + "angle": 0, + "content": "Fig. 1 demonstrates the superior real-time adaptation performance of HAMLET compared to slower methods such as CoTTA [57], which experience significant drops in performance when forced to maintain a similar framerate by adapting only once every 50 frames. In contrast, HAMLET achieves an impressive 29 FPS while maintaining high accuracy. Additionally, Fig. 2 offers a glimpse of HAMLET's performance on the Increasing Storm benchmark [41], further highlighting its favorable accuracy-speed trade-off." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.529, + 0.642, + 0.544 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.554, + 0.892, + 0.599 + ], + "angle": 0, + "content": "We review the literature relevant to our work, about semantic segmentation and UDA, with particular attention to continuous and online methodologies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.893, + 0.84 + ], + "angle": 0, + "content": "Semantic Segmentation. Very much like classification, deep learning plays a fundamental role in semantic segmentation. Fully Convolutional Network (FCN) [36] represents the pivotal step in this field, adapting common networks by means of learned upsample operators (deconvolutions). Several works aimed at improving FCN both in terms of speed [68, 38] and accuracy [5, 6, 7], with a large body of literature focusing on the latter. Major improvements have been achieved by enlarging the receptive field [72, 66, 5, 6, 7], introducing refinement modules [14, 73, 17], exploiting boundary cues [3, 10, 46] or using attention mechanisms in different flavors [13, 31, 58, 64]. The recent spread of Transformers in computer vision [11] reached semantic segmentation as well [64, 69, 65], with SegFormer [65] representing the state-of-the-art in the field and being the object of studies in the domain adaptation literature as well [20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Unsupervised Domain Adaptation (UDA). This body of research aims at adapting a network trained on a source, labeled domain to a target, unlabeled one. Early approaches rely on the notion of \"style\" and learn how to transfer it" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16549" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "across domains [74, 61, 18, 32, 12, 67]. Common strategies consist of learning domain-invariant features [15, 25], often using adversarial learning in the process [15, 55, 8, 19, 51]. A popular trend in UDA is Self-Training. These methods rely on self-supervision to learn from unlabelled data. In UDA, a successful strategy consists of leveraging target-curated pseudo-labels. Popular approaches for this purpose make use of confidence [77, 37, 76], try to balance the class predictions [75, 20], or use prototypes [2, 71, 70] to improve the quality of the pseudo-labels. Among many domain shifts, the synthetic-to-real one is the most studied, since the earliest works [74, 61, 18] to the latest [60, 30, 21, 28, 16, 40, 24]. However, this shift is one of a kind since it occurs only once after training, and without the requirement of avoiding forgetting the source domain." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.473, + 0.579 + ], + "angle": 0, + "content": "Continuous/Test-Time UDA. This family of approaches marries UDA with continuous learning, thus dealing with the catastrophic forgetting issue ignored in the synthetic-to-real case. Most continuous UDA approaches deal with it by introducing a Replay Buffer [1, 29, 27], while additional strategies make use of style transfer [62], contrastive [44, 53] or adversarial learning [63]. Despite the definition, continuous UDA often deals with offline adaptation, with well-defined target domains over which to adapt. Conceptually similar to it, is the branch of test-time adaptation, or source-free UDA, although tackling the problem in deployment rather than offline - i.e. with no access to the data from the source domain [43]. Popular strategies to deal with it consist of generating pseudo-source data to avoid forgetting [35], freezing the final layers in the model [33], aligning features [34], batch norm retraining through entropy minimization [54] or prototypes adaptation [22]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.473, + 0.779 + ], + "angle": 0, + "content": "Online UDA. Although similar in principle to test-time adaptation, online UDA [45, 41, 52] aims to tackle multiple domain shifts, occurring unpredictably during deployment in real applications and without clear boundaries between them. On this track, the SHIFT dataset [45] provides a synthetic benchmark specifically thought for this scenario, while OASIS [52] proposes a novel protocol to evaluate UDA approaches, considering an online setting and constraining the evaluated methods to deal with frame-by-frame sequences. As for methods, OnDA [41] implements self-training as the orchestration of a static and a dynamic teacher to achieve effective online adaptation while avoiding forgetting, yet introducing massive overhead." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Real-time performance is an essential aspect of online adaptation, particularly in applications such as autonomous driving where slow models are impractical. A slow adaptation process not only limits the practicality of real-world applications but also fails to provide high accuracy until the adaptation is complete, thereby defeating the original purpose. Therefore, accelerating the adaptation process is crucial for achieving high accuracy in real-time scenarios." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.09, + 0.6, + 0.107 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.404 + ], + "angle": 0, + "content": "This section introduces HAMLET, a framework for Hardware-Aware Modular Least Expensive Training. The framework aims to solve the problem of online domain adaptation with real-time performance through several synergistic strategies. First, we introduce a Hardware-Aware Modular Training (HAMT) agent able to optimize online a trade-off between model accuracy and adaptation time. HAMT allows us to significantly reduce online training time and GFLOPS. Nevertheless, the cheapest training consists of no training at all. Therefore, as the second strategy, we introduce a formal geometric model for online domain shifts that enable reliable domain shift detection and domain estimator signals (Adaptive Domain Detection, Sec. 3.3.1). These can be easily integrated to activate the adaptation process only at specific times, as least as possible. Moreover, we can further leverage these signals by designing adaptive training policies that dynamically adapt domain-sensitive hyperparameters. We refer to these as Active Training Modulations. We present an overview of HAMLET in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.41, + 0.636, + 0.426 + ], + "angle": 0, + "content": "3.1. Model Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.895, + 0.766 + ], + "angle": 0, + "content": "Our approach builds on the recent progress in unsupervised domain adaptation and segmentation networks. We start with DAFormer [20], a state-of-the-art UDA method, and adopt SegFormer [65] as our segmentation backbone due to its strong generalization capacity. We use three instances of the backbone, all pre-trained on the source domain: a student, a teacher, and a static (i.e. frozen) teacher. During training, the student receives a mix of target and source images [49] and is supervised with a \"mixed-sample\" cross-entropy loss, \\(\\mathcal{L}_T\\) (represented by green, blue and red dashed lines, in Fig. 3). This loss is computed by mixing the teacher's pseudo-labels and source annotations. To improve training stability, the teacher is updated as the exponential moving average (EMA) of the student. To further regularize the student, we use source samples stored in a replay buffer and apply two additional losses (blue lines in Fig. 3). First, we minimize the feature distance (Euclidean) between the student and the static teacher's encoder, \\(\\mathcal{L}_{FD}\\). Then, we employ a supervised cross-entropy task loss \\(\\mathcal{L}_S\\). Our complete objective is \\(\\mathcal{L} = \\mathcal{L}_S + \\mathcal{L}_T + \\lambda_{FD}\\mathcal{L}_{FD}\\), with \\(\\lambda_{FD}\\) being a weight factor. During inference on the target domain, only the student is used (red lines in Fig. 3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.885, + 0.789 + ], + "angle": 0, + "content": "3.2. Hardware-Aware Modular Training (HAMT)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Online adaptation requires updating the parameters during deployment time. However, back-propagation is computationally expensive and hence too slow to be continuously applied on a deployed agent. Opting for a partial weight update, for example by finetuning the last module of the network, would enable much more efficient training time. However, domain shifts can manifest as changes" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "16550" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.094, + 0.87, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.331, + 0.895, + 0.378 + ], + "angle": 0, + "content": "Figure 3: HAMLET framework. We employ a student-teacher model with an EMA and a static teacher. HAMT orchestrates the back-propagation over the student restricting it to a network subsection. The Active Training Modulation instead controls the adaptation process by selectively enabling it only when necessary as well as tweaking sensitive training parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.473, + 0.862 + ], + "angle": 0, + "content": "in both the data input distribution (such as attributes of the images, e.g. day/night) and the output distribution (e.g. class priors). This information could be encoded in different parts of the network, therefore just updating the very last segment might not suffice. This motivates the need for orchestrating the training process, to ensure sufficient training while minimizing the computational overhead. Inspired by reward-punishment [48] and reinforcement learning [56] policies, we introduce an orchestration agent in charge of deciding how deeply the network shall be finetuned through a trade-off between the pseudo-loss minimization rate and the computational time. In contrast to previous efficient back-propagation approaches [59, 23, 9], our model is pre-trained on the task and thus requires smaller updates to adapt. Let us start by modeling the problem. Our model backbone, \\( f \\), is composed of four different modules: \\( f = m_4 \\circ m_3 \\circ m_2 \\circ m_1 \\). This defines our action space \\( \\mathcal{A} = \\{\\mathrm{T}_1, \\mathrm{T}_2, \\mathrm{T}_3, \\mathrm{T}_4\\} \\) where \\( \\mathrm{T}_4 \\) corresponds to training just the last module of the network, \\( m_4 \\), while \\( \\mathrm{T}_3 \\) the last two modules, i.e. \\( m_4 \\circ m_3 \\), \\( \\mathrm{T}_2 \\) the last three, i.e. \\( m_4 \\circ m_3 \\circ m_2 \\), and \\( \\mathrm{T}_1 \\) the whole network \\( f \\). We also define a continuous state space \\( S = \\{\\mathrm{R}, \\mathrm{V}\\} \\) where \\( \\mathrm{R} \\) is the second derivative of the EMA teacher pseudo-loss, \\( l_t \\), over time, hence \\( R_t = -\\frac{\\Delta^2 l}{(\\Delta t)^2} \\), computed in discrete form as \\( R_t = -(l_t - 2l_{t-1} + l_{t-2}) \\). V represents a cumulative vector with the same dimension as the action space \\( \\mathcal{A} \\), initialized at zero. Now we have everything in place to employ an expected-improvement based decision model. At each time-step \\( t \\), action \\( T_j \\) is selected for \\( j = \\operatorname{argmax} V_t \\). During training step \\( t \\), \\( V[j] \\) is updated as:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.885, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathrm {V} [ j ] _ {t + 1} = \\alpha R _ {t} + (1 - \\alpha) \\mathrm {V} [ j ] _ {t} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.895, + 0.64 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a smoothing factor, e.g. 0.1. i.e. \\(\\mathrm{V_t}\\) hold a discrete exponential moving average of \\(R_{t}\\). Therefore, our policy can be seen as a greedy module selection based on the highest expected loss improvement over its linear approximation. A notable drawback of this policy is that we will inevitably converge towards picking more rewarding, yet expensive, actions i.e. \\(\\mathrm{T_1,T_2}\\) compared to more efficient but potentially less effective actions i.e. \\(\\mathrm{T}_3,\\mathrm{T}_4\\). However, our goal is not to maximize \\(-\\frac{\\Delta^2l}{(\\Delta t)^2}\\) where \\(\\Delta t\\) is the number of updates, our goal is instead to maximize \\(-\\frac{\\Delta^2l}{(\\Delta\\tau)^2}\\) where \\(\\Delta \\tau\\) is a real-time interval. Therefore, we have to introduce in the optimization policy some notion of the actual training cost of each action in \\(\\mathcal{A}\\) on the target device. To start with, we measure the training time associated with each action, obtaining \\(\\omega_{T} = \\{\\omega_{T_{1}},\\omega_{T_{2}},\\omega_{T_{3}},\\omega_{T_{4}}\\}\\). With this we can compute the time-conditioning vector \\(\\gamma\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.65, + 0.892, + 0.698 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {j} = \\frac {e ^ {\\frac {1}{\\beta \\omega_ {T _ {j}}}}}{\\sum_ {k = 1} ^ {K} e ^ {\\frac {1}{\\beta \\omega_ {T _ {k}}}}} \\quad \\text {f o r} j = 1, \\dots , K \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.707, + 0.895, + 0.768 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the softmax temperature, and \\(K\\) the number of actions, i.e. 4 in our model. We modify our update policy to favor less computationally expensive modules by scaling the updates with \\(\\gamma\\), replacing Eq. 1 with:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.792, + 0.892, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathrm {V} [ j ] _ {t + 1} = \\left\\{ \\begin{array}{l l} \\gamma_ {j} \\alpha R _ {t} + (1 - \\alpha) \\mathrm {V} [ j ] _ {t} & \\text {i f} R _ {t} \\geq 0 \\\\ (1 - \\gamma_ {j}) \\alpha R _ {t} + (1 - \\alpha) \\mathrm {V} [ j ] _ {t} & \\text {i f} R _ {t} < 0 \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.895, + 0.902 + ], + "angle": 0, + "content": "This policy makes it so that more expensive actions receive smaller rewards and larger punishments. Despite its simplicity, this leads to a significant reduction in FLOPS for an average back-propagation \\(\\beta\\), i.e. \\(-30\\%\\) with \\(\\beta = 2.75\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16551" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.47, + 0.138 + ], + "angle": 0, + "content": "or \\(-43\\%\\) with \\(\\beta = 1\\). We finally choose \\(\\beta = 1.75\\) to obtain a FLOPS reduction of \\(-34\\%\\). Exhaustive ablations on HAMT are presented in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.153, + 0.33, + 0.169 + ], + "angle": 0, + "content": "3.3. Active Training Modulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.179, + 0.473, + 0.436 + ], + "angle": 0, + "content": "Continuous and test-time adaptation methods tackle online learning as a continuous and constant process carried out on the data stream. Nevertheless, this approach presents several shortcomings when it comes to real-world deployments. Performing adaptation when the deployment domain is unchanged does not lead to further performance improvements on the current domain; instead, it might cause significant forgetting on previous domains, hence hindering model generalization (we present evidence of this in the supplementary material). Even if mitigated by HAMT, online training remains a computationally expensive procedure, also due to several teachers' necessary forward passes. However, knowing when and what kind of adaptation is needed is not a trivial task. We tackle this by introducing an Adaptive Domain Detection mechanism, in Sec. 3.3.1, and then a set of strategies to reduce the training time while optimizing the learning rate accordingly, in Sec. 3.3.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.466, + 0.327, + 0.48 + ], + "angle": 0, + "content": "3.3.1 Adaptive Domain Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.473, + 0.903 + ], + "angle": 0, + "content": "A key element of an online adaptation system consists of acquiring awareness of the trajectory in the data distribution space, i.e. domains, traveled by the student model during deployment. We can model the problem by setting the trajectory origin in the source domain. With high dimensional data, the data distribution is not tractable, therefore the trajectory cannot be described in closed form. Recent work [41] introduced the notion of distance between the current deployed domain and source by approximating it with the confidence drop of a source pre-trained model. This approach heavily relies on the assumption that the pretrained model is well-calibrated. While this might hold for domains close to source, the calibration quickly degrades in farther domains [45, 41]. This myopic behavior dampen the simple use of confidence for domain detection. Furthermore, the additional forward pass increases the computational cost during deployment. We tackle these limitations with an equivalently simple, yet more robust, approach. We modify the backbone of the static teacher \\( f^{\\mathrm{st}} \\) used for the feature distance loss \\( \\mathcal{L}_{FD} \\) by connecting a lightweight segmentation head, \\( d_1^{\\mathrm{st}} \\), after the first encoder module \\( m_1^{\\mathrm{st}} \\): \\( h_1^{\\mathrm{st}} = d_1^{\\mathrm{st}} \\circ m_1^{\\mathrm{st}} \\). This additional decoder, \\( h_1^{\\mathrm{st}} \\), is trained offline, on source data, without propagating gradients in the backbone (\\( m_1^{\\mathrm{st}} \\) is frozen). Given a target sample \\( x_T \\), we propose to compute the cross-entropy between the one-hot encoded student prediction \\( p(x_T) = 1_{\\arg \\max(f(x_T))} \\) and the lightweight decoder prediction \\( g(x_T) = h_1^{\\mathrm{st}}(x_T) \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.113, + 0.894, + 0.157 + ], + "angle": 0, + "content": "\\[\nH _ {T} ^ {(i)} = - \\sum_ {p = 1} ^ {H \\times W} \\sum_ {c = 1} ^ {C} p \\left(x _ {T} ^ {(i)}\\right) \\log g \\left(x _ {T} ^ {(i)}\\right) \\Big | _ {p, c} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.166, + 0.895, + 0.359 + ], + "angle": 0, + "content": "Thanks to the student model's higher generalization capability (both due to a larger number of parameters and the unsupervised adaptation process), it will always outperform the lightweight decoder head. Nevertheless, since now the distance is measured in the prediction space, we are not subjected to model miscalibration. Furthermore, since the student model is in constant adaptation, the domain distance accuracy actually improves over time, leading to better results. We present evidence of these claims in the supplementary material. We now define a denoised signal by using bin-averaging \\( A_T^{(i)} = \\sum_{j=m}^{m(i+1)-1} \\frac{H_T^{(j)}}{m} \\) where \\( m \\) is the bin size. Domains are modeled as discrete steps of \\( A_T^{(i)} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.376, + 0.893, + 0.413 + ], + "angle": 0, + "content": "\\[\nB _ {0} = A _ {0} \\quad B _ {i} = \\left\\{ \\begin{array}{l l} A _ {i} & \\text {i f} | B _ {i - 1} - A _ {i} | > z \\\\ B _ {i - 1} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.894, + 0.482 + ], + "angle": 0, + "content": "where \\(B\\) is the discretized signal and \\(z\\) is the minimum distance used to identify new domains. Finally, we refer to the signed amplitude of domain shifts as \\(\\Delta B_{i} = B_{i} - B_{i - 1}\\) and a domain change is detected whenever \\(|\\Delta B_i| > z\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.499, + 0.86, + 0.515 + ], + "angle": 0, + "content": "3.3.2 Least Training and Adaptive Learning Rate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The definitions of \\( B \\) allow us to customize the training process. To this end, we adopt a Least Training (LT) strategy and trigger adaptation only when facing a new domain, which occurs when \\( |\\Delta B_i| > z \\). Effective online learning performance depends heavily on the choice of hyperparameters such as the learning rate \\( \\eta \\) and learning rate decay rate. Therefore, we can adjust these parameters to facilitate adaptation according to the nature and intensity of domain shifts we encounter, we refer to this orchestration as Adaptive Learning Rate (ALR). For example, the larger the domain shift (i.e. \\( |\\Delta B_i| \\)), the more we need to adapt to counteract its effect. This can be achieved by either running more optimization steps or using a higher learning rate. Whenever a domain shift is detected, we compute the number of adaptation iterations \\( L = K_l\\frac{|\\Delta B_i|}{z} \\), hence proportionally to the amplitude of the shift \\( |\\Delta B_i| \\) relative to the threshold \\( z \\). \\( K_l \\) is a multiplicative factor representing the minimum adaptation iterations. If a new domain shift takes place before the adaptation process completes, we accumulate the required optimization steps. Then, we can play on two further parameters: \\( K_l \\) and the learning rate schedule. We argue that proper scheduling is crucial for attaining a smoother adaptation. The learning rate, \\( \\eta \\), is linearly decayed until the adaptation is concluded – the smaller the domain shift, the faster the decay. While the initial learning rate, \\( K_{\\eta} \\), should" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16552" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.088, + 0.9, + 0.229 + ], + "angle": 0, + "content": "
HAMTLTALRDCMRCS200mm (mIoU)All-domains (mIoU)FPSAverage GFLOPSAdaptation GFLOPS
TotalFwd.Bwd.Fwd.Bwd.
(A)-----62.2 ± 0.969.5 ± 0.35.9 ± 0.0125.2 ± 0.094.4 ± 0.030.8 ± 0.056.6 ± 0.030.8 ± 0.0
(B)✓----60.2 ± 0.568.7 ± 0.37.0 ± 0.1114.7 ± 0.094.4 ± 0.020.3 ± 0.056.6 ± 0.020.3 ± 0.0
(C)✓✓---51.8 ± 0.565.7 ± 0.229.5 ± 0.644.4 ± 0.542.6 ± 0.41.8 ± 0.256.6 ± 0.020.2 ± 0.2
(D)✓✓✓--54.1 ± 1.265.9 ± 0.229.5 ± 0.544.4 ± 0.342.7 ± 0.21.8 ± 0.156.6 ± 0.020.3 ± 0.1
(E)✓✓✓✓-56.6 ± 0.866.3 ± 0.128.9 ± 0.344.7 ± 0.242.9 ± 0.21.8 ± 0.156.6 ± 0.020.2 ± 0.0
(F)✓✓✓-✓55.8 ± 1.066.3 ± 0.229.1 ± 1.145.2 ± 0.143.2 ± 0.12.0 ± 0.056.6 ± 0.020.3 ± 0.0
(G)✓✓✓✓✓58.2 ± 0.866.9 ± 0.329.7 ± 0.645.7 ± 0.343.6 ± 0.22.1 ± 0.156.6 ± 0.020.2 ± 0.1
(a)
clear 1200mmclear 2100mmclear 375mmclear 4clear h-meantarget h-meantotal h-meanFPSGFLOPS
(A)72.952.273.664.273.067.673.473.260.667.25.6125.2
(B)73.050.473.462.173.067.373.273.159.166.46.8114.7
(C)73.446.073.561.573.666.173.873.656.565.17.2100.0
(G)73.453.673.165.273.568.273.273.361.667.89.182.2
(b)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.233, + 0.892, + 0.263 + ], + "angle": 0, + "content": "Table 1: Ablation studies - HAMLET components. Top: Increasing Storm (8925 frames per domain) [41], bottom: Fast Storm C [41] (2975 frames per domain). For each configuration, we report mIoU, framerate, and GFLOPS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.468, + 0.307 + ], + "angle": 0, + "content": "be higher when the domain shift is triggered in domains farther from the source" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.33, + 0.469, + 0.358 + ], + "angle": 0, + "content": "\\[\nK _ {\\eta} = K _ {\\eta , \\min } + \\frac {\\left(B _ {\\mathrm {i}} - B _ {\\text {s o u r c e}}\\right) \\left(K _ {\\eta , \\max } - K _ {\\eta , \\min }\\right)}{B _ {\\text {h a r d}} - B _ {\\text {s o u r c e}}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.368, + 0.469, + 0.489 + ], + "angle": 0, + "content": "where \\( B_{\\mathrm{source}} \\) (resp. \\( B_{\\mathrm{hard}} \\)) is an estimate of \\( B \\) when the network is close to (resp. far from) the source domain; and \\( K_{\\eta, \\min} \\) (resp. \\( K_{\\eta, \\max} \\)) is the value of \\( K_{\\eta} \\) assigned when the network is close to (resp. far away from) the source. Concerning \\( K_{l} \\), we posit that moving towards the source requires less adaptation than going towards harder domains: the model shows good recalling of previously explored domains and thanks to the employed regularization strategies" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.512, + 0.469, + 0.549 + ], + "angle": 0, + "content": "\\[\nK _ {l} = \\left\\{ \\begin{array}{l l} K _ {l, \\max } & \\text {i f} \\Delta B _ {i} \\geq 0 \\\\ K _ {1, \\min } + \\frac {\\left(B _ {\\mathrm {i}} - B _ {\\text {s o u r c e}}\\right) \\left(K _ {1 , \\max } - K _ {1 , \\min }\\right)}{B _ {\\text {h a r d}} - B _ {\\text {s o u r c e}}} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.558, + 0.469, + 0.633 + ], + "angle": 0, + "content": "where \\( K_{l,\\min} \\) (resp. \\( K_{l,\\max} \\)) is the value of \\( K_{l} \\) assigned when the model is close to (resp. far away from) the source domain. Extensive ablations in the supplementary material will highlight how the orchestration of the adaptation hyperparameters improves the accuracy-speed trade-off." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.655, + 0.317, + 0.669 + ], + "angle": 0, + "content": "3.3.3 Dynamic ClassMix (DCM)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.469, + 0.86 + ], + "angle": 0, + "content": "ClassMix [39] provides a simple mechanism for data augmentation by mixing classes from the source dataset into target images. Usually \\(50\\%\\) of the classes in the source dataset are selected, however we notice that this percentage is a highly sensitive hyperparameter in online domain adaptation. Injecting a significant portion of source classes has a beneficial impact when adapting to domains closer to the source domain, whereas when adapting to domains further from the source the opposite effect can be observed, as it effectively slows down the adaptation process. We therefore exploit once more the deployment domain awareness to control the mixing augmentation:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.877, + 0.469, + 0.906 + ], + "angle": 0, + "content": "\\[\nK _ {\\mathrm {C M}} = K _ {\\mathrm {C M}, \\min } + \\frac {\\left(B _ {\\mathrm {i}} - B _ {\\text {s o u r c e}}\\right) \\left(K _ {\\mathrm {C M} , \\max } - K _ {\\mathrm {C M} , \\min }\\right)}{B _ {\\text {h a r d}} - B _ {\\text {s o u r c e}}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.278, + 0.892, + 0.337 + ], + "angle": 0, + "content": "where \\( K_{\\mathrm{CM}} \\) is the percentage of source classes used during adaptation; and \\( K_{\\mathrm{CM,min}} \\) (resp. \\( K_{\\mathrm{CM,max}} \\)) is the value of \\( K_{\\mathrm{CM}} \\) assigned when the network is close to (resp. far away from) the source domain." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.356, + 0.671, + 0.371 + ], + "angle": 0, + "content": "3.3.4 Buffer Sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Following [41], to simulate real deployment, we limit our access to the source domain by using a replay buffer. Additionally, instead of initializing at random (with a uniform prior), we apply Rare Class Sampling (RCS) (skewed priors) as in [20]. This incentives a more balanced class distribution over the buffer, ultimately leading to better accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.482, + 0.705, + 0.499 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.704 + ], + "angle": 0, + "content": "The experiments are carried out on (a) the OnDA benchmarks [41] and (b) the SHIFT dataset [45]. (a) is a semi-syntheptic benchmark, as it applies synthetic rain and fog [50] over 4 different intensities profiles. The main benchmark, Increasing Storm, presents a storm with a pyramidal intensity profile; see Fig. 4. In contrast, (b) is a purely synthetic dataset, where both the underlying image and the weather are synthetically generated and thus domain change is fully controllable. All models are evaluated using mIoU: following [41], we report the harmonic mean over domains to present the overall adaptation performance. All experiments were carried out using an Nvidia™ RTX 3090 GPU. We refer to supplementary material for further details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.665, + 0.726 + ], + "angle": 0, + "content": "4.1. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.9 + ], + "angle": 0, + "content": "In Tab. 1 we study the impact of each contribution to adaptation performance, both in terms of accuracy and efficiency. For each configuration, we report mIoU over different portions of the sequence, the framerate and the amount of GFLOPS - respectively averages of: total, forward and backward passes, and dedicated adaptation only, also divided in forward (Fwd) and backward (Bwd). Tab. 1 (a) shows results on the Increasing Storm scenario [41]. Here, we show mIoU over the \\(200\\mathrm{mm}\\) domain, i.e. the hardest in the sequence, as well as the mIoU averaged over forward and backward adaptation, i.e., from clear to \\(200\\mathrm{mm}\\) rain" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16553" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.872, + 0.246 + ], + "angle": 0, + "content": "
clear25mm50mm75mm100mm200mmh-meanFPSGFLOPS
FBFBFBFBFBFBFBT
(A)DeepLabV2 (no adaptation)64.5-57.1-48.7-41.5-34.4-18.537.3--39.4-
(B)DeepLabV2 fully supervised (oracle)64.5-64.1-63.7-63.0-62.4-58.262.6--39.4-
(C)OnDA64.564.860.457.157.354.554.852.252.049.142.254.255.1-1.3-
(D)SegFormer MiT-B1 (no adaptation)73.4-68.8-64.2-58.0-51.8-31.257.8--48.434.9
(E)SegFormer MiT-B5 (no adaptation)77.6-73.9-71.0-67.2-62.6-46.764.7--11.5240.4
(F)SegFormer MiT-B1 fully supervised (oracle)72.9-72.4-72.1-71.5-70.7-68.671.3--48.434.9
(G)TENT73.072.868.568.664.564.859.760.254.554.835.956.263.659.910.0-
(H)TENT + Replay Buffer73.072.868.568.664.564.859.760.254.454.735.856.163.659.97.8-
(I)CoTTA72.574.469.570.965.968.266.164.764.663.557.265.668.166.80.6593.8
(J)CoTTA real-time73.375.470.370.666.966.462.561.457.656.939.759.265.562.327.041.7
(K)HAMLET (ours)73.471.070.168.867.767.566.666.465.564.659.266.867.667.229.145.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.247, + 0.892, + 0.278 + ], + "angle": 0, + "content": "Table 2: Comparison against other models – Increasing storm scenario. (A-C) methods built over DeepLabv2, (D-E) SegFormer variants trained on source, (F) oracle, (G-K) models adapted online. We report mIoU, framerate, and GFLOPS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.473, + 0.835 + ], + "angle": 0, + "content": "and backward. Results are averaged over 3 runs with different seeds, with standard deviation being reported. (A) reports the results achieved by naively performing full adaptation of the model. HAMT can increase the framerate by roughly \\(15\\%\\) by reducing the Bwd GFLOPS of \\(34\\%\\), at the expense of as few as \\(0.7\\mathrm{mIoU}\\) on average, i.e., about 2 points on the \\(200\\mathrm{mm}\\) domain. The main boost in terms of speed is obviously given by LT (C), which inhibits the training in absence of detected domain shifts. LT increases the framerate by approximately \\(4\\times\\) by decimating the total GFLOPS, yet not affecting the adaptation Bwd GFLOPS. This comes with a price in terms of mIoU, dropping by about 4 points on average and more than 10 points on \\(200\\mathrm{mm}\\) - not a moderate drop anymore. LT impact highly depends on the domain sequence experienced during deployment: frequent domain changes could prevent training inhibition, thus neglecting LT gains in terms of efficiency, as we will appreciate later. The loss in accuracy is progressively regained by adding ALR (D), with further improvements yielded by one between DCM (E) and RCS (F), or both together (G) leading to the full HAMLET configuration. The three together allow for reducing the gap to 2.5 points mIoU - 4 over the \\(200\\mathrm{mm}\\) domain - without sacrificing any efficiency. Tab. 1 (b) shows further results, on a faster version of Storm C [41]. This represents a much more challenging scenario, with harsher and \\(3\\times\\) more frequent domain shifts. Here we show the single domains mIoU, as well as harmonic mean on source and target domains, and all frames. As expected, in this benchmark, LT alone (C) results much less effective than before, with a much lower gain in FPS and GFLOPS. Here, the synergy between the HAMT, LT, and the other components (G) allows for the best accuracy and speedup - even outperforming the full training variant (A) - highlighting their complementarity. Further ablations are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.332, + 0.863 + ], + "angle": 0, + "content": "4.2. Results on Increasing Storm" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Tab. 2 shows a direct comparison between HAMLET and relevant approaches. The presented test-time adaptation" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.311, + 0.851, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.35, + 0.849, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.421, + 0.848, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.5, + 0.848, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.583, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Figure 4: HAMLET on the Increasing Storm. We show rain intensity (in millimetres), mIoU over active (bold) and inactive (dashed) domains, learning rate and FPS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "strategies namely - TENT and CoTTA - were revised to handle the online setting and be fairly compared with HAMLET. All methods start with the same exact initial weights - with HAMLET requiring the additional lightweight decoder, not needed by TENT and CoTTA - using SegFormer MiT-B1 as the backbone, since it is \\(4 \\times\\) faster than SegFormer MiT-B5 and thus better suited to keep real-time performance even during adaptation. We report results achieved by DeepLabv2 trained on source data only (A), an oracle model trained with full supervision (B), as well as OnDA [41] (C) as a reference. Then, we report SegFormer models trained on the source domain only (D) and (E). In (F) we show the performance achieved by an oracle SegFormer, trained on all domains fully supervised. Following [41], columns \"F\" concern forward adaptation from clear to \\(200\\mathrm{mm}\\), while columns \"B\" show backward adap" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16554" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.86, + 0.149 + ], + "angle": 0, + "content": "
clear750m375m150m75mh-meanFPSGFLOPS
FBFBFBFBFBFBT
OnDA64.965.863.362.360.758.851.649.142.155.154.1-1.3-
SegFormer MiT-B1 (no adaptation)71.1-70.0-67.5-58.8-46.961.3--48.434.9
Full training71.572.172.974.771.973.167.668.161.368.771.970.35.6125.2
HAMLET (ours)71.171.670.370.868.869.264.364.357.065.968.967.424.850.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.151, + 0.891, + 0.166 + ], + "angle": 0, + "content": "Table 3: Results on foggy domains. Comparison between OnDA, Source SegFormer, full training adaptation, and HAMLET." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.169, + 0.887, + 0.229 + ], + "angle": 0, + "content": "
ClearCloudyOvercastSmall rainMid rainHeavy rainh-meanFPSGFLOPS
FBFBFBFBFBFBFBT
SegFormer MiT-B1 fully supervised (oracle)80.1-79.9-79.8-78.9-78.7-77.179.1--48.434.93
SegFormer MiT-B1 (no adaptation)79.6-77.1-75.4-73.4-71.4-66.773.7--48.434.93
Full training78.979.376.776.876.877.974.874.876.376.574.076.277.076.65.0125.1
HAMLET (ours)79.678.976.976.676.177.473.374.374.276.074.275.776.676.126.843.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.23, + 0.888, + 0.245 + ], + "angle": 0, + "content": "Table 4: Results on SHIFT dataset [45]. Comparison between Source SegFormer, full training adaptation, and HAMLET." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.272, + 0.473, + 0.896 + ], + "angle": 0, + "content": "tation from \\(200\\mathrm{mm}\\) to clear, while the h-mean T refers to the overall harmonic mean. We can notice how SegFormer results are much more robust to domain changes with respect to DeepLabv2. Indeed, SegFormer MiT-B5 (E), without any adaptation, results more accurate than DeepLabv2 oracle (B), as well as better and faster than OnDA (C). The faster variant (D) outperforms OnDA both in speed and accuracy, reaching 48 FPS. Nevertheless, domain changes still dampen the full potential of SegFormer. Indeed, the oracle (F) outperforms (D) by about \\(+14\\) mIoU. However, this is not meaningful for real deployment experiencing unpredictable domain shifts, as it assumes to have data available in advance. Concerning test-time models, TENT starts adapting properly only beyond \\(50\\mathrm{mm}\\), both with (G) and without (H) frame buffer, while it loses some accuracy on \\(25\\mathrm{mm}\\). This makes its overall forward adaptation performance slightly worse compared to the pre-trained model (D), while being better at backward adaptation. Despite outperforming SegFormer MiT-B1, TENT is both slower and less accurate than SegFormer MiT-B5 running without any adaptation, further suggesting the robustness of the latter and making TENT not suitable for real-world deployment. On the contrary, CoTTA (I) outperforms both SegFormer models trained on source only, at the expense of dropping the framerate below 1FPS. It is worth mentioning that these metrics were collected after each domain was completed by each model individually. In an evaluation setup imposing a shared time frame, slower models would present much lower metrics, since their adaptation process would result constantly lagged. In fact, forcing CoTTA to run in real-time, at nearly 30FPS – i.e. by training once every 50 frames – dramatically reduces the effectiveness of the adaptation process (J), with drastic drops in the hardest domains. Finally, HAMLET (K) succeeds on any fronts, improving the baseline (D) by about 10 points with only a cost of \\(25\\%\\) in terms of speed, while outperforming SegFormer MiT-B5 (E) both on accuracy \\((+2.5\\) mIoU) and speed \\((3\\times\\) faster) – being the only method achieving this, and thus the only suitable choice for real-time applications. Fig. 4 shows the overall behavior of HAMLET while adapting over the Increasing Storm. In addition to the rain intensity and the" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.279, + 0.852, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.551, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Figure 5: HAMLET on the SHIFT benchmark. We show mIoU over active (bold) and inactive (dashed) domains, learning rate and FPS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.777 + ], + "angle": 0, + "content": "mIoU achieved on each domain - active (bold) or inactive (dashed), i.e. respectively the mIoU on the domain being currently faced during deployment, and how the current adaptation affects the performance on the other domains to highlight the robustness to forgetting - we also report how the learning rate is modulated in correspondence of detected domain shifts, with a consequent drop in FPS due to the short training process taking place. For further experiments on harsher and sudden adaptation cycles, we include results of Storms A, B, C [41] in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.812, + 0.803 + ], + "angle": 0, + "content": "4.3. Additional Results: Fog and SHIFT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Fog. In Tab. 3, we investigate adaptation on the Increasing Fog scenario in the OnDA benchmark [41]. Crucially, for this experiment, we keep the same hyperparameters used for the Increasing Storm, since in both cases the starting SegFormer model is trained on the same source domain. This allows for validating how the proposed setting general-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16555" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.138, + 0.089, + 0.835, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.317, + 0.895, + 0.351 + ], + "angle": 0, + "content": "Figure 6: Qualitative results - HAMLET in action. From left to right, we show frames from clean, \\(50\\mathrm{mm}\\), \\(100\\mathrm{mm}\\), and \\(200\\mathrm{m}\\) domains. From top to bottom: input image, prediction by SegFormer trained on source domain and HAMLET." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.473, + 0.51 + ], + "angle": 0, + "content": "izes at dealing with different kind of domain shifts, beyond those considered in the main experiments. We effectively use Increasing Fog as test set, and compare against Seg-Former trained on source (no adaptation) and a model that has been adapted by means of full online training optimization (configuration (A) of Table 1). HAMLET is able to adapt almost as well as the full online training model, with less than a \\(3\\mathrm{mIoU}\\) gap, while enjoying real-time adaptation at nearly \\(5\\times\\) the speed using just \\(40\\%\\) of the FLOPS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.47, + 0.722 + ], + "angle": 0, + "content": "SHIFT. We further test HAMLET on the SHIFT dataset [45]. Tab. 4 collects the results achieved by Seg-Former trained on source, full online training and HAMLET respectively, both at forward and backward adaptation across Clear, Cloudy, Overcast, Small rain, Mid rain and Heavy rain domains. Here HAMLET results highly competitive with the full training regime, with only 0.5 drop in average mIoU, while being more than \\(5 \\times\\) faster. Fig. 5 depicts, from top to bottom, the rain intensity characterizing any domain encountered on SHIFT, the mIoU achieved both on current (bold) and inactive (dashed) domains, the learning rate changes based on the domain shift detection, and the framerate achieved at any step. We refer to the supplementary material for a deeper analysis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.723, + 0.47, + 0.8 + ], + "angle": 0, + "content": "Qualitative results. To conclude, Fig. 6 shows some qualitative examples from CityScapes. We can notice how SegFormer accuracy (second tow) drops with severe rain, whereas HAMLET (third row) is capable of keeping the same segmentation quality across the storm." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.814, + 0.192, + 0.83 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Orthogonality. HAMT and LT act independently. Indeed, by strongly constraining the adaptation periods through LT, HAMT has a limited margin of action. The impact of HAMT also depends on the backbone and by care" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.892, + 0.435 + ], + "angle": 0, + "content": "fully crafting modular architectures, one can achieve further optimization. Nevertheless, in a deployment environment where domain shifts occur at high frequencies (e.g., Storm C), LT is ineffective, while HAMT thrives." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.892, + 0.496 + ], + "angle": 0, + "content": "Measuring forgetting. An interesting topic we have not investigated consists of introducing an explicit awareness of which domains have been explored and how well we can recall them, expanding the distance \\( B \\) to multiple dimensions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.892, + 0.589 + ], + "angle": 0, + "content": "Safety. We believe dynamic adaptation has the potential to enhance safety, but we acknowledge the necessity for rigorous testing and verification to safeguard against drift or catastrophic forgetting. This mandates a comprehensive effort from academia, industry, and certification authorities for ensuring the integrity of dynamically adapting models." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.605, + 0.727, + 0.621 + ], + "angle": 0, + "content": "6. Summary & Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.892, + 0.798 + ], + "angle": 0, + "content": "We have presented HAMLET, a framework for real-time adaptation for semantic segmentation that achieves state-of-the-art performance on established benchmarks with continuous domain changes. Our approach combines a hardware-aware backpropagation orchestrator and a specialized domain-shift detector to enable active control over the model's adaptation, resulting in high framereates on a consumer-grade GPU. These advancements enable HAMLET to be a promising solution for in-the-wild deployment, making it a valuable tool for applications that require robust performance in the face of unforeseen domain changes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. The authors thank Gianluca Villani for the insightful discussion on reward-punishment policies, Leonardo Ravaglia for his expertise on hardware-aware training, and Lorenzo Andraghetti for exceptional technical support throughout the project. Their assistance was invaluable in the completion of this work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16556" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Andreea Bobu, Judy Hoffman, Eric Tzeng, and Trevor Darrell. Adapting to continuously shifting domains. In ICLR 2018 Workshop Program Chairs, 2018. 00000." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Chaoqi Chen, Weiping Xie, Wenbing Huang, Yu Rong, Xinghao Ding, Yue Huang, Tingyang Xu, and Junzhou Huang. Progressive feature alignment for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 627-636, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.472, + 0.325 + ], + "angle": 0, + "content": "[3] Liang-Chieh Chen, Jonathan T Barron, George Papandreou, Kevin Murphy, and Alan L Yuille. Semantic image segmentation with task-specific edge detection using cnns and a discriminatively trained domain transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4545-4554, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.47, + 0.409 + ], + "angle": 0, + "content": "[4] Liang-Chieh Chen, Raphael Gontijo Lopes, Bowen Cheng, Maxwell D Collins, Ekin D Cubuk, Barret Zoph, Hartwig Adam, and Jonathon Shlens. Naive-student: Leveraging semi-supervised learning in video sequences for urban scene segmentation. In European Conference on Computer Vision, pages 695-714. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.47, + 0.48 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(4):834-848, Apr 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.633 + ], + "angle": 0, + "content": "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[8] Yi-Hsin Chen, Wei-Yu Chen, Yu-Ting Chen, Bo-Cheng Tsai, Yu-Chiang Frank Wang, and Min Sun. No more discrimination: Cross city adaptation of road scene segmenters. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2011-2020. IEEE, 2017. 00000." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.705, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[9] Feng Cheng, Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Li, and Wei Xia. Stochastic backpropagation: A memory efficient strategy for training video models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8301-8310, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[10] Henghui Ding, Xudong Jiang, Ai Qun Liu, Nadia Magnenat Thalmann, and Gang Wang. Boundary-aware feature propagation for scene segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6819-6829, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[12] A. Dundar, M. Y. Liu, Z. Yu, T. C. Wang, J. Zedlewski, and J. Kautz. Domain stylization: A fast covariance matching framework towards domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[13] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3146-3154, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[14] Jun Fu, Jing Liu, Yuhang Wang, Yong Li, Yongjun Bao, Jinhui Tang, and Hanqing Lu. Adaptive context network for scene parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6748-6757, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[15] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, Jan. 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[16] Rui Gong, Martin Danelljan, Dengxin Dai, Danda Pani Paudel, Ajad Chhatkuli, Fisher Yu, and Luc Van Gool. Tacs: Taxonomy adaptive cross-domain semantic segmentation. In European Conference on Computer Vision, pages 19-35. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[17] Junjun He, Zhongying Deng, Lei Zhou, Yali Wang, and Yu Qiao. Adaptive pyramid context network for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7519-7528, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[18] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. CyCADA: Cycle-consistent adversarial domain adaptation. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 1989-1998, Stockholm, Sweden, 10-15 Jul 2018. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[19] Judy Hoffman, Dequan Wang, Fisher Yu, and Trevor Darryll. FCNs in the wild: Pixel-level adversarial and constraint-based adaptation. CoRR, 2016. 00000." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[20] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. arXiv preprint arXiv:2111.14887, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[21] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. *Hrda: Context-aware high-resolution domain-adaptive semantic segmentation*. In European Conference on Computer Vision (ECCV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[22] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Angela H Jiang, Daniel L-K Wong, Giulio Zhou, David G Andersen, Jeffrey Dean, Gregory R Ganger, Gauri Joshi," + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16557" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Michael Kaminsky, Michael Kozuch, Zachary C Lipton, et al. Accelerating deep learning by focusing on the biggest losers. arXiv preprint arXiv:1910.00762, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[24] Zhengkai Jiang, Yuxi Li, Ceyuan Yang, Peng Gao, Yabiao Wang, Ying Tai, and Chengjie Wang. Prototypical contrast adaptation for domain adaptive semantic segmentation. In European Conference on Computer Vision, pages 36-54. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[25] Myeongjin Kim and Hyeran Byun. Learning texture invariant representation for domain adaptation of semantic segmentation. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.469, + 0.359 + ], + "angle": 0, + "content": "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114, 12 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[27] Yevhen Kuznietsov, Marc Proesmans, and Luc Van Gool. Towards unsupervised online domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 261-271, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[28] Xin Lai, Zhuotao Tian, Xiaogang Xu, Yingcong Chen, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Decouplenet: Decoupled network for domain adaptive semantic segmentation. In European Conference on Computer Vision, pages 369-387. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[29] Qicheng Lao, Xiang Jiang, Mohammad Havaei, and Yoshua Bengio. Continuous domain adaptation with variational domain-agnostic feature replay. arXiv preprint arXiv:2003.04382, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[30] Geon Lee, Chanho Eom, Wonkyung Lee, Hyekang Park, and Bumsub Ham. Bi-directional contrastive learning for domain adaptive semantic segmentation. In European Conference on Computer Vision, pages 38-55. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[31] Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention networks for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9167-9176, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[32] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[33] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. CoRR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[34] Yuejiang Liu, Parth Kothari, Bastien Germain van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++: When does self-supervised test-time training fail or thrive? In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[35] Yang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[36] Jonathan Long, Evan Shelhamer, and Trevor Darrel. Fully convolutional networks for semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[37] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. Lecture Notes in Computer Science, page 415-430, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[38] Vladimir Nekrasov, Chunhua Shen, and Ian Reid. Lightweight refinenet for real-time semantic segmentation. In British Conference on Computer Vision (BMVC), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[39] Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1369-1378, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[40] Fei Pan, Sungsu Hur, Seokju Lee, Junsik Kim, and In So Kweon. Ml-bpm: Multi-teacher learning with bidirectional photometric mixing for open compound domain adaptation in semantic segmentation. In European Conference on Computer Vision, pages 236–251. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[41] Theodoros Panagiotakopoulos, Pier Luigi Dovesi, Linus Harenstam-Nielsen, and Matteo Poggi. Online domain adaptation for semantic segmentation in ever-changing conditions. In European Conference on Computer Vision (ECCV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[42] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In European conference on computer vision, pages 102-118. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[43] Serban Stan and Mohammad Rostami. Unsupervised model adaptation for continual semantic segmentation. In AAAI, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[44] Peng Su, Shixiang Tang, Peng Gao, Di Qiu, Ni Zhao, and Xiaogang Wang. Gradient regularized contrastive learning for continual domain adaptation. 2020. 00000." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[45] Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. SHIFT: a synthetic driving dataset for continuous multi-task domain adaptation. In Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[46] Towaki Takikawa, David Acuna, Varun Jampani, and Sanja Fidler. Gated-scnn: Gated shape cnns for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5229-5238, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[47] Phillip Thomas, Lars Pandikow, Alex Kim, Michael Stanley, and James Grieve. Open synthetic dataset for improving cyclist detection, Nov 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[48] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 195-204, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.833, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[49] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1379-1389, January 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16558" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[50] Maxime Tremblay, Shirsendu S. Halder, Raoul de Charette, and Jean-François Lalonde. Rain rendering for evaluating and improving robustness to bad weather. International Journal of Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[51] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Jun 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.273 + ], + "angle": 0, + "content": "[52] Riccardo Volpi, Pau de Jorge, Diane Larlus, and Gabriela Csurka. On the road to online adaptation for semantic image segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.327 + ], + "angle": 0, + "content": "[53] Vibashan VS, Poojan Oza, and Vishal M. Patel. Towards online domain adaptive object detection. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Jan 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.328, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[54] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.47, + 0.441 + ], + "angle": 0, + "content": "[55] Haoran Wang, Tong Shen, Wei Zhang, Lingyu Duan, and Tao Mei. Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In The European Conference on Computer Vision (ECCV), August 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.47, + 0.482 + ], + "angle": 0, + "content": "[56] Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. Haq: Hardware-aware automated quantization with mixed precision, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.47, + 0.51 + ], + "angle": 0, + "content": "[57] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[58] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[59] Bingzhen Wei, Xu Sun, Xuancheng Ren, and Jingjing Xu. Minimal effort back propagation for convolutional neural networks. arXiv preprint arXiv:1709.05804, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[60] Tsung-Han Wu, Yi-Syuan Liou, Shao-Ji Yuan, Hsin-Ying Lee, Tung-I Chen, Kuan-Chih Huang, and Winston H Hsu. D2ada: Dynamic density-aware active domain adaptation for semantic segmentation. In European Conference on Computer Vision (ECCV), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[61] Zuxuan Wu, Xintong Han, Yen-Liang Lin, Mustafa Gokhan Uzunbas, Tom Goldstein, Ser Nam Lim, and Larry S. Davis. Dcan: Dual channel-wise alignment networks for unsupervised scene adaptation. Lecture Notes in Computer Science, page 535-552, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[62] Zuxuan Wu, Xin Wang, Joseph Gonzalez, Tom Goldstein, and Larry Davis. ACE: Adapting to changing environments for semantic segmentation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2121-2130. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[63] Markus Wulfmeier, Alex Bewley, and Ingmar Posner. Incremental adversarial domain adaptation for continually changing environments. 2018. 00000." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[64] Enze Xie, Wenjia Wang, Wenhai Wang, Peize Sun, Hang Xu, Ding Liang, and Ping Luo. Segmenting transparent objects in the wild with transformer. In IJCAI, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.16 + ], + "angle": 0, + "content": "[65] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.218 + ], + "angle": 0, + "content": "[66] Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic segmentation in street scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3684-3692, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.893, + 0.275 + ], + "angle": 0, + "content": "[67] Yanchao Yang and Stefano Soatto. FDA: Fourier domain adaptation for semantic segmentation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4084-4094. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.893, + 0.345 + ], + "angle": 0, + "content": "[68] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 325-341, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.893, + 0.402 + ], + "angle": 0, + "content": "[69] Yuhui Yuan, Xiaokang Chen, Xilin Chen, and Jingdong Wang. Segmentation transformer: Object-contextual representations for semantic segmentation. In European Conference on Computer Vision (ECCV), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.893, + 0.458 + ], + "angle": 0, + "content": "[70] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.459, + 0.893, + 0.515 + ], + "angle": 0, + "content": "[71] Qiming Zhang, Jing Zhang, Wei Liu, and Dacheng Tao. Category anchor-guided unsupervised domain adaptation for semantic segmentation. Advances in Neural Information Processing Systems, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.516, + 0.893, + 0.557 + ], + "angle": 0, + "content": "[72] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.559, + 0.893, + 0.615 + ], + "angle": 0, + "content": "[73] Yizhou Zhou, Xiaoyan Sun, Zheng-Jun Zha, and Wenjun Zeng. Context-reinforced semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4046-4055, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.616, + 0.893, + 0.671 + ], + "angle": 0, + "content": "[74] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), Oct 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.893, + 0.74 + ], + "angle": 0, + "content": "[75] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.742, + 0.893, + 0.798 + ], + "angle": 0, + "content": "[76] Yang Zou, Zhiding Yu, Xiaofeng Liu, BVK Kumar, and Jinsong Wang. Confidence regularized self-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5982-5991, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.799, + 0.893, + 0.854 + ], + "angle": 0, + "content": "[77] Yang Zou, Zhiding Yu, Xiaofeng Liu, B. V. K. Vijaya Kumar, and Jinsong Wang. Confidence regularized self-training. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16559" + } + ] +] \ No newline at end of file diff --git a/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/d5cc5a6e-449b-4085-98b6-42b0071a0627_origin.pdf b/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/d5cc5a6e-449b-4085-98b6-42b0071a0627_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..64ad023f7bca74674a376ac0bc3bab3fe73f8f4e --- /dev/null +++ b/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/d5cc5a6e-449b-4085-98b6-42b0071a0627_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c1461cf022dbdbbca9dd7244091e844dce45efff1ffe46e9ab52dd06fd0208e +size 9889703 diff --git a/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/full.md b/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e1c98b1849cfeed30532b4a36b3e87ac3c3ee3f3 --- /dev/null +++ b/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/full.md @@ -0,0 +1,305 @@ +# To Adapt or Not to Adapt? + +# Real-Time Adaptation for Semantic Segmentation + +![](images/94a51354b0137b899c5ed9b46e74d97ea543d6355691a13a73f671aadecada51.jpg) + +![](images/9b531da2a21f0adb6eccf38eac398688d2049a0530162967e18a2c900ed1113b.jpg) +Figure 1. Real-time adaptation with HAMLET. Online adaptation to continuous and unforeseeable domain shifts is hard and computationally expensive. HAMLET can deal with it at almost 30FPS outperforming much slower online methods - e.g. OnDA and CoTTA. + +# Abstract + +The goal of Online Domain Adaptation for semantic segmentation is to handle unforeseeable domain changes that occur during deployment, like sudden weather events. However, the high computational costs associated with brute-force adaptation make this paradigm unfeasible for real-world applications. In this paper we propose HAMLET, a Hardware-Aware Modular Least Expensive Training framework for real-time domain adaptation. Our approach includes a hardware-aware back-propagation orchestration agent (HAMT) and a dedicated domain-shift detector that enables active control over when and how the model is adapted (LT). Thanks to these advancements, our approach is capable of performing semantic segmentation while simultaneously adapting at more than 29FPS on a single consumer-grade GPU. Our framework's encouraging accuracy and speed trade-off is demonstrated on OnDA and SHIFT benchmarks through experimental results. + +# 1. Introduction + +Semantic segmentation aims at classifying an image at a pixel level, based on the local and global context, to enable a higher level of understanding of the depicted + +![](images/4bc99e80e8f50d2a7405896a09971660fe857c9fdcbe03c4b1d93eb72c00b6c4.jpg) +Figure 2: Online adaptation methods on the Increasing Storm. We plot mIoUs achieved on single domains. Colors from colder to warmer encode slower to faster methods. + +scene. In recent years, deep learning has become the dominant paradigm to tackle this task effectively employing CNNs [5, 69, 4] or, more recently, transformers [65], at the expense of requiring large quantities of annotated images for training. Specifically, annotating for this task needs per-pixel labeling, which is an expensive and time-consuming task, severely limiting the availability of training data. + +The use of simulations and graphics engines [42] to generate annotated frames enabled a marked decrease in the time and cost necessary to gather labeled data thanks + +to the availability of the ground truth. However, despite the increasing quality in data realism [47], there is a substantial difference between simulated data generated by graphics engines and real-world images, such that leveraging these data for real-world applications requires adapting over a significant domain shift. The promise of unlocking this cheap and plentiful source of training data has provided a major impulse behind the development of a large body of work on Unsupervised Domain Adaptation (UDA) techniques [74, 61, 18, 15, 55], consisting of training semantic segmentation networks on labelled synthetic frames – the source domain – and then adapting the network to operate on real images, representing the target domain, without requiring human annotation. However, the synthetic-to-real shift represents only one of many possible domain transitions; specifically, when dealing with real-world deployment, domain shifts can occur from various causes, from different camera placements to different lighting, weather conditions, urban scenario, or any possible combination of the above. Because of the combinatorial nature of the problem, it is simply impossible to evenly represent all possible deployment domains in a dataset. This curse of dimensionality prevents having generalized robust performances [41, 45]. However, the recent advent of online domain adaptation [41] potentially allows us to face continuous and unpredictable domain shifts at deployment time, without requiring data associated with such domain shifts beforehand. Nonetheless, despite its potential, several severe limitations still hamper the online adaptation paradigm. In particular, continuously performing backpropagation on a frame-by-frame schedule [41] incurs a high computational cost, which negatively affects the performance of the network, dropping its overall fraterate to accommodate the need for continuous adaptation. Various factors are involved in this matter: first, the severity of this overhead is proportional to the complexity of the network itself – the larger the number of parameters, the heavier the adaptation process becomes; second, we argue that frame-by-frame optimization is an excessive process for the adaptation itself – not only the network might need much fewer optimization steps to effectively counter domain shifts, but also such an intense adaptation definitely increases the likelihood of catastrophic forgetting over previous domains [26, 45]. In summary, a practical solution for online domain adaptation in semantic segmentation that can effectively operate in real-world environments and applications still seems to be a distant goal. + +In this paper, we propose a novel framework aimed at overcoming these issues and thus allowing for real-time, online domain adaptation: + +- We address the problem of online training by designing an automatic lightweight mechanism capable of significantly reducing back-propagation complex + +ity. We exploit the model modularity to automatically choose to train the network subset which yields the highest improvement for the allocated optimisation time. This approach reduces back-propagation FLOPS by $34\%$ while minimizing the impact on accuracy. + +- In an orthogonal fashion to the previous contribution, we introduce a lightweight domain detector. This allows us to design principled strategies to activate training only when it really matters as well as setting hyperparameters to maximize adaptation speed. Overall, these strategies increase our speed by over $5 \times$ while sacrificing less than $2.6\%$ in mIoU. +- We evaluate our method on multiple online domain adaptation benchmarks both fully synthetic [45] and semi-synthetic CityScapes domain sequences [41], showing superior accuracy and speed compared to other test-time adaptation strategies. + +Fig. 1 demonstrates the superior real-time adaptation performance of HAMLET compared to slower methods such as CoTTA [57], which experience significant drops in performance when forced to maintain a similar framerate by adapting only once every 50 frames. In contrast, HAMLET achieves an impressive 29 FPS while maintaining high accuracy. Additionally, Fig. 2 offers a glimpse of HAMLET's performance on the Increasing Storm benchmark [41], further highlighting its favorable accuracy-speed trade-off. + +# 2. Related Work + +We review the literature relevant to our work, about semantic segmentation and UDA, with particular attention to continuous and online methodologies. + +Semantic Segmentation. Very much like classification, deep learning plays a fundamental role in semantic segmentation. Fully Convolutional Network (FCN) [36] represents the pivotal step in this field, adapting common networks by means of learned upsample operators (deconvolutions). Several works aimed at improving FCN both in terms of speed [68, 38] and accuracy [5, 6, 7], with a large body of literature focusing on the latter. Major improvements have been achieved by enlarging the receptive field [72, 66, 5, 6, 7], introducing refinement modules [14, 73, 17], exploiting boundary cues [3, 10, 46] or using attention mechanisms in different flavors [13, 31, 58, 64]. The recent spread of Transformers in computer vision [11] reached semantic segmentation as well [64, 69, 65], with SegFormer [65] representing the state-of-the-art in the field and being the object of studies in the domain adaptation literature as well [20]. + +Unsupervised Domain Adaptation (UDA). This body of research aims at adapting a network trained on a source, labeled domain to a target, unlabeled one. Early approaches rely on the notion of "style" and learn how to transfer it + +across domains [74, 61, 18, 32, 12, 67]. Common strategies consist of learning domain-invariant features [15, 25], often using adversarial learning in the process [15, 55, 8, 19, 51]. A popular trend in UDA is Self-Training. These methods rely on self-supervision to learn from unlabelled data. In UDA, a successful strategy consists of leveraging target-curated pseudo-labels. Popular approaches for this purpose make use of confidence [77, 37, 76], try to balance the class predictions [75, 20], or use prototypes [2, 71, 70] to improve the quality of the pseudo-labels. Among many domain shifts, the synthetic-to-real one is the most studied, since the earliest works [74, 61, 18] to the latest [60, 30, 21, 28, 16, 40, 24]. However, this shift is one of a kind since it occurs only once after training, and without the requirement of avoiding forgetting the source domain. + +Continuous/Test-Time UDA. This family of approaches marries UDA with continuous learning, thus dealing with the catastrophic forgetting issue ignored in the synthetic-to-real case. Most continuous UDA approaches deal with it by introducing a Replay Buffer [1, 29, 27], while additional strategies make use of style transfer [62], contrastive [44, 53] or adversarial learning [63]. Despite the definition, continuous UDA often deals with offline adaptation, with well-defined target domains over which to adapt. Conceptually similar to it, is the branch of test-time adaptation, or source-free UDA, although tackling the problem in deployment rather than offline - i.e. with no access to the data from the source domain [43]. Popular strategies to deal with it consist of generating pseudo-source data to avoid forgetting [35], freezing the final layers in the model [33], aligning features [34], batch norm retraining through entropy minimization [54] or prototypes adaptation [22]. + +Online UDA. Although similar in principle to test-time adaptation, online UDA [45, 41, 52] aims to tackle multiple domain shifts, occurring unpredictably during deployment in real applications and without clear boundaries between them. On this track, the SHIFT dataset [45] provides a synthetic benchmark specifically thought for this scenario, while OASIS [52] proposes a novel protocol to evaluate UDA approaches, considering an online setting and constraining the evaluated methods to deal with frame-by-frame sequences. As for methods, OnDA [41] implements self-training as the orchestration of a static and a dynamic teacher to achieve effective online adaptation while avoiding forgetting, yet introducing massive overhead. + +Real-time performance is an essential aspect of online adaptation, particularly in applications such as autonomous driving where slow models are impractical. A slow adaptation process not only limits the practicality of real-world applications but also fails to provide high accuracy until the adaptation is complete, thereby defeating the original purpose. Therefore, accelerating the adaptation process is crucial for achieving high accuracy in real-time scenarios. + +# 3. Methods + +This section introduces HAMLET, a framework for Hardware-Aware Modular Least Expensive Training. The framework aims to solve the problem of online domain adaptation with real-time performance through several synergistic strategies. First, we introduce a Hardware-Aware Modular Training (HAMT) agent able to optimize online a trade-off between model accuracy and adaptation time. HAMT allows us to significantly reduce online training time and GFLOPS. Nevertheless, the cheapest training consists of no training at all. Therefore, as the second strategy, we introduce a formal geometric model for online domain shifts that enable reliable domain shift detection and domain estimator signals (Adaptive Domain Detection, Sec. 3.3.1). These can be easily integrated to activate the adaptation process only at specific times, as least as possible. Moreover, we can further leverage these signals by designing adaptive training policies that dynamically adapt domain-sensitive hyperparameters. We refer to these as Active Training Modulations. We present an overview of HAMLET in Fig. 3. + +# 3.1. Model Setup + +Our approach builds on the recent progress in unsupervised domain adaptation and segmentation networks. We start with DAFormer [20], a state-of-the-art UDA method, and adopt SegFormer [65] as our segmentation backbone due to its strong generalization capacity. We use three instances of the backbone, all pre-trained on the source domain: a student, a teacher, and a static (i.e. frozen) teacher. During training, the student receives a mix of target and source images [49] and is supervised with a "mixed-sample" cross-entropy loss, $\mathcal{L}_T$ (represented by green, blue and red dashed lines, in Fig. 3). This loss is computed by mixing the teacher's pseudo-labels and source annotations. To improve training stability, the teacher is updated as the exponential moving average (EMA) of the student. To further regularize the student, we use source samples stored in a replay buffer and apply two additional losses (blue lines in Fig. 3). First, we minimize the feature distance (Euclidean) between the student and the static teacher's encoder, $\mathcal{L}_{FD}$ . Then, we employ a supervised cross-entropy task loss $\mathcal{L}_S$ . Our complete objective is $\mathcal{L} = \mathcal{L}_S + \mathcal{L}_T + \lambda_{FD}\mathcal{L}_{FD}$ , with $\lambda_{FD}$ being a weight factor. During inference on the target domain, only the student is used (red lines in Fig. 3). + +# 3.2. Hardware-Aware Modular Training (HAMT) + +Online adaptation requires updating the parameters during deployment time. However, back-propagation is computationally expensive and hence too slow to be continuously applied on a deployed agent. Opting for a partial weight update, for example by finetuning the last module of the network, would enable much more efficient training time. However, domain shifts can manifest as changes + +![](images/622101d847939fac3dfcfdabded3d0f986e12affbefee80ea1b3898f02dd95a3.jpg) +Figure 3: HAMLET framework. We employ a student-teacher model with an EMA and a static teacher. HAMT orchestrates the back-propagation over the student restricting it to a network subsection. The Active Training Modulation instead controls the adaptation process by selectively enabling it only when necessary as well as tweaking sensitive training parameters. + +in both the data input distribution (such as attributes of the images, e.g. day/night) and the output distribution (e.g. class priors). This information could be encoded in different parts of the network, therefore just updating the very last segment might not suffice. This motivates the need for orchestrating the training process, to ensure sufficient training while minimizing the computational overhead. Inspired by reward-punishment [48] and reinforcement learning [56] policies, we introduce an orchestration agent in charge of deciding how deeply the network shall be finetuned through a trade-off between the pseudo-loss minimization rate and the computational time. In contrast to previous efficient back-propagation approaches [59, 23, 9], our model is pre-trained on the task and thus requires smaller updates to adapt. Let us start by modeling the problem. Our model backbone, $f$ , is composed of four different modules: $f = m_4 \circ m_3 \circ m_2 \circ m_1$ . This defines our action space $\mathcal{A} = \{\mathrm{T}_1, \mathrm{T}_2, \mathrm{T}_3, \mathrm{T}_4\}$ where $\mathrm{T}_4$ corresponds to training just the last module of the network, $m_4$ , while $\mathrm{T}_3$ the last two modules, i.e. $m_4 \circ m_3$ , $\mathrm{T}_2$ the last three, i.e. $m_4 \circ m_3 \circ m_2$ , and $\mathrm{T}_1$ the whole network $f$ . We also define a continuous state space $S = \{\mathrm{R}, \mathrm{V}\}$ where $\mathrm{R}$ is the second derivative of the EMA teacher pseudo-loss, $l_t$ , over time, hence $R_t = -\frac{\Delta^2 l}{(\Delta t)^2}$ , computed in discrete form as $R_t = -(l_t - 2l_{t-1} + l_{t-2})$ . V represents a cumulative vector with the same dimension as the action space $\mathcal{A}$ , initialized at zero. Now we have everything in place to employ an expected-improvement based decision model. At each time-step $t$ , action $T_j$ is selected for $j = \operatorname{argmax} V_t$ . During training step $t$ , $V[j]$ is updated as: + +$$ +\mathrm {V} [ j ] _ {t + 1} = \alpha R _ {t} + (1 - \alpha) \mathrm {V} [ j ] _ {t} \tag {1} +$$ + +where $\alpha$ is a smoothing factor, e.g. 0.1. i.e. $\mathrm{V_t}$ hold a discrete exponential moving average of $R_{t}$ . Therefore, our policy can be seen as a greedy module selection based on the highest expected loss improvement over its linear approximation. A notable drawback of this policy is that we will inevitably converge towards picking more rewarding, yet expensive, actions i.e. $\mathrm{T_1,T_2}$ compared to more efficient but potentially less effective actions i.e. $\mathrm{T}_3,\mathrm{T}_4$ . However, our goal is not to maximize $-\frac{\Delta^2l}{(\Delta t)^2}$ where $\Delta t$ is the number of updates, our goal is instead to maximize $-\frac{\Delta^2l}{(\Delta\tau)^2}$ where $\Delta \tau$ is a real-time interval. Therefore, we have to introduce in the optimization policy some notion of the actual training cost of each action in $\mathcal{A}$ on the target device. To start with, we measure the training time associated with each action, obtaining $\omega_{T} = \{\omega_{T_{1}},\omega_{T_{2}},\omega_{T_{3}},\omega_{T_{4}}\}$ . With this we can compute the time-conditioning vector $\gamma$ as + +$$ +\gamma_ {j} = \frac {e ^ {\frac {1}{\beta \omega_ {T _ {j}}}}}{\sum_ {k = 1} ^ {K} e ^ {\frac {1}{\beta \omega_ {T _ {k}}}}} \quad \text {f o r} j = 1, \dots , K \tag {2} +$$ + +where $\beta$ is the softmax temperature, and $K$ the number of actions, i.e. 4 in our model. We modify our update policy to favor less computationally expensive modules by scaling the updates with $\gamma$ , replacing Eq. 1 with: + +$$ +\mathrm {V} [ j ] _ {t + 1} = \left\{ \begin{array}{l l} \gamma_ {j} \alpha R _ {t} + (1 - \alpha) \mathrm {V} [ j ] _ {t} & \text {i f} R _ {t} \geq 0 \\ (1 - \gamma_ {j}) \alpha R _ {t} + (1 - \alpha) \mathrm {V} [ j ] _ {t} & \text {i f} R _ {t} < 0 \end{array} \right. \tag {3} +$$ + +This policy makes it so that more expensive actions receive smaller rewards and larger punishments. Despite its simplicity, this leads to a significant reduction in FLOPS for an average back-propagation $\beta$ , i.e. $-30\%$ with $\beta = 2.75$ + +or $-43\%$ with $\beta = 1$ . We finally choose $\beta = 1.75$ to obtain a FLOPS reduction of $-34\%$ . Exhaustive ablations on HAMT are presented in the supplementary material. + +# 3.3. Active Training Modulation + +Continuous and test-time adaptation methods tackle online learning as a continuous and constant process carried out on the data stream. Nevertheless, this approach presents several shortcomings when it comes to real-world deployments. Performing adaptation when the deployment domain is unchanged does not lead to further performance improvements on the current domain; instead, it might cause significant forgetting on previous domains, hence hindering model generalization (we present evidence of this in the supplementary material). Even if mitigated by HAMT, online training remains a computationally expensive procedure, also due to several teachers' necessary forward passes. However, knowing when and what kind of adaptation is needed is not a trivial task. We tackle this by introducing an Adaptive Domain Detection mechanism, in Sec. 3.3.1, and then a set of strategies to reduce the training time while optimizing the learning rate accordingly, in Sec. 3.3.2. + +# 3.3.1 Adaptive Domain Detection + +A key element of an online adaptation system consists of acquiring awareness of the trajectory in the data distribution space, i.e. domains, traveled by the student model during deployment. We can model the problem by setting the trajectory origin in the source domain. With high dimensional data, the data distribution is not tractable, therefore the trajectory cannot be described in closed form. Recent work [41] introduced the notion of distance between the current deployed domain and source by approximating it with the confidence drop of a source pre-trained model. This approach heavily relies on the assumption that the pretrained model is well-calibrated. While this might hold for domains close to source, the calibration quickly degrades in farther domains [45, 41]. This myopic behavior dampen the simple use of confidence for domain detection. Furthermore, the additional forward pass increases the computational cost during deployment. We tackle these limitations with an equivalently simple, yet more robust, approach. We modify the backbone of the static teacher $f^{\mathrm{st}}$ used for the feature distance loss $\mathcal{L}_{FD}$ by connecting a lightweight segmentation head, $d_1^{\mathrm{st}}$ , after the first encoder module $m_1^{\mathrm{st}}$ : $h_1^{\mathrm{st}} = d_1^{\mathrm{st}} \circ m_1^{\mathrm{st}}$ . This additional decoder, $h_1^{\mathrm{st}}$ , is trained offline, on source data, without propagating gradients in the backbone ( $m_1^{\mathrm{st}}$ is frozen). Given a target sample $x_T$ , we propose to compute the cross-entropy between the one-hot encoded student prediction $p(x_T) = 1_{\arg \max(f(x_T))}$ and the lightweight decoder prediction $g(x_T) = h_1^{\mathrm{st}}(x_T)$ as + +$$ +H _ {T} ^ {(i)} = - \sum_ {p = 1} ^ {H \times W} \sum_ {c = 1} ^ {C} p \left(x _ {T} ^ {(i)}\right) \log g \left(x _ {T} ^ {(i)}\right) \Big | _ {p, c} \tag {4} +$$ + +Thanks to the student model's higher generalization capability (both due to a larger number of parameters and the unsupervised adaptation process), it will always outperform the lightweight decoder head. Nevertheless, since now the distance is measured in the prediction space, we are not subjected to model miscalibration. Furthermore, since the student model is in constant adaptation, the domain distance accuracy actually improves over time, leading to better results. We present evidence of these claims in the supplementary material. We now define a denoised signal by using bin-averaging $A_T^{(i)} = \sum_{j=m}^{m(i+1)-1} \frac{H_T^{(j)}}{m}$ where $m$ is the bin size. Domains are modeled as discrete steps of $A_T^{(i)}$ + +$$ +B _ {0} = A _ {0} \quad B _ {i} = \left\{ \begin{array}{l l} A _ {i} & \text {i f} | B _ {i - 1} - A _ {i} | > z \\ B _ {i - 1} & \text {o t h e r w i s e} \end{array} \right. \tag {5} +$$ + +where $B$ is the discretized signal and $z$ is the minimum distance used to identify new domains. Finally, we refer to the signed amplitude of domain shifts as $\Delta B_{i} = B_{i} - B_{i - 1}$ and a domain change is detected whenever $|\Delta B_i| > z$ . + +# 3.3.2 Least Training and Adaptive Learning Rate + +The definitions of $B$ allow us to customize the training process. To this end, we adopt a Least Training (LT) strategy and trigger adaptation only when facing a new domain, which occurs when $|\Delta B_i| > z$ . Effective online learning performance depends heavily on the choice of hyperparameters such as the learning rate $\eta$ and learning rate decay rate. Therefore, we can adjust these parameters to facilitate adaptation according to the nature and intensity of domain shifts we encounter, we refer to this orchestration as Adaptive Learning Rate (ALR). For example, the larger the domain shift (i.e. $|\Delta B_i|$ ), the more we need to adapt to counteract its effect. This can be achieved by either running more optimization steps or using a higher learning rate. Whenever a domain shift is detected, we compute the number of adaptation iterations $L = K_l\frac{|\Delta B_i|}{z}$ , hence proportionally to the amplitude of the shift $|\Delta B_i|$ relative to the threshold $z$ . $K_l$ is a multiplicative factor representing the minimum adaptation iterations. If a new domain shift takes place before the adaptation process completes, we accumulate the required optimization steps. Then, we can play on two further parameters: $K_l$ and the learning rate schedule. We argue that proper scheduling is crucial for attaining a smoother adaptation. The learning rate, $\eta$ , is linearly decayed until the adaptation is concluded – the smaller the domain shift, the faster the decay. While the initial learning rate, $K_{\eta}$ , should + +
HAMTLTALRDCMRCS200mm (mIoU)All-domains (mIoU)FPSAverage GFLOPSAdaptation GFLOPS
TotalFwd.Bwd.Fwd.Bwd.
(A)-----62.2 ± 0.969.5 ± 0.35.9 ± 0.0125.2 ± 0.094.4 ± 0.030.8 ± 0.056.6 ± 0.030.8 ± 0.0
(B)✓----60.2 ± 0.568.7 ± 0.37.0 ± 0.1114.7 ± 0.094.4 ± 0.020.3 ± 0.056.6 ± 0.020.3 ± 0.0
(C)✓✓---51.8 ± 0.565.7 ± 0.229.5 ± 0.644.4 ± 0.542.6 ± 0.41.8 ± 0.256.6 ± 0.020.2 ± 0.2
(D)✓✓✓--54.1 ± 1.265.9 ± 0.229.5 ± 0.544.4 ± 0.342.7 ± 0.21.8 ± 0.156.6 ± 0.020.3 ± 0.1
(E)✓✓✓✓-56.6 ± 0.866.3 ± 0.128.9 ± 0.344.7 ± 0.242.9 ± 0.21.8 ± 0.156.6 ± 0.020.2 ± 0.0
(F)✓✓✓-✓55.8 ± 1.066.3 ± 0.229.1 ± 1.145.2 ± 0.143.2 ± 0.12.0 ± 0.056.6 ± 0.020.3 ± 0.0
(G)✓✓✓✓✓58.2 ± 0.866.9 ± 0.329.7 ± 0.645.7 ± 0.343.6 ± 0.22.1 ± 0.156.6 ± 0.020.2 ± 0.1
(a)
clear 1200mmclear 2100mmclear 375mmclear 4clear h-meantarget h-meantotal h-meanFPSGFLOPS
(A)72.952.273.664.273.067.673.473.260.667.25.6125.2
(B)73.050.473.462.173.067.373.273.159.166.46.8114.7
(C)73.446.073.561.573.666.173.873.656.565.17.2100.0
(G)73.453.673.165.273.568.273.273.361.667.89.182.2
(b)
+ +Table 1: Ablation studies - HAMLET components. Top: Increasing Storm (8925 frames per domain) [41], bottom: Fast Storm C [41] (2975 frames per domain). For each configuration, we report mIoU, framerate, and GFLOPS. + +be higher when the domain shift is triggered in domains farther from the source + +$$ +K _ {\eta} = K _ {\eta , \min } + \frac {\left(B _ {\mathrm {i}} - B _ {\text {s o u r c e}}\right) \left(K _ {\eta , \max } - K _ {\eta , \min }\right)}{B _ {\text {h a r d}} - B _ {\text {s o u r c e}}} \tag {6} +$$ + +where $B_{\mathrm{source}}$ (resp. $B_{\mathrm{hard}}$ ) is an estimate of $B$ when the network is close to (resp. far from) the source domain; and $K_{\eta, \min}$ (resp. $K_{\eta, \max}$ ) is the value of $K_{\eta}$ assigned when the network is close to (resp. far away from) the source. Concerning $K_{l}$ , we posit that moving towards the source requires less adaptation than going towards harder domains: the model shows good recalling of previously explored domains and thanks to the employed regularization strategies + +$$ +K _ {l} = \left\{ \begin{array}{l l} K _ {l, \max } & \text {i f} \Delta B _ {i} \geq 0 \\ K _ {1, \min } + \frac {\left(B _ {\mathrm {i}} - B _ {\text {s o u r c e}}\right) \left(K _ {1 , \max } - K _ {1 , \min }\right)}{B _ {\text {h a r d}} - B _ {\text {s o u r c e}}} & \text {o t h e r w i s e} \end{array} \right. \tag {7} +$$ + +where $K_{l,\min}$ (resp. $K_{l,\max}$ ) is the value of $K_{l}$ assigned when the model is close to (resp. far away from) the source domain. Extensive ablations in the supplementary material will highlight how the orchestration of the adaptation hyperparameters improves the accuracy-speed trade-off. + +# 3.3.3 Dynamic ClassMix (DCM) + +ClassMix [39] provides a simple mechanism for data augmentation by mixing classes from the source dataset into target images. Usually $50\%$ of the classes in the source dataset are selected, however we notice that this percentage is a highly sensitive hyperparameter in online domain adaptation. Injecting a significant portion of source classes has a beneficial impact when adapting to domains closer to the source domain, whereas when adapting to domains further from the source the opposite effect can be observed, as it effectively slows down the adaptation process. We therefore exploit once more the deployment domain awareness to control the mixing augmentation: + +$$ +K _ {\mathrm {C M}} = K _ {\mathrm {C M}, \min } + \frac {\left(B _ {\mathrm {i}} - B _ {\text {s o u r c e}}\right) \left(K _ {\mathrm {C M} , \max } - K _ {\mathrm {C M} , \min }\right)}{B _ {\text {h a r d}} - B _ {\text {s o u r c e}}}. \tag {8} +$$ + +where $K_{\mathrm{CM}}$ is the percentage of source classes used during adaptation; and $K_{\mathrm{CM,min}}$ (resp. $K_{\mathrm{CM,max}}$ ) is the value of $K_{\mathrm{CM}}$ assigned when the network is close to (resp. far away from) the source domain. + +# 3.3.4 Buffer Sampling + +Following [41], to simulate real deployment, we limit our access to the source domain by using a replay buffer. Additionally, instead of initializing at random (with a uniform prior), we apply Rare Class Sampling (RCS) (skewed priors) as in [20]. This incentives a more balanced class distribution over the buffer, ultimately leading to better accuracy. + +# 4. Experimental Results + +The experiments are carried out on (a) the OnDA benchmarks [41] and (b) the SHIFT dataset [45]. (a) is a semi-syntheptic benchmark, as it applies synthetic rain and fog [50] over 4 different intensities profiles. The main benchmark, Increasing Storm, presents a storm with a pyramidal intensity profile; see Fig. 4. In contrast, (b) is a purely synthetic dataset, where both the underlying image and the weather are synthetically generated and thus domain change is fully controllable. All models are evaluated using mIoU: following [41], we report the harmonic mean over domains to present the overall adaptation performance. All experiments were carried out using an Nvidia™ RTX 3090 GPU. We refer to supplementary material for further details. + +# 4.1. Ablation Studies + +In Tab. 1 we study the impact of each contribution to adaptation performance, both in terms of accuracy and efficiency. For each configuration, we report mIoU over different portions of the sequence, the framerate and the amount of GFLOPS - respectively averages of: total, forward and backward passes, and dedicated adaptation only, also divided in forward (Fwd) and backward (Bwd). Tab. 1 (a) shows results on the Increasing Storm scenario [41]. Here, we show mIoU over the $200\mathrm{mm}$ domain, i.e. the hardest in the sequence, as well as the mIoU averaged over forward and backward adaptation, i.e., from clear to $200\mathrm{mm}$ rain + +
clear25mm50mm75mm100mm200mmh-meanFPSGFLOPS
FBFBFBFBFBFBFBT
(A)DeepLabV2 (no adaptation)64.5-57.1-48.7-41.5-34.4-18.537.3--39.4-
(B)DeepLabV2 fully supervised (oracle)64.5-64.1-63.7-63.0-62.4-58.262.6--39.4-
(C)OnDA64.564.860.457.157.354.554.852.252.049.142.254.255.1-1.3-
(D)SegFormer MiT-B1 (no adaptation)73.4-68.8-64.2-58.0-51.8-31.257.8--48.434.9
(E)SegFormer MiT-B5 (no adaptation)77.6-73.9-71.0-67.2-62.6-46.764.7--11.5240.4
(F)SegFormer MiT-B1 fully supervised (oracle)72.9-72.4-72.1-71.5-70.7-68.671.3--48.434.9
(G)TENT73.072.868.568.664.564.859.760.254.554.835.956.263.659.910.0-
(H)TENT + Replay Buffer73.072.868.568.664.564.859.760.254.454.735.856.163.659.97.8-
(I)CoTTA72.574.469.570.965.968.266.164.764.663.557.265.668.166.80.6593.8
(J)CoTTA real-time73.375.470.370.666.966.462.561.457.656.939.759.265.562.327.041.7
(K)HAMLET (ours)73.471.070.168.867.767.566.666.465.564.659.266.867.667.229.145.7
+ +Table 2: Comparison against other models – Increasing storm scenario. (A-C) methods built over DeepLabv2, (D-E) SegFormer variants trained on source, (F) oracle, (G-K) models adapted online. We report mIoU, framerate, and GFLOPS. + +and backward. Results are averaged over 3 runs with different seeds, with standard deviation being reported. (A) reports the results achieved by naively performing full adaptation of the model. HAMT can increase the framerate by roughly $15\%$ by reducing the Bwd GFLOPS of $34\%$ , at the expense of as few as $0.7\mathrm{mIoU}$ on average, i.e., about 2 points on the $200\mathrm{mm}$ domain. The main boost in terms of speed is obviously given by LT (C), which inhibits the training in absence of detected domain shifts. LT increases the framerate by approximately $4\times$ by decimating the total GFLOPS, yet not affecting the adaptation Bwd GFLOPS. This comes with a price in terms of mIoU, dropping by about 4 points on average and more than 10 points on $200\mathrm{mm}$ - not a moderate drop anymore. LT impact highly depends on the domain sequence experienced during deployment: frequent domain changes could prevent training inhibition, thus neglecting LT gains in terms of efficiency, as we will appreciate later. The loss in accuracy is progressively regained by adding ALR (D), with further improvements yielded by one between DCM (E) and RCS (F), or both together (G) leading to the full HAMLET configuration. The three together allow for reducing the gap to 2.5 points mIoU - 4 over the $200\mathrm{mm}$ domain - without sacrificing any efficiency. Tab. 1 (b) shows further results, on a faster version of Storm C [41]. This represents a much more challenging scenario, with harsher and $3\times$ more frequent domain shifts. Here we show the single domains mIoU, as well as harmonic mean on source and target domains, and all frames. As expected, in this benchmark, LT alone (C) results much less effective than before, with a much lower gain in FPS and GFLOPS. Here, the synergy between the HAMT, LT, and the other components (G) allows for the best accuracy and speedup - even outperforming the full training variant (A) - highlighting their complementarity. Further ablations are in the supplementary material. + +# 4.2. Results on Increasing Storm + +Tab. 2 shows a direct comparison between HAMLET and relevant approaches. The presented test-time adaptation + +![](images/1f4433bc49d15129f594d5c8ec125e604e8568ee8e10b04e24fd6d40576a4f56.jpg) + +![](images/d24ecd2323e20c9b68c67191586e436a765e6a0dbea0a39967845645c400b4bb.jpg) + +![](images/79e0895ecd0d4a25ec03c1c4836f78786d7d64dc4e4bb9e31e46c9d83370760d.jpg) + +![](images/fda9ff3d1862761f0f33b02a91c509e2a6e7eae557fe386105aad3905486660b.jpg) +Figure 4: HAMLET on the Increasing Storm. We show rain intensity (in millimetres), mIoU over active (bold) and inactive (dashed) domains, learning rate and FPS. + +strategies namely - TENT and CoTTA - were revised to handle the online setting and be fairly compared with HAMLET. All methods start with the same exact initial weights - with HAMLET requiring the additional lightweight decoder, not needed by TENT and CoTTA - using SegFormer MiT-B1 as the backbone, since it is $4 \times$ faster than SegFormer MiT-B5 and thus better suited to keep real-time performance even during adaptation. We report results achieved by DeepLabv2 trained on source data only (A), an oracle model trained with full supervision (B), as well as OnDA [41] (C) as a reference. Then, we report SegFormer models trained on the source domain only (D) and (E). In (F) we show the performance achieved by an oracle SegFormer, trained on all domains fully supervised. Following [41], columns "F" concern forward adaptation from clear to $200\mathrm{mm}$ , while columns "B" show backward adap + +
clear750m375m150m75mh-meanFPSGFLOPS
FBFBFBFBFBFBT
OnDA64.965.863.362.360.758.851.649.142.155.154.1-1.3-
SegFormer MiT-B1 (no adaptation)71.1-70.0-67.5-58.8-46.961.3--48.434.9
Full training71.572.172.974.771.973.167.668.161.368.771.970.35.6125.2
HAMLET (ours)71.171.670.370.868.869.264.364.357.065.968.967.424.850.7
+ +Table 3: Results on foggy domains. Comparison between OnDA, Source SegFormer, full training adaptation, and HAMLET. + +
ClearCloudyOvercastSmall rainMid rainHeavy rainh-meanFPSGFLOPS
FBFBFBFBFBFBFBT
SegFormer MiT-B1 fully supervised (oracle)80.1-79.9-79.8-78.9-78.7-77.179.1--48.434.93
SegFormer MiT-B1 (no adaptation)79.6-77.1-75.4-73.4-71.4-66.773.7--48.434.93
Full training78.979.376.776.876.877.974.874.876.376.574.076.277.076.65.0125.1
HAMLET (ours)79.678.976.976.676.177.473.374.374.276.074.275.776.676.126.843.9
+ +Table 4: Results on SHIFT dataset [45]. Comparison between Source SegFormer, full training adaptation, and HAMLET. + +tation from $200\mathrm{mm}$ to clear, while the h-mean T refers to the overall harmonic mean. We can notice how SegFormer results are much more robust to domain changes with respect to DeepLabv2. Indeed, SegFormer MiT-B5 (E), without any adaptation, results more accurate than DeepLabv2 oracle (B), as well as better and faster than OnDA (C). The faster variant (D) outperforms OnDA both in speed and accuracy, reaching 48 FPS. Nevertheless, domain changes still dampen the full potential of SegFormer. Indeed, the oracle (F) outperforms (D) by about $+14$ mIoU. However, this is not meaningful for real deployment experiencing unpredictable domain shifts, as it assumes to have data available in advance. Concerning test-time models, TENT starts adapting properly only beyond $50\mathrm{mm}$ , both with (G) and without (H) frame buffer, while it loses some accuracy on $25\mathrm{mm}$ . This makes its overall forward adaptation performance slightly worse compared to the pre-trained model (D), while being better at backward adaptation. Despite outperforming SegFormer MiT-B1, TENT is both slower and less accurate than SegFormer MiT-B5 running without any adaptation, further suggesting the robustness of the latter and making TENT not suitable for real-world deployment. On the contrary, CoTTA (I) outperforms both SegFormer models trained on source only, at the expense of dropping the framerate below 1FPS. It is worth mentioning that these metrics were collected after each domain was completed by each model individually. In an evaluation setup imposing a shared time frame, slower models would present much lower metrics, since their adaptation process would result constantly lagged. In fact, forcing CoTTA to run in real-time, at nearly 30FPS – i.e. by training once every 50 frames – dramatically reduces the effectiveness of the adaptation process (J), with drastic drops in the hardest domains. Finally, HAMLET (K) succeeds on any fronts, improving the baseline (D) by about 10 points with only a cost of $25\%$ in terms of speed, while outperforming SegFormer MiT-B5 (E) both on accuracy $(+2.5$ mIoU) and speed $(3\times$ faster) – being the only method achieving this, and thus the only suitable choice for real-time applications. Fig. 4 shows the overall behavior of HAMLET while adapting over the Increasing Storm. In addition to the rain intensity and the + +![](images/0b8e0c1eeba7fae64b64798ae555b3b1cccde62d7dee9dc25e223452ccce2c5d.jpg) +Figure 5: HAMLET on the SHIFT benchmark. We show mIoU over active (bold) and inactive (dashed) domains, learning rate and FPS. + +mIoU achieved on each domain - active (bold) or inactive (dashed), i.e. respectively the mIoU on the domain being currently faced during deployment, and how the current adaptation affects the performance on the other domains to highlight the robustness to forgetting - we also report how the learning rate is modulated in correspondence of detected domain shifts, with a consequent drop in FPS due to the short training process taking place. For further experiments on harsher and sudden adaptation cycles, we include results of Storms A, B, C [41] in the supplementary material. + +# 4.3. Additional Results: Fog and SHIFT + +Fog. In Tab. 3, we investigate adaptation on the Increasing Fog scenario in the OnDA benchmark [41]. Crucially, for this experiment, we keep the same hyperparameters used for the Increasing Storm, since in both cases the starting SegFormer model is trained on the same source domain. This allows for validating how the proposed setting general- + +![](images/754b478f08cbb8f9ef25a77d64555c04cb1d29a05f6758abea85409722214ff3.jpg) +Figure 6: Qualitative results - HAMLET in action. From left to right, we show frames from clean, $50\mathrm{mm}$ , $100\mathrm{mm}$ , and $200\mathrm{m}$ domains. From top to bottom: input image, prediction by SegFormer trained on source domain and HAMLET. + +izes at dealing with different kind of domain shifts, beyond those considered in the main experiments. We effectively use Increasing Fog as test set, and compare against Seg-Former trained on source (no adaptation) and a model that has been adapted by means of full online training optimization (configuration (A) of Table 1). HAMLET is able to adapt almost as well as the full online training model, with less than a $3\mathrm{mIoU}$ gap, while enjoying real-time adaptation at nearly $5\times$ the speed using just $40\%$ of the FLOPS. + +SHIFT. We further test HAMLET on the SHIFT dataset [45]. Tab. 4 collects the results achieved by Seg-Former trained on source, full online training and HAMLET respectively, both at forward and backward adaptation across Clear, Cloudy, Overcast, Small rain, Mid rain and Heavy rain domains. Here HAMLET results highly competitive with the full training regime, with only 0.5 drop in average mIoU, while being more than $5 \times$ faster. Fig. 5 depicts, from top to bottom, the rain intensity characterizing any domain encountered on SHIFT, the mIoU achieved both on current (bold) and inactive (dashed) domains, the learning rate changes based on the domain shift detection, and the framerate achieved at any step. We refer to the supplementary material for a deeper analysis. + +Qualitative results. To conclude, Fig. 6 shows some qualitative examples from CityScapes. We can notice how SegFormer accuracy (second tow) drops with severe rain, whereas HAMLET (third row) is capable of keeping the same segmentation quality across the storm. + +# 5. Discussion + +Orthogonality. HAMT and LT act independently. Indeed, by strongly constraining the adaptation periods through LT, HAMT has a limited margin of action. The impact of HAMT also depends on the backbone and by care + +fully crafting modular architectures, one can achieve further optimization. Nevertheless, in a deployment environment where domain shifts occur at high frequencies (e.g., Storm C), LT is ineffective, while HAMT thrives. + +Measuring forgetting. An interesting topic we have not investigated consists of introducing an explicit awareness of which domains have been explored and how well we can recall them, expanding the distance $B$ to multiple dimensions. + +Safety. We believe dynamic adaptation has the potential to enhance safety, but we acknowledge the necessity for rigorous testing and verification to safeguard against drift or catastrophic forgetting. This mandates a comprehensive effort from academia, industry, and certification authorities for ensuring the integrity of dynamically adapting models. + +# 6. Summary & Conclusion + +We have presented HAMLET, a framework for real-time adaptation for semantic segmentation that achieves state-of-the-art performance on established benchmarks with continuous domain changes. Our approach combines a hardware-aware backpropagation orchestrator and a specialized domain-shift detector to enable active control over the model's adaptation, resulting in high framereates on a consumer-grade GPU. These advancements enable HAMLET to be a promising solution for in-the-wild deployment, making it a valuable tool for applications that require robust performance in the face of unforeseen domain changes. + +Acknowledgement. The authors thank Gianluca Villani for the insightful discussion on reward-punishment policies, Leonardo Ravaglia for his expertise on hardware-aware training, and Lorenzo Andraghetti for exceptional technical support throughout the project. Their assistance was invaluable in the completion of this work. + +# References + +[1] Andreea Bobu, Judy Hoffman, Eric Tzeng, and Trevor Darrell. Adapting to continuously shifting domains. In ICLR 2018 Workshop Program Chairs, 2018. 00000. +[2] Chaoqi Chen, Weiping Xie, Wenbing Huang, Yu Rong, Xinghao Ding, Yue Huang, Tingyang Xu, and Junzhou Huang. Progressive feature alignment for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 627-636, 2019. +[3] Liang-Chieh Chen, Jonathan T Barron, George Papandreou, Kevin Murphy, and Alan L Yuille. Semantic image segmentation with task-specific edge detection using cnns and a discriminatively trained domain transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4545-4554, 2016. +[4] Liang-Chieh Chen, Raphael Gontijo Lopes, Bowen Cheng, Maxwell D Collins, Ekin D Cubuk, Barret Zoph, Hartwig Adam, and Jonathon Shlens. Naive-student: Leveraging semi-supervised learning in video sequences for urban scene segmentation. In European Conference on Computer Vision, pages 695-714. Springer, 2020. +[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. +[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(4):834-848, Apr 2018. +[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. +[8] Yi-Hsin Chen, Wei-Yu Chen, Yu-Ting Chen, Bo-Cheng Tsai, Yu-Chiang Frank Wang, and Min Sun. No more discrimination: Cross city adaptation of road scene segmenters. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2011-2020. IEEE, 2017. 00000. +[9] Feng Cheng, Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Li, and Wei Xia. Stochastic backpropagation: A memory efficient strategy for training video models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8301-8310, 2022. +[10] Henghui Ding, Xudong Jiang, Ai Qun Liu, Nadia Magnenat Thalmann, and Gang Wang. Boundary-aware feature propagation for scene segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6819-6829, 2019. +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans + +formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[12] A. Dundar, M. Y. Liu, Z. Yu, T. C. Wang, J. Zedlewski, and J. Kautz. Domain stylization: A fast covariance matching framework towards domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2020. +[13] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3146-3154, 2019. +[14] Jun Fu, Jing Liu, Yuhang Wang, Yong Li, Yongjun Bao, Jinhui Tang, and Hanqing Lu. Adaptive context network for scene parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6748-6757, 2019. +[15] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, Jan. 2016. +[16] Rui Gong, Martin Danelljan, Dengxin Dai, Danda Pani Paudel, Ajad Chhatkuli, Fisher Yu, and Luc Van Gool. Tacs: Taxonomy adaptive cross-domain semantic segmentation. In European Conference on Computer Vision, pages 19-35. Springer, 2022. +[17] Junjun He, Zhongying Deng, Lei Zhou, Yali Wang, and Yu Qiao. Adaptive pyramid context network for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7519-7528, 2019. +[18] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. CyCADA: Cycle-consistent adversarial domain adaptation. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 1989-1998, Stockholm, Sweden, 10-15 Jul 2018. PMLR. +[19] Judy Hoffman, Dequan Wang, Fisher Yu, and Trevor Darryll. FCNs in the wild: Pixel-level adversarial and constraint-based adaptation. CoRR, 2016. 00000. +[20] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. arXiv preprint arXiv:2111.14887, 2021. +[21] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. *Hrda: Context-aware high-resolution domain-adaptive semantic segmentation*. In European Conference on Computer Vision (ECCV), 2022. +[22] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. +[23] Angela H Jiang, Daniel L-K Wong, Giulio Zhou, David G Andersen, Jeffrey Dean, Gregory R Ganger, Gauri Joshi, + +Michael Kaminsky, Michael Kozuch, Zachary C Lipton, et al. Accelerating deep learning by focusing on the biggest losers. arXiv preprint arXiv:1910.00762, 2019. +[24] Zhengkai Jiang, Yuxi Li, Ceyuan Yang, Peng Gao, Yabiao Wang, Ying Tai, and Chengjie Wang. Prototypical contrast adaptation for domain adaptive semantic segmentation. In European Conference on Computer Vision, pages 36-54. Springer, 2022. +[25] Myeongjin Kim and Hyeran Byun. Learning texture invariant representation for domain adaptation of semantic segmentation. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2020. +[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114, 12 2016. +[27] Yevhen Kuznietsov, Marc Proesmans, and Luc Van Gool. Towards unsupervised online domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 261-271, 2022. +[28] Xin Lai, Zhuotao Tian, Xiaogang Xu, Yingcong Chen, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Decouplenet: Decoupled network for domain adaptive semantic segmentation. In European Conference on Computer Vision, pages 369-387. Springer, 2022. +[29] Qicheng Lao, Xiang Jiang, Mohammad Havaei, and Yoshua Bengio. Continuous domain adaptation with variational domain-agnostic feature replay. arXiv preprint arXiv:2003.04382, 2020. +[30] Geon Lee, Chanho Eom, Wonkyung Lee, Hyekang Park, and Bumsub Ham. Bi-directional contrastive learning for domain adaptive semantic segmentation. In European Conference on Computer Vision, pages 38-55. Springer, 2022. +[31] Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention networks for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9167-9176, 2019. +[32] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2019. +[33] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. CoRR, 2020. +[34] Yuejiang Liu, Parth Kothari, Bastien Germain van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++: When does self-supervised test-time training fail or thrive? In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. +[35] Yang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. 2021. + +[36] Jonathan Long, Evan Shelhamer, and Trevor Darrel. Fully convolutional networks for semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. +[37] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. Lecture Notes in Computer Science, page 415-430, 2020. +[38] Vladimir Nekrasov, Chunhua Shen, and Ian Reid. Lightweight refinenet for real-time semantic segmentation. In British Conference on Computer Vision (BMVC), 2018. +[39] Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1369-1378, 2021. +[40] Fei Pan, Sungsu Hur, Seokju Lee, Junsik Kim, and In So Kweon. Ml-bpm: Multi-teacher learning with bidirectional photometric mixing for open compound domain adaptation in semantic segmentation. In European Conference on Computer Vision, pages 236–251. Springer, 2022. +[41] Theodoros Panagiotakopoulos, Pier Luigi Dovesi, Linus Harenstam-Nielsen, and Matteo Poggi. Online domain adaptation for semantic segmentation in ever-changing conditions. In European Conference on Computer Vision (ECCV), 2022. +[42] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In European conference on computer vision, pages 102-118. Springer, 2016. +[43] Serban Stan and Mohammad Rostami. Unsupervised model adaptation for continual semantic segmentation. In AAAI, 2021. +[44] Peng Su, Shixiang Tang, Peng Gao, Di Qiu, Ni Zhao, and Xiaogang Wang. Gradient regularized contrastive learning for continual domain adaptation. 2020. 00000. +[45] Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. SHIFT: a synthetic driving dataset for continuous multi-task domain adaptation. In Computer Vision and Pattern Recognition, 2022. +[46] Towaki Takikawa, David Acuna, Varun Jampani, and Sanja Fidler. Gated-scnn: Gated shape cnns for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5229-5238, 2019. +[47] Phillip Thomas, Lars Pandikow, Alex Kim, Michael Stanley, and James Grieve. Open synthetic dataset for improving cyclist detection, Nov 2021. +[48] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 195-204, 2019. +[49] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1379-1389, January 2021. + +[50] Maxime Tremblay, Shirsendu S. Halder, Raoul de Charette, and Jean-François Lalonde. Rain rendering for evaluating and improving robustness to bad weather. International Journal of Computer Vision, 2020. +[51] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Jun 2018. +[52] Riccardo Volpi, Pau de Jorge, Diane Larlus, and Gabriela Csurka. On the road to online adaptation for semantic image segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. +[53] Vibashan VS, Poojan Oza, and Vishal M. Patel. Towards online domain adaptive object detection. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Jan 2023. +[54] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. +[55] Haoran Wang, Tong Shen, Wei Zhang, Lingyu Duan, and Tao Mei. Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In The European Conference on Computer Vision (ECCV), August 2020. +[56] Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. Haq: Hardware-aware automated quantization with mixed precision, 2018. +[57] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation, 2022. +[58] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. +[59] Bingzhen Wei, Xu Sun, Xuancheng Ren, and Jingjing Xu. Minimal effort back propagation for convolutional neural networks. arXiv preprint arXiv:1709.05804, 2017. +[60] Tsung-Han Wu, Yi-Syuan Liou, Shao-Ji Yuan, Hsin-Ying Lee, Tung-I Chen, Kuan-Chih Huang, and Winston H Hsu. D2ada: Dynamic density-aware active domain adaptation for semantic segmentation. In European Conference on Computer Vision (ECCV), 2022. +[61] Zuxuan Wu, Xintong Han, Yen-Liang Lin, Mustafa Gokhan Uzunbas, Tom Goldstein, Ser Nam Lim, and Larry S. Davis. Dcan: Dual channel-wise alignment networks for unsupervised scene adaptation. Lecture Notes in Computer Science, page 535-552, 2018. +[62] Zuxuan Wu, Xin Wang, Joseph Gonzalez, Tom Goldstein, and Larry Davis. ACE: Adapting to changing environments for semantic segmentation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2121-2130. IEEE, 2019. +[63] Markus Wulfmeier, Alex Bewley, and Ingmar Posner. Incremental adversarial domain adaptation for continually changing environments. 2018. 00000. +[64] Enze Xie, Wenjia Wang, Wenhai Wang, Peize Sun, Hang Xu, Ding Liang, and Ping Luo. Segmenting transparent objects in the wild with transformer. In IJCAI, 2021. + +[65] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. +[66] Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic segmentation in street scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3684-3692, 2018. +[67] Yanchao Yang and Stefano Soatto. FDA: Fourier domain adaptation for semantic segmentation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4084-4094. IEEE, 2020. +[68] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 325-341, 2018. +[69] Yuhui Yuan, Xiaokang Chen, Xilin Chen, and Jingdong Wang. Segmentation transformer: Object-contextual representations for semantic segmentation. In European Conference on Computer Vision (ECCV), 2020. +[70] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. 2021. +[71] Qiming Zhang, Jing Zhang, Wei Liu, and Dacheng Tao. Category anchor-guided unsupervised domain adaptation for semantic segmentation. Advances in Neural Information Processing Systems, 2019. +[72] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. +[73] Yizhou Zhou, Xiaoyan Sun, Zheng-Jun Zha, and Wenjun Zeng. Context-reinforced semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4046-4055, 2019. +[74] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), Oct 2017. +[75] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. +[76] Yang Zou, Zhiding Yu, Xiaofeng Liu, BVK Kumar, and Jinsong Wang. Confidence regularized self-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5982-5991, 2019. +[77] Yang Zou, Zhiding Yu, Xiaofeng Liu, B. V. K. Vijaya Kumar, and Jinsong Wang. Confidence regularized self-training. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019. \ No newline at end of file diff --git a/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/images.zip b/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..99cca50303469604b7a91b806f6b614d95572bde --- /dev/null +++ b/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e34eb479bb946e1c15f96bda6bb7f1930b19745257f8a09562845f57b4bfdb8f +size 624202 diff --git a/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/layout.json b/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9516551ee907529767a0812ed6b9e17f32e15f13 --- /dev/null +++ b/2023/To Adapt or Not to Adapt_ Real-Time Adaptation for Semantic Segmentation/layout.json @@ -0,0 +1,9436 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 213, + 103, + 381, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 103, + 381, + 121 + ], + "spans": [ + { + "bbox": [ + 213, + 103, + 381, + 121 + ], + "type": "text", + "content": "To Adapt or Not to Adapt?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 122, + 452, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 122, + 452, + 140 + ], + "spans": [ + { + "bbox": [ + 142, + 122, + 452, + 140 + ], + "type": "text", + "content": "Real-Time Adaptation for Semantic Segmentation" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 159, + 536, + 254 + ], + "blocks": [ + { + "bbox": [ + 58, + 159, + 536, + 254 + ], + "lines": [ + { + "bbox": [ + 58, + 159, + 536, + 254 + ], + "spans": [ + { + "bbox": [ + 58, + 159, + 536, + 254 + ], + "type": "image", + "image_path": "94a51354b0137b899c5ed9b46e74d97ea543d6355691a13a73f671aadecada51.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 267, + 545, + 335 + ], + "blocks": [ + { + "bbox": [ + 48, + 267, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 48, + 267, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 267, + 545, + 335 + ], + "type": "image", + "image_path": "9b531da2a21f0adb6eccf38eac398688d2049a0530162967e18a2c900ed1113b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 346, + 546, + 371 + ], + "lines": [ + { + "bbox": [ + 45, + 346, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 45, + 346, + 546, + 371 + ], + "type": "text", + "content": "Figure 1. Real-time adaptation with HAMLET. Online adaptation to continuous and unforeseeable domain shifts is hard and computationally expensive. HAMLET can deal with it at almost 30FPS outperforming much slower online methods - e.g. OnDA and CoTTA." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 383, + 192, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 383, + 192, + 396 + ], + "spans": [ + { + "bbox": [ + 143, + 383, + 192, + 396 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 412, + 290, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 412, + 290, + 617 + ], + "spans": [ + { + "bbox": [ + 45, + 412, + 290, + 617 + ], + "type": "text", + "content": "The goal of Online Domain Adaptation for semantic segmentation is to handle unforeseeable domain changes that occur during deployment, like sudden weather events. However, the high computational costs associated with brute-force adaptation make this paradigm unfeasible for real-world applications. In this paper we propose HAMLET, a Hardware-Aware Modular Least Expensive Training framework for real-time domain adaptation. Our approach includes a hardware-aware back-propagation orchestration agent (HAMT) and a dedicated domain-shift detector that enables active control over when and how the model is adapted (LT). Thanks to these advancements, our approach is capable of performing semantic segmentation while simultaneously adapting at more than 29FPS on a single consumer-grade GPU. Our framework's encouraging accuracy and speed trade-off is demonstrated on OnDA and SHIFT benchmarks through experimental results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 652, + 128, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 652, + 128, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 652, + 128, + 665 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Semantic segmentation aims at classifying an image at a pixel level, based on the local and global context, to enable a higher level of understanding of the depicted" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 340, + 391, + 534, + 510 + ], + "blocks": [ + { + "bbox": [ + 340, + 391, + 534, + 510 + ], + "lines": [ + { + "bbox": [ + 340, + 391, + 534, + 510 + ], + "spans": [ + { + "bbox": [ + 340, + 391, + 534, + 510 + ], + "type": "image", + "image_path": "4bc99e80e8f50d2a7405896a09971660fe857c9fdcbe03c4b1d93eb72c00b6c4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 514, + 547, + 549 + ], + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 549 + ], + "type": "text", + "content": "Figure 2: Online adaptation methods on the Increasing Storm. We plot mIoUs achieved on single domains. Colors from colder to warmer encode slower to faster methods." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 572, + 546, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 546, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 546, + 655 + ], + "type": "text", + "content": "scene. In recent years, deep learning has become the dominant paradigm to tackle this task effectively employing CNNs [5, 69, 4] or, more recently, transformers [65], at the expense of requiring large quantities of annotated images for training. Specifically, annotating for this task needs per-pixel labeling, which is an expensive and time-consuming task, severely limiting the availability of training data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 656, + 547, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 656, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 656, + 547, + 693 + ], + "type": "text", + "content": "The use of simulations and graphics engines [42] to generate annotated frames enabled a marked decrease in the time and cost necessary to gather labeled data thanks" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 320, + 702, + 541, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 702, + 541, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 702, + 541, + 713 + ], + "type": "text", + "content": "* Joint first authorship † Part of the work done while at Univres" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16548" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 633 + ], + "type": "text", + "content": "to the availability of the ground truth. However, despite the increasing quality in data realism [47], there is a substantial difference between simulated data generated by graphics engines and real-world images, such that leveraging these data for real-world applications requires adapting over a significant domain shift. The promise of unlocking this cheap and plentiful source of training data has provided a major impulse behind the development of a large body of work on Unsupervised Domain Adaptation (UDA) techniques [74, 61, 18, 15, 55], consisting of training semantic segmentation networks on labelled synthetic frames – the source domain – and then adapting the network to operate on real images, representing the target domain, without requiring human annotation. However, the synthetic-to-real shift represents only one of many possible domain transitions; specifically, when dealing with real-world deployment, domain shifts can occur from various causes, from different camera placements to different lighting, weather conditions, urban scenario, or any possible combination of the above. Because of the combinatorial nature of the problem, it is simply impossible to evenly represent all possible deployment domains in a dataset. This curse of dimensionality prevents having generalized robust performances [41, 45]. However, the recent advent of online domain adaptation [41] potentially allows us to face continuous and unpredictable domain shifts at deployment time, without requiring data associated with such domain shifts beforehand. Nonetheless, despite its potential, several severe limitations still hamper the online adaptation paradigm. In particular, continuously performing backpropagation on a frame-by-frame schedule [41] incurs a high computational cost, which negatively affects the performance of the network, dropping its overall fraterate to accommodate the need for continuous adaptation. Various factors are involved in this matter: first, the severity of this overhead is proportional to the complexity of the network itself – the larger the number of parameters, the heavier the adaptation process becomes; second, we argue that frame-by-frame optimization is an excessive process for the adaptation itself – not only the network might need much fewer optimization steps to effectively counter domain shifts, but also such an intense adaptation definitely increases the likelihood of catastrophic forgetting over previous domains [26, 45]. In summary, a practical solution for online domain adaptation in semantic segmentation that can effectively operate in real-world environments and applications still seems to be a distant goal." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 634, + 287, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 287, + 672 + ], + "type": "text", + "content": "In this paper, we propose a novel framework aimed at overcoming these issues and thus allowing for real-time, online domain adaptation:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 287, + 715 + ], + "type": "text", + "content": "- We address the problem of online training by designing an automatic lightweight mechanism capable of significantly reducing back-propagation complex" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 324, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 133 + ], + "type": "text", + "content": "ity. We exploit the model modularity to automatically choose to train the network subset which yields the highest improvement for the allocated optimisation time. This approach reduces back-propagation FLOPS by " + }, + { + "bbox": [ + 324, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "34\\%" + }, + { + "bbox": [ + 324, + 72, + 545, + 133 + ], + "type": "text", + "content": " while minimizing the impact on accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 140, + 545, + 292 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 317, + 140, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 545, + 224 + ], + "type": "text", + "content": "- In an orthogonal fashion to the previous contribution, we introduce a lightweight domain detector. This allows us to design principled strategies to activate training only when it really matters as well as setting hyperparameters to maximize adaptation speed. Overall, these strategies increase our speed by over " + }, + { + "bbox": [ + 317, + 140, + 545, + 224 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 317, + 140, + 545, + 224 + ], + "type": "text", + "content": " while sacrificing less than " + }, + { + "bbox": [ + 317, + 140, + 545, + 224 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 317, + 140, + 545, + 224 + ], + "type": "text", + "content": " in mIoU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 232, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 545, + 292 + ], + "type": "text", + "content": "- We evaluate our method on multiple online domain adaptation benchmarks both fully synthetic [45] and semi-synthetic CityScapes domain sequences [41], showing superior accuracy and speed compared to other test-time adaptation strategies." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 300, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 408 + ], + "type": "text", + "content": "Fig. 1 demonstrates the superior real-time adaptation performance of HAMLET compared to slower methods such as CoTTA [57], which experience significant drops in performance when forced to maintain a similar framerate by adapting only once every 50 frames. In contrast, HAMLET achieves an impressive 29 FPS while maintaining high accuracy. Additionally, Fig. 2 offers a glimpse of HAMLET's performance on the Increasing Storm benchmark [41], further highlighting its favorable accuracy-speed trade-off." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 418, + 392, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 392, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 392, + 430 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 438, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 438, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 545, + 474 + ], + "type": "text", + "content": "We review the literature relevant to our work, about semantic segmentation and UDA, with particular attention to continuous and online methodologies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 475, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 546, + 665 + ], + "type": "text", + "content": "Semantic Segmentation. Very much like classification, deep learning plays a fundamental role in semantic segmentation. Fully Convolutional Network (FCN) [36] represents the pivotal step in this field, adapting common networks by means of learned upsample operators (deconvolutions). Several works aimed at improving FCN both in terms of speed [68, 38] and accuracy [5, 6, 7], with a large body of literature focusing on the latter. Major improvements have been achieved by enlarging the receptive field [72, 66, 5, 6, 7], introducing refinement modules [14, 73, 17], exploiting boundary cues [3, 10, 46] or using attention mechanisms in different flavors [13, 31, 58, 64]. The recent spread of Transformers in computer vision [11] reached semantic segmentation as well [64, 69, 65], with SegFormer [65] representing the state-of-the-art in the field and being the object of studies in the domain adaptation literature as well [20]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Unsupervised Domain Adaptation (UDA). This body of research aims at adapting a network trained on a source, labeled domain to a target, unlabeled one. Early approaches rely on the notion of \"style\" and learn how to transfer it" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16549" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "across domains [74, 61, 18, 32, 12, 67]. Common strategies consist of learning domain-invariant features [15, 25], often using adversarial learning in the process [15, 55, 8, 19, 51]. A popular trend in UDA is Self-Training. These methods rely on self-supervision to learn from unlabelled data. In UDA, a successful strategy consists of leveraging target-curated pseudo-labels. Popular approaches for this purpose make use of confidence [77, 37, 76], try to balance the class predictions [75, 20], or use prototypes [2, 71, 70] to improve the quality of the pseudo-labels. Among many domain shifts, the synthetic-to-real one is the most studied, since the earliest works [74, 61, 18] to the latest [60, 30, 21, 28, 16, 40, 24]. However, this shift is one of a kind since it occurs only once after training, and without the requirement of avoiding forgetting the source domain." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 254, + 289, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 289, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 289, + 458 + ], + "type": "text", + "content": "Continuous/Test-Time UDA. This family of approaches marries UDA with continuous learning, thus dealing with the catastrophic forgetting issue ignored in the synthetic-to-real case. Most continuous UDA approaches deal with it by introducing a Replay Buffer [1, 29, 27], while additional strategies make use of style transfer [62], contrastive [44, 53] or adversarial learning [63]. Despite the definition, continuous UDA often deals with offline adaptation, with well-defined target domains over which to adapt. Conceptually similar to it, is the branch of test-time adaptation, or source-free UDA, although tackling the problem in deployment rather than offline - i.e. with no access to the data from the source domain [43]. Popular strategies to deal with it consist of generating pseudo-source data to avoid forgetting [35], freezing the final layers in the model [33], aligning features [34], batch norm retraining through entropy minimization [54] or prototypes adaptation [22]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 460, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 289, + 616 + ], + "type": "text", + "content": "Online UDA. Although similar in principle to test-time adaptation, online UDA [45, 41, 52] aims to tackle multiple domain shifts, occurring unpredictably during deployment in real applications and without clear boundaries between them. On this track, the SHIFT dataset [45] provides a synthetic benchmark specifically thought for this scenario, while OASIS [52] proposes a novel protocol to evaluate UDA approaches, considering an online setting and constraining the evaluated methods to deal with frame-by-frame sequences. As for methods, OnDA [41] implements self-training as the orchestration of a static and a dynamic teacher to achieve effective online adaptation while avoiding forgetting, yet introducing massive overhead." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Real-time performance is an essential aspect of online adaptation, particularly in applications such as autonomous driving where slow models are impractical. A slow adaptation process not only limits the practicality of real-world applications but also fails to provide high accuracy until the adaptation is complete, thereby defeating the original purpose. Therefore, accelerating the adaptation process is crucial for achieving high accuracy in real-time scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 71, + 367, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 367, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 367, + 84 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 91, + 547, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 319 + ], + "type": "text", + "content": "This section introduces HAMLET, a framework for Hardware-Aware Modular Least Expensive Training. The framework aims to solve the problem of online domain adaptation with real-time performance through several synergistic strategies. First, we introduce a Hardware-Aware Modular Training (HAMT) agent able to optimize online a trade-off between model accuracy and adaptation time. HAMT allows us to significantly reduce online training time and GFLOPS. Nevertheless, the cheapest training consists of no training at all. Therefore, as the second strategy, we introduce a formal geometric model for online domain shifts that enable reliable domain shift detection and domain estimator signals (Adaptive Domain Detection, Sec. 3.3.1). These can be easily integrated to activate the adaptation process only at specific times, as least as possible. Moreover, we can further leverage these signals by designing adaptive training policies that dynamically adapt domain-sensitive hyperparameters. We refer to these as Active Training Modulations. We present an overview of HAMLET in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 324, + 389, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 389, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 389, + 337 + ], + "type": "text", + "content": "3.1. Model Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 342, + 547, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 606 + ], + "type": "text", + "content": "Our approach builds on the recent progress in unsupervised domain adaptation and segmentation networks. We start with DAFormer [20], a state-of-the-art UDA method, and adopt SegFormer [65] as our segmentation backbone due to its strong generalization capacity. We use three instances of the backbone, all pre-trained on the source domain: a student, a teacher, and a static (i.e. frozen) teacher. During training, the student receives a mix of target and source images [49] and is supervised with a \"mixed-sample\" cross-entropy loss, " + }, + { + "bbox": [ + 304, + 342, + 547, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_T" + }, + { + "bbox": [ + 304, + 342, + 547, + 606 + ], + "type": "text", + "content": " (represented by green, blue and red dashed lines, in Fig. 3). This loss is computed by mixing the teacher's pseudo-labels and source annotations. To improve training stability, the teacher is updated as the exponential moving average (EMA) of the student. To further regularize the student, we use source samples stored in a replay buffer and apply two additional losses (blue lines in Fig. 3). First, we minimize the feature distance (Euclidean) between the student and the static teacher's encoder, " + }, + { + "bbox": [ + 304, + 342, + 547, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{FD}" + }, + { + "bbox": [ + 304, + 342, + 547, + 606 + ], + "type": "text", + "content": ". Then, we employ a supervised cross-entropy task loss " + }, + { + "bbox": [ + 304, + 342, + 547, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_S" + }, + { + "bbox": [ + 304, + 342, + 547, + 606 + ], + "type": "text", + "content": ". Our complete objective is " + }, + { + "bbox": [ + 304, + 342, + 547, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_S + \\mathcal{L}_T + \\lambda_{FD}\\mathcal{L}_{FD}" + }, + { + "bbox": [ + 304, + 342, + 547, + 606 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 342, + 547, + 606 + ], + "type": "inline_equation", + "content": "\\lambda_{FD}" + }, + { + "bbox": [ + 304, + 342, + 547, + 606 + ], + "type": "text", + "content": " being a weight factor. During inference on the target domain, only the student is used (red lines in Fig. 3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 611, + 541, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 541, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 541, + 624 + ], + "type": "text", + "content": "3.2. Hardware-Aware Modular Training (HAMT)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Online adaptation requires updating the parameters during deployment time. However, back-propagation is computationally expensive and hence too slow to be continuously applied on a deployed agent. Opting for a partial weight update, for example by finetuning the last module of the network, would enable much more efficient training time. However, domain shifts can manifest as changes" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16550" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 74, + 532, + 259 + ], + "blocks": [ + { + "bbox": [ + 61, + 74, + 532, + 259 + ], + "lines": [ + { + "bbox": [ + 61, + 74, + 532, + 259 + ], + "spans": [ + { + "bbox": [ + 61, + 74, + 532, + 259 + ], + "type": "image", + "image_path": "622101d847939fac3dfcfdabded3d0f986e12affbefee80ea1b3898f02dd95a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 547, + 299 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 547, + 299 + ], + "type": "text", + "content": "Figure 3: HAMLET framework. We employ a student-teacher model with an EMA and a static teacher. HAMT orchestrates the back-propagation over the student restricting it to a network subsection. The Active Training Modulation instead controls the adaptation process by selectively enabling it only when necessary as well as tweaking sensitive training parameters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": "in both the data input distribution (such as attributes of the images, e.g. day/night) and the output distribution (e.g. class priors). This information could be encoded in different parts of the network, therefore just updating the very last segment might not suffice. This motivates the need for orchestrating the training process, to ensure sufficient training while minimizing the computational overhead. Inspired by reward-punishment [48] and reinforcement learning [56] policies, we introduce an orchestration agent in charge of deciding how deeply the network shall be finetuned through a trade-off between the pseudo-loss minimization rate and the computational time. In contrast to previous efficient back-propagation approaches [59, 23, 9], our model is pre-trained on the task and thus requires smaller updates to adapt. Let us start by modeling the problem. Our model backbone, " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": ", is composed of four different modules: " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "f = m_4 \\circ m_3 \\circ m_2 \\circ m_1" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": ". This defines our action space " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\{\\mathrm{T}_1, \\mathrm{T}_2, \\mathrm{T}_3, \\mathrm{T}_4\\}" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_4" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": " corresponds to training just the last module of the network, " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "m_4" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_3" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": " the last two modules, i.e. " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "m_4 \\circ m_3" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_2" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": " the last three, i.e. " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "m_4 \\circ m_3 \\circ m_2" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_1" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": " the whole network " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": ". We also define a continuous state space " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "S = \\{\\mathrm{R}, \\mathrm{V}\\}" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "\\mathrm{R}" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": " is the second derivative of the EMA teacher pseudo-loss, " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "l_t" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": ", over time, hence " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "R_t = -\\frac{\\Delta^2 l}{(\\Delta t)^2}" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": ", computed in discrete form as " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "R_t = -(l_t - 2l_{t-1} + l_{t-2})" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": ". V represents a cumulative vector with the same dimension as the action space " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": ", initialized at zero. Now we have everything in place to employ an expected-improvement based decision model. At each time-step " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": ", action " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "T_j" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": " is selected for " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "j = \\operatorname{argmax} V_t" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": ". During training step " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "inline_equation", + "content": "V[j]" + }, + { + "bbox": [ + 46, + 319, + 289, + 682 + ], + "type": "text", + "content": " is updated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 102, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathrm {V} [ j ] _ {t + 1} = \\alpha R _ {t} + (1 - \\alpha) \\mathrm {V} [ j ] _ {t} \\tag {1}", + "image_path": "f6d170990c01553ac036abee00914e61255cb9dd03e3afdc88512a092b95f42b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "content": " is a smoothing factor, e.g. 0.1. i.e. " + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{V_t}" + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "content": " hold a discrete exponential moving average of " + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "inline_equation", + "content": "R_{t}" + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "content": ". Therefore, our policy can be seen as a greedy module selection based on the highest expected loss improvement over its linear approximation. A notable drawback of this policy is that we will inevitably converge towards picking more rewarding, yet expensive, actions i.e. " + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{T_1,T_2}" + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "content": " compared to more efficient but potentially less effective actions i.e. " + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_3,\\mathrm{T}_4" + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "content": ". However, our goal is not to maximize " + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "inline_equation", + "content": "-\\frac{\\Delta^2l}{(\\Delta t)^2}" + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "content": " is the number of updates, our goal is instead to maximize " + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "inline_equation", + "content": "-\\frac{\\Delta^2l}{(\\Delta\\tau)^2}" + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\Delta \\tau" + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "content": " is a real-time interval. Therefore, we have to introduce in the optimization policy some notion of the actual training cost of each action in " + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "content": " on the target device. To start with, we measure the training time associated with each action, obtaining " + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\omega_{T} = \\{\\omega_{T_{1}},\\omega_{T_{2}},\\omega_{T_{3}},\\omega_{T_{4}}\\}" + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "content": ". With this we can compute the time-conditioning vector " + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 319, + 547, + 506 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 350, + 514, + 545, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 514, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 350, + 514, + 545, + 552 + ], + "type": "interline_equation", + "content": "\\gamma_ {j} = \\frac {e ^ {\\frac {1}{\\beta \\omega_ {T _ {j}}}}}{\\sum_ {k = 1} ^ {K} e ^ {\\frac {1}{\\beta \\omega_ {T _ {k}}}}} \\quad \\text {f o r} j = 1, \\dots , K \\tag {2}", + "image_path": "da918dc60a581fc12da9b4a7293c0642098df19d96a13bb27dc89e98dbd68430.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 559, + 547, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 547, + 608 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 559, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 559, + 547, + 608 + ], + "type": "text", + "content": " is the softmax temperature, and " + }, + { + "bbox": [ + 304, + 559, + 547, + 608 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 559, + 547, + 608 + ], + "type": "text", + "content": " the number of actions, i.e. 4 in our model. We modify our update policy to favor less computationally expensive modules by scaling the updates with " + }, + { + "bbox": [ + 304, + 559, + 547, + 608 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 559, + 547, + 608 + ], + "type": "text", + "content": ", replacing Eq. 1 with:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 627, + 545, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 627, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 318, + 627, + 545, + 657 + ], + "type": "interline_equation", + "content": "\\mathrm {V} [ j ] _ {t + 1} = \\left\\{ \\begin{array}{l l} \\gamma_ {j} \\alpha R _ {t} + (1 - \\alpha) \\mathrm {V} [ j ] _ {t} & \\text {i f} R _ {t} \\geq 0 \\\\ (1 - \\gamma_ {j}) \\alpha R _ {t} + (1 - \\alpha) \\mathrm {V} [ j ] _ {t} & \\text {i f} R _ {t} < 0 \\end{array} \\right. \\tag {3}", + "image_path": "92ff6e4fb2e74eaf542cb945517dfabc2017afa76c23dfe7ec0f71712aa3254c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "content": "This policy makes it so that more expensive actions receive smaller rewards and larger punishments. Despite its simplicity, this leads to a significant reduction in FLOPS for an average back-propagation " + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "inline_equation", + "content": "-30\\%" + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\beta = 2.75" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16551" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "or " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "-43\\%" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": ". We finally choose " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\beta = 1.75" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " to obtain a FLOPS reduction of " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "-34\\%" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": ". Exhaustive ablations on HAMT are presented in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 121, + 201, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 201, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 201, + 133 + ], + "type": "text", + "content": "3.3. Active Training Modulation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 141, + 289, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 141, + 289, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 141, + 289, + 345 + ], + "type": "text", + "content": "Continuous and test-time adaptation methods tackle online learning as a continuous and constant process carried out on the data stream. Nevertheless, this approach presents several shortcomings when it comes to real-world deployments. Performing adaptation when the deployment domain is unchanged does not lead to further performance improvements on the current domain; instead, it might cause significant forgetting on previous domains, hence hindering model generalization (we present evidence of this in the supplementary material). Even if mitigated by HAMT, online training remains a computationally expensive procedure, also due to several teachers' necessary forward passes. However, knowing when and what kind of adaptation is needed is not a trivial task. We tackle this by introducing an Adaptive Domain Detection mechanism, in Sec. 3.3.1, and then a set of strategies to reduce the training time while optimizing the learning rate accordingly, in Sec. 3.3.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 369, + 200, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 200, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 200, + 380 + ], + "type": "text", + "content": "3.3.1 Adaptive Domain Detection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "content": "A key element of an online adaptation system consists of acquiring awareness of the trajectory in the data distribution space, i.e. domains, traveled by the student model during deployment. We can model the problem by setting the trajectory origin in the source domain. With high dimensional data, the data distribution is not tractable, therefore the trajectory cannot be described in closed form. Recent work [41] introduced the notion of distance between the current deployed domain and source by approximating it with the confidence drop of a source pre-trained model. This approach heavily relies on the assumption that the pretrained model is well-calibrated. While this might hold for domains close to source, the calibration quickly degrades in farther domains [45, 41]. This myopic behavior dampen the simple use of confidence for domain detection. Furthermore, the additional forward pass increases the computational cost during deployment. We tackle these limitations with an equivalently simple, yet more robust, approach. We modify the backbone of the static teacher " + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "inline_equation", + "content": "f^{\\mathrm{st}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "content": " used for the feature distance loss " + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{FD}" + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "content": " by connecting a lightweight segmentation head, " + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "inline_equation", + "content": "d_1^{\\mathrm{st}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "content": ", after the first encoder module " + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "inline_equation", + "content": "m_1^{\\mathrm{st}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "inline_equation", + "content": "h_1^{\\mathrm{st}} = d_1^{\\mathrm{st}} \\circ m_1^{\\mathrm{st}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "content": ". This additional decoder, " + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "inline_equation", + "content": "h_1^{\\mathrm{st}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "content": ", is trained offline, on source data, without propagating gradients in the backbone (" + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "inline_equation", + "content": "m_1^{\\mathrm{st}}" + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "content": " is frozen). Given a target sample " + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "inline_equation", + "content": "x_T" + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "content": ", we propose to compute the cross-entropy between the one-hot encoded student prediction " + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "inline_equation", + "content": "p(x_T) = 1_{\\arg \\max(f(x_T))}" + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "content": " and the lightweight decoder prediction " + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "inline_equation", + "content": "g(x_T) = h_1^{\\mathrm{st}}(x_T)" + }, + { + "bbox": [ + 46, + 391, + 289, + 715 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 334, + 89, + 547, + 124 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 89, + 547, + 124 + ], + "spans": [ + { + "bbox": [ + 334, + 89, + 547, + 124 + ], + "type": "interline_equation", + "content": "H _ {T} ^ {(i)} = - \\sum_ {p = 1} ^ {H \\times W} \\sum_ {c = 1} ^ {C} p \\left(x _ {T} ^ {(i)}\\right) \\log g \\left(x _ {T} ^ {(i)}\\right) \\Big | _ {p, c} \\tag {4}", + "image_path": "8befff2218ed7fe9fa2b5da8598b8ca1db6198dd623881c0edafb4744decebac.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 131, + 547, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 131, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 304, + 131, + 547, + 284 + ], + "type": "text", + "content": "Thanks to the student model's higher generalization capability (both due to a larger number of parameters and the unsupervised adaptation process), it will always outperform the lightweight decoder head. Nevertheless, since now the distance is measured in the prediction space, we are not subjected to model miscalibration. Furthermore, since the student model is in constant adaptation, the domain distance accuracy actually improves over time, leading to better results. We present evidence of these claims in the supplementary material. We now define a denoised signal by using bin-averaging " + }, + { + "bbox": [ + 304, + 131, + 547, + 284 + ], + "type": "inline_equation", + "content": "A_T^{(i)} = \\sum_{j=m}^{m(i+1)-1} \\frac{H_T^{(j)}}{m}" + }, + { + "bbox": [ + 304, + 131, + 547, + 284 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 131, + 547, + 284 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 131, + 547, + 284 + ], + "type": "text", + "content": " is the bin size. Domains are modeled as discrete steps of " + }, + { + "bbox": [ + 304, + 131, + 547, + 284 + ], + "type": "inline_equation", + "content": "A_T^{(i)}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 332, + 297, + 546, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 297, + 546, + 327 + ], + "spans": [ + { + "bbox": [ + 332, + 297, + 546, + 327 + ], + "type": "interline_equation", + "content": "B _ {0} = A _ {0} \\quad B _ {i} = \\left\\{ \\begin{array}{l l} A _ {i} & \\text {i f} | B _ {i - 1} - A _ {i} | > z \\\\ B _ {i - 1} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}", + "image_path": "14ccca4499067715fc42c9dee430fb6b183054d4ace61e107a921b972b5e2feb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 332, + 547, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 547, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 332, + 547, + 381 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 332, + 547, + 381 + ], + "type": "text", + "content": " is the discretized signal and " + }, + { + "bbox": [ + 304, + 332, + 547, + 381 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 332, + 547, + 381 + ], + "type": "text", + "content": " is the minimum distance used to identify new domains. Finally, we refer to the signed amplitude of domain shifts as " + }, + { + "bbox": [ + 304, + 332, + 547, + 381 + ], + "type": "inline_equation", + "content": "\\Delta B_{i} = B_{i} - B_{i - 1}" + }, + { + "bbox": [ + 304, + 332, + 547, + 381 + ], + "type": "text", + "content": " and a domain change is detected whenever " + }, + { + "bbox": [ + 304, + 332, + 547, + 381 + ], + "type": "inline_equation", + "content": "|\\Delta B_i| > z" + }, + { + "bbox": [ + 304, + 332, + 547, + 381 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 395, + 526, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 395, + 526, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 395, + 526, + 407 + ], + "type": "text", + "content": "3.3.2 Least Training and Adaptive Learning Rate" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "text", + "content": "The definitions of " + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "text", + "content": " allow us to customize the training process. To this end, we adopt a Least Training (LT) strategy and trigger adaptation only when facing a new domain, which occurs when " + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "inline_equation", + "content": "|\\Delta B_i| > z" + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "text", + "content": ". Effective online learning performance depends heavily on the choice of hyperparameters such as the learning rate " + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "text", + "content": " and learning rate decay rate. Therefore, we can adjust these parameters to facilitate adaptation according to the nature and intensity of domain shifts we encounter, we refer to this orchestration as Adaptive Learning Rate (ALR). For example, the larger the domain shift (i.e. " + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "inline_equation", + "content": "|\\Delta B_i|" + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "text", + "content": "), the more we need to adapt to counteract its effect. This can be achieved by either running more optimization steps or using a higher learning rate. Whenever a domain shift is detected, we compute the number of adaptation iterations " + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "inline_equation", + "content": "L = K_l\\frac{|\\Delta B_i|}{z}" + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "text", + "content": ", hence proportionally to the amplitude of the shift " + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "inline_equation", + "content": "|\\Delta B_i|" + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "text", + "content": " relative to the threshold " + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "inline_equation", + "content": "K_l" + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "text", + "content": " is a multiplicative factor representing the minimum adaptation iterations. If a new domain shift takes place before the adaptation process completes, we accumulate the required optimization steps. Then, we can play on two further parameters: " + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "inline_equation", + "content": "K_l" + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "text", + "content": " and the learning rate schedule. We argue that proper scheduling is crucial for attaining a smoother adaptation. The learning rate, " + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "text", + "content": ", is linearly decayed until the adaptation is concluded – the smaller the domain shift, the faster the decay. While the initial learning rate, " + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "inline_equation", + "content": "K_{\\eta}" + }, + { + "bbox": [ + 304, + 413, + 547, + 715 + ], + "type": "text", + "content": ", should" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16552" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 69, + 550, + 181 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 550, + 181 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 550, + 181 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 550, + 181 + ], + "type": "table", + "html": "
HAMTLTALRDCMRCS200mm (mIoU)All-domains (mIoU)FPSAverage GFLOPSAdaptation GFLOPS
TotalFwd.Bwd.Fwd.Bwd.
(A)-----62.2 ± 0.969.5 ± 0.35.9 ± 0.0125.2 ± 0.094.4 ± 0.030.8 ± 0.056.6 ± 0.030.8 ± 0.0
(B)✓----60.2 ± 0.568.7 ± 0.37.0 ± 0.1114.7 ± 0.094.4 ± 0.020.3 ± 0.056.6 ± 0.020.3 ± 0.0
(C)✓✓---51.8 ± 0.565.7 ± 0.229.5 ± 0.644.4 ± 0.542.6 ± 0.41.8 ± 0.256.6 ± 0.020.2 ± 0.2
(D)✓✓✓--54.1 ± 1.265.9 ± 0.229.5 ± 0.544.4 ± 0.342.7 ± 0.21.8 ± 0.156.6 ± 0.020.3 ± 0.1
(E)✓✓✓✓-56.6 ± 0.866.3 ± 0.128.9 ± 0.344.7 ± 0.242.9 ± 0.21.8 ± 0.156.6 ± 0.020.2 ± 0.0
(F)✓✓✓-✓55.8 ± 1.066.3 ± 0.229.1 ± 1.145.2 ± 0.143.2 ± 0.12.0 ± 0.056.6 ± 0.020.3 ± 0.0
(G)✓✓✓✓✓58.2 ± 0.866.9 ± 0.329.7 ± 0.645.7 ± 0.343.6 ± 0.22.1 ± 0.156.6 ± 0.020.2 ± 0.1
(a)
clear 1200mmclear 2100mmclear 375mmclear 4clear h-meantarget h-meantotal h-meanFPSGFLOPS
(A)72.952.273.664.273.067.673.473.260.667.25.6125.2
(B)73.050.473.462.173.067.373.273.159.166.46.8114.7
(C)73.446.073.561.573.666.173.873.656.565.17.2100.0
(G)73.453.673.165.273.568.273.273.361.667.89.182.2
(b)
", + "image_path": "44f7b326ae0dbd48758eddb9e71d95fc184ee4ecd72b7456baac2e6512710f22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 184, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 46, + 184, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 184, + 545, + 208 + ], + "type": "text", + "content": "Table 1: Ablation studies - HAMLET components. Top: Increasing Storm (8925 frames per domain) [41], bottom: Fast Storm C [41] (2975 frames per domain). For each configuration, we report mIoU, framerate, and GFLOPS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 220, + 286, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 286, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 286, + 243 + ], + "type": "text", + "content": "be higher when the domain shift is triggered in domains farther from the source" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 261, + 287, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 261, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 75, + 261, + 287, + 283 + ], + "type": "interline_equation", + "content": "K _ {\\eta} = K _ {\\eta , \\min } + \\frac {\\left(B _ {\\mathrm {i}} - B _ {\\text {s o u r c e}}\\right) \\left(K _ {\\eta , \\max } - K _ {\\eta , \\min }\\right)}{B _ {\\text {h a r d}} - B _ {\\text {s o u r c e}}} \\tag {6}", + "image_path": "1c22e35d68910ef5a0a14d2b6612d1fd14d81c016d1ca59912434f59a5cdb291.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "inline_equation", + "content": "B_{\\mathrm{source}}" + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "text", + "content": " (resp. " + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "inline_equation", + "content": "B_{\\mathrm{hard}}" + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "text", + "content": ") is an estimate of " + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "text", + "content": " when the network is close to (resp. far from) the source domain; and " + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "inline_equation", + "content": "K_{\\eta, \\min}" + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "text", + "content": " (resp. " + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "inline_equation", + "content": "K_{\\eta, \\max}" + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "text", + "content": ") is the value of " + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "inline_equation", + "content": "K_{\\eta}" + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "text", + "content": " assigned when the network is close to (resp. far away from) the source. Concerning " + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "inline_equation", + "content": "K_{l}" + }, + { + "bbox": [ + 47, + 291, + 287, + 387 + ], + "type": "text", + "content": ", we posit that moving towards the source requires less adaptation than going towards harder domains: the model shows good recalling of previously explored domains and thanks to the employed regularization strategies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 405, + 287, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 405, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 59, + 405, + 287, + 434 + ], + "type": "interline_equation", + "content": "K _ {l} = \\left\\{ \\begin{array}{l l} K _ {l, \\max } & \\text {i f} \\Delta B _ {i} \\geq 0 \\\\ K _ {1, \\min } + \\frac {\\left(B _ {\\mathrm {i}} - B _ {\\text {s o u r c e}}\\right) \\left(K _ {1 , \\max } - K _ {1 , \\min }\\right)}{B _ {\\text {h a r d}} - B _ {\\text {s o u r c e}}} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {7}", + "image_path": "b62b18e1e28f91908ea49fb98faf9686694d6d1d3c061cced8c589f4f6702cd0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 441, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 287, + 501 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 441, + 287, + 501 + ], + "type": "inline_equation", + "content": "K_{l,\\min}" + }, + { + "bbox": [ + 47, + 441, + 287, + 501 + ], + "type": "text", + "content": " (resp. " + }, + { + "bbox": [ + 47, + 441, + 287, + 501 + ], + "type": "inline_equation", + "content": "K_{l,\\max}" + }, + { + "bbox": [ + 47, + 441, + 287, + 501 + ], + "type": "text", + "content": ") is the value of " + }, + { + "bbox": [ + 47, + 441, + 287, + 501 + ], + "type": "inline_equation", + "content": "K_{l}" + }, + { + "bbox": [ + 47, + 441, + 287, + 501 + ], + "type": "text", + "content": " assigned when the model is close to (resp. far away from) the source domain. Extensive ablations in the supplementary material will highlight how the orchestration of the adaptation hyperparameters improves the accuracy-speed trade-off." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 518, + 194, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 194, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 194, + 529 + ], + "type": "text", + "content": "3.3.3 Dynamic ClassMix (DCM)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 538, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 681 + ], + "type": "text", + "content": "ClassMix [39] provides a simple mechanism for data augmentation by mixing classes from the source dataset into target images. Usually " + }, + { + "bbox": [ + 46, + 538, + 287, + 681 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 538, + 287, + 681 + ], + "type": "text", + "content": " of the classes in the source dataset are selected, however we notice that this percentage is a highly sensitive hyperparameter in online domain adaptation. Injecting a significant portion of source classes has a beneficial impact when adapting to domains closer to the source domain, whereas when adapting to domains further from the source the opposite effect can be observed, as it effectively slows down the adaptation process. We therefore exploit once more the deployment domain awareness to control the mixing augmentation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 694, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 694, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 68, + 694, + 287, + 717 + ], + "type": "interline_equation", + "content": "K _ {\\mathrm {C M}} = K _ {\\mathrm {C M}, \\min } + \\frac {\\left(B _ {\\mathrm {i}} - B _ {\\text {s o u r c e}}\\right) \\left(K _ {\\mathrm {C M} , \\max } - K _ {\\mathrm {C M} , \\min }\\right)}{B _ {\\text {h a r d}} - B _ {\\text {s o u r c e}}}. \\tag {8}", + "image_path": "4941f7ebaebafd877c50baed6f344bde596f76e209a40f9734be9ca10f8f58a0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 220, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 545, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 220, + 545, + 266 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{CM}}" + }, + { + "bbox": [ + 306, + 220, + 545, + 266 + ], + "type": "text", + "content": " is the percentage of source classes used during adaptation; and " + }, + { + "bbox": [ + 306, + 220, + 545, + 266 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{CM,min}}" + }, + { + "bbox": [ + 306, + 220, + 545, + 266 + ], + "type": "text", + "content": " (resp. " + }, + { + "bbox": [ + 306, + 220, + 545, + 266 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{CM,max}}" + }, + { + "bbox": [ + 306, + 220, + 545, + 266 + ], + "type": "text", + "content": ") is the value of " + }, + { + "bbox": [ + 306, + 220, + 545, + 266 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{CM}}" + }, + { + "bbox": [ + 306, + 220, + 545, + 266 + ], + "type": "text", + "content": " assigned when the network is close to (resp. far away from) the source domain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 281, + 410, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 410, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 410, + 293 + ], + "type": "text", + "content": "3.3.4 Buffer Sampling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 371 + ], + "type": "text", + "content": "Following [41], to simulate real deployment, we limit our access to the source domain by using a replay buffer. Additionally, instead of initializing at random (with a uniform prior), we apply Rare Class Sampling (RCS) (skewed priors) as in [20]. This incentives a more balanced class distribution over the buffer, ultimately leading to better accuracy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 381, + 431, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 431, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 431, + 395 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 402, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 557 + ], + "type": "text", + "content": "The experiments are carried out on (a) the OnDA benchmarks [41] and (b) the SHIFT dataset [45]. (a) is a semi-syntheptic benchmark, as it applies synthetic rain and fog [50] over 4 different intensities profiles. The main benchmark, Increasing Storm, presents a storm with a pyramidal intensity profile; see Fig. 4. In contrast, (b) is a purely synthetic dataset, where both the underlying image and the weather are synthetically generated and thus domain change is fully controllable. All models are evaluated using mIoU: following [41], we report the harmonic mean over domains to present the overall adaptation performance. All experiments were carried out using an Nvidia™ RTX 3090 GPU. We refer to supplementary material for further details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 563, + 406, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 406, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 406, + 574 + ], + "type": "text", + "content": "4.1. Ablation Studies" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "content": "In Tab. 1 we study the impact of each contribution to adaptation performance, both in terms of accuracy and efficiency. For each configuration, we report mIoU over different portions of the sequence, the framerate and the amount of GFLOPS - respectively averages of: total, forward and backward passes, and dedicated adaptation only, also divided in forward (Fwd) and backward (Bwd). Tab. 1 (a) shows results on the Increasing Storm scenario [41]. Here, we show mIoU over the " + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "inline_equation", + "content": "200\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "content": " domain, i.e. the hardest in the sequence, as well as the mIoU averaged over forward and backward adaptation, i.e., from clear to " + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "inline_equation", + "content": "200\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "content": " rain" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16553" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 533, + 194 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 533, + 194 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 533, + 194 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 533, + 194 + ], + "type": "table", + "html": "
clear25mm50mm75mm100mm200mmh-meanFPSGFLOPS
FBFBFBFBFBFBFBT
(A)DeepLabV2 (no adaptation)64.5-57.1-48.7-41.5-34.4-18.537.3--39.4-
(B)DeepLabV2 fully supervised (oracle)64.5-64.1-63.7-63.0-62.4-58.262.6--39.4-
(C)OnDA64.564.860.457.157.354.554.852.252.049.142.254.255.1-1.3-
(D)SegFormer MiT-B1 (no adaptation)73.4-68.8-64.2-58.0-51.8-31.257.8--48.434.9
(E)SegFormer MiT-B5 (no adaptation)77.6-73.9-71.0-67.2-62.6-46.764.7--11.5240.4
(F)SegFormer MiT-B1 fully supervised (oracle)72.9-72.4-72.1-71.5-70.7-68.671.3--48.434.9
(G)TENT73.072.868.568.664.564.859.760.254.554.835.956.263.659.910.0-
(H)TENT + Replay Buffer73.072.868.568.664.564.859.760.254.454.735.856.163.659.97.8-
(I)CoTTA72.574.469.570.965.968.266.164.764.663.557.265.668.166.80.6593.8
(J)CoTTA real-time73.375.470.370.666.966.462.561.457.656.939.759.265.562.327.041.7
(K)HAMLET (ours)73.471.070.168.867.767.566.666.465.564.659.266.867.667.229.145.7
", + "image_path": "6c6f9b71ed359d9f1d07076d2643823f75623ee47b1119997bca1ceaf559c2b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 195, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 195, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 545, + 220 + ], + "type": "text", + "content": "Table 2: Comparison against other models – Increasing storm scenario. (A-C) methods built over DeepLabv2, (D-E) SegFormer variants trained on source, (F) oracle, (G-K) models adapted online. We report mIoU, framerate, and GFLOPS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "text", + "content": "and backward. Results are averaged over 3 runs with different seeds, with standard deviation being reported. (A) reports the results achieved by naively performing full adaptation of the model. HAMT can increase the framerate by roughly " + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "text", + "content": " by reducing the Bwd GFLOPS of " + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "inline_equation", + "content": "34\\%" + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "text", + "content": ", at the expense of as few as " + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "inline_equation", + "content": "0.7\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "text", + "content": " on average, i.e., about 2 points on the " + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "inline_equation", + "content": "200\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "text", + "content": " domain. The main boost in terms of speed is obviously given by LT (C), which inhibits the training in absence of detected domain shifts. LT increases the framerate by approximately " + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "text", + "content": " by decimating the total GFLOPS, yet not affecting the adaptation Bwd GFLOPS. This comes with a price in terms of mIoU, dropping by about 4 points on average and more than 10 points on " + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "inline_equation", + "content": "200\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "text", + "content": " - not a moderate drop anymore. LT impact highly depends on the domain sequence experienced during deployment: frequent domain changes could prevent training inhibition, thus neglecting LT gains in terms of efficiency, as we will appreciate later. The loss in accuracy is progressively regained by adding ALR (D), with further improvements yielded by one between DCM (E) and RCS (F), or both together (G) leading to the full HAMLET configuration. The three together allow for reducing the gap to 2.5 points mIoU - 4 over the " + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "inline_equation", + "content": "200\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "text", + "content": " domain - without sacrificing any efficiency. Tab. 1 (b) shows further results, on a faster version of Storm C [41]. This represents a much more challenging scenario, with harsher and " + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "inline_equation", + "content": "3\\times" + }, + { + "bbox": [ + 46, + 240, + 289, + 661 + ], + "type": "text", + "content": " more frequent domain shifts. Here we show the single domains mIoU, as well as harmonic mean on source and target domains, and all frames. As expected, in this benchmark, LT alone (C) results much less effective than before, with a much lower gain in FPS and GFLOPS. Here, the synergy between the HAMT, LT, and the other components (G) allows for the best accuracy and speedup - even outperforming the full training variant (A) - highlighting their complementarity. Further ablations are in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 670, + 203, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 203, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 203, + 683 + ], + "type": "text", + "content": "4.2. Results on Increasing Storm" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Tab. 2 shows a direct comparison between HAMLET and relevant approaches. The presented test-time adaptation" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 332, + 246, + 520, + 275 + ], + "blocks": [ + { + "bbox": [ + 332, + 246, + 520, + 275 + ], + "lines": [ + { + "bbox": [ + 332, + 246, + 520, + 275 + ], + "spans": [ + { + "bbox": [ + 332, + 246, + 520, + 275 + ], + "type": "image", + "image_path": "1f4433bc49d15129f594d5c8ec125e604e8568ee8e10b04e24fd6d40576a4f56.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 334, + 277, + 519, + 334 + ], + "blocks": [ + { + "bbox": [ + 334, + 277, + 519, + 334 + ], + "lines": [ + { + "bbox": [ + 334, + 277, + 519, + 334 + ], + "spans": [ + { + "bbox": [ + 334, + 277, + 519, + 334 + ], + "type": "image", + "image_path": "d24ecd2323e20c9b68c67191586e436a765e6a0dbea0a39967845645c400b4bb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 338, + 333, + 518, + 393 + ], + "blocks": [ + { + "bbox": [ + 338, + 333, + 518, + 393 + ], + "lines": [ + { + "bbox": [ + 338, + 333, + 518, + 393 + ], + "spans": [ + { + "bbox": [ + 338, + 333, + 518, + 393 + ], + "type": "image", + "image_path": "79e0895ecd0d4a25ec03c1c4836f78786d7d64dc4e4bb9e31e46c9d83370760d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 335, + 396, + 518, + 460 + ], + "blocks": [ + { + "bbox": [ + 335, + 396, + 518, + 460 + ], + "lines": [ + { + "bbox": [ + 335, + 396, + 518, + 460 + ], + "spans": [ + { + "bbox": [ + 335, + 396, + 518, + 460 + ], + "type": "image", + "image_path": "fda9ff3d1862761f0f33b02a91c509e2a6e7eae557fe386105aad3905486660b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 461, + 545, + 498 + ], + "lines": [ + { + "bbox": [ + 305, + 461, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 461, + 545, + 498 + ], + "type": "text", + "content": "Figure 4: HAMLET on the Increasing Storm. We show rain intensity (in millimetres), mIoU over active (bold) and inactive (dashed) domains, learning rate and FPS." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "strategies namely - TENT and CoTTA - were revised to handle the online setting and be fairly compared with HAMLET. All methods start with the same exact initial weights - with HAMLET requiring the additional lightweight decoder, not needed by TENT and CoTTA - using SegFormer MiT-B1 as the backbone, since it is " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": " faster than SegFormer MiT-B5 and thus better suited to keep real-time performance even during adaptation. We report results achieved by DeepLabv2 trained on source data only (A), an oracle model trained with full supervision (B), as well as OnDA [41] (C) as a reference. Then, we report SegFormer models trained on the source domain only (D) and (E). In (F) we show the performance achieved by an oracle SegFormer, trained on all domains fully supervised. Following [41], columns \"F\" concern forward adaptation from clear to " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "200\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": ", while columns \"B\" show backward adap" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16554" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 526, + 118 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 526, + 118 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 526, + 118 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 526, + 118 + ], + "type": "table", + "html": "
clear750m375m150m75mh-meanFPSGFLOPS
FBFBFBFBFBFBT
OnDA64.965.863.362.360.758.851.649.142.155.154.1-1.3-
SegFormer MiT-B1 (no adaptation)71.1-70.0-67.5-58.8-46.961.3--48.434.9
Full training71.572.172.974.771.973.167.668.161.368.771.970.35.6125.2
HAMLET (ours)71.171.670.370.868.869.264.364.357.065.968.967.424.850.7
", + "image_path": "c3e17f1a06f9bbc9db60b7e986743fd76d51f1b6498ce31fe1661a731e2e38f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 133, + 542, + 181 + ], + "blocks": [ + { + "bbox": [ + 48, + 119, + 545, + 131 + ], + "lines": [ + { + "bbox": [ + 48, + 119, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 545, + 131 + ], + "type": "text", + "content": "Table 3: Results on foggy domains. Comparison between OnDA, Source SegFormer, full training adaptation, and HAMLET." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 133, + 542, + 181 + ], + "lines": [ + { + "bbox": [ + 52, + 133, + 542, + 181 + ], + "spans": [ + { + "bbox": [ + 52, + 133, + 542, + 181 + ], + "type": "table", + "html": "
ClearCloudyOvercastSmall rainMid rainHeavy rainh-meanFPSGFLOPS
FBFBFBFBFBFBFBT
SegFormer MiT-B1 fully supervised (oracle)80.1-79.9-79.8-78.9-78.7-77.179.1--48.434.93
SegFormer MiT-B1 (no adaptation)79.6-77.1-75.4-73.4-71.4-66.773.7--48.434.93
Full training78.979.376.776.876.877.974.874.876.376.574.076.277.076.65.0125.1
HAMLET (ours)79.678.976.976.676.177.473.374.374.276.074.275.776.676.126.843.9
", + "image_path": "cde8bd9b6f9bd3c966bb0af02c355124e705c5f8361ff5cd23d622250eda5156.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 182, + 543, + 194 + ], + "lines": [ + { + "bbox": [ + 49, + 182, + 543, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 182, + 543, + 194 + ], + "type": "text", + "content": "Table 4: Results on SHIFT dataset [45]. Comparison between Source SegFormer, full training adaptation, and HAMLET." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "text", + "content": "tation from " + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "inline_equation", + "content": "200\\mathrm{mm}" + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "text", + "content": " to clear, while the h-mean T refers to the overall harmonic mean. We can notice how SegFormer results are much more robust to domain changes with respect to DeepLabv2. Indeed, SegFormer MiT-B5 (E), without any adaptation, results more accurate than DeepLabv2 oracle (B), as well as better and faster than OnDA (C). The faster variant (D) outperforms OnDA both in speed and accuracy, reaching 48 FPS. Nevertheless, domain changes still dampen the full potential of SegFormer. Indeed, the oracle (F) outperforms (D) by about " + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "inline_equation", + "content": "+14" + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "text", + "content": " mIoU. However, this is not meaningful for real deployment experiencing unpredictable domain shifts, as it assumes to have data available in advance. Concerning test-time models, TENT starts adapting properly only beyond " + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "inline_equation", + "content": "50\\mathrm{mm}" + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "text", + "content": ", both with (G) and without (H) frame buffer, while it loses some accuracy on " + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "inline_equation", + "content": "25\\mathrm{mm}" + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "text", + "content": ". This makes its overall forward adaptation performance slightly worse compared to the pre-trained model (D), while being better at backward adaptation. Despite outperforming SegFormer MiT-B1, TENT is both slower and less accurate than SegFormer MiT-B5 running without any adaptation, further suggesting the robustness of the latter and making TENT not suitable for real-world deployment. On the contrary, CoTTA (I) outperforms both SegFormer models trained on source only, at the expense of dropping the framerate below 1FPS. It is worth mentioning that these metrics were collected after each domain was completed by each model individually. In an evaluation setup imposing a shared time frame, slower models would present much lower metrics, since their adaptation process would result constantly lagged. In fact, forcing CoTTA to run in real-time, at nearly 30FPS – i.e. by training once every 50 frames – dramatically reduces the effectiveness of the adaptation process (J), with drastic drops in the hardest domains. Finally, HAMLET (K) succeeds on any fronts, improving the baseline (D) by about 10 points with only a cost of " + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "text", + "content": " in terms of speed, while outperforming SegFormer MiT-B5 (E) both on accuracy " + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "inline_equation", + "content": "(+2.5" + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "text", + "content": " mIoU) and speed " + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "inline_equation", + "content": "(3\\times" + }, + { + "bbox": [ + 47, + 215, + 289, + 709 + ], + "type": "text", + "content": " faster) – being the only method achieving this, and thus the only suitable choice for real-time applications. Fig. 4 shows the overall behavior of HAMLET while adapting over the Increasing Storm. In addition to the rain intensity and the" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 331, + 220, + 521, + 434 + ], + "blocks": [ + { + "bbox": [ + 331, + 220, + 521, + 434 + ], + "lines": [ + { + "bbox": [ + 331, + 220, + 521, + 434 + ], + "spans": [ + { + "bbox": [ + 331, + 220, + 521, + 434 + ], + "type": "image", + "image_path": "0b8e0c1eeba7fae64b64798ae555b3b1cccde62d7dee9dc25e223452ccce2c5d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "lines": [ + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "text", + "content": "Figure 5: HAMLET on the SHIFT benchmark. We show mIoU over active (bold) and inactive (dashed) domains, learning rate and FPS." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 494, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 615 + ], + "type": "text", + "content": "mIoU achieved on each domain - active (bold) or inactive (dashed), i.e. respectively the mIoU on the domain being currently faced during deployment, and how the current adaptation affects the performance on the other domains to highlight the robustness to forgetting - we also report how the learning rate is modulated in correspondence of detected domain shifts, with a consequent drop in FPS due to the short training process taking place. For further experiments on harsher and sudden adaptation cycles, we include results of Storms A, B, C [41] in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 623, + 496, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 496, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 496, + 635 + ], + "type": "text", + "content": "4.3. Additional Results: Fog and SHIFT" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 715 + ], + "type": "text", + "content": "Fog. In Tab. 3, we investigate adaptation on the Increasing Fog scenario in the OnDA benchmark [41]. Crucially, for this experiment, we keep the same hyperparameters used for the Increasing Storm, since in both cases the starting SegFormer model is trained on the same source domain. This allows for validating how the proposed setting general-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16555" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 84, + 70, + 511, + 247 + ], + "blocks": [ + { + "bbox": [ + 84, + 70, + 511, + 247 + ], + "lines": [ + { + "bbox": [ + 84, + 70, + 511, + 247 + ], + "spans": [ + { + "bbox": [ + 84, + 70, + 511, + 247 + ], + "type": "image", + "image_path": "754b478f08cbb8f9ef25a77d64555c04cb1d29a05f6758abea85409722214ff3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 547, + 277 + ], + "type": "text", + "content": "Figure 6: Qualitative results - HAMLET in action. From left to right, we show frames from clean, " + }, + { + "bbox": [ + 46, + 251, + 547, + 277 + ], + "type": "inline_equation", + "content": "50\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 251, + 547, + 277 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 251, + 547, + 277 + ], + "type": "inline_equation", + "content": "100\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 251, + 547, + 277 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 251, + 547, + 277 + ], + "type": "inline_equation", + "content": "200\\mathrm{m}" + }, + { + "bbox": [ + 46, + 251, + 547, + 277 + ], + "type": "text", + "content": " domains. From top to bottom: input image, prediction by SegFormer trained on source domain and HAMLET." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 296, + 289, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 289, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 289, + 403 + ], + "type": "text", + "content": "izes at dealing with different kind of domain shifts, beyond those considered in the main experiments. We effectively use Increasing Fog as test set, and compare against Seg-Former trained on source (no adaptation) and a model that has been adapted by means of full online training optimization (configuration (A) of Table 1). HAMLET is able to adapt almost as well as the full online training model, with less than a " + }, + { + "bbox": [ + 46, + 296, + 289, + 403 + ], + "type": "inline_equation", + "content": "3\\mathrm{mIoU}" + }, + { + "bbox": [ + 46, + 296, + 289, + 403 + ], + "type": "text", + "content": " gap, while enjoying real-time adaptation at nearly " + }, + { + "bbox": [ + 46, + 296, + 289, + 403 + ], + "type": "inline_equation", + "content": "5\\times" + }, + { + "bbox": [ + 46, + 296, + 289, + 403 + ], + "type": "text", + "content": " the speed using just " + }, + { + "bbox": [ + 46, + 296, + 289, + 403 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 296, + 289, + 403 + ], + "type": "text", + "content": " of the FLOPS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 403, + 287, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 571 + ], + "type": "text", + "content": "SHIFT. We further test HAMLET on the SHIFT dataset [45]. Tab. 4 collects the results achieved by Seg-Former trained on source, full online training and HAMLET respectively, both at forward and backward adaptation across Clear, Cloudy, Overcast, Small rain, Mid rain and Heavy rain domains. Here HAMLET results highly competitive with the full training regime, with only 0.5 drop in average mIoU, while being more than " + }, + { + "bbox": [ + 46, + 403, + 287, + 571 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 46, + 403, + 287, + 571 + ], + "type": "text", + "content": " faster. Fig. 5 depicts, from top to bottom, the rain intensity characterizing any domain encountered on SHIFT, the mIoU achieved both on current (bold) and inactive (dashed) domains, the learning rate changes based on the domain shift detection, and the framerate achieved at any step. We refer to the supplementary material for a deeper analysis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 572, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 572, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 287, + 633 + ], + "type": "text", + "content": "Qualitative results. To conclude, Fig. 6 shows some qualitative examples from CityScapes. We can notice how SegFormer accuracy (second tow) drops with severe rain, whereas HAMLET (third row) is capable of keeping the same segmentation quality across the storm." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 644, + 117, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 117, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 117, + 657 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Orthogonality. HAMT and LT act independently. Indeed, by strongly constraining the adaptation periods through LT, HAMT has a limited margin of action. The impact of HAMT also depends on the backbone and by care" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 296, + 545, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 545, + 344 + ], + "type": "text", + "content": "fully crafting modular architectures, one can achieve further optimization. Nevertheless, in a deployment environment where domain shifts occur at high frequencies (e.g., Storm C), LT is ineffective, while HAMT thrives." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 345, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 545, + 392 + ], + "type": "text", + "content": "Measuring forgetting. An interesting topic we have not investigated consists of introducing an explicit awareness of which domains have been explored and how well we can recall them, expanding the distance " + }, + { + "bbox": [ + 304, + 345, + 545, + 392 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 345, + 545, + 392 + ], + "type": "text", + "content": " to multiple dimensions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 393, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 466 + ], + "type": "text", + "content": "Safety. We believe dynamic adaptation has the potential to enhance safety, but we acknowledge the necessity for rigorous testing and verification to safeguard against drift or catastrophic forgetting. This mandates a comprehensive effort from academia, industry, and certification authorities for ensuring the integrity of dynamically adapting models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 479, + 444, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 444, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 444, + 491 + ], + "type": "text", + "content": "6. Summary & Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 499, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 545, + 632 + ], + "type": "text", + "content": "We have presented HAMLET, a framework for real-time adaptation for semantic segmentation that achieves state-of-the-art performance on established benchmarks with continuous domain changes. Our approach combines a hardware-aware backpropagation orchestrator and a specialized domain-shift detector to enable active control over the model's adaptation, resulting in high framereates on a consumer-grade GPU. These advancements enable HAMLET to be a promising solution for in-the-wild deployment, making it a valuable tool for applications that require robust performance in the face of unforeseen domain changes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. The authors thank Gianluca Villani for the insightful discussion on reward-punishment policies, Leonardo Ravaglia for his expertise on hardware-aware training, and Lorenzo Andraghetti for exceptional technical support throughout the project. Their assistance was invaluable in the completion of this work." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16556" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Andreea Bobu, Judy Hoffman, Eric Tzeng, and Trevor Darrell. Adapting to continuously shifting domains. In ICLR 2018 Workshop Program Chairs, 2018. 00000." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 190 + ], + "type": "text", + "content": "[2] Chaoqi Chen, Weiping Xie, Wenbing Huang, Yu Rong, Xinghao Ding, Yue Huang, Tingyang Xu, and Junzhou Huang. Progressive feature alignment for unsupervised domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 627-636, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 257 + ], + "type": "text", + "content": "[3] Liang-Chieh Chen, Jonathan T Barron, George Papandreou, Kevin Murphy, and Alan L Yuille. Semantic image segmentation with task-specific edge detection using cnns and a discriminatively trained domain transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4545-4554, 2016." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 323 + ], + "type": "text", + "content": "[4] Liang-Chieh Chen, Raphael Gontijo Lopes, Bowen Cheng, Maxwell D Collins, Ekin D Cubuk, Barret Zoph, Hartwig Adam, and Jonathon Shlens. Naive-student: Leveraging semi-supervised learning in video sequences for urban scene segmentation. In European Conference on Computer Vision, pages 695-714. Springer, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 325, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 287, + 380 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 446 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(4):834-848, Apr 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 447, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 501 + ], + "type": "text", + "content": "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 502, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 557 + ], + "type": "text", + "content": "[8] Yi-Hsin Chen, Wei-Yu Chen, Yu-Ting Chen, Bo-Cheng Tsai, Yu-Chiang Frank Wang, and Min Sun. No more discrimination: Cross city adaptation of road scene segmenters. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2011-2020. IEEE, 2017. 00000." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 558, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 558, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 558, + 287, + 612 + ], + "type": "text", + "content": "[9] Feng Cheng, Mingze Xu, Yuanjun Xiong, Hao Chen, Xinyu Li, Wei Li, and Wei Xia. Stochastic backpropagation: A memory efficient strategy for training video models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8301-8310, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 668 + ], + "type": "text", + "content": "[10] Henghui Ding, Xudong Jiang, Ai Qun Liu, Nadia Magnenat Thalmann, and Gang Wang. Boundary-aware feature propagation for scene segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6819-6829, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Trans" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "formers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "text", + "content": "[12] A. Dundar, M. Y. Liu, Z. Yu, T. C. Wang, J. Zedlewski, and J. Kautz. Domain stylization: A fast covariance matching framework towards domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "type": "text", + "content": "[13] Jun Fu, Jing Liu, Hajjie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3146-3154, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "type": "text", + "content": "[14] Jun Fu, Jing Liu, Yuhang Wang, Yong Li, Yongjun Bao, Jinhui Tang, and Hanqing Lu. Adaptive context network for scene parsing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6748-6757, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "text", + "content": "[15] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, Jan. 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "text", + "content": "[16] Rui Gong, Martin Danelljan, Dengxin Dai, Danda Pani Paudel, Ajad Chhatkuli, Fisher Yu, and Luc Van Gool. Tacs: Taxonomy adaptive cross-domain semantic segmentation. In European Conference on Computer Vision, pages 19-35. Springer, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "text", + "content": "[17] Junjun He, Zhongying Deng, Lei Zhou, Yali Wang, and Yu Qiao. Adaptive pyramid context network for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7519-7528, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 426, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 513 + ], + "type": "text", + "content": "[18] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. CyCADA: Cycle-consistent adversarial domain adaptation. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 1989-1998, Stockholm, Sweden, 10-15 Jul 2018. PMLR." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 547 + ], + "type": "text", + "content": "[19] Judy Hoffman, Dequan Wang, Fisher Yu, and Trevor Darryll. FCNs in the wild: Pixel-level adversarial and constraint-based adaptation. CoRR, 2016. 00000." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 548, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 590 + ], + "type": "text", + "content": "[20] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. arXiv preprint arXiv:2111.14887, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[21] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. *Hrda: Context-aware high-resolution domain-adaptive semantic segmentation*. In European Conference on Computer Vision (ECCV), 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 690 + ], + "type": "text", + "content": "[22] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[23] Angela H Jiang, Daniel L-K Wong, Giulio Zhou, David G Andersen, Jeffrey Dean, Gregory R Ganger, Gauri Joshi," + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16557" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Michael Kaminsky, Michael Kozuch, Zachary C Lipton, et al. Accelerating deep learning by focusing on the biggest losers. arXiv preprint arXiv:1910.00762, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[24] Zhengkai Jiang, Yuxi Li, Ceyuan Yang, Peng Gao, Yabiao Wang, Ying Tai, and Chengjie Wang. Prototypical contrast adaptation for domain adaptive semantic segmentation. In European Conference on Computer Vision, pages 36-54. Springer, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 207 + ], + "type": "text", + "content": "[25] Myeongjin Kim and Hyeran Byun. Learning texture invariant representation for domain adaptation of semantic segmentation. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 284 + ], + "type": "text", + "content": "[26] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114, 12 2016." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 341 + ], + "type": "text", + "content": "[27] Yevhen Kuznietsov, Marc Proesmans, and Luc Van Gool. Towards unsupervised online domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 261-271, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 397 + ], + "type": "text", + "content": "[28] Xin Lai, Zhuotao Tian, Xiaogang Xu, Yingcong Chen, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Decouplenet: Decoupled network for domain adaptive semantic segmentation. In European Conference on Computer Vision, pages 369-387. Springer, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "text", + "content": "[29] Qicheng Lao, Xiang Jiang, Mohammad Havaei, and Yoshua Bengio. Continuous domain adaptation with variational domain-agnostic feature replay. arXiv preprint arXiv:2003.04382, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "text", + "content": "[30] Geon Lee, Chanho Eom, Wonkyung Lee, Hyekang Park, and Bumsub Ham. Bi-directional contrastive learning for domain adaptive semantic segmentation. In European Conference on Computer Vision, pages 38-55. Springer, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 544 + ], + "type": "text", + "content": "[31] Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention networks for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9167-9176, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "text", + "content": "[32] Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "text", + "content": "[33] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. CoRR, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 689 + ], + "type": "text", + "content": "[34] Yuejiang Liu, Parth Kothari, Bastien Germain van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++: When does self-supervised test-time training fail or thrive? In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[35] Yang Liu, Wei Zhang, and Jun Wang. Source-free domain adaptation for semantic segmentation. 2021." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[36] Jonathan Long, Evan Shelhamer, and Trevor Darrel. Fully convolutional networks for semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[37] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. Lecture Notes in Computer Science, page 415-430, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[38] Vladimir Nekrasov, Chunhua Shen, and Ian Reid. Lightweight refinenet for real-time semantic segmentation. In British Conference on Computer Vision (BMVC), 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 198, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 251 + ], + "type": "text", + "content": "[39] Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1369-1378, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 308 + ], + "type": "text", + "content": "[40] Fei Pan, Sungsu Hur, Seokju Lee, Junsik Kim, and In So Kweon. Ml-bpm: Multi-teacher learning with bidirectional photometric mixing for open compound domain adaptation in semantic segmentation. In European Conference on Computer Vision, pages 236–251. Springer, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 363 + ], + "type": "text", + "content": "[41] Theodoros Panagiotakopoulos, Pier Luigi Dovesi, Linus Harenstam-Nielsen, and Matteo Poggi. Online domain adaptation for semantic segmentation in ever-changing conditions. In European Conference on Computer Vision (ECCV), 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 409 + ], + "type": "text", + "content": "[42] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In European conference on computer vision, pages 102-118. Springer, 2016." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 411, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 442 + ], + "type": "text", + "content": "[43] Serban Stan and Mohammad Rostami. Unsupervised model adaptation for continual semantic segmentation. In AAAI, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 445, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 476 + ], + "type": "text", + "content": "[44] Peng Su, Shixiang Tang, Peng Gao, Di Qiu, Ni Zhao, and Xiaogang Wang. Gradient regularized contrastive learning for continual domain adaptation. 2020. 00000." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 533 + ], + "type": "text", + "content": "[45] Tao Sun, Mattia Segu, Janis Postels, Yuxuan Wang, Luc Van Gool, Bernt Schiele, Federico Tombari, and Fisher Yu. SHIFT: a synthetic driving dataset for continuous multi-task domain adaptation. In Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 578 + ], + "type": "text", + "content": "[46] Towaki Takikawa, David Acuna, Varun Jampani, and Sanja Fidler. Gated-scnn: Gated shape cnns for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5229-5238, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 580, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 612 + ], + "type": "text", + "content": "[47] Phillip Thomas, Lars Pandikow, Alex Kim, Michael Stanley, and James Grieve. Open synthetic dataset for improving cyclist detection, Nov 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 658 + ], + "type": "text", + "content": "[48] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 195-204, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 659, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 659, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 659, + 545, + 713 + ], + "type": "text", + "content": "[49] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1379-1389, January 2021." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16558" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[50] Maxime Tremblay, Shirsendu S. Halder, Raoul de Charette, and Jean-François Lalonde. Rain rendering for evaluating and improving robustness to bad weather. International Journal of Computer Vision, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 172 + ], + "type": "text", + "content": "[51] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Jun 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "text", + "content": "[52] Riccardo Volpi, Pau de Jorge, Diane Larlus, and Gabriela Csurka. On the road to online adaptation for semantic image segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 258 + ], + "type": "text", + "content": "[53] Vibashan VS, Poojan Oza, and Vishal M. Patel. Towards online domain adaptive object detection. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Jan 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 259, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 259, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 259, + 287, + 304 + ], + "type": "text", + "content": "[54] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 349 + ], + "type": "text", + "content": "[55] Haoran Wang, Tong Shen, Wei Zhang, Lingyu Duan, and Tao Mei. Classes matter: A fine-grained adversarial approach to cross-domain semantic segmentation. In The European Conference on Computer Vision (ECCV), August 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 381 + ], + "type": "text", + "content": "[56] Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. Haq: Hardware-aware automated quantization with mixed precision, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 382, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 403 + ], + "type": "text", + "content": "[57] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "text", + "content": "[58] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "text", + "content": "[59] Bingzhen Wei, Xu Sun, Xuancheng Ren, and Jingjing Xu. Minimal effort back propagation for convolutional neural networks. arXiv preprint arXiv:1709.05804, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 481, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 536 + ], + "type": "text", + "content": "[60] Tsung-Han Wu, Yi-Syuan Liou, Shao-Ji Yuan, Hsin-Ying Lee, Tung-I Chen, Kuan-Chih Huang, and Winston H Hsu. D2ada: Dynamic density-aware active domain adaptation for semantic segmentation. In European Conference on Computer Vision (ECCV), 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "type": "text", + "content": "[61] Zuxuan Wu, Xintong Han, Yen-Liang Lin, Mustafa Gokhan Uzunbas, Tom Goldstein, Ser Nam Lim, and Larry S. Davis. Dcan: Dual channel-wise alignment networks for unsupervised scene adaptation. Lecture Notes in Computer Science, page 535-552, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 646 + ], + "type": "text", + "content": "[62] Zuxuan Wu, Xin Wang, Joseph Gonzalez, Tom Goldstein, and Larry Davis. ACE: Adapting to changing environments for semantic segmentation. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2121-2130. IEEE, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[63] Markus Wulfmeier, Alex Bewley, and Ingmar Posner. Incremental adversarial domain adaptation for continually changing environments. 2018. 00000." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[64] Enze Xie, Wenjia Wang, Wenhai Wang, Peize Sun, Hang Xu, Ding Liang, and Ping Luo. Segmenting transparent objects in the wild with transformer. In IJCAI, 2021." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 676 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 126 + ], + "type": "text", + "content": "[65] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 546, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 172 + ], + "type": "text", + "content": "[66] Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic segmentation in street scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3684-3692, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 546, + 217 + ], + "type": "text", + "content": "[67] Yanchao Yang and Stefano Soatto. FDA: Fourier domain adaptation for semantic segmentation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4084-4094. IEEE, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 218, + 546, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 546, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 546, + 273 + ], + "type": "text", + "content": "[68] Changqian Yu, Jingbo Wang, Chao Peng, Changxin Gao, Gang Yu, and Nong Sang. Bisenet: Bilateral segmentation network for real-time semantic segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 325-341, 2018." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 274, + 546, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 546, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 546, + 318 + ], + "type": "text", + "content": "[69] Yuhui Yuan, Xiaokang Chen, Xilin Chen, and Jingdong Wang. Segmentation transformer: Object-contextual representations for semantic segmentation. In European Conference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 319, + 546, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 546, + 362 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 546, + 362 + ], + "type": "text", + "content": "[70] Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 363, + 546, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 363, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 363, + 546, + 407 + ], + "type": "text", + "content": "[71] Qiming Zhang, Jing Zhang, Wei Liu, and Dacheng Tao. Category anchor-guided unsupervised domain adaptation for semantic segmentation. Advances in Neural Information Processing Systems, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 408, + 546, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 546, + 441 + ], + "type": "text", + "content": "[72] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 442, + 546, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 442, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 442, + 546, + 487 + ], + "type": "text", + "content": "[73] Yizhou Zhou, Xiaoyan Sun, Zheng-Jun Zha, and Wenjun Zeng. Context-reinforced semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4046-4055, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 487, + 546, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 546, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 546, + 531 + ], + "type": "text", + "content": "[74] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), Oct 2017." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 532, + 546, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 546, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 546, + 586 + ], + "type": "text", + "content": "[75] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 587, + 546, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 546, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 546, + 632 + ], + "type": "text", + "content": "[76] Yang Zou, Zhiding Yu, Xiaofeng Liu, BVK Kumar, and Jinsong Wang. Confidence regularized self-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5982-5991, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 632, + 546, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 632, + 546, + 676 + ], + "spans": [ + { + "bbox": [ + 307, + 632, + 546, + 676 + ], + "type": "text", + "content": "[77] Yang Zou, Zhiding Yu, Xiaofeng Liu, B. V. K. Vijaya Kumar, and Jinsong Wang. Confidence regularized self-training. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2019." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16559" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Token-Label Alignment for Vision Transformers/ef46bf1f-34a3-4e70-bf1a-b3f5f27b821d_content_list.json b/2023/Token-Label Alignment for Vision Transformers/ef46bf1f-34a3-4e70-bf1a-b3f5f27b821d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..19fd9ffdb334d39987ade434b1093f7e75e7cc41 --- /dev/null +++ b/2023/Token-Label Alignment for Vision Transformers/ef46bf1f-34a3-4e70-bf1a-b3f5f27b821d_content_list.json @@ -0,0 +1,1543 @@ +[ + { + "type": "text", + "text": "Token-Label Alignment for Vision Transformers", + "text_level": 1, + "bbox": [ + 240, + 130, + 728, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Han Xiao $^{1,2,*}$ Wenzhao Zheng $^{1,2,*}$ Zheng Zhu $^{3}$ Jie Zhou $^{1,2}$ Jiwen Lu $^{1,2,\\dagger}$ $^{1}$ Beijing National Research Center for Information Science and Technology, China \n $^{2}$ Department of Automation, Tsinghua University, China \n $^{3}$ PhiGent Robotics \n{h-xiao20, zhengwz18}@mails.tsinghua.edu.cn; zhengzhu@ieee.org; {jzhou, lujiwen} @tsinghua.edu.cn", + "bbox": [ + 122, + 178, + 844, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Data mixing strategies (e.g., CutMix) have shown the ability to greatly improve the performance of convolutional neural networks (CNNs). They mix two images as inputs for training and assign them with a mixed label with the same ratio. While they are shown effective for vision transformers (ViTs), we identify a token fluctuation phenomenon that has suppressed the potential of data mixing strategies. We empirically observe that the contributions of input tokens fluctuate as forward propagating, which might induce a different mixing ratio in the output tokens. The training target computed by the original data mixing strategy can thus be inaccurate, resulting in less effective training. To address this, we propose a token-label alignment (TL-Align) method to trace the correspondence between transformed tokens and the original tokens to maintain a label for each token. We reuse the computed attention at each layer for efficient token-label alignment, introducing only negligible additional training costs. Extensive experiments demonstrate that our method improves the performance of ViTs on image classification, semantic segmentation, objective detection, and transfer learning tasks. Code is available at: https://github.com/Euphorial6/TL-Align.", + "bbox": [ + 75, + 335, + 473, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 700, + 209, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent developments of vision transformers (ViTs) have revolutionized the computer vision field and set new state-of-the-arts in various tasks, such as image classification [16, 43, 31, 10], object detection [5, 56, 12, 13], and semantic segmentation [28, 39, 54, 9]. The successful structure of alternative spatial mixing and channel mixing in ViTs also motivates the arising of high-performance MLP-like deep architectures [40, 41, 42] and promotes the evolution of better CNNs [15, 32, 17]. In addition to architec", + "bbox": [ + 75, + 726, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ture designs, an improved training strategy can also greatly boost the performance of a trained deep model [23, 44, 8, 7].", + "bbox": [ + 500, + 305, + 890, + 335 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The training of modern deep architecture almost all adopts data mixing strategies for data augmentation [48, 45, 24, 47, 52, 53], which have been proven to consistently improve the generalization performance. They randomly mix two images and their labels with the same mixing ratio to produce mixed data. As the most commonly used data mixing strategy, CutMix [52] performs a copy-and-paste operation on the spatial domain to produce spatially mixed images. While data mixing strategies have been widely studied for CNNs [48, 45, 24], few works have explored their compatibilities with ViTs [7]. We find that self-attention in ViTs causes a fluctuation of the original spatial structure. Unlike the translation equivalence that ensures a global label consistency for CNNs, self-attention in ViTs undermines this global consistency and causes a misalignment between the token and label. This misalignment induces a different mixing ratio in the output tokens. The training targets computed by the original data mixing strategies can then be inaccurate, resulting in less effective training.", + "bbox": [ + 496, + 338, + 892, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this, we propose a token-label alignment (TL-Align) method for ViTs to obtain a more accurate target for training. We present an overview of our method in Figure 1. We first assign a label to each input token in the mixed image according to the source of the token. We then trace the correspondence between the input tokens and the transformed tokens and align the labels accordingly. We assume that only the spatial self-attention and residual connection operation alter the presence of input tokens since channel MLP and layer normalization process each token independently. We reuse the computed attentions to linearly mix the labels of input tokens to obtain those of transformed tokens. The token-label alignment is performed iteratively to obtain a label for each output token. For class-token-based classification (e.g., ViT [16] and DeiT [43]), we directly use the aligned label for the output class token as the training target. For global-pooling-based classification (e.g., Swin [31]), we similarly average the labels of output tokens as the train", + "bbox": [ + 496, + 628, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 205, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5495", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a75d82a46cf1b38b0be800352cada4aca186c90985e32fccc083502cd5e3af83.jpg", + "image_caption": [ + "(a) CutMix" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 240, + 367 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d34ab6d53d42770cb0ac9219357b66933ab60f35a6c7acb00557ef2a3336d492.jpg", + "image_caption": [ + "(b) Token misalignment for ViTs" + ], + "image_footnote": [], + "bbox": [ + 243, + 89, + 496, + 364 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/affaf24816e0d1ff852247366f51daa33c17b7a80f7201d6bdd47bbe24de14a9.jpg", + "image_caption": [ + "(c) Different ways to obtain the mixed labels", + "Figure 1. An overview of the proposed TL-Align. (a) CutMix-like methods [52] are widely used in model training, which spatially mix the tokens and their labels in the input space. (b) They are originally designed for CNNs and assume the processed tokens are spatially aligned with the input tokens. We show that it does not hold true for ViTs due to the global receptive field and adaptive weights. (c) Compared with existing methods, our method can effectively and efficiently align the tokens and labels without requiring a pretrained teacher network." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 885, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing target. The proposed TL-Align is only used for training to improve performance and introduces no additional workload for inference. We apply the proposed TL-Align to various ViT variants with CutMix including plain ViTs (DeiT [43]) and hierarchical ViTs (Swin [31]). We observe a consistent performance boost across different models on ImageNet-1K [14]. Specifically, our TL-Align improves DeiT-S by $0.8\\%$ using the same training recipe. We evaluate the ImageNet-pretrained models on various downstream tasks including semantic segmentation, objection detection, and transfer learning. Experimental results also verify the robustness and generalization ability of our method.", + "bbox": [ + 75, + 444, + 472, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 646, + 218, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformer. Transformers have been widely used in natural language processing and achieved great success on many language tasks. Recently, Vision Transformers (ViTs) have aroused extensive interest in computer vision due to their competitive performance compared with CNNs [16, 43, 31, 10]. Dosovitskiy et al. [16] firstly introduced transformers into the image classification task. They split the input image into non-overlapped patches and then feed them into the transformer encoders. Liu et al. [31] proposed a shifted windowing scheme to produce hierarchical feature maps suitable for dense prediction tasks. The great potential of vision transformer has motivated its adaptation to many challenging tasks including object detection [12, 56, 5], segmentation [9, 39], image enhancement [6, 27] and video understanding [33, 1].", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, some efforts have been devoted to producing better training targets to improve the performance of vision transformers [23, 44]. For example, DeiT [43] introduces a knowledge distillation procedure to reduce the training cost of ViTs and achieves a better accuracy/speed trade-off. TokenLabeling [23] employs a pretrained teacher annotator to predict a label for each token for dense knowledge distillation. Differently, we do not require a pretrained network to obtain the training targets. Our TL-Align maintains an aligned label for each token layer by layer and can be trained efficiently in an end-to-end manner.", + "bbox": [ + 496, + 444, + 890, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data Mixing Strategy. As an important type of data augmentation, data mixing strategies have demonstrated a consistent improvement in the generalization performance of CNNs. Zhang et al. [53] first proposed to combine a training pair to create augmented samples for model regularization. They perform linear interpolations on both the input images and associated targets. Following MixUp, CutMix [52] also utilizes the mixture of two input images but adopts a region copy-and-paste operation. Later methods including Puzzle Mix [24], SaliencyMix [45] and Attentive CutMix [48] leverage the salient regions for informative mixture generation. AlignMixUp [46] aligns two images geometrically in the feature space to improve the mixing representations. Recently, Yang et al. [51] proposed a RecursiveMix strategy which employs the historical input-prediction-label triplets for scale-invariant feature learning. Despite the better performance, a drawback of these methods is the heavily increased training cost due to the saliency extraction or historical information exploitation.", + "bbox": [ + 496, + 613, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5496", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most existing data mixing methods are originally designed for CNNs, and their effectiveness on ViTs has not been well explored. TransMix [7] uses the class attention map at the last layer to re-weight the mixing targets and assumes the output tokens to keep spatial correspondence with the input tokens. However, we identify a token fluctuation phenomenon for ViTs which may cause a mismatch between tokens and labels, leading to inaccurate label assignments in both the original CutMix and TransMix. To address this, we propose to align the label and token space by tracing their correspondence in a layerwise manner.", + "bbox": [ + 75, + 90, + 472, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 76, + 271, + 267, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 296, + 217, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The convolution neural network (CNN) has been the dominant architecture for computer vision in the deep learning era, greatly improving the performance of many tasks. Its monopoly has been challenged by the recent emergence of vision transformers (ViTs), which first \"patchify\" each image into tokens and process them with alternating self-attention (SA) and multi-layer perceptron (MLP).", + "bbox": [ + 75, + 320, + 468, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to architecture design, training strategy also has a large effect on the model performance, especially the data augmentation strategy. Data mixing [48, 45, 24, 47, 52, 53] is an important set of data augmentation for the training of both CNNs and ViTs, as it significantly improves the generalization ability of models. As the most commonly used data mixing strategy, CutMix [52] aims to create virtual training samples from the given training samples $(\\mathbf{X}, y)$ , where $\\mathbf{X} \\in \\mathcal{R}^{H \\times W \\times C}$ denotes the input image and $y$ is the corresponding label. CutMix randomly selects a local region from one input $\\mathbf{X}_1$ and uses it to replace the pixels in the same region of another input $\\mathbf{X}_2$ to generate a new sample $\\tilde{\\mathbf{X}}$ . Similarly, the label $\\tilde{y}$ of $\\tilde{\\mathbf{x}}$ is also the combination of the original labels $y_1$ and $y_2$ :", + "bbox": [ + 75, + 426, + 472, + 638 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {X}} = \\mathbf {M} \\odot \\mathbf {X} _ {1} + (\\mathbf {1} - \\mathbf {M}) \\odot \\mathbf {X} _ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 648, + 468, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} = \\lambda y _ {1} + (1 - \\lambda) y _ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 671, + 313, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M \\in \\{0,1\\}^{H \\times W}$ is a binary mask indicating the image each pixel belongs to, $\\mathbf{1}$ is an all-one matrix, and $\\odot$ is the element-wise multiplication. $\\lambda$ reflects the mixing ratio of two labels and is the proportion of pixels cropped from $\\mathbf{X}_1$ in the mixed image $\\widetilde{\\mathbf{X}}$ . For a cropped region $[r_x, r_x + r_w] \\times [r_y, r_y + r_h]$ from $\\mathbf{X}_1$ , we compute $\\lambda = \\frac{r_w r_h}{WH}$ to obtain the initial mixed target $\\tilde{y}$ .", + "bbox": [ + 76, + 699, + 468, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. The Token Fluctuation Phenomenon", + "text_level": 1, + "bbox": [ + 76, + 816, + 393, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CutMix is originally designed for CNNs and assumes the feature extraction process does not alter the mixing ratio. However, we discover that different from CNNs, self-attention in ViTs can lead to the fluctuation of some tokens.", + "bbox": [ + 75, + 839, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The fluctuation further results in the mismatch between the token space and label space, which hinders the effective training of the network.", + "bbox": [ + 496, + 90, + 890, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, we use $\\mathbf{z}_i$ to denote a token of the image $\\mathbf{Z}$ , i.e., $\\mathbf{z}_i$ is the transposed $i$ -th column vector of $\\mathbf{Z}$ . We can then compute the $i$ -th transformed token $\\hat{\\mathbf{z}}_i$ after the spatial operation as $\\hat{\\mathbf{z}}_i = \\sum_{j=1}^{N} w_{i,j}^s \\mathbf{z}_j$ , where $w_{i,j}^s$ is the $i, j$ -th element of the computed spatial mixing matrix $\\mathbf{w}^s(\\mathbf{z})$ .", + "bbox": [ + 496, + 136, + 890, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the assumption of the linear information integration, we define the contribution of an original token $\\mathbf{z}_i$ to a mixed token $\\hat{\\mathbf{z}}_j$ as $c(\\mathbf{z}_i,\\hat{\\mathbf{z}}_j) = \\frac{|w_{i,j}^s|}{\\sum_{k = 1}^{N}|w_{k,j}^s|}$ , where $|\\cdot |$ denotes the absolute value. We can then compute the presence of a token $\\mathbf{z}_i$ in all the mixed image tokens as:", + "bbox": [ + 496, + 212, + 890, + 295 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {z} _ {i}\\right) = \\sum_ {j = 1} ^ {N} c \\left(\\mathbf {z} _ {i}, \\hat {\\mathbf {z}} _ {j}\\right) = \\sum_ {j = 1} ^ {N} \\frac {\\left| w _ {i , j} ^ {s} \\right|}{\\sum_ {k = 1} ^ {N} \\left| w _ {k , j} ^ {s} \\right|}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 308, + 890, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For non-strided depth-wise convolution, each token is multiplied by each element in the convolutional kernel due to the translation invariance. We can thus obtain:", + "bbox": [ + 496, + 362, + 890, + 406 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {l = 1} ^ {N} \\left| w _ {i, l} ^ {s} \\right| = \\sum_ {j = 1} ^ {N} \\left| w _ {k, j} ^ {s} \\right| = \\sum_ {k = 1, l = 1} ^ {M} \\left| K _ {k, l} \\right|, \\quad \\forall i, j \\in \\mathbf {P} _ {N E}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 417, + 890, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{P}_{NE}$ denotes the set of positions that are not at the edge of the image, $K_{k,l}$ denotes the value of the $k,l$ -th position of the convolution kernel $\\mathbf{K}$ and $M$ is the kernel size. We can infer that $p(\\mathbf{z}_i) = 1$ , $\\forall i \\in \\mathbf{P}_{NE}$ , i.e., the effect of all the internal tokens does not change during the convolution process. However, for self-attention in ViTs, (3) does not hold due to the non-existence of translation invariance. The fluctuation of $p(\\mathbf{z})$ is further amplified by the input dependency of the spatial mixing matrix $\\mathbf{w}^s (\\mathbf{z})$ induced by self-attention. As an extreme case, we may obtain $p(\\mathbf{z}) \\sim 0$ for certain tokens. The fluctuation of tokens will alter the proportion of mixing (i.e., $\\lambda$ ) and the network might even completely ignore one of the mixed images. The actual label of the processed tokens can then deviate from the mixed label computed by (1), resulting in less effective training.", + "bbox": [ + 496, + 474, + 892, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Token-Label Alignment", + "text_level": 1, + "bbox": [ + 500, + 710, + 718, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each token in ViTs interacts with other tokens using the self-attention mechanism. The input-dependent weights empower ViTs with more flexibility but also result in a mismatch between the processed token and the initial token. To address this, we propose a token-label alignment (TL-Align) method to trace the correspondence between the input and transformed tokens to obtain the aligned labels for the resulting representations, as illustrated in Figure 2.", + "bbox": [ + 496, + 733, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, ViTs first split the mixed input $\\tilde{\\mathbf{X}}$ after CutMix (1) to a sequence of $N$ non-overlapped patches and then flatten them to obtain the original image tokens", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5497", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/293d805b64d2790b2045bbebfb4c2817de8ae864f0940b4bd76a5fa85ce1fba2.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed TL-Align. We trace the correspondence between the input tokens and the transformed tokens and align the labels accordingly. We reuse the computed attentions to linearly mix the labels of input tokens to obtain those of transformed tokens. The token-label alignment is performed iteratively to obtain a label for each output token." + ], + "image_footnote": [], + "bbox": [ + 112, + 85, + 851, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\{\\tilde{\\mathbf{x}}_1,\\tilde{\\mathbf{x}}_2,\\dots ,\\tilde{\\mathbf{x}}_N\\}$ . We then project them into a proper dimension and add positional embeddings:", + "bbox": [ + 76, + 368, + 470, + 398 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} ^ {0} = [ \\tilde {\\mathbf {z}} _ {c l s}; \\tilde {\\mathbf {x}} _ {1} \\cdot \\mathbf {E}; \\tilde {\\mathbf {x}} _ {2} \\cdot \\mathbf {E}; \\dots ; \\tilde {\\mathbf {x}} _ {N} \\cdot \\mathbf {E} ] + \\mathbf {E} _ {p o s}, (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 417, + 468, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{z}}_{cls} \\in \\mathcal{R}^{1 \\times d}$ denotes the class token, $\\mathbf{N}$ is the number of tokens, $\\mathbf{E}$ represents the patch projector, and $\\mathbf{E}_{pos} \\in \\mathcal{R}^{(N + 1) \\times d}$ is the position embeddings. Note that we adopt the process of the original transformer architecture [16] as an example without loss of generality. Other models may omit the class token and use a relative positional embedding instead, which does not affect the utility of the proposed TL-Align method.", + "bbox": [ + 75, + 455, + 468, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first assign each token $\\mathbf{z}_i\\in \\mathcal{R}^{1\\times d}$ with a label embedding $\\mathbf{y}_i\\in \\mathcal{R}^{1\\times C}$ :", + "bbox": [ + 75, + 580, + 468, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} ^ {0} = \\left[ \\tilde {\\mathbf {y}} _ {c l s} ^ {0}; \\tilde {\\mathbf {y}} _ {1} ^ {0}; \\tilde {\\mathbf {y}} _ {2} ^ {0}; \\dots ; \\tilde {\\mathbf {y}} _ {N} ^ {0} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 631, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the sum of elements in each $\\mathbf{y}_i$ equals 1 (i.e., $\\sum_{j=1}^{C} y_{i,j} = 1$ ) and $y_{i,j}$ indicates how much the $i$ -th token belong to the $j$ -th class. We initialize the label embedding following the conventional data mixing paradigm. For example, when using CutMix to mix two images $\\mathbf{X}_1$ and $\\mathbf{X}_2$ from the $j$ -th class and the $k$ -th class with a mixing ratio of $\\lambda$ , we set $\\tilde{y}_{cls,j} = \\lambda$ and $\\tilde{y}_{cls,k} = 1 - \\lambda$ for the class token. For each patch token, we set $\\tilde{y}_{i,j} = 1$ if it comes from $\\mathbf{X}_1$ and $\\tilde{y}_{i,k} = 1$ if it comes from $\\mathbf{X}_2$ . If a patch token contains both the mixed images, we use the mixing ratio within this patch as the label. For MixUp, we can simply set all label embeddings $\\{\\tilde{\\mathbf{y}}_i\\}$ with $\\tilde{y}_{j,k} = \\lambda$ and $\\tilde{y}_{j,j} = 1 - \\lambda$ .", + "bbox": [ + 75, + 669, + 470, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We perform TL-Align in a layer-wise manner and compute the aligned labels based on the operation on the tokens. Formally, ViTs use self-attention to perform spatial mixing", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the input tokens $\\mathbf{Z}$", + "bbox": [ + 498, + 368, + 648, + 383 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathbf {Z} \\cdot \\mathbf {W} _ {\\mathbf {Q}}, \\mathbf {K} = \\mathbf {Z} \\cdot \\mathbf {W} _ {\\mathbf {K}}, \\mathbf {V} = \\mathbf {Z} \\cdot \\mathbf {W} _ {\\mathbf {V}},\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 388, + 846, + 406 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} (\\mathbf {Q}, \\mathbf {K}) = \\operatorname {S o f t m a x} \\left(\\mathbf {Q} \\cdot \\mathbf {K} ^ {T} / \\sqrt {d}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 409, + 890, + 428 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Z}} = \\operatorname {S A} (\\mathbf {Z}) = \\mathcal {A} (\\mathbf {Q}, \\mathbf {K}) \\cdot \\mathbf {V}.\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 431, + 745, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To align the labels, we update the label embeddings $\\mathbf{Y}$ using the same attention matrix $\\mathcal{A}(\\mathbf{Q},\\mathbf{K})$", + "bbox": [ + 498, + 455, + 890, + 486 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Y}} = \\mathcal {A} (\\mathbf {Q}, \\mathbf {K}) \\cdot \\mathbf {Y}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 492, + 890, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ViTs usually adopt multi-head self-attention (MSA) to perform multiple self-attention parallelly:", + "bbox": [ + 498, + 517, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Z}} = \\operatorname {M S A} (\\mathbf {Z}) = \\left[ \\mathrm {S A} _ {1} (\\mathbf {Z}); \\mathrm {S A} _ {2} (\\mathbf {Z}); \\dots ; \\mathrm {S A} _ {H} (\\mathbf {Z}) \\right] \\cdot \\mathbf {w} _ {h}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 553, + 890, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H$ is the number of heads and $\\mathbf{w}_h\\in \\mathcal{R}^{d\\times d}$ . We then adapt our label alignment to MSA by simply taking the average of all the attention matrices for alignment:", + "bbox": [ + 498, + 584, + 890, + 631 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Y}} = \\operatorname {T L} - \\operatorname {A l i g n} - \\mathbf {S} (\\mathbf {Z}, \\mathbf {Y}) := \\frac {1}{H} \\sum_ {i = 1} ^ {H} \\mathcal {A} _ {i} (\\mathbf {Q}, \\mathbf {K}) \\cdot \\mathbf {Y}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 636, + 890, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{A}_i$ is the attention matrix corresponding to the $i$ -th head $\\mathrm{SA}_i$ .", + "bbox": [ + 498, + 683, + 890, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each transformer block $l$ processes the tokens by both spatial and channel mixing:", + "bbox": [ + 498, + 714, + 890, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Z}} ^ {l - 1} = \\operatorname {M S A} (\\mathrm {L N} (\\mathbf {Z} ^ {l - 1})), \\quad \\mathbf {Z} ^ {\\prime l - 1} = \\hat {\\mathbf {Z}} ^ {l - 1} + \\mathbf {Z} ^ {l - 1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 750, + 890, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Z}} ^ {l} = \\operatorname {M L P} (\\operatorname {L N} \\left(\\mathbf {Z} ^ {\\prime l - 1}\\right)), \\quad \\mathbf {Z} ^ {l} = \\hat {\\mathbf {Z}} ^ {l} + \\mathbf {Z} ^ {\\prime l - 1},\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 771, + 803, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where MLP and LN denote the MLP module and layer normalization [2], respectively. Our TL-Align then aligns the label embeddings in a similar manner:", + "bbox": [ + 498, + 795, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Y}} ^ {l - 1} = \\operatorname {T L} \\text {- A l i g n - S} \\left(\\mathbf {Y} ^ {l - 1}\\right), \\mathbf {Y} ^ {\\prime l - 1} = \\operatorname {N o r m} \\left(\\hat {\\mathbf {Y}} ^ {l - 1} + \\mathbf {Y} ^ {l - 1}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 847, + 903, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Y}} ^ {l} = \\mathbf {Y} ^ {\\prime l - 1}, \\quad \\mathbf {Y} ^ {l} = \\operatorname {N o r m} \\left(\\hat {\\mathbf {Y}} ^ {l} + \\mathbf {Y} ^ {\\prime l - 1}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 868, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5498", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Norm denotes the normalization operation. We implement Norm by a simple average.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hierarchical vision transformers such as Swin [31] further introduce a patch aggregation operation to merge multiple patches. They usually concatenate multiple tokens across the channels to reduce the spatial resolution. Instead of concatenation, we simply add the label embeddings of the merged tokens followed by normalization as the aligned labels. The proposed TL-Align can be generalized to different architectures composed of spatial mixing, channel mixing, point-wise transformation, residual connection, and spatial aggregation. We provide detailed illustrations of alignment with these operations in the supplementary.", + "bbox": [ + 75, + 121, + 468, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We synchronously align the labels with the processed tokens layer by layer and obtain the aligned tokens $\\mathbf{Z}^L$ and labels $\\mathbf{Y}^L$ . The final representation of the image $\\mathbf{z}$ is either the class token $\\mathbf{z}_{cls}^L$ [16, 43] or the average pooling of all the spatial tokens $\\frac{1}{N}\\sum_{i=1}^{N}\\mathbf{z}_i^L$ [31]. The aligned label $\\mathbf{y}_{align}$ for the image is then $\\mathbf{y}_{cls}^L$ or $\\frac{1}{N}\\sum_{i=1}^{N}\\mathbf{y}_i^L$ depending on the specific model. We then adopt the aligned label $\\mathbf{y}_{align}$ to train the network and can adapt to different loss functions and training schemes:", + "bbox": [ + 75, + 287, + 470, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nJ = J (\\mathbf {z}, \\text {s t o p - g r a d i e n t} (\\mathbf {y} _ {\\text {a l l i n g}})). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 439, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We do not back-propagate through the aligned label as they only serve as a more accurate target.", + "bbox": [ + 75, + 465, + 468, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our TL-Align serves as a plug-and-play module on various vision transformers while only introducing negligible training costs. We adjust the label of each token adaptively during the layer-by-layer propagation and preserve alignment between tokens and labels throughout the forward process. TL-Align is only used during training and introduces no additional computation cost when inference.", + "bbox": [ + 75, + 497, + 468, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 616, + 209, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conducted extensive experiments to evaluate the proposed TL-Align method. We demonstrate the improvement of TL-Align on various vision transformers and compare it with state-of-the-art training strategies concerning the accuracy, network complexity, and training speed. We also examine the transferability on downstream tasks including semantic segmentation, object detection, and transfer learning. We further provide in-depth analysis to evaluate the effectiveness of TL-Align.", + "bbox": [ + 75, + 641, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. ImageNet Classification", + "text_level": 1, + "bbox": [ + 76, + 786, + 294, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We first evaluate TL-Align on ImageNet [38] for image classification. We implement our method based on PyTorch [35] and the timm library [50]. We conduct experiments on various transformer architectures: three variants of DeiT [43] (DeiT-T, DeiT-S, and DeiT-B), two variants of PVT [49] (PVT-T, PVT-S) and", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/dc40af6d1ce3a61a9ede21307db38057b4fb1e8caf8790e1814ecfcff0c89f48.jpg", + "table_caption": [ + "Table 1. Results on ImageNet classification task. We compare the parameters, FLOPs, and accuracy of different vision transformer backbones without and with our TL-Align." + ], + "table_footnote": [], + "table_body": "
ModelImage SizeParamsFLOPsTop-1(%)Top-5(%)
DeiT-T22425.7M1.6G72.291.3
+TL-Align73.291.7
PVT-T224213.2M1.9G75.192.4
+TL-Align75.593.0
DeiT-S224222M4.6G79.895.0
+TL-Align80.695.0
PVT-S224224.5M3.8G79.895.0
+TL-Align80.495.5
Swin-T224229M4.5G81.295.5
+TL-Align81.495.7
Swin-S224250M8.8G83.096.3
+TL-Align83.496.5
DeiT-B224286M17.5G81.895.5
+TL-Align82.395.8
Swin-B224288M15.4G83.596.4
+TL-Align83.796.5
", + "bbox": [ + 501, + 130, + 901, + 395 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2651409140fa97e067440fbb4abbc5ffb7a86ef74a683d5ce133158025ee9d37.jpg", + "table_caption": [ + "Table 2. Comparison of our TL-Align with other training strategies on ImageNet." + ], + "table_footnote": [], + "table_body": "
MethodParamsTrainig TimeAcc.(%)
Vanilla22M1×76.4
CutMix [52]22M1×79.8
Puzzle-Mix [24]22M2.24×79.8
F-Mix [18]22M1.08×77.4
ResizeMix [36]22M1.03×78.6
SaliencyMix [45]22M1.26×79.2
Attentive-CutMix [48]46M1.31×77.5
TransMix [7]22M1.16×80.1
TokenMix [30]22M1.65×80.3
CutMix + TL-Align22M1.24×80.6
", + "bbox": [ + 517, + 430, + 870, + 611 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/57af12b6b7aed1baed4fb5518f622093680d26f0716e36e493e0e7b57e1743c2.jpg", + "table_caption": [ + "Table 3. Results on semantic segmentation ADE20K." + ], + "table_footnote": [], + "table_body": "
BackboneParamsFLOPsmIoUmIoU (MS)mAcc
DeiT-S58M1032G43.845.155.2
+TL-Align44.545.755.5
Swin-T60M945G44.445.855.6
+TL-Align44.746.556.4
Swin-S81M1038G47.649.558.8
+TL-Align48.049.759.5
Swin-B121M1188G48.149.759.1
+TL-Align48.350.159.7
", + "bbox": [ + 513, + 631, + 875, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "three variants of Swin Transformer [31] (Swin-T, Swin-S, and Swin-B). For tiny and small models, we train from scratch for 300 epochs following the same training recipe using CutMix as DeiT [43], PVT [49] and Swin [31]. We keep all the data augmentation policies and all hyperparameter settings unchanged for fair comparisons. The only modification is that we replace the mixing targets in Cut", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5499", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/edc30d1a787d1ea1bd14da599d54693b20d85b263f18db9fc90af5486b3c6d5e.jpg", + "table_caption": [ + "Table 4. Experimental results on object detection and instance segmentation on COCO." + ], + "table_footnote": [], + "table_body": "
BackboneParamsFLOPsSchedule\\(AP^{box}\\)\\(AP^{box}_{50}\\)\\(AP^{box}_{75}\\)\\(AP^{mask}\\)\\(AP^{mask}_{50}\\)\\(AP^{mask}_{75}\\)
Swin-T+TL-Align86M745G3x50.469.254.743.766.647.3
50.569.454.943.866.647.3
Swin-S+TL-Align107M838G3x51.970.756.345.068.248.8
52.271.156.745.268.449.1
Swin-B+TL-Align145M982G3x51.970.556.445.068.148.9
52.371.256.945.368.749.1
", + "bbox": [ + 86, + 103, + 879, + 229 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/670875c6252a1071265dafff7b19b0f2a5103e87c2593bc2df82efb1fcfba067.jpg", + "table_caption": [ + "Table 5. Results on different transfer learning datasets." + ], + "table_footnote": [], + "table_body": "
ModelParamsFLOPsC-10C-100FlowersCars
ResNet5026M4.1G--96.290.0
ViT-B/1686M55.4G98.187.189.5-
ViT-L/16307M190.7G97.986.489.7-
DeiT5.7M1.6G97.685.797.190.1
+TL-Align5.7M1.6G97.886.497.990.7
DeiT-S22M4.6G97.990.298.191.4
+TL-Align22M4.6G98.890.498.391.8
DeiT-B86M17.5G99.190.898.492.1
+TL-Align86M17.5G99.190.598.693.0
", + "bbox": [ + 84, + 250, + 459, + 409 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fec01cb0a93a68f3ac2faae047365d102018f2fef6fdadb42d492ced3e204b1a.jpg", + "image_caption": [ + "Figure 3. The Root Mean Square Error (RMSE) between original CutMix targets and labels obtained by T-L Align." + ], + "image_footnote": [], + "bbox": [ + 83, + 415, + 457, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MiX with the labels obtained by our TL-Align. For base models (i.e., Deit-B and Swin-B), we finetune the official pre-trained models for 40 epochs with a constant learning rate of 1e-5 and a weight decay of 1e-8.", + "bbox": [ + 75, + 642, + 468, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on Different Architectures. As shown in Table 1, TL-Align steadily improves the performance of different vision transformer architectures. Specifically, TL-Align boosts the top-1 accuracy of DeiT-T, DeiT-S, and DeiT-B by $1.0\\%$ , $0.8\\%$ , and $0.5\\%$ , respectively, in a parameter-free manner. Moreover, our method is generalizable and can be directly applied to hierarchical vision transformers like Swin. It is worth noting that most existing methods need either architecture modifications (adding a class token in [7]) or extra computations (saliency map extraction in [45]) when applied to Swin. In contrast, our TL-Align method can be used as a plug-and-play module and achieves consistent improvement on variants of Swin.", + "bbox": [ + 75, + 704, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/792f8205bf4f9fa77f216fc97556f10c8eb50c38c5b774ff62be9edf7ad55d41.jpg", + "image_caption": [ + "Figure 4. Visualization of mixing ratio $\\lambda$ of fluctuating tokens from different layers. We compare the results of TL-Align with CutMix, token similarity, TransMix, and TokenLabeling." + ], + "image_footnote": [], + "bbox": [ + 517, + 234, + 867, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Other Training Strategies. We also compare our method with the state-of-the-art training strategies for data mixing on DeiT-S, as shown in Table 2. Specifically, we train the DeiT-S model while only disabling CutMix as the baseline method, which is denoted as Vanilla. Moreover, since TransMix [7] and TokenMix [30] report the EMA accuracy with different hyperparameters, we reproduce it under the same training recipe [43] for a fair comparison. We see that TL-Align shows better performance than the other mixup variants while maintaining the number of parameters and training speed. Puzzle-Mix obtains the same classification accuracy as CutMix but results in a much lower training speed as it relies on an extra model to get the optimal solution. SaliencyMix and Attentive-CutMix lead to performance degeneration when built upon DeiT-S backbone. Notably, our method also achieves higher top-1 accuracy than ViT-targeted TransMix and TokenMix. While the teacher annotator in TokenMix introduces heavy training costs, TL-Align obtains accurate token-label alignment in an efficient manner. Moreover, TransMix has limitations in handling token fluctuation and cannot be applied to ViTs without a class token, such as Swin.", + "bbox": [ + 496, + 474, + 892, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Downstream Tasks", + "text_level": 1, + "bbox": [ + 500, + 816, + 681, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic Segmentation. We evaluate our TL-Align on ADE20K dataset [55] for semantic segmentation. We adopt DeiT-S and three variants of Swin Transformer as backbones equipped with UpperNet for segmentation. As pre", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5500", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8e9c3962b662580f7ba72edd7c02f145e24a0ae4f184476121bab7dee78529e1.jpg", + "table_caption": [ + "Table 6. Comparison results of model generalization ability and robustness. We evaluate them on various out-of-distribution/corrupted datasets and against adversarial attacks. $\\uparrow$ denotes higher is better, and $\\downarrow$ denotes lower is better." + ], + "table_footnote": [], + "table_body": "
ModelParamsFLOPsImageNetGeneralizationRobustness
Top-1↑Top-5↑IN-V2↑IN-A↑IN-C↓IN-R↑AutoAttack↑
DeiT-T5.7M1.6G72.291.360.47.769.134.13.9
+TL-Align5.7M1.6G73.291.761.46.168.034.64.4
DeiT-S22M4.6G79.895.068.518.954.742.56.9
+TL-Align22M4.6G80.695.068.919.253.243.27.5
DeiT-B86M17.5G81.895.570.527.948.545.3-
+TL-Align86M17.5G82.395.870.929.047.144.4-
", + "bbox": [ + 104, + 117, + 859, + 263 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4661141df34944a9ee63e80676adb1d7cd48e3e966c070b36e2d4cc76ecfe8b3.jpg", + "table_caption": [ + "Table 7. Ablation of using different data mixing strategies." + ], + "table_footnote": [], + "table_body": "
MixUpCutMixRandomBlock-wiseTop-1 (%)+TL-Align Top-1 (%)
××××76.4-
×✓××79.880.6
✓✓××79.880.2
××✓×79.780.2
×××✓80.080.3
", + "bbox": [ + 78, + 282, + 480, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sented in Table 3, TL-Align improves the segmentation performance on both DeiT and Swin at different model scales.", + "bbox": [ + 76, + 398, + 467, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Object Detection and Instance Segmentation. We also examine the performance of TL-Align on object detection and instance segmentation on the COCO 2017 dataset [29]. We apply our TL-Align to Swin [31] due to the advantage of the hierarchical representations on object detection tasks. We adopt the Cascade Mask-RCNN [4] framework and use the training strategy of $3 \\times$ schedule. As shown in Table 4, we observe consistent improvements on all variants of Swin Transformer. This demonstrates the advantages to learn token-level meaningful features suitable for dense prediction tasks.", + "bbox": [ + 75, + 429, + 467, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Transfer Learning. We further evaluate the transferred classification performance of TL-Align on CIFAR-10 [26], CIFAR-100 [26], Flowers [34] and Cars [25]. We use pretrained models on ImageNet and finetune them on these datasets following existing works [43]. We compare the performance with and without TL-Align on three variants of DeiT [43], as shown in Table 5. TL-Align obtains significant performance gains for all variants on the four datasets.", + "bbox": [ + 75, + 595, + 467, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Performance Analysis and Visualization", + "text_level": 1, + "bbox": [ + 76, + 726, + 419, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Token-Label Alignment. We first quantize the difference between the original targets and aligned labels and investigate its correlation with the model. Specifically, we compute the Root Mean Square Error (RMSE) between the original targets and labels obtained by our TL-Align. As shown in Figure 3, the RMSE decreases when enlarging the model size. This indicates that larger models demonstrate less token fluctuation since the self-correction ability is also enhanced as the model capacity scales up. Moreover, the RMSE for Swin Transformer", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/17d7926aafdecf807109451487201208a5c4ba8d8bd8e309147e99512a8bc82f.jpg", + "table_caption": [ + "Table 8. Ablation of different TL-Align operations." + ], + "table_footnote": [], + "table_body": "
AlignmentTop-1 Acc.(%)Δ (%)
None (DeiT-S baseline)79.8-
TL-Align-S (Layer 12)80.1+0.3
TL-Align-S (Layer 2,4,6,8)80.2+0.4
Normalization Disabled80.3+0.5
Default (TL-Align)80.6+0.8
", + "bbox": [ + 531, + 282, + 857, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tends to be lower compared with DeiT of similar model size. This is due to the local-window self-attention in Swin which preserves more local information. These observations are consistent with our experimental results: the improvements on small models and DeiT-like backbones tend to be more significant as they encounter more token fluctuation.", + "bbox": [ + 496, + 383, + 890, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of the Layer-wise Mixing Ratio of Fluctuated Tokens. To investigate the effectiveness of TL-Align, we compute a similarity-based \"ground-truth\" mixing ratio for each layer. Specifically, we compute the similarities of tokens between the mixed and unmixed images and use them as the label of each token. We compare them with the mixing ratios produced by TL-Align, CutMix [52], TransMix [7], and TokenLabeling [23]. As shown in Figure 4, the similarity-based mixing ratio changes at each layer, resulting from token fluctuation. However, CutMix, TransMix, and TokenLabeling assume the output tokens keep spatial correspondence with the input tokens and compute a fixed mixing ratio. TL-Align assigns dynamic labels to tokens using layer-wise alignment, which is more accurate compared with other methods.", + "bbox": [ + 496, + 477, + 890, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation of Robustness and Generalization. We further conduct experiments to validate the generalization and robustness of TL-Align, as shown in Table 6. We employ four corrupted and out-of-distribution datasets for robustness evaluation, including ImageNet-A [22], ImageNet-C [21], and ImageNet-R [20]. We also adopt AutoAttack [11] to evaluate the adversarial robustness on the ImageNet validation set. Due to memory limitation, we do not experiment with DeiT-B on AutoAttack. We use mean Corruption Error (mCE) for ImageNet-C and Top-1 Accuracy for others as the evaluation metric. For generalization evaluation, we adopt the ImageNet-V2 dataset [37]. We see that TL-Align improves both robustness and general", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5501", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8bc8153a112392f9b8621f9444fda5e2399360da810779c8a2e58b6a808cfcd.jpg", + "image_caption": [ + "Figure 5. The visualization of results on DeiT-S and Swin-S. We visualize the input images, the mixed image, the original label embedding, and the label embedding after token-label alignment." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 883, + 315 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ization, showing the superiority of adopting TL-Align for pre-training.", + "bbox": [ + 75, + 357, + 468, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study on different Data Mixing Strategies. Due to the efficiency of the proposed layer-wise alignment, TL-Align can be directly applied to a wide range of data mixing strategies. We adopt MixUp, CutMix, a random mixing strategy and a block-wise mixing strategy to evaluate the generalizability of TL-Align. The random mixing and block-wise mixing strategies are inspired by MAE [19] and BEiT [3] and we replace the masking operation with image mixing on patch-level and block-level (both of size $16 \\times 16$ ) respectively. The comparison results of training DeiT-S with and without our approach is demonstrated in Table 7. Specifically, TL-Align improves CutMix by $0.8\\%$ , MixUp+CutMix by $0.4\\%$ , random mixing by $0.5\\%$ and block-wise mixing by $0.3\\%$ respectively, further verifying the generalizability of the proposed TL-Align.", + "bbox": [ + 75, + 387, + 468, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study on Different Label Alignment Operations. Our TL-Align aligns the labels with tokens transformed by spatial self-attention and residual connection layer-by-layer. To investigate the effect of reusing attention maps and normalization, we conduct an ablation study regarding different alignment operations on DeiT-S. We try aligning the labels only by using the attention map of Layer 12, which is equivalent to TransMix [7]. We also test the performance of applying alignment to several middle transformer layers and disabling normalization. As presented in Table 8, incomplete alignment at a part of layers marginally boosts the performance as it cannot well handle the token fluctuation issue. Disabling normalization leads to $0.3\\%$ accuracy drop due to the inaccurate alignment at the presence of residual connections. This demonstrates the significance of the token-label alignment by attention utilization and normalization in a layer-wise manner.", + "bbox": [ + 75, + 614, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualizations of Aligned Labels. We visualize the labels obtained by TL-Align on DeiT-S [43] and Swin-S [31]", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "as shown in Figure 5. Specifically, the aligned label embedding is obtained after the final transformer block for both DeiT-S and Swin-S. The value of the label embedding represents the probability of the belonged class of the corresponding token. We use red to denote larger probabilities towards the first image and blue for the second image. We observe that the aligned labels can deviate from the original labels and result in different mixing ratios for training. Therefore, using the original ratio as the training target may produce false training signals and lead to inferior performance. We see that TL-Align can correct the labels when the images are mixed with uninformative tokens. More visualization results are included in the supplementary.", + "bbox": [ + 496, + 357, + 890, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 571, + 617, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have presented a token-label alignment method for training better vision transformers. As important subsets of data augmentation methods, data mixing strategies can generally improve the performance of both CNNs and ViTs. We identify a token fading issue for ViTs and address it by tracing the correspondence between transformed tokens and the original tokens to obtain a label for each output token to obtain more accurate training signals. Experimental results have demonstrated that our TL-Align can consistently improve the performance of various ViT models. The generalization performance of TL-Align to other architectures such as MLP-like models remains unknown and is a promising future direction to explore.", + "bbox": [ + 496, + 598, + 890, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 813, + 666, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160102 and in part by the National Natural Science Foundation of China under Grant 62125603.", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5502", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 2", + "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv, abs/1607.06450, 2016. 4", + "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.8", + "[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, pages 6154-6162, 2018. 7", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 2", + "[6] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310, 2021. 2", + "[7] Jie-Neng Chen, Shuyang Sun, Ju He, Philip Torr, Alan Yuille, and Song Bai. Transmix: Attend to mix for vision transformers. arXiv preprint arXiv:2111.09833, 2021. 1, 3, 5, 6, 7, 8", + "[8] Tianlong Chen, Zhenyu Zhang, Yu Cheng, Ahmed Awadallah, and Zhangyang Wang. The principle of diversity: Training stronger vision transformers calls for reducing all levels of redundancy. In CVPR, 2022. 1", + "[9] Bowen Cheng, Alexander G Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In NeurIPS, 2021. 1, 2", + "[10] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In NeurIPS, 2021. 1, 2", + "[11] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216, 2020. 7", + "[12] Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In ICCV, pages 2988-2997, 2021. 1, 2", + "[13] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In CVPR, pages 1601-1610, 2021. 1", + "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 2", + "[15] Xiaohan Ding, Xiangyu Zhang, Yizhuang Zhou, Jungong Han, Guiguang Ding, and Jian Sun. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 1", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 78, + 116, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 2, 4, 5", + "[17] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 1", + "[18] Ethan Harris, Antonia Marcu, Matthew Painter, Mahesan Niranjan, Adam Prügel-Bennett, and Jonathon Hare. Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047, 2020.5", + "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 8", + "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 7", + "[21] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 7", + "[22] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 7", + "[23] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. In NeurIPS, 2021. 1, 2, 7", + "[24] Jang-Hyun Kim, Wonho Choo, and Hyun Oh Song. Puzzle mix: Exploiting saliency and local statistics for optimal mixup. In ICML, pages 5275-5285, 2020. 1, 2, 3, 5", + "[25] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCVW, pages 554–561, 2013. 7", + "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7", + "[27] Wenbo Li, Xin Lu, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2", + "[28] Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, and Xiaou Tang. Not all pixels are equal: Difficulty-aware semantic segmentation via deep layer cascade. In CVPR, pages 3193-3202, 2017. 1", + "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 7", + "[30] Jihao Liu, Boxiao Liu, Hang Zhou, Hongsheng Li, and Yu Liu. Tokenmix: Rethinking image mixing for data augmentation in vision transformers. In ECCV, pages 455-471, 2022. 5, 6" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5503", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 5, 7, 8", + "[32] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. arXiv preprint arXiv:2201.03545, 2022. 1", + "[33] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. arXiv preprint arXiv:2106.13230, 2021. 2", + "[34] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 7", + "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NIPS, pages 8026-8037, 2019. 5", + "[36] Jie Qin, Jiemin Fang, Qian Zhang, Wenyu Liu, Xingang Wang, and Xinggang Wang. Resizemix: Mixing data with preserved object information and true labels. arXiv preprint arXiv:2012.11101, 2020. 5", + "[37] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning, pages 5389-5400, 2019. 7", + "[38] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 5", + "[39] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, 2021. 1, 2", + "[40] Yehui Tang, Kai Han, Jianyuan Guo, Chang Xu, Yanxi Li, Chao Xu, and Yunhe Wang. An image patch is a wave: Quantum inspired vision mlp. In CVPR, 2022. 1", + "[41] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. NeurIPS, 34, 2021. 1", + "[42] Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Gautier Izacard, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, et al. Resmlp: Feedforward networks for image classification with data-efficient training. arXiv preprint arXiv:2105.03404, 2021. 1", + "[43] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 1, 2, 5, 6, 7, 8", + "[44] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] AFM Uddin, Mst Monira, Wheemyung Shin, TaeChoong Chung, Sung-Ho Bae, et al. Saliencymix: A saliency guided data augmentation strategy for better regularization. arXiv preprint arXiv:2006.01791, 2020. 1, 2, 3, 5, 6", + "[46] Shashanka Venkataramanan, Ewa Kijak, Laurent Amsaleg, and Yannis Avrithis. Alignmixup: Improving representations by interpolating aligned features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19174-19183, 2022. 2", + "[47] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In ICML, pages 6438-6447, 2019. 1, 3", + "[48] Devesh Walawalkar, Zhiqiang Shen, Zechun Liu, and Marios Savvides. Attentive cutmix: An enhanced data augmentation approach for deep learning based image classification. arXiv preprint arXiv:2003.13048, 2020. 1, 2, 3, 5", + "[49] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. 5", + "[50] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.5", + "[51] Lingfeng Yang, Xiang Li, Borui Zhao, Renjie Song, and Jian Yang. Recursivelymix: Mixed learning with history. arXiv preprint arXiv:2203.06844, 2022. 2", + "[52] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, pages 6023-6032, 2019. 1, 2, 3, 5, 7", + "[53] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 1, 2, 3", + "[54] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 1", + "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. IJCV, 127:302-321, 2019. 6", + "[56] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 1, 2" + ], + "bbox": [ + 501, + 92, + 893, + 741 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5504", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Token-Label Alignment for Vision Transformers/ef46bf1f-34a3-4e70-bf1a-b3f5f27b821d_model.json b/2023/Token-Label Alignment for Vision Transformers/ef46bf1f-34a3-4e70-bf1a-b3f5f27b821d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dfcd6092daf402a3dafba87db23820fab25a1ff2 --- /dev/null +++ b/2023/Token-Label Alignment for Vision Transformers/ef46bf1f-34a3-4e70-bf1a-b3f5f27b821d_model.json @@ -0,0 +1,2211 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.241, + 0.131, + 0.73, + 0.154 + ], + "angle": 0, + "content": "Token-Label Alignment for Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.179, + 0.846, + 0.273 + ], + "angle": 0, + "content": "Han Xiao\\(^{1,2,*}\\) Wenzhao Zheng\\(^{1,2,*}\\) Zheng Zhu\\(^{3}\\) Jie Zhou\\(^{1,2}\\) Jiwen Lu\\(^{1,2,\\dagger}\\) \n\\(^{1}\\)Beijing National Research Center for Information Science and Technology, China \n\\(^{2}\\)Department of Automation, Tsinghua University, China \n\\(^{3}\\)PhiGent Robotics \n{h-xiao20, zhengwz18}@mails.tsinghua.edu.cn; zhengzhu@ieee.org; {jzhou, lujiwen} @tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.474, + 0.671 + ], + "angle": 0, + "content": "Data mixing strategies (e.g., CutMix) have shown the ability to greatly improve the performance of convolutional neural networks (CNNs). They mix two images as inputs for training and assign them with a mixed label with the same ratio. While they are shown effective for vision transformers (ViTs), we identify a token fluctuation phenomenon that has suppressed the potential of data mixing strategies. We empirically observe that the contributions of input tokens fluctuate as forward propagating, which might induce a different mixing ratio in the output tokens. The training target computed by the original data mixing strategy can thus be inaccurate, resulting in less effective training. To address this, we propose a token-label alignment (TL-Align) method to trace the correspondence between transformed tokens and the original tokens to maintain a label for each token. We reuse the computed attention at each layer for efficient token-label alignment, introducing only negligible additional training costs. Extensive experiments demonstrate that our method improves the performance of ViTs on image classification, semantic segmentation, objective detection, and transfer learning tasks. Code is available at: https://github.com/Euphorial6/TL-Align." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.701, + 0.21, + 0.716 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.47, + 0.864 + ], + "angle": 0, + "content": "The recent developments of vision transformers (ViTs) have revolutionized the computer vision field and set new state-of-the-arts in various tasks, such as image classification [16, 43, 31, 10], object detection [5, 56, 12, 13], and semantic segmentation [28, 39, 54, 9]. The successful structure of alternative spatial mixing and channel mixing in ViTs also motivates the arising of high-performance MLP-like deep architectures [40, 41, 42] and promotes the evolution of better CNNs [15, 32, 17]. In addition to architec" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.306, + 0.892, + 0.336 + ], + "angle": 0, + "content": "ture designs, an improved training strategy can also greatly boost the performance of a trained deep model [23, 44, 8, 7]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.893, + 0.627 + ], + "angle": 0, + "content": "The training of modern deep architecture almost all adopts data mixing strategies for data augmentation [48, 45, 24, 47, 52, 53], which have been proven to consistently improve the generalization performance. They randomly mix two images and their labels with the same mixing ratio to produce mixed data. As the most commonly used data mixing strategy, CutMix [52] performs a copy-and-paste operation on the spatial domain to produce spatially mixed images. While data mixing strategies have been widely studied for CNNs [48, 45, 24], few works have explored their compatibilities with ViTs [7]. We find that self-attention in ViTs causes a fluctuation of the original spatial structure. Unlike the translation equivalence that ensures a global label consistency for CNNs, self-attention in ViTs undermines this global consistency and causes a misalignment between the token and label. This misalignment induces a different mixing ratio in the output tokens. The training targets computed by the original data mixing strategies can then be inaccurate, resulting in less effective training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To address this, we propose a token-label alignment (TL-Align) method for ViTs to obtain a more accurate target for training. We present an overview of our method in Figure 1. We first assign a label to each input token in the mixed image according to the source of the token. We then trace the correspondence between the input tokens and the transformed tokens and align the labels accordingly. We assume that only the spatial self-attention and residual connection operation alter the presence of input tokens since channel MLP and layer normalization process each token independently. We reuse the computed attentions to linearly mix the labels of input tokens to obtain those of transformed tokens. The token-label alignment is performed iteratively to obtain a label for each output token. For class-token-based classification (e.g., ViT [16] and DeiT [43]), we directly use the aligned label for the output class token as the training target. For global-pooling-based classification (e.g., Swin [31]), we similarly average the labels of output tokens as the train" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.206, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5495" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.241, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.371, + 0.182, + 0.382 + ], + "angle": 0, + "content": "(a) CutMix" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.09, + 0.498, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.37, + 0.466, + 0.383 + ], + "angle": 0, + "content": "(b) Token misalignment for ViTs" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.887, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.37, + 0.833, + 0.383 + ], + "angle": 0, + "content": "(c) Different ways to obtain the mixed labels" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.385, + 0.895, + 0.442 + ], + "angle": 0, + "content": "Figure 1. An overview of the proposed TL-Align. (a) CutMix-like methods [52] are widely used in model training, which spatially mix the tokens and their labels in the input space. (b) They are originally designed for CNNs and assume the processed tokens are spatially aligned with the input tokens. We show that it does not hold true for ViTs due to the global receptive field and adaptive weights. (c) Compared with existing methods, our method can effectively and efficiently align the tokens and labels without requiring a pretrained teacher network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.473, + 0.626 + ], + "angle": 0, + "content": "ing target. The proposed TL-Align is only used for training to improve performance and introduces no additional workload for inference. We apply the proposed TL-Align to various ViT variants with CutMix including plain ViTs (DeiT [43]) and hierarchical ViTs (Swin [31]). We observe a consistent performance boost across different models on ImageNet-1K [14]. Specifically, our TL-Align improves DeiT-S by \\(0.8\\%\\) using the same training recipe. We evaluate the ImageNet-pretrained models on various downstream tasks including semantic segmentation, objection detection, and transfer learning. Experimental results also verify the robustness and generalization ability of our method." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.647, + 0.22, + 0.663 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Vision Transformer. Transformers have been widely used in natural language processing and achieved great success on many language tasks. Recently, Vision Transformers (ViTs) have aroused extensive interest in computer vision due to their competitive performance compared with CNNs [16, 43, 31, 10]. Dosovitskiy et al. [16] firstly introduced transformers into the image classification task. They split the input image into non-overlapped patches and then feed them into the transformer encoders. Liu et al. [31] proposed a shifted windowing scheme to produce hierarchical feature maps suitable for dense prediction tasks. The great potential of vision transformer has motivated its adaptation to many challenging tasks including object detection [12, 56, 5], segmentation [9, 39], image enhancement [6, 27] and video understanding [33, 1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Recently, some efforts have been devoted to producing better training targets to improve the performance of vision transformers [23, 44]. For example, DeiT [43] introduces a knowledge distillation procedure to reduce the training cost of ViTs and achieves a better accuracy/speed trade-off. TokenLabeling [23] employs a pretrained teacher annotator to predict a label for each token for dense knowledge distillation. Differently, we do not require a pretrained network to obtain the training targets. Our TL-Align maintains an aligned label for each token layer by layer and can be trained efficiently in an end-to-end manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Data Mixing Strategy. As an important type of data augmentation, data mixing strategies have demonstrated a consistent improvement in the generalization performance of CNNs. Zhang et al. [53] first proposed to combine a training pair to create augmented samples for model regularization. They perform linear interpolations on both the input images and associated targets. Following MixUp, CutMix [52] also utilizes the mixture of two input images but adopts a region copy-and-paste operation. Later methods including Puzzle Mix [24], SaliencyMix [45] and Attentive CutMix [48] leverage the salient regions for informative mixture generation. AlignMixUp [46] aligns two images geometrically in the feature space to improve the mixing representations. Recently, Yang et al. [51] proposed a RecursiveMix strategy which employs the historical input-prediction-label triplets for scale-invariant feature learning. Despite the better performance, a drawback of these methods is the heavily increased training cost due to the saliency extraction or historical information exploitation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5496" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.259 + ], + "angle": 0, + "content": "Most existing data mixing methods are originally designed for CNNs, and their effectiveness on ViTs has not been well explored. TransMix [7] uses the class attention map at the last layer to re-weight the mixing targets and assumes the output tokens to keep spatial correspondence with the input tokens. However, we identify a token fluctuation phenomenon for ViTs which may cause a mismatch between tokens and labels, leading to inaccurate label assignments in both the original CutMix and TransMix. To address this, we propose to align the label and token space by tracing their correspondence in a layerwise manner." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.272, + 0.269, + 0.289 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.297, + 0.218, + 0.312 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.469, + 0.426 + ], + "angle": 0, + "content": "The convolution neural network (CNN) has been the dominant architecture for computer vision in the deep learning era, greatly improving the performance of many tasks. Its monopoly has been challenged by the recent emergence of vision transformers (ViTs), which first \"patchify\" each image into tokens and process them with alternating self-attention (SA) and multi-layer perceptron (MLP)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.473, + 0.639 + ], + "angle": 0, + "content": "In addition to architecture design, training strategy also has a large effect on the model performance, especially the data augmentation strategy. Data mixing [48, 45, 24, 47, 52, 53] is an important set of data augmentation for the training of both CNNs and ViTs, as it significantly improves the generalization ability of models. As the most commonly used data mixing strategy, CutMix [52] aims to create virtual training samples from the given training samples \\((\\mathbf{X}, y)\\), where \\(\\mathbf{X} \\in \\mathcal{R}^{H \\times W \\times C}\\) denotes the input image and \\(y\\) is the corresponding label. CutMix randomly selects a local region from one input \\(\\mathbf{X}_1\\) and uses it to replace the pixels in the same region of another input \\(\\mathbf{X}_2\\) to generate a new sample \\(\\tilde{\\mathbf{X}}\\). Similarly, the label \\(\\tilde{y}\\) of \\(\\tilde{\\mathbf{x}}\\) is also the combination of the original labels \\(y_1\\) and \\(y_2\\):" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.65, + 0.47, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {X}} = \\mathbf {M} \\odot \\mathbf {X} _ {1} + (\\mathbf {1} - \\mathbf {M}) \\odot \\mathbf {X} _ {2} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.672, + 0.314, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} = \\lambda y _ {1} + (1 - \\lambda) y _ {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.7, + 0.47, + 0.807 + ], + "angle": 0, + "content": "where \\( M \\in \\{0,1\\}^{H \\times W} \\) is a binary mask indicating the image each pixel belongs to, \\( \\mathbf{1} \\) is an all-one matrix, and \\( \\odot \\) is the element-wise multiplication. \\( \\lambda \\) reflects the mixing ratio of two labels and is the proportion of pixels cropped from \\( \\mathbf{X}_1 \\) in the mixed image \\( \\widetilde{\\mathbf{X}} \\). For a cropped region \\( [r_x, r_x + r_w] \\times [r_y, r_y + r_h] \\) from \\( \\mathbf{X}_1 \\), we compute \\( \\lambda = \\frac{r_w r_h}{WH} \\) to obtain the initial mixed target \\( \\tilde{y} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.394, + 0.832 + ], + "angle": 0, + "content": "3.2. The Token Fluctuation Phenomenon" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.9 + ], + "angle": 0, + "content": "CutMix is originally designed for CNNs and assumes the feature extraction process does not alter the mixing ratio. However, we discover that different from CNNs, self-attention in ViTs can lead to the fluctuation of some tokens." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "The fluctuation further results in the mismatch between the token space and label space, which hinders the effective training of the network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.213 + ], + "angle": 0, + "content": "Formally, we use \\(\\mathbf{z}_i\\) to denote a token of the image \\(\\mathbf{Z}\\), i.e., \\(\\mathbf{z}_i\\) is the transposed \\(i\\)-th column vector of \\(\\mathbf{Z}\\). We can then compute the \\(i\\)-th transformed token \\(\\hat{\\mathbf{z}}_i\\) after the spatial operation as \\(\\hat{\\mathbf{z}}_i = \\sum_{j=1}^{N} w_{i,j}^s \\mathbf{z}_j\\), where \\(w_{i,j}^s\\) is the \\(i, j\\)-th element of the computed spatial mixing matrix \\(\\mathbf{w}^s(\\mathbf{z})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.296 + ], + "angle": 0, + "content": "With the assumption of the linear information integration, we define the contribution of an original token \\(\\mathbf{z}_i\\) to a mixed token \\(\\hat{\\mathbf{z}}_j\\) as \\(c(\\mathbf{z}_i,\\hat{\\mathbf{z}}_j) = \\frac{|w_{i,j}^s|}{\\sum_{k = 1}^{N}|w_{k,j}^s|}\\), where \\(|\\cdot |\\) denotes the absolute value. We can then compute the presence of a token \\(\\mathbf{z}_i\\) in all the mixed image tokens as:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.309, + 0.892, + 0.352 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {z} _ {i}\\right) = \\sum_ {j = 1} ^ {N} c \\left(\\mathbf {z} _ {i}, \\hat {\\mathbf {z}} _ {j}\\right) = \\sum_ {j = 1} ^ {N} \\frac {\\left| w _ {i , j} ^ {s} \\right|}{\\sum_ {k = 1} ^ {N} \\left| w _ {k , j} ^ {s} \\right|}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.892, + 0.407 + ], + "angle": 0, + "content": "For non-strided depth-wise convolution, each token is multiplied by each element in the convolutional kernel due to the translation invariance. We can thus obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.418, + 0.891, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\sum_ {l = 1} ^ {N} \\left| w _ {i, l} ^ {s} \\right| = \\sum_ {j = 1} ^ {N} \\left| w _ {k, j} ^ {s} \\right| = \\sum_ {k = 1, l = 1} ^ {M} \\left| K _ {k, l} \\right|, \\quad \\forall i, j \\in \\mathbf {P} _ {N E}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.893, + 0.701 + ], + "angle": 0, + "content": "where \\(\\mathbf{P}_{NE}\\) denotes the set of positions that are not at the edge of the image, \\(K_{k,l}\\) denotes the value of the \\(k,l\\)-th position of the convolution kernel \\(\\mathbf{K}\\) and \\(M\\) is the kernel size. We can infer that \\(p(\\mathbf{z}_i) = 1\\), \\(\\forall i \\in \\mathbf{P}_{NE}\\), i.e., the effect of all the internal tokens does not change during the convolution process. However, for self-attention in ViTs, (3) does not hold due to the non-existence of translation invariance. The fluctuation of \\(p(\\mathbf{z})\\) is further amplified by the input dependency of the spatial mixing matrix \\(\\mathbf{w}^s (\\mathbf{z})\\) induced by self-attention. As an extreme case, we may obtain \\(p(\\mathbf{z}) \\sim 0\\) for certain tokens. The fluctuation of tokens will alter the proportion of mixing (i.e., \\(\\lambda\\)) and the network might even completely ignore one of the mixed images. The actual label of the processed tokens can then deviate from the mixed label computed by (1), resulting in less effective training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.72, + 0.727 + ], + "angle": 0, + "content": "3.3. Token-Label Alignment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Each token in ViTs interacts with other tokens using the self-attention mechanism. The input-dependent weights empower ViTs with more flexibility but also result in a mismatch between the processed token and the initial token. To address this, we propose a token-label alignment (TL-Align) method to trace the correspondence between the input and transformed tokens to obtain the aligned labels for the resulting representations, as illustrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Specifically, ViTs first split the mixed input \\(\\tilde{\\mathbf{X}}\\) after CutMix (1) to a sequence of \\(N\\) non-overlapped patches and then flatten them to obtain the original image tokens" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5497" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.086, + 0.852, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.325, + 0.895, + 0.366 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed TL-Align. We trace the correspondence between the input tokens and the transformed tokens and align the labels accordingly. We reuse the computed attentions to linearly mix the labels of input tokens to obtain those of transformed tokens. The token-label alignment is performed iteratively to obtain a label for each output token." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.369, + 0.471, + 0.4 + ], + "angle": 0, + "content": "\\(\\{\\tilde{\\mathbf{x}}_1,\\tilde{\\mathbf{x}}_2,\\dots ,\\tilde{\\mathbf{x}}_N\\}\\) . We then project them into a proper dimension and add positional embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.419, + 0.47, + 0.438 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} ^ {0} = [ \\tilde {\\mathbf {z}} _ {c l s}; \\tilde {\\mathbf {x}} _ {1} \\cdot \\mathbf {E}; \\tilde {\\mathbf {x}} _ {2} \\cdot \\mathbf {E}; \\dots ; \\tilde {\\mathbf {x}} _ {N} \\cdot \\mathbf {E} ] + \\mathbf {E} _ {p o s}, (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.47, + 0.578 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{z}}_{cls} \\in \\mathcal{R}^{1 \\times d}\\) denotes the class token, \\(\\mathbf{N}\\) is the number of tokens, \\(\\mathbf{E}\\) represents the patch projector, and \\(\\mathbf{E}_{pos} \\in \\mathcal{R}^{(N + 1) \\times d}\\) is the position embeddings. Note that we adopt the process of the original transformer architecture [16] as an example without loss of generality. Other models may omit the class token and use a relative positional embedding instead, which does not affect the utility of the proposed TL-Align method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.613 + ], + "angle": 0, + "content": "We first assign each token \\(\\mathbf{z}_i\\in \\mathcal{R}^{1\\times d}\\) with a label embedding \\(\\mathbf{y}_i\\in \\mathcal{R}^{1\\times C}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.632, + 0.47, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} ^ {0} = \\left[ \\tilde {\\mathbf {y}} _ {c l s} ^ {0}; \\tilde {\\mathbf {y}} _ {1} ^ {0}; \\tilde {\\mathbf {y}} _ {2} ^ {0}; \\dots ; \\tilde {\\mathbf {y}} _ {N} ^ {0} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.471, + 0.852 + ], + "angle": 0, + "content": "where the sum of elements in each \\(\\mathbf{y}_i\\) equals 1 (i.e., \\(\\sum_{j=1}^{C} y_{i,j} = 1\\)) and \\(y_{i,j}\\) indicates how much the \\(i\\)-th token belong to the \\(j\\)-th class. We initialize the label embedding following the conventional data mixing paradigm. For example, when using CutMix to mix two images \\(\\mathbf{X}_1\\) and \\(\\mathbf{X}_2\\) from the \\(j\\)-th class and the \\(k\\)-th class with a mixing ratio of \\(\\lambda\\), we set \\(\\tilde{y}_{cls,j} = \\lambda\\) and \\(\\tilde{y}_{cls,k} = 1 - \\lambda\\) for the class token. For each patch token, we set \\(\\tilde{y}_{i,j} = 1\\) if it comes from \\(\\mathbf{X}_1\\) and \\(\\tilde{y}_{i,k} = 1\\) if it comes from \\(\\mathbf{X}_2\\). If a patch token contains both the mixed images, we use the mixing ratio within this patch as the label. For MixUp, we can simply set all label embeddings \\(\\{\\tilde{\\mathbf{y}}_i\\}\\) with \\(\\tilde{y}_{j,k} = \\lambda\\) and \\(\\tilde{y}_{j,j} = 1 - \\lambda\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We perform TL-Align in a layer-wise manner and compute the aligned labels based on the operation on the tokens. Formally, ViTs use self-attention to perform spatial mixing" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.369, + 0.65, + 0.385 + ], + "angle": 0, + "content": "of the input tokens \\(\\mathbf{Z}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.39, + 0.847, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathbf {Z} \\cdot \\mathbf {W} _ {\\mathbf {Q}}, \\mathbf {K} = \\mathbf {Z} \\cdot \\mathbf {W} _ {\\mathbf {K}}, \\mathbf {V} = \\mathbf {Z} \\cdot \\mathbf {W} _ {\\mathbf {V}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.41, + 0.892, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} (\\mathbf {Q}, \\mathbf {K}) = \\operatorname {S o f t m a x} \\left(\\mathbf {Q} \\cdot \\mathbf {K} ^ {T} / \\sqrt {d}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.433, + 0.746, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Z}} = \\operatorname {S A} (\\mathbf {Z}) = \\mathcal {A} (\\mathbf {Q}, \\mathbf {K}) \\cdot \\mathbf {V}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.456, + 0.892, + 0.487 + ], + "angle": 0, + "content": "To align the labels, we update the label embeddings \\(\\mathbf{Y}\\) using the same attention matrix \\(\\mathcal{A}(\\mathbf{Q},\\mathbf{K})\\)" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.493, + 0.892, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Y}} = \\mathcal {A} (\\mathbf {Q}, \\mathbf {K}) \\cdot \\mathbf {Y}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.518, + 0.892, + 0.549 + ], + "angle": 0, + "content": "ViTs usually adopt multi-head self-attention (MSA) to perform multiple self-attention parallelly:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.554, + 0.892, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Z}} = \\operatorname {M S A} (\\mathbf {Z}) = \\left[ \\mathrm {S A} _ {1} (\\mathbf {Z}); \\mathrm {S A} _ {2} (\\mathbf {Z}); \\dots ; \\mathrm {S A} _ {H} (\\mathbf {Z}) \\right] \\cdot \\mathbf {w} _ {h}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.585, + 0.892, + 0.632 + ], + "angle": 0, + "content": "where \\(H\\) is the number of heads and \\(\\mathbf{w}_h\\in \\mathcal{R}^{d\\times d}\\). We then adapt our label alignment to MSA by simply taking the average of all the attention matrices for alignment:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.637, + 0.892, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Y}} = \\operatorname {T L} - \\operatorname {A l i g n} - \\mathbf {S} (\\mathbf {Z}, \\mathbf {Y}) := \\frac {1}{H} \\sum_ {i = 1} ^ {H} \\mathcal {A} _ {i} (\\mathbf {Q}, \\mathbf {K}) \\cdot \\mathbf {Y}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.684, + 0.892, + 0.713 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}_i\\) is the attention matrix corresponding to the \\(i\\)-th head \\(\\mathrm{SA}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.715, + 0.892, + 0.745 + ], + "angle": 0, + "content": "Each transformer block \\( l \\) processes the tokens by both spatial and channel mixing:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.75, + 0.892, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Z}} ^ {l - 1} = \\operatorname {M S A} (\\mathrm {L N} (\\mathbf {Z} ^ {l - 1})), \\quad \\mathbf {Z} ^ {\\prime l - 1} = \\hat {\\mathbf {Z}} ^ {l - 1} + \\mathbf {Z} ^ {l - 1}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.772, + 0.805, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Z}} ^ {l} = \\operatorname {M L P} (\\operatorname {L N} \\left(\\mathbf {Z} ^ {\\prime l - 1}\\right)), \\quad \\mathbf {Z} ^ {l} = \\hat {\\mathbf {Z}} ^ {l} + \\mathbf {Z} ^ {\\prime l - 1},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.796, + 0.892, + 0.842 + ], + "angle": 0, + "content": "where MLP and LN denote the MLP module and layer normalization [2], respectively. Our TL-Align then aligns the label embeddings in a similar manner:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.848, + 0.905, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Y}} ^ {l - 1} = \\operatorname {T L} \\text {- A l i g n - S} \\left(\\mathbf {Y} ^ {l - 1}\\right), \\mathbf {Y} ^ {\\prime l - 1} = \\operatorname {N o r m} \\left(\\hat {\\mathbf {Y}} ^ {l - 1} + \\mathbf {Y} ^ {l - 1}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.869, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Y}} ^ {l} = \\mathbf {Y} ^ {\\prime l - 1}, \\quad \\mathbf {Y} ^ {l} = \\operatorname {N o r m} \\left(\\hat {\\mathbf {Y}} ^ {l} + \\mathbf {Y} ^ {\\prime l - 1}\\right), \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5498" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "where Norm denotes the normalization operation. We implement Norm by a simple average." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.47, + 0.288 + ], + "angle": 0, + "content": "Hierarchical vision transformers such as Swin [31] further introduce a patch aggregation operation to merge multiple patches. They usually concatenate multiple tokens across the channels to reduce the spatial resolution. Instead of concatenation, we simply add the label embeddings of the merged tokens followed by normalization as the aligned labels. The proposed TL-Align can be generalized to different architectures composed of spatial mixing, channel mixing, point-wise transformation, residual connection, and spatial aggregation. We provide detailed illustrations of alignment with these operations in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.471, + 0.429 + ], + "angle": 0, + "content": "We synchronously align the labels with the processed tokens layer by layer and obtain the aligned tokens \\(\\mathbf{Z}^L\\) and labels \\(\\mathbf{Y}^L\\). The final representation of the image \\(\\mathbf{z}\\) is either the class token \\(\\mathbf{z}_{cls}^L\\) [16, 43] or the average pooling of all the spatial tokens \\(\\frac{1}{N}\\sum_{i=1}^{N}\\mathbf{z}_i^L\\) [31]. The aligned label \\(\\mathbf{y}_{align}\\) for the image is then \\(\\mathbf{y}_{cls}^L\\) or \\(\\frac{1}{N}\\sum_{i=1}^{N}\\mathbf{y}_i^L\\) depending on the specific model. We then adopt the aligned label \\(\\mathbf{y}_{align}\\) to train the network and can adapt to different loss functions and training schemes:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.44, + 0.469, + 0.457 + ], + "angle": 0, + "content": "\\[\nJ = J (\\mathbf {z}, \\text {s t o p - g r a d i e n t} (\\mathbf {y} _ {\\text {a l l i n g}})). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.469, + 0.497 + ], + "angle": 0, + "content": "We do not back-propagate through the aligned label as they only serve as a more accurate target." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.469, + 0.603 + ], + "angle": 0, + "content": "Our TL-Align serves as a plug-and-play module on various vision transformers while only introducing negligible training costs. We adjust the label of each token adaptively during the layer-by-layer propagation and preserve alignment between tokens and labels throughout the forward process. TL-Align is only used during training and introduces no additional computation cost when inference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.617, + 0.21, + 0.634 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.469, + 0.779 + ], + "angle": 0, + "content": "In this section, we conducted extensive experiments to evaluate the proposed TL-Align method. We demonstrate the improvement of TL-Align on various vision transformers and compare it with state-of-the-art training strategies concerning the accuracy, network complexity, and training speed. We also examine the transferability on downstream tasks including semantic segmentation, object detection, and transfer learning. We further provide in-depth analysis to evaluate the effectiveness of TL-Align." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.295, + 0.803 + ], + "angle": 0, + "content": "4.1. ImageNet Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We first evaluate TL-Align on ImageNet [38] for image classification. We implement our method based on PyTorch [35] and the timm library [50]. We conduct experiments on various transformer architectures: three variants of DeiT [43] (DeiT-T, DeiT-S, and DeiT-B), two variants of PVT [49] (PVT-T, PVT-S) and" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.087, + 0.892, + 0.129 + ], + "angle": 0, + "content": "Table 1. Results on ImageNet classification task. We compare the parameters, FLOPs, and accuracy of different vision transformer backbones without and with our TL-Align." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.131, + 0.903, + 0.396 + ], + "angle": 0, + "content": "
ModelImage SizeParamsFLOPsTop-1(%)Top-5(%)
DeiT-T22425.7M1.6G72.291.3
+TL-Align73.291.7
PVT-T224213.2M1.9G75.192.4
+TL-Align75.593.0
DeiT-S224222M4.6G79.895.0
+TL-Align80.695.0
PVT-S224224.5M3.8G79.895.0
+TL-Align80.495.5
Swin-T224229M4.5G81.295.5
+TL-Align81.495.7
Swin-S224250M8.8G83.096.3
+TL-Align83.496.5
DeiT-B224286M17.5G81.895.5
+TL-Align82.395.8
Swin-B224288M15.4G83.596.4
+TL-Align83.796.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.401, + 0.892, + 0.429 + ], + "angle": 0, + "content": "Table 2. Comparison of our TL-Align with other training strategies on ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.431, + 0.872, + 0.612 + ], + "angle": 0, + "content": "
MethodParamsTrainig TimeAcc.(%)
Vanilla22M1×76.4
CutMix [52]22M1×79.8
Puzzle-Mix [24]22M2.24×79.8
F-Mix [18]22M1.08×77.4
ResizeMix [36]22M1.03×78.6
SaliencyMix [45]22M1.26×79.2
Attentive-CutMix [48]46M1.31×77.5
TransMix [7]22M1.16×80.1
TokenMix [30]22M1.65×80.3
CutMix + TL-Align22M1.24×80.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.529, + 0.616, + 0.859, + 0.63 + ], + "angle": 0, + "content": "Table 3. Results on semantic segmentation ADE20K." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.632, + 0.876, + 0.791 + ], + "angle": 0, + "content": "
BackboneParamsFLOPsmIoUmIoU (MS)mAcc
DeiT-S58M1032G43.845.155.2
+TL-Align44.545.755.5
Swin-T60M945G44.445.855.6
+TL-Align44.746.556.4
Swin-S81M1038G47.649.558.8
+TL-Align48.049.759.5
Swin-B121M1188G48.149.759.1
+TL-Align48.350.159.7
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "three variants of Swin Transformer [31] (Swin-T, Swin-S, and Swin-B). For tiny and small models, we train from scratch for 300 epochs following the same training recipe using CutMix as DeiT [43], PVT [49] and Swin [31]. We keep all the data augmentation policies and all hyperparameter settings unchanged for fair comparisons. The only modification is that we replace the mixing targets in Cut" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5499" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.209, + 0.088, + 0.757, + 0.102 + ], + "angle": 0, + "content": "Table 4. Experimental results on object detection and instance segmentation on COCO." + }, + { + "type": "table", + "bbox": [ + 0.087, + 0.104, + 0.88, + 0.23 + ], + "angle": 0, + "content": "
BackboneParamsFLOPsSchedule\\(AP^{box}\\)\\(AP^{box}_{50}\\)\\(AP^{box}_{75}\\)\\(AP^{mask}\\)\\(AP^{mask}_{50}\\)\\(AP^{mask}_{75}\\)
Swin-T+TL-Align86M745G3x50.469.254.743.766.647.3
50.569.454.943.866.647.3
Swin-S+TL-Align107M838G3x51.970.756.345.068.248.8
52.271.156.745.268.449.1
Swin-B+TL-Align145M982G3x51.970.556.445.068.148.9
52.371.256.945.368.749.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.234, + 0.443, + 0.247 + ], + "angle": 0, + "content": "Table 5. Results on different transfer learning datasets." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.25, + 0.46, + 0.41 + ], + "angle": 0, + "content": "
ModelParamsFLOPsC-10C-100FlowersCars
ResNet5026M4.1G--96.290.0
ViT-B/1686M55.4G98.187.189.5-
ViT-L/16307M190.7G97.986.489.7-
DeiT5.7M1.6G97.685.797.190.1
+TL-Align5.7M1.6G97.886.497.990.7
DeiT-S22M4.6G97.990.298.191.4
+TL-Align22M4.6G98.890.498.391.8
DeiT-B86M17.5G99.190.898.492.1
+TL-Align86M17.5G99.190.598.693.0
" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.416, + 0.459, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.641 + ], + "angle": 0, + "content": "Figure 3. The Root Mean Square Error (RMSE) between original CutMix targets and labels obtained by T-L Align." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.703 + ], + "angle": 0, + "content": "MiX with the labels obtained by our TL-Align. For base models (i.e., Deit-B and Swin-B), we finetune the official pre-trained models for 40 epochs with a constant learning rate of 1e-5 and a weight decay of 1e-8." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Performance on Different Architectures. As shown in Table 1, TL-Align steadily improves the performance of different vision transformer architectures. Specifically, TL-Align boosts the top-1 accuracy of DeiT-T, DeiT-S, and DeiT-B by \\(1.0\\%\\), \\(0.8\\%\\), and \\(0.5\\%\\), respectively, in a parameter-free manner. Moreover, our method is generalizable and can be directly applied to hierarchical vision transformers like Swin. It is worth noting that most existing methods need either architecture modifications (adding a class token in [7]) or extra computations (saliency map extraction in [45]) when applied to Swin. In contrast, our TL-Align method can be used as a plug-and-play module and achieves consistent improvement on variants of Swin." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.235, + 0.868, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.433, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Figure 4. Visualization of mixing ratio \\(\\lambda\\) of fluctuating tokens from different layers. We compare the results of TL-Align with CutMix, token similarity, TransMix, and TokenLabeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.893, + 0.807 + ], + "angle": 0, + "content": "Comparison with Other Training Strategies. We also compare our method with the state-of-the-art training strategies for data mixing on DeiT-S, as shown in Table 2. Specifically, we train the DeiT-S model while only disabling CutMix as the baseline method, which is denoted as Vanilla. Moreover, since TransMix [7] and TokenMix [30] report the EMA accuracy with different hyperparameters, we reproduce it under the same training recipe [43] for a fair comparison. We see that TL-Align shows better performance than the other mixup variants while maintaining the number of parameters and training speed. Puzzle-Mix obtains the same classification accuracy as CutMix but results in a much lower training speed as it relies on an extra model to get the optimal solution. SaliencyMix and Attentive-CutMix lead to performance degeneration when built upon DeiT-S backbone. Notably, our method also achieves higher top-1 accuracy than ViT-targeted TransMix and TokenMix. While the teacher annotator in TokenMix introduces heavy training costs, TL-Align obtains accurate token-label alignment in an efficient manner. Moreover, TransMix has limitations in handling token fluctuation and cannot be applied to ViTs without a class token, such as Swin." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.683, + 0.832 + ], + "angle": 0, + "content": "4.2. Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Semantic Segmentation. We evaluate our TL-Align on ADE20K dataset [55] for semantic segmentation. We adopt DeiT-S and three variants of Swin Transformer as backbones equipped with UpperNet for segmentation. As pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5500" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.088, + 0.892, + 0.116 + ], + "angle": 0, + "content": "Table 6. Comparison results of model generalization ability and robustness. We evaluate them on various out-of-distribution/corrupted datasets and against adversarial attacks. \\(\\uparrow\\) denotes higher is better, and \\(\\downarrow\\) denotes lower is better." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.118, + 0.861, + 0.264 + ], + "angle": 0, + "content": "
ModelParamsFLOPsImageNetGeneralizationRobustness
Top-1↑Top-5↑IN-V2↑IN-A↑IN-C↓IN-R↑AutoAttack↑
DeiT-T5.7M1.6G72.291.360.47.769.134.13.9
+TL-Align5.7M1.6G73.291.761.46.168.034.64.4
DeiT-S22M4.6G79.895.068.518.954.742.56.9
+TL-Align22M4.6G80.695.068.919.253.243.27.5
DeiT-B86M17.5G81.895.570.527.948.545.3-
+TL-Align86M17.5G82.395.870.929.047.144.4-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.267, + 0.443, + 0.281 + ], + "angle": 0, + "content": "Table 7. Ablation of using different data mixing strategies." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.284, + 0.482, + 0.397 + ], + "angle": 0, + "content": "
MixUpCutMixRandomBlock-wiseTop-1 (%)+TL-Align Top-1 (%)
××××76.4-
×✓××79.880.6
✓✓××79.880.2
××✓×79.780.2
×××✓80.080.3
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.399, + 0.468, + 0.429 + ], + "angle": 0, + "content": "sented in Table 3, TL-Align improves the segmentation performance on both DeiT and Swin at different model scales." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.468, + 0.594 + ], + "angle": 0, + "content": "Object Detection and Instance Segmentation. We also examine the performance of TL-Align on object detection and instance segmentation on the COCO 2017 dataset [29]. We apply our TL-Align to Swin [31] due to the advantage of the hierarchical representations on object detection tasks. We adopt the Cascade Mask-RCNN [4] framework and use the training strategy of \\(3 \\times\\) schedule. As shown in Table 4, we observe consistent improvements on all variants of Swin Transformer. This demonstrates the advantages to learn token-level meaningful features suitable for dense prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.468, + 0.716 + ], + "angle": 0, + "content": "Transfer Learning. We further evaluate the transferred classification performance of TL-Align on CIFAR-10 [26], CIFAR-100 [26], Flowers [34] and Cars [25]. We use pretrained models on ImageNet and finetune them on these datasets following existing works [43]. We compare the performance with and without TL-Align on three variants of DeiT [43], as shown in Table 5. TL-Align obtains significant performance gains for all variants on the four datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.42, + 0.743 + ], + "angle": 0, + "content": "4.3. Performance Analysis and Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Token-Label Alignment. We first quantize the difference between the original targets and aligned labels and investigate its correlation with the model. Specifically, we compute the Root Mean Square Error (RMSE) between the original targets and labels obtained by our TL-Align. As shown in Figure 3, the RMSE decreases when enlarging the model size. This indicates that larger models demonstrate less token fluctuation since the self-correction ability is also enhanced as the model capacity scales up. Moreover, the RMSE for Swin Transformer" + }, + { + "type": "table_caption", + "bbox": [ + 0.534, + 0.267, + 0.854, + 0.281 + ], + "angle": 0, + "content": "Table 8. Ablation of different TL-Align operations." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.284, + 0.858, + 0.383 + ], + "angle": 0, + "content": "
AlignmentTop-1 Acc.(%)Δ (%)
None (DeiT-S baseline)79.8-
TL-Align-S (Layer 12)80.1+0.3
TL-Align-S (Layer 2,4,6,8)80.2+0.4
Normalization Disabled80.3+0.5
Default (TL-Align)80.6+0.8
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.476 + ], + "angle": 0, + "content": "tends to be lower compared with DeiT of similar model size. This is due to the local-window self-attention in Swin which preserves more local information. These observations are consistent with our experimental results: the improvements on small models and DeiT-like backbones tend to be more significant as they encounter more token fluctuation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Visualization of the Layer-wise Mixing Ratio of Fluctuated Tokens. To investigate the effectiveness of TL-Align, we compute a similarity-based \"ground-truth\" mixing ratio for each layer. Specifically, we compute the similarities of tokens between the mixed and unmixed images and use them as the label of each token. We compare them with the mixing ratios produced by TL-Align, CutMix [52], TransMix [7], and TokenLabeling [23]. As shown in Figure 4, the similarity-based mixing ratio changes at each layer, resulting from token fluctuation. However, CutMix, TransMix, and TokenLabeling assume the output tokens keep spatial correspondence with the input tokens and compute a fixed mixing ratio. TL-Align assigns dynamic labels to tokens using layer-wise alignment, which is more accurate compared with other methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation of Robustness and Generalization. We further conduct experiments to validate the generalization and robustness of TL-Align, as shown in Table 6. We employ four corrupted and out-of-distribution datasets for robustness evaluation, including ImageNet-A [22], ImageNet-C [21], and ImageNet-R [20]. We also adopt AutoAttack [11] to evaluate the adversarial robustness on the ImageNet validation set. Due to memory limitation, we do not experiment with DeiT-B on AutoAttack. We use mean Corruption Error (mCE) for ImageNet-C and Top-1 Accuracy for others as the evaluation metric. For generalization evaluation, we adopt the ImageNet-V2 dataset [37]. We see that TL-Align improves both robustness and general" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5501" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.885, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.326, + 0.893, + 0.354 + ], + "angle": 0, + "content": "Figure 5. The visualization of results on DeiT-S and Swin-S. We visualize the input images, the mixed image, the original label embedding, and the label embedding after token-label alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.388 + ], + "angle": 0, + "content": "ization, showing the superiority of adopting TL-Align for pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.47, + 0.614 + ], + "angle": 0, + "content": "Ablation Study on different Data Mixing Strategies. Due to the efficiency of the proposed layer-wise alignment, TL-Align can be directly applied to a wide range of data mixing strategies. We adopt MixUp, CutMix, a random mixing strategy and a block-wise mixing strategy to evaluate the generalizability of TL-Align. The random mixing and block-wise mixing strategies are inspired by MAE [19] and BEiT [3] and we replace the masking operation with image mixing on patch-level and block-level (both of size \\(16 \\times 16\\)) respectively. The comparison results of training DeiT-S with and without our approach is demonstrated in Table 7. Specifically, TL-Align improves CutMix by \\(0.8\\%\\), MixUp+CutMix by \\(0.4\\%\\), random mixing by \\(0.5\\%\\) and block-wise mixing by \\(0.3\\%\\) respectively, further verifying the generalizability of the proposed TL-Align." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Ablation Study on Different Label Alignment Operations. Our TL-Align aligns the labels with tokens transformed by spatial self-attention and residual connection layer-by-layer. To investigate the effect of reusing attention maps and normalization, we conduct an ablation study regarding different alignment operations on DeiT-S. We try aligning the labels only by using the attention map of Layer 12, which is equivalent to TransMix [7]. We also test the performance of applying alignment to several middle transformer layers and disabling normalization. As presented in Table 8, incomplete alignment at a part of layers marginally boosts the performance as it cannot well handle the token fluctuation issue. Disabling normalization leads to \\(0.3\\%\\) accuracy drop due to the inaccurate alignment at the presence of residual connections. This demonstrates the significance of the token-label alignment by attention utilization and normalization in a layer-wise manner." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Visualizations of Aligned Labels. We visualize the labels obtained by TL-Align on DeiT-S [43] and Swin-S [31]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.892, + 0.553 + ], + "angle": 0, + "content": "as shown in Figure 5. Specifically, the aligned label embedding is obtained after the final transformer block for both DeiT-S and Swin-S. The value of the label embedding represents the probability of the belonged class of the corresponding token. We use red to denote larger probabilities towards the first image and blue for the second image. We observe that the aligned labels can deviate from the original labels and result in different mixing ratios for training. Therefore, using the original ratio as the training target may produce false training signals and lead to inferior performance. We see that TL-Align can correct the labels when the images are mixed with uninformative tokens. More visualization results are included in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.572, + 0.619, + 0.587 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.795 + ], + "angle": 0, + "content": "In this paper, we have presented a token-label alignment method for training better vision transformers. As important subsets of data augmentation methods, data mixing strategies can generally improve the performance of both CNNs and ViTs. We identify a token fading issue for ViTs and address it by tracing the correspondence between transformed tokens and the original tokens to obtain a label for each output token to obtain more accurate training signals. Experimental results have demonstrated that our TL-Align can consistently improve the performance of various ViT models. The generalization performance of TL-Align to other architectures such as MLP-like models remains unknown and is a promising future direction to explore." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.814, + 0.667, + 0.831 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160102 and in part by the National Natural Science Foundation of China under Grant 62125603." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5502" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.468, + 0.158 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.161, + 0.468, + 0.188 + ], + "angle": 0, + "content": "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv, abs/1607.06450, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.191, + 0.468, + 0.231 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.235, + 0.468, + 0.275 + ], + "angle": 0, + "content": "[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, pages 6154-6162, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.279, + 0.468, + 0.332 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.337, + 0.468, + 0.391 + ], + "angle": 0, + "content": "[6] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.395, + 0.468, + 0.449 + ], + "angle": 0, + "content": "[7] Jie-Neng Chen, Shuyang Sun, Ju He, Philip Torr, Alan Yuille, and Song Bai. Transmix: Attend to mix for vision transformers. arXiv preprint arXiv:2111.09833, 2021. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.468, + 0.507 + ], + "angle": 0, + "content": "[8] Tianlong Chen, Zhenyu Zhang, Yu Cheng, Ahmed Awadallah, and Zhangyang Wang. The principle of diversity: Training stronger vision transformers calls for reducing all levels of redundancy. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.468, + 0.55 + ], + "angle": 0, + "content": "[9] Bowen Cheng, Alexander G Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In NeurIPS, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.554, + 0.468, + 0.608 + ], + "angle": 0, + "content": "[10] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In NeurIPS, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.612, + 0.468, + 0.667 + ], + "angle": 0, + "content": "[11] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.67, + 0.468, + 0.724 + ], + "angle": 0, + "content": "[12] Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In ICCV, pages 2988-2997, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.728, + 0.468, + 0.769 + ], + "angle": 0, + "content": "[13] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In CVPR, pages 1601-1610, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.468, + 0.812 + ], + "angle": 0, + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.468, + 0.869 + ], + "angle": 0, + "content": "[15] Xiaohan Ding, Xiangyu Zhang, Yizhuang Zhou, Jungong Han, Guiguang Ding, and Jian Sun. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.151, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[17] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.195, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[18] Ethan Harris, Antonia Marcu, Matthew Painter, Mahesan Niranjan, Adam Prügel-Bennett, and Jonathon Hare. Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.252, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.324, + 0.892, + 0.406 + ], + "angle": 0, + "content": "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.409, + 0.892, + 0.45 + ], + "angle": 0, + "content": "[21] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.453, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[22] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[23] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. In NeurIPS, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.569, + 0.892, + 0.609 + ], + "angle": 0, + "content": "[24] Jang-Hyun Kim, Wonho Choo, and Hyun Oh Song. Puzzle mix: Exploiting saliency and local statistics for optimal mixup. In ICML, pages 5275-5285, 2020. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.613, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[25] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCVW, pages 554–561, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.657, + 0.892, + 0.684 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[27] Wenbo Li, Xin Lu, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.892, + 0.784 + ], + "angle": 0, + "content": "[28] Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, and Xiaou Tang. Not all pixels are equal: Difficulty-aware semantic segmentation via deep layer cascade. In CVPR, pages 3193-3202, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[30] Jihao Liu, Boxiao Liu, Hang Zhou, Hongsheng Li, and Yu Liu. Tokenmix: Rethinking image mixing for data augmentation in vision transformers. In ECCV, pages 455-471, 2022. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "5503" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[32] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. arXiv preprint arXiv:2201.03545, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.47, + 0.234 + ], + "angle": 0, + "content": "[33] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. arXiv preprint arXiv:2106.13230, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.47, + 0.291 + ], + "angle": 0, + "content": "[34] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NIPS, pages 8026-8037, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[36] Jie Qin, Jiemin Fang, Qian Zhang, Wenyu Liu, Xingang Wang, and Xinggang Wang. Resizemix: Mixing data with preserved object information and true labels. arXiv preprint arXiv:2012.11101, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[37] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning, pages 5389-5400, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[38] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[39] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.632 + ], + "angle": 0, + "content": "[40] Yehui Tang, Kai Han, Jianyuan Guo, Chang Xu, Yanxi Li, Chao Xu, and Yunhe Wang. An image patch is a wave: Quantum inspired vision mlp. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[41] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. NeurIPS, 34, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[42] Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Gautier Izacard, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, et al. Resmlp: Feedforward networks for image classification with data-efficient training. arXiv preprint arXiv:2105.03404, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.855 + ], + "angle": 0, + "content": "[43] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[44] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[45] AFM Uddin, Mst Monira, Wheemyung Shin, TaeChoong Chung, Sung-Ho Bae, et al. Saliencymix: A saliency guided data augmentation strategy for better regularization. arXiv preprint arXiv:2006.01791, 2020. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[46] Shashanka Venkataramanan, Ewa Kijak, Laurent Amsaleg, and Yannis Avrithis. Alignmixup: Improving representations by interpolating aligned features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19174-19183, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[47] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In ICML, pages 6438-6447, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[48] Devesh Walawalkar, Zhiqiang Shen, Zechun Liu, and Marios Savvides. Attentive cutmix: An enhanced data augmentation approach for deep learning based image classification. arXiv preprint arXiv:2003.13048, 2020. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[49] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[50] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[51] Lingfeng Yang, Xiang Li, Borui Zhao, Renjie Song, and Jian Yang. Recursivelymix: Mixed learning with history. arXiv preprint arXiv:2203.06844, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[52] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, pages 6023-6032, 2019. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[53] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[54] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. IJCV, 127:302-321, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[56] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5504" + } + ] +] \ No newline at end of file diff --git a/2023/Token-Label Alignment for Vision Transformers/ef46bf1f-34a3-4e70-bf1a-b3f5f27b821d_origin.pdf b/2023/Token-Label Alignment for Vision Transformers/ef46bf1f-34a3-4e70-bf1a-b3f5f27b821d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..537cc724dfb462a3cb43177f9ab498dc48498a17 --- /dev/null +++ b/2023/Token-Label Alignment for Vision Transformers/ef46bf1f-34a3-4e70-bf1a-b3f5f27b821d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9db85290c8e1f46715492f77aa785e6f8a41435183b7f4a1253d88ba13631f45 +size 2740176 diff --git a/2023/Token-Label Alignment for Vision Transformers/full.md b/2023/Token-Label Alignment for Vision Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c2c27eb688e3921fea32110ab1c6853aa5cc28bf --- /dev/null +++ b/2023/Token-Label Alignment for Vision Transformers/full.md @@ -0,0 +1,335 @@ +# Token-Label Alignment for Vision Transformers + +Han Xiao $^{1,2,*}$ Wenzhao Zheng $^{1,2,*}$ Zheng Zhu $^{3}$ Jie Zhou $^{1,2}$ Jiwen Lu $^{1,2,\dagger}$ $^{1}$ Beijing National Research Center for Information Science and Technology, China + $^{2}$ Department of Automation, Tsinghua University, China + $^{3}$ PhiGent Robotics +{h-xiao20, zhengwz18}@mails.tsinghua.edu.cn; zhengzhu@ieee.org; {jzhou, lujiwen} @tsinghua.edu.cn + +# Abstract + +Data mixing strategies (e.g., CutMix) have shown the ability to greatly improve the performance of convolutional neural networks (CNNs). They mix two images as inputs for training and assign them with a mixed label with the same ratio. While they are shown effective for vision transformers (ViTs), we identify a token fluctuation phenomenon that has suppressed the potential of data mixing strategies. We empirically observe that the contributions of input tokens fluctuate as forward propagating, which might induce a different mixing ratio in the output tokens. The training target computed by the original data mixing strategy can thus be inaccurate, resulting in less effective training. To address this, we propose a token-label alignment (TL-Align) method to trace the correspondence between transformed tokens and the original tokens to maintain a label for each token. We reuse the computed attention at each layer for efficient token-label alignment, introducing only negligible additional training costs. Extensive experiments demonstrate that our method improves the performance of ViTs on image classification, semantic segmentation, objective detection, and transfer learning tasks. Code is available at: https://github.com/Euphorial6/TL-Align. + +# 1. Introduction + +The recent developments of vision transformers (ViTs) have revolutionized the computer vision field and set new state-of-the-arts in various tasks, such as image classification [16, 43, 31, 10], object detection [5, 56, 12, 13], and semantic segmentation [28, 39, 54, 9]. The successful structure of alternative spatial mixing and channel mixing in ViTs also motivates the arising of high-performance MLP-like deep architectures [40, 41, 42] and promotes the evolution of better CNNs [15, 32, 17]. In addition to architec + +ture designs, an improved training strategy can also greatly boost the performance of a trained deep model [23, 44, 8, 7]. + +The training of modern deep architecture almost all adopts data mixing strategies for data augmentation [48, 45, 24, 47, 52, 53], which have been proven to consistently improve the generalization performance. They randomly mix two images and their labels with the same mixing ratio to produce mixed data. As the most commonly used data mixing strategy, CutMix [52] performs a copy-and-paste operation on the spatial domain to produce spatially mixed images. While data mixing strategies have been widely studied for CNNs [48, 45, 24], few works have explored their compatibilities with ViTs [7]. We find that self-attention in ViTs causes a fluctuation of the original spatial structure. Unlike the translation equivalence that ensures a global label consistency for CNNs, self-attention in ViTs undermines this global consistency and causes a misalignment between the token and label. This misalignment induces a different mixing ratio in the output tokens. The training targets computed by the original data mixing strategies can then be inaccurate, resulting in less effective training. + +To address this, we propose a token-label alignment (TL-Align) method for ViTs to obtain a more accurate target for training. We present an overview of our method in Figure 1. We first assign a label to each input token in the mixed image according to the source of the token. We then trace the correspondence between the input tokens and the transformed tokens and align the labels accordingly. We assume that only the spatial self-attention and residual connection operation alter the presence of input tokens since channel MLP and layer normalization process each token independently. We reuse the computed attentions to linearly mix the labels of input tokens to obtain those of transformed tokens. The token-label alignment is performed iteratively to obtain a label for each output token. For class-token-based classification (e.g., ViT [16] and DeiT [43]), we directly use the aligned label for the output class token as the training target. For global-pooling-based classification (e.g., Swin [31]), we similarly average the labels of output tokens as the train + +![](images/a75d82a46cf1b38b0be800352cada4aca186c90985e32fccc083502cd5e3af83.jpg) +(a) CutMix + +![](images/d34ab6d53d42770cb0ac9219357b66933ab60f35a6c7acb00557ef2a3336d492.jpg) +(b) Token misalignment for ViTs + +![](images/affaf24816e0d1ff852247366f51daa33c17b7a80f7201d6bdd47bbe24de14a9.jpg) +(c) Different ways to obtain the mixed labels +Figure 1. An overview of the proposed TL-Align. (a) CutMix-like methods [52] are widely used in model training, which spatially mix the tokens and their labels in the input space. (b) They are originally designed for CNNs and assume the processed tokens are spatially aligned with the input tokens. We show that it does not hold true for ViTs due to the global receptive field and adaptive weights. (c) Compared with existing methods, our method can effectively and efficiently align the tokens and labels without requiring a pretrained teacher network. + +ing target. The proposed TL-Align is only used for training to improve performance and introduces no additional workload for inference. We apply the proposed TL-Align to various ViT variants with CutMix including plain ViTs (DeiT [43]) and hierarchical ViTs (Swin [31]). We observe a consistent performance boost across different models on ImageNet-1K [14]. Specifically, our TL-Align improves DeiT-S by $0.8\%$ using the same training recipe. We evaluate the ImageNet-pretrained models on various downstream tasks including semantic segmentation, objection detection, and transfer learning. Experimental results also verify the robustness and generalization ability of our method. + +# 2. Related Work + +Vision Transformer. Transformers have been widely used in natural language processing and achieved great success on many language tasks. Recently, Vision Transformers (ViTs) have aroused extensive interest in computer vision due to their competitive performance compared with CNNs [16, 43, 31, 10]. Dosovitskiy et al. [16] firstly introduced transformers into the image classification task. They split the input image into non-overlapped patches and then feed them into the transformer encoders. Liu et al. [31] proposed a shifted windowing scheme to produce hierarchical feature maps suitable for dense prediction tasks. The great potential of vision transformer has motivated its adaptation to many challenging tasks including object detection [12, 56, 5], segmentation [9, 39], image enhancement [6, 27] and video understanding [33, 1]. + +Recently, some efforts have been devoted to producing better training targets to improve the performance of vision transformers [23, 44]. For example, DeiT [43] introduces a knowledge distillation procedure to reduce the training cost of ViTs and achieves a better accuracy/speed trade-off. TokenLabeling [23] employs a pretrained teacher annotator to predict a label for each token for dense knowledge distillation. Differently, we do not require a pretrained network to obtain the training targets. Our TL-Align maintains an aligned label for each token layer by layer and can be trained efficiently in an end-to-end manner. + +Data Mixing Strategy. As an important type of data augmentation, data mixing strategies have demonstrated a consistent improvement in the generalization performance of CNNs. Zhang et al. [53] first proposed to combine a training pair to create augmented samples for model regularization. They perform linear interpolations on both the input images and associated targets. Following MixUp, CutMix [52] also utilizes the mixture of two input images but adopts a region copy-and-paste operation. Later methods including Puzzle Mix [24], SaliencyMix [45] and Attentive CutMix [48] leverage the salient regions for informative mixture generation. AlignMixUp [46] aligns two images geometrically in the feature space to improve the mixing representations. Recently, Yang et al. [51] proposed a RecursiveMix strategy which employs the historical input-prediction-label triplets for scale-invariant feature learning. Despite the better performance, a drawback of these methods is the heavily increased training cost due to the saliency extraction or historical information exploitation. + +Most existing data mixing methods are originally designed for CNNs, and their effectiveness on ViTs has not been well explored. TransMix [7] uses the class attention map at the last layer to re-weight the mixing targets and assumes the output tokens to keep spatial correspondence with the input tokens. However, we identify a token fluctuation phenomenon for ViTs which may cause a mismatch between tokens and labels, leading to inaccurate label assignments in both the original CutMix and TransMix. To address this, we propose to align the label and token space by tracing their correspondence in a layerwise manner. + +# 3. Proposed Approach + +# 3.1. Preliminaries + +The convolution neural network (CNN) has been the dominant architecture for computer vision in the deep learning era, greatly improving the performance of many tasks. Its monopoly has been challenged by the recent emergence of vision transformers (ViTs), which first "patchify" each image into tokens and process them with alternating self-attention (SA) and multi-layer perceptron (MLP). + +In addition to architecture design, training strategy also has a large effect on the model performance, especially the data augmentation strategy. Data mixing [48, 45, 24, 47, 52, 53] is an important set of data augmentation for the training of both CNNs and ViTs, as it significantly improves the generalization ability of models. As the most commonly used data mixing strategy, CutMix [52] aims to create virtual training samples from the given training samples $(\mathbf{X}, y)$ , where $\mathbf{X} \in \mathcal{R}^{H \times W \times C}$ denotes the input image and $y$ is the corresponding label. CutMix randomly selects a local region from one input $\mathbf{X}_1$ and uses it to replace the pixels in the same region of another input $\mathbf{X}_2$ to generate a new sample $\tilde{\mathbf{X}}$ . Similarly, the label $\tilde{y}$ of $\tilde{\mathbf{x}}$ is also the combination of the original labels $y_1$ and $y_2$ : + +$$ +\tilde {\mathbf {X}} = \mathbf {M} \odot \mathbf {X} _ {1} + (\mathbf {1} - \mathbf {M}) \odot \mathbf {X} _ {2} \tag {1} +$$ + +$$ +\tilde {y} = \lambda y _ {1} + (1 - \lambda) y _ {2} +$$ + +where $M \in \{0,1\}^{H \times W}$ is a binary mask indicating the image each pixel belongs to, $\mathbf{1}$ is an all-one matrix, and $\odot$ is the element-wise multiplication. $\lambda$ reflects the mixing ratio of two labels and is the proportion of pixels cropped from $\mathbf{X}_1$ in the mixed image $\widetilde{\mathbf{X}}$ . For a cropped region $[r_x, r_x + r_w] \times [r_y, r_y + r_h]$ from $\mathbf{X}_1$ , we compute $\lambda = \frac{r_w r_h}{WH}$ to obtain the initial mixed target $\tilde{y}$ . + +# 3.2. The Token Fluctuation Phenomenon + +CutMix is originally designed for CNNs and assumes the feature extraction process does not alter the mixing ratio. However, we discover that different from CNNs, self-attention in ViTs can lead to the fluctuation of some tokens. + +The fluctuation further results in the mismatch between the token space and label space, which hinders the effective training of the network. + +Formally, we use $\mathbf{z}_i$ to denote a token of the image $\mathbf{Z}$ , i.e., $\mathbf{z}_i$ is the transposed $i$ -th column vector of $\mathbf{Z}$ . We can then compute the $i$ -th transformed token $\hat{\mathbf{z}}_i$ after the spatial operation as $\hat{\mathbf{z}}_i = \sum_{j=1}^{N} w_{i,j}^s \mathbf{z}_j$ , where $w_{i,j}^s$ is the $i, j$ -th element of the computed spatial mixing matrix $\mathbf{w}^s(\mathbf{z})$ . + +With the assumption of the linear information integration, we define the contribution of an original token $\mathbf{z}_i$ to a mixed token $\hat{\mathbf{z}}_j$ as $c(\mathbf{z}_i,\hat{\mathbf{z}}_j) = \frac{|w_{i,j}^s|}{\sum_{k = 1}^{N}|w_{k,j}^s|}$ , where $|\cdot |$ denotes the absolute value. We can then compute the presence of a token $\mathbf{z}_i$ in all the mixed image tokens as: + +$$ +p \left(\mathbf {z} _ {i}\right) = \sum_ {j = 1} ^ {N} c \left(\mathbf {z} _ {i}, \hat {\mathbf {z}} _ {j}\right) = \sum_ {j = 1} ^ {N} \frac {\left| w _ {i , j} ^ {s} \right|}{\sum_ {k = 1} ^ {N} \left| w _ {k , j} ^ {s} \right|}. \tag {2} +$$ + +For non-strided depth-wise convolution, each token is multiplied by each element in the convolutional kernel due to the translation invariance. We can thus obtain: + +$$ +\sum_ {l = 1} ^ {N} \left| w _ {i, l} ^ {s} \right| = \sum_ {j = 1} ^ {N} \left| w _ {k, j} ^ {s} \right| = \sum_ {k = 1, l = 1} ^ {M} \left| K _ {k, l} \right|, \quad \forall i, j \in \mathbf {P} _ {N E}, \tag {3} +$$ + +where $\mathbf{P}_{NE}$ denotes the set of positions that are not at the edge of the image, $K_{k,l}$ denotes the value of the $k,l$ -th position of the convolution kernel $\mathbf{K}$ and $M$ is the kernel size. We can infer that $p(\mathbf{z}_i) = 1$ , $\forall i \in \mathbf{P}_{NE}$ , i.e., the effect of all the internal tokens does not change during the convolution process. However, for self-attention in ViTs, (3) does not hold due to the non-existence of translation invariance. The fluctuation of $p(\mathbf{z})$ is further amplified by the input dependency of the spatial mixing matrix $\mathbf{w}^s (\mathbf{z})$ induced by self-attention. As an extreme case, we may obtain $p(\mathbf{z}) \sim 0$ for certain tokens. The fluctuation of tokens will alter the proportion of mixing (i.e., $\lambda$ ) and the network might even completely ignore one of the mixed images. The actual label of the processed tokens can then deviate from the mixed label computed by (1), resulting in less effective training. + +# 3.3. Token-Label Alignment + +Each token in ViTs interacts with other tokens using the self-attention mechanism. The input-dependent weights empower ViTs with more flexibility but also result in a mismatch between the processed token and the initial token. To address this, we propose a token-label alignment (TL-Align) method to trace the correspondence between the input and transformed tokens to obtain the aligned labels for the resulting representations, as illustrated in Figure 2. + +Specifically, ViTs first split the mixed input $\tilde{\mathbf{X}}$ after CutMix (1) to a sequence of $N$ non-overlapped patches and then flatten them to obtain the original image tokens + +![](images/293d805b64d2790b2045bbebfb4c2817de8ae864f0940b4bd76a5fa85ce1fba2.jpg) +Figure 2. Illustration of the proposed TL-Align. We trace the correspondence between the input tokens and the transformed tokens and align the labels accordingly. We reuse the computed attentions to linearly mix the labels of input tokens to obtain those of transformed tokens. The token-label alignment is performed iteratively to obtain a label for each output token. + +$\{\tilde{\mathbf{x}}_1,\tilde{\mathbf{x}}_2,\dots ,\tilde{\mathbf{x}}_N\}$ . We then project them into a proper dimension and add positional embeddings: + +$$ +\mathbf {Z} ^ {0} = [ \tilde {\mathbf {z}} _ {c l s}; \tilde {\mathbf {x}} _ {1} \cdot \mathbf {E}; \tilde {\mathbf {x}} _ {2} \cdot \mathbf {E}; \dots ; \tilde {\mathbf {x}} _ {N} \cdot \mathbf {E} ] + \mathbf {E} _ {p o s}, (4) +$$ + +where $\tilde{\mathbf{z}}_{cls} \in \mathcal{R}^{1 \times d}$ denotes the class token, $\mathbf{N}$ is the number of tokens, $\mathbf{E}$ represents the patch projector, and $\mathbf{E}_{pos} \in \mathcal{R}^{(N + 1) \times d}$ is the position embeddings. Note that we adopt the process of the original transformer architecture [16] as an example without loss of generality. Other models may omit the class token and use a relative positional embedding instead, which does not affect the utility of the proposed TL-Align method. + +We first assign each token $\mathbf{z}_i\in \mathcal{R}^{1\times d}$ with a label embedding $\mathbf{y}_i\in \mathcal{R}^{1\times C}$ : + +$$ +\mathbf {Y} ^ {0} = \left[ \tilde {\mathbf {y}} _ {c l s} ^ {0}; \tilde {\mathbf {y}} _ {1} ^ {0}; \tilde {\mathbf {y}} _ {2} ^ {0}; \dots ; \tilde {\mathbf {y}} _ {N} ^ {0} \right], \tag {5} +$$ + +where the sum of elements in each $\mathbf{y}_i$ equals 1 (i.e., $\sum_{j=1}^{C} y_{i,j} = 1$ ) and $y_{i,j}$ indicates how much the $i$ -th token belong to the $j$ -th class. We initialize the label embedding following the conventional data mixing paradigm. For example, when using CutMix to mix two images $\mathbf{X}_1$ and $\mathbf{X}_2$ from the $j$ -th class and the $k$ -th class with a mixing ratio of $\lambda$ , we set $\tilde{y}_{cls,j} = \lambda$ and $\tilde{y}_{cls,k} = 1 - \lambda$ for the class token. For each patch token, we set $\tilde{y}_{i,j} = 1$ if it comes from $\mathbf{X}_1$ and $\tilde{y}_{i,k} = 1$ if it comes from $\mathbf{X}_2$ . If a patch token contains both the mixed images, we use the mixing ratio within this patch as the label. For MixUp, we can simply set all label embeddings $\{\tilde{\mathbf{y}}_i\}$ with $\tilde{y}_{j,k} = \lambda$ and $\tilde{y}_{j,j} = 1 - \lambda$ . + +We perform TL-Align in a layer-wise manner and compute the aligned labels based on the operation on the tokens. Formally, ViTs use self-attention to perform spatial mixing + +of the input tokens $\mathbf{Z}$ + +$$ +\mathbf {Q} = \mathbf {Z} \cdot \mathbf {W} _ {\mathbf {Q}}, \mathbf {K} = \mathbf {Z} \cdot \mathbf {W} _ {\mathbf {K}}, \mathbf {V} = \mathbf {Z} \cdot \mathbf {W} _ {\mathbf {V}}, +$$ + +$$ +\mathcal {A} (\mathbf {Q}, \mathbf {K}) = \operatorname {S o f t m a x} \left(\mathbf {Q} \cdot \mathbf {K} ^ {T} / \sqrt {d}\right), \tag {6} +$$ + +$$ +\hat {\mathbf {Z}} = \operatorname {S A} (\mathbf {Z}) = \mathcal {A} (\mathbf {Q}, \mathbf {K}) \cdot \mathbf {V}. +$$ + +To align the labels, we update the label embeddings $\mathbf{Y}$ using the same attention matrix $\mathcal{A}(\mathbf{Q},\mathbf{K})$ + +$$ +\hat {\mathbf {Y}} = \mathcal {A} (\mathbf {Q}, \mathbf {K}) \cdot \mathbf {Y}. \tag {7} +$$ + +ViTs usually adopt multi-head self-attention (MSA) to perform multiple self-attention parallelly: + +$$ +\hat {\mathbf {Z}} = \operatorname {M S A} (\mathbf {Z}) = \left[ \mathrm {S A} _ {1} (\mathbf {Z}); \mathrm {S A} _ {2} (\mathbf {Z}); \dots ; \mathrm {S A} _ {H} (\mathbf {Z}) \right] \cdot \mathbf {w} _ {h}, \tag {8} +$$ + +where $H$ is the number of heads and $\mathbf{w}_h\in \mathcal{R}^{d\times d}$ . We then adapt our label alignment to MSA by simply taking the average of all the attention matrices for alignment: + +$$ +\hat {\mathbf {Y}} = \operatorname {T L} - \operatorname {A l i g n} - \mathbf {S} (\mathbf {Z}, \mathbf {Y}) := \frac {1}{H} \sum_ {i = 1} ^ {H} \mathcal {A} _ {i} (\mathbf {Q}, \mathbf {K}) \cdot \mathbf {Y}, \tag {9} +$$ + +where $\mathcal{A}_i$ is the attention matrix corresponding to the $i$ -th head $\mathrm{SA}_i$ . + +Each transformer block $l$ processes the tokens by both spatial and channel mixing: + +$$ +\hat {\mathbf {Z}} ^ {l - 1} = \operatorname {M S A} (\mathrm {L N} (\mathbf {Z} ^ {l - 1})), \quad \mathbf {Z} ^ {\prime l - 1} = \hat {\mathbf {Z}} ^ {l - 1} + \mathbf {Z} ^ {l - 1}, \tag {10} +$$ + +$$ +\hat {\mathbf {Z}} ^ {l} = \operatorname {M L P} (\operatorname {L N} \left(\mathbf {Z} ^ {\prime l - 1}\right)), \quad \mathbf {Z} ^ {l} = \hat {\mathbf {Z}} ^ {l} + \mathbf {Z} ^ {\prime l - 1}, +$$ + +where MLP and LN denote the MLP module and layer normalization [2], respectively. Our TL-Align then aligns the label embeddings in a similar manner: + +$$ +\hat {\mathbf {Y}} ^ {l - 1} = \operatorname {T L} \text {- A l i g n - S} \left(\mathbf {Y} ^ {l - 1}\right), \mathbf {Y} ^ {\prime l - 1} = \operatorname {N o r m} \left(\hat {\mathbf {Y}} ^ {l - 1} + \mathbf {Y} ^ {l - 1}\right), +$$ + +$$ +\hat {\mathbf {Y}} ^ {l} = \mathbf {Y} ^ {\prime l - 1}, \quad \mathbf {Y} ^ {l} = \operatorname {N o r m} \left(\hat {\mathbf {Y}} ^ {l} + \mathbf {Y} ^ {\prime l - 1}\right), \tag {11} +$$ + +where Norm denotes the normalization operation. We implement Norm by a simple average. + +Hierarchical vision transformers such as Swin [31] further introduce a patch aggregation operation to merge multiple patches. They usually concatenate multiple tokens across the channels to reduce the spatial resolution. Instead of concatenation, we simply add the label embeddings of the merged tokens followed by normalization as the aligned labels. The proposed TL-Align can be generalized to different architectures composed of spatial mixing, channel mixing, point-wise transformation, residual connection, and spatial aggregation. We provide detailed illustrations of alignment with these operations in the supplementary. + +We synchronously align the labels with the processed tokens layer by layer and obtain the aligned tokens $\mathbf{Z}^L$ and labels $\mathbf{Y}^L$ . The final representation of the image $\mathbf{z}$ is either the class token $\mathbf{z}_{cls}^L$ [16, 43] or the average pooling of all the spatial tokens $\frac{1}{N}\sum_{i=1}^{N}\mathbf{z}_i^L$ [31]. The aligned label $\mathbf{y}_{align}$ for the image is then $\mathbf{y}_{cls}^L$ or $\frac{1}{N}\sum_{i=1}^{N}\mathbf{y}_i^L$ depending on the specific model. We then adopt the aligned label $\mathbf{y}_{align}$ to train the network and can adapt to different loss functions and training schemes: + +$$ +J = J (\mathbf {z}, \text {s t o p - g r a d i e n t} (\mathbf {y} _ {\text {a l l i n g}})). \tag {12} +$$ + +We do not back-propagate through the aligned label as they only serve as a more accurate target. + +Our TL-Align serves as a plug-and-play module on various vision transformers while only introducing negligible training costs. We adjust the label of each token adaptively during the layer-by-layer propagation and preserve alignment between tokens and labels throughout the forward process. TL-Align is only used during training and introduces no additional computation cost when inference. + +# 4. Experiments + +In this section, we conducted extensive experiments to evaluate the proposed TL-Align method. We demonstrate the improvement of TL-Align on various vision transformers and compare it with state-of-the-art training strategies concerning the accuracy, network complexity, and training speed. We also examine the transferability on downstream tasks including semantic segmentation, object detection, and transfer learning. We further provide in-depth analysis to evaluate the effectiveness of TL-Align. + +# 4.1. ImageNet Classification + +Implementation Details. We first evaluate TL-Align on ImageNet [38] for image classification. We implement our method based on PyTorch [35] and the timm library [50]. We conduct experiments on various transformer architectures: three variants of DeiT [43] (DeiT-T, DeiT-S, and DeiT-B), two variants of PVT [49] (PVT-T, PVT-S) and + +Table 1. Results on ImageNet classification task. We compare the parameters, FLOPs, and accuracy of different vision transformer backbones without and with our TL-Align. + +
ModelImage SizeParamsFLOPsTop-1(%)Top-5(%)
DeiT-T22425.7M1.6G72.291.3
+TL-Align73.291.7
PVT-T224213.2M1.9G75.192.4
+TL-Align75.593.0
DeiT-S224222M4.6G79.895.0
+TL-Align80.695.0
PVT-S224224.5M3.8G79.895.0
+TL-Align80.495.5
Swin-T224229M4.5G81.295.5
+TL-Align81.495.7
Swin-S224250M8.8G83.096.3
+TL-Align83.496.5
DeiT-B224286M17.5G81.895.5
+TL-Align82.395.8
Swin-B224288M15.4G83.596.4
+TL-Align83.796.5
+ +Table 2. Comparison of our TL-Align with other training strategies on ImageNet. + +
MethodParamsTrainig TimeAcc.(%)
Vanilla22M1×76.4
CutMix [52]22M1×79.8
Puzzle-Mix [24]22M2.24×79.8
F-Mix [18]22M1.08×77.4
ResizeMix [36]22M1.03×78.6
SaliencyMix [45]22M1.26×79.2
Attentive-CutMix [48]46M1.31×77.5
TransMix [7]22M1.16×80.1
TokenMix [30]22M1.65×80.3
CutMix + TL-Align22M1.24×80.6
+ +Table 3. Results on semantic segmentation ADE20K. + +
BackboneParamsFLOPsmIoUmIoU (MS)mAcc
DeiT-S58M1032G43.845.155.2
+TL-Align44.545.755.5
Swin-T60M945G44.445.855.6
+TL-Align44.746.556.4
Swin-S81M1038G47.649.558.8
+TL-Align48.049.759.5
Swin-B121M1188G48.149.759.1
+TL-Align48.350.159.7
+ +three variants of Swin Transformer [31] (Swin-T, Swin-S, and Swin-B). For tiny and small models, we train from scratch for 300 epochs following the same training recipe using CutMix as DeiT [43], PVT [49] and Swin [31]. We keep all the data augmentation policies and all hyperparameter settings unchanged for fair comparisons. The only modification is that we replace the mixing targets in Cut + +Table 4. Experimental results on object detection and instance segmentation on COCO. + +
BackboneParamsFLOPsSchedule\(AP^{box}\)\(AP^{box}_{50}\)\(AP^{box}_{75}\)\(AP^{mask}\)\(AP^{mask}_{50}\)\(AP^{mask}_{75}\)
Swin-T+TL-Align86M745G3x50.469.254.743.766.647.3
50.569.454.943.866.647.3
Swin-S+TL-Align107M838G3x51.970.756.345.068.248.8
52.271.156.745.268.449.1
Swin-B+TL-Align145M982G3x51.970.556.445.068.148.9
52.371.256.945.368.749.1
+ +Table 5. Results on different transfer learning datasets. + +
ModelParamsFLOPsC-10C-100FlowersCars
ResNet5026M4.1G--96.290.0
ViT-B/1686M55.4G98.187.189.5-
ViT-L/16307M190.7G97.986.489.7-
DeiT5.7M1.6G97.685.797.190.1
+TL-Align5.7M1.6G97.886.497.990.7
DeiT-S22M4.6G97.990.298.191.4
+TL-Align22M4.6G98.890.498.391.8
DeiT-B86M17.5G99.190.898.492.1
+TL-Align86M17.5G99.190.598.693.0
+ +![](images/fec01cb0a93a68f3ac2faae047365d102018f2fef6fdadb42d492ced3e204b1a.jpg) +Figure 3. The Root Mean Square Error (RMSE) between original CutMix targets and labels obtained by T-L Align. + +MiX with the labels obtained by our TL-Align. For base models (i.e., Deit-B and Swin-B), we finetune the official pre-trained models for 40 epochs with a constant learning rate of 1e-5 and a weight decay of 1e-8. + +Performance on Different Architectures. As shown in Table 1, TL-Align steadily improves the performance of different vision transformer architectures. Specifically, TL-Align boosts the top-1 accuracy of DeiT-T, DeiT-S, and DeiT-B by $1.0\%$ , $0.8\%$ , and $0.5\%$ , respectively, in a parameter-free manner. Moreover, our method is generalizable and can be directly applied to hierarchical vision transformers like Swin. It is worth noting that most existing methods need either architecture modifications (adding a class token in [7]) or extra computations (saliency map extraction in [45]) when applied to Swin. In contrast, our TL-Align method can be used as a plug-and-play module and achieves consistent improvement on variants of Swin. + +![](images/792f8205bf4f9fa77f216fc97556f10c8eb50c38c5b774ff62be9edf7ad55d41.jpg) +Figure 4. Visualization of mixing ratio $\lambda$ of fluctuating tokens from different layers. We compare the results of TL-Align with CutMix, token similarity, TransMix, and TokenLabeling. + +Comparison with Other Training Strategies. We also compare our method with the state-of-the-art training strategies for data mixing on DeiT-S, as shown in Table 2. Specifically, we train the DeiT-S model while only disabling CutMix as the baseline method, which is denoted as Vanilla. Moreover, since TransMix [7] and TokenMix [30] report the EMA accuracy with different hyperparameters, we reproduce it under the same training recipe [43] for a fair comparison. We see that TL-Align shows better performance than the other mixup variants while maintaining the number of parameters and training speed. Puzzle-Mix obtains the same classification accuracy as CutMix but results in a much lower training speed as it relies on an extra model to get the optimal solution. SaliencyMix and Attentive-CutMix lead to performance degeneration when built upon DeiT-S backbone. Notably, our method also achieves higher top-1 accuracy than ViT-targeted TransMix and TokenMix. While the teacher annotator in TokenMix introduces heavy training costs, TL-Align obtains accurate token-label alignment in an efficient manner. Moreover, TransMix has limitations in handling token fluctuation and cannot be applied to ViTs without a class token, such as Swin. + +# 4.2. Downstream Tasks + +Semantic Segmentation. We evaluate our TL-Align on ADE20K dataset [55] for semantic segmentation. We adopt DeiT-S and three variants of Swin Transformer as backbones equipped with UpperNet for segmentation. As pre + +Table 6. Comparison results of model generalization ability and robustness. We evaluate them on various out-of-distribution/corrupted datasets and against adversarial attacks. $\uparrow$ denotes higher is better, and $\downarrow$ denotes lower is better. + +
ModelParamsFLOPsImageNetGeneralizationRobustness
Top-1↑Top-5↑IN-V2↑IN-A↑IN-C↓IN-R↑AutoAttack↑
DeiT-T5.7M1.6G72.291.360.47.769.134.13.9
+TL-Align5.7M1.6G73.291.761.46.168.034.64.4
DeiT-S22M4.6G79.895.068.518.954.742.56.9
+TL-Align22M4.6G80.695.068.919.253.243.27.5
DeiT-B86M17.5G81.895.570.527.948.545.3-
+TL-Align86M17.5G82.395.870.929.047.144.4-
+ +Table 7. Ablation of using different data mixing strategies. + +
MixUpCutMixRandomBlock-wiseTop-1 (%)+TL-Align Top-1 (%)
××××76.4-
×✓××79.880.6
✓✓××79.880.2
××✓×79.780.2
×××✓80.080.3
+ +sented in Table 3, TL-Align improves the segmentation performance on both DeiT and Swin at different model scales. + +Object Detection and Instance Segmentation. We also examine the performance of TL-Align on object detection and instance segmentation on the COCO 2017 dataset [29]. We apply our TL-Align to Swin [31] due to the advantage of the hierarchical representations on object detection tasks. We adopt the Cascade Mask-RCNN [4] framework and use the training strategy of $3 \times$ schedule. As shown in Table 4, we observe consistent improvements on all variants of Swin Transformer. This demonstrates the advantages to learn token-level meaningful features suitable for dense prediction tasks. + +Transfer Learning. We further evaluate the transferred classification performance of TL-Align on CIFAR-10 [26], CIFAR-100 [26], Flowers [34] and Cars [25]. We use pretrained models on ImageNet and finetune them on these datasets following existing works [43]. We compare the performance with and without TL-Align on three variants of DeiT [43], as shown in Table 5. TL-Align obtains significant performance gains for all variants on the four datasets. + +# 4.3. Performance Analysis and Visualization + +Effectiveness of Token-Label Alignment. We first quantize the difference between the original targets and aligned labels and investigate its correlation with the model. Specifically, we compute the Root Mean Square Error (RMSE) between the original targets and labels obtained by our TL-Align. As shown in Figure 3, the RMSE decreases when enlarging the model size. This indicates that larger models demonstrate less token fluctuation since the self-correction ability is also enhanced as the model capacity scales up. Moreover, the RMSE for Swin Transformer + +Table 8. Ablation of different TL-Align operations. + +
AlignmentTop-1 Acc.(%)Δ (%)
None (DeiT-S baseline)79.8-
TL-Align-S (Layer 12)80.1+0.3
TL-Align-S (Layer 2,4,6,8)80.2+0.4
Normalization Disabled80.3+0.5
Default (TL-Align)80.6+0.8
+ +tends to be lower compared with DeiT of similar model size. This is due to the local-window self-attention in Swin which preserves more local information. These observations are consistent with our experimental results: the improvements on small models and DeiT-like backbones tend to be more significant as they encounter more token fluctuation. + +Visualization of the Layer-wise Mixing Ratio of Fluctuated Tokens. To investigate the effectiveness of TL-Align, we compute a similarity-based "ground-truth" mixing ratio for each layer. Specifically, we compute the similarities of tokens between the mixed and unmixed images and use them as the label of each token. We compare them with the mixing ratios produced by TL-Align, CutMix [52], TransMix [7], and TokenLabeling [23]. As shown in Figure 4, the similarity-based mixing ratio changes at each layer, resulting from token fluctuation. However, CutMix, TransMix, and TokenLabeling assume the output tokens keep spatial correspondence with the input tokens and compute a fixed mixing ratio. TL-Align assigns dynamic labels to tokens using layer-wise alignment, which is more accurate compared with other methods. + +Evaluation of Robustness and Generalization. We further conduct experiments to validate the generalization and robustness of TL-Align, as shown in Table 6. We employ four corrupted and out-of-distribution datasets for robustness evaluation, including ImageNet-A [22], ImageNet-C [21], and ImageNet-R [20]. We also adopt AutoAttack [11] to evaluate the adversarial robustness on the ImageNet validation set. Due to memory limitation, we do not experiment with DeiT-B on AutoAttack. We use mean Corruption Error (mCE) for ImageNet-C and Top-1 Accuracy for others as the evaluation metric. For generalization evaluation, we adopt the ImageNet-V2 dataset [37]. We see that TL-Align improves both robustness and general + +![](images/b8bc8153a112392f9b8621f9444fda5e2399360da810779c8a2e58b6a808cfcd.jpg) +Figure 5. The visualization of results on DeiT-S and Swin-S. We visualize the input images, the mixed image, the original label embedding, and the label embedding after token-label alignment. + +ization, showing the superiority of adopting TL-Align for pre-training. + +Ablation Study on different Data Mixing Strategies. Due to the efficiency of the proposed layer-wise alignment, TL-Align can be directly applied to a wide range of data mixing strategies. We adopt MixUp, CutMix, a random mixing strategy and a block-wise mixing strategy to evaluate the generalizability of TL-Align. The random mixing and block-wise mixing strategies are inspired by MAE [19] and BEiT [3] and we replace the masking operation with image mixing on patch-level and block-level (both of size $16 \times 16$ ) respectively. The comparison results of training DeiT-S with and without our approach is demonstrated in Table 7. Specifically, TL-Align improves CutMix by $0.8\%$ , MixUp+CutMix by $0.4\%$ , random mixing by $0.5\%$ and block-wise mixing by $0.3\%$ respectively, further verifying the generalizability of the proposed TL-Align. + +Ablation Study on Different Label Alignment Operations. Our TL-Align aligns the labels with tokens transformed by spatial self-attention and residual connection layer-by-layer. To investigate the effect of reusing attention maps and normalization, we conduct an ablation study regarding different alignment operations on DeiT-S. We try aligning the labels only by using the attention map of Layer 12, which is equivalent to TransMix [7]. We also test the performance of applying alignment to several middle transformer layers and disabling normalization. As presented in Table 8, incomplete alignment at a part of layers marginally boosts the performance as it cannot well handle the token fluctuation issue. Disabling normalization leads to $0.3\%$ accuracy drop due to the inaccurate alignment at the presence of residual connections. This demonstrates the significance of the token-label alignment by attention utilization and normalization in a layer-wise manner. + +Visualizations of Aligned Labels. We visualize the labels obtained by TL-Align on DeiT-S [43] and Swin-S [31] + +as shown in Figure 5. Specifically, the aligned label embedding is obtained after the final transformer block for both DeiT-S and Swin-S. The value of the label embedding represents the probability of the belonged class of the corresponding token. We use red to denote larger probabilities towards the first image and blue for the second image. We observe that the aligned labels can deviate from the original labels and result in different mixing ratios for training. Therefore, using the original ratio as the training target may produce false training signals and lead to inferior performance. We see that TL-Align can correct the labels when the images are mixed with uninformative tokens. More visualization results are included in the supplementary. + +# 5. Conclusion + +In this paper, we have presented a token-label alignment method for training better vision transformers. As important subsets of data augmentation methods, data mixing strategies can generally improve the performance of both CNNs and ViTs. We identify a token fading issue for ViTs and address it by tracing the correspondence between transformed tokens and the original tokens to obtain a label for each output token to obtain more accurate training signals. Experimental results have demonstrated that our TL-Align can consistently improve the performance of various ViT models. The generalization performance of TL-Align to other architectures such as MLP-like models remains unknown and is a promising future direction to explore. + +# Acknowledgements + +This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160102 and in part by the National Natural Science Foundation of China under Grant 62125603. + +# References + +[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 2 +[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv, abs/1607.06450, 2016. 4 +[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.8 +[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, pages 6154-6162, 2018. 7 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 2 +[6] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310, 2021. 2 +[7] Jie-Neng Chen, Shuyang Sun, Ju He, Philip Torr, Alan Yuille, and Song Bai. Transmix: Attend to mix for vision transformers. arXiv preprint arXiv:2111.09833, 2021. 1, 3, 5, 6, 7, 8 +[8] Tianlong Chen, Zhenyu Zhang, Yu Cheng, Ahmed Awadallah, and Zhangyang Wang. The principle of diversity: Training stronger vision transformers calls for reducing all levels of redundancy. In CVPR, 2022. 1 +[9] Bowen Cheng, Alexander G Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In NeurIPS, 2021. 1, 2 +[10] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In NeurIPS, 2021. 1, 2 +[11] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216, 2020. 7 +[12] Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In ICCV, pages 2988-2997, 2021. 1, 2 +[13] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In CVPR, pages 1601-1610, 2021. 1 +[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 2 +[15] Xiaohan Ding, Xiangyu Zhang, Yizhuang Zhou, Jungong Han, Guiguang Ding, and Jian Sun. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 1 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 2, 4, 5 +[17] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 1 +[18] Ethan Harris, Antonia Marcu, Matthew Painter, Mahesan Niranjan, Adam Prügel-Bennett, and Jonathon Hare. Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047, 2020.5 +[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 8 +[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 7 +[21] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 7 +[22] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 7 +[23] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. In NeurIPS, 2021. 1, 2, 7 +[24] Jang-Hyun Kim, Wonho Choo, and Hyun Oh Song. Puzzle mix: Exploiting saliency and local statistics for optimal mixup. In ICML, pages 5275-5285, 2020. 1, 2, 3, 5 +[25] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCVW, pages 554–561, 2013. 7 +[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7 +[27] Wenbo Li, Xin Lu, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2 +[28] Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, and Xiaou Tang. Not all pixels are equal: Difficulty-aware semantic segmentation via deep layer cascade. In CVPR, pages 3193-3202, 2017. 1 +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 7 +[30] Jihao Liu, Boxiao Liu, Hang Zhou, Hongsheng Li, and Yu Liu. Tokenmix: Rethinking image mixing for data augmentation in vision transformers. In ECCV, pages 455-471, 2022. 5, 6 + +[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 5, 7, 8 +[32] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. arXiv preprint arXiv:2201.03545, 2022. 1 +[33] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. arXiv preprint arXiv:2106.13230, 2021. 2 +[34] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 7 +[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NIPS, pages 8026-8037, 2019. 5 +[36] Jie Qin, Jiemin Fang, Qian Zhang, Wenyu Liu, Xingang Wang, and Xinggang Wang. Resizemix: Mixing data with preserved object information and true labels. arXiv preprint arXiv:2012.11101, 2020. 5 +[37] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning, pages 5389-5400, 2019. 7 +[38] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 5 +[39] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, 2021. 1, 2 +[40] Yehui Tang, Kai Han, Jianyuan Guo, Chang Xu, Yanxi Li, Chao Xu, and Yunhe Wang. An image patch is a wave: Quantum inspired vision mlp. In CVPR, 2022. 1 +[41] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. NeurIPS, 34, 2021. 1 +[42] Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Gautier Izacard, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, et al. Resmlp: Feedforward networks for image classification with data-efficient training. arXiv preprint arXiv:2105.03404, 2021. 1 +[43] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 1, 2, 5, 6, 7, 8 +[44] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 1, 2 + +[45] AFM Uddin, Mst Monira, Wheemyung Shin, TaeChoong Chung, Sung-Ho Bae, et al. Saliencymix: A saliency guided data augmentation strategy for better regularization. arXiv preprint arXiv:2006.01791, 2020. 1, 2, 3, 5, 6 +[46] Shashanka Venkataramanan, Ewa Kijak, Laurent Amsaleg, and Yannis Avrithis. Alignmixup: Improving representations by interpolating aligned features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19174-19183, 2022. 2 +[47] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In ICML, pages 6438-6447, 2019. 1, 3 +[48] Devesh Walawalkar, Zhiqiang Shen, Zechun Liu, and Marios Savvides. Attentive cutmix: An enhanced data augmentation approach for deep learning based image classification. arXiv preprint arXiv:2003.13048, 2020. 1, 2, 3, 5 +[49] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. 5 +[50] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.5 +[51] Lingfeng Yang, Xiang Li, Borui Zhao, Renjie Song, and Jian Yang. Recursivelymix: Mixed learning with history. arXiv preprint arXiv:2203.06844, 2022. 2 +[52] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, pages 6023-6032, 2019. 1, 2, 3, 5, 7 +[53] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 1, 2, 3 +[54] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 1 +[55] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. IJCV, 127:302-321, 2019. 6 +[56] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 1, 2 \ No newline at end of file diff --git a/2023/Token-Label Alignment for Vision Transformers/images.zip b/2023/Token-Label Alignment for Vision Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ad84940839face9510130e9debb6921698bd9e11 --- /dev/null +++ b/2023/Token-Label Alignment for Vision Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4471d10e5d3c7cc63494fa8602cf073c1702d4aa26aa7d1ab98b45f7afdc1524 +size 706390 diff --git a/2023/Token-Label Alignment for Vision Transformers/layout.json b/2023/Token-Label Alignment for Vision Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a99ba3a76df288f317fed386fb016c58d3f54108 --- /dev/null +++ b/2023/Token-Label Alignment for Vision Transformers/layout.json @@ -0,0 +1,9039 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 147, + 103, + 446, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 103, + 446, + 121 + ], + "spans": [ + { + "bbox": [ + 147, + 103, + 446, + 121 + ], + "type": "text", + "content": "Token-Label Alignment for Vision Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "spans": [ + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "text", + "content": "Han Xiao" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "text", + "content": " Wenzhao Zheng" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "text", + "content": " Zheng Zhu" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "text", + "content": " Jie Zhou" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "text", + "content": " Jiwen Lu" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger}" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "text", + "content": "Beijing National Research Center for Information Science and Technology, China \n" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "text", + "content": "Department of Automation, Tsinghua University, China \n" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 75, + 141, + 517, + 216 + ], + "type": "text", + "content": "PhiGent Robotics \n{h-xiao20, zhengwz18}@mails.tsinghua.edu.cn; zhengzhu@ieee.org; {jzhou, lujiwen} @tsinghua.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 290, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 531 + ], + "type": "text", + "content": "Data mixing strategies (e.g., CutMix) have shown the ability to greatly improve the performance of convolutional neural networks (CNNs). They mix two images as inputs for training and assign them with a mixed label with the same ratio. While they are shown effective for vision transformers (ViTs), we identify a token fluctuation phenomenon that has suppressed the potential of data mixing strategies. We empirically observe that the contributions of input tokens fluctuate as forward propagating, which might induce a different mixing ratio in the output tokens. The training target computed by the original data mixing strategy can thus be inaccurate, resulting in less effective training. To address this, we propose a token-label alignment (TL-Align) method to trace the correspondence between transformed tokens and the original tokens to maintain a label for each token. We reuse the computed attention at each layer for efficient token-label alignment, introducing only negligible additional training costs. Extensive experiments demonstrate that our method improves the performance of ViTs on image classification, semantic segmentation, objective detection, and transfer learning tasks. Code is available at: https://github.com/Euphorial6/TL-Align." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 575, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 684 + ], + "type": "text", + "content": "The recent developments of vision transformers (ViTs) have revolutionized the computer vision field and set new state-of-the-arts in various tasks, such as image classification [16, 43, 31, 10], object detection [5, 56, 12, 13], and semantic segmentation [28, 39, 54, 9]. The successful structure of alternative spatial mixing and channel mixing in ViTs also motivates the arising of high-performance MLP-like deep architectures [40, 41, 42] and promotes the evolution of better CNNs [15, 32, 17]. In addition to architec" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 242, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 545, + 266 + ], + "type": "text", + "content": "ture designs, an improved training strategy can also greatly boost the performance of a trained deep model [23, 44, 8, 7]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 268, + 546, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 546, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 546, + 496 + ], + "type": "text", + "content": "The training of modern deep architecture almost all adopts data mixing strategies for data augmentation [48, 45, 24, 47, 52, 53], which have been proven to consistently improve the generalization performance. They randomly mix two images and their labels with the same mixing ratio to produce mixed data. As the most commonly used data mixing strategy, CutMix [52] performs a copy-and-paste operation on the spatial domain to produce spatially mixed images. While data mixing strategies have been widely studied for CNNs [48, 45, 24], few works have explored their compatibilities with ViTs [7]. We find that self-attention in ViTs causes a fluctuation of the original spatial structure. Unlike the translation equivalence that ensures a global label consistency for CNNs, self-attention in ViTs undermines this global consistency and causes a misalignment between the token and label. This misalignment induces a different mixing ratio in the output tokens. The training targets computed by the original data mixing strategies can then be inaccurate, resulting in less effective training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "content": "To address this, we propose a token-label alignment (TL-Align) method for ViTs to obtain a more accurate target for training. We present an overview of our method in Figure 1. We first assign a label to each input token in the mixed image according to the source of the token. We then trace the correspondence between the input tokens and the transformed tokens and align the labels accordingly. We assume that only the spatial self-attention and residual connection operation alter the presence of input tokens since channel MLP and layer normalization process each token independently. We reuse the computed attentions to linearly mix the labels of input tokens to obtain those of transformed tokens. The token-label alignment is performed iteratively to obtain a label for each output token. For class-token-based classification (e.g., ViT [16] and DeiT [43]), we directly use the aligned label for the output class token as the training target. For global-pooling-based classification (e.g., Swin [31]), we similarly average the labels of output tokens as the train" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5495" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 147, + 291 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 147, + 291 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 147, + 291 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 147, + 291 + ], + "type": "image", + "image_path": "a75d82a46cf1b38b0be800352cada4aca186c90985e32fccc083502cd5e3af83.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 293, + 111, + 302 + ], + "lines": [ + { + "bbox": [ + 72, + 293, + 111, + 302 + ], + "spans": [ + { + "bbox": [ + 72, + 293, + 111, + 302 + ], + "type": "text", + "content": "(a) CutMix" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 149, + 71, + 304, + 289 + ], + "blocks": [ + { + "bbox": [ + 149, + 71, + 304, + 289 + ], + "lines": [ + { + "bbox": [ + 149, + 71, + 304, + 289 + ], + "spans": [ + { + "bbox": [ + 149, + 71, + 304, + 289 + ], + "type": "image", + "image_path": "d34ab6d53d42770cb0ac9219357b66933ab60f35a6c7acb00557ef2a3336d492.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 293, + 285, + 303 + ], + "lines": [ + { + "bbox": [ + 174, + 293, + 285, + 303 + ], + "spans": [ + { + "bbox": [ + 174, + 293, + 285, + 303 + ], + "type": "text", + "content": "(b) Token misalignment for ViTs" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 542, + 290 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 542, + 290 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 542, + 290 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 542, + 290 + ], + "type": "image", + "image_path": "affaf24816e0d1ff852247366f51daa33c17b7a80f7201d6bdd47bbe24de14a9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 293, + 509, + 303 + ], + "lines": [ + { + "bbox": [ + 358, + 293, + 509, + 303 + ], + "spans": [ + { + "bbox": [ + 358, + 293, + 509, + 303 + ], + "type": "text", + "content": "(c) Different ways to obtain the mixed labels" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 304, + 547, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 547, + 350 + ], + "type": "text", + "content": "Figure 1. An overview of the proposed TL-Align. (a) CutMix-like methods [52] are widely used in model training, which spatially mix the tokens and their labels in the input space. (b) They are originally designed for CNNs and assume the processed tokens are spatially aligned with the input tokens. We show that it does not hold true for ViTs due to the global receptive field and adaptive weights. (c) Compared with existing methods, our method can effectively and efficiently align the tokens and labels without requiring a pretrained teacher network." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "type": "text", + "content": "ing target. The proposed TL-Align is only used for training to improve performance and introduces no additional workload for inference. We apply the proposed TL-Align to various ViT variants with CutMix including plain ViTs (DeiT [43]) and hierarchical ViTs (Swin [31]). We observe a consistent performance boost across different models on ImageNet-1K [14]. Specifically, our TL-Align improves DeiT-S by " + }, + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 46, + 352, + 289, + 495 + ], + "type": "text", + "content": " using the same training recipe. We evaluate the ImageNet-pretrained models on various downstream tasks including semantic segmentation, objection detection, and transfer learning. Experimental results also verify the robustness and generalization ability of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 512, + 134, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 134, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 134, + 525 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "Vision Transformer. Transformers have been widely used in natural language processing and achieved great success on many language tasks. Recently, Vision Transformers (ViTs) have aroused extensive interest in computer vision due to their competitive performance compared with CNNs [16, 43, 31, 10]. Dosovitskiy et al. [16] firstly introduced transformers into the image classification task. They split the input image into non-overlapped patches and then feed them into the transformer encoders. Liu et al. [31] proposed a shifted windowing scheme to produce hierarchical feature maps suitable for dense prediction tasks. The great potential of vision transformer has motivated its adaptation to many challenging tasks including object detection [12, 56, 5], segmentation [9, 39], image enhancement [6, 27] and video understanding [33, 1]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 352, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 483 + ], + "type": "text", + "content": "Recently, some efforts have been devoted to producing better training targets to improve the performance of vision transformers [23, 44]. For example, DeiT [43] introduces a knowledge distillation procedure to reduce the training cost of ViTs and achieves a better accuracy/speed trade-off. TokenLabeling [23] employs a pretrained teacher annotator to predict a label for each token for dense knowledge distillation. Differently, we do not require a pretrained network to obtain the training targets. Our TL-Align maintains an aligned label for each token layer by layer and can be trained efficiently in an end-to-end manner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": "Data Mixing Strategy. As an important type of data augmentation, data mixing strategies have demonstrated a consistent improvement in the generalization performance of CNNs. Zhang et al. [53] first proposed to combine a training pair to create augmented samples for model regularization. They perform linear interpolations on both the input images and associated targets. Following MixUp, CutMix [52] also utilizes the mixture of two input images but adopts a region copy-and-paste operation. Later methods including Puzzle Mix [24], SaliencyMix [45] and Attentive CutMix [48] leverage the salient regions for informative mixture generation. AlignMixUp [46] aligns two images geometrically in the feature space to improve the mixing representations. Recently, Yang et al. [51] proposed a RecursiveMix strategy which employs the historical input-prediction-label triplets for scale-invariant feature learning. Despite the better performance, a drawback of these methods is the heavily increased training cost due to the saliency extraction or historical information exploitation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5496" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 205 + ], + "type": "text", + "content": "Most existing data mixing methods are originally designed for CNNs, and their effectiveness on ViTs has not been well explored. TransMix [7] uses the class attention map at the last layer to re-weight the mixing targets and assumes the output tokens to keep spatial correspondence with the input tokens. However, we identify a token fluctuation phenomenon for ViTs which may cause a mismatch between tokens and labels, leading to inaccurate label assignments in both the original CutMix and TransMix. To address this, we propose to align the label and token space by tracing their correspondence in a layerwise manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 215, + 164, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 215, + 164, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 164, + 228 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 235, + 133, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 235, + 133, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 235, + 133, + 247 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 254, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 287, + 337 + ], + "type": "text", + "content": "The convolution neural network (CNN) has been the dominant architecture for computer vision in the deep learning era, greatly improving the performance of many tasks. Its monopoly has been challenged by the recent emergence of vision transformers (ViTs), which first \"patchify\" each image into tokens and process them with alternating self-attention (SA) and multi-layer perceptron (MLP)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "content": "In addition to architecture design, training strategy also has a large effect on the model performance, especially the data augmentation strategy. Data mixing [48, 45, 24, 47, 52, 53] is an important set of data augmentation for the training of both CNNs and ViTs, as it significantly improves the generalization ability of models. As the most commonly used data mixing strategy, CutMix [52] aims to create virtual training samples from the given training samples " + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}, y)" + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathcal{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "content": " denotes the input image and " + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "content": " is the corresponding label. CutMix randomly selects a local region from one input " + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1" + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "content": " and uses it to replace the pixels in the same region of another input " + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_2" + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "content": " to generate a new sample " + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}" + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "content": ". Similarly, the label " + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "content": " is also the combination of the original labels " + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "inline_equation", + "content": "y_1" + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "inline_equation", + "content": "y_2" + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 514, + 287, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 514, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 99, + 514, + 287, + 534 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {X}} = \\mathbf {M} \\odot \\mathbf {X} _ {1} + (\\mathbf {1} - \\mathbf {M}) \\odot \\mathbf {X} _ {2} \\tag {1}", + "image_path": "c97af1bb8f6101c2c3eaa49a4ed9bf1f72c036955deb58f754af95f91bc97f66.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 103, + 532, + 192, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 532, + 192, + 544 + ], + "spans": [ + { + "bbox": [ + 103, + 532, + 192, + 544 + ], + "type": "interline_equation", + "content": "\\tilde {y} = \\lambda y _ {1} + (1 - \\lambda) y _ {2}", + "image_path": "0eabb31dafa9cce1202a75d7b4bb9c949ba66585a5bdfd6e833f9ff08e5d3080.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "inline_equation", + "content": "M \\in \\{0,1\\}^{H \\times W}" + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "text", + "content": " is a binary mask indicating the image each pixel belongs to, " + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "text", + "content": " is an all-one matrix, and " + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "text", + "content": " is the element-wise multiplication. " + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "text", + "content": " reflects the mixing ratio of two labels and is the proportion of pixels cropped from " + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1" + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "text", + "content": " in the mixed image " + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{X}}" + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "text", + "content": ". For a cropped region " + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "inline_equation", + "content": "[r_x, r_x + r_w] \\times [r_y, r_y + r_h]" + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1" + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "text", + "content": ", we compute " + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\lambda = \\frac{r_w r_h}{WH}" + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "text", + "content": " to obtain the initial mixed target " + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\tilde{y}" + }, + { + "bbox": [ + 47, + 554, + 287, + 639 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 647, + 241, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 241, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 241, + 658 + ], + "type": "text", + "content": "3.2. The Token Fluctuation Phenomenon" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": "CutMix is originally designed for CNNs and assumes the feature extraction process does not alter the mixing ratio. However, we discover that different from CNNs, self-attention in ViTs can lead to the fluctuation of some tokens." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": "The fluctuation further results in the mismatch between the token space and label space, which hinders the effective training of the network." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": "Formally, we use " + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": " to denote a token of the image " + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": " is the transposed " + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": "-th column vector of " + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": ". We can then compute the " + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": "-th transformed token " + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_i" + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": " after the spatial operation as " + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_i = \\sum_{j=1}^{N} w_{i,j}^s \\mathbf{z}_j" + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "inline_equation", + "content": "w_{i,j}^s" + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": "-th element of the computed spatial mixing matrix " + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^s(\\mathbf{z})" + }, + { + "bbox": [ + 304, + 108, + 545, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 168, + 545, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 234 + ], + "type": "text", + "content": "With the assumption of the linear information integration, we define the contribution of an original token " + }, + { + "bbox": [ + 304, + 168, + 545, + 234 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 304, + 168, + 545, + 234 + ], + "type": "text", + "content": " to a mixed token " + }, + { + "bbox": [ + 304, + 168, + 545, + 234 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_j" + }, + { + "bbox": [ + 304, + 168, + 545, + 234 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 168, + 545, + 234 + ], + "type": "inline_equation", + "content": "c(\\mathbf{z}_i,\\hat{\\mathbf{z}}_j) = \\frac{|w_{i,j}^s|}{\\sum_{k = 1}^{N}|w_{k,j}^s|}" + }, + { + "bbox": [ + 304, + 168, + 545, + 234 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 168, + 545, + 234 + ], + "type": "inline_equation", + "content": "|\\cdot |" + }, + { + "bbox": [ + 304, + 168, + 545, + 234 + ], + "type": "text", + "content": " denotes the absolute value. We can then compute the presence of a token " + }, + { + "bbox": [ + 304, + 168, + 545, + 234 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 304, + 168, + 545, + 234 + ], + "type": "text", + "content": " in all the mixed image tokens as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 338, + 244, + 545, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 244, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 338, + 244, + 545, + 278 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {z} _ {i}\\right) = \\sum_ {j = 1} ^ {N} c \\left(\\mathbf {z} _ {i}, \\hat {\\mathbf {z}} _ {j}\\right) = \\sum_ {j = 1} ^ {N} \\frac {\\left| w _ {i , j} ^ {s} \\right|}{\\sum_ {k = 1} ^ {N} \\left| w _ {k , j} ^ {s} \\right|}. \\tag {2}", + "image_path": "a840c38a37f1c8aa9fb14ea035516e7f864d5ce0876aacc3ad194cb7280e215b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 287, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 545, + 322 + ], + "type": "text", + "content": "For non-strided depth-wise convolution, each token is multiplied by each element in the convolutional kernel due to the translation invariance. We can thus obtain:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "interline_equation", + "content": "\\sum_ {l = 1} ^ {N} \\left| w _ {i, l} ^ {s} \\right| = \\sum_ {j = 1} ^ {N} \\left| w _ {k, j} ^ {s} \\right| = \\sum_ {k = 1, l = 1} ^ {M} \\left| K _ {k, l} \\right|, \\quad \\forall i, j \\in \\mathbf {P} _ {N E}, \\tag {3}", + "image_path": "5ba1cda45d25b0979d65a704665718e3beeb0054912ff321694f59a63162d385.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{NE}" + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "text", + "content": " denotes the set of positions that are not at the edge of the image, " + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "inline_equation", + "content": "K_{k,l}" + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "text", + "content": " denotes the value of the " + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "inline_equation", + "content": "k,l" + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "text", + "content": "-th position of the convolution kernel " + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "text", + "content": " is the kernel size. We can infer that " + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z}_i) = 1" + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "inline_equation", + "content": "\\forall i \\in \\mathbf{P}_{NE}" + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "text", + "content": ", i.e., the effect of all the internal tokens does not change during the convolution process. However, for self-attention in ViTs, (3) does not hold due to the non-existence of translation invariance. The fluctuation of " + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z})" + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "text", + "content": " is further amplified by the input dependency of the spatial mixing matrix " + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^s (\\mathbf{z})" + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "text", + "content": " induced by self-attention. As an extreme case, we may obtain " + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z}) \\sim 0" + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "text", + "content": " for certain tokens. The fluctuation of tokens will alter the proportion of mixing (i.e., " + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 376, + 546, + 555 + ], + "type": "text", + "content": ") and the network might even completely ignore one of the mixed images. The actual label of the processed tokens can then deviate from the mixed label computed by (1), resulting in less effective training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 563, + 440, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 440, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 440, + 575 + ], + "type": "text", + "content": "3.3. Token-Label Alignment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 581, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 677 + ], + "type": "text", + "content": "Each token in ViTs interacts with other tokens using the self-attention mechanism. The input-dependent weights empower ViTs with more flexibility but also result in a mismatch between the processed token and the initial token. To address this, we propose a token-label alignment (TL-Align) method to trace the correspondence between the input and transformed tokens to obtain the aligned labels for the resulting representations, as illustrated in Figure 2." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Specifically, ViTs first split the mixed input " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " after CutMix (1) to a sequence of " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " non-overlapped patches and then flatten them to obtain the original image tokens" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5497" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 521, + 257 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 521, + 257 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 521, + 257 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 521, + 257 + ], + "type": "image", + "image_path": "293d805b64d2790b2045bbebfb4c2817de8ae864f0940b4bd76a5fa85ce1fba2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 257, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 547, + 289 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed TL-Align. We trace the correspondence between the input tokens and the transformed tokens and align the labels accordingly. We reuse the computed attentions to linearly mix the labels of input tokens to obtain those of transformed tokens. The token-label alignment is performed iteratively to obtain a label for each output token." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 292, + 288, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 288, + 316 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{x}}_1,\\tilde{\\mathbf{x}}_2,\\dots ,\\tilde{\\mathbf{x}}_N\\}" + }, + { + "bbox": [ + 47, + 292, + 288, + 316 + ], + "type": "text", + "content": " . We then project them into a proper dimension and add positional embeddings:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 331, + 287, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 331, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 63, + 331, + 287, + 346 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} ^ {0} = [ \\tilde {\\mathbf {z}} _ {c l s}; \\tilde {\\mathbf {x}} _ {1} \\cdot \\mathbf {E}; \\tilde {\\mathbf {x}} _ {2} \\cdot \\mathbf {E}; \\dots ; \\tilde {\\mathbf {x}} _ {N} \\cdot \\mathbf {E} ] + \\mathbf {E} _ {p o s}, (4)", + "image_path": "89a9e08fd71c54b47397296616977f437580235241b555bf5167b367b51b44cd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}_{cls} \\in \\mathcal{R}^{1 \\times d}" + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": " denotes the class token, " + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": " is the number of tokens, " + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": " represents the patch projector, and " + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{pos} \\in \\mathcal{R}^{(N + 1) \\times d}" + }, + { + "bbox": [ + 46, + 361, + 287, + 457 + ], + "type": "text", + "content": " is the position embeddings. Note that we adopt the process of the original transformer architecture [16] as an example without loss of generality. Other models may omit the class token and use a relative positional embedding instead, which does not affect the utility of the proposed TL-Align method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 485 + ], + "type": "text", + "content": "We first assign each token " + }, + { + "bbox": [ + 46, + 460, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i\\in \\mathcal{R}^{1\\times d}" + }, + { + "bbox": [ + 46, + 460, + 287, + 485 + ], + "type": "text", + "content": " with a label embedding " + }, + { + "bbox": [ + 46, + 460, + 287, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i\\in \\mathcal{R}^{1\\times C}" + }, + { + "bbox": [ + 46, + 460, + 287, + 485 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 500, + 287, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 500, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 108, + 500, + 287, + 515 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} ^ {0} = \\left[ \\tilde {\\mathbf {y}} _ {c l s} ^ {0}; \\tilde {\\mathbf {y}} _ {1} ^ {0}; \\tilde {\\mathbf {y}} _ {2} ^ {0}; \\dots ; \\tilde {\\mathbf {y}} _ {N} ^ {0} \\right], \\tag {5}", + "image_path": "224e018b39ec35fcc09f481e941a2be6e1afdf1ec6884face5820fa5ce994587.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": "where the sum of elements in each " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": " equals 1 (i.e., " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\sum_{j=1}^{C} y_{i,j} = 1" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "y_{i,j}" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": " indicates how much the " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": "-th token belong to the " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": "-th class. We initialize the label embedding following the conventional data mixing paradigm. For example, when using CutMix to mix two images " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_2" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": "-th class and the " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": "-th class with a mixing ratio of " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{cls,j} = \\lambda" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{cls,k} = 1 - \\lambda" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": " for the class token. For each patch token, we set " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{i,j} = 1" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": " if it comes from " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{i,k} = 1" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": " if it comes from " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_2" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": ". If a patch token contains both the mixed images, we use the mixing ratio within this patch as the label. For MixUp, we can simply set all label embeddings " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{y}}_i\\}" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{j,k} = \\lambda" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{j,j} = 1 - \\lambda" + }, + { + "bbox": [ + 46, + 530, + 288, + 674 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "We perform TL-Align in a layer-wise manner and compute the aligned labels based on the operation on the tokens. Formally, ViTs use self-attention to perform spatial mixing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 292, + 397, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 292, + 397, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 397, + 304 + ], + "type": "text", + "content": "of the input tokens " + }, + { + "bbox": [ + 305, + 292, + 397, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 334, + 308, + 518, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 308, + 518, + 322 + ], + "spans": [ + { + "bbox": [ + 334, + 308, + 518, + 322 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathbf {Z} \\cdot \\mathbf {W} _ {\\mathbf {Q}}, \\mathbf {K} = \\mathbf {Z} \\cdot \\mathbf {W} _ {\\mathbf {K}}, \\mathbf {V} = \\mathbf {Z} \\cdot \\mathbf {W} _ {\\mathbf {V}},", + "image_path": "1442bb1ea14f0194d1dd7e82bc17ae17b1f50c42d29d72562aed9db7f6bfa2da.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 334, + 324, + 545, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 324, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 334, + 324, + 545, + 339 + ], + "type": "interline_equation", + "content": "\\mathcal {A} (\\mathbf {Q}, \\mathbf {K}) = \\operatorname {S o f t m a x} \\left(\\mathbf {Q} \\cdot \\mathbf {K} ^ {T} / \\sqrt {d}\\right), \\tag {6}", + "image_path": "5385da2f6b416a25aa09ffb93cd822e2802de61817276f66cc677e020fe124e1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 334, + 342, + 456, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 342, + 456, + 357 + ], + "spans": [ + { + "bbox": [ + 334, + 342, + 456, + 357 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Z}} = \\operatorname {S A} (\\mathbf {Z}) = \\mathcal {A} (\\mathbf {Q}, \\mathbf {K}) \\cdot \\mathbf {V}.", + "image_path": "91fc9f30e630c941f06122f2a50fb73bc1d3459c87d3c1e3f1f85754c759de2f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 361, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 545, + 385 + ], + "type": "text", + "content": "To align the labels, we update the label embeddings " + }, + { + "bbox": [ + 305, + 361, + 545, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 305, + 361, + 545, + 385 + ], + "type": "text", + "content": " using the same attention matrix " + }, + { + "bbox": [ + 305, + 361, + 545, + 385 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\mathbf{Q},\\mathbf{K})" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 384, + 390, + 545, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 390, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 384, + 390, + 545, + 405 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Y}} = \\mathcal {A} (\\mathbf {Q}, \\mathbf {K}) \\cdot \\mathbf {Y}. \\tag {7}", + "image_path": "dbe65314fb3ddb539b4269c4180c1c11dffa73c4d027c979e9cef48a4d3f5358.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 410, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 410, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 410, + 545, + 434 + ], + "type": "text", + "content": "ViTs usually adopt multi-head self-attention (MSA) to perform multiple self-attention parallelly:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 310, + 438, + 545, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 438, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 310, + 438, + 545, + 463 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Z}} = \\operatorname {M S A} (\\mathbf {Z}) = \\left[ \\mathrm {S A} _ {1} (\\mathbf {Z}); \\mathrm {S A} _ {2} (\\mathbf {Z}); \\dots ; \\mathrm {S A} _ {H} (\\mathbf {Z}) \\right] \\cdot \\mathbf {w} _ {h}, \\tag {8}", + "image_path": "7187cce6a6de9b1cff4a77b848904d18b42947aeb2b9dc5365354a09766f1f71.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 463, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 463, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 305, + 463, + 545, + 500 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 463, + 545, + 500 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 305, + 463, + 545, + 500 + ], + "type": "text", + "content": " is the number of heads and " + }, + { + "bbox": [ + 305, + 463, + 545, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_h\\in \\mathcal{R}^{d\\times d}" + }, + { + "bbox": [ + 305, + 463, + 545, + 500 + ], + "type": "text", + "content": ". We then adapt our label alignment to MSA by simply taking the average of all the attention matrices for alignment:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 504, + 545, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 545, + 537 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Y}} = \\operatorname {T L} - \\operatorname {A l i g n} - \\mathbf {S} (\\mathbf {Z}, \\mathbf {Y}) := \\frac {1}{H} \\sum_ {i = 1} ^ {H} \\mathcal {A} _ {i} (\\mathbf {Q}, \\mathbf {K}) \\cdot \\mathbf {Y}, \\tag {9}", + "image_path": "3481989bdaa8a30e428b00f21db418279bc7f668a898a471ed2ed8c896ac0538.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 541, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 541, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 305, + 541, + 545, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 541, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 305, + 541, + 545, + 564 + ], + "type": "text", + "content": " is the attention matrix corresponding to the " + }, + { + "bbox": [ + 305, + 541, + 545, + 564 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 541, + 545, + 564 + ], + "type": "text", + "content": "-th head " + }, + { + "bbox": [ + 305, + 541, + 545, + 564 + ], + "type": "inline_equation", + "content": "\\mathrm{SA}_i" + }, + { + "bbox": [ + 305, + 541, + 545, + 564 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 566, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 566, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 305, + 566, + 545, + 590 + ], + "type": "text", + "content": "Each transformer block " + }, + { + "bbox": [ + 305, + 566, + 545, + 590 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 566, + 545, + 590 + ], + "type": "text", + "content": " processes the tokens by both spatial and channel mixing:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 594, + 545, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 545, + 613 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Z}} ^ {l - 1} = \\operatorname {M S A} (\\mathrm {L N} (\\mathbf {Z} ^ {l - 1})), \\quad \\mathbf {Z} ^ {\\prime l - 1} = \\hat {\\mathbf {Z}} ^ {l - 1} + \\mathbf {Z} ^ {l - 1}, \\tag {10}", + "image_path": "bea3b83d4b1cb1cc298a2ccc7cd71095385173c71396d8a4632db6783b514fe4.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 611, + 492, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 492, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 492, + 626 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Z}} ^ {l} = \\operatorname {M L P} (\\operatorname {L N} \\left(\\mathbf {Z} ^ {\\prime l - 1}\\right)), \\quad \\mathbf {Z} ^ {l} = \\hat {\\mathbf {Z}} ^ {l} + \\mathbf {Z} ^ {\\prime l - 1},", + "image_path": "4062701fe188a81589a9711c0e3cd9145f970c9d1e730bb9b2375c7ec8fae22d.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 630, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 630, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 305, + 630, + 545, + 666 + ], + "type": "text", + "content": "where MLP and LN denote the MLP module and layer normalization [2], respectively. Our TL-Align then aligns the label embeddings in a similar manner:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 671, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 553, + 685 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Y}} ^ {l - 1} = \\operatorname {T L} \\text {- A l i g n - S} \\left(\\mathbf {Y} ^ {l - 1}\\right), \\mathbf {Y} ^ {\\prime l - 1} = \\operatorname {N o r m} \\left(\\hat {\\mathbf {Y}} ^ {l - 1} + \\mathbf {Y} ^ {l - 1}\\right),", + "image_path": "f47104ff53e819b863da6b29a1612079c233906508d26bf0c6230a27868aed66.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 688, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 688, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 688, + 545, + 713 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Y}} ^ {l} = \\mathbf {Y} ^ {\\prime l - 1}, \\quad \\mathbf {Y} ^ {l} = \\operatorname {N o r m} \\left(\\hat {\\mathbf {Y}} ^ {l} + \\mathbf {Y} ^ {\\prime l - 1}\\right), \\tag {11}", + "image_path": "e6763619eddf6ae571a9602aaa975892cc942a376e508115246dd1f2d36021bb.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5498" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "where Norm denotes the normalization operation. We implement Norm by a simple average." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 228 + ], + "type": "text", + "content": "Hierarchical vision transformers such as Swin [31] further introduce a patch aggregation operation to merge multiple patches. They usually concatenate multiple tokens across the channels to reduce the spatial resolution. Instead of concatenation, we simply add the label embeddings of the merged tokens followed by normalization as the aligned labels. The proposed TL-Align can be generalized to different architectures composed of spatial mixing, channel mixing, point-wise transformation, residual connection, and spatial aggregation. We provide detailed illustrations of alignment with these operations in the supplementary." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "text", + "content": "We synchronously align the labels with the processed tokens layer by layer and obtain the aligned tokens " + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^L" + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "text", + "content": " and labels " + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^L" + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "text", + "content": ". The final representation of the image " + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "text", + "content": " is either the class token " + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{cls}^L" + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "text", + "content": " [16, 43] or the average pooling of all the spatial tokens " + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "inline_equation", + "content": "\\frac{1}{N}\\sum_{i=1}^{N}\\mathbf{z}_i^L" + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "text", + "content": " [31]. The aligned label " + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{align}" + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "text", + "content": " for the image is then " + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{cls}^L" + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "inline_equation", + "content": "\\frac{1}{N}\\sum_{i=1}^{N}\\mathbf{y}_i^L" + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "text", + "content": " depending on the specific model. We then adopt the aligned label " + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{align}" + }, + { + "bbox": [ + 46, + 228, + 288, + 339 + ], + "type": "text", + "content": " to train the network and can adapt to different loss functions and training schemes:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 348, + 287, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 348, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 99, + 348, + 287, + 361 + ], + "type": "interline_equation", + "content": "J = J (\\mathbf {z}, \\text {s t o p - g r a d i e n t} (\\mathbf {y} _ {\\text {a l l i n g}})). \\tag {12}", + "image_path": "5a1399418528e62fc08a0dba41965a9ad7bfd18c8dbea8fdea01206b3fcab916.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 287, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 393 + ], + "type": "text", + "content": "We do not back-propagate through the aligned label as they only serve as a more accurate target." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 287, + 477 + ], + "type": "text", + "content": "Our TL-Align serves as a plug-and-play module on various vision transformers while only introducing negligible training costs. We adjust the label of each token adaptively during the layer-by-layer propagation and preserve alignment between tokens and labels throughout the forward process. TL-Align is only used during training and introduces no additional computation cost when inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 488, + 128, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 488, + 128, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 488, + 128, + 502 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 508, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 616 + ], + "type": "text", + "content": "In this section, we conducted extensive experiments to evaluate the proposed TL-Align method. We demonstrate the improvement of TL-Align on various vision transformers and compare it with state-of-the-art training strategies concerning the accuracy, network complexity, and training speed. We also examine the transferability on downstream tasks including semantic segmentation, object detection, and transfer learning. We further provide in-depth analysis to evaluate the effectiveness of TL-Align." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 623, + 180, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 180, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 180, + 635 + ], + "type": "text", + "content": "4.1. ImageNet Classification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "Implementation Details. We first evaluate TL-Align on ImageNet [38] for image classification. We implement our method based on PyTorch [35] and the timm library [50]. We conduct experiments on various transformer architectures: three variants of DeiT [43] (DeiT-T, DeiT-S, and DeiT-B), two variants of PVT [49] (PVT-T, PVT-S) and" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 103, + 552, + 313 + ], + "blocks": [ + { + "bbox": [ + 305, + 68, + 545, + 102 + ], + "lines": [ + { + "bbox": [ + 305, + 68, + 545, + 102 + ], + "spans": [ + { + "bbox": [ + 305, + 68, + 545, + 102 + ], + "type": "text", + "content": "Table 1. Results on ImageNet classification task. We compare the parameters, FLOPs, and accuracy of different vision transformer backbones without and with our TL-Align." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 103, + 552, + 313 + ], + "lines": [ + { + "bbox": [ + 307, + 103, + 552, + 313 + ], + "spans": [ + { + "bbox": [ + 307, + 103, + 552, + 313 + ], + "type": "table", + "html": "
ModelImage SizeParamsFLOPsTop-1(%)Top-5(%)
DeiT-T22425.7M1.6G72.291.3
+TL-Align73.291.7
PVT-T224213.2M1.9G75.192.4
+TL-Align75.593.0
DeiT-S224222M4.6G79.895.0
+TL-Align80.695.0
PVT-S224224.5M3.8G79.895.0
+TL-Align80.495.5
Swin-T224229M4.5G81.295.5
+TL-Align81.495.7
Swin-S224250M8.8G83.096.3
+TL-Align83.496.5
DeiT-B224286M17.5G81.895.5
+TL-Align82.395.8
Swin-B224288M15.4G83.596.4
+TL-Align83.796.5
", + "image_path": "dc40af6d1ce3a61a9ede21307db38057b4fb1e8caf8790e1814ecfcff0c89f48.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 341, + 533, + 484 + ], + "blocks": [ + { + "bbox": [ + 305, + 317, + 545, + 339 + ], + "lines": [ + { + "bbox": [ + 305, + 317, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 305, + 317, + 545, + 339 + ], + "type": "text", + "content": "Table 2. Comparison of our TL-Align with other training strategies on ImageNet." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 341, + 533, + 484 + ], + "lines": [ + { + "bbox": [ + 317, + 341, + 533, + 484 + ], + "spans": [ + { + "bbox": [ + 317, + 341, + 533, + 484 + ], + "type": "table", + "html": "
MethodParamsTrainig TimeAcc.(%)
Vanilla22M1×76.4
CutMix [52]22M1×79.8
Puzzle-Mix [24]22M2.24×79.8
F-Mix [18]22M1.08×77.4
ResizeMix [36]22M1.03×78.6
SaliencyMix [45]22M1.26×79.2
Attentive-CutMix [48]46M1.31×77.5
TransMix [7]22M1.16×80.1
TokenMix [30]22M1.65×80.3
CutMix + TL-Align22M1.24×80.6
", + "image_path": "2651409140fa97e067440fbb4abbc5ffb7a86ef74a683d5ce133158025ee9d37.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 314, + 500, + 536, + 626 + ], + "blocks": [ + { + "bbox": [ + 323, + 487, + 525, + 498 + ], + "lines": [ + { + "bbox": [ + 323, + 487, + 525, + 498 + ], + "spans": [ + { + "bbox": [ + 323, + 487, + 525, + 498 + ], + "type": "text", + "content": "Table 3. Results on semantic segmentation ADE20K." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 500, + 536, + 626 + ], + "lines": [ + { + "bbox": [ + 314, + 500, + 536, + 626 + ], + "spans": [ + { + "bbox": [ + 314, + 500, + 536, + 626 + ], + "type": "table", + "html": "
BackboneParamsFLOPsmIoUmIoU (MS)mAcc
DeiT-S58M1032G43.845.155.2
+TL-Align44.545.755.5
Swin-T60M945G44.445.855.6
+TL-Align44.746.556.4
Swin-S81M1038G47.649.558.8
+TL-Align48.049.759.5
Swin-B121M1188G48.149.759.1
+TL-Align48.350.159.7
", + "image_path": "57af12b6b7aed1baed4fb5518f622093680d26f0716e36e493e0e7b57e1743c2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "three variants of Swin Transformer [31] (Swin-T, Swin-S, and Swin-B). For tiny and small models, we train from scratch for 300 epochs following the same training recipe using CutMix as DeiT [43], PVT [49] and Swin [31]. We keep all the data augmentation policies and all hyperparameter settings unchanged for fair comparisons. The only modification is that we replace the mixing targets in Cut" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5499" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 82, + 538, + 182 + ], + "blocks": [ + { + "bbox": [ + 127, + 69, + 463, + 80 + ], + "lines": [ + { + "bbox": [ + 127, + 69, + 463, + 80 + ], + "spans": [ + { + "bbox": [ + 127, + 69, + 463, + 80 + ], + "type": "text", + "content": "Table 4. Experimental results on object detection and instance segmentation on COCO." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 82, + 538, + 182 + ], + "lines": [ + { + "bbox": [ + 53, + 82, + 538, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 82, + 538, + 182 + ], + "type": "table", + "html": "
BackboneParamsFLOPsSchedule\\(AP^{box}\\)\\(AP^{box}_{50}\\)\\(AP^{box}_{75}\\)\\(AP^{mask}\\)\\(AP^{mask}_{50}\\)\\(AP^{mask}_{75}\\)
Swin-T+TL-Align86M745G3x50.469.254.743.766.647.3
50.569.454.943.866.647.3
Swin-S+TL-Align107M838G3x51.970.756.345.068.248.8
52.271.156.745.268.449.1
Swin-B+TL-Align145M982G3x51.970.556.445.068.148.9
52.371.256.945.368.749.1
", + "image_path": "edc30d1a787d1ea1bd14da599d54693b20d85b263f18db9fc90af5486b3c6d5e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 52, + 198, + 281, + 324 + ], + "blocks": [ + { + "bbox": [ + 59, + 185, + 271, + 195 + ], + "lines": [ + { + "bbox": [ + 59, + 185, + 271, + 195 + ], + "spans": [ + { + "bbox": [ + 59, + 185, + 271, + 195 + ], + "type": "text", + "content": "Table 5. Results on different transfer learning datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 198, + 281, + 324 + ], + "lines": [ + { + "bbox": [ + 52, + 198, + 281, + 324 + ], + "spans": [ + { + "bbox": [ + 52, + 198, + 281, + 324 + ], + "type": "table", + "html": "
ModelParamsFLOPsC-10C-100FlowersCars
ResNet5026M4.1G--96.290.0
ViT-B/1686M55.4G98.187.189.5-
ViT-L/16307M190.7G97.986.489.7-
DeiT5.7M1.6G97.685.797.190.1
+TL-Align5.7M1.6G97.886.497.990.7
DeiT-S22M4.6G97.990.298.191.4
+TL-Align22M4.6G98.890.498.391.8
DeiT-B86M17.5G99.190.898.492.1
+TL-Align86M17.5G99.190.598.693.0
", + "image_path": "670875c6252a1071265dafff7b19b0f2a5103e87c2593bc2df82efb1fcfba067.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 51, + 329, + 280, + 485 + ], + "blocks": [ + { + "bbox": [ + 51, + 329, + 280, + 485 + ], + "lines": [ + { + "bbox": [ + 51, + 329, + 280, + 485 + ], + "spans": [ + { + "bbox": [ + 51, + 329, + 280, + 485 + ], + "type": "image", + "image_path": "fec01cb0a93a68f3ac2faae047365d102018f2fef6fdadb42d492ced3e204b1a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 485, + 287, + 507 + ], + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 507 + ], + "type": "text", + "content": "Figure 3. The Root Mean Square Error (RMSE) between original CutMix targets and labels obtained by T-L Align." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 509, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 556 + ], + "type": "text", + "content": "MiX with the labels obtained by our TL-Align. For base models (i.e., Deit-B and Swin-B), we finetune the official pre-trained models for 40 epochs with a constant learning rate of 1e-5 and a weight decay of 1e-8." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": "Performance on Different Architectures. As shown in Table 1, TL-Align steadily improves the performance of different vision transformer architectures. Specifically, TL-Align boosts the top-1 accuracy of DeiT-T, DeiT-S, and DeiT-B by " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": ", respectively, in a parameter-free manner. Moreover, our method is generalizable and can be directly applied to hierarchical vision transformers like Swin. It is worth noting that most existing methods need either architecture modifications (adding a class token in [7]) or extra computations (saliency map extraction in [45]) when applied to Swin. In contrast, our TL-Align method can be used as a plug-and-play module and achieves consistent improvement on variants of Swin." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 186, + 531, + 342 + ], + "blocks": [ + { + "bbox": [ + 317, + 186, + 531, + 342 + ], + "lines": [ + { + "bbox": [ + 317, + 186, + 531, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 531, + 342 + ], + "type": "image", + "image_path": "792f8205bf4f9fa77f216fc97556f10c8eb50c38c5b774ff62be9edf7ad55d41.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 342, + 545, + 374 + ], + "lines": [ + { + "bbox": [ + 305, + 342, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 545, + 374 + ], + "type": "text", + "content": "Figure 4. Visualization of mixing ratio " + }, + { + "bbox": [ + 305, + 342, + 545, + 374 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 342, + 545, + 374 + ], + "type": "text", + "content": " of fluctuating tokens from different layers. We compare the results of TL-Align with CutMix, token similarity, TransMix, and TokenLabeling." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 376, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 639 + ], + "type": "text", + "content": "Comparison with Other Training Strategies. We also compare our method with the state-of-the-art training strategies for data mixing on DeiT-S, as shown in Table 2. Specifically, we train the DeiT-S model while only disabling CutMix as the baseline method, which is denoted as Vanilla. Moreover, since TransMix [7] and TokenMix [30] report the EMA accuracy with different hyperparameters, we reproduce it under the same training recipe [43] for a fair comparison. We see that TL-Align shows better performance than the other mixup variants while maintaining the number of parameters and training speed. Puzzle-Mix obtains the same classification accuracy as CutMix but results in a much lower training speed as it relies on an extra model to get the optimal solution. SaliencyMix and Attentive-CutMix lead to performance degeneration when built upon DeiT-S backbone. Notably, our method also achieves higher top-1 accuracy than ViT-targeted TransMix and TokenMix. While the teacher annotator in TokenMix introduces heavy training costs, TL-Align obtains accurate token-label alignment in an efficient manner. Moreover, TransMix has limitations in handling token fluctuation and cannot be applied to ViTs without a class token, such as Swin." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 647, + 417, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 417, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 417, + 658 + ], + "type": "text", + "content": "4.2. Downstream Tasks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "Semantic Segmentation. We evaluate our TL-Align on ADE20K dataset [55] for semantic segmentation. We adopt DeiT-S and three variants of Swin Transformer as backbones equipped with UpperNet for segmentation. As pre" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5500" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 93, + 526, + 209 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "type": "text", + "content": "Table 6. Comparison results of model generalization ability and robustness. We evaluate them on various out-of-distribution/corrupted datasets and against adversarial attacks. " + }, + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "type": "text", + "content": " denotes higher is better, and " + }, + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 47, + 69, + 545, + 91 + ], + "type": "text", + "content": " denotes lower is better." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 93, + 526, + 209 + ], + "lines": [ + { + "bbox": [ + 64, + 93, + 526, + 209 + ], + "spans": [ + { + "bbox": [ + 64, + 93, + 526, + 209 + ], + "type": "table", + "html": "
ModelParamsFLOPsImageNetGeneralizationRobustness
Top-1↑Top-5↑IN-V2↑IN-A↑IN-C↓IN-R↑AutoAttack↑
DeiT-T5.7M1.6G72.291.360.47.769.134.13.9
+TL-Align5.7M1.6G73.291.761.46.168.034.64.4
DeiT-S22M4.6G79.895.068.518.954.742.56.9
+TL-Align22M4.6G80.695.068.919.253.243.27.5
DeiT-B86M17.5G81.895.570.527.948.545.3-
+TL-Align86M17.5G82.395.870.929.047.144.4-
", + "image_path": "8e9c3962b662580f7ba72edd7c02f145e24a0ae4f184476121bab7dee78529e1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 224, + 294, + 314 + ], + "blocks": [ + { + "bbox": [ + 47, + 211, + 271, + 222 + ], + "lines": [ + { + "bbox": [ + 47, + 211, + 271, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 271, + 222 + ], + "type": "text", + "content": "Table 7. Ablation of using different data mixing strategies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 224, + 294, + 314 + ], + "lines": [ + { + "bbox": [ + 48, + 224, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 48, + 224, + 294, + 314 + ], + "type": "table", + "html": "
MixUpCutMixRandomBlock-wiseTop-1 (%)+TL-Align Top-1 (%)
××××76.4-
×✓××79.880.6
✓✓××79.880.2
××✓×79.780.2
×××✓80.080.3
", + "image_path": "4661141df34944a9ee63e80676adb1d7cd48e3e966c070b36e2d4cc76ecfe8b3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 316, + 286, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 286, + 339 + ], + "type": "text", + "content": "sented in Table 3, TL-Align improves the segmentation performance on both DeiT and Swin at different model scales." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 340, + 286, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 286, + 470 + ], + "type": "text", + "content": "Object Detection and Instance Segmentation. We also examine the performance of TL-Align on object detection and instance segmentation on the COCO 2017 dataset [29]. We apply our TL-Align to Swin [31] due to the advantage of the hierarchical representations on object detection tasks. We adopt the Cascade Mask-RCNN [4] framework and use the training strategy of " + }, + { + "bbox": [ + 46, + 340, + 286, + 470 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 340, + 286, + 470 + ], + "type": "text", + "content": " schedule. As shown in Table 4, we observe consistent improvements on all variants of Swin Transformer. This demonstrates the advantages to learn token-level meaningful features suitable for dense prediction tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 472, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 286, + 567 + ], + "type": "text", + "content": "Transfer Learning. We further evaluate the transferred classification performance of TL-Align on CIFAR-10 [26], CIFAR-100 [26], Flowers [34] and Cars [25]. We use pretrained models on ImageNet and finetune them on these datasets following existing works [43]. We compare the performance with and without TL-Align on three variants of DeiT [43], as shown in Table 5. TL-Align obtains significant performance gains for all variants on the four datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 575, + 257, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 257, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 257, + 588 + ], + "type": "text", + "content": "4.3. Performance Analysis and Visualization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Effectiveness of Token-Label Alignment. We first quantize the difference between the original targets and aligned labels and investigate its correlation with the model. Specifically, we compute the Root Mean Square Error (RMSE) between the original targets and labels obtained by our TL-Align. As shown in Figure 3, the RMSE decreases when enlarging the model size. This indicates that larger models demonstrate less token fluctuation since the self-correction ability is also enhanced as the model capacity scales up. Moreover, the RMSE for Swin Transformer" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 325, + 224, + 525, + 303 + ], + "blocks": [ + { + "bbox": [ + 326, + 211, + 522, + 222 + ], + "lines": [ + { + "bbox": [ + 326, + 211, + 522, + 222 + ], + "spans": [ + { + "bbox": [ + 326, + 211, + 522, + 222 + ], + "type": "text", + "content": "Table 8. Ablation of different TL-Align operations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 224, + 525, + 303 + ], + "lines": [ + { + "bbox": [ + 325, + 224, + 525, + 303 + ], + "spans": [ + { + "bbox": [ + 325, + 224, + 525, + 303 + ], + "type": "table", + "html": "
AlignmentTop-1 Acc.(%)Δ (%)
None (DeiT-S baseline)79.8-
TL-Align-S (Layer 12)80.1+0.3
TL-Align-S (Layer 2,4,6,8)80.2+0.4
Normalization Disabled80.3+0.5
Default (TL-Align)80.6+0.8
", + "image_path": "17d7926aafdecf807109451487201208a5c4ba8d8bd8e309147e99512a8bc82f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 376 + ], + "type": "text", + "content": "tends to be lower compared with DeiT of similar model size. This is due to the local-window self-attention in Swin which preserves more local information. These observations are consistent with our experimental results: the improvements on small models and DeiT-like backbones tend to be more significant as they encounter more token fluctuation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 378, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 556 + ], + "type": "text", + "content": "Visualization of the Layer-wise Mixing Ratio of Fluctuated Tokens. To investigate the effectiveness of TL-Align, we compute a similarity-based \"ground-truth\" mixing ratio for each layer. Specifically, we compute the similarities of tokens between the mixed and unmixed images and use them as the label of each token. We compare them with the mixing ratios produced by TL-Align, CutMix [52], TransMix [7], and TokenLabeling [23]. As shown in Figure 4, the similarity-based mixing ratio changes at each layer, resulting from token fluctuation. However, CutMix, TransMix, and TokenLabeling assume the output tokens keep spatial correspondence with the input tokens and compute a fixed mixing ratio. TL-Align assigns dynamic labels to tokens using layer-wise alignment, which is more accurate compared with other methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Evaluation of Robustness and Generalization. We further conduct experiments to validate the generalization and robustness of TL-Align, as shown in Table 6. We employ four corrupted and out-of-distribution datasets for robustness evaluation, including ImageNet-A [22], ImageNet-C [21], and ImageNet-R [20]. We also adopt AutoAttack [11] to evaluate the adversarial robustness on the ImageNet validation set. Due to memory limitation, we do not experiment with DeiT-B on AutoAttack. We use mean Corruption Error (mCE) for ImageNet-C and Top-1 Accuracy for others as the evaluation metric. For generalization evaluation, we adopt the ImageNet-V2 dataset [37]. We see that TL-Align improves both robustness and general" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5501" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 541, + 250 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 541, + 250 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 541, + 250 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 541, + 250 + ], + "type": "image", + "image_path": "b8bc8153a112392f9b8621f9444fda5e2399360da810779c8a2e58b6a808cfcd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 546, + 280 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 546, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 546, + 280 + ], + "type": "text", + "content": "Figure 5. The visualization of results on DeiT-S and Swin-S. We visualize the input images, the mixed image, the original label embedding, and the label embedding after token-label alignment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 283, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 307 + ], + "type": "text", + "content": "ization, showing the superiority of adopting TL-Align for pre-training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 307, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 486 + ], + "type": "text", + "content": "Ablation Study on different Data Mixing Strategies. Due to the efficiency of the proposed layer-wise alignment, TL-Align can be directly applied to a wide range of data mixing strategies. We adopt MixUp, CutMix, a random mixing strategy and a block-wise mixing strategy to evaluate the generalizability of TL-Align. The random mixing and block-wise mixing strategies are inspired by MAE [19] and BEiT [3] and we replace the masking operation with image mixing on patch-level and block-level (both of size " + }, + { + "bbox": [ + 46, + 307, + 287, + 486 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 307, + 287, + 486 + ], + "type": "text", + "content": ") respectively. The comparison results of training DeiT-S with and without our approach is demonstrated in Table 7. Specifically, TL-Align improves CutMix by " + }, + { + "bbox": [ + 46, + 307, + 287, + 486 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 46, + 307, + 287, + 486 + ], + "type": "text", + "content": ", MixUp+CutMix by " + }, + { + "bbox": [ + 46, + 307, + 287, + 486 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 46, + 307, + 287, + 486 + ], + "type": "text", + "content": ", random mixing by " + }, + { + "bbox": [ + 46, + 307, + 287, + 486 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 307, + 287, + 486 + ], + "type": "text", + "content": " and block-wise mixing by " + }, + { + "bbox": [ + 46, + 307, + 287, + 486 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 46, + 307, + 287, + 486 + ], + "type": "text", + "content": " respectively, further verifying the generalizability of the proposed TL-Align." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 487, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 689 + ], + "type": "text", + "content": "Ablation Study on Different Label Alignment Operations. Our TL-Align aligns the labels with tokens transformed by spatial self-attention and residual connection layer-by-layer. To investigate the effect of reusing attention maps and normalization, we conduct an ablation study regarding different alignment operations on DeiT-S. We try aligning the labels only by using the attention map of Layer 12, which is equivalent to TransMix [7]. We also test the performance of applying alignment to several middle transformer layers and disabling normalization. As presented in Table 8, incomplete alignment at a part of layers marginally boosts the performance as it cannot well handle the token fluctuation issue. Disabling normalization leads to " + }, + { + "bbox": [ + 46, + 487, + 287, + 689 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 46, + 487, + 287, + 689 + ], + "type": "text", + "content": " accuracy drop due to the inaccurate alignment at the presence of residual connections. This demonstrates the significance of the token-label alignment by attention utilization and normalization in a layer-wise manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Visualizations of Aligned Labels. We visualize the labels obtained by TL-Align on DeiT-S [43] and Swin-S [31]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 283, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 545, + 437 + ], + "type": "text", + "content": "as shown in Figure 5. Specifically, the aligned label embedding is obtained after the final transformer block for both DeiT-S and Swin-S. The value of the label embedding represents the probability of the belonged class of the corresponding token. We use red to denote larger probabilities towards the first image and blue for the second image. We observe that the aligned labels can deviate from the original labels and result in different mixing ratios for training. Therefore, using the original ratio as the training target may produce false training signals and lead to inferior performance. We see that TL-Align can correct the labels when the images are mixed with uninformative tokens. More visualization results are included in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 453, + 378, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 378, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 378, + 464 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 474, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 629 + ], + "type": "text", + "content": "In this paper, we have presented a token-label alignment method for training better vision transformers. As important subsets of data augmentation methods, data mixing strategies can generally improve the performance of both CNNs and ViTs. We identify a token fading issue for ViTs and address it by tracing the correspondence between transformed tokens and the original tokens to obtain a label for each output token to obtain more accurate training signals. Experimental results have demonstrated that our TL-Align can consistently improve the performance of various ViT models. The generalization performance of TL-Align to other architectures such as MLP-like models remains unknown and is a promising future direction to explore." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 644, + 408, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 644, + 408, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 644, + 408, + 658 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "This work was supported in part by the National Key Research and Development Program of China under Grant 2022ZD0160102 and in part by the National Natural Science Foundation of China under Grant 62125603." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5502" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 92, + 286, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 286, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 286, + 125 + ], + "type": "text", + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 127, + 286, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 127, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 127, + 286, + 148 + ], + "type": "text", + "content": "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv, abs/1607.06450, 2016. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 151, + 286, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 151, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 53, + 151, + 286, + 182 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 186, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 286, + 217 + ], + "type": "text", + "content": "[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, pages 6154-6162, 2018. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 220, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 286, + 262 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 266, + 286, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 266, + 286, + 309 + ], + "spans": [ + { + "bbox": [ + 53, + 266, + 286, + 309 + ], + "type": "text", + "content": "[6] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 312, + 286, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 312, + 286, + 355 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 286, + 355 + ], + "type": "text", + "content": "[7] Jie-Neng Chen, Shuyang Sun, Ju He, Philip Torr, Alan Yuille, and Song Bai. Transmix: Attend to mix for vision transformers. arXiv preprint arXiv:2111.09833, 2021. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 357, + 286, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 286, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 286, + 401 + ], + "type": "text", + "content": "[8] Tianlong Chen, Zhenyu Zhang, Yu Cheng, Ahmed Awadallah, and Zhangyang Wang. The principle of diversity: Training stronger vision transformers calls for reducing all levels of redundancy. In CVPR, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 404, + 286, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 286, + 435 + ], + "type": "text", + "content": "[9] Bowen Cheng, Alexander G Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In NeurIPS, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 438, + 286, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 438, + 286, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 438, + 286, + 481 + ], + "type": "text", + "content": "[10] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In NeurIPS, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 484, + 286, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 484, + 286, + 528 + ], + "spans": [ + { + "bbox": [ + 48, + 484, + 286, + 528 + ], + "type": "text", + "content": "[11] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216, 2020. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 530, + 286, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 530, + 286, + 573 + ], + "spans": [ + { + "bbox": [ + 48, + 530, + 286, + 573 + ], + "type": "text", + "content": "[12] Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In ICCV, pages 2988-2997, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 576, + 286, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 576, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 576, + 286, + 609 + ], + "type": "text", + "content": "[13] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In CVPR, pages 1601-1610, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 611, + 286, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 286, + 643 + ], + "type": "text", + "content": "[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 646, + 286, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 286, + 688 + ], + "type": "text", + "content": "[15] Xiaohan Ding, Xiangyu Zhang, Yizhuang Zhou, Jungong Han, Guiguang Ding, and Jian Sun. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1, 2, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 119, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 119, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 308, + 119, + 545, + 151 + ], + "type": "text", + "content": "[17] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 154, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 154, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 154, + 545, + 196 + ], + "type": "text", + "content": "[18] Ethan Harris, Antonia Marcu, Matthew Painter, Mahesan Niranjan, Adam Prügel-Bennett, and Jonathon Hare. Fmix: Enhancing mixed sample data augmentation. arXiv preprint arXiv:2002.12047, 2020.5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 199, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 199, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 199, + 545, + 252 + ], + "type": "text", + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 256, + 545, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 256, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 308, + 256, + 545, + 321 + ], + "type": "text", + "content": "[20] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 323, + 545, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 323, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 308, + 323, + 545, + 356 + ], + "type": "text", + "content": "[21] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 358, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 358, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 308, + 358, + 545, + 402 + ], + "type": "text", + "content": "[22] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 404, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 446 + ], + "type": "text", + "content": "[23] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. In NeurIPS, 2021. 1, 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 450, + 545, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 450, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 308, + 450, + 545, + 482 + ], + "type": "text", + "content": "[24] Jang-Hyun Kim, Wonho Choo, and Hyun Oh Song. Puzzle mix: Exploiting saliency and local statistics for optimal mixup. In ICML, pages 5275-5285, 2020. 1, 2, 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 485, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 485, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 308, + 485, + 545, + 517 + ], + "type": "text", + "content": "[25] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCVW, pages 554–561, 2013. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 520, + 545, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 520, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 308, + 520, + 545, + 541 + ], + "type": "text", + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 544, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 575 + ], + "type": "text", + "content": "[27] Wenbo Li, Xin Lu, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 578, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 620 + ], + "type": "text", + "content": "[28] Xiaoxiao Li, Ziwei Liu, Ping Luo, Chen Change Loy, and Xiaou Tang. Not all pixels are equal: Difficulty-aware semantic segmentation via deep layer cascade. In CVPR, pages 3193-3202, 2017. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[30] Jihao Liu, Boxiao Liu, Hang Zhou, Hongsheng Li, and Yu Liu. Tokenmix: Rethinking image mixing for data augmentation in vision transformers. In ECCV, pages 455-471, 2022. 5, 6" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5503" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 151 + ], + "type": "text", + "content": "[32] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. arXiv preprint arXiv:2201.03545, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[33] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. arXiv preprint arXiv:2106.13230, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 230 + ], + "type": "text", + "content": "[34] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "type": "text", + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NIPS, pages 8026-8037, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "type": "text", + "content": "[36] Jie Qin, Jiemin Fang, Qian Zhang, Wenyu Liu, Xingang Wang, and Xinggang Wang. Resizemix: Mixing data with preserved object information and true labels. arXiv preprint arXiv:2012.11101, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "text", + "content": "[37] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning, pages 5389-5400, 2019. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 378, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 430 + ], + "type": "text", + "content": "[38] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 465 + ], + "type": "text", + "content": "[39] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 500 + ], + "type": "text", + "content": "[40] Yehui Tang, Kai Han, Jianyuan Guo, Chang Xu, Yanxi Li, Chao Xu, and Yunhe Wang. An image patch is a wave: Quantum inspired vision mlp. In CVPR, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 555 + ], + "type": "text", + "content": "[41] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. NeurIPS, 34, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 621 + ], + "type": "text", + "content": "[42] Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Gautier Izacard, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, et al. Resmlp: Feedforward networks for image classification with data-efficient training. arXiv preprint arXiv:2105.03404, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[43] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[44] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. arXiv preprint arXiv:2204.07118, 2022. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 587 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[45] AFM Uddin, Mst Monira, Wheemyung Shin, TaeChoong Chung, Sung-Ho Bae, et al. Saliencymix: A saliency guided data augmentation strategy for better regularization. arXiv preprint arXiv:2006.01791, 2020. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": "[46] Shashanka Venkataramanan, Ewa Kijak, Laurent Amsaleg, and Yannis Avrithis. Alignmixup: Improving representations by interpolating aligned features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19174-19183, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[47] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitlagkas, David Lopez-Paz, and Yoshua Bengio. Manifold mixup: Better representations by interpolating hidden states. In ICML, pages 6438-6447, 2019. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[48] Devesh Walawalkar, Zhiqiang Shen, Zechun Liu, and Marios Savvides. Attentive cutmix: An enhanced data augmentation approach for deep learning based image classification. arXiv preprint arXiv:2003.13048, 2020. 1, 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 306 + ], + "type": "text", + "content": "[49] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "type": "text", + "content": "[50] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 343, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 374 + ], + "type": "text", + "content": "[51] Lingfeng Yang, Xiang Li, Borui Zhao, Renjie Song, and Jian Yang. Recursivelymix: Mixed learning with history. arXiv preprint arXiv:2203.06844, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 419 + ], + "type": "text", + "content": "[52] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, pages 6023-6032, 2019. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "text", + "content": "[53] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 455, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 509 + ], + "type": "text", + "content": "[54] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, pages 6881-6890, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 511, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 553 + ], + "type": "text", + "content": "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. IJCV, 127:302-321, 2019. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 555, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 587 + ], + "type": "text", + "content": "[56] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 1, 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5504" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Too Large; Data Reduction for Vision-Language Pre-Training/bf4cd6c9-c260-42c2-88a5-f694382bd718_content_list.json b/2023/Too Large; Data Reduction for Vision-Language Pre-Training/bf4cd6c9-c260-42c2-88a5-f694382bd718_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d58a3c28838904a1bd4e98a1c093ef6283a13a9a --- /dev/null +++ b/2023/Too Large; Data Reduction for Vision-Language Pre-Training/bf4cd6c9-c260-42c2-88a5-f694382bd718_content_list.json @@ -0,0 +1,1773 @@ +[ + { + "type": "text", + "text": "Too Large; Data Reduction for Vision-Language Pre-Training", + "text_level": 1, + "bbox": [ + 158, + 128, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alex Jinpeng Wang, Kevin Qinghong Lin, David Junhao Zhang, Stan Weixian Lei, Mike Zheng Shou* Show Lab, National University of Singapore", + "bbox": [ + 230, + 180, + 738, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 272, + 313, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper examines the problems of severe image-text misalignment and high redundancy in the widely-used large-scale Vision-Language Pre-Training (VLP) datasets. To address these issues, we propose an efficient and straightforward Vision-Language learning algorithm called TL;DR, which aims to compress the existing large VLP data into a small, high-quality set. Our approach consists of two major steps. First, a codebook-based encoder-decoder captioner is developed to select representative samples. Second, a new caption is generated to complement the original captions for selected samples, mitigating the text-image misalignment problem while maintaining uniqueness. As the result, TL;DR enables us to reduce the large dataset into a small set of high-quality data, which can serve as an alternative pre-training dataset. This algorithm significantly speeds up the time-consuming pretraining process. Specifically, TL;DR can compress the mainstream VLP datasets at a high ratio, e.g., reduce well-cleaned CC3M dataset from 2.82M to 0.67M ( $\\sim 24\\%$ ) and noisy YFCC15M from 15M to 2.5M ( $\\sim 16.7\\%$ ). Extensive experiments with three popular VLP models over seven downstream tasks show that VLP model trained on the compressed dataset provided by TL;DR can perform similar or even better results compared with training on the full-scale dataset1.", + "bbox": [ + 76, + 305, + 473, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 699, + 209, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent \"scale-is-everything\" viewpoint has become a widely accepted notion in the Vision-language Pre-training (VLP) community [40, 7, 34, 17, 1]. According to this view, the scale of the data has increased from the original tens of thousands-level (e.g., COCO [26] and VG [20]) to millions-level (e.g., CC3M [40] and CC12M [7]), and even up to billions-level (e.g., YFCC100M [43], WIT400M [34], and LAION400M [39]). Approaches [57, 34, 17] trained on these large-scale data show remarkable performance im", + "bbox": [ + 75, + 724, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c884edc4e9f990ead5d34f18305048dcfaea13cd7fa2361ca38e50714552c154.jpg", + "image_caption": [ + "Figure 1: Does using more data really lead to better performance in VLP? Instead of training on the full-scale CC3M dataset, we delete data with low image-text matching score. We find that BLIP [22] model pretrained on $50\\%$ reserved data even obtains better result than full-scale dataset on downstream COCO retrieval [26]. This observation exposes there exists serious misalignment between text&visual modalities and data redundancy in dataset." + ], + "image_footnote": [], + "bbox": [ + 506, + 271, + 689, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ed4e85d8e0b95abebcd24f052ef1588ad1880b9ba5e8ace5e93115732e73a3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 271, + 879, + 392 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "provement in various downstream tasks.", + "bbox": [ + 500, + 534, + 766, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, simply scaling-up data brings two critical challenges: $i$ . Larger image-text datasets lead to more training cost (e.g., Pretraining CoCa takes about 5 days on 2,048 CloudTPUv4 chips [57]) and storage overhead, which is difficult to afford. $ii$ . Obtaining high-quality VLP data requires massive data and well-designed collecting/filtering pipeline, which is expensive. For instance, the CC3M [40] data was obtained after filtering 5 billion collected images. These challenges are daunting and may impede the participation of numerous researchers in the VLP community.", + "bbox": [ + 496, + 551, + 890, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this study, we stop hunting for larger-scale data blindly and ask an important question: Does employing a larger dataset always result in better performance in VLP? To explore and answer this question, we begin with a simple experiment. First, we utilize a pre-trained BLIP [22] model to calculate the Image-Text Matching (ITM) scores for all samples in the clean CC3M dataset. Subsequently, we remove a portion of the samples with the lowest ITM scores and evaluate the transfer learning results, as shown in Figure 1. Surprisingly, discarding $50\\%$ of the samples slightly improves performance. This remarkable finding challenges the prevailing belief that employing larger amounts of data invariably leads to superior VLP outcomes.", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 94, + 875, + 225, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/showlab/datacentric.vlp", + "bbox": [ + 94, + 887, + 433, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3147", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/b48def29d89400a3a7fedf6117730a6b64628a6a3b557b4368d586fc50df1286.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodYearData TypeCompression Ratio↑Task AgnosticLarge-scaleSupervisionGeneration/Selection
Dataset Distillation [51]2017Image99%-99.99%XXClass LabelGeneration
Data Pruning [41]2022Image20%-30%X✓Class LabelSelection
Neural Data Server [53]2020Multi-modality94%-98%X✓Image-text PairsSelection
TL;DR (ours)-Multi-modality75%-90%✓✓Image-text PairsGeneration+Selection
", + "bbox": [ + 76, + 87, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1: Data-efficient learning methods. \"Large-scale\" means that the methods are effective when used on datasets that are very large in size. The \"task agnostic\" means that the methods can be used regardless of the specific downstream task, and without any prior exposure to the associated data.", + "bbox": [ + 75, + 154, + 892, + 199 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This experiment suggests removing certain data points can actually improve the model's ability to learn and generalize. Moreover, considering the performance improvements after removing the low ITM score data, we can infer the existence of significant misalignment between the textual and visual modalities in many text-image data pairs (see Figure 7 and the supplementary material for more evidences). These discoveries present promising potentiality to enhance the performance of models that depend on a smaller volume of VLP data.", + "bbox": [ + 75, + 215, + 468, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Driven by above analysis and recent advance in dataset pruning [41], we present a simple, effective and scalable algorithm called $TL;DR$ that aims to improve data efficiency for visual-language pretraining. The $TL;DR$ has a powerful codebook-based captioner, which contains a visual encoder, a look-up codebook and a text decoder. Here is how it works: First, $TL;DR$ feeds each image into the visual encoder and determines the corresponding codes of the image by measuring the similarity between the codebook and the embedding generated by the encoder. Given a large pool of image-text pairs, $TL;DR$ clusters the samples based on their image corresponding codes and selects a representative subset of samples from each cluster. Then, $TL;DR$ further refines the caption of the selected samples via text decoder to reduce text-image misalignment. By doing so, $TL;DR$ is able to significantly reduce the size of the training dataset while maintaining the high quality.", + "bbox": [ + 75, + 367, + 470, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we employ $TL; DR$ on widely-used CC3M, CC12M, YFCC100M and LAION400M datasets and evaluate small size data on three widely-used frameworks including CLIP [34], ViLT [19], and BLIP [22] for data efficiency pretraining with seven representative visual-language downstream tasks. The results show that, with only $10\\% - 25\\%$ data obtained by $TL; DR$ , frameworks achieve similar or even better performance compared with the full-scale dataset. We hope our findings can inspire the community to reconsider data efficiency for VLP rather than blindly utilizing increasingly massive datasets.", + "bbox": [ + 75, + 625, + 468, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 806, + 218, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Data-Efficient Learning", + "text_level": 1, + "bbox": [ + 76, + 830, + 295, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent successes in deep learning are largely attributed to the vast amount of data [10, 34]. However, collecting massive amounts of data is expensive and raises con", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cerns about privacy and copyright [59]. As a result, the research community has become increasingly interested in data-efficient learning, which includes:", + "bbox": [ + 496, + 215, + 890, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dataset Distillation [51, 61, 50, 33] compress a large dataset into a small set of synthetic samples, enabling models trained on the smaller dataset to achieve competitive performance with those trained on the original dataset. However, these techniques are only effective on relatively small datasets at low resolutions, such as CIFAR [21], and their performance deteriorates significantly when applied to larger-scale datasets. For example, the accuracy of a model trained on the state-of-the-art MMT's generated data is only $33.8\\%$ on the ImageNet-1K [10] test result [6], while pre-training on real ImageNet-1K achieves over $80\\%$ accuracy [9]. Furthermore, these methods necessitate supervised class labels, which are not suitable for multimodal data.", + "bbox": [ + 496, + 261, + 892, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data Pruning [44, 31] assumes high redundancy in large datasets, selecting only a subset of challenging samples. [29, 31] observed that during the entire training process, some examples are learned early and never forgotten, while others can be repeatedly learned and forgotten. The related work [41] uses a hard sample selection method to select $80\\%$ samples of the ImageNet dataset, and the model trained on selected samples approximating training on all data. Another recent work, CiT [52], also proposes to train models with dynamic training data.", + "bbox": [ + 496, + 458, + 890, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Data Server (NDS) [53, 27, 5] proposes a large-scale search engine to identify the most useful transfer learning data from large corpus. While these methods can be extended to multi-modality data, a similar idea has also been applied in NLP [54]. However, this setting assumes that the user has access to all downstream data and needs to train the downstream task using additional retrieval data.", + "bbox": [ + 496, + 609, + 890, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we are different from previous techniques in that we attempt to compress large-scale multimodal data for the first time, leading to comparable performance between the compressed and original vision-language datasets. We provide a comparison of our approach with these related works in Table 1.", + "bbox": [ + 496, + 715, + 890, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Visual-Language Pre-training", + "text_level": 1, + "bbox": [ + 500, + 816, + 764, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large-scale Vision-Language Pre-training (VLP) involves training on extensive multi-modality data and evaluating performance on various downstream vision-language tasks. Conventional frameworks include the dual-stream", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3148", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3677a06d50e4d08d552894d71c68f2e5555385dbf6d36aa502b373e52afaaa91.jpg", + "image_caption": [ + "Figure 2: Our TL;DR architecture. We first train a codebook-based captioner in Stage1. Then the learned codebook and captioner are used to reduce VLP data in Stage 2. Pre-training on the reduced dataset achieves similar performance to the original full-scale dataset across downstream tasks." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 413, + 251 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8417fe376969c3596a6b7a530f6f0746a205a561eb10606be6bfa6b79ade3aec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 88, + 857, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "architecture [34], the one-stream architecture [19, 24, 48, 25, 49], and the encoder-decoder architecture [22]. Previous works have relied on high-quality, human-annotated datasets such as COCO [26] (110K images) and Visual Genome [20] (100K). As model sizes continue to increase, pre-training requires even more data than before [26, 17, 49], resulting in an extremely high computational cost. However, obtaining large and high-quality multi-modality data is challenging due to the difficulties in annotation. In this paper, we aim to democratize VLP research by proposing a general compression method for existing VLP data.", + "bbox": [ + 75, + 316, + 472, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 502, + 169, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our $TL;DR$ is a simple yet effective approach for compressing the Vision-Language Pre-training dataset, leading to further reduction of the training cost. Our approach consists of two stages: (1) codebook-based captioner training and (2) data reduction including samples selection and caption refining. Figure 2 illustrates the idea, introduced next.", + "bbox": [ + 75, + 529, + 468, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Codebook-based Captioner", + "text_level": 1, + "bbox": [ + 76, + 633, + 323, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The captioner consists of a visual encoder, a codebook and a text decoder. The visual encoder is employed to extract image features. Inspired by vector quantisation technique [56, 45], we try to quantize the image feature for further clustering by utilizing a learnable codebook. Codebook comprises $K$ learnable embedding vectors, each of which can be regarded as a code. Each token of image features conducts a nearest neighbor look-up from codebook and finds its corresponding code. In this way, image features are quantized into a couple of codes (quantized vectors). The quantized vectors are then sent into a text decoder, generating a caption. In order to enhance the quality of text generation, we initialize the codebook with the text embedding of $K$ most frequently occurring keywords/keyphrases in the entire dataset, which enables the codebook to contain meaningful and intuitively understandable semantics.", + "bbox": [ + 75, + 659, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train the whole captioner, we utilize a Language Modeling loss [11], which maximizes the likelihood of the text in an autoregressive manner, and a symmetric commitment loss [56], which is specifically designed for codebook. We initially train this captioner on noisy source data and subsequently fine-tune it on smaller-scale datasets, such as COCO [26] and VisualGenome [20].", + "bbox": [ + 496, + 316, + 893, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data Reduction", + "text_level": 1, + "bbox": [ + 500, + 435, + 656, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Currently, large-scale datasets exist with serious redundancy [41]. Meanwhile, a large part of texts is noisy and misaligned with images in VLP data. See Figure 2 for the example (the caption \"You need to think twice before buying a pet as present\" does not match the image). To overcome these limitations, we use the learned codebook to condense large-scale noisy data and the learned captioner to reduce the misalignment over image-text pairs.", + "bbox": [ + 496, + 459, + 890, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Samples selection. For an encoded image feature with $L$ tokens, we compute an index vector with length $L$ . Each value is the index of the code, which is the closest to each token. This vector maps the features from image space to semantic space so that it reduces the complexity of the image, benefiting and accelerating the cluster process. Subsequently, each image sample in the dataset is equipped with an index vector according to the above process and we cluster these vectors into $N$ clusters with K-Means (speed up by Faiss [18]). Then we uniformly sample $M\\%$ data points from each cluster, producing a small subset of the dataset. We examine various sampling methods and observe that uniform sampling is stable across different scales.", + "bbox": [ + 496, + 580, + 892, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Caption refining. To alleviate the misalignment problem, we want to improve the text quality using the generated caption. Generated text $T_{g}$ is from the text decoder, which takes the quantized vector of the image as input. We simply concatenate $T_{g}$ with original text $T_{o}$ together, denoted as $T = T_{o} + T_{g}$ , to refine and preserve the original caption's uniqueness while maintaining data diversity.", + "bbox": [ + 496, + 777, + 890, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The compressed small-scale dataset with refined cap", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3149", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/4fb3ce84ba4e03e8a155d97e9bd177c63c58b1c75b8b3771795359ab2e67f3a2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
samplingrefiningTR@1IR@1
65.349.8
✓68.551.9
✓69.452.3
✓✓72.854.8
", + "bbox": [ + 102, + 88, + 315, + 155 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/f8625fbc4986755bc0bd74aa0da93931d97548ac490647c6807e93ee9f30a99d.jpg", + "table_caption": [ + "(a) Component ablation. Both the sampling and refining operation are important to the downstream retrieval." + ], + "table_footnote": [], + "table_body": "
caseTR@1IR@1
Image Embedding70.652.3
Text Embedding69.050.4
BLIP Image Embedding [22]72.354.5
Codebook72.854.8
", + "bbox": [ + 86, + 208, + 354, + 276 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/fbf9328d46e7e1c444146f452c4d1afaebbd8166e514830aacb319fdcfe8bf3c.jpg", + "table_caption": [ + "(d) Clustering feature. Codebook is better than Image Embedding at same scale." + ], + "table_footnote": [], + "table_body": "
caseTR@1IR@1
gradient-based72.954.8
hard-sample [41]73.154.5
uniform72.854.8
large distance72.353.1
", + "bbox": [ + 370, + 88, + 604, + 154 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/9bbe61beaf812ac9e1d035793872c8a55862c332f88991fe52ca77e890e2cad1.jpg", + "table_caption": [ + "(b) Sample-selection strategy. The different way to select samples." + ], + "table_footnote": [], + "table_body": "
caseTR@1IR@1
full-scale baseline70.654.0
10%68.952.3
25%72.854.8
50%74.855.2
100%75.157.7
", + "bbox": [ + 370, + 208, + 604, + 285 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/61fde20a00c9ad89cc7f12a589923af002de07caad06e33baf332b7cc2818e7a.jpg", + "table_caption": [ + "(e) Sampling ratio. Sampling $25\\%$ data is enough to beats with full scale." + ], + "table_footnote": [], + "table_body": "
caseTR@1IR@1
xavier [14]72.054.1
key words/phrases72.854.8
object tags72.554.4
", + "bbox": [ + 666, + 88, + 867, + 142 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/f0683456a15a9c80c4b849547e69e42ad9c21cba3663ea984dc27ba42ef92110.jpg", + "table_caption": [ + "(c) Codebook Initialization. An codebook initialized with keywords is more stable." + ], + "table_footnote": [], + "table_body": "
caseTR@1IR@1
10071.853.9
100072.454.5
300072.854.8
500072.954.4
1000072.354.2
", + "bbox": [ + 666, + 208, + 867, + 285 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(f) Cluster Number. More clusters not equals better result.", + "bbox": [ + 638, + 292, + 887, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2: TL;DR ablation experiments with BLIP model [22] on CC3M. We report image-to-text retrieval top-1 (TR@1) and text-to-image retrieval top-1 (IR@1) accuracy (\\%) on COCO [26] dataset. If not specified, the default baseline is: pretraining BLIP model based on ViT-B/16 with $25\\%$ sample of CC3M. Default settings are marked in gray.", + "bbox": [ + 75, + 327, + 892, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tions is recorded as dataset $D_{c}$ . At last, we train VLP models on this high quality dataset $D_{c}$ and expect the model to achieve comparable performance with original full-scale dataset $D$ on downstream Vision-Language tasks.", + "bbox": [ + 75, + 388, + 468, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion. Considering the serious misalignment problem, it seems quite straightforward to use pure generated high-quality caption $T_{g}$ to replace original noisy text. Driven by this idea, we try to pretrain BLIP [22] models with $T_{o}, T_{g}$ and $T_{o} + T_{g}$ independently and show the train curve of Image-Text Contrastive (ITC) loss in Figure 3. However, we find the model trained with $T_{g}$ fails into model collapse [36]. This phenomenon can be explained by captioning collapse [46, 47] and one-to-many problem [55] in image cponing. That is, the trained captioner will generate fixed or similar captions for different images, which limits diversity in the output and easily leads to trivial solutions for contrastive loss. On the contrary, the ITC loss for both $T_{o}$ and $T_{o} + T_{g}$ works well and the $T_{o} + T_{g}$ converges better. We also observe the loss of $T_{g}$ is smaller than other two variants at epoch 0-2, which indicates the generated caption matches well with the image. Note that this simple stitching operation on caption does not bring additional computation cost for VLP as the max length in text-encoder Bert [11] model keeps unchanged for all setting.", + "bbox": [ + 75, + 450, + 468, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Technical Details.", + "text_level": 1, + "bbox": [ + 76, + 762, + 248, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our $TL;DR$ can be implemented efficiently, and importantly, does not require any large auxiliary model. The codebook size $K$ is 3000 as default. The selection of keywords/phrases is implemented using the NLTK $^2$ . We adopt ViT-B/16 [12] as image encoder and BertLMHead Model [11] as text decoder. In this way, the token length $L$", + "bbox": [ + 75, + 785, + 468, + 878 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3b2d50dd45fe00565bc4320af1f574e73c1c93d71130509bf68eafe1613810b2.jpg", + "image_caption": [ + "Figure 3: Training curve with CC3M dataset. Simply stitching generated text and original text together solved the model collapse problem in Image-text Contrastive Loss." + ], + "image_footnote": [], + "bbox": [ + 501, + 386, + 633, + 486 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0e099d21e14963a74f9f3ada66cba5e119877e8b25f64fe54e2867eacd9202dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 386, + 759, + 484 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1f228c7b4840f080049a7b4b322e4b0d41c83677aef8ce3bef64ca0fbe36502d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 386, + 888, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is 196 as default. The cross-attention is computed over image embedding and text embedding. To show the generality of compressed dataset, we test $D_{c}$ on three different and representative VLP architectures: dual-stream CLIP [34], one-stream ViLT [19] and Fusion-encoder Blip [22] on various downstream tasks. All these models are trained under the same setting with different datasets.", + "bbox": [ + 496, + 551, + 892, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. CC3M Experiments", + "text_level": 1, + "bbox": [ + 500, + 672, + 692, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first study dataset reduction on well-cleaned $CC3M$ [40] which heavily filters web crawled pairs and only keeps $0.1\\%$ of the raw data. This dataset contains a total of 2.8 million images. We employ our TL;DR to compress the CC3M dataset, then conduct pre-training and fine-tuning evaluations on both original and compressed datasets. Following our ablation study, we transfer the pre-trained model to seven Vision-Language tasks downstream and fine-tune it through end-to-end training to evaluate its performance.", + "bbox": [ + 496, + 699, + 892, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training. We utilize PyTorch [30] to implement our models and trained them on 8 NVIDIA A100 GPUs to reduce the data samples. For Vision-Language Pre-training, we utilize 2 nodes, each equipped with 16 GPUs. The im", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2https://github.com/pltk/pltk", + "bbox": [ + 94, + 886, + 248, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3150", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/7227784f8256e7f7f88367796d0442d7680121e5b97c735acb8be72c1dbce26e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDataset#SamplesMSCOCO (5K test set)Flickr30K (1K test set)
Image→TextText→ImageImage→TextText→Image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP [34]CC3M [40]2.82M60.485.393.248.975.484.777.391.193.271.690.191.4
TL;DR-CC3M0.67M60.385.693.849.477.486.082.591.892.272.090.592.1
CC3M [40]2.82M36.264.380.129.957.966.967.483.292.454.384.190.8
TL;DR-CC3M0.67M37.764.680.830.758.468.268.585.492.055.682.190.8
ViLT [19]CC3M [40]2.82M66.789.293.852.579.387.183.892.093.274.092.092.8
TL;DR-CC3M0.67M67.188.794.153.178.988.285.392.493.675.692.192.5
CC3M [40]2.82M39.268.677.830.453.266.170.588.792.157.684.992.6
TL;DR-CC3M0.67M43.570.881.433.957.966.873.290.593.358.684.792.4
BLIP [22]CC3M [40]2.82M70.991.396.154.380.288.086.394.194.874.891.692.6
TL;DR-CC3M0.67M72.891.995.954.880.689.487.594.895.375.792.293.4
CC3M [40]2.82M42.367.877.431.555.766.375.191.293.660.685.991.8
TL;DR-CC3M0.67M48.773.182.736.760.670.476.391.993.961.087.793.0
", + "bbox": [ + 78, + 85, + 897, + 272 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/42e4741624d41ad0d9206291dcabbf8dcf6019750bc21b925331ec12013a4ff6.jpg", + "table_caption": [ + "Table 3: Fine-tuning and zero-shot image-text retrieval results on MSCOCO and Flickr30K dataset." + ], + "table_footnote": [], + "table_body": "
Dataset#SamplesVQANLVR2valRefCOCO+COCO Caption
test-devtest-standarddevtest-PtestAtestBB@4CIDEr
Random-CC3M0.67M68.366.273.673.868.671.862.835.9118.8
CC3M [40]2.8M71.571.876.076.272.476.165.336.8121.6
TL;DR-CC3M0.67M73.1+1.673.2+1.477.7+1.778.0+1.875.1+2.778.5+2.468.4+3.137.6+0.8123.8+2.2
", + "bbox": [ + 81, + 299, + 883, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "age transformer is initialized from ViT pre-trained on ImageNet [10], and the text transformer is initialized from BERTbase [11](BLIP,CLIP) and DistillBERT [38] (ViLT). The model is pre-trained for 20 epochs with a batch size of 1260 and an AdamW [28] optimizer with a weight decay of 0.05. During training, we apply a learning rate warm-up to 3e-4 and a linear decay with a rate of 0.85. For image augmentation, we utilize RandAugment [8] and apply all of the original policies except color inversion. This decision is based on the recognition of the crucial role that color information plays in the data. For pre-training, images were randomly cropped to a resolution of $224 \\times 224$ . We then increase this to $384 \\times 384$ for fine-tuning downstream tasks. Further information about the training hyperparameters for downstream tasks can be found in the supplementary material.", + "bbox": [ + 75, + 421, + 472, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Main Properties", + "text_level": 1, + "bbox": [ + 76, + 671, + 238, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We ablate our $TL;DR$ using the default setting in Table 2 (see caption). Several intriguing properties are observed.", + "bbox": [ + 75, + 695, + 468, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Module deconstruction. In Table 2a we analyze the impact of different components in $TL;DR$ . We establish a baseline by randomly selecting $25\\%$ of the data from CC3M (first row). Our results show that codebook-based sampling outperforms random selection by $3.2\\%$ in TR@1. We also observe that both codebook-based sampling and caption refinement are crucial and the combination of them achieves optimal downstream performance.", + "bbox": [ + 73, + 729, + 468, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sample selection. In Table 2b we study the sample selection strategy in Stage 2. We sample $25\\%$ data in each cluster by default. For Gradient-based, we train a tiny network", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4f5d60ea3aa24d5028783a0d8632d6b6cd046def480ae52d39e2bfa694acb3fc.jpg", + "table_caption": [ + "Table 4: Comparison with BLIP model pre-trained on different data sources for VQA, NLVR², RefCOCO+ and COCO Captioning. ViLT and CLIP architectures can not be evaluated on part of these tasks since structural limitations." + ], + "table_footnote": [], + "table_body": "
MethodDatasetR@1↑R@5↑R@10↑MdR↓
CLIP [34]Rand-CC3M15.334.846.313.0
CC3M [40]19.437.347.511.0
TL;DR-CC3M21.838.648.510.0
ViLT [19]Rand-CC3M18.838.249.511.0
CC3M [40]21.040.551.510.0
TL;DR-CC3M22.542.752.48.0
BLIP [22]Rand-CC3M23.342.853.38.0
CC3M [40]26.046.358.07.0
TL;DR-CC3M27.448.759.46.0
", + "bbox": [ + 500, + 417, + 888, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 5: MSRVTT-1K retrieval using three architectures. We created a subset of the CC3M dataset called Rand-CC3M by randomly selecting the same number of samples as in TL;DR-CC3M.", + "bbox": [ + 498, + 544, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to conduct VLP pretrained with ITC [24], ITM [24] and LM [11]. Then we select samples which contribute most to the gradients in each cluster. Large distance: Another perspective is that data points on the border of each cluster are more important than those at the center [4]. So we first compute the center of each cluster and then choose the sample that has the largest distance from the center of each cluster. We also report the result of hard-sample selection from [41]. We observe that all these variants produce similar results except large distances. This suggests that the clustering step, rather than the selection step, plays a key role in data compression during Stage 2. To maintain simplicity, we choose uniform sampling as the default method.", + "bbox": [ + 496, + 619, + 890, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Codebook initialization. In Table 2c we compare different initialization strategies. The xavier means all parameters in the codebook are initialized with xavier initialization [14]. For the object tags initialization, following previous works [2, 60], we use the 1600 object tags from", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3151", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/47d2443050d4ff275131faf8caa869362609839b84ec46870ca6fe76eb434e2e.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6: Zero-shot image classification results on ImageNet [10], ImageNet-A [16], ImageNet-R [15]. There is no free lunch, as selecting partial samples reduces the visual diversity crucial for classification. Despite this, TL;DR still performs significantly better than random selection." + ], + "table_body": "
ModelDataset#SamplesImNetImNet-AImNet-R
CLIP [34]Rand-CC3M0.67M58.361.862.3
CC3M [40]2.82M62.265.266.9
TL;DR-CC3M0.67M61.465.065.7
ViLT [19]Rand-CC3M0.67M54.359.858.4
CC3M [40]2.82M58.662.964.2
TL;DR-CC3M0.67M59.163.364.0
BLIP [22]Rand-CC3M0.67M57.361.865.2
CC3M [40]2.82M62.565.568.1
TL;DR-CC3M0.67M62.063.967.4
", + "bbox": [ + 78, + 88, + 468, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Genome [20] and extract text feature with a pretrained BERT [11]. With same training setting, the keywords achieve a $0.8\\%$ TR@1 improvement and a $0.7\\%$ IR@1 improvement over xavier. This result is expected as the text embeddings provide contextual information and simplify the learning process.", + "bbox": [ + 75, + 315, + 470, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Codebook vs. Image embedding. In Table 2d, we investigate different ways of cluster sampling. First, we remove the codebook from Stage-1 and use image embedding instead. Alternatively, we directly cluster images using the image embedding [22] of images from BLIP model (pretrained on 200M Image-text pairs). We observe the image embedding leads to much better result than text embedding. This is reasonable because clustering visual-similarity samples with text only is difficult. We observe that clustering depended on our codebook performs better than both image embedding and text embedding. This demonstrates that our codebook can efficiently project image embedding to semantic space, benefiting cluster process.", + "bbox": [ + 75, + 414, + 468, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cluster sampling ratio. Table 2e varies the sampling ratio of each cluster from $10\\%$ to $100\\%$ . We are surprised to find that a low sampling ratio can still produce effective results. With only $25\\%$ of the data and the TL;DR model, we are able to achieve a $1.9\\%$ improvement on TR@1 and a $0.8\\%$ improvement on IR@1 over the full-scale baseline. Additionally, we observe that larger sampling ratios lead to even better results. Since our focus is on achieving similar transfer learning results with fewer samples, we use a default sampling ratio of $25\\%$ to minimize computation costs.", + "bbox": [ + 73, + 619, + 468, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cluster numbers. In Table 2f, we investigate the impact of cluster number on Stage 2 by increasing it from 300 to $30\\mathrm{K}$ . We observe that using more clusters results in a slight improvement at the beginning and becomes stable when the number of clusters exceeds $3\\mathrm{K}$ . Moreover, all results consistently outperform the random selection baseline. Therefore, we use $3\\mathrm{K}$ clusters as the default in this work, as it performs well on fine-tuning tasks.", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb4d08a74d726e2c19aab1d2bb512d5f5fd9f94a890950993bd85c7176195275.jpg", + "image_caption": [], + "image_footnote": [ + "Captioner: A racer jumps over a railing on a horse." + ], + "bbox": [ + 504, + 90, + 581, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d61092376a66fc5dcb355e7359d596571283f1342776441317f33c10fadf4f50.jpg", + "image_caption": [ + "Figure 4: The generated caption match the image well." + ], + "image_footnote": [], + "bbox": [ + 629, + 90, + 743, + 174 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/29efb37dd3e58a3e611286fdf1fa5cb2bc7d2d1b560069d6be6c54e10b5fb60e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 758, + 90, + 867, + 174 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Transfer Learning Experiments.", + "text_level": 1, + "bbox": [ + 500, + 209, + 784, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct an extensive evaluation of transfer learning in downstream tasks using the model pre-trained on our compressed TL;DR-CC3M and source CC3M with 3 architectures. Our evaluation primarily focuses on the core tasks of three categories that examine: (1) cross-modality alignment, (2) image captioning and multi-modality understanding capabilities, and (3) visual recognition. The baseline in this section is the model trained on CC3M dataset.", + "bbox": [ + 496, + 233, + 893, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Cross-modality Alignment Task", + "text_level": 1, + "bbox": [ + 498, + 377, + 774, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image-Text retrieval. Fine-grained world region alignment plays a critical role in this task. We report both image-to-text retrieval (TR) and text-to-image retrieval (IR) on the COCO [26] and Flickr30K [32] benchmarks. For the BLIP [22] model, we adopt an additional re-ranking strategy, following the original implementation. In Table 3, we also report zero-shot retrieval results. We found that $TL; DR$ achieves comparable results with the baselines on all metrics and surprisingly performs quite well on zero-shot results. For example, for the BLIP [22] architecture, our method leads to a $6.4\\%$ improvement (from $42.3\\%$ to $48.7\\%$ ) in Recall@1 of image-to-text retrieval on MSCOCO. All results suggest that a small part of refined image-text pairs is enough to learn good alignment.", + "bbox": [ + 496, + 406, + 892, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot video retrieval. In this experiment, we analyze the generalization ability of our method to video-language tasks. Specifically, we perform zero-shot transfer to text-to-video retrieval and evaluate the models trained on COCO-retrieval in Table 5. We uniformly sample 8 frames from each video to process the video input and concatenate the frame features into a single sequence. These models trained on our compressed dataset outperform the baseline on all metrics, demonstrating the generality of $TL; DR$ .", + "bbox": [ + 496, + 623, + 893, + 758 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a636e3c42ab7ea13a7115b8c9c644393d168fab2806dca47ae6f579d805f5f92.jpg", + "image_caption": [ + "Figure 5: The codebook-based clusters visualization. The samples within each cluster exhibit similar contextual characteristics, as opposed to mere visual appearance. For example, the \"Christmas elements\" cluster located at the right." + ], + "image_footnote": [], + "bbox": [ + 501, + 773, + 691, + 825 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bdc5d88023c00c6a8dc1bcc8abd2de0ca3060f21f09dc14156d57c309caf8473.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 773, + 826, + 825 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/441946df53e490025b215e92156544169bc701c92f569e6cb98ca673ddb9b94f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 773, + 890, + 825 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3152", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ffdb52a15d2ef9e150e33b0e9ae0a75d2631b936484146dd24102825cd515e8a.jpg", + "image_caption": [ + "VQ-GAN: 5.13s" + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 207, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97cbdbff14d12199f45ee15a7bbef5f1ad326c7625f9bc4aa6995d35b5c3e844.jpg", + "image_caption": [ + "DALL E2: 2.65s" + ], + "image_footnote": [], + "bbox": [ + 210, + 88, + 331, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57a1d2243c7355a484af46cdfc1cc5dbf61468cc608f54254b55c89d806d9aed.jpg", + "image_caption": [ + "Stable Diffusion: 0.17s", + "Figure 6: Image generation result with strong Text-to-image Model. The generation time is also reported." + ], + "image_footnote": [], + "bbox": [ + 334, + 88, + 455, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Image Captioning and Multi-modality Understanding Tasks", + "text_level": 1, + "bbox": [ + 76, + 239, + 467, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image captioning. The task involves describing an input image, which we evaluate using No-Caps and COCO datasets. Both datasets are fine-tuned on COCO with the Language Modeling (LM) loss. We adopt a zero-shot setting for No-Caps dataset, and start each caption with the phrase \"a picture of\" for the BLIP architecture. We do not pre-train using COCO to avoid information leakage. Our results outperform baseline with a much smaller quantity of pre-training data, as shown in Table 4.", + "bbox": [ + 75, + 282, + 467, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual question answering (VQA). We evaluate our model's performance on the VQA task [3], where the model needs to provide an answer based on an image and a question. We consider it as an answer generation task that allows open-vocabulary VQA for better results, following previous works [23, 22]. The results are presented in Table 4. The BLIP trained on TL;DR-CC3M outperforms baseline by $1.4\\%$ on test-dev splits, demonstrating the effectiveness of our compressed dataset for improving VQA performance.", + "bbox": [ + 75, + 422, + 467, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual reasoning. The Natural Language Visual Reasoning $(\\mathrm{NLVR}^2)$ [42] task is a binary classification task that requires the model to reason about two images and a question in natural language. Multi-modal reasoning is crucial for the completion of this task. We observe that BLIP trained on our dataset achieved $78.0\\%$ accuracy compared to $76.2\\%$ achieved by the CC3M, as shown in Table 4.", + "bbox": [ + 75, + 564, + 467, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-modality grounding. Referring Expression (RE) Comprehension requires the model to select the target object from a set of image regions proposals, based on the query description. This task heavily relies on visual-grounding ability. The models are evaluated on ground-truth objects, and we evaluate RE Comprehension on RefCOCO+ [58]. The results are reported in Table 4, and we observe that TL;DR-CC3M achieves better results.", + "bbox": [ + 75, + 674, + 467, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3 Visual Recognition Tasks", + "text_level": 1, + "bbox": [ + 76, + 816, + 305, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides the cross-modality task, we also explore a unimodality task, mainly image classification. Specifically, we fix the image encoder and explore zero-shot image classification. Since the vision encoder is loaded from", + "bbox": [ + 75, + 839, + 467, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a163e91b1eab68a9e6b3ef8e547a527c55c7b5af681a6817cbc6331947d470a0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTR@1IR@1
real data58.344.0
VQ-GAN [13]35.232.4
DALLE2 [35] (implement from 3)44.338.3
Stable Diffusion [37] (implement from 4)52.440.7
", + "bbox": [ + 529, + 88, + 856, + 152 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7: Compare different sample generation methods over 0.3M subset of CC3M. We first pre-train BLIP model on these generated data and then evaluation on COCO.", + "bbox": [ + 498, + 152, + 890, + 198 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pretrained model, this task demonstrates the impact of training a Vision-Language model with noisy image-text pairs. Specifically, it shows how such training affects the well-learned representation derived from a human-crafted dataset. We show the results in Table 6 and observe noisy data leads to significant Catastrophic Forgetting. For example, the CLIP model drops down to only $58.3\\%$ accuracy with noisy data training. We also observe our TL;DR shows steady improvement for all architectures over random selection. Unfortunately, the classification performance for TL;DR-CC3M is slightly worse than the full-scale CC3M for the CLIP and BLIP architectures. Both of these architectures have independent image encoders like ViT to extract image embeddings. This indicates that this task heavily relies on visual diversity, which is different from multimodal tasks, and our method reduces the visual diversity potentially. For the ViLT model, this architecture adopts a shared backbone for both visual and text, and we observe the slightly different results. We guess that multi-modality interaction in early-fusion affects the classification result.", + "bbox": [ + 496, + 215, + 890, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Visualization", + "text_level": 1, + "bbox": [ + 500, + 529, + 635, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generated caption visualization. We show the generated caption in Figure 4. It is evident that the original captions can be highly abstract and difficult to match their respective images, even for human observers sometimes. For instance, when the ITM score is as low as 0.04, matching the figure with its corresponding caption becomes arduous. Such challenging cases can potentially harm the cross-modality alignment. In contrast, we observe that the generated captions describe the image very well and sometimes offer helpful complementary information. For example, \"bus\" and \"castle\" in the middle example.", + "bbox": [ + 496, + 556, + 890, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Codebook-based cluster visualization. Figure 5 displays the codebook grouping result achieved with simple KMeans. Clusters are sets of data points with similar characteristics, often defined by their features or attributes. Interestingly, we observe that the model cluster samples \"accurate\", meaning that these samples have semantic similarity rather than simple appearance. For instance, the model classifies \"dollars\" and \"piggy bank\" together, even though they differ significantly in appearance.", + "bbox": [ + 496, + 728, + 890, + 864 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{3}$ https://github.com/LAION-AI/dalle2-laion", + "bbox": [ + 514, + 875, + 750, + 887 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{4}$ https://github.com/huggingface/diffusers", + "bbox": [ + 517, + 887, + 740, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3153", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/927d86d3ae134e157e05f31675b6da0916a54566ba68ee4281a80a27dfe0cce1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset#Samples TimeVQA test-devNLVR2 test-PRefCOCO valNocaps Captioning B@4CIDErFlcikr30K Retrieval TR@1IR@1Imagenet Acc
Rand-CC12M2.4M14h71.876.272.536.8121.082.973.361.2
CC12M [7]10.8M65h73.578.974.137.5122.984.775.365.3
TL;DR-CC12M2.4M14h74.1+0.678.5-0.474.0-0.138.1+0.6124.1+1.285.5+0.876.3+1.063.8-1.5
Rand-YFCC15M2.5M15h67.270.568.135.2116.378.870.565.4
YFCC15M [43]15M90h70.574.270.635.9118.481.572.467.8
TL;DR-YFCC15M2.5M15h70.3-0.275.3+1.172.6+2.037.2+1.3122.5+4.182.3+0.874.3+1.967.3-0.5
Rand-LAION40M(128)8M48h70.775.373.434.8113.280.472.568.5
LAION40M(128) [39]40M120h74.579.176.635.2117.483.274.971.3
TL;DR-LAION40M(128)8M48h76.3+1.880.5+1.477.4+0.836.8+1.6120.9+3.582.8-0.476.1+1.270.4-0.9
", + "bbox": [ + 78, + 87, + 893, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8: Comparison with different source of data on 6 downstream tasks. BLIP [22] is adopted as baseline and (128) means the image resolution is ${128} \\times {128}$ . We also list the pre-training time,which can be significantly reduced via ${TL};{DR}$ .", + "bbox": [ + 75, + 226, + 893, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9bda6184e9448fb0fd18b91f69cf5f7fb0592e7edbde4e83215fd5cdbe75508d.jpg", + "image_caption": [ + "Figure 7: ITM score distribution. $TL; DR$ alleviates the issue of misalignment in VLP data." + ], + "image_footnote": [], + "bbox": [ + 98, + 270, + 274, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b57301f7f7960dbca96e636febef744721af53ddbc7346350b7bc1bcbc87430.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 271, + 446, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. More Investigation", + "text_level": 1, + "bbox": [ + 76, + 439, + 259, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Is image generation possible? To ease the misalignment problem of image-text pairs, instead of simply selecting representative samples, a potential and naive idea is to generate images from text. To this end, we randomly sample 0.3M subset of CC3M and generate image from text with three popular text to image generation models, VQ-GAN [56], DALLE 2 [35] and Stable Diffusion [37]. We display the generated samples in Figure 6. We observe that the generative models struggle with complex scenarios, but are capable of generating simple prompts like \"dog\" proficiently. In addition, generation methods only produce visual cues in a fixed vocabulary, potentially reducing data diversity.", + "bbox": [ + 75, + 468, + 468, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, we pre-train BLIP models on these generated data and evaluate it on COCO Retrieval. In Table 7 we observe the results of transfer learning depend on the quality of generated samples, with those generated by stable diffusion being particularly effective. However, there still exists a significant gap between the generated data and the real dataset (e.g., $52.4\\%$ vs. $58.3\\%$ on TR@1). We believe that higher-quality and diverse generated images may lead to comparable results with real images in the near future.", + "bbox": [ + 75, + 652, + 468, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Explore the misalignment problem. Figure 7 shows the Image-text Matching (ITM) score distribution for both $CC3M$ and our $TL;DR-CC3M$ data (the visualization about more datasets is reported in the supplementary). We observe a lot of samples of original $CC3M$ at low matching score even tends to zero, which indicates the current dataset has serious misalignment problems. Since image", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "text matching (ITM) loss and image-text contrastive (ITC) loss are used in all architectures, these samples will damage the multimodal representation learning. When adopting our $TL;DR$ , we observe that the matching score tends to be higher and has very few samples with low ITM score.", + "bbox": [ + 496, + 273, + 893, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Transfer to other VLP datasets", + "text_level": 1, + "bbox": [ + 498, + 369, + 784, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We study data compression performed in two categories shown below: clean data that involves human-based offline filter pipelines and raw data that has not undergone cleaning. For clean data, in addition to $CC3M$ , we explore the well-cleaned, high-quality dataset $CC12M$ [7]. Then, we study the raw data YFCC100M [43] and LAION400M [39]. $CC12M$ [7] contains 12 million image-text pairs specifically meant to be used for vision-and-language pre-training. These data are collected by relaxing the data collection pipeline as in CC3M. YFCC15M [34] is a subset of the multilingual and noisy YFCC100M [43] that contains English captions. LAION400M [39] is a large-scale noisy dataset that provides URLs with captions for download. To control the computation cost and reduce the storage overhead, we randomly sample a 40M subset of LAION400M and download images at a resolution of $128 \\times 128$ . So, we record the dataset as $TL;DR-LAION40M(128)$ , and the performance over downstream tasks could be improved with higher resolution. More exploration about video-text datasets is reported in the supplementary material", + "bbox": [ + 496, + 398, + 890, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We use BLIP as the default architecture and evaluate our $TL;DR$ on different datasets and show the results in Table 8. Surprisingly, with only $2.5\\mathrm{M}$ $(16.7\\%)$ data, $TL;DR$ -YFCC15M leads to similar results with $15\\mathrm{M}$ raw data over all metrics except Imagenet. More results with different backbones are reported in the supplementary material. For LAION40M(128), when using $8\\mathrm{M}$ data $(20\\%)$ , the model trained on our dataset consistently outperforms the baseline method on six downstream tasks. We noticed that the compression rate of LAION40M(128) is less than that of YFCC15M. This may be due to the fact that the collection of LAION40M(128) has already been filtered with CLIP similarity, reducing the impact of the misalignment problem.", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3154", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 76, + 89, + 326, + 106 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper presents $TL;DR$ , a novel and pioneering algorithm for selecting and generating high-quality image-text pairs from noisy Vision-Language Pre-training (VLP) data, thereby contributing to the field of VLP. $TL;DR$ incorporates a text generation process into learning to reduce serious misalignment problems. Our experiments demonstrate three widely-used architectures leads to comparable results and much smaller training cost when learning from our generated dataset. Additionally, we demonstrate that the misalignment problem can be effectively addressed using our simple $TL;DR$ . However, the choice of the highest compression ratio is done manually rather than learned. Furthermore, achieving even higher compression ratios for VLP models remains a challenge, and text-to-image generation models may be helpful in this regard. We hope that this perspective will inspire future research.", + "bbox": [ + 75, + 114, + 472, + 357 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 76, + 373, + 236, + 390 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008, and Meta, Mike Zheng Shou's Start-Up Grant from NUS. The computational work for this article was partially performed on resources of the National Supercomputing Centre, Singapore.", + "bbox": [ + 75, + 398, + 468, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 76, + 505, + 174, + 520 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022.", + "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018.", + "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015.", + "[4] Dibya Jyoti Bora, Dr Gupta, and Anil Kumar. A comparative study between fuzzy clustering algorithm and hard clustering algorithm. arXiv preprint arXiv:1404.6059, 2014.", + "[5] Tianshi Cao, Sasha Alexandre Doubov, David Acuna, and Sanja Fidler. Scalable neural data server: A data recommender for transfer learning. Advances in Neural Information Processing Systems, 34:8984-8997, 2021.", + "[6] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Proceedings of the" + ], + "bbox": [ + 84, + 530, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4750-4759, 2022.", + "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021.", + "[8] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020.", + "[9] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation to imagenet-1k with constant memory. arXiv preprint arXiv:2211.10586, 2022.", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009.", + "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021.", + "[14] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249-256. JMLR Workshop and Conference Proceedings, 2010.", + "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021.", + "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021.", + "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021.", + "[18] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535-547, 2019." + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3155", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021.", + "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017.", + "[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009.", + "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022.", + "[23] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021.", + "[24] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020.", + "[25] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022.", + "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014.", + "[27] Yiqi Lin, Huabin Zheng, Huaping Zhong, Jinjing Zhu, Wei-jia Li, Conghui He, and Lin Wang. Sept: Towards scalable and efficient visual pre-training. AAAI, 2023.", + "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017.", + "[29] Rafid Mahmood, James Lucas, David Acuna, Daiqing Li, Jonah Philion, Jose M Alvarez, Zhiding Yu, Sanja Fidler, and Marc T Law. How much more data do i need? estimating requirements for downstream tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 275-284, 2022.", + "[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019.", + "[31] Mansheej Paul, Surya Ganguli, and Gintare Karolina Dziugaite. Deep learning on a data diet: Finding important examples early in training. Advances in Neural Information Processing Systems, 34:20596-20607, 2021." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015.", + "[33] Ziheng Qin, Kai Wang, Zangwei Zheng, Jianyang Gu, Xi-angyu Peng, Daquan Zhou, and Yang You. Infobatch: Lossless training speed up by unbiased dynamic data pruning. arXiv preprint arXiv:2303.04947, 2023.", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021.", + "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.", + "[36] Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, and Suvrit Sra. Can contrastive learning avoid shortcut solutions? Advances in neural information processing systems, 34:4974-4986, 2021.", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022.", + "[38] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.", + "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021.", + "[40] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018.", + "[41] Ben Sorscher, Robert Geirhos, Shashank Shekhar, Surya Ganguli, and Ari S Morcos. Beyond neural scaling laws: beating power law scaling via data pruning. NeurIPS, 2022.", + "[42] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Hua-jun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. arXiv preprint arXiv:1811.00491, 2018.", + "[43] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016.", + "[44] Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J Gordon. An empirical study of example forget-" + ], + "bbox": [ + 501, + 90, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3156", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ting during deep neural network learning. arXiv preprint arXiv:1812.05159, 2018.", + "[45] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017.", + "[46] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2015.", + "[47] Haoran Wang, Yue Zhang, Xiaosheng Yu, et al. An overview of image caption generation methods. Computational intelligence and neuroscience, 2020, 2020.", + "[48] Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Kevin Qinghong Lin, Satoshi Tsutsui, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, et al. All in one: Exploring unified video-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6598-6608, 2023.", + "[49] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022.", + "[50] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12196-12205, 2022.", + "[51] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018.", + "[52] Hu Xu, Saining Xie, Po-Yao Huang, Licheng Yu, Russell Howes, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Cit: Curation in training for effective vision-language data. arXiv preprint arXiv:2301.02241, 2023.", + "[53] Xi Yan, David Acuna, and Sanja Fidler. Neural data server: A large-scale search engine for transfer learning data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3893-3902, 2020.", + "[54] Xingcheng Yao, Yanan Zheng, Xiaocoong Yang, and Zhilin Yang. Nlp from scratch without large-scale pretraining: A simple and efficient framework. In International Conference on Machine Learning, pages 25438-25451. PMLR, 2022.", + "[55] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014.", + "[56] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. ICLR, 2022.", + "[57] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022." + ], + "bbox": [ + 76, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016.", + "[59] Ruonan Yu, Songhua Liu, and Xinchao Wang. Dataset distillation: A comprehensive review. arXiv preprint arXiv:2301.07014, 2023.", + "[60] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. 2021.", + "[61] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. arXiv preprint arXiv:2006.05929, 2020." + ], + "bbox": [ + 501, + 90, + 892, + 303 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3157", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Too Large; Data Reduction for Vision-Language Pre-Training/bf4cd6c9-c260-42c2-88a5-f694382bd718_model.json b/2023/Too Large; Data Reduction for Vision-Language Pre-Training/bf4cd6c9-c260-42c2-88a5-f694382bd718_model.json new file mode 100644 index 0000000000000000000000000000000000000000..86475764236c511791c85a9166eea31c11bcd042 --- /dev/null +++ b/2023/Too Large; Data Reduction for Vision-Language Pre-Training/bf4cd6c9-c260-42c2-88a5-f694382bd718_model.json @@ -0,0 +1,2510 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.129, + 0.812, + 0.154 + ], + "angle": 0, + "content": "Too Large; Data Reduction for Vision-Language Pre-Training" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.181, + 0.74, + 0.24 + ], + "angle": 0, + "content": "Alex Jinpeng Wang, Kevin Qinghong Lin, David Junhao Zhang, Stan Weixian Lei, Mike Zheng Shou* Show Lab, National University of Singapore" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.273, + 0.314, + 0.289 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.306, + 0.474, + 0.669 + ], + "angle": 0, + "content": "This paper examines the problems of severe image-text misalignment and high redundancy in the widely-used large-scale Vision-Language Pre-Training (VLP) datasets. To address these issues, we propose an efficient and straightforward Vision-Language learning algorithm called TL;DR, which aims to compress the existing large VLP data into a small, high-quality set. Our approach consists of two major steps. First, a codebook-based encoder-decoder captioner is developed to select representative samples. Second, a new caption is generated to complement the original captions for selected samples, mitigating the text-image misalignment problem while maintaining uniqueness. As the result, TL;DR enables us to reduce the large dataset into a small set of high-quality data, which can serve as an alternative pre-training dataset. This algorithm significantly speeds up the time-consuming pretraining process. Specifically, TL;DR can compress the mainstream VLP datasets at a high ratio, e.g., reduce well-cleaned CC3M dataset from 2.82M to 0.67M (\\(\\sim 24\\%\\)) and noisy YFCC15M from 15M to 2.5M (\\(\\sim 16.7\\%\\)). Extensive experiments with three popular VLP models over seven downstream tasks show that VLP model trained on the compressed dataset provided by TL;DR can perform similar or even better results compared with training on the full-scale dataset1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.7, + 0.21, + 0.715 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.47, + 0.864 + ], + "angle": 0, + "content": "The recent \"scale-is-everything\" viewpoint has become a widely accepted notion in the Vision-language Pre-training (VLP) community [40, 7, 34, 17, 1]. According to this view, the scale of the data has increased from the original tens of thousands-level (e.g., COCO [26] and VG [20]) to millions-level (e.g., CC3M [40] and CC12M [7]), and even up to billions-level (e.g., YFCC100M [43], WIT400M [34], and LAION400M [39]). Approaches [57, 34, 17] trained on these large-scale data show remarkable performance im" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.272, + 0.691, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.272, + 0.88, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.397, + 0.895, + 0.519 + ], + "angle": 0, + "content": "Figure 1: Does using more data really lead to better performance in VLP? Instead of training on the full-scale CC3M dataset, we delete data with low image-text matching score. We find that BLIP [22] model pretrained on \\(50\\%\\) reserved data even obtains better result than full-scale dataset on downstream COCO retrieval [26]. This observation exposes there exists serious misalignment between text&visual modalities and data redundancy in dataset." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.535, + 0.767, + 0.549 + ], + "angle": 0, + "content": "provement in various downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.703 + ], + "angle": 0, + "content": "However, simply scaling-up data brings two critical challenges: \\( i \\). Larger image-text datasets lead to more training cost (e.g., Pretraining CoCa takes about 5 days on 2,048 CloudTPUv4 chips [57]) and storage overhead, which is difficult to afford. \\( ii \\). Obtaining high-quality VLP data requires massive data and well-designed collecting/filtering pipeline, which is expensive. For instance, the CC3M [40] data was obtained after filtering 5 billion collected images. These challenges are daunting and may impede the participation of numerous researchers in the VLP community." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In this study, we stop hunting for larger-scale data blindly and ask an important question: Does employing a larger dataset always result in better performance in VLP? To explore and answer this question, we begin with a simple experiment. First, we utilize a pre-trained BLIP [22] model to calculate the Image-Text Matching (ITM) scores for all samples in the clean CC3M dataset. Subsequently, we remove a portion of the samples with the lowest ITM scores and evaluate the transfer learning results, as shown in Figure 1. Surprisingly, discarding \\(50\\%\\) of the samples slightly improves performance. This remarkable finding challenges the prevailing belief that employing larger amounts of data invariably leads to superior VLP outcomes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.227, + 0.888 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.434, + 0.901 + ], + "angle": 0, + "content": "1https://github.com/showlab/datacentric.vlp" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.434, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3147" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.088, + 0.892, + 0.152 + ], + "angle": 0, + "content": "
MethodYearData TypeCompression Ratio↑Task AgnosticLarge-scaleSupervisionGeneration/Selection
Dataset Distillation [51]2017Image99%-99.99%XXClass LabelGeneration
Data Pruning [41]2022Image20%-30%X✓Class LabelSelection
Neural Data Server [53]2020Multi-modality94%-98%X✓Image-text PairsSelection
TL;DR (ours)-Multi-modality75%-90%✓✓Image-text PairsGeneration+Selection
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.155, + 0.893, + 0.2 + ], + "angle": 0, + "content": "Table 1: Data-efficient learning methods. \"Large-scale\" means that the methods are effective when used on datasets that are very large in size. The \"task agnostic\" means that the methods can be used regardless of the specific downstream task, and without any prior exposure to the associated data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.47, + 0.367 + ], + "angle": 0, + "content": "This experiment suggests removing certain data points can actually improve the model's ability to learn and generalize. Moreover, considering the performance improvements after removing the low ITM score data, we can infer the existence of significant misalignment between the textual and visual modalities in many text-image data pairs (see Figure 7 and the supplementary material for more evidences). These discoveries present promising potentiality to enhance the performance of models that depend on a smaller volume of VLP data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.471, + 0.625 + ], + "angle": 0, + "content": "Driven by above analysis and recent advance in dataset pruning [41], we present a simple, effective and scalable algorithm called \\(TL;DR\\) that aims to improve data efficiency for visual-language pretraining. The \\(TL;DR\\) has a powerful codebook-based captioner, which contains a visual encoder, a look-up codebook and a text decoder. Here is how it works: First, \\(TL;DR\\) feeds each image into the visual encoder and determines the corresponding codes of the image by measuring the similarity between the codebook and the embedding generated by the encoder. Given a large pool of image-text pairs, \\(TL;DR\\) clusters the samples based on their image corresponding codes and selects a representative subset of samples from each cluster. Then, \\(TL;DR\\) further refines the caption of the selected samples via text decoder to reduce text-image misalignment. By doing so, \\(TL;DR\\) is able to significantly reduce the size of the training dataset while maintaining the high quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.792 + ], + "angle": 0, + "content": "In this work, we employ \\(TL; DR\\) on widely-used CC3M, CC12M, YFCC100M and LAION400M datasets and evaluate small size data on three widely-used frameworks including CLIP [34], ViLT [19], and BLIP [22] for data efficiency pretraining with seven representative visual-language downstream tasks. The results show that, with only \\(10\\% - 25\\%\\) data obtained by \\(TL; DR\\), frameworks achieve similar or even better performance compared with the full-scale dataset. We hope our findings can inspire the community to reconsider data efficiency for VLP rather than blindly utilizing increasingly massive datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.807, + 0.22, + 0.822 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.297, + 0.849 + ], + "angle": 0, + "content": "2.1. Data-Efficient Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Recent successes in deep learning are largely attributed to the vast amount of data [10, 34]. However, collecting massive amounts of data is expensive and raises con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.261 + ], + "angle": 0, + "content": "cerns about privacy and copyright [59]. As a result, the research community has become increasingly interested in data-efficient learning, which includes:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.893, + 0.457 + ], + "angle": 0, + "content": "Dataset Distillation [51, 61, 50, 33] compress a large dataset into a small set of synthetic samples, enabling models trained on the smaller dataset to achieve competitive performance with those trained on the original dataset. However, these techniques are only effective on relatively small datasets at low resolutions, such as CIFAR [21], and their performance deteriorates significantly when applied to larger-scale datasets. For example, the accuracy of a model trained on the state-of-the-art MMT's generated data is only \\(33.8\\%\\) on the ImageNet-1K [10] test result [6], while pre-training on real ImageNet-1K achieves over \\(80\\%\\) accuracy [9]. Furthermore, these methods necessitate supervised class labels, which are not suitable for multimodal data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Data Pruning [44, 31] assumes high redundancy in large datasets, selecting only a subset of challenging samples. [29, 31] observed that during the entire training process, some examples are learned early and never forgotten, while others can be repeatedly learned and forgotten. The related work [41] uses a hard sample selection method to select \\(80\\%\\) samples of the ImageNet dataset, and the model trained on selected samples approximating training on all data. Another recent work, CiT [52], also proposes to train models with dynamic training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.716 + ], + "angle": 0, + "content": "Neural Data Server (NDS) [53, 27, 5] proposes a large-scale search engine to identify the most useful transfer learning data from large corpus. While these methods can be extended to multi-modality data, a similar idea has also been applied in NLP [54]. However, this setting assumes that the user has access to all downstream data and needs to train the downstream task using additional retrieval data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.892, + 0.807 + ], + "angle": 0, + "content": "In this work, we are different from previous techniques in that we attempt to compress large-scale multimodal data for the first time, leading to comparable performance between the compressed and original vision-language datasets. We provide a comparison of our approach with these related works in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.766, + 0.833 + ], + "angle": 0, + "content": "2.2. Visual-Language Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Large-scale Vision-Language Pre-training (VLP) involves training on extensive multi-modality data and evaluating performance on various downstream vision-language tasks. Conventional frameworks include the dual-stream" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3148" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.415, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.089, + 0.858, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.259, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Figure 2: Our TL;DR architecture. We first train a codebook-based captioner in Stage1. Then the learned codebook and captioner are used to reduce VLP data in Stage 2. Pre-training on the reduced dataset achieves similar performance to the original full-scale dataset across downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.473, + 0.485 + ], + "angle": 0, + "content": "architecture [34], the one-stream architecture [19, 24, 48, 25, 49], and the encoder-decoder architecture [22]. Previous works have relied on high-quality, human-annotated datasets such as COCO [26] (110K images) and Visual Genome [20] (100K). As model sizes continue to increase, pre-training requires even more data than before [26, 17, 49], resulting in an extremely high computational cost. However, obtaining large and high-quality multi-modality data is challenging due to the difficulties in annotation. In this paper, we aim to democratize VLP research by proposing a general compression method for existing VLP data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.503, + 0.17, + 0.518 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.47, + 0.621 + ], + "angle": 0, + "content": "Our \\(TL;DR\\) is a simple yet effective approach for compressing the Vision-Language Pre-training dataset, leading to further reduction of the training cost. Our approach consists of two stages: (1) codebook-based captioner training and (2) data reduction including samples selection and caption refining. Figure 2 illustrates the idea, introduced next." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.635, + 0.325, + 0.65 + ], + "angle": 0, + "content": "3.1. Codebook-based Captioner" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.901 + ], + "angle": 0, + "content": "The captioner consists of a visual encoder, a codebook and a text decoder. The visual encoder is employed to extract image features. Inspired by vector quantisation technique [56, 45], we try to quantize the image feature for further clustering by utilizing a learnable codebook. Codebook comprises \\( K \\) learnable embedding vectors, each of which can be regarded as a code. Each token of image features conducts a nearest neighbor look-up from codebook and finds its corresponding code. In this way, image features are quantized into a couple of codes (quantized vectors). The quantized vectors are then sent into a text decoder, generating a caption. In order to enhance the quality of text generation, we initialize the codebook with the text embedding of \\( K \\) most frequently occurring keywords/keyphrases in the entire dataset, which enables the codebook to contain meaningful and intuitively understandable semantics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.895, + 0.426 + ], + "angle": 0, + "content": "To train the whole captioner, we utilize a Language Modeling loss [11], which maximizes the likelihood of the text in an autoregressive manner, and a symmetric commitment loss [56], which is specifically designed for codebook. We initially train this captioner on noisy source data and subsequently fine-tune it on smaller-scale datasets, such as COCO [26] and VisualGenome [20]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.436, + 0.658, + 0.451 + ], + "angle": 0, + "content": "3.2. Data Reduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Currently, large-scale datasets exist with serious redundancy [41]. Meanwhile, a large part of texts is noisy and misaligned with images in VLP data. See Figure 2 for the example (the caption \"You need to think twice before buying a pet as present\" does not match the image). To overcome these limitations, we use the learned codebook to condense large-scale noisy data and the learned captioner to reduce the misalignment over image-text pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.893, + 0.779 + ], + "angle": 0, + "content": "Samples selection. For an encoded image feature with \\( L \\) tokens, we compute an index vector with length \\( L \\). Each value is the index of the code, which is the closest to each token. This vector maps the features from image space to semantic space so that it reduces the complexity of the image, benefiting and accelerating the cluster process. Subsequently, each image sample in the dataset is equipped with an index vector according to the above process and we cluster these vectors into \\( N \\) clusters with K-Means (speed up by Faiss [18]). Then we uniformly sample \\( M\\% \\) data points from each cluster, producing a small subset of the dataset. We examine various sampling methods and observe that uniform sampling is stable across different scales." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Caption refining. To alleviate the misalignment problem, we want to improve the text quality using the generated caption. Generated text \\( T_{g} \\) is from the text decoder, which takes the quantized vector of the image as input. We simply concatenate \\( T_{g} \\) with original text \\( T_{o} \\) together, denoted as \\( T = T_{o} + T_{g} \\), to refine and preserve the original caption's uniqueness while maintaining data diversity." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "The compressed small-scale dataset with refined cap" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3149" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.316, + 0.156 + ], + "angle": 0, + "content": "
samplingrefiningTR@1IR@1
65.349.8
✓68.551.9
✓69.452.3
✓✓72.854.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.162, + 0.331, + 0.202 + ], + "angle": 0, + "content": "(a) Component ablation. Both the sampling and refining operation are important to the downstream retrieval." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.209, + 0.356, + 0.277 + ], + "angle": 0, + "content": "
caseTR@1IR@1
Image Embedding70.652.3
Text Embedding69.050.4
BLIP Image Embedding [22]72.354.5
Codebook72.854.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.293, + 0.331, + 0.321 + ], + "angle": 0, + "content": "(d) Clustering feature. Codebook is better than Image Embedding at same scale." + }, + { + "type": "table", + "bbox": [ + 0.372, + 0.089, + 0.605, + 0.155 + ], + "angle": 0, + "content": "
caseTR@1IR@1
gradient-based72.954.8
hard-sample [41]73.154.5
uniform72.854.8
large distance72.353.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.361, + 0.161, + 0.609, + 0.189 + ], + "angle": 0, + "content": "(b) Sample-selection strategy. The different way to select samples." + }, + { + "type": "table", + "bbox": [ + 0.372, + 0.209, + 0.605, + 0.286 + ], + "angle": 0, + "content": "
caseTR@1IR@1
full-scale baseline70.654.0
10%68.952.3
25%72.854.8
50%74.855.2
100%75.157.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.361, + 0.293, + 0.608, + 0.321 + ], + "angle": 0, + "content": "(e) Sampling ratio. Sampling \\(25\\%\\) data is enough to beats with full scale." + }, + { + "type": "table", + "bbox": [ + 0.667, + 0.089, + 0.868, + 0.143 + ], + "angle": 0, + "content": "
caseTR@1IR@1
xavier [14]72.054.1
key words/phrases72.854.8
object tags72.554.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.64, + 0.161, + 0.888, + 0.201 + ], + "angle": 0, + "content": "(c) Codebook Initialization. An codebook initialized with keywords is more stable." + }, + { + "type": "table", + "bbox": [ + 0.667, + 0.209, + 0.868, + 0.286 + ], + "angle": 0, + "content": "
caseTR@1IR@1
10071.853.9
100072.454.5
300072.854.8
500072.954.4
1000072.354.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.64, + 0.293, + 0.888, + 0.321 + ], + "angle": 0, + "content": "(f) Cluster Number. More clusters not equals better result." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.328, + 0.893, + 0.375 + ], + "angle": 0, + "content": "Table 2: TL;DR ablation experiments with BLIP model [22] on CC3M. We report image-to-text retrieval top-1 (TR@1) and text-to-image retrieval top-1 (IR@1) accuracy (\\%) on COCO [26] dataset. If not specified, the default baseline is: pretraining BLIP model based on ViT-B/16 with \\(25\\%\\) sample of CC3M. Default settings are marked in gray." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.469, + 0.451 + ], + "angle": 0, + "content": "tions is recorded as dataset \\(D_{c}\\). At last, we train VLP models on this high quality dataset \\(D_{c}\\) and expect the model to achieve comparable performance with original full-scale dataset \\(D\\) on downstream Vision-Language tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.47, + 0.753 + ], + "angle": 0, + "content": "Discussion. Considering the serious misalignment problem, it seems quite straightforward to use pure generated high-quality caption \\( T_{g} \\) to replace original noisy text. Driven by this idea, we try to pretrain BLIP [22] models with \\( T_{o}, T_{g} \\) and \\( T_{o} + T_{g} \\) independently and show the train curve of Image-Text Contrastive (ITC) loss in Figure 3. However, we find the model trained with \\( T_{g} \\) fails into model collapse [36]. This phenomenon can be explained by captioning collapse [46, 47] and one-to-many problem [55] in image cponing. That is, the trained captioner will generate fixed or similar captions for different images, which limits diversity in the output and easily leads to trivial solutions for contrastive loss. On the contrary, the ITC loss for both \\( T_{o} \\) and \\( T_{o} + T_{g} \\) works well and the \\( T_{o} + T_{g} \\) converges better. We also observe the loss of \\( T_{g} \\) is smaller than other two variants at epoch 0-2, which indicates the generated caption matches well with the image. Note that this simple stitching operation on caption does not bring additional computation cost for VLP as the max length in text-encoder Bert [11] model keeps unchanged for all setting." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.763, + 0.249, + 0.779 + ], + "angle": 0, + "content": "3.3. Technical Details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.786, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Our \\(TL;DR\\) can be implemented efficiently, and importantly, does not require any large auxiliary model. The codebook size \\(K\\) is 3000 as default. The selection of keywords/phrases is implemented using the NLTK\\(^2\\). We adopt ViT-B/16 [12] as image encoder and BertLMHead Model [11] as text decoder. In this way, the token length \\(L\\)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.387, + 0.635, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.387, + 0.761, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.387, + 0.89, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.493, + 0.892, + 0.54 + ], + "angle": 0, + "content": "Figure 3: Training curve with CC3M dataset. Simply stitching generated text and original text together solved the model collapse problem in Image-text Contrastive Loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.893, + 0.658 + ], + "angle": 0, + "content": "is 196 as default. The cross-attention is computed over image embedding and text embedding. To show the generality of compressed dataset, we test \\( D_{c} \\) on three different and representative VLP architectures: dual-stream CLIP [34], one-stream ViLT [19] and Fusion-encoder Blip [22] on various downstream tasks. All these models are trained under the same setting with different datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.673, + 0.694, + 0.691 + ], + "angle": 0, + "content": "4. CC3M Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.893, + 0.836 + ], + "angle": 0, + "content": "We first study dataset reduction on well-cleaned \\( CC3M \\) [40] which heavily filters web crawled pairs and only keeps \\( 0.1\\% \\) of the raw data. This dataset contains a total of 2.8 million images. We employ our TL;DR to compress the CC3M dataset, then conduct pre-training and fine-tuning evaluations on both original and compressed datasets. Following our ablation study, we transfer the pre-trained model to seven Vision-Language tasks downstream and fine-tune it through end-to-end training to evaluate its performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Training. We utilize PyTorch [30] to implement our models and trained them on 8 NVIDIA A100 GPUs to reduce the data samples. For Vision-Language Pre-training, we utilize 2 nodes, each equipped with 16 GPUs. The im" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.249, + 0.901 + ], + "angle": 0, + "content": "2https://github.com/pltk/pltk" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3150" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.087, + 0.898, + 0.273 + ], + "angle": 0, + "content": "
MethodDataset#SamplesMSCOCO (5K test set)Flickr30K (1K test set)
Image→TextText→ImageImage→TextText→Image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP [34]CC3M [40]2.82M60.485.393.248.975.484.777.391.193.271.690.191.4
TL;DR-CC3M0.67M60.385.693.849.477.486.082.591.892.272.090.592.1
CC3M [40]2.82M36.264.380.129.957.966.967.483.292.454.384.190.8
TL;DR-CC3M0.67M37.764.680.830.758.468.268.585.492.055.682.190.8
ViLT [19]CC3M [40]2.82M66.789.293.852.579.387.183.892.093.274.092.092.8
TL;DR-CC3M0.67M67.188.794.153.178.988.285.392.493.675.692.192.5
CC3M [40]2.82M39.268.677.830.453.266.170.588.792.157.684.992.6
TL;DR-CC3M0.67M43.570.881.433.957.966.873.290.593.358.684.792.4
BLIP [22]CC3M [40]2.82M70.991.396.154.380.288.086.394.194.874.891.692.6
TL;DR-CC3M0.67M72.891.995.954.880.689.487.594.895.375.792.293.4
CC3M [40]2.82M42.367.877.431.555.766.375.191.293.660.685.991.8
TL;DR-CC3M0.67M48.773.182.736.760.670.476.391.993.961.087.793.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.141, + 0.277, + 0.825, + 0.292 + ], + "angle": 0, + "content": "Table 3: Fine-tuning and zero-shot image-text retrieval results on MSCOCO and Flickr30K dataset." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.3, + 0.885, + 0.365 + ], + "angle": 0, + "content": "
Dataset#SamplesVQANLVR2valRefCOCO+COCO Caption
test-devtest-standarddevtest-PtestAtestBB@4CIDEr
Random-CC3M0.67M68.366.273.673.868.671.862.835.9118.8
CC3M [40]2.8M71.571.876.076.272.476.165.336.8121.6
TL;DR-CC3M0.67M73.1+1.673.2+1.477.7+1.778.0+1.875.1+2.778.5+2.468.4+3.137.6+0.8123.8+2.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.365, + 0.891, + 0.396 + ], + "angle": 0, + "content": "Table 4: Comparison with BLIP model pre-trained on different data sources for VQA, NLVR², RefCOCO+ and COCO Captioning. ViLT and CLIP architectures can not be evaluated on part of these tasks since structural limitations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.473, + 0.665 + ], + "angle": 0, + "content": "age transformer is initialized from ViT pre-trained on ImageNet [10], and the text transformer is initialized from BERTbase [11](BLIP,CLIP) and DistillBERT [38] (ViLT). The model is pre-trained for 20 epochs with a batch size of 1260 and an AdamW [28] optimizer with a weight decay of 0.05. During training, we apply a learning rate warm-up to 3e-4 and a linear decay with a rate of 0.85. For image augmentation, we utilize RandAugment [8] and apply all of the original policies except color inversion. This decision is based on the recognition of the crucial role that color information plays in the data. For pre-training, images were randomly cropped to a resolution of \\(224 \\times 224\\). We then increase this to \\(384 \\times 384\\) for fine-tuning downstream tasks. Further information about the training hyperparameters for downstream tasks can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.672, + 0.24, + 0.689 + ], + "angle": 0, + "content": "4.1. Main Properties" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.47, + 0.726 + ], + "angle": 0, + "content": "We ablate our \\(TL;DR\\) using the default setting in Table 2 (see caption). Several intriguing properties are observed." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.73, + 0.47, + 0.852 + ], + "angle": 0, + "content": "Module deconstruction. In Table 2a we analyze the impact of different components in \\(TL;DR\\). We establish a baseline by randomly selecting \\(25\\%\\) of the data from CC3M (first row). Our results show that codebook-based sampling outperforms random selection by \\(3.2\\%\\) in TR@1. We also observe that both codebook-based sampling and caption refinement are crucial and the combination of them achieves optimal downstream performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Sample selection. In Table 2b we study the sample selection strategy in Stage 2. We sample \\(25\\%\\) data in each cluster by default. For Gradient-based, we train a tiny network" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.419, + 0.89, + 0.544 + ], + "angle": 0, + "content": "
MethodDatasetR@1↑R@5↑R@10↑MdR↓
CLIP [34]Rand-CC3M15.334.846.313.0
CC3M [40]19.437.347.511.0
TL;DR-CC3M21.838.648.510.0
ViLT [19]Rand-CC3M18.838.249.511.0
CC3M [40]21.040.551.510.0
TL;DR-CC3M22.542.752.48.0
BLIP [22]Rand-CC3M23.342.853.38.0
CC3M [40]26.046.358.07.0
TL;DR-CC3M27.448.759.46.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.545, + 0.892, + 0.605 + ], + "angle": 0, + "content": "Table 5: MSRVTT-1K retrieval using three architectures. We created a subset of the CC3M dataset called Rand-CC3M by randomly selecting the same number of samples as in TL;DR-CC3M." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.818 + ], + "angle": 0, + "content": "to conduct VLP pretrained with ITC [24], ITM [24] and LM [11]. Then we select samples which contribute most to the gradients in each cluster. Large distance: Another perspective is that data points on the border of each cluster are more important than those at the center [4]. So we first compute the center of each cluster and then choose the sample that has the largest distance from the center of each cluster. We also report the result of hard-sample selection from [41]. We observe that all these variants produce similar results except large distances. This suggests that the clustering step, rather than the selection step, plays a key role in data compression during Stage 2. To maintain simplicity, we choose uniform sampling as the default method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Codebook initialization. In Table 2c we compare different initialization strategies. The xavier means all parameters in the codebook are initialized with xavier initialization [14]. For the object tags initialization, following previous works [2, 60], we use the 1600 object tags from" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3151" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.47, + 0.213 + ], + "angle": 0, + "content": "
ModelDataset#SamplesImNetImNet-AImNet-R
CLIP [34]Rand-CC3M0.67M58.361.862.3
CC3M [40]2.82M62.265.266.9
TL;DR-CC3M0.67M61.465.065.7
ViLT [19]Rand-CC3M0.67M54.359.858.4
CC3M [40]2.82M58.662.964.2
TL;DR-CC3M0.67M59.163.364.0
BLIP [22]Rand-CC3M0.67M57.361.865.2
CC3M [40]2.82M62.565.568.1
TL;DR-CC3M0.67M62.063.967.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.215, + 0.471, + 0.292 + ], + "angle": 0, + "content": "Table 6: Zero-shot image classification results on ImageNet [10], ImageNet-A [16], ImageNet-R [15]. There is no free lunch, as selecting partial samples reduces the visual diversity crucial for classification. Despite this, TL;DR still performs significantly better than random selection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.471, + 0.408 + ], + "angle": 0, + "content": "Visual Genome [20] and extract text feature with a pretrained BERT [11]. With same training setting, the keywords achieve a \\(0.8\\%\\) TR@1 improvement and a \\(0.7\\%\\) IR@1 improvement over xavier. This result is expected as the text embeddings provide contextual information and simplify the learning process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.613 + ], + "angle": 0, + "content": "Codebook vs. Image embedding. In Table 2d, we investigate different ways of cluster sampling. First, we remove the codebook from Stage-1 and use image embedding instead. Alternatively, we directly cluster images using the image embedding [22] of images from BLIP model (pretrained on 200M Image-text pairs). We observe the image embedding leads to much better result than text embedding. This is reasonable because clustering visual-similarity samples with text only is difficult. We observe that clustering depended on our codebook performs better than both image embedding and text embedding. This demonstrates that our codebook can efficiently project image embedding to semantic space, benefiting cluster process." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.62, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Cluster sampling ratio. Table 2e varies the sampling ratio of each cluster from \\(10\\%\\) to \\(100\\%\\). We are surprised to find that a low sampling ratio can still produce effective results. With only \\(25\\%\\) of the data and the TL;DR model, we are able to achieve a \\(1.9\\%\\) improvement on TR@1 and a \\(0.8\\%\\) improvement on IR@1 over the full-scale baseline. Additionally, we observe that larger sampling ratios lead to even better results. Since our focus is on achieving similar transfer learning results with fewer samples, we use a default sampling ratio of \\(25\\%\\) to minimize computation costs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Cluster numbers. In Table 2f, we investigate the impact of cluster number on Stage 2 by increasing it from 300 to \\(30\\mathrm{K}\\). We observe that using more clusters results in a slight improvement at the beginning and becomes stable when the number of clusters exceeds \\(3\\mathrm{K}\\). Moreover, all results consistently outperform the random selection baseline. Therefore, we use \\(3\\mathrm{K}\\) clusters as the default in this work, as it performs well on fine-tuning tasks." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.582, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.507, + 0.158, + 0.621, + 0.175 + ], + "angle": 0, + "content": "Captioner: A racer jumps over a railing on a horse." + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.092, + 0.745, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.092, + 0.868, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.18, + 0.882, + 0.196 + ], + "angle": 0, + "content": "Figure 4: The generated caption match the image well." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.21, + 0.785, + 0.227 + ], + "angle": 0, + "content": "4.2. Transfer Learning Experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.894, + 0.355 + ], + "angle": 0, + "content": "We conduct an extensive evaluation of transfer learning in downstream tasks using the model pre-trained on our compressed TL;DR-CC3M and source CC3M with 3 architectures. Our evaluation primarily focuses on the core tasks of three categories that examine: (1) cross-modality alignment, (2) image captioning and multi-modality understanding capabilities, and (3) visual recognition. The baseline in this section is the model trained on CC3M dataset." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.378, + 0.776, + 0.393 + ], + "angle": 0, + "content": "4.2.1 Cross-modality Alignment Task" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.893, + 0.62 + ], + "angle": 0, + "content": "Image-Text retrieval. Fine-grained world region alignment plays a critical role in this task. We report both image-to-text retrieval (TR) and text-to-image retrieval (IR) on the COCO [26] and Flickr30K [32] benchmarks. For the BLIP [22] model, we adopt an additional re-ranking strategy, following the original implementation. In Table 3, we also report zero-shot retrieval results. We found that \\(TL; DR\\) achieves comparable results with the baselines on all metrics and surprisingly performs quite well on zero-shot results. For example, for the BLIP [22] architecture, our method leads to a \\(6.4\\%\\) improvement (from \\(42.3\\%\\) to \\(48.7\\%\\)) in Recall@1 of image-to-text retrieval on MSCOCO. All results suggest that a small part of refined image-text pairs is enough to learn good alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.894, + 0.76 + ], + "angle": 0, + "content": "Zero-shot video retrieval. In this experiment, we analyze the generalization ability of our method to video-language tasks. Specifically, we perform zero-shot transfer to text-to-video retrieval and evaluate the models trained on COCO-retrieval in Table 5. We uniformly sample 8 frames from each video to process the video input and concatenate the frame features into a single sequence. These models trained on our compressed dataset outperform the baseline on all metrics, demonstrating the generality of \\(TL; DR\\)." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.774, + 0.692, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.774, + 0.828, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.775, + 0.892, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.833, + 0.894, + 0.894 + ], + "angle": 0, + "content": "Figure 5: The codebook-based clusters visualization. The samples within each cluster exhibit similar contextual characteristics, as opposed to mere visual appearance. For example, the \"Christmas elements\" cluster located at the right." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3152" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.208, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.184, + 0.175, + 0.192 + ], + "angle": 0, + "content": "VQ-GAN: 5.13s" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.089, + 0.332, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.184, + 0.301, + 0.192 + ], + "angle": 0, + "content": "DALL E2: 2.65s" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.089, + 0.457, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.184, + 0.445, + 0.191 + ], + "angle": 0, + "content": "Stable Diffusion: 0.17s" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.195, + 0.468, + 0.224 + ], + "angle": 0, + "content": "Figure 6: Image generation result with strong Text-to-image Model. The generation time is also reported." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.24, + 0.468, + 0.269 + ], + "angle": 0, + "content": "4.2.2 Image Captioning and Multi-modality Understanding Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.468, + 0.418 + ], + "angle": 0, + "content": "Image captioning. The task involves describing an input image, which we evaluate using No-Caps and COCO datasets. Both datasets are fine-tuned on COCO with the Language Modeling (LM) loss. We adopt a zero-shot setting for No-Caps dataset, and start each caption with the phrase \"a picture of\" for the BLIP architecture. We do not pre-train using COCO to avoid information leakage. Our results outperform baseline with a much smaller quantity of pre-training data, as shown in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.468, + 0.56 + ], + "angle": 0, + "content": "Visual question answering (VQA). We evaluate our model's performance on the VQA task [3], where the model needs to provide an answer based on an image and a question. We consider it as an answer generation task that allows open-vocabulary VQA for better results, following previous works [23, 22]. The results are presented in Table 4. The BLIP trained on TL;DR-CC3M outperforms baseline by \\(1.4\\%\\) on test-dev splits, demonstrating the effectiveness of our compressed dataset for improving VQA performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.468, + 0.67 + ], + "angle": 0, + "content": "Visual reasoning. The Natural Language Visual Reasoning \\((\\mathrm{NLVR}^2)\\) [42] task is a binary classification task that requires the model to reason about two images and a question in natural language. Multi-modal reasoning is crucial for the completion of this task. We observe that BLIP trained on our dataset achieved \\(78.0\\%\\) accuracy compared to \\(76.2\\%\\) achieved by the CC3M, as shown in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.468, + 0.794 + ], + "angle": 0, + "content": "Cross-modality grounding. Referring Expression (RE) Comprehension requires the model to select the target object from a set of image regions proposals, based on the query description. This task heavily relies on visual-grounding ability. The models are evaluated on ground-truth objects, and we evaluate RE Comprehension on RefCOCO+ [58]. The results are reported in Table 4, and we observe that TL;DR-CC3M achieves better results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.306, + 0.831 + ], + "angle": 0, + "content": "4.2.3 Visual Recognition Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.899 + ], + "angle": 0, + "content": "Besides the cross-modality task, we also explore a unimodality task, mainly image classification. Specifically, we fix the image encoder and explore zero-shot image classification. Since the vision encoder is loaded from" + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.089, + 0.857, + 0.153 + ], + "angle": 0, + "content": "
MethodTR@1IR@1
real data58.344.0
VQ-GAN [13]35.232.4
DALLE2 [35] (implement from 3)44.338.3
Stable Diffusion [37] (implement from 4)52.440.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.154, + 0.892, + 0.199 + ], + "angle": 0, + "content": "Table 7: Compare different sample generation methods over 0.3M subset of CC3M. We first pre-train BLIP model on these generated data and then evaluation on COCO." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.217, + 0.892, + 0.518 + ], + "angle": 0, + "content": "pretrained model, this task demonstrates the impact of training a Vision-Language model with noisy image-text pairs. Specifically, it shows how such training affects the well-learned representation derived from a human-crafted dataset. We show the results in Table 6 and observe noisy data leads to significant Catastrophic Forgetting. For example, the CLIP model drops down to only \\(58.3\\%\\) accuracy with noisy data training. We also observe our TL;DR shows steady improvement for all architectures over random selection. Unfortunately, the classification performance for TL;DR-CC3M is slightly worse than the full-scale CC3M for the CLIP and BLIP architectures. Both of these architectures have independent image encoders like ViT to extract image embeddings. This indicates that this task heavily relies on visual diversity, which is different from multimodal tasks, and our method reduces the visual diversity potentially. For the ViLT model, this architecture adopts a shared backbone for both visual and text, and we observe the slightly different results. We guess that multi-modality interaction in early-fusion affects the classification result." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.53, + 0.637, + 0.543 + ], + "angle": 0, + "content": "4.3. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.724 + ], + "angle": 0, + "content": "Generated caption visualization. We show the generated caption in Figure 4. It is evident that the original captions can be highly abstract and difficult to match their respective images, even for human observers sometimes. For instance, when the ITM score is as low as 0.04, matching the figure with its corresponding caption becomes arduous. Such challenging cases can potentially harm the cross-modality alignment. In contrast, we observe that the generated captions describe the image very well and sometimes offer helpful complementary information. For example, \"bus\" and \"castle\" in the middle example." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.892, + 0.865 + ], + "angle": 0, + "content": "Codebook-based cluster visualization. Figure 5 displays the codebook grouping result achieved with simple KMeans. Clusters are sets of data points with similar characteristics, often defined by their features or attributes. Interestingly, we observe that the model cluster samples \"accurate\", meaning that these samples have semantic similarity rather than simple appearance. For instance, the model classifies \"dollars\" and \"piggy bank\" together, even though they differ significantly in appearance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.875, + 0.751, + 0.888 + ], + "angle": 0, + "content": "\\(^{3}\\)https://github.com/LAION-AI/dalle2-laion" + }, + { + "type": "page_footnote", + "bbox": [ + 0.519, + 0.888, + 0.741, + 0.901 + ], + "angle": 0, + "content": "\\(^{4}\\)https://github.com/huggingface/diffusers" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.875, + 0.751, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "3153" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.088, + 0.895, + 0.226 + ], + "angle": 0, + "content": "
Dataset#Samples TimeVQA test-devNLVR2 test-PRefCOCO valNocaps Captioning B@4CIDErFlcikr30K Retrieval TR@1IR@1Imagenet Acc
Rand-CC12M2.4M14h71.876.272.536.8121.082.973.361.2
CC12M [7]10.8M65h73.578.974.137.5122.984.775.365.3
TL;DR-CC12M2.4M14h74.1+0.678.5-0.474.0-0.138.1+0.6124.1+1.285.5+0.876.3+1.063.8-1.5
Rand-YFCC15M2.5M15h67.270.568.135.2116.378.870.565.4
YFCC15M [43]15M90h70.574.270.635.9118.481.572.467.8
TL;DR-YFCC15M2.5M15h70.3-0.275.3+1.172.6+2.037.2+1.3122.5+4.182.3+0.874.3+1.967.3-0.5
Rand-LAION40M(128)8M48h70.775.373.434.8113.280.472.568.5
LAION40M(128) [39]40M120h74.579.176.635.2117.483.274.971.3
TL;DR-LAION40M(128)8M48h76.3+1.880.5+1.477.4+0.836.8+1.6120.9+3.582.8-0.476.1+1.270.4-0.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.227, + 0.895, + 0.259 + ], + "angle": 0, + "content": "Table 8: Comparison with different source of data on 6 downstream tasks. BLIP [22] is adopted as baseline and (128) means the image resolution is \\( {128} \\times {128} \\) . We also list the pre-training time,which can be significantly reduced via \\( {TL};{DR} \\) ." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.271, + 0.275, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.272, + 0.447, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.391, + 0.471, + 0.422 + ], + "angle": 0, + "content": "Figure 7: ITM score distribution. \\(TL; DR\\) alleviates the issue of misalignment in VLP data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.44, + 0.26, + 0.456 + ], + "angle": 0, + "content": "4.4. More Investigation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.651 + ], + "angle": 0, + "content": "Is image generation possible? To ease the misalignment problem of image-text pairs, instead of simply selecting representative samples, a potential and naive idea is to generate images from text. To this end, we randomly sample 0.3M subset of CC3M and generate image from text with three popular text to image generation models, VQ-GAN [56], DALLE 2 [35] and Stable Diffusion [37]. We display the generated samples in Figure 6. We observe that the generative models struggle with complex scenarios, but are capable of generating simple prompts like \"dog\" proficiently. In addition, generation methods only produce visual cues in a fixed vocabulary, potentially reducing data diversity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.47, + 0.789 + ], + "angle": 0, + "content": "Next, we pre-train BLIP models on these generated data and evaluate it on COCO Retrieval. In Table 7 we observe the results of transfer learning depend on the quality of generated samples, with those generated by stable diffusion being particularly effective. However, there still exists a significant gap between the generated data and the real dataset (e.g., \\(52.4\\%\\) vs. \\(58.3\\%\\) on TR@1). We believe that higher-quality and diverse generated images may lead to comparable results with real images in the near future." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Explore the misalignment problem. Figure 7 shows the Image-text Matching (ITM) score distribution for both \\( CC3M \\) and our \\( TL;DR-CC3M \\) data (the visualization about more datasets is reported in the supplementary). We observe a lot of samples of original \\( CC3M \\) at low matching score even tends to zero, which indicates the current dataset has serious misalignment problems. Since image" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.895, + 0.351 + ], + "angle": 0, + "content": "text matching (ITM) loss and image-text contrastive (ITC) loss are used in all architectures, these samples will damage the multimodal representation learning. When adopting our \\(TL;DR\\), we observe that the matching score tends to be higher and has very few samples with low ITM score." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.37, + 0.785, + 0.387 + ], + "angle": 0, + "content": "5. Transfer to other VLP datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.892, + 0.702 + ], + "angle": 0, + "content": "We study data compression performed in two categories shown below: clean data that involves human-based offline filter pipelines and raw data that has not undergone cleaning. For clean data, in addition to \\( CC3M \\), we explore the well-cleaned, high-quality dataset \\( CC12M \\) [7]. Then, we study the raw data YFCC100M [43] and LAION400M [39]. \\( CC12M \\) [7] contains 12 million image-text pairs specifically meant to be used for vision-and-language pre-training. These data are collected by relaxing the data collection pipeline as in CC3M. YFCC15M [34] is a subset of the multilingual and noisy YFCC100M [43] that contains English captions. LAION400M [39] is a large-scale noisy dataset that provides URLs with captions for download. To control the computation cost and reduce the storage overhead, we randomly sample a 40M subset of LAION400M and download images at a resolution of \\( 128 \\times 128 \\). So, we record the dataset as \\( TL;DR-LAION40M(128) \\), and the performance over downstream tasks could be improved with higher resolution. More exploration about video-text datasets is reported in the supplementary material" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We use BLIP as the default architecture and evaluate our \\(TL;DR\\) on different datasets and show the results in Table 8. Surprisingly, with only \\(2.5\\mathrm{M}\\) \\((16.7\\%)\\) data, \\(TL;DR\\)-YFCC15M leads to similar results with \\(15\\mathrm{M}\\) raw data over all metrics except Imagenet. More results with different backbones are reported in the supplementary material. For LAION40M(128), when using \\(8\\mathrm{M}\\) data \\((20\\%)\\), the model trained on our dataset consistently outperforms the baseline method on six downstream tasks. We noticed that the compression rate of LAION40M(128) is less than that of YFCC15M. This may be due to the fact that the collection of LAION40M(128) has already been filtered with CLIP similarity, reducing the impact of the misalignment problem." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3154" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.327, + 0.107 + ], + "angle": 0, + "content": "6. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.473, + 0.358 + ], + "angle": 0, + "content": "This paper presents \\(TL;DR\\), a novel and pioneering algorithm for selecting and generating high-quality image-text pairs from noisy Vision-Language Pre-training (VLP) data, thereby contributing to the field of VLP. \\(TL;DR\\) incorporates a text generation process into learning to reduce serious misalignment problems. Our experiments demonstrate three widely-used architectures leads to comparable results and much smaller training cost when learning from our generated dataset. Additionally, we demonstrate that the misalignment problem can be effectively addressed using our simple \\(TL;DR\\). However, the choice of the highest compression ratio is done manually rather than learned. Furthermore, achieving even higher compression ratios for VLP models remains a challenge, and text-to-image generation models may be helpful in this regard. We hope that this perspective will inspire future research." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.374, + 0.238, + 0.391 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.47, + 0.492 + ], + "angle": 0, + "content": "This project is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008, and Meta, Mike Zheng Shou's Start-Up Grant from NUS. The computational work for this article was partially performed on resources of the National Supercomputing Centre, Singapore." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.506, + 0.175, + 0.521 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.531, + 0.47, + 0.6 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.603, + 0.472, + 0.685 + ], + "angle": 0, + "content": "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.688, + 0.473, + 0.757 + ], + "angle": 0, + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.759, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[4] Dibya Jyoti Bora, Dr Gupta, and Anil Kumar. A comparative study between fuzzy clustering algorithm and hard clustering algorithm. arXiv preprint arXiv:1404.6059, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.802, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[5] Tianshi Cao, Sasha Alexandre Doubov, David Acuna, and Sanja Fidler. Scalable neural data server: A data recommender for transfer learning. Advances in Neural Information Processing Systems, 34:8984-8997, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.86, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[6] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.531, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4750-4759, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.123, + 0.894, + 0.191 + ], + "angle": 0, + "content": "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.193, + 0.895, + 0.262 + ], + "angle": 0, + "content": "[8] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[9] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation to imagenet-1k with constant memory. arXiv preprint arXiv:2211.10586, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.632 + ], + "angle": 0, + "content": "[14] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249-256. JMLR Workshop and Conference Proceedings, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.634, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[18] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535-547, 2019." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3155" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[19] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.47, + 0.266 + ], + "angle": 0, + "content": "[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.267, + 0.47, + 0.322 + ], + "angle": 0, + "content": "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[23] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.47, + 0.467 + ], + "angle": 0, + "content": "[24] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[25] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.541, + 0.47, + 0.61 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.612, + 0.47, + 0.655 + ], + "angle": 0, + "content": "[27] Yiqi Lin, Huabin Zheng, Huaping Zhong, Jinjing Zhu, Wei-jia Li, Conghui He, and Lin Wang. Sept: Towards scalable and efficient visual pre-training. AAAI, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.468, + 0.685 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[29] Rafid Mahmood, James Lucas, David Acuna, Daiqing Li, Jonah Philion, Jose M Alvarez, Zhiding Yu, Sanja Fidler, and Marc T Law. How much more data do i need? estimating requirements for downstream tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 275-284, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[31] Mansheej Paul, Surya Ganguli, and Gintare Karolina Dziugaite. Deep learning on a data diet: Finding important examples early in training. Advances in Neural Information Processing Systems, 34:20596-20607, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.176 + ], + "angle": 0, + "content": "[32] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[33] Ziheng Qin, Kai Wang, Zangwei Zheng, Jianyang Gu, Xi-angyu Peng, Daquan Zhou, and Yang You. Infobatch: Lossless training speed up by unbiased dynamic data pruning. arXiv preprint arXiv:2303.04947, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.893, + 0.303 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.893, + 0.358 + ], + "angle": 0, + "content": "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.893, + 0.417 + ], + "angle": 0, + "content": "[36] Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, and Suvrit Sra. Can contrastive learning avoid shortcut solutions? Advances in neural information processing systems, 34:4974-4986, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.894, + 0.488 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.893, + 0.544 + ], + "angle": 0, + "content": "[38] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.893, + 0.616 + ], + "angle": 0, + "content": "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.893, + 0.699 + ], + "angle": 0, + "content": "[40] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.893, + 0.744 + ], + "angle": 0, + "content": "[41] Ben Sorscher, Robert Geirhos, Shashank Shekhar, Surya Ganguli, and Ari S Morcos. Beyond neural scaling laws: beating power law scaling via data pruning. NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.893, + 0.8 + ], + "angle": 0, + "content": "[42] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Hua-jun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. arXiv preprint arXiv:1811.00491, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[43] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[44] Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J Gordon. An empirical study of example forget-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3156" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "ting during deep neural network learning. arXiv preprint arXiv:1812.05159, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.122, + 0.472, + 0.164 + ], + "angle": 0, + "content": "[45] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.167, + 0.472, + 0.222 + ], + "angle": 0, + "content": "[46] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.224, + 0.472, + 0.266 + ], + "angle": 0, + "content": "[47] Haoran Wang, Yue Zhang, Xiaosheng Yu, et al. An overview of image caption generation methods. Computational intelligence and neuroscience, 2020, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.268, + 0.472, + 0.352 + ], + "angle": 0, + "content": "[48] Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Kevin Qinghong Lin, Satoshi Tsutsui, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, et al. All in one: Exploring unified video-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6598-6608, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.354, + 0.472, + 0.41 + ], + "angle": 0, + "content": "[49] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.412, + 0.472, + 0.495 + ], + "angle": 0, + "content": "[50] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12196-12205, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.497, + 0.472, + 0.539 + ], + "angle": 0, + "content": "[51] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.541, + 0.472, + 0.598 + ], + "angle": 0, + "content": "[52] Hu Xu, Saining Xie, Po-Yao Huang, Licheng Yu, Russell Howes, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Cit: Curation in training for effective vision-language data. arXiv preprint arXiv:2301.02241, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.6, + 0.472, + 0.655 + ], + "angle": 0, + "content": "[53] Xi Yan, David Acuna, and Sanja Fidler. Neural data server: A large-scale search engine for transfer learning data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3893-3902, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.657, + 0.472, + 0.713 + ], + "angle": 0, + "content": "[54] Xingcheng Yao, Yanan Zheng, Xiaocoong Yang, and Zhilin Yang. Nlp from scratch without large-scale pretraining: A simple and efficient framework. In International Conference on Machine Learning, pages 25438-25451. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.715, + 0.472, + 0.785 + ], + "angle": 0, + "content": "[55] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.787, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[56] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.845, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[57] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[58] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.204 + ], + "angle": 0, + "content": "[59] Ruonan Yu, Songhua Liu, and Xinchao Wang. Dataset distillation: A comprehensive review. arXiv preprint arXiv:2301.07014, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.893, + 0.261 + ], + "angle": 0, + "content": "[60] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.893, + 0.304 + ], + "angle": 0, + "content": "[61] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. arXiv preprint arXiv:2006.05929, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3157" + } + ] +] \ No newline at end of file diff --git a/2023/Too Large; Data Reduction for Vision-Language Pre-Training/bf4cd6c9-c260-42c2-88a5-f694382bd718_origin.pdf b/2023/Too Large; Data Reduction for Vision-Language Pre-Training/bf4cd6c9-c260-42c2-88a5-f694382bd718_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..62dc58fd9669a63b22ee2c4b31c8db0897ea35d2 --- /dev/null +++ b/2023/Too Large; Data Reduction for Vision-Language Pre-Training/bf4cd6c9-c260-42c2-88a5-f694382bd718_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fcb6e85b4669c5d8f23f1ce4d5c262920829be19fc4675bcccdeee7f29a72ff +size 2460028 diff --git a/2023/Too Large; Data Reduction for Vision-Language Pre-Training/full.md b/2023/Too Large; Data Reduction for Vision-Language Pre-Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1fc48fdf83bb4ccb5c946c5675eb42db9ec27d39 --- /dev/null +++ b/2023/Too Large; Data Reduction for Vision-Language Pre-Training/full.md @@ -0,0 +1,331 @@ +# Too Large; Data Reduction for Vision-Language Pre-Training + +Alex Jinpeng Wang, Kevin Qinghong Lin, David Junhao Zhang, Stan Weixian Lei, Mike Zheng Shou* Show Lab, National University of Singapore + +# Abstract + +This paper examines the problems of severe image-text misalignment and high redundancy in the widely-used large-scale Vision-Language Pre-Training (VLP) datasets. To address these issues, we propose an efficient and straightforward Vision-Language learning algorithm called TL;DR, which aims to compress the existing large VLP data into a small, high-quality set. Our approach consists of two major steps. First, a codebook-based encoder-decoder captioner is developed to select representative samples. Second, a new caption is generated to complement the original captions for selected samples, mitigating the text-image misalignment problem while maintaining uniqueness. As the result, TL;DR enables us to reduce the large dataset into a small set of high-quality data, which can serve as an alternative pre-training dataset. This algorithm significantly speeds up the time-consuming pretraining process. Specifically, TL;DR can compress the mainstream VLP datasets at a high ratio, e.g., reduce well-cleaned CC3M dataset from 2.82M to 0.67M ( $\sim 24\%$ ) and noisy YFCC15M from 15M to 2.5M ( $\sim 16.7\%$ ). Extensive experiments with three popular VLP models over seven downstream tasks show that VLP model trained on the compressed dataset provided by TL;DR can perform similar or even better results compared with training on the full-scale dataset1. + +# 1. Introduction + +The recent "scale-is-everything" viewpoint has become a widely accepted notion in the Vision-language Pre-training (VLP) community [40, 7, 34, 17, 1]. According to this view, the scale of the data has increased from the original tens of thousands-level (e.g., COCO [26] and VG [20]) to millions-level (e.g., CC3M [40] and CC12M [7]), and even up to billions-level (e.g., YFCC100M [43], WIT400M [34], and LAION400M [39]). Approaches [57, 34, 17] trained on these large-scale data show remarkable performance im + +![](images/c884edc4e9f990ead5d34f18305048dcfaea13cd7fa2361ca38e50714552c154.jpg) +Figure 1: Does using more data really lead to better performance in VLP? Instead of training on the full-scale CC3M dataset, we delete data with low image-text matching score. We find that BLIP [22] model pretrained on $50\%$ reserved data even obtains better result than full-scale dataset on downstream COCO retrieval [26]. This observation exposes there exists serious misalignment between text&visual modalities and data redundancy in dataset. + +![](images/2ed4e85d8e0b95abebcd24f052ef1588ad1880b9ba5e8ace5e93115732e73a3e.jpg) + +provement in various downstream tasks. + +However, simply scaling-up data brings two critical challenges: $i$ . Larger image-text datasets lead to more training cost (e.g., Pretraining CoCa takes about 5 days on 2,048 CloudTPUv4 chips [57]) and storage overhead, which is difficult to afford. $ii$ . Obtaining high-quality VLP data requires massive data and well-designed collecting/filtering pipeline, which is expensive. For instance, the CC3M [40] data was obtained after filtering 5 billion collected images. These challenges are daunting and may impede the participation of numerous researchers in the VLP community. + +In this study, we stop hunting for larger-scale data blindly and ask an important question: Does employing a larger dataset always result in better performance in VLP? To explore and answer this question, we begin with a simple experiment. First, we utilize a pre-trained BLIP [22] model to calculate the Image-Text Matching (ITM) scores for all samples in the clean CC3M dataset. Subsequently, we remove a portion of the samples with the lowest ITM scores and evaluate the transfer learning results, as shown in Figure 1. Surprisingly, discarding $50\%$ of the samples slightly improves performance. This remarkable finding challenges the prevailing belief that employing larger amounts of data invariably leads to superior VLP outcomes. + +
MethodYearData TypeCompression Ratio↑Task AgnosticLarge-scaleSupervisionGeneration/Selection
Dataset Distillation [51]2017Image99%-99.99%XXClass LabelGeneration
Data Pruning [41]2022Image20%-30%X✓Class LabelSelection
Neural Data Server [53]2020Multi-modality94%-98%X✓Image-text PairsSelection
TL;DR (ours)-Multi-modality75%-90%✓✓Image-text PairsGeneration+Selection
+ +Table 1: Data-efficient learning methods. "Large-scale" means that the methods are effective when used on datasets that are very large in size. The "task agnostic" means that the methods can be used regardless of the specific downstream task, and without any prior exposure to the associated data. + +This experiment suggests removing certain data points can actually improve the model's ability to learn and generalize. Moreover, considering the performance improvements after removing the low ITM score data, we can infer the existence of significant misalignment between the textual and visual modalities in many text-image data pairs (see Figure 7 and the supplementary material for more evidences). These discoveries present promising potentiality to enhance the performance of models that depend on a smaller volume of VLP data. + +Driven by above analysis and recent advance in dataset pruning [41], we present a simple, effective and scalable algorithm called $TL;DR$ that aims to improve data efficiency for visual-language pretraining. The $TL;DR$ has a powerful codebook-based captioner, which contains a visual encoder, a look-up codebook and a text decoder. Here is how it works: First, $TL;DR$ feeds each image into the visual encoder and determines the corresponding codes of the image by measuring the similarity between the codebook and the embedding generated by the encoder. Given a large pool of image-text pairs, $TL;DR$ clusters the samples based on their image corresponding codes and selects a representative subset of samples from each cluster. Then, $TL;DR$ further refines the caption of the selected samples via text decoder to reduce text-image misalignment. By doing so, $TL;DR$ is able to significantly reduce the size of the training dataset while maintaining the high quality. + +In this work, we employ $TL; DR$ on widely-used CC3M, CC12M, YFCC100M and LAION400M datasets and evaluate small size data on three widely-used frameworks including CLIP [34], ViLT [19], and BLIP [22] for data efficiency pretraining with seven representative visual-language downstream tasks. The results show that, with only $10\% - 25\%$ data obtained by $TL; DR$ , frameworks achieve similar or even better performance compared with the full-scale dataset. We hope our findings can inspire the community to reconsider data efficiency for VLP rather than blindly utilizing increasingly massive datasets. + +# 2. Related Work + +# 2.1. Data-Efficient Learning + +Recent successes in deep learning are largely attributed to the vast amount of data [10, 34]. However, collecting massive amounts of data is expensive and raises con + +cerns about privacy and copyright [59]. As a result, the research community has become increasingly interested in data-efficient learning, which includes: + +Dataset Distillation [51, 61, 50, 33] compress a large dataset into a small set of synthetic samples, enabling models trained on the smaller dataset to achieve competitive performance with those trained on the original dataset. However, these techniques are only effective on relatively small datasets at low resolutions, such as CIFAR [21], and their performance deteriorates significantly when applied to larger-scale datasets. For example, the accuracy of a model trained on the state-of-the-art MMT's generated data is only $33.8\%$ on the ImageNet-1K [10] test result [6], while pre-training on real ImageNet-1K achieves over $80\%$ accuracy [9]. Furthermore, these methods necessitate supervised class labels, which are not suitable for multimodal data. + +Data Pruning [44, 31] assumes high redundancy in large datasets, selecting only a subset of challenging samples. [29, 31] observed that during the entire training process, some examples are learned early and never forgotten, while others can be repeatedly learned and forgotten. The related work [41] uses a hard sample selection method to select $80\%$ samples of the ImageNet dataset, and the model trained on selected samples approximating training on all data. Another recent work, CiT [52], also proposes to train models with dynamic training data. + +Neural Data Server (NDS) [53, 27, 5] proposes a large-scale search engine to identify the most useful transfer learning data from large corpus. While these methods can be extended to multi-modality data, a similar idea has also been applied in NLP [54]. However, this setting assumes that the user has access to all downstream data and needs to train the downstream task using additional retrieval data. + +In this work, we are different from previous techniques in that we attempt to compress large-scale multimodal data for the first time, leading to comparable performance between the compressed and original vision-language datasets. We provide a comparison of our approach with these related works in Table 1. + +# 2.2. Visual-Language Pre-training + +Large-scale Vision-Language Pre-training (VLP) involves training on extensive multi-modality data and evaluating performance on various downstream vision-language tasks. Conventional frameworks include the dual-stream + +![](images/3677a06d50e4d08d552894d71c68f2e5555385dbf6d36aa502b373e52afaaa91.jpg) +Figure 2: Our TL;DR architecture. We first train a codebook-based captioner in Stage1. Then the learned codebook and captioner are used to reduce VLP data in Stage 2. Pre-training on the reduced dataset achieves similar performance to the original full-scale dataset across downstream tasks. + +![](images/8417fe376969c3596a6b7a530f6f0746a205a561eb10606be6bfa6b79ade3aec.jpg) + +architecture [34], the one-stream architecture [19, 24, 48, 25, 49], and the encoder-decoder architecture [22]. Previous works have relied on high-quality, human-annotated datasets such as COCO [26] (110K images) and Visual Genome [20] (100K). As model sizes continue to increase, pre-training requires even more data than before [26, 17, 49], resulting in an extremely high computational cost. However, obtaining large and high-quality multi-modality data is challenging due to the difficulties in annotation. In this paper, we aim to democratize VLP research by proposing a general compression method for existing VLP data. + +# 3. Method + +Our $TL;DR$ is a simple yet effective approach for compressing the Vision-Language Pre-training dataset, leading to further reduction of the training cost. Our approach consists of two stages: (1) codebook-based captioner training and (2) data reduction including samples selection and caption refining. Figure 2 illustrates the idea, introduced next. + +# 3.1. Codebook-based Captioner + +The captioner consists of a visual encoder, a codebook and a text decoder. The visual encoder is employed to extract image features. Inspired by vector quantisation technique [56, 45], we try to quantize the image feature for further clustering by utilizing a learnable codebook. Codebook comprises $K$ learnable embedding vectors, each of which can be regarded as a code. Each token of image features conducts a nearest neighbor look-up from codebook and finds its corresponding code. In this way, image features are quantized into a couple of codes (quantized vectors). The quantized vectors are then sent into a text decoder, generating a caption. In order to enhance the quality of text generation, we initialize the codebook with the text embedding of $K$ most frequently occurring keywords/keyphrases in the entire dataset, which enables the codebook to contain meaningful and intuitively understandable semantics. + +To train the whole captioner, we utilize a Language Modeling loss [11], which maximizes the likelihood of the text in an autoregressive manner, and a symmetric commitment loss [56], which is specifically designed for codebook. We initially train this captioner on noisy source data and subsequently fine-tune it on smaller-scale datasets, such as COCO [26] and VisualGenome [20]. + +# 3.2. Data Reduction + +Currently, large-scale datasets exist with serious redundancy [41]. Meanwhile, a large part of texts is noisy and misaligned with images in VLP data. See Figure 2 for the example (the caption "You need to think twice before buying a pet as present" does not match the image). To overcome these limitations, we use the learned codebook to condense large-scale noisy data and the learned captioner to reduce the misalignment over image-text pairs. + +Samples selection. For an encoded image feature with $L$ tokens, we compute an index vector with length $L$ . Each value is the index of the code, which is the closest to each token. This vector maps the features from image space to semantic space so that it reduces the complexity of the image, benefiting and accelerating the cluster process. Subsequently, each image sample in the dataset is equipped with an index vector according to the above process and we cluster these vectors into $N$ clusters with K-Means (speed up by Faiss [18]). Then we uniformly sample $M\%$ data points from each cluster, producing a small subset of the dataset. We examine various sampling methods and observe that uniform sampling is stable across different scales. + +Caption refining. To alleviate the misalignment problem, we want to improve the text quality using the generated caption. Generated text $T_{g}$ is from the text decoder, which takes the quantized vector of the image as input. We simply concatenate $T_{g}$ with original text $T_{o}$ together, denoted as $T = T_{o} + T_{g}$ , to refine and preserve the original caption's uniqueness while maintaining data diversity. + +The compressed small-scale dataset with refined cap + +
samplingrefiningTR@1IR@1
65.349.8
✓68.551.9
✓69.452.3
✓✓72.854.8
+ +(a) Component ablation. Both the sampling and refining operation are important to the downstream retrieval. + +
caseTR@1IR@1
Image Embedding70.652.3
Text Embedding69.050.4
BLIP Image Embedding [22]72.354.5
Codebook72.854.8
+ +(d) Clustering feature. Codebook is better than Image Embedding at same scale. + +
caseTR@1IR@1
gradient-based72.954.8
hard-sample [41]73.154.5
uniform72.854.8
large distance72.353.1
+ +(b) Sample-selection strategy. The different way to select samples. + +
caseTR@1IR@1
full-scale baseline70.654.0
10%68.952.3
25%72.854.8
50%74.855.2
100%75.157.7
+ +(e) Sampling ratio. Sampling $25\%$ data is enough to beats with full scale. + +
caseTR@1IR@1
xavier [14]72.054.1
key words/phrases72.854.8
object tags72.554.4
+ +(c) Codebook Initialization. An codebook initialized with keywords is more stable. + +
caseTR@1IR@1
10071.853.9
100072.454.5
300072.854.8
500072.954.4
1000072.354.2
+ +(f) Cluster Number. More clusters not equals better result. + +Table 2: TL;DR ablation experiments with BLIP model [22] on CC3M. We report image-to-text retrieval top-1 (TR@1) and text-to-image retrieval top-1 (IR@1) accuracy (\%) on COCO [26] dataset. If not specified, the default baseline is: pretraining BLIP model based on ViT-B/16 with $25\%$ sample of CC3M. Default settings are marked in gray. + +tions is recorded as dataset $D_{c}$ . At last, we train VLP models on this high quality dataset $D_{c}$ and expect the model to achieve comparable performance with original full-scale dataset $D$ on downstream Vision-Language tasks. + +Discussion. Considering the serious misalignment problem, it seems quite straightforward to use pure generated high-quality caption $T_{g}$ to replace original noisy text. Driven by this idea, we try to pretrain BLIP [22] models with $T_{o}, T_{g}$ and $T_{o} + T_{g}$ independently and show the train curve of Image-Text Contrastive (ITC) loss in Figure 3. However, we find the model trained with $T_{g}$ fails into model collapse [36]. This phenomenon can be explained by captioning collapse [46, 47] and one-to-many problem [55] in image cponing. That is, the trained captioner will generate fixed or similar captions for different images, which limits diversity in the output and easily leads to trivial solutions for contrastive loss. On the contrary, the ITC loss for both $T_{o}$ and $T_{o} + T_{g}$ works well and the $T_{o} + T_{g}$ converges better. We also observe the loss of $T_{g}$ is smaller than other two variants at epoch 0-2, which indicates the generated caption matches well with the image. Note that this simple stitching operation on caption does not bring additional computation cost for VLP as the max length in text-encoder Bert [11] model keeps unchanged for all setting. + +# 3.3. Technical Details. + +Our $TL;DR$ can be implemented efficiently, and importantly, does not require any large auxiliary model. The codebook size $K$ is 3000 as default. The selection of keywords/phrases is implemented using the NLTK $^2$ . We adopt ViT-B/16 [12] as image encoder and BertLMHead Model [11] as text decoder. In this way, the token length $L$ + +![](images/3b2d50dd45fe00565bc4320af1f574e73c1c93d71130509bf68eafe1613810b2.jpg) +Figure 3: Training curve with CC3M dataset. Simply stitching generated text and original text together solved the model collapse problem in Image-text Contrastive Loss. + +![](images/0e099d21e14963a74f9f3ada66cba5e119877e8b25f64fe54e2867eacd9202dd.jpg) + +![](images/1f228c7b4840f080049a7b4b322e4b0d41c83677aef8ce3bef64ca0fbe36502d.jpg) + +is 196 as default. The cross-attention is computed over image embedding and text embedding. To show the generality of compressed dataset, we test $D_{c}$ on three different and representative VLP architectures: dual-stream CLIP [34], one-stream ViLT [19] and Fusion-encoder Blip [22] on various downstream tasks. All these models are trained under the same setting with different datasets. + +# 4. CC3M Experiments + +We first study dataset reduction on well-cleaned $CC3M$ [40] which heavily filters web crawled pairs and only keeps $0.1\%$ of the raw data. This dataset contains a total of 2.8 million images. We employ our TL;DR to compress the CC3M dataset, then conduct pre-training and fine-tuning evaluations on both original and compressed datasets. Following our ablation study, we transfer the pre-trained model to seven Vision-Language tasks downstream and fine-tune it through end-to-end training to evaluate its performance. + +Training. We utilize PyTorch [30] to implement our models and trained them on 8 NVIDIA A100 GPUs to reduce the data samples. For Vision-Language Pre-training, we utilize 2 nodes, each equipped with 16 GPUs. The im + +
MethodDataset#SamplesMSCOCO (5K test set)Flickr30K (1K test set)
Image→TextText→ImageImage→TextText→Image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP [34]CC3M [40]2.82M60.485.393.248.975.484.777.391.193.271.690.191.4
TL;DR-CC3M0.67M60.385.693.849.477.486.082.591.892.272.090.592.1
CC3M [40]2.82M36.264.380.129.957.966.967.483.292.454.384.190.8
TL;DR-CC3M0.67M37.764.680.830.758.468.268.585.492.055.682.190.8
ViLT [19]CC3M [40]2.82M66.789.293.852.579.387.183.892.093.274.092.092.8
TL;DR-CC3M0.67M67.188.794.153.178.988.285.392.493.675.692.192.5
CC3M [40]2.82M39.268.677.830.453.266.170.588.792.157.684.992.6
TL;DR-CC3M0.67M43.570.881.433.957.966.873.290.593.358.684.792.4
BLIP [22]CC3M [40]2.82M70.991.396.154.380.288.086.394.194.874.891.692.6
TL;DR-CC3M0.67M72.891.995.954.880.689.487.594.895.375.792.293.4
CC3M [40]2.82M42.367.877.431.555.766.375.191.293.660.685.991.8
TL;DR-CC3M0.67M48.773.182.736.760.670.476.391.993.961.087.793.0
+ +Table 3: Fine-tuning and zero-shot image-text retrieval results on MSCOCO and Flickr30K dataset. + +
Dataset#SamplesVQANLVR2valRefCOCO+COCO Caption
test-devtest-standarddevtest-PtestAtestBB@4CIDEr
Random-CC3M0.67M68.366.273.673.868.671.862.835.9118.8
CC3M [40]2.8M71.571.876.076.272.476.165.336.8121.6
TL;DR-CC3M0.67M73.1+1.673.2+1.477.7+1.778.0+1.875.1+2.778.5+2.468.4+3.137.6+0.8123.8+2.2
+ +age transformer is initialized from ViT pre-trained on ImageNet [10], and the text transformer is initialized from BERTbase [11](BLIP,CLIP) and DistillBERT [38] (ViLT). The model is pre-trained for 20 epochs with a batch size of 1260 and an AdamW [28] optimizer with a weight decay of 0.05. During training, we apply a learning rate warm-up to 3e-4 and a linear decay with a rate of 0.85. For image augmentation, we utilize RandAugment [8] and apply all of the original policies except color inversion. This decision is based on the recognition of the crucial role that color information plays in the data. For pre-training, images were randomly cropped to a resolution of $224 \times 224$ . We then increase this to $384 \times 384$ for fine-tuning downstream tasks. Further information about the training hyperparameters for downstream tasks can be found in the supplementary material. + +# 4.1. Main Properties + +We ablate our $TL;DR$ using the default setting in Table 2 (see caption). Several intriguing properties are observed. + +Module deconstruction. In Table 2a we analyze the impact of different components in $TL;DR$ . We establish a baseline by randomly selecting $25\%$ of the data from CC3M (first row). Our results show that codebook-based sampling outperforms random selection by $3.2\%$ in TR@1. We also observe that both codebook-based sampling and caption refinement are crucial and the combination of them achieves optimal downstream performance. + +Sample selection. In Table 2b we study the sample selection strategy in Stage 2. We sample $25\%$ data in each cluster by default. For Gradient-based, we train a tiny network + +Table 4: Comparison with BLIP model pre-trained on different data sources for VQA, NLVR², RefCOCO+ and COCO Captioning. ViLT and CLIP architectures can not be evaluated on part of these tasks since structural limitations. + +
MethodDatasetR@1↑R@5↑R@10↑MdR↓
CLIP [34]Rand-CC3M15.334.846.313.0
CC3M [40]19.437.347.511.0
TL;DR-CC3M21.838.648.510.0
ViLT [19]Rand-CC3M18.838.249.511.0
CC3M [40]21.040.551.510.0
TL;DR-CC3M22.542.752.48.0
BLIP [22]Rand-CC3M23.342.853.38.0
CC3M [40]26.046.358.07.0
TL;DR-CC3M27.448.759.46.0
+ +Table 5: MSRVTT-1K retrieval using three architectures. We created a subset of the CC3M dataset called Rand-CC3M by randomly selecting the same number of samples as in TL;DR-CC3M. + +to conduct VLP pretrained with ITC [24], ITM [24] and LM [11]. Then we select samples which contribute most to the gradients in each cluster. Large distance: Another perspective is that data points on the border of each cluster are more important than those at the center [4]. So we first compute the center of each cluster and then choose the sample that has the largest distance from the center of each cluster. We also report the result of hard-sample selection from [41]. We observe that all these variants produce similar results except large distances. This suggests that the clustering step, rather than the selection step, plays a key role in data compression during Stage 2. To maintain simplicity, we choose uniform sampling as the default method. + +Codebook initialization. In Table 2c we compare different initialization strategies. The xavier means all parameters in the codebook are initialized with xavier initialization [14]. For the object tags initialization, following previous works [2, 60], we use the 1600 object tags from + +
ModelDataset#SamplesImNetImNet-AImNet-R
CLIP [34]Rand-CC3M0.67M58.361.862.3
CC3M [40]2.82M62.265.266.9
TL;DR-CC3M0.67M61.465.065.7
ViLT [19]Rand-CC3M0.67M54.359.858.4
CC3M [40]2.82M58.662.964.2
TL;DR-CC3M0.67M59.163.364.0
BLIP [22]Rand-CC3M0.67M57.361.865.2
CC3M [40]2.82M62.565.568.1
TL;DR-CC3M0.67M62.063.967.4
+ +Table 6: Zero-shot image classification results on ImageNet [10], ImageNet-A [16], ImageNet-R [15]. There is no free lunch, as selecting partial samples reduces the visual diversity crucial for classification. Despite this, TL;DR still performs significantly better than random selection. + +Visual Genome [20] and extract text feature with a pretrained BERT [11]. With same training setting, the keywords achieve a $0.8\%$ TR@1 improvement and a $0.7\%$ IR@1 improvement over xavier. This result is expected as the text embeddings provide contextual information and simplify the learning process. + +Codebook vs. Image embedding. In Table 2d, we investigate different ways of cluster sampling. First, we remove the codebook from Stage-1 and use image embedding instead. Alternatively, we directly cluster images using the image embedding [22] of images from BLIP model (pretrained on 200M Image-text pairs). We observe the image embedding leads to much better result than text embedding. This is reasonable because clustering visual-similarity samples with text only is difficult. We observe that clustering depended on our codebook performs better than both image embedding and text embedding. This demonstrates that our codebook can efficiently project image embedding to semantic space, benefiting cluster process. + +Cluster sampling ratio. Table 2e varies the sampling ratio of each cluster from $10\%$ to $100\%$ . We are surprised to find that a low sampling ratio can still produce effective results. With only $25\%$ of the data and the TL;DR model, we are able to achieve a $1.9\%$ improvement on TR@1 and a $0.8\%$ improvement on IR@1 over the full-scale baseline. Additionally, we observe that larger sampling ratios lead to even better results. Since our focus is on achieving similar transfer learning results with fewer samples, we use a default sampling ratio of $25\%$ to minimize computation costs. + +Cluster numbers. In Table 2f, we investigate the impact of cluster number on Stage 2 by increasing it from 300 to $30\mathrm{K}$ . We observe that using more clusters results in a slight improvement at the beginning and becomes stable when the number of clusters exceeds $3\mathrm{K}$ . Moreover, all results consistently outperform the random selection baseline. Therefore, we use $3\mathrm{K}$ clusters as the default in this work, as it performs well on fine-tuning tasks. + +![](images/fb4d08a74d726e2c19aab1d2bb512d5f5fd9f94a890950993bd85c7176195275.jpg) +Captioner: A racer jumps over a railing on a horse. + +![](images/d61092376a66fc5dcb355e7359d596571283f1342776441317f33c10fadf4f50.jpg) +Figure 4: The generated caption match the image well. + +![](images/29efb37dd3e58a3e611286fdf1fa5cb2bc7d2d1b560069d6be6c54e10b5fb60e.jpg) + +# 4.2. Transfer Learning Experiments. + +We conduct an extensive evaluation of transfer learning in downstream tasks using the model pre-trained on our compressed TL;DR-CC3M and source CC3M with 3 architectures. Our evaluation primarily focuses on the core tasks of three categories that examine: (1) cross-modality alignment, (2) image captioning and multi-modality understanding capabilities, and (3) visual recognition. The baseline in this section is the model trained on CC3M dataset. + +# 4.2.1 Cross-modality Alignment Task + +Image-Text retrieval. Fine-grained world region alignment plays a critical role in this task. We report both image-to-text retrieval (TR) and text-to-image retrieval (IR) on the COCO [26] and Flickr30K [32] benchmarks. For the BLIP [22] model, we adopt an additional re-ranking strategy, following the original implementation. In Table 3, we also report zero-shot retrieval results. We found that $TL; DR$ achieves comparable results with the baselines on all metrics and surprisingly performs quite well on zero-shot results. For example, for the BLIP [22] architecture, our method leads to a $6.4\%$ improvement (from $42.3\%$ to $48.7\%$ ) in Recall@1 of image-to-text retrieval on MSCOCO. All results suggest that a small part of refined image-text pairs is enough to learn good alignment. + +Zero-shot video retrieval. In this experiment, we analyze the generalization ability of our method to video-language tasks. Specifically, we perform zero-shot transfer to text-to-video retrieval and evaluate the models trained on COCO-retrieval in Table 5. We uniformly sample 8 frames from each video to process the video input and concatenate the frame features into a single sequence. These models trained on our compressed dataset outperform the baseline on all metrics, demonstrating the generality of $TL; DR$ . + +![](images/a636e3c42ab7ea13a7115b8c9c644393d168fab2806dca47ae6f579d805f5f92.jpg) +Figure 5: The codebook-based clusters visualization. The samples within each cluster exhibit similar contextual characteristics, as opposed to mere visual appearance. For example, the "Christmas elements" cluster located at the right. + +![](images/bdc5d88023c00c6a8dc1bcc8abd2de0ca3060f21f09dc14156d57c309caf8473.jpg) + +![](images/441946df53e490025b215e92156544169bc701c92f569e6cb98ca673ddb9b94f.jpg) + +![](images/ffdb52a15d2ef9e150e33b0e9ae0a75d2631b936484146dd24102825cd515e8a.jpg) +VQ-GAN: 5.13s + +![](images/97cbdbff14d12199f45ee15a7bbef5f1ad326c7625f9bc4aa6995d35b5c3e844.jpg) +DALL E2: 2.65s + +![](images/57a1d2243c7355a484af46cdfc1cc5dbf61468cc608f54254b55c89d806d9aed.jpg) +Stable Diffusion: 0.17s +Figure 6: Image generation result with strong Text-to-image Model. The generation time is also reported. + +# 4.2.2 Image Captioning and Multi-modality Understanding Tasks + +Image captioning. The task involves describing an input image, which we evaluate using No-Caps and COCO datasets. Both datasets are fine-tuned on COCO with the Language Modeling (LM) loss. We adopt a zero-shot setting for No-Caps dataset, and start each caption with the phrase "a picture of" for the BLIP architecture. We do not pre-train using COCO to avoid information leakage. Our results outperform baseline with a much smaller quantity of pre-training data, as shown in Table 4. + +Visual question answering (VQA). We evaluate our model's performance on the VQA task [3], where the model needs to provide an answer based on an image and a question. We consider it as an answer generation task that allows open-vocabulary VQA for better results, following previous works [23, 22]. The results are presented in Table 4. The BLIP trained on TL;DR-CC3M outperforms baseline by $1.4\%$ on test-dev splits, demonstrating the effectiveness of our compressed dataset for improving VQA performance. + +Visual reasoning. The Natural Language Visual Reasoning $(\mathrm{NLVR}^2)$ [42] task is a binary classification task that requires the model to reason about two images and a question in natural language. Multi-modal reasoning is crucial for the completion of this task. We observe that BLIP trained on our dataset achieved $78.0\%$ accuracy compared to $76.2\%$ achieved by the CC3M, as shown in Table 4. + +Cross-modality grounding. Referring Expression (RE) Comprehension requires the model to select the target object from a set of image regions proposals, based on the query description. This task heavily relies on visual-grounding ability. The models are evaluated on ground-truth objects, and we evaluate RE Comprehension on RefCOCO+ [58]. The results are reported in Table 4, and we observe that TL;DR-CC3M achieves better results. + +# 4.2.3 Visual Recognition Tasks + +Besides the cross-modality task, we also explore a unimodality task, mainly image classification. Specifically, we fix the image encoder and explore zero-shot image classification. Since the vision encoder is loaded from + +
MethodTR@1IR@1
real data58.344.0
VQ-GAN [13]35.232.4
DALLE2 [35] (implement from 3)44.338.3
Stable Diffusion [37] (implement from 4)52.440.7
+ +Table 7: Compare different sample generation methods over 0.3M subset of CC3M. We first pre-train BLIP model on these generated data and then evaluation on COCO. + +pretrained model, this task demonstrates the impact of training a Vision-Language model with noisy image-text pairs. Specifically, it shows how such training affects the well-learned representation derived from a human-crafted dataset. We show the results in Table 6 and observe noisy data leads to significant Catastrophic Forgetting. For example, the CLIP model drops down to only $58.3\%$ accuracy with noisy data training. We also observe our TL;DR shows steady improvement for all architectures over random selection. Unfortunately, the classification performance for TL;DR-CC3M is slightly worse than the full-scale CC3M for the CLIP and BLIP architectures. Both of these architectures have independent image encoders like ViT to extract image embeddings. This indicates that this task heavily relies on visual diversity, which is different from multimodal tasks, and our method reduces the visual diversity potentially. For the ViLT model, this architecture adopts a shared backbone for both visual and text, and we observe the slightly different results. We guess that multi-modality interaction in early-fusion affects the classification result. + +# 4.3. Visualization + +Generated caption visualization. We show the generated caption in Figure 4. It is evident that the original captions can be highly abstract and difficult to match their respective images, even for human observers sometimes. For instance, when the ITM score is as low as 0.04, matching the figure with its corresponding caption becomes arduous. Such challenging cases can potentially harm the cross-modality alignment. In contrast, we observe that the generated captions describe the image very well and sometimes offer helpful complementary information. For example, "bus" and "castle" in the middle example. + +Codebook-based cluster visualization. Figure 5 displays the codebook grouping result achieved with simple KMeans. Clusters are sets of data points with similar characteristics, often defined by their features or attributes. Interestingly, we observe that the model cluster samples "accurate", meaning that these samples have semantic similarity rather than simple appearance. For instance, the model classifies "dollars" and "piggy bank" together, even though they differ significantly in appearance. + +
Dataset#Samples TimeVQA test-devNLVR2 test-PRefCOCO valNocaps Captioning B@4CIDErFlcikr30K Retrieval TR@1IR@1Imagenet Acc
Rand-CC12M2.4M14h71.876.272.536.8121.082.973.361.2
CC12M [7]10.8M65h73.578.974.137.5122.984.775.365.3
TL;DR-CC12M2.4M14h74.1+0.678.5-0.474.0-0.138.1+0.6124.1+1.285.5+0.876.3+1.063.8-1.5
Rand-YFCC15M2.5M15h67.270.568.135.2116.378.870.565.4
YFCC15M [43]15M90h70.574.270.635.9118.481.572.467.8
TL;DR-YFCC15M2.5M15h70.3-0.275.3+1.172.6+2.037.2+1.3122.5+4.182.3+0.874.3+1.967.3-0.5
Rand-LAION40M(128)8M48h70.775.373.434.8113.280.472.568.5
LAION40M(128) [39]40M120h74.579.176.635.2117.483.274.971.3
TL;DR-LAION40M(128)8M48h76.3+1.880.5+1.477.4+0.836.8+1.6120.9+3.582.8-0.476.1+1.270.4-0.9
+ +Table 8: Comparison with different source of data on 6 downstream tasks. BLIP [22] is adopted as baseline and (128) means the image resolution is ${128} \times {128}$ . We also list the pre-training time,which can be significantly reduced via ${TL};{DR}$ . + +![](images/9bda6184e9448fb0fd18b91f69cf5f7fb0592e7edbde4e83215fd5cdbe75508d.jpg) +Figure 7: ITM score distribution. $TL; DR$ alleviates the issue of misalignment in VLP data. + +![](images/6b57301f7f7960dbca96e636febef744721af53ddbc7346350b7bc1bcbc87430.jpg) + +# 4.4. More Investigation + +Is image generation possible? To ease the misalignment problem of image-text pairs, instead of simply selecting representative samples, a potential and naive idea is to generate images from text. To this end, we randomly sample 0.3M subset of CC3M and generate image from text with three popular text to image generation models, VQ-GAN [56], DALLE 2 [35] and Stable Diffusion [37]. We display the generated samples in Figure 6. We observe that the generative models struggle with complex scenarios, but are capable of generating simple prompts like "dog" proficiently. In addition, generation methods only produce visual cues in a fixed vocabulary, potentially reducing data diversity. + +Next, we pre-train BLIP models on these generated data and evaluate it on COCO Retrieval. In Table 7 we observe the results of transfer learning depend on the quality of generated samples, with those generated by stable diffusion being particularly effective. However, there still exists a significant gap between the generated data and the real dataset (e.g., $52.4\%$ vs. $58.3\%$ on TR@1). We believe that higher-quality and diverse generated images may lead to comparable results with real images in the near future. + +Explore the misalignment problem. Figure 7 shows the Image-text Matching (ITM) score distribution for both $CC3M$ and our $TL;DR-CC3M$ data (the visualization about more datasets is reported in the supplementary). We observe a lot of samples of original $CC3M$ at low matching score even tends to zero, which indicates the current dataset has serious misalignment problems. Since image + +text matching (ITM) loss and image-text contrastive (ITC) loss are used in all architectures, these samples will damage the multimodal representation learning. When adopting our $TL;DR$ , we observe that the matching score tends to be higher and has very few samples with low ITM score. + +# 5. Transfer to other VLP datasets + +We study data compression performed in two categories shown below: clean data that involves human-based offline filter pipelines and raw data that has not undergone cleaning. For clean data, in addition to $CC3M$ , we explore the well-cleaned, high-quality dataset $CC12M$ [7]. Then, we study the raw data YFCC100M [43] and LAION400M [39]. $CC12M$ [7] contains 12 million image-text pairs specifically meant to be used for vision-and-language pre-training. These data are collected by relaxing the data collection pipeline as in CC3M. YFCC15M [34] is a subset of the multilingual and noisy YFCC100M [43] that contains English captions. LAION400M [39] is a large-scale noisy dataset that provides URLs with captions for download. To control the computation cost and reduce the storage overhead, we randomly sample a 40M subset of LAION400M and download images at a resolution of $128 \times 128$ . So, we record the dataset as $TL;DR-LAION40M(128)$ , and the performance over downstream tasks could be improved with higher resolution. More exploration about video-text datasets is reported in the supplementary material + +We use BLIP as the default architecture and evaluate our $TL;DR$ on different datasets and show the results in Table 8. Surprisingly, with only $2.5\mathrm{M}$ $(16.7\%)$ data, $TL;DR$ -YFCC15M leads to similar results with $15\mathrm{M}$ raw data over all metrics except Imagenet. More results with different backbones are reported in the supplementary material. For LAION40M(128), when using $8\mathrm{M}$ data $(20\%)$ , the model trained on our dataset consistently outperforms the baseline method on six downstream tasks. We noticed that the compression rate of LAION40M(128) is less than that of YFCC15M. This may be due to the fact that the collection of LAION40M(128) has already been filtered with CLIP similarity, reducing the impact of the misalignment problem. + +# 6. Conclusion and Discussion + +This paper presents $TL;DR$ , a novel and pioneering algorithm for selecting and generating high-quality image-text pairs from noisy Vision-Language Pre-training (VLP) data, thereby contributing to the field of VLP. $TL;DR$ incorporates a text generation process into learning to reduce serious misalignment problems. Our experiments demonstrate three widely-used architectures leads to comparable results and much smaller training cost when learning from our generated dataset. Additionally, we demonstrate that the misalignment problem can be effectively addressed using our simple $TL;DR$ . However, the choice of the highest compression ratio is done manually rather than learned. Furthermore, achieving even higher compression ratios for VLP models remains a challenge, and text-to-image generation models may be helpful in this regard. We hope that this perspective will inspire future research. + +# Acknowledgement + +This project is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008, and Meta, Mike Zheng Shou's Start-Up Grant from NUS. The computational work for this article was partially performed on resources of the National Supercomputing Centre, Singapore. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. +[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018. +[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. +[4] Dibya Jyoti Bora, Dr Gupta, and Anil Kumar. A comparative study between fuzzy clustering algorithm and hard clustering algorithm. arXiv preprint arXiv:1404.6059, 2014. +[5] Tianshi Cao, Sasha Alexandre Doubov, David Acuna, and Sanja Fidler. Scalable neural data server: A data recommender for transfer learning. Advances in Neural Information Processing Systems, 34:8984-8997, 2021. +[6] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Proceedings of the + +IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4750-4759, 2022. +[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. +[8] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. +[9] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation to imagenet-1k with constant memory. arXiv preprint arXiv:2211.10586, 2022. +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. +[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. +[14] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249-256. JMLR Workshop and Conference Proceedings, 2010. +[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. +[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. +[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. +[18] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535-547, 2019. + +[19] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. +[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017. +[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. +[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. +[23] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. +[24] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020. +[25] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022. +[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. +[27] Yiqi Lin, Huabin Zheng, Huaping Zhong, Jinjing Zhu, Wei-jia Li, Conghui He, and Lin Wang. Sept: Towards scalable and efficient visual pre-training. AAAI, 2023. +[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. +[29] Rafid Mahmood, James Lucas, David Acuna, Daiqing Li, Jonah Philion, Jose M Alvarez, Zhiding Yu, Sanja Fidler, and Marc T Law. How much more data do i need? estimating requirements for downstream tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 275-284, 2022. +[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. +[31] Mansheej Paul, Surya Ganguli, and Gintare Karolina Dziugaite. Deep learning on a data diet: Finding important examples early in training. Advances in Neural Information Processing Systems, 34:20596-20607, 2021. + +[32] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. +[33] Ziheng Qin, Kai Wang, Zangwei Zheng, Jianyang Gu, Xi-angyu Peng, Daquan Zhou, and Yang You. Infobatch: Lossless training speed up by unbiased dynamic data pruning. arXiv preprint arXiv:2303.04947, 2023. +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. +[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. +[36] Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, and Suvrit Sra. Can contrastive learning avoid shortcut solutions? Advances in neural information processing systems, 34:4974-4986, 2021. +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. +[38] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. +[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. +[40] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. +[41] Ben Sorscher, Robert Geirhos, Shashank Shekhar, Surya Ganguli, and Ari S Morcos. Beyond neural scaling laws: beating power law scaling via data pruning. NeurIPS, 2022. +[42] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Hua-jun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. arXiv preprint arXiv:1811.00491, 2018. +[43] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. +[44] Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J Gordon. An empirical study of example forget- + +ting during deep neural network learning. arXiv preprint arXiv:1812.05159, 2018. +[45] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. +[46] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2015. +[47] Haoran Wang, Yue Zhang, Xiaosheng Yu, et al. An overview of image caption generation methods. Computational intelligence and neuroscience, 2020, 2020. +[48] Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Kevin Qinghong Lin, Satoshi Tsutsui, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, et al. All in one: Exploring unified video-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6598-6608, 2023. +[49] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. +[50] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12196-12205, 2022. +[51] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. +[52] Hu Xu, Saining Xie, Po-Yao Huang, Licheng Yu, Russell Howes, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Cit: Curation in training for effective vision-language data. arXiv preprint arXiv:2301.02241, 2023. +[53] Xi Yan, David Acuna, and Sanja Fidler. Neural data server: A large-scale search engine for transfer learning data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3893-3902, 2020. +[54] Xingcheng Yao, Yanan Zheng, Xiaocoong Yang, and Zhilin Yang. Nlp from scratch without large-scale pretraining: A simple and efficient framework. In International Conference on Machine Learning, pages 25438-25451. PMLR, 2022. +[55] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. +[56] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. ICLR, 2022. +[57] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. + +[58] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. +[59] Ruonan Yu, Songhua Liu, and Xinchao Wang. Dataset distillation: A comprehensive review. arXiv preprint arXiv:2301.07014, 2023. +[60] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. 2021. +[61] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. arXiv preprint arXiv:2006.05929, 2020. \ No newline at end of file diff --git a/2023/Too Large; Data Reduction for Vision-Language Pre-Training/images.zip b/2023/Too Large; Data Reduction for Vision-Language Pre-Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..169abcaa94b2c2198f11fe33f5380f292e8ca092 --- /dev/null +++ b/2023/Too Large; Data Reduction for Vision-Language Pre-Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07596918bec0b306b759df0c65544b57af6c98fd5f0bd76994df85815897409a +size 540059 diff --git a/2023/Too Large; Data Reduction for Vision-Language Pre-Training/layout.json b/2023/Too Large; Data Reduction for Vision-Language Pre-Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fcdcce435cfbd989a5481ee59e44369df8ccd11b --- /dev/null +++ b/2023/Too Large; Data Reduction for Vision-Language Pre-Training/layout.json @@ -0,0 +1,9810 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 102, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 102, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 97, + 102, + 496, + 121 + ], + "type": "text", + "content": "Too Large; Data Reduction for Vision-Language Pre-Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 143, + 452, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 143, + 452, + 190 + ], + "spans": [ + { + "bbox": [ + 141, + 143, + 452, + 190 + ], + "type": "text", + "content": "Alex Jinpeng Wang, Kevin Qinghong Lin, David Junhao Zhang, Stan Weixian Lei, Mike Zheng Shou* Show Lab, National University of Singapore" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 216, + 192, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 216, + 192, + 228 + ], + "spans": [ + { + "bbox": [ + 143, + 216, + 192, + 228 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 242, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 242, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 290, + 529 + ], + "type": "text", + "content": "This paper examines the problems of severe image-text misalignment and high redundancy in the widely-used large-scale Vision-Language Pre-Training (VLP) datasets. To address these issues, we propose an efficient and straightforward Vision-Language learning algorithm called TL;DR, which aims to compress the existing large VLP data into a small, high-quality set. Our approach consists of two major steps. First, a codebook-based encoder-decoder captioner is developed to select representative samples. Second, a new caption is generated to complement the original captions for selected samples, mitigating the text-image misalignment problem while maintaining uniqueness. As the result, TL;DR enables us to reduce the large dataset into a small set of high-quality data, which can serve as an alternative pre-training dataset. This algorithm significantly speeds up the time-consuming pretraining process. Specifically, TL;DR can compress the mainstream VLP datasets at a high ratio, e.g., reduce well-cleaned CC3M dataset from 2.82M to 0.67M (" + }, + { + "bbox": [ + 47, + 242, + 290, + 529 + ], + "type": "inline_equation", + "content": "\\sim 24\\%" + }, + { + "bbox": [ + 47, + 242, + 290, + 529 + ], + "type": "text", + "content": ") and noisy YFCC15M from 15M to 2.5M (" + }, + { + "bbox": [ + 47, + 242, + 290, + 529 + ], + "type": "inline_equation", + "content": "\\sim 16.7\\%" + }, + { + "bbox": [ + 47, + 242, + 290, + 529 + ], + "type": "text", + "content": "). Extensive experiments with three popular VLP models over seven downstream tasks show that VLP model trained on the compressed dataset provided by TL;DR can perform similar or even better results compared with training on the full-scale dataset1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 554, + 128, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 128, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 128, + 566 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 574, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 684 + ], + "type": "text", + "content": "The recent \"scale-is-everything\" viewpoint has become a widely accepted notion in the Vision-language Pre-training (VLP) community [40, 7, 34, 17, 1]. According to this view, the scale of the data has increased from the original tens of thousands-level (e.g., COCO [26] and VG [20]) to millions-level (e.g., CC3M [40] and CC12M [7]), and even up to billions-level (e.g., YFCC100M [43], WIT400M [34], and LAION400M [39]). Approaches [57, 34, 17] trained on these large-scale data show remarkable performance im" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 215, + 422, + 312 + ], + "blocks": [ + { + "bbox": [ + 310, + 215, + 422, + 312 + ], + "lines": [ + { + "bbox": [ + 310, + 215, + 422, + 312 + ], + "spans": [ + { + "bbox": [ + 310, + 215, + 422, + 312 + ], + "type": "image", + "image_path": "c884edc4e9f990ead5d34f18305048dcfaea13cd7fa2361ca38e50714552c154.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 314, + 547, + 411 + ], + "lines": [ + { + "bbox": [ + 304, + 314, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 547, + 411 + ], + "type": "text", + "content": "Figure 1: Does using more data really lead to better performance in VLP? Instead of training on the full-scale CC3M dataset, we delete data with low image-text matching score. We find that BLIP [22] model pretrained on " + }, + { + "bbox": [ + 304, + 314, + 547, + 411 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 314, + 547, + 411 + ], + "type": "text", + "content": " reserved data even obtains better result than full-scale dataset on downstream COCO retrieval [26]. This observation exposes there exists serious misalignment between text&visual modalities and data redundancy in dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 425, + 215, + 538, + 311 + ], + "blocks": [ + { + "bbox": [ + 425, + 215, + 538, + 311 + ], + "lines": [ + { + "bbox": [ + 425, + 215, + 538, + 311 + ], + "spans": [ + { + "bbox": [ + 425, + 215, + 538, + 311 + ], + "type": "image", + "image_path": "2ed4e85d8e0b95abebcd24f052ef1588ad1880b9ba5e8ace5e93115732e73a3e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 423, + 469, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 423, + 469, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 423, + 469, + 434 + ], + "type": "text", + "content": "provement in various downstream tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "content": "However, simply scaling-up data brings two critical challenges: " + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "content": ". Larger image-text datasets lead to more training cost (e.g., Pretraining CoCa takes about 5 days on 2,048 CloudTPUv4 chips [57]) and storage overhead, which is difficult to afford. " + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "inline_equation", + "content": "ii" + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "content": ". Obtaining high-quality VLP data requires massive data and well-designed collecting/filtering pipeline, which is expensive. For instance, the CC3M [40] data was obtained after filtering 5 billion collected images. These challenges are daunting and may impede the participation of numerous researchers in the VLP community." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "In this study, we stop hunting for larger-scale data blindly and ask an important question: Does employing a larger dataset always result in better performance in VLP? To explore and answer this question, we begin with a simple experiment. First, we utilize a pre-trained BLIP [22] model to calculate the Image-Text Matching (ITM) scores for all samples in the clean CC3M dataset. Subsequently, we remove a portion of the samples with the lowest ITM scores and evaluate the transfer learning results, as shown in Figure 1. Surprisingly, discarding " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": " of the samples slightly improves performance. This remarkable finding challenges the prevailing belief that employing larger amounts of data invariably leads to superior VLP outcomes." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 138, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 138, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 138, + 703 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 58, + 703, + 265, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 265, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 265, + 713 + ], + "type": "text", + "content": "1https://github.com/showlab/datacentric.vlp" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3147" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 69, + 545, + 120 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 545, + 120 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 545, + 120 + ], + "type": "table", + "html": "
MethodYearData TypeCompression Ratio↑Task AgnosticLarge-scaleSupervisionGeneration/Selection
Dataset Distillation [51]2017Image99%-99.99%XXClass LabelGeneration
Data Pruning [41]2022Image20%-30%X✓Class LabelSelection
Neural Data Server [53]2020Multi-modality94%-98%X✓Image-text PairsSelection
TL;DR (ours)-Multi-modality75%-90%✓✓Image-text PairsGeneration+Selection
", + "image_path": "b48def29d89400a3a7fedf6117730a6b64628a6a3b557b4368d586fc50df1286.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 122, + 546, + 158 + ], + "lines": [ + { + "bbox": [ + 46, + 122, + 546, + 158 + ], + "spans": [ + { + "bbox": [ + 46, + 122, + 546, + 158 + ], + "type": "text", + "content": "Table 1: Data-efficient learning methods. \"Large-scale\" means that the methods are effective when used on datasets that are very large in size. The \"task agnostic\" means that the methods can be used regardless of the specific downstream task, and without any prior exposure to the associated data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 171, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 290 + ], + "type": "text", + "content": "This experiment suggests removing certain data points can actually improve the model's ability to learn and generalize. Moreover, considering the performance improvements after removing the low ITM score data, we can infer the existence of significant misalignment between the textual and visual modalities in many text-image data pairs (see Figure 7 and the supplementary material for more evidences). These discoveries present promising potentiality to enhance the performance of models that depend on a smaller volume of VLP data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "text", + "content": "Driven by above analysis and recent advance in dataset pruning [41], we present a simple, effective and scalable algorithm called " + }, + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "text", + "content": " that aims to improve data efficiency for visual-language pretraining. The " + }, + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "text", + "content": " has a powerful codebook-based captioner, which contains a visual encoder, a look-up codebook and a text decoder. Here is how it works: First, " + }, + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "text", + "content": " feeds each image into the visual encoder and determines the corresponding codes of the image by measuring the similarity between the codebook and the embedding generated by the encoder. Given a large pool of image-text pairs, " + }, + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "text", + "content": " clusters the samples based on their image corresponding codes and selects a representative subset of samples from each cluster. Then, " + }, + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "text", + "content": " further refines the caption of the selected samples via text decoder to reduce text-image misalignment. By doing so, " + }, + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 46, + 291, + 288, + 495 + ], + "type": "text", + "content": " is able to significantly reduce the size of the training dataset while maintaining the high quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 495, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 627 + ], + "type": "text", + "content": "In this work, we employ " + }, + { + "bbox": [ + 46, + 495, + 287, + 627 + ], + "type": "inline_equation", + "content": "TL; DR" + }, + { + "bbox": [ + 46, + 495, + 287, + 627 + ], + "type": "text", + "content": " on widely-used CC3M, CC12M, YFCC100M and LAION400M datasets and evaluate small size data on three widely-used frameworks including CLIP [34], ViLT [19], and BLIP [22] for data efficiency pretraining with seven representative visual-language downstream tasks. The results show that, with only " + }, + { + "bbox": [ + 46, + 495, + 287, + 627 + ], + "type": "inline_equation", + "content": "10\\% - 25\\%" + }, + { + "bbox": [ + 46, + 495, + 287, + 627 + ], + "type": "text", + "content": " data obtained by " + }, + { + "bbox": [ + 46, + 495, + 287, + 627 + ], + "type": "inline_equation", + "content": "TL; DR" + }, + { + "bbox": [ + 46, + 495, + 287, + 627 + ], + "type": "text", + "content": ", frameworks achieve similar or even better performance compared with the full-scale dataset. We hope our findings can inspire the community to reconsider data efficiency for VLP rather than blindly utilizing increasingly massive datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 639, + 134, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 134, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 134, + 651 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 181, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 181, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 181, + 672 + ], + "type": "text", + "content": "2.1. Data-Efficient Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Recent successes in deep learning are largely attributed to the vast amount of data [10, 34]. However, collecting massive amounts of data is expensive and raises con" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 171, + 545, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 206 + ], + "type": "text", + "content": "cerns about privacy and copyright [59]. As a result, the research community has become increasingly interested in data-efficient learning, which includes:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 207, + 546, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 546, + 361 + ], + "type": "text", + "content": "Dataset Distillation [51, 61, 50, 33] compress a large dataset into a small set of synthetic samples, enabling models trained on the smaller dataset to achieve competitive performance with those trained on the original dataset. However, these techniques are only effective on relatively small datasets at low resolutions, such as CIFAR [21], and their performance deteriorates significantly when applied to larger-scale datasets. For example, the accuracy of a model trained on the state-of-the-art MMT's generated data is only " + }, + { + "bbox": [ + 304, + 207, + 546, + 361 + ], + "type": "inline_equation", + "content": "33.8\\%" + }, + { + "bbox": [ + 304, + 207, + 546, + 361 + ], + "type": "text", + "content": " on the ImageNet-1K [10] test result [6], while pre-training on real ImageNet-1K achieves over " + }, + { + "bbox": [ + 304, + 207, + 546, + 361 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 207, + 546, + 361 + ], + "type": "text", + "content": " accuracy [9]. Furthermore, these methods necessitate supervised class labels, which are not suitable for multimodal data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 363, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 483 + ], + "type": "text", + "content": "Data Pruning [44, 31] assumes high redundancy in large datasets, selecting only a subset of challenging samples. [29, 31] observed that during the entire training process, some examples are learned early and never forgotten, while others can be repeatedly learned and forgotten. The related work [41] uses a hard sample selection method to select " + }, + { + "bbox": [ + 304, + 363, + 545, + 483 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 363, + 545, + 483 + ], + "type": "text", + "content": " samples of the ImageNet dataset, and the model trained on selected samples approximating training on all data. Another recent work, CiT [52], also proposes to train models with dynamic training data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 567 + ], + "type": "text", + "content": "Neural Data Server (NDS) [53, 27, 5] proposes a large-scale search engine to identify the most useful transfer learning data from large corpus. While these methods can be extended to multi-modality data, a similar idea has also been applied in NLP [54]. However, this setting assumes that the user has access to all downstream data and needs to train the downstream task using additional retrieval data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 639 + ], + "type": "text", + "content": "In this work, we are different from previous techniques in that we attempt to compress large-scale multimodal data for the first time, leading to comparable performance between the compressed and original vision-language datasets. We provide a comparison of our approach with these related works in Table 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 647, + 468, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 468, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 468, + 659 + ], + "type": "text", + "content": "2.2. Visual-Language Pre-training" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "Large-scale Vision-Language Pre-training (VLP) involves training on extensive multi-modality data and evaluating performance on various downstream vision-language tasks. Conventional frameworks include the dual-stream" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3148" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 253, + 199 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 253, + 199 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 253, + 199 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 253, + 199 + ], + "type": "image", + "image_path": "3677a06d50e4d08d552894d71c68f2e5555385dbf6d36aa502b373e52afaaa91.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 547, + 240 + ], + "type": "text", + "content": "Figure 2: Our TL;DR architecture. We first train a codebook-based captioner in Stage1. Then the learned codebook and captioner are used to reduce VLP data in Stage 2. Pre-training on the reduced dataset achieves similar performance to the original full-scale dataset across downstream tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 261, + 70, + 525, + 199 + ], + "blocks": [ + { + "bbox": [ + 261, + 70, + 525, + 199 + ], + "lines": [ + { + "bbox": [ + 261, + 70, + 525, + 199 + ], + "spans": [ + { + "bbox": [ + 261, + 70, + 525, + 199 + ], + "type": "image", + "image_path": "8417fe376969c3596a6b7a530f6f0746a205a561eb10606be6bfa6b79ade3aec.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 384 + ], + "type": "text", + "content": "architecture [34], the one-stream architecture [19, 24, 48, 25, 49], and the encoder-decoder architecture [22]. Previous works have relied on high-quality, human-annotated datasets such as COCO [26] (110K images) and Visual Genome [20] (100K). As model sizes continue to increase, pre-training requires even more data than before [26, 17, 49], resulting in an extremely high computational cost. However, obtaining large and high-quality multi-modality data is challenging due to the difficulties in annotation. In this paper, we aim to democratize VLP research by proposing a general compression method for existing VLP data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 398, + 104, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 398, + 104, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 398, + 104, + 410 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 419, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 287, + 491 + ], + "type": "text", + "content": "Our " + }, + { + "bbox": [ + 46, + 419, + 287, + 491 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 46, + 419, + 287, + 491 + ], + "type": "text", + "content": " is a simple yet effective approach for compressing the Vision-Language Pre-training dataset, leading to further reduction of the training cost. Our approach consists of two stages: (1) codebook-based captioner training and (2) data reduction including samples selection and caption refining. Figure 2 illustrates the idea, introduced next." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 502, + 198, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 198, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 198, + 514 + ], + "type": "text", + "content": "3.1. Codebook-based Captioner" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": "The captioner consists of a visual encoder, a codebook and a text decoder. The visual encoder is employed to extract image features. Inspired by vector quantisation technique [56, 45], we try to quantize the image feature for further clustering by utilizing a learnable codebook. Codebook comprises " + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": " learnable embedding vectors, each of which can be regarded as a code. Each token of image features conducts a nearest neighbor look-up from codebook and finds its corresponding code. In this way, image features are quantized into a couple of codes (quantized vectors). The quantized vectors are then sent into a text decoder, generating a caption. In order to enhance the quality of text generation, we initialize the codebook with the text embedding of " + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": " most frequently occurring keywords/keyphrases in the entire dataset, which enables the codebook to contain meaningful and intuitively understandable semantics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 251, + 547, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 547, + 337 + ], + "type": "text", + "content": "To train the whole captioner, we utilize a Language Modeling loss [11], which maximizes the likelihood of the text in an autoregressive manner, and a symmetric commitment loss [56], which is specifically designed for codebook. We initially train this captioner on noisy source data and subsequently fine-tune it on smaller-scale datasets, such as COCO [26] and VisualGenome [20]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 345, + 402, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 345, + 402, + 357 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 402, + 357 + ], + "type": "text", + "content": "3.2. Data Reduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 364, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 460 + ], + "type": "text", + "content": "Currently, large-scale datasets exist with serious redundancy [41]. Meanwhile, a large part of texts is noisy and misaligned with images in VLP data. See Figure 2 for the example (the caption \"You need to think twice before buying a pet as present\" does not match the image). To overcome these limitations, we use the learned codebook to condense large-scale noisy data and the learned captioner to reduce the misalignment over image-text pairs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": "Samples selection. For an encoded image feature with " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": " tokens, we compute an index vector with length " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": ". Each value is the index of the code, which is the closest to each token. This vector maps the features from image space to semantic space so that it reduces the complexity of the image, benefiting and accelerating the cluster process. Subsequently, each image sample in the dataset is equipped with an index vector according to the above process and we cluster these vectors into " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": " clusters with K-Means (speed up by Faiss [18]). Then we uniformly sample " + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "inline_equation", + "content": "M\\%" + }, + { + "bbox": [ + 304, + 460, + 546, + 616 + ], + "type": "text", + "content": " data points from each cluster, producing a small subset of the dataset. We examine various sampling methods and observe that uniform sampling is stable across different scales." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "text", + "content": "Caption refining. To alleviate the misalignment problem, we want to improve the text quality using the generated caption. Generated text " + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "inline_equation", + "content": "T_{g}" + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "text", + "content": " is from the text decoder, which takes the quantized vector of the image as input. We simply concatenate " + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "inline_equation", + "content": "T_{g}" + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "text", + "content": " with original text " + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "inline_equation", + "content": "T_{o}" + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "text", + "content": " together, denoted as " + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "inline_equation", + "content": "T = T_{o} + T_{g}" + }, + { + "bbox": [ + 304, + 616, + 545, + 700 + ], + "type": "text", + "content": ", to refine and preserve the original caption's uniqueness while maintaining data diversity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "The compressed small-scale dataset with refined cap" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3149" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 193, + 123 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 193, + 123 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 193, + 123 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 193, + 123 + ], + "type": "table", + "html": "
samplingrefiningTR@1IR@1
65.349.8
✓68.551.9
✓69.452.3
✓✓72.854.8
", + "image_path": "4fb3ce84ba4e03e8a155d97e9bd177c63c58b1c75b8b3771795359ab2e67f3a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 165, + 217, + 219 + ], + "blocks": [ + { + "bbox": [ + 49, + 128, + 202, + 159 + ], + "lines": [ + { + "bbox": [ + 49, + 128, + 202, + 159 + ], + "spans": [ + { + "bbox": [ + 49, + 128, + 202, + 159 + ], + "type": "text", + "content": "(a) Component ablation. Both the sampling and refining operation are important to the downstream retrieval." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 165, + 217, + 219 + ], + "lines": [ + { + "bbox": [ + 53, + 165, + 217, + 219 + ], + "spans": [ + { + "bbox": [ + 53, + 165, + 217, + 219 + ], + "type": "table", + "html": "
caseTR@1IR@1
Image Embedding70.652.3
Text Embedding69.050.4
BLIP Image Embedding [22]72.354.5
Codebook72.854.8
", + "image_path": "f8625fbc4986755bc0bd74aa0da93931d97548ac490647c6807e93ee9f30a99d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 227, + 70, + 370, + 122 + ], + "blocks": [ + { + "bbox": [ + 49, + 232, + 202, + 254 + ], + "lines": [ + { + "bbox": [ + 49, + 232, + 202, + 254 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 202, + 254 + ], + "type": "text", + "content": "(d) Clustering feature. Codebook is better than Image Embedding at same scale." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 227, + 70, + 370, + 122 + ], + "lines": [ + { + "bbox": [ + 227, + 70, + 370, + 122 + ], + "spans": [ + { + "bbox": [ + 227, + 70, + 370, + 122 + ], + "type": "table", + "html": "
caseTR@1IR@1
gradient-based72.954.8
hard-sample [41]73.154.5
uniform72.854.8
large distance72.353.1
", + "image_path": "fbf9328d46e7e1c444146f452c4d1afaebbd8166e514830aacb319fdcfe8bf3c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 227, + 165, + 370, + 226 + ], + "blocks": [ + { + "bbox": [ + 220, + 127, + 372, + 149 + ], + "lines": [ + { + "bbox": [ + 220, + 127, + 372, + 149 + ], + "spans": [ + { + "bbox": [ + 220, + 127, + 372, + 149 + ], + "type": "text", + "content": "(b) Sample-selection strategy. The different way to select samples." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 227, + 165, + 370, + 226 + ], + "lines": [ + { + "bbox": [ + 227, + 165, + 370, + 226 + ], + "spans": [ + { + "bbox": [ + 227, + 165, + 370, + 226 + ], + "type": "table", + "html": "
caseTR@1IR@1
full-scale baseline70.654.0
10%68.952.3
25%72.854.8
50%74.855.2
100%75.157.7
", + "image_path": "9bbe61beaf812ac9e1d035793872c8a55862c332f88991fe52ca77e890e2cad1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 408, + 70, + 531, + 113 + ], + "blocks": [ + { + "bbox": [ + 220, + 232, + 372, + 254 + ], + "lines": [ + { + "bbox": [ + 220, + 232, + 372, + 254 + ], + "spans": [ + { + "bbox": [ + 220, + 232, + 372, + 254 + ], + "type": "text", + "content": "(e) Sampling ratio. Sampling " + }, + { + "bbox": [ + 220, + 232, + 372, + 254 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 220, + 232, + 372, + 254 + ], + "type": "text", + "content": " data is enough to beats with full scale." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 408, + 70, + 531, + 113 + ], + "lines": [ + { + "bbox": [ + 408, + 70, + 531, + 113 + ], + "spans": [ + { + "bbox": [ + 408, + 70, + 531, + 113 + ], + "type": "table", + "html": "
caseTR@1IR@1
xavier [14]72.054.1
key words/phrases72.854.8
object tags72.554.4
", + "image_path": "61fde20a00c9ad89cc7f12a589923af002de07caad06e33baf332b7cc2818e7a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 408, + 165, + 531, + 226 + ], + "blocks": [ + { + "bbox": [ + 391, + 127, + 543, + 159 + ], + "lines": [ + { + "bbox": [ + 391, + 127, + 543, + 159 + ], + "spans": [ + { + "bbox": [ + 391, + 127, + 543, + 159 + ], + "type": "text", + "content": "(c) Codebook Initialization. An codebook initialized with keywords is more stable." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 408, + 165, + 531, + 226 + ], + "lines": [ + { + "bbox": [ + 408, + 165, + 531, + 226 + ], + "spans": [ + { + "bbox": [ + 408, + 165, + 531, + 226 + ], + "type": "table", + "html": "
caseTR@1IR@1
10071.853.9
100072.454.5
300072.854.8
500072.954.4
1000072.354.2
", + "image_path": "f0683456a15a9c80c4b849547e69e42ad9c21cba3663ea984dc27ba42ef92110.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 391, + 232, + 543, + 254 + ], + "lines": [ + { + "bbox": [ + 391, + 232, + 543, + 254 + ], + "spans": [ + { + "bbox": [ + 391, + 232, + 543, + 254 + ], + "type": "text", + "content": "(f) Cluster Number. More clusters not equals better result." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 259, + 546, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 259, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 546, + 297 + ], + "type": "text", + "content": "Table 2: TL;DR ablation experiments with BLIP model [22] on CC3M. We report image-to-text retrieval top-1 (TR@1) and text-to-image retrieval top-1 (IR@1) accuracy (\\%) on COCO [26] dataset. If not specified, the default baseline is: pretraining BLIP model based on ViT-B/16 with " + }, + { + "bbox": [ + 46, + 259, + 546, + 297 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 259, + 546, + 297 + ], + "type": "text", + "content": " sample of CC3M. Default settings are marked in gray." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 308, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 357 + ], + "type": "text", + "content": "tions is recorded as dataset " + }, + { + "bbox": [ + 46, + 308, + 287, + 357 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 46, + 308, + 287, + 357 + ], + "type": "text", + "content": ". At last, we train VLP models on this high quality dataset " + }, + { + "bbox": [ + 46, + 308, + 287, + 357 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 46, + 308, + 287, + 357 + ], + "type": "text", + "content": " and expect the model to achieve comparable performance with original full-scale dataset " + }, + { + "bbox": [ + 46, + 308, + 287, + 357 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 308, + 287, + 357 + ], + "type": "text", + "content": " on downstream Vision-Language tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "text", + "content": "Discussion. Considering the serious misalignment problem, it seems quite straightforward to use pure generated high-quality caption " + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "inline_equation", + "content": "T_{g}" + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "text", + "content": " to replace original noisy text. Driven by this idea, we try to pretrain BLIP [22] models with " + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "inline_equation", + "content": "T_{o}, T_{g}" + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "inline_equation", + "content": "T_{o} + T_{g}" + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "text", + "content": " independently and show the train curve of Image-Text Contrastive (ITC) loss in Figure 3. However, we find the model trained with " + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "inline_equation", + "content": "T_{g}" + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "text", + "content": " fails into model collapse [36]. This phenomenon can be explained by captioning collapse [46, 47] and one-to-many problem [55] in image cponing. That is, the trained captioner will generate fixed or similar captions for different images, which limits diversity in the output and easily leads to trivial solutions for contrastive loss. On the contrary, the ITC loss for both " + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "inline_equation", + "content": "T_{o}" + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "inline_equation", + "content": "T_{o} + T_{g}" + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "text", + "content": " works well and the " + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "inline_equation", + "content": "T_{o} + T_{g}" + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "text", + "content": " converges better. We also observe the loss of " + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "inline_equation", + "content": "T_{g}" + }, + { + "bbox": [ + 46, + 357, + 287, + 596 + ], + "type": "text", + "content": " is smaller than other two variants at epoch 0-2, which indicates the generated caption matches well with the image. Note that this simple stitching operation on caption does not bring additional computation cost for VLP as the max length in text-encoder Bert [11] model keeps unchanged for all setting." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 604, + 152, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 152, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 152, + 616 + ], + "type": "text", + "content": "3.3. Technical Details." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 622, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 622, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 622, + 287, + 696 + ], + "type": "text", + "content": "Our " + }, + { + "bbox": [ + 46, + 622, + 287, + 696 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 46, + 622, + 287, + 696 + ], + "type": "text", + "content": " can be implemented efficiently, and importantly, does not require any large auxiliary model. The codebook size " + }, + { + "bbox": [ + 46, + 622, + 287, + 696 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 622, + 287, + 696 + ], + "type": "text", + "content": " is 3000 as default. The selection of keywords/phrases is implemented using the NLTK" + }, + { + "bbox": [ + 46, + 622, + 287, + 696 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 622, + 287, + 696 + ], + "type": "text", + "content": ". We adopt ViT-B/16 [12] as image encoder and BertLMHead Model [11] as text decoder. In this way, the token length " + }, + { + "bbox": [ + 46, + 622, + 287, + 696 + ], + "type": "inline_equation", + "content": "L" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 306, + 388, + 385 + ], + "blocks": [ + { + "bbox": [ + 307, + 306, + 388, + 385 + ], + "lines": [ + { + "bbox": [ + 307, + 306, + 388, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 306, + 388, + 385 + ], + "type": "image", + "image_path": "3b2d50dd45fe00565bc4320af1f574e73c1c93d71130509bf68eafe1613810b2.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 390, + 545, + 427 + ], + "lines": [ + { + "bbox": [ + 305, + 390, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 305, + 390, + 545, + 427 + ], + "type": "text", + "content": "Figure 3: Training curve with CC3M dataset. Simply stitching generated text and original text together solved the model collapse problem in Image-text Contrastive Loss." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 391, + 306, + 465, + 384 + ], + "blocks": [ + { + "bbox": [ + 391, + 306, + 465, + 384 + ], + "lines": [ + { + "bbox": [ + 391, + 306, + 465, + 384 + ], + "spans": [ + { + "bbox": [ + 391, + 306, + 465, + 384 + ], + "type": "image", + "image_path": "0e099d21e14963a74f9f3ada66cba5e119877e8b25f64fe54e2867eacd9202dd.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 468, + 306, + 544, + 384 + ], + "blocks": [ + { + "bbox": [ + 468, + 306, + 544, + 384 + ], + "lines": [ + { + "bbox": [ + 468, + 306, + 544, + 384 + ], + "spans": [ + { + "bbox": [ + 468, + 306, + 544, + 384 + ], + "type": "image", + "image_path": "1f228c7b4840f080049a7b4b322e4b0d41c83677aef8ce3bef64ca0fbe36502d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 437, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 521 + ], + "type": "text", + "content": "is 196 as default. The cross-attention is computed over image embedding and text embedding. To show the generality of compressed dataset, we test " + }, + { + "bbox": [ + 304, + 437, + 546, + 521 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 304, + 437, + 546, + 521 + ], + "type": "text", + "content": " on three different and representative VLP architectures: dual-stream CLIP [34], one-stream ViLT [19] and Fusion-encoder Blip [22] on various downstream tasks. All these models are trained under the same setting with different datasets." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 533, + 424, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 424, + 547 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 424, + 547 + ], + "type": "text", + "content": "4. CC3M Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 554, + 546, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 546, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 546, + 662 + ], + "type": "text", + "content": "We first study dataset reduction on well-cleaned " + }, + { + "bbox": [ + 304, + 554, + 546, + 662 + ], + "type": "inline_equation", + "content": "CC3M" + }, + { + "bbox": [ + 304, + 554, + 546, + 662 + ], + "type": "text", + "content": " [40] which heavily filters web crawled pairs and only keeps " + }, + { + "bbox": [ + 304, + 554, + 546, + 662 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 304, + 554, + 546, + 662 + ], + "type": "text", + "content": " of the raw data. This dataset contains a total of 2.8 million images. We employ our TL;DR to compress the CC3M dataset, then conduct pre-training and fine-tuning evaluations on both original and compressed datasets. Following our ablation study, we transfer the pre-trained model to seven Vision-Language tasks downstream and fine-tune it through end-to-end training to evaluate its performance." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Training. We utilize PyTorch [30] to implement our models and trained them on 8 NVIDIA A100 GPUs to reduce the data samples. For Vision-Language Pre-training, we utilize 2 nodes, each equipped with 16 GPUs. The im" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 152, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 152, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 152, + 713 + ], + "type": "text", + "content": "2https://github.com/pltk/pltk" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3150" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 68, + 549, + 216 + ], + "blocks": [ + { + "bbox": [ + 48, + 68, + 549, + 216 + ], + "lines": [ + { + "bbox": [ + 48, + 68, + 549, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 68, + 549, + 216 + ], + "type": "table", + "html": "
MethodDataset#SamplesMSCOCO (5K test set)Flickr30K (1K test set)
Image→TextText→ImageImage→TextText→Image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP [34]CC3M [40]2.82M60.485.393.248.975.484.777.391.193.271.690.191.4
TL;DR-CC3M0.67M60.385.693.849.477.486.082.591.892.272.090.592.1
CC3M [40]2.82M36.264.380.129.957.966.967.483.292.454.384.190.8
TL;DR-CC3M0.67M37.764.680.830.758.468.268.585.492.055.682.190.8
ViLT [19]CC3M [40]2.82M66.789.293.852.579.387.183.892.093.274.092.092.8
TL;DR-CC3M0.67M67.188.794.153.178.988.285.392.493.675.692.192.5
CC3M [40]2.82M39.268.677.830.453.266.170.588.792.157.684.992.6
TL;DR-CC3M0.67M43.570.881.433.957.966.873.290.593.358.684.792.4
BLIP [22]CC3M [40]2.82M70.991.396.154.380.288.086.394.194.874.891.692.6
TL;DR-CC3M0.67M72.891.995.954.880.689.487.594.895.375.792.293.4
CC3M [40]2.82M42.367.877.431.555.766.375.191.293.660.685.991.8
TL;DR-CC3M0.67M48.773.182.736.760.670.476.391.993.961.087.793.0
", + "image_path": "7227784f8256e7f7f88367796d0442d7680121e5b97c735acb8be72c1dbce26e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 237, + 541, + 289 + ], + "blocks": [ + { + "bbox": [ + 86, + 219, + 504, + 231 + ], + "lines": [ + { + "bbox": [ + 86, + 219, + 504, + 231 + ], + "spans": [ + { + "bbox": [ + 86, + 219, + 504, + 231 + ], + "type": "text", + "content": "Table 3: Fine-tuning and zero-shot image-text retrieval results on MSCOCO and Flickr30K dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 237, + 541, + 289 + ], + "lines": [ + { + "bbox": [ + 50, + 237, + 541, + 289 + ], + "spans": [ + { + "bbox": [ + 50, + 237, + 541, + 289 + ], + "type": "table", + "html": "
Dataset#SamplesVQANLVR2valRefCOCO+COCO Caption
test-devtest-standarddevtest-PtestAtestBB@4CIDEr
Random-CC3M0.67M68.366.273.673.868.671.862.835.9118.8
CC3M [40]2.8M71.571.876.076.272.476.165.336.8121.6
TL;DR-CC3M0.67M73.1+1.673.2+1.477.7+1.778.0+1.875.1+2.778.5+2.468.4+3.137.6+0.8123.8+2.2
", + "image_path": "42e4741624d41ad0d9206291dcabbf8dcf6019750bc21b925331ec12013a4ff6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 334, + 289, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 289, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 289, + 526 + ], + "type": "text", + "content": "age transformer is initialized from ViT pre-trained on ImageNet [10], and the text transformer is initialized from BERTbase [11](BLIP,CLIP) and DistillBERT [38] (ViLT). The model is pre-trained for 20 epochs with a batch size of 1260 and an AdamW [28] optimizer with a weight decay of 0.05. During training, we apply a learning rate warm-up to 3e-4 and a linear decay with a rate of 0.85. For image augmentation, we utilize RandAugment [8] and apply all of the original policies except color inversion. This decision is based on the recognition of the crucial role that color information plays in the data. For pre-training, images were randomly cropped to a resolution of " + }, + { + "bbox": [ + 46, + 334, + 289, + 526 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 334, + 289, + 526 + ], + "type": "text", + "content": ". We then increase this to " + }, + { + "bbox": [ + 46, + 334, + 289, + 526 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 46, + 334, + 289, + 526 + ], + "type": "text", + "content": " for fine-tuning downstream tasks. Further information about the training hyperparameters for downstream tasks can be found in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 532, + 146, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 146, + 545 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 146, + 545 + ], + "type": "text", + "content": "4.1. Main Properties" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 551, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 574 + ], + "type": "text", + "content": "We ablate our " + }, + { + "bbox": [ + 46, + 551, + 287, + 574 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 46, + 551, + 287, + 574 + ], + "type": "text", + "content": " using the default setting in Table 2 (see caption). Several intriguing properties are observed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 578, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 578, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 45, + 578, + 287, + 674 + ], + "type": "text", + "content": "Module deconstruction. In Table 2a we analyze the impact of different components in " + }, + { + "bbox": [ + 45, + 578, + 287, + 674 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 45, + 578, + 287, + 674 + ], + "type": "text", + "content": ". We establish a baseline by randomly selecting " + }, + { + "bbox": [ + 45, + 578, + 287, + 674 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 45, + 578, + 287, + 674 + ], + "type": "text", + "content": " of the data from CC3M (first row). Our results show that codebook-based sampling outperforms random selection by " + }, + { + "bbox": [ + 45, + 578, + 287, + 674 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 45, + 578, + 287, + 674 + ], + "type": "text", + "content": " in TR@1. We also observe that both codebook-based sampling and caption refinement are crucial and the combination of them achieves optimal downstream performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Sample selection. In Table 2b we study the sample selection strategy in Stage 2. We sample " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": " data in each cluster by default. For Gradient-based, we train a tiny network" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 331, + 544, + 430 + ], + "blocks": [ + { + "bbox": [ + 46, + 289, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 289, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 545, + 313 + ], + "type": "text", + "content": "Table 4: Comparison with BLIP model pre-trained on different data sources for VQA, NLVR², RefCOCO+ and COCO Captioning. ViLT and CLIP architectures can not be evaluated on part of these tasks since structural limitations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 331, + 544, + 430 + ], + "lines": [ + { + "bbox": [ + 306, + 331, + 544, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 544, + 430 + ], + "type": "table", + "html": "
MethodDatasetR@1↑R@5↑R@10↑MdR↓
CLIP [34]Rand-CC3M15.334.846.313.0
CC3M [40]19.437.347.511.0
TL;DR-CC3M21.838.648.510.0
ViLT [19]Rand-CC3M18.838.249.511.0
CC3M [40]21.040.551.510.0
TL;DR-CC3M22.542.752.48.0
BLIP [22]Rand-CC3M23.342.853.38.0
CC3M [40]26.046.358.07.0
TL;DR-CC3M27.448.759.46.0
", + "image_path": "4f5d60ea3aa24d5028783a0d8632d6b6cd046def480ae52d39e2bfa694acb3fc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 431, + 545, + 479 + ], + "lines": [ + { + "bbox": [ + 305, + 431, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 545, + 479 + ], + "type": "text", + "content": "Table 5: MSRVTT-1K retrieval using three architectures. We created a subset of the CC3M dataset called Rand-CC3M by randomly selecting the same number of samples as in TL;DR-CC3M." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 647 + ], + "type": "text", + "content": "to conduct VLP pretrained with ITC [24], ITM [24] and LM [11]. Then we select samples which contribute most to the gradients in each cluster. Large distance: Another perspective is that data points on the border of each cluster are more important than those at the center [4]. So we first compute the center of each cluster and then choose the sample that has the largest distance from the center of each cluster. We also report the result of hard-sample selection from [41]. We observe that all these variants produce similar results except large distances. This suggests that the clustering step, rather than the selection step, plays a key role in data compression during Stage 2. To maintain simplicity, we choose uniform sampling as the default method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Codebook initialization. In Table 2c we compare different initialization strategies. The xavier means all parameters in the codebook are initialized with xavier initialization [14]. For the object tags initialization, following previous works [2, 60], we use the 1600 object tags from" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3151" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 287, + 168 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 287, + 168 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 287, + 168 + ], + "type": "table", + "html": "
ModelDataset#SamplesImNetImNet-AImNet-R
CLIP [34]Rand-CC3M0.67M58.361.862.3
CC3M [40]2.82M62.265.266.9
TL;DR-CC3M0.67M61.465.065.7
ViLT [19]Rand-CC3M0.67M54.359.858.4
CC3M [40]2.82M58.662.964.2
TL;DR-CC3M0.67M59.163.364.0
BLIP [22]Rand-CC3M0.67M57.361.865.2
CC3M [40]2.82M62.565.568.1
TL;DR-CC3M0.67M62.063.967.4
", + "image_path": "47d2443050d4ff275131faf8caa869362609839b84ec46870ca6fe76eb434e2e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 170, + 288, + 231 + ], + "lines": [ + { + "bbox": [ + 46, + 170, + 288, + 231 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 288, + 231 + ], + "type": "text", + "content": "Table 6: Zero-shot image classification results on ImageNet [10], ImageNet-A [16], ImageNet-R [15]. There is no free lunch, as selecting partial samples reduces the visual diversity crucial for classification. Despite this, TL;DR still performs significantly better than random selection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 250, + 288, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 288, + 323 + ], + "type": "text", + "content": "Visual Genome [20] and extract text feature with a pretrained BERT [11]. With same training setting, the keywords achieve a " + }, + { + "bbox": [ + 46, + 250, + 288, + 323 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 46, + 250, + 288, + 323 + ], + "type": "text", + "content": " TR@1 improvement and a " + }, + { + "bbox": [ + 46, + 250, + 288, + 323 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 46, + 250, + 288, + 323 + ], + "type": "text", + "content": " IR@1 improvement over xavier. This result is expected as the text embeddings provide contextual information and simplify the learning process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 328, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 485 + ], + "type": "text", + "content": "Codebook vs. Image embedding. In Table 2d, we investigate different ways of cluster sampling. First, we remove the codebook from Stage-1 and use image embedding instead. Alternatively, we directly cluster images using the image embedding [22] of images from BLIP model (pretrained on 200M Image-text pairs). We observe the image embedding leads to much better result than text embedding. This is reasonable because clustering visual-similarity samples with text only is difficult. We observe that clustering depended on our codebook performs better than both image embedding and text embedding. This demonstrates that our codebook can efficiently project image embedding to semantic space, benefiting cluster process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "text", + "content": "Cluster sampling ratio. Table 2e varies the sampling ratio of each cluster from " + }, + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "text", + "content": ". We are surprised to find that a low sampling ratio can still produce effective results. With only " + }, + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "text", + "content": " of the data and the TL;DR model, we are able to achieve a " + }, + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "text", + "content": " improvement on TR@1 and a " + }, + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "text", + "content": " improvement on IR@1 over the full-scale baseline. Additionally, we observe that larger sampling ratios lead to even better results. Since our focus is on achieving similar transfer learning results with fewer samples, we use a default sampling ratio of " + }, + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 45, + 491, + 287, + 612 + ], + "type": "text", + "content": " to minimize computation costs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Cluster numbers. In Table 2f, we investigate the impact of cluster number on Stage 2 by increasing it from 300 to " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "30\\mathrm{K}" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": ". We observe that using more clusters results in a slight improvement at the beginning and becomes stable when the number of clusters exceeds " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "3\\mathrm{K}" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": ". Moreover, all results consistently outperform the random selection baseline. Therefore, we use " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "3\\mathrm{K}" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " clusters as the default in this work, as it performs well on fine-tuning tasks." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 356, + 121 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 356, + 121 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 356, + 121 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 356, + 121 + ], + "type": "image", + "image_path": "fb4d08a74d726e2c19aab1d2bb512d5f5fd9f94a890950993bd85c7176195275.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 125, + 380, + 138 + ], + "lines": [ + { + "bbox": [ + 310, + 125, + 380, + 138 + ], + "spans": [ + { + "bbox": [ + 310, + 125, + 380, + 138 + ], + "type": "text", + "content": "Captioner: A racer jumps over a railing on a horse." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 385, + 72, + 455, + 138 + ], + "blocks": [ + { + "bbox": [ + 385, + 72, + 455, + 138 + ], + "lines": [ + { + "bbox": [ + 385, + 72, + 455, + 138 + ], + "spans": [ + { + "bbox": [ + 385, + 72, + 455, + 138 + ], + "type": "image", + "image_path": "d61092376a66fc5dcb355e7359d596571283f1342776441317f33c10fadf4f50.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 142, + 539, + 155 + ], + "lines": [ + { + "bbox": [ + 311, + 142, + 539, + 155 + ], + "spans": [ + { + "bbox": [ + 311, + 142, + 539, + 155 + ], + "type": "text", + "content": "Figure 4: The generated caption match the image well." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 464, + 72, + 531, + 138 + ], + "blocks": [ + { + "bbox": [ + 464, + 72, + 531, + 138 + ], + "lines": [ + { + "bbox": [ + 464, + 72, + 531, + 138 + ], + "spans": [ + { + "bbox": [ + 464, + 72, + 531, + 138 + ], + "type": "image", + "image_path": "29efb37dd3e58a3e611286fdf1fa5cb2bc7d2d1b560069d6be6c54e10b5fb60e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 166, + 480, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 166, + 480, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 166, + 480, + 179 + ], + "type": "text", + "content": "4.2. Transfer Learning Experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 185, + 547, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 547, + 281 + ], + "type": "text", + "content": "We conduct an extensive evaluation of transfer learning in downstream tasks using the model pre-trained on our compressed TL;DR-CC3M and source CC3M with 3 architectures. Our evaluation primarily focuses on the core tasks of three categories that examine: (1) cross-modality alignment, (2) image captioning and multi-modality understanding capabilities, and (3) visual recognition. The baseline in this section is the model trained on CC3M dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 299, + 474, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 299, + 474, + 311 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 474, + 311 + ], + "type": "text", + "content": "4.2.1 Cross-modality Alignment Task" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 322, + 546, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 546, + 491 + ], + "type": "text", + "content": "Image-Text retrieval. Fine-grained world region alignment plays a critical role in this task. We report both image-to-text retrieval (TR) and text-to-image retrieval (IR) on the COCO [26] and Flickr30K [32] benchmarks. For the BLIP [22] model, we adopt an additional re-ranking strategy, following the original implementation. In Table 3, we also report zero-shot retrieval results. We found that " + }, + { + "bbox": [ + 304, + 322, + 546, + 491 + ], + "type": "inline_equation", + "content": "TL; DR" + }, + { + "bbox": [ + 304, + 322, + 546, + 491 + ], + "type": "text", + "content": " achieves comparable results with the baselines on all metrics and surprisingly performs quite well on zero-shot results. For example, for the BLIP [22] architecture, our method leads to a " + }, + { + "bbox": [ + 304, + 322, + 546, + 491 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 304, + 322, + 546, + 491 + ], + "type": "text", + "content": " improvement (from " + }, + { + "bbox": [ + 304, + 322, + 546, + 491 + ], + "type": "inline_equation", + "content": "42.3\\%" + }, + { + "bbox": [ + 304, + 322, + 546, + 491 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 322, + 546, + 491 + ], + "type": "inline_equation", + "content": "48.7\\%" + }, + { + "bbox": [ + 304, + 322, + 546, + 491 + ], + "type": "text", + "content": ") in Recall@1 of image-to-text retrieval on MSCOCO. All results suggest that a small part of refined image-text pairs is enough to learn good alignment." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 494, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 601 + ], + "type": "text", + "content": "Zero-shot video retrieval. In this experiment, we analyze the generalization ability of our method to video-language tasks. Specifically, we perform zero-shot transfer to text-to-video retrieval and evaluate the models trained on COCO-retrieval in Table 5. We uniformly sample 8 frames from each video to process the video input and concatenate the frame features into a single sequence. These models trained on our compressed dataset outperform the baseline on all metrics, demonstrating the generality of " + }, + { + "bbox": [ + 304, + 494, + 547, + 601 + ], + "type": "inline_equation", + "content": "TL; DR" + }, + { + "bbox": [ + 304, + 494, + 547, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 613, + 423, + 654 + ], + "blocks": [ + { + "bbox": [ + 307, + 613, + 423, + 654 + ], + "lines": [ + { + "bbox": [ + 307, + 613, + 423, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 423, + 654 + ], + "type": "image", + "image_path": "a636e3c42ab7ea13a7115b8c9c644393d168fab2806dca47ae6f579d805f5f92.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 659, + 547, + 708 + ], + "lines": [ + { + "bbox": [ + 305, + 659, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 305, + 659, + 547, + 708 + ], + "type": "text", + "content": "Figure 5: The codebook-based clusters visualization. The samples within each cluster exhibit similar contextual characteristics, as opposed to mere visual appearance. For example, the \"Christmas elements\" cluster located at the right." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 428, + 613, + 506, + 654 + ], + "blocks": [ + { + "bbox": [ + 428, + 613, + 506, + 654 + ], + "lines": [ + { + "bbox": [ + 428, + 613, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 428, + 613, + 506, + 654 + ], + "type": "image", + "image_path": "bdc5d88023c00c6a8dc1bcc8abd2de0ca3060f21f09dc14156d57c309caf8473.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 507, + 613, + 545, + 654 + ], + "blocks": [ + { + "bbox": [ + 507, + 613, + 545, + 654 + ], + "lines": [ + { + "bbox": [ + 507, + 613, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 507, + 613, + 545, + 654 + ], + "type": "image", + "image_path": "441946df53e490025b215e92156544169bc701c92f569e6cb98ca673ddb9b94f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3152" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 127, + 144 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 127, + 144 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 127, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 127, + 144 + ], + "type": "image", + "image_path": "ffdb52a15d2ef9e150e33b0e9ae0a75d2631b936484146dd24102825cd515e8a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 145, + 107, + 152 + ], + "lines": [ + { + "bbox": [ + 70, + 145, + 107, + 152 + ], + "spans": [ + { + "bbox": [ + 70, + 145, + 107, + 152 + ], + "type": "text", + "content": "VQ-GAN: 5.13s" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 203, + 144 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 203, + 144 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 203, + 144 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 203, + 144 + ], + "type": "image", + "image_path": "97cbdbff14d12199f45ee15a7bbef5f1ad326c7625f9bc4aa6995d35b5c3e844.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 145, + 184, + 152 + ], + "lines": [ + { + "bbox": [ + 148, + 145, + 184, + 152 + ], + "spans": [ + { + "bbox": [ + 148, + 145, + 184, + 152 + ], + "type": "text", + "content": "DALL E2: 2.65s" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 70, + 279, + 144 + ], + "blocks": [ + { + "bbox": [ + 205, + 70, + 279, + 144 + ], + "lines": [ + { + "bbox": [ + 205, + 70, + 279, + 144 + ], + "spans": [ + { + "bbox": [ + 205, + 70, + 279, + 144 + ], + "type": "image", + "image_path": "57a1d2243c7355a484af46cdfc1cc5dbf61468cc608f54254b55c89d806d9aed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 145, + 272, + 151 + ], + "lines": [ + { + "bbox": [ + 221, + 145, + 272, + 151 + ], + "spans": [ + { + "bbox": [ + 221, + 145, + 272, + 151 + ], + "type": "text", + "content": "Stable Diffusion: 0.17s" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 154, + 286, + 177 + ], + "lines": [ + { + "bbox": [ + 47, + 154, + 286, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 286, + 177 + ], + "type": "text", + "content": "Figure 6: Image generation result with strong Text-to-image Model. The generation time is also reported." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 190, + 286, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 190, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 286, + 213 + ], + "type": "text", + "content": "4.2.2 Image Captioning and Multi-modality Understanding Tasks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 224, + 286, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 286, + 331 + ], + "type": "text", + "content": "Image captioning. The task involves describing an input image, which we evaluate using No-Caps and COCO datasets. Both datasets are fine-tuned on COCO with the Language Modeling (LM) loss. We adopt a zero-shot setting for No-Caps dataset, and start each caption with the phrase \"a picture of\" for the BLIP architecture. We do not pre-train using COCO to avoid information leakage. Our results outperform baseline with a much smaller quantity of pre-training data, as shown in Table 4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 335, + 286, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 286, + 443 + ], + "type": "text", + "content": "Visual question answering (VQA). We evaluate our model's performance on the VQA task [3], where the model needs to provide an answer based on an image and a question. We consider it as an answer generation task that allows open-vocabulary VQA for better results, following previous works [23, 22]. The results are presented in Table 4. The BLIP trained on TL;DR-CC3M outperforms baseline by " + }, + { + "bbox": [ + 46, + 335, + 286, + 443 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 46, + 335, + 286, + 443 + ], + "type": "text", + "content": " on test-dev splits, demonstrating the effectiveness of our compressed dataset for improving VQA performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 447, + 286, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 286, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 286, + 530 + ], + "type": "text", + "content": "Visual reasoning. The Natural Language Visual Reasoning " + }, + { + "bbox": [ + 46, + 447, + 286, + 530 + ], + "type": "inline_equation", + "content": "(\\mathrm{NLVR}^2)" + }, + { + "bbox": [ + 46, + 447, + 286, + 530 + ], + "type": "text", + "content": " [42] task is a binary classification task that requires the model to reason about two images and a question in natural language. Multi-modal reasoning is crucial for the completion of this task. We observe that BLIP trained on our dataset achieved " + }, + { + "bbox": [ + 46, + 447, + 286, + 530 + ], + "type": "inline_equation", + "content": "78.0\\%" + }, + { + "bbox": [ + 46, + 447, + 286, + 530 + ], + "type": "text", + "content": " accuracy compared to " + }, + { + "bbox": [ + 46, + 447, + 286, + 530 + ], + "type": "inline_equation", + "content": "76.2\\%" + }, + { + "bbox": [ + 46, + 447, + 286, + 530 + ], + "type": "text", + "content": " achieved by the CC3M, as shown in Table 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 534, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 286, + 628 + ], + "type": "text", + "content": "Cross-modality grounding. Referring Expression (RE) Comprehension requires the model to select the target object from a set of image regions proposals, based on the query description. This task heavily relies on visual-grounding ability. The models are evaluated on ground-truth objects, and we evaluate RE Comprehension on RefCOCO+ [58]. The results are reported in Table 4, and we observe that TL;DR-CC3M achieves better results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 647, + 187, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 187, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 187, + 658 + ], + "type": "text", + "content": "4.2.3 Visual Recognition Tasks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "content": "Besides the cross-modality task, we also explore a unimodality task, mainly image classification. Specifically, we fix the image encoder and explore zero-shot image classification. Since the vision encoder is loaded from" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 324, + 70, + 524, + 121 + ], + "blocks": [ + { + "bbox": [ + 324, + 70, + 524, + 121 + ], + "lines": [ + { + "bbox": [ + 324, + 70, + 524, + 121 + ], + "spans": [ + { + "bbox": [ + 324, + 70, + 524, + 121 + ], + "type": "table", + "html": "
MethodTR@1IR@1
real data58.344.0
VQ-GAN [13]35.232.4
DALLE2 [35] (implement from 3)44.338.3
Stable Diffusion [37] (implement from 4)52.440.7
", + "image_path": "a163e91b1eab68a9e6b3ef8e547a527c55c7b5af681a6817cbc6331947d470a0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 121, + 545, + 157 + ], + "lines": [ + { + "bbox": [ + 305, + 121, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 305, + 121, + 545, + 157 + ], + "type": "text", + "content": "Table 7: Compare different sample generation methods over 0.3M subset of CC3M. We first pre-train BLIP model on these generated data and then evaluation on COCO." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 171, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 410 + ], + "type": "text", + "content": "pretrained model, this task demonstrates the impact of training a Vision-Language model with noisy image-text pairs. Specifically, it shows how such training affects the well-learned representation derived from a human-crafted dataset. We show the results in Table 6 and observe noisy data leads to significant Catastrophic Forgetting. For example, the CLIP model drops down to only " + }, + { + "bbox": [ + 304, + 171, + 545, + 410 + ], + "type": "inline_equation", + "content": "58.3\\%" + }, + { + "bbox": [ + 304, + 171, + 545, + 410 + ], + "type": "text", + "content": " accuracy with noisy data training. We also observe our TL;DR shows steady improvement for all architectures over random selection. Unfortunately, the classification performance for TL;DR-CC3M is slightly worse than the full-scale CC3M for the CLIP and BLIP architectures. Both of these architectures have independent image encoders like ViT to extract image embeddings. This indicates that this task heavily relies on visual diversity, which is different from multimodal tasks, and our method reduces the visual diversity potentially. For the ViLT model, this architecture adopts a shared backbone for both visual and text, and we observe the slightly different results. We guess that multi-modality interaction in early-fusion affects the classification result." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 419, + 389, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 389, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 389, + 430 + ], + "type": "text", + "content": "4.3. Visualization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 441, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 573 + ], + "type": "text", + "content": "Generated caption visualization. We show the generated caption in Figure 4. It is evident that the original captions can be highly abstract and difficult to match their respective images, even for human observers sometimes. For instance, when the ITM score is as low as 0.04, matching the figure with its corresponding caption becomes arduous. Such challenging cases can potentially harm the cross-modality alignment. In contrast, we observe that the generated captions describe the image very well and sometimes offer helpful complementary information. For example, \"bus\" and \"castle\" in the middle example." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 577, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 545, + 685 + ], + "type": "text", + "content": "Codebook-based cluster visualization. Figure 5 displays the codebook grouping result achieved with simple KMeans. Clusters are sets of data points with similar characteristics, often defined by their features or attributes. Interestingly, we observe that the model cluster samples \"accurate\", meaning that these samples have semantic similarity rather than simple appearance. For instance, the model classifies \"dollars\" and \"piggy bank\" together, even though they differ significantly in appearance." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 693, + 459, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 693, + 459, + 703 + ], + "spans": [ + { + "bbox": [ + 315, + 693, + 459, + 703 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 315, + 693, + 459, + 703 + ], + "type": "text", + "content": "https://github.com/LAION-AI/dalle2-laion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 703, + 453, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 453, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 453, + 713 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 317, + 703, + 453, + 713 + ], + "type": "text", + "content": "https://github.com/huggingface/diffusers" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3153" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 69, + 547, + 178 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 547, + 178 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 547, + 178 + ], + "type": "table", + "html": "
Dataset#Samples TimeVQA test-devNLVR2 test-PRefCOCO valNocaps Captioning B@4CIDErFlcikr30K Retrieval TR@1IR@1Imagenet Acc
Rand-CC12M2.4M14h71.876.272.536.8121.082.973.361.2
CC12M [7]10.8M65h73.578.974.137.5122.984.775.365.3
TL;DR-CC12M2.4M14h74.1+0.678.5-0.474.0-0.138.1+0.6124.1+1.285.5+0.876.3+1.063.8-1.5
Rand-YFCC15M2.5M15h67.270.568.135.2116.378.870.565.4
YFCC15M [43]15M90h70.574.270.635.9118.481.572.467.8
TL;DR-YFCC15M2.5M15h70.3-0.275.3+1.172.6+2.037.2+1.3122.5+4.182.3+0.874.3+1.967.3-0.5
Rand-LAION40M(128)8M48h70.775.373.434.8113.280.472.568.5
LAION40M(128) [39]40M120h74.579.176.635.2117.483.274.971.3
TL;DR-LAION40M(128)8M48h76.3+1.880.5+1.477.4+0.836.8+1.6120.9+3.582.8-0.476.1+1.270.4-0.9
", + "image_path": "927d86d3ae134e157e05f31675b6da0916a54566ba68ee4281a80a27dfe0cce1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 179, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 46, + 179, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 547, + 205 + ], + "type": "text", + "content": "Table 8: Comparison with different source of data on 6 downstream tasks. BLIP [22] is adopted as baseline and (128) means the image resolution is " + }, + { + "bbox": [ + 46, + 179, + 547, + 205 + ], + "type": "inline_equation", + "content": "{128} \\times {128}" + }, + { + "bbox": [ + 46, + 179, + 547, + 205 + ], + "type": "text", + "content": " . We also list the pre-training time,which can be significantly reduced via " + }, + { + "bbox": [ + 46, + 179, + 547, + 205 + ], + "type": "inline_equation", + "content": "{TL};{DR}" + }, + { + "bbox": [ + 46, + 179, + 547, + 205 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 60, + 214, + 168, + 306 + ], + "blocks": [ + { + "bbox": [ + 60, + 214, + 168, + 306 + ], + "lines": [ + { + "bbox": [ + 60, + 214, + 168, + 306 + ], + "spans": [ + { + "bbox": [ + 60, + 214, + 168, + 306 + ], + "type": "image", + "image_path": "9bda6184e9448fb0fd18b91f69cf5f7fb0592e7edbde4e83215fd5cdbe75508d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 309, + 288, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 309, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 288, + 334 + ], + "type": "text", + "content": "Figure 7: ITM score distribution. " + }, + { + "bbox": [ + 46, + 309, + 288, + 334 + ], + "type": "inline_equation", + "content": "TL; DR" + }, + { + "bbox": [ + 46, + 309, + 288, + 334 + ], + "type": "text", + "content": " alleviates the issue of misalignment in VLP data." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 170, + 215, + 273, + 306 + ], + "blocks": [ + { + "bbox": [ + 170, + 215, + 273, + 306 + ], + "lines": [ + { + "bbox": [ + 170, + 215, + 273, + 306 + ], + "spans": [ + { + "bbox": [ + 170, + 215, + 273, + 306 + ], + "type": "image", + "image_path": "6b57301f7f7960dbca96e636febef744721af53ddbc7346350b7bc1bcbc87430.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 348, + 159, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 159, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 159, + 361 + ], + "type": "text", + "content": "4.4. More Investigation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 371, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 515 + ], + "type": "text", + "content": "Is image generation possible? To ease the misalignment problem of image-text pairs, instead of simply selecting representative samples, a potential and naive idea is to generate images from text. To this end, we randomly sample 0.3M subset of CC3M and generate image from text with three popular text to image generation models, VQ-GAN [56], DALLE 2 [35] and Stable Diffusion [37]. We display the generated samples in Figure 6. We observe that the generative models struggle with complex scenarios, but are capable of generating simple prompts like \"dog\" proficiently. In addition, generation methods only produce visual cues in a fixed vocabulary, potentially reducing data diversity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "content": "Next, we pre-train BLIP models on these generated data and evaluate it on COCO Retrieval. In Table 7 we observe the results of transfer learning depend on the quality of generated samples, with those generated by stable diffusion being particularly effective. However, there still exists a significant gap between the generated data and the real dataset (e.g., " + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "inline_equation", + "content": "52.4\\%" + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "inline_equation", + "content": "58.3\\%" + }, + { + "bbox": [ + 46, + 517, + 287, + 624 + ], + "type": "text", + "content": " on TR@1). We believe that higher-quality and diverse generated images may lead to comparable results with real images in the near future." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Explore the misalignment problem. Figure 7 shows the Image-text Matching (ITM) score distribution for both " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "CC3M" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " and our " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "TL;DR-CC3M" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " data (the visualization about more datasets is reported in the supplementary). We observe a lot of samples of original " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "CC3M" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " at low matching score even tends to zero, which indicates the current dataset has serious misalignment problems. Since image" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 217, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 277 + ], + "type": "text", + "content": "text matching (ITM) loss and image-text contrastive (ITC) loss are used in all architectures, these samples will damage the multimodal representation learning. When adopting our " + }, + { + "bbox": [ + 304, + 217, + 547, + 277 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 304, + 217, + 547, + 277 + ], + "type": "text", + "content": ", we observe that the matching score tends to be higher and has very few samples with low ITM score." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 293, + 480, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 480, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 480, + 306 + ], + "type": "text", + "content": "5. Transfer to other VLP datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 316, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 555 + ], + "type": "text", + "content": "We study data compression performed in two categories shown below: clean data that involves human-based offline filter pipelines and raw data that has not undergone cleaning. For clean data, in addition to " + }, + { + "bbox": [ + 304, + 316, + 545, + 555 + ], + "type": "inline_equation", + "content": "CC3M" + }, + { + "bbox": [ + 304, + 316, + 545, + 555 + ], + "type": "text", + "content": ", we explore the well-cleaned, high-quality dataset " + }, + { + "bbox": [ + 304, + 316, + 545, + 555 + ], + "type": "inline_equation", + "content": "CC12M" + }, + { + "bbox": [ + 304, + 316, + 545, + 555 + ], + "type": "text", + "content": " [7]. Then, we study the raw data YFCC100M [43] and LAION400M [39]. " + }, + { + "bbox": [ + 304, + 316, + 545, + 555 + ], + "type": "inline_equation", + "content": "CC12M" + }, + { + "bbox": [ + 304, + 316, + 545, + 555 + ], + "type": "text", + "content": " [7] contains 12 million image-text pairs specifically meant to be used for vision-and-language pre-training. These data are collected by relaxing the data collection pipeline as in CC3M. YFCC15M [34] is a subset of the multilingual and noisy YFCC100M [43] that contains English captions. LAION400M [39] is a large-scale noisy dataset that provides URLs with captions for download. To control the computation cost and reduce the storage overhead, we randomly sample a 40M subset of LAION400M and download images at a resolution of " + }, + { + "bbox": [ + 304, + 316, + 545, + 555 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 316, + 545, + 555 + ], + "type": "text", + "content": ". So, we record the dataset as " + }, + { + "bbox": [ + 304, + 316, + 545, + 555 + ], + "type": "inline_equation", + "content": "TL;DR-LAION40M(128)" + }, + { + "bbox": [ + 304, + 316, + 545, + 555 + ], + "type": "text", + "content": ", and the performance over downstream tasks could be improved with higher resolution. More exploration about video-text datasets is reported in the supplementary material" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "We use BLIP as the default architecture and evaluate our " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": " on different datasets and show the results in Table 8. Surprisingly, with only " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "2.5\\mathrm{M}" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "(16.7\\%)" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": " data, " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "-YFCC15M leads to similar results with " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "15\\mathrm{M}" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": " raw data over all metrics except Imagenet. More results with different backbones are reported in the supplementary material. For LAION40M(128), when using " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "8\\mathrm{M}" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": " data " + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "inline_equation", + "content": "(20\\%)" + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": ", the model trained on our dataset consistently outperforms the baseline method on six downstream tasks. We noticed that the compression rate of LAION40M(128) is less than that of YFCC15M. This may be due to the fact that the collection of LAION40M(128) has already been filtered with CLIP similarity, reducing the impact of the misalignment problem." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3154" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 200, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 200, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 200, + 84 + ], + "type": "text", + "content": "6. Conclusion and Discussion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 283 + ], + "type": "text", + "content": "This paper presents " + }, + { + "bbox": [ + 46, + 91, + 289, + 283 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 46, + 91, + 289, + 283 + ], + "type": "text", + "content": ", a novel and pioneering algorithm for selecting and generating high-quality image-text pairs from noisy Vision-Language Pre-training (VLP) data, thereby contributing to the field of VLP. " + }, + { + "bbox": [ + 46, + 91, + 289, + 283 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 46, + 91, + 289, + 283 + ], + "type": "text", + "content": " incorporates a text generation process into learning to reduce serious misalignment problems. Our experiments demonstrate three widely-used architectures leads to comparable results and much smaller training cost when learning from our generated dataset. Additionally, we demonstrate that the misalignment problem can be effectively addressed using our simple " + }, + { + "bbox": [ + 46, + 91, + 289, + 283 + ], + "type": "inline_equation", + "content": "TL;DR" + }, + { + "bbox": [ + 46, + 91, + 289, + 283 + ], + "type": "text", + "content": ". However, the choice of the highest compression ratio is done manually rather than learned. Furthermore, achieving even higher compression ratios for VLP models remains a challenge, and text-to-image generation models may be helpful in this regard. We hope that this perspective will inspire future research." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 296, + 145, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 296, + 145, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 296, + 145, + 309 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 389 + ], + "type": "text", + "content": "This project is supported by the National Research Foundation, Singapore under its NRFF Award NRF-NRFF13-2021-0008, and Meta, Mike Zheng Shou's Start-Up Grant from NUS. The computational work for this article was partially performed on resources of the National Supercomputing Centre, Singapore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 400, + 107, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 400, + 107, + 412 + ], + "spans": [ + { + "bbox": [ + 47, + 400, + 107, + 412 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 420, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 52, + 420, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 420, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 52, + 420, + 287, + 475 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 477, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 477, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 477, + 288, + 542 + ], + "type": "text", + "content": "[2] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077-6086, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 544, + 289, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 289, + 599 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 289, + 599 + ], + "type": "text", + "content": "[3] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 601, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 288, + 633 + ], + "type": "text", + "content": "[4] Dibya Jyoti Bora, Dr Gupta, and Anil Kumar. A comparative study between fuzzy clustering algorithm and hard clustering algorithm. arXiv preprint arXiv:1404.6059, 2014." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 288, + 678 + ], + "type": "text", + "content": "[5] Tianshi Cao, Sasha Alexandre Doubov, David Acuna, and Sanja Fidler. Scalable neural data server: A data recommender for transfer learning. Advances in Neural Information Processing Systems, 34:8984-8997, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 681, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 288, + 714 + ], + "type": "text", + "content": "[6] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Proceedings of the" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4750-4759, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 97, + 547, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 97, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 312, + 97, + 547, + 151 + ], + "type": "text", + "content": "[7] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 152, + 547, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 547, + 207 + ], + "type": "text", + "content": "[8] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 312, + 209, + 545, + 241 + ], + "type": "text", + "content": "[9] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation to imagenet-1k with constant memory. arXiv preprint arXiv:2211.10586, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 243, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 286 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 288, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 545, + 331 + ], + "type": "text", + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 333, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 333, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 333, + 545, + 398 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 399, + 545, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 444 + ], + "type": "text", + "content": "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 445, + 545, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 500 + ], + "type": "text", + "content": "[14] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249-256. JMLR Workshop and Conference Proceedings, 2010." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 502, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 502, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 502, + 545, + 567 + ], + "type": "text", + "content": "[15] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "text", + "content": "[16] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "type": "text", + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[18] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billionscale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535-547, 2019." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3155" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[19] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 288, + 185 + ], + "type": "text", + "content": "[20] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 287, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 210 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 210 + ], + "type": "text", + "content": "[21] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 211, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 211, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 211, + 287, + 255 + ], + "type": "text", + "content": "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 287, + 312 + ], + "type": "text", + "content": "[23] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 314, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 287, + 369 + ], + "type": "text", + "content": "[24] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121-137. Springer, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "text", + "content": "[25] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 428, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 428, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 428, + 287, + 483 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 484, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 484, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 484, + 287, + 518 + ], + "type": "text", + "content": "[27] Yiqi Lin, Huabin Zheng, Huaping Zhong, Jinjing Zhu, Wei-jia Li, Conghui He, and Lin Wang. Sept: Towards scalable and efficient visual pre-training. AAAI, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 520, + 286, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 286, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 286, + 542 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 609 + ], + "type": "text", + "content": "[29] Rafid Mahmood, James Lucas, David Acuna, Daiqing Li, Jonah Philion, Jose M Alvarez, Zhiding Yu, Sanja Fidler, and Marc T Law. How much more data do i need? estimating requirements for downstream tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 275-284, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 667 + ], + "type": "text", + "content": "[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 286, + 713 + ], + "type": "text", + "content": "[31] Mansheej Paul, Surya Ganguli, and Gintare Karolina Dziugaite. Deep learning on a data diet: Finding important examples early in training. Advances in Neural Information Processing Systems, 34:20596-20607, 2021." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 139 + ], + "type": "text", + "content": "[32] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 546, + 184 + ], + "type": "text", + "content": "[33] Ziheng Qin, Kai Wang, Zangwei Zheng, Jianyang Gu, Xi-angyu Peng, Daquan Zhou, and Yang You. Infobatch: Lossless training speed up by unbiased dynamic data pruning. arXiv preprint arXiv:2303.04947, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 241, + 546, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 546, + 283 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 546, + 283 + ], + "type": "text", + "content": "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 285, + 546, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 546, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 546, + 330 + ], + "type": "text", + "content": "[36] Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, and Suvrit Sra. Can contrastive learning avoid shortcut solutions? Advances in neural information processing systems, 34:4974-4986, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 547, + 386 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 388, + 546, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 546, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 546, + 430 + ], + "type": "text", + "content": "[38] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 433, + 546, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 546, + 487 + ], + "type": "text", + "content": "[39] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 488, + 546, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 546, + 553 + ], + "type": "text", + "content": "[40] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 555, + 546, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 546, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 546, + 589 + ], + "type": "text", + "content": "[41] Ben Sorscher, Robert Geirhos, Shashank Shekhar, Surya Ganguli, and Ari S Morcos. Beyond neural scaling laws: beating power law scaling via data pruning. NeurIPS, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 590, + 546, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 546, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 546, + 633 + ], + "type": "text", + "content": "[42] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Hua-jun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. arXiv preprint arXiv:1811.00491, 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 635, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 546, + 678 + ], + "type": "text", + "content": "[43] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 680, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 546, + 713 + ], + "type": "text", + "content": "[44] Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J Gordon. An empirical study of example forget-" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3156" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "ting during deep neural network learning. arXiv preprint arXiv:1812.05159, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 96, + 288, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 96, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 47, + 96, + 288, + 129 + ], + "type": "text", + "content": "[45] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 132, + 288, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 132, + 288, + 175 + ], + "spans": [ + { + "bbox": [ + 47, + 132, + 288, + 175 + ], + "type": "text", + "content": "[46] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2015." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 177, + 288, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 288, + 210 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 288, + 210 + ], + "type": "text", + "content": "[47] Haoran Wang, Yue Zhang, Xiaosheng Yu, et al. An overview of image caption generation methods. Computational intelligence and neuroscience, 2020, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 212, + 288, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 212, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 212, + 288, + 278 + ], + "type": "text", + "content": "[48] Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Kevin Qinghong Lin, Satoshi Tsutsui, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, et al. All in one: Exploring unified video-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6598-6608, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 280, + 288, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 280, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 280, + 288, + 324 + ], + "type": "text", + "content": "[49] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 326, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 288, + 392 + ], + "type": "text", + "content": "[50] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12196-12205, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 393, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 288, + 426 + ], + "type": "text", + "content": "[51] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 428, + 288, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 428, + 288, + 473 + ], + "spans": [ + { + "bbox": [ + 47, + 428, + 288, + 473 + ], + "type": "text", + "content": "[52] Hu Xu, Saining Xie, Po-Yao Huang, Licheng Yu, Russell Howes, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Cit: Curation in training for effective vision-language data. arXiv preprint arXiv:2301.02241, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 475, + 288, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 288, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 288, + 518 + ], + "type": "text", + "content": "[53] Xi Yan, David Acuna, and Sanja Fidler. Neural data server: A large-scale search engine for transfer learning data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3893-3902, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 520, + 288, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 288, + 564 + ], + "type": "text", + "content": "[54] Xingcheng Yao, Yanan Zheng, Xiaocoong Yang, and Zhilin Yang. Nlp from scratch without large-scale pretraining: A simple and efficient framework. In International Conference on Machine Learning, pages 25438-25451. PMLR, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 566, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 288, + 621 + ], + "type": "text", + "content": "[55] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 623, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 288, + 667 + ], + "type": "text", + "content": "[56] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. ICLR, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 669, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 288, + 712 + ], + "type": "text", + "content": "[57] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 240 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "text", + "content": "[58] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 546, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 161 + ], + "type": "text", + "content": "[59] Ruonan Yu, Songhua Liu, and Xinchao Wang. Dataset distillation: A comprehensive review. arXiv preprint arXiv:2301.07014, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 546, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 546, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 546, + 206 + ], + "type": "text", + "content": "[60] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Making visual representations matter in vision-language models. 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 546, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 546, + 240 + ], + "type": "text", + "content": "[61] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. arXiv preprint arXiv:2006.05929, 2020." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3157" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/ToonTalker_ Cross-Domain Face Reenactment/d3014894-36d6-4bc9-bdc4-403449ecf9bd_content_list.json b/2023/ToonTalker_ Cross-Domain Face Reenactment/d3014894-36d6-4bc9-bdc4-403449ecf9bd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ba59f3a5a45819ced3dc42902e2441d8c91bf8d9 --- /dev/null +++ b/2023/ToonTalker_ Cross-Domain Face Reenactment/d3014894-36d6-4bc9-bdc4-403449ecf9bd_content_list.json @@ -0,0 +1,2195 @@ +[ + { + "type": "text", + "text": "ToonTalker: Cross-Domain Face Reenactment", + "text_level": 1, + "bbox": [ + 250, + 130, + 722, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuan Gong $^{1}$ Yong Zhang $^{2*}$ Xiaodong Cun $^{2}$ Fei Yin $^{1}$ Yanbo Fan $^{2}$ Xuan Wang $^{3}$ Baoyuan Wu $^{4}$ Yujiu Yang $^{1*}$", + "bbox": [ + 256, + 179, + 720, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shenzhen International Graduate School, Tsinghua University $^{2}$ Tencent AI Lab $^{3}$ Ant Group $^{4}$ The School of Data Science, Shenzhen Research Institute of Big Data, The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen)", + "bbox": [ + 155, + 215, + 815, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We target cross-domain face reenactment in this paper, i.e., driving a cartoon image with the video of a real person and vice versa. Recently, many works have focused on one-shot talking face generation to drive a portrait with a real video, i.e., within-domain reenactment. Straightforwardly applying those methods to cross-domain animation will cause inaccurate expression transfer, blur effects, and even apparent artifacts due to the domain shift between cartoon and real faces. Only a few works attempt to settle cross-domain face reenactment. The most related work AnimeCeleb [13] requires constructing a dataset with pose vector and cartoon image pairs by animating 3D characters, which makes it inapplicable anymore if no paired data is available. In this paper, we propose a novel method for cross-domain reenactment without paired data. Specifically, we propose a transformer-based framework to align the motions from different domains into a common latent space where motion transfer is conducted via latent code addition. Two domain-specific motion encoders and two learnable motion base memories are used to capture domain properties. A source query transformer and a driving one are exploited to project domain-specific motion to the canonical space. The edited motion is projected back to the domain of the source with a transformer. Moreover, since no paired data is provided, we propose a novel cross-domain training scheme using data from two domains with the designed analogy constraint. Besides, we contribute a cartoon dataset in Disney style. Extensive evaluations demonstrate the superiority of our method over competing methods.", + "bbox": [ + 75, + 335, + 473, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 804, + 209, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Online video conferences and live streaming are spreading rapidly since the population of smartphones and communication techniques. Speakers sometimes present them", + "bbox": [ + 76, + 830, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/61b8aaf7d892918700650c5e39005a92db18b3cab23873327a4b3b758e625819.jpg", + "image_caption": [ + "Figure 1. Cross-domain reenactment examples of our method. Given a cartoon face, we animate it by transferring the pose and expression from a real face." + ], + "image_footnote": [], + "bbox": [ + 501, + 303, + 890, + 483 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "selves as cartoon characters for privacy and entertainment consideration. However, conventional cartoon animation depends on a complex pipeline of 3D modeling and retargeting based on a powerful render engine, which always takes a long period [24]. This limitation makes it difficult to drive a new model or directly drive a single comic picture.", + "bbox": [ + 496, + 553, + 890, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are many works [17, 19, 10, 23, 27] that focus on one-shot talking face generation, i.e., driving a portrait with a video of a real person, which belongs to within-domain reenactment. Their models are trained with a large amount of talking videos and perform well in the same-identity and cross-identity driving tasks. However, when those models are straightforwardly applied to cross-domain reenactment, i.e., driving a cartoon face with a real video as in Fig. 1, they persistently encounter the following issues, including inaccurate expression transfer, blur effects, and even apparent artifacts around facial components and the background. The reason is the distribution shift between different domains. There is a large appearance and motion gap between the cartoon and real datasets. Though fine-tuning a pre-trained model on a cartoon dataset can improve the visual quality to a certain extent, those issues still cannot be wiped out.", + "bbox": [ + 496, + 643, + 892, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Only a few methods focus on visually-driven cross", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 94, + 886, + 227, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7690", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "domain face reenactment. The early method RecycleGAN [1] overfits two videos for unsupervised video translation, which cannot be used for one-shot reenactment. MAA [28] uses keypoints as an intermediate representation and designs an angle consistency loss to mimic the driving motion. But only focusing on keypoints will result in ignoring the subtle motion of facial components, especially the mouth and eyes. AnimeCeleb [13] is based on pre-defined 3D Morphable Models (3DMMs) [2] and additional pose annotations1 of cartoon images. However, it is inapplicable when paired data is not available, which limits its range of use. Moreover, the pre-defined pose vector has limited capability of capturing fine-grained facial expressions.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel method for cross-domain reenactment without the requirement of paired data. All we need is a set of real videos and another set of cartoon videos. To solve the domain shift problem, we propose a transformer-based framework to align the motions from two domains in a shared canonical latent space. Specifically, two motion encoders and two motion bases are designed to discover the domain-specific properties of appearance and motion. To align the domain-specific motions in a common latent space, a source query transformer and a target take the responsibility of projecting the motions into a canonical motion space, where motion transfer can be conducted via the addition of motion codes. The edited motion is then projected from the canonical space back to the domain of the source by a transformer. A generator takes the projected motion and the source features as input to synthesize a face in the domain of the source.", + "bbox": [ + 75, + 287, + 472, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, to overcome the lack of paired data, we propose a novel cross-domain training scheme. The core idea is the analogy constraint (see Fig. 2). Specifically, given two real faces and a cartoon face, we use the relative motion between the two real faces to drive the cartoon one, producing a synthetic cartoon face. The relative motion from one real face to the other real one is analogous to that from the cartoon face to the synthetic one. We measure the distance between the two types of analogical motions in the aligned canonical motion space. They are supposed to be identical.", + "bbox": [ + 75, + 544, + 472, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To conduct cross-domain reenactment experiments, we collect a cartoon video dataset in Disney style, which contains 344 videos and about 47k frames in total. The cartoon dataset will be released for research purposes.", + "bbox": [ + 75, + 695, + 472, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows:", + "bbox": [ + 96, + 756, + 316, + 770 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel framework for cross-domain face reenactment without using paired data. Several transformers are designed to align the motions of different domains in a common latent space.", + "- We propose a training scheme to compensate for the" + ], + "bbox": [ + 94, + 780, + 468, + 867 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0383b6a34e320079cd79497edce396c30ecc481b2957108b08d23cca1d25549a.jpg", + "image_caption": [ + "Figure 2. Visualization of analogy constraint. $\\Delta \\omega_{real}$ and $\\Delta \\omega_{cartoon}$ represent the relative motion in the real and cartoon domain, respectively. $\\Delta \\omega_{can}$ represents the projected motion in the canonical space where we can perform cross-domain motion transfer by the addition of motion codes. Given a cartoon face, we utilize the relative motion between two real faces to animate the cartoon one and generate a synthetic cartoon face. We hold that the relative motion between two real faces and that between the cartoon face and the synthetic are equivalent within the canonical motion space." + ], + "image_footnote": [], + "bbox": [ + 545, + 89, + 844, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lack of paired data by using the analogy constraint.", + "bbox": [ + 531, + 446, + 870, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We collect a cartoon video dataset and conduct extensive experiments to demonstrate the superiority of our method in the cross-domain reenactment.", + "bbox": [ + 517, + 473, + 890, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 532, + 640, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Within-domain Face Reenactment", + "text_level": 1, + "bbox": [ + 500, + 559, + 799, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most one-shot talking face generation works focus on within-domain face reenactment, i.e., driving a portrait image using another talking head video, where both the driving video and source image is from real humans.", + "bbox": [ + 496, + 582, + 890, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keypoint-based Methods. Many reenactment algorithms [26, 30, 31, 5] are based on pre-defined facial landmarks that are used for motion conversion and repositioning. For example, ReenactGAN [26] encodes the face into the boundary hidden space defined by the keypoint heat map and transfers motion by a boundary transformer. Although ReenactGAN can produce high-quality target faces, it needs to retrain the model for the new individual images. FReeNet [31] trained a keypoint converter to convert the driven expression onto the source face. However, it can only deal with expression change, but not attitude changes, so the generated face image will keep the pose of the source face.", + "bbox": [ + 496, + 643, + 892, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3DMM-based Methods. A number of works [17, 7, 4, 29, 33] take advantage of the 3D deformation model [3] that can decouple pose, expression, and identity. For instance, Pirenderer [17] mapped the target 3DMM parameters into hidden codes and injected them into the source", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Note that Pose in [13] refers to a 20 dimension (20D) vector that contains the status of facial expression and head pose.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7691", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image features to guide the generation of distorted streams. HeadGAN [7] renders 3D parameters to images as motion representations. [4] map 3DMM parameters to StyleGAN [11, 12] latent space, and a transformation matrix is learned to map 3DMM parameters to latent code. Style-HEAT [29] used 3DMM to generate an optical flow map to change the feature map of StyleGAN. Using 3DMM parameters as motion representation can avoid the problem of video identity information leakage in cross-identity reenactment, but the 3D model only encodes facial regions, such as teeth, eyes, hair and other information that is difficult to synthesize.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised Methods. Another large class of methods [25, 18, 19, 22, 10, 23, 8] do not use existing face models, but use unsupervised training to obtain facial motion representations. A typical method is FOMM [19], which proposes a first-order motion method by representing the motion as the key points and its local transformation. Facevid2vid [22] further extends this method to 3D facial key points. DAGAN [10] used the self-supervised geometry learning method to obtain the depth map from the video. Motion transformer [21] uses the interactive capabilities of transformer to predict key points. Differing from the above methods, LIA [23] does not require explicit structural representation but directly manipulates the latent space of the generative model. Megapotrait [8] follows a similar way and adds a super-resolution network for higher resolution.", + "bbox": [ + 75, + 273, + 472, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goals of these within-domain reenactment methods are different from ours since they do not need to consider the domain gap. They can be trained in a self-supervised way from the same video, while there are no cartoon-real training pairs in our cross-domain settings.", + "bbox": [ + 76, + 501, + 472, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Cross-domain Face Reenactment", + "text_level": 1, + "bbox": [ + 76, + 588, + 366, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Only a few methods focus on cross-domain face reenactment. The early study Recycle-GAN [1] is based on image translation networks. It introduces a recurrent loss by training a temporal predictor to monitor the temporal consistency. And each character requires two face videos as two domains for training, which cannot be used for one-shot reenactment. Everythingtalking [20] proposes a parametric shape modeling technique parametric models based on Bezier Curve and transfers motion by adapting the human face's control points to the pareidolia face. But it is unable to change the facial posture. MAA [28] is a method that relies on keypoints and incorporates an angle consistency loss to replicate driving motion. However, the differences in keypoints only account for absolute motion transfer, which might significantly alter the appearance of source images and fail to capture intricate facial expressions well.", + "bbox": [ + 75, + 612, + 468, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The most related work AnimeCeleb [13] proposes a new cartoon talking-head video dataset. It uses a 3D animation model to generate a large number of animated face im", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ages and corresponding pose annotations. To achieve cross-domain reenactment, it proposes to learn the mapping relationship from animation pose vectors to 3DMM parameters by paired data. One shortcoming of AnimeCeleb is that it becomes inapplicable if no paired data is available, which limits its application scope. Moreover, it cannot be generalized to other cartoon styles that require rich expressions. Because AnimeCeleb uses the manually designed 20D vector to represent the status of pose and expression, which is far from enough. Differently, our method does not depend on 3DMMs and can be trained without being paired by using a designed cross-domain training scheme.", + "bbox": [ + 496, + 90, + 893, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 284, + 591, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a novel transformer-based framework to align the motions from different domains in a shared canonical motion space for motion transfer. The framework is shown in Fig. 3. The task becomes simple if there exist paired data, i.e., faces in different domains on the same expression and pose. Unfortunately, no such dataset is available. To settle the lack of paired data, we propose a training scheme with the designed analogy constraint, using faces from two domains jointly. We first introduce the architecture of the framework in Sec. 3.1. Then, we present the training scheme in Sec. 3.2. Finally, we show a two-stage learning strategy to alleviate the issue of imbalanced training data in Sec. 3.3, i.e., the number of real videos is much larger than that of cartoon videos.", + "bbox": [ + 496, + 310, + 893, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Transformer-based Framework", + "text_level": 1, + "bbox": [ + 500, + 530, + 779, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 3, given a source image and a driving one from two domains, we extract the motion information from the two images with two domain-specific motion encoders. Then, a source query transformer and a driving one are used to project the extracted motion information from different domains into the shared canonical motion space, with the domain-specific motion code base and the motion information as input. The motion transfer is realized by adding motion codes in the shared motion space. A domain-specific backward transformer projects the edited motion code back to the domain of the source image with the assistance of its features extracted via a domain-specific appearance encoder. Finally, a generator takes the projected motion code and the source features as input to render an image in the domain of the source.", + "bbox": [ + 496, + 553, + 893, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Please note that we use the cartoon as the source and the real as the driving for illustration in Fig. 3. We can exchange the roles of the cartoon and the real, i.e., the real as the source while the cartoon is the driving. Therefore, in our full framework, we have two appearance encoders, two motion encoders, two backward transformers, two generators, and two motion code bases, which are domain-specific. We also have one source query transformer and one driv-", + "bbox": [ + 496, + 780, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7692", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0777385d3f6a2231b65bfba5ed2ce616432d0731491f57c0b85a6c0be50fab7e.jpg", + "image_caption": [ + "Figure 3. An illustration of our model, which can be mainly divided into three parts: (1) Feature extraction. We use $E_{a}$ and $E_{m}$ to extract local features and latent motion descriptor respectively. (2) Latent space transformation. A source query transformer and a driving one are used to project the motion descriptor into a canonical motion space. Motion transfer is conducted via the addition of motion codes in the shared motion space. (3) Image generation. The edited motion is projected back to the domain of the source with the assistant of the source features. A progressive image generator is used to render an image with the projected motion. We also apply a spatial feature transformation (SFT) to the last three high-resolution features to reduce the artifacts caused by feature warping." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 509, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/09459a17696775d8e6964ca82fe3ddf068706635ead81e9f0b3e63e29fc2c8bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 88, + 696, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c9ef07372e0a0ed88e1e5436431cc756cafaaaa3668728268b63842b3f57c9d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 89, + 885, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing query transformer, which are shared by two domains. The bounding box in color indicates whether a component is domain-specific in Fig. 3. We present the components in detail as follows.", + "bbox": [ + 75, + 421, + 468, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Appearance and Motion Encoders. We have domain-specific encoders to extract appearance and motion information from an image. Let $E_{a}^{C}$ and $E_{m}^{C}$ denote the appearance and motion encoder for the cartoon domain, respectively, while $E_{a}^{R}$ and $E_{m}^{R}$ for the real domain. As the roles of the two domains as the source and the target can be exchanged, we use $E_{a}$ and $E_{m}$ for simplicity. The appearance encoder takes the source image as input and outputs multi-scale spatial features, while the motion encoder extracts motion descriptors from the source and driving image, respectively. Thus, the feature extraction can be defined as:", + "bbox": [ + 75, + 481, + 468, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS = E _ {a} \\left(x _ {s}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 657, + 468, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {s} = E _ {m} \\left(x _ {s}\\right), F _ {d} = E _ {m} \\left(x _ {d}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 676, + 468, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F_{s}$ and $F_{d}$ denote the motion descriptors of the source $x_{s}$ and the driving $x_{d}$ . $S = \\{S^{i}\\}_{i = 1}^{N}$ represents the appearance features of the source.", + "bbox": [ + 75, + 703, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion Query Transformer. In addition to the domain-specific motion encoder, we use a motion base for each domain for further discovering domain-specific motion properties. The motion base consists of learnable embeddings to memorize typical motions, denoted as $p = \\{p^k\\}_{k=1}^K$ , where $p_k \\in R^d$ represents the $k$ -th motion embedding and $d$ is the dimension. We adopt transformer blocks to project the latent descriptor into a canonical motion space due to their capability of capturing long-range dependencies. As shown in the middle of Fig. 3, the embedding is linearly projected", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to the query features while the latent descriptor from the image is projected to the key and value features. Inspired by [19], since it is challenging to model motion transformation $x_{s} \\rightarrow x_{d}$ straightforwardly, we choose to model transformation $x_{s} \\rightarrow x_{r}$ and $x_{r} \\rightarrow x_{d}$ separately by leveraging a virtual reference $x_{r}$ :", + "bbox": [ + 496, + 421, + 890, + 512 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {s \\rightarrow d} = w _ {s \\rightarrow r} + w _ {r \\rightarrow d}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 526, + 890, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{s\\rightarrow r} = T_{s\\rightarrow r}(F_s,p)$ and $w_{r\\rightarrow d} = T_{r\\rightarrow d}(F_d,p)$ . $w_{s\\rightarrow r},w_{r\\rightarrow d}$ , and $w_{s\\rightarrow d}$ are three motion codes in canonical motion space. $T_{s\\rightarrow r}$ denotes the source query transformer that models the motion transformation from the source to the reference. While $T_{r\\rightarrow d}$ denotes the driving query transformer that models the motion transformation from the reference to the driving. The two query transformers are shared by the two domains.", + "bbox": [ + 496, + 551, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Backward Transformer. As we obtain the motion transformation $w_{s\\rightarrow d}$ in the canonical motion space, the next step is to project the motion code back to the domain of the source image. We use a domain-specific backward transformer to take over this task with the assistance of the information of the source image, i.e., the multi-scale features $S$ of the source image from the appearance encoder $E_{a}$ . We have two such transformers, i.e., $T_B^C$ and $T_B^R$ , and take the notation $T_{B}$ for simplicity. To let the motion code obtain information from image features in different resolutions, we design the backward transformer as a stack of three transformer blocks. As shown on the right of Fig. 3, in each transformer block, the query features are derived from the output of the previous module, while the key and value features are computed from the corresponding multi-scale fea", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7693", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tures. The first query features are the linear projection of the motion transformation $w_{s\\rightarrow d}$ from the canonical space. And the resolutions of the feature maps are $16\\times 16$ , $32\\times 32$ , and $64\\times 64$ , respectively.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Generator. After obtaining the projected motion code in the source domain, we use a domain-specific generator to progressively warp and refine the source feature maps layer by layer conditioned on the motion code $w_{s \\to d}$ . We have two generators, i.e., $G^{C}$ and $G^{R}$ , and take the notation $G$ for simplicity. Considering that warping them only will make the final output suffer from artifacts, we add refinement modules in the last three layers of the generator to enhance three high-resolution feature maps. Given the motion code $w_{s \\to d}$ and the previous feature maps $o^{i-1}$ , a StyleConv block is used to predict a flow field that is used to warp the feature map $S^{i}$ . With the warped feature map $f_{w}^{i}$ and the motion code $w_{s \\to d}$ as input, the spatial feature transformation (SFT) for refinement is defined as:", + "bbox": [ + 75, + 152, + 470, + 363 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\no ^ {i} = \\alpha^ {i} \\odot f _ {w} ^ {i} + \\beta^ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 369, + 468, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha^i = \\mathrm{StyleConv}(f_w^i, w_{s \\to d})$ is the spatial scale and $\\beta^i = \\mathrm{StyleConv}(f_w^i, w_{s \\to d})$ is the bias.", + "bbox": [ + 76, + 395, + 468, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Cross-domain Training Scheme", + "text_level": 1, + "bbox": [ + 76, + 433, + 354, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If there are paired data where cartoon and real video pairs share the same head pose and expression, cross-domain training can be settled simply following the way of training a conventional within-domain model. However, no such paired data is available. To tackle this issue, we propose an analogy constraint to train the model with unpaired cartoons and real videos.", + "bbox": [ + 75, + 457, + 468, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analogy Constraint. As shown in Fig. 2, the core idea is to project the relative motions of different domains into the canonical motion space. For instance, given a tuple consisting of two real faces and a cartoon one, we can calculate the relative motion from one real face to the other real one. Then, the relative motion can be used to animate the cartoon face to generate a synthetic cartoon face. We can compute the relative motion from the cartoon face to the synthetic cartoon one. The two types of relative motions are supposed to be identical in canonical motion space.", + "bbox": [ + 75, + 563, + 468, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the canonical motion space, following [19], the relative motion transfer can be defined as:", + "bbox": [ + 76, + 714, + 468, + 743 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {s \\rightarrow d} = \\left(w _ {s \\rightarrow r} + w _ {r \\rightarrow s}\\right) + \\left(w _ {r \\rightarrow d} - w _ {r \\rightarrow 1}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 753, + 468, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where '1' refers to a selected frame in the driving video. For illustration, let $x_{a}^{R}$ and $x_{b}^{R}$ denote two real images and $x_{s}^{C}$ denote a cartoon image. Then, the relative motion transfer is applied to the cartoon, which can be defined as:", + "bbox": [ + 76, + 776, + 468, + 838 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {s \\rightarrow g} = \\left(w _ {s \\rightarrow r} + w _ {r \\rightarrow s}\\right) + \\left(w _ {r \\rightarrow b} - w _ {r \\rightarrow a}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 845, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{s\\rightarrow g}$ represents the motion transformation from $x_{s}^{C}$ to $x_{g}^{C}$ . 'g' indicates the synthetic cartoon face. Each item", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "can be defined as follows:", + "bbox": [ + 500, + 90, + 673, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {s \\rightarrow r} = T _ {s \\rightarrow r} \\left(E _ {m} ^ {C} \\left(x _ {s} ^ {C}\\right), p ^ {C}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 112, + 795, + 130 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {r \\rightarrow s} = T _ {r \\rightarrow d} \\left(E _ {m} ^ {C} \\left(x _ {s} ^ {C}\\right), p ^ {C}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 133, + 890, + 159 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {r \\rightarrow a} = T _ {r \\rightarrow d} \\left(E _ {m} ^ {R} \\left(x _ {a} ^ {R}\\right), p ^ {R}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 152, + 795, + 171 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {r \\rightarrow b} = T _ {r \\rightarrow d} \\left(E _ {m} ^ {R} \\left(x _ {b} ^ {R}\\right), p ^ {R}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 175, + 795, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the motion code $w_{s \\to g}$ as well as the source features $S$ , the generator $G^{C}$ can produce a synthetic cartoon face $x_{g}^{C}$ , i.e., $x_{g}^{C} = G^{C}(T_{B}^{C}(w_{s \\to g}), E_{a}^{C}(x_{s}^{C}))$ . Then, we can calculate the motion code of the synthetic cartoon face in the canonical motion space, i.e.,", + "bbox": [ + 500, + 200, + 890, + 276 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {r \\rightarrow g} = T _ {r \\rightarrow d} \\left(E _ {m} ^ {C} \\left(x _ {g} ^ {C}\\right), p ^ {C}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 285, + 890, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The relative motion between the cartoon face and the synthetic one is $(w_{r\\rightarrow g} - w_{r\\rightarrow s})$ , which is supposed to be identical to the relative motion between the two real faces, i.e., $(w_{r\\rightarrow b} - w_{r\\rightarrow a})$ . Therefore, we can measure the distance between the two types of relative motions by", + "bbox": [ + 498, + 311, + 890, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m o t i o n}} ^ {R \\rightarrow C} = \\mathbb {E} [ \\| \\left(w _ {r \\rightarrow g} - w _ {r \\rightarrow s}\\right) - \\left(w _ {r \\rightarrow b} - w _ {r \\rightarrow a}\\right) \\| _ {1} ]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 396, + 890, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, given two cartoon faces and one real face, we have $\\mathcal{L}_{motion}^{C\\rightarrow R}$ following the same formulation. Further, to maintain the high fidelity of face generation, we impose adversarial loss. The adversarial loss can be defined as:", + "bbox": [ + 498, + 429, + 890, + 488 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} ^ {R} = \\mathbb {E} _ {x _ {g} ^ {R}} [ - \\log (D ^ {R} (x _ {g} ^ {R})) ], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 494, + 890, + 522 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a d v} ^ {C} = \\mathbb {E} _ {x _ {g} ^ {C}} [ - \\log (D ^ {C} (x _ {g} ^ {C})) ],\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 517, + 795, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D^{R}$ and $D^{C}$ are two discriminators for the real and cartoon domains, respectively.", + "bbox": [ + 498, + 547, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reconstruction. Besides the analogy constraint that emphasizes the interaction between two domains, we can also perform within-domain reenactment to assist the model learning. Specifically, we can sample two images within the same domain as the source and the driving, e.g., a pair of $x_{s}^{R}$ and $x_{d}^{R}$ or a pair of $x_{s}^{C}$ and $x_{d}^{C}$ . The reconstruction loss of the real pairs can be used to update the parameters of the real domain-related components as well as the shared motion query transformers. Similarly, the reconstruction loss of the cartoon pairs is for the corresponding components. We exploit the $L_{1}$ loss $\\mathcal{L}_{1}(x_{s\\rightarrow d},x_{d})$ , the perceptual loss $\\mathcal{L}_{per}(x_{s\\rightarrow d},x_{d})$ , and the GAN loss $\\mathcal{L}_{adv}(x_{s\\rightarrow d},x_{d})$ . Hence, the objective of reconstruction can be defined as:", + "bbox": [ + 498, + 579, + 890, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c}} = \\mathcal {L} _ {1} + \\lambda_ {\\text {p e r}} \\mathcal {L} _ {\\text {p e r}} + \\mathcal {L} _ {\\text {a d v}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 786, + 890, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we omit the domain notation R&C for simplicity.", + "bbox": [ + 500, + 814, + 867, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall objective is defined as follows:", + "bbox": [ + 517, + 830, + 803, + 844 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {r e c} ^ {R} + \\mathcal {L} _ {r e c} ^ {C} + \\lambda_ {m} * \\left(\\mathcal {L} _ {\\text {m o t i o n}} ^ {R \\rightarrow C} + \\mathcal {L} _ {\\text {m o t i o n}} ^ {C \\rightarrow R}\\right) + \\mathcal {L} _ {\\text {a d v}} ^ {C} + \\mathcal {L} _ {\\text {a d v}} ^ {R}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 852, + 890, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{m}$ is the trade-off hyperparameter.", + "bbox": [ + 500, + 886, + 779, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7694", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/15f70bb361daaa1a641397abcd294b5cbd9ffa1ef1ee91e79cd5130122b76def.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 90, + 196, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3b46cfe5329d3f720eb2f6d9278e006515ea869530013db651fab816b7cddeb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 90, + 312, + 180 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/87aa5dee397bdaa1b0a102b7e0994dfd648455158baea16f6bde68530b429b05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 90, + 426, + 180 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/136da6006347cc990177518b3b3c6d931299a4e1309d8ebb0534bee71f20edd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 90, + 542, + 180 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6c069edc8893e8c0e0ba48f7c0963c3244409d1404eaecffa3a0ac858d6cc1a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 90, + 656, + 180 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b509db33864fe08b4a0bd05d6649fe287039a2dfd9f8d9e143f0e92af29424fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 90, + 774, + 180 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cc699f785c6cf9dbeb8ea4cc2249ee97a3917f4509a22d59b558dabe1d833afb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 90, + 888, + 180 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/746e393aad85bcbeb019c27c71fd14734277542152dd1e7ce67ddf05cf4ee0aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 181, + 196, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f89aa8696d50fee10818afc819880c41ad83b1959f622fc1e5a8a09effe9d1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 181, + 310, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f368ecf62861dfca60a7bb8b97249ccd96510c391a0d4a018f9df881a74a3a0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 181, + 426, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71c21fef78925ed4f31b9298c72139b2c9db077266f489e52aec4ba5f41bc042.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 181, + 542, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/697f6cf14c2953d83f79e9e7860d3fcf212600884e66d567aa9497e88865b068.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 181, + 656, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3fc7ff01840344ed3ed15b4cc58c12bf931bf8879b0a6dd4e25d0d385581e0b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 181, + 774, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55bb4d1f34ce718754f447c43f0852e5a0ebbab9fddc547fc1dbc6753c7e9ca1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 181, + 888, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/447a4e4982aa89abc459274d2d6d1532f7a734a2fddef937924e783b1effac32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 270, + 196, + 357 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43ba0535483f2f922220a5734f3b26a5b2c646d4bd1b6f4d4c3f5019c1796855.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 270, + 310, + 357 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/edef89eb42e5eca7ef21b6bafbb9d909df8244cb8d902561b9f317a5d4a54fc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 312, + 270, + 426, + 357 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c740cac3ad7a2624609f08ba34c21ee1d79970bd68e7b3173c6736f558b0e689.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 270, + 542, + 357 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ddfab59f7739ab470275329d3f79433e26cabeef8672a2e0b6446be49a0206ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 270, + 656, + 357 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/042bf4f241b73434e9bc458d4a163ac99a6e7dc6aa7d514ff09d85406eaab4d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 270, + 772, + 357 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b00c0afab42688f0b6c36ae790193fcd468c5adf8c3ea37b44e8f2185cc0668a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 270, + 888, + 357 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/820948162f062253a2961b5676b81ea2d0256bf2c2f45ff6c0cec199e9a191fe.jpg", + "image_caption": [ + "(a) Source" + ], + "image_footnote": [], + "bbox": [ + 80, + 358, + 196, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d9380c89cebe77dffccd731670d636bf9e8a2264eefa4f1bf7a26cabfa89dc3.jpg", + "image_caption": [ + "(b) FOMM" + ], + "image_footnote": [], + "bbox": [ + 197, + 358, + 310, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ca83750f818e07c2d3be39b96e6fe5e895f404a96e766823a3eef4284b7126e.jpg", + "image_caption": [ + "(c) LIA" + ], + "image_footnote": [], + "bbox": [ + 312, + 358, + 426, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd80d55c042ea52f16861b81b2a0a689541b41a943711f3152dffb222cb0142f.jpg", + "image_caption": [ + "(d) DaGAN", + "Figure 4. Qualitative comparisons with state-of-the-art methods on cross-domain reenactment. Competing methods are finetuned with the cartoon dataset. Hence all models are trained on the same data. It can be seen that our results have superior image quality and motion consistency. More visual results are in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 426, + 358, + 542, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6ad27e5292613041d305a7a92c8b4dd5afb534b6abc86665443122629d7f9531.jpg", + "image_caption": [ + "(e) Ours" + ], + "image_footnote": [], + "bbox": [ + 542, + 358, + 656, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/53aaaf1480412b5106be672953983cd06feed2dbdeb5ac559b2b157554ecb779.jpg", + "image_caption": [ + "(f) Start" + ], + "image_footnote": [], + "bbox": [ + 656, + 358, + 774, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e3a3b10f91610c298080c8d7d5a9c7650cf9353cf8b34460403e2200acd71e8a.jpg", + "image_caption": [ + "(g) Driving" + ], + "image_footnote": [], + "bbox": [ + 774, + 358, + 888, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Two-stage Learning on Imbalanced Data.", + "text_level": 1, + "bbox": [ + 76, + 536, + 429, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The number of real videos is much larger than that of cartoon videos. Directly training with mixed imbalanced data cannot promote the performance of cross-domain reenactment because real faces dominate the data distribution. To alleviate the issue of the imbalanced dataset, we propose to learn the model in two stages.", + "bbox": [ + 75, + 569, + 468, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the first stage, we train the model with real videos by performing the reconstruction task, through which the model can learn the general knowledge of face motion from a large number of talking head videos. The obtained model cannot accurately transfer motion from real faces to cartoon ones due to domain shifts, especially facial expressions.", + "bbox": [ + 75, + 670, + 468, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the second stage, we copy the domain-specific components of the real domain as initialization for those of the cartoon domain. Then, we jointly learn all the components in the framework with the analogy constraint and the reconstruction loss.", + "bbox": [ + 75, + 770, + 468, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Please note that our final model can be used to reenact a cartoon with a real video as well as to reenact a real face with a cartoon video.", + "bbox": [ + 75, + 854, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 536, + 630, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 500, + 564, + 596, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. For cross-domain reenactment, we collect a cartoon video dataset in Disney style which contains 344 videos with an FPS of 30 and the total number of frames is about $47\\mathrm{k}$ . To train our model, we take a subset of videos from the VoxCeleb [15] dataset containing 17K videos, and a subset from self-collected cartoon videos containing 289 videos. For evaluation, we use the rest 473 videos from the VoxCeleb [15] dataset and 55 videos from the cartoon dataset. We also use a set of images from CelebA-HQ [14] for cross-identity reenactment evaluation.", + "bbox": [ + 496, + 592, + 890, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We evaluate the performance from four aspects, i.e., reconstruction, fidelity, identity preservation, and motion transfer. $L_{1}$ , Learned Perceptual Image Patch Similarity (LPIPS)[32], and Peak Signal-to-Noise Ratio (PSNR) are used to measure the reconstruction accuracy. Frechet Inception Distance (FID) [9] and Cumulative probability of Blur Detection (CPBD) [16] are used for measuring generation realism. The cosine similarity (CSIM) between synthetic and source images through ArcFace [6] is used for identity preservation. To evaluate the accuracy of mo", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7695", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/61f6b0f7df7251baed3e5399b2db63c86febbfb7d86d022898b7120db9aad243.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons with MAA [28]. Our method can better preserve the face shape of the original image. And more accurately imitated the driving image, especially the eyes and mouth." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 460, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion transfer, following [10], we use Average Keypoint Distance (AKD) and Average Expression Distance (AED) to compute the distance between synthetic and target images in terms of expression and pose.", + "bbox": [ + 75, + 369, + 468, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Competing methods. We compare our method with several recent state-of-the-art methods, including FOMM [19], DAGAN [10], and LIA [23]. FOMM is a typical motion transfer method based on unsupervised keypoints. DAGAN introduces depth into the estimation of key points and has a similar formulation of motion transfer as FOMM. LIA is a representative method that uses a reference space similar to FOMM but conducts motion transfer in a latent space. For a fair comparison, we initialize the models using their pretrained official models and finetune them on the cartoon dataset. Therefore, we learn with the same data. Furthermore, we also compare with MAA [28], a keypoint-based cross-domain motion transfer method.", + "bbox": [ + 75, + 435, + 468, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. The model is trained in two stages to handle the imbalanced dataset where real videos are much more than cartoons. In the first stage, we trained the model on the training set of VoxCeleb for 450K iterations, where the batch size is 64 and the optimizer is Adam with an initial learning rate of $5 \\times 10^{-4}$ . In the second stage, we fine-tune the model with Disney videos with 10K iterations, where the batch size is 8 and the optimizer is Adam with an initial learning rate of $1 \\times 10^{-4}$ . The hyperparameters are set as $\\lambda_{per} = 10$ and $\\lambda_{m} = 20$ to balance generative fidelity and motion consistency. A detailed description of model components is in the supplementary.", + "bbox": [ + 75, + 635, + 468, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Cross-Domain Face Reenactment", + "text_level": 1, + "bbox": [ + 76, + 830, + 369, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As mentioned in Sec. 3.3, our method can handle the bidirectional cross-domain reenactment. We execute motion transfer task for evaluation i.e., real $\\rightarrow$ cartoon", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/848961eeb304ca2c4bc49c97d198fd47a3d8d524188137313e7edd50b0152306.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Real→CartoonCartoon→Real
FID↓CPBD↑FID↓CSIM↑CPBD↑
FOMM30.3690.043651.9290.5360.0626
LIA30.4420.049348.6110.5150.0727
DaGAN29.7790.073547.2220.5610.0985
Ours28.5310.084031.8520.6550.1230
", + "bbox": [ + 501, + 89, + 893, + 190 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison on cross-domain reenactment.", + "bbox": [ + 500, + 200, + 879, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and cartoon $\\rightarrow$ real. The former represents driving a cartoon face with a real video, while the latter represents driving a real face with a cartoon video. As for real $\\rightarrow$ cartoon, 60 videos from VoxCeleb are used to drive 300 cartoon images from the cartoon dataset. For a fair comparison, all competing methods are fine-tuned on the cartoon training set. As for cartoon $\\rightarrow$ real, 55 cartoon videos are employed to drive 275 images of different identities from CelebA-HQ [14].", + "bbox": [ + 496, + 242, + 890, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Evaluation. The results of real $\\rightarrow$ cartoon are shown in the first two rows of Fig. 4. First, our method achieves better image quality than other methods in terms of sharpness and artifacts. Second, our method outperforms others in motion transfer, especially facial expression. For instance, in the first row, the eyes of the cartoon should be closed according to the driving image. Only our method mimics driving while the eyes of others stay open.", + "bbox": [ + 496, + 381, + 890, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results of cartoon $\\rightarrow$ real are shown in the last two rows of Fig. 4. We can observe a similar phenomenon as the real $\\rightarrow$ cartoon case in terms of image quality, where our method works the best. FOMM and DaGAN become underperformed when there is a large head pose difference between the source and driving, as shown in the third and fourth rows. LIA is better than them but still worse than ours. For motion and identity preservation, our method surpasses others.", + "bbox": [ + 496, + 503, + 890, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The comparison with the cross-domain motion transfer method MAA [28] is shown in Fig. 5. As the source code of MAA is not released, its results are adapted from the original paper. Our method has better, especially in the mouth, eyes and image quality. For instance, in the third column, MAA ignores the motion of the eyes, whereas our method accurately mimics the driving image. In the last column, our mouth shape is more consistent with the driving one.", + "bbox": [ + 496, + 641, + 890, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Evaluation. The quantitative results are reported in Tab. 1. Our method achieves the best FID in the two tasks, indicating that the synthetic results are most consistent with the source distribution. Our method also performs the best in sharpness and identity preservation, i.e., the highest CPBD and CSIM, which is consistent with the observation in visual results. These results demonstrate that our model can capture the appearance and motion properties of different domains and the motion transfer in the canoni", + "bbox": [ + 496, + 763, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7696", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5a34432666915a490b55848cd7685217972d53067b98ea48d4a616ef37f330b.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons on within-domain reenactment. Our method achieves better image quality than other methods in terms of sharpness, distortion and artifacts." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 467, + 294 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7b3a09333358a71e5823ffffd57b5d7a34bb018e6c4961b7498a713bd1e033af.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
L1↓LPIPS↓CPBD↑AKD↓AED↓
FOMM0.04660.1390.03081.3880.142
LIA0.04610.1230.03721.5120.153
DaGAN0.04420.1220.04151.3000.129
Ours0.04570.1210.04291.3350.139
", + "bbox": [ + 84, + 354, + 460, + 441 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f09a5a85c6c8e4996ace405ef7d650a9f584c79c37d18cc54dba138e74de5013.jpg", + "table_caption": [ + "Table 2. Same-identity reconstruction in real domain" + ], + "table_footnote": [], + "table_body": "
FID↓CSIM↑CPBD ↑
FOMM54.4110.5410.0689
LIA53.1610.5220.0787
DaGAN49.9200.5480.1069
Ours48.7130.5260.1132
", + "bbox": [ + 84, + 478, + 460, + 565 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/699f02b23cb4b9fd652d85106482f4742c69916b1c49ed2e26984b24b5c2cc94.jpg", + "table_caption": [ + "Table 3. Cross-identity reenactment in real domain" + ], + "table_footnote": [], + "table_body": "
FOMMLIADAGANOurs
FID↓40.85334.06636.21827.467
CPBD↑0.03680.05230.05980.0846
", + "bbox": [ + 83, + 603, + 464, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Cross-identity reenactment in cartoon domain.", + "bbox": [ + 109, + 669, + 434, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cal motion space works well.", + "bbox": [ + 76, + 710, + 272, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study. To further evaluate the visual quality, we conduct a user study by asking 20 human raters to answer 20 multiple-choice questions. In each question, a rater chooses the best from 4 synthetic cartoon videos generated by the three competing methods and ours. Our method is the most favorable with a selection rate of $72.75\\%$ . While the rates are $13.75\\%$ , $9\\%$ , and $4.5\\%$ for LIA, DaGAN, and FOMM.", + "bbox": [ + 75, + 728, + 468, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Within-Domain Face Reenactment", + "text_level": 1, + "bbox": [ + 76, + 845, + 380, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides the cross-domain evaluation, we also compare our method with competing methods in the scenario of", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3662346031899134b75f2bd9a0390cfe06d5f3fcd137fef5472807059c70b29f.jpg", + "image_caption": [ + "Figure 7. Qualitative ablation study on the analogy constraint. We observe that the image quality and motion consistency is much better by using analogy constraint." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "within-domain reenactment though our method is designed for the cross-domain case. For the same-identity evaluation, we use 200 test videos from VoxCeleb. For cross-identity evaluation, we use the 200 videos to drive 1,000 images from CelebA-HQ. Each video drives five images. The obtained 1,000 videos are used to compute the metrics. Note that we use the released pre-trained models of competing methods here.", + "bbox": [ + 496, + 290, + 890, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The quantitative results are shown in Tab. 2 and Tab. 3. For same-identity evaluation, our method is overall comparable with DaGAN and better than LIA and FOMM. For cross-identity evaluation, our method achieves the best FID and CPBD. Our CSIM is comparable with others.", + "bbox": [ + 496, + 411, + 890, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the visual difference among different methods in same-identity reenactment is marginal, we only present a visual comparison of cross-identity reenactment in Fig. 6. It can be observed that our results have the best image quality, presenting more details in the face area, e.g., eyes and teeth. Besides, our method is comparable with LIA in preserving face shape and sometimes even better (see the last row in Fig. 6).", + "bbox": [ + 496, + 487, + 890, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition, we also perform a cross-identity face reenactment in the cartoon domain. We use 55 cartoon videos to drive 275 cartoon images. Results are shown in Tab. 4. Our method outperforms others in both FID and CPBD.", + "bbox": [ + 496, + 609, + 890, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 680, + 653, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the training scheme, we perform an ablation study by dropping the analogy constraint. The quantitative results are shown in Tab. 5. 'w/o analogy' represents dropping the analogy constraint, i.e., finetuning the obtained model from the first stage with the cartoon dataset using only $\\mathcal{L}_{rec}^{R}$ and $\\mathcal{L}_{rec}^{C}$ in Eq.12. It can be observed that dropping the analogy constraint causes overall performance degradation, especially for the cartoon→real task. The qualitative evaluation is shown in Fig. 7. In the second row, the result without analogy constraints cannot accurately mimic the motion of the driving image. The eyes of the cartoon should be closed, but it is opened when dropping constraints. The dropping operation has a side effect", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7697", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9ce5e5f9a5f4041e23895359174be0ebf9496f0d07a18dce2d13d0ac9b95da52.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Real →CartoonCartoon → Real
FID↓CPBD↑FID↓CPBD↑
our full model28.5310.084031.8520.123
w/o analogy29.2300.074643.3530.107
", + "bbox": [ + 81, + 89, + 486, + 157 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f621967d77daec51e6ea4b3e5f0994c4cc66638e7b55ab90a0aba1e80793e204.jpg", + "table_caption": [ + "Table 5. Ablation study on the analogy constraint." + ], + "table_footnote": [], + "table_body": "
02040100
Real→Cartoon26.88528.53132.09441.936
Cartoon→Real33.46331.85230.17628.166
", + "bbox": [ + 78, + 196, + 480, + 253 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/de91fd4f7254b42fd87c0361761ab765851ae68c843c241afb1a819f6ded3e19.jpg", + "table_caption": [ + "Table 6. Ablation study on hyper-parameter ${\\lambda }_{m}$ ." + ], + "table_footnote": [], + "table_body": "
Fullmlpw/o backw/o sft
Real→Cartoon28.531116.74028.90431.291
Cartoon→Real31.85231.88535.26541.077
", + "bbox": [ + 78, + 292, + 483, + 348 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 7. Ablation study on model architecture.", + "bbox": [ + 133, + 358, + 410, + 372 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "on the motion consistency, which indicates that the analogy constraint contributes.", + "bbox": [ + 76, + 398, + 467, + 428 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "To further analyze the influence of analogy constraint, we perform an ablation on the trade-off hyperparameter $\\lambda_{m}$ . As is shown in Tab. 6, it can be observed that $\\lambda_{m}$ trades off the performance of the two cross-domain tasks, i.e., Cartoon $\\rightarrow$ Real and Real $\\rightarrow$ Cartoon. We empirically choose $\\lambda_{m} = 20$ in practical.", + "bbox": [ + 76, + 430, + 467, + 521 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct ablation studies to demonstrate the effectiveness of several modules in the architecture, i.e., the query transformer, the backward transformer, and the spatial feature transformation. The ablation results are shown in Tab. 7. The metric is FID. \"mlp\" indicates that we replace the query transformer with multi-layer perception (MLP). \"w/o back\" denotes that we drop the backward transformer. And \"w/o sft\" drops the spatial feature transformation module. In Tab. 7, we can observe that query transformer brings significant improvement in the Real $\\rightarrow$ Cartoon case. It verifies that the proposed query transformer can project the latent descriptor from different domains into a canonical motion space. Additionally, removing the backward transformer and sft will damage the quality of generated images.", + "bbox": [ + 75, + 522, + 467, + 733 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We also conduct a visualization experiment to verify the proposed canonical space. Since no paired data is available, we select three real images with different poses and expressions as templates to drive real and cartoon images, generating 450 real image pairs and 150 cartoon pairs. Each pair contains faces before and after driving. We compute their relative motions in the common motion space. In Fig. 8, 0, 1, and 2 represent real pairs, while 3, 4, and 5 represent cartoon pairs. $<0,3>$ , $<1,4>$ , and $<2,5>$ are driven by the same template, respectively. This shows that the motions of the two domains are aligned.", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/37bf8e1ef987eda8b3c0ef1d40c3d203b6dea69936a14ace0f8aa0b8886c52f9.jpg", + "image_caption": [ + "Figure 8. Visualizing the canonical motion space. We visualize the t-SNE plot of motion codes in canonical space. $<0,1,2>$ and $<3,4,5>$ are from the relative motions of real domain and cartoon domain, respectively. And $<0,3>, <1,4>$ , and $<2,5>$ are driven by the same template, respectively." + ], + "image_footnote": [], + "bbox": [ + 586, + 90, + 807, + 262 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/ffb368eead750d1db10c1531e05c4adf1e50196251112240f8e5f1d21a150f8a.jpg", + "image_caption": [ + "Figure 9. Failure cases. We observed that it is challenging to handle extreme poses." + ], + "image_footnote": [], + "bbox": [ + 501, + 357, + 888, + 452 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 508, + 617, + 523 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We propose a novel transformer-based framework for cross-domain face reenactment. To settle the domain shift issue, we propose to align the motions of different domains in a canonical motion space. A set of domain-specific and domain-shared components are designed for motion alignment. Motion transfer can be implemented by adding motion codes in the shared motion space. Since there is no paired data for training, we propose a cross-domain training scheme with the analogy constraint using the two domain datasets. To alleviate the influence of imbalanced data, we use a two-stage learning strategy. Extensive experiments are conducted to demonstrate the superiority of our method over competing methods in cross-domain reenactment.", + "bbox": [ + 496, + 536, + 890, + 731 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations. Like other SOTA models, our model cannot handle extreme poses well. As is shown in Fig 9, our method suffers from artifacts and deformations when the driving image is an extreme pose. This may be due to the fact that extreme poses rarely appear in the dataset.", + "bbox": [ + 496, + 748, + 890, + 824 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgement. This work was partly supported by the National Natural Science Foundation of China (Grant No. U1903213) and the Shenzhen Science and Technology Program (JCYJ20220818101014030).", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7698", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aayush Bansal, Shugao Ma, Deva Ramanan, and Yaser Sheikh. Recycle-gan: Unsupervised video retargeting. In Proceedings of the European conference on computer vision (ECCV), pages 119-135, 2018. 2, 3", + "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In ACM SIGGRAPH, pages 187-194, 1999, 2", + "[3] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5543-5552. ISSN: 1063-6919. 2", + "[4] Stella Bounareli, Vasileios Argyriou, and Georgios Tzimiropoulos. Finding directions in GAN's latent space for neural face reenactment. 2, 3", + "[5] Zhuo Chen, Chaoyue Wang, Bo Yuan, and Dacheng Tao. PuppeteerGAN: Arbitrary portrait animation with semantic-aware appearance transformation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13515-13524. IEEE. 2", + "[6] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 6", + "[7] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoria Sharmanska. HeadGAN: One-shot neural head synthesis and editing. 2, 3", + "[8] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. MegaPortraits: One-shot megapixel neural head avatars. 3", + "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[10] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. 1, 3, 7", + "[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 3", + "[12] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. 3", + "[13] Kangyeol Kim, Sunghyun Park, Jaeseong Lee, Sunghyo Chung, Junsoo Lee, and Jaegul Choo. AnimeCeleb: Large-scale animation CelebHeads dataset for head reenactment. 1, 2, 3", + "[14] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5549-5558, 2020. 6, 7", + "[15] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. VoxCeleb: a large-scale speaker identification dataset. In Interspeech 2017, pages 2616-2620. 6" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Niranjan D Narvekar and Lina J Karam. A no-reference image blur metric based on the cumulative probability of blur detection (cpbd). IEEE Transactions on Image Processing, 20(9):2678-2683, 2011. 6", + "[17] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. PIRenderer: Controllable portrait image generation via semantic neural rendering. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 13739-13748. IEEE. 1, 2", + "[18] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. version: 2.3", + "[19] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. 1, 3, 4, 5, 7", + "[20] Linsen Song, Wayne Wu, Chaoyou Fu, Chen Qian, Chen Change Loy, and Ran He. Everything's talkin': Pareidolia face reenactment. 3", + "[21] Jiale Tao, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Motion transformer for unsupervised image animation. 3", + "[22] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10034-10044. IEEE. 3", + "[23] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. LATIVE IMAGE ANIMATOR: LEARNING TO ANIMATE IMAGES VIA LATENT SPACE NAVIGATION. page 17. 1, 3, 7", + "[24] Wikipedia contributors. Morph target animation — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Morph_target Animation&oldid=1078665970, 2022. [Online; accessed 6-November-2022]. 1", + "[25] Olivia Wiles, A. Sophia Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, volume 11217, pages 690–706. Springer International Publishing. Series Title: Lecture Notes in Computer Science. 3", + "[26] Wayne Wu, Yunxuan Zhang, Cheng Li, Chen Qian, and Chen Change Loy. ReenactGAN: Learning to reenact faces via boundary transfer. 2", + "[27] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 1", + "[28] Borun Xu, Biao Wang, Jinhong Deng, Jiale Tao, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Motion and appearance adaptation for cross-domain motion transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 529-545. Springer, 2022. 2, 3, 7", + "[29] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7699", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yujiu Yang. StyleHEAT: One-shot high-resolution editable talking face generation via pre-trained StyleGAN. 2, 3", + "[30] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9458-9467. IEEE. 2", + "[31] Jiangning Zhang, Xianfang Zeng, Mengmeng Wang, Yusu Pan, Liang Liu, Yong Liu, Yu Ding, and Changjie Fan. FReeNet: Multi-identity face reenactment. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5325-5334. IEEE. 2", + "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[33] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 415 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7700", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/ToonTalker_ Cross-Domain Face Reenactment/d3014894-36d6-4bc9-bdc4-403449ecf9bd_model.json b/2023/ToonTalker_ Cross-Domain Face Reenactment/d3014894-36d6-4bc9-bdc4-403449ecf9bd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..669adc598e8bb1aa5120ac42584853675a2ec2ef --- /dev/null +++ b/2023/ToonTalker_ Cross-Domain Face Reenactment/d3014894-36d6-4bc9-bdc4-403449ecf9bd_model.json @@ -0,0 +1,2620 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.131, + 0.723, + 0.151 + ], + "angle": 0, + "content": "ToonTalker: Cross-Domain Face Reenactment" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.18, + 0.721, + 0.217 + ], + "angle": 0, + "content": "Yuan Gong\\(^{1}\\) Yong Zhang\\(^{2*}\\) Xiaodong Cun\\(^{2}\\) Fei Yin\\(^{1}\\) Yanbo Fan\\(^{2}\\) Xuan Wang\\(^{3}\\) Baoyuan Wu\\(^{4}\\) Yujiu Yang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.217, + 0.816, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\)Shenzhen International Graduate School, Tsinghua University \\(^{2}\\)Tencent AI Lab \\(^{3}\\)Ant Group \\(^{4}\\)The School of Data Science, Shenzhen Research Institute of Big Data, The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen)" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.474, + 0.776 + ], + "angle": 0, + "content": "We target cross-domain face reenactment in this paper, i.e., driving a cartoon image with the video of a real person and vice versa. Recently, many works have focused on one-shot talking face generation to drive a portrait with a real video, i.e., within-domain reenactment. Straightforwardly applying those methods to cross-domain animation will cause inaccurate expression transfer, blur effects, and even apparent artifacts due to the domain shift between cartoon and real faces. Only a few works attempt to settle cross-domain face reenactment. The most related work AnimeCeleb [13] requires constructing a dataset with pose vector and cartoon image pairs by animating 3D characters, which makes it inapplicable anymore if no paired data is available. In this paper, we propose a novel method for cross-domain reenactment without paired data. Specifically, we propose a transformer-based framework to align the motions from different domains into a common latent space where motion transfer is conducted via latent code addition. Two domain-specific motion encoders and two learnable motion base memories are used to capture domain properties. A source query transformer and a driving one are exploited to project domain-specific motion to the canonical space. The edited motion is projected back to the domain of the source with a transformer. Moreover, since no paired data is provided, we propose a novel cross-domain training scheme using data from two domains with the designed analogy constraint. Besides, we contribute a cartoon dataset in Disney style. Extensive evaluations demonstrate the superiority of our method over competing methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.805, + 0.21, + 0.82 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.831, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Online video conferences and live streaming are spreading rapidly since the population of smartphones and communication techniques. Speakers sometimes present them" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.304, + 0.892, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.487, + 0.892, + 0.528 + ], + "angle": 0, + "content": "Figure 1. Cross-domain reenactment examples of our method. Given a cartoon face, we animate it by transferring the pose and expression from a real face." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.645 + ], + "angle": 0, + "content": "selves as cartoon characters for privacy and entertainment consideration. However, conventional cartoon animation depends on a complex pipeline of 3D modeling and retargeting based on a powerful render engine, which always takes a long period [24]. This limitation makes it difficult to drive a new model or directly drive a single comic picture." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.893, + 0.886 + ], + "angle": 0, + "content": "There are many works [17, 19, 10, 23, 27] that focus on one-shot talking face generation, i.e., driving a portrait with a video of a real person, which belongs to within-domain reenactment. Their models are trained with a large amount of talking videos and perform well in the same-identity and cross-identity driving tasks. However, when those models are straightforwardly applied to cross-domain reenactment, i.e., driving a cartoon face with a real video as in Fig. 1, they persistently encounter the following issues, including inaccurate expression transfer, blur effects, and even apparent artifacts around facial components and the background. The reason is the distribution shift between different domains. There is a large appearance and motion gap between the cartoon and real datasets. Though fine-tuning a pre-trained model on a cartoon dataset can improve the visual quality to a certain extent, those issues still cannot be wiped out." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Only a few methods focus on visually-driven cross" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.228, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7690" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "domain face reenactment. The early method RecycleGAN [1] overfits two videos for unsupervised video translation, which cannot be used for one-shot reenactment. MAA [28] uses keypoints as an intermediate representation and designs an angle consistency loss to mimic the driving motion. But only focusing on keypoints will result in ignoring the subtle motion of facial components, especially the mouth and eyes. AnimeCeleb [13] is based on pre-defined 3D Morphable Models (3DMMs) [2] and additional pose annotations1 of cartoon images. However, it is inapplicable when paired data is not available, which limits its range of use. Moreover, the pre-defined pose vector has limited capability of capturing fine-grained facial expressions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.473, + 0.544 + ], + "angle": 0, + "content": "In this paper, we propose a novel method for cross-domain reenactment without the requirement of paired data. All we need is a set of real videos and another set of cartoon videos. To solve the domain shift problem, we propose a transformer-based framework to align the motions from two domains in a shared canonical latent space. Specifically, two motion encoders and two motion bases are designed to discover the domain-specific properties of appearance and motion. To align the domain-specific motions in a common latent space, a source query transformer and a target take the responsibility of projecting the motions into a canonical motion space, where motion transfer can be conducted via the addition of motion codes. The edited motion is then projected from the canonical space back to the domain of the source by a transformer. A generator takes the projected motion and the source features as input to synthesize a face in the domain of the source." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.473, + 0.696 + ], + "angle": 0, + "content": "Moreover, to overcome the lack of paired data, we propose a novel cross-domain training scheme. The core idea is the analogy constraint (see Fig. 2). Specifically, given two real faces and a cartoon face, we use the relative motion between the two real faces to drive the cartoon one, producing a synthetic cartoon face. The relative motion from one real face to the other real one is analogous to that from the cartoon face to the synthetic one. We measure the distance between the two types of analogical motions in the aligned canonical motion space. They are supposed to be identical." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.473, + 0.757 + ], + "angle": 0, + "content": "To conduct cross-domain reenactment experiments, we collect a cartoon video dataset in Disney style, which contains 344 videos and about 47k frames in total. The cartoon dataset will be released for research purposes." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.757, + 0.318, + 0.771 + ], + "angle": 0, + "content": "Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.781, + 0.47, + 0.842 + ], + "angle": 0, + "content": "- We propose a novel framework for cross-domain face reenactment without using paired data. Several transformers are designed to align the motions of different domains in a common latent space." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.852, + 0.47, + 0.868 + ], + "angle": 0, + "content": "- We propose a training scheme to compensate for the" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.781, + 0.47, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.09, + 0.845, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.282, + 0.895, + 0.422 + ], + "angle": 0, + "content": "Figure 2. Visualization of analogy constraint. \\(\\Delta \\omega_{real}\\) and \\(\\Delta \\omega_{cartoon}\\) represent the relative motion in the real and cartoon domain, respectively. \\(\\Delta \\omega_{can}\\) represents the projected motion in the canonical space where we can perform cross-domain motion transfer by the addition of motion codes. Given a cartoon face, we utilize the relative motion between two real faces to animate the cartoon one and generate a synthetic cartoon face. We hold that the relative motion between two real faces and that between the cartoon face and the synthetic are equivalent within the canonical motion space." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.447, + 0.871, + 0.463 + ], + "angle": 0, + "content": "lack of paired data by using the analogy constraint." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.474, + 0.892, + 0.518 + ], + "angle": 0, + "content": "- We collect a cartoon video dataset and conduct extensive experiments to demonstrate the superiority of our method in the cross-domain reenactment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.534, + 0.642, + 0.55 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.8, + 0.575 + ], + "angle": 0, + "content": "2.1. Within-domain Face Reenactment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Most one-shot talking face generation works focus on within-domain face reenactment, i.e., driving a portrait image using another talking head video, where both the driving video and source image is from real humans." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.825 + ], + "angle": 0, + "content": "Keypoint-based Methods. Many reenactment algorithms [26, 30, 31, 5] are based on pre-defined facial landmarks that are used for motion conversion and repositioning. For example, ReenactGAN [26] encodes the face into the boundary hidden space defined by the keypoint heat map and transfers motion by a boundary transformer. Although ReenactGAN can produce high-quality target faces, it needs to retrain the model for the new individual images. FReeNet [31] trained a keypoint converter to convert the driven expression onto the source face. However, it can only deal with expression change, but not attitude changes, so the generated face image will keep the pose of the source face." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "3DMM-based Methods. A number of works [17, 7, 4, 29, 33] take advantage of the 3D deformation model [3] that can decouple pose, expression, and identity. For instance, Pirenderer [17] mapped the target 3DMM parameters into hidden codes and injected them into the source" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "1Note that Pose in [13] refers to a 20 dimension (20D) vector that contains the status of facial expression and head pose." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7691" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "image features to guide the generation of distorted streams. HeadGAN [7] renders 3D parameters to images as motion representations. [4] map 3DMM parameters to StyleGAN [11, 12] latent space, and a transformation matrix is learned to map 3DMM parameters to latent code. Style-HEAT [29] used 3DMM to generate an optical flow map to change the feature map of StyleGAN. Using 3DMM parameters as motion representation can avoid the problem of video identity information leakage in cross-identity reenactment, but the 3D model only encodes facial regions, such as teeth, eyes, hair and other information that is difficult to synthesize." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.501 + ], + "angle": 0, + "content": "Unsupervised Methods. Another large class of methods [25, 18, 19, 22, 10, 23, 8] do not use existing face models, but use unsupervised training to obtain facial motion representations. A typical method is FOMM [19], which proposes a first-order motion method by representing the motion as the key points and its local transformation. Facevid2vid [22] further extends this method to 3D facial key points. DAGAN [10] used the self-supervised geometry learning method to obtain the depth map from the video. Motion transformer [21] uses the interactive capabilities of transformer to predict key points. Differing from the above methods, LIA [23] does not require explicit structural representation but directly manipulates the latent space of the generative model. Megapotrait [8] follows a similar way and adds a super-resolution network for higher resolution." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.502, + 0.473, + 0.577 + ], + "angle": 0, + "content": "The goals of these within-domain reenactment methods are different from ours since they do not need to consider the domain gap. They can be trained in a self-supervised way from the same video, while there are no cartoon-real training pairs in our cross-domain settings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.367, + 0.605 + ], + "angle": 0, + "content": "2.2. Cross-domain Face Reenactment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.855 + ], + "angle": 0, + "content": "Only a few methods focus on cross-domain face reenactment. The early study Recycle-GAN [1] is based on image translation networks. It introduces a recurrent loss by training a temporal predictor to monitor the temporal consistency. And each character requires two face videos as two domains for training, which cannot be used for one-shot reenactment. Everythingtalking [20] proposes a parametric shape modeling technique parametric models based on Bezier Curve and transfers motion by adapting the human face's control points to the pareidolia face. But it is unable to change the facial posture. MAA [28] is a method that relies on keypoints and incorporates an angle consistency loss to replicate driving motion. However, the differences in keypoints only account for absolute motion transfer, which might significantly alter the appearance of source images and fail to capture intricate facial expressions well." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The most related work AnimeCeleb [13] proposes a new cartoon talking-head video dataset. It uses a 3D animation model to generate a large number of animated face im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.274 + ], + "angle": 0, + "content": "ages and corresponding pose annotations. To achieve cross-domain reenactment, it proposes to learn the mapping relationship from animation pose vectors to 3DMM parameters by paired data. One shortcoming of AnimeCeleb is that it becomes inapplicable if no paired data is available, which limits its application scope. Moreover, it cannot be generalized to other cartoon styles that require rich expressions. Because AnimeCeleb uses the manually designed 20D vector to represent the status of pose and expression, which is far from enough. Differently, our method does not depend on 3DMMs and can be trained without being paired by using a designed cross-domain training scheme." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.285, + 0.593, + 0.301 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.895, + 0.522 + ], + "angle": 0, + "content": "We propose a novel transformer-based framework to align the motions from different domains in a shared canonical motion space for motion transfer. The framework is shown in Fig. 3. The task becomes simple if there exist paired data, i.e., faces in different domains on the same expression and pose. Unfortunately, no such dataset is available. To settle the lack of paired data, we propose a training scheme with the designed analogy constraint, using faces from two domains jointly. We first introduce the architecture of the framework in Sec. 3.1. Then, we present the training scheme in Sec. 3.2. Finally, we show a two-stage learning strategy to alleviate the issue of imbalanced training data in Sec. 3.3, i.e., the number of real videos is much larger than that of cartoon videos." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.78, + 0.546 + ], + "angle": 0, + "content": "3.1. Transformer-based Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.894, + 0.78 + ], + "angle": 0, + "content": "As shown in Fig. 3, given a source image and a driving one from two domains, we extract the motion information from the two images with two domain-specific motion encoders. Then, a source query transformer and a driving one are used to project the extracted motion information from different domains into the shared canonical motion space, with the domain-specific motion code base and the motion information as input. The motion transfer is realized by adding motion codes in the shared motion space. A domain-specific backward transformer projects the edited motion code back to the domain of the source image with the assistance of its features extracted via a domain-specific appearance encoder. Finally, a generator takes the projected motion code and the source features as input to render an image in the domain of the source." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Please note that we use the cartoon as the source and the real as the driving for illustration in Fig. 3. We can exchange the roles of the cartoon and the real, i.e., the real as the source while the cartoon is the driving. Therefore, in our full framework, we have two appearance encoders, two motion encoders, two backward transformers, two generators, and two motion code bases, which are domain-specific. We also have one source query transformer and one driv-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7692" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.511, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.089, + 0.697, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.09, + 0.887, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.313, + 0.894, + 0.398 + ], + "angle": 0, + "content": "Figure 3. An illustration of our model, which can be mainly divided into three parts: (1) Feature extraction. We use \\( E_{a} \\) and \\( E_{m} \\) to extract local features and latent motion descriptor respectively. (2) Latent space transformation. A source query transformer and a driving one are used to project the motion descriptor into a canonical motion space. Motion transfer is conducted via the addition of motion codes in the shared motion space. (3) Image generation. The edited motion is projected back to the domain of the source with the assistant of the source features. A progressive image generator is used to render an image with the projected motion. We also apply a spatial feature transformation (SFT) to the last three high-resolution features to reduce the artifacts caused by feature warping." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.469, + 0.48 + ], + "angle": 0, + "content": "ing query transformer, which are shared by two domains. The bounding box in color indicates whether a component is domain-specific in Fig. 3. We present the components in detail as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.648 + ], + "angle": 0, + "content": "Appearance and Motion Encoders. We have domain-specific encoders to extract appearance and motion information from an image. Let \\( E_{a}^{C} \\) and \\( E_{m}^{C} \\) denote the appearance and motion encoder for the cartoon domain, respectively, while \\( E_{a}^{R} \\) and \\( E_{m}^{R} \\) for the real domain. As the roles of the two domains as the source and the target can be exchanged, we use \\( E_{a} \\) and \\( E_{m} \\) for simplicity. The appearance encoder takes the source image as input and outputs multi-scale spatial features, while the motion encoder extracts motion descriptors from the source and driving image, respectively. Thus, the feature extraction can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.659, + 0.469, + 0.675 + ], + "angle": 0, + "content": "\\[\nS = E _ {a} \\left(x _ {s}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.678, + 0.469, + 0.695 + ], + "angle": 0, + "content": "\\[\nF _ {s} = E _ {m} \\left(x _ {s}\\right), F _ {d} = E _ {m} \\left(x _ {d}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.749 + ], + "angle": 0, + "content": "where \\( F_{s} \\) and \\( F_{d} \\) denote the motion descriptors of the source \\( x_{s} \\) and the driving \\( x_{d} \\). \\( S = \\{S^{i}\\}_{i = 1}^{N} \\) represents the appearance features of the source." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Motion Query Transformer. In addition to the domain-specific motion encoder, we use a motion base for each domain for further discovering domain-specific motion properties. The motion base consists of learnable embeddings to memorize typical motions, denoted as \\( p = \\{p^k\\}_{k=1}^K \\), where \\( p_k \\in R^d \\) represents the \\( k \\)-th motion embedding and \\( d \\) is the dimension. We adopt transformer blocks to project the latent descriptor into a canonical motion space due to their capability of capturing long-range dependencies. As shown in the middle of Fig. 3, the embedding is linearly projected" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.892, + 0.513 + ], + "angle": 0, + "content": "to the query features while the latent descriptor from the image is projected to the key and value features. Inspired by [19], since it is challenging to model motion transformation \\( x_{s} \\rightarrow x_{d} \\) straightforwardly, we choose to model transformation \\( x_{s} \\rightarrow x_{r} \\) and \\( x_{r} \\rightarrow x_{d} \\) separately by leveraging a virtual reference \\( x_{r} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.527, + 0.892, + 0.542 + ], + "angle": 0, + "content": "\\[\nw _ {s \\rightarrow d} = w _ {s \\rightarrow r} + w _ {r \\rightarrow d}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.674 + ], + "angle": 0, + "content": "where \\(w_{s\\rightarrow r} = T_{s\\rightarrow r}(F_s,p)\\) and \\(w_{r\\rightarrow d} = T_{r\\rightarrow d}(F_d,p)\\). \\(w_{s\\rightarrow r},w_{r\\rightarrow d}\\), and \\(w_{s\\rightarrow d}\\) are three motion codes in canonical motion space. \\(T_{s\\rightarrow r}\\) denotes the source query transformer that models the motion transformation from the source to the reference. While \\(T_{r\\rightarrow d}\\) denotes the driving query transformer that models the motion transformation from the reference to the driving. The two query transformers are shared by the two domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Backward Transformer. As we obtain the motion transformation \\( w_{s\\rightarrow d} \\) in the canonical motion space, the next step is to project the motion code back to the domain of the source image. We use a domain-specific backward transformer to take over this task with the assistance of the information of the source image, i.e., the multi-scale features \\( S \\) of the source image from the appearance encoder \\( E_{a} \\). We have two such transformers, i.e., \\( T_B^C \\) and \\( T_B^R \\), and take the notation \\( T_{B} \\) for simplicity. To let the motion code obtain information from image features in different resolutions, we design the backward transformer as a stack of three transformer blocks. As shown on the right of Fig. 3, in each transformer block, the query features are derived from the output of the previous module, while the key and value features are computed from the corresponding multi-scale fea" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7693" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "tures. The first query features are the linear projection of the motion transformation \\( w_{s\\rightarrow d} \\) from the canonical space. And the resolutions of the feature maps are \\( 16\\times 16 \\), \\( 32\\times 32 \\), and \\( 64\\times 64 \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.364 + ], + "angle": 0, + "content": "Image Generator. After obtaining the projected motion code in the source domain, we use a domain-specific generator to progressively warp and refine the source feature maps layer by layer conditioned on the motion code \\( w_{s \\to d} \\). We have two generators, i.e., \\( G^{C} \\) and \\( G^{R} \\), and take the notation \\( G \\) for simplicity. Considering that warping them only will make the final output suffer from artifacts, we add refinement modules in the last three layers of the generator to enhance three high-resolution feature maps. Given the motion code \\( w_{s \\to d} \\) and the previous feature maps \\( o^{i-1} \\), a StyleConv block is used to predict a flow field that is used to warp the feature map \\( S^{i} \\). With the warped feature map \\( f_{w}^{i} \\) and the motion code \\( w_{s \\to d} \\) as input, the spatial feature transformation (SFT) for refinement is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.371, + 0.469, + 0.389 + ], + "angle": 0, + "content": "\\[\no ^ {i} = \\alpha^ {i} \\odot f _ {w} ^ {i} + \\beta^ {i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.396, + 0.47, + 0.427 + ], + "angle": 0, + "content": "where \\(\\alpha^i = \\mathrm{StyleConv}(f_w^i, w_{s \\to d})\\) is the spatial scale and \\(\\beta^i = \\mathrm{StyleConv}(f_w^i, w_{s \\to d})\\) is the bias." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.434, + 0.356, + 0.451 + ], + "angle": 0, + "content": "3.2. Cross-domain Training Scheme" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.47, + 0.563 + ], + "angle": 0, + "content": "If there are paired data where cartoon and real video pairs share the same head pose and expression, cross-domain training can be settled simply following the way of training a conventional within-domain model. However, no such paired data is available. To tackle this issue, we propose an analogy constraint to train the model with unpaired cartoons and real videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.715 + ], + "angle": 0, + "content": "Analogy Constraint. As shown in Fig. 2, the core idea is to project the relative motions of different domains into the canonical motion space. For instance, given a tuple consisting of two real faces and a cartoon one, we can calculate the relative motion from one real face to the other real one. Then, the relative motion can be used to animate the cartoon face to generate a synthetic cartoon face. We can compute the relative motion from the cartoon face to the synthetic cartoon one. The two types of relative motions are supposed to be identical in canonical motion space." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.715, + 0.469, + 0.744 + ], + "angle": 0, + "content": "In the canonical motion space, following [19], the relative motion transfer can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.754, + 0.469, + 0.77 + ], + "angle": 0, + "content": "\\[\nw _ {s \\rightarrow d} = \\left(w _ {s \\rightarrow r} + w _ {r \\rightarrow s}\\right) + \\left(w _ {r \\rightarrow d} - w _ {r \\rightarrow 1}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.777, + 0.47, + 0.839 + ], + "angle": 0, + "content": "where '1' refers to a selected frame in the driving video. For illustration, let \\( x_{a}^{R} \\) and \\( x_{b}^{R} \\) denote two real images and \\( x_{s}^{C} \\) denote a cartoon image. Then, the relative motion transfer is applied to the cartoon, which can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.847, + 0.469, + 0.865 + ], + "angle": 0, + "content": "\\[\nw _ {s \\rightarrow g} = \\left(w _ {s \\rightarrow r} + w _ {r \\rightarrow s}\\right) + \\left(w _ {r \\rightarrow b} - w _ {r \\rightarrow a}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.469, + 0.904 + ], + "angle": 0, + "content": "where \\(w_{s\\rightarrow g}\\) represents the motion transformation from \\(x_{s}^{C}\\) to \\(x_{g}^{C}\\). 'g' indicates the synthetic cartoon face. Each item" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.674, + 0.106 + ], + "angle": 0, + "content": "can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.113, + 0.796, + 0.131 + ], + "angle": 0, + "content": "\\[\nw _ {s \\rightarrow r} = T _ {s \\rightarrow r} \\left(E _ {m} ^ {C} \\left(x _ {s} ^ {C}\\right), p ^ {C}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.134, + 0.892, + 0.16 + ], + "angle": 0, + "content": "\\[\nw _ {r \\rightarrow s} = T _ {r \\rightarrow d} \\left(E _ {m} ^ {C} \\left(x _ {s} ^ {C}\\right), p ^ {C}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.154, + 0.796, + 0.172 + ], + "angle": 0, + "content": "\\[\nw _ {r \\rightarrow a} = T _ {r \\rightarrow d} \\left(E _ {m} ^ {R} \\left(x _ {a} ^ {R}\\right), p ^ {R}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.176, + 0.796, + 0.193 + ], + "angle": 0, + "content": "\\[\nw _ {r \\rightarrow b} = T _ {r \\rightarrow d} \\left(E _ {m} ^ {R} \\left(x _ {b} ^ {R}\\right), p ^ {R}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.202, + 0.892, + 0.277 + ], + "angle": 0, + "content": "With the motion code \\( w_{s \\to g} \\) as well as the source features \\( S \\), the generator \\( G^{C} \\) can produce a synthetic cartoon face \\( x_{g}^{C} \\), i.e., \\( x_{g}^{C} = G^{C}(T_{B}^{C}(w_{s \\to g}), E_{a}^{C}(x_{s}^{C})) \\). Then, we can calculate the motion code of the synthetic cartoon face in the canonical motion space, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.286, + 0.891, + 0.305 + ], + "angle": 0, + "content": "\\[\nw _ {r \\rightarrow g} = T _ {r \\rightarrow d} \\left(E _ {m} ^ {C} \\left(x _ {g} ^ {C}\\right), p ^ {C}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.313, + 0.892, + 0.389 + ], + "angle": 0, + "content": "The relative motion between the cartoon face and the synthetic one is \\((w_{r\\rightarrow g} - w_{r\\rightarrow s})\\), which is supposed to be identical to the relative motion between the two real faces, i.e., \\((w_{r\\rightarrow b} - w_{r\\rightarrow a})\\). Therefore, we can measure the distance between the two types of relative motions by" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.397, + 0.891, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m o t i o n}} ^ {R \\rightarrow C} = \\mathbb {E} [ \\| \\left(w _ {r \\rightarrow g} - w _ {r \\rightarrow s}\\right) - \\left(w _ {r \\rightarrow b} - w _ {r \\rightarrow a}\\right) \\| _ {1} ]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.43, + 0.892, + 0.489 + ], + "angle": 0, + "content": "Similarly, given two cartoon faces and one real face, we have \\(\\mathcal{L}_{motion}^{C\\rightarrow R}\\) following the same formulation. Further, to maintain the high fidelity of face generation, we impose adversarial loss. The adversarial loss can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.495, + 0.891, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} ^ {R} = \\mathbb {E} _ {x _ {g} ^ {R}} [ - \\log (D ^ {R} (x _ {g} ^ {R})) ], \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.518, + 0.797, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a d v} ^ {C} = \\mathbb {E} _ {x _ {g} ^ {C}} [ - \\log (D ^ {C} (x _ {g} ^ {C})) ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.548, + 0.892, + 0.579 + ], + "angle": 0, + "content": "where \\( D^{R} \\) and \\( D^{C} \\) are two discriminators for the real and cartoon domains, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.58, + 0.892, + 0.775 + ], + "angle": 0, + "content": "Reconstruction. Besides the analogy constraint that emphasizes the interaction between two domains, we can also perform within-domain reenactment to assist the model learning. Specifically, we can sample two images within the same domain as the source and the driving, e.g., a pair of \\( x_{s}^{R} \\) and \\( x_{d}^{R} \\) or a pair of \\( x_{s}^{C} \\) and \\( x_{d}^{C} \\). The reconstruction loss of the real pairs can be used to update the parameters of the real domain-related components as well as the shared motion query transformers. Similarly, the reconstruction loss of the cartoon pairs is for the corresponding components. We exploit the \\( L_{1} \\) loss \\( \\mathcal{L}_{1}(x_{s\\rightarrow d},x_{d}) \\), the perceptual loss \\( \\mathcal{L}_{per}(x_{s\\rightarrow d},x_{d}) \\), and the GAN loss \\( \\mathcal{L}_{adv}(x_{s\\rightarrow d},x_{d}) \\). Hence, the objective of reconstruction can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.787, + 0.891, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c}} = \\mathcal {L} _ {1} + \\lambda_ {\\text {p e r}} \\mathcal {L} _ {\\text {p e r}} + \\mathcal {L} _ {\\text {a d v}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.815, + 0.869, + 0.83 + ], + "angle": 0, + "content": "where we omit the domain notation R&C for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.831, + 0.804, + 0.845 + ], + "angle": 0, + "content": "The overall objective is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.853, + 0.892, + 0.885 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {r e c} ^ {R} + \\mathcal {L} _ {r e c} ^ {C} + \\lambda_ {m} * \\left(\\mathcal {L} _ {\\text {m o t i o n}} ^ {R \\rightarrow C} + \\mathcal {L} _ {\\text {m o t i o n}} ^ {C \\rightarrow R}\\right) + \\mathcal {L} _ {\\text {a d v}} ^ {C} + \\mathcal {L} _ {\\text {a d v}} ^ {R}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.887, + 0.78, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda_{m}\\) is the trade-off hyperparameter." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7694" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.092, + 0.197, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.092, + 0.313, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.092, + 0.427, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.092, + 0.543, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.092, + 0.657, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.092, + 0.775, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.092, + 0.89, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.182, + 0.197, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.182, + 0.312, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.182, + 0.427, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.182, + 0.543, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.182, + 0.657, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.182, + 0.775, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.182, + 0.89, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.271, + 0.197, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.271, + 0.312, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.271, + 0.427, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.271, + 0.543, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.271, + 0.657, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.271, + 0.774, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.271, + 0.89, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.359, + 0.197, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.359, + 0.312, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.359, + 0.427, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.359, + 0.543, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.359, + 0.657, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.359, + 0.775, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.359, + 0.89, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.449, + 0.178, + 0.463 + ], + "angle": 0, + "content": "(a) Source" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.449, + 0.295, + 0.463 + ], + "angle": 0, + "content": "(b) FOMM" + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.449, + 0.397, + 0.463 + ], + "angle": 0, + "content": "(c) LIA" + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.449, + 0.528, + 0.463 + ], + "angle": 0, + "content": "(d) DaGAN" + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.449, + 0.635, + 0.463 + ], + "angle": 0, + "content": "(e) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.449, + 0.746, + 0.463 + ], + "angle": 0, + "content": "(f) Start" + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.449, + 0.873, + 0.464 + ], + "angle": 0, + "content": "(g) Driving" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.471, + 0.893, + 0.514 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons with state-of-the-art methods on cross-domain reenactment. Competing methods are finetuned with the cartoon dataset. Hence all models are trained on the same data. It can be seen that our results have superior image quality and motion consistency. More visual results are in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.537, + 0.43, + 0.552 + ], + "angle": 0, + "content": "3.3. Two-stage Learning on Imbalanced Data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.661 + ], + "angle": 0, + "content": "The number of real videos is much larger than that of cartoon videos. Directly training with mixed imbalanced data cannot promote the performance of cross-domain reenactment because real faces dominate the data distribution. To alleviate the issue of the imbalanced dataset, we propose to learn the model in two stages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.469, + 0.761 + ], + "angle": 0, + "content": "In the first stage, we train the model with real videos by performing the reconstruction task, through which the model can learn the general knowledge of face motion from a large number of talking head videos. The obtained model cannot accurately transfer motion from real faces to cartoon ones due to domain shifts, especially facial expressions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.771, + 0.469, + 0.845 + ], + "angle": 0, + "content": "In the second stage, we copy the domain-specific components of the real domain as initialization for those of the cartoon domain. Then, we jointly learn all the components in the framework with the analogy constraint and the reconstruction loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Please note that our final model can be used to reenact a cartoon with a real video as well as to reenact a real face with a cartoon video." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.537, + 0.632, + 0.553 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.565, + 0.597, + 0.581 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.743 + ], + "angle": 0, + "content": "Datasets. For cross-domain reenactment, we collect a cartoon video dataset in Disney style which contains 344 videos with an FPS of 30 and the total number of frames is about \\(47\\mathrm{k}\\). To train our model, we take a subset of videos from the VoxCeleb [15] dataset containing 17K videos, and a subset from self-collected cartoon videos containing 289 videos. For evaluation, we use the rest 473 videos from the VoxCeleb [15] dataset and 55 videos from the cartoon dataset. We also use a set of images from CelebA-HQ [14] for cross-identity reenactment evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Metrics. We evaluate the performance from four aspects, i.e., reconstruction, fidelity, identity preservation, and motion transfer. \\(L_{1}\\), Learned Perceptual Image Patch Similarity (LPIPS)[32], and Peak Signal-to-Noise Ratio (PSNR) are used to measure the reconstruction accuracy. Frechet Inception Distance (FID) [9] and Cumulative probability of Blur Detection (CPBD) [16] are used for measuring generation realism. The cosine similarity (CSIM) between synthetic and source images through ArcFace [6] is used for identity preservation. To evaluate the accuracy of mo" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7695" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.089, + 0.462, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.47, + 0.346 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons with MAA [28]. Our method can better preserve the face shape of the original image. And more accurately imitated the driving image, especially the eyes and mouth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.469, + 0.432 + ], + "angle": 0, + "content": "tion transfer, following [10], we use Average Keypoint Distance (AKD) and Average Expression Distance (AED) to compute the distance between synthetic and target images in terms of expression and pose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.469, + 0.631 + ], + "angle": 0, + "content": "Competing methods. We compare our method with several recent state-of-the-art methods, including FOMM [19], DAGAN [10], and LIA [23]. FOMM is a typical motion transfer method based on unsupervised keypoints. DAGAN introduces depth into the estimation of key points and has a similar formulation of motion transfer as FOMM. LIA is a representative method that uses a reference space similar to FOMM but conducts motion transfer in a latent space. For a fair comparison, we initialize the models using their pretrained official models and finetune them on the cartoon dataset. Therefore, we learn with the same data. Furthermore, we also compare with MAA [28], a keypoint-based cross-domain motion transfer method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.469, + 0.818 + ], + "angle": 0, + "content": "Implementation Details. The model is trained in two stages to handle the imbalanced dataset where real videos are much more than cartoons. In the first stage, we trained the model on the training set of VoxCeleb for 450K iterations, where the batch size is 64 and the optimizer is Adam with an initial learning rate of \\(5 \\times 10^{-4}\\). In the second stage, we fine-tune the model with Disney videos with 10K iterations, where the batch size is 8 and the optimizer is Adam with an initial learning rate of \\(1 \\times 10^{-4}\\). The hyperparameters are set as \\(\\lambda_{per} = 10\\) and \\(\\lambda_{m} = 20\\) to balance generative fidelity and motion consistency. A detailed description of model components is in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.37, + 0.847 + ], + "angle": 0, + "content": "4.2. Cross-Domain Face Reenactment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "As mentioned in Sec. 3.3, our method can handle the bidirectional cross-domain reenactment. We execute motion transfer task for evaluation i.e., real \\(\\rightarrow\\) cartoon" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.09, + 0.895, + 0.191 + ], + "angle": 0, + "content": "
Real→CartoonCartoon→Real
FID↓CPBD↑FID↓CSIM↑CPBD↑
FOMM30.3690.043651.9290.5360.0626
LIA30.4420.049348.6110.5150.0727
DaGAN29.7790.073547.2220.5610.0985
Ours28.5310.084031.8520.6550.1230
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.201, + 0.88, + 0.216 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on cross-domain reenactment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.379 + ], + "angle": 0, + "content": "and cartoon \\(\\rightarrow\\) real. The former represents driving a cartoon face with a real video, while the latter represents driving a real face with a cartoon video. As for real \\(\\rightarrow\\) cartoon, 60 videos from VoxCeleb are used to drive 300 cartoon images from the cartoon dataset. For a fair comparison, all competing methods are fine-tuned on the cartoon training set. As for cartoon \\(\\rightarrow\\) real, 55 cartoon videos are employed to drive 275 images of different identities from CelebA-HQ [14]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.892, + 0.503 + ], + "angle": 0, + "content": "Qualitative Evaluation. The results of real \\(\\rightarrow\\) cartoon are shown in the first two rows of Fig. 4. First, our method achieves better image quality than other methods in terms of sharpness and artifacts. Second, our method outperforms others in motion transfer, especially facial expression. For instance, in the first row, the eyes of the cartoon should be closed according to the driving image. Only our method mimics driving while the eyes of others stay open." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.639 + ], + "angle": 0, + "content": "The results of cartoon \\(\\rightarrow\\) real are shown in the last two rows of Fig. 4. We can observe a similar phenomenon as the real \\(\\rightarrow\\) cartoon case in terms of image quality, where our method works the best. FOMM and DaGAN become underperformed when there is a large head pose difference between the source and driving, as shown in the third and fourth rows. LIA is better than them but still worse than ours. For motion and identity preservation, our method surpasses others." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.762 + ], + "angle": 0, + "content": "The comparison with the cross-domain motion transfer method MAA [28] is shown in Fig. 5. As the source code of MAA is not released, its results are adapted from the original paper. Our method has better, especially in the mouth, eyes and image quality. For instance, in the third column, MAA ignores the motion of the eyes, whereas our method accurately mimics the driving image. In the last column, our mouth shape is more consistent with the driving one." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Quantitative Evaluation. The quantitative results are reported in Tab. 1. Our method achieves the best FID in the two tasks, indicating that the synthetic results are most consistent with the source distribution. Our method also performs the best in sharpness and identity preservation, i.e., the highest CPBD and CSIM, which is consistent with the observation in visual results. These results demonstrate that our model can capture the appearance and motion properties of different domains and the motion transfer in the canoni" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7696" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.3, + 0.47, + 0.342 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons on within-domain reenactment. Our method achieves better image quality than other methods in terms of sharpness, distortion and artifacts." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.355, + 0.462, + 0.442 + ], + "angle": 0, + "content": "
L1↓LPIPS↓CPBD↑AKD↓AED↓
FOMM0.04660.1390.03081.3880.142
LIA0.04610.1230.03721.5120.153
DaGAN0.04420.1220.04151.3000.129
Ours0.04570.1210.04291.3350.139
" + }, + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.451, + 0.431, + 0.465 + ], + "angle": 0, + "content": "Table 2. Same-identity reconstruction in real domain" + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.479, + 0.462, + 0.566 + ], + "angle": 0, + "content": "
FID↓CSIM↑CPBD ↑
FOMM54.4110.5410.0689
LIA53.1610.5220.0787
DaGAN49.9200.5480.1069
Ours48.7130.5260.1132
" + }, + { + "type": "table_caption", + "bbox": [ + 0.122, + 0.576, + 0.424, + 0.59 + ], + "angle": 0, + "content": "Table 3. Cross-identity reenactment in real domain" + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.604, + 0.465, + 0.66 + ], + "angle": 0, + "content": "
FOMMLIADAGANOurs
FID↓40.85334.06636.21827.467
CPBD↑0.03680.05230.05980.0846
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.67, + 0.436, + 0.684 + ], + "angle": 0, + "content": "Table 4. Cross-identity reenactment in cartoon domain." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.711, + 0.273, + 0.726 + ], + "angle": 0, + "content": "cal motion space works well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.729, + 0.47, + 0.835 + ], + "angle": 0, + "content": "User Study. To further evaluate the visual quality, we conduct a user study by asking 20 human raters to answer 20 multiple-choice questions. In each question, a rater chooses the best from 4 synthetic cartoon videos generated by the three competing methods and ours. Our method is the most favorable with a selection rate of \\(72.75\\%\\). While the rates are \\(13.75\\%\\), \\(9\\%\\), and \\(4.5\\%\\) for LIA, DaGAN, and FOMM." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.381, + 0.862 + ], + "angle": 0, + "content": "4.3. Within-Domain Face Reenactment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Besides the cross-domain evaluation, we also compare our method with competing methods in the scenario of" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.223, + 0.892, + 0.265 + ], + "angle": 0, + "content": "Figure 7. Qualitative ablation study on the analogy constraint. We observe that the image quality and motion consistency is much better by using analogy constraint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.41 + ], + "angle": 0, + "content": "within-domain reenactment though our method is designed for the cross-domain case. For the same-identity evaluation, we use 200 test videos from VoxCeleb. For cross-identity evaluation, we use the 200 videos to drive 1,000 images from CelebA-HQ. Each video drives five images. The obtained 1,000 videos are used to compute the metrics. Note that we use the released pre-trained models of competing methods here." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.487 + ], + "angle": 0, + "content": "The quantitative results are shown in Tab. 2 and Tab. 3. For same-identity evaluation, our method is overall comparable with DaGAN and better than LIA and FOMM. For cross-identity evaluation, our method achieves the best FID and CPBD. Our CSIM is comparable with others." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.608 + ], + "angle": 0, + "content": "Since the visual difference among different methods in same-identity reenactment is marginal, we only present a visual comparison of cross-identity reenactment in Fig. 6. It can be observed that our results have the best image quality, presenting more details in the face area, e.g., eyes and teeth. Besides, our method is comparable with LIA in preserving face shape and sometimes even better (see the last row in Fig. 6)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.67 + ], + "angle": 0, + "content": "In addition, we also perform a cross-identity face reenactment in the cartoon domain. We use 55 cartoon videos to drive 275 cartoon images. Results are shown in Tab. 4. Our method outperforms others in both FID and CPBD." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.654, + 0.697 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To evaluate the training scheme, we perform an ablation study by dropping the analogy constraint. The quantitative results are shown in Tab. 5. 'w/o analogy' represents dropping the analogy constraint, i.e., finetuning the obtained model from the first stage with the cartoon dataset using only \\(\\mathcal{L}_{rec}^{R}\\) and \\(\\mathcal{L}_{rec}^{C}\\) in Eq.12. It can be observed that dropping the analogy constraint causes overall performance degradation, especially for the cartoon→real task. The qualitative evaluation is shown in Fig. 7. In the second row, the result without analogy constraints cannot accurately mimic the motion of the driving image. The eyes of the cartoon should be closed, but it is opened when dropping constraints. The dropping operation has a side effect" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7697" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.487, + 0.158 + ], + "angle": 0, + "content": "
Real →CartoonCartoon → Real
FID↓CPBD↑FID↓CPBD↑
our full model28.5310.084031.8520.123
w/o analogy29.2300.074643.3530.107
" + }, + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.168, + 0.421, + 0.182 + ], + "angle": 0, + "content": "Table 5. Ablation study on the analogy constraint." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.197, + 0.482, + 0.254 + ], + "angle": 0, + "content": "
02040100
Real→Cartoon26.88528.53132.09441.936
Cartoon→Real33.46331.85230.17628.166
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.264, + 0.416, + 0.277 + ], + "angle": 0, + "content": "Table 6. Ablation study on hyper-parameter \\( {\\lambda }_{m} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.293, + 0.485, + 0.349 + ], + "angle": 0, + "content": "
Fullmlpw/o backw/o sft
Real→Cartoon28.531116.74028.90431.291
Cartoon→Real31.85231.88535.26541.077
" + }, + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.359, + 0.411, + 0.373 + ], + "angle": 0, + "content": "Table 7. Ablation study on model architecture." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.4, + 0.468, + 0.429 + ], + "angle": 0, + "content": "on the motion consistency, which indicates that the analogy constraint contributes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.468, + 0.522 + ], + "angle": 0, + "content": "To further analyze the influence of analogy constraint, we perform an ablation on the trade-off hyperparameter \\(\\lambda_{m}\\). As is shown in Tab. 6, it can be observed that \\(\\lambda_{m}\\) trades off the performance of the two cross-domain tasks, i.e., Cartoon \\(\\rightarrow\\) Real and Real \\(\\rightarrow\\) Cartoon. We empirically choose \\(\\lambda_{m} = 20\\) in practical." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.468, + 0.734 + ], + "angle": 0, + "content": "We conduct ablation studies to demonstrate the effectiveness of several modules in the architecture, i.e., the query transformer, the backward transformer, and the spatial feature transformation. The ablation results are shown in Tab. 7. The metric is FID. \"mlp\" indicates that we replace the query transformer with multi-layer perception (MLP). \"w/o back\" denotes that we drop the backward transformer. And \"w/o sft\" drops the spatial feature transformation module. In Tab. 7, we can observe that query transformer brings significant improvement in the Real \\(\\rightarrow\\) Cartoon case. It verifies that the proposed query transformer can project the latent descriptor from different domains into a canonical motion space. Additionally, removing the backward transformer and sft will damage the quality of generated images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We also conduct a visualization experiment to verify the proposed canonical space. Since no paired data is available, we select three real images with different poses and expressions as templates to drive real and cartoon images, generating 450 real image pairs and 150 cartoon pairs. Each pair contains faces before and after driving. We compute their relative motions in the common motion space. In Fig. 8, 0, 1, and 2 represent real pairs, while 3, 4, and 5 represent cartoon pairs. \\(<0,3>\\), \\(<1,4>\\), and \\(<2,5>\\) are driven by the same template, respectively. This shows that the motions of the two domains are aligned." + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.092, + 0.808, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.343 + ], + "angle": 0, + "content": "Figure 8. Visualizing the canonical motion space. We visualize the t-SNE plot of motion codes in canonical space. \\( <0,1,2> \\) and \\( <3,4,5> \\) are from the relative motions of real domain and cartoon domain, respectively. And \\( <0,3>, <1,4> \\), and \\( <2,5> \\) are driven by the same template, respectively." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.358, + 0.89, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.455, + 0.891, + 0.484 + ], + "angle": 0, + "content": "Figure 9. Failure cases. We observed that it is challenging to handle extreme poses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.51, + 0.618, + 0.525 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.732 + ], + "angle": 0, + "content": "We propose a novel transformer-based framework for cross-domain face reenactment. To settle the domain shift issue, we propose to align the motions of different domains in a canonical motion space. A set of domain-specific and domain-shared components are designed for motion alignment. Motion transfer can be implemented by adding motion codes in the shared motion space. Since there is no paired data for training, we propose a cross-domain training scheme with the analogy constraint using the two domain datasets. To alleviate the influence of imbalanced data, we use a two-stage learning strategy. Extensive experiments are conducted to demonstrate the superiority of our method over competing methods in cross-domain reenactment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Limitations. Like other SOTA models, our model cannot handle extreme poses well. As is shown in Fig 9, our method suffers from artifacts and deformations when the driving image is an extreme pose. This may be due to the fact that extreme poses rarely appear in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was partly supported by the National Natural Science Foundation of China (Grant No. U1903213) and the Shenzhen Science and Technology Program (JCYJ20220818101014030)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7698" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Aayush Bansal, Shugao Ma, Deva Ramanan, and Yaser Sheikh. Recycle-gan: Unsupervised video retargeting. In Proceedings of the European conference on computer vision (ECCV), pages 119-135, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.211 + ], + "angle": 0, + "content": "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In ACM SIGGRAPH, pages 187-194, 1999, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5543-5552. ISSN: 1063-6919. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.324 + ], + "angle": 0, + "content": "[4] Stella Bounareli, Vasileios Argyriou, and Georgios Tzimiropoulos. Finding directions in GAN's latent space for neural face reenactment. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[5] Zhuo Chen, Chaoyue Wang, Bo Yuan, and Dacheng Tao. PuppeteerGAN: Arbitrary portrait animation with semantic-aware appearance transformation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13515-13524. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.464 + ], + "angle": 0, + "content": "[6] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[7] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoria Sharmanska. HeadGAN: One-shot neural head synthesis and editing. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[8] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. MegaPortraits: One-shot megapixel neural head avatars. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.471, + 0.618 + ], + "angle": 0, + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[10] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.47, + 0.69 + ], + "angle": 0, + "content": "[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.47, + 0.733 + ], + "angle": 0, + "content": "[12] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[13] Kangyeol Kim, Sunghyun Park, Jaeseong Lee, Sunghyo Chung, Junsoo Lee, and Jaegul Choo. AnimeCeleb: Large-scale animation CelebHeads dataset for head reenactment. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[14] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5549-5558, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. VoxCeleb: a large-scale speaker identification dataset. In Interspeech 2017, pages 2616-2620. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[16] Niranjan D Narvekar and Lina J Karam. A no-reference image blur metric based on the cumulative probability of blur detection (cpbd). IEEE Transactions on Image Processing, 20(9):2678-2683, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[17] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. PIRenderer: Controllable portrait image generation via semantic neural rendering. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 13739-13748. IEEE. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[18] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. version: 2.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[19] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. 1, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[20] Linsen Song, Wayne Wu, Chaoyou Fu, Chen Qian, Chen Change Loy, and Ran He. Everything's talkin': Pareidolia face reenactment. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[21] Jiale Tao, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Motion transformer for unsupervised image animation. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[22] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10034-10044. IEEE. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[23] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. LATIVE IMAGE ANIMATOR: LEARNING TO ANIMATE IMAGES VIA LATENT SPACE NAVIGATION. page 17. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.506, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[24] Wikipedia contributors. Morph target animation — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Morph_target Animation&oldid=1078665970, 2022. [Online; accessed 6-November-2022]. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[25] Olivia Wiles, A. Sophia Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, volume 11217, pages 690–706. Springer International Publishing. Series Title: Lecture Notes in Computer Science. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[26] Wayne Wu, Yunxuan Zhang, Cheng Li, Chen Qian, and Chen Change Loy. ReenactGAN: Learning to reenact faces via boundary transfer. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[27] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[28] Borun Xu, Biao Wang, Jinhong Deng, Jiale Tao, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Motion and appearance adaptation for cross-domain motion transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 529-545. Springer, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7699" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Yujiu Yang. StyleHEAT: One-shot high-resolution editable talking face generation via pre-trained StyleGAN. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[30] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9458-9467. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[31] Jiangning Zhang, Xianfang Zeng, Mengmeng Wang, Yusu Pan, Liang Liu, Yong Liu, Yu Ding, and Changjie Fan. FReeNet: Multi-identity face reenactment. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5325-5334. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[33] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7700" + } + ] +] \ No newline at end of file diff --git a/2023/ToonTalker_ Cross-Domain Face Reenactment/d3014894-36d6-4bc9-bdc4-403449ecf9bd_origin.pdf b/2023/ToonTalker_ Cross-Domain Face Reenactment/d3014894-36d6-4bc9-bdc4-403449ecf9bd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..84d215833060d091fc4456577b4d8bf037380cbc --- /dev/null +++ b/2023/ToonTalker_ Cross-Domain Face Reenactment/d3014894-36d6-4bc9-bdc4-403449ecf9bd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afeee71f599aae878c7d3f811021e484c9bcf190040476f45366420661c85716 +size 3475551 diff --git a/2023/ToonTalker_ Cross-Domain Face Reenactment/full.md b/2023/ToonTalker_ Cross-Domain Face Reenactment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eaa6e26274d748b6fb59e66d0b93aca2e32c20a8 --- /dev/null +++ b/2023/ToonTalker_ Cross-Domain Face Reenactment/full.md @@ -0,0 +1,417 @@ +# ToonTalker: Cross-Domain Face Reenactment + +Yuan Gong $^{1}$ Yong Zhang $^{2*}$ Xiaodong Cun $^{2}$ Fei Yin $^{1}$ Yanbo Fan $^{2}$ Xuan Wang $^{3}$ Baoyuan Wu $^{4}$ Yujiu Yang $^{1*}$ + +$^{1}$ Shenzhen International Graduate School, Tsinghua University $^{2}$ Tencent AI Lab $^{3}$ Ant Group $^{4}$ The School of Data Science, Shenzhen Research Institute of Big Data, The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen) + +# Abstract + +We target cross-domain face reenactment in this paper, i.e., driving a cartoon image with the video of a real person and vice versa. Recently, many works have focused on one-shot talking face generation to drive a portrait with a real video, i.e., within-domain reenactment. Straightforwardly applying those methods to cross-domain animation will cause inaccurate expression transfer, blur effects, and even apparent artifacts due to the domain shift between cartoon and real faces. Only a few works attempt to settle cross-domain face reenactment. The most related work AnimeCeleb [13] requires constructing a dataset with pose vector and cartoon image pairs by animating 3D characters, which makes it inapplicable anymore if no paired data is available. In this paper, we propose a novel method for cross-domain reenactment without paired data. Specifically, we propose a transformer-based framework to align the motions from different domains into a common latent space where motion transfer is conducted via latent code addition. Two domain-specific motion encoders and two learnable motion base memories are used to capture domain properties. A source query transformer and a driving one are exploited to project domain-specific motion to the canonical space. The edited motion is projected back to the domain of the source with a transformer. Moreover, since no paired data is provided, we propose a novel cross-domain training scheme using data from two domains with the designed analogy constraint. Besides, we contribute a cartoon dataset in Disney style. Extensive evaluations demonstrate the superiority of our method over competing methods. + +# 1. Introduction + +Online video conferences and live streaming are spreading rapidly since the population of smartphones and communication techniques. Speakers sometimes present them + +![](images/61b8aaf7d892918700650c5e39005a92db18b3cab23873327a4b3b758e625819.jpg) +Figure 1. Cross-domain reenactment examples of our method. Given a cartoon face, we animate it by transferring the pose and expression from a real face. + +selves as cartoon characters for privacy and entertainment consideration. However, conventional cartoon animation depends on a complex pipeline of 3D modeling and retargeting based on a powerful render engine, which always takes a long period [24]. This limitation makes it difficult to drive a new model or directly drive a single comic picture. + +There are many works [17, 19, 10, 23, 27] that focus on one-shot talking face generation, i.e., driving a portrait with a video of a real person, which belongs to within-domain reenactment. Their models are trained with a large amount of talking videos and perform well in the same-identity and cross-identity driving tasks. However, when those models are straightforwardly applied to cross-domain reenactment, i.e., driving a cartoon face with a real video as in Fig. 1, they persistently encounter the following issues, including inaccurate expression transfer, blur effects, and even apparent artifacts around facial components and the background. The reason is the distribution shift between different domains. There is a large appearance and motion gap between the cartoon and real datasets. Though fine-tuning a pre-trained model on a cartoon dataset can improve the visual quality to a certain extent, those issues still cannot be wiped out. + +Only a few methods focus on visually-driven cross + +domain face reenactment. The early method RecycleGAN [1] overfits two videos for unsupervised video translation, which cannot be used for one-shot reenactment. MAA [28] uses keypoints as an intermediate representation and designs an angle consistency loss to mimic the driving motion. But only focusing on keypoints will result in ignoring the subtle motion of facial components, especially the mouth and eyes. AnimeCeleb [13] is based on pre-defined 3D Morphable Models (3DMMs) [2] and additional pose annotations1 of cartoon images. However, it is inapplicable when paired data is not available, which limits its range of use. Moreover, the pre-defined pose vector has limited capability of capturing fine-grained facial expressions. + +In this paper, we propose a novel method for cross-domain reenactment without the requirement of paired data. All we need is a set of real videos and another set of cartoon videos. To solve the domain shift problem, we propose a transformer-based framework to align the motions from two domains in a shared canonical latent space. Specifically, two motion encoders and two motion bases are designed to discover the domain-specific properties of appearance and motion. To align the domain-specific motions in a common latent space, a source query transformer and a target take the responsibility of projecting the motions into a canonical motion space, where motion transfer can be conducted via the addition of motion codes. The edited motion is then projected from the canonical space back to the domain of the source by a transformer. A generator takes the projected motion and the source features as input to synthesize a face in the domain of the source. + +Moreover, to overcome the lack of paired data, we propose a novel cross-domain training scheme. The core idea is the analogy constraint (see Fig. 2). Specifically, given two real faces and a cartoon face, we use the relative motion between the two real faces to drive the cartoon one, producing a synthetic cartoon face. The relative motion from one real face to the other real one is analogous to that from the cartoon face to the synthetic one. We measure the distance between the two types of analogical motions in the aligned canonical motion space. They are supposed to be identical. + +To conduct cross-domain reenactment experiments, we collect a cartoon video dataset in Disney style, which contains 344 videos and about 47k frames in total. The cartoon dataset will be released for research purposes. + +Our contributions are as follows: + +- We propose a novel framework for cross-domain face reenactment without using paired data. Several transformers are designed to align the motions of different domains in a common latent space. +- We propose a training scheme to compensate for the + +![](images/0383b6a34e320079cd79497edce396c30ecc481b2957108b08d23cca1d25549a.jpg) +Figure 2. Visualization of analogy constraint. $\Delta \omega_{real}$ and $\Delta \omega_{cartoon}$ represent the relative motion in the real and cartoon domain, respectively. $\Delta \omega_{can}$ represents the projected motion in the canonical space where we can perform cross-domain motion transfer by the addition of motion codes. Given a cartoon face, we utilize the relative motion between two real faces to animate the cartoon one and generate a synthetic cartoon face. We hold that the relative motion between two real faces and that between the cartoon face and the synthetic are equivalent within the canonical motion space. + +lack of paired data by using the analogy constraint. + +- We collect a cartoon video dataset and conduct extensive experiments to demonstrate the superiority of our method in the cross-domain reenactment. + +# 2. Related Work + +# 2.1. Within-domain Face Reenactment + +Most one-shot talking face generation works focus on within-domain face reenactment, i.e., driving a portrait image using another talking head video, where both the driving video and source image is from real humans. + +Keypoint-based Methods. Many reenactment algorithms [26, 30, 31, 5] are based on pre-defined facial landmarks that are used for motion conversion and repositioning. For example, ReenactGAN [26] encodes the face into the boundary hidden space defined by the keypoint heat map and transfers motion by a boundary transformer. Although ReenactGAN can produce high-quality target faces, it needs to retrain the model for the new individual images. FReeNet [31] trained a keypoint converter to convert the driven expression onto the source face. However, it can only deal with expression change, but not attitude changes, so the generated face image will keep the pose of the source face. + +3DMM-based Methods. A number of works [17, 7, 4, 29, 33] take advantage of the 3D deformation model [3] that can decouple pose, expression, and identity. For instance, Pirenderer [17] mapped the target 3DMM parameters into hidden codes and injected them into the source + +image features to guide the generation of distorted streams. HeadGAN [7] renders 3D parameters to images as motion representations. [4] map 3DMM parameters to StyleGAN [11, 12] latent space, and a transformation matrix is learned to map 3DMM parameters to latent code. Style-HEAT [29] used 3DMM to generate an optical flow map to change the feature map of StyleGAN. Using 3DMM parameters as motion representation can avoid the problem of video identity information leakage in cross-identity reenactment, but the 3D model only encodes facial regions, such as teeth, eyes, hair and other information that is difficult to synthesize. + +Unsupervised Methods. Another large class of methods [25, 18, 19, 22, 10, 23, 8] do not use existing face models, but use unsupervised training to obtain facial motion representations. A typical method is FOMM [19], which proposes a first-order motion method by representing the motion as the key points and its local transformation. Facevid2vid [22] further extends this method to 3D facial key points. DAGAN [10] used the self-supervised geometry learning method to obtain the depth map from the video. Motion transformer [21] uses the interactive capabilities of transformer to predict key points. Differing from the above methods, LIA [23] does not require explicit structural representation but directly manipulates the latent space of the generative model. Megapotrait [8] follows a similar way and adds a super-resolution network for higher resolution. + +The goals of these within-domain reenactment methods are different from ours since they do not need to consider the domain gap. They can be trained in a self-supervised way from the same video, while there are no cartoon-real training pairs in our cross-domain settings. + +# 2.2. Cross-domain Face Reenactment + +Only a few methods focus on cross-domain face reenactment. The early study Recycle-GAN [1] is based on image translation networks. It introduces a recurrent loss by training a temporal predictor to monitor the temporal consistency. And each character requires two face videos as two domains for training, which cannot be used for one-shot reenactment. Everythingtalking [20] proposes a parametric shape modeling technique parametric models based on Bezier Curve and transfers motion by adapting the human face's control points to the pareidolia face. But it is unable to change the facial posture. MAA [28] is a method that relies on keypoints and incorporates an angle consistency loss to replicate driving motion. However, the differences in keypoints only account for absolute motion transfer, which might significantly alter the appearance of source images and fail to capture intricate facial expressions well. + +The most related work AnimeCeleb [13] proposes a new cartoon talking-head video dataset. It uses a 3D animation model to generate a large number of animated face im + +ages and corresponding pose annotations. To achieve cross-domain reenactment, it proposes to learn the mapping relationship from animation pose vectors to 3DMM parameters by paired data. One shortcoming of AnimeCeleb is that it becomes inapplicable if no paired data is available, which limits its application scope. Moreover, it cannot be generalized to other cartoon styles that require rich expressions. Because AnimeCeleb uses the manually designed 20D vector to represent the status of pose and expression, which is far from enough. Differently, our method does not depend on 3DMMs and can be trained without being paired by using a designed cross-domain training scheme. + +# 3. Method + +We propose a novel transformer-based framework to align the motions from different domains in a shared canonical motion space for motion transfer. The framework is shown in Fig. 3. The task becomes simple if there exist paired data, i.e., faces in different domains on the same expression and pose. Unfortunately, no such dataset is available. To settle the lack of paired data, we propose a training scheme with the designed analogy constraint, using faces from two domains jointly. We first introduce the architecture of the framework in Sec. 3.1. Then, we present the training scheme in Sec. 3.2. Finally, we show a two-stage learning strategy to alleviate the issue of imbalanced training data in Sec. 3.3, i.e., the number of real videos is much larger than that of cartoon videos. + +# 3.1. Transformer-based Framework + +As shown in Fig. 3, given a source image and a driving one from two domains, we extract the motion information from the two images with two domain-specific motion encoders. Then, a source query transformer and a driving one are used to project the extracted motion information from different domains into the shared canonical motion space, with the domain-specific motion code base and the motion information as input. The motion transfer is realized by adding motion codes in the shared motion space. A domain-specific backward transformer projects the edited motion code back to the domain of the source image with the assistance of its features extracted via a domain-specific appearance encoder. Finally, a generator takes the projected motion code and the source features as input to render an image in the domain of the source. + +Please note that we use the cartoon as the source and the real as the driving for illustration in Fig. 3. We can exchange the roles of the cartoon and the real, i.e., the real as the source while the cartoon is the driving. Therefore, in our full framework, we have two appearance encoders, two motion encoders, two backward transformers, two generators, and two motion code bases, which are domain-specific. We also have one source query transformer and one driv- + +![](images/0777385d3f6a2231b65bfba5ed2ce616432d0731491f57c0b85a6c0be50fab7e.jpg) +Figure 3. An illustration of our model, which can be mainly divided into three parts: (1) Feature extraction. We use $E_{a}$ and $E_{m}$ to extract local features and latent motion descriptor respectively. (2) Latent space transformation. A source query transformer and a driving one are used to project the motion descriptor into a canonical motion space. Motion transfer is conducted via the addition of motion codes in the shared motion space. (3) Image generation. The edited motion is projected back to the domain of the source with the assistant of the source features. A progressive image generator is used to render an image with the projected motion. We also apply a spatial feature transformation (SFT) to the last three high-resolution features to reduce the artifacts caused by feature warping. + +![](images/09459a17696775d8e6964ca82fe3ddf068706635ead81e9f0b3e63e29fc2c8bd.jpg) + +![](images/c9ef07372e0a0ed88e1e5436431cc756cafaaaa3668728268b63842b3f57c9d6.jpg) + +ing query transformer, which are shared by two domains. The bounding box in color indicates whether a component is domain-specific in Fig. 3. We present the components in detail as follows. + +Appearance and Motion Encoders. We have domain-specific encoders to extract appearance and motion information from an image. Let $E_{a}^{C}$ and $E_{m}^{C}$ denote the appearance and motion encoder for the cartoon domain, respectively, while $E_{a}^{R}$ and $E_{m}^{R}$ for the real domain. As the roles of the two domains as the source and the target can be exchanged, we use $E_{a}$ and $E_{m}$ for simplicity. The appearance encoder takes the source image as input and outputs multi-scale spatial features, while the motion encoder extracts motion descriptors from the source and driving image, respectively. Thus, the feature extraction can be defined as: + +$$ +S = E _ {a} \left(x _ {s}\right), \tag {1} +$$ + +$$ +F _ {s} = E _ {m} \left(x _ {s}\right), F _ {d} = E _ {m} \left(x _ {d}\right), \tag {2} +$$ + +where $F_{s}$ and $F_{d}$ denote the motion descriptors of the source $x_{s}$ and the driving $x_{d}$ . $S = \{S^{i}\}_{i = 1}^{N}$ represents the appearance features of the source. + +Motion Query Transformer. In addition to the domain-specific motion encoder, we use a motion base for each domain for further discovering domain-specific motion properties. The motion base consists of learnable embeddings to memorize typical motions, denoted as $p = \{p^k\}_{k=1}^K$ , where $p_k \in R^d$ represents the $k$ -th motion embedding and $d$ is the dimension. We adopt transformer blocks to project the latent descriptor into a canonical motion space due to their capability of capturing long-range dependencies. As shown in the middle of Fig. 3, the embedding is linearly projected + +to the query features while the latent descriptor from the image is projected to the key and value features. Inspired by [19], since it is challenging to model motion transformation $x_{s} \rightarrow x_{d}$ straightforwardly, we choose to model transformation $x_{s} \rightarrow x_{r}$ and $x_{r} \rightarrow x_{d}$ separately by leveraging a virtual reference $x_{r}$ : + +$$ +w _ {s \rightarrow d} = w _ {s \rightarrow r} + w _ {r \rightarrow d}, \tag {3} +$$ + +where $w_{s\rightarrow r} = T_{s\rightarrow r}(F_s,p)$ and $w_{r\rightarrow d} = T_{r\rightarrow d}(F_d,p)$ . $w_{s\rightarrow r},w_{r\rightarrow d}$ , and $w_{s\rightarrow d}$ are three motion codes in canonical motion space. $T_{s\rightarrow r}$ denotes the source query transformer that models the motion transformation from the source to the reference. While $T_{r\rightarrow d}$ denotes the driving query transformer that models the motion transformation from the reference to the driving. The two query transformers are shared by the two domains. + +Backward Transformer. As we obtain the motion transformation $w_{s\rightarrow d}$ in the canonical motion space, the next step is to project the motion code back to the domain of the source image. We use a domain-specific backward transformer to take over this task with the assistance of the information of the source image, i.e., the multi-scale features $S$ of the source image from the appearance encoder $E_{a}$ . We have two such transformers, i.e., $T_B^C$ and $T_B^R$ , and take the notation $T_{B}$ for simplicity. To let the motion code obtain information from image features in different resolutions, we design the backward transformer as a stack of three transformer blocks. As shown on the right of Fig. 3, in each transformer block, the query features are derived from the output of the previous module, while the key and value features are computed from the corresponding multi-scale fea + +tures. The first query features are the linear projection of the motion transformation $w_{s\rightarrow d}$ from the canonical space. And the resolutions of the feature maps are $16\times 16$ , $32\times 32$ , and $64\times 64$ , respectively. + +Image Generator. After obtaining the projected motion code in the source domain, we use a domain-specific generator to progressively warp and refine the source feature maps layer by layer conditioned on the motion code $w_{s \to d}$ . We have two generators, i.e., $G^{C}$ and $G^{R}$ , and take the notation $G$ for simplicity. Considering that warping them only will make the final output suffer from artifacts, we add refinement modules in the last three layers of the generator to enhance three high-resolution feature maps. Given the motion code $w_{s \to d}$ and the previous feature maps $o^{i-1}$ , a StyleConv block is used to predict a flow field that is used to warp the feature map $S^{i}$ . With the warped feature map $f_{w}^{i}$ and the motion code $w_{s \to d}$ as input, the spatial feature transformation (SFT) for refinement is defined as: + +$$ +o ^ {i} = \alpha^ {i} \odot f _ {w} ^ {i} + \beta^ {i}, \tag {4} +$$ + +where $\alpha^i = \mathrm{StyleConv}(f_w^i, w_{s \to d})$ is the spatial scale and $\beta^i = \mathrm{StyleConv}(f_w^i, w_{s \to d})$ is the bias. + +# 3.2. Cross-domain Training Scheme + +If there are paired data where cartoon and real video pairs share the same head pose and expression, cross-domain training can be settled simply following the way of training a conventional within-domain model. However, no such paired data is available. To tackle this issue, we propose an analogy constraint to train the model with unpaired cartoons and real videos. + +Analogy Constraint. As shown in Fig. 2, the core idea is to project the relative motions of different domains into the canonical motion space. For instance, given a tuple consisting of two real faces and a cartoon one, we can calculate the relative motion from one real face to the other real one. Then, the relative motion can be used to animate the cartoon face to generate a synthetic cartoon face. We can compute the relative motion from the cartoon face to the synthetic cartoon one. The two types of relative motions are supposed to be identical in canonical motion space. + +In the canonical motion space, following [19], the relative motion transfer can be defined as: + +$$ +w _ {s \rightarrow d} = \left(w _ {s \rightarrow r} + w _ {r \rightarrow s}\right) + \left(w _ {r \rightarrow d} - w _ {r \rightarrow 1}\right), \tag {5} +$$ + +where '1' refers to a selected frame in the driving video. For illustration, let $x_{a}^{R}$ and $x_{b}^{R}$ denote two real images and $x_{s}^{C}$ denote a cartoon image. Then, the relative motion transfer is applied to the cartoon, which can be defined as: + +$$ +w _ {s \rightarrow g} = \left(w _ {s \rightarrow r} + w _ {r \rightarrow s}\right) + \left(w _ {r \rightarrow b} - w _ {r \rightarrow a}\right), \tag {6} +$$ + +where $w_{s\rightarrow g}$ represents the motion transformation from $x_{s}^{C}$ to $x_{g}^{C}$ . 'g' indicates the synthetic cartoon face. Each item + +can be defined as follows: + +$$ +w _ {s \rightarrow r} = T _ {s \rightarrow r} \left(E _ {m} ^ {C} \left(x _ {s} ^ {C}\right), p ^ {C}\right), +$$ + +$$ +w _ {r \rightarrow s} = T _ {r \rightarrow d} \left(E _ {m} ^ {C} \left(x _ {s} ^ {C}\right), p ^ {C}\right), \tag {7} +$$ + +$$ +w _ {r \rightarrow a} = T _ {r \rightarrow d} \left(E _ {m} ^ {R} \left(x _ {a} ^ {R}\right), p ^ {R}\right), +$$ + +$$ +w _ {r \rightarrow b} = T _ {r \rightarrow d} \left(E _ {m} ^ {R} \left(x _ {b} ^ {R}\right), p ^ {R}\right). +$$ + +With the motion code $w_{s \to g}$ as well as the source features $S$ , the generator $G^{C}$ can produce a synthetic cartoon face $x_{g}^{C}$ , i.e., $x_{g}^{C} = G^{C}(T_{B}^{C}(w_{s \to g}), E_{a}^{C}(x_{s}^{C}))$ . Then, we can calculate the motion code of the synthetic cartoon face in the canonical motion space, i.e., + +$$ +w _ {r \rightarrow g} = T _ {r \rightarrow d} \left(E _ {m} ^ {C} \left(x _ {g} ^ {C}\right), p ^ {C}\right). \tag {8} +$$ + +The relative motion between the cartoon face and the synthetic one is $(w_{r\rightarrow g} - w_{r\rightarrow s})$ , which is supposed to be identical to the relative motion between the two real faces, i.e., $(w_{r\rightarrow b} - w_{r\rightarrow a})$ . Therefore, we can measure the distance between the two types of relative motions by + +$$ +\mathcal {L} _ {\text {m o t i o n}} ^ {R \rightarrow C} = \mathbb {E} [ \| \left(w _ {r \rightarrow g} - w _ {r \rightarrow s}\right) - \left(w _ {r \rightarrow b} - w _ {r \rightarrow a}\right) \| _ {1} ]. \tag {9} +$$ + +Similarly, given two cartoon faces and one real face, we have $\mathcal{L}_{motion}^{C\rightarrow R}$ following the same formulation. Further, to maintain the high fidelity of face generation, we impose adversarial loss. The adversarial loss can be defined as: + +$$ +\mathcal {L} _ {a d v} ^ {R} = \mathbb {E} _ {x _ {g} ^ {R}} [ - \log (D ^ {R} (x _ {g} ^ {R})) ], \tag {10} +$$ + +$$ +\mathcal {L} _ {a d v} ^ {C} = \mathbb {E} _ {x _ {g} ^ {C}} [ - \log (D ^ {C} (x _ {g} ^ {C})) ], +$$ + +where $D^{R}$ and $D^{C}$ are two discriminators for the real and cartoon domains, respectively. + +Reconstruction. Besides the analogy constraint that emphasizes the interaction between two domains, we can also perform within-domain reenactment to assist the model learning. Specifically, we can sample two images within the same domain as the source and the driving, e.g., a pair of $x_{s}^{R}$ and $x_{d}^{R}$ or a pair of $x_{s}^{C}$ and $x_{d}^{C}$ . The reconstruction loss of the real pairs can be used to update the parameters of the real domain-related components as well as the shared motion query transformers. Similarly, the reconstruction loss of the cartoon pairs is for the corresponding components. We exploit the $L_{1}$ loss $\mathcal{L}_{1}(x_{s\rightarrow d},x_{d})$ , the perceptual loss $\mathcal{L}_{per}(x_{s\rightarrow d},x_{d})$ , and the GAN loss $\mathcal{L}_{adv}(x_{s\rightarrow d},x_{d})$ . Hence, the objective of reconstruction can be defined as: + +$$ +\mathcal {L} _ {\text {r e c}} = \mathcal {L} _ {1} + \lambda_ {\text {p e r}} \mathcal {L} _ {\text {p e r}} + \mathcal {L} _ {\text {a d v}}, \tag {11} +$$ + +where we omit the domain notation R&C for simplicity. + +The overall objective is defined as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {r e c} ^ {R} + \mathcal {L} _ {r e c} ^ {C} + \lambda_ {m} * \left(\mathcal {L} _ {\text {m o t i o n}} ^ {R \rightarrow C} + \mathcal {L} _ {\text {m o t i o n}} ^ {C \rightarrow R}\right) + \mathcal {L} _ {\text {a d v}} ^ {C} + \mathcal {L} _ {\text {a d v}} ^ {R}, \tag {12} +$$ + +where $\lambda_{m}$ is the trade-off hyperparameter. + +![](images/15f70bb361daaa1a641397abcd294b5cbd9ffa1ef1ee91e79cd5130122b76def.jpg) + +![](images/3b46cfe5329d3f720eb2f6d9278e006515ea869530013db651fab816b7cddeb6.jpg) + +![](images/87aa5dee397bdaa1b0a102b7e0994dfd648455158baea16f6bde68530b429b05.jpg) + +![](images/136da6006347cc990177518b3b3c6d931299a4e1309d8ebb0534bee71f20edd4.jpg) + +![](images/6c069edc8893e8c0e0ba48f7c0963c3244409d1404eaecffa3a0ac858d6cc1a5.jpg) + +![](images/b509db33864fe08b4a0bd05d6649fe287039a2dfd9f8d9e143f0e92af29424fe.jpg) + +![](images/cc699f785c6cf9dbeb8ea4cc2249ee97a3917f4509a22d59b558dabe1d833afb.jpg) + +![](images/746e393aad85bcbeb019c27c71fd14734277542152dd1e7ce67ddf05cf4ee0aa.jpg) + +![](images/4f89aa8696d50fee10818afc819880c41ad83b1959f622fc1e5a8a09effe9d1b.jpg) + +![](images/f368ecf62861dfca60a7bb8b97249ccd96510c391a0d4a018f9df881a74a3a0d.jpg) + +![](images/71c21fef78925ed4f31b9298c72139b2c9db077266f489e52aec4ba5f41bc042.jpg) + +![](images/697f6cf14c2953d83f79e9e7860d3fcf212600884e66d567aa9497e88865b068.jpg) + +![](images/3fc7ff01840344ed3ed15b4cc58c12bf931bf8879b0a6dd4e25d0d385581e0b2.jpg) + +![](images/55bb4d1f34ce718754f447c43f0852e5a0ebbab9fddc547fc1dbc6753c7e9ca1.jpg) + +![](images/447a4e4982aa89abc459274d2d6d1532f7a734a2fddef937924e783b1effac32.jpg) + +![](images/43ba0535483f2f922220a5734f3b26a5b2c646d4bd1b6f4d4c3f5019c1796855.jpg) + +![](images/edef89eb42e5eca7ef21b6bafbb9d909df8244cb8d902561b9f317a5d4a54fc4.jpg) + +![](images/c740cac3ad7a2624609f08ba34c21ee1d79970bd68e7b3173c6736f558b0e689.jpg) + +![](images/ddfab59f7739ab470275329d3f79433e26cabeef8672a2e0b6446be49a0206ed.jpg) + +![](images/042bf4f241b73434e9bc458d4a163ac99a6e7dc6aa7d514ff09d85406eaab4d1.jpg) + +![](images/b00c0afab42688f0b6c36ae790193fcd468c5adf8c3ea37b44e8f2185cc0668a.jpg) + +![](images/820948162f062253a2961b5676b81ea2d0256bf2c2f45ff6c0cec199e9a191fe.jpg) +(a) Source + +![](images/3d9380c89cebe77dffccd731670d636bf9e8a2264eefa4f1bf7a26cabfa89dc3.jpg) +(b) FOMM + +![](images/1ca83750f818e07c2d3be39b96e6fe5e895f404a96e766823a3eef4284b7126e.jpg) +(c) LIA + +![](images/dd80d55c042ea52f16861b81b2a0a689541b41a943711f3152dffb222cb0142f.jpg) +(d) DaGAN +Figure 4. Qualitative comparisons with state-of-the-art methods on cross-domain reenactment. Competing methods are finetuned with the cartoon dataset. Hence all models are trained on the same data. It can be seen that our results have superior image quality and motion consistency. More visual results are in the supplementary. + +![](images/6ad27e5292613041d305a7a92c8b4dd5afb534b6abc86665443122629d7f9531.jpg) +(e) Ours + +![](images/53aaaf1480412b5106be672953983cd06feed2dbdeb5ac559b2b157554ecb779.jpg) +(f) Start + +![](images/e3a3b10f91610c298080c8d7d5a9c7650cf9353cf8b34460403e2200acd71e8a.jpg) +(g) Driving + +# 3.3. Two-stage Learning on Imbalanced Data. + +The number of real videos is much larger than that of cartoon videos. Directly training with mixed imbalanced data cannot promote the performance of cross-domain reenactment because real faces dominate the data distribution. To alleviate the issue of the imbalanced dataset, we propose to learn the model in two stages. + +In the first stage, we train the model with real videos by performing the reconstruction task, through which the model can learn the general knowledge of face motion from a large number of talking head videos. The obtained model cannot accurately transfer motion from real faces to cartoon ones due to domain shifts, especially facial expressions. + +In the second stage, we copy the domain-specific components of the real domain as initialization for those of the cartoon domain. Then, we jointly learn all the components in the framework with the analogy constraint and the reconstruction loss. + +Please note that our final model can be used to reenact a cartoon with a real video as well as to reenact a real face with a cartoon video. + +# 4. Experiments + +# 4.1. Settings + +Datasets. For cross-domain reenactment, we collect a cartoon video dataset in Disney style which contains 344 videos with an FPS of 30 and the total number of frames is about $47\mathrm{k}$ . To train our model, we take a subset of videos from the VoxCeleb [15] dataset containing 17K videos, and a subset from self-collected cartoon videos containing 289 videos. For evaluation, we use the rest 473 videos from the VoxCeleb [15] dataset and 55 videos from the cartoon dataset. We also use a set of images from CelebA-HQ [14] for cross-identity reenactment evaluation. + +Metrics. We evaluate the performance from four aspects, i.e., reconstruction, fidelity, identity preservation, and motion transfer. $L_{1}$ , Learned Perceptual Image Patch Similarity (LPIPS)[32], and Peak Signal-to-Noise Ratio (PSNR) are used to measure the reconstruction accuracy. Frechet Inception Distance (FID) [9] and Cumulative probability of Blur Detection (CPBD) [16] are used for measuring generation realism. The cosine similarity (CSIM) between synthetic and source images through ArcFace [6] is used for identity preservation. To evaluate the accuracy of mo + +![](images/61f6b0f7df7251baed3e5399b2db63c86febbfb7d86d022898b7120db9aad243.jpg) +Figure 5. Qualitative comparisons with MAA [28]. Our method can better preserve the face shape of the original image. And more accurately imitated the driving image, especially the eyes and mouth. + +tion transfer, following [10], we use Average Keypoint Distance (AKD) and Average Expression Distance (AED) to compute the distance between synthetic and target images in terms of expression and pose. + +Competing methods. We compare our method with several recent state-of-the-art methods, including FOMM [19], DAGAN [10], and LIA [23]. FOMM is a typical motion transfer method based on unsupervised keypoints. DAGAN introduces depth into the estimation of key points and has a similar formulation of motion transfer as FOMM. LIA is a representative method that uses a reference space similar to FOMM but conducts motion transfer in a latent space. For a fair comparison, we initialize the models using their pretrained official models and finetune them on the cartoon dataset. Therefore, we learn with the same data. Furthermore, we also compare with MAA [28], a keypoint-based cross-domain motion transfer method. + +Implementation Details. The model is trained in two stages to handle the imbalanced dataset where real videos are much more than cartoons. In the first stage, we trained the model on the training set of VoxCeleb for 450K iterations, where the batch size is 64 and the optimizer is Adam with an initial learning rate of $5 \times 10^{-4}$ . In the second stage, we fine-tune the model with Disney videos with 10K iterations, where the batch size is 8 and the optimizer is Adam with an initial learning rate of $1 \times 10^{-4}$ . The hyperparameters are set as $\lambda_{per} = 10$ and $\lambda_{m} = 20$ to balance generative fidelity and motion consistency. A detailed description of model components is in the supplementary. + +# 4.2. Cross-Domain Face Reenactment + +As mentioned in Sec. 3.3, our method can handle the bidirectional cross-domain reenactment. We execute motion transfer task for evaluation i.e., real $\rightarrow$ cartoon + +
Real→CartoonCartoon→Real
FID↓CPBD↑FID↓CSIM↑CPBD↑
FOMM30.3690.043651.9290.5360.0626
LIA30.4420.049348.6110.5150.0727
DaGAN29.7790.073547.2220.5610.0985
Ours28.5310.084031.8520.6550.1230
+ +Table 1. Quantitative comparison on cross-domain reenactment. + +and cartoon $\rightarrow$ real. The former represents driving a cartoon face with a real video, while the latter represents driving a real face with a cartoon video. As for real $\rightarrow$ cartoon, 60 videos from VoxCeleb are used to drive 300 cartoon images from the cartoon dataset. For a fair comparison, all competing methods are fine-tuned on the cartoon training set. As for cartoon $\rightarrow$ real, 55 cartoon videos are employed to drive 275 images of different identities from CelebA-HQ [14]. + +Qualitative Evaluation. The results of real $\rightarrow$ cartoon are shown in the first two rows of Fig. 4. First, our method achieves better image quality than other methods in terms of sharpness and artifacts. Second, our method outperforms others in motion transfer, especially facial expression. For instance, in the first row, the eyes of the cartoon should be closed according to the driving image. Only our method mimics driving while the eyes of others stay open. + +The results of cartoon $\rightarrow$ real are shown in the last two rows of Fig. 4. We can observe a similar phenomenon as the real $\rightarrow$ cartoon case in terms of image quality, where our method works the best. FOMM and DaGAN become underperformed when there is a large head pose difference between the source and driving, as shown in the third and fourth rows. LIA is better than them but still worse than ours. For motion and identity preservation, our method surpasses others. + +The comparison with the cross-domain motion transfer method MAA [28] is shown in Fig. 5. As the source code of MAA is not released, its results are adapted from the original paper. Our method has better, especially in the mouth, eyes and image quality. For instance, in the third column, MAA ignores the motion of the eyes, whereas our method accurately mimics the driving image. In the last column, our mouth shape is more consistent with the driving one. + +Quantitative Evaluation. The quantitative results are reported in Tab. 1. Our method achieves the best FID in the two tasks, indicating that the synthetic results are most consistent with the source distribution. Our method also performs the best in sharpness and identity preservation, i.e., the highest CPBD and CSIM, which is consistent with the observation in visual results. These results demonstrate that our model can capture the appearance and motion properties of different domains and the motion transfer in the canoni + +![](images/d5a34432666915a490b55848cd7685217972d53067b98ea48d4a616ef37f330b.jpg) +Figure 6. Qualitative comparisons on within-domain reenactment. Our method achieves better image quality than other methods in terms of sharpness, distortion and artifacts. + +
L1↓LPIPS↓CPBD↑AKD↓AED↓
FOMM0.04660.1390.03081.3880.142
LIA0.04610.1230.03721.5120.153
DaGAN0.04420.1220.04151.3000.129
Ours0.04570.1210.04291.3350.139
+ +Table 2. Same-identity reconstruction in real domain + +
FID↓CSIM↑CPBD ↑
FOMM54.4110.5410.0689
LIA53.1610.5220.0787
DaGAN49.9200.5480.1069
Ours48.7130.5260.1132
+ +Table 3. Cross-identity reenactment in real domain + +
FOMMLIADAGANOurs
FID↓40.85334.06636.21827.467
CPBD↑0.03680.05230.05980.0846
+ +Table 4. Cross-identity reenactment in cartoon domain. + +cal motion space works well. + +User Study. To further evaluate the visual quality, we conduct a user study by asking 20 human raters to answer 20 multiple-choice questions. In each question, a rater chooses the best from 4 synthetic cartoon videos generated by the three competing methods and ours. Our method is the most favorable with a selection rate of $72.75\%$ . While the rates are $13.75\%$ , $9\%$ , and $4.5\%$ for LIA, DaGAN, and FOMM. + +# 4.3. Within-Domain Face Reenactment + +Besides the cross-domain evaluation, we also compare our method with competing methods in the scenario of + +![](images/3662346031899134b75f2bd9a0390cfe06d5f3fcd137fef5472807059c70b29f.jpg) +Figure 7. Qualitative ablation study on the analogy constraint. We observe that the image quality and motion consistency is much better by using analogy constraint. + +within-domain reenactment though our method is designed for the cross-domain case. For the same-identity evaluation, we use 200 test videos from VoxCeleb. For cross-identity evaluation, we use the 200 videos to drive 1,000 images from CelebA-HQ. Each video drives five images. The obtained 1,000 videos are used to compute the metrics. Note that we use the released pre-trained models of competing methods here. + +The quantitative results are shown in Tab. 2 and Tab. 3. For same-identity evaluation, our method is overall comparable with DaGAN and better than LIA and FOMM. For cross-identity evaluation, our method achieves the best FID and CPBD. Our CSIM is comparable with others. + +Since the visual difference among different methods in same-identity reenactment is marginal, we only present a visual comparison of cross-identity reenactment in Fig. 6. It can be observed that our results have the best image quality, presenting more details in the face area, e.g., eyes and teeth. Besides, our method is comparable with LIA in preserving face shape and sometimes even better (see the last row in Fig. 6). + +In addition, we also perform a cross-identity face reenactment in the cartoon domain. We use 55 cartoon videos to drive 275 cartoon images. Results are shown in Tab. 4. Our method outperforms others in both FID and CPBD. + +# 4.4. Ablation Study + +To evaluate the training scheme, we perform an ablation study by dropping the analogy constraint. The quantitative results are shown in Tab. 5. 'w/o analogy' represents dropping the analogy constraint, i.e., finetuning the obtained model from the first stage with the cartoon dataset using only $\mathcal{L}_{rec}^{R}$ and $\mathcal{L}_{rec}^{C}$ in Eq.12. It can be observed that dropping the analogy constraint causes overall performance degradation, especially for the cartoon→real task. The qualitative evaluation is shown in Fig. 7. In the second row, the result without analogy constraints cannot accurately mimic the motion of the driving image. The eyes of the cartoon should be closed, but it is opened when dropping constraints. The dropping operation has a side effect + +
Real →CartoonCartoon → Real
FID↓CPBD↑FID↓CPBD↑
our full model28.5310.084031.8520.123
w/o analogy29.2300.074643.3530.107
+ +Table 5. Ablation study on the analogy constraint. + +
02040100
Real→Cartoon26.88528.53132.09441.936
Cartoon→Real33.46331.85230.17628.166
+ +Table 6. Ablation study on hyper-parameter ${\lambda }_{m}$ . + +
Fullmlpw/o backw/o sft
Real→Cartoon28.531116.74028.90431.291
Cartoon→Real31.85231.88535.26541.077
+ +Table 7. Ablation study on model architecture. + +on the motion consistency, which indicates that the analogy constraint contributes. + +To further analyze the influence of analogy constraint, we perform an ablation on the trade-off hyperparameter $\lambda_{m}$ . As is shown in Tab. 6, it can be observed that $\lambda_{m}$ trades off the performance of the two cross-domain tasks, i.e., Cartoon $\rightarrow$ Real and Real $\rightarrow$ Cartoon. We empirically choose $\lambda_{m} = 20$ in practical. + +We conduct ablation studies to demonstrate the effectiveness of several modules in the architecture, i.e., the query transformer, the backward transformer, and the spatial feature transformation. The ablation results are shown in Tab. 7. The metric is FID. "mlp" indicates that we replace the query transformer with multi-layer perception (MLP). "w/o back" denotes that we drop the backward transformer. And "w/o sft" drops the spatial feature transformation module. In Tab. 7, we can observe that query transformer brings significant improvement in the Real $\rightarrow$ Cartoon case. It verifies that the proposed query transformer can project the latent descriptor from different domains into a canonical motion space. Additionally, removing the backward transformer and sft will damage the quality of generated images. + +We also conduct a visualization experiment to verify the proposed canonical space. Since no paired data is available, we select three real images with different poses and expressions as templates to drive real and cartoon images, generating 450 real image pairs and 150 cartoon pairs. Each pair contains faces before and after driving. We compute their relative motions in the common motion space. In Fig. 8, 0, 1, and 2 represent real pairs, while 3, 4, and 5 represent cartoon pairs. $<0,3>$ , $<1,4>$ , and $<2,5>$ are driven by the same template, respectively. This shows that the motions of the two domains are aligned. + +![](images/37bf8e1ef987eda8b3c0ef1d40c3d203b6dea69936a14ace0f8aa0b8886c52f9.jpg) +Figure 8. Visualizing the canonical motion space. We visualize the t-SNE plot of motion codes in canonical space. $<0,1,2>$ and $<3,4,5>$ are from the relative motions of real domain and cartoon domain, respectively. And $<0,3>, <1,4>$ , and $<2,5>$ are driven by the same template, respectively. + +![](images/ffb368eead750d1db10c1531e05c4adf1e50196251112240f8e5f1d21a150f8a.jpg) +Figure 9. Failure cases. We observed that it is challenging to handle extreme poses. + +# 5. Conclusion + +We propose a novel transformer-based framework for cross-domain face reenactment. To settle the domain shift issue, we propose to align the motions of different domains in a canonical motion space. A set of domain-specific and domain-shared components are designed for motion alignment. Motion transfer can be implemented by adding motion codes in the shared motion space. Since there is no paired data for training, we propose a cross-domain training scheme with the analogy constraint using the two domain datasets. To alleviate the influence of imbalanced data, we use a two-stage learning strategy. Extensive experiments are conducted to demonstrate the superiority of our method over competing methods in cross-domain reenactment. + +Limitations. Like other SOTA models, our model cannot handle extreme poses well. As is shown in Fig 9, our method suffers from artifacts and deformations when the driving image is an extreme pose. This may be due to the fact that extreme poses rarely appear in the dataset. + +Acknowledgement. This work was partly supported by the National Natural Science Foundation of China (Grant No. U1903213) and the Shenzhen Science and Technology Program (JCYJ20220818101014030). + +# References + +[1] Aayush Bansal, Shugao Ma, Deva Ramanan, and Yaser Sheikh. Recycle-gan: Unsupervised video retargeting. In Proceedings of the European conference on computer vision (ECCV), pages 119-135, 2018. 2, 3 +[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In ACM SIGGRAPH, pages 187-194, 1999, 2 +[3] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5543-5552. ISSN: 1063-6919. 2 +[4] Stella Bounareli, Vasileios Argyriou, and Georgios Tzimiropoulos. Finding directions in GAN's latent space for neural face reenactment. 2, 3 +[5] Zhuo Chen, Chaoyue Wang, Bo Yuan, and Dacheng Tao. PuppeteerGAN: Arbitrary portrait animation with semantic-aware appearance transformation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13515-13524. IEEE. 2 +[6] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 6 +[7] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoria Sharmanska. HeadGAN: One-shot neural head synthesis and editing. 2, 3 +[8] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. MegaPortraits: One-shot megapixel neural head avatars. 3 +[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[10] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. 1, 3, 7 +[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 3 +[12] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. 3 +[13] Kangyeol Kim, Sunghyun Park, Jaeseong Lee, Sunghyo Chung, Junsoo Lee, and Jaegul Choo. AnimeCeleb: Large-scale animation CelebHeads dataset for head reenactment. 1, 2, 3 +[14] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5549-5558, 2020. 6, 7 +[15] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. VoxCeleb: a large-scale speaker identification dataset. In Interspeech 2017, pages 2616-2620. 6 + +[16] Niranjan D Narvekar and Lina J Karam. A no-reference image blur metric based on the cumulative probability of blur detection (cpbd). IEEE Transactions on Image Processing, 20(9):2678-2683, 2011. 6 +[17] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. PIRenderer: Controllable portrait image generation via semantic neural rendering. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 13739-13748. IEEE. 1, 2 +[18] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. version: 2.3 +[19] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. 1, 3, 4, 5, 7 +[20] Linsen Song, Wayne Wu, Chaoyou Fu, Chen Qian, Chen Change Loy, and Ran He. Everything's talkin': Pareidolia face reenactment. 3 +[21] Jiale Tao, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Motion transformer for unsupervised image animation. 3 +[22] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10034-10044. IEEE. 3 +[23] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. LATIVE IMAGE ANIMATOR: LEARNING TO ANIMATE IMAGES VIA LATENT SPACE NAVIGATION. page 17. 1, 3, 7 +[24] Wikipedia contributors. Morph target animation — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Morph_target Animation&oldid=1078665970, 2022. [Online; accessed 6-November-2022]. 1 +[25] Olivia Wiles, A. Sophia Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, volume 11217, pages 690–706. Springer International Publishing. Series Title: Lecture Notes in Computer Science. 3 +[26] Wayne Wu, Yunxuan Zhang, Cheng Li, Chen Qian, and Chen Change Loy. ReenactGAN: Learning to reenact faces via boundary transfer. 2 +[27] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 1 +[28] Borun Xu, Biao Wang, Jinhong Deng, Jiale Tao, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Motion and appearance adaptation for cross-domain motion transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 529-545. Springer, 2022. 2, 3, 7 +[29] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and + +Yujiu Yang. StyleHEAT: One-shot high-resolution editable talking face generation via pre-trained StyleGAN. 2, 3 +[30] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9458-9467. IEEE. 2 +[31] Jiangning Zhang, Xianfang Zeng, Mengmeng Wang, Yusu Pan, Liang Liu, Yong Liu, Yu Ding, and Changjie Fan. FReeNet: Multi-identity face reenactment. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5325-5334. IEEE. 2 +[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[33] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2 \ No newline at end of file diff --git a/2023/ToonTalker_ Cross-Domain Face Reenactment/images.zip b/2023/ToonTalker_ Cross-Domain Face Reenactment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6e777664b19637483cb6b0bed5f1b8bd81dd4f25 --- /dev/null +++ b/2023/ToonTalker_ Cross-Domain Face Reenactment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a85f3ee6d8de83e26a3be0598a80d38d9bdcbfbcfbd1632350d4b9fca9fc9d72 +size 739657 diff --git a/2023/ToonTalker_ Cross-Domain Face Reenactment/layout.json b/2023/ToonTalker_ Cross-Domain Face Reenactment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9e7f4daf1913c5b3e732e6ad0a7457c9d659aa76 --- /dev/null +++ b/2023/ToonTalker_ Cross-Domain Face Reenactment/layout.json @@ -0,0 +1,10802 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 103, + 442, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 103, + 442, + 119 + ], + "spans": [ + { + "bbox": [ + 153, + 103, + 442, + 119 + ], + "type": "text", + "content": "ToonTalker: Cross-Domain Face Reenactment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "spans": [ + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "text", + "content": "Yuan Gong" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "text", + "content": " Yong Zhang" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "text", + "content": " Xiaodong Cun" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "text", + "content": " Fei Yin" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "text", + "content": " Yanbo Fan" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "text", + "content": " Xuan Wang" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "text", + "content": " Baoyuan Wu" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "text", + "content": " Yujiu Yang" + }, + { + "bbox": [ + 157, + 142, + 441, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 95, + 171, + 499, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 171, + 499, + 213 + ], + "spans": [ + { + "bbox": [ + 95, + 171, + 499, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 171, + 499, + 213 + ], + "type": "text", + "content": "Shenzhen International Graduate School, Tsinghua University " + }, + { + "bbox": [ + 95, + 171, + 499, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 171, + 499, + 213 + ], + "type": "text", + "content": "Tencent AI Lab " + }, + { + "bbox": [ + 95, + 171, + 499, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 95, + 171, + 499, + 213 + ], + "type": "text", + "content": "Ant Group " + }, + { + "bbox": [ + 95, + 171, + 499, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 95, + 171, + 499, + 213 + ], + "type": "text", + "content": "The School of Data Science, Shenzhen Research Institute of Big Data, The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 266, + 290, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 614 + ], + "type": "text", + "content": "We target cross-domain face reenactment in this paper, i.e., driving a cartoon image with the video of a real person and vice versa. Recently, many works have focused on one-shot talking face generation to drive a portrait with a real video, i.e., within-domain reenactment. Straightforwardly applying those methods to cross-domain animation will cause inaccurate expression transfer, blur effects, and even apparent artifacts due to the domain shift between cartoon and real faces. Only a few works attempt to settle cross-domain face reenactment. The most related work AnimeCeleb [13] requires constructing a dataset with pose vector and cartoon image pairs by animating 3D characters, which makes it inapplicable anymore if no paired data is available. In this paper, we propose a novel method for cross-domain reenactment without paired data. Specifically, we propose a transformer-based framework to align the motions from different domains into a common latent space where motion transfer is conducted via latent code addition. Two domain-specific motion encoders and two learnable motion base memories are used to capture domain properties. A source query transformer and a driving one are exploited to project domain-specific motion to the canonical space. The edited motion is projected back to the domain of the source with a transformer. Moreover, since no paired data is provided, we propose a novel cross-domain training scheme using data from two domains with the designed analogy constraint. Besides, we contribute a cartoon dataset in Disney style. Extensive evaluations demonstrate the superiority of our method over competing methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 637, + 128, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 637, + 128, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 637, + 128, + 649 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 694 + ], + "type": "text", + "content": "Online video conferences and live streaming are spreading rapidly since the population of smartphones and communication techniques. Speakers sometimes present them" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 240, + 545, + 383 + ], + "blocks": [ + { + "bbox": [ + 307, + 240, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 307, + 240, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 545, + 383 + ], + "type": "image", + "image_path": "61b8aaf7d892918700650c5e39005a92db18b3cab23873327a4b3b758e625819.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 385, + 545, + 418 + ], + "lines": [ + { + "bbox": [ + 305, + 385, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 545, + 418 + ], + "type": "text", + "content": "Figure 1. Cross-domain reenactment examples of our method. Given a cartoon face, we animate it by transferring the pose and expression from a real face." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "content": "selves as cartoon characters for privacy and entertainment consideration. However, conventional cartoon animation depends on a complex pipeline of 3D modeling and retargeting based on a powerful render engine, which always takes a long period [24]. This limitation makes it difficult to drive a new model or directly drive a single comic picture." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 546, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 701 + ], + "type": "text", + "content": "There are many works [17, 19, 10, 23, 27] that focus on one-shot talking face generation, i.e., driving a portrait with a video of a real person, which belongs to within-domain reenactment. Their models are trained with a large amount of talking videos and perform well in the same-identity and cross-identity driving tasks. However, when those models are straightforwardly applied to cross-domain reenactment, i.e., driving a cartoon face with a real video as in Fig. 1, they persistently encounter the following issues, including inaccurate expression transfer, blur effects, and even apparent artifacts around facial components and the background. The reason is the distribution shift between different domains. There is a large appearance and motion gap between the cartoon and real datasets. Though fine-tuning a pre-trained model on a cartoon dataset can improve the visual quality to a certain extent, those issues still cannot be wiped out." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Only a few methods focus on visually-driven cross" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7690" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "domain face reenactment. The early method RecycleGAN [1] overfits two videos for unsupervised video translation, which cannot be used for one-shot reenactment. MAA [28] uses keypoints as an intermediate representation and designs an angle consistency loss to mimic the driving motion. But only focusing on keypoints will result in ignoring the subtle motion of facial components, especially the mouth and eyes. AnimeCeleb [13] is based on pre-defined 3D Morphable Models (3DMMs) [2] and additional pose annotations1 of cartoon images. However, it is inapplicable when paired data is not available, which limits its range of use. Moreover, the pre-defined pose vector has limited capability of capturing fine-grained facial expressions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 430 + ], + "type": "text", + "content": "In this paper, we propose a novel method for cross-domain reenactment without the requirement of paired data. All we need is a set of real videos and another set of cartoon videos. To solve the domain shift problem, we propose a transformer-based framework to align the motions from two domains in a shared canonical latent space. Specifically, two motion encoders and two motion bases are designed to discover the domain-specific properties of appearance and motion. To align the domain-specific motions in a common latent space, a source query transformer and a target take the responsibility of projecting the motions into a canonical motion space, where motion transfer can be conducted via the addition of motion codes. The edited motion is then projected from the canonical space back to the domain of the source by a transformer. A generator takes the projected motion and the source features as input to synthesize a face in the domain of the source." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 431, + 289, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 289, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 289, + 551 + ], + "type": "text", + "content": "Moreover, to overcome the lack of paired data, we propose a novel cross-domain training scheme. The core idea is the analogy constraint (see Fig. 2). Specifically, given two real faces and a cartoon face, we use the relative motion between the two real faces to drive the cartoon one, producing a synthetic cartoon face. The relative motion from one real face to the other real one is analogous to that from the cartoon face to the synthetic one. We measure the distance between the two types of analogical motions in the aligned canonical motion space. They are supposed to be identical." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 551, + 289, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 289, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 289, + 599 + ], + "type": "text", + "content": "To conduct cross-domain reenactment experiments, we collect a cartoon video dataset in Disney style, which contains 344 videos and about 47k frames in total. The cartoon dataset will be released for research purposes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 599, + 194, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 599, + 194, + 610 + ], + "spans": [ + { + "bbox": [ + 59, + 599, + 194, + 610 + ], + "type": "text", + "content": "Our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 618, + 287, + 687 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 618, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 618, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 58, + 618, + 287, + 666 + ], + "type": "text", + "content": "- We propose a novel framework for cross-domain face reenactment without using paired data. Several transformers are designed to align the motions of different domains in a common latent space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 674, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 674, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 58, + 674, + 287, + 687 + ], + "type": "text", + "content": "- We propose a training scheme to compensate for the" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 334, + 71, + 517, + 219 + ], + "blocks": [ + { + "bbox": [ + 334, + 71, + 517, + 219 + ], + "lines": [ + { + "bbox": [ + 334, + 71, + 517, + 219 + ], + "spans": [ + { + "bbox": [ + 334, + 71, + 517, + 219 + ], + "type": "image", + "image_path": "0383b6a34e320079cd79497edce396c30ecc481b2957108b08d23cca1d25549a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 223, + 547, + 334 + ], + "lines": [ + { + "bbox": [ + 304, + 223, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 547, + 334 + ], + "type": "text", + "content": "Figure 2. Visualization of analogy constraint. " + }, + { + "bbox": [ + 304, + 223, + 547, + 334 + ], + "type": "inline_equation", + "content": "\\Delta \\omega_{real}" + }, + { + "bbox": [ + 304, + 223, + 547, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 223, + 547, + 334 + ], + "type": "inline_equation", + "content": "\\Delta \\omega_{cartoon}" + }, + { + "bbox": [ + 304, + 223, + 547, + 334 + ], + "type": "text", + "content": " represent the relative motion in the real and cartoon domain, respectively. " + }, + { + "bbox": [ + 304, + 223, + 547, + 334 + ], + "type": "inline_equation", + "content": "\\Delta \\omega_{can}" + }, + { + "bbox": [ + 304, + 223, + 547, + 334 + ], + "type": "text", + "content": " represents the projected motion in the canonical space where we can perform cross-domain motion transfer by the addition of motion codes. Given a cartoon face, we utilize the relative motion between two real faces to animate the cartoon one and generate a synthetic cartoon face. We hold that the relative motion between two real faces and that between the cartoon face and the synthetic are equivalent within the canonical motion space." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 354, + 533, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 354, + 533, + 366 + ], + "spans": [ + { + "bbox": [ + 325, + 354, + 533, + 366 + ], + "type": "text", + "content": "lack of paired data by using the analogy constraint." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 375, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 375, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 317, + 375, + 545, + 410 + ], + "type": "text", + "content": "- We collect a cartoon video dataset and conduct extensive experiments to demonstrate the superiority of our method in the cross-domain reenactment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 422, + 392, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 392, + 435 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 392, + 435 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 443, + 489, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 489, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 489, + 455 + ], + "type": "text", + "content": "2.1. Within-domain Face Reenactment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 461, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 510 + ], + "type": "text", + "content": "Most one-shot talking face generation works focus on within-domain face reenactment, i.e., driving a portrait image using another talking head video, where both the driving video and source image is from real humans." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 653 + ], + "type": "text", + "content": "Keypoint-based Methods. Many reenactment algorithms [26, 30, 31, 5] are based on pre-defined facial landmarks that are used for motion conversion and repositioning. For example, ReenactGAN [26] encodes the face into the boundary hidden space defined by the keypoint heat map and transfers motion by a boundary transformer. Although ReenactGAN can produce high-quality target faces, it needs to retrain the model for the new individual images. FReeNet [31] trained a keypoint converter to convert the driven expression onto the source face. However, it can only deal with expression change, but not attitude changes, so the generated face image will keep the pose of the source face." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "3DMM-based Methods. A number of works [17, 7, 4, 29, 33] take advantage of the 3D deformation model [3] that can decouple pose, expression, and identity. For instance, Pirenderer [17] mapped the target 3DMM parameters into hidden codes and injected them into the source" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": "1Note that Pose in [13] refers to a 20 dimension (20D) vector that contains the status of facial expression and head pose." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7691" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "image features to guide the generation of distorted streams. HeadGAN [7] renders 3D parameters to images as motion representations. [4] map 3DMM parameters to StyleGAN [11, 12] latent space, and a transformation matrix is learned to map 3DMM parameters to latent code. Style-HEAT [29] used 3DMM to generate an optical flow map to change the feature map of StyleGAN. Using 3DMM parameters as motion representation can avoid the problem of video identity information leakage in cross-identity reenactment, but the 3D model only encodes facial regions, such as teeth, eyes, hair and other information that is difficult to synthesize." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 396 + ], + "type": "text", + "content": "Unsupervised Methods. Another large class of methods [25, 18, 19, 22, 10, 23, 8] do not use existing face models, but use unsupervised training to obtain facial motion representations. A typical method is FOMM [19], which proposes a first-order motion method by representing the motion as the key points and its local transformation. Facevid2vid [22] further extends this method to 3D facial key points. DAGAN [10] used the self-supervised geometry learning method to obtain the depth map from the video. Motion transformer [21] uses the interactive capabilities of transformer to predict key points. Differing from the above methods, LIA [23] does not require explicit structural representation but directly manipulates the latent space of the generative model. Megapotrait [8] follows a similar way and adds a super-resolution network for higher resolution." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 397, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 289, + 456 + ], + "type": "text", + "content": "The goals of these within-domain reenactment methods are different from ours since they do not need to consider the domain gap. They can be trained in a self-supervised way from the same video, while there are no cartoon-real training pairs in our cross-domain settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 466, + 224, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 224, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 224, + 479 + ], + "type": "text", + "content": "2.2. Cross-domain Face Reenactment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 485, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 677 + ], + "type": "text", + "content": "Only a few methods focus on cross-domain face reenactment. The early study Recycle-GAN [1] is based on image translation networks. It introduces a recurrent loss by training a temporal predictor to monitor the temporal consistency. And each character requires two face videos as two domains for training, which cannot be used for one-shot reenactment. Everythingtalking [20] proposes a parametric shape modeling technique parametric models based on Bezier Curve and transfers motion by adapting the human face's control points to the pareidolia face. But it is unable to change the facial posture. MAA [28] is a method that relies on keypoints and incorporates an angle consistency loss to replicate driving motion. However, the differences in keypoints only account for absolute motion transfer, which might significantly alter the appearance of source images and fail to capture intricate facial expressions well." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "The most related work AnimeCeleb [13] proposes a new cartoon talking-head video dataset. It uses a 3D animation model to generate a large number of animated face im" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 217 + ], + "type": "text", + "content": "ages and corresponding pose annotations. To achieve cross-domain reenactment, it proposes to learn the mapping relationship from animation pose vectors to 3DMM parameters by paired data. One shortcoming of AnimeCeleb is that it becomes inapplicable if no paired data is available, which limits its application scope. Moreover, it cannot be generalized to other cartoon styles that require rich expressions. Because AnimeCeleb uses the manually designed 20D vector to represent the status of pose and expression, which is far from enough. Differently, our method does not depend on 3DMMs and can be trained without being paired by using a designed cross-domain training scheme." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 225, + 362, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 225, + 362, + 238 + ], + "spans": [ + { + "bbox": [ + 306, + 225, + 362, + 238 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 246, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 547, + 413 + ], + "type": "text", + "content": "We propose a novel transformer-based framework to align the motions from different domains in a shared canonical motion space for motion transfer. The framework is shown in Fig. 3. The task becomes simple if there exist paired data, i.e., faces in different domains on the same expression and pose. Unfortunately, no such dataset is available. To settle the lack of paired data, we propose a training scheme with the designed analogy constraint, using faces from two domains jointly. We first introduce the architecture of the framework in Sec. 3.1. Then, we present the training scheme in Sec. 3.2. Finally, we show a two-stage learning strategy to alleviate the issue of imbalanced training data in Sec. 3.3, i.e., the number of real videos is much larger than that of cartoon videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 420, + 477, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 477, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 477, + 432 + ], + "type": "text", + "content": "3.1. Transformer-based Framework" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 438, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 617 + ], + "type": "text", + "content": "As shown in Fig. 3, given a source image and a driving one from two domains, we extract the motion information from the two images with two domain-specific motion encoders. Then, a source query transformer and a driving one are used to project the extracted motion information from different domains into the shared canonical motion space, with the domain-specific motion code base and the motion information as input. The motion transfer is realized by adding motion codes in the shared motion space. A domain-specific backward transformer projects the edited motion code back to the domain of the source image with the assistance of its features extracted via a domain-specific appearance encoder. Finally, a generator takes the projected motion code and the source features as input to render an image in the domain of the source." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": "Please note that we use the cartoon as the source and the real as the driving for illustration in Fig. 3. We can exchange the roles of the cartoon and the real, i.e., the real as the source while the cartoon is the driving. Therefore, in our full framework, we have two appearance encoders, two motion encoders, two backward transformers, two generators, and two motion code bases, which are domain-specific. We also have one source query transformer and one driv-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7692" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 312, + 243 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 312, + 243 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 312, + 243 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 312, + 243 + ], + "type": "image", + "image_path": "0777385d3f6a2231b65bfba5ed2ce616432d0731491f57c0b85a6c0be50fab7e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 547, + 315 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 547, + 315 + ], + "type": "text", + "content": "Figure 3. An illustration of our model, which can be mainly divided into three parts: (1) Feature extraction. We use " + }, + { + "bbox": [ + 46, + 247, + 547, + 315 + ], + "type": "inline_equation", + "content": "E_{a}" + }, + { + "bbox": [ + 46, + 247, + 547, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 247, + 547, + 315 + ], + "type": "inline_equation", + "content": "E_{m}" + }, + { + "bbox": [ + 46, + 247, + 547, + 315 + ], + "type": "text", + "content": " to extract local features and latent motion descriptor respectively. (2) Latent space transformation. A source query transformer and a driving one are used to project the motion descriptor into a canonical motion space. Motion transfer is conducted via the addition of motion codes in the shared motion space. (3) Image generation. The edited motion is projected back to the domain of the source with the assistant of the source features. A progressive image generator is used to render an image with the projected motion. We also apply a spatial feature transformation (SFT) to the last three high-resolution features to reduce the artifacts caused by feature warping." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 319, + 70, + 426, + 243 + ], + "blocks": [ + { + "bbox": [ + 319, + 70, + 426, + 243 + ], + "lines": [ + { + "bbox": [ + 319, + 70, + 426, + 243 + ], + "spans": [ + { + "bbox": [ + 319, + 70, + 426, + 243 + ], + "type": "image", + "image_path": "09459a17696775d8e6964ca82fe3ddf068706635ead81e9f0b3e63e29fc2c8bd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 434, + 71, + 542, + 243 + ], + "blocks": [ + { + "bbox": [ + 434, + 71, + 542, + 243 + ], + "lines": [ + { + "bbox": [ + 434, + 71, + 542, + 243 + ], + "spans": [ + { + "bbox": [ + 434, + 71, + 542, + 243 + ], + "type": "image", + "image_path": "c9ef07372e0a0ed88e1e5436431cc756cafaaaa3668728268b63842b3f57c9d6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 334, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 380 + ], + "type": "text", + "content": "ing query transformer, which are shared by two domains. The bounding box in color indicates whether a component is domain-specific in Fig. 3. We present the components in detail as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "text", + "content": "Appearance and Motion Encoders. We have domain-specific encoders to extract appearance and motion information from an image. Let " + }, + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "inline_equation", + "content": "E_{a}^{C}" + }, + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "inline_equation", + "content": "E_{m}^{C}" + }, + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "text", + "content": " denote the appearance and motion encoder for the cartoon domain, respectively, while " + }, + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "inline_equation", + "content": "E_{a}^{R}" + }, + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "inline_equation", + "content": "E_{m}^{R}" + }, + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "text", + "content": " for the real domain. As the roles of the two domains as the source and the target can be exchanged, we use " + }, + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "inline_equation", + "content": "E_{a}" + }, + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "inline_equation", + "content": "E_{m}" + }, + { + "bbox": [ + 46, + 381, + 287, + 513 + ], + "type": "text", + "content": " for simplicity. The appearance encoder takes the source image as input and outputs multi-scale spatial features, while the motion encoder extracts motion descriptors from the source and driving image, respectively. Thus, the feature extraction can be defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 521, + 287, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 521, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 138, + 521, + 287, + 534 + ], + "type": "interline_equation", + "content": "S = E _ {a} \\left(x _ {s}\\right), \\tag {1}", + "image_path": "3275221d51952a8e3136aeae13294a749ab5879cf842398bad79f2e380d8b246.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 536, + 287, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 536, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 287, + 550 + ], + "type": "interline_equation", + "content": "F _ {s} = E _ {m} \\left(x _ {s}\\right), F _ {d} = E _ {m} \\left(x _ {d}\\right), \\tag {2}", + "image_path": "0536e86bd6022719d7dda3c1058f05741eb1b40ebb781aebebcdcb772a8aee66.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "inline_equation", + "content": "F_{d}" + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "text", + "content": " denote the motion descriptors of the source " + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "text", + "content": " and the driving " + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "inline_equation", + "content": "S = \\{S^{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 46, + 557, + 287, + 593 + ], + "type": "text", + "content": " represents the appearance features of the source." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Motion Query Transformer. In addition to the domain-specific motion encoder, we use a motion base for each domain for further discovering domain-specific motion properties. The motion base consists of learnable embeddings to memorize typical motions, denoted as " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "p = \\{p^k\\}_{k=1}^K" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "p_k \\in R^d" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "-th motion embedding and " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " is the dimension. We adopt transformer blocks to project the latent descriptor into a canonical motion space due to their capability of capturing long-range dependencies. As shown in the middle of Fig. 3, the embedding is linearly projected" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 334, + 545, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 545, + 406 + ], + "type": "text", + "content": "to the query features while the latent descriptor from the image is projected to the key and value features. Inspired by [19], since it is challenging to model motion transformation " + }, + { + "bbox": [ + 304, + 334, + 545, + 406 + ], + "type": "inline_equation", + "content": "x_{s} \\rightarrow x_{d}" + }, + { + "bbox": [ + 304, + 334, + 545, + 406 + ], + "type": "text", + "content": " straightforwardly, we choose to model transformation " + }, + { + "bbox": [ + 304, + 334, + 545, + 406 + ], + "type": "inline_equation", + "content": "x_{s} \\rightarrow x_{r}" + }, + { + "bbox": [ + 304, + 334, + 545, + 406 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 334, + 545, + 406 + ], + "type": "inline_equation", + "content": "x_{r} \\rightarrow x_{d}" + }, + { + "bbox": [ + 304, + 334, + 545, + 406 + ], + "type": "text", + "content": " separately by leveraging a virtual reference " + }, + { + "bbox": [ + 304, + 334, + 545, + 406 + ], + "type": "inline_equation", + "content": "x_{r}" + }, + { + "bbox": [ + 304, + 334, + 545, + 406 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 375, + 417, + 545, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 417, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 375, + 417, + 545, + 429 + ], + "type": "interline_equation", + "content": "w _ {s \\rightarrow d} = w _ {s \\rightarrow r} + w _ {r \\rightarrow d}, \\tag {3}", + "image_path": "b931d5ce050422029af875e63ece444fabfffbe2f5bdc532749e3571f4ae2205.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "w_{s\\rightarrow r} = T_{s\\rightarrow r}(F_s,p)" + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "w_{r\\rightarrow d} = T_{r\\rightarrow d}(F_d,p)" + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "w_{s\\rightarrow r},w_{r\\rightarrow d}" + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "w_{s\\rightarrow d}" + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": " are three motion codes in canonical motion space. " + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "T_{s\\rightarrow r}" + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": " denotes the source query transformer that models the motion transformation from the source to the reference. While " + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "T_{r\\rightarrow d}" + }, + { + "bbox": [ + 304, + 437, + 545, + 533 + ], + "type": "text", + "content": " denotes the driving query transformer that models the motion transformation from the reference to the driving. The two query transformers are shared by the two domains." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "Backward Transformer. As we obtain the motion transformation " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "w_{s\\rightarrow d}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " in the canonical motion space, the next step is to project the motion code back to the domain of the source image. We use a domain-specific backward transformer to take over this task with the assistance of the information of the source image, i.e., the multi-scale features " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " of the source image from the appearance encoder " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "E_{a}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ". We have two such transformers, i.e., " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "T_B^C" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "T_B^R" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": ", and take the notation " + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "inline_equation", + "content": "T_{B}" + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": " for simplicity. To let the motion code obtain information from image features in different resolutions, we design the backward transformer as a stack of three transformer blocks. As shown on the right of Fig. 3, in each transformer block, the query features are derived from the output of the previous module, while the key and value features are computed from the corresponding multi-scale fea" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7693" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "tures. The first query features are the linear projection of the motion transformation " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "w_{s\\rightarrow d}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " from the canonical space. And the resolutions of the feature maps are " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "16\\times 16" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "32\\times 32" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "text", + "content": "Image Generator. After obtaining the projected motion code in the source domain, we use a domain-specific generator to progressively warp and refine the source feature maps layer by layer conditioned on the motion code " + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "inline_equation", + "content": "w_{s \\to d}" + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "text", + "content": ". We have two generators, i.e., " + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "inline_equation", + "content": "G^{C}" + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "inline_equation", + "content": "G^{R}" + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "text", + "content": ", and take the notation " + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "text", + "content": " for simplicity. Considering that warping them only will make the final output suffer from artifacts, we add refinement modules in the last three layers of the generator to enhance three high-resolution feature maps. Given the motion code " + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "inline_equation", + "content": "w_{s \\to d}" + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "text", + "content": " and the previous feature maps " + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "inline_equation", + "content": "o^{i-1}" + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "text", + "content": ", a StyleConv block is used to predict a flow field that is used to warp the feature map " + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "inline_equation", + "content": "S^{i}" + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "text", + "content": ". With the warped feature map " + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "inline_equation", + "content": "f_{w}^{i}" + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "text", + "content": " and the motion code " + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "inline_equation", + "content": "w_{s \\to d}" + }, + { + "bbox": [ + 46, + 121, + 288, + 288 + ], + "type": "text", + "content": " as input, the spatial feature transformation (SFT) for refinement is defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 293, + 287, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 293, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 126, + 293, + 287, + 308 + ], + "type": "interline_equation", + "content": "o ^ {i} = \\alpha^ {i} \\odot f _ {w} ^ {i} + \\beta^ {i}, \\tag {4}", + "image_path": "0bf3f2ebf516555fb956bc714264aa81ab33e0ec2117f3d4a083069933968cde.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 313, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 313, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 287, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 313, + 287, + 338 + ], + "type": "inline_equation", + "content": "\\alpha^i = \\mathrm{StyleConv}(f_w^i, w_{s \\to d})" + }, + { + "bbox": [ + 47, + 313, + 287, + 338 + ], + "type": "text", + "content": " is the spatial scale and " + }, + { + "bbox": [ + 47, + 313, + 287, + 338 + ], + "type": "inline_equation", + "content": "\\beta^i = \\mathrm{StyleConv}(f_w^i, w_{s \\to d})" + }, + { + "bbox": [ + 47, + 313, + 287, + 338 + ], + "type": "text", + "content": " is the bias." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 343, + 217, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 217, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 217, + 357 + ], + "type": "text", + "content": "3.2. Cross-domain Training Scheme" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 362, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 445 + ], + "type": "text", + "content": "If there are paired data where cartoon and real video pairs share the same head pose and expression, cross-domain training can be settled simply following the way of training a conventional within-domain model. However, no such paired data is available. To tackle this issue, we propose an analogy constraint to train the model with unpaired cartoons and real videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 446, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 566 + ], + "type": "text", + "content": "Analogy Constraint. As shown in Fig. 2, the core idea is to project the relative motions of different domains into the canonical motion space. For instance, given a tuple consisting of two real faces and a cartoon one, we can calculate the relative motion from one real face to the other real one. Then, the relative motion can be used to animate the cartoon face to generate a synthetic cartoon face. We can compute the relative motion from the cartoon face to the synthetic cartoon one. The two types of relative motions are supposed to be identical in canonical motion space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 566, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 287, + 589 + ], + "type": "text", + "content": "In the canonical motion space, following [19], the relative motion transfer can be defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 73, + 597, + 287, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 597, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 73, + 597, + 287, + 609 + ], + "type": "interline_equation", + "content": "w _ {s \\rightarrow d} = \\left(w _ {s \\rightarrow r} + w _ {r \\rightarrow s}\\right) + \\left(w _ {r \\rightarrow d} - w _ {r \\rightarrow 1}\\right), \\tag {5}", + "image_path": "7a0855bb7d30c8291d56bd8051f6472d1e956715e0b17cbb2d775c53c10b6792.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 615, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 287, + 664 + ], + "type": "text", + "content": "where '1' refers to a selected frame in the driving video. For illustration, let " + }, + { + "bbox": [ + 47, + 615, + 287, + 664 + ], + "type": "inline_equation", + "content": "x_{a}^{R}" + }, + { + "bbox": [ + 47, + 615, + 287, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 615, + 287, + 664 + ], + "type": "inline_equation", + "content": "x_{b}^{R}" + }, + { + "bbox": [ + 47, + 615, + 287, + 664 + ], + "type": "text", + "content": " denote two real images and " + }, + { + "bbox": [ + 47, + 615, + 287, + 664 + ], + "type": "inline_equation", + "content": "x_{s}^{C}" + }, + { + "bbox": [ + 47, + 615, + 287, + 664 + ], + "type": "text", + "content": " denote a cartoon image. Then, the relative motion transfer is applied to the cartoon, which can be defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 73, + 670, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 670, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 73, + 670, + 287, + 685 + ], + "type": "interline_equation", + "content": "w _ {s \\rightarrow g} = \\left(w _ {s \\rightarrow r} + w _ {r \\rightarrow s}\\right) + \\left(w _ {r \\rightarrow b} - w _ {r \\rightarrow a}\\right), \\tag {6}", + "image_path": "fbff345c0dc66b45e62c29177d73100f4db4c2557b37dc168a9c461e8fff8e6b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "w_{s\\rightarrow g}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " represents the motion transformation from " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "x_{s}^{C}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "x_{g}^{C}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": ". 'g' indicates the synthetic cartoon face. Each item" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 412, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 412, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 412, + 83 + ], + "type": "text", + "content": "can be defined as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 89, + 487, + 103 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 89, + 487, + 103 + ], + "spans": [ + { + "bbox": [ + 365, + 89, + 487, + 103 + ], + "type": "interline_equation", + "content": "w _ {s \\rightarrow r} = T _ {s \\rightarrow r} \\left(E _ {m} ^ {C} \\left(x _ {s} ^ {C}\\right), p ^ {C}\\right),", + "image_path": "d585be48348d7c2de2113ea763a0f9d1aa5a269521ecdc04f992ffb319bbe2ee.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 106, + 545, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 106, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 365, + 106, + 545, + 126 + ], + "type": "interline_equation", + "content": "w _ {r \\rightarrow s} = T _ {r \\rightarrow d} \\left(E _ {m} ^ {C} \\left(x _ {s} ^ {C}\\right), p ^ {C}\\right), \\tag {7}", + "image_path": "29ce7e44321e936a54ab6973e4af916fd37b3771adc67711eb16f1124cd07b86.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 365, + 121, + 487, + 136 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 121, + 487, + 136 + ], + "spans": [ + { + "bbox": [ + 365, + 121, + 487, + 136 + ], + "type": "interline_equation", + "content": "w _ {r \\rightarrow a} = T _ {r \\rightarrow d} \\left(E _ {m} ^ {R} \\left(x _ {a} ^ {R}\\right), p ^ {R}\\right),", + "image_path": "681d8e584e8aff06e68b0a5860b74920cd5f0cc7e9e103dfabd0cef9ce3cad29.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 365, + 139, + 487, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 139, + 487, + 152 + ], + "spans": [ + { + "bbox": [ + 365, + 139, + 487, + 152 + ], + "type": "interline_equation", + "content": "w _ {r \\rightarrow b} = T _ {r \\rightarrow d} \\left(E _ {m} ^ {R} \\left(x _ {b} ^ {R}\\right), p ^ {R}\\right).", + "image_path": "e05212860e0dc9096533eed1ab0b05e17bf71cf2831412916bdc69dd541a15ed.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 159, + 545, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 159, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 306, + 159, + 545, + 219 + ], + "type": "text", + "content": "With the motion code " + }, + { + "bbox": [ + 306, + 159, + 545, + 219 + ], + "type": "inline_equation", + "content": "w_{s \\to g}" + }, + { + "bbox": [ + 306, + 159, + 545, + 219 + ], + "type": "text", + "content": " as well as the source features " + }, + { + "bbox": [ + 306, + 159, + 545, + 219 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 306, + 159, + 545, + 219 + ], + "type": "text", + "content": ", the generator " + }, + { + "bbox": [ + 306, + 159, + 545, + 219 + ], + "type": "inline_equation", + "content": "G^{C}" + }, + { + "bbox": [ + 306, + 159, + 545, + 219 + ], + "type": "text", + "content": " can produce a synthetic cartoon face " + }, + { + "bbox": [ + 306, + 159, + 545, + 219 + ], + "type": "inline_equation", + "content": "x_{g}^{C}" + }, + { + "bbox": [ + 306, + 159, + 545, + 219 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 306, + 159, + 545, + 219 + ], + "type": "inline_equation", + "content": "x_{g}^{C} = G^{C}(T_{B}^{C}(w_{s \\to g}), E_{a}^{C}(x_{s}^{C}))" + }, + { + "bbox": [ + 306, + 159, + 545, + 219 + ], + "type": "text", + "content": ". Then, we can calculate the motion code of the synthetic cartoon face in the canonical motion space, i.e.," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 364, + 226, + 545, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 226, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 364, + 226, + 545, + 241 + ], + "type": "interline_equation", + "content": "w _ {r \\rightarrow g} = T _ {r \\rightarrow d} \\left(E _ {m} ^ {C} \\left(x _ {g} ^ {C}\\right), p ^ {C}\\right). \\tag {8}", + "image_path": "bb0404f1f2235a2fc975bd5f2d878befcf5b69cc95c9a37766eeb9807afda292.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 247, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 308 + ], + "type": "text", + "content": "The relative motion between the cartoon face and the synthetic one is " + }, + { + "bbox": [ + 305, + 247, + 545, + 308 + ], + "type": "inline_equation", + "content": "(w_{r\\rightarrow g} - w_{r\\rightarrow s})" + }, + { + "bbox": [ + 305, + 247, + 545, + 308 + ], + "type": "text", + "content": ", which is supposed to be identical to the relative motion between the two real faces, i.e., " + }, + { + "bbox": [ + 305, + 247, + 545, + 308 + ], + "type": "inline_equation", + "content": "(w_{r\\rightarrow b} - w_{r\\rightarrow a})" + }, + { + "bbox": [ + 305, + 247, + 545, + 308 + ], + "type": "text", + "content": ". Therefore, we can measure the distance between the two types of relative motions by" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 314, + 545, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 545, + 339 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m o t i o n}} ^ {R \\rightarrow C} = \\mathbb {E} [ \\| \\left(w _ {r \\rightarrow g} - w _ {r \\rightarrow s}\\right) - \\left(w _ {r \\rightarrow b} - w _ {r \\rightarrow a}\\right) \\| _ {1} ]. \\tag {9}", + "image_path": "b35e2b4318eb0e513b71869c7715adbd8b02058b020c65f9c817f923b8d27355.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 340, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 340, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 305, + 340, + 545, + 387 + ], + "type": "text", + "content": "Similarly, given two cartoon faces and one real face, we have " + }, + { + "bbox": [ + 305, + 340, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{motion}^{C\\rightarrow R}" + }, + { + "bbox": [ + 305, + 340, + 545, + 387 + ], + "type": "text", + "content": " following the same formulation. Further, to maintain the high fidelity of face generation, we impose adversarial loss. The adversarial loss can be defined as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 362, + 392, + 545, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 392, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 362, + 392, + 545, + 414 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} ^ {R} = \\mathbb {E} _ {x _ {g} ^ {R}} [ - \\log (D ^ {R} (x _ {g} ^ {R})) ], \\tag {10}", + "image_path": "9d496724817a82963ed82fdb688bccd9ff41ac8d7f3c2bdca45d3b858d6c7302.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 364, + 410, + 487, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 410, + 487, + 426 + ], + "spans": [ + { + "bbox": [ + 364, + 410, + 487, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a d v} ^ {C} = \\mathbb {E} _ {x _ {g} ^ {C}} [ - \\log (D ^ {C} (x _ {g} ^ {C})) ],", + "image_path": "09a987ef9f181899f957a36d601b553031efb8865bc0bd74a6abb5fb9f17a3b7.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 434, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 545, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 434, + 545, + 458 + ], + "type": "inline_equation", + "content": "D^{R}" + }, + { + "bbox": [ + 305, + 434, + 545, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 434, + 545, + 458 + ], + "type": "inline_equation", + "content": "D^{C}" + }, + { + "bbox": [ + 305, + 434, + 545, + 458 + ], + "type": "text", + "content": " are two discriminators for the real and cartoon domains, respectively." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "text", + "content": "Reconstruction. Besides the analogy constraint that emphasizes the interaction between two domains, we can also perform within-domain reenactment to assist the model learning. Specifically, we can sample two images within the same domain as the source and the driving, e.g., a pair of " + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "inline_equation", + "content": "x_{s}^{R}" + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "inline_equation", + "content": "x_{d}^{R}" + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "text", + "content": " or a pair of " + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "inline_equation", + "content": "x_{s}^{C}" + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "inline_equation", + "content": "x_{d}^{C}" + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "text", + "content": ". The reconstruction loss of the real pairs can be used to update the parameters of the real domain-related components as well as the shared motion query transformers. Similarly, the reconstruction loss of the cartoon pairs is for the corresponding components. We exploit the " + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "text", + "content": " loss " + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{1}(x_{s\\rightarrow d},x_{d})" + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "text", + "content": ", the perceptual loss " + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{per}(x_{s\\rightarrow d},x_{d})" + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "text", + "content": ", and the GAN loss " + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{adv}(x_{s\\rightarrow d},x_{d})" + }, + { + "bbox": [ + 305, + 459, + 545, + 613 + ], + "type": "text", + "content": ". Hence, the objective of reconstruction can be defined as:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 358, + 623, + 545, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 623, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 358, + 623, + 545, + 636 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c}} = \\mathcal {L} _ {1} + \\lambda_ {\\text {p e r}} \\mathcal {L} _ {\\text {p e r}} + \\mathcal {L} _ {\\text {a d v}}, \\tag {11}", + "image_path": "b85ecb528393080934849bc62a9976665ccb21cabcffee73a32ac884d004c4ea.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 645, + 531, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 531, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 531, + 657 + ], + "type": "text", + "content": "where we omit the domain notation R&C for simplicity." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 492, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 492, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 492, + 669 + ], + "type": "text", + "content": "The overall objective is defined as follows:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 675, + 545, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 675, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 306, + 675, + 545, + 700 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {r e c} ^ {R} + \\mathcal {L} _ {r e c} ^ {C} + \\lambda_ {m} * \\left(\\mathcal {L} _ {\\text {m o t i o n}} ^ {R \\rightarrow C} + \\mathcal {L} _ {\\text {m o t i o n}} ^ {C \\rightarrow R}\\right) + \\mathcal {L} _ {\\text {a d v}} ^ {C} + \\mathcal {L} _ {\\text {a d v}} ^ {R}, \\tag {12}", + "image_path": "3e8db580d2bdbcc5bb0b5b2bb4fce6b0a2fec22607dfade5252eff5b4aa4b567.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 702, + 477, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 702, + 477, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 702, + 477, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 702, + 477, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 306, + 702, + 477, + 714 + ], + "type": "text", + "content": " is the trade-off hyperparameter." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7694" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 120, + 144 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 120, + 144 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 120, + 144 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 120, + 144 + ], + "type": "image", + "image_path": "15f70bb361daaa1a641397abcd294b5cbd9ffa1ef1ee91e79cd5130122b76def.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 72, + 191, + 143 + ], + "blocks": [ + { + "bbox": [ + 121, + 72, + 191, + 143 + ], + "lines": [ + { + "bbox": [ + 121, + 72, + 191, + 143 + ], + "spans": [ + { + "bbox": [ + 121, + 72, + 191, + 143 + ], + "type": "image", + "image_path": "3b46cfe5329d3f720eb2f6d9278e006515ea869530013db651fab816b7cddeb6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 192, + 72, + 261, + 143 + ], + "blocks": [ + { + "bbox": [ + 192, + 72, + 261, + 143 + ], + "lines": [ + { + "bbox": [ + 192, + 72, + 261, + 143 + ], + "spans": [ + { + "bbox": [ + 192, + 72, + 261, + 143 + ], + "type": "image", + "image_path": "87aa5dee397bdaa1b0a102b7e0994dfd648455158baea16f6bde68530b429b05.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 261, + 72, + 332, + 143 + ], + "blocks": [ + { + "bbox": [ + 261, + 72, + 332, + 143 + ], + "lines": [ + { + "bbox": [ + 261, + 72, + 332, + 143 + ], + "spans": [ + { + "bbox": [ + 261, + 72, + 332, + 143 + ], + "type": "image", + "image_path": "136da6006347cc990177518b3b3c6d931299a4e1309d8ebb0534bee71f20edd4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 332, + 72, + 402, + 143 + ], + "blocks": [ + { + "bbox": [ + 332, + 72, + 402, + 143 + ], + "lines": [ + { + "bbox": [ + 332, + 72, + 402, + 143 + ], + "spans": [ + { + "bbox": [ + 332, + 72, + 402, + 143 + ], + "type": "image", + "image_path": "6c069edc8893e8c0e0ba48f7c0963c3244409d1404eaecffa3a0ac858d6cc1a5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 402, + 72, + 474, + 143 + ], + "blocks": [ + { + "bbox": [ + 402, + 72, + 474, + 143 + ], + "lines": [ + { + "bbox": [ + 402, + 72, + 474, + 143 + ], + "spans": [ + { + "bbox": [ + 402, + 72, + 474, + 143 + ], + "type": "image", + "image_path": "b509db33864fe08b4a0bd05d6649fe287039a2dfd9f8d9e143f0e92af29424fe.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 474, + 72, + 544, + 143 + ], + "blocks": [ + { + "bbox": [ + 474, + 72, + 544, + 143 + ], + "lines": [ + { + "bbox": [ + 474, + 72, + 544, + 143 + ], + "spans": [ + { + "bbox": [ + 474, + 72, + 544, + 143 + ], + "type": "image", + "image_path": "cc699f785c6cf9dbeb8ea4cc2249ee97a3917f4509a22d59b558dabe1d833afb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 144, + 120, + 213 + ], + "blocks": [ + { + "bbox": [ + 49, + 144, + 120, + 213 + ], + "lines": [ + { + "bbox": [ + 49, + 144, + 120, + 213 + ], + "spans": [ + { + "bbox": [ + 49, + 144, + 120, + 213 + ], + "type": "image", + "image_path": "746e393aad85bcbeb019c27c71fd14734277542152dd1e7ce67ddf05cf4ee0aa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 121, + 144, + 190, + 213 + ], + "blocks": [ + { + "bbox": [ + 121, + 144, + 190, + 213 + ], + "lines": [ + { + "bbox": [ + 121, + 144, + 190, + 213 + ], + "spans": [ + { + "bbox": [ + 121, + 144, + 190, + 213 + ], + "type": "image", + "image_path": "4f89aa8696d50fee10818afc819880c41ad83b1959f622fc1e5a8a09effe9d1b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 191, + 144, + 261, + 213 + ], + "blocks": [ + { + "bbox": [ + 191, + 144, + 261, + 213 + ], + "lines": [ + { + "bbox": [ + 191, + 144, + 261, + 213 + ], + "spans": [ + { + "bbox": [ + 191, + 144, + 261, + 213 + ], + "type": "image", + "image_path": "f368ecf62861dfca60a7bb8b97249ccd96510c391a0d4a018f9df881a74a3a0d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 261, + 144, + 332, + 213 + ], + "blocks": [ + { + "bbox": [ + 261, + 144, + 332, + 213 + ], + "lines": [ + { + "bbox": [ + 261, + 144, + 332, + 213 + ], + "spans": [ + { + "bbox": [ + 261, + 144, + 332, + 213 + ], + "type": "image", + "image_path": "71c21fef78925ed4f31b9298c72139b2c9db077266f489e52aec4ba5f41bc042.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 332, + 144, + 402, + 213 + ], + "blocks": [ + { + "bbox": [ + 332, + 144, + 402, + 213 + ], + "lines": [ + { + "bbox": [ + 332, + 144, + 402, + 213 + ], + "spans": [ + { + "bbox": [ + 332, + 144, + 402, + 213 + ], + "type": "image", + "image_path": "697f6cf14c2953d83f79e9e7860d3fcf212600884e66d567aa9497e88865b068.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 402, + 144, + 474, + 213 + ], + "blocks": [ + { + "bbox": [ + 402, + 144, + 474, + 213 + ], + "lines": [ + { + "bbox": [ + 402, + 144, + 474, + 213 + ], + "spans": [ + { + "bbox": [ + 402, + 144, + 474, + 213 + ], + "type": "image", + "image_path": "3fc7ff01840344ed3ed15b4cc58c12bf931bf8879b0a6dd4e25d0d385581e0b2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 474, + 144, + 544, + 213 + ], + "blocks": [ + { + "bbox": [ + 474, + 144, + 544, + 213 + ], + "lines": [ + { + "bbox": [ + 474, + 144, + 544, + 213 + ], + "spans": [ + { + "bbox": [ + 474, + 144, + 544, + 213 + ], + "type": "image", + "image_path": "55bb4d1f34ce718754f447c43f0852e5a0ebbab9fddc547fc1dbc6753c7e9ca1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 49, + 214, + 120, + 283 + ], + "blocks": [ + { + "bbox": [ + 49, + 214, + 120, + 283 + ], + "lines": [ + { + "bbox": [ + 49, + 214, + 120, + 283 + ], + "spans": [ + { + "bbox": [ + 49, + 214, + 120, + 283 + ], + "type": "image", + "image_path": "447a4e4982aa89abc459274d2d6d1532f7a734a2fddef937924e783b1effac32.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 121, + 214, + 190, + 283 + ], + "blocks": [ + { + "bbox": [ + 121, + 214, + 190, + 283 + ], + "lines": [ + { + "bbox": [ + 121, + 214, + 190, + 283 + ], + "spans": [ + { + "bbox": [ + 121, + 214, + 190, + 283 + ], + "type": "image", + "image_path": "43ba0535483f2f922220a5734f3b26a5b2c646d4bd1b6f4d4c3f5019c1796855.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 191, + 214, + 261, + 283 + ], + "blocks": [ + { + "bbox": [ + 191, + 214, + 261, + 283 + ], + "lines": [ + { + "bbox": [ + 191, + 214, + 261, + 283 + ], + "spans": [ + { + "bbox": [ + 191, + 214, + 261, + 283 + ], + "type": "image", + "image_path": "edef89eb42e5eca7ef21b6bafbb9d909df8244cb8d902561b9f317a5d4a54fc4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 261, + 214, + 332, + 283 + ], + "blocks": [ + { + "bbox": [ + 261, + 214, + 332, + 283 + ], + "lines": [ + { + "bbox": [ + 261, + 214, + 332, + 283 + ], + "spans": [ + { + "bbox": [ + 261, + 214, + 332, + 283 + ], + "type": "image", + "image_path": "c740cac3ad7a2624609f08ba34c21ee1d79970bd68e7b3173c6736f558b0e689.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 332, + 214, + 402, + 283 + ], + "blocks": [ + { + "bbox": [ + 332, + 214, + 402, + 283 + ], + "lines": [ + { + "bbox": [ + 332, + 214, + 402, + 283 + ], + "spans": [ + { + "bbox": [ + 332, + 214, + 402, + 283 + ], + "type": "image", + "image_path": "ddfab59f7739ab470275329d3f79433e26cabeef8672a2e0b6446be49a0206ed.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 402, + 214, + 473, + 283 + ], + "blocks": [ + { + "bbox": [ + 402, + 214, + 473, + 283 + ], + "lines": [ + { + "bbox": [ + 402, + 214, + 473, + 283 + ], + "spans": [ + { + "bbox": [ + 402, + 214, + 473, + 283 + ], + "type": "image", + "image_path": "042bf4f241b73434e9bc458d4a163ac99a6e7dc6aa7d514ff09d85406eaab4d1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 474, + 214, + 544, + 283 + ], + "blocks": [ + { + "bbox": [ + 474, + 214, + 544, + 283 + ], + "lines": [ + { + "bbox": [ + 474, + 214, + 544, + 283 + ], + "spans": [ + { + "bbox": [ + 474, + 214, + 544, + 283 + ], + "type": "image", + "image_path": "b00c0afab42688f0b6c36ae790193fcd468c5adf8c3ea37b44e8f2185cc0668a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 49, + 284, + 120, + 353 + ], + "blocks": [ + { + "bbox": [ + 49, + 284, + 120, + 353 + ], + "lines": [ + { + "bbox": [ + 49, + 284, + 120, + 353 + ], + "spans": [ + { + "bbox": [ + 49, + 284, + 120, + 353 + ], + "type": "image", + "image_path": "820948162f062253a2961b5676b81ea2d0256bf2c2f45ff6c0cec199e9a191fe.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 355, + 108, + 366 + ], + "lines": [ + { + "bbox": [ + 62, + 355, + 108, + 366 + ], + "spans": [ + { + "bbox": [ + 62, + 355, + 108, + 366 + ], + "type": "text", + "content": "(a) Source" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 121, + 284, + 190, + 353 + ], + "blocks": [ + { + "bbox": [ + 121, + 284, + 190, + 353 + ], + "lines": [ + { + "bbox": [ + 121, + 284, + 190, + 353 + ], + "spans": [ + { + "bbox": [ + 121, + 284, + 190, + 353 + ], + "type": "image", + "image_path": "3d9380c89cebe77dffccd731670d636bf9e8a2264eefa4f1bf7a26cabfa89dc3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 355, + 180, + 366 + ], + "lines": [ + { + "bbox": [ + 130, + 355, + 180, + 366 + ], + "spans": [ + { + "bbox": [ + 130, + 355, + 180, + 366 + ], + "type": "text", + "content": "(b) FOMM" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 191, + 284, + 261, + 353 + ], + "blocks": [ + { + "bbox": [ + 191, + 284, + 261, + 353 + ], + "lines": [ + { + "bbox": [ + 191, + 284, + 261, + 353 + ], + "spans": [ + { + "bbox": [ + 191, + 284, + 261, + 353 + ], + "type": "image", + "image_path": "1ca83750f818e07c2d3be39b96e6fe5e895f404a96e766823a3eef4284b7126e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 355, + 242, + 366 + ], + "lines": [ + { + "bbox": [ + 209, + 355, + 242, + 366 + ], + "spans": [ + { + "bbox": [ + 209, + 355, + 242, + 366 + ], + "type": "text", + "content": "(c) LIA" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 261, + 284, + 332, + 353 + ], + "blocks": [ + { + "bbox": [ + 261, + 284, + 332, + 353 + ], + "lines": [ + { + "bbox": [ + 261, + 284, + 332, + 353 + ], + "spans": [ + { + "bbox": [ + 261, + 284, + 332, + 353 + ], + "type": "image", + "image_path": "dd80d55c042ea52f16861b81b2a0a689541b41a943711f3152dffb222cb0142f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 355, + 323, + 366 + ], + "lines": [ + { + "bbox": [ + 270, + 355, + 323, + 366 + ], + "spans": [ + { + "bbox": [ + 270, + 355, + 323, + 366 + ], + "type": "text", + "content": "(d) DaGAN" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons with state-of-the-art methods on cross-domain reenactment. Competing methods are finetuned with the cartoon dataset. Hence all models are trained on the same data. It can be seen that our results have superior image quality and motion consistency. More visual results are in the supplementary." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 332, + 284, + 402, + 353 + ], + "blocks": [ + { + "bbox": [ + 332, + 284, + 402, + 353 + ], + "lines": [ + { + "bbox": [ + 332, + 284, + 402, + 353 + ], + "spans": [ + { + "bbox": [ + 332, + 284, + 402, + 353 + ], + "type": "image", + "image_path": "6ad27e5292613041d305a7a92c8b4dd5afb534b6abc86665443122629d7f9531.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 355, + 388, + 366 + ], + "lines": [ + { + "bbox": [ + 351, + 355, + 388, + 366 + ], + "spans": [ + { + "bbox": [ + 351, + 355, + 388, + 366 + ], + "type": "text", + "content": "(e) Ours" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 402, + 284, + 474, + 353 + ], + "blocks": [ + { + "bbox": [ + 402, + 284, + 474, + 353 + ], + "lines": [ + { + "bbox": [ + 402, + 284, + 474, + 353 + ], + "spans": [ + { + "bbox": [ + 402, + 284, + 474, + 353 + ], + "type": "image", + "image_path": "53aaaf1480412b5106be672953983cd06feed2dbdeb5ac559b2b157554ecb779.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 355, + 456, + 366 + ], + "lines": [ + { + "bbox": [ + 420, + 355, + 456, + 366 + ], + "spans": [ + { + "bbox": [ + 420, + 355, + 456, + 366 + ], + "type": "text", + "content": "(f) Start" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 474, + 284, + 544, + 353 + ], + "blocks": [ + { + "bbox": [ + 474, + 284, + 544, + 353 + ], + "lines": [ + { + "bbox": [ + 474, + 284, + 544, + 353 + ], + "spans": [ + { + "bbox": [ + 474, + 284, + 544, + 353 + ], + "type": "image", + "image_path": "e3a3b10f91610c298080c8d7d5a9c7650cf9353cf8b34460403e2200acd71e8a.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 355, + 534, + 367 + ], + "lines": [ + { + "bbox": [ + 484, + 355, + 534, + 367 + ], + "spans": [ + { + "bbox": [ + 484, + 355, + 534, + 367 + ], + "type": "text", + "content": "(g) Driving" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 47, + 425, + 263, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 425, + 263, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 425, + 263, + 437 + ], + "type": "text", + "content": "3.3. Two-stage Learning on Imbalanced Data." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 523 + ], + "type": "text", + "content": "The number of real videos is much larger than that of cartoon videos. Directly training with mixed imbalanced data cannot promote the performance of cross-domain reenactment because real faces dominate the data distribution. To alleviate the issue of the imbalanced dataset, we propose to learn the model in two stages." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 46, + 531, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 602 + ], + "type": "text", + "content": "In the first stage, we train the model with real videos by performing the reconstruction task, through which the model can learn the general knowledge of face motion from a large number of talking head videos. The obtained model cannot accurately transfer motion from real faces to cartoon ones due to domain shifts, especially facial expressions." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 46, + 610, + 287, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 610, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 46, + 610, + 287, + 669 + ], + "type": "text", + "content": "In the second stage, we copy the domain-specific components of the real domain as initialization for those of the cartoon domain. Then, we jointly learn all the components in the framework with the analogy constraint and the reconstruction loss." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 712 + ], + "type": "text", + "content": "Please note that our final model can be used to reenact a cartoon with a real video as well as to reenact a real face with a cartoon video." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 306, + 425, + 386, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 425, + 386, + 437 + ], + "spans": [ + { + "bbox": [ + 306, + 425, + 386, + 437 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 306, + 447, + 365, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 447, + 365, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 447, + 365, + 460 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 469, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 588 + ], + "type": "text", + "content": "Datasets. For cross-domain reenactment, we collect a cartoon video dataset in Disney style which contains 344 videos with an FPS of 30 and the total number of frames is about " + }, + { + "bbox": [ + 304, + 469, + 545, + 588 + ], + "type": "inline_equation", + "content": "47\\mathrm{k}" + }, + { + "bbox": [ + 304, + 469, + 545, + 588 + ], + "type": "text", + "content": ". To train our model, we take a subset of videos from the VoxCeleb [15] dataset containing 17K videos, and a subset from self-collected cartoon videos containing 289 videos. For evaluation, we use the rest 473 videos from the VoxCeleb [15] dataset and 55 videos from the cartoon dataset. We also use a set of images from CelebA-HQ [14] for cross-identity reenactment evaluation." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Metrics. We evaluate the performance from four aspects, i.e., reconstruction, fidelity, identity preservation, and motion transfer. " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ", Learned Perceptual Image Patch Similarity (LPIPS)[32], and Peak Signal-to-Noise Ratio (PSNR) are used to measure the reconstruction accuracy. Frechet Inception Distance (FID) [9] and Cumulative probability of Blur Detection (CPBD) [16] are used for measuring generation realism. The cosine similarity (CSIM) between synthetic and source images through ArcFace [6] is used for identity preservation. To evaluate the accuracy of mo" + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7695" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 282, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 282, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 282, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 282, + 228 + ], + "type": "image", + "image_path": "61f6b0f7df7251baed3e5399b2db63c86febbfb7d86d022898b7120db9aad243.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 287, + 274 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 287, + 274 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons with MAA [28]. Our method can better preserve the face shape of the original image. And more accurately imitated the driving image, especially the eyes and mouth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 342 + ], + "type": "text", + "content": "tion transfer, following [10], we use Average Keypoint Distance (AKD) and Average Expression Distance (AED) to compute the distance between synthetic and target images in terms of expression and pose." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 345, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 345, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 287, + 499 + ], + "type": "text", + "content": "Competing methods. We compare our method with several recent state-of-the-art methods, including FOMM [19], DAGAN [10], and LIA [23]. FOMM is a typical motion transfer method based on unsupervised keypoints. DAGAN introduces depth into the estimation of key points and has a similar formulation of motion transfer as FOMM. LIA is a representative method that uses a reference space similar to FOMM but conducts motion transfer in a latent space. For a fair comparison, we initialize the models using their pretrained official models and finetune them on the cartoon dataset. Therefore, we learn with the same data. Furthermore, we also compare with MAA [28], a keypoint-based cross-domain motion transfer method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 503, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 647 + ], + "type": "text", + "content": "Implementation Details. The model is trained in two stages to handle the imbalanced dataset where real videos are much more than cartoons. In the first stage, we trained the model on the training set of VoxCeleb for 450K iterations, where the batch size is 64 and the optimizer is Adam with an initial learning rate of " + }, + { + "bbox": [ + 46, + 503, + 287, + 647 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 503, + 287, + 647 + ], + "type": "text", + "content": ". In the second stage, we fine-tune the model with Disney videos with 10K iterations, where the batch size is 8 and the optimizer is Adam with an initial learning rate of " + }, + { + "bbox": [ + 46, + 503, + 287, + 647 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 503, + 287, + 647 + ], + "type": "text", + "content": ". The hyperparameters are set as " + }, + { + "bbox": [ + 46, + 503, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\lambda_{per} = 10" + }, + { + "bbox": [ + 46, + 503, + 287, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 503, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\lambda_{m} = 20" + }, + { + "bbox": [ + 46, + 503, + 287, + 647 + ], + "type": "text", + "content": " to balance generative fidelity and motion consistency. A detailed description of model components is in the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 226, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 226, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 226, + 670 + ], + "type": "text", + "content": "4.2. Cross-Domain Face Reenactment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "As mentioned in Sec. 3.3, our method can handle the bidirectional cross-domain reenactment. We execute motion transfer task for evaluation i.e., real " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " cartoon" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 547, + 151 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 547, + 151 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 547, + 151 + ], + "type": "table", + "html": "
Real→CartoonCartoon→Real
FID↓CPBD↑FID↓CSIM↑CPBD↑
FOMM30.3690.043651.9290.5360.0626
LIA30.4420.049348.6110.5150.0727
DaGAN29.7790.073547.2220.5610.0985
Ours28.5310.084031.8520.6550.1230
", + "image_path": "848961eeb304ca2c4bc49c97d198fd47a3d8d524188137313e7edd50b0152306.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 159, + 538, + 171 + ], + "lines": [ + { + "bbox": [ + 306, + 159, + 538, + 171 + ], + "spans": [ + { + "bbox": [ + 306, + 159, + 538, + 171 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on cross-domain reenactment." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 192, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 300 + ], + "type": "text", + "content": "and cartoon " + }, + { + "bbox": [ + 304, + 192, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 192, + 545, + 300 + ], + "type": "text", + "content": " real. The former represents driving a cartoon face with a real video, while the latter represents driving a real face with a cartoon video. As for real " + }, + { + "bbox": [ + 304, + 192, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 192, + 545, + 300 + ], + "type": "text", + "content": " cartoon, 60 videos from VoxCeleb are used to drive 300 cartoon images from the cartoon dataset. For a fair comparison, all competing methods are fine-tuned on the cartoon training set. As for cartoon " + }, + { + "bbox": [ + 304, + 192, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 192, + 545, + 300 + ], + "type": "text", + "content": " real, 55 cartoon videos are employed to drive 275 images of different identities from CelebA-HQ [14]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": "Qualitative Evaluation. The results of real " + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 302, + 545, + 398 + ], + "type": "text", + "content": " cartoon are shown in the first two rows of Fig. 4. First, our method achieves better image quality than other methods in terms of sharpness and artifacts. Second, our method outperforms others in motion transfer, especially facial expression. For instance, in the first row, the eyes of the cartoon should be closed according to the driving image. Only our method mimics driving while the eyes of others stay open." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "content": "The results of cartoon " + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "content": " real are shown in the last two rows of Fig. 4. We can observe a similar phenomenon as the real " + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 399, + 545, + 506 + ], + "type": "text", + "content": " cartoon case in terms of image quality, where our method works the best. FOMM and DaGAN become underperformed when there is a large head pose difference between the source and driving, as shown in the third and fourth rows. LIA is better than them but still worse than ours. For motion and identity preservation, our method surpasses others." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 508, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 603 + ], + "type": "text", + "content": "The comparison with the cross-domain motion transfer method MAA [28] is shown in Fig. 5. As the source code of MAA is not released, its results are adapted from the original paper. Our method has better, especially in the mouth, eyes and image quality. For instance, in the third column, MAA ignores the motion of the eyes, whereas our method accurately mimics the driving image. In the last column, our mouth shape is more consistent with the driving one." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "content": "Quantitative Evaluation. The quantitative results are reported in Tab. 1. Our method achieves the best FID in the two tasks, indicating that the synthetic results are most consistent with the source distribution. Our method also performs the best in sharpness and identity preservation, i.e., the highest CPBD and CSIM, which is consistent with the observation in visual results. These results demonstrate that our model can capture the appearance and motion properties of different domains and the motion transfer in the canoni" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7696" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 286, + 233 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 233 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 233 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 233 + ], + "type": "image", + "image_path": "d5a34432666915a490b55848cd7685217972d53067b98ea48d4a616ef37f330b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 237, + 287, + 270 + ], + "lines": [ + { + "bbox": [ + 47, + 237, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 287, + 270 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons on within-domain reenactment. Our method achieves better image quality than other methods in terms of sharpness, distortion and artifacts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 281, + 282, + 350 + ], + "blocks": [ + { + "bbox": [ + 52, + 281, + 282, + 350 + ], + "lines": [ + { + "bbox": [ + 52, + 281, + 282, + 350 + ], + "spans": [ + { + "bbox": [ + 52, + 281, + 282, + 350 + ], + "type": "table", + "html": "
L1↓LPIPS↓CPBD↑AKD↓AED↓
FOMM0.04660.1390.03081.3880.142
LIA0.04610.1230.03721.5120.153
DaGAN0.04420.1220.04151.3000.129
Ours0.04570.1210.04291.3350.139
", + "image_path": "7b3a09333358a71e5823ffffd57b5d7a34bb018e6c4961b7498a713bd1e033af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 52, + 379, + 282, + 448 + ], + "blocks": [ + { + "bbox": [ + 70, + 357, + 263, + 368 + ], + "lines": [ + { + "bbox": [ + 70, + 357, + 263, + 368 + ], + "spans": [ + { + "bbox": [ + 70, + 357, + 263, + 368 + ], + "type": "text", + "content": "Table 2. Same-identity reconstruction in real domain" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 379, + 282, + 448 + ], + "lines": [ + { + "bbox": [ + 52, + 379, + 282, + 448 + ], + "spans": [ + { + "bbox": [ + 52, + 379, + 282, + 448 + ], + "type": "table", + "html": "
FID↓CSIM↑CPBD ↑
FOMM54.4110.5410.0689
LIA53.1610.5220.0787
DaGAN49.9200.5480.1069
Ours48.7130.5260.1132
", + "image_path": "f09a5a85c6c8e4996ace405ef7d650a9f584c79c37d18cc54dba138e74de5013.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 51, + 478, + 284, + 522 + ], + "blocks": [ + { + "bbox": [ + 74, + 456, + 259, + 467 + ], + "lines": [ + { + "bbox": [ + 74, + 456, + 259, + 467 + ], + "spans": [ + { + "bbox": [ + 74, + 456, + 259, + 467 + ], + "type": "text", + "content": "Table 3. Cross-identity reenactment in real domain" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 478, + 284, + 522 + ], + "lines": [ + { + "bbox": [ + 51, + 478, + 284, + 522 + ], + "spans": [ + { + "bbox": [ + 51, + 478, + 284, + 522 + ], + "type": "table", + "html": "
FOMMLIADAGANOurs
FID↓40.85334.06636.21827.467
CPBD↑0.03680.05230.05980.0846
", + "image_path": "699f02b23cb4b9fd652d85106482f4742c69916b1c49ed2e26984b24b5c2cc94.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 530, + 266, + 541 + ], + "lines": [ + { + "bbox": [ + 67, + 530, + 266, + 541 + ], + "spans": [ + { + "bbox": [ + 67, + 530, + 266, + 541 + ], + "type": "text", + "content": "Table 4. Cross-identity reenactment in cartoon domain." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 563, + 167, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 167, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 167, + 574 + ], + "type": "text", + "content": "cal motion space works well." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "type": "text", + "content": "User Study. To further evaluate the visual quality, we conduct a user study by asking 20 human raters to answer 20 multiple-choice questions. In each question, a rater chooses the best from 4 synthetic cartoon videos generated by the three competing methods and ours. Our method is the most favorable with a selection rate of " + }, + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "type": "inline_equation", + "content": "72.75\\%" + }, + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "type": "text", + "content": ". While the rates are " + }, + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "type": "inline_equation", + "content": "13.75\\%" + }, + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "type": "text", + "content": " for LIA, DaGAN, and FOMM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 670, + 233, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 233, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 233, + 682 + ], + "type": "text", + "content": "4.3. Within-Domain Face Reenactment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Besides the cross-domain evaluation, we also compare our method with competing methods in the scenario of" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 174 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 174 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 174 + ], + "type": "image", + "image_path": "3662346031899134b75f2bd9a0390cfe06d5f3fcd137fef5472807059c70b29f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 176, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 305, + 176, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 176, + 545, + 209 + ], + "type": "text", + "content": "Figure 7. Qualitative ablation study on the analogy constraint. We observe that the image quality and motion consistency is much better by using analogy constraint." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 230, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 324 + ], + "type": "text", + "content": "within-domain reenactment though our method is designed for the cross-domain case. For the same-identity evaluation, we use 200 test videos from VoxCeleb. For cross-identity evaluation, we use the 200 videos to drive 1,000 images from CelebA-HQ. Each video drives five images. The obtained 1,000 videos are used to compute the metrics. Note that we use the released pre-trained models of competing methods here." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 326, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 385 + ], + "type": "text", + "content": "The quantitative results are shown in Tab. 2 and Tab. 3. For same-identity evaluation, our method is overall comparable with DaGAN and better than LIA and FOMM. For cross-identity evaluation, our method achieves the best FID and CPBD. Our CSIM is comparable with others." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 386, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 481 + ], + "type": "text", + "content": "Since the visual difference among different methods in same-identity reenactment is marginal, we only present a visual comparison of cross-identity reenactment in Fig. 6. It can be observed that our results have the best image quality, presenting more details in the face area, e.g., eyes and teeth. Besides, our method is comparable with LIA in preserving face shape and sometimes even better (see the last row in Fig. 6)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 483, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 530 + ], + "type": "text", + "content": "In addition, we also perform a cross-identity face reenactment in the cartoon domain. We use 55 cartoon videos to drive 275 cartoon images. Results are shown in Tab. 4. Our method outperforms others in both FID and CPBD." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 539, + 400, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 400, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 400, + 552 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "To evaluate the training scheme, we perform an ablation study by dropping the analogy constraint. The quantitative results are shown in Tab. 5. 'w/o analogy' represents dropping the analogy constraint, i.e., finetuning the obtained model from the first stage with the cartoon dataset using only " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}^{R}" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}^{C}" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " in Eq.12. It can be observed that dropping the analogy constraint causes overall performance degradation, especially for the cartoon→real task. The qualitative evaluation is shown in Fig. 7. In the second row, the result without analogy constraints cannot accurately mimic the motion of the driving image. The eyes of the cartoon should be closed, but it is opened when dropping constraints. The dropping operation has a side effect" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7697" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 298, + 125 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 298, + 125 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 298, + 125 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 298, + 125 + ], + "type": "table", + "html": "
Real →CartoonCartoon → Real
FID↓CPBD↑FID↓CPBD↑
our full model28.5310.084031.8520.123
w/o analogy29.2300.074643.3530.107
", + "image_path": "9ce5e5f9a5f4041e23895359174be0ebf9496f0d07a18dce2d13d0ac9b95da52.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 156, + 294, + 201 + ], + "blocks": [ + { + "bbox": [ + 76, + 133, + 257, + 144 + ], + "lines": [ + { + "bbox": [ + 76, + 133, + 257, + 144 + ], + "spans": [ + { + "bbox": [ + 76, + 133, + 257, + 144 + ], + "type": "text", + "content": "Table 5. Ablation study on the analogy constraint." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 156, + 294, + 201 + ], + "lines": [ + { + "bbox": [ + 48, + 156, + 294, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 156, + 294, + 201 + ], + "type": "table", + "html": "
02040100
Real→Cartoon26.88528.53132.09441.936
Cartoon→Real33.46331.85230.17628.166
", + "image_path": "f621967d77daec51e6ea4b3e5f0994c4cc66638e7b55ab90a0aba1e80793e204.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 232, + 296, + 276 + ], + "blocks": [ + { + "bbox": [ + 79, + 209, + 254, + 219 + ], + "lines": [ + { + "bbox": [ + 79, + 209, + 254, + 219 + ], + "spans": [ + { + "bbox": [ + 79, + 209, + 254, + 219 + ], + "type": "text", + "content": "Table 6. Ablation study on hyper-parameter " + }, + { + "bbox": [ + 79, + 209, + 254, + 219 + ], + "type": "inline_equation", + "content": "{\\lambda }_{m}" + }, + { + "bbox": [ + 79, + 209, + 254, + 219 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 232, + 296, + 276 + ], + "lines": [ + { + "bbox": [ + 48, + 232, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 296, + 276 + ], + "type": "table", + "html": "
Fullmlpw/o backw/o sft
Real→Cartoon28.531116.74028.90431.291
Cartoon→Real31.85231.88535.26541.077
", + "image_path": "de91fd4f7254b42fd87c0361761ab765851ae68c843c241afb1a819f6ded3e19.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 284, + 251, + 295 + ], + "lines": [ + { + "bbox": [ + 82, + 284, + 251, + 295 + ], + "spans": [ + { + "bbox": [ + 82, + 284, + 251, + 295 + ], + "type": "text", + "content": "Table 7. Ablation study on model architecture." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 316, + 286, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 286, + 339 + ], + "type": "text", + "content": "on the motion consistency, which indicates that the analogy constraint contributes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 341, + 286, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 286, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 286, + 413 + ], + "type": "text", + "content": "To further analyze the influence of analogy constraint, we perform an ablation on the trade-off hyperparameter " + }, + { + "bbox": [ + 47, + 341, + 286, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 47, + 341, + 286, + 413 + ], + "type": "text", + "content": ". As is shown in Tab. 6, it can be observed that " + }, + { + "bbox": [ + 47, + 341, + 286, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{m}" + }, + { + "bbox": [ + 47, + 341, + 286, + 413 + ], + "type": "text", + "content": " trades off the performance of the two cross-domain tasks, i.e., Cartoon " + }, + { + "bbox": [ + 47, + 341, + 286, + 413 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 47, + 341, + 286, + 413 + ], + "type": "text", + "content": " Real and Real " + }, + { + "bbox": [ + 47, + 341, + 286, + 413 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 47, + 341, + 286, + 413 + ], + "type": "text", + "content": " Cartoon. We empirically choose " + }, + { + "bbox": [ + 47, + 341, + 286, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_{m} = 20" + }, + { + "bbox": [ + 47, + 341, + 286, + 413 + ], + "type": "text", + "content": " in practical." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 414, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 286, + 581 + ], + "type": "text", + "content": "We conduct ablation studies to demonstrate the effectiveness of several modules in the architecture, i.e., the query transformer, the backward transformer, and the spatial feature transformation. The ablation results are shown in Tab. 7. The metric is FID. \"mlp\" indicates that we replace the query transformer with multi-layer perception (MLP). \"w/o back\" denotes that we drop the backward transformer. And \"w/o sft\" drops the spatial feature transformation module. In Tab. 7, we can observe that query transformer brings significant improvement in the Real " + }, + { + "bbox": [ + 46, + 414, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 414, + 286, + 581 + ], + "type": "text", + "content": " Cartoon case. It verifies that the proposed query transformer can project the latent descriptor from different domains into a canonical motion space. Additionally, removing the backward transformer and sft will damage the quality of generated images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "We also conduct a visualization experiment to verify the proposed canonical space. Since no paired data is available, we select three real images with different poses and expressions as templates to drive real and cartoon images, generating 450 real image pairs and 150 cartoon pairs. Each pair contains faces before and after driving. We compute their relative motions in the common motion space. In Fig. 8, 0, 1, and 2 represent real pairs, while 3, 4, and 5 represent cartoon pairs. " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "<0,3>" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "<1,4>" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "<2,5>" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " are driven by the same template, respectively. This shows that the motions of the two domains are aligned." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 359, + 72, + 494, + 208 + ], + "blocks": [ + { + "bbox": [ + 359, + 72, + 494, + 208 + ], + "lines": [ + { + "bbox": [ + 359, + 72, + 494, + 208 + ], + "spans": [ + { + "bbox": [ + 359, + 72, + 494, + 208 + ], + "type": "image", + "image_path": "37bf8e1ef987eda8b3c0ef1d40c3d203b6dea69936a14ace0f8aa0b8886c52f9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 216, + 545, + 271 + ], + "lines": [ + { + "bbox": [ + 305, + 216, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 545, + 271 + ], + "type": "text", + "content": "Figure 8. Visualizing the canonical motion space. We visualize the t-SNE plot of motion codes in canonical space. " + }, + { + "bbox": [ + 305, + 216, + 545, + 271 + ], + "type": "inline_equation", + "content": "<0,1,2>" + }, + { + "bbox": [ + 305, + 216, + 545, + 271 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 216, + 545, + 271 + ], + "type": "inline_equation", + "content": "<3,4,5>" + }, + { + "bbox": [ + 305, + 216, + 545, + 271 + ], + "type": "text", + "content": " are from the relative motions of real domain and cartoon domain, respectively. And " + }, + { + "bbox": [ + 305, + 216, + 545, + 271 + ], + "type": "inline_equation", + "content": "<0,3>, <1,4>" + }, + { + "bbox": [ + 305, + 216, + 545, + 271 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 216, + 545, + 271 + ], + "type": "inline_equation", + "content": "<2,5>" + }, + { + "bbox": [ + 305, + 216, + 545, + 271 + ], + "type": "text", + "content": " are driven by the same template, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 283, + 544, + 358 + ], + "blocks": [ + { + "bbox": [ + 307, + 283, + 544, + 358 + ], + "lines": [ + { + "bbox": [ + 307, + 283, + 544, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 544, + 358 + ], + "type": "image", + "image_path": "ffb368eead750d1db10c1531e05c4adf1e50196251112240f8e5f1d21a150f8a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 360, + 545, + 383 + ], + "lines": [ + { + "bbox": [ + 306, + 360, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 545, + 383 + ], + "type": "text", + "content": "Figure 9. Failure cases. We observed that it is challenging to handle extreme poses." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 403, + 378, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 378, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 378, + 415 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 425, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 579 + ], + "type": "text", + "content": "We propose a novel transformer-based framework for cross-domain face reenactment. To settle the domain shift issue, we propose to align the motions of different domains in a canonical motion space. A set of domain-specific and domain-shared components are designed for motion alignment. Motion transfer can be implemented by adding motion codes in the shared motion space. Since there is no paired data for training, we propose a cross-domain training scheme with the analogy constraint using the two domain datasets. To alleviate the influence of imbalanced data, we use a two-stage learning strategy. Extensive experiments are conducted to demonstrate the superiority of our method over competing methods in cross-domain reenactment." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 593, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 653 + ], + "type": "text", + "content": "Limitations. Like other SOTA models, our model cannot handle extreme poses well. As is shown in Fig 9, our method suffers from artifacts and deformations when the driving image is an extreme pose. This may be due to the fact that extreme poses rarely appear in the dataset." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was partly supported by the National Natural Science Foundation of China (Grant No. U1903213) and the Shenzhen Science and Technology Program (JCYJ20220818101014030)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7698" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Aayush Bansal, Shugao Ma, Deva Ramanan, and Yaser Sheikh. Recycle-gan: Unsupervised video retargeting. In Proceedings of the European conference on computer vision (ECCV), pages 119-135, 2018. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 167 + ], + "type": "text", + "content": "[2] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In ACM SIGGRAPH, pages 187-194, 1999, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "type": "text", + "content": "[3] James Booth, Anastasios Roussos, Stefanos Zafeiriou, Allan Ponniah, and David Dunaway. A 3d morphable model learnt from 10,000 faces. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5543-5552. ISSN: 1063-6919. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 256 + ], + "type": "text", + "content": "[4] Stella Bounareli, Vasileios Argyriou, and Georgios Tzimiropoulos. Finding directions in GAN's latent space for neural face reenactment. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "text", + "content": "[5] Zhuo Chen, Chaoyue Wang, Bo Yuan, and Dacheng Tao. PuppeteerGAN: Arbitrary portrait animation with semantic-aware appearance transformation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13515-13524. IEEE. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 288, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 367 + ], + "type": "text", + "content": "[6] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "type": "text", + "content": "[7] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoria Sharmanska. HeadGAN: One-shot neural head synthesis and editing. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "type": "text", + "content": "[8] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. MegaPortraits: One-shot megapixel neural head avatars. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 288, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 489 + ], + "type": "text", + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 523 + ], + "type": "text", + "content": "[10] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. 1, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 546 + ], + "type": "text", + "content": "[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 547, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 580 + ], + "type": "text", + "content": "[12] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 581, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 623 + ], + "type": "text", + "content": "[13] Kangyeol Kim, Sunghyun Park, Jaeseong Lee, Sunghyo Chung, Junsoo Lee, and Jaegul Choo. AnimeCeleb: Large-scale animation CelebHeads dataset for head reenactment. 1, 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[14] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5549-5558, 2020. 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[15] Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. VoxCeleb: a large-scale speaker identification dataset. In Interspeech 2017, pages 2616-2620. 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[16] Niranjan D Narvekar and Lina J Karam. A no-reference image blur metric based on the cumulative probability of blur detection (cpbd). IEEE Transactions on Image Processing, 20(9):2678-2683, 2011. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[17] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H. Li, and Shan Liu. PIRenderer: Controllable portrait image generation via semantic neural rendering. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 13739-13748. IEEE. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 173, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 206 + ], + "type": "text", + "content": "[18] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. Animating arbitrary objects via deep motion transfer. version: 2.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 208, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 240 + ], + "type": "text", + "content": "[19] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. 1, 3, 4, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[20] Linsen Song, Wayne Wu, Chaoyou Fu, Chen Qian, Chen Change Loy, and Ran He. Everything's talkin': Pareidolia face reenactment. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[21] Jiale Tao, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Motion transformer for unsupervised image animation. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 354 + ], + "type": "text", + "content": "[22] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10034-10044. IEEE. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 399 + ], + "type": "text", + "content": "[23] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. LATIVE IMAGE ANIMATOR: LEARNING TO ANIMATE IMAGES VIA LATENT SPACE NAVIGATION. page 17. 1, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 400, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 545, + 454 + ], + "type": "text", + "content": "[24] Wikipedia contributors. Morph target animation — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Morph_target Animation&oldid=1078665970, 2022. [Online; accessed 6-November-2022]. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 456, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 533 + ], + "type": "text", + "content": "[25] Olivia Wiles, A. Sophia Koepke, and Andrew Zisserman. X2face: A network for controlling face generation using images, audio, and pose codes. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, volume 11217, pages 690–706. Springer International Publishing. Series Title: Lecture Notes in Computer Science. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "type": "text", + "content": "[26] Wayne Wu, Yunxuan Zhang, Cheng Li, Chen Qian, and Chen Change Loy. ReenactGAN: Learning to reenact faces via boundary transfer. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 623 + ], + "type": "text", + "content": "[27] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven 3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 689 + ], + "type": "text", + "content": "[28] Borun Xu, Biao Wang, Jinhong Deng, Jiale Tao, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Motion and appearance adaptation for cross-domain motion transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 529-545. Springer, 2022. 2, 3, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[29] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7699" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 329 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "Yujiu Yang. StyleHEAT: One-shot high-resolution editable talking face generation via pre-trained StyleGAN. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[30] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9458-9467. IEEE. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "text", + "content": "[31] Jiangning Zhang, Xianfang Zeng, Mengmeng Wang, Yusu Pan, Liang Liu, Yong Liu, Yu Ding, and Changjie Fan. FReeNet: Multi-identity face reenactment. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5325-5334. IEEE. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "type": "text", + "content": "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 329 + ], + "type": "text", + "content": "[33] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7700" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/a134e8ff-997a-47c5-810d-0149806d46fc_content_list.json b/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/a134e8ff-997a-47c5-810d-0149806d46fc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f27f45773bc2e83d79ae460ff59a2bd19cedd015 --- /dev/null +++ b/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/a134e8ff-997a-47c5-810d-0149806d46fc_content_list.json @@ -0,0 +1,1736 @@ +[ + { + "type": "text", + "text": "TopoSeg: Topology-Aware Nuclear Instance Segmentation", + "text_level": 1, + "bbox": [ + 191, + 130, + 779, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongliang He $^{1,2*}$ Jun Wang $^{1*}$ Pengxu Wei $^{3}$ Fan Xu $^{4}$ Xiangyang Ji $^{5}$ Chang Liu $^{5}$ Jie Chen $^{1,4}$", + "bbox": [ + 238, + 179, + 730, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Electronic and Computer Engineering, Peking University, Shenzhen, China \n $^{2}$ School of Computer Science and Technology, Soochow University, Suzhou, China \n $^{3}$ Sun Yat-sen University, Guangzhou, China \n $^{4}$ Peng Cheng Laboratory, Shenzhen, China \n $^{5}$ Department of Automation and BNRist, Tsinghua University, Beijing, China", + "bbox": [ + 129, + 222, + 839, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hehl@pku.edu.cn, junw@stu.pku.edu.cn, weipx3@mail.sysu.edu.cn, {xuf01,chenj}@pcl.ac.cn, {xyji,liuchang2022}@tsinghua.edu.cn", + "bbox": [ + 212, + 297, + 761, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 364, + 313, + 380 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nuclear instance segmentation has been critical for pathology image analysis in medical science, e.g., cancer diagnosis. Current methods typically adopt pixel-wise optimization for nuclei boundary exploration, where rich structural information could be lost for subsequent quantitative morphology assessment. To address this issue, we develop a topology-aware segmentation approach, termed TopoSeg, which exploits topological structure information to keep the predictions rational, especially in common situations with densely touching and overlapping nucleus instances. Concretely, TopoSeg builds on a topology-aware module (TAM), which encodes dynamic changes of different topology structures within the three-class probability maps (inside, boundary, and background) of the nuclei to persistence barcodes and makes the topology-aware loss function. To efficiently focus on regions with high topological errors, we propose an adaptive topology-aware selection (ATS) strategy to enhance the topology-aware optimization procedure further. Experiments on three nuclear instance segmentation datasets justify the superiority of TopoSeg, which achieves state-of-the-art performance. The code is available at https://github.com/hhlisme/toposeg.", + "bbox": [ + 75, + 397, + 473, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 758, + 209, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pathological image analysis is considered the gold standard for cancer diagnosis, treatment, and prevention, where nuclear instance segmentation serves as the core step. It is because the morphology of nuclei including the shape, appearance, and distribution within pathology images provides interpretable features of diagnostic and prognostic", + "bbox": [ + 75, + 784, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad304ca217be5ad48b46bb88830e5649fecc03d361e7a3b459c790b29d378d6b.jpg", + "image_caption": [ + "Figure 1. The schematic illustration of pixel-wise loss and topology-aware loss. (Left) A patch in the image and its corresponding boundary ground truth, as well as the original boundary segmentation probability map and its corresponding instance prediction. (Top-right) The gradient map was obtained by pixel-wise loss, the 3D map of the boundary probability map and the optimization trend, as well as the final instance prediction. (Bottom-right) The gradient map obtained from topology-aware loss, the 3D image of the boundary probability map and the optimization trend, as well as the final instance prediction." + ], + "image_footnote": [], + "bbox": [ + 506, + 366, + 888, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cancer indicators[22].", + "bbox": [ + 500, + 684, + 648, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compared to general image object segmentation, nuclear instance segmentation faces with some particular challenges: (i) the crowded nuclei distribution leads to clustered overlapping instances. (ii) The blurry boundaries between touching or overlapping nuclei increase the difficulty of distinguishing individual instances. (iii) Large variability in size, shape, texture and intensity of nuclei between different images requires more robust annotation and prediction [18].", + "bbox": [ + 496, + 702, + 892, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To cope with these challenges, several studies have focused on exploring innovative network architectures [12, 23, 25] and incorporating distance maps [3, 9, 19, 21] to enhance segmentation performance. However, existing works typically rely on pixel-wise losses, such as CE/Dice loss,", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution ☑ Corresponding Author", + "bbox": [ + 94, + 886, + 346, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21307", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which enforce predictions to align with GT pixel-by-pixel yet ignore the crucial structure information of instances. Since treat pixels independently, they lack sensitivity towards high-level structural characteristics, such as the connectivity of boundaries, which are essential for the rational segmentation of individual instances. This inability of pixel-wise loss often leads to unreasonable results such as broken boundaries, bringing about failure to distinguish touching or overlapping instances. This consequently results in undesirable effects on the subsequent quantitative morphology analysis. As depicted in Fig. 1 (top-right), pixel-wise loss focuses on misaligned pixels thus reducing pixel errors. However, due to the lack of perception of structural information, the boundaries between overlapping instances with smaller pixel proportions are not given enough consideration and result in incorrect instance predictions.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To capture structural information of nuclear instances, we propose a novel topology-aware nuclear instance segmentation approach, termed TopoSeg. Specifically, we introduce a topology encoding method to encode topological features of three-class maps which are utilized to separate individual instances. Building upon this, we propose a topology-aware module (TAM), which encodes dynamic changes and persistence of different topology structures within the predicted three-class probability maps to persistence barcodes and develops a topology-aware loss function, integrating topology information into model learning. As shown in Fig. 1 (bottom-right), with topology constraints, the critical boundaries are identified and optimized, resulting in a reasonable segmentation outcome. To further adapt to the task of dense and small instances like nuclei, we introduce an adaptive topology-aware selection module (ATS) to concentrate on images and regions requiring topological optimization. Notably, TopoSeg effectively promotes instance segmentation performance without requiring any changes to the original network structure.", + "bbox": [ + 75, + 333, + 472, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions can be summarized as follows:", + "bbox": [ + 96, + 636, + 421, + 650 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose TopoSeg, which provides a new perspective that goes beyond the common pixel-wise optimization paradigm and incorporates meaningful topology constraints for nuclear instance segmentation.", + "- We present a TAM to encode the persistence of topology structures in probability maps and introduce a topology-aware loss as a further constraint.", + "- We develop an ATS to identify regions that most require topology optimization, enabling the model to efficiently accommodate segmentation tasks involving small and dense objects.", + "- We justify the superiority of our TopoSeg on three nuclei datasets with state-of-the-art performance." + ], + "bbox": [ + 96, + 664, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 642, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Nuclei Segmentation", + "text_level": 1, + "bbox": [ + 500, + 114, + 694, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning has been extensively utilized for nuclear instance segmentation [2, 9, 11, 24]. In 2015, Ronneberger et al. [25] proposed the U-Net model, which has become one of the most fundamental models in medical image segmentation. Raza et al. [24] proposed Micro-Net, which achieves robustness to large internal/external variances in nucleus size by utilizing multi-resolution and weighted loss functions. Qu et al. [23] proposed a full-resolution convolutional neural network (FullNet) with no downsampling operation employed in the network structure to enhance localization accuracy. He et al. [12] proposed a hybrid attention nested U-shaped network (Han-Net) to extract effective feature information from multiple layers.", + "bbox": [ + 496, + 138, + 890, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To take advantage of contour information to differentiate contact/overlapping nuclei, Chen et al. [2] initially proposed incorporating contour information into a multi-level FCN to create a deep contour-aware network for nuclear instance segmentation. Subsequently, Zhou et al. [28] proposed the contour-aware information aggregation network to combine spatial and textural features between nuclei and contours. Additionally, some works [3, 9, 19, 21] introduced distance maps to separate contact/overlapping nuclei. Naylor et al. [21] addressed the issue of segmenting touching nuclei by formulating the segmentation task as a regression task of intra-nuclear distance maps. Graham et al. [9] proposed a network (Hover-Net) for simultaneous nucleus segmentation and classification, which uses the vertical and horizontal distances between a nucleus pixel and its center of mass to separate clusters of nuclei. Moreover, He et al. [11] proposed a centripetal directional network (CDNet) for nuclear instance segmentation, combining directional information into the network.", + "bbox": [ + 496, + 335, + 890, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The above models have achieved some performance improvements in nuclear segmentation, but these models using pixel-wise loss tend to ignore structural information of nuclear boundaries.", + "bbox": [ + 496, + 623, + 890, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Segmentation Topology", + "text_level": 1, + "bbox": [ + 500, + 695, + 718, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several topology-based approaches [1, 14, 16, 20, 26] have exploited ways to guarantee the topological accuracy of objects in images. Mosinska et al. [20] considered topological information by constraining the similarity between feature maps and predicted probability maps generated by a pre-trained VGG-19 network. Shit et al. [26] proposed a similarity metric (clDice) to calculate the intersection between segmentation masks and their morphological skeletons, and developed a novel loss function (Soft-clDice) to train the neural network based on the skeleton structure.", + "bbox": [ + 496, + 718, + 890, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, the differentiability of persistent homology [6] allows it to be formulated as a loss function and ap-", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21308", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aad59e3199bb222ed617aacb0807f12618ee40d2a222fdc8b2a864451edc0441.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 90, + 890, + 255 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/af60d89dded5e5049dfa0c31a65e1eb0870b30661e568c58986d22d928aa184f.jpg", + "image_caption": [ + "Figure 2. The overall framework of the proposed TopoSeg. The network contains two basic prediction branches: the semantic branch and the three-class (inside, boundary, background) branch. In order to be aware of topology, the framework contains two important modules: the adaptive topology-aware selection module and the topology-aware module." + ], + "image_footnote": [], + "bbox": [ + 83, + 256, + 890, + 401 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "plied to end-to-end deep neural networks. Hu et al. [14] developed a loss function that directly constrains the prediction results and topological GT based on persistent homology. Furthermore, Hu et al. [16] used Discrete Morse Theory [7, 8] to capture the singularities of the gradient vector field of the likelihood function to find all potential skeletons of the object. They also leveraged Discrete Morse Theory and persistent homology to construct a topological representation space and trained a probabilistic model for performing inference tasks in such a topological representation space [15]. In cardiac magnetic resonance (CMR), Clough et al. [1, 5, 4] investigated various methods to couple topological information, where the number of components in the required segmentation is known a priori.", + "bbox": [ + 75, + 465, + 468, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the above methods have certain limitations in terms of computational efficiency, topology completeness, and task flexibility. Therefore, we propose a topology-aware model for nuclear instance segmentation, which encodes structural information to construct topological losses, thus improving the instance segmentation performance.", + "bbox": [ + 75, + 678, + 468, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 787, + 210, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 814, + 187, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 provides a flowchart of our TopoSeg, which adopts a dual-branch architecture with two types of predictions: semantic prediction of nuclear semantic categories (e.g. carcinoma, epithelial, etc.) and three-class prediction of struc", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tural categories (inside, boundary, background). The output of three-class branch is sequentially processed through an adaptive topology-aware selection module (ATS) and a topology-aware module (TAM) to efficiently capture the topology information of nuclear instances.", + "bbox": [ + 496, + 465, + 890, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, to quantify the topological features of an image, we present topology encoding in Sec. 3.2. Topological information contained in three-class images, such as the number of holes and connected components of the boundary, can reflect the number of instances and overlapping characteristics, which is meaningful for reasonable instance segmentation. Therefore, we decompose the three-class image into three channels (inside, boundary, and union) and encode their topological features. This allows the direct representation of the structural correctness, but its inapplicability to predicated probability maps makes it challenging to incorporate into the training process.", + "bbox": [ + 496, + 544, + 892, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this issue, we introduce TAM in Sec. 3.3, which extracts topological structures of predicted three-class probability maps into persistence barcode through a persistence encoding method and constrains the topology similarity between GTs and predictions by a topology-aware loss.", + "bbox": [ + 496, + 729, + 890, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To further improve the topology perception of the model and reduce the computational cost, we develop ATS in Sec. 3.4 to adaptively select regions with greater topology importance in both image- and patch-level, which makes the topological constraints more suitable for the task of seg", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21309", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b323df670db911b86f3dc58dfd9a1b544cabf8465bf8ba3f980f42a86d55f9c9.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 83, + 102, + 176, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dc29468e81133cb7c07a101e05a81dc3f2af7662f7cb1d36f0f96948ba72cce7.jpg", + "image_caption": [ + "Prediction 1" + ], + "image_footnote": [], + "bbox": [ + 179, + 102, + 272, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9f08e60781f98b4bc5b8677527b0b7ed0c4667400c941332451b8df16a77ae54.jpg", + "image_caption": [ + "Prediction 2" + ], + "image_footnote": [], + "bbox": [ + 276, + 102, + 369, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/845bfed57d0f6fb281172a474f4cf172bea3dd5f377bad0f4220dc9c423e16fb.jpg", + "image_caption": [ + "Prediction 3" + ], + "image_footnote": [], + "bbox": [ + 372, + 102, + 464, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d9b6f36fbaef075ffbe9068d710dd3347f1cb54adeea3a8e0c77efd15d6c548a.jpg", + "image_caption": [ + "Topology encoding" + ], + "image_footnote": [], + "bbox": [ + 117, + 172, + 138, + 184 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/77a5769a0794d305397b1d43a797459dbf56ca01cea8c4e5319e0bdc8cf65d14.jpg", + "image_caption": [ + "Figure 3. (Top) The ground truth and different prediction results of two overlapping nuclei in the pathological image. (Bottom) An example of topology encoding of ground truth." + ], + "image_footnote": [], + "bbox": [ + 83, + 188, + 464, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "menting dense and small instances.", + "bbox": [ + 76, + 320, + 312, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In general, during the training phase, the total loss function of TopoSeg is", + "bbox": [ + 75, + 335, + 468, + 366 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p i x e l}} + \\lambda \\mathcal {L} _ {\\text {t o p o}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 378, + 468, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{pixel}$ denotes pixel-wise loss with respect to the output of both branches, $\\mathcal{L}_{topo}$ denotes the proposed topology-aware in TAM with respect to the output at three-class branch and $\\lambda$ denotes the weight of topology-aware loss.", + "bbox": [ + 75, + 407, + 468, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The design of this total loss leverages the strengths of both types of losses. Initially, the model is trained with only the pixel-wise loss for a certain number of iterations ( $\\lambda = 0$ ) to obtain accurate pixel-level predictions. Subsequently, the topology loss is introduced ( $\\lambda > 0$ ) to further enhance the model's ability to capture rational object structure.", + "bbox": [ + 75, + 469, + 468, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the inference phase, the final predicted instance segmentation map is obtained by utilizing boundaries predicted in the three-class branch to separate the semantic result of the semantic branch.", + "bbox": [ + 75, + 561, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Topology Encoding", + "text_level": 1, + "bbox": [ + 76, + 632, + 261, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As previously noted, pixel-wise loss is insensitive to the differences in a few pixels at the boundary between overlapping nuclei, as distinct examples depicted in Fig. 3, despite significant structural differences. To quantitatively represent them, we present a topology encoding approach whose general representation is as follows.", + "bbox": [ + 75, + 656, + 468, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For image with $K$ classes, consider pixels of class $i$ as set $S_{i}$ . The zero-dimensional topology structure refers to the connection components formed by the connected pixels in $S_{i}$ , and the one-dimensional structure refers to the holes. The topology of $S_{i}$ can be considered as a family of subsets $\\{S_{ij}^{d}|d = 0,1;j = 1,2,\\dots,b_{d}\\}$ formed by topological structures of different dimensions, where $d$ represents topology dimension and $b_{d}$ represents the number of subsets of dimension $d$ . Then a subset $S_{ij}^{0}$ represents a connected component and $S_{ij}^{1}$ represents a hole, Fig. 3 (Bottom).", + "bbox": [ + 75, + 748, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To quantify the topological characteristics of $S_{i}$ , we introduce the Betti number from algebraic topology. For a 2D image, the Betti numbers can be expressed as $B = (b_{0}, b_{1})$ where $b_{0}$ denotes the number of connected components and $b_{1}$ denotes the number of 2D holes. Given a K-class image, the topology encoding can be formulated as", + "bbox": [ + 498, + 90, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} = \\left\\{B _ {i} ^ {d} | i = 1, 2, \\dots , K; d = 0, 1 \\right\\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 188, + 890, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B_{i}^{d}$ represents Betti number of subset $S_{i}^{d}$ .", + "bbox": [ + 498, + 210, + 816, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the nuclear instance segmentation task, the topology encoding of instance maps are mapped to three-class maps with three categories (inside, boundary and their union):", + "bbox": [ + 496, + 227, + 890, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} = \\left[ \\begin{array}{l} B _ {i n} \\\\ B _ {b o} \\\\ B _ {u n} \\end{array} \\right] = \\left[ \\begin{array}{c c} N _ {i n} & 0 \\\\ N _ {b o} & N _ {i n} \\\\ N _ {b o} & 0 \\end{array} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 279, + 890, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B_{in}, B_{bo}$ and $B_{un}$ respectively represent the Betti numbers of the corresponding class. The specific values of $B_{i}^{d}$ are defined based on the number of instances $N_{in}$ and non-connected boundaries $N_{bo}$ in the image.", + "bbox": [ + 498, + 335, + 890, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike the fixed topology prior of semantic given in [10] and [1], our topology encoding represents dynamically changing instance information for each image through three channels. As an example, the corresponding topology encoding of images in Fig. 3 (Top) is:", + "bbox": [ + 496, + 396, + 890, + 472 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} (G T) = \\left[ \\begin{array}{c c} 2 & 0 \\\\ 1 & 2 \\\\ 1 & 0 \\end{array} \\right], \\mathbf {B} (P 1) = \\left[ \\begin{array}{c c} 2 & 0 \\\\ 1 & 2 \\\\ 1 & 0 \\end{array} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 489, + 841, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} (P 2) = \\left[ \\begin{array}{c c} 1 & 0 \\\\ 1 & 1 \\\\ 1 & 0 \\end{array} \\right], \\mathbf {B} (P 3) = \\left[ \\begin{array}{c c} 2 & 0 \\\\ 2 & 2 \\\\ 2 & 0 \\end{array} \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 541, + 844, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Topology-Aware Module", + "text_level": 1, + "bbox": [ + 500, + 597, + 723, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Persistence Encoding", + "text_level": 1, + "bbox": [ + 500, + 619, + 700, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consider a probability map of class $i$ as a topology set $S_{i}$ with topology subsets $\\{S_{ij}^{d}|d = 0,1;j = 1,2,\\dots,J_{d}\\}$ , where $J_{d}$ represents the total number of subsets of dimension $d$ . Here we use persistent homology to extract meaningful topology structures which are robust to noise and small perturbations in the probability map.", + "bbox": [ + 496, + 643, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first step of persistent homology is to construct a sequence of combinations $\\{C_n\\}$ of subsets, where each combination is obtained by adding more and more elements from $S_{i}$ , which can be obtained by thresholding the probability map at a sequence of decreasing threshold values from 1 to 0. During this process, some subsets disappear with the gradual increase of pixels, while others appear. In Fig. 4, subsets of different dimensions are marked with letters of different colors. The subset $a$ and $b$ both appear in $C_1$ , and $a$ survives since then while $b$ disappears in $C_2$ because the added pixels connect $b$ and $a$ into a common subset.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21310", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/411528c836fcc76276cab4b88943bd4b0242fc91875dfdbcdba76c3f7d1fae3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 90, + 447, + 166 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0df07790ea3c4adf4b72eaa9ebe438d5122c16940531b0167a10f733ac7c8256.jpg", + "image_caption": [ + "Figure 4. The binary images obtained by taking different threshold $\\mathbf{p}$ of the boundary probability map and corresponding persistence barcode for topology structures. The red letters indicate connected components $(d = 0)$ , which is signified by the blue solid bars in the persistence barcode map. The golden letters indicate holes $(d = 1)$ , which are signified by the blue open bars." + ], + "image_footnote": [], + "bbox": [ + 81, + 167, + 464, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the persistence of $S_{ij}^{d}$ is measured by how long it survives in the sequence of combinations, which can be expressed through the persistence barcode whose length represents the corresponding persistence, Fig. 4. The longer the persistence barcode, the more robust the subset. Thereby, the persistence $P_{ij}^{d}$ of $S_{ij}^{d}$ can be calculated as the difference between the threshold values at which $S_{ij}^{d}$ first appears $(p_{birth}(S_{ij}^{d}))$ and when it disappears $(p_{death}(S_{ij}^{d}))$ :", + "bbox": [ + 75, + 430, + 468, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i j} ^ {d} = p _ {\\text {b i r t h}} \\left(S _ {i j} ^ {d}\\right) - p _ {\\text {d e a t h}} \\left(S _ {i j} ^ {d}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 565, + 468, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given prediction probabilities with K classes of a 2D image, the persistence encoding $P$ is", + "bbox": [ + 76, + 588, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} = \\left\\{P _ {i j} ^ {d} \\mid i = 1, 2, \\dots , K; d = 0, 1; j = 1, 2, \\dots , J _ {d} \\right\\}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 627, + 468, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that since ground truth can be regarded as a probability map containing only binary values of 0 and 1, the persistence encoding of it is", + "bbox": [ + 76, + 655, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = \\left\\{Y _ {i j} ^ {d} \\mid i = 1, 2, \\dots , K; d = 0, 1; j = 1, 2, \\dots , B _ {i} ^ {d} \\right\\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 708, + 468, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the value of $Y_{ij}^{d}$ is 1. As predicted probability maps typically contain more topological structures than those in ground truth, we assume that $J_{d} > B_{i}^{d}$ .", + "bbox": [ + 76, + 737, + 468, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Topology-Aware Loss", + "text_level": 1, + "bbox": [ + 76, + 801, + 281, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on persistence encoding, a better prediction should meet the condition that the subsets of $C_n$ consistent with ground truth have greater persistence. We match the top $B_i^d$ persistence barcode $P_{ij}^d$ with GT and regard the remaining codes as unmatched. Our purpose is to reduce the error", + "bbox": [ + 76, + 824, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "between the persistence encoding of prediction and ground truth, so as to make the topology correctness of prediction more robust. Therefore, the difference between them can be calculated in two parts:", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i} = \\sum_ {d \\in \\{0, 1 \\}} \\sum_ {j = 1} ^ {B _ {i} ^ {d}} \\left(Y _ {i j} ^ {d} - P _ {i j} ^ {d}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 160, + 890, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU _ {i} = \\sum_ {d \\in \\{0, 1 \\}} \\sum_ {j = B _ {i} ^ {d} + 1} ^ {J _ {d}} (P _ {i j} ^ {d}),\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 210, + 782, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{i}$ measures the length difference between the matched barcode and GT barcode for the $i$ th class, and $U_{i}$ measures the length of the unmatched barcode.", + "bbox": [ + 500, + 267, + 890, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given prediction probabilities with $K$ classes of a 2D image, the topology-aware loss $\\mathcal{L}_{topo}$ is", + "bbox": [ + 500, + 313, + 890, + 343 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t o p o} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\left(M _ {i} + U _ {i}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 349, + 890, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Minimizing this loss is equivalent to maximizing the persistence of topology features that are consistent with ground truth, and minimizing the persistence of undesired features. This provides a structure perspective to optimize the model, which overcomes the limitation of pixel-wise loss and achieves superior nucleus segmentation performance.", + "bbox": [ + 498, + 398, + 890, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Adaptive Topology-Aware Selection (ATS)", + "text_level": 1, + "bbox": [ + 500, + 500, + 857, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ATS aims at efficiently optimizing images with topology-important regions by providing image- and patch-level selection for loss calculation, Fig. 5.", + "bbox": [ + 498, + 522, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the image level, we calculate the topology error between prediction and GT, then selects and stores images with an adaptive threshold, Fig. 5 (Left). If the topology error exceeds the threshold, it indicates that the overall structure of the image is significantly different from ground truth, then the image is preferentially selected and stored for further learning. On the contrary, if the topology error does not exceed the threshold, it will be dropped out in the next iteration. The adaptive threshold is determined based on the current number of iterations and topology error rate. The more iterations, the higher the adaptive threshold.", + "bbox": [ + 496, + 568, + 890, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the patch level, we split the pipeline into three main parts, Fig. 5 (Right). First, we divide the original image into hierarchical patches of varying scales. Then the topology error between predicted patches and corresponding ground truth is calculated. Moreover, a sorted list of patches is built based on these. According to their topology errors, an adaptive screen method is adopted. This allows the calculation of topology-aware loss on selected patches, which drives the model to focus on regions with more topology errors.", + "bbox": [ + 496, + 734, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In more detail, the Betti error is calculated by comparing the difference in the number of disjoint pixel sets of each", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21311", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a3ebc062e8aa456b6df9b73b400a5b05c634be81c91e1a2eeeba8a4372aab33c.jpg", + "image_caption": [ + "Figure 5. The pipeline of adaptive topology-aware selection module (ATS). In the topological Error Calculation and Sort and Adaptive Selection, the darker the square, the larger the Betti error." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 885, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "class between the prediction and GT, Fig. 5 (Bottom right). Note that the Betti error calculation used in the selection at the image and patch levels is the same. The difference is that one uses the whole image and the other uses the patch.", + "bbox": [ + 75, + 339, + 470, + 401 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 414, + 200, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 438, + 284, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. I) MoNuSeg [17] contains 30 H&E-stained pathological images of size $1000 \\times 1000$ from seven organs, including a total of 21,623 nuclei labeled without distinction between various categories. According to [17], we divide 30 images into three sets: 12 for training, 4 for validation, and 14 for testing (combining the same organ and different organ test sets). II) CPM17 [27] contains 64 H&E stained histopathology images with 7,570 annotated nuclear boundaries. It is from the MICCAI 2017 Digital Pathology Challenge [27] and images on two different scales: $500 \\times 500$ and $600 \\times 600$ . This is also a dataset for single-class segmentation task. III) PBNulei is a nuclear segmentation dataset from Peking University Affiliated Hospital, marked by professional pathologists. It contains 106 H&E stained histopathology images of $512 \\times 512$ sizes, with a total of 36,659 nuclei, 78 images for training and 28 for testing. The dataset is divided into five categories, namely carcinoma, epithelial, inflammatory, mesenchymal, and myoepithelium.", + "bbox": [ + 75, + 460, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use three instance-level evaluation metrics to measure the instance segmentation performance of the comparison models, which are: mean Panoptic Quality (PQ), the Aggregated Jaccard Index (AJI), and average Hausdorff distance (Haus).", + "bbox": [ + 75, + 750, + 468, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. The proposed TopoSeg is trained and tested using the open-source software library Pytorch 1.8.0 on NVIDIA Tesla V100 with CUDA 10.1.", + "bbox": [ + 75, + 825, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The MoNuSeg dataset has a relatively large original image size and the CPM17 dataset has an inconsistent original", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bb4dba25e6c96288524d3b93dbb08b6934865142392501d7f171e588a8c3e646.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 339, + 890, + 450 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be6c2c2a2701b4164a4664aa572f25ec156feda52202e78f6e0293d8f41a37fa.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison before and after adding the topology-aware loss. (a) the probability map obtained by the model just using pixel-wise loss; (b) the probability map obtained by the model using pixel-wise loss and topology-aware loss." + ], + "image_footnote": [], + "bbox": [ + 503, + 453, + 890, + 579 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6390d9900a7df4bb5a2b6cdd2d787b5c264b2510c9de07fcd922d67c434de1e8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetBackboneLossPQ ↑AJI ↑Haus ↓
pixeltopo
MoNuSegU-Net✓0.5810.57912.80
U-Net✓✓0.5970.6199.647
CPMU-Net✓0.6250.6667.740
U-Net✓✓0.6620.7185.961
PBNuleiU-Net✓0.4260.3862.770
U-Net✓✓0.4250.4161.953
", + "bbox": [ + 501, + 655, + 901, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of instance segmentation results obtained using only pixel-wise loss and adding topology-aware loss based on pixel-wise loss in three datasets.", + "bbox": [ + 498, + 794, + 890, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "image size, so it is divided into multiple patches, thus allowing the network to be uniform and fully trained. As a result, the MoNuSeg dataset generates 192 training patches, the", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21312", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CPM17 dataset generates 512 training patches. In the training phase, we apply a series of data augmentation methods, which include random flip, random rotation, random color, random crop, and Gaussian blur. The training epoch is set as 300, and the initial learning rate is set as $10^{-3}$ .", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation studies", + "text_level": 1, + "bbox": [ + 76, + 178, + 238, + 193 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted ablation studies using three datasets, MoNuSeg, CPM17, and PBNulei, to validate the effectiveness of the proposed TopoSeg. We selected U-Net as the baseline segmentation method, and we improved the U-Net by adding two output branches: a semantic prediction branch and a three-class prediction branch.", + "bbox": [ + 75, + 200, + 468, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Topology-Aware Loss. When only pixel-wise losses are present, the model's performance on the MoNuSeg dataset is 0.581 PQ, 0.579 AJI, and 12.8 Haus, Table 1. After incorporating topological awareness losses, the model's PQ, AJI, and Haus on the MoNuSeg dataset are 0.597, 0.619, and 9.647, respectively, which have increased by $1.7\\%$ , $3.9\\%$ , and 3.153 compared to the original results.", + "bbox": [ + 75, + 292, + 470, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to the performance obtained by using only pixel-wise loss, the model trained with the addition of pixelwise and topological loss has increased by $3.7\\%$ PQ, $5.2\\%$ AJI, and 1.779 Haus respectively on the CPM17 dataset and by $3.0\\%$ AJI, and 0.817 Haus respectively on the PBNulei dataset compared to the original results.", + "bbox": [ + 75, + 398, + 470, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 6 makes a qualitative comparison between the boundary probability map obtained before and after adding the topology-aware loss and its corresponding persistence barcode. It shows that the corresponding structure obtained after adding topology-aware loss is more persistent. As an example, when $p = 0.5$ , three instances are correctly segmented, but when only pixel-wise loss is used, one instance can not be predicted. This indicates that adding topology-aware loss on the basis of pixel-wise loss helps the model learn object structure information, thus more accurately segmenting nuclear instances.", + "bbox": [ + 75, + 489, + 470, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptive Image Selection. We set the threshold of $25\\%$ , $50\\%$ , $75\\%$ and the adaptive threshold for the comparisons. According to the threshold, select the image with Betti error greater than the threshold as the new training dataset, and ensure that the number of training data obtained by different thresholds is the same.", + "bbox": [ + 75, + 657, + 470, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 2, we compare the performance before and after image selection. The comparison results show that the segmentation results obtained after image selection are better than those without image selection. This proves that image selection promotes the model to focus on learning images with more structural prediction errors and improves the instance segmentation performance of the model.", + "bbox": [ + 75, + 750, + 470, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to training with data selected with a threshold of $25\\%$ or $75\\%$ , training with data selected with a threshold of $50\\%$ yielded more competitive model performance,", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bf304f5f6a203394cdbc555051c544d05cdf3cdf5ffaadded89b453db09d239d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsBackboneSimgPQ↑AJI↑Haus↓
MoNuSegU-Net100%0.5970.6199.647
75%0.6040.6269.358
50%0.6060.6289.328
25%0.6020.6259.452
Adaptive0.6100.6329.113
CPMU-Net100%0.6620.7185.961
75%0.6650.7235.835
50%0.6700.7245.887
25%0.6720.7235.847
Adaptive0.6790.7315.805
PBNuleiU-Net100%0.4250.4161.953
75%0.4270.4192.014
50%0.4190.4142.028
25%0.4290.4182.009
Adaptive0.4460.4351.931
", + "bbox": [ + 501, + 90, + 911, + 340 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5703239aaa07b4d30e3e8c678e937af3937c3c4df7016d5959db0fc8086bf335.jpg", + "table_caption": [ + "Table 2. Comparison results of the models obtained by selecting images based on different thresholds. Here, $100\\%$ represents the use of raw data without any image selection operation. $25\\%$ , $50\\%$ , and $75\\%$ represent selecting images with threshold values greater than $25\\%$ , $50\\%$ , and $75\\%$ , respectively, as new input data. Adaptive represents the proposed adaptive image selection method." + ], + "table_footnote": [], + "table_body": "
DatasetsBackbone\\( S_{patch} \\)PQ↑AJI↑Haus ↓
MoNuSegU-NetWhole0.5970.6199.647
Non-hierc0.6090.62810.063
Hierarchical0.6130.6398.842
Adaptive0.6160.6438.787
CPMU-NetWhole0.6620.7185.961
Non-hierc0.6750.7296.108
Hierarchical0.6620.7335.510
Adaptive0.6780.7415.605
PBNuleiU-NetWhole0.4250.4161.953
Non-hierc0.4980.4792.320
Hierarchical0.5140.4992.011
Adaptive0.5200.5042.185
", + "bbox": [ + 501, + 455, + 890, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation studies on different patch selection methods. $S_{patch}$ denotes the method of dividing the image, \"Whole\" denotes the whole image for calculation, \"Non-hierc\" means non-hierarchical and refers to single size patch division, \"Hierarchical\" refers to multi-scale patch division, and \"Adaptive\" refers to the adaptive approach of dividing the patch.", + "bbox": [ + 498, + 671, + 890, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "indicating that overly high or low thresholds can limit the learning of the model. Therefore, selecting a single threshold for image selection has limitations.", + "bbox": [ + 498, + 776, + 890, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to the three fixed threshold image filtering methods, our method has the ability to flexibly select input data. As shown in Table 2, the adaptive image selection approach achieved superior segmentation performance. In the MoNuSeg dataset, the performance metrics PQ, AJI,", + "bbox": [ + 498, + 824, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21313", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85994c6154f84a21905c204e932f9161284a922d443813a83480e32a9d1d7ecc.jpg", + "image_caption": [ + "Figure 7. Visualization of instance segmentation results with different methods. (a) Input image; (b) Ground Truth; (c) U-Net [25]; (d) Micro-Net [24]; (e) FullNet [23]; (f) Hover-Net [9]; (g) CDNet [11], (h) Our TopoSeg. Different colors indicate different instances in the images. The white rectangles are drawn for a clear comparison." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 890, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and Haus improved by $1.2\\%$ , $1.3\\%$ , and $0.534$ , respectively, compared to the original results. Similarly, compared with the original results, the performance of the CPM17 dataset is improved by $1.7\\%$ PQ, $1.3\\%$ AJI and $0.16$ Haus respectively, and the performance of the PBNulei dataset is improved by $2\\%$ PQ, $1.9\\%$ AJI and $0.022$ Haus, respectively.", + "bbox": [ + 75, + 570, + 468, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This reflects that, compared with the three fixed threshold image selection methods, our adaptive image selection approach has the ability to select input data more flexibly, thus promoting the model to obtain more correct instance segmentation results.", + "bbox": [ + 75, + 667, + 468, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive Hierarchical Patch Selection. As shown in Table 3, we compare the performance of the topology loss calculation after patch selection with that of the whole image. First of all, we verify the effect of patch selection on segmentation performance. Comparing the first row (Whole) and the second row (Non-hierc) in each dataset in Table 3, it is concluded that the segmentation performance of topology loss calculation after selecting patch is better than that of using the whole image on both PQ and AJI. This shows that by selecting the appropriate patch, the model", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pays more attention to the regions that need structural optimization, so as to improve the segmentation performance.", + "bbox": [ + 500, + 571, + 888, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then, we verified the effect of using hierarchical patches with different scales for calculating topological loss in instance segmentation. We compared the performance of the third row (Hierarchical) and the second row (Non-hierc) within each dataset in Table 3 and found that hierarchical patches outperformed single-layer patches. This also shows that selecting the appropriate region that needs structural optimization helps the model better segment instances.", + "bbox": [ + 496, + 628, + 890, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 shows that compared with the topology loss calculated by the whole image, the segmentation performance obtained by adaptive hierarchical selection of patch is significantly improved. In the MoNuSeg dataset, PQ increased by $1.8\\%$ , AJI increased by $2.4\\%$ , Haus increased by 0.86. In the CPM dataset, PQ increased by $1.6\\%$ , AJI increased by $2.3\\%$ , Haus increased by 0.356. In the PBNulei dataset, PQ increased by $9.5\\%$ , AJI increased by $8.8\\%$ .", + "bbox": [ + 496, + 779, + 892, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21314", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7a2e06e1de18ae690c9aba05e468486187bb1583c6a785bf81f2653ed82d1acf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMoNuSegCPM17PBNuclei
PQ↑AJI↑PQ↑AJI↑PQ↑AJI↑
U-Net [25]0.5810.5790.6250.6660.3550.312
Mask-RCNN [13]0.5090.5460.6740.684--
DCAN [2]0.4920.5250.5450.5610.2430.249
DIST [21]0.4430.5590.5040.6160.3250.301
Micro-Net [24]0.5190.560.6610.6680.3060.287
Full-Net [23]-0.604-0.6610.3850.378
PFF-Net [18]0.5870.611----
Hover-Net [9]0.5970.6180.6970.7050.4160.363
CDNet [11]-0.637-0.7330.4130.381
Ours0.6250.6430.7050.7560.5280.509
", + "bbox": [ + 78, + 90, + 470, + 280 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 4. Performance comparisons on three nuclei datasets.", + "bbox": [ + 76, + 285, + 429, + 299 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3. Performance Comparison", + "text_level": 1, + "bbox": [ + 76, + 313, + 312, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In Table 4, the proposed TopoSeg is compared with the state-of-the-art nuclear segmentation model on MoNuSeg and CPM17. Our TopoSeg achieves 0.625 PQ and 0.643 AJI on the MoNuSeg dataset and 0.705 PQ and 0.756 AJI on CPM17 dataset, respectively. The TopoSeg achieves competitive results for nuclei segmentation task.", + "bbox": [ + 76, + 338, + 468, + 429 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We further carry out a qualitative visualization analysis of our method and different existing methods. Compared to all the other methods in Fig. 7, our TopoSeg achieves better instance segmentation results for nuclear instance segmentation, especially for the correct structure segmentation of dense or overlapping regions. Both quantitative and qualitative results demonstrate that our TopoSeg is competitive and able to segment dense nuclei instances.", + "bbox": [ + 75, + 431, + 468, + 551 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 569, + 194, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper proposes a novel topology-aware segmentation approach, termed TopoSeg, for nuclear instance segmentation tasks to address the challenge of touching and overlapping nuclear segmentation. We develop a TAM, which contains a persistence encoding method that accounts for topology information in predicted probability maps, and a topology-aware loss that is flexible and suitable for topology optimization. We further improve it by applying an ATS, which enforces the model to focus on regions with high topology errors. The experimental results in three nuclei datasets show that TopoSeg has achieved superior performance. As a general method of topology awareness, TopoSeg provides fresh insight into the instance segmentation problem of small and dense objects.", + "bbox": [ + 75, + 597, + 468, + 809 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments This work was supported in part by the National Key R&D Program of China (No. 2022ZD0118201), Natural Science Foundation of China (No. 61972217, 32071459, 62176249, 62006133, 62271465).", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 90, + 596, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nick Byrne, James R Clough, Israel Valverde, Giovanni Montana, and Andrew P King. A persistent homology-based topological loss for cnn-based multiclass segmentation of cmr. IEEE Transactions on Medical Imaging, 42(1):3-14, 2022. 2, 3, 4", + "[2] Hao Chen, Xiaojuan Qi, Lequan Yu, Qi Dou, Jing Qin, and Pheng-Ann Heng. Dcan: Deep contour-aware networks for object instance segmentation from histology images. Medical Image Analysis, 36:135–146, 2017. 2, 9", + "[3] Shengcong Chen, Changxing Ding, Minfeng Liu, Jun Cheng, and Dacheng Tao. CPP-net: Context-aware polygon proposal network for nucleus segmentation. IEEE Transactions on Image Processing, 32:980-994, 2023. 1, 2", + "[4] James R Clough, Nicholas Byrne, Ilkay Oksuz, Veronika A Zimmer, Julia A Schnabel, and Andrew P King. A topological loss function for deep-learning based image segmentation using persistent homology. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):8766-8778, 2020. 3", + "[5] James R Clough, Ilkay Oksuz, Nicholas Byrne, Julia A Schnabel, and Andrew P King. Explicit topological priors for deep-learning based image segmentation using persistent homology. In International Conference on Information Processing in Medical Imaging, pages 16-28, 2019. 3", + "[6] Herbert Edelsbrunner, David Letscher, and Afra Zomorodian. Topological persistence and simplification. In Proceedings of the 41st Annual Symposium on Foundations of Computer Science, pages 454-463, 2000. 2", + "[7] Robin Forman. Morse theory for cell complexes. Advances in Mathematics, 134(1):90-145, 1998. 3", + "[8] Robin Forman. A user's guide to discrete Morse theory. Séminaire Lotharingien de Combinatoire [electronic only], 48:1-35, 2002. 3", + "[9] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:1-15, 2019. 1, 2, 8, 9", + "[10] Saumya Gupta, Xiaoling Hu, James Kaan, Michael Jin, Muth shipay Mpoy, Katherine Chung, Gagandeep Singh, Mary Saltz, Tahsin Kurc, Joel Saltz, et al. Learning topological interactions for multi-class medical image segmentation. In Proceedings of the European Conference on Computer Vision, pages 701-718, 2022. 4", + "[11] Hongliang He, Zhongyi Huang, Yao Ding, Guoli Song, Lin Wang, Qian Ren, Pengxu Wei, Zhiqiang Gao, and Jie Chen. Cdnet: Centripetal direction network for nuclear instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4026-4035, 2021. 2, 8, 9", + "[12] Hongliang He, Chi Zhang, Jie Chen, Ruizhe Geng, Luyang Chen, Yongsheng Liang, Yanchang Lu, Jihua Wu, and Yongjie Xu. A hybrid-attention nested unet for nuclear segmentation in histopathological images. Frontiers in Molecular Biosciences, 8:1–9, 2021. 1, 2" + ], + "bbox": [ + 501, + 114, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21315", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2961-2969, 2017. 9", + "[14] Xiaoling Hu, Fuxin Li, Dimitris Samaras, and Chao Chen. Topology-preserving deep image segmentation. Advances in Neural Information Processing Systems, 32:5658-5669, 2019. 2, 3", + "[15] Xiaoling Hu, Dimitris Samaras, and Chao Chen. Learning probabilistic topological representations using discrete Morse theory. arXiv preprint arXiv:2206.01742, 2022. 3", + "[16] Xiaoling Hu, Yusu Wang, Li Fuxin, Dimitris Samaras, and Chao Chen. Topology-aware segmentation using discrete Morse theory. In Proceedings of the International Conference on Learning Representation, pages 1-19, 2021. 2, 3", + "[17] Neeraj Kumar, Ruchika Verma, Sanuj Sharma, Surabhi Bhargava, Abhishek Vahadane, and Amit Sethi. A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Transactions on Medical Imaging, 36(7):1550-1560, 2017. 6", + "[18] Dongnan Liu, Donghao Zhang, Yang Song, Heng Huang, and Weidong Cai. Panoptic feature fusion net: A novel instance segmentation paradigm for biomedical and biological images. IEEE Transactions on Image Processing, 30:2045-2059, 2021. 1, 9", + "[19] Xiaoming Liu, Zhengsheng Guo, Jun Cao, and Jinshan Tang. Mdc-net: a new convolutional neural network for nucleus segmentation in histopathology images with distance maps and contour information. Computers in Biology and Medicine, 135:1-12, 2021. 1, 2", + "[20] Agata Mosinska, Pablo Marquez-Neila, Mateusz Kozinski, and Pascal Fua. Beyond the pixel-wise loss for topology-aware delineation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3136–3145, 2018. 2", + "[21] Peter Naylor, Marick Laé, Fabien Reyal, and Thomas Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE Transactions on Medical Imaging, 38(2):448-459, 2018. 1, 2, 9", + "[22] H. Pinckaers, W. Bulten, J. van der Laak, and G. Litjens. Detection of prostate cancer in whole-slide images through end-to-end training with image-level labels. IEEE Transactions on Medical Imaging, 40(7):1817-1826, 2021. 1", + "[23] Hui Qu, Zhennan Yan, Gregory M Riedlinger, Subhajyoti De, and Dimitris N Metaxas. Improving nuclei/gland instance segmentation in histopathology images by full resolution neural network and spatial constrained loss. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 378-386, 2019. 1, 2, 8, 9", + "[24] Shan E Ahmed Raza, Linda Cheung, Muhammad Shaban, Simon Graham, David Epstein, Stella Pelengaris, Michael Khan, and Nasir M Rajpoot. Micro-net: A unified model for segmentation of various objects in microscopy images. Medical Image Analysis, 52:160-173, 2019. 2, 8, 9", + "[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 1, 2, 8, 9", + "[26] Suprosanna Shit, Johannes C Paetzold, Anjany Sekuboyina, Ivan Ezhov, Alexander Unger, Andrey Zhylka, Josien PW Pluim, Ulrich Bauer, and Bjoern H Menze. cldice-a novel topology-preserving loss function for tubular structure segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16560-16569, 2021. 2", + "[27] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in Bioengineering and Biotechnology, 53:1-15, 2019. 6", + "[28] Yanning Zhou, Omer Fahri Onder, Qi Dou, Efstratios Tsougenis, Hao Chen, and Pheng-Ann Heng. Cia-net: Robust nuclei instance segmentation with contour-aware information aggregation. In International Conference on Information Processing in Medical Imaging, pages 682–693, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 398 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21316", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/a134e8ff-997a-47c5-810d-0149806d46fc_model.json b/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/a134e8ff-997a-47c5-810d-0149806d46fc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cbd744c8ca1734f462b8b5ece914145d5c7ae54a --- /dev/null +++ b/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/a134e8ff-997a-47c5-810d-0149806d46fc_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.192, + 0.131, + 0.78, + 0.154 + ], + "angle": 0, + "content": "TopoSeg: Topology-Aware Nuclear Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.18, + 0.731, + 0.218 + ], + "angle": 0, + "content": "Hongliang He\\(^{1,2*}\\) Jun Wang\\(^{1*}\\) Pengxu Wei\\(^{3}\\) Fan Xu\\(^{4}\\) Xiangyang Ji\\(^{5}\\) Chang Liu\\(^{5}\\) Jie Chen\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.223, + 0.841, + 0.295 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Electronic and Computer Engineering, Peking University, Shenzhen, China \n\\(^{2}\\)School of Computer Science and Technology, Soochow University, Suzhou, China \n\\(^{3}\\)Sun Yat-sen University, Guangzhou, China \n\\(^{4}\\)Peng Cheng Laboratory, Shenzhen, China \n\\(^{5}\\)Department of Automation and BNRist, Tsinghua University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.298, + 0.762, + 0.331 + ], + "angle": 0, + "content": "hehl@pku.edu.cn, junw@stu.pku.edu.cn, weipx3@mail.sysu.edu.cn, {xuf01,chenj}@pcl.ac.cn, {xyji,liuchang2022}@tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.365, + 0.314, + 0.381 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.474, + 0.73 + ], + "angle": 0, + "content": "Nuclear instance segmentation has been critical for pathology image analysis in medical science, e.g., cancer diagnosis. Current methods typically adopt pixel-wise optimization for nuclei boundary exploration, where rich structural information could be lost for subsequent quantitative morphology assessment. To address this issue, we develop a topology-aware segmentation approach, termed TopoSeg, which exploits topological structure information to keep the predictions rational, especially in common situations with densely touching and overlapping nucleus instances. Concretely, TopoSeg builds on a topology-aware module (TAM), which encodes dynamic changes of different topology structures within the three-class probability maps (inside, boundary, and background) of the nuclei to persistence barcodes and makes the topology-aware loss function. To efficiently focus on regions with high topological errors, we propose an adaptive topology-aware selection (ATS) strategy to enhance the topology-aware optimization procedure further. Experiments on three nuclear instance segmentation datasets justify the superiority of TopoSeg, which achieves state-of-the-art performance. The code is available at https://github.com/hhlisme/toposeg." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.759, + 0.21, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.785, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Pathological image analysis is considered the gold standard for cancer diagnosis, treatment, and prevention, where nuclear instance segmentation serves as the core step. It is because the morphology of nuclei including the shape, appearance, and distribution within pathology images provides interpretable features of diagnostic and prognostic" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.367, + 0.889, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.527, + 0.895, + 0.667 + ], + "angle": 0, + "content": "Figure 1. The schematic illustration of pixel-wise loss and topology-aware loss. (Left) A patch in the image and its corresponding boundary ground truth, as well as the original boundary segmentation probability map and its corresponding instance prediction. (Top-right) The gradient map was obtained by pixel-wise loss, the 3D map of the boundary probability map and the optimization trend, as well as the final instance prediction. (Bottom-right) The gradient map obtained from topology-aware loss, the 3D image of the boundary probability map and the optimization trend, as well as the final instance prediction." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.685, + 0.649, + 0.7 + ], + "angle": 0, + "content": "cancer indicators[22]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.893, + 0.824 + ], + "angle": 0, + "content": "Compared to general image object segmentation, nuclear instance segmentation faces with some particular challenges: (i) the crowded nuclei distribution leads to clustered overlapping instances. (ii) The blurry boundaries between touching or overlapping nuclei increase the difficulty of distinguishing individual instances. (iii) Large variability in size, shape, texture and intensity of nuclei between different images requires more robust annotation and prediction [18]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "To cope with these challenges, several studies have focused on exploring innovative network architectures [12, 23, 25] and incorporating distance maps [3, 9, 19, 21] to enhance segmentation performance. However, existing works typically rely on pixel-wise losses, such as CE/Dice loss," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.348, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution ☑ Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21307" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "which enforce predictions to align with GT pixel-by-pixel yet ignore the crucial structure information of instances. Since treat pixels independently, they lack sensitivity towards high-level structural characteristics, such as the connectivity of boundaries, which are essential for the rational segmentation of individual instances. This inability of pixel-wise loss often leads to unreasonable results such as broken boundaries, bringing about failure to distinguish touching or overlapping instances. This consequently results in undesirable effects on the subsequent quantitative morphology analysis. As depicted in Fig. 1 (top-right), pixel-wise loss focuses on misaligned pixels thus reducing pixel errors. However, due to the lack of perception of structural information, the boundaries between overlapping instances with smaller pixel proportions are not given enough consideration and result in incorrect instance predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.636 + ], + "angle": 0, + "content": "To capture structural information of nuclear instances, we propose a novel topology-aware nuclear instance segmentation approach, termed TopoSeg. Specifically, we introduce a topology encoding method to encode topological features of three-class maps which are utilized to separate individual instances. Building upon this, we propose a topology-aware module (TAM), which encodes dynamic changes and persistence of different topology structures within the predicted three-class probability maps to persistence barcodes and develops a topology-aware loss function, integrating topology information into model learning. As shown in Fig. 1 (bottom-right), with topology constraints, the critical boundaries are identified and optimized, resulting in a reasonable segmentation outcome. To further adapt to the task of dense and small instances like nuclei, we introduce an adaptive topology-aware selection module (ATS) to concentrate on images and regions requiring topological optimization. Notably, TopoSeg effectively promotes instance segmentation performance without requiring any changes to the original network structure." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.637, + 0.423, + 0.651 + ], + "angle": 0, + "content": "The contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.665, + 0.47, + 0.725 + ], + "angle": 0, + "content": "- We propose TopoSeg, which provides a new perspective that goes beyond the common pixel-wise optimization paradigm and incorporates meaningful topology constraints for nuclear instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.738, + 0.47, + 0.784 + ], + "angle": 0, + "content": "- We present a TAM to encode the persistence of topology structures in probability maps and introduce a topology-aware loss as a further constraint." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.797, + 0.47, + 0.858 + ], + "angle": 0, + "content": "- We develop an ATS to identify regions that most require topology optimization, enabling the model to efficiently accommodate segmentation tasks involving small and dense objects." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We justify the superiority of our TopoSeg on three nuclei datasets with state-of-the-art performance." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.665, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.643, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.695, + 0.133 + ], + "angle": 0, + "content": "2.1. Nuclei Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.336 + ], + "angle": 0, + "content": "Deep learning has been extensively utilized for nuclear instance segmentation [2, 9, 11, 24]. In 2015, Ronneberger et al. [25] proposed the U-Net model, which has become one of the most fundamental models in medical image segmentation. Raza et al. [24] proposed Micro-Net, which achieves robustness to large internal/external variances in nucleus size by utilizing multi-resolution and weighted loss functions. Qu et al. [23] proposed a full-resolution convolutional neural network (FullNet) with no downsampling operation employed in the network structure to enhance localization accuracy. He et al. [12] proposed a hybrid attention nested U-shaped network (Han-Net) to extract effective feature information from multiple layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.892, + 0.622 + ], + "angle": 0, + "content": "To take advantage of contour information to differentiate contact/overlapping nuclei, Chen et al. [2] initially proposed incorporating contour information into a multi-level FCN to create a deep contour-aware network for nuclear instance segmentation. Subsequently, Zhou et al. [28] proposed the contour-aware information aggregation network to combine spatial and textural features between nuclei and contours. Additionally, some works [3, 9, 19, 21] introduced distance maps to separate contact/overlapping nuclei. Naylor et al. [21] addressed the issue of segmenting touching nuclei by formulating the segmentation task as a regression task of intra-nuclear distance maps. Graham et al. [9] proposed a network (Hover-Net) for simultaneous nucleus segmentation and classification, which uses the vertical and horizontal distances between a nucleus pixel and its center of mass to separate clusters of nuclei. Moreover, He et al. [11] proposed a centripetal directional network (CDNet) for nuclear instance segmentation, combining directional information into the network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.892, + 0.684 + ], + "angle": 0, + "content": "The above models have achieved some performance improvements in nuclear segmentation, but these models using pixel-wise loss tend to ignore structural information of nuclear boundaries." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.719, + 0.712 + ], + "angle": 0, + "content": "2.2. Segmentation Topology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Several topology-based approaches [1, 14, 16, 20, 26] have exploited ways to guarantee the topological accuracy of objects in images. Mosinska et al. [20] considered topological information by constraining the similarity between feature maps and predicted probability maps generated by a pre-trained VGG-19 network. Shit et al. [26] proposed a similarity metric (clDice) to calculate the intersection between segmentation masks and their morphological skeletons, and developed a novel loss function (Soft-clDice) to train the neural network based on the skeleton structure." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In addition, the differentiability of persistent homology [6] allows it to be formulated as a loss function and ap-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21308" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.092, + 0.891, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.257, + 0.891, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.406, + 0.893, + 0.448 + ], + "angle": 0, + "content": "Figure 2. The overall framework of the proposed TopoSeg. The network contains two basic prediction branches: the semantic branch and the three-class (inside, boundary, background) branch. In order to be aware of topology, the framework contains two important modules: the adaptive topology-aware selection module and the topology-aware module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.47, + 0.677 + ], + "angle": 0, + "content": "plied to end-to-end deep neural networks. Hu et al. [14] developed a loss function that directly constrains the prediction results and topological GT based on persistent homology. Furthermore, Hu et al. [16] used Discrete Morse Theory [7, 8] to capture the singularities of the gradient vector field of the likelihood function to find all potential skeletons of the object. They also leveraged Discrete Morse Theory and persistent homology to construct a topological representation space and trained a probabilistic model for performing inference tasks in such a topological representation space [15]. In cardiac magnetic resonance (CMR), Clough et al. [1, 5, 4] investigated various methods to couple topological information, where the number of components in the required segmentation is known a priori." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.47, + 0.77 + ], + "angle": 0, + "content": "However, the above methods have certain limitations in terms of computational efficiency, topology completeness, and task flexibility. Therefore, we propose a topology-aware model for nuclear instance segmentation, which encodes structural information to construct topological losses, thus improving the instance segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.789, + 0.212, + 0.806 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.188, + 0.83 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Fig. 2 provides a flowchart of our TopoSeg, which adopts a dual-branch architecture with two types of predictions: semantic prediction of nuclear semantic categories (e.g. carcinoma, epithelial, etc.) and three-class prediction of struc" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.892, + 0.542 + ], + "angle": 0, + "content": "tural categories (inside, boundary, background). The output of three-class branch is sequentially processed through an adaptive topology-aware selection module (ATS) and a topology-aware module (TAM) to efficiently capture the topology information of nuclear instances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.893, + 0.727 + ], + "angle": 0, + "content": "Specifically, to quantify the topological features of an image, we present topology encoding in Sec. 3.2. Topological information contained in three-class images, such as the number of holes and connected components of the boundary, can reflect the number of instances and overlapping characteristics, which is meaningful for reasonable instance segmentation. Therefore, we decompose the three-class image into three channels (inside, boundary, and union) and encode their topological features. This allows the direct representation of the structural correctness, but its inapplicability to predicated probability maps makes it challenging to incorporate into the training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.892, + 0.821 + ], + "angle": 0, + "content": "To address this issue, we introduce TAM in Sec. 3.3, which extracts topological structures of predicted three-class probability maps into persistence barcode through a persistence encoding method and constrains the topology similarity between GTs and predictions by a topology-aware loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To further improve the topology perception of the model and reduce the computational cost, we develop ATS in Sec. 3.4 to adaptively select regions with greater topology importance in both image- and patch-level, which makes the topological constraints more suitable for the task of seg" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21309" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.09, + 0.168, + 0.1 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.09, + 0.261, + 0.1 + ], + "angle": 0, + "content": "Prediction 1" + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.09, + 0.358, + 0.1 + ], + "angle": 0, + "content": "Prediction 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.09, + 0.464, + 0.1 + ], + "angle": 0, + "content": "Prediction 3" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.103, + 0.177, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.103, + 0.273, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.103, + 0.37, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.103, + 0.465, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.173, + 0.139, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.176, + 0.234, + 0.186 + ], + "angle": 0, + "content": "Topology encoding" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.189, + 0.465, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.259, + 0.47, + 0.302 + ], + "angle": 0, + "content": "Figure 3. (Top) The ground truth and different prediction results of two overlapping nuclei in the pathological image. (Bottom) An example of topology encoding of ground truth." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.321, + 0.313, + 0.335 + ], + "angle": 0, + "content": "menting dense and small instances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.469, + 0.367 + ], + "angle": 0, + "content": "In general, during the training phase, the total loss function of TopoSeg is" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.38, + 0.469, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p i x e l}} + \\lambda \\mathcal {L} _ {\\text {t o p o}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.469, + 0.47 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{pixel}\\) denotes pixel-wise loss with respect to the output of both branches, \\(\\mathcal{L}_{topo}\\) denotes the proposed topology-aware in TAM with respect to the output at three-class branch and \\(\\lambda\\) denotes the weight of topology-aware loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.561 + ], + "angle": 0, + "content": "The design of this total loss leverages the strengths of both types of losses. Initially, the model is trained with only the pixel-wise loss for a certain number of iterations (\\(\\lambda = 0\\)) to obtain accurate pixel-level predictions. Subsequently, the topology loss is introduced (\\(\\lambda > 0\\)) to further enhance the model's ability to capture rational object structure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.47, + 0.621 + ], + "angle": 0, + "content": "In the inference phase, the final predicted instance segmentation map is obtained by utilizing boundaries predicted in the three-class branch to separate the semantic result of the semantic branch." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.633, + 0.262, + 0.65 + ], + "angle": 0, + "content": "3.2. Topology Encoding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.469, + 0.747 + ], + "angle": 0, + "content": "As previously noted, pixel-wise loss is insensitive to the differences in a few pixels at the boundary between overlapping nuclei, as distinct examples depicted in Fig. 3, despite significant structural differences. To quantitatively represent them, we present a topology encoding approach whose general representation is as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.47, + 0.904 + ], + "angle": 0, + "content": "For image with \\(K\\) classes, consider pixels of class \\(i\\) as set \\(S_{i}\\). The zero-dimensional topology structure refers to the connection components formed by the connected pixels in \\(S_{i}\\), and the one-dimensional structure refers to the holes. The topology of \\(S_{i}\\) can be considered as a family of subsets \\(\\{S_{ij}^{d}|d = 0,1;j = 1,2,\\dots,b_{d}\\}\\) formed by topological structures of different dimensions, where \\(d\\) represents topology dimension and \\(b_{d}\\) represents the number of subsets of dimension \\(d\\). Then a subset \\(S_{ij}^{0}\\) represents a connected component and \\(S_{ij}^{1}\\) represents a hole, Fig. 3 (Bottom)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "To quantify the topological characteristics of \\( S_{i} \\), we introduce the Betti number from algebraic topology. For a 2D image, the Betti numbers can be expressed as \\( B = (b_{0}, b_{1}) \\) where \\( b_{0} \\) denotes the number of connected components and \\( b_{1} \\) denotes the number of 2D holes. Given a K-class image, the topology encoding can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.189, + 0.892, + 0.207 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} = \\left\\{B _ {i} ^ {d} | i = 1, 2, \\dots , K; d = 0, 1 \\right\\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.212, + 0.818, + 0.228 + ], + "angle": 0, + "content": "where \\( B_{i}^{d} \\) represents Betti number of subset \\( S_{i}^{d} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.273 + ], + "angle": 0, + "content": "In the nuclear instance segmentation task, the topology encoding of instance maps are mapped to three-class maps with three categories (inside, boundary and their union):" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.28, + 0.892, + 0.33 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} = \\left[ \\begin{array}{l} B _ {i n} \\\\ B _ {b o} \\\\ B _ {u n} \\end{array} \\right] = \\left[ \\begin{array}{c c} N _ {i n} & 0 \\\\ N _ {b o} & N _ {i n} \\\\ N _ {b o} & 0 \\end{array} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.336, + 0.892, + 0.396 + ], + "angle": 0, + "content": "where \\( B_{in}, B_{bo} \\) and \\( B_{un} \\) respectively represent the Betti numbers of the corresponding class. The specific values of \\( B_{i}^{d} \\) are defined based on the number of instances \\( N_{in} \\) and non-connected boundaries \\( N_{bo} \\) in the image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Unlike the fixed topology prior of semantic given in [10] and [1], our topology encoding represents dynamically changing instance information for each image through three channels. As an example, the corresponding topology encoding of images in Fig. 3 (Top) is:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.491, + 0.843, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} (G T) = \\left[ \\begin{array}{c c} 2 & 0 \\\\ 1 & 2 \\\\ 1 & 0 \\end{array} \\right], \\mathbf {B} (P 1) = \\left[ \\begin{array}{c c} 2 & 0 \\\\ 1 & 2 \\\\ 1 & 0 \\end{array} \\right],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.542, + 0.845, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} (P 2) = \\left[ \\begin{array}{c c} 1 & 0 \\\\ 1 & 1 \\\\ 1 & 0 \\end{array} \\right], \\mathbf {B} (P 3) = \\left[ \\begin{array}{c c} 2 & 0 \\\\ 2 & 2 \\\\ 2 & 0 \\end{array} \\right].\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.598, + 0.725, + 0.614 + ], + "angle": 0, + "content": "3.3. Topology-Aware Module" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.702, + 0.637 + ], + "angle": 0, + "content": "3.3.1 Persistence Encoding" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.735 + ], + "angle": 0, + "content": "Consider a probability map of class \\(i\\) as a topology set \\(S_{i}\\) with topology subsets \\(\\{S_{ij}^{d}|d = 0,1;j = 1,2,\\dots,J_{d}\\}\\), where \\(J_{d}\\) represents the total number of subsets of dimension \\(d\\). Here we use persistent homology to extract meaningful topology structures which are robust to noise and small perturbations in the probability map." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The first step of persistent homology is to construct a sequence of combinations \\(\\{C_n\\}\\) of subsets, where each combination is obtained by adding more and more elements from \\(S_{i}\\), which can be obtained by thresholding the probability map at a sequence of decreasing threshold values from 1 to 0. During this process, some subsets disappear with the gradual increase of pixels, while others appear. In Fig. 4, subsets of different dimensions are marked with letters of different colors. The subset \\(a\\) and \\(b\\) both appear in \\(C_1\\), and \\(a\\) survives since then while \\(b\\) disappears in \\(C_2\\) because the added pixels connect \\(b\\) and \\(a\\) into a common subset." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21310" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.091, + 0.449, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.168, + 0.465, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.47, + 0.416 + ], + "angle": 0, + "content": "Figure 4. The binary images obtained by taking different threshold \\(\\mathbf{p}\\) of the boundary probability map and corresponding persistence barcode for topology structures. The red letters indicate connected components \\((d = 0)\\), which is signified by the blue solid bars in the persistence barcode map. The golden letters indicate holes \\((d = 1)\\), which are signified by the blue open bars." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.47, + 0.558 + ], + "angle": 0, + "content": "Then the persistence of \\( S_{ij}^{d} \\) is measured by how long it survives in the sequence of combinations, which can be expressed through the persistence barcode whose length represents the corresponding persistence, Fig. 4. The longer the persistence barcode, the more robust the subset. Thereby, the persistence \\( P_{ij}^{d} \\) of \\( S_{ij}^{d} \\) can be calculated as the difference between the threshold values at which \\( S_{ij}^{d} \\) first appears \\( (p_{birth}(S_{ij}^{d})) \\) and when it disappears \\( (p_{death}(S_{ij}^{d})) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.566, + 0.469, + 0.585 + ], + "angle": 0, + "content": "\\[\nP _ {i j} ^ {d} = p _ {\\text {b i r t h}} \\left(S _ {i j} ^ {d}\\right) - p _ {\\text {d e a t h}} \\left(S _ {i j} ^ {d}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.589, + 0.469, + 0.62 + ], + "angle": 0, + "content": "Given prediction probabilities with K classes of a 2D image, the persistence encoding \\( P \\) is" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.628, + 0.469, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} = \\left\\{P _ {i j} ^ {d} \\mid i = 1, 2, \\dots , K; d = 0, 1; j = 1, 2, \\dots , J _ {d} \\right\\}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.656, + 0.469, + 0.701 + ], + "angle": 0, + "content": "Note that since ground truth can be regarded as a probability map containing only binary values of 0 and 1, the persistence encoding of it is" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.709, + 0.469, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = \\left\\{Y _ {i j} ^ {d} \\mid i = 1, 2, \\dots , K; d = 0, 1; j = 1, 2, \\dots , B _ {i} ^ {d} \\right\\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.738, + 0.469, + 0.785 + ], + "angle": 0, + "content": "where the value of \\( Y_{ij}^{d} \\) is 1. As predicted probability maps typically contain more topological structures than those in ground truth, we assume that \\( J_{d} > B_{i}^{d} \\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.282, + 0.817 + ], + "angle": 0, + "content": "3.3.2 Topology-Aware Loss" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Based on persistence encoding, a better prediction should meet the condition that the subsets of \\( C_n \\) consistent with ground truth have greater persistence. We match the top \\( B_i^d \\) persistence barcode \\( P_{ij}^d \\) with GT and regard the remaining codes as unmatched. Our purpose is to reduce the error" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "between the persistence encoding of prediction and ground truth, so as to make the topology correctness of prediction more robust. Therefore, the difference between them can be calculated in two parts:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.161, + 0.892, + 0.214 + ], + "angle": 0, + "content": "\\[\nM _ {i} = \\sum_ {d \\in \\{0, 1 \\}} \\sum_ {j = 1} ^ {B _ {i} ^ {d}} \\left(Y _ {i j} ^ {d} - P _ {i j} ^ {d}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.212, + 0.783, + 0.256 + ], + "angle": 0, + "content": "\\[\nU _ {i} = \\sum_ {d \\in \\{0, 1 \\}} \\sum_ {j = B _ {i} ^ {d} + 1} ^ {J _ {d}} (P _ {i j} ^ {d}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.268, + 0.891, + 0.312 + ], + "angle": 0, + "content": "where \\( M_{i} \\) measures the length difference between the matched barcode and GT barcode for the \\( i \\)th class, and \\( U_{i} \\) measures the length of the unmatched barcode." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.314, + 0.892, + 0.344 + ], + "angle": 0, + "content": "Given prediction probabilities with \\(K\\) classes of a 2D image, the topology-aware loss \\(\\mathcal{L}_{topo}\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.351, + 0.891, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t o p o} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\left(M _ {i} + U _ {i}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.399, + 0.892, + 0.49 + ], + "angle": 0, + "content": "Minimizing this loss is equivalent to maximizing the persistence of topology features that are consistent with ground truth, and minimizing the persistence of undesired features. This provides a structure perspective to optimize the model, which overcomes the limitation of pixel-wise loss and achieves superior nucleus segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.5, + 0.858, + 0.516 + ], + "angle": 0, + "content": "3.4. Adaptive Topology-Aware Selection (ATS)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.523, + 0.891, + 0.568 + ], + "angle": 0, + "content": "ATS aims at efficiently optimizing images with topology-important regions by providing image- and patch-level selection for loss calculation, Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.734 + ], + "angle": 0, + "content": "At the image level, we calculate the topology error between prediction and GT, then selects and stores images with an adaptive threshold, Fig. 5 (Left). If the topology error exceeds the threshold, it indicates that the overall structure of the image is significantly different from ground truth, then the image is preferentially selected and stored for further learning. On the contrary, if the topology error does not exceed the threshold, it will be dropped out in the next iteration. The adaptive threshold is determined based on the current number of iterations and topology error rate. The more iterations, the higher the adaptive threshold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.87 + ], + "angle": 0, + "content": "At the patch level, we split the pipeline into three main parts, Fig. 5 (Right). First, we divide the original image into hierarchical patches of varying scales. Then the topology error between predicted patches and corresponding ground truth is calculated. Moreover, a sorted list of patches is built based on these. According to their topology errors, an adaptive screen method is adopted. This allows the calculation of topology-aware loss on selected patches, which drives the model to focus on regions with more topology errors." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In more detail, the Betti error is calculated by comparing the difference in the number of disjoint pixel sets of each" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21311" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.887, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.893, + 0.327 + ], + "angle": 0, + "content": "Figure 5. The pipeline of adaptive topology-aware selection module (ATS). In the topological Error Calculation and Sort and Adaptive Selection, the darker the square, the larger the Betti error." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.471, + 0.402 + ], + "angle": 0, + "content": "class between the prediction and GT, Fig. 5 (Bottom right). Note that the Betti error calculation used in the selection at the image and patch levels is the same. The difference is that one uses the whole image and the other uses the patch." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.415, + 0.202, + 0.431 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.439, + 0.285, + 0.456 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Datasets. I) MoNuSeg [17] contains 30 H&E-stained pathological images of size \\(1000 \\times 1000\\) from seven organs, including a total of 21,623 nuclei labeled without distinction between various categories. According to [17], we divide 30 images into three sets: 12 for training, 4 for validation, and 14 for testing (combining the same organ and different organ test sets). II) CPM17 [27] contains 64 H&E stained histopathology images with 7,570 annotated nuclear boundaries. It is from the MICCAI 2017 Digital Pathology Challenge [27] and images on two different scales: \\(500 \\times 500\\) and \\(600 \\times 600\\). This is also a dataset for single-class segmentation task. III) PBNulei is a nuclear segmentation dataset from Peking University Affiliated Hospital, marked by professional pathologists. It contains 106 H&E stained histopathology images of \\(512 \\times 512\\) sizes, with a total of 36,659 nuclei, 78 images for training and 28 for testing. The dataset is divided into five categories, namely carcinoma, epithelial, inflammatory, mesenchymal, and myoepithelium." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Evaluation Metrics. We use three instance-level evaluation metrics to measure the instance segmentation performance of the comparison models, which are: mean Panoptic Quality (PQ), the Aggregated Jaccard Index (AJI), and average Hausdorff distance (Haus)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Implementation details. The proposed TopoSeg is trained and tested using the open-source software library Pytorch 1.8.0 on NVIDIA Tesla V100 with CUDA 10.1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The MoNuSeg dataset has a relatively large original image size and the CPM17 dataset has an inconsistent original" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.34, + 0.892, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.454, + 0.892, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.584, + 0.893, + 0.641 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison before and after adding the topology-aware loss. (a) the probability map obtained by the model just using pixel-wise loss; (b) the probability map obtained by the model using pixel-wise loss and topology-aware loss." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.656, + 0.903, + 0.786 + ], + "angle": 0, + "content": "
DatasetBackboneLossPQ ↑AJI ↑Haus ↓
pixeltopo
MoNuSegU-Net✓0.5810.57912.80
U-Net✓✓0.5970.6199.647
CPMU-Net✓0.6250.6667.740
U-Net✓✓0.6620.7185.961
PBNuleiU-Net✓0.4260.3862.770
U-Net✓✓0.4250.4161.953
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Table 1. Comparison of instance segmentation results obtained using only pixel-wise loss and adding topology-aware loss based on pixel-wise loss in three datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "image size, so it is divided into multiple patches, thus allowing the network to be uniform and fully trained. As a result, the MoNuSeg dataset generates 192 training patches, the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21312" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "CPM17 dataset generates 512 training patches. In the training phase, we apply a series of data augmentation methods, which include random flip, random rotation, random color, random crop, and Gaussian blur. The training epoch is set as 300, and the initial learning rate is set as \\(10^{-3}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.179, + 0.24, + 0.194 + ], + "angle": 0, + "content": "4.2. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.202, + 0.47, + 0.293 + ], + "angle": 0, + "content": "We conducted ablation studies using three datasets, MoNuSeg, CPM17, and PBNulei, to validate the effectiveness of the proposed TopoSeg. We selected U-Net as the baseline segmentation method, and we improved the U-Net by adding two output branches: a semantic prediction branch and a three-class prediction branch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.471, + 0.4 + ], + "angle": 0, + "content": "Topology-Aware Loss. When only pixel-wise losses are present, the model's performance on the MoNuSeg dataset is 0.581 PQ, 0.579 AJI, and 12.8 Haus, Table 1. After incorporating topological awareness losses, the model's PQ, AJI, and Haus on the MoNuSeg dataset are 0.597, 0.619, and 9.647, respectively, which have increased by \\(1.7\\%\\), \\(3.9\\%\\), and 3.153 compared to the original results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.491 + ], + "angle": 0, + "content": "Compared to the performance obtained by using only pixel-wise loss, the model trained with the addition of pixelwise and topological loss has increased by \\(3.7\\%\\) PQ, \\(5.2\\%\\) AJI, and 1.779 Haus respectively on the CPM17 dataset and by \\(3.0\\%\\) AJI, and 0.817 Haus respectively on the PBNulei dataset compared to the original results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.471, + 0.657 + ], + "angle": 0, + "content": "Fig. 6 makes a qualitative comparison between the boundary probability map obtained before and after adding the topology-aware loss and its corresponding persistence barcode. It shows that the corresponding structure obtained after adding topology-aware loss is more persistent. As an example, when \\( p = 0.5 \\), three instances are correctly segmented, but when only pixel-wise loss is used, one instance can not be predicted. This indicates that adding topology-aware loss on the basis of pixel-wise loss helps the model learn object structure information, thus more accurately segmenting nuclear instances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.471, + 0.748 + ], + "angle": 0, + "content": "Adaptive Image Selection. We set the threshold of \\(25\\%\\), \\(50\\%\\), \\(75\\%\\) and the adaptive threshold for the comparisons. According to the threshold, select the image with Betti error greater than the threshold as the new training dataset, and ensure that the number of training data obtained by different thresholds is the same." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.855 + ], + "angle": 0, + "content": "In Table 2, we compare the performance before and after image selection. The comparison results show that the segmentation results obtained after image selection are better than those without image selection. This proves that image selection promotes the model to focus on learning images with more structural prediction errors and improves the instance segmentation performance of the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Compared to training with data selected with a threshold of \\(25\\%\\) or \\(75\\%\\), training with data selected with a threshold of \\(50\\%\\) yielded more competitive model performance," + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.092, + 0.912, + 0.342 + ], + "angle": 0, + "content": "
DatasetsBackboneSimgPQ↑AJI↑Haus↓
MoNuSegU-Net100%0.5970.6199.647
75%0.6040.6269.358
50%0.6060.6289.328
25%0.6020.6259.452
Adaptive0.6100.6329.113
CPMU-Net100%0.6620.7185.961
75%0.6650.7235.835
50%0.6700.7245.887
25%0.6720.7235.847
Adaptive0.6790.7315.805
PBNuleiU-Net100%0.4250.4161.953
75%0.4270.4192.014
50%0.4190.4142.028
25%0.4290.4182.009
Adaptive0.4460.4351.931
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.352, + 0.895, + 0.449 + ], + "angle": 0, + "content": "Table 2. Comparison results of the models obtained by selecting images based on different thresholds. Here, \\(100\\%\\) represents the use of raw data without any image selection operation. \\(25\\%\\), \\(50\\%\\), and \\(75\\%\\) represent selecting images with threshold values greater than \\(25\\%\\), \\(50\\%\\), and \\(75\\%\\), respectively, as new input data. Adaptive represents the proposed adaptive image selection method." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.457, + 0.892, + 0.661 + ], + "angle": 0, + "content": "
DatasetsBackbone\\( S_{patch} \\)PQ↑AJI↑Haus ↓
MoNuSegU-NetWhole0.5970.6199.647
Non-hierc0.6090.62810.063
Hierarchical0.6130.6398.842
Adaptive0.6160.6438.787
CPMU-NetWhole0.6620.7185.961
Non-hierc0.6750.7296.108
Hierarchical0.6620.7335.510
Adaptive0.6780.7415.605
PBNuleiU-NetWhole0.4250.4161.953
Non-hierc0.4980.4792.320
Hierarchical0.5140.4992.011
Adaptive0.5200.5042.185
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.672, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Table 3. Ablation studies on different patch selection methods. \\(S_{patch}\\) denotes the method of dividing the image, \"Whole\" denotes the whole image for calculation, \"Non-hierc\" means non-hierarchical and refers to single size patch division, \"Hierarchical\" refers to multi-scale patch division, and \"Adaptive\" refers to the adaptive approach of dividing the patch." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.777, + 0.892, + 0.824 + ], + "angle": 0, + "content": "indicating that overly high or low thresholds can limit the learning of the model. Therefore, selecting a single threshold for image selection has limitations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Compared to the three fixed threshold image filtering methods, our method has the ability to flexibly select input data. As shown in Table 2, the adaptive image selection approach achieved superior segmentation performance. In the MoNuSeg dataset, the performance metrics PQ, AJI," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21313" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.891, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.516, + 0.893, + 0.558 + ], + "angle": 0, + "content": "Figure 7. Visualization of instance segmentation results with different methods. (a) Input image; (b) Ground Truth; (c) U-Net [25]; (d) Micro-Net [24]; (e) FullNet [23]; (f) Hover-Net [9]; (g) CDNet [11], (h) Our TopoSeg. Different colors indicate different instances in the images. The white rectangles are drawn for a clear comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.571, + 0.47, + 0.663 + ], + "angle": 0, + "content": "and Haus improved by \\(1.2\\%\\), \\(1.3\\%\\), and \\(0.534\\), respectively, compared to the original results. Similarly, compared with the original results, the performance of the CPM17 dataset is improved by \\(1.7\\%\\) PQ, \\(1.3\\%\\) AJI and \\(0.16\\) Haus respectively, and the performance of the PBNulei dataset is improved by \\(2\\%\\) PQ, \\(1.9\\%\\) AJI and \\(0.022\\) Haus, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.47, + 0.744 + ], + "angle": 0, + "content": "This reflects that, compared with the three fixed threshold image selection methods, our adaptive image selection approach has the ability to select input data more flexibly, thus promoting the model to obtain more correct instance segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Adaptive Hierarchical Patch Selection. As shown in Table 3, we compare the performance of the topology loss calculation after patch selection with that of the whole image. First of all, we verify the effect of patch selection on segmentation performance. Comparing the first row (Whole) and the second row (Non-hierc) in each dataset in Table 3, it is concluded that the segmentation performance of topology loss calculation after selecting patch is better than that of using the whole image on both PQ and AJI. This shows that by selecting the appropriate patch, the model" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.572, + 0.89, + 0.603 + ], + "angle": 0, + "content": "pays more attention to the regions that need structural optimization, so as to improve the segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.753 + ], + "angle": 0, + "content": "Then, we verified the effect of using hierarchical patches with different scales for calculating topological loss in instance segmentation. We compared the performance of the third row (Hierarchical) and the second row (Non-hierc) within each dataset in Table 3 and found that hierarchical patches outperformed single-layer patches. This also shows that selecting the appropriate region that needs structural optimization helps the model better segment instances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Table 3 shows that compared with the topology loss calculated by the whole image, the segmentation performance obtained by adaptive hierarchical selection of patch is significantly improved. In the MoNuSeg dataset, PQ increased by \\(1.8\\%\\), AJI increased by \\(2.4\\%\\), Haus increased by 0.86. In the CPM dataset, PQ increased by \\(1.6\\%\\), AJI increased by \\(2.3\\%\\), Haus increased by 0.356. In the PBNulei dataset, PQ increased by \\(9.5\\%\\), AJI increased by \\(8.8\\%\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21314" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.281 + ], + "angle": 0, + "content": "
MethodMoNuSegCPM17PBNuclei
PQ↑AJI↑PQ↑AJI↑PQ↑AJI↑
U-Net [25]0.5810.5790.6250.6660.3550.312
Mask-RCNN [13]0.5090.5460.6740.684--
DCAN [2]0.4920.5250.5450.5610.2430.249
DIST [21]0.4430.5590.5040.6160.3250.301
Micro-Net [24]0.5190.560.6610.6680.3060.287
Full-Net [23]-0.604-0.6610.3850.378
PFF-Net [18]0.5870.611----
Hover-Net [9]0.5970.6180.6970.7050.4160.363
CDNet [11]-0.637-0.7330.4130.381
Ours0.6250.6430.7050.7560.5280.509
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.286, + 0.431, + 0.3 + ], + "angle": 0, + "content": "Table 4. Performance comparisons on three nuclei datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.314, + 0.313, + 0.33 + ], + "angle": 0, + "content": "4.3. Performance Comparison" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.339, + 0.469, + 0.43 + ], + "angle": 0, + "content": "In Table 4, the proposed TopoSeg is compared with the state-of-the-art nuclear segmentation model on MoNuSeg and CPM17. Our TopoSeg achieves 0.625 PQ and 0.643 AJI on the MoNuSeg dataset and 0.705 PQ and 0.756 AJI on CPM17 dataset, respectively. The TopoSeg achieves competitive results for nuclei segmentation task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.47, + 0.552 + ], + "angle": 0, + "content": "We further carry out a qualitative visualization analysis of our method and different existing methods. Compared to all the other methods in Fig. 7, our TopoSeg achieves better instance segmentation results for nuclear instance segmentation, especially for the correct structure segmentation of dense or overlapping regions. Both quantitative and qualitative results demonstrate that our TopoSeg is competitive and able to segment dense nuclei instances." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.57, + 0.196, + 0.587 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.81 + ], + "angle": 0, + "content": "This paper proposes a novel topology-aware segmentation approach, termed TopoSeg, for nuclear instance segmentation tasks to address the challenge of touching and overlapping nuclear segmentation. We develop a TAM, which contains a persistence encoding method that accounts for topology information in predicted probability maps, and a topology-aware loss that is flexible and suitable for topology optimization. We further improve it by applying an ATS, which enforces the model to focus on regions with high topology errors. The experimental results in three nuclei datasets show that TopoSeg has achieved superior performance. As a general method of topology awareness, TopoSeg provides fresh insight into the instance segmentation problem of small and dense objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments This work was supported in part by the National Key R&D Program of China (No. 2022ZD0118201), Natural Science Foundation of China (No. 61972217, 32071459, 62176249, 62006133, 62271465)." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.091, + 0.597, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.116, + 0.892, + 0.184 + ], + "angle": 0, + "content": "[1] Nick Byrne, James R Clough, Israel Valverde, Giovanni Montana, and Andrew P King. A persistent homology-based topological loss for cnn-based multiclass segmentation of cmr. IEEE Transactions on Medical Imaging, 42(1):3-14, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.187, + 0.894, + 0.243 + ], + "angle": 0, + "content": "[2] Hao Chen, Xiaojuan Qi, Lequan Yu, Qi Dou, Jing Qin, and Pheng-Ann Heng. Dcan: Deep contour-aware networks for object instance segmentation from histology images. Medical Image Analysis, 36:135–146, 2017. 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.246, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[3] Shengcong Chen, Changxing Ding, Minfeng Liu, Jun Cheng, and Dacheng Tao. CPP-net: Context-aware polygon proposal network for nucleus segmentation. IEEE Transactions on Image Processing, 32:980-994, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.302, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[4] James R Clough, Nicholas Byrne, Ilkay Oksuz, Veronika A Zimmer, Julia A Schnabel, and Andrew P King. A topological loss function for deep-learning based image segmentation using persistent homology. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):8766-8778, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.387, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[5] James R Clough, Ilkay Oksuz, Nicholas Byrne, Julia A Schnabel, and Andrew P King. Explicit topological priors for deep-learning based image segmentation using persistent homology. In International Conference on Information Processing in Medical Imaging, pages 16-28, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.459, + 0.894, + 0.514 + ], + "angle": 0, + "content": "[6] Herbert Edelsbrunner, David Letscher, and Afra Zomorodian. Topological persistence and simplification. In Proceedings of the 41st Annual Symposium on Foundations of Computer Science, pages 454-463, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.517, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[7] Robin Forman. Morse theory for cell complexes. Advances in Mathematics, 134(1):90-145, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.546, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[8] Robin Forman. A user's guide to discrete Morse theory. Séminaire Lotharingien de Combinatoire [electronic only], 48:1-35, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.59, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[9] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:1-15, 2019. 1, 2, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.661, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[10] Saumya Gupta, Xiaoling Hu, James Kaan, Michael Jin, Muth shipay Mpoy, Katherine Chung, Gagandeep Singh, Mary Saltz, Tahsin Kurc, Joel Saltz, et al. Learning topological interactions for multi-class medical image segmentation. In Proceedings of the European Conference on Computer Vision, pages 701-718, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.746, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[11] Hongliang He, Zhongyi Huang, Yao Ding, Guoli Song, Lin Wang, Qian Ren, Pengxu Wei, Zhiqiang Gao, and Jie Chen. Cdnet: Centripetal direction network for nuclear instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4026-4035, 2021. 2, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.831, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[12] Hongliang He, Chi Zhang, Jie Chen, Ruizhe Geng, Luyang Chen, Yongsheng Liang, Yanchang Lu, Jihua Wu, and Yongjie Xu. A hybrid-attention nested unet for nuclear segmentation in histopathological images. Frontiers in Molecular Biosciences, 8:1–9, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.116, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21315" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[13] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2961-2969, 2017. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.148, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[14] Xiaoling Hu, Fuxin Li, Dimitris Samaras, and Chao Chen. Topology-preserving deep image segmentation. Advances in Neural Information Processing Systems, 32:5658-5669, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[15] Xiaoling Hu, Dimitris Samaras, and Chao Chen. Learning probabilistic topological representations using discrete Morse theory. arXiv preprint arXiv:2206.01742, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.248, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[16] Xiaoling Hu, Yusu Wang, Li Fuxin, Dimitris Samaras, and Chao Chen. Topology-aware segmentation using discrete Morse theory. In Proceedings of the International Conference on Learning Representation, pages 1-19, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.303, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[17] Neeraj Kumar, Ruchika Verma, Sanuj Sharma, Surabhi Bhargava, Abhishek Vahadane, and Amit Sethi. A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Transactions on Medical Imaging, 36(7):1550-1560, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[18] Dongnan Liu, Donghao Zhang, Yang Song, Heng Huang, and Weidong Cai. Panoptic feature fusion net: A novel instance segmentation paradigm for biomedical and biological images. IEEE Transactions on Image Processing, 30:2045-2059, 2021. 1, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[19] Xiaoming Liu, Zhengsheng Guo, Jun Cao, and Jinshan Tang. Mdc-net: a new convolutional neural network for nucleus segmentation in histopathology images with distance maps and contour information. Computers in Biology and Medicine, 135:1-12, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[20] Agata Mosinska, Pablo Marquez-Neila, Mateusz Kozinski, and Pascal Fua. Beyond the pixel-wise loss for topology-aware delineation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3136–3145, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.636 + ], + "angle": 0, + "content": "[21] Peter Naylor, Marick Laé, Fabien Reyal, and Thomas Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE Transactions on Medical Imaging, 38(2):448-459, 2018. 1, 2, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[22] H. Pinckaers, W. Bulten, J. van der Laak, and G. Litjens. Detection of prostate cancer in whole-slide images through end-to-end training with image-level labels. IEEE Transactions on Medical Imaging, 40(7):1817-1826, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[23] Hui Qu, Zhennan Yan, Gregory M Riedlinger, Subhajyoti De, and Dimitris N Metaxas. Improving nuclei/gland instance segmentation in histopathology images by full resolution neural network and spatial constrained loss. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 378-386, 2019. 1, 2, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[24] Shan E Ahmed Raza, Linda Cheung, Muhammad Shaban, Simon Graham, David Epstein, Stella Pelengaris, Michael Khan, and Nasir M Rajpoot. Micro-net: A unified model for segmentation of various objects in microscopy images. Medical Image Analysis, 52:160-173, 2019. 2, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 1, 2, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[26] Suprosanna Shit, Johannes C Paetzold, Anjany Sekuboyina, Ivan Ezhov, Alexander Unger, Andrey Zhylka, Josien PW Pluim, Ulrich Bauer, and Bjoern H Menze. cldice-a novel topology-preserving loss function for tubular structure segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16560-16569, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[27] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in Bioengineering and Biotechnology, 53:1-15, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.318, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[28] Yanning Zhou, Omer Fahri Onder, Qi Dou, Efstratios Tsougenis, Hao Chen, and Pheng-Ann Heng. Cia-net: Robust nuclei instance segmentation with contour-aware information aggregation. In International Conference on Information Processing in Medical Imaging, pages 682–693, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21316" + } + ] +] \ No newline at end of file diff --git a/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/a134e8ff-997a-47c5-810d-0149806d46fc_origin.pdf b/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/a134e8ff-997a-47c5-810d-0149806d46fc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e1e8a83a86cc274c59afed713c0ec7614fd29119 --- /dev/null +++ b/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/a134e8ff-997a-47c5-810d-0149806d46fc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bcaf7668497bfc338b616fbab986202ecf9e39374566f47f296ad5de0bf13d7 +size 3719101 diff --git a/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/full.md b/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..413a88b15b0d8746b1b34fce0e3f4d73427a81ae --- /dev/null +++ b/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/full.md @@ -0,0 +1,332 @@ +# TopoSeg: Topology-Aware Nuclear Instance Segmentation + +Hongliang He $^{1,2*}$ Jun Wang $^{1*}$ Pengxu Wei $^{3}$ Fan Xu $^{4}$ Xiangyang Ji $^{5}$ Chang Liu $^{5}$ Jie Chen $^{1,4}$ + +$^{1}$ School of Electronic and Computer Engineering, Peking University, Shenzhen, China + $^{2}$ School of Computer Science and Technology, Soochow University, Suzhou, China + $^{3}$ Sun Yat-sen University, Guangzhou, China + $^{4}$ Peng Cheng Laboratory, Shenzhen, China + $^{5}$ Department of Automation and BNRist, Tsinghua University, Beijing, China + +hehl@pku.edu.cn, junw@stu.pku.edu.cn, weipx3@mail.sysu.edu.cn, {xuf01,chenj}@pcl.ac.cn, {xyji,liuchang2022}@tsinghua.edu.cn + +# Abstract + +Nuclear instance segmentation has been critical for pathology image analysis in medical science, e.g., cancer diagnosis. Current methods typically adopt pixel-wise optimization for nuclei boundary exploration, where rich structural information could be lost for subsequent quantitative morphology assessment. To address this issue, we develop a topology-aware segmentation approach, termed TopoSeg, which exploits topological structure information to keep the predictions rational, especially in common situations with densely touching and overlapping nucleus instances. Concretely, TopoSeg builds on a topology-aware module (TAM), which encodes dynamic changes of different topology structures within the three-class probability maps (inside, boundary, and background) of the nuclei to persistence barcodes and makes the topology-aware loss function. To efficiently focus on regions with high topological errors, we propose an adaptive topology-aware selection (ATS) strategy to enhance the topology-aware optimization procedure further. Experiments on three nuclear instance segmentation datasets justify the superiority of TopoSeg, which achieves state-of-the-art performance. The code is available at https://github.com/hhlisme/toposeg. + +# 1. Introduction + +Pathological image analysis is considered the gold standard for cancer diagnosis, treatment, and prevention, where nuclear instance segmentation serves as the core step. It is because the morphology of nuclei including the shape, appearance, and distribution within pathology images provides interpretable features of diagnostic and prognostic + +![](images/ad304ca217be5ad48b46bb88830e5649fecc03d361e7a3b459c790b29d378d6b.jpg) +Figure 1. The schematic illustration of pixel-wise loss and topology-aware loss. (Left) A patch in the image and its corresponding boundary ground truth, as well as the original boundary segmentation probability map and its corresponding instance prediction. (Top-right) The gradient map was obtained by pixel-wise loss, the 3D map of the boundary probability map and the optimization trend, as well as the final instance prediction. (Bottom-right) The gradient map obtained from topology-aware loss, the 3D image of the boundary probability map and the optimization trend, as well as the final instance prediction. + +cancer indicators[22]. + +Compared to general image object segmentation, nuclear instance segmentation faces with some particular challenges: (i) the crowded nuclei distribution leads to clustered overlapping instances. (ii) The blurry boundaries between touching or overlapping nuclei increase the difficulty of distinguishing individual instances. (iii) Large variability in size, shape, texture and intensity of nuclei between different images requires more robust annotation and prediction [18]. + +To cope with these challenges, several studies have focused on exploring innovative network architectures [12, 23, 25] and incorporating distance maps [3, 9, 19, 21] to enhance segmentation performance. However, existing works typically rely on pixel-wise losses, such as CE/Dice loss, + +which enforce predictions to align with GT pixel-by-pixel yet ignore the crucial structure information of instances. Since treat pixels independently, they lack sensitivity towards high-level structural characteristics, such as the connectivity of boundaries, which are essential for the rational segmentation of individual instances. This inability of pixel-wise loss often leads to unreasonable results such as broken boundaries, bringing about failure to distinguish touching or overlapping instances. This consequently results in undesirable effects on the subsequent quantitative morphology analysis. As depicted in Fig. 1 (top-right), pixel-wise loss focuses on misaligned pixels thus reducing pixel errors. However, due to the lack of perception of structural information, the boundaries between overlapping instances with smaller pixel proportions are not given enough consideration and result in incorrect instance predictions. + +To capture structural information of nuclear instances, we propose a novel topology-aware nuclear instance segmentation approach, termed TopoSeg. Specifically, we introduce a topology encoding method to encode topological features of three-class maps which are utilized to separate individual instances. Building upon this, we propose a topology-aware module (TAM), which encodes dynamic changes and persistence of different topology structures within the predicted three-class probability maps to persistence barcodes and develops a topology-aware loss function, integrating topology information into model learning. As shown in Fig. 1 (bottom-right), with topology constraints, the critical boundaries are identified and optimized, resulting in a reasonable segmentation outcome. To further adapt to the task of dense and small instances like nuclei, we introduce an adaptive topology-aware selection module (ATS) to concentrate on images and regions requiring topological optimization. Notably, TopoSeg effectively promotes instance segmentation performance without requiring any changes to the original network structure. + +The contributions can be summarized as follows: + +- We propose TopoSeg, which provides a new perspective that goes beyond the common pixel-wise optimization paradigm and incorporates meaningful topology constraints for nuclear instance segmentation. +- We present a TAM to encode the persistence of topology structures in probability maps and introduce a topology-aware loss as a further constraint. +- We develop an ATS to identify regions that most require topology optimization, enabling the model to efficiently accommodate segmentation tasks involving small and dense objects. +- We justify the superiority of our TopoSeg on three nuclei datasets with state-of-the-art performance. + +# 2. Related Work + +# 2.1. Nuclei Segmentation + +Deep learning has been extensively utilized for nuclear instance segmentation [2, 9, 11, 24]. In 2015, Ronneberger et al. [25] proposed the U-Net model, which has become one of the most fundamental models in medical image segmentation. Raza et al. [24] proposed Micro-Net, which achieves robustness to large internal/external variances in nucleus size by utilizing multi-resolution and weighted loss functions. Qu et al. [23] proposed a full-resolution convolutional neural network (FullNet) with no downsampling operation employed in the network structure to enhance localization accuracy. He et al. [12] proposed a hybrid attention nested U-shaped network (Han-Net) to extract effective feature information from multiple layers. + +To take advantage of contour information to differentiate contact/overlapping nuclei, Chen et al. [2] initially proposed incorporating contour information into a multi-level FCN to create a deep contour-aware network for nuclear instance segmentation. Subsequently, Zhou et al. [28] proposed the contour-aware information aggregation network to combine spatial and textural features between nuclei and contours. Additionally, some works [3, 9, 19, 21] introduced distance maps to separate contact/overlapping nuclei. Naylor et al. [21] addressed the issue of segmenting touching nuclei by formulating the segmentation task as a regression task of intra-nuclear distance maps. Graham et al. [9] proposed a network (Hover-Net) for simultaneous nucleus segmentation and classification, which uses the vertical and horizontal distances between a nucleus pixel and its center of mass to separate clusters of nuclei. Moreover, He et al. [11] proposed a centripetal directional network (CDNet) for nuclear instance segmentation, combining directional information into the network. + +The above models have achieved some performance improvements in nuclear segmentation, but these models using pixel-wise loss tend to ignore structural information of nuclear boundaries. + +# 2.2. Segmentation Topology + +Several topology-based approaches [1, 14, 16, 20, 26] have exploited ways to guarantee the topological accuracy of objects in images. Mosinska et al. [20] considered topological information by constraining the similarity between feature maps and predicted probability maps generated by a pre-trained VGG-19 network. Shit et al. [26] proposed a similarity metric (clDice) to calculate the intersection between segmentation masks and their morphological skeletons, and developed a novel loss function (Soft-clDice) to train the neural network based on the skeleton structure. + +In addition, the differentiability of persistent homology [6] allows it to be formulated as a loss function and ap- + +![](images/aad59e3199bb222ed617aacb0807f12618ee40d2a222fdc8b2a864451edc0441.jpg) + +![](images/af60d89dded5e5049dfa0c31a65e1eb0870b30661e568c58986d22d928aa184f.jpg) +Figure 2. The overall framework of the proposed TopoSeg. The network contains two basic prediction branches: the semantic branch and the three-class (inside, boundary, background) branch. In order to be aware of topology, the framework contains two important modules: the adaptive topology-aware selection module and the topology-aware module. + +plied to end-to-end deep neural networks. Hu et al. [14] developed a loss function that directly constrains the prediction results and topological GT based on persistent homology. Furthermore, Hu et al. [16] used Discrete Morse Theory [7, 8] to capture the singularities of the gradient vector field of the likelihood function to find all potential skeletons of the object. They also leveraged Discrete Morse Theory and persistent homology to construct a topological representation space and trained a probabilistic model for performing inference tasks in such a topological representation space [15]. In cardiac magnetic resonance (CMR), Clough et al. [1, 5, 4] investigated various methods to couple topological information, where the number of components in the required segmentation is known a priori. + +However, the above methods have certain limitations in terms of computational efficiency, topology completeness, and task flexibility. Therefore, we propose a topology-aware model for nuclear instance segmentation, which encodes structural information to construct topological losses, thus improving the instance segmentation performance. + +# 3. Methodology + +# 3.1. Overview + +Fig. 2 provides a flowchart of our TopoSeg, which adopts a dual-branch architecture with two types of predictions: semantic prediction of nuclear semantic categories (e.g. carcinoma, epithelial, etc.) and three-class prediction of struc + +tural categories (inside, boundary, background). The output of three-class branch is sequentially processed through an adaptive topology-aware selection module (ATS) and a topology-aware module (TAM) to efficiently capture the topology information of nuclear instances. + +Specifically, to quantify the topological features of an image, we present topology encoding in Sec. 3.2. Topological information contained in three-class images, such as the number of holes and connected components of the boundary, can reflect the number of instances and overlapping characteristics, which is meaningful for reasonable instance segmentation. Therefore, we decompose the three-class image into three channels (inside, boundary, and union) and encode their topological features. This allows the direct representation of the structural correctness, but its inapplicability to predicated probability maps makes it challenging to incorporate into the training process. + +To address this issue, we introduce TAM in Sec. 3.3, which extracts topological structures of predicted three-class probability maps into persistence barcode through a persistence encoding method and constrains the topology similarity between GTs and predictions by a topology-aware loss. + +To further improve the topology perception of the model and reduce the computational cost, we develop ATS in Sec. 3.4 to adaptively select regions with greater topology importance in both image- and patch-level, which makes the topological constraints more suitable for the task of seg + +![](images/b323df670db911b86f3dc58dfd9a1b544cabf8465bf8ba3f980f42a86d55f9c9.jpg) +Ground Truth + +![](images/dc29468e81133cb7c07a101e05a81dc3f2af7662f7cb1d36f0f96948ba72cce7.jpg) +Prediction 1 + +![](images/9f08e60781f98b4bc5b8677527b0b7ed0c4667400c941332451b8df16a77ae54.jpg) +Prediction 2 + +![](images/845bfed57d0f6fb281172a474f4cf172bea3dd5f377bad0f4220dc9c423e16fb.jpg) +Prediction 3 + +![](images/d9b6f36fbaef075ffbe9068d710dd3347f1cb54adeea3a8e0c77efd15d6c548a.jpg) +Topology encoding + +![](images/77a5769a0794d305397b1d43a797459dbf56ca01cea8c4e5319e0bdc8cf65d14.jpg) +Figure 3. (Top) The ground truth and different prediction results of two overlapping nuclei in the pathological image. (Bottom) An example of topology encoding of ground truth. + +menting dense and small instances. + +In general, during the training phase, the total loss function of TopoSeg is + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {p i x e l}} + \lambda \mathcal {L} _ {\text {t o p o}}, \tag {1} +$$ + +where $\mathcal{L}_{pixel}$ denotes pixel-wise loss with respect to the output of both branches, $\mathcal{L}_{topo}$ denotes the proposed topology-aware in TAM with respect to the output at three-class branch and $\lambda$ denotes the weight of topology-aware loss. + +The design of this total loss leverages the strengths of both types of losses. Initially, the model is trained with only the pixel-wise loss for a certain number of iterations ( $\lambda = 0$ ) to obtain accurate pixel-level predictions. Subsequently, the topology loss is introduced ( $\lambda > 0$ ) to further enhance the model's ability to capture rational object structure. + +In the inference phase, the final predicted instance segmentation map is obtained by utilizing boundaries predicted in the three-class branch to separate the semantic result of the semantic branch. + +# 3.2. Topology Encoding + +As previously noted, pixel-wise loss is insensitive to the differences in a few pixels at the boundary between overlapping nuclei, as distinct examples depicted in Fig. 3, despite significant structural differences. To quantitatively represent them, we present a topology encoding approach whose general representation is as follows. + +For image with $K$ classes, consider pixels of class $i$ as set $S_{i}$ . The zero-dimensional topology structure refers to the connection components formed by the connected pixels in $S_{i}$ , and the one-dimensional structure refers to the holes. The topology of $S_{i}$ can be considered as a family of subsets $\{S_{ij}^{d}|d = 0,1;j = 1,2,\dots,b_{d}\}$ formed by topological structures of different dimensions, where $d$ represents topology dimension and $b_{d}$ represents the number of subsets of dimension $d$ . Then a subset $S_{ij}^{0}$ represents a connected component and $S_{ij}^{1}$ represents a hole, Fig. 3 (Bottom). + +To quantify the topological characteristics of $S_{i}$ , we introduce the Betti number from algebraic topology. For a 2D image, the Betti numbers can be expressed as $B = (b_{0}, b_{1})$ where $b_{0}$ denotes the number of connected components and $b_{1}$ denotes the number of 2D holes. Given a K-class image, the topology encoding can be formulated as + +$$ +\mathbf {B} = \left\{B _ {i} ^ {d} | i = 1, 2, \dots , K; d = 0, 1 \right\}, \tag {2} +$$ + +where $B_{i}^{d}$ represents Betti number of subset $S_{i}^{d}$ . + +In the nuclear instance segmentation task, the topology encoding of instance maps are mapped to three-class maps with three categories (inside, boundary and their union): + +$$ +\mathbf {B} = \left[ \begin{array}{l} B _ {i n} \\ B _ {b o} \\ B _ {u n} \end{array} \right] = \left[ \begin{array}{c c} N _ {i n} & 0 \\ N _ {b o} & N _ {i n} \\ N _ {b o} & 0 \end{array} \right], \tag {3} +$$ + +where $B_{in}, B_{bo}$ and $B_{un}$ respectively represent the Betti numbers of the corresponding class. The specific values of $B_{i}^{d}$ are defined based on the number of instances $N_{in}$ and non-connected boundaries $N_{bo}$ in the image. + +Unlike the fixed topology prior of semantic given in [10] and [1], our topology encoding represents dynamically changing instance information for each image through three channels. As an example, the corresponding topology encoding of images in Fig. 3 (Top) is: + +$$ +\mathbf {B} (G T) = \left[ \begin{array}{c c} 2 & 0 \\ 1 & 2 \\ 1 & 0 \end{array} \right], \mathbf {B} (P 1) = \left[ \begin{array}{c c} 2 & 0 \\ 1 & 2 \\ 1 & 0 \end{array} \right], +$$ + +$$ +\mathbf {B} (P 2) = \left[ \begin{array}{c c} 1 & 0 \\ 1 & 1 \\ 1 & 0 \end{array} \right], \mathbf {B} (P 3) = \left[ \begin{array}{c c} 2 & 0 \\ 2 & 2 \\ 2 & 0 \end{array} \right]. +$$ + +# 3.3. Topology-Aware Module + +# 3.3.1 Persistence Encoding + +Consider a probability map of class $i$ as a topology set $S_{i}$ with topology subsets $\{S_{ij}^{d}|d = 0,1;j = 1,2,\dots,J_{d}\}$ , where $J_{d}$ represents the total number of subsets of dimension $d$ . Here we use persistent homology to extract meaningful topology structures which are robust to noise and small perturbations in the probability map. + +The first step of persistent homology is to construct a sequence of combinations $\{C_n\}$ of subsets, where each combination is obtained by adding more and more elements from $S_{i}$ , which can be obtained by thresholding the probability map at a sequence of decreasing threshold values from 1 to 0. During this process, some subsets disappear with the gradual increase of pixels, while others appear. In Fig. 4, subsets of different dimensions are marked with letters of different colors. The subset $a$ and $b$ both appear in $C_1$ , and $a$ survives since then while $b$ disappears in $C_2$ because the added pixels connect $b$ and $a$ into a common subset. + +![](images/411528c836fcc76276cab4b88943bd4b0242fc91875dfdbcdba76c3f7d1fae3c.jpg) + +![](images/0df07790ea3c4adf4b72eaa9ebe438d5122c16940531b0167a10f733ac7c8256.jpg) +Figure 4. The binary images obtained by taking different threshold $\mathbf{p}$ of the boundary probability map and corresponding persistence barcode for topology structures. The red letters indicate connected components $(d = 0)$ , which is signified by the blue solid bars in the persistence barcode map. The golden letters indicate holes $(d = 1)$ , which are signified by the blue open bars. + +Then the persistence of $S_{ij}^{d}$ is measured by how long it survives in the sequence of combinations, which can be expressed through the persistence barcode whose length represents the corresponding persistence, Fig. 4. The longer the persistence barcode, the more robust the subset. Thereby, the persistence $P_{ij}^{d}$ of $S_{ij}^{d}$ can be calculated as the difference between the threshold values at which $S_{ij}^{d}$ first appears $(p_{birth}(S_{ij}^{d}))$ and when it disappears $(p_{death}(S_{ij}^{d}))$ : + +$$ +P _ {i j} ^ {d} = p _ {\text {b i r t h}} \left(S _ {i j} ^ {d}\right) - p _ {\text {d e a t h}} \left(S _ {i j} ^ {d}\right). \tag {5} +$$ + +Given prediction probabilities with K classes of a 2D image, the persistence encoding $P$ is + +$$ +\mathbf {P} = \left\{P _ {i j} ^ {d} \mid i = 1, 2, \dots , K; d = 0, 1; j = 1, 2, \dots , J _ {d} \right\}. \tag {6} +$$ + +Note that since ground truth can be regarded as a probability map containing only binary values of 0 and 1, the persistence encoding of it is + +$$ +\mathbf {Y} = \left\{Y _ {i j} ^ {d} \mid i = 1, 2, \dots , K; d = 0, 1; j = 1, 2, \dots , B _ {i} ^ {d} \right\}, \tag {7} +$$ + +where the value of $Y_{ij}^{d}$ is 1. As predicted probability maps typically contain more topological structures than those in ground truth, we assume that $J_{d} > B_{i}^{d}$ . + +# 3.3.2 Topology-Aware Loss + +Based on persistence encoding, a better prediction should meet the condition that the subsets of $C_n$ consistent with ground truth have greater persistence. We match the top $B_i^d$ persistence barcode $P_{ij}^d$ with GT and regard the remaining codes as unmatched. Our purpose is to reduce the error + +between the persistence encoding of prediction and ground truth, so as to make the topology correctness of prediction more robust. Therefore, the difference between them can be calculated in two parts: + +$$ +M _ {i} = \sum_ {d \in \{0, 1 \}} \sum_ {j = 1} ^ {B _ {i} ^ {d}} \left(Y _ {i j} ^ {d} - P _ {i j} ^ {d}\right), \tag {8} +$$ + +$$ +U _ {i} = \sum_ {d \in \{0, 1 \}} \sum_ {j = B _ {i} ^ {d} + 1} ^ {J _ {d}} (P _ {i j} ^ {d}), +$$ + +where $M_{i}$ measures the length difference between the matched barcode and GT barcode for the $i$ th class, and $U_{i}$ measures the length of the unmatched barcode. + +Given prediction probabilities with $K$ classes of a 2D image, the topology-aware loss $\mathcal{L}_{topo}$ is + +$$ +\mathcal {L} _ {t o p o} = \frac {1}{K} \sum_ {i = 1} ^ {K} \left(M _ {i} + U _ {i}\right). \tag {9} +$$ + +Minimizing this loss is equivalent to maximizing the persistence of topology features that are consistent with ground truth, and minimizing the persistence of undesired features. This provides a structure perspective to optimize the model, which overcomes the limitation of pixel-wise loss and achieves superior nucleus segmentation performance. + +# 3.4. Adaptive Topology-Aware Selection (ATS) + +ATS aims at efficiently optimizing images with topology-important regions by providing image- and patch-level selection for loss calculation, Fig. 5. + +At the image level, we calculate the topology error between prediction and GT, then selects and stores images with an adaptive threshold, Fig. 5 (Left). If the topology error exceeds the threshold, it indicates that the overall structure of the image is significantly different from ground truth, then the image is preferentially selected and stored for further learning. On the contrary, if the topology error does not exceed the threshold, it will be dropped out in the next iteration. The adaptive threshold is determined based on the current number of iterations and topology error rate. The more iterations, the higher the adaptive threshold. + +At the patch level, we split the pipeline into three main parts, Fig. 5 (Right). First, we divide the original image into hierarchical patches of varying scales. Then the topology error between predicted patches and corresponding ground truth is calculated. Moreover, a sorted list of patches is built based on these. According to their topology errors, an adaptive screen method is adopted. This allows the calculation of topology-aware loss on selected patches, which drives the model to focus on regions with more topology errors. + +In more detail, the Betti error is calculated by comparing the difference in the number of disjoint pixel sets of each + +![](images/a3ebc062e8aa456b6df9b73b400a5b05c634be81c91e1a2eeeba8a4372aab33c.jpg) +Figure 5. The pipeline of adaptive topology-aware selection module (ATS). In the topological Error Calculation and Sort and Adaptive Selection, the darker the square, the larger the Betti error. + +class between the prediction and GT, Fig. 5 (Bottom right). Note that the Betti error calculation used in the selection at the image and patch levels is the same. The difference is that one uses the whole image and the other uses the patch. + +# 4. Experiment + +# 4.1. Experimental Settings + +Datasets. I) MoNuSeg [17] contains 30 H&E-stained pathological images of size $1000 \times 1000$ from seven organs, including a total of 21,623 nuclei labeled without distinction between various categories. According to [17], we divide 30 images into three sets: 12 for training, 4 for validation, and 14 for testing (combining the same organ and different organ test sets). II) CPM17 [27] contains 64 H&E stained histopathology images with 7,570 annotated nuclear boundaries. It is from the MICCAI 2017 Digital Pathology Challenge [27] and images on two different scales: $500 \times 500$ and $600 \times 600$ . This is also a dataset for single-class segmentation task. III) PBNulei is a nuclear segmentation dataset from Peking University Affiliated Hospital, marked by professional pathologists. It contains 106 H&E stained histopathology images of $512 \times 512$ sizes, with a total of 36,659 nuclei, 78 images for training and 28 for testing. The dataset is divided into five categories, namely carcinoma, epithelial, inflammatory, mesenchymal, and myoepithelium. + +Evaluation Metrics. We use three instance-level evaluation metrics to measure the instance segmentation performance of the comparison models, which are: mean Panoptic Quality (PQ), the Aggregated Jaccard Index (AJI), and average Hausdorff distance (Haus). + +Implementation details. The proposed TopoSeg is trained and tested using the open-source software library Pytorch 1.8.0 on NVIDIA Tesla V100 with CUDA 10.1. + +The MoNuSeg dataset has a relatively large original image size and the CPM17 dataset has an inconsistent original + +![](images/bb4dba25e6c96288524d3b93dbb08b6934865142392501d7f171e588a8c3e646.jpg) + +![](images/be6c2c2a2701b4164a4664aa572f25ec156feda52202e78f6e0293d8f41a37fa.jpg) +Figure 6. Qualitative comparison before and after adding the topology-aware loss. (a) the probability map obtained by the model just using pixel-wise loss; (b) the probability map obtained by the model using pixel-wise loss and topology-aware loss. + +
DatasetBackboneLossPQ ↑AJI ↑Haus ↓
pixeltopo
MoNuSegU-Net✓0.5810.57912.80
U-Net✓✓0.5970.6199.647
CPMU-Net✓0.6250.6667.740
U-Net✓✓0.6620.7185.961
PBNuleiU-Net✓0.4260.3862.770
U-Net✓✓0.4250.4161.953
+ +Table 1. Comparison of instance segmentation results obtained using only pixel-wise loss and adding topology-aware loss based on pixel-wise loss in three datasets. + +image size, so it is divided into multiple patches, thus allowing the network to be uniform and fully trained. As a result, the MoNuSeg dataset generates 192 training patches, the + +CPM17 dataset generates 512 training patches. In the training phase, we apply a series of data augmentation methods, which include random flip, random rotation, random color, random crop, and Gaussian blur. The training epoch is set as 300, and the initial learning rate is set as $10^{-3}$ . + +# 4.2. Ablation studies + +We conducted ablation studies using three datasets, MoNuSeg, CPM17, and PBNulei, to validate the effectiveness of the proposed TopoSeg. We selected U-Net as the baseline segmentation method, and we improved the U-Net by adding two output branches: a semantic prediction branch and a three-class prediction branch. + +Topology-Aware Loss. When only pixel-wise losses are present, the model's performance on the MoNuSeg dataset is 0.581 PQ, 0.579 AJI, and 12.8 Haus, Table 1. After incorporating topological awareness losses, the model's PQ, AJI, and Haus on the MoNuSeg dataset are 0.597, 0.619, and 9.647, respectively, which have increased by $1.7\%$ , $3.9\%$ , and 3.153 compared to the original results. + +Compared to the performance obtained by using only pixel-wise loss, the model trained with the addition of pixelwise and topological loss has increased by $3.7\%$ PQ, $5.2\%$ AJI, and 1.779 Haus respectively on the CPM17 dataset and by $3.0\%$ AJI, and 0.817 Haus respectively on the PBNulei dataset compared to the original results. + +Fig. 6 makes a qualitative comparison between the boundary probability map obtained before and after adding the topology-aware loss and its corresponding persistence barcode. It shows that the corresponding structure obtained after adding topology-aware loss is more persistent. As an example, when $p = 0.5$ , three instances are correctly segmented, but when only pixel-wise loss is used, one instance can not be predicted. This indicates that adding topology-aware loss on the basis of pixel-wise loss helps the model learn object structure information, thus more accurately segmenting nuclear instances. + +Adaptive Image Selection. We set the threshold of $25\%$ , $50\%$ , $75\%$ and the adaptive threshold for the comparisons. According to the threshold, select the image with Betti error greater than the threshold as the new training dataset, and ensure that the number of training data obtained by different thresholds is the same. + +In Table 2, we compare the performance before and after image selection. The comparison results show that the segmentation results obtained after image selection are better than those without image selection. This proves that image selection promotes the model to focus on learning images with more structural prediction errors and improves the instance segmentation performance of the model. + +Compared to training with data selected with a threshold of $25\%$ or $75\%$ , training with data selected with a threshold of $50\%$ yielded more competitive model performance, + +
DatasetsBackboneSimgPQ↑AJI↑Haus↓
MoNuSegU-Net100%0.5970.6199.647
75%0.6040.6269.358
50%0.6060.6289.328
25%0.6020.6259.452
Adaptive0.6100.6329.113
CPMU-Net100%0.6620.7185.961
75%0.6650.7235.835
50%0.6700.7245.887
25%0.6720.7235.847
Adaptive0.6790.7315.805
PBNuleiU-Net100%0.4250.4161.953
75%0.4270.4192.014
50%0.4190.4142.028
25%0.4290.4182.009
Adaptive0.4460.4351.931
+ +Table 2. Comparison results of the models obtained by selecting images based on different thresholds. Here, $100\%$ represents the use of raw data without any image selection operation. $25\%$ , $50\%$ , and $75\%$ represent selecting images with threshold values greater than $25\%$ , $50\%$ , and $75\%$ , respectively, as new input data. Adaptive represents the proposed adaptive image selection method. + +
DatasetsBackbone\( S_{patch} \)PQ↑AJI↑Haus ↓
MoNuSegU-NetWhole0.5970.6199.647
Non-hierc0.6090.62810.063
Hierarchical0.6130.6398.842
Adaptive0.6160.6438.787
CPMU-NetWhole0.6620.7185.961
Non-hierc0.6750.7296.108
Hierarchical0.6620.7335.510
Adaptive0.6780.7415.605
PBNuleiU-NetWhole0.4250.4161.953
Non-hierc0.4980.4792.320
Hierarchical0.5140.4992.011
Adaptive0.5200.5042.185
+ +Table 3. Ablation studies on different patch selection methods. $S_{patch}$ denotes the method of dividing the image, "Whole" denotes the whole image for calculation, "Non-hierc" means non-hierarchical and refers to single size patch division, "Hierarchical" refers to multi-scale patch division, and "Adaptive" refers to the adaptive approach of dividing the patch. + +indicating that overly high or low thresholds can limit the learning of the model. Therefore, selecting a single threshold for image selection has limitations. + +Compared to the three fixed threshold image filtering methods, our method has the ability to flexibly select input data. As shown in Table 2, the adaptive image selection approach achieved superior segmentation performance. In the MoNuSeg dataset, the performance metrics PQ, AJI, + +![](images/85994c6154f84a21905c204e932f9161284a922d443813a83480e32a9d1d7ecc.jpg) +Figure 7. Visualization of instance segmentation results with different methods. (a) Input image; (b) Ground Truth; (c) U-Net [25]; (d) Micro-Net [24]; (e) FullNet [23]; (f) Hover-Net [9]; (g) CDNet [11], (h) Our TopoSeg. Different colors indicate different instances in the images. The white rectangles are drawn for a clear comparison. + +and Haus improved by $1.2\%$ , $1.3\%$ , and $0.534$ , respectively, compared to the original results. Similarly, compared with the original results, the performance of the CPM17 dataset is improved by $1.7\%$ PQ, $1.3\%$ AJI and $0.16$ Haus respectively, and the performance of the PBNulei dataset is improved by $2\%$ PQ, $1.9\%$ AJI and $0.022$ Haus, respectively. + +This reflects that, compared with the three fixed threshold image selection methods, our adaptive image selection approach has the ability to select input data more flexibly, thus promoting the model to obtain more correct instance segmentation results. + +Adaptive Hierarchical Patch Selection. As shown in Table 3, we compare the performance of the topology loss calculation after patch selection with that of the whole image. First of all, we verify the effect of patch selection on segmentation performance. Comparing the first row (Whole) and the second row (Non-hierc) in each dataset in Table 3, it is concluded that the segmentation performance of topology loss calculation after selecting patch is better than that of using the whole image on both PQ and AJI. This shows that by selecting the appropriate patch, the model + +pays more attention to the regions that need structural optimization, so as to improve the segmentation performance. + +Then, we verified the effect of using hierarchical patches with different scales for calculating topological loss in instance segmentation. We compared the performance of the third row (Hierarchical) and the second row (Non-hierc) within each dataset in Table 3 and found that hierarchical patches outperformed single-layer patches. This also shows that selecting the appropriate region that needs structural optimization helps the model better segment instances. + +Table 3 shows that compared with the topology loss calculated by the whole image, the segmentation performance obtained by adaptive hierarchical selection of patch is significantly improved. In the MoNuSeg dataset, PQ increased by $1.8\%$ , AJI increased by $2.4\%$ , Haus increased by 0.86. In the CPM dataset, PQ increased by $1.6\%$ , AJI increased by $2.3\%$ , Haus increased by 0.356. In the PBNulei dataset, PQ increased by $9.5\%$ , AJI increased by $8.8\%$ . + +
MethodMoNuSegCPM17PBNuclei
PQ↑AJI↑PQ↑AJI↑PQ↑AJI↑
U-Net [25]0.5810.5790.6250.6660.3550.312
Mask-RCNN [13]0.5090.5460.6740.684--
DCAN [2]0.4920.5250.5450.5610.2430.249
DIST [21]0.4430.5590.5040.6160.3250.301
Micro-Net [24]0.5190.560.6610.6680.3060.287
Full-Net [23]-0.604-0.6610.3850.378
PFF-Net [18]0.5870.611----
Hover-Net [9]0.5970.6180.6970.7050.4160.363
CDNet [11]-0.637-0.7330.4130.381
Ours0.6250.6430.7050.7560.5280.509
+ +Table 4. Performance comparisons on three nuclei datasets. + +# 4.3. Performance Comparison + +In Table 4, the proposed TopoSeg is compared with the state-of-the-art nuclear segmentation model on MoNuSeg and CPM17. Our TopoSeg achieves 0.625 PQ and 0.643 AJI on the MoNuSeg dataset and 0.705 PQ and 0.756 AJI on CPM17 dataset, respectively. The TopoSeg achieves competitive results for nuclei segmentation task. + +We further carry out a qualitative visualization analysis of our method and different existing methods. Compared to all the other methods in Fig. 7, our TopoSeg achieves better instance segmentation results for nuclear instance segmentation, especially for the correct structure segmentation of dense or overlapping regions. Both quantitative and qualitative results demonstrate that our TopoSeg is competitive and able to segment dense nuclei instances. + +# 5. Conclusion + +This paper proposes a novel topology-aware segmentation approach, termed TopoSeg, for nuclear instance segmentation tasks to address the challenge of touching and overlapping nuclear segmentation. We develop a TAM, which contains a persistence encoding method that accounts for topology information in predicted probability maps, and a topology-aware loss that is flexible and suitable for topology optimization. We further improve it by applying an ATS, which enforces the model to focus on regions with high topology errors. The experimental results in three nuclei datasets show that TopoSeg has achieved superior performance. As a general method of topology awareness, TopoSeg provides fresh insight into the instance segmentation problem of small and dense objects. + +Acknowledgments This work was supported in part by the National Key R&D Program of China (No. 2022ZD0118201), Natural Science Foundation of China (No. 61972217, 32071459, 62176249, 62006133, 62271465). + +# References + +[1] Nick Byrne, James R Clough, Israel Valverde, Giovanni Montana, and Andrew P King. A persistent homology-based topological loss for cnn-based multiclass segmentation of cmr. IEEE Transactions on Medical Imaging, 42(1):3-14, 2022. 2, 3, 4 +[2] Hao Chen, Xiaojuan Qi, Lequan Yu, Qi Dou, Jing Qin, and Pheng-Ann Heng. Dcan: Deep contour-aware networks for object instance segmentation from histology images. Medical Image Analysis, 36:135–146, 2017. 2, 9 +[3] Shengcong Chen, Changxing Ding, Minfeng Liu, Jun Cheng, and Dacheng Tao. CPP-net: Context-aware polygon proposal network for nucleus segmentation. IEEE Transactions on Image Processing, 32:980-994, 2023. 1, 2 +[4] James R Clough, Nicholas Byrne, Ilkay Oksuz, Veronika A Zimmer, Julia A Schnabel, and Andrew P King. A topological loss function for deep-learning based image segmentation using persistent homology. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):8766-8778, 2020. 3 +[5] James R Clough, Ilkay Oksuz, Nicholas Byrne, Julia A Schnabel, and Andrew P King. Explicit topological priors for deep-learning based image segmentation using persistent homology. In International Conference on Information Processing in Medical Imaging, pages 16-28, 2019. 3 +[6] Herbert Edelsbrunner, David Letscher, and Afra Zomorodian. Topological persistence and simplification. In Proceedings of the 41st Annual Symposium on Foundations of Computer Science, pages 454-463, 2000. 2 +[7] Robin Forman. Morse theory for cell complexes. Advances in Mathematics, 134(1):90-145, 1998. 3 +[8] Robin Forman. A user's guide to discrete Morse theory. Séminaire Lotharingien de Combinatoire [electronic only], 48:1-35, 2002. 3 +[9] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:1-15, 2019. 1, 2, 8, 9 +[10] Saumya Gupta, Xiaoling Hu, James Kaan, Michael Jin, Muth shipay Mpoy, Katherine Chung, Gagandeep Singh, Mary Saltz, Tahsin Kurc, Joel Saltz, et al. Learning topological interactions for multi-class medical image segmentation. In Proceedings of the European Conference on Computer Vision, pages 701-718, 2022. 4 +[11] Hongliang He, Zhongyi Huang, Yao Ding, Guoli Song, Lin Wang, Qian Ren, Pengxu Wei, Zhiqiang Gao, and Jie Chen. Cdnet: Centripetal direction network for nuclear instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4026-4035, 2021. 2, 8, 9 +[12] Hongliang He, Chi Zhang, Jie Chen, Ruizhe Geng, Luyang Chen, Yongsheng Liang, Yanchang Lu, Jihua Wu, and Yongjie Xu. A hybrid-attention nested unet for nuclear segmentation in histopathological images. Frontiers in Molecular Biosciences, 8:1–9, 2021. 1, 2 + +[13] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2961-2969, 2017. 9 +[14] Xiaoling Hu, Fuxin Li, Dimitris Samaras, and Chao Chen. Topology-preserving deep image segmentation. Advances in Neural Information Processing Systems, 32:5658-5669, 2019. 2, 3 +[15] Xiaoling Hu, Dimitris Samaras, and Chao Chen. Learning probabilistic topological representations using discrete Morse theory. arXiv preprint arXiv:2206.01742, 2022. 3 +[16] Xiaoling Hu, Yusu Wang, Li Fuxin, Dimitris Samaras, and Chao Chen. Topology-aware segmentation using discrete Morse theory. In Proceedings of the International Conference on Learning Representation, pages 1-19, 2021. 2, 3 +[17] Neeraj Kumar, Ruchika Verma, Sanuj Sharma, Surabhi Bhargava, Abhishek Vahadane, and Amit Sethi. A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Transactions on Medical Imaging, 36(7):1550-1560, 2017. 6 +[18] Dongnan Liu, Donghao Zhang, Yang Song, Heng Huang, and Weidong Cai. Panoptic feature fusion net: A novel instance segmentation paradigm for biomedical and biological images. IEEE Transactions on Image Processing, 30:2045-2059, 2021. 1, 9 +[19] Xiaoming Liu, Zhengsheng Guo, Jun Cao, and Jinshan Tang. Mdc-net: a new convolutional neural network for nucleus segmentation in histopathology images with distance maps and contour information. Computers in Biology and Medicine, 135:1-12, 2021. 1, 2 +[20] Agata Mosinska, Pablo Marquez-Neila, Mateusz Kozinski, and Pascal Fua. Beyond the pixel-wise loss for topology-aware delineation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3136–3145, 2018. 2 +[21] Peter Naylor, Marick Laé, Fabien Reyal, and Thomas Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE Transactions on Medical Imaging, 38(2):448-459, 2018. 1, 2, 9 +[22] H. Pinckaers, W. Bulten, J. van der Laak, and G. Litjens. Detection of prostate cancer in whole-slide images through end-to-end training with image-level labels. IEEE Transactions on Medical Imaging, 40(7):1817-1826, 2021. 1 +[23] Hui Qu, Zhennan Yan, Gregory M Riedlinger, Subhajyoti De, and Dimitris N Metaxas. Improving nuclei/gland instance segmentation in histopathology images by full resolution neural network and spatial constrained loss. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 378-386, 2019. 1, 2, 8, 9 +[24] Shan E Ahmed Raza, Linda Cheung, Muhammad Shaban, Simon Graham, David Epstein, Stella Pelengaris, Michael Khan, and Nasir M Rajpoot. Micro-net: A unified model for segmentation of various objects in microscopy images. Medical Image Analysis, 52:160-173, 2019. 2, 8, 9 +[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on + +Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 1, 2, 8, 9 +[26] Suprosanna Shit, Johannes C Paetzold, Anjany Sekuboyina, Ivan Ezhov, Alexander Unger, Andrey Zhylka, Josien PW Pluim, Ulrich Bauer, and Bjoern H Menze. cldice-a novel topology-preserving loss function for tubular structure segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16560-16569, 2021. 2 +[27] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in Bioengineering and Biotechnology, 53:1-15, 2019. 6 +[28] Yanning Zhou, Omer Fahri Onder, Qi Dou, Efstratios Tsougenis, Hao Chen, and Pheng-Ann Heng. Cia-net: Robust nuclei instance segmentation with contour-aware information aggregation. In International Conference on Information Processing in Medical Imaging, pages 682–693, 2019. 2 \ No newline at end of file diff --git a/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/images.zip b/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..063b3f4bc7383e59bdd0b4eff7124086169578c7 --- /dev/null +++ b/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ef618f3f09d95f24c75515a46c4b50bce279891ff5b08b91d92f9aa419874a4 +size 1089006 diff --git a/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/layout.json b/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a74f436e35d6be8af5afe206a376d71830cadd63 --- /dev/null +++ b/2023/TopoSeg_ Topology-Aware Nuclear Instance Segmentation/layout.json @@ -0,0 +1,8846 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 117, + 103, + 477, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 103, + 477, + 121 + ], + "spans": [ + { + "bbox": [ + 117, + 103, + 477, + 121 + ], + "type": "text", + "content": "TopoSeg: Topology-Aware Nuclear Instance Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "spans": [ + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "text", + "content": "Hongliang He" + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "text", + "content": " Jun Wang" + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "text", + "content": " Pengxu Wei" + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "text", + "content": " Fan Xu" + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "text", + "content": " Xiangyang Ji" + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "text", + "content": " Chang Liu" + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "text", + "content": " Jie Chen" + }, + { + "bbox": [ + 146, + 142, + 447, + 172 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 176, + 514, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 176, + 514, + 233 + ], + "spans": [ + { + "bbox": [ + 79, + 176, + 514, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 176, + 514, + 233 + ], + "type": "text", + "content": "School of Electronic and Computer Engineering, Peking University, Shenzhen, China \n" + }, + { + "bbox": [ + 79, + 176, + 514, + 233 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 176, + 514, + 233 + ], + "type": "text", + "content": "School of Computer Science and Technology, Soochow University, Suzhou, China \n" + }, + { + "bbox": [ + 79, + 176, + 514, + 233 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 79, + 176, + 514, + 233 + ], + "type": "text", + "content": "Sun Yat-sen University, Guangzhou, China \n" + }, + { + "bbox": [ + 79, + 176, + 514, + 233 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 79, + 176, + 514, + 233 + ], + "type": "text", + "content": "Peng Cheng Laboratory, Shenzhen, China \n" + }, + { + "bbox": [ + 79, + 176, + 514, + 233 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 79, + 176, + 514, + 233 + ], + "type": "text", + "content": "Department of Automation and BNRist, Tsinghua University, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 236, + 466, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 236, + 466, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 236, + 466, + 262 + ], + "type": "text", + "content": "hehl@pku.edu.cn, junw@stu.pku.edu.cn, weipx3@mail.sysu.edu.cn, {xuf01,chenj}@pcl.ac.cn, {xyji,liuchang2022}@tsinghua.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 289, + 192, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 289, + 192, + 301 + ], + "spans": [ + { + "bbox": [ + 143, + 289, + 192, + 301 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 315, + 290, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 290, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 290, + 578 + ], + "type": "text", + "content": "Nuclear instance segmentation has been critical for pathology image analysis in medical science, e.g., cancer diagnosis. Current methods typically adopt pixel-wise optimization for nuclei boundary exploration, where rich structural information could be lost for subsequent quantitative morphology assessment. To address this issue, we develop a topology-aware segmentation approach, termed TopoSeg, which exploits topological structure information to keep the predictions rational, especially in common situations with densely touching and overlapping nucleus instances. Concretely, TopoSeg builds on a topology-aware module (TAM), which encodes dynamic changes of different topology structures within the three-class probability maps (inside, boundary, and background) of the nuclei to persistence barcodes and makes the topology-aware loss function. To efficiently focus on regions with high topological errors, we propose an adaptive topology-aware selection (ATS) strategy to enhance the topology-aware optimization procedure further. Experiments on three nuclear instance segmentation datasets justify the superiority of TopoSeg, which achieves state-of-the-art performance. The code is available at https://github.com/hhlisme/toposeg." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 621, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 621, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 287, + 694 + ], + "type": "text", + "content": "Pathological image analysis is considered the gold standard for cancer diagnosis, treatment, and prevention, where nuclear instance segmentation serves as the core step. It is because the morphology of nuclei including the shape, appearance, and distribution within pathology images provides interpretable features of diagnostic and prognostic" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 290, + 544, + 414 + ], + "blocks": [ + { + "bbox": [ + 310, + 290, + 544, + 414 + ], + "lines": [ + { + "bbox": [ + 310, + 290, + 544, + 414 + ], + "spans": [ + { + "bbox": [ + 310, + 290, + 544, + 414 + ], + "type": "image", + "image_path": "ad304ca217be5ad48b46bb88830e5649fecc03d361e7a3b459c790b29d378d6b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 417, + 547, + 528 + ], + "lines": [ + { + "bbox": [ + 304, + 417, + 547, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 547, + 528 + ], + "type": "text", + "content": "Figure 1. The schematic illustration of pixel-wise loss and topology-aware loss. (Left) A patch in the image and its corresponding boundary ground truth, as well as the original boundary segmentation probability map and its corresponding instance prediction. (Top-right) The gradient map was obtained by pixel-wise loss, the 3D map of the boundary probability map and the optimization trend, as well as the final instance prediction. (Bottom-right) The gradient map obtained from topology-aware loss, the 3D image of the boundary probability map and the optimization trend, as well as the final instance prediction." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 542, + 397, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 542, + 397, + 554 + ], + "spans": [ + { + "bbox": [ + 306, + 542, + 397, + 554 + ], + "type": "text", + "content": "cancer indicators[22]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 556, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 546, + 652 + ], + "type": "text", + "content": "Compared to general image object segmentation, nuclear instance segmentation faces with some particular challenges: (i) the crowded nuclei distribution leads to clustered overlapping instances. (ii) The blurry boundaries between touching or overlapping nuclei increase the difficulty of distinguishing individual instances. (iii) Large variability in size, shape, texture and intensity of nuclei between different images requires more robust annotation and prediction [18]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "To cope with these challenges, several studies have focused on exploring innovative network architectures [12, 23, 25] and incorporating distance maps [3, 9, 19, 21] to enhance segmentation performance. However, existing works typically rely on pixel-wise losses, such as CE/Dice loss," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 212, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 212, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 212, + 713 + ], + "type": "text", + "content": "*Equal contribution ☑ Corresponding Author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21307" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "which enforce predictions to align with GT pixel-by-pixel yet ignore the crucial structure information of instances. Since treat pixels independently, they lack sensitivity towards high-level structural characteristics, such as the connectivity of boundaries, which are essential for the rational segmentation of individual instances. This inability of pixel-wise loss often leads to unreasonable results such as broken boundaries, bringing about failure to distinguish touching or overlapping instances. This consequently results in undesirable effects on the subsequent quantitative morphology analysis. As depicted in Fig. 1 (top-right), pixel-wise loss focuses on misaligned pixels thus reducing pixel errors. However, due to the lack of perception of structural information, the boundaries between overlapping instances with smaller pixel proportions are not given enough consideration and result in incorrect instance predictions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 264, + 289, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 503 + ], + "type": "text", + "content": "To capture structural information of nuclear instances, we propose a novel topology-aware nuclear instance segmentation approach, termed TopoSeg. Specifically, we introduce a topology encoding method to encode topological features of three-class maps which are utilized to separate individual instances. Building upon this, we propose a topology-aware module (TAM), which encodes dynamic changes and persistence of different topology structures within the predicted three-class probability maps to persistence barcodes and develops a topology-aware loss function, integrating topology information into model learning. As shown in Fig. 1 (bottom-right), with topology constraints, the critical boundaries are identified and optimized, resulting in a reasonable segmentation outcome. To further adapt to the task of dense and small instances like nuclei, we introduce an adaptive topology-aware selection module (ATS) to concentrate on images and regions requiring topological optimization. Notably, TopoSeg effectively promotes instance segmentation performance without requiring any changes to the original network structure." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 504, + 258, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 504, + 258, + 515 + ], + "spans": [ + { + "bbox": [ + 59, + 504, + 258, + 515 + ], + "type": "text", + "content": "The contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 526, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 526, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 526, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 59, + 526, + 287, + 574 + ], + "type": "text", + "content": "- We propose TopoSeg, which provides a new perspective that goes beyond the common pixel-wise optimization paradigm and incorporates meaningful topology constraints for nuclear instance segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 584, + 287, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 584, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 59, + 584, + 287, + 620 + ], + "type": "text", + "content": "- We present a TAM to encode the persistence of topology structures in probability maps and introduce a topology-aware loss as a further constraint." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 631, + 287, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 631, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 59, + 631, + 287, + 679 + ], + "type": "text", + "content": "- We develop an ATS to identify regions that most require topology optimization, enabling the model to efficiently accommodate segmentation tasks involving small and dense objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 689, + 287, + 714 + ], + "type": "text", + "content": "- We justify the superiority of our TopoSeg on three nuclei datasets with state-of-the-art performance." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 393, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 91, + 425, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 425, + 105 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 425, + 105 + ], + "type": "text", + "content": "2.1. Nuclei Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 110, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 266 + ], + "type": "text", + "content": "Deep learning has been extensively utilized for nuclear instance segmentation [2, 9, 11, 24]. In 2015, Ronneberger et al. [25] proposed the U-Net model, which has become one of the most fundamental models in medical image segmentation. Raza et al. [24] proposed Micro-Net, which achieves robustness to large internal/external variances in nucleus size by utilizing multi-resolution and weighted loss functions. Qu et al. [23] proposed a full-resolution convolutional neural network (FullNet) with no downsampling operation employed in the network structure to enhance localization accuracy. He et al. [12] proposed a hybrid attention nested U-shaped network (Han-Net) to extract effective feature information from multiple layers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 266, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 492 + ], + "type": "text", + "content": "To take advantage of contour information to differentiate contact/overlapping nuclei, Chen et al. [2] initially proposed incorporating contour information into a multi-level FCN to create a deep contour-aware network for nuclear instance segmentation. Subsequently, Zhou et al. [28] proposed the contour-aware information aggregation network to combine spatial and textural features between nuclei and contours. Additionally, some works [3, 9, 19, 21] introduced distance maps to separate contact/overlapping nuclei. Naylor et al. [21] addressed the issue of segmenting touching nuclei by formulating the segmentation task as a regression task of intra-nuclear distance maps. Graham et al. [9] proposed a network (Hover-Net) for simultaneous nucleus segmentation and classification, which uses the vertical and horizontal distances between a nucleus pixel and its center of mass to separate clusters of nuclei. Moreover, He et al. [11] proposed a centripetal directional network (CDNet) for nuclear instance segmentation, combining directional information into the network." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 494, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 545, + 541 + ], + "type": "text", + "content": "The above models have achieved some performance improvements in nuclear segmentation, but these models using pixel-wise loss tend to ignore structural information of nuclear boundaries." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 551, + 440, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 440, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 440, + 563 + ], + "type": "text", + "content": "2.2. Segmentation Topology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 569, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 688 + ], + "type": "text", + "content": "Several topology-based approaches [1, 14, 16, 20, 26] have exploited ways to guarantee the topological accuracy of objects in images. Mosinska et al. [20] considered topological information by constraining the similarity between feature maps and predicted probability maps generated by a pre-trained VGG-19 network. Shit et al. [26] proposed a similarity metric (clDice) to calculate the intersection between segmentation masks and their morphological skeletons, and developed a novel loss function (Soft-clDice) to train the neural network based on the skeleton structure." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "In addition, the differentiability of persistent homology [6] allows it to be formulated as a loss function and ap-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21308" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 545, + 202 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 545, + 202 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 545, + 202 + ], + "type": "image", + "image_path": "aad59e3199bb222ed617aacb0807f12618ee40d2a222fdc8b2a864451edc0441.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 203, + 545, + 318 + ], + "blocks": [ + { + "bbox": [ + 51, + 203, + 545, + 318 + ], + "lines": [ + { + "bbox": [ + 51, + 203, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 51, + 203, + 545, + 318 + ], + "type": "image", + "image_path": "af60d89dded5e5049dfa0c31a65e1eb0870b30661e568c58986d22d928aa184f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 321, + 546, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 321, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 546, + 354 + ], + "type": "text", + "content": "Figure 2. The overall framework of the proposed TopoSeg. The network contains two basic prediction branches: the semantic branch and the three-class (inside, boundary, background) branch. In order to be aware of topology, the framework contains two important modules: the adaptive topology-aware selection module and the topology-aware module." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 369, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 536 + ], + "type": "text", + "content": "plied to end-to-end deep neural networks. Hu et al. [14] developed a loss function that directly constrains the prediction results and topological GT based on persistent homology. Furthermore, Hu et al. [16] used Discrete Morse Theory [7, 8] to capture the singularities of the gradient vector field of the likelihood function to find all potential skeletons of the object. They also leveraged Discrete Morse Theory and persistent homology to construct a topological representation space and trained a probabilistic model for performing inference tasks in such a topological representation space [15]. In cardiac magnetic resonance (CMR), Clough et al. [1, 5, 4] investigated various methods to couple topological information, where the number of components in the required segmentation is known a priori." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 537, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 287, + 609 + ], + "type": "text", + "content": "However, the above methods have certain limitations in terms of computational efficiency, topology completeness, and task flexibility. Therefore, we propose a topology-aware model for nuclear instance segmentation, which encodes structural information to construct topological losses, thus improving the instance segmentation performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 624, + 129, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 129, + 638 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 129, + 638 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 645, + 115, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 115, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 115, + 657 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Fig. 2 provides a flowchart of our TopoSeg, which adopts a dual-branch architecture with two types of predictions: semantic prediction of nuclear semantic categories (e.g. carcinoma, epithelial, etc.) and three-class prediction of struc" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 369, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 429 + ], + "type": "text", + "content": "tural categories (inside, boundary, background). The output of three-class branch is sequentially processed through an adaptive topology-aware selection module (ATS) and a topology-aware module (TAM) to efficiently capture the topology information of nuclear instances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 431, + 546, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 546, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 546, + 575 + ], + "type": "text", + "content": "Specifically, to quantify the topological features of an image, we present topology encoding in Sec. 3.2. Topological information contained in three-class images, such as the number of holes and connected components of the boundary, can reflect the number of instances and overlapping characteristics, which is meaningful for reasonable instance segmentation. Therefore, we decompose the three-class image into three channels (inside, boundary, and union) and encode their topological features. This allows the direct representation of the structural correctness, but its inapplicability to predicated probability maps makes it challenging to incorporate into the training process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 578, + 545, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 650 + ], + "type": "text", + "content": "To address this issue, we introduce TAM in Sec. 3.3, which extracts topological structures of predicted three-class probability maps into persistence barcode through a persistence encoding method and constrains the topology similarity between GTs and predictions by a topology-aware loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "To further improve the topology perception of the model and reduce the computational cost, we develop ATS in Sec. 3.4 to adaptively select regions with greater topology importance in both image- and patch-level, which makes the topological constraints more suitable for the task of seg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21309" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 81, + 108, + 137 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 102, + 79 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 102, + 79 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 102, + 79 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 81, + 108, + 137 + ], + "lines": [ + { + "bbox": [ + 51, + 81, + 108, + 137 + ], + "spans": [ + { + "bbox": [ + 51, + 81, + 108, + 137 + ], + "type": "image", + "image_path": "b323df670db911b86f3dc58dfd9a1b544cabf8465bf8ba3f980f42a86d55f9c9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 110, + 81, + 167, + 137 + ], + "blocks": [ + { + "bbox": [ + 109, + 71, + 159, + 79 + ], + "lines": [ + { + "bbox": [ + 109, + 71, + 159, + 79 + ], + "spans": [ + { + "bbox": [ + 109, + 71, + 159, + 79 + ], + "type": "text", + "content": "Prediction 1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 110, + 81, + 167, + 137 + ], + "lines": [ + { + "bbox": [ + 110, + 81, + 167, + 137 + ], + "spans": [ + { + "bbox": [ + 110, + 81, + 167, + 137 + ], + "type": "image", + "image_path": "dc29468e81133cb7c07a101e05a81dc3f2af7662f7cb1d36f0f96948ba72cce7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 169, + 81, + 226, + 137 + ], + "blocks": [ + { + "bbox": [ + 168, + 71, + 219, + 79 + ], + "lines": [ + { + "bbox": [ + 168, + 71, + 219, + 79 + ], + "spans": [ + { + "bbox": [ + 168, + 71, + 219, + 79 + ], + "type": "text", + "content": "Prediction 2" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 81, + 226, + 137 + ], + "lines": [ + { + "bbox": [ + 169, + 81, + 226, + 137 + ], + "spans": [ + { + "bbox": [ + 169, + 81, + 226, + 137 + ], + "type": "image", + "image_path": "9f08e60781f98b4bc5b8677527b0b7ed0c4667400c941332451b8df16a77ae54.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 228, + 81, + 284, + 137 + ], + "blocks": [ + { + "bbox": [ + 227, + 71, + 283, + 79 + ], + "lines": [ + { + "bbox": [ + 227, + 71, + 283, + 79 + ], + "spans": [ + { + "bbox": [ + 227, + 71, + 283, + 79 + ], + "type": "text", + "content": "Prediction 3" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 81, + 284, + 137 + ], + "lines": [ + { + "bbox": [ + 228, + 81, + 284, + 137 + ], + "spans": [ + { + "bbox": [ + 228, + 81, + 284, + 137 + ], + "type": "image", + "image_path": "845bfed57d0f6fb281172a474f4cf172bea3dd5f377bad0f4220dc9c423e16fb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 72, + 137, + 85, + 146 + ], + "blocks": [ + { + "bbox": [ + 72, + 137, + 85, + 146 + ], + "lines": [ + { + "bbox": [ + 72, + 137, + 85, + 146 + ], + "spans": [ + { + "bbox": [ + 72, + 137, + 85, + 146 + ], + "type": "image", + "image_path": "d9b6f36fbaef075ffbe9068d710dd3347f1cb54adeea3a8e0c77efd15d6c548a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 139, + 143, + 147 + ], + "lines": [ + { + "bbox": [ + 88, + 139, + 143, + 147 + ], + "spans": [ + { + "bbox": [ + 88, + 139, + 143, + 147 + ], + "type": "text", + "content": "Topology encoding" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 51, + 149, + 284, + 201 + ], + "blocks": [ + { + "bbox": [ + 51, + 149, + 284, + 201 + ], + "lines": [ + { + "bbox": [ + 51, + 149, + 284, + 201 + ], + "spans": [ + { + "bbox": [ + 51, + 149, + 284, + 201 + ], + "type": "image", + "image_path": "77a5769a0794d305397b1d43a797459dbf56ca01cea8c4e5319e0bdc8cf65d14.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 287, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 239 + ], + "type": "text", + "content": "Figure 3. (Top) The ground truth and different prediction results of two overlapping nuclei in the pathological image. (Bottom) An example of topology encoding of ground truth." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 254, + 191, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 254, + 191, + 265 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 191, + 265 + ], + "type": "text", + "content": "menting dense and small instances." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 266, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 290 + ], + "type": "text", + "content": "In general, during the training phase, the total loss function of TopoSeg is" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 112, + 300, + 287, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 300, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 112, + 300, + 287, + 314 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {p i x e l}} + \\lambda \\mathcal {L} _ {\\text {t o p o}}, \\tag {1}", + "image_path": "d9cd4a2289ac51fa05010c80391a257e140b69660cb3a8dc61ec5705000e6e22.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 323, + 287, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 372 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 323, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pixel}" + }, + { + "bbox": [ + 46, + 323, + 287, + 372 + ], + "type": "text", + "content": " denotes pixel-wise loss with respect to the output of both branches, " + }, + { + "bbox": [ + 46, + 323, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{topo}" + }, + { + "bbox": [ + 46, + 323, + 287, + 372 + ], + "type": "text", + "content": " denotes the proposed topology-aware in TAM with respect to the output at three-class branch and " + }, + { + "bbox": [ + 46, + 323, + 287, + 372 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 323, + 287, + 372 + ], + "type": "text", + "content": " denotes the weight of topology-aware loss." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 372, + 287, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 444 + ], + "type": "text", + "content": "The design of this total loss leverages the strengths of both types of losses. Initially, the model is trained with only the pixel-wise loss for a certain number of iterations (" + }, + { + "bbox": [ + 46, + 372, + 287, + 444 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 46, + 372, + 287, + 444 + ], + "type": "text", + "content": ") to obtain accurate pixel-level predictions. Subsequently, the topology loss is introduced (" + }, + { + "bbox": [ + 46, + 372, + 287, + 444 + ], + "type": "inline_equation", + "content": "\\lambda > 0" + }, + { + "bbox": [ + 46, + 372, + 287, + 444 + ], + "type": "text", + "content": ") to further enhance the model's ability to capture rational object structure." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 445, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 491 + ], + "type": "text", + "content": "In the inference phase, the final predicted instance segmentation map is obtained by utilizing boundaries predicted in the three-class branch to separate the semantic result of the semantic branch." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 501, + 160, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 160, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 160, + 514 + ], + "type": "text", + "content": "3.2. Topology Encoding" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 520, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 591 + ], + "type": "text", + "content": "As previously noted, pixel-wise loss is insensitive to the differences in a few pixels at the boundary between overlapping nuclei, as distinct examples depicted in Fig. 3, despite significant structural differences. To quantitatively represent them, we present a topology encoding approach whose general representation is as follows." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "text", + "content": "For image with " + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "text", + "content": " classes, consider pixels of class " + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "text", + "content": " as set " + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "text", + "content": ". The zero-dimensional topology structure refers to the connection components formed by the connected pixels in " + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "text", + "content": ", and the one-dimensional structure refers to the holes. The topology of " + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "text", + "content": " can be considered as a family of subsets " + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\{S_{ij}^{d}|d = 0,1;j = 1,2,\\dots,b_{d}\\}" + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "text", + "content": " formed by topological structures of different dimensions, where " + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "text", + "content": " represents topology dimension and " + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "inline_equation", + "content": "b_{d}" + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "text", + "content": " represents the number of subsets of dimension " + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "text", + "content": ". Then a subset " + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "inline_equation", + "content": "S_{ij}^{0}" + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "text", + "content": " represents a connected component and " + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "inline_equation", + "content": "S_{ij}^{1}" + }, + { + "bbox": [ + 46, + 593, + 287, + 715 + ], + "type": "text", + "content": " represents a hole, Fig. 3 (Bottom)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": "To quantify the topological characteristics of " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": ", we introduce the Betti number from algebraic topology. For a 2D image, the Betti numbers can be expressed as " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "B = (b_{0}, b_{1})" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "b_{0}" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " denotes the number of connected components and " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "b_{1}" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " denotes the number of 2D holes. Given a K-class image, the topology encoding can be formulated as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 352, + 149, + 545, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 149, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 352, + 149, + 545, + 163 + ], + "type": "interline_equation", + "content": "\\mathbf {B} = \\left\\{B _ {i} ^ {d} | i = 1, 2, \\dots , K; d = 0, 1 \\right\\}, \\tag {2}", + "image_path": "d290955bfc0d0fc857d9d9f7b179ea81e9f394e5c9e59cc5e2f58825164c8941.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 167, + 500, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 167, + 500, + 180 + ], + "spans": [ + { + "bbox": [ + 305, + 167, + 500, + 180 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 167, + 500, + 180 + ], + "type": "inline_equation", + "content": "B_{i}^{d}" + }, + { + "bbox": [ + 305, + 167, + 500, + 180 + ], + "type": "text", + "content": " represents Betti number of subset " + }, + { + "bbox": [ + 305, + 167, + 500, + 180 + ], + "type": "inline_equation", + "content": "S_{i}^{d}" + }, + { + "bbox": [ + 305, + 167, + 500, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 180, + 545, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 216 + ], + "type": "text", + "content": "In the nuclear instance segmentation task, the topology encoding of instance maps are mapped to three-class maps with three categories (inside, boundary and their union):" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 351, + 221, + 545, + 261 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 221, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 351, + 221, + 545, + 261 + ], + "type": "interline_equation", + "content": "\\mathbf {B} = \\left[ \\begin{array}{l} B _ {i n} \\\\ B _ {b o} \\\\ B _ {u n} \\end{array} \\right] = \\left[ \\begin{array}{c c} N _ {i n} & 0 \\\\ N _ {b o} & N _ {i n} \\\\ N _ {b o} & 0 \\end{array} \\right], \\tag {3}", + "image_path": "fccb5d746c0f0dc4606fbf12fecf74a3fcb1f4bffc710bfb2406a625860e064c.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 266, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 545, + 313 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 266, + 545, + 313 + ], + "type": "inline_equation", + "content": "B_{in}, B_{bo}" + }, + { + "bbox": [ + 305, + 266, + 545, + 313 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 266, + 545, + 313 + ], + "type": "inline_equation", + "content": "B_{un}" + }, + { + "bbox": [ + 305, + 266, + 545, + 313 + ], + "type": "text", + "content": " respectively represent the Betti numbers of the corresponding class. The specific values of " + }, + { + "bbox": [ + 305, + 266, + 545, + 313 + ], + "type": "inline_equation", + "content": "B_{i}^{d}" + }, + { + "bbox": [ + 305, + 266, + 545, + 313 + ], + "type": "text", + "content": " are defined based on the number of instances " + }, + { + "bbox": [ + 305, + 266, + 545, + 313 + ], + "type": "inline_equation", + "content": "N_{in}" + }, + { + "bbox": [ + 305, + 266, + 545, + 313 + ], + "type": "text", + "content": " and non-connected boundaries " + }, + { + "bbox": [ + 305, + 266, + 545, + 313 + ], + "type": "inline_equation", + "content": "N_{bo}" + }, + { + "bbox": [ + 305, + 266, + 545, + 313 + ], + "type": "text", + "content": " in the image." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 314, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 374 + ], + "type": "text", + "content": "Unlike the fixed topology prior of semantic given in [10] and [1], our topology encoding represents dynamically changing instance information for each image through three channels. As an example, the corresponding topology encoding of images in Fig. 3 (Top) is:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 333, + 388, + 515, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 388, + 515, + 432 + ], + "spans": [ + { + "bbox": [ + 333, + 388, + 515, + 432 + ], + "type": "interline_equation", + "content": "\\mathbf {B} (G T) = \\left[ \\begin{array}{c c} 2 & 0 \\\\ 1 & 2 \\\\ 1 & 0 \\end{array} \\right], \\mathbf {B} (P 1) = \\left[ \\begin{array}{c c} 2 & 0 \\\\ 1 & 2 \\\\ 1 & 0 \\end{array} \\right],", + "image_path": "e41acc22ead838aa879112f4c43be033a40e9aa49ff01e7722906e52ca702a8a.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 337, + 429, + 517, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 429, + 517, + 468 + ], + "spans": [ + { + "bbox": [ + 337, + 429, + 517, + 468 + ], + "type": "interline_equation", + "content": "\\mathbf {B} (P 2) = \\left[ \\begin{array}{c c} 1 & 0 \\\\ 1 & 1 \\\\ 1 & 0 \\end{array} \\right], \\mathbf {B} (P 3) = \\left[ \\begin{array}{c c} 2 & 0 \\\\ 2 & 2 \\\\ 2 & 0 \\end{array} \\right].", + "image_path": "ad387924fc84f7eb417dc2b2759c3f00fed2dfc10263a1b484cb3ec5c243ad05.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 473, + 443, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 473, + 443, + 486 + ], + "spans": [ + { + "bbox": [ + 306, + 473, + 443, + 486 + ], + "type": "text", + "content": "3.3. Topology-Aware Module" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 491, + 429, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 429, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 429, + 504 + ], + "type": "text", + "content": "3.3.1 Persistence Encoding" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "text", + "content": "Consider a probability map of class " + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "text", + "content": " as a topology set " + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "text", + "content": " with topology subsets " + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\{S_{ij}^{d}|d = 0,1;j = 1,2,\\dots,J_{d}\\}" + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "inline_equation", + "content": "J_{d}" + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "text", + "content": " represents the total number of subsets of dimension " + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 510, + 545, + 582 + ], + "type": "text", + "content": ". Here we use persistent homology to extract meaningful topology structures which are robust to noise and small perturbations in the probability map." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "The first step of persistent homology is to construct a sequence of combinations " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{C_n\\}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " of subsets, where each combination is obtained by adding more and more elements from " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", which can be obtained by thresholding the probability map at a sequence of decreasing threshold values from 1 to 0. During this process, some subsets disappear with the gradual increase of pixels, while others appear. In Fig. 4, subsets of different dimensions are marked with letters of different colors. The subset " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " both appear in " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " survives since then while " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " disappears in " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "C_2" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " because the added pixels connect " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " into a common subset." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21310" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 72, + 274, + 132 + ], + "blocks": [ + { + "bbox": [ + 78, + 72, + 274, + 132 + ], + "lines": [ + { + "bbox": [ + 78, + 72, + 274, + 132 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 274, + 132 + ], + "type": "image", + "image_path": "411528c836fcc76276cab4b88943bd4b0242fc91875dfdbcdba76c3f7d1fae3c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 133, + 284, + 259 + ], + "blocks": [ + { + "bbox": [ + 50, + 133, + 284, + 259 + ], + "lines": [ + { + "bbox": [ + 50, + 133, + 284, + 259 + ], + "spans": [ + { + "bbox": [ + 50, + 133, + 284, + 259 + ], + "type": "image", + "image_path": "0df07790ea3c4adf4b72eaa9ebe438d5122c16940531b0167a10f733ac7c8256.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 287, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 329 + ], + "type": "text", + "content": "Figure 4. The binary images obtained by taking different threshold " + }, + { + "bbox": [ + 46, + 262, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 262, + 287, + 329 + ], + "type": "text", + "content": " of the boundary probability map and corresponding persistence barcode for topology structures. The red letters indicate connected components " + }, + { + "bbox": [ + 46, + 262, + 287, + 329 + ], + "type": "inline_equation", + "content": "(d = 0)" + }, + { + "bbox": [ + 46, + 262, + 287, + 329 + ], + "type": "text", + "content": ", which is signified by the blue solid bars in the persistence barcode map. The golden letters indicate holes " + }, + { + "bbox": [ + 46, + 262, + 287, + 329 + ], + "type": "inline_equation", + "content": "(d = 1)" + }, + { + "bbox": [ + 46, + 262, + 287, + 329 + ], + "type": "text", + "content": ", which are signified by the blue open bars." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "text", + "content": "Then the persistence of " + }, + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "inline_equation", + "content": "S_{ij}^{d}" + }, + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "text", + "content": " is measured by how long it survives in the sequence of combinations, which can be expressed through the persistence barcode whose length represents the corresponding persistence, Fig. 4. The longer the persistence barcode, the more robust the subset. Thereby, the persistence " + }, + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "inline_equation", + "content": "P_{ij}^{d}" + }, + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "inline_equation", + "content": "S_{ij}^{d}" + }, + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "text", + "content": " can be calculated as the difference between the threshold values at which " + }, + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "inline_equation", + "content": "S_{ij}^{d}" + }, + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "text", + "content": " first appears " + }, + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "inline_equation", + "content": "(p_{birth}(S_{ij}^{d}))" + }, + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "text", + "content": " and when it disappears " + }, + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "inline_equation", + "content": "(p_{death}(S_{ij}^{d}))" + }, + { + "bbox": [ + 46, + 341, + 287, + 441 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 448, + 287, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 448, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 99, + 448, + 287, + 463 + ], + "type": "interline_equation", + "content": "P _ {i j} ^ {d} = p _ {\\text {b i r t h}} \\left(S _ {i j} ^ {d}\\right) - p _ {\\text {d e a t h}} \\left(S _ {i j} ^ {d}\\right). \\tag {5}", + "image_path": "22aab5b1012cbb06ff01152c242e4ec349f31be4d0f3108ccd5d74678528514a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 466, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 287, + 491 + ], + "type": "text", + "content": "Given prediction probabilities with K classes of a 2D image, the persistence encoding " + }, + { + "bbox": [ + 47, + 466, + 287, + 491 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 47, + 466, + 287, + 491 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 497, + 287, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 287, + 512 + ], + "type": "interline_equation", + "content": "\\mathbf {P} = \\left\\{P _ {i j} ^ {d} \\mid i = 1, 2, \\dots , K; d = 0, 1; j = 1, 2, \\dots , J _ {d} \\right\\}. \\tag {6}", + "image_path": "ddb91dc33f03901e0218c851dcfc045c10f68d3a0076dfd6456bf7191ecbd4c9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 519, + 287, + 555 + ], + "type": "text", + "content": "Note that since ground truth can be regarded as a probability map containing only binary values of 0 and 1, the persistence encoding of it is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 561, + 287, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 561, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 54, + 561, + 287, + 576 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = \\left\\{Y _ {i j} ^ {d} \\mid i = 1, 2, \\dots , K; d = 0, 1; j = 1, 2, \\dots , B _ {i} ^ {d} \\right\\}, \\tag {7}", + "image_path": "c9148c8a0f293448654cd23adf4337d10a8bb77a49abaf1d1534e9387c8524ba.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 584, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 287, + 621 + ], + "type": "text", + "content": "where the value of " + }, + { + "bbox": [ + 47, + 584, + 287, + 621 + ], + "type": "inline_equation", + "content": "Y_{ij}^{d}" + }, + { + "bbox": [ + 47, + 584, + 287, + 621 + ], + "type": "text", + "content": " is 1. As predicted probability maps typically contain more topological structures than those in ground truth, we assume that " + }, + { + "bbox": [ + 47, + 584, + 287, + 621 + ], + "type": "inline_equation", + "content": "J_{d} > B_{i}^{d}" + }, + { + "bbox": [ + 47, + 584, + 287, + 621 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 635, + 172, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 172, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 172, + 647 + ], + "type": "text", + "content": "3.3.2 Topology-Aware Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": "Based on persistence encoding, a better prediction should meet the condition that the subsets of " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": " consistent with ground truth have greater persistence. We match the top " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "B_i^d" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": " persistence barcode " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "P_{ij}^d" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": " with GT and regard the remaining codes as unmatched. Our purpose is to reduce the error" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "between the persistence encoding of prediction and ground truth, so as to make the topology correctness of prediction more robust. Therefore, the difference between them can be calculated in two parts:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 364, + 127, + 545, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 127, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 364, + 127, + 545, + 169 + ], + "type": "interline_equation", + "content": "M _ {i} = \\sum_ {d \\in \\{0, 1 \\}} \\sum_ {j = 1} ^ {B _ {i} ^ {d}} \\left(Y _ {i j} ^ {d} - P _ {i j} ^ {d}\\right), \\tag {8}", + "image_path": "2a583d44c7d0ee043fbd49ab0876cac06c15d1ec3b16e4953dfa859ffb2ae588.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 167, + 479, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 167, + 479, + 202 + ], + "spans": [ + { + "bbox": [ + 365, + 167, + 479, + 202 + ], + "type": "interline_equation", + "content": "U _ {i} = \\sum_ {d \\in \\{0, 1 \\}} \\sum_ {j = B _ {i} ^ {d} + 1} ^ {J _ {d}} (P _ {i j} ^ {d}),", + "image_path": "a99f0536caaadb4e5f79142f7c8d4ef7e7e32d86c6bc679f82e1ad09c2904c5b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 212, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 212, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 306, + 212, + 545, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 212, + 545, + 247 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 306, + 212, + 545, + 247 + ], + "type": "text", + "content": " measures the length difference between the matched barcode and GT barcode for the " + }, + { + "bbox": [ + 306, + 212, + 545, + 247 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 306, + 212, + 545, + 247 + ], + "type": "text", + "content": "th class, and " + }, + { + "bbox": [ + 306, + 212, + 545, + 247 + ], + "type": "inline_equation", + "content": "U_{i}" + }, + { + "bbox": [ + 306, + 212, + 545, + 247 + ], + "type": "text", + "content": " measures the length of the unmatched barcode." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 248, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 248, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 306, + 248, + 545, + 272 + ], + "type": "text", + "content": "Given prediction probabilities with " + }, + { + "bbox": [ + 306, + 248, + 545, + 272 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 306, + 248, + 545, + 272 + ], + "type": "text", + "content": " classes of a 2D image, the topology-aware loss " + }, + { + "bbox": [ + 306, + 248, + 545, + 272 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{topo}" + }, + { + "bbox": [ + 306, + 248, + 545, + 272 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 368, + 277, + 545, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 277, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 368, + 277, + 545, + 311 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t o p o} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\left(M _ {i} + U _ {i}\\right). \\tag {9}", + "image_path": "dfec65cb69a158d2521ba3b935cf851cd0e1b6391a8f5bf2731d2f3a52dbec6e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 316, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 388 + ], + "type": "text", + "content": "Minimizing this loss is equivalent to maximizing the persistence of topology features that are consistent with ground truth, and minimizing the persistence of undesired features. This provides a structure perspective to optimize the model, which overcomes the limitation of pixel-wise loss and achieves superior nucleus segmentation performance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 396, + 525, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 525, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 525, + 408 + ], + "type": "text", + "content": "3.4. Adaptive Topology-Aware Selection (ATS)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 414, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 414, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 414, + 545, + 449 + ], + "type": "text", + "content": "ATS aims at efficiently optimizing images with topology-important regions by providing image- and patch-level selection for loss calculation, Fig. 5." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 450, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 581 + ], + "type": "text", + "content": "At the image level, we calculate the topology error between prediction and GT, then selects and stores images with an adaptive threshold, Fig. 5 (Left). If the topology error exceeds the threshold, it indicates that the overall structure of the image is significantly different from ground truth, then the image is preferentially selected and stored for further learning. On the contrary, if the topology error does not exceed the threshold, it will be dropped out in the next iteration. The adaptive threshold is determined based on the current number of iterations and topology error rate. The more iterations, the higher the adaptive threshold." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 689 + ], + "type": "text", + "content": "At the patch level, we split the pipeline into three main parts, Fig. 5 (Right). First, we divide the original image into hierarchical patches of varying scales. Then the topology error between predicted patches and corresponding ground truth is calculated. Moreover, a sorted list of patches is built based on these. According to their topology errors, an adaptive screen method is adopted. This allows the calculation of topology-aware loss on selected patches, which drives the model to focus on regions with more topology errors." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "In more detail, the Betti error is calculated by comparing the difference in the number of disjoint pixel sets of each" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21311" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 542, + 232 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 542, + 232 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 542, + 232 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 542, + 232 + ], + "type": "image", + "image_path": "a3ebc062e8aa456b6df9b73b400a5b05c634be81c91e1a2eeeba8a4372aab33c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 546, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 546, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 546, + 258 + ], + "type": "text", + "content": "Figure 5. The pipeline of adaptive topology-aware selection module (ATS). In the topological Error Calculation and Sort and Adaptive Selection, the darker the square, the larger the Betti error." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 269, + 288, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 288, + 318 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 288, + 318 + ], + "type": "text", + "content": "class between the prediction and GT, Fig. 5 (Bottom right). Note that the Betti error calculation used in the selection at the image and patch levels is the same. The difference is that one uses the whole image and the other uses the patch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 328, + 123, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 123, + 341 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 123, + 341 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 347, + 174, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 174, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 174, + 361 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 365, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 593 + ], + "type": "text", + "content": "Datasets. I) MoNuSeg [17] contains 30 H&E-stained pathological images of size " + }, + { + "bbox": [ + 46, + 365, + 287, + 593 + ], + "type": "inline_equation", + "content": "1000 \\times 1000" + }, + { + "bbox": [ + 46, + 365, + 287, + 593 + ], + "type": "text", + "content": " from seven organs, including a total of 21,623 nuclei labeled without distinction between various categories. According to [17], we divide 30 images into three sets: 12 for training, 4 for validation, and 14 for testing (combining the same organ and different organ test sets). II) CPM17 [27] contains 64 H&E stained histopathology images with 7,570 annotated nuclear boundaries. It is from the MICCAI 2017 Digital Pathology Challenge [27] and images on two different scales: " + }, + { + "bbox": [ + 46, + 365, + 287, + 593 + ], + "type": "inline_equation", + "content": "500 \\times 500" + }, + { + "bbox": [ + 46, + 365, + 287, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 365, + 287, + 593 + ], + "type": "inline_equation", + "content": "600 \\times 600" + }, + { + "bbox": [ + 46, + 365, + 287, + 593 + ], + "type": "text", + "content": ". This is also a dataset for single-class segmentation task. III) PBNulei is a nuclear segmentation dataset from Peking University Affiliated Hospital, marked by professional pathologists. It contains 106 H&E stained histopathology images of " + }, + { + "bbox": [ + 46, + 365, + 287, + 593 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 365, + 287, + 593 + ], + "type": "text", + "content": " sizes, with a total of 36,659 nuclei, 78 images for training and 28 for testing. The dataset is divided into five categories, namely carcinoma, epithelial, inflammatory, mesenchymal, and myoepithelium." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 653 + ], + "type": "text", + "content": "Evaluation Metrics. We use three instance-level evaluation metrics to measure the instance segmentation performance of the comparison models, which are: mean Panoptic Quality (PQ), the Aggregated Jaccard Index (AJI), and average Hausdorff distance (Haus)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 689 + ], + "type": "text", + "content": "Implementation details. The proposed TopoSeg is trained and tested using the open-source software library Pytorch 1.8.0 on NVIDIA Tesla V100 with CUDA 10.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "The MoNuSeg dataset has a relatively large original image size and the CPM17 dataset has an inconsistent original" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 269, + 545, + 357 + ], + "blocks": [ + { + "bbox": [ + 309, + 269, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 309, + 269, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 309, + 269, + 545, + 357 + ], + "type": "image", + "image_path": "bb4dba25e6c96288524d3b93dbb08b6934865142392501d7f171e588a8c3e646.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 359, + 545, + 459 + ], + "blocks": [ + { + "bbox": [ + 308, + 359, + 545, + 459 + ], + "lines": [ + { + "bbox": [ + 308, + 359, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 359, + 545, + 459 + ], + "type": "image", + "image_path": "be6c2c2a2701b4164a4664aa572f25ec156feda52202e78f6e0293d8f41a37fa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 462, + 546, + 507 + ], + "lines": [ + { + "bbox": [ + 305, + 462, + 546, + 507 + ], + "spans": [ + { + "bbox": [ + 305, + 462, + 546, + 507 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison before and after adding the topology-aware loss. (a) the probability map obtained by the model just using pixel-wise loss; (b) the probability map obtained by the model using pixel-wise loss and topology-aware loss." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 307, + 519, + 552, + 622 + ], + "blocks": [ + { + "bbox": [ + 307, + 519, + 552, + 622 + ], + "lines": [ + { + "bbox": [ + 307, + 519, + 552, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 519, + 552, + 622 + ], + "type": "table", + "html": "
DatasetBackboneLossPQ ↑AJI ↑Haus ↓
pixeltopo
MoNuSegU-Net✓0.5810.57912.80
U-Net✓✓0.5970.6199.647
CPMU-Net✓0.6250.6667.740
U-Net✓✓0.6620.7185.961
PBNuleiU-Net✓0.4260.3862.770
U-Net✓✓0.4250.4161.953
", + "image_path": "6390d9900a7df4bb5a2b6cdd2d787b5c264b2510c9de07fcd922d67c434de1e8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 629, + 545, + 663 + ], + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 663 + ], + "type": "text", + "content": "Table 1. Comparison of instance segmentation results obtained using only pixel-wise loss and adding topology-aware loss based on pixel-wise loss in three datasets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "image size, so it is divided into multiple patches, thus allowing the network to be uniform and fully trained. As a result, the MoNuSeg dataset generates 192 training patches, the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21312" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "CPM17 dataset generates 512 training patches. In the training phase, we apply a series of data augmentation methods, which include random flip, random rotation, random color, random crop, and Gaussian blur. The training epoch is set as 300, and the initial learning rate is set as " + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 141, + 146, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 141, + 146, + 153 + ], + "spans": [ + { + "bbox": [ + 47, + 141, + 146, + 153 + ], + "type": "text", + "content": "4.2. Ablation studies" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 159, + 287, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 159, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 287, + 232 + ], + "type": "text", + "content": "We conducted ablation studies using three datasets, MoNuSeg, CPM17, and PBNulei, to validate the effectiveness of the proposed TopoSeg. We selected U-Net as the baseline segmentation method, and we improved the U-Net by adding two output branches: a semantic prediction branch and a three-class prediction branch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 232, + 288, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 288, + 316 + ], + "type": "text", + "content": "Topology-Aware Loss. When only pixel-wise losses are present, the model's performance on the MoNuSeg dataset is 0.581 PQ, 0.579 AJI, and 12.8 Haus, Table 1. After incorporating topological awareness losses, the model's PQ, AJI, and Haus on the MoNuSeg dataset are 0.597, 0.619, and 9.647, respectively, which have increased by " + }, + { + "bbox": [ + 46, + 232, + 288, + 316 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 46, + 232, + 288, + 316 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 232, + 288, + 316 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 46, + 232, + 288, + 316 + ], + "type": "text", + "content": ", and 3.153 compared to the original results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "text", + "content": "Compared to the performance obtained by using only pixel-wise loss, the model trained with the addition of pixelwise and topological loss has increased by " + }, + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "text", + "content": " PQ, " + }, + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "text", + "content": " AJI, and 1.779 Haus respectively on the CPM17 dataset and by " + }, + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "inline_equation", + "content": "3.0\\%" + }, + { + "bbox": [ + 46, + 316, + 288, + 388 + ], + "type": "text", + "content": " AJI, and 0.817 Haus respectively on the PBNulei dataset compared to the original results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 388, + 288, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 288, + 520 + ], + "type": "text", + "content": "Fig. 6 makes a qualitative comparison between the boundary probability map obtained before and after adding the topology-aware loss and its corresponding persistence barcode. It shows that the corresponding structure obtained after adding topology-aware loss is more persistent. As an example, when " + }, + { + "bbox": [ + 46, + 388, + 288, + 520 + ], + "type": "inline_equation", + "content": "p = 0.5" + }, + { + "bbox": [ + 46, + 388, + 288, + 520 + ], + "type": "text", + "content": ", three instances are correctly segmented, but when only pixel-wise loss is used, one instance can not be predicted. This indicates that adding topology-aware loss on the basis of pixel-wise loss helps the model learn object structure information, thus more accurately segmenting nuclear instances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 288, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 592 + ], + "type": "text", + "content": "Adaptive Image Selection. We set the threshold of " + }, + { + "bbox": [ + 46, + 521, + 288, + 592 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 521, + 288, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 521, + 288, + 592 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 521, + 288, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 521, + 288, + 592 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 521, + 288, + 592 + ], + "type": "text", + "content": " and the adaptive threshold for the comparisons. According to the threshold, select the image with Betti error greater than the threshold as the new training dataset, and ensure that the number of training data obtained by different thresholds is the same." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 677 + ], + "type": "text", + "content": "In Table 2, we compare the performance before and after image selection. The comparison results show that the segmentation results obtained after image selection are better than those without image selection. This proves that image selection promotes the model to focus on learning images with more structural prediction errors and improves the instance segmentation performance of the model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Compared to training with data selected with a threshold of " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": ", training with data selected with a threshold of " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " yielded more competitive model performance," + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 72, + 558, + 270 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 558, + 270 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 558, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 558, + 270 + ], + "type": "table", + "html": "
DatasetsBackboneSimgPQ↑AJI↑Haus↓
MoNuSegU-Net100%0.5970.6199.647
75%0.6040.6269.358
50%0.6060.6289.328
25%0.6020.6259.452
Adaptive0.6100.6329.113
CPMU-Net100%0.6620.7185.961
75%0.6650.7235.835
50%0.6700.7245.887
25%0.6720.7235.847
Adaptive0.6790.7315.805
PBNuleiU-Net100%0.4250.4161.953
75%0.4270.4192.014
50%0.4190.4142.028
25%0.4290.4182.009
Adaptive0.4460.4351.931
", + "image_path": "bf304f5f6a203394cdbc555051c544d05cdf3cdf5ffaadded89b453db09d239d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 361, + 545, + 523 + ], + "blocks": [ + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "lines": [ + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "text", + "content": "Table 2. Comparison results of the models obtained by selecting images based on different thresholds. Here, " + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "text", + "content": " represents the use of raw data without any image selection operation. " + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "text", + "content": " represent selecting images with threshold values greater than " + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 305, + 278, + 547, + 355 + ], + "type": "text", + "content": ", respectively, as new input data. Adaptive represents the proposed adaptive image selection method." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 361, + 545, + 523 + ], + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 523 + ], + "type": "table", + "html": "
DatasetsBackbone\\( S_{patch} \\)PQ↑AJI↑Haus ↓
MoNuSegU-NetWhole0.5970.6199.647
Non-hierc0.6090.62810.063
Hierarchical0.6130.6398.842
Adaptive0.6160.6438.787
CPMU-NetWhole0.6620.7185.961
Non-hierc0.6750.7296.108
Hierarchical0.6620.7335.510
Adaptive0.6780.7415.605
PBNuleiU-NetWhole0.4250.4161.953
Non-hierc0.4980.4792.320
Hierarchical0.5140.4992.011
Adaptive0.5200.5042.185
", + "image_path": "5703239aaa07b4d30e3e8c678e937af3937c3c4df7016d5959db0fc8086bf335.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 532, + 545, + 598 + ], + "lines": [ + { + "bbox": [ + 305, + 532, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 305, + 532, + 545, + 598 + ], + "type": "text", + "content": "Table 3. Ablation studies on different patch selection methods. " + }, + { + "bbox": [ + 305, + 532, + 545, + 598 + ], + "type": "inline_equation", + "content": "S_{patch}" + }, + { + "bbox": [ + 305, + 532, + 545, + 598 + ], + "type": "text", + "content": " denotes the method of dividing the image, \"Whole\" denotes the whole image for calculation, \"Non-hierc\" means non-hierarchical and refers to single size patch division, \"Hierarchical\" refers to multi-scale patch division, and \"Adaptive\" refers to the adaptive approach of dividing the patch." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 615, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 615, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 305, + 615, + 545, + 652 + ], + "type": "text", + "content": "indicating that overly high or low thresholds can limit the learning of the model. Therefore, selecting a single threshold for image selection has limitations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 653, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 715 + ], + "type": "text", + "content": "Compared to the three fixed threshold image filtering methods, our method has the ability to flexibly select input data. As shown in Table 2, the adaptive image selection approach achieved superior segmentation performance. In the MoNuSeg dataset, the performance metrics PQ, AJI," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21313" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 545, + 403 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 545, + 403 + ], + "type": "image", + "image_path": "85994c6154f84a21905c204e932f9161284a922d443813a83480e32a9d1d7ecc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 408, + 546, + 441 + ], + "lines": [ + { + "bbox": [ + 46, + 408, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 546, + 441 + ], + "type": "text", + "content": "Figure 7. Visualization of instance segmentation results with different methods. (a) Input image; (b) Ground Truth; (c) U-Net [25]; (d) Micro-Net [24]; (e) FullNet [23]; (f) Hover-Net [9]; (g) CDNet [11], (h) Our TopoSeg. Different colors indicate different instances in the images. The white rectangles are drawn for a clear comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "text", + "content": "and Haus improved by " + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "inline_equation", + "content": "0.534" + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "text", + "content": ", respectively, compared to the original results. Similarly, compared with the original results, the performance of the CPM17 dataset is improved by " + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "text", + "content": " PQ, " + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "text", + "content": " AJI and " + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "inline_equation", + "content": "0.16" + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "text", + "content": " Haus respectively, and the performance of the PBNulei dataset is improved by " + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "text", + "content": " PQ, " + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "text", + "content": " AJI and " + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "inline_equation", + "content": "0.022" + }, + { + "bbox": [ + 46, + 452, + 287, + 525 + ], + "type": "text", + "content": " Haus, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 529, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 589 + ], + "type": "text", + "content": "This reflects that, compared with the three fixed threshold image selection methods, our adaptive image selection approach has the ability to select input data more flexibly, thus promoting the model to obtain more correct instance segmentation results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Adaptive Hierarchical Patch Selection. As shown in Table 3, we compare the performance of the topology loss calculation after patch selection with that of the whole image. First of all, we verify the effect of patch selection on segmentation performance. Comparing the first row (Whole) and the second row (Non-hierc) in each dataset in Table 3, it is concluded that the segmentation performance of topology loss calculation after selecting patch is better than that of using the whole image on both PQ and AJI. This shows that by selecting the appropriate patch, the model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 453, + 544, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 544, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 544, + 477 + ], + "type": "text", + "content": "pays more attention to the regions that need structural optimization, so as to improve the segmentation performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 596 + ], + "type": "text", + "content": "Then, we verified the effect of using hierarchical patches with different scales for calculating topological loss in instance segmentation. We compared the performance of the third row (Hierarchical) and the second row (Non-hierc) within each dataset in Table 3 and found that hierarchical patches outperformed single-layer patches. This also shows that selecting the appropriate region that needs structural optimization helps the model better segment instances." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "Table 3 shows that compared with the topology loss calculated by the whole image, the segmentation performance obtained by adaptive hierarchical selection of patch is significantly improved. In the MoNuSeg dataset, PQ increased by " + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": ", AJI increased by " + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": ", Haus increased by 0.86. In the CPM dataset, PQ increased by " + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": ", AJI increased by " + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": ", Haus increased by 0.356. In the PBNulei dataset, PQ increased by " + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": ", AJI increased by " + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "inline_equation", + "content": "8.8\\%" + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21314" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 72, + 288, + 222 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 222 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 222 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 222 + ], + "type": "table", + "html": "
MethodMoNuSegCPM17PBNuclei
PQ↑AJI↑PQ↑AJI↑PQ↑AJI↑
U-Net [25]0.5810.5790.6250.6660.3550.312
Mask-RCNN [13]0.5090.5460.6740.684--
DCAN [2]0.4920.5250.5450.5610.2430.249
DIST [21]0.4430.5590.5040.6160.3250.301
Micro-Net [24]0.5190.560.6610.6680.3060.287
Full-Net [23]-0.604-0.6610.3850.378
PFF-Net [18]0.5870.611----
Hover-Net [9]0.5970.6180.6970.7050.4160.363
CDNet [11]-0.637-0.7330.4130.381
Ours0.6250.6430.7050.7560.5280.509
", + "image_path": "7a2e06e1de18ae690c9aba05e468486187bb1583c6a785bf81f2653ed82d1acf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 226, + 263, + 237 + ], + "lines": [ + { + "bbox": [ + 47, + 226, + 263, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 226, + 263, + 237 + ], + "type": "text", + "content": "Table 4. Performance comparisons on three nuclei datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 248, + 191, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 191, + 261 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 191, + 261 + ], + "type": "text", + "content": "4.3. Performance Comparison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 268, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 268, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 268, + 287, + 340 + ], + "type": "text", + "content": "In Table 4, the proposed TopoSeg is compared with the state-of-the-art nuclear segmentation model on MoNuSeg and CPM17. Our TopoSeg achieves 0.625 PQ and 0.643 AJI on the MoNuSeg dataset and 0.705 PQ and 0.756 AJI on CPM17 dataset, respectively. The TopoSeg achieves competitive results for nuclei segmentation task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 437 + ], + "type": "text", + "content": "We further carry out a qualitative visualization analysis of our method and different existing methods. Compared to all the other methods in Fig. 7, our TopoSeg achieves better instance segmentation results for nuclear instance segmentation, especially for the correct structure segmentation of dense or overlapping regions. Both quantitative and qualitative results demonstrate that our TopoSeg is competitive and able to segment dense nuclei instances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 451, + 119, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 451, + 119, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 451, + 119, + 464 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 473, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 641 + ], + "type": "text", + "content": "This paper proposes a novel topology-aware segmentation approach, termed TopoSeg, for nuclear instance segmentation tasks to address the challenge of touching and overlapping nuclear segmentation. We develop a TAM, which contains a persistence encoding method that accounts for topology information in predicted probability maps, and a topology-aware loss that is flexible and suitable for topology optimization. We further improve it by applying an ATS, which enforces the model to focus on regions with high topology errors. The experimental results in three nuclei datasets show that TopoSeg has achieved superior performance. As a general method of topology awareness, TopoSeg provides fresh insight into the instance segmentation problem of small and dense objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Acknowledgments This work was supported in part by the National Key R&D Program of China (No. 2022ZD0118201), Natural Science Foundation of China (No. 61972217, 32071459, 62176249, 62006133, 62271465)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 365, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 91, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 311, + 91, + 545, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 91, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 311, + 91, + 545, + 145 + ], + "type": "text", + "content": "[1] Nick Byrne, James R Clough, Israel Valverde, Giovanni Montana, and Andrew P King. A persistent homology-based topological loss for cnn-based multiclass segmentation of cmr. IEEE Transactions on Medical Imaging, 42(1):3-14, 2022. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 148, + 547, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 148, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 311, + 148, + 547, + 192 + ], + "type": "text", + "content": "[2] Hao Chen, Xiaojuan Qi, Lequan Yu, Qi Dou, Jing Qin, and Pheng-Ann Heng. Dcan: Deep contour-aware networks for object instance segmentation from histology images. Medical Image Analysis, 36:135–146, 2017. 2, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 194, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 194, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 311, + 194, + 545, + 237 + ], + "type": "text", + "content": "[3] Shengcong Chen, Changxing Ding, Minfeng Liu, Jun Cheng, and Dacheng Tao. CPP-net: Context-aware polygon proposal network for nucleus segmentation. IEEE Transactions on Image Processing, 32:980-994, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 239, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 239, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 311, + 239, + 545, + 303 + ], + "type": "text", + "content": "[4] James R Clough, Nicholas Byrne, Ilkay Oksuz, Veronika A Zimmer, Julia A Schnabel, and Andrew P King. A topological loss function for deep-learning based image segmentation using persistent homology. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):8766-8778, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 306, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 306, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 311, + 306, + 545, + 361 + ], + "type": "text", + "content": "[5] James R Clough, Ilkay Oksuz, Nicholas Byrne, Julia A Schnabel, and Andrew P King. Explicit topological priors for deep-learning based image segmentation using persistent homology. In International Conference on Information Processing in Medical Imaging, pages 16-28, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 363, + 547, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 363, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 311, + 363, + 547, + 407 + ], + "type": "text", + "content": "[6] Herbert Edelsbrunner, David Letscher, and Afra Zomorodian. Topological persistence and simplification. In Proceedings of the 41st Annual Symposium on Foundations of Computer Science, pages 454-463, 2000. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 409, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 409, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 311, + 409, + 545, + 430 + ], + "type": "text", + "content": "[7] Robin Forman. Morse theory for cell complexes. Advances in Mathematics, 134(1):90-145, 1998. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 432, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 432, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 311, + 432, + 545, + 464 + ], + "type": "text", + "content": "[8] Robin Forman. A user's guide to discrete Morse theory. Séminaire Lotharingien de Combinatoire [electronic only], 48:1-35, 2002. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 467, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 467, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 311, + 467, + 545, + 521 + ], + "type": "text", + "content": "[9] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:1-15, 2019. 1, 2, 8, 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 523, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 589 + ], + "type": "text", + "content": "[10] Saumya Gupta, Xiaoling Hu, James Kaan, Michael Jin, Muth shipay Mpoy, Katherine Chung, Gagandeep Singh, Mary Saltz, Tahsin Kurc, Joel Saltz, et al. Learning topological interactions for multi-class medical image segmentation. In Proceedings of the European Conference on Computer Vision, pages 701-718, 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 590, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 656 + ], + "type": "text", + "content": "[11] Hongliang He, Zhongyi Huang, Yao Ding, Guoli Song, Lin Wang, Qian Ren, Pengxu Wei, Zhiqiang Gao, and Jie Chen. Cdnet: Centripetal direction network for nuclear instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4026-4035, 2021. 2, 8, 9" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[12] Hongliang He, Chi Zhang, Jie Chen, Ruizhe Geng, Luyang Chen, Yongsheng Liang, Yanchang Lu, Jihua Wu, and Yongjie Xu. A hybrid-attention nested unet for nuclear segmentation in histopathological images. Frontiers in Molecular Biosciences, 8:1–9, 2021. 1, 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21315" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[13] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2961-2969, 2017. 9" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 117, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 117, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 117, + 287, + 160 + ], + "type": "text", + "content": "[14] Xiaoling Hu, Fuxin Li, Dimitris Samaras, and Chao Chen. Topology-preserving deep image segmentation. Advances in Neural Information Processing Systems, 32:5658-5669, 2019. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 195 + ], + "type": "text", + "content": "[15] Xiaoling Hu, Dimitris Samaras, and Chao Chen. Learning probabilistic topological representations using discrete Morse theory. arXiv preprint arXiv:2206.01742, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 196, + 287, + 239 + ], + "type": "text", + "content": "[16] Xiaoling Hu, Yusu Wang, Li Fuxin, Dimitris Samaras, and Chao Chen. Topology-aware segmentation using discrete Morse theory. In Proceedings of the International Conference on Learning Representation, pages 1-19, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "text", + "content": "[17] Neeraj Kumar, Ruchika Verma, Sanuj Sharma, Surabhi Bhargava, Abhishek Vahadane, and Amit Sethi. A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Transactions on Medical Imaging, 36(7):1550-1560, 2017. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 348 + ], + "type": "text", + "content": "[18] Dongnan Liu, Donghao Zhang, Yang Song, Heng Huang, and Weidong Cai. Panoptic feature fusion net: A novel instance segmentation paradigm for biomedical and biological images. IEEE Transactions on Image Processing, 30:2045-2059, 2021. 1, 9" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 403 + ], + "type": "text", + "content": "[19] Xiaoming Liu, Zhengsheng Guo, Jun Cao, and Jinshan Tang. Mdc-net: a new convolutional neural network for nucleus segmentation in histopathology images with distance maps and contour information. Computers in Biology and Medicine, 135:1-12, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "text", + "content": "[20] Agata Mosinska, Pablo Marquez-Neila, Mateusz Kozinski, and Pascal Fua. Beyond the pixel-wise loss for topology-aware delineation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3136–3145, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 503 + ], + "type": "text", + "content": "[21] Peter Naylor, Marick Laé, Fabien Reyal, and Thomas Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE Transactions on Medical Imaging, 38(2):448-459, 2018. 1, 2, 9" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 504, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 548 + ], + "type": "text", + "content": "[22] H. Pinckaers, W. Bulten, J. van der Laak, and G. Litjens. Detection of prostate cancer in whole-slide images through end-to-end training with image-level labels. IEEE Transactions on Medical Imaging, 40(7):1817-1826, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 624 + ], + "type": "text", + "content": "[23] Hui Qu, Zhennan Yan, Gregory M Riedlinger, Subhajyoti De, and Dimitris N Metaxas. Improving nuclei/gland instance segmentation in histopathology images by full resolution neural network and spatial constrained loss. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 378-386, 2019. 1, 2, 8, 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "text", + "content": "[24] Shan E Ahmed Raza, Linda Cheung, Muhammad Shaban, Simon Graham, David Epstein, Stella Pelengaris, Michael Khan, and Nasir M Rajpoot. Micro-net: A unified model for segmentation of various objects in microscopy images. Medical Image Analysis, 52:160-173, 2019. 2, 8, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the International Conference on" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 316 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 1, 2, 8, 9" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 172 + ], + "type": "text", + "content": "[26] Suprosanna Shit, Johannes C Paetzold, Anjany Sekuboyina, Ivan Ezhov, Alexander Unger, Andrey Zhylka, Josien PW Pluim, Ulrich Bauer, and Bjoern H Menze. cldice-a novel topology-preserving loss function for tubular structure segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16560-16569, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 174, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 250 + ], + "type": "text", + "content": "[27] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in Bioengineering and Biotechnology, 53:1-15, 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 251, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 316 + ], + "type": "text", + "content": "[28] Yanning Zhou, Omer Fahri Onder, Qi Dou, Efstratios Tsougenis, Hao Chen, and Pheng-Ann Heng. Cia-net: Robust nuclei instance segmentation with contour-aware information aggregation. In International Conference on Information Processing in Medical Imaging, pages 682–693, 2019. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21316" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/49fdb57e-3f25-4f0f-85ac-577ce80d4f48_content_list.json b/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/49fdb57e-3f25-4f0f-85ac-577ce80d4f48_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ddd1dd9e4d80855027d03c03b0c9a1dbee9fbd8e --- /dev/null +++ b/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/49fdb57e-3f25-4f0f-85ac-577ce80d4f48_content_list.json @@ -0,0 +1,1381 @@ +[ + { + "type": "text", + "text": "Total-Recon: Deformable Scene Reconstruction for Embodied View Synthesis", + "text_level": 1, + "bbox": [ + 94, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chonghyuk Song Gengshan Yang Kangle Deng Jun-Yan Zhu Deva Ramanan Carnegie Mellon University", + "bbox": [ + 158, + 180, + 810, + 219 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/baeec2e9755f02e3f6586435c235cbe25ffdfdc9edf954fd3dac8ff708f9f7d2.jpg", + "image_caption": [ + "Figure 1: Embodied View Synthesis. Given a long video of deformable objects captured by a handheld RGBD sensor, Total-Recon renders the scene from novel camera trajectories derived from the in-scene motion of actors: (1) egocentric cameras that simulate the point-of-view of a target actor (such as the pet) and (2) 3rd-person (or pet) cameras that follow the actor from behind. Our method also enables (3) 3D video filters that attach virtual 3D assets to the actor. Total-Recon achieves this by reconstructing the geometry, appearance, root-body- and articulated motion of each deformable object in the scene and the background." + ], + "image_footnote": [], + "bbox": [ + 78, + 220, + 893, + 409 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 508, + 313, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We explore the task of embodied view synthesis from monocular videos of deformable scenes. Given a minute-long RGBD video of people interacting with their pets, we render the scene from novel camera trajectories derived from the in-scene motion of actors: (1) egocentric cameras that simulate the point of view of a target actor and (2) 3rd-person cameras that follow the actor. Building such a system requires reconstructing the root-body and articulated motion of every actor, as well as a scene representation that supports free-viewpoint synthesis. Longer videos are more likely to capture the scene from diverse viewpoints (which helps reconstruction) but are also more likely to contain larger motions (which complicates reconstruction). To address these challenges, we present Total-Recon, the first method to photorealistically reconstruct deformable scenes from long monocular RGBD videos. Crucially, to scale to long videos, our method hierarchically decomposes the scene into the background and objects, whose motion is decomposed into carefully initialized root-body motion and local articulations. To quantify such \"in-the-wild\" reconstruction and view synthesis, we collect ground-truth data from a specialized stereo RGBD capture rig for 11 challenging videos, significantly outperforming prior methods.", + "bbox": [ + 73, + 537, + 473, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 508, + 632, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We explore embodied view synthesis, a new class of novel-view synthesis tasks that renders deformable scenes from novel 6-DOF trajectories reconstructed from the in-scene motion of actors: egocentric cameras [45, 7] that simulate the point-of-view of moving actors and 3rd-person-follow cameras [54, 7] that track a moving actor from behind (Figure 1). We focus on everyday scenes of people interacting with their pets, producing renderings from the point-of-view of the person and pet (Figure 1). While such camera trajectories could be manually constructed (e.g., by artists via keyframing), building an automated system is an interesting problem of its own: spatial cognition theory [57] suggests that the ability to visualize behavior from another actor's perspective is necessary for action learning and imitation; in the context of gaming and virtual reality [7, 45], egocentric cameras offer high levels of user immersion, while 3rd-person-follow cameras provide a large field of view that is useful for exploring a user's environment.", + "bbox": [ + 496, + 532, + 893, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Challenges. Building a system for embodied view synthesis is challenging for many reasons. First, to reconstruct everyday-but-interesting content, it needs to process long, monocular captures of multiple interacting actors. How-", + "bbox": [ + 496, + 840, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17671", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e8218f59c1037fb89cfa0dc8f68b116edcc85c26a28a3c8133ea47da4a8eb54.jpg", + "image_caption": [ + "Figure 2: Method Overview. Total-Recon represents the entire scene as a composition of $M$ object-centric neural fields, one for the rigid background and each of the $M - 1$ deformable objects. To render a scene, (1) each object field $j$ is transformed into the camera space with a rigid transformation $\\left(\\mathbf{G}_j^t\\right)^{-1}$ that encodes root-body motion and, for each deformable object, an additional deformation field $\\mathbf{J}_j^{t,\\rightarrow}$ that encodes articulated motion. Next, all (2) posed object fields are combined into a (3) composite field, which is then volume-rendered into (4) color, depth, optical flow, and object silhouettes. Each rendered output defines a reconstruction loss that derives supervision from a monocular RGBD video captured by a moving iPad Pro." + ], + "image_footnote": [], + "bbox": [ + 160, + 89, + 808, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ever, such videos are likely to contain large scene motions, which we demonstrate are difficult to reconstruct with current approaches. Second, it needs to produce a deformable 3D scene representation that supports free-viewpoint synthesis, which also would benefit from long videos likely to capture the scene from diverse viewpoints. Recent approaches have extended Neural Radiance Fields (NeRFs) [28] to deformable scenes, but such work is often limited to rigid-only object motion [18, 33], short videos with limited scene motion [41, 35, 21, 55, 36, 60, 10, 61, 58], or reconstructing single objects as opposed the entire scene [65, 66, 67, 4]. Third, it needs to compute global 6-DOF trajectories of root-bodies and articulated body parts (e.g., head) of multiple actors.", + "bbox": [ + 75, + 445, + 472, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Key Ideas. To address these challenges, we introduce Total-Recon, the first monocular NeRF that enables embodied view synthesis for deformable scenes with large motions. Given a monocular RGBD video, Total-Recon reconstructs the scene as a composition of object-centric representations, which encode the 3D appearance, geometry, and motion of each deformable object and the background. Crucially, Total-Recon hierarchically decomposes scene motion into the motion of individual objects, which itself is decomposed into global root-body movement and the local deformation of articulated body parts. We demonstrate that such decomposition of object motion, along with appropriate initialization of root-body pose, allows reconstruction to scale to longer videos, enabling free-viewpoint synthesis. By reconstructing such motions in a globally-consistent", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "coordinate frame, Total-Recon can generate renderings from egocentric and 3rd-person-follow cameras, as well as static but extreme viewpoints like bird's-eye-views.", + "bbox": [ + 498, + 445, + 893, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evaluation. Due to the difficulty of collecting ground-truth data for embodied view synthesis on in-the-wild videos, we evaluate our method on the proxy task of stereo-view synthesis [35], which compares rendered views to those captured from a stereo pair. To this end, we build a stereo RGBD sensor capture rig for ground-truthing and collect a dataset of 11 long video sequences in various indoor environments, including people interacting with their pets. Total-Recon outperforms the state-of-the-art monocular deformable NeRF methods [36, 60], even when modified to use depth sensor measurements.", + "bbox": [ + 496, + 515, + 893, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. In summary, our contributions are: (1) Total-Recon, a hierarchical 3D representation that models deformable scenes as a composition of object-centric representations, each of which decomposes object motion into its global root-body motion and its local articulations; (2) a system based on Total-Recon for automated embodied view synthesis from casual, minute-long RGBD videos of highly dynamic scenes; (3) a dataset of stereo RGBD videos containing various deformable objects, such as humans and pets, in a host of different background environments. Our code, models, and data can be found at https://andrewsonga.github.io/totalrecon.", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17672", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/049ff2708edfcc24bdac102133780f47771de364cc1d3e7e8d45eb22a6c73dc8.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Comparison to Related Work. Unlike prior work, Total-Recon targets embodied view synthesis of scenes containing humans and pets, requiring the ability to (1) reconstruct entire scenes, (2) model deformable objects, (3) extend beyond humans, (4) recover global 6-DOF trajectories of objects' root-bodies and their articulated parts, (5) process minute-long videos of dynamic scenes, and (6) render extreme views." + ], + "table_body": "
MethodEntire ScenesDeform. ObjectsBeyond HumansGlobal 6-DOF Traj.Long VideosExtreme Views
BANMo [67]X✓✓X✓✓
PNF [18]✓X✓✓XX
NeuMan [16]✓✓X✓XX
SLAHMR [69]X✓X✓XX
HyperNeRF [36]✓✓✓XXX
D2NeRF [60]✓✓✓XXX
DynIBaR [22]✓✓✓X✓X
SUDS [56]✓✓✓X✓X
Ours✓✓✓✓✓✓
", + "bbox": [ + 78, + 88, + 470, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 371, + 218, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Fields. Prior works on Neural Radiance Fields (NeRF) optimize a continuous scene function for novel view synthesis given a set of multi-view images, usually under the assumption of a rigid scene and densely sampled views [28, 26, 27, 23, 14, 59]. DS-NeRF [6] and Dense Depth Priors [44] extend NeRFs to the sparse-view setting by introducing depth as additional supervision. Total-Recon also operates in the sparse-view regime and uses depth supervision to reduce the ambiguities inherent to monocular, multibody, non-rigid reconstruction [62, 34]. Another line of work [18, 33] represents rigidly moving scenes as a composition of multiple object-level NeRFs. Total-Recon also leverages such an object-centric scene representation, but models scenes containing non-rigidly moving objects, such as humans and pets.", + "bbox": [ + 75, + 397, + 468, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deformable NeRFs. Recent approaches extend NeRF to monocular deformable scene reconstruction either by learning an additional function that deforms observed points in the camera space to a time-independent canonical space [41, 35, 55, 36, 60] or explicitly modeling density changes over time [10, 61, 58, 21]. Such methods are typically limited to short videos containing little scene and camera motion. They also perform novel-view synthesis only over small baselines. Total-Recon belongs to the former category of prior monocular deformable NeRFs, but unlike them, our method hierarchically decomposes scene motion into the motion of each object, which is further decomposed into global root-body motion and local articulations. The proposed motion decomposition is what enables embodied view synthesis: it allows Total-Recon to scale to minute-long videos and reconstruct a deformable 3D scene representation that supports free-viewpoint syn", + "bbox": [ + 75, + 643, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "thesis; it also makes it easy to extract an object's root-body motion, the key motion primitive required for 3rd-person-follow view synthesis. Several works have taken different approaches to making non-rigid reconstruction more tractable. One group of work leverages human-specific priors [38, 53, 32, 39, 24, 16, 19, 37] such as human body models (e.g., SMPL), 3D skeletons, or 2D poses to achieve high reconstruction quality. We achieve similar levels of fidelity without relying on such shape priors, allowing Total-Recon to generalize to pets and, by extension, reconstruct human-pet interaction videos. Another body of work [15, 48, 20] achieves high-fidelity scene reconstructions by relying on synchronized multi-view video captured from a specialized camera rig ranging from 8 to 18 static cameras. In contrast, Total-Recon only requires a single video captured from a moving RGBD camera equipped with inertial sensors, which has now become widely accessible in consumer products with the advent of Apple's iPhone and iPad Pro.", + "bbox": [ + 496, + 90, + 890, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reconstruction with RGBD Sensors. Depth sensors represent the third class of attempts to make non-rigid reconstruction more tractable, reducing the need for a predefined shape template. Kinect-fusion [30] creates a real-time system for indoor scene localization and mapping. Dynamic Fusion [29] builds a template-free dense SLAM system for dynamic objects. Later works improve RGBD reconstruction to be able to deal with topology changes [50, 51] and use correspondence matching for registration over large motions [2, 3]. Recent works have incorporated neural implicit representations to reconstruct the surface geometry and 3D motion fields for deformable objects [43, 4] or largescale rigid scenes [1, 42] in isolation. Other works have reconstructed humans alongside small-scale objects and furniture [8, 2], but not the entire background. We aim to go even further by reconstructing the entire scene, which includes the background and multiple deformable targets such as humans and pets; not only do we reconstruct the geometry, but we also recover a radiance field that allows for photorealistic scene rendering from embodied viewpoints and other novel 6-DOF trajectories.", + "bbox": [ + 496, + 382, + 890, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concurrent Work. Concurrent work exhibits a subset of the design choices necessary for embodied view synthesis. SLAHMR [69] reconstructs the geometry and in-scene motion of human actors but not the scene appearance. Nerflets [71] models the appearance, geometry, and motion of each scene element but is limited to rigidly moving objects. RoDynRF [25], NeRF-DS [63], HexPlane [5], and K-planes [9] reconstruct other types of dynamic scene elements, such as deformable or specular objects, but these methods have been demonstrated on only short videos and/or videos containing limited object root-body motion [40, 47, 63, 11]. DynIBaR [22] scales dynamic view synthesis to longer", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17673", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "videos with complex camera and scene motion, and SUDS [56] scales reconstruction to urban-scale dynamic scenes captured from 1.2 million frames. However, neither demonstrates extreme-view synthesis, a prerequisite for rendering embodied views. We summarize and compare prior work to Total-Recon in Table 1.", + "bbox": [ + 75, + 90, + 472, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 194, + 168, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Limitations of Prior Art", + "text_level": 1, + "bbox": [ + 76, + 218, + 299, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The state-of-the-art monocular deformable NeRFs [36, 60] decompose a deformable scene into a rigid, canonical template model and a deformation field $\\mathbf{J}^{t,\\leftarrow}$ that maps the world space $\\mathbf{G}_0^t\\mathbf{X}^t$ to the canonical space $\\mathbf{X}^*$ , where $\\mathbf{G}_0^t$ is the known camera pose at time $t$ , and $\\mathbf{X}^t$ is a camera space point at time $t$ :", + "bbox": [ + 75, + 242, + 468, + 332 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {*} = \\mathcal {W} ^ {t, \\leftarrow} \\left(\\mathbf {X} ^ {t}\\right) = \\mathbf {J} ^ {t, \\leftarrow} \\left(\\mathbf {G} _ {0} ^ {t} \\mathbf {X} ^ {t}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 340, + 468, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In theory, this formulation is sufficient to represent all continuous motion; it performs well on short videos containing near-rigid scenes, as the deformation field only has to learn minute deviations from the template model. However, this motion model is difficult to scale to minute-long videos, which are more likely to contain deformable objects undergoing large translations (e.g., a person walking into another room) and pose changes (e.g., a person sitting down). Here, the deformation field must learn large deviations from the canonical model, significantly complicating optimization.", + "bbox": [ + 75, + 367, + 468, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another critical limitation of HyperNeRF and $\\mathrm{D}^2\\mathrm{NeRF}$ is that they cannot track separate deformable objects and therefore cannot perform 3rd-person-follow view synthesis for scenes with multiple actors.", + "bbox": [ + 75, + 518, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Component Radiance Fields", + "text_level": 1, + "bbox": [ + 76, + 587, + 331, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the limitations of existing monocular deformable NeRFs, we propose Total-Recon, a novel 3D representation that models a deformable scene as a composition of $M$ object-centric neural fields, one for the rigid background and each of the $M - 1$ deformable objects (Figure 2). Crucially, Total-Recon hierarchically decomposes scene motion into the motion of each object, which itself is decomposed into global root-body motion and local articulations. This key design choice scales our method to minute-long videos containing highly dynamic and deformable objects.", + "bbox": [ + 75, + 609, + 470, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Background Radiance Field. We begin by modeling the background environment as a Neural Radiance Field (NeRF) [28]. For a 3D point $\\mathbf{X}^{*} \\in \\mathbb{R}^{3}$ and a viewing direction $\\mathbf{v}^{*}$ in the canonical world space, NeRF defines a color $\\mathbf{c}$ and density $\\sigma$ represented by an MLP. We follow contemporary variants [26] that include a time-specific embedding code $\\omega_{e}^{t}$ to model illumination changes over time", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and model density with as a function of a neural signed distance function (SDF) $\\mathbf{MLP}_{\\sigma}(\\cdot) = \\alpha \\Gamma_{\\beta}(\\mathbf{MLP}_{\\mathrm{SDF}}(\\cdot))$ [68] to encourage the reconstruction of a valid surface:", + "bbox": [ + 496, + 90, + 892, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma = \\mathbf {M L P} _ {\\sigma} \\left(\\mathbf {X} ^ {*}\\right), \\quad \\mathbf {c} ^ {t} = \\mathbf {M L P} _ {\\mathbf {c}} \\left(\\mathbf {X} ^ {*}, \\mathbf {v} ^ {*}, \\omega_ {e} ^ {t}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 147, + 892, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pixel color can then be computed with differentiable volume rendering equations (Section 3.3).", + "bbox": [ + 496, + 176, + 890, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most NeRF methods, including HyperNeRF [36] and $\\mathrm{D}^2\\mathrm{NeRF}$ [60], assume images with known cameras. While our capture devices are equipped with inertial sensors, we find their self-reported camera poses have room for improvement. As such, we also model camera pose as an opti- mizable rigid-body transformation $\\mathbf{G}_0^t\\in SE(3)$ that maps points in a time-specific camera space $\\mathbf{X}^t\\in \\mathbb{R}^3$ to the world space (where we assume homogenous notation):", + "bbox": [ + 496, + 207, + 892, + 327 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {*} = \\mathbf {G} _ {0} ^ {t} \\mathbf {X} ^ {t}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 338, + 890, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deformable Field (for Object $j$ ). We model the deformable radiance field of object $j \\in \\{1, \\dots, M - 1\\}$ with BANMo [67], which consists of a canonical rest shape and time-dependent deformation field. The canonical rest shape is represented by the same formulation described by Equation 2, but now defined in a local object-centric canonical space rather than the world space. BANMo represents object motion with a warping function $\\mathcal{W}_j^{t,\\leftarrow} : \\mathbf{X}^t \\rightarrow \\mathbf{X}_j^*$ that maps the camera space points $\\mathbf{X}^t$ to canonical space points $\\mathbf{X}_j^*$ with a rigid-body transformation $\\mathbf{G}_j^t \\in SE(3)$ and a deformation field $\\mathbf{J}_j^{t,\\leftarrow}$ modeled by linear blend skinning [13]:", + "bbox": [ + 496, + 373, + 892, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {j} ^ {*} = \\mathcal {W} _ {j} ^ {t, \\leftarrow} \\left(\\mathbf {X} ^ {t}\\right) = \\mathbf {J} _ {j} ^ {t, \\leftarrow} \\left(\\mathbf {G} _ {j} ^ {t} \\mathbf {X} ^ {t}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 568, + 890, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that our choice of deformation field differs from the $SE(3)$ -field used in HyperNeRF and $\\mathrm{D}^2\\mathrm{NeRF}$ , which has been shown to produce irregular deformation in the presence of complex scene motion [67]. Intuitively, rigid-body transformation $\\mathbf{G}_j^t$ captures the global root-body pose of object $j$ relative to the camera at time $t$ , while deformation field $\\mathbf{J}_j^{t,\\leftarrow}$ aligns more fine-grained articulations relative to its local canonical space (Figure 2). Explicitly disentangling these two sources of object motion (as opposed to conflating them) enables easier optimization of the deformation field, because local articulations are significantly easier to learn than those modeled relative to the world space (Equation 1). Furthermore, this motion decomposition makes the deformation field invariant to rigid-body transformations of the object. A motion model similar to ours was proposed by ST-NeRF [15], but their model encodes an object's global root-body motion with a 3D axis-aligned bounding box that does not explicitly represent object orientation, a prerequisite for embodied view synthesis from 3rd-person-follow cameras.", + "bbox": [ + 496, + 598, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17674", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/54cc8b99de65c0224c31cc96927559fa548da8e795441d48c61efc1f3a389e8c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Rendered features f at pixel xtCorresponding 3D features fij(Xt)
color帽子(xt)ct( Wj,←(Xt))
flow F( xt, t → t')Πt' (Wj', → (Wj, ←(Xt)) - xt
depth d( xt)[0, 0, 1] · Xt
", + "bbox": [ + 81, + 88, + 459, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2: Rendered 2D features $\\hat{\\mathbf{f}}$ and their corresponding 3D features $\\mathbf{f}_{ij}$ . $\\Pi^{t'}$ denotes the camera intrinsics at time $t'$ .", + "bbox": [ + 76, + 172, + 468, + 203 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As did BANMo, Total-Recon also models a forward warp $\\mathbf{X}_j^t = \\mathcal{W}_j^{t,\\rightarrow}(\\mathbf{X}^*) = (\\mathbf{G}_j^t)^{-1}\\mathbf{J}_j^{t,\\rightarrow}(\\mathbf{X}^*)$ that maps the canonical space to the camera space, which is used to establish the surface correspondences required for egocentric view synthesis and 3D video filters.", + "bbox": [ + 76, + 222, + 470, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Composite Rendering of Multiple Objects", + "text_level": 1, + "bbox": [ + 76, + 309, + 433, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a set of $M$ object representations (the background is treated as an object as well), we use the composite rendering scheme from prior work [31, 52] to combine the outputs of all object representations and volume-render the entire scene. To volumetrically render the image at frame $t$ , we sample multiple points along each camera ray $\\mathbf{v}^t$ . Denoting the $i^{th}$ sample as $\\mathbf{X}_i^t$ , we write the density and color observed at sample $i$ due to object $j$ as:", + "bbox": [ + 75, + 333, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {i j} = \\mathbf {M L P} _ {\\sigma , j} \\left(\\mathbf {X} _ {i j} ^ {*}\\right), \\qquad \\mathbf {c} _ {i j} = \\mathbf {M L P} _ {\\mathbf {c}, j} \\left(\\mathbf {X} _ {i j} ^ {*}, \\mathbf {v} _ {j} ^ {*}, \\boldsymbol {\\omega} _ {e} ^ {t}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 465, + 470, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_{ij}^{*} = \\mathcal{W}_{j}^{t,\\leftarrow}(\\mathbf{X}_{i}^{t})$ and $\\mathbf{v}_j^* = \\mathcal{W}_j^{t,\\leftarrow}(\\mathbf{v}^t)$ are sample $i$ and camera ray $\\mathbf{v}^t$ backward-warped into object $j$ 's canonical space, respectively. The composite density $\\sigma_{i}$ at sample $i$ along the ray is then computed as the sum of each object's density $\\sigma_{ij}$ ; the composite color $\\mathbf{c}_i$ is computed as the weighted sum of each object's color $\\mathbf{c}_{ij}$ , where the weights are the normalized object densities $\\sigma_{ij} / \\sigma_i$ :", + "bbox": [ + 75, + 497, + 468, + 607 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {i} = \\sum_ {j = 0} ^ {M - 1} \\sigma_ {i j}, \\quad \\mathbf {c} _ {i} = \\frac {1}{\\sigma_ {i}} \\sum_ {j = 0} ^ {M - 1} \\sigma_ {i j} \\mathbf {c} _ {i j}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 618, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can then use the standard volume rendering equations to generate an RGB image of the scene, where $N$ is the number of sampled points along camera ray $\\mathbf{v}^t$ , $\\tau_{i}$ is the transmittance, $\\alpha_{i}$ is the alpha value for sample point $i$ and $\\delta_{i}$ is the distance between sample point $i$ and the $(i + 1)$ :", + "bbox": [ + 76, + 674, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {c}} = \\sum_ {i = 1} ^ {N} \\tau_ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\quad \\tau_ {i} = \\prod_ {k = 1} ^ {i - 1} (1 - \\alpha_ {k}), \\quad \\alpha_ {i} = 1 - e ^ {- \\sigma_ {i} \\delta_ {i}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 773, + 462, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rendering Flow, Depth, and Silhouettes. Our composite rendering scheme can be used to render different quantities by replacing the object color $\\mathbf{c}_{ij}$ in Equation 5 with the", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dee609955c7eb87a0593595a6c91eb580240b9426e3498568638cb241961a32e.jpg", + "image_caption": [ + "Figure 3: 6-DOF Trajectories for Embodied View Synthesis. To synthesize embodied views from egocentric or actor-following cameras, Total-Recon reconstructs the entire background, every individual actor in the scene, as well as global 6-DOF trajectories of its root-body and its articulated body parts (e.g. head)." + ], + "image_footnote": [], + "bbox": [ + 501, + 92, + 890, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "appropriately defined 3D feature $\\mathbf{f}_{ij}$ (Table 2) and rendering the resulting composite feature $\\mathbf{f}_i$ . To render occlusion-aware object silhouettes, we follow ObSURF [52] to produce a categorical distribution over the $M$ objects:", + "bbox": [ + 496, + 484, + 890, + 546 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {o}} _ {j} = \\sum_ {i = 1} ^ {N} \\tau_ {i} \\alpha_ {i j}, \\quad \\text {w h e r e} \\quad \\tau_ {i} = \\prod_ {k = 1} ^ {i - 1} (1 - \\alpha_ {k}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 575, + 890, + 616 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = 1 - e ^ {- \\sigma_ {i} \\delta_ {i}}, \\quad \\alpha_ {i j} = 1 - e ^ {- \\sigma_ {i j} \\delta_ {i}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 618, + 890, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Losses. Given a monocular RGBD video, we optimize all parameters in our composite scene representation, which for each of the $M$ objects includes the appearance and shape MLPs $(\\mathbf{MLP}_{c,j}, \\mathbf{MLP}_{\\sigma,j})$ , rigid-body transformations $\\mathbf{G}_j^t$ , and forward, backward deformation fields $\\mathbf{J}_j^{\\leftarrow}$ , $\\mathbf{J}_j^{\\rightarrow}$ . We optimize three reconstruction losses: a color loss $\\mathcal{L}_{\\mathrm{rgb}}$ , a flow loss $\\mathcal{L}_{\\mathrm{flow}}$ , and a depth loss $\\mathcal{L}_{\\mathrm{depth}}$ , where the ground truth color $\\mathbf{c}$ and depth $\\mathbf{d}$ are provided by the RGBD video, and the \"ground truth\" flow $\\mathcal{F}$ is computed by an off-the-shelf network [64]. The model also optimizes a 3D-cycle consistency loss $\\mathcal{L}_{\\mathrm{cyc},j}$ [67] for each deformable object to encourage their forward and backward warps to be consistent, where $\\mathbf{x}^t \\in \\mathbb{R}^2$ denotes pixel location at time $t$ :", + "bbox": [ + 496, + 703, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17675", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r g b}} = \\sum_ {\\mathbf {x} ^ {t}} \\left| \\left| \\mathbf {c} \\left(\\mathbf {x} ^ {t}\\right) - \\hat {\\mathbf {c}} \\left(\\mathbf {x} ^ {t}\\right) \\right| \\right| ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 116, + 468, + 148 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f l o w}} = \\sum_ {\\mathbf {x} ^ {t}} \\left| \\left| \\mathcal {F} \\left(\\mathbf {x} ^ {t}\\right) - \\hat {\\mathcal {F}} \\left(\\mathbf {x} ^ {t}\\right) \\right| \\right| ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 151, + 468, + 183 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e p t h}} = \\sum_ {\\mathbf {x} ^ {t}} \\left| \\left| \\mathbf {d} \\left(\\mathbf {x} ^ {t}\\right) - \\hat {\\mathbf {d}} \\left(\\mathbf {x} ^ {t}\\right) \\right| \\right| ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 186, + 468, + 218 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c y c}, j} = \\sum_ {i} \\tau_ {i} \\alpha_ {i j} \\left\\| \\mathcal {W} _ {j} ^ {t ^ {\\prime}, \\rightarrow} \\left(\\mathcal {W} _ {j} ^ {t}, \\leftarrow (\\mathbf {X} _ {i} ^ {t})\\right) - \\mathbf {X} _ {i} ^ {t} \\right\\| ^ {2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 220, + 468, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Initialization. We initialize the rigid-body transformations of each deformable object $\\mathbf{G}_j^t$ using a pre-trained category-specific PoseNet [67]; we initialize the rigid-body transformation of the background $\\mathbf{G}_0^t$ with the camera poses provided by the iPad Pro.", + "bbox": [ + 75, + 276, + 468, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Embodied View Synthesis and 3D Filters. To enable embodied view synthesis and 3D video filters (Figure 3), we design a simple interface that allows a user to select a point on a target object's surface in its reconstructed canonical mesh, and use its forward warping function $\\mathcal{W}_j^{t,\\rightarrow}:\\mathbf{X}^* \\to \\mathbf{X}^t$ followed by the rigid-body transformation $\\mathbf{G}_0^t$ to place the egocentric camera (or virtual 3D asset) in the world space. The surface normal to the object's mesh at the user-defined point provides a reference frame to align the egocentric camera's viewing direction and place the 3D asset. To implement a 3rd-person-follow camera, we add a user-defined offset to the object's local reference frame, which is defined by its root-body pose.", + "bbox": [ + 75, + 373, + 472, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 587, + 209, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. In practice, we train our composite scene representation by first pre-training each object field separately. For deformable objects, we pre-train using a depth loss (Equation 10) combined with the losses optimized by BANMo [67]. This includes a silhouette loss $\\mathcal{L}_{\\mathrm{mask}} = \\sum_{\\mathbf{x}^t}||\\mathbf{o}_j(\\mathbf{x}^t) - \\hat{\\mathbf{o}}_j(\\mathbf{x}^t)||^2$ , where the \"ground truth\" object silhouette $\\mathbf{o}_j$ is computed by an off-the-shelf instance segmentation engine [17]. For pre-training the background, we optimize color, flow, and depth losses (Equations 8, 9, 10) on pixels outside the ground truth object silhouettes. Importantly, we don't supervise the object fields on frames that are not provided an object silhouette since it cannot be determined whether the absence of detection is a true or false negative.", + "bbox": [ + 75, + 613, + 468, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After pre-training, we composite-render the pre-trained object fields and jointly finetune them using only the color, depth, flow, and object-specific 3D-cycle consistency losses. Since the silhouette loss is no longer used, the scene representation is supervised on all frames of the training se", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "quence during joint-finetuning. We provide a complete description of the implementation details in the supplement.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We evaluate Total-Recon on novel-view synthesis for deformable scenes. To enable quantitative evaluation, we built a stereo rig comprised of two iPad-Pros rigidly attached to a camera mount, a setup similar to that of Nerfies [35]. Using the stereo rig, we captured 11 RGBD sequences containing 3 different cats, 1 dog, and 2 human subjects in 4 different indoor environments. The RGBD videos were captured using the Record3D iOS App [49], which also automatically registers the frames captured by each camera. These video sequences, which were subsampled at 10 fps, range from 392 to 901 frames, amount to, on average minute-long videos that are significantly longer and contain more dynamic motion than the datasets introduced by [35, 36, 60, 11]. The left and right cameras were registered by solving a Perspective-n-Point (PnP) problem using manually annotated correspondences, and their videos were synchronized based on audio. We provide a complete description of our dataset in the supplement.", + "bbox": [ + 496, + 138, + 890, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reconstruction and Applications. By hierarchically decomposing scene motion into the motion of each object, which itself is decomposed into root-body motion and local articulations, Total-Recon automatically computes novel 6-DoF trajectories such as those traversed by egocentric cameras and 3rd-person follow cameras (Figure 3). In turn, these trajectories enable automated embodied view synthesis and 3D occlusion-aware video filters (Figure 4). These tasks are also enabled by Total-Recon's ability to recover an accurate deformable 3D scene representation, which is currently out of reach for the best of related methods (Figure 5). As shown in the bird's eye view, each reconstructed object is properly situated with respect to the background and other objects, a direct consequence of our use of depth supervision. Furthermore, even though the iPad Pro can only measure depth up to $4\\mathrm{m}$ , Total-Recon can render depth beyond this sensor limit by pooling the measurements from other frames into a single metric scene reconstruction. We provide results on additional sequences in the supplement.", + "bbox": [ + 496, + 429, + 890, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and Evaluation. In Figure 5 and Table 3, we compare Total-Recon to $\\mathrm{D}^2\\mathrm{NeRF}$ [60] and HyperNeRF [36], and their depth-supervised equivalents on the proxy task of stereo-view synthesis, a prerequisite for embodied view synthesis: we train each method on the RGBD frames captured from the left camera of our dataset and evaluate the images rendered from the viewpoint of the right camera. The depth-supervised versions of the baselines contain the same depth loss used in Total-Recon. We report LPIPS [70] and the average (depth) accuracy at $0.1\\mathrm{m}$ [42] in all subsequent experiments, and we include a more complete set", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17676", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9e8a93237ca95f8a6992de4d2c5afcd32218436318eb7d4d9ba12bc36911a487.jpg", + "image_caption": [ + "Figure 4: Embodied View Synthesis and 3D Filters. For select sequences of our RGBD dataset, we visualize the scene geometry and appearance reconstructed by our method (3D reconstruction) and the resulting downstream applications. The yellow and blue camera meshes in the mesh renderings represent the egocentric and 3rd-person-follow cameras, respectively. To showcase the 3D video filter we attach a sky-blue unicorn horn to the forehead of the target object, which is then automatically propagated across all frames. Full-length videos can be found at https://andrewsonga.github.io/totalrecon/applications.html." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 872, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of metrics (PSNR, SSIM, RMS depth error) in the supplement. Because $\\mathrm{D}^2\\mathrm{NeRF}$ and HyperNeRF were not designed to recover a metric scene representation, we replaced their COLMAP [46] camera poses with those provided by the iPad Pro (which are metric measurements) for the sake of fair comparison.", + "bbox": [ + 75, + 699, + 472, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons. Total-Recon qualitatively and quantitatively outperforms all of the baselines. As shown in Figure 5, Total-Recon successfully reconstructs the entire scene, whereas the baselines are only able to reconstruct the rigid background at best. As shown in Table 3, Total-Recon significantly outperforms all baselines in terms of LPIPS and", + "bbox": [ + 75, + 810, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the average accuracy at $0.1\\mathrm{m}$ (Acc@0.1m). We attribute this huge gap to the baselines' inability to reconstruct highly dynamic objects. We provide more details regarding the baselines and additional visualizations in the supplement.", + "bbox": [ + 496, + 699, + 893, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 770, + 665, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 (and Figures 6 and 7) analyzes the importance of Total-Recon's design choices (see Section 3) by ablating its key components: the depth loss $\\mathcal{L}_{\\mathrm{depth}}$ (row 2), the deformation field $\\mathbf{J}_j^t$ (row 3), PoseNet-initialization of the root-body pose (row 4), and the root-body pose $\\mathbf{G}_j^t$ itself (row 5), where $j$ denotes a deformable actor. For all ablations, we use the same set of training losses used in Total-", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17677", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/055a6e13508eefc5facc78127dea3f7acf802e5bcb7d093f91a96cfa618dd142.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DOG 1 (626 images)DOG 1 (v2) (531 images)CAT 1 (641 images)CAT 1 (v2) (632 images)CAT 2 (632 images) (834 images)CAT 2 (v2) (901 images)CAT 3 (767 images)HUMAN 1 (550 images)HUMAN 2 (483 images)HUMAN - DOG (392 images)HUMAN - CAT (431 images)MEAN
LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑
HyperNeRF.634.107.432.176.521.316.438.314.641.277.397.252.592.213.632.053.585.067.487.072.462.162.531.198
D2NeRF.540.219.546.220.687.346.588.403.556.333.595.339.759.231.588.066.630.128.576.078.628.126.611.247
HyperNeRF (+depth).373.352.425.357.532.552.371.596.330.605.376.612.514.451.501.211.445.249.450.283.456.214.428.439
D2NeRF (+depth).507.338.532.270.685.510.580.362.561.438.553.376.730.243.585.086.609.131.608.154.645.176.599.302
Total-Recon.271.841.313.790.382.889.333.894.237.967.281.925.261.949.213.909.264.849.256.827.233.914.278.895
", + "bbox": [ + 84, + 89, + 885, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/86e2c39980dc895e8c6568960f3e9a74f94daf6a9704a87d1ed1a17bd2fcc4d3.jpg", + "image_caption": [ + "Figure 5: Baseline Comparisons. We compare Total-Recon to depth-supervised variants of HyperNeRF [36] and $\\mathrm{D}^2\\mathrm{NeRF}$ [60] on the task of stereo-view synthesis (the left camera is used for training and the right is used for testing). While the baselines are able to reconstruct only the background, Total-Recon can reconstruct both the background and the moving deformable object(s), demonstrating holistic scene reconstruction." + ], + "image_footnote": [], + "bbox": [ + 81, + 257, + 467, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recon and initialize camera pose $\\mathbf{G}_0^t$ with those reported by ARKit. For ablations that model root-body motions, we initialize each deformable actor's root-body pose $\\mathbf{G}_j^t$ with predictions made by PoseNet [67] and optimize them during reconstruction; for row 4, we replace the PoseNet predictions with identity rotations. We report the novel-view metrics averaged over 6 select sequences of our dataset: DOG 1 (v1), CAT 1 (v1), CAT 2 (v1), HUMAN 1, HUMAN 1 & DOG 1, and HUMAN 2 & CAT 1.", + "bbox": [ + 75, + 617, + 468, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Depth Supervision. Table 4 shows that removing depth supervision (row 2) significantly reduces the average accuracy at $0.1\\mathrm{m}$ (Acc). Figure 6 indicates that this reflects the incorrect arrangement of objects stemming from their scale inconsistency - while removing depth supervision does not significantly deteriorate the training-view RGB renderings, it induces critical failure modes as shown in the novel-view 3D reconstructions: (a) floating foreground objects, as evi", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/089512a22f661b1412e2a5ae194d7eaf1dfc078d00c46ecb0f3a45b7085f643b.jpg", + "table_caption": [ + "Table 3: Baseline Comparisons. We train Total-Recon, HyperNeRF [36], D²NeRF [60], and their depth-supervised variants on the left video captured with our stereo rig, and evaluate the novel view synthesis results on the held-out right video. Total-Recon significantly outperforms all of the baselines for all 11 sequences. These sequences are sampled at 10 fps, amounting to minute-long videos, on average." + ], + "table_footnote": [], + "table_body": "
MethodsDepth LossDeform. Obj.Root Init.Root MotionLPIPS↓Acc@0.1m↑
(1) Full model✓✓✓✓.268.898
(2) w/o loss LdepthX✓✓✓.372.154
(3) w/o deform. Jj✓X✓✓.296.867
(4) w/o root-body init.✓✓X✓.293.870
(5) w/o root-body Gj✓✓XXN/AN/A
", + "bbox": [ + 511, + 252, + 880, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: Ablation Study. Removing depth supervision (2) significantly hurts performance, while removing the deformation field (3) and PoseNet-initialization of root-body poses (4) hurts moderately. Most importantly, removing root-body poses entirely (5) prevents convergence (N/A) as the deformation field alone has to explain global object motion (see Figure 2). These experiments justify our hierarchical modeling of motion, as even root-bodies without a deformation field (3) or poorly initialized root-bodies (4) are better than no root-bodies (5). We visualize these ablations in Figure 7 and explore other ablations in the supplement.", + "bbox": [ + 498, + 356, + 890, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "denced by their shadows, and (b) the human incorrectly occluding the dog. In other words, without depth supervision, Total-Recon overfits the training view and learns a degenerate scene representation where the reconstructed objects fail to converge to the same scale. We show results on additional sequences in the supplement.", + "bbox": [ + 498, + 525, + 890, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motion Modeling. Table 4 shows that removing the deformation field (row 3) also hurts performance. This is because, without the deformation field, our method has to explain an object's non-rigid motion solely with its rigid, root-body poses. As a result, this ablation can only recover coarse object reconstructions that fail to model moving body parts such as limbs. Removing PoseNet-initialization of root-body poses (row 4) is just as detrimental, resulting in noisy appearance and geometry artifacts; see Figure 7 and additional visuals in the supplement. Most notably, Table 4 shows that removing object root-bodies entirely (row 5) causes the optimization to fail to converge (N/A), even though the deformation field can (in theory) represent all continuous motion. It appears difficult for deformation fields alone to explain global root-body motion because such motions can deviate significantly from a canonical model, complicating optimization.", + "bbox": [ + 496, + 643, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17678", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6cf3c6572a73a8bf1922c93283acb8d76d41ea7ba320ab4494f4cc8c3d316329.jpg", + "image_caption": [ + "Figure 6: Ablation Study on Depth Supervision. While removing depth supervision does not significantly deteriorate the training-view RGB renderings, it significantly hurts the novel-view 3D reconstructions, as characterized by the following: (a) floating foreground objects (as evidenced by their shadows) and (b) the human incorrectly occluding the dog. These failure modes indicate that without depth supervision, Total-Recon overfits the training view, and the reconstructed objects fail to converge to the same scale." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 467, + 252 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c03c6cd8035ea19355438b7fbc3e13dbf5abbc226a7546780f83999cf555a483.jpg", + "image_caption": [ + "Figure 7: Ablation Study. We visualize the ablations from Table 4. Removing the depth loss causes the cat to sink into the ground due to inconsistent object scales. Removing the deformation field produces coarse object reconstructions that fail to capture articulated body parts such as moving limbs. Removing PoseNet-initialization of root-body poses results in noisier appearance and geometry, as shown by the human actor's left hand. We omit the ablation without root-body poses as it does not converge. We present additional visualizations in the supplement." + ], + "image_footnote": [], + "bbox": [ + 78, + 404, + 468, + 588 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "These diagnostics justify Total-Recon's hierarchical motion representation, which explicitly models objects' root-body poses; root-bodies without articulated deformations (row 3) or poorly initialized root-bodies (row 4) are better than no root-bodies at all (row 5). Our ablations also suggest that the poor performance of the baseline methods (on our challenging dataset) may be attributed to the lack of object-centric motion modeling. We provide a more detailed analysis with additional experiments and RGBD sequences in the supplement.", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5. Discussion and Limitations", + "text_level": 1, + "bbox": [ + 500, + 89, + 751, + 106 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have presented a new system for automated embodied view synthesis from monocular RGBD videos, focusing on videos of people interacting with their pets. Our main technical contribution is Total-Recon, a 3D representation for deformable scenes that hierarchically decomposes scene motion into the motion of each object, which is further decomposed into its root-body motion and local articulations; this key design choice enables appropriate initialization of root-body poses and hence easier optimization over long videos containing large motions. By explicitly reconstructing the geometry, appearance, root-body- and articulated motion of each object, Total-Recon enables seeing through the eyes of people and pets and generating game-like traversals of deformable scenes from behind a target actor.", + "bbox": [ + 496, + 114, + 893, + 327 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations. In Total-Recon, scene decomposition is primarily supervised by object silhouettes computed by an off-the-shelf segmentation model [17], which may be inaccurate, especially in partial occlusion scenarios. This may damage the resulting reconstructions and embodied-view renderings. We believe that incorporating the latest advances in video instance segmentation [12] will enable Total-Recon to be applied to more challenging scenarios. Second, Total-Recon initializes the root-body pose of each deformable object using a PoseNet [67] trained for humans and quadruped animals, which does not generalize to other object categories (e.g., birds, fish). We reserve the reconstruction of generic scenes for future work. Finally, our model needs to be optimized on a per-sequence basis for roughly 15 hours with 4 NVIDIA RTX A5000 GPUs and is therefore not suitable for real-time applications. Incorporating recent advances in fast neural field training methods is an interesting avenue for future research.", + "bbox": [ + 496, + 347, + 893, + 619 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments. We thank Nathaniel Chodosh, Jeff Tan, George Cazenavette, and Jason Zhang for proofreading and Songwei Ge for reviewing the codebase. We also thank Sheng-Yu Wang, Daohan (Fred) Lu, Tamaki Kojima, Krishna Wadhwani, Takuya Narihira, and Tatsuo Fujiwara for providing valuable feedback. This work is supported in part by the Sony Corporation, Cisco Research, and the CMU Argo AI Center for Autonomous Vehicle Research.", + "bbox": [ + 496, + 640, + 893, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 776, + 596, + 791 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[2] Aljaz Bozic, Pablo Palafox, Michael Zollhofer, Justus Thies, Angela Dai, and Matthias Nießner. Neural deformation graphs for globally-consistent non-rigid reconstruction. In" + ], + "bbox": [ + 506, + 801, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17679", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[3] Aljaz Bozic, Michael Zollhofer, Christian Theobalt, and Matthias Nießner. Deepdeform: Learning non-rigid rgb-d reconstruction with semi-supervised data. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[4] Hongrui Cai, Wanquan Feng, Xuetao Feng, Yan Wang, and Juyong Zhang. Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3", + "[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[7] Alena Denisova and Paul Cairns. First person vs. third person perspective in digital games: Do player preferences affect immersion? In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, page 145-148, 2015. 1", + "[8] Mingsong Dou, Jonathan Taylor, Henry Fuchs, Andrew Fitzgibbon, and Shahram Izadi. 3d scanning deformable objects with a single rgbd sensor. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3", + "[9] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbaek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[10] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3", + "[11] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3, 6", + "[12] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 9", + "[13] Alec Jacobson, Zhigang Deng, Ladislav Kavan, and J. P. Lewis. Skinning: Real-time shape deformation (full text not available). In ACM SIGGRAPH 2014 Courses, SIGGRAPH '14. Association for Computing Machinery, 2014. 4", + "[14] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 3", + "[15] Zhang Jiakai, Liu Xinhang, Ye Xinyi, Zhao Fuqiang, Zhang Yanshun, Wu Minye, Zhang Yingliang, Xu Lan, and Yu Jingyi. Editable free-viewpoint video using a layered neural representation. In ACM SIGGRAPH, 2021. 3, 4", + "[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "from a single video. In European Conference on Computer Vision (ECCV), 2022. 3", + "[17] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6, 9", + "[18] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J. Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3", + "[19] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In European Conference on Computer Vision (ECCV), 2022. 3", + "[20] Tianye Li, Mira Slavcheva, Michael Zollhöfer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[21] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3", + "[22] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[23] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 3", + "[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM SIGGRAPH Asia, 2021. 3", + "[25] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[26] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 4", + "[27] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In IEEE International Conference on Computer Vision (ICCV), 2021. 3", + "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 4" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17680", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3", + "[30] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE International Symposium on Mixed and Augmented Reality, pages 127-136. IEEE, 2011. 3", + "[31] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5", + "[32] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In IEEE International Conference on Computer Vision (ICCV), 2021. 3", + "[33] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3", + "[34] Kemal Egemen Ozden, Kurt Cornelis, Luc Van Eycken, and Luc Van Gool. Reconstructing 3d trajectories of independently moving objects using generic constraints. Comput. Vis. Image Underst., 96(3):453-471, 2004. 3", + "[35] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3, 6", + "[36] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 40(6),2021.2,3,4,6,8", + "[37] Georgios Pavlakos*, Ethan Weber*, , Matthew Tancik, and Angjoo Kanazawa. The one where they reconstructed 3d humans and environments in tv shows. In European Conference on Computer Vision (ECCV), 2022. 3", + "[38] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In IEEE International Conference on Computer Vision (ICCV), 2021. 3", + "[39] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[40] F. Peruzzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3", + "[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for Dynamic Scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3", + "[42] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 6", + "[43] Zhongzheng Ren, Xiaoming Zhao, and Alex Schwing. Class-agnostic reconstruction of dynamic objects from videos. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[44] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[45] Richard Rouse. What's your perspective? ACM SIGGRAPH Computer Graphics, 33(3):9-12, August 1999. 1", + "[46] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 7", + "[47] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[48] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In SIGGRAPH Conference Proceedings, 2022. 3", + "[49] Marek Simonik. Record3d. https://record3d.app. 6", + "[50] Miroslava Slavcheva, Maximilian Baust, Daniel Cremers, and Slobodan Ilic. Killingfusion: Non-rigid 3d reconstruction without correspondences. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3", + "[51] Miroslava Slavcheva, Maximilian Baust, and Slobodan Ilic. Sobolevfusion: 3d reconstruction of scenes undergoing free non-rigid motion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3", + "[52] Karl Stelzner, Kristian Kersting, and Adam R Kosiorek. Decomposing 3d scenes into objects via unsupervised volume segmentation. arXiv preprint arXiv:2104.01148, 2021. 5", + "[53] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[54] Unity Technologies. 3rd person follow. https://docs.unity3d.com/Packages/ com.unity.cinemachine@2.8/manual/ Cinemachine3rdPersonFollow.html.1", + "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17681", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4", + "[57] Barbara Tversky. Spatial Cognition Embodied and Situated. Cambridge, 2008. 1", + "[58] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021. 2, 3", + "[59] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 3", + "[60] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. $\\mathbf{D}^2$ nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 4, 6, 8", + "[61] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3", + "[62] Jing Xiao, Jinxiang Chai, and Takeo Kanade. A closed-form solution to non-rigid shape and motion recovery. International Journal of Computer Vision (IJCV), 67(2):233-246, April 2006. 3", + "[63] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[64] Gengshan Yang and Deva Ramanan. Volumetric correspondence networks for optical flow. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 5", + "[65] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T Freeman, and Ce Liu. LASR: Learning articulated shape reconstruction from a monocular video. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[66] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2", + "[67] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6, 8, 9", + "[68] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4", + "[69] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[70] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6", + "[71] Xiaoshuai Zhang, Abhijit Kundu, Thomas Funkhouser, Leonidas Guibas, Hao Su, and Kyle Genova. Nerflets: Local radiance fields for efficient structure-aware 3d scene representation from 2d supervision. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + ], + "bbox": [ + 503, + 92, + 890, + 189 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "17682", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/49fdb57e-3f25-4f0f-85ac-577ce80d4f48_model.json b/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/49fdb57e-3f25-4f0f-85ac-577ce80d4f48_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f869d5ce7eb2c92f31147fde97652d1b303ca50a --- /dev/null +++ b/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/49fdb57e-3f25-4f0f-85ac-577ce80d4f48_model.json @@ -0,0 +1,2171 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "Total-Recon: Deformable Scene Reconstruction for Embodied View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.181, + 0.812, + 0.22 + ], + "angle": 0, + "content": "Chonghyuk Song Gengshan Yang Kangle Deng Jun-Yan Zhu Deva Ramanan Carnegie Mellon University" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.222, + 0.894, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.419, + 0.893, + 0.492 + ], + "angle": 0, + "content": "Figure 1: Embodied View Synthesis. Given a long video of deformable objects captured by a handheld RGBD sensor, Total-Recon renders the scene from novel camera trajectories derived from the in-scene motion of actors: (1) egocentric cameras that simulate the point-of-view of a target actor (such as the pet) and (2) 3rd-person (or pet) cameras that follow the actor from behind. Our method also enables (3) 3D video filters that attach virtual 3D assets to the actor. Total-Recon achieves this by reconstructing the geometry, appearance, root-body- and articulated motion of each deformable object in the scene and the background." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.509, + 0.314, + 0.525 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.538, + 0.474, + 0.886 + ], + "angle": 0, + "content": "We explore the task of embodied view synthesis from monocular videos of deformable scenes. Given a minute-long RGBD video of people interacting with their pets, we render the scene from novel camera trajectories derived from the in-scene motion of actors: (1) egocentric cameras that simulate the point of view of a target actor and (2) 3rd-person cameras that follow the actor. Building such a system requires reconstructing the root-body and articulated motion of every actor, as well as a scene representation that supports free-viewpoint synthesis. Longer videos are more likely to capture the scene from diverse viewpoints (which helps reconstruction) but are also more likely to contain larger motions (which complicates reconstruction). To address these challenges, we present Total-Recon, the first method to photorealistically reconstruct deformable scenes from long monocular RGBD videos. Crucially, to scale to long videos, our method hierarchically decomposes the scene into the background and objects, whose motion is decomposed into carefully initialized root-body motion and local articulations. To quantify such \"in-the-wild\" reconstruction and view synthesis, we collect ground-truth data from a specialized stereo RGBD capture rig for 11 challenging videos, significantly outperforming prior methods." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.509, + 0.633, + 0.525 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.895, + 0.822 + ], + "angle": 0, + "content": "We explore embodied view synthesis, a new class of novel-view synthesis tasks that renders deformable scenes from novel 6-DOF trajectories reconstructed from the in-scene motion of actors: egocentric cameras [45, 7] that simulate the point-of-view of moving actors and 3rd-person-follow cameras [54, 7] that track a moving actor from behind (Figure 1). We focus on everyday scenes of people interacting with their pets, producing renderings from the point-of-view of the person and pet (Figure 1). While such camera trajectories could be manually constructed (e.g., by artists via keyframing), building an automated system is an interesting problem of its own: spatial cognition theory [57] suggests that the ability to visualize behavior from another actor's perspective is necessary for action learning and imitation; in the context of gaming and virtual reality [7, 45], egocentric cameras offer high levels of user immersion, while 3rd-person-follow cameras provide a large field of view that is useful for exploring a user's environment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.842, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Challenges. Building a system for embodied view synthesis is challenging for many reasons. First, to reconstruct everyday-but-interesting content, it needs to process long, monocular captures of multiple interacting actors. How-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17671" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.161, + 0.09, + 0.81, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.347, + 0.896, + 0.434 + ], + "angle": 0, + "content": "Figure 2: Method Overview. Total-Recon represents the entire scene as a composition of \\(M\\) object-centric neural fields, one for the rigid background and each of the \\(M - 1\\) deformable objects. To render a scene, (1) each object field \\(j\\) is transformed into the camera space with a rigid transformation \\(\\left(\\mathbf{G}_j^t\\right)^{-1}\\) that encodes root-body motion and, for each deformable object, an additional deformation field \\(\\mathbf{J}_j^{t,\\rightarrow}\\) that encodes articulated motion. Next, all (2) posed object fields are combined into a (3) composite field, which is then volume-rendered into (4) color, depth, optical flow, and object silhouettes. Each rendered output defines a reconstruction loss that derives supervision from a monocular RGBD video captured by a moving iPad Pro." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.473, + 0.658 + ], + "angle": 0, + "content": "ever, such videos are likely to contain large scene motions, which we demonstrate are difficult to reconstruct with current approaches. Second, it needs to produce a deformable 3D scene representation that supports free-viewpoint synthesis, which also would benefit from long videos likely to capture the scene from diverse viewpoints. Recent approaches have extended Neural Radiance Fields (NeRFs) [28] to deformable scenes, but such work is often limited to rigid-only object motion [18, 33], short videos with limited scene motion [41, 35, 21, 55, 36, 60, 10, 61, 58], or reconstructing single objects as opposed the entire scene [65, 66, 67, 4]. Third, it needs to compute global 6-DOF trajectories of root-bodies and articulated body parts (e.g., head) of multiple actors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Key Ideas. To address these challenges, we introduce Total-Recon, the first monocular NeRF that enables embodied view synthesis for deformable scenes with large motions. Given a monocular RGBD video, Total-Recon reconstructs the scene as a composition of object-centric representations, which encode the 3D appearance, geometry, and motion of each deformable object and the background. Crucially, Total-Recon hierarchically decomposes scene motion into the motion of individual objects, which itself is decomposed into global root-body movement and the local deformation of articulated body parts. We demonstrate that such decomposition of object motion, along with appropriate initialization of root-body pose, allows reconstruction to scale to longer videos, enabling free-viewpoint synthesis. By reconstructing such motions in a globally-consistent" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.446, + 0.895, + 0.493 + ], + "angle": 0, + "content": "coordinate frame, Total-Recon can generate renderings from egocentric and 3rd-person-follow cameras, as well as static but extreme viewpoints like bird's-eye-views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.895, + 0.683 + ], + "angle": 0, + "content": "Evaluation. Due to the difficulty of collecting ground-truth data for embodied view synthesis on in-the-wild videos, we evaluate our method on the proxy task of stereo-view synthesis [35], which compares rendered views to those captured from a stereo pair. To this end, we build a stereo RGBD sensor capture rig for ground-truthing and collect a dataset of 11 long video sequences in various indoor environments, including people interacting with their pets. Total-Recon outperforms the state-of-the-art monocular deformable NeRF methods [36, 60], even when modified to use depth sensor measurements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Contributions. In summary, our contributions are: (1) Total-Recon, a hierarchical 3D representation that models deformable scenes as a composition of object-centric representations, each of which decomposes object motion into its global root-body motion and its local articulations; (2) a system based on Total-Recon for automated embodied view synthesis from casual, minute-long RGBD videos of highly dynamic scenes; (3) a dataset of stereo RGBD videos containing various deformable objects, such as humans and pets, in a host of different background environments. Our code, models, and data can be found at https://andrewsonga.github.io/totalrecon." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17672" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.472, + 0.238 + ], + "angle": 0, + "content": "
MethodEntire ScenesDeform. ObjectsBeyond HumansGlobal 6-DOF Traj.Long VideosExtreme Views
BANMo [67]X✓✓X✓✓
PNF [18]✓X✓✓XX
NeuMan [16]✓✓X✓XX
SLAHMR [69]X✓X✓XX
HyperNeRF [36]✓✓✓XXX
D2NeRF [60]✓✓✓XXX
DynIBaR [22]✓✓✓X✓X
SUDS [56]✓✓✓X✓X
Ours✓✓✓✓✓✓
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.348 + ], + "angle": 0, + "content": "Table 1: Comparison to Related Work. Unlike prior work, Total-Recon targets embodied view synthesis of scenes containing humans and pets, requiring the ability to (1) reconstruct entire scenes, (2) model deformable objects, (3) extend beyond humans, (4) recover global 6-DOF trajectories of objects' root-bodies and their articulated parts, (5) process minute-long videos of dynamic scenes, and (6) render extreme views." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.372, + 0.22, + 0.387 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.47, + 0.626 + ], + "angle": 0, + "content": "Neural Radiance Fields. Prior works on Neural Radiance Fields (NeRF) optimize a continuous scene function for novel view synthesis given a set of multi-view images, usually under the assumption of a rigid scene and densely sampled views [28, 26, 27, 23, 14, 59]. DS-NeRF [6] and Dense Depth Priors [44] extend NeRFs to the sparse-view setting by introducing depth as additional supervision. Total-Recon also operates in the sparse-view regime and uses depth supervision to reduce the ambiguities inherent to monocular, multibody, non-rigid reconstruction [62, 34]. Another line of work [18, 33] represents rigidly moving scenes as a composition of multiple object-level NeRFs. Total-Recon also leverages such an object-centric scene representation, but models scenes containing non-rigidly moving objects, such as humans and pets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Deformable NeRFs. Recent approaches extend NeRF to monocular deformable scene reconstruction either by learning an additional function that deforms observed points in the camera space to a time-independent canonical space [41, 35, 55, 36, 60] or explicitly modeling density changes over time [10, 61, 58, 21]. Such methods are typically limited to short videos containing little scene and camera motion. They also perform novel-view synthesis only over small baselines. Total-Recon belongs to the former category of prior monocular deformable NeRFs, but unlike them, our method hierarchically decomposes scene motion into the motion of each object, which is further decomposed into global root-body motion and local articulations. The proposed motion decomposition is what enables embodied view synthesis: it allows Total-Recon to scale to minute-long videos and reconstruct a deformable 3D scene representation that supports free-viewpoint syn" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.364 + ], + "angle": 0, + "content": "thesis; it also makes it easy to extract an object's root-body motion, the key motion primitive required for 3rd-person-follow view synthesis. Several works have taken different approaches to making non-rigid reconstruction more tractable. One group of work leverages human-specific priors [38, 53, 32, 39, 24, 16, 19, 37] such as human body models (e.g., SMPL), 3D skeletons, or 2D poses to achieve high reconstruction quality. We achieve similar levels of fidelity without relying on such shape priors, allowing Total-Recon to generalize to pets and, by extension, reconstruct human-pet interaction videos. Another body of work [15, 48, 20] achieves high-fidelity scene reconstructions by relying on synchronized multi-view video captured from a specialized camera rig ranging from 8 to 18 static cameras. In contrast, Total-Recon only requires a single video captured from a moving RGBD camera equipped with inertial sensors, which has now become widely accessible in consumer products with the advent of Apple's iPhone and iPad Pro." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Reconstruction with RGBD Sensors. Depth sensors represent the third class of attempts to make non-rigid reconstruction more tractable, reducing the need for a predefined shape template. Kinect-fusion [30] creates a real-time system for indoor scene localization and mapping. Dynamic Fusion [29] builds a template-free dense SLAM system for dynamic objects. Later works improve RGBD reconstruction to be able to deal with topology changes [50, 51] and use correspondence matching for registration over large motions [2, 3]. Recent works have incorporated neural implicit representations to reconstruct the surface geometry and 3D motion fields for deformable objects [43, 4] or largescale rigid scenes [1, 42] in isolation. Other works have reconstructed humans alongside small-scale objects and furniture [8, 2], but not the entire background. We aim to go even further by reconstructing the entire scene, which includes the background and multiple deformable targets such as humans and pets; not only do we reconstruct the geometry, but we also recover a radiance field that allows for photorealistic scene rendering from embodied viewpoints and other novel 6-DOF trajectories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Concurrent Work. Concurrent work exhibits a subset of the design choices necessary for embodied view synthesis. SLAHMR [69] reconstructs the geometry and in-scene motion of human actors but not the scene appearance. Nerflets [71] models the appearance, geometry, and motion of each scene element but is limited to rigidly moving objects. RoDynRF [25], NeRF-DS [63], HexPlane [5], and K-planes [9] reconstruct other types of dynamic scene elements, such as deformable or specular objects, but these methods have been demonstrated on only short videos and/or videos containing limited object root-body motion [40, 47, 63, 11]. DynIBaR [22] scales dynamic view synthesis to longer" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17673" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.182 + ], + "angle": 0, + "content": "videos with complex camera and scene motion, and SUDS [56] scales reconstruction to urban-scale dynamic scenes captured from 1.2 million frames. However, neither demonstrates extreme-view synthesis, a prerequisite for rendering embodied views. We summarize and compare prior work to Total-Recon in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.195, + 0.169, + 0.21 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.219, + 0.3, + 0.235 + ], + "angle": 0, + "content": "3.1. Limitations of Prior Art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.47, + 0.333 + ], + "angle": 0, + "content": "The state-of-the-art monocular deformable NeRFs [36, 60] decompose a deformable scene into a rigid, canonical template model and a deformation field \\(\\mathbf{J}^{t,\\leftarrow}\\) that maps the world space \\(\\mathbf{G}_0^t\\mathbf{X}^t\\) to the canonical space \\(\\mathbf{X}^*\\), where \\(\\mathbf{G}_0^t\\) is the known camera pose at time \\(t\\), and \\(\\mathbf{X}^t\\) is a camera space point at time \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.342, + 0.47, + 0.361 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {*} = \\mathcal {W} ^ {t, \\leftarrow} \\left(\\mathbf {X} ^ {t}\\right) = \\mathbf {J} ^ {t, \\leftarrow} \\left(\\mathbf {G} _ {0} ^ {t} \\mathbf {X} ^ {t}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.47, + 0.519 + ], + "angle": 0, + "content": "In theory, this formulation is sufficient to represent all continuous motion; it performs well on short videos containing near-rigid scenes, as the deformation field only has to learn minute deviations from the template model. However, this motion model is difficult to scale to minute-long videos, which are more likely to contain deformable objects undergoing large translations (e.g., a person walking into another room) and pose changes (e.g., a person sitting down). Here, the deformation field must learn large deviations from the canonical model, significantly complicating optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.47, + 0.581 + ], + "angle": 0, + "content": "Another critical limitation of HyperNeRF and \\(\\mathrm{D}^2\\mathrm{NeRF}\\) is that they cannot track separate deformable objects and therefore cannot perform 3rd-person-follow view synthesis for scenes with multiple actors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.588, + 0.332, + 0.604 + ], + "angle": 0, + "content": "3.2. Component Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.611, + 0.471, + 0.779 + ], + "angle": 0, + "content": "To address the limitations of existing monocular deformable NeRFs, we propose Total-Recon, a novel 3D representation that models a deformable scene as a composition of \\(M\\) object-centric neural fields, one for the rigid background and each of the \\(M - 1\\) deformable objects (Figure 2). Crucially, Total-Recon hierarchically decomposes scene motion into the motion of each object, which itself is decomposed into global root-body motion and local articulations. This key design choice scales our method to minute-long videos containing highly dynamic and deformable objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Background Radiance Field. We begin by modeling the background environment as a Neural Radiance Field (NeRF) [28]. For a 3D point \\(\\mathbf{X}^{*} \\in \\mathbb{R}^{3}\\) and a viewing direction \\(\\mathbf{v}^{*}\\) in the canonical world space, NeRF defines a color \\(\\mathbf{c}\\) and density \\(\\sigma\\) represented by an MLP. We follow contemporary variants [26] that include a time-specific embedding code \\(\\omega_{e}^{t}\\) to model illumination changes over time" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.138 + ], + "angle": 0, + "content": "and model density with as a function of a neural signed distance function (SDF) \\(\\mathbf{MLP}_{\\sigma}(\\cdot) = \\alpha \\Gamma_{\\beta}(\\mathbf{MLP}_{\\mathrm{SDF}}(\\cdot))\\) [68] to encourage the reconstruction of a valid surface:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.148, + 0.893, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\sigma = \\mathbf {M L P} _ {\\sigma} \\left(\\mathbf {X} ^ {*}\\right), \\quad \\mathbf {c} ^ {t} = \\mathbf {M L P} _ {\\mathbf {c}} \\left(\\mathbf {X} ^ {*}, \\mathbf {v} ^ {*}, \\omega_ {e} ^ {t}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.177, + 0.892, + 0.207 + ], + "angle": 0, + "content": "The pixel color can then be computed with differentiable volume rendering equations (Section 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.208, + 0.893, + 0.328 + ], + "angle": 0, + "content": "Most NeRF methods, including HyperNeRF [36] and \\(\\mathrm{D}^2\\mathrm{NeRF}\\) [60], assume images with known cameras. While our capture devices are equipped with inertial sensors, we find their self-reported camera poses have room for improvement. As such, we also model camera pose as an opti- mizable rigid-body transformation \\(\\mathbf{G}_0^t\\in SE(3)\\) that maps points in a time-specific camera space \\(\\mathbf{X}^t\\in \\mathbb{R}^3\\) to the world space (where we assume homogenous notation):" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.339, + 0.892, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {*} = \\mathbf {G} _ {0} ^ {t} \\mathbf {X} ^ {t}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.893, + 0.547 + ], + "angle": 0, + "content": "Deformable Field (for Object \\(j\\)). We model the deformable radiance field of object \\(j \\in \\{1, \\dots, M - 1\\}\\) with BANMo [67], which consists of a canonical rest shape and time-dependent deformation field. The canonical rest shape is represented by the same formulation described by Equation 2, but now defined in a local object-centric canonical space rather than the world space. BANMo represents object motion with a warping function \\(\\mathcal{W}_j^{t,\\leftarrow} : \\mathbf{X}^t \\rightarrow \\mathbf{X}_j^*\\) that maps the camera space points \\(\\mathbf{X}^t\\) to canonical space points \\(\\mathbf{X}_j^*\\) with a rigid-body transformation \\(\\mathbf{G}_j^t \\in SE(3)\\) and a deformation field \\(\\mathbf{J}_j^{t,\\leftarrow}\\) modeled by linear blend skinning [13]:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.569, + 0.892, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {j} ^ {*} = \\mathcal {W} _ {j} ^ {t, \\leftarrow} \\left(\\mathbf {X} ^ {t}\\right) = \\mathbf {J} _ {j} ^ {t, \\leftarrow} \\left(\\mathbf {G} _ {j} ^ {t} \\mathbf {X} ^ {t}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Note that our choice of deformation field differs from the \\(SE(3)\\)-field used in HyperNeRF and \\(\\mathrm{D}^2\\mathrm{NeRF}\\), which has been shown to produce irregular deformation in the presence of complex scene motion [67]. Intuitively, rigid-body transformation \\(\\mathbf{G}_j^t\\) captures the global root-body pose of object \\(j\\) relative to the camera at time \\(t\\), while deformation field \\(\\mathbf{J}_j^{t,\\leftarrow}\\) aligns more fine-grained articulations relative to its local canonical space (Figure 2). Explicitly disentangling these two sources of object motion (as opposed to conflating them) enables easier optimization of the deformation field, because local articulations are significantly easier to learn than those modeled relative to the world space (Equation 1). Furthermore, this motion decomposition makes the deformation field invariant to rigid-body transformations of the object. A motion model similar to ours was proposed by ST-NeRF [15], but their model encodes an object's global root-body motion with a 3D axis-aligned bounding box that does not explicitly represent object orientation, a prerequisite for embodied view synthesis from 3rd-person-follow cameras." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17674" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.46, + 0.164 + ], + "angle": 0, + "content": "
Rendered features f at pixel xtCorresponding 3D features fij(Xt)
color帽子(xt)ct( Wj,←(Xt))
flow F( xt, t → t')Πt' (Wj', → (Wj, ←(Xt)) - xt
depth d( xt)[0, 0, 1] · Xt
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.173, + 0.47, + 0.204 + ], + "angle": 0, + "content": "Table 2: Rendered 2D features \\(\\hat{\\mathbf{f}}\\) and their corresponding 3D features \\(\\mathbf{f}_{ij}\\). \\(\\Pi^{t'}\\) denotes the camera intrinsics at time \\(t'\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.223, + 0.471, + 0.3 + ], + "angle": 0, + "content": "As did BANMo, Total-Recon also models a forward warp \\(\\mathbf{X}_j^t = \\mathcal{W}_j^{t,\\rightarrow}(\\mathbf{X}^*) = (\\mathbf{G}_j^t)^{-1}\\mathbf{J}_j^{t,\\rightarrow}(\\mathbf{X}^*)\\) that maps the canonical space to the camera space, which is used to establish the surface correspondences required for egocentric view synthesis and 3D video filters." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.31, + 0.434, + 0.328 + ], + "angle": 0, + "content": "3.3. Composite Rendering of Multiple Objects" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.47, + 0.456 + ], + "angle": 0, + "content": "Given a set of \\(M\\) object representations (the background is treated as an object as well), we use the composite rendering scheme from prior work [31, 52] to combine the outputs of all object representations and volume-render the entire scene. To volumetrically render the image at frame \\(t\\), we sample multiple points along each camera ray \\(\\mathbf{v}^t\\). Denoting the \\(i^{th}\\) sample as \\(\\mathbf{X}_i^t\\), we write the density and color observed at sample \\(i\\) due to object \\(j\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.467, + 0.472, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {i j} = \\mathbf {M L P} _ {\\sigma , j} \\left(\\mathbf {X} _ {i j} ^ {*}\\right), \\qquad \\mathbf {c} _ {i j} = \\mathbf {M L P} _ {\\mathbf {c}, j} \\left(\\mathbf {X} _ {i j} ^ {*}, \\mathbf {v} _ {j} ^ {*}, \\boldsymbol {\\omega} _ {e} ^ {t}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.47, + 0.608 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_{ij}^{*} = \\mathcal{W}_{j}^{t,\\leftarrow}(\\mathbf{X}_{i}^{t})\\) and \\(\\mathbf{v}_j^* = \\mathcal{W}_j^{t,\\leftarrow}(\\mathbf{v}^t)\\) are sample \\(i\\) and camera ray \\(\\mathbf{v}^t\\) backward-warped into object \\(j\\)'s canonical space, respectively. The composite density \\(\\sigma_{i}\\) at sample \\(i\\) along the ray is then computed as the sum of each object's density \\(\\sigma_{ij}\\); the composite color \\(\\mathbf{c}_i\\) is computed as the weighted sum of each object's color \\(\\mathbf{c}_{ij}\\), where the weights are the normalized object densities \\(\\sigma_{ij} / \\sigma_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.619, + 0.47, + 0.664 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {i} = \\sum_ {j = 0} ^ {M - 1} \\sigma_ {i j}, \\quad \\mathbf {c} _ {i} = \\frac {1}{\\sigma_ {i}} \\sum_ {j = 0} ^ {M - 1} \\sigma_ {i j} \\mathbf {c} _ {i j}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.675, + 0.47, + 0.751 + ], + "angle": 0, + "content": "We can then use the standard volume rendering equations to generate an RGB image of the scene, where \\(N\\) is the number of sampled points along camera ray \\(\\mathbf{v}^t\\), \\(\\tau_{i}\\) is the transmittance, \\(\\alpha_{i}\\) is the alpha value for sample point \\(i\\) and \\(\\delta_{i}\\) is the distance between sample point \\(i\\) and the \\((i + 1)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.775, + 0.463, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {c}} = \\sum_ {i = 1} ^ {N} \\tau_ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\quad \\tau_ {i} = \\prod_ {k = 1} ^ {i - 1} (1 - \\alpha_ {k}), \\quad \\alpha_ {i} = 1 - e ^ {- \\sigma_ {i} \\delta_ {i}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Rendering Flow, Depth, and Silhouettes. Our composite rendering scheme can be used to render different quantities by replacing the object color \\(\\mathbf{c}_{ij}\\) in Equation 5 with the" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.45 + ], + "angle": 0, + "content": "Figure 3: 6-DOF Trajectories for Embodied View Synthesis. To synthesize embodied views from egocentric or actor-following cameras, Total-Recon reconstructs the entire background, every individual actor in the scene, as well as global 6-DOF trajectories of its root-body and its articulated body parts (e.g. head)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.547 + ], + "angle": 0, + "content": "appropriately defined 3D feature \\(\\mathbf{f}_{ij}\\) (Table 2) and rendering the resulting composite feature \\(\\mathbf{f}_i\\). To render occlusion-aware object silhouettes, we follow ObSURF [52] to produce a categorical distribution over the \\(M\\) objects:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.576, + 0.891, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {o}} _ {j} = \\sum_ {i = 1} ^ {N} \\tau_ {i} \\alpha_ {i j}, \\quad \\text {w h e r e} \\quad \\tau_ {i} = \\prod_ {k = 1} ^ {i - 1} (1 - \\alpha_ {k}), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.619, + 0.892, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = 1 - e ^ {- \\sigma_ {i} \\delta_ {i}}, \\quad \\alpha_ {i j} = 1 - e ^ {- \\sigma_ {i j} \\delta_ {i}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Losses. Given a monocular RGBD video, we optimize all parameters in our composite scene representation, which for each of the \\(M\\) objects includes the appearance and shape MLPs \\((\\mathbf{MLP}_{c,j}, \\mathbf{MLP}_{\\sigma,j})\\), rigid-body transformations \\(\\mathbf{G}_j^t\\), and forward, backward deformation fields \\(\\mathbf{J}_j^{\\leftarrow}\\), \\(\\mathbf{J}_j^{\\rightarrow}\\). We optimize three reconstruction losses: a color loss \\(\\mathcal{L}_{\\mathrm{rgb}}\\), a flow loss \\(\\mathcal{L}_{\\mathrm{flow}}\\), and a depth loss \\(\\mathcal{L}_{\\mathrm{depth}}\\), where the ground truth color \\(\\mathbf{c}\\) and depth \\(\\mathbf{d}\\) are provided by the RGBD video, and the \"ground truth\" flow \\(\\mathcal{F}\\) is computed by an off-the-shelf network [64]. The model also optimizes a 3D-cycle consistency loss \\(\\mathcal{L}_{\\mathrm{cyc},j}\\) [67] for each deformable object to encourage their forward and backward warps to be consistent, where \\(\\mathbf{x}^t \\in \\mathbb{R}^2\\) denotes pixel location at time \\(t\\):" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17675" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.1, + 0.117, + 0.469, + 0.149 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r g b}} = \\sum_ {\\mathbf {x} ^ {t}} \\left| \\left| \\mathbf {c} \\left(\\mathbf {x} ^ {t}\\right) - \\hat {\\mathbf {c}} \\left(\\mathbf {x} ^ {t}\\right) \\right| \\right| ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.152, + 0.469, + 0.184 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f l o w}} = \\sum_ {\\mathbf {x} ^ {t}} \\left| \\left| \\mathcal {F} \\left(\\mathbf {x} ^ {t}\\right) - \\hat {\\mathcal {F}} \\left(\\mathbf {x} ^ {t}\\right) \\right| \\right| ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.187, + 0.469, + 0.219 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e p t h}} = \\sum_ {\\mathbf {x} ^ {t}} \\left| \\left| \\mathbf {d} \\left(\\mathbf {x} ^ {t}\\right) - \\hat {\\mathbf {d}} \\left(\\mathbf {x} ^ {t}\\right) \\right| \\right| ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.222, + 0.469, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c y c}, j} = \\sum_ {i} \\tau_ {i} \\alpha_ {i j} \\left\\| \\mathcal {W} _ {j} ^ {t ^ {\\prime}, \\rightarrow} \\left(\\mathcal {W} _ {j} ^ {t}, \\leftarrow (\\mathbf {X} _ {i} ^ {t})\\right) - \\mathbf {X} _ {i} ^ {t} \\right\\| ^ {2}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.47, + 0.352 + ], + "angle": 0, + "content": "Initialization. We initialize the rigid-body transformations of each deformable object \\(\\mathbf{G}_j^t\\) using a pre-trained category-specific PoseNet [67]; we initialize the rigid-body transformation of the background \\(\\mathbf{G}_0^t\\) with the camera poses provided by the iPad Pro." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.473, + 0.573 + ], + "angle": 0, + "content": "Embodied View Synthesis and 3D Filters. To enable embodied view synthesis and 3D video filters (Figure 3), we design a simple interface that allows a user to select a point on a target object's surface in its reconstructed canonical mesh, and use its forward warping function \\(\\mathcal{W}_j^{t,\\rightarrow}:\\mathbf{X}^* \\to \\mathbf{X}^t\\) followed by the rigid-body transformation \\(\\mathbf{G}_0^t\\) to place the egocentric camera (or virtual 3D asset) in the world space. The surface normal to the object's mesh at the user-defined point provides a reference frame to align the egocentric camera's viewing direction and place the 3D asset. To implement a 3rd-person-follow camera, we add a user-defined offset to the object's local reference frame, which is defined by its root-body pose." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.588, + 0.21, + 0.604 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Implementation Details. In practice, we train our composite scene representation by first pre-training each object field separately. For deformable objects, we pre-train using a depth loss (Equation 10) combined with the losses optimized by BANMo [67]. This includes a silhouette loss \\(\\mathcal{L}_{\\mathrm{mask}} = \\sum_{\\mathbf{x}^t}||\\mathbf{o}_j(\\mathbf{x}^t) - \\hat{\\mathbf{o}}_j(\\mathbf{x}^t)||^2\\), where the \"ground truth\" object silhouette \\(\\mathbf{o}_j\\) is computed by an off-the-shelf instance segmentation engine [17]. For pre-training the background, we optimize color, flow, and depth losses (Equations 8, 9, 10) on pixels outside the ground truth object silhouettes. Importantly, we don't supervise the object fields on frames that are not provided an object silhouette since it cannot be determined whether the absence of detection is a true or false negative." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "After pre-training, we composite-render the pre-trained object fields and jointly finetune them using only the color, depth, flow, and object-specific 3D-cycle consistency losses. Since the silhouette loss is no longer used, the scene representation is supervised on all frames of the training se" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "quence during joint-finetuning. We provide a complete description of the implementation details in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.413 + ], + "angle": 0, + "content": "Dataset. We evaluate Total-Recon on novel-view synthesis for deformable scenes. To enable quantitative evaluation, we built a stereo rig comprised of two iPad-Pros rigidly attached to a camera mount, a setup similar to that of Nerfies [35]. Using the stereo rig, we captured 11 RGBD sequences containing 3 different cats, 1 dog, and 2 human subjects in 4 different indoor environments. The RGBD videos were captured using the Record3D iOS App [49], which also automatically registers the frames captured by each camera. These video sequences, which were subsampled at 10 fps, range from 392 to 901 frames, amount to, on average minute-long videos that are significantly longer and contain more dynamic motion than the datasets introduced by [35, 36, 60, 11]. The left and right cameras were registered by solving a Perspective-n-Point (PnP) problem using manually annotated correspondences, and their videos were synchronized based on audio. We provide a complete description of our dataset in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Reconstruction and Applications. By hierarchically decomposing scene motion into the motion of each object, which itself is decomposed into root-body motion and local articulations, Total-Recon automatically computes novel 6-DoF trajectories such as those traversed by egocentric cameras and 3rd-person follow cameras (Figure 3). In turn, these trajectories enable automated embodied view synthesis and 3D occlusion-aware video filters (Figure 4). These tasks are also enabled by Total-Recon's ability to recover an accurate deformable 3D scene representation, which is currently out of reach for the best of related methods (Figure 5). As shown in the bird's eye view, each reconstructed object is properly situated with respect to the background and other objects, a direct consequence of our use of depth supervision. Furthermore, even though the iPad Pro can only measure depth up to \\(4\\mathrm{m}\\), Total-Recon can render depth beyond this sensor limit by pooling the measurements from other frames into a single metric scene reconstruction. We provide results on additional sequences in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Baselines and Evaluation. In Figure 5 and Table 3, we compare Total-Recon to \\(\\mathrm{D}^2\\mathrm{NeRF}\\) [60] and HyperNeRF [36], and their depth-supervised equivalents on the proxy task of stereo-view synthesis, a prerequisite for embodied view synthesis: we train each method on the RGBD frames captured from the left camera of our dataset and evaluate the images rendered from the viewpoint of the right camera. The depth-supervised versions of the baselines contain the same depth loss used in Total-Recon. We report LPIPS [70] and the average (depth) accuracy at \\(0.1\\mathrm{m}\\) [42] in all subsequent experiments, and we include a more complete set" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17676" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.874, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.622, + 0.895, + 0.693 + ], + "angle": 0, + "content": "Figure 4: Embodied View Synthesis and 3D Filters. For select sequences of our RGBD dataset, we visualize the scene geometry and appearance reconstructed by our method (3D reconstruction) and the resulting downstream applications. The yellow and blue camera meshes in the mesh renderings represent the egocentric and 3rd-person-follow cameras, respectively. To showcase the 3D video filter we attach a sky-blue unicorn horn to the forehead of the target object, which is then automatically propagated across all frames. Full-length videos can be found at https://andrewsonga.github.io/totalrecon/applications.html." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.473, + 0.79 + ], + "angle": 0, + "content": "of metrics (PSNR, SSIM, RMS depth error) in the supplement. Because \\(\\mathrm{D}^2\\mathrm{NeRF}\\) and HyperNeRF were not designed to recover a metric scene representation, we replaced their COLMAP [46] camera poses with those provided by the iPad Pro (which are metric measurements) for the sake of fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Comparisons. Total-Recon qualitatively and quantitatively outperforms all of the baselines. As shown in Figure 5, Total-Recon successfully reconstructs the entire scene, whereas the baselines are only able to reconstruct the rigid background at best. As shown in Table 3, Total-Recon significantly outperforms all baselines in terms of LPIPS and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.895, + 0.763 + ], + "angle": 0, + "content": "the average accuracy at \\(0.1\\mathrm{m}\\) (Acc@0.1m). We attribute this huge gap to the baselines' inability to reconstruct highly dynamic objects. We provide more details regarding the baselines and additional visualizations in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.666, + 0.787 + ], + "angle": 0, + "content": "4.1. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Table 4 (and Figures 6 and 7) analyzes the importance of Total-Recon's design choices (see Section 3) by ablating its key components: the depth loss \\(\\mathcal{L}_{\\mathrm{depth}}\\) (row 2), the deformation field \\(\\mathbf{J}_j^t\\) (row 3), PoseNet-initialization of the root-body pose (row 4), and the root-body pose \\(\\mathbf{G}_j^t\\) itself (row 5), where \\(j\\) denotes a deformable actor. For all ablations, we use the same set of training losses used in Total-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17677" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.09, + 0.887, + 0.176 + ], + "angle": 0, + "content": "
DOG 1 (626 images)DOG 1 (v2) (531 images)CAT 1 (641 images)CAT 1 (v2) (632 images)CAT 2 (632 images) (834 images)CAT 2 (v2) (901 images)CAT 3 (767 images)HUMAN 1 (550 images)HUMAN 2 (483 images)HUMAN - DOG (392 images)HUMAN - CAT (431 images)MEAN
LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑
HyperNeRF.634.107.432.176.521.316.438.314.641.277.397.252.592.213.632.053.585.067.487.072.462.162.531.198
D2NeRF.540.219.546.220.687.346.588.403.556.333.595.339.759.231.588.066.630.128.576.078.628.126.611.247
HyperNeRF (+depth).373.352.425.357.532.552.371.596.330.605.376.612.514.451.501.211.445.249.450.283.456.214.428.439
D2NeRF (+depth).507.338.532.270.685.510.580.362.561.438.553.376.730.243.585.086.609.131.608.154.645.176.599.302
Total-Recon.271.841.313.790.382.889.333.894.237.967.281.925.261.949.213.909.264.849.256.827.233.914.278.895
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.187, + 0.895, + 0.232 + ], + "angle": 0, + "content": "Table 3: Baseline Comparisons. We train Total-Recon, HyperNeRF [36], D²NeRF [60], and their depth-supervised variants on the left video captured with our stereo rig, and evaluate the novel view synthesis results on the held-out right video. Total-Recon significantly outperforms all of the baselines for all 11 sequences. These sequences are sampled at 10 fps, amounting to minute-long videos, on average." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.258, + 0.468, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.499, + 0.472, + 0.598 + ], + "angle": 0, + "content": "Figure 5: Baseline Comparisons. We compare Total-Recon to depth-supervised variants of HyperNeRF [36] and \\(\\mathrm{D}^2\\mathrm{NeRF}\\) [60] on the task of stereo-view synthesis (the left camera is used for training and the right is used for testing). While the baselines are able to reconstruct only the background, Total-Recon can reconstruct both the background and the moving deformable object(s), demonstrating holistic scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.47, + 0.753 + ], + "angle": 0, + "content": "Recon and initialize camera pose \\(\\mathbf{G}_0^t\\) with those reported by ARKit. For ablations that model root-body motions, we initialize each deformable actor's root-body pose \\(\\mathbf{G}_j^t\\) with predictions made by PoseNet [67] and optimize them during reconstruction; for row 4, we replace the PoseNet predictions with identity rotations. We report the novel-view metrics averaged over 6 select sequences of our dataset: DOG 1 (v1), CAT 1 (v1), CAT 2 (v1), HUMAN 1, HUMAN 1 & DOG 1, and HUMAN 2 & CAT 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Depth Supervision. Table 4 shows that removing depth supervision (row 2) significantly reduces the average accuracy at \\(0.1\\mathrm{m}\\) (Acc). Figure 6 indicates that this reflects the incorrect arrangement of objects stemming from their scale inconsistency - while removing depth supervision does not significantly deteriorate the training-view RGB renderings, it induces critical failure modes as shown in the novel-view 3D reconstructions: (a) floating foreground objects, as evi" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.253, + 0.882, + 0.346 + ], + "angle": 0, + "content": "
MethodsDepth LossDeform. Obj.Root Init.Root MotionLPIPS↓Acc@0.1m↑
(1) Full model✓✓✓✓.268.898
(2) w/o loss LdepthX✓✓✓.372.154
(3) w/o deform. Jj✓X✓✓.296.867
(4) w/o root-body init.✓✓X✓.293.870
(5) w/o root-body Gj✓✓XXN/AN/A
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.357, + 0.892, + 0.496 + ], + "angle": 0, + "content": "Table 4: Ablation Study. Removing depth supervision (2) significantly hurts performance, while removing the deformation field (3) and PoseNet-initialization of root-body poses (4) hurts moderately. Most importantly, removing root-body poses entirely (5) prevents convergence (N/A) as the deformation field alone has to explain global object motion (see Figure 2). These experiments justify our hierarchical modeling of motion, as even root-bodies without a deformation field (3) or poorly initialized root-bodies (4) are better than no root-bodies (5). We visualize these ablations in Figure 7 and explore other ablations in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.526, + 0.892, + 0.618 + ], + "angle": 0, + "content": "denced by their shadows, and (b) the human incorrectly occluding the dog. In other words, without depth supervision, Total-Recon overfits the training view and learns a degenerate scene representation where the reconstructed objects fail to converge to the same scale. We show results on additional sequences in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Motion Modeling. Table 4 shows that removing the deformation field (row 3) also hurts performance. This is because, without the deformation field, our method has to explain an object's non-rigid motion solely with its rigid, root-body poses. As a result, this ablation can only recover coarse object reconstructions that fail to model moving body parts such as limbs. Removing PoseNet-initialization of root-body poses (row 4) is just as detrimental, resulting in noisy appearance and geometry artifacts; see Figure 7 and additional visuals in the supplement. Most notably, Table 4 shows that removing object root-bodies entirely (row 5) causes the optimization to fail to converge (N/A), even though the deformation field can (in theory) represent all continuous motion. It appears difficult for deformation fields alone to explain global root-body motion because such motions can deviate significantly from a canonical model, complicating optimization." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17678" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.468, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.47, + 0.378 + ], + "angle": 0, + "content": "Figure 6: Ablation Study on Depth Supervision. While removing depth supervision does not significantly deteriorate the training-view RGB renderings, it significantly hurts the novel-view 3D reconstructions, as characterized by the following: (a) floating foreground objects (as evidenced by their shadows) and (b) the human incorrectly occluding the dog. These failure modes indicate that without depth supervision, Total-Recon overfits the training view, and the reconstructed objects fail to converge to the same scale." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.725 + ], + "angle": 0, + "content": "Figure 7: Ablation Study. We visualize the ablations from Table 4. Removing the depth loss causes the cat to sink into the ground due to inconsistent object scales. Removing the deformation field produces coarse object reconstructions that fail to capture articulated body parts such as moving limbs. Removing PoseNet-initialization of root-body poses results in noisier appearance and geometry, as shown by the human actor's left hand. We omit the ablation without root-body poses as it does not converge. We present additional visualizations in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "These diagnostics justify Total-Recon's hierarchical motion representation, which explicitly models objects' root-body poses; root-bodies without articulated deformations (row 3) or poorly initialized root-bodies (row 4) are better than no root-bodies at all (row 5). Our ablations also suggest that the poor performance of the baseline methods (on our challenging dataset) may be attributed to the lack of object-centric motion modeling. We provide a more detailed analysis with additional experiments and RGBD sequences in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.753, + 0.107 + ], + "angle": 0, + "content": "5. Discussion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.895, + 0.328 + ], + "angle": 0, + "content": "We have presented a new system for automated embodied view synthesis from monocular RGBD videos, focusing on videos of people interacting with their pets. Our main technical contribution is Total-Recon, a 3D representation for deformable scenes that hierarchically decomposes scene motion into the motion of each object, which is further decomposed into its root-body motion and local articulations; this key design choice enables appropriate initialization of root-body poses and hence easier optimization over long videos containing large motions. By explicitly reconstructing the geometry, appearance, root-body- and articulated motion of each object, Total-Recon enables seeing through the eyes of people and pets and generating game-like traversals of deformable scenes from behind a target actor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.895, + 0.621 + ], + "angle": 0, + "content": "Limitations. In Total-Recon, scene decomposition is primarily supervised by object silhouettes computed by an off-the-shelf segmentation model [17], which may be inaccurate, especially in partial occlusion scenarios. This may damage the resulting reconstructions and embodied-view renderings. We believe that incorporating the latest advances in video instance segmentation [12] will enable Total-Recon to be applied to more challenging scenarios. Second, Total-Recon initializes the root-body pose of each deformable object using a PoseNet [67] trained for humans and quadruped animals, which does not generalize to other object categories (e.g., birds, fish). We reserve the reconstruction of generic scenes for future work. Finally, our model needs to be optimized on a per-sequence basis for roughly 15 hours with 4 NVIDIA RTX A5000 GPUs and is therefore not suitable for real-time applications. Incorporating recent advances in fast neural field training methods is an interesting avenue for future research." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.894, + 0.763 + ], + "angle": 0, + "content": "Acknowledgments. We thank Nathaniel Chodosh, Jeff Tan, George Cazenavette, and Jason Zhang for proofreading and Songwei Ge for reviewing the codebase. We also thank Sheng-Yu Wang, Daohan (Fred) Lu, Tamaki Kojima, Krishna Wadhwani, Takuya Narihira, and Tatsuo Fujiwara for providing valuable feedback. This work is supported in part by the Sony Corporation, Cisco Research, and the CMU Argo AI Center for Autonomous Vehicle Research." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.777, + 0.597, + 0.792 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.802, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.859, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[2] Aljaz Bozic, Pablo Palafox, Michael Zollhofer, Justus Thies, Angela Dai, and Matthias Nießner. Neural deformation graphs for globally-consistent non-rigid reconstruction. In" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.802, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17679" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.121, + 0.47, + 0.188 + ], + "angle": 0, + "content": "[3] Aljaz Bozic, Michael Zollhofer, Christian Theobalt, and Matthias Nießner. Deepdeform: Learning non-rigid rgb-d reconstruction with semi-supervised data. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.19, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[4] Hongrui Cai, Wanquan Feng, Xuetao Feng, Yan Wang, and Juyong Zhang. Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.247, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.344, + 0.471, + 0.412 + ], + "angle": 0, + "content": "[7] Alena Denisova and Paul Cairns. First person vs. third person perspective in digital games: Do player preferences affect immersion? In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, page 145-148, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.413, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[8] Mingsong Dou, Jonathan Taylor, Henry Fuchs, Andrew Fitzgibbon, and Shahram Izadi. 3d scanning deformable objects with a single rgbd sensor. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.469, + 0.471, + 0.537 + ], + "angle": 0, + "content": "[9] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbaek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[10] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[11] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[12] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[13] Alec Jacobson, Zhigang Deng, Ladislav Kavan, and J. P. Lewis. Skinning: Real-time shape deformation (full text not available). In ACM SIGGRAPH 2014 Courses, SIGGRAPH '14. Association for Computing Machinery, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[14] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[15] Zhang Jiakai, Liu Xinhang, Ye Xinyi, Zhao Fuqiang, Zhang Yanshun, Wu Minye, Zhang Yingliang, Xu Lan, and Yu Jingyi. Editable free-viewpoint video using a layered neural representation. In ACM SIGGRAPH, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "from a single video. In European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[17] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[18] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J. Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[19] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[20] Tianye Li, Mira Slavcheva, Michael Zollhöfer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[21] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[22] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[23] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM SIGGRAPH Asia, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[25] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[26] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[27] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17680" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[30] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE International Symposium on Mixed and Augmented Reality, pages 127-136. IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[31] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.329 + ], + "angle": 0, + "content": "[32] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[33] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.471, + 0.441 + ], + "angle": 0, + "content": "[34] Kemal Egemen Ozden, Kurt Cornelis, Luc Van Eycken, and Luc Van Gool. Reconstructing 3d trajectories of independently moving objects using generic constraints. Comput. Vis. Image Underst., 96(3):453-471, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[35] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[36] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 40(6),2021.2,3,4,6,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.65 + ], + "angle": 0, + "content": "[37] Georgios Pavlakos*, Ethan Weber*, , Matthew Tancik, and Angjoo Kanazawa. The one where they reconstructed 3d humans and environments in tv shows. In European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.719 + ], + "angle": 0, + "content": "[38] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[39] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[40] F. Peruzzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "for Dynamic Scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[42] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[43] Zhongzheng Ren, Xiaoming Zhao, and Alex Schwing. Class-agnostic reconstruction of dynamic objects from videos. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[44] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[45] Richard Rouse. What's your perspective? ACM SIGGRAPH Computer Graphics, 33(3):9-12, August 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[46] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[47] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[48] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In SIGGRAPH Conference Proceedings, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[49] Marek Simonik. Record3d. https://record3d.app. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.604 + ], + "angle": 0, + "content": "[50] Miroslava Slavcheva, Maximilian Baust, Daniel Cremers, and Slobodan Ilic. Killingfusion: Non-rigid 3d reconstruction without correspondences. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.605, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[51] Miroslava Slavcheva, Maximilian Baust, and Slobodan Ilic. Sobolevfusion: 3d reconstruction of scenes undergoing free non-rigid motion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[52] Karl Stelzner, Kristian Kersting, and Adam R Kosiorek. Decomposing 3d scenes into objects via unsupervised volume segmentation. arXiv preprint arXiv:2104.01148, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[53] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[54] Unity Technologies. 3rd person follow. https://docs.unity3d.com/Packages/ com.unity.cinemachine@2.8/manual/ Cinemachine3rdPersonFollow.html.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17681" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[56] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.175 + ], + "angle": 0, + "content": "[57] Barbara Tversky. Spatial Cognition Embodied and Situated. Cambridge, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.178, + 0.468, + 0.217 + ], + "angle": 0, + "content": "[58] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.219, + 0.468, + 0.272 + ], + "angle": 0, + "content": "[59] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.275, + 0.468, + 0.342 + ], + "angle": 0, + "content": "[60] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. \\(\\mathbf{D}^2\\) nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.345, + 0.468, + 0.398 + ], + "angle": 0, + "content": "[61] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.401, + 0.468, + 0.453 + ], + "angle": 0, + "content": "[62] Jing Xiao, Jinxiang Chai, and Takeo Kanade. A closed-form solution to non-rigid shape and motion recovery. International Journal of Computer Vision (IJCV), 67(2):233-246, April 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.456, + 0.468, + 0.509 + ], + "angle": 0, + "content": "[63] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.512, + 0.468, + 0.551 + ], + "angle": 0, + "content": "[64] Gengshan Yang and Deva Ramanan. Volumetric correspondence networks for optical flow. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.554, + 0.468, + 0.633 + ], + "angle": 0, + "content": "[65] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T Freeman, and Ce Liu. LASR: Learning articulated shape reconstruction from a monocular video. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.637, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[66] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.707, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[67] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6, 8, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[68] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[69] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.874, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[70] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[71] Xiaoshuai Zhang, Abhijit Kundu, Thomas Funkhouser, Leonidas Guibas, Hao Su, and Kyle Genova. Nerflets: Local radiance fields for efficient structure-aware 3d scene representation from 2d supervision. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17682" + } + ] +] \ No newline at end of file diff --git a/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/49fdb57e-3f25-4f0f-85ac-577ce80d4f48_origin.pdf b/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/49fdb57e-3f25-4f0f-85ac-577ce80d4f48_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6906e79dd2a63e20726088773325b69217525e50 --- /dev/null +++ b/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/49fdb57e-3f25-4f0f-85ac-577ce80d4f48_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45431ba162436409a77a46a4368f93ecbe9be27c516b38978da631334be20a7c +size 5245337 diff --git a/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/full.md b/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..50e874e6254968be328424713c0ab41df42c70bd --- /dev/null +++ b/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/full.md @@ -0,0 +1,298 @@ +# Total-Recon: Deformable Scene Reconstruction for Embodied View Synthesis + +Chonghyuk Song Gengshan Yang Kangle Deng Jun-Yan Zhu Deva Ramanan Carnegie Mellon University + +![](images/baeec2e9755f02e3f6586435c235cbe25ffdfdc9edf954fd3dac8ff708f9f7d2.jpg) +Figure 1: Embodied View Synthesis. Given a long video of deformable objects captured by a handheld RGBD sensor, Total-Recon renders the scene from novel camera trajectories derived from the in-scene motion of actors: (1) egocentric cameras that simulate the point-of-view of a target actor (such as the pet) and (2) 3rd-person (or pet) cameras that follow the actor from behind. Our method also enables (3) 3D video filters that attach virtual 3D assets to the actor. Total-Recon achieves this by reconstructing the geometry, appearance, root-body- and articulated motion of each deformable object in the scene and the background. + +# Abstract + +We explore the task of embodied view synthesis from monocular videos of deformable scenes. Given a minute-long RGBD video of people interacting with their pets, we render the scene from novel camera trajectories derived from the in-scene motion of actors: (1) egocentric cameras that simulate the point of view of a target actor and (2) 3rd-person cameras that follow the actor. Building such a system requires reconstructing the root-body and articulated motion of every actor, as well as a scene representation that supports free-viewpoint synthesis. Longer videos are more likely to capture the scene from diverse viewpoints (which helps reconstruction) but are also more likely to contain larger motions (which complicates reconstruction). To address these challenges, we present Total-Recon, the first method to photorealistically reconstruct deformable scenes from long monocular RGBD videos. Crucially, to scale to long videos, our method hierarchically decomposes the scene into the background and objects, whose motion is decomposed into carefully initialized root-body motion and local articulations. To quantify such "in-the-wild" reconstruction and view synthesis, we collect ground-truth data from a specialized stereo RGBD capture rig for 11 challenging videos, significantly outperforming prior methods. + +# 1. Introduction + +We explore embodied view synthesis, a new class of novel-view synthesis tasks that renders deformable scenes from novel 6-DOF trajectories reconstructed from the in-scene motion of actors: egocentric cameras [45, 7] that simulate the point-of-view of moving actors and 3rd-person-follow cameras [54, 7] that track a moving actor from behind (Figure 1). We focus on everyday scenes of people interacting with their pets, producing renderings from the point-of-view of the person and pet (Figure 1). While such camera trajectories could be manually constructed (e.g., by artists via keyframing), building an automated system is an interesting problem of its own: spatial cognition theory [57] suggests that the ability to visualize behavior from another actor's perspective is necessary for action learning and imitation; in the context of gaming and virtual reality [7, 45], egocentric cameras offer high levels of user immersion, while 3rd-person-follow cameras provide a large field of view that is useful for exploring a user's environment. + +Challenges. Building a system for embodied view synthesis is challenging for many reasons. First, to reconstruct everyday-but-interesting content, it needs to process long, monocular captures of multiple interacting actors. How- + +![](images/5e8218f59c1037fb89cfa0dc8f68b116edcc85c26a28a3c8133ea47da4a8eb54.jpg) +Figure 2: Method Overview. Total-Recon represents the entire scene as a composition of $M$ object-centric neural fields, one for the rigid background and each of the $M - 1$ deformable objects. To render a scene, (1) each object field $j$ is transformed into the camera space with a rigid transformation $\left(\mathbf{G}_j^t\right)^{-1}$ that encodes root-body motion and, for each deformable object, an additional deformation field $\mathbf{J}_j^{t,\rightarrow}$ that encodes articulated motion. Next, all (2) posed object fields are combined into a (3) composite field, which is then volume-rendered into (4) color, depth, optical flow, and object silhouettes. Each rendered output defines a reconstruction loss that derives supervision from a monocular RGBD video captured by a moving iPad Pro. + +ever, such videos are likely to contain large scene motions, which we demonstrate are difficult to reconstruct with current approaches. Second, it needs to produce a deformable 3D scene representation that supports free-viewpoint synthesis, which also would benefit from long videos likely to capture the scene from diverse viewpoints. Recent approaches have extended Neural Radiance Fields (NeRFs) [28] to deformable scenes, but such work is often limited to rigid-only object motion [18, 33], short videos with limited scene motion [41, 35, 21, 55, 36, 60, 10, 61, 58], or reconstructing single objects as opposed the entire scene [65, 66, 67, 4]. Third, it needs to compute global 6-DOF trajectories of root-bodies and articulated body parts (e.g., head) of multiple actors. + +Key Ideas. To address these challenges, we introduce Total-Recon, the first monocular NeRF that enables embodied view synthesis for deformable scenes with large motions. Given a monocular RGBD video, Total-Recon reconstructs the scene as a composition of object-centric representations, which encode the 3D appearance, geometry, and motion of each deformable object and the background. Crucially, Total-Recon hierarchically decomposes scene motion into the motion of individual objects, which itself is decomposed into global root-body movement and the local deformation of articulated body parts. We demonstrate that such decomposition of object motion, along with appropriate initialization of root-body pose, allows reconstruction to scale to longer videos, enabling free-viewpoint synthesis. By reconstructing such motions in a globally-consistent + +coordinate frame, Total-Recon can generate renderings from egocentric and 3rd-person-follow cameras, as well as static but extreme viewpoints like bird's-eye-views. + +Evaluation. Due to the difficulty of collecting ground-truth data for embodied view synthesis on in-the-wild videos, we evaluate our method on the proxy task of stereo-view synthesis [35], which compares rendered views to those captured from a stereo pair. To this end, we build a stereo RGBD sensor capture rig for ground-truthing and collect a dataset of 11 long video sequences in various indoor environments, including people interacting with their pets. Total-Recon outperforms the state-of-the-art monocular deformable NeRF methods [36, 60], even when modified to use depth sensor measurements. + +Contributions. In summary, our contributions are: (1) Total-Recon, a hierarchical 3D representation that models deformable scenes as a composition of object-centric representations, each of which decomposes object motion into its global root-body motion and its local articulations; (2) a system based on Total-Recon for automated embodied view synthesis from casual, minute-long RGBD videos of highly dynamic scenes; (3) a dataset of stereo RGBD videos containing various deformable objects, such as humans and pets, in a host of different background environments. Our code, models, and data can be found at https://andrewsonga.github.io/totalrecon. + +
MethodEntire ScenesDeform. ObjectsBeyond HumansGlobal 6-DOF Traj.Long VideosExtreme Views
BANMo [67]X✓✓X✓✓
PNF [18]✓X✓✓XX
NeuMan [16]✓✓X✓XX
SLAHMR [69]X✓X✓XX
HyperNeRF [36]✓✓✓XXX
D2NeRF [60]✓✓✓XXX
DynIBaR [22]✓✓✓X✓X
SUDS [56]✓✓✓X✓X
Ours✓✓✓✓✓✓
+ +Table 1: Comparison to Related Work. Unlike prior work, Total-Recon targets embodied view synthesis of scenes containing humans and pets, requiring the ability to (1) reconstruct entire scenes, (2) model deformable objects, (3) extend beyond humans, (4) recover global 6-DOF trajectories of objects' root-bodies and their articulated parts, (5) process minute-long videos of dynamic scenes, and (6) render extreme views. + +# 2. Related Work + +Neural Radiance Fields. Prior works on Neural Radiance Fields (NeRF) optimize a continuous scene function for novel view synthesis given a set of multi-view images, usually under the assumption of a rigid scene and densely sampled views [28, 26, 27, 23, 14, 59]. DS-NeRF [6] and Dense Depth Priors [44] extend NeRFs to the sparse-view setting by introducing depth as additional supervision. Total-Recon also operates in the sparse-view regime and uses depth supervision to reduce the ambiguities inherent to monocular, multibody, non-rigid reconstruction [62, 34]. Another line of work [18, 33] represents rigidly moving scenes as a composition of multiple object-level NeRFs. Total-Recon also leverages such an object-centric scene representation, but models scenes containing non-rigidly moving objects, such as humans and pets. + +Deformable NeRFs. Recent approaches extend NeRF to monocular deformable scene reconstruction either by learning an additional function that deforms observed points in the camera space to a time-independent canonical space [41, 35, 55, 36, 60] or explicitly modeling density changes over time [10, 61, 58, 21]. Such methods are typically limited to short videos containing little scene and camera motion. They also perform novel-view synthesis only over small baselines. Total-Recon belongs to the former category of prior monocular deformable NeRFs, but unlike them, our method hierarchically decomposes scene motion into the motion of each object, which is further decomposed into global root-body motion and local articulations. The proposed motion decomposition is what enables embodied view synthesis: it allows Total-Recon to scale to minute-long videos and reconstruct a deformable 3D scene representation that supports free-viewpoint syn + +thesis; it also makes it easy to extract an object's root-body motion, the key motion primitive required for 3rd-person-follow view synthesis. Several works have taken different approaches to making non-rigid reconstruction more tractable. One group of work leverages human-specific priors [38, 53, 32, 39, 24, 16, 19, 37] such as human body models (e.g., SMPL), 3D skeletons, or 2D poses to achieve high reconstruction quality. We achieve similar levels of fidelity without relying on such shape priors, allowing Total-Recon to generalize to pets and, by extension, reconstruct human-pet interaction videos. Another body of work [15, 48, 20] achieves high-fidelity scene reconstructions by relying on synchronized multi-view video captured from a specialized camera rig ranging from 8 to 18 static cameras. In contrast, Total-Recon only requires a single video captured from a moving RGBD camera equipped with inertial sensors, which has now become widely accessible in consumer products with the advent of Apple's iPhone and iPad Pro. + +Reconstruction with RGBD Sensors. Depth sensors represent the third class of attempts to make non-rigid reconstruction more tractable, reducing the need for a predefined shape template. Kinect-fusion [30] creates a real-time system for indoor scene localization and mapping. Dynamic Fusion [29] builds a template-free dense SLAM system for dynamic objects. Later works improve RGBD reconstruction to be able to deal with topology changes [50, 51] and use correspondence matching for registration over large motions [2, 3]. Recent works have incorporated neural implicit representations to reconstruct the surface geometry and 3D motion fields for deformable objects [43, 4] or largescale rigid scenes [1, 42] in isolation. Other works have reconstructed humans alongside small-scale objects and furniture [8, 2], but not the entire background. We aim to go even further by reconstructing the entire scene, which includes the background and multiple deformable targets such as humans and pets; not only do we reconstruct the geometry, but we also recover a radiance field that allows for photorealistic scene rendering from embodied viewpoints and other novel 6-DOF trajectories. + +Concurrent Work. Concurrent work exhibits a subset of the design choices necessary for embodied view synthesis. SLAHMR [69] reconstructs the geometry and in-scene motion of human actors but not the scene appearance. Nerflets [71] models the appearance, geometry, and motion of each scene element but is limited to rigidly moving objects. RoDynRF [25], NeRF-DS [63], HexPlane [5], and K-planes [9] reconstruct other types of dynamic scene elements, such as deformable or specular objects, but these methods have been demonstrated on only short videos and/or videos containing limited object root-body motion [40, 47, 63, 11]. DynIBaR [22] scales dynamic view synthesis to longer + +videos with complex camera and scene motion, and SUDS [56] scales reconstruction to urban-scale dynamic scenes captured from 1.2 million frames. However, neither demonstrates extreme-view synthesis, a prerequisite for rendering embodied views. We summarize and compare prior work to Total-Recon in Table 1. + +# 3. Method + +# 3.1. Limitations of Prior Art + +The state-of-the-art monocular deformable NeRFs [36, 60] decompose a deformable scene into a rigid, canonical template model and a deformation field $\mathbf{J}^{t,\leftarrow}$ that maps the world space $\mathbf{G}_0^t\mathbf{X}^t$ to the canonical space $\mathbf{X}^*$ , where $\mathbf{G}_0^t$ is the known camera pose at time $t$ , and $\mathbf{X}^t$ is a camera space point at time $t$ : + +$$ +\mathbf {X} ^ {*} = \mathcal {W} ^ {t, \leftarrow} \left(\mathbf {X} ^ {t}\right) = \mathbf {J} ^ {t, \leftarrow} \left(\mathbf {G} _ {0} ^ {t} \mathbf {X} ^ {t}\right). \tag {1} +$$ + +In theory, this formulation is sufficient to represent all continuous motion; it performs well on short videos containing near-rigid scenes, as the deformation field only has to learn minute deviations from the template model. However, this motion model is difficult to scale to minute-long videos, which are more likely to contain deformable objects undergoing large translations (e.g., a person walking into another room) and pose changes (e.g., a person sitting down). Here, the deformation field must learn large deviations from the canonical model, significantly complicating optimization. + +Another critical limitation of HyperNeRF and $\mathrm{D}^2\mathrm{NeRF}$ is that they cannot track separate deformable objects and therefore cannot perform 3rd-person-follow view synthesis for scenes with multiple actors. + +# 3.2. Component Radiance Fields + +To address the limitations of existing monocular deformable NeRFs, we propose Total-Recon, a novel 3D representation that models a deformable scene as a composition of $M$ object-centric neural fields, one for the rigid background and each of the $M - 1$ deformable objects (Figure 2). Crucially, Total-Recon hierarchically decomposes scene motion into the motion of each object, which itself is decomposed into global root-body motion and local articulations. This key design choice scales our method to minute-long videos containing highly dynamic and deformable objects. + +Background Radiance Field. We begin by modeling the background environment as a Neural Radiance Field (NeRF) [28]. For a 3D point $\mathbf{X}^{*} \in \mathbb{R}^{3}$ and a viewing direction $\mathbf{v}^{*}$ in the canonical world space, NeRF defines a color $\mathbf{c}$ and density $\sigma$ represented by an MLP. We follow contemporary variants [26] that include a time-specific embedding code $\omega_{e}^{t}$ to model illumination changes over time + +and model density with as a function of a neural signed distance function (SDF) $\mathbf{MLP}_{\sigma}(\cdot) = \alpha \Gamma_{\beta}(\mathbf{MLP}_{\mathrm{SDF}}(\cdot))$ [68] to encourage the reconstruction of a valid surface: + +$$ +\sigma = \mathbf {M L P} _ {\sigma} \left(\mathbf {X} ^ {*}\right), \quad \mathbf {c} ^ {t} = \mathbf {M L P} _ {\mathbf {c}} \left(\mathbf {X} ^ {*}, \mathbf {v} ^ {*}, \omega_ {e} ^ {t}\right). \tag {2} +$$ + +The pixel color can then be computed with differentiable volume rendering equations (Section 3.3). + +Most NeRF methods, including HyperNeRF [36] and $\mathrm{D}^2\mathrm{NeRF}$ [60], assume images with known cameras. While our capture devices are equipped with inertial sensors, we find their self-reported camera poses have room for improvement. As such, we also model camera pose as an opti- mizable rigid-body transformation $\mathbf{G}_0^t\in SE(3)$ that maps points in a time-specific camera space $\mathbf{X}^t\in \mathbb{R}^3$ to the world space (where we assume homogenous notation): + +$$ +\mathbf {X} ^ {*} = \mathbf {G} _ {0} ^ {t} \mathbf {X} ^ {t}. \tag {3} +$$ + +Deformable Field (for Object $j$ ). We model the deformable radiance field of object $j \in \{1, \dots, M - 1\}$ with BANMo [67], which consists of a canonical rest shape and time-dependent deformation field. The canonical rest shape is represented by the same formulation described by Equation 2, but now defined in a local object-centric canonical space rather than the world space. BANMo represents object motion with a warping function $\mathcal{W}_j^{t,\leftarrow} : \mathbf{X}^t \rightarrow \mathbf{X}_j^*$ that maps the camera space points $\mathbf{X}^t$ to canonical space points $\mathbf{X}_j^*$ with a rigid-body transformation $\mathbf{G}_j^t \in SE(3)$ and a deformation field $\mathbf{J}_j^{t,\leftarrow}$ modeled by linear blend skinning [13]: + +$$ +\mathbf {X} _ {j} ^ {*} = \mathcal {W} _ {j} ^ {t, \leftarrow} \left(\mathbf {X} ^ {t}\right) = \mathbf {J} _ {j} ^ {t, \leftarrow} \left(\mathbf {G} _ {j} ^ {t} \mathbf {X} ^ {t}\right). \tag {4} +$$ + +Note that our choice of deformation field differs from the $SE(3)$ -field used in HyperNeRF and $\mathrm{D}^2\mathrm{NeRF}$ , which has been shown to produce irregular deformation in the presence of complex scene motion [67]. Intuitively, rigid-body transformation $\mathbf{G}_j^t$ captures the global root-body pose of object $j$ relative to the camera at time $t$ , while deformation field $\mathbf{J}_j^{t,\leftarrow}$ aligns more fine-grained articulations relative to its local canonical space (Figure 2). Explicitly disentangling these two sources of object motion (as opposed to conflating them) enables easier optimization of the deformation field, because local articulations are significantly easier to learn than those modeled relative to the world space (Equation 1). Furthermore, this motion decomposition makes the deformation field invariant to rigid-body transformations of the object. A motion model similar to ours was proposed by ST-NeRF [15], but their model encodes an object's global root-body motion with a 3D axis-aligned bounding box that does not explicitly represent object orientation, a prerequisite for embodied view synthesis from 3rd-person-follow cameras. + +
Rendered features f at pixel xtCorresponding 3D features fij(Xt)
color帽子(xt)ct( Wj,←(Xt))
flow F( xt, t → t')Πt' (Wj', → (Wj, ←(Xt)) - xt
depth d( xt)[0, 0, 1] · Xt
+ +Table 2: Rendered 2D features $\hat{\mathbf{f}}$ and their corresponding 3D features $\mathbf{f}_{ij}$ . $\Pi^{t'}$ denotes the camera intrinsics at time $t'$ . + +As did BANMo, Total-Recon also models a forward warp $\mathbf{X}_j^t = \mathcal{W}_j^{t,\rightarrow}(\mathbf{X}^*) = (\mathbf{G}_j^t)^{-1}\mathbf{J}_j^{t,\rightarrow}(\mathbf{X}^*)$ that maps the canonical space to the camera space, which is used to establish the surface correspondences required for egocentric view synthesis and 3D video filters. + +# 3.3. Composite Rendering of Multiple Objects + +Given a set of $M$ object representations (the background is treated as an object as well), we use the composite rendering scheme from prior work [31, 52] to combine the outputs of all object representations and volume-render the entire scene. To volumetrically render the image at frame $t$ , we sample multiple points along each camera ray $\mathbf{v}^t$ . Denoting the $i^{th}$ sample as $\mathbf{X}_i^t$ , we write the density and color observed at sample $i$ due to object $j$ as: + +$$ +\sigma_ {i j} = \mathbf {M L P} _ {\sigma , j} \left(\mathbf {X} _ {i j} ^ {*}\right), \qquad \mathbf {c} _ {i j} = \mathbf {M L P} _ {\mathbf {c}, j} \left(\mathbf {X} _ {i j} ^ {*}, \mathbf {v} _ {j} ^ {*}, \boldsymbol {\omega} _ {e} ^ {t}\right), +$$ + +where $\mathbf{X}_{ij}^{*} = \mathcal{W}_{j}^{t,\leftarrow}(\mathbf{X}_{i}^{t})$ and $\mathbf{v}_j^* = \mathcal{W}_j^{t,\leftarrow}(\mathbf{v}^t)$ are sample $i$ and camera ray $\mathbf{v}^t$ backward-warped into object $j$ 's canonical space, respectively. The composite density $\sigma_{i}$ at sample $i$ along the ray is then computed as the sum of each object's density $\sigma_{ij}$ ; the composite color $\mathbf{c}_i$ is computed as the weighted sum of each object's color $\mathbf{c}_{ij}$ , where the weights are the normalized object densities $\sigma_{ij} / \sigma_i$ : + +$$ +\sigma_ {i} = \sum_ {j = 0} ^ {M - 1} \sigma_ {i j}, \quad \mathbf {c} _ {i} = \frac {1}{\sigma_ {i}} \sum_ {j = 0} ^ {M - 1} \sigma_ {i j} \mathbf {c} _ {i j}. \tag {5} +$$ + +We can then use the standard volume rendering equations to generate an RGB image of the scene, where $N$ is the number of sampled points along camera ray $\mathbf{v}^t$ , $\tau_{i}$ is the transmittance, $\alpha_{i}$ is the alpha value for sample point $i$ and $\delta_{i}$ is the distance between sample point $i$ and the $(i + 1)$ : + +$$ +\hat {\mathbf {c}} = \sum_ {i = 1} ^ {N} \tau_ {i} \alpha_ {i} \mathbf {c} _ {i}, \quad \tau_ {i} = \prod_ {k = 1} ^ {i - 1} (1 - \alpha_ {k}), \quad \alpha_ {i} = 1 - e ^ {- \sigma_ {i} \delta_ {i}}. +$$ + +Rendering Flow, Depth, and Silhouettes. Our composite rendering scheme can be used to render different quantities by replacing the object color $\mathbf{c}_{ij}$ in Equation 5 with the + +![](images/dee609955c7eb87a0593595a6c91eb580240b9426e3498568638cb241961a32e.jpg) +Figure 3: 6-DOF Trajectories for Embodied View Synthesis. To synthesize embodied views from egocentric or actor-following cameras, Total-Recon reconstructs the entire background, every individual actor in the scene, as well as global 6-DOF trajectories of its root-body and its articulated body parts (e.g. head). + +appropriately defined 3D feature $\mathbf{f}_{ij}$ (Table 2) and rendering the resulting composite feature $\mathbf{f}_i$ . To render occlusion-aware object silhouettes, we follow ObSURF [52] to produce a categorical distribution over the $M$ objects: + +$$ +\hat {\mathbf {o}} _ {j} = \sum_ {i = 1} ^ {N} \tau_ {i} \alpha_ {i j}, \quad \text {w h e r e} \quad \tau_ {i} = \prod_ {k = 1} ^ {i - 1} (1 - \alpha_ {k}), \tag {6} +$$ + +$$ +\alpha_ {i} = 1 - e ^ {- \sigma_ {i} \delta_ {i}}, \quad \alpha_ {i j} = 1 - e ^ {- \sigma_ {i j} \delta_ {i}}. \tag {7} +$$ + +Losses. Given a monocular RGBD video, we optimize all parameters in our composite scene representation, which for each of the $M$ objects includes the appearance and shape MLPs $(\mathbf{MLP}_{c,j}, \mathbf{MLP}_{\sigma,j})$ , rigid-body transformations $\mathbf{G}_j^t$ , and forward, backward deformation fields $\mathbf{J}_j^{\leftarrow}$ , $\mathbf{J}_j^{\rightarrow}$ . We optimize three reconstruction losses: a color loss $\mathcal{L}_{\mathrm{rgb}}$ , a flow loss $\mathcal{L}_{\mathrm{flow}}$ , and a depth loss $\mathcal{L}_{\mathrm{depth}}$ , where the ground truth color $\mathbf{c}$ and depth $\mathbf{d}$ are provided by the RGBD video, and the "ground truth" flow $\mathcal{F}$ is computed by an off-the-shelf network [64]. The model also optimizes a 3D-cycle consistency loss $\mathcal{L}_{\mathrm{cyc},j}$ [67] for each deformable object to encourage their forward and backward warps to be consistent, where $\mathbf{x}^t \in \mathbb{R}^2$ denotes pixel location at time $t$ : + +$$ +\mathcal {L} _ {\mathrm {r g b}} = \sum_ {\mathbf {x} ^ {t}} \left| \left| \mathbf {c} \left(\mathbf {x} ^ {t}\right) - \hat {\mathbf {c}} \left(\mathbf {x} ^ {t}\right) \right| \right| ^ {2}, \tag {8} +$$ + +$$ +\mathcal {L} _ {\text {f l o w}} = \sum_ {\mathbf {x} ^ {t}} \left| \left| \mathcal {F} \left(\mathbf {x} ^ {t}\right) - \hat {\mathcal {F}} \left(\mathbf {x} ^ {t}\right) \right| \right| ^ {2}, \tag {9} +$$ + +$$ +\mathcal {L} _ {\text {d e p t h}} = \sum_ {\mathbf {x} ^ {t}} \left| \left| \mathbf {d} \left(\mathbf {x} ^ {t}\right) - \hat {\mathbf {d}} \left(\mathbf {x} ^ {t}\right) \right| \right| ^ {2}, \tag {10} +$$ + +$$ +\mathcal {L} _ {\text {c y c}, j} = \sum_ {i} \tau_ {i} \alpha_ {i j} \left\| \mathcal {W} _ {j} ^ {t ^ {\prime}, \rightarrow} \left(\mathcal {W} _ {j} ^ {t}, \leftarrow (\mathbf {X} _ {i} ^ {t})\right) - \mathbf {X} _ {i} ^ {t} \right\| ^ {2}. \tag {11} +$$ + +Initialization. We initialize the rigid-body transformations of each deformable object $\mathbf{G}_j^t$ using a pre-trained category-specific PoseNet [67]; we initialize the rigid-body transformation of the background $\mathbf{G}_0^t$ with the camera poses provided by the iPad Pro. + +Embodied View Synthesis and 3D Filters. To enable embodied view synthesis and 3D video filters (Figure 3), we design a simple interface that allows a user to select a point on a target object's surface in its reconstructed canonical mesh, and use its forward warping function $\mathcal{W}_j^{t,\rightarrow}:\mathbf{X}^* \to \mathbf{X}^t$ followed by the rigid-body transformation $\mathbf{G}_0^t$ to place the egocentric camera (or virtual 3D asset) in the world space. The surface normal to the object's mesh at the user-defined point provides a reference frame to align the egocentric camera's viewing direction and place the 3D asset. To implement a 3rd-person-follow camera, we add a user-defined offset to the object's local reference frame, which is defined by its root-body pose. + +# 4. Experiments + +Implementation Details. In practice, we train our composite scene representation by first pre-training each object field separately. For deformable objects, we pre-train using a depth loss (Equation 10) combined with the losses optimized by BANMo [67]. This includes a silhouette loss $\mathcal{L}_{\mathrm{mask}} = \sum_{\mathbf{x}^t}||\mathbf{o}_j(\mathbf{x}^t) - \hat{\mathbf{o}}_j(\mathbf{x}^t)||^2$ , where the "ground truth" object silhouette $\mathbf{o}_j$ is computed by an off-the-shelf instance segmentation engine [17]. For pre-training the background, we optimize color, flow, and depth losses (Equations 8, 9, 10) on pixels outside the ground truth object silhouettes. Importantly, we don't supervise the object fields on frames that are not provided an object silhouette since it cannot be determined whether the absence of detection is a true or false negative. + +After pre-training, we composite-render the pre-trained object fields and jointly finetune them using only the color, depth, flow, and object-specific 3D-cycle consistency losses. Since the silhouette loss is no longer used, the scene representation is supervised on all frames of the training se + +quence during joint-finetuning. We provide a complete description of the implementation details in the supplement. + +Dataset. We evaluate Total-Recon on novel-view synthesis for deformable scenes. To enable quantitative evaluation, we built a stereo rig comprised of two iPad-Pros rigidly attached to a camera mount, a setup similar to that of Nerfies [35]. Using the stereo rig, we captured 11 RGBD sequences containing 3 different cats, 1 dog, and 2 human subjects in 4 different indoor environments. The RGBD videos were captured using the Record3D iOS App [49], which also automatically registers the frames captured by each camera. These video sequences, which were subsampled at 10 fps, range from 392 to 901 frames, amount to, on average minute-long videos that are significantly longer and contain more dynamic motion than the datasets introduced by [35, 36, 60, 11]. The left and right cameras were registered by solving a Perspective-n-Point (PnP) problem using manually annotated correspondences, and their videos were synchronized based on audio. We provide a complete description of our dataset in the supplement. + +Reconstruction and Applications. By hierarchically decomposing scene motion into the motion of each object, which itself is decomposed into root-body motion and local articulations, Total-Recon automatically computes novel 6-DoF trajectories such as those traversed by egocentric cameras and 3rd-person follow cameras (Figure 3). In turn, these trajectories enable automated embodied view synthesis and 3D occlusion-aware video filters (Figure 4). These tasks are also enabled by Total-Recon's ability to recover an accurate deformable 3D scene representation, which is currently out of reach for the best of related methods (Figure 5). As shown in the bird's eye view, each reconstructed object is properly situated with respect to the background and other objects, a direct consequence of our use of depth supervision. Furthermore, even though the iPad Pro can only measure depth up to $4\mathrm{m}$ , Total-Recon can render depth beyond this sensor limit by pooling the measurements from other frames into a single metric scene reconstruction. We provide results on additional sequences in the supplement. + +Baselines and Evaluation. In Figure 5 and Table 3, we compare Total-Recon to $\mathrm{D}^2\mathrm{NeRF}$ [60] and HyperNeRF [36], and their depth-supervised equivalents on the proxy task of stereo-view synthesis, a prerequisite for embodied view synthesis: we train each method on the RGBD frames captured from the left camera of our dataset and evaluate the images rendered from the viewpoint of the right camera. The depth-supervised versions of the baselines contain the same depth loss used in Total-Recon. We report LPIPS [70] and the average (depth) accuracy at $0.1\mathrm{m}$ [42] in all subsequent experiments, and we include a more complete set + +![](images/9e8a93237ca95f8a6992de4d2c5afcd32218436318eb7d4d9ba12bc36911a487.jpg) +Figure 4: Embodied View Synthesis and 3D Filters. For select sequences of our RGBD dataset, we visualize the scene geometry and appearance reconstructed by our method (3D reconstruction) and the resulting downstream applications. The yellow and blue camera meshes in the mesh renderings represent the egocentric and 3rd-person-follow cameras, respectively. To showcase the 3D video filter we attach a sky-blue unicorn horn to the forehead of the target object, which is then automatically propagated across all frames. Full-length videos can be found at https://andrewsonga.github.io/totalrecon/applications.html. + +of metrics (PSNR, SSIM, RMS depth error) in the supplement. Because $\mathrm{D}^2\mathrm{NeRF}$ and HyperNeRF were not designed to recover a metric scene representation, we replaced their COLMAP [46] camera poses with those provided by the iPad Pro (which are metric measurements) for the sake of fair comparison. + +Comparisons. Total-Recon qualitatively and quantitatively outperforms all of the baselines. As shown in Figure 5, Total-Recon successfully reconstructs the entire scene, whereas the baselines are only able to reconstruct the rigid background at best. As shown in Table 3, Total-Recon significantly outperforms all baselines in terms of LPIPS and + +the average accuracy at $0.1\mathrm{m}$ (Acc@0.1m). We attribute this huge gap to the baselines' inability to reconstruct highly dynamic objects. We provide more details regarding the baselines and additional visualizations in the supplement. + +# 4.1. Ablation Studies + +Table 4 (and Figures 6 and 7) analyzes the importance of Total-Recon's design choices (see Section 3) by ablating its key components: the depth loss $\mathcal{L}_{\mathrm{depth}}$ (row 2), the deformation field $\mathbf{J}_j^t$ (row 3), PoseNet-initialization of the root-body pose (row 4), and the root-body pose $\mathbf{G}_j^t$ itself (row 5), where $j$ denotes a deformable actor. For all ablations, we use the same set of training losses used in Total- + +
DOG 1 (626 images)DOG 1 (v2) (531 images)CAT 1 (641 images)CAT 1 (v2) (632 images)CAT 2 (632 images) (834 images)CAT 2 (v2) (901 images)CAT 3 (767 images)HUMAN 1 (550 images)HUMAN 2 (483 images)HUMAN - DOG (392 images)HUMAN - CAT (431 images)MEAN
LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑
HyperNeRF.634.107.432.176.521.316.438.314.641.277.397.252.592.213.632.053.585.067.487.072.462.162.531.198
D2NeRF.540.219.546.220.687.346.588.403.556.333.595.339.759.231.588.066.630.128.576.078.628.126.611.247
HyperNeRF (+depth).373.352.425.357.532.552.371.596.330.605.376.612.514.451.501.211.445.249.450.283.456.214.428.439
D2NeRF (+depth).507.338.532.270.685.510.580.362.561.438.553.376.730.243.585.086.609.131.608.154.645.176.599.302
Total-Recon.271.841.313.790.382.889.333.894.237.967.281.925.261.949.213.909.264.849.256.827.233.914.278.895
+ +![](images/86e2c39980dc895e8c6568960f3e9a74f94daf6a9704a87d1ed1a17bd2fcc4d3.jpg) +Figure 5: Baseline Comparisons. We compare Total-Recon to depth-supervised variants of HyperNeRF [36] and $\mathrm{D}^2\mathrm{NeRF}$ [60] on the task of stereo-view synthesis (the left camera is used for training and the right is used for testing). While the baselines are able to reconstruct only the background, Total-Recon can reconstruct both the background and the moving deformable object(s), demonstrating holistic scene reconstruction. + +Recon and initialize camera pose $\mathbf{G}_0^t$ with those reported by ARKit. For ablations that model root-body motions, we initialize each deformable actor's root-body pose $\mathbf{G}_j^t$ with predictions made by PoseNet [67] and optimize them during reconstruction; for row 4, we replace the PoseNet predictions with identity rotations. We report the novel-view metrics averaged over 6 select sequences of our dataset: DOG 1 (v1), CAT 1 (v1), CAT 2 (v1), HUMAN 1, HUMAN 1 & DOG 1, and HUMAN 2 & CAT 1. + +Depth Supervision. Table 4 shows that removing depth supervision (row 2) significantly reduces the average accuracy at $0.1\mathrm{m}$ (Acc). Figure 6 indicates that this reflects the incorrect arrangement of objects stemming from their scale inconsistency - while removing depth supervision does not significantly deteriorate the training-view RGB renderings, it induces critical failure modes as shown in the novel-view 3D reconstructions: (a) floating foreground objects, as evi + +Table 3: Baseline Comparisons. We train Total-Recon, HyperNeRF [36], D²NeRF [60], and their depth-supervised variants on the left video captured with our stereo rig, and evaluate the novel view synthesis results on the held-out right video. Total-Recon significantly outperforms all of the baselines for all 11 sequences. These sequences are sampled at 10 fps, amounting to minute-long videos, on average. + +
MethodsDepth LossDeform. Obj.Root Init.Root MotionLPIPS↓Acc@0.1m↑
(1) Full model✓✓✓✓.268.898
(2) w/o loss LdepthX✓✓✓.372.154
(3) w/o deform. Jj✓X✓✓.296.867
(4) w/o root-body init.✓✓X✓.293.870
(5) w/o root-body Gj✓✓XXN/AN/A
+ +Table 4: Ablation Study. Removing depth supervision (2) significantly hurts performance, while removing the deformation field (3) and PoseNet-initialization of root-body poses (4) hurts moderately. Most importantly, removing root-body poses entirely (5) prevents convergence (N/A) as the deformation field alone has to explain global object motion (see Figure 2). These experiments justify our hierarchical modeling of motion, as even root-bodies without a deformation field (3) or poorly initialized root-bodies (4) are better than no root-bodies (5). We visualize these ablations in Figure 7 and explore other ablations in the supplement. + +denced by their shadows, and (b) the human incorrectly occluding the dog. In other words, without depth supervision, Total-Recon overfits the training view and learns a degenerate scene representation where the reconstructed objects fail to converge to the same scale. We show results on additional sequences in the supplement. + +Motion Modeling. Table 4 shows that removing the deformation field (row 3) also hurts performance. This is because, without the deformation field, our method has to explain an object's non-rigid motion solely with its rigid, root-body poses. As a result, this ablation can only recover coarse object reconstructions that fail to model moving body parts such as limbs. Removing PoseNet-initialization of root-body poses (row 4) is just as detrimental, resulting in noisy appearance and geometry artifacts; see Figure 7 and additional visuals in the supplement. Most notably, Table 4 shows that removing object root-bodies entirely (row 5) causes the optimization to fail to converge (N/A), even though the deformation field can (in theory) represent all continuous motion. It appears difficult for deformation fields alone to explain global root-body motion because such motions can deviate significantly from a canonical model, complicating optimization. + +![](images/6cf3c6572a73a8bf1922c93283acb8d76d41ea7ba320ab4494f4cc8c3d316329.jpg) +Figure 6: Ablation Study on Depth Supervision. While removing depth supervision does not significantly deteriorate the training-view RGB renderings, it significantly hurts the novel-view 3D reconstructions, as characterized by the following: (a) floating foreground objects (as evidenced by their shadows) and (b) the human incorrectly occluding the dog. These failure modes indicate that without depth supervision, Total-Recon overfits the training view, and the reconstructed objects fail to converge to the same scale. + +![](images/c03c6cd8035ea19355438b7fbc3e13dbf5abbc226a7546780f83999cf555a483.jpg) +Figure 7: Ablation Study. We visualize the ablations from Table 4. Removing the depth loss causes the cat to sink into the ground due to inconsistent object scales. Removing the deformation field produces coarse object reconstructions that fail to capture articulated body parts such as moving limbs. Removing PoseNet-initialization of root-body poses results in noisier appearance and geometry, as shown by the human actor's left hand. We omit the ablation without root-body poses as it does not converge. We present additional visualizations in the supplement. + +These diagnostics justify Total-Recon's hierarchical motion representation, which explicitly models objects' root-body poses; root-bodies without articulated deformations (row 3) or poorly initialized root-bodies (row 4) are better than no root-bodies at all (row 5). Our ablations also suggest that the poor performance of the baseline methods (on our challenging dataset) may be attributed to the lack of object-centric motion modeling. We provide a more detailed analysis with additional experiments and RGBD sequences in the supplement. + +# 5. Discussion and Limitations + +We have presented a new system for automated embodied view synthesis from monocular RGBD videos, focusing on videos of people interacting with their pets. Our main technical contribution is Total-Recon, a 3D representation for deformable scenes that hierarchically decomposes scene motion into the motion of each object, which is further decomposed into its root-body motion and local articulations; this key design choice enables appropriate initialization of root-body poses and hence easier optimization over long videos containing large motions. By explicitly reconstructing the geometry, appearance, root-body- and articulated motion of each object, Total-Recon enables seeing through the eyes of people and pets and generating game-like traversals of deformable scenes from behind a target actor. + +Limitations. In Total-Recon, scene decomposition is primarily supervised by object silhouettes computed by an off-the-shelf segmentation model [17], which may be inaccurate, especially in partial occlusion scenarios. This may damage the resulting reconstructions and embodied-view renderings. We believe that incorporating the latest advances in video instance segmentation [12] will enable Total-Recon to be applied to more challenging scenarios. Second, Total-Recon initializes the root-body pose of each deformable object using a PoseNet [67] trained for humans and quadruped animals, which does not generalize to other object categories (e.g., birds, fish). We reserve the reconstruction of generic scenes for future work. Finally, our model needs to be optimized on a per-sequence basis for roughly 15 hours with 4 NVIDIA RTX A5000 GPUs and is therefore not suitable for real-time applications. Incorporating recent advances in fast neural field training methods is an interesting avenue for future research. + +Acknowledgments. We thank Nathaniel Chodosh, Jeff Tan, George Cazenavette, and Jason Zhang for proofreading and Songwei Ge for reviewing the codebase. We also thank Sheng-Yu Wang, Daohan (Fred) Lu, Tamaki Kojima, Krishna Wadhwani, Takuya Narihira, and Tatsuo Fujiwara for providing valuable feedback. This work is supported in part by the Sony Corporation, Cisco Research, and the CMU Argo AI Center for Autonomous Vehicle Research. + +# References + +[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[2] Aljaz Bozic, Pablo Palafox, Michael Zollhofer, Justus Thies, Angela Dai, and Matthias Nießner. Neural deformation graphs for globally-consistent non-rigid reconstruction. In + +IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[3] Aljaz Bozic, Michael Zollhofer, Christian Theobalt, and Matthias Nießner. Deepdeform: Learning non-rigid rgb-d reconstruction with semi-supervised data. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[4] Hongrui Cai, Wanquan Feng, Xuetao Feng, Yan Wang, and Juyong Zhang. Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3 +[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[7] Alena Denisova and Paul Cairns. First person vs. third person perspective in digital games: Do player preferences affect immersion? In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, page 145-148, 2015. 1 +[8] Mingsong Dou, Jonathan Taylor, Henry Fuchs, Andrew Fitzgibbon, and Shahram Izadi. 3d scanning deformable objects with a single rgbd sensor. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3 +[9] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbaek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[10] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3 +[11] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3, 6 +[12] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 9 +[13] Alec Jacobson, Zhigang Deng, Ladislav Kavan, and J. P. Lewis. Skinning: Real-time shape deformation (full text not available). In ACM SIGGRAPH 2014 Courses, SIGGRAPH '14. Association for Computing Machinery, 2014. 4 +[14] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 3 +[15] Zhang Jiakai, Liu Xinhang, Ye Xinyi, Zhao Fuqiang, Zhang Yanshun, Wu Minye, Zhang Yingliang, Xu Lan, and Yu Jingyi. Editable free-viewpoint video using a layered neural representation. In ACM SIGGRAPH, 2021. 3, 4 +[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field + +from a single video. In European Conference on Computer Vision (ECCV), 2022. 3 +[17] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6, 9 +[18] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J. Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3 +[19] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In European Conference on Computer Vision (ECCV), 2022. 3 +[20] Tianye Li, Mira Slavcheva, Michael Zollhöfer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[21] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3 +[22] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[23] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 3 +[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM SIGGRAPH Asia, 2021. 3 +[25] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[26] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 4 +[27] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In IEEE International Conference on Computer Vision (ICCV), 2021. 3 +[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 4 + +[29] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3 +[30] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE International Symposium on Mixed and Augmented Reality, pages 127-136. IEEE, 2011. 3 +[31] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5 +[32] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In IEEE International Conference on Computer Vision (ICCV), 2021. 3 +[33] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3 +[34] Kemal Egemen Ozden, Kurt Cornelis, Luc Van Eycken, and Luc Van Gool. Reconstructing 3d trajectories of independently moving objects using generic constraints. Comput. Vis. Image Underst., 96(3):453-471, 2004. 3 +[35] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3, 6 +[36] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 40(6),2021.2,3,4,6,8 +[37] Georgios Pavlakos*, Ethan Weber*, , Matthew Tancik, and Angjoo Kanazawa. The one where they reconstructed 3d humans and environments in tv shows. In European Conference on Computer Vision (ECCV), 2022. 3 +[38] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In IEEE International Conference on Computer Vision (ICCV), 2021. 3 +[39] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[40] F. Peruzzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3 +[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields + +for Dynamic Scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3 +[42] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 6 +[43] Zhongzheng Ren, Xiaoming Zhao, and Alex Schwing. Class-agnostic reconstruction of dynamic objects from videos. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[44] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[45] Richard Rouse. What's your perspective? ACM SIGGRAPH Computer Graphics, 33(3):9-12, August 1999. 1 +[46] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 7 +[47] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[48] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In SIGGRAPH Conference Proceedings, 2022. 3 +[49] Marek Simonik. Record3d. https://record3d.app. 6 +[50] Miroslava Slavcheva, Maximilian Baust, Daniel Cremers, and Slobodan Ilic. Killingfusion: Non-rigid 3d reconstruction without correspondences. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3 +[51] Miroslava Slavcheva, Maximilian Baust, and Slobodan Ilic. Sobolevfusion: 3d reconstruction of scenes undergoing free non-rigid motion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3 +[52] Karl Stelzner, Kristian Kersting, and Adam R Kosiorek. Decomposing 3d scenes into objects via unsupervised volume segmentation. arXiv preprint arXiv:2104.01148, 2021. 5 +[53] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[54] Unity Technologies. 3rd person follow. https://docs.unity3d.com/Packages/ com.unity.cinemachine@2.8/manual/ Cinemachine3rdPersonFollow.html.1 +[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3 + +[56] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4 +[57] Barbara Tversky. Spatial Cognition Embodied and Situated. Cambridge, 2008. 1 +[58] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021. 2, 3 +[59] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 3 +[60] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. $\mathbf{D}^2$ nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 4, 6, 8 +[61] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3 +[62] Jing Xiao, Jinxiang Chai, and Takeo Kanade. A closed-form solution to non-rigid shape and motion recovery. International Journal of Computer Vision (IJCV), 67(2):233-246, April 2006. 3 +[63] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[64] Gengshan Yang and Deva Ramanan. Volumetric correspondence networks for optical flow. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 5 +[65] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T Freeman, and Ce Liu. LASR: Learning articulated shape reconstruction from a monocular video. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[66] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2 +[67] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6, 8, 9 +[68] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4 +[69] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[70] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of + +deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6 +[71] Xiaoshuai Zhang, Abhijit Kundu, Thomas Funkhouser, Leonidas Guibas, Hao Su, and Kyle Genova. Nerflets: Local radiance fields for efficient structure-aware 3d scene representation from 2d supervision. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 \ No newline at end of file diff --git a/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/images.zip b/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8fabc24ced0e6dfefc82896c192dfc6386c58daa --- /dev/null +++ b/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f37aa1ab53114a970b18603481c559cf75777391a94ca5592b15bd72fe8f15db +size 738410 diff --git a/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/layout.json b/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d8ca17604da24a02fc9e4ad7d006c69e44e9ac9f --- /dev/null +++ b/2023/Total-Recon_ Deformable Scene Reconstruction for Embodied View Synthesis/layout.json @@ -0,0 +1,8709 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "text", + "content": "Total-Recon: Deformable Scene Reconstruction for Embodied View Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 143, + 496, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 143, + 496, + 174 + ], + "spans": [ + { + "bbox": [ + 97, + 143, + 496, + 174 + ], + "type": "text", + "content": "Chonghyuk Song Gengshan Yang Kangle Deng Jun-Yan Zhu Deva Ramanan Carnegie Mellon University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 175, + 547, + 324 + ], + "blocks": [ + { + "bbox": [ + 48, + 175, + 547, + 324 + ], + "lines": [ + { + "bbox": [ + 48, + 175, + 547, + 324 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 547, + 324 + ], + "type": "image", + "image_path": "baeec2e9755f02e3f6586435c235cbe25ffdfdc9edf954fd3dac8ff708f9f7d2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 546, + 389 + ], + "type": "text", + "content": "Figure 1: Embodied View Synthesis. Given a long video of deformable objects captured by a handheld RGBD sensor, Total-Recon renders the scene from novel camera trajectories derived from the in-scene motion of actors: (1) egocentric cameras that simulate the point-of-view of a target actor (such as the pet) and (2) 3rd-person (or pet) cameras that follow the actor from behind. Our method also enables (3) 3D video filters that attach virtual 3D assets to the actor. Total-Recon achieves this by reconstructing the geometry, appearance, root-body- and articulated motion of each deformable object in the scene and the background." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 403, + 192, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 403, + 192, + 415 + ], + "spans": [ + { + "bbox": [ + 143, + 403, + 192, + 415 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 426, + 290, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 426, + 290, + 701 + ], + "spans": [ + { + "bbox": [ + 45, + 426, + 290, + 701 + ], + "type": "text", + "content": "We explore the task of embodied view synthesis from monocular videos of deformable scenes. Given a minute-long RGBD video of people interacting with their pets, we render the scene from novel camera trajectories derived from the in-scene motion of actors: (1) egocentric cameras that simulate the point of view of a target actor and (2) 3rd-person cameras that follow the actor. Building such a system requires reconstructing the root-body and articulated motion of every actor, as well as a scene representation that supports free-viewpoint synthesis. Longer videos are more likely to capture the scene from diverse viewpoints (which helps reconstruction) but are also more likely to contain larger motions (which complicates reconstruction). To address these challenges, we present Total-Recon, the first method to photorealistically reconstruct deformable scenes from long monocular RGBD videos. Crucially, to scale to long videos, our method hierarchically decomposes the scene into the background and objects, whose motion is decomposed into carefully initialized root-body motion and local articulations. To quantify such \"in-the-wild\" reconstruction and view synthesis, we collect ground-truth data from a specialized stereo RGBD capture rig for 11 challenging videos, significantly outperforming prior methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 403, + 387, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 403, + 387, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 403, + 387, + 415 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 422, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 651 + ], + "type": "text", + "content": "We explore embodied view synthesis, a new class of novel-view synthesis tasks that renders deformable scenes from novel 6-DOF trajectories reconstructed from the in-scene motion of actors: egocentric cameras [45, 7] that simulate the point-of-view of moving actors and 3rd-person-follow cameras [54, 7] that track a moving actor from behind (Figure 1). We focus on everyday scenes of people interacting with their pets, producing renderings from the point-of-view of the person and pet (Figure 1). While such camera trajectories could be manually constructed (e.g., by artists via keyframing), building an automated system is an interesting problem of its own: spatial cognition theory [57] suggests that the ability to visualize behavior from another actor's perspective is necessary for action learning and imitation; in the context of gaming and virtual reality [7, 45], egocentric cameras offer high levels of user immersion, while 3rd-person-follow cameras provide a large field of view that is useful for exploring a user's environment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "content": "Challenges. Building a system for embodied view synthesis is challenging for many reasons. First, to reconstruct everyday-but-interesting content, it needs to process long, monocular captures of multiple interacting actors. How-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17671" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 98, + 71, + 495, + 268 + ], + "blocks": [ + { + "bbox": [ + 98, + 71, + 495, + 268 + ], + "lines": [ + { + "bbox": [ + 98, + 71, + 495, + 268 + ], + "spans": [ + { + "bbox": [ + 98, + 71, + 495, + 268 + ], + "type": "image", + "image_path": "5e8218f59c1037fb89cfa0dc8f68b116edcc85c26a28a3c8133ea47da4a8eb54.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 548, + 343 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 548, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 548, + 343 + ], + "type": "text", + "content": "Figure 2: Method Overview. Total-Recon represents the entire scene as a composition of " + }, + { + "bbox": [ + 46, + 274, + 548, + 343 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 274, + 548, + 343 + ], + "type": "text", + "content": " object-centric neural fields, one for the rigid background and each of the " + }, + { + "bbox": [ + 46, + 274, + 548, + 343 + ], + "type": "inline_equation", + "content": "M - 1" + }, + { + "bbox": [ + 46, + 274, + 548, + 343 + ], + "type": "text", + "content": " deformable objects. To render a scene, (1) each object field " + }, + { + "bbox": [ + 46, + 274, + 548, + 343 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 274, + 548, + 343 + ], + "type": "text", + "content": " is transformed into the camera space with a rigid transformation " + }, + { + "bbox": [ + 46, + 274, + 548, + 343 + ], + "type": "inline_equation", + "content": "\\left(\\mathbf{G}_j^t\\right)^{-1}" + }, + { + "bbox": [ + 46, + 274, + 548, + 343 + ], + "type": "text", + "content": " that encodes root-body motion and, for each deformable object, an additional deformation field " + }, + { + "bbox": [ + 46, + 274, + 548, + 343 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_j^{t,\\rightarrow}" + }, + { + "bbox": [ + 46, + 274, + 548, + 343 + ], + "type": "text", + "content": " that encodes articulated motion. Next, all (2) posed object fields are combined into a (3) composite field, which is then volume-rendered into (4) color, depth, optical flow, and object silhouettes. Each rendered output defines a reconstruction loss that derives supervision from a monocular RGBD video captured by a moving iPad Pro." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 353, + 289, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 289, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 289, + 521 + ], + "type": "text", + "content": "ever, such videos are likely to contain large scene motions, which we demonstrate are difficult to reconstruct with current approaches. Second, it needs to produce a deformable 3D scene representation that supports free-viewpoint synthesis, which also would benefit from long videos likely to capture the scene from diverse viewpoints. Recent approaches have extended Neural Radiance Fields (NeRFs) [28] to deformable scenes, but such work is often limited to rigid-only object motion [18, 33], short videos with limited scene motion [41, 35, 21, 55, 36, 60, 10, 61, 58], or reconstructing single objects as opposed the entire scene [65, 66, 67, 4]. Third, it needs to compute global 6-DOF trajectories of root-bodies and articulated body parts (e.g., head) of multiple actors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "Key Ideas. To address these challenges, we introduce Total-Recon, the first monocular NeRF that enables embodied view synthesis for deformable scenes with large motions. Given a monocular RGBD video, Total-Recon reconstructs the scene as a composition of object-centric representations, which encode the 3D appearance, geometry, and motion of each deformable object and the background. Crucially, Total-Recon hierarchically decomposes scene motion into the motion of individual objects, which itself is decomposed into global root-body movement and the local deformation of articulated body parts. We demonstrate that such decomposition of object motion, along with appropriate initialization of root-body pose, allows reconstruction to scale to longer videos, enabling free-viewpoint synthesis. By reconstructing such motions in a globally-consistent" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 353, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 353, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 547, + 390 + ], + "type": "text", + "content": "coordinate frame, Total-Recon can generate renderings from egocentric and 3rd-person-follow cameras, as well as static but extreme viewpoints like bird's-eye-views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 408, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 547, + 540 + ], + "type": "text", + "content": "Evaluation. Due to the difficulty of collecting ground-truth data for embodied view synthesis on in-the-wild videos, we evaluate our method on the proxy task of stereo-view synthesis [35], which compares rendered views to those captured from a stereo pair. To this end, we build a stereo RGBD sensor capture rig for ground-truthing and collect a dataset of 11 long video sequences in various indoor environments, including people interacting with their pets. Total-Recon outperforms the state-of-the-art monocular deformable NeRF methods [36, 60], even when modified to use depth sensor measurements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "Contributions. In summary, our contributions are: (1) Total-Recon, a hierarchical 3D representation that models deformable scenes as a composition of object-centric representations, each of which decomposes object motion into its global root-body motion and its local articulations; (2) a system based on Total-Recon for automated embodied view synthesis from casual, minute-long RGBD videos of highly dynamic scenes; (3) a dataset of stereo RGBD videos containing various deformable objects, such as humans and pets, in a host of different background environments. Our code, models, and data can be found at https://andrewsonga.github.io/totalrecon." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17672" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 188 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 188 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 188 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 188 + ], + "type": "table", + "html": "
MethodEntire ScenesDeform. ObjectsBeyond HumansGlobal 6-DOF Traj.Long VideosExtreme Views
BANMo [67]X✓✓X✓✓
PNF [18]✓X✓✓XX
NeuMan [16]✓✓X✓XX
SLAHMR [69]X✓X✓XX
HyperNeRF [36]✓✓✓XXX
D2NeRF [60]✓✓✓XXX
DynIBaR [22]✓✓✓X✓X
SUDS [56]✓✓✓X✓X
Ours✓✓✓✓✓✓
", + "image_path": "049ff2708edfcc24bdac102133780f47771de364cc1d3e7e8d45eb22a6c73dc8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 198, + 287, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 275 + ], + "type": "text", + "content": "Table 1: Comparison to Related Work. Unlike prior work, Total-Recon targets embodied view synthesis of scenes containing humans and pets, requiring the ability to (1) reconstruct entire scenes, (2) model deformable objects, (3) extend beyond humans, (4) recover global 6-DOF trajectories of objects' root-bodies and their articulated parts, (5) process minute-long videos of dynamic scenes, and (6) render extreme views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 294, + 134, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 294, + 134, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 294, + 134, + 306 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 315, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 495 + ], + "type": "text", + "content": "Neural Radiance Fields. Prior works on Neural Radiance Fields (NeRF) optimize a continuous scene function for novel view synthesis given a set of multi-view images, usually under the assumption of a rigid scene and densely sampled views [28, 26, 27, 23, 14, 59]. DS-NeRF [6] and Dense Depth Priors [44] extend NeRFs to the sparse-view setting by introducing depth as additional supervision. Total-Recon also operates in the sparse-view regime and uses depth supervision to reduce the ambiguities inherent to monocular, multibody, non-rigid reconstruction [62, 34]. Another line of work [18, 33] represents rigidly moving scenes as a composition of multiple object-level NeRFs. Total-Recon also leverages such an object-centric scene representation, but models scenes containing non-rigidly moving objects, such as humans and pets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 715 + ], + "type": "text", + "content": "Deformable NeRFs. Recent approaches extend NeRF to monocular deformable scene reconstruction either by learning an additional function that deforms observed points in the camera space to a time-independent canonical space [41, 35, 55, 36, 60] or explicitly modeling density changes over time [10, 61, 58, 21]. Such methods are typically limited to short videos containing little scene and camera motion. They also perform novel-view synthesis only over small baselines. Total-Recon belongs to the former category of prior monocular deformable NeRFs, but unlike them, our method hierarchically decomposes scene motion into the motion of each object, which is further decomposed into global root-body motion and local articulations. The proposed motion decomposition is what enables embodied view synthesis: it allows Total-Recon to scale to minute-long videos and reconstruct a deformable 3D scene representation that supports free-viewpoint syn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "content": "thesis; it also makes it easy to extract an object's root-body motion, the key motion primitive required for 3rd-person-follow view synthesis. Several works have taken different approaches to making non-rigid reconstruction more tractable. One group of work leverages human-specific priors [38, 53, 32, 39, 24, 16, 19, 37] such as human body models (e.g., SMPL), 3D skeletons, or 2D poses to achieve high reconstruction quality. We achieve similar levels of fidelity without relying on such shape priors, allowing Total-Recon to generalize to pets and, by extension, reconstruct human-pet interaction videos. Another body of work [15, 48, 20] achieves high-fidelity scene reconstructions by relying on synchronized multi-view video captured from a specialized camera rig ranging from 8 to 18 static cameras. In contrast, Total-Recon only requires a single video captured from a moving RGBD camera equipped with inertial sensors, which has now become widely accessible in consumer products with the advent of Apple's iPhone and iPad Pro." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 303, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 555 + ], + "type": "text", + "content": "Reconstruction with RGBD Sensors. Depth sensors represent the third class of attempts to make non-rigid reconstruction more tractable, reducing the need for a predefined shape template. Kinect-fusion [30] creates a real-time system for indoor scene localization and mapping. Dynamic Fusion [29] builds a template-free dense SLAM system for dynamic objects. Later works improve RGBD reconstruction to be able to deal with topology changes [50, 51] and use correspondence matching for registration over large motions [2, 3]. Recent works have incorporated neural implicit representations to reconstruct the surface geometry and 3D motion fields for deformable objects [43, 4] or largescale rigid scenes [1, 42] in isolation. Other works have reconstructed humans alongside small-scale objects and furniture [8, 2], but not the entire background. We aim to go even further by reconstructing the entire scene, which includes the background and multiple deformable targets such as humans and pets; not only do we reconstruct the geometry, but we also recover a radiance field that allows for photorealistic scene rendering from embodied viewpoints and other novel 6-DOF trajectories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Concurrent Work. Concurrent work exhibits a subset of the design choices necessary for embodied view synthesis. SLAHMR [69] reconstructs the geometry and in-scene motion of human actors but not the scene appearance. Nerflets [71] models the appearance, geometry, and motion of each scene element but is limited to rigidly moving objects. RoDynRF [25], NeRF-DS [63], HexPlane [5], and K-planes [9] reconstruct other types of dynamic scene elements, such as deformable or specular objects, but these methods have been demonstrated on only short videos and/or videos containing limited object root-body motion [40, 47, 63, 11]. DynIBaR [22] scales dynamic view synthesis to longer" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17673" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": "videos with complex camera and scene motion, and SUDS [56] scales reconstruction to urban-scale dynamic scenes captured from 1.2 million frames. However, neither demonstrates extreme-view synthesis, a prerequisite for rendering embodied views. We summarize and compare prior work to Total-Recon in Table 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 154, + 103, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 154, + 103, + 166 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 103, + 166 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 173, + 183, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 173, + 183, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 173, + 183, + 186 + ], + "type": "text", + "content": "3.1. Limitations of Prior Art" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "text", + "content": "The state-of-the-art monocular deformable NeRFs [36, 60] decompose a deformable scene into a rigid, canonical template model and a deformation field " + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{J}^{t,\\leftarrow}" + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "text", + "content": " that maps the world space " + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_0^t\\mathbf{X}^t" + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "text", + "content": " to the canonical space " + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^*" + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_0^t" + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "text", + "content": " is the known camera pose at time " + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t" + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "text", + "content": " is a camera space point at time " + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 192, + 287, + 263 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 270, + 287, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 270, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 93, + 270, + 287, + 285 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {*} = \\mathcal {W} ^ {t, \\leftarrow} \\left(\\mathbf {X} ^ {t}\\right) = \\mathbf {J} ^ {t, \\leftarrow} \\left(\\mathbf {G} _ {0} ^ {t} \\mathbf {X} ^ {t}\\right). \\tag {1}", + "image_path": "359efde39c461f1beced3d9d28a821a845629068c0b3a63f1bb6a96797cf120e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 291, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 287, + 411 + ], + "type": "text", + "content": "In theory, this formulation is sufficient to represent all continuous motion; it performs well on short videos containing near-rigid scenes, as the deformation field only has to learn minute deviations from the template model. However, this motion model is difficult to scale to minute-long videos, which are more likely to contain deformable objects undergoing large translations (e.g., a person walking into another room) and pose changes (e.g., a person sitting down). Here, the deformation field must learn large deviations from the canonical model, significantly complicating optimization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "type": "text", + "content": "Another critical limitation of HyperNeRF and " + }, + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 46, + 411, + 287, + 460 + ], + "type": "text", + "content": " is that they cannot track separate deformable objects and therefore cannot perform 3rd-person-follow view synthesis for scenes with multiple actors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 465, + 203, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 203, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 203, + 478 + ], + "type": "text", + "content": "3.2. Component Radiance Fields" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 483, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 288, + 616 + ], + "type": "text", + "content": "To address the limitations of existing monocular deformable NeRFs, we propose Total-Recon, a novel 3D representation that models a deformable scene as a composition of " + }, + { + "bbox": [ + 46, + 483, + 288, + 616 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 483, + 288, + 616 + ], + "type": "text", + "content": " object-centric neural fields, one for the rigid background and each of the " + }, + { + "bbox": [ + 46, + 483, + 288, + 616 + ], + "type": "inline_equation", + "content": "M - 1" + }, + { + "bbox": [ + 46, + 483, + 288, + 616 + ], + "type": "text", + "content": " deformable objects (Figure 2). Crucially, Total-Recon hierarchically decomposes scene motion into the motion of each object, which itself is decomposed into global root-body motion and local articulations. This key design choice scales our method to minute-long videos containing highly dynamic and deformable objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Background Radiance Field. We begin by modeling the background environment as a Neural Radiance Field (NeRF) [28]. For a 3D point " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{*} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " and a viewing direction " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{*}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " in the canonical world space, NeRF defines a color " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " represented by an MLP. We follow contemporary variants [26] that include a time-specific embedding code " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\omega_{e}^{t}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " to model illumination changes over time" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "content": "and model density with as a function of a neural signed distance function (SDF) " + }, + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{MLP}_{\\sigma}(\\cdot) = \\alpha \\Gamma_{\\beta}(\\mathbf{MLP}_{\\mathrm{SDF}}(\\cdot))" + }, + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "content": " [68] to encourage the reconstruction of a valid surface:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 117, + 546, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 117, + 546, + 131 + ], + "spans": [ + { + "bbox": [ + 317, + 117, + 546, + 131 + ], + "type": "interline_equation", + "content": "\\sigma = \\mathbf {M L P} _ {\\sigma} \\left(\\mathbf {X} ^ {*}\\right), \\quad \\mathbf {c} ^ {t} = \\mathbf {M L P} _ {\\mathbf {c}} \\left(\\mathbf {X} ^ {*}, \\mathbf {v} ^ {*}, \\omega_ {e} ^ {t}\\right). \\tag {2}", + "image_path": "6d36f2ff68d20b490fa26dce8345ae3ea5fd6086d4213a0498fdb86e3f41e061.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 140, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 140, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 545, + 163 + ], + "type": "text", + "content": "The pixel color can then be computed with differentiable volume rendering equations (Section 3.3)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 164, + 546, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 164, + 546, + 259 + ], + "spans": [ + { + "bbox": [ + 304, + 164, + 546, + 259 + ], + "type": "text", + "content": "Most NeRF methods, including HyperNeRF [36] and " + }, + { + "bbox": [ + 304, + 164, + 546, + 259 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 304, + 164, + 546, + 259 + ], + "type": "text", + "content": " [60], assume images with known cameras. While our capture devices are equipped with inertial sensors, we find their self-reported camera poses have room for improvement. As such, we also model camera pose as an opti- mizable rigid-body transformation " + }, + { + "bbox": [ + 304, + 164, + 546, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_0^t\\in SE(3)" + }, + { + "bbox": [ + 304, + 164, + 546, + 259 + ], + "type": "text", + "content": " that maps points in a time-specific camera space " + }, + { + "bbox": [ + 304, + 164, + 546, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 164, + 546, + 259 + ], + "type": "text", + "content": " to the world space (where we assume homogenous notation):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 397, + 268, + 545, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 268, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 397, + 268, + 545, + 283 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {*} = \\mathbf {G} _ {0} ^ {t} \\mathbf {X} ^ {t}. \\tag {3}", + "image_path": "e674622cfe2d087c6f373a5d2dd5542b47410ba63457958a3ce6fb6c2c60dd7d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "text", + "content": "Deformable Field (for Object " + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "text", + "content": "). We model the deformable radiance field of object " + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "inline_equation", + "content": "j \\in \\{1, \\dots, M - 1\\}" + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "text", + "content": " with BANMo [67], which consists of a canonical rest shape and time-dependent deformation field. The canonical rest shape is represented by the same formulation described by Equation 2, but now defined in a local object-centric canonical space rather than the world space. BANMo represents object motion with a warping function " + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j^{t,\\leftarrow} : \\mathbf{X}^t \\rightarrow \\mathbf{X}_j^*" + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "text", + "content": " that maps the camera space points " + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t" + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "text", + "content": " to canonical space points " + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^*" + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "text", + "content": " with a rigid-body transformation " + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_j^t \\in SE(3)" + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "text", + "content": " and a deformation field " + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_j^{t,\\leftarrow}" + }, + { + "bbox": [ + 304, + 296, + 546, + 433 + ], + "type": "text", + "content": " modeled by linear blend skinning [13]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 350, + 450, + 545, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 450, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 350, + 450, + 545, + 467 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {j} ^ {*} = \\mathcal {W} _ {j} ^ {t, \\leftarrow} \\left(\\mathbf {X} ^ {t}\\right) = \\mathbf {J} _ {j} ^ {t, \\leftarrow} \\left(\\mathbf {G} _ {j} ^ {t} \\mathbf {X} ^ {t}\\right). \\tag {4}", + "image_path": "b39c2b7d43a0267c00c8089e5bcd99c20fd189497aa89022ecda0891206c5b19.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": "Note that our choice of deformation field differs from the " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": "-field used in HyperNeRF and " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": ", which has been shown to produce irregular deformation in the presence of complex scene motion [67]. Intuitively, rigid-body transformation " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_j^t" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": " captures the global root-body pose of object " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": " relative to the camera at time " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": ", while deformation field " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_j^{t,\\leftarrow}" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": " aligns more fine-grained articulations relative to its local canonical space (Figure 2). Explicitly disentangling these two sources of object motion (as opposed to conflating them) enables easier optimization of the deformation field, because local articulations are significantly easier to learn than those modeled relative to the world space (Equation 1). Furthermore, this motion decomposition makes the deformation field invariant to rigid-body transformations of the object. A motion model similar to ours was proposed by ST-NeRF [15], but their model encodes an object's global root-body motion with a 3D axis-aligned bounding box that does not explicitly represent object orientation, a prerequisite for embodied view synthesis from 3rd-person-follow cameras." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17674" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 281, + 129 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 281, + 129 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 281, + 129 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 281, + 129 + ], + "type": "table", + "html": "
Rendered features f at pixel xtCorresponding 3D features fij(Xt)
color帽子(xt)ct( Wj,←(Xt))
flow F( xt, t → t')Πt' (Wj', → (Wj, ←(Xt)) - xt
depth d( xt)[0, 0, 1] · Xt
", + "image_path": "54cc8b99de65c0224c31cc96927559fa548da8e795441d48c61efc1f3a389e8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "lines": [ + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "type": "text", + "content": "Table 2: Rendered 2D features " + }, + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}" + }, + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "type": "text", + "content": " and their corresponding 3D features " + }, + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ij}" + }, + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "type": "inline_equation", + "content": "\\Pi^{t'}" + }, + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "type": "text", + "content": " denotes the camera intrinsics at time " + }, + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 47, + 137, + 287, + 161 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 176, + 288, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 176, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 176, + 288, + 237 + ], + "type": "text", + "content": "As did BANMo, Total-Recon also models a forward warp " + }, + { + "bbox": [ + 47, + 176, + 288, + 237 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j^t = \\mathcal{W}_j^{t,\\rightarrow}(\\mathbf{X}^*) = (\\mathbf{G}_j^t)^{-1}\\mathbf{J}_j^{t,\\rightarrow}(\\mathbf{X}^*)" + }, + { + "bbox": [ + 47, + 176, + 288, + 237 + ], + "type": "text", + "content": " that maps the canonical space to the camera space, which is used to establish the surface correspondences required for egocentric view synthesis and 3D video filters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 245, + 265, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 245, + 265, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 245, + 265, + 259 + ], + "type": "text", + "content": "3.3. Composite Rendering of Multiple Objects" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "text", + "content": "Given a set of " + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "text", + "content": " object representations (the background is treated as an object as well), we use the composite rendering scheme from prior work [31, 52] to combine the outputs of all object representations and volume-render the entire scene. To volumetrically render the image at frame " + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "text", + "content": ", we sample multiple points along each camera ray " + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^t" + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "text", + "content": ". Denoting the " + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "text", + "content": " sample as " + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i^t" + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "text", + "content": ", we write the density and color observed at sample " + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "text", + "content": " due to object " + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 264, + 287, + 361 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 369, + 288, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 288, + 386 + ], + "type": "interline_equation", + "content": "\\sigma_ {i j} = \\mathbf {M L P} _ {\\sigma , j} \\left(\\mathbf {X} _ {i j} ^ {*}\\right), \\qquad \\mathbf {c} _ {i j} = \\mathbf {M L P} _ {\\mathbf {c}, j} \\left(\\mathbf {X} _ {i j} ^ {*}, \\mathbf {v} _ {j} ^ {*}, \\boldsymbol {\\omega} _ {e} ^ {t}\\right),", + "image_path": "e1226378c11857e3ac75a81a4560d4c6dfbaf434b31483386a3937d3bca9d4b6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{ij}^{*} = \\mathcal{W}_{j}^{t,\\leftarrow}(\\mathbf{X}_{i}^{t})" + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_j^* = \\mathcal{W}_j^{t,\\leftarrow}(\\mathbf{v}^t)" + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "text", + "content": " are sample " + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "text", + "content": " and camera ray " + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^t" + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "text", + "content": " backward-warped into object " + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "text", + "content": "'s canonical space, respectively. The composite density " + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "text", + "content": " at sample " + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "text", + "content": " along the ray is then computed as the sum of each object's density " + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\sigma_{ij}" + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "text", + "content": "; the composite color " + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "text", + "content": " is computed as the weighted sum of each object's color " + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{ij}" + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "text", + "content": ", where the weights are the normalized object densities " + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "inline_equation", + "content": "\\sigma_{ij} / \\sigma_i" + }, + { + "bbox": [ + 46, + 394, + 287, + 481 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 490, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 490, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 88, + 490, + 287, + 525 + ], + "type": "interline_equation", + "content": "\\sigma_ {i} = \\sum_ {j = 0} ^ {M - 1} \\sigma_ {i j}, \\quad \\mathbf {c} _ {i} = \\frac {1}{\\sigma_ {i}} \\sum_ {j = 0} ^ {M - 1} \\sigma_ {i j} \\mathbf {c} _ {i j}. \\tag {5}", + "image_path": "28668f3b98df153c9dc4679d2519b2954618cf34e7047ebe594f8e89c483e2a3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "text", + "content": "We can then use the standard volume rendering equations to generate an RGB image of the scene, where " + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "text", + "content": " is the number of sampled points along camera ray " + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^t" + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\tau_{i}" + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "text", + "content": " is the transmittance, " + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "text", + "content": " is the alpha value for sample point " + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "text", + "content": " is the distance between sample point " + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 47, + 534, + 287, + 594 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 50, + 613, + 283, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 613, + 283, + 647 + ], + "spans": [ + { + "bbox": [ + 50, + 613, + 283, + 647 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {c}} = \\sum_ {i = 1} ^ {N} \\tau_ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\quad \\tau_ {i} = \\prod_ {k = 1} ^ {i - 1} (1 - \\alpha_ {k}), \\quad \\alpha_ {i} = 1 - e ^ {- \\sigma_ {i} \\delta_ {i}}.", + "image_path": "ad8aed81f8eb4d2552c535d7c4891a215c83fc643c88bf66db60790b81e2eb37.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "Rendering Flow, Depth, and Silhouettes. Our composite rendering scheme can be used to render different quantities by replacing the object color " + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{ij}" + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": " in Equation 5 with the" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 73, + 545, + 292 + ], + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 292 + ], + "type": "image", + "image_path": "dee609955c7eb87a0593595a6c91eb580240b9426e3498568638cb241961a32e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 300, + 545, + 356 + ], + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 356 + ], + "type": "text", + "content": "Figure 3: 6-DOF Trajectories for Embodied View Synthesis. To synthesize embodied views from egocentric or actor-following cameras, Total-Recon reconstructs the entire background, every individual actor in the scene, as well as global 6-DOF trajectories of its root-body and its articulated body parts (e.g. head)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 384, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 433 + ], + "type": "text", + "content": "appropriately defined 3D feature " + }, + { + "bbox": [ + 304, + 384, + 545, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ij}" + }, + { + "bbox": [ + 304, + 384, + 545, + 433 + ], + "type": "text", + "content": " (Table 2) and rendering the resulting composite feature " + }, + { + "bbox": [ + 304, + 384, + 545, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 304, + 384, + 545, + 433 + ], + "type": "text", + "content": ". To render occlusion-aware object silhouettes, we follow ObSURF [52] to produce a categorical distribution over the " + }, + { + "bbox": [ + 304, + 384, + 545, + 433 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 384, + 545, + 433 + ], + "type": "text", + "content": " objects:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 456, + 545, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 331, + 456, + 545, + 488 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {o}} _ {j} = \\sum_ {i = 1} ^ {N} \\tau_ {i} \\alpha_ {i j}, \\quad \\text {w h e r e} \\quad \\tau_ {i} = \\prod_ {k = 1} ^ {i - 1} (1 - \\alpha_ {k}), \\tag {6}", + "image_path": "5c064c214f8db51fe996fad86625a25bbb05dd41253fbc8de1e8742b3775fb88.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 332, + 490, + 545, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 490, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 332, + 490, + 545, + 506 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = 1 - e ^ {- \\sigma_ {i} \\delta_ {i}}, \\quad \\alpha_ {i j} = 1 - e ^ {- \\sigma_ {i j} \\delta_ {i}}. \\tag {7}", + "image_path": "062d0797ae7f37fd45c151ffafb7a69e377bbbc60c93d564ff76b5dc25e9a7c7.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": "Losses. Given a monocular RGBD video, we optimize all parameters in our composite scene representation, which for each of the " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " objects includes the appearance and shape MLPs " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{MLP}_{c,j}, \\mathbf{MLP}_{\\sigma,j})" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ", rigid-body transformations " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_j^t" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ", and forward, backward deformation fields " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_j^{\\leftarrow}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_j^{\\rightarrow}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ". We optimize three reconstruction losses: a color loss " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ", a flow loss " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{flow}}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ", and a depth loss " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{depth}}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ", where the ground truth color " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " are provided by the RGBD video, and the \"ground truth\" flow " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " is computed by an off-the-shelf network [64]. The model also optimizes a 3D-cycle consistency loss " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cyc},j}" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " [67] for each deformable object to encourage their forward and backward warps to be consistent, where " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^t \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": " denotes pixel location at time " + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 557, + 546, + 713 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17675" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 92, + 287, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 92, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 61, + 92, + 287, + 118 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r g b}} = \\sum_ {\\mathbf {x} ^ {t}} \\left| \\left| \\mathbf {c} \\left(\\mathbf {x} ^ {t}\\right) - \\hat {\\mathbf {c}} \\left(\\mathbf {x} ^ {t}\\right) \\right| \\right| ^ {2}, \\tag {8}", + "image_path": "6a43df4cb45cf6084c7ec6b08c63951589f1679bf84859d52e56863b5a0a97a4.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 120, + 287, + 145 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 120, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 58, + 120, + 287, + 145 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f l o w}} = \\sum_ {\\mathbf {x} ^ {t}} \\left| \\left| \\mathcal {F} \\left(\\mathbf {x} ^ {t}\\right) - \\hat {\\mathcal {F}} \\left(\\mathbf {x} ^ {t}\\right) \\right| \\right| ^ {2}, \\tag {9}", + "image_path": "7cb17a01d8235b4098673134fd6e1e4f4e058b2cfb617bbccf13143fbb8570dd.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 148, + 287, + 173 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 287, + 173 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e p t h}} = \\sum_ {\\mathbf {x} ^ {t}} \\left| \\left| \\mathbf {d} \\left(\\mathbf {x} ^ {t}\\right) - \\hat {\\mathbf {d}} \\left(\\mathbf {x} ^ {t}\\right) \\right| \\right| ^ {2}, \\tag {10}", + "image_path": "8be9f29f7d9bf1f7d149e7e285e42da70c094b9358df50ee73639108ce369301.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 175, + 287, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 287, + 202 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c y c}, j} = \\sum_ {i} \\tau_ {i} \\alpha_ {i j} \\left\\| \\mathcal {W} _ {j} ^ {t ^ {\\prime}, \\rightarrow} \\left(\\mathcal {W} _ {j} ^ {t}, \\leftarrow (\\mathbf {X} _ {i} ^ {t})\\right) - \\mathbf {X} _ {i} ^ {t} \\right\\| ^ {2}. \\tag {11}", + "image_path": "2efae66ab9f24e4b1b2485430b0d13a191c1d630315352b1b081a20167a91499.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 219, + 287, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 287, + 278 + ], + "type": "text", + "content": "Initialization. We initialize the rigid-body transformations of each deformable object " + }, + { + "bbox": [ + 46, + 219, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_j^t" + }, + { + "bbox": [ + 46, + 219, + 287, + 278 + ], + "type": "text", + "content": " using a pre-trained category-specific PoseNet [67]; we initialize the rigid-body transformation of the background " + }, + { + "bbox": [ + 46, + 219, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_0^t" + }, + { + "bbox": [ + 46, + 219, + 287, + 278 + ], + "type": "text", + "content": " with the camera poses provided by the iPad Pro." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 296, + 289, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 289, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 289, + 453 + ], + "type": "text", + "content": "Embodied View Synthesis and 3D Filters. To enable embodied view synthesis and 3D video filters (Figure 3), we design a simple interface that allows a user to select a point on a target object's surface in its reconstructed canonical mesh, and use its forward warping function " + }, + { + "bbox": [ + 46, + 296, + 289, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j^{t,\\rightarrow}:\\mathbf{X}^* \\to \\mathbf{X}^t" + }, + { + "bbox": [ + 46, + 296, + 289, + 453 + ], + "type": "text", + "content": " followed by the rigid-body transformation " + }, + { + "bbox": [ + 46, + 296, + 289, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_0^t" + }, + { + "bbox": [ + 46, + 296, + 289, + 453 + ], + "type": "text", + "content": " to place the egocentric camera (or virtual 3D asset) in the world space. The surface normal to the object's mesh at the user-defined point provides a reference frame to align the egocentric camera's viewing direction and place the 3D asset. To implement a 3rd-person-follow camera, we add a user-defined offset to the object's local reference frame, which is defined by its root-body pose." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 465, + 128, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 128, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 128, + 478 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 653 + ], + "type": "text", + "content": "Implementation Details. In practice, we train our composite scene representation by first pre-training each object field separately. For deformable objects, we pre-train using a depth loss (Equation 10) combined with the losses optimized by BANMo [67]. This includes a silhouette loss " + }, + { + "bbox": [ + 46, + 486, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}} = \\sum_{\\mathbf{x}^t}||\\mathbf{o}_j(\\mathbf{x}^t) - \\hat{\\mathbf{o}}_j(\\mathbf{x}^t)||^2" + }, + { + "bbox": [ + 46, + 486, + 287, + 653 + ], + "type": "text", + "content": ", where the \"ground truth\" object silhouette " + }, + { + "bbox": [ + 46, + 486, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_j" + }, + { + "bbox": [ + 46, + 486, + 287, + 653 + ], + "type": "text", + "content": " is computed by an off-the-shelf instance segmentation engine [17]. For pre-training the background, we optimize color, flow, and depth losses (Equations 8, 9, 10) on pixels outside the ground truth object silhouettes. Importantly, we don't supervise the object fields on frames that are not provided an object silhouette since it cannot be determined whether the absence of detection is a true or false negative." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "After pre-training, we composite-render the pre-trained object fields and jointly finetune them using only the color, depth, flow, and object-specific 3D-cycle consistency losses. Since the silhouette loss is no longer used, the scene representation is supervised on all frames of the training se" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "quence during joint-finetuning. We provide a complete description of the implementation details in the supplement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 110, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 327 + ], + "type": "text", + "content": "Dataset. We evaluate Total-Recon on novel-view synthesis for deformable scenes. To enable quantitative evaluation, we built a stereo rig comprised of two iPad-Pros rigidly attached to a camera mount, a setup similar to that of Nerfies [35]. Using the stereo rig, we captured 11 RGBD sequences containing 3 different cats, 1 dog, and 2 human subjects in 4 different indoor environments. The RGBD videos were captured using the Record3D iOS App [49], which also automatically registers the frames captured by each camera. These video sequences, which were subsampled at 10 fps, range from 392 to 901 frames, amount to, on average minute-long videos that are significantly longer and contain more dynamic motion than the datasets introduced by [35, 36, 60, 11]. The left and right cameras were registered by solving a Perspective-n-Point (PnP) problem using manually annotated correspondences, and their videos were synchronized based on audio. We provide a complete description of our dataset in the supplement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 340, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 545, + 569 + ], + "type": "text", + "content": "Reconstruction and Applications. By hierarchically decomposing scene motion into the motion of each object, which itself is decomposed into root-body motion and local articulations, Total-Recon automatically computes novel 6-DoF trajectories such as those traversed by egocentric cameras and 3rd-person follow cameras (Figure 3). In turn, these trajectories enable automated embodied view synthesis and 3D occlusion-aware video filters (Figure 4). These tasks are also enabled by Total-Recon's ability to recover an accurate deformable 3D scene representation, which is currently out of reach for the best of related methods (Figure 5). As shown in the bird's eye view, each reconstructed object is properly situated with respect to the background and other objects, a direct consequence of our use of depth supervision. Furthermore, even though the iPad Pro can only measure depth up to " + }, + { + "bbox": [ + 304, + 340, + 545, + 569 + ], + "type": "inline_equation", + "content": "4\\mathrm{m}" + }, + { + "bbox": [ + 304, + 340, + 545, + 569 + ], + "type": "text", + "content": ", Total-Recon can render depth beyond this sensor limit by pooling the measurements from other frames into a single metric scene reconstruction. We provide results on additional sequences in the supplement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Baselines and Evaluation. In Figure 5 and Table 3, we compare Total-Recon to " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " [60] and HyperNeRF [36], and their depth-supervised equivalents on the proxy task of stereo-view synthesis, a prerequisite for embodied view synthesis: we train each method on the RGBD frames captured from the left camera of our dataset and evaluate the images rendered from the viewpoint of the right camera. The depth-supervised versions of the baselines contain the same depth loss used in Total-Recon. We report LPIPS [70] and the average (depth) accuracy at " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " [42] in all subsequent experiments, and we include a more complete set" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17676" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 534, + 485 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 534, + 485 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 534, + 485 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 534, + 485 + ], + "type": "image", + "image_path": "9e8a93237ca95f8a6992de4d2c5afcd32218436318eb7d4d9ba12bc36911a487.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 492, + 547, + 548 + ], + "lines": [ + { + "bbox": [ + 46, + 492, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 547, + 548 + ], + "type": "text", + "content": "Figure 4: Embodied View Synthesis and 3D Filters. For select sequences of our RGBD dataset, we visualize the scene geometry and appearance reconstructed by our method (3D reconstruction) and the resulting downstream applications. The yellow and blue camera meshes in the mesh renderings represent the egocentric and 3rd-person-follow cameras, respectively. To showcase the 3D video filter we attach a sky-blue unicorn horn to the forehead of the target object, which is then automatically propagated across all frames. Full-length videos can be found at https://andrewsonga.github.io/totalrecon/applications.html." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 554, + 289, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 289, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 289, + 625 + ], + "type": "text", + "content": "of metrics (PSNR, SSIM, RMS depth error) in the supplement. Because " + }, + { + "bbox": [ + 46, + 554, + 289, + 625 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 46, + 554, + 289, + 625 + ], + "type": "text", + "content": " and HyperNeRF were not designed to recover a metric scene representation, we replaced their COLMAP [46] camera poses with those provided by the iPad Pro (which are metric measurements) for the sake of fair comparison." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": "Comparisons. Total-Recon qualitatively and quantitatively outperforms all of the baselines. As shown in Figure 5, Total-Recon successfully reconstructs the entire scene, whereas the baselines are only able to reconstruct the rigid background at best. As shown in Table 3, Total-Recon significantly outperforms all baselines in terms of LPIPS and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 554, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 547, + 604 + ], + "type": "text", + "content": "the average accuracy at " + }, + { + "bbox": [ + 304, + 554, + 547, + 604 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 304, + 554, + 547, + 604 + ], + "type": "text", + "content": " (Acc@0.1m). We attribute this huge gap to the baselines' inability to reconstruct highly dynamic objects. We provide more details regarding the baselines and additional visualizations in the supplement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 610, + 407, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 407, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 407, + 623 + ], + "type": "text", + "content": "4.1. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Table 4 (and Figures 6 and 7) analyzes the importance of Total-Recon's design choices (see Section 3) by ablating its key components: the depth loss " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{depth}}" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " (row 2), the deformation field " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_j^t" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " (row 3), PoseNet-initialization of the root-body pose (row 4), and the root-body pose " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_j^t" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " itself (row 5), where " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " denotes a deformable actor. For all ablations, we use the same set of training losses used in Total-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17677" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 71, + 542, + 139 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 542, + 139 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 542, + 139 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 542, + 139 + ], + "type": "table", + "html": "
DOG 1 (626 images)DOG 1 (v2) (531 images)CAT 1 (641 images)CAT 1 (v2) (632 images)CAT 2 (632 images) (834 images)CAT 2 (v2) (901 images)CAT 3 (767 images)HUMAN 1 (550 images)HUMAN 2 (483 images)HUMAN - DOG (392 images)HUMAN - CAT (431 images)MEAN
LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑LPIPS↓Acc↑
HyperNeRF.634.107.432.176.521.316.438.314.641.277.397.252.592.213.632.053.585.067.487.072.462.162.531.198
D2NeRF.540.219.546.220.687.346.588.403.556.333.595.339.759.231.588.066.630.128.576.078.628.126.611.247
HyperNeRF (+depth).373.352.425.357.532.552.371.596.330.605.376.612.514.451.501.211.445.249.450.283.456.214.428.439
D2NeRF (+depth).507.338.532.270.685.510.580.362.561.438.553.376.730.243.585.086.609.131.608.154.645.176.599.302
Total-Recon.271.841.313.790.382.889.333.894.237.967.281.925.261.949.213.909.264.849.256.827.233.914.278.895
", + "image_path": "055a6e13508eefc5facc78127dea3f7acf802e5bcb7d093f91a96cfa618dd142.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 204, + 286, + 386 + ], + "blocks": [ + { + "bbox": [ + 50, + 204, + 286, + 386 + ], + "lines": [ + { + "bbox": [ + 50, + 204, + 286, + 386 + ], + "spans": [ + { + "bbox": [ + 50, + 204, + 286, + 386 + ], + "type": "image", + "image_path": "86e2c39980dc895e8c6568960f3e9a74f94daf6a9704a87d1ed1a17bd2fcc4d3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 395, + 288, + 473 + ], + "lines": [ + { + "bbox": [ + 46, + 395, + 288, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 288, + 473 + ], + "type": "text", + "content": "Figure 5: Baseline Comparisons. We compare Total-Recon to depth-supervised variants of HyperNeRF [36] and " + }, + { + "bbox": [ + 46, + 395, + 288, + 473 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 46, + 395, + 288, + 473 + ], + "type": "text", + "content": " [60] on the task of stereo-view synthesis (the left camera is used for training and the right is used for testing). While the baselines are able to reconstruct only the background, Total-Recon can reconstruct both the background and the moving deformable object(s), demonstrating holistic scene reconstruction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 489, + 287, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 287, + 596 + ], + "type": "text", + "content": "Recon and initialize camera pose " + }, + { + "bbox": [ + 46, + 489, + 287, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_0^t" + }, + { + "bbox": [ + 46, + 489, + 287, + 596 + ], + "type": "text", + "content": " with those reported by ARKit. For ablations that model root-body motions, we initialize each deformable actor's root-body pose " + }, + { + "bbox": [ + 46, + 489, + 287, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_j^t" + }, + { + "bbox": [ + 46, + 489, + 287, + 596 + ], + "type": "text", + "content": " with predictions made by PoseNet [67] and optimize them during reconstruction; for row 4, we replace the PoseNet predictions with identity rotations. We report the novel-view metrics averaged over 6 select sequences of our dataset: DOG 1 (v1), CAT 1 (v1), CAT 2 (v1), HUMAN 1, HUMAN 1 & DOG 1, and HUMAN 2 & CAT 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Depth Supervision. Table 4 shows that removing depth supervision (row 2) significantly reduces the average accuracy at " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{m}" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " (Acc). Figure 6 indicates that this reflects the incorrect arrangement of objects stemming from their scale inconsistency - while removing depth supervision does not significantly deteriorate the training-view RGB renderings, it induces critical failure modes as shown in the novel-view 3D reconstructions: (a) floating foreground objects, as evi" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 313, + 200, + 539, + 274 + ], + "blocks": [ + { + "bbox": [ + 46, + 148, + 547, + 183 + ], + "lines": [ + { + "bbox": [ + 46, + 148, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 547, + 183 + ], + "type": "text", + "content": "Table 3: Baseline Comparisons. We train Total-Recon, HyperNeRF [36], D²NeRF [60], and their depth-supervised variants on the left video captured with our stereo rig, and evaluate the novel view synthesis results on the held-out right video. Total-Recon significantly outperforms all of the baselines for all 11 sequences. These sequences are sampled at 10 fps, amounting to minute-long videos, on average." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 200, + 539, + 274 + ], + "lines": [ + { + "bbox": [ + 313, + 200, + 539, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 539, + 274 + ], + "type": "table", + "html": "
MethodsDepth LossDeform. Obj.Root Init.Root MotionLPIPS↓Acc@0.1m↑
(1) Full model✓✓✓✓.268.898
(2) w/o loss LdepthX✓✓✓.372.154
(3) w/o deform. Jj✓X✓✓.296.867
(4) w/o root-body init.✓✓X✓.293.870
(5) w/o root-body Gj✓✓XXN/AN/A
", + "image_path": "089512a22f661b1412e2a5ae194d7eaf1dfc078d00c46ecb0f3a45b7085f643b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 282, + 545, + 392 + ], + "lines": [ + { + "bbox": [ + 305, + 282, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 545, + 392 + ], + "type": "text", + "content": "Table 4: Ablation Study. Removing depth supervision (2) significantly hurts performance, while removing the deformation field (3) and PoseNet-initialization of root-body poses (4) hurts moderately. Most importantly, removing root-body poses entirely (5) prevents convergence (N/A) as the deformation field alone has to explain global object motion (see Figure 2). These experiments justify our hierarchical modeling of motion, as even root-bodies without a deformation field (3) or poorly initialized root-bodies (4) are better than no root-bodies (5). We visualize these ablations in Figure 7 and explore other ablations in the supplement." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 416, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 416, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 305, + 416, + 545, + 489 + ], + "type": "text", + "content": "denced by their shadows, and (b) the human incorrectly occluding the dog. In other words, without depth supervision, Total-Recon overfits the training view and learns a degenerate scene representation where the reconstructed objects fail to converge to the same scale. We show results on additional sequences in the supplement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": "Motion Modeling. Table 4 shows that removing the deformation field (row 3) also hurts performance. This is because, without the deformation field, our method has to explain an object's non-rigid motion solely with its rigid, root-body poses. As a result, this ablation can only recover coarse object reconstructions that fail to model moving body parts such as limbs. Removing PoseNet-initialization of root-body poses (row 4) is just as detrimental, resulting in noisy appearance and geometry artifacts; see Figure 7 and additional visuals in the supplement. Most notably, Table 4 shows that removing object root-bodies entirely (row 5) causes the optimization to fail to converge (N/A), even though the deformation field can (in theory) represent all continuous motion. It appears difficult for deformation fields alone to explain global root-body motion because such motions can deviate significantly from a canonical model, complicating optimization." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17678" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 286, + 200 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 286, + 200 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 286, + 200 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 286, + 200 + ], + "type": "image", + "image_path": "6cf3c6572a73a8bf1922c93283acb8d76d41ea7ba320ab4494f4cc8c3d316329.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 287, + 299 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 299 + ], + "type": "text", + "content": "Figure 6: Ablation Study on Depth Supervision. While removing depth supervision does not significantly deteriorate the training-view RGB renderings, it significantly hurts the novel-view 3D reconstructions, as characterized by the following: (a) floating foreground objects (as evidenced by their shadows) and (b) the human incorrectly occluding the dog. These failure modes indicate that without depth supervision, Total-Recon overfits the training view, and the reconstructed objects fail to converge to the same scale." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 320, + 287, + 466 + ], + "blocks": [ + { + "bbox": [ + 48, + 320, + 287, + 466 + ], + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 466 + ], + "type": "image", + "image_path": "c03c6cd8035ea19355438b7fbc3e13dbf5abbc226a7546780f83999cf555a483.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 474, + 287, + 574 + ], + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 574 + ], + "type": "text", + "content": "Figure 7: Ablation Study. We visualize the ablations from Table 4. Removing the depth loss causes the cat to sink into the ground due to inconsistent object scales. Removing the deformation field produces coarse object reconstructions that fail to capture articulated body parts such as moving limbs. Removing PoseNet-initialization of root-body poses results in noisier appearance and geometry, as shown by the human actor's left hand. We omit the ablation without root-body poses as it does not converge. We present additional visualizations in the supplement." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "These diagnostics justify Total-Recon's hierarchical motion representation, which explicitly models objects' root-body poses; root-bodies without articulated deformations (row 3) or poorly initialized root-bodies (row 4) are better than no root-bodies at all (row 5). Our ablations also suggest that the poor performance of the baseline methods (on our challenging dataset) may be attributed to the lack of object-centric motion modeling. We provide a more detailed analysis with additional experiments and RGBD sequences in the supplement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 71, + 460, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 460, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 460, + 84 + ], + "type": "text", + "content": "5. Discussion and Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 91, + 547, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 259 + ], + "type": "text", + "content": "We have presented a new system for automated embodied view synthesis from monocular RGBD videos, focusing on videos of people interacting with their pets. Our main technical contribution is Total-Recon, a 3D representation for deformable scenes that hierarchically decomposes scene motion into the motion of each object, which is further decomposed into its root-body motion and local articulations; this key design choice enables appropriate initialization of root-body poses and hence easier optimization over long videos containing large motions. By explicitly reconstructing the geometry, appearance, root-body- and articulated motion of each object, Total-Recon enables seeing through the eyes of people and pets and generating game-like traversals of deformable scenes from behind a target actor." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 275, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 547, + 491 + ], + "type": "text", + "content": "Limitations. In Total-Recon, scene decomposition is primarily supervised by object silhouettes computed by an off-the-shelf segmentation model [17], which may be inaccurate, especially in partial occlusion scenarios. This may damage the resulting reconstructions and embodied-view renderings. We believe that incorporating the latest advances in video instance segmentation [12] will enable Total-Recon to be applied to more challenging scenarios. Second, Total-Recon initializes the root-body pose of each deformable object using a PoseNet [67] trained for humans and quadruped animals, which does not generalize to other object categories (e.g., birds, fish). We reserve the reconstruction of generic scenes for future work. Finally, our model needs to be optimized on a per-sequence basis for roughly 15 hours with 4 NVIDIA RTX A5000 GPUs and is therefore not suitable for real-time applications. Incorporating recent advances in fast neural field training methods is an interesting avenue for future research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 507, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 547, + 604 + ], + "type": "text", + "content": "Acknowledgments. We thank Nathaniel Chodosh, Jeff Tan, George Cazenavette, and Jason Zhang for proofreading and Songwei Ge for reviewing the codebase. We also thank Sheng-Yu Wang, Daohan (Fred) Lu, Tamaki Kojima, Krishna Wadhwani, Takuya Narihira, and Tatsuo Fujiwara for providing valuable feedback. This work is supported in part by the Sony Corporation, Cisco Research, and the CMU Argo AI Center for Autonomous Vehicle Research." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 615, + 365, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 615, + 365, + 627 + ], + "spans": [ + { + "bbox": [ + 306, + 615, + 365, + 627 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 635, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 310, + 635, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 635, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 310, + 635, + 545, + 677 + ], + "type": "text", + "content": "[1] Dejan Azinović, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 680, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 680, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 310, + 680, + 547, + 713 + ], + "type": "text", + "content": "[2] Aljaz Bozic, Pablo Palafox, Michael Zollhofer, Justus Thies, Angela Dai, and Matthias Nießner. Neural deformation graphs for globally-consistent non-rigid reconstruction. In" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17679" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 95, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 95, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 95, + 287, + 148 + ], + "type": "text", + "content": "[3] Aljaz Bozic, Michael Zollhofer, Christian Theobalt, and Matthias Nießner. Deepdeform: Learning non-rigid rgb-d reconstruction with semi-supervised data. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 150, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 288, + 194 + ], + "type": "text", + "content": "[4] Hongrui Cai, Wanquan Feng, Xuetao Feng, Yan Wang, and Juyong Zhang. Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 195, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 288, + 227 + ], + "type": "text", + "content": "[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "text", + "content": "[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 326 + ], + "type": "text", + "content": "[7] Alena Denisova and Paul Cairns. First person vs. third person perspective in digital games: Do player preferences affect immersion? In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, page 145-148, 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 327, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 288, + 370 + ], + "type": "text", + "content": "[8] Mingsong Dou, Jonathan Taylor, Henry Fuchs, Andrew Fitzgibbon, and Shahram Izadi. 3d scanning deformable objects with a single rgbd sensor. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 371, + 288, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 371, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 53, + 371, + 288, + 425 + ], + "type": "text", + "content": "[9] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbaek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "type": "text", + "content": "[10] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 470, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 287, + 514 + ], + "type": "text", + "content": "[11] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 557 + ], + "type": "text", + "content": "[12] De-An Huang, Zhiding Yu, and Anima Anandkumar. Minvis: A minimal video instance segmentation framework without video-based training. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 9" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "text", + "content": "[13] Alec Jacobson, Zhigang Deng, Ladislav Kavan, and J. P. Lewis. Skinning: Real-time shape deformation (full text not available). In ACM SIGGRAPH 2014 Courses, SIGGRAPH '14. Association for Computing Machinery, 2014. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 646 + ], + "type": "text", + "content": "[14] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[15] Zhang Jiakai, Liu Xinhang, Ye Xinyi, Zhao Fuqiang, Zhang Yanshun, Wu Minye, Zhang Yingliang, Xu Lan, and Yu Jingyi. Editable free-viewpoint video using a layered neural representation. In ACM SIGGRAPH, 2021. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "from a single video. In European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[17] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6, 9" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 206 + ], + "type": "text", + "content": "[18] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J. Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 208, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 252 + ], + "type": "text", + "content": "[19] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 254, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 319 + ], + "type": "text", + "content": "[20] Tianye Li, Mira Slavcheva, Michael Zollhöfer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 320, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 364 + ], + "type": "text", + "content": "[21] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "text", + "content": "[22] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "text", + "content": "[23] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 498 + ], + "type": "text", + "content": "[24] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM SIGGRAPH Asia, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "text", + "content": "[25] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "text", + "content": "[26] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[27] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 3, 4" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17680" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 183 + ], + "type": "text", + "content": "[30] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE International Symposium on Mixed and Augmented Reality, pages 127-136. IEEE, 2011. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "text", + "content": "[31] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 260 + ], + "type": "text", + "content": "[32] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "type": "text", + "content": "[33] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 288, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 288, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 288, + 349 + ], + "type": "text", + "content": "[34] Kemal Egemen Ozden, Kurt Cornelis, Luc Van Eycken, and Luc Van Gool. Reconstructing 3d trajectories of independently moving objects using generic constraints. Comput. Vis. Image Underst., 96(3):453-471, 2004. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "type": "text", + "content": "[35] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 470 + ], + "type": "text", + "content": "[36] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics (TOG), 40(6),2021.2,3,4,6,8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "text", + "content": "[37] Georgios Pavlakos*, Ethan Weber*, , Matthew Tancik, and Angjoo Kanazawa. The one where they reconstructed 3d humans and environments in tv shows. In European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 288, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 569 + ], + "type": "text", + "content": "[38] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In IEEE International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 635 + ], + "type": "text", + "content": "[39] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "text", + "content": "[40] F. Peruzzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "for Dynamic Scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[42] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[43] Zhongzheng Ren, Xiaoming Zhao, and Alex Schwing. Class-agnostic reconstruction of dynamic objects from videos. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[44] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 274 + ], + "type": "text", + "content": "[45] Richard Rouse. What's your perspective? ACM SIGGRAPH Computer Graphics, 33(3):9-12, August 1999. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[46] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 364 + ], + "type": "text", + "content": "[47] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 410 + ], + "type": "text", + "content": "[48] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In SIGGRAPH Conference Proceedings, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 411, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 431 + ], + "type": "text", + "content": "[49] Marek Simonik. Record3d. https://record3d.app. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 434, + 545, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 478 + ], + "type": "text", + "content": "[50] Miroslava Slavcheva, Maximilian Baust, Daniel Cremers, and Slobodan Ilic. Killingfusion: Non-rigid 3d reconstruction without correspondences. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 479, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 522 + ], + "type": "text", + "content": "[51] Miroslava Slavcheva, Maximilian Baust, and Slobodan Ilic. Sobolevfusion: 3d reconstruction of scenes undergoing free non-rigid motion. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 523, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 556 + ], + "type": "text", + "content": "[52] Karl Stelzner, Kristian Kersting, and Adam R Kosiorek. Decomposing 3d scenes into objects via unsupervised volume segmentation. arXiv preprint arXiv:2104.01148, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "text", + "content": "[53] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-nerf: Articulated neural radiance fields for learning human shape, appearance, and pose. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 602, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 646 + ], + "type": "text", + "content": "[54] Unity Technologies. 3rd person follow. https://docs.unity3d.com/Packages/ com.unity.cinemachine@2.8/manual/ Cinemachine3rdPersonFollow.html.1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17681" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[56] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "type": "text", + "content": "[57] Barbara Tversky. Spatial Cognition Embodied and Situated. Cambridge, 2008. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 140, + 286, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 140, + 286, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 286, + 171 + ], + "type": "text", + "content": "[58] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 173, + 286, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 173, + 286, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 173, + 286, + 215 + ], + "type": "text", + "content": "[59] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 217, + 286, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 286, + 270 + ], + "type": "text", + "content": "[60] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. " + }, + { + "bbox": [ + 49, + 217, + 286, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^2" + }, + { + "bbox": [ + 49, + 217, + 286, + 270 + ], + "type": "text", + "content": " nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 273, + 286, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 273, + 286, + 315 + ], + "spans": [ + { + "bbox": [ + 49, + 273, + 286, + 315 + ], + "type": "text", + "content": "[61] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 317, + 286, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 317, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 317, + 286, + 358 + ], + "type": "text", + "content": "[62] Jing Xiao, Jinxiang Chai, and Takeo Kanade. A closed-form solution to non-rigid shape and motion recovery. International Journal of Computer Vision (IJCV), 67(2):233-246, April 2006. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 361, + 286, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 361, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 286, + 403 + ], + "type": "text", + "content": "[63] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 405, + 286, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 405, + 286, + 436 + ], + "spans": [ + { + "bbox": [ + 49, + 405, + 286, + 436 + ], + "type": "text", + "content": "[64] Gengshan Yang and Deva Ramanan. Volumetric correspondence networks for optical flow. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 438, + 286, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 438, + 286, + 501 + ], + "spans": [ + { + "bbox": [ + 49, + 438, + 286, + 501 + ], + "type": "text", + "content": "[65] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T Freeman, and Ce Liu. LASR: Learning articulated shape reconstruction from a monocular video. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 504, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 504, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 49, + 504, + 286, + 557 + ], + "type": "text", + "content": "[66] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 559, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 559, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 49, + 559, + 286, + 612 + ], + "type": "text", + "content": "[67] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 4, 5, 6, 8, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 614, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 286, + 646 + ], + "type": "text", + "content": "[68] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 286, + 689 + ], + "type": "text", + "content": "[69] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 692, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 692, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 692, + 286, + 712 + ], + "type": "text", + "content": "[70] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 150 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[71] Xiaoshuai Zhang, Abhijit Kundu, Thomas Funkhouser, Leonidas Guibas, Hao Su, and Kyle Genova. Nerflets: Local radiance fields for efficient structure-aware 3d scene representation from 2d supervision. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17682" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file